1 /* SPDX-License-Identifier: MIT */
2 /*
3  * Copyright 2023 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  *
23  * Authors: AMD
24  *
25  */
26 
27 #include "display_mode_core.h"
28 #include "display_mode_util.h"
29 #include "display_mode_lib_defines.h"
30 
31 #include "dml_assert.h"
32 
33 #define DML2_MAX_FMT_420_BUFFER_WIDTH 4096
34 #define TB_BORROWED_MAX 400
35 
36 // ---------------------------
37 //  Declaration Begins
38 // ---------------------------
39 static void CalculateBytePerPixelAndBlockSizes(
40 	enum dml_source_format_class SourcePixelFormat,
41 	enum dml_swizzle_mode SurfaceTiling,
42 	// Output
43 	dml_uint_t *BytePerPixelY,
44 	dml_uint_t *BytePerPixelC,
45 	dml_float_t *BytePerPixelDETY,
46 	dml_float_t *BytePerPixelDETC,
47 	dml_uint_t *BlockHeight256BytesY,
48 	dml_uint_t *BlockHeight256BytesC,
49 	dml_uint_t *BlockWidth256BytesY,
50 	dml_uint_t *BlockWidth256BytesC,
51 	dml_uint_t *MacroTileHeightY,
52 	dml_uint_t *MacroTileHeightC,
53 	dml_uint_t *MacroTileWidthY,
54 	dml_uint_t *MacroTileWidthC);
55 
56 static dml_float_t CalculateWriteBackDISPCLK(
57 	enum dml_source_format_class WritebackPixelFormat,
58 	dml_float_t PixelClock,
59 	dml_float_t WritebackHRatio,
60 	dml_float_t WritebackVRatio,
61 	dml_uint_t WritebackHTaps,
62 	dml_uint_t WritebackVTaps,
63 	dml_uint_t WritebackSourceWidth,
64 	dml_uint_t WritebackDestinationWidth,
65 	dml_uint_t HTotal,
66 	dml_uint_t WritebackLineBufferSize,
67 	dml_float_t DISPCLKDPPCLKVCOSpeed);
68 
69 static void CalculateVMRowAndSwath(
70 	struct display_mode_lib_scratch_st *s,
71 	struct CalculateVMRowAndSwath_params_st *p);
72 
73 static void CalculateOutputLink(
74 	dml_float_t PHYCLKPerState,
75 	dml_float_t PHYCLKD18PerState,
76 	dml_float_t PHYCLKD32PerState,
77 	dml_float_t Downspreading,
78 	dml_bool_t IsMainSurfaceUsingTheIndicatedTiming,
79 	enum dml_output_encoder_class Output,
80 	enum dml_output_format_class OutputFormat,
81 	dml_uint_t HTotal,
82 	dml_uint_t HActive,
83 	dml_float_t PixelClockBackEnd,
84 	dml_float_t ForcedOutputLinkBPP,
85 	dml_uint_t DSCInputBitPerComponent,
86 	dml_uint_t NumberOfDSCSlices,
87 	dml_float_t AudioSampleRate,
88 	dml_uint_t AudioSampleLayout,
89 	enum dml_odm_mode ODMModeNoDSC,
90 	enum dml_odm_mode ODMModeDSC,
91 	enum dml_dsc_enable DSCEnable,
92 	dml_uint_t OutputLinkDPLanes,
93 	enum dml_output_link_dp_rate OutputLinkDPRate,
94 
95 	// Output
96 	dml_bool_t *RequiresDSC,
97 	dml_bool_t *RequiresFEC,
98 	dml_float_t *OutBpp,
99 	enum dml_output_type_and_rate__type *OutputType,
100 	enum dml_output_type_and_rate__rate *OutputRate,
101 	dml_uint_t *RequiredSlots);
102 
103 static void CalculateODMMode(
104 	dml_uint_t MaximumPixelsPerLinePerDSCUnit,
105 	dml_uint_t HActive,
106 	enum dml_output_encoder_class Output,
107 	enum dml_output_format_class OutputFormat,
108 	enum dml_odm_use_policy ODMUse,
109 	dml_float_t StateDispclk,
110 	dml_float_t MaxDispclk,
111 	dml_bool_t DSCEnable,
112 	dml_uint_t TotalNumberOfActiveDPP,
113 	dml_uint_t MaxNumDPP,
114 	dml_float_t PixelClock,
115 	dml_float_t DISPCLKDPPCLKDSCCLKDownSpreading,
116 	dml_float_t DISPCLKRampingMargin,
117 	dml_float_t DISPCLKDPPCLKVCOSpeed,
118 	dml_uint_t NumberOfDSCSlices,
119 
120 	// Output
121 	dml_bool_t *TotalAvailablePipesSupport,
122 	dml_uint_t *NumberOfDPP,
123 	enum dml_odm_mode *ODMMode,
124 	dml_float_t *RequiredDISPCLKPerSurface);
125 
126 static dml_float_t CalculateRequiredDispclk(
127 	enum dml_odm_mode ODMMode,
128 	dml_float_t PixelClock,
129 	dml_float_t DISPCLKDPPCLKDSCCLKDownSpreading,
130 	dml_float_t DISPCLKRampingMargin,
131 	dml_float_t DISPCLKDPPCLKVCOSpeed,
132 	dml_float_t MaxDispclkSingle);
133 
134 static void CalculateSinglePipeDPPCLKAndSCLThroughput(
135 	dml_float_t HRatio,
136 	dml_float_t HRatioChroma,
137 	dml_float_t VRatio,
138 	dml_float_t VRatioChroma,
139 	dml_float_t MaxDCHUBToPSCLThroughput,
140 	dml_float_t MaxPSCLToLBThroughput,
141 	dml_float_t PixelClock,
142 	enum dml_source_format_class SourcePixelFormat,
143 	dml_uint_t HTaps,
144 	dml_uint_t HTapsChroma,
145 	dml_uint_t VTaps,
146 	dml_uint_t VTapsChroma,
147 
148 	// Output
149 	dml_float_t *PSCL_THROUGHPUT,
150 	dml_float_t *PSCL_THROUGHPUT_CHROMA,
151 	dml_float_t *DPPCLKUsingSingleDPP);
152 
153 static void CalculateDPPCLK(
154 	dml_uint_t NumberOfActiveSurfaces,
155 	dml_float_t DISPCLKDPPCLKDSCCLKDownSpreading,
156 	dml_float_t DISPCLKDPPCLKVCOSpeed,
157 	dml_float_t DPPCLKUsingSingleDPP[],
158 	dml_uint_t DPPPerSurface[],
159 
160 	// Output
161 	dml_float_t *GlobalDPPCLK,
162 	dml_float_t Dppclk[]);
163 
164 static void CalculateMALLUseForStaticScreen(
165 	dml_uint_t NumberOfActiveSurfaces,
166 	dml_uint_t MALLAllocatedForDCNFinal,
167 	enum dml_use_mall_for_static_screen_mode *UseMALLForStaticScreen,
168 	dml_uint_t SurfaceSizeInMALL[],
169 	dml_bool_t one_row_per_frame_fits_in_buffer[],
170 
171 	// Output
172 	dml_bool_t UsesMALLForStaticScreen[]);
173 
174 static dml_uint_t dscceComputeDelay(
175 	dml_uint_t bpc,
176 	dml_float_t BPP,
177 	dml_uint_t sliceWidth,
178 	dml_uint_t numSlices,
179 	enum dml_output_format_class pixelFormat,
180 	enum dml_output_encoder_class Output);
181 
182 static dml_uint_t dscComputeDelay(enum dml_output_format_class pixelFormat,
183 	enum dml_output_encoder_class Output);
184 
185 static dml_bool_t CalculatePrefetchSchedule(struct display_mode_lib_scratch_st *scratch,
186 	struct CalculatePrefetchSchedule_params_st *p);
187 
188 static dml_float_t RoundToDFSGranularity(dml_float_t Clock, dml_bool_t round_up, dml_float_t VCOSpeed);
189 
190 static void CalculateDCCConfiguration(
191 	dml_bool_t DCCEnabled,
192 	dml_bool_t DCCProgrammingAssumesScanDirectionUnknown,
193 	enum dml_source_format_class SourcePixelFormat,
194 	dml_uint_t SurfaceWidthLuma,
195 	dml_uint_t SurfaceWidthChroma,
196 	dml_uint_t SurfaceHeightLuma,
197 	dml_uint_t SurfaceHeightChroma,
198 	dml_uint_t nomDETInKByte,
199 	dml_uint_t RequestHeight256ByteLuma,
200 	dml_uint_t RequestHeight256ByteChroma,
201 	enum dml_swizzle_mode TilingFormat,
202 	dml_uint_t BytePerPixelY,
203 	dml_uint_t BytePerPixelC,
204 	dml_float_t BytePerPixelDETY,
205 	dml_float_t BytePerPixelDETC,
206 	enum dml_rotation_angle SourceScan,
207 	// Output
208 	dml_uint_t *MaxUncompressedBlockLuma,
209 	dml_uint_t *MaxUncompressedBlockChroma,
210 	dml_uint_t *MaxCompressedBlockLuma,
211 	dml_uint_t *MaxCompressedBlockChroma,
212 	dml_uint_t *IndependentBlockLuma,
213 	dml_uint_t *IndependentBlockChroma);
214 
215 static dml_uint_t CalculatePrefetchSourceLines(
216 	dml_float_t VRatio,
217 	dml_uint_t VTaps,
218 	dml_bool_t Interlace,
219 	dml_bool_t ProgressiveToInterlaceUnitInOPP,
220 	dml_uint_t SwathHeight,
221 	enum dml_rotation_angle SourceScan,
222 	dml_bool_t ViewportStationary,
223 	dml_uint_t SwathWidth,
224 	dml_uint_t ViewportHeight,
225 	dml_uint_t ViewportXStart,
226 	dml_uint_t ViewportYStart,
227 
228 	// Output
229 	dml_uint_t *VInitPreFill,
230 	dml_uint_t *MaxNumSwath);
231 
232 static dml_uint_t CalculateVMAndRowBytes(
233 	dml_bool_t ViewportStationary,
234 	dml_bool_t DCCEnable,
235 	dml_uint_t NumberOfDPPs,
236 	dml_uint_t BlockHeight256Bytes,
237 	dml_uint_t BlockWidth256Bytes,
238 	enum dml_source_format_class SourcePixelFormat,
239 	dml_uint_t SurfaceTiling,
240 	dml_uint_t BytePerPixel,
241 	enum dml_rotation_angle SourceScan,
242 	dml_uint_t SwathWidth,
243 	dml_uint_t ViewportHeight,
244 	dml_uint_t ViewportXStart,
245 	dml_uint_t ViewportYStart,
246 	dml_bool_t GPUVMEnable,
247 	dml_uint_t GPUVMMaxPageTableLevels,
248 	dml_uint_t GPUVMMinPageSizeKBytes,
249 	dml_uint_t PTEBufferSizeInRequests,
250 	dml_uint_t Pitch,
251 	dml_uint_t DCCMetaPitch,
252 	dml_uint_t MacroTileWidth,
253 	dml_uint_t MacroTileHeight,
254 
255 	// Output
256 	dml_uint_t *MetaRowByte,
257 	dml_uint_t *PixelPTEBytesPerRow,
258 	dml_uint_t *PixelPTEBytesPerRowStorage, // for PTE buffer size check
259 	dml_uint_t *dpte_row_width_ub,
260 	dml_uint_t *dpte_row_height,
261 	dml_uint_t *dpte_row_height_linear,
262 	dml_uint_t *PixelPTEBytesPerRow_one_row_per_frame,
263 	dml_uint_t *dpte_row_width_ub_one_row_per_frame,
264 	dml_uint_t *dpte_row_height_one_row_per_frame,
265 	dml_uint_t *MetaRequestWidth,
266 	dml_uint_t *MetaRequestHeight,
267 	dml_uint_t *meta_row_width,
268 	dml_uint_t *meta_row_height,
269 	dml_uint_t *PixelPTEReqWidth,
270 	dml_uint_t *PixelPTEReqHeight,
271 	dml_uint_t *PTERequestSize,
272 	dml_uint_t *DPDE0BytesFrame,
273 	dml_uint_t *MetaPTEBytesFrame);
274 
275 static dml_float_t CalculateTWait(
276 	dml_uint_t PrefetchMode,
277 	enum dml_use_mall_for_pstate_change_mode UseMALLForPStateChange,
278 	dml_bool_t SynchronizeDRRDisplaysForUCLKPStateChangeFinal,
279 	dml_bool_t DRRDisplay,
280 	dml_float_t DRAMClockChangeLatency,
281 	dml_float_t FCLKChangeLatency,
282 	dml_float_t UrgentLatency,
283 	dml_float_t SREnterPlusExitTime);
284 
285 static void CalculatePrefetchMode(
286 	enum dml_prefetch_modes AllowForPStateChangeOrStutterInVBlank,
287 	dml_uint_t *MinPrefetchMode,
288 	dml_uint_t *MaxPrefetchMode);
289 
290 static void CalculateRowBandwidth(
291 	dml_bool_t GPUVMEnable,
292 	enum dml_source_format_class SourcePixelFormat,
293 	dml_float_t VRatio,
294 	dml_float_t VRatioChroma,
295 	dml_bool_t DCCEnable,
296 	dml_float_t LineTime,
297 	dml_uint_t MetaRowByteLuma,
298 	dml_uint_t MetaRowByteChroma,
299 	dml_uint_t meta_row_height_luma,
300 	dml_uint_t meta_row_height_chroma,
301 	dml_uint_t PixelPTEBytesPerRowLuma,
302 	dml_uint_t PixelPTEBytesPerRowChroma,
303 	dml_uint_t dpte_row_height_luma,
304 	dml_uint_t dpte_row_height_chroma,
305 	// Output
306 	dml_float_t *meta_row_bw,
307 	dml_float_t *dpte_row_bw);
308 
309 static void CalculateFlipSchedule(
310 	dml_float_t HostVMInefficiencyFactor,
311 	dml_float_t UrgentExtraLatency,
312 	dml_float_t UrgentLatency,
313 	dml_uint_t GPUVMMaxPageTableLevels,
314 	dml_bool_t HostVMEnable,
315 	dml_uint_t HostVMMaxNonCachedPageTableLevels,
316 	dml_bool_t GPUVMEnable,
317 	dml_uint_t HostVMMinPageSize,
318 	dml_float_t PDEAndMetaPTEBytesPerFrame,
319 	dml_float_t MetaRowBytes,
320 	dml_float_t DPTEBytesPerRow,
321 	dml_float_t BandwidthAvailableForImmediateFlip,
322 	dml_uint_t TotImmediateFlipBytes,
323 	enum dml_source_format_class SourcePixelFormat,
324 	dml_float_t LineTime,
325 	dml_float_t VRatio,
326 	dml_float_t VRatioChroma,
327 	dml_float_t Tno_bw,
328 	dml_bool_t DCCEnable,
329 	dml_uint_t dpte_row_height,
330 	dml_uint_t meta_row_height,
331 	dml_uint_t dpte_row_height_chroma,
332 	dml_uint_t meta_row_height_chroma,
333 	dml_bool_t use_one_row_for_frame_flip,
334 
335 	// Output
336 	dml_float_t *DestinationLinesToRequestVMInImmediateFlip,
337 	dml_float_t *DestinationLinesToRequestRowInImmediateFlip,
338 	dml_float_t *final_flip_bw,
339 	dml_bool_t *ImmediateFlipSupportedForPipe);
340 
341 static dml_float_t CalculateWriteBackDelay(
342 	enum dml_source_format_class WritebackPixelFormat,
343 	dml_float_t WritebackHRatio,
344 	dml_float_t WritebackVRatio,
345 	dml_uint_t WritebackVTaps,
346 	dml_uint_t WritebackDestinationWidth,
347 	dml_uint_t WritebackDestinationHeight,
348 	dml_uint_t WritebackSourceHeight,
349 	dml_uint_t HTotal);
350 
351 static void CalculateVUpdateAndDynamicMetadataParameters(
352 	dml_uint_t MaxInterDCNTileRepeaters,
353 	dml_float_t Dppclk,
354 	dml_float_t DISPCLK,
355 	dml_float_t DCFClkDeepSleep,
356 	dml_float_t PixelClock,
357 	dml_uint_t HTotal,
358 	dml_uint_t VBlank,
359 	dml_uint_t DynamicMetadataTransmittedBytes,
360 	dml_uint_t DynamicMetadataLinesBeforeActiveRequired,
361 	dml_uint_t InterlaceEnable,
362 	dml_bool_t ProgressiveToInterlaceUnitInOPP,
363 	dml_float_t *TSetup,
364 	dml_float_t *Tdmbf,
365 	dml_float_t *Tdmec,
366 	dml_float_t *Tdmsks,
367 	dml_uint_t *VUpdateOffsetPix,
368 	dml_uint_t *VUpdateWidthPix,
369 	dml_uint_t *VReadyOffsetPix);
370 
371 static void PixelClockAdjustmentForProgressiveToInterlaceUnit(struct dml_display_cfg_st *display_cfg, dml_bool_t ptoi_supported);
372 
373 static dml_float_t TruncToValidBPP(
374 	dml_float_t LinkBitRate,
375 	dml_uint_t Lanes,
376 	dml_uint_t HTotal,
377 	dml_uint_t HActive,
378 	dml_float_t PixelClock,
379 	dml_float_t DesiredBPP,
380 	dml_bool_t DSCEnable,
381 	enum dml_output_encoder_class Output,
382 	enum dml_output_format_class Format,
383 	dml_uint_t DSCInputBitPerComponent,
384 	dml_uint_t DSCSlices,
385 	dml_uint_t AudioRate,
386 	dml_uint_t AudioLayout,
387 	enum dml_odm_mode ODMModeNoDSC,
388 	enum dml_odm_mode ODMModeDSC,
389 	// Output
390 	dml_uint_t *RequiredSlotsSingle);
391 
392 static void CalculateWatermarksMALLUseAndDRAMSpeedChangeSupport(
393 	struct display_mode_lib_scratch_st *s,
394 	struct CalculateWatermarksMALLUseAndDRAMSpeedChangeSupport_params_st *p);
395 
396 static void CalculateDCFCLKDeepSleep(
397 	dml_uint_t NumberOfActiveSurfaces,
398 	dml_uint_t BytePerPixelY[],
399 	dml_uint_t BytePerPixelC[],
400 	dml_float_t VRatio[],
401 	dml_float_t VRatioChroma[],
402 	dml_uint_t SwathWidthY[],
403 	dml_uint_t SwathWidthC[],
404 	dml_uint_t DPPPerSurface[],
405 	dml_float_t HRatio[],
406 	dml_float_t HRatioChroma[],
407 	dml_float_t PixelClock[],
408 	dml_float_t PSCL_THROUGHPUT[],
409 	dml_float_t PSCL_THROUGHPUT_CHROMA[],
410 	dml_float_t Dppclk[],
411 	dml_float_t ReadBandwidthLuma[],
412 	dml_float_t ReadBandwidthChroma[],
413 	dml_uint_t ReturnBusWidth,
414 
415 	// Output
416 	dml_float_t *DCFCLKDeepSleep);
417 
418 static void CalculateUrgentBurstFactor(
419 	enum dml_use_mall_for_pstate_change_mode UseMALLForPStateChange,
420 	dml_uint_t swath_width_luma_ub,
421 	dml_uint_t swath_width_chroma_ub,
422 	dml_uint_t SwathHeightY,
423 	dml_uint_t SwathHeightC,
424 	dml_float_t LineTime,
425 	dml_float_t UrgentLatency,
426 	dml_float_t CursorBufferSize,
427 	dml_uint_t CursorWidth,
428 	dml_uint_t CursorBPP,
429 	dml_float_t VRatio,
430 	dml_float_t VRatioC,
431 	dml_float_t BytePerPixelInDETY,
432 	dml_float_t BytePerPixelInDETC,
433 	dml_uint_t DETBufferSizeY,
434 	dml_uint_t DETBufferSizeC,
435 	// Output
436 	dml_float_t *UrgentBurstFactorCursor,
437 	dml_float_t *UrgentBurstFactorLuma,
438 	dml_float_t *UrgentBurstFactorChroma,
439 	dml_bool_t *NotEnoughUrgentLatencyHiding);
440 
441 static dml_float_t RequiredDTBCLK(
442 	dml_bool_t DSCEnable,
443 	dml_float_t PixelClock,
444 	enum dml_output_format_class OutputFormat,
445 	dml_float_t OutputBpp,
446 	dml_uint_t DSCSlices,
447 	dml_uint_t HTotal,
448 	dml_uint_t HActive,
449 	dml_uint_t AudioRate,
450 	dml_uint_t AudioLayoutSingle);
451 
452 static void UseMinimumDCFCLK(
453 	struct display_mode_lib_scratch_st *scratch,
454 	struct UseMinimumDCFCLK_params_st *p);
455 
456 static void CalculatePixelDeliveryTimes(
457 	dml_uint_t NumberOfActiveSurfaces,
458 	dml_float_t VRatio[],
459 	dml_float_t VRatioChroma[],
460 	dml_float_t VRatioPrefetchY[],
461 	dml_float_t VRatioPrefetchC[],
462 	dml_uint_t swath_width_luma_ub[],
463 	dml_uint_t swath_width_chroma_ub[],
464 	dml_uint_t DPPPerSurface[],
465 	dml_float_t HRatio[],
466 	dml_float_t HRatioChroma[],
467 	dml_float_t PixelClock[],
468 	dml_float_t PSCL_THROUGHPUT[],
469 	dml_float_t PSCL_THROUGHPUT_CHROMA[],
470 	dml_float_t Dppclk[],
471 	dml_uint_t BytePerPixelC[],
472 	enum dml_rotation_angle SourceScan[],
473 	dml_uint_t NumberOfCursors[],
474 	dml_uint_t CursorWidth[],
475 	dml_uint_t CursorBPP[],
476 	dml_uint_t BlockWidth256BytesY[],
477 	dml_uint_t BlockHeight256BytesY[],
478 	dml_uint_t BlockWidth256BytesC[],
479 	dml_uint_t BlockHeight256BytesC[],
480 
481 	// Output
482 	dml_float_t DisplayPipeLineDeliveryTimeLuma[],
483 	dml_float_t DisplayPipeLineDeliveryTimeChroma[],
484 	dml_float_t DisplayPipeLineDeliveryTimeLumaPrefetch[],
485 	dml_float_t DisplayPipeLineDeliveryTimeChromaPrefetch[],
486 	dml_float_t DisplayPipeRequestDeliveryTimeLuma[],
487 	dml_float_t DisplayPipeRequestDeliveryTimeChroma[],
488 	dml_float_t DisplayPipeRequestDeliveryTimeLumaPrefetch[],
489 	dml_float_t DisplayPipeRequestDeliveryTimeChromaPrefetch[],
490 	dml_float_t CursorRequestDeliveryTime[],
491 	dml_float_t CursorRequestDeliveryTimePrefetch[]);
492 
493 static void CalculateMetaAndPTETimes(
494 	dml_bool_t use_one_row_for_frame[],
495 	dml_uint_t NumberOfActiveSurfaces,
496 	dml_bool_t GPUVMEnable,
497 	dml_uint_t MetaChunkSize,
498 	dml_uint_t MinMetaChunkSizeBytes,
499 	dml_uint_t HTotal[],
500 	dml_float_t VRatio[],
501 	dml_float_t VRatioChroma[],
502 	dml_float_t DestinationLinesToRequestRowInVBlank[],
503 	dml_float_t DestinationLinesToRequestRowInImmediateFlip[],
504 	dml_bool_t DCCEnable[],
505 	dml_float_t PixelClock[],
506 	dml_uint_t BytePerPixelY[],
507 	dml_uint_t BytePerPixelC[],
508 	enum dml_rotation_angle SourceScan[],
509 	dml_uint_t dpte_row_height[],
510 	dml_uint_t dpte_row_height_chroma[],
511 	dml_uint_t meta_row_width[],
512 	dml_uint_t meta_row_width_chroma[],
513 	dml_uint_t meta_row_height[],
514 	dml_uint_t meta_row_height_chroma[],
515 	dml_uint_t meta_req_width[],
516 	dml_uint_t meta_req_width_chroma[],
517 	dml_uint_t meta_req_height[],
518 	dml_uint_t meta_req_height_chroma[],
519 	dml_uint_t dpte_group_bytes[],
520 	dml_uint_t PTERequestSizeY[],
521 	dml_uint_t PTERequestSizeC[],
522 	dml_uint_t PixelPTEReqWidthY[],
523 	dml_uint_t PixelPTEReqHeightY[],
524 	dml_uint_t PixelPTEReqWidthC[],
525 	dml_uint_t PixelPTEReqHeightC[],
526 	dml_uint_t dpte_row_width_luma_ub[],
527 	dml_uint_t dpte_row_width_chroma_ub[],
528 
529 	// Output
530 	dml_float_t DST_Y_PER_PTE_ROW_NOM_L[],
531 	dml_float_t DST_Y_PER_PTE_ROW_NOM_C[],
532 	dml_float_t DST_Y_PER_META_ROW_NOM_L[],
533 	dml_float_t DST_Y_PER_META_ROW_NOM_C[],
534 	dml_float_t TimePerMetaChunkNominal[],
535 	dml_float_t TimePerChromaMetaChunkNominal[],
536 	dml_float_t TimePerMetaChunkVBlank[],
537 	dml_float_t TimePerChromaMetaChunkVBlank[],
538 	dml_float_t TimePerMetaChunkFlip[],
539 	dml_float_t TimePerChromaMetaChunkFlip[],
540 	dml_float_t time_per_pte_group_nom_luma[],
541 	dml_float_t time_per_pte_group_vblank_luma[],
542 	dml_float_t time_per_pte_group_flip_luma[],
543 	dml_float_t time_per_pte_group_nom_chroma[],
544 	dml_float_t time_per_pte_group_vblank_chroma[],
545 	dml_float_t time_per_pte_group_flip_chroma[]);
546 
547 static void CalculateVMGroupAndRequestTimes(
548 	dml_uint_t NumberOfActiveSurfaces,
549 	dml_bool_t GPUVMEnable,
550 	dml_uint_t GPUVMMaxPageTableLevels,
551 	dml_uint_t HTotal[],
552 	dml_uint_t BytePerPixelC[],
553 	dml_float_t DestinationLinesToRequestVMInVBlank[],
554 	dml_float_t DestinationLinesToRequestVMInImmediateFlip[],
555 	dml_bool_t DCCEnable[],
556 	dml_float_t PixelClock[],
557 	dml_uint_t dpte_row_width_luma_ub[],
558 	dml_uint_t dpte_row_width_chroma_ub[],
559 	dml_uint_t vm_group_bytes[],
560 	dml_uint_t dpde0_bytes_per_frame_ub_l[],
561 	dml_uint_t dpde0_bytes_per_frame_ub_c[],
562 	dml_uint_t meta_pte_bytes_per_frame_ub_l[],
563 	dml_uint_t meta_pte_bytes_per_frame_ub_c[],
564 
565 	// Output
566 	dml_float_t TimePerVMGroupVBlank[],
567 	dml_float_t TimePerVMGroupFlip[],
568 	dml_float_t TimePerVMRequestVBlank[],
569 	dml_float_t TimePerVMRequestFlip[]);
570 
571 static void CalculateStutterEfficiency(
572 	struct display_mode_lib_scratch_st *scratch,
573 	struct CalculateStutterEfficiency_params_st *p);
574 
575 static void CalculateSwathAndDETConfiguration(
576 	struct display_mode_lib_scratch_st *scratch,
577 	struct CalculateSwathAndDETConfiguration_params_st *p);
578 
579 static void CalculateSwathWidth(
580 	dml_bool_t ForceSingleDPP,
581 	dml_uint_t NumberOfActiveSurfaces,
582 	enum dml_source_format_class SourcePixelFormat[],
583 	enum dml_rotation_angle SourceScan[],
584 	dml_bool_t ViewportStationary[],
585 	dml_uint_t ViewportWidth[],
586 	dml_uint_t ViewportHeight[],
587 	dml_uint_t ViewportXStart[],
588 	dml_uint_t ViewportYStart[],
589 	dml_uint_t ViewportXStartC[],
590 	dml_uint_t ViewportYStartC[],
591 	dml_uint_t SurfaceWidthY[],
592 	dml_uint_t SurfaceWidthC[],
593 	dml_uint_t SurfaceHeightY[],
594 	dml_uint_t SurfaceHeightC[],
595 	enum dml_odm_mode ODMMode[],
596 	dml_uint_t BytePerPixY[],
597 	dml_uint_t BytePerPixC[],
598 	dml_uint_t Read256BytesBlockHeightY[],
599 	dml_uint_t Read256BytesBlockHeightC[],
600 	dml_uint_t Read256BytesBlockWidthY[],
601 	dml_uint_t Read256BytesBlockWidthC[],
602 	dml_uint_t BlendingAndTiming[],
603 	dml_uint_t HActive[],
604 	dml_float_t HRatio[],
605 	dml_uint_t DPPPerSurface[],
606 
607 	// Output
608 	dml_uint_t SwathWidthSingleDPPY[],
609 	dml_uint_t SwathWidthSingleDPPC[],
610 	dml_uint_t SwathWidthY[],
611 	dml_uint_t SwathWidthC[],
612 	dml_uint_t MaximumSwathHeightY[],
613 	dml_uint_t MaximumSwathHeightC[],
614 	dml_uint_t swath_width_luma_ub[],
615 	dml_uint_t swath_width_chroma_ub[]);
616 
617 static dml_float_t CalculateExtraLatency(
618 	dml_uint_t RoundTripPingLatencyCycles,
619 	dml_uint_t ReorderingBytes,
620 	dml_float_t DCFCLK,
621 	dml_uint_t TotalNumberOfActiveDPP,
622 	dml_uint_t PixelChunkSizeInKByte,
623 	dml_uint_t TotalNumberOfDCCActiveDPP,
624 	dml_uint_t MetaChunkSize,
625 	dml_float_t ReturnBW,
626 	dml_bool_t GPUVMEnable,
627 	dml_bool_t HostVMEnable,
628 	dml_uint_t NumberOfActiveSurfaces,
629 	dml_uint_t NumberOfDPP[],
630 	dml_uint_t dpte_group_bytes[],
631 	dml_float_t HostVMInefficiencyFactor,
632 	dml_uint_t HostVMMinPageSize,
633 	dml_uint_t HostVMMaxNonCachedPageTableLevels);
634 
635 static dml_uint_t CalculateExtraLatencyBytes(
636 	dml_uint_t ReorderingBytes,
637 	dml_uint_t TotalNumberOfActiveDPP,
638 	dml_uint_t PixelChunkSizeInKByte,
639 	dml_uint_t TotalNumberOfDCCActiveDPP,
640 	dml_uint_t MetaChunkSize,
641 	dml_bool_t GPUVMEnable,
642 	dml_bool_t HostVMEnable,
643 	dml_uint_t NumberOfActiveSurfaces,
644 	dml_uint_t NumberOfDPP[],
645 	dml_uint_t dpte_group_bytes[],
646 	dml_float_t HostVMInefficiencyFactor,
647 	dml_uint_t HostVMMinPageSize,
648 	dml_uint_t HostVMMaxNonCachedPageTableLevels);
649 
650 static dml_float_t CalculateUrgentLatency(
651 	dml_float_t UrgentLatencyPixelDataOnly,
652 	dml_float_t UrgentLatencyPixelMixedWithVMData,
653 	dml_float_t UrgentLatencyVMDataOnly,
654 	dml_bool_t DoUrgentLatencyAdjustment,
655 	dml_float_t UrgentLatencyAdjustmentFabricClockComponent,
656 	dml_float_t UrgentLatencyAdjustmentFabricClockReference,
657 	dml_float_t FabricClockSingle);
658 
659 static dml_bool_t UnboundedRequest(
660 	enum dml_unbounded_requesting_policy UseUnboundedRequestingFinal,
661 	dml_uint_t TotalNumberOfActiveDPP,
662 	dml_bool_t NoChromaOrLinear,
663 	enum dml_output_encoder_class Output);
664 
665 static void CalculateSurfaceSizeInMall(
666 	dml_uint_t NumberOfActiveSurfaces,
667 	dml_uint_t MALLAllocatedForDCN,
668 	enum dml_use_mall_for_static_screen_mode UseMALLForStaticScreen[],
669 	dml_bool_t DCCEnable[],
670 	dml_bool_t ViewportStationary[],
671 	dml_uint_t ViewportXStartY[],
672 	dml_uint_t ViewportYStartY[],
673 	dml_uint_t ViewportXStartC[],
674 	dml_uint_t ViewportYStartC[],
675 	dml_uint_t ViewportWidthY[],
676 	dml_uint_t ViewportHeightY[],
677 	dml_uint_t BytesPerPixelY[],
678 	dml_uint_t ViewportWidthC[],
679 	dml_uint_t ViewportHeightC[],
680 	dml_uint_t BytesPerPixelC[],
681 	dml_uint_t SurfaceWidthY[],
682 	dml_uint_t SurfaceWidthC[],
683 	dml_uint_t SurfaceHeightY[],
684 	dml_uint_t SurfaceHeightC[],
685 	dml_uint_t Read256BytesBlockWidthY[],
686 	dml_uint_t Read256BytesBlockWidthC[],
687 	dml_uint_t Read256BytesBlockHeightY[],
688 	dml_uint_t Read256BytesBlockHeightC[],
689 	dml_uint_t ReadBlockWidthY[],
690 	dml_uint_t ReadBlockWidthC[],
691 	dml_uint_t ReadBlockHeightY[],
692 	dml_uint_t ReadBlockHeightC[],
693 
694 	// Output
695 	dml_uint_t SurfaceSizeInMALL[],
696 	dml_bool_t *ExceededMALLSize);
697 
698 static void CalculateDETBufferSize(
699 	dml_uint_t DETSizeOverride[],
700 	enum dml_use_mall_for_pstate_change_mode UseMALLForPStateChange[],
701 	dml_bool_t ForceSingleDPP,
702 	dml_uint_t NumberOfActiveSurfaces,
703 	dml_bool_t UnboundedRequestEnabled,
704 	dml_uint_t nomDETInKByte,
705 	dml_uint_t MaxTotalDETInKByte,
706 	dml_uint_t ConfigReturnBufferSizeInKByte,
707 	dml_uint_t MinCompressedBufferSizeInKByte,
708 	dml_uint_t ConfigReturnBufferSegmentSizeInkByte,
709 	dml_uint_t CompressedBufferSegmentSizeInkByteFinal,
710 	enum dml_source_format_class SourcePixelFormat[],
711 	dml_float_t ReadBandwidthLuma[],
712 	dml_float_t ReadBandwidthChroma[],
713 	dml_uint_t RotesY[],
714 	dml_uint_t RoundedUpMaxSwathSizeBytesC[],
715 	dml_uint_t DPPPerSurface[],
716 	// Output
717 	dml_uint_t DETBufferSizeInKByte[],
718 	dml_uint_t *CompressedBufferSizeInkByte);
719 
720 static void CalculateMaxDETAndMinCompressedBufferSize(
721 	dml_uint_t ConfigReturnBufferSizeInKByte,
722 	dml_uint_t ConfigReturnBufferSegmentSizeInKByte,
723 	dml_uint_t ROBBufferSizeInKByte,
724 	dml_uint_t MaxNumDPP,
725 	dml_bool_t nomDETInKByteOverrideEnable,
726 	dml_uint_t nomDETInKByteOverrideValue,
727 
728 	// Output
729 	dml_uint_t *MaxTotalDETInKByte,
730 	dml_uint_t *nomDETInKByte,
731 	dml_uint_t *MinCompressedBufferSizeInKByte);
732 
733 static dml_uint_t DSCDelayRequirement(
734 	dml_bool_t DSCEnabled,
735 	enum dml_odm_mode ODMMode,
736 	dml_uint_t DSCInputBitPerComponent,
737 	dml_float_t OutputBpp,
738 	dml_uint_t HActive,
739 	dml_uint_t HTotal,
740 	dml_uint_t NumberOfDSCSlices,
741 	enum dml_output_format_class OutputFormat,
742 	enum dml_output_encoder_class Output,
743 	dml_float_t PixelClock,
744 	dml_float_t PixelClockBackEnd);
745 
746 static dml_bool_t CalculateVActiveBandwithSupport(
747 	dml_uint_t NumberOfActiveSurfaces,
748 	dml_float_t ReturnBW,
749 	dml_bool_t NotUrgentLatencyHiding[],
750 	dml_float_t ReadBandwidthLuma[],
751 	dml_float_t ReadBandwidthChroma[],
752 	dml_float_t cursor_bw[],
753 	dml_float_t meta_row_bandwidth[],
754 	dml_float_t dpte_row_bandwidth[],
755 	dml_uint_t NumberOfDPP[],
756 	dml_float_t UrgentBurstFactorLuma[],
757 	dml_float_t UrgentBurstFactorChroma[],
758 	dml_float_t UrgentBurstFactorCursor[]);
759 
760 static void CalculatePrefetchBandwithSupport(
761 	dml_uint_t NumberOfActiveSurfaces,
762 	dml_float_t ReturnBW,
763 	enum dml_use_mall_for_pstate_change_mode UseMALLForPStateChange[],
764 	dml_bool_t NotUrgentLatencyHiding[],
765 	dml_float_t ReadBandwidthLuma[],
766 	dml_float_t ReadBandwidthChroma[],
767 	dml_float_t PrefetchBandwidthLuma[],
768 	dml_float_t PrefetchBandwidthChroma[],
769 	dml_float_t cursor_bw[],
770 	dml_float_t meta_row_bandwidth[],
771 	dml_float_t dpte_row_bandwidth[],
772 	dml_float_t cursor_bw_pre[],
773 	dml_float_t prefetch_vmrow_bw[],
774 	dml_uint_t NumberOfDPP[],
775 	dml_float_t UrgentBurstFactorLuma[],
776 	dml_float_t UrgentBurstFactorChroma[],
777 	dml_float_t UrgentBurstFactorCursor[],
778 	dml_float_t UrgentBurstFactorLumaPre[],
779 	dml_float_t UrgentBurstFactorChromaPre[],
780 	dml_float_t UrgentBurstFactorCursorPre[],
781 
782 	// Output
783 	dml_float_t *PrefetchBandwidth,
784 	dml_float_t *PrefetchBandwidthNotIncludingMALLPrefetch,
785 	dml_float_t *FractionOfUrgentBandwidth,
786 	dml_bool_t *PrefetchBandwidthSupport);
787 
788 static dml_float_t CalculateBandwidthAvailableForImmediateFlip(
789 	dml_uint_t NumberOfActiveSurfaces,
790 	dml_float_t ReturnBW,
791 	dml_float_t ReadBandwidthLuma[],
792 	dml_float_t ReadBandwidthChroma[],
793 	dml_float_t PrefetchBandwidthLuma[],
794 	dml_float_t PrefetchBandwidthChroma[],
795 	dml_float_t cursor_bw[],
796 	dml_float_t cursor_bw_pre[],
797 	dml_uint_t NumberOfDPP[],
798 	dml_float_t UrgentBurstFactorLuma[],
799 	dml_float_t UrgentBurstFactorChroma[],
800 	dml_float_t UrgentBurstFactorCursor[],
801 	dml_float_t UrgentBurstFactorLumaPre[],
802 	dml_float_t UrgentBurstFactorChromaPre[],
803 	dml_float_t UrgentBurstFactorCursorPre[]);
804 
805 static void CalculateImmediateFlipBandwithSupport(
806 	dml_uint_t NumberOfActiveSurfaces,
807 	dml_float_t ReturnBW,
808 	enum dml_use_mall_for_pstate_change_mode UseMALLForPStateChange[],
809 	enum dml_immediate_flip_requirement ImmediateFlipRequirement[],
810 	dml_float_t final_flip_bw[],
811 	dml_float_t ReadBandwidthLuma[],
812 	dml_float_t ReadBandwidthChroma[],
813 	dml_float_t PrefetchBandwidthLuma[],
814 	dml_float_t PrefetchBandwidthChroma[],
815 	dml_float_t cursor_bw[],
816 	dml_float_t meta_row_bandwidth[],
817 	dml_float_t dpte_row_bandwidth[],
818 	dml_float_t cursor_bw_pre[],
819 	dml_float_t prefetch_vmrow_bw[],
820 	dml_uint_t NumberOfDPP[],
821 	dml_float_t UrgentBurstFactorLuma[],
822 	dml_float_t UrgentBurstFactorChroma[],
823 	dml_float_t UrgentBurstFactorCursor[],
824 	dml_float_t UrgentBurstFactorLumaPre[],
825 	dml_float_t UrgentBurstFactorChromaPre[],
826 	dml_float_t UrgentBurstFactorCursorPre[],
827 
828 	// Output
829 	dml_float_t *TotalBandwidth,
830 	dml_float_t *TotalBandwidthNotIncludingMALLPrefetch,
831 	dml_float_t *FractionOfUrgentBandwidth,
832 	dml_bool_t *ImmediateFlipBandwidthSupport);
833 
834 // ---------------------------
835 //  Declaration Ends
836 // ---------------------------
837 
dscceComputeDelay(dml_uint_t bpc,dml_float_t BPP,dml_uint_t sliceWidth,dml_uint_t numSlices,enum dml_output_format_class pixelFormat,enum dml_output_encoder_class Output)838 static dml_uint_t dscceComputeDelay(
839 	dml_uint_t bpc,
840 	dml_float_t BPP,
841 	dml_uint_t sliceWidth,
842 	dml_uint_t numSlices,
843 	enum dml_output_format_class pixelFormat,
844 	enum dml_output_encoder_class Output)
845 {
846 	// valid bpc = source bits per component in the set of {8, 10, 12}
847 	// valid bpp = increments of 1/16 of a bit
848 	// min = 6/7/8 in N420/N422/444, respectively
849 	// max = such that compression is 1:1
850 	//valid sliceWidth = number of pixels per slice line, must be less than or equal to 5184/numSlices (or 4096/numSlices in 420 mode)
851 	//valid numSlices = number of slices in the horiziontal direction per DSC engine in the set of {1, 2, 3, 4}
852 	//valid pixelFormat = pixel/color format in the set of {:N444_RGB, :S422, :N422, :N420}
853 
854 	// fixed value
855 	dml_uint_t rcModelSize = 8192;
856 
857 	// N422/N420 operate at 2 pixels per clock
858 	dml_uint_t pixelsPerClock, lstall, D, initalXmitDelay, w, s, ix, wx, p, l0, a, ax, L,
859 		Delay, pixels;
860 
861 	if (pixelFormat == dml_420)
862 		pixelsPerClock = 2;
863 	// #all other modes operate at 1 pixel per clock
864 	else if (pixelFormat == dml_444)
865 		pixelsPerClock = 1;
866 	else if (pixelFormat == dml_n422)
867 		pixelsPerClock = 2;
868 	else
869 		pixelsPerClock = 1;
870 
871 	//initial transmit delay as per PPS
872 	initalXmitDelay = (dml_uint_t)(dml_round(rcModelSize / 2.0 / BPP / pixelsPerClock, 1));
873 
874 	//compute ssm delay
875 	if (bpc == 8)
876 		D = 81;
877 	else if (bpc == 10)
878 		D = 89;
879 	else
880 		D = 113;
881 
882 	//divide by pixel per cycle to compute slice width as seen by DSC
883 	w = sliceWidth / pixelsPerClock;
884 
885 	//422 mode has an additional cycle of delay
886 	if (pixelFormat == dml_420 || pixelFormat == dml_444 || pixelFormat == dml_n422)
887 		s = 0;
888 	else
889 		s = 1;
890 
891 	//main calculation for the dscce
892 	ix = initalXmitDelay + 45;
893 	wx = (w + 2) / 3;
894 	p = 3 * wx - w;
895 	l0 = ix / w;
896 	a = ix + p * l0;
897 	ax = (a + 2) / 3 + D + 6 + 1;
898 	L = (ax + wx - 1) / wx;
899 	if ((ix % w) == 0 && p != 0)
900 		lstall = 1;
901 	else
902 		lstall = 0;
903 	Delay = L * wx * (numSlices - 1) + ax + s + lstall + 22;
904 
905 	//dsc processes 3 pixel containers per cycle and a container can contain 1 or 2 pixels
906 	pixels = Delay * 3 * pixelsPerClock;
907 
908 #ifdef __DML_VBA_DEBUG__
909 	dml_print("DML::%s: bpc: %u\n", __func__, bpc);
910 	dml_print("DML::%s: BPP: %f\n", __func__, BPP);
911 	dml_print("DML::%s: sliceWidth: %u\n", __func__, sliceWidth);
912 	dml_print("DML::%s: numSlices: %u\n", __func__, numSlices);
913 	dml_print("DML::%s: pixelFormat: %u\n", __func__, pixelFormat);
914 	dml_print("DML::%s: Output: %u\n", __func__, Output);
915 	dml_print("DML::%s: pixels: %u\n", __func__, pixels);
916 #endif
917 	return pixels;
918 }
919 
dscComputeDelay(enum dml_output_format_class pixelFormat,enum dml_output_encoder_class Output)920 static dml_uint_t dscComputeDelay(enum dml_output_format_class pixelFormat, enum dml_output_encoder_class Output)
921 {
922 	dml_uint_t Delay = 0;
923 
924 	if (pixelFormat == dml_420) {
925 		// sfr
926 		Delay = Delay + 2;
927 		// dsccif
928 		Delay = Delay + 0;
929 		// dscc - input deserializer
930 		Delay = Delay + 3;
931 		// dscc gets pixels every other cycle
932 		Delay = Delay + 2;
933 		// dscc - input cdc fifo
934 		Delay = Delay + 12;
935 		// dscc gets pixels every other cycle
936 		Delay = Delay + 13;
937 		// dscc - cdc uncertainty
938 		Delay = Delay + 2;
939 		// dscc - output cdc fifo
940 		Delay = Delay + 7;
941 		// dscc gets pixels every other cycle
942 		Delay = Delay + 3;
943 		// dscc - cdc uncertainty
944 		Delay = Delay + 2;
945 		// dscc - output serializer
946 		Delay = Delay + 1;
947 		// sft
948 		Delay = Delay + 1;
949 	} else if (pixelFormat == dml_n422) {
950 	// sfr
951 	Delay = Delay + 2;
952 	// dsccif
953 	Delay = Delay + 1;
954 	// dscc - input deserializer
955 	Delay = Delay + 5;
956 	// dscc - input cdc fifo
957 	Delay = Delay + 25;
958 	// dscc - cdc uncertainty
959 	Delay = Delay + 2;
960 	// dscc - output cdc fifo
961 	Delay = Delay + 10;
962 	// dscc - cdc uncertainty
963 	Delay = Delay + 2;
964 	// dscc - output serializer
965 	Delay = Delay + 1;
966 	// sft
967 	Delay = Delay + 1;
968 	} else {
969 	// sfr
970 	Delay = Delay + 2;
971 	// dsccif
972 	Delay = Delay + 0;
973 	// dscc - input deserializer
974 	Delay = Delay + 3;
975 	// dscc - input cdc fifo
976 	Delay = Delay + 12;
977 	// dscc - cdc uncertainty
978 	Delay = Delay + 2;
979 	// dscc - output cdc fifo
980 	Delay = Delay + 7;
981 	// dscc - output serializer
982 	Delay = Delay + 1;
983 	// dscc - cdc uncertainty
984 	Delay = Delay + 2;
985 	// sft
986 	Delay = Delay + 1;
987 	}
988 #ifdef __DML_VBA_DEBUG__
989 	dml_print("DML::%s: pixelFormat = %u\n", __func__, pixelFormat);
990 	dml_print("DML::%s: Delay = %u\n", __func__, Delay);
991 #endif
992 
993 	return Delay;
994 }
995 
CalculatePrefetchSchedule(struct display_mode_lib_scratch_st * scratch,struct CalculatePrefetchSchedule_params_st * p)996 static dml_bool_t CalculatePrefetchSchedule(struct display_mode_lib_scratch_st *scratch,
997 	struct CalculatePrefetchSchedule_params_st *p)
998 {
999 	struct CalculatePrefetchSchedule_locals_st *s = &scratch->CalculatePrefetchSchedule_locals;
1000 
1001 	s->MyError = false;
1002 	s->DPPCycles = 0;
1003 	s->DISPCLKCycles = 0;
1004 	s->DSTTotalPixelsAfterScaler = 0.0;
1005 	s->LineTime = 0.0;
1006 	s->dst_y_prefetch_equ = 0.0;
1007 	s->prefetch_bw_oto = 0.0;
1008 	s->Tvm_oto = 0.0;
1009 	s->Tr0_oto = 0.0;
1010 	s->Tvm_oto_lines = 0.0;
1011 	s->Tr0_oto_lines = 0.0;
1012 	s->dst_y_prefetch_oto = 0.0;
1013 	s->TimeForFetchingMetaPTE = 0.0;
1014 	s->TimeForFetchingRowInVBlank = 0.0;
1015 	s->LinesToRequestPrefetchPixelData = 0.0;
1016 	s->HostVMDynamicLevelsTrips = 0;
1017 	s->trip_to_mem = 0.0;
1018 	s->Tvm_trips = 0.0;
1019 	s->Tr0_trips = 0.0;
1020 	s->Tvm_trips_rounded = 0.0;
1021 	s->Tr0_trips_rounded = 0.0;
1022 	s->max_Tsw = 0.0;
1023 	s->Lsw_oto = 0.0;
1024 	s->Tpre_rounded = 0.0;
1025 	s->prefetch_bw_equ = 0.0;
1026 	s->Tvm_equ = 0.0;
1027 	s->Tr0_equ = 0.0;
1028 	s->Tdmbf = 0.0;
1029 	s->Tdmec = 0.0;
1030 	s->Tdmsks = 0.0;
1031 	s->prefetch_sw_bytes = 0.0;
1032 	s->prefetch_bw_pr = 0.0;
1033 	s->bytes_pp = 0.0;
1034 	s->dep_bytes = 0.0;
1035 	s->min_Lsw_oto = 0.0;
1036 	s->Tsw_est1 = 0.0;
1037 	s->Tsw_est3 = 0.0;
1038 
1039 	if (p->GPUVMEnable == true && p->HostVMEnable == true) {
1040 		s->HostVMDynamicLevelsTrips = p->HostVMMaxNonCachedPageTableLevels;
1041 	} else {
1042 		s->HostVMDynamicLevelsTrips = 0;
1043 	}
1044 #ifdef __DML_VBA_DEBUG__
1045 	dml_print("DML::%s: GPUVMEnable = %u\n", __func__, p->GPUVMEnable);
1046 	dml_print("DML::%s: GPUVMPageTableLevels = %u\n", __func__, p->GPUVMPageTableLevels);
1047 	dml_print("DML::%s: DCCEnable = %u\n", __func__, p->myPipe->DCCEnable);
1048 	dml_print("DML::%s: VStartup = %u\n", __func__, p->VStartup);
1049 	dml_print("DML::%s: MaxVStartup = %u\n", __func__, p->MaxVStartup);
1050 	dml_print("DML::%s: HostVMEnable = %u\n", __func__, p->HostVMEnable);
1051 	dml_print("DML::%s: HostVMInefficiencyFactor= %f\n", __func__, p->HostVMInefficiencyFactor);
1052 	dml_print("DML::%s: myPipe->Dppclk = %f\n", __func__, p->myPipe->Dppclk);
1053 #endif
1054 	CalculateVUpdateAndDynamicMetadataParameters(
1055 	p->MaxInterDCNTileRepeaters,
1056 	p->myPipe->Dppclk,
1057 	p->myPipe->Dispclk,
1058 	p->myPipe->DCFClkDeepSleep,
1059 	p->myPipe->PixelClock,
1060 	p->myPipe->HTotal,
1061 	p->myPipe->VBlank,
1062 	p->DynamicMetadataTransmittedBytes,
1063 	p->DynamicMetadataLinesBeforeActiveRequired,
1064 	p->myPipe->InterlaceEnable,
1065 	p->myPipe->ProgressiveToInterlaceUnitInOPP,
1066 	p->TSetup,
1067 
1068 	// Output
1069 	&s->Tdmbf,
1070 	&s->Tdmec,
1071 	&s->Tdmsks,
1072 	p->VUpdateOffsetPix,
1073 	p->VUpdateWidthPix,
1074 	p->VReadyOffsetPix);
1075 
1076 	s->LineTime = p->myPipe->HTotal / p->myPipe->PixelClock;
1077 	s->trip_to_mem = p->UrgentLatency;
1078 	s->Tvm_trips = p->UrgentExtraLatency + s->trip_to_mem * (p->GPUVMPageTableLevels * (s->HostVMDynamicLevelsTrips + 1) - 1);
1079 
1080 	if (p->DynamicMetadataVMEnabled == true) {
1081 		*p->Tdmdl = p->TWait + s->Tvm_trips + s->trip_to_mem;
1082 	} else {
1083 		*p->Tdmdl = p->TWait + p->UrgentExtraLatency;
1084 	}
1085 
1086 #ifdef __DML_VBA_ALLOW_DELTA__
1087 	if (DynamicMetadataEnable == false) {
1088 		*Tdmdl = 0.0;
1089 	}
1090 #endif
1091 
1092 	if (p->DynamicMetadataEnable == true) {
1093 		if (p->VStartup * s->LineTime < *p->TSetup + *p->Tdmdl + s->Tdmbf + s->Tdmec + s->Tdmsks) {
1094 			*p->NotEnoughTimeForDynamicMetadata = true;
1095 			dml_print("DML::%s: Not Enough Time for Dynamic Meta!\n", __func__);
1096 			dml_print("DML::%s: Tdmbf: %fus - time for dmd transfer from dchub to dio output buffer\n", __func__, s->Tdmbf);
1097 			dml_print("DML::%s: Tdmec: %fus - time dio takes to transfer dmd\n", __func__, s->Tdmec);
1098 			dml_print("DML::%s: Tdmsks: %fus - time before active dmd must complete transmission at dio\n", __func__, s->Tdmsks);
1099 			dml_print("DML::%s: Tdmdl: %fus - time for fabric to become ready and fetch dmd \n", __func__, *p->Tdmdl);
1100 		} else {
1101 			*p->NotEnoughTimeForDynamicMetadata = false;
1102 		}
1103 	} else {
1104 		*p->NotEnoughTimeForDynamicMetadata = false;
1105 	}
1106 
1107 	*p->Tdmdl_vm = (p->DynamicMetadataEnable == true && p->DynamicMetadataVMEnabled == true && p->GPUVMEnable == true ? p->TWait + s->Tvm_trips : 0);
1108 
1109 	if (p->myPipe->ScalerEnabled)
1110 		s->DPPCycles = (dml_uint_t)(p->DPPCLKDelaySubtotalPlusCNVCFormater + p->DPPCLKDelaySCL);
1111 	else
1112 		s->DPPCycles = (dml_uint_t)(p->DPPCLKDelaySubtotalPlusCNVCFormater + p->DPPCLKDelaySCLLBOnly);
1113 
1114 	s->DPPCycles = (dml_uint_t)(s->DPPCycles + p->myPipe->NumberOfCursors * p->DPPCLKDelayCNVCCursor);
1115 
1116 	s->DISPCLKCycles = (dml_uint_t)p->DISPCLKDelaySubtotal;
1117 
1118 	if (p->myPipe->Dppclk == 0.0 || p->myPipe->Dispclk == 0.0)
1119 		return true;
1120 
1121 	*p->DSTXAfterScaler = (dml_uint_t) dml_round(s->DPPCycles * p->myPipe->PixelClock / p->myPipe->Dppclk + s->DISPCLKCycles * p->myPipe->PixelClock / p->myPipe->Dispclk + p->DSCDelay, 1.0);
1122 	*p->DSTXAfterScaler = (dml_uint_t) dml_round(*p->DSTXAfterScaler + (p->myPipe->ODMMode != dml_odm_mode_bypass ? 18 : 0) + (p->myPipe->DPPPerSurface - 1) * p->DPP_RECOUT_WIDTH +
1123 						((p->myPipe->ODMMode == dml_odm_mode_split_1to2 || p->myPipe->ODMMode == dml_odm_mode_mso_1to2) ? (dml_float_t)p->myPipe->HActive / 2.0 : 0) +
1124 						((p->myPipe->ODMMode == dml_odm_mode_mso_1to4) ? (dml_float_t)p->myPipe->HActive * 3.0 / 4.0 : 0), 1.0);
1125 
1126 #ifdef __DML_VBA_DEBUG__
1127 	dml_print("DML::%s: DPPCycles = %u\n", __func__, s->DPPCycles);
1128 	dml_print("DML::%s: PixelClock = %f\n", __func__, p->myPipe->PixelClock);
1129 	dml_print("DML::%s: Dppclk = %f\n", __func__, p->myPipe->Dppclk);
1130 	dml_print("DML::%s: DISPCLKCycles = %u\n", __func__, s->DISPCLKCycles);
1131 	dml_print("DML::%s: DISPCLK = %f\n", __func__, p->myPipe->Dispclk);
1132 	dml_print("DML::%s: DSCDelay = %u\n", __func__, p->DSCDelay);
1133 	dml_print("DML::%s: ODMMode = %u\n", __func__, p->myPipe->ODMMode);
1134 	dml_print("DML::%s: DPP_RECOUT_WIDTH = %u\n", __func__, p->DPP_RECOUT_WIDTH);
1135 	dml_print("DML::%s: DSTXAfterScaler = %u\n", __func__, *p->DSTXAfterScaler);
1136 #endif
1137 
1138 	if (p->OutputFormat == dml_420 || (p->myPipe->InterlaceEnable && p->myPipe->ProgressiveToInterlaceUnitInOPP))
1139 		*p->DSTYAfterScaler = 1;
1140 	else
1141 		*p->DSTYAfterScaler = 0;
1142 
1143 	s->DSTTotalPixelsAfterScaler = *p->DSTYAfterScaler * p->myPipe->HTotal + *p->DSTXAfterScaler;
1144 	*p->DSTYAfterScaler = (dml_uint_t)(dml_floor(s->DSTTotalPixelsAfterScaler / p->myPipe->HTotal, 1));
1145 	*p->DSTXAfterScaler = (dml_uint_t)(s->DSTTotalPixelsAfterScaler - ((dml_float_t) (*p->DSTYAfterScaler * p->myPipe->HTotal)));
1146 #ifdef __DML_VBA_DEBUG__
1147 	dml_print("DML::%s: DSTXAfterScaler = %u (final)\n", __func__,  *p->DSTXAfterScaler);
1148 	dml_print("DML::%s: DSTYAfterScaler = %u (final)\n", __func__, *p->DSTYAfterScaler);
1149 #endif
1150 
1151 	s->MyError = false;
1152 
1153 	s->Tr0_trips = s->trip_to_mem * (s->HostVMDynamicLevelsTrips + 1);
1154 
1155 	if (p->GPUVMEnable == true) {
1156 		s->Tvm_trips_rounded = dml_ceil(4.0 * s->Tvm_trips / s->LineTime, 1.0) / 4.0 * s->LineTime;
1157 		s->Tr0_trips_rounded = dml_ceil(4.0 * s->Tr0_trips / s->LineTime, 1.0) / 4.0 * s->LineTime;
1158 		if (p->GPUVMPageTableLevels >= 3) {
1159 			*p->Tno_bw = p->UrgentExtraLatency + s->trip_to_mem * (dml_float_t) ((p->GPUVMPageTableLevels - 2) * (s->HostVMDynamicLevelsTrips + 1) - 1);
1160 	} else if (p->GPUVMPageTableLevels == 1 && p->myPipe->DCCEnable != true) {
1161 			s->Tr0_trips_rounded = dml_ceil(4.0 * p->UrgentExtraLatency / s->LineTime, 1.0) / 4.0 * s->LineTime;
1162 			*p->Tno_bw = p->UrgentExtraLatency;
1163 		} else {
1164 			*p->Tno_bw = 0;
1165 		}
1166 	} else if (p->myPipe->DCCEnable == true) {
1167 		s->Tvm_trips_rounded = s->LineTime / 4.0;
1168 		s->Tr0_trips_rounded = dml_ceil(4.0 * s->Tr0_trips / s->LineTime, 1.0) / 4.0 * s->LineTime;
1169 		*p->Tno_bw = 0;
1170 	} else {
1171 		s->Tvm_trips_rounded = s->LineTime / 4.0;
1172 		s->Tr0_trips_rounded = s->LineTime / 2.0;
1173 		*p->Tno_bw = 0;
1174 	}
1175 	s->Tvm_trips_rounded = dml_max(s->Tvm_trips_rounded, s->LineTime / 4.0);
1176 	s->Tr0_trips_rounded = dml_max(s->Tr0_trips_rounded, s->LineTime / 4.0);
1177 
1178 	if (p->myPipe->SourcePixelFormat == dml_420_8 || p->myPipe->SourcePixelFormat == dml_420_10 || p->myPipe->SourcePixelFormat == dml_420_12) {
1179 		s->bytes_pp = p->myPipe->BytePerPixelY + p->myPipe->BytePerPixelC / 4;
1180 	} else {
1181 		s->bytes_pp = p->myPipe->BytePerPixelY + p->myPipe->BytePerPixelC;
1182 	}
1183 
1184 	s->prefetch_bw_pr = s->bytes_pp * p->myPipe->PixelClock / (dml_float_t)p->myPipe->DPPPerSurface;
1185 	if (p->myPipe->VRatio < 1.0)
1186 		s->prefetch_bw_pr = p->myPipe->VRatio * s->prefetch_bw_pr;
1187 
1188 	s->max_Tsw = (dml_max(p->PrefetchSourceLinesY, p->PrefetchSourceLinesC) * s->LineTime);
1189 
1190 	s->prefetch_sw_bytes = p->PrefetchSourceLinesY * p->swath_width_luma_ub * p->myPipe->BytePerPixelY + p->PrefetchSourceLinesC * p->swath_width_chroma_ub * p->myPipe->BytePerPixelC;
1191 	s->prefetch_bw_oto = dml_max(s->prefetch_bw_pr, s->prefetch_sw_bytes / s->max_Tsw);
1192 
1193 	s->min_Lsw_oto = dml_max(p->PrefetchSourceLinesY, p->PrefetchSourceLinesC) / __DML_MAX_VRATIO_PRE_OTO__;
1194 	s->min_Lsw_oto = dml_max(s->min_Lsw_oto, 1.0);
1195 	s->Lsw_oto = dml_ceil(4.0 * dml_max(s->prefetch_sw_bytes / s->prefetch_bw_oto / s->LineTime, s->min_Lsw_oto), 1.0) / 4.0;
1196 
1197 	if (p->GPUVMEnable == true) {
1198 		s->Tvm_oto = dml_max3(
1199 			s->Tvm_trips,
1200 			*p->Tno_bw + p->PDEAndMetaPTEBytesFrame * p->HostVMInefficiencyFactor / s->prefetch_bw_oto,
1201 			s->LineTime / 4.0);
1202 	} else
1203 		s->Tvm_oto = s->LineTime / 4.0;
1204 
1205 	if ((p->GPUVMEnable == true || p->myPipe->DCCEnable == true)) {
1206 		s->Tr0_oto = dml_max4(
1207 			s->Tr0_trips,
1208 			(p->MetaRowByte + p->PixelPTEBytesPerRow * p->HostVMInefficiencyFactor) / s->prefetch_bw_oto,
1209 			(s->LineTime - s->Tvm_oto)/2.0,
1210 			s->LineTime / 4.0);
1211 #ifdef __DML_VBA_DEBUG__
1212 	dml_print("DML::%s: Tr0_oto max0 = %f\n", __func__, (p->MetaRowByte + p->PixelPTEBytesPerRow * p->HostVMInefficiencyFactor) / s->prefetch_bw_oto);
1213 	dml_print("DML::%s: Tr0_oto max1 = %f\n", __func__, s->Tr0_trips);
1214 	dml_print("DML::%s: Tr0_oto max2 = %f\n", __func__, s->LineTime - s->Tvm_oto);
1215 	dml_print("DML::%s: Tr0_oto max3 = %f\n", __func__, s->LineTime / 4);
1216 #endif
1217 	} else
1218 		s->Tr0_oto = (s->LineTime - s->Tvm_oto) / 2.0;
1219 
1220 	s->Tvm_oto_lines = dml_ceil(4.0 * s->Tvm_oto / s->LineTime, 1) / 4.0;
1221 	s->Tr0_oto_lines = dml_ceil(4.0 * s->Tr0_oto / s->LineTime, 1) / 4.0;
1222 	s->dst_y_prefetch_oto = s->Tvm_oto_lines + 2 * s->Tr0_oto_lines + s->Lsw_oto;
1223 
1224 	s->dst_y_prefetch_equ = p->VStartup - (*p->TSetup + dml_max(p->TWait + p->TCalc, *p->Tdmdl)) / s->LineTime - (*p->DSTYAfterScaler + (dml_float_t) *p->DSTXAfterScaler / (dml_float_t)p->myPipe->HTotal);
1225 
1226 #ifdef __DML_VBA_DEBUG__
1227 	dml_print("DML::%s: HTotal = %u\n", __func__, p->myPipe->HTotal);
1228 	dml_print("DML::%s: min_Lsw_oto = %f\n", __func__, s->min_Lsw_oto);
1229 	dml_print("DML::%s: *Tno_bw = %f\n", __func__, *p->Tno_bw);
1230 	dml_print("DML::%s: UrgentExtraLatency = %f\n", __func__, p->UrgentExtraLatency);
1231 	dml_print("DML::%s: trip_to_mem = %f\n", __func__, s->trip_to_mem);
1232 	dml_print("DML::%s: BytePerPixelY = %u\n", __func__, p->myPipe->BytePerPixelY);
1233 	dml_print("DML::%s: PrefetchSourceLinesY = %f\n", __func__, p->PrefetchSourceLinesY);
1234 	dml_print("DML::%s: swath_width_luma_ub = %u\n", __func__, p->swath_width_luma_ub);
1235 	dml_print("DML::%s: BytePerPixelC = %u\n", __func__, p->myPipe->BytePerPixelC);
1236 	dml_print("DML::%s: PrefetchSourceLinesC = %f\n", __func__, p->PrefetchSourceLinesC);
1237 	dml_print("DML::%s: swath_width_chroma_ub = %u\n", __func__, p->swath_width_chroma_ub);
1238 	dml_print("DML::%s: prefetch_sw_bytes = %f\n", __func__, s->prefetch_sw_bytes);
1239 	dml_print("DML::%s: bytes_pp = %f\n", __func__, s->bytes_pp);
1240 	dml_print("DML::%s: PDEAndMetaPTEBytesFrame = %u\n", __func__, p->PDEAndMetaPTEBytesFrame);
1241 	dml_print("DML::%s: MetaRowByte = %u\n", __func__, p->MetaRowByte);
1242 	dml_print("DML::%s: PixelPTEBytesPerRow = %u\n", __func__, p->PixelPTEBytesPerRow);
1243 	dml_print("DML::%s: HostVMInefficiencyFactor = %f\n", __func__, p->HostVMInefficiencyFactor);
1244 	dml_print("DML::%s: Tvm_trips = %f\n", __func__, s->Tvm_trips);
1245 	dml_print("DML::%s: Tr0_trips = %f\n", __func__, s->Tr0_trips);
1246 	dml_print("DML::%s: prefetch_bw_oto = %f\n", __func__, s->prefetch_bw_oto);
1247 	dml_print("DML::%s: Tr0_oto = %f\n", __func__, s->Tr0_oto);
1248 	dml_print("DML::%s: Tvm_oto = %f\n", __func__, s->Tvm_oto);
1249 	dml_print("DML::%s: Tvm_oto_lines = %f\n", __func__, s->Tvm_oto_lines);
1250 	dml_print("DML::%s: Tr0_oto_lines = %f\n", __func__, s->Tr0_oto_lines);
1251 	dml_print("DML::%s: Lsw_oto = %f\n", __func__, s->Lsw_oto);
1252 	dml_print("DML::%s: dst_y_prefetch_oto = %f\n", __func__, s->dst_y_prefetch_oto);
1253 	dml_print("DML::%s: dst_y_prefetch_equ = %f\n", __func__, s->dst_y_prefetch_equ);
1254 #endif
1255 
1256 	s->dst_y_prefetch_equ = dml_floor(4.0 * (s->dst_y_prefetch_equ + 0.125), 1) / 4.0;
1257 	s->Tpre_rounded = s->dst_y_prefetch_equ * s->LineTime;
1258 
1259 	dml_print("DML::%s: dst_y_prefetch_equ: %f (after round)\n", __func__, s->dst_y_prefetch_equ);
1260 
1261 	dml_print("DML::%s: LineTime: %f\n", __func__, s->LineTime);
1262 	dml_print("DML::%s: VStartup: %u\n", __func__, p->VStartup);
1263 	dml_print("DML::%s: Tvstartup: %fus - time between vstartup and first pixel of active\n", __func__, p->VStartup * s->LineTime);
1264 	dml_print("DML::%s: TSetup: %fus - time from vstartup to vready\n", __func__, *p->TSetup);
1265 	dml_print("DML::%s: TCalc: %fus - time for calculations in dchub starting at vready\n", __func__, p->TCalc);
1266 	dml_print("DML::%s: TWait: %fus - time for fabric to become ready max(pstate exit,cstate enter/exit, urgent latency) after TCalc\n", __func__, p->TWait);
1267 	dml_print("DML::%s: Tdmbf: %fus - time for dmd transfer from dchub to dio output buffer\n", __func__, s->Tdmbf);
1268 	dml_print("DML::%s: Tdmec: %fus - time dio takes to transfer dmd\n", __func__, s->Tdmec);
1269 	dml_print("DML::%s: Tdmsks: %fus - time before active dmd must complete transmission at dio\n", __func__, s->Tdmsks);
1270 	dml_print("DML::%s: Tdmdl_vm: %fus - time for vm stages of dmd \n", __func__, *p->Tdmdl_vm);
1271 	dml_print("DML::%s: Tdmdl: %fus - time for fabric to become ready and fetch dmd \n", __func__, *p->Tdmdl);
1272 	dml_print("DML::%s: DSTXAfterScaler: %u pixels - number of pixel clocks pipeline and buffer delay after scaler \n", __func__, *p->DSTXAfterScaler);
1273 	dml_print("DML::%s: DSTYAfterScaler: %u lines - number of lines of pipeline and buffer delay after scaler \n", __func__, *p->DSTYAfterScaler);
1274 
1275 	s->dep_bytes = dml_max(p->PDEAndMetaPTEBytesFrame * p->HostVMInefficiencyFactor, p->MetaRowByte + p->PixelPTEBytesPerRow * p->HostVMInefficiencyFactor);
1276 
1277 	if (s->prefetch_sw_bytes < s->dep_bytes) {
1278 		s->prefetch_sw_bytes = 2 * s->dep_bytes;
1279 	}
1280 
1281 	*p->DestinationLinesToRequestVMInVBlank = 0;
1282 	*p->DestinationLinesToRequestRowInVBlank = 0;
1283 	*p->VRatioPrefetchY = 0;
1284 	*p->VRatioPrefetchC = 0;
1285 	*p->RequiredPrefetchPixDataBWLuma = 0;
1286 	if (s->dst_y_prefetch_equ > 1) {
1287 
1288 		if (s->Tpre_rounded - *p->Tno_bw > 0) {
1289 		s->PrefetchBandwidth1 = (p->PDEAndMetaPTEBytesFrame * p->HostVMInefficiencyFactor + 2 * p->MetaRowByte
1290 					+ 2 * p->PixelPTEBytesPerRow * p->HostVMInefficiencyFactor
1291 					+ s->prefetch_sw_bytes)
1292 					/ (s->Tpre_rounded - *p->Tno_bw);
1293 			s->Tsw_est1 = s->prefetch_sw_bytes / s->PrefetchBandwidth1;
1294 		} else
1295 			s->PrefetchBandwidth1 = 0;
1296 
1297 		if (p->VStartup == p->MaxVStartup && (s->Tsw_est1 / s->LineTime < s->min_Lsw_oto) && s->Tpre_rounded - s->min_Lsw_oto * s->LineTime - 0.75 * s->LineTime - *p->Tno_bw > 0) {
1298 			s->PrefetchBandwidth1 = (p->PDEAndMetaPTEBytesFrame * p->HostVMInefficiencyFactor + 2 * p->MetaRowByte + 2 * p->PixelPTEBytesPerRow * p->HostVMInefficiencyFactor) /
1299 								(s->Tpre_rounded - s->min_Lsw_oto * s->LineTime - 0.75 * s->LineTime - *p->Tno_bw);
1300 		}
1301 
1302 		if (s->Tpre_rounded - *p->Tno_bw - 2 * s->Tr0_trips_rounded > 0)
1303 			s->PrefetchBandwidth2 = (p->PDEAndMetaPTEBytesFrame * p->HostVMInefficiencyFactor + s->prefetch_sw_bytes) /
1304 									(s->Tpre_rounded - *p->Tno_bw - 2 * s->Tr0_trips_rounded);
1305 		else
1306 			s->PrefetchBandwidth2 = 0;
1307 
1308 		if (s->Tpre_rounded - s->Tvm_trips_rounded > 0) {
1309 			s->PrefetchBandwidth3 = (2 * p->MetaRowByte + 2 * p->PixelPTEBytesPerRow * p->HostVMInefficiencyFactor + s->prefetch_sw_bytes) /
1310 									(s->Tpre_rounded - s->Tvm_trips_rounded);
1311 			s->Tsw_est3 = s->prefetch_sw_bytes / s->PrefetchBandwidth3;
1312 	}
1313 		else
1314 			s->PrefetchBandwidth3 = 0;
1315 
1316 
1317 		if (p->VStartup == p->MaxVStartup && (s->Tsw_est3 / s->LineTime < s->min_Lsw_oto) && s->Tpre_rounded - s->min_Lsw_oto * s->LineTime - 0.5 * s->LineTime - s->Tvm_trips_rounded > 0) {
1318 			s->PrefetchBandwidth3 = (2 * p->MetaRowByte + 2 * p->PixelPTEBytesPerRow * p->HostVMInefficiencyFactor) / (s->Tpre_rounded - s->min_Lsw_oto * s->LineTime - 0.5 * s->LineTime - s->Tvm_trips_rounded);
1319 		}
1320 
1321 		if (s->Tpre_rounded - s->Tvm_trips_rounded - 2 * s->Tr0_trips_rounded > 0)
1322 			s->PrefetchBandwidth4 = s->prefetch_sw_bytes / (s->Tpre_rounded - s->Tvm_trips_rounded - 2 * s->Tr0_trips_rounded);
1323 		else
1324 			s->PrefetchBandwidth4 = 0;
1325 
1326 #ifdef __DML_VBA_DEBUG__
1327 		dml_print("DML::%s: Tpre_rounded: %f\n", __func__, s->Tpre_rounded);
1328 		dml_print("DML::%s: Tno_bw: %f\n", __func__, *p->Tno_bw);
1329 		dml_print("DML::%s: Tvm_trips_rounded: %f\n", __func__, s->Tvm_trips_rounded);
1330 		dml_print("DML::%s: Tsw_est1: %f\n", __func__, s->Tsw_est1);
1331 		dml_print("DML::%s: Tsw_est3: %f\n", __func__, s->Tsw_est3);
1332 		dml_print("DML::%s: PrefetchBandwidth1: %f\n", __func__, s->PrefetchBandwidth1);
1333 		dml_print("DML::%s: PrefetchBandwidth2: %f\n", __func__, s->PrefetchBandwidth2);
1334 		dml_print("DML::%s: PrefetchBandwidth3: %f\n", __func__, s->PrefetchBandwidth3);
1335 		dml_print("DML::%s: PrefetchBandwidth4: %f\n", __func__, s->PrefetchBandwidth4);
1336 #endif
1337 		{
1338 			dml_bool_t Case1OK;
1339 			dml_bool_t Case2OK;
1340 			dml_bool_t Case3OK;
1341 
1342 			if (s->PrefetchBandwidth1 > 0) {
1343 				if (*p->Tno_bw + p->PDEAndMetaPTEBytesFrame * p->HostVMInefficiencyFactor / s->PrefetchBandwidth1 >= s->Tvm_trips_rounded && (p->MetaRowByte + p->PixelPTEBytesPerRow * p->HostVMInefficiencyFactor) / s->PrefetchBandwidth1 >= s->Tr0_trips_rounded) {
1344 					Case1OK = true;
1345 				} else {
1346 					Case1OK = false;
1347 				}
1348 			} else {
1349 				Case1OK = false;
1350 			}
1351 
1352 			if (s->PrefetchBandwidth2 > 0) {
1353 				if (*p->Tno_bw + p->PDEAndMetaPTEBytesFrame * p->HostVMInefficiencyFactor / s->PrefetchBandwidth2 >= s->Tvm_trips_rounded && (p->MetaRowByte + p->PixelPTEBytesPerRow * p->HostVMInefficiencyFactor) / s->PrefetchBandwidth2 < s->Tr0_trips_rounded) {
1354 					Case2OK = true;
1355 				} else {
1356 					Case2OK = false;
1357 				}
1358 			} else {
1359 				Case2OK = false;
1360 			}
1361 
1362 			if (s->PrefetchBandwidth3 > 0) {
1363 				if (*p->Tno_bw + p->PDEAndMetaPTEBytesFrame * p->HostVMInefficiencyFactor / s->PrefetchBandwidth3 < s->Tvm_trips_rounded && (p->MetaRowByte + p->PixelPTEBytesPerRow * p->HostVMInefficiencyFactor) / s->PrefetchBandwidth3 >= s->Tr0_trips_rounded) {
1364 					Case3OK = true;
1365 				} else {
1366 					Case3OK = false;
1367 				}
1368 			} else {
1369 				Case3OK = false;
1370 			}
1371 
1372 			if (Case1OK) {
1373 				s->prefetch_bw_equ = s->PrefetchBandwidth1;
1374 			} else if (Case2OK) {
1375 				s->prefetch_bw_equ = s->PrefetchBandwidth2;
1376 			} else if (Case3OK) {
1377 				s->prefetch_bw_equ = s->PrefetchBandwidth3;
1378 			} else {
1379 				s->prefetch_bw_equ = s->PrefetchBandwidth4;
1380 			}
1381 
1382 #ifdef __DML_VBA_DEBUG__
1383 			dml_print("DML::%s: Case1OK: %u\n", __func__, Case1OK);
1384 			dml_print("DML::%s: Case2OK: %u\n", __func__, Case2OK);
1385 			dml_print("DML::%s: Case3OK: %u\n", __func__, Case3OK);
1386 			dml_print("DML::%s: prefetch_bw_equ: %f\n", __func__, s->prefetch_bw_equ);
1387 #endif
1388 
1389 			if (s->prefetch_bw_equ > 0) {
1390 				if (p->GPUVMEnable == true) {
1391 					s->Tvm_equ = dml_max3(*p->Tno_bw + p->PDEAndMetaPTEBytesFrame * p->HostVMInefficiencyFactor / s->prefetch_bw_equ, s->Tvm_trips, s->LineTime / 4);
1392 				} else {
1393 					s->Tvm_equ = s->LineTime / 4;
1394 				}
1395 
1396 				if ((p->GPUVMEnable == true || p->myPipe->DCCEnable == true)) {
1397 					s->Tr0_equ = dml_max4((p->MetaRowByte + p->PixelPTEBytesPerRow * p->HostVMInefficiencyFactor) / s->prefetch_bw_equ, s->Tr0_trips, (s->LineTime - s->Tvm_equ) / 2, s->LineTime / 4);
1398 				} else {
1399 					s->Tr0_equ = (s->LineTime - s->Tvm_equ) / 2;
1400 				}
1401 			} else {
1402 				s->Tvm_equ = 0;
1403 				s->Tr0_equ = 0;
1404 				dml_print("DML::%s: prefetch_bw_equ equals 0!\n", __func__);
1405 			}
1406 		}
1407 
1408 
1409 		if (s->dst_y_prefetch_oto < s->dst_y_prefetch_equ) {
1410 			*p->DestinationLinesForPrefetch = s->dst_y_prefetch_oto;
1411 			s->TimeForFetchingMetaPTE = s->Tvm_oto;
1412 			s->TimeForFetchingRowInVBlank = s->Tr0_oto;
1413 
1414 			*p->DestinationLinesToRequestVMInVBlank = dml_ceil(4.0 * s->TimeForFetchingMetaPTE / s->LineTime, 1.0) / 4.0;
1415 			*p->DestinationLinesToRequestRowInVBlank = dml_ceil(4.0 * s->TimeForFetchingRowInVBlank / s->LineTime, 1.0) / 4.0;
1416 		} else {
1417 			*p->DestinationLinesForPrefetch = s->dst_y_prefetch_equ;
1418 			s->TimeForFetchingMetaPTE = s->Tvm_equ;
1419 			s->TimeForFetchingRowInVBlank = s->Tr0_equ;
1420 
1421 			if (p->VStartup == p->MaxVStartup && p->EnhancedPrefetchScheduleAccelerationFinal != 0) {
1422 				*p->DestinationLinesToRequestVMInVBlank = dml_floor(4.0 * s->TimeForFetchingMetaPTE / s->LineTime, 1.0) / 4.0;
1423 				*p->DestinationLinesToRequestRowInVBlank = dml_floor(4.0 * s->TimeForFetchingRowInVBlank / s->LineTime, 1.0) / 4.0;
1424 			} else {
1425 				*p->DestinationLinesToRequestVMInVBlank = dml_ceil(4.0 * s->TimeForFetchingMetaPTE / s->LineTime, 1.0) / 4.0;
1426 				*p->DestinationLinesToRequestRowInVBlank = dml_ceil(4.0 * s->TimeForFetchingRowInVBlank / s->LineTime, 1.0) / 4.0;
1427 			}
1428 		}
1429 
1430 		s->LinesToRequestPrefetchPixelData = *p->DestinationLinesForPrefetch - *p->DestinationLinesToRequestVMInVBlank - 2 * *p->DestinationLinesToRequestRowInVBlank;
1431 
1432 #ifdef __DML_VBA_DEBUG__
1433 		dml_print("DML::%s: DestinationLinesForPrefetch = %f\n", __func__, *p->DestinationLinesForPrefetch);
1434 		dml_print("DML::%s: DestinationLinesToRequestVMInVBlank = %f\n", __func__, *p->DestinationLinesToRequestVMInVBlank);
1435 		dml_print("DML::%s: TimeForFetchingRowInVBlank = %f\n", __func__, s->TimeForFetchingRowInVBlank);
1436 		dml_print("DML::%s: LineTime = %f\n", __func__, s->LineTime);
1437 		dml_print("DML::%s: DestinationLinesToRequestRowInVBlank = %f\n", __func__, *p->DestinationLinesToRequestRowInVBlank);
1438 		dml_print("DML::%s: PrefetchSourceLinesY = %f\n", __func__, p->PrefetchSourceLinesY);
1439 		dml_print("DML::%s: LinesToRequestPrefetchPixelData = %f\n", __func__, s->LinesToRequestPrefetchPixelData);
1440 #endif
1441 
1442 		if (s->LinesToRequestPrefetchPixelData >= 1 && s->prefetch_bw_equ > 0) {
1443 			*p->VRatioPrefetchY = (dml_float_t)p->PrefetchSourceLinesY / s->LinesToRequestPrefetchPixelData;
1444 			*p->VRatioPrefetchY = dml_max(*p->VRatioPrefetchY, 1.0);
1445 #ifdef __DML_VBA_DEBUG__
1446 			dml_print("DML::%s: VRatioPrefetchY = %f\n", __func__, *p->VRatioPrefetchY);
1447 			dml_print("DML::%s: SwathHeightY = %u\n", __func__, p->SwathHeightY);
1448 			dml_print("DML::%s: VInitPreFillY = %u\n", __func__, p->VInitPreFillY);
1449 #endif
1450 			if ((p->SwathHeightY > 4) && (p->VInitPreFillY > 3)) {
1451 				if (s->LinesToRequestPrefetchPixelData > (p->VInitPreFillY - 3.0) / 2.0) {
1452 					*p->VRatioPrefetchY = dml_max(*p->VRatioPrefetchY,
1453 										(dml_float_t)p->MaxNumSwathY * p->SwathHeightY / (s->LinesToRequestPrefetchPixelData - (p->VInitPreFillY - 3.0) / 2.0));
1454 			} else {
1455 				s->MyError = true;
1456 				dml_print("DML::%s: MyErr set. LinesToRequestPrefetchPixelData=%f VinitPreFillY=%u\n", __func__, s->LinesToRequestPrefetchPixelData, p->VInitPreFillY);
1457 				*p->VRatioPrefetchY = 0;
1458 			}
1459 #ifdef __DML_VBA_DEBUG__
1460 			dml_print("DML::%s: VRatioPrefetchY = %f\n", __func__, *p->VRatioPrefetchY);
1461 			dml_print("DML::%s: PrefetchSourceLinesY = %f\n", __func__, p->PrefetchSourceLinesY);
1462 			dml_print("DML::%s: MaxNumSwathY = %u\n", __func__, p->MaxNumSwathY);
1463 #endif
1464 			}
1465 
1466 			*p->VRatioPrefetchC = (dml_float_t)p->PrefetchSourceLinesC / s->LinesToRequestPrefetchPixelData;
1467 			*p->VRatioPrefetchC = dml_max(*p->VRatioPrefetchC, 1.0);
1468 
1469 #ifdef __DML_VBA_DEBUG__
1470 			dml_print("DML::%s: VRatioPrefetchC = %f\n", __func__, *p->VRatioPrefetchC);
1471 			dml_print("DML::%s: SwathHeightC = %u\n", __func__, p->SwathHeightC);
1472 			dml_print("DML::%s: VInitPreFillC = %u\n", __func__, p->VInitPreFillC);
1473 #endif
1474 			if ((p->SwathHeightC > 4) && (p->VInitPreFillC > 3)) {
1475 				if (s->LinesToRequestPrefetchPixelData > (p->VInitPreFillC - 3.0) / 2.0) {
1476 					*p->VRatioPrefetchC = dml_max(*p->VRatioPrefetchC, (dml_float_t)p->MaxNumSwathC * p->SwathHeightC / (s->LinesToRequestPrefetchPixelData - (p->VInitPreFillC - 3.0) / 2.0));
1477 				} else {
1478 					s->MyError = true;
1479 					dml_print("DML::%s: MyErr set. LinesToRequestPrefetchPixelData=%f VInitPreFillC=%u\n", __func__, s->LinesToRequestPrefetchPixelData, p->VInitPreFillC);
1480 					*p->VRatioPrefetchC = 0;
1481 				}
1482 #ifdef __DML_VBA_DEBUG__
1483 				dml_print("DML::%s: VRatioPrefetchC = %f\n", __func__, *p->VRatioPrefetchC);
1484 				dml_print("DML::%s: PrefetchSourceLinesC = %f\n", __func__, p->PrefetchSourceLinesC);
1485 				dml_print("DML::%s: MaxNumSwathC = %u\n", __func__, p->MaxNumSwathC);
1486 #endif
1487 			}
1488 
1489 			*p->RequiredPrefetchPixDataBWLuma = (dml_float_t)p->PrefetchSourceLinesY / s->LinesToRequestPrefetchPixelData
1490 				* p->myPipe->BytePerPixelY
1491 				* p->swath_width_luma_ub / s->LineTime;
1492 
1493 #ifdef __DML_VBA_DEBUG__
1494 			dml_print("DML::%s: BytePerPixelY = %u\n", __func__, p->myPipe->BytePerPixelY);
1495 			dml_print("DML::%s: swath_width_luma_ub = %u\n", __func__, p->swath_width_luma_ub);
1496 			dml_print("DML::%s: LineTime = %f\n", __func__, s->LineTime);
1497 			dml_print("DML::%s: RequiredPrefetchPixDataBWLuma = %f\n", __func__, *p->RequiredPrefetchPixDataBWLuma);
1498 #endif
1499 			*p->RequiredPrefetchPixDataBWChroma = (dml_float_t)p->PrefetchSourceLinesC / s->LinesToRequestPrefetchPixelData
1500 				*p->myPipe->BytePerPixelC
1501 				*p->swath_width_chroma_ub / s->LineTime;
1502 		} else {
1503 			s->MyError = true;
1504 			dml_print("DML:%s: MyErr set. LinesToRequestPrefetchPixelData: %f, should be > 0\n", __func__, s->LinesToRequestPrefetchPixelData);
1505 			*p->VRatioPrefetchY = 0;
1506 			*p->VRatioPrefetchC = 0;
1507 			*p->RequiredPrefetchPixDataBWLuma = 0;
1508 			*p->RequiredPrefetchPixDataBWChroma = 0;
1509 		}
1510 
1511 		dml_print("DML: Tpre: %fus - sum of time to request meta pte, 2 x data pte + meta data, swaths\n", (dml_float_t)s->LinesToRequestPrefetchPixelData * s->LineTime + 2.0 * s->TimeForFetchingRowInVBlank + s->TimeForFetchingMetaPTE);
1512 		dml_print("DML: Tvm: %fus - time to fetch page tables for meta surface\n", s->TimeForFetchingMetaPTE);
1513 		dml_print("DML: Tr0: %fus - time to fetch first row of data pagetables and first row of meta data (done in parallel)\n", s->TimeForFetchingRowInVBlank);
1514 		dml_print("DML: Tsw: %fus = time to fetch enough pixel data and cursor data to feed the scalers init position and detile\n", (dml_float_t)s->LinesToRequestPrefetchPixelData * s->LineTime);
1515 		dml_print("DML: To: %fus - time for propagation from scaler to optc\n", (*p->DSTYAfterScaler + ((dml_float_t) (*p->DSTXAfterScaler) / (dml_float_t)p->myPipe->HTotal)) * s->LineTime);
1516 		dml_print("DML: Tvstartup - TSetup - Tcalc - Twait - Tpre - To > 0\n");
1517 		dml_print("DML: Tslack(pre): %fus - time left over in schedule\n", p->VStartup * s->LineTime - s->TimeForFetchingMetaPTE - 2 * s->TimeForFetchingRowInVBlank - (*p->DSTYAfterScaler + ((dml_float_t) (*p->DSTXAfterScaler) / (dml_float_t)p->myPipe->HTotal)) * s->LineTime - p->TWait - p->TCalc - *p->TSetup);
1518 		dml_print("DML: row_bytes = dpte_row_bytes (per_pipe) = PixelPTEBytesPerRow = : %u\n", p->PixelPTEBytesPerRow);
1519 
1520 	} else {
1521 		s->MyError = true;
1522 		dml_print("DML::%s: MyErr set, dst_y_prefetch_equ = %f (should be > 1)\n", __func__, s->dst_y_prefetch_equ);
1523 		s->TimeForFetchingMetaPTE = 0;
1524 		s->TimeForFetchingRowInVBlank = 0;
1525 		*p->DestinationLinesToRequestVMInVBlank = 0;
1526 		*p->DestinationLinesToRequestRowInVBlank = 0;
1527 		s->LinesToRequestPrefetchPixelData = 0;
1528 		*p->VRatioPrefetchY = 0;
1529 		*p->VRatioPrefetchC = 0;
1530 		*p->RequiredPrefetchPixDataBWLuma = 0;
1531 		*p->RequiredPrefetchPixDataBWChroma = 0;
1532 	}
1533 
1534 	{
1535 		dml_float_t prefetch_vm_bw;
1536 		dml_float_t prefetch_row_bw;
1537 
1538 		if (p->PDEAndMetaPTEBytesFrame == 0) {
1539 			prefetch_vm_bw = 0;
1540 		} else if (*p->DestinationLinesToRequestVMInVBlank > 0) {
1541 #ifdef __DML_VBA_DEBUG__
1542 			dml_print("DML::%s: PDEAndMetaPTEBytesFrame = %u\n", __func__, p->PDEAndMetaPTEBytesFrame);
1543 			dml_print("DML::%s: HostVMInefficiencyFactor = %f\n", __func__, p->HostVMInefficiencyFactor);
1544 			dml_print("DML::%s: DestinationLinesToRequestVMInVBlank = %f\n", __func__, *p->DestinationLinesToRequestVMInVBlank);
1545 			dml_print("DML::%s: LineTime = %f\n", __func__, s->LineTime);
1546 #endif
1547 		prefetch_vm_bw = p->PDEAndMetaPTEBytesFrame * p->HostVMInefficiencyFactor / (*p->DestinationLinesToRequestVMInVBlank * s->LineTime);
1548 #ifdef __DML_VBA_DEBUG__
1549 			dml_print("DML::%s: prefetch_vm_bw = %f\n", __func__, prefetch_vm_bw);
1550 #endif
1551 		} else {
1552 			prefetch_vm_bw = 0;
1553 			s->MyError = true;
1554 			dml_print("DML::%s: MyErr set. DestinationLinesToRequestVMInVBlank=%f (should be > 0)\n", __func__, *p->DestinationLinesToRequestVMInVBlank);
1555 		}
1556 
1557 		if (p->MetaRowByte + p->PixelPTEBytesPerRow == 0) {
1558 			prefetch_row_bw = 0;
1559 		} else if (*p->DestinationLinesToRequestRowInVBlank > 0) {
1560 			prefetch_row_bw = (p->MetaRowByte + p->PixelPTEBytesPerRow * p->HostVMInefficiencyFactor) / (*p->DestinationLinesToRequestRowInVBlank * s->LineTime);
1561 
1562 #ifdef __DML_VBA_DEBUG__
1563 		dml_print("DML::%s: MetaRowByte = %u\n", __func__, p->MetaRowByte);
1564 		dml_print("DML::%s: PixelPTEBytesPerRow = %u\n", __func__, p->PixelPTEBytesPerRow);
1565 		dml_print("DML::%s: DestinationLinesToRequestRowInVBlank = %f\n", __func__, *p->DestinationLinesToRequestRowInVBlank);
1566 		dml_print("DML::%s: prefetch_row_bw = %f\n", __func__, prefetch_row_bw);
1567 #endif
1568 		} else {
1569 			prefetch_row_bw = 0;
1570 			s->MyError = true;
1571 			dml_print("DML::%s: MyErr set. DestinationLinesToRequestRowInVBlank=%f (should be > 0)\n", __func__, *p->DestinationLinesToRequestRowInVBlank);
1572 		}
1573 
1574 		*p->prefetch_vmrow_bw = dml_max(prefetch_vm_bw, prefetch_row_bw);
1575 	}
1576 
1577 	if (s->MyError) {
1578 		s->TimeForFetchingMetaPTE = 0;
1579 		s->TimeForFetchingRowInVBlank = 0;
1580 		*p->DestinationLinesToRequestVMInVBlank = 0;
1581 		*p->DestinationLinesToRequestRowInVBlank = 0;
1582 		*p->DestinationLinesForPrefetch = 0;
1583 		s->LinesToRequestPrefetchPixelData = 0;
1584 		*p->VRatioPrefetchY = 0;
1585 		*p->VRatioPrefetchC = 0;
1586 		*p->RequiredPrefetchPixDataBWLuma = 0;
1587 		*p->RequiredPrefetchPixDataBWChroma = 0;
1588 	}
1589 
1590 	return s->MyError;
1591 } // CalculatePrefetchSchedule
1592 
CalculateBytePerPixelAndBlockSizes(enum dml_source_format_class SourcePixelFormat,enum dml_swizzle_mode SurfaceTiling,dml_uint_t * BytePerPixelY,dml_uint_t * BytePerPixelC,dml_float_t * BytePerPixelDETY,dml_float_t * BytePerPixelDETC,dml_uint_t * BlockHeight256BytesY,dml_uint_t * BlockHeight256BytesC,dml_uint_t * BlockWidth256BytesY,dml_uint_t * BlockWidth256BytesC,dml_uint_t * MacroTileHeightY,dml_uint_t * MacroTileHeightC,dml_uint_t * MacroTileWidthY,dml_uint_t * MacroTileWidthC)1593 static void CalculateBytePerPixelAndBlockSizes(
1594 	enum dml_source_format_class SourcePixelFormat,
1595 	enum dml_swizzle_mode SurfaceTiling,
1596 
1597 	// Output
1598 	dml_uint_t *BytePerPixelY,
1599 	dml_uint_t *BytePerPixelC,
1600 	dml_float_t *BytePerPixelDETY,
1601 	dml_float_t *BytePerPixelDETC,
1602 	dml_uint_t *BlockHeight256BytesY,
1603 	dml_uint_t *BlockHeight256BytesC,
1604 	dml_uint_t *BlockWidth256BytesY,
1605 	dml_uint_t *BlockWidth256BytesC,
1606 	dml_uint_t *MacroTileHeightY,
1607 	dml_uint_t *MacroTileHeightC,
1608 	dml_uint_t *MacroTileWidthY,
1609 	dml_uint_t *MacroTileWidthC)
1610 {
1611 	if (SourcePixelFormat == dml_444_64) {
1612 		*BytePerPixelDETY = 8;
1613 		*BytePerPixelDETC = 0;
1614 		*BytePerPixelY = 8;
1615 		*BytePerPixelC = 0;
1616 	} else if (SourcePixelFormat == dml_444_32 || SourcePixelFormat == dml_rgbe) {
1617 		*BytePerPixelDETY = 4;
1618 		*BytePerPixelDETC = 0;
1619 		*BytePerPixelY = 4;
1620 		*BytePerPixelC = 0;
1621 	} else if (SourcePixelFormat == dml_444_16 || SourcePixelFormat == dml_mono_16) {
1622 		*BytePerPixelDETY = 2;
1623 		*BytePerPixelDETC = 0;
1624 		*BytePerPixelY = 2;
1625 		*BytePerPixelC = 0;
1626 	} else if (SourcePixelFormat == dml_444_8 || SourcePixelFormat == dml_mono_8) {
1627 		*BytePerPixelDETY = 1;
1628 		*BytePerPixelDETC = 0;
1629 		*BytePerPixelY = 1;
1630 		*BytePerPixelC = 0;
1631 	} else if (SourcePixelFormat == dml_rgbe_alpha) {
1632 		*BytePerPixelDETY = 4;
1633 		*BytePerPixelDETC = 1;
1634 		*BytePerPixelY = 4;
1635 		*BytePerPixelC = 1;
1636 	} else if (SourcePixelFormat == dml_420_8) {
1637 		*BytePerPixelDETY = 1;
1638 		*BytePerPixelDETC = 2;
1639 		*BytePerPixelY = 1;
1640 		*BytePerPixelC = 2;
1641 	} else if (SourcePixelFormat == dml_420_12) {
1642 		*BytePerPixelDETY = 2;
1643 		*BytePerPixelDETC = 4;
1644 		*BytePerPixelY = 2;
1645 		*BytePerPixelC = 4;
1646 	} else {
1647 		*BytePerPixelDETY = (dml_float_t) (4.0 / 3);
1648 		*BytePerPixelDETC = (dml_float_t) (8.0 / 3);
1649 		*BytePerPixelY = 2;
1650 		*BytePerPixelC = 4;
1651 	}
1652 #ifdef __DML_VBA_DEBUG__
1653 	dml_print("DML::%s: SourcePixelFormat = %u\n", __func__, SourcePixelFormat);
1654 	dml_print("DML::%s: BytePerPixelDETY = %f\n", __func__, *BytePerPixelDETY);
1655 	dml_print("DML::%s: BytePerPixelDETC = %f\n", __func__, *BytePerPixelDETC);
1656 	dml_print("DML::%s: BytePerPixelY = %u\n", __func__, *BytePerPixelY);
1657 	dml_print("DML::%s: BytePerPixelC = %u\n", __func__, *BytePerPixelC);
1658 #endif
1659 	if ((SourcePixelFormat == dml_444_64 || SourcePixelFormat == dml_444_32
1660 		|| SourcePixelFormat == dml_444_16
1661 		|| SourcePixelFormat == dml_444_8
1662 		|| SourcePixelFormat == dml_mono_16
1663 		|| SourcePixelFormat == dml_mono_8
1664 		|| SourcePixelFormat == dml_rgbe)) {
1665 		if (SurfaceTiling == dml_sw_linear) {
1666 			*BlockHeight256BytesY = 1;
1667 		} else if (SourcePixelFormat == dml_444_64) {
1668 			*BlockHeight256BytesY = 4;
1669 		} else if (SourcePixelFormat == dml_444_8) {
1670 			*BlockHeight256BytesY = 16;
1671 		} else {
1672 			*BlockHeight256BytesY = 8;
1673 		}
1674 		*BlockWidth256BytesY = 256U / *BytePerPixelY / *BlockHeight256BytesY;
1675 		*BlockHeight256BytesC = 0;
1676 		*BlockWidth256BytesC = 0;
1677 	} else {
1678 		if (SurfaceTiling == dml_sw_linear) {
1679 			*BlockHeight256BytesY = 1;
1680 			*BlockHeight256BytesC = 1;
1681 		} else if (SourcePixelFormat == dml_rgbe_alpha) {
1682 			*BlockHeight256BytesY = 8;
1683 			*BlockHeight256BytesC = 16;
1684 		} else if (SourcePixelFormat == dml_420_8) {
1685 			*BlockHeight256BytesY = 16;
1686 			*BlockHeight256BytesC = 8;
1687 		} else {
1688 			*BlockHeight256BytesY = 8;
1689 			*BlockHeight256BytesC = 8;
1690 		}
1691 		*BlockWidth256BytesY = 256U / *BytePerPixelY / *BlockHeight256BytesY;
1692 		*BlockWidth256BytesC = 256U / *BytePerPixelC / *BlockHeight256BytesC;
1693 	}
1694 #ifdef __DML_VBA_DEBUG__
1695 	dml_print("DML::%s: BlockWidth256BytesY = %u\n", __func__, *BlockWidth256BytesY);
1696 	dml_print("DML::%s: BlockHeight256BytesY = %u\n", __func__, *BlockHeight256BytesY);
1697 	dml_print("DML::%s: BlockWidth256BytesC = %u\n", __func__, *BlockWidth256BytesC);
1698 	dml_print("DML::%s: BlockHeight256BytesC = %u\n", __func__, *BlockHeight256BytesC);
1699 #endif
1700 
1701 	if (SurfaceTiling == dml_sw_linear) {
1702 		*MacroTileHeightY = *BlockHeight256BytesY;
1703 		*MacroTileWidthY = 256 / *BytePerPixelY / *MacroTileHeightY;
1704 		*MacroTileHeightC = *BlockHeight256BytesC;
1705 		if (*MacroTileHeightC == 0) {
1706 			*MacroTileWidthC = 0;
1707 		} else {
1708 			*MacroTileWidthC = 256 / *BytePerPixelC / *MacroTileHeightC;
1709 	}
1710 	} else if (SurfaceTiling == dml_sw_64kb_d || SurfaceTiling == dml_sw_64kb_d_t || SurfaceTiling == dml_sw_64kb_d_x || SurfaceTiling == dml_sw_64kb_r_x) {
1711 		*MacroTileHeightY = 16 * *BlockHeight256BytesY;
1712 		*MacroTileWidthY = 65536 / *BytePerPixelY / *MacroTileHeightY;
1713 		*MacroTileHeightC = 16 * *BlockHeight256BytesC;
1714 		if (*MacroTileHeightC == 0) {
1715 			*MacroTileWidthC = 0;
1716 		} else {
1717 			*MacroTileWidthC = 65536 / *BytePerPixelC / *MacroTileHeightC;
1718 		}
1719 	} else {
1720 		*MacroTileHeightY = 32 * *BlockHeight256BytesY;
1721 		*MacroTileWidthY = 65536 * 4 / *BytePerPixelY / *MacroTileHeightY;
1722 		*MacroTileHeightC = 32 * *BlockHeight256BytesC;
1723 		if (*MacroTileHeightC == 0) {
1724 			*MacroTileWidthC = 0;
1725 		} else {
1726 			*MacroTileWidthC = 65536 * 4 / *BytePerPixelC / *MacroTileHeightC;
1727 		}
1728 	}
1729 
1730 #ifdef __DML_VBA_DEBUG__
1731 	dml_print("DML::%s: MacroTileWidthY = %u\n", __func__, *MacroTileWidthY);
1732 	dml_print("DML::%s: MacroTileHeightY = %u\n", __func__, *MacroTileHeightY);
1733 	dml_print("DML::%s: MacroTileWidthC = %u\n", __func__, *MacroTileWidthC);
1734 	dml_print("DML::%s: MacroTileHeightC = %u\n", __func__, *MacroTileHeightC);
1735 #endif
1736 } // CalculateBytePerPixelAndBlockSizes
1737 
CalculateTWait(dml_uint_t PrefetchMode,enum dml_use_mall_for_pstate_change_mode UseMALLForPStateChange,dml_bool_t SynchronizeDRRDisplaysForUCLKPStateChangeFinal,dml_bool_t DRRDisplay,dml_float_t DRAMClockChangeLatency,dml_float_t FCLKChangeLatency,dml_float_t UrgentLatency,dml_float_t SREnterPlusExitTime)1738 static dml_float_t CalculateTWait(
1739 		dml_uint_t PrefetchMode,
1740 		enum dml_use_mall_for_pstate_change_mode UseMALLForPStateChange,
1741 		dml_bool_t SynchronizeDRRDisplaysForUCLKPStateChangeFinal,
1742 		dml_bool_t DRRDisplay,
1743 		dml_float_t DRAMClockChangeLatency,
1744 		dml_float_t FCLKChangeLatency,
1745 		dml_float_t UrgentLatency,
1746 		dml_float_t SREnterPlusExitTime)
1747 {
1748 	dml_float_t TWait = 0.0;
1749 
1750 	if (PrefetchMode == 0 &&
1751 			!(UseMALLForPStateChange == dml_use_mall_pstate_change_full_frame) && !(UseMALLForPStateChange == dml_use_mall_pstate_change_sub_viewport) &&
1752 			!(UseMALLForPStateChange == dml_use_mall_pstate_change_phantom_pipe) && !(SynchronizeDRRDisplaysForUCLKPStateChangeFinal && DRRDisplay)) {
1753 		TWait = dml_max3(DRAMClockChangeLatency + UrgentLatency, SREnterPlusExitTime, UrgentLatency);
1754 	} else if (PrefetchMode <= 1 && !(UseMALLForPStateChange == dml_use_mall_pstate_change_phantom_pipe)) {
1755 		TWait = dml_max3(FCLKChangeLatency + UrgentLatency, SREnterPlusExitTime, UrgentLatency);
1756 	} else if (PrefetchMode <= 2 && !(UseMALLForPStateChange == dml_use_mall_pstate_change_phantom_pipe)) {
1757 		TWait = dml_max(SREnterPlusExitTime, UrgentLatency);
1758 	} else {
1759 		TWait = UrgentLatency;
1760 	}
1761 
1762 #ifdef __DML_VBA_DEBUG__
1763 	dml_print("DML::%s: PrefetchMode = %u\n", __func__, PrefetchMode);
1764 	dml_print("DML::%s: TWait = %f\n", __func__, TWait);
1765 #endif
1766 	return TWait;
1767 } // CalculateTWait
1768 
1769 
1770 /// @brief Calculate the "starting point" for prefetch calculation
1771 ///  if AllowForPStateChangeOrStutterInVBlank is set as a particular requirement, then the mode evalulation
1772 ///  will only be done at the given mode. If no specific requirement (i.e. *_if_possible), then will just go from
1773 ///  try all the prefetch mode in decreasing order of "difficulty" (start from 0 which means all power saving
1774 ///  features).
CalculatePrefetchMode(enum dml_prefetch_modes AllowForPStateChangeOrStutterInVBlank,dml_uint_t * MinPrefetchMode,dml_uint_t * MaxPrefetchMode)1775 static void CalculatePrefetchMode(
1776 		enum dml_prefetch_modes AllowForPStateChangeOrStutterInVBlank,
1777 		dml_uint_t *MinPrefetchMode,
1778 		dml_uint_t *MaxPrefetchMode)
1779 {
1780 	if (AllowForPStateChangeOrStutterInVBlank == dml_prefetch_support_uclk_fclk_and_stutter_if_possible) {
1781 		*MinPrefetchMode = 0;   // consider all pwr saving features
1782 		*MaxPrefetchMode = 3;   // consider just urgent latency
1783 	} else {
1784 		if (AllowForPStateChangeOrStutterInVBlank == dml_prefetch_support_none) {
1785 			*MinPrefetchMode = 3;
1786 		} else if (AllowForPStateChangeOrStutterInVBlank == dml_prefetch_support_stutter) {
1787 			*MinPrefetchMode = 2;
1788 		} else if (AllowForPStateChangeOrStutterInVBlank == dml_prefetch_support_fclk_and_stutter) {
1789 			*MinPrefetchMode = 1;
1790 		} else if (AllowForPStateChangeOrStutterInVBlank == dml_prefetch_support_uclk_fclk_and_stutter) {
1791 			*MinPrefetchMode = 0;
1792 		} else {
1793 			dml_print("ERROR: Invalid AllowForPStateChangeOrStutterInVBlank setting! val=%u\n", AllowForPStateChangeOrStutterInVBlank);
1794 			ASSERT(0);
1795 		}
1796 		*MaxPrefetchMode = *MinPrefetchMode;
1797 	}
1798 } // CalculatePrefetchMode
1799 
CalculateWriteBackDISPCLK(enum dml_source_format_class WritebackPixelFormat,dml_float_t PixelClock,dml_float_t WritebackHRatio,dml_float_t WritebackVRatio,dml_uint_t WritebackHTaps,dml_uint_t WritebackVTaps,dml_uint_t WritebackSourceWidth,dml_uint_t WritebackDestinationWidth,dml_uint_t HTotal,dml_uint_t WritebackLineBufferSize,dml_float_t DISPCLKDPPCLKVCOSpeed)1800 static dml_float_t CalculateWriteBackDISPCLK(
1801 		enum dml_source_format_class WritebackPixelFormat,
1802 		dml_float_t PixelClock,
1803 		dml_float_t WritebackHRatio,
1804 		dml_float_t WritebackVRatio,
1805 		dml_uint_t WritebackHTaps,
1806 		dml_uint_t WritebackVTaps,
1807 		dml_uint_t WritebackSourceWidth,
1808 		dml_uint_t WritebackDestinationWidth,
1809 		dml_uint_t HTotal,
1810 		dml_uint_t WritebackLineBufferSize,
1811 		dml_float_t DISPCLKDPPCLKVCOSpeed)
1812 {
1813 	dml_float_t DISPCLK_H, DISPCLK_V, DISPCLK_HB;
1814 
1815 	DISPCLK_H = PixelClock * dml_ceil(WritebackHTaps / 8.0, 1) / WritebackHRatio;
1816 	DISPCLK_V = PixelClock * (WritebackVTaps * dml_ceil(WritebackDestinationWidth / 6.0, 1) + 8.0) / (dml_float_t) HTotal;
1817 	DISPCLK_HB = PixelClock * WritebackVTaps * (WritebackDestinationWidth * WritebackVTaps - WritebackLineBufferSize / 57.0) / 6.0 / (dml_float_t) WritebackSourceWidth;
1818 	return RoundToDFSGranularity(dml_max3(DISPCLK_H, DISPCLK_V, DISPCLK_HB), 1, DISPCLKDPPCLKVCOSpeed);
1819 }
1820 
CalculateWriteBackDelay(enum dml_source_format_class WritebackPixelFormat,dml_float_t WritebackHRatio,dml_float_t WritebackVRatio,dml_uint_t WritebackVTaps,dml_uint_t WritebackDestinationWidth,dml_uint_t WritebackDestinationHeight,dml_uint_t WritebackSourceHeight,dml_uint_t HTotal)1821 static dml_float_t CalculateWriteBackDelay(
1822 		enum dml_source_format_class WritebackPixelFormat,
1823 		dml_float_t WritebackHRatio,
1824 		dml_float_t WritebackVRatio,
1825 		dml_uint_t WritebackVTaps,
1826 		dml_uint_t WritebackDestinationWidth,
1827 		dml_uint_t WritebackDestinationHeight,
1828 		dml_uint_t WritebackSourceHeight,
1829 		dml_uint_t HTotal)
1830 {
1831 	dml_float_t CalculateWriteBackDelay;
1832 	dml_float_t Line_length;
1833 	dml_float_t Output_lines_last_notclamped;
1834 	dml_float_t WritebackVInit;
1835 
1836 	WritebackVInit = (WritebackVRatio + WritebackVTaps + 1) / 2;
1837 	Line_length = dml_max((dml_float_t) WritebackDestinationWidth, dml_ceil((dml_float_t)WritebackDestinationWidth / 6.0, 1.0) * WritebackVTaps);
1838 	Output_lines_last_notclamped = WritebackDestinationHeight - 1 - dml_ceil(((dml_float_t)WritebackSourceHeight - (dml_float_t) WritebackVInit) / (dml_float_t)WritebackVRatio, 1.0);
1839 	if (Output_lines_last_notclamped < 0) {
1840 		CalculateWriteBackDelay = 0;
1841 	} else {
1842 		CalculateWriteBackDelay = Output_lines_last_notclamped * Line_length + (HTotal - WritebackDestinationWidth) + 80;
1843 	}
1844 	return CalculateWriteBackDelay;
1845 }
1846 
CalculateVUpdateAndDynamicMetadataParameters(dml_uint_t MaxInterDCNTileRepeaters,dml_float_t Dppclk,dml_float_t Dispclk,dml_float_t DCFClkDeepSleep,dml_float_t PixelClock,dml_uint_t HTotal,dml_uint_t VBlank,dml_uint_t DynamicMetadataTransmittedBytes,dml_uint_t DynamicMetadataLinesBeforeActiveRequired,dml_uint_t InterlaceEnable,dml_bool_t ProgressiveToInterlaceUnitInOPP,dml_float_t * TSetup,dml_float_t * Tdmbf,dml_float_t * Tdmec,dml_float_t * Tdmsks,dml_uint_t * VUpdateOffsetPix,dml_uint_t * VUpdateWidthPix,dml_uint_t * VReadyOffsetPix)1847 static void CalculateVUpdateAndDynamicMetadataParameters(
1848 		dml_uint_t MaxInterDCNTileRepeaters,
1849 		dml_float_t Dppclk,
1850 		dml_float_t Dispclk,
1851 		dml_float_t DCFClkDeepSleep,
1852 		dml_float_t PixelClock,
1853 		dml_uint_t HTotal,
1854 		dml_uint_t VBlank,
1855 		dml_uint_t DynamicMetadataTransmittedBytes,
1856 		dml_uint_t DynamicMetadataLinesBeforeActiveRequired,
1857 		dml_uint_t InterlaceEnable,
1858 		dml_bool_t ProgressiveToInterlaceUnitInOPP,
1859 
1860 		// Output
1861 		dml_float_t *TSetup,
1862 		dml_float_t *Tdmbf,
1863 		dml_float_t *Tdmec,
1864 		dml_float_t *Tdmsks,
1865 		dml_uint_t *VUpdateOffsetPix,
1866 		dml_uint_t *VUpdateWidthPix,
1867 		dml_uint_t *VReadyOffsetPix)
1868 {
1869 	dml_float_t TotalRepeaterDelayTime;
1870 	TotalRepeaterDelayTime = MaxInterDCNTileRepeaters * (2 / Dppclk + 3 / Dispclk);
1871 	*VUpdateWidthPix = (dml_uint_t)(dml_ceil((14.0 / DCFClkDeepSleep + 12.0 / Dppclk + TotalRepeaterDelayTime) * PixelClock, 1.0));
1872 	*VReadyOffsetPix = (dml_uint_t)(dml_ceil(dml_max(150.0 / Dppclk, TotalRepeaterDelayTime + 20.0 / DCFClkDeepSleep + 10.0 / Dppclk) * PixelClock, 1.0));
1873 	*VUpdateOffsetPix = (dml_uint_t)(dml_ceil(HTotal / 4.0, 1.0));
1874 	*TSetup = (*VUpdateOffsetPix + *VUpdateWidthPix + *VReadyOffsetPix) / PixelClock;
1875 	*Tdmbf = DynamicMetadataTransmittedBytes / 4.0 / Dispclk;
1876 	*Tdmec = HTotal / PixelClock;
1877 
1878 	if (DynamicMetadataLinesBeforeActiveRequired == 0) {
1879 		*Tdmsks = VBlank * HTotal / PixelClock / 2.0;
1880 	} else {
1881 		*Tdmsks = DynamicMetadataLinesBeforeActiveRequired * HTotal / PixelClock;
1882 	}
1883 	if (InterlaceEnable == 1 && ProgressiveToInterlaceUnitInOPP == false) {
1884 		*Tdmsks = *Tdmsks / 2;
1885 	}
1886 #ifdef __DML_VBA_DEBUG__
1887 	dml_print("DML::%s: DynamicMetadataLinesBeforeActiveRequired = %u\n", __func__, DynamicMetadataLinesBeforeActiveRequired);
1888 	dml_print("DML::%s: VBlank = %u\n", __func__, VBlank);
1889 	dml_print("DML::%s: HTotal = %u\n", __func__, HTotal);
1890 	dml_print("DML::%s: PixelClock = %f\n", __func__, PixelClock);
1891 	dml_print("DML::%s: Dppclk = %f\n", __func__, Dppclk);
1892 	dml_print("DML::%s: DCFClkDeepSleep = %f\n", __func__, DCFClkDeepSleep);
1893 	dml_print("DML::%s: MaxInterDCNTileRepeaters = %u\n", __func__, MaxInterDCNTileRepeaters);
1894 	dml_print("DML::%s: TotalRepeaterDelayTime = %f\n", __func__, TotalRepeaterDelayTime);
1895 
1896 	dml_print("DML::%s: VUpdateWidthPix = %u\n", __func__, *VUpdateWidthPix);
1897 	dml_print("DML::%s: VReadyOffsetPix = %u\n", __func__, *VReadyOffsetPix);
1898 	dml_print("DML::%s: VUpdateOffsetPix = %u\n", __func__, *VUpdateOffsetPix);
1899 
1900 	dml_print("DML::%s: Tdmsks = %f\n", __func__, *Tdmsks);
1901 #endif
1902 }
1903 
CalculateRowBandwidth(dml_bool_t GPUVMEnable,enum dml_source_format_class SourcePixelFormat,dml_float_t VRatio,dml_float_t VRatioChroma,dml_bool_t DCCEnable,dml_float_t LineTime,dml_uint_t MetaRowByteLuma,dml_uint_t MetaRowByteChroma,dml_uint_t meta_row_height_luma,dml_uint_t meta_row_height_chroma,dml_uint_t PixelPTEBytesPerRowLuma,dml_uint_t PixelPTEBytesPerRowChroma,dml_uint_t dpte_row_height_luma,dml_uint_t dpte_row_height_chroma,dml_float_t * meta_row_bw,dml_float_t * dpte_row_bw)1904 static void CalculateRowBandwidth(
1905 		dml_bool_t GPUVMEnable,
1906 		enum dml_source_format_class SourcePixelFormat,
1907 		dml_float_t VRatio,
1908 		dml_float_t VRatioChroma,
1909 		dml_bool_t DCCEnable,
1910 		dml_float_t LineTime,
1911 		dml_uint_t MetaRowByteLuma,
1912 		dml_uint_t MetaRowByteChroma,
1913 		dml_uint_t meta_row_height_luma,
1914 		dml_uint_t meta_row_height_chroma,
1915 		dml_uint_t PixelPTEBytesPerRowLuma,
1916 		dml_uint_t PixelPTEBytesPerRowChroma,
1917 		dml_uint_t dpte_row_height_luma,
1918 		dml_uint_t dpte_row_height_chroma,
1919 		// Output
1920 		dml_float_t *meta_row_bw,
1921 		dml_float_t *dpte_row_bw)
1922 {
1923 	if (DCCEnable != true) {
1924 		*meta_row_bw = 0;
1925 	} else if (SourcePixelFormat == dml_420_8 || SourcePixelFormat == dml_420_10 || SourcePixelFormat == dml_420_12 || SourcePixelFormat == dml_rgbe_alpha) {
1926 		*meta_row_bw = VRatio * MetaRowByteLuma / (meta_row_height_luma * LineTime)
1927 				+ VRatioChroma * MetaRowByteChroma
1928 					/ (meta_row_height_chroma * LineTime);
1929 	} else {
1930 		*meta_row_bw = VRatio * MetaRowByteLuma / (meta_row_height_luma * LineTime);
1931 	}
1932 
1933 	if (GPUVMEnable != true) {
1934 		*dpte_row_bw = 0;
1935 	} else if (SourcePixelFormat == dml_420_8 || SourcePixelFormat == dml_420_10 || SourcePixelFormat == dml_420_12 || SourcePixelFormat == dml_rgbe_alpha) {
1936 		*dpte_row_bw = VRatio * PixelPTEBytesPerRowLuma / (dpte_row_height_luma * LineTime)
1937 				+ VRatioChroma * PixelPTEBytesPerRowChroma
1938 					/ (dpte_row_height_chroma * LineTime);
1939 	} else {
1940 		*dpte_row_bw = VRatio * PixelPTEBytesPerRowLuma / (dpte_row_height_luma * LineTime);
1941 	}
1942 }
1943 
1944 /// @brief Determine immediate flip schedule given bw remaining after considering the prefetch schedule
1945 /// @param BandwidthAvailableForImmediateFlip Bandwidth available for iflip for all planes
CalculateFlipSchedule(dml_float_t HostVMInefficiencyFactor,dml_float_t UrgentExtraLatency,dml_float_t UrgentLatency,dml_uint_t GPUVMMaxPageTableLevels,dml_bool_t HostVMEnable,dml_uint_t HostVMMaxNonCachedPageTableLevels,dml_bool_t GPUVMEnable,dml_uint_t HostVMMinPageSize,dml_float_t PDEAndMetaPTEBytesPerFrame,dml_float_t MetaRowBytes,dml_float_t DPTEBytesPerRow,dml_float_t BandwidthAvailableForImmediateFlip,dml_uint_t TotImmediateFlipBytes,enum dml_source_format_class SourcePixelFormat,dml_float_t LineTime,dml_float_t VRatio,dml_float_t VRatioChroma,dml_float_t Tno_bw,dml_bool_t DCCEnable,dml_uint_t dpte_row_height,dml_uint_t meta_row_height,dml_uint_t dpte_row_height_chroma,dml_uint_t meta_row_height_chroma,dml_bool_t use_one_row_for_frame_flip,dml_float_t * DestinationLinesToRequestVMInImmediateFlip,dml_float_t * DestinationLinesToRequestRowInImmediateFlip,dml_float_t * final_flip_bw,dml_bool_t * ImmediateFlipSupportedForPipe)1946 static void CalculateFlipSchedule(
1947 		dml_float_t HostVMInefficiencyFactor,
1948 		dml_float_t UrgentExtraLatency,
1949 		dml_float_t UrgentLatency,
1950 		dml_uint_t GPUVMMaxPageTableLevels,
1951 		dml_bool_t HostVMEnable,
1952 		dml_uint_t HostVMMaxNonCachedPageTableLevels,
1953 		dml_bool_t GPUVMEnable,
1954 		dml_uint_t HostVMMinPageSize,
1955 		dml_float_t PDEAndMetaPTEBytesPerFrame,
1956 		dml_float_t MetaRowBytes,
1957 		dml_float_t DPTEBytesPerRow,
1958 		dml_float_t BandwidthAvailableForImmediateFlip,
1959 		dml_uint_t TotImmediateFlipBytes,
1960 		enum dml_source_format_class SourcePixelFormat,
1961 		dml_float_t LineTime,
1962 		dml_float_t VRatio,
1963 		dml_float_t VRatioChroma,
1964 		dml_float_t Tno_bw,
1965 		dml_bool_t DCCEnable,
1966 		dml_uint_t dpte_row_height,
1967 		dml_uint_t meta_row_height,
1968 		dml_uint_t dpte_row_height_chroma,
1969 		dml_uint_t meta_row_height_chroma,
1970 		dml_bool_t use_one_row_for_frame_flip,
1971 
1972 		// Output
1973 		dml_float_t *DestinationLinesToRequestVMInImmediateFlip,
1974 		dml_float_t *DestinationLinesToRequestRowInImmediateFlip,
1975 		dml_float_t *final_flip_bw,
1976 		dml_bool_t *ImmediateFlipSupportedForPipe)
1977 {
1978 	dml_float_t min_row_time = 0.0;
1979 	dml_uint_t HostVMDynamicLevelsTrips = 0;
1980 	dml_float_t TimeForFetchingMetaPTEImmediateFlip = 0;
1981 	dml_float_t TimeForFetchingRowInVBlankImmediateFlip = 0;
1982 	dml_float_t ImmediateFlipBW = 0; // @brief The immediate flip bandwidth for this pipe
1983 
1984 	if (GPUVMEnable == true && HostVMEnable == true) {
1985 		HostVMDynamicLevelsTrips = HostVMMaxNonCachedPageTableLevels;
1986 	} else {
1987 		HostVMDynamicLevelsTrips = 0;
1988 	}
1989 
1990 #ifdef __DML_VBA_DEBUG__
1991 	dml_print("DML::%s: TotImmediateFlipBytes = %u\n", __func__, TotImmediateFlipBytes);
1992 	dml_print("DML::%s: HostVMInefficiencyFactor = %f\n", __func__, HostVMInefficiencyFactor);
1993 	dml_print("DML::%s: UrgentLatency = %f\n", __func__, UrgentLatency);
1994 	dml_print("DML::%s: BandwidthAvailableForImmediateFlip = %f\n", __func__, BandwidthAvailableForImmediateFlip);
1995 #endif
1996 
1997 	if (TotImmediateFlipBytes > 0) {
1998 		if (use_one_row_for_frame_flip) {
1999 			ImmediateFlipBW = (PDEAndMetaPTEBytesPerFrame + MetaRowBytes + 2.0 * DPTEBytesPerRow) * BandwidthAvailableForImmediateFlip / (dml_float_t) TotImmediateFlipBytes;
2000 		} else {
2001 			ImmediateFlipBW = (PDEAndMetaPTEBytesPerFrame + MetaRowBytes + DPTEBytesPerRow) * BandwidthAvailableForImmediateFlip / (dml_float_t) TotImmediateFlipBytes;
2002 		}
2003 		if (GPUVMEnable == true) {
2004 			TimeForFetchingMetaPTEImmediateFlip = dml_max3(Tno_bw + PDEAndMetaPTEBytesPerFrame * HostVMInefficiencyFactor / ImmediateFlipBW,
2005 														UrgentExtraLatency + UrgentLatency * (GPUVMMaxPageTableLevels * (HostVMDynamicLevelsTrips + 1) - 1),
2006 														LineTime / 4.0);
2007 		} else {
2008 			TimeForFetchingMetaPTEImmediateFlip = 0;
2009 		}
2010 		if ((GPUVMEnable == true || DCCEnable == true)) {
2011 			TimeForFetchingRowInVBlankImmediateFlip = dml_max3((MetaRowBytes + DPTEBytesPerRow * HostVMInefficiencyFactor) / ImmediateFlipBW, UrgentLatency * (HostVMDynamicLevelsTrips + 1), LineTime / 4.0);
2012 		} else {
2013 			TimeForFetchingRowInVBlankImmediateFlip = 0;
2014 		}
2015 
2016 		*DestinationLinesToRequestVMInImmediateFlip = dml_ceil(4.0 * (TimeForFetchingMetaPTEImmediateFlip / LineTime), 1.0) / 4.0;
2017 		*DestinationLinesToRequestRowInImmediateFlip = dml_ceil(4.0 * (TimeForFetchingRowInVBlankImmediateFlip / LineTime), 1.0) / 4.0;
2018 
2019 		if (GPUVMEnable == true) {
2020 			*final_flip_bw = dml_max(PDEAndMetaPTEBytesPerFrame * HostVMInefficiencyFactor / (*DestinationLinesToRequestVMInImmediateFlip * LineTime),
2021 								(MetaRowBytes + DPTEBytesPerRow * HostVMInefficiencyFactor) / (*DestinationLinesToRequestRowInImmediateFlip * LineTime));
2022 		} else if ((GPUVMEnable == true || DCCEnable == true)) {
2023 			*final_flip_bw = (MetaRowBytes + DPTEBytesPerRow * HostVMInefficiencyFactor) / (*DestinationLinesToRequestRowInImmediateFlip * LineTime);
2024 		} else {
2025 			*final_flip_bw = 0;
2026 		}
2027 	} else {
2028 		TimeForFetchingMetaPTEImmediateFlip = 0;
2029 		TimeForFetchingRowInVBlankImmediateFlip = 0;
2030 		*DestinationLinesToRequestVMInImmediateFlip = 0;
2031 		*DestinationLinesToRequestRowInImmediateFlip = 0;
2032 		*final_flip_bw = 0;
2033 	}
2034 
2035 	if (SourcePixelFormat == dml_420_8 || SourcePixelFormat == dml_420_10 || SourcePixelFormat == dml_rgbe_alpha) {
2036 		if (GPUVMEnable == true && DCCEnable != true) {
2037 			min_row_time = dml_min(dpte_row_height * LineTime / VRatio, dpte_row_height_chroma * LineTime / VRatioChroma);
2038 		} else if (GPUVMEnable != true && DCCEnable == true) {
2039 			min_row_time = dml_min(meta_row_height * LineTime / VRatio, meta_row_height_chroma * LineTime / VRatioChroma);
2040 		} else {
2041 			min_row_time = dml_min4(dpte_row_height * LineTime / VRatio, meta_row_height * LineTime / VRatio, dpte_row_height_chroma * LineTime / VRatioChroma, meta_row_height_chroma * LineTime / VRatioChroma);
2042 		}
2043 	} else {
2044 		if (GPUVMEnable == true && DCCEnable != true) {
2045 			min_row_time = dpte_row_height * LineTime / VRatio;
2046 		} else if (GPUVMEnable != true && DCCEnable == true) {
2047 			min_row_time = meta_row_height * LineTime / VRatio;
2048 		} else {
2049 			min_row_time = dml_min(dpte_row_height * LineTime / VRatio, meta_row_height * LineTime / VRatio);
2050 		}
2051 	}
2052 
2053 	if (*DestinationLinesToRequestVMInImmediateFlip >= 32 || *DestinationLinesToRequestRowInImmediateFlip >= 16 || TimeForFetchingMetaPTEImmediateFlip + 2 * TimeForFetchingRowInVBlankImmediateFlip > min_row_time) {
2054 		*ImmediateFlipSupportedForPipe = false;
2055 	} else {
2056 		*ImmediateFlipSupportedForPipe = true;
2057 	}
2058 
2059 #ifdef __DML_VBA_DEBUG__
2060 	dml_print("DML::%s: GPUVMEnable = %u\n", __func__, GPUVMEnable);
2061 	dml_print("DML::%s: DCCEnable = %u\n", __func__, DCCEnable);
2062 
2063 	dml_print("DML::%s: MetaRowBytes = %f\n", __func__, MetaRowBytes);
2064 	dml_print("DML::%s: DPTEBytesPerRow = %f\n", __func__, DPTEBytesPerRow);
2065 	dml_print("DML::%s: BandwidthAvailableForImmediateFlip = %f\n", __func__, BandwidthAvailableForImmediateFlip);
2066 	dml_print("DML::%s: TotImmediateFlipBytes = %u\n", __func__, TotImmediateFlipBytes);
2067 	dml_print("DML::%s: ImmediateFlipBW = %f\n", __func__, ImmediateFlipBW);
2068 	dml_print("DML::%s: PDEAndMetaPTEBytesPerFrame = %f\n", __func__, PDEAndMetaPTEBytesPerFrame);
2069 	dml_print("DML::%s: HostVMInefficiencyFactor = %f\n", __func__, HostVMInefficiencyFactor);
2070 	dml_print("DML::%s: LineTime = %f\n", __func__, LineTime);
2071 	dml_print("DML::%s: final_flip_bw = %f\n", __func__, *final_flip_bw);
2072 
2073 	dml_print("DML::%s: DestinationLinesToRequestVMInImmediateFlip = %f\n", __func__, *DestinationLinesToRequestVMInImmediateFlip);
2074 	dml_print("DML::%s: DestinationLinesToRequestRowInImmediateFlip = %f\n", __func__, *DestinationLinesToRequestRowInImmediateFlip);
2075 	dml_print("DML::%s: TimeForFetchingMetaPTEImmediateFlip = %f\n", __func__, TimeForFetchingMetaPTEImmediateFlip);
2076 	dml_print("DML::%s: TimeForFetchingRowInVBlankImmediateFlip = %f\n", __func__, TimeForFetchingRowInVBlankImmediateFlip);
2077 	dml_print("DML::%s: min_row_time = %f\n", __func__, min_row_time);
2078 	dml_print("DML::%s: ImmediateFlipSupportedForPipe = %u\n", __func__, *ImmediateFlipSupportedForPipe);
2079 #endif
2080 } // CalculateFlipSchedule
2081 
RoundToDFSGranularity(dml_float_t Clock,dml_bool_t round_up,dml_float_t VCOSpeed)2082 static dml_float_t RoundToDFSGranularity(dml_float_t Clock, dml_bool_t round_up, dml_float_t VCOSpeed)
2083 {
2084 	if (Clock <= 0.0)
2085 		return 0.0;
2086 	else {
2087 		if (round_up)
2088 			return VCOSpeed * 4.0 / dml_floor(VCOSpeed * 4.0 / Clock, 1.0);
2089 		else
2090 			return VCOSpeed * 4.0 / dml_ceil(VCOSpeed * 4.0 / Clock, 1.0);
2091 	}
2092 }
2093 
CalculateDCCConfiguration(dml_bool_t DCCEnabled,dml_bool_t DCCProgrammingAssumesScanDirectionUnknown,enum dml_source_format_class SourcePixelFormat,dml_uint_t SurfaceWidthLuma,dml_uint_t SurfaceWidthChroma,dml_uint_t SurfaceHeightLuma,dml_uint_t SurfaceHeightChroma,dml_uint_t nomDETInKByte,dml_uint_t RequestHeight256ByteLuma,dml_uint_t RequestHeight256ByteChroma,enum dml_swizzle_mode TilingFormat,dml_uint_t BytePerPixelY,dml_uint_t BytePerPixelC,dml_float_t BytePerPixelDETY,dml_float_t BytePerPixelDETC,enum dml_rotation_angle SourceScan,dml_uint_t * MaxUncompressedBlockLuma,dml_uint_t * MaxUncompressedBlockChroma,dml_uint_t * MaxCompressedBlockLuma,dml_uint_t * MaxCompressedBlockChroma,dml_uint_t * IndependentBlockLuma,dml_uint_t * IndependentBlockChroma)2094 static void CalculateDCCConfiguration(
2095 		dml_bool_t DCCEnabled,
2096 		dml_bool_t DCCProgrammingAssumesScanDirectionUnknown,
2097 		enum dml_source_format_class SourcePixelFormat,
2098 		dml_uint_t SurfaceWidthLuma,
2099 		dml_uint_t SurfaceWidthChroma,
2100 		dml_uint_t SurfaceHeightLuma,
2101 		dml_uint_t SurfaceHeightChroma,
2102 		dml_uint_t nomDETInKByte,
2103 		dml_uint_t RequestHeight256ByteLuma,
2104 		dml_uint_t RequestHeight256ByteChroma,
2105 		enum dml_swizzle_mode TilingFormat,
2106 		dml_uint_t BytePerPixelY,
2107 		dml_uint_t BytePerPixelC,
2108 		dml_float_t BytePerPixelDETY,
2109 		dml_float_t BytePerPixelDETC,
2110 		enum dml_rotation_angle SourceScan,
2111 		// Output
2112 		dml_uint_t *MaxUncompressedBlockLuma,
2113 		dml_uint_t *MaxUncompressedBlockChroma,
2114 		dml_uint_t *MaxCompressedBlockLuma,
2115 		dml_uint_t *MaxCompressedBlockChroma,
2116 		dml_uint_t *IndependentBlockLuma,
2117 		dml_uint_t *IndependentBlockChroma)
2118 {
2119 	dml_uint_t DETBufferSizeForDCC = nomDETInKByte * 1024;
2120 
2121 	dml_uint_t yuv420;
2122 	dml_uint_t horz_div_l;
2123 	dml_uint_t horz_div_c;
2124 	dml_uint_t vert_div_l;
2125 	dml_uint_t vert_div_c;
2126 
2127 	dml_uint_t swath_buf_size;
2128 	dml_float_t detile_buf_vp_horz_limit;
2129 	dml_float_t detile_buf_vp_vert_limit;
2130 
2131 	dml_uint_t MAS_vp_horz_limit;
2132 	dml_uint_t MAS_vp_vert_limit;
2133 	dml_uint_t max_vp_horz_width;
2134 	dml_uint_t max_vp_vert_height;
2135 	dml_uint_t eff_surf_width_l;
2136 	dml_uint_t eff_surf_width_c;
2137 	dml_uint_t eff_surf_height_l;
2138 	dml_uint_t eff_surf_height_c;
2139 
2140 	dml_uint_t full_swath_bytes_horz_wc_l;
2141 	dml_uint_t full_swath_bytes_horz_wc_c;
2142 	dml_uint_t full_swath_bytes_vert_wc_l;
2143 	dml_uint_t full_swath_bytes_vert_wc_c;
2144 
2145 	dml_uint_t req128_horz_wc_l;
2146 	dml_uint_t req128_horz_wc_c;
2147 	dml_uint_t req128_vert_wc_l;
2148 	dml_uint_t req128_vert_wc_c;
2149 
2150 	dml_uint_t   segment_order_horz_contiguous_luma;
2151 	dml_uint_t   segment_order_horz_contiguous_chroma;
2152 	dml_uint_t   segment_order_vert_contiguous_luma;
2153 	dml_uint_t   segment_order_vert_contiguous_chroma;
2154 
2155 	typedef enum{
2156 		REQ_256Bytes,
2157 		REQ_128BytesNonContiguous,
2158 		REQ_128BytesContiguous,
2159 		REQ_NA
2160 	} RequestType;
2161 
2162 	RequestType   RequestLuma;
2163 	RequestType   RequestChroma;
2164 
2165 	yuv420 = ((SourcePixelFormat == dml_420_8 || SourcePixelFormat == dml_420_10 || SourcePixelFormat == dml_420_12) ? 1 : 0);
2166 	horz_div_l = 1;
2167 	horz_div_c = 1;
2168 	vert_div_l = 1;
2169 	vert_div_c = 1;
2170 
2171 	if (BytePerPixelY == 1)
2172 		vert_div_l = 0;
2173 	if (BytePerPixelC == 1)
2174 		vert_div_c = 0;
2175 
2176 	if (BytePerPixelC == 0) {
2177 		swath_buf_size = DETBufferSizeForDCC / 2 - 2 * 256;
2178 		detile_buf_vp_horz_limit = (dml_float_t) swath_buf_size / ((dml_float_t) RequestHeight256ByteLuma * BytePerPixelY / (1 + horz_div_l));
2179 		detile_buf_vp_vert_limit = (dml_float_t) swath_buf_size / (256.0 / RequestHeight256ByteLuma / (1 + vert_div_l));
2180 	} else {
2181 		swath_buf_size = DETBufferSizeForDCC / 2 - 2 * 2 * 256;
2182 		detile_buf_vp_horz_limit = (dml_float_t) swath_buf_size / ((dml_float_t) RequestHeight256ByteLuma * BytePerPixelY / (1 + horz_div_l) + (dml_float_t) RequestHeight256ByteChroma * BytePerPixelC / (1 + horz_div_c) / (1 + yuv420));
2183 		detile_buf_vp_vert_limit = (dml_float_t) swath_buf_size / (256.0 / RequestHeight256ByteLuma / (1 + vert_div_l) + 256.0 / RequestHeight256ByteChroma / (1 + vert_div_c) / (1 + yuv420));
2184 	}
2185 
2186 	if (SourcePixelFormat == dml_420_10) {
2187 		detile_buf_vp_horz_limit = 1.5 * detile_buf_vp_horz_limit;
2188 		detile_buf_vp_vert_limit = 1.5 * detile_buf_vp_vert_limit;
2189 	}
2190 
2191 	detile_buf_vp_horz_limit = dml_floor(detile_buf_vp_horz_limit - 1, 16);
2192 	detile_buf_vp_vert_limit = dml_floor(detile_buf_vp_vert_limit - 1, 16);
2193 
2194 	MAS_vp_horz_limit = SourcePixelFormat == dml_rgbe_alpha ? 3840 : 6144;
2195 	MAS_vp_vert_limit = SourcePixelFormat == dml_rgbe_alpha ? 3840 : (BytePerPixelY == 8 ? 3072 : 6144);
2196 	max_vp_horz_width = (dml_uint_t)(dml_min((dml_float_t) MAS_vp_horz_limit, detile_buf_vp_horz_limit));
2197 	max_vp_vert_height = (dml_uint_t)(dml_min((dml_float_t) MAS_vp_vert_limit, detile_buf_vp_vert_limit));
2198 	eff_surf_width_l = (SurfaceWidthLuma > max_vp_horz_width ? max_vp_horz_width : SurfaceWidthLuma);
2199 	eff_surf_width_c = eff_surf_width_l / (1 + yuv420);
2200 	eff_surf_height_l = (SurfaceHeightLuma > max_vp_vert_height ? max_vp_vert_height : SurfaceHeightLuma);
2201 	eff_surf_height_c = eff_surf_height_l / (1 + yuv420);
2202 
2203 	full_swath_bytes_horz_wc_l = eff_surf_width_l * RequestHeight256ByteLuma * BytePerPixelY;
2204 	full_swath_bytes_vert_wc_l = eff_surf_height_l * 256 / RequestHeight256ByteLuma;
2205 	if (BytePerPixelC > 0) {
2206 		full_swath_bytes_horz_wc_c = eff_surf_width_c * RequestHeight256ByteChroma * BytePerPixelC;
2207 		full_swath_bytes_vert_wc_c = eff_surf_height_c * 256 / RequestHeight256ByteChroma;
2208 	} else {
2209 		full_swath_bytes_horz_wc_c = 0;
2210 		full_swath_bytes_vert_wc_c = 0;
2211 	}
2212 
2213 	if (SourcePixelFormat == dml_420_10) {
2214 		full_swath_bytes_horz_wc_l = (dml_uint_t)(dml_ceil((dml_float_t) full_swath_bytes_horz_wc_l * 2.0 / 3.0, 256.0));
2215 		full_swath_bytes_horz_wc_c = (dml_uint_t)(dml_ceil((dml_float_t) full_swath_bytes_horz_wc_c * 2.0 / 3.0, 256.0));
2216 		full_swath_bytes_vert_wc_l = (dml_uint_t)(dml_ceil((dml_float_t) full_swath_bytes_vert_wc_l * 2.0 / 3.0, 256.0));
2217 		full_swath_bytes_vert_wc_c = (dml_uint_t)(dml_ceil((dml_float_t) full_swath_bytes_vert_wc_c * 2.0 / 3.0, 256.0));
2218 	}
2219 
2220 	if (2 * full_swath_bytes_horz_wc_l + 2 * full_swath_bytes_horz_wc_c <= DETBufferSizeForDCC) {
2221 		req128_horz_wc_l = 0;
2222 		req128_horz_wc_c = 0;
2223 	} else if (full_swath_bytes_horz_wc_l < 1.5 * full_swath_bytes_horz_wc_c && 2 * full_swath_bytes_horz_wc_l + full_swath_bytes_horz_wc_c <= DETBufferSizeForDCC) {
2224 		req128_horz_wc_l = 0;
2225 		req128_horz_wc_c = 1;
2226 	} else if (full_swath_bytes_horz_wc_l >= 1.5 * full_swath_bytes_horz_wc_c && full_swath_bytes_horz_wc_l + 2 * full_swath_bytes_horz_wc_c <= DETBufferSizeForDCC) {
2227 		req128_horz_wc_l = 1;
2228 		req128_horz_wc_c = 0;
2229 	} else {
2230 		req128_horz_wc_l = 1;
2231 		req128_horz_wc_c = 1;
2232 	}
2233 
2234 	if (2 * full_swath_bytes_vert_wc_l + 2 * full_swath_bytes_vert_wc_c <= DETBufferSizeForDCC) {
2235 		req128_vert_wc_l = 0;
2236 		req128_vert_wc_c = 0;
2237 	} else if (full_swath_bytes_vert_wc_l < 1.5 * full_swath_bytes_vert_wc_c && 2 * full_swath_bytes_vert_wc_l + full_swath_bytes_vert_wc_c <= DETBufferSizeForDCC) {
2238 		req128_vert_wc_l = 0;
2239 		req128_vert_wc_c = 1;
2240 	} else if (full_swath_bytes_vert_wc_l >= 1.5 * full_swath_bytes_vert_wc_c && full_swath_bytes_vert_wc_l + 2 * full_swath_bytes_vert_wc_c <= DETBufferSizeForDCC) {
2241 		req128_vert_wc_l = 1;
2242 		req128_vert_wc_c = 0;
2243 	} else {
2244 		req128_vert_wc_l = 1;
2245 		req128_vert_wc_c = 1;
2246 	}
2247 
2248 	if (BytePerPixelY == 2) {
2249 		segment_order_horz_contiguous_luma = 0;
2250 		segment_order_vert_contiguous_luma = 1;
2251 	} else {
2252 		segment_order_horz_contiguous_luma = 1;
2253 		segment_order_vert_contiguous_luma = 0;
2254 	}
2255 
2256 	if (BytePerPixelC == 2) {
2257 		segment_order_horz_contiguous_chroma = 0;
2258 		segment_order_vert_contiguous_chroma = 1;
2259 	} else {
2260 		segment_order_horz_contiguous_chroma = 1;
2261 		segment_order_vert_contiguous_chroma = 0;
2262 	}
2263 #ifdef __DML_VBA_DEBUG__
2264 	dml_print("DML::%s: DCCEnabled = %u\n", __func__, DCCEnabled);
2265 	dml_print("DML::%s: nomDETInKByte = %u\n", __func__, nomDETInKByte);
2266 	dml_print("DML::%s: DETBufferSizeForDCC = %u\n", __func__, DETBufferSizeForDCC);
2267 	dml_print("DML::%s: req128_horz_wc_l = %u\n", __func__, req128_horz_wc_l);
2268 	dml_print("DML::%s: req128_horz_wc_c = %u\n", __func__, req128_horz_wc_c);
2269 	dml_print("DML::%s: full_swath_bytes_horz_wc_l = %u\n", __func__, full_swath_bytes_horz_wc_l);
2270 	dml_print("DML::%s: full_swath_bytes_vert_wc_c = %u\n", __func__, full_swath_bytes_vert_wc_c);
2271 	dml_print("DML::%s: segment_order_horz_contiguous_luma = %u\n", __func__, segment_order_horz_contiguous_luma);
2272 	dml_print("DML::%s: segment_order_horz_contiguous_chroma = %u\n", __func__, segment_order_horz_contiguous_chroma);
2273 #endif
2274 
2275 	if (DCCProgrammingAssumesScanDirectionUnknown == true) {
2276 		if (req128_horz_wc_l == 0 && req128_vert_wc_l == 0) {
2277 			RequestLuma = REQ_256Bytes;
2278 		} else if ((req128_horz_wc_l == 1 && segment_order_horz_contiguous_luma == 0) || (req128_vert_wc_l == 1 && segment_order_vert_contiguous_luma == 0)) {
2279 			RequestLuma = REQ_128BytesNonContiguous;
2280 		} else {
2281 			RequestLuma = REQ_128BytesContiguous;
2282 		}
2283 		if (req128_horz_wc_c == 0 && req128_vert_wc_c == 0) {
2284 			RequestChroma = REQ_256Bytes;
2285 		} else if ((req128_horz_wc_c == 1 && segment_order_horz_contiguous_chroma == 0) || (req128_vert_wc_c == 1 && segment_order_vert_contiguous_chroma == 0)) {
2286 			RequestChroma = REQ_128BytesNonContiguous;
2287 		} else {
2288 			RequestChroma = REQ_128BytesContiguous;
2289 		}
2290 	} else if (!dml_is_vertical_rotation(SourceScan)) {
2291 		if (req128_horz_wc_l == 0) {
2292 			RequestLuma = REQ_256Bytes;
2293 		} else if (segment_order_horz_contiguous_luma == 0) {
2294 			RequestLuma = REQ_128BytesNonContiguous;
2295 		} else {
2296 			RequestLuma = REQ_128BytesContiguous;
2297 		}
2298 		if (req128_horz_wc_c == 0) {
2299 			RequestChroma = REQ_256Bytes;
2300 		} else if (segment_order_horz_contiguous_chroma == 0) {
2301 			RequestChroma = REQ_128BytesNonContiguous;
2302 		} else {
2303 			RequestChroma = REQ_128BytesContiguous;
2304 		}
2305 	} else {
2306 		if (req128_vert_wc_l == 0) {
2307 			RequestLuma = REQ_256Bytes;
2308 		} else if (segment_order_vert_contiguous_luma == 0) {
2309 			RequestLuma = REQ_128BytesNonContiguous;
2310 		} else {
2311 			RequestLuma = REQ_128BytesContiguous;
2312 		}
2313 		if (req128_vert_wc_c == 0) {
2314 			RequestChroma = REQ_256Bytes;
2315 		} else if (segment_order_vert_contiguous_chroma == 0) {
2316 			RequestChroma = REQ_128BytesNonContiguous;
2317 		} else {
2318 			RequestChroma = REQ_128BytesContiguous;
2319 		}
2320 	}
2321 
2322 	if (RequestLuma == REQ_256Bytes) {
2323 		*MaxUncompressedBlockLuma = 256;
2324 		*MaxCompressedBlockLuma = 256;
2325 		*IndependentBlockLuma = 0;
2326 	} else if (RequestLuma == REQ_128BytesContiguous) {
2327 		*MaxUncompressedBlockLuma = 256;
2328 		*MaxCompressedBlockLuma = 128;
2329 		*IndependentBlockLuma = 128;
2330 	} else {
2331 		*MaxUncompressedBlockLuma = 256;
2332 		*MaxCompressedBlockLuma = 64;
2333 		*IndependentBlockLuma = 64;
2334 	}
2335 
2336 	if (RequestChroma == REQ_256Bytes) {
2337 		*MaxUncompressedBlockChroma = 256;
2338 		*MaxCompressedBlockChroma = 256;
2339 		*IndependentBlockChroma = 0;
2340 	} else if (RequestChroma == REQ_128BytesContiguous) {
2341 		*MaxUncompressedBlockChroma = 256;
2342 		*MaxCompressedBlockChroma = 128;
2343 		*IndependentBlockChroma = 128;
2344 	} else {
2345 		*MaxUncompressedBlockChroma = 256;
2346 		*MaxCompressedBlockChroma = 64;
2347 		*IndependentBlockChroma = 64;
2348 	}
2349 
2350 	if (DCCEnabled != true || BytePerPixelC == 0) {
2351 		*MaxUncompressedBlockChroma = 0;
2352 		*MaxCompressedBlockChroma = 0;
2353 		*IndependentBlockChroma = 0;
2354 	}
2355 
2356 	if (DCCEnabled != true) {
2357 		*MaxUncompressedBlockLuma = 0;
2358 		*MaxCompressedBlockLuma = 0;
2359 		*IndependentBlockLuma = 0;
2360 	}
2361 
2362 #ifdef __DML_VBA_DEBUG__
2363 	dml_print("DML::%s: MaxUncompressedBlockLuma = %u\n", __func__, *MaxUncompressedBlockLuma);
2364 	dml_print("DML::%s: MaxCompressedBlockLuma = %u\n", __func__, *MaxCompressedBlockLuma);
2365 	dml_print("DML::%s: IndependentBlockLuma = %u\n", __func__, *IndependentBlockLuma);
2366 	dml_print("DML::%s: MaxUncompressedBlockChroma = %u\n", __func__, *MaxUncompressedBlockChroma);
2367 	dml_print("DML::%s: MaxCompressedBlockChroma = %u\n", __func__, *MaxCompressedBlockChroma);
2368 	dml_print("DML::%s: IndependentBlockChroma = %u\n", __func__, *IndependentBlockChroma);
2369 #endif
2370 
2371 } // CalculateDCCConfiguration
2372 
CalculatePrefetchSourceLines(dml_float_t VRatio,dml_uint_t VTaps,dml_bool_t Interlace,dml_bool_t ProgressiveToInterlaceUnitInOPP,dml_uint_t SwathHeight,enum dml_rotation_angle SourceScan,dml_bool_t ViewportStationary,dml_uint_t SwathWidth,dml_uint_t ViewportHeight,dml_uint_t ViewportXStart,dml_uint_t ViewportYStart,dml_uint_t * VInitPreFill,dml_uint_t * MaxNumSwath)2373 static dml_uint_t CalculatePrefetchSourceLines(
2374 		dml_float_t VRatio,
2375 		dml_uint_t VTaps,
2376 		dml_bool_t Interlace,
2377 		dml_bool_t ProgressiveToInterlaceUnitInOPP,
2378 		dml_uint_t SwathHeight,
2379 		enum dml_rotation_angle SourceScan,
2380 		dml_bool_t ViewportStationary,
2381 		dml_uint_t SwathWidth,
2382 		dml_uint_t ViewportHeight,
2383 		dml_uint_t ViewportXStart,
2384 		dml_uint_t ViewportYStart,
2385 
2386 		// Output
2387 		dml_uint_t *VInitPreFill,
2388 		dml_uint_t *MaxNumSwath)
2389 {
2390 
2391 	dml_uint_t vp_start_rot = 0;
2392 	dml_uint_t sw0_tmp = 0;
2393 	dml_uint_t MaxPartialSwath = 0;
2394 	dml_float_t numLines = 0;
2395 
2396 #ifdef __DML_VBA_DEBUG__
2397 	dml_print("DML::%s: VRatio = %f\n", __func__, VRatio);
2398 	dml_print("DML::%s: VTaps = %u\n", __func__, VTaps);
2399 	dml_print("DML::%s: ViewportXStart = %u\n", __func__, ViewportXStart);
2400 	dml_print("DML::%s: ViewportYStart = %u\n", __func__, ViewportYStart);
2401 	dml_print("DML::%s: ViewportStationary = %u\n", __func__, ViewportStationary);
2402 	dml_print("DML::%s: SwathHeight = %u\n", __func__, SwathHeight);
2403 #endif
2404 	if (ProgressiveToInterlaceUnitInOPP)
2405 		*VInitPreFill = (dml_uint_t)(dml_floor((VRatio + (dml_float_t) VTaps + 1) / 2.0, 1));
2406 	else
2407 		*VInitPreFill = (dml_uint_t)(dml_floor((VRatio + (dml_float_t) VTaps + 1 + Interlace * 0.5 * VRatio) / 2.0, 1));
2408 
2409 	if (ViewportStationary) {
2410 		if (SourceScan == dml_rotation_180 || SourceScan == dml_rotation_180m) {
2411 			vp_start_rot = SwathHeight - (((dml_uint_t) (ViewportYStart + ViewportHeight - 1) % SwathHeight) + 1);
2412 		} else if (SourceScan == dml_rotation_270 || SourceScan == dml_rotation_90m) {
2413 			vp_start_rot = ViewportXStart;
2414 		} else if (SourceScan == dml_rotation_90 || SourceScan == dml_rotation_270m) {
2415 			vp_start_rot = SwathHeight - (((dml_uint_t)(ViewportYStart + SwathWidth - 1) % SwathHeight) + 1);
2416 		} else {
2417 			vp_start_rot = ViewportYStart;
2418 		}
2419 		sw0_tmp = SwathHeight - (vp_start_rot % SwathHeight);
2420 		if (sw0_tmp < *VInitPreFill) {
2421 			*MaxNumSwath = (dml_uint_t)(dml_ceil((*VInitPreFill - sw0_tmp) / (dml_float_t) SwathHeight, 1) + 1);
2422 		} else {
2423 			*MaxNumSwath = 1;
2424 		}
2425 		MaxPartialSwath = (dml_uint_t)(dml_max(1, (dml_uint_t) (vp_start_rot + *VInitPreFill - 1) % SwathHeight));
2426 	} else {
2427 		*MaxNumSwath = (dml_uint_t)(dml_ceil((*VInitPreFill - 1.0) / (dml_float_t) SwathHeight, 1) + 1);
2428 		if (*VInitPreFill > 1) {
2429 			MaxPartialSwath = (dml_uint_t)(dml_max(1, (dml_uint_t) (*VInitPreFill - 2) % SwathHeight));
2430 		} else {
2431 			MaxPartialSwath = (dml_uint_t)(dml_max(1, (dml_uint_t) (*VInitPreFill + SwathHeight - 2) % SwathHeight));
2432 		}
2433 	}
2434 	numLines = *MaxNumSwath * SwathHeight + MaxPartialSwath;
2435 
2436 #ifdef __DML_VBA_DEBUG__
2437 	dml_print("DML::%s: vp_start_rot = %u\n", __func__, vp_start_rot);
2438 	dml_print("DML::%s: VInitPreFill = %u\n", __func__, *VInitPreFill);
2439 	dml_print("DML::%s: MaxPartialSwath = %u\n", __func__, MaxPartialSwath);
2440 	dml_print("DML::%s: MaxNumSwath = %u\n", __func__, *MaxNumSwath);
2441 	dml_print("DML::%s: Prefetch source lines = %3.2f\n", __func__, numLines);
2442 #endif
2443 	return (dml_uint_t)(numLines);
2444 
2445 } // CalculatePrefetchSourceLines
2446 
CalculateVMAndRowBytes(dml_bool_t ViewportStationary,dml_bool_t DCCEnable,dml_uint_t NumberOfDPPs,dml_uint_t BlockHeight256Bytes,dml_uint_t BlockWidth256Bytes,enum dml_source_format_class SourcePixelFormat,dml_uint_t SurfaceTiling,dml_uint_t BytePerPixel,enum dml_rotation_angle SourceScan,dml_uint_t SwathWidth,dml_uint_t ViewportHeight,dml_uint_t ViewportXStart,dml_uint_t ViewportYStart,dml_bool_t GPUVMEnable,dml_uint_t GPUVMMaxPageTableLevels,dml_uint_t GPUVMMinPageSizeKBytes,dml_uint_t PTEBufferSizeInRequests,dml_uint_t Pitch,dml_uint_t DCCMetaPitch,dml_uint_t MacroTileWidth,dml_uint_t MacroTileHeight,dml_uint_t * MetaRowByte,dml_uint_t * PixelPTEBytesPerRow,dml_uint_t * PixelPTEBytesPerRowStorage,dml_uint_t * dpte_row_width_ub,dml_uint_t * dpte_row_height,dml_uint_t * dpte_row_height_linear,dml_uint_t * PixelPTEBytesPerRow_one_row_per_frame,dml_uint_t * dpte_row_width_ub_one_row_per_frame,dml_uint_t * dpte_row_height_one_row_per_frame,dml_uint_t * MetaRequestWidth,dml_uint_t * MetaRequestHeight,dml_uint_t * meta_row_width,dml_uint_t * meta_row_height,dml_uint_t * PixelPTEReqWidth,dml_uint_t * PixelPTEReqHeight,dml_uint_t * PTERequestSize,dml_uint_t * DPDE0BytesFrame,dml_uint_t * MetaPTEBytesFrame)2447 static dml_uint_t CalculateVMAndRowBytes(
2448 		dml_bool_t ViewportStationary,
2449 		dml_bool_t DCCEnable,
2450 		dml_uint_t NumberOfDPPs,
2451 		dml_uint_t BlockHeight256Bytes,
2452 		dml_uint_t BlockWidth256Bytes,
2453 		enum dml_source_format_class SourcePixelFormat,
2454 		dml_uint_t SurfaceTiling,
2455 		dml_uint_t BytePerPixel,
2456 		enum dml_rotation_angle SourceScan,
2457 		dml_uint_t SwathWidth,
2458 		dml_uint_t ViewportHeight,
2459 		dml_uint_t ViewportXStart,
2460 		dml_uint_t ViewportYStart,
2461 		dml_bool_t GPUVMEnable,
2462 		dml_uint_t GPUVMMaxPageTableLevels,
2463 		dml_uint_t GPUVMMinPageSizeKBytes,
2464 		dml_uint_t PTEBufferSizeInRequests,
2465 		dml_uint_t Pitch,
2466 		dml_uint_t DCCMetaPitch,
2467 		dml_uint_t MacroTileWidth,
2468 		dml_uint_t MacroTileHeight,
2469 
2470 		// Output
2471 		dml_uint_t *MetaRowByte,
2472 		dml_uint_t *PixelPTEBytesPerRow, // for bandwidth calculation
2473 		dml_uint_t *PixelPTEBytesPerRowStorage, // for PTE buffer size check
2474 		dml_uint_t *dpte_row_width_ub,
2475 		dml_uint_t *dpte_row_height,
2476 		dml_uint_t *dpte_row_height_linear,
2477 		dml_uint_t *PixelPTEBytesPerRow_one_row_per_frame,
2478 		dml_uint_t *dpte_row_width_ub_one_row_per_frame,
2479 		dml_uint_t *dpte_row_height_one_row_per_frame,
2480 		dml_uint_t *MetaRequestWidth,
2481 		dml_uint_t *MetaRequestHeight,
2482 		dml_uint_t *meta_row_width,
2483 		dml_uint_t *meta_row_height,
2484 		dml_uint_t *PixelPTEReqWidth,
2485 		dml_uint_t *PixelPTEReqHeight,
2486 		dml_uint_t *PTERequestSize,
2487 		dml_uint_t *DPDE0BytesFrame,
2488 		dml_uint_t *MetaPTEBytesFrame)
2489 {
2490 	dml_uint_t MPDEBytesFrame;
2491 	dml_uint_t DCCMetaSurfaceBytes;
2492 	dml_uint_t ExtraDPDEBytesFrame;
2493 	dml_uint_t PDEAndMetaPTEBytesFrame;
2494 	dml_uint_t MacroTileSizeBytes;
2495 	dml_uint_t vp_height_meta_ub;
2496 	dml_uint_t vp_height_dpte_ub;
2497 
2498 	dml_uint_t PixelPTEReqWidth_linear = 0; // VBA_DELTA. VBA doesn't calculate this
2499 
2500 	*MetaRequestHeight = 8 * BlockHeight256Bytes;
2501 	*MetaRequestWidth = 8 * BlockWidth256Bytes;
2502 	if (SurfaceTiling == dml_sw_linear) {
2503 		*meta_row_height = 32;
2504 		*meta_row_width = (dml_uint_t)(dml_floor(ViewportXStart + SwathWidth + *MetaRequestWidth - 1, *MetaRequestWidth) - dml_floor(ViewportXStart, *MetaRequestWidth));
2505 	} else if (!dml_is_vertical_rotation(SourceScan)) {
2506 		*meta_row_height = *MetaRequestHeight;
2507 		if (ViewportStationary && NumberOfDPPs == 1) {
2508 			*meta_row_width = (dml_uint_t)(dml_floor(ViewportXStart + SwathWidth + *MetaRequestWidth - 1, *MetaRequestWidth) - dml_floor(ViewportXStart, *MetaRequestWidth));
2509 		} else {
2510 			*meta_row_width = (dml_uint_t)(dml_ceil(SwathWidth - 1, *MetaRequestWidth) + *MetaRequestWidth);
2511 		}
2512 		*MetaRowByte = (dml_uint_t)(*meta_row_width * *MetaRequestHeight * BytePerPixel / 256.0);
2513 	} else {
2514 		*meta_row_height = *MetaRequestWidth;
2515 		if (ViewportStationary && NumberOfDPPs == 1) {
2516 			*meta_row_width = (dml_uint_t)(dml_floor(ViewportYStart + ViewportHeight + *MetaRequestHeight - 1, *MetaRequestHeight) - dml_floor(ViewportYStart, *MetaRequestHeight));
2517 		} else {
2518 			*meta_row_width = (dml_uint_t)(dml_ceil(SwathWidth - 1, *MetaRequestHeight) + *MetaRequestHeight);
2519 		}
2520 		*MetaRowByte = (dml_uint_t)(*meta_row_width * *MetaRequestWidth * BytePerPixel / 256.0);
2521 	}
2522 
2523 	if (ViewportStationary && (NumberOfDPPs == 1 || !dml_is_vertical_rotation(SourceScan))) {
2524 		vp_height_meta_ub = (dml_uint_t)(dml_floor(ViewportYStart + ViewportHeight + 64 * BlockHeight256Bytes - 1, 64 * BlockHeight256Bytes) - dml_floor(ViewportYStart, 64 * BlockHeight256Bytes));
2525 	} else if (!dml_is_vertical_rotation(SourceScan)) {
2526 		vp_height_meta_ub = (dml_uint_t)(dml_ceil(ViewportHeight - 1, 64 * BlockHeight256Bytes) + 64 * BlockHeight256Bytes);
2527 	} else {
2528 		vp_height_meta_ub = (dml_uint_t)(dml_ceil(SwathWidth - 1, 64 * BlockHeight256Bytes) + 64 * BlockHeight256Bytes);
2529 	}
2530 
2531 	DCCMetaSurfaceBytes = (dml_uint_t)(DCCMetaPitch * vp_height_meta_ub * BytePerPixel / 256.0);
2532 
2533 	if (GPUVMEnable == true) {
2534 		*MetaPTEBytesFrame = (dml_uint_t)((dml_ceil((dml_float_t) (DCCMetaSurfaceBytes - 4.0 * 1024.0) / (8 * 4.0 * 1024), 1) + 1) * 64);
2535 		MPDEBytesFrame = 128 * (GPUVMMaxPageTableLevels - 1);
2536 	} else {
2537 		*MetaPTEBytesFrame = 0;
2538 		MPDEBytesFrame = 0;
2539 	}
2540 
2541 	if (DCCEnable != true) {
2542 		*MetaPTEBytesFrame = 0;
2543 		MPDEBytesFrame = 0;
2544 		*MetaRowByte = 0;
2545 	}
2546 
2547 	MacroTileSizeBytes = MacroTileWidth * BytePerPixel * MacroTileHeight;
2548 
2549 	if (ViewportStationary && (NumberOfDPPs == 1 || !dml_is_vertical_rotation(SourceScan))) {
2550 		vp_height_dpte_ub = (dml_uint_t)(dml_floor(ViewportYStart + ViewportHeight + MacroTileHeight - 1, MacroTileHeight) - dml_floor(ViewportYStart, MacroTileHeight));
2551 	} else if (!dml_is_vertical_rotation(SourceScan)) {
2552 		vp_height_dpte_ub = (dml_uint_t)(dml_ceil(ViewportHeight - 1, MacroTileHeight) + MacroTileHeight);
2553 	} else {
2554 		vp_height_dpte_ub = (dml_uint_t)(dml_ceil(SwathWidth - 1, MacroTileHeight) + MacroTileHeight);
2555 	}
2556 
2557 	if (GPUVMEnable == true && GPUVMMaxPageTableLevels > 1) {
2558 		*DPDE0BytesFrame = (dml_uint_t)(64 * (dml_ceil((dml_float_t) (Pitch * vp_height_dpte_ub * BytePerPixel - MacroTileSizeBytes) / (dml_float_t) (8 * 2097152), 1) + 1));
2559 		ExtraDPDEBytesFrame = 128 * (GPUVMMaxPageTableLevels - 2);
2560 	} else {
2561 		*DPDE0BytesFrame = 0;
2562 		ExtraDPDEBytesFrame = 0;
2563 	}
2564 
2565 	PDEAndMetaPTEBytesFrame = *MetaPTEBytesFrame + MPDEBytesFrame + *DPDE0BytesFrame + ExtraDPDEBytesFrame;
2566 
2567 #ifdef __DML_VBA_DEBUG__
2568 	dml_print("DML::%s: DCCEnable = %u\n", __func__, DCCEnable);
2569 	dml_print("DML::%s: GPUVMEnable = %u\n", __func__, GPUVMEnable);
2570 	dml_print("DML::%s: SwModeLinear = %u\n", __func__, SurfaceTiling == dml_sw_linear);
2571 	dml_print("DML::%s: BytePerPixel = %u\n", __func__, BytePerPixel);
2572 	dml_print("DML::%s: GPUVMMaxPageTableLevels = %u\n", __func__, GPUVMMaxPageTableLevels);
2573 	dml_print("DML::%s: BlockHeight256Bytes = %u\n", __func__, BlockHeight256Bytes);
2574 	dml_print("DML::%s: BlockWidth256Bytes = %u\n", __func__, BlockWidth256Bytes);
2575 	dml_print("DML::%s: MacroTileHeight = %u\n", __func__, MacroTileHeight);
2576 	dml_print("DML::%s: MacroTileWidth = %u\n", __func__, MacroTileWidth);
2577 	dml_print("DML::%s: MetaPTEBytesFrame = %u\n", __func__, *MetaPTEBytesFrame);
2578 	dml_print("DML::%s: MPDEBytesFrame = %u\n", __func__, MPDEBytesFrame);
2579 	dml_print("DML::%s: DPDE0BytesFrame = %u\n", __func__, *DPDE0BytesFrame);
2580 	dml_print("DML::%s: ExtraDPDEBytesFrame= %u\n", __func__, ExtraDPDEBytesFrame);
2581 	dml_print("DML::%s: PDEAndMetaPTEBytesFrame = %u\n", __func__, PDEAndMetaPTEBytesFrame);
2582 	dml_print("DML::%s: ViewportHeight = %u\n", __func__, ViewportHeight);
2583 	dml_print("DML::%s: SwathWidth = %u\n", __func__, SwathWidth);
2584 	dml_print("DML::%s: vp_height_dpte_ub = %u\n", __func__, vp_height_dpte_ub);
2585 #endif
2586 
2587 	if (SurfaceTiling == dml_sw_linear) {
2588 		*PixelPTEReqHeight = 1;
2589 		*PixelPTEReqWidth = GPUVMMinPageSizeKBytes * 1024 * 8 / BytePerPixel;
2590 		PixelPTEReqWidth_linear = GPUVMMinPageSizeKBytes * 1024 * 8 / BytePerPixel;
2591 		*PTERequestSize = 64;
2592 	} else if (GPUVMMinPageSizeKBytes == 4) {
2593 		*PixelPTEReqHeight = 16 * BlockHeight256Bytes;
2594 		*PixelPTEReqWidth = 16 * BlockWidth256Bytes;
2595 		*PTERequestSize = 128;
2596 	} else {
2597 		*PixelPTEReqHeight = MacroTileHeight;
2598 		*PixelPTEReqWidth = 8 *  1024 * GPUVMMinPageSizeKBytes / (MacroTileHeight * BytePerPixel);
2599 		*PTERequestSize = 64;
2600 	}
2601 #ifdef __DML_VBA_DEBUG__
2602 	dml_print("DML::%s: GPUVMMinPageSizeKBytes = %u\n", __func__, GPUVMMinPageSizeKBytes);
2603 	dml_print("DML::%s: PDEAndMetaPTEBytesFrame = %u (after HostVM factor)\n", __func__, PDEAndMetaPTEBytesFrame);
2604 	dml_print("DML::%s: PixelPTEReqHeight = %u\n", __func__, *PixelPTEReqHeight);
2605 	dml_print("DML::%s: PixelPTEReqWidth = %u\n", __func__, *PixelPTEReqWidth);
2606 	dml_print("DML::%s: PixelPTEReqWidth_linear = %u\n", __func__, PixelPTEReqWidth_linear);
2607 	dml_print("DML::%s: PTERequestSize = %u\n", __func__, *PTERequestSize);
2608 	dml_print("DML::%s: Pitch = %u\n", __func__, Pitch);
2609 #endif
2610 
2611 	*dpte_row_height_one_row_per_frame = vp_height_dpte_ub;
2612 	*dpte_row_width_ub_one_row_per_frame = (dml_uint_t)((dml_ceil(((dml_float_t)Pitch * (dml_float_t) *dpte_row_height_one_row_per_frame / (dml_float_t) *PixelPTEReqHeight - 1) / (dml_float_t) *PixelPTEReqWidth, 1) + 1) * (dml_float_t) *PixelPTEReqWidth);
2613 	*PixelPTEBytesPerRow_one_row_per_frame = (dml_uint_t)((dml_float_t) *dpte_row_width_ub_one_row_per_frame / (dml_float_t) *PixelPTEReqWidth * *PTERequestSize);
2614 
2615 	if (SurfaceTiling == dml_sw_linear) {
2616 		*dpte_row_height = (dml_uint_t)(dml_min(128, 1 << (dml_uint_t) dml_floor(dml_log2(PTEBufferSizeInRequests * *PixelPTEReqWidth / Pitch), 1)));
2617 		dml_print("DML::%s: dpte_row_height term 1 = %u\n", __func__, PTEBufferSizeInRequests * *PixelPTEReqWidth / Pitch);
2618 		dml_print("DML::%s: dpte_row_height term 2 = %f\n", __func__, dml_log2(PTEBufferSizeInRequests * *PixelPTEReqWidth / Pitch));
2619 		dml_print("DML::%s: dpte_row_height term 3 = %f\n", __func__, dml_floor(dml_log2(PTEBufferSizeInRequests * *PixelPTEReqWidth / Pitch), 1));
2620 		dml_print("DML::%s: dpte_row_height term 4 = %u\n", __func__, 1 << (dml_uint_t) dml_floor(dml_log2(PTEBufferSizeInRequests * *PixelPTEReqWidth / Pitch), 1));
2621 		dml_print("DML::%s: dpte_row_height = %u\n", __func__, *dpte_row_height);
2622 
2623 		*dpte_row_width_ub = (dml_uint_t)(dml_ceil(((dml_float_t) Pitch * (dml_float_t) *dpte_row_height - 1), (dml_float_t) *PixelPTEReqWidth) + *PixelPTEReqWidth);
2624 		*PixelPTEBytesPerRow = (dml_uint_t)((dml_float_t) *dpte_row_width_ub / (dml_float_t) *PixelPTEReqWidth * *PTERequestSize);
2625 
2626 		// VBA_DELTA, VBA doesn't have programming value for pte row height linear.
2627 		*dpte_row_height_linear = 1 << (dml_uint_t) dml_floor(dml_log2(PTEBufferSizeInRequests * PixelPTEReqWidth_linear / Pitch), 1);
2628 		if (*dpte_row_height_linear > 128)
2629 			*dpte_row_height_linear = 128;
2630 
2631 #ifdef __DML_VBA_DEBUG__
2632 		dml_print("DML::%s: dpte_row_width_ub = %u (linear)\n", __func__, *dpte_row_width_ub);
2633 #endif
2634 
2635 	} else if (!dml_is_vertical_rotation(SourceScan)) {
2636 		*dpte_row_height = *PixelPTEReqHeight;
2637 
2638 		if (GPUVMMinPageSizeKBytes > 64) {
2639 			*dpte_row_width_ub = (dml_uint_t)((dml_ceil(((dml_float_t) Pitch * (dml_float_t) *dpte_row_height / (dml_float_t) *PixelPTEReqHeight - 1) / (dml_float_t) *PixelPTEReqWidth, 1) + 1) * *PixelPTEReqWidth);
2640 		} else if (ViewportStationary && (NumberOfDPPs == 1)) {
2641 			*dpte_row_width_ub = (dml_uint_t)(dml_floor(ViewportXStart + SwathWidth + *PixelPTEReqWidth - 1, *PixelPTEReqWidth) - dml_floor(ViewportXStart, *PixelPTEReqWidth));
2642 		} else {
2643 			*dpte_row_width_ub = (dml_uint_t)((dml_ceil((dml_float_t) (SwathWidth - 1) / (dml_float_t)*PixelPTEReqWidth, 1) + 1.0) * *PixelPTEReqWidth);
2644 		}
2645 #ifdef __DML_VBA_DEBUG__
2646 		dml_print("DML::%s: dpte_row_width_ub = %u (tiled horz)\n", __func__, *dpte_row_width_ub);
2647 #endif
2648 
2649 		ASSERT(*PixelPTEReqWidth);
2650 		if (*PixelPTEReqWidth != 0)
2651 			*PixelPTEBytesPerRow = *dpte_row_width_ub / *PixelPTEReqWidth * *PTERequestSize;
2652 	} else {
2653 		*dpte_row_height = (dml_uint_t)(dml_min(*PixelPTEReqWidth, MacroTileWidth));
2654 
2655 		if (ViewportStationary && (NumberOfDPPs == 1)) {
2656 			*dpte_row_width_ub = (dml_uint_t)(dml_floor(ViewportYStart + ViewportHeight + *PixelPTEReqHeight - 1, *PixelPTEReqHeight) - dml_floor(ViewportYStart, *PixelPTEReqHeight));
2657 		} else {
2658 			*dpte_row_width_ub = (dml_uint_t)((dml_ceil((dml_float_t) (SwathWidth - 1) / (dml_float_t) *PixelPTEReqHeight, 1) + 1) * *PixelPTEReqHeight);
2659 		}
2660 
2661 		*PixelPTEBytesPerRow = (dml_uint_t)((dml_float_t) *dpte_row_width_ub / (dml_float_t) *PixelPTEReqHeight * *PTERequestSize);
2662 #ifdef __DML_VBA_DEBUG__
2663 		dml_print("DML::%s: dpte_row_width_ub = %u (tiled vert)\n", __func__, *dpte_row_width_ub);
2664 #endif
2665 	}
2666 
2667 	if (GPUVMEnable != true)
2668 		*PixelPTEBytesPerRow = 0;
2669 
2670 	*PixelPTEBytesPerRowStorage = *PixelPTEBytesPerRow;
2671 
2672 #ifdef __DML_VBA_DEBUG__
2673 	dml_print("DML::%s: GPUVMMinPageSizeKBytes = %u\n", __func__, GPUVMMinPageSizeKBytes);
2674 	dml_print("DML::%s: GPUVMEnable = %u\n", __func__, GPUVMEnable);
2675 	dml_print("DML::%s: dpte_row_height = %u\n", __func__, *dpte_row_height);
2676 	dml_print("DML::%s: dpte_row_height_linear = %u\n", __func__, *dpte_row_height_linear);
2677 	dml_print("DML::%s: dpte_row_width_ub = %u\n", __func__, *dpte_row_width_ub);
2678 	dml_print("DML::%s: PixelPTEBytesPerRow = %u\n", __func__, *PixelPTEBytesPerRow);
2679 	dml_print("DML::%s: PixelPTEBytesPerRowStorage = %u\n", __func__, *PixelPTEBytesPerRowStorage);
2680 	dml_print("DML::%s: PTEBufferSizeInRequests = %u\n", __func__, PTEBufferSizeInRequests);
2681 	dml_print("DML::%s: dpte_row_height_one_row_per_frame = %u\n", __func__, *dpte_row_height_one_row_per_frame);
2682 	dml_print("DML::%s: dpte_row_width_ub_one_row_per_frame = %u\n", __func__, *dpte_row_width_ub_one_row_per_frame);
2683 	dml_print("DML::%s: PixelPTEBytesPerRow_one_row_per_frame = %u\n", __func__, *PixelPTEBytesPerRow_one_row_per_frame);
2684 #endif
2685 
2686 	dml_print("DML: vm_bytes = meta_pte_bytes_per_frame (per_pipe) = MetaPTEBytesFrame = : %i\n", *MetaPTEBytesFrame);
2687 
2688 	return PDEAndMetaPTEBytesFrame;
2689 } // CalculateVMAndRowBytes
2690 
PixelClockAdjustmentForProgressiveToInterlaceUnit(struct dml_display_cfg_st * display_cfg,dml_bool_t ptoi_supported)2691 static void PixelClockAdjustmentForProgressiveToInterlaceUnit(struct dml_display_cfg_st *display_cfg, dml_bool_t ptoi_supported)
2692 {
2693 	dml_uint_t num_active_planes = dml_get_num_active_planes(display_cfg);
2694 
2695 	//Progressive To Interlace Unit Effect
2696 	for (dml_uint_t k = 0; k < num_active_planes; ++k) {
2697 		display_cfg->output.PixelClockBackEnd[k] = display_cfg->timing.PixelClock[k];
2698 		if (display_cfg->timing.Interlace[k] == 1 && ptoi_supported == true) {
2699 			display_cfg->timing.PixelClock[k] = 2 * display_cfg->timing.PixelClock[k];
2700 		}
2701 	}
2702 }
2703 
TruncToValidBPP(dml_float_t LinkBitRate,dml_uint_t Lanes,dml_uint_t HTotal,dml_uint_t HActive,dml_float_t PixelClock,dml_float_t DesiredBPP,dml_bool_t DSCEnable,enum dml_output_encoder_class Output,enum dml_output_format_class Format,dml_uint_t DSCInputBitPerComponent,dml_uint_t DSCSlices,dml_uint_t AudioRate,dml_uint_t AudioLayout,enum dml_odm_mode ODMModeNoDSC,enum dml_odm_mode ODMModeDSC,dml_uint_t * RequiredSlots)2704 static dml_float_t TruncToValidBPP(
2705 		dml_float_t LinkBitRate,
2706 		dml_uint_t Lanes,
2707 		dml_uint_t HTotal,
2708 		dml_uint_t HActive,
2709 		dml_float_t PixelClock,
2710 		dml_float_t DesiredBPP,
2711 		dml_bool_t DSCEnable,
2712 		enum dml_output_encoder_class Output,
2713 		enum dml_output_format_class Format,
2714 		dml_uint_t DSCInputBitPerComponent,
2715 		dml_uint_t DSCSlices,
2716 		dml_uint_t AudioRate,
2717 		dml_uint_t AudioLayout,
2718 		enum dml_odm_mode ODMModeNoDSC,
2719 		enum dml_odm_mode ODMModeDSC,
2720 
2721 		// Output
2722 		dml_uint_t *RequiredSlots)
2723 {
2724 	dml_float_t MaxLinkBPP;
2725 	dml_uint_t MinDSCBPP;
2726 	dml_float_t MaxDSCBPP;
2727 	dml_uint_t NonDSCBPP0;
2728 	dml_uint_t NonDSCBPP1;
2729 	dml_uint_t NonDSCBPP2;
2730 
2731 	if (Format == dml_420) {
2732 		NonDSCBPP0 = 12;
2733 		NonDSCBPP1 = 15;
2734 		NonDSCBPP2 = 18;
2735 		MinDSCBPP = 6;
2736 		MaxDSCBPP = 1.5 * DSCInputBitPerComponent - 1.0 / 16;
2737 	} else if (Format == dml_444) {
2738 		NonDSCBPP0 = 24;
2739 		NonDSCBPP1 = 30;
2740 		NonDSCBPP2 = 36;
2741 		MinDSCBPP = 8;
2742 		MaxDSCBPP = 3 * DSCInputBitPerComponent - 1.0 / 16;
2743 	} else {
2744 		if (Output == dml_hdmi) {
2745 			NonDSCBPP0 = 24;
2746 			NonDSCBPP1 = 24;
2747 			NonDSCBPP2 = 24;
2748 		} else {
2749 			NonDSCBPP0 = 16;
2750 			NonDSCBPP1 = 20;
2751 			NonDSCBPP2 = 24;
2752 	}
2753 	if (Format == dml_n422) {
2754 		MinDSCBPP = 7;
2755 			MaxDSCBPP = 2 * DSCInputBitPerComponent - 1.0 / 16.0;
2756 		} else {
2757 			MinDSCBPP = 8;
2758 			MaxDSCBPP = 3 * DSCInputBitPerComponent - 1.0 / 16.0;
2759 		}
2760 	}
2761 
2762 	if (Output == dml_dp2p0) {
2763 		MaxLinkBPP = LinkBitRate * Lanes / PixelClock * 128.0 / 132.0 * 383.0 / 384.0 * 65536.0 / 65540.0;
2764 	} else if (DSCEnable && Output == dml_dp) {
2765 		MaxLinkBPP = LinkBitRate / 10.0 * 8.0 * Lanes / PixelClock * (1 - 2.4 / 100);
2766 	} else {
2767 		MaxLinkBPP = LinkBitRate / 10.0 * 8.0 * Lanes / PixelClock;
2768 	}
2769 
2770 	if (DSCEnable) {
2771 		if (ODMModeDSC == dml_odm_mode_combine_4to1) {
2772 			MaxLinkBPP = dml_min(MaxLinkBPP, 16);
2773 		} else if (ODMModeDSC == dml_odm_mode_combine_2to1) {
2774 			MaxLinkBPP = dml_min(MaxLinkBPP, 32);
2775 		} else if (ODMModeDSC == dml_odm_mode_split_1to2) {
2776 			MaxLinkBPP = 2 * MaxLinkBPP;
2777 		}
2778 	} else {
2779 		if (ODMModeNoDSC == dml_odm_mode_combine_4to1) {
2780 			MaxLinkBPP = dml_min(MaxLinkBPP, 16);
2781 		} else if (ODMModeNoDSC == dml_odm_mode_combine_2to1) {
2782 			MaxLinkBPP = dml_min(MaxLinkBPP, 32);
2783 		} else if (ODMModeNoDSC == dml_odm_mode_split_1to2) {
2784 			MaxLinkBPP = 2 * MaxLinkBPP;
2785 		}
2786 	}
2787 
2788 	*RequiredSlots = (dml_uint_t)(dml_ceil(DesiredBPP / MaxLinkBPP * 64, 1));
2789 
2790 	if (DesiredBPP == 0) {
2791 		if (DSCEnable) {
2792 			if (MaxLinkBPP < MinDSCBPP) {
2793 				return __DML_DPP_INVALID__;
2794 			} else if (MaxLinkBPP >= MaxDSCBPP) {
2795 				return MaxDSCBPP;
2796 			} else {
2797 				return dml_floor(16.0 * MaxLinkBPP, 1.0) / 16.0;
2798 			}
2799 		} else {
2800 			if (MaxLinkBPP >= NonDSCBPP2) {
2801 				return NonDSCBPP2;
2802 			} else if (MaxLinkBPP >= NonDSCBPP1) {
2803 				return NonDSCBPP1;
2804 			} else if (MaxLinkBPP >= NonDSCBPP0) {
2805 				return NonDSCBPP0;
2806 			} else {
2807 				return __DML_DPP_INVALID__;
2808 			}
2809 		}
2810 	} else {
2811 		if (!((DSCEnable == false && (DesiredBPP == NonDSCBPP2 || DesiredBPP == NonDSCBPP1 || DesiredBPP == NonDSCBPP0)) ||
2812 				(DSCEnable && DesiredBPP >= MinDSCBPP && DesiredBPP <= MaxDSCBPP))) {
2813 			return __DML_DPP_INVALID__;
2814 		} else {
2815 			return DesiredBPP;
2816 		}
2817 	}
2818 } // TruncToValidBPP
2819 
CalculateWatermarksMALLUseAndDRAMSpeedChangeSupport(struct display_mode_lib_scratch_st * scratch,struct CalculateWatermarksMALLUseAndDRAMSpeedChangeSupport_params_st * p)2820 static void CalculateWatermarksMALLUseAndDRAMSpeedChangeSupport(
2821 		struct display_mode_lib_scratch_st *scratch,
2822 		struct CalculateWatermarksMALLUseAndDRAMSpeedChangeSupport_params_st *p)
2823 {
2824 	struct CalculateWatermarksMALLUseAndDRAMSpeedChangeSupport_locals_st *s = &scratch->CalculateWatermarksMALLUseAndDRAMSpeedChangeSupport_locals;
2825 
2826 	s->TotalActiveWriteback = 0;
2827 	p->Watermark->UrgentWatermark = p->mmSOCParameters.UrgentLatency + p->mmSOCParameters.ExtraLatency;
2828 	p->Watermark->USRRetrainingWatermark = p->mmSOCParameters.UrgentLatency + p->mmSOCParameters.ExtraLatency + p->mmSOCParameters.USRRetrainingLatency + p->mmSOCParameters.SMNLatency;
2829 	p->Watermark->DRAMClockChangeWatermark = p->mmSOCParameters.DRAMClockChangeLatency + p->Watermark->UrgentWatermark;
2830 	p->Watermark->FCLKChangeWatermark = p->mmSOCParameters.FCLKChangeLatency + p->Watermark->UrgentWatermark;
2831 	p->Watermark->StutterExitWatermark = p->mmSOCParameters.SRExitTime + p->mmSOCParameters.ExtraLatency + 10 / p->DCFClkDeepSleep;
2832 	p->Watermark->StutterEnterPlusExitWatermark = p->mmSOCParameters.SREnterPlusExitTime + p->mmSOCParameters.ExtraLatency + 10 / p->DCFClkDeepSleep;
2833 	p->Watermark->Z8StutterExitWatermark = p->mmSOCParameters.SRExitZ8Time + p->mmSOCParameters.ExtraLatency + 10 / p->DCFClkDeepSleep;
2834 	p->Watermark->Z8StutterEnterPlusExitWatermark = p->mmSOCParameters.SREnterPlusExitZ8Time + p->mmSOCParameters.ExtraLatency + 10 / p->DCFClkDeepSleep;
2835 
2836 #ifdef __DML_VBA_DEBUG__
2837 	dml_print("DML::%s: UrgentLatency = %f\n", __func__, p->mmSOCParameters.UrgentLatency);
2838 	dml_print("DML::%s: ExtraLatency = %f\n", __func__, p->mmSOCParameters.ExtraLatency);
2839 	dml_print("DML::%s: DRAMClockChangeLatency = %f\n", __func__, p->mmSOCParameters.DRAMClockChangeLatency);
2840 	dml_print("DML::%s: UrgentWatermark = %f\n", __func__, p->Watermark->UrgentWatermark);
2841 	dml_print("DML::%s: USRRetrainingWatermark = %f\n", __func__, p->Watermark->USRRetrainingWatermark);
2842 	dml_print("DML::%s: DRAMClockChangeWatermark = %f\n", __func__, p->Watermark->DRAMClockChangeWatermark);
2843 	dml_print("DML::%s: FCLKChangeWatermark = %f\n", __func__, p->Watermark->FCLKChangeWatermark);
2844 	dml_print("DML::%s: StutterExitWatermark = %f\n", __func__, p->Watermark->StutterExitWatermark);
2845 	dml_print("DML::%s: StutterEnterPlusExitWatermark = %f\n", __func__, p->Watermark->StutterEnterPlusExitWatermark);
2846 	dml_print("DML::%s: Z8StutterExitWatermark = %f\n", __func__, p->Watermark->Z8StutterExitWatermark);
2847 	dml_print("DML::%s: Z8StutterEnterPlusExitWatermark = %f\n", __func__, p->Watermark->Z8StutterEnterPlusExitWatermark);
2848 #endif
2849 
2850 	s->TotalActiveWriteback = 0;
2851 	for (dml_uint_t k = 0; k < p->NumberOfActiveSurfaces; ++k) {
2852 		if (p->WritebackEnable[k] == true) {
2853 			s->TotalActiveWriteback = s->TotalActiveWriteback + 1;
2854 		}
2855 	}
2856 
2857 	if (s->TotalActiveWriteback <= 1) {
2858 		p->Watermark->WritebackUrgentWatermark = p->mmSOCParameters.WritebackLatency;
2859 	} else {
2860 		p->Watermark->WritebackUrgentWatermark = p->mmSOCParameters.WritebackLatency + p->WritebackChunkSize * 1024.0 / 32.0 / p->SOCCLK;
2861 	}
2862 	if (p->USRRetrainingRequiredFinal)
2863 		p->Watermark->WritebackUrgentWatermark = p->Watermark->WritebackUrgentWatermark + p->mmSOCParameters.USRRetrainingLatency;
2864 
2865 	if (s->TotalActiveWriteback <= 1) {
2866 		p->Watermark->WritebackDRAMClockChangeWatermark = p->mmSOCParameters.DRAMClockChangeLatency + p->mmSOCParameters.WritebackLatency;
2867 		p->Watermark->WritebackFCLKChangeWatermark = p->mmSOCParameters.FCLKChangeLatency + p->mmSOCParameters.WritebackLatency;
2868 	} else {
2869 		p->Watermark->WritebackDRAMClockChangeWatermark = p->mmSOCParameters.DRAMClockChangeLatency + p->mmSOCParameters.WritebackLatency + p->WritebackChunkSize * 1024.0 / 32.0 / p->SOCCLK;
2870 		p->Watermark->WritebackFCLKChangeWatermark = p->mmSOCParameters.FCLKChangeLatency + p->mmSOCParameters.WritebackLatency + p->WritebackChunkSize * 1024 / 32 / p->SOCCLK;
2871 	}
2872 
2873 	if (p->USRRetrainingRequiredFinal)
2874 		p->Watermark->WritebackDRAMClockChangeWatermark = p->Watermark->WritebackDRAMClockChangeWatermark + p->mmSOCParameters.USRRetrainingLatency;
2875 
2876 	if (p->USRRetrainingRequiredFinal)
2877 		p->Watermark->WritebackFCLKChangeWatermark = p->Watermark->WritebackFCLKChangeWatermark + p->mmSOCParameters.USRRetrainingLatency;
2878 
2879 #ifdef __DML_VBA_DEBUG__
2880 	dml_print("DML::%s: WritebackDRAMClockChangeWatermark = %f\n", __func__, p->Watermark->WritebackDRAMClockChangeWatermark);
2881 	dml_print("DML::%s: WritebackFCLKChangeWatermark = %f\n", __func__, p->Watermark->WritebackFCLKChangeWatermark);
2882 	dml_print("DML::%s: WritebackUrgentWatermark = %f\n", __func__, p->Watermark->WritebackUrgentWatermark);
2883 	dml_print("DML::%s: USRRetrainingRequiredFinal = %u\n", __func__, p->USRRetrainingRequiredFinal);
2884 	dml_print("DML::%s: USRRetrainingLatency = %f\n", __func__, p->mmSOCParameters.USRRetrainingLatency);
2885 #endif
2886 
2887 	s->TotalPixelBW = 0.0;
2888 	for (dml_uint_t k = 0; k < p->NumberOfActiveSurfaces; ++k) {
2889 		s->TotalPixelBW = s->TotalPixelBW + p->DPPPerSurface[k]
2890 					* (p->SwathWidthY[k] * p->BytePerPixelDETY[k] * p->VRatio[k] + p->SwathWidthC[k] * p->BytePerPixelDETC[k] * p->VRatioChroma[k]) / (p->HTotal[k] / p->PixelClock[k]);
2891 	}
2892 
2893 	for (dml_uint_t k = 0; k < p->NumberOfActiveSurfaces; ++k) {
2894 
2895 		s->LBLatencyHidingSourceLinesY[k] = (dml_uint_t)(dml_min((dml_float_t)p->MaxLineBufferLines, dml_floor((dml_float_t)p->LineBufferSize / (dml_float_t)p->LBBitPerPixel[k] / ((dml_float_t)p->SwathWidthY[k] / dml_max(p->HRatio[k], 1.0)), 1)) - (p->VTaps[k] - 1));
2896 		s->LBLatencyHidingSourceLinesC[k] = (dml_uint_t)(dml_min((dml_float_t)p->MaxLineBufferLines, dml_floor((dml_float_t)p->LineBufferSize / (dml_float_t)p->LBBitPerPixel[k] / ((dml_float_t)p->SwathWidthC[k] / dml_max(p->HRatioChroma[k], 1.0)), 1)) - (p->VTapsChroma[k] - 1));
2897 
2898 
2899 #ifdef __DML_VBA_DEBUG__
2900 		dml_print("DML::%s: k=%u, MaxLineBufferLines = %u\n", __func__, k, p->MaxLineBufferLines);
2901 		dml_print("DML::%s: k=%u, LineBufferSize = %u\n", __func__, k, p->LineBufferSize);
2902 		dml_print("DML::%s: k=%u, LBBitPerPixel = %u\n", __func__, k, p->LBBitPerPixel[k]);
2903 		dml_print("DML::%s: k=%u, HRatio = %f\n", __func__, k, p->HRatio[k]);
2904 		dml_print("DML::%s: k=%u, VTaps              = %u\n", __func__, k, p->VTaps[k]);
2905 #endif
2906 
2907 		s->EffectiveLBLatencyHidingY = s->LBLatencyHidingSourceLinesY[k] / p->VRatio[k] * (p->HTotal[k] / p->PixelClock[k]);
2908 		s->EffectiveLBLatencyHidingC = s->LBLatencyHidingSourceLinesC[k] / p->VRatioChroma[k] * (p->HTotal[k] / p->PixelClock[k]);
2909 
2910 		s->EffectiveDETBufferSizeY = p->DETBufferSizeY[k];
2911 		if (p->UnboundedRequestEnabled) {
2912 			s->EffectiveDETBufferSizeY = s->EffectiveDETBufferSizeY + p->CompressedBufferSizeInkByte * 1024 * (p->SwathWidthY[k] * p->BytePerPixelDETY[k] * p->VRatio[k]) / (p->HTotal[k] / p->PixelClock[k]) / s->TotalPixelBW;
2913 		}
2914 
2915 		s->LinesInDETY[k] = (dml_float_t)s->EffectiveDETBufferSizeY / p->BytePerPixelDETY[k] / p->SwathWidthY[k];
2916 		s->LinesInDETYRoundedDownToSwath[k] = (dml_uint_t)(dml_floor(s->LinesInDETY[k], p->SwathHeightY[k]));
2917 		s->FullDETBufferingTimeY = s->LinesInDETYRoundedDownToSwath[k] * (p->HTotal[k] / p->PixelClock[k]) / p->VRatio[k];
2918 
2919 		s->ActiveClockChangeLatencyHidingY = s->EffectiveLBLatencyHidingY + s->FullDETBufferingTimeY - ((dml_float_t)p->DSTXAfterScaler[k] / (dml_float_t)p->HTotal[k] + (dml_float_t)p->DSTYAfterScaler[k]) * (dml_float_t)p->HTotal[k] / p->PixelClock[k];
2920 
2921 		if (p->NumberOfActiveSurfaces > 1) {
2922 			s->ActiveClockChangeLatencyHidingY = s->ActiveClockChangeLatencyHidingY - (1.0 - 1.0 / (dml_float_t)p->NumberOfActiveSurfaces) * (dml_float_t)p->SwathHeightY[k] * (dml_float_t)p->HTotal[k] / p->PixelClock[k] / p->VRatio[k];
2923 	}
2924 
2925 		if (p->BytePerPixelDETC[k] > 0) {
2926 			s->LinesInDETC[k] = p->DETBufferSizeC[k] / p->BytePerPixelDETC[k] / p->SwathWidthC[k];
2927 			s->LinesInDETCRoundedDownToSwath[k] = (dml_uint_t)(dml_floor(s->LinesInDETC[k], p->SwathHeightC[k]));
2928 			s->FullDETBufferingTimeC = s->LinesInDETCRoundedDownToSwath[k] * (p->HTotal[k] / p->PixelClock[k]) / p->VRatioChroma[k];
2929 			s->ActiveClockChangeLatencyHidingC = s->EffectiveLBLatencyHidingC + s->FullDETBufferingTimeC - ((dml_float_t)p->DSTXAfterScaler[k] / (dml_float_t)p->HTotal[k] + (dml_float_t)p->DSTYAfterScaler[k]) * (dml_float_t)p->HTotal[k] / p->PixelClock[k];
2930 			if (p->NumberOfActiveSurfaces > 1) {
2931 			s->ActiveClockChangeLatencyHidingC = s->ActiveClockChangeLatencyHidingC - (1.0 - 1.0 / (dml_float_t)p->NumberOfActiveSurfaces) * (dml_float_t)p->SwathHeightC[k] * (dml_float_t)p->HTotal[k] / p->PixelClock[k] / p->VRatioChroma[k];
2932 			}
2933 			s->ActiveClockChangeLatencyHiding = dml_min(s->ActiveClockChangeLatencyHidingY, s->ActiveClockChangeLatencyHidingC);
2934 		} else {
2935 			s->ActiveClockChangeLatencyHiding = s->ActiveClockChangeLatencyHidingY;
2936 		}
2937 
2938 		s->ActiveDRAMClockChangeLatencyMargin[k] = s->ActiveClockChangeLatencyHiding - p->Watermark->UrgentWatermark - p->Watermark->DRAMClockChangeWatermark;
2939 		s->ActiveFCLKChangeLatencyMargin[k] = s->ActiveClockChangeLatencyHiding - p->Watermark->UrgentWatermark - p->Watermark->FCLKChangeWatermark;
2940 		s->USRRetrainingLatencyMargin[k] = s->ActiveClockChangeLatencyHiding - p->Watermark->USRRetrainingWatermark;
2941 
2942 		if (p->WritebackEnable[k]) {
2943 			s->WritebackLatencyHiding = (dml_float_t)p->WritebackInterfaceBufferSize * 1024.0 / ((dml_float_t)p->WritebackDestinationWidth[k] * (dml_float_t)p->WritebackDestinationHeight[k] / ((dml_float_t)p->WritebackSourceHeight[k] * (dml_float_t)p->HTotal[k] / p->PixelClock[k]) * 4.0);
2944 			if (p->WritebackPixelFormat[k] == dml_444_64) {
2945 				s->WritebackLatencyHiding = s->WritebackLatencyHiding / 2;
2946 			}
2947 			s->WritebackDRAMClockChangeLatencyMargin = s->WritebackLatencyHiding - p->Watermark->WritebackDRAMClockChangeWatermark;
2948 
2949 			s->WritebackFCLKChangeLatencyMargin = s->WritebackLatencyHiding - p->Watermark->WritebackFCLKChangeWatermark;
2950 
2951 			s->ActiveDRAMClockChangeLatencyMargin[k] = dml_min(s->ActiveDRAMClockChangeLatencyMargin[k], s->WritebackFCLKChangeLatencyMargin);
2952 			s->ActiveFCLKChangeLatencyMargin[k] = dml_min(s->ActiveFCLKChangeLatencyMargin[k], s->WritebackDRAMClockChangeLatencyMargin);
2953 		}
2954 		p->MaxActiveDRAMClockChangeLatencySupported[k] =  (p->UseMALLForPStateChange[k] == dml_use_mall_pstate_change_phantom_pipe) ? 0 : (s->ActiveDRAMClockChangeLatencyMargin[k] + p->mmSOCParameters.DRAMClockChangeLatency);
2955 		p->ActiveDRAMClockChangeLatencyMargin[k] = s->ActiveDRAMClockChangeLatencyMargin[k];
2956 	}
2957 
2958 	*p->USRRetrainingSupport = true;
2959 	for (dml_uint_t k = 0; k < p->NumberOfActiveSurfaces; ++k) {
2960 		if ((p->UseMALLForPStateChange[k] != dml_use_mall_pstate_change_phantom_pipe) && (s->USRRetrainingLatencyMargin[k] < 0)) {
2961 			*p->USRRetrainingSupport = false;
2962 		}
2963 	}
2964 
2965 	s->FoundCriticalSurface = false;
2966 	for (dml_uint_t k = 0; k < p->NumberOfActiveSurfaces; ++k) {
2967 		if ((p->UseMALLForPStateChange[k] != dml_use_mall_pstate_change_phantom_pipe) && ((!s->FoundCriticalSurface)
2968 			|| ((s->ActiveFCLKChangeLatencyMargin[k] + p->mmSOCParameters.FCLKChangeLatency) < *p->MaxActiveFCLKChangeLatencySupported))) {
2969 			s->FoundCriticalSurface = true;
2970 			*p->MaxActiveFCLKChangeLatencySupported = s->ActiveFCLKChangeLatencyMargin[k] + p->mmSOCParameters.FCLKChangeLatency;
2971 		}
2972 	}
2973 
2974 	for (dml_uint_t i = 0; i < p->NumberOfActiveSurfaces; ++i) {
2975 		for (dml_uint_t j = 0; j < p->NumberOfActiveSurfaces; ++j) {
2976 			if (i == j ||
2977 				(p->BlendingAndTiming[i] == i && p->BlendingAndTiming[j] == i) ||
2978 				(p->BlendingAndTiming[j] == j && p->BlendingAndTiming[i] == j) ||
2979 				(p->BlendingAndTiming[i] == p->BlendingAndTiming[j] && p->BlendingAndTiming[i] != i) ||
2980 				(p->SynchronizeTimingsFinal && p->PixelClock[i] == p->PixelClock[j] && p->HTotal[i] == p->HTotal[j] && p->VTotal[i] == p->VTotal[j] && p->VActive[i] == p->VActive[j]) ||
2981 				(p->SynchronizeDRRDisplaysForUCLKPStateChangeFinal && (p->DRRDisplay[i] || p->DRRDisplay[j]))) {
2982 				s->SynchronizedSurfaces[i][j] = true;
2983 			} else {
2984 				s->SynchronizedSurfaces[i][j] = false;
2985 			}
2986 		}
2987 	}
2988 
2989 	s->FCLKChangeSupportNumber = 0;
2990 	for (dml_uint_t k = 0; k < p->NumberOfActiveSurfaces; ++k) {
2991 		if ((p->UseMALLForPStateChange[k] != dml_use_mall_pstate_change_phantom_pipe) &&  (s->ActiveFCLKChangeLatencyMargin[k] < 0)) {
2992 			if (!(p->PrefetchMode[k] <= 1)) {
2993 				s->FCLKChangeSupportNumber = 3;
2994 			} else if (s->FCLKChangeSupportNumber == 0) {
2995 				s->FCLKChangeSupportNumber = ((p->SynchronizeDRRDisplaysForUCLKPStateChangeFinal && p->DRRDisplay[k]) ? 2 : 1);
2996 				s->LastSurfaceWithoutMargin = k;
2997 			} else if (((s->FCLKChangeSupportNumber == 1) && (p->DRRDisplay[k] || (!s->SynchronizedSurfaces[s->LastSurfaceWithoutMargin][k]))) || (s->FCLKChangeSupportNumber == 2))
2998 				s->FCLKChangeSupportNumber = 3;
2999 		}
3000 	}
3001 
3002 	if (s->FCLKChangeSupportNumber == 0) {
3003 		*p->FCLKChangeSupport = dml_fclock_change_vactive;
3004 	} else if ((s->FCLKChangeSupportNumber == 1) || (s->FCLKChangeSupportNumber == 2)) {
3005 		*p->FCLKChangeSupport = dml_fclock_change_vblank;
3006 	} else {
3007 		*p->FCLKChangeSupport = dml_fclock_change_unsupported;
3008 	}
3009 
3010 	s->DRAMClockChangeMethod = 0;
3011 	for (dml_uint_t k = 0; k < p->NumberOfActiveSurfaces; ++k) {
3012 		if (p->UseMALLForPStateChange[k] == dml_use_mall_pstate_change_full_frame)
3013 			s->DRAMClockChangeMethod = 1;
3014 		else if (p->UseMALLForPStateChange[k] == dml_use_mall_pstate_change_sub_viewport)
3015 			s->DRAMClockChangeMethod = 2;
3016 	}
3017 
3018 	s->DRAMClockChangeSupportNumber = 0;
3019 	for (dml_uint_t k = 0; k < p->NumberOfActiveSurfaces; ++k) {
3020 		if (((s->DRAMClockChangeMethod == 0) && (s->ActiveDRAMClockChangeLatencyMargin[k] < 0)) ||
3021 			((s->DRAMClockChangeMethod == 1) && (p->UseMALLForPStateChange[k] != dml_use_mall_pstate_change_full_frame)) ||
3022 			((s->DRAMClockChangeMethod == 2) && (p->UseMALLForPStateChange[k] != dml_use_mall_pstate_change_sub_viewport) && (p->UseMALLForPStateChange[k] != dml_use_mall_pstate_change_phantom_pipe))) {
3023 			if (p->PrefetchMode[k] != 0) { // Don't need to support DRAM clock change, PrefetchMode 0 means needs DRAM clock change support
3024 				s->DRAMClockChangeSupportNumber = 3;
3025 			} else if (s->DRAMClockChangeSupportNumber == 0) {
3026 				s->DRAMClockChangeSupportNumber =  (p->SynchronizeDRRDisplaysForUCLKPStateChangeFinal && p->DRRDisplay[k]) ? 2 : 1;
3027 				s->LastSurfaceWithoutMargin = k;
3028 			} else if (((s->DRAMClockChangeSupportNumber == 1) && (p->DRRDisplay[k] || !s->SynchronizedSurfaces[s->LastSurfaceWithoutMargin][k])) || (s->DRAMClockChangeSupportNumber == 2)) {
3029 				s->DRAMClockChangeSupportNumber = 3;
3030 			}
3031 		}
3032 	}
3033 
3034 	if (s->DRAMClockChangeMethod == 0) { // No MALL usage
3035 		if (s->DRAMClockChangeSupportNumber == 0) {
3036 			*p->DRAMClockChangeSupport = dml_dram_clock_change_vactive;
3037 		} else if (s->DRAMClockChangeSupportNumber == 1) {
3038 			*p->DRAMClockChangeSupport = dml_dram_clock_change_vblank;
3039 		} else if (s->DRAMClockChangeSupportNumber == 2) {
3040 			*p->DRAMClockChangeSupport = dml_dram_clock_change_vblank_drr;
3041 		} else {
3042 			*p->DRAMClockChangeSupport = dml_dram_clock_change_unsupported;
3043 		}
3044 	} else if (s->DRAMClockChangeMethod == 1) { // Any pipe using MALL full frame
3045 		if (s->DRAMClockChangeSupportNumber == 0) {
3046 			*p->DRAMClockChangeSupport = dml_dram_clock_change_vactive_w_mall_full_frame;
3047 		} else if (s->DRAMClockChangeSupportNumber == 1) {
3048 			*p->DRAMClockChangeSupport = dml_dram_clock_change_vblank_w_mall_full_frame;
3049 		} else if (s->DRAMClockChangeSupportNumber == 2) {
3050 			*p->DRAMClockChangeSupport = dml_dram_clock_change_vblank_drr_w_mall_full_frame;
3051 		} else {
3052 			*p->DRAMClockChangeSupport = dml_dram_clock_change_unsupported;
3053 		}
3054 	} else { // Any pipe using MALL subviewport
3055 		if (s->DRAMClockChangeSupportNumber == 0) {
3056 			*p->DRAMClockChangeSupport = dml_dram_clock_change_vactive_w_mall_sub_vp;
3057 		} else if (s->DRAMClockChangeSupportNumber == 1) {
3058 			*p->DRAMClockChangeSupport = dml_dram_clock_change_vblank_w_mall_sub_vp;
3059 		} else if (s->DRAMClockChangeSupportNumber == 2) {
3060 			*p->DRAMClockChangeSupport = dml_dram_clock_change_vblank_drr_w_mall_sub_vp;
3061 		} else {
3062 			*p->DRAMClockChangeSupport = dml_dram_clock_change_unsupported;
3063 		}
3064 	}
3065 
3066 	for (dml_uint_t k = 0; k < p->NumberOfActiveSurfaces; ++k) {
3067 		s->dst_y_pstate = (dml_uint_t)(dml_ceil((p->mmSOCParameters.DRAMClockChangeLatency + p->mmSOCParameters.UrgentLatency) / (p->HTotal[k] / p->PixelClock[k]), 1));
3068 		s->src_y_pstate_l = (dml_uint_t)(dml_ceil(s->dst_y_pstate * p->VRatio[k], p->SwathHeightY[k]));
3069 		s->src_y_ahead_l = (dml_uint_t)(dml_floor(p->DETBufferSizeY[k] / p->BytePerPixelDETY[k] / p->SwathWidthY[k], p->SwathHeightY[k]) + s->LBLatencyHidingSourceLinesY[k]);
3070 		s->sub_vp_lines_l = s->src_y_pstate_l + s->src_y_ahead_l + p->meta_row_height[k];
3071 
3072 #ifdef __DML_VBA_DEBUG__
3073 		dml_print("DML::%s: k=%u, DETBufferSizeY = %u\n", __func__, k, p->DETBufferSizeY[k]);
3074 		dml_print("DML::%s: k=%u, BytePerPixelDETY = %f\n", __func__, k, p->BytePerPixelDETY[k]);
3075 		dml_print("DML::%s: k=%u, SwathWidthY = %u\n", __func__, k, p->SwathWidthY[k]);
3076 		dml_print("DML::%s: k=%u, SwathHeightY = %u\n", __func__, k, p->SwathHeightY[k]);
3077 		dml_print("DML::%s: k=%u, LBLatencyHidingSourceLinesY = %u\n", __func__, k, s->LBLatencyHidingSourceLinesY[k]);
3078 		dml_print("DML::%s: k=%u, dst_y_pstate = %u\n", __func__, k, s->dst_y_pstate);
3079 		dml_print("DML::%s: k=%u, src_y_pstate_l = %u\n", __func__, k, s->src_y_pstate_l);
3080 		dml_print("DML::%s: k=%u, src_y_ahead_l = %u\n", __func__, k, s->src_y_ahead_l);
3081 		dml_print("DML::%s: k=%u, meta_row_height = %u\n", __func__, k, p->meta_row_height[k]);
3082 		dml_print("DML::%s: k=%u, sub_vp_lines_l  = %u\n", __func__, k, s->sub_vp_lines_l);
3083 #endif
3084 		p->SubViewportLinesNeededInMALL[k] = s->sub_vp_lines_l;
3085 
3086 		if (p->BytePerPixelDETC[k] > 0) {
3087 		s->src_y_pstate_c = (dml_uint_t)(dml_ceil(s->dst_y_pstate * p->VRatioChroma[k], p->SwathHeightC[k]));
3088 		s->src_y_ahead_c = (dml_uint_t)(dml_floor(p->DETBufferSizeC[k] / p->BytePerPixelDETC[k] / p->SwathWidthC[k], p->SwathHeightC[k]) + s->LBLatencyHidingSourceLinesC[k]);
3089 		s->sub_vp_lines_c = s->src_y_pstate_c + s->src_y_ahead_c + p->meta_row_height_chroma[k];
3090 		p->SubViewportLinesNeededInMALL[k] = (dml_uint_t)(dml_max(s->sub_vp_lines_l, s->sub_vp_lines_c));
3091 
3092 #ifdef __DML_VBA_DEBUG__
3093 		dml_print("DML::%s: k=%u, src_y_pstate_c = %u\n", __func__, k, s->src_y_pstate_c);
3094 		dml_print("DML::%s: k=%u, src_y_ahead_c = %u\n", __func__, k, s->src_y_ahead_c);
3095 		dml_print("DML::%s: k=%u, meta_row_height_chroma = %u\n", __func__, k, p->meta_row_height_chroma[k]);
3096 		dml_print("DML::%s: k=%u, sub_vp_lines_c            = %u\n", __func__, k, s->sub_vp_lines_c);
3097 #endif
3098 		}
3099 	}
3100 
3101 #ifdef __DML_VBA_DEBUG__
3102 	dml_print("DML::%s: DRAMClockChangeSupport = %u\n", __func__, *p->DRAMClockChangeSupport);
3103 	dml_print("DML::%s: FCLKChangeSupport = %u\n", __func__, *p->FCLKChangeSupport);
3104 	dml_print("DML::%s: MaxActiveFCLKChangeLatencySupported = %f\n", __func__, *p->MaxActiveFCLKChangeLatencySupported);
3105 	dml_print("DML::%s: USRRetrainingSupport                        = %u\n", __func__, *p->USRRetrainingSupport);
3106 #endif
3107 } // CalculateWatermarksMALLUseAndDRAMSpeedChangeSupport
3108 
CalculateDCFCLKDeepSleep(dml_uint_t NumberOfActiveSurfaces,dml_uint_t BytePerPixelY[],dml_uint_t BytePerPixelC[],dml_float_t VRatio[],dml_float_t VRatioChroma[],dml_uint_t SwathWidthY[],dml_uint_t SwathWidthC[],dml_uint_t DPPPerSurface[],dml_float_t HRatio[],dml_float_t HRatioChroma[],dml_float_t PixelClock[],dml_float_t PSCL_THROUGHPUT[],dml_float_t PSCL_THROUGHPUT_CHROMA[],dml_float_t Dppclk[],dml_float_t ReadBandwidthLuma[],dml_float_t ReadBandwidthChroma[],dml_uint_t ReturnBusWidth,dml_float_t * DCFClkDeepSleep)3109 static void CalculateDCFCLKDeepSleep(
3110 		dml_uint_t NumberOfActiveSurfaces,
3111 		dml_uint_t BytePerPixelY[],
3112 		dml_uint_t BytePerPixelC[],
3113 		dml_float_t VRatio[],
3114 		dml_float_t VRatioChroma[],
3115 		dml_uint_t SwathWidthY[],
3116 		dml_uint_t SwathWidthC[],
3117 		dml_uint_t DPPPerSurface[],
3118 		dml_float_t HRatio[],
3119 		dml_float_t HRatioChroma[],
3120 		dml_float_t PixelClock[],
3121 		dml_float_t PSCL_THROUGHPUT[],
3122 		dml_float_t PSCL_THROUGHPUT_CHROMA[],
3123 		dml_float_t Dppclk[],
3124 		dml_float_t ReadBandwidthLuma[],
3125 		dml_float_t ReadBandwidthChroma[],
3126 		dml_uint_t ReturnBusWidth,
3127 
3128 		// Output
3129 		dml_float_t *DCFClkDeepSleep)
3130 {
3131 	dml_float_t DisplayPipeLineDeliveryTimeLuma;
3132 	dml_float_t DisplayPipeLineDeliveryTimeChroma;
3133 	dml_float_t DCFClkDeepSleepPerSurface[__DML_NUM_PLANES__];
3134 	dml_float_t ReadBandwidth = 0.0;
3135 
3136 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
3137 
3138 		if (VRatio[k] <= 1) {
3139 			DisplayPipeLineDeliveryTimeLuma = SwathWidthY[k] * DPPPerSurface[k] / HRatio[k] / PixelClock[k];
3140 		} else {
3141 			DisplayPipeLineDeliveryTimeLuma = SwathWidthY[k] / PSCL_THROUGHPUT[k] / Dppclk[k];
3142 		}
3143 		if (BytePerPixelC[k] == 0) {
3144 			DisplayPipeLineDeliveryTimeChroma = 0;
3145 		} else {
3146 			if (VRatioChroma[k] <= 1) {
3147 				DisplayPipeLineDeliveryTimeChroma = SwathWidthC[k] * DPPPerSurface[k] / HRatioChroma[k] / PixelClock[k];
3148 			} else {
3149 				DisplayPipeLineDeliveryTimeChroma = SwathWidthC[k] / PSCL_THROUGHPUT_CHROMA[k] / Dppclk[k];
3150 			}
3151 		}
3152 
3153 		if (BytePerPixelC[k] > 0) {
3154 			DCFClkDeepSleepPerSurface[k] = dml_max(__DML_MIN_DCFCLK_FACTOR__ * SwathWidthY[k] * BytePerPixelY[k] / 32.0 / DisplayPipeLineDeliveryTimeLuma,
3155 														__DML_MIN_DCFCLK_FACTOR__ * SwathWidthC[k] * BytePerPixelC[k] / 32.0 / DisplayPipeLineDeliveryTimeChroma);
3156 		} else {
3157 			DCFClkDeepSleepPerSurface[k] = __DML_MIN_DCFCLK_FACTOR__ * SwathWidthY[k] * BytePerPixelY[k] / 64.0 / DisplayPipeLineDeliveryTimeLuma;
3158 		}
3159 		DCFClkDeepSleepPerSurface[k] = dml_max(DCFClkDeepSleepPerSurface[k], PixelClock[k] / 16);
3160 
3161 #ifdef __DML_VBA_DEBUG__
3162 		dml_print("DML::%s: k=%u, PixelClock = %f\n", __func__, k, PixelClock[k]);
3163 		dml_print("DML::%s: k=%u, DCFClkDeepSleepPerSurface = %f\n", __func__, k, DCFClkDeepSleepPerSurface[k]);
3164 #endif
3165 	}
3166 
3167 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
3168 		ReadBandwidth = ReadBandwidth + ReadBandwidthLuma[k] + ReadBandwidthChroma[k];
3169 	}
3170 
3171 	*DCFClkDeepSleep = dml_max(8.0, __DML_MIN_DCFCLK_FACTOR__ * ReadBandwidth / (dml_float_t) ReturnBusWidth);
3172 
3173 #ifdef __DML_VBA_DEBUG__
3174 		dml_print("DML::%s: __DML_MIN_DCFCLK_FACTOR__ = %f\n", __func__, __DML_MIN_DCFCLK_FACTOR__);
3175 		dml_print("DML::%s: ReadBandwidth = %f\n", __func__, ReadBandwidth);
3176 		dml_print("DML::%s: ReturnBusWidth = %u\n", __func__, ReturnBusWidth);
3177 		dml_print("DML::%s: DCFClkDeepSleep = %f\n", __func__, *DCFClkDeepSleep);
3178 #endif
3179 
3180 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
3181 		*DCFClkDeepSleep = dml_max(*DCFClkDeepSleep, DCFClkDeepSleepPerSurface[k]);
3182 	}
3183 	dml_print("DML::%s: DCFClkDeepSleep = %f (final)\n", __func__, *DCFClkDeepSleep);
3184 } // CalculateDCFCLKDeepSleep
3185 
CalculateUrgentBurstFactor(enum dml_use_mall_for_pstate_change_mode UseMALLForPStateChange,dml_uint_t swath_width_luma_ub,dml_uint_t swath_width_chroma_ub,dml_uint_t SwathHeightY,dml_uint_t SwathHeightC,dml_float_t LineTime,dml_float_t UrgentLatency,dml_float_t CursorBufferSize,dml_uint_t CursorWidth,dml_uint_t CursorBPP,dml_float_t VRatio,dml_float_t VRatioC,dml_float_t BytePerPixelInDETY,dml_float_t BytePerPixelInDETC,dml_uint_t DETBufferSizeY,dml_uint_t DETBufferSizeC,dml_float_t * UrgentBurstFactorCursor,dml_float_t * UrgentBurstFactorLuma,dml_float_t * UrgentBurstFactorChroma,dml_bool_t * NotEnoughUrgentLatencyHiding)3186 static void CalculateUrgentBurstFactor(
3187 		enum dml_use_mall_for_pstate_change_mode UseMALLForPStateChange,
3188 		dml_uint_t swath_width_luma_ub,
3189 		dml_uint_t swath_width_chroma_ub,
3190 		dml_uint_t SwathHeightY,
3191 		dml_uint_t SwathHeightC,
3192 		dml_float_t LineTime,
3193 		dml_float_t UrgentLatency,
3194 		dml_float_t CursorBufferSize,
3195 		dml_uint_t CursorWidth,
3196 		dml_uint_t CursorBPP,
3197 		dml_float_t VRatio,
3198 		dml_float_t VRatioC,
3199 		dml_float_t BytePerPixelInDETY,
3200 		dml_float_t BytePerPixelInDETC,
3201 		dml_uint_t DETBufferSizeY,
3202 		dml_uint_t DETBufferSizeC,
3203 		// Output
3204 		dml_float_t *UrgentBurstFactorCursor,
3205 		dml_float_t *UrgentBurstFactorLuma,
3206 		dml_float_t *UrgentBurstFactorChroma,
3207 		dml_bool_t *NotEnoughUrgentLatencyHiding)
3208 {
3209 	dml_float_t LinesInDETLuma;
3210 	dml_float_t LinesInDETChroma;
3211 	dml_uint_t LinesInCursorBuffer;
3212 	dml_float_t CursorBufferSizeInTime;
3213 	dml_float_t DETBufferSizeInTimeLuma;
3214 	dml_float_t DETBufferSizeInTimeChroma;
3215 
3216 	*NotEnoughUrgentLatencyHiding = 0;
3217 
3218 	if (CursorWidth > 0) {
3219 		LinesInCursorBuffer = 1 << (dml_uint_t) dml_floor(dml_log2(CursorBufferSize * 1024.0 / (CursorWidth * CursorBPP / 8.0)), 1.0);
3220 		if (VRatio > 0) {
3221 			CursorBufferSizeInTime = LinesInCursorBuffer * LineTime / VRatio;
3222 			if (CursorBufferSizeInTime - UrgentLatency <= 0) {
3223 				*NotEnoughUrgentLatencyHiding = 1;
3224 				*UrgentBurstFactorCursor = 0;
3225 			} else {
3226 				*UrgentBurstFactorCursor = CursorBufferSizeInTime / (CursorBufferSizeInTime - UrgentLatency);
3227 			}
3228 		} else {
3229 			*UrgentBurstFactorCursor = 1;
3230 		}
3231 	}
3232 
3233 	LinesInDETLuma = (UseMALLForPStateChange == dml_use_mall_pstate_change_phantom_pipe ? 1024*1024 : DETBufferSizeY) / BytePerPixelInDETY / swath_width_luma_ub;
3234 
3235 	if (VRatio > 0) {
3236 		DETBufferSizeInTimeLuma = dml_floor(LinesInDETLuma, SwathHeightY) * LineTime / VRatio;
3237 		if (DETBufferSizeInTimeLuma - UrgentLatency <= 0) {
3238 			*NotEnoughUrgentLatencyHiding = 1;
3239 			*UrgentBurstFactorLuma = 0;
3240 		} else {
3241 			*UrgentBurstFactorLuma = DETBufferSizeInTimeLuma / (DETBufferSizeInTimeLuma - UrgentLatency);
3242 		}
3243 	} else {
3244 		*UrgentBurstFactorLuma = 1;
3245 	}
3246 
3247 	if (BytePerPixelInDETC > 0) {
3248 		LinesInDETChroma = (UseMALLForPStateChange == dml_use_mall_pstate_change_phantom_pipe ? 1024*1024 : DETBufferSizeC) / BytePerPixelInDETC / swath_width_chroma_ub;
3249 
3250 		if (VRatioC > 0) {
3251 			DETBufferSizeInTimeChroma = dml_floor(LinesInDETChroma, SwathHeightC) * LineTime / VRatioC;
3252 			if (DETBufferSizeInTimeChroma - UrgentLatency <= 0) {
3253 				*NotEnoughUrgentLatencyHiding = 1;
3254 				*UrgentBurstFactorChroma = 0;
3255 			} else {
3256 				*UrgentBurstFactorChroma = DETBufferSizeInTimeChroma / (DETBufferSizeInTimeChroma - UrgentLatency);
3257 			}
3258 		} else {
3259 			*UrgentBurstFactorChroma = 1;
3260 		}
3261 	}
3262 } // CalculateUrgentBurstFactor
3263 
CalculatePixelDeliveryTimes(dml_uint_t NumberOfActiveSurfaces,dml_float_t VRatio[],dml_float_t VRatioChroma[],dml_float_t VRatioPrefetchY[],dml_float_t VRatioPrefetchC[],dml_uint_t swath_width_luma_ub[],dml_uint_t swath_width_chroma_ub[],dml_uint_t DPPPerSurface[],dml_float_t HRatio[],dml_float_t HRatioChroma[],dml_float_t PixelClock[],dml_float_t PSCL_THROUGHPUT[],dml_float_t PSCL_THROUGHPUT_CHROMA[],dml_float_t Dppclk[],dml_uint_t BytePerPixelC[],enum dml_rotation_angle SourceScan[],dml_uint_t NumberOfCursors[],dml_uint_t CursorWidth[],dml_uint_t CursorBPP[],dml_uint_t BlockWidth256BytesY[],dml_uint_t BlockHeight256BytesY[],dml_uint_t BlockWidth256BytesC[],dml_uint_t BlockHeight256BytesC[],dml_float_t DisplayPipeLineDeliveryTimeLuma[],dml_float_t DisplayPipeLineDeliveryTimeChroma[],dml_float_t DisplayPipeLineDeliveryTimeLumaPrefetch[],dml_float_t DisplayPipeLineDeliveryTimeChromaPrefetch[],dml_float_t DisplayPipeRequestDeliveryTimeLuma[],dml_float_t DisplayPipeRequestDeliveryTimeChroma[],dml_float_t DisplayPipeRequestDeliveryTimeLumaPrefetch[],dml_float_t DisplayPipeRequestDeliveryTimeChromaPrefetch[],dml_float_t CursorRequestDeliveryTime[],dml_float_t CursorRequestDeliveryTimePrefetch[])3264 static void CalculatePixelDeliveryTimes(
3265 		dml_uint_t NumberOfActiveSurfaces,
3266 		dml_float_t VRatio[],
3267 		dml_float_t VRatioChroma[],
3268 		dml_float_t VRatioPrefetchY[],
3269 		dml_float_t VRatioPrefetchC[],
3270 		dml_uint_t swath_width_luma_ub[],
3271 		dml_uint_t swath_width_chroma_ub[],
3272 		dml_uint_t DPPPerSurface[],
3273 		dml_float_t HRatio[],
3274 		dml_float_t HRatioChroma[],
3275 		dml_float_t PixelClock[],
3276 		dml_float_t PSCL_THROUGHPUT[],
3277 		dml_float_t PSCL_THROUGHPUT_CHROMA[],
3278 		dml_float_t Dppclk[],
3279 		dml_uint_t BytePerPixelC[],
3280 		enum dml_rotation_angle SourceScan[],
3281 		dml_uint_t NumberOfCursors[],
3282 		dml_uint_t CursorWidth[],
3283 		dml_uint_t CursorBPP[],
3284 		dml_uint_t BlockWidth256BytesY[],
3285 		dml_uint_t BlockHeight256BytesY[],
3286 		dml_uint_t BlockWidth256BytesC[],
3287 		dml_uint_t BlockHeight256BytesC[],
3288 
3289 		// Output
3290 		dml_float_t DisplayPipeLineDeliveryTimeLuma[],
3291 		dml_float_t DisplayPipeLineDeliveryTimeChroma[],
3292 		dml_float_t DisplayPipeLineDeliveryTimeLumaPrefetch[],
3293 		dml_float_t DisplayPipeLineDeliveryTimeChromaPrefetch[],
3294 		dml_float_t DisplayPipeRequestDeliveryTimeLuma[],
3295 		dml_float_t DisplayPipeRequestDeliveryTimeChroma[],
3296 		dml_float_t DisplayPipeRequestDeliveryTimeLumaPrefetch[],
3297 		dml_float_t DisplayPipeRequestDeliveryTimeChromaPrefetch[],
3298 		dml_float_t CursorRequestDeliveryTime[],
3299 		dml_float_t CursorRequestDeliveryTimePrefetch[])
3300 {
3301 		dml_float_t req_per_swath_ub;
3302 
3303 		for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
3304 
3305 #ifdef __DML_VBA_DEBUG__
3306 		dml_print("DML::%s: k=%u : HRatio = %f\n", __func__, k, HRatio[k]);
3307 		dml_print("DML::%s: k=%u : VRatio = %f\n", __func__, k, VRatio[k]);
3308 		dml_print("DML::%s: k=%u : HRatioChroma = %f\n", __func__, k, HRatioChroma[k]);
3309 		dml_print("DML::%s: k=%u : VRatioChroma = %f\n", __func__, k, VRatioChroma[k]);
3310 		dml_print("DML::%s: k=%u : swath_width_luma_ub = %u\n", __func__, k, swath_width_luma_ub[k]);
3311 		dml_print("DML::%s: k=%u : swath_width_chroma_ub = %u\n", __func__, k, swath_width_chroma_ub[k]);
3312 		dml_print("DML::%s: k=%u : PSCL_THROUGHPUT = %f\n", __func__, k, PSCL_THROUGHPUT[k]);
3313 		dml_print("DML::%s: k=%u : PSCL_THROUGHPUT_CHROMA = %f\n", __func__, k, PSCL_THROUGHPUT_CHROMA[k]);
3314 		dml_print("DML::%s: k=%u : DPPPerSurface = %u\n", __func__, k, DPPPerSurface[k]);
3315 		dml_print("DML::%s: k=%u : PixelClock = %f\n", __func__, k, PixelClock[k]);
3316 		dml_print("DML::%s: k=%u : Dppclk = %f\n", __func__, k, Dppclk[k]);
3317 #endif
3318 
3319 		if (VRatio[k] <= 1) {
3320 			DisplayPipeLineDeliveryTimeLuma[k] = swath_width_luma_ub[k] * DPPPerSurface[k] / HRatio[k] / PixelClock[k];
3321 		} else {
3322 			DisplayPipeLineDeliveryTimeLuma[k] = swath_width_luma_ub[k] / PSCL_THROUGHPUT[k] / Dppclk[k];
3323 		}
3324 
3325 		if (BytePerPixelC[k] == 0) {
3326 			DisplayPipeLineDeliveryTimeChroma[k] = 0;
3327 		} else {
3328 			if (VRatioChroma[k] <= 1) {
3329 				DisplayPipeLineDeliveryTimeChroma[k] = swath_width_chroma_ub[k] * DPPPerSurface[k] / HRatioChroma[k] / PixelClock[k];
3330 			} else {
3331 				DisplayPipeLineDeliveryTimeChroma[k] = swath_width_chroma_ub[k] / PSCL_THROUGHPUT_CHROMA[k] / Dppclk[k];
3332 			}
3333 		}
3334 
3335 		if (VRatioPrefetchY[k] <= 1) {
3336 			DisplayPipeLineDeliveryTimeLumaPrefetch[k] = swath_width_luma_ub[k] * DPPPerSurface[k] / HRatio[k] / PixelClock[k];
3337 		} else {
3338 			DisplayPipeLineDeliveryTimeLumaPrefetch[k] = swath_width_luma_ub[k] / PSCL_THROUGHPUT[k] / Dppclk[k];
3339 		}
3340 
3341 		if (BytePerPixelC[k] == 0) {
3342 			DisplayPipeLineDeliveryTimeChromaPrefetch[k] = 0;
3343 		} else {
3344 			if (VRatioPrefetchC[k] <= 1) {
3345 				DisplayPipeLineDeliveryTimeChromaPrefetch[k] = swath_width_chroma_ub[k] * DPPPerSurface[k] / HRatioChroma[k] / PixelClock[k];
3346 			} else {
3347 				DisplayPipeLineDeliveryTimeChromaPrefetch[k] = swath_width_chroma_ub[k] / PSCL_THROUGHPUT_CHROMA[k] / Dppclk[k];
3348 			}
3349 		}
3350 #ifdef __DML_VBA_DEBUG__
3351 		dml_print("DML::%s: k=%u : DisplayPipeLineDeliveryTimeLuma = %f\n", __func__, k, DisplayPipeLineDeliveryTimeLuma[k]);
3352 		dml_print("DML::%s: k=%u : DisplayPipeLineDeliveryTimeLumaPrefetch = %f\n", __func__, k, DisplayPipeLineDeliveryTimeLumaPrefetch[k]);
3353 		dml_print("DML::%s: k=%u : DisplayPipeLineDeliveryTimeChroma = %f\n", __func__, k, DisplayPipeLineDeliveryTimeChroma[k]);
3354 		dml_print("DML::%s: k=%u : DisplayPipeLineDeliveryTimeChromaPrefetch = %f\n", __func__, k, DisplayPipeLineDeliveryTimeChromaPrefetch[k]);
3355 #endif
3356 	}
3357 
3358 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
3359 		if (!dml_is_vertical_rotation(SourceScan[k])) {
3360 			req_per_swath_ub = swath_width_luma_ub[k] / BlockWidth256BytesY[k];
3361 		} else {
3362 			req_per_swath_ub = swath_width_luma_ub[k] / BlockHeight256BytesY[k];
3363 		}
3364 #ifdef __DML_VBA_DEBUG__
3365 		dml_print("DML::%s: k=%u : req_per_swath_ub = %f (Luma)\n", __func__, k, req_per_swath_ub);
3366 #endif
3367 
3368 		DisplayPipeRequestDeliveryTimeLuma[k] = DisplayPipeLineDeliveryTimeLuma[k] / req_per_swath_ub;
3369 		DisplayPipeRequestDeliveryTimeLumaPrefetch[k] = DisplayPipeLineDeliveryTimeLumaPrefetch[k] / req_per_swath_ub;
3370 		if (BytePerPixelC[k] == 0) {
3371 			DisplayPipeRequestDeliveryTimeChroma[k] = 0;
3372 			DisplayPipeRequestDeliveryTimeChromaPrefetch[k] = 0;
3373 		} else {
3374 			if (!dml_is_vertical_rotation(SourceScan[k])) {
3375 				req_per_swath_ub = swath_width_chroma_ub[k] / BlockWidth256BytesC[k];
3376 			} else {
3377 				req_per_swath_ub = swath_width_chroma_ub[k] / BlockHeight256BytesC[k];
3378 			}
3379 #ifdef __DML_VBA_DEBUG__
3380 			dml_print("DML::%s: k=%u : req_per_swath_ub = %f (Chroma)\n", __func__, k, req_per_swath_ub);
3381 #endif
3382 			DisplayPipeRequestDeliveryTimeChroma[k] = DisplayPipeLineDeliveryTimeChroma[k] / req_per_swath_ub;
3383 			DisplayPipeRequestDeliveryTimeChromaPrefetch[k] = DisplayPipeLineDeliveryTimeChromaPrefetch[k] / req_per_swath_ub;
3384 		}
3385 #ifdef __DML_VBA_DEBUG__
3386 		dml_print("DML::%s: k=%u : DisplayPipeRequestDeliveryTimeLuma = %f\n", __func__, k, DisplayPipeRequestDeliveryTimeLuma[k]);
3387 		dml_print("DML::%s: k=%u : DisplayPipeRequestDeliveryTimeLumaPrefetch = %f\n", __func__, k, DisplayPipeRequestDeliveryTimeLumaPrefetch[k]);
3388 		dml_print("DML::%s: k=%u : DisplayPipeRequestDeliveryTimeChroma = %f\n", __func__, k, DisplayPipeRequestDeliveryTimeChroma[k]);
3389 		dml_print("DML::%s: k=%u : DisplayPipeRequestDeliveryTimeChromaPrefetch = %f\n", __func__, k, DisplayPipeRequestDeliveryTimeChromaPrefetch[k]);
3390 #endif
3391 	}
3392 
3393 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
3394 		dml_uint_t cursor_req_per_width;
3395 		cursor_req_per_width = (dml_uint_t)(dml_ceil((dml_float_t) CursorWidth[k] * (dml_float_t) CursorBPP[k] / 256.0 / 8.0, 1.0));
3396 		if (NumberOfCursors[k] > 0) {
3397 			if (VRatio[k] <= 1) {
3398 				CursorRequestDeliveryTime[k] = (dml_float_t) CursorWidth[k] / HRatio[k] / PixelClock[k] / cursor_req_per_width;
3399 			} else {
3400 				CursorRequestDeliveryTime[k] = (dml_float_t) CursorWidth[k] / PSCL_THROUGHPUT[k] / Dppclk[k] / cursor_req_per_width;
3401 			}
3402 			if (VRatioPrefetchY[k] <= 1) {
3403 				CursorRequestDeliveryTimePrefetch[k] = (dml_float_t) CursorWidth[k] / HRatio[k] / PixelClock[k] / cursor_req_per_width;
3404 		} else {
3405 			CursorRequestDeliveryTimePrefetch[k] = (dml_float_t) CursorWidth[k] / PSCL_THROUGHPUT[k] / Dppclk[k] / cursor_req_per_width;
3406 			}
3407 		} else {
3408 			CursorRequestDeliveryTime[k] = 0;
3409 			CursorRequestDeliveryTimePrefetch[k] = 0;
3410 		}
3411 #ifdef __DML_VBA_DEBUG__
3412 		dml_print("DML::%s: k=%u : NumberOfCursors = %u\n", __func__, k, NumberOfCursors[k]);
3413 		dml_print("DML::%s: k=%u : CursorRequestDeliveryTime = %f\n", __func__, k, CursorRequestDeliveryTime[k]);
3414 		dml_print("DML::%s: k=%u : CursorRequestDeliveryTimePrefetch = %f\n", __func__, k, CursorRequestDeliveryTimePrefetch[k]);
3415 #endif
3416 	}
3417 } // CalculatePixelDeliveryTimes
3418 
CalculateMetaAndPTETimes(dml_bool_t use_one_row_for_frame[],dml_uint_t NumberOfActiveSurfaces,dml_bool_t GPUVMEnable,dml_uint_t MetaChunkSize,dml_uint_t MinMetaChunkSizeBytes,dml_uint_t HTotal[],dml_float_t VRatio[],dml_float_t VRatioChroma[],dml_float_t DestinationLinesToRequestRowInVBlank[],dml_float_t DestinationLinesToRequestRowInImmediateFlip[],dml_bool_t DCCEnable[],dml_float_t PixelClock[],dml_uint_t BytePerPixelY[],dml_uint_t BytePerPixelC[],enum dml_rotation_angle SourceScan[],dml_uint_t dpte_row_height[],dml_uint_t dpte_row_height_chroma[],dml_uint_t meta_row_width[],dml_uint_t meta_row_width_chroma[],dml_uint_t meta_row_height[],dml_uint_t meta_row_height_chroma[],dml_uint_t meta_req_width[],dml_uint_t meta_req_width_chroma[],dml_uint_t meta_req_height[],dml_uint_t meta_req_height_chroma[],dml_uint_t dpte_group_bytes[],dml_uint_t PTERequestSizeY[],dml_uint_t PTERequestSizeC[],dml_uint_t PixelPTEReqWidthY[],dml_uint_t PixelPTEReqHeightY[],dml_uint_t PixelPTEReqWidthC[],dml_uint_t PixelPTEReqHeightC[],dml_uint_t dpte_row_width_luma_ub[],dml_uint_t dpte_row_width_chroma_ub[],dml_float_t DST_Y_PER_PTE_ROW_NOM_L[],dml_float_t DST_Y_PER_PTE_ROW_NOM_C[],dml_float_t DST_Y_PER_META_ROW_NOM_L[],dml_float_t DST_Y_PER_META_ROW_NOM_C[],dml_float_t TimePerMetaChunkNominal[],dml_float_t TimePerChromaMetaChunkNominal[],dml_float_t TimePerMetaChunkVBlank[],dml_float_t TimePerChromaMetaChunkVBlank[],dml_float_t TimePerMetaChunkFlip[],dml_float_t TimePerChromaMetaChunkFlip[],dml_float_t time_per_pte_group_nom_luma[],dml_float_t time_per_pte_group_vblank_luma[],dml_float_t time_per_pte_group_flip_luma[],dml_float_t time_per_pte_group_nom_chroma[],dml_float_t time_per_pte_group_vblank_chroma[],dml_float_t time_per_pte_group_flip_chroma[])3419 static void CalculateMetaAndPTETimes(
3420 		dml_bool_t use_one_row_for_frame[],
3421 		dml_uint_t NumberOfActiveSurfaces,
3422 		dml_bool_t GPUVMEnable,
3423 		dml_uint_t MetaChunkSize,
3424 		dml_uint_t MinMetaChunkSizeBytes,
3425 		dml_uint_t HTotal[],
3426 		dml_float_t VRatio[],
3427 		dml_float_t VRatioChroma[],
3428 		dml_float_t DestinationLinesToRequestRowInVBlank[],
3429 		dml_float_t DestinationLinesToRequestRowInImmediateFlip[],
3430 		dml_bool_t DCCEnable[],
3431 		dml_float_t PixelClock[],
3432 		dml_uint_t BytePerPixelY[],
3433 		dml_uint_t BytePerPixelC[],
3434 		enum dml_rotation_angle SourceScan[],
3435 		dml_uint_t dpte_row_height[],
3436 		dml_uint_t dpte_row_height_chroma[],
3437 		dml_uint_t meta_row_width[],
3438 		dml_uint_t meta_row_width_chroma[],
3439 		dml_uint_t meta_row_height[],
3440 		dml_uint_t meta_row_height_chroma[],
3441 		dml_uint_t meta_req_width[],
3442 		dml_uint_t meta_req_width_chroma[],
3443 		dml_uint_t meta_req_height[],
3444 		dml_uint_t meta_req_height_chroma[],
3445 		dml_uint_t dpte_group_bytes[],
3446 		dml_uint_t PTERequestSizeY[],
3447 		dml_uint_t PTERequestSizeC[],
3448 		dml_uint_t PixelPTEReqWidthY[],
3449 		dml_uint_t PixelPTEReqHeightY[],
3450 		dml_uint_t PixelPTEReqWidthC[],
3451 		dml_uint_t PixelPTEReqHeightC[],
3452 		dml_uint_t dpte_row_width_luma_ub[],
3453 		dml_uint_t dpte_row_width_chroma_ub[],
3454 
3455 		// Output
3456 		dml_float_t DST_Y_PER_PTE_ROW_NOM_L[],
3457 		dml_float_t DST_Y_PER_PTE_ROW_NOM_C[],
3458 		dml_float_t DST_Y_PER_META_ROW_NOM_L[],
3459 		dml_float_t DST_Y_PER_META_ROW_NOM_C[],
3460 		dml_float_t TimePerMetaChunkNominal[],
3461 		dml_float_t TimePerChromaMetaChunkNominal[],
3462 		dml_float_t TimePerMetaChunkVBlank[],
3463 		dml_float_t TimePerChromaMetaChunkVBlank[],
3464 		dml_float_t TimePerMetaChunkFlip[],
3465 		dml_float_t TimePerChromaMetaChunkFlip[],
3466 		dml_float_t time_per_pte_group_nom_luma[],
3467 		dml_float_t time_per_pte_group_vblank_luma[],
3468 		dml_float_t time_per_pte_group_flip_luma[],
3469 		dml_float_t time_per_pte_group_nom_chroma[],
3470 		dml_float_t time_per_pte_group_vblank_chroma[],
3471 		dml_float_t time_per_pte_group_flip_chroma[])
3472 {
3473 	dml_uint_t meta_chunk_width;
3474 	dml_uint_t min_meta_chunk_width;
3475 	dml_uint_t meta_chunk_per_row_int;
3476 	dml_uint_t meta_row_remainder;
3477 	dml_uint_t meta_chunk_threshold;
3478 	dml_uint_t meta_chunks_per_row_ub;
3479 	dml_uint_t meta_chunk_width_chroma;
3480 	dml_uint_t min_meta_chunk_width_chroma;
3481 	dml_uint_t meta_chunk_per_row_int_chroma;
3482 	dml_uint_t meta_row_remainder_chroma;
3483 	dml_uint_t meta_chunk_threshold_chroma;
3484 	dml_uint_t meta_chunks_per_row_ub_chroma;
3485 	dml_uint_t dpte_group_width_luma;
3486 	dml_uint_t dpte_groups_per_row_luma_ub;
3487 	dml_uint_t dpte_group_width_chroma;
3488 	dml_uint_t dpte_groups_per_row_chroma_ub;
3489 
3490 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
3491 		DST_Y_PER_PTE_ROW_NOM_L[k] = dpte_row_height[k] / VRatio[k];
3492 		if (BytePerPixelC[k] == 0) {
3493 			DST_Y_PER_PTE_ROW_NOM_C[k] = 0;
3494 		} else {
3495 			DST_Y_PER_PTE_ROW_NOM_C[k] = dpte_row_height_chroma[k] / VRatioChroma[k];
3496 		}
3497 		DST_Y_PER_META_ROW_NOM_L[k] = meta_row_height[k] / VRatio[k];
3498 		if (BytePerPixelC[k] == 0) {
3499 			DST_Y_PER_META_ROW_NOM_C[k] = 0;
3500 		} else {
3501 			DST_Y_PER_META_ROW_NOM_C[k] = meta_row_height_chroma[k] / VRatioChroma[k];
3502 		}
3503 	}
3504 
3505 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
3506 		if (DCCEnable[k] == true) {
3507 			meta_chunk_width = MetaChunkSize * 1024 * 256 / BytePerPixelY[k] / meta_row_height[k];
3508 			min_meta_chunk_width = MinMetaChunkSizeBytes * 256 / BytePerPixelY[k] / meta_row_height[k];
3509 			meta_chunk_per_row_int = meta_row_width[k] / meta_chunk_width;
3510 			meta_row_remainder = meta_row_width[k] % meta_chunk_width;
3511 			if (!dml_is_vertical_rotation(SourceScan[k])) {
3512 				meta_chunk_threshold = 2 * min_meta_chunk_width - meta_req_width[k];
3513 			} else {
3514 				meta_chunk_threshold = 2 * min_meta_chunk_width - meta_req_height[k];
3515 			}
3516 			if (meta_row_remainder <= meta_chunk_threshold) {
3517 				meta_chunks_per_row_ub = meta_chunk_per_row_int + 1;
3518 			} else {
3519 				meta_chunks_per_row_ub = meta_chunk_per_row_int + 2;
3520 			}
3521 			TimePerMetaChunkNominal[k] = meta_row_height[k] / VRatio[k] * HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub;
3522 			TimePerMetaChunkVBlank[k] = DestinationLinesToRequestRowInVBlank[k] * HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub;
3523 			TimePerMetaChunkFlip[k] = DestinationLinesToRequestRowInImmediateFlip[k] * HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub;
3524 			if (BytePerPixelC[k] == 0) {
3525 				TimePerChromaMetaChunkNominal[k] = 0;
3526 				TimePerChromaMetaChunkVBlank[k] = 0;
3527 				TimePerChromaMetaChunkFlip[k] = 0;
3528 			} else {
3529 				meta_chunk_width_chroma = MetaChunkSize * 1024 * 256 / BytePerPixelC[k] / meta_row_height_chroma[k];
3530 				min_meta_chunk_width_chroma = MinMetaChunkSizeBytes * 256 / BytePerPixelC[k] / meta_row_height_chroma[k];
3531 				meta_chunk_per_row_int_chroma = (dml_uint_t)((dml_float_t) meta_row_width_chroma[k] / meta_chunk_width_chroma);
3532 				meta_row_remainder_chroma = meta_row_width_chroma[k] % meta_chunk_width_chroma;
3533 				if (!dml_is_vertical_rotation(SourceScan[k])) {
3534 					meta_chunk_threshold_chroma = 2 * min_meta_chunk_width_chroma - meta_req_width_chroma[k];
3535 				} else {
3536 					meta_chunk_threshold_chroma = 2 * min_meta_chunk_width_chroma - meta_req_height_chroma[k];
3537 				}
3538 				if (meta_row_remainder_chroma <= meta_chunk_threshold_chroma) {
3539 					meta_chunks_per_row_ub_chroma = meta_chunk_per_row_int_chroma + 1;
3540 				} else {
3541 					meta_chunks_per_row_ub_chroma = meta_chunk_per_row_int_chroma + 2;
3542 				}
3543 				TimePerChromaMetaChunkNominal[k] = meta_row_height_chroma[k] / VRatioChroma[k] * HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub_chroma;
3544 				TimePerChromaMetaChunkVBlank[k] = DestinationLinesToRequestRowInVBlank[k] * HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub_chroma;
3545 				TimePerChromaMetaChunkFlip[k] = DestinationLinesToRequestRowInImmediateFlip[k] * HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub_chroma;
3546 			}
3547 		} else {
3548 			TimePerMetaChunkNominal[k] = 0;
3549 			TimePerMetaChunkVBlank[k] = 0;
3550 			TimePerMetaChunkFlip[k] = 0;
3551 			TimePerChromaMetaChunkNominal[k] = 0;
3552 			TimePerChromaMetaChunkVBlank[k] = 0;
3553 			TimePerChromaMetaChunkFlip[k] = 0;
3554 		}
3555 	}
3556 
3557 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
3558 		if (GPUVMEnable == true) {
3559 			if (!dml_is_vertical_rotation(SourceScan[k])) {
3560 				dpte_group_width_luma = (dml_uint_t)((dml_float_t) dpte_group_bytes[k] / (dml_float_t) PTERequestSizeY[k] * PixelPTEReqWidthY[k]);
3561 			} else {
3562 				dpte_group_width_luma = (dml_uint_t)((dml_float_t) dpte_group_bytes[k] / (dml_float_t) PTERequestSizeY[k] * PixelPTEReqHeightY[k]);
3563 			}
3564 
3565 			if (use_one_row_for_frame[k]) {
3566 				dpte_groups_per_row_luma_ub = (dml_uint_t)(dml_ceil((dml_float_t) dpte_row_width_luma_ub[k] / (dml_float_t) dpte_group_width_luma / 2.0, 1.0));
3567 			} else {
3568 				dpte_groups_per_row_luma_ub = (dml_uint_t)(dml_ceil((dml_float_t) dpte_row_width_luma_ub[k] / (dml_float_t) dpte_group_width_luma, 1.0));
3569 			}
3570 
3571 			dml_print("DML::%s: k=%u, use_one_row_for_frame = %u\n", __func__, k, use_one_row_for_frame[k]);
3572 			dml_print("DML::%s: k=%u, dpte_group_bytes = %u\n", __func__, k, dpte_group_bytes[k]);
3573 			dml_print("DML::%s: k=%u, PTERequestSizeY = %u\n", __func__, k, PTERequestSizeY[k]);
3574 			dml_print("DML::%s: k=%u, PixelPTEReqWidthY = %u\n", __func__, k, PixelPTEReqWidthY[k]);
3575 			dml_print("DML::%s: k=%u, PixelPTEReqHeightY = %u\n", __func__, k, PixelPTEReqHeightY[k]);
3576 			dml_print("DML::%s: k=%u, dpte_row_width_luma_ub = %u\n", __func__, k, dpte_row_width_luma_ub[k]);
3577 			dml_print("DML::%s: k=%u, dpte_group_width_luma = %u\n", __func__, k, dpte_group_width_luma);
3578 			dml_print("DML::%s: k=%u, dpte_groups_per_row_luma_ub = %u\n", __func__, k, dpte_groups_per_row_luma_ub);
3579 
3580 			time_per_pte_group_nom_luma[k] = DST_Y_PER_PTE_ROW_NOM_L[k] * HTotal[k] / PixelClock[k] / dpte_groups_per_row_luma_ub;
3581 			time_per_pte_group_vblank_luma[k] = DestinationLinesToRequestRowInVBlank[k] * HTotal[k] / PixelClock[k] / dpte_groups_per_row_luma_ub;
3582 			time_per_pte_group_flip_luma[k] = DestinationLinesToRequestRowInImmediateFlip[k] * HTotal[k] / PixelClock[k] / dpte_groups_per_row_luma_ub;
3583 			if (BytePerPixelC[k] == 0) {
3584 				time_per_pte_group_nom_chroma[k] = 0;
3585 				time_per_pte_group_vblank_chroma[k] = 0;
3586 				time_per_pte_group_flip_chroma[k] = 0;
3587 			} else {
3588 				if (!dml_is_vertical_rotation(SourceScan[k])) {
3589 					dpte_group_width_chroma = (dml_uint_t)((dml_float_t) dpte_group_bytes[k] / (dml_float_t) PTERequestSizeC[k] * PixelPTEReqWidthC[k]);
3590 				} else {
3591 					dpte_group_width_chroma = (dml_uint_t)((dml_float_t) dpte_group_bytes[k] / (dml_float_t) PTERequestSizeC[k] * PixelPTEReqHeightC[k]);
3592 				}
3593 
3594 				if (use_one_row_for_frame[k]) {
3595 					dpte_groups_per_row_chroma_ub = (dml_uint_t)(dml_ceil((dml_float_t) dpte_row_width_chroma_ub[k] / (dml_float_t) dpte_group_width_chroma / 2.0, 1.0));
3596 				} else {
3597 					dpte_groups_per_row_chroma_ub = (dml_uint_t)(dml_ceil((dml_float_t) dpte_row_width_chroma_ub[k] / (dml_float_t) dpte_group_width_chroma, 1.0));
3598 				}
3599 				dml_print("DML::%s: k=%u, dpte_row_width_chroma_ub = %u\n", __func__, k, dpte_row_width_chroma_ub[k]);
3600 				dml_print("DML::%s: k=%u, dpte_group_width_chroma = %u\n", __func__, k, dpte_group_width_chroma);
3601 				dml_print("DML::%s: k=%u, dpte_groups_per_row_chroma_ub = %u\n", __func__, k, dpte_groups_per_row_chroma_ub);
3602 
3603 				time_per_pte_group_nom_chroma[k] = DST_Y_PER_PTE_ROW_NOM_C[k] * HTotal[k] / PixelClock[k] / dpte_groups_per_row_chroma_ub;
3604 				time_per_pte_group_vblank_chroma[k] = DestinationLinesToRequestRowInVBlank[k] * HTotal[k] / PixelClock[k] / dpte_groups_per_row_chroma_ub;
3605 				time_per_pte_group_flip_chroma[k] = DestinationLinesToRequestRowInImmediateFlip[k] * HTotal[k] / PixelClock[k] / dpte_groups_per_row_chroma_ub;
3606 			}
3607 		} else {
3608 			time_per_pte_group_nom_luma[k] = 0;
3609 			time_per_pte_group_vblank_luma[k] = 0;
3610 			time_per_pte_group_flip_luma[k] = 0;
3611 			time_per_pte_group_nom_chroma[k] = 0;
3612 			time_per_pte_group_vblank_chroma[k] = 0;
3613 			time_per_pte_group_flip_chroma[k] = 0;
3614 		}
3615 #ifdef __DML_VBA_DEBUG__
3616 		dml_print("DML::%s: k=%u, DestinationLinesToRequestRowInVBlank = %f\n", __func__, k, DestinationLinesToRequestRowInVBlank[k]);
3617 		dml_print("DML::%s: k=%u, DestinationLinesToRequestRowInImmediateFlip = %f\n", __func__, k, DestinationLinesToRequestRowInImmediateFlip[k]);
3618 
3619 		dml_print("DML::%s: k=%u, DST_Y_PER_PTE_ROW_NOM_L = %f\n", __func__, k, DST_Y_PER_PTE_ROW_NOM_L[k]);
3620 		dml_print("DML::%s: k=%u, DST_Y_PER_PTE_ROW_NOM_C = %f\n", __func__, k, DST_Y_PER_PTE_ROW_NOM_C[k]);
3621 		dml_print("DML::%s: k=%u, DST_Y_PER_META_ROW_NOM_L = %f\n", __func__, k, DST_Y_PER_META_ROW_NOM_L[k]);
3622 		dml_print("DML::%s: k=%u, DST_Y_PER_META_ROW_NOM_C = %f\n", __func__, k, DST_Y_PER_META_ROW_NOM_C[k]);
3623 		dml_print("DML::%s: k=%u, TimePerMetaChunkNominal = %f\n", __func__, k, TimePerMetaChunkNominal[k]);
3624 		dml_print("DML::%s: k=%u, TimePerMetaChunkVBlank = %f\n", __func__, k, TimePerMetaChunkVBlank[k]);
3625 		dml_print("DML::%s: k=%u, TimePerMetaChunkFlip = %f\n", __func__, k, TimePerMetaChunkFlip[k]);
3626 		dml_print("DML::%s: k=%u, TimePerChromaMetaChunkNominal = %f\n", __func__, k, TimePerChromaMetaChunkNominal[k]);
3627 		dml_print("DML::%s: k=%u, TimePerChromaMetaChunkVBlank = %f\n", __func__, k, TimePerChromaMetaChunkVBlank[k]);
3628 		dml_print("DML::%s: k=%u, TimePerChromaMetaChunkFlip = %f\n", __func__, k, TimePerChromaMetaChunkFlip[k]);
3629 		dml_print("DML::%s: k=%u, time_per_pte_group_nom_luma = %f\n", __func__, k, time_per_pte_group_nom_luma[k]);
3630 		dml_print("DML::%s: k=%u, time_per_pte_group_vblank_luma = %f\n", __func__, k, time_per_pte_group_vblank_luma[k]);
3631 		dml_print("DML::%s: k=%u, time_per_pte_group_flip_luma = %f\n", __func__, k, time_per_pte_group_flip_luma[k]);
3632 		dml_print("DML::%s: k=%u, time_per_pte_group_nom_chroma = %f\n", __func__, k, time_per_pte_group_nom_chroma[k]);
3633 		dml_print("DML::%s: k=%u, time_per_pte_group_vblank_chroma = %f\n", __func__, k, time_per_pte_group_vblank_chroma[k]);
3634 		dml_print("DML::%s: k=%u, time_per_pte_group_flip_chroma   = %f\n", __func__, k, time_per_pte_group_flip_chroma[k]);
3635 #endif
3636 	}
3637 } // CalculateMetaAndPTETimes
3638 
CalculateVMGroupAndRequestTimes(dml_uint_t NumberOfActiveSurfaces,dml_bool_t GPUVMEnable,dml_uint_t GPUVMMaxPageTableLevels,dml_uint_t HTotal[],dml_uint_t BytePerPixelC[],dml_float_t DestinationLinesToRequestVMInVBlank[],dml_float_t DestinationLinesToRequestVMInImmediateFlip[],dml_bool_t DCCEnable[],dml_float_t PixelClock[],dml_uint_t dpte_row_width_luma_ub[],dml_uint_t dpte_row_width_chroma_ub[],dml_uint_t vm_group_bytes[],dml_uint_t dpde0_bytes_per_frame_ub_l[],dml_uint_t dpde0_bytes_per_frame_ub_c[],dml_uint_t meta_pte_bytes_per_frame_ub_l[],dml_uint_t meta_pte_bytes_per_frame_ub_c[],dml_float_t TimePerVMGroupVBlank[],dml_float_t TimePerVMGroupFlip[],dml_float_t TimePerVMRequestVBlank[],dml_float_t TimePerVMRequestFlip[])3639 static void CalculateVMGroupAndRequestTimes(
3640 		dml_uint_t NumberOfActiveSurfaces,
3641 		dml_bool_t GPUVMEnable,
3642 		dml_uint_t GPUVMMaxPageTableLevels,
3643 		dml_uint_t HTotal[],
3644 		dml_uint_t BytePerPixelC[],
3645 		dml_float_t DestinationLinesToRequestVMInVBlank[],
3646 		dml_float_t DestinationLinesToRequestVMInImmediateFlip[],
3647 		dml_bool_t DCCEnable[],
3648 		dml_float_t PixelClock[],
3649 		dml_uint_t dpte_row_width_luma_ub[],
3650 		dml_uint_t dpte_row_width_chroma_ub[],
3651 		dml_uint_t vm_group_bytes[],
3652 		dml_uint_t dpde0_bytes_per_frame_ub_l[],
3653 		dml_uint_t dpde0_bytes_per_frame_ub_c[],
3654 		dml_uint_t meta_pte_bytes_per_frame_ub_l[],
3655 		dml_uint_t meta_pte_bytes_per_frame_ub_c[],
3656 
3657 		// Output
3658 		dml_float_t TimePerVMGroupVBlank[],
3659 		dml_float_t TimePerVMGroupFlip[],
3660 		dml_float_t TimePerVMRequestVBlank[],
3661 		dml_float_t TimePerVMRequestFlip[])
3662 {
3663 	dml_uint_t num_group_per_lower_vm_stage;
3664 	dml_uint_t num_req_per_lower_vm_stage;
3665 
3666 #ifdef __DML_VBA_DEBUG__
3667 	dml_print("DML::%s: NumberOfActiveSurfaces = %u\n", __func__, NumberOfActiveSurfaces);
3668 	dml_print("DML::%s: GPUVMEnable = %u\n", __func__, GPUVMEnable);
3669 #endif
3670 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
3671 
3672 #ifdef __DML_VBA_DEBUG__
3673 		dml_print("DML::%s: k=%u, DCCEnable = %u\n", __func__, k, DCCEnable[k]);
3674 		dml_print("DML::%s: k=%u, vm_group_bytes = %u\n", __func__, k, vm_group_bytes[k]);
3675 		dml_print("DML::%s: k=%u, dpde0_bytes_per_frame_ub_l = %u\n", __func__, k, dpde0_bytes_per_frame_ub_l[k]);
3676 		dml_print("DML::%s: k=%u, dpde0_bytes_per_frame_ub_c = %u\n", __func__, k, dpde0_bytes_per_frame_ub_c[k]);
3677 		dml_print("DML::%s: k=%u, meta_pte_bytes_per_frame_ub_l = %u\n", __func__, k, meta_pte_bytes_per_frame_ub_l[k]);
3678 		dml_print("DML::%s: k=%u, meta_pte_bytes_per_frame_ub_c = %u\n", __func__, k, meta_pte_bytes_per_frame_ub_c[k]);
3679 #endif
3680 
3681 		if (GPUVMEnable == true && (DCCEnable[k] == true || GPUVMMaxPageTableLevels > 1)) {
3682 			if (DCCEnable[k] == false) {
3683 				if (BytePerPixelC[k] > 0) {
3684 					num_group_per_lower_vm_stage = (dml_uint_t) (dml_ceil((dml_float_t) dpde0_bytes_per_frame_ub_l[k] / (dml_float_t) vm_group_bytes[k], 1.0) +
3685 													dml_ceil((dml_float_t) dpde0_bytes_per_frame_ub_c[k] / (dml_float_t) vm_group_bytes[k], 1.0));
3686 				} else {
3687 					num_group_per_lower_vm_stage = (dml_uint_t) (dml_ceil((dml_float_t) dpde0_bytes_per_frame_ub_l[k] / (dml_float_t) vm_group_bytes[k], 1.0));
3688 				}
3689 			} else {
3690 				if (GPUVMMaxPageTableLevels == 1) {
3691 					if (BytePerPixelC[k] > 0) {
3692 						num_group_per_lower_vm_stage = (dml_uint_t)(dml_ceil((dml_float_t) (meta_pte_bytes_per_frame_ub_l[k]) / (dml_float_t) (vm_group_bytes[k]), 1.0) +
3693 																	dml_ceil((dml_float_t) (meta_pte_bytes_per_frame_ub_c[k]) / (dml_float_t) (vm_group_bytes[k]), 1.0));
3694 					} else {
3695 						num_group_per_lower_vm_stage = (dml_uint_t)(dml_ceil((dml_float_t) (meta_pte_bytes_per_frame_ub_l[k]) / (dml_float_t) (vm_group_bytes[k]), 1.0));
3696 					}
3697 				} else {
3698 					if (BytePerPixelC[k] > 0) {
3699 						num_group_per_lower_vm_stage = (dml_uint_t)(2.0 + dml_ceil((dml_float_t) (dpde0_bytes_per_frame_ub_l[k]) / (dml_float_t) (vm_group_bytes[k]), 1) +
3700 														dml_ceil((dml_float_t) (dpde0_bytes_per_frame_ub_c[k]) / (dml_float_t) (vm_group_bytes[k]), 1) +
3701 														dml_ceil((dml_float_t) (meta_pte_bytes_per_frame_ub_l[k]) / (dml_float_t) (vm_group_bytes[k]), 1) +
3702 														dml_ceil((dml_float_t) (meta_pte_bytes_per_frame_ub_c[k]) / (dml_float_t) (vm_group_bytes[k]), 1));
3703 					} else {
3704 						num_group_per_lower_vm_stage = (dml_uint_t)(1.0 + dml_ceil((dml_float_t) (dpde0_bytes_per_frame_ub_l[k]) / (dml_float_t) (vm_group_bytes[k]), 1) +
3705 																	dml_ceil((dml_float_t) (meta_pte_bytes_per_frame_ub_l[k]) / (dml_float_t) (vm_group_bytes[k]), 1));
3706 					}
3707 				}
3708 			}
3709 
3710 			if (DCCEnable[k] == false) {
3711 				if (BytePerPixelC[k] > 0) {
3712 					num_req_per_lower_vm_stage = dpde0_bytes_per_frame_ub_l[k] / 64 + dpde0_bytes_per_frame_ub_c[k] / 64;
3713 				} else {
3714 					num_req_per_lower_vm_stage = dpde0_bytes_per_frame_ub_l[k] / 64;
3715 				}
3716 			} else {
3717 				if (GPUVMMaxPageTableLevels == 1) {
3718 					if (BytePerPixelC[k] > 0) {
3719 						num_req_per_lower_vm_stage = meta_pte_bytes_per_frame_ub_l[k] / 64 + meta_pte_bytes_per_frame_ub_c[k] / 64;
3720 					} else {
3721 						num_req_per_lower_vm_stage = meta_pte_bytes_per_frame_ub_l[k] / 64;
3722 					}
3723 				} else {
3724 					if (BytePerPixelC[k] > 0) {
3725 						num_req_per_lower_vm_stage = dpde0_bytes_per_frame_ub_l[k] / 64 + dpde0_bytes_per_frame_ub_c[k] / 64 + meta_pte_bytes_per_frame_ub_l[k] / 64 + meta_pte_bytes_per_frame_ub_c[k] / 64;
3726 					} else {
3727 						num_req_per_lower_vm_stage = dpde0_bytes_per_frame_ub_l[k] / 64 + meta_pte_bytes_per_frame_ub_l[k] / 64;
3728 					}
3729 				}
3730 			}
3731 
3732 			TimePerVMGroupVBlank[k] = DestinationLinesToRequestVMInVBlank[k] * HTotal[k] / PixelClock[k] / num_group_per_lower_vm_stage;
3733 			TimePerVMGroupFlip[k] = DestinationLinesToRequestVMInImmediateFlip[k] * HTotal[k] / PixelClock[k] / num_group_per_lower_vm_stage;
3734 			TimePerVMRequestVBlank[k] = DestinationLinesToRequestVMInVBlank[k] * HTotal[k] / PixelClock[k] / num_req_per_lower_vm_stage;
3735 			TimePerVMRequestFlip[k] = DestinationLinesToRequestVMInImmediateFlip[k] * HTotal[k] / PixelClock[k] / num_req_per_lower_vm_stage;
3736 
3737 			if (GPUVMMaxPageTableLevels > 2) {
3738 				TimePerVMGroupVBlank[k] = TimePerVMGroupVBlank[k] / 2;
3739 				TimePerVMGroupFlip[k] = TimePerVMGroupFlip[k] / 2;
3740 				TimePerVMRequestVBlank[k] = TimePerVMRequestVBlank[k] / 2;
3741 				TimePerVMRequestFlip[k]    = TimePerVMRequestFlip[k] / 2;
3742 			}
3743 
3744 		} else {
3745 			TimePerVMGroupVBlank[k] = 0;
3746 			TimePerVMGroupFlip[k] = 0;
3747 			TimePerVMRequestVBlank[k] = 0;
3748 			TimePerVMRequestFlip[k] = 0;
3749 		}
3750 
3751 #ifdef __DML_VBA_DEBUG__
3752 			dml_print("DML::%s: k=%u, TimePerVMGroupVBlank = %f\n", __func__, k, TimePerVMGroupVBlank[k]);
3753 			dml_print("DML::%s: k=%u, TimePerVMGroupFlip = %f\n", __func__, k, TimePerVMGroupFlip[k]);
3754 			dml_print("DML::%s: k=%u, TimePerVMRequestVBlank = %f\n", __func__, k, TimePerVMRequestVBlank[k]);
3755 			dml_print("DML::%s: k=%u, TimePerVMRequestFlip = %f\n", __func__, k, TimePerVMRequestFlip[k]);
3756 #endif
3757 	}
3758 } // CalculateVMGroupAndRequestTimes
3759 
CalculateStutterEfficiency(struct display_mode_lib_scratch_st * scratch,struct CalculateStutterEfficiency_params_st * p)3760 static void CalculateStutterEfficiency(struct display_mode_lib_scratch_st *scratch,
3761 		struct CalculateStutterEfficiency_params_st *p)
3762 {
3763 	dml_float_t DETBufferingTimeY = 0;
3764 	dml_float_t SwathWidthYCriticalSurface = 0;
3765 	dml_float_t SwathHeightYCriticalSurface = 0;
3766 	dml_float_t VActiveTimeCriticalSurface = 0;
3767 	dml_float_t FrameTimeCriticalSurface = 0;
3768 	dml_uint_t BytePerPixelYCriticalSurface = 0;
3769 	dml_float_t LinesToFinishSwathTransferStutterCriticalSurface = 0;
3770 	dml_uint_t DETBufferSizeYCriticalSurface = 0;
3771 	dml_float_t MinTTUVBlankCriticalSurface = 0;
3772 	dml_uint_t BlockWidth256BytesYCriticalSurface = 0;
3773 	dml_bool_t SinglePlaneCriticalSurface = 0;
3774 	dml_bool_t SinglePipeCriticalSurface = 0;
3775 	dml_float_t TotalCompressedReadBandwidth = 0;
3776 	dml_float_t TotalRowReadBandwidth = 0;
3777 	dml_float_t AverageDCCCompressionRate = 0;
3778 	dml_float_t EffectiveCompressedBufferSize = 0;
3779 	dml_float_t PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer = 0;
3780 	dml_float_t StutterBurstTime = 0;
3781 	dml_uint_t TotalActiveWriteback = 0;
3782 	dml_float_t LinesInDETY = 0;
3783 	dml_float_t LinesInDETYRoundedDownToSwath = 0;
3784 	dml_float_t MaximumEffectiveCompressionLuma = 0;
3785 	dml_float_t MaximumEffectiveCompressionChroma = 0;
3786 	dml_float_t TotalZeroSizeRequestReadBandwidth = 0;
3787 	dml_float_t TotalZeroSizeCompressedReadBandwidth = 0;
3788 	dml_float_t AverageDCCZeroSizeFraction = 0;
3789 	dml_float_t AverageZeroSizeCompressionRate = 0;
3790 
3791 	dml_bool_t FoundCriticalSurface = false;
3792 
3793 	dml_uint_t TotalNumberOfActiveOTG = 0;
3794 	dml_float_t SinglePixelClock = 0;
3795 	dml_uint_t SingleHTotal = 0;
3796 	dml_uint_t SingleVTotal = 0;
3797 	dml_bool_t SameTiming = true;
3798 
3799 	dml_float_t LastStutterPeriod = 0.0;
3800 	dml_float_t LastZ8StutterPeriod = 0.0;
3801 
3802 	dml_uint_t SwathSizeCriticalSurface;
3803 	dml_uint_t LastChunkOfSwathSize;
3804 	dml_uint_t MissingPartOfLastSwathOfDETSize;
3805 
3806 	TotalZeroSizeRequestReadBandwidth = 0;
3807 	TotalZeroSizeCompressedReadBandwidth = 0;
3808 	TotalRowReadBandwidth = 0;
3809 	TotalCompressedReadBandwidth = 0;
3810 
3811 	for (dml_uint_t k = 0; k < p->NumberOfActiveSurfaces; ++k) {
3812 	if (p->UseMALLForPStateChange[k] != dml_use_mall_pstate_change_phantom_pipe) {
3813 		if (p->DCCEnable[k] == true) {
3814 			if ((dml_is_vertical_rotation(p->SourceScan[k]) && p->BlockWidth256BytesY[k] > p->SwathHeightY[k]) || (!dml_is_vertical_rotation(p->SourceScan[k]) && p->BlockHeight256BytesY[k] > p->SwathHeightY[k]) || p->DCCYMaxUncompressedBlock[k] < 256) {
3815 				MaximumEffectiveCompressionLuma = 2;
3816 			} else {
3817 				MaximumEffectiveCompressionLuma = 4;
3818 			}
3819 			TotalCompressedReadBandwidth = TotalCompressedReadBandwidth + p->ReadBandwidthSurfaceLuma[k] / dml_min(p->NetDCCRateLuma[k], MaximumEffectiveCompressionLuma);
3820 #ifdef __DML_VBA_DEBUG__
3821 			dml_print("DML::%s: k=%u, ReadBandwidthSurfaceLuma = %f\n", __func__, k, p->ReadBandwidthSurfaceLuma[k]);
3822 			dml_print("DML::%s: k=%u, NetDCCRateLuma = %f\n", __func__, k, p->NetDCCRateLuma[k]);
3823 			dml_print("DML::%s: k=%u, MaximumEffectiveCompressionLuma = %f\n", __func__, k, MaximumEffectiveCompressionLuma);
3824 #endif
3825 			TotalZeroSizeRequestReadBandwidth = TotalZeroSizeRequestReadBandwidth + p->ReadBandwidthSurfaceLuma[k] * p->DCCFractionOfZeroSizeRequestsLuma[k];
3826 			TotalZeroSizeCompressedReadBandwidth = TotalZeroSizeCompressedReadBandwidth + p->ReadBandwidthSurfaceLuma[k] * p->DCCFractionOfZeroSizeRequestsLuma[k] / MaximumEffectiveCompressionLuma;
3827 
3828 			if (p->ReadBandwidthSurfaceChroma[k] > 0) {
3829 				if ((dml_is_vertical_rotation(p->SourceScan[k]) && p->BlockWidth256BytesC[k] > p->SwathHeightC[k]) || (!dml_is_vertical_rotation(p->SourceScan[k]) && p->BlockHeight256BytesC[k] > p->SwathHeightC[k]) || p->DCCCMaxUncompressedBlock[k] < 256) {
3830 					MaximumEffectiveCompressionChroma = 2;
3831 				} else {
3832 					MaximumEffectiveCompressionChroma = 4;
3833 				}
3834 				TotalCompressedReadBandwidth = TotalCompressedReadBandwidth + p->ReadBandwidthSurfaceChroma[k] / dml_min(p->NetDCCRateChroma[k], MaximumEffectiveCompressionChroma);
3835 #ifdef __DML_VBA_DEBUG__
3836 				dml_print("DML::%s: k=%u, ReadBandwidthSurfaceChroma = %f\n", __func__, k, p->ReadBandwidthSurfaceChroma[k]);
3837 				dml_print("DML::%s: k=%u, NetDCCRateChroma = %f\n", __func__, k, p->NetDCCRateChroma[k]);
3838 				dml_print("DML::%s: k=%u, MaximumEffectiveCompressionChroma = %f\n", __func__, k, MaximumEffectiveCompressionChroma);
3839 #endif
3840 				TotalZeroSizeRequestReadBandwidth = TotalZeroSizeRequestReadBandwidth + p->ReadBandwidthSurfaceChroma[k] * p->DCCFractionOfZeroSizeRequestsChroma[k];
3841 				TotalZeroSizeCompressedReadBandwidth = TotalZeroSizeCompressedReadBandwidth + p->ReadBandwidthSurfaceChroma[k] * p->DCCFractionOfZeroSizeRequestsChroma[k] / MaximumEffectiveCompressionChroma;
3842 			}
3843 		} else {
3844 			TotalCompressedReadBandwidth = TotalCompressedReadBandwidth + p->ReadBandwidthSurfaceLuma[k] + p->ReadBandwidthSurfaceChroma[k];
3845 		}
3846 		TotalRowReadBandwidth = TotalRowReadBandwidth + p->DPPPerSurface[k] * (p->meta_row_bw[k] + p->dpte_row_bw[k]);
3847 	}
3848 	}
3849 
3850 	AverageDCCCompressionRate = p->TotalDataReadBandwidth / TotalCompressedReadBandwidth;
3851 	AverageDCCZeroSizeFraction = TotalZeroSizeRequestReadBandwidth / p->TotalDataReadBandwidth;
3852 
3853 #ifdef __DML_VBA_DEBUG__
3854 	dml_print("DML::%s: UnboundedRequestEnabled = %u\n", __func__, p->UnboundedRequestEnabled);
3855 	dml_print("DML::%s: TotalCompressedReadBandwidth = %f\n", __func__, TotalCompressedReadBandwidth);
3856 	dml_print("DML::%s: TotalZeroSizeRequestReadBandwidth = %f\n", __func__, TotalZeroSizeRequestReadBandwidth);
3857 	dml_print("DML::%s: TotalZeroSizeCompressedReadBandwidth = %f\n", __func__, TotalZeroSizeCompressedReadBandwidth);
3858 	dml_print("DML::%s: MaximumEffectiveCompressionLuma = %f\n", __func__, MaximumEffectiveCompressionLuma);
3859 	dml_print("DML::%s: MaximumEffectiveCompressionChroma = %f\n", __func__, MaximumEffectiveCompressionChroma);
3860 	dml_print("DML::%s: AverageDCCCompressionRate = %f\n", __func__, AverageDCCCompressionRate);
3861 	dml_print("DML::%s: AverageDCCZeroSizeFraction = %f\n", __func__, AverageDCCZeroSizeFraction);
3862 	dml_print("DML::%s: CompbufReservedSpace64B = %u\n", __func__, p->CompbufReservedSpace64B);
3863 	dml_print("DML::%s: CompbufReservedSpaceZs = %u\n", __func__, p->CompbufReservedSpaceZs);
3864 	dml_print("DML::%s: CompressedBufferSizeInkByte = %u\n", __func__, p->CompressedBufferSizeInkByte);
3865 #endif
3866 	if (AverageDCCZeroSizeFraction == 1) {
3867 		AverageZeroSizeCompressionRate = TotalZeroSizeRequestReadBandwidth / TotalZeroSizeCompressedReadBandwidth;
3868 		EffectiveCompressedBufferSize = (dml_float_t)p->MetaFIFOSizeInKEntries * 1024 * 64 * AverageZeroSizeCompressionRate + ((dml_float_t)p->ZeroSizeBufferEntries - p->CompbufReservedSpaceZs) * 64 * AverageZeroSizeCompressionRate;
3869 	} 	else if (AverageDCCZeroSizeFraction > 0) {
3870 		AverageZeroSizeCompressionRate = TotalZeroSizeRequestReadBandwidth / TotalZeroSizeCompressedReadBandwidth;
3871 		EffectiveCompressedBufferSize = dml_min((dml_float_t)p->CompressedBufferSizeInkByte * 1024 * AverageDCCCompressionRate,
3872 											(dml_float_t)p->MetaFIFOSizeInKEntries * 1024 * 64 / (AverageDCCZeroSizeFraction / AverageZeroSizeCompressionRate + 1 / AverageDCCCompressionRate)) +
3873 										dml_min(((dml_float_t)p->ROBBufferSizeInKByte * 1024 - p->CompbufReservedSpace64B * 64) * AverageDCCCompressionRate,
3874 											((dml_float_t)p->ZeroSizeBufferEntries - p->CompbufReservedSpaceZs) * 64 / (AverageDCCZeroSizeFraction / AverageZeroSizeCompressionRate));
3875 
3876 #ifdef __DML_VBA_DEBUG__
3877 		dml_print("DML::%s: min 1 = %f\n", __func__, p->CompressedBufferSizeInkByte * 1024 * AverageDCCCompressionRate);
3878 		dml_print("DML::%s: min 2 = %f\n", __func__, p->MetaFIFOSizeInKEntries * 1024 * 64 / (AverageDCCZeroSizeFraction / AverageZeroSizeCompressionRate + 1 / AverageDCCCompressionRate));
3879 		dml_print("DML::%s: min 3 = %f\n", __func__, (p->ROBBufferSizeInKByte * 1024 - p->CompbufReservedSpace64B * 64) * AverageDCCCompressionRate);
3880 		dml_print("DML::%s: min 4 = %f\n", __func__, (p->ZeroSizeBufferEntries - p->CompbufReservedSpaceZs) * 64 / (AverageDCCZeroSizeFraction / AverageZeroSizeCompressionRate));
3881 #endif
3882 	} else {
3883 		EffectiveCompressedBufferSize = dml_min((dml_float_t)p->CompressedBufferSizeInkByte * 1024 * AverageDCCCompressionRate,
3884 												(dml_float_t)p->MetaFIFOSizeInKEntries * 1024 * 64 * AverageDCCCompressionRate) +
3885 												((dml_float_t)p->ROBBufferSizeInKByte * 1024 - p->CompbufReservedSpace64B * 64) * AverageDCCCompressionRate;
3886 
3887 #ifdef __DML_VBA_DEBUG__
3888 		dml_print("DML::%s: min 1 = %f\n", __func__, p->CompressedBufferSizeInkByte * 1024 * AverageDCCCompressionRate);
3889 		dml_print("DML::%s: min 2 = %f\n", __func__, p->MetaFIFOSizeInKEntries * 1024 * 64 * AverageDCCCompressionRate);
3890 #endif
3891 	}
3892 
3893 #ifdef __DML_VBA_DEBUG__
3894 	dml_print("DML::%s: MetaFIFOSizeInKEntries = %u\n", __func__, p->MetaFIFOSizeInKEntries);
3895 	dml_print("DML::%s: AverageZeroSizeCompressionRate = %f\n", __func__, AverageZeroSizeCompressionRate);
3896 	dml_print("DML::%s: EffectiveCompressedBufferSize = %f\n", __func__, EffectiveCompressedBufferSize);
3897 #endif
3898 
3899 	*p->StutterPeriod = 0;
3900 
3901 	for (dml_uint_t k = 0; k < p->NumberOfActiveSurfaces; ++k) {
3902 	if (p->UseMALLForPStateChange[k] != dml_use_mall_pstate_change_phantom_pipe) {
3903 		LinesInDETY = ((dml_float_t)p->DETBufferSizeY[k] + (p->UnboundedRequestEnabled == true ? EffectiveCompressedBufferSize : 0) * p->ReadBandwidthSurfaceLuma[k] / p->TotalDataReadBandwidth) / p->BytePerPixelDETY[k] / p->SwathWidthY[k];
3904 		LinesInDETYRoundedDownToSwath = dml_floor(LinesInDETY, p->SwathHeightY[k]);
3905 		DETBufferingTimeY = LinesInDETYRoundedDownToSwath * ((dml_float_t)p->HTotal[k] / p->PixelClock[k]) / p->VRatio[k];
3906 #ifdef __DML_VBA_DEBUG__
3907 		dml_print("DML::%s: k=%u, DETBufferSizeY = %u\n", __func__, k, p->DETBufferSizeY[k]);
3908 		dml_print("DML::%s: k=%u, BytePerPixelDETY = %f\n", __func__, k, p->BytePerPixelDETY[k]);
3909 		dml_print("DML::%s: k=%u, SwathWidthY = %u\n", __func__, k, p->SwathWidthY[k]);
3910 		dml_print("DML::%s: k=%u, ReadBandwidthSurfaceLuma = %f\n", __func__, k, p->ReadBandwidthSurfaceLuma[k]);
3911 		dml_print("DML::%s: k=%u, TotalDataReadBandwidth = %f\n", __func__, k, p->TotalDataReadBandwidth);
3912 		dml_print("DML::%s: k=%u, LinesInDETY = %f\n", __func__, k, LinesInDETY);
3913 		dml_print("DML::%s: k=%u, LinesInDETYRoundedDownToSwath = %f\n", __func__, k, LinesInDETYRoundedDownToSwath);
3914 		dml_print("DML::%s: k=%u, HTotal = %u\n", __func__, k, p->HTotal[k]);
3915 		dml_print("DML::%s: k=%u, PixelClock = %f\n", __func__, k, p->PixelClock[k]);
3916 		dml_print("DML::%s: k=%u, VRatio = %f\n", __func__, k, p->VRatio[k]);
3917 		dml_print("DML::%s: k=%u, DETBufferingTimeY = %f\n", __func__, k, DETBufferingTimeY);
3918 		dml_print("DML::%s: k=%u,PixelClock = %f\n", __func__, k, p->PixelClock[k]);
3919 #endif
3920 
3921 		if (!FoundCriticalSurface || DETBufferingTimeY < *p->StutterPeriod) {
3922 			dml_bool_t isInterlaceTiming = p->Interlace[k] && !p->ProgressiveToInterlaceUnitInOPP;
3923 
3924 			FoundCriticalSurface = true;
3925 			*p->StutterPeriod = DETBufferingTimeY;
3926 			FrameTimeCriticalSurface = (isInterlaceTiming ? dml_floor((dml_float_t)p->VTotal[k]/2.0, 1.0) : p->VTotal[k]) * (dml_float_t)p->HTotal[k] / p->PixelClock[k];
3927 			VActiveTimeCriticalSurface = (isInterlaceTiming ? dml_floor((dml_float_t)p->VActive[k]/2.0, 1.0) : p->VActive[k]) * (dml_float_t)p->HTotal[k] / p->PixelClock[k];
3928 			BytePerPixelYCriticalSurface = p->BytePerPixelY[k];
3929 			SwathWidthYCriticalSurface = p->SwathWidthY[k];
3930 			SwathHeightYCriticalSurface = p->SwathHeightY[k];
3931 			BlockWidth256BytesYCriticalSurface = p->BlockWidth256BytesY[k];
3932 			LinesToFinishSwathTransferStutterCriticalSurface = p->SwathHeightY[k] - (LinesInDETY - LinesInDETYRoundedDownToSwath);
3933 			DETBufferSizeYCriticalSurface = p->DETBufferSizeY[k];
3934 			MinTTUVBlankCriticalSurface = p->MinTTUVBlank[k];
3935 			SinglePlaneCriticalSurface = (p->ReadBandwidthSurfaceChroma[k] == 0);
3936 			SinglePipeCriticalSurface = (p->DPPPerSurface[k] == 1);
3937 
3938 #ifdef __DML_VBA_DEBUG__
3939 			dml_print("DML::%s: k=%u, FoundCriticalSurface = %u\n", __func__, k, FoundCriticalSurface);
3940 			dml_print("DML::%s: k=%u, StutterPeriod = %f\n", __func__, k, *p->StutterPeriod);
3941 			dml_print("DML::%s: k=%u, MinTTUVBlankCriticalSurface = %f\n", __func__, k, MinTTUVBlankCriticalSurface);
3942 			dml_print("DML::%s: k=%u, FrameTimeCriticalSurface = %f\n", __func__, k, FrameTimeCriticalSurface);
3943 			dml_print("DML::%s: k=%u, VActiveTimeCriticalSurface = %f\n", __func__, k, VActiveTimeCriticalSurface);
3944 			dml_print("DML::%s: k=%u, BytePerPixelYCriticalSurface = %u\n", __func__, k, BytePerPixelYCriticalSurface);
3945 			dml_print("DML::%s: k=%u, SwathWidthYCriticalSurface = %f\n", __func__, k, SwathWidthYCriticalSurface);
3946 			dml_print("DML::%s: k=%u, SwathHeightYCriticalSurface = %f\n", __func__, k, SwathHeightYCriticalSurface);
3947 			dml_print("DML::%s: k=%u, BlockWidth256BytesYCriticalSurface = %u\n", __func__, k, BlockWidth256BytesYCriticalSurface);
3948 			dml_print("DML::%s: k=%u, SinglePlaneCriticalSurface = %u\n", __func__, k, SinglePlaneCriticalSurface);
3949 			dml_print("DML::%s: k=%u, SinglePipeCriticalSurface = %u\n", __func__, k, SinglePipeCriticalSurface);
3950 			dml_print("DML::%s: k=%u, LinesToFinishSwathTransferStutterCriticalSurface = %f\n", __func__, k, LinesToFinishSwathTransferStutterCriticalSurface);
3951 #endif
3952 		}
3953 	}
3954 	}
3955 
3956 	PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer = dml_min(*p->StutterPeriod * p->TotalDataReadBandwidth, EffectiveCompressedBufferSize);
3957 #ifdef __DML_VBA_DEBUG__
3958 	dml_print("DML::%s: ROBBufferSizeInKByte = %u\n", __func__, p->ROBBufferSizeInKByte);
3959 	dml_print("DML::%s: AverageDCCCompressionRate = %f\n", __func__, AverageDCCCompressionRate);
3960 	dml_print("DML::%s: StutterPeriod * TotalDataReadBandwidth = %f\n", __func__, *p->StutterPeriod * p->TotalDataReadBandwidth);
3961 	dml_print("DML::%s: ROBBufferSizeInKByte * 1024 * AverageDCCCompressionRate + EffectiveCompressedBufferSize = %f\n", __func__, p->ROBBufferSizeInKByte * 1024 * AverageDCCCompressionRate + EffectiveCompressedBufferSize);
3962 	dml_print("DML::%s: EffectiveCompressedBufferSize = %f\n", __func__, EffectiveCompressedBufferSize);
3963 	dml_print("DML::%s: PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer = %f\n", __func__, PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer);
3964 	dml_print("DML::%s: ReturnBW = %f\n", __func__, p->ReturnBW);
3965 	dml_print("DML::%s: TotalDataReadBandwidth = %f\n", __func__, p->TotalDataReadBandwidth);
3966 	dml_print("DML::%s: TotalRowReadBandwidth = %f\n", __func__, TotalRowReadBandwidth);
3967 	dml_print("DML::%s: DCFCLK = %f\n", __func__, p->DCFCLK);
3968 #endif
3969 
3970 	StutterBurstTime = PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer / AverageDCCCompressionRate / p->ReturnBW + (*p->StutterPeriod * p->TotalDataReadBandwidth - PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer) / (p->DCFCLK * 64) + *p->StutterPeriod * TotalRowReadBandwidth / p->ReturnBW;
3971 #ifdef __DML_VBA_DEBUG__
3972 	dml_print("DML::%s: Part 1 = %f\n", __func__, PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer / AverageDCCCompressionRate / p->ReturnBW);
3973 	dml_print("DML::%s: StutterPeriod * TotalDataReadBandwidth = %f\n", __func__, (*p->StutterPeriod * p->TotalDataReadBandwidth));
3974 	dml_print("DML::%s: Part 2 = %f\n", __func__, (*p->StutterPeriod * p->TotalDataReadBandwidth - PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer) / (p->DCFCLK * 64));
3975 	dml_print("DML::%s: Part 3 = %f\n", __func__, *p->StutterPeriod * TotalRowReadBandwidth / p->ReturnBW);
3976 	dml_print("DML::%s: StutterBurstTime = %f\n", __func__, StutterBurstTime);
3977 #endif
3978 	StutterBurstTime = dml_max(StutterBurstTime, LinesToFinishSwathTransferStutterCriticalSurface * BytePerPixelYCriticalSurface * SwathWidthYCriticalSurface / p->ReturnBW);
3979 
3980 	dml_print("DML::%s: Time to finish residue swath=%f\n", __func__, LinesToFinishSwathTransferStutterCriticalSurface * BytePerPixelYCriticalSurface * SwathWidthYCriticalSurface / p->ReturnBW);
3981 
3982 	TotalActiveWriteback = 0;
3983 	for (dml_uint_t k = 0; k < p->NumberOfActiveSurfaces; ++k) {
3984 		if (p->WritebackEnable[k]) {
3985 			TotalActiveWriteback = TotalActiveWriteback + 1;
3986 		}
3987 	}
3988 
3989 	if (TotalActiveWriteback == 0) {
3990 #ifdef __DML_VBA_DEBUG__
3991 		dml_print("DML::%s: SRExitTime = %f\n", __func__, p->SRExitTime);
3992 		dml_print("DML::%s: SRExitZ8Time = %f\n", __func__, p->SRExitZ8Time);
3993 		dml_print("DML::%s: StutterBurstTime = %f (final)\n", __func__, StutterBurstTime);
3994 		dml_print("DML::%s: StutterPeriod = %f\n", __func__, *p->StutterPeriod);
3995 #endif
3996 		*p->StutterEfficiencyNotIncludingVBlank = dml_max(0., 1 - (p->SRExitTime + StutterBurstTime) / *p->StutterPeriod) * 100;
3997 		*p->Z8StutterEfficiencyNotIncludingVBlank = dml_max(0., 1 - (p->SRExitZ8Time + StutterBurstTime) / *p->StutterPeriod) * 100;
3998 		*p->NumberOfStutterBurstsPerFrame = (*p->StutterEfficiencyNotIncludingVBlank > 0 ? (dml_uint_t)(dml_ceil(VActiveTimeCriticalSurface / *p->StutterPeriod, 1)) : 0);
3999 		*p->Z8NumberOfStutterBurstsPerFrame = (*p->Z8StutterEfficiencyNotIncludingVBlank > 0 ? (dml_uint_t)(dml_ceil(VActiveTimeCriticalSurface / *p->StutterPeriod, 1)) : 0);
4000 	} else {
4001 		*p->StutterEfficiencyNotIncludingVBlank = 0.;
4002 		*p->Z8StutterEfficiencyNotIncludingVBlank = 0.;
4003 		*p->NumberOfStutterBurstsPerFrame = 0;
4004 		*p->Z8NumberOfStutterBurstsPerFrame = 0;
4005 	}
4006 #ifdef __DML_VBA_DEBUG__
4007 	dml_print("DML::%s: VActiveTimeCriticalSurface = %f\n", __func__, VActiveTimeCriticalSurface);
4008 	dml_print("DML::%s: StutterEfficiencyNotIncludingVBlank = %f\n", __func__, *p->StutterEfficiencyNotIncludingVBlank);
4009 	dml_print("DML::%s: Z8StutterEfficiencyNotIncludingVBlank = %f\n", __func__, *p->Z8StutterEfficiencyNotIncludingVBlank);
4010 	dml_print("DML::%s: NumberOfStutterBurstsPerFrame = %u\n", __func__, *p->NumberOfStutterBurstsPerFrame);
4011 	dml_print("DML::%s: Z8NumberOfStutterBurstsPerFrame = %u\n", __func__, *p->Z8NumberOfStutterBurstsPerFrame);
4012 #endif
4013 
4014 	for (dml_uint_t k = 0; k < p->NumberOfActiveSurfaces; ++k) {
4015 		if (p->UseMALLForPStateChange[k] != dml_use_mall_pstate_change_phantom_pipe) {
4016 			if (p->BlendingAndTiming[k] == k) {
4017 				if (TotalNumberOfActiveOTG == 0) {
4018 					SinglePixelClock = p->PixelClock[k];
4019 					SingleHTotal = p->HTotal[k];
4020 					SingleVTotal = p->VTotal[k];
4021 				} else if (SinglePixelClock != p->PixelClock[k] || SingleHTotal != p->HTotal[k] || SingleVTotal != p->VTotal[k]) {
4022 					SameTiming = false;
4023 				}
4024 				TotalNumberOfActiveOTG = TotalNumberOfActiveOTG + 1;
4025 			}
4026 		}
4027 	}
4028 
4029 	if (*p->StutterEfficiencyNotIncludingVBlank > 0) {
4030 		LastStutterPeriod = VActiveTimeCriticalSurface - (*p->NumberOfStutterBurstsPerFrame - 1) * *p->StutterPeriod;
4031 
4032 		if ((p->SynchronizeTimingsFinal || TotalNumberOfActiveOTG == 1) && SameTiming &&
4033 			LastStutterPeriod + MinTTUVBlankCriticalSurface > p->StutterEnterPlusExitWatermark) {
4034 			*p->StutterEfficiency = (1 - (*p->NumberOfStutterBurstsPerFrame * p->SRExitTime + StutterBurstTime * VActiveTimeCriticalSurface / *p->StutterPeriod) / FrameTimeCriticalSurface) * 100;
4035 		} else {
4036 			*p->StutterEfficiency = *p->StutterEfficiencyNotIncludingVBlank;
4037 		}
4038 	} else {
4039 		*p->StutterEfficiency = 0;
4040 	}
4041 
4042 	if (*p->Z8StutterEfficiencyNotIncludingVBlank > 0) {
4043 		LastZ8StutterPeriod = VActiveTimeCriticalSurface - (*p->NumberOfStutterBurstsPerFrame - 1) * *p->StutterPeriod;
4044 		if ((p->SynchronizeTimingsFinal || TotalNumberOfActiveOTG == 1) && SameTiming && LastZ8StutterPeriod + MinTTUVBlankCriticalSurface > p->Z8StutterEnterPlusExitWatermark) {
4045 			*p->Z8StutterEfficiency = (1 - (*p->NumberOfStutterBurstsPerFrame * p->SRExitZ8Time + StutterBurstTime * VActiveTimeCriticalSurface / *p->StutterPeriod) / FrameTimeCriticalSurface) * 100;
4046 		} else {
4047 			*p->Z8StutterEfficiency = *p->Z8StutterEfficiencyNotIncludingVBlank;
4048 		}
4049 	} else {
4050 		*p->Z8StutterEfficiency = 0.;
4051 	}
4052 
4053 #ifdef __DML_VBA_DEBUG__
4054 		dml_print("DML::%s: LastZ8StutterPeriod = %f\n", __func__, LastZ8StutterPeriod);
4055 		dml_print("DML::%s: Z8StutterEnterPlusExitWatermark = %f\n", __func__, p->Z8StutterEnterPlusExitWatermark);
4056 		dml_print("DML::%s: StutterBurstTime = %f\n", __func__, StutterBurstTime);
4057 		dml_print("DML::%s: StutterPeriod = %f\n", __func__, *p->StutterPeriod);
4058 		dml_print("DML::%s: StutterEfficiency = %f\n", __func__, *p->StutterEfficiency);
4059 		dml_print("DML::%s: Z8StutterEfficiency = %f\n", __func__, *p->Z8StutterEfficiency);
4060 		dml_print("DML::%s: StutterEfficiencyNotIncludingVBlank = %f\n", __func__, *p->StutterEfficiencyNotIncludingVBlank);
4061 		dml_print("DML::%s: Z8NumberOfStutterBurstsPerFrame = %u\n", __func__, *p->Z8NumberOfStutterBurstsPerFrame);
4062 #endif
4063 
4064 		SwathSizeCriticalSurface = (dml_uint_t)(BytePerPixelYCriticalSurface * SwathHeightYCriticalSurface * dml_ceil(SwathWidthYCriticalSurface, BlockWidth256BytesYCriticalSurface));
4065 		LastChunkOfSwathSize = SwathSizeCriticalSurface % (p->PixelChunkSizeInKByte * 1024);
4066 		MissingPartOfLastSwathOfDETSize = (dml_uint_t)(dml_ceil(DETBufferSizeYCriticalSurface, SwathSizeCriticalSurface) - DETBufferSizeYCriticalSurface);
4067 
4068 		*p->DCHUBBUB_ARB_CSTATE_MAX_CAP_MODE = !(!p->UnboundedRequestEnabled && (p->NumberOfActiveSurfaces == 1) && SinglePlaneCriticalSurface && SinglePipeCriticalSurface && (LastChunkOfSwathSize > 0) &&
4069 			(LastChunkOfSwathSize <= 4096) && (MissingPartOfLastSwathOfDETSize > 0) && (MissingPartOfLastSwathOfDETSize <= LastChunkOfSwathSize));
4070 
4071 #ifdef __DML_VBA_DEBUG__
4072 	dml_print("DML::%s: SwathSizeCriticalSurface = %u\n", __func__, SwathSizeCriticalSurface);
4073 	dml_print("DML::%s: DETBufferSizeYCriticalSurface = %u\n", __func__, DETBufferSizeYCriticalSurface);
4074 	dml_print("DML::%s: PixelChunkSizeInKByte = %u\n", __func__, p->PixelChunkSizeInKByte);
4075 	dml_print("DML::%s: LastChunkOfSwathSize = %u\n", __func__, LastChunkOfSwathSize);
4076 	dml_print("DML::%s: MissingPartOfLastSwathOfDETSize = %u\n", __func__, MissingPartOfLastSwathOfDETSize);
4077 	dml_print("DML::%s: DCHUBBUB_ARB_CSTATE_MAX_CAP_MODE = %u\n", __func__, *p->DCHUBBUB_ARB_CSTATE_MAX_CAP_MODE);
4078 #endif
4079 } // CalculateStutterEfficiency
4080 
4081 /// \CalculateSwathAndDETConfiguration
4082 /// @brief Calculates swath width and different return buffers sizing (DET, CDB, etc.)
CalculateSwathAndDETConfiguration(struct display_mode_lib_scratch_st * scratch,struct CalculateSwathAndDETConfiguration_params_st * p)4083 static void CalculateSwathAndDETConfiguration(struct display_mode_lib_scratch_st *scratch,
4084 	struct CalculateSwathAndDETConfiguration_params_st *p)
4085 {
4086 	dml_uint_t MaximumSwathHeightY[__DML_NUM_PLANES__];
4087 	dml_uint_t MaximumSwathHeightC[__DML_NUM_PLANES__];
4088 	dml_uint_t RoundedUpMaxSwathSizeBytesY[__DML_NUM_PLANES__];
4089 	dml_uint_t RoundedUpMaxSwathSizeBytesC[__DML_NUM_PLANES__];
4090 	dml_uint_t RoundedUpSwathSizeBytesY[__DML_NUM_PLANES__];
4091 	dml_uint_t RoundedUpSwathSizeBytesC[__DML_NUM_PLANES__];
4092 	dml_uint_t SwathWidthSingleDPP[__DML_NUM_PLANES__];
4093 	dml_uint_t SwathWidthSingleDPPChroma[__DML_NUM_PLANES__];
4094 
4095 	dml_uint_t TotalActiveDPP = 0;
4096 	dml_bool_t NoChromaOrLinearSurfaces = true;
4097 	dml_uint_t SurfaceDoingUnboundedRequest = 0;
4098 
4099 	dml_uint_t DETBufferSizeInKByteForSwathCalculation;
4100 
4101 	const long TTUFIFODEPTH = 8;
4102 	const long MAXIMUMCOMPRESSION = 4;
4103 
4104 #ifdef __DML_VBA_DEBUG__
4105 	dml_print("DML::%s: ForceSingleDPP = %u\n", __func__, p->ForceSingleDPP);
4106 	for (dml_uint_t k = 0; k < p->NumberOfActiveSurfaces; ++k) {
4107 		dml_print("DML::%s: DPPPerSurface[%u] = %u\n", __func__, k, p->DPPPerSurface[k]);
4108 	}
4109 #endif
4110 	CalculateSwathWidth(p->ForceSingleDPP,
4111 		p->NumberOfActiveSurfaces,
4112 		p->SourcePixelFormat,
4113 		p->SourceScan,
4114 		p->ViewportStationary,
4115 		p->ViewportWidth,
4116 		p->ViewportHeight,
4117 		p->ViewportXStart,
4118 		p->ViewportYStart,
4119 		p->ViewportXStartC,
4120 		p->ViewportYStartC,
4121 		p->SurfaceWidthY,
4122 		p->SurfaceWidthC,
4123 		p->SurfaceHeightY,
4124 		p->SurfaceHeightC,
4125 		p->ODMMode,
4126 		p->BytePerPixY,
4127 		p->BytePerPixC,
4128 		p->Read256BytesBlockHeightY,
4129 		p->Read256BytesBlockHeightC,
4130 		p->Read256BytesBlockWidthY,
4131 		p->Read256BytesBlockWidthC,
4132 		p->BlendingAndTiming,
4133 		p->HActive,
4134 		p->HRatio,
4135 		p->DPPPerSurface,
4136 
4137 		// Output
4138 		SwathWidthSingleDPP,
4139 		SwathWidthSingleDPPChroma,
4140 		p->SwathWidth,
4141 		p->SwathWidthChroma,
4142 		MaximumSwathHeightY,
4143 		MaximumSwathHeightC,
4144 		p->swath_width_luma_ub,
4145 		p->swath_width_chroma_ub);
4146 
4147 	for (dml_uint_t k = 0; k < p->NumberOfActiveSurfaces; ++k) {
4148 		RoundedUpMaxSwathSizeBytesY[k] = (dml_uint_t)(p->swath_width_luma_ub[k] * p->BytePerPixDETY[k] * MaximumSwathHeightY[k]);
4149 		RoundedUpMaxSwathSizeBytesC[k] = (dml_uint_t)(p->swath_width_chroma_ub[k] * p->BytePerPixDETC[k] * MaximumSwathHeightC[k]);
4150 #ifdef __DML_VBA_DEBUG__
4151 		dml_print("DML::%s: k=%u DPPPerSurface = %u\n", __func__, k, p->DPPPerSurface[k]);
4152 		dml_print("DML::%s: k=%u swath_width_luma_ub = %u\n", __func__, k, p->swath_width_luma_ub[k]);
4153 		dml_print("DML::%s: k=%u BytePerPixDETY = %f\n", __func__, k, p->BytePerPixDETY[k]);
4154 		dml_print("DML::%s: k=%u MaximumSwathHeightY = %u\n", __func__, k, MaximumSwathHeightY[k]);
4155 		dml_print("DML::%s: k=%u RoundedUpMaxSwathSizeBytesY = %u\n", __func__, k, RoundedUpMaxSwathSizeBytesY[k]);
4156 		dml_print("DML::%s: k=%u swath_width_chroma_ub = %u\n", __func__, k, p->swath_width_chroma_ub[k]);
4157 		dml_print("DML::%s: k=%u BytePerPixDETC = %f\n", __func__, k, p->BytePerPixDETC[k]);
4158 		dml_print("DML::%s: k=%u MaximumSwathHeightC = %u\n", __func__, k, MaximumSwathHeightC[k]);
4159 		dml_print("DML::%s: k=%u RoundedUpMaxSwathSizeBytesC = %u\n", __func__, k, RoundedUpMaxSwathSizeBytesC[k]);
4160 #endif
4161 		if (p->SourcePixelFormat[k] == dml_420_10) {
4162 			RoundedUpMaxSwathSizeBytesY[k] = (dml_uint_t)(dml_ceil((dml_float_t) RoundedUpMaxSwathSizeBytesY[k], 256));
4163 			RoundedUpMaxSwathSizeBytesC[k] = (dml_uint_t)(dml_ceil((dml_float_t) RoundedUpMaxSwathSizeBytesC[k], 256));
4164 		}
4165 	}
4166 
4167 	for (dml_uint_t k = 0; k < p->NumberOfActiveSurfaces; ++k) {
4168 		TotalActiveDPP = TotalActiveDPP + (p->ForceSingleDPP ? 1 : p->DPPPerSurface[k]);
4169 		if (p->DPPPerSurface[k] > 0)
4170 			SurfaceDoingUnboundedRequest = k;
4171 		if (p->SourcePixelFormat[k] == dml_420_8 || p->SourcePixelFormat[k] == dml_420_10 ||
4172 				p->SourcePixelFormat[k] == dml_420_12 || p->SourcePixelFormat[k] == dml_rgbe_alpha
4173 				|| p->SurfaceTiling[k] == dml_sw_linear) {
4174 			NoChromaOrLinearSurfaces = false;
4175 		}
4176 	}
4177 
4178 	*p->UnboundedRequestEnabled = UnboundedRequest(p->UseUnboundedRequestingFinal, TotalActiveDPP,
4179 			NoChromaOrLinearSurfaces, p->Output[0]);
4180 
4181 	CalculateDETBufferSize(p->DETSizeOverride,
4182 		p->UseMALLForPStateChange,
4183 		p->ForceSingleDPP,
4184 		p->NumberOfActiveSurfaces,
4185 		*p->UnboundedRequestEnabled,
4186 		p->nomDETInKByte,
4187 		p->MaxTotalDETInKByte,
4188 		p->ConfigReturnBufferSizeInKByte,
4189 		p->MinCompressedBufferSizeInKByte,
4190 		p->ConfigReturnBufferSegmentSizeInkByte,
4191 		p->CompressedBufferSegmentSizeInkByteFinal,
4192 		p->SourcePixelFormat,
4193 		p->ReadBandwidthLuma,
4194 		p->ReadBandwidthChroma,
4195 		RoundedUpMaxSwathSizeBytesY,
4196 		RoundedUpMaxSwathSizeBytesC,
4197 		p->DPPPerSurface,
4198 
4199 		// Output
4200 		p->DETBufferSizeInKByte, // per hubp pipe
4201 		p->CompressedBufferSizeInkByte);
4202 
4203 #ifdef __DML_VBA_DEBUG__
4204 		dml_print("DML::%s: TotalActiveDPP = %u\n", __func__, TotalActiveDPP);
4205 		dml_print("DML::%s: nomDETInKByte = %u\n", __func__, p->nomDETInKByte);
4206 		dml_print("DML::%s: ConfigReturnBufferSizeInKByte = %u\n", __func__, p->ConfigReturnBufferSizeInKByte);
4207 		dml_print("DML::%s: UseUnboundedRequestingFinal = %u\n", __func__, p->UseUnboundedRequestingFinal);
4208 		dml_print("DML::%s: UnboundedRequestEnabled = %u\n", __func__, *p->UnboundedRequestEnabled);
4209 		dml_print("DML::%s: CompressedBufferSizeInkByte = %u\n", __func__, *p->CompressedBufferSizeInkByte);
4210 #endif
4211 
4212 	*p->ViewportSizeSupport = true;
4213 	for (dml_uint_t k = 0; k < p->NumberOfActiveSurfaces; ++k) {
4214 
4215 		DETBufferSizeInKByteForSwathCalculation =  (p->UseMALLForPStateChange[k] == dml_use_mall_pstate_change_phantom_pipe ? 1024 : p->DETBufferSizeInKByte[k]);
4216 #ifdef __DML_VBA_DEBUG__
4217 	dml_print("DML::%s: k=%u DETBufferSizeInKByteForSwathCalculation = %u\n", __func__, k, DETBufferSizeInKByteForSwathCalculation);
4218 #endif
4219 
4220 		if (RoundedUpMaxSwathSizeBytesY[k] + RoundedUpMaxSwathSizeBytesC[k] <= DETBufferSizeInKByteForSwathCalculation * 1024 / 2) {
4221 			p->SwathHeightY[k] = MaximumSwathHeightY[k];
4222 			p->SwathHeightC[k] = MaximumSwathHeightC[k];
4223 			RoundedUpSwathSizeBytesY[k] = RoundedUpMaxSwathSizeBytesY[k];
4224 			RoundedUpSwathSizeBytesC[k] = RoundedUpMaxSwathSizeBytesC[k];
4225 		} else if (RoundedUpMaxSwathSizeBytesY[k] >= 1.5 * RoundedUpMaxSwathSizeBytesC[k] && RoundedUpMaxSwathSizeBytesY[k] / 2 + RoundedUpMaxSwathSizeBytesC[k] <= DETBufferSizeInKByteForSwathCalculation * 1024 / 2) {
4226 			p->SwathHeightY[k] = MaximumSwathHeightY[k] / 2;
4227 			p->SwathHeightC[k] = MaximumSwathHeightC[k];
4228 			RoundedUpSwathSizeBytesY[k] = RoundedUpMaxSwathSizeBytesY[k] / 2;
4229 			RoundedUpSwathSizeBytesC[k] = RoundedUpMaxSwathSizeBytesC[k];
4230 		} else if (RoundedUpMaxSwathSizeBytesY[k] < 1.5 * RoundedUpMaxSwathSizeBytesC[k] && RoundedUpMaxSwathSizeBytesY[k] + RoundedUpMaxSwathSizeBytesC[k] / 2 <= DETBufferSizeInKByteForSwathCalculation * 1024 / 2) {
4231 			p->SwathHeightY[k] = MaximumSwathHeightY[k];
4232 			p->SwathHeightC[k] = MaximumSwathHeightC[k] / 2;
4233 			RoundedUpSwathSizeBytesY[k] = RoundedUpMaxSwathSizeBytesY[k];
4234 			RoundedUpSwathSizeBytesC[k] = RoundedUpMaxSwathSizeBytesC[k] / 2;
4235 		} else {
4236 			p->SwathHeightY[k] = MaximumSwathHeightY[k] / 2;
4237 			p->SwathHeightC[k] = MaximumSwathHeightC[k] / 2;
4238 			RoundedUpSwathSizeBytesY[k] = RoundedUpMaxSwathSizeBytesY[k] / 2;
4239 			RoundedUpSwathSizeBytesC[k] = RoundedUpMaxSwathSizeBytesC[k] / 2;
4240 		}
4241 
4242 		if ((RoundedUpMaxSwathSizeBytesY[k] / 2 + RoundedUpMaxSwathSizeBytesC[k] / 2 > DETBufferSizeInKByteForSwathCalculation * 1024 / 2) ||
4243 			p->SwathWidth[k] > p->MaximumSwathWidthLuma[k] || (p->SwathHeightC[k] > 0 && p->SwathWidthChroma[k] > p->MaximumSwathWidthChroma[k])) {
4244 			*p->ViewportSizeSupport = false;
4245 			p->ViewportSizeSupportPerSurface[k] = false;
4246 		} else {
4247 			p->ViewportSizeSupportPerSurface[k] = true;
4248 		}
4249 
4250 	if (p->SwathHeightC[k] == 0) {
4251 #ifdef __DML_VBA_DEBUG__
4252 			dml_print("DML::%s: k=%u All DET for plane0\n", __func__, k);
4253 #endif
4254 			p->DETBufferSizeY[k] = p->DETBufferSizeInKByte[k] * 1024;
4255 			p->DETBufferSizeC[k] = 0;
4256 		} else if (RoundedUpSwathSizeBytesY[k] <= 1.5 * RoundedUpSwathSizeBytesC[k]) {
4257 #ifdef __DML_VBA_DEBUG__
4258 			dml_print("DML::%s: k=%u Half DET for plane0, half for plane1\n", __func__, k);
4259 #endif
4260 			p->DETBufferSizeY[k] = p->DETBufferSizeInKByte[k] * 1024 / 2;
4261 			p->DETBufferSizeC[k] = p->DETBufferSizeInKByte[k] * 1024 / 2;
4262 		} else {
4263 #ifdef __DML_VBA_DEBUG__
4264 			dml_print("DML::%s: k=%u 2/3 DET for plane0, 1/3 for plane1\n", __func__, k);
4265 #endif
4266 			p->DETBufferSizeY[k] = (dml_uint_t)(dml_floor(p->DETBufferSizeInKByte[k] * 1024 * 2 / 3, 1024));
4267 			p->DETBufferSizeC[k] = p->DETBufferSizeInKByte[k] * 1024 - p->DETBufferSizeY[k];
4268 		}
4269 
4270 #ifdef __DML_VBA_DEBUG__
4271 		dml_print("DML::%s: k=%u SwathHeightY = %u\n", __func__, k, p->SwathHeightY[k]);
4272 		dml_print("DML::%s: k=%u SwathHeightC = %u\n", __func__, k, p->SwathHeightC[k]);
4273 		dml_print("DML::%s: k=%u RoundedUpMaxSwathSizeBytesY = %u\n", __func__, k, RoundedUpMaxSwathSizeBytesY[k]);
4274 		dml_print("DML::%s: k=%u RoundedUpMaxSwathSizeBytesC = %u\n", __func__, k, RoundedUpMaxSwathSizeBytesC[k]);
4275 		dml_print("DML::%s: k=%u RoundedUpSwathSizeBytesY = %u\n", __func__, k, RoundedUpSwathSizeBytesY[k]);
4276 		dml_print("DML::%s: k=%u RoundedUpSwathSizeBytesC = %u\n", __func__, k, RoundedUpSwathSizeBytesC[k]);
4277 		dml_print("DML::%s: k=%u DETBufferSizeInKByte = %u\n", __func__, k, p->DETBufferSizeInKByte[k]);
4278 		dml_print("DML::%s: k=%u DETBufferSizeY = %u\n", __func__, k, p->DETBufferSizeY[k]);
4279 		dml_print("DML::%s: k=%u DETBufferSizeC = %u\n", __func__, k, p->DETBufferSizeC[k]);
4280 		dml_print("DML::%s: k=%u ViewportSizeSupportPerSurface = %u\n", __func__, k, p->ViewportSizeSupportPerSurface[k]);
4281 #endif
4282 
4283 	}
4284 
4285 	*p->compbuf_reserved_space_64b = 2 * p->PixelChunkSizeInKByte * 1024 / 64;
4286 	if (*p->UnboundedRequestEnabled) {
4287 		*p->compbuf_reserved_space_64b = dml_max(*p->compbuf_reserved_space_64b,
4288 				(dml_float_t)(p->ROBBufferSizeInKByte * 1024/64)
4289 				- (dml_float_t)(RoundedUpSwathSizeBytesY[SurfaceDoingUnboundedRequest] * TTUFIFODEPTH / MAXIMUMCOMPRESSION/64));
4290 	}
4291 	*p->compbuf_reserved_space_zs = 2 * p->PixelChunkSizeInKByte * 1024 / 256;
4292 } // CalculateSwathAndDETConfiguration
4293 
CalculateSwathWidth(dml_bool_t ForceSingleDPP,dml_uint_t NumberOfActiveSurfaces,enum dml_source_format_class SourcePixelFormat[],enum dml_rotation_angle SourceScan[],dml_bool_t ViewportStationary[],dml_uint_t ViewportWidth[],dml_uint_t ViewportHeight[],dml_uint_t ViewportXStart[],dml_uint_t ViewportYStart[],dml_uint_t ViewportXStartC[],dml_uint_t ViewportYStartC[],dml_uint_t SurfaceWidthY[],dml_uint_t SurfaceWidthC[],dml_uint_t SurfaceHeightY[],dml_uint_t SurfaceHeightC[],enum dml_odm_mode ODMMode[],dml_uint_t BytePerPixY[],dml_uint_t BytePerPixC[],dml_uint_t Read256BytesBlockHeightY[],dml_uint_t Read256BytesBlockHeightC[],dml_uint_t Read256BytesBlockWidthY[],dml_uint_t Read256BytesBlockWidthC[],dml_uint_t BlendingAndTiming[],dml_uint_t HActive[],dml_float_t HRatio[],dml_uint_t DPPPerSurface[],dml_uint_t SwathWidthSingleDPPY[],dml_uint_t SwathWidthSingleDPPC[],dml_uint_t SwathWidthY[],dml_uint_t SwathWidthC[],dml_uint_t MaximumSwathHeightY[],dml_uint_t MaximumSwathHeightC[],dml_uint_t swath_width_luma_ub[],dml_uint_t swath_width_chroma_ub[])4294 static void CalculateSwathWidth(
4295 		dml_bool_t ForceSingleDPP,
4296 		dml_uint_t NumberOfActiveSurfaces,
4297 		enum dml_source_format_class SourcePixelFormat[],
4298 		enum dml_rotation_angle SourceScan[],
4299 		dml_bool_t ViewportStationary[],
4300 		dml_uint_t ViewportWidth[],
4301 		dml_uint_t ViewportHeight[],
4302 		dml_uint_t ViewportXStart[],
4303 		dml_uint_t ViewportYStart[],
4304 		dml_uint_t ViewportXStartC[],
4305 		dml_uint_t ViewportYStartC[],
4306 		dml_uint_t SurfaceWidthY[],
4307 		dml_uint_t SurfaceWidthC[],
4308 		dml_uint_t SurfaceHeightY[],
4309 		dml_uint_t SurfaceHeightC[],
4310 		enum dml_odm_mode ODMMode[],
4311 		dml_uint_t BytePerPixY[],
4312 		dml_uint_t BytePerPixC[],
4313 		dml_uint_t Read256BytesBlockHeightY[],
4314 		dml_uint_t Read256BytesBlockHeightC[],
4315 		dml_uint_t Read256BytesBlockWidthY[],
4316 		dml_uint_t Read256BytesBlockWidthC[],
4317 		dml_uint_t BlendingAndTiming[],
4318 		dml_uint_t HActive[],
4319 		dml_float_t HRatio[],
4320 		dml_uint_t DPPPerSurface[],
4321 
4322 		// Output
4323 		dml_uint_t SwathWidthSingleDPPY[],
4324 		dml_uint_t SwathWidthSingleDPPC[],
4325 		dml_uint_t SwathWidthY[], // per-pipe
4326 		dml_uint_t SwathWidthC[], // per-pipe
4327 		dml_uint_t MaximumSwathHeightY[],
4328 		dml_uint_t MaximumSwathHeightC[],
4329 		dml_uint_t swath_width_luma_ub[], // per-pipe
4330 		dml_uint_t swath_width_chroma_ub[]) // per-pipe
4331 {
4332 	enum dml_odm_mode   MainSurfaceODMMode;
4333 	dml_uint_t surface_width_ub_l;
4334 	dml_uint_t surface_height_ub_l;
4335 	dml_uint_t surface_width_ub_c = 0;
4336 	dml_uint_t surface_height_ub_c = 0;
4337 
4338 #ifdef __DML_VBA_DEBUG__
4339 	dml_print("DML::%s: ForceSingleDPP = %u\n", __func__, ForceSingleDPP);
4340 	dml_print("DML::%s: NumberOfActiveSurfaces = %u\n", __func__, NumberOfActiveSurfaces);
4341 #endif
4342 
4343 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
4344 		if (!dml_is_vertical_rotation(SourceScan[k])) {
4345 			SwathWidthSingleDPPY[k] = ViewportWidth[k];
4346 		} else {
4347 			SwathWidthSingleDPPY[k] = ViewportHeight[k];
4348 		}
4349 
4350 #ifdef __DML_VBA_DEBUG__
4351 		dml_print("DML::%s: k=%u ViewportWidth=%u\n", __func__, k, ViewportWidth[k]);
4352 		dml_print("DML::%s: k=%u ViewportHeight=%u\n", __func__, k, ViewportHeight[k]);
4353 		dml_print("DML::%s: k=%u DPPPerSurface=%u\n", __func__, k, DPPPerSurface[k]);
4354 #endif
4355 
4356 		MainSurfaceODMMode = ODMMode[k];
4357 		for (dml_uint_t j = 0; j < NumberOfActiveSurfaces; ++j) {
4358 			if (BlendingAndTiming[k] == j) {
4359 				MainSurfaceODMMode = ODMMode[j];
4360 			}
4361 		}
4362 
4363 		if (ForceSingleDPP) {
4364 			SwathWidthY[k] = SwathWidthSingleDPPY[k];
4365 		} else {
4366 			if (MainSurfaceODMMode == dml_odm_mode_combine_4to1) {
4367 				SwathWidthY[k] = (dml_uint_t)(dml_min(SwathWidthSingleDPPY[k], dml_round(HActive[k] / 4.0 * HRatio[k], true)));
4368 			} else if (MainSurfaceODMMode == dml_odm_mode_combine_2to1) {
4369 				SwathWidthY[k] = (dml_uint_t)(dml_min(SwathWidthSingleDPPY[k], dml_round(HActive[k] / 2.0 * HRatio[k], true)));
4370 			} else if (DPPPerSurface[k] == 2) {
4371 				SwathWidthY[k] = SwathWidthSingleDPPY[k] / 2;
4372 			} else {
4373 				SwathWidthY[k] = SwathWidthSingleDPPY[k];
4374 			}
4375 		}
4376 
4377 #ifdef __DML_VBA_DEBUG__
4378 		dml_print("DML::%s: k=%u HActive=%u\n", __func__, k, HActive[k]);
4379 		dml_print("DML::%s: k=%u HRatio=%f\n", __func__, k, HRatio[k]);
4380 		dml_print("DML::%s: k=%u MainSurfaceODMMode=%u\n", __func__, k, MainSurfaceODMMode);
4381 		dml_print("DML::%s: k=%u SwathWidthSingleDPPY=%u\n", __func__, k, SwathWidthSingleDPPY[k]);
4382 		dml_print("DML::%s: k=%u SwathWidthY=%u\n", __func__, k, SwathWidthY[k]);
4383 #endif
4384 
4385 		if (SourcePixelFormat[k] == dml_420_8 || SourcePixelFormat[k] == dml_420_10 || SourcePixelFormat[k] == dml_420_12) {
4386 			SwathWidthC[k] = SwathWidthY[k] / 2;
4387 			SwathWidthSingleDPPC[k] = SwathWidthSingleDPPY[k] / 2;
4388 		} else {
4389 			SwathWidthC[k] = SwathWidthY[k];
4390 			SwathWidthSingleDPPC[k] = SwathWidthSingleDPPY[k];
4391 		}
4392 
4393 		if (ForceSingleDPP == true) {
4394 			SwathWidthY[k] = SwathWidthSingleDPPY[k];
4395 			SwathWidthC[k] = SwathWidthSingleDPPC[k];
4396 		}
4397 
4398 		surface_width_ub_l = (dml_uint_t)dml_ceil(SurfaceWidthY[k], Read256BytesBlockWidthY[k]);
4399 		surface_height_ub_l = (dml_uint_t)dml_ceil(SurfaceHeightY[k], Read256BytesBlockHeightY[k]);
4400 
4401 		if (!dml_is_vertical_rotation(SourceScan[k])) {
4402 			MaximumSwathHeightY[k] = Read256BytesBlockHeightY[k];
4403 			MaximumSwathHeightC[k] = Read256BytesBlockHeightC[k];
4404 			if (ViewportStationary[k] && DPPPerSurface[k] == 1) {
4405 				swath_width_luma_ub[k] = (dml_uint_t)(dml_min(surface_width_ub_l, dml_floor(ViewportXStart[k] + SwathWidthY[k] + Read256BytesBlockWidthY[k] - 1, Read256BytesBlockWidthY[k]) - dml_floor(ViewportXStart[k], Read256BytesBlockWidthY[k])));
4406 			} else {
4407 				swath_width_luma_ub[k] = (dml_uint_t)(dml_min(surface_width_ub_l, dml_ceil(SwathWidthY[k] - 1, Read256BytesBlockWidthY[k]) + Read256BytesBlockWidthY[k]));
4408 			}
4409 			if (BytePerPixC[k] > 0) {
4410 				surface_width_ub_c = (dml_uint_t)dml_ceil(SurfaceWidthC[k], Read256BytesBlockWidthC[k]);
4411 				if (ViewportStationary[k] && DPPPerSurface[k] == 1) {
4412 					swath_width_chroma_ub[k] = (dml_uint_t)(dml_min(surface_width_ub_c, dml_floor(ViewportXStartC[k] + SwathWidthC[k] + Read256BytesBlockWidthC[k] - 1, Read256BytesBlockWidthC[k]) - dml_floor(ViewportXStartC[k], Read256BytesBlockWidthC[k])));
4413 				} else {
4414 					swath_width_chroma_ub[k] = (dml_uint_t)(dml_min(surface_width_ub_c, dml_ceil(SwathWidthC[k] - 1, Read256BytesBlockWidthC[k]) + Read256BytesBlockWidthC[k]));
4415 				}
4416 			} else {
4417 				swath_width_chroma_ub[k] = 0;
4418 			}
4419 		} else {
4420 			MaximumSwathHeightY[k] = Read256BytesBlockWidthY[k];
4421 			MaximumSwathHeightC[k] = Read256BytesBlockWidthC[k];
4422 
4423 			if (ViewportStationary[k] && DPPPerSurface[k] == 1) {
4424 				swath_width_luma_ub[k] = (dml_uint_t)(dml_min(surface_height_ub_l, dml_floor(ViewportYStart[k] + SwathWidthY[k] + Read256BytesBlockHeightY[k] - 1, Read256BytesBlockHeightY[k]) - dml_floor(ViewportYStart[k], Read256BytesBlockHeightY[k])));
4425 			} else {
4426 				swath_width_luma_ub[k] = (dml_uint_t)(dml_min(surface_height_ub_l, dml_ceil(SwathWidthY[k] - 1, Read256BytesBlockHeightY[k]) + Read256BytesBlockHeightY[k]));
4427 			}
4428 			if (BytePerPixC[k] > 0) {
4429 				surface_height_ub_c = (dml_uint_t)dml_ceil(SurfaceHeightC[k], Read256BytesBlockHeightC[k]);
4430 				if (ViewportStationary[k] && DPPPerSurface[k] == 1) {
4431 					swath_width_chroma_ub[k] = (dml_uint_t)(dml_min(surface_height_ub_c, dml_floor(ViewportYStartC[k] + SwathWidthC[k] + Read256BytesBlockHeightC[k] - 1, Read256BytesBlockHeightC[k]) - dml_floor(ViewportYStartC[k], Read256BytesBlockHeightC[k])));
4432 				} else {
4433 					swath_width_chroma_ub[k] = (dml_uint_t)(dml_min(surface_height_ub_c, dml_ceil(SwathWidthC[k] - 1, Read256BytesBlockHeightC[k]) + Read256BytesBlockHeightC[k]));
4434 				}
4435 			} else {
4436 				swath_width_chroma_ub[k] = 0;
4437 			}
4438 		}
4439 
4440 #ifdef __DML_VBA_DEBUG__
4441 		dml_print("DML::%s: k=%u surface_width_ub_l=%u\n", __func__, k, surface_width_ub_l);
4442 		dml_print("DML::%s: k=%u surface_height_ub_l=%u\n", __func__, k, surface_height_ub_l);
4443 		dml_print("DML::%s: k=%u surface_width_ub_c=%u\n", __func__, k, surface_width_ub_c);
4444 		dml_print("DML::%s: k=%u surface_height_ub_c=%u\n", __func__, k, surface_height_ub_c);
4445 		dml_print("DML::%s: k=%u Read256BytesBlockWidthY=%u\n", __func__, k, Read256BytesBlockWidthY[k]);
4446 		dml_print("DML::%s: k=%u Read256BytesBlockHeightY=%u\n", __func__, k, Read256BytesBlockHeightY[k]);
4447 		dml_print("DML::%s: k=%u Read256BytesBlockWidthC=%u\n", __func__, k, Read256BytesBlockWidthC[k]);
4448 		dml_print("DML::%s: k=%u Read256BytesBlockHeightC=%u\n", __func__, k, Read256BytesBlockHeightC[k]);
4449 		dml_print("DML::%s: k=%u ViewportStationary=%u\n", __func__, k, ViewportStationary[k]);
4450 		dml_print("DML::%s: k=%u DPPPerSurface=%u\n", __func__, k, DPPPerSurface[k]);
4451 		dml_print("DML::%s: k=%u swath_width_luma_ub=%u\n", __func__, k, swath_width_luma_ub[k]);
4452 		dml_print("DML::%s: k=%u swath_width_chroma_ub=%u\n", __func__, k, swath_width_chroma_ub[k]);
4453 		dml_print("DML::%s: k=%u MaximumSwathHeightY=%u\n", __func__, k, MaximumSwathHeightY[k]);
4454 		dml_print("DML::%s: k=%u MaximumSwathHeightC=%u\n", __func__, k, MaximumSwathHeightC[k]);
4455 #endif
4456 
4457 	}
4458 } // CalculateSwathWidth
4459 
CalculateExtraLatency(dml_uint_t RoundTripPingLatencyCycles,dml_uint_t ReorderingBytes,dml_float_t DCFCLK,dml_uint_t TotalNumberOfActiveDPP,dml_uint_t PixelChunkSizeInKByte,dml_uint_t TotalNumberOfDCCActiveDPP,dml_uint_t MetaChunkSize,dml_float_t ReturnBW,dml_bool_t GPUVMEnable,dml_bool_t HostVMEnable,dml_uint_t NumberOfActiveSurfaces,dml_uint_t NumberOfDPP[],dml_uint_t dpte_group_bytes[],dml_float_t HostVMInefficiencyFactor,dml_uint_t HostVMMinPageSize,dml_uint_t HostVMMaxNonCachedPageTableLevels)4460 static  dml_float_t CalculateExtraLatency(
4461 		dml_uint_t RoundTripPingLatencyCycles,
4462 		dml_uint_t ReorderingBytes,
4463 		dml_float_t DCFCLK,
4464 		dml_uint_t TotalNumberOfActiveDPP,
4465 		dml_uint_t PixelChunkSizeInKByte,
4466 		dml_uint_t TotalNumberOfDCCActiveDPP,
4467 		dml_uint_t MetaChunkSize,
4468 		dml_float_t ReturnBW,
4469 		dml_bool_t GPUVMEnable,
4470 		dml_bool_t HostVMEnable,
4471 		dml_uint_t NumberOfActiveSurfaces,
4472 		dml_uint_t NumberOfDPP[],
4473 		dml_uint_t dpte_group_bytes[],
4474 		dml_float_t HostVMInefficiencyFactor,
4475 		dml_uint_t HostVMMinPageSize,
4476 		dml_uint_t HostVMMaxNonCachedPageTableLevels)
4477 {
4478 	dml_float_t ExtraLatencyBytes;
4479 	dml_float_t ExtraLatency;
4480 
4481 	ExtraLatencyBytes = CalculateExtraLatencyBytes(
4482 			ReorderingBytes,
4483 			TotalNumberOfActiveDPP,
4484 			PixelChunkSizeInKByte,
4485 			TotalNumberOfDCCActiveDPP,
4486 			MetaChunkSize,
4487 			GPUVMEnable,
4488 			HostVMEnable,
4489 			NumberOfActiveSurfaces,
4490 			NumberOfDPP,
4491 			dpte_group_bytes,
4492 			HostVMInefficiencyFactor,
4493 			HostVMMinPageSize,
4494 			HostVMMaxNonCachedPageTableLevels);
4495 
4496 	ExtraLatency = (RoundTripPingLatencyCycles + __DML_ARB_TO_RET_DELAY__) / DCFCLK + ExtraLatencyBytes / ReturnBW;
4497 
4498 #ifdef __DML_VBA_DEBUG__
4499 	dml_print("DML::%s: RoundTripPingLatencyCycles=%u\n", __func__, RoundTripPingLatencyCycles);
4500 	dml_print("DML::%s: DCFCLK=%f\n", __func__, DCFCLK);
4501 	dml_print("DML::%s: ExtraLatencyBytes=%f\n", __func__, ExtraLatencyBytes);
4502 	dml_print("DML::%s: ReturnBW=%f\n", __func__, ReturnBW);
4503 	dml_print("DML::%s: ExtraLatency=%f\n", __func__, ExtraLatency);
4504 #endif
4505 
4506 	return ExtraLatency;
4507 } // CalculateExtraLatency
4508 
CalculateHostVMDynamicLevels(dml_bool_t GPUVMEnable,dml_bool_t HostVMEnable,dml_uint_t HostVMMinPageSize,dml_uint_t HostVMMaxNonCachedPageTableLevels)4509 static dml_uint_t CalculateHostVMDynamicLevels(
4510 									dml_bool_t GPUVMEnable,
4511 									dml_bool_t HostVMEnable,
4512 									dml_uint_t HostVMMinPageSize,
4513 									dml_uint_t HostVMMaxNonCachedPageTableLevels)
4514 {
4515 	dml_uint_t HostVMDynamicLevels = 0;
4516 
4517 	if (GPUVMEnable && HostVMEnable) {
4518 		if (HostVMMinPageSize < 2048)
4519 			HostVMDynamicLevels = HostVMMaxNonCachedPageTableLevels;
4520 		else if (HostVMMinPageSize >= 2048 && HostVMMinPageSize < 1048576)
4521 			HostVMDynamicLevels = (dml_uint_t) dml_max(0, (dml_float_t) HostVMMaxNonCachedPageTableLevels - 1);
4522 		else
4523 			HostVMDynamicLevels = (dml_uint_t) dml_max(0, (dml_float_t) HostVMMaxNonCachedPageTableLevels - 2);
4524 	} else {
4525 		HostVMDynamicLevels = 0;
4526 	}
4527 	return HostVMDynamicLevels;
4528 }
4529 
CalculateExtraLatencyBytes(dml_uint_t ReorderingBytes,dml_uint_t TotalNumberOfActiveDPP,dml_uint_t PixelChunkSizeInKByte,dml_uint_t TotalNumberOfDCCActiveDPP,dml_uint_t MetaChunkSize,dml_bool_t GPUVMEnable,dml_bool_t HostVMEnable,dml_uint_t NumberOfActiveSurfaces,dml_uint_t NumberOfDPP[],dml_uint_t dpte_group_bytes[],dml_float_t HostVMInefficiencyFactor,dml_uint_t HostVMMinPageSize,dml_uint_t HostVMMaxNonCachedPageTableLevels)4530 static dml_uint_t CalculateExtraLatencyBytes(dml_uint_t ReorderingBytes,
4531 										dml_uint_t TotalNumberOfActiveDPP,
4532 										dml_uint_t PixelChunkSizeInKByte,
4533 										dml_uint_t TotalNumberOfDCCActiveDPP,
4534 										dml_uint_t MetaChunkSize,
4535 										dml_bool_t GPUVMEnable,
4536 										dml_bool_t HostVMEnable,
4537 										dml_uint_t NumberOfActiveSurfaces,
4538 										dml_uint_t NumberOfDPP[],
4539 										dml_uint_t dpte_group_bytes[],
4540 										dml_float_t HostVMInefficiencyFactor,
4541 										dml_uint_t HostVMMinPageSize,
4542 										dml_uint_t HostVMMaxNonCachedPageTableLevels)
4543 {
4544 	dml_uint_t  HostVMDynamicLevels = CalculateHostVMDynamicLevels(GPUVMEnable, HostVMEnable, HostVMMinPageSize, HostVMMaxNonCachedPageTableLevels);
4545 	dml_float_t ret                 = ReorderingBytes + (TotalNumberOfActiveDPP * PixelChunkSizeInKByte + TotalNumberOfDCCActiveDPP * MetaChunkSize) * 1024.0;
4546 
4547 	if (GPUVMEnable == true) {
4548 		for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
4549 			ret = ret + NumberOfDPP[k] * dpte_group_bytes[k] * (1 + 8 * HostVMDynamicLevels) * HostVMInefficiencyFactor;
4550 		}
4551 	}
4552 	return (dml_uint_t)(ret);
4553 }
4554 
CalculateUrgentLatency(dml_float_t UrgentLatencyPixelDataOnly,dml_float_t UrgentLatencyPixelMixedWithVMData,dml_float_t UrgentLatencyVMDataOnly,dml_bool_t DoUrgentLatencyAdjustment,dml_float_t UrgentLatencyAdjustmentFabricClockComponent,dml_float_t UrgentLatencyAdjustmentFabricClockReference,dml_float_t FabricClock)4555 static dml_float_t CalculateUrgentLatency(
4556 		dml_float_t UrgentLatencyPixelDataOnly,
4557 		dml_float_t UrgentLatencyPixelMixedWithVMData,
4558 		dml_float_t UrgentLatencyVMDataOnly,
4559 		dml_bool_t DoUrgentLatencyAdjustment,
4560 		dml_float_t UrgentLatencyAdjustmentFabricClockComponent,
4561 		dml_float_t UrgentLatencyAdjustmentFabricClockReference,
4562 		dml_float_t FabricClock)
4563 {
4564 	dml_float_t   ret;
4565 
4566 	ret = dml_max3(UrgentLatencyPixelDataOnly, UrgentLatencyPixelMixedWithVMData, UrgentLatencyVMDataOnly);
4567 	if (DoUrgentLatencyAdjustment == true) {
4568 		ret = ret + UrgentLatencyAdjustmentFabricClockComponent * (UrgentLatencyAdjustmentFabricClockReference / FabricClock - 1);
4569 	}
4570 	return ret;
4571 }
4572 
RequiredDTBCLK(dml_bool_t DSCEnable,dml_float_t PixelClock,enum dml_output_format_class OutputFormat,dml_float_t OutputBpp,dml_uint_t DSCSlices,dml_uint_t HTotal,dml_uint_t HActive,dml_uint_t AudioRate,dml_uint_t AudioLayout)4573 static dml_float_t RequiredDTBCLK(
4574 		dml_bool_t DSCEnable,
4575 		dml_float_t PixelClock,
4576 		enum dml_output_format_class OutputFormat,
4577 		dml_float_t OutputBpp,
4578 		dml_uint_t DSCSlices,
4579 		dml_uint_t HTotal,
4580 		dml_uint_t HActive,
4581 		dml_uint_t AudioRate,
4582 		dml_uint_t AudioLayout)
4583 {
4584 	if (DSCEnable != true) {
4585 		return dml_max(PixelClock / 4.0 * OutputBpp / 24.0, 25.0);
4586 	} else {
4587 		dml_float_t PixelWordRate = PixelClock / (OutputFormat == dml_444 ? 1 : 2);
4588 		dml_float_t HCActive = dml_ceil(DSCSlices * dml_ceil(OutputBpp * dml_ceil(HActive / DSCSlices, 1) / 8.0, 1) / 3.0, 1);
4589 		dml_float_t HCBlank = 64 + 32 * dml_ceil(AudioRate * (AudioLayout == 1 ? 1 : 0.25) * HTotal / (PixelClock * 1000), 1);
4590 		dml_float_t AverageTribyteRate = PixelWordRate * (HCActive + HCBlank) / HTotal;
4591 		dml_float_t HActiveTribyteRate = PixelWordRate * HCActive / HActive;
4592 		return dml_max4(PixelWordRate / 4.0, AverageTribyteRate / 4.0, HActiveTribyteRate / 4.0, 25.0) * 1.002;
4593 	}
4594 }
4595 
UseMinimumDCFCLK(struct display_mode_lib_scratch_st * scratch,struct UseMinimumDCFCLK_params_st * p)4596 static void UseMinimumDCFCLK(struct display_mode_lib_scratch_st *scratch, struct UseMinimumDCFCLK_params_st *p)
4597 {
4598 	struct UseMinimumDCFCLK_locals_st *s = &scratch->UseMinimumDCFCLK_locals;
4599 
4600 	s->NormalEfficiency = p->PercentOfIdealSDPPortBWReceivedAfterUrgLatency / 100.0;
4601 	for (dml_uint_t j = 0; j < 2; ++j) {
4602 
4603 
4604 		s->TotalMaxPrefetchFlipDPTERowBandwidth[j] = 0;
4605 		for (dml_uint_t k = 0; k < p->NumberOfActiveSurfaces; ++k) {
4606 			s->TotalMaxPrefetchFlipDPTERowBandwidth[j] = s->TotalMaxPrefetchFlipDPTERowBandwidth[j] + p->NoOfDPP[j][k] * p->DPTEBytesPerRow[j][k] / (15.75 * p->HTotal[k] / p->PixelClock[k]);
4607 		}
4608 
4609 		for (dml_uint_t k = 0; k <= p->NumberOfActiveSurfaces - 1; ++k) {
4610 			s->NoOfDPPState[k] = p->NoOfDPP[j][k];
4611 		}
4612 
4613 		s->DPTEBandwidth = s->TotalMaxPrefetchFlipDPTERowBandwidth[j];
4614 
4615 		s->DCFCLKRequiredForAverageBandwidth = dml_max(p->ProjectedDCFCLKDeepSleep[j], s->DPTEBandwidth / s->NormalEfficiency / p->ReturnBusWidth);
4616 
4617 		s->ExtraLatencyBytes = CalculateExtraLatencyBytes(p->ReorderingBytes, p->TotalNumberOfActiveDPP[j], p->PixelChunkSizeInKByte, p->TotalNumberOfDCCActiveDPP[j],
4618 			p->MetaChunkSize, p->GPUVMEnable, p->HostVMEnable, p->NumberOfActiveSurfaces, s->NoOfDPPState, p->dpte_group_bytes,
4619 												1, p->HostVMMinPageSize, p->HostVMMaxNonCachedPageTableLevels);
4620 		s->ExtraLatencyCycles = p->RoundTripPingLatencyCycles + __DML_ARB_TO_RET_DELAY__ + s->ExtraLatencyBytes / s->NormalEfficiency / p->ReturnBusWidth;
4621 		for (dml_uint_t k = 0; k < p->NumberOfActiveSurfaces; ++k) {
4622 			dml_float_t DCFCLKCyclesRequiredInPrefetch;
4623 			dml_float_t PrefetchTime;
4624 
4625 			s->PixelDCFCLKCyclesRequiredInPrefetch[k] = (p->PrefetchLinesY[j][k] * p->swath_width_luma_ub_all_states[j][k] * p->BytePerPixelY[k] + p->PrefetchLinesC[j][k] * p->swath_width_chroma_ub_all_states[j][k] * p->BytePerPixelC[k]) / s->NormalEfficiency / p->ReturnBusWidth;
4626 			DCFCLKCyclesRequiredInPrefetch = 2 * s->ExtraLatencyCycles / s->NoOfDPPState[k] + p->PDEAndMetaPTEBytesPerFrame[j][k] / s->NormalEfficiency / s->NormalEfficiency / p->ReturnBusWidth * (p->GPUVMMaxPageTableLevels > 2 ? 1 : 0) + 2 * p->DPTEBytesPerRow[j][k] / s->NormalEfficiency / s->NormalEfficiency / p->ReturnBusWidth + 2 * p->MetaRowBytes[j][k] / s->NormalEfficiency / p->ReturnBusWidth + s->PixelDCFCLKCyclesRequiredInPrefetch[k];
4627 			s->PrefetchPixelLinesTime[k] = dml_max(p->PrefetchLinesY[j][k], p->PrefetchLinesC[j][k]) * p->HTotal[k] / p->PixelClock[k];
4628 			s->DynamicMetadataVMExtraLatency[k] = (p->GPUVMEnable == true && p->DynamicMetadataEnable[k] == true && p->DynamicMetadataVMEnabled == true) ? p->UrgLatency * p->GPUVMMaxPageTableLevels * (p->HostVMEnable == true ? p->HostVMMaxNonCachedPageTableLevels + 1 : 1) : 0;
4629 
4630 			s->MinimumTWait = CalculateTWait(p->MaxPrefetchMode,
4631 				p->UseMALLForPStateChange[k],
4632 				p->SynchronizeDRRDisplaysForUCLKPStateChangeFinal,
4633 				p->DRRDisplay[k],
4634 				p->DRAMClockChangeLatencyFinal,
4635 				p->FCLKChangeLatency,
4636 				p->UrgLatency,
4637 				p->SREnterPlusExitTime);
4638 
4639 			PrefetchTime = (p->MaximumVStartup[j][k] - 1) * p->HTotal[k] / p->PixelClock[k] - s->MinimumTWait - p->UrgLatency * ((p->GPUVMMaxPageTableLevels <= 2 ? p->GPUVMMaxPageTableLevels : p->GPUVMMaxPageTableLevels - 2) *  (p->HostVMEnable == true ? p->HostVMMaxNonCachedPageTableLevels + 1 : 1) - 1) - s->DynamicMetadataVMExtraLatency[k];
4640 
4641 			if (PrefetchTime > 0) {
4642 				dml_float_t ExpectedVRatioPrefetch;
4643 				ExpectedVRatioPrefetch = s->PrefetchPixelLinesTime[k] / (PrefetchTime * s->PixelDCFCLKCyclesRequiredInPrefetch[k] / DCFCLKCyclesRequiredInPrefetch);
4644 				s->DCFCLKRequiredForPeakBandwidthPerSurface[k] = s->NoOfDPPState[k] * s->PixelDCFCLKCyclesRequiredInPrefetch[k] / s->PrefetchPixelLinesTime[k] * dml_max(1.0, ExpectedVRatioPrefetch) * dml_max(1.0, ExpectedVRatioPrefetch / 4);
4645 				if (p->HostVMEnable == true || p->ImmediateFlipRequirement == true) {
4646 					s->DCFCLKRequiredForPeakBandwidthPerSurface[k] = s->DCFCLKRequiredForPeakBandwidthPerSurface[k] + s->NoOfDPPState[k] * s->DPTEBandwidth / s->NormalEfficiency / s->NormalEfficiency / p->ReturnBusWidth;
4647 				}
4648 			} else {
4649 				s->DCFCLKRequiredForPeakBandwidthPerSurface[k] = p->DCFCLKPerState;
4650 			}
4651 			if (p->DynamicMetadataEnable[k] == true) {
4652 				dml_float_t TSetupPipe;
4653 				dml_float_t TdmbfPipe;
4654 				dml_float_t TdmsksPipe;
4655 				dml_float_t TdmecPipe;
4656 				dml_float_t AllowedTimeForUrgentExtraLatency;
4657 
4658 				CalculateVUpdateAndDynamicMetadataParameters(
4659 					p->MaxInterDCNTileRepeaters,
4660 					p->RequiredDPPCLKPerSurface[j][k],
4661 					p->RequiredDISPCLK[j],
4662 					p->ProjectedDCFCLKDeepSleep[j],
4663 					p->PixelClock[k],
4664 					p->HTotal[k],
4665 					p->VTotal[k] - p->VActive[k],
4666 					p->DynamicMetadataTransmittedBytes[k],
4667 					p->DynamicMetadataLinesBeforeActiveRequired[k],
4668 					p->Interlace[k],
4669 					p->ProgressiveToInterlaceUnitInOPP,
4670 
4671 					// Output
4672 					&TSetupPipe,
4673 					&TdmbfPipe,
4674 					&TdmecPipe,
4675 					&TdmsksPipe,
4676 					&s->dummy1,
4677 					&s->dummy2,
4678 					&s->dummy3);
4679 
4680 				AllowedTimeForUrgentExtraLatency = p->MaximumVStartup[j][k] * p->HTotal[k] / p->PixelClock[k] - s->MinimumTWait - TSetupPipe - TdmbfPipe - TdmecPipe - TdmsksPipe - s->DynamicMetadataVMExtraLatency[k];
4681 				if (AllowedTimeForUrgentExtraLatency > 0) {
4682 					s->DCFCLKRequiredForPeakBandwidthPerSurface[k] = dml_max(s->DCFCLKRequiredForPeakBandwidthPerSurface[k], s->ExtraLatencyCycles / AllowedTimeForUrgentExtraLatency);
4683 				} else {
4684 					s->DCFCLKRequiredForPeakBandwidthPerSurface[k] = p->DCFCLKPerState;
4685 				}
4686 			}
4687 		}
4688 		s->DCFCLKRequiredForPeakBandwidth = 0;
4689 		for (dml_uint_t k = 0; k <= p->NumberOfActiveSurfaces - 1; ++k) {
4690 			s->DCFCLKRequiredForPeakBandwidth = s->DCFCLKRequiredForPeakBandwidth + s->DCFCLKRequiredForPeakBandwidthPerSurface[k];
4691 		}
4692 		s->MinimumTvmPlus2Tr0 = p->UrgLatency * (p->GPUVMEnable == true ? (p->HostVMEnable == true ? (p->GPUVMMaxPageTableLevels + 2) * (p->HostVMMaxNonCachedPageTableLevels + 1) - 1 : p->GPUVMMaxPageTableLevels + 1) : 0);
4693 		for (dml_uint_t k = 0; k < p->NumberOfActiveSurfaces; ++k) {
4694 			dml_float_t MaximumTvmPlus2Tr0PlusTsw;
4695 			MaximumTvmPlus2Tr0PlusTsw = (p->MaximumVStartup[j][k] - 2) * p->HTotal[k] / p->PixelClock[k] - s->MinimumTWait - s->DynamicMetadataVMExtraLatency[k];
4696 			if (MaximumTvmPlus2Tr0PlusTsw <= s->MinimumTvmPlus2Tr0 + s->PrefetchPixelLinesTime[k] / 4) {
4697 				s->DCFCLKRequiredForPeakBandwidth = p->DCFCLKPerState;
4698 			} else {
4699 				s->DCFCLKRequiredForPeakBandwidth = dml_max3(s->DCFCLKRequiredForPeakBandwidth,
4700 														2 * s->ExtraLatencyCycles / (MaximumTvmPlus2Tr0PlusTsw - s->MinimumTvmPlus2Tr0 - s->PrefetchPixelLinesTime[k] / 4),
4701 														(2 * s->ExtraLatencyCycles + s->PixelDCFCLKCyclesRequiredInPrefetch[k]) / (MaximumTvmPlus2Tr0PlusTsw - s->MinimumTvmPlus2Tr0));
4702 			}
4703 		}
4704 		p->DCFCLKState[j] = dml_min(p->DCFCLKPerState, 1.05 * dml_max(s->DCFCLKRequiredForAverageBandwidth, s->DCFCLKRequiredForPeakBandwidth));
4705 	}
4706 }
4707 
4708 
UnboundedRequest(enum dml_unbounded_requesting_policy UseUnboundedRequestingFinal,dml_uint_t TotalNumberOfActiveDPP,dml_bool_t NoChromaOrLinear,enum dml_output_encoder_class Output)4709 static dml_bool_t UnboundedRequest(enum dml_unbounded_requesting_policy UseUnboundedRequestingFinal,
4710 						dml_uint_t TotalNumberOfActiveDPP,
4711 						dml_bool_t NoChromaOrLinear,
4712 						enum dml_output_encoder_class Output)
4713 {
4714 	dml_bool_t ret_val = false;
4715 
4716 	ret_val = (UseUnboundedRequestingFinal != dml_unbounded_requesting_disable
4717 			&& TotalNumberOfActiveDPP == 1 && NoChromaOrLinear);
4718 	if (UseUnboundedRequestingFinal == dml_unbounded_requesting_edp_only && Output != dml_edp) {
4719 		ret_val = false;
4720 	}
4721 	return (ret_val);
4722 }
4723 
CalculateSurfaceSizeInMall(dml_uint_t NumberOfActiveSurfaces,dml_uint_t MALLAllocatedForDCN,enum dml_use_mall_for_static_screen_mode UseMALLForStaticScreen[],dml_bool_t DCCEnable[],dml_bool_t ViewportStationary[],dml_uint_t ViewportXStartY[],dml_uint_t ViewportYStartY[],dml_uint_t ViewportXStartC[],dml_uint_t ViewportYStartC[],dml_uint_t ViewportWidthY[],dml_uint_t ViewportHeightY[],dml_uint_t BytesPerPixelY[],dml_uint_t ViewportWidthC[],dml_uint_t ViewportHeightC[],dml_uint_t BytesPerPixelC[],dml_uint_t SurfaceWidthY[],dml_uint_t SurfaceWidthC[],dml_uint_t SurfaceHeightY[],dml_uint_t SurfaceHeightC[],dml_uint_t Read256BytesBlockWidthY[],dml_uint_t Read256BytesBlockWidthC[],dml_uint_t Read256BytesBlockHeightY[],dml_uint_t Read256BytesBlockHeightC[],dml_uint_t ReadBlockWidthY[],dml_uint_t ReadBlockWidthC[],dml_uint_t ReadBlockHeightY[],dml_uint_t ReadBlockHeightC[],dml_uint_t SurfaceSizeInMALL[],dml_bool_t * ExceededMALLSize)4724 static void CalculateSurfaceSizeInMall(
4725 		dml_uint_t NumberOfActiveSurfaces,
4726 		dml_uint_t MALLAllocatedForDCN,
4727 		enum dml_use_mall_for_static_screen_mode UseMALLForStaticScreen[],
4728 		dml_bool_t DCCEnable[],
4729 		dml_bool_t ViewportStationary[],
4730 		dml_uint_t ViewportXStartY[],
4731 		dml_uint_t ViewportYStartY[],
4732 		dml_uint_t ViewportXStartC[],
4733 		dml_uint_t ViewportYStartC[],
4734 		dml_uint_t ViewportWidthY[],
4735 		dml_uint_t ViewportHeightY[],
4736 		dml_uint_t BytesPerPixelY[],
4737 		dml_uint_t ViewportWidthC[],
4738 		dml_uint_t ViewportHeightC[],
4739 		dml_uint_t BytesPerPixelC[],
4740 		dml_uint_t SurfaceWidthY[],
4741 		dml_uint_t SurfaceWidthC[],
4742 		dml_uint_t SurfaceHeightY[],
4743 		dml_uint_t SurfaceHeightC[],
4744 		dml_uint_t Read256BytesBlockWidthY[],
4745 		dml_uint_t Read256BytesBlockWidthC[],
4746 		dml_uint_t Read256BytesBlockHeightY[],
4747 		dml_uint_t Read256BytesBlockHeightC[],
4748 		dml_uint_t ReadBlockWidthY[],
4749 		dml_uint_t ReadBlockWidthC[],
4750 		dml_uint_t ReadBlockHeightY[],
4751 		dml_uint_t ReadBlockHeightC[],
4752 
4753 		// Output
4754 		dml_uint_t SurfaceSizeInMALL[],
4755 		dml_bool_t *ExceededMALLSize)
4756 {
4757 	dml_uint_t TotalSurfaceSizeInMALL  = 0;
4758 
4759 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
4760 		if (ViewportStationary[k]) {
4761 			SurfaceSizeInMALL[k] = (dml_uint_t)(dml_min(dml_ceil(SurfaceWidthY[k], ReadBlockWidthY[k]), dml_floor(ViewportXStartY[k] + ViewportWidthY[k] + ReadBlockWidthY[k] - 1, ReadBlockWidthY[k]) - dml_floor(ViewportXStartY[k], ReadBlockWidthY[k])) *
4762 									dml_min(dml_ceil(SurfaceHeightY[k], ReadBlockHeightY[k]), dml_floor(ViewportYStartY[k] + ViewportHeightY[k] + ReadBlockHeightY[k] - 1, ReadBlockHeightY[k]) - dml_floor(ViewportYStartY[k], ReadBlockHeightY[k])) *
4763 									BytesPerPixelY[k]);
4764 
4765 			if (ReadBlockWidthC[k] > 0) {
4766 				SurfaceSizeInMALL[k] = (dml_uint_t)(SurfaceSizeInMALL[k] +
4767 										dml_min(dml_ceil(SurfaceWidthC[k], ReadBlockWidthC[k]), dml_floor(ViewportXStartC[k] + ViewportWidthC[k] + ReadBlockWidthC[k] - 1, ReadBlockWidthC[k]) - dml_floor(ViewportXStartC[k], ReadBlockWidthC[k])) *
4768 										dml_min(dml_ceil(SurfaceHeightC[k], ReadBlockHeightC[k]), dml_floor(ViewportYStartC[k] + ViewportHeightC[k] + ReadBlockHeightC[k] - 1, ReadBlockHeightC[k]) - dml_floor(ViewportYStartC[k], ReadBlockHeightC[k])) * BytesPerPixelC[k]);
4769 			}
4770 			if (DCCEnable[k] == true) {
4771 				SurfaceSizeInMALL[k] = (dml_uint_t)(SurfaceSizeInMALL[k] +
4772 					dml_min(dml_ceil(SurfaceWidthY[k], 8 * Read256BytesBlockWidthY[k]), dml_floor(ViewportXStartY[k] + ViewportWidthY[k] + 8 * Read256BytesBlockWidthY[k] - 1, 8 * Read256BytesBlockWidthY[k]) - dml_floor(ViewportXStartY[k], 8 * Read256BytesBlockWidthY[k])) *
4773 					dml_min(dml_ceil(SurfaceHeightY[k], 8 * Read256BytesBlockHeightY[k]), dml_floor(ViewportYStartY[k] + ViewportHeightY[k] + 8 * Read256BytesBlockHeightY[k] - 1, 8 * Read256BytesBlockHeightY[k]) - dml_floor(ViewportYStartY[k], 8 * Read256BytesBlockHeightY[k])) * BytesPerPixelY[k] / 256);
4774 				if (Read256BytesBlockWidthC[k] > 0) {
4775 					SurfaceSizeInMALL[k] = (dml_uint_t)(SurfaceSizeInMALL[k] +
4776 						dml_min(dml_ceil(SurfaceWidthC[k], 8 * Read256BytesBlockWidthC[k]), dml_floor(ViewportXStartC[k] + ViewportWidthC[k] + 8 * Read256BytesBlockWidthC[k] - 1, 8 * Read256BytesBlockWidthC[k]) - dml_floor(ViewportXStartC[k], 8 * Read256BytesBlockWidthC[k])) *
4777 						dml_min(dml_ceil(SurfaceHeightC[k], 8 * Read256BytesBlockHeightC[k]), dml_floor(ViewportYStartC[k] + ViewportHeightC[k] + 8 * Read256BytesBlockHeightC[k] - 1, 8 * Read256BytesBlockHeightC[k]) - dml_floor(ViewportYStartC[k], 8 * Read256BytesBlockHeightC[k])) * BytesPerPixelC[k] / 256);
4778 				}
4779 			}
4780 		} else {
4781 			SurfaceSizeInMALL[k] = (dml_uint_t)(dml_ceil(dml_min(SurfaceWidthY[k], ViewportWidthY[k] + ReadBlockWidthY[k] - 1), ReadBlockWidthY[k]) * dml_ceil(dml_min(SurfaceHeightY[k], ViewportHeightY[k] + ReadBlockHeightY[k] - 1), ReadBlockHeightY[k]) * BytesPerPixelY[k]);
4782 			if (ReadBlockWidthC[k] > 0) {
4783 				SurfaceSizeInMALL[k] = (dml_uint_t)(SurfaceSizeInMALL[k] +
4784 										dml_ceil(dml_min(SurfaceWidthC[k], ViewportWidthC[k] + ReadBlockWidthC[k] - 1), ReadBlockWidthC[k]) *
4785 										dml_ceil(dml_min(SurfaceHeightC[k], ViewportHeightC[k] + ReadBlockHeightC[k] - 1), ReadBlockHeightC[k]) * BytesPerPixelC[k]);
4786 		}
4787 		if (DCCEnable[k] == true) {
4788 			SurfaceSizeInMALL[k] = (dml_uint_t)(SurfaceSizeInMALL[k] +
4789 					dml_ceil(dml_min(SurfaceWidthY[k], ViewportWidthY[k] + 8 * Read256BytesBlockWidthY[k] - 1), 8 * Read256BytesBlockWidthY[k]) *
4790 					dml_ceil(dml_min(SurfaceHeightY[k], ViewportHeightY[k] + 8 * Read256BytesBlockHeightY[k] - 1), 8 * Read256BytesBlockHeightY[k]) * BytesPerPixelY[k] / 256);
4791 
4792 				if (Read256BytesBlockWidthC[k] > 0) {
4793 					SurfaceSizeInMALL[k] = (dml_uint_t)(SurfaceSizeInMALL[k] +
4794 						dml_ceil(dml_min(SurfaceWidthC[k], ViewportWidthC[k] + 8 * Read256BytesBlockWidthC[k] - 1), 8 * Read256BytesBlockWidthC[k]) *
4795 						dml_ceil(dml_min(SurfaceHeightC[k], ViewportHeightC[k] + 8 * Read256BytesBlockHeightC[k] - 1), 8 * Read256BytesBlockHeightC[k]) * BytesPerPixelC[k] / 256);
4796 				}
4797 			}
4798 		}
4799 	}
4800 
4801 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
4802 		if (UseMALLForStaticScreen[k] == dml_use_mall_static_screen_enable)
4803 			TotalSurfaceSizeInMALL = TotalSurfaceSizeInMALL + SurfaceSizeInMALL[k];
4804 	}
4805 	*ExceededMALLSize = (TotalSurfaceSizeInMALL > MALLAllocatedForDCN * 1024 * 1024);
4806 } // CalculateSurfaceSizeInMall
4807 
CalculateDETBufferSize(dml_uint_t DETSizeOverride[],enum dml_use_mall_for_pstate_change_mode UseMALLForPStateChange[],dml_bool_t ForceSingleDPP,dml_uint_t NumberOfActiveSurfaces,dml_bool_t UnboundedRequestEnabled,dml_uint_t nomDETInKByte,dml_uint_t MaxTotalDETInKByte,dml_uint_t ConfigReturnBufferSizeInKByte,dml_uint_t MinCompressedBufferSizeInKByte,dml_uint_t ConfigReturnBufferSegmentSizeInkByte,dml_uint_t CompressedBufferSegmentSizeInkByteFinal,enum dml_source_format_class SourcePixelFormat[],dml_float_t ReadBandwidthLuma[],dml_float_t ReadBandwidthChroma[],dml_uint_t RoundedUpMaxSwathSizeBytesY[],dml_uint_t RoundedUpMaxSwathSizeBytesC[],dml_uint_t DPPPerSurface[],dml_uint_t DETBufferSizeInKByte[],dml_uint_t * CompressedBufferSizeInkByte)4808 static void CalculateDETBufferSize(
4809 						dml_uint_t DETSizeOverride[],
4810 						enum dml_use_mall_for_pstate_change_mode UseMALLForPStateChange[],
4811 						dml_bool_t ForceSingleDPP,
4812 						dml_uint_t NumberOfActiveSurfaces,
4813 						dml_bool_t UnboundedRequestEnabled,
4814 						dml_uint_t nomDETInKByte,
4815 						dml_uint_t MaxTotalDETInKByte,
4816 						dml_uint_t ConfigReturnBufferSizeInKByte,
4817 						dml_uint_t MinCompressedBufferSizeInKByte,
4818 						dml_uint_t ConfigReturnBufferSegmentSizeInkByte,
4819 						dml_uint_t CompressedBufferSegmentSizeInkByteFinal,
4820 						enum dml_source_format_class SourcePixelFormat[],
4821 						dml_float_t ReadBandwidthLuma[],
4822 						dml_float_t ReadBandwidthChroma[],
4823 						dml_uint_t RoundedUpMaxSwathSizeBytesY[],
4824 						dml_uint_t RoundedUpMaxSwathSizeBytesC[],
4825 						dml_uint_t DPPPerSurface[],
4826 						// Output
4827 						dml_uint_t DETBufferSizeInKByte[],
4828 						dml_uint_t *CompressedBufferSizeInkByte)
4829 {
4830 	dml_uint_t DETBufferSizePoolInKByte;
4831 	dml_uint_t NextDETBufferPieceInKByte;
4832 	dml_bool_t DETPieceAssignedToThisSurfaceAlready[__DML_NUM_PLANES__];
4833 	dml_bool_t NextPotentialSurfaceToAssignDETPieceFound;
4834 	dml_uint_t NextSurfaceToAssignDETPiece;
4835 	dml_float_t TotalBandwidth;
4836 	dml_float_t BandwidthOfSurfacesNotAssignedDETPiece;
4837 	dml_uint_t max_minDET;
4838 	dml_uint_t minDET;
4839 	dml_uint_t minDET_pipe;
4840 
4841 #ifdef __DML_VBA_DEBUG__
4842 	dml_print("DML::%s: ForceSingleDPP = %u\n", __func__, ForceSingleDPP);
4843 	dml_print("DML::%s: nomDETInKByte = %u\n", __func__, nomDETInKByte);
4844 	dml_print("DML::%s: NumberOfActiveSurfaces = %u\n", __func__, NumberOfActiveSurfaces);
4845 	dml_print("DML::%s: UnboundedRequestEnabled = %u\n", __func__, UnboundedRequestEnabled);
4846 	dml_print("DML::%s: MaxTotalDETInKByte = %u\n", __func__, MaxTotalDETInKByte);
4847 	dml_print("DML::%s: ConfigReturnBufferSizeInKByte = %u\n", __func__, ConfigReturnBufferSizeInKByte);
4848 	dml_print("DML::%s: MinCompressedBufferSizeInKByte = %u\n", __func__, MinCompressedBufferSizeInKByte);
4849 	dml_print("DML::%s: CompressedBufferSegmentSizeInkByteFinal = %u\n", __func__, CompressedBufferSegmentSizeInkByteFinal);
4850 #endif
4851 
4852 	// Note: Will use default det size if that fits 2 swaths
4853 	if (UnboundedRequestEnabled) {
4854 		if (DETSizeOverride[0] > 0) {
4855 			DETBufferSizeInKByte[0] = DETSizeOverride[0];
4856 		} else {
4857 			DETBufferSizeInKByte[0] = (dml_uint_t) dml_max(128.0, dml_ceil(2.0 * ((dml_float_t) RoundedUpMaxSwathSizeBytesY[0] + (dml_float_t) RoundedUpMaxSwathSizeBytesC[0]) / 1024.0, ConfigReturnBufferSegmentSizeInkByte));
4858 		}
4859 		*CompressedBufferSizeInkByte = ConfigReturnBufferSizeInKByte - DETBufferSizeInKByte[0];
4860 	} else {
4861 		DETBufferSizePoolInKByte = MaxTotalDETInKByte;
4862 		for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
4863 			DETBufferSizeInKByte[k] = 0;
4864 			if (SourcePixelFormat[k] == dml_420_8 || SourcePixelFormat[k] == dml_420_10 || SourcePixelFormat[k] == dml_420_12) {
4865 				max_minDET = nomDETInKByte - ConfigReturnBufferSegmentSizeInkByte;
4866 			} else {
4867 				max_minDET = nomDETInKByte;
4868 			}
4869 			minDET = 128;
4870 			minDET_pipe = 0;
4871 
4872 			// add DET resource until can hold 2 full swaths
4873 			while (minDET <= max_minDET && minDET_pipe == 0) {
4874 				if (2.0 * ((dml_float_t) RoundedUpMaxSwathSizeBytesY[k] + (dml_float_t) RoundedUpMaxSwathSizeBytesC[k]) / 1024.0 <= minDET)
4875 					minDET_pipe = minDET;
4876 				minDET = minDET + ConfigReturnBufferSegmentSizeInkByte;
4877 			}
4878 
4879 #ifdef __DML_VBA_DEBUG__
4880 			dml_print("DML::%s: k=%u minDET = %u\n", __func__, k, minDET);
4881 			dml_print("DML::%s: k=%u max_minDET = %u\n", __func__, k, max_minDET);
4882 			dml_print("DML::%s: k=%u minDET_pipe = %u\n", __func__, k, minDET_pipe);
4883 			dml_print("DML::%s: k=%u RoundedUpMaxSwathSizeBytesY = %u\n", __func__, k, RoundedUpMaxSwathSizeBytesY[k]);
4884 			dml_print("DML::%s: k=%u RoundedUpMaxSwathSizeBytesC = %u\n", __func__, k, RoundedUpMaxSwathSizeBytesC[k]);
4885 #endif
4886 
4887 			if (minDET_pipe == 0) {
4888 				minDET_pipe = (dml_uint_t)(dml_max(128, dml_ceil(((dml_float_t)RoundedUpMaxSwathSizeBytesY[k] + (dml_float_t)RoundedUpMaxSwathSizeBytesC[k]) / 1024.0, ConfigReturnBufferSegmentSizeInkByte)));
4889 #ifdef __DML_VBA_DEBUG__
4890 				dml_print("DML::%s: k=%u minDET_pipe = %u (assume each plane take half DET)\n", __func__, k, minDET_pipe);
4891 #endif
4892 			}
4893 
4894 			if (UseMALLForPStateChange[k] == dml_use_mall_pstate_change_phantom_pipe) {
4895 				DETBufferSizeInKByte[k] = 0;
4896 			} else if (DETSizeOverride[k] > 0) {
4897 				DETBufferSizeInKByte[k] = DETSizeOverride[k];
4898 				DETBufferSizePoolInKByte = DETBufferSizePoolInKByte - (ForceSingleDPP ? 1 : DPPPerSurface[k]) * DETSizeOverride[k];
4899 			} else if ((ForceSingleDPP ? 1 : DPPPerSurface[k]) * minDET_pipe <= DETBufferSizePoolInKByte) {
4900 				DETBufferSizeInKByte[k] = minDET_pipe;
4901 				DETBufferSizePoolInKByte = DETBufferSizePoolInKByte - (ForceSingleDPP ? 1 : DPPPerSurface[k]) * minDET_pipe;
4902 			}
4903 
4904 #ifdef __DML_VBA_DEBUG__
4905 			dml_print("DML::%s: k=%u DPPPerSurface = %u\n", __func__, k, DPPPerSurface[k]);
4906 			dml_print("DML::%s: k=%u DETSizeOverride = %u\n", __func__, k, DETSizeOverride[k]);
4907 			dml_print("DML::%s: k=%u DETBufferSizeInKByte = %u\n", __func__, k, DETBufferSizeInKByte[k]);
4908 			dml_print("DML::%s: DETBufferSizePoolInKByte = %u\n", __func__, DETBufferSizePoolInKByte);
4909 #endif
4910 		}
4911 
4912 		TotalBandwidth = 0;
4913 		for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
4914 			if (UseMALLForPStateChange[k] != dml_use_mall_pstate_change_phantom_pipe)
4915 				TotalBandwidth = TotalBandwidth + ReadBandwidthLuma[k] + ReadBandwidthChroma[k];
4916 		}
4917 #ifdef __DML_VBA_DEBUG__
4918 		dml_print("DML::%s: --- Before bandwidth adjustment ---\n", __func__);
4919 		for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
4920 			dml_print("DML::%s: k=%u DETBufferSizeInKByte   = %u\n", __func__, k, DETBufferSizeInKByte[k]);
4921 		}
4922 		dml_print("DML::%s: --- DET allocation with bandwidth ---\n", __func__);
4923 #endif
4924 		dml_print("DML::%s: TotalBandwidth = %f\n", __func__, TotalBandwidth);
4925 		BandwidthOfSurfacesNotAssignedDETPiece = TotalBandwidth;
4926 		for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
4927 
4928 			if (UseMALLForPStateChange[k] == dml_use_mall_pstate_change_phantom_pipe) {
4929 				DETPieceAssignedToThisSurfaceAlready[k] = true;
4930 			} else if (DETSizeOverride[k] > 0 || (((dml_float_t) (ForceSingleDPP ? 1 : DPPPerSurface[k]) * (dml_float_t) DETBufferSizeInKByte[k] / (dml_float_t) MaxTotalDETInKByte) >= ((ReadBandwidthLuma[k] + ReadBandwidthChroma[k]) / TotalBandwidth))) {
4931 				DETPieceAssignedToThisSurfaceAlready[k] = true;
4932 				BandwidthOfSurfacesNotAssignedDETPiece = BandwidthOfSurfacesNotAssignedDETPiece - ReadBandwidthLuma[k] - ReadBandwidthChroma[k];
4933 			} else {
4934 				DETPieceAssignedToThisSurfaceAlready[k] = false;
4935 			}
4936 #ifdef __DML_VBA_DEBUG__
4937 			dml_print("DML::%s: k=%u DETPieceAssignedToThisSurfaceAlready = %u\n", __func__, k, DETPieceAssignedToThisSurfaceAlready[k]);
4938 			dml_print("DML::%s: k=%u BandwidthOfSurfacesNotAssignedDETPiece = %f\n", __func__, k, BandwidthOfSurfacesNotAssignedDETPiece);
4939 #endif
4940 		}
4941 
4942 		for (dml_uint_t j = 0; j < NumberOfActiveSurfaces; ++j) {
4943 			NextPotentialSurfaceToAssignDETPieceFound = false;
4944 			NextSurfaceToAssignDETPiece = 0;
4945 
4946 			for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
4947 #ifdef __DML_VBA_DEBUG__
4948 				dml_print("DML::%s: j=%u k=%u, ReadBandwidthLuma[k] = %f\n", __func__, j, k, ReadBandwidthLuma[k]);
4949 				dml_print("DML::%s: j=%u k=%u, ReadBandwidthChroma[k] = %f\n", __func__, j, k, ReadBandwidthChroma[k]);
4950 				dml_print("DML::%s: j=%u k=%u, ReadBandwidthLuma[Next] = %f\n", __func__, j, k, ReadBandwidthLuma[NextSurfaceToAssignDETPiece]);
4951 				dml_print("DML::%s: j=%u k=%u, ReadBandwidthChroma[Next] = %f\n", __func__, j, k, ReadBandwidthChroma[NextSurfaceToAssignDETPiece]);
4952 				dml_print("DML::%s: j=%u k=%u, NextSurfaceToAssignDETPiece = %u\n", __func__, j, k, NextSurfaceToAssignDETPiece);
4953 #endif
4954 				if (!DETPieceAssignedToThisSurfaceAlready[k] && (!NextPotentialSurfaceToAssignDETPieceFound ||
4955 					ReadBandwidthLuma[k] + ReadBandwidthChroma[k] < ReadBandwidthLuma[NextSurfaceToAssignDETPiece] + ReadBandwidthChroma[NextSurfaceToAssignDETPiece])) {
4956 					NextSurfaceToAssignDETPiece = k;
4957 					NextPotentialSurfaceToAssignDETPieceFound = true;
4958 				}
4959 #ifdef __DML_VBA_DEBUG__
4960 				dml_print("DML::%s: j=%u k=%u, DETPieceAssignedToThisSurfaceAlready = %u\n", __func__, j, k, DETPieceAssignedToThisSurfaceAlready[k]);
4961 				dml_print("DML::%s: j=%u k=%u, NextPotentialSurfaceToAssignDETPieceFound = %u\n", __func__, j, k, NextPotentialSurfaceToAssignDETPieceFound);
4962 #endif
4963 			}
4964 
4965 			if (NextPotentialSurfaceToAssignDETPieceFound) {
4966 				// Note: To show the banker's rounding behavior in VBA and also the fact that the DET buffer size varies due to precision issue
4967 				//
4968 				//dml_float_t tmp1 =  ((dml_float_t) DETBufferSizePoolInKByte * (ReadBandwidthLuma[NextSurfaceToAssignDETPiece] + ReadBandwidthChroma[NextSurfaceToAssignDETPiece]) / BandwidthOfSurfacesNotAssignedDETPiece /
4969 				//                         ((ForceSingleDPP ? 1 : DPPPerSurface[NextSurfaceToAssignDETPiece]) * 64.0));
4970 				//dml_float_t tmp2 =  dml_round((dml_float_t) DETBufferSizePoolInKByte * (ReadBandwidthLuma[NextSurfaceToAssignDETPiece] + ReadBandwidthChroma[NextSurfaceToAssignDETPiece]) / BandwidthOfSurfacesNotAssignedDETPiece /
4971 				//                         ((ForceSingleDPP ? 1 : DPPPerSurface[NextSurfaceToAssignDETPiece]) * 64.0));
4972 				//
4973 				//dml_print("DML::%s: j=%u, tmp1 = %f\n", __func__, j, tmp1);
4974 				//dml_print("DML::%s: j=%u, tmp2 = %f\n", __func__, j, tmp2);
4975 
4976 				NextDETBufferPieceInKByte = (dml_uint_t)(dml_min(
4977 											dml_round((dml_float_t) DETBufferSizePoolInKByte * (ReadBandwidthLuma[NextSurfaceToAssignDETPiece] + ReadBandwidthChroma[NextSurfaceToAssignDETPiece]) / BandwidthOfSurfacesNotAssignedDETPiece /
4978 												((ForceSingleDPP ? 1 : DPPPerSurface[NextSurfaceToAssignDETPiece]) * ConfigReturnBufferSegmentSizeInkByte), true)
4979 												* (ForceSingleDPP ? 1 : DPPPerSurface[NextSurfaceToAssignDETPiece]) * ConfigReturnBufferSegmentSizeInkByte,
4980 											dml_floor((dml_float_t) DETBufferSizePoolInKByte, (ForceSingleDPP ? 1 : DPPPerSurface[NextSurfaceToAssignDETPiece]) * ConfigReturnBufferSegmentSizeInkByte)));
4981 
4982 #ifdef __DML_VBA_DEBUG__
4983 				dml_print("DML::%s: j=%u, DETBufferSizePoolInKByte = %u\n", __func__, j, DETBufferSizePoolInKByte);
4984 				dml_print("DML::%s: j=%u, NextSurfaceToAssignDETPiece = %u\n", __func__, j, NextSurfaceToAssignDETPiece);
4985 				dml_print("DML::%s: j=%u, ReadBandwidthLuma[%u] = %f\n", __func__, j, NextSurfaceToAssignDETPiece, ReadBandwidthLuma[NextSurfaceToAssignDETPiece]);
4986 				dml_print("DML::%s: j=%u, ReadBandwidthChroma[%u] = %f\n", __func__, j, NextSurfaceToAssignDETPiece, ReadBandwidthChroma[NextSurfaceToAssignDETPiece]);
4987 				dml_print("DML::%s: j=%u, BandwidthOfSurfacesNotAssignedDETPiece = %f\n", __func__, j, BandwidthOfSurfacesNotAssignedDETPiece);
4988 				dml_print("DML::%s: j=%u, NextDETBufferPieceInKByte = %u\n", __func__, j, NextDETBufferPieceInKByte);
4989 				dml_print("DML::%s: j=%u, DETBufferSizeInKByte[%u] increases from %u ", __func__, j, NextSurfaceToAssignDETPiece, DETBufferSizeInKByte[NextSurfaceToAssignDETPiece]);
4990 #endif
4991 
4992 				DETBufferSizeInKByte[NextSurfaceToAssignDETPiece] = DETBufferSizeInKByte[NextSurfaceToAssignDETPiece] + NextDETBufferPieceInKByte / (ForceSingleDPP ? 1 : DPPPerSurface[NextSurfaceToAssignDETPiece]);
4993 #ifdef __DML_VBA_DEBUG__
4994 				dml_print("to %u\n", DETBufferSizeInKByte[NextSurfaceToAssignDETPiece]);
4995 #endif
4996 
4997 				DETBufferSizePoolInKByte = DETBufferSizePoolInKByte - NextDETBufferPieceInKByte;
4998 				DETPieceAssignedToThisSurfaceAlready[NextSurfaceToAssignDETPiece] = true;
4999 				BandwidthOfSurfacesNotAssignedDETPiece = BandwidthOfSurfacesNotAssignedDETPiece - (ReadBandwidthLuma[NextSurfaceToAssignDETPiece] + ReadBandwidthChroma[NextSurfaceToAssignDETPiece]);
5000 			}
5001 		}
5002 		*CompressedBufferSizeInkByte = MinCompressedBufferSizeInKByte;
5003 	}
5004 	*CompressedBufferSizeInkByte = *CompressedBufferSizeInkByte * CompressedBufferSegmentSizeInkByteFinal / ConfigReturnBufferSegmentSizeInkByte;
5005 
5006 #ifdef __DML_VBA_DEBUG__
5007 	dml_print("DML::%s: --- After bandwidth adjustment ---\n", __func__);
5008 	dml_print("DML::%s: CompressedBufferSizeInkByte = %u\n", __func__, *CompressedBufferSizeInkByte);
5009 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
5010 		dml_print("DML::%s: k=%u DETBufferSizeInKByte = %u (TotalReadBandWidth=%f)\n", __func__, k, DETBufferSizeInKByte[k], ReadBandwidthLuma[k] + ReadBandwidthChroma[k]);
5011 	}
5012 #endif
5013 } // CalculateDETBufferSize
5014 
5015 
5016 /// @brief Calculate the bound for return buffer sizing
CalculateMaxDETAndMinCompressedBufferSize(dml_uint_t ConfigReturnBufferSizeInKByte,dml_uint_t ConfigReturnBufferSegmentSizeInKByte,dml_uint_t ROBBufferSizeInKByte,dml_uint_t MaxNumDPP,dml_bool_t nomDETInKByteOverrideEnable,dml_uint_t nomDETInKByteOverrideValue,dml_uint_t * MaxTotalDETInKByte,dml_uint_t * nomDETInKByte,dml_uint_t * MinCompressedBufferSizeInKByte)5017 static void CalculateMaxDETAndMinCompressedBufferSize(
5018 		dml_uint_t  ConfigReturnBufferSizeInKByte,
5019 		dml_uint_t  ConfigReturnBufferSegmentSizeInKByte,
5020 		dml_uint_t  ROBBufferSizeInKByte,
5021 		dml_uint_t MaxNumDPP,
5022 		dml_bool_t nomDETInKByteOverrideEnable, // VBA_DELTA, allow DV to override default DET size
5023 		dml_uint_t nomDETInKByteOverrideValue,  // VBA_DELTA
5024 
5025 		// Output
5026 		dml_uint_t *MaxTotalDETInKByte,
5027 		dml_uint_t *nomDETInKByte,
5028 		dml_uint_t *MinCompressedBufferSizeInKByte)
5029 {
5030 	*MaxTotalDETInKByte = ConfigReturnBufferSizeInKByte - ConfigReturnBufferSegmentSizeInKByte;
5031 	*nomDETInKByte = (dml_uint_t)(dml_floor((dml_float_t) *MaxTotalDETInKByte / (dml_float_t) MaxNumDPP, ConfigReturnBufferSegmentSizeInKByte));
5032 	*MinCompressedBufferSizeInKByte = ConfigReturnBufferSizeInKByte - *MaxTotalDETInKByte;
5033 
5034 #ifdef __DML_VBA_DEBUG__
5035 	dml_print("DML::%s: ConfigReturnBufferSizeInKByte = %u\n", __func__, ConfigReturnBufferSizeInKByte);
5036 	dml_print("DML::%s: ROBBufferSizeInKByte = %u\n", __func__, ROBBufferSizeInKByte);
5037 	dml_print("DML::%s: MaxNumDPP = %u\n", __func__, MaxNumDPP);
5038 	dml_print("DML::%s: MaxTotalDETInKByte = %u\n", __func__, *MaxTotalDETInKByte);
5039 	dml_print("DML::%s: nomDETInKByte = %u\n", __func__, *nomDETInKByte);
5040 	dml_print("DML::%s: MinCompressedBufferSizeInKByte = %u\n", __func__, *MinCompressedBufferSizeInKByte);
5041 #endif
5042 
5043 	if (nomDETInKByteOverrideEnable) {
5044 		*nomDETInKByte = nomDETInKByteOverrideValue;
5045 		dml_print("DML::%s: nomDETInKByte = %u (overrided)\n", __func__, *nomDETInKByte);
5046 	}
5047 } // CalculateMaxDETAndMinCompressedBufferSize
5048 
5049 /// @brief Calculate all the RQ request attributes, like row height and # swath
CalculateVMRowAndSwath(struct display_mode_lib_scratch_st * scratch,struct CalculateVMRowAndSwath_params_st * p)5050 static void CalculateVMRowAndSwath(struct display_mode_lib_scratch_st *scratch,
5051 		struct CalculateVMRowAndSwath_params_st *p)
5052 {
5053 	struct CalculateVMRowAndSwath_locals_st *s = &scratch->CalculateVMRowAndSwath_locals;
5054 
5055 	s->HostVMDynamicLevels = CalculateHostVMDynamicLevels(p->GPUVMEnable, p->HostVMEnable, p->HostVMMinPageSize, p->HostVMMaxNonCachedPageTableLevels);
5056 
5057 	for (dml_uint_t k = 0; k < p->NumberOfActiveSurfaces; ++k) {
5058 		if (p->HostVMEnable == true) {
5059 			p->vm_group_bytes[k] = 512;
5060 			p->dpte_group_bytes[k] = 512;
5061 		} else if (p->GPUVMEnable == true) {
5062 			p->vm_group_bytes[k] = 2048;
5063 			if (p->GPUVMMinPageSizeKBytes[k] >= 64 && dml_is_vertical_rotation(p->myPipe[k].SourceScan)) {
5064 				p->dpte_group_bytes[k] = 512;
5065 			} else {
5066 				p->dpte_group_bytes[k] = 2048;
5067 			}
5068 		} else {
5069 			p->vm_group_bytes[k] = 0;
5070 			p->dpte_group_bytes[k] = 0;
5071 		}
5072 
5073 		if (p->myPipe[k].SourcePixelFormat == dml_420_8 || p->myPipe[k].SourcePixelFormat == dml_420_10 ||
5074 			p->myPipe[k].SourcePixelFormat == dml_420_12 || p->myPipe[k].SourcePixelFormat == dml_rgbe_alpha) {
5075 			if ((p->myPipe[k].SourcePixelFormat == dml_420_10 || p->myPipe[k].SourcePixelFormat == dml_420_12) && !dml_is_vertical_rotation(p->myPipe[k].SourceScan)) {
5076 				s->PTEBufferSizeInRequestsForLuma[k] = (p->PTEBufferSizeInRequestsLuma + p->PTEBufferSizeInRequestsChroma) / 2;
5077 				s->PTEBufferSizeInRequestsForChroma[k] = s->PTEBufferSizeInRequestsForLuma[k];
5078 			} else {
5079 				s->PTEBufferSizeInRequestsForLuma[k] = p->PTEBufferSizeInRequestsLuma;
5080 				s->PTEBufferSizeInRequestsForChroma[k] = p->PTEBufferSizeInRequestsChroma;
5081 			}
5082 
5083 			s->PDEAndMetaPTEBytesFrameC = CalculateVMAndRowBytes(
5084 				p->myPipe[k].ViewportStationary,
5085 				p->myPipe[k].DCCEnable,
5086 				p->myPipe[k].DPPPerSurface,
5087 				p->myPipe[k].BlockHeight256BytesC,
5088 				p->myPipe[k].BlockWidth256BytesC,
5089 				p->myPipe[k].SourcePixelFormat,
5090 				p->myPipe[k].SurfaceTiling,
5091 				p->myPipe[k].BytePerPixelC,
5092 				p->myPipe[k].SourceScan,
5093 				p->SwathWidthC[k],
5094 				p->myPipe[k].ViewportHeightChroma,
5095 				p->myPipe[k].ViewportXStartC,
5096 				p->myPipe[k].ViewportYStartC,
5097 				p->GPUVMEnable,
5098 				p->GPUVMMaxPageTableLevels,
5099 				p->GPUVMMinPageSizeKBytes[k],
5100 				s->PTEBufferSizeInRequestsForChroma[k],
5101 				p->myPipe[k].PitchC,
5102 				p->myPipe[k].DCCMetaPitchC,
5103 				p->myPipe[k].BlockWidthC,
5104 				p->myPipe[k].BlockHeightC,
5105 
5106 				// Output
5107 				&s->MetaRowByteC[k],
5108 				&s->PixelPTEBytesPerRowC[k],
5109 				&s->PixelPTEBytesPerRowStorageC[k],
5110 				&p->dpte_row_width_chroma_ub[k],
5111 				&p->dpte_row_height_chroma[k],
5112 				&p->dpte_row_height_linear_chroma[k],
5113 				&s->PixelPTEBytesPerRowC_one_row_per_frame[k],
5114 				&s->dpte_row_width_chroma_ub_one_row_per_frame[k],
5115 				&s->dpte_row_height_chroma_one_row_per_frame[k],
5116 				&p->meta_req_width_chroma[k],
5117 				&p->meta_req_height_chroma[k],
5118 				&p->meta_row_width_chroma[k],
5119 				&p->meta_row_height_chroma[k],
5120 				&p->PixelPTEReqWidthC[k],
5121 				&p->PixelPTEReqHeightC[k],
5122 				&p->PTERequestSizeC[k],
5123 				&p->dpde0_bytes_per_frame_ub_c[k],
5124 				&p->meta_pte_bytes_per_frame_ub_c[k]);
5125 
5126 			p->PrefetchSourceLinesC[k] = CalculatePrefetchSourceLines (
5127 				p->myPipe[k].VRatioChroma,
5128 				p->myPipe[k].VTapsChroma,
5129 				p->myPipe[k].InterlaceEnable,
5130 				p->myPipe[k].ProgressiveToInterlaceUnitInOPP,
5131 				p->myPipe[k].SwathHeightC,
5132 				p->myPipe[k].SourceScan,
5133 				p->myPipe[k].ViewportStationary,
5134 				p->SwathWidthC[k],
5135 				p->myPipe[k].ViewportHeightChroma,
5136 				p->myPipe[k].ViewportXStartC,
5137 				p->myPipe[k].ViewportYStartC,
5138 
5139 				// Output
5140 				&p->VInitPreFillC[k],
5141 				&p->MaxNumSwathC[k]);
5142 		} else {
5143 			s->PTEBufferSizeInRequestsForLuma[k] = p->PTEBufferSizeInRequestsLuma + p->PTEBufferSizeInRequestsChroma;
5144 			s->PTEBufferSizeInRequestsForChroma[k] = 0;
5145 			s->PixelPTEBytesPerRowC[k] = 0;
5146 			s->PixelPTEBytesPerRowStorageC[k] = 0;
5147 			s->PDEAndMetaPTEBytesFrameC = 0;
5148 			s->MetaRowByteC[k] = 0;
5149 			p->MaxNumSwathC[k] = 0;
5150 			p->PrefetchSourceLinesC[k] = 0;
5151 			s->dpte_row_height_chroma_one_row_per_frame[k] = 0;
5152 			s->dpte_row_width_chroma_ub_one_row_per_frame[k] = 0;
5153 			s->PixelPTEBytesPerRowC_one_row_per_frame[k] = 0;
5154 		}
5155 
5156 		s->PDEAndMetaPTEBytesFrameY = CalculateVMAndRowBytes(
5157 			p->myPipe[k].ViewportStationary,
5158 			p->myPipe[k].DCCEnable,
5159 			p->myPipe[k].DPPPerSurface,
5160 			p->myPipe[k].BlockHeight256BytesY,
5161 			p->myPipe[k].BlockWidth256BytesY,
5162 			p->myPipe[k].SourcePixelFormat,
5163 			p->myPipe[k].SurfaceTiling,
5164 			p->myPipe[k].BytePerPixelY,
5165 			p->myPipe[k].SourceScan,
5166 			p->SwathWidthY[k],
5167 			p->myPipe[k].ViewportHeight,
5168 			p->myPipe[k].ViewportXStart,
5169 			p->myPipe[k].ViewportYStart,
5170 			p->GPUVMEnable,
5171 			p->GPUVMMaxPageTableLevels,
5172 			p->GPUVMMinPageSizeKBytes[k],
5173 			s->PTEBufferSizeInRequestsForLuma[k],
5174 			p->myPipe[k].PitchY,
5175 			p->myPipe[k].DCCMetaPitchY,
5176 			p->myPipe[k].BlockWidthY,
5177 			p->myPipe[k].BlockHeightY,
5178 
5179 			// Output
5180 			&s->MetaRowByteY[k],
5181 			&s->PixelPTEBytesPerRowY[k],
5182 			&s->PixelPTEBytesPerRowStorageY[k],
5183 			&p->dpte_row_width_luma_ub[k],
5184 			&p->dpte_row_height_luma[k],
5185 			&p->dpte_row_height_linear_luma[k],
5186 			&s->PixelPTEBytesPerRowY_one_row_per_frame[k],
5187 			&s->dpte_row_width_luma_ub_one_row_per_frame[k],
5188 			&s->dpte_row_height_luma_one_row_per_frame[k],
5189 			&p->meta_req_width[k],
5190 			&p->meta_req_height[k],
5191 			&p->meta_row_width[k],
5192 			&p->meta_row_height[k],
5193 			&p->PixelPTEReqWidthY[k],
5194 			&p->PixelPTEReqHeightY[k],
5195 			&p->PTERequestSizeY[k],
5196 			&p->dpde0_bytes_per_frame_ub_l[k],
5197 			&p->meta_pte_bytes_per_frame_ub_l[k]);
5198 
5199 			p->PrefetchSourceLinesY[k] = CalculatePrefetchSourceLines(
5200 			p->myPipe[k].VRatio,
5201 			p->myPipe[k].VTaps,
5202 			p->myPipe[k].InterlaceEnable,
5203 			p->myPipe[k].ProgressiveToInterlaceUnitInOPP,
5204 			p->myPipe[k].SwathHeightY,
5205 			p->myPipe[k].SourceScan,
5206 			p->myPipe[k].ViewportStationary,
5207 			p->SwathWidthY[k],
5208 			p->myPipe[k].ViewportHeight,
5209 			p->myPipe[k].ViewportXStart,
5210 			p->myPipe[k].ViewportYStart,
5211 
5212 			// Output
5213 			&p->VInitPreFillY[k],
5214 			&p->MaxNumSwathY[k]);
5215 
5216 		p->PDEAndMetaPTEBytesFrame[k] = (s->PDEAndMetaPTEBytesFrameY + s->PDEAndMetaPTEBytesFrameC) * (1 + 8 * s->HostVMDynamicLevels);
5217 		p->MetaRowByte[k] = s->MetaRowByteY[k] + s->MetaRowByteC[k];
5218 
5219 		if (s->PixelPTEBytesPerRowStorageY[k] <= 64 * s->PTEBufferSizeInRequestsForLuma[k] && s->PixelPTEBytesPerRowStorageC[k] <= 64 * s->PTEBufferSizeInRequestsForChroma[k]) {
5220 			p->PTEBufferSizeNotExceeded[k] = true;
5221 		} else {
5222 			p->PTEBufferSizeNotExceeded[k] = false;
5223 #ifdef __DML_VBA_DEBUG__
5224 			dml_print("DML::%s: k=%u, PixelPTEBytesPerRowY = %u\n", __func__, k, s->PixelPTEBytesPerRowY[k]);
5225 			dml_print("DML::%s: k=%u, PixelPTEBytesPerRowC = %u\n", __func__, k, s->PixelPTEBytesPerRowC[k]);
5226 			dml_print("DML::%s: k=%u, PixelPTEBytesPerRowStorageY = %u\n", __func__, k, s->PixelPTEBytesPerRowStorageY[k]);
5227 			dml_print("DML::%s: k=%u, PixelPTEBytesPerRowStorageC = %u\n", __func__, k, s->PixelPTEBytesPerRowStorageC[k]);
5228 			dml_print("DML::%s: k=%u, PTEBufferSizeInRequestsForLuma = %u\n", __func__, k, s->PTEBufferSizeInRequestsForLuma[k]);
5229 			dml_print("DML::%s: k=%u, PTEBufferSizeInRequestsForChroma = %u\n", __func__, k, s->PTEBufferSizeInRequestsForChroma[k]);
5230 			dml_print("DML::%s: k=%u, PTEBufferSizeNotExceeded          = %u\n",  __func__, k, p->PTEBufferSizeNotExceeded[k]);
5231 #endif
5232 		}
5233 		s->one_row_per_frame_fits_in_buffer[k] = (s->PixelPTEBytesPerRowY_one_row_per_frame[k] <= 64 * 2 * s->PTEBufferSizeInRequestsForLuma[k] &&
5234 			s->PixelPTEBytesPerRowC_one_row_per_frame[k] <= 64 * 2 * s->PTEBufferSizeInRequestsForChroma[k]);
5235 
5236 #ifdef __DML_VBA_DEBUG__
5237 		dml_print("DML::%s: k=%u, PDEAndMetaPTEBytesFrame = %u\n", __func__, k, p->PDEAndMetaPTEBytesFrame[k]);
5238 		dml_print("DML::%s: k=%u, PDEAndMetaPTEBytesFrameY = %u\n", __func__, k, s->PDEAndMetaPTEBytesFrameY);
5239 		dml_print("DML::%s: k=%u, PDEAndMetaPTEBytesFrameC = %u\n", __func__, k, s->PDEAndMetaPTEBytesFrameC);
5240 		dml_print("DML::%s: k=%u, HostVMDynamicLevels = %u\n", __func__, k, s->HostVMDynamicLevels);
5241 		dml_print("DML::%s: k=%u, one_row_per_frame_fits_in_buffer = %u\n", __func__, k, s->one_row_per_frame_fits_in_buffer[k]);
5242 		dml_print("DML::%s: k=%u, PixelPTEBytesPerRowY_one_row_per_frame = %u\n", __func__, k, s->PixelPTEBytesPerRowY_one_row_per_frame[k]);
5243 		dml_print("DML::%s: k=%u, PixelPTEBytesPerRowC_one_row_per_frame    = %u\n",  __func__, k, s->PixelPTEBytesPerRowC_one_row_per_frame[k]);
5244 #endif
5245 	}
5246 
5247 	CalculateMALLUseForStaticScreen(
5248 		p->NumberOfActiveSurfaces,
5249 		p->MALLAllocatedForDCN,
5250 		p->UseMALLForStaticScreen,   // mode
5251 		p->SurfaceSizeInMALL,
5252 		s->one_row_per_frame_fits_in_buffer,
5253 		// Output
5254 		p->UsesMALLForStaticScreen); // boolen
5255 
5256 	for (dml_uint_t k = 0; k < p->NumberOfActiveSurfaces; ++k) {
5257 		if (p->PTEBufferModeOverrideEn[k] == 1) {
5258 			p->PTE_BUFFER_MODE[k] = p->PTEBufferModeOverrideVal[k];
5259 		}
5260 		p->PTE_BUFFER_MODE[k] = p->myPipe[k].FORCE_ONE_ROW_FOR_FRAME || p->UsesMALLForStaticScreen[k] || (p->UseMALLForPStateChange[k] == dml_use_mall_pstate_change_sub_viewport) ||
5261 			(p->UseMALLForPStateChange[k] == dml_use_mall_pstate_change_phantom_pipe) || (p->GPUVMMinPageSizeKBytes[k] > 64);
5262 		p->BIGK_FRAGMENT_SIZE[k] = (dml_uint_t)(dml_log2(p->GPUVMMinPageSizeKBytes[k] * 1024) - 12);
5263 	}
5264 
5265 	for (dml_uint_t k = 0; k < p->NumberOfActiveSurfaces; ++k) {
5266 #ifdef __DML_VBA_DEBUG__
5267 		dml_print("DML::%s: k=%u, SurfaceSizeInMALL         = %u\n",  __func__, k, p->SurfaceSizeInMALL[k]);
5268 		dml_print("DML::%s: k=%u, UsesMALLForStaticScreen   = %u\n",  __func__, k, p->UsesMALLForStaticScreen[k]);
5269 #endif
5270 		p->use_one_row_for_frame[k] = p->myPipe[k].FORCE_ONE_ROW_FOR_FRAME || p->UsesMALLForStaticScreen[k] || (p->UseMALLForPStateChange[k] == dml_use_mall_pstate_change_sub_viewport) ||
5271 									(p->UseMALLForPStateChange[k] == dml_use_mall_pstate_change_phantom_pipe) || (p->GPUVMMinPageSizeKBytes[k] > 64 && dml_is_vertical_rotation(p->myPipe[k].SourceScan));
5272 
5273 		p->use_one_row_for_frame_flip[k] = p->use_one_row_for_frame[k] && !(p->UseMALLForPStateChange[k] == dml_use_mall_pstate_change_full_frame);
5274 
5275 		if (p->use_one_row_for_frame[k]) {
5276 			p->dpte_row_height_luma[k] = s->dpte_row_height_luma_one_row_per_frame[k];
5277 			p->dpte_row_width_luma_ub[k] = s->dpte_row_width_luma_ub_one_row_per_frame[k];
5278 			s->PixelPTEBytesPerRowY[k] = s->PixelPTEBytesPerRowY_one_row_per_frame[k];
5279 			p->dpte_row_height_chroma[k] = s->dpte_row_height_chroma_one_row_per_frame[k];
5280 			p->dpte_row_width_chroma_ub[k] = s->dpte_row_width_chroma_ub_one_row_per_frame[k];
5281 			s->PixelPTEBytesPerRowC[k] = s->PixelPTEBytesPerRowC_one_row_per_frame[k];
5282 			p->PTEBufferSizeNotExceeded[k] = s->one_row_per_frame_fits_in_buffer[k];
5283 		}
5284 
5285 		if (p->MetaRowByte[k] <= p->DCCMetaBufferSizeBytes) {
5286 			p->DCCMetaBufferSizeNotExceeded[k] = true;
5287 		} else {
5288 			p->DCCMetaBufferSizeNotExceeded[k] = false;
5289 
5290 #ifdef __DML_VBA_DEBUG__
5291 			dml_print("DML::%s: k=%u, MetaRowByte                   = %u\n",  __func__, k, p->MetaRowByte[k]);
5292 			dml_print("DML::%s: k=%u, DCCMetaBufferSizeBytes        = %u\n",  __func__, k, p->DCCMetaBufferSizeBytes);
5293 			dml_print("DML::%s: k=%u, DCCMetaBufferSizeNotExceeded  = %u\n",  __func__, k, p->DCCMetaBufferSizeNotExceeded[k]);
5294 #endif
5295 		}
5296 		s->PixelPTEBytesPerRowY[k] = s->PixelPTEBytesPerRowY[k] * (1 + 8 * s->HostVMDynamicLevels);
5297 		s->PixelPTEBytesPerRowC[k] = s->PixelPTEBytesPerRowC[k] * (1 + 8 * s->HostVMDynamicLevels);
5298 		p->PixelPTEBytesPerRow[k] = s->PixelPTEBytesPerRowY[k] + s->PixelPTEBytesPerRowC[k];
5299 		if (p->use_one_row_for_frame[k])
5300 			p->PixelPTEBytesPerRow[k] = p->PixelPTEBytesPerRow[k] / 2;
5301 
5302 		CalculateRowBandwidth(
5303 			p->GPUVMEnable,
5304 			p->myPipe[k].SourcePixelFormat,
5305 			p->myPipe[k].VRatio,
5306 			p->myPipe[k].VRatioChroma,
5307 			p->myPipe[k].DCCEnable,
5308 			p->myPipe[k].HTotal / p->myPipe[k].PixelClock,
5309 			s->MetaRowByteY[k],
5310 			s->MetaRowByteC[k],
5311 			p->meta_row_height[k],
5312 			p->meta_row_height_chroma[k],
5313 			s->PixelPTEBytesPerRowY[k],
5314 			s->PixelPTEBytesPerRowC[k],
5315 			p->dpte_row_height_luma[k],
5316 			p->dpte_row_height_chroma[k],
5317 
5318 			// Output
5319 			&p->meta_row_bw[k],
5320 			&p->dpte_row_bw[k]);
5321 #ifdef __DML_VBA_DEBUG__
5322 		dml_print("DML::%s: k=%u, use_one_row_for_frame = %u\n", __func__, k, p->use_one_row_for_frame[k]);
5323 		dml_print("DML::%s: k=%u, use_one_row_for_frame_flip = %u\n", __func__, k, p->use_one_row_for_frame_flip[k]);
5324 		dml_print("DML::%s: k=%u, UseMALLForPStateChange = %u\n", __func__, k, p->UseMALLForPStateChange[k]);
5325 		dml_print("DML::%s: k=%u, dpte_row_height_luma = %u\n", __func__, k, p->dpte_row_height_luma[k]);
5326 		dml_print("DML::%s: k=%u, dpte_row_width_luma_ub = %u\n", __func__, k, p->dpte_row_width_luma_ub[k]);
5327 		dml_print("DML::%s: k=%u, PixelPTEBytesPerRowY = %u\n", __func__, k, s->PixelPTEBytesPerRowY[k]);
5328 		dml_print("DML::%s: k=%u, dpte_row_height_chroma = %u\n", __func__, k, p->dpte_row_height_chroma[k]);
5329 		dml_print("DML::%s: k=%u, dpte_row_width_chroma_ub = %u\n", __func__, k, p->dpte_row_width_chroma_ub[k]);
5330 		dml_print("DML::%s: k=%u, PixelPTEBytesPerRowC = %u\n", __func__, k, s->PixelPTEBytesPerRowC[k]);
5331 		dml_print("DML::%s: k=%u, PixelPTEBytesPerRow = %u\n", __func__, k, p->PixelPTEBytesPerRow[k]);
5332 		dml_print("DML::%s: k=%u, PTEBufferSizeNotExceeded = %u\n", __func__, k, p->PTEBufferSizeNotExceeded[k]);
5333 		dml_print("DML::%s: k=%u, PTE_BUFFER_MODE = %u\n", __func__, k, p->PTE_BUFFER_MODE[k]);
5334 		dml_print("DML::%s: k=%u, BIGK_FRAGMENT_SIZE     = %u\n", __func__, k, p->BIGK_FRAGMENT_SIZE[k]);
5335 #endif
5336 	}
5337 }
5338 
CalculateOutputLink(dml_float_t PHYCLKPerState,dml_float_t PHYCLKD18PerState,dml_float_t PHYCLKD32PerState,dml_float_t Downspreading,dml_bool_t IsMainSurfaceUsingTheIndicatedTiming,enum dml_output_encoder_class Output,enum dml_output_format_class OutputFormat,dml_uint_t HTotal,dml_uint_t HActive,dml_float_t PixelClockBackEnd,dml_float_t ForcedOutputLinkBPP,dml_uint_t DSCInputBitPerComponent,dml_uint_t NumberOfDSCSlices,dml_float_t AudioSampleRate,dml_uint_t AudioSampleLayout,enum dml_odm_mode ODMModeNoDSC,enum dml_odm_mode ODMModeDSC,enum dml_dsc_enable DSCEnable,dml_uint_t OutputLinkDPLanes,enum dml_output_link_dp_rate OutputLinkDPRate,dml_bool_t * RequiresDSC,dml_bool_t * RequiresFEC,dml_float_t * OutBpp,enum dml_output_type_and_rate__type * OutputType,enum dml_output_type_and_rate__rate * OutputRate,dml_uint_t * RequiredSlots)5339 static void CalculateOutputLink(
5340 		dml_float_t PHYCLKPerState,
5341 		dml_float_t PHYCLKD18PerState,
5342 		dml_float_t PHYCLKD32PerState,
5343 		dml_float_t Downspreading,
5344 		dml_bool_t IsMainSurfaceUsingTheIndicatedTiming,
5345 		enum dml_output_encoder_class Output,
5346 		enum dml_output_format_class OutputFormat,
5347 		dml_uint_t HTotal,
5348 		dml_uint_t HActive,
5349 		dml_float_t PixelClockBackEnd,
5350 		dml_float_t ForcedOutputLinkBPP,
5351 		dml_uint_t DSCInputBitPerComponent,
5352 		dml_uint_t NumberOfDSCSlices,
5353 		dml_float_t AudioSampleRate,
5354 		dml_uint_t AudioSampleLayout,
5355 		enum dml_odm_mode ODMModeNoDSC,
5356 		enum dml_odm_mode ODMModeDSC,
5357 		enum dml_dsc_enable DSCEnable,
5358 		dml_uint_t OutputLinkDPLanes,
5359 		enum dml_output_link_dp_rate OutputLinkDPRate,
5360 
5361 		// Output
5362 		dml_bool_t *RequiresDSC,
5363 		dml_bool_t *RequiresFEC,
5364 		dml_float_t *OutBpp,
5365 		enum dml_output_type_and_rate__type *OutputType,
5366 		enum dml_output_type_and_rate__rate *OutputRate,
5367 		dml_uint_t *RequiredSlots)
5368 {
5369 	dml_bool_t LinkDSCEnable;
5370 	dml_uint_t dummy;
5371 	*RequiresDSC = false;
5372 	*RequiresFEC = false;
5373 	*OutBpp = 0;
5374 
5375 	*OutputType = dml_output_type_unknown;
5376 	*OutputRate = dml_output_rate_unknown;
5377 
5378 	if (IsMainSurfaceUsingTheIndicatedTiming) {
5379 		if (Output == dml_hdmi) {
5380 			*RequiresDSC = false;
5381 			*RequiresFEC = false;
5382 			*OutBpp = TruncToValidBPP(dml_min(600, PHYCLKPerState) * 10, 3, HTotal, HActive, PixelClockBackEnd, ForcedOutputLinkBPP, false, Output,
5383 									OutputFormat, DSCInputBitPerComponent, NumberOfDSCSlices, (dml_uint_t)AudioSampleRate, AudioSampleLayout, ODMModeNoDSC, ODMModeDSC, &dummy);
5384 			//OutputTypeAndRate = "HDMI";
5385 			*OutputType = dml_output_type_hdmi;
5386 
5387 		} else if (Output == dml_dp || Output == dml_dp2p0 || Output == dml_edp) {
5388 			if (DSCEnable == dml_dsc_enable) {
5389 				*RequiresDSC = true;
5390 				LinkDSCEnable = true;
5391 				if (Output == dml_dp || Output == dml_dp2p0) {
5392 					*RequiresFEC = true;
5393 				} else {
5394 					*RequiresFEC = false;
5395 				}
5396 			} else {
5397 				*RequiresDSC = false;
5398 				LinkDSCEnable = false;
5399 				if (Output == dml_dp2p0) {
5400 					*RequiresFEC = true;
5401 				} else {
5402 					*RequiresFEC = false;
5403 				}
5404 			}
5405 			if (Output == dml_dp2p0) {
5406 				*OutBpp = 0;
5407 				if ((OutputLinkDPRate == dml_dp_rate_na || OutputLinkDPRate == dml_dp_rate_uhbr10) && PHYCLKD32PerState >= 10000 / 32.0) {
5408 					*OutBpp = TruncToValidBPP((1 - Downspreading / 100) * 10000, OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, ForcedOutputLinkBPP, LinkDSCEnable, Output,
5409 												OutputFormat, DSCInputBitPerComponent, NumberOfDSCSlices, (dml_uint_t)AudioSampleRate, AudioSampleLayout, ODMModeNoDSC, ODMModeDSC, RequiredSlots);
5410 					if (*OutBpp == 0 && PHYCLKD32PerState < 13500 / 32.0 && DSCEnable == dml_dsc_enable_if_necessary && ForcedOutputLinkBPP == 0) {
5411 						*RequiresDSC = true;
5412 						LinkDSCEnable = true;
5413 						*OutBpp = TruncToValidBPP((1 - Downspreading / 100) * 10000, OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, ForcedOutputLinkBPP, LinkDSCEnable, Output,
5414 													OutputFormat, DSCInputBitPerComponent, NumberOfDSCSlices, (dml_uint_t)AudioSampleRate, AudioSampleLayout, ODMModeNoDSC, ODMModeDSC, RequiredSlots);
5415 					}
5416 					//OutputTypeAndRate = Output & " UHBR10";
5417 					*OutputType = dml_output_type_dp2p0;
5418 					*OutputRate = dml_output_rate_dp_rate_uhbr10;
5419 				}
5420 				if ((OutputLinkDPRate == dml_dp_rate_na || OutputLinkDPRate == dml_dp_rate_uhbr13p5) && *OutBpp == 0 && PHYCLKD32PerState >= 13500 / 32.0) {
5421 					*OutBpp = TruncToValidBPP((1 - Downspreading / 100) * 13500, OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, ForcedOutputLinkBPP, LinkDSCEnable, Output,
5422 												OutputFormat, DSCInputBitPerComponent, NumberOfDSCSlices, (dml_uint_t)AudioSampleRate, AudioSampleLayout, ODMModeNoDSC, ODMModeDSC, RequiredSlots);
5423 
5424 					if (*OutBpp == 0 && PHYCLKD32PerState < 20000 / 32 && DSCEnable == dml_dsc_enable_if_necessary && ForcedOutputLinkBPP == 0) {
5425 						*RequiresDSC = true;
5426 						LinkDSCEnable = true;
5427 						*OutBpp = TruncToValidBPP((1 - Downspreading / 100) * 13500, OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, ForcedOutputLinkBPP, LinkDSCEnable, Output,
5428 												OutputFormat, DSCInputBitPerComponent, NumberOfDSCSlices, (dml_uint_t)AudioSampleRate, AudioSampleLayout, ODMModeNoDSC, ODMModeDSC, RequiredSlots);
5429 					}
5430 					//OutputTypeAndRate = Output & " UHBR13p5";
5431 					*OutputType = dml_output_type_dp2p0;
5432 					*OutputRate = dml_output_rate_dp_rate_uhbr13p5;
5433 				}
5434 				if ((OutputLinkDPRate == dml_dp_rate_na || OutputLinkDPRate == dml_dp_rate_uhbr20) && *OutBpp == 0 && PHYCLKD32PerState >= 20000 / 32) {
5435 					*OutBpp = TruncToValidBPP((1 - Downspreading / 100) * 20000, OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, ForcedOutputLinkBPP, LinkDSCEnable, Output,
5436 											OutputFormat, DSCInputBitPerComponent, NumberOfDSCSlices, (dml_uint_t)AudioSampleRate, AudioSampleLayout, ODMModeNoDSC, ODMModeDSC, RequiredSlots);
5437 					if (*OutBpp == 0 && DSCEnable == dml_dsc_enable_if_necessary && ForcedOutputLinkBPP == 0) {
5438 						*RequiresDSC = true;
5439 						LinkDSCEnable = true;
5440 						*OutBpp = TruncToValidBPP((1 - Downspreading / 100) * 20000, OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, ForcedOutputLinkBPP, LinkDSCEnable, Output,
5441 												OutputFormat, DSCInputBitPerComponent, NumberOfDSCSlices, (dml_uint_t)AudioSampleRate, AudioSampleLayout, ODMModeNoDSC, ODMModeDSC, RequiredSlots);
5442 					}
5443 					//OutputTypeAndRate = Output & " UHBR20";
5444 					*OutputType = dml_output_type_dp2p0;
5445 					*OutputRate = dml_output_rate_dp_rate_uhbr20;
5446 				}
5447 			} else { // output is dp or edp
5448 				*OutBpp = 0;
5449 				if ((OutputLinkDPRate == dml_dp_rate_na || OutputLinkDPRate == dml_dp_rate_hbr) && PHYCLKPerState >= 270) {
5450 					*OutBpp = TruncToValidBPP((1 - Downspreading / 100) * 2700, OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, ForcedOutputLinkBPP, LinkDSCEnable, Output,
5451 											OutputFormat, DSCInputBitPerComponent, NumberOfDSCSlices, (dml_uint_t)AudioSampleRate, AudioSampleLayout, ODMModeNoDSC, ODMModeDSC, RequiredSlots);
5452 					if (*OutBpp == 0 && PHYCLKPerState < 540 && DSCEnable == dml_dsc_enable_if_necessary && ForcedOutputLinkBPP == 0) {
5453 						*RequiresDSC = true;
5454 						LinkDSCEnable = true;
5455 						if (Output == dml_dp) {
5456 							*RequiresFEC = true;
5457 						}
5458 						*OutBpp = TruncToValidBPP((1 - Downspreading / 100) * 2700, OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, ForcedOutputLinkBPP, LinkDSCEnable, Output,
5459 												OutputFormat, DSCInputBitPerComponent, NumberOfDSCSlices, (dml_uint_t)AudioSampleRate, AudioSampleLayout, ODMModeNoDSC, ODMModeDSC, RequiredSlots);
5460 					}
5461 					//OutputTypeAndRate = Output & " HBR";
5462 					*OutputType = (Output == dml_dp) ? dml_output_type_dp : dml_output_type_edp;
5463 					*OutputRate = dml_output_rate_dp_rate_hbr;
5464 				}
5465 				if ((OutputLinkDPRate == dml_dp_rate_na || OutputLinkDPRate == dml_dp_rate_hbr2) && *OutBpp == 0 && PHYCLKPerState >= 540) {
5466 					*OutBpp = TruncToValidBPP((1 - Downspreading / 100) * 5400, OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, ForcedOutputLinkBPP, LinkDSCEnable, Output,
5467 											OutputFormat, DSCInputBitPerComponent, NumberOfDSCSlices, (dml_uint_t)AudioSampleRate, AudioSampleLayout, ODMModeNoDSC, ODMModeDSC, RequiredSlots);
5468 
5469 					if (*OutBpp == 0 && PHYCLKPerState < 810 && DSCEnable == dml_dsc_enable_if_necessary && ForcedOutputLinkBPP == 0) {
5470 						*RequiresDSC = true;
5471 						LinkDSCEnable = true;
5472 						if (Output == dml_dp) {
5473 							*RequiresFEC = true;
5474 						}
5475 						*OutBpp = TruncToValidBPP((1 - Downspreading / 100) * 5400, OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, ForcedOutputLinkBPP, LinkDSCEnable, Output,
5476 												OutputFormat, DSCInputBitPerComponent, NumberOfDSCSlices, (dml_uint_t)AudioSampleRate, AudioSampleLayout, ODMModeNoDSC, ODMModeDSC, RequiredSlots);
5477 					}
5478 					//OutputTypeAndRate = Output & " HBR2";
5479 					*OutputType = (Output == dml_dp) ? dml_output_type_dp : dml_output_type_edp;
5480 					*OutputRate = dml_output_rate_dp_rate_hbr2;
5481 				}
5482 				if ((OutputLinkDPRate == dml_dp_rate_na || OutputLinkDPRate == dml_dp_rate_hbr3) && *OutBpp == 0 && PHYCLKPerState >= 810) { // VBA_ERROR, vba code doesn't have hbr3 check
5483 					*OutBpp = TruncToValidBPP((1 - Downspreading / 100) * 8100, OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, ForcedOutputLinkBPP, LinkDSCEnable, Output,
5484 											OutputFormat, DSCInputBitPerComponent, NumberOfDSCSlices, (dml_uint_t)AudioSampleRate, AudioSampleLayout, ODMModeNoDSC, ODMModeDSC, RequiredSlots);
5485 
5486 					if (*OutBpp == 0 && DSCEnable == dml_dsc_enable_if_necessary && ForcedOutputLinkBPP == 0) {
5487 						*RequiresDSC = true;
5488 						LinkDSCEnable = true;
5489 						if (Output == dml_dp) {
5490 							*RequiresFEC = true;
5491 						}
5492 						*OutBpp = TruncToValidBPP((1 - Downspreading / 100) * 8100, OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, ForcedOutputLinkBPP, LinkDSCEnable, Output,
5493 												OutputFormat, DSCInputBitPerComponent, NumberOfDSCSlices, (dml_uint_t)AudioSampleRate, AudioSampleLayout, ODMModeNoDSC, ODMModeDSC, RequiredSlots);
5494 					}
5495 					//OutputTypeAndRate = Output & " HBR3";
5496 					*OutputType = (Output == dml_dp) ? dml_output_type_dp : dml_output_type_edp;
5497 					*OutputRate = dml_output_rate_dp_rate_hbr3;
5498 				}
5499 			}
5500 		}
5501 	}
5502 }
5503 
5504 /// @brief Determine the ODM mode and number of DPP used per plane based on dispclk, dsc usage, odm usage policy
CalculateODMMode(dml_uint_t MaximumPixelsPerLinePerDSCUnit,dml_uint_t HActive,enum dml_output_encoder_class Output,enum dml_output_format_class OutputFormat,enum dml_odm_use_policy ODMUse,dml_float_t StateDispclk,dml_float_t MaxDispclk,dml_bool_t DSCEnable,dml_uint_t TotalNumberOfActiveDPP,dml_uint_t MaxNumDPP,dml_float_t PixelClock,dml_float_t DISPCLKDPPCLKDSCCLKDownSpreading,dml_float_t DISPCLKRampingMargin,dml_float_t DISPCLKDPPCLKVCOSpeed,dml_uint_t NumberOfDSCSlices,dml_bool_t * TotalAvailablePipesSupport,dml_uint_t * NumberOfDPP,enum dml_odm_mode * ODMMode,dml_float_t * RequiredDISPCLKPerSurface)5505 static void CalculateODMMode(
5506 		dml_uint_t MaximumPixelsPerLinePerDSCUnit,
5507 		dml_uint_t HActive,
5508 		enum dml_output_encoder_class Output,
5509 		enum dml_output_format_class OutputFormat,
5510 		enum dml_odm_use_policy ODMUse,
5511 		dml_float_t StateDispclk,
5512 		dml_float_t MaxDispclk,
5513 		dml_bool_t DSCEnable,
5514 		dml_uint_t TotalNumberOfActiveDPP,
5515 		dml_uint_t MaxNumDPP,
5516 		dml_float_t PixelClock,
5517 		dml_float_t DISPCLKDPPCLKDSCCLKDownSpreading,
5518 		dml_float_t DISPCLKRampingMargin,
5519 		dml_float_t DISPCLKDPPCLKVCOSpeed,
5520 		dml_uint_t NumberOfDSCSlices,
5521 
5522 		// Output
5523 		dml_bool_t *TotalAvailablePipesSupport,
5524 		dml_uint_t *NumberOfDPP,
5525 		enum dml_odm_mode *ODMMode,
5526 		dml_float_t *RequiredDISPCLKPerSurface)
5527 {
5528 
5529 	dml_float_t SurfaceRequiredDISPCLKWithoutODMCombine;
5530 	dml_float_t SurfaceRequiredDISPCLKWithODMCombineTwoToOne;
5531 	dml_float_t SurfaceRequiredDISPCLKWithODMCombineFourToOne;
5532 
5533 	SurfaceRequiredDISPCLKWithoutODMCombine = CalculateRequiredDispclk(dml_odm_mode_bypass, PixelClock, DISPCLKDPPCLKDSCCLKDownSpreading, DISPCLKRampingMargin, DISPCLKDPPCLKVCOSpeed, MaxDispclk);
5534 	SurfaceRequiredDISPCLKWithODMCombineTwoToOne = CalculateRequiredDispclk(dml_odm_mode_combine_2to1, PixelClock, DISPCLKDPPCLKDSCCLKDownSpreading, DISPCLKRampingMargin, DISPCLKDPPCLKVCOSpeed, MaxDispclk);
5535 	SurfaceRequiredDISPCLKWithODMCombineFourToOne = CalculateRequiredDispclk(dml_odm_mode_combine_4to1, PixelClock, DISPCLKDPPCLKDSCCLKDownSpreading, DISPCLKRampingMargin, DISPCLKDPPCLKVCOSpeed, MaxDispclk);
5536 	*TotalAvailablePipesSupport = true;
5537 
5538 	if (OutputFormat == dml_420) {
5539 		if (HActive > 4 * DML2_MAX_FMT_420_BUFFER_WIDTH)
5540 			*TotalAvailablePipesSupport = false;
5541 		else if (HActive > 2 * DML2_MAX_FMT_420_BUFFER_WIDTH)
5542 			ODMUse = dml_odm_use_policy_combine_4to1;
5543 		else if (HActive > DML2_MAX_FMT_420_BUFFER_WIDTH && ODMUse != dml_odm_use_policy_combine_4to1)
5544 			ODMUse = dml_odm_use_policy_combine_2to1;
5545 		if (Output == dml_hdmi && ODMUse == dml_odm_use_policy_combine_2to1)
5546 			*TotalAvailablePipesSupport = false;
5547 		if ((Output == dml_hdmi || Output == dml_dp || Output == dml_edp) && ODMUse == dml_odm_use_policy_combine_4to1)
5548 			*TotalAvailablePipesSupport = false;
5549 	}
5550 
5551 	if (ODMUse == dml_odm_use_policy_bypass || ODMUse == dml_odm_use_policy_combine_as_needed)
5552 		*ODMMode = dml_odm_mode_bypass;
5553 	else if (ODMUse == dml_odm_use_policy_combine_2to1)
5554 		*ODMMode = dml_odm_mode_combine_2to1;
5555 	else if (ODMUse == dml_odm_use_policy_combine_4to1)
5556 		*ODMMode = dml_odm_mode_combine_4to1;
5557 	else if (ODMUse == dml_odm_use_policy_split_1to2)
5558 		*ODMMode = dml_odm_mode_split_1to2;
5559 	else if (ODMUse == dml_odm_use_policy_mso_1to2)
5560 		*ODMMode = dml_odm_mode_mso_1to2;
5561 	else if (ODMUse == dml_odm_use_policy_mso_1to4)
5562 		*ODMMode = dml_odm_mode_mso_1to4;
5563 
5564 	*RequiredDISPCLKPerSurface = SurfaceRequiredDISPCLKWithoutODMCombine;
5565 	*NumberOfDPP = 0;
5566 
5567 	if (!(Output == dml_hdmi || Output == dml_dp || Output == dml_edp) && (ODMUse == dml_odm_use_policy_combine_4to1 || (ODMUse == dml_odm_use_policy_combine_as_needed &&
5568 		(SurfaceRequiredDISPCLKWithODMCombineTwoToOne > StateDispclk || (DSCEnable && (HActive > 2 * MaximumPixelsPerLinePerDSCUnit)) || NumberOfDSCSlices > 8)))) {
5569 		if (TotalNumberOfActiveDPP + 4 <= MaxNumDPP) {
5570 			*ODMMode = dml_odm_mode_combine_4to1;
5571 			*RequiredDISPCLKPerSurface = SurfaceRequiredDISPCLKWithODMCombineFourToOne;
5572 			*NumberOfDPP = 4;
5573 		} else {
5574 			*TotalAvailablePipesSupport = false;
5575 		}
5576 	} else if (Output != dml_hdmi && (ODMUse == dml_odm_use_policy_combine_2to1 || (ODMUse == dml_odm_use_policy_combine_as_needed &&
5577 				((SurfaceRequiredDISPCLKWithoutODMCombine > StateDispclk && SurfaceRequiredDISPCLKWithODMCombineTwoToOne <= StateDispclk) ||
5578 				(DSCEnable && (HActive > MaximumPixelsPerLinePerDSCUnit)) || (NumberOfDSCSlices <= 8 && NumberOfDSCSlices > 4))))) {
5579 		if (TotalNumberOfActiveDPP + 2 <= MaxNumDPP) {
5580 			*ODMMode = dml_odm_mode_combine_2to1;
5581 			*RequiredDISPCLKPerSurface = SurfaceRequiredDISPCLKWithODMCombineTwoToOne;
5582 			*NumberOfDPP = 2;
5583 		} else {
5584 			*TotalAvailablePipesSupport = false;
5585 		}
5586 	} else {
5587 		if (TotalNumberOfActiveDPP + 1 <= MaxNumDPP) {
5588 			*NumberOfDPP = 1;
5589 		} else {
5590 			*TotalAvailablePipesSupport = false;
5591 		}
5592 	}
5593 }
5594 
5595 /// @brief Calculate the required DISPCLK given the odm mode and pixclk
CalculateRequiredDispclk(enum dml_odm_mode ODMMode,dml_float_t PixelClock,dml_float_t DISPCLKDPPCLKDSCCLKDownSpreading,dml_float_t DISPCLKRampingMargin,dml_float_t DISPCLKDPPCLKVCOSpeed,dml_float_t MaxDispclk)5596 static dml_float_t CalculateRequiredDispclk(
5597 		enum dml_odm_mode ODMMode,
5598 		dml_float_t PixelClock,
5599 		dml_float_t DISPCLKDPPCLKDSCCLKDownSpreading,
5600 		dml_float_t DISPCLKRampingMargin,
5601 		dml_float_t DISPCLKDPPCLKVCOSpeed,
5602 		dml_float_t MaxDispclk)
5603 {
5604 	dml_float_t RequiredDispclk = 0.;
5605 	dml_float_t PixelClockAfterODM;
5606 
5607 	dml_float_t DISPCLKWithRampingRoundedToDFSGranularity;
5608 	dml_float_t DISPCLKWithoutRampingRoundedToDFSGranularity;
5609 	dml_float_t MaxDispclkRoundedDownToDFSGranularity;
5610 
5611 	if (ODMMode == dml_odm_mode_combine_4to1) {
5612 		PixelClockAfterODM = PixelClock / 4;
5613 	} else if (ODMMode == dml_odm_mode_combine_2to1) {
5614 		PixelClockAfterODM = PixelClock / 2;
5615 	} else {
5616 		PixelClockAfterODM = PixelClock;
5617 	}
5618 
5619 	DISPCLKWithRampingRoundedToDFSGranularity = RoundToDFSGranularity(PixelClockAfterODM * (1.0 + DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) * (1 + DISPCLKRampingMargin / 100.0), 1, DISPCLKDPPCLKVCOSpeed);
5620 	DISPCLKWithoutRampingRoundedToDFSGranularity = RoundToDFSGranularity(PixelClockAfterODM * (1.0 + DISPCLKDPPCLKDSCCLKDownSpreading / 100.0), 1, DISPCLKDPPCLKVCOSpeed);
5621 	MaxDispclkRoundedDownToDFSGranularity = RoundToDFSGranularity(MaxDispclk, 0, DISPCLKDPPCLKVCOSpeed);
5622 
5623 	if (DISPCLKWithoutRampingRoundedToDFSGranularity > MaxDispclkRoundedDownToDFSGranularity) {
5624 		RequiredDispclk = DISPCLKWithoutRampingRoundedToDFSGranularity;
5625 	} else if (DISPCLKWithRampingRoundedToDFSGranularity > MaxDispclkRoundedDownToDFSGranularity) {
5626 		RequiredDispclk = MaxDispclkRoundedDownToDFSGranularity;
5627 	} else {
5628 		RequiredDispclk = DISPCLKWithRampingRoundedToDFSGranularity;
5629 	}
5630 
5631 	return RequiredDispclk;
5632 }
5633 
5634 /// @brief Determine DPPCLK if there only one DPP per plane, main factor is the pixel rate and DPP scaling parameter
CalculateSinglePipeDPPCLKAndSCLThroughput(dml_float_t HRatio,dml_float_t HRatioChroma,dml_float_t VRatio,dml_float_t VRatioChroma,dml_float_t MaxDCHUBToPSCLThroughput,dml_float_t MaxPSCLToLBThroughput,dml_float_t PixelClock,enum dml_source_format_class SourcePixelFormat,dml_uint_t HTaps,dml_uint_t HTapsChroma,dml_uint_t VTaps,dml_uint_t VTapsChroma,dml_float_t * PSCL_THROUGHPUT,dml_float_t * PSCL_THROUGHPUT_CHROMA,dml_float_t * DPPCLKUsingSingleDPP)5635 static void CalculateSinglePipeDPPCLKAndSCLThroughput(
5636 		dml_float_t HRatio,
5637 		dml_float_t HRatioChroma,
5638 		dml_float_t VRatio,
5639 		dml_float_t VRatioChroma,
5640 		dml_float_t MaxDCHUBToPSCLThroughput,
5641 		dml_float_t MaxPSCLToLBThroughput,
5642 		dml_float_t PixelClock,
5643 		enum dml_source_format_class SourcePixelFormat,
5644 		dml_uint_t HTaps,
5645 		dml_uint_t HTapsChroma,
5646 		dml_uint_t VTaps,
5647 		dml_uint_t VTapsChroma,
5648 
5649 		// Output
5650 		dml_float_t *PSCL_THROUGHPUT,
5651 		dml_float_t *PSCL_THROUGHPUT_CHROMA,
5652 		dml_float_t *DPPCLKUsingSingleDPP)
5653 {
5654 	dml_float_t DPPCLKUsingSingleDPPLuma;
5655 	dml_float_t DPPCLKUsingSingleDPPChroma;
5656 
5657 	if (HRatio > 1) {
5658 		*PSCL_THROUGHPUT = dml_min(MaxDCHUBToPSCLThroughput, MaxPSCLToLBThroughput * HRatio / dml_ceil((dml_float_t) HTaps / 6.0, 1.0));
5659 	} else {
5660 		*PSCL_THROUGHPUT = dml_min(MaxDCHUBToPSCLThroughput, MaxPSCLToLBThroughput);
5661 	}
5662 
5663 	DPPCLKUsingSingleDPPLuma = PixelClock * dml_max3(VTaps / 6 * dml_min(1, HRatio), HRatio * VRatio / *PSCL_THROUGHPUT, 1);
5664 
5665 	if ((HTaps > 6 || VTaps > 6) && DPPCLKUsingSingleDPPLuma < 2 * PixelClock)
5666 		DPPCLKUsingSingleDPPLuma = 2 * PixelClock;
5667 
5668 	if ((SourcePixelFormat != dml_420_8 && SourcePixelFormat != dml_420_10 && SourcePixelFormat != dml_420_12 && SourcePixelFormat != dml_rgbe_alpha)) {
5669 		*PSCL_THROUGHPUT_CHROMA = 0;
5670 		*DPPCLKUsingSingleDPP = DPPCLKUsingSingleDPPLuma;
5671 	} else {
5672 		if (HRatioChroma > 1) {
5673 			*PSCL_THROUGHPUT_CHROMA = dml_min(MaxDCHUBToPSCLThroughput, MaxPSCLToLBThroughput * HRatioChroma / dml_ceil((dml_float_t) HTapsChroma / 6.0, 1.0));
5674 		} else {
5675 			*PSCL_THROUGHPUT_CHROMA = dml_min(MaxDCHUBToPSCLThroughput, MaxPSCLToLBThroughput);
5676 		}
5677 		DPPCLKUsingSingleDPPChroma = PixelClock * dml_max3(VTapsChroma / 6 * dml_min(1, HRatioChroma),
5678 															HRatioChroma * VRatioChroma / *PSCL_THROUGHPUT_CHROMA, 1);
5679 		if ((HTapsChroma > 6 || VTapsChroma > 6) && DPPCLKUsingSingleDPPChroma < 2 * PixelClock)
5680 			DPPCLKUsingSingleDPPChroma = 2 * PixelClock;
5681 		*DPPCLKUsingSingleDPP = dml_max(DPPCLKUsingSingleDPPLuma, DPPCLKUsingSingleDPPChroma);
5682 	}
5683 }
5684 
5685 /// @brief Calculate the actual dppclk freq
5686 /// @param DPPCLKUsingSingleDPP DppClk freq required if there is only 1 DPP per plane
5687 /// @param DPPPerSurface Number of DPP for each plane
CalculateDPPCLK(dml_uint_t NumberOfActiveSurfaces,dml_float_t DISPCLKDPPCLKDSCCLKDownSpreading,dml_float_t DISPCLKDPPCLKVCOSpeed,dml_float_t DPPCLKUsingSingleDPP[],dml_uint_t DPPPerSurface[],dml_float_t * GlobalDPPCLK,dml_float_t Dppclk[])5688 static void CalculateDPPCLK(
5689 		dml_uint_t NumberOfActiveSurfaces,
5690 		dml_float_t DISPCLKDPPCLKDSCCLKDownSpreading,
5691 		dml_float_t DISPCLKDPPCLKVCOSpeed,
5692 		dml_float_t DPPCLKUsingSingleDPP[],
5693 		dml_uint_t DPPPerSurface[],
5694 
5695 		// Output
5696 		dml_float_t *GlobalDPPCLK,
5697 		dml_float_t Dppclk[])
5698 {
5699 	*GlobalDPPCLK = 0;
5700 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
5701 		Dppclk[k] = DPPCLKUsingSingleDPP[k] / DPPPerSurface[k] * (1 + DISPCLKDPPCLKDSCCLKDownSpreading / 100.0);
5702 		*GlobalDPPCLK = dml_max(*GlobalDPPCLK, Dppclk[k]);
5703 	}
5704 	*GlobalDPPCLK = RoundToDFSGranularity(*GlobalDPPCLK, 1, DISPCLKDPPCLKVCOSpeed);
5705 
5706 	dml_print("DML::%s: GlobalDPPCLK = %f\n", __func__, *GlobalDPPCLK);
5707 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
5708 		Dppclk[k] = *GlobalDPPCLK / 255.0 * dml_ceil(Dppclk[k] * 255.0 / *GlobalDPPCLK, 1.0);
5709 		dml_print("DML::%s: Dppclk[%0d] = %f\n", __func__, k, Dppclk[k]);
5710 	}
5711 }
5712 
CalculateMALLUseForStaticScreen(dml_uint_t NumberOfActiveSurfaces,dml_uint_t MALLAllocatedForDCNFinal,enum dml_use_mall_for_static_screen_mode * UseMALLForStaticScreen,dml_uint_t SurfaceSizeInMALL[],dml_bool_t one_row_per_frame_fits_in_buffer[],dml_bool_t UsesMALLForStaticScreen[])5713 static void CalculateMALLUseForStaticScreen(
5714 		dml_uint_t NumberOfActiveSurfaces,
5715 		dml_uint_t MALLAllocatedForDCNFinal,
5716 		enum dml_use_mall_for_static_screen_mode *UseMALLForStaticScreen,
5717 		dml_uint_t SurfaceSizeInMALL[],
5718 		dml_bool_t one_row_per_frame_fits_in_buffer[],
5719 
5720 		// Output
5721 		dml_bool_t UsesMALLForStaticScreen[])
5722 {
5723 
5724 	dml_uint_t SurfaceToAddToMALL;
5725 	dml_bool_t CanAddAnotherSurfaceToMALL;
5726 	dml_uint_t TotalSurfaceSizeInMALL;
5727 
5728 	TotalSurfaceSizeInMALL = 0;
5729 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
5730 		UsesMALLForStaticScreen[k] = (UseMALLForStaticScreen[k] == dml_use_mall_static_screen_enable);
5731 		if (UsesMALLForStaticScreen[k])
5732 			TotalSurfaceSizeInMALL = TotalSurfaceSizeInMALL + SurfaceSizeInMALL[k];
5733 #ifdef __DML_VBA_DEBUG__
5734 		dml_print("DML::%s: k=%u, UsesMALLForStaticScreen = %u\n",  __func__, k, UsesMALLForStaticScreen[k]);
5735 		dml_print("DML::%s: k=%u, TotalSurfaceSizeInMALL = %u\n",  __func__, k, TotalSurfaceSizeInMALL);
5736 #endif
5737 	}
5738 
5739 	SurfaceToAddToMALL = 0;
5740 	CanAddAnotherSurfaceToMALL = true;
5741 	while (CanAddAnotherSurfaceToMALL) {
5742 		CanAddAnotherSurfaceToMALL = false;
5743 		for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
5744 			if (TotalSurfaceSizeInMALL + SurfaceSizeInMALL[k] <= MALLAllocatedForDCNFinal * 1024 * 1024 &&
5745 				!UsesMALLForStaticScreen[k] && UseMALLForStaticScreen[k] != dml_use_mall_static_screen_disable && one_row_per_frame_fits_in_buffer[k] &&
5746 				(!CanAddAnotherSurfaceToMALL || SurfaceSizeInMALL[k] < SurfaceSizeInMALL[SurfaceToAddToMALL])) {
5747 				CanAddAnotherSurfaceToMALL = true;
5748 				SurfaceToAddToMALL = k;
5749 				dml_print("DML::%s: k=%u, UseMALLForStaticScreen = %u (dis, en, optimize)\n",  __func__, k, UseMALLForStaticScreen[k]);
5750 			}
5751 		}
5752 		if (CanAddAnotherSurfaceToMALL) {
5753 			UsesMALLForStaticScreen[SurfaceToAddToMALL] = true;
5754 			TotalSurfaceSizeInMALL = TotalSurfaceSizeInMALL + SurfaceSizeInMALL[SurfaceToAddToMALL];
5755 
5756 #ifdef __DML_VBA_DEBUG__
5757 			dml_print("DML::%s: SurfaceToAddToMALL       = %u\n",  __func__, SurfaceToAddToMALL);
5758 			dml_print("DML::%s: TotalSurfaceSizeInMALL   = %u\n",  __func__, TotalSurfaceSizeInMALL);
5759 #endif
5760 		}
5761 	}
5762 }
5763 
5764 // @brief Calculate return bw for VM only traffic
dml_get_return_bw_mbps_vm_only(const struct soc_bounding_box_st * soc,dml_bool_t use_ideal_dram_bw_strobe,dml_bool_t HostVMEnable,dml_float_t DCFCLK,dml_float_t FabricClock,dml_float_t DRAMSpeed)5765 dml_float_t dml_get_return_bw_mbps_vm_only(
5766 								const struct soc_bounding_box_st *soc,
5767 								dml_bool_t use_ideal_dram_bw_strobe,
5768 								dml_bool_t HostVMEnable,
5769 								dml_float_t DCFCLK,
5770 								dml_float_t FabricClock,
5771 								dml_float_t DRAMSpeed)
5772 {
5773 	dml_float_t VMDataOnlyReturnBW =
5774 		dml_min3(soc->return_bus_width_bytes * DCFCLK * soc->pct_ideal_sdp_bw_after_urgent / 100.0,
5775 				FabricClock * soc->fabric_datapath_to_dcn_data_return_bytes * soc->pct_ideal_sdp_bw_after_urgent / 100.0,
5776 				DRAMSpeed * soc->num_chans *  soc->dram_channel_width_bytes *
5777 				((use_ideal_dram_bw_strobe && !HostVMEnable) ? soc->pct_ideal_dram_bw_after_urgent_strobe :  soc->pct_ideal_dram_bw_after_urgent_vm_only) / 100.0);
5778 #ifdef __DML_VBA_DEBUG__
5779 	dml_print("DML::%s: use_ideal_dram_bw_strobe = %u\n", __func__, use_ideal_dram_bw_strobe);
5780 	dml_print("DML::%s: HostVMEnable = %u\n", __func__, HostVMEnable);
5781 	dml_print("DML::%s: DCFCLK = %f\n", __func__, DCFCLK);
5782 	dml_print("DML::%s: FabricClock = %f\n", __func__, FabricClock);
5783 	dml_print("DML::%s: DRAMSpeed = %f\n", __func__, DRAMSpeed);
5784 	dml_print("DML::%s: VMDataOnlyReturnBW = %f\n", __func__, VMDataOnlyReturnBW);
5785 #endif
5786 	return VMDataOnlyReturnBW;
5787 }
5788 
5789 // Function: dml_get_return_bw_mbps
5790 // Megabyte per second
dml_get_return_bw_mbps(const struct soc_bounding_box_st * soc,dml_bool_t use_ideal_dram_bw_strobe,dml_bool_t HostVMEnable,dml_float_t DCFCLK,dml_float_t FabricClock,dml_float_t DRAMSpeed)5791 dml_float_t dml_get_return_bw_mbps(
5792 						const struct soc_bounding_box_st *soc,
5793 						dml_bool_t use_ideal_dram_bw_strobe,
5794 						dml_bool_t HostVMEnable,
5795 						dml_float_t DCFCLK,
5796 						dml_float_t FabricClock,
5797 						dml_float_t DRAMSpeed)
5798 {
5799 	dml_float_t ReturnBW = 0.;
5800 	dml_float_t IdealSDPPortBandwidth    = soc->return_bus_width_bytes * DCFCLK;
5801 	dml_float_t IdealFabricBandwidth     = FabricClock * soc->fabric_datapath_to_dcn_data_return_bytes;
5802 	dml_float_t IdealDRAMBandwidth       = DRAMSpeed * soc->num_chans * soc->dram_channel_width_bytes;
5803 	dml_float_t PixelDataOnlyReturnBW    = dml_min3(IdealSDPPortBandwidth * soc->pct_ideal_sdp_bw_after_urgent / 100,
5804 												IdealFabricBandwidth * soc->pct_ideal_fabric_bw_after_urgent / 100,
5805 												IdealDRAMBandwidth * ((use_ideal_dram_bw_strobe && !HostVMEnable) ? soc->pct_ideal_dram_bw_after_urgent_strobe :
5806 																						soc->pct_ideal_dram_bw_after_urgent_pixel_only) / 100);
5807 	dml_float_t PixelMixedWithVMDataReturnBW = dml_min3(IdealSDPPortBandwidth * soc->pct_ideal_sdp_bw_after_urgent / 100,
5808 													IdealFabricBandwidth * soc->pct_ideal_fabric_bw_after_urgent / 100,
5809 													IdealDRAMBandwidth * ((use_ideal_dram_bw_strobe && !HostVMEnable) ? soc->pct_ideal_dram_bw_after_urgent_strobe :
5810 																							soc->pct_ideal_dram_bw_after_urgent_pixel_and_vm) / 100);
5811 
5812 	if (HostVMEnable != true) {
5813 		ReturnBW = PixelDataOnlyReturnBW;
5814 	} else {
5815 		ReturnBW = PixelMixedWithVMDataReturnBW;
5816 	}
5817 
5818 #ifdef __DML_VBA_DEBUG__
5819 	dml_print("DML::%s: use_ideal_dram_bw_strobe = %u\n", __func__, use_ideal_dram_bw_strobe);
5820 	dml_print("DML::%s: HostVMEnable = %u\n", __func__, HostVMEnable);
5821 	dml_print("DML::%s: DCFCLK = %f\n", __func__, DCFCLK);
5822 	dml_print("DML::%s: FabricClock = %f\n", __func__, FabricClock);
5823 	dml_print("DML::%s: DRAMSpeed = %f\n", __func__, DRAMSpeed);
5824 	dml_print("DML::%s: IdealSDPPortBandwidth = %f\n", __func__, IdealSDPPortBandwidth);
5825 	dml_print("DML::%s: IdealFabricBandwidth = %f\n", __func__, IdealFabricBandwidth);
5826 	dml_print("DML::%s: IdealDRAMBandwidth = %f\n", __func__, IdealDRAMBandwidth);
5827 	dml_print("DML::%s: PixelDataOnlyReturnBW = %f\n", __func__, PixelDataOnlyReturnBW);
5828 	dml_print("DML::%s: PixelMixedWithVMDataReturnBW = %f\n", __func__, PixelMixedWithVMDataReturnBW);
5829 	dml_print("DML::%s: ReturnBW                  = %f MBps\n", __func__, ReturnBW);
5830 #endif
5831 	return ReturnBW;
5832 }
5833 
5834 // Function: dml_get_return_dram_bw_mbps
5835 // Megabyte per second
dml_get_return_dram_bw_mbps(const struct soc_bounding_box_st * soc,dml_bool_t use_ideal_dram_bw_strobe,dml_bool_t HostVMEnable,dml_float_t DRAMSpeed)5836 static dml_float_t dml_get_return_dram_bw_mbps(
5837 						const struct soc_bounding_box_st *soc,
5838 						dml_bool_t use_ideal_dram_bw_strobe,
5839 						dml_bool_t HostVMEnable,
5840 						dml_float_t DRAMSpeed)
5841 {
5842 	dml_float_t ReturnDRAMBW = 0.;
5843 	dml_float_t IdealDRAMBandwidth       = DRAMSpeed * soc->num_chans * soc->dram_channel_width_bytes;
5844 	dml_float_t PixelDataOnlyReturnBW    = IdealDRAMBandwidth * ((use_ideal_dram_bw_strobe && !HostVMEnable) ? soc->pct_ideal_dram_bw_after_urgent_strobe :
5845 																						soc->pct_ideal_dram_bw_after_urgent_pixel_only) / 100;
5846 	dml_float_t PixelMixedWithVMDataReturnBW =  IdealDRAMBandwidth * ((use_ideal_dram_bw_strobe && !HostVMEnable) ? soc->pct_ideal_dram_bw_after_urgent_strobe :
5847 																							soc->pct_ideal_dram_bw_after_urgent_pixel_and_vm) / 100;
5848 
5849 	if (HostVMEnable != true) {
5850 		ReturnDRAMBW = PixelDataOnlyReturnBW;
5851 	} else {
5852 		ReturnDRAMBW = PixelMixedWithVMDataReturnBW;
5853 	}
5854 
5855 #ifdef __DML_VBA_DEBUG__
5856 	dml_print("DML::%s: use_ideal_dram_bw_strobe = %u\n", __func__, use_ideal_dram_bw_strobe);
5857 	dml_print("DML::%s: HostVMEnable = %u\n", __func__, HostVMEnable);
5858 	dml_print("DML::%s: DRAMSpeed = %f\n", __func__, DRAMSpeed);
5859 	dml_print("DML::%s: IdealDRAMBandwidth = %f\n", __func__, IdealDRAMBandwidth);
5860 	dml_print("DML::%s: PixelDataOnlyReturnBW = %f\n", __func__, PixelDataOnlyReturnBW);
5861 	dml_print("DML::%s: PixelMixedWithVMDataReturnBW = %f\n", __func__, PixelMixedWithVMDataReturnBW);
5862 	dml_print("DML::%s: ReturnDRAMBW                     = %f MBps\n", __func__, ReturnDRAMBW);
5863 #endif
5864 	return ReturnDRAMBW;
5865 }
5866 
5867 /// @brief BACKEND
DSCDelayRequirement(dml_bool_t DSCEnabled,enum dml_odm_mode ODMMode,dml_uint_t DSCInputBitPerComponent,dml_float_t OutputBpp,dml_uint_t HActive,dml_uint_t HTotal,dml_uint_t NumberOfDSCSlices,enum dml_output_format_class OutputFormat,enum dml_output_encoder_class Output,dml_float_t PixelClock,dml_float_t PixelClockBackEnd)5868 static dml_uint_t DSCDelayRequirement(
5869 						dml_bool_t DSCEnabled,
5870 						enum dml_odm_mode ODMMode,
5871 						dml_uint_t DSCInputBitPerComponent,
5872 						dml_float_t OutputBpp,
5873 						dml_uint_t HActive,
5874 						dml_uint_t HTotal,
5875 						dml_uint_t NumberOfDSCSlices,
5876 						enum dml_output_format_class OutputFormat,
5877 						enum dml_output_encoder_class Output,
5878 						dml_float_t PixelClock,
5879 						dml_float_t PixelClockBackEnd)
5880 {
5881 	dml_uint_t DSCDelayRequirement_val = 0;
5882 
5883 	if (DSCEnabled == true && OutputBpp != 0) {
5884 		if (ODMMode == dml_odm_mode_combine_4to1) {
5885 			DSCDelayRequirement_val = dscceComputeDelay(DSCInputBitPerComponent, OutputBpp, (dml_uint_t)(dml_ceil((dml_float_t) HActive / (dml_float_t) NumberOfDSCSlices, 1.0)),
5886 												(dml_uint_t) (NumberOfDSCSlices / 4.0), OutputFormat, Output) + dscComputeDelay(OutputFormat, Output);
5887 		} else if (ODMMode == dml_odm_mode_combine_2to1) {
5888 			DSCDelayRequirement_val = dscceComputeDelay(DSCInputBitPerComponent, OutputBpp, (dml_uint_t)(dml_ceil((dml_float_t) HActive / (dml_float_t) NumberOfDSCSlices, 1.0)),
5889 												(dml_uint_t) (NumberOfDSCSlices / 2.0), OutputFormat, Output) + dscComputeDelay(OutputFormat, Output);
5890 		} else {
5891 			DSCDelayRequirement_val = dscceComputeDelay(DSCInputBitPerComponent, OutputBpp, (dml_uint_t)((dml_float_t) dml_ceil(HActive / (dml_float_t) NumberOfDSCSlices, 1.0)),
5892 										NumberOfDSCSlices, OutputFormat, Output) + dscComputeDelay(OutputFormat, Output);
5893 		}
5894 		DSCDelayRequirement_val = (dml_uint_t)(DSCDelayRequirement_val + (HTotal - HActive) * dml_ceil((dml_float_t) DSCDelayRequirement_val / (dml_float_t) HActive, 1.0));
5895 		DSCDelayRequirement_val = (dml_uint_t)(DSCDelayRequirement_val * PixelClock / PixelClockBackEnd);
5896 
5897 	} else {
5898 		DSCDelayRequirement_val = 0;
5899 	}
5900 #ifdef __DML_VBA_DEBUG__
5901 	dml_print("DML::%s: DSCEnabled = %u\n", __func__, DSCEnabled);
5902 	dml_print("DML::%s: ODMMode = %u\n", __func__, ODMMode);
5903 	dml_print("DML::%s: OutputBpp = %f\n", __func__, OutputBpp);
5904 	dml_print("DML::%s: HActive = %u\n", __func__, HActive);
5905 	dml_print("DML::%s: HTotal = %u\n", __func__, HTotal);
5906 	dml_print("DML::%s: PixelClock = %f\n", __func__, PixelClock);
5907 	dml_print("DML::%s: PixelClockBackEnd = %f\n", __func__, PixelClockBackEnd);
5908 	dml_print("DML::%s: OutputFormat = %u\n", __func__, OutputFormat);
5909 	dml_print("DML::%s: DSCInputBitPerComponent = %u\n", __func__, DSCInputBitPerComponent);
5910 	dml_print("DML::%s: NumberOfDSCSlices = %u\n", __func__, NumberOfDSCSlices);
5911 	dml_print("DML::%s: DSCDelayRequirement_val = %u\n", __func__, DSCDelayRequirement_val);
5912 #endif
5913 
5914 	return DSCDelayRequirement_val;
5915 }
5916 
CalculateVActiveBandwithSupport(dml_uint_t NumberOfActiveSurfaces,dml_float_t ReturnBW,dml_bool_t NotUrgentLatencyHiding[],dml_float_t ReadBandwidthLuma[],dml_float_t ReadBandwidthChroma[],dml_float_t cursor_bw[],dml_float_t meta_row_bandwidth[],dml_float_t dpte_row_bandwidth[],dml_uint_t NumberOfDPP[],dml_float_t UrgentBurstFactorLuma[],dml_float_t UrgentBurstFactorChroma[],dml_float_t UrgentBurstFactorCursor[])5917 static dml_bool_t CalculateVActiveBandwithSupport(dml_uint_t NumberOfActiveSurfaces,
5918 										dml_float_t ReturnBW,
5919 										dml_bool_t NotUrgentLatencyHiding[],
5920 										dml_float_t ReadBandwidthLuma[],
5921 										dml_float_t ReadBandwidthChroma[],
5922 										dml_float_t cursor_bw[],
5923 										dml_float_t meta_row_bandwidth[],
5924 										dml_float_t dpte_row_bandwidth[],
5925 										dml_uint_t NumberOfDPP[],
5926 										dml_float_t UrgentBurstFactorLuma[],
5927 										dml_float_t UrgentBurstFactorChroma[],
5928 										dml_float_t UrgentBurstFactorCursor[])
5929 {
5930 	dml_bool_t NotEnoughUrgentLatencyHiding = false;
5931 	dml_bool_t CalculateVActiveBandwithSupport_val = false;
5932 	dml_float_t VActiveBandwith = 0;
5933 
5934 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
5935 		if (NotUrgentLatencyHiding[k]) {
5936 			NotEnoughUrgentLatencyHiding = true;
5937 		}
5938 	}
5939 
5940 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
5941 		VActiveBandwith = VActiveBandwith + ReadBandwidthLuma[k] * UrgentBurstFactorLuma[k] + ReadBandwidthChroma[k] * UrgentBurstFactorChroma[k] + cursor_bw[k] * UrgentBurstFactorCursor[k] + NumberOfDPP[k] * meta_row_bandwidth[k] + NumberOfDPP[k] * dpte_row_bandwidth[k];
5942 	}
5943 
5944 	CalculateVActiveBandwithSupport_val = (VActiveBandwith <= ReturnBW) && !NotEnoughUrgentLatencyHiding;
5945 
5946 #ifdef __DML_VBA_DEBUG__
5947 	dml_print("DML::%s: NotEnoughUrgentLatencyHiding        = %u\n", __func__, NotEnoughUrgentLatencyHiding);
5948 	dml_print("DML::%s: VActiveBandwith                     = %f\n", __func__, VActiveBandwith);
5949 	dml_print("DML::%s: ReturnBW                            = %f\n", __func__, ReturnBW);
5950 	dml_print("DML::%s: CalculateVActiveBandwithSupport_val = %u\n", __func__, CalculateVActiveBandwithSupport_val);
5951 #endif
5952 	return CalculateVActiveBandwithSupport_val;
5953 }
5954 
CalculatePrefetchBandwithSupport(dml_uint_t NumberOfActiveSurfaces,dml_float_t ReturnBW,enum dml_use_mall_for_pstate_change_mode UseMALLForPStateChange[],dml_bool_t NotUrgentLatencyHiding[],dml_float_t ReadBandwidthLuma[],dml_float_t ReadBandwidthChroma[],dml_float_t PrefetchBandwidthLuma[],dml_float_t PrefetchBandwidthChroma[],dml_float_t cursor_bw[],dml_float_t meta_row_bandwidth[],dml_float_t dpte_row_bandwidth[],dml_float_t cursor_bw_pre[],dml_float_t prefetch_vmrow_bw[],dml_uint_t NumberOfDPP[],dml_float_t UrgentBurstFactorLuma[],dml_float_t UrgentBurstFactorChroma[],dml_float_t UrgentBurstFactorCursor[],dml_float_t UrgentBurstFactorLumaPre[],dml_float_t UrgentBurstFactorChromaPre[],dml_float_t UrgentBurstFactorCursorPre[],dml_float_t * PrefetchBandwidth,dml_float_t * PrefetchBandwidthNotIncludingMALLPrefetch,dml_float_t * FractionOfUrgentBandwidth,dml_bool_t * PrefetchBandwidthSupport)5955 static void CalculatePrefetchBandwithSupport(
5956 										dml_uint_t NumberOfActiveSurfaces,
5957 										dml_float_t ReturnBW,
5958 										enum dml_use_mall_for_pstate_change_mode UseMALLForPStateChange[],
5959 										dml_bool_t NotUrgentLatencyHiding[],
5960 										dml_float_t ReadBandwidthLuma[],
5961 										dml_float_t ReadBandwidthChroma[],
5962 										dml_float_t PrefetchBandwidthLuma[],
5963 										dml_float_t PrefetchBandwidthChroma[],
5964 										dml_float_t cursor_bw[],
5965 										dml_float_t meta_row_bandwidth[],
5966 										dml_float_t dpte_row_bandwidth[],
5967 										dml_float_t cursor_bw_pre[],
5968 										dml_float_t prefetch_vmrow_bw[],
5969 										dml_uint_t NumberOfDPP[],
5970 										dml_float_t UrgentBurstFactorLuma[],
5971 										dml_float_t UrgentBurstFactorChroma[],
5972 										dml_float_t UrgentBurstFactorCursor[],
5973 										dml_float_t UrgentBurstFactorLumaPre[],
5974 										dml_float_t UrgentBurstFactorChromaPre[],
5975 										dml_float_t UrgentBurstFactorCursorPre[],
5976 
5977 										// Output
5978 										dml_float_t *PrefetchBandwidth,
5979 										dml_float_t *PrefetchBandwidthNotIncludingMALLPrefetch,
5980 										dml_float_t *FractionOfUrgentBandwidth,
5981 										dml_bool_t *PrefetchBandwidthSupport)
5982 {
5983 	dml_bool_t NotEnoughUrgentLatencyHiding = false;
5984 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
5985 		if (NotUrgentLatencyHiding[k]) {
5986 			NotEnoughUrgentLatencyHiding = true;
5987 		}
5988 	}
5989 
5990 	*PrefetchBandwidth = 0;
5991 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
5992 		*PrefetchBandwidth = *PrefetchBandwidth + dml_max3(NumberOfDPP[k] * prefetch_vmrow_bw[k],
5993 														ReadBandwidthLuma[k] * UrgentBurstFactorLuma[k] + ReadBandwidthChroma[k] * UrgentBurstFactorChroma[k] + cursor_bw[k] * UrgentBurstFactorCursor[k] + NumberOfDPP[k] * (meta_row_bandwidth[k] + dpte_row_bandwidth[k]),
5994 														NumberOfDPP[k] * (PrefetchBandwidthLuma[k] * UrgentBurstFactorLumaPre[k] + PrefetchBandwidthChroma[k] * UrgentBurstFactorChromaPre[k]) + cursor_bw_pre[k] * UrgentBurstFactorCursorPre[k]);
5995 	}
5996 
5997 	*PrefetchBandwidthNotIncludingMALLPrefetch = 0;
5998 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
5999 		if (UseMALLForPStateChange[k] != dml_use_mall_pstate_change_phantom_pipe)
6000 			*PrefetchBandwidthNotIncludingMALLPrefetch = *PrefetchBandwidthNotIncludingMALLPrefetch
6001 				+ dml_max3(NumberOfDPP[k] * prefetch_vmrow_bw[k],
6002 				ReadBandwidthLuma[k] * UrgentBurstFactorLuma[k] + ReadBandwidthChroma[k] * UrgentBurstFactorChroma[k]
6003 				+ cursor_bw[k] * UrgentBurstFactorCursor[k]
6004 				+ NumberOfDPP[k] * (meta_row_bandwidth[k] + dpte_row_bandwidth[k]),
6005 				NumberOfDPP[k] * (PrefetchBandwidthLuma[k] * UrgentBurstFactorLumaPre[k]
6006 				+ PrefetchBandwidthChroma[k] * UrgentBurstFactorChromaPre[k])
6007 				+ cursor_bw_pre[k] * UrgentBurstFactorCursorPre[k]);
6008 	}
6009 
6010 	*PrefetchBandwidthSupport = (*PrefetchBandwidth <= ReturnBW) && !NotEnoughUrgentLatencyHiding;
6011 	*FractionOfUrgentBandwidth = *PrefetchBandwidth / ReturnBW;
6012 
6013 #ifdef __DML_VBA_DEBUG__
6014 	dml_print("DML::%s: ReturnBW = %f\n", __func__, ReturnBW);
6015 	dml_print("DML::%s: PrefetchBandwidth = %f\n", __func__, *PrefetchBandwidth);
6016 	dml_print("DML::%s: FractionOfUrgentBandwidth = %f\n", __func__, *FractionOfUrgentBandwidth);
6017    dml_print("DML::%s: PrefetchBandwidthSupport = %u\n", __func__, *PrefetchBandwidthSupport);
6018 #endif
6019 }
6020 
CalculateBandwidthAvailableForImmediateFlip(dml_uint_t NumberOfActiveSurfaces,dml_float_t ReturnBW,dml_float_t ReadBandwidthLuma[],dml_float_t ReadBandwidthChroma[],dml_float_t PrefetchBandwidthLuma[],dml_float_t PrefetchBandwidthChroma[],dml_float_t cursor_bw[],dml_float_t cursor_bw_pre[],dml_uint_t NumberOfDPP[],dml_float_t UrgentBurstFactorLuma[],dml_float_t UrgentBurstFactorChroma[],dml_float_t UrgentBurstFactorCursor[],dml_float_t UrgentBurstFactorLumaPre[],dml_float_t UrgentBurstFactorChromaPre[],dml_float_t UrgentBurstFactorCursorPre[])6021 static dml_float_t CalculateBandwidthAvailableForImmediateFlip(
6022 													dml_uint_t NumberOfActiveSurfaces,
6023 													dml_float_t ReturnBW,
6024 													dml_float_t ReadBandwidthLuma[],
6025 													dml_float_t ReadBandwidthChroma[],
6026 													dml_float_t PrefetchBandwidthLuma[],
6027 													dml_float_t PrefetchBandwidthChroma[],
6028 													dml_float_t cursor_bw[],
6029 													dml_float_t cursor_bw_pre[],
6030 													dml_uint_t NumberOfDPP[],
6031 													dml_float_t UrgentBurstFactorLuma[],
6032 													dml_float_t UrgentBurstFactorChroma[],
6033 													dml_float_t UrgentBurstFactorCursor[],
6034 													dml_float_t UrgentBurstFactorLumaPre[],
6035 													dml_float_t UrgentBurstFactorChromaPre[],
6036 													dml_float_t UrgentBurstFactorCursorPre[])
6037 {
6038 	dml_float_t ret_val = ReturnBW;
6039 
6040 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
6041 		ret_val = ret_val - dml_max(ReadBandwidthLuma[k] * UrgentBurstFactorLuma[k] + ReadBandwidthChroma[k] * UrgentBurstFactorChroma[k] + cursor_bw[k] * UrgentBurstFactorCursor[k],
6042 									NumberOfDPP[k] * (PrefetchBandwidthLuma[k] * UrgentBurstFactorLumaPre[k] + PrefetchBandwidthChroma[k] * UrgentBurstFactorChromaPre[k]) +
6043 									cursor_bw_pre[k] * UrgentBurstFactorCursorPre[k]);
6044 #ifdef __DML_VBA_DEBUG__
6045 		dml_print("DML::%s: k=%u\n", __func__, k);
6046 		dml_print("DML::%s: NumberOfDPP = %u\n", __func__, NumberOfDPP[k]);
6047 		dml_print("DML::%s: ReadBandwidthLuma = %f\n", __func__, ReadBandwidthLuma[k]);
6048 		dml_print("DML::%s: UrgentBurstFactorLuma = %f\n", __func__, UrgentBurstFactorLuma[k]);
6049 		dml_print("DML::%s: ReadBandwidthChroma = %f\n", __func__, ReadBandwidthChroma[k]);
6050 		dml_print("DML::%s: UrgentBurstFactorChroma = %f\n", __func__, UrgentBurstFactorChroma[k]);
6051 		dml_print("DML::%s: cursor_bw = %f\n", __func__, cursor_bw[k]);
6052 		dml_print("DML::%s: UrgentBurstFactorCursor = %f\n", __func__, UrgentBurstFactorCursor[k]);
6053 
6054 		dml_print("DML::%s: PrefetchBandwidthLuma = %f\n", __func__, PrefetchBandwidthLuma[k]);
6055 		dml_print("DML::%s: UrgentBurstFactorLumaPre = %f\n", __func__, UrgentBurstFactorLumaPre[k]);
6056 		dml_print("DML::%s: PrefetchBandwidthChroma = %f\n", __func__, PrefetchBandwidthChroma[k]);
6057 		dml_print("DML::%s: UrgentBurstFactorChromaPre = %f\n", __func__, UrgentBurstFactorChromaPre[k]);
6058 		dml_print("DML::%s: cursor_bw_pre = %f\n", __func__, cursor_bw_pre[k]);
6059 		dml_print("DML::%s: UrgentBurstFactorCursorPre = %f\n", __func__, UrgentBurstFactorCursorPre[k]);
6060 		dml_print("DML::%s: ret_val              = %f\n", __func__, ret_val);
6061 #endif
6062 	}
6063 
6064 	return ret_val;
6065 }
6066 
CalculateImmediateFlipBandwithSupport(dml_uint_t NumberOfActiveSurfaces,dml_float_t ReturnBW,enum dml_use_mall_for_pstate_change_mode UseMALLForPStateChange[],enum dml_immediate_flip_requirement ImmediateFlipRequirement[],dml_float_t final_flip_bw[],dml_float_t ReadBandwidthLuma[],dml_float_t ReadBandwidthChroma[],dml_float_t PrefetchBandwidthLuma[],dml_float_t PrefetchBandwidthChroma[],dml_float_t cursor_bw[],dml_float_t meta_row_bandwidth[],dml_float_t dpte_row_bandwidth[],dml_float_t cursor_bw_pre[],dml_float_t prefetch_vmrow_bw[],dml_uint_t NumberOfDPP[],dml_float_t UrgentBurstFactorLuma[],dml_float_t UrgentBurstFactorChroma[],dml_float_t UrgentBurstFactorCursor[],dml_float_t UrgentBurstFactorLumaPre[],dml_float_t UrgentBurstFactorChromaPre[],dml_float_t UrgentBurstFactorCursorPre[],dml_float_t * TotalBandwidth,dml_float_t * TotalBandwidthNotIncludingMALLPrefetch,dml_float_t * FractionOfUrgentBandwidth,dml_bool_t * ImmediateFlipBandwidthSupport)6067 static void CalculateImmediateFlipBandwithSupport(
6068 											dml_uint_t NumberOfActiveSurfaces,
6069 											dml_float_t ReturnBW,
6070 											enum dml_use_mall_for_pstate_change_mode UseMALLForPStateChange[],
6071 											enum dml_immediate_flip_requirement ImmediateFlipRequirement[],
6072 											dml_float_t final_flip_bw[],
6073 											dml_float_t ReadBandwidthLuma[],
6074 											dml_float_t ReadBandwidthChroma[],
6075 											dml_float_t PrefetchBandwidthLuma[],
6076 											dml_float_t PrefetchBandwidthChroma[],
6077 											dml_float_t cursor_bw[],
6078 											dml_float_t meta_row_bandwidth[],
6079 											dml_float_t dpte_row_bandwidth[],
6080 											dml_float_t cursor_bw_pre[],
6081 											dml_float_t prefetch_vmrow_bw[],
6082 											dml_uint_t NumberOfDPP[],
6083 											dml_float_t UrgentBurstFactorLuma[],
6084 											dml_float_t UrgentBurstFactorChroma[],
6085 											dml_float_t UrgentBurstFactorCursor[],
6086 											dml_float_t UrgentBurstFactorLumaPre[],
6087 											dml_float_t UrgentBurstFactorChromaPre[],
6088 											dml_float_t UrgentBurstFactorCursorPre[],
6089 
6090 											// Output
6091 											dml_float_t *TotalBandwidth,
6092 											dml_float_t *TotalBandwidthNotIncludingMALLPrefetch,
6093 											dml_float_t *FractionOfUrgentBandwidth,
6094 											dml_bool_t *ImmediateFlipBandwidthSupport)
6095 {
6096 	*TotalBandwidth = 0;
6097 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
6098 		if (ImmediateFlipRequirement[k] != dml_immediate_flip_not_required) {
6099 
6100 
6101 
6102 			*TotalBandwidth = *TotalBandwidth + dml_max3(NumberOfDPP[k] * prefetch_vmrow_bw[k],
6103 														NumberOfDPP[k] * final_flip_bw[k] + ReadBandwidthLuma[k] * UrgentBurstFactorLuma[k] + ReadBandwidthChroma[k] * UrgentBurstFactorChroma[k] + cursor_bw[k] * UrgentBurstFactorCursor[k],
6104 														NumberOfDPP[k] * (final_flip_bw[k] + PrefetchBandwidthLuma[k] * UrgentBurstFactorLumaPre[k] + PrefetchBandwidthChroma[k] * UrgentBurstFactorChromaPre[k]) + cursor_bw_pre[k] * UrgentBurstFactorCursorPre[k]);
6105 		} else {
6106 			*TotalBandwidth = *TotalBandwidth + dml_max3(NumberOfDPP[k] * prefetch_vmrow_bw[k],
6107 														NumberOfDPP[k] * (meta_row_bandwidth[k] + dpte_row_bandwidth[k]) + ReadBandwidthLuma[k] * UrgentBurstFactorLuma[k] + ReadBandwidthChroma[k] * UrgentBurstFactorChroma[k] + cursor_bw[k] * UrgentBurstFactorCursor[k],
6108 														NumberOfDPP[k] * (PrefetchBandwidthLuma[k] * UrgentBurstFactorLumaPre[k] + PrefetchBandwidthChroma[k] * UrgentBurstFactorChromaPre[k]) + cursor_bw_pre[k] * UrgentBurstFactorCursorPre[k]);
6109 		}
6110 #ifdef __DML_VBA_DEBUG__
6111 		dml_print("DML::%s: k = %u\n", __func__, k);
6112 		dml_print("DML::%s: ImmediateFlipRequirement = %u\n", __func__, ImmediateFlipRequirement[k]);
6113 		dml_print("DML::%s: TotalBandwidth = %f\n", __func__, *TotalBandwidth);
6114 		dml_print("DML::%s: NumberOfDPP = %u\n", __func__, NumberOfDPP[k]);
6115 		dml_print("DML::%s: prefetch_vmrow_bw = %f\n", __func__, prefetch_vmrow_bw[k]);
6116 		dml_print("DML::%s: final_flip_bw = %f\n", __func__, final_flip_bw[k]);
6117 		dml_print("DML::%s: ReadBandwidthLuma = %f\n", __func__, ReadBandwidthLuma[k]);
6118 		dml_print("DML::%s: UrgentBurstFactorLuma = %f\n", __func__, UrgentBurstFactorLuma[k]);
6119 		dml_print("DML::%s: ReadBandwidthChroma = %f\n", __func__, ReadBandwidthChroma[k]);
6120 		dml_print("DML::%s: UrgentBurstFactorChroma = %f\n", __func__, UrgentBurstFactorChroma[k]);
6121 		dml_print("DML::%s: cursor_bw = %f\n", __func__, cursor_bw[k]);
6122 		dml_print("DML::%s: UrgentBurstFactorCursor = %f\n", __func__, UrgentBurstFactorCursor[k]);
6123 		dml_print("DML::%s: PrefetchBandwidthLuma = %f\n", __func__, PrefetchBandwidthLuma[k]);
6124 		dml_print("DML::%s: UrgentBurstFactorLumaPre = %f\n", __func__, UrgentBurstFactorLumaPre[k]);
6125 		dml_print("DML::%s: PrefetchBandwidthChroma = %f\n", __func__, PrefetchBandwidthChroma[k]);
6126 		dml_print("DML::%s: UrgentBurstFactorChromaPre = %f\n", __func__, UrgentBurstFactorChromaPre[k]);
6127 		dml_print("DML::%s: cursor_bw_pre = %f\n", __func__, cursor_bw_pre[k]);
6128 		dml_print("DML::%s: UrgentBurstFactorCursorPre = %f\n", __func__, UrgentBurstFactorCursorPre[k]);
6129 		dml_print("DML::%s: meta_row_bandwidth = %f\n", __func__, meta_row_bandwidth[k]);
6130 		dml_print("DML::%s: dpte_row_bandwidth          = %f\n", __func__, dpte_row_bandwidth[k]);
6131 #endif
6132 	}
6133 
6134 	*TotalBandwidthNotIncludingMALLPrefetch = 0;
6135 	for (dml_uint_t k = 0; k < NumberOfActiveSurfaces; ++k) {
6136 		if (UseMALLForPStateChange[k] != dml_use_mall_pstate_change_phantom_pipe) {
6137 			if (ImmediateFlipRequirement[k] != dml_immediate_flip_not_required)
6138 				*TotalBandwidthNotIncludingMALLPrefetch = *TotalBandwidthNotIncludingMALLPrefetch + dml_max3(NumberOfDPP[k] * prefetch_vmrow_bw[k],
6139 					NumberOfDPP[k] * final_flip_bw[k] + ReadBandwidthLuma[k] * UrgentBurstFactorLuma[k] + ReadBandwidthChroma[k] * UrgentBurstFactorChroma[k] + cursor_bw[k] * UrgentBurstFactorCursor[k],
6140 					NumberOfDPP[k] * (final_flip_bw[k] + PrefetchBandwidthLuma[k] * UrgentBurstFactorLumaPre[k] + PrefetchBandwidthChroma[k] * UrgentBurstFactorChromaPre[k])
6141 					+ cursor_bw_pre[k] * UrgentBurstFactorCursorPre[k]);
6142 			else
6143 				*TotalBandwidthNotIncludingMALLPrefetch = *TotalBandwidthNotIncludingMALLPrefetch + dml_max3(NumberOfDPP[k] * prefetch_vmrow_bw[k],
6144 					NumberOfDPP[k] * (meta_row_bandwidth[k] + dpte_row_bandwidth[k])
6145 					+ ReadBandwidthLuma[k] * UrgentBurstFactorLuma[k] + ReadBandwidthChroma[k] * UrgentBurstFactorChroma[k] + cursor_bw[k] * UrgentBurstFactorCursor[k],
6146 					NumberOfDPP[k] * (PrefetchBandwidthLuma[k] * UrgentBurstFactorLumaPre[k] + PrefetchBandwidthChroma[k] * UrgentBurstFactorChromaPre[k])
6147 					+ cursor_bw_pre[k] * UrgentBurstFactorCursorPre[k]);
6148 		}
6149 	}
6150 
6151 	*ImmediateFlipBandwidthSupport = (*TotalBandwidth <= ReturnBW);
6152 	*FractionOfUrgentBandwidth = *TotalBandwidth / ReturnBW;
6153 #ifdef __DML_VBA_DEBUG__
6154 	dml_print("DML::%s: ReturnBW = %f\n", __func__, ReturnBW);
6155 	dml_print("DML::%s: TotalBandwidth = %f\n", __func__, *TotalBandwidth);
6156 	dml_print("DML::%s: ImmediateFlipBandwidthSupport = %u\n", __func__, *ImmediateFlipBandwidthSupport);
6157 #endif
6158 }
6159 
MicroSecToVertLines(dml_uint_t num_us,dml_uint_t h_total,dml_float_t pixel_clock)6160 static dml_uint_t MicroSecToVertLines(dml_uint_t num_us, dml_uint_t h_total, dml_float_t pixel_clock)
6161 {
6162 	dml_uint_t lines_time_in_ns = 1000.0 * (h_total * 1000.0) / (pixel_clock * 1000.0);
6163 
6164 	return dml_ceil(1000.0 * num_us / lines_time_in_ns, 1.0);
6165 }
6166 
6167 /// @brief Calculate the maximum vstartup for mode support and mode programming consideration
6168 ///         Bounded by min of actual vblank and input vblank_nom, dont want vstartup/ready to start too early if actual vbllank is huge
CalculateMaxVStartup(dml_uint_t plane_idx,dml_bool_t ptoi_supported,dml_uint_t vblank_nom_default_us,struct dml_timing_cfg_st * timing,dml_float_t write_back_delay_us)6169 static dml_uint_t CalculateMaxVStartup(
6170 			dml_uint_t          plane_idx,
6171 			dml_bool_t          ptoi_supported,
6172 			dml_uint_t          vblank_nom_default_us,
6173 			struct dml_timing_cfg_st  *timing,
6174 			dml_float_t         write_back_delay_us)
6175 {
6176 	dml_uint_t vblank_size = 0;
6177 	dml_uint_t max_vstartup_lines = 0;
6178 	const dml_uint_t max_allowed_vblank_nom = 1023;
6179 
6180 	dml_float_t line_time_us = (dml_float_t) timing->HTotal[plane_idx] / timing->PixelClock[plane_idx];
6181 	dml_uint_t vblank_actual = timing->VTotal[plane_idx] - timing->VActive[plane_idx];
6182 
6183 	dml_uint_t vblank_nom_default_in_line = MicroSecToVertLines(vblank_nom_default_us, timing->HTotal[plane_idx],
6184 			timing->PixelClock[plane_idx]);
6185 	dml_uint_t vblank_nom_input = (dml_uint_t)dml_min(vblank_actual, vblank_nom_default_in_line);
6186 
6187 	// vblank_nom should not be smaller than (VSync (VTotal - VActive - VFrontPorch) + 2)
6188 	// + 2 is because
6189 	// 1 -> VStartup_start should be 1 line before VSync
6190 	// 1 -> always reserve 1 line between start of VBlank to VStartup signal
6191 	dml_uint_t vblank_nom_vsync_capped = dml_max(vblank_nom_input,
6192 			timing->VTotal[plane_idx] - timing->VActive[plane_idx] - timing->VFrontPorch[plane_idx] + 2);
6193 	dml_uint_t vblank_nom_max_allowed_capped = dml_min(vblank_nom_vsync_capped, max_allowed_vblank_nom);
6194 	dml_uint_t vblank_avail = (vblank_nom_max_allowed_capped == 0) ?
6195 			vblank_nom_default_in_line : vblank_nom_max_allowed_capped;
6196 
6197 	vblank_size = (dml_uint_t) dml_min(vblank_actual, vblank_avail);
6198 
6199    if (timing->Interlace[plane_idx] && !ptoi_supported)
6200 		max_vstartup_lines = (dml_uint_t) (dml_floor(vblank_size/2.0, 1.0));
6201 	else
6202 		max_vstartup_lines = vblank_size - (dml_uint_t) dml_max(1.0, dml_ceil(write_back_delay_us/line_time_us, 1.0));
6203 #ifdef __DML_VBA_DEBUG__
6204 	dml_print("DML::%s: plane_idx = %u\n", __func__, plane_idx);
6205 	dml_print("DML::%s: VBlankNom = %u\n", __func__, timing->VBlankNom[plane_idx]);
6206 	dml_print("DML::%s: vblank_nom_default_us = %u\n", __func__, vblank_nom_default_us);
6207 	dml_print("DML::%s: line_time_us = %f\n", __func__, line_time_us);
6208 	dml_print("DML::%s: vblank_actual = %u\n", __func__, vblank_actual);
6209 	dml_print("DML::%s: vblank_avail = %u\n", __func__, vblank_avail);
6210 	dml_print("DML::%s: max_vstartup_lines = %u\n", __func__, max_vstartup_lines);
6211 #endif
6212 	return max_vstartup_lines;
6213 }
6214 
set_calculate_prefetch_schedule_params(struct display_mode_lib_st * mode_lib,struct CalculatePrefetchSchedule_params_st * CalculatePrefetchSchedule_params,dml_uint_t j,dml_uint_t k)6215 static void set_calculate_prefetch_schedule_params(struct display_mode_lib_st *mode_lib,
6216 						   struct CalculatePrefetchSchedule_params_st *CalculatePrefetchSchedule_params,
6217 						   dml_uint_t j,
6218 						   dml_uint_t k)
6219 {
6220 				CalculatePrefetchSchedule_params->DSCDelay = mode_lib->ms.DSCDelayPerState[k];
6221 				CalculatePrefetchSchedule_params->EnhancedPrefetchScheduleAccelerationFinal = mode_lib->ms.policy.EnhancedPrefetchScheduleAccelerationFinal;
6222 				CalculatePrefetchSchedule_params->DPPCLKDelaySubtotalPlusCNVCFormater = mode_lib->ms.ip.dppclk_delay_subtotal + mode_lib->ms.ip.dppclk_delay_cnvc_formatter;
6223 				CalculatePrefetchSchedule_params->DPPCLKDelaySCL = mode_lib->ms.ip.dppclk_delay_scl;
6224 				CalculatePrefetchSchedule_params->DPPCLKDelaySCLLBOnly = mode_lib->ms.ip.dppclk_delay_scl_lb_only;
6225 				CalculatePrefetchSchedule_params->DPPCLKDelayCNVCCursor = mode_lib->ms.ip.dppclk_delay_cnvc_cursor;
6226 				CalculatePrefetchSchedule_params->DISPCLKDelaySubtotal = mode_lib->ms.ip.dispclk_delay_subtotal;
6227 				CalculatePrefetchSchedule_params->DPP_RECOUT_WIDTH = (dml_uint_t)(mode_lib->ms.SwathWidthYThisState[k] / mode_lib->ms.cache_display_cfg.plane.HRatio[k]);
6228 				CalculatePrefetchSchedule_params->OutputFormat = mode_lib->ms.cache_display_cfg.output.OutputFormat[k];
6229 				CalculatePrefetchSchedule_params->MaxInterDCNTileRepeaters = mode_lib->ms.ip.max_inter_dcn_tile_repeaters;
6230 				CalculatePrefetchSchedule_params->GPUVMPageTableLevels = mode_lib->ms.cache_display_cfg.plane.GPUVMMaxPageTableLevels;
6231 				CalculatePrefetchSchedule_params->GPUVMEnable = mode_lib->ms.cache_display_cfg.plane.GPUVMEnable;
6232 				CalculatePrefetchSchedule_params->HostVMEnable = mode_lib->ms.cache_display_cfg.plane.HostVMEnable;
6233 				CalculatePrefetchSchedule_params->HostVMMaxNonCachedPageTableLevels = mode_lib->ms.cache_display_cfg.plane.HostVMMaxPageTableLevels;
6234 				CalculatePrefetchSchedule_params->HostVMMinPageSize = mode_lib->ms.soc.hostvm_min_page_size_kbytes * 1024;
6235 				CalculatePrefetchSchedule_params->DynamicMetadataEnable = mode_lib->ms.cache_display_cfg.plane.DynamicMetadataEnable[k];
6236 				CalculatePrefetchSchedule_params->DynamicMetadataVMEnabled = mode_lib->ms.ip.dynamic_metadata_vm_enabled;
6237 				CalculatePrefetchSchedule_params->DynamicMetadataLinesBeforeActiveRequired = mode_lib->ms.cache_display_cfg.plane.DynamicMetadataLinesBeforeActiveRequired[k];
6238 				CalculatePrefetchSchedule_params->DynamicMetadataTransmittedBytes = mode_lib->ms.cache_display_cfg.plane.DynamicMetadataTransmittedBytes[k];
6239 				CalculatePrefetchSchedule_params->UrgentLatency = mode_lib->ms.UrgLatency;
6240 				CalculatePrefetchSchedule_params->UrgentExtraLatency = mode_lib->ms.ExtraLatency;
6241 				CalculatePrefetchSchedule_params->TCalc = mode_lib->ms.TimeCalc;
6242 				CalculatePrefetchSchedule_params->PDEAndMetaPTEBytesFrame = mode_lib->ms.PDEAndMetaPTEBytesPerFrame[j][k];
6243 				CalculatePrefetchSchedule_params->MetaRowByte = mode_lib->ms.MetaRowBytes[j][k];
6244 				CalculatePrefetchSchedule_params->PixelPTEBytesPerRow = mode_lib->ms.DPTEBytesPerRow[j][k];
6245 				CalculatePrefetchSchedule_params->PrefetchSourceLinesY = mode_lib->ms.PrefetchLinesY[j][k];
6246 				CalculatePrefetchSchedule_params->VInitPreFillY = mode_lib->ms.PrefillY[k];
6247 				CalculatePrefetchSchedule_params->MaxNumSwathY = mode_lib->ms.MaxNumSwY[k];
6248 				CalculatePrefetchSchedule_params->PrefetchSourceLinesC = mode_lib->ms.PrefetchLinesC[j][k];
6249 				CalculatePrefetchSchedule_params->VInitPreFillC = mode_lib->ms.PrefillC[k];
6250 				CalculatePrefetchSchedule_params->MaxNumSwathC = mode_lib->ms.MaxNumSwC[k];
6251 				CalculatePrefetchSchedule_params->swath_width_luma_ub = mode_lib->ms.swath_width_luma_ub_this_state[k];
6252 				CalculatePrefetchSchedule_params->swath_width_chroma_ub = mode_lib->ms.swath_width_chroma_ub_this_state[k];
6253 				CalculatePrefetchSchedule_params->SwathHeightY = mode_lib->ms.SwathHeightYThisState[k];
6254 				CalculatePrefetchSchedule_params->SwathHeightC = mode_lib->ms.SwathHeightCThisState[k];
6255 				CalculatePrefetchSchedule_params->TWait = mode_lib->ms.TWait;
6256 				CalculatePrefetchSchedule_params->DestinationLinesForPrefetch = &mode_lib->ms.LineTimesForPrefetch[k];
6257 				CalculatePrefetchSchedule_params->DestinationLinesToRequestVMInVBlank = &mode_lib->ms.LinesForMetaPTE[k];
6258 				CalculatePrefetchSchedule_params->DestinationLinesToRequestRowInVBlank = &mode_lib->ms.LinesForMetaAndDPTERow[k];
6259 				CalculatePrefetchSchedule_params->VRatioPrefetchY = &mode_lib->ms.VRatioPreY[j][k];
6260 				CalculatePrefetchSchedule_params->VRatioPrefetchC = &mode_lib->ms.VRatioPreC[j][k];
6261 				CalculatePrefetchSchedule_params->RequiredPrefetchPixDataBWLuma = &mode_lib->ms.RequiredPrefetchPixelDataBWLuma[k];
6262 				CalculatePrefetchSchedule_params->RequiredPrefetchPixDataBWChroma = &mode_lib->ms.RequiredPrefetchPixelDataBWChroma[k];
6263 				CalculatePrefetchSchedule_params->NotEnoughTimeForDynamicMetadata = &mode_lib->ms.support.NoTimeForDynamicMetadata[j][k];
6264 				CalculatePrefetchSchedule_params->Tno_bw = &mode_lib->ms.Tno_bw[k];
6265 }
6266 
dml_prefetch_check(struct display_mode_lib_st * mode_lib)6267 static void dml_prefetch_check(struct display_mode_lib_st *mode_lib)
6268 {
6269 	struct dml_core_mode_support_locals_st *s = &mode_lib->scratch.dml_core_mode_support_locals;
6270 	struct CalculatePrefetchSchedule_params_st *CalculatePrefetchSchedule_params = &mode_lib->scratch.CalculatePrefetchSchedule_params;
6271 	struct CalculateWatermarksMALLUseAndDRAMSpeedChangeSupport_params_st *CalculateWatermarks_params = &mode_lib->scratch.CalculateWatermarksMALLUseAndDRAMSpeedChangeSupport_params;
6272 	struct DmlPipe *myPipe;
6273 	dml_uint_t j, k;
6274 
6275 	for (j = 0; j < 2; ++j) {
6276 		mode_lib->ms.TimeCalc = 24 / mode_lib->ms.ProjectedDCFCLKDeepSleep[j];
6277 
6278 		for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
6279 			mode_lib->ms.NoOfDPPThisState[k] = mode_lib->ms.NoOfDPP[j][k];
6280 			mode_lib->ms.swath_width_luma_ub_this_state[k] = mode_lib->ms.swath_width_luma_ub_all_states[j][k];
6281 			mode_lib->ms.swath_width_chroma_ub_this_state[k] = mode_lib->ms.swath_width_chroma_ub_all_states[j][k];
6282 			mode_lib->ms.SwathWidthYThisState[k] = mode_lib->ms.SwathWidthYAllStates[j][k];
6283 			mode_lib->ms.SwathWidthCThisState[k] = mode_lib->ms.SwathWidthCAllStates[j][k];
6284 			mode_lib->ms.SwathHeightYThisState[k] = mode_lib->ms.SwathHeightYAllStates[j][k];
6285 			mode_lib->ms.SwathHeightCThisState[k] = mode_lib->ms.SwathHeightCAllStates[j][k];
6286 			mode_lib->ms.UnboundedRequestEnabledThisState = mode_lib->ms.UnboundedRequestEnabledAllStates[j];
6287 			mode_lib->ms.CompressedBufferSizeInkByteThisState = mode_lib->ms.CompressedBufferSizeInkByteAllStates[j];
6288 			mode_lib->ms.DETBufferSizeInKByteThisState[k] = mode_lib->ms.DETBufferSizeInKByteAllStates[j][k];
6289 			mode_lib->ms.DETBufferSizeYThisState[k] = mode_lib->ms.DETBufferSizeYAllStates[j][k];
6290 			mode_lib->ms.DETBufferSizeCThisState[k] = mode_lib->ms.DETBufferSizeCAllStates[j][k];
6291 		}
6292 
6293 		mode_lib->ms.support.VActiveBandwithSupport[j] = CalculateVActiveBandwithSupport(
6294 			mode_lib->ms.num_active_planes,
6295 			mode_lib->ms.ReturnBWPerState[j],
6296 			mode_lib->ms.NotUrgentLatencyHiding,
6297 			mode_lib->ms.ReadBandwidthLuma,
6298 			mode_lib->ms.ReadBandwidthChroma,
6299 			mode_lib->ms.cursor_bw,
6300 			mode_lib->ms.meta_row_bandwidth_this_state,
6301 			mode_lib->ms.dpte_row_bandwidth_this_state,
6302 			mode_lib->ms.NoOfDPPThisState,
6303 			mode_lib->ms.UrgentBurstFactorLuma,
6304 			mode_lib->ms.UrgentBurstFactorChroma,
6305 			mode_lib->ms.UrgentBurstFactorCursor);
6306 
6307 		s->VMDataOnlyReturnBWPerState = dml_get_return_bw_mbps_vm_only(
6308 																	&mode_lib->ms.soc,
6309 																	mode_lib->ms.state.use_ideal_dram_bw_strobe,
6310 																	mode_lib->ms.cache_display_cfg.plane.HostVMEnable,
6311 																	mode_lib->ms.DCFCLKState[j],
6312 																	mode_lib->ms.state.fabricclk_mhz,
6313 																	mode_lib->ms.state.dram_speed_mts);
6314 
6315 		s->HostVMInefficiencyFactor = 1;
6316 		if (mode_lib->ms.cache_display_cfg.plane.GPUVMEnable && mode_lib->ms.cache_display_cfg.plane.HostVMEnable)
6317 			s->HostVMInefficiencyFactor = mode_lib->ms.ReturnBWPerState[j] / s->VMDataOnlyReturnBWPerState;
6318 
6319 		mode_lib->ms.ExtraLatency = CalculateExtraLatency(
6320 				mode_lib->ms.soc.round_trip_ping_latency_dcfclk_cycles,
6321 				s->ReorderingBytes,
6322 				mode_lib->ms.DCFCLKState[j],
6323 				mode_lib->ms.TotalNumberOfActiveDPP[j],
6324 				mode_lib->ms.ip.pixel_chunk_size_kbytes,
6325 				mode_lib->ms.TotalNumberOfDCCActiveDPP[j],
6326 				mode_lib->ms.ip.meta_chunk_size_kbytes,
6327 				mode_lib->ms.ReturnBWPerState[j],
6328 				mode_lib->ms.cache_display_cfg.plane.GPUVMEnable,
6329 				mode_lib->ms.cache_display_cfg.plane.HostVMEnable,
6330 				mode_lib->ms.num_active_planes,
6331 				mode_lib->ms.NoOfDPPThisState,
6332 				mode_lib->ms.dpte_group_bytes,
6333 				s->HostVMInefficiencyFactor,
6334 				mode_lib->ms.soc.hostvm_min_page_size_kbytes * 1024,
6335 				mode_lib->ms.cache_display_cfg.plane.HostVMMaxPageTableLevels);
6336 
6337 		s->NextMaxVStartup = s->MaxVStartupAllPlanes[j];
6338 		s->MaxVStartup = 0;
6339 		s->AllPrefetchModeTested = true;
6340 		for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
6341 			CalculatePrefetchMode(mode_lib->ms.policy.AllowForPStateChangeOrStutterInVBlank[k], &s->MinPrefetchMode[k], &s->MaxPrefetchMode[k]);
6342 			s->NextPrefetchMode[k] = s->MinPrefetchMode[k];
6343 		}
6344 
6345 		do {
6346 			s->MaxVStartup = s->NextMaxVStartup;
6347 			s->AllPrefetchModeTested = true;
6348 
6349 			for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
6350 				mode_lib->ms.PrefetchMode[k] = s->NextPrefetchMode[k];
6351 				mode_lib->ms.TWait = CalculateTWait(
6352 								mode_lib->ms.PrefetchMode[k],
6353 								mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange[k],
6354 								mode_lib->ms.policy.SynchronizeDRRDisplaysForUCLKPStateChangeFinal,
6355 								mode_lib->ms.cache_display_cfg.timing.DRRDisplay[k],
6356 								mode_lib->ms.state.dram_clock_change_latency_us,
6357 								mode_lib->ms.state.fclk_change_latency_us,
6358 								mode_lib->ms.UrgLatency,
6359 								mode_lib->ms.state.sr_enter_plus_exit_time_us);
6360 
6361 				myPipe = &s->myPipe;
6362 				myPipe->Dppclk = mode_lib->ms.RequiredDPPCLKPerSurface[j][k];
6363 				myPipe->Dispclk = mode_lib->ms.RequiredDISPCLK[j];
6364 				myPipe->PixelClock = mode_lib->ms.cache_display_cfg.timing.PixelClock[k];
6365 				myPipe->DCFClkDeepSleep = mode_lib->ms.ProjectedDCFCLKDeepSleep[j];
6366 				myPipe->DPPPerSurface = mode_lib->ms.NoOfDPP[j][k];
6367 				myPipe->ScalerEnabled = mode_lib->ms.cache_display_cfg.plane.ScalerEnabled[k];
6368 				myPipe->SourceScan = mode_lib->ms.cache_display_cfg.plane.SourceScan[k];
6369 				myPipe->BlockWidth256BytesY = mode_lib->ms.Read256BlockWidthY[k];
6370 				myPipe->BlockHeight256BytesY = mode_lib->ms.Read256BlockHeightY[k];
6371 				myPipe->BlockWidth256BytesC = mode_lib->ms.Read256BlockWidthC[k];
6372 				myPipe->BlockHeight256BytesC = mode_lib->ms.Read256BlockHeightC[k];
6373 				myPipe->InterlaceEnable = mode_lib->ms.cache_display_cfg.timing.Interlace[k];
6374 				myPipe->NumberOfCursors = mode_lib->ms.cache_display_cfg.plane.NumberOfCursors[k];
6375 				myPipe->VBlank = mode_lib->ms.cache_display_cfg.timing.VTotal[k] - mode_lib->ms.cache_display_cfg.timing.VActive[k];
6376 				myPipe->HTotal = mode_lib->ms.cache_display_cfg.timing.HTotal[k];
6377 				myPipe->HActive = mode_lib->ms.cache_display_cfg.timing.HActive[k];
6378 				myPipe->DCCEnable = mode_lib->ms.cache_display_cfg.surface.DCCEnable[k];
6379 				myPipe->ODMMode = mode_lib->ms.ODMModePerState[k];
6380 				myPipe->SourcePixelFormat = mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k];
6381 				myPipe->BytePerPixelY = mode_lib->ms.BytePerPixelY[k];
6382 				myPipe->BytePerPixelC = mode_lib->ms.BytePerPixelC[k];
6383 				myPipe->ProgressiveToInterlaceUnitInOPP = mode_lib->ms.ip.ptoi_supported;
6384 
6385 #ifdef __DML_VBA_DEBUG__
6386 				dml_print("DML::%s: Calling CalculatePrefetchSchedule for j=%u, k=%u\n", __func__, j, k);
6387 				dml_print("DML::%s: MaximumVStartup = %u\n", __func__, s->MaximumVStartup[j][k]);
6388 				dml_print("DML::%s: MaxVStartup = %u\n", __func__, s->MaxVStartup);
6389 				dml_print("DML::%s: NextPrefetchMode = %u\n", __func__, s->NextPrefetchMode[k]);
6390 				dml_print("DML::%s: AllowForPStateChangeOrStutterInVBlank = %u\n", __func__, mode_lib->ms.policy.AllowForPStateChangeOrStutterInVBlank[k]);
6391 				dml_print("DML::%s: PrefetchMode = %u\n", __func__, mode_lib->ms.PrefetchMode[k]);
6392 #endif
6393 
6394 				CalculatePrefetchSchedule_params->HostVMInefficiencyFactor = s->HostVMInefficiencyFactor;
6395 				CalculatePrefetchSchedule_params->myPipe = myPipe;
6396 				CalculatePrefetchSchedule_params->VStartup = (dml_uint_t)(dml_min(s->MaxVStartup, s->MaximumVStartup[j][k]));
6397 				CalculatePrefetchSchedule_params->MaxVStartup = s->MaximumVStartup[j][k];
6398 				CalculatePrefetchSchedule_params->DSTXAfterScaler = &s->DSTXAfterScaler[k];
6399 				CalculatePrefetchSchedule_params->DSTYAfterScaler = &s->DSTYAfterScaler[k];
6400 				CalculatePrefetchSchedule_params->prefetch_vmrow_bw = &mode_lib->ms.prefetch_vmrow_bw[k];
6401 				CalculatePrefetchSchedule_params->Tdmdl_vm = &s->dummy_single[0];
6402 				CalculatePrefetchSchedule_params->Tdmdl = &s->dummy_single[1];
6403 				CalculatePrefetchSchedule_params->TSetup = &s->dummy_single[2];
6404 				CalculatePrefetchSchedule_params->VUpdateOffsetPix = &s->dummy_integer[0];
6405 				CalculatePrefetchSchedule_params->VUpdateWidthPix = &s->dummy_integer[1];
6406 				CalculatePrefetchSchedule_params->VReadyOffsetPix = &s->dummy_integer[2];
6407 
6408 				set_calculate_prefetch_schedule_params(mode_lib, CalculatePrefetchSchedule_params, j, k);
6409 
6410 				mode_lib->ms.support.NoTimeForPrefetch[j][k] =
6411 								CalculatePrefetchSchedule(&mode_lib->scratch,
6412 								CalculatePrefetchSchedule_params);
6413 			}
6414 
6415 			for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
6416 					CalculateUrgentBurstFactor(
6417 							mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange[k],
6418 							mode_lib->ms.swath_width_luma_ub_this_state[k],
6419 							mode_lib->ms.swath_width_chroma_ub_this_state[k],
6420 							mode_lib->ms.SwathHeightYThisState[k],
6421 							mode_lib->ms.SwathHeightCThisState[k],
6422 							mode_lib->ms.cache_display_cfg.timing.HTotal[k] / mode_lib->ms.cache_display_cfg.timing.PixelClock[k],
6423 							mode_lib->ms.UrgLatency,
6424 							mode_lib->ms.ip.cursor_buffer_size,
6425 							mode_lib->ms.cache_display_cfg.plane.CursorWidth[k],
6426 							mode_lib->ms.cache_display_cfg.plane.CursorBPP[k],
6427 							mode_lib->ms.VRatioPreY[j][k],
6428 							mode_lib->ms.VRatioPreC[j][k],
6429 							mode_lib->ms.BytePerPixelInDETY[k],
6430 							mode_lib->ms.BytePerPixelInDETC[k],
6431 							mode_lib->ms.DETBufferSizeYThisState[k],
6432 							mode_lib->ms.DETBufferSizeCThisState[k],
6433 							/* Output */
6434 							&mode_lib->ms.UrgentBurstFactorCursorPre[k],
6435 							&mode_lib->ms.UrgentBurstFactorLumaPre[k],
6436 							&mode_lib->ms.UrgentBurstFactorChroma[k],
6437 							&mode_lib->ms.NotUrgentLatencyHidingPre[k]);
6438 
6439 					mode_lib->ms.cursor_bw_pre[k] = mode_lib->ms.cache_display_cfg.plane.NumberOfCursors[k] * mode_lib->ms.cache_display_cfg.plane.CursorWidth[k] *
6440 													mode_lib->ms.cache_display_cfg.plane.CursorBPP[k] / 8.0 / (mode_lib->ms.cache_display_cfg.timing.HTotal[k] /
6441 													mode_lib->ms.cache_display_cfg.timing.PixelClock[k]) * mode_lib->ms.VRatioPreY[j][k];
6442 			}
6443 
6444 			{
6445 			CalculatePrefetchBandwithSupport(
6446 				mode_lib->ms.num_active_planes,
6447 				mode_lib->ms.ReturnBWPerState[j],
6448 				mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange,
6449 				mode_lib->ms.NotUrgentLatencyHidingPre,
6450 				mode_lib->ms.ReadBandwidthLuma,
6451 				mode_lib->ms.ReadBandwidthChroma,
6452 				mode_lib->ms.RequiredPrefetchPixelDataBWLuma,
6453 				mode_lib->ms.RequiredPrefetchPixelDataBWChroma,
6454 				mode_lib->ms.cursor_bw,
6455 				mode_lib->ms.meta_row_bandwidth_this_state,
6456 				mode_lib->ms.dpte_row_bandwidth_this_state,
6457 				mode_lib->ms.cursor_bw_pre,
6458 				mode_lib->ms.prefetch_vmrow_bw,
6459 				mode_lib->ms.NoOfDPPThisState,
6460 				mode_lib->ms.UrgentBurstFactorLuma,
6461 				mode_lib->ms.UrgentBurstFactorChroma,
6462 				mode_lib->ms.UrgentBurstFactorCursor,
6463 				mode_lib->ms.UrgentBurstFactorLumaPre,
6464 				mode_lib->ms.UrgentBurstFactorChromaPre,
6465 				mode_lib->ms.UrgentBurstFactorCursorPre,
6466 
6467 				/* output */
6468 				&s->dummy_single[0], // dml_float_t *PrefetchBandwidth
6469 				&s->dummy_single[1], // dml_float_t *PrefetchBandwidthNotIncludingMALLPrefetch
6470 				&mode_lib->mp.FractionOfUrgentBandwidth, // dml_float_t *FractionOfUrgentBandwidth
6471 				&mode_lib->ms.support.PrefetchSupported[j]);
6472 			}
6473 
6474 			for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
6475 				if (mode_lib->ms.LineTimesForPrefetch[k] < 2.0
6476 					|| mode_lib->ms.LinesForMetaPTE[k] >= 32.0
6477 					|| mode_lib->ms.LinesForMetaAndDPTERow[k] >= 16.0
6478 					|| mode_lib->ms.support.NoTimeForPrefetch[j][k] == true) {
6479 						mode_lib->ms.support.PrefetchSupported[j] = false;
6480 				}
6481 			}
6482 
6483 			mode_lib->ms.support.DynamicMetadataSupported[j] = true;
6484 			for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
6485 				if (mode_lib->ms.support.NoTimeForDynamicMetadata[j][k] == true) {
6486 					mode_lib->ms.support.DynamicMetadataSupported[j] = false;
6487 				}
6488 			}
6489 
6490 			mode_lib->ms.support.VRatioInPrefetchSupported[j] = true;
6491 			for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
6492 				if (mode_lib->ms.support.NoTimeForPrefetch[j][k] == true ||
6493 					mode_lib->ms.VRatioPreY[j][k] > __DML_MAX_VRATIO_PRE_ENHANCE_PREFETCH_ACC__ ||
6494 					mode_lib->ms.VRatioPreC[j][k] > __DML_MAX_VRATIO_PRE_ENHANCE_PREFETCH_ACC__ ||
6495 					((s->MaxVStartup < s->MaximumVStartup[j][k] || mode_lib->ms.policy.EnhancedPrefetchScheduleAccelerationFinal == 0) &&
6496 						(mode_lib->ms.VRatioPreY[j][k] > __DML_MAX_VRATIO_PRE__ || mode_lib->ms.VRatioPreC[j][k] > __DML_MAX_VRATIO_PRE__))) {
6497 							mode_lib->ms.support.VRatioInPrefetchSupported[j] = false;
6498 				}
6499 			}
6500 
6501 			s->AnyLinesForVMOrRowTooLarge = false;
6502 			for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
6503 				if (mode_lib->ms.LinesForMetaAndDPTERow[k] >= 16 || mode_lib->ms.LinesForMetaPTE[k] >= 32) {
6504 					s->AnyLinesForVMOrRowTooLarge = true;
6505 				}
6506 			}
6507 
6508 			if (mode_lib->ms.support.PrefetchSupported[j] == true && mode_lib->ms.support.VRatioInPrefetchSupported[j] == true) {
6509 				mode_lib->ms.BandwidthAvailableForImmediateFlip = CalculateBandwidthAvailableForImmediateFlip(
6510 						mode_lib->ms.num_active_planes,
6511 						mode_lib->ms.ReturnBWPerState[j],
6512 						mode_lib->ms.ReadBandwidthLuma,
6513 						mode_lib->ms.ReadBandwidthChroma,
6514 						mode_lib->ms.RequiredPrefetchPixelDataBWLuma,
6515 						mode_lib->ms.RequiredPrefetchPixelDataBWChroma,
6516 						mode_lib->ms.cursor_bw,
6517 						mode_lib->ms.cursor_bw_pre,
6518 						mode_lib->ms.NoOfDPPThisState,
6519 						mode_lib->ms.UrgentBurstFactorLuma,
6520 						mode_lib->ms.UrgentBurstFactorChroma,
6521 						mode_lib->ms.UrgentBurstFactorCursor,
6522 						mode_lib->ms.UrgentBurstFactorLumaPre,
6523 						mode_lib->ms.UrgentBurstFactorChromaPre,
6524 						mode_lib->ms.UrgentBurstFactorCursorPre);
6525 
6526 				mode_lib->ms.TotImmediateFlipBytes = 0;
6527 				for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
6528 					if (!(mode_lib->ms.policy.ImmediateFlipRequirement[k] == dml_immediate_flip_not_required)) {
6529 						mode_lib->ms.TotImmediateFlipBytes = mode_lib->ms.TotImmediateFlipBytes + mode_lib->ms.NoOfDPP[j][k] * mode_lib->ms.PDEAndMetaPTEBytesPerFrame[j][k] + mode_lib->ms.MetaRowBytes[j][k];
6530 						if (mode_lib->ms.use_one_row_for_frame_flip[j][k]) {
6531 							mode_lib->ms.TotImmediateFlipBytes = mode_lib->ms.TotImmediateFlipBytes + mode_lib->ms.NoOfDPP[j][k] * (2 * mode_lib->ms.DPTEBytesPerRow[j][k]);
6532 						} else {
6533 							mode_lib->ms.TotImmediateFlipBytes = mode_lib->ms.TotImmediateFlipBytes + mode_lib->ms.NoOfDPP[j][k] * mode_lib->ms.DPTEBytesPerRow[j][k];
6534 						}
6535 					}
6536 				}
6537 
6538 				for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
6539 					CalculateFlipSchedule(
6540 						s->HostVMInefficiencyFactor,
6541 						mode_lib->ms.ExtraLatency,
6542 						mode_lib->ms.UrgLatency,
6543 						mode_lib->ms.cache_display_cfg.plane.GPUVMMaxPageTableLevels,
6544 						mode_lib->ms.cache_display_cfg.plane.HostVMEnable,
6545 						mode_lib->ms.cache_display_cfg.plane.HostVMMaxPageTableLevels,
6546 						mode_lib->ms.cache_display_cfg.plane.GPUVMEnable,
6547 						mode_lib->ms.soc.hostvm_min_page_size_kbytes * 1024,
6548 						mode_lib->ms.PDEAndMetaPTEBytesPerFrame[j][k],
6549 						mode_lib->ms.MetaRowBytes[j][k],
6550 						mode_lib->ms.DPTEBytesPerRow[j][k],
6551 						mode_lib->ms.BandwidthAvailableForImmediateFlip,
6552 						mode_lib->ms.TotImmediateFlipBytes,
6553 						mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k],
6554 						(mode_lib->ms.cache_display_cfg.timing.HTotal[k] / mode_lib->ms.cache_display_cfg.timing.PixelClock[k]),
6555 						mode_lib->ms.cache_display_cfg.plane.VRatio[k],
6556 						mode_lib->ms.cache_display_cfg.plane.VRatioChroma[k],
6557 						mode_lib->ms.Tno_bw[k],
6558 						mode_lib->ms.cache_display_cfg.surface.DCCEnable[k],
6559 						mode_lib->ms.dpte_row_height[k],
6560 						mode_lib->ms.meta_row_height[k],
6561 						mode_lib->ms.dpte_row_height_chroma[k],
6562 						mode_lib->ms.meta_row_height_chroma[k],
6563 						mode_lib->ms.use_one_row_for_frame_flip[j][k], // 24
6564 
6565 						/* Output */
6566 						&mode_lib->ms.DestinationLinesToRequestVMInImmediateFlip[k],
6567 						&mode_lib->ms.DestinationLinesToRequestRowInImmediateFlip[k],
6568 						&mode_lib->ms.final_flip_bw[k],
6569 						&mode_lib->ms.ImmediateFlipSupportedForPipe[k]);
6570 				}
6571 
6572 				{
6573 				CalculateImmediateFlipBandwithSupport(mode_lib->ms.num_active_planes,
6574 													mode_lib->ms.ReturnBWPerState[j],
6575 													mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange,
6576 													mode_lib->ms.policy.ImmediateFlipRequirement,
6577 													mode_lib->ms.final_flip_bw,
6578 													mode_lib->ms.ReadBandwidthLuma,
6579 													mode_lib->ms.ReadBandwidthChroma,
6580 													mode_lib->ms.RequiredPrefetchPixelDataBWLuma,
6581 													mode_lib->ms.RequiredPrefetchPixelDataBWChroma,
6582 													mode_lib->ms.cursor_bw,
6583 													mode_lib->ms.meta_row_bandwidth_this_state,
6584 													mode_lib->ms.dpte_row_bandwidth_this_state,
6585 													mode_lib->ms.cursor_bw_pre,
6586 													mode_lib->ms.prefetch_vmrow_bw,
6587 													mode_lib->ms.NoOfDPP[j], // VBA_ERROR DPPPerSurface is not assigned at this point, should use NoOfDpp here
6588 													mode_lib->ms.UrgentBurstFactorLuma,
6589 													mode_lib->ms.UrgentBurstFactorChroma,
6590 													mode_lib->ms.UrgentBurstFactorCursor,
6591 													mode_lib->ms.UrgentBurstFactorLumaPre,
6592 													mode_lib->ms.UrgentBurstFactorChromaPre,
6593 													mode_lib->ms.UrgentBurstFactorCursorPre,
6594 
6595 													/* output */
6596 													&s->dummy_single[0], // dml_float_t *TotalBandwidth
6597 													&s->dummy_single[1], // dml_float_t *TotalBandwidthNotIncludingMALLPrefetch
6598 													&s->dummy_single[2], // dml_float_t *FractionOfUrgentBandwidth
6599 													&mode_lib->ms.support.ImmediateFlipSupportedForState[j]); // dml_bool_t *ImmediateFlipBandwidthSupport
6600 				}
6601 
6602 				for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
6603 					if (!(mode_lib->ms.policy.ImmediateFlipRequirement[k] == dml_immediate_flip_not_required) && (mode_lib->ms.ImmediateFlipSupportedForPipe[k] == false))
6604 						mode_lib->ms.support.ImmediateFlipSupportedForState[j] = false;
6605 				}
6606 
6607 			} else { // if prefetch not support, assume iflip not supported
6608 				mode_lib->ms.support.ImmediateFlipSupportedForState[j] = false;
6609 			}
6610 
6611 			if (s->MaxVStartup <= __DML_VBA_MIN_VSTARTUP__ || s->AnyLinesForVMOrRowTooLarge == false) {
6612 				s->NextMaxVStartup = s->MaxVStartupAllPlanes[j];
6613 				for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
6614 					s->NextPrefetchMode[k] = s->NextPrefetchMode[k] + 1;
6615 
6616 					if (s->NextPrefetchMode[k] <= s->MaxPrefetchMode[k])
6617 						s->AllPrefetchModeTested = false;
6618 				}
6619 			} else {
6620 				s->NextMaxVStartup = s->NextMaxVStartup - 1;
6621 			}
6622 		} while (!((mode_lib->ms.support.PrefetchSupported[j] == true && mode_lib->ms.support.DynamicMetadataSupported[j] == true &&
6623 					mode_lib->ms.support.VRatioInPrefetchSupported[j] == true &&
6624 					// consider flip support is okay if when there is no hostvm and the user does't require a iflip OR the flip bw is ok
6625 					// If there is hostvm, DCN needs to support iflip for invalidation
6626 					((s->ImmediateFlipRequiredFinal) || mode_lib->ms.support.ImmediateFlipSupportedForState[j] == true)) ||
6627 					(s->NextMaxVStartup == s->MaxVStartupAllPlanes[j] && s->AllPrefetchModeTested)));
6628 
6629 		for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
6630 			mode_lib->ms.use_one_row_for_frame_this_state[k] = mode_lib->ms.use_one_row_for_frame[j][k];
6631 		}
6632 
6633 		s->mSOCParameters.UrgentLatency = mode_lib->ms.UrgLatency;
6634 		s->mSOCParameters.ExtraLatency = mode_lib->ms.ExtraLatency;
6635 		s->mSOCParameters.WritebackLatency = mode_lib->ms.state.writeback_latency_us;
6636 		s->mSOCParameters.DRAMClockChangeLatency = mode_lib->ms.state.dram_clock_change_latency_us;
6637 		s->mSOCParameters.FCLKChangeLatency = mode_lib->ms.state.fclk_change_latency_us;
6638 		s->mSOCParameters.SRExitTime = mode_lib->ms.state.sr_exit_time_us;
6639 		s->mSOCParameters.SREnterPlusExitTime = mode_lib->ms.state.sr_enter_plus_exit_time_us;
6640 		s->mSOCParameters.SRExitZ8Time = mode_lib->ms.state.sr_exit_z8_time_us;
6641 		s->mSOCParameters.SREnterPlusExitZ8Time = mode_lib->ms.state.sr_enter_plus_exit_z8_time_us;
6642 		s->mSOCParameters.USRRetrainingLatency = mode_lib->ms.state.usr_retraining_latency_us;
6643 		s->mSOCParameters.SMNLatency = mode_lib->ms.soc.smn_latency_us;
6644 
6645 		CalculateWatermarks_params->USRRetrainingRequiredFinal = mode_lib->ms.policy.USRRetrainingRequiredFinal;
6646 		CalculateWatermarks_params->UseMALLForPStateChange = mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange;
6647 		CalculateWatermarks_params->PrefetchMode = mode_lib->ms.PrefetchMode;
6648 		CalculateWatermarks_params->NumberOfActiveSurfaces = mode_lib->ms.num_active_planes;
6649 		CalculateWatermarks_params->MaxLineBufferLines = mode_lib->ms.ip.max_line_buffer_lines;
6650 		CalculateWatermarks_params->LineBufferSize = mode_lib->ms.ip.line_buffer_size_bits;
6651 		CalculateWatermarks_params->WritebackInterfaceBufferSize = mode_lib->ms.ip.writeback_interface_buffer_size_kbytes;
6652 		CalculateWatermarks_params->DCFCLK = mode_lib->ms.DCFCLKState[j];
6653 		CalculateWatermarks_params->ReturnBW = mode_lib->ms.ReturnBWPerState[j];
6654 		CalculateWatermarks_params->SynchronizeTimingsFinal = mode_lib->ms.policy.SynchronizeTimingsFinal;
6655 		CalculateWatermarks_params->SynchronizeDRRDisplaysForUCLKPStateChangeFinal = mode_lib->ms.policy.SynchronizeDRRDisplaysForUCLKPStateChangeFinal;
6656 		CalculateWatermarks_params->DRRDisplay = mode_lib->ms.cache_display_cfg.timing.DRRDisplay;
6657 		CalculateWatermarks_params->dpte_group_bytes = mode_lib->ms.dpte_group_bytes;
6658 		CalculateWatermarks_params->meta_row_height = mode_lib->ms.meta_row_height;
6659 		CalculateWatermarks_params->meta_row_height_chroma = mode_lib->ms.meta_row_height_chroma;
6660 		CalculateWatermarks_params->mmSOCParameters = s->mSOCParameters;
6661 		CalculateWatermarks_params->WritebackChunkSize = mode_lib->ms.ip.writeback_chunk_size_kbytes;
6662 		CalculateWatermarks_params->SOCCLK = mode_lib->ms.state.socclk_mhz;
6663 		CalculateWatermarks_params->DCFClkDeepSleep = mode_lib->ms.ProjectedDCFCLKDeepSleep[j];
6664 		CalculateWatermarks_params->DETBufferSizeY = mode_lib->ms.DETBufferSizeYThisState;
6665 		CalculateWatermarks_params->DETBufferSizeC = mode_lib->ms.DETBufferSizeCThisState;
6666 		CalculateWatermarks_params->SwathHeightY = mode_lib->ms.SwathHeightYThisState;
6667 		CalculateWatermarks_params->SwathHeightC = mode_lib->ms.SwathHeightCThisState;
6668 		CalculateWatermarks_params->LBBitPerPixel = mode_lib->ms.cache_display_cfg.plane.LBBitPerPixel;
6669 		CalculateWatermarks_params->SwathWidthY = mode_lib->ms.SwathWidthYThisState;
6670 		CalculateWatermarks_params->SwathWidthC = mode_lib->ms.SwathWidthCThisState;
6671 		CalculateWatermarks_params->HRatio = mode_lib->ms.cache_display_cfg.plane.HRatio;
6672 		CalculateWatermarks_params->HRatioChroma = mode_lib->ms.cache_display_cfg.plane.HRatioChroma;
6673 		CalculateWatermarks_params->VTaps = mode_lib->ms.cache_display_cfg.plane.VTaps;
6674 		CalculateWatermarks_params->VTapsChroma = mode_lib->ms.cache_display_cfg.plane.VTapsChroma;
6675 		CalculateWatermarks_params->VRatio = mode_lib->ms.cache_display_cfg.plane.VRatio;
6676 		CalculateWatermarks_params->VRatioChroma = mode_lib->ms.cache_display_cfg.plane.VRatioChroma;
6677 		CalculateWatermarks_params->HTotal = mode_lib->ms.cache_display_cfg.timing.HTotal;
6678 		CalculateWatermarks_params->VTotal = mode_lib->ms.cache_display_cfg.timing.VTotal;
6679 		CalculateWatermarks_params->VActive = mode_lib->ms.cache_display_cfg.timing.VActive;
6680 		CalculateWatermarks_params->PixelClock = mode_lib->ms.cache_display_cfg.timing.PixelClock;
6681 		CalculateWatermarks_params->BlendingAndTiming = mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming;
6682 		CalculateWatermarks_params->DPPPerSurface = mode_lib->ms.NoOfDPPThisState;
6683 		CalculateWatermarks_params->BytePerPixelDETY = mode_lib->ms.BytePerPixelInDETY;
6684 		CalculateWatermarks_params->BytePerPixelDETC = mode_lib->ms.BytePerPixelInDETC;
6685 		CalculateWatermarks_params->DSTXAfterScaler = s->DSTXAfterScaler;
6686 		CalculateWatermarks_params->DSTYAfterScaler = s->DSTYAfterScaler;
6687 		CalculateWatermarks_params->WritebackEnable = mode_lib->ms.cache_display_cfg.writeback.WritebackEnable;
6688 		CalculateWatermarks_params->WritebackPixelFormat = mode_lib->ms.cache_display_cfg.writeback.WritebackPixelFormat;
6689 		CalculateWatermarks_params->WritebackDestinationWidth = mode_lib->ms.cache_display_cfg.writeback.WritebackDestinationWidth;
6690 		CalculateWatermarks_params->WritebackDestinationHeight = mode_lib->ms.cache_display_cfg.writeback.WritebackDestinationHeight;
6691 		CalculateWatermarks_params->WritebackSourceHeight = mode_lib->ms.cache_display_cfg.writeback.WritebackSourceHeight;
6692 		CalculateWatermarks_params->UnboundedRequestEnabled = mode_lib->ms.UnboundedRequestEnabledThisState;
6693 		CalculateWatermarks_params->CompressedBufferSizeInkByte = mode_lib->ms.CompressedBufferSizeInkByteThisState;
6694 
6695 		// Output
6696 		CalculateWatermarks_params->Watermark = &s->dummy_watermark; // Watermarks *Watermark
6697 		CalculateWatermarks_params->DRAMClockChangeSupport = &mode_lib->ms.support.DRAMClockChangeSupport[j];
6698 		CalculateWatermarks_params->MaxActiveDRAMClockChangeLatencySupported = &s->dummy_single_array[0]; // dml_float_t *MaxActiveDRAMClockChangeLatencySupported[]
6699 		CalculateWatermarks_params->SubViewportLinesNeededInMALL = &mode_lib->ms.SubViewportLinesNeededInMALL[j]; // dml_uint_t SubViewportLinesNeededInMALL[]
6700 		CalculateWatermarks_params->FCLKChangeSupport = &mode_lib->ms.support.FCLKChangeSupport[j];
6701 		CalculateWatermarks_params->MaxActiveFCLKChangeLatencySupported = &s->dummy_single[0]; // dml_float_t *MaxActiveFCLKChangeLatencySupported
6702 		CalculateWatermarks_params->USRRetrainingSupport = &mode_lib->ms.support.USRRetrainingSupport[j];
6703 		CalculateWatermarks_params->ActiveDRAMClockChangeLatencyMargin = mode_lib->ms.support.ActiveDRAMClockChangeLatencyMargin;
6704 
6705 		CalculateWatermarksMALLUseAndDRAMSpeedChangeSupport(&mode_lib->scratch,
6706 			CalculateWatermarks_params);
6707 
6708 	} // for j
6709 }
6710 
6711 /// @brief The Mode Support function.
dml_core_mode_support(struct display_mode_lib_st * mode_lib)6712 dml_bool_t dml_core_mode_support(struct display_mode_lib_st *mode_lib)
6713 {
6714 	struct dml_core_mode_support_locals_st *s = &mode_lib->scratch.dml_core_mode_support_locals;
6715 	struct UseMinimumDCFCLK_params_st *UseMinimumDCFCLK_params = &mode_lib->scratch.UseMinimumDCFCLK_params;
6716 	struct CalculateSwathAndDETConfiguration_params_st *CalculateSwathAndDETConfiguration_params = &mode_lib->scratch.CalculateSwathAndDETConfiguration_params;
6717 	struct CalculateVMRowAndSwath_params_st *CalculateVMRowAndSwath_params = &mode_lib->scratch.CalculateVMRowAndSwath_params;
6718 
6719 	dml_uint_t j, k, m;
6720 
6721 	mode_lib->ms.num_active_planes = dml_get_num_active_planes(&mode_lib->ms.cache_display_cfg);
6722 	dml_print("DML::%s: num_active_planes = %u\n", __func__, mode_lib->ms.num_active_planes);
6723 
6724 	CalculateMaxDETAndMinCompressedBufferSize(
6725 								mode_lib->ms.ip.config_return_buffer_size_in_kbytes,
6726 								mode_lib->ms.ip.config_return_buffer_segment_size_in_kbytes,
6727 								mode_lib->ms.ip.rob_buffer_size_kbytes,
6728 								mode_lib->ms.ip.max_num_dpp,
6729 								mode_lib->ms.policy.NomDETInKByteOverrideEnable,   // VBA_DELTA
6730 								mode_lib->ms.policy.NomDETInKByteOverrideValue,    // VBA_DELTA
6731 
6732 								/* Output */
6733 								&mode_lib->ms.MaxTotalDETInKByte,
6734 								&mode_lib->ms.NomDETInKByte,
6735 								&mode_lib->ms.MinCompressedBufferSizeInKByte);
6736 
6737 	PixelClockAdjustmentForProgressiveToInterlaceUnit(&mode_lib->ms.cache_display_cfg, mode_lib->ms.ip.ptoi_supported);
6738 
6739 
6740 	/*MODE SUPPORT, VOLTAGE STATE AND SOC CONFIGURATION*/
6741 
6742 	/*Scale Ratio, taps Support Check*/
6743 	mode_lib->ms.support.ScaleRatioAndTapsSupport = true;
6744 	for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
6745 		if (mode_lib->ms.cache_display_cfg.plane.ScalerEnabled[k] == false
6746 				&& ((mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_444_64
6747 						&& mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_444_32
6748 						&& mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_444_16
6749 						&& mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_mono_16
6750 						&& mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_mono_8
6751 						&& mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_rgbe
6752 						&& mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_rgbe_alpha)
6753 						|| mode_lib->ms.cache_display_cfg.plane.HRatio[k] != 1.0
6754 						|| mode_lib->ms.cache_display_cfg.plane.HTaps[k] != 1.0
6755 						|| mode_lib->ms.cache_display_cfg.plane.VRatio[k] != 1.0
6756 						|| mode_lib->ms.cache_display_cfg.plane.VTaps[k] != 1.0)) {
6757 			mode_lib->ms.support.ScaleRatioAndTapsSupport = false;
6758 		} else if (mode_lib->ms.cache_display_cfg.plane.VTaps[k] < 1.0 || mode_lib->ms.cache_display_cfg.plane.VTaps[k] > 8.0
6759 				|| mode_lib->ms.cache_display_cfg.plane.HTaps[k] < 1.0 || mode_lib->ms.cache_display_cfg.plane.HTaps[k] > 8.0
6760 				|| (mode_lib->ms.cache_display_cfg.plane.HTaps[k] > 1.0 && (mode_lib->ms.cache_display_cfg.plane.HTaps[k] % 2) == 1)
6761 				|| mode_lib->ms.cache_display_cfg.plane.HRatio[k] > mode_lib->ms.ip.max_hscl_ratio
6762 				|| mode_lib->ms.cache_display_cfg.plane.VRatio[k] > mode_lib->ms.ip.max_vscl_ratio
6763 				|| mode_lib->ms.cache_display_cfg.plane.HRatio[k] > mode_lib->ms.cache_display_cfg.plane.HTaps[k]
6764 				|| mode_lib->ms.cache_display_cfg.plane.VRatio[k] > mode_lib->ms.cache_display_cfg.plane.VTaps[k]
6765 				|| (mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_444_64
6766 					&& mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_444_32
6767 					&& mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_444_16
6768 					&& mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_mono_16
6769 					&& mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_mono_8
6770 					&& mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_rgbe
6771 					&& (mode_lib->ms.cache_display_cfg.plane.VTapsChroma[k] < 1 || mode_lib->ms.cache_display_cfg.plane.VTapsChroma[k] > 8 || mode_lib->ms.cache_display_cfg.plane.HTapsChroma[k] < 1 || mode_lib->ms.cache_display_cfg.plane.HTapsChroma[k] > 8 ||
6772 						(mode_lib->ms.cache_display_cfg.plane.HTapsChroma[k] > 1 && mode_lib->ms.cache_display_cfg.plane.HTapsChroma[k] % 2 == 1) ||
6773 					mode_lib->ms.cache_display_cfg.plane.HRatioChroma[k] > mode_lib->ms.ip.max_hscl_ratio ||
6774 					mode_lib->ms.cache_display_cfg.plane.VRatioChroma[k] > mode_lib->ms.ip.max_vscl_ratio ||
6775 					mode_lib->ms.cache_display_cfg.plane.HRatioChroma[k] > mode_lib->ms.cache_display_cfg.plane.HTapsChroma[k] ||
6776 					mode_lib->ms.cache_display_cfg.plane.VRatioChroma[k] > mode_lib->ms.cache_display_cfg.plane.VTapsChroma[k]))) {
6777 			mode_lib->ms.support.ScaleRatioAndTapsSupport = false;
6778 		}
6779 	}
6780 
6781 	/*Source Format, Pixel Format and Scan Support Check*/
6782 	mode_lib->ms.support.SourceFormatPixelAndScanSupport = true;
6783 	for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
6784 			if (mode_lib->ms.cache_display_cfg.surface.SurfaceTiling[k] == dml_sw_linear && (!(!dml_is_vertical_rotation(mode_lib->ms.cache_display_cfg.plane.SourceScan[k])) || mode_lib->ms.cache_display_cfg.surface.DCCEnable[k] == true)) {
6785 			mode_lib->ms.support.SourceFormatPixelAndScanSupport = false;
6786 		}
6787 	}
6788 
6789 	for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
6790 		CalculateBytePerPixelAndBlockSizes(
6791 								mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k],
6792 								mode_lib->ms.cache_display_cfg.surface.SurfaceTiling[k],
6793 
6794 								/* Output */
6795 								&mode_lib->ms.BytePerPixelY[k],
6796 								&mode_lib->ms.BytePerPixelC[k],
6797 								&mode_lib->ms.BytePerPixelInDETY[k],
6798 								&mode_lib->ms.BytePerPixelInDETC[k],
6799 								&mode_lib->ms.Read256BlockHeightY[k],
6800 								&mode_lib->ms.Read256BlockHeightC[k],
6801 								&mode_lib->ms.Read256BlockWidthY[k],
6802 								&mode_lib->ms.Read256BlockWidthC[k],
6803 								&mode_lib->ms.MacroTileHeightY[k],
6804 								&mode_lib->ms.MacroTileHeightC[k],
6805 								&mode_lib->ms.MacroTileWidthY[k],
6806 								&mode_lib->ms.MacroTileWidthC[k]);
6807 	}
6808 
6809 	/*Bandwidth Support Check*/
6810 	for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
6811 		if (!dml_is_vertical_rotation(mode_lib->ms.cache_display_cfg.plane.SourceScan[k])) {
6812 			mode_lib->ms.SwathWidthYSingleDPP[k] = mode_lib->ms.cache_display_cfg.plane.ViewportWidth[k];
6813 			mode_lib->ms.SwathWidthCSingleDPP[k] = mode_lib->ms.cache_display_cfg.plane.ViewportWidthChroma[k];
6814 		} else {
6815 			mode_lib->ms.SwathWidthYSingleDPP[k] = mode_lib->ms.cache_display_cfg.plane.ViewportHeight[k];
6816 			mode_lib->ms.SwathWidthCSingleDPP[k] = mode_lib->ms.cache_display_cfg.plane.ViewportHeightChroma[k];
6817 		}
6818 	}
6819 	for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
6820 		mode_lib->ms.ReadBandwidthLuma[k] = mode_lib->ms.SwathWidthYSingleDPP[k] * dml_ceil(mode_lib->ms.BytePerPixelInDETY[k], 1.0) / (mode_lib->ms.cache_display_cfg.timing.HTotal[k] / mode_lib->ms.cache_display_cfg.timing.PixelClock[k]) * mode_lib->ms.cache_display_cfg.plane.VRatio[k];
6821 		mode_lib->ms.ReadBandwidthChroma[k] = mode_lib->ms.SwathWidthYSingleDPP[k] / 2 * dml_ceil(mode_lib->ms.BytePerPixelInDETC[k], 2.0) / (mode_lib->ms.cache_display_cfg.timing.HTotal[k] / mode_lib->ms.cache_display_cfg.timing.PixelClock[k]) * mode_lib->ms.cache_display_cfg.plane.VRatio[k] / 2.0;
6822 	}
6823 	for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
6824 		if (mode_lib->ms.cache_display_cfg.writeback.WritebackEnable[k] == true
6825 				&& mode_lib->ms.cache_display_cfg.writeback.WritebackPixelFormat[k] == dml_444_64) {
6826 			mode_lib->ms.WriteBandwidth[k] = mode_lib->ms.cache_display_cfg.writeback.WritebackDestinationWidth[k]
6827 					* mode_lib->ms.cache_display_cfg.writeback.WritebackDestinationHeight[k]
6828 					/ (mode_lib->ms.cache_display_cfg.writeback.WritebackSourceHeight[k]
6829 							* mode_lib->ms.cache_display_cfg.timing.HTotal[k]
6830 							/ mode_lib->ms.cache_display_cfg.timing.PixelClock[k]) * 8.0;
6831 		} else if (mode_lib->ms.cache_display_cfg.writeback.WritebackEnable[k] == true) {
6832 			mode_lib->ms.WriteBandwidth[k] = mode_lib->ms.cache_display_cfg.writeback.WritebackDestinationWidth[k]
6833 					* mode_lib->ms.cache_display_cfg.writeback.WritebackDestinationHeight[k]
6834 					/ (mode_lib->ms.cache_display_cfg.writeback.WritebackSourceHeight[k]
6835 							* mode_lib->ms.cache_display_cfg.timing.HTotal[k]
6836 							/ mode_lib->ms.cache_display_cfg.timing.PixelClock[k]) * 4.0;
6837 		} else {
6838 			mode_lib->ms.WriteBandwidth[k] = 0.0;
6839 		}
6840 	}
6841 
6842 	/*Writeback Latency support check*/
6843 	mode_lib->ms.support.WritebackLatencySupport = true;
6844 	for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
6845 		if (mode_lib->ms.cache_display_cfg.writeback.WritebackEnable[k] == true &&
6846 			(mode_lib->ms.WriteBandwidth[k] > mode_lib->ms.ip.writeback_interface_buffer_size_kbytes * 1024 / mode_lib->ms.state.writeback_latency_us)) {
6847 			mode_lib->ms.support.WritebackLatencySupport = false;
6848 		}
6849 	}
6850 
6851 	/*Writeback Mode Support Check*/
6852 	s->TotalNumberOfActiveWriteback = 0;
6853 	for (k = 0; k <= (dml_uint_t) mode_lib->ms.num_active_planes - 1; k++) {
6854 		if (mode_lib->ms.cache_display_cfg.writeback.WritebackEnable[k] == true) {
6855 			s->TotalNumberOfActiveWriteback = s->TotalNumberOfActiveWriteback + 1;
6856 		}
6857    }
6858 
6859 	mode_lib->ms.support.EnoughWritebackUnits = 1;
6860 	if (s->TotalNumberOfActiveWriteback > (dml_uint_t) mode_lib->ms.ip.max_num_wb) {
6861 		mode_lib->ms.support.EnoughWritebackUnits = false;
6862 	}
6863 
6864 	/*Writeback Scale Ratio and Taps Support Check*/
6865 	mode_lib->ms.support.WritebackScaleRatioAndTapsSupport = true;
6866 	for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
6867 		if (mode_lib->ms.cache_display_cfg.writeback.WritebackEnable[k] == true) {
6868 			if (mode_lib->ms.cache_display_cfg.writeback.WritebackHRatio[k] > mode_lib->ms.ip.writeback_max_hscl_ratio
6869 				|| mode_lib->ms.cache_display_cfg.writeback.WritebackVRatio[k] > mode_lib->ms.ip.writeback_max_vscl_ratio
6870 				|| mode_lib->ms.cache_display_cfg.writeback.WritebackHRatio[k] < mode_lib->ms.ip.writeback_min_hscl_ratio
6871 				|| mode_lib->ms.cache_display_cfg.writeback.WritebackVRatio[k] < mode_lib->ms.ip.writeback_min_vscl_ratio
6872 				|| mode_lib->ms.cache_display_cfg.writeback.WritebackHTaps[k] > (dml_uint_t) mode_lib->ms.ip.writeback_max_hscl_taps
6873 				|| mode_lib->ms.cache_display_cfg.writeback.WritebackVTaps[k] > (dml_uint_t) mode_lib->ms.ip.writeback_max_vscl_taps
6874 				|| mode_lib->ms.cache_display_cfg.writeback.WritebackHRatio[k] > (dml_uint_t) mode_lib->ms.cache_display_cfg.writeback.WritebackHTaps[k]
6875 				|| mode_lib->ms.cache_display_cfg.writeback.WritebackVRatio[k] > (dml_uint_t) mode_lib->ms.cache_display_cfg.writeback.WritebackVTaps[k]
6876 				|| (mode_lib->ms.cache_display_cfg.writeback.WritebackHTaps[k] > 2.0 && ((mode_lib->ms.cache_display_cfg.writeback.WritebackHTaps[k] % 2) == 1))) {
6877 				mode_lib->ms.support.WritebackScaleRatioAndTapsSupport = false;
6878 			}
6879 			if (2.0 * mode_lib->ms.cache_display_cfg.writeback.WritebackDestinationWidth[k] * (mode_lib->ms.cache_display_cfg.writeback.WritebackVTaps[k] - 1) * 57 > mode_lib->ms.ip.writeback_line_buffer_buffer_size) {
6880 				mode_lib->ms.support.WritebackScaleRatioAndTapsSupport = false;
6881 			}
6882 		}
6883 	}
6884 
6885 	for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
6886 		CalculateSinglePipeDPPCLKAndSCLThroughput(
6887 				mode_lib->ms.cache_display_cfg.plane.HRatio[k],
6888 				mode_lib->ms.cache_display_cfg.plane.HRatioChroma[k],
6889 				mode_lib->ms.cache_display_cfg.plane.VRatio[k],
6890 				mode_lib->ms.cache_display_cfg.plane.VRatioChroma[k],
6891 				mode_lib->ms.ip.max_dchub_pscl_bw_pix_per_clk,
6892 				mode_lib->ms.ip.max_pscl_lb_bw_pix_per_clk,
6893 				mode_lib->ms.cache_display_cfg.timing.PixelClock[k],
6894 				mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k],
6895 				mode_lib->ms.cache_display_cfg.plane.HTaps[k],
6896 				mode_lib->ms.cache_display_cfg.plane.HTapsChroma[k],
6897 				mode_lib->ms.cache_display_cfg.plane.VTaps[k],
6898 				mode_lib->ms.cache_display_cfg.plane.VTapsChroma[k],
6899 				/* Output */
6900 				&mode_lib->ms.PSCL_FACTOR[k],
6901 				&mode_lib->ms.PSCL_FACTOR_CHROMA[k],
6902 				&mode_lib->ms.MinDPPCLKUsingSingleDPP[k]);
6903 	}
6904 
6905 	for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
6906 		if (mode_lib->ms.cache_display_cfg.surface.SurfaceTiling[k] == dml_sw_linear) {
6907 			s->MaximumSwathWidthSupportLuma = 8192;
6908 		} else if (!dml_is_vertical_rotation(mode_lib->ms.cache_display_cfg.plane.SourceScan[k]) && mode_lib->ms.BytePerPixelC[k] > 0 && mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_rgbe_alpha) {
6909 			s->MaximumSwathWidthSupportLuma = 7680;
6910 		} else if (dml_is_vertical_rotation(mode_lib->ms.cache_display_cfg.plane.SourceScan[k]) && mode_lib->ms.BytePerPixelC[k] > 0 && mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_rgbe_alpha) {
6911 			s->MaximumSwathWidthSupportLuma = 4320;
6912 		} else if (mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] == dml_rgbe_alpha) {
6913 			s->MaximumSwathWidthSupportLuma = 3840;
6914 		} else if (dml_is_vertical_rotation(mode_lib->ms.cache_display_cfg.plane.SourceScan[k]) && mode_lib->ms.BytePerPixelY[k] == 8 && mode_lib->ms.cache_display_cfg.surface.DCCEnable[k] == true) {
6915 			s->MaximumSwathWidthSupportLuma = 3072;
6916 		} else {
6917 			s->MaximumSwathWidthSupportLuma = 6144;
6918 		}
6919 
6920        if (mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] == dml_420_8 || mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] == dml_420_10 || mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] == dml_420_12) {
6921 			s->MaximumSwathWidthSupportChroma = (dml_uint_t)(s->MaximumSwathWidthSupportLuma / 2.0);
6922 		} else {
6923 			s->MaximumSwathWidthSupportChroma = s->MaximumSwathWidthSupportLuma;
6924        }
6925 		mode_lib->ms.MaximumSwathWidthInLineBufferLuma = mode_lib->ms.ip.line_buffer_size_bits * dml_max(mode_lib->ms.cache_display_cfg.plane.HRatio[k], 1.0) / mode_lib->ms.cache_display_cfg.plane.LBBitPerPixel[k] /
6926 															(mode_lib->ms.cache_display_cfg.plane.VTaps[k] + dml_max(dml_ceil(mode_lib->ms.cache_display_cfg.plane.VRatio[k], 1.0) - 2, 0.0));
6927 		if (mode_lib->ms.BytePerPixelC[k] == 0.0) {
6928 			mode_lib->ms.MaximumSwathWidthInLineBufferChroma = 0;
6929 		} else {
6930 			mode_lib->ms.MaximumSwathWidthInLineBufferChroma =
6931 							mode_lib->ms.ip.line_buffer_size_bits
6932 									* dml_max(mode_lib->ms.cache_display_cfg.plane.HRatioChroma[k], 1.0)
6933 									/ mode_lib->ms.cache_display_cfg.plane.LBBitPerPixel[k]
6934 									/ (mode_lib->ms.cache_display_cfg.plane.VTapsChroma[k]
6935 									+ dml_max(dml_ceil(mode_lib->ms.cache_display_cfg.plane.VRatioChroma[k], 1.0) - 2, 0.0));
6936 		}
6937 		mode_lib->ms.MaximumSwathWidthLuma[k] = dml_min(s->MaximumSwathWidthSupportLuma, mode_lib->ms.MaximumSwathWidthInLineBufferLuma);
6938 		mode_lib->ms.MaximumSwathWidthChroma[k] = dml_min(s->MaximumSwathWidthSupportChroma, mode_lib->ms.MaximumSwathWidthInLineBufferChroma);
6939 	}
6940 
6941 	/*Number Of DSC Slices*/
6942 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
6943 		if (mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming[k] == k &&
6944 			mode_lib->ms.cache_display_cfg.output.DSCEnable[k] != dml_dsc_disable) {
6945 			mode_lib->ms.support.NumberOfDSCSlices[k] = mode_lib->ms.cache_display_cfg.output.DSCSlices[k];
6946 
6947 			if (mode_lib->ms.support.NumberOfDSCSlices[k] == 0) {
6948 				if (mode_lib->ms.cache_display_cfg.output.PixelClockBackEnd[k] > 4800) {
6949 					mode_lib->ms.support.NumberOfDSCSlices[k] = (dml_uint_t)(dml_ceil(mode_lib->ms.cache_display_cfg.output.PixelClockBackEnd[k] / 600, 4));
6950 				} else if (mode_lib->ms.cache_display_cfg.output.PixelClockBackEnd[k] > 2400) {
6951 					mode_lib->ms.support.NumberOfDSCSlices[k] = 8;
6952 				} else if (mode_lib->ms.cache_display_cfg.output.PixelClockBackEnd[k] > 1200) {
6953 					mode_lib->ms.support.NumberOfDSCSlices[k] = 4;
6954 				} else if (mode_lib->ms.cache_display_cfg.output.PixelClockBackEnd[k] > 340) {
6955 					mode_lib->ms.support.NumberOfDSCSlices[k] = 2;
6956 				} else {
6957 					mode_lib->ms.support.NumberOfDSCSlices[k] = 1;
6958 				}
6959 			}
6960 		} else {
6961 			mode_lib->ms.support.NumberOfDSCSlices[k] = 1;
6962 		}
6963 	}
6964 
6965 	CalculateSwathAndDETConfiguration_params->DETSizeOverride = mode_lib->ms.cache_display_cfg.plane.DETSizeOverride;
6966 	CalculateSwathAndDETConfiguration_params->UseMALLForPStateChange = mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange;
6967 	CalculateSwathAndDETConfiguration_params->ConfigReturnBufferSizeInKByte = mode_lib->ms.ip.config_return_buffer_size_in_kbytes;
6968 	CalculateSwathAndDETConfiguration_params->ROBBufferSizeInKByte = mode_lib->ms.ip.rob_buffer_size_kbytes;
6969 	CalculateSwathAndDETConfiguration_params->MaxTotalDETInKByte = mode_lib->ms.MaxTotalDETInKByte;
6970 	CalculateSwathAndDETConfiguration_params->MinCompressedBufferSizeInKByte = mode_lib->ms.MinCompressedBufferSizeInKByte;
6971 	CalculateSwathAndDETConfiguration_params->PixelChunkSizeInKByte = mode_lib->ms.ip.pixel_chunk_size_kbytes;
6972 	CalculateSwathAndDETConfiguration_params->ForceSingleDPP = 1;
6973 	CalculateSwathAndDETConfiguration_params->NumberOfActiveSurfaces = mode_lib->ms.num_active_planes;
6974 	CalculateSwathAndDETConfiguration_params->nomDETInKByte = mode_lib->ms.NomDETInKByte;
6975 	CalculateSwathAndDETConfiguration_params->UseUnboundedRequestingFinal = mode_lib->ms.policy.UseUnboundedRequesting;
6976 	CalculateSwathAndDETConfiguration_params->ConfigReturnBufferSegmentSizeInkByte = mode_lib->ms.ip.config_return_buffer_segment_size_in_kbytes;
6977 	CalculateSwathAndDETConfiguration_params->CompressedBufferSegmentSizeInkByteFinal = mode_lib->ms.ip.compressed_buffer_segment_size_in_kbytes;
6978 	CalculateSwathAndDETConfiguration_params->Output = mode_lib->ms.cache_display_cfg.output.OutputEncoder;
6979 	CalculateSwathAndDETConfiguration_params->ReadBandwidthLuma = mode_lib->ms.ReadBandwidthLuma;
6980 	CalculateSwathAndDETConfiguration_params->ReadBandwidthChroma = mode_lib->ms.ReadBandwidthChroma;
6981 	CalculateSwathAndDETConfiguration_params->MaximumSwathWidthLuma = mode_lib->ms.MaximumSwathWidthLuma;
6982 	CalculateSwathAndDETConfiguration_params->MaximumSwathWidthChroma = mode_lib->ms.MaximumSwathWidthChroma;
6983 	CalculateSwathAndDETConfiguration_params->SourceScan = mode_lib->ms.cache_display_cfg.plane.SourceScan;
6984 	CalculateSwathAndDETConfiguration_params->ViewportStationary = mode_lib->ms.cache_display_cfg.plane.ViewportStationary;
6985 	CalculateSwathAndDETConfiguration_params->SourcePixelFormat = mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat;
6986 	CalculateSwathAndDETConfiguration_params->SurfaceTiling = mode_lib->ms.cache_display_cfg.surface.SurfaceTiling;
6987 	CalculateSwathAndDETConfiguration_params->ViewportWidth = mode_lib->ms.cache_display_cfg.plane.ViewportWidth;
6988 	CalculateSwathAndDETConfiguration_params->ViewportHeight = mode_lib->ms.cache_display_cfg.plane.ViewportHeight;
6989 	CalculateSwathAndDETConfiguration_params->ViewportXStart = mode_lib->ms.cache_display_cfg.plane.ViewportXStart;
6990 	CalculateSwathAndDETConfiguration_params->ViewportYStart = mode_lib->ms.cache_display_cfg.plane.ViewportYStart;
6991 	CalculateSwathAndDETConfiguration_params->ViewportXStartC = mode_lib->ms.cache_display_cfg.plane.ViewportXStartC;
6992 	CalculateSwathAndDETConfiguration_params->ViewportYStartC = mode_lib->ms.cache_display_cfg.plane.ViewportYStartC;
6993 	CalculateSwathAndDETConfiguration_params->SurfaceWidthY = mode_lib->ms.cache_display_cfg.surface.SurfaceWidthY;
6994 	CalculateSwathAndDETConfiguration_params->SurfaceWidthC = mode_lib->ms.cache_display_cfg.surface.SurfaceWidthC;
6995 	CalculateSwathAndDETConfiguration_params->SurfaceHeightY = mode_lib->ms.cache_display_cfg.surface.SurfaceHeightY;
6996 	CalculateSwathAndDETConfiguration_params->SurfaceHeightC = mode_lib->ms.cache_display_cfg.surface.SurfaceHeightC;
6997 	CalculateSwathAndDETConfiguration_params->Read256BytesBlockHeightY = mode_lib->ms.Read256BlockHeightY;
6998 	CalculateSwathAndDETConfiguration_params->Read256BytesBlockHeightC = mode_lib->ms.Read256BlockHeightC;
6999 	CalculateSwathAndDETConfiguration_params->Read256BytesBlockWidthY = mode_lib->ms.Read256BlockWidthY;
7000 	CalculateSwathAndDETConfiguration_params->Read256BytesBlockWidthC = mode_lib->ms.Read256BlockWidthC;
7001 	CalculateSwathAndDETConfiguration_params->ODMMode = s->dummy_odm_mode;
7002 	CalculateSwathAndDETConfiguration_params->BlendingAndTiming = mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming;
7003 	CalculateSwathAndDETConfiguration_params->BytePerPixY = mode_lib->ms.BytePerPixelY;
7004 	CalculateSwathAndDETConfiguration_params->BytePerPixC = mode_lib->ms.BytePerPixelC;
7005 	CalculateSwathAndDETConfiguration_params->BytePerPixDETY = mode_lib->ms.BytePerPixelInDETY;
7006 	CalculateSwathAndDETConfiguration_params->BytePerPixDETC = mode_lib->ms.BytePerPixelInDETC;
7007 	CalculateSwathAndDETConfiguration_params->HActive = mode_lib->ms.cache_display_cfg.timing.HActive;
7008 	CalculateSwathAndDETConfiguration_params->HRatio = mode_lib->ms.cache_display_cfg.plane.HRatio;
7009 	CalculateSwathAndDETConfiguration_params->HRatioChroma = mode_lib->ms.cache_display_cfg.plane.HRatioChroma;
7010 	CalculateSwathAndDETConfiguration_params->DPPPerSurface = s->dummy_integer_array[0];
7011 	CalculateSwathAndDETConfiguration_params->swath_width_luma_ub = s->dummy_integer_array[1];
7012 	CalculateSwathAndDETConfiguration_params->swath_width_chroma_ub = s->dummy_integer_array[2];
7013 	CalculateSwathAndDETConfiguration_params->SwathWidth = s->dummy_integer_array[3];
7014 	CalculateSwathAndDETConfiguration_params->SwathWidthChroma = s->dummy_integer_array[4];
7015 	CalculateSwathAndDETConfiguration_params->SwathHeightY = s->dummy_integer_array[5];
7016 	CalculateSwathAndDETConfiguration_params->SwathHeightC = s->dummy_integer_array[6];
7017 	CalculateSwathAndDETConfiguration_params->DETBufferSizeInKByte = s->dummy_integer_array[7];
7018 	CalculateSwathAndDETConfiguration_params->DETBufferSizeY = mode_lib->ms.DETBufferSizeY;
7019 	CalculateSwathAndDETConfiguration_params->DETBufferSizeC = mode_lib->ms.DETBufferSizeC;
7020 	CalculateSwathAndDETConfiguration_params->UnboundedRequestEnabled = &s->dummy_boolean[0];
7021 	CalculateSwathAndDETConfiguration_params->compbuf_reserved_space_64b = &s->dummy_integer[2];
7022 	CalculateSwathAndDETConfiguration_params->compbuf_reserved_space_zs = &s->dummy_integer[1];
7023 	CalculateSwathAndDETConfiguration_params->CompressedBufferSizeInkByte = &s->dummy_integer[0];
7024 	CalculateSwathAndDETConfiguration_params->ViewportSizeSupportPerSurface = mode_lib->ms.SingleDPPViewportSizeSupportPerSurface;
7025 	CalculateSwathAndDETConfiguration_params->ViewportSizeSupport = &s->dummy_boolean[1];
7026 
7027 	CalculateSwathAndDETConfiguration(&mode_lib->scratch,
7028 	CalculateSwathAndDETConfiguration_params); /* dml_bool_t *ViewportSizeSupport */
7029 
7030 	s->MPCCombineMethodAsNeededForPStateChangeAndVoltage = false;
7031 	s->MPCCombineMethodAsPossible = false;
7032 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7033 		if (mode_lib->ms.policy.MPCCombineUse[k] == dml_mpc_as_needed_for_pstate_and_voltage)
7034 			s->MPCCombineMethodAsNeededForPStateChangeAndVoltage = true;
7035 		if (mode_lib->ms.policy.MPCCombineUse[k] == dml_mpc_as_possible)
7036 			s->MPCCombineMethodAsPossible = true;
7037 	}
7038 	mode_lib->ms.support.MPCCombineMethodIncompatible = s->MPCCombineMethodAsNeededForPStateChangeAndVoltage && s->MPCCombineMethodAsPossible;
7039 
7040 	for (j = 0; j < 2; j++) {
7041 		mode_lib->ms.TotalNumberOfActiveDPP[j] = 0;
7042 		mode_lib->ms.support.TotalAvailablePipesSupport[j] = true;
7043 
7044 		for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7045 			CalculateODMMode(
7046 					mode_lib->ms.ip.maximum_pixels_per_line_per_dsc_unit,
7047 					mode_lib->ms.cache_display_cfg.timing.HActive[k],
7048 					mode_lib->ms.cache_display_cfg.output.OutputEncoder[k],
7049 					mode_lib->ms.cache_display_cfg.output.OutputFormat[k],
7050 					mode_lib->ms.policy.ODMUse[k],
7051 					mode_lib->ms.state.dispclk_mhz,
7052 					mode_lib->ms.max_state.dispclk_mhz,
7053 					false, // DSCEnable
7054 					mode_lib->ms.TotalNumberOfActiveDPP[j],
7055 					mode_lib->ms.ip.max_num_dpp,
7056 					mode_lib->ms.cache_display_cfg.timing.PixelClock[k],
7057 					mode_lib->ms.soc.dcn_downspread_percent,
7058 					mode_lib->ms.ip.dispclk_ramp_margin_percent,
7059 					mode_lib->ms.soc.dispclk_dppclk_vco_speed_mhz,
7060 					mode_lib->ms.support.NumberOfDSCSlices[k],
7061 
7062 					/* Output */
7063 					&s->TotalAvailablePipesSupportNoDSC,
7064 					&s->NumberOfDPPNoDSC,
7065 					&s->ODMModeNoDSC,
7066 					&s->RequiredDISPCLKPerSurfaceNoDSC);
7067 
7068 			CalculateODMMode(
7069 					mode_lib->ms.ip.maximum_pixels_per_line_per_dsc_unit,
7070 					mode_lib->ms.cache_display_cfg.timing.HActive[k],
7071 					mode_lib->ms.cache_display_cfg.output.OutputEncoder[k],
7072 					mode_lib->ms.cache_display_cfg.output.OutputFormat[k],
7073 					mode_lib->ms.policy.ODMUse[k],
7074 					mode_lib->ms.state.dispclk_mhz,
7075 					mode_lib->ms.max_state.dispclk_mhz,
7076 					true, // DSCEnable
7077 					mode_lib->ms.TotalNumberOfActiveDPP[j],
7078 					mode_lib->ms.ip.max_num_dpp,
7079 					mode_lib->ms.cache_display_cfg.timing.PixelClock[k],
7080 					mode_lib->ms.soc.dcn_downspread_percent,
7081 					mode_lib->ms.ip.dispclk_ramp_margin_percent,
7082 					mode_lib->ms.soc.dispclk_dppclk_vco_speed_mhz,
7083 					mode_lib->ms.support.NumberOfDSCSlices[k],
7084 
7085 					/* Output */
7086 					&s->TotalAvailablePipesSupportDSC,
7087 					&s->NumberOfDPPDSC,
7088 					&s->ODMModeDSC,
7089 					&s->RequiredDISPCLKPerSurfaceDSC);
7090 
7091 			CalculateOutputLink(
7092 					mode_lib->ms.state.phyclk_mhz,
7093 					mode_lib->ms.state.phyclk_d18_mhz,
7094 					mode_lib->ms.state.phyclk_d32_mhz,
7095 					mode_lib->ms.soc.phy_downspread_percent,
7096 					(mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming[k] == k),
7097 					mode_lib->ms.cache_display_cfg.output.OutputEncoder[k],
7098 					mode_lib->ms.cache_display_cfg.output.OutputFormat[k],
7099 					mode_lib->ms.cache_display_cfg.timing.HTotal[k],
7100 					mode_lib->ms.cache_display_cfg.timing.HActive[k],
7101 					mode_lib->ms.cache_display_cfg.output.PixelClockBackEnd[k],
7102 					mode_lib->ms.cache_display_cfg.output.ForcedOutputLinkBPP[k],
7103 					mode_lib->ms.cache_display_cfg.output.DSCInputBitPerComponent[k],
7104 					mode_lib->ms.support.NumberOfDSCSlices[k],
7105 					mode_lib->ms.cache_display_cfg.output.AudioSampleRate[k],
7106 					mode_lib->ms.cache_display_cfg.output.AudioSampleLayout[k],
7107 					s->ODMModeNoDSC,
7108 					s->ODMModeDSC,
7109 					mode_lib->ms.cache_display_cfg.output.DSCEnable[k],
7110 					mode_lib->ms.cache_display_cfg.output.OutputLinkDPLanes[k],
7111 					mode_lib->ms.cache_display_cfg.output.OutputLinkDPRate[k],
7112 
7113 					/* Output */
7114 					&mode_lib->ms.RequiresDSC[k],
7115 					&mode_lib->ms.RequiresFEC[k],
7116 					&mode_lib->ms.OutputBppPerState[k],
7117 					&mode_lib->ms.OutputTypePerState[k], // VBA_DELTA, VBA uses a string to represent type and rate, but DML uses enum, don't want to rely on strng
7118 					&mode_lib->ms.OutputRatePerState[k],
7119 					&mode_lib->ms.RequiredSlots[k]);
7120 
7121 			if (mode_lib->ms.RequiresDSC[k] == false) {
7122 				mode_lib->ms.ODMModePerState[k] = s->ODMModeNoDSC;
7123 				mode_lib->ms.RequiredDISPCLKPerSurface[j][k] = s->RequiredDISPCLKPerSurfaceNoDSC;
7124 				if (!s->TotalAvailablePipesSupportNoDSC)
7125 					mode_lib->ms.support.TotalAvailablePipesSupport[j] = false;
7126 				mode_lib->ms.TotalNumberOfActiveDPP[j] = mode_lib->ms.TotalNumberOfActiveDPP[j] + s->NumberOfDPPNoDSC;
7127 			} else {
7128 				mode_lib->ms.ODMModePerState[k] = s->ODMModeDSC;
7129 				mode_lib->ms.RequiredDISPCLKPerSurface[j][k] = s->RequiredDISPCLKPerSurfaceDSC;
7130 				if (!s->TotalAvailablePipesSupportDSC)
7131 					mode_lib->ms.support.TotalAvailablePipesSupport[j] = false;
7132 				mode_lib->ms.TotalNumberOfActiveDPP[j] = mode_lib->ms.TotalNumberOfActiveDPP[j] + s->NumberOfDPPDSC;
7133 			}
7134 		}
7135 
7136 		for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7137 			if (mode_lib->ms.ODMModePerState[k] == dml_odm_mode_combine_4to1) {
7138 				mode_lib->ms.MPCCombine[j][k] = false;
7139 				mode_lib->ms.NoOfDPP[j][k] = 4;
7140 			} else if (mode_lib->ms.ODMModePerState[k] == dml_odm_mode_combine_2to1) {
7141 				mode_lib->ms.MPCCombine[j][k] = false;
7142 				mode_lib->ms.NoOfDPP[j][k] = 2;
7143 			} else if (mode_lib->ms.policy.MPCCombineUse[k] == dml_mpc_disabled) {
7144 				mode_lib->ms.MPCCombine[j][k] = false;
7145 				mode_lib->ms.NoOfDPP[j][k] = 1;
7146 			} else if (RoundToDFSGranularity(mode_lib->ms.MinDPPCLKUsingSingleDPP[k] * (1 + mode_lib->ms.soc.dcn_downspread_percent / 100),
7147 											1, mode_lib->ms.soc.dispclk_dppclk_vco_speed_mhz) <= mode_lib->ms.state.dppclk_mhz &&
7148 											mode_lib->ms.SingleDPPViewportSizeSupportPerSurface[k] == true) {
7149 				mode_lib->ms.MPCCombine[j][k] = false;
7150 				mode_lib->ms.NoOfDPP[j][k] = 1;
7151 			} else if (mode_lib->ms.TotalNumberOfActiveDPP[j] < (dml_uint_t) mode_lib->ms.ip.max_num_dpp) {
7152 				mode_lib->ms.MPCCombine[j][k] = true;
7153 				mode_lib->ms.NoOfDPP[j][k] = 2;
7154 				mode_lib->ms.TotalNumberOfActiveDPP[j] = (dml_uint_t) mode_lib->ms.TotalNumberOfActiveDPP[j] + 1;
7155 			} else {
7156 				mode_lib->ms.MPCCombine[j][k] = false;
7157 				mode_lib->ms.NoOfDPP[j][k] = 1;
7158 				mode_lib->ms.support.TotalAvailablePipesSupport[j] = false;
7159 			}
7160 		}
7161 
7162 		mode_lib->ms.TotalNumberOfSingleDPPSurfaces[j] = 0;
7163 		s->NoChromaOrLinear = true;
7164 		for (k = 0; k < (dml_uint_t) mode_lib->ms.num_active_planes; ++k) {
7165 			if (mode_lib->ms.NoOfDPP[j][k] == 1)
7166 				mode_lib->ms.TotalNumberOfSingleDPPSurfaces[j] = mode_lib->ms.TotalNumberOfSingleDPPSurfaces[j] + 1;
7167 			if (mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] == dml_420_8
7168 					|| mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] == dml_420_10
7169 					|| mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] == dml_420_12
7170 					|| mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] == dml_rgbe_alpha
7171 					|| mode_lib->ms.cache_display_cfg.surface.SurfaceTiling[k] == dml_sw_linear) {
7172 				s->NoChromaOrLinear = false;
7173 			}
7174 		}
7175 
7176 		if (j == 1 && !UnboundedRequest(mode_lib->ms.policy.UseUnboundedRequesting,
7177 				mode_lib->ms.TotalNumberOfActiveDPP[j], s->NoChromaOrLinear,
7178 				mode_lib->ms.cache_display_cfg.output.OutputEncoder[0])) {
7179 			while (!(mode_lib->ms.TotalNumberOfActiveDPP[j] >= (dml_uint_t) mode_lib->ms.ip.max_num_dpp || mode_lib->ms.TotalNumberOfSingleDPPSurfaces[j] == 0)) {
7180 				s->BWOfNonCombinedSurfaceOfMaximumBandwidth = 0;
7181 				s->NumberOfNonCombinedSurfaceOfMaximumBandwidth = 0;
7182 				for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7183 					if (mode_lib->ms.policy.MPCCombineUse[k] != dml_mpc_disabled && mode_lib->ms.policy.MPCCombineUse[k] != dml_mpc_as_needed_for_voltage &&
7184 						mode_lib->ms.ReadBandwidthLuma[k] + mode_lib->ms.ReadBandwidthChroma[k] > s->BWOfNonCombinedSurfaceOfMaximumBandwidth &&
7185 						(mode_lib->ms.ODMModePerState[k] != dml_odm_mode_combine_2to1 && mode_lib->ms.ODMModePerState[k] != dml_odm_mode_combine_4to1) &&
7186 						mode_lib->ms.MPCCombine[j][k] == false) {
7187 						s->BWOfNonCombinedSurfaceOfMaximumBandwidth = mode_lib->ms.ReadBandwidthLuma[k] + mode_lib->ms.ReadBandwidthChroma[k];
7188 						s->NumberOfNonCombinedSurfaceOfMaximumBandwidth = k;
7189 					}
7190 				}
7191 				mode_lib->ms.MPCCombine[j][s->NumberOfNonCombinedSurfaceOfMaximumBandwidth] = true;
7192 				mode_lib->ms.NoOfDPP[j][s->NumberOfNonCombinedSurfaceOfMaximumBandwidth] = 2;
7193 				mode_lib->ms.TotalNumberOfActiveDPP[j] = mode_lib->ms.TotalNumberOfActiveDPP[j] + 1;
7194 				mode_lib->ms.TotalNumberOfSingleDPPSurfaces[j] = mode_lib->ms.TotalNumberOfSingleDPPSurfaces[j] - 1;
7195 			}
7196 		}
7197 
7198 		//DISPCLK/DPPCLK
7199 		mode_lib->ms.WritebackRequiredDISPCLK = 0;
7200 		for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7201 			if (mode_lib->ms.cache_display_cfg.writeback.WritebackEnable[k]) {
7202 				mode_lib->ms.WritebackRequiredDISPCLK = dml_max(mode_lib->ms.WritebackRequiredDISPCLK,
7203 																	CalculateWriteBackDISPCLK(mode_lib->ms.cache_display_cfg.writeback.WritebackPixelFormat[k],
7204 																							mode_lib->ms.cache_display_cfg.timing.PixelClock[k],
7205 																							mode_lib->ms.cache_display_cfg.writeback.WritebackHRatio[k],
7206 																							mode_lib->ms.cache_display_cfg.writeback.WritebackVRatio[k],
7207 																							mode_lib->ms.cache_display_cfg.writeback.WritebackHTaps[k],
7208 																							mode_lib->ms.cache_display_cfg.writeback.WritebackVTaps[k],
7209 																							mode_lib->ms.cache_display_cfg.writeback.WritebackSourceWidth[k],
7210 																							mode_lib->ms.cache_display_cfg.writeback.WritebackDestinationWidth[k],
7211 																							mode_lib->ms.cache_display_cfg.timing.HTotal[k],
7212 																							mode_lib->ms.ip.writeback_line_buffer_buffer_size,
7213 																							mode_lib->ms.soc.dispclk_dppclk_vco_speed_mhz));
7214 			}
7215 		}
7216 
7217 		 mode_lib->ms.RequiredDISPCLK[j] = mode_lib->ms.WritebackRequiredDISPCLK;
7218 		for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7219 			mode_lib->ms.RequiredDISPCLK[j] = dml_max(mode_lib->ms.RequiredDISPCLK[j], mode_lib->ms.RequiredDISPCLKPerSurface[j][k]);
7220 		}
7221 
7222 		for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7223 			mode_lib->ms.NoOfDPPThisState[k] = mode_lib->ms.NoOfDPP[j][k];
7224 		}
7225 
7226 		CalculateDPPCLK(mode_lib->ms.num_active_planes,
7227 					mode_lib->ms.soc.dcn_downspread_percent,
7228 					mode_lib->ms.soc.dispclk_dppclk_vco_speed_mhz,
7229 					mode_lib->ms.MinDPPCLKUsingSingleDPP,
7230 					mode_lib->ms.NoOfDPPThisState,
7231 					/* Output */
7232 					&mode_lib->ms.GlobalDPPCLK,
7233 					mode_lib->ms.RequiredDPPCLKThisState);
7234 
7235 		for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7236 			mode_lib->ms.RequiredDPPCLKPerSurface[j][k] = mode_lib->ms.RequiredDPPCLKThisState[k];
7237 		}
7238 
7239 		mode_lib->ms.support.DISPCLK_DPPCLK_Support[j] = !((mode_lib->ms.RequiredDISPCLK[j] > mode_lib->ms.state.dispclk_mhz) || (mode_lib->ms.GlobalDPPCLK > mode_lib->ms.state.dppclk_mhz));
7240 
7241 		if (mode_lib->ms.TotalNumberOfActiveDPP[j] > (dml_uint_t) mode_lib->ms.ip.max_num_dpp) {
7242 			mode_lib->ms.support.TotalAvailablePipesSupport[j] = false;
7243 		}
7244 	} // j
7245 
7246 	/* Total Available OTG, HDMIFRL, DP Support Check */
7247 	s->TotalNumberOfActiveOTG = 0;
7248 	s->TotalNumberOfActiveHDMIFRL = 0;
7249 	s->TotalNumberOfActiveDP2p0 = 0;
7250 	s->TotalNumberOfActiveDP2p0Outputs = 0;
7251 
7252 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7253 		if (mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming[k] == k) {
7254 			s->TotalNumberOfActiveOTG = s->TotalNumberOfActiveOTG + 1;
7255 			if (mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_hdmifrl)
7256 				s->TotalNumberOfActiveHDMIFRL = s->TotalNumberOfActiveHDMIFRL + 1;
7257 			if (mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_dp2p0) {
7258 				s->TotalNumberOfActiveDP2p0 = s->TotalNumberOfActiveDP2p0 + 1;
7259 				if (mode_lib->ms.cache_display_cfg.output.OutputMultistreamId[k] == k || mode_lib->ms.cache_display_cfg.output.OutputMultistreamEn[k] == false) {
7260 					s->TotalNumberOfActiveDP2p0Outputs = s->TotalNumberOfActiveDP2p0Outputs + 1;
7261 				}
7262 			}
7263 		}
7264 	}
7265 
7266 	mode_lib->ms.support.NumberOfOTGSupport      = (s->TotalNumberOfActiveOTG <= (dml_uint_t) mode_lib->ms.ip.max_num_otg);
7267 	mode_lib->ms.support.NumberOfHDMIFRLSupport  = (s->TotalNumberOfActiveHDMIFRL <= (dml_uint_t) mode_lib->ms.ip.max_num_hdmi_frl_outputs);
7268 	mode_lib->ms.support.NumberOfDP2p0Support    = (s->TotalNumberOfActiveDP2p0 <= (dml_uint_t) mode_lib->ms.ip.max_num_dp2p0_streams && s->TotalNumberOfActiveDP2p0Outputs <= (dml_uint_t) mode_lib->ms.ip.max_num_dp2p0_outputs);
7269 
7270 	/* Display IO and DSC Support Check */
7271 	mode_lib->ms.support.NonsupportedDSCInputBPC = false;
7272 	for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
7273 		if (mode_lib->ms.cache_display_cfg.output.OutputDisabled[k] == false &&
7274 			!(mode_lib->ms.cache_display_cfg.output.DSCInputBitPerComponent[k] == 12.0
7275 				|| mode_lib->ms.cache_display_cfg.output.DSCInputBitPerComponent[k] == 10.0
7276 				|| mode_lib->ms.cache_display_cfg.output.DSCInputBitPerComponent[k] == 8.0
7277 				|| mode_lib->ms.cache_display_cfg.output.DSCInputBitPerComponent[k] > (dml_uint_t) mode_lib->ms.ip.maximum_dsc_bits_per_component
7278 				)) {
7279 			mode_lib->ms.support.NonsupportedDSCInputBPC = true;
7280 		}
7281 	}
7282 
7283 	mode_lib->ms.support.ExceededMultistreamSlots = false;
7284 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7285 		if (mode_lib->ms.cache_display_cfg.output.OutputMultistreamId[k] == k) {
7286 			s->TotalSlots = mode_lib->ms.RequiredSlots[k];
7287 			for (j = 0; j < mode_lib->ms.num_active_planes; ++j) {
7288 				if (mode_lib->ms.cache_display_cfg.output.OutputMultistreamId[j] == k)
7289 					s->TotalSlots = s->TotalSlots + mode_lib->ms.RequiredSlots[j];
7290 			}
7291 			if (mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_dp && s->TotalSlots > 63)
7292 				mode_lib->ms.support.ExceededMultistreamSlots = true;
7293 			if (mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_dp2p0 && s->TotalSlots > 64)
7294 				mode_lib->ms.support.ExceededMultistreamSlots = true;
7295 		}
7296 	}
7297 	mode_lib->ms.support.LinkCapacitySupport = true;
7298 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7299 		if (mode_lib->ms.cache_display_cfg.output.OutputDisabled[k] == false &&
7300 			mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming[k] == k && (mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_dp || mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_dp2p0 || mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_edp ||
7301 			mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_hdmi || mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_hdmifrl) && mode_lib->ms.OutputBppPerState[k] == 0) {
7302 			mode_lib->ms.support.LinkCapacitySupport = false;
7303 		}
7304 	}
7305 
7306 	mode_lib->ms.support.P2IWith420 = false;
7307 	mode_lib->ms.support.DSCOnlyIfNecessaryWithBPP = false;
7308 	mode_lib->ms.support.DSC422NativeNotSupported = false;
7309 	mode_lib->ms.support.LinkRateDoesNotMatchDPVersion = false;
7310 	mode_lib->ms.support.LinkRateForMultistreamNotIndicated = false;
7311 	mode_lib->ms.support.BPPForMultistreamNotIndicated = false;
7312 	mode_lib->ms.support.MultistreamWithHDMIOreDP = false;
7313 	mode_lib->ms.support.MSOOrODMSplitWithNonDPLink = false;
7314 	mode_lib->ms.support.NotEnoughLanesForMSO = false;
7315 
7316 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7317 		if (mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming[k] == k && (mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_dp || mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_dp2p0 || mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_edp ||
7318 														mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_hdmi || mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_hdmifrl)) {
7319 			if (mode_lib->ms.cache_display_cfg.output.OutputFormat[k] == dml_420 && mode_lib->ms.cache_display_cfg.timing.Interlace[k] == 1 && mode_lib->ms.ip.ptoi_supported == true)
7320 				mode_lib->ms.support.P2IWith420 = true;
7321 
7322 			if (mode_lib->ms.cache_display_cfg.output.DSCEnable[k] == dml_dsc_enable_if_necessary && mode_lib->ms.cache_display_cfg.output.ForcedOutputLinkBPP[k] != 0)
7323 				mode_lib->ms.support.DSCOnlyIfNecessaryWithBPP = true;
7324 			if ((mode_lib->ms.cache_display_cfg.output.DSCEnable[k] == dml_dsc_enable || mode_lib->ms.cache_display_cfg.output.DSCEnable[k] == dml_dsc_enable_if_necessary) && mode_lib->ms.cache_display_cfg.output.OutputFormat[k] == dml_n422 && !mode_lib->ms.ip.dsc422_native_support)
7325 				mode_lib->ms.support.DSC422NativeNotSupported = true;
7326 
7327 			if (((mode_lib->ms.cache_display_cfg.output.OutputLinkDPRate[k] == dml_dp_rate_hbr || mode_lib->ms.cache_display_cfg.output.OutputLinkDPRate[k] == dml_dp_rate_hbr2 || mode_lib->ms.cache_display_cfg.output.OutputLinkDPRate[k] == dml_dp_rate_hbr3) &&
7328 					mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] != dml_dp && mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] != dml_edp) ||
7329 					((mode_lib->ms.cache_display_cfg.output.OutputLinkDPRate[k] == dml_dp_rate_uhbr10 || mode_lib->ms.cache_display_cfg.output.OutputLinkDPRate[k] == dml_dp_rate_uhbr13p5 || mode_lib->ms.cache_display_cfg.output.OutputLinkDPRate[k] == dml_dp_rate_uhbr20) &&
7330 					mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] != dml_dp2p0))
7331 				mode_lib->ms.support.LinkRateDoesNotMatchDPVersion = true;
7332 
7333 			if (mode_lib->ms.cache_display_cfg.output.OutputMultistreamEn[k] == 1) {
7334 				if (mode_lib->ms.cache_display_cfg.output.OutputMultistreamId[k] == k && mode_lib->ms.cache_display_cfg.output.OutputLinkDPRate[k] == dml_dp_rate_na)
7335 					mode_lib->ms.support.LinkRateForMultistreamNotIndicated = true;
7336 				if (mode_lib->ms.cache_display_cfg.output.OutputMultistreamId[k] == k && mode_lib->ms.cache_display_cfg.output.ForcedOutputLinkBPP[k] == 0)
7337 					mode_lib->ms.support.BPPForMultistreamNotIndicated = true;
7338 				for (j = 0; j < mode_lib->ms.num_active_planes; ++j) {
7339 					if (mode_lib->ms.cache_display_cfg.output.OutputMultistreamId[k] == j && mode_lib->ms.cache_display_cfg.output.ForcedOutputLinkBPP[k] == 0)
7340 						mode_lib->ms.support.BPPForMultistreamNotIndicated = true;
7341 				}
7342 			}
7343 
7344 			if ((mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_edp || mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_hdmi || mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_hdmifrl)) {
7345 				if (mode_lib->ms.cache_display_cfg.output.OutputMultistreamEn[k] == 1 && mode_lib->ms.cache_display_cfg.output.OutputMultistreamId[k] == k)
7346 					mode_lib->ms.support.MultistreamWithHDMIOreDP = true;
7347 				for (j = 0; j < mode_lib->ms.num_active_planes; ++j) {
7348 					if (mode_lib->ms.cache_display_cfg.output.OutputMultistreamEn[k] == 1 && mode_lib->ms.cache_display_cfg.output.OutputMultistreamId[k] == j)
7349 						mode_lib->ms.support.MultistreamWithHDMIOreDP = true;
7350 				}
7351 			}
7352 			if (mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] != dml_dp && (mode_lib->ms.policy.ODMUse[k] == dml_odm_use_policy_split_1to2 ||
7353 				mode_lib->ms.policy.ODMUse[k] == dml_odm_use_policy_mso_1to2 || mode_lib->ms.policy.ODMUse[k] == dml_odm_use_policy_mso_1to4))
7354 				mode_lib->ms.support.MSOOrODMSplitWithNonDPLink = true;
7355 
7356 			if ((mode_lib->ms.policy.ODMUse[k] == dml_odm_use_policy_mso_1to2 && mode_lib->ms.cache_display_cfg.output.OutputLinkDPLanes[k] < 2) ||
7357 				(mode_lib->ms.policy.ODMUse[k] == dml_odm_use_policy_mso_1to4 && mode_lib->ms.cache_display_cfg.output.OutputLinkDPLanes[k] < 4))
7358 				mode_lib->ms.support.NotEnoughLanesForMSO = true;
7359 		}
7360 	}
7361 
7362 	mode_lib->ms.support.DTBCLKRequiredMoreThanSupported = false;
7363 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7364 		if (mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming[k] == k &&
7365 				mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_hdmifrl &&
7366 				RequiredDTBCLK(
7367 							mode_lib->ms.RequiresDSC[k],
7368 							mode_lib->ms.cache_display_cfg.output.PixelClockBackEnd[k],
7369 							mode_lib->ms.cache_display_cfg.output.OutputFormat[k],
7370 							mode_lib->ms.OutputBppPerState[k],
7371 							mode_lib->ms.support.NumberOfDSCSlices[k],
7372 							mode_lib->ms.cache_display_cfg.timing.HTotal[k],
7373 							mode_lib->ms.cache_display_cfg.timing.HActive[k],
7374 							mode_lib->ms.cache_display_cfg.output.AudioSampleRate[k],
7375 							mode_lib->ms.cache_display_cfg.output.AudioSampleLayout[k]) > mode_lib->ms.state.dtbclk_mhz) {
7376 								mode_lib->ms.support.DTBCLKRequiredMoreThanSupported = true;
7377 							}
7378 	}
7379 
7380 	mode_lib->ms.support.ODMCombineTwoToOneSupportCheckOK = true;
7381 	mode_lib->ms.support.ODMCombineFourToOneSupportCheckOK = true;
7382 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7383 		if (mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming[k] == k && mode_lib->ms.ODMModePerState[k] == dml_odm_mode_combine_2to1 && mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_hdmi) {
7384 			mode_lib->ms.support.ODMCombineTwoToOneSupportCheckOK = false;
7385 		}
7386 		if (mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming[k] == k && mode_lib->ms.ODMModePerState[k] == dml_odm_mode_combine_4to1 && (mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_dp ||
7387 			mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_edp || mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_hdmi)) {
7388 			mode_lib->ms.support.ODMCombineFourToOneSupportCheckOK = false;
7389 		}
7390 	}
7391 
7392 	mode_lib->ms.support.DSCCLKRequiredMoreThanSupported = false;
7393 	for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
7394 		if (mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming[k] == k) {
7395 			if (mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_dp ||
7396 				mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_dp2p0 ||
7397 				mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_edp ||
7398 				mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_hdmifrl) {
7399 				if (mode_lib->ms.cache_display_cfg.output.OutputFormat[k] == dml_420) {
7400 					s->DSCFormatFactor = 2;
7401 				} else if (mode_lib->ms.cache_display_cfg.output.OutputFormat[k] == dml_444) {
7402 					s->DSCFormatFactor = 1;
7403 				} else if (mode_lib->ms.cache_display_cfg.output.OutputFormat[k] == dml_n422 || mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_hdmifrl) {
7404 					s->DSCFormatFactor = 2;
7405 				} else {
7406 					s->DSCFormatFactor = 1;
7407 				}
7408 #ifdef __DML_VBA_DEBUG__
7409 				dml_print("DML::%s: k=%u, RequiresDSC = %u\n",  __func__, k, mode_lib->ms.RequiresDSC[k]);
7410 #endif
7411 				if (mode_lib->ms.RequiresDSC[k] == true) {
7412 					if (mode_lib->ms.ODMModePerState[k] == dml_odm_mode_combine_4to1) {
7413 						if (mode_lib->ms.cache_display_cfg.output.PixelClockBackEnd[k] / 12.0 / (dml_float_t)s->DSCFormatFactor > (1.0 - mode_lib->ms.soc.dcn_downspread_percent / 100.0) * mode_lib->ms.state.dscclk_mhz) {
7414 #ifdef __DML_VBA_DEBUG__
7415 							dml_print("DML::%s: k=%u, PixelClockBackEnd     = %f\n",  __func__, k, mode_lib->ms.cache_display_cfg.output.PixelClockBackEnd[k]);
7416 							dml_print("DML::%s: k=%u, DSCCLKPerState        = %f\n",  __func__, k, mode_lib->ms.state.dscclk_mhz);
7417 							dml_print("DML::%s: k=%u, DSCFormatFactor       = %u\n",  __func__, k, s->DSCFormatFactor);
7418 #endif
7419 							mode_lib->ms.support.DSCCLKRequiredMoreThanSupported = true;
7420 						}
7421 					} else if (mode_lib->ms.ODMModePerState[k] == dml_odm_mode_combine_2to1) {
7422 						if (mode_lib->ms.cache_display_cfg.output.PixelClockBackEnd[k] / 6.0 / (dml_float_t)s->DSCFormatFactor > (1.0 - mode_lib->ms.soc.dcn_downspread_percent / 100.0) * mode_lib->ms.state.dscclk_mhz) {
7423 							mode_lib->ms.support.DSCCLKRequiredMoreThanSupported = true;
7424 						}
7425 					} else {
7426 						if (mode_lib->ms.cache_display_cfg.output.PixelClockBackEnd[k] / 3.0 / (dml_float_t)s->DSCFormatFactor > (1.0 - mode_lib->ms.soc.dcn_downspread_percent / 100.0) * mode_lib->ms.state.dscclk_mhz) {
7427 							mode_lib->ms.support.DSCCLKRequiredMoreThanSupported = true;
7428 						}
7429 					}
7430 				}
7431 			}
7432 		}
7433 	}
7434 #ifdef __DML_VBA_DEBUG__
7435 	dml_print("DML::%s: DSCCLKRequiredMoreThanSupported = %u\n",  __func__, mode_lib->ms.support.DSCCLKRequiredMoreThanSupported);
7436 #endif
7437 
7438 	/* Check DSC Unit and Slices Support */
7439 	mode_lib->ms.support.NotEnoughDSCUnits = false;
7440 	mode_lib->ms.support.NotEnoughDSCSlices = false;
7441 	s->TotalDSCUnitsRequired = 0;
7442 	mode_lib->ms.support.PixelsPerLinePerDSCUnitSupport = true;
7443 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7444 		if (mode_lib->ms.RequiresDSC[k] == true) {
7445 			if (mode_lib->ms.ODMModePerState[k] == dml_odm_mode_combine_4to1) {
7446 				if (mode_lib->ms.cache_display_cfg.timing.HActive[k] > 4 * (dml_uint_t) mode_lib->ms.ip.maximum_pixels_per_line_per_dsc_unit)
7447 					mode_lib->ms.support.PixelsPerLinePerDSCUnitSupport = false;
7448 				s->TotalDSCUnitsRequired = s->TotalDSCUnitsRequired + 4;
7449 				if (mode_lib->ms.support.NumberOfDSCSlices[k] > 16)
7450 					mode_lib->ms.support.NotEnoughDSCSlices = true;
7451 			} else if (mode_lib->ms.ODMModePerState[k] == dml_odm_mode_combine_2to1) {
7452 				if (mode_lib->ms.cache_display_cfg.timing.HActive[k] > 2 * (dml_uint_t) mode_lib->ms.ip.maximum_pixels_per_line_per_dsc_unit)
7453 					mode_lib->ms.support.PixelsPerLinePerDSCUnitSupport = false;
7454 				s->TotalDSCUnitsRequired = s->TotalDSCUnitsRequired + 2;
7455 				if (mode_lib->ms.support.NumberOfDSCSlices[k] > 8)
7456 					mode_lib->ms.support.NotEnoughDSCSlices = true;
7457 			} else {
7458 				if (mode_lib->ms.cache_display_cfg.timing.HActive[k] > (dml_uint_t) mode_lib->ms.ip.maximum_pixels_per_line_per_dsc_unit)
7459 					mode_lib->ms.support.PixelsPerLinePerDSCUnitSupport = false;
7460 				s->TotalDSCUnitsRequired = s->TotalDSCUnitsRequired + 1;
7461 				if (mode_lib->ms.support.NumberOfDSCSlices[k] > 4)
7462 					mode_lib->ms.support.NotEnoughDSCSlices = true;
7463 			}
7464 		}
7465 	}
7466    if (s->TotalDSCUnitsRequired > (dml_uint_t) mode_lib->ms.ip.num_dsc) {
7467 		mode_lib->ms.support.NotEnoughDSCUnits = true;
7468 	}
7469 
7470 	/*DSC Delay per state*/
7471 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7472 		mode_lib->ms.DSCDelayPerState[k] = DSCDelayRequirement(mode_lib->ms.RequiresDSC[k],
7473 													mode_lib->ms.ODMModePerState[k],
7474 													mode_lib->ms.cache_display_cfg.output.DSCInputBitPerComponent[k],
7475 													mode_lib->ms.OutputBppPerState[k],
7476 													mode_lib->ms.cache_display_cfg.timing.HActive[k],
7477 													mode_lib->ms.cache_display_cfg.timing.HTotal[k],
7478 													mode_lib->ms.support.NumberOfDSCSlices[k],
7479 													mode_lib->ms.cache_display_cfg.output.OutputFormat[k],
7480 													mode_lib->ms.cache_display_cfg.output.OutputEncoder[k],
7481 													mode_lib->ms.cache_display_cfg.timing.PixelClock[k],
7482 													mode_lib->ms.cache_display_cfg.output.PixelClockBackEnd[k]);
7483 	}
7484 
7485 	for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
7486 		for (m = 0; m <= mode_lib->ms.num_active_planes - 1; m++) {
7487 			for (j = 0; j <= mode_lib->ms.num_active_planes - 1; j++) {
7488 				if (mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming[k] == m && mode_lib->ms.RequiresDSC[m] == true) {
7489 					mode_lib->ms.DSCDelayPerState[k] = mode_lib->ms.DSCDelayPerState[m];
7490 				}
7491 			}
7492 		}
7493 	}
7494 
7495 	//Calculate Swath, DET Configuration, DCFCLKDeepSleep
7496 	//
7497 	for (j = 0; j < 2; ++j) {
7498 		for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7499 			mode_lib->ms.RequiredDPPCLKThisState[k] = mode_lib->ms.RequiredDPPCLKPerSurface[j][k];
7500 			mode_lib->ms.NoOfDPPThisState[k] = mode_lib->ms.NoOfDPP[j][k];
7501 			mode_lib->ms.ODMModeThisState[k] = mode_lib->ms.ODMModePerState[k];
7502 		}
7503 
7504 		CalculateSwathAndDETConfiguration_params->DETSizeOverride = mode_lib->ms.cache_display_cfg.plane.DETSizeOverride;
7505 		CalculateSwathAndDETConfiguration_params->UseMALLForPStateChange = mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange;
7506 		CalculateSwathAndDETConfiguration_params->ConfigReturnBufferSizeInKByte = mode_lib->ms.ip.config_return_buffer_size_in_kbytes;
7507 		CalculateSwathAndDETConfiguration_params->ROBBufferSizeInKByte = mode_lib->ms.ip.rob_buffer_size_kbytes;
7508 		CalculateSwathAndDETConfiguration_params->MaxTotalDETInKByte = mode_lib->ms.MaxTotalDETInKByte;
7509 		CalculateSwathAndDETConfiguration_params->MinCompressedBufferSizeInKByte = mode_lib->ms.MinCompressedBufferSizeInKByte;
7510 		CalculateSwathAndDETConfiguration_params->PixelChunkSizeInKByte = mode_lib->ms.ip.pixel_chunk_size_kbytes;
7511 		CalculateSwathAndDETConfiguration_params->ForceSingleDPP = false;
7512 		CalculateSwathAndDETConfiguration_params->NumberOfActiveSurfaces = mode_lib->ms.num_active_planes;
7513 		CalculateSwathAndDETConfiguration_params->nomDETInKByte = mode_lib->ms.NomDETInKByte;
7514 		CalculateSwathAndDETConfiguration_params->UseUnboundedRequestingFinal = mode_lib->ms.policy.UseUnboundedRequesting;
7515 		CalculateSwathAndDETConfiguration_params->ConfigReturnBufferSegmentSizeInkByte = mode_lib->ms.ip.config_return_buffer_segment_size_in_kbytes;
7516 		CalculateSwathAndDETConfiguration_params->CompressedBufferSegmentSizeInkByteFinal = mode_lib->ms.ip.compressed_buffer_segment_size_in_kbytes;
7517 		CalculateSwathAndDETConfiguration_params->Output = mode_lib->ms.cache_display_cfg.output.OutputEncoder;
7518 		CalculateSwathAndDETConfiguration_params->ReadBandwidthLuma = mode_lib->ms.ReadBandwidthLuma;
7519 		CalculateSwathAndDETConfiguration_params->ReadBandwidthChroma = mode_lib->ms.ReadBandwidthChroma;
7520 		CalculateSwathAndDETConfiguration_params->MaximumSwathWidthLuma = mode_lib->ms.MaximumSwathWidthLuma;
7521 		CalculateSwathAndDETConfiguration_params->MaximumSwathWidthChroma = mode_lib->ms.MaximumSwathWidthChroma;
7522 		CalculateSwathAndDETConfiguration_params->SourceScan = mode_lib->ms.cache_display_cfg.plane.SourceScan;
7523 		CalculateSwathAndDETConfiguration_params->ViewportStationary = mode_lib->ms.cache_display_cfg.plane.ViewportStationary;
7524 		CalculateSwathAndDETConfiguration_params->SourcePixelFormat = mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat;
7525 		CalculateSwathAndDETConfiguration_params->SurfaceTiling = mode_lib->ms.cache_display_cfg.surface.SurfaceTiling;
7526 		CalculateSwathAndDETConfiguration_params->ViewportWidth = mode_lib->ms.cache_display_cfg.plane.ViewportWidth;
7527 		CalculateSwathAndDETConfiguration_params->ViewportHeight = mode_lib->ms.cache_display_cfg.plane.ViewportHeight;
7528 		CalculateSwathAndDETConfiguration_params->ViewportXStart = mode_lib->ms.cache_display_cfg.plane.ViewportXStart;
7529 		CalculateSwathAndDETConfiguration_params->ViewportYStart = mode_lib->ms.cache_display_cfg.plane.ViewportYStart;
7530 		CalculateSwathAndDETConfiguration_params->ViewportXStartC = mode_lib->ms.cache_display_cfg.plane.ViewportXStartC;
7531 		CalculateSwathAndDETConfiguration_params->ViewportYStartC = mode_lib->ms.cache_display_cfg.plane.ViewportYStartC;
7532 		CalculateSwathAndDETConfiguration_params->SurfaceWidthY = mode_lib->ms.cache_display_cfg.surface.SurfaceWidthY;
7533 		CalculateSwathAndDETConfiguration_params->SurfaceWidthC = mode_lib->ms.cache_display_cfg.surface.SurfaceWidthC;
7534 		CalculateSwathAndDETConfiguration_params->SurfaceHeightY = mode_lib->ms.cache_display_cfg.surface.SurfaceHeightY;
7535 		CalculateSwathAndDETConfiguration_params->SurfaceHeightC = mode_lib->ms.cache_display_cfg.surface.SurfaceHeightC;
7536 		CalculateSwathAndDETConfiguration_params->Read256BytesBlockHeightY = mode_lib->ms.Read256BlockHeightY;
7537 		CalculateSwathAndDETConfiguration_params->Read256BytesBlockHeightC = mode_lib->ms.Read256BlockHeightC;
7538 		CalculateSwathAndDETConfiguration_params->Read256BytesBlockWidthY = mode_lib->ms.Read256BlockWidthY;
7539 		CalculateSwathAndDETConfiguration_params->Read256BytesBlockWidthC = mode_lib->ms.Read256BlockWidthC;
7540 		CalculateSwathAndDETConfiguration_params->ODMMode = mode_lib->ms.ODMModeThisState;
7541 		CalculateSwathAndDETConfiguration_params->BlendingAndTiming = mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming;
7542 		CalculateSwathAndDETConfiguration_params->BytePerPixY = mode_lib->ms.BytePerPixelY;
7543 		CalculateSwathAndDETConfiguration_params->BytePerPixC = mode_lib->ms.BytePerPixelC;
7544 		CalculateSwathAndDETConfiguration_params->BytePerPixDETY = mode_lib->ms.BytePerPixelInDETY;
7545 		CalculateSwathAndDETConfiguration_params->BytePerPixDETC = mode_lib->ms.BytePerPixelInDETC;
7546 		CalculateSwathAndDETConfiguration_params->HActive = mode_lib->ms.cache_display_cfg.timing.HActive;
7547 		CalculateSwathAndDETConfiguration_params->HRatio = mode_lib->ms.cache_display_cfg.plane.HRatio;
7548 		CalculateSwathAndDETConfiguration_params->HRatioChroma = mode_lib->ms.cache_display_cfg.plane.HRatioChroma;
7549 		CalculateSwathAndDETConfiguration_params->DPPPerSurface = mode_lib->ms.NoOfDPPThisState;
7550 		CalculateSwathAndDETConfiguration_params->swath_width_luma_ub = mode_lib->ms.swath_width_luma_ub_this_state;
7551 		CalculateSwathAndDETConfiguration_params->swath_width_chroma_ub = mode_lib->ms.swath_width_chroma_ub_this_state;
7552 		CalculateSwathAndDETConfiguration_params->SwathWidth = mode_lib->ms.SwathWidthYThisState;
7553 		CalculateSwathAndDETConfiguration_params->SwathWidthChroma = mode_lib->ms.SwathWidthCThisState;
7554 		CalculateSwathAndDETConfiguration_params->SwathHeightY = mode_lib->ms.SwathHeightYThisState;
7555 		CalculateSwathAndDETConfiguration_params->SwathHeightC = mode_lib->ms.SwathHeightCThisState;
7556 		CalculateSwathAndDETConfiguration_params->DETBufferSizeInKByte = mode_lib->ms.DETBufferSizeInKByteThisState;
7557 		CalculateSwathAndDETConfiguration_params->DETBufferSizeY = mode_lib->ms.DETBufferSizeYThisState;
7558 		CalculateSwathAndDETConfiguration_params->DETBufferSizeC = mode_lib->ms.DETBufferSizeCThisState;
7559 		CalculateSwathAndDETConfiguration_params->UnboundedRequestEnabled = &mode_lib->ms.UnboundedRequestEnabledThisState;
7560 		CalculateSwathAndDETConfiguration_params->compbuf_reserved_space_64b = &s->dummy_integer[2];
7561 		CalculateSwathAndDETConfiguration_params->compbuf_reserved_space_zs = &s->dummy_integer[1];
7562 		CalculateSwathAndDETConfiguration_params->CompressedBufferSizeInkByte = &mode_lib->ms.CompressedBufferSizeInkByteThisState;
7563 		CalculateSwathAndDETConfiguration_params->ViewportSizeSupportPerSurface = s->dummy_boolean_array[0];
7564 		CalculateSwathAndDETConfiguration_params->ViewportSizeSupport = &mode_lib->ms.support.ViewportSizeSupport[j];
7565 
7566 		CalculateSwathAndDETConfiguration(&mode_lib->scratch,
7567 		CalculateSwathAndDETConfiguration_params);
7568 
7569 		for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7570 			mode_lib->ms.swath_width_luma_ub_all_states[j][k] = mode_lib->ms.swath_width_luma_ub_this_state[k];
7571 			mode_lib->ms.swath_width_chroma_ub_all_states[j][k] = mode_lib->ms.swath_width_chroma_ub_this_state[k];
7572 			mode_lib->ms.SwathWidthYAllStates[j][k] = mode_lib->ms.SwathWidthYThisState[k];
7573 			mode_lib->ms.SwathWidthCAllStates[j][k] = mode_lib->ms.SwathWidthCThisState[k];
7574 			mode_lib->ms.SwathHeightYAllStates[j][k] = mode_lib->ms.SwathHeightYThisState[k];
7575 			mode_lib->ms.SwathHeightCAllStates[j][k] = mode_lib->ms.SwathHeightCThisState[k];
7576 			mode_lib->ms.UnboundedRequestEnabledAllStates[j] = mode_lib->ms.UnboundedRequestEnabledThisState;
7577 			mode_lib->ms.CompressedBufferSizeInkByteAllStates[j] = mode_lib->ms.CompressedBufferSizeInkByteThisState;
7578 			mode_lib->ms.DETBufferSizeInKByteAllStates[j][k] = mode_lib->ms.DETBufferSizeInKByteThisState[k];
7579 			mode_lib->ms.DETBufferSizeYAllStates[j][k] = mode_lib->ms.DETBufferSizeYThisState[k];
7580 			mode_lib->ms.DETBufferSizeCAllStates[j][k] = mode_lib->ms.DETBufferSizeCThisState[k];
7581 		}
7582 	}
7583 
7584 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7585 		mode_lib->ms.cursor_bw[k] = mode_lib->ms.cache_display_cfg.plane.NumberOfCursors[k] * mode_lib->ms.cache_display_cfg.plane.CursorWidth[k] * mode_lib->ms.cache_display_cfg.plane.CursorBPP[k] / 8.0 / (mode_lib->ms.cache_display_cfg.timing.HTotal[k] / mode_lib->ms.cache_display_cfg.timing.PixelClock[k]) * mode_lib->ms.cache_display_cfg.plane.VRatio[k];
7586 	}
7587 
7588 	CalculateSurfaceSizeInMall(
7589 			mode_lib->ms.num_active_planes,
7590 			mode_lib->ms.soc.mall_allocated_for_dcn_mbytes,
7591 			mode_lib->ms.cache_display_cfg.plane.UseMALLForStaticScreen,
7592 			mode_lib->ms.cache_display_cfg.surface.DCCEnable,
7593 			mode_lib->ms.cache_display_cfg.plane.ViewportStationary,
7594 			mode_lib->ms.cache_display_cfg.plane.ViewportXStart,
7595 			mode_lib->ms.cache_display_cfg.plane.ViewportYStart,
7596 			mode_lib->ms.cache_display_cfg.plane.ViewportXStartC,
7597 			mode_lib->ms.cache_display_cfg.plane.ViewportYStartC,
7598 			mode_lib->ms.cache_display_cfg.plane.ViewportWidth,
7599 			mode_lib->ms.cache_display_cfg.plane.ViewportHeight,
7600 			mode_lib->ms.BytePerPixelY,
7601 			mode_lib->ms.cache_display_cfg.plane.ViewportWidthChroma,
7602 			mode_lib->ms.cache_display_cfg.plane.ViewportHeightChroma,
7603 			mode_lib->ms.BytePerPixelC,
7604 			mode_lib->ms.cache_display_cfg.surface.SurfaceWidthY,
7605 			mode_lib->ms.cache_display_cfg.surface.SurfaceWidthC,
7606 			mode_lib->ms.cache_display_cfg.surface.SurfaceHeightY,
7607 			mode_lib->ms.cache_display_cfg.surface.SurfaceHeightC,
7608 			mode_lib->ms.Read256BlockWidthY,
7609 			mode_lib->ms.Read256BlockWidthC,
7610 			mode_lib->ms.Read256BlockHeightY,
7611 			mode_lib->ms.Read256BlockHeightC,
7612 			mode_lib->ms.MacroTileWidthY,
7613 			mode_lib->ms.MacroTileWidthC,
7614 			mode_lib->ms.MacroTileHeightY,
7615 			mode_lib->ms.MacroTileHeightC,
7616 
7617 			/* Output */
7618 			mode_lib->ms.SurfaceSizeInMALL,
7619 			&mode_lib->ms.support.ExceededMALLSize);
7620 
7621 	for (j = 0; j < 2; j++) {
7622 		for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
7623 			mode_lib->ms.swath_width_luma_ub_this_state[k] = mode_lib->ms.swath_width_luma_ub_all_states[j][k];
7624 			mode_lib->ms.swath_width_chroma_ub_this_state[k] = mode_lib->ms.swath_width_chroma_ub_all_states[j][k];
7625 			mode_lib->ms.SwathWidthYThisState[k] = mode_lib->ms.SwathWidthYAllStates[j][k];
7626 			mode_lib->ms.SwathWidthCThisState[k] = mode_lib->ms.SwathWidthCAllStates[j][k];
7627 			mode_lib->ms.SwathHeightYThisState[k] = mode_lib->ms.SwathHeightYAllStates[j][k];
7628 			mode_lib->ms.SwathHeightCThisState[k] = mode_lib->ms.SwathHeightCAllStates[j][k];
7629 			mode_lib->ms.DETBufferSizeInKByteThisState[k] = mode_lib->ms.DETBufferSizeInKByteAllStates[j][k];
7630 			mode_lib->ms.DETBufferSizeYThisState[k] = mode_lib->ms.DETBufferSizeYAllStates[j][k];
7631 			mode_lib->ms.DETBufferSizeCThisState[k] = mode_lib->ms.DETBufferSizeCAllStates[j][k];
7632 			mode_lib->ms.RequiredDPPCLKThisState[k] = mode_lib->ms.RequiredDPPCLKPerSurface[j][k];
7633 			mode_lib->ms.NoOfDPPThisState[k] = mode_lib->ms.NoOfDPP[j][k];
7634 		}
7635 
7636 		mode_lib->ms.TotalNumberOfDCCActiveDPP[j] = 0;
7637 		for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7638 			if (mode_lib->ms.cache_display_cfg.surface.DCCEnable[k] == true) {
7639 				mode_lib->ms.TotalNumberOfDCCActiveDPP[j] = mode_lib->ms.TotalNumberOfDCCActiveDPP[j] + mode_lib->ms.NoOfDPP[j][k];
7640 			}
7641 		}
7642 
7643 		for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7644 			s->SurfParameters[k].PixelClock = mode_lib->ms.cache_display_cfg.timing.PixelClock[k];
7645 			s->SurfParameters[k].DPPPerSurface = mode_lib->ms.NoOfDPP[j][k];
7646 			s->SurfParameters[k].SourceScan = mode_lib->ms.cache_display_cfg.plane.SourceScan[k];
7647 			s->SurfParameters[k].ViewportHeight = mode_lib->ms.cache_display_cfg.plane.ViewportHeight[k];
7648 			s->SurfParameters[k].ViewportHeightChroma = mode_lib->ms.cache_display_cfg.plane.ViewportHeightChroma[k];
7649 			s->SurfParameters[k].BlockWidth256BytesY = mode_lib->ms.Read256BlockWidthY[k];
7650 			s->SurfParameters[k].BlockHeight256BytesY = mode_lib->ms.Read256BlockHeightY[k];
7651 			s->SurfParameters[k].BlockWidth256BytesC = mode_lib->ms.Read256BlockWidthC[k];
7652 			s->SurfParameters[k].BlockHeight256BytesC = mode_lib->ms.Read256BlockHeightC[k];
7653 			s->SurfParameters[k].BlockWidthY = mode_lib->ms.MacroTileWidthY[k];
7654 			s->SurfParameters[k].BlockHeightY = mode_lib->ms.MacroTileHeightY[k];
7655 			s->SurfParameters[k].BlockWidthC = mode_lib->ms.MacroTileWidthC[k];
7656 			s->SurfParameters[k].BlockHeightC = mode_lib->ms.MacroTileHeightC[k];
7657 			s->SurfParameters[k].InterlaceEnable = mode_lib->ms.cache_display_cfg.timing.Interlace[k];
7658 			s->SurfParameters[k].HTotal = mode_lib->ms.cache_display_cfg.timing.HTotal[k];
7659 			s->SurfParameters[k].DCCEnable = mode_lib->ms.cache_display_cfg.surface.DCCEnable[k];
7660 			s->SurfParameters[k].SourcePixelFormat = mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k];
7661 			s->SurfParameters[k].SurfaceTiling = mode_lib->ms.cache_display_cfg.surface.SurfaceTiling[k];
7662 			s->SurfParameters[k].BytePerPixelY = mode_lib->ms.BytePerPixelY[k];
7663 			s->SurfParameters[k].BytePerPixelC = mode_lib->ms.BytePerPixelC[k];
7664 			s->SurfParameters[k].ProgressiveToInterlaceUnitInOPP = mode_lib->ms.ip.ptoi_supported;
7665 			s->SurfParameters[k].VRatio = mode_lib->ms.cache_display_cfg.plane.VRatio[k];
7666 			s->SurfParameters[k].VRatioChroma = mode_lib->ms.cache_display_cfg.plane.VRatioChroma[k];
7667 			s->SurfParameters[k].VTaps = mode_lib->ms.cache_display_cfg.plane.VTaps[k];
7668 			s->SurfParameters[k].VTapsChroma = mode_lib->ms.cache_display_cfg.plane.VTapsChroma[k];
7669 			s->SurfParameters[k].PitchY = mode_lib->ms.cache_display_cfg.surface.PitchY[k];
7670 			s->SurfParameters[k].DCCMetaPitchY = mode_lib->ms.cache_display_cfg.surface.DCCMetaPitchY[k];
7671 			s->SurfParameters[k].PitchC = mode_lib->ms.cache_display_cfg.surface.PitchC[k];
7672 			s->SurfParameters[k].DCCMetaPitchC = mode_lib->ms.cache_display_cfg.surface.DCCMetaPitchC[k];
7673 			s->SurfParameters[k].ViewportStationary = mode_lib->ms.cache_display_cfg.plane.ViewportStationary[k];
7674 			s->SurfParameters[k].ViewportXStart = mode_lib->ms.cache_display_cfg.plane.ViewportXStart[k];
7675 			s->SurfParameters[k].ViewportYStart = mode_lib->ms.cache_display_cfg.plane.ViewportYStart[k];
7676 			s->SurfParameters[k].ViewportXStartC = mode_lib->ms.cache_display_cfg.plane.ViewportXStartC[k];
7677 			s->SurfParameters[k].ViewportYStartC = mode_lib->ms.cache_display_cfg.plane.ViewportYStartC[k];
7678 			s->SurfParameters[k].FORCE_ONE_ROW_FOR_FRAME = mode_lib->ms.cache_display_cfg.plane.ForceOneRowForFrame[k];
7679 			s->SurfParameters[k].SwathHeightY = mode_lib->ms.SwathHeightYThisState[k];
7680 			s->SurfParameters[k].SwathHeightC = mode_lib->ms.SwathHeightCThisState[k];
7681 		}
7682 
7683 		CalculateVMRowAndSwath_params->NumberOfActiveSurfaces = mode_lib->ms.num_active_planes;
7684 		CalculateVMRowAndSwath_params->myPipe = s->SurfParameters;
7685 		CalculateVMRowAndSwath_params->SurfaceSizeInMALL = mode_lib->ms.SurfaceSizeInMALL;
7686 		CalculateVMRowAndSwath_params->PTEBufferSizeInRequestsLuma = mode_lib->ms.ip.dpte_buffer_size_in_pte_reqs_luma;
7687 		CalculateVMRowAndSwath_params->PTEBufferSizeInRequestsChroma = mode_lib->ms.ip.dpte_buffer_size_in_pte_reqs_chroma;
7688 		CalculateVMRowAndSwath_params->DCCMetaBufferSizeBytes = mode_lib->ms.ip.dcc_meta_buffer_size_bytes;
7689 		CalculateVMRowAndSwath_params->UseMALLForStaticScreen = mode_lib->ms.cache_display_cfg.plane.UseMALLForStaticScreen;
7690 		CalculateVMRowAndSwath_params->UseMALLForPStateChange = mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange;
7691 		CalculateVMRowAndSwath_params->MALLAllocatedForDCN = mode_lib->ms.soc.mall_allocated_for_dcn_mbytes;
7692 		CalculateVMRowAndSwath_params->SwathWidthY = mode_lib->ms.SwathWidthYThisState;
7693 		CalculateVMRowAndSwath_params->SwathWidthC = mode_lib->ms.SwathWidthCThisState;
7694 		CalculateVMRowAndSwath_params->GPUVMEnable = mode_lib->ms.cache_display_cfg.plane.GPUVMEnable;
7695 		CalculateVMRowAndSwath_params->HostVMEnable = mode_lib->ms.cache_display_cfg.plane.HostVMEnable;
7696 		CalculateVMRowAndSwath_params->HostVMMaxNonCachedPageTableLevels = mode_lib->ms.cache_display_cfg.plane.HostVMMaxPageTableLevels;
7697 		CalculateVMRowAndSwath_params->GPUVMMaxPageTableLevels = mode_lib->ms.cache_display_cfg.plane.GPUVMMaxPageTableLevels;
7698 		CalculateVMRowAndSwath_params->GPUVMMinPageSizeKBytes = mode_lib->ms.cache_display_cfg.plane.GPUVMMinPageSizeKBytes;
7699 		CalculateVMRowAndSwath_params->HostVMMinPageSize = mode_lib->ms.soc.hostvm_min_page_size_kbytes * 1024;
7700 		CalculateVMRowAndSwath_params->PTEBufferModeOverrideEn = mode_lib->ms.cache_display_cfg.plane.PTEBufferModeOverrideEn;
7701 		CalculateVMRowAndSwath_params->PTEBufferModeOverrideVal = mode_lib->ms.cache_display_cfg.plane.PTEBufferMode;
7702 		CalculateVMRowAndSwath_params->PTEBufferSizeNotExceeded = mode_lib->ms.PTEBufferSizeNotExceededPerState;
7703 		CalculateVMRowAndSwath_params->DCCMetaBufferSizeNotExceeded = mode_lib->ms.DCCMetaBufferSizeNotExceededPerState;
7704 		CalculateVMRowAndSwath_params->dpte_row_width_luma_ub = s->dummy_integer_array[0];
7705 		CalculateVMRowAndSwath_params->dpte_row_width_chroma_ub = s->dummy_integer_array[1];
7706 		CalculateVMRowAndSwath_params->dpte_row_height_luma = mode_lib->ms.dpte_row_height;
7707 		CalculateVMRowAndSwath_params->dpte_row_height_chroma = mode_lib->ms.dpte_row_height_chroma;
7708 		CalculateVMRowAndSwath_params->dpte_row_height_linear_luma = s->dummy_integer_array[2]; // VBA_DELTA
7709 		CalculateVMRowAndSwath_params->dpte_row_height_linear_chroma = s->dummy_integer_array[3]; // VBA_DELTA
7710 		CalculateVMRowAndSwath_params->meta_req_width = s->dummy_integer_array[4];
7711 		CalculateVMRowAndSwath_params->meta_req_width_chroma = s->dummy_integer_array[5];
7712 		CalculateVMRowAndSwath_params->meta_req_height = s->dummy_integer_array[6];
7713 		CalculateVMRowAndSwath_params->meta_req_height_chroma = s->dummy_integer_array[7];
7714 		CalculateVMRowAndSwath_params->meta_row_width = s->dummy_integer_array[8];
7715 		CalculateVMRowAndSwath_params->meta_row_width_chroma = s->dummy_integer_array[9];
7716 		CalculateVMRowAndSwath_params->meta_row_height = mode_lib->ms.meta_row_height;
7717 		CalculateVMRowAndSwath_params->meta_row_height_chroma = mode_lib->ms.meta_row_height_chroma;
7718 		CalculateVMRowAndSwath_params->vm_group_bytes = s->dummy_integer_array[10];
7719 		CalculateVMRowAndSwath_params->dpte_group_bytes = mode_lib->ms.dpte_group_bytes;
7720 		CalculateVMRowAndSwath_params->PixelPTEReqWidthY = s->dummy_integer_array[11];
7721 		CalculateVMRowAndSwath_params->PixelPTEReqHeightY = s->dummy_integer_array[12];
7722 		CalculateVMRowAndSwath_params->PTERequestSizeY = s->dummy_integer_array[13];
7723 		CalculateVMRowAndSwath_params->PixelPTEReqWidthC = s->dummy_integer_array[14];
7724 		CalculateVMRowAndSwath_params->PixelPTEReqHeightC = s->dummy_integer_array[15];
7725 		CalculateVMRowAndSwath_params->PTERequestSizeC = s->dummy_integer_array[16];
7726 		CalculateVMRowAndSwath_params->dpde0_bytes_per_frame_ub_l = s->dummy_integer_array[17];
7727 		CalculateVMRowAndSwath_params->meta_pte_bytes_per_frame_ub_l = s->dummy_integer_array[18];
7728 		CalculateVMRowAndSwath_params->dpde0_bytes_per_frame_ub_c = s->dummy_integer_array[19];
7729 		CalculateVMRowAndSwath_params->meta_pte_bytes_per_frame_ub_c = s->dummy_integer_array[20];
7730 		CalculateVMRowAndSwath_params->PrefetchSourceLinesY = mode_lib->ms.PrefetchLinesYThisState;
7731 		CalculateVMRowAndSwath_params->PrefetchSourceLinesC = mode_lib->ms.PrefetchLinesCThisState;
7732 		CalculateVMRowAndSwath_params->VInitPreFillY = mode_lib->ms.PrefillY;
7733 		CalculateVMRowAndSwath_params->VInitPreFillC = mode_lib->ms.PrefillC;
7734 		CalculateVMRowAndSwath_params->MaxNumSwathY = mode_lib->ms.MaxNumSwY;
7735 		CalculateVMRowAndSwath_params->MaxNumSwathC = mode_lib->ms.MaxNumSwC;
7736 		CalculateVMRowAndSwath_params->meta_row_bw = mode_lib->ms.meta_row_bandwidth_this_state;
7737 		CalculateVMRowAndSwath_params->dpte_row_bw = mode_lib->ms.dpte_row_bandwidth_this_state;
7738 		CalculateVMRowAndSwath_params->PixelPTEBytesPerRow = mode_lib->ms.DPTEBytesPerRowThisState;
7739 		CalculateVMRowAndSwath_params->PDEAndMetaPTEBytesFrame = mode_lib->ms.PDEAndMetaPTEBytesPerFrameThisState;
7740 		CalculateVMRowAndSwath_params->MetaRowByte = mode_lib->ms.MetaRowBytesThisState;
7741 		CalculateVMRowAndSwath_params->use_one_row_for_frame = mode_lib->ms.use_one_row_for_frame_this_state;
7742 		CalculateVMRowAndSwath_params->use_one_row_for_frame_flip = mode_lib->ms.use_one_row_for_frame_flip_this_state;
7743 		CalculateVMRowAndSwath_params->UsesMALLForStaticScreen = s->dummy_boolean_array[0];
7744 		CalculateVMRowAndSwath_params->PTE_BUFFER_MODE = s->dummy_boolean_array[1];
7745 		CalculateVMRowAndSwath_params->BIGK_FRAGMENT_SIZE = s->dummy_integer_array[21];
7746 
7747 		CalculateVMRowAndSwath(&mode_lib->scratch,
7748 			CalculateVMRowAndSwath_params);
7749 
7750 		for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7751 			mode_lib->ms.PrefetchLinesY[j][k] = mode_lib->ms.PrefetchLinesYThisState[k];
7752 			mode_lib->ms.PrefetchLinesC[j][k] = mode_lib->ms.PrefetchLinesCThisState[k];
7753 			mode_lib->ms.meta_row_bandwidth[j][k] = mode_lib->ms.meta_row_bandwidth_this_state[k];
7754 			mode_lib->ms.dpte_row_bandwidth[j][k] = mode_lib->ms.dpte_row_bandwidth_this_state[k];
7755 			mode_lib->ms.DPTEBytesPerRow[j][k] = mode_lib->ms.DPTEBytesPerRowThisState[k];
7756 			mode_lib->ms.PDEAndMetaPTEBytesPerFrame[j][k] = mode_lib->ms.PDEAndMetaPTEBytesPerFrameThisState[k];
7757 			mode_lib->ms.MetaRowBytes[j][k] = mode_lib->ms.MetaRowBytesThisState[k];
7758 			mode_lib->ms.use_one_row_for_frame[j][k] = mode_lib->ms.use_one_row_for_frame_this_state[k];
7759 			mode_lib->ms.use_one_row_for_frame_flip[j][k] = mode_lib->ms.use_one_row_for_frame_flip_this_state[k];
7760 		}
7761 
7762 		mode_lib->ms.support.PTEBufferSizeNotExceeded[j] = true;
7763 
7764 		for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7765 			if (mode_lib->ms.PTEBufferSizeNotExceededPerState[k] == false)
7766 				mode_lib->ms.support.PTEBufferSizeNotExceeded[j] = false;
7767 #ifdef __DML_VBA_DEBUG__
7768 			dml_print("DML::%s: j=%u k=%u, PTEBufferSizeNotExceededPerState[%u] = %u\n",  __func__, j, k, k, mode_lib->ms.PTEBufferSizeNotExceededPerState[k]);
7769 #endif
7770 		}
7771 #ifdef __DML_VBA_DEBUG__
7772 		dml_print("DML::%s: PTEBufferSizeNotExceeded[%u] = %u\n",  __func__, j, mode_lib->ms.support.PTEBufferSizeNotExceeded[j]);
7773 #endif
7774 
7775 		mode_lib->ms.support.DCCMetaBufferSizeNotExceeded[j] = true;
7776 		for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7777 			if (mode_lib->ms.DCCMetaBufferSizeNotExceededPerState[k] == false)
7778 				mode_lib->ms.support.DCCMetaBufferSizeNotExceeded[j] = false;
7779 		}
7780 
7781 		mode_lib->ms.UrgLatency = CalculateUrgentLatency(mode_lib->ms.state.urgent_latency_pixel_data_only_us,
7782 													mode_lib->ms.state.urgent_latency_pixel_mixed_with_vm_data_us,
7783 													mode_lib->ms.state.urgent_latency_vm_data_only_us,
7784 													mode_lib->ms.soc.do_urgent_latency_adjustment,
7785 													mode_lib->ms.state.urgent_latency_adjustment_fabric_clock_component_us,
7786 													mode_lib->ms.state.urgent_latency_adjustment_fabric_clock_reference_mhz,
7787 													mode_lib->ms.state.fabricclk_mhz);
7788 
7789 		/* Getter functions work at mp interface so copy the urgent latency to mp*/
7790 		mode_lib->mp.UrgentLatency = mode_lib->ms.UrgLatency;
7791 
7792 		for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7793 			CalculateUrgentBurstFactor(
7794 				mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange[k],
7795 				mode_lib->ms.swath_width_luma_ub_this_state[k],
7796 				mode_lib->ms.swath_width_chroma_ub_this_state[k],
7797 				mode_lib->ms.SwathHeightYThisState[k],
7798 				mode_lib->ms.SwathHeightCThisState[k],
7799 				(dml_float_t) mode_lib->ms.cache_display_cfg.timing.HTotal[k] / mode_lib->ms.cache_display_cfg.timing.PixelClock[k],
7800 				mode_lib->ms.UrgLatency,
7801 				mode_lib->ms.ip.cursor_buffer_size,
7802 				mode_lib->ms.cache_display_cfg.plane.CursorWidth[k],
7803 				mode_lib->ms.cache_display_cfg.plane.CursorBPP[k],
7804 				mode_lib->ms.cache_display_cfg.plane.VRatio[k],
7805 				mode_lib->ms.cache_display_cfg.plane.VRatioChroma[k],
7806 				mode_lib->ms.BytePerPixelInDETY[k],
7807 				mode_lib->ms.BytePerPixelInDETC[k],
7808 				mode_lib->ms.DETBufferSizeYThisState[k],
7809 				mode_lib->ms.DETBufferSizeCThisState[k],
7810 				/* Output */
7811 				&mode_lib->ms.UrgentBurstFactorCursor[k],
7812 				&mode_lib->ms.UrgentBurstFactorLuma[k],
7813 				&mode_lib->ms.UrgentBurstFactorChroma[k],
7814 				&mode_lib->ms.NotUrgentLatencyHiding[k]);
7815 		}
7816 
7817 		CalculateDCFCLKDeepSleep(
7818 				mode_lib->ms.num_active_planes,
7819 				mode_lib->ms.BytePerPixelY,
7820 				mode_lib->ms.BytePerPixelC,
7821 				mode_lib->ms.cache_display_cfg.plane.VRatio,
7822 				mode_lib->ms.cache_display_cfg.plane.VRatioChroma,
7823 				mode_lib->ms.SwathWidthYThisState,
7824 				mode_lib->ms.SwathWidthCThisState,
7825 				mode_lib->ms.NoOfDPPThisState,
7826 				mode_lib->ms.cache_display_cfg.plane.HRatio,
7827 				mode_lib->ms.cache_display_cfg.plane.HRatioChroma,
7828 				mode_lib->ms.cache_display_cfg.timing.PixelClock,
7829 				mode_lib->ms.PSCL_FACTOR,
7830 				mode_lib->ms.PSCL_FACTOR_CHROMA,
7831 				mode_lib->ms.RequiredDPPCLKThisState,
7832 				mode_lib->ms.ReadBandwidthLuma,
7833 				mode_lib->ms.ReadBandwidthChroma,
7834 				mode_lib->ms.soc.return_bus_width_bytes,
7835 
7836 				/* Output */
7837 				&mode_lib->ms.ProjectedDCFCLKDeepSleep[j]);
7838 	}
7839 
7840 	//Calculate Return BW
7841 	for (j = 0; j < 2; ++j) {
7842 		for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
7843 			if (mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming[k] == k) {
7844 				if (mode_lib->ms.cache_display_cfg.writeback.WritebackEnable[k] == true) {
7845 					mode_lib->ms.WritebackDelayTime[k] = mode_lib->ms.state.writeback_latency_us + CalculateWriteBackDelay(
7846 									mode_lib->ms.cache_display_cfg.writeback.WritebackPixelFormat[k],
7847 									mode_lib->ms.cache_display_cfg.writeback.WritebackHRatio[k],
7848 									mode_lib->ms.cache_display_cfg.writeback.WritebackVRatio[k],
7849 									mode_lib->ms.cache_display_cfg.writeback.WritebackVTaps[k],
7850 									mode_lib->ms.cache_display_cfg.writeback.WritebackDestinationWidth[k],
7851 									mode_lib->ms.cache_display_cfg.writeback.WritebackDestinationHeight[k],
7852 									mode_lib->ms.cache_display_cfg.writeback.WritebackSourceHeight[k],
7853 									mode_lib->ms.cache_display_cfg.timing.HTotal[k]) / mode_lib->ms.RequiredDISPCLK[j];
7854 				} else {
7855 					mode_lib->ms.WritebackDelayTime[k] = 0.0;
7856 				}
7857 				for (m = 0; m <= mode_lib->ms.num_active_planes - 1; m++) {
7858 					if (mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming[m] == k && mode_lib->ms.cache_display_cfg.writeback.WritebackEnable[m] == true) {
7859 						mode_lib->ms.WritebackDelayTime[k] = dml_max(mode_lib->ms.WritebackDelayTime[k],
7860 											mode_lib->ms.state.writeback_latency_us + CalculateWriteBackDelay(
7861 											mode_lib->ms.cache_display_cfg.writeback.WritebackPixelFormat[m],
7862 											mode_lib->ms.cache_display_cfg.writeback.WritebackHRatio[m],
7863 											mode_lib->ms.cache_display_cfg.writeback.WritebackVRatio[m],
7864 											mode_lib->ms.cache_display_cfg.writeback.WritebackVTaps[m],
7865 											mode_lib->ms.cache_display_cfg.writeback.WritebackDestinationWidth[m],
7866 											mode_lib->ms.cache_display_cfg.writeback.WritebackDestinationHeight[m],
7867 											mode_lib->ms.cache_display_cfg.writeback.WritebackSourceHeight[m],
7868 											mode_lib->ms.cache_display_cfg.timing.HTotal[m]) / mode_lib->ms.RequiredDISPCLK[j]);
7869 					}
7870 				}
7871 			}
7872 		}
7873 		for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
7874 			for (m = 0; m <= mode_lib->ms.num_active_planes - 1; m++) {
7875 				if (mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming[k] == m) {
7876 					mode_lib->ms.WritebackDelayTime[k] = mode_lib->ms.WritebackDelayTime[m];
7877 				}
7878 			}
7879 		}
7880        s->MaxVStartupAllPlanes[j] = 0;  // max vstartup among all planes
7881 
7882 		for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
7883 			s->MaximumVStartup[j][k] = CalculateMaxVStartup(k,
7884 														mode_lib->ms.ip.ptoi_supported,
7885 														mode_lib->ms.ip.vblank_nom_default_us,
7886 														&mode_lib->ms.cache_display_cfg.timing,
7887 														mode_lib->ms.WritebackDelayTime[k]);
7888 
7889 			s->MaxVStartupAllPlanes[j] = (dml_uint_t)(dml_max(s->MaxVStartupAllPlanes[j], s->MaximumVStartup[j][k]));
7890 #ifdef __DML_VBA_DEBUG__
7891 			dml_print("DML::%s: k=%u, MaxVStartupAllPlanes[%u] = %u\n", __func__, k, j, s->MaxVStartupAllPlanes[j]);
7892 			dml_print("DML::%s: k=%u, MaximumVStartup[%u][%u] = %u\n", __func__, k, j, k, s->MaximumVStartup[j][k]);
7893 #endif
7894 		}
7895 	}
7896 
7897 	s->ReorderingBytes = (dml_uint_t)(mode_lib->ms.soc.num_chans * dml_max3(mode_lib->ms.soc.urgent_out_of_order_return_per_channel_pixel_only_bytes,
7898 																mode_lib->ms.soc.urgent_out_of_order_return_per_channel_pixel_and_vm_bytes,
7899 																mode_lib->ms.soc.urgent_out_of_order_return_per_channel_vm_only_bytes));
7900 
7901 	for (j = 0; j < 2; ++j) {
7902 		mode_lib->ms.DCFCLKState[j] = mode_lib->ms.state.dcfclk_mhz;
7903 	}
7904 
7905 	/* Immediate Flip and MALL parameters */
7906 	s->ImmediateFlipRequiredFinal = false;
7907 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7908 		s->ImmediateFlipRequiredFinal = s->ImmediateFlipRequiredFinal || (mode_lib->ms.policy.ImmediateFlipRequirement[k] == dml_immediate_flip_required);
7909 	}
7910 
7911 	mode_lib->ms.support.ImmediateFlipRequiredButTheRequirementForEachSurfaceIsNotSpecified = false;
7912 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7913 		mode_lib->ms.support.ImmediateFlipRequiredButTheRequirementForEachSurfaceIsNotSpecified = mode_lib->ms.support.ImmediateFlipRequiredButTheRequirementForEachSurfaceIsNotSpecified ||
7914 																							((mode_lib->ms.policy.ImmediateFlipRequirement[k] != dml_immediate_flip_required) &&
7915 																							(mode_lib->ms.policy.ImmediateFlipRequirement[k] != dml_immediate_flip_not_required));
7916 	}
7917 	mode_lib->ms.support.ImmediateFlipRequiredButTheRequirementForEachSurfaceIsNotSpecified = mode_lib->ms.support.ImmediateFlipRequiredButTheRequirementForEachSurfaceIsNotSpecified && s->ImmediateFlipRequiredFinal;
7918 
7919 	mode_lib->ms.support.ImmediateFlipOrHostVMAndPStateWithMALLFullFrameOrPhantomPipe = false;
7920 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7921 		mode_lib->ms.support.ImmediateFlipOrHostVMAndPStateWithMALLFullFrameOrPhantomPipe =
7922 										mode_lib->ms.support.ImmediateFlipOrHostVMAndPStateWithMALLFullFrameOrPhantomPipe || ((mode_lib->ms.cache_display_cfg.plane.HostVMEnable == true || mode_lib->ms.policy.ImmediateFlipRequirement[k] != dml_immediate_flip_not_required) &&
7923 										(mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange[k] == dml_use_mall_pstate_change_full_frame || mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange[k] == dml_use_mall_pstate_change_phantom_pipe));
7924 	}
7925 
7926 	mode_lib->ms.support.InvalidCombinationOfMALLUseForPStateAndStaticScreen = false;
7927 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7928 		mode_lib->ms.support.InvalidCombinationOfMALLUseForPStateAndStaticScreen = mode_lib->ms.support.InvalidCombinationOfMALLUseForPStateAndStaticScreen ||
7929 																((mode_lib->ms.cache_display_cfg.plane.UseMALLForStaticScreen[k] == dml_use_mall_static_screen_enable || mode_lib->ms.cache_display_cfg.plane.UseMALLForStaticScreen[k] == dml_use_mall_static_screen_optimize) && (mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange[k] == dml_use_mall_pstate_change_phantom_pipe)) ||
7930 																((mode_lib->ms.cache_display_cfg.plane.UseMALLForStaticScreen[k] == dml_use_mall_static_screen_disable || mode_lib->ms.cache_display_cfg.plane.UseMALLForStaticScreen[k] == dml_use_mall_static_screen_optimize) && (mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange[k] == dml_use_mall_pstate_change_full_frame));
7931 	}
7932 
7933 	s->FullFrameMALLPStateMethod = false;
7934 	s->SubViewportMALLPStateMethod = false;
7935 	s->PhantomPipeMALLPStateMethod = false;
7936 	s->SubViewportMALLRefreshGreaterThan120Hz = false;
7937 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
7938 		if (mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange[k] == dml_use_mall_pstate_change_full_frame)
7939 			s->FullFrameMALLPStateMethod = true;
7940 		if (mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange[k] == dml_use_mall_pstate_change_sub_viewport) {
7941 			s->SubViewportMALLPStateMethod = true;
7942 			if (mode_lib->ms.cache_display_cfg.timing.RefreshRate[k] > 120)
7943 				s->SubViewportMALLRefreshGreaterThan120Hz = true;
7944 		}
7945 		if (mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange[k] == dml_use_mall_pstate_change_phantom_pipe)
7946 			s->PhantomPipeMALLPStateMethod = true;
7947 	}
7948 	mode_lib->ms.support.InvalidCombinationOfMALLUseForPState = (s->SubViewportMALLPStateMethod != s->PhantomPipeMALLPStateMethod)
7949 	|| (s->SubViewportMALLPStateMethod && s->FullFrameMALLPStateMethod) || s->SubViewportMALLRefreshGreaterThan120Hz;
7950 
7951     if (mode_lib->ms.policy.UseMinimumRequiredDCFCLK == true) {
7952 		UseMinimumDCFCLK_params->UseMALLForPStateChange = mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange;
7953 		UseMinimumDCFCLK_params->DRRDisplay = mode_lib->ms.cache_display_cfg.timing.DRRDisplay;
7954 		UseMinimumDCFCLK_params->SynchronizeDRRDisplaysForUCLKPStateChangeFinal = mode_lib->ms.policy.SynchronizeDRRDisplaysForUCLKPStateChangeFinal;
7955 		UseMinimumDCFCLK_params->MaxInterDCNTileRepeaters = mode_lib->ms.ip.max_inter_dcn_tile_repeaters;
7956 		UseMinimumDCFCLK_params->MaxPrefetchMode = dml_prefetch_support_stutter;
7957 		UseMinimumDCFCLK_params->DRAMClockChangeLatencyFinal = mode_lib->ms.state.dram_clock_change_latency_us;
7958 		UseMinimumDCFCLK_params->FCLKChangeLatency = mode_lib->ms.state.fclk_change_latency_us;
7959 		UseMinimumDCFCLK_params->SREnterPlusExitTime = mode_lib->ms.state.sr_enter_plus_exit_time_us;
7960 		UseMinimumDCFCLK_params->ReturnBusWidth = mode_lib->ms.soc.return_bus_width_bytes;
7961 		UseMinimumDCFCLK_params->RoundTripPingLatencyCycles = mode_lib->ms.soc.round_trip_ping_latency_dcfclk_cycles;
7962 		UseMinimumDCFCLK_params->ReorderingBytes = s->ReorderingBytes;
7963 		UseMinimumDCFCLK_params->PixelChunkSizeInKByte = mode_lib->ms.ip.pixel_chunk_size_kbytes;
7964 		UseMinimumDCFCLK_params->MetaChunkSize = mode_lib->ms.ip.meta_chunk_size_kbytes;
7965 		UseMinimumDCFCLK_params->GPUVMEnable = mode_lib->ms.cache_display_cfg.plane.GPUVMEnable;
7966 		UseMinimumDCFCLK_params->GPUVMMaxPageTableLevels = mode_lib->ms.cache_display_cfg.plane.GPUVMMaxPageTableLevels;
7967 		UseMinimumDCFCLK_params->HostVMEnable = mode_lib->ms.cache_display_cfg.plane.HostVMEnable;
7968 		UseMinimumDCFCLK_params->NumberOfActiveSurfaces = mode_lib->ms.num_active_planes;
7969 		UseMinimumDCFCLK_params->HostVMMinPageSize = mode_lib->ms.soc.hostvm_min_page_size_kbytes * 1024;
7970 		UseMinimumDCFCLK_params->HostVMMaxNonCachedPageTableLevels = mode_lib->ms.cache_display_cfg.plane.HostVMMaxPageTableLevels;
7971 		UseMinimumDCFCLK_params->DynamicMetadataVMEnabled = mode_lib->ms.ip.dynamic_metadata_vm_enabled;
7972 		UseMinimumDCFCLK_params->ImmediateFlipRequirement = s->ImmediateFlipRequiredFinal;
7973 		UseMinimumDCFCLK_params->ProgressiveToInterlaceUnitInOPP = mode_lib->ms.ip.ptoi_supported;
7974 		UseMinimumDCFCLK_params->MaxAveragePercentOfIdealSDPPortBWDisplayCanUseInNormalSystemOperation = mode_lib->ms.soc.max_avg_sdp_bw_use_normal_percent;
7975 		UseMinimumDCFCLK_params->PercentOfIdealSDPPortBWReceivedAfterUrgLatency = mode_lib->ms.soc.pct_ideal_sdp_bw_after_urgent;
7976 		UseMinimumDCFCLK_params->VTotal = mode_lib->ms.cache_display_cfg.timing.VTotal;
7977 		UseMinimumDCFCLK_params->VActive = mode_lib->ms.cache_display_cfg.timing.VActive;
7978 		UseMinimumDCFCLK_params->DynamicMetadataTransmittedBytes = mode_lib->ms.cache_display_cfg.plane.DynamicMetadataTransmittedBytes;
7979 		UseMinimumDCFCLK_params->DynamicMetadataLinesBeforeActiveRequired = mode_lib->ms.cache_display_cfg.plane.DynamicMetadataLinesBeforeActiveRequired;
7980 		UseMinimumDCFCLK_params->Interlace = mode_lib->ms.cache_display_cfg.timing.Interlace;
7981 		UseMinimumDCFCLK_params->RequiredDPPCLKPerSurface = mode_lib->ms.RequiredDPPCLKPerSurface;
7982 		UseMinimumDCFCLK_params->RequiredDISPCLK = mode_lib->ms.RequiredDISPCLK;
7983 		UseMinimumDCFCLK_params->UrgLatency = mode_lib->ms.UrgLatency;
7984 		UseMinimumDCFCLK_params->NoOfDPP = mode_lib->ms.NoOfDPP;
7985 		UseMinimumDCFCLK_params->ProjectedDCFCLKDeepSleep = mode_lib->ms.ProjectedDCFCLKDeepSleep;
7986 		UseMinimumDCFCLK_params->MaximumVStartup = s->MaximumVStartup;
7987 		UseMinimumDCFCLK_params->TotalNumberOfActiveDPP = mode_lib->ms.TotalNumberOfActiveDPP;
7988 		UseMinimumDCFCLK_params->TotalNumberOfDCCActiveDPP = mode_lib->ms.TotalNumberOfDCCActiveDPP;
7989 		UseMinimumDCFCLK_params->dpte_group_bytes = mode_lib->ms.dpte_group_bytes;
7990 		UseMinimumDCFCLK_params->PrefetchLinesY = mode_lib->ms.PrefetchLinesY;
7991 		UseMinimumDCFCLK_params->PrefetchLinesC = mode_lib->ms.PrefetchLinesC;
7992 		UseMinimumDCFCLK_params->swath_width_luma_ub_all_states = mode_lib->ms.swath_width_luma_ub_all_states;
7993 		UseMinimumDCFCLK_params->swath_width_chroma_ub_all_states = mode_lib->ms.swath_width_chroma_ub_all_states;
7994 		UseMinimumDCFCLK_params->BytePerPixelY = mode_lib->ms.BytePerPixelY;
7995 		UseMinimumDCFCLK_params->BytePerPixelC = mode_lib->ms.BytePerPixelC;
7996 		UseMinimumDCFCLK_params->HTotal = mode_lib->ms.cache_display_cfg.timing.HTotal;
7997 		UseMinimumDCFCLK_params->PixelClock = mode_lib->ms.cache_display_cfg.timing.PixelClock;
7998 		UseMinimumDCFCLK_params->PDEAndMetaPTEBytesPerFrame = mode_lib->ms.PDEAndMetaPTEBytesPerFrame;
7999 		UseMinimumDCFCLK_params->DPTEBytesPerRow = mode_lib->ms.DPTEBytesPerRow;
8000 		UseMinimumDCFCLK_params->MetaRowBytes = mode_lib->ms.MetaRowBytes;
8001 		UseMinimumDCFCLK_params->DynamicMetadataEnable = mode_lib->ms.cache_display_cfg.plane.DynamicMetadataEnable;
8002 		UseMinimumDCFCLK_params->ReadBandwidthLuma = mode_lib->ms.ReadBandwidthLuma;
8003 		UseMinimumDCFCLK_params->ReadBandwidthChroma = mode_lib->ms.ReadBandwidthChroma;
8004 		UseMinimumDCFCLK_params->DCFCLKPerState = mode_lib->ms.state.dcfclk_mhz;
8005 		UseMinimumDCFCLK_params->DCFCLKState = mode_lib->ms.DCFCLKState;
8006 
8007 		UseMinimumDCFCLK(&mode_lib->scratch,
8008 		UseMinimumDCFCLK_params);
8009 
8010 	 } // UseMinimumRequiredDCFCLK == true
8011 
8012 	for (j = 0; j < 2; ++j) {
8013 		mode_lib->ms.ReturnBWPerState[j] = dml_get_return_bw_mbps(&mode_lib->ms.soc, mode_lib->ms.state.use_ideal_dram_bw_strobe,
8014 																mode_lib->ms.cache_display_cfg.plane.HostVMEnable, mode_lib->ms.DCFCLKState[j], mode_lib->ms.state.fabricclk_mhz,
8015 																mode_lib->ms.state.dram_speed_mts);
8016 		mode_lib->ms.ReturnDRAMBWPerState[j] = dml_get_return_dram_bw_mbps(&mode_lib->ms.soc, mode_lib->ms.state.use_ideal_dram_bw_strobe,
8017 																mode_lib->ms.cache_display_cfg.plane.HostVMEnable,
8018 																mode_lib->ms.state.dram_speed_mts);
8019 	}
8020 
8021 	//Re-ordering Buffer Support Check
8022 	for (j = 0; j < 2; ++j) {
8023 		if ((mode_lib->ms.ip.rob_buffer_size_kbytes - mode_lib->ms.ip.pixel_chunk_size_kbytes) * 1024 / mode_lib->ms.ReturnBWPerState[j] >
8024 			(mode_lib->ms.soc.round_trip_ping_latency_dcfclk_cycles + 32) / mode_lib->ms.DCFCLKState[j] + s->ReorderingBytes / mode_lib->ms.ReturnBWPerState[j]) {
8025 			mode_lib->ms.support.ROBSupport[j] = true;
8026 		} else {
8027 			mode_lib->ms.support.ROBSupport[j] = false;
8028 		}
8029 		dml_print("DML::%s: DEBUG ROBSupport[%u] = %u (%u)\n",  __func__, j, mode_lib->ms.support.ROBSupport[j], __LINE__);
8030 	}
8031 
8032 	//Vertical Active BW support check
8033 	s->MaxTotalVActiveRDBandwidth = 0;
8034 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
8035 		s->MaxTotalVActiveRDBandwidth = s->MaxTotalVActiveRDBandwidth + mode_lib->ms.ReadBandwidthLuma[k] + mode_lib->ms.ReadBandwidthChroma[k];
8036 	}
8037 
8038 	for (j = 0; j < 2; ++j) {
8039 		mode_lib->ms.support.MaxTotalVerticalActiveAvailableBandwidth[j] = dml_min3(mode_lib->ms.soc.return_bus_width_bytes * mode_lib->ms.DCFCLKState[j] * mode_lib->ms.soc.max_avg_sdp_bw_use_normal_percent / 100.0,
8040 																	mode_lib->ms.state.fabricclk_mhz * mode_lib->ms.soc.fabric_datapath_to_dcn_data_return_bytes * mode_lib->ms.soc.max_avg_fabric_bw_use_normal_percent / 100.0,
8041 																	mode_lib->ms.state.dram_speed_mts * mode_lib->ms.soc.num_chans * mode_lib->ms.soc.dram_channel_width_bytes *
8042 																	((mode_lib->ms.state.use_ideal_dram_bw_strobe && !mode_lib->ms.cache_display_cfg.plane.HostVMEnable) ?
8043 																	mode_lib->ms.soc.max_avg_dram_bw_use_normal_strobe_percent : mode_lib->ms.soc.max_avg_dram_bw_use_normal_percent) / 100.0);
8044 
8045 		if (s->MaxTotalVActiveRDBandwidth <= mode_lib->ms.support.MaxTotalVerticalActiveAvailableBandwidth[j]) {
8046 			mode_lib->ms.support.TotalVerticalActiveBandwidthSupport[j] = true;
8047 		} else {
8048 			mode_lib->ms.support.TotalVerticalActiveBandwidthSupport[j] = false;
8049 		}
8050 	}
8051 
8052 	/* Prefetch Check */
8053 	dml_prefetch_check(mode_lib);
8054 
8055 	// End of Prefetch Check
8056 	dml_print("DML::%s: Done prefetch calculation\n", __func__);
8057 
8058 	/*Cursor Support Check*/
8059 	mode_lib->ms.support.CursorSupport = true;
8060 	for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
8061 		if (mode_lib->ms.cache_display_cfg.plane.CursorWidth[k] > 0.0) {
8062 			if (mode_lib->ms.cache_display_cfg.plane.CursorBPP[k] == 64 && mode_lib->ms.ip.cursor_64bpp_support == false) {
8063 				mode_lib->ms.support.CursorSupport = false;
8064 			}
8065 		}
8066 	}
8067 
8068 	/*Valid Pitch Check*/
8069 	mode_lib->ms.support.PitchSupport = true;
8070 	for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
8071 		mode_lib->ms.support.AlignedYPitch[k] = dml_ceil(
8072 				dml_max(mode_lib->ms.cache_display_cfg.surface.PitchY[k], mode_lib->ms.cache_display_cfg.surface.SurfaceWidthY[k]),
8073 				mode_lib->ms.MacroTileWidthY[k]);
8074 		if (mode_lib->ms.cache_display_cfg.surface.DCCEnable[k] == true) {
8075 			mode_lib->ms.support.AlignedDCCMetaPitchY[k] = dml_ceil(dml_max(mode_lib->ms.cache_display_cfg.surface.DCCMetaPitchY[k], mode_lib->ms.cache_display_cfg.surface.SurfaceWidthY[k]), 64.0 * mode_lib->ms.Read256BlockWidthY[k]);
8076 		} else {
8077 			mode_lib->ms.support.AlignedDCCMetaPitchY[k] = mode_lib->ms.cache_display_cfg.surface.DCCMetaPitchY[k];
8078 		}
8079 		if (mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_444_64
8080 			&& mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_444_32
8081 			&& mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_444_16
8082 			&& mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_mono_16
8083 			&& mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_rgbe
8084 			&& mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_mono_8) {
8085 			mode_lib->ms.support.AlignedCPitch[k] = dml_ceil(dml_max(mode_lib->ms.cache_display_cfg.surface.PitchC[k], mode_lib->ms.cache_display_cfg.surface.SurfaceWidthC[k]), mode_lib->ms.MacroTileWidthC[k]);
8086 			if (mode_lib->ms.cache_display_cfg.surface.DCCEnable[k] == true) {
8087 				mode_lib->ms.support.AlignedDCCMetaPitchC[k] = dml_ceil(dml_max(mode_lib->ms.cache_display_cfg.surface.DCCMetaPitchC[k], mode_lib->ms.cache_display_cfg.surface.SurfaceWidthC[k]), 64.0 * mode_lib->ms.Read256BlockWidthC[k]);
8088 			} else {
8089 				mode_lib->ms.support.AlignedDCCMetaPitchC[k] = mode_lib->ms.cache_display_cfg.surface.DCCMetaPitchC[k];
8090 			}
8091 		} else {
8092 			mode_lib->ms.support.AlignedCPitch[k] = mode_lib->ms.cache_display_cfg.surface.PitchC[k];
8093 			mode_lib->ms.support.AlignedDCCMetaPitchC[k] = mode_lib->ms.cache_display_cfg.surface.DCCMetaPitchC[k];
8094 		}
8095 		if (mode_lib->ms.support.AlignedYPitch[k] > mode_lib->ms.cache_display_cfg.surface.PitchY[k] || mode_lib->ms.support.AlignedCPitch[k] > mode_lib->ms.cache_display_cfg.surface.PitchC[k] ||
8096 			mode_lib->ms.support.AlignedDCCMetaPitchY[k] > mode_lib->ms.cache_display_cfg.surface.DCCMetaPitchY[k] || mode_lib->ms.support.AlignedDCCMetaPitchC[k] > mode_lib->ms.cache_display_cfg.surface.DCCMetaPitchC[k]) {
8097 			mode_lib->ms.support.PitchSupport = false;
8098 		}
8099 	}
8100 
8101 	mode_lib->ms.support.ViewportExceedsSurface = false;
8102 	for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
8103 		if (mode_lib->ms.cache_display_cfg.plane.ViewportWidth[k] > mode_lib->ms.cache_display_cfg.surface.SurfaceWidthY[k] || mode_lib->ms.cache_display_cfg.plane.ViewportHeight[k] > mode_lib->ms.cache_display_cfg.surface.SurfaceHeightY[k]) {
8104 			mode_lib->ms.support.ViewportExceedsSurface = true;
8105 			if (mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_444_64 && mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_444_32 &&
8106 				mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_444_16 && mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_444_8 && mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k] != dml_rgbe) {
8107 				if (mode_lib->ms.cache_display_cfg.plane.ViewportWidthChroma[k] > mode_lib->ms.cache_display_cfg.surface.SurfaceWidthC[k] || mode_lib->ms.cache_display_cfg.plane.ViewportHeightChroma[k] > mode_lib->ms.cache_display_cfg.surface.SurfaceHeightC[k]) {
8108 					mode_lib->ms.support.ViewportExceedsSurface = true;
8109 				}
8110 			}
8111 		}
8112 	}
8113 
8114 	/*Mode Support, Voltage State and SOC Configuration*/
8115 	for (j = 0; j < 2; j++) { // j iterator is for the combine mode off or on
8116 		dml_print("DML::%s: checking support for j=%u\n", __func__, j);
8117 		dml_print("DML::%s: state_idx=%0d max_state_idx=%0d\n", __func__, mode_lib->ms.state_idx, mode_lib->ms.max_state_idx);
8118 
8119 		s->is_max_pwr_state = (mode_lib->ms.max_state_idx == mode_lib->ms.state_idx);
8120 		s->is_max_dram_pwr_state = (mode_lib->ms.max_state.dram_speed_mts == mode_lib->ms.state.dram_speed_mts);
8121 
8122 		s->dram_clock_change_support = (!mode_lib->ms.policy.DRAMClockChangeRequirementFinal ||
8123 											(s->is_max_dram_pwr_state && mode_lib->policy.AssumeModeSupportAtMaxPwrStateEvenDRAMClockChangeNotSupported) ||
8124 											mode_lib->ms.support.DRAMClockChangeSupport[j] != dml_dram_clock_change_unsupported);
8125 		s->f_clock_change_support = (!mode_lib->ms.policy.FCLKChangeRequirementFinal ||
8126 											(s->is_max_pwr_state && mode_lib->policy.AssumeModeSupportAtMaxPwrStateEvenFClockChangeNotSupported) ||
8127 											mode_lib->ms.support.FCLKChangeSupport[j] != dml_fclock_change_unsupported);
8128 
8129 		if (mode_lib->ms.support.ScaleRatioAndTapsSupport == true
8130 			&& mode_lib->ms.support.SourceFormatPixelAndScanSupport == true
8131 			&& mode_lib->ms.support.ViewportSizeSupport[j] == true
8132 			&& !mode_lib->ms.support.LinkRateDoesNotMatchDPVersion
8133 			&& !mode_lib->ms.support.LinkRateForMultistreamNotIndicated
8134 			&& !mode_lib->ms.support.BPPForMultistreamNotIndicated
8135 			&& !mode_lib->ms.support.MultistreamWithHDMIOreDP
8136 			&& !mode_lib->ms.support.ExceededMultistreamSlots
8137 			&& !mode_lib->ms.support.MSOOrODMSplitWithNonDPLink
8138 			&& !mode_lib->ms.support.NotEnoughLanesForMSO
8139 			&& mode_lib->ms.support.LinkCapacitySupport == true
8140 			&& !mode_lib->ms.support.P2IWith420
8141 			&& !mode_lib->ms.support.DSCOnlyIfNecessaryWithBPP
8142 			&& !mode_lib->ms.support.DSC422NativeNotSupported
8143 			&& !mode_lib->ms.support.MPCCombineMethodIncompatible
8144 			&& mode_lib->ms.support.ODMCombineTwoToOneSupportCheckOK == true
8145 			&& mode_lib->ms.support.ODMCombineFourToOneSupportCheckOK == true
8146 			&& mode_lib->ms.support.NotEnoughDSCUnits == false
8147 			&& !mode_lib->ms.support.NotEnoughDSCSlices
8148 			&& !mode_lib->ms.support.ImmediateFlipOrHostVMAndPStateWithMALLFullFrameOrPhantomPipe
8149 			&& !mode_lib->ms.support.InvalidCombinationOfMALLUseForPStateAndStaticScreen
8150 			&& mode_lib->ms.support.DSCCLKRequiredMoreThanSupported == false
8151 			&& mode_lib->ms.support.PixelsPerLinePerDSCUnitSupport
8152 			&& mode_lib->ms.support.DTBCLKRequiredMoreThanSupported == false
8153 			&& !mode_lib->ms.support.InvalidCombinationOfMALLUseForPState
8154 			&& !mode_lib->ms.support.ImmediateFlipRequiredButTheRequirementForEachSurfaceIsNotSpecified
8155 			&& mode_lib->ms.support.ROBSupport[j] == true
8156 			&& mode_lib->ms.support.DISPCLK_DPPCLK_Support[j] == true
8157 			&& mode_lib->ms.support.TotalAvailablePipesSupport[j] == true
8158 			&& mode_lib->ms.support.NumberOfOTGSupport == true
8159 			&& mode_lib->ms.support.NumberOfHDMIFRLSupport == true
8160 			&& mode_lib->ms.support.NumberOfDP2p0Support == true
8161 			&& mode_lib->ms.support.EnoughWritebackUnits == true
8162 			&& mode_lib->ms.support.WritebackLatencySupport == true
8163 			&& mode_lib->ms.support.WritebackScaleRatioAndTapsSupport == true
8164 			&& mode_lib->ms.support.CursorSupport == true
8165 			&& mode_lib->ms.support.PitchSupport == true
8166 			&& mode_lib->ms.support.ViewportExceedsSurface == false
8167 			&& mode_lib->ms.support.PrefetchSupported[j] == true
8168 			&& mode_lib->ms.support.VActiveBandwithSupport[j] == true
8169 			&& mode_lib->ms.support.DynamicMetadataSupported[j] == true
8170 			&& mode_lib->ms.support.TotalVerticalActiveBandwidthSupport[j] == true
8171 			&& mode_lib->ms.support.VRatioInPrefetchSupported[j] == true
8172 			&& mode_lib->ms.support.PTEBufferSizeNotExceeded[j] == true
8173 			&& mode_lib->ms.support.DCCMetaBufferSizeNotExceeded[j] == true
8174 			&& mode_lib->ms.support.NonsupportedDSCInputBPC == false
8175 			&& !mode_lib->ms.support.ExceededMALLSize
8176 			&& ((mode_lib->ms.cache_display_cfg.plane.HostVMEnable == false && !s->ImmediateFlipRequiredFinal) || mode_lib->ms.support.ImmediateFlipSupportedForState[j])
8177 			&& s->dram_clock_change_support == true
8178 			&& s->f_clock_change_support == true
8179 			&& (!mode_lib->ms.policy.USRRetrainingRequiredFinal || mode_lib->ms.support.USRRetrainingSupport[j])) {
8180 			dml_print("DML::%s: mode is supported\n", __func__);
8181 			mode_lib->ms.support.ModeSupport[j] = true;
8182 		} else {
8183 			dml_print("DML::%s: mode is NOT supported\n", __func__);
8184 			mode_lib->ms.support.ModeSupport[j] = false;
8185 			dml_print_mode_support(mode_lib, j);
8186 		}
8187 	}
8188 
8189 	mode_lib->ms.support.MaximumMPCCombine = 0;
8190 	mode_lib->ms.support.ModeIsSupported = 0;
8191 	if (mode_lib->ms.support.ModeSupport[0] == true || mode_lib->ms.support.ModeSupport[1] == true) {  // if the mode is supported by either no combine or mpccombine
8192 		mode_lib->ms.support.ModeIsSupported = mode_lib->ms.support.ModeSupport[0] == true || mode_lib->ms.support.ModeSupport[1] == true;
8193 
8194 		// Determine if MPC combine is necessary, depends on if using MPC combine will help dram clock change or fclk change, etc.
8195 		if ((mode_lib->ms.support.ModeSupport[0] == false && mode_lib->ms.support.ModeSupport[1] == true) || s->MPCCombineMethodAsPossible ||
8196 			(s->MPCCombineMethodAsNeededForPStateChangeAndVoltage && mode_lib->ms.policy.DRAMClockChangeRequirementFinal &&
8197 			(((mode_lib->ms.support.DRAMClockChangeSupport[1] == dml_dram_clock_change_vactive || mode_lib->ms.support.DRAMClockChangeSupport[1] == dml_dram_clock_change_vactive_w_mall_full_frame || mode_lib->ms.support.DRAMClockChangeSupport[1] == dml_dram_clock_change_vactive_w_mall_sub_vp) &&
8198 			!(mode_lib->ms.support.DRAMClockChangeSupport[0] == dml_dram_clock_change_vactive || mode_lib->ms.support.DRAMClockChangeSupport[0] == dml_dram_clock_change_vactive_w_mall_full_frame || mode_lib->ms.support.DRAMClockChangeSupport[0] == dml_dram_clock_change_vactive_w_mall_sub_vp)) ||
8199 			((mode_lib->ms.support.DRAMClockChangeSupport[1] == dml_dram_clock_change_vblank || mode_lib->ms.support.DRAMClockChangeSupport[1] == dml_dram_clock_change_vblank_drr
8200 		|| mode_lib->ms.support.DRAMClockChangeSupport[1] == dml_dram_clock_change_vblank_w_mall_full_frame || mode_lib->ms.support.DRAMClockChangeSupport[1] == dml_dram_clock_change_vblank_drr_w_mall_full_frame
8201 		|| mode_lib->ms.support.DRAMClockChangeSupport[1] == dml_dram_clock_change_vblank_w_mall_sub_vp || mode_lib->ms.support.DRAMClockChangeSupport[1] == dml_dram_clock_change_vblank_drr_w_mall_sub_vp
8202 		) &&
8203 				mode_lib->ms.support.DRAMClockChangeSupport[0] == dml_dram_clock_change_unsupported)))
8204 			|| (s->MPCCombineMethodAsNeededForPStateChangeAndVoltage && mode_lib->ms.policy.FCLKChangeRequirementFinal &&
8205 				((mode_lib->ms.support.FCLKChangeSupport[1] == dml_fclock_change_vactive && mode_lib->ms.support.FCLKChangeSupport[0] != dml_fclock_change_vactive) ||
8206 				(mode_lib->ms.support.FCLKChangeSupport[1] == dml_fclock_change_vblank && mode_lib->ms.support.FCLKChangeSupport[0] == dml_fclock_change_unsupported)))) {
8207 			mode_lib->ms.support.MaximumMPCCombine = 1;
8208 		} else {
8209 			mode_lib->ms.support.MaximumMPCCombine = 0;
8210 		}
8211 	}
8212 
8213 	// Since now the mode_support work on 1 particular power state, so there is only 1 state idx (index 0).
8214 	mode_lib->ms.support.ImmediateFlipSupport          = mode_lib->ms.support.ImmediateFlipSupportedForState[mode_lib->ms.support.MaximumMPCCombine];   // Consider flip support if max combine support imm flip
8215 	mode_lib->ms.support.UnboundedRequestEnabled       = mode_lib->ms.UnboundedRequestEnabledAllStates[mode_lib->ms.support.MaximumMPCCombine];         // Not used, informational
8216 	mode_lib->ms.support.CompressedBufferSizeInkByte   = mode_lib->ms.CompressedBufferSizeInkByteAllStates[mode_lib->ms.support.MaximumMPCCombine];     // Not used, informational
8217 
8218 	dml_print("DML::%s: ModeIsSupported                = %u\n", __func__, mode_lib->ms.support.ModeIsSupported);
8219 	dml_print("DML::%s: MaximumMPCCombine              = %u\n", __func__, mode_lib->ms.support.MaximumMPCCombine);
8220 	dml_print("DML::%s: ImmediateFlipSupport           = %u\n", __func__, mode_lib->ms.support.ImmediateFlipSupport);
8221 	dml_print("DML::%s: UnboundedRequestEnabled        = %u\n", __func__, mode_lib->ms.support.UnboundedRequestEnabled);
8222 	dml_print("DML::%s: CompressedBufferSizeInkByte    = %u\n", __func__, mode_lib->ms.support.CompressedBufferSizeInkByte);
8223 
8224 	for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
8225 		mode_lib->ms.support.MPCCombineEnable[k]   = mode_lib->ms.MPCCombine[mode_lib->ms.support.MaximumMPCCombine][k];
8226 		mode_lib->ms.support.DPPPerSurface[k]      = mode_lib->ms.NoOfDPP[mode_lib->ms.support.MaximumMPCCombine][k];
8227 		mode_lib->ms.SwathHeightY[k]               = mode_lib->ms.SwathHeightYAllStates[mode_lib->ms.support.MaximumMPCCombine][k];
8228 		mode_lib->ms.SwathHeightC[k]               = mode_lib->ms.SwathHeightCAllStates[mode_lib->ms.support.MaximumMPCCombine][k];
8229 		mode_lib->ms.DETBufferSizeInKByte[k]       = mode_lib->ms.DETBufferSizeInKByteAllStates[mode_lib->ms.support.MaximumMPCCombine][k];
8230 		mode_lib->ms.DETBufferSizeY[k]             = mode_lib->ms.DETBufferSizeYAllStates[mode_lib->ms.support.MaximumMPCCombine][k];
8231 		mode_lib->ms.DETBufferSizeC[k]             = mode_lib->ms.DETBufferSizeCAllStates[mode_lib->ms.support.MaximumMPCCombine][k];
8232 	}
8233 
8234 	mode_lib->ms.DRAMSpeed     = mode_lib->ms.state.dram_speed_mts;
8235 	mode_lib->ms.FabricClock   = mode_lib->ms.state.fabricclk_mhz;
8236 	mode_lib->ms.SOCCLK        = mode_lib->ms.state.socclk_mhz;
8237 	mode_lib->ms.DCFCLK        = mode_lib->ms.DCFCLKState[mode_lib->ms.support.MaximumMPCCombine];
8238 	mode_lib->ms.ReturnBW      = mode_lib->ms.ReturnBWPerState[mode_lib->ms.support.MaximumMPCCombine];
8239 	mode_lib->ms.ReturnDRAMBW  = mode_lib->ms.ReturnDRAMBWPerState[mode_lib->ms.support.MaximumMPCCombine];
8240 
8241 	for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
8242 		if (mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming[k] == k) {
8243 			mode_lib->ms.support.ODMMode[k] = mode_lib->ms.ODMModePerState[k];
8244 		} else {
8245 			mode_lib->ms.support.ODMMode[k] = dml_odm_mode_bypass;
8246 		}
8247 
8248 		mode_lib->ms.support.DSCEnabled[k] = mode_lib->ms.RequiresDSC[k];
8249 		mode_lib->ms.support.FECEnabled[k] = mode_lib->ms.RequiresFEC[k];
8250 		mode_lib->ms.support.OutputBpp[k] = mode_lib->ms.OutputBppPerState[k];
8251 		mode_lib->ms.support.OutputType[k] = mode_lib->ms.OutputTypePerState[k];
8252 		mode_lib->ms.support.OutputRate[k] = mode_lib->ms.OutputRatePerState[k];
8253 		mode_lib->ms.support.SubViewportLinesNeededInMALL[k] = mode_lib->ms.SubViewportLinesNeededInMALL[k];
8254 	}
8255 
8256 	return mode_lib->ms.support.ModeIsSupported;
8257 } // dml_core_mode_support
8258 
8259 /// @brief This function calculates some parameters thats are needed ahead of the mode programming function all
dml_core_mode_support_partial(struct display_mode_lib_st * mode_lib)8260 void dml_core_mode_support_partial(struct display_mode_lib_st *mode_lib)
8261 {
8262 	CalculateMaxDETAndMinCompressedBufferSize(
8263 								mode_lib->ms.ip.config_return_buffer_size_in_kbytes,
8264 								mode_lib->ms.ip.config_return_buffer_segment_size_in_kbytes,
8265 								mode_lib->ms.ip.rob_buffer_size_kbytes,
8266 								mode_lib->ms.ip.max_num_dpp,
8267 								mode_lib->ms.policy.NomDETInKByteOverrideEnable,
8268 								mode_lib->ms.policy.NomDETInKByteOverrideValue,
8269 
8270 								/* Output */
8271 								&mode_lib->ms.MaxTotalDETInKByte,
8272 								&mode_lib->ms.NomDETInKByte,
8273 								&mode_lib->ms.MinCompressedBufferSizeInKByte);
8274 
8275 	PixelClockAdjustmentForProgressiveToInterlaceUnit(&mode_lib->ms.cache_display_cfg, mode_lib->ms.ip.ptoi_supported);
8276 
8277 	mode_lib->ms.ReturnBW = dml_get_return_bw_mbps(&mode_lib->ms.soc,
8278 													mode_lib->ms.state.use_ideal_dram_bw_strobe,
8279 													mode_lib->ms.cache_display_cfg.plane.HostVMEnable,
8280 													mode_lib->ms.DCFCLK,
8281 													mode_lib->ms.FabricClock,
8282 													mode_lib->ms.DRAMSpeed);
8283 	dml_print("DML::%s: ReturnBW = %f\n", __func__, mode_lib->ms.ReturnBW);
8284 
8285 } // dml_core_mode_support_partial
8286 
8287 /// @brief This is the mode programming function. It is assumed the display cfg is support at the given power state
dml_core_mode_programming(struct display_mode_lib_st * mode_lib,const struct dml_clk_cfg_st * clk_cfg)8288 void dml_core_mode_programming(struct display_mode_lib_st *mode_lib, const struct dml_clk_cfg_st *clk_cfg)
8289 {
8290 	struct dml_core_mode_programming_locals_st *s = &mode_lib->scratch.dml_core_mode_programming_locals;
8291 	struct CalculateWatermarksMALLUseAndDRAMSpeedChangeSupport_params_st *CalculateWatermarks_params = &mode_lib->scratch.CalculateWatermarksMALLUseAndDRAMSpeedChangeSupport_params;
8292 	struct CalculateVMRowAndSwath_params_st *CalculateVMRowAndSwath_params = &mode_lib->scratch.CalculateVMRowAndSwath_params;
8293 	struct CalculateSwathAndDETConfiguration_params_st *CalculateSwathAndDETConfiguration_params = &mode_lib->scratch.CalculateSwathAndDETConfiguration_params;
8294 	struct CalculateStutterEfficiency_params_st *CalculateStutterEfficiency_params = &mode_lib->scratch.CalculateStutterEfficiency_params;
8295 	struct CalculatePrefetchSchedule_params_st *CalculatePrefetchSchedule_params = &mode_lib->scratch.CalculatePrefetchSchedule_params;
8296 
8297 	struct mode_program_st   *locals    = &mode_lib->mp;
8298 	struct DmlPipe *myPipe;
8299 	dml_uint_t j = 0, k = 0;
8300 	dml_float_t TWait;
8301 	dml_bool_t isInterlaceTiming;
8302 
8303 	mode_lib->ms.num_active_planes = dml_get_num_active_planes(&mode_lib->ms.cache_display_cfg);
8304 	mode_lib->mp.num_active_pipes = dml_get_num_active_pipes(&mode_lib->ms.cache_display_cfg);
8305 	dml_calc_pipe_plane_mapping(&mode_lib->ms.cache_display_cfg.hw, mode_lib->mp.pipe_plane);
8306 
8307 #ifdef __DML_VBA_DEBUG__
8308 	dml_print("DML::%s: --- START --- \n",  __func__);
8309 	dml_print("DML::%s: num_active_planes = %u\n", __func__, mode_lib->ms.num_active_planes);
8310 	dml_print("DML::%s: num_active_pipes = %u\n", __func__, mode_lib->mp.num_active_pipes);
8311 #endif
8312 
8313 	s->DSCFormatFactor          = 0;
8314 
8315 	// Unlike dppclk and dispclk which can be calculated in mode_programming
8316 	// DCFCLK is calculated in mode_support (which is the state bbox dcfclk or min dcfclk if min dcfclk option is used in mode support calculation)
8317 	if (clk_cfg->dcfclk_option != dml_use_override_freq)
8318 		locals->Dcfclk = mode_lib->ms.DCFCLK;
8319 	else
8320 		locals->Dcfclk = clk_cfg->dcfclk_freq_mhz;
8321 
8322 #ifdef __DML_VBA_DEBUG__
8323 	dml_print_dml_policy(&mode_lib->ms.policy);
8324 	dml_print_soc_state_bounding_box(&mode_lib->ms.state);
8325 	dml_print_soc_bounding_box(&mode_lib->ms.soc);
8326 	dml_print_clk_cfg(clk_cfg);
8327 
8328 	dml_print("DML::%s: ImmediateFlipSupport = %u\n", __func__, mode_lib->ms.support.ImmediateFlipSupport);
8329 	dml_print("DML::%s: Using DCFCLK = %f\n", __func__, locals->Dcfclk);
8330 	dml_print("DML::%s: Using SOCCLK = %f\n", __func__, mode_lib->ms.SOCCLK);
8331 #endif
8332 
8333 	locals->WritebackDISPCLK = 0.0;
8334 	locals->GlobalDPPCLK = 0.0;
8335 
8336 	// DISPCLK and DPPCLK Calculation
8337 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
8338 		if (mode_lib->ms.cache_display_cfg.writeback.WritebackEnable[k]) {
8339 			locals->WritebackDISPCLK =
8340 					dml_max(
8341 							locals->WritebackDISPCLK,
8342 							CalculateWriteBackDISPCLK(
8343 									mode_lib->ms.cache_display_cfg.writeback.WritebackPixelFormat[k],
8344 									mode_lib->ms.cache_display_cfg.timing.PixelClock[k],
8345 									mode_lib->ms.cache_display_cfg.writeback.WritebackHRatio[k],
8346 									mode_lib->ms.cache_display_cfg.writeback.WritebackVRatio[k],
8347 									mode_lib->ms.cache_display_cfg.writeback.WritebackHTaps[k],
8348 									mode_lib->ms.cache_display_cfg.writeback.WritebackVTaps[k],
8349 									mode_lib->ms.cache_display_cfg.writeback.WritebackSourceWidth[k],
8350 									mode_lib->ms.cache_display_cfg.writeback.WritebackDestinationWidth[k],
8351 									mode_lib->ms.cache_display_cfg.timing.HTotal[k],
8352 									mode_lib->ms.ip.writeback_line_buffer_buffer_size,
8353 									mode_lib->ms.soc.dispclk_dppclk_vco_speed_mhz));
8354 		}
8355 	}
8356 
8357 	locals->Dispclk_calculated = locals->WritebackDISPCLK;
8358 
8359 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
8360 		if (mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming[k] == k) {
8361 				locals->Dispclk_calculated = dml_max(locals->Dispclk_calculated, CalculateRequiredDispclk(
8362 																					mode_lib->ms.cache_display_cfg.hw.ODMMode[k],
8363 																					mode_lib->ms.cache_display_cfg.timing.PixelClock[k],
8364 																					mode_lib->ms.soc.dcn_downspread_percent,
8365 																					mode_lib->ms.ip.dispclk_ramp_margin_percent,
8366 																					mode_lib->ms.soc.dispclk_dppclk_vco_speed_mhz,
8367 																					mode_lib->ms.max_state.dispclk_mhz));
8368 		}
8369 	}
8370 	if (clk_cfg->dispclk_option == dml_use_required_freq)
8371 		locals->Dispclk = locals->Dispclk_calculated;
8372 	else if (clk_cfg->dispclk_option == dml_use_override_freq)
8373 		locals->Dispclk = clk_cfg->dispclk_freq_mhz;
8374 	else
8375 		locals->Dispclk = mode_lib->ms.state.dispclk_mhz;
8376 #ifdef __DML_VBA_DEBUG__
8377 	 dml_print("DML::%s: Using Dispclk = %f\n", __func__, locals->Dispclk);
8378 #endif
8379 
8380 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
8381 		CalculateSinglePipeDPPCLKAndSCLThroughput(
8382 				mode_lib->ms.cache_display_cfg.plane.HRatio[k],
8383 				mode_lib->ms.cache_display_cfg.plane.HRatioChroma[k],
8384 				mode_lib->ms.cache_display_cfg.plane.VRatio[k],
8385 				mode_lib->ms.cache_display_cfg.plane.VRatioChroma[k],
8386 				mode_lib->ms.ip.max_dchub_pscl_bw_pix_per_clk,
8387 				mode_lib->ms.ip.max_pscl_lb_bw_pix_per_clk,
8388 				mode_lib->ms.cache_display_cfg.timing.PixelClock[k],
8389 				mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k],
8390 				mode_lib->ms.cache_display_cfg.plane.HTaps[k],
8391 				mode_lib->ms.cache_display_cfg.plane.HTapsChroma[k],
8392 				mode_lib->ms.cache_display_cfg.plane.VTaps[k],
8393 				mode_lib->ms.cache_display_cfg.plane.VTapsChroma[k],
8394 
8395 				/* Output */
8396 				&locals->PSCL_THROUGHPUT[k],
8397 				&locals->PSCL_THROUGHPUT_CHROMA[k],
8398 				&locals->DPPCLKUsingSingleDPP[k]);
8399 	}
8400 
8401 	CalculateDPPCLK(mode_lib->ms.num_active_planes,
8402 					mode_lib->ms.soc.dcn_downspread_percent,
8403 					mode_lib->ms.soc.dispclk_dppclk_vco_speed_mhz,
8404 					locals->DPPCLKUsingSingleDPP,
8405 					mode_lib->ms.cache_display_cfg.hw.DPPPerSurface,
8406 					/* Output */
8407 					&locals->GlobalDPPCLK,
8408 					locals->Dppclk_calculated);
8409 
8410 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
8411 		if (clk_cfg->dppclk_option[k] == dml_use_required_freq)
8412 			locals->Dppclk[k] = locals->Dppclk_calculated[k];
8413 		else if (clk_cfg->dppclk_option[k] == dml_use_override_freq)
8414 			locals->Dppclk[k] = clk_cfg->dppclk_freq_mhz[k];
8415 		else
8416 			locals->Dppclk[k] = mode_lib->ms.state.dppclk_mhz;
8417 #ifdef __DML_VBA_DEBUG__
8418 		dml_print("DML::%s: Using Dppclk[%0d] = %f\n", __func__, k, locals->Dppclk[k]);
8419 #endif
8420 	}
8421 
8422 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
8423 		CalculateBytePerPixelAndBlockSizes(
8424 				mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k],
8425 				mode_lib->ms.cache_display_cfg.surface.SurfaceTiling[k],
8426 
8427 				/* Output */
8428 				&locals->BytePerPixelY[k],
8429 				&locals->BytePerPixelC[k],
8430 				&locals->BytePerPixelDETY[k],
8431 				&locals->BytePerPixelDETC[k],
8432 				&locals->BlockHeight256BytesY[k],
8433 				&locals->BlockHeight256BytesC[k],
8434 				&locals->BlockWidth256BytesY[k],
8435 				&locals->BlockWidth256BytesC[k],
8436 				&locals->BlockHeightY[k],
8437 				&locals->BlockHeightC[k],
8438 				&locals->BlockWidthY[k],
8439 				&locals->BlockWidthC[k]);
8440 	}
8441 
8442 
8443 	dml_print("DML::%s: %u\n", __func__, __LINE__);
8444 	CalculateSwathWidth(
8445 		false,  // ForceSingleDPP
8446 		mode_lib->ms.num_active_planes,
8447 		mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat,
8448 		mode_lib->ms.cache_display_cfg.plane.SourceScan,
8449 		mode_lib->ms.cache_display_cfg.plane.ViewportStationary,
8450 		mode_lib->ms.cache_display_cfg.plane.ViewportWidth,
8451 		mode_lib->ms.cache_display_cfg.plane.ViewportHeight,
8452 		mode_lib->ms.cache_display_cfg.plane.ViewportXStart,
8453 		mode_lib->ms.cache_display_cfg.plane.ViewportYStart,
8454 		mode_lib->ms.cache_display_cfg.plane.ViewportXStartC,
8455 		mode_lib->ms.cache_display_cfg.plane.ViewportYStartC,
8456 		mode_lib->ms.cache_display_cfg.surface.SurfaceWidthY,
8457 		mode_lib->ms.cache_display_cfg.surface.SurfaceWidthC,
8458 		mode_lib->ms.cache_display_cfg.surface.SurfaceHeightY,
8459 		mode_lib->ms.cache_display_cfg.surface.SurfaceHeightC,
8460 		mode_lib->ms.cache_display_cfg.hw.ODMMode,
8461 		locals->BytePerPixelY,
8462 		locals->BytePerPixelC,
8463 		locals->BlockHeight256BytesY,
8464 		locals->BlockHeight256BytesC,
8465 		locals->BlockWidth256BytesY,
8466 		locals->BlockWidth256BytesC,
8467 		mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming,
8468 		mode_lib->ms.cache_display_cfg.timing.HActive,
8469 		mode_lib->ms.cache_display_cfg.plane.HRatio,
8470 		mode_lib->ms.cache_display_cfg.hw.DPPPerSurface,
8471 
8472 		/* Output */
8473 		locals->SwathWidthSingleDPPY,
8474 		locals->SwathWidthSingleDPPC,
8475 		locals->SwathWidthY,
8476 		locals->SwathWidthC,
8477 		s->dummy_integer_array[0], // dml_uint_t MaximumSwathHeightY[]
8478 		s->dummy_integer_array[1], // dml_uint_t MaximumSwathHeightC[]
8479 		locals->swath_width_luma_ub,
8480 		locals->swath_width_chroma_ub);
8481 
8482 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
8483 		locals->ReadBandwidthSurfaceLuma[k] = locals->SwathWidthSingleDPPY[k] * locals->BytePerPixelY[k] / (mode_lib->ms.cache_display_cfg.timing.HTotal[k] / mode_lib->ms.cache_display_cfg.timing.PixelClock[k]) * mode_lib->ms.cache_display_cfg.plane.VRatio[k];
8484 		locals->ReadBandwidthSurfaceChroma[k] = locals->SwathWidthSingleDPPC[k] * locals->BytePerPixelC[k] / (mode_lib->ms.cache_display_cfg.timing.HTotal[k] / mode_lib->ms.cache_display_cfg.timing.PixelClock[k]) * mode_lib->ms.cache_display_cfg.plane.VRatioChroma[k];
8485 		dml_print("DML::%s: ReadBandwidthSurfaceLuma[%i] = %fBps\n", __func__, k, locals->ReadBandwidthSurfaceLuma[k]);
8486 		dml_print("DML::%s: ReadBandwidthSurfaceChroma[%i] = %fBps\n", __func__, k, locals->ReadBandwidthSurfaceChroma[k]);
8487 	}
8488 
8489 	CalculateSwathAndDETConfiguration_params->DETSizeOverride = mode_lib->ms.cache_display_cfg.plane.DETSizeOverride;
8490 	CalculateSwathAndDETConfiguration_params->UseMALLForPStateChange = mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange;
8491 	CalculateSwathAndDETConfiguration_params->ConfigReturnBufferSizeInKByte = mode_lib->ms.ip.config_return_buffer_size_in_kbytes;
8492 	CalculateSwathAndDETConfiguration_params->ROBBufferSizeInKByte = mode_lib->ms.ip.rob_buffer_size_kbytes;
8493 	CalculateSwathAndDETConfiguration_params->MaxTotalDETInKByte = mode_lib->ms.MaxTotalDETInKByte;
8494 	CalculateSwathAndDETConfiguration_params->MinCompressedBufferSizeInKByte = mode_lib->ms.MinCompressedBufferSizeInKByte;
8495 	CalculateSwathAndDETConfiguration_params->PixelChunkSizeInKByte = mode_lib->ms.ip.pixel_chunk_size_kbytes;
8496 	CalculateSwathAndDETConfiguration_params->ForceSingleDPP = false;
8497 	CalculateSwathAndDETConfiguration_params->NumberOfActiveSurfaces = mode_lib->ms.num_active_planes;
8498 	CalculateSwathAndDETConfiguration_params->nomDETInKByte = mode_lib->ms.NomDETInKByte;
8499 	CalculateSwathAndDETConfiguration_params->UseUnboundedRequestingFinal = mode_lib->ms.policy.UseUnboundedRequesting;
8500 	CalculateSwathAndDETConfiguration_params->ConfigReturnBufferSegmentSizeInkByte = mode_lib->ms.ip.config_return_buffer_segment_size_in_kbytes;
8501 	CalculateSwathAndDETConfiguration_params->CompressedBufferSegmentSizeInkByteFinal = mode_lib->ms.ip.compressed_buffer_segment_size_in_kbytes;
8502 	CalculateSwathAndDETConfiguration_params->Output = s->dummy_output_encoder_array;
8503 	CalculateSwathAndDETConfiguration_params->ReadBandwidthLuma = locals->ReadBandwidthSurfaceLuma;
8504 	CalculateSwathAndDETConfiguration_params->ReadBandwidthChroma = locals->ReadBandwidthSurfaceChroma;
8505 	CalculateSwathAndDETConfiguration_params->MaximumSwathWidthLuma = s->dummy_single_array[0];
8506 	CalculateSwathAndDETConfiguration_params->MaximumSwathWidthChroma = s->dummy_single_array[1];
8507 	CalculateSwathAndDETConfiguration_params->SourceScan = mode_lib->ms.cache_display_cfg.plane.SourceScan;
8508 	CalculateSwathAndDETConfiguration_params->ViewportStationary = mode_lib->ms.cache_display_cfg.plane.ViewportStationary;
8509 	CalculateSwathAndDETConfiguration_params->SourcePixelFormat = mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat;
8510 	CalculateSwathAndDETConfiguration_params->SurfaceTiling = mode_lib->ms.cache_display_cfg.surface.SurfaceTiling;
8511 	CalculateSwathAndDETConfiguration_params->ViewportWidth = mode_lib->ms.cache_display_cfg.plane.ViewportWidth;
8512 	CalculateSwathAndDETConfiguration_params->ViewportHeight = mode_lib->ms.cache_display_cfg.plane.ViewportHeight;
8513 	CalculateSwathAndDETConfiguration_params->ViewportXStart = mode_lib->ms.cache_display_cfg.plane.ViewportXStart;
8514 	CalculateSwathAndDETConfiguration_params->ViewportYStart = mode_lib->ms.cache_display_cfg.plane.ViewportYStart;
8515 	CalculateSwathAndDETConfiguration_params->ViewportXStartC = mode_lib->ms.cache_display_cfg.plane.ViewportXStartC;
8516 	CalculateSwathAndDETConfiguration_params->ViewportYStartC = mode_lib->ms.cache_display_cfg.plane.ViewportYStartC;
8517 	CalculateSwathAndDETConfiguration_params->SurfaceWidthY = mode_lib->ms.cache_display_cfg.surface.SurfaceWidthY;
8518 	CalculateSwathAndDETConfiguration_params->SurfaceWidthC = mode_lib->ms.cache_display_cfg.surface.SurfaceWidthC;
8519 	CalculateSwathAndDETConfiguration_params->SurfaceHeightY = mode_lib->ms.cache_display_cfg.surface.SurfaceHeightY;
8520 	CalculateSwathAndDETConfiguration_params->SurfaceHeightC = mode_lib->ms.cache_display_cfg.surface.SurfaceHeightC;
8521 	CalculateSwathAndDETConfiguration_params->Read256BytesBlockHeightY = locals->BlockHeight256BytesY;
8522 	CalculateSwathAndDETConfiguration_params->Read256BytesBlockHeightC = locals->BlockHeight256BytesC;
8523 	CalculateSwathAndDETConfiguration_params->Read256BytesBlockWidthY = locals->BlockWidth256BytesY;
8524 	CalculateSwathAndDETConfiguration_params->Read256BytesBlockWidthC = locals->BlockWidth256BytesC;
8525 	CalculateSwathAndDETConfiguration_params->ODMMode = mode_lib->ms.cache_display_cfg.hw.ODMMode;
8526 	CalculateSwathAndDETConfiguration_params->BlendingAndTiming = mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming;
8527 	CalculateSwathAndDETConfiguration_params->BytePerPixY = locals->BytePerPixelY;
8528 	CalculateSwathAndDETConfiguration_params->BytePerPixC = locals->BytePerPixelC;
8529 	CalculateSwathAndDETConfiguration_params->BytePerPixDETY = locals->BytePerPixelDETY;
8530 	CalculateSwathAndDETConfiguration_params->BytePerPixDETC = locals->BytePerPixelDETC;
8531 	CalculateSwathAndDETConfiguration_params->HActive = mode_lib->ms.cache_display_cfg.timing.HActive;
8532 	CalculateSwathAndDETConfiguration_params->HRatio = mode_lib->ms.cache_display_cfg.plane.HRatio;
8533 	CalculateSwathAndDETConfiguration_params->HRatioChroma = mode_lib->ms.cache_display_cfg.plane.HRatioChroma;
8534 	CalculateSwathAndDETConfiguration_params->DPPPerSurface = mode_lib->ms.cache_display_cfg.hw.DPPPerSurface;
8535 	CalculateSwathAndDETConfiguration_params->swath_width_luma_ub = s->dummy_long_array[0];
8536 	CalculateSwathAndDETConfiguration_params->swath_width_chroma_ub = s->dummy_long_array[1];
8537 	CalculateSwathAndDETConfiguration_params->SwathWidth = s->dummy_long_array[2];
8538 	CalculateSwathAndDETConfiguration_params->SwathWidthChroma = s->dummy_long_array[3];
8539 	CalculateSwathAndDETConfiguration_params->SwathHeightY = locals->SwathHeightY;
8540 	CalculateSwathAndDETConfiguration_params->SwathHeightC = locals->SwathHeightC;
8541 	CalculateSwathAndDETConfiguration_params->DETBufferSizeInKByte = locals->DETBufferSizeInKByte;
8542 	CalculateSwathAndDETConfiguration_params->DETBufferSizeY = locals->DETBufferSizeY;
8543 	CalculateSwathAndDETConfiguration_params->DETBufferSizeC = locals->DETBufferSizeC;
8544 	CalculateSwathAndDETConfiguration_params->UnboundedRequestEnabled = &locals->UnboundedRequestEnabled;
8545 	CalculateSwathAndDETConfiguration_params->compbuf_reserved_space_64b = &locals->compbuf_reserved_space_64b;
8546 	CalculateSwathAndDETConfiguration_params->compbuf_reserved_space_zs = &locals->compbuf_reserved_space_zs;
8547 	CalculateSwathAndDETConfiguration_params->CompressedBufferSizeInkByte = &locals->CompressedBufferSizeInkByte;
8548 	CalculateSwathAndDETConfiguration_params->ViewportSizeSupportPerSurface = &s->dummy_boolean_array[0][0];
8549 	CalculateSwathAndDETConfiguration_params->ViewportSizeSupport = &s->dummy_boolean[0];
8550 
8551 	// VBA_DELTA
8552 	// Calculate DET size, swath height here. In VBA, they are calculated in mode check stage
8553 	CalculateSwathAndDETConfiguration(&mode_lib->scratch,
8554 		CalculateSwathAndDETConfiguration_params);
8555 
8556 	// DCFCLK Deep Sleep
8557 	CalculateDCFCLKDeepSleep(
8558 			mode_lib->ms.num_active_planes,
8559 			locals->BytePerPixelY,
8560 			locals->BytePerPixelC,
8561 			mode_lib->ms.cache_display_cfg.plane.VRatio,
8562 			mode_lib->ms.cache_display_cfg.plane.VRatioChroma,
8563 			locals->SwathWidthY,
8564 			locals->SwathWidthC,
8565 			mode_lib->ms.cache_display_cfg.hw.DPPPerSurface,
8566 			mode_lib->ms.cache_display_cfg.plane.HRatio,
8567 			mode_lib->ms.cache_display_cfg.plane.HRatioChroma,
8568 			mode_lib->ms.cache_display_cfg.timing.PixelClock,
8569 			locals->PSCL_THROUGHPUT,
8570 			locals->PSCL_THROUGHPUT_CHROMA,
8571 			locals->Dppclk,
8572 			locals->ReadBandwidthSurfaceLuma,
8573 			locals->ReadBandwidthSurfaceChroma,
8574 			mode_lib->ms.soc.return_bus_width_bytes,
8575 
8576 			/* Output */
8577 			&locals->DCFCLKDeepSleep);
8578 
8579 	// DSCCLK
8580 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
8581 		if ((mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming[k] != k) || !mode_lib->ms.cache_display_cfg.hw.DSCEnabled[k]) {
8582 			locals->DSCCLK_calculated[k] = 0.0;
8583 		} else {
8584 			if (mode_lib->ms.cache_display_cfg.output.OutputFormat[k] == dml_420)
8585 				s->DSCFormatFactor = 2;
8586 			else if (mode_lib->ms.cache_display_cfg.output.OutputFormat[k] == dml_444)
8587 				s->DSCFormatFactor = 1;
8588 			else if (mode_lib->ms.cache_display_cfg.output.OutputFormat[k] == dml_n422 || mode_lib->ms.cache_display_cfg.output.OutputEncoder[k] == dml_hdmifrl)
8589 				s->DSCFormatFactor = 2;
8590 			else
8591 				s->DSCFormatFactor = 1;
8592 			if (mode_lib->ms.cache_display_cfg.hw.ODMMode[k] == dml_odm_mode_combine_4to1)
8593 				locals->DSCCLK_calculated[k] = mode_lib->ms.cache_display_cfg.output.PixelClockBackEnd[k] / 12 / s->DSCFormatFactor / (1 - mode_lib->ms.soc.dcn_downspread_percent / 100);
8594 			else if (mode_lib->ms.cache_display_cfg.hw.ODMMode[k] == dml_odm_mode_combine_2to1)
8595 				locals->DSCCLK_calculated[k] = mode_lib->ms.cache_display_cfg.output.PixelClockBackEnd[k] / 6 / s->DSCFormatFactor / (1 - mode_lib->ms.soc.dcn_downspread_percent / 100);
8596 			else
8597 				locals->DSCCLK_calculated[k] = mode_lib->ms.cache_display_cfg.output.PixelClockBackEnd[k] / 3 / s->DSCFormatFactor / (1 - mode_lib->ms.soc.dcn_downspread_percent / 100);
8598 		}
8599 	}
8600 
8601 	// DSC Delay
8602 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
8603 		locals->DSCDelay[k] = DSCDelayRequirement(mode_lib->ms.cache_display_cfg.hw.DSCEnabled[k],
8604 												mode_lib->ms.cache_display_cfg.hw.ODMMode[k],
8605 												mode_lib->ms.cache_display_cfg.output.DSCInputBitPerComponent[k],
8606 												mode_lib->ms.cache_display_cfg.output.OutputBpp[k],
8607 												mode_lib->ms.cache_display_cfg.timing.HActive[k],
8608 												mode_lib->ms.cache_display_cfg.timing.HTotal[k],
8609 												mode_lib->ms.cache_display_cfg.hw.NumberOfDSCSlices[k],
8610 												mode_lib->ms.cache_display_cfg.output.OutputFormat[k],
8611 												mode_lib->ms.cache_display_cfg.output.OutputEncoder[k],
8612 												mode_lib->ms.cache_display_cfg.timing.PixelClock[k],
8613 												mode_lib->ms.cache_display_cfg.output.PixelClockBackEnd[k]);
8614 	}
8615 
8616 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k)
8617 		for (j = 0; j < mode_lib->ms.num_active_planes; ++j) // NumberOfSurfaces
8618 			if (j != k && mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming[k] == j && mode_lib->ms.cache_display_cfg.hw.DSCEnabled[j])
8619 				locals->DSCDelay[k] = locals->DSCDelay[j];
8620 
8621 	// Prefetch
8622 	CalculateSurfaceSizeInMall(
8623 		mode_lib->ms.num_active_planes,
8624 		mode_lib->ms.soc.mall_allocated_for_dcn_mbytes,
8625 		mode_lib->ms.cache_display_cfg.plane.UseMALLForStaticScreen,
8626 		mode_lib->ms.cache_display_cfg.surface.DCCEnable,
8627 		mode_lib->ms.cache_display_cfg.plane.ViewportStationary,
8628 		mode_lib->ms.cache_display_cfg.plane.ViewportXStart,
8629 		mode_lib->ms.cache_display_cfg.plane.ViewportYStart,
8630 		mode_lib->ms.cache_display_cfg.plane.ViewportXStartC,
8631 		mode_lib->ms.cache_display_cfg.plane.ViewportYStartC,
8632 		mode_lib->ms.cache_display_cfg.plane.ViewportWidth,
8633 		mode_lib->ms.cache_display_cfg.plane.ViewportHeight,
8634 		locals->BytePerPixelY,
8635 		mode_lib->ms.cache_display_cfg.plane.ViewportWidthChroma,
8636 		mode_lib->ms.cache_display_cfg.plane.ViewportHeightChroma,
8637 		locals->BytePerPixelC,
8638 		mode_lib->ms.cache_display_cfg.surface.SurfaceWidthY,
8639 		mode_lib->ms.cache_display_cfg.surface.SurfaceWidthC,
8640 		mode_lib->ms.cache_display_cfg.surface.SurfaceHeightY,
8641 		mode_lib->ms.cache_display_cfg.surface.SurfaceHeightC,
8642 		locals->BlockWidth256BytesY,
8643 		locals->BlockWidth256BytesC,
8644 		locals->BlockHeight256BytesY,
8645 		locals->BlockHeight256BytesC,
8646 		locals->BlockWidthY,
8647 		locals->BlockWidthC,
8648 		locals->BlockHeightY,
8649 		locals->BlockHeightC,
8650 
8651 		/* Output */
8652 		locals->SurfaceSizeInTheMALL,
8653 		&s->dummy_boolean[0]); /* dml_bool_t *ExceededMALLSize */
8654 
8655 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
8656 		s->SurfaceParameters[k].PixelClock = mode_lib->ms.cache_display_cfg.timing.PixelClock[k];
8657 		s->SurfaceParameters[k].DPPPerSurface = mode_lib->ms.cache_display_cfg.hw.DPPPerSurface[k];
8658 		s->SurfaceParameters[k].SourceScan = mode_lib->ms.cache_display_cfg.plane.SourceScan[k];
8659 		s->SurfaceParameters[k].ViewportHeight = mode_lib->ms.cache_display_cfg.plane.ViewportHeight[k];
8660 		s->SurfaceParameters[k].ViewportHeightChroma = mode_lib->ms.cache_display_cfg.plane.ViewportHeightChroma[k];
8661 		s->SurfaceParameters[k].BlockWidth256BytesY = locals->BlockWidth256BytesY[k];
8662 		s->SurfaceParameters[k].BlockHeight256BytesY = locals->BlockHeight256BytesY[k];
8663 		s->SurfaceParameters[k].BlockWidth256BytesC = locals->BlockWidth256BytesC[k];
8664 		s->SurfaceParameters[k].BlockHeight256BytesC = locals->BlockHeight256BytesC[k];
8665 		s->SurfaceParameters[k].BlockWidthY = locals->BlockWidthY[k];
8666 		s->SurfaceParameters[k].BlockHeightY = locals->BlockHeightY[k];
8667 		s->SurfaceParameters[k].BlockWidthC = locals->BlockWidthC[k];
8668 		s->SurfaceParameters[k].BlockHeightC = locals->BlockHeightC[k];
8669 		s->SurfaceParameters[k].InterlaceEnable = mode_lib->ms.cache_display_cfg.timing.Interlace[k];
8670 		s->SurfaceParameters[k].HTotal = mode_lib->ms.cache_display_cfg.timing.HTotal[k];
8671 		s->SurfaceParameters[k].DCCEnable = mode_lib->ms.cache_display_cfg.surface.DCCEnable[k];
8672 		s->SurfaceParameters[k].SourcePixelFormat = mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k];
8673 		s->SurfaceParameters[k].SurfaceTiling = mode_lib->ms.cache_display_cfg.surface.SurfaceTiling[k];
8674 		s->SurfaceParameters[k].BytePerPixelY = locals->BytePerPixelY[k];
8675 		s->SurfaceParameters[k].BytePerPixelC = locals->BytePerPixelC[k];
8676 		s->SurfaceParameters[k].ProgressiveToInterlaceUnitInOPP = mode_lib->ms.ip.ptoi_supported;
8677 		s->SurfaceParameters[k].VRatio = mode_lib->ms.cache_display_cfg.plane.VRatio[k];
8678 		s->SurfaceParameters[k].VRatioChroma = mode_lib->ms.cache_display_cfg.plane.VRatioChroma[k];
8679 		s->SurfaceParameters[k].VTaps = mode_lib->ms.cache_display_cfg.plane.VTaps[k];
8680 		s->SurfaceParameters[k].VTapsChroma = mode_lib->ms.cache_display_cfg.plane.VTapsChroma[k];
8681 		s->SurfaceParameters[k].PitchY = mode_lib->ms.cache_display_cfg.surface.PitchY[k];
8682 		s->SurfaceParameters[k].DCCMetaPitchY = mode_lib->ms.cache_display_cfg.surface.DCCMetaPitchY[k];
8683 		s->SurfaceParameters[k].PitchC = mode_lib->ms.cache_display_cfg.surface.PitchC[k];
8684 		s->SurfaceParameters[k].DCCMetaPitchC = mode_lib->ms.cache_display_cfg.surface.DCCMetaPitchC[k];
8685 		s->SurfaceParameters[k].ViewportStationary = mode_lib->ms.cache_display_cfg.plane.ViewportStationary[k];
8686 		s->SurfaceParameters[k].ViewportXStart = mode_lib->ms.cache_display_cfg.plane.ViewportXStart[k];
8687 		s->SurfaceParameters[k].ViewportYStart = mode_lib->ms.cache_display_cfg.plane.ViewportYStart[k];
8688 		s->SurfaceParameters[k].ViewportXStartC = mode_lib->ms.cache_display_cfg.plane.ViewportXStartC[k];
8689 		s->SurfaceParameters[k].ViewportYStartC = mode_lib->ms.cache_display_cfg.plane.ViewportYStartC[k];
8690 		s->SurfaceParameters[k].FORCE_ONE_ROW_FOR_FRAME = mode_lib->ms.cache_display_cfg.plane.ForceOneRowForFrame[k];
8691 		s->SurfaceParameters[k].SwathHeightY = locals->SwathHeightY[k];
8692 		s->SurfaceParameters[k].SwathHeightC = locals->SwathHeightC[k];
8693 	}
8694 
8695 	CalculateVMRowAndSwath_params->NumberOfActiveSurfaces = mode_lib->ms.num_active_planes;
8696 	CalculateVMRowAndSwath_params->myPipe = s->SurfaceParameters;
8697 	CalculateVMRowAndSwath_params->SurfaceSizeInMALL = locals->SurfaceSizeInTheMALL;
8698 	CalculateVMRowAndSwath_params->PTEBufferSizeInRequestsLuma = mode_lib->ms.ip.dpte_buffer_size_in_pte_reqs_luma;
8699 	CalculateVMRowAndSwath_params->PTEBufferSizeInRequestsChroma = mode_lib->ms.ip.dpte_buffer_size_in_pte_reqs_chroma;
8700 	CalculateVMRowAndSwath_params->DCCMetaBufferSizeBytes = mode_lib->ms.ip.dcc_meta_buffer_size_bytes;
8701 	CalculateVMRowAndSwath_params->UseMALLForStaticScreen = mode_lib->ms.cache_display_cfg.plane.UseMALLForStaticScreen;
8702 	CalculateVMRowAndSwath_params->UseMALLForPStateChange = mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange;
8703 	CalculateVMRowAndSwath_params->MALLAllocatedForDCN = mode_lib->ms.soc.mall_allocated_for_dcn_mbytes;
8704 	CalculateVMRowAndSwath_params->SwathWidthY = locals->SwathWidthY;
8705 	CalculateVMRowAndSwath_params->SwathWidthC = locals->SwathWidthC;
8706 	CalculateVMRowAndSwath_params->GPUVMEnable = mode_lib->ms.cache_display_cfg.plane.GPUVMEnable;
8707 	CalculateVMRowAndSwath_params->HostVMEnable = mode_lib->ms.cache_display_cfg.plane.HostVMEnable;
8708 	CalculateVMRowAndSwath_params->HostVMMaxNonCachedPageTableLevels = mode_lib->ms.cache_display_cfg.plane.HostVMMaxPageTableLevels;
8709 	CalculateVMRowAndSwath_params->GPUVMMaxPageTableLevels = mode_lib->ms.cache_display_cfg.plane.GPUVMMaxPageTableLevels;
8710 	CalculateVMRowAndSwath_params->GPUVMMinPageSizeKBytes = mode_lib->ms.cache_display_cfg.plane.GPUVMMinPageSizeKBytes;
8711 	CalculateVMRowAndSwath_params->HostVMMinPageSize = mode_lib->ms.soc.hostvm_min_page_size_kbytes * 1024;
8712 	CalculateVMRowAndSwath_params->PTEBufferModeOverrideEn = mode_lib->ms.cache_display_cfg.plane.PTEBufferModeOverrideEn;
8713 	CalculateVMRowAndSwath_params->PTEBufferModeOverrideVal = mode_lib->ms.cache_display_cfg.plane.PTEBufferMode;
8714 	CalculateVMRowAndSwath_params->PTEBufferSizeNotExceeded = s->dummy_boolean_array[0];
8715 	CalculateVMRowAndSwath_params->DCCMetaBufferSizeNotExceeded = s->dummy_boolean_array[1];
8716 	CalculateVMRowAndSwath_params->dpte_row_width_luma_ub = locals->dpte_row_width_luma_ub;
8717 	CalculateVMRowAndSwath_params->dpte_row_width_chroma_ub = locals->dpte_row_width_chroma_ub;
8718 	CalculateVMRowAndSwath_params->dpte_row_height_luma = locals->dpte_row_height;
8719 	CalculateVMRowAndSwath_params->dpte_row_height_chroma = locals->dpte_row_height_chroma;
8720 	CalculateVMRowAndSwath_params->dpte_row_height_linear_luma = locals->dpte_row_height_linear;
8721 	CalculateVMRowAndSwath_params->dpte_row_height_linear_chroma = locals->dpte_row_height_linear_chroma;
8722 	CalculateVMRowAndSwath_params->meta_req_width = locals->meta_req_width;
8723 	CalculateVMRowAndSwath_params->meta_req_width_chroma = locals->meta_req_width_chroma;
8724 	CalculateVMRowAndSwath_params->meta_req_height = locals->meta_req_height;
8725 	CalculateVMRowAndSwath_params->meta_req_height_chroma = locals->meta_req_height_chroma;
8726 	CalculateVMRowAndSwath_params->meta_row_width = locals->meta_row_width;
8727 	CalculateVMRowAndSwath_params->meta_row_width_chroma = locals->meta_row_width_chroma;
8728 	CalculateVMRowAndSwath_params->meta_row_height = locals->meta_row_height;
8729 	CalculateVMRowAndSwath_params->meta_row_height_chroma = locals->meta_row_height_chroma;
8730 	CalculateVMRowAndSwath_params->vm_group_bytes = locals->vm_group_bytes;
8731 	CalculateVMRowAndSwath_params->dpte_group_bytes = locals->dpte_group_bytes;
8732 	CalculateVMRowAndSwath_params->PixelPTEReqWidthY = locals->PixelPTEReqWidthY;
8733 	CalculateVMRowAndSwath_params->PixelPTEReqHeightY = locals->PixelPTEReqHeightY;
8734 	CalculateVMRowAndSwath_params->PTERequestSizeY = locals->PTERequestSizeY;
8735 	CalculateVMRowAndSwath_params->PixelPTEReqWidthC = locals->PixelPTEReqWidthC;
8736 	CalculateVMRowAndSwath_params->PixelPTEReqHeightC = locals->PixelPTEReqHeightC;
8737 	CalculateVMRowAndSwath_params->PTERequestSizeC = locals->PTERequestSizeC;
8738 	CalculateVMRowAndSwath_params->dpde0_bytes_per_frame_ub_l = locals->dpde0_bytes_per_frame_ub_l;
8739 	CalculateVMRowAndSwath_params->meta_pte_bytes_per_frame_ub_l = locals->meta_pte_bytes_per_frame_ub_l;
8740 	CalculateVMRowAndSwath_params->dpde0_bytes_per_frame_ub_c = locals->dpde0_bytes_per_frame_ub_c;
8741 	CalculateVMRowAndSwath_params->meta_pte_bytes_per_frame_ub_c = locals->meta_pte_bytes_per_frame_ub_c;
8742 	CalculateVMRowAndSwath_params->PrefetchSourceLinesY = locals->PrefetchSourceLinesY;
8743 	CalculateVMRowAndSwath_params->PrefetchSourceLinesC = locals->PrefetchSourceLinesC;
8744 	CalculateVMRowAndSwath_params->VInitPreFillY = locals->VInitPreFillY;
8745 	CalculateVMRowAndSwath_params->VInitPreFillC = locals->VInitPreFillC;
8746 	CalculateVMRowAndSwath_params->MaxNumSwathY = locals->MaxNumSwathY;
8747 	CalculateVMRowAndSwath_params->MaxNumSwathC = locals->MaxNumSwathC;
8748 	CalculateVMRowAndSwath_params->meta_row_bw = locals->meta_row_bw;
8749 	CalculateVMRowAndSwath_params->dpte_row_bw = locals->dpte_row_bw;
8750 	CalculateVMRowAndSwath_params->PixelPTEBytesPerRow = locals->PixelPTEBytesPerRow;
8751 	CalculateVMRowAndSwath_params->PDEAndMetaPTEBytesFrame = locals->PDEAndMetaPTEBytesFrame;
8752 	CalculateVMRowAndSwath_params->MetaRowByte = locals->MetaRowByte;
8753 	CalculateVMRowAndSwath_params->use_one_row_for_frame = locals->use_one_row_for_frame;
8754 	CalculateVMRowAndSwath_params->use_one_row_for_frame_flip = locals->use_one_row_for_frame_flip;
8755 	CalculateVMRowAndSwath_params->UsesMALLForStaticScreen = locals->UsesMALLForStaticScreen;
8756 	CalculateVMRowAndSwath_params->PTE_BUFFER_MODE = locals->PTE_BUFFER_MODE;
8757 	CalculateVMRowAndSwath_params->BIGK_FRAGMENT_SIZE = locals->BIGK_FRAGMENT_SIZE;
8758 
8759 	CalculateVMRowAndSwath(&mode_lib->scratch,
8760 	CalculateVMRowAndSwath_params);
8761 
8762 	s->ReorderBytes = (dml_uint_t)(mode_lib->ms.soc.num_chans * dml_max3(
8763 			mode_lib->ms.soc.urgent_out_of_order_return_per_channel_pixel_only_bytes,
8764 			mode_lib->ms.soc.urgent_out_of_order_return_per_channel_pixel_and_vm_bytes,
8765 			mode_lib->ms.soc.urgent_out_of_order_return_per_channel_vm_only_bytes));
8766 
8767 	s->VMDataOnlyReturnBW = dml_get_return_bw_mbps_vm_only(&mode_lib->ms.soc,
8768 																	mode_lib->ms.state.use_ideal_dram_bw_strobe,
8769 																	mode_lib->ms.cache_display_cfg.plane.HostVMEnable,
8770 																	locals->Dcfclk,
8771 																	mode_lib->ms.FabricClock,
8772 																	mode_lib->ms.DRAMSpeed);
8773 
8774 #ifdef __DML_VBA_DEBUG__
8775 	dml_print("DML::%s: locals->Dcfclk = %f\n", __func__, locals->Dcfclk);
8776 	dml_print("DML::%s: mode_lib->ms.soc.return_bus_width_bytes = %u\n", __func__, mode_lib->ms.soc.return_bus_width_bytes);
8777 	dml_print("DML::%s: mode_lib->ms.FabricClock = %f\n", __func__, mode_lib->ms.FabricClock);
8778 	dml_print("DML::%s: mode_lib->ms.soc.fabric_datapath_to_dcn_data_return_bytes = %u\n", __func__, mode_lib->ms.soc.fabric_datapath_to_dcn_data_return_bytes);
8779 	dml_print("DML::%s: mode_lib->ms.soc.pct_ideal_sdp_bw_after_urgent = %f\n", __func__, mode_lib->ms.soc.pct_ideal_sdp_bw_after_urgent);
8780 	dml_print("DML::%s: mode_lib->ms.DRAMSpeed = %f\n", __func__, mode_lib->ms.DRAMSpeed);
8781 	dml_print("DML::%s: mode_lib->ms.soc.num_chans = %u\n", __func__, mode_lib->ms.soc.num_chans);
8782 	dml_print("DML::%s: mode_lib->ms.soc.dram_channel_width_bytes = %u\n", __func__, mode_lib->ms.soc.dram_channel_width_bytes);
8783 	dml_print("DML::%s: mode_lib->ms.state_idx = %u\n", __func__, mode_lib->ms.state_idx);
8784 	dml_print("DML::%s: mode_lib->ms.max_state_idx = %u\n", __func__, mode_lib->ms.max_state_idx);
8785 	dml_print("DML::%s: mode_lib->ms.state.use_ideal_dram_bw_strobe = %u\n", __func__, mode_lib->ms.state.use_ideal_dram_bw_strobe);
8786 	dml_print("DML::%s: VMDataOnlyReturnBW = %f\n", __func__, s->VMDataOnlyReturnBW);
8787 	dml_print("DML::%s: ReturnBW = %f\n",  __func__, mode_lib->ms.ReturnBW);
8788 #endif
8789 
8790 	s->HostVMInefficiencyFactor = 1.0;
8791 	if (mode_lib->ms.cache_display_cfg.plane.GPUVMEnable && mode_lib->ms.cache_display_cfg.plane.HostVMEnable)
8792 		s->HostVMInefficiencyFactor = mode_lib->ms.ReturnBW / s->VMDataOnlyReturnBW;
8793 
8794 	s->TotalDCCActiveDPP = 0;
8795 	s->TotalActiveDPP = 0;
8796 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
8797 		s->TotalActiveDPP = s->TotalActiveDPP + mode_lib->ms.cache_display_cfg.hw.DPPPerSurface[k];
8798 		if (mode_lib->ms.cache_display_cfg.surface.DCCEnable[k])
8799 			s->TotalDCCActiveDPP = s->TotalDCCActiveDPP + mode_lib->ms.cache_display_cfg.hw.DPPPerSurface[k];
8800 	}
8801 
8802 	locals->UrgentExtraLatency = CalculateExtraLatency(
8803 			mode_lib->ms.soc.round_trip_ping_latency_dcfclk_cycles,
8804 			s->ReorderBytes,
8805 			locals->Dcfclk,
8806 			s->TotalActiveDPP,
8807 			mode_lib->ms.ip.pixel_chunk_size_kbytes,
8808 			s->TotalDCCActiveDPP,
8809 			mode_lib->ms.ip.meta_chunk_size_kbytes,
8810 			mode_lib->ms.ReturnBW,
8811 			mode_lib->ms.cache_display_cfg.plane.GPUVMEnable,
8812 			mode_lib->ms.cache_display_cfg.plane.HostVMEnable,
8813 			mode_lib->ms.num_active_planes,
8814 			mode_lib->ms.cache_display_cfg.hw.DPPPerSurface,
8815 			locals->dpte_group_bytes,
8816 			s->HostVMInefficiencyFactor,
8817 			mode_lib->ms.soc.hostvm_min_page_size_kbytes * 1024,
8818 			mode_lib->ms.cache_display_cfg.plane.HostVMMaxPageTableLevels);
8819 
8820 	locals->TCalc = 24.0 / locals->DCFCLKDeepSleep;
8821 
8822 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
8823 		if (mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming[k] == k) {
8824 			if (mode_lib->ms.cache_display_cfg.writeback.WritebackEnable[k] == true) {
8825 				locals->WritebackDelay[k] =
8826 						mode_lib->ms.state.writeback_latency_us
8827 								+ CalculateWriteBackDelay(
8828 										mode_lib->ms.cache_display_cfg.writeback.WritebackPixelFormat[k],
8829 										mode_lib->ms.cache_display_cfg.writeback.WritebackHRatio[k],
8830 										mode_lib->ms.cache_display_cfg.writeback.WritebackVRatio[k],
8831 										mode_lib->ms.cache_display_cfg.writeback.WritebackVTaps[k],
8832 										mode_lib->ms.cache_display_cfg.writeback.WritebackDestinationWidth[k],
8833 										mode_lib->ms.cache_display_cfg.writeback.WritebackDestinationHeight[k],
8834 										mode_lib->ms.cache_display_cfg.writeback.WritebackSourceHeight[k],
8835 										mode_lib->ms.cache_display_cfg.timing.HTotal[k]) / locals->Dispclk;
8836 			} else
8837 				locals->WritebackDelay[k] = 0;
8838 			for (j = 0; j < mode_lib->ms.num_active_planes; ++j) {
8839 				if (mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming[j] == k
8840 						&& mode_lib->ms.cache_display_cfg.writeback.WritebackEnable[j] == true) {
8841 					locals->WritebackDelay[k] =
8842 							dml_max(
8843 									locals->WritebackDelay[k],
8844 									mode_lib->ms.state.writeback_latency_us
8845 											+ CalculateWriteBackDelay(
8846 													mode_lib->ms.cache_display_cfg.writeback.WritebackPixelFormat[j],
8847 													mode_lib->ms.cache_display_cfg.writeback.WritebackHRatio[j],
8848 													mode_lib->ms.cache_display_cfg.writeback.WritebackVRatio[j],
8849 													mode_lib->ms.cache_display_cfg.writeback.WritebackVTaps[j],
8850 													mode_lib->ms.cache_display_cfg.writeback.WritebackDestinationWidth[j],
8851 													mode_lib->ms.cache_display_cfg.writeback.WritebackDestinationHeight[j],
8852 													mode_lib->ms.cache_display_cfg.writeback.WritebackSourceHeight[j],
8853 													mode_lib->ms.cache_display_cfg.timing.HTotal[k]) / locals->Dispclk);
8854 				}
8855 			}
8856 		}
8857 	}
8858 
8859 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k)
8860 		for (j = 0; j < mode_lib->ms.num_active_planes; ++j)
8861 			if (mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming[k] == j)
8862 				locals->WritebackDelay[k] = locals->WritebackDelay[j];
8863 
8864 	locals->UrgentLatency = CalculateUrgentLatency(mode_lib->ms.state.urgent_latency_pixel_data_only_us,
8865 												mode_lib->ms.state.urgent_latency_pixel_mixed_with_vm_data_us,
8866 												mode_lib->ms.state.urgent_latency_vm_data_only_us,
8867 												mode_lib->ms.soc.do_urgent_latency_adjustment,
8868 												mode_lib->ms.state.urgent_latency_adjustment_fabric_clock_component_us,
8869 												mode_lib->ms.state.urgent_latency_adjustment_fabric_clock_reference_mhz,
8870 												mode_lib->ms.FabricClock);
8871 
8872 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
8873 		CalculateUrgentBurstFactor(mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange[k],
8874 									locals->swath_width_luma_ub[k],
8875 									locals->swath_width_chroma_ub[k],
8876 									locals->SwathHeightY[k],
8877 									locals->SwathHeightC[k],
8878 									mode_lib->ms.cache_display_cfg.timing.HTotal[k] / mode_lib->ms.cache_display_cfg.timing.PixelClock[k],
8879 									locals->UrgentLatency,
8880 									mode_lib->ms.ip.cursor_buffer_size,
8881 									mode_lib->ms.cache_display_cfg.plane.CursorWidth[k],
8882 									mode_lib->ms.cache_display_cfg.plane.CursorBPP[k],
8883 									mode_lib->ms.cache_display_cfg.plane.VRatio[k],
8884 									mode_lib->ms.cache_display_cfg.plane.VRatioChroma[k],
8885 									locals->BytePerPixelDETY[k],
8886 									locals->BytePerPixelDETC[k],
8887 									locals->DETBufferSizeY[k],
8888 									locals->DETBufferSizeC[k],
8889 
8890 									/* output */
8891 									&locals->UrgBurstFactorCursor[k],
8892 									&locals->UrgBurstFactorLuma[k],
8893 									&locals->UrgBurstFactorChroma[k],
8894 									&locals->NoUrgentLatencyHiding[k]);
8895 
8896 		locals->cursor_bw[k] = mode_lib->ms.cache_display_cfg.plane.NumberOfCursors[k] * mode_lib->ms.cache_display_cfg.plane.CursorWidth[k] * mode_lib->ms.cache_display_cfg.plane.CursorBPP[k] / 8.0 /
8897 								((dml_float_t) mode_lib->ms.cache_display_cfg.timing.HTotal[k] / mode_lib->ms.cache_display_cfg.timing.PixelClock[k]) * mode_lib->ms.cache_display_cfg.plane.VRatio[k];
8898 	}
8899 
8900 	s->VStartupLines = __DML_VBA_MIN_VSTARTUP__;
8901 	s->MaxVStartupAllPlanes = 0;
8902 
8903 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
8904 		s->MaxVStartupLines[k] = CalculateMaxVStartup(k,
8905 													mode_lib->ms.ip.ptoi_supported,
8906 													mode_lib->ms.ip.vblank_nom_default_us,
8907 													&mode_lib->ms.cache_display_cfg.timing,
8908 													locals->WritebackDelay[k]);
8909 
8910 #ifdef __DML_VBA_DEBUG__
8911 		dml_print("DML::%s: k=%u MaxVStartupLines   = %u\n", __func__, k, s->MaxVStartupLines[k]);
8912 		dml_print("DML::%s: k=%u WritebackDelay     = %f\n", __func__, k, locals->WritebackDelay[k]);
8913 #endif
8914 	}
8915 
8916 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k)
8917 		s->MaxVStartupAllPlanes = (dml_uint_t)(dml_max(s->MaxVStartupAllPlanes, s->MaxVStartupLines[k]));
8918 
8919 	s->ImmediateFlipRequirementFinal = false;
8920 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
8921 		s->ImmediateFlipRequirementFinal = s->ImmediateFlipRequirementFinal || (mode_lib->ms.policy.ImmediateFlipRequirement[k] == dml_immediate_flip_required);
8922 	}
8923 #ifdef __DML_VBA_DEBUG__
8924 	dml_print("DML::%s: ImmediateFlipRequirementFinal = %u\n", __func__, s->ImmediateFlipRequirementFinal);
8925 #endif
8926 
8927 	// The prefetch scheduling should only be calculated once as per AllowForPStateChangeOrStutterInVBlank requirement
8928 	// If the AllowForPStateChangeOrStutterInVBlank requirement is not strict (i.e. only try those power saving feature
8929 	// if possible, then will try to program for the best power saving features in order of difficulty (dram, fclk, stutter)
8930 	s->iteration = 0;
8931 	s->MaxTotalRDBandwidth = 0;
8932 	s->AllPrefetchModeTested = false;
8933 	for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
8934 		CalculatePrefetchMode(mode_lib->ms.policy.AllowForPStateChangeOrStutterInVBlank[k], &s->MinPrefetchMode[k], &s->MaxPrefetchMode[k]);
8935 		s->NextPrefetchMode[k] = s->MinPrefetchMode[k];
8936 	}
8937 
8938 	do {
8939 		s->MaxTotalRDBandwidthNoUrgentBurst = 0.0;
8940 		s->DestinationLineTimesForPrefetchLessThan2 = false;
8941 		s->VRatioPrefetchMoreThanMax = false;
8942 
8943 		dml_print("DML::%s: Start one iteration: VStartupLines = %u\n", __func__, s->VStartupLines);
8944 
8945 		s->AllPrefetchModeTested = true;
8946 		s->MaxTotalRDBandwidth = 0;
8947 		for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
8948 			locals->PrefetchMode[k] = s->NextPrefetchMode[k];
8949 			TWait = CalculateTWait(
8950 					locals->PrefetchMode[k],
8951 					mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange[k],
8952 					mode_lib->ms.policy.SynchronizeDRRDisplaysForUCLKPStateChangeFinal,
8953 					mode_lib->ms.cache_display_cfg.timing.DRRDisplay[k],
8954 					mode_lib->ms.state.dram_clock_change_latency_us,
8955 					mode_lib->ms.state.fclk_change_latency_us,
8956 					locals->UrgentLatency,
8957 					mode_lib->ms.state.sr_enter_plus_exit_time_us);
8958 
8959 			myPipe = &s->myPipe;
8960 			myPipe->Dppclk = locals->Dppclk[k];
8961 			myPipe->Dispclk = locals->Dispclk;
8962 			myPipe->PixelClock = mode_lib->ms.cache_display_cfg.timing.PixelClock[k];
8963 			myPipe->DCFClkDeepSleep = locals->DCFCLKDeepSleep;
8964 			myPipe->DPPPerSurface = mode_lib->ms.cache_display_cfg.hw.DPPPerSurface[k];
8965 			myPipe->ScalerEnabled = mode_lib->ms.cache_display_cfg.plane.ScalerEnabled[k];
8966 			myPipe->SourceScan = mode_lib->ms.cache_display_cfg.plane.SourceScan[k];
8967 			myPipe->BlockWidth256BytesY = locals->BlockWidth256BytesY[k];
8968 			myPipe->BlockHeight256BytesY = locals->BlockHeight256BytesY[k];
8969 			myPipe->BlockWidth256BytesC = locals->BlockWidth256BytesC[k];
8970 			myPipe->BlockHeight256BytesC = locals->BlockHeight256BytesC[k];
8971 			myPipe->InterlaceEnable = mode_lib->ms.cache_display_cfg.timing.Interlace[k];
8972 			myPipe->NumberOfCursors = mode_lib->ms.cache_display_cfg.plane.NumberOfCursors[k];
8973 			myPipe->VBlank = mode_lib->ms.cache_display_cfg.timing.VTotal[k] - mode_lib->ms.cache_display_cfg.timing.VActive[k];
8974 			myPipe->HTotal = mode_lib->ms.cache_display_cfg.timing.HTotal[k];
8975 			myPipe->HActive = mode_lib->ms.cache_display_cfg.timing.HActive[k];
8976 			myPipe->DCCEnable = mode_lib->ms.cache_display_cfg.surface.DCCEnable[k];
8977 			myPipe->ODMMode = mode_lib->ms.cache_display_cfg.hw.ODMMode[k];
8978 			myPipe->SourcePixelFormat = mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k];
8979 			myPipe->BytePerPixelY = locals->BytePerPixelY[k];
8980 			myPipe->BytePerPixelC = locals->BytePerPixelC[k];
8981 			myPipe->ProgressiveToInterlaceUnitInOPP = mode_lib->ms.ip.ptoi_supported;
8982 
8983 #ifdef __DML_VBA_DEBUG__
8984 			dml_print("DML::%s: Calling CalculatePrefetchSchedule for k=%u\n", __func__, k);
8985 			dml_print("DML::%s: AllowForPStateChangeOrStutterInVBlank = %u\n", __func__, mode_lib->ms.policy.AllowForPStateChangeOrStutterInVBlank[k]);
8986 			dml_print("DML::%s: PrefetchMode[k] = %u (Min=%u Max=%u)\n", __func__, locals->PrefetchMode[k], s->MinPrefetchMode[k], s->MaxPrefetchMode[k]);
8987 #endif
8988 
8989 			CalculatePrefetchSchedule_params->EnhancedPrefetchScheduleAccelerationFinal = mode_lib->ms.policy.EnhancedPrefetchScheduleAccelerationFinal;
8990 			CalculatePrefetchSchedule_params->HostVMInefficiencyFactor = s->HostVMInefficiencyFactor;
8991 			CalculatePrefetchSchedule_params->myPipe = myPipe;
8992 			CalculatePrefetchSchedule_params->DSCDelay = locals->DSCDelay[k];
8993 			CalculatePrefetchSchedule_params->DPPCLKDelaySubtotalPlusCNVCFormater = mode_lib->ms.ip.dppclk_delay_subtotal + mode_lib->ms.ip.dppclk_delay_cnvc_formatter;
8994 			CalculatePrefetchSchedule_params->DPPCLKDelaySCL = mode_lib->ms.ip.dppclk_delay_scl;
8995 			CalculatePrefetchSchedule_params->DPPCLKDelaySCLLBOnly = mode_lib->ms.ip.dppclk_delay_scl_lb_only;
8996 			CalculatePrefetchSchedule_params->DPPCLKDelayCNVCCursor = mode_lib->ms.ip.dppclk_delay_cnvc_cursor;
8997 			CalculatePrefetchSchedule_params->DISPCLKDelaySubtotal = mode_lib->ms.ip.dispclk_delay_subtotal;
8998 			CalculatePrefetchSchedule_params->DPP_RECOUT_WIDTH = (dml_uint_t)(locals->SwathWidthY[k] / mode_lib->ms.cache_display_cfg.plane.HRatio[k]);
8999 			CalculatePrefetchSchedule_params->OutputFormat = mode_lib->ms.cache_display_cfg.output.OutputFormat[k];
9000 			CalculatePrefetchSchedule_params->MaxInterDCNTileRepeaters = mode_lib->ms.ip.max_inter_dcn_tile_repeaters;
9001 			CalculatePrefetchSchedule_params->VStartup = (dml_uint_t)(dml_min(s->VStartupLines, s->MaxVStartupLines[k]));
9002 			CalculatePrefetchSchedule_params->MaxVStartup = s->MaxVStartupLines[k];
9003 			CalculatePrefetchSchedule_params->GPUVMPageTableLevels = mode_lib->ms.cache_display_cfg.plane.GPUVMMaxPageTableLevels;
9004 			CalculatePrefetchSchedule_params->GPUVMEnable = mode_lib->ms.cache_display_cfg.plane.GPUVMEnable;
9005 			CalculatePrefetchSchedule_params->HostVMEnable = mode_lib->ms.cache_display_cfg.plane.HostVMEnable;
9006 			CalculatePrefetchSchedule_params->HostVMMaxNonCachedPageTableLevels = mode_lib->ms.cache_display_cfg.plane.HostVMMaxPageTableLevels;
9007 			CalculatePrefetchSchedule_params->HostVMMinPageSize = mode_lib->ms.soc.hostvm_min_page_size_kbytes * 1024;
9008 			CalculatePrefetchSchedule_params->DynamicMetadataEnable = mode_lib->ms.cache_display_cfg.plane.DynamicMetadataEnable[k];
9009 			CalculatePrefetchSchedule_params->DynamicMetadataVMEnabled = mode_lib->ms.ip.dynamic_metadata_vm_enabled;
9010 			CalculatePrefetchSchedule_params->DynamicMetadataLinesBeforeActiveRequired = mode_lib->ms.cache_display_cfg.plane.DynamicMetadataLinesBeforeActiveRequired[k];
9011 			CalculatePrefetchSchedule_params->DynamicMetadataTransmittedBytes = mode_lib->ms.cache_display_cfg.plane.DynamicMetadataTransmittedBytes[k];
9012 			CalculatePrefetchSchedule_params->UrgentLatency = locals->UrgentLatency;
9013 			CalculatePrefetchSchedule_params->UrgentExtraLatency = locals->UrgentExtraLatency;
9014 			CalculatePrefetchSchedule_params->TCalc = locals->TCalc;
9015 			CalculatePrefetchSchedule_params->PDEAndMetaPTEBytesFrame = locals->PDEAndMetaPTEBytesFrame[k];
9016 			CalculatePrefetchSchedule_params->MetaRowByte = locals->MetaRowByte[k];
9017 			CalculatePrefetchSchedule_params->PixelPTEBytesPerRow = locals->PixelPTEBytesPerRow[k];
9018 			CalculatePrefetchSchedule_params->PrefetchSourceLinesY = locals->PrefetchSourceLinesY[k];
9019 			CalculatePrefetchSchedule_params->VInitPreFillY = locals->VInitPreFillY[k];
9020 			CalculatePrefetchSchedule_params->MaxNumSwathY = locals->MaxNumSwathY[k];
9021 			CalculatePrefetchSchedule_params->PrefetchSourceLinesC = locals->PrefetchSourceLinesC[k];
9022 			CalculatePrefetchSchedule_params->VInitPreFillC = locals->VInitPreFillC[k];
9023 			CalculatePrefetchSchedule_params->MaxNumSwathC = locals->MaxNumSwathC[k];
9024 			CalculatePrefetchSchedule_params->swath_width_luma_ub = locals->swath_width_luma_ub[k];
9025 			CalculatePrefetchSchedule_params->swath_width_chroma_ub = locals->swath_width_chroma_ub[k];
9026 			CalculatePrefetchSchedule_params->SwathHeightY = locals->SwathHeightY[k];
9027 			CalculatePrefetchSchedule_params->SwathHeightC = locals->SwathHeightC[k];
9028 			CalculatePrefetchSchedule_params->TWait = TWait;
9029 			CalculatePrefetchSchedule_params->DSTXAfterScaler = &locals->DSTXAfterScaler[k];
9030 			CalculatePrefetchSchedule_params->DSTYAfterScaler = &locals->DSTYAfterScaler[k];
9031 			CalculatePrefetchSchedule_params->DestinationLinesForPrefetch = &locals->DestinationLinesForPrefetch[k];
9032 			CalculatePrefetchSchedule_params->DestinationLinesToRequestVMInVBlank = &locals->DestinationLinesToRequestVMInVBlank[k];
9033 			CalculatePrefetchSchedule_params->DestinationLinesToRequestRowInVBlank = &locals->DestinationLinesToRequestRowInVBlank[k];
9034 			CalculatePrefetchSchedule_params->VRatioPrefetchY = &locals->VRatioPrefetchY[k];
9035 			CalculatePrefetchSchedule_params->VRatioPrefetchC = &locals->VRatioPrefetchC[k];
9036 			CalculatePrefetchSchedule_params->RequiredPrefetchPixDataBWLuma = &locals->RequiredPrefetchPixDataBWLuma[k];
9037 			CalculatePrefetchSchedule_params->RequiredPrefetchPixDataBWChroma = &locals->RequiredPrefetchPixDataBWChroma[k];
9038 			CalculatePrefetchSchedule_params->NotEnoughTimeForDynamicMetadata = &locals->NotEnoughTimeForDynamicMetadata[k];
9039 			CalculatePrefetchSchedule_params->Tno_bw = &locals->Tno_bw[k];
9040 			CalculatePrefetchSchedule_params->prefetch_vmrow_bw = &locals->prefetch_vmrow_bw[k];
9041 			CalculatePrefetchSchedule_params->Tdmdl_vm = &locals->Tdmdl_vm[k];
9042 			CalculatePrefetchSchedule_params->Tdmdl = &locals->Tdmdl[k];
9043 			CalculatePrefetchSchedule_params->TSetup = &locals->TSetup[k];
9044 			CalculatePrefetchSchedule_params->VUpdateOffsetPix = &locals->VUpdateOffsetPix[k];
9045 			CalculatePrefetchSchedule_params->VUpdateWidthPix = &locals->VUpdateWidthPix[k];
9046 			CalculatePrefetchSchedule_params->VReadyOffsetPix = &locals->VReadyOffsetPix[k];
9047 
9048 			locals->NoTimeToPrefetch[k] =
9049 				CalculatePrefetchSchedule(&mode_lib->scratch,
9050 					CalculatePrefetchSchedule_params);
9051 
9052 #ifdef __DML_VBA_DEBUG__
9053 			dml_print("DML::%s: k=%0u NoTimeToPrefetch=%0d\n", __func__, k, locals->NoTimeToPrefetch[k]);
9054 #endif
9055 			locals->VStartup[k] = (dml_uint_t)(dml_min(s->VStartupLines, s->MaxVStartupLines[k]));
9056 			locals->VStartupMin[k] = locals->VStartup[k];
9057 		}
9058 
9059 		for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
9060 			CalculateUrgentBurstFactor(
9061 				mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange[k],
9062 				locals->swath_width_luma_ub[k],
9063 				locals->swath_width_chroma_ub[k],
9064 				locals->SwathHeightY[k],
9065 				locals->SwathHeightC[k],
9066 				mode_lib->ms.cache_display_cfg.timing.HTotal[k] / mode_lib->ms.cache_display_cfg.timing.PixelClock[k],
9067 				locals->UrgentLatency,
9068 				mode_lib->ms.ip.cursor_buffer_size,
9069 				mode_lib->ms.cache_display_cfg.plane.CursorWidth[k],
9070 				mode_lib->ms.cache_display_cfg.plane.CursorBPP[k],
9071 				locals->VRatioPrefetchY[k],
9072 				locals->VRatioPrefetchC[k],
9073 				locals->BytePerPixelDETY[k],
9074 				locals->BytePerPixelDETC[k],
9075 				locals->DETBufferSizeY[k],
9076 				locals->DETBufferSizeC[k],
9077 				/* Output */
9078 				&locals->UrgBurstFactorCursorPre[k],
9079 				&locals->UrgBurstFactorLumaPre[k],
9080 				&locals->UrgBurstFactorChromaPre[k],
9081 				&locals->NoUrgentLatencyHidingPre[k]);
9082 
9083 			locals->cursor_bw_pre[k] = mode_lib->ms.cache_display_cfg.plane.NumberOfCursors[k] * mode_lib->ms.cache_display_cfg.plane.CursorWidth[k] * mode_lib->ms.cache_display_cfg.plane.CursorBPP[k] / 8.0 / (mode_lib->ms.cache_display_cfg.timing.HTotal[k] / mode_lib->ms.cache_display_cfg.timing.PixelClock[k]) * locals->VRatioPrefetchY[k];
9084 
9085 #ifdef __DML_VBA_DEBUG__
9086 			dml_print("DML::%s: k=%0u DPPPerSurface=%u\n", __func__, k, mode_lib->ms.cache_display_cfg.hw.DPPPerSurface[k]);
9087 			dml_print("DML::%s: k=%0u UrgBurstFactorLuma=%f\n", __func__, k, locals->UrgBurstFactorLuma[k]);
9088 			dml_print("DML::%s: k=%0u UrgBurstFactorChroma=%f\n", __func__, k, locals->UrgBurstFactorChroma[k]);
9089 			dml_print("DML::%s: k=%0u UrgBurstFactorLumaPre=%f\n", __func__, k, locals->UrgBurstFactorLumaPre[k]);
9090 			dml_print("DML::%s: k=%0u UrgBurstFactorChromaPre=%f\n", __func__, k, locals->UrgBurstFactorChromaPre[k]);
9091 
9092 			dml_print("DML::%s: k=%0u VRatioPrefetchY=%f\n", __func__, k, locals->VRatioPrefetchY[k]);
9093 			dml_print("DML::%s: k=%0u VRatioY=%f\n", __func__, k, mode_lib->ms.cache_display_cfg.plane.VRatio[k]);
9094 
9095 			dml_print("DML::%s: k=%0u prefetch_vmrow_bw=%f\n", __func__, k, locals->prefetch_vmrow_bw[k]);
9096 			dml_print("DML::%s: k=%0u ReadBandwidthSurfaceLuma=%f\n", __func__, k, locals->ReadBandwidthSurfaceLuma[k]);
9097 			dml_print("DML::%s: k=%0u ReadBandwidthSurfaceChroma=%f\n", __func__, k, locals->ReadBandwidthSurfaceChroma[k]);
9098 			dml_print("DML::%s: k=%0u cursor_bw=%f\n", __func__, k, locals->cursor_bw[k]);
9099 			dml_print("DML::%s: k=%0u meta_row_bw=%f\n", __func__, k, locals->meta_row_bw[k]);
9100 			dml_print("DML::%s: k=%0u dpte_row_bw=%f\n", __func__, k, locals->dpte_row_bw[k]);
9101 			dml_print("DML::%s: k=%0u RequiredPrefetchPixDataBWLuma=%f\n", __func__, k, locals->RequiredPrefetchPixDataBWLuma[k]);
9102 			dml_print("DML::%s: k=%0u RequiredPrefetchPixDataBWChroma=%f\n", __func__, k, locals->RequiredPrefetchPixDataBWChroma[k]);
9103 			dml_print("DML::%s: k=%0u cursor_bw_pre=%f\n", __func__, k, locals->cursor_bw_pre[k]);
9104 			dml_print("DML::%s: k=%0u MaxTotalRDBandwidthNoUrgentBurst=%f\n", __func__, k, s->MaxTotalRDBandwidthNoUrgentBurst);
9105 #endif
9106 			if (locals->DestinationLinesForPrefetch[k] < 2)
9107 				s->DestinationLineTimesForPrefetchLessThan2 = true;
9108 
9109 			if (locals->VRatioPrefetchY[k] > __DML_MAX_VRATIO_PRE_ENHANCE_PREFETCH_ACC__ ||
9110 				locals->VRatioPrefetchC[k] > __DML_MAX_VRATIO_PRE_ENHANCE_PREFETCH_ACC__ ||
9111 				((s->VStartupLines < s->MaxVStartupLines[k] || mode_lib->ms.policy.EnhancedPrefetchScheduleAccelerationFinal == 0) &&
9112 					(locals->VRatioPrefetchY[k] > __DML_MAX_VRATIO_PRE__ || locals->VRatioPrefetchC[k] > __DML_MAX_VRATIO_PRE__)))
9113 				s->VRatioPrefetchMoreThanMax = true;
9114 
9115 			//dml_bool_t DestinationLinesToRequestVMInVBlankEqualOrMoreThan32 = false;
9116 			//dml_bool_t DestinationLinesToRequestRowInVBlankEqualOrMoreThan16 = false;
9117 			//if (locals->DestinationLinesToRequestVMInVBlank[k] >= 32) {
9118 			//    DestinationLinesToRequestVMInVBlankEqualOrMoreThan32 = true;
9119 			//}
9120 
9121 			//if (locals->DestinationLinesToRequestRowInVBlank[k] >= 16) {
9122 			//    DestinationLinesToRequestRowInVBlankEqualOrMoreThan16 = true;
9123 			//}
9124 		}
9125 
9126 		locals->FractionOfUrgentBandwidth = s->MaxTotalRDBandwidthNoUrgentBurst / mode_lib->ms.ReturnBW;
9127 
9128 #ifdef __DML_VBA_DEBUG__
9129 		dml_print("DML::%s: MaxTotalRDBandwidthNoUrgentBurst=%f \n", __func__, s->MaxTotalRDBandwidthNoUrgentBurst);
9130 		dml_print("DML::%s: ReturnBW=%f \n", __func__, mode_lib->ms.ReturnBW);
9131 		dml_print("DML::%s: FractionOfUrgentBandwidth=%f \n", __func__, locals->FractionOfUrgentBandwidth);
9132 #endif
9133 
9134 		CalculatePrefetchBandwithSupport(
9135 			mode_lib->ms.num_active_planes,
9136 			mode_lib->ms.ReturnBW,
9137 			mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange,
9138 			locals->NoUrgentLatencyHidingPre,
9139 			locals->ReadBandwidthSurfaceLuma,
9140 			locals->ReadBandwidthSurfaceChroma,
9141 			locals->RequiredPrefetchPixDataBWLuma,
9142 			locals->RequiredPrefetchPixDataBWChroma,
9143 			locals->cursor_bw,
9144 			locals->meta_row_bw,
9145 			locals->dpte_row_bw,
9146 			locals->cursor_bw_pre,
9147 			locals->prefetch_vmrow_bw,
9148 			mode_lib->ms.cache_display_cfg.hw.DPPPerSurface,
9149 			locals->UrgBurstFactorLuma,
9150 			locals->UrgBurstFactorChroma,
9151 			locals->UrgBurstFactorCursor,
9152 			locals->UrgBurstFactorLumaPre,
9153 			locals->UrgBurstFactorChromaPre,
9154 			locals->UrgBurstFactorCursorPre,
9155 
9156 			/* output */
9157 			&s->MaxTotalRDBandwidth, // dml_float_t *PrefetchBandwidth
9158 			&s->MaxTotalRDBandwidthNotIncludingMALLPrefetch, // dml_float_t *PrefetchBandwidthNotIncludingMALLPrefetch
9159 			&s->dummy_single[0], // dml_float_t *FractionOfUrgentBandwidth
9160 			&locals->PrefetchModeSupported);
9161 
9162 		for (k = 0; k < mode_lib->ms.num_active_planes; ++k)
9163 			s->dummy_unit_vector[k] = 1.0;
9164 
9165 		CalculatePrefetchBandwithSupport(mode_lib->ms.num_active_planes,
9166 			mode_lib->ms.ReturnBW,
9167 			mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange,
9168 			locals->NoUrgentLatencyHidingPre,
9169 			locals->ReadBandwidthSurfaceLuma,
9170 			locals->ReadBandwidthSurfaceChroma,
9171 			locals->RequiredPrefetchPixDataBWLuma,
9172 			locals->RequiredPrefetchPixDataBWChroma,
9173 			locals->cursor_bw,
9174 			locals->meta_row_bw,
9175 			locals->dpte_row_bw,
9176 			locals->cursor_bw_pre,
9177 			locals->prefetch_vmrow_bw,
9178 			mode_lib->ms.cache_display_cfg.hw.DPPPerSurface,
9179 			s->dummy_unit_vector,
9180 			s->dummy_unit_vector,
9181 			s->dummy_unit_vector,
9182 			s->dummy_unit_vector,
9183 			s->dummy_unit_vector,
9184 			s->dummy_unit_vector,
9185 
9186 			/* output */
9187 			&s->NonUrgentMaxTotalRDBandwidth, // dml_float_t *PrefetchBandwidth
9188 			&s->NonUrgentMaxTotalRDBandwidthNotIncludingMALLPrefetch, // dml_float_t *PrefetchBandwidthNotIncludingMALLPrefetch
9189 			&locals->FractionOfUrgentBandwidth,
9190 			&s->dummy_boolean[0]); // dml_bool_t *PrefetchBandwidthSupport
9191 
9192 		if (s->VRatioPrefetchMoreThanMax != false || s->DestinationLineTimesForPrefetchLessThan2 != false) {
9193 			dml_print("DML::%s: VRatioPrefetchMoreThanMax                   = %u\n", __func__, s->VRatioPrefetchMoreThanMax);
9194 			dml_print("DML::%s: DestinationLineTimesForPrefetchLessThan2    = %u\n", __func__, s->DestinationLineTimesForPrefetchLessThan2);
9195 			locals->PrefetchModeSupported = false;
9196 		}
9197 
9198 		for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
9199 			if (locals->NoTimeToPrefetch[k] == true || locals->NotEnoughTimeForDynamicMetadata[k]) {
9200 				dml_print("DML::%s: k=%u, NoTimeToPrefetch = %0d\n", __func__, k, locals->NoTimeToPrefetch[k]);
9201 				dml_print("DML::%s: k=%u, NotEnoughTimeForDynamicMetadata=%u\n", __func__, k, locals->NotEnoughTimeForDynamicMetadata[k]);
9202 				locals->PrefetchModeSupported = false;
9203 			}
9204 		}
9205 
9206 		if (locals->PrefetchModeSupported == true && mode_lib->ms.support.ImmediateFlipSupport == true) {
9207 			locals->BandwidthAvailableForImmediateFlip = CalculateBandwidthAvailableForImmediateFlip(
9208 																	mode_lib->ms.num_active_planes,
9209 																	mode_lib->ms.ReturnBW,
9210 																	locals->ReadBandwidthSurfaceLuma,
9211 																	locals->ReadBandwidthSurfaceChroma,
9212 																	locals->RequiredPrefetchPixDataBWLuma,
9213 																	locals->RequiredPrefetchPixDataBWChroma,
9214 																	locals->cursor_bw,
9215 																	locals->cursor_bw_pre,
9216 																	mode_lib->ms.cache_display_cfg.hw.DPPPerSurface,
9217 																	locals->UrgBurstFactorLuma,
9218 																	locals->UrgBurstFactorChroma,
9219 																	locals->UrgBurstFactorCursor,
9220 																	locals->UrgBurstFactorLumaPre,
9221 																	locals->UrgBurstFactorChromaPre,
9222 																	locals->UrgBurstFactorCursorPre);
9223 
9224 			locals->TotImmediateFlipBytes = 0;
9225 			for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
9226 				if (mode_lib->ms.policy.ImmediateFlipRequirement[k] != dml_immediate_flip_not_required) {
9227 					locals->TotImmediateFlipBytes = locals->TotImmediateFlipBytes + mode_lib->ms.cache_display_cfg.hw.DPPPerSurface[k] * (locals->PDEAndMetaPTEBytesFrame[k] + locals->MetaRowByte[k]);
9228 					if (locals->use_one_row_for_frame_flip[k]) {
9229 						locals->TotImmediateFlipBytes = locals->TotImmediateFlipBytes + mode_lib->ms.cache_display_cfg.hw.DPPPerSurface[k] * (2 * locals->PixelPTEBytesPerRow[k]);
9230 					} else {
9231 						locals->TotImmediateFlipBytes = locals->TotImmediateFlipBytes + mode_lib->ms.cache_display_cfg.hw.DPPPerSurface[k] * locals->PixelPTEBytesPerRow[k];
9232 					}
9233 #ifdef __DML_VBA_DEBUG__
9234 					dml_print("DML::%s: k = %u\n", __func__, k);
9235 					dml_print("DML::%s: DPPPerSurface = %u\n", __func__, mode_lib->ms.cache_display_cfg.hw.DPPPerSurface[k]);
9236 					dml_print("DML::%s: PDEAndMetaPTEBytesFrame = %u\n", __func__, locals->PDEAndMetaPTEBytesFrame[k]);
9237 					dml_print("DML::%s: MetaRowByte = %u\n", __func__, locals->MetaRowByte[k]);
9238 					dml_print("DML::%s: PixelPTEBytesPerRow = %u\n", __func__, locals->PixelPTEBytesPerRow[k]);
9239 					dml_print("DML::%s: TotImmediateFlipBytes = %u\n", __func__, locals->TotImmediateFlipBytes);
9240 #endif
9241 				}
9242 			}
9243 			for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
9244 				CalculateFlipSchedule(
9245 						s->HostVMInefficiencyFactor,
9246 						locals->UrgentExtraLatency,
9247 						locals->UrgentLatency,
9248 						mode_lib->ms.cache_display_cfg.plane.GPUVMMaxPageTableLevels,
9249 						mode_lib->ms.cache_display_cfg.plane.HostVMEnable,
9250 						mode_lib->ms.cache_display_cfg.plane.HostVMMaxPageTableLevels,
9251 						mode_lib->ms.cache_display_cfg.plane.GPUVMEnable,
9252 						mode_lib->ms.soc.hostvm_min_page_size_kbytes * 1024,
9253 						locals->PDEAndMetaPTEBytesFrame[k],
9254 						locals->MetaRowByte[k],
9255 						locals->PixelPTEBytesPerRow[k],
9256 						locals->BandwidthAvailableForImmediateFlip,
9257 						locals->TotImmediateFlipBytes,
9258 						mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k],
9259 						mode_lib->ms.cache_display_cfg.timing.HTotal[k] / mode_lib->ms.cache_display_cfg.timing.PixelClock[k],
9260 						mode_lib->ms.cache_display_cfg.plane.VRatio[k],
9261 						mode_lib->ms.cache_display_cfg.plane.VRatioChroma[k],
9262 						locals->Tno_bw[k],
9263 						mode_lib->ms.cache_display_cfg.surface.DCCEnable[k],
9264 						locals->dpte_row_height[k],
9265 						locals->meta_row_height[k],
9266 						locals->dpte_row_height_chroma[k],
9267 						locals->meta_row_height_chroma[k],
9268 						locals->use_one_row_for_frame_flip[k],
9269 
9270 						/* Output */
9271 						&locals->DestinationLinesToRequestVMInImmediateFlip[k],
9272 						&locals->DestinationLinesToRequestRowInImmediateFlip[k],
9273 						&locals->final_flip_bw[k],
9274 						&locals->ImmediateFlipSupportedForPipe[k]);
9275 			}
9276 
9277 			CalculateImmediateFlipBandwithSupport(mode_lib->ms.num_active_planes,
9278 												mode_lib->ms.ReturnBW,
9279 												mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange,
9280 												mode_lib->ms.policy.ImmediateFlipRequirement,
9281 												locals->final_flip_bw,
9282 												locals->ReadBandwidthSurfaceLuma,
9283 												locals->ReadBandwidthSurfaceChroma,
9284 												locals->RequiredPrefetchPixDataBWLuma,
9285 												locals->RequiredPrefetchPixDataBWChroma,
9286 												locals->cursor_bw,
9287 												locals->meta_row_bw,
9288 												locals->dpte_row_bw,
9289 												locals->cursor_bw_pre,
9290 												locals->prefetch_vmrow_bw,
9291 												mode_lib->ms.cache_display_cfg.hw.DPPPerSurface,
9292 												locals->UrgBurstFactorLuma,
9293 												locals->UrgBurstFactorChroma,
9294 												locals->UrgBurstFactorCursor,
9295 												locals->UrgBurstFactorLumaPre,
9296 												locals->UrgBurstFactorChromaPre,
9297 												locals->UrgBurstFactorCursorPre,
9298 
9299 												/* output */
9300 												&locals->total_dcn_read_bw_with_flip, // dml_float_t *TotalBandwidth
9301 												&locals->total_dcn_read_bw_with_flip_not_including_MALL_prefetch, // dml_float_t TotalBandwidthNotIncludingMALLPrefetch
9302 												&s->dummy_single[0], // dml_float_t *FractionOfUrgentBandwidth
9303 												&locals->ImmediateFlipSupported); // dml_bool_t *ImmediateFlipBandwidthSupport
9304 
9305 			CalculateImmediateFlipBandwithSupport(mode_lib->ms.num_active_planes,
9306 				mode_lib->ms.ReturnBW,
9307 				mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange,
9308 				mode_lib->ms.policy.ImmediateFlipRequirement,
9309 				locals->final_flip_bw,
9310 				locals->ReadBandwidthSurfaceLuma,
9311 				locals->ReadBandwidthSurfaceChroma,
9312 				locals->RequiredPrefetchPixDataBWLuma,
9313 				locals->RequiredPrefetchPixDataBWChroma,
9314 				locals->cursor_bw,
9315 				locals->meta_row_bw,
9316 				locals->dpte_row_bw,
9317 				locals->cursor_bw_pre,
9318 				locals->prefetch_vmrow_bw,
9319 				mode_lib->ms.cache_display_cfg.hw.DPPPerSurface,
9320 				s->dummy_unit_vector,
9321 				s->dummy_unit_vector,
9322 				s->dummy_unit_vector,
9323 				s->dummy_unit_vector,
9324 				s->dummy_unit_vector,
9325 				s->dummy_unit_vector,
9326 
9327 				/* output */
9328 				&locals->non_urgent_total_dcn_read_bw_with_flip, // dml_float_t *TotalBandwidth
9329 				&locals->non_urgent_total_dcn_read_bw_with_flip_not_including_MALL_prefetch, // dml_float_t TotalBandwidthNotIncludingMALLPrefetch
9330 				&locals->FractionOfUrgentBandwidthImmediateFlip, // dml_float_t *FractionOfUrgentBandwidth
9331 				&s->dummy_boolean[0]); // dml_bool_t *ImmediateFlipBandwidthSupport
9332 
9333 			for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
9334 				if (mode_lib->ms.policy.ImmediateFlipRequirement[k] != dml_immediate_flip_not_required && locals->ImmediateFlipSupportedForPipe[k] == false) {
9335 					locals->ImmediateFlipSupported = false;
9336 #ifdef __DML_VBA_DEBUG__
9337 					dml_print("DML::%s: Pipe %0d not supporting iflip\n", __func__, k);
9338 #endif
9339 				}
9340 			}
9341 		} else {
9342 			locals->ImmediateFlipSupported = false;
9343 			locals->total_dcn_read_bw_with_flip = s->MaxTotalRDBandwidth;
9344 			locals->total_dcn_read_bw_with_flip_not_including_MALL_prefetch = s->MaxTotalRDBandwidthNotIncludingMALLPrefetch;
9345 			locals->non_urgent_total_dcn_read_bw_with_flip = s->NonUrgentMaxTotalRDBandwidth;
9346 			locals->non_urgent_total_dcn_read_bw_with_flip_not_including_MALL_prefetch = s->NonUrgentMaxTotalRDBandwidthNotIncludingMALLPrefetch;
9347 		}
9348 
9349 		/* consider flip support is okay if the flip bw is ok or (when user does't require a iflip and there is no host vm) */
9350 		locals->PrefetchAndImmediateFlipSupported = (locals->PrefetchModeSupported == true &&
9351 													((!mode_lib->ms.support.ImmediateFlipSupport && !mode_lib->ms.cache_display_cfg.plane.HostVMEnable && !s->ImmediateFlipRequirementFinal) ||
9352 													locals->ImmediateFlipSupported)) ? true : false;
9353 
9354 #ifdef __DML_VBA_DEBUG__
9355 		dml_print("DML::%s: PrefetchModeSupported = %u\n", __func__, locals->PrefetchModeSupported);
9356 		for (k = 0; k < mode_lib->ms.num_active_planes; ++k)
9357 		dml_print("DML::%s: ImmediateFlipRequirement[%u] = %u\n", __func__, k, mode_lib->ms.policy.ImmediateFlipRequirement[k] == dml_immediate_flip_required);
9358 		dml_print("DML::%s: HostVMEnable = %u\n", __func__, mode_lib->ms.cache_display_cfg.plane.HostVMEnable);
9359 		dml_print("DML::%s: ImmediateFlipSupport = %u (from mode_support)\n", __func__, mode_lib->ms.support.ImmediateFlipSupport);
9360 		dml_print("DML::%s: ImmediateFlipSupported = %u\n", __func__, locals->ImmediateFlipSupported);
9361 		dml_print("DML::%s: PrefetchAndImmediateFlipSupported = %u\n", __func__, locals->PrefetchAndImmediateFlipSupported);
9362 #endif
9363 		dml_print("DML::%s: Done one iteration: VStartupLines=%u, MaxVStartupAllPlanes=%u\n", __func__, s->VStartupLines, s->MaxVStartupAllPlanes);
9364 
9365 		s->VStartupLines = s->VStartupLines + 1;
9366 
9367 		if (s->VStartupLines > s->MaxVStartupAllPlanes) {
9368 			s->VStartupLines = __DML_VBA_MIN_VSTARTUP__;
9369 
9370 			for (k = 0; k <= mode_lib->ms.num_active_planes - 1; k++) {
9371 				s->NextPrefetchMode[k] = s->NextPrefetchMode[k] + 1;
9372 
9373 				if (s->NextPrefetchMode[k] <= s->MaxPrefetchMode[k])
9374 					s->AllPrefetchModeTested = false;
9375 				dml_print("DML::%s: VStartupLines=%u, reaches max vstartup, try next prefetch mode=%u\n", __func__, s->VStartupLines-1, s->AllPrefetchModeTested);
9376 			}
9377 		} else {
9378 			s->AllPrefetchModeTested = false;
9379 		}
9380 		s->iteration++;
9381 		if (s->iteration > 2500) {
9382 			dml_print("ERROR: DML::%s: Too many errors, exit now\n", __func__);
9383 			ASSERT(0);
9384 		}
9385 	} while (!(locals->PrefetchAndImmediateFlipSupported || s->AllPrefetchModeTested));
9386 
9387 	if (locals->PrefetchAndImmediateFlipSupported) {
9388 		dml_print("DML::%s: Good, Prefetch and flip scheduling solution found at VStartupLines=%u (MaxVStartupAllPlanes=%u)\n", __func__, s->VStartupLines-1, s->MaxVStartupAllPlanes);
9389 	} else {
9390 		dml_print("DML::%s: Bad, Prefetch and flip scheduling solution did NOT find solution! (MaxVStartupAllPlanes=%u)\n", __func__, s->MaxVStartupAllPlanes);
9391 	}
9392 
9393 	//Watermarks and NB P-State/DRAM Clock Change Support
9394 	{
9395 		s->mmSOCParameters.UrgentLatency = locals->UrgentLatency;
9396 		s->mmSOCParameters.ExtraLatency = locals->UrgentExtraLatency;
9397 		s->mmSOCParameters.WritebackLatency = mode_lib->ms.state.writeback_latency_us;
9398 		s->mmSOCParameters.DRAMClockChangeLatency = mode_lib->ms.state.dram_clock_change_latency_us;
9399 		s->mmSOCParameters.FCLKChangeLatency = mode_lib->ms.state.fclk_change_latency_us;
9400 		s->mmSOCParameters.SRExitTime = mode_lib->ms.state.sr_exit_time_us;
9401 		s->mmSOCParameters.SREnterPlusExitTime = mode_lib->ms.state.sr_enter_plus_exit_time_us;
9402 		s->mmSOCParameters.SRExitZ8Time = mode_lib->ms.state.sr_exit_z8_time_us;
9403 		s->mmSOCParameters.SREnterPlusExitZ8Time = mode_lib->ms.state.sr_enter_plus_exit_z8_time_us;
9404 		s->mmSOCParameters.USRRetrainingLatency = mode_lib->ms.state.usr_retraining_latency_us;
9405 		s->mmSOCParameters.SMNLatency = mode_lib->ms.soc.smn_latency_us;
9406 
9407 		CalculateWatermarks_params->USRRetrainingRequiredFinal = mode_lib->ms.policy.USRRetrainingRequiredFinal;
9408 		CalculateWatermarks_params->UseMALLForPStateChange = mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange;
9409 		CalculateWatermarks_params->PrefetchMode = locals->PrefetchMode;
9410 		CalculateWatermarks_params->NumberOfActiveSurfaces = mode_lib->ms.num_active_planes;
9411 		CalculateWatermarks_params->MaxLineBufferLines = mode_lib->ms.ip.max_line_buffer_lines;
9412 		CalculateWatermarks_params->LineBufferSize = mode_lib->ms.ip.line_buffer_size_bits;
9413 		CalculateWatermarks_params->WritebackInterfaceBufferSize = mode_lib->ms.ip.writeback_interface_buffer_size_kbytes;
9414 		CalculateWatermarks_params->DCFCLK = locals->Dcfclk;
9415 		CalculateWatermarks_params->ReturnBW = mode_lib->ms.ReturnBW;
9416 		CalculateWatermarks_params->SynchronizeTimingsFinal = mode_lib->ms.policy.SynchronizeTimingsFinal;
9417 		CalculateWatermarks_params->SynchronizeDRRDisplaysForUCLKPStateChangeFinal = mode_lib->ms.policy.SynchronizeDRRDisplaysForUCLKPStateChangeFinal;
9418 		CalculateWatermarks_params->DRRDisplay = mode_lib->ms.cache_display_cfg.timing.DRRDisplay;
9419 		CalculateWatermarks_params->dpte_group_bytes = locals->dpte_group_bytes;
9420 		CalculateWatermarks_params->meta_row_height = locals->meta_row_height;
9421 		CalculateWatermarks_params->meta_row_height_chroma = locals->meta_row_height_chroma;
9422 		CalculateWatermarks_params->mmSOCParameters = s->mmSOCParameters;
9423 		CalculateWatermarks_params->WritebackChunkSize = mode_lib->ms.ip.writeback_chunk_size_kbytes;
9424 		CalculateWatermarks_params->SOCCLK = mode_lib->ms.SOCCLK;
9425 		CalculateWatermarks_params->DCFClkDeepSleep = locals->DCFCLKDeepSleep;
9426 		CalculateWatermarks_params->DETBufferSizeY = locals->DETBufferSizeY;
9427 		CalculateWatermarks_params->DETBufferSizeC = locals->DETBufferSizeC;
9428 		CalculateWatermarks_params->SwathHeightY = locals->SwathHeightY;
9429 		CalculateWatermarks_params->SwathHeightC = locals->SwathHeightC;
9430 		CalculateWatermarks_params->LBBitPerPixel = mode_lib->ms.cache_display_cfg.plane.LBBitPerPixel;
9431 		CalculateWatermarks_params->SwathWidthY = locals->SwathWidthY;
9432 		CalculateWatermarks_params->SwathWidthC = locals->SwathWidthC;
9433 		CalculateWatermarks_params->HRatio = mode_lib->ms.cache_display_cfg.plane.HRatio;
9434 		CalculateWatermarks_params->HRatioChroma = mode_lib->ms.cache_display_cfg.plane.HRatioChroma;
9435 		CalculateWatermarks_params->VTaps = mode_lib->ms.cache_display_cfg.plane.VTaps;
9436 		CalculateWatermarks_params->VTapsChroma = mode_lib->ms.cache_display_cfg.plane.VTapsChroma;
9437 		CalculateWatermarks_params->VRatio = mode_lib->ms.cache_display_cfg.plane.VRatio;
9438 		CalculateWatermarks_params->VRatioChroma = mode_lib->ms.cache_display_cfg.plane.VRatioChroma;
9439 		CalculateWatermarks_params->HTotal = mode_lib->ms.cache_display_cfg.timing.HTotal;
9440 		CalculateWatermarks_params->VTotal = mode_lib->ms.cache_display_cfg.timing.VTotal;
9441 		CalculateWatermarks_params->VActive = mode_lib->ms.cache_display_cfg.timing.VActive;
9442 		CalculateWatermarks_params->PixelClock = mode_lib->ms.cache_display_cfg.timing.PixelClock;
9443 		CalculateWatermarks_params->BlendingAndTiming = mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming;
9444 		CalculateWatermarks_params->DPPPerSurface = mode_lib->ms.cache_display_cfg.hw.DPPPerSurface;
9445 		CalculateWatermarks_params->BytePerPixelDETY = locals->BytePerPixelDETY;
9446 		CalculateWatermarks_params->BytePerPixelDETC = locals->BytePerPixelDETC;
9447 		CalculateWatermarks_params->DSTXAfterScaler = locals->DSTXAfterScaler;
9448 		CalculateWatermarks_params->DSTYAfterScaler = locals->DSTYAfterScaler;
9449 		CalculateWatermarks_params->WritebackEnable = mode_lib->ms.cache_display_cfg.writeback.WritebackEnable;
9450 		CalculateWatermarks_params->WritebackPixelFormat = mode_lib->ms.cache_display_cfg.writeback.WritebackPixelFormat;
9451 		CalculateWatermarks_params->WritebackDestinationWidth = mode_lib->ms.cache_display_cfg.writeback.WritebackDestinationWidth;
9452 		CalculateWatermarks_params->WritebackDestinationHeight = mode_lib->ms.cache_display_cfg.writeback.WritebackDestinationHeight;
9453 		CalculateWatermarks_params->WritebackSourceHeight = mode_lib->ms.cache_display_cfg.writeback.WritebackSourceHeight;
9454 		CalculateWatermarks_params->UnboundedRequestEnabled = locals->UnboundedRequestEnabled;
9455 		CalculateWatermarks_params->CompressedBufferSizeInkByte = locals->CompressedBufferSizeInkByte;
9456 
9457 		// Output
9458 		CalculateWatermarks_params->Watermark = &locals->Watermark; // Watermarks *Watermark
9459 		CalculateWatermarks_params->DRAMClockChangeSupport = &locals->DRAMClockChangeSupport;
9460 		CalculateWatermarks_params->MaxActiveDRAMClockChangeLatencySupported = locals->MaxActiveDRAMClockChangeLatencySupported; // dml_float_t *MaxActiveDRAMClockChangeLatencySupported[]
9461 		CalculateWatermarks_params->SubViewportLinesNeededInMALL = locals->SubViewportLinesNeededInMALL; // dml_uint_t SubViewportLinesNeededInMALL[]
9462 		CalculateWatermarks_params->FCLKChangeSupport = &locals->FCLKChangeSupport;
9463 		CalculateWatermarks_params->MaxActiveFCLKChangeLatencySupported = &locals->MaxActiveFCLKChangeLatencySupported; // dml_float_t *MaxActiveFCLKChangeLatencySupported
9464 		CalculateWatermarks_params->USRRetrainingSupport = &locals->USRRetrainingSupport;
9465 
9466 		CalculateWatermarksMALLUseAndDRAMSpeedChangeSupport(
9467 			&mode_lib->scratch,
9468 			CalculateWatermarks_params);
9469 
9470 		/* Copy the calculated watermarks to mp.Watermark as the getter functions are
9471 		 * implemented by the DML team to copy the calculated values from the mp.Watermark interface.
9472 		 * &mode_lib->mp.Watermark and &locals->Watermark are the same address, memcpy may lead to
9473 		 * unexpected behavior. memmove should be used.
9474 		 */
9475 		memmove(&mode_lib->mp.Watermark, CalculateWatermarks_params->Watermark, sizeof(struct Watermarks));
9476 
9477 		for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
9478 			if (mode_lib->ms.cache_display_cfg.writeback.WritebackEnable[k] == true) {
9479 				locals->WritebackAllowDRAMClockChangeEndPosition[k] = dml_max(0, locals->VStartupMin[k] * mode_lib->ms.cache_display_cfg.timing.HTotal[k] /
9480 																			mode_lib->ms.cache_display_cfg.timing.PixelClock[k] - locals->Watermark.WritebackDRAMClockChangeWatermark);
9481 				locals->WritebackAllowFCLKChangeEndPosition[k] = dml_max(0, locals->VStartupMin[k] * mode_lib->ms.cache_display_cfg.timing.HTotal[k] /
9482 																			mode_lib->ms.cache_display_cfg.timing.PixelClock[k] - locals->Watermark.WritebackFCLKChangeWatermark);
9483 			} else {
9484 				locals->WritebackAllowDRAMClockChangeEndPosition[k] = 0;
9485 				locals->WritebackAllowFCLKChangeEndPosition[k] = 0;
9486 			}
9487 		}
9488 	}
9489 
9490 	//Display Pipeline Delivery Time in Prefetch, Groups
9491 	CalculatePixelDeliveryTimes(
9492 			mode_lib->ms.num_active_planes,
9493 			mode_lib->ms.cache_display_cfg.plane.VRatio,
9494 			mode_lib->ms.cache_display_cfg.plane.VRatioChroma,
9495 			locals->VRatioPrefetchY,
9496 			locals->VRatioPrefetchC,
9497 			locals->swath_width_luma_ub,
9498 			locals->swath_width_chroma_ub,
9499 			mode_lib->ms.cache_display_cfg.hw.DPPPerSurface,
9500 			mode_lib->ms.cache_display_cfg.plane.HRatio,
9501 			mode_lib->ms.cache_display_cfg.plane.HRatioChroma,
9502 			mode_lib->ms.cache_display_cfg.timing.PixelClock,
9503 			locals->PSCL_THROUGHPUT,
9504 			locals->PSCL_THROUGHPUT_CHROMA,
9505 			locals->Dppclk,
9506 			locals->BytePerPixelC,
9507 			mode_lib->ms.cache_display_cfg.plane.SourceScan,
9508 			mode_lib->ms.cache_display_cfg.plane.NumberOfCursors,
9509 			mode_lib->ms.cache_display_cfg.plane.CursorWidth,
9510 			mode_lib->ms.cache_display_cfg.plane.CursorBPP,
9511 			locals->BlockWidth256BytesY,
9512 			locals->BlockHeight256BytesY,
9513 			locals->BlockWidth256BytesC,
9514 			locals->BlockHeight256BytesC,
9515 
9516 			/* Output */
9517 			locals->DisplayPipeLineDeliveryTimeLuma,
9518 			locals->DisplayPipeLineDeliveryTimeChroma,
9519 			locals->DisplayPipeLineDeliveryTimeLumaPrefetch,
9520 			locals->DisplayPipeLineDeliveryTimeChromaPrefetch,
9521 			locals->DisplayPipeRequestDeliveryTimeLuma,
9522 			locals->DisplayPipeRequestDeliveryTimeChroma,
9523 			locals->DisplayPipeRequestDeliveryTimeLumaPrefetch,
9524 			locals->DisplayPipeRequestDeliveryTimeChromaPrefetch,
9525 			locals->CursorRequestDeliveryTime,
9526 			locals->CursorRequestDeliveryTimePrefetch);
9527 
9528 	CalculateMetaAndPTETimes(
9529 			locals->use_one_row_for_frame,
9530 			mode_lib->ms.num_active_planes,
9531 			mode_lib->ms.cache_display_cfg.plane.GPUVMEnable,
9532 			mode_lib->ms.ip.meta_chunk_size_kbytes,
9533 			mode_lib->ms.ip.min_meta_chunk_size_bytes,
9534 			mode_lib->ms.cache_display_cfg.timing.HTotal,
9535 			mode_lib->ms.cache_display_cfg.plane.VRatio,
9536 			mode_lib->ms.cache_display_cfg.plane.VRatioChroma,
9537 			locals->DestinationLinesToRequestRowInVBlank,
9538 			locals->DestinationLinesToRequestRowInImmediateFlip,
9539 			mode_lib->ms.cache_display_cfg.surface.DCCEnable,
9540 			mode_lib->ms.cache_display_cfg.timing.PixelClock,
9541 			locals->BytePerPixelY,
9542 			locals->BytePerPixelC,
9543 			mode_lib->ms.cache_display_cfg.plane.SourceScan,
9544 			locals->dpte_row_height,
9545 			locals->dpte_row_height_chroma,
9546 			locals->meta_row_width,
9547 			locals->meta_row_width_chroma,
9548 			locals->meta_row_height,
9549 			locals->meta_row_height_chroma,
9550 			locals->meta_req_width,
9551 			locals->meta_req_width_chroma,
9552 			locals->meta_req_height,
9553 			locals->meta_req_height_chroma,
9554 			locals->dpte_group_bytes,
9555 			locals->PTERequestSizeY,
9556 			locals->PTERequestSizeC,
9557 			locals->PixelPTEReqWidthY,
9558 			locals->PixelPTEReqHeightY,
9559 			locals->PixelPTEReqWidthC,
9560 			locals->PixelPTEReqHeightC,
9561 			locals->dpte_row_width_luma_ub,
9562 			locals->dpte_row_width_chroma_ub,
9563 
9564 			/* Output */
9565 			locals->DST_Y_PER_PTE_ROW_NOM_L,
9566 			locals->DST_Y_PER_PTE_ROW_NOM_C,
9567 			locals->DST_Y_PER_META_ROW_NOM_L,
9568 			locals->DST_Y_PER_META_ROW_NOM_C,
9569 			locals->TimePerMetaChunkNominal,
9570 			locals->TimePerChromaMetaChunkNominal,
9571 			locals->TimePerMetaChunkVBlank,
9572 			locals->TimePerChromaMetaChunkVBlank,
9573 			locals->TimePerMetaChunkFlip,
9574 			locals->TimePerChromaMetaChunkFlip,
9575 			locals->time_per_pte_group_nom_luma,
9576 			locals->time_per_pte_group_vblank_luma,
9577 			locals->time_per_pte_group_flip_luma,
9578 			locals->time_per_pte_group_nom_chroma,
9579 			locals->time_per_pte_group_vblank_chroma,
9580 			locals->time_per_pte_group_flip_chroma);
9581 
9582 	CalculateVMGroupAndRequestTimes(
9583 			mode_lib->ms.num_active_planes,
9584 			mode_lib->ms.cache_display_cfg.plane.GPUVMEnable,
9585 			mode_lib->ms.cache_display_cfg.plane.GPUVMMaxPageTableLevels,
9586 			mode_lib->ms.cache_display_cfg.timing.HTotal,
9587 			locals->BytePerPixelC,
9588 			locals->DestinationLinesToRequestVMInVBlank,
9589 			locals->DestinationLinesToRequestVMInImmediateFlip,
9590 			mode_lib->ms.cache_display_cfg.surface.DCCEnable,
9591 			mode_lib->ms.cache_display_cfg.timing.PixelClock,
9592 			locals->dpte_row_width_luma_ub,
9593 			locals->dpte_row_width_chroma_ub,
9594 			locals->vm_group_bytes,
9595 			locals->dpde0_bytes_per_frame_ub_l,
9596 			locals->dpde0_bytes_per_frame_ub_c,
9597 			locals->meta_pte_bytes_per_frame_ub_l,
9598 			locals->meta_pte_bytes_per_frame_ub_c,
9599 
9600 			/* Output */
9601 			locals->TimePerVMGroupVBlank,
9602 			locals->TimePerVMGroupFlip,
9603 			locals->TimePerVMRequestVBlank,
9604 			locals->TimePerVMRequestFlip);
9605 
9606 	// Min TTUVBlank
9607 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
9608 		if (locals->PrefetchMode[k] == 0) {
9609 			locals->MinTTUVBlank[k] = dml_max4(
9610 					locals->Watermark.DRAMClockChangeWatermark,
9611 					locals->Watermark.FCLKChangeWatermark,
9612 					locals->Watermark.StutterEnterPlusExitWatermark,
9613 					locals->Watermark.UrgentWatermark);
9614 		} else if (locals->PrefetchMode[k] == 1) {
9615 			locals->MinTTUVBlank[k] = dml_max3(
9616 					locals->Watermark.FCLKChangeWatermark,
9617 					locals->Watermark.StutterEnterPlusExitWatermark,
9618 					locals->Watermark.UrgentWatermark);
9619 		} else if (locals->PrefetchMode[k] == 2) {
9620 			locals->MinTTUVBlank[k] = dml_max(
9621 					locals->Watermark.StutterEnterPlusExitWatermark,
9622 					locals->Watermark.UrgentWatermark);
9623 		} else {
9624 			locals->MinTTUVBlank[k] = locals->Watermark.UrgentWatermark;
9625 		}
9626 		if (!mode_lib->ms.cache_display_cfg.plane.DynamicMetadataEnable[k])
9627 			locals->MinTTUVBlank[k] = locals->TCalc + locals->MinTTUVBlank[k];
9628 	}
9629 
9630 	// DCC Configuration
9631 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
9632 #ifdef __DML_VBA_DEBUG__
9633 		dml_print("DML::%s: Calculate DCC configuration for surface k=%u\n", __func__, k);
9634 #endif
9635 		CalculateDCCConfiguration(
9636 			mode_lib->ms.cache_display_cfg.surface.DCCEnable[k],
9637 			mode_lib->ms.policy.DCCProgrammingAssumesScanDirectionUnknownFinal,
9638 			mode_lib->ms.cache_display_cfg.surface.SourcePixelFormat[k],
9639 			mode_lib->ms.cache_display_cfg.surface.SurfaceWidthY[k],
9640 			mode_lib->ms.cache_display_cfg.surface.SurfaceWidthC[k],
9641 			mode_lib->ms.cache_display_cfg.surface.SurfaceHeightY[k],
9642 			mode_lib->ms.cache_display_cfg.surface.SurfaceHeightC[k],
9643 			mode_lib->ms.NomDETInKByte,
9644 			locals->BlockHeight256BytesY[k],
9645 			locals->BlockHeight256BytesC[k],
9646 			mode_lib->ms.cache_display_cfg.surface.SurfaceTiling[k],
9647 			locals->BytePerPixelY[k],
9648 			locals->BytePerPixelC[k],
9649 			locals->BytePerPixelDETY[k],
9650 			locals->BytePerPixelDETC[k],
9651 			mode_lib->ms.cache_display_cfg.plane.SourceScan[k],
9652 			/* Output */
9653 			&locals->DCCYMaxUncompressedBlock[k],
9654 			&locals->DCCCMaxUncompressedBlock[k],
9655 			&locals->DCCYMaxCompressedBlock[k],
9656 			&locals->DCCCMaxCompressedBlock[k],
9657 			&locals->DCCYIndependentBlock[k],
9658 			&locals->DCCCIndependentBlock[k]);
9659 	}
9660 
9661 	// VStartup Adjustment
9662 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
9663 		s->Tvstartup_margin = (s->MaxVStartupLines[k] - locals->VStartupMin[k]) * mode_lib->ms.cache_display_cfg.timing.HTotal[k] / mode_lib->ms.cache_display_cfg.timing.PixelClock[k];
9664 #ifdef __DML_VBA_DEBUG__
9665 		dml_print("DML::%s: k=%u, MinTTUVBlank = %f (before vstartup margin)\n", __func__, k, locals->MinTTUVBlank[k]);
9666 #endif
9667 
9668 		locals->MinTTUVBlank[k] = locals->MinTTUVBlank[k] + s->Tvstartup_margin;
9669 
9670 #ifdef __DML_VBA_DEBUG__
9671 		dml_print("DML::%s: k=%u, Tvstartup_margin = %f\n", __func__, k, s->Tvstartup_margin);
9672 		dml_print("DML::%s: k=%u, MaxVStartupLines = %u\n", __func__, k, s->MaxVStartupLines[k]);
9673 		dml_print("DML::%s: k=%u, MinTTUVBlank = %f\n", __func__, k, locals->MinTTUVBlank[k]);
9674 #endif
9675 
9676 		locals->Tdmdl[k] = locals->Tdmdl[k] + s->Tvstartup_margin;
9677 		if (mode_lib->ms.cache_display_cfg.plane.DynamicMetadataEnable[k] && mode_lib->ms.ip.dynamic_metadata_vm_enabled) {
9678 			locals->Tdmdl_vm[k] = locals->Tdmdl_vm[k] + s->Tvstartup_margin;
9679 		}
9680 
9681 		isInterlaceTiming        = (mode_lib->ms.cache_display_cfg.timing.Interlace[k] && !mode_lib->ms.ip.ptoi_supported);
9682 
9683 		// The actual positioning of the vstartup
9684 		locals->VStartup[k] = (isInterlaceTiming ? (2 * s->MaxVStartupLines[k]) : s->MaxVStartupLines[k]);
9685 
9686 		s->dlg_vblank_start        =  ((isInterlaceTiming ? dml_floor((mode_lib->ms.cache_display_cfg.timing.VTotal[k] - mode_lib->ms.cache_display_cfg.timing.VFrontPorch[k]) / 2.0, 1.0) :
9687 																	mode_lib->ms.cache_display_cfg.timing.VTotal[k]) - mode_lib->ms.cache_display_cfg.timing.VFrontPorch[k]);
9688 		s->LSetup                  = dml_floor(4.0 * locals->TSetup[k] / ((dml_float_t) mode_lib->ms.cache_display_cfg.timing.HTotal[k] / mode_lib->ms.cache_display_cfg.timing.PixelClock[k]), 1.0) / 4.0;
9689 		s->blank_lines_remaining   = (mode_lib->ms.cache_display_cfg.timing.VTotal[k] - mode_lib->ms.cache_display_cfg.timing.VActive[k]) - locals->VStartup[k];
9690 
9691 		if (s->blank_lines_remaining < 0) {
9692 			dml_print("ERROR: Vstartup is larger than vblank!?\n");
9693 			s->blank_lines_remaining = 0;
9694 			ASSERT(0);
9695 		}
9696 		locals->MIN_DST_Y_NEXT_START[k] = s->dlg_vblank_start + s->blank_lines_remaining + s->LSetup;
9697 
9698 		// debug only
9699 		s->old_MIN_DST_Y_NEXT_START = ((isInterlaceTiming ? dml_floor((mode_lib->ms.cache_display_cfg.timing.VTotal[k] - mode_lib->ms.cache_display_cfg.timing.VFrontPorch[k]) / 2.0, 1.0) :
9700 																	mode_lib->ms.cache_display_cfg.timing.VTotal[k]) - mode_lib->ms.cache_display_cfg.timing.VFrontPorch[k])
9701 											+ dml_max(1.0, dml_ceil((dml_float_t) locals->WritebackDelay[k] / ((dml_float_t) mode_lib->ms.cache_display_cfg.timing.HTotal[k] / mode_lib->ms.cache_display_cfg.timing.PixelClock[k]), 1.0))
9702 											+ dml_floor(4.0 * locals->TSetup[k] / ((dml_float_t) mode_lib->ms.cache_display_cfg.timing.HTotal[k] / mode_lib->ms.cache_display_cfg.timing.PixelClock[k]), 1.0) / 4.0;
9703 
9704 		if (((locals->VUpdateOffsetPix[k] + locals->VUpdateWidthPix[k] + locals->VReadyOffsetPix[k]) / (double) mode_lib->ms.cache_display_cfg.timing.HTotal[k]) <=
9705 			(isInterlaceTiming ?
9706 				dml_floor((mode_lib->ms.cache_display_cfg.timing.VTotal[k] - mode_lib->ms.cache_display_cfg.timing.VActive[k] - mode_lib->ms.cache_display_cfg.timing.VFrontPorch[k] - locals->VStartup[k]) / 2.0, 1.0) :
9707 				(int) (mode_lib->ms.cache_display_cfg.timing.VTotal[k] - mode_lib->ms.cache_display_cfg.timing.VActive[k] - mode_lib->ms.cache_display_cfg.timing.VFrontPorch[k] - locals->VStartup[k]))) {
9708 			locals->VREADY_AT_OR_AFTER_VSYNC[k] = true;
9709 		} else {
9710 			locals->VREADY_AT_OR_AFTER_VSYNC[k] = false;
9711 		}
9712 #ifdef __DML_VBA_DEBUG__
9713 		dml_print("DML::%s: k=%u, VStartup = %u (max)\n", __func__, k, locals->VStartup[k]);
9714 		dml_print("DML::%s: k=%u, VStartupMin = %u (max)\n", __func__, k, locals->VStartupMin[k]);
9715 		dml_print("DML::%s: k=%u, VUpdateOffsetPix = %u\n", __func__, k, locals->VUpdateOffsetPix[k]);
9716 		dml_print("DML::%s: k=%u, VUpdateWidthPix = %u\n", __func__, k, locals->VUpdateWidthPix[k]);
9717 		dml_print("DML::%s: k=%u, VReadyOffsetPix = %u\n", __func__, k, locals->VReadyOffsetPix[k]);
9718 		dml_print("DML::%s: k=%u, HTotal = %u\n", __func__, k, mode_lib->ms.cache_display_cfg.timing.HTotal[k]);
9719 		dml_print("DML::%s: k=%u, VTotal = %u\n", __func__, k, mode_lib->ms.cache_display_cfg.timing.VTotal[k]);
9720 		dml_print("DML::%s: k=%u, VActive = %u\n", __func__, k, mode_lib->ms.cache_display_cfg.timing.VActive[k]);
9721 		dml_print("DML::%s: k=%u, VFrontPorch = %u\n", __func__, k, mode_lib->ms.cache_display_cfg.timing.VFrontPorch[k]);
9722 		dml_print("DML::%s: k=%u, TSetup = %f\n", __func__, k, locals->TSetup[k]);
9723 		dml_print("DML::%s: k=%u, MIN_DST_Y_NEXT_START = %f\n", __func__, k, locals->MIN_DST_Y_NEXT_START[k]);
9724 		dml_print("DML::%s: k=%u, MIN_DST_Y_NEXT_START = %f (old)\n", __func__, k, s->old_MIN_DST_Y_NEXT_START);
9725 		dml_print("DML::%s: k=%u, VREADY_AT_OR_AFTER_VSYNC = %u\n", __func__, k, locals->VREADY_AT_OR_AFTER_VSYNC[k]);
9726 #endif
9727 	}
9728 
9729 	//Maximum Bandwidth Used
9730 	s->TotalWRBandwidth = 0;
9731 	s->WRBandwidth = 0;
9732 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
9733 		if (mode_lib->ms.cache_display_cfg.writeback.WritebackEnable[k] == true && mode_lib->ms.cache_display_cfg.writeback.WritebackPixelFormat[k] == dml_444_32) {
9734 			s->WRBandwidth = mode_lib->ms.cache_display_cfg.writeback.WritebackDestinationWidth[k] * mode_lib->ms.cache_display_cfg.writeback.WritebackDestinationHeight[k] /
9735 							(mode_lib->ms.cache_display_cfg.timing.HTotal[k] * mode_lib->ms.cache_display_cfg.writeback.WritebackSourceHeight[k] / mode_lib->ms.cache_display_cfg.timing.PixelClock[k]) * 4;
9736 		} else if (mode_lib->ms.cache_display_cfg.writeback.WritebackEnable[k] == true) {
9737 			s->WRBandwidth = mode_lib->ms.cache_display_cfg.writeback.WritebackDestinationWidth[k] * mode_lib->ms.cache_display_cfg.writeback.WritebackDestinationHeight[k] /
9738 							(mode_lib->ms.cache_display_cfg.timing.HTotal[k] * mode_lib->ms.cache_display_cfg.writeback.WritebackSourceHeight[k] / mode_lib->ms.cache_display_cfg.timing.PixelClock[k]) * 8;
9739 		}
9740 		s->TotalWRBandwidth = s->TotalWRBandwidth + s->WRBandwidth;
9741 	}
9742 
9743 	locals->TotalDataReadBandwidth = 0;
9744 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
9745 		locals->TotalDataReadBandwidth = locals->TotalDataReadBandwidth + locals->ReadBandwidthSurfaceLuma[k] + locals->ReadBandwidthSurfaceChroma[k];
9746 
9747 #ifdef __DML_VBA_DEBUG__
9748 		dml_print("DML::%s: k=%u, TotalDataReadBandwidth = %f\n", __func__, k, locals->TotalDataReadBandwidth);
9749 		dml_print("DML::%s: k=%u, ReadBandwidthSurfaceLuma = %f\n", __func__, k, locals->ReadBandwidthSurfaceLuma[k]);
9750 		dml_print("DML::%s: k=%u, ReadBandwidthSurfaceChroma = %f\n", __func__, k, locals->ReadBandwidthSurfaceChroma[k]);
9751 #endif
9752 	}
9753 
9754 	locals->TotalDataReadBandwidthNotIncludingMALLPrefetch = 0;
9755 	for (k = 0; k < mode_lib->ms.num_active_planes; ++k) {
9756 		if (mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange[k] != dml_use_mall_pstate_change_phantom_pipe) {
9757 			locals->TotalDataReadBandwidthNotIncludingMALLPrefetch = locals->TotalDataReadBandwidthNotIncludingMALLPrefetch
9758 		+ locals->ReadBandwidthSurfaceLuma[k] + locals->ReadBandwidthSurfaceChroma[k];
9759 		}
9760 	}
9761 
9762 	CalculateStutterEfficiency_params->CompressedBufferSizeInkByte = locals->CompressedBufferSizeInkByte;
9763 	CalculateStutterEfficiency_params->UseMALLForPStateChange = mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange;
9764 	CalculateStutterEfficiency_params->UnboundedRequestEnabled = locals->UnboundedRequestEnabled;
9765 	CalculateStutterEfficiency_params->MetaFIFOSizeInKEntries = mode_lib->ms.ip.meta_fifo_size_in_kentries;
9766 	CalculateStutterEfficiency_params->ZeroSizeBufferEntries = mode_lib->ms.ip.zero_size_buffer_entries;
9767 	CalculateStutterEfficiency_params->PixelChunkSizeInKByte = mode_lib->ms.ip.pixel_chunk_size_kbytes;
9768 	CalculateStutterEfficiency_params->NumberOfActiveSurfaces = mode_lib->ms.num_active_planes;
9769 	CalculateStutterEfficiency_params->ROBBufferSizeInKByte = mode_lib->ms.ip.rob_buffer_size_kbytes;
9770 	CalculateStutterEfficiency_params->TotalDataReadBandwidth = locals->TotalDataReadBandwidth;
9771 	CalculateStutterEfficiency_params->DCFCLK = locals->Dcfclk;
9772 	CalculateStutterEfficiency_params->ReturnBW = mode_lib->ms.ReturnBW;
9773 	CalculateStutterEfficiency_params->CompbufReservedSpace64B = locals->compbuf_reserved_space_64b;
9774 	CalculateStutterEfficiency_params->CompbufReservedSpaceZs = locals->compbuf_reserved_space_zs;
9775 	CalculateStutterEfficiency_params->SRExitTime = mode_lib->ms.state.sr_exit_time_us;
9776 	CalculateStutterEfficiency_params->SRExitZ8Time = mode_lib->ms.state.sr_exit_z8_time_us;
9777 	CalculateStutterEfficiency_params->SynchronizeTimingsFinal = mode_lib->ms.policy.SynchronizeTimingsFinal;
9778 	CalculateStutterEfficiency_params->BlendingAndTiming = mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming;
9779 	CalculateStutterEfficiency_params->StutterEnterPlusExitWatermark = locals->Watermark.StutterEnterPlusExitWatermark;
9780 	CalculateStutterEfficiency_params->Z8StutterEnterPlusExitWatermark = locals->Watermark.Z8StutterEnterPlusExitWatermark;
9781 	CalculateStutterEfficiency_params->ProgressiveToInterlaceUnitInOPP = mode_lib->ms.ip.ptoi_supported;
9782 	CalculateStutterEfficiency_params->Interlace = mode_lib->ms.cache_display_cfg.timing.Interlace;
9783 	CalculateStutterEfficiency_params->MinTTUVBlank = locals->MinTTUVBlank;
9784 	CalculateStutterEfficiency_params->DPPPerSurface = mode_lib->ms.cache_display_cfg.hw.DPPPerSurface;
9785 	CalculateStutterEfficiency_params->DETBufferSizeY = locals->DETBufferSizeY;
9786 	CalculateStutterEfficiency_params->BytePerPixelY = locals->BytePerPixelY;
9787 	CalculateStutterEfficiency_params->BytePerPixelDETY = locals->BytePerPixelDETY;
9788 	CalculateStutterEfficiency_params->SwathWidthY = locals->SwathWidthY;
9789 	CalculateStutterEfficiency_params->SwathHeightY = locals->SwathHeightY;
9790 	CalculateStutterEfficiency_params->SwathHeightC = locals->SwathHeightC;
9791 	CalculateStutterEfficiency_params->NetDCCRateLuma = mode_lib->ms.cache_display_cfg.surface.DCCRateLuma;
9792 	CalculateStutterEfficiency_params->NetDCCRateChroma = mode_lib->ms.cache_display_cfg.surface.DCCRateChroma;
9793 	CalculateStutterEfficiency_params->DCCFractionOfZeroSizeRequestsLuma = mode_lib->ms.cache_display_cfg.surface.DCCFractionOfZeroSizeRequestsLuma;
9794 	CalculateStutterEfficiency_params->DCCFractionOfZeroSizeRequestsChroma = mode_lib->ms.cache_display_cfg.surface.DCCFractionOfZeroSizeRequestsChroma;
9795 	CalculateStutterEfficiency_params->HTotal = mode_lib->ms.cache_display_cfg.timing.HTotal;
9796 	CalculateStutterEfficiency_params->VTotal = mode_lib->ms.cache_display_cfg.timing.VTotal;
9797 	CalculateStutterEfficiency_params->PixelClock = mode_lib->ms.cache_display_cfg.timing.PixelClock;
9798 	CalculateStutterEfficiency_params->VRatio = mode_lib->ms.cache_display_cfg.plane.VRatio;
9799 	CalculateStutterEfficiency_params->SourceScan = mode_lib->ms.cache_display_cfg.plane.SourceScan;
9800 	CalculateStutterEfficiency_params->BlockHeight256BytesY = locals->BlockHeight256BytesY;
9801 	CalculateStutterEfficiency_params->BlockWidth256BytesY = locals->BlockWidth256BytesY;
9802 	CalculateStutterEfficiency_params->BlockHeight256BytesC = locals->BlockHeight256BytesC;
9803 	CalculateStutterEfficiency_params->BlockWidth256BytesC = locals->BlockWidth256BytesC;
9804 	CalculateStutterEfficiency_params->DCCYMaxUncompressedBlock = locals->DCCYMaxUncompressedBlock;
9805 	CalculateStutterEfficiency_params->DCCCMaxUncompressedBlock = locals->DCCCMaxUncompressedBlock;
9806 	CalculateStutterEfficiency_params->VActive = mode_lib->ms.cache_display_cfg.timing.VActive;
9807 	CalculateStutterEfficiency_params->DCCEnable = mode_lib->ms.cache_display_cfg.surface.DCCEnable;
9808 	CalculateStutterEfficiency_params->WritebackEnable = mode_lib->ms.cache_display_cfg.writeback.WritebackEnable;
9809 	CalculateStutterEfficiency_params->ReadBandwidthSurfaceLuma = locals->ReadBandwidthSurfaceLuma;
9810 	CalculateStutterEfficiency_params->ReadBandwidthSurfaceChroma = locals->ReadBandwidthSurfaceChroma;
9811 	CalculateStutterEfficiency_params->meta_row_bw = locals->meta_row_bw;
9812 	CalculateStutterEfficiency_params->dpte_row_bw = locals->dpte_row_bw;
9813 	CalculateStutterEfficiency_params->StutterEfficiencyNotIncludingVBlank = &locals->StutterEfficiencyNotIncludingVBlank;
9814 	CalculateStutterEfficiency_params->StutterEfficiency = &locals->StutterEfficiency;
9815 	CalculateStutterEfficiency_params->NumberOfStutterBurstsPerFrame = &locals->NumberOfStutterBurstsPerFrame;
9816 	CalculateStutterEfficiency_params->Z8StutterEfficiencyNotIncludingVBlank = &locals->Z8StutterEfficiencyNotIncludingVBlank;
9817 	CalculateStutterEfficiency_params->Z8StutterEfficiency = &locals->Z8StutterEfficiency;
9818 	CalculateStutterEfficiency_params->Z8NumberOfStutterBurstsPerFrame = &locals->Z8NumberOfStutterBurstsPerFrame;
9819 	CalculateStutterEfficiency_params->StutterPeriod = &locals->StutterPeriod;
9820 	CalculateStutterEfficiency_params->DCHUBBUB_ARB_CSTATE_MAX_CAP_MODE = &locals->DCHUBBUB_ARB_CSTATE_MAX_CAP_MODE;
9821 
9822 	// Stutter Efficiency
9823 	CalculateStutterEfficiency(&mode_lib->scratch,
9824 		CalculateStutterEfficiency_params);
9825 
9826 #ifdef __DML_VBA_ALLOW_DELTA__
9827 	{
9828 	dml_float_t dummy_single[2];
9829 	dml_uint_t dummy_integer[1];
9830 	dml_bool_t dummy_boolean[1];
9831 
9832 	// Calculate z8 stutter eff assuming 0 reserved space
9833 	CalculateStutterEfficiency(
9834 			locals->CompressedBufferSizeInkByte,
9835 			mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange,
9836 			locals->UnboundedRequestEnabled,
9837 			mode_lib->ms.ip.meta_fifo_size_in_kentries,
9838 			mode_lib->ms.ip.zero_size_buffer_entries,
9839 			mode_lib->ms.ip.pixel_chunk_size_kbytes,
9840 			mode_lib->ms.num_active_planes,
9841 			mode_lib->ms.ip.rob_buffer_size_kbytes,
9842 			locals->TotalDataReadBandwidth,
9843 			locals->Dcfclk,
9844 			mode_lib->ms.ReturnBW,
9845 			0, //mode_lib->ms.ip.compbuf_reserved_space_64b,
9846 			0, //mode_lib->ms.ip.compbuf_reserved_space_zs,
9847 			mode_lib->ms.state.sr_exit_time_us,
9848 			mode_lib->ms.state.sr_exit_z8_time_us,
9849 			mode_lib->ms.policy.SynchronizeTimingsFinal,
9850 			mode_lib->ms.cache_display_cfg.plane.BlendingAndTiming,
9851 			locals->Watermark.StutterEnterPlusExitWatermark,
9852 			locals->Watermark.Z8StutterEnterPlusExitWatermark,
9853 			mode_lib->ms.ip.ptoi_supported,
9854 			mode_lib->ms.cache_display_cfg.timing.Interlace,
9855 			locals->MinTTUVBlank,
9856 			mode_lib->ms.cache_display_cfg.hw.DPPPerSurface,
9857 			mode_lib->ms.DETBufferSizeY,
9858 			locals->BytePerPixelY,
9859 			locals->BytePerPixelDETY,
9860 			locals->SwathWidthY,
9861 			mode_lib->ms.SwathHeightY,
9862 			mode_lib->ms.SwathHeightC,
9863 			mode_lib->ms.cache_display_cfg.surface.DCCRateLuma,
9864 			mode_lib->ms.cache_display_cfg.surface.DCCRateChroma,
9865 			mode_lib->ms.cache_display_cfg.surface.DCCFractionOfZeroSizeRequestsLuma,
9866 			mode_lib->ms.cache_display_cfg.surface.DCCFractionOfZeroSizeRequestsChroma,
9867 			mode_lib->ms.cache_display_cfg.timing.HTotal,
9868 			mode_lib->ms.cache_display_cfg.timing.VTotal,
9869 			mode_lib->ms.cache_display_cfg.timing.PixelClock,
9870 			mode_lib->ms.cache_display_cfg.plane.VRatio,
9871 			mode_lib->ms.cache_display_cfg.plane.SourceScan,
9872 			locals->BlockHeight256BytesY,
9873 			locals->BlockWidth256BytesY,
9874 			locals->BlockHeight256BytesC,
9875 			locals->BlockWidth256BytesC,
9876 			locals->DCCYMaxUncompressedBlock,
9877 			locals->DCCCMaxUncompressedBlock,
9878 			mode_lib->ms.cache_display_cfg.timing.VActive,
9879 			mode_lib->ms.cache_display_cfg.surface.DCCEnable,
9880 			mode_lib->ms.cache_display_cfg.writeback.WritebackEnable,
9881 			locals->ReadBandwidthSurfaceLuma,
9882 			locals->ReadBandwidthSurfaceChroma,
9883 			locals->meta_row_bw,
9884 			locals->dpte_row_bw,
9885 
9886 			/* Output */
9887 			&dummy_single[0],
9888 			&dummy_single[1],
9889 			&dummy_integer[0],
9890 			&locals->Z8StutterEfficiencyNotIncludingVBlankBestCase,
9891 			&locals->Z8StutterEfficiencyBestCase,
9892 			&locals->Z8NumberOfStutterBurstsPerFrameBestCase,
9893 			&locals->StutterPeriodBestCase,
9894 			&dummy_boolean[0]);
9895 	}
9896 #else
9897 	locals->Z8StutterEfficiencyNotIncludingVBlankBestCase = locals->Z8StutterEfficiencyNotIncludingVBlank;
9898 	locals->Z8StutterEfficiencyBestCase                   = locals->Z8StutterEfficiency;
9899 	locals->Z8NumberOfStutterBurstsPerFrameBestCase       = locals->Z8NumberOfStutterBurstsPerFrame;
9900 	locals->StutterPeriodBestCase                         = locals->StutterPeriod;
9901 #endif
9902 
9903 #ifdef __DML_VBA_DEBUG__
9904 	dml_print("DML::%s: --- END --- \n",  __func__);
9905 #endif
9906 } // dml_core_mode_programming
9907 
9908 /// Function: dml_core_get_row_heights
9909 /// @brief Get row height for DPTE and META with minimal input.
dml_core_get_row_heights(dml_uint_t * dpte_row_height,dml_uint_t * meta_row_height,const struct display_mode_lib_st * mode_lib,dml_bool_t is_plane1,enum dml_source_format_class SourcePixelFormat,enum dml_swizzle_mode SurfaceTiling,enum dml_rotation_angle ScanDirection,dml_uint_t pitch,dml_uint_t GPUVMMinPageSizeKBytes)9910 void dml_core_get_row_heights(
9911 						dml_uint_t                         *dpte_row_height,
9912 						dml_uint_t                         *meta_row_height,
9913 						const struct display_mode_lib_st   *mode_lib,
9914 						dml_bool_t                         is_plane1,
9915 						enum dml_source_format_class        SourcePixelFormat,
9916 						enum dml_swizzle_mode            SurfaceTiling,
9917 						enum dml_rotation_angle          ScanDirection,
9918 						dml_uint_t                         pitch,
9919 						dml_uint_t                         GPUVMMinPageSizeKBytes)
9920 {
9921 	dml_uint_t BytePerPixelY;
9922 	dml_uint_t BytePerPixelC;
9923 	dml_float_t BytePerPixelInDETY;
9924 	dml_float_t BytePerPixelInDETC;
9925 	dml_uint_t BlockHeight256BytesY;
9926 	dml_uint_t BlockHeight256BytesC;
9927 	dml_uint_t BlockWidth256BytesY;
9928 	dml_uint_t BlockWidth256BytesC;
9929 	dml_uint_t MacroTileWidthY;
9930 	dml_uint_t MacroTileWidthC;
9931 	dml_uint_t MacroTileHeightY;
9932 	dml_uint_t MacroTileHeightC;
9933 
9934 	dml_uint_t BytePerPixel;
9935 	dml_uint_t BlockHeight256Bytes;
9936 	dml_uint_t BlockWidth256Bytes;
9937 	dml_uint_t MacroTileWidth;
9938 	dml_uint_t MacroTileHeight;
9939 	dml_uint_t PTEBufferSizeInRequests;
9940 
9941 	dml_uint_t dummy_integer[16];
9942 
9943 	CalculateBytePerPixelAndBlockSizes(
9944 		SourcePixelFormat,
9945 		SurfaceTiling,
9946 
9947 		/* Output */
9948 		&BytePerPixelY,
9949 		&BytePerPixelC,
9950 		&BytePerPixelInDETY,
9951 		&BytePerPixelInDETC,
9952 		&BlockHeight256BytesY,
9953 		&BlockHeight256BytesC,
9954 		&BlockWidth256BytesY,
9955 		&BlockWidth256BytesC,
9956 		&MacroTileHeightY,
9957 		&MacroTileHeightC,
9958 		&MacroTileWidthY,
9959 		&MacroTileWidthC);
9960 
9961 	BytePerPixel = is_plane1 ? BytePerPixelC : BytePerPixelY;
9962 	BlockHeight256Bytes = is_plane1 ? BlockHeight256BytesC : BlockHeight256BytesY;
9963 	BlockWidth256Bytes = is_plane1 ? BlockWidth256BytesC : BlockWidth256BytesY;
9964 	MacroTileWidth = is_plane1 ? MacroTileWidthC : MacroTileWidthY;
9965 	MacroTileHeight = is_plane1 ? MacroTileHeightC : MacroTileHeightY;
9966 	PTEBufferSizeInRequests = is_plane1 ? mode_lib->ip.dpte_buffer_size_in_pte_reqs_chroma : mode_lib->ip.dpte_buffer_size_in_pte_reqs_luma;
9967 #ifdef __DML_RQ_DLG_CALC_DEBUG__
9968 	dml_print("DML_DLG: %s: is_plane1 = %u\n", __func__, is_plane1);
9969 	dml_print("DML_DLG: %s: BytePerPixel = %u\n", __func__, BytePerPixel);
9970 	dml_print("DML_DLG: %s: BlockHeight256Bytes = %u\n", __func__, BlockHeight256Bytes);
9971 	dml_print("DML_DLG: %s: BlockWidth256Bytes = %u\n", __func__, BlockWidth256Bytes);
9972 	dml_print("DML_DLG: %s: MacroTileWidth = %u\n", __func__, MacroTileWidth);
9973 	dml_print("DML_DLG: %s: MacroTileHeight = %u\n", __func__, MacroTileHeight);
9974 	dml_print("DML_DLG: %s: PTEBufferSizeInRequests = %u\n", __func__, PTEBufferSizeInRequests);
9975 	dml_print("DML_DLG: %s: dpte_buffer_size_in_pte_reqs_luma = %u\n", __func__, mode_lib->ip.dpte_buffer_size_in_pte_reqs_luma);
9976 	dml_print("DML_DLG: %s: dpte_buffer_size_in_pte_reqs_chroma = %u\n", __func__, mode_lib->ip.dpte_buffer_size_in_pte_reqs_chroma);
9977 	dml_print("DML_DLG: %s: GPUVMMinPageSizeKBytes = %u\n", __func__, GPUVMMinPageSizeKBytes);
9978 #endif
9979 
9980 	// just supply with enough parameters to calculate meta and dte
9981 	CalculateVMAndRowBytes(
9982 			0, // dml_bool_t ViewportStationary,
9983 			1, // dml_bool_t DCCEnable,
9984 			1, // dml_uint_t NumberOfDPPs,
9985 			BlockHeight256Bytes,
9986 			BlockWidth256Bytes,
9987 			SourcePixelFormat,
9988 			SurfaceTiling,
9989 			BytePerPixel,
9990 			ScanDirection,
9991 			0, // dml_uint_t SwathWidth,
9992 			0, // dml_uint_t ViewportHeight, (Note: DML calculates one_row_for_frame height regardless, would need test input if that height is useful)
9993 			0, // dml_uint_t ViewportXStart,
9994 			0, // dml_uint_t ViewportYStart,
9995 			1, // dml_bool_t GPUVMEnable,
9996 			4, // dml_uint_t GPUVMMaxPageTableLevels,
9997 			GPUVMMinPageSizeKBytes,
9998 			PTEBufferSizeInRequests,
9999 			pitch,
10000 			0, // dml_uint_t DCCMetaPitch,
10001 			MacroTileWidth,
10002 			MacroTileHeight,
10003 
10004 			// /* Output */
10005 			&dummy_integer[0], // dml_uint_t *MetaRowByte,
10006 			&dummy_integer[1], // dml_uint_t *PixelPTEBytesPerRow,
10007 			&dummy_integer[2], // dml_uint_t *PixelPTEBytesPerRowStorage,
10008 			&dummy_integer[3], // dml_uint_t *dpte_row_width_ub,
10009 			dpte_row_height,
10010 			&dummy_integer[4], // dml_uint_t *dpte_row_height_linear
10011 			&dummy_integer[5], // dml_uint_t *PixelPTEBytesPerRow_one_row_per_frame,
10012 			&dummy_integer[6], // dml_uint_t *dpte_row_width_ub_one_row_per_frame,
10013 			&dummy_integer[7], // dml_uint_t *dpte_row_height_one_row_per_frame,
10014 			&dummy_integer[8], // dml_uint_t *MetaRequestWidth,
10015 			&dummy_integer[9], // dml_uint_t *MetaRequestHeight,
10016 			&dummy_integer[10], // dml_uint_t *meta_row_width,
10017 			meta_row_height,
10018 			&dummy_integer[11], // dml_uint_t *PixelPTEReqWidth,
10019 			&dummy_integer[12], // dml_uint_t *PixelPTEReqHeight,
10020 			&dummy_integer[13], // dml_uint_t *PTERequestSize,
10021 			&dummy_integer[14], // dml_uint_t *DPDE0BytesFrame,
10022 			&dummy_integer[15]); // dml_uint_t *MetaPTEBytesFrame)
10023 
10024 #ifdef __DML_RQ_DLG_CALC_DEBUG__
10025 	dml_print("DML_DLG: %s: dpte_row_height = %u\n", __func__, *dpte_row_height);
10026 	dml_print("DML_DLG: %s: meta_row_height = %u\n", __func__, *meta_row_height);
10027 #endif
10028 }
10029 
dml_get_soc_state_bounding_box(const struct soc_states_st * states,dml_uint_t state_idx)10030 static struct soc_state_bounding_box_st dml_get_soc_state_bounding_box(
10031 	const struct soc_states_st *states,
10032 	dml_uint_t state_idx)
10033 {
10034 	dml_print("DML::%s: state_idx=%u (num_states=%u)\n", __func__, state_idx, states->num_states);
10035 
10036 	if (state_idx >= (dml_uint_t)states->num_states) {
10037 		dml_print("DML::%s: ERROR: Invalid state_idx=%u! num_states=%u\n", __func__, state_idx, states->num_states);
10038 		ASSERT(0);
10039 	}
10040 	return (states->state_array[state_idx]);
10041 }
10042 
10043 /// @brief Copy the parameters to a calculation struct, it actually only need when the DML needs to have
10044 ///        the intelligence to re-calculate when any of display cfg, bbox, or policy changes since last calculated.
10045 ///
cache_ip_soc_cfg(struct display_mode_lib_st * mode_lib,dml_uint_t state_idx)10046 static void cache_ip_soc_cfg(struct display_mode_lib_st *mode_lib,
10047 						dml_uint_t state_idx)
10048 {
10049 	mode_lib->ms.state_idx = state_idx;
10050 	mode_lib->ms.max_state_idx = mode_lib->states.num_states - 1;
10051 	mode_lib->ms.soc = mode_lib->soc;
10052 	mode_lib->ms.ip = mode_lib->ip;
10053 	mode_lib->ms.policy = mode_lib->policy;
10054 	mode_lib->ms.state = dml_get_soc_state_bounding_box(&mode_lib->states, state_idx);
10055 	mode_lib->ms.max_state = dml_get_soc_state_bounding_box(&mode_lib->states, mode_lib->states.num_states - 1);
10056 }
10057 
cache_display_cfg(struct display_mode_lib_st * mode_lib,const struct dml_display_cfg_st * display_cfg)10058 static void cache_display_cfg(struct display_mode_lib_st *mode_lib,
10059 	const struct dml_display_cfg_st *display_cfg)
10060 {
10061 	mode_lib->ms.cache_display_cfg = *display_cfg;
10062 }
10063 
fetch_socbb_params(struct display_mode_lib_st * mode_lib)10064 static void fetch_socbb_params(struct display_mode_lib_st *mode_lib)
10065 {
10066 	struct soc_state_bounding_box_st *state = &mode_lib->ms.state;
10067 
10068 	// Default values, SOCCLK, DRAMSpeed, and FabricClock will be reassigned to the same state value in mode_check step
10069 	// If UseMinimumRequiredDCFCLK is used, the DCFCLK will be the min dcflk for the mode support
10070 	mode_lib->ms.SOCCLK = (dml_float_t)state->socclk_mhz;
10071 	mode_lib->ms.DRAMSpeed = (dml_float_t)state->dram_speed_mts;
10072 	mode_lib->ms.FabricClock = (dml_float_t)state->fabricclk_mhz;
10073 	mode_lib->ms.DCFCLK = (dml_float_t)state->dcfclk_mhz;
10074 }
10075 
10076 /// @brief Use display_cfg directly for mode_support calculation
10077 ///        Calculated values and informational output are stored in mode_lib.vba data struct
10078 ///        The display configuration is described with pipes struct and num_pipes
10079 ///        This function is used when physical resource mapping is not finalized (for example,
10080 ///        don't know how many pipes to represent a surface)
10081 /// @param mode_lib Contains the bounding box and policy setting.
10082 /// @param state_idx Power state index
10083 /// @param display_cfg Display configurations. A display
dml_mode_support(struct display_mode_lib_st * mode_lib,dml_uint_t state_idx,const struct dml_display_cfg_st * display_cfg)10084 dml_bool_t dml_mode_support(
10085 	struct display_mode_lib_st *mode_lib,
10086 	dml_uint_t                        state_idx,
10087 	const struct dml_display_cfg_st *display_cfg)
10088 {
10089 	dml_bool_t is_mode_support;
10090 
10091 	dml_print("DML::%s: ------------- START ----------\n", __func__);
10092 	cache_ip_soc_cfg(mode_lib, state_idx);
10093 	cache_display_cfg(mode_lib, display_cfg);
10094 
10095 	fetch_socbb_params(mode_lib);
10096 
10097 	dml_print("DML::%s: state_idx          = %u\n", __func__, state_idx);
10098 
10099 	is_mode_support = dml_core_mode_support(mode_lib);
10100 
10101 	dml_print("DML::%s: is_mode_support = %u\n", __func__, is_mode_support);
10102 	dml_print("DML::%s: ------------- DONE ----------\n", __func__);
10103 	return is_mode_support;
10104 }
10105 
10106 /// @Brief A function to calculate the programming values for DCN DCHUB (Assume mode is supported)
10107 /// The output will be stored in the mode_lib.mp (mode_program_st) data struct and those can be accessed via the getter functions
10108 /// Calculated values include: watermarks, dlg, rq reg, different clock frequency
10109 /// This function returns 1 when there is no error.
10110 /// Note: In this function, it is assumed that DCFCLK, SOCCLK freq are the state values, and mode_program will just use the DML calculated DPPCLK and DISPCLK
10111 /// @param mode_lib mode_lib data struct that house all the input/output/bbox and calculation values.
10112 /// @param state_idx Power state idx chosen
10113 /// @param display_cfg Display Configuration
10114 /// @param call_standalone Calling mode_programming without calling mode support.  Some of the "support" struct member will be pre-calculated before doing mode programming
10115 /// TODO: Add clk_cfg input, could be useful for standalone mode
dml_mode_programming(struct display_mode_lib_st * mode_lib,dml_uint_t state_idx,const struct dml_display_cfg_st * display_cfg,bool call_standalone)10116 dml_bool_t dml_mode_programming(
10117 	struct display_mode_lib_st *mode_lib,
10118 	dml_uint_t                         state_idx,
10119 	const struct dml_display_cfg_st *display_cfg,
10120 	bool                               call_standalone)
10121 {
10122 	struct dml_clk_cfg_st clk_cfg;
10123 	memset(&clk_cfg, 0, sizeof(clk_cfg));
10124 
10125 	clk_cfg.dcfclk_option = dml_use_required_freq;
10126 	clk_cfg.dispclk_option = dml_use_required_freq;
10127 	for (dml_uint_t k = 0; k < __DML_NUM_PLANES__; ++k)
10128 		clk_cfg.dppclk_option[k] = dml_use_required_freq;
10129 
10130 	dml_print("DML::%s: ------------- START ----------\n", __func__);
10131 	dml_print("DML::%s: state_idx       = %u\n", __func__, state_idx);
10132 	dml_print("DML::%s: call_standalone = %u\n", __func__, call_standalone);
10133 
10134 	cache_ip_soc_cfg(mode_lib, state_idx);
10135 	cache_display_cfg(mode_lib, display_cfg);
10136 
10137 	fetch_socbb_params(mode_lib);
10138 	if (call_standalone) {
10139 		mode_lib->ms.support.ImmediateFlipSupport = 1; // assume mode support say immediate flip ok at max state/combine
10140 		dml_core_mode_support_partial(mode_lib);
10141 	}
10142 
10143 	dml_core_mode_programming(mode_lib, &clk_cfg);
10144 
10145 	dml_print("DML::%s: ------------- DONE ----------\n", __func__);
10146 	dml_print("DML::%s: PrefetchAndImmediateFlipSupported = %0d\n", __func__, mode_lib->mp.PrefetchAndImmediateFlipSupported);
10147 	return mode_lib->mp.PrefetchAndImmediateFlipSupported;
10148 }
10149 
mode_support_pwr_states(dml_uint_t * lowest_state_idx,struct display_mode_lib_st * mode_lib,const struct dml_display_cfg_st * display_cfg,dml_uint_t start_state_idx,dml_uint_t end_state_idx)10150 static dml_uint_t mode_support_pwr_states(
10151 	dml_uint_t *lowest_state_idx,
10152 	struct display_mode_lib_st *mode_lib,
10153 	const struct dml_display_cfg_st *display_cfg,
10154 	dml_uint_t start_state_idx,
10155 	dml_uint_t end_state_idx)
10156 {
10157 	dml_uint_t state_idx = 0;
10158 	dml_bool_t mode_is_supported = 0;
10159 	*lowest_state_idx = end_state_idx;
10160 
10161 	if (end_state_idx < start_state_idx)
10162 		ASSERT(0);
10163 
10164 	if (end_state_idx >= mode_lib->states.num_states) // idx is 0-based
10165 		ASSERT(0);
10166 
10167 	for (state_idx = start_state_idx; state_idx <= end_state_idx; state_idx++) {
10168 		if (dml_mode_support(mode_lib, state_idx, display_cfg)) {
10169 			dml_print("DML::%s: Mode is supported at power state_idx = %u\n", __func__, state_idx);
10170 			mode_is_supported = 1;
10171 			*lowest_state_idx = state_idx;
10172 			break;
10173 		}
10174 	}
10175 
10176 	return mode_is_supported;
10177 }
10178 
dml_mode_support_ex(struct dml_mode_support_ex_params_st * in_out_params)10179 dml_uint_t dml_mode_support_ex(struct dml_mode_support_ex_params_st *in_out_params)
10180 {
10181 	dml_uint_t result;
10182 
10183 	result = mode_support_pwr_states(&in_out_params->out_lowest_state_idx,
10184 		in_out_params->mode_lib,
10185 		in_out_params->in_display_cfg,
10186 		0,
10187 		in_out_params->mode_lib->states.num_states - 1);
10188 
10189 	if (result)
10190 		*in_out_params->out_evaluation_info = in_out_params->mode_lib->ms.support;
10191 
10192 	return result;
10193 }
10194 
dml_get_is_phantom_pipe(struct display_mode_lib_st * mode_lib,dml_uint_t pipe_idx)10195 dml_bool_t dml_get_is_phantom_pipe(struct display_mode_lib_st *mode_lib, dml_uint_t pipe_idx)
10196 {
10197 	dml_uint_t plane_idx = mode_lib->mp.pipe_plane[pipe_idx];
10198 	dml_print("DML::%s: pipe_idx=%d UseMALLForPStateChange=%0d\n", __func__, pipe_idx, mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange[plane_idx]);
10199 	return (mode_lib->ms.cache_display_cfg.plane.UseMALLForPStateChange[plane_idx] == dml_use_mall_pstate_change_phantom_pipe);
10200 }
10201 
10202 #define dml_get_per_surface_var_func(variable, type, interval_var) type dml_get_##variable(struct display_mode_lib_st *mode_lib, dml_uint_t surface_idx) \
10203 { \
10204 	dml_uint_t plane_idx; \
10205 	plane_idx = mode_lib->mp.pipe_plane[surface_idx]; \
10206 	return (type) interval_var[plane_idx]; \
10207 }
10208 
10209 #define dml_get_var_func(var, type, internal_var)  type dml_get_##var(struct display_mode_lib_st *mode_lib) \
10210 { \
10211 	return (type) internal_var; \
10212 }
10213 
10214 dml_get_var_func(wm_urgent, dml_float_t, mode_lib->mp.Watermark.UrgentWatermark);
10215 dml_get_var_func(wm_stutter_exit, dml_float_t, mode_lib->mp.Watermark.StutterExitWatermark);
10216 dml_get_var_func(wm_stutter_enter_exit, dml_float_t, mode_lib->mp.Watermark.StutterEnterPlusExitWatermark);
10217 dml_get_var_func(wm_memory_trip, dml_float_t, mode_lib->mp.UrgentLatency);
10218 dml_get_var_func(wm_fclk_change, dml_float_t, mode_lib->mp.Watermark.FCLKChangeWatermark);
10219 dml_get_var_func(wm_usr_retraining, dml_float_t, mode_lib->mp.Watermark.USRRetrainingWatermark);
10220 dml_get_var_func(wm_dram_clock_change, dml_float_t, mode_lib->mp.Watermark.DRAMClockChangeWatermark);
10221 dml_get_var_func(wm_z8_stutter_enter_exit, dml_float_t, mode_lib->mp.Watermark.Z8StutterEnterPlusExitWatermark);
10222 dml_get_var_func(wm_z8_stutter, dml_float_t, mode_lib->mp.Watermark.Z8StutterExitWatermark);
10223 dml_get_var_func(fraction_of_urgent_bandwidth, dml_float_t, mode_lib->mp.FractionOfUrgentBandwidth);
10224 dml_get_var_func(fraction_of_urgent_bandwidth_imm_flip, dml_float_t, mode_lib->mp.FractionOfUrgentBandwidthImmediateFlip);
10225 dml_get_var_func(urgent_latency, dml_float_t, mode_lib->mp.UrgentLatency);
10226 dml_get_var_func(clk_dcf_deepsleep, dml_float_t, mode_lib->mp.DCFCLKDeepSleep);
10227 dml_get_var_func(wm_writeback_dram_clock_change, dml_float_t, mode_lib->mp.Watermark.WritebackDRAMClockChangeWatermark);
10228 dml_get_var_func(wm_writeback_urgent, dml_float_t, mode_lib->mp.Watermark.WritebackUrgentWatermark);
10229 dml_get_var_func(stutter_efficiency, dml_float_t, mode_lib->mp.StutterEfficiency);
10230 dml_get_var_func(stutter_efficiency_no_vblank, dml_float_t, mode_lib->mp.StutterEfficiencyNotIncludingVBlank);
10231 dml_get_var_func(stutter_efficiency_z8, dml_float_t, mode_lib->mp.Z8StutterEfficiency);
10232 dml_get_var_func(stutter_num_bursts_z8, dml_float_t, mode_lib->mp.Z8NumberOfStutterBurstsPerFrame);
10233 dml_get_var_func(stutter_period, dml_float_t, mode_lib->mp.StutterPeriod);
10234 dml_get_var_func(stutter_efficiency_z8_bestcase, dml_float_t, mode_lib->mp.Z8StutterEfficiencyBestCase);
10235 dml_get_var_func(stutter_num_bursts_z8_bestcase, dml_float_t, mode_lib->mp.Z8NumberOfStutterBurstsPerFrameBestCase);
10236 dml_get_var_func(stutter_period_bestcase, dml_float_t, mode_lib->mp.StutterPeriodBestCase);
10237 dml_get_var_func(urgent_extra_latency, dml_float_t, mode_lib->mp.UrgentExtraLatency);
10238 dml_get_var_func(fclk_change_latency, dml_float_t, mode_lib->mp.MaxActiveFCLKChangeLatencySupported);
10239 dml_get_var_func(dispclk_calculated, dml_float_t, mode_lib->mp.Dispclk_calculated);
10240 dml_get_var_func(total_data_read_bw, dml_float_t, mode_lib->mp.TotalDataReadBandwidth);
10241 dml_get_var_func(return_bw, dml_float_t, mode_lib->ms.ReturnBW);
10242 dml_get_var_func(return_dram_bw, dml_float_t, mode_lib->ms.ReturnDRAMBW);
10243 dml_get_var_func(tcalc, dml_float_t, mode_lib->mp.TCalc);
10244 dml_get_var_func(comp_buffer_size_kbytes, dml_uint_t, mode_lib->mp.CompressedBufferSizeInkByte);
10245 dml_get_var_func(pixel_chunk_size_in_kbyte, dml_uint_t, mode_lib->ms.ip.pixel_chunk_size_kbytes);
10246 dml_get_var_func(alpha_pixel_chunk_size_in_kbyte, dml_uint_t, mode_lib->ms.ip.alpha_pixel_chunk_size_kbytes);
10247 dml_get_var_func(meta_chunk_size_in_kbyte, dml_uint_t, mode_lib->ms.ip.meta_chunk_size_kbytes);
10248 dml_get_var_func(min_pixel_chunk_size_in_byte, dml_uint_t, mode_lib->ms.ip.min_pixel_chunk_size_bytes);
10249 dml_get_var_func(min_meta_chunk_size_in_byte, dml_uint_t, mode_lib->ms.ip.min_meta_chunk_size_bytes);
10250 dml_get_var_func(total_immediate_flip_bytes, dml_uint_t, mode_lib->mp.TotImmediateFlipBytes);
10251 
10252 dml_get_per_surface_var_func(dsc_delay, dml_uint_t, mode_lib->mp.DSCDelay); // this is the dsc latency
10253 dml_get_per_surface_var_func(dppclk_calculated, dml_float_t, mode_lib->mp.Dppclk_calculated);
10254 dml_get_per_surface_var_func(dscclk_calculated, dml_float_t, mode_lib->mp.DSCCLK_calculated);
10255 dml_get_per_surface_var_func(min_ttu_vblank_in_us, dml_float_t, mode_lib->mp.MinTTUVBlank);
10256 dml_get_per_surface_var_func(vratio_prefetch_l, dml_float_t, mode_lib->mp.VRatioPrefetchY);
10257 dml_get_per_surface_var_func(vratio_prefetch_c, dml_float_t, mode_lib->mp.VRatioPrefetchC);
10258 dml_get_per_surface_var_func(dst_x_after_scaler, dml_uint_t, mode_lib->mp.DSTXAfterScaler);
10259 dml_get_per_surface_var_func(dst_y_after_scaler, dml_uint_t, mode_lib->mp.DSTYAfterScaler);
10260 dml_get_per_surface_var_func(dst_y_per_vm_vblank, dml_float_t, mode_lib->mp.DestinationLinesToRequestVMInVBlank);
10261 dml_get_per_surface_var_func(dst_y_per_row_vblank, dml_float_t, mode_lib->mp.DestinationLinesToRequestRowInVBlank);
10262 dml_get_per_surface_var_func(dst_y_prefetch, dml_float_t, mode_lib->mp.DestinationLinesForPrefetch);
10263 dml_get_per_surface_var_func(dst_y_per_vm_flip, dml_float_t, mode_lib->mp.DestinationLinesToRequestVMInImmediateFlip);
10264 dml_get_per_surface_var_func(dst_y_per_row_flip, dml_float_t, mode_lib->mp.DestinationLinesToRequestRowInImmediateFlip);
10265 dml_get_per_surface_var_func(dst_y_per_pte_row_nom_l, dml_float_t, mode_lib->mp.DST_Y_PER_PTE_ROW_NOM_L);
10266 dml_get_per_surface_var_func(dst_y_per_pte_row_nom_c, dml_float_t, mode_lib->mp.DST_Y_PER_PTE_ROW_NOM_C);
10267 dml_get_per_surface_var_func(dst_y_per_meta_row_nom_l, dml_float_t, mode_lib->mp.DST_Y_PER_META_ROW_NOM_L);
10268 dml_get_per_surface_var_func(dst_y_per_meta_row_nom_c, dml_float_t, mode_lib->mp.DST_Y_PER_META_ROW_NOM_C);
10269 dml_get_per_surface_var_func(refcyc_per_vm_group_vblank_in_us, dml_float_t, mode_lib->mp.TimePerVMGroupVBlank);
10270 dml_get_per_surface_var_func(refcyc_per_vm_group_flip_in_us, dml_float_t, mode_lib->mp.TimePerVMGroupFlip);
10271 dml_get_per_surface_var_func(refcyc_per_vm_req_vblank_in_us, dml_float_t, mode_lib->mp.TimePerVMRequestVBlank);
10272 dml_get_per_surface_var_func(refcyc_per_vm_req_flip_in_us, dml_float_t, mode_lib->mp.TimePerVMRequestFlip);
10273 dml_get_per_surface_var_func(refcyc_per_vm_dmdata_in_us, dml_float_t, mode_lib->mp.Tdmdl_vm);
10274 dml_get_per_surface_var_func(dmdata_dl_delta_in_us, dml_float_t, mode_lib->mp.Tdmdl);
10275 dml_get_per_surface_var_func(refcyc_per_line_delivery_l_in_us, dml_float_t, mode_lib->mp.DisplayPipeLineDeliveryTimeLuma);
10276 dml_get_per_surface_var_func(refcyc_per_line_delivery_c_in_us, dml_float_t, mode_lib->mp.DisplayPipeLineDeliveryTimeChroma);
10277 dml_get_per_surface_var_func(refcyc_per_line_delivery_pre_l_in_us, dml_float_t, mode_lib->mp.DisplayPipeLineDeliveryTimeLumaPrefetch);
10278 dml_get_per_surface_var_func(refcyc_per_line_delivery_pre_c_in_us, dml_float_t, mode_lib->mp.DisplayPipeLineDeliveryTimeChromaPrefetch);
10279 dml_get_per_surface_var_func(refcyc_per_req_delivery_l_in_us, dml_float_t, mode_lib->mp.DisplayPipeRequestDeliveryTimeLuma);
10280 dml_get_per_surface_var_func(refcyc_per_req_delivery_c_in_us, dml_float_t, mode_lib->mp.DisplayPipeRequestDeliveryTimeChroma);
10281 dml_get_per_surface_var_func(refcyc_per_req_delivery_pre_l_in_us, dml_float_t, mode_lib->mp.DisplayPipeRequestDeliveryTimeLumaPrefetch);
10282 dml_get_per_surface_var_func(refcyc_per_req_delivery_pre_c_in_us, dml_float_t, mode_lib->mp.DisplayPipeRequestDeliveryTimeChromaPrefetch);
10283 dml_get_per_surface_var_func(refcyc_per_cursor_req_delivery_in_us, dml_float_t, mode_lib->mp.CursorRequestDeliveryTime);
10284 dml_get_per_surface_var_func(refcyc_per_cursor_req_delivery_pre_in_us, dml_float_t, mode_lib->mp.CursorRequestDeliveryTimePrefetch);
10285 dml_get_per_surface_var_func(refcyc_per_meta_chunk_nom_l_in_us, dml_float_t, mode_lib->mp.TimePerMetaChunkNominal);
10286 dml_get_per_surface_var_func(refcyc_per_meta_chunk_nom_c_in_us, dml_float_t, mode_lib->mp.TimePerChromaMetaChunkNominal);
10287 dml_get_per_surface_var_func(refcyc_per_meta_chunk_vblank_l_in_us, dml_float_t, mode_lib->mp.TimePerMetaChunkVBlank);
10288 dml_get_per_surface_var_func(refcyc_per_meta_chunk_vblank_c_in_us, dml_float_t, mode_lib->mp.TimePerChromaMetaChunkVBlank);
10289 dml_get_per_surface_var_func(refcyc_per_meta_chunk_flip_l_in_us, dml_float_t, mode_lib->mp.TimePerMetaChunkFlip);
10290 dml_get_per_surface_var_func(refcyc_per_meta_chunk_flip_c_in_us, dml_float_t, mode_lib->mp.TimePerChromaMetaChunkFlip);
10291 dml_get_per_surface_var_func(refcyc_per_pte_group_nom_l_in_us, dml_float_t, mode_lib->mp.time_per_pte_group_nom_luma);
10292 dml_get_per_surface_var_func(refcyc_per_pte_group_nom_c_in_us, dml_float_t, mode_lib->mp.time_per_pte_group_nom_chroma);
10293 dml_get_per_surface_var_func(refcyc_per_pte_group_vblank_l_in_us, dml_float_t, mode_lib->mp.time_per_pte_group_vblank_luma);
10294 dml_get_per_surface_var_func(refcyc_per_pte_group_vblank_c_in_us, dml_float_t, mode_lib->mp.time_per_pte_group_vblank_chroma);
10295 dml_get_per_surface_var_func(refcyc_per_pte_group_flip_l_in_us, dml_float_t, mode_lib->mp.time_per_pte_group_flip_luma);
10296 dml_get_per_surface_var_func(refcyc_per_pte_group_flip_c_in_us, dml_float_t, mode_lib->mp.time_per_pte_group_flip_chroma);
10297 dml_get_per_surface_var_func(dpte_group_size_in_bytes, dml_uint_t, mode_lib->mp.dpte_group_bytes);
10298 dml_get_per_surface_var_func(vm_group_size_in_bytes, dml_uint_t, mode_lib->mp.vm_group_bytes);
10299 dml_get_per_surface_var_func(swath_height_l, dml_uint_t, mode_lib->ms.SwathHeightY);
10300 dml_get_per_surface_var_func(swath_height_c, dml_uint_t, mode_lib->ms.SwathHeightC);
10301 dml_get_per_surface_var_func(dpte_row_height_l, dml_uint_t, mode_lib->mp.dpte_row_height);
10302 dml_get_per_surface_var_func(dpte_row_height_c, dml_uint_t, mode_lib->mp.dpte_row_height_chroma);
10303 dml_get_per_surface_var_func(dpte_row_height_linear_l, dml_uint_t, mode_lib->mp.dpte_row_height_linear);
10304 dml_get_per_surface_var_func(dpte_row_height_linear_c, dml_uint_t, mode_lib->mp.dpte_row_height_linear_chroma);
10305 dml_get_per_surface_var_func(meta_row_height_l, dml_uint_t, mode_lib->mp.meta_row_height);
10306 dml_get_per_surface_var_func(meta_row_height_c, dml_uint_t, mode_lib->mp.meta_row_height_chroma);
10307 
10308 dml_get_per_surface_var_func(vstartup_calculated, dml_uint_t, mode_lib->mp.VStartup);
10309 dml_get_per_surface_var_func(vupdate_offset, dml_uint_t, mode_lib->mp.VUpdateOffsetPix);
10310 dml_get_per_surface_var_func(vupdate_width, dml_uint_t, mode_lib->mp.VUpdateWidthPix);
10311 dml_get_per_surface_var_func(vready_offset, dml_uint_t, mode_lib->mp.VReadyOffsetPix);
10312 dml_get_per_surface_var_func(vready_at_or_after_vsync, dml_uint_t, mode_lib->mp.VREADY_AT_OR_AFTER_VSYNC);
10313 dml_get_per_surface_var_func(min_dst_y_next_start, dml_uint_t, mode_lib->mp.MIN_DST_Y_NEXT_START);
10314 dml_get_per_surface_var_func(det_stored_buffer_size_l_bytes, dml_uint_t, mode_lib->ms.DETBufferSizeY);
10315 dml_get_per_surface_var_func(det_stored_buffer_size_c_bytes, dml_uint_t, mode_lib->ms.DETBufferSizeC);
10316 dml_get_per_surface_var_func(use_mall_for_static_screen, dml_uint_t, mode_lib->mp.UsesMALLForStaticScreen);
10317 dml_get_per_surface_var_func(surface_size_for_mall, dml_uint_t, mode_lib->mp.SurfaceSizeInTheMALL);
10318 dml_get_per_surface_var_func(dcc_max_uncompressed_block_l, dml_uint_t, mode_lib->mp.DCCYMaxUncompressedBlock);
10319 dml_get_per_surface_var_func(dcc_max_compressed_block_l, dml_uint_t, mode_lib->mp.DCCYMaxCompressedBlock);
10320 dml_get_per_surface_var_func(dcc_independent_block_l, dml_uint_t, mode_lib->mp.DCCYIndependentBlock);
10321 dml_get_per_surface_var_func(dcc_max_uncompressed_block_c, dml_uint_t, mode_lib->mp.DCCCMaxUncompressedBlock);
10322 dml_get_per_surface_var_func(dcc_max_compressed_block_c, dml_uint_t, mode_lib->mp.DCCCMaxCompressedBlock);
10323 dml_get_per_surface_var_func(dcc_independent_block_c, dml_uint_t, mode_lib->mp.DCCCIndependentBlock);
10324 dml_get_per_surface_var_func(max_active_dram_clock_change_latency_supported, dml_uint_t, mode_lib->mp.MaxActiveDRAMClockChangeLatencySupported);
10325 dml_get_per_surface_var_func(pte_buffer_mode, dml_uint_t, mode_lib->mp.PTE_BUFFER_MODE);
10326 dml_get_per_surface_var_func(bigk_fragment_size, dml_uint_t, mode_lib->mp.BIGK_FRAGMENT_SIZE);
10327 dml_get_per_surface_var_func(dpte_bytes_per_row, dml_uint_t, mode_lib->mp.PixelPTEBytesPerRow);
10328 dml_get_per_surface_var_func(meta_bytes_per_row, dml_uint_t, mode_lib->mp.MetaRowByte);
10329 dml_get_per_surface_var_func(det_buffer_size_kbytes, dml_uint_t, mode_lib->ms.DETBufferSizeInKByte);
10330