2 * Copyright 2018 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
26 #include "../display_mode_lib.h"
27 #include "display_mode_vba_20.h"
28 #include "../dml_inline_defs.h"
32 * This file is gcc-parseable HW gospel, coming straight from HW engineers.
34 * It doesn't adhere to Linux kernel style and sometimes will do things in odd
35 * ways. Unless there is something clearly wrong with it the code should
36 * remain as-is as it provides us with a guarantee from HW that it is correct.
40 #define BPP_BLENDED_PIPE 0xffffffff
42 static double adjust_ReturnBW(
43 struct display_mode_lib *mode_lib,
45 bool DCCEnabledAnyPlane,
46 double ReturnBandwidthToDCN);
47 static unsigned int dscceComputeDelay(
50 unsigned int sliceWidth,
51 unsigned int numSlices,
52 enum output_format_class pixelFormat);
53 static unsigned int dscComputeDelay(enum output_format_class pixelFormat);
54 // Super monster function with some 45 argument
55 static bool CalculatePrefetchSchedule(
56 struct display_mode_lib *mode_lib,
60 double DCFCLKDeepSleep,
61 unsigned int DSCDelay,
62 unsigned int DPPPerPlane,
64 unsigned int NumberOfCursors,
65 double DPPCLKDelaySubtotal,
66 double DPPCLKDelaySCL,
67 double DPPCLKDelaySCLLBOnly,
68 double DPPCLKDelayCNVCFormater,
69 double DPPCLKDelayCNVCCursor,
70 double DISPCLKDelaySubtotal,
71 unsigned int ScalerRecoutWidth,
72 enum output_format_class OutputFormat,
75 unsigned int MaxInterDCNTileRepeaters,
76 unsigned int VStartup,
77 unsigned int PageTableLevels,
79 bool DynamicMetadataEnable,
80 unsigned int DynamicMetadataLinesBeforeActiveRequired,
81 unsigned int DynamicMetadataTransmittedBytes,
83 double UrgentLatencyPixelDataOnly,
84 double UrgentExtraLatency,
86 unsigned int PDEAndMetaPTEBytesFrame,
87 unsigned int MetaRowByte,
88 unsigned int PixelPTEBytesPerRow,
89 double PrefetchSourceLinesY,
90 unsigned int SwathWidthY,
91 double BytePerPixelDETY,
93 unsigned int MaxNumSwathY,
94 double PrefetchSourceLinesC,
95 double BytePerPixelDETC,
97 unsigned int MaxNumSwathC,
98 unsigned int SwathHeightY,
99 unsigned int SwathHeightC,
102 double XFCRemoteSurfaceFlipDelay,
103 bool InterlaceEnable,
104 bool ProgressiveToInterlaceUnitInOPP,
105 double *DSTXAfterScaler,
106 double *DSTYAfterScaler,
107 double *DestinationLinesForPrefetch,
108 double *PrefetchBandwidth,
109 double *DestinationLinesToRequestVMInVBlank,
110 double *DestinationLinesToRequestRowInVBlank,
111 double *VRatioPrefetchY,
112 double *VRatioPrefetchC,
113 double *RequiredPrefetchPixDataBW,
114 unsigned int *VStartupRequiredWhenNotEnoughTimeForDynamicMetadata,
116 unsigned int *VUpdateOffsetPix,
117 double *VUpdateWidthPix,
118 double *VReadyOffsetPix);
119 static double RoundToDFSGranularityUp(double Clock, double VCOSpeed);
120 static double RoundToDFSGranularityDown(double Clock, double VCOSpeed);
121 static double CalculatePrefetchSourceLines(
122 struct display_mode_lib *mode_lib,
126 bool ProgressiveToInterlaceUnitInOPP,
127 unsigned int SwathHeight,
128 unsigned int ViewportYStart,
129 double *VInitPreFill,
130 unsigned int *MaxNumSwath);
131 static unsigned int CalculateVMAndRowBytes(
132 struct display_mode_lib *mode_lib,
134 unsigned int BlockHeight256Bytes,
135 unsigned int BlockWidth256Bytes,
136 enum source_format_class SourcePixelFormat,
137 unsigned int SurfaceTiling,
138 unsigned int BytePerPixel,
139 enum scan_direction_class ScanDirection,
140 unsigned int ViewportWidth,
141 unsigned int ViewportHeight,
142 unsigned int SwathWidthY,
144 unsigned int VMMPageSize,
145 unsigned int PTEBufferSizeInRequestsLuma,
146 unsigned int PDEProcessingBufIn64KBReqs,
148 unsigned int DCCMetaPitch,
149 unsigned int *MacroTileWidth,
150 unsigned int *MetaRowByte,
151 unsigned int *PixelPTEBytesPerRow,
152 bool *PTEBufferSizeNotExceeded,
153 unsigned int *dpte_row_height,
154 unsigned int *meta_row_height);
155 static double CalculateTWait(
156 unsigned int PrefetchMode,
157 double DRAMClockChangeLatency,
158 double UrgentLatencyPixelDataOnly,
159 double SREnterPlusExitTime);
160 static double CalculateRemoteSurfaceFlipDelay(
161 struct display_mode_lib *mode_lib,
166 double XFCTSlvVupdateOffset,
167 double XFCTSlvVupdateWidth,
168 double XFCTSlvVreadyOffset,
169 double XFCXBUFLatencyTolerance,
170 double XFCFillBWOverhead,
171 double XFCSlvChunkSize,
172 double XFCBusTransportTime,
175 double *SrcActiveDrainRate,
178 static void CalculateActiveRowBandwidth(
180 enum source_format_class SourcePixelFormat,
184 unsigned int MetaRowByteLuma,
185 unsigned int MetaRowByteChroma,
186 unsigned int meta_row_height_luma,
187 unsigned int meta_row_height_chroma,
188 unsigned int PixelPTEBytesPerRowLuma,
189 unsigned int PixelPTEBytesPerRowChroma,
190 unsigned int dpte_row_height_luma,
191 unsigned int dpte_row_height_chroma,
194 double *qual_row_bw);
195 static void CalculateFlipSchedule(
196 struct display_mode_lib *mode_lib,
197 double UrgentExtraLatency,
198 double UrgentLatencyPixelDataOnly,
199 unsigned int GPUVMMaxPageTableLevels,
201 double BandwidthAvailableForImmediateFlip,
202 unsigned int TotImmediateFlipBytes,
203 enum source_format_class SourcePixelFormat,
204 unsigned int ImmediateFlipBytes,
208 double PDEAndMetaPTEBytesFrame,
209 unsigned int MetaRowByte,
210 unsigned int PixelPTEBytesPerRow,
212 unsigned int dpte_row_height,
213 unsigned int meta_row_height,
215 double *DestinationLinesToRequestVMInImmediateFlip,
216 double *DestinationLinesToRequestRowInImmediateFlip,
217 double *final_flip_bw,
218 bool *ImmediateFlipSupportedForPipe);
219 static double CalculateWriteBackDelay(
220 enum source_format_class WritebackPixelFormat,
221 double WritebackHRatio,
222 double WritebackVRatio,
223 unsigned int WritebackLumaHTaps,
224 unsigned int WritebackLumaVTaps,
225 unsigned int WritebackChromaHTaps,
226 unsigned int WritebackChromaVTaps,
227 unsigned int WritebackDestinationWidth);
229 static void dml20_DisplayPipeConfiguration(struct display_mode_lib *mode_lib);
230 static void dml20_DISPCLKDPPCLKDCFCLKDeepSleepPrefetchParametersWatermarksAndPerformanceCalculation(
231 struct display_mode_lib *mode_lib);
233 void dml20_recalculate(struct display_mode_lib *mode_lib)
235 ModeSupportAndSystemConfiguration(mode_lib);
236 mode_lib->vba.FabricAndDRAMBandwidth = dml_min(
237 mode_lib->vba.DRAMSpeed * mode_lib->vba.NumberOfChannels * mode_lib->vba.DRAMChannelWidth,
238 mode_lib->vba.FabricClock * mode_lib->vba.FabricDatapathToDCNDataReturn) / 1000.0;
239 PixelClockAdjustmentForProgressiveToInterlaceUnit(mode_lib);
240 dml20_DisplayPipeConfiguration(mode_lib);
241 dml20_DISPCLKDPPCLKDCFCLKDeepSleepPrefetchParametersWatermarksAndPerformanceCalculation(mode_lib);
244 static double adjust_ReturnBW(
245 struct display_mode_lib *mode_lib,
247 bool DCCEnabledAnyPlane,
248 double ReturnBandwidthToDCN)
250 double CriticalCompression;
252 if (DCCEnabledAnyPlane
253 && ReturnBandwidthToDCN
254 > mode_lib->vba.DCFCLK * mode_lib->vba.ReturnBusWidth / 4.0)
258 ReturnBandwidthToDCN * 4
260 - mode_lib->vba.UrgentLatencyPixelDataOnly
261 / ((mode_lib->vba.ROBBufferSizeInKByte
262 - mode_lib->vba.PixelChunkSizeInKByte)
264 / ReturnBandwidthToDCN
265 - mode_lib->vba.DCFCLK
266 * mode_lib->vba.ReturnBusWidth
268 + mode_lib->vba.UrgentLatencyPixelDataOnly));
270 CriticalCompression = 2.0 * mode_lib->vba.ReturnBusWidth * mode_lib->vba.DCFCLK
271 * mode_lib->vba.UrgentLatencyPixelDataOnly
272 / (ReturnBandwidthToDCN * mode_lib->vba.UrgentLatencyPixelDataOnly
273 + (mode_lib->vba.ROBBufferSizeInKByte
274 - mode_lib->vba.PixelChunkSizeInKByte)
277 if (DCCEnabledAnyPlane && CriticalCompression > 1.0 && CriticalCompression < 4.0)
281 4.0 * ReturnBandwidthToDCN
282 * (mode_lib->vba.ROBBufferSizeInKByte
283 - mode_lib->vba.PixelChunkSizeInKByte)
285 * mode_lib->vba.ReturnBusWidth
286 * mode_lib->vba.DCFCLK
287 * mode_lib->vba.UrgentLatencyPixelDataOnly
289 (ReturnBandwidthToDCN
290 * mode_lib->vba.UrgentLatencyPixelDataOnly
291 + (mode_lib->vba.ROBBufferSizeInKByte
292 - mode_lib->vba.PixelChunkSizeInKByte)
299 static unsigned int dscceComputeDelay(
302 unsigned int sliceWidth,
303 unsigned int numSlices,
304 enum output_format_class pixelFormat)
306 // valid bpc = source bits per component in the set of {8, 10, 12}
307 // valid bpp = increments of 1/16 of a bit
308 // min = 6/7/8 in N420/N422/444, respectively
309 // max = such that compression is 1:1
310 //valid sliceWidth = number of pixels per slice line, must be less than or equal to 5184/numSlices (or 4096/numSlices in 420 mode)
311 //valid numSlices = number of slices in the horiziontal direction per DSC engine in the set of {1, 2, 3, 4}
312 //valid pixelFormat = pixel/color format in the set of {:N444_RGB, :S422, :N422, :N420}
315 unsigned int rcModelSize = 8192;
317 // N422/N420 operate at 2 pixels per clock
318 unsigned int pixelsPerClock, lstall, D, initalXmitDelay, w, s, ix, wx, p, l0, a, ax, l,
321 if (pixelFormat == dm_n422 || pixelFormat == dm_420)
323 // #all other modes operate at 1 pixel per clock
327 //initial transmit delay as per PPS
328 initalXmitDelay = dml_round(rcModelSize / 2.0 / bpp / pixelsPerClock);
338 //divide by pixel per cycle to compute slice width as seen by DSC
339 w = sliceWidth / pixelsPerClock;
341 //422 mode has an additional cycle of delay
342 if (pixelFormat == dm_s422)
347 //main calculation for the dscce
348 ix = initalXmitDelay + 45;
353 ax = (a + 2) / 3 + D + 6 + 1;
354 l = (ax + wx - 1) / wx;
355 if ((ix % w) == 0 && p != 0)
359 Delay = l * wx * (numSlices - 1) + ax + s + lstall + 22;
361 //dsc processes 3 pixel containers per cycle and a container can contain 1 or 2 pixels
362 pixels = Delay * 3 * pixelsPerClock;
366 static unsigned int dscComputeDelay(enum output_format_class pixelFormat)
368 unsigned int Delay = 0;
370 if (pixelFormat == dm_420) {
375 // dscc - input deserializer
377 // dscc gets pixels every other cycle
379 // dscc - input cdc fifo
381 // dscc gets pixels every other cycle
383 // dscc - cdc uncertainty
385 // dscc - output cdc fifo
387 // dscc gets pixels every other cycle
389 // dscc - cdc uncertainty
391 // dscc - output serializer
395 } else if (pixelFormat == dm_n422) {
400 // dscc - input deserializer
402 // dscc - input cdc fifo
404 // dscc - cdc uncertainty
406 // dscc - output cdc fifo
408 // dscc - cdc uncertainty
410 // dscc - output serializer
419 // dscc - input deserializer
421 // dscc - input cdc fifo
423 // dscc - cdc uncertainty
425 // dscc - output cdc fifo
427 // dscc - output serializer
429 // dscc - cdc uncertainty
438 static bool CalculatePrefetchSchedule(
439 struct display_mode_lib *mode_lib,
443 double DCFCLKDeepSleep,
444 unsigned int DSCDelay,
445 unsigned int DPPPerPlane,
447 unsigned int NumberOfCursors,
448 double DPPCLKDelaySubtotal,
449 double DPPCLKDelaySCL,
450 double DPPCLKDelaySCLLBOnly,
451 double DPPCLKDelayCNVCFormater,
452 double DPPCLKDelayCNVCCursor,
453 double DISPCLKDelaySubtotal,
454 unsigned int ScalerRecoutWidth,
455 enum output_format_class OutputFormat,
458 unsigned int MaxInterDCNTileRepeaters,
459 unsigned int VStartup,
460 unsigned int PageTableLevels,
462 bool DynamicMetadataEnable,
463 unsigned int DynamicMetadataLinesBeforeActiveRequired,
464 unsigned int DynamicMetadataTransmittedBytes,
466 double UrgentLatencyPixelDataOnly,
467 double UrgentExtraLatency,
469 unsigned int PDEAndMetaPTEBytesFrame,
470 unsigned int MetaRowByte,
471 unsigned int PixelPTEBytesPerRow,
472 double PrefetchSourceLinesY,
473 unsigned int SwathWidthY,
474 double BytePerPixelDETY,
475 double VInitPreFillY,
476 unsigned int MaxNumSwathY,
477 double PrefetchSourceLinesC,
478 double BytePerPixelDETC,
479 double VInitPreFillC,
480 unsigned int MaxNumSwathC,
481 unsigned int SwathHeightY,
482 unsigned int SwathHeightC,
485 double XFCRemoteSurfaceFlipDelay,
486 bool InterlaceEnable,
487 bool ProgressiveToInterlaceUnitInOPP,
488 double *DSTXAfterScaler,
489 double *DSTYAfterScaler,
490 double *DestinationLinesForPrefetch,
491 double *PrefetchBandwidth,
492 double *DestinationLinesToRequestVMInVBlank,
493 double *DestinationLinesToRequestRowInVBlank,
494 double *VRatioPrefetchY,
495 double *VRatioPrefetchC,
496 double *RequiredPrefetchPixDataBW,
497 unsigned int *VStartupRequiredWhenNotEnoughTimeForDynamicMetadata,
499 unsigned int *VUpdateOffsetPix,
500 double *VUpdateWidthPix,
501 double *VReadyOffsetPix)
503 bool MyError = false;
504 unsigned int DPPCycles, DISPCLKCycles;
505 double DSTTotalPixelsAfterScaler, TotalRepeaterDelayTime;
506 double Tdm, LineTime, Tsetup;
507 double dst_y_prefetch_equ;
509 double prefetch_bw_oto;
513 double dst_y_prefetch_oto;
514 double TimeForFetchingMetaPTE = 0;
515 double TimeForFetchingRowInVBlank = 0;
516 double LinesToRequestPrefetchPixelData = 0;
519 DPPCycles = DPPCLKDelaySubtotal + DPPCLKDelaySCL;
521 DPPCycles = DPPCLKDelaySubtotal + DPPCLKDelaySCLLBOnly;
523 DPPCycles = DPPCycles + DPPCLKDelayCNVCFormater + NumberOfCursors * DPPCLKDelayCNVCCursor;
525 DISPCLKCycles = DISPCLKDelaySubtotal;
527 if (DPPCLK == 0.0 || DISPCLK == 0.0)
530 *DSTXAfterScaler = DPPCycles * PixelClock / DPPCLK + DISPCLKCycles * PixelClock / DISPCLK
534 *DSTXAfterScaler = *DSTXAfterScaler + ScalerRecoutWidth;
536 if (OutputFormat == dm_420 || (InterlaceEnable && ProgressiveToInterlaceUnitInOPP))
537 *DSTYAfterScaler = 1;
539 *DSTYAfterScaler = 0;
541 DSTTotalPixelsAfterScaler = ((double) (*DSTYAfterScaler * HTotal)) + *DSTXAfterScaler;
542 *DSTYAfterScaler = dml_floor(DSTTotalPixelsAfterScaler / HTotal, 1);
543 *DSTXAfterScaler = DSTTotalPixelsAfterScaler - ((double) (*DSTYAfterScaler * HTotal));
545 *VUpdateOffsetPix = dml_ceil(HTotal / 4.0, 1);
546 TotalRepeaterDelayTime = MaxInterDCNTileRepeaters * (2.0 / DPPCLK + 3.0 / DISPCLK);
547 *VUpdateWidthPix = (14.0 / DCFCLKDeepSleep + 12.0 / DPPCLK + TotalRepeaterDelayTime)
550 *VReadyOffsetPix = dml_max(
552 TotalRepeaterDelayTime + 20.0 / DCFCLKDeepSleep + 10.0 / DPPCLK)
555 Tsetup = (double) (*VUpdateOffsetPix + *VUpdateWidthPix + *VReadyOffsetPix) / PixelClock;
557 LineTime = (double) HTotal / PixelClock;
559 if (DynamicMetadataEnable) {
560 double Tdmbf, Tdmec, Tdmsks;
562 Tdm = dml_max(0.0, UrgentExtraLatency - TCalc);
563 Tdmbf = DynamicMetadataTransmittedBytes / 4.0 / DISPCLK;
565 if (DynamicMetadataLinesBeforeActiveRequired == 0)
566 Tdmsks = VBlank * LineTime / 2.0;
568 Tdmsks = DynamicMetadataLinesBeforeActiveRequired * LineTime;
569 if (InterlaceEnable && !ProgressiveToInterlaceUnitInOPP)
571 if (VStartup * LineTime
572 < Tsetup + TWait + UrgentExtraLatency + Tdmbf + Tdmec + Tdmsks) {
574 *VStartupRequiredWhenNotEnoughTimeForDynamicMetadata = (Tsetup + TWait
575 + UrgentExtraLatency + Tdmbf + Tdmec + Tdmsks) / LineTime;
577 *VStartupRequiredWhenNotEnoughTimeForDynamicMetadata = 0.0;
582 if (PageTableLevels == 4)
583 *Tno_bw = UrgentExtraLatency + UrgentLatencyPixelDataOnly;
584 else if (PageTableLevels == 3)
585 *Tno_bw = UrgentExtraLatency;
588 } else if (DCCEnable)
591 *Tno_bw = LineTime / 4;
593 dst_y_prefetch_equ = VStartup - dml_max(TCalc + TWait, XFCRemoteSurfaceFlipDelay) / LineTime
594 - (Tsetup + Tdm) / LineTime
595 - (*DSTYAfterScaler + *DSTXAfterScaler / HTotal);
597 Tsw_oto = dml_max(PrefetchSourceLinesY, PrefetchSourceLinesC) * LineTime;
599 prefetch_bw_oto = (MetaRowByte + PixelPTEBytesPerRow
600 + PrefetchSourceLinesY * SwathWidthY * dml_ceil(BytePerPixelDETY, 1)
601 + PrefetchSourceLinesC * SwathWidthY / 2 * dml_ceil(BytePerPixelDETC, 2))
604 if (GPUVMEnable == true) {
607 *Tno_bw + PDEAndMetaPTEBytesFrame / prefetch_bw_oto,
610 + UrgentLatencyPixelDataOnly
615 Tvm_oto = LineTime / 4.0;
617 if ((GPUVMEnable == true || DCCEnable == true)) {
619 (MetaRowByte + PixelPTEBytesPerRow) / prefetch_bw_oto,
620 dml_max(UrgentLatencyPixelDataOnly, dml_max(LineTime - Tvm_oto, LineTime / 4)));
622 Tr0_oto = LineTime - Tvm_oto;
624 Tpre_oto = Tvm_oto + Tr0_oto + Tsw_oto;
626 dst_y_prefetch_oto = Tpre_oto / LineTime;
628 if (dst_y_prefetch_oto < dst_y_prefetch_equ)
629 *DestinationLinesForPrefetch = dst_y_prefetch_oto;
631 *DestinationLinesForPrefetch = dst_y_prefetch_equ;
633 *DestinationLinesForPrefetch = dml_floor(4.0 * (*DestinationLinesForPrefetch + 0.125), 1)
636 dml_print("DML: VStartup: %d\n", VStartup);
637 dml_print("DML: TCalc: %f\n", TCalc);
638 dml_print("DML: TWait: %f\n", TWait);
639 dml_print("DML: XFCRemoteSurfaceFlipDelay: %f\n", XFCRemoteSurfaceFlipDelay);
640 dml_print("DML: LineTime: %f\n", LineTime);
641 dml_print("DML: Tsetup: %f\n", Tsetup);
642 dml_print("DML: Tdm: %f\n", Tdm);
643 dml_print("DML: DSTYAfterScaler: %f\n", *DSTYAfterScaler);
644 dml_print("DML: DSTXAfterScaler: %f\n", *DSTXAfterScaler);
645 dml_print("DML: HTotal: %d\n", HTotal);
647 *PrefetchBandwidth = 0;
648 *DestinationLinesToRequestVMInVBlank = 0;
649 *DestinationLinesToRequestRowInVBlank = 0;
650 *VRatioPrefetchY = 0;
651 *VRatioPrefetchC = 0;
652 *RequiredPrefetchPixDataBW = 0;
653 if (*DestinationLinesForPrefetch > 1) {
654 *PrefetchBandwidth = (PDEAndMetaPTEBytesFrame + 2 * MetaRowByte
655 + 2 * PixelPTEBytesPerRow
656 + PrefetchSourceLinesY * SwathWidthY * dml_ceil(BytePerPixelDETY, 1)
657 + PrefetchSourceLinesC * SwathWidthY / 2
658 * dml_ceil(BytePerPixelDETC, 2))
659 / (*DestinationLinesForPrefetch * LineTime - *Tno_bw);
661 TimeForFetchingMetaPTE =
664 + (double) PDEAndMetaPTEBytesFrame
665 / *PrefetchBandwidth,
668 + UrgentLatencyPixelDataOnly
673 if (NumberOfCursors > 0 || XFCEnabled)
674 TimeForFetchingMetaPTE = LineTime / 4;
676 TimeForFetchingMetaPTE = 0.0;
679 if ((GPUVMEnable == true || DCCEnable == true)) {
680 TimeForFetchingRowInVBlank =
682 (MetaRowByte + PixelPTEBytesPerRow)
683 / *PrefetchBandwidth,
685 UrgentLatencyPixelDataOnly,
688 - TimeForFetchingMetaPTE,
692 if (NumberOfCursors > 0 || XFCEnabled)
693 TimeForFetchingRowInVBlank = LineTime - TimeForFetchingMetaPTE;
695 TimeForFetchingRowInVBlank = 0.0;
698 *DestinationLinesToRequestVMInVBlank = dml_floor(
699 4.0 * (TimeForFetchingMetaPTE / LineTime + 0.125),
702 *DestinationLinesToRequestRowInVBlank = dml_floor(
703 4.0 * (TimeForFetchingRowInVBlank / LineTime + 0.125),
706 LinesToRequestPrefetchPixelData =
707 *DestinationLinesForPrefetch
708 - ((NumberOfCursors > 0 || GPUVMEnable
710 (*DestinationLinesToRequestVMInVBlank
711 + *DestinationLinesToRequestRowInVBlank) :
714 if (LinesToRequestPrefetchPixelData > 0) {
716 *VRatioPrefetchY = (double) PrefetchSourceLinesY
717 / LinesToRequestPrefetchPixelData;
718 *VRatioPrefetchY = dml_max(*VRatioPrefetchY, 1.0);
719 if ((SwathHeightY > 4) && (VInitPreFillY > 3)) {
720 if (LinesToRequestPrefetchPixelData > (VInitPreFillY - 3.0) / 2.0) {
723 (double) PrefetchSourceLinesY
724 / LinesToRequestPrefetchPixelData,
725 (double) MaxNumSwathY
727 / (LinesToRequestPrefetchPixelData
731 *VRatioPrefetchY = dml_max(*VRatioPrefetchY, 1.0);
734 *VRatioPrefetchY = 0;
738 *VRatioPrefetchC = (double) PrefetchSourceLinesC
739 / LinesToRequestPrefetchPixelData;
740 *VRatioPrefetchC = dml_max(*VRatioPrefetchC, 1.0);
742 if ((SwathHeightC > 4)) {
743 if (LinesToRequestPrefetchPixelData > (VInitPreFillC - 3.0) / 2.0) {
747 (double) MaxNumSwathC
749 / (LinesToRequestPrefetchPixelData
753 *VRatioPrefetchC = dml_max(*VRatioPrefetchC, 1.0);
756 *VRatioPrefetchC = 0;
760 *RequiredPrefetchPixDataBW =
762 * ((double) PrefetchSourceLinesY
763 / LinesToRequestPrefetchPixelData
767 + (double) PrefetchSourceLinesC
768 / LinesToRequestPrefetchPixelData
773 * SwathWidthY / LineTime;
776 *VRatioPrefetchY = 0;
777 *VRatioPrefetchC = 0;
778 *RequiredPrefetchPixDataBW = 0;
786 *PrefetchBandwidth = 0;
787 TimeForFetchingMetaPTE = 0;
788 TimeForFetchingRowInVBlank = 0;
789 *DestinationLinesToRequestVMInVBlank = 0;
790 *DestinationLinesToRequestRowInVBlank = 0;
791 *DestinationLinesForPrefetch = 0;
792 LinesToRequestPrefetchPixelData = 0;
793 *VRatioPrefetchY = 0;
794 *VRatioPrefetchC = 0;
795 *RequiredPrefetchPixDataBW = 0;
801 static double RoundToDFSGranularityUp(double Clock, double VCOSpeed)
803 return VCOSpeed * 4 / dml_floor(VCOSpeed * 4 / Clock, 1);
806 static double RoundToDFSGranularityDown(double Clock, double VCOSpeed)
808 return VCOSpeed * 4 / dml_ceil(VCOSpeed * 4 / Clock, 1);
811 static double CalculatePrefetchSourceLines(
812 struct display_mode_lib *mode_lib,
816 bool ProgressiveToInterlaceUnitInOPP,
817 unsigned int SwathHeight,
818 unsigned int ViewportYStart,
819 double *VInitPreFill,
820 unsigned int *MaxNumSwath)
822 unsigned int MaxPartialSwath;
824 if (ProgressiveToInterlaceUnitInOPP)
825 *VInitPreFill = dml_floor((VRatio + vtaps + 1) / 2.0, 1);
827 *VInitPreFill = dml_floor((VRatio + vtaps + 1 + Interlace * 0.5 * VRatio) / 2.0, 1);
829 if (!mode_lib->vba.IgnoreViewportPositioning) {
831 *MaxNumSwath = dml_ceil((*VInitPreFill - 1.0) / SwathHeight, 1) + 1.0;
833 if (*VInitPreFill > 1.0)
834 MaxPartialSwath = (unsigned int) (*VInitPreFill - 2) % SwathHeight;
836 MaxPartialSwath = (unsigned int) (*VInitPreFill + SwathHeight - 2)
838 MaxPartialSwath = dml_max(1U, MaxPartialSwath);
842 if (ViewportYStart != 0)
844 "WARNING DML: using viewport y position of 0 even though actual viewport y position is non-zero in prefetch source lines calculation\n");
846 *MaxNumSwath = dml_ceil(*VInitPreFill / SwathHeight, 1);
848 if (*VInitPreFill > 1.0)
849 MaxPartialSwath = (unsigned int) (*VInitPreFill - 1) % SwathHeight;
851 MaxPartialSwath = (unsigned int) (*VInitPreFill + SwathHeight - 1)
855 return *MaxNumSwath * SwathHeight + MaxPartialSwath;
858 static unsigned int CalculateVMAndRowBytes(
859 struct display_mode_lib *mode_lib,
861 unsigned int BlockHeight256Bytes,
862 unsigned int BlockWidth256Bytes,
863 enum source_format_class SourcePixelFormat,
864 unsigned int SurfaceTiling,
865 unsigned int BytePerPixel,
866 enum scan_direction_class ScanDirection,
867 unsigned int ViewportWidth,
868 unsigned int ViewportHeight,
869 unsigned int SwathWidth,
871 unsigned int VMMPageSize,
872 unsigned int PTEBufferSizeInRequestsLuma,
873 unsigned int PDEProcessingBufIn64KBReqs,
875 unsigned int DCCMetaPitch,
876 unsigned int *MacroTileWidth,
877 unsigned int *MetaRowByte,
878 unsigned int *PixelPTEBytesPerRow,
879 bool *PTEBufferSizeNotExceeded,
880 unsigned int *dpte_row_height,
881 unsigned int *meta_row_height)
883 unsigned int MetaRequestHeight;
884 unsigned int MetaRequestWidth;
885 unsigned int MetaSurfWidth;
886 unsigned int MetaSurfHeight;
887 unsigned int MPDEBytesFrame;
888 unsigned int MetaPTEBytesFrame;
889 unsigned int DCCMetaSurfaceBytes;
891 unsigned int MacroTileSizeBytes;
892 unsigned int MacroTileHeight;
893 unsigned int DPDE0BytesFrame;
894 unsigned int ExtraDPDEBytesFrame;
895 unsigned int PDEAndMetaPTEBytesFrame;
897 if (DCCEnable == true) {
898 MetaRequestHeight = 8 * BlockHeight256Bytes;
899 MetaRequestWidth = 8 * BlockWidth256Bytes;
900 if (ScanDirection == dm_horz) {
901 *meta_row_height = MetaRequestHeight;
902 MetaSurfWidth = dml_ceil((double) SwathWidth - 1, MetaRequestWidth)
904 *MetaRowByte = MetaSurfWidth * MetaRequestHeight * BytePerPixel / 256.0;
906 *meta_row_height = MetaRequestWidth;
907 MetaSurfHeight = dml_ceil((double) SwathWidth - 1, MetaRequestHeight)
909 *MetaRowByte = MetaSurfHeight * MetaRequestWidth * BytePerPixel / 256.0;
911 if (ScanDirection == dm_horz) {
912 DCCMetaSurfaceBytes = DCCMetaPitch
913 * (dml_ceil(ViewportHeight - 1, 64 * BlockHeight256Bytes)
914 + 64 * BlockHeight256Bytes) * BytePerPixel
917 DCCMetaSurfaceBytes = DCCMetaPitch
919 (double) ViewportHeight - 1,
920 64 * BlockHeight256Bytes)
921 + 64 * BlockHeight256Bytes) * BytePerPixel
924 if (GPUVMEnable == true) {
925 MetaPTEBytesFrame = (dml_ceil(
926 (double) (DCCMetaSurfaceBytes - VMMPageSize)
929 MPDEBytesFrame = 128 * (mode_lib->vba.GPUVMMaxPageTableLevels - 1);
931 MetaPTEBytesFrame = 0;
935 MetaPTEBytesFrame = 0;
940 if (SurfaceTiling == dm_sw_linear || SurfaceTiling == dm_sw_gfx7_2d_thin_gl || SurfaceTiling == dm_sw_gfx7_2d_thin_lvp) {
941 MacroTileSizeBytes = 256;
942 MacroTileHeight = BlockHeight256Bytes;
943 } else if (SurfaceTiling == dm_sw_4kb_s || SurfaceTiling == dm_sw_4kb_s_x
944 || SurfaceTiling == dm_sw_4kb_d || SurfaceTiling == dm_sw_4kb_d_x) {
945 MacroTileSizeBytes = 4096;
946 MacroTileHeight = 4 * BlockHeight256Bytes;
947 } else if (SurfaceTiling == dm_sw_64kb_s || SurfaceTiling == dm_sw_64kb_s_t
948 || SurfaceTiling == dm_sw_64kb_s_x || SurfaceTiling == dm_sw_64kb_d
949 || SurfaceTiling == dm_sw_64kb_d_t || SurfaceTiling == dm_sw_64kb_d_x
950 || SurfaceTiling == dm_sw_64kb_r_x) {
951 MacroTileSizeBytes = 65536;
952 MacroTileHeight = 16 * BlockHeight256Bytes;
954 MacroTileSizeBytes = 262144;
955 MacroTileHeight = 32 * BlockHeight256Bytes;
957 *MacroTileWidth = MacroTileSizeBytes / BytePerPixel / MacroTileHeight;
959 if (GPUVMEnable == true && mode_lib->vba.GPUVMMaxPageTableLevels > 1) {
960 if (ScanDirection == dm_horz) {
971 - MacroTileSizeBytes)
986 - MacroTileSizeBytes)
991 ExtraDPDEBytesFrame = 128 * (mode_lib->vba.GPUVMMaxPageTableLevels - 2);
994 ExtraDPDEBytesFrame = 0;
997 PDEAndMetaPTEBytesFrame = MetaPTEBytesFrame + MPDEBytesFrame + DPDE0BytesFrame
998 + ExtraDPDEBytesFrame;
1000 if (GPUVMEnable == true) {
1001 unsigned int PTERequestSize;
1002 unsigned int PixelPTEReqHeight;
1003 unsigned int PixelPTEReqWidth;
1004 double FractionOfPTEReturnDrop;
1005 unsigned int EffectivePDEProcessingBufIn64KBReqs;
1007 if (SurfaceTiling == dm_sw_linear) {
1008 PixelPTEReqHeight = 1;
1009 PixelPTEReqWidth = 8.0 * VMMPageSize / BytePerPixel;
1010 PTERequestSize = 64;
1011 FractionOfPTEReturnDrop = 0;
1012 } else if (MacroTileSizeBytes == 4096) {
1013 PixelPTEReqHeight = MacroTileHeight;
1014 PixelPTEReqWidth = 8 * *MacroTileWidth;
1015 PTERequestSize = 64;
1016 if (ScanDirection == dm_horz)
1017 FractionOfPTEReturnDrop = 0;
1019 FractionOfPTEReturnDrop = 7 / 8;
1020 } else if (VMMPageSize == 4096 && MacroTileSizeBytes > 4096) {
1021 PixelPTEReqHeight = 16 * BlockHeight256Bytes;
1022 PixelPTEReqWidth = 16 * BlockWidth256Bytes;
1023 PTERequestSize = 128;
1024 FractionOfPTEReturnDrop = 0;
1026 PixelPTEReqHeight = MacroTileHeight;
1027 PixelPTEReqWidth = 8 * *MacroTileWidth;
1028 PTERequestSize = 64;
1029 FractionOfPTEReturnDrop = 0;
1032 if (SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10)
1033 EffectivePDEProcessingBufIn64KBReqs = PDEProcessingBufIn64KBReqs / 2;
1035 EffectivePDEProcessingBufIn64KBReqs = PDEProcessingBufIn64KBReqs;
1037 if (SurfaceTiling == dm_sw_linear) {
1042 << (unsigned int) dml_floor(
1045 (double) PTEBufferSizeInRequestsLuma
1047 EffectivePDEProcessingBufIn64KBReqs
1052 *PixelPTEBytesPerRow = PTERequestSize
1054 (double) (Pitch * *dpte_row_height - 1)
1057 } else if (ScanDirection == dm_horz) {
1058 *dpte_row_height = PixelPTEReqHeight;
1059 *PixelPTEBytesPerRow = PTERequestSize
1060 * (dml_ceil(((double) SwathWidth - 1) / PixelPTEReqWidth, 1)
1063 *dpte_row_height = dml_min(PixelPTEReqWidth, *MacroTileWidth);
1064 *PixelPTEBytesPerRow = PTERequestSize
1066 ((double) SwathWidth - 1)
1067 / PixelPTEReqHeight,
1070 if (*PixelPTEBytesPerRow * (1 - FractionOfPTEReturnDrop)
1071 <= 64 * PTEBufferSizeInRequestsLuma) {
1072 *PTEBufferSizeNotExceeded = true;
1074 *PTEBufferSizeNotExceeded = false;
1077 *PixelPTEBytesPerRow = 0;
1078 *PTEBufferSizeNotExceeded = true;
1081 return PDEAndMetaPTEBytesFrame;
1084 static void dml20_DISPCLKDPPCLKDCFCLKDeepSleepPrefetchParametersWatermarksAndPerformanceCalculation(
1085 struct display_mode_lib *mode_lib)
1089 mode_lib->vba.WritebackDISPCLK = 0.0;
1090 mode_lib->vba.DISPCLKWithRamping = 0;
1091 mode_lib->vba.DISPCLKWithoutRamping = 0;
1092 mode_lib->vba.GlobalDPPCLK = 0.0;
1094 // dml_ml->vba.DISPCLK and dml_ml->vba.DPPCLK Calculation
1096 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
1097 if (mode_lib->vba.WritebackEnable[k]) {
1098 mode_lib->vba.WritebackDISPCLK =
1100 mode_lib->vba.WritebackDISPCLK,
1101 CalculateWriteBackDISPCLK(
1102 mode_lib->vba.WritebackPixelFormat[k],
1103 mode_lib->vba.PixelClock[k],
1104 mode_lib->vba.WritebackHRatio[k],
1105 mode_lib->vba.WritebackVRatio[k],
1106 mode_lib->vba.WritebackLumaHTaps[k],
1107 mode_lib->vba.WritebackLumaVTaps[k],
1108 mode_lib->vba.WritebackChromaHTaps[k],
1109 mode_lib->vba.WritebackChromaVTaps[k],
1110 mode_lib->vba.WritebackDestinationWidth[k],
1111 mode_lib->vba.HTotal[k],
1112 mode_lib->vba.WritebackChromaLineBufferWidth));
1116 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
1117 if (mode_lib->vba.HRatio[k] > 1) {
1118 mode_lib->vba.PSCL_THROUGHPUT_LUMA[k] = dml_min(
1119 mode_lib->vba.MaxDCHUBToPSCLThroughput,
1120 mode_lib->vba.MaxPSCLToLBThroughput
1121 * mode_lib->vba.HRatio[k]
1123 mode_lib->vba.htaps[k]
1127 mode_lib->vba.PSCL_THROUGHPUT_LUMA[k] = dml_min(
1128 mode_lib->vba.MaxDCHUBToPSCLThroughput,
1129 mode_lib->vba.MaxPSCLToLBThroughput);
1132 mode_lib->vba.DPPCLKUsingSingleDPPLuma =
1133 mode_lib->vba.PixelClock[k]
1135 mode_lib->vba.vtaps[k] / 6.0
1138 mode_lib->vba.HRatio[k]),
1140 mode_lib->vba.HRatio[k]
1141 * mode_lib->vba.VRatio[k]
1142 / mode_lib->vba.PSCL_THROUGHPUT_LUMA[k],
1145 if ((mode_lib->vba.htaps[k] > 6 || mode_lib->vba.vtaps[k] > 6)
1146 && mode_lib->vba.DPPCLKUsingSingleDPPLuma
1147 < 2 * mode_lib->vba.PixelClock[k]) {
1148 mode_lib->vba.DPPCLKUsingSingleDPPLuma = 2 * mode_lib->vba.PixelClock[k];
1151 if ((mode_lib->vba.SourcePixelFormat[k] != dm_420_8
1152 && mode_lib->vba.SourcePixelFormat[k] != dm_420_10)) {
1153 mode_lib->vba.PSCL_THROUGHPUT_CHROMA[k] = 0.0;
1154 mode_lib->vba.DPPCLKUsingSingleDPP[k] =
1155 mode_lib->vba.DPPCLKUsingSingleDPPLuma;
1157 if (mode_lib->vba.HRatio[k] > 1) {
1158 mode_lib->vba.PSCL_THROUGHPUT_CHROMA[k] =
1160 mode_lib->vba.MaxDCHUBToPSCLThroughput,
1161 mode_lib->vba.MaxPSCLToLBThroughput
1162 * mode_lib->vba.HRatio[k]
1165 mode_lib->vba.HTAPsChroma[k]
1169 mode_lib->vba.PSCL_THROUGHPUT_CHROMA[k] = dml_min(
1170 mode_lib->vba.MaxDCHUBToPSCLThroughput,
1171 mode_lib->vba.MaxPSCLToLBThroughput);
1173 mode_lib->vba.DPPCLKUsingSingleDPPChroma =
1174 mode_lib->vba.PixelClock[k]
1176 mode_lib->vba.VTAPsChroma[k]
1180 mode_lib->vba.HRatio[k]
1183 mode_lib->vba.HRatio[k]
1184 * mode_lib->vba.VRatio[k]
1186 / mode_lib->vba.PSCL_THROUGHPUT_CHROMA[k],
1189 if ((mode_lib->vba.HTAPsChroma[k] > 6 || mode_lib->vba.VTAPsChroma[k] > 6)
1190 && mode_lib->vba.DPPCLKUsingSingleDPPChroma
1191 < 2 * mode_lib->vba.PixelClock[k]) {
1192 mode_lib->vba.DPPCLKUsingSingleDPPChroma = 2
1193 * mode_lib->vba.PixelClock[k];
1196 mode_lib->vba.DPPCLKUsingSingleDPP[k] = dml_max(
1197 mode_lib->vba.DPPCLKUsingSingleDPPLuma,
1198 mode_lib->vba.DPPCLKUsingSingleDPPChroma);
1202 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
1203 if (mode_lib->vba.BlendingAndTiming[k] != k)
1205 if (mode_lib->vba.ODMCombineEnabled[k]) {
1206 mode_lib->vba.DISPCLKWithRamping =
1208 mode_lib->vba.DISPCLKWithRamping,
1209 mode_lib->vba.PixelClock[k] / 2
1211 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading
1214 + mode_lib->vba.DISPCLKRampingMargin
1216 mode_lib->vba.DISPCLKWithoutRamping =
1218 mode_lib->vba.DISPCLKWithoutRamping,
1219 mode_lib->vba.PixelClock[k] / 2
1221 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading
1223 } else if (!mode_lib->vba.ODMCombineEnabled[k]) {
1224 mode_lib->vba.DISPCLKWithRamping =
1226 mode_lib->vba.DISPCLKWithRamping,
1227 mode_lib->vba.PixelClock[k]
1229 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading
1232 + mode_lib->vba.DISPCLKRampingMargin
1234 mode_lib->vba.DISPCLKWithoutRamping =
1236 mode_lib->vba.DISPCLKWithoutRamping,
1237 mode_lib->vba.PixelClock[k]
1239 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading
1244 mode_lib->vba.DISPCLKWithRamping = dml_max(
1245 mode_lib->vba.DISPCLKWithRamping,
1246 mode_lib->vba.WritebackDISPCLK);
1247 mode_lib->vba.DISPCLKWithoutRamping = dml_max(
1248 mode_lib->vba.DISPCLKWithoutRamping,
1249 mode_lib->vba.WritebackDISPCLK);
1251 ASSERT(mode_lib->vba.DISPCLKDPPCLKVCOSpeed != 0);
1252 mode_lib->vba.DISPCLKWithRampingRoundedToDFSGranularity = RoundToDFSGranularityUp(
1253 mode_lib->vba.DISPCLKWithRamping,
1254 mode_lib->vba.DISPCLKDPPCLKVCOSpeed);
1255 mode_lib->vba.DISPCLKWithoutRampingRoundedToDFSGranularity = RoundToDFSGranularityUp(
1256 mode_lib->vba.DISPCLKWithoutRamping,
1257 mode_lib->vba.DISPCLKDPPCLKVCOSpeed);
1258 mode_lib->vba.MaxDispclkRoundedToDFSGranularity = RoundToDFSGranularityDown(
1259 mode_lib->vba.soc.clock_limits[mode_lib->vba.soc.num_states].dispclk_mhz,
1260 mode_lib->vba.DISPCLKDPPCLKVCOSpeed);
1261 if (mode_lib->vba.DISPCLKWithoutRampingRoundedToDFSGranularity
1262 > mode_lib->vba.MaxDispclkRoundedToDFSGranularity) {
1263 mode_lib->vba.DISPCLK_calculated =
1264 mode_lib->vba.DISPCLKWithoutRampingRoundedToDFSGranularity;
1265 } else if (mode_lib->vba.DISPCLKWithRampingRoundedToDFSGranularity
1266 > mode_lib->vba.MaxDispclkRoundedToDFSGranularity) {
1267 mode_lib->vba.DISPCLK_calculated = mode_lib->vba.MaxDispclkRoundedToDFSGranularity;
1269 mode_lib->vba.DISPCLK_calculated =
1270 mode_lib->vba.DISPCLKWithRampingRoundedToDFSGranularity;
1272 DTRACE(" dispclk_mhz (calculated) = %f", mode_lib->vba.DISPCLK_calculated);
1274 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
1275 if (mode_lib->vba.DPPPerPlane[k] == 0) {
1276 mode_lib->vba.DPPCLK_calculated[k] = 0;
1278 mode_lib->vba.DPPCLK_calculated[k] = mode_lib->vba.DPPCLKUsingSingleDPP[k]
1279 / mode_lib->vba.DPPPerPlane[k]
1280 * (1 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100);
1282 mode_lib->vba.GlobalDPPCLK = dml_max(
1283 mode_lib->vba.GlobalDPPCLK,
1284 mode_lib->vba.DPPCLK_calculated[k]);
1286 mode_lib->vba.GlobalDPPCLK = RoundToDFSGranularityUp(
1287 mode_lib->vba.GlobalDPPCLK,
1288 mode_lib->vba.DISPCLKDPPCLKVCOSpeed);
1289 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
1290 mode_lib->vba.DPPCLK_calculated[k] = mode_lib->vba.GlobalDPPCLK / 255
1292 mode_lib->vba.DPPCLK_calculated[k] * 255
1293 / mode_lib->vba.GlobalDPPCLK,
1295 DTRACE(" dppclk_mhz[%i] (calculated) = %f", k, mode_lib->vba.DPPCLK_calculated[k]);
1299 mode_lib->vba.DCCEnabledAnyPlane = false;
1300 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k)
1301 if (mode_lib->vba.DCCEnable[k])
1302 mode_lib->vba.DCCEnabledAnyPlane = true;
1304 mode_lib->vba.ReturnBandwidthToDCN = dml_min(
1305 mode_lib->vba.ReturnBusWidth * mode_lib->vba.DCFCLK,
1306 mode_lib->vba.FabricAndDRAMBandwidth * 1000)
1307 * mode_lib->vba.PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelDataOnly / 100;
1309 mode_lib->vba.ReturnBW = mode_lib->vba.ReturnBandwidthToDCN;
1310 mode_lib->vba.ReturnBW = adjust_ReturnBW(
1312 mode_lib->vba.ReturnBW,
1313 mode_lib->vba.DCCEnabledAnyPlane,
1314 mode_lib->vba.ReturnBandwidthToDCN);
1316 // Let's do this calculation again??
1317 mode_lib->vba.ReturnBandwidthToDCN = dml_min(
1318 mode_lib->vba.ReturnBusWidth * mode_lib->vba.DCFCLK,
1319 mode_lib->vba.FabricAndDRAMBandwidth * 1000);
1320 mode_lib->vba.ReturnBW = adjust_ReturnBW(
1322 mode_lib->vba.ReturnBW,
1323 mode_lib->vba.DCCEnabledAnyPlane,
1324 mode_lib->vba.ReturnBandwidthToDCN);
1326 DTRACE(" dcfclk_mhz = %f", mode_lib->vba.DCFCLK);
1327 DTRACE(" return_bw_to_dcn = %f", mode_lib->vba.ReturnBandwidthToDCN);
1328 DTRACE(" return_bus_bw = %f", mode_lib->vba.ReturnBW);
1330 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
1331 bool MainPlaneDoesODMCombine = false;
1333 if (mode_lib->vba.SourceScan[k] == dm_horz)
1334 mode_lib->vba.SwathWidthSingleDPPY[k] = mode_lib->vba.ViewportWidth[k];
1336 mode_lib->vba.SwathWidthSingleDPPY[k] = mode_lib->vba.ViewportHeight[k];
1338 if (mode_lib->vba.ODMCombineEnabled[k] == true)
1339 MainPlaneDoesODMCombine = true;
1340 for (j = 0; j < mode_lib->vba.NumberOfActivePlanes; ++j)
1341 if (mode_lib->vba.BlendingAndTiming[k] == j
1342 && mode_lib->vba.ODMCombineEnabled[j] == true)
1343 MainPlaneDoesODMCombine = true;
1345 if (MainPlaneDoesODMCombine == true)
1346 mode_lib->vba.SwathWidthY[k] = dml_min(
1347 (double) mode_lib->vba.SwathWidthSingleDPPY[k],
1349 mode_lib->vba.HActive[k] / 2.0
1350 * mode_lib->vba.HRatio[k]));
1352 if (mode_lib->vba.DPPPerPlane[k] == 0) {
1353 mode_lib->vba.SwathWidthY[k] = 0;
1355 mode_lib->vba.SwathWidthY[k] = mode_lib->vba.SwathWidthSingleDPPY[k]
1356 / mode_lib->vba.DPPPerPlane[k];
1361 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
1362 if (mode_lib->vba.SourcePixelFormat[k] == dm_444_64) {
1363 mode_lib->vba.BytePerPixelDETY[k] = 8;
1364 mode_lib->vba.BytePerPixelDETC[k] = 0;
1365 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_32) {
1366 mode_lib->vba.BytePerPixelDETY[k] = 4;
1367 mode_lib->vba.BytePerPixelDETC[k] = 0;
1368 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_16) {
1369 mode_lib->vba.BytePerPixelDETY[k] = 2;
1370 mode_lib->vba.BytePerPixelDETC[k] = 0;
1371 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_8) {
1372 mode_lib->vba.BytePerPixelDETY[k] = 1;
1373 mode_lib->vba.BytePerPixelDETC[k] = 0;
1374 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_8) {
1375 mode_lib->vba.BytePerPixelDETY[k] = 1;
1376 mode_lib->vba.BytePerPixelDETC[k] = 2;
1377 } else { // dm_420_10
1378 mode_lib->vba.BytePerPixelDETY[k] = 4.0 / 3.0;
1379 mode_lib->vba.BytePerPixelDETC[k] = 8.0 / 3.0;
1383 mode_lib->vba.TotalDataReadBandwidth = 0.0;
1384 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
1385 mode_lib->vba.ReadBandwidthPlaneLuma[k] = mode_lib->vba.SwathWidthSingleDPPY[k]
1386 * dml_ceil(mode_lib->vba.BytePerPixelDETY[k], 1)
1387 / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k])
1388 * mode_lib->vba.VRatio[k];
1389 mode_lib->vba.ReadBandwidthPlaneChroma[k] = mode_lib->vba.SwathWidthSingleDPPY[k]
1390 / 2 * dml_ceil(mode_lib->vba.BytePerPixelDETC[k], 2)
1391 / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k])
1392 * mode_lib->vba.VRatio[k] / 2;
1394 " read_bw[%i] = %fBps",
1396 mode_lib->vba.ReadBandwidthPlaneLuma[k]
1397 + mode_lib->vba.ReadBandwidthPlaneChroma[k]);
1398 mode_lib->vba.TotalDataReadBandwidth += mode_lib->vba.ReadBandwidthPlaneLuma[k]
1399 + mode_lib->vba.ReadBandwidthPlaneChroma[k];
1402 mode_lib->vba.TotalDCCActiveDPP = 0;
1403 mode_lib->vba.TotalActiveDPP = 0;
1404 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
1405 mode_lib->vba.TotalActiveDPP = mode_lib->vba.TotalActiveDPP
1406 + mode_lib->vba.DPPPerPlane[k];
1407 if (mode_lib->vba.DCCEnable[k])
1408 mode_lib->vba.TotalDCCActiveDPP = mode_lib->vba.TotalDCCActiveDPP
1409 + mode_lib->vba.DPPPerPlane[k];
1412 mode_lib->vba.UrgentRoundTripAndOutOfOrderLatency =
1413 (mode_lib->vba.RoundTripPingLatencyCycles + 32) / mode_lib->vba.DCFCLK
1414 + mode_lib->vba.UrgentOutOfOrderReturnPerChannelPixelDataOnly
1415 * mode_lib->vba.NumberOfChannels
1416 / mode_lib->vba.ReturnBW;
1418 mode_lib->vba.LastPixelOfLineExtraWatermark = 0;
1419 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
1420 double DataFabricLineDeliveryTimeLuma, DataFabricLineDeliveryTimeChroma;
1422 if (mode_lib->vba.VRatio[k] <= 1.0)
1423 mode_lib->vba.DisplayPipeLineDeliveryTimeLuma[k] =
1424 (double) mode_lib->vba.SwathWidthY[k]
1425 * mode_lib->vba.DPPPerPlane[k]
1426 / mode_lib->vba.HRatio[k]
1427 / mode_lib->vba.PixelClock[k];
1429 mode_lib->vba.DisplayPipeLineDeliveryTimeLuma[k] =
1430 (double) mode_lib->vba.SwathWidthY[k]
1431 / mode_lib->vba.PSCL_THROUGHPUT_LUMA[k]
1432 / mode_lib->vba.DPPCLK[k];
1434 DataFabricLineDeliveryTimeLuma = mode_lib->vba.SwathWidthSingleDPPY[k]
1435 * mode_lib->vba.SwathHeightY[k]
1436 * dml_ceil(mode_lib->vba.BytePerPixelDETY[k], 1)
1437 / (mode_lib->vba.ReturnBW * mode_lib->vba.ReadBandwidthPlaneLuma[k]
1438 / mode_lib->vba.TotalDataReadBandwidth);
1439 mode_lib->vba.LastPixelOfLineExtraWatermark = dml_max(
1440 mode_lib->vba.LastPixelOfLineExtraWatermark,
1441 DataFabricLineDeliveryTimeLuma
1442 - mode_lib->vba.DisplayPipeLineDeliveryTimeLuma[k]);
1444 if (mode_lib->vba.BytePerPixelDETC[k] == 0)
1445 mode_lib->vba.DisplayPipeLineDeliveryTimeChroma[k] = 0.0;
1446 else if (mode_lib->vba.VRatio[k] / 2.0 <= 1.0)
1447 mode_lib->vba.DisplayPipeLineDeliveryTimeChroma[k] =
1448 mode_lib->vba.SwathWidthY[k] / 2.0
1449 * mode_lib->vba.DPPPerPlane[k]
1450 / (mode_lib->vba.HRatio[k] / 2.0)
1451 / mode_lib->vba.PixelClock[k];
1453 mode_lib->vba.DisplayPipeLineDeliveryTimeChroma[k] =
1454 mode_lib->vba.SwathWidthY[k] / 2.0
1455 / mode_lib->vba.PSCL_THROUGHPUT_CHROMA[k]
1456 / mode_lib->vba.DPPCLK[k];
1458 DataFabricLineDeliveryTimeChroma = mode_lib->vba.SwathWidthSingleDPPY[k] / 2.0
1459 * mode_lib->vba.SwathHeightC[k]
1460 * dml_ceil(mode_lib->vba.BytePerPixelDETC[k], 2)
1461 / (mode_lib->vba.ReturnBW
1462 * mode_lib->vba.ReadBandwidthPlaneChroma[k]
1463 / mode_lib->vba.TotalDataReadBandwidth);
1464 mode_lib->vba.LastPixelOfLineExtraWatermark =
1466 mode_lib->vba.LastPixelOfLineExtraWatermark,
1467 DataFabricLineDeliveryTimeChroma
1468 - mode_lib->vba.DisplayPipeLineDeliveryTimeChroma[k]);
1471 mode_lib->vba.UrgentExtraLatency = mode_lib->vba.UrgentRoundTripAndOutOfOrderLatency
1472 + (mode_lib->vba.TotalActiveDPP * mode_lib->vba.PixelChunkSizeInKByte
1473 + mode_lib->vba.TotalDCCActiveDPP
1474 * mode_lib->vba.MetaChunkSize) * 1024.0
1475 / mode_lib->vba.ReturnBW;
1477 if (mode_lib->vba.GPUVMEnable)
1478 mode_lib->vba.UrgentExtraLatency += mode_lib->vba.TotalActiveDPP
1479 * mode_lib->vba.PTEGroupSize / mode_lib->vba.ReturnBW;
1481 mode_lib->vba.UrgentWatermark = mode_lib->vba.UrgentLatencyPixelDataOnly
1482 + mode_lib->vba.LastPixelOfLineExtraWatermark
1483 + mode_lib->vba.UrgentExtraLatency;
1485 DTRACE(" urgent_extra_latency = %fus", mode_lib->vba.UrgentExtraLatency);
1486 DTRACE(" wm_urgent = %fus", mode_lib->vba.UrgentWatermark);
1488 mode_lib->vba.UrgentLatency = mode_lib->vba.UrgentLatencyPixelDataOnly;
1490 mode_lib->vba.TotalActiveWriteback = 0;
1491 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
1492 if (mode_lib->vba.WritebackEnable[k])
1493 mode_lib->vba.TotalActiveWriteback = mode_lib->vba.TotalActiveWriteback + mode_lib->vba.ActiveWritebacksPerPlane[k];
1496 if (mode_lib->vba.TotalActiveWriteback <= 1)
1497 mode_lib->vba.WritebackUrgentWatermark = mode_lib->vba.WritebackLatency;
1499 mode_lib->vba.WritebackUrgentWatermark = mode_lib->vba.WritebackLatency
1500 + mode_lib->vba.WritebackChunkSize * 1024.0 / 32
1501 / mode_lib->vba.SOCCLK;
1503 DTRACE(" wm_wb_urgent = %fus", mode_lib->vba.WritebackUrgentWatermark);
1505 // NB P-State/DRAM Clock Change Watermark
1506 mode_lib->vba.DRAMClockChangeWatermark = mode_lib->vba.DRAMClockChangeLatency
1507 + mode_lib->vba.UrgentWatermark;
1509 DTRACE(" wm_pstate_change = %fus", mode_lib->vba.DRAMClockChangeWatermark);
1511 DTRACE(" calculating wb pstate watermark");
1512 DTRACE(" total wb outputs %d", mode_lib->vba.TotalActiveWriteback);
1513 DTRACE(" socclk frequency %f Mhz", mode_lib->vba.SOCCLK);
1515 if (mode_lib->vba.TotalActiveWriteback <= 1)
1516 mode_lib->vba.WritebackDRAMClockChangeWatermark =
1517 mode_lib->vba.DRAMClockChangeLatency
1518 + mode_lib->vba.WritebackLatency;
1520 mode_lib->vba.WritebackDRAMClockChangeWatermark =
1521 mode_lib->vba.DRAMClockChangeLatency
1522 + mode_lib->vba.WritebackLatency
1523 + mode_lib->vba.WritebackChunkSize * 1024.0 / 32
1524 / mode_lib->vba.SOCCLK;
1526 DTRACE(" wm_wb_pstate %fus", mode_lib->vba.WritebackDRAMClockChangeWatermark);
1528 // Stutter Efficiency
1529 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
1530 mode_lib->vba.LinesInDETY[k] = mode_lib->vba.DETBufferSizeY[k]
1531 / mode_lib->vba.BytePerPixelDETY[k] / mode_lib->vba.SwathWidthY[k];
1532 mode_lib->vba.LinesInDETYRoundedDownToSwath[k] = dml_floor(
1533 mode_lib->vba.LinesInDETY[k],
1534 mode_lib->vba.SwathHeightY[k]);
1535 mode_lib->vba.FullDETBufferingTimeY[k] =
1536 mode_lib->vba.LinesInDETYRoundedDownToSwath[k]
1537 * (mode_lib->vba.HTotal[k]
1538 / mode_lib->vba.PixelClock[k])
1539 / mode_lib->vba.VRatio[k];
1540 if (mode_lib->vba.BytePerPixelDETC[k] > 0) {
1541 mode_lib->vba.LinesInDETC[k] = mode_lib->vba.DETBufferSizeC[k]
1542 / mode_lib->vba.BytePerPixelDETC[k]
1543 / (mode_lib->vba.SwathWidthY[k] / 2);
1544 mode_lib->vba.LinesInDETCRoundedDownToSwath[k] = dml_floor(
1545 mode_lib->vba.LinesInDETC[k],
1546 mode_lib->vba.SwathHeightC[k]);
1547 mode_lib->vba.FullDETBufferingTimeC[k] =
1548 mode_lib->vba.LinesInDETCRoundedDownToSwath[k]
1549 * (mode_lib->vba.HTotal[k]
1550 / mode_lib->vba.PixelClock[k])
1551 / (mode_lib->vba.VRatio[k] / 2);
1553 mode_lib->vba.LinesInDETC[k] = 0;
1554 mode_lib->vba.LinesInDETCRoundedDownToSwath[k] = 0;
1555 mode_lib->vba.FullDETBufferingTimeC[k] = 999999;
1559 mode_lib->vba.MinFullDETBufferingTime = 999999.0;
1560 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
1561 if (mode_lib->vba.FullDETBufferingTimeY[k]
1562 < mode_lib->vba.MinFullDETBufferingTime) {
1563 mode_lib->vba.MinFullDETBufferingTime =
1564 mode_lib->vba.FullDETBufferingTimeY[k];
1565 mode_lib->vba.FrameTimeForMinFullDETBufferingTime =
1566 (double) mode_lib->vba.VTotal[k] * mode_lib->vba.HTotal[k]
1567 / mode_lib->vba.PixelClock[k];
1569 if (mode_lib->vba.FullDETBufferingTimeC[k]
1570 < mode_lib->vba.MinFullDETBufferingTime) {
1571 mode_lib->vba.MinFullDETBufferingTime =
1572 mode_lib->vba.FullDETBufferingTimeC[k];
1573 mode_lib->vba.FrameTimeForMinFullDETBufferingTime =
1574 (double) mode_lib->vba.VTotal[k] * mode_lib->vba.HTotal[k]
1575 / mode_lib->vba.PixelClock[k];
1579 mode_lib->vba.AverageReadBandwidthGBytePerSecond = 0.0;
1580 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
1581 if (mode_lib->vba.DCCEnable[k]) {
1582 mode_lib->vba.AverageReadBandwidthGBytePerSecond =
1583 mode_lib->vba.AverageReadBandwidthGBytePerSecond
1584 + mode_lib->vba.ReadBandwidthPlaneLuma[k]
1585 / mode_lib->vba.DCCRate[k]
1587 + mode_lib->vba.ReadBandwidthPlaneChroma[k]
1588 / mode_lib->vba.DCCRate[k]
1591 mode_lib->vba.AverageReadBandwidthGBytePerSecond =
1592 mode_lib->vba.AverageReadBandwidthGBytePerSecond
1593 + mode_lib->vba.ReadBandwidthPlaneLuma[k]
1595 + mode_lib->vba.ReadBandwidthPlaneChroma[k]
1598 if (mode_lib->vba.DCCEnable[k]) {
1599 mode_lib->vba.AverageReadBandwidthGBytePerSecond =
1600 mode_lib->vba.AverageReadBandwidthGBytePerSecond
1601 + mode_lib->vba.ReadBandwidthPlaneLuma[k]
1603 + mode_lib->vba.ReadBandwidthPlaneChroma[k]
1606 if (mode_lib->vba.GPUVMEnable) {
1607 mode_lib->vba.AverageReadBandwidthGBytePerSecond =
1608 mode_lib->vba.AverageReadBandwidthGBytePerSecond
1609 + mode_lib->vba.ReadBandwidthPlaneLuma[k]
1611 + mode_lib->vba.ReadBandwidthPlaneChroma[k]
1616 mode_lib->vba.PartOfBurstThatFitsInROB =
1618 mode_lib->vba.MinFullDETBufferingTime
1619 * mode_lib->vba.TotalDataReadBandwidth,
1620 mode_lib->vba.ROBBufferSizeInKByte * 1024
1621 * mode_lib->vba.TotalDataReadBandwidth
1622 / (mode_lib->vba.AverageReadBandwidthGBytePerSecond
1624 mode_lib->vba.StutterBurstTime = mode_lib->vba.PartOfBurstThatFitsInROB
1625 * (mode_lib->vba.AverageReadBandwidthGBytePerSecond * 1000)
1626 / mode_lib->vba.TotalDataReadBandwidth / mode_lib->vba.ReturnBW
1627 + (mode_lib->vba.MinFullDETBufferingTime
1628 * mode_lib->vba.TotalDataReadBandwidth
1629 - mode_lib->vba.PartOfBurstThatFitsInROB)
1630 / (mode_lib->vba.DCFCLK * 64);
1631 if (mode_lib->vba.TotalActiveWriteback == 0) {
1632 mode_lib->vba.StutterEfficiencyNotIncludingVBlank = (1
1633 - (mode_lib->vba.SRExitTime + mode_lib->vba.StutterBurstTime)
1634 / mode_lib->vba.MinFullDETBufferingTime) * 100;
1636 mode_lib->vba.StutterEfficiencyNotIncludingVBlank = 0;
1639 mode_lib->vba.SmallestVBlank = 999999;
1640 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
1641 if (mode_lib->vba.SynchronizedVBlank || mode_lib->vba.NumberOfActivePlanes == 1) {
1642 mode_lib->vba.VBlankTime = (double) (mode_lib->vba.VTotal[k]
1643 - mode_lib->vba.VActive[k]) * mode_lib->vba.HTotal[k]
1644 / mode_lib->vba.PixelClock[k];
1646 mode_lib->vba.VBlankTime = 0;
1648 mode_lib->vba.SmallestVBlank = dml_min(
1649 mode_lib->vba.SmallestVBlank,
1650 mode_lib->vba.VBlankTime);
1653 mode_lib->vba.StutterEfficiency = (mode_lib->vba.StutterEfficiencyNotIncludingVBlank / 100
1654 * (mode_lib->vba.FrameTimeForMinFullDETBufferingTime
1655 - mode_lib->vba.SmallestVBlank)
1656 + mode_lib->vba.SmallestVBlank)
1657 / mode_lib->vba.FrameTimeForMinFullDETBufferingTime * 100;
1659 // dml_ml->vba.DCFCLK Deep Sleep
1660 mode_lib->vba.DCFCLKDeepSleep = 8.0;
1662 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; k++) {
1663 if (mode_lib->vba.BytePerPixelDETC[k] > 0) {
1664 mode_lib->vba.DCFCLKDeepSleepPerPlane[k] =
1666 1.1 * mode_lib->vba.SwathWidthY[k]
1668 mode_lib->vba.BytePerPixelDETY[k],
1670 / mode_lib->vba.DisplayPipeLineDeliveryTimeLuma[k],
1671 1.1 * mode_lib->vba.SwathWidthY[k] / 2.0
1673 mode_lib->vba.BytePerPixelDETC[k],
1675 / mode_lib->vba.DisplayPipeLineDeliveryTimeChroma[k]);
1677 mode_lib->vba.DCFCLKDeepSleepPerPlane[k] = 1.1 * mode_lib->vba.SwathWidthY[k]
1678 * dml_ceil(mode_lib->vba.BytePerPixelDETY[k], 1) / 64.0
1679 / mode_lib->vba.DisplayPipeLineDeliveryTimeLuma[k];
1680 mode_lib->vba.DCFCLKDeepSleepPerPlane[k] = dml_max(
1681 mode_lib->vba.DCFCLKDeepSleepPerPlane[k],
1682 mode_lib->vba.PixelClock[k] / 16.0);
1683 mode_lib->vba.DCFCLKDeepSleep = dml_max(
1684 mode_lib->vba.DCFCLKDeepSleep,
1685 mode_lib->vba.DCFCLKDeepSleepPerPlane[k]);
1688 " dcfclk_deepsleep_per_plane[%i] = %fMHz",
1690 mode_lib->vba.DCFCLKDeepSleepPerPlane[k]);
1693 DTRACE(" dcfclk_deepsleep_mhz = %fMHz", mode_lib->vba.DCFCLKDeepSleep);
1695 // Stutter Watermark
1696 mode_lib->vba.StutterExitWatermark = mode_lib->vba.SRExitTime
1697 + mode_lib->vba.LastPixelOfLineExtraWatermark
1698 + mode_lib->vba.UrgentExtraLatency + 10 / mode_lib->vba.DCFCLKDeepSleep;
1699 mode_lib->vba.StutterEnterPlusExitWatermark = mode_lib->vba.SREnterPlusExitTime
1700 + mode_lib->vba.LastPixelOfLineExtraWatermark
1701 + mode_lib->vba.UrgentExtraLatency;
1703 DTRACE(" wm_cstate_exit = %fus", mode_lib->vba.StutterExitWatermark);
1704 DTRACE(" wm_cstate_enter_exit = %fus", mode_lib->vba.StutterEnterPlusExitWatermark);
1706 // Urgent Latency Supported
1707 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
1708 mode_lib->vba.EffectiveDETPlusLBLinesLuma =
1710 mode_lib->vba.LinesInDETY[k]
1712 mode_lib->vba.LinesInDETY[k]
1713 * mode_lib->vba.DPPCLK[k]
1714 * mode_lib->vba.BytePerPixelDETY[k]
1715 * mode_lib->vba.PSCL_THROUGHPUT_LUMA[k]
1716 / (mode_lib->vba.ReturnBW
1717 / mode_lib->vba.DPPPerPlane[k]),
1718 (double) mode_lib->vba.EffectiveLBLatencyHidingSourceLinesLuma),
1719 mode_lib->vba.SwathHeightY[k]);
1721 mode_lib->vba.UrgentLatencySupportUsLuma = mode_lib->vba.EffectiveDETPlusLBLinesLuma
1722 * (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k])
1723 / mode_lib->vba.VRatio[k]
1724 - mode_lib->vba.EffectiveDETPlusLBLinesLuma
1725 * mode_lib->vba.SwathWidthY[k]
1726 * mode_lib->vba.BytePerPixelDETY[k]
1727 / (mode_lib->vba.ReturnBW
1728 / mode_lib->vba.DPPPerPlane[k]);
1730 if (mode_lib->vba.BytePerPixelDETC[k] > 0) {
1731 mode_lib->vba.EffectiveDETPlusLBLinesChroma =
1733 mode_lib->vba.LinesInDETC[k]
1735 mode_lib->vba.LinesInDETC[k]
1736 * mode_lib->vba.DPPCLK[k]
1737 * mode_lib->vba.BytePerPixelDETC[k]
1738 * mode_lib->vba.PSCL_THROUGHPUT_CHROMA[k]
1739 / (mode_lib->vba.ReturnBW
1740 / mode_lib->vba.DPPPerPlane[k]),
1741 (double) mode_lib->vba.EffectiveLBLatencyHidingSourceLinesChroma),
1742 mode_lib->vba.SwathHeightC[k]);
1743 mode_lib->vba.UrgentLatencySupportUsChroma =
1744 mode_lib->vba.EffectiveDETPlusLBLinesChroma
1745 * (mode_lib->vba.HTotal[k]
1746 / mode_lib->vba.PixelClock[k])
1747 / (mode_lib->vba.VRatio[k] / 2)
1748 - mode_lib->vba.EffectiveDETPlusLBLinesChroma
1749 * (mode_lib->vba.SwathWidthY[k]
1751 * mode_lib->vba.BytePerPixelDETC[k]
1752 / (mode_lib->vba.ReturnBW
1753 / mode_lib->vba.DPPPerPlane[k]);
1754 mode_lib->vba.UrgentLatencySupportUs[k] = dml_min(
1755 mode_lib->vba.UrgentLatencySupportUsLuma,
1756 mode_lib->vba.UrgentLatencySupportUsChroma);
1758 mode_lib->vba.UrgentLatencySupportUs[k] =
1759 mode_lib->vba.UrgentLatencySupportUsLuma;
1763 mode_lib->vba.MinUrgentLatencySupportUs = 999999;
1764 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
1765 mode_lib->vba.MinUrgentLatencySupportUs = dml_min(
1766 mode_lib->vba.MinUrgentLatencySupportUs,
1767 mode_lib->vba.UrgentLatencySupportUs[k]);
1770 // Non-Urgent Latency Tolerance
1771 mode_lib->vba.NonUrgentLatencyTolerance = mode_lib->vba.MinUrgentLatencySupportUs
1772 - mode_lib->vba.UrgentWatermark;
1775 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
1776 if ((mode_lib->vba.BlendingAndTiming[k] != k) || !mode_lib->vba.DSCEnabled[k]) {
1777 mode_lib->vba.DSCCLK_calculated[k] = 0.0;
1779 if (mode_lib->vba.OutputFormat[k] == dm_420
1780 || mode_lib->vba.OutputFormat[k] == dm_n422)
1781 mode_lib->vba.DSCFormatFactor = 2;
1783 mode_lib->vba.DSCFormatFactor = 1;
1784 if (mode_lib->vba.ODMCombineEnabled[k])
1785 mode_lib->vba.DSCCLK_calculated[k] =
1786 mode_lib->vba.PixelClockBackEnd[k] / 6
1787 / mode_lib->vba.DSCFormatFactor
1789 - mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading
1792 mode_lib->vba.DSCCLK_calculated[k] =
1793 mode_lib->vba.PixelClockBackEnd[k] / 3
1794 / mode_lib->vba.DSCFormatFactor
1796 - mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading
1803 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
1804 double bpp = mode_lib->vba.OutputBpp[k];
1805 unsigned int slices = mode_lib->vba.NumberOfDSCSlices[k];
1807 if (mode_lib->vba.DSCEnabled[k] && bpp != 0) {
1808 if (!mode_lib->vba.ODMCombineEnabled[k]) {
1809 mode_lib->vba.DSCDelay[k] =
1811 mode_lib->vba.DSCInputBitPerComponent[k],
1814 (double) mode_lib->vba.HActive[k]
1815 / mode_lib->vba.NumberOfDSCSlices[k],
1818 mode_lib->vba.OutputFormat[k])
1820 mode_lib->vba.OutputFormat[k]);
1822 mode_lib->vba.DSCDelay[k] =
1824 * (dscceComputeDelay(
1825 mode_lib->vba.DSCInputBitPerComponent[k],
1828 (double) mode_lib->vba.HActive[k]
1829 / mode_lib->vba.NumberOfDSCSlices[k],
1832 mode_lib->vba.OutputFormat[k])
1834 mode_lib->vba.OutputFormat[k]));
1836 mode_lib->vba.DSCDelay[k] = mode_lib->vba.DSCDelay[k]
1837 * mode_lib->vba.PixelClock[k]
1838 / mode_lib->vba.PixelClockBackEnd[k];
1840 mode_lib->vba.DSCDelay[k] = 0;
1844 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k)
1845 for (j = 0; j < mode_lib->vba.NumberOfActivePlanes; ++j) // NumberOfPlanes
1846 if (j != k && mode_lib->vba.BlendingAndTiming[k] == j
1847 && mode_lib->vba.DSCEnabled[j])
1848 mode_lib->vba.DSCDelay[k] = mode_lib->vba.DSCDelay[j];
1851 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
1852 unsigned int PDEAndMetaPTEBytesFrameY;
1853 unsigned int PixelPTEBytesPerRowY;
1854 unsigned int MetaRowByteY;
1855 unsigned int MetaRowByteC;
1856 unsigned int PDEAndMetaPTEBytesFrameC;
1857 unsigned int PixelPTEBytesPerRowC;
1859 Calculate256BBlockSizes(
1860 mode_lib->vba.SourcePixelFormat[k],
1861 mode_lib->vba.SurfaceTiling[k],
1862 dml_ceil(mode_lib->vba.BytePerPixelDETY[k], 1),
1863 dml_ceil(mode_lib->vba.BytePerPixelDETC[k], 2),
1864 &mode_lib->vba.BlockHeight256BytesY[k],
1865 &mode_lib->vba.BlockHeight256BytesC[k],
1866 &mode_lib->vba.BlockWidth256BytesY[k],
1867 &mode_lib->vba.BlockWidth256BytesC[k]);
1868 PDEAndMetaPTEBytesFrameY = CalculateVMAndRowBytes(
1870 mode_lib->vba.DCCEnable[k],
1871 mode_lib->vba.BlockHeight256BytesY[k],
1872 mode_lib->vba.BlockWidth256BytesY[k],
1873 mode_lib->vba.SourcePixelFormat[k],
1874 mode_lib->vba.SurfaceTiling[k],
1875 dml_ceil(mode_lib->vba.BytePerPixelDETY[k], 1),
1876 mode_lib->vba.SourceScan[k],
1877 mode_lib->vba.ViewportWidth[k],
1878 mode_lib->vba.ViewportHeight[k],
1879 mode_lib->vba.SwathWidthY[k],
1880 mode_lib->vba.GPUVMEnable,
1881 mode_lib->vba.VMMPageSize,
1882 mode_lib->vba.PTEBufferSizeInRequestsLuma,
1883 mode_lib->vba.PDEProcessingBufIn64KBReqs,
1884 mode_lib->vba.PitchY[k],
1885 mode_lib->vba.DCCMetaPitchY[k],
1886 &mode_lib->vba.MacroTileWidthY[k],
1888 &PixelPTEBytesPerRowY,
1889 &mode_lib->vba.PTEBufferSizeNotExceeded[mode_lib->vba.VoltageLevel][0],
1890 &mode_lib->vba.dpte_row_height[k],
1891 &mode_lib->vba.meta_row_height[k]);
1892 mode_lib->vba.PrefetchSourceLinesY[k] = CalculatePrefetchSourceLines(
1894 mode_lib->vba.VRatio[k],
1895 mode_lib->vba.vtaps[k],
1896 mode_lib->vba.Interlace[k],
1897 mode_lib->vba.ProgressiveToInterlaceUnitInOPP,
1898 mode_lib->vba.SwathHeightY[k],
1899 mode_lib->vba.ViewportYStartY[k],
1900 &mode_lib->vba.VInitPreFillY[k],
1901 &mode_lib->vba.MaxNumSwathY[k]);
1903 if ((mode_lib->vba.SourcePixelFormat[k] != dm_444_64
1904 && mode_lib->vba.SourcePixelFormat[k] != dm_444_32
1905 && mode_lib->vba.SourcePixelFormat[k] != dm_444_16
1906 && mode_lib->vba.SourcePixelFormat[k] != dm_444_8)) {
1907 PDEAndMetaPTEBytesFrameC =
1908 CalculateVMAndRowBytes(
1910 mode_lib->vba.DCCEnable[k],
1911 mode_lib->vba.BlockHeight256BytesC[k],
1912 mode_lib->vba.BlockWidth256BytesC[k],
1913 mode_lib->vba.SourcePixelFormat[k],
1914 mode_lib->vba.SurfaceTiling[k],
1916 mode_lib->vba.BytePerPixelDETC[k],
1918 mode_lib->vba.SourceScan[k],
1919 mode_lib->vba.ViewportWidth[k] / 2,
1920 mode_lib->vba.ViewportHeight[k] / 2,
1921 mode_lib->vba.SwathWidthY[k] / 2,
1922 mode_lib->vba.GPUVMEnable,
1923 mode_lib->vba.VMMPageSize,
1924 mode_lib->vba.PTEBufferSizeInRequestsLuma,
1925 mode_lib->vba.PDEProcessingBufIn64KBReqs,
1926 mode_lib->vba.PitchC[k],
1928 &mode_lib->vba.MacroTileWidthC[k],
1930 &PixelPTEBytesPerRowC,
1931 &mode_lib->vba.PTEBufferSizeNotExceeded[mode_lib->vba.VoltageLevel][0],
1932 &mode_lib->vba.dpte_row_height_chroma[k],
1933 &mode_lib->vba.meta_row_height_chroma[k]);
1934 mode_lib->vba.PrefetchSourceLinesC[k] = CalculatePrefetchSourceLines(
1936 mode_lib->vba.VRatio[k] / 2,
1937 mode_lib->vba.VTAPsChroma[k],
1938 mode_lib->vba.Interlace[k],
1939 mode_lib->vba.ProgressiveToInterlaceUnitInOPP,
1940 mode_lib->vba.SwathHeightC[k],
1941 mode_lib->vba.ViewportYStartC[k],
1942 &mode_lib->vba.VInitPreFillC[k],
1943 &mode_lib->vba.MaxNumSwathC[k]);
1945 PixelPTEBytesPerRowC = 0;
1946 PDEAndMetaPTEBytesFrameC = 0;
1948 mode_lib->vba.MaxNumSwathC[k] = 0;
1949 mode_lib->vba.PrefetchSourceLinesC[k] = 0;
1952 mode_lib->vba.PixelPTEBytesPerRow[k] = PixelPTEBytesPerRowY + PixelPTEBytesPerRowC;
1953 mode_lib->vba.PDEAndMetaPTEBytesFrame[k] = PDEAndMetaPTEBytesFrameY
1954 + PDEAndMetaPTEBytesFrameC;
1955 mode_lib->vba.MetaRowByte[k] = MetaRowByteY + MetaRowByteC;
1957 CalculateActiveRowBandwidth(
1958 mode_lib->vba.GPUVMEnable,
1959 mode_lib->vba.SourcePixelFormat[k],
1960 mode_lib->vba.VRatio[k],
1961 mode_lib->vba.DCCEnable[k],
1962 mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k],
1965 mode_lib->vba.meta_row_height[k],
1966 mode_lib->vba.meta_row_height_chroma[k],
1967 PixelPTEBytesPerRowY,
1968 PixelPTEBytesPerRowC,
1969 mode_lib->vba.dpte_row_height[k],
1970 mode_lib->vba.dpte_row_height_chroma[k],
1971 &mode_lib->vba.meta_row_bw[k],
1972 &mode_lib->vba.dpte_row_bw[k],
1973 &mode_lib->vba.qual_row_bw[k]);
1976 mode_lib->vba.TCalc = 24.0 / mode_lib->vba.DCFCLKDeepSleep;
1978 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
1979 if (mode_lib->vba.BlendingAndTiming[k] == k) {
1980 if (mode_lib->vba.WritebackEnable[k] == true) {
1981 mode_lib->vba.WritebackDelay[mode_lib->vba.VoltageLevel][k] =
1982 mode_lib->vba.WritebackLatency
1983 + CalculateWriteBackDelay(
1984 mode_lib->vba.WritebackPixelFormat[k],
1985 mode_lib->vba.WritebackHRatio[k],
1986 mode_lib->vba.WritebackVRatio[k],
1987 mode_lib->vba.WritebackLumaHTaps[k],
1988 mode_lib->vba.WritebackLumaVTaps[k],
1989 mode_lib->vba.WritebackChromaHTaps[k],
1990 mode_lib->vba.WritebackChromaVTaps[k],
1991 mode_lib->vba.WritebackDestinationWidth[k])
1992 / mode_lib->vba.DISPCLK;
1994 mode_lib->vba.WritebackDelay[mode_lib->vba.VoltageLevel][k] = 0;
1995 for (j = 0; j < mode_lib->vba.NumberOfActivePlanes; ++j) {
1996 if (mode_lib->vba.BlendingAndTiming[j] == k
1997 && mode_lib->vba.WritebackEnable[j] == true) {
1998 mode_lib->vba.WritebackDelay[mode_lib->vba.VoltageLevel][k] =
2000 mode_lib->vba.WritebackDelay[mode_lib->vba.VoltageLevel][k],
2001 mode_lib->vba.WritebackLatency
2002 + CalculateWriteBackDelay(
2003 mode_lib->vba.WritebackPixelFormat[j],
2004 mode_lib->vba.WritebackHRatio[j],
2005 mode_lib->vba.WritebackVRatio[j],
2006 mode_lib->vba.WritebackLumaHTaps[j],
2007 mode_lib->vba.WritebackLumaVTaps[j],
2008 mode_lib->vba.WritebackChromaHTaps[j],
2009 mode_lib->vba.WritebackChromaVTaps[j],
2010 mode_lib->vba.WritebackDestinationWidth[j])
2011 / mode_lib->vba.DISPCLK);
2017 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k)
2018 for (j = 0; j < mode_lib->vba.NumberOfActivePlanes; ++j)
2019 if (mode_lib->vba.BlendingAndTiming[k] == j)
2020 mode_lib->vba.WritebackDelay[mode_lib->vba.VoltageLevel][k] =
2021 mode_lib->vba.WritebackDelay[mode_lib->vba.VoltageLevel][j];
2023 mode_lib->vba.VStartupLines = 13;
2024 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
2025 mode_lib->vba.MaxVStartupLines[k] =
2026 mode_lib->vba.VTotal[k] - mode_lib->vba.VActive[k]
2030 mode_lib->vba.WritebackDelay[mode_lib->vba.VoltageLevel][k]
2031 / (mode_lib->vba.HTotal[k]
2032 / mode_lib->vba.PixelClock[k]),
2036 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k)
2037 mode_lib->vba.MaximumMaxVStartupLines = dml_max(
2038 mode_lib->vba.MaximumMaxVStartupLines,
2039 mode_lib->vba.MaxVStartupLines[k]);
2041 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
2042 mode_lib->vba.cursor_bw[k] = 0.0;
2043 for (j = 0; j < mode_lib->vba.NumberOfCursors[k]; ++j)
2044 mode_lib->vba.cursor_bw[k] += mode_lib->vba.CursorWidth[k][j]
2045 * mode_lib->vba.CursorBPP[k][j] / 8.0
2046 / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k])
2047 * mode_lib->vba.VRatio[k];
2051 double MaxTotalRDBandwidth = 0;
2052 bool DestinationLineTimesForPrefetchLessThan2 = false;
2053 bool VRatioPrefetchMoreThan4 = false;
2054 bool prefetch_vm_bw_valid = true;
2055 bool prefetch_row_bw_valid = true;
2056 double TWait = CalculateTWait(
2057 mode_lib->vba.PrefetchMode[mode_lib->vba.VoltageLevel][mode_lib->vba.maxMpcComb],
2058 mode_lib->vba.DRAMClockChangeLatency,
2059 mode_lib->vba.UrgentLatencyPixelDataOnly,
2060 mode_lib->vba.SREnterPlusExitTime);
2062 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
2063 if (mode_lib->vba.XFCEnabled[k] == true) {
2064 mode_lib->vba.XFCRemoteSurfaceFlipDelay =
2065 CalculateRemoteSurfaceFlipDelay(
2067 mode_lib->vba.VRatio[k],
2068 mode_lib->vba.SwathWidthY[k],
2070 mode_lib->vba.BytePerPixelDETY[k],
2072 mode_lib->vba.HTotal[k]
2073 / mode_lib->vba.PixelClock[k],
2074 mode_lib->vba.XFCTSlvVupdateOffset,
2075 mode_lib->vba.XFCTSlvVupdateWidth,
2076 mode_lib->vba.XFCTSlvVreadyOffset,
2077 mode_lib->vba.XFCXBUFLatencyTolerance,
2078 mode_lib->vba.XFCFillBWOverhead,
2079 mode_lib->vba.XFCSlvChunkSize,
2080 mode_lib->vba.XFCBusTransportTime,
2081 mode_lib->vba.TCalc,
2083 &mode_lib->vba.SrcActiveDrainRate,
2084 &mode_lib->vba.TInitXFill,
2085 &mode_lib->vba.TslvChk);
2087 mode_lib->vba.XFCRemoteSurfaceFlipDelay = 0;
2089 mode_lib->vba.ErrorResult[k] =
2090 CalculatePrefetchSchedule(
2092 mode_lib->vba.DPPCLK[k],
2093 mode_lib->vba.DISPCLK,
2094 mode_lib->vba.PixelClock[k],
2095 mode_lib->vba.DCFCLKDeepSleep,
2096 mode_lib->vba.DSCDelay[k],
2097 mode_lib->vba.DPPPerPlane[k],
2098 mode_lib->vba.ScalerEnabled[k],
2099 mode_lib->vba.NumberOfCursors[k],
2100 mode_lib->vba.DPPCLKDelaySubtotal,
2101 mode_lib->vba.DPPCLKDelaySCL,
2102 mode_lib->vba.DPPCLKDelaySCLLBOnly,
2103 mode_lib->vba.DPPCLKDelayCNVCFormater,
2104 mode_lib->vba.DPPCLKDelayCNVCCursor,
2105 mode_lib->vba.DISPCLKDelaySubtotal,
2106 (unsigned int) (mode_lib->vba.SwathWidthY[k]
2107 / mode_lib->vba.HRatio[k]),
2108 mode_lib->vba.OutputFormat[k],
2109 mode_lib->vba.VTotal[k]
2110 - mode_lib->vba.VActive[k],
2111 mode_lib->vba.HTotal[k],
2112 mode_lib->vba.MaxInterDCNTileRepeaters,
2114 mode_lib->vba.VStartupLines,
2115 mode_lib->vba.MaxVStartupLines[k]),
2116 mode_lib->vba.GPUVMMaxPageTableLevels,
2117 mode_lib->vba.GPUVMEnable,
2118 mode_lib->vba.DynamicMetadataEnable[k],
2119 mode_lib->vba.DynamicMetadataLinesBeforeActiveRequired[k],
2120 mode_lib->vba.DynamicMetadataTransmittedBytes[k],
2121 mode_lib->vba.DCCEnable[k],
2122 mode_lib->vba.UrgentLatencyPixelDataOnly,
2123 mode_lib->vba.UrgentExtraLatency,
2124 mode_lib->vba.TCalc,
2125 mode_lib->vba.PDEAndMetaPTEBytesFrame[k],
2126 mode_lib->vba.MetaRowByte[k],
2127 mode_lib->vba.PixelPTEBytesPerRow[k],
2128 mode_lib->vba.PrefetchSourceLinesY[k],
2129 mode_lib->vba.SwathWidthY[k],
2130 mode_lib->vba.BytePerPixelDETY[k],
2131 mode_lib->vba.VInitPreFillY[k],
2132 mode_lib->vba.MaxNumSwathY[k],
2133 mode_lib->vba.PrefetchSourceLinesC[k],
2134 mode_lib->vba.BytePerPixelDETC[k],
2135 mode_lib->vba.VInitPreFillC[k],
2136 mode_lib->vba.MaxNumSwathC[k],
2137 mode_lib->vba.SwathHeightY[k],
2138 mode_lib->vba.SwathHeightC[k],
2140 mode_lib->vba.XFCEnabled[k],
2141 mode_lib->vba.XFCRemoteSurfaceFlipDelay,
2142 mode_lib->vba.Interlace[k],
2143 mode_lib->vba.ProgressiveToInterlaceUnitInOPP,
2144 &mode_lib->vba.DSTXAfterScaler[k],
2145 &mode_lib->vba.DSTYAfterScaler[k],
2146 &mode_lib->vba.DestinationLinesForPrefetch[k],
2147 &mode_lib->vba.PrefetchBandwidth[k],
2148 &mode_lib->vba.DestinationLinesToRequestVMInVBlank[k],
2149 &mode_lib->vba.DestinationLinesToRequestRowInVBlank[k],
2150 &mode_lib->vba.VRatioPrefetchY[k],
2151 &mode_lib->vba.VRatioPrefetchC[k],
2152 &mode_lib->vba.RequiredPrefetchPixDataBWLuma[k],
2153 &mode_lib->vba.VStartupRequiredWhenNotEnoughTimeForDynamicMetadata,
2154 &mode_lib->vba.Tno_bw[k],
2155 &mode_lib->vba.VUpdateOffsetPix[k],
2156 &mode_lib->vba.VUpdateWidthPix[k],
2157 &mode_lib->vba.VReadyOffsetPix[k]);
2158 if (mode_lib->vba.BlendingAndTiming[k] == k) {
2159 mode_lib->vba.VStartup[k] = dml_min(
2160 mode_lib->vba.VStartupLines,
2161 mode_lib->vba.MaxVStartupLines[k]);
2162 if (mode_lib->vba.VStartupRequiredWhenNotEnoughTimeForDynamicMetadata
2164 mode_lib->vba.VStartup[k] =
2165 mode_lib->vba.VStartupRequiredWhenNotEnoughTimeForDynamicMetadata;
2168 mode_lib->vba.VStartup[k] =
2170 mode_lib->vba.VStartupLines,
2171 mode_lib->vba.MaxVStartupLines[mode_lib->vba.BlendingAndTiming[k]]);
2175 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
2177 if (mode_lib->vba.PDEAndMetaPTEBytesFrame[k] == 0)
2178 mode_lib->vba.prefetch_vm_bw[k] = 0;
2179 else if (mode_lib->vba.DestinationLinesToRequestVMInVBlank[k] > 0) {
2180 mode_lib->vba.prefetch_vm_bw[k] =
2181 (double) mode_lib->vba.PDEAndMetaPTEBytesFrame[k]
2182 / (mode_lib->vba.DestinationLinesToRequestVMInVBlank[k]
2183 * mode_lib->vba.HTotal[k]
2184 / mode_lib->vba.PixelClock[k]);
2186 mode_lib->vba.prefetch_vm_bw[k] = 0;
2187 prefetch_vm_bw_valid = false;
2189 if (mode_lib->vba.MetaRowByte[k] + mode_lib->vba.PixelPTEBytesPerRow[k]
2191 mode_lib->vba.prefetch_row_bw[k] = 0;
2192 else if (mode_lib->vba.DestinationLinesToRequestRowInVBlank[k] > 0) {
2193 mode_lib->vba.prefetch_row_bw[k] =
2194 (double) (mode_lib->vba.MetaRowByte[k]
2195 + mode_lib->vba.PixelPTEBytesPerRow[k])
2196 / (mode_lib->vba.DestinationLinesToRequestRowInVBlank[k]
2197 * mode_lib->vba.HTotal[k]
2198 / mode_lib->vba.PixelClock[k]);
2200 mode_lib->vba.prefetch_row_bw[k] = 0;
2201 prefetch_row_bw_valid = false;
2204 MaxTotalRDBandwidth =
2205 MaxTotalRDBandwidth + mode_lib->vba.cursor_bw[k]
2207 mode_lib->vba.prefetch_vm_bw[k],
2209 mode_lib->vba.prefetch_row_bw[k],
2211 mode_lib->vba.ReadBandwidthPlaneLuma[k]
2212 + mode_lib->vba.ReadBandwidthPlaneChroma[k],
2213 mode_lib->vba.RequiredPrefetchPixDataBWLuma[k])
2214 + mode_lib->vba.meta_row_bw[k]
2215 + mode_lib->vba.dpte_row_bw[k]));
2217 if (mode_lib->vba.DestinationLinesForPrefetch[k] < 2)
2218 DestinationLineTimesForPrefetchLessThan2 = true;
2219 if (mode_lib->vba.VRatioPrefetchY[k] > 4
2220 || mode_lib->vba.VRatioPrefetchC[k] > 4)
2221 VRatioPrefetchMoreThan4 = true;
2224 if (MaxTotalRDBandwidth <= mode_lib->vba.ReturnBW && prefetch_vm_bw_valid
2225 && prefetch_row_bw_valid && !VRatioPrefetchMoreThan4
2226 && !DestinationLineTimesForPrefetchLessThan2)
2227 mode_lib->vba.PrefetchModeSupported = true;
2229 mode_lib->vba.PrefetchModeSupported = false;
2231 "DML: CalculatePrefetchSchedule ***failed***. Bandwidth violation. Results are NOT valid\n");
2234 if (mode_lib->vba.PrefetchModeSupported == true) {
2235 double final_flip_bw[DC__NUM_DPP__MAX];
2236 unsigned int ImmediateFlipBytes[DC__NUM_DPP__MAX];
2237 double total_dcn_read_bw_with_flip = 0;
2239 mode_lib->vba.BandwidthAvailableForImmediateFlip = mode_lib->vba.ReturnBW;
2240 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
2241 mode_lib->vba.BandwidthAvailableForImmediateFlip =
2242 mode_lib->vba.BandwidthAvailableForImmediateFlip
2243 - mode_lib->vba.cursor_bw[k]
2245 mode_lib->vba.ReadBandwidthPlaneLuma[k]
2246 + mode_lib->vba.ReadBandwidthPlaneChroma[k]
2247 + mode_lib->vba.qual_row_bw[k],
2248 mode_lib->vba.PrefetchBandwidth[k]);
2251 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
2252 ImmediateFlipBytes[k] = 0;
2253 if ((mode_lib->vba.SourcePixelFormat[k] != dm_420_8
2254 && mode_lib->vba.SourcePixelFormat[k] != dm_420_10)) {
2255 ImmediateFlipBytes[k] =
2256 mode_lib->vba.PDEAndMetaPTEBytesFrame[k]
2257 + mode_lib->vba.MetaRowByte[k]
2258 + mode_lib->vba.PixelPTEBytesPerRow[k];
2261 mode_lib->vba.TotImmediateFlipBytes = 0;
2262 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
2263 if ((mode_lib->vba.SourcePixelFormat[k] != dm_420_8
2264 && mode_lib->vba.SourcePixelFormat[k] != dm_420_10)) {
2265 mode_lib->vba.TotImmediateFlipBytes =
2266 mode_lib->vba.TotImmediateFlipBytes
2267 + ImmediateFlipBytes[k];
2270 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
2271 CalculateFlipSchedule(
2273 mode_lib->vba.UrgentExtraLatency,
2274 mode_lib->vba.UrgentLatencyPixelDataOnly,
2275 mode_lib->vba.GPUVMMaxPageTableLevels,
2276 mode_lib->vba.GPUVMEnable,
2277 mode_lib->vba.BandwidthAvailableForImmediateFlip,
2278 mode_lib->vba.TotImmediateFlipBytes,
2279 mode_lib->vba.SourcePixelFormat[k],
2280 ImmediateFlipBytes[k],
2281 mode_lib->vba.HTotal[k]
2282 / mode_lib->vba.PixelClock[k],
2283 mode_lib->vba.VRatio[k],
2284 mode_lib->vba.Tno_bw[k],
2285 mode_lib->vba.PDEAndMetaPTEBytesFrame[k],
2286 mode_lib->vba.MetaRowByte[k],
2287 mode_lib->vba.PixelPTEBytesPerRow[k],
2288 mode_lib->vba.DCCEnable[k],
2289 mode_lib->vba.dpte_row_height[k],
2290 mode_lib->vba.meta_row_height[k],
2291 mode_lib->vba.qual_row_bw[k],
2292 &mode_lib->vba.DestinationLinesToRequestVMInImmediateFlip[k],
2293 &mode_lib->vba.DestinationLinesToRequestRowInImmediateFlip[k],
2295 &mode_lib->vba.ImmediateFlipSupportedForPipe[k]);
2297 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
2298 total_dcn_read_bw_with_flip =
2299 total_dcn_read_bw_with_flip
2300 + mode_lib->vba.cursor_bw[k]
2302 mode_lib->vba.prefetch_vm_bw[k],
2304 mode_lib->vba.prefetch_row_bw[k],
2307 mode_lib->vba.ReadBandwidthPlaneLuma[k]
2308 + mode_lib->vba.ReadBandwidthPlaneChroma[k],
2309 mode_lib->vba.RequiredPrefetchPixDataBWLuma[k])));
2311 mode_lib->vba.ImmediateFlipSupported = true;
2312 if (total_dcn_read_bw_with_flip > mode_lib->vba.ReturnBW) {
2313 mode_lib->vba.ImmediateFlipSupported = false;
2315 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
2316 if (mode_lib->vba.ImmediateFlipSupportedForPipe[k] == false) {
2317 mode_lib->vba.ImmediateFlipSupported = false;
2321 mode_lib->vba.ImmediateFlipSupported = false;
2324 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
2325 if (mode_lib->vba.ErrorResult[k]) {
2326 mode_lib->vba.PrefetchModeSupported = false;
2328 "DML: CalculatePrefetchSchedule ***failed***. Prefetch schedule violation. Results are NOT valid\n");
2332 mode_lib->vba.VStartupLines = mode_lib->vba.VStartupLines + 1;
2333 } while (!((mode_lib->vba.PrefetchModeSupported
2334 && (!mode_lib->vba.ImmediateFlipSupport
2335 || mode_lib->vba.ImmediateFlipSupported))
2336 || mode_lib->vba.MaximumMaxVStartupLines < mode_lib->vba.VStartupLines));
2338 //Display Pipeline Delivery Time in Prefetch
2339 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
2340 if (mode_lib->vba.VRatioPrefetchY[k] <= 1) {
2341 mode_lib->vba.DisplayPipeLineDeliveryTimeLumaPrefetch[k] =
2342 mode_lib->vba.SwathWidthY[k] * mode_lib->vba.DPPPerPlane[k]
2343 / mode_lib->vba.HRatio[k]
2344 / mode_lib->vba.PixelClock[k];
2346 mode_lib->vba.DisplayPipeLineDeliveryTimeLumaPrefetch[k] =
2347 mode_lib->vba.SwathWidthY[k]
2348 / mode_lib->vba.PSCL_THROUGHPUT_LUMA[k]
2349 / mode_lib->vba.DPPCLK[k];
2351 if (mode_lib->vba.BytePerPixelDETC[k] == 0) {
2352 mode_lib->vba.DisplayPipeLineDeliveryTimeChromaPrefetch[k] = 0;
2354 if (mode_lib->vba.VRatioPrefetchC[k] <= 1) {
2355 mode_lib->vba.DisplayPipeLineDeliveryTimeChromaPrefetch[k] =
2356 mode_lib->vba.SwathWidthY[k]
2357 * mode_lib->vba.DPPPerPlane[k]
2358 / mode_lib->vba.HRatio[k]
2359 / mode_lib->vba.PixelClock[k];
2361 mode_lib->vba.DisplayPipeLineDeliveryTimeChromaPrefetch[k] =
2362 mode_lib->vba.SwathWidthY[k]
2363 / mode_lib->vba.PSCL_THROUGHPUT_LUMA[k]
2364 / mode_lib->vba.DPPCLK[k];
2370 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
2371 if (mode_lib->vba.PrefetchMode[mode_lib->vba.VoltageLevel][mode_lib->vba.maxMpcComb] == 0) {
2372 mode_lib->vba.AllowDRAMClockChangeDuringVBlank[k] = true;
2373 mode_lib->vba.AllowDRAMSelfRefreshDuringVBlank[k] = true;
2374 mode_lib->vba.MinTTUVBlank[k] = dml_max(
2375 mode_lib->vba.DRAMClockChangeWatermark,
2377 mode_lib->vba.StutterEnterPlusExitWatermark,
2378 mode_lib->vba.UrgentWatermark));
2379 } else if (mode_lib->vba.PrefetchMode[mode_lib->vba.VoltageLevel][mode_lib->vba.maxMpcComb] == 1) {
2380 mode_lib->vba.AllowDRAMClockChangeDuringVBlank[k] = false;
2381 mode_lib->vba.AllowDRAMSelfRefreshDuringVBlank[k] = true;
2382 mode_lib->vba.MinTTUVBlank[k] = dml_max(
2383 mode_lib->vba.StutterEnterPlusExitWatermark,
2384 mode_lib->vba.UrgentWatermark);
2386 mode_lib->vba.AllowDRAMClockChangeDuringVBlank[k] = false;
2387 mode_lib->vba.AllowDRAMSelfRefreshDuringVBlank[k] = false;
2388 mode_lib->vba.MinTTUVBlank[k] = mode_lib->vba.UrgentWatermark;
2390 if (!mode_lib->vba.DynamicMetadataEnable[k])
2391 mode_lib->vba.MinTTUVBlank[k] = mode_lib->vba.TCalc
2392 + mode_lib->vba.MinTTUVBlank[k];
2395 // DCC Configuration
2396 mode_lib->vba.ActiveDPPs = 0;
2397 // NB P-State/DRAM Clock Change Support
2398 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
2399 mode_lib->vba.ActiveDPPs = mode_lib->vba.ActiveDPPs + mode_lib->vba.DPPPerPlane[k];
2402 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
2403 double EffectiveLBLatencyHidingY;
2404 double EffectiveLBLatencyHidingC;
2405 double DPPOutputBufferLinesY;
2406 double DPPOutputBufferLinesC;
2407 double DPPOPPBufferingY;
2408 double MaxDETBufferingTimeY;
2409 double ActiveDRAMClockChangeLatencyMarginY;
2411 mode_lib->vba.LBLatencyHidingSourceLinesY =
2413 mode_lib->vba.MaxLineBufferLines,
2414 (unsigned int) dml_floor(
2415 (double) mode_lib->vba.LineBufferSize
2416 / mode_lib->vba.LBBitPerPixel[k]
2417 / (mode_lib->vba.SwathWidthY[k]
2419 mode_lib->vba.HRatio[k],
2421 1)) - (mode_lib->vba.vtaps[k] - 1);
2423 mode_lib->vba.LBLatencyHidingSourceLinesC =
2425 mode_lib->vba.MaxLineBufferLines,
2426 (unsigned int) dml_floor(
2427 (double) mode_lib->vba.LineBufferSize
2428 / mode_lib->vba.LBBitPerPixel[k]
2429 / (mode_lib->vba.SwathWidthY[k]
2432 mode_lib->vba.HRatio[k]
2436 - (mode_lib->vba.VTAPsChroma[k] - 1);
2438 EffectiveLBLatencyHidingY = mode_lib->vba.LBLatencyHidingSourceLinesY
2439 / mode_lib->vba.VRatio[k]
2440 * (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]);
2442 EffectiveLBLatencyHidingC = mode_lib->vba.LBLatencyHidingSourceLinesC
2443 / (mode_lib->vba.VRatio[k] / 2)
2444 * (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]);
2446 if (mode_lib->vba.SwathWidthY[k] > 2 * mode_lib->vba.DPPOutputBufferPixels) {
2447 DPPOutputBufferLinesY = mode_lib->vba.DPPOutputBufferPixels
2448 / mode_lib->vba.SwathWidthY[k];
2449 } else if (mode_lib->vba.SwathWidthY[k] > mode_lib->vba.DPPOutputBufferPixels) {
2450 DPPOutputBufferLinesY = 0.5;
2452 DPPOutputBufferLinesY = 1;
2455 if (mode_lib->vba.SwathWidthY[k] / 2 > 2 * mode_lib->vba.DPPOutputBufferPixels) {
2456 DPPOutputBufferLinesC = mode_lib->vba.DPPOutputBufferPixels
2457 / (mode_lib->vba.SwathWidthY[k] / 2);
2458 } else if (mode_lib->vba.SwathWidthY[k] / 2 > mode_lib->vba.DPPOutputBufferPixels) {
2459 DPPOutputBufferLinesC = 0.5;
2461 DPPOutputBufferLinesC = 1;
2464 DPPOPPBufferingY = (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k])
2465 * (DPPOutputBufferLinesY + mode_lib->vba.OPPOutputBufferLines);
2466 MaxDETBufferingTimeY = mode_lib->vba.FullDETBufferingTimeY[k]
2467 + (mode_lib->vba.LinesInDETY[k]
2468 - mode_lib->vba.LinesInDETYRoundedDownToSwath[k])
2469 / mode_lib->vba.SwathHeightY[k]
2470 * (mode_lib->vba.HTotal[k]
2471 / mode_lib->vba.PixelClock[k]);
2473 ActiveDRAMClockChangeLatencyMarginY = DPPOPPBufferingY + EffectiveLBLatencyHidingY
2474 + MaxDETBufferingTimeY - mode_lib->vba.DRAMClockChangeWatermark;
2476 if (mode_lib->vba.ActiveDPPs > 1) {
2477 ActiveDRAMClockChangeLatencyMarginY =
2478 ActiveDRAMClockChangeLatencyMarginY
2479 - (1 - 1 / (mode_lib->vba.ActiveDPPs - 1))
2480 * mode_lib->vba.SwathHeightY[k]
2481 * (mode_lib->vba.HTotal[k]
2482 / mode_lib->vba.PixelClock[k]);
2485 if (mode_lib->vba.BytePerPixelDETC[k] > 0) {
2486 double DPPOPPBufferingC = (mode_lib->vba.HTotal[k]
2487 / mode_lib->vba.PixelClock[k])
2488 * (DPPOutputBufferLinesC
2489 + mode_lib->vba.OPPOutputBufferLines);
2490 double MaxDETBufferingTimeC =
2491 mode_lib->vba.FullDETBufferingTimeC[k]
2492 + (mode_lib->vba.LinesInDETC[k]
2493 - mode_lib->vba.LinesInDETCRoundedDownToSwath[k])
2494 / mode_lib->vba.SwathHeightC[k]
2495 * (mode_lib->vba.HTotal[k]
2496 / mode_lib->vba.PixelClock[k]);
2497 double ActiveDRAMClockChangeLatencyMarginC = DPPOPPBufferingC
2498 + EffectiveLBLatencyHidingC + MaxDETBufferingTimeC
2499 - mode_lib->vba.DRAMClockChangeWatermark;
2501 if (mode_lib->vba.ActiveDPPs > 1) {
2502 ActiveDRAMClockChangeLatencyMarginC =
2503 ActiveDRAMClockChangeLatencyMarginC
2506 / (mode_lib->vba.ActiveDPPs
2508 * mode_lib->vba.SwathHeightC[k]
2509 * (mode_lib->vba.HTotal[k]
2510 / mode_lib->vba.PixelClock[k]);
2512 mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k] = dml_min(
2513 ActiveDRAMClockChangeLatencyMarginY,
2514 ActiveDRAMClockChangeLatencyMarginC);
2516 mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k] =
2517 ActiveDRAMClockChangeLatencyMarginY;
2520 if (mode_lib->vba.WritebackEnable[k]) {
2521 double WritebackDRAMClockChangeLatencyMargin;
2523 if (mode_lib->vba.WritebackPixelFormat[k] == dm_444_32) {
2524 WritebackDRAMClockChangeLatencyMargin =
2525 (double) (mode_lib->vba.WritebackInterfaceLumaBufferSize
2526 + mode_lib->vba.WritebackInterfaceChromaBufferSize)
2527 / (mode_lib->vba.WritebackDestinationWidth[k]
2528 * mode_lib->vba.WritebackDestinationHeight[k]
2529 / (mode_lib->vba.WritebackSourceHeight[k]
2530 * mode_lib->vba.HTotal[k]
2531 / mode_lib->vba.PixelClock[k])
2533 - mode_lib->vba.WritebackDRAMClockChangeWatermark;
2534 } else if (mode_lib->vba.WritebackPixelFormat[k] == dm_420_10) {
2535 WritebackDRAMClockChangeLatencyMargin =
2537 (double) mode_lib->vba.WritebackInterfaceLumaBufferSize
2540 * mode_lib->vba.WritebackInterfaceChromaBufferSize
2542 / (mode_lib->vba.WritebackDestinationWidth[k]
2543 * mode_lib->vba.WritebackDestinationHeight[k]
2544 / (mode_lib->vba.WritebackSourceHeight[k]
2545 * mode_lib->vba.HTotal[k]
2546 / mode_lib->vba.PixelClock[k]))
2547 - mode_lib->vba.WritebackDRAMClockChangeWatermark;
2549 WritebackDRAMClockChangeLatencyMargin =
2551 (double) mode_lib->vba.WritebackInterfaceLumaBufferSize,
2553 * mode_lib->vba.WritebackInterfaceChromaBufferSize)
2554 / (mode_lib->vba.WritebackDestinationWidth[k]
2555 * mode_lib->vba.WritebackDestinationHeight[k]
2556 / (mode_lib->vba.WritebackSourceHeight[k]
2557 * mode_lib->vba.HTotal[k]
2558 / mode_lib->vba.PixelClock[k]))
2559 - mode_lib->vba.WritebackDRAMClockChangeWatermark;
2561 mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k] = dml_min(
2562 mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k],
2563 WritebackDRAMClockChangeLatencyMargin);
2567 mode_lib->vba.MinActiveDRAMClockChangeMargin = 999999;
2568 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
2569 if (mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k]
2570 < mode_lib->vba.MinActiveDRAMClockChangeMargin) {
2571 mode_lib->vba.MinActiveDRAMClockChangeMargin =
2572 mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k];
2576 mode_lib->vba.MinActiveDRAMClockChangeLatencySupported =
2577 mode_lib->vba.MinActiveDRAMClockChangeMargin
2578 + mode_lib->vba.DRAMClockChangeLatency;
2580 if (mode_lib->vba.MinActiveDRAMClockChangeMargin > 0) {
2581 mode_lib->vba.DRAMClockChangeSupport[0][0] = dm_dram_clock_change_vactive;
2583 if (mode_lib->vba.SynchronizedVBlank || mode_lib->vba.NumberOfActivePlanes == 1) {
2584 mode_lib->vba.DRAMClockChangeSupport[0][0] = dm_dram_clock_change_vblank;
2585 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
2586 if (!mode_lib->vba.AllowDRAMClockChangeDuringVBlank[k]) {
2587 mode_lib->vba.DRAMClockChangeSupport[0][0] =
2588 dm_dram_clock_change_unsupported;
2592 mode_lib->vba.DRAMClockChangeSupport[0][0] = dm_dram_clock_change_unsupported;
2595 for (k = 0; k <= mode_lib->vba.soc.num_states; k++)
2596 for (j = 0; j < 2; j++)
2597 mode_lib->vba.DRAMClockChangeSupport[k][j] = mode_lib->vba.DRAMClockChangeSupport[0][0];
2600 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
2601 if (mode_lib->vba.XFCEnabled[k] == true) {
2604 mode_lib->vba.XFCSlaveVUpdateOffset[k] = mode_lib->vba.XFCTSlvVupdateOffset;
2605 mode_lib->vba.XFCSlaveVupdateWidth[k] = mode_lib->vba.XFCTSlvVupdateWidth;
2606 mode_lib->vba.XFCSlaveVReadyOffset[k] = mode_lib->vba.XFCTSlvVreadyOffset;
2607 TWait = CalculateTWait(
2608 mode_lib->vba.PrefetchMode[mode_lib->vba.VoltageLevel][mode_lib->vba.maxMpcComb],
2609 mode_lib->vba.DRAMClockChangeLatency,
2610 mode_lib->vba.UrgentLatencyPixelDataOnly,
2611 mode_lib->vba.SREnterPlusExitTime);
2612 mode_lib->vba.XFCRemoteSurfaceFlipDelay = CalculateRemoteSurfaceFlipDelay(
2614 mode_lib->vba.VRatio[k],
2615 mode_lib->vba.SwathWidthY[k],
2616 dml_ceil(mode_lib->vba.BytePerPixelDETY[k], 1),
2617 mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k],
2618 mode_lib->vba.XFCTSlvVupdateOffset,
2619 mode_lib->vba.XFCTSlvVupdateWidth,
2620 mode_lib->vba.XFCTSlvVreadyOffset,
2621 mode_lib->vba.XFCXBUFLatencyTolerance,
2622 mode_lib->vba.XFCFillBWOverhead,
2623 mode_lib->vba.XFCSlvChunkSize,
2624 mode_lib->vba.XFCBusTransportTime,
2625 mode_lib->vba.TCalc,
2627 &mode_lib->vba.SrcActiveDrainRate,
2628 &mode_lib->vba.TInitXFill,
2629 &mode_lib->vba.TslvChk);
2630 mode_lib->vba.XFCRemoteSurfaceFlipLatency[k] =
2632 mode_lib->vba.XFCRemoteSurfaceFlipDelay
2633 / (mode_lib->vba.HTotal[k]
2634 / mode_lib->vba.PixelClock[k]),
2636 mode_lib->vba.XFCTransferDelay[k] =
2638 mode_lib->vba.XFCBusTransportTime
2639 / (mode_lib->vba.HTotal[k]
2640 / mode_lib->vba.PixelClock[k]),
2642 mode_lib->vba.XFCPrechargeDelay[k] =
2644 (mode_lib->vba.XFCBusTransportTime
2645 + mode_lib->vba.TInitXFill
2646 + mode_lib->vba.TslvChk)
2647 / (mode_lib->vba.HTotal[k]
2648 / mode_lib->vba.PixelClock[k]),
2650 mode_lib->vba.InitFillLevel = mode_lib->vba.XFCXBUFLatencyTolerance
2651 * mode_lib->vba.SrcActiveDrainRate;
2652 mode_lib->vba.FinalFillMargin =
2653 (mode_lib->vba.DestinationLinesToRequestVMInVBlank[k]
2654 + mode_lib->vba.DestinationLinesToRequestRowInVBlank[k])
2655 * mode_lib->vba.HTotal[k]
2656 / mode_lib->vba.PixelClock[k]
2657 * mode_lib->vba.SrcActiveDrainRate
2658 + mode_lib->vba.XFCFillConstant;
2659 mode_lib->vba.FinalFillLevel = mode_lib->vba.XFCRemoteSurfaceFlipDelay
2660 * mode_lib->vba.SrcActiveDrainRate
2661 + mode_lib->vba.FinalFillMargin;
2662 mode_lib->vba.RemainingFillLevel = dml_max(
2664 mode_lib->vba.FinalFillLevel - mode_lib->vba.InitFillLevel);
2665 mode_lib->vba.TFinalxFill = mode_lib->vba.RemainingFillLevel
2666 / (mode_lib->vba.SrcActiveDrainRate
2667 * mode_lib->vba.XFCFillBWOverhead / 100);
2668 mode_lib->vba.XFCPrefetchMargin[k] =
2669 mode_lib->vba.XFCRemoteSurfaceFlipDelay
2670 + mode_lib->vba.TFinalxFill
2671 + (mode_lib->vba.DestinationLinesToRequestVMInVBlank[k]
2672 + mode_lib->vba.DestinationLinesToRequestRowInVBlank[k])
2673 * mode_lib->vba.HTotal[k]
2674 / mode_lib->vba.PixelClock[k];
2676 mode_lib->vba.XFCSlaveVUpdateOffset[k] = 0;
2677 mode_lib->vba.XFCSlaveVupdateWidth[k] = 0;
2678 mode_lib->vba.XFCSlaveVReadyOffset[k] = 0;
2679 mode_lib->vba.XFCRemoteSurfaceFlipLatency[k] = 0;
2680 mode_lib->vba.XFCPrechargeDelay[k] = 0;
2681 mode_lib->vba.XFCTransferDelay[k] = 0;
2682 mode_lib->vba.XFCPrefetchMargin[k] = 0;
2686 unsigned int VStartupMargin = 0;
2687 bool FirstMainPlane = true;
2689 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
2690 if (mode_lib->vba.BlendingAndTiming[k] == k) {
2691 unsigned int Margin = (mode_lib->vba.MaxVStartupLines[k] - mode_lib->vba.VStartup[k])
2692 * mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k];
2694 if (FirstMainPlane) {
2695 VStartupMargin = Margin;
2696 FirstMainPlane = false;
2698 VStartupMargin = dml_min(VStartupMargin, Margin);
2701 if (mode_lib->vba.UseMaximumVStartup) {
2702 if (mode_lib->vba.VTotal_Max[k] == mode_lib->vba.VTotal[k]) {
2703 //only use max vstart if it is not drr or lateflip.
2704 mode_lib->vba.VStartup[k] = mode_lib->vba.MaxVStartupLines[mode_lib->vba.BlendingAndTiming[k]];
2711 static void dml20_DisplayPipeConfiguration(struct display_mode_lib *mode_lib)
2713 double BytePerPixDETY;
2714 double BytePerPixDETC;
2715 double Read256BytesBlockHeightY;
2716 double Read256BytesBlockHeightC;
2717 double Read256BytesBlockWidthY;
2718 double Read256BytesBlockWidthC;
2719 double MaximumSwathHeightY;
2720 double MaximumSwathHeightC;
2721 double MinimumSwathHeightY;
2722 double MinimumSwathHeightC;
2724 double SwathWidthGranularityY;
2725 double SwathWidthGranularityC;
2726 double RoundedUpMaxSwathSizeBytesY;
2727 double RoundedUpMaxSwathSizeBytesC;
2730 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
2731 bool MainPlaneDoesODMCombine = false;
2733 if (mode_lib->vba.SourcePixelFormat[k] == dm_444_64) {
2736 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_32) {
2739 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_16) {
2742 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_8) {
2745 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_8) {
2749 BytePerPixDETY = 4.0 / 3.0;
2750 BytePerPixDETC = 8.0 / 3.0;
2753 if ((mode_lib->vba.SourcePixelFormat[k] == dm_444_64
2754 || mode_lib->vba.SourcePixelFormat[k] == dm_444_32
2755 || mode_lib->vba.SourcePixelFormat[k] == dm_444_16
2756 || mode_lib->vba.SourcePixelFormat[k] == dm_444_8)) {
2757 if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear) {
2758 Read256BytesBlockHeightY = 1;
2759 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_64) {
2760 Read256BytesBlockHeightY = 4;
2761 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_32
2762 || mode_lib->vba.SourcePixelFormat[k] == dm_444_16) {
2763 Read256BytesBlockHeightY = 8;
2765 Read256BytesBlockHeightY = 16;
2767 Read256BytesBlockWidthY = 256 / dml_ceil(BytePerPixDETY, 1)
2768 / Read256BytesBlockHeightY;
2769 Read256BytesBlockHeightC = 0;
2770 Read256BytesBlockWidthC = 0;
2772 if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear) {
2773 Read256BytesBlockHeightY = 1;
2774 Read256BytesBlockHeightC = 1;
2775 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_8) {
2776 Read256BytesBlockHeightY = 16;
2777 Read256BytesBlockHeightC = 8;
2779 Read256BytesBlockHeightY = 8;
2780 Read256BytesBlockHeightC = 8;
2782 Read256BytesBlockWidthY = 256 / dml_ceil(BytePerPixDETY, 1)
2783 / Read256BytesBlockHeightY;
2784 Read256BytesBlockWidthC = 256 / dml_ceil(BytePerPixDETC, 2)
2785 / Read256BytesBlockHeightC;
2788 if (mode_lib->vba.SourceScan[k] == dm_horz) {
2789 MaximumSwathHeightY = Read256BytesBlockHeightY;
2790 MaximumSwathHeightC = Read256BytesBlockHeightC;
2792 MaximumSwathHeightY = Read256BytesBlockWidthY;
2793 MaximumSwathHeightC = Read256BytesBlockWidthC;
2796 if ((mode_lib->vba.SourcePixelFormat[k] == dm_444_64
2797 || mode_lib->vba.SourcePixelFormat[k] == dm_444_32
2798 || mode_lib->vba.SourcePixelFormat[k] == dm_444_16
2799 || mode_lib->vba.SourcePixelFormat[k] == dm_444_8)) {
2800 if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear
2801 || (mode_lib->vba.SourcePixelFormat[k] == dm_444_64
2802 && (mode_lib->vba.SurfaceTiling[k]
2804 || mode_lib->vba.SurfaceTiling[k]
2806 || mode_lib->vba.SurfaceTiling[k]
2808 || mode_lib->vba.SurfaceTiling[k]
2810 || mode_lib->vba.SurfaceTiling[k]
2812 || mode_lib->vba.SurfaceTiling[k]
2814 || mode_lib->vba.SurfaceTiling[k]
2816 && mode_lib->vba.SourceScan[k] == dm_horz)) {
2817 MinimumSwathHeightY = MaximumSwathHeightY;
2818 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_8
2819 && mode_lib->vba.SourceScan[k] != dm_horz) {
2820 MinimumSwathHeightY = MaximumSwathHeightY;
2822 MinimumSwathHeightY = MaximumSwathHeightY / 2.0;
2824 MinimumSwathHeightC = MaximumSwathHeightC;
2826 if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear) {
2827 MinimumSwathHeightY = MaximumSwathHeightY;
2828 MinimumSwathHeightC = MaximumSwathHeightC;
2829 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_8
2830 && mode_lib->vba.SourceScan[k] == dm_horz) {
2831 MinimumSwathHeightY = MaximumSwathHeightY / 2.0;
2832 MinimumSwathHeightC = MaximumSwathHeightC;
2833 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_10
2834 && mode_lib->vba.SourceScan[k] == dm_horz) {
2835 MinimumSwathHeightC = MaximumSwathHeightC / 2.0;
2836 MinimumSwathHeightY = MaximumSwathHeightY;
2838 MinimumSwathHeightY = MaximumSwathHeightY;
2839 MinimumSwathHeightC = MaximumSwathHeightC;
2843 if (mode_lib->vba.SourceScan[k] == dm_horz) {
2844 SwathWidth = mode_lib->vba.ViewportWidth[k];
2846 SwathWidth = mode_lib->vba.ViewportHeight[k];
2849 if (mode_lib->vba.ODMCombineEnabled[k] == true) {
2850 MainPlaneDoesODMCombine = true;
2852 for (j = 0; j < mode_lib->vba.NumberOfActivePlanes; ++j) {
2853 if (mode_lib->vba.BlendingAndTiming[k] == j
2854 && mode_lib->vba.ODMCombineEnabled[j] == true) {
2855 MainPlaneDoesODMCombine = true;
2859 if (MainPlaneDoesODMCombine == true) {
2860 SwathWidth = dml_min(
2862 mode_lib->vba.HActive[k] / 2.0 * mode_lib->vba.HRatio[k]);
2864 if (mode_lib->vba.DPPPerPlane[k] == 0)
2867 SwathWidth = SwathWidth / mode_lib->vba.DPPPerPlane[k];
2870 SwathWidthGranularityY = 256 / dml_ceil(BytePerPixDETY, 1) / MaximumSwathHeightY;
2871 RoundedUpMaxSwathSizeBytesY = (dml_ceil(
2872 (double) (SwathWidth - 1),
2873 SwathWidthGranularityY) + SwathWidthGranularityY) * BytePerPixDETY
2874 * MaximumSwathHeightY;
2875 if (mode_lib->vba.SourcePixelFormat[k] == dm_420_10) {
2876 RoundedUpMaxSwathSizeBytesY = dml_ceil(RoundedUpMaxSwathSizeBytesY, 256)
2879 if (MaximumSwathHeightC > 0) {
2880 SwathWidthGranularityC = 256.0 / dml_ceil(BytePerPixDETC, 2)
2881 / MaximumSwathHeightC;
2882 RoundedUpMaxSwathSizeBytesC = (dml_ceil(
2883 (double) (SwathWidth / 2.0 - 1),
2884 SwathWidthGranularityC) + SwathWidthGranularityC)
2885 * BytePerPixDETC * MaximumSwathHeightC;
2886 if (mode_lib->vba.SourcePixelFormat[k] == dm_420_10) {
2887 RoundedUpMaxSwathSizeBytesC = dml_ceil(
2888 RoundedUpMaxSwathSizeBytesC,
2892 RoundedUpMaxSwathSizeBytesC = 0.0;
2894 if (RoundedUpMaxSwathSizeBytesY + RoundedUpMaxSwathSizeBytesC
2895 <= mode_lib->vba.DETBufferSizeInKByte * 1024.0 / 2.0) {
2896 mode_lib->vba.SwathHeightY[k] = MaximumSwathHeightY;
2897 mode_lib->vba.SwathHeightC[k] = MaximumSwathHeightC;
2899 mode_lib->vba.SwathHeightY[k] = MinimumSwathHeightY;
2900 mode_lib->vba.SwathHeightC[k] = MinimumSwathHeightC;
2903 if (mode_lib->vba.SwathHeightC[k] == 0) {
2904 mode_lib->vba.DETBufferSizeY[k] = mode_lib->vba.DETBufferSizeInKByte * 1024;
2905 mode_lib->vba.DETBufferSizeC[k] = 0;
2906 } else if (mode_lib->vba.SwathHeightY[k] <= mode_lib->vba.SwathHeightC[k]) {
2907 mode_lib->vba.DETBufferSizeY[k] = mode_lib->vba.DETBufferSizeInKByte
2909 mode_lib->vba.DETBufferSizeC[k] = mode_lib->vba.DETBufferSizeInKByte
2912 mode_lib->vba.DETBufferSizeY[k] = mode_lib->vba.DETBufferSizeInKByte
2914 mode_lib->vba.DETBufferSizeC[k] = mode_lib->vba.DETBufferSizeInKByte
2920 static double CalculateTWait(
2921 unsigned int PrefetchMode,
2922 double DRAMClockChangeLatency,
2923 double UrgentLatencyPixelDataOnly,
2924 double SREnterPlusExitTime)
2926 if (PrefetchMode == 0) {
2928 DRAMClockChangeLatency + UrgentLatencyPixelDataOnly,
2929 dml_max(SREnterPlusExitTime, UrgentLatencyPixelDataOnly));
2930 } else if (PrefetchMode == 1) {
2931 return dml_max(SREnterPlusExitTime, UrgentLatencyPixelDataOnly);
2933 return UrgentLatencyPixelDataOnly;
2937 static double CalculateRemoteSurfaceFlipDelay(
2938 struct display_mode_lib *mode_lib,
2943 double XFCTSlvVupdateOffset,
2944 double XFCTSlvVupdateWidth,
2945 double XFCTSlvVreadyOffset,
2946 double XFCXBUFLatencyTolerance,
2947 double XFCFillBWOverhead,
2948 double XFCSlvChunkSize,
2949 double XFCBusTransportTime,
2952 double *SrcActiveDrainRate,
2956 double TSlvSetup, AvgfillRate, result;
2958 *SrcActiveDrainRate = VRatio * SwathWidth * Bpp / LineTime;
2959 TSlvSetup = XFCTSlvVupdateOffset + XFCTSlvVupdateWidth + XFCTSlvVreadyOffset;
2960 *TInitXFill = XFCXBUFLatencyTolerance / (1 + XFCFillBWOverhead / 100);
2961 AvgfillRate = *SrcActiveDrainRate * (1 + XFCFillBWOverhead / 100);
2962 *TslvChk = XFCSlvChunkSize / AvgfillRate;
2964 "DML::CalculateRemoteSurfaceFlipDelay: SrcActiveDrainRate: %f\n",
2965 *SrcActiveDrainRate);
2966 dml_print("DML::CalculateRemoteSurfaceFlipDelay: TSlvSetup: %f\n", TSlvSetup);
2967 dml_print("DML::CalculateRemoteSurfaceFlipDelay: TInitXFill: %f\n", *TInitXFill);
2968 dml_print("DML::CalculateRemoteSurfaceFlipDelay: AvgfillRate: %f\n", AvgfillRate);
2969 dml_print("DML::CalculateRemoteSurfaceFlipDelay: TslvChk: %f\n", *TslvChk);
2970 result = 2 * XFCBusTransportTime + TSlvSetup + TCalc + TWait + *TslvChk + *TInitXFill; // TODO: This doesn't seem to match programming guide
2971 dml_print("DML::CalculateRemoteSurfaceFlipDelay: RemoteSurfaceFlipDelay: %f\n", result);
2975 static double CalculateWriteBackDelay(
2976 enum source_format_class WritebackPixelFormat,
2977 double WritebackHRatio,
2978 double WritebackVRatio,
2979 unsigned int WritebackLumaHTaps,
2980 unsigned int WritebackLumaVTaps,
2981 unsigned int WritebackChromaHTaps,
2982 unsigned int WritebackChromaVTaps,
2983 unsigned int WritebackDestinationWidth)
2985 double CalculateWriteBackDelay =
2987 dml_ceil(WritebackLumaHTaps / 4.0, 1) / WritebackHRatio,
2988 WritebackLumaVTaps * dml_ceil(1.0 / WritebackVRatio, 1)
2990 WritebackDestinationWidth
2993 + dml_ceil(1.0 / WritebackVRatio, 1)
2999 if (WritebackPixelFormat != dm_444_32) {
3000 CalculateWriteBackDelay =
3002 CalculateWriteBackDelay,
3005 WritebackChromaHTaps
3010 WritebackChromaVTaps
3017 WritebackDestinationWidth
3027 WritebackChromaVTaps
3032 return CalculateWriteBackDelay;
3035 static void CalculateActiveRowBandwidth(
3037 enum source_format_class SourcePixelFormat,
3041 unsigned int MetaRowByteLuma,
3042 unsigned int MetaRowByteChroma,
3043 unsigned int meta_row_height_luma,
3044 unsigned int meta_row_height_chroma,
3045 unsigned int PixelPTEBytesPerRowLuma,
3046 unsigned int PixelPTEBytesPerRowChroma,
3047 unsigned int dpte_row_height_luma,
3048 unsigned int dpte_row_height_chroma,
3049 double *meta_row_bw,
3050 double *dpte_row_bw,
3051 double *qual_row_bw)
3053 if (DCCEnable != true) {
3055 } else if (SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10) {
3056 *meta_row_bw = VRatio * MetaRowByteLuma / (meta_row_height_luma * LineTime)
3057 + VRatio / 2 * MetaRowByteChroma
3058 / (meta_row_height_chroma * LineTime);
3060 *meta_row_bw = VRatio * MetaRowByteLuma / (meta_row_height_luma * LineTime);
3063 if (GPUVMEnable != true) {
3065 } else if (SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10) {
3066 *dpte_row_bw = VRatio * PixelPTEBytesPerRowLuma / (dpte_row_height_luma * LineTime)
3067 + VRatio / 2 * PixelPTEBytesPerRowChroma
3068 / (dpte_row_height_chroma * LineTime);
3070 *dpte_row_bw = VRatio * PixelPTEBytesPerRowLuma / (dpte_row_height_luma * LineTime);
3073 if ((SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10)) {
3074 *qual_row_bw = *meta_row_bw + *dpte_row_bw;
3080 static void CalculateFlipSchedule(
3081 struct display_mode_lib *mode_lib,
3082 double UrgentExtraLatency,
3083 double UrgentLatencyPixelDataOnly,
3084 unsigned int GPUVMMaxPageTableLevels,
3086 double BandwidthAvailableForImmediateFlip,
3087 unsigned int TotImmediateFlipBytes,
3088 enum source_format_class SourcePixelFormat,
3089 unsigned int ImmediateFlipBytes,
3093 double PDEAndMetaPTEBytesFrame,
3094 unsigned int MetaRowByte,
3095 unsigned int PixelPTEBytesPerRow,
3097 unsigned int dpte_row_height,
3098 unsigned int meta_row_height,
3100 double *DestinationLinesToRequestVMInImmediateFlip,
3101 double *DestinationLinesToRequestRowInImmediateFlip,
3102 double *final_flip_bw,
3103 bool *ImmediateFlipSupportedForPipe)
3105 double min_row_time = 0.0;
3107 if (SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10) {
3108 *DestinationLinesToRequestVMInImmediateFlip = 0.0;
3109 *DestinationLinesToRequestRowInImmediateFlip = 0.0;
3110 *final_flip_bw = qual_row_bw;
3111 *ImmediateFlipSupportedForPipe = true;
3113 double TimeForFetchingMetaPTEImmediateFlip;
3114 double TimeForFetchingRowInVBlankImmediateFlip;
3116 if (GPUVMEnable == true) {
3117 mode_lib->vba.ImmediateFlipBW[0] = BandwidthAvailableForImmediateFlip
3118 * ImmediateFlipBytes / TotImmediateFlipBytes;
3119 TimeForFetchingMetaPTEImmediateFlip =
3122 + PDEAndMetaPTEBytesFrame
3123 / mode_lib->vba.ImmediateFlipBW[0],
3126 + UrgentLatencyPixelDataOnly
3127 * (GPUVMMaxPageTableLevels
3131 TimeForFetchingMetaPTEImmediateFlip = 0;
3134 *DestinationLinesToRequestVMInImmediateFlip = dml_floor(
3135 4.0 * (TimeForFetchingMetaPTEImmediateFlip / LineTime + 0.125),
3138 if ((GPUVMEnable == true || DCCEnable == true)) {
3139 mode_lib->vba.ImmediateFlipBW[0] = BandwidthAvailableForImmediateFlip
3140 * ImmediateFlipBytes / TotImmediateFlipBytes;
3141 TimeForFetchingRowInVBlankImmediateFlip = dml_max(
3142 (MetaRowByte + PixelPTEBytesPerRow)
3143 / mode_lib->vba.ImmediateFlipBW[0],
3144 dml_max(UrgentLatencyPixelDataOnly, LineTime / 4.0));
3146 TimeForFetchingRowInVBlankImmediateFlip = 0;
3149 *DestinationLinesToRequestRowInImmediateFlip = dml_floor(
3150 4.0 * (TimeForFetchingRowInVBlankImmediateFlip / LineTime + 0.125),
3153 if (GPUVMEnable == true) {
3156 PDEAndMetaPTEBytesFrame
3157 / (*DestinationLinesToRequestVMInImmediateFlip
3159 (MetaRowByte + PixelPTEBytesPerRow)
3160 / (TimeForFetchingRowInVBlankImmediateFlip
3162 } else if (MetaRowByte + PixelPTEBytesPerRow > 0) {
3163 *final_flip_bw = (MetaRowByte + PixelPTEBytesPerRow)
3164 / (TimeForFetchingRowInVBlankImmediateFlip * LineTime);
3169 if (GPUVMEnable && !DCCEnable)
3170 min_row_time = dpte_row_height * LineTime / VRatio;
3171 else if (!GPUVMEnable && DCCEnable)
3172 min_row_time = meta_row_height * LineTime / VRatio;
3174 min_row_time = dml_min(dpte_row_height, meta_row_height) * LineTime
3177 if (*DestinationLinesToRequestVMInImmediateFlip >= 8
3178 || *DestinationLinesToRequestRowInImmediateFlip >= 16
3179 || TimeForFetchingMetaPTEImmediateFlip
3180 + 2 * TimeForFetchingRowInVBlankImmediateFlip
3182 *ImmediateFlipSupportedForPipe = false;
3184 *ImmediateFlipSupportedForPipe = true;
3188 static unsigned int TruncToValidBPP(
3191 enum output_encoder_class Output,
3192 enum output_format_class Format,
3193 unsigned int DSCInputBitPerComponent)
3195 if (Output == dm_hdmi) {
3196 if (Format == dm_420) {
3197 if (DecimalBPP >= 18)
3199 else if (DecimalBPP >= 15)
3201 else if (DecimalBPP >= 12)
3205 } else if (Format == dm_444) {
3206 if (DecimalBPP >= 36)
3208 else if (DecimalBPP >= 30)
3210 else if (DecimalBPP >= 24)
3212 else if (DecimalBPP >= 18)
3217 if (DecimalBPP / 1.5 >= 24)
3219 else if (DecimalBPP / 1.5 >= 20)
3221 else if (DecimalBPP / 1.5 >= 16)
3228 if (Format == dm_420) {
3231 else if (DecimalBPP >= 1.5 * DSCInputBitPerComponent - 1 / 16)
3232 return 1.5 * DSCInputBitPerComponent - 1 / 16;
3234 return dml_floor(16 * DecimalBPP, 1) / 16;
3235 } else if (Format == dm_n422) {
3238 else if (DecimalBPP >= 2 * DSCInputBitPerComponent - 1 / 16)
3239 return 2 * DSCInputBitPerComponent - 1 / 16;
3241 return dml_floor(16 * DecimalBPP, 1) / 16;
3245 else if (DecimalBPP >= 3 * DSCInputBitPerComponent - 1 / 16)
3246 return 3 * DSCInputBitPerComponent - 1 / 16;
3248 return dml_floor(16 * DecimalBPP, 1) / 16;
3250 } else if (Format == dm_420) {
3251 if (DecimalBPP >= 18)
3253 else if (DecimalBPP >= 15)
3255 else if (DecimalBPP >= 12)
3259 } else if (Format == dm_s422 || Format == dm_n422) {
3260 if (DecimalBPP >= 24)
3262 else if (DecimalBPP >= 20)
3264 else if (DecimalBPP >= 16)
3269 if (DecimalBPP >= 36)
3271 else if (DecimalBPP >= 30)
3273 else if (DecimalBPP >= 24)
3275 else if (DecimalBPP >= 18)
3283 void dml20_ModeSupportAndSystemConfigurationFull(struct display_mode_lib *mode_lib)
3285 struct vba_vars_st *locals = &mode_lib->vba;
3288 unsigned int j, k, m;
3290 /*MODE SUPPORT, VOLTAGE STATE AND SOC CONFIGURATION*/
3292 /*Scale Ratio, taps Support Check*/
3294 mode_lib->vba.ScaleRatioAndTapsSupport = true;
3295 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
3296 if (mode_lib->vba.ScalerEnabled[k] == false
3297 && ((mode_lib->vba.SourcePixelFormat[k] != dm_444_64
3298 && mode_lib->vba.SourcePixelFormat[k] != dm_444_32
3299 && mode_lib->vba.SourcePixelFormat[k] != dm_444_16
3300 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_16
3301 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_8)
3302 || mode_lib->vba.HRatio[k] != 1.0
3303 || mode_lib->vba.htaps[k] != 1.0
3304 || mode_lib->vba.VRatio[k] != 1.0
3305 || mode_lib->vba.vtaps[k] != 1.0)) {
3306 mode_lib->vba.ScaleRatioAndTapsSupport = false;
3307 } else if (mode_lib->vba.vtaps[k] < 1.0 || mode_lib->vba.vtaps[k] > 8.0
3308 || mode_lib->vba.htaps[k] < 1.0 || mode_lib->vba.htaps[k] > 8.0
3309 || (mode_lib->vba.htaps[k] > 1.0
3310 && (mode_lib->vba.htaps[k] % 2) == 1)
3311 || mode_lib->vba.HRatio[k] > mode_lib->vba.MaxHSCLRatio
3312 || mode_lib->vba.VRatio[k] > mode_lib->vba.MaxVSCLRatio
3313 || mode_lib->vba.HRatio[k] > mode_lib->vba.htaps[k]
3314 || mode_lib->vba.VRatio[k] > mode_lib->vba.vtaps[k]
3315 || (mode_lib->vba.SourcePixelFormat[k] != dm_444_64
3316 && mode_lib->vba.SourcePixelFormat[k] != dm_444_32
3317 && mode_lib->vba.SourcePixelFormat[k] != dm_444_16
3318 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_16
3319 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_8
3320 && (mode_lib->vba.HRatio[k] / 2.0
3321 > mode_lib->vba.HTAPsChroma[k]
3322 || mode_lib->vba.VRatio[k] / 2.0
3323 > mode_lib->vba.VTAPsChroma[k]))) {
3324 mode_lib->vba.ScaleRatioAndTapsSupport = false;
3327 /*Source Format, Pixel Format and Scan Support Check*/
3329 mode_lib->vba.SourceFormatPixelAndScanSupport = true;
3330 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
3331 if ((mode_lib->vba.SurfaceTiling[k] == dm_sw_linear
3332 && mode_lib->vba.SourceScan[k] != dm_horz)
3333 || ((mode_lib->vba.SurfaceTiling[k] == dm_sw_4kb_d
3334 || mode_lib->vba.SurfaceTiling[k] == dm_sw_4kb_d_x
3335 || mode_lib->vba.SurfaceTiling[k] == dm_sw_64kb_d
3336 || mode_lib->vba.SurfaceTiling[k] == dm_sw_64kb_d_t
3337 || mode_lib->vba.SurfaceTiling[k] == dm_sw_64kb_d_x
3338 || mode_lib->vba.SurfaceTiling[k] == dm_sw_var_d
3339 || mode_lib->vba.SurfaceTiling[k] == dm_sw_var_d_x)
3340 && mode_lib->vba.SourcePixelFormat[k] != dm_444_64)
3341 || (mode_lib->vba.SurfaceTiling[k] == dm_sw_64kb_r_x
3342 && (mode_lib->vba.SourcePixelFormat[k] == dm_mono_8
3343 || mode_lib->vba.SourcePixelFormat[k]
3345 || mode_lib->vba.SourcePixelFormat[k]
3347 || (((mode_lib->vba.SurfaceTiling[k] == dm_sw_gfx7_2d_thin_gl
3348 || mode_lib->vba.SurfaceTiling[k]
3349 == dm_sw_gfx7_2d_thin_lvp)
3350 && !((mode_lib->vba.SourcePixelFormat[k]
3352 || mode_lib->vba.SourcePixelFormat[k]
3354 && mode_lib->vba.SourceScan[k]
3356 && mode_lib->vba.SupportGFX7CompatibleTilingIn32bppAnd64bpp
3358 && mode_lib->vba.DCCEnable[k]
3360 || (mode_lib->vba.DCCEnable[k] == true
3361 && (mode_lib->vba.SurfaceTiling[k]
3363 || mode_lib->vba.SourcePixelFormat[k]
3365 || mode_lib->vba.SourcePixelFormat[k]
3367 mode_lib->vba.SourceFormatPixelAndScanSupport = false;
3370 /*Bandwidth Support Check*/
3372 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
3373 if (mode_lib->vba.SourcePixelFormat[k] == dm_444_64) {
3374 locals->BytePerPixelInDETY[k] = 8.0;
3375 locals->BytePerPixelInDETC[k] = 0.0;
3376 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_32) {
3377 locals->BytePerPixelInDETY[k] = 4.0;
3378 locals->BytePerPixelInDETC[k] = 0.0;
3379 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_16
3380 || mode_lib->vba.SourcePixelFormat[k] == dm_mono_16) {
3381 locals->BytePerPixelInDETY[k] = 2.0;
3382 locals->BytePerPixelInDETC[k] = 0.0;
3383 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_mono_8) {
3384 locals->BytePerPixelInDETY[k] = 1.0;
3385 locals->BytePerPixelInDETC[k] = 0.0;
3386 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_8) {
3387 locals->BytePerPixelInDETY[k] = 1.0;
3388 locals->BytePerPixelInDETC[k] = 2.0;
3390 locals->BytePerPixelInDETY[k] = 4.0 / 3;
3391 locals->BytePerPixelInDETC[k] = 8.0 / 3;
3393 if (mode_lib->vba.SourceScan[k] == dm_horz) {
3394 locals->SwathWidthYSingleDPP[k] = mode_lib->vba.ViewportWidth[k];
3396 locals->SwathWidthYSingleDPP[k] = mode_lib->vba.ViewportHeight[k];
3399 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
3400 locals->ReadBandwidthLuma[k] = locals->SwathWidthYSingleDPP[k] * dml_ceil(locals->BytePerPixelInDETY[k], 1.0)
3401 / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]) * mode_lib->vba.VRatio[k];
3402 locals->ReadBandwidthChroma[k] = locals->SwathWidthYSingleDPP[k] / 2 * dml_ceil(locals->BytePerPixelInDETC[k], 2.0)
3403 / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]) * mode_lib->vba.VRatio[k] / 2.0;
3404 locals->ReadBandwidth[k] = locals->ReadBandwidthLuma[k] + locals->ReadBandwidthChroma[k];
3406 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
3407 if (mode_lib->vba.WritebackEnable[k] == true
3408 && mode_lib->vba.WritebackPixelFormat[k] == dm_444_32) {
3409 locals->WriteBandwidth[k] = mode_lib->vba.WritebackDestinationWidth[k]
3410 * mode_lib->vba.WritebackDestinationHeight[k]
3411 / (mode_lib->vba.WritebackSourceHeight[k]
3412 * mode_lib->vba.HTotal[k]
3413 / mode_lib->vba.PixelClock[k]) * 4.0;
3414 } else if (mode_lib->vba.WritebackEnable[k] == true
3415 && mode_lib->vba.WritebackPixelFormat[k] == dm_420_10) {
3416 locals->WriteBandwidth[k] = mode_lib->vba.WritebackDestinationWidth[k]
3417 * mode_lib->vba.WritebackDestinationHeight[k]
3418 / (mode_lib->vba.WritebackSourceHeight[k]
3419 * mode_lib->vba.HTotal[k]
3420 / mode_lib->vba.PixelClock[k]) * 3.0;
3421 } else if (mode_lib->vba.WritebackEnable[k] == true) {
3422 locals->WriteBandwidth[k] = mode_lib->vba.WritebackDestinationWidth[k]
3423 * mode_lib->vba.WritebackDestinationHeight[k]
3424 / (mode_lib->vba.WritebackSourceHeight[k]
3425 * mode_lib->vba.HTotal[k]
3426 / mode_lib->vba.PixelClock[k]) * 1.5;
3428 locals->WriteBandwidth[k] = 0.0;
3431 mode_lib->vba.DCCEnabledInAnyPlane = false;
3432 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
3433 if (mode_lib->vba.DCCEnable[k] == true) {
3434 mode_lib->vba.DCCEnabledInAnyPlane = true;
3437 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) {
3438 locals->FabricAndDRAMBandwidthPerState[i] = dml_min(
3439 mode_lib->vba.DRAMSpeedPerState[i] * mode_lib->vba.NumberOfChannels
3440 * mode_lib->vba.DRAMChannelWidth,
3441 mode_lib->vba.FabricClockPerState[i]
3442 * mode_lib->vba.FabricDatapathToDCNDataReturn) / 1000;
3443 locals->ReturnBWToDCNPerState = dml_min(locals->ReturnBusWidth * locals->DCFCLKPerState[i],
3444 locals->FabricAndDRAMBandwidthPerState[i] * 1000)
3445 * locals->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelDataOnly / 100;
3447 locals->ReturnBWPerState[i] = locals->ReturnBWToDCNPerState;
3449 if (locals->DCCEnabledInAnyPlane == true && locals->ReturnBWToDCNPerState > locals->DCFCLKPerState[i] * locals->ReturnBusWidth / 4) {
3450 locals->ReturnBWPerState[i] = dml_min(locals->ReturnBWPerState[i],
3451 locals->ReturnBWToDCNPerState * 4 * (1 - locals->UrgentLatency /
3452 ((locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024
3453 / (locals->ReturnBWToDCNPerState - locals->DCFCLKPerState[i]
3454 * locals->ReturnBusWidth / 4) + locals->UrgentLatency)));
3456 locals->CriticalPoint = 2 * locals->ReturnBusWidth * locals->DCFCLKPerState[i] *
3457 locals->UrgentLatency / (locals->ReturnBWToDCNPerState * locals->UrgentLatency
3458 + (locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024);
3460 if (locals->DCCEnabledInAnyPlane && locals->CriticalPoint > 1 && locals->CriticalPoint < 4) {
3461 locals->ReturnBWPerState[i] = dml_min(locals->ReturnBWPerState[i],
3462 4 * locals->ReturnBWToDCNPerState *
3463 (locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024
3464 * locals->ReturnBusWidth * locals->DCFCLKPerState[i] * locals->UrgentLatency /
3465 dml_pow((locals->ReturnBWToDCNPerState * locals->UrgentLatency
3466 + (locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024), 2));
3469 locals->ReturnBWToDCNPerState = dml_min(locals->ReturnBusWidth *
3470 locals->DCFCLKPerState[i], locals->FabricAndDRAMBandwidthPerState[i] * 1000);
3472 if (locals->DCCEnabledInAnyPlane == true && locals->ReturnBWToDCNPerState > locals->DCFCLKPerState[i] * locals->ReturnBusWidth / 4) {
3473 locals->ReturnBWPerState[i] = dml_min(locals->ReturnBWPerState[i],
3474 locals->ReturnBWToDCNPerState * 4 * (1 - locals->UrgentLatency /
3475 ((locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024
3476 / (locals->ReturnBWToDCNPerState - locals->DCFCLKPerState[i]
3477 * locals->ReturnBusWidth / 4) + locals->UrgentLatency)));
3479 locals->CriticalPoint = 2 * locals->ReturnBusWidth * locals->DCFCLKPerState[i] *
3480 locals->UrgentLatency / (locals->ReturnBWToDCNPerState * locals->UrgentLatency
3481 + (locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024);
3483 if (locals->DCCEnabledInAnyPlane && locals->CriticalPoint > 1 && locals->CriticalPoint < 4) {
3484 locals->ReturnBWPerState[i] = dml_min(locals->ReturnBWPerState[i],
3485 4 * locals->ReturnBWToDCNPerState *
3486 (locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024
3487 * locals->ReturnBusWidth * locals->DCFCLKPerState[i] * locals->UrgentLatency /
3488 dml_pow((locals->ReturnBWToDCNPerState * locals->UrgentLatency
3489 + (locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024), 2));
3492 /*Writeback Latency support check*/
3494 mode_lib->vba.WritebackLatencySupport = true;
3495 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
3496 if (mode_lib->vba.WritebackEnable[k] == true) {
3497 if (mode_lib->vba.WritebackPixelFormat[k] == dm_444_32) {
3498 if (locals->WriteBandwidth[k]
3499 > (mode_lib->vba.WritebackInterfaceLumaBufferSize
3500 + mode_lib->vba.WritebackInterfaceChromaBufferSize)
3501 / mode_lib->vba.WritebackLatency) {
3502 mode_lib->vba.WritebackLatencySupport = false;
3505 if (locals->WriteBandwidth[k]
3508 mode_lib->vba.WritebackInterfaceLumaBufferSize,
3510 * mode_lib->vba.WritebackInterfaceChromaBufferSize)
3511 / mode_lib->vba.WritebackLatency) {
3512 mode_lib->vba.WritebackLatencySupport = false;
3517 /*Re-ordering Buffer Support Check*/
3519 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) {
3520 locals->UrgentRoundTripAndOutOfOrderLatencyPerState[i] =
3521 (mode_lib->vba.RoundTripPingLatencyCycles + 32.0) / mode_lib->vba.DCFCLKPerState[i]
3522 + locals->UrgentOutOfOrderReturnPerChannel * mode_lib->vba.NumberOfChannels / locals->ReturnBWPerState[i];
3523 if ((mode_lib->vba.ROBBufferSizeInKByte - mode_lib->vba.PixelChunkSizeInKByte) * 1024.0 / locals->ReturnBWPerState[i]
3524 > locals->UrgentRoundTripAndOutOfOrderLatencyPerState[i]) {
3525 locals->ROBSupport[i] = true;
3527 locals->ROBSupport[i] = false;
3530 /*Writeback Mode Support Check*/
3532 mode_lib->vba.TotalNumberOfActiveWriteback = 0;
3533 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
3534 if (mode_lib->vba.WritebackEnable[k] == true) {
3535 if (mode_lib->vba.ActiveWritebacksPerPlane[k] == 0)
3536 mode_lib->vba.ActiveWritebacksPerPlane[k] = 1;
3537 mode_lib->vba.TotalNumberOfActiveWriteback =
3538 mode_lib->vba.TotalNumberOfActiveWriteback
3539 + mode_lib->vba.ActiveWritebacksPerPlane[k];
3542 mode_lib->vba.WritebackModeSupport = true;
3543 if (mode_lib->vba.TotalNumberOfActiveWriteback > mode_lib->vba.MaxNumWriteback) {
3544 mode_lib->vba.WritebackModeSupport = false;
3546 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
3547 if (mode_lib->vba.WritebackEnable[k] == true
3548 && mode_lib->vba.Writeback10bpc420Supported != true
3549 && mode_lib->vba.WritebackPixelFormat[k] == dm_420_10) {
3550 mode_lib->vba.WritebackModeSupport = false;
3553 /*Writeback Scale Ratio and Taps Support Check*/
3555 mode_lib->vba.WritebackScaleRatioAndTapsSupport = true;
3556 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
3557 if (mode_lib->vba.WritebackEnable[k] == true) {
3558 if (mode_lib->vba.WritebackLumaAndChromaScalingSupported == false
3559 && (mode_lib->vba.WritebackHRatio[k] != 1.0
3560 || mode_lib->vba.WritebackVRatio[k] != 1.0)) {
3561 mode_lib->vba.WritebackScaleRatioAndTapsSupport = false;
3563 if (mode_lib->vba.WritebackHRatio[k] > mode_lib->vba.WritebackMaxHSCLRatio
3564 || mode_lib->vba.WritebackVRatio[k]
3565 > mode_lib->vba.WritebackMaxVSCLRatio
3566 || mode_lib->vba.WritebackHRatio[k]
3567 < mode_lib->vba.WritebackMinHSCLRatio
3568 || mode_lib->vba.WritebackVRatio[k]
3569 < mode_lib->vba.WritebackMinVSCLRatio
3570 || mode_lib->vba.WritebackLumaHTaps[k]
3571 > mode_lib->vba.WritebackMaxHSCLTaps
3572 || mode_lib->vba.WritebackLumaVTaps[k]
3573 > mode_lib->vba.WritebackMaxVSCLTaps
3574 || mode_lib->vba.WritebackHRatio[k]
3575 > mode_lib->vba.WritebackLumaHTaps[k]
3576 || mode_lib->vba.WritebackVRatio[k]
3577 > mode_lib->vba.WritebackLumaVTaps[k]
3578 || (mode_lib->vba.WritebackLumaHTaps[k] > 2.0
3579 && ((mode_lib->vba.WritebackLumaHTaps[k] % 2)
3581 || (mode_lib->vba.WritebackPixelFormat[k] != dm_444_32
3582 && (mode_lib->vba.WritebackChromaHTaps[k]
3583 > mode_lib->vba.WritebackMaxHSCLTaps
3584 || mode_lib->vba.WritebackChromaVTaps[k]
3585 > mode_lib->vba.WritebackMaxVSCLTaps
3587 * mode_lib->vba.WritebackHRatio[k]
3588 > mode_lib->vba.WritebackChromaHTaps[k]
3590 * mode_lib->vba.WritebackVRatio[k]
3591 > mode_lib->vba.WritebackChromaVTaps[k]
3592 || (mode_lib->vba.WritebackChromaHTaps[k] > 2.0
3593 && ((mode_lib->vba.WritebackChromaHTaps[k] % 2) == 1))))) {
3594 mode_lib->vba.WritebackScaleRatioAndTapsSupport = false;
3596 if (mode_lib->vba.WritebackVRatio[k] < 1.0) {
3597 mode_lib->vba.WritebackLumaVExtra =
3598 dml_max(1.0 - 2.0 / dml_ceil(1.0 / mode_lib->vba.WritebackVRatio[k], 1.0), 0.0);
3600 mode_lib->vba.WritebackLumaVExtra = -1;
3602 if ((mode_lib->vba.WritebackPixelFormat[k] == dm_444_32
3603 && mode_lib->vba.WritebackLumaVTaps[k]
3604 > (mode_lib->vba.WritebackLineBufferLumaBufferSize
3605 + mode_lib->vba.WritebackLineBufferChromaBufferSize)
3607 / mode_lib->vba.WritebackDestinationWidth[k]
3608 - mode_lib->vba.WritebackLumaVExtra)
3609 || (mode_lib->vba.WritebackPixelFormat[k] == dm_420_8
3610 && mode_lib->vba.WritebackLumaVTaps[k]
3611 > mode_lib->vba.WritebackLineBufferLumaBufferSize
3612 * 8.0 / 10.0 / mode_lib->vba.WritebackDestinationWidth[k]
3613 - mode_lib->vba.WritebackLumaVExtra)
3614 || (mode_lib->vba.WritebackPixelFormat[k] == dm_420_10
3615 && mode_lib->vba.WritebackLumaVTaps[k]
3616 > mode_lib->vba.WritebackLineBufferLumaBufferSize
3618 / mode_lib->vba.WritebackDestinationWidth[k]
3619 - mode_lib->vba.WritebackLumaVExtra)) {
3620 mode_lib->vba.WritebackScaleRatioAndTapsSupport = false;
3622 if (2.0 * mode_lib->vba.WritebackVRatio[k] < 1) {
3623 mode_lib->vba.WritebackChromaVExtra = 0.0;
3625 mode_lib->vba.WritebackChromaVExtra = -1;
3627 if ((mode_lib->vba.WritebackPixelFormat[k] == dm_420_8
3628 && mode_lib->vba.WritebackChromaVTaps[k]
3629 > mode_lib->vba.WritebackLineBufferChromaBufferSize
3630 * 8.0 / 10.0 / mode_lib->vba.WritebackDestinationWidth[k]
3631 - mode_lib->vba.WritebackChromaVExtra)
3632 || (mode_lib->vba.WritebackPixelFormat[k] == dm_420_10
3633 && mode_lib->vba.WritebackChromaVTaps[k]
3634 > mode_lib->vba.WritebackLineBufferChromaBufferSize
3636 / mode_lib->vba.WritebackDestinationWidth[k]
3637 - mode_lib->vba.WritebackChromaVExtra)) {
3638 mode_lib->vba.WritebackScaleRatioAndTapsSupport = false;
3642 /*Maximum DISPCLK/DPPCLK Support check*/
3644 mode_lib->vba.WritebackRequiredDISPCLK = 0.0;
3645 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
3646 if (mode_lib->vba.WritebackEnable[k] == true) {
3647 mode_lib->vba.WritebackRequiredDISPCLK =
3649 mode_lib->vba.WritebackRequiredDISPCLK,
3650 CalculateWriteBackDISPCLK(
3651 mode_lib->vba.WritebackPixelFormat[k],
3652 mode_lib->vba.PixelClock[k],
3653 mode_lib->vba.WritebackHRatio[k],
3654 mode_lib->vba.WritebackVRatio[k],
3655 mode_lib->vba.WritebackLumaHTaps[k],
3656 mode_lib->vba.WritebackLumaVTaps[k],
3657 mode_lib->vba.WritebackChromaHTaps[k],
3658 mode_lib->vba.WritebackChromaVTaps[k],
3659 mode_lib->vba.WritebackDestinationWidth[k],
3660 mode_lib->vba.HTotal[k],
3661 mode_lib->vba.WritebackChromaLineBufferWidth));
3664 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
3665 if (mode_lib->vba.HRatio[k] > 1.0) {
3666 locals->PSCL_FACTOR[k] = dml_min(
3667 mode_lib->vba.MaxDCHUBToPSCLThroughput,
3668 mode_lib->vba.MaxPSCLToLBThroughput
3669 * mode_lib->vba.HRatio[k]
3671 mode_lib->vba.htaps[k]
3675 locals->PSCL_FACTOR[k] = dml_min(
3676 mode_lib->vba.MaxDCHUBToPSCLThroughput,
3677 mode_lib->vba.MaxPSCLToLBThroughput);
3679 if (locals->BytePerPixelInDETC[k] == 0.0) {
3680 locals->PSCL_FACTOR_CHROMA[k] = 0.0;
3681 locals->MinDPPCLKUsingSingleDPP[k] =
3682 mode_lib->vba.PixelClock[k]
3684 mode_lib->vba.vtaps[k] / 6.0
3687 mode_lib->vba.HRatio[k]),
3688 mode_lib->vba.HRatio[k]
3689 * mode_lib->vba.VRatio[k]
3690 / locals->PSCL_FACTOR[k],
3692 if ((mode_lib->vba.htaps[k] > 6.0 || mode_lib->vba.vtaps[k] > 6.0)
3693 && locals->MinDPPCLKUsingSingleDPP[k]
3694 < 2.0 * mode_lib->vba.PixelClock[k]) {
3695 locals->MinDPPCLKUsingSingleDPP[k] = 2.0
3696 * mode_lib->vba.PixelClock[k];
3699 if (mode_lib->vba.HRatio[k] / 2.0 > 1.0) {
3700 locals->PSCL_FACTOR_CHROMA[k] =
3702 mode_lib->vba.MaxDCHUBToPSCLThroughput,
3703 mode_lib->vba.MaxPSCLToLBThroughput
3704 * mode_lib->vba.HRatio[k]
3707 mode_lib->vba.HTAPsChroma[k]
3711 locals->PSCL_FACTOR_CHROMA[k] = dml_min(
3712 mode_lib->vba.MaxDCHUBToPSCLThroughput,
3713 mode_lib->vba.MaxPSCLToLBThroughput);
3715 locals->MinDPPCLKUsingSingleDPP[k] =
3716 mode_lib->vba.PixelClock[k]
3718 mode_lib->vba.vtaps[k] / 6.0
3721 mode_lib->vba.HRatio[k]),
3722 mode_lib->vba.HRatio[k]
3723 * mode_lib->vba.VRatio[k]
3724 / locals->PSCL_FACTOR[k],
3725 mode_lib->vba.VTAPsChroma[k]
3729 mode_lib->vba.HRatio[k]
3731 mode_lib->vba.HRatio[k]
3732 * mode_lib->vba.VRatio[k]
3734 / locals->PSCL_FACTOR_CHROMA[k],
3736 if ((mode_lib->vba.htaps[k] > 6.0 || mode_lib->vba.vtaps[k] > 6.0
3737 || mode_lib->vba.HTAPsChroma[k] > 6.0
3738 || mode_lib->vba.VTAPsChroma[k] > 6.0)
3739 && locals->MinDPPCLKUsingSingleDPP[k]
3740 < 2.0 * mode_lib->vba.PixelClock[k]) {
3741 locals->MinDPPCLKUsingSingleDPP[k] = 2.0
3742 * mode_lib->vba.PixelClock[k];
3746 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
3747 Calculate256BBlockSizes(
3748 mode_lib->vba.SourcePixelFormat[k],
3749 mode_lib->vba.SurfaceTiling[k],
3750 dml_ceil(locals->BytePerPixelInDETY[k], 1.0),
3751 dml_ceil(locals->BytePerPixelInDETC[k], 2.0),
3752 &locals->Read256BlockHeightY[k],
3753 &locals->Read256BlockHeightC[k],
3754 &locals->Read256BlockWidthY[k],
3755 &locals->Read256BlockWidthC[k]);
3756 if (mode_lib->vba.SourceScan[k] == dm_horz) {
3757 locals->MaxSwathHeightY[k] = locals->Read256BlockHeightY[k];
3758 locals->MaxSwathHeightC[k] = locals->Read256BlockHeightC[k];
3760 locals->MaxSwathHeightY[k] = locals->Read256BlockWidthY[k];
3761 locals->MaxSwathHeightC[k] = locals->Read256BlockWidthC[k];
3763 if ((mode_lib->vba.SourcePixelFormat[k] == dm_444_64
3764 || mode_lib->vba.SourcePixelFormat[k] == dm_444_32
3765 || mode_lib->vba.SourcePixelFormat[k] == dm_444_16
3766 || mode_lib->vba.SourcePixelFormat[k] == dm_mono_16
3767 || mode_lib->vba.SourcePixelFormat[k] == dm_mono_8)) {
3768 if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear
3769 || (mode_lib->vba.SourcePixelFormat[k] == dm_444_64
3770 && (mode_lib->vba.SurfaceTiling[k]
3772 || mode_lib->vba.SurfaceTiling[k]
3774 || mode_lib->vba.SurfaceTiling[k]
3776 || mode_lib->vba.SurfaceTiling[k]
3778 || mode_lib->vba.SurfaceTiling[k]
3780 || mode_lib->vba.SurfaceTiling[k]
3782 || mode_lib->vba.SurfaceTiling[k]
3784 && mode_lib->vba.SourceScan[k] == dm_horz)) {
3785 locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k];
3787 locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k]
3790 locals->MinSwathHeightC[k] = locals->MaxSwathHeightC[k];
3792 if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear) {
3793 locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k];
3794 locals->MinSwathHeightC[k] = locals->MaxSwathHeightC[k];
3795 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_8
3796 && mode_lib->vba.SourceScan[k] == dm_horz) {
3797 locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k]
3799 locals->MinSwathHeightC[k] = locals->MaxSwathHeightC[k];
3800 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_10
3801 && mode_lib->vba.SourceScan[k] == dm_horz) {
3802 locals->MinSwathHeightC[k] = locals->MaxSwathHeightC[k]
3804 locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k];
3806 locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k];
3807 locals->MinSwathHeightC[k] = locals->MaxSwathHeightC[k];
3810 if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear) {
3811 mode_lib->vba.MaximumSwathWidthSupport = 8192.0;
3813 mode_lib->vba.MaximumSwathWidthSupport = 5120.0;
3815 mode_lib->vba.MaximumSwathWidthInDETBuffer =
3817 mode_lib->vba.MaximumSwathWidthSupport,
3818 mode_lib->vba.DETBufferSizeInKByte * 1024.0 / 2.0
3819 / (locals->BytePerPixelInDETY[k]
3820 * locals->MinSwathHeightY[k]
3821 + locals->BytePerPixelInDETC[k]
3823 * locals->MinSwathHeightC[k]));
3824 if (locals->BytePerPixelInDETC[k] == 0.0) {
3825 mode_lib->vba.MaximumSwathWidthInLineBuffer =
3826 mode_lib->vba.LineBufferSize
3827 * dml_max(mode_lib->vba.HRatio[k], 1.0)
3828 / mode_lib->vba.LBBitPerPixel[k]
3829 / (mode_lib->vba.vtaps[k]
3832 mode_lib->vba.VRatio[k],
3837 mode_lib->vba.MaximumSwathWidthInLineBuffer =
3839 mode_lib->vba.LineBufferSize
3841 mode_lib->vba.HRatio[k],
3843 / mode_lib->vba.LBBitPerPixel[k]
3844 / (mode_lib->vba.vtaps[k]
3847 mode_lib->vba.VRatio[k],
3851 2.0 * mode_lib->vba.LineBufferSize
3853 mode_lib->vba.HRatio[k]
3856 / mode_lib->vba.LBBitPerPixel[k]
3857 / (mode_lib->vba.VTAPsChroma[k]
3860 mode_lib->vba.VRatio[k]
3866 locals->MaximumSwathWidth[k] = dml_min(
3867 mode_lib->vba.MaximumSwathWidthInDETBuffer,
3868 mode_lib->vba.MaximumSwathWidthInLineBuffer);
3870 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) {
3871 for (j = 0; j < 2; j++) {
3872 mode_lib->vba.MaxDispclkRoundedDownToDFSGranularity = RoundToDFSGranularityDown(
3873 mode_lib->vba.MaxDispclk[i],
3874 mode_lib->vba.DISPCLKDPPCLKVCOSpeed);
3875 mode_lib->vba.MaxDppclkRoundedDownToDFSGranularity = RoundToDFSGranularityDown(
3876 mode_lib->vba.MaxDppclk[i],
3877 mode_lib->vba.DISPCLKDPPCLKVCOSpeed);
3878 locals->RequiredDISPCLK[i][j] = 0.0;
3879 locals->DISPCLK_DPPCLK_Support[i][j] = true;
3880 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
3881 mode_lib->vba.PlaneRequiredDISPCLKWithoutODMCombine =
3882 mode_lib->vba.PixelClock[k] * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0)
3883 * (1.0 + mode_lib->vba.DISPCLKRampingMargin / 100.0);
3884 if (mode_lib->vba.PlaneRequiredDISPCLKWithoutODMCombine >= mode_lib->vba.MaxDispclk[i]
3885 && i == mode_lib->vba.soc.num_states)
3886 mode_lib->vba.PlaneRequiredDISPCLKWithoutODMCombine = mode_lib->vba.PixelClock[k]
3887 * (1 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0);
3889 mode_lib->vba.PlaneRequiredDISPCLKWithODMCombine = mode_lib->vba.PixelClock[k] / 2
3890 * (1 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) * (1 + mode_lib->vba.DISPCLKRampingMargin / 100.0);
3891 if (mode_lib->vba.PlaneRequiredDISPCLKWithODMCombine >= mode_lib->vba.MaxDispclk[i]
3892 && i == mode_lib->vba.soc.num_states)
3893 mode_lib->vba.PlaneRequiredDISPCLKWithODMCombine = mode_lib->vba.PixelClock[k] / 2
3894 * (1 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0);
3895 if (mode_lib->vba.ODMCapability == false || mode_lib->vba.PlaneRequiredDISPCLKWithoutODMCombine <= mode_lib->vba.MaxDispclkRoundedDownToDFSGranularity) {
3896 locals->ODMCombineEnablePerState[i][k] = false;
3897 mode_lib->vba.PlaneRequiredDISPCLK = mode_lib->vba.PlaneRequiredDISPCLKWithoutODMCombine;
3899 locals->ODMCombineEnablePerState[i][k] = true;
3900 mode_lib->vba.PlaneRequiredDISPCLK = mode_lib->vba.PlaneRequiredDISPCLKWithODMCombine;
3902 if (locals->MinDPPCLKUsingSingleDPP[k] * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) <= mode_lib->vba.MaxDppclkRoundedDownToDFSGranularity
3903 && locals->SwathWidthYSingleDPP[k] <= locals->MaximumSwathWidth[k]
3904 && locals->ODMCombineEnablePerState[i][k] == false) {
3905 locals->NoOfDPP[i][j][k] = 1;
3906 locals->RequiredDPPCLK[i][j][k] =
3907 locals->MinDPPCLKUsingSingleDPP[k] * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0);
3909 locals->NoOfDPP[i][j][k] = 2;
3910 locals->RequiredDPPCLK[i][j][k] =
3911 locals->MinDPPCLKUsingSingleDPP[k] * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) / 2.0;
3913 locals->RequiredDISPCLK[i][j] = dml_max(
3914 locals->RequiredDISPCLK[i][j],
3915 mode_lib->vba.PlaneRequiredDISPCLK);
3916 if ((locals->MinDPPCLKUsingSingleDPP[k] / locals->NoOfDPP[i][j][k] * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0)
3917 > mode_lib->vba.MaxDppclkRoundedDownToDFSGranularity)
3918 || (mode_lib->vba.PlaneRequiredDISPCLK > mode_lib->vba.MaxDispclkRoundedDownToDFSGranularity)) {
3919 locals->DISPCLK_DPPCLK_Support[i][j] = false;
3922 locals->TotalNumberOfActiveDPP[i][j] = 0.0;
3923 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++)
3924 locals->TotalNumberOfActiveDPP[i][j] = locals->TotalNumberOfActiveDPP[i][j] + locals->NoOfDPP[i][j][k];
3926 while (locals->TotalNumberOfActiveDPP[i][j] < mode_lib->vba.MaxNumDPP
3927 && locals->TotalNumberOfActiveDPP[i][j] < 2 * mode_lib->vba.NumberOfActivePlanes) {
3928 double BWOfNonSplitPlaneOfMaximumBandwidth;
3929 unsigned int NumberOfNonSplitPlaneOfMaximumBandwidth;
3931 BWOfNonSplitPlaneOfMaximumBandwidth = 0;
3932 NumberOfNonSplitPlaneOfMaximumBandwidth = 0;
3933 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) {
3934 if (locals->ReadBandwidth[k] > BWOfNonSplitPlaneOfMaximumBandwidth && locals->NoOfDPP[i][j][k] == 1) {
3935 BWOfNonSplitPlaneOfMaximumBandwidth = locals->ReadBandwidth[k];
3936 NumberOfNonSplitPlaneOfMaximumBandwidth = k;
3939 locals->NoOfDPP[i][j][NumberOfNonSplitPlaneOfMaximumBandwidth] = 2;
3940 locals->RequiredDPPCLK[i][j][NumberOfNonSplitPlaneOfMaximumBandwidth] =
3941 locals->MinDPPCLKUsingSingleDPP[NumberOfNonSplitPlaneOfMaximumBandwidth]
3942 * (1 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100) / 2;
3943 locals->TotalNumberOfActiveDPP[i][j] = locals->TotalNumberOfActiveDPP[i][j] + 1;
3946 if (locals->TotalNumberOfActiveDPP[i][j] > mode_lib->vba.MaxNumDPP) {
3947 locals->RequiredDISPCLK[i][j] = 0.0;
3948 locals->DISPCLK_DPPCLK_Support[i][j] = true;
3949 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
3950 locals->ODMCombineEnablePerState[i][k] = false;
3951 if (locals->SwathWidthYSingleDPP[k] <= locals->MaximumSwathWidth[k]) {
3952 locals->NoOfDPP[i][j][k] = 1;
3953 locals->RequiredDPPCLK[i][j][k] = locals->MinDPPCLKUsingSingleDPP[k]
3954 * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0);
3956 locals->NoOfDPP[i][j][k] = 2;
3957 locals->RequiredDPPCLK[i][j][k] = locals->MinDPPCLKUsingSingleDPP[k]
3958 * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) / 2.0;
3960 if (i != mode_lib->vba.soc.num_states) {
3961 mode_lib->vba.PlaneRequiredDISPCLK =
3962 mode_lib->vba.PixelClock[k]
3963 * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0)
3964 * (1.0 + mode_lib->vba.DISPCLKRampingMargin / 100.0);
3966 mode_lib->vba.PlaneRequiredDISPCLK = mode_lib->vba.PixelClock[k]
3967 * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0);
3969 locals->RequiredDISPCLK[i][j] = dml_max(
3970 locals->RequiredDISPCLK[i][j],
3971 mode_lib->vba.PlaneRequiredDISPCLK);
3972 if (locals->MinDPPCLKUsingSingleDPP[k] / locals->NoOfDPP[i][j][k] * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0)
3973 > mode_lib->vba.MaxDppclkRoundedDownToDFSGranularity
3974 || mode_lib->vba.PlaneRequiredDISPCLK > mode_lib->vba.MaxDispclkRoundedDownToDFSGranularity)
3975 locals->DISPCLK_DPPCLK_Support[i][j] = false;
3977 locals->TotalNumberOfActiveDPP[i][j] = 0.0;
3978 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++)
3979 locals->TotalNumberOfActiveDPP[i][j] = locals->TotalNumberOfActiveDPP[i][j] + locals->NoOfDPP[i][j][k];
3981 locals->RequiredDISPCLK[i][j] = dml_max(
3982 locals->RequiredDISPCLK[i][j],
3983 mode_lib->vba.WritebackRequiredDISPCLK);
3984 if (mode_lib->vba.MaxDispclkRoundedDownToDFSGranularity
3985 < mode_lib->vba.WritebackRequiredDISPCLK) {
3986 locals->DISPCLK_DPPCLK_Support[i][j] = false;
3990 /*Viewport Size Check*/
3992 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) {
3993 locals->ViewportSizeSupport[i] = true;
3994 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
3995 if (locals->ODMCombineEnablePerState[i][k] == true) {
3996 if (dml_min(locals->SwathWidthYSingleDPP[k], dml_round(mode_lib->vba.HActive[k] / 2.0 * mode_lib->vba.HRatio[k]))
3997 > locals->MaximumSwathWidth[k]) {
3998 locals->ViewportSizeSupport[i] = false;
4001 if (locals->SwathWidthYSingleDPP[k] / 2.0 > locals->MaximumSwathWidth[k]) {
4002 locals->ViewportSizeSupport[i] = false;
4007 /*Total Available Pipes Support Check*/
4009 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) {
4010 for (j = 0; j < 2; j++) {
4011 if (locals->TotalNumberOfActiveDPP[i][j] <= mode_lib->vba.MaxNumDPP)
4012 locals->TotalAvailablePipesSupport[i][j] = true;
4014 locals->TotalAvailablePipesSupport[i][j] = false;
4017 /*Total Available OTG Support Check*/
4019 mode_lib->vba.TotalNumberOfActiveOTG = 0.0;
4020 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4021 if (mode_lib->vba.BlendingAndTiming[k] == k) {
4022 mode_lib->vba.TotalNumberOfActiveOTG = mode_lib->vba.TotalNumberOfActiveOTG
4026 if (mode_lib->vba.TotalNumberOfActiveOTG <= mode_lib->vba.MaxNumOTG) {
4027 mode_lib->vba.NumberOfOTGSupport = true;
4029 mode_lib->vba.NumberOfOTGSupport = false;
4031 /*Display IO and DSC Support Check*/
4033 mode_lib->vba.NonsupportedDSCInputBPC = false;
4034 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4035 if (!(mode_lib->vba.DSCInputBitPerComponent[k] == 12.0
4036 || mode_lib->vba.DSCInputBitPerComponent[k] == 10.0
4037 || mode_lib->vba.DSCInputBitPerComponent[k] == 8.0)) {
4038 mode_lib->vba.NonsupportedDSCInputBPC = true;
4041 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) {
4042 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4043 locals->RequiresDSC[i][k] = 0;
4044 locals->RequiresFEC[i][k] = 0;
4045 if (mode_lib->vba.BlendingAndTiming[k] == k) {
4046 if (mode_lib->vba.Output[k] == dm_hdmi) {
4047 locals->RequiresDSC[i][k] = 0;
4048 locals->RequiresFEC[i][k] = 0;
4049 locals->OutputBppPerState[i][k] = TruncToValidBPP(
4050 dml_min(600.0, mode_lib->vba.PHYCLKPerState[i]) / mode_lib->vba.PixelClockBackEnd[k] * 24,
4052 mode_lib->vba.Output[k],
4053 mode_lib->vba.OutputFormat[k],
4054 mode_lib->vba.DSCInputBitPerComponent[k]);
4055 } else if (mode_lib->vba.Output[k] == dm_dp
4056 || mode_lib->vba.Output[k] == dm_edp) {
4057 if (mode_lib->vba.Output[k] == dm_edp) {
4058 mode_lib->vba.EffectiveFECOverhead = 0.0;
4060 mode_lib->vba.EffectiveFECOverhead =
4061 mode_lib->vba.FECOverhead;
4063 if (mode_lib->vba.PHYCLKPerState[i] >= 270.0) {
4064 mode_lib->vba.Outbpp = TruncToValidBPP(
4065 (1.0 - mode_lib->vba.Downspreading / 100.0) * 270.0
4066 * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0,
4068 mode_lib->vba.Output[k],
4069 mode_lib->vba.OutputFormat[k],
4070 mode_lib->vba.DSCInputBitPerComponent[k]);
4071 mode_lib->vba.OutbppDSC = TruncToValidBPP(
4072 (1.0 - mode_lib->vba.Downspreading / 100.0) * (1.0 - mode_lib->vba.EffectiveFECOverhead / 100.0) * 270.0
4073 * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0,
4075 mode_lib->vba.Output[k],
4076 mode_lib->vba.OutputFormat[k],
4077 mode_lib->vba.DSCInputBitPerComponent[k]);
4078 if (mode_lib->vba.DSCEnabled[k] == true) {
4079 locals->RequiresDSC[i][k] = true;
4080 if (mode_lib->vba.Output[k] == dm_dp) {
4081 locals->RequiresFEC[i][k] = true;
4083 locals->RequiresFEC[i][k] = false;
4085 mode_lib->vba.Outbpp = mode_lib->vba.OutbppDSC;
4087 locals->RequiresDSC[i][k] = false;
4088 locals->RequiresFEC[i][k] = false;
4090 locals->OutputBppPerState[i][k] = mode_lib->vba.Outbpp;
4092 if (mode_lib->vba.Outbpp == BPP_INVALID && mode_lib->vba.PHYCLKPerState[i] >= 540.0) {
4093 mode_lib->vba.Outbpp = TruncToValidBPP(
4094 (1.0 - mode_lib->vba.Downspreading / 100.0) * 540.0
4095 * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0,
4097 mode_lib->vba.Output[k],
4098 mode_lib->vba.OutputFormat[k],
4099 mode_lib->vba.DSCInputBitPerComponent[k]);
4100 mode_lib->vba.OutbppDSC = TruncToValidBPP(
4101 (1.0 - mode_lib->vba.Downspreading / 100.0) * (1.0 - mode_lib->vba.EffectiveFECOverhead / 100.0) * 540.0
4102 * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0,
4104 mode_lib->vba.Output[k],
4105 mode_lib->vba.OutputFormat[k],
4106 mode_lib->vba.DSCInputBitPerComponent[k]);
4107 if (mode_lib->vba.DSCEnabled[k] == true) {
4108 locals->RequiresDSC[i][k] = true;
4109 if (mode_lib->vba.Output[k] == dm_dp) {
4110 locals->RequiresFEC[i][k] = true;
4112 locals->RequiresFEC[i][k] = false;
4114 mode_lib->vba.Outbpp = mode_lib->vba.OutbppDSC;
4116 locals->RequiresDSC[i][k] = false;
4117 locals->RequiresFEC[i][k] = false;
4119 locals->OutputBppPerState[i][k] = mode_lib->vba.Outbpp;
4121 if (mode_lib->vba.Outbpp == BPP_INVALID
4122 && mode_lib->vba.PHYCLKPerState[i]
4124 mode_lib->vba.Outbpp = TruncToValidBPP(
4125 (1.0 - mode_lib->vba.Downspreading / 100.0) * 810.0
4126 * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0,
4128 mode_lib->vba.Output[k],
4129 mode_lib->vba.OutputFormat[k],
4130 mode_lib->vba.DSCInputBitPerComponent[k]);
4131 mode_lib->vba.OutbppDSC = TruncToValidBPP(
4132 (1.0 - mode_lib->vba.Downspreading / 100.0) * (1.0 - mode_lib->vba.EffectiveFECOverhead / 100.0) * 810.0
4133 * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0,
4135 mode_lib->vba.Output[k],
4136 mode_lib->vba.OutputFormat[k],
4137 mode_lib->vba.DSCInputBitPerComponent[k]);
4138 if (mode_lib->vba.DSCEnabled[k] == true || mode_lib->vba.Outbpp == BPP_INVALID) {
4139 locals->RequiresDSC[i][k] = true;
4140 if (mode_lib->vba.Output[k] == dm_dp) {
4141 locals->RequiresFEC[i][k] = true;
4143 locals->RequiresFEC[i][k] = false;
4145 mode_lib->vba.Outbpp = mode_lib->vba.OutbppDSC;
4147 locals->RequiresDSC[i][k] = false;
4148 locals->RequiresFEC[i][k] = false;
4150 locals->OutputBppPerState[i][k] =
4151 mode_lib->vba.Outbpp;
4155 locals->OutputBppPerState[i][k] = BPP_BLENDED_PIPE;
4159 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) {
4160 locals->DIOSupport[i] = true;
4161 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4162 if (locals->OutputBppPerState[i][k] == BPP_INVALID
4163 || (mode_lib->vba.OutputFormat[k] == dm_420
4164 && mode_lib->vba.Interlace[k] == true
4165 && mode_lib->vba.ProgressiveToInterlaceUnitInOPP == true)) {
4166 locals->DIOSupport[i] = false;
4170 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) {
4171 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4172 locals->DSCCLKRequiredMoreThanSupported[i] = false;
4173 if (mode_lib->vba.BlendingAndTiming[k] == k) {
4174 if ((mode_lib->vba.Output[k] == dm_dp
4175 || mode_lib->vba.Output[k] == dm_edp)) {
4176 if (mode_lib->vba.OutputFormat[k] == dm_420
4177 || mode_lib->vba.OutputFormat[k]
4179 mode_lib->vba.DSCFormatFactor = 2;
4181 mode_lib->vba.DSCFormatFactor = 1;
4183 if (locals->RequiresDSC[i][k] == true) {
4184 if (locals->ODMCombineEnablePerState[i][k]
4186 if (mode_lib->vba.PixelClockBackEnd[k] / 6.0 / mode_lib->vba.DSCFormatFactor
4187 > (1.0 - mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) * mode_lib->vba.MaxDSCCLK[i]) {
4188 locals->DSCCLKRequiredMoreThanSupported[i] =
4192 if (mode_lib->vba.PixelClockBackEnd[k] / 3.0 / mode_lib->vba.DSCFormatFactor
4193 > (1.0 - mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) * mode_lib->vba.MaxDSCCLK[i]) {
4194 locals->DSCCLKRequiredMoreThanSupported[i] =
4203 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) {
4204 locals->NotEnoughDSCUnits[i] = false;
4205 mode_lib->vba.TotalDSCUnitsRequired = 0.0;
4206 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4207 if (locals->RequiresDSC[i][k] == true) {
4208 if (locals->ODMCombineEnablePerState[i][k] == true) {
4209 mode_lib->vba.TotalDSCUnitsRequired =
4210 mode_lib->vba.TotalDSCUnitsRequired + 2.0;
4212 mode_lib->vba.TotalDSCUnitsRequired =
4213 mode_lib->vba.TotalDSCUnitsRequired + 1.0;
4217 if (mode_lib->vba.TotalDSCUnitsRequired > mode_lib->vba.NumberOfDSC) {
4218 locals->NotEnoughDSCUnits[i] = true;
4221 /*DSC Delay per state*/
4223 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) {
4224 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4225 if (mode_lib->vba.BlendingAndTiming[k] != k) {
4226 mode_lib->vba.slices = 0;
4227 } else if (locals->RequiresDSC[i][k] == 0
4228 || locals->RequiresDSC[i][k] == false) {
4229 mode_lib->vba.slices = 0;
4230 } else if (mode_lib->vba.PixelClockBackEnd[k] > 3200.0) {
4231 mode_lib->vba.slices = dml_ceil(
4232 mode_lib->vba.PixelClockBackEnd[k] / 400.0,
4234 } else if (mode_lib->vba.PixelClockBackEnd[k] > 1360.0) {
4235 mode_lib->vba.slices = 8.0;
4236 } else if (mode_lib->vba.PixelClockBackEnd[k] > 680.0) {
4237 mode_lib->vba.slices = 4.0;
4238 } else if (mode_lib->vba.PixelClockBackEnd[k] > 340.0) {
4239 mode_lib->vba.slices = 2.0;
4241 mode_lib->vba.slices = 1.0;
4243 if (locals->OutputBppPerState[i][k] == BPP_BLENDED_PIPE
4244 || locals->OutputBppPerState[i][k] == BPP_INVALID) {
4245 mode_lib->vba.bpp = 0.0;
4247 mode_lib->vba.bpp = locals->OutputBppPerState[i][k];
4249 if (locals->RequiresDSC[i][k] == true && mode_lib->vba.bpp != 0.0) {
4250 if (locals->ODMCombineEnablePerState[i][k] == false) {
4251 locals->DSCDelayPerState[i][k] =
4253 mode_lib->vba.DSCInputBitPerComponent[k],
4256 mode_lib->vba.HActive[k]
4257 / mode_lib->vba.slices,
4259 mode_lib->vba.slices,
4260 mode_lib->vba.OutputFormat[k])
4262 mode_lib->vba.OutputFormat[k]);
4264 locals->DSCDelayPerState[i][k] =
4265 2.0 * (dscceComputeDelay(
4266 mode_lib->vba.DSCInputBitPerComponent[k],
4268 dml_ceil(mode_lib->vba.HActive[k] / mode_lib->vba.slices, 1.0),
4269 mode_lib->vba.slices / 2,
4270 mode_lib->vba.OutputFormat[k])
4271 + dscComputeDelay(mode_lib->vba.OutputFormat[k]));
4273 locals->DSCDelayPerState[i][k] =
4274 locals->DSCDelayPerState[i][k] * mode_lib->vba.PixelClock[k] / mode_lib->vba.PixelClockBackEnd[k];
4276 locals->DSCDelayPerState[i][k] = 0.0;
4279 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4280 for (m = 0; m <= mode_lib->vba.NumberOfActivePlanes - 1; m++) {
4281 for (j = 0; j <= mode_lib->vba.NumberOfActivePlanes - 1; j++) {
4282 if (mode_lib->vba.BlendingAndTiming[k] == m && locals->RequiresDSC[i][m] == true)
4283 locals->DSCDelayPerState[i][k] = locals->DSCDelayPerState[i][m];
4290 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) {
4291 for (j = 0; j < 2; j++) {
4292 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4293 if (locals->ODMCombineEnablePerState[i][k] == true)
4294 locals->SwathWidthYPerState[i][j][k] = dml_min(locals->SwathWidthYSingleDPP[k], dml_round(locals->HActive[k] / 2 * locals->HRatio[k]));
4296 locals->SwathWidthYPerState[i][j][k] = locals->SwathWidthYSingleDPP[k] / locals->NoOfDPP[i][j][k];
4297 locals->SwathWidthGranularityY = 256 / dml_ceil(locals->BytePerPixelInDETY[k], 1) / locals->MaxSwathHeightY[k];
4298 locals->RoundedUpMaxSwathSizeBytesY = (dml_ceil(locals->SwathWidthYPerState[i][j][k] - 1, locals->SwathWidthGranularityY)
4299 + locals->SwathWidthGranularityY) * locals->BytePerPixelInDETY[k] * locals->MaxSwathHeightY[k];
4300 if (locals->SourcePixelFormat[k] == dm_420_10) {
4301 locals->RoundedUpMaxSwathSizeBytesY = dml_ceil(locals->RoundedUpMaxSwathSizeBytesY, 256) + 256;
4303 if (locals->MaxSwathHeightC[k] > 0) {
4304 locals->SwathWidthGranularityC = 256 / dml_ceil(locals->BytePerPixelInDETC[k], 2) / locals->MaxSwathHeightC[k];
4306 locals->RoundedUpMaxSwathSizeBytesC = (dml_ceil(locals->SwathWidthYPerState[i][j][k] / 2 - 1, locals->SwathWidthGranularityC)
4307 + locals->SwathWidthGranularityC) * locals->BytePerPixelInDETC[k] * locals->MaxSwathHeightC[k];
4309 if (locals->SourcePixelFormat[k] == dm_420_10) {
4310 locals->RoundedUpMaxSwathSizeBytesC = dml_ceil(locals->RoundedUpMaxSwathSizeBytesC, 256) + 256;
4312 locals->RoundedUpMaxSwathSizeBytesC = 0;
4315 if (locals->RoundedUpMaxSwathSizeBytesY + locals->RoundedUpMaxSwathSizeBytesC <= locals->DETBufferSizeInKByte * 1024 / 2) {
4316 locals->SwathHeightYPerState[i][j][k] = locals->MaxSwathHeightY[k];
4317 locals->SwathHeightCPerState[i][j][k] = locals->MaxSwathHeightC[k];
4319 locals->SwathHeightYPerState[i][j][k] = locals->MinSwathHeightY[k];
4320 locals->SwathHeightCPerState[i][j][k] = locals->MinSwathHeightC[k];
4323 if (locals->BytePerPixelInDETC[k] == 0) {
4324 locals->LinesInDETLuma = locals->DETBufferSizeInKByte * 1024 / locals->BytePerPixelInDETY[k] / locals->SwathWidthYPerState[i][j][k];
4325 locals->LinesInDETChroma = 0;
4326 } else if (locals->SwathHeightYPerState[i][j][k] <= locals->SwathHeightCPerState[i][j][k]) {
4327 locals->LinesInDETLuma = locals->DETBufferSizeInKByte * 1024 / 2 / locals->BytePerPixelInDETY[k] /
4328 locals->SwathWidthYPerState[i][j][k];
4329 locals->LinesInDETChroma = locals->DETBufferSizeInKByte * 1024 / 2 / locals->BytePerPixelInDETC[k] / (locals->SwathWidthYPerState[i][j][k] / 2);
4331 locals->LinesInDETLuma = locals->DETBufferSizeInKByte * 1024 * 2 / 3 / locals->BytePerPixelInDETY[k] / locals->SwathWidthYPerState[i][j][k];
4332 locals->LinesInDETChroma = locals->DETBufferSizeInKByte * 1024 / 3 / locals->BytePerPixelInDETY[k] / (locals->SwathWidthYPerState[i][j][k] / 2);
4335 locals->EffectiveLBLatencyHidingSourceLinesLuma = dml_min(locals->MaxLineBufferLines,
4336 dml_floor(locals->LineBufferSize / locals->LBBitPerPixel[k] / (locals->SwathWidthYPerState[i][j][k]
4337 / dml_max(locals->HRatio[k], 1)), 1)) - (locals->vtaps[k] - 1);
4339 locals->EffectiveLBLatencyHidingSourceLinesChroma = dml_min(locals->MaxLineBufferLines,
4340 dml_floor(locals->LineBufferSize / locals->LBBitPerPixel[k]
4341 / (locals->SwathWidthYPerState[i][j][k] / 2
4342 / dml_max(locals->HRatio[k] / 2, 1)), 1)) - (locals->VTAPsChroma[k] - 1);
4344 locals->EffectiveDETLBLinesLuma = dml_floor(locals->LinesInDETLuma + dml_min(
4345 locals->LinesInDETLuma * locals->RequiredDISPCLK[i][j] * locals->BytePerPixelInDETY[k] *
4346 locals->PSCL_FACTOR[k] / locals->ReturnBWPerState[i],
4347 locals->EffectiveLBLatencyHidingSourceLinesLuma),
4348 locals->SwathHeightYPerState[i][j][k]);
4350 locals->EffectiveDETLBLinesChroma = dml_floor(locals->LinesInDETChroma + dml_min(
4351 locals->LinesInDETChroma * locals->RequiredDISPCLK[i][j] * locals->BytePerPixelInDETC[k] *
4352 locals->PSCL_FACTOR_CHROMA[k] / locals->ReturnBWPerState[i],
4353 locals->EffectiveLBLatencyHidingSourceLinesChroma),
4354 locals->SwathHeightCPerState[i][j][k]);
4356 if (locals->BytePerPixelInDETC[k] == 0) {
4357 locals->UrgentLatencySupportUsPerState[i][j][k] = locals->EffectiveDETLBLinesLuma * (locals->HTotal[k] / locals->PixelClock[k])
4358 / locals->VRatio[k] - locals->EffectiveDETLBLinesLuma * locals->SwathWidthYPerState[i][j][k] *
4359 dml_ceil(locals->BytePerPixelInDETY[k], 1) / (locals->ReturnBWPerState[i] / locals->NoOfDPP[i][j][k]);
4361 locals->UrgentLatencySupportUsPerState[i][j][k] = dml_min(
4362 locals->EffectiveDETLBLinesLuma * (locals->HTotal[k] / locals->PixelClock[k])
4363 / locals->VRatio[k] - locals->EffectiveDETLBLinesLuma * locals->SwathWidthYPerState[i][j][k] *
4364 dml_ceil(locals->BytePerPixelInDETY[k], 1) / (locals->ReturnBWPerState[i] / locals->NoOfDPP[i][j][k]),
4365 locals->EffectiveDETLBLinesChroma * (locals->HTotal[k] / locals->PixelClock[k]) / (locals->VRatio[k] / 2) -
4366 locals->EffectiveDETLBLinesChroma * locals->SwathWidthYPerState[i][j][k] / 2 *
4367 dml_ceil(locals->BytePerPixelInDETC[k], 2) / (locals->ReturnBWPerState[i] / locals->NoOfDPP[i][j][k]));
4373 for (i = 0; i <= locals->soc.num_states; i++) {
4374 for (j = 0; j < 2; j++) {
4375 locals->UrgentLatencySupport[i][j] = true;
4376 for (k = 0; k < locals->NumberOfActivePlanes; k++) {
4377 if (locals->UrgentLatencySupportUsPerState[i][j][k] < locals->UrgentLatency)
4378 locals->UrgentLatencySupport[i][j] = false;
4385 for (i = 0; i <= locals->soc.num_states; i++) {
4386 for (j = 0; j < 2; j++) {
4387 locals->TotalNumberOfDCCActiveDPP[i][j] = 0;
4388 for (k = 0; k < locals->NumberOfActivePlanes; k++) {
4389 if (locals->DCCEnable[k] == true) {
4390 locals->TotalNumberOfDCCActiveDPP[i][j] =
4391 locals->TotalNumberOfDCCActiveDPP[i][j] + locals->NoOfDPP[i][j][k];
4397 CalculateMinAndMaxPrefetchMode(locals->AllowDRAMSelfRefreshOrDRAMClockChangeInVblank, &locals->MinPrefetchMode, &locals->MaxPrefetchMode);
4399 for (i = 0; i <= locals->soc.num_states; i++) {
4400 for (j = 0; j < 2; j++) {
4401 for (k = 0; k < locals->NumberOfActivePlanes; k++) {
4402 locals->NoOfDPPThisState[k] = locals->NoOfDPP[i][j][k];
4403 locals->RequiredDPPCLKThisState[k] = locals->RequiredDPPCLK[i][j][k];
4404 locals->SwathHeightYThisState[k] = locals->SwathHeightYPerState[i][j][k];
4405 locals->SwathHeightCThisState[k] = locals->SwathHeightCPerState[i][j][k];
4406 locals->SwathWidthYThisState[k] = locals->SwathWidthYPerState[i][j][k];
4407 mode_lib->vba.ProjectedDCFCLKDeepSleep = dml_max(
4408 mode_lib->vba.ProjectedDCFCLKDeepSleep,
4409 mode_lib->vba.PixelClock[k] / 16.0);
4410 if (mode_lib->vba.BytePerPixelInDETC[k] == 0.0) {
4411 if (mode_lib->vba.VRatio[k] <= 1.0) {
4412 mode_lib->vba.ProjectedDCFCLKDeepSleep =
4414 mode_lib->vba.ProjectedDCFCLKDeepSleep,
4417 mode_lib->vba.BytePerPixelInDETY[k],
4420 * mode_lib->vba.HRatio[k]
4421 * mode_lib->vba.PixelClock[k]
4422 / mode_lib->vba.NoOfDPP[i][j][k]);
4424 mode_lib->vba.ProjectedDCFCLKDeepSleep =
4426 mode_lib->vba.ProjectedDCFCLKDeepSleep,
4429 mode_lib->vba.BytePerPixelInDETY[k],
4432 * mode_lib->vba.PSCL_FACTOR[k]
4433 * mode_lib->vba.RequiredDPPCLK[i][j][k]);
4436 if (mode_lib->vba.VRatio[k] <= 1.0) {
4437 mode_lib->vba.ProjectedDCFCLKDeepSleep =
4439 mode_lib->vba.ProjectedDCFCLKDeepSleep,
4442 mode_lib->vba.BytePerPixelInDETY[k],
4445 * mode_lib->vba.HRatio[k]
4446 * mode_lib->vba.PixelClock[k]
4447 / mode_lib->vba.NoOfDPP[i][j][k]);
4449 mode_lib->vba.ProjectedDCFCLKDeepSleep =
4451 mode_lib->vba.ProjectedDCFCLKDeepSleep,
4454 mode_lib->vba.BytePerPixelInDETY[k],
4457 * mode_lib->vba.PSCL_FACTOR[k]
4458 * mode_lib->vba.RequiredDPPCLK[i][j][k]);
4460 if (mode_lib->vba.VRatio[k] / 2.0 <= 1.0) {
4461 mode_lib->vba.ProjectedDCFCLKDeepSleep =
4463 mode_lib->vba.ProjectedDCFCLKDeepSleep,
4466 mode_lib->vba.BytePerPixelInDETC[k],
4469 * mode_lib->vba.HRatio[k]
4471 * mode_lib->vba.PixelClock[k]
4472 / mode_lib->vba.NoOfDPP[i][j][k]);
4474 mode_lib->vba.ProjectedDCFCLKDeepSleep =
4476 mode_lib->vba.ProjectedDCFCLKDeepSleep,
4479 mode_lib->vba.BytePerPixelInDETC[k],
4482 * mode_lib->vba.PSCL_FACTOR_CHROMA[k]
4483 * mode_lib->vba.RequiredDPPCLK[i][j][k]);
4487 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4488 mode_lib->vba.PDEAndMetaPTEBytesPerFrameY = CalculateVMAndRowBytes(
4490 mode_lib->vba.DCCEnable[k],
4491 mode_lib->vba.Read256BlockHeightY[k],
4492 mode_lib->vba.Read256BlockWidthY[k],
4493 mode_lib->vba.SourcePixelFormat[k],
4494 mode_lib->vba.SurfaceTiling[k],
4495 dml_ceil(mode_lib->vba.BytePerPixelInDETY[k], 1.0),
4496 mode_lib->vba.SourceScan[k],
4497 mode_lib->vba.ViewportWidth[k],
4498 mode_lib->vba.ViewportHeight[k],
4499 mode_lib->vba.SwathWidthYPerState[i][j][k],
4500 mode_lib->vba.GPUVMEnable,
4501 mode_lib->vba.VMMPageSize,
4502 mode_lib->vba.PTEBufferSizeInRequestsLuma,
4503 mode_lib->vba.PDEProcessingBufIn64KBReqs,
4504 mode_lib->vba.PitchY[k],
4505 mode_lib->vba.DCCMetaPitchY[k],
4506 &mode_lib->vba.MacroTileWidthY[k],
4507 &mode_lib->vba.MetaRowBytesY,
4508 &mode_lib->vba.DPTEBytesPerRowY,
4509 &mode_lib->vba.PTEBufferSizeNotExceededY[i][j][k],
4510 &mode_lib->vba.dpte_row_height[k],
4511 &mode_lib->vba.meta_row_height[k]);
4512 mode_lib->vba.PrefetchLinesY[k] = CalculatePrefetchSourceLines(
4514 mode_lib->vba.VRatio[k],
4515 mode_lib->vba.vtaps[k],
4516 mode_lib->vba.Interlace[k],
4517 mode_lib->vba.ProgressiveToInterlaceUnitInOPP,
4518 mode_lib->vba.SwathHeightYPerState[i][j][k],
4519 mode_lib->vba.ViewportYStartY[k],
4520 &mode_lib->vba.PrefillY[k],
4521 &mode_lib->vba.MaxNumSwY[k]);
4522 if ((mode_lib->vba.SourcePixelFormat[k] != dm_444_64
4523 && mode_lib->vba.SourcePixelFormat[k] != dm_444_32
4524 && mode_lib->vba.SourcePixelFormat[k] != dm_444_16
4525 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_16
4526 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_8)) {
4527 mode_lib->vba.PDEAndMetaPTEBytesPerFrameC = CalculateVMAndRowBytes(
4529 mode_lib->vba.DCCEnable[k],
4530 mode_lib->vba.Read256BlockHeightY[k],
4531 mode_lib->vba.Read256BlockWidthY[k],
4532 mode_lib->vba.SourcePixelFormat[k],
4533 mode_lib->vba.SurfaceTiling[k],
4534 dml_ceil(mode_lib->vba.BytePerPixelInDETC[k], 2.0),
4535 mode_lib->vba.SourceScan[k],
4536 mode_lib->vba.ViewportWidth[k] / 2.0,
4537 mode_lib->vba.ViewportHeight[k] / 2.0,
4538 mode_lib->vba.SwathWidthYPerState[i][j][k] / 2.0,
4539 mode_lib->vba.GPUVMEnable,
4540 mode_lib->vba.VMMPageSize,
4541 mode_lib->vba.PTEBufferSizeInRequestsLuma,
4542 mode_lib->vba.PDEProcessingBufIn64KBReqs,
4543 mode_lib->vba.PitchC[k],
4545 &mode_lib->vba.MacroTileWidthC[k],
4546 &mode_lib->vba.MetaRowBytesC,
4547 &mode_lib->vba.DPTEBytesPerRowC,
4548 &mode_lib->vba.PTEBufferSizeNotExceededC[i][j][k],
4549 &mode_lib->vba.dpte_row_height_chroma[k],
4550 &mode_lib->vba.meta_row_height_chroma[k]);
4551 mode_lib->vba.PrefetchLinesC[k] = CalculatePrefetchSourceLines(
4553 mode_lib->vba.VRatio[k] / 2.0,
4554 mode_lib->vba.VTAPsChroma[k],
4555 mode_lib->vba.Interlace[k],
4556 mode_lib->vba.ProgressiveToInterlaceUnitInOPP,
4557 mode_lib->vba.SwathHeightCPerState[i][j][k],
4558 mode_lib->vba.ViewportYStartC[k],
4559 &mode_lib->vba.PrefillC[k],
4560 &mode_lib->vba.MaxNumSwC[k]);
4562 mode_lib->vba.PDEAndMetaPTEBytesPerFrameC = 0.0;
4563 mode_lib->vba.MetaRowBytesC = 0.0;
4564 mode_lib->vba.DPTEBytesPerRowC = 0.0;
4565 locals->PrefetchLinesC[k] = 0.0;
4566 locals->PTEBufferSizeNotExceededC[i][j][k] = true;
4567 locals->PTEBufferSizeInRequestsForLuma = mode_lib->vba.PTEBufferSizeInRequestsLuma + mode_lib->vba.PTEBufferSizeInRequestsChroma;
4569 locals->PDEAndMetaPTEBytesPerFrame[k] =
4570 mode_lib->vba.PDEAndMetaPTEBytesPerFrameY + mode_lib->vba.PDEAndMetaPTEBytesPerFrameC;
4571 locals->MetaRowBytes[k] = mode_lib->vba.MetaRowBytesY + mode_lib->vba.MetaRowBytesC;
4572 locals->DPTEBytesPerRow[k] = mode_lib->vba.DPTEBytesPerRowY + mode_lib->vba.DPTEBytesPerRowC;
4574 CalculateActiveRowBandwidth(
4575 mode_lib->vba.GPUVMEnable,
4576 mode_lib->vba.SourcePixelFormat[k],
4577 mode_lib->vba.VRatio[k],
4578 mode_lib->vba.DCCEnable[k],
4579 mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k],
4580 mode_lib->vba.MetaRowBytesY,
4581 mode_lib->vba.MetaRowBytesC,
4582 mode_lib->vba.meta_row_height[k],
4583 mode_lib->vba.meta_row_height_chroma[k],
4584 mode_lib->vba.DPTEBytesPerRowY,
4585 mode_lib->vba.DPTEBytesPerRowC,
4586 mode_lib->vba.dpte_row_height[k],
4587 mode_lib->vba.dpte_row_height_chroma[k],
4588 &mode_lib->vba.meta_row_bw[k],
4589 &mode_lib->vba.dpte_row_bw[k],
4590 &mode_lib->vba.qual_row_bw[k]);
4592 mode_lib->vba.ExtraLatency =
4593 mode_lib->vba.UrgentRoundTripAndOutOfOrderLatencyPerState[i]
4594 + (mode_lib->vba.TotalNumberOfActiveDPP[i][j]
4595 * mode_lib->vba.PixelChunkSizeInKByte
4596 + mode_lib->vba.TotalNumberOfDCCActiveDPP[i][j]
4597 * mode_lib->vba.MetaChunkSize)
4599 / mode_lib->vba.ReturnBWPerState[i];
4600 if (mode_lib->vba.GPUVMEnable == true) {
4601 mode_lib->vba.ExtraLatency = mode_lib->vba.ExtraLatency
4602 + mode_lib->vba.TotalNumberOfActiveDPP[i][j]
4603 * mode_lib->vba.PTEGroupSize
4604 / mode_lib->vba.ReturnBWPerState[i];
4606 mode_lib->vba.TimeCalc = 24.0 / mode_lib->vba.ProjectedDCFCLKDeepSleep;
4608 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4609 if (mode_lib->vba.BlendingAndTiming[k] == k) {
4610 if (mode_lib->vba.WritebackEnable[k] == true) {
4611 locals->WritebackDelay[i][k] = mode_lib->vba.WritebackLatency
4612 + CalculateWriteBackDelay(
4613 mode_lib->vba.WritebackPixelFormat[k],
4614 mode_lib->vba.WritebackHRatio[k],
4615 mode_lib->vba.WritebackVRatio[k],
4616 mode_lib->vba.WritebackLumaHTaps[k],
4617 mode_lib->vba.WritebackLumaVTaps[k],
4618 mode_lib->vba.WritebackChromaHTaps[k],
4619 mode_lib->vba.WritebackChromaVTaps[k],
4620 mode_lib->vba.WritebackDestinationWidth[k]) / locals->RequiredDISPCLK[i][j];
4622 locals->WritebackDelay[i][k] = 0.0;
4624 for (m = 0; m <= mode_lib->vba.NumberOfActivePlanes - 1; m++) {
4625 if (mode_lib->vba.BlendingAndTiming[m] == k
4626 && mode_lib->vba.WritebackEnable[m]
4628 locals->WritebackDelay[i][k] = dml_max(locals->WritebackDelay[i][k],
4629 mode_lib->vba.WritebackLatency + CalculateWriteBackDelay(
4630 mode_lib->vba.WritebackPixelFormat[m],
4631 mode_lib->vba.WritebackHRatio[m],
4632 mode_lib->vba.WritebackVRatio[m],
4633 mode_lib->vba.WritebackLumaHTaps[m],
4634 mode_lib->vba.WritebackLumaVTaps[m],
4635 mode_lib->vba.WritebackChromaHTaps[m],
4636 mode_lib->vba.WritebackChromaVTaps[m],
4637 mode_lib->vba.WritebackDestinationWidth[m]) / locals->RequiredDISPCLK[i][j]);
4642 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4643 for (m = 0; m <= mode_lib->vba.NumberOfActivePlanes - 1; m++) {
4644 if (mode_lib->vba.BlendingAndTiming[k] == m) {
4645 locals->WritebackDelay[i][k] = locals->WritebackDelay[i][m];
4649 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4650 for (m = 0; m < locals->NumberOfCursors[k]; m++)
4651 locals->cursor_bw[k] = locals->NumberOfCursors[k] * locals->CursorWidth[k][m] * locals->CursorBPP[k][m]
4652 / 8 / (locals->HTotal[k] / locals->PixelClock[k]) * locals->VRatio[k];
4655 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4656 locals->MaximumVStartup[k] = mode_lib->vba.VTotal[k] - mode_lib->vba.VActive[k]
4657 - dml_max(1.0, dml_ceil(locals->WritebackDelay[i][k] / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]), 1.0));
4660 mode_lib->vba.NextPrefetchMode = mode_lib->vba.MinPrefetchMode;
4662 mode_lib->vba.PrefetchMode[i][j] = mode_lib->vba.NextPrefetchMode;
4663 mode_lib->vba.NextPrefetchMode = mode_lib->vba.NextPrefetchMode + 1;
4665 mode_lib->vba.TWait = CalculateTWait(
4666 mode_lib->vba.PrefetchMode[i][j],
4667 mode_lib->vba.DRAMClockChangeLatency,
4668 mode_lib->vba.UrgentLatency,
4669 mode_lib->vba.SREnterPlusExitTime);
4670 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4672 if (mode_lib->vba.XFCEnabled[k] == true) {
4673 mode_lib->vba.XFCRemoteSurfaceFlipDelay =
4674 CalculateRemoteSurfaceFlipDelay(
4676 mode_lib->vba.VRatio[k],
4677 locals->SwathWidthYPerState[i][j][k],
4678 dml_ceil(locals->BytePerPixelInDETY[k], 1.0),
4679 mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k],
4680 mode_lib->vba.XFCTSlvVupdateOffset,
4681 mode_lib->vba.XFCTSlvVupdateWidth,
4682 mode_lib->vba.XFCTSlvVreadyOffset,
4683 mode_lib->vba.XFCXBUFLatencyTolerance,
4684 mode_lib->vba.XFCFillBWOverhead,
4685 mode_lib->vba.XFCSlvChunkSize,
4686 mode_lib->vba.XFCBusTransportTime,
4687 mode_lib->vba.TimeCalc,
4688 mode_lib->vba.TWait,
4689 &mode_lib->vba.SrcActiveDrainRate,
4690 &mode_lib->vba.TInitXFill,
4691 &mode_lib->vba.TslvChk);
4693 mode_lib->vba.XFCRemoteSurfaceFlipDelay = 0.0;
4695 mode_lib->vba.IsErrorResult[i][j][k] =
4696 CalculatePrefetchSchedule(
4698 mode_lib->vba.RequiredDPPCLK[i][j][k],
4699 mode_lib->vba.RequiredDISPCLK[i][j],
4700 mode_lib->vba.PixelClock[k],
4701 mode_lib->vba.ProjectedDCFCLKDeepSleep,
4702 mode_lib->vba.DSCDelayPerState[i][k],
4703 mode_lib->vba.NoOfDPP[i][j][k],
4704 mode_lib->vba.ScalerEnabled[k],
4705 mode_lib->vba.NumberOfCursors[k],
4706 mode_lib->vba.DPPCLKDelaySubtotal,
4707 mode_lib->vba.DPPCLKDelaySCL,
4708 mode_lib->vba.DPPCLKDelaySCLLBOnly,
4709 mode_lib->vba.DPPCLKDelayCNVCFormater,
4710 mode_lib->vba.DPPCLKDelayCNVCCursor,
4711 mode_lib->vba.DISPCLKDelaySubtotal,
4712 mode_lib->vba.SwathWidthYPerState[i][j][k]
4713 / mode_lib->vba.HRatio[k],
4714 mode_lib->vba.OutputFormat[k],
4715 mode_lib->vba.VTotal[k]
4716 - mode_lib->vba.VActive[k],
4717 mode_lib->vba.HTotal[k],
4718 mode_lib->vba.MaxInterDCNTileRepeaters,
4719 mode_lib->vba.MaximumVStartup[k],
4720 mode_lib->vba.GPUVMMaxPageTableLevels,
4721 mode_lib->vba.GPUVMEnable,
4722 mode_lib->vba.DynamicMetadataEnable[k],
4723 mode_lib->vba.DynamicMetadataLinesBeforeActiveRequired[k],
4724 mode_lib->vba.DynamicMetadataTransmittedBytes[k],
4725 mode_lib->vba.DCCEnable[k],
4726 mode_lib->vba.UrgentLatencyPixelDataOnly,
4727 mode_lib->vba.ExtraLatency,
4728 mode_lib->vba.TimeCalc,
4729 mode_lib->vba.PDEAndMetaPTEBytesPerFrame[k],
4730 mode_lib->vba.MetaRowBytes[k],
4731 mode_lib->vba.DPTEBytesPerRow[k],
4732 mode_lib->vba.PrefetchLinesY[k],
4733 mode_lib->vba.SwathWidthYPerState[i][j][k],
4734 mode_lib->vba.BytePerPixelInDETY[k],
4735 mode_lib->vba.PrefillY[k],
4736 mode_lib->vba.MaxNumSwY[k],
4737 mode_lib->vba.PrefetchLinesC[k],
4738 mode_lib->vba.BytePerPixelInDETC[k],
4739 mode_lib->vba.PrefillC[k],
4740 mode_lib->vba.MaxNumSwC[k],
4741 mode_lib->vba.SwathHeightYPerState[i][j][k],
4742 mode_lib->vba.SwathHeightCPerState[i][j][k],
4743 mode_lib->vba.TWait,
4744 mode_lib->vba.XFCEnabled[k],
4745 mode_lib->vba.XFCRemoteSurfaceFlipDelay,
4746 mode_lib->vba.Interlace[k],
4747 mode_lib->vba.ProgressiveToInterlaceUnitInOPP,
4748 mode_lib->vba.DSTXAfterScaler,
4749 mode_lib->vba.DSTYAfterScaler,
4750 &mode_lib->vba.LineTimesForPrefetch[k],
4751 &mode_lib->vba.PrefetchBW[k],
4752 &mode_lib->vba.LinesForMetaPTE[k],
4753 &mode_lib->vba.LinesForMetaAndDPTERow[k],
4754 &mode_lib->vba.VRatioPreY[i][j][k],
4755 &mode_lib->vba.VRatioPreC[i][j][k],
4756 &mode_lib->vba.RequiredPrefetchPixelDataBWLuma[i][j][k],
4757 &mode_lib->vba.VStartupRequiredWhenNotEnoughTimeForDynamicMetadata,
4758 &mode_lib->vba.Tno_bw[k],
4759 &mode_lib->vba.VUpdateOffsetPix[k],
4760 &mode_lib->vba.VUpdateWidthPix[k],
4761 &mode_lib->vba.VReadyOffsetPix[k]);
4763 mode_lib->vba.MaximumReadBandwidthWithoutPrefetch = 0.0;
4764 mode_lib->vba.MaximumReadBandwidthWithPrefetch = 0.0;
4765 locals->prefetch_vm_bw_valid = true;
4766 locals->prefetch_row_bw_valid = true;
4767 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4768 if (locals->PDEAndMetaPTEBytesPerFrame[k] == 0)
4769 locals->prefetch_vm_bw[k] = 0;
4770 else if (locals->LinesForMetaPTE[k] > 0)
4771 locals->prefetch_vm_bw[k] = locals->PDEAndMetaPTEBytesPerFrame[k]
4772 / (locals->LinesForMetaPTE[k] * locals->HTotal[k] / locals->PixelClock[k]);
4774 locals->prefetch_vm_bw[k] = 0;
4775 locals->prefetch_vm_bw_valid = false;
4777 if (locals->MetaRowBytes[k] + locals->DPTEBytesPerRow[k] == 0)
4778 locals->prefetch_row_bw[k] = 0;
4779 else if (locals->LinesForMetaAndDPTERow[k] > 0)
4780 locals->prefetch_row_bw[k] = (locals->MetaRowBytes[k] + locals->DPTEBytesPerRow[k])
4781 / (locals->LinesForMetaAndDPTERow[k] * locals->HTotal[k] / locals->PixelClock[k]);
4783 locals->prefetch_row_bw[k] = 0;
4784 locals->prefetch_row_bw_valid = false;
4787 mode_lib->vba.MaximumReadBandwidthWithoutPrefetch = mode_lib->vba.MaximumReadBandwidthWithPrefetch
4788 + mode_lib->vba.cursor_bw[k] + mode_lib->vba.ReadBandwidth[k] + mode_lib->vba.meta_row_bw[k] + mode_lib->vba.dpte_row_bw[k];
4789 mode_lib->vba.MaximumReadBandwidthWithPrefetch =
4790 mode_lib->vba.MaximumReadBandwidthWithPrefetch
4791 + mode_lib->vba.cursor_bw[k]
4793 mode_lib->vba.prefetch_vm_bw[k],
4794 mode_lib->vba.prefetch_row_bw[k],
4795 dml_max(mode_lib->vba.ReadBandwidth[k],
4796 mode_lib->vba.RequiredPrefetchPixelDataBWLuma[i][j][k])
4797 + mode_lib->vba.meta_row_bw[k] + mode_lib->vba.dpte_row_bw[k]);
4799 locals->BandwidthWithoutPrefetchSupported[i] = true;
4800 if (mode_lib->vba.MaximumReadBandwidthWithoutPrefetch > locals->ReturnBWPerState[i]) {
4801 locals->BandwidthWithoutPrefetchSupported[i] = false;
4804 locals->PrefetchSupported[i][j] = true;
4805 if (mode_lib->vba.MaximumReadBandwidthWithPrefetch > locals->ReturnBWPerState[i]) {
4806 locals->PrefetchSupported[i][j] = false;
4808 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4809 if (locals->LineTimesForPrefetch[k] < 2.0
4810 || locals->LinesForMetaPTE[k] >= 8.0
4811 || locals->LinesForMetaAndDPTERow[k] >= 16.0
4812 || mode_lib->vba.IsErrorResult[i][j][k] == true) {
4813 locals->PrefetchSupported[i][j] = false;
4816 locals->VRatioInPrefetchSupported[i][j] = true;
4817 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4818 if (locals->VRatioPreY[i][j][k] > 4.0
4819 || locals->VRatioPreC[i][j][k] > 4.0
4820 || mode_lib->vba.IsErrorResult[i][j][k] == true) {
4821 locals->VRatioInPrefetchSupported[i][j] = false;
4824 } while ((locals->PrefetchSupported[i][j] != true || locals->VRatioInPrefetchSupported[i][j] != true)
4825 && mode_lib->vba.NextPrefetchMode < mode_lib->vba.MaxPrefetchMode);
4827 if (mode_lib->vba.PrefetchSupported[i][j] == true
4828 && mode_lib->vba.VRatioInPrefetchSupported[i][j] == true) {
4829 mode_lib->vba.BandwidthAvailableForImmediateFlip =
4830 mode_lib->vba.ReturnBWPerState[i];
4831 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4832 mode_lib->vba.BandwidthAvailableForImmediateFlip =
4833 mode_lib->vba.BandwidthAvailableForImmediateFlip
4834 - mode_lib->vba.cursor_bw[k]
4836 mode_lib->vba.ReadBandwidth[k] + mode_lib->vba.qual_row_bw[k],
4837 mode_lib->vba.PrefetchBW[k]);
4839 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4840 mode_lib->vba.ImmediateFlipBytes[k] = 0.0;
4841 if ((mode_lib->vba.SourcePixelFormat[k] != dm_420_8
4842 && mode_lib->vba.SourcePixelFormat[k] != dm_420_10)) {
4843 mode_lib->vba.ImmediateFlipBytes[k] =
4844 mode_lib->vba.PDEAndMetaPTEBytesPerFrame[k]
4845 + mode_lib->vba.MetaRowBytes[k]
4846 + mode_lib->vba.DPTEBytesPerRow[k];
4849 mode_lib->vba.TotImmediateFlipBytes = 0.0;
4850 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4851 if ((mode_lib->vba.SourcePixelFormat[k] != dm_420_8
4852 && mode_lib->vba.SourcePixelFormat[k] != dm_420_10)) {
4853 mode_lib->vba.TotImmediateFlipBytes =
4854 mode_lib->vba.TotImmediateFlipBytes
4855 + mode_lib->vba.ImmediateFlipBytes[k];
4859 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4860 CalculateFlipSchedule(
4862 mode_lib->vba.ExtraLatency,
4863 mode_lib->vba.UrgentLatencyPixelDataOnly,
4864 mode_lib->vba.GPUVMMaxPageTableLevels,
4865 mode_lib->vba.GPUVMEnable,
4866 mode_lib->vba.BandwidthAvailableForImmediateFlip,
4867 mode_lib->vba.TotImmediateFlipBytes,
4868 mode_lib->vba.SourcePixelFormat[k],
4869 mode_lib->vba.ImmediateFlipBytes[k],
4870 mode_lib->vba.HTotal[k]
4871 / mode_lib->vba.PixelClock[k],
4872 mode_lib->vba.VRatio[k],
4873 mode_lib->vba.Tno_bw[k],
4874 mode_lib->vba.PDEAndMetaPTEBytesPerFrame[k],
4875 mode_lib->vba.MetaRowBytes[k],
4876 mode_lib->vba.DPTEBytesPerRow[k],
4877 mode_lib->vba.DCCEnable[k],
4878 mode_lib->vba.dpte_row_height[k],
4879 mode_lib->vba.meta_row_height[k],
4880 mode_lib->vba.qual_row_bw[k],
4881 &mode_lib->vba.DestinationLinesToRequestVMInImmediateFlip[k],
4882 &mode_lib->vba.DestinationLinesToRequestRowInImmediateFlip[k],
4883 &mode_lib->vba.final_flip_bw[k],
4884 &mode_lib->vba.ImmediateFlipSupportedForPipe[k]);
4886 mode_lib->vba.total_dcn_read_bw_with_flip = 0.0;
4887 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4888 mode_lib->vba.total_dcn_read_bw_with_flip =
4889 mode_lib->vba.total_dcn_read_bw_with_flip
4890 + mode_lib->vba.cursor_bw[k]
4892 mode_lib->vba.prefetch_vm_bw[k],
4893 mode_lib->vba.prefetch_row_bw[k],
4894 mode_lib->vba.final_flip_bw[k]
4896 mode_lib->vba.ReadBandwidth[k],
4897 mode_lib->vba.RequiredPrefetchPixelDataBWLuma[i][j][k]));
4899 mode_lib->vba.ImmediateFlipSupportedForState[i][j] = true;
4900 if (mode_lib->vba.total_dcn_read_bw_with_flip
4901 > mode_lib->vba.ReturnBWPerState[i]) {
4902 mode_lib->vba.ImmediateFlipSupportedForState[i][j] = false;
4904 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4905 if (mode_lib->vba.ImmediateFlipSupportedForPipe[k] == false) {
4906 mode_lib->vba.ImmediateFlipSupportedForState[i][j] = false;
4910 mode_lib->vba.ImmediateFlipSupportedForState[i][j] = false;
4915 /*Vertical Active BW support*/
4916 mode_lib->vba.MaxTotalVActiveRDBandwidth = 0;
4917 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; k++)
4918 mode_lib->vba.MaxTotalVActiveRDBandwidth = mode_lib->vba.MaxTotalVActiveRDBandwidth + mode_lib->vba.ReadBandwidth[k];
4919 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) {
4920 mode_lib->vba.MaxTotalVerticalActiveAvailableBandwidth[i] = dml_min(mode_lib->vba.ReturnBusWidth *
4921 mode_lib->vba.DCFCLKPerState[i], mode_lib->vba.FabricAndDRAMBandwidthPerState[i] * 1000) *
4922 mode_lib->vba.MaxAveragePercentOfIdealDRAMBWDisplayCanUseInNormalSystemOperation / 100;
4923 if (mode_lib->vba.MaxTotalVActiveRDBandwidth <= mode_lib->vba.MaxTotalVerticalActiveAvailableBandwidth[i])
4924 mode_lib->vba.TotalVerticalActiveBandwidthSupport[i] = true;
4926 mode_lib->vba.TotalVerticalActiveBandwidthSupport[i] = false;
4929 /*PTE Buffer Size Check*/
4931 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) {
4932 for (j = 0; j < 2; j++) {
4933 locals->PTEBufferSizeNotExceeded[i][j] = true;
4934 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4935 if (locals->PTEBufferSizeNotExceededY[i][j][k] == false
4936 || locals->PTEBufferSizeNotExceededC[i][j][k] == false) {
4937 locals->PTEBufferSizeNotExceeded[i][j] = false;
4942 /*Cursor Support Check*/
4943 mode_lib->vba.CursorSupport = true;
4944 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4945 for (j = 0; j < 2; j++) {
4946 if (mode_lib->vba.CursorWidth[k][j] > 0.0) {
4949 mode_lib->vba.CursorBufferSize
4950 - mode_lib->vba.CursorChunkSize,
4951 mode_lib->vba.CursorChunkSize) * 1024.0
4952 / (mode_lib->vba.CursorWidth[k][j]
4953 * mode_lib->vba.CursorBPP[k][j]
4956 * (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k])
4957 / mode_lib->vba.VRatio[k] < mode_lib->vba.UrgentLatencyPixelDataOnly
4958 || (mode_lib->vba.CursorBPP[k][j] == 64.0
4959 && mode_lib->vba.Cursor64BppSupport == false)) {
4960 mode_lib->vba.CursorSupport = false;
4965 /*Valid Pitch Check*/
4967 mode_lib->vba.PitchSupport = true;
4968 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
4969 locals->AlignedYPitch[k] = dml_ceil(
4970 dml_max(mode_lib->vba.PitchY[k], mode_lib->vba.ViewportWidth[k]),
4971 locals->MacroTileWidthY[k]);
4972 if (locals->AlignedYPitch[k] > mode_lib->vba.PitchY[k]) {
4973 mode_lib->vba.PitchSupport = false;
4975 if (mode_lib->vba.DCCEnable[k] == true) {
4976 locals->AlignedDCCMetaPitch[k] = dml_ceil(
4978 mode_lib->vba.DCCMetaPitchY[k],
4979 mode_lib->vba.ViewportWidth[k]),
4980 64.0 * locals->Read256BlockWidthY[k]);
4982 locals->AlignedDCCMetaPitch[k] = mode_lib->vba.DCCMetaPitchY[k];
4984 if (locals->AlignedDCCMetaPitch[k] > mode_lib->vba.DCCMetaPitchY[k]) {
4985 mode_lib->vba.PitchSupport = false;
4987 if (mode_lib->vba.SourcePixelFormat[k] != dm_444_64
4988 && mode_lib->vba.SourcePixelFormat[k] != dm_444_32
4989 && mode_lib->vba.SourcePixelFormat[k] != dm_444_16
4990 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_16
4991 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_8) {
4992 locals->AlignedCPitch[k] = dml_ceil(
4994 mode_lib->vba.PitchC[k],
4995 mode_lib->vba.ViewportWidth[k] / 2.0),
4996 locals->MacroTileWidthC[k]);
4998 locals->AlignedCPitch[k] = mode_lib->vba.PitchC[k];
5000 if (locals->AlignedCPitch[k] > mode_lib->vba.PitchC[k]) {
5001 mode_lib->vba.PitchSupport = false;
5004 /*Mode Support, Voltage State and SOC Configuration*/
5006 for (i = mode_lib->vba.soc.num_states; i >= 0; i--) {
5007 for (j = 0; j < 2; j++) {
5008 enum dm_validation_status status = DML_VALIDATION_OK;
5010 if (mode_lib->vba.ScaleRatioAndTapsSupport != true) {
5011 status = DML_FAIL_SCALE_RATIO_TAP;
5012 } else if (mode_lib->vba.SourceFormatPixelAndScanSupport != true) {
5013 status = DML_FAIL_SOURCE_PIXEL_FORMAT;
5014 } else if (locals->ViewportSizeSupport[i] != true) {
5015 status = DML_FAIL_VIEWPORT_SIZE;
5016 } else if (locals->DIOSupport[i] != true) {
5017 status = DML_FAIL_DIO_SUPPORT;
5018 } else if (locals->NotEnoughDSCUnits[i] != false) {
5019 status = DML_FAIL_NOT_ENOUGH_DSC;
5020 } else if (locals->DSCCLKRequiredMoreThanSupported[i] != false) {
5021 status = DML_FAIL_DSC_CLK_REQUIRED;
5022 } else if (locals->UrgentLatencySupport[i][j] != true) {
5023 status = DML_FAIL_URGENT_LATENCY;
5024 } else if (locals->ROBSupport[i] != true) {
5025 status = DML_FAIL_REORDERING_BUFFER;
5026 } else if (locals->DISPCLK_DPPCLK_Support[i][j] != true) {
5027 status = DML_FAIL_DISPCLK_DPPCLK;
5028 } else if (locals->TotalAvailablePipesSupport[i][j] != true) {
5029 status = DML_FAIL_TOTAL_AVAILABLE_PIPES;
5030 } else if (mode_lib->vba.NumberOfOTGSupport != true) {
5031 status = DML_FAIL_NUM_OTG;
5032 } else if (mode_lib->vba.WritebackModeSupport != true) {
5033 status = DML_FAIL_WRITEBACK_MODE;
5034 } else if (mode_lib->vba.WritebackLatencySupport != true) {
5035 status = DML_FAIL_WRITEBACK_LATENCY;
5036 } else if (mode_lib->vba.WritebackScaleRatioAndTapsSupport != true) {
5037 status = DML_FAIL_WRITEBACK_SCALE_RATIO_TAP;
5038 } else if (mode_lib->vba.CursorSupport != true) {
5039 status = DML_FAIL_CURSOR_SUPPORT;
5040 } else if (mode_lib->vba.PitchSupport != true) {
5041 status = DML_FAIL_PITCH_SUPPORT;
5042 } else if (locals->PrefetchSupported[i][j] != true) {
5043 status = DML_FAIL_PREFETCH_SUPPORT;
5044 } else if (locals->TotalVerticalActiveBandwidthSupport[i] != true) {
5045 status = DML_FAIL_TOTAL_V_ACTIVE_BW;
5046 } else if (locals->VRatioInPrefetchSupported[i][j] != true) {
5047 status = DML_FAIL_V_RATIO_PREFETCH;
5048 } else if (locals->PTEBufferSizeNotExceeded[i][j] != true) {
5049 status = DML_FAIL_PTE_BUFFER_SIZE;
5050 } else if (mode_lib->vba.NonsupportedDSCInputBPC != false) {
5051 status = DML_FAIL_DSC_INPUT_BPC;
5054 if (status == DML_VALIDATION_OK) {
5055 locals->ModeSupport[i][j] = true;
5057 locals->ModeSupport[i][j] = false;
5059 locals->ValidationStatus[i] = status;
5063 unsigned int MaximumMPCCombine = 0;
5064 mode_lib->vba.VoltageLevel = mode_lib->vba.soc.num_states + 1;
5065 for (i = mode_lib->vba.VoltageOverrideLevel; i <= mode_lib->vba.soc.num_states; i++) {
5066 if (locals->ModeSupport[i][0] == true || locals->ModeSupport[i][1] == true) {
5067 mode_lib->vba.VoltageLevel = i;
5068 if (locals->ModeSupport[i][1] == true && (locals->ModeSupport[i][0] == false
5069 || mode_lib->vba.WhenToDoMPCCombine == dm_mpc_always_when_possible)) {
5070 MaximumMPCCombine = 1;
5072 MaximumMPCCombine = 0;
5077 mode_lib->vba.ImmediateFlipSupport =
5078 locals->ImmediateFlipSupportedForState[mode_lib->vba.VoltageLevel][MaximumMPCCombine];
5079 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
5080 mode_lib->vba.DPPPerPlane[k] = locals->NoOfDPP[mode_lib->vba.VoltageLevel][MaximumMPCCombine][k];
5081 locals->DPPCLK[k] = locals->RequiredDPPCLK[mode_lib->vba.VoltageLevel][MaximumMPCCombine][k];
5083 mode_lib->vba.DISPCLK = locals->RequiredDISPCLK[mode_lib->vba.VoltageLevel][MaximumMPCCombine];
5084 mode_lib->vba.maxMpcComb = MaximumMPCCombine;
5086 mode_lib->vba.DCFCLK = mode_lib->vba.DCFCLKPerState[mode_lib->vba.VoltageLevel];
5087 mode_lib->vba.DRAMSpeed = mode_lib->vba.DRAMSpeedPerState[mode_lib->vba.VoltageLevel];
5088 mode_lib->vba.FabricClock = mode_lib->vba.FabricClockPerState[mode_lib->vba.VoltageLevel];
5089 mode_lib->vba.SOCCLK = mode_lib->vba.SOCCLKPerState[mode_lib->vba.VoltageLevel];
5090 mode_lib->vba.ReturnBW = locals->ReturnBWPerState[mode_lib->vba.VoltageLevel];
5091 mode_lib->vba.FabricAndDRAMBandwidth = locals->FabricAndDRAMBandwidthPerState[mode_lib->vba.VoltageLevel];
5092 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) {
5093 if (mode_lib->vba.BlendingAndTiming[k] == k) {
5094 mode_lib->vba.ODMCombineEnabled[k] =
5095 locals->ODMCombineEnablePerState[mode_lib->vba.VoltageLevel][k];
5097 mode_lib->vba.ODMCombineEnabled[k] = 0;
5099 mode_lib->vba.DSCEnabled[k] =
5100 locals->RequiresDSC[mode_lib->vba.VoltageLevel][k];
5101 mode_lib->vba.OutputBpp[k] =
5102 locals->OutputBppPerState[mode_lib->vba.VoltageLevel][k];