1 /* 2 * Copyright 2018 Advanced Micro Devices, Inc. 3 * 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: 10 * 11 * The above copyright notice and this permission notice shall be included in 12 * all copies or substantial portions of the Software. 13 * 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. 21 * 22 * Authors: AMD 23 * 24 */ 25 26 #include "../display_mode_lib.h" 27 #include "display_mode_vba_20v2.h" 28 #include "../dml_inline_defs.h" 29 30 /* 31 * NOTE: 32 * This file is gcc-parseable HW gospel, coming straight from HW engineers. 33 * 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. 37 */ 38 39 #define BPP_INVALID 0 40 #define BPP_BLENDED_PIPE 0xffffffff 41 #define DCN20_MAX_DSC_IMAGE_WIDTH 5184 42 #define DCN20_MAX_420_IMAGE_WIDTH 4096 43 44 static double adjust_ReturnBW( 45 struct display_mode_lib *mode_lib, 46 double ReturnBW, 47 bool DCCEnabledAnyPlane, 48 double ReturnBandwidthToDCN); 49 static unsigned int dscceComputeDelay( 50 unsigned int bpc, 51 double bpp, 52 unsigned int sliceWidth, 53 unsigned int numSlices, 54 enum output_format_class pixelFormat); 55 static unsigned int dscComputeDelay(enum output_format_class pixelFormat); 56 static bool CalculateDelayAfterScaler( 57 struct display_mode_lib *mode_lib, 58 double ReturnBW, 59 double ReadBandwidthPlaneLuma, 60 double ReadBandwidthPlaneChroma, 61 double TotalDataReadBandwidth, 62 double DisplayPipeLineDeliveryTimeLuma, 63 double DisplayPipeLineDeliveryTimeChroma, 64 double DPPCLK, 65 double DISPCLK, 66 double PixelClock, 67 unsigned int DSCDelay, 68 unsigned int DPPPerPlane, 69 bool ScalerEnabled, 70 unsigned int NumberOfCursors, 71 double DPPCLKDelaySubtotal, 72 double DPPCLKDelaySCL, 73 double DPPCLKDelaySCLLBOnly, 74 double DPPCLKDelayCNVCFormater, 75 double DPPCLKDelayCNVCCursor, 76 double DISPCLKDelaySubtotal, 77 unsigned int ScalerRecoutWidth, 78 enum output_format_class OutputFormat, 79 unsigned int HTotal, 80 unsigned int SwathWidthSingleDPPY, 81 double BytePerPixelDETY, 82 double BytePerPixelDETC, 83 unsigned int SwathHeightY, 84 unsigned int SwathHeightC, 85 bool Interlace, 86 bool ProgressiveToInterlaceUnitInOPP, 87 double *DSTXAfterScaler, 88 double *DSTYAfterScaler 89 ); 90 // Super monster function with some 45 argument 91 static bool CalculatePrefetchSchedule( 92 struct display_mode_lib *mode_lib, 93 double DPPCLK, 94 double DISPCLK, 95 double PixelClock, 96 double DCFCLKDeepSleep, 97 unsigned int DPPPerPlane, 98 unsigned int NumberOfCursors, 99 unsigned int VBlank, 100 unsigned int HTotal, 101 unsigned int MaxInterDCNTileRepeaters, 102 unsigned int VStartup, 103 unsigned int PageTableLevels, 104 bool GPUVMEnable, 105 bool DynamicMetadataEnable, 106 unsigned int DynamicMetadataLinesBeforeActiveRequired, 107 unsigned int DynamicMetadataTransmittedBytes, 108 bool DCCEnable, 109 double UrgentLatencyPixelDataOnly, 110 double UrgentExtraLatency, 111 double TCalc, 112 unsigned int PDEAndMetaPTEBytesFrame, 113 unsigned int MetaRowByte, 114 unsigned int PixelPTEBytesPerRow, 115 double PrefetchSourceLinesY, 116 unsigned int SwathWidthY, 117 double BytePerPixelDETY, 118 double VInitPreFillY, 119 unsigned int MaxNumSwathY, 120 double PrefetchSourceLinesC, 121 double BytePerPixelDETC, 122 double VInitPreFillC, 123 unsigned int MaxNumSwathC, 124 unsigned int SwathHeightY, 125 unsigned int SwathHeightC, 126 double TWait, 127 bool XFCEnabled, 128 double XFCRemoteSurfaceFlipDelay, 129 bool InterlaceEnable, 130 bool ProgressiveToInterlaceUnitInOPP, 131 double DSTXAfterScaler, 132 double DSTYAfterScaler, 133 double *DestinationLinesForPrefetch, 134 double *PrefetchBandwidth, 135 double *DestinationLinesToRequestVMInVBlank, 136 double *DestinationLinesToRequestRowInVBlank, 137 double *VRatioPrefetchY, 138 double *VRatioPrefetchC, 139 double *RequiredPrefetchPixDataBW, 140 double *Tno_bw, 141 unsigned int *VUpdateOffsetPix, 142 double *VUpdateWidthPix, 143 double *VReadyOffsetPix); 144 static double RoundToDFSGranularityUp(double Clock, double VCOSpeed); 145 static double RoundToDFSGranularityDown(double Clock, double VCOSpeed); 146 static double CalculatePrefetchSourceLines( 147 struct display_mode_lib *mode_lib, 148 double VRatio, 149 double vtaps, 150 bool Interlace, 151 bool ProgressiveToInterlaceUnitInOPP, 152 unsigned int SwathHeight, 153 unsigned int ViewportYStart, 154 double *VInitPreFill, 155 unsigned int *MaxNumSwath); 156 static unsigned int CalculateVMAndRowBytes( 157 struct display_mode_lib *mode_lib, 158 bool DCCEnable, 159 unsigned int BlockHeight256Bytes, 160 unsigned int BlockWidth256Bytes, 161 enum source_format_class SourcePixelFormat, 162 unsigned int SurfaceTiling, 163 unsigned int BytePerPixel, 164 enum scan_direction_class ScanDirection, 165 unsigned int ViewportWidth, 166 unsigned int ViewportHeight, 167 unsigned int SwathWidthY, 168 bool GPUVMEnable, 169 unsigned int VMMPageSize, 170 unsigned int PTEBufferSizeInRequestsLuma, 171 unsigned int PDEProcessingBufIn64KBReqs, 172 unsigned int Pitch, 173 unsigned int DCCMetaPitch, 174 unsigned int *MacroTileWidth, 175 unsigned int *MetaRowByte, 176 unsigned int *PixelPTEBytesPerRow, 177 bool *PTEBufferSizeNotExceeded, 178 unsigned int *dpte_row_height, 179 unsigned int *meta_row_height); 180 static double CalculateTWait( 181 unsigned int PrefetchMode, 182 double DRAMClockChangeLatency, 183 double UrgentLatencyPixelDataOnly, 184 double SREnterPlusExitTime); 185 static double CalculateRemoteSurfaceFlipDelay( 186 struct display_mode_lib *mode_lib, 187 double VRatio, 188 double SwathWidth, 189 double Bpp, 190 double LineTime, 191 double XFCTSlvVupdateOffset, 192 double XFCTSlvVupdateWidth, 193 double XFCTSlvVreadyOffset, 194 double XFCXBUFLatencyTolerance, 195 double XFCFillBWOverhead, 196 double XFCSlvChunkSize, 197 double XFCBusTransportTime, 198 double TCalc, 199 double TWait, 200 double *SrcActiveDrainRate, 201 double *TInitXFill, 202 double *TslvChk); 203 static void CalculateActiveRowBandwidth( 204 bool GPUVMEnable, 205 enum source_format_class SourcePixelFormat, 206 double VRatio, 207 bool DCCEnable, 208 double LineTime, 209 unsigned int MetaRowByteLuma, 210 unsigned int MetaRowByteChroma, 211 unsigned int meta_row_height_luma, 212 unsigned int meta_row_height_chroma, 213 unsigned int PixelPTEBytesPerRowLuma, 214 unsigned int PixelPTEBytesPerRowChroma, 215 unsigned int dpte_row_height_luma, 216 unsigned int dpte_row_height_chroma, 217 double *meta_row_bw, 218 double *dpte_row_bw, 219 double *qual_row_bw); 220 static void CalculateFlipSchedule( 221 struct display_mode_lib *mode_lib, 222 double UrgentExtraLatency, 223 double UrgentLatencyPixelDataOnly, 224 unsigned int GPUVMMaxPageTableLevels, 225 bool GPUVMEnable, 226 double BandwidthAvailableForImmediateFlip, 227 unsigned int TotImmediateFlipBytes, 228 enum source_format_class SourcePixelFormat, 229 unsigned int ImmediateFlipBytes, 230 double LineTime, 231 double VRatio, 232 double Tno_bw, 233 double PDEAndMetaPTEBytesFrame, 234 unsigned int MetaRowByte, 235 unsigned int PixelPTEBytesPerRow, 236 bool DCCEnable, 237 unsigned int dpte_row_height, 238 unsigned int meta_row_height, 239 double qual_row_bw, 240 double *DestinationLinesToRequestVMInImmediateFlip, 241 double *DestinationLinesToRequestRowInImmediateFlip, 242 double *final_flip_bw, 243 bool *ImmediateFlipSupportedForPipe); 244 static double CalculateWriteBackDelay( 245 enum source_format_class WritebackPixelFormat, 246 double WritebackHRatio, 247 double WritebackVRatio, 248 unsigned int WritebackLumaHTaps, 249 unsigned int WritebackLumaVTaps, 250 unsigned int WritebackChromaHTaps, 251 unsigned int WritebackChromaVTaps, 252 unsigned int WritebackDestinationWidth); 253 254 static void dml20v2_DisplayPipeConfiguration(struct display_mode_lib *mode_lib); 255 static void dml20v2_DISPCLKDPPCLKDCFCLKDeepSleepPrefetchParametersWatermarksAndPerformanceCalculation( 256 struct display_mode_lib *mode_lib); 257 258 void dml20v2_recalculate(struct display_mode_lib *mode_lib) 259 { 260 ModeSupportAndSystemConfiguration(mode_lib); 261 mode_lib->vba.FabricAndDRAMBandwidth = dml_min( 262 mode_lib->vba.DRAMSpeed * mode_lib->vba.NumberOfChannels * mode_lib->vba.DRAMChannelWidth, 263 mode_lib->vba.FabricClock * mode_lib->vba.FabricDatapathToDCNDataReturn) / 1000.0; 264 PixelClockAdjustmentForProgressiveToInterlaceUnit(mode_lib); 265 dml20v2_DisplayPipeConfiguration(mode_lib); 266 dml20v2_DISPCLKDPPCLKDCFCLKDeepSleepPrefetchParametersWatermarksAndPerformanceCalculation(mode_lib); 267 } 268 269 static double adjust_ReturnBW( 270 struct display_mode_lib *mode_lib, 271 double ReturnBW, 272 bool DCCEnabledAnyPlane, 273 double ReturnBandwidthToDCN) 274 { 275 double CriticalCompression; 276 277 if (DCCEnabledAnyPlane 278 && ReturnBandwidthToDCN 279 > mode_lib->vba.DCFCLK * mode_lib->vba.ReturnBusWidth / 4.0) 280 ReturnBW = 281 dml_min( 282 ReturnBW, 283 ReturnBandwidthToDCN * 4 284 * (1.0 285 - mode_lib->vba.UrgentLatencyPixelDataOnly 286 / ((mode_lib->vba.ROBBufferSizeInKByte 287 - mode_lib->vba.PixelChunkSizeInKByte) 288 * 1024 289 / ReturnBandwidthToDCN 290 - mode_lib->vba.DCFCLK 291 * mode_lib->vba.ReturnBusWidth 292 / 4) 293 + mode_lib->vba.UrgentLatencyPixelDataOnly)); 294 295 CriticalCompression = 2.0 * mode_lib->vba.ReturnBusWidth * mode_lib->vba.DCFCLK 296 * mode_lib->vba.UrgentLatencyPixelDataOnly 297 / (ReturnBandwidthToDCN * mode_lib->vba.UrgentLatencyPixelDataOnly 298 + (mode_lib->vba.ROBBufferSizeInKByte 299 - mode_lib->vba.PixelChunkSizeInKByte) 300 * 1024); 301 302 if (DCCEnabledAnyPlane && CriticalCompression > 1.0 && CriticalCompression < 4.0) 303 ReturnBW = 304 dml_min( 305 ReturnBW, 306 4.0 * ReturnBandwidthToDCN 307 * (mode_lib->vba.ROBBufferSizeInKByte 308 - mode_lib->vba.PixelChunkSizeInKByte) 309 * 1024 310 * mode_lib->vba.ReturnBusWidth 311 * mode_lib->vba.DCFCLK 312 * mode_lib->vba.UrgentLatencyPixelDataOnly 313 / dml_pow( 314 (ReturnBandwidthToDCN 315 * mode_lib->vba.UrgentLatencyPixelDataOnly 316 + (mode_lib->vba.ROBBufferSizeInKByte 317 - mode_lib->vba.PixelChunkSizeInKByte) 318 * 1024), 319 2)); 320 321 return ReturnBW; 322 } 323 324 static unsigned int dscceComputeDelay( 325 unsigned int bpc, 326 double bpp, 327 unsigned int sliceWidth, 328 unsigned int numSlices, 329 enum output_format_class pixelFormat) 330 { 331 // valid bpc = source bits per component in the set of {8, 10, 12} 332 // valid bpp = increments of 1/16 of a bit 333 // min = 6/7/8 in N420/N422/444, respectively 334 // max = such that compression is 1:1 335 //valid sliceWidth = number of pixels per slice line, must be less than or equal to 5184/numSlices (or 4096/numSlices in 420 mode) 336 //valid numSlices = number of slices in the horiziontal direction per DSC engine in the set of {1, 2, 3, 4} 337 //valid pixelFormat = pixel/color format in the set of {:N444_RGB, :S422, :N422, :N420} 338 339 // fixed value 340 unsigned int rcModelSize = 8192; 341 342 // N422/N420 operate at 2 pixels per clock 343 unsigned int pixelsPerClock, lstall, D, initalXmitDelay, w, s, ix, wx, p, l0, a, ax, l, 344 Delay, pixels; 345 346 if (pixelFormat == dm_n422 || pixelFormat == dm_420) 347 pixelsPerClock = 2; 348 // #all other modes operate at 1 pixel per clock 349 else 350 pixelsPerClock = 1; 351 352 //initial transmit delay as per PPS 353 initalXmitDelay = dml_round(rcModelSize / 2.0 / bpp / pixelsPerClock); 354 355 //compute ssm delay 356 if (bpc == 8) 357 D = 81; 358 else if (bpc == 10) 359 D = 89; 360 else 361 D = 113; 362 363 //divide by pixel per cycle to compute slice width as seen by DSC 364 w = sliceWidth / pixelsPerClock; 365 366 //422 mode has an additional cycle of delay 367 if (pixelFormat == dm_s422) 368 s = 1; 369 else 370 s = 0; 371 372 //main calculation for the dscce 373 ix = initalXmitDelay + 45; 374 wx = (w + 2) / 3; 375 p = 3 * wx - w; 376 l0 = ix / w; 377 a = ix + p * l0; 378 ax = (a + 2) / 3 + D + 6 + 1; 379 l = (ax + wx - 1) / wx; 380 if ((ix % w) == 0 && p != 0) 381 lstall = 1; 382 else 383 lstall = 0; 384 Delay = l * wx * (numSlices - 1) + ax + s + lstall + 22; 385 386 //dsc processes 3 pixel containers per cycle and a container can contain 1 or 2 pixels 387 pixels = Delay * 3 * pixelsPerClock; 388 return pixels; 389 } 390 391 static unsigned int dscComputeDelay(enum output_format_class pixelFormat) 392 { 393 unsigned int Delay = 0; 394 395 if (pixelFormat == dm_420) { 396 // sfr 397 Delay = Delay + 2; 398 // dsccif 399 Delay = Delay + 0; 400 // dscc - input deserializer 401 Delay = Delay + 3; 402 // dscc gets pixels every other cycle 403 Delay = Delay + 2; 404 // dscc - input cdc fifo 405 Delay = Delay + 12; 406 // dscc gets pixels every other cycle 407 Delay = Delay + 13; 408 // dscc - cdc uncertainty 409 Delay = Delay + 2; 410 // dscc - output cdc fifo 411 Delay = Delay + 7; 412 // dscc gets pixels every other cycle 413 Delay = Delay + 3; 414 // dscc - cdc uncertainty 415 Delay = Delay + 2; 416 // dscc - output serializer 417 Delay = Delay + 1; 418 // sft 419 Delay = Delay + 1; 420 } else if (pixelFormat == dm_n422) { 421 // sfr 422 Delay = Delay + 2; 423 // dsccif 424 Delay = Delay + 1; 425 // dscc - input deserializer 426 Delay = Delay + 5; 427 // dscc - input cdc fifo 428 Delay = Delay + 25; 429 // dscc - cdc uncertainty 430 Delay = Delay + 2; 431 // dscc - output cdc fifo 432 Delay = Delay + 10; 433 // dscc - cdc uncertainty 434 Delay = Delay + 2; 435 // dscc - output serializer 436 Delay = Delay + 1; 437 // sft 438 Delay = Delay + 1; 439 } else { 440 // sfr 441 Delay = Delay + 2; 442 // dsccif 443 Delay = Delay + 0; 444 // dscc - input deserializer 445 Delay = Delay + 3; 446 // dscc - input cdc fifo 447 Delay = Delay + 12; 448 // dscc - cdc uncertainty 449 Delay = Delay + 2; 450 // dscc - output cdc fifo 451 Delay = Delay + 7; 452 // dscc - output serializer 453 Delay = Delay + 1; 454 // dscc - cdc uncertainty 455 Delay = Delay + 2; 456 // sft 457 Delay = Delay + 1; 458 } 459 460 return Delay; 461 } 462 463 static bool CalculateDelayAfterScaler( 464 struct display_mode_lib *mode_lib, 465 double ReturnBW, 466 double ReadBandwidthPlaneLuma, 467 double ReadBandwidthPlaneChroma, 468 double TotalDataReadBandwidth, 469 double DisplayPipeLineDeliveryTimeLuma, 470 double DisplayPipeLineDeliveryTimeChroma, 471 double DPPCLK, 472 double DISPCLK, 473 double PixelClock, 474 unsigned int DSCDelay, 475 unsigned int DPPPerPlane, 476 bool ScalerEnabled, 477 unsigned int NumberOfCursors, 478 double DPPCLKDelaySubtotal, 479 double DPPCLKDelaySCL, 480 double DPPCLKDelaySCLLBOnly, 481 double DPPCLKDelayCNVCFormater, 482 double DPPCLKDelayCNVCCursor, 483 double DISPCLKDelaySubtotal, 484 unsigned int ScalerRecoutWidth, 485 enum output_format_class OutputFormat, 486 unsigned int HTotal, 487 unsigned int SwathWidthSingleDPPY, 488 double BytePerPixelDETY, 489 double BytePerPixelDETC, 490 unsigned int SwathHeightY, 491 unsigned int SwathHeightC, 492 bool Interlace, 493 bool ProgressiveToInterlaceUnitInOPP, 494 double *DSTXAfterScaler, 495 double *DSTYAfterScaler 496 ) 497 { 498 unsigned int DPPCycles, DISPCLKCycles; 499 double DataFabricLineDeliveryTimeLuma; 500 double DataFabricLineDeliveryTimeChroma; 501 double DSTTotalPixelsAfterScaler; 502 503 DataFabricLineDeliveryTimeLuma = SwathWidthSingleDPPY * SwathHeightY * dml_ceil(BytePerPixelDETY, 1) / (mode_lib->vba.ReturnBW * ReadBandwidthPlaneLuma / TotalDataReadBandwidth); 504 mode_lib->vba.LastPixelOfLineExtraWatermark = dml_max(mode_lib->vba.LastPixelOfLineExtraWatermark, DataFabricLineDeliveryTimeLuma - DisplayPipeLineDeliveryTimeLuma); 505 506 if (BytePerPixelDETC != 0) { 507 DataFabricLineDeliveryTimeChroma = SwathWidthSingleDPPY / 2 * SwathHeightC * dml_ceil(BytePerPixelDETC, 2) / (mode_lib->vba.ReturnBW * ReadBandwidthPlaneChroma / TotalDataReadBandwidth); 508 mode_lib->vba.LastPixelOfLineExtraWatermark = dml_max(mode_lib->vba.LastPixelOfLineExtraWatermark, DataFabricLineDeliveryTimeChroma - DisplayPipeLineDeliveryTimeChroma); 509 } 510 511 if (ScalerEnabled) 512 DPPCycles = DPPCLKDelaySubtotal + DPPCLKDelaySCL; 513 else 514 DPPCycles = DPPCLKDelaySubtotal + DPPCLKDelaySCLLBOnly; 515 516 DPPCycles = DPPCycles + DPPCLKDelayCNVCFormater + NumberOfCursors * DPPCLKDelayCNVCCursor; 517 518 DISPCLKCycles = DISPCLKDelaySubtotal; 519 520 if (DPPCLK == 0.0 || DISPCLK == 0.0) 521 return true; 522 523 *DSTXAfterScaler = DPPCycles * PixelClock / DPPCLK + DISPCLKCycles * PixelClock / DISPCLK 524 + DSCDelay; 525 526 if (DPPPerPlane > 1) 527 *DSTXAfterScaler = *DSTXAfterScaler + ScalerRecoutWidth; 528 529 if (OutputFormat == dm_420 || (Interlace && ProgressiveToInterlaceUnitInOPP)) 530 *DSTYAfterScaler = 1; 531 else 532 *DSTYAfterScaler = 0; 533 534 DSTTotalPixelsAfterScaler = ((double) (*DSTYAfterScaler * HTotal)) + *DSTXAfterScaler; 535 *DSTYAfterScaler = dml_floor(DSTTotalPixelsAfterScaler / HTotal, 1); 536 *DSTXAfterScaler = DSTTotalPixelsAfterScaler - ((double) (*DSTYAfterScaler * HTotal)); 537 538 return true; 539 } 540 541 static bool CalculatePrefetchSchedule( 542 struct display_mode_lib *mode_lib, 543 double DPPCLK, 544 double DISPCLK, 545 double PixelClock, 546 double DCFCLKDeepSleep, 547 unsigned int DPPPerPlane, 548 unsigned int NumberOfCursors, 549 unsigned int VBlank, 550 unsigned int HTotal, 551 unsigned int MaxInterDCNTileRepeaters, 552 unsigned int VStartup, 553 unsigned int PageTableLevels, 554 bool GPUVMEnable, 555 bool DynamicMetadataEnable, 556 unsigned int DynamicMetadataLinesBeforeActiveRequired, 557 unsigned int DynamicMetadataTransmittedBytes, 558 bool DCCEnable, 559 double UrgentLatencyPixelDataOnly, 560 double UrgentExtraLatency, 561 double TCalc, 562 unsigned int PDEAndMetaPTEBytesFrame, 563 unsigned int MetaRowByte, 564 unsigned int PixelPTEBytesPerRow, 565 double PrefetchSourceLinesY, 566 unsigned int SwathWidthY, 567 double BytePerPixelDETY, 568 double VInitPreFillY, 569 unsigned int MaxNumSwathY, 570 double PrefetchSourceLinesC, 571 double BytePerPixelDETC, 572 double VInitPreFillC, 573 unsigned int MaxNumSwathC, 574 unsigned int SwathHeightY, 575 unsigned int SwathHeightC, 576 double TWait, 577 bool XFCEnabled, 578 double XFCRemoteSurfaceFlipDelay, 579 bool InterlaceEnable, 580 bool ProgressiveToInterlaceUnitInOPP, 581 double DSTXAfterScaler, 582 double DSTYAfterScaler, 583 double *DestinationLinesForPrefetch, 584 double *PrefetchBandwidth, 585 double *DestinationLinesToRequestVMInVBlank, 586 double *DestinationLinesToRequestRowInVBlank, 587 double *VRatioPrefetchY, 588 double *VRatioPrefetchC, 589 double *RequiredPrefetchPixDataBW, 590 double *Tno_bw, 591 unsigned int *VUpdateOffsetPix, 592 double *VUpdateWidthPix, 593 double *VReadyOffsetPix) 594 { 595 bool MyError = false; 596 double TotalRepeaterDelayTime; 597 double Tdm, LineTime, Tsetup; 598 double dst_y_prefetch_equ; 599 double Tsw_oto; 600 double prefetch_bw_oto; 601 double Tvm_oto; 602 double Tr0_oto; 603 double Tpre_oto; 604 double dst_y_prefetch_oto; 605 double TimeForFetchingMetaPTE = 0; 606 double TimeForFetchingRowInVBlank = 0; 607 double LinesToRequestPrefetchPixelData = 0; 608 609 *VUpdateOffsetPix = dml_ceil(HTotal / 4.0, 1); 610 TotalRepeaterDelayTime = MaxInterDCNTileRepeaters * (2.0 / DPPCLK + 3.0 / DISPCLK); 611 *VUpdateWidthPix = (14.0 / DCFCLKDeepSleep + 12.0 / DPPCLK + TotalRepeaterDelayTime) 612 * PixelClock; 613 614 *VReadyOffsetPix = dml_max( 615 150.0 / DPPCLK, 616 TotalRepeaterDelayTime + 20.0 / DCFCLKDeepSleep + 10.0 / DPPCLK) 617 * PixelClock; 618 619 Tsetup = (double) (*VUpdateOffsetPix + *VUpdateWidthPix + *VReadyOffsetPix) / PixelClock; 620 621 LineTime = (double) HTotal / PixelClock; 622 623 if (DynamicMetadataEnable) { 624 double Tdmbf, Tdmec, Tdmsks; 625 626 Tdm = dml_max(0.0, UrgentExtraLatency - TCalc); 627 Tdmbf = DynamicMetadataTransmittedBytes / 4.0 / DISPCLK; 628 Tdmec = LineTime; 629 if (DynamicMetadataLinesBeforeActiveRequired == 0) 630 Tdmsks = VBlank * LineTime / 2.0; 631 else 632 Tdmsks = DynamicMetadataLinesBeforeActiveRequired * LineTime; 633 if (InterlaceEnable && !ProgressiveToInterlaceUnitInOPP) 634 Tdmsks = Tdmsks / 2; 635 if (VStartup * LineTime 636 < Tsetup + TWait + UrgentExtraLatency + Tdmbf + Tdmec + Tdmsks) { 637 MyError = true; 638 } 639 } else 640 Tdm = 0; 641 642 if (GPUVMEnable) { 643 if (PageTableLevels == 4) 644 *Tno_bw = UrgentExtraLatency + UrgentLatencyPixelDataOnly; 645 else if (PageTableLevels == 3) 646 *Tno_bw = UrgentExtraLatency; 647 else 648 *Tno_bw = 0; 649 } else if (DCCEnable) 650 *Tno_bw = LineTime; 651 else 652 *Tno_bw = LineTime / 4; 653 654 dst_y_prefetch_equ = VStartup - dml_max(TCalc + TWait, XFCRemoteSurfaceFlipDelay) / LineTime 655 - (Tsetup + Tdm) / LineTime 656 - (DSTYAfterScaler + DSTXAfterScaler / HTotal); 657 658 Tsw_oto = dml_max(PrefetchSourceLinesY, PrefetchSourceLinesC) * LineTime; 659 660 prefetch_bw_oto = (MetaRowByte + PixelPTEBytesPerRow 661 + PrefetchSourceLinesY * SwathWidthY * dml_ceil(BytePerPixelDETY, 1) 662 + PrefetchSourceLinesC * SwathWidthY / 2 * dml_ceil(BytePerPixelDETC, 2)) 663 / Tsw_oto; 664 665 if (GPUVMEnable == true) { 666 Tvm_oto = 667 dml_max( 668 *Tno_bw + PDEAndMetaPTEBytesFrame / prefetch_bw_oto, 669 dml_max( 670 UrgentExtraLatency 671 + UrgentLatencyPixelDataOnly 672 * (PageTableLevels 673 - 1), 674 LineTime / 4.0)); 675 } else 676 Tvm_oto = LineTime / 4.0; 677 678 if ((GPUVMEnable == true || DCCEnable == true)) { 679 Tr0_oto = dml_max( 680 (MetaRowByte + PixelPTEBytesPerRow) / prefetch_bw_oto, 681 dml_max(UrgentLatencyPixelDataOnly, dml_max(LineTime - Tvm_oto, LineTime / 4))); 682 } else 683 Tr0_oto = LineTime - Tvm_oto; 684 685 Tpre_oto = Tvm_oto + Tr0_oto + Tsw_oto; 686 687 dst_y_prefetch_oto = Tpre_oto / LineTime; 688 689 if (dst_y_prefetch_oto < dst_y_prefetch_equ) 690 *DestinationLinesForPrefetch = dst_y_prefetch_oto; 691 else 692 *DestinationLinesForPrefetch = dst_y_prefetch_equ; 693 694 *DestinationLinesForPrefetch = dml_floor(4.0 * (*DestinationLinesForPrefetch + 0.125), 1) 695 / 4; 696 697 dml_print("DML: VStartup: %d\n", VStartup); 698 dml_print("DML: TCalc: %f\n", TCalc); 699 dml_print("DML: TWait: %f\n", TWait); 700 dml_print("DML: XFCRemoteSurfaceFlipDelay: %f\n", XFCRemoteSurfaceFlipDelay); 701 dml_print("DML: LineTime: %f\n", LineTime); 702 dml_print("DML: Tsetup: %f\n", Tsetup); 703 dml_print("DML: Tdm: %f\n", Tdm); 704 dml_print("DML: DSTYAfterScaler: %f\n", DSTYAfterScaler); 705 dml_print("DML: DSTXAfterScaler: %f\n", DSTXAfterScaler); 706 dml_print("DML: HTotal: %d\n", HTotal); 707 708 *PrefetchBandwidth = 0; 709 *DestinationLinesToRequestVMInVBlank = 0; 710 *DestinationLinesToRequestRowInVBlank = 0; 711 *VRatioPrefetchY = 0; 712 *VRatioPrefetchC = 0; 713 *RequiredPrefetchPixDataBW = 0; 714 if (*DestinationLinesForPrefetch > 1) { 715 *PrefetchBandwidth = (PDEAndMetaPTEBytesFrame + 2 * MetaRowByte 716 + 2 * PixelPTEBytesPerRow 717 + PrefetchSourceLinesY * SwathWidthY * dml_ceil(BytePerPixelDETY, 1) 718 + PrefetchSourceLinesC * SwathWidthY / 2 719 * dml_ceil(BytePerPixelDETC, 2)) 720 / (*DestinationLinesForPrefetch * LineTime - *Tno_bw); 721 if (GPUVMEnable) { 722 TimeForFetchingMetaPTE = 723 dml_max( 724 *Tno_bw 725 + (double) PDEAndMetaPTEBytesFrame 726 / *PrefetchBandwidth, 727 dml_max( 728 UrgentExtraLatency 729 + UrgentLatencyPixelDataOnly 730 * (PageTableLevels 731 - 1), 732 LineTime / 4)); 733 } else { 734 if (NumberOfCursors > 0 || XFCEnabled) 735 TimeForFetchingMetaPTE = LineTime / 4; 736 else 737 TimeForFetchingMetaPTE = 0.0; 738 } 739 740 if ((GPUVMEnable == true || DCCEnable == true)) { 741 TimeForFetchingRowInVBlank = 742 dml_max( 743 (MetaRowByte + PixelPTEBytesPerRow) 744 / *PrefetchBandwidth, 745 dml_max( 746 UrgentLatencyPixelDataOnly, 747 dml_max( 748 LineTime 749 - TimeForFetchingMetaPTE, 750 LineTime 751 / 4.0))); 752 } else { 753 if (NumberOfCursors > 0 || XFCEnabled) 754 TimeForFetchingRowInVBlank = LineTime - TimeForFetchingMetaPTE; 755 else 756 TimeForFetchingRowInVBlank = 0.0; 757 } 758 759 *DestinationLinesToRequestVMInVBlank = dml_floor( 760 4.0 * (TimeForFetchingMetaPTE / LineTime + 0.125), 761 1) / 4.0; 762 763 *DestinationLinesToRequestRowInVBlank = dml_floor( 764 4.0 * (TimeForFetchingRowInVBlank / LineTime + 0.125), 765 1) / 4.0; 766 767 LinesToRequestPrefetchPixelData = 768 *DestinationLinesForPrefetch 769 - ((NumberOfCursors > 0 || GPUVMEnable 770 || DCCEnable) ? 771 (*DestinationLinesToRequestVMInVBlank 772 + *DestinationLinesToRequestRowInVBlank) : 773 0.0); 774 775 if (LinesToRequestPrefetchPixelData > 0) { 776 777 *VRatioPrefetchY = (double) PrefetchSourceLinesY 778 / LinesToRequestPrefetchPixelData; 779 *VRatioPrefetchY = dml_max(*VRatioPrefetchY, 1.0); 780 if ((SwathHeightY > 4) && (VInitPreFillY > 3)) { 781 if (LinesToRequestPrefetchPixelData > (VInitPreFillY - 3.0) / 2.0) { 782 *VRatioPrefetchY = 783 dml_max( 784 (double) PrefetchSourceLinesY 785 / LinesToRequestPrefetchPixelData, 786 (double) MaxNumSwathY 787 * SwathHeightY 788 / (LinesToRequestPrefetchPixelData 789 - (VInitPreFillY 790 - 3.0) 791 / 2.0)); 792 *VRatioPrefetchY = dml_max(*VRatioPrefetchY, 1.0); 793 } else { 794 MyError = true; 795 *VRatioPrefetchY = 0; 796 } 797 } 798 799 *VRatioPrefetchC = (double) PrefetchSourceLinesC 800 / LinesToRequestPrefetchPixelData; 801 *VRatioPrefetchC = dml_max(*VRatioPrefetchC, 1.0); 802 803 if ((SwathHeightC > 4)) { 804 if (LinesToRequestPrefetchPixelData > (VInitPreFillC - 3.0) / 2.0) { 805 *VRatioPrefetchC = 806 dml_max( 807 *VRatioPrefetchC, 808 (double) MaxNumSwathC 809 * SwathHeightC 810 / (LinesToRequestPrefetchPixelData 811 - (VInitPreFillC 812 - 3.0) 813 / 2.0)); 814 *VRatioPrefetchC = dml_max(*VRatioPrefetchC, 1.0); 815 } else { 816 MyError = true; 817 *VRatioPrefetchC = 0; 818 } 819 } 820 821 *RequiredPrefetchPixDataBW = 822 DPPPerPlane 823 * ((double) PrefetchSourceLinesY 824 / LinesToRequestPrefetchPixelData 825 * dml_ceil( 826 BytePerPixelDETY, 827 1) 828 + (double) PrefetchSourceLinesC 829 / LinesToRequestPrefetchPixelData 830 * dml_ceil( 831 BytePerPixelDETC, 832 2) 833 / 2) 834 * SwathWidthY / LineTime; 835 } else { 836 MyError = true; 837 *VRatioPrefetchY = 0; 838 *VRatioPrefetchC = 0; 839 *RequiredPrefetchPixDataBW = 0; 840 } 841 842 } else { 843 MyError = true; 844 } 845 846 if (MyError) { 847 *PrefetchBandwidth = 0; 848 TimeForFetchingMetaPTE = 0; 849 TimeForFetchingRowInVBlank = 0; 850 *DestinationLinesToRequestVMInVBlank = 0; 851 *DestinationLinesToRequestRowInVBlank = 0; 852 *DestinationLinesForPrefetch = 0; 853 LinesToRequestPrefetchPixelData = 0; 854 *VRatioPrefetchY = 0; 855 *VRatioPrefetchC = 0; 856 *RequiredPrefetchPixDataBW = 0; 857 } 858 859 return MyError; 860 } 861 862 static double RoundToDFSGranularityUp(double Clock, double VCOSpeed) 863 { 864 return VCOSpeed * 4 / dml_floor(VCOSpeed * 4 / Clock, 1); 865 } 866 867 static double RoundToDFSGranularityDown(double Clock, double VCOSpeed) 868 { 869 return VCOSpeed * 4 / dml_ceil(VCOSpeed * 4 / Clock, 1); 870 } 871 872 static double CalculatePrefetchSourceLines( 873 struct display_mode_lib *mode_lib, 874 double VRatio, 875 double vtaps, 876 bool Interlace, 877 bool ProgressiveToInterlaceUnitInOPP, 878 unsigned int SwathHeight, 879 unsigned int ViewportYStart, 880 double *VInitPreFill, 881 unsigned int *MaxNumSwath) 882 { 883 unsigned int MaxPartialSwath; 884 885 if (ProgressiveToInterlaceUnitInOPP) 886 *VInitPreFill = dml_floor((VRatio + vtaps + 1) / 2.0, 1); 887 else 888 *VInitPreFill = dml_floor((VRatio + vtaps + 1 + Interlace * 0.5 * VRatio) / 2.0, 1); 889 890 if (!mode_lib->vba.IgnoreViewportPositioning) { 891 892 *MaxNumSwath = dml_ceil((*VInitPreFill - 1.0) / SwathHeight, 1) + 1.0; 893 894 if (*VInitPreFill > 1.0) 895 MaxPartialSwath = (unsigned int) (*VInitPreFill - 2) % SwathHeight; 896 else 897 MaxPartialSwath = (unsigned int) (*VInitPreFill + SwathHeight - 2) 898 % SwathHeight; 899 MaxPartialSwath = dml_max(1U, MaxPartialSwath); 900 901 } else { 902 903 if (ViewportYStart != 0) 904 dml_print( 905 "WARNING DML: using viewport y position of 0 even though actual viewport y position is non-zero in prefetch source lines calculation\n"); 906 907 *MaxNumSwath = dml_ceil(*VInitPreFill / SwathHeight, 1); 908 909 if (*VInitPreFill > 1.0) 910 MaxPartialSwath = (unsigned int) (*VInitPreFill - 1) % SwathHeight; 911 else 912 MaxPartialSwath = (unsigned int) (*VInitPreFill + SwathHeight - 1) 913 % SwathHeight; 914 } 915 916 return *MaxNumSwath * SwathHeight + MaxPartialSwath; 917 } 918 919 static unsigned int CalculateVMAndRowBytes( 920 struct display_mode_lib *mode_lib, 921 bool DCCEnable, 922 unsigned int BlockHeight256Bytes, 923 unsigned int BlockWidth256Bytes, 924 enum source_format_class SourcePixelFormat, 925 unsigned int SurfaceTiling, 926 unsigned int BytePerPixel, 927 enum scan_direction_class ScanDirection, 928 unsigned int ViewportWidth, 929 unsigned int ViewportHeight, 930 unsigned int SwathWidth, 931 bool GPUVMEnable, 932 unsigned int VMMPageSize, 933 unsigned int PTEBufferSizeInRequestsLuma, 934 unsigned int PDEProcessingBufIn64KBReqs, 935 unsigned int Pitch, 936 unsigned int DCCMetaPitch, 937 unsigned int *MacroTileWidth, 938 unsigned int *MetaRowByte, 939 unsigned int *PixelPTEBytesPerRow, 940 bool *PTEBufferSizeNotExceeded, 941 unsigned int *dpte_row_height, 942 unsigned int *meta_row_height) 943 { 944 unsigned int MetaRequestHeight; 945 unsigned int MetaRequestWidth; 946 unsigned int MetaSurfWidth; 947 unsigned int MetaSurfHeight; 948 unsigned int MPDEBytesFrame; 949 unsigned int MetaPTEBytesFrame; 950 unsigned int DCCMetaSurfaceBytes; 951 952 unsigned int MacroTileSizeBytes; 953 unsigned int MacroTileHeight; 954 unsigned int DPDE0BytesFrame; 955 unsigned int ExtraDPDEBytesFrame; 956 unsigned int PDEAndMetaPTEBytesFrame; 957 958 if (DCCEnable == true) { 959 MetaRequestHeight = 8 * BlockHeight256Bytes; 960 MetaRequestWidth = 8 * BlockWidth256Bytes; 961 if (ScanDirection == dm_horz) { 962 *meta_row_height = MetaRequestHeight; 963 MetaSurfWidth = dml_ceil((double) SwathWidth - 1, MetaRequestWidth) 964 + MetaRequestWidth; 965 *MetaRowByte = MetaSurfWidth * MetaRequestHeight * BytePerPixel / 256.0; 966 } else { 967 *meta_row_height = MetaRequestWidth; 968 MetaSurfHeight = dml_ceil((double) SwathWidth - 1, MetaRequestHeight) 969 + MetaRequestHeight; 970 *MetaRowByte = MetaSurfHeight * MetaRequestWidth * BytePerPixel / 256.0; 971 } 972 if (ScanDirection == dm_horz) { 973 DCCMetaSurfaceBytes = DCCMetaPitch 974 * (dml_ceil(ViewportHeight - 1, 64 * BlockHeight256Bytes) 975 + 64 * BlockHeight256Bytes) * BytePerPixel 976 / 256; 977 } else { 978 DCCMetaSurfaceBytes = DCCMetaPitch 979 * (dml_ceil( 980 (double) ViewportHeight - 1, 981 64 * BlockHeight256Bytes) 982 + 64 * BlockHeight256Bytes) * BytePerPixel 983 / 256; 984 } 985 if (GPUVMEnable == true) { 986 MetaPTEBytesFrame = (dml_ceil( 987 (double) (DCCMetaSurfaceBytes - VMMPageSize) 988 / (8 * VMMPageSize), 989 1) + 1) * 64; 990 MPDEBytesFrame = 128 * (mode_lib->vba.GPUVMMaxPageTableLevels - 1); 991 } else { 992 MetaPTEBytesFrame = 0; 993 MPDEBytesFrame = 0; 994 } 995 } else { 996 MetaPTEBytesFrame = 0; 997 MPDEBytesFrame = 0; 998 *MetaRowByte = 0; 999 } 1000 1001 if (SurfaceTiling == dm_sw_linear || SurfaceTiling == dm_sw_gfx7_2d_thin_gl || SurfaceTiling == dm_sw_gfx7_2d_thin_l_vp) { 1002 MacroTileSizeBytes = 256; 1003 MacroTileHeight = BlockHeight256Bytes; 1004 } else if (SurfaceTiling == dm_sw_4kb_s || SurfaceTiling == dm_sw_4kb_s_x 1005 || SurfaceTiling == dm_sw_4kb_d || SurfaceTiling == dm_sw_4kb_d_x) { 1006 MacroTileSizeBytes = 4096; 1007 MacroTileHeight = 4 * BlockHeight256Bytes; 1008 } else if (SurfaceTiling == dm_sw_64kb_s || SurfaceTiling == dm_sw_64kb_s_t 1009 || SurfaceTiling == dm_sw_64kb_s_x || SurfaceTiling == dm_sw_64kb_d 1010 || SurfaceTiling == dm_sw_64kb_d_t || SurfaceTiling == dm_sw_64kb_d_x 1011 || SurfaceTiling == dm_sw_64kb_r_x) { 1012 MacroTileSizeBytes = 65536; 1013 MacroTileHeight = 16 * BlockHeight256Bytes; 1014 } else { 1015 MacroTileSizeBytes = 262144; 1016 MacroTileHeight = 32 * BlockHeight256Bytes; 1017 } 1018 *MacroTileWidth = MacroTileSizeBytes / BytePerPixel / MacroTileHeight; 1019 1020 if (GPUVMEnable == true && mode_lib->vba.GPUVMMaxPageTableLevels > 1) { 1021 if (ScanDirection == dm_horz) { 1022 DPDE0BytesFrame = 1023 64 1024 * (dml_ceil( 1025 ((Pitch 1026 * (dml_ceil( 1027 ViewportHeight 1028 - 1, 1029 MacroTileHeight) 1030 + MacroTileHeight) 1031 * BytePerPixel) 1032 - MacroTileSizeBytes) 1033 / (8 1034 * 2097152), 1035 1) + 1); 1036 } else { 1037 DPDE0BytesFrame = 1038 64 1039 * (dml_ceil( 1040 ((Pitch 1041 * (dml_ceil( 1042 (double) SwathWidth 1043 - 1, 1044 MacroTileHeight) 1045 + MacroTileHeight) 1046 * BytePerPixel) 1047 - MacroTileSizeBytes) 1048 / (8 1049 * 2097152), 1050 1) + 1); 1051 } 1052 ExtraDPDEBytesFrame = 128 * (mode_lib->vba.GPUVMMaxPageTableLevels - 2); 1053 } else { 1054 DPDE0BytesFrame = 0; 1055 ExtraDPDEBytesFrame = 0; 1056 } 1057 1058 PDEAndMetaPTEBytesFrame = MetaPTEBytesFrame + MPDEBytesFrame + DPDE0BytesFrame 1059 + ExtraDPDEBytesFrame; 1060 1061 if (GPUVMEnable == true) { 1062 unsigned int PTERequestSize; 1063 unsigned int PixelPTEReqHeight; 1064 unsigned int PixelPTEReqWidth; 1065 double FractionOfPTEReturnDrop; 1066 unsigned int EffectivePDEProcessingBufIn64KBReqs; 1067 1068 if (SurfaceTiling == dm_sw_linear) { 1069 PixelPTEReqHeight = 1; 1070 PixelPTEReqWidth = 8.0 * VMMPageSize / BytePerPixel; 1071 PTERequestSize = 64; 1072 FractionOfPTEReturnDrop = 0; 1073 } else if (MacroTileSizeBytes == 4096) { 1074 PixelPTEReqHeight = MacroTileHeight; 1075 PixelPTEReqWidth = 8 * *MacroTileWidth; 1076 PTERequestSize = 64; 1077 if (ScanDirection == dm_horz) 1078 FractionOfPTEReturnDrop = 0; 1079 else 1080 FractionOfPTEReturnDrop = 7 / 8; 1081 } else if (VMMPageSize == 4096 && MacroTileSizeBytes > 4096) { 1082 PixelPTEReqHeight = 16 * BlockHeight256Bytes; 1083 PixelPTEReqWidth = 16 * BlockWidth256Bytes; 1084 PTERequestSize = 128; 1085 FractionOfPTEReturnDrop = 0; 1086 } else { 1087 PixelPTEReqHeight = MacroTileHeight; 1088 PixelPTEReqWidth = 8 * *MacroTileWidth; 1089 PTERequestSize = 64; 1090 FractionOfPTEReturnDrop = 0; 1091 } 1092 1093 if (SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10) 1094 EffectivePDEProcessingBufIn64KBReqs = PDEProcessingBufIn64KBReqs / 2; 1095 else 1096 EffectivePDEProcessingBufIn64KBReqs = PDEProcessingBufIn64KBReqs; 1097 1098 if (SurfaceTiling == dm_sw_linear) { 1099 *dpte_row_height = 1100 dml_min( 1101 128, 1102 1 1103 << (unsigned int) dml_floor( 1104 dml_log2( 1105 dml_min( 1106 (double) PTEBufferSizeInRequestsLuma 1107 * PixelPTEReqWidth, 1108 EffectivePDEProcessingBufIn64KBReqs 1109 * 65536.0 1110 / BytePerPixel) 1111 / Pitch), 1112 1)); 1113 *PixelPTEBytesPerRow = PTERequestSize 1114 * (dml_ceil( 1115 (double) (Pitch * *dpte_row_height - 1) 1116 / PixelPTEReqWidth, 1117 1) + 1); 1118 } else if (ScanDirection == dm_horz) { 1119 *dpte_row_height = PixelPTEReqHeight; 1120 *PixelPTEBytesPerRow = PTERequestSize 1121 * (dml_ceil(((double) SwathWidth - 1) / PixelPTEReqWidth, 1) 1122 + 1); 1123 } else { 1124 *dpte_row_height = dml_min(PixelPTEReqWidth, *MacroTileWidth); 1125 *PixelPTEBytesPerRow = PTERequestSize 1126 * (dml_ceil( 1127 ((double) SwathWidth - 1) 1128 / PixelPTEReqHeight, 1129 1) + 1); 1130 } 1131 if (*PixelPTEBytesPerRow * (1 - FractionOfPTEReturnDrop) 1132 <= 64 * PTEBufferSizeInRequestsLuma) { 1133 *PTEBufferSizeNotExceeded = true; 1134 } else { 1135 *PTEBufferSizeNotExceeded = false; 1136 } 1137 } else { 1138 *PixelPTEBytesPerRow = 0; 1139 *PTEBufferSizeNotExceeded = true; 1140 } 1141 1142 return PDEAndMetaPTEBytesFrame; 1143 } 1144 1145 static void dml20v2_DISPCLKDPPCLKDCFCLKDeepSleepPrefetchParametersWatermarksAndPerformanceCalculation( 1146 struct display_mode_lib *mode_lib) 1147 { 1148 unsigned int j, k; 1149 1150 mode_lib->vba.WritebackDISPCLK = 0.0; 1151 mode_lib->vba.DISPCLKWithRamping = 0; 1152 mode_lib->vba.DISPCLKWithoutRamping = 0; 1153 mode_lib->vba.GlobalDPPCLK = 0.0; 1154 1155 // dml_ml->vba.DISPCLK and dml_ml->vba.DPPCLK Calculation 1156 // 1157 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 1158 if (mode_lib->vba.WritebackEnable[k]) { 1159 mode_lib->vba.WritebackDISPCLK = 1160 dml_max( 1161 mode_lib->vba.WritebackDISPCLK, 1162 CalculateWriteBackDISPCLK( 1163 mode_lib->vba.WritebackPixelFormat[k], 1164 mode_lib->vba.PixelClock[k], 1165 mode_lib->vba.WritebackHRatio[k], 1166 mode_lib->vba.WritebackVRatio[k], 1167 mode_lib->vba.WritebackLumaHTaps[k], 1168 mode_lib->vba.WritebackLumaVTaps[k], 1169 mode_lib->vba.WritebackChromaHTaps[k], 1170 mode_lib->vba.WritebackChromaVTaps[k], 1171 mode_lib->vba.WritebackDestinationWidth[k], 1172 mode_lib->vba.HTotal[k], 1173 mode_lib->vba.WritebackChromaLineBufferWidth)); 1174 } 1175 } 1176 1177 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 1178 if (mode_lib->vba.HRatio[k] > 1) { 1179 mode_lib->vba.PSCL_THROUGHPUT_LUMA[k] = dml_min( 1180 mode_lib->vba.MaxDCHUBToPSCLThroughput, 1181 mode_lib->vba.MaxPSCLToLBThroughput 1182 * mode_lib->vba.HRatio[k] 1183 / dml_ceil( 1184 mode_lib->vba.htaps[k] 1185 / 6.0, 1186 1)); 1187 } else { 1188 mode_lib->vba.PSCL_THROUGHPUT_LUMA[k] = dml_min( 1189 mode_lib->vba.MaxDCHUBToPSCLThroughput, 1190 mode_lib->vba.MaxPSCLToLBThroughput); 1191 } 1192 1193 mode_lib->vba.DPPCLKUsingSingleDPPLuma = 1194 mode_lib->vba.PixelClock[k] 1195 * dml_max( 1196 mode_lib->vba.vtaps[k] / 6.0 1197 * dml_min( 1198 1.0, 1199 mode_lib->vba.HRatio[k]), 1200 dml_max( 1201 mode_lib->vba.HRatio[k] 1202 * mode_lib->vba.VRatio[k] 1203 / mode_lib->vba.PSCL_THROUGHPUT_LUMA[k], 1204 1.0)); 1205 1206 if ((mode_lib->vba.htaps[k] > 6 || mode_lib->vba.vtaps[k] > 6) 1207 && mode_lib->vba.DPPCLKUsingSingleDPPLuma 1208 < 2 * mode_lib->vba.PixelClock[k]) { 1209 mode_lib->vba.DPPCLKUsingSingleDPPLuma = 2 * mode_lib->vba.PixelClock[k]; 1210 } 1211 1212 if ((mode_lib->vba.SourcePixelFormat[k] != dm_420_8 1213 && mode_lib->vba.SourcePixelFormat[k] != dm_420_10)) { 1214 mode_lib->vba.PSCL_THROUGHPUT_CHROMA[k] = 0.0; 1215 mode_lib->vba.DPPCLKUsingSingleDPP[k] = 1216 mode_lib->vba.DPPCLKUsingSingleDPPLuma; 1217 } else { 1218 if (mode_lib->vba.HRatio[k] > 1) { 1219 mode_lib->vba.PSCL_THROUGHPUT_CHROMA[k] = 1220 dml_min( 1221 mode_lib->vba.MaxDCHUBToPSCLThroughput, 1222 mode_lib->vba.MaxPSCLToLBThroughput 1223 * mode_lib->vba.HRatio[k] 1224 / 2 1225 / dml_ceil( 1226 mode_lib->vba.HTAPsChroma[k] 1227 / 6.0, 1228 1.0)); 1229 } else { 1230 mode_lib->vba.PSCL_THROUGHPUT_CHROMA[k] = dml_min( 1231 mode_lib->vba.MaxDCHUBToPSCLThroughput, 1232 mode_lib->vba.MaxPSCLToLBThroughput); 1233 } 1234 mode_lib->vba.DPPCLKUsingSingleDPPChroma = 1235 mode_lib->vba.PixelClock[k] 1236 * dml_max( 1237 mode_lib->vba.VTAPsChroma[k] 1238 / 6.0 1239 * dml_min( 1240 1.0, 1241 mode_lib->vba.HRatio[k] 1242 / 2), 1243 dml_max( 1244 mode_lib->vba.HRatio[k] 1245 * mode_lib->vba.VRatio[k] 1246 / 4 1247 / mode_lib->vba.PSCL_THROUGHPUT_CHROMA[k], 1248 1.0)); 1249 1250 if ((mode_lib->vba.HTAPsChroma[k] > 6 || mode_lib->vba.VTAPsChroma[k] > 6) 1251 && mode_lib->vba.DPPCLKUsingSingleDPPChroma 1252 < 2 * mode_lib->vba.PixelClock[k]) { 1253 mode_lib->vba.DPPCLKUsingSingleDPPChroma = 2 1254 * mode_lib->vba.PixelClock[k]; 1255 } 1256 1257 mode_lib->vba.DPPCLKUsingSingleDPP[k] = dml_max( 1258 mode_lib->vba.DPPCLKUsingSingleDPPLuma, 1259 mode_lib->vba.DPPCLKUsingSingleDPPChroma); 1260 } 1261 } 1262 1263 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 1264 if (mode_lib->vba.BlendingAndTiming[k] != k) 1265 continue; 1266 if (mode_lib->vba.ODMCombineEnabled[k]) { 1267 mode_lib->vba.DISPCLKWithRamping = 1268 dml_max( 1269 mode_lib->vba.DISPCLKWithRamping, 1270 mode_lib->vba.PixelClock[k] / 2 1271 * (1 1272 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading 1273 / 100) 1274 * (1 1275 + mode_lib->vba.DISPCLKRampingMargin 1276 / 100)); 1277 mode_lib->vba.DISPCLKWithoutRamping = 1278 dml_max( 1279 mode_lib->vba.DISPCLKWithoutRamping, 1280 mode_lib->vba.PixelClock[k] / 2 1281 * (1 1282 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading 1283 / 100)); 1284 } else if (!mode_lib->vba.ODMCombineEnabled[k]) { 1285 mode_lib->vba.DISPCLKWithRamping = 1286 dml_max( 1287 mode_lib->vba.DISPCLKWithRamping, 1288 mode_lib->vba.PixelClock[k] 1289 * (1 1290 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading 1291 / 100) 1292 * (1 1293 + mode_lib->vba.DISPCLKRampingMargin 1294 / 100)); 1295 mode_lib->vba.DISPCLKWithoutRamping = 1296 dml_max( 1297 mode_lib->vba.DISPCLKWithoutRamping, 1298 mode_lib->vba.PixelClock[k] 1299 * (1 1300 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading 1301 / 100)); 1302 } 1303 } 1304 1305 mode_lib->vba.DISPCLKWithRamping = dml_max( 1306 mode_lib->vba.DISPCLKWithRamping, 1307 mode_lib->vba.WritebackDISPCLK); 1308 mode_lib->vba.DISPCLKWithoutRamping = dml_max( 1309 mode_lib->vba.DISPCLKWithoutRamping, 1310 mode_lib->vba.WritebackDISPCLK); 1311 1312 ASSERT(mode_lib->vba.DISPCLKDPPCLKVCOSpeed != 0); 1313 mode_lib->vba.DISPCLKWithRampingRoundedToDFSGranularity = RoundToDFSGranularityUp( 1314 mode_lib->vba.DISPCLKWithRamping, 1315 mode_lib->vba.DISPCLKDPPCLKVCOSpeed); 1316 mode_lib->vba.DISPCLKWithoutRampingRoundedToDFSGranularity = RoundToDFSGranularityUp( 1317 mode_lib->vba.DISPCLKWithoutRamping, 1318 mode_lib->vba.DISPCLKDPPCLKVCOSpeed); 1319 mode_lib->vba.MaxDispclkRoundedToDFSGranularity = RoundToDFSGranularityDown( 1320 mode_lib->vba.soc.clock_limits[mode_lib->vba.soc.num_states].dispclk_mhz, 1321 mode_lib->vba.DISPCLKDPPCLKVCOSpeed); 1322 if (mode_lib->vba.DISPCLKWithoutRampingRoundedToDFSGranularity 1323 > mode_lib->vba.MaxDispclkRoundedToDFSGranularity) { 1324 mode_lib->vba.DISPCLK_calculated = 1325 mode_lib->vba.DISPCLKWithoutRampingRoundedToDFSGranularity; 1326 } else if (mode_lib->vba.DISPCLKWithRampingRoundedToDFSGranularity 1327 > mode_lib->vba.MaxDispclkRoundedToDFSGranularity) { 1328 mode_lib->vba.DISPCLK_calculated = mode_lib->vba.MaxDispclkRoundedToDFSGranularity; 1329 } else { 1330 mode_lib->vba.DISPCLK_calculated = 1331 mode_lib->vba.DISPCLKWithRampingRoundedToDFSGranularity; 1332 } 1333 DTRACE(" dispclk_mhz (calculated) = %f", mode_lib->vba.DISPCLK_calculated); 1334 1335 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 1336 if (mode_lib->vba.DPPPerPlane[k] == 0) { 1337 mode_lib->vba.DPPCLK_calculated[k] = 0; 1338 } else { 1339 mode_lib->vba.DPPCLK_calculated[k] = mode_lib->vba.DPPCLKUsingSingleDPP[k] 1340 / mode_lib->vba.DPPPerPlane[k] 1341 * (1 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100); 1342 } 1343 mode_lib->vba.GlobalDPPCLK = dml_max( 1344 mode_lib->vba.GlobalDPPCLK, 1345 mode_lib->vba.DPPCLK_calculated[k]); 1346 } 1347 mode_lib->vba.GlobalDPPCLK = RoundToDFSGranularityUp( 1348 mode_lib->vba.GlobalDPPCLK, 1349 mode_lib->vba.DISPCLKDPPCLKVCOSpeed); 1350 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 1351 mode_lib->vba.DPPCLK_calculated[k] = mode_lib->vba.GlobalDPPCLK / 255 1352 * dml_ceil( 1353 mode_lib->vba.DPPCLK_calculated[k] * 255 1354 / mode_lib->vba.GlobalDPPCLK, 1355 1); 1356 DTRACE(" dppclk_mhz[%i] (calculated) = %f", k, mode_lib->vba.DPPCLK_calculated[k]); 1357 } 1358 1359 // Urgent Watermark 1360 mode_lib->vba.DCCEnabledAnyPlane = false; 1361 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) 1362 if (mode_lib->vba.DCCEnable[k]) 1363 mode_lib->vba.DCCEnabledAnyPlane = true; 1364 1365 mode_lib->vba.ReturnBandwidthToDCN = dml_min( 1366 mode_lib->vba.ReturnBusWidth * mode_lib->vba.DCFCLK, 1367 mode_lib->vba.FabricAndDRAMBandwidth * 1000) 1368 * mode_lib->vba.PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelDataOnly / 100; 1369 1370 mode_lib->vba.ReturnBW = mode_lib->vba.ReturnBandwidthToDCN; 1371 mode_lib->vba.ReturnBW = adjust_ReturnBW( 1372 mode_lib, 1373 mode_lib->vba.ReturnBW, 1374 mode_lib->vba.DCCEnabledAnyPlane, 1375 mode_lib->vba.ReturnBandwidthToDCN); 1376 1377 // Let's do this calculation again?? 1378 mode_lib->vba.ReturnBandwidthToDCN = dml_min( 1379 mode_lib->vba.ReturnBusWidth * mode_lib->vba.DCFCLK, 1380 mode_lib->vba.FabricAndDRAMBandwidth * 1000); 1381 mode_lib->vba.ReturnBW = adjust_ReturnBW( 1382 mode_lib, 1383 mode_lib->vba.ReturnBW, 1384 mode_lib->vba.DCCEnabledAnyPlane, 1385 mode_lib->vba.ReturnBandwidthToDCN); 1386 1387 DTRACE(" dcfclk_mhz = %f", mode_lib->vba.DCFCLK); 1388 DTRACE(" return_bw_to_dcn = %f", mode_lib->vba.ReturnBandwidthToDCN); 1389 DTRACE(" return_bus_bw = %f", mode_lib->vba.ReturnBW); 1390 1391 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 1392 bool MainPlaneDoesODMCombine = false; 1393 1394 if (mode_lib->vba.SourceScan[k] == dm_horz) 1395 mode_lib->vba.SwathWidthSingleDPPY[k] = mode_lib->vba.ViewportWidth[k]; 1396 else 1397 mode_lib->vba.SwathWidthSingleDPPY[k] = mode_lib->vba.ViewportHeight[k]; 1398 1399 if (mode_lib->vba.ODMCombineEnabled[k] == dm_odm_combine_mode_2to1) 1400 MainPlaneDoesODMCombine = true; 1401 for (j = 0; j < mode_lib->vba.NumberOfActivePlanes; ++j) 1402 if (mode_lib->vba.BlendingAndTiming[k] == j 1403 && mode_lib->vba.ODMCombineEnabled[k] == dm_odm_combine_mode_2to1) 1404 MainPlaneDoesODMCombine = true; 1405 1406 if (MainPlaneDoesODMCombine == true) 1407 mode_lib->vba.SwathWidthY[k] = dml_min( 1408 (double) mode_lib->vba.SwathWidthSingleDPPY[k], 1409 dml_round( 1410 mode_lib->vba.HActive[k] / 2.0 1411 * mode_lib->vba.HRatio[k])); 1412 else { 1413 if (mode_lib->vba.DPPPerPlane[k] == 0) { 1414 mode_lib->vba.SwathWidthY[k] = 0; 1415 } else { 1416 mode_lib->vba.SwathWidthY[k] = mode_lib->vba.SwathWidthSingleDPPY[k] 1417 / mode_lib->vba.DPPPerPlane[k]; 1418 } 1419 } 1420 } 1421 1422 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 1423 if (mode_lib->vba.SourcePixelFormat[k] == dm_444_64) { 1424 mode_lib->vba.BytePerPixelDETY[k] = 8; 1425 mode_lib->vba.BytePerPixelDETC[k] = 0; 1426 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_32) { 1427 mode_lib->vba.BytePerPixelDETY[k] = 4; 1428 mode_lib->vba.BytePerPixelDETC[k] = 0; 1429 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_16) { 1430 mode_lib->vba.BytePerPixelDETY[k] = 2; 1431 mode_lib->vba.BytePerPixelDETC[k] = 0; 1432 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_8) { 1433 mode_lib->vba.BytePerPixelDETY[k] = 1; 1434 mode_lib->vba.BytePerPixelDETC[k] = 0; 1435 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_8) { 1436 mode_lib->vba.BytePerPixelDETY[k] = 1; 1437 mode_lib->vba.BytePerPixelDETC[k] = 2; 1438 } else { // dm_420_10 1439 mode_lib->vba.BytePerPixelDETY[k] = 4.0 / 3.0; 1440 mode_lib->vba.BytePerPixelDETC[k] = 8.0 / 3.0; 1441 } 1442 } 1443 1444 mode_lib->vba.TotalDataReadBandwidth = 0.0; 1445 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 1446 mode_lib->vba.ReadBandwidthPlaneLuma[k] = mode_lib->vba.SwathWidthSingleDPPY[k] 1447 * dml_ceil(mode_lib->vba.BytePerPixelDETY[k], 1) 1448 / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]) 1449 * mode_lib->vba.VRatio[k]; 1450 mode_lib->vba.ReadBandwidthPlaneChroma[k] = mode_lib->vba.SwathWidthSingleDPPY[k] 1451 / 2 * dml_ceil(mode_lib->vba.BytePerPixelDETC[k], 2) 1452 / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]) 1453 * mode_lib->vba.VRatio[k] / 2; 1454 DTRACE( 1455 " read_bw[%i] = %fBps", 1456 k, 1457 mode_lib->vba.ReadBandwidthPlaneLuma[k] 1458 + mode_lib->vba.ReadBandwidthPlaneChroma[k]); 1459 mode_lib->vba.TotalDataReadBandwidth += mode_lib->vba.ReadBandwidthPlaneLuma[k] 1460 + mode_lib->vba.ReadBandwidthPlaneChroma[k]; 1461 } 1462 1463 mode_lib->vba.TotalDCCActiveDPP = 0; 1464 mode_lib->vba.TotalActiveDPP = 0; 1465 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 1466 mode_lib->vba.TotalActiveDPP = mode_lib->vba.TotalActiveDPP 1467 + mode_lib->vba.DPPPerPlane[k]; 1468 if (mode_lib->vba.DCCEnable[k]) 1469 mode_lib->vba.TotalDCCActiveDPP = mode_lib->vba.TotalDCCActiveDPP 1470 + mode_lib->vba.DPPPerPlane[k]; 1471 } 1472 1473 mode_lib->vba.UrgentRoundTripAndOutOfOrderLatency = 1474 (mode_lib->vba.RoundTripPingLatencyCycles + 32) / mode_lib->vba.DCFCLK 1475 + mode_lib->vba.UrgentOutOfOrderReturnPerChannelPixelDataOnly 1476 * mode_lib->vba.NumberOfChannels 1477 / mode_lib->vba.ReturnBW; 1478 1479 mode_lib->vba.LastPixelOfLineExtraWatermark = 0; 1480 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 1481 if (mode_lib->vba.VRatio[k] <= 1.0) 1482 mode_lib->vba.DisplayPipeLineDeliveryTimeLuma[k] = 1483 (double) mode_lib->vba.SwathWidthY[k] 1484 * mode_lib->vba.DPPPerPlane[k] 1485 / mode_lib->vba.HRatio[k] 1486 / mode_lib->vba.PixelClock[k]; 1487 else 1488 mode_lib->vba.DisplayPipeLineDeliveryTimeLuma[k] = 1489 (double) mode_lib->vba.SwathWidthY[k] 1490 / mode_lib->vba.PSCL_THROUGHPUT_LUMA[k] 1491 / mode_lib->vba.DPPCLK[k]; 1492 1493 if (mode_lib->vba.BytePerPixelDETC[k] == 0) 1494 mode_lib->vba.DisplayPipeLineDeliveryTimeChroma[k] = 0.0; 1495 else if (mode_lib->vba.VRatio[k] / 2.0 <= 1.0) 1496 mode_lib->vba.DisplayPipeLineDeliveryTimeChroma[k] = 1497 mode_lib->vba.SwathWidthY[k] / 2.0 1498 * mode_lib->vba.DPPPerPlane[k] 1499 / (mode_lib->vba.HRatio[k] / 2.0) 1500 / mode_lib->vba.PixelClock[k]; 1501 else 1502 mode_lib->vba.DisplayPipeLineDeliveryTimeChroma[k] = 1503 mode_lib->vba.SwathWidthY[k] / 2.0 1504 / mode_lib->vba.PSCL_THROUGHPUT_CHROMA[k] 1505 / mode_lib->vba.DPPCLK[k]; 1506 } 1507 1508 mode_lib->vba.UrgentExtraLatency = mode_lib->vba.UrgentRoundTripAndOutOfOrderLatency 1509 + (mode_lib->vba.TotalActiveDPP * mode_lib->vba.PixelChunkSizeInKByte 1510 + mode_lib->vba.TotalDCCActiveDPP 1511 * mode_lib->vba.MetaChunkSize) * 1024.0 1512 / mode_lib->vba.ReturnBW; 1513 1514 if (mode_lib->vba.GPUVMEnable) 1515 mode_lib->vba.UrgentExtraLatency += mode_lib->vba.TotalActiveDPP 1516 * mode_lib->vba.PTEGroupSize / mode_lib->vba.ReturnBW; 1517 1518 mode_lib->vba.UrgentWatermark = mode_lib->vba.UrgentLatencyPixelDataOnly 1519 + mode_lib->vba.LastPixelOfLineExtraWatermark 1520 + mode_lib->vba.UrgentExtraLatency; 1521 1522 DTRACE(" urgent_extra_latency = %fus", mode_lib->vba.UrgentExtraLatency); 1523 DTRACE(" wm_urgent = %fus", mode_lib->vba.UrgentWatermark); 1524 1525 mode_lib->vba.UrgentLatency = mode_lib->vba.UrgentLatencyPixelDataOnly; 1526 1527 mode_lib->vba.TotalActiveWriteback = 0; 1528 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 1529 if (mode_lib->vba.WritebackEnable[k]) 1530 mode_lib->vba.TotalActiveWriteback = mode_lib->vba.TotalActiveWriteback + mode_lib->vba.ActiveWritebacksPerPlane[k]; 1531 } 1532 1533 if (mode_lib->vba.TotalActiveWriteback <= 1) 1534 mode_lib->vba.WritebackUrgentWatermark = mode_lib->vba.WritebackLatency; 1535 else 1536 mode_lib->vba.WritebackUrgentWatermark = mode_lib->vba.WritebackLatency 1537 + mode_lib->vba.WritebackChunkSize * 1024.0 / 32 1538 / mode_lib->vba.SOCCLK; 1539 1540 DTRACE(" wm_wb_urgent = %fus", mode_lib->vba.WritebackUrgentWatermark); 1541 1542 // NB P-State/DRAM Clock Change Watermark 1543 mode_lib->vba.DRAMClockChangeWatermark = mode_lib->vba.DRAMClockChangeLatency 1544 + mode_lib->vba.UrgentWatermark; 1545 1546 DTRACE(" wm_pstate_change = %fus", mode_lib->vba.DRAMClockChangeWatermark); 1547 1548 DTRACE(" calculating wb pstate watermark"); 1549 DTRACE(" total wb outputs %d", mode_lib->vba.TotalActiveWriteback); 1550 DTRACE(" socclk frequency %f Mhz", mode_lib->vba.SOCCLK); 1551 1552 if (mode_lib->vba.TotalActiveWriteback <= 1) 1553 mode_lib->vba.WritebackDRAMClockChangeWatermark = 1554 mode_lib->vba.DRAMClockChangeLatency 1555 + mode_lib->vba.WritebackLatency; 1556 else 1557 mode_lib->vba.WritebackDRAMClockChangeWatermark = 1558 mode_lib->vba.DRAMClockChangeLatency 1559 + mode_lib->vba.WritebackLatency 1560 + mode_lib->vba.WritebackChunkSize * 1024.0 / 32 1561 / mode_lib->vba.SOCCLK; 1562 1563 DTRACE(" wm_wb_pstate %fus", mode_lib->vba.WritebackDRAMClockChangeWatermark); 1564 1565 // Stutter Efficiency 1566 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 1567 mode_lib->vba.LinesInDETY[k] = mode_lib->vba.DETBufferSizeY[k] 1568 / mode_lib->vba.BytePerPixelDETY[k] / mode_lib->vba.SwathWidthY[k]; 1569 mode_lib->vba.LinesInDETYRoundedDownToSwath[k] = dml_floor( 1570 mode_lib->vba.LinesInDETY[k], 1571 mode_lib->vba.SwathHeightY[k]); 1572 mode_lib->vba.FullDETBufferingTimeY[k] = 1573 mode_lib->vba.LinesInDETYRoundedDownToSwath[k] 1574 * (mode_lib->vba.HTotal[k] 1575 / mode_lib->vba.PixelClock[k]) 1576 / mode_lib->vba.VRatio[k]; 1577 if (mode_lib->vba.BytePerPixelDETC[k] > 0) { 1578 mode_lib->vba.LinesInDETC[k] = mode_lib->vba.DETBufferSizeC[k] 1579 / mode_lib->vba.BytePerPixelDETC[k] 1580 / (mode_lib->vba.SwathWidthY[k] / 2); 1581 mode_lib->vba.LinesInDETCRoundedDownToSwath[k] = dml_floor( 1582 mode_lib->vba.LinesInDETC[k], 1583 mode_lib->vba.SwathHeightC[k]); 1584 mode_lib->vba.FullDETBufferingTimeC[k] = 1585 mode_lib->vba.LinesInDETCRoundedDownToSwath[k] 1586 * (mode_lib->vba.HTotal[k] 1587 / mode_lib->vba.PixelClock[k]) 1588 / (mode_lib->vba.VRatio[k] / 2); 1589 } else { 1590 mode_lib->vba.LinesInDETC[k] = 0; 1591 mode_lib->vba.LinesInDETCRoundedDownToSwath[k] = 0; 1592 mode_lib->vba.FullDETBufferingTimeC[k] = 999999; 1593 } 1594 } 1595 1596 mode_lib->vba.MinFullDETBufferingTime = 999999.0; 1597 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 1598 if (mode_lib->vba.FullDETBufferingTimeY[k] 1599 < mode_lib->vba.MinFullDETBufferingTime) { 1600 mode_lib->vba.MinFullDETBufferingTime = 1601 mode_lib->vba.FullDETBufferingTimeY[k]; 1602 mode_lib->vba.FrameTimeForMinFullDETBufferingTime = 1603 (double) mode_lib->vba.VTotal[k] * mode_lib->vba.HTotal[k] 1604 / mode_lib->vba.PixelClock[k]; 1605 } 1606 if (mode_lib->vba.FullDETBufferingTimeC[k] 1607 < mode_lib->vba.MinFullDETBufferingTime) { 1608 mode_lib->vba.MinFullDETBufferingTime = 1609 mode_lib->vba.FullDETBufferingTimeC[k]; 1610 mode_lib->vba.FrameTimeForMinFullDETBufferingTime = 1611 (double) mode_lib->vba.VTotal[k] * mode_lib->vba.HTotal[k] 1612 / mode_lib->vba.PixelClock[k]; 1613 } 1614 } 1615 1616 mode_lib->vba.AverageReadBandwidthGBytePerSecond = 0.0; 1617 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 1618 if (mode_lib->vba.DCCEnable[k]) { 1619 mode_lib->vba.AverageReadBandwidthGBytePerSecond = 1620 mode_lib->vba.AverageReadBandwidthGBytePerSecond 1621 + mode_lib->vba.ReadBandwidthPlaneLuma[k] 1622 / mode_lib->vba.DCCRate[k] 1623 / 1000 1624 + mode_lib->vba.ReadBandwidthPlaneChroma[k] 1625 / mode_lib->vba.DCCRate[k] 1626 / 1000; 1627 } else { 1628 mode_lib->vba.AverageReadBandwidthGBytePerSecond = 1629 mode_lib->vba.AverageReadBandwidthGBytePerSecond 1630 + mode_lib->vba.ReadBandwidthPlaneLuma[k] 1631 / 1000 1632 + mode_lib->vba.ReadBandwidthPlaneChroma[k] 1633 / 1000; 1634 } 1635 if (mode_lib->vba.DCCEnable[k]) { 1636 mode_lib->vba.AverageReadBandwidthGBytePerSecond = 1637 mode_lib->vba.AverageReadBandwidthGBytePerSecond 1638 + mode_lib->vba.ReadBandwidthPlaneLuma[k] 1639 / 1000 / 256 1640 + mode_lib->vba.ReadBandwidthPlaneChroma[k] 1641 / 1000 / 256; 1642 } 1643 if (mode_lib->vba.GPUVMEnable) { 1644 mode_lib->vba.AverageReadBandwidthGBytePerSecond = 1645 mode_lib->vba.AverageReadBandwidthGBytePerSecond 1646 + mode_lib->vba.ReadBandwidthPlaneLuma[k] 1647 / 1000 / 512 1648 + mode_lib->vba.ReadBandwidthPlaneChroma[k] 1649 / 1000 / 512; 1650 } 1651 } 1652 1653 mode_lib->vba.PartOfBurstThatFitsInROB = 1654 dml_min( 1655 mode_lib->vba.MinFullDETBufferingTime 1656 * mode_lib->vba.TotalDataReadBandwidth, 1657 mode_lib->vba.ROBBufferSizeInKByte * 1024 1658 * mode_lib->vba.TotalDataReadBandwidth 1659 / (mode_lib->vba.AverageReadBandwidthGBytePerSecond 1660 * 1000)); 1661 mode_lib->vba.StutterBurstTime = mode_lib->vba.PartOfBurstThatFitsInROB 1662 * (mode_lib->vba.AverageReadBandwidthGBytePerSecond * 1000) 1663 / mode_lib->vba.TotalDataReadBandwidth / mode_lib->vba.ReturnBW 1664 + (mode_lib->vba.MinFullDETBufferingTime 1665 * mode_lib->vba.TotalDataReadBandwidth 1666 - mode_lib->vba.PartOfBurstThatFitsInROB) 1667 / (mode_lib->vba.DCFCLK * 64); 1668 if (mode_lib->vba.TotalActiveWriteback == 0) { 1669 mode_lib->vba.StutterEfficiencyNotIncludingVBlank = (1 1670 - (mode_lib->vba.SRExitTime + mode_lib->vba.StutterBurstTime) 1671 / mode_lib->vba.MinFullDETBufferingTime) * 100; 1672 } else { 1673 mode_lib->vba.StutterEfficiencyNotIncludingVBlank = 0; 1674 } 1675 1676 mode_lib->vba.SmallestVBlank = 999999; 1677 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 1678 if (mode_lib->vba.SynchronizedVBlank || mode_lib->vba.NumberOfActivePlanes == 1) { 1679 mode_lib->vba.VBlankTime = (double) (mode_lib->vba.VTotal[k] 1680 - mode_lib->vba.VActive[k]) * mode_lib->vba.HTotal[k] 1681 / mode_lib->vba.PixelClock[k]; 1682 } else { 1683 mode_lib->vba.VBlankTime = 0; 1684 } 1685 mode_lib->vba.SmallestVBlank = dml_min( 1686 mode_lib->vba.SmallestVBlank, 1687 mode_lib->vba.VBlankTime); 1688 } 1689 1690 mode_lib->vba.StutterEfficiency = (mode_lib->vba.StutterEfficiencyNotIncludingVBlank / 100 1691 * (mode_lib->vba.FrameTimeForMinFullDETBufferingTime 1692 - mode_lib->vba.SmallestVBlank) 1693 + mode_lib->vba.SmallestVBlank) 1694 / mode_lib->vba.FrameTimeForMinFullDETBufferingTime * 100; 1695 1696 // dml_ml->vba.DCFCLK Deep Sleep 1697 mode_lib->vba.DCFCLKDeepSleep = 8.0; 1698 1699 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; k++) { 1700 if (mode_lib->vba.BytePerPixelDETC[k] > 0) { 1701 mode_lib->vba.DCFCLKDeepSleepPerPlane[k] = 1702 dml_max( 1703 1.1 * mode_lib->vba.SwathWidthY[k] 1704 * dml_ceil( 1705 mode_lib->vba.BytePerPixelDETY[k], 1706 1) / 32 1707 / mode_lib->vba.DisplayPipeLineDeliveryTimeLuma[k], 1708 1.1 * mode_lib->vba.SwathWidthY[k] / 2.0 1709 * dml_ceil( 1710 mode_lib->vba.BytePerPixelDETC[k], 1711 2) / 32 1712 / mode_lib->vba.DisplayPipeLineDeliveryTimeChroma[k]); 1713 } else 1714 mode_lib->vba.DCFCLKDeepSleepPerPlane[k] = 1.1 * mode_lib->vba.SwathWidthY[k] 1715 * dml_ceil(mode_lib->vba.BytePerPixelDETY[k], 1) / 64.0 1716 / mode_lib->vba.DisplayPipeLineDeliveryTimeLuma[k]; 1717 mode_lib->vba.DCFCLKDeepSleepPerPlane[k] = dml_max( 1718 mode_lib->vba.DCFCLKDeepSleepPerPlane[k], 1719 mode_lib->vba.PixelClock[k] / 16.0); 1720 mode_lib->vba.DCFCLKDeepSleep = dml_max( 1721 mode_lib->vba.DCFCLKDeepSleep, 1722 mode_lib->vba.DCFCLKDeepSleepPerPlane[k]); 1723 1724 DTRACE( 1725 " dcfclk_deepsleep_per_plane[%i] = %fMHz", 1726 k, 1727 mode_lib->vba.DCFCLKDeepSleepPerPlane[k]); 1728 } 1729 1730 DTRACE(" dcfclk_deepsleep_mhz = %fMHz", mode_lib->vba.DCFCLKDeepSleep); 1731 1732 // Stutter Watermark 1733 mode_lib->vba.StutterExitWatermark = mode_lib->vba.SRExitTime 1734 + mode_lib->vba.LastPixelOfLineExtraWatermark 1735 + mode_lib->vba.UrgentExtraLatency + 10 / mode_lib->vba.DCFCLKDeepSleep; 1736 mode_lib->vba.StutterEnterPlusExitWatermark = mode_lib->vba.SREnterPlusExitTime 1737 + mode_lib->vba.LastPixelOfLineExtraWatermark 1738 + mode_lib->vba.UrgentExtraLatency; 1739 1740 DTRACE(" wm_cstate_exit = %fus", mode_lib->vba.StutterExitWatermark); 1741 DTRACE(" wm_cstate_enter_exit = %fus", mode_lib->vba.StutterEnterPlusExitWatermark); 1742 1743 // Urgent Latency Supported 1744 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 1745 mode_lib->vba.EffectiveDETPlusLBLinesLuma = 1746 dml_floor( 1747 mode_lib->vba.LinesInDETY[k] 1748 + dml_min( 1749 mode_lib->vba.LinesInDETY[k] 1750 * mode_lib->vba.DPPCLK[k] 1751 * mode_lib->vba.BytePerPixelDETY[k] 1752 * mode_lib->vba.PSCL_THROUGHPUT_LUMA[k] 1753 / (mode_lib->vba.ReturnBW 1754 / mode_lib->vba.DPPPerPlane[k]), 1755 (double) mode_lib->vba.EffectiveLBLatencyHidingSourceLinesLuma), 1756 mode_lib->vba.SwathHeightY[k]); 1757 1758 mode_lib->vba.UrgentLatencySupportUsLuma = mode_lib->vba.EffectiveDETPlusLBLinesLuma 1759 * (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]) 1760 / mode_lib->vba.VRatio[k] 1761 - mode_lib->vba.EffectiveDETPlusLBLinesLuma 1762 * mode_lib->vba.SwathWidthY[k] 1763 * mode_lib->vba.BytePerPixelDETY[k] 1764 / (mode_lib->vba.ReturnBW 1765 / mode_lib->vba.DPPPerPlane[k]); 1766 1767 if (mode_lib->vba.BytePerPixelDETC[k] > 0) { 1768 mode_lib->vba.EffectiveDETPlusLBLinesChroma = 1769 dml_floor( 1770 mode_lib->vba.LinesInDETC[k] 1771 + dml_min( 1772 mode_lib->vba.LinesInDETC[k] 1773 * mode_lib->vba.DPPCLK[k] 1774 * mode_lib->vba.BytePerPixelDETC[k] 1775 * mode_lib->vba.PSCL_THROUGHPUT_CHROMA[k] 1776 / (mode_lib->vba.ReturnBW 1777 / mode_lib->vba.DPPPerPlane[k]), 1778 (double) mode_lib->vba.EffectiveLBLatencyHidingSourceLinesChroma), 1779 mode_lib->vba.SwathHeightC[k]); 1780 mode_lib->vba.UrgentLatencySupportUsChroma = 1781 mode_lib->vba.EffectiveDETPlusLBLinesChroma 1782 * (mode_lib->vba.HTotal[k] 1783 / mode_lib->vba.PixelClock[k]) 1784 / (mode_lib->vba.VRatio[k] / 2) 1785 - mode_lib->vba.EffectiveDETPlusLBLinesChroma 1786 * (mode_lib->vba.SwathWidthY[k] 1787 / 2) 1788 * mode_lib->vba.BytePerPixelDETC[k] 1789 / (mode_lib->vba.ReturnBW 1790 / mode_lib->vba.DPPPerPlane[k]); 1791 mode_lib->vba.UrgentLatencySupportUs[k] = dml_min( 1792 mode_lib->vba.UrgentLatencySupportUsLuma, 1793 mode_lib->vba.UrgentLatencySupportUsChroma); 1794 } else { 1795 mode_lib->vba.UrgentLatencySupportUs[k] = 1796 mode_lib->vba.UrgentLatencySupportUsLuma; 1797 } 1798 } 1799 1800 mode_lib->vba.MinUrgentLatencySupportUs = 999999; 1801 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 1802 mode_lib->vba.MinUrgentLatencySupportUs = dml_min( 1803 mode_lib->vba.MinUrgentLatencySupportUs, 1804 mode_lib->vba.UrgentLatencySupportUs[k]); 1805 } 1806 1807 // Non-Urgent Latency Tolerance 1808 mode_lib->vba.NonUrgentLatencyTolerance = mode_lib->vba.MinUrgentLatencySupportUs 1809 - mode_lib->vba.UrgentWatermark; 1810 1811 // DSCCLK 1812 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 1813 if ((mode_lib->vba.BlendingAndTiming[k] != k) || !mode_lib->vba.DSCEnabled[k]) { 1814 mode_lib->vba.DSCCLK_calculated[k] = 0.0; 1815 } else { 1816 if (mode_lib->vba.OutputFormat[k] == dm_420 1817 || mode_lib->vba.OutputFormat[k] == dm_n422) 1818 mode_lib->vba.DSCFormatFactor = 2; 1819 else 1820 mode_lib->vba.DSCFormatFactor = 1; 1821 if (mode_lib->vba.ODMCombineEnabled[k]) 1822 mode_lib->vba.DSCCLK_calculated[k] = 1823 mode_lib->vba.PixelClockBackEnd[k] / 6 1824 / mode_lib->vba.DSCFormatFactor 1825 / (1 1826 - mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading 1827 / 100); 1828 else 1829 mode_lib->vba.DSCCLK_calculated[k] = 1830 mode_lib->vba.PixelClockBackEnd[k] / 3 1831 / mode_lib->vba.DSCFormatFactor 1832 / (1 1833 - mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading 1834 / 100); 1835 } 1836 } 1837 1838 // DSC Delay 1839 // TODO 1840 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 1841 double bpp = mode_lib->vba.OutputBpp[k]; 1842 unsigned int slices = mode_lib->vba.NumberOfDSCSlices[k]; 1843 1844 if (mode_lib->vba.DSCEnabled[k] && bpp != 0) { 1845 if (!mode_lib->vba.ODMCombineEnabled[k]) { 1846 mode_lib->vba.DSCDelay[k] = 1847 dscceComputeDelay( 1848 mode_lib->vba.DSCInputBitPerComponent[k], 1849 bpp, 1850 dml_ceil( 1851 (double) mode_lib->vba.HActive[k] 1852 / mode_lib->vba.NumberOfDSCSlices[k], 1853 1), 1854 slices, 1855 mode_lib->vba.OutputFormat[k]) 1856 + dscComputeDelay( 1857 mode_lib->vba.OutputFormat[k]); 1858 } else { 1859 mode_lib->vba.DSCDelay[k] = 1860 2 1861 * (dscceComputeDelay( 1862 mode_lib->vba.DSCInputBitPerComponent[k], 1863 bpp, 1864 dml_ceil( 1865 (double) mode_lib->vba.HActive[k] 1866 / mode_lib->vba.NumberOfDSCSlices[k], 1867 1), 1868 slices / 2.0, 1869 mode_lib->vba.OutputFormat[k]) 1870 + dscComputeDelay( 1871 mode_lib->vba.OutputFormat[k])); 1872 } 1873 mode_lib->vba.DSCDelay[k] = mode_lib->vba.DSCDelay[k] 1874 * mode_lib->vba.PixelClock[k] 1875 / mode_lib->vba.PixelClockBackEnd[k]; 1876 } else { 1877 mode_lib->vba.DSCDelay[k] = 0; 1878 } 1879 } 1880 1881 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) 1882 for (j = 0; j < mode_lib->vba.NumberOfActivePlanes; ++j) // NumberOfPlanes 1883 if (j != k && mode_lib->vba.BlendingAndTiming[k] == j 1884 && mode_lib->vba.DSCEnabled[j]) 1885 mode_lib->vba.DSCDelay[k] = mode_lib->vba.DSCDelay[j]; 1886 1887 // Prefetch 1888 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 1889 unsigned int PDEAndMetaPTEBytesFrameY; 1890 unsigned int PixelPTEBytesPerRowY; 1891 unsigned int MetaRowByteY; 1892 unsigned int MetaRowByteC; 1893 unsigned int PDEAndMetaPTEBytesFrameC; 1894 unsigned int PixelPTEBytesPerRowC; 1895 1896 Calculate256BBlockSizes( 1897 mode_lib->vba.SourcePixelFormat[k], 1898 mode_lib->vba.SurfaceTiling[k], 1899 dml_ceil(mode_lib->vba.BytePerPixelDETY[k], 1), 1900 dml_ceil(mode_lib->vba.BytePerPixelDETC[k], 2), 1901 &mode_lib->vba.BlockHeight256BytesY[k], 1902 &mode_lib->vba.BlockHeight256BytesC[k], 1903 &mode_lib->vba.BlockWidth256BytesY[k], 1904 &mode_lib->vba.BlockWidth256BytesC[k]); 1905 PDEAndMetaPTEBytesFrameY = CalculateVMAndRowBytes( 1906 mode_lib, 1907 mode_lib->vba.DCCEnable[k], 1908 mode_lib->vba.BlockHeight256BytesY[k], 1909 mode_lib->vba.BlockWidth256BytesY[k], 1910 mode_lib->vba.SourcePixelFormat[k], 1911 mode_lib->vba.SurfaceTiling[k], 1912 dml_ceil(mode_lib->vba.BytePerPixelDETY[k], 1), 1913 mode_lib->vba.SourceScan[k], 1914 mode_lib->vba.ViewportWidth[k], 1915 mode_lib->vba.ViewportHeight[k], 1916 mode_lib->vba.SwathWidthY[k], 1917 mode_lib->vba.GPUVMEnable, 1918 mode_lib->vba.VMMPageSize, 1919 mode_lib->vba.PTEBufferSizeInRequestsLuma, 1920 mode_lib->vba.PDEProcessingBufIn64KBReqs, 1921 mode_lib->vba.PitchY[k], 1922 mode_lib->vba.DCCMetaPitchY[k], 1923 &mode_lib->vba.MacroTileWidthY[k], 1924 &MetaRowByteY, 1925 &PixelPTEBytesPerRowY, 1926 &mode_lib->vba.PTEBufferSizeNotExceeded[mode_lib->vba.VoltageLevel][0], 1927 &mode_lib->vba.dpte_row_height[k], 1928 &mode_lib->vba.meta_row_height[k]); 1929 mode_lib->vba.PrefetchSourceLinesY[k] = CalculatePrefetchSourceLines( 1930 mode_lib, 1931 mode_lib->vba.VRatio[k], 1932 mode_lib->vba.vtaps[k], 1933 mode_lib->vba.Interlace[k], 1934 mode_lib->vba.ProgressiveToInterlaceUnitInOPP, 1935 mode_lib->vba.SwathHeightY[k], 1936 mode_lib->vba.ViewportYStartY[k], 1937 &mode_lib->vba.VInitPreFillY[k], 1938 &mode_lib->vba.MaxNumSwathY[k]); 1939 1940 if ((mode_lib->vba.SourcePixelFormat[k] != dm_444_64 1941 && mode_lib->vba.SourcePixelFormat[k] != dm_444_32 1942 && mode_lib->vba.SourcePixelFormat[k] != dm_444_16 1943 && mode_lib->vba.SourcePixelFormat[k] != dm_444_8)) { 1944 PDEAndMetaPTEBytesFrameC = 1945 CalculateVMAndRowBytes( 1946 mode_lib, 1947 mode_lib->vba.DCCEnable[k], 1948 mode_lib->vba.BlockHeight256BytesC[k], 1949 mode_lib->vba.BlockWidth256BytesC[k], 1950 mode_lib->vba.SourcePixelFormat[k], 1951 mode_lib->vba.SurfaceTiling[k], 1952 dml_ceil( 1953 mode_lib->vba.BytePerPixelDETC[k], 1954 2), 1955 mode_lib->vba.SourceScan[k], 1956 mode_lib->vba.ViewportWidth[k] / 2, 1957 mode_lib->vba.ViewportHeight[k] / 2, 1958 mode_lib->vba.SwathWidthY[k] / 2, 1959 mode_lib->vba.GPUVMEnable, 1960 mode_lib->vba.VMMPageSize, 1961 mode_lib->vba.PTEBufferSizeInRequestsLuma, 1962 mode_lib->vba.PDEProcessingBufIn64KBReqs, 1963 mode_lib->vba.PitchC[k], 1964 0, 1965 &mode_lib->vba.MacroTileWidthC[k], 1966 &MetaRowByteC, 1967 &PixelPTEBytesPerRowC, 1968 &mode_lib->vba.PTEBufferSizeNotExceeded[mode_lib->vba.VoltageLevel][0], 1969 &mode_lib->vba.dpte_row_height_chroma[k], 1970 &mode_lib->vba.meta_row_height_chroma[k]); 1971 mode_lib->vba.PrefetchSourceLinesC[k] = CalculatePrefetchSourceLines( 1972 mode_lib, 1973 mode_lib->vba.VRatio[k] / 2, 1974 mode_lib->vba.VTAPsChroma[k], 1975 mode_lib->vba.Interlace[k], 1976 mode_lib->vba.ProgressiveToInterlaceUnitInOPP, 1977 mode_lib->vba.SwathHeightC[k], 1978 mode_lib->vba.ViewportYStartC[k], 1979 &mode_lib->vba.VInitPreFillC[k], 1980 &mode_lib->vba.MaxNumSwathC[k]); 1981 } else { 1982 PixelPTEBytesPerRowC = 0; 1983 PDEAndMetaPTEBytesFrameC = 0; 1984 MetaRowByteC = 0; 1985 mode_lib->vba.MaxNumSwathC[k] = 0; 1986 mode_lib->vba.PrefetchSourceLinesC[k] = 0; 1987 } 1988 1989 mode_lib->vba.PixelPTEBytesPerRow[k] = PixelPTEBytesPerRowY + PixelPTEBytesPerRowC; 1990 mode_lib->vba.PDEAndMetaPTEBytesFrame[k] = PDEAndMetaPTEBytesFrameY 1991 + PDEAndMetaPTEBytesFrameC; 1992 mode_lib->vba.MetaRowByte[k] = MetaRowByteY + MetaRowByteC; 1993 1994 CalculateActiveRowBandwidth( 1995 mode_lib->vba.GPUVMEnable, 1996 mode_lib->vba.SourcePixelFormat[k], 1997 mode_lib->vba.VRatio[k], 1998 mode_lib->vba.DCCEnable[k], 1999 mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k], 2000 MetaRowByteY, 2001 MetaRowByteC, 2002 mode_lib->vba.meta_row_height[k], 2003 mode_lib->vba.meta_row_height_chroma[k], 2004 PixelPTEBytesPerRowY, 2005 PixelPTEBytesPerRowC, 2006 mode_lib->vba.dpte_row_height[k], 2007 mode_lib->vba.dpte_row_height_chroma[k], 2008 &mode_lib->vba.meta_row_bw[k], 2009 &mode_lib->vba.dpte_row_bw[k], 2010 &mode_lib->vba.qual_row_bw[k]); 2011 } 2012 2013 mode_lib->vba.TCalc = 24.0 / mode_lib->vba.DCFCLKDeepSleep; 2014 2015 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2016 if (mode_lib->vba.BlendingAndTiming[k] == k) { 2017 if (mode_lib->vba.WritebackEnable[k] == true) { 2018 mode_lib->vba.WritebackDelay[mode_lib->vba.VoltageLevel][k] = 2019 mode_lib->vba.WritebackLatency 2020 + CalculateWriteBackDelay( 2021 mode_lib->vba.WritebackPixelFormat[k], 2022 mode_lib->vba.WritebackHRatio[k], 2023 mode_lib->vba.WritebackVRatio[k], 2024 mode_lib->vba.WritebackLumaHTaps[k], 2025 mode_lib->vba.WritebackLumaVTaps[k], 2026 mode_lib->vba.WritebackChromaHTaps[k], 2027 mode_lib->vba.WritebackChromaVTaps[k], 2028 mode_lib->vba.WritebackDestinationWidth[k]) 2029 / mode_lib->vba.DISPCLK; 2030 } else 2031 mode_lib->vba.WritebackDelay[mode_lib->vba.VoltageLevel][k] = 0; 2032 for (j = 0; j < mode_lib->vba.NumberOfActivePlanes; ++j) { 2033 if (mode_lib->vba.BlendingAndTiming[j] == k 2034 && mode_lib->vba.WritebackEnable[j] == true) { 2035 mode_lib->vba.WritebackDelay[mode_lib->vba.VoltageLevel][k] = 2036 dml_max( 2037 mode_lib->vba.WritebackDelay[mode_lib->vba.VoltageLevel][k], 2038 mode_lib->vba.WritebackLatency 2039 + CalculateWriteBackDelay( 2040 mode_lib->vba.WritebackPixelFormat[j], 2041 mode_lib->vba.WritebackHRatio[j], 2042 mode_lib->vba.WritebackVRatio[j], 2043 mode_lib->vba.WritebackLumaHTaps[j], 2044 mode_lib->vba.WritebackLumaVTaps[j], 2045 mode_lib->vba.WritebackChromaHTaps[j], 2046 mode_lib->vba.WritebackChromaVTaps[j], 2047 mode_lib->vba.WritebackDestinationWidth[j]) 2048 / mode_lib->vba.DISPCLK); 2049 } 2050 } 2051 } 2052 } 2053 2054 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) 2055 for (j = 0; j < mode_lib->vba.NumberOfActivePlanes; ++j) 2056 if (mode_lib->vba.BlendingAndTiming[k] == j) 2057 mode_lib->vba.WritebackDelay[mode_lib->vba.VoltageLevel][k] = 2058 mode_lib->vba.WritebackDelay[mode_lib->vba.VoltageLevel][j]; 2059 2060 mode_lib->vba.VStartupLines = 13; 2061 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2062 mode_lib->vba.MaxVStartupLines[k] = 2063 mode_lib->vba.VTotal[k] - mode_lib->vba.VActive[k] 2064 - dml_max( 2065 1.0, 2066 dml_ceil( 2067 mode_lib->vba.WritebackDelay[mode_lib->vba.VoltageLevel][k] 2068 / (mode_lib->vba.HTotal[k] 2069 / mode_lib->vba.PixelClock[k]), 2070 1)); 2071 } 2072 2073 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) 2074 mode_lib->vba.MaximumMaxVStartupLines = dml_max( 2075 mode_lib->vba.MaximumMaxVStartupLines, 2076 mode_lib->vba.MaxVStartupLines[k]); 2077 2078 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2079 mode_lib->vba.cursor_bw[k] = 0.0; 2080 for (j = 0; j < mode_lib->vba.NumberOfCursors[k]; ++j) 2081 mode_lib->vba.cursor_bw[k] += mode_lib->vba.CursorWidth[k][j] 2082 * mode_lib->vba.CursorBPP[k][j] / 8.0 2083 / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]) 2084 * mode_lib->vba.VRatio[k]; 2085 } 2086 2087 do { 2088 double MaxTotalRDBandwidth = 0; 2089 bool DestinationLineTimesForPrefetchLessThan2 = false; 2090 bool VRatioPrefetchMoreThan4 = false; 2091 bool prefetch_vm_bw_valid = true; 2092 bool prefetch_row_bw_valid = true; 2093 double TWait = CalculateTWait( 2094 mode_lib->vba.PrefetchMode[mode_lib->vba.VoltageLevel][mode_lib->vba.maxMpcComb], 2095 mode_lib->vba.DRAMClockChangeLatency, 2096 mode_lib->vba.UrgentLatencyPixelDataOnly, 2097 mode_lib->vba.SREnterPlusExitTime); 2098 2099 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2100 if (mode_lib->vba.XFCEnabled[k] == true) { 2101 mode_lib->vba.XFCRemoteSurfaceFlipDelay = 2102 CalculateRemoteSurfaceFlipDelay( 2103 mode_lib, 2104 mode_lib->vba.VRatio[k], 2105 mode_lib->vba.SwathWidthY[k], 2106 dml_ceil( 2107 mode_lib->vba.BytePerPixelDETY[k], 2108 1), 2109 mode_lib->vba.HTotal[k] 2110 / mode_lib->vba.PixelClock[k], 2111 mode_lib->vba.XFCTSlvVupdateOffset, 2112 mode_lib->vba.XFCTSlvVupdateWidth, 2113 mode_lib->vba.XFCTSlvVreadyOffset, 2114 mode_lib->vba.XFCXBUFLatencyTolerance, 2115 mode_lib->vba.XFCFillBWOverhead, 2116 mode_lib->vba.XFCSlvChunkSize, 2117 mode_lib->vba.XFCBusTransportTime, 2118 mode_lib->vba.TCalc, 2119 TWait, 2120 &mode_lib->vba.SrcActiveDrainRate, 2121 &mode_lib->vba.TInitXFill, 2122 &mode_lib->vba.TslvChk); 2123 } else { 2124 mode_lib->vba.XFCRemoteSurfaceFlipDelay = 0; 2125 } 2126 2127 CalculateDelayAfterScaler(mode_lib, mode_lib->vba.ReturnBW, mode_lib->vba.ReadBandwidthPlaneLuma[k], mode_lib->vba.ReadBandwidthPlaneChroma[k], mode_lib->vba.TotalDataReadBandwidth, 2128 mode_lib->vba.DisplayPipeLineDeliveryTimeLuma[k], mode_lib->vba.DisplayPipeLineDeliveryTimeChroma[k], 2129 mode_lib->vba.DPPCLK[k], mode_lib->vba.DISPCLK, mode_lib->vba.PixelClock[k], mode_lib->vba.DSCDelay[k], mode_lib->vba.DPPPerPlane[k], mode_lib->vba.ScalerEnabled[k], mode_lib->vba.NumberOfCursors[k], 2130 mode_lib->vba.DPPCLKDelaySubtotal, mode_lib->vba.DPPCLKDelaySCL, mode_lib->vba.DPPCLKDelaySCLLBOnly, mode_lib->vba.DPPCLKDelayCNVCFormater, mode_lib->vba.DPPCLKDelayCNVCCursor, mode_lib->vba.DISPCLKDelaySubtotal, 2131 mode_lib->vba.SwathWidthY[k] / mode_lib->vba.HRatio[k], mode_lib->vba.OutputFormat[k], mode_lib->vba.HTotal[k], 2132 mode_lib->vba.SwathWidthSingleDPPY[k], mode_lib->vba.BytePerPixelDETY[k], mode_lib->vba.BytePerPixelDETC[k], mode_lib->vba.SwathHeightY[k], mode_lib->vba.SwathHeightC[k], mode_lib->vba.Interlace[k], 2133 mode_lib->vba.ProgressiveToInterlaceUnitInOPP, &mode_lib->vba.DSTXAfterScaler[k], &mode_lib->vba.DSTYAfterScaler[k]); 2134 2135 mode_lib->vba.ErrorResult[k] = 2136 CalculatePrefetchSchedule( 2137 mode_lib, 2138 mode_lib->vba.DPPCLK[k], 2139 mode_lib->vba.DISPCLK, 2140 mode_lib->vba.PixelClock[k], 2141 mode_lib->vba.DCFCLKDeepSleep, 2142 mode_lib->vba.DPPPerPlane[k], 2143 mode_lib->vba.NumberOfCursors[k], 2144 mode_lib->vba.VTotal[k] 2145 - mode_lib->vba.VActive[k], 2146 mode_lib->vba.HTotal[k], 2147 mode_lib->vba.MaxInterDCNTileRepeaters, 2148 dml_min( 2149 mode_lib->vba.VStartupLines, 2150 mode_lib->vba.MaxVStartupLines[k]), 2151 mode_lib->vba.GPUVMMaxPageTableLevels, 2152 mode_lib->vba.GPUVMEnable, 2153 mode_lib->vba.DynamicMetadataEnable[k], 2154 mode_lib->vba.DynamicMetadataLinesBeforeActiveRequired[k], 2155 mode_lib->vba.DynamicMetadataTransmittedBytes[k], 2156 mode_lib->vba.DCCEnable[k], 2157 mode_lib->vba.UrgentLatencyPixelDataOnly, 2158 mode_lib->vba.UrgentExtraLatency, 2159 mode_lib->vba.TCalc, 2160 mode_lib->vba.PDEAndMetaPTEBytesFrame[k], 2161 mode_lib->vba.MetaRowByte[k], 2162 mode_lib->vba.PixelPTEBytesPerRow[k], 2163 mode_lib->vba.PrefetchSourceLinesY[k], 2164 mode_lib->vba.SwathWidthY[k], 2165 mode_lib->vba.BytePerPixelDETY[k], 2166 mode_lib->vba.VInitPreFillY[k], 2167 mode_lib->vba.MaxNumSwathY[k], 2168 mode_lib->vba.PrefetchSourceLinesC[k], 2169 mode_lib->vba.BytePerPixelDETC[k], 2170 mode_lib->vba.VInitPreFillC[k], 2171 mode_lib->vba.MaxNumSwathC[k], 2172 mode_lib->vba.SwathHeightY[k], 2173 mode_lib->vba.SwathHeightC[k], 2174 TWait, 2175 mode_lib->vba.XFCEnabled[k], 2176 mode_lib->vba.XFCRemoteSurfaceFlipDelay, 2177 mode_lib->vba.Interlace[k], 2178 mode_lib->vba.ProgressiveToInterlaceUnitInOPP, 2179 mode_lib->vba.DSTXAfterScaler[k], 2180 mode_lib->vba.DSTYAfterScaler[k], 2181 &mode_lib->vba.DestinationLinesForPrefetch[k], 2182 &mode_lib->vba.PrefetchBandwidth[k], 2183 &mode_lib->vba.DestinationLinesToRequestVMInVBlank[k], 2184 &mode_lib->vba.DestinationLinesToRequestRowInVBlank[k], 2185 &mode_lib->vba.VRatioPrefetchY[k], 2186 &mode_lib->vba.VRatioPrefetchC[k], 2187 &mode_lib->vba.RequiredPrefetchPixDataBWLuma[k], 2188 &mode_lib->vba.Tno_bw[k], 2189 &mode_lib->vba.VUpdateOffsetPix[k], 2190 &mode_lib->vba.VUpdateWidthPix[k], 2191 &mode_lib->vba.VReadyOffsetPix[k]); 2192 2193 if (mode_lib->vba.BlendingAndTiming[k] == k) { 2194 mode_lib->vba.VStartup[k] = dml_min( 2195 mode_lib->vba.VStartupLines, 2196 mode_lib->vba.MaxVStartupLines[k]); 2197 if (mode_lib->vba.VStartupRequiredWhenNotEnoughTimeForDynamicMetadata 2198 != 0) { 2199 mode_lib->vba.VStartup[k] = 2200 mode_lib->vba.VStartupRequiredWhenNotEnoughTimeForDynamicMetadata; 2201 } 2202 } else { 2203 mode_lib->vba.VStartup[k] = 2204 dml_min( 2205 mode_lib->vba.VStartupLines, 2206 mode_lib->vba.MaxVStartupLines[mode_lib->vba.BlendingAndTiming[k]]); 2207 } 2208 } 2209 2210 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2211 2212 if (mode_lib->vba.PDEAndMetaPTEBytesFrame[k] == 0) 2213 mode_lib->vba.prefetch_vm_bw[k] = 0; 2214 else if (mode_lib->vba.DestinationLinesToRequestVMInVBlank[k] > 0) { 2215 mode_lib->vba.prefetch_vm_bw[k] = 2216 (double) mode_lib->vba.PDEAndMetaPTEBytesFrame[k] 2217 / (mode_lib->vba.DestinationLinesToRequestVMInVBlank[k] 2218 * mode_lib->vba.HTotal[k] 2219 / mode_lib->vba.PixelClock[k]); 2220 } else { 2221 mode_lib->vba.prefetch_vm_bw[k] = 0; 2222 prefetch_vm_bw_valid = false; 2223 } 2224 if (mode_lib->vba.MetaRowByte[k] + mode_lib->vba.PixelPTEBytesPerRow[k] 2225 == 0) 2226 mode_lib->vba.prefetch_row_bw[k] = 0; 2227 else if (mode_lib->vba.DestinationLinesToRequestRowInVBlank[k] > 0) { 2228 mode_lib->vba.prefetch_row_bw[k] = 2229 (double) (mode_lib->vba.MetaRowByte[k] 2230 + mode_lib->vba.PixelPTEBytesPerRow[k]) 2231 / (mode_lib->vba.DestinationLinesToRequestRowInVBlank[k] 2232 * mode_lib->vba.HTotal[k] 2233 / mode_lib->vba.PixelClock[k]); 2234 } else { 2235 mode_lib->vba.prefetch_row_bw[k] = 0; 2236 prefetch_row_bw_valid = false; 2237 } 2238 2239 MaxTotalRDBandwidth = 2240 MaxTotalRDBandwidth + mode_lib->vba.cursor_bw[k] 2241 + dml_max( 2242 mode_lib->vba.prefetch_vm_bw[k], 2243 dml_max( 2244 mode_lib->vba.prefetch_row_bw[k], 2245 dml_max( 2246 mode_lib->vba.ReadBandwidthPlaneLuma[k] 2247 + mode_lib->vba.ReadBandwidthPlaneChroma[k], 2248 mode_lib->vba.RequiredPrefetchPixDataBWLuma[k]) 2249 + mode_lib->vba.meta_row_bw[k] 2250 + mode_lib->vba.dpte_row_bw[k])); 2251 2252 if (mode_lib->vba.DestinationLinesForPrefetch[k] < 2) 2253 DestinationLineTimesForPrefetchLessThan2 = true; 2254 if (mode_lib->vba.VRatioPrefetchY[k] > 4 2255 || mode_lib->vba.VRatioPrefetchC[k] > 4) 2256 VRatioPrefetchMoreThan4 = true; 2257 } 2258 2259 if (MaxTotalRDBandwidth <= mode_lib->vba.ReturnBW && prefetch_vm_bw_valid 2260 && prefetch_row_bw_valid && !VRatioPrefetchMoreThan4 2261 && !DestinationLineTimesForPrefetchLessThan2) 2262 mode_lib->vba.PrefetchModeSupported = true; 2263 else { 2264 mode_lib->vba.PrefetchModeSupported = false; 2265 dml_print( 2266 "DML: CalculatePrefetchSchedule ***failed***. Bandwidth violation. Results are NOT valid\n"); 2267 } 2268 2269 if (mode_lib->vba.PrefetchModeSupported == true) { 2270 double final_flip_bw[DC__NUM_DPP__MAX]; 2271 unsigned int ImmediateFlipBytes[DC__NUM_DPP__MAX]; 2272 double total_dcn_read_bw_with_flip = 0; 2273 2274 mode_lib->vba.BandwidthAvailableForImmediateFlip = mode_lib->vba.ReturnBW; 2275 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2276 mode_lib->vba.BandwidthAvailableForImmediateFlip = 2277 mode_lib->vba.BandwidthAvailableForImmediateFlip 2278 - mode_lib->vba.cursor_bw[k] 2279 - dml_max( 2280 mode_lib->vba.ReadBandwidthPlaneLuma[k] 2281 + mode_lib->vba.ReadBandwidthPlaneChroma[k] 2282 + mode_lib->vba.qual_row_bw[k], 2283 mode_lib->vba.PrefetchBandwidth[k]); 2284 } 2285 2286 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2287 ImmediateFlipBytes[k] = 0; 2288 if ((mode_lib->vba.SourcePixelFormat[k] != dm_420_8 2289 && mode_lib->vba.SourcePixelFormat[k] != dm_420_10)) { 2290 ImmediateFlipBytes[k] = 2291 mode_lib->vba.PDEAndMetaPTEBytesFrame[k] 2292 + mode_lib->vba.MetaRowByte[k] 2293 + mode_lib->vba.PixelPTEBytesPerRow[k]; 2294 } 2295 } 2296 mode_lib->vba.TotImmediateFlipBytes = 0; 2297 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2298 if ((mode_lib->vba.SourcePixelFormat[k] != dm_420_8 2299 && mode_lib->vba.SourcePixelFormat[k] != dm_420_10)) { 2300 mode_lib->vba.TotImmediateFlipBytes = 2301 mode_lib->vba.TotImmediateFlipBytes 2302 + ImmediateFlipBytes[k]; 2303 } 2304 } 2305 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2306 CalculateFlipSchedule( 2307 mode_lib, 2308 mode_lib->vba.UrgentExtraLatency, 2309 mode_lib->vba.UrgentLatencyPixelDataOnly, 2310 mode_lib->vba.GPUVMMaxPageTableLevels, 2311 mode_lib->vba.GPUVMEnable, 2312 mode_lib->vba.BandwidthAvailableForImmediateFlip, 2313 mode_lib->vba.TotImmediateFlipBytes, 2314 mode_lib->vba.SourcePixelFormat[k], 2315 ImmediateFlipBytes[k], 2316 mode_lib->vba.HTotal[k] 2317 / mode_lib->vba.PixelClock[k], 2318 mode_lib->vba.VRatio[k], 2319 mode_lib->vba.Tno_bw[k], 2320 mode_lib->vba.PDEAndMetaPTEBytesFrame[k], 2321 mode_lib->vba.MetaRowByte[k], 2322 mode_lib->vba.PixelPTEBytesPerRow[k], 2323 mode_lib->vba.DCCEnable[k], 2324 mode_lib->vba.dpte_row_height[k], 2325 mode_lib->vba.meta_row_height[k], 2326 mode_lib->vba.qual_row_bw[k], 2327 &mode_lib->vba.DestinationLinesToRequestVMInImmediateFlip[k], 2328 &mode_lib->vba.DestinationLinesToRequestRowInImmediateFlip[k], 2329 &final_flip_bw[k], 2330 &mode_lib->vba.ImmediateFlipSupportedForPipe[k]); 2331 } 2332 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2333 total_dcn_read_bw_with_flip = 2334 total_dcn_read_bw_with_flip 2335 + mode_lib->vba.cursor_bw[k] 2336 + dml_max( 2337 mode_lib->vba.prefetch_vm_bw[k], 2338 dml_max( 2339 mode_lib->vba.prefetch_row_bw[k], 2340 final_flip_bw[k] 2341 + dml_max( 2342 mode_lib->vba.ReadBandwidthPlaneLuma[k] 2343 + mode_lib->vba.ReadBandwidthPlaneChroma[k], 2344 mode_lib->vba.RequiredPrefetchPixDataBWLuma[k]))); 2345 } 2346 mode_lib->vba.ImmediateFlipSupported = true; 2347 if (total_dcn_read_bw_with_flip > mode_lib->vba.ReturnBW) { 2348 mode_lib->vba.ImmediateFlipSupported = false; 2349 } 2350 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2351 if (mode_lib->vba.ImmediateFlipSupportedForPipe[k] == false) { 2352 mode_lib->vba.ImmediateFlipSupported = false; 2353 } 2354 } 2355 } else { 2356 mode_lib->vba.ImmediateFlipSupported = false; 2357 } 2358 2359 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2360 if (mode_lib->vba.ErrorResult[k]) { 2361 mode_lib->vba.PrefetchModeSupported = false; 2362 dml_print( 2363 "DML: CalculatePrefetchSchedule ***failed***. Prefetch schedule violation. Results are NOT valid\n"); 2364 } 2365 } 2366 2367 mode_lib->vba.VStartupLines = mode_lib->vba.VStartupLines + 1; 2368 } while (!((mode_lib->vba.PrefetchModeSupported 2369 && (!mode_lib->vba.ImmediateFlipSupport 2370 || mode_lib->vba.ImmediateFlipSupported)) 2371 || mode_lib->vba.MaximumMaxVStartupLines < mode_lib->vba.VStartupLines)); 2372 2373 //Display Pipeline Delivery Time in Prefetch 2374 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2375 if (mode_lib->vba.VRatioPrefetchY[k] <= 1) { 2376 mode_lib->vba.DisplayPipeLineDeliveryTimeLumaPrefetch[k] = 2377 mode_lib->vba.SwathWidthY[k] * mode_lib->vba.DPPPerPlane[k] 2378 / mode_lib->vba.HRatio[k] 2379 / mode_lib->vba.PixelClock[k]; 2380 } else { 2381 mode_lib->vba.DisplayPipeLineDeliveryTimeLumaPrefetch[k] = 2382 mode_lib->vba.SwathWidthY[k] 2383 / mode_lib->vba.PSCL_THROUGHPUT_LUMA[k] 2384 / mode_lib->vba.DPPCLK[k]; 2385 } 2386 if (mode_lib->vba.BytePerPixelDETC[k] == 0) { 2387 mode_lib->vba.DisplayPipeLineDeliveryTimeChromaPrefetch[k] = 0; 2388 } else { 2389 if (mode_lib->vba.VRatioPrefetchC[k] <= 1) { 2390 mode_lib->vba.DisplayPipeLineDeliveryTimeChromaPrefetch[k] = 2391 mode_lib->vba.SwathWidthY[k] 2392 * mode_lib->vba.DPPPerPlane[k] 2393 / mode_lib->vba.HRatio[k] 2394 / mode_lib->vba.PixelClock[k]; 2395 } else { 2396 mode_lib->vba.DisplayPipeLineDeliveryTimeChromaPrefetch[k] = 2397 mode_lib->vba.SwathWidthY[k] 2398 / mode_lib->vba.PSCL_THROUGHPUT_LUMA[k] 2399 / mode_lib->vba.DPPCLK[k]; 2400 } 2401 } 2402 } 2403 2404 // Min TTUVBlank 2405 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2406 if (mode_lib->vba.PrefetchMode[mode_lib->vba.VoltageLevel][mode_lib->vba.maxMpcComb] == 0) { 2407 mode_lib->vba.AllowDRAMClockChangeDuringVBlank[k] = true; 2408 mode_lib->vba.AllowDRAMSelfRefreshDuringVBlank[k] = true; 2409 mode_lib->vba.MinTTUVBlank[k] = dml_max( 2410 mode_lib->vba.DRAMClockChangeWatermark, 2411 dml_max( 2412 mode_lib->vba.StutterEnterPlusExitWatermark, 2413 mode_lib->vba.UrgentWatermark)); 2414 } else if (mode_lib->vba.PrefetchMode[mode_lib->vba.VoltageLevel][mode_lib->vba.maxMpcComb] == 1) { 2415 mode_lib->vba.AllowDRAMClockChangeDuringVBlank[k] = false; 2416 mode_lib->vba.AllowDRAMSelfRefreshDuringVBlank[k] = true; 2417 mode_lib->vba.MinTTUVBlank[k] = dml_max( 2418 mode_lib->vba.StutterEnterPlusExitWatermark, 2419 mode_lib->vba.UrgentWatermark); 2420 } else { 2421 mode_lib->vba.AllowDRAMClockChangeDuringVBlank[k] = false; 2422 mode_lib->vba.AllowDRAMSelfRefreshDuringVBlank[k] = false; 2423 mode_lib->vba.MinTTUVBlank[k] = mode_lib->vba.UrgentWatermark; 2424 } 2425 if (!mode_lib->vba.DynamicMetadataEnable[k]) 2426 mode_lib->vba.MinTTUVBlank[k] = mode_lib->vba.TCalc 2427 + mode_lib->vba.MinTTUVBlank[k]; 2428 } 2429 2430 // DCC Configuration 2431 mode_lib->vba.ActiveDPPs = 0; 2432 // NB P-State/DRAM Clock Change Support 2433 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2434 mode_lib->vba.ActiveDPPs = mode_lib->vba.ActiveDPPs + mode_lib->vba.DPPPerPlane[k]; 2435 } 2436 2437 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2438 double EffectiveLBLatencyHidingY; 2439 double EffectiveLBLatencyHidingC; 2440 double DPPOutputBufferLinesY; 2441 double DPPOutputBufferLinesC; 2442 double DPPOPPBufferingY; 2443 double MaxDETBufferingTimeY; 2444 double ActiveDRAMClockChangeLatencyMarginY; 2445 2446 mode_lib->vba.LBLatencyHidingSourceLinesY = 2447 dml_min( 2448 mode_lib->vba.MaxLineBufferLines, 2449 (unsigned int) dml_floor( 2450 (double) mode_lib->vba.LineBufferSize 2451 / mode_lib->vba.LBBitPerPixel[k] 2452 / (mode_lib->vba.SwathWidthY[k] 2453 / dml_max( 2454 mode_lib->vba.HRatio[k], 2455 1.0)), 2456 1)) - (mode_lib->vba.vtaps[k] - 1); 2457 2458 mode_lib->vba.LBLatencyHidingSourceLinesC = 2459 dml_min( 2460 mode_lib->vba.MaxLineBufferLines, 2461 (unsigned int) dml_floor( 2462 (double) mode_lib->vba.LineBufferSize 2463 / mode_lib->vba.LBBitPerPixel[k] 2464 / (mode_lib->vba.SwathWidthY[k] 2465 / 2.0 2466 / dml_max( 2467 mode_lib->vba.HRatio[k] 2468 / 2, 2469 1.0)), 2470 1)) 2471 - (mode_lib->vba.VTAPsChroma[k] - 1); 2472 2473 EffectiveLBLatencyHidingY = mode_lib->vba.LBLatencyHidingSourceLinesY 2474 / mode_lib->vba.VRatio[k] 2475 * (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]); 2476 2477 EffectiveLBLatencyHidingC = mode_lib->vba.LBLatencyHidingSourceLinesC 2478 / (mode_lib->vba.VRatio[k] / 2) 2479 * (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]); 2480 2481 if (mode_lib->vba.SwathWidthY[k] > 2 * mode_lib->vba.DPPOutputBufferPixels) { 2482 DPPOutputBufferLinesY = mode_lib->vba.DPPOutputBufferPixels 2483 / mode_lib->vba.SwathWidthY[k]; 2484 } else if (mode_lib->vba.SwathWidthY[k] > mode_lib->vba.DPPOutputBufferPixels) { 2485 DPPOutputBufferLinesY = 0.5; 2486 } else { 2487 DPPOutputBufferLinesY = 1; 2488 } 2489 2490 if (mode_lib->vba.SwathWidthY[k] / 2 > 2 * mode_lib->vba.DPPOutputBufferPixels) { 2491 DPPOutputBufferLinesC = mode_lib->vba.DPPOutputBufferPixels 2492 / (mode_lib->vba.SwathWidthY[k] / 2); 2493 } else if (mode_lib->vba.SwathWidthY[k] / 2 > mode_lib->vba.DPPOutputBufferPixels) { 2494 DPPOutputBufferLinesC = 0.5; 2495 } else { 2496 DPPOutputBufferLinesC = 1; 2497 } 2498 2499 DPPOPPBufferingY = (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]) 2500 * (DPPOutputBufferLinesY + mode_lib->vba.OPPOutputBufferLines); 2501 MaxDETBufferingTimeY = mode_lib->vba.FullDETBufferingTimeY[k] 2502 + (mode_lib->vba.LinesInDETY[k] 2503 - mode_lib->vba.LinesInDETYRoundedDownToSwath[k]) 2504 / mode_lib->vba.SwathHeightY[k] 2505 * (mode_lib->vba.HTotal[k] 2506 / mode_lib->vba.PixelClock[k]); 2507 2508 ActiveDRAMClockChangeLatencyMarginY = DPPOPPBufferingY + EffectiveLBLatencyHidingY 2509 + MaxDETBufferingTimeY - mode_lib->vba.DRAMClockChangeWatermark; 2510 2511 if (mode_lib->vba.ActiveDPPs > 1) { 2512 ActiveDRAMClockChangeLatencyMarginY = 2513 ActiveDRAMClockChangeLatencyMarginY 2514 - (1 - 1 / (mode_lib->vba.ActiveDPPs - 1)) 2515 * mode_lib->vba.SwathHeightY[k] 2516 * (mode_lib->vba.HTotal[k] 2517 / mode_lib->vba.PixelClock[k]); 2518 } 2519 2520 if (mode_lib->vba.BytePerPixelDETC[k] > 0) { 2521 double DPPOPPBufferingC = (mode_lib->vba.HTotal[k] 2522 / mode_lib->vba.PixelClock[k]) 2523 * (DPPOutputBufferLinesC 2524 + mode_lib->vba.OPPOutputBufferLines); 2525 double MaxDETBufferingTimeC = 2526 mode_lib->vba.FullDETBufferingTimeC[k] 2527 + (mode_lib->vba.LinesInDETC[k] 2528 - mode_lib->vba.LinesInDETCRoundedDownToSwath[k]) 2529 / mode_lib->vba.SwathHeightC[k] 2530 * (mode_lib->vba.HTotal[k] 2531 / mode_lib->vba.PixelClock[k]); 2532 double ActiveDRAMClockChangeLatencyMarginC = DPPOPPBufferingC 2533 + EffectiveLBLatencyHidingC + MaxDETBufferingTimeC 2534 - mode_lib->vba.DRAMClockChangeWatermark; 2535 2536 if (mode_lib->vba.ActiveDPPs > 1) { 2537 ActiveDRAMClockChangeLatencyMarginC = 2538 ActiveDRAMClockChangeLatencyMarginC 2539 - (1 2540 - 1 2541 / (mode_lib->vba.ActiveDPPs 2542 - 1)) 2543 * mode_lib->vba.SwathHeightC[k] 2544 * (mode_lib->vba.HTotal[k] 2545 / mode_lib->vba.PixelClock[k]); 2546 } 2547 mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k] = dml_min( 2548 ActiveDRAMClockChangeLatencyMarginY, 2549 ActiveDRAMClockChangeLatencyMarginC); 2550 } else { 2551 mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k] = 2552 ActiveDRAMClockChangeLatencyMarginY; 2553 } 2554 2555 if (mode_lib->vba.WritebackEnable[k]) { 2556 double WritebackDRAMClockChangeLatencyMargin; 2557 2558 if (mode_lib->vba.WritebackPixelFormat[k] == dm_444_32) { 2559 WritebackDRAMClockChangeLatencyMargin = 2560 (double) (mode_lib->vba.WritebackInterfaceLumaBufferSize 2561 + mode_lib->vba.WritebackInterfaceChromaBufferSize) 2562 / (mode_lib->vba.WritebackDestinationWidth[k] 2563 * mode_lib->vba.WritebackDestinationHeight[k] 2564 / (mode_lib->vba.WritebackSourceHeight[k] 2565 * mode_lib->vba.HTotal[k] 2566 / mode_lib->vba.PixelClock[k]) 2567 * 4) 2568 - mode_lib->vba.WritebackDRAMClockChangeWatermark; 2569 } else if (mode_lib->vba.WritebackPixelFormat[k] == dm_420_10) { 2570 WritebackDRAMClockChangeLatencyMargin = 2571 dml_min( 2572 (double) mode_lib->vba.WritebackInterfaceLumaBufferSize 2573 * 8.0 / 10, 2574 2.0 2575 * mode_lib->vba.WritebackInterfaceChromaBufferSize 2576 * 8 / 10) 2577 / (mode_lib->vba.WritebackDestinationWidth[k] 2578 * mode_lib->vba.WritebackDestinationHeight[k] 2579 / (mode_lib->vba.WritebackSourceHeight[k] 2580 * mode_lib->vba.HTotal[k] 2581 / mode_lib->vba.PixelClock[k])) 2582 - mode_lib->vba.WritebackDRAMClockChangeWatermark; 2583 } else { 2584 WritebackDRAMClockChangeLatencyMargin = 2585 dml_min( 2586 (double) mode_lib->vba.WritebackInterfaceLumaBufferSize, 2587 2.0 2588 * mode_lib->vba.WritebackInterfaceChromaBufferSize) 2589 / (mode_lib->vba.WritebackDestinationWidth[k] 2590 * mode_lib->vba.WritebackDestinationHeight[k] 2591 / (mode_lib->vba.WritebackSourceHeight[k] 2592 * mode_lib->vba.HTotal[k] 2593 / mode_lib->vba.PixelClock[k])) 2594 - mode_lib->vba.WritebackDRAMClockChangeWatermark; 2595 } 2596 mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k] = dml_min( 2597 mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k], 2598 WritebackDRAMClockChangeLatencyMargin); 2599 } 2600 } 2601 2602 { 2603 float SecondMinActiveDRAMClockChangeMarginOneDisplayInVBLank = 999999; 2604 int PlaneWithMinActiveDRAMClockChangeMargin = -1; 2605 2606 mode_lib->vba.MinActiveDRAMClockChangeMargin = 999999; 2607 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2608 if (mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k] 2609 < mode_lib->vba.MinActiveDRAMClockChangeMargin) { 2610 mode_lib->vba.MinActiveDRAMClockChangeMargin = 2611 mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k]; 2612 if (mode_lib->vba.BlendingAndTiming[k] == k) { 2613 PlaneWithMinActiveDRAMClockChangeMargin = k; 2614 } else { 2615 for (j = 0; j < mode_lib->vba.NumberOfActivePlanes; ++j) { 2616 if (mode_lib->vba.BlendingAndTiming[k] == j) { 2617 PlaneWithMinActiveDRAMClockChangeMargin = j; 2618 } 2619 } 2620 } 2621 } 2622 } 2623 2624 mode_lib->vba.MinActiveDRAMClockChangeLatencySupported = 2625 mode_lib->vba.MinActiveDRAMClockChangeMargin 2626 + mode_lib->vba.DRAMClockChangeLatency; 2627 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2628 if (!((k == PlaneWithMinActiveDRAMClockChangeMargin) && (mode_lib->vba.BlendingAndTiming[k] == k)) 2629 && !(mode_lib->vba.BlendingAndTiming[k] == PlaneWithMinActiveDRAMClockChangeMargin) 2630 && mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k] 2631 < SecondMinActiveDRAMClockChangeMarginOneDisplayInVBLank) { 2632 SecondMinActiveDRAMClockChangeMarginOneDisplayInVBLank = 2633 mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k]; 2634 } 2635 } 2636 2637 if (mode_lib->vba.DRAMClockChangeSupportsVActive && 2638 mode_lib->vba.MinActiveDRAMClockChangeMargin > 60) { 2639 mode_lib->vba.DRAMClockChangeWatermark += 25; 2640 2641 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2642 if (mode_lib->vba.PrefetchMode[mode_lib->vba.VoltageLevel][mode_lib->vba.maxMpcComb] == 0) { 2643 if (mode_lib->vba.DRAMClockChangeWatermark > 2644 dml_max(mode_lib->vba.StutterEnterPlusExitWatermark, mode_lib->vba.UrgentWatermark)) 2645 mode_lib->vba.MinTTUVBlank[k] += 25; 2646 } 2647 } 2648 2649 mode_lib->vba.DRAMClockChangeSupport[0][0] = dm_dram_clock_change_vactive; 2650 } else if (mode_lib->vba.DummyPStateCheck && 2651 mode_lib->vba.MinActiveDRAMClockChangeMargin > 0) { 2652 mode_lib->vba.DRAMClockChangeSupport[0][0] = dm_dram_clock_change_vactive; 2653 } else { 2654 if ((mode_lib->vba.SynchronizedVBlank 2655 || mode_lib->vba.NumberOfActivePlanes == 1 2656 || (SecondMinActiveDRAMClockChangeMarginOneDisplayInVBLank > 0 && 2657 mode_lib->vba.AllowDramClockChangeOneDisplayVactive)) 2658 && mode_lib->vba.PrefetchMode[mode_lib->vba.VoltageLevel][mode_lib->vba.maxMpcComb] == 0) { 2659 mode_lib->vba.DRAMClockChangeSupport[0][0] = dm_dram_clock_change_vblank; 2660 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2661 if (!mode_lib->vba.AllowDRAMClockChangeDuringVBlank[k]) { 2662 mode_lib->vba.DRAMClockChangeSupport[0][0] = 2663 dm_dram_clock_change_unsupported; 2664 } 2665 } 2666 } else { 2667 mode_lib->vba.DRAMClockChangeSupport[0][0] = dm_dram_clock_change_unsupported; 2668 } 2669 } 2670 } 2671 for (k = 0; k <= mode_lib->vba.soc.num_states; k++) 2672 for (j = 0; j < 2; j++) 2673 mode_lib->vba.DRAMClockChangeSupport[k][j] = mode_lib->vba.DRAMClockChangeSupport[0][0]; 2674 2675 //XFC Parameters: 2676 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2677 if (mode_lib->vba.XFCEnabled[k] == true) { 2678 double TWait; 2679 2680 mode_lib->vba.XFCSlaveVUpdateOffset[k] = mode_lib->vba.XFCTSlvVupdateOffset; 2681 mode_lib->vba.XFCSlaveVupdateWidth[k] = mode_lib->vba.XFCTSlvVupdateWidth; 2682 mode_lib->vba.XFCSlaveVReadyOffset[k] = mode_lib->vba.XFCTSlvVreadyOffset; 2683 TWait = CalculateTWait( 2684 mode_lib->vba.PrefetchMode[mode_lib->vba.VoltageLevel][mode_lib->vba.maxMpcComb], 2685 mode_lib->vba.DRAMClockChangeLatency, 2686 mode_lib->vba.UrgentLatencyPixelDataOnly, 2687 mode_lib->vba.SREnterPlusExitTime); 2688 mode_lib->vba.XFCRemoteSurfaceFlipDelay = CalculateRemoteSurfaceFlipDelay( 2689 mode_lib, 2690 mode_lib->vba.VRatio[k], 2691 mode_lib->vba.SwathWidthY[k], 2692 dml_ceil(mode_lib->vba.BytePerPixelDETY[k], 1), 2693 mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k], 2694 mode_lib->vba.XFCTSlvVupdateOffset, 2695 mode_lib->vba.XFCTSlvVupdateWidth, 2696 mode_lib->vba.XFCTSlvVreadyOffset, 2697 mode_lib->vba.XFCXBUFLatencyTolerance, 2698 mode_lib->vba.XFCFillBWOverhead, 2699 mode_lib->vba.XFCSlvChunkSize, 2700 mode_lib->vba.XFCBusTransportTime, 2701 mode_lib->vba.TCalc, 2702 TWait, 2703 &mode_lib->vba.SrcActiveDrainRate, 2704 &mode_lib->vba.TInitXFill, 2705 &mode_lib->vba.TslvChk); 2706 mode_lib->vba.XFCRemoteSurfaceFlipLatency[k] = 2707 dml_floor( 2708 mode_lib->vba.XFCRemoteSurfaceFlipDelay 2709 / (mode_lib->vba.HTotal[k] 2710 / mode_lib->vba.PixelClock[k]), 2711 1); 2712 mode_lib->vba.XFCTransferDelay[k] = 2713 dml_ceil( 2714 mode_lib->vba.XFCBusTransportTime 2715 / (mode_lib->vba.HTotal[k] 2716 / mode_lib->vba.PixelClock[k]), 2717 1); 2718 mode_lib->vba.XFCPrechargeDelay[k] = 2719 dml_ceil( 2720 (mode_lib->vba.XFCBusTransportTime 2721 + mode_lib->vba.TInitXFill 2722 + mode_lib->vba.TslvChk) 2723 / (mode_lib->vba.HTotal[k] 2724 / mode_lib->vba.PixelClock[k]), 2725 1); 2726 mode_lib->vba.InitFillLevel = mode_lib->vba.XFCXBUFLatencyTolerance 2727 * mode_lib->vba.SrcActiveDrainRate; 2728 mode_lib->vba.FinalFillMargin = 2729 (mode_lib->vba.DestinationLinesToRequestVMInVBlank[k] 2730 + mode_lib->vba.DestinationLinesToRequestRowInVBlank[k]) 2731 * mode_lib->vba.HTotal[k] 2732 / mode_lib->vba.PixelClock[k] 2733 * mode_lib->vba.SrcActiveDrainRate 2734 + mode_lib->vba.XFCFillConstant; 2735 mode_lib->vba.FinalFillLevel = mode_lib->vba.XFCRemoteSurfaceFlipDelay 2736 * mode_lib->vba.SrcActiveDrainRate 2737 + mode_lib->vba.FinalFillMargin; 2738 mode_lib->vba.RemainingFillLevel = dml_max( 2739 0.0, 2740 mode_lib->vba.FinalFillLevel - mode_lib->vba.InitFillLevel); 2741 mode_lib->vba.TFinalxFill = mode_lib->vba.RemainingFillLevel 2742 / (mode_lib->vba.SrcActiveDrainRate 2743 * mode_lib->vba.XFCFillBWOverhead / 100); 2744 mode_lib->vba.XFCPrefetchMargin[k] = 2745 mode_lib->vba.XFCRemoteSurfaceFlipDelay 2746 + mode_lib->vba.TFinalxFill 2747 + (mode_lib->vba.DestinationLinesToRequestVMInVBlank[k] 2748 + mode_lib->vba.DestinationLinesToRequestRowInVBlank[k]) 2749 * mode_lib->vba.HTotal[k] 2750 / mode_lib->vba.PixelClock[k]; 2751 } else { 2752 mode_lib->vba.XFCSlaveVUpdateOffset[k] = 0; 2753 mode_lib->vba.XFCSlaveVupdateWidth[k] = 0; 2754 mode_lib->vba.XFCSlaveVReadyOffset[k] = 0; 2755 mode_lib->vba.XFCRemoteSurfaceFlipLatency[k] = 0; 2756 mode_lib->vba.XFCPrechargeDelay[k] = 0; 2757 mode_lib->vba.XFCTransferDelay[k] = 0; 2758 mode_lib->vba.XFCPrefetchMargin[k] = 0; 2759 } 2760 } 2761 { 2762 unsigned int VStartupMargin = 0; 2763 bool FirstMainPlane = true; 2764 2765 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2766 if (mode_lib->vba.BlendingAndTiming[k] == k) { 2767 unsigned int Margin = (mode_lib->vba.MaxVStartupLines[k] - mode_lib->vba.VStartup[k]) 2768 * mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]; 2769 2770 if (FirstMainPlane) { 2771 VStartupMargin = Margin; 2772 FirstMainPlane = false; 2773 } else 2774 VStartupMargin = dml_min(VStartupMargin, Margin); 2775 } 2776 2777 if (mode_lib->vba.UseMaximumVStartup) { 2778 if (mode_lib->vba.VTotal_Max[k] == mode_lib->vba.VTotal[k]) { 2779 //only use max vstart if it is not drr or lateflip. 2780 mode_lib->vba.VStartup[k] = mode_lib->vba.MaxVStartupLines[mode_lib->vba.BlendingAndTiming[k]]; 2781 } 2782 } 2783 } 2784 } 2785 } 2786 2787 static void dml20v2_DisplayPipeConfiguration(struct display_mode_lib *mode_lib) 2788 { 2789 double BytePerPixDETY; 2790 double BytePerPixDETC; 2791 double Read256BytesBlockHeightY; 2792 double Read256BytesBlockHeightC; 2793 double Read256BytesBlockWidthY; 2794 double Read256BytesBlockWidthC; 2795 double MaximumSwathHeightY; 2796 double MaximumSwathHeightC; 2797 double MinimumSwathHeightY; 2798 double MinimumSwathHeightC; 2799 double SwathWidth; 2800 double SwathWidthGranularityY; 2801 double SwathWidthGranularityC; 2802 double RoundedUpMaxSwathSizeBytesY; 2803 double RoundedUpMaxSwathSizeBytesC; 2804 unsigned int j, k; 2805 2806 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2807 bool MainPlaneDoesODMCombine = false; 2808 2809 if (mode_lib->vba.SourcePixelFormat[k] == dm_444_64) { 2810 BytePerPixDETY = 8; 2811 BytePerPixDETC = 0; 2812 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_32) { 2813 BytePerPixDETY = 4; 2814 BytePerPixDETC = 0; 2815 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_16) { 2816 BytePerPixDETY = 2; 2817 BytePerPixDETC = 0; 2818 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_8) { 2819 BytePerPixDETY = 1; 2820 BytePerPixDETC = 0; 2821 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_8) { 2822 BytePerPixDETY = 1; 2823 BytePerPixDETC = 2; 2824 } else { 2825 BytePerPixDETY = 4.0 / 3.0; 2826 BytePerPixDETC = 8.0 / 3.0; 2827 } 2828 2829 if ((mode_lib->vba.SourcePixelFormat[k] == dm_444_64 2830 || mode_lib->vba.SourcePixelFormat[k] == dm_444_32 2831 || mode_lib->vba.SourcePixelFormat[k] == dm_444_16 2832 || mode_lib->vba.SourcePixelFormat[k] == dm_444_8)) { 2833 if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear) { 2834 Read256BytesBlockHeightY = 1; 2835 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_64) { 2836 Read256BytesBlockHeightY = 4; 2837 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_32 2838 || mode_lib->vba.SourcePixelFormat[k] == dm_444_16) { 2839 Read256BytesBlockHeightY = 8; 2840 } else { 2841 Read256BytesBlockHeightY = 16; 2842 } 2843 Read256BytesBlockWidthY = 256 / dml_ceil(BytePerPixDETY, 1) 2844 / Read256BytesBlockHeightY; 2845 Read256BytesBlockHeightC = 0; 2846 Read256BytesBlockWidthC = 0; 2847 } else { 2848 if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear) { 2849 Read256BytesBlockHeightY = 1; 2850 Read256BytesBlockHeightC = 1; 2851 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_8) { 2852 Read256BytesBlockHeightY = 16; 2853 Read256BytesBlockHeightC = 8; 2854 } else { 2855 Read256BytesBlockHeightY = 8; 2856 Read256BytesBlockHeightC = 8; 2857 } 2858 Read256BytesBlockWidthY = 256 / dml_ceil(BytePerPixDETY, 1) 2859 / Read256BytesBlockHeightY; 2860 Read256BytesBlockWidthC = 256 / dml_ceil(BytePerPixDETC, 2) 2861 / Read256BytesBlockHeightC; 2862 } 2863 2864 if (mode_lib->vba.SourceScan[k] == dm_horz) { 2865 MaximumSwathHeightY = Read256BytesBlockHeightY; 2866 MaximumSwathHeightC = Read256BytesBlockHeightC; 2867 } else { 2868 MaximumSwathHeightY = Read256BytesBlockWidthY; 2869 MaximumSwathHeightC = Read256BytesBlockWidthC; 2870 } 2871 2872 if ((mode_lib->vba.SourcePixelFormat[k] == dm_444_64 2873 || mode_lib->vba.SourcePixelFormat[k] == dm_444_32 2874 || mode_lib->vba.SourcePixelFormat[k] == dm_444_16 2875 || mode_lib->vba.SourcePixelFormat[k] == dm_444_8)) { 2876 if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear 2877 || (mode_lib->vba.SourcePixelFormat[k] == dm_444_64 2878 && (mode_lib->vba.SurfaceTiling[k] 2879 == dm_sw_4kb_s 2880 || mode_lib->vba.SurfaceTiling[k] 2881 == dm_sw_4kb_s_x 2882 || mode_lib->vba.SurfaceTiling[k] 2883 == dm_sw_64kb_s 2884 || mode_lib->vba.SurfaceTiling[k] 2885 == dm_sw_64kb_s_t 2886 || mode_lib->vba.SurfaceTiling[k] 2887 == dm_sw_64kb_s_x 2888 || mode_lib->vba.SurfaceTiling[k] 2889 == dm_sw_var_s 2890 || mode_lib->vba.SurfaceTiling[k] 2891 == dm_sw_var_s_x) 2892 && mode_lib->vba.SourceScan[k] == dm_horz)) { 2893 MinimumSwathHeightY = MaximumSwathHeightY; 2894 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_8 2895 && mode_lib->vba.SourceScan[k] != dm_horz) { 2896 MinimumSwathHeightY = MaximumSwathHeightY; 2897 } else { 2898 MinimumSwathHeightY = MaximumSwathHeightY / 2.0; 2899 } 2900 MinimumSwathHeightC = MaximumSwathHeightC; 2901 } else { 2902 if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear) { 2903 MinimumSwathHeightY = MaximumSwathHeightY; 2904 MinimumSwathHeightC = MaximumSwathHeightC; 2905 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_8 2906 && mode_lib->vba.SourceScan[k] == dm_horz) { 2907 MinimumSwathHeightY = MaximumSwathHeightY / 2.0; 2908 MinimumSwathHeightC = MaximumSwathHeightC; 2909 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_10 2910 && mode_lib->vba.SourceScan[k] == dm_horz) { 2911 MinimumSwathHeightC = MaximumSwathHeightC / 2.0; 2912 MinimumSwathHeightY = MaximumSwathHeightY; 2913 } else { 2914 MinimumSwathHeightY = MaximumSwathHeightY; 2915 MinimumSwathHeightC = MaximumSwathHeightC; 2916 } 2917 } 2918 2919 if (mode_lib->vba.SourceScan[k] == dm_horz) { 2920 SwathWidth = mode_lib->vba.ViewportWidth[k]; 2921 } else { 2922 SwathWidth = mode_lib->vba.ViewportHeight[k]; 2923 } 2924 2925 if (mode_lib->vba.ODMCombineEnabled[k] == dm_odm_combine_mode_2to1) { 2926 MainPlaneDoesODMCombine = true; 2927 } 2928 for (j = 0; j < mode_lib->vba.NumberOfActivePlanes; ++j) { 2929 if (mode_lib->vba.BlendingAndTiming[k] == j 2930 && mode_lib->vba.ODMCombineEnabled[k] == dm_odm_combine_mode_2to1) { 2931 MainPlaneDoesODMCombine = true; 2932 } 2933 } 2934 2935 if (MainPlaneDoesODMCombine == true) { 2936 SwathWidth = dml_min( 2937 SwathWidth, 2938 mode_lib->vba.HActive[k] / 2.0 * mode_lib->vba.HRatio[k]); 2939 } else { 2940 if (mode_lib->vba.DPPPerPlane[k] == 0) 2941 SwathWidth = 0; 2942 else 2943 SwathWidth = SwathWidth / mode_lib->vba.DPPPerPlane[k]; 2944 } 2945 2946 SwathWidthGranularityY = 256 / dml_ceil(BytePerPixDETY, 1) / MaximumSwathHeightY; 2947 RoundedUpMaxSwathSizeBytesY = (dml_ceil( 2948 (double) (SwathWidth - 1), 2949 SwathWidthGranularityY) + SwathWidthGranularityY) * BytePerPixDETY 2950 * MaximumSwathHeightY; 2951 if (mode_lib->vba.SourcePixelFormat[k] == dm_420_10) { 2952 RoundedUpMaxSwathSizeBytesY = dml_ceil(RoundedUpMaxSwathSizeBytesY, 256) 2953 + 256; 2954 } 2955 if (MaximumSwathHeightC > 0) { 2956 SwathWidthGranularityC = 256.0 / dml_ceil(BytePerPixDETC, 2) 2957 / MaximumSwathHeightC; 2958 RoundedUpMaxSwathSizeBytesC = (dml_ceil( 2959 (double) (SwathWidth / 2.0 - 1), 2960 SwathWidthGranularityC) + SwathWidthGranularityC) 2961 * BytePerPixDETC * MaximumSwathHeightC; 2962 if (mode_lib->vba.SourcePixelFormat[k] == dm_420_10) { 2963 RoundedUpMaxSwathSizeBytesC = dml_ceil( 2964 RoundedUpMaxSwathSizeBytesC, 2965 256) + 256; 2966 } 2967 } else 2968 RoundedUpMaxSwathSizeBytesC = 0.0; 2969 2970 if (RoundedUpMaxSwathSizeBytesY + RoundedUpMaxSwathSizeBytesC 2971 <= mode_lib->vba.DETBufferSizeInKByte[0] * 1024.0 / 2.0) { 2972 mode_lib->vba.SwathHeightY[k] = MaximumSwathHeightY; 2973 mode_lib->vba.SwathHeightC[k] = MaximumSwathHeightC; 2974 } else { 2975 mode_lib->vba.SwathHeightY[k] = MinimumSwathHeightY; 2976 mode_lib->vba.SwathHeightC[k] = MinimumSwathHeightC; 2977 } 2978 2979 if (mode_lib->vba.SwathHeightC[k] == 0) { 2980 mode_lib->vba.DETBufferSizeY[k] = mode_lib->vba.DETBufferSizeInKByte[0] * 1024; 2981 mode_lib->vba.DETBufferSizeC[k] = 0; 2982 } else if (mode_lib->vba.SwathHeightY[k] <= mode_lib->vba.SwathHeightC[k]) { 2983 mode_lib->vba.DETBufferSizeY[k] = mode_lib->vba.DETBufferSizeInKByte[0] 2984 * 1024.0 / 2; 2985 mode_lib->vba.DETBufferSizeC[k] = mode_lib->vba.DETBufferSizeInKByte[0] 2986 * 1024.0 / 2; 2987 } else { 2988 mode_lib->vba.DETBufferSizeY[k] = mode_lib->vba.DETBufferSizeInKByte[0] 2989 * 1024.0 * 2 / 3; 2990 mode_lib->vba.DETBufferSizeC[k] = mode_lib->vba.DETBufferSizeInKByte[0] 2991 * 1024.0 / 3; 2992 } 2993 } 2994 } 2995 2996 static double CalculateTWait( 2997 unsigned int PrefetchMode, 2998 double DRAMClockChangeLatency, 2999 double UrgentLatencyPixelDataOnly, 3000 double SREnterPlusExitTime) 3001 { 3002 if (PrefetchMode == 0) { 3003 return dml_max( 3004 DRAMClockChangeLatency + UrgentLatencyPixelDataOnly, 3005 dml_max(SREnterPlusExitTime, UrgentLatencyPixelDataOnly)); 3006 } else if (PrefetchMode == 1) { 3007 return dml_max(SREnterPlusExitTime, UrgentLatencyPixelDataOnly); 3008 } else { 3009 return UrgentLatencyPixelDataOnly; 3010 } 3011 } 3012 3013 static double CalculateRemoteSurfaceFlipDelay( 3014 struct display_mode_lib *mode_lib, 3015 double VRatio, 3016 double SwathWidth, 3017 double Bpp, 3018 double LineTime, 3019 double XFCTSlvVupdateOffset, 3020 double XFCTSlvVupdateWidth, 3021 double XFCTSlvVreadyOffset, 3022 double XFCXBUFLatencyTolerance, 3023 double XFCFillBWOverhead, 3024 double XFCSlvChunkSize, 3025 double XFCBusTransportTime, 3026 double TCalc, 3027 double TWait, 3028 double *SrcActiveDrainRate, 3029 double *TInitXFill, 3030 double *TslvChk) 3031 { 3032 double TSlvSetup, AvgfillRate, result; 3033 3034 *SrcActiveDrainRate = VRatio * SwathWidth * Bpp / LineTime; 3035 TSlvSetup = XFCTSlvVupdateOffset + XFCTSlvVupdateWidth + XFCTSlvVreadyOffset; 3036 *TInitXFill = XFCXBUFLatencyTolerance / (1 + XFCFillBWOverhead / 100); 3037 AvgfillRate = *SrcActiveDrainRate * (1 + XFCFillBWOverhead / 100); 3038 *TslvChk = XFCSlvChunkSize / AvgfillRate; 3039 dml_print( 3040 "DML::CalculateRemoteSurfaceFlipDelay: SrcActiveDrainRate: %f\n", 3041 *SrcActiveDrainRate); 3042 dml_print("DML::CalculateRemoteSurfaceFlipDelay: TSlvSetup: %f\n", TSlvSetup); 3043 dml_print("DML::CalculateRemoteSurfaceFlipDelay: TInitXFill: %f\n", *TInitXFill); 3044 dml_print("DML::CalculateRemoteSurfaceFlipDelay: AvgfillRate: %f\n", AvgfillRate); 3045 dml_print("DML::CalculateRemoteSurfaceFlipDelay: TslvChk: %f\n", *TslvChk); 3046 result = 2 * XFCBusTransportTime + TSlvSetup + TCalc + TWait + *TslvChk + *TInitXFill; // TODO: This doesn't seem to match programming guide 3047 dml_print("DML::CalculateRemoteSurfaceFlipDelay: RemoteSurfaceFlipDelay: %f\n", result); 3048 return result; 3049 } 3050 3051 static double CalculateWriteBackDelay( 3052 enum source_format_class WritebackPixelFormat, 3053 double WritebackHRatio, 3054 double WritebackVRatio, 3055 unsigned int WritebackLumaHTaps, 3056 unsigned int WritebackLumaVTaps, 3057 unsigned int WritebackChromaHTaps, 3058 unsigned int WritebackChromaVTaps, 3059 unsigned int WritebackDestinationWidth) 3060 { 3061 double CalculateWriteBackDelay = 3062 dml_max( 3063 dml_ceil(WritebackLumaHTaps / 4.0, 1) / WritebackHRatio, 3064 WritebackLumaVTaps * dml_ceil(1.0 / WritebackVRatio, 1) 3065 * dml_ceil( 3066 WritebackDestinationWidth 3067 / 4.0, 3068 1) 3069 + dml_ceil(1.0 / WritebackVRatio, 1) 3070 * (dml_ceil( 3071 WritebackLumaVTaps 3072 / 4.0, 3073 1) + 4)); 3074 3075 if (WritebackPixelFormat != dm_444_32) { 3076 CalculateWriteBackDelay = 3077 dml_max( 3078 CalculateWriteBackDelay, 3079 dml_max( 3080 dml_ceil( 3081 WritebackChromaHTaps 3082 / 2.0, 3083 1) 3084 / (2 3085 * WritebackHRatio), 3086 WritebackChromaVTaps 3087 * dml_ceil( 3088 1 3089 / (2 3090 * WritebackVRatio), 3091 1) 3092 * dml_ceil( 3093 WritebackDestinationWidth 3094 / 2.0 3095 / 2.0, 3096 1) 3097 + dml_ceil( 3098 1 3099 / (2 3100 * WritebackVRatio), 3101 1) 3102 * (dml_ceil( 3103 WritebackChromaVTaps 3104 / 4.0, 3105 1) 3106 + 4))); 3107 } 3108 return CalculateWriteBackDelay; 3109 } 3110 3111 static void CalculateActiveRowBandwidth( 3112 bool GPUVMEnable, 3113 enum source_format_class SourcePixelFormat, 3114 double VRatio, 3115 bool DCCEnable, 3116 double LineTime, 3117 unsigned int MetaRowByteLuma, 3118 unsigned int MetaRowByteChroma, 3119 unsigned int meta_row_height_luma, 3120 unsigned int meta_row_height_chroma, 3121 unsigned int PixelPTEBytesPerRowLuma, 3122 unsigned int PixelPTEBytesPerRowChroma, 3123 unsigned int dpte_row_height_luma, 3124 unsigned int dpte_row_height_chroma, 3125 double *meta_row_bw, 3126 double *dpte_row_bw, 3127 double *qual_row_bw) 3128 { 3129 if (DCCEnable != true) { 3130 *meta_row_bw = 0; 3131 } else if (SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10) { 3132 *meta_row_bw = VRatio * MetaRowByteLuma / (meta_row_height_luma * LineTime) 3133 + VRatio / 2 * MetaRowByteChroma 3134 / (meta_row_height_chroma * LineTime); 3135 } else { 3136 *meta_row_bw = VRatio * MetaRowByteLuma / (meta_row_height_luma * LineTime); 3137 } 3138 3139 if (GPUVMEnable != true) { 3140 *dpte_row_bw = 0; 3141 } else if (SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10) { 3142 *dpte_row_bw = VRatio * PixelPTEBytesPerRowLuma / (dpte_row_height_luma * LineTime) 3143 + VRatio / 2 * PixelPTEBytesPerRowChroma 3144 / (dpte_row_height_chroma * LineTime); 3145 } else { 3146 *dpte_row_bw = VRatio * PixelPTEBytesPerRowLuma / (dpte_row_height_luma * LineTime); 3147 } 3148 3149 if ((SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10)) { 3150 *qual_row_bw = *meta_row_bw + *dpte_row_bw; 3151 } else { 3152 *qual_row_bw = 0; 3153 } 3154 } 3155 3156 static void CalculateFlipSchedule( 3157 struct display_mode_lib *mode_lib, 3158 double UrgentExtraLatency, 3159 double UrgentLatencyPixelDataOnly, 3160 unsigned int GPUVMMaxPageTableLevels, 3161 bool GPUVMEnable, 3162 double BandwidthAvailableForImmediateFlip, 3163 unsigned int TotImmediateFlipBytes, 3164 enum source_format_class SourcePixelFormat, 3165 unsigned int ImmediateFlipBytes, 3166 double LineTime, 3167 double VRatio, 3168 double Tno_bw, 3169 double PDEAndMetaPTEBytesFrame, 3170 unsigned int MetaRowByte, 3171 unsigned int PixelPTEBytesPerRow, 3172 bool DCCEnable, 3173 unsigned int dpte_row_height, 3174 unsigned int meta_row_height, 3175 double qual_row_bw, 3176 double *DestinationLinesToRequestVMInImmediateFlip, 3177 double *DestinationLinesToRequestRowInImmediateFlip, 3178 double *final_flip_bw, 3179 bool *ImmediateFlipSupportedForPipe) 3180 { 3181 double min_row_time = 0.0; 3182 3183 if (SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10) { 3184 *DestinationLinesToRequestVMInImmediateFlip = 0.0; 3185 *DestinationLinesToRequestRowInImmediateFlip = 0.0; 3186 *final_flip_bw = qual_row_bw; 3187 *ImmediateFlipSupportedForPipe = true; 3188 } else { 3189 double TimeForFetchingMetaPTEImmediateFlip; 3190 double TimeForFetchingRowInVBlankImmediateFlip; 3191 3192 if (GPUVMEnable == true) { 3193 mode_lib->vba.ImmediateFlipBW[0] = BandwidthAvailableForImmediateFlip 3194 * ImmediateFlipBytes / TotImmediateFlipBytes; 3195 TimeForFetchingMetaPTEImmediateFlip = 3196 dml_max( 3197 Tno_bw 3198 + PDEAndMetaPTEBytesFrame 3199 / mode_lib->vba.ImmediateFlipBW[0], 3200 dml_max( 3201 UrgentExtraLatency 3202 + UrgentLatencyPixelDataOnly 3203 * (GPUVMMaxPageTableLevels 3204 - 1), 3205 LineTime / 4.0)); 3206 } else { 3207 TimeForFetchingMetaPTEImmediateFlip = 0; 3208 } 3209 3210 *DestinationLinesToRequestVMInImmediateFlip = dml_floor( 3211 4.0 * (TimeForFetchingMetaPTEImmediateFlip / LineTime + 0.125), 3212 1) / 4.0; 3213 3214 if ((GPUVMEnable == true || DCCEnable == true)) { 3215 mode_lib->vba.ImmediateFlipBW[0] = BandwidthAvailableForImmediateFlip 3216 * ImmediateFlipBytes / TotImmediateFlipBytes; 3217 TimeForFetchingRowInVBlankImmediateFlip = dml_max( 3218 (MetaRowByte + PixelPTEBytesPerRow) 3219 / mode_lib->vba.ImmediateFlipBW[0], 3220 dml_max(UrgentLatencyPixelDataOnly, LineTime / 4.0)); 3221 } else { 3222 TimeForFetchingRowInVBlankImmediateFlip = 0; 3223 } 3224 3225 *DestinationLinesToRequestRowInImmediateFlip = dml_floor( 3226 4.0 * (TimeForFetchingRowInVBlankImmediateFlip / LineTime + 0.125), 3227 1) / 4.0; 3228 3229 if (GPUVMEnable == true) { 3230 *final_flip_bw = 3231 dml_max( 3232 PDEAndMetaPTEBytesFrame 3233 / (*DestinationLinesToRequestVMInImmediateFlip 3234 * LineTime), 3235 (MetaRowByte + PixelPTEBytesPerRow) 3236 / (TimeForFetchingRowInVBlankImmediateFlip 3237 * LineTime)); 3238 } else if (MetaRowByte + PixelPTEBytesPerRow > 0) { 3239 *final_flip_bw = (MetaRowByte + PixelPTEBytesPerRow) 3240 / (TimeForFetchingRowInVBlankImmediateFlip * LineTime); 3241 } else { 3242 *final_flip_bw = 0; 3243 } 3244 3245 if (GPUVMEnable && !DCCEnable) 3246 min_row_time = dpte_row_height * LineTime / VRatio; 3247 else if (!GPUVMEnable && DCCEnable) 3248 min_row_time = meta_row_height * LineTime / VRatio; 3249 else 3250 min_row_time = dml_min(dpte_row_height, meta_row_height) * LineTime 3251 / VRatio; 3252 3253 if (*DestinationLinesToRequestVMInImmediateFlip >= 8 3254 || *DestinationLinesToRequestRowInImmediateFlip >= 16 3255 || TimeForFetchingMetaPTEImmediateFlip 3256 + 2 * TimeForFetchingRowInVBlankImmediateFlip 3257 > min_row_time) 3258 *ImmediateFlipSupportedForPipe = false; 3259 else 3260 *ImmediateFlipSupportedForPipe = true; 3261 } 3262 } 3263 3264 static unsigned int TruncToValidBPP( 3265 double DecimalBPP, 3266 double DesiredBPP, 3267 bool DSCEnabled, 3268 enum output_encoder_class Output, 3269 enum output_format_class Format, 3270 unsigned int DSCInputBitPerComponent) 3271 { 3272 if (Output == dm_hdmi) { 3273 if (Format == dm_420) { 3274 if (DecimalBPP >= 18 && (DesiredBPP == 0 || DesiredBPP == 18)) 3275 return 18; 3276 else if (DecimalBPP >= 15 && (DesiredBPP == 0 || DesiredBPP == 15)) 3277 return 15; 3278 else if (DecimalBPP >= 12 && (DesiredBPP == 0 || DesiredBPP == 12)) 3279 return 12; 3280 else 3281 return BPP_INVALID; 3282 } else if (Format == dm_444) { 3283 if (DecimalBPP >= 36 && (DesiredBPP == 0 || DesiredBPP == 36)) 3284 return 36; 3285 else if (DecimalBPP >= 30 && (DesiredBPP == 0 || DesiredBPP == 30)) 3286 return 30; 3287 else if (DecimalBPP >= 24 && (DesiredBPP == 0 || DesiredBPP == 24)) 3288 return 24; 3289 else if (DecimalBPP >= 18 && (DesiredBPP == 0 || DesiredBPP == 18)) 3290 return 18; 3291 else 3292 return BPP_INVALID; 3293 } else { 3294 if (DecimalBPP / 1.5 >= 24 && (DesiredBPP == 0 || DesiredBPP == 24)) 3295 return 24; 3296 else if (DecimalBPP / 1.5 >= 20 && (DesiredBPP == 0 || DesiredBPP == 20)) 3297 return 20; 3298 else if (DecimalBPP / 1.5 >= 16 && (DesiredBPP == 0 || DesiredBPP == 16)) 3299 return 16; 3300 else 3301 return BPP_INVALID; 3302 } 3303 } else { 3304 if (DSCEnabled) { 3305 if (Format == dm_420) { 3306 if (DesiredBPP == 0) { 3307 if (DecimalBPP < 6) 3308 return BPP_INVALID; 3309 else if (DecimalBPP >= 1.5 * DSCInputBitPerComponent - 1.0 / 16.0) 3310 return 1.5 * DSCInputBitPerComponent - 1.0 / 16.0; 3311 else 3312 return dml_floor(16 * DecimalBPP, 1) / 16.0; 3313 } else { 3314 if (DecimalBPP < 6 3315 || DesiredBPP < 6 3316 || DesiredBPP > 1.5 * DSCInputBitPerComponent - 1.0 / 16.0 3317 || DecimalBPP < DesiredBPP) { 3318 return BPP_INVALID; 3319 } else { 3320 return DesiredBPP; 3321 } 3322 } 3323 } else if (Format == dm_n422) { 3324 if (DesiredBPP == 0) { 3325 if (DecimalBPP < 7) 3326 return BPP_INVALID; 3327 else if (DecimalBPP >= 2 * DSCInputBitPerComponent - 1.0 / 16.0) 3328 return 2 * DSCInputBitPerComponent - 1.0 / 16.0; 3329 else 3330 return dml_floor(16 * DecimalBPP, 1) / 16.0; 3331 } else { 3332 if (DecimalBPP < 7 3333 || DesiredBPP < 7 3334 || DesiredBPP > 2 * DSCInputBitPerComponent - 1.0 / 16.0 3335 || DecimalBPP < DesiredBPP) { 3336 return BPP_INVALID; 3337 } else { 3338 return DesiredBPP; 3339 } 3340 } 3341 } else { 3342 if (DesiredBPP == 0) { 3343 if (DecimalBPP < 8) 3344 return BPP_INVALID; 3345 else if (DecimalBPP >= 3 * DSCInputBitPerComponent - 1.0 / 16.0) 3346 return 3 * DSCInputBitPerComponent - 1.0 / 16.0; 3347 else 3348 return dml_floor(16 * DecimalBPP, 1) / 16.0; 3349 } else { 3350 if (DecimalBPP < 8 3351 || DesiredBPP < 8 3352 || DesiredBPP > 3 * DSCInputBitPerComponent - 1.0 / 16.0 3353 || DecimalBPP < DesiredBPP) { 3354 return BPP_INVALID; 3355 } else { 3356 return DesiredBPP; 3357 } 3358 } 3359 } 3360 } else if (Format == dm_420) { 3361 if (DecimalBPP >= 18 && (DesiredBPP == 0 || DesiredBPP == 18)) 3362 return 18; 3363 else if (DecimalBPP >= 15 && (DesiredBPP == 0 || DesiredBPP == 15)) 3364 return 15; 3365 else if (DecimalBPP >= 12 && (DesiredBPP == 0 || DesiredBPP == 12)) 3366 return 12; 3367 else 3368 return BPP_INVALID; 3369 } else if (Format == dm_s422 || Format == dm_n422) { 3370 if (DecimalBPP >= 24 && (DesiredBPP == 0 || DesiredBPP == 24)) 3371 return 24; 3372 else if (DecimalBPP >= 20 && (DesiredBPP == 0 || DesiredBPP == 20)) 3373 return 20; 3374 else if (DecimalBPP >= 16 && (DesiredBPP == 0 || DesiredBPP == 16)) 3375 return 16; 3376 else 3377 return BPP_INVALID; 3378 } else { 3379 if (DecimalBPP >= 36 && (DesiredBPP == 0 || DesiredBPP == 36)) 3380 return 36; 3381 else if (DecimalBPP >= 30 && (DesiredBPP == 0 || DesiredBPP == 30)) 3382 return 30; 3383 else if (DecimalBPP >= 24 && (DesiredBPP == 0 || DesiredBPP == 24)) 3384 return 24; 3385 else if (DecimalBPP >= 18 && (DesiredBPP == 0 || DesiredBPP == 18)) 3386 return 18; 3387 else 3388 return BPP_INVALID; 3389 } 3390 } 3391 } 3392 3393 void dml20v2_ModeSupportAndSystemConfigurationFull(struct display_mode_lib *mode_lib) 3394 { 3395 struct vba_vars_st *locals = &mode_lib->vba; 3396 3397 int i; 3398 unsigned int j, k, m; 3399 3400 /*MODE SUPPORT, VOLTAGE STATE AND SOC CONFIGURATION*/ 3401 3402 /*Scale Ratio, taps Support Check*/ 3403 3404 mode_lib->vba.ScaleRatioAndTapsSupport = true; 3405 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3406 if (mode_lib->vba.ScalerEnabled[k] == false 3407 && ((mode_lib->vba.SourcePixelFormat[k] != dm_444_64 3408 && mode_lib->vba.SourcePixelFormat[k] != dm_444_32 3409 && mode_lib->vba.SourcePixelFormat[k] != dm_444_16 3410 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_16 3411 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_8) 3412 || mode_lib->vba.HRatio[k] != 1.0 3413 || mode_lib->vba.htaps[k] != 1.0 3414 || mode_lib->vba.VRatio[k] != 1.0 3415 || mode_lib->vba.vtaps[k] != 1.0)) { 3416 mode_lib->vba.ScaleRatioAndTapsSupport = false; 3417 } else if (mode_lib->vba.vtaps[k] < 1.0 || mode_lib->vba.vtaps[k] > 8.0 3418 || mode_lib->vba.htaps[k] < 1.0 || mode_lib->vba.htaps[k] > 8.0 3419 || (mode_lib->vba.htaps[k] > 1.0 3420 && (mode_lib->vba.htaps[k] % 2) == 1) 3421 || mode_lib->vba.HRatio[k] > mode_lib->vba.MaxHSCLRatio 3422 || mode_lib->vba.VRatio[k] > mode_lib->vba.MaxVSCLRatio 3423 || mode_lib->vba.HRatio[k] > mode_lib->vba.htaps[k] 3424 || mode_lib->vba.VRatio[k] > mode_lib->vba.vtaps[k] 3425 || (mode_lib->vba.SourcePixelFormat[k] != dm_444_64 3426 && mode_lib->vba.SourcePixelFormat[k] != dm_444_32 3427 && mode_lib->vba.SourcePixelFormat[k] != dm_444_16 3428 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_16 3429 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_8 3430 && (mode_lib->vba.HRatio[k] / 2.0 3431 > mode_lib->vba.HTAPsChroma[k] 3432 || mode_lib->vba.VRatio[k] / 2.0 3433 > mode_lib->vba.VTAPsChroma[k]))) { 3434 mode_lib->vba.ScaleRatioAndTapsSupport = false; 3435 } 3436 } 3437 /*Source Format, Pixel Format and Scan Support Check*/ 3438 3439 mode_lib->vba.SourceFormatPixelAndScanSupport = true; 3440 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3441 if ((mode_lib->vba.SurfaceTiling[k] == dm_sw_linear 3442 && mode_lib->vba.SourceScan[k] != dm_horz) 3443 || ((mode_lib->vba.SurfaceTiling[k] == dm_sw_4kb_d 3444 || mode_lib->vba.SurfaceTiling[k] == dm_sw_4kb_d_x 3445 || mode_lib->vba.SurfaceTiling[k] == dm_sw_64kb_d 3446 || mode_lib->vba.SurfaceTiling[k] == dm_sw_64kb_d_t 3447 || mode_lib->vba.SurfaceTiling[k] == dm_sw_64kb_d_x 3448 || mode_lib->vba.SurfaceTiling[k] == dm_sw_var_d 3449 || mode_lib->vba.SurfaceTiling[k] == dm_sw_var_d_x) 3450 && mode_lib->vba.SourcePixelFormat[k] != dm_444_64) 3451 || (mode_lib->vba.SurfaceTiling[k] == dm_sw_64kb_r_x 3452 && (mode_lib->vba.SourcePixelFormat[k] == dm_mono_8 3453 || mode_lib->vba.SourcePixelFormat[k] 3454 == dm_420_8 3455 || mode_lib->vba.SourcePixelFormat[k] 3456 == dm_420_10)) 3457 || (((mode_lib->vba.SurfaceTiling[k] == dm_sw_gfx7_2d_thin_gl 3458 || mode_lib->vba.SurfaceTiling[k] 3459 == dm_sw_gfx7_2d_thin_l_vp) 3460 && !((mode_lib->vba.SourcePixelFormat[k] 3461 == dm_444_64 3462 || mode_lib->vba.SourcePixelFormat[k] 3463 == dm_444_32) 3464 && mode_lib->vba.SourceScan[k] 3465 == dm_horz 3466 && mode_lib->vba.SupportGFX7CompatibleTilingIn32bppAnd64bpp 3467 == true 3468 && mode_lib->vba.DCCEnable[k] 3469 == false)) 3470 || (mode_lib->vba.DCCEnable[k] == true 3471 && (mode_lib->vba.SurfaceTiling[k] 3472 == dm_sw_linear 3473 || mode_lib->vba.SourcePixelFormat[k] 3474 == dm_420_8 3475 || mode_lib->vba.SourcePixelFormat[k] 3476 == dm_420_10)))) { 3477 mode_lib->vba.SourceFormatPixelAndScanSupport = false; 3478 } 3479 } 3480 /*Bandwidth Support Check*/ 3481 3482 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3483 if (mode_lib->vba.SourcePixelFormat[k] == dm_444_64) { 3484 locals->BytePerPixelInDETY[k] = 8.0; 3485 locals->BytePerPixelInDETC[k] = 0.0; 3486 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_32) { 3487 locals->BytePerPixelInDETY[k] = 4.0; 3488 locals->BytePerPixelInDETC[k] = 0.0; 3489 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_16 3490 || mode_lib->vba.SourcePixelFormat[k] == dm_mono_16) { 3491 locals->BytePerPixelInDETY[k] = 2.0; 3492 locals->BytePerPixelInDETC[k] = 0.0; 3493 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_mono_8) { 3494 locals->BytePerPixelInDETY[k] = 1.0; 3495 locals->BytePerPixelInDETC[k] = 0.0; 3496 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_8) { 3497 locals->BytePerPixelInDETY[k] = 1.0; 3498 locals->BytePerPixelInDETC[k] = 2.0; 3499 } else { 3500 locals->BytePerPixelInDETY[k] = 4.0 / 3; 3501 locals->BytePerPixelInDETC[k] = 8.0 / 3; 3502 } 3503 if (mode_lib->vba.SourceScan[k] == dm_horz) { 3504 locals->SwathWidthYSingleDPP[k] = mode_lib->vba.ViewportWidth[k]; 3505 } else { 3506 locals->SwathWidthYSingleDPP[k] = mode_lib->vba.ViewportHeight[k]; 3507 } 3508 } 3509 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3510 locals->ReadBandwidthLuma[k] = locals->SwathWidthYSingleDPP[k] * dml_ceil(locals->BytePerPixelInDETY[k], 1.0) 3511 / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]) * mode_lib->vba.VRatio[k]; 3512 locals->ReadBandwidthChroma[k] = locals->SwathWidthYSingleDPP[k] / 2 * dml_ceil(locals->BytePerPixelInDETC[k], 2.0) 3513 / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]) * mode_lib->vba.VRatio[k] / 2.0; 3514 locals->ReadBandwidth[k] = locals->ReadBandwidthLuma[k] + locals->ReadBandwidthChroma[k]; 3515 } 3516 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3517 if (mode_lib->vba.WritebackEnable[k] == true 3518 && mode_lib->vba.WritebackPixelFormat[k] == dm_444_32) { 3519 locals->WriteBandwidth[k] = mode_lib->vba.WritebackDestinationWidth[k] 3520 * mode_lib->vba.WritebackDestinationHeight[k] 3521 / (mode_lib->vba.WritebackSourceHeight[k] 3522 * mode_lib->vba.HTotal[k] 3523 / mode_lib->vba.PixelClock[k]) * 4.0; 3524 } else if (mode_lib->vba.WritebackEnable[k] == true 3525 && mode_lib->vba.WritebackPixelFormat[k] == dm_420_10) { 3526 locals->WriteBandwidth[k] = mode_lib->vba.WritebackDestinationWidth[k] 3527 * mode_lib->vba.WritebackDestinationHeight[k] 3528 / (mode_lib->vba.WritebackSourceHeight[k] 3529 * mode_lib->vba.HTotal[k] 3530 / mode_lib->vba.PixelClock[k]) * 3.0; 3531 } else if (mode_lib->vba.WritebackEnable[k] == true) { 3532 locals->WriteBandwidth[k] = mode_lib->vba.WritebackDestinationWidth[k] 3533 * mode_lib->vba.WritebackDestinationHeight[k] 3534 / (mode_lib->vba.WritebackSourceHeight[k] 3535 * mode_lib->vba.HTotal[k] 3536 / mode_lib->vba.PixelClock[k]) * 1.5; 3537 } else { 3538 locals->WriteBandwidth[k] = 0.0; 3539 } 3540 } 3541 mode_lib->vba.DCCEnabledInAnyPlane = false; 3542 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3543 if (mode_lib->vba.DCCEnable[k] == true) { 3544 mode_lib->vba.DCCEnabledInAnyPlane = true; 3545 } 3546 } 3547 mode_lib->vba.UrgentLatency = mode_lib->vba.UrgentLatencyPixelDataOnly; 3548 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 3549 locals->FabricAndDRAMBandwidthPerState[i] = dml_min( 3550 mode_lib->vba.DRAMSpeedPerState[i] * mode_lib->vba.NumberOfChannels 3551 * mode_lib->vba.DRAMChannelWidth, 3552 mode_lib->vba.FabricClockPerState[i] 3553 * mode_lib->vba.FabricDatapathToDCNDataReturn) / 1000; 3554 locals->ReturnBWToDCNPerState = dml_min(locals->ReturnBusWidth * locals->DCFCLKPerState[i], 3555 locals->FabricAndDRAMBandwidthPerState[i] * 1000) 3556 * locals->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelDataOnly / 100; 3557 3558 locals->ReturnBWPerState[i][0] = locals->ReturnBWToDCNPerState; 3559 3560 if (locals->DCCEnabledInAnyPlane == true && locals->ReturnBWToDCNPerState > locals->DCFCLKPerState[i] * locals->ReturnBusWidth / 4) { 3561 locals->ReturnBWPerState[i][0] = dml_min(locals->ReturnBWPerState[i][0], 3562 locals->ReturnBWToDCNPerState * 4 * (1 - locals->UrgentLatency / 3563 ((locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024 3564 / (locals->ReturnBWToDCNPerState - locals->DCFCLKPerState[i] 3565 * locals->ReturnBusWidth / 4) + locals->UrgentLatency))); 3566 } 3567 locals->CriticalPoint = 2 * locals->ReturnBusWidth * locals->DCFCLKPerState[i] * 3568 locals->UrgentLatency / (locals->ReturnBWToDCNPerState * locals->UrgentLatency 3569 + (locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024); 3570 3571 if (locals->DCCEnabledInAnyPlane && locals->CriticalPoint > 1 && locals->CriticalPoint < 4) { 3572 locals->ReturnBWPerState[i][0] = dml_min(locals->ReturnBWPerState[i][0], 3573 4 * locals->ReturnBWToDCNPerState * 3574 (locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024 3575 * locals->ReturnBusWidth * locals->DCFCLKPerState[i] * locals->UrgentLatency / 3576 dml_pow((locals->ReturnBWToDCNPerState * locals->UrgentLatency 3577 + (locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024), 2)); 3578 } 3579 3580 locals->ReturnBWToDCNPerState = dml_min(locals->ReturnBusWidth * 3581 locals->DCFCLKPerState[i], locals->FabricAndDRAMBandwidthPerState[i] * 1000); 3582 3583 if (locals->DCCEnabledInAnyPlane == true && locals->ReturnBWToDCNPerState > locals->DCFCLKPerState[i] * locals->ReturnBusWidth / 4) { 3584 locals->ReturnBWPerState[i][0] = dml_min(locals->ReturnBWPerState[i][0], 3585 locals->ReturnBWToDCNPerState * 4 * (1 - locals->UrgentLatency / 3586 ((locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024 3587 / (locals->ReturnBWToDCNPerState - locals->DCFCLKPerState[i] 3588 * locals->ReturnBusWidth / 4) + locals->UrgentLatency))); 3589 } 3590 locals->CriticalPoint = 2 * locals->ReturnBusWidth * locals->DCFCLKPerState[i] * 3591 locals->UrgentLatency / (locals->ReturnBWToDCNPerState * locals->UrgentLatency 3592 + (locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024); 3593 3594 if (locals->DCCEnabledInAnyPlane && locals->CriticalPoint > 1 && locals->CriticalPoint < 4) { 3595 locals->ReturnBWPerState[i][0] = dml_min(locals->ReturnBWPerState[i][0], 3596 4 * locals->ReturnBWToDCNPerState * 3597 (locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024 3598 * locals->ReturnBusWidth * locals->DCFCLKPerState[i] * locals->UrgentLatency / 3599 dml_pow((locals->ReturnBWToDCNPerState * locals->UrgentLatency 3600 + (locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024), 2)); 3601 } 3602 } 3603 /*Writeback Latency support check*/ 3604 3605 mode_lib->vba.WritebackLatencySupport = true; 3606 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3607 if (mode_lib->vba.WritebackEnable[k] == true) { 3608 if (mode_lib->vba.WritebackPixelFormat[k] == dm_444_32) { 3609 if (locals->WriteBandwidth[k] 3610 > (mode_lib->vba.WritebackInterfaceLumaBufferSize 3611 + mode_lib->vba.WritebackInterfaceChromaBufferSize) 3612 / mode_lib->vba.WritebackLatency) { 3613 mode_lib->vba.WritebackLatencySupport = false; 3614 } 3615 } else { 3616 if (locals->WriteBandwidth[k] 3617 > 1.5 3618 * dml_min( 3619 mode_lib->vba.WritebackInterfaceLumaBufferSize, 3620 2.0 3621 * mode_lib->vba.WritebackInterfaceChromaBufferSize) 3622 / mode_lib->vba.WritebackLatency) { 3623 mode_lib->vba.WritebackLatencySupport = false; 3624 } 3625 } 3626 } 3627 } 3628 /*Re-ordering Buffer Support Check*/ 3629 3630 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 3631 locals->UrgentRoundTripAndOutOfOrderLatencyPerState[i] = 3632 (mode_lib->vba.RoundTripPingLatencyCycles + 32.0) / mode_lib->vba.DCFCLKPerState[i] 3633 + locals->UrgentOutOfOrderReturnPerChannel * mode_lib->vba.NumberOfChannels / locals->ReturnBWPerState[i][0]; 3634 if ((mode_lib->vba.ROBBufferSizeInKByte - mode_lib->vba.PixelChunkSizeInKByte) * 1024.0 / locals->ReturnBWPerState[i][0] 3635 > locals->UrgentRoundTripAndOutOfOrderLatencyPerState[i]) { 3636 locals->ROBSupport[i][0] = true; 3637 } else { 3638 locals->ROBSupport[i][0] = false; 3639 } 3640 } 3641 /*Writeback Mode Support Check*/ 3642 3643 mode_lib->vba.TotalNumberOfActiveWriteback = 0; 3644 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3645 if (mode_lib->vba.WritebackEnable[k] == true) { 3646 if (mode_lib->vba.ActiveWritebacksPerPlane[k] == 0) 3647 mode_lib->vba.ActiveWritebacksPerPlane[k] = 1; 3648 mode_lib->vba.TotalNumberOfActiveWriteback = 3649 mode_lib->vba.TotalNumberOfActiveWriteback 3650 + mode_lib->vba.ActiveWritebacksPerPlane[k]; 3651 } 3652 } 3653 mode_lib->vba.WritebackModeSupport = true; 3654 if (mode_lib->vba.TotalNumberOfActiveWriteback > mode_lib->vba.MaxNumWriteback) { 3655 mode_lib->vba.WritebackModeSupport = false; 3656 } 3657 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3658 if (mode_lib->vba.WritebackEnable[k] == true 3659 && mode_lib->vba.Writeback10bpc420Supported != true 3660 && mode_lib->vba.WritebackPixelFormat[k] == dm_420_10) { 3661 mode_lib->vba.WritebackModeSupport = false; 3662 } 3663 } 3664 /*Writeback Scale Ratio and Taps Support Check*/ 3665 3666 mode_lib->vba.WritebackScaleRatioAndTapsSupport = true; 3667 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3668 if (mode_lib->vba.WritebackEnable[k] == true) { 3669 if (mode_lib->vba.WritebackLumaAndChromaScalingSupported == false 3670 && (mode_lib->vba.WritebackHRatio[k] != 1.0 3671 || mode_lib->vba.WritebackVRatio[k] != 1.0)) { 3672 mode_lib->vba.WritebackScaleRatioAndTapsSupport = false; 3673 } 3674 if (mode_lib->vba.WritebackHRatio[k] > mode_lib->vba.WritebackMaxHSCLRatio 3675 || mode_lib->vba.WritebackVRatio[k] 3676 > mode_lib->vba.WritebackMaxVSCLRatio 3677 || mode_lib->vba.WritebackHRatio[k] 3678 < mode_lib->vba.WritebackMinHSCLRatio 3679 || mode_lib->vba.WritebackVRatio[k] 3680 < mode_lib->vba.WritebackMinVSCLRatio 3681 || mode_lib->vba.WritebackLumaHTaps[k] 3682 > mode_lib->vba.WritebackMaxHSCLTaps 3683 || mode_lib->vba.WritebackLumaVTaps[k] 3684 > mode_lib->vba.WritebackMaxVSCLTaps 3685 || mode_lib->vba.WritebackHRatio[k] 3686 > mode_lib->vba.WritebackLumaHTaps[k] 3687 || mode_lib->vba.WritebackVRatio[k] 3688 > mode_lib->vba.WritebackLumaVTaps[k] 3689 || (mode_lib->vba.WritebackLumaHTaps[k] > 2.0 3690 && ((mode_lib->vba.WritebackLumaHTaps[k] % 2) 3691 == 1)) 3692 || (mode_lib->vba.WritebackPixelFormat[k] != dm_444_32 3693 && (mode_lib->vba.WritebackChromaHTaps[k] 3694 > mode_lib->vba.WritebackMaxHSCLTaps 3695 || mode_lib->vba.WritebackChromaVTaps[k] 3696 > mode_lib->vba.WritebackMaxVSCLTaps 3697 || 2.0 3698 * mode_lib->vba.WritebackHRatio[k] 3699 > mode_lib->vba.WritebackChromaHTaps[k] 3700 || 2.0 3701 * mode_lib->vba.WritebackVRatio[k] 3702 > mode_lib->vba.WritebackChromaVTaps[k] 3703 || (mode_lib->vba.WritebackChromaHTaps[k] > 2.0 3704 && ((mode_lib->vba.WritebackChromaHTaps[k] % 2) == 1))))) { 3705 mode_lib->vba.WritebackScaleRatioAndTapsSupport = false; 3706 } 3707 if (mode_lib->vba.WritebackVRatio[k] < 1.0) { 3708 mode_lib->vba.WritebackLumaVExtra = 3709 dml_max(1.0 - 2.0 / dml_ceil(1.0 / mode_lib->vba.WritebackVRatio[k], 1.0), 0.0); 3710 } else { 3711 mode_lib->vba.WritebackLumaVExtra = -1; 3712 } 3713 if ((mode_lib->vba.WritebackPixelFormat[k] == dm_444_32 3714 && mode_lib->vba.WritebackLumaVTaps[k] 3715 > (mode_lib->vba.WritebackLineBufferLumaBufferSize 3716 + mode_lib->vba.WritebackLineBufferChromaBufferSize) 3717 / 3.0 3718 / mode_lib->vba.WritebackDestinationWidth[k] 3719 - mode_lib->vba.WritebackLumaVExtra) 3720 || (mode_lib->vba.WritebackPixelFormat[k] == dm_420_8 3721 && mode_lib->vba.WritebackLumaVTaps[k] 3722 > mode_lib->vba.WritebackLineBufferLumaBufferSize 3723 * 8.0 / 10.0 / mode_lib->vba.WritebackDestinationWidth[k] 3724 - mode_lib->vba.WritebackLumaVExtra) 3725 || (mode_lib->vba.WritebackPixelFormat[k] == dm_420_10 3726 && mode_lib->vba.WritebackLumaVTaps[k] 3727 > mode_lib->vba.WritebackLineBufferLumaBufferSize 3728 * 8.0 / 10.0 3729 / mode_lib->vba.WritebackDestinationWidth[k] 3730 - mode_lib->vba.WritebackLumaVExtra)) { 3731 mode_lib->vba.WritebackScaleRatioAndTapsSupport = false; 3732 } 3733 if (2.0 * mode_lib->vba.WritebackVRatio[k] < 1) { 3734 mode_lib->vba.WritebackChromaVExtra = 0.0; 3735 } else { 3736 mode_lib->vba.WritebackChromaVExtra = -1; 3737 } 3738 if ((mode_lib->vba.WritebackPixelFormat[k] == dm_420_8 3739 && mode_lib->vba.WritebackChromaVTaps[k] 3740 > mode_lib->vba.WritebackLineBufferChromaBufferSize 3741 * 8.0 / 10.0 / mode_lib->vba.WritebackDestinationWidth[k] 3742 - mode_lib->vba.WritebackChromaVExtra) 3743 || (mode_lib->vba.WritebackPixelFormat[k] == dm_420_10 3744 && mode_lib->vba.WritebackChromaVTaps[k] 3745 > mode_lib->vba.WritebackLineBufferChromaBufferSize 3746 * 8.0 / 10.0 3747 / mode_lib->vba.WritebackDestinationWidth[k] 3748 - mode_lib->vba.WritebackChromaVExtra)) { 3749 mode_lib->vba.WritebackScaleRatioAndTapsSupport = false; 3750 } 3751 } 3752 } 3753 /*Maximum DISPCLK/DPPCLK Support check*/ 3754 3755 mode_lib->vba.WritebackRequiredDISPCLK = 0.0; 3756 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3757 if (mode_lib->vba.WritebackEnable[k] == true) { 3758 mode_lib->vba.WritebackRequiredDISPCLK = 3759 dml_max( 3760 mode_lib->vba.WritebackRequiredDISPCLK, 3761 CalculateWriteBackDISPCLK( 3762 mode_lib->vba.WritebackPixelFormat[k], 3763 mode_lib->vba.PixelClock[k], 3764 mode_lib->vba.WritebackHRatio[k], 3765 mode_lib->vba.WritebackVRatio[k], 3766 mode_lib->vba.WritebackLumaHTaps[k], 3767 mode_lib->vba.WritebackLumaVTaps[k], 3768 mode_lib->vba.WritebackChromaHTaps[k], 3769 mode_lib->vba.WritebackChromaVTaps[k], 3770 mode_lib->vba.WritebackDestinationWidth[k], 3771 mode_lib->vba.HTotal[k], 3772 mode_lib->vba.WritebackChromaLineBufferWidth)); 3773 } 3774 } 3775 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3776 if (mode_lib->vba.HRatio[k] > 1.0) { 3777 locals->PSCL_FACTOR[k] = dml_min( 3778 mode_lib->vba.MaxDCHUBToPSCLThroughput, 3779 mode_lib->vba.MaxPSCLToLBThroughput 3780 * mode_lib->vba.HRatio[k] 3781 / dml_ceil( 3782 mode_lib->vba.htaps[k] 3783 / 6.0, 3784 1.0)); 3785 } else { 3786 locals->PSCL_FACTOR[k] = dml_min( 3787 mode_lib->vba.MaxDCHUBToPSCLThroughput, 3788 mode_lib->vba.MaxPSCLToLBThroughput); 3789 } 3790 if (locals->BytePerPixelInDETC[k] == 0.0) { 3791 locals->PSCL_FACTOR_CHROMA[k] = 0.0; 3792 locals->MinDPPCLKUsingSingleDPP[k] = 3793 mode_lib->vba.PixelClock[k] 3794 * dml_max3( 3795 mode_lib->vba.vtaps[k] / 6.0 3796 * dml_min( 3797 1.0, 3798 mode_lib->vba.HRatio[k]), 3799 mode_lib->vba.HRatio[k] 3800 * mode_lib->vba.VRatio[k] 3801 / locals->PSCL_FACTOR[k], 3802 1.0); 3803 if ((mode_lib->vba.htaps[k] > 6.0 || mode_lib->vba.vtaps[k] > 6.0) 3804 && locals->MinDPPCLKUsingSingleDPP[k] 3805 < 2.0 * mode_lib->vba.PixelClock[k]) { 3806 locals->MinDPPCLKUsingSingleDPP[k] = 2.0 3807 * mode_lib->vba.PixelClock[k]; 3808 } 3809 } else { 3810 if (mode_lib->vba.HRatio[k] / 2.0 > 1.0) { 3811 locals->PSCL_FACTOR_CHROMA[k] = 3812 dml_min( 3813 mode_lib->vba.MaxDCHUBToPSCLThroughput, 3814 mode_lib->vba.MaxPSCLToLBThroughput 3815 * mode_lib->vba.HRatio[k] 3816 / 2.0 3817 / dml_ceil( 3818 mode_lib->vba.HTAPsChroma[k] 3819 / 6.0, 3820 1.0)); 3821 } else { 3822 locals->PSCL_FACTOR_CHROMA[k] = dml_min( 3823 mode_lib->vba.MaxDCHUBToPSCLThroughput, 3824 mode_lib->vba.MaxPSCLToLBThroughput); 3825 } 3826 locals->MinDPPCLKUsingSingleDPP[k] = 3827 mode_lib->vba.PixelClock[k] 3828 * dml_max5( 3829 mode_lib->vba.vtaps[k] / 6.0 3830 * dml_min( 3831 1.0, 3832 mode_lib->vba.HRatio[k]), 3833 mode_lib->vba.HRatio[k] 3834 * mode_lib->vba.VRatio[k] 3835 / locals->PSCL_FACTOR[k], 3836 mode_lib->vba.VTAPsChroma[k] 3837 / 6.0 3838 * dml_min( 3839 1.0, 3840 mode_lib->vba.HRatio[k] 3841 / 2.0), 3842 mode_lib->vba.HRatio[k] 3843 * mode_lib->vba.VRatio[k] 3844 / 4.0 3845 / locals->PSCL_FACTOR_CHROMA[k], 3846 1.0); 3847 if ((mode_lib->vba.htaps[k] > 6.0 || mode_lib->vba.vtaps[k] > 6.0 3848 || mode_lib->vba.HTAPsChroma[k] > 6.0 3849 || mode_lib->vba.VTAPsChroma[k] > 6.0) 3850 && locals->MinDPPCLKUsingSingleDPP[k] 3851 < 2.0 * mode_lib->vba.PixelClock[k]) { 3852 locals->MinDPPCLKUsingSingleDPP[k] = 2.0 3853 * mode_lib->vba.PixelClock[k]; 3854 } 3855 } 3856 } 3857 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3858 Calculate256BBlockSizes( 3859 mode_lib->vba.SourcePixelFormat[k], 3860 mode_lib->vba.SurfaceTiling[k], 3861 dml_ceil(locals->BytePerPixelInDETY[k], 1.0), 3862 dml_ceil(locals->BytePerPixelInDETC[k], 2.0), 3863 &locals->Read256BlockHeightY[k], 3864 &locals->Read256BlockHeightC[k], 3865 &locals->Read256BlockWidthY[k], 3866 &locals->Read256BlockWidthC[k]); 3867 if (mode_lib->vba.SourceScan[k] == dm_horz) { 3868 locals->MaxSwathHeightY[k] = locals->Read256BlockHeightY[k]; 3869 locals->MaxSwathHeightC[k] = locals->Read256BlockHeightC[k]; 3870 } else { 3871 locals->MaxSwathHeightY[k] = locals->Read256BlockWidthY[k]; 3872 locals->MaxSwathHeightC[k] = locals->Read256BlockWidthC[k]; 3873 } 3874 if ((mode_lib->vba.SourcePixelFormat[k] == dm_444_64 3875 || mode_lib->vba.SourcePixelFormat[k] == dm_444_32 3876 || mode_lib->vba.SourcePixelFormat[k] == dm_444_16 3877 || mode_lib->vba.SourcePixelFormat[k] == dm_mono_16 3878 || mode_lib->vba.SourcePixelFormat[k] == dm_mono_8)) { 3879 if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear 3880 || (mode_lib->vba.SourcePixelFormat[k] == dm_444_64 3881 && (mode_lib->vba.SurfaceTiling[k] 3882 == dm_sw_4kb_s 3883 || mode_lib->vba.SurfaceTiling[k] 3884 == dm_sw_4kb_s_x 3885 || mode_lib->vba.SurfaceTiling[k] 3886 == dm_sw_64kb_s 3887 || mode_lib->vba.SurfaceTiling[k] 3888 == dm_sw_64kb_s_t 3889 || mode_lib->vba.SurfaceTiling[k] 3890 == dm_sw_64kb_s_x 3891 || mode_lib->vba.SurfaceTiling[k] 3892 == dm_sw_var_s 3893 || mode_lib->vba.SurfaceTiling[k] 3894 == dm_sw_var_s_x) 3895 && mode_lib->vba.SourceScan[k] == dm_horz)) { 3896 locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k]; 3897 } else { 3898 locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k] 3899 / 2.0; 3900 } 3901 locals->MinSwathHeightC[k] = locals->MaxSwathHeightC[k]; 3902 } else { 3903 if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear) { 3904 locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k]; 3905 locals->MinSwathHeightC[k] = locals->MaxSwathHeightC[k]; 3906 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_8 3907 && mode_lib->vba.SourceScan[k] == dm_horz) { 3908 locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k] 3909 / 2.0; 3910 locals->MinSwathHeightC[k] = locals->MaxSwathHeightC[k]; 3911 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_10 3912 && mode_lib->vba.SourceScan[k] == dm_horz) { 3913 locals->MinSwathHeightC[k] = locals->MaxSwathHeightC[k] 3914 / 2.0; 3915 locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k]; 3916 } else { 3917 locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k]; 3918 locals->MinSwathHeightC[k] = locals->MaxSwathHeightC[k]; 3919 } 3920 } 3921 if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear) { 3922 mode_lib->vba.MaximumSwathWidthSupport = 8192.0; 3923 } else { 3924 mode_lib->vba.MaximumSwathWidthSupport = 5120.0; 3925 } 3926 mode_lib->vba.MaximumSwathWidthInDETBuffer = 3927 dml_min( 3928 mode_lib->vba.MaximumSwathWidthSupport, 3929 mode_lib->vba.DETBufferSizeInKByte[0] * 1024.0 / 2.0 3930 / (locals->BytePerPixelInDETY[k] 3931 * locals->MinSwathHeightY[k] 3932 + locals->BytePerPixelInDETC[k] 3933 / 2.0 3934 * locals->MinSwathHeightC[k])); 3935 if (locals->BytePerPixelInDETC[k] == 0.0) { 3936 mode_lib->vba.MaximumSwathWidthInLineBuffer = 3937 mode_lib->vba.LineBufferSize 3938 * dml_max(mode_lib->vba.HRatio[k], 1.0) 3939 / mode_lib->vba.LBBitPerPixel[k] 3940 / (mode_lib->vba.vtaps[k] 3941 + dml_max( 3942 dml_ceil( 3943 mode_lib->vba.VRatio[k], 3944 1.0) 3945 - 2, 3946 0.0)); 3947 } else { 3948 mode_lib->vba.MaximumSwathWidthInLineBuffer = 3949 dml_min( 3950 mode_lib->vba.LineBufferSize 3951 * dml_max( 3952 mode_lib->vba.HRatio[k], 3953 1.0) 3954 / mode_lib->vba.LBBitPerPixel[k] 3955 / (mode_lib->vba.vtaps[k] 3956 + dml_max( 3957 dml_ceil( 3958 mode_lib->vba.VRatio[k], 3959 1.0) 3960 - 2, 3961 0.0)), 3962 2.0 * mode_lib->vba.LineBufferSize 3963 * dml_max( 3964 mode_lib->vba.HRatio[k] 3965 / 2.0, 3966 1.0) 3967 / mode_lib->vba.LBBitPerPixel[k] 3968 / (mode_lib->vba.VTAPsChroma[k] 3969 + dml_max( 3970 dml_ceil( 3971 mode_lib->vba.VRatio[k] 3972 / 2.0, 3973 1.0) 3974 - 2, 3975 0.0))); 3976 } 3977 locals->MaximumSwathWidth[k] = dml_min( 3978 mode_lib->vba.MaximumSwathWidthInDETBuffer, 3979 mode_lib->vba.MaximumSwathWidthInLineBuffer); 3980 } 3981 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 3982 double MaxMaxDispclkRoundedDown = RoundToDFSGranularityDown( 3983 mode_lib->vba.MaxDispclk[mode_lib->vba.soc.num_states], 3984 mode_lib->vba.DISPCLKDPPCLKVCOSpeed); 3985 3986 for (j = 0; j < 2; j++) { 3987 mode_lib->vba.MaxDispclkRoundedDownToDFSGranularity = RoundToDFSGranularityDown( 3988 mode_lib->vba.MaxDispclk[i], 3989 mode_lib->vba.DISPCLKDPPCLKVCOSpeed); 3990 mode_lib->vba.MaxDppclkRoundedDownToDFSGranularity = RoundToDFSGranularityDown( 3991 mode_lib->vba.MaxDppclk[i], 3992 mode_lib->vba.DISPCLKDPPCLKVCOSpeed); 3993 locals->RequiredDISPCLK[i][j] = 0.0; 3994 locals->DISPCLK_DPPCLK_Support[i][j] = true; 3995 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3996 mode_lib->vba.PlaneRequiredDISPCLKWithoutODMCombine = 3997 mode_lib->vba.PixelClock[k] * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) 3998 * (1.0 + mode_lib->vba.DISPCLKRampingMargin / 100.0); 3999 if (mode_lib->vba.PlaneRequiredDISPCLKWithoutODMCombine >= mode_lib->vba.MaxDispclk[i] 4000 && i == mode_lib->vba.soc.num_states) 4001 mode_lib->vba.PlaneRequiredDISPCLKWithoutODMCombine = mode_lib->vba.PixelClock[k] 4002 * (1 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); 4003 4004 mode_lib->vba.PlaneRequiredDISPCLKWithODMCombine = mode_lib->vba.PixelClock[k] / 2 4005 * (1 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) * (1 + mode_lib->vba.DISPCLKRampingMargin / 100.0); 4006 if (mode_lib->vba.PlaneRequiredDISPCLKWithODMCombine >= mode_lib->vba.MaxDispclk[i] 4007 && i == mode_lib->vba.soc.num_states) 4008 mode_lib->vba.PlaneRequiredDISPCLKWithODMCombine = mode_lib->vba.PixelClock[k] / 2 4009 * (1 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); 4010 4011 locals->ODMCombineEnablePerState[i][k] = false; 4012 mode_lib->vba.PlaneRequiredDISPCLK = mode_lib->vba.PlaneRequiredDISPCLKWithoutODMCombine; 4013 if (mode_lib->vba.ODMCapability) { 4014 if (locals->PlaneRequiredDISPCLKWithoutODMCombine > MaxMaxDispclkRoundedDown) { 4015 locals->ODMCombineEnablePerState[i][k] = true; 4016 mode_lib->vba.PlaneRequiredDISPCLK = mode_lib->vba.PlaneRequiredDISPCLKWithODMCombine; 4017 } else if (locals->DSCEnabled[k] && (locals->HActive[k] > DCN20_MAX_DSC_IMAGE_WIDTH)) { 4018 locals->ODMCombineEnablePerState[i][k] = true; 4019 mode_lib->vba.PlaneRequiredDISPCLK = mode_lib->vba.PlaneRequiredDISPCLKWithODMCombine; 4020 } else if (locals->HActive[k] > DCN20_MAX_420_IMAGE_WIDTH && locals->OutputFormat[k] == dm_420) { 4021 locals->ODMCombineEnablePerState[i][k] = true; 4022 mode_lib->vba.PlaneRequiredDISPCLK = mode_lib->vba.PlaneRequiredDISPCLKWithODMCombine; 4023 } 4024 } 4025 4026 if (locals->MinDPPCLKUsingSingleDPP[k] * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) <= mode_lib->vba.MaxDppclkRoundedDownToDFSGranularity 4027 && locals->SwathWidthYSingleDPP[k] <= locals->MaximumSwathWidth[k] 4028 && locals->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_disabled) { 4029 locals->NoOfDPP[i][j][k] = 1; 4030 locals->RequiredDPPCLK[i][j][k] = 4031 locals->MinDPPCLKUsingSingleDPP[k] * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); 4032 } else { 4033 locals->NoOfDPP[i][j][k] = 2; 4034 locals->RequiredDPPCLK[i][j][k] = 4035 locals->MinDPPCLKUsingSingleDPP[k] * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) / 2.0; 4036 } 4037 locals->RequiredDISPCLK[i][j] = dml_max( 4038 locals->RequiredDISPCLK[i][j], 4039 mode_lib->vba.PlaneRequiredDISPCLK); 4040 if ((locals->MinDPPCLKUsingSingleDPP[k] / locals->NoOfDPP[i][j][k] * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) 4041 > mode_lib->vba.MaxDppclkRoundedDownToDFSGranularity) 4042 || (mode_lib->vba.PlaneRequiredDISPCLK > mode_lib->vba.MaxDispclkRoundedDownToDFSGranularity)) { 4043 locals->DISPCLK_DPPCLK_Support[i][j] = false; 4044 } 4045 } 4046 locals->TotalNumberOfActiveDPP[i][j] = 0.0; 4047 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) 4048 locals->TotalNumberOfActiveDPP[i][j] = locals->TotalNumberOfActiveDPP[i][j] + locals->NoOfDPP[i][j][k]; 4049 if (j == 1) { 4050 while (locals->TotalNumberOfActiveDPP[i][j] < mode_lib->vba.MaxNumDPP 4051 && locals->TotalNumberOfActiveDPP[i][j] < 2 * mode_lib->vba.NumberOfActivePlanes) { 4052 double BWOfNonSplitPlaneOfMaximumBandwidth; 4053 unsigned int NumberOfNonSplitPlaneOfMaximumBandwidth; 4054 4055 BWOfNonSplitPlaneOfMaximumBandwidth = 0; 4056 NumberOfNonSplitPlaneOfMaximumBandwidth = 0; 4057 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 4058 if (locals->ReadBandwidth[k] > BWOfNonSplitPlaneOfMaximumBandwidth && locals->NoOfDPP[i][j][k] == 1) { 4059 BWOfNonSplitPlaneOfMaximumBandwidth = locals->ReadBandwidth[k]; 4060 NumberOfNonSplitPlaneOfMaximumBandwidth = k; 4061 } 4062 } 4063 locals->NoOfDPP[i][j][NumberOfNonSplitPlaneOfMaximumBandwidth] = 2; 4064 locals->RequiredDPPCLK[i][j][NumberOfNonSplitPlaneOfMaximumBandwidth] = 4065 locals->MinDPPCLKUsingSingleDPP[NumberOfNonSplitPlaneOfMaximumBandwidth] 4066 * (1 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100) / 2; 4067 locals->TotalNumberOfActiveDPP[i][j] = locals->TotalNumberOfActiveDPP[i][j] + 1; 4068 } 4069 } 4070 if (locals->TotalNumberOfActiveDPP[i][j] > mode_lib->vba.MaxNumDPP) { 4071 locals->RequiredDISPCLK[i][j] = 0.0; 4072 locals->DISPCLK_DPPCLK_Support[i][j] = true; 4073 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4074 locals->ODMCombineEnablePerState[i][k] = false; 4075 if (locals->SwathWidthYSingleDPP[k] <= locals->MaximumSwathWidth[k]) { 4076 locals->NoOfDPP[i][j][k] = 1; 4077 locals->RequiredDPPCLK[i][j][k] = locals->MinDPPCLKUsingSingleDPP[k] 4078 * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); 4079 } else { 4080 locals->NoOfDPP[i][j][k] = 2; 4081 locals->RequiredDPPCLK[i][j][k] = locals->MinDPPCLKUsingSingleDPP[k] 4082 * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) / 2.0; 4083 } 4084 if (i != mode_lib->vba.soc.num_states) { 4085 mode_lib->vba.PlaneRequiredDISPCLK = 4086 mode_lib->vba.PixelClock[k] 4087 * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) 4088 * (1.0 + mode_lib->vba.DISPCLKRampingMargin / 100.0); 4089 } else { 4090 mode_lib->vba.PlaneRequiredDISPCLK = mode_lib->vba.PixelClock[k] 4091 * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); 4092 } 4093 locals->RequiredDISPCLK[i][j] = dml_max( 4094 locals->RequiredDISPCLK[i][j], 4095 mode_lib->vba.PlaneRequiredDISPCLK); 4096 if (locals->MinDPPCLKUsingSingleDPP[k] / locals->NoOfDPP[i][j][k] * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) 4097 > mode_lib->vba.MaxDppclkRoundedDownToDFSGranularity 4098 || mode_lib->vba.PlaneRequiredDISPCLK > mode_lib->vba.MaxDispclkRoundedDownToDFSGranularity) 4099 locals->DISPCLK_DPPCLK_Support[i][j] = false; 4100 } 4101 locals->TotalNumberOfActiveDPP[i][j] = 0.0; 4102 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) 4103 locals->TotalNumberOfActiveDPP[i][j] = locals->TotalNumberOfActiveDPP[i][j] + locals->NoOfDPP[i][j][k]; 4104 } 4105 locals->RequiredDISPCLK[i][j] = dml_max( 4106 locals->RequiredDISPCLK[i][j], 4107 mode_lib->vba.WritebackRequiredDISPCLK); 4108 if (mode_lib->vba.MaxDispclkRoundedDownToDFSGranularity 4109 < mode_lib->vba.WritebackRequiredDISPCLK) { 4110 locals->DISPCLK_DPPCLK_Support[i][j] = false; 4111 } 4112 } 4113 } 4114 /*Viewport Size Check*/ 4115 4116 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 4117 locals->ViewportSizeSupport[i][0] = true; 4118 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4119 if (locals->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_2to1) { 4120 if (dml_min(locals->SwathWidthYSingleDPP[k], dml_round(mode_lib->vba.HActive[k] / 2.0 * mode_lib->vba.HRatio[k])) 4121 > locals->MaximumSwathWidth[k]) { 4122 locals->ViewportSizeSupport[i][0] = false; 4123 } 4124 } else { 4125 if (locals->SwathWidthYSingleDPP[k] / 2.0 > locals->MaximumSwathWidth[k]) { 4126 locals->ViewportSizeSupport[i][0] = false; 4127 } 4128 } 4129 } 4130 } 4131 /*Total Available Pipes Support Check*/ 4132 4133 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 4134 for (j = 0; j < 2; j++) { 4135 if (locals->TotalNumberOfActiveDPP[i][j] <= mode_lib->vba.MaxNumDPP) 4136 locals->TotalAvailablePipesSupport[i][j] = true; 4137 else 4138 locals->TotalAvailablePipesSupport[i][j] = false; 4139 } 4140 } 4141 /*Total Available OTG Support Check*/ 4142 4143 mode_lib->vba.TotalNumberOfActiveOTG = 0.0; 4144 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4145 if (mode_lib->vba.BlendingAndTiming[k] == k) { 4146 mode_lib->vba.TotalNumberOfActiveOTG = mode_lib->vba.TotalNumberOfActiveOTG 4147 + 1.0; 4148 } 4149 } 4150 if (mode_lib->vba.TotalNumberOfActiveOTG <= mode_lib->vba.MaxNumOTG) { 4151 mode_lib->vba.NumberOfOTGSupport = true; 4152 } else { 4153 mode_lib->vba.NumberOfOTGSupport = false; 4154 } 4155 /*Display IO and DSC Support Check*/ 4156 4157 mode_lib->vba.NonsupportedDSCInputBPC = false; 4158 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4159 if (!(mode_lib->vba.DSCInputBitPerComponent[k] == 12.0 4160 || mode_lib->vba.DSCInputBitPerComponent[k] == 10.0 4161 || mode_lib->vba.DSCInputBitPerComponent[k] == 8.0)) { 4162 mode_lib->vba.NonsupportedDSCInputBPC = true; 4163 } 4164 } 4165 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 4166 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4167 locals->RequiresDSC[i][k] = 0; 4168 locals->RequiresFEC[i][k] = 0; 4169 if (mode_lib->vba.BlendingAndTiming[k] == k) { 4170 if (mode_lib->vba.Output[k] == dm_hdmi) { 4171 locals->RequiresDSC[i][k] = 0; 4172 locals->RequiresFEC[i][k] = 0; 4173 locals->OutputBppPerState[i][k] = TruncToValidBPP( 4174 dml_min(600.0, mode_lib->vba.PHYCLKPerState[i]) / mode_lib->vba.PixelClockBackEnd[k] * 24, 4175 mode_lib->vba.ForcedOutputLinkBPP[k], 4176 false, 4177 mode_lib->vba.Output[k], 4178 mode_lib->vba.OutputFormat[k], 4179 mode_lib->vba.DSCInputBitPerComponent[k]); 4180 } else if (mode_lib->vba.Output[k] == dm_dp 4181 || mode_lib->vba.Output[k] == dm_edp) { 4182 if (mode_lib->vba.Output[k] == dm_edp) { 4183 mode_lib->vba.EffectiveFECOverhead = 0.0; 4184 } else { 4185 mode_lib->vba.EffectiveFECOverhead = 4186 mode_lib->vba.FECOverhead; 4187 } 4188 if (mode_lib->vba.PHYCLKPerState[i] >= 270.0) { 4189 mode_lib->vba.Outbpp = TruncToValidBPP( 4190 (1.0 - mode_lib->vba.Downspreading / 100.0) * 270.0 4191 * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0, 4192 mode_lib->vba.ForcedOutputLinkBPP[k], 4193 false, 4194 mode_lib->vba.Output[k], 4195 mode_lib->vba.OutputFormat[k], 4196 mode_lib->vba.DSCInputBitPerComponent[k]); 4197 mode_lib->vba.OutbppDSC = TruncToValidBPP( 4198 (1.0 - mode_lib->vba.Downspreading / 100.0) * (1.0 - mode_lib->vba.EffectiveFECOverhead / 100.0) * 270.0 4199 * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0, 4200 mode_lib->vba.ForcedOutputLinkBPP[k], 4201 true, 4202 mode_lib->vba.Output[k], 4203 mode_lib->vba.OutputFormat[k], 4204 mode_lib->vba.DSCInputBitPerComponent[k]); 4205 if (mode_lib->vba.DSCEnabled[k] == true) { 4206 locals->RequiresDSC[i][k] = true; 4207 if (mode_lib->vba.Output[k] == dm_dp) { 4208 locals->RequiresFEC[i][k] = true; 4209 } else { 4210 locals->RequiresFEC[i][k] = false; 4211 } 4212 mode_lib->vba.Outbpp = mode_lib->vba.OutbppDSC; 4213 } else { 4214 locals->RequiresDSC[i][k] = false; 4215 locals->RequiresFEC[i][k] = false; 4216 } 4217 locals->OutputBppPerState[i][k] = mode_lib->vba.Outbpp; 4218 } 4219 if (mode_lib->vba.Outbpp == BPP_INVALID && mode_lib->vba.PHYCLKPerState[i] >= 540.0) { 4220 mode_lib->vba.Outbpp = TruncToValidBPP( 4221 (1.0 - mode_lib->vba.Downspreading / 100.0) * 540.0 4222 * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0, 4223 mode_lib->vba.ForcedOutputLinkBPP[k], 4224 false, 4225 mode_lib->vba.Output[k], 4226 mode_lib->vba.OutputFormat[k], 4227 mode_lib->vba.DSCInputBitPerComponent[k]); 4228 mode_lib->vba.OutbppDSC = TruncToValidBPP( 4229 (1.0 - mode_lib->vba.Downspreading / 100.0) * (1.0 - mode_lib->vba.EffectiveFECOverhead / 100.0) * 540.0 4230 * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0, 4231 mode_lib->vba.ForcedOutputLinkBPP[k], 4232 true, 4233 mode_lib->vba.Output[k], 4234 mode_lib->vba.OutputFormat[k], 4235 mode_lib->vba.DSCInputBitPerComponent[k]); 4236 if (mode_lib->vba.DSCEnabled[k] == true) { 4237 locals->RequiresDSC[i][k] = true; 4238 if (mode_lib->vba.Output[k] == dm_dp) { 4239 locals->RequiresFEC[i][k] = true; 4240 } else { 4241 locals->RequiresFEC[i][k] = false; 4242 } 4243 mode_lib->vba.Outbpp = mode_lib->vba.OutbppDSC; 4244 } else { 4245 locals->RequiresDSC[i][k] = false; 4246 locals->RequiresFEC[i][k] = false; 4247 } 4248 locals->OutputBppPerState[i][k] = mode_lib->vba.Outbpp; 4249 } 4250 if (mode_lib->vba.Outbpp == BPP_INVALID 4251 && mode_lib->vba.PHYCLKPerState[i] 4252 >= 810.0) { 4253 mode_lib->vba.Outbpp = TruncToValidBPP( 4254 (1.0 - mode_lib->vba.Downspreading / 100.0) * 810.0 4255 * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0, 4256 mode_lib->vba.ForcedOutputLinkBPP[k], 4257 false, 4258 mode_lib->vba.Output[k], 4259 mode_lib->vba.OutputFormat[k], 4260 mode_lib->vba.DSCInputBitPerComponent[k]); 4261 mode_lib->vba.OutbppDSC = TruncToValidBPP( 4262 (1.0 - mode_lib->vba.Downspreading / 100.0) * (1.0 - mode_lib->vba.EffectiveFECOverhead / 100.0) * 810.0 4263 * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0, 4264 mode_lib->vba.ForcedOutputLinkBPP[k], 4265 true, 4266 mode_lib->vba.Output[k], 4267 mode_lib->vba.OutputFormat[k], 4268 mode_lib->vba.DSCInputBitPerComponent[k]); 4269 if (mode_lib->vba.DSCEnabled[k] == true || mode_lib->vba.Outbpp == BPP_INVALID) { 4270 locals->RequiresDSC[i][k] = true; 4271 if (mode_lib->vba.Output[k] == dm_dp) { 4272 locals->RequiresFEC[i][k] = true; 4273 } else { 4274 locals->RequiresFEC[i][k] = false; 4275 } 4276 mode_lib->vba.Outbpp = mode_lib->vba.OutbppDSC; 4277 } else { 4278 locals->RequiresDSC[i][k] = false; 4279 locals->RequiresFEC[i][k] = false; 4280 } 4281 locals->OutputBppPerState[i][k] = 4282 mode_lib->vba.Outbpp; 4283 } 4284 } 4285 } else { 4286 locals->OutputBppPerState[i][k] = BPP_BLENDED_PIPE; 4287 } 4288 } 4289 } 4290 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 4291 locals->DIOSupport[i] = true; 4292 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4293 if (!mode_lib->vba.skip_dio_check[k] 4294 && (locals->OutputBppPerState[i][k] == BPP_INVALID 4295 || (mode_lib->vba.OutputFormat[k] == dm_420 4296 && mode_lib->vba.Interlace[k] == true 4297 && mode_lib->vba.ProgressiveToInterlaceUnitInOPP == true))) { 4298 locals->DIOSupport[i] = false; 4299 } 4300 } 4301 } 4302 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 4303 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4304 locals->DSCCLKRequiredMoreThanSupported[i] = false; 4305 if (mode_lib->vba.BlendingAndTiming[k] == k) { 4306 if ((mode_lib->vba.Output[k] == dm_dp 4307 || mode_lib->vba.Output[k] == dm_edp)) { 4308 if (mode_lib->vba.OutputFormat[k] == dm_420 4309 || mode_lib->vba.OutputFormat[k] 4310 == dm_n422) { 4311 mode_lib->vba.DSCFormatFactor = 2; 4312 } else { 4313 mode_lib->vba.DSCFormatFactor = 1; 4314 } 4315 if (locals->RequiresDSC[i][k] == true) { 4316 if (locals->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_2to1) { 4317 if (mode_lib->vba.PixelClockBackEnd[k] / 6.0 / mode_lib->vba.DSCFormatFactor 4318 > (1.0 - mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) * mode_lib->vba.MaxDSCCLK[i]) { 4319 locals->DSCCLKRequiredMoreThanSupported[i] = 4320 true; 4321 } 4322 } else { 4323 if (mode_lib->vba.PixelClockBackEnd[k] / 3.0 / mode_lib->vba.DSCFormatFactor 4324 > (1.0 - mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) * mode_lib->vba.MaxDSCCLK[i]) { 4325 locals->DSCCLKRequiredMoreThanSupported[i] = 4326 true; 4327 } 4328 } 4329 } 4330 } 4331 } 4332 } 4333 } 4334 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 4335 locals->NotEnoughDSCUnits[i] = false; 4336 mode_lib->vba.TotalDSCUnitsRequired = 0.0; 4337 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4338 if (locals->RequiresDSC[i][k] == true) { 4339 if (locals->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_2to1) { 4340 mode_lib->vba.TotalDSCUnitsRequired = 4341 mode_lib->vba.TotalDSCUnitsRequired + 2.0; 4342 } else { 4343 mode_lib->vba.TotalDSCUnitsRequired = 4344 mode_lib->vba.TotalDSCUnitsRequired + 1.0; 4345 } 4346 } 4347 } 4348 if (mode_lib->vba.TotalDSCUnitsRequired > mode_lib->vba.NumberOfDSC) { 4349 locals->NotEnoughDSCUnits[i] = true; 4350 } 4351 } 4352 /*DSC Delay per state*/ 4353 4354 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 4355 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4356 if (mode_lib->vba.BlendingAndTiming[k] != k) { 4357 mode_lib->vba.slices = 0; 4358 } else if (locals->RequiresDSC[i][k] == 0 4359 || locals->RequiresDSC[i][k] == false) { 4360 mode_lib->vba.slices = 0; 4361 } else if (mode_lib->vba.PixelClockBackEnd[k] > 3200.0) { 4362 mode_lib->vba.slices = dml_ceil( 4363 mode_lib->vba.PixelClockBackEnd[k] / 400.0, 4364 4.0); 4365 } else if (mode_lib->vba.PixelClockBackEnd[k] > 1360.0) { 4366 mode_lib->vba.slices = 8.0; 4367 } else if (mode_lib->vba.PixelClockBackEnd[k] > 680.0) { 4368 mode_lib->vba.slices = 4.0; 4369 } else if (mode_lib->vba.PixelClockBackEnd[k] > 340.0) { 4370 mode_lib->vba.slices = 2.0; 4371 } else { 4372 mode_lib->vba.slices = 1.0; 4373 } 4374 if (locals->OutputBppPerState[i][k] == BPP_BLENDED_PIPE 4375 || locals->OutputBppPerState[i][k] == BPP_INVALID) { 4376 mode_lib->vba.bpp = 0.0; 4377 } else { 4378 mode_lib->vba.bpp = locals->OutputBppPerState[i][k]; 4379 } 4380 if (locals->RequiresDSC[i][k] == true && mode_lib->vba.bpp != 0.0) { 4381 if (locals->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_disabled) { 4382 locals->DSCDelayPerState[i][k] = 4383 dscceComputeDelay( 4384 mode_lib->vba.DSCInputBitPerComponent[k], 4385 mode_lib->vba.bpp, 4386 dml_ceil( 4387 mode_lib->vba.HActive[k] 4388 / mode_lib->vba.slices, 4389 1.0), 4390 mode_lib->vba.slices, 4391 mode_lib->vba.OutputFormat[k]) 4392 + dscComputeDelay( 4393 mode_lib->vba.OutputFormat[k]); 4394 } else { 4395 locals->DSCDelayPerState[i][k] = 4396 2.0 * (dscceComputeDelay( 4397 mode_lib->vba.DSCInputBitPerComponent[k], 4398 mode_lib->vba.bpp, 4399 dml_ceil(mode_lib->vba.HActive[k] / mode_lib->vba.slices, 1.0), 4400 mode_lib->vba.slices / 2, 4401 mode_lib->vba.OutputFormat[k]) 4402 + dscComputeDelay(mode_lib->vba.OutputFormat[k])); 4403 } 4404 locals->DSCDelayPerState[i][k] = 4405 locals->DSCDelayPerState[i][k] * mode_lib->vba.PixelClock[k] / mode_lib->vba.PixelClockBackEnd[k]; 4406 } else { 4407 locals->DSCDelayPerState[i][k] = 0.0; 4408 } 4409 } 4410 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4411 for (m = 0; m <= mode_lib->vba.NumberOfActivePlanes - 1; m++) { 4412 for (j = 0; j <= mode_lib->vba.NumberOfActivePlanes - 1; j++) { 4413 if (mode_lib->vba.BlendingAndTiming[k] == m && locals->RequiresDSC[i][m] == true) 4414 locals->DSCDelayPerState[i][k] = locals->DSCDelayPerState[i][m]; 4415 } 4416 } 4417 } 4418 } 4419 4420 //Prefetch Check 4421 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 4422 for (j = 0; j < 2; j++) { 4423 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4424 if (locals->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_2to1) 4425 locals->SwathWidthYPerState[i][j][k] = dml_min(locals->SwathWidthYSingleDPP[k], dml_round(locals->HActive[k] / 2 * locals->HRatio[k])); 4426 else 4427 locals->SwathWidthYPerState[i][j][k] = locals->SwathWidthYSingleDPP[k] / locals->NoOfDPP[i][j][k]; 4428 locals->SwathWidthGranularityY = 256 / dml_ceil(locals->BytePerPixelInDETY[k], 1) / locals->MaxSwathHeightY[k]; 4429 locals->RoundedUpMaxSwathSizeBytesY = (dml_ceil(locals->SwathWidthYPerState[i][j][k] - 1, locals->SwathWidthGranularityY) 4430 + locals->SwathWidthGranularityY) * locals->BytePerPixelInDETY[k] * locals->MaxSwathHeightY[k]; 4431 if (locals->SourcePixelFormat[k] == dm_420_10) { 4432 locals->RoundedUpMaxSwathSizeBytesY = dml_ceil(locals->RoundedUpMaxSwathSizeBytesY, 256) + 256; 4433 } 4434 if (locals->MaxSwathHeightC[k] > 0) { 4435 locals->SwathWidthGranularityC = 256 / dml_ceil(locals->BytePerPixelInDETC[k], 2) / locals->MaxSwathHeightC[k]; 4436 4437 locals->RoundedUpMaxSwathSizeBytesC = (dml_ceil(locals->SwathWidthYPerState[i][j][k] / 2 - 1, locals->SwathWidthGranularityC) 4438 + locals->SwathWidthGranularityC) * locals->BytePerPixelInDETC[k] * locals->MaxSwathHeightC[k]; 4439 } 4440 if (locals->SourcePixelFormat[k] == dm_420_10) { 4441 locals->RoundedUpMaxSwathSizeBytesC = dml_ceil(locals->RoundedUpMaxSwathSizeBytesC, 256) + 256; 4442 } else { 4443 locals->RoundedUpMaxSwathSizeBytesC = 0; 4444 } 4445 4446 if (locals->RoundedUpMaxSwathSizeBytesY + locals->RoundedUpMaxSwathSizeBytesC <= locals->DETBufferSizeInKByte[0] * 1024 / 2) { 4447 locals->SwathHeightYPerState[i][j][k] = locals->MaxSwathHeightY[k]; 4448 locals->SwathHeightCPerState[i][j][k] = locals->MaxSwathHeightC[k]; 4449 } else { 4450 locals->SwathHeightYPerState[i][j][k] = locals->MinSwathHeightY[k]; 4451 locals->SwathHeightCPerState[i][j][k] = locals->MinSwathHeightC[k]; 4452 } 4453 4454 if (locals->BytePerPixelInDETC[k] == 0) { 4455 locals->LinesInDETLuma = locals->DETBufferSizeInKByte[0] * 1024 / locals->BytePerPixelInDETY[k] / locals->SwathWidthYPerState[i][j][k]; 4456 locals->LinesInDETChroma = 0; 4457 } else if (locals->SwathHeightYPerState[i][j][k] <= locals->SwathHeightCPerState[i][j][k]) { 4458 locals->LinesInDETLuma = locals->DETBufferSizeInKByte[0] * 1024 / 2 / locals->BytePerPixelInDETY[k] / 4459 locals->SwathWidthYPerState[i][j][k]; 4460 locals->LinesInDETChroma = locals->DETBufferSizeInKByte[0] * 1024 / 2 / locals->BytePerPixelInDETC[k] / (locals->SwathWidthYPerState[i][j][k] / 2); 4461 } else { 4462 locals->LinesInDETLuma = locals->DETBufferSizeInKByte[0] * 1024 * 2 / 3 / locals->BytePerPixelInDETY[k] / locals->SwathWidthYPerState[i][j][k]; 4463 locals->LinesInDETChroma = locals->DETBufferSizeInKByte[0] * 1024 / 3 / locals->BytePerPixelInDETY[k] / (locals->SwathWidthYPerState[i][j][k] / 2); 4464 } 4465 4466 locals->EffectiveLBLatencyHidingSourceLinesLuma = dml_min(locals->MaxLineBufferLines, 4467 dml_floor(locals->LineBufferSize / locals->LBBitPerPixel[k] / (locals->SwathWidthYPerState[i][j][k] 4468 / dml_max(locals->HRatio[k], 1)), 1)) - (locals->vtaps[k] - 1); 4469 4470 locals->EffectiveLBLatencyHidingSourceLinesChroma = dml_min(locals->MaxLineBufferLines, 4471 dml_floor(locals->LineBufferSize / locals->LBBitPerPixel[k] 4472 / (locals->SwathWidthYPerState[i][j][k] / 2 4473 / dml_max(locals->HRatio[k] / 2, 1)), 1)) - (locals->VTAPsChroma[k] - 1); 4474 4475 locals->EffectiveDETLBLinesLuma = dml_floor(locals->LinesInDETLuma + dml_min( 4476 locals->LinesInDETLuma * locals->RequiredDISPCLK[i][j] * locals->BytePerPixelInDETY[k] * 4477 locals->PSCL_FACTOR[k] / locals->ReturnBWPerState[i][0], 4478 locals->EffectiveLBLatencyHidingSourceLinesLuma), 4479 locals->SwathHeightYPerState[i][j][k]); 4480 4481 locals->EffectiveDETLBLinesChroma = dml_floor(locals->LinesInDETChroma + dml_min( 4482 locals->LinesInDETChroma * locals->RequiredDISPCLK[i][j] * locals->BytePerPixelInDETC[k] * 4483 locals->PSCL_FACTOR_CHROMA[k] / locals->ReturnBWPerState[i][0], 4484 locals->EffectiveLBLatencyHidingSourceLinesChroma), 4485 locals->SwathHeightCPerState[i][j][k]); 4486 4487 if (locals->BytePerPixelInDETC[k] == 0) { 4488 locals->UrgentLatencySupportUsPerState[i][j][k] = locals->EffectiveDETLBLinesLuma * (locals->HTotal[k] / locals->PixelClock[k]) 4489 / locals->VRatio[k] - locals->EffectiveDETLBLinesLuma * locals->SwathWidthYPerState[i][j][k] * 4490 dml_ceil(locals->BytePerPixelInDETY[k], 1) / (locals->ReturnBWPerState[i][0] / locals->NoOfDPP[i][j][k]); 4491 } else { 4492 locals->UrgentLatencySupportUsPerState[i][j][k] = dml_min( 4493 locals->EffectiveDETLBLinesLuma * (locals->HTotal[k] / locals->PixelClock[k]) 4494 / locals->VRatio[k] - locals->EffectiveDETLBLinesLuma * locals->SwathWidthYPerState[i][j][k] * 4495 dml_ceil(locals->BytePerPixelInDETY[k], 1) / (locals->ReturnBWPerState[i][0] / locals->NoOfDPP[i][j][k]), 4496 locals->EffectiveDETLBLinesChroma * (locals->HTotal[k] / locals->PixelClock[k]) / (locals->VRatio[k] / 2) - 4497 locals->EffectiveDETLBLinesChroma * locals->SwathWidthYPerState[i][j][k] / 2 * 4498 dml_ceil(locals->BytePerPixelInDETC[k], 2) / (locals->ReturnBWPerState[i][0] / locals->NoOfDPP[i][j][k])); 4499 } 4500 } 4501 } 4502 } 4503 4504 for (i = 0; i <= locals->soc.num_states; i++) { 4505 for (j = 0; j < 2; j++) { 4506 locals->UrgentLatencySupport[i][j] = true; 4507 for (k = 0; k < locals->NumberOfActivePlanes; k++) { 4508 if (locals->UrgentLatencySupportUsPerState[i][j][k] < locals->UrgentLatency) 4509 locals->UrgentLatencySupport[i][j] = false; 4510 } 4511 } 4512 } 4513 4514 4515 /*Prefetch Check*/ 4516 for (i = 0; i <= locals->soc.num_states; i++) { 4517 for (j = 0; j < 2; j++) { 4518 locals->TotalNumberOfDCCActiveDPP[i][j] = 0; 4519 for (k = 0; k < locals->NumberOfActivePlanes; k++) { 4520 if (locals->DCCEnable[k] == true) { 4521 locals->TotalNumberOfDCCActiveDPP[i][j] = 4522 locals->TotalNumberOfDCCActiveDPP[i][j] + locals->NoOfDPP[i][j][k]; 4523 } 4524 } 4525 } 4526 } 4527 4528 CalculateMinAndMaxPrefetchMode(locals->AllowDRAMSelfRefreshOrDRAMClockChangeInVblank, &locals->MinPrefetchMode, &locals->MaxPrefetchMode); 4529 4530 locals->MaxTotalVActiveRDBandwidth = 0; 4531 for (k = 0; k < locals->NumberOfActivePlanes; k++) { 4532 locals->MaxTotalVActiveRDBandwidth = locals->MaxTotalVActiveRDBandwidth + locals->ReadBandwidth[k]; 4533 } 4534 4535 for (i = 0; i <= locals->soc.num_states; i++) { 4536 for (j = 0; j < 2; j++) { 4537 for (k = 0; k < locals->NumberOfActivePlanes; k++) { 4538 locals->NoOfDPPThisState[k] = locals->NoOfDPP[i][j][k]; 4539 locals->RequiredDPPCLKThisState[k] = locals->RequiredDPPCLK[i][j][k]; 4540 locals->SwathHeightYThisState[k] = locals->SwathHeightYPerState[i][j][k]; 4541 locals->SwathHeightCThisState[k] = locals->SwathHeightCPerState[i][j][k]; 4542 locals->SwathWidthYThisState[k] = locals->SwathWidthYPerState[i][j][k]; 4543 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0] = dml_max( 4544 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0], 4545 mode_lib->vba.PixelClock[k] / 16.0); 4546 if (mode_lib->vba.BytePerPixelInDETC[k] == 0.0) { 4547 if (mode_lib->vba.VRatio[k] <= 1.0) { 4548 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0] = 4549 dml_max( 4550 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0], 4551 1.1 4552 * dml_ceil( 4553 mode_lib->vba.BytePerPixelInDETY[k], 4554 1.0) 4555 / 64.0 4556 * mode_lib->vba.HRatio[k] 4557 * mode_lib->vba.PixelClock[k] 4558 / mode_lib->vba.NoOfDPP[i][j][k]); 4559 } else { 4560 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0] = 4561 dml_max( 4562 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0], 4563 1.1 4564 * dml_ceil( 4565 mode_lib->vba.BytePerPixelInDETY[k], 4566 1.0) 4567 / 64.0 4568 * mode_lib->vba.PSCL_FACTOR[k] 4569 * mode_lib->vba.RequiredDPPCLK[i][j][k]); 4570 } 4571 } else { 4572 if (mode_lib->vba.VRatio[k] <= 1.0) { 4573 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0] = 4574 dml_max( 4575 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0], 4576 1.1 4577 * dml_ceil( 4578 mode_lib->vba.BytePerPixelInDETY[k], 4579 1.0) 4580 / 32.0 4581 * mode_lib->vba.HRatio[k] 4582 * mode_lib->vba.PixelClock[k] 4583 / mode_lib->vba.NoOfDPP[i][j][k]); 4584 } else { 4585 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0] = 4586 dml_max( 4587 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0], 4588 1.1 4589 * dml_ceil( 4590 mode_lib->vba.BytePerPixelInDETY[k], 4591 1.0) 4592 / 32.0 4593 * mode_lib->vba.PSCL_FACTOR[k] 4594 * mode_lib->vba.RequiredDPPCLK[i][j][k]); 4595 } 4596 if (mode_lib->vba.VRatio[k] / 2.0 <= 1.0) { 4597 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0] = 4598 dml_max( 4599 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0], 4600 1.1 4601 * dml_ceil( 4602 mode_lib->vba.BytePerPixelInDETC[k], 4603 2.0) 4604 / 32.0 4605 * mode_lib->vba.HRatio[k] 4606 / 2.0 4607 * mode_lib->vba.PixelClock[k] 4608 / mode_lib->vba.NoOfDPP[i][j][k]); 4609 } else { 4610 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0] = 4611 dml_max( 4612 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0], 4613 1.1 4614 * dml_ceil( 4615 mode_lib->vba.BytePerPixelInDETC[k], 4616 2.0) 4617 / 32.0 4618 * mode_lib->vba.PSCL_FACTOR_CHROMA[k] 4619 * mode_lib->vba.RequiredDPPCLK[i][j][k]); 4620 } 4621 } 4622 } 4623 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4624 mode_lib->vba.PDEAndMetaPTEBytesPerFrameY = CalculateVMAndRowBytes( 4625 mode_lib, 4626 mode_lib->vba.DCCEnable[k], 4627 mode_lib->vba.Read256BlockHeightY[k], 4628 mode_lib->vba.Read256BlockWidthY[k], 4629 mode_lib->vba.SourcePixelFormat[k], 4630 mode_lib->vba.SurfaceTiling[k], 4631 dml_ceil(mode_lib->vba.BytePerPixelInDETY[k], 1.0), 4632 mode_lib->vba.SourceScan[k], 4633 mode_lib->vba.ViewportWidth[k], 4634 mode_lib->vba.ViewportHeight[k], 4635 mode_lib->vba.SwathWidthYPerState[i][j][k], 4636 mode_lib->vba.GPUVMEnable, 4637 mode_lib->vba.VMMPageSize, 4638 mode_lib->vba.PTEBufferSizeInRequestsLuma, 4639 mode_lib->vba.PDEProcessingBufIn64KBReqs, 4640 mode_lib->vba.PitchY[k], 4641 mode_lib->vba.DCCMetaPitchY[k], 4642 &mode_lib->vba.MacroTileWidthY[k], 4643 &mode_lib->vba.MetaRowBytesY, 4644 &mode_lib->vba.DPTEBytesPerRowY, 4645 &mode_lib->vba.PTEBufferSizeNotExceededY[i][j][k], 4646 &mode_lib->vba.dpte_row_height[k], 4647 &mode_lib->vba.meta_row_height[k]); 4648 mode_lib->vba.PrefetchLinesY[0][0][k] = CalculatePrefetchSourceLines( 4649 mode_lib, 4650 mode_lib->vba.VRatio[k], 4651 mode_lib->vba.vtaps[k], 4652 mode_lib->vba.Interlace[k], 4653 mode_lib->vba.ProgressiveToInterlaceUnitInOPP, 4654 mode_lib->vba.SwathHeightYPerState[i][j][k], 4655 mode_lib->vba.ViewportYStartY[k], 4656 &mode_lib->vba.PrefillY[k], 4657 &mode_lib->vba.MaxNumSwY[k]); 4658 if ((mode_lib->vba.SourcePixelFormat[k] != dm_444_64 4659 && mode_lib->vba.SourcePixelFormat[k] != dm_444_32 4660 && mode_lib->vba.SourcePixelFormat[k] != dm_444_16 4661 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_16 4662 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_8)) { 4663 mode_lib->vba.PDEAndMetaPTEBytesPerFrameC = CalculateVMAndRowBytes( 4664 mode_lib, 4665 mode_lib->vba.DCCEnable[k], 4666 mode_lib->vba.Read256BlockHeightY[k], 4667 mode_lib->vba.Read256BlockWidthY[k], 4668 mode_lib->vba.SourcePixelFormat[k], 4669 mode_lib->vba.SurfaceTiling[k], 4670 dml_ceil(mode_lib->vba.BytePerPixelInDETC[k], 2.0), 4671 mode_lib->vba.SourceScan[k], 4672 mode_lib->vba.ViewportWidth[k] / 2.0, 4673 mode_lib->vba.ViewportHeight[k] / 2.0, 4674 mode_lib->vba.SwathWidthYPerState[i][j][k] / 2.0, 4675 mode_lib->vba.GPUVMEnable, 4676 mode_lib->vba.VMMPageSize, 4677 mode_lib->vba.PTEBufferSizeInRequestsLuma, 4678 mode_lib->vba.PDEProcessingBufIn64KBReqs, 4679 mode_lib->vba.PitchC[k], 4680 0.0, 4681 &mode_lib->vba.MacroTileWidthC[k], 4682 &mode_lib->vba.MetaRowBytesC, 4683 &mode_lib->vba.DPTEBytesPerRowC, 4684 &mode_lib->vba.PTEBufferSizeNotExceededC[i][j][k], 4685 &mode_lib->vba.dpte_row_height_chroma[k], 4686 &mode_lib->vba.meta_row_height_chroma[k]); 4687 mode_lib->vba.PrefetchLinesC[0][0][k] = CalculatePrefetchSourceLines( 4688 mode_lib, 4689 mode_lib->vba.VRatio[k] / 2.0, 4690 mode_lib->vba.VTAPsChroma[k], 4691 mode_lib->vba.Interlace[k], 4692 mode_lib->vba.ProgressiveToInterlaceUnitInOPP, 4693 mode_lib->vba.SwathHeightCPerState[i][j][k], 4694 mode_lib->vba.ViewportYStartC[k], 4695 &mode_lib->vba.PrefillC[k], 4696 &mode_lib->vba.MaxNumSwC[k]); 4697 } else { 4698 mode_lib->vba.PDEAndMetaPTEBytesPerFrameC = 0.0; 4699 mode_lib->vba.MetaRowBytesC = 0.0; 4700 mode_lib->vba.DPTEBytesPerRowC = 0.0; 4701 locals->PrefetchLinesC[0][0][k] = 0.0; 4702 locals->PTEBufferSizeNotExceededC[i][j][k] = true; 4703 locals->PTEBufferSizeInRequestsForLuma = mode_lib->vba.PTEBufferSizeInRequestsLuma + mode_lib->vba.PTEBufferSizeInRequestsChroma; 4704 } 4705 locals->PDEAndMetaPTEBytesPerFrame[0][0][k] = 4706 mode_lib->vba.PDEAndMetaPTEBytesPerFrameY + mode_lib->vba.PDEAndMetaPTEBytesPerFrameC; 4707 locals->MetaRowBytes[0][0][k] = mode_lib->vba.MetaRowBytesY + mode_lib->vba.MetaRowBytesC; 4708 locals->DPTEBytesPerRow[0][0][k] = mode_lib->vba.DPTEBytesPerRowY + mode_lib->vba.DPTEBytesPerRowC; 4709 4710 CalculateActiveRowBandwidth( 4711 mode_lib->vba.GPUVMEnable, 4712 mode_lib->vba.SourcePixelFormat[k], 4713 mode_lib->vba.VRatio[k], 4714 mode_lib->vba.DCCEnable[k], 4715 mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k], 4716 mode_lib->vba.MetaRowBytesY, 4717 mode_lib->vba.MetaRowBytesC, 4718 mode_lib->vba.meta_row_height[k], 4719 mode_lib->vba.meta_row_height_chroma[k], 4720 mode_lib->vba.DPTEBytesPerRowY, 4721 mode_lib->vba.DPTEBytesPerRowC, 4722 mode_lib->vba.dpte_row_height[k], 4723 mode_lib->vba.dpte_row_height_chroma[k], 4724 &mode_lib->vba.meta_row_bw[k], 4725 &mode_lib->vba.dpte_row_bw[k], 4726 &mode_lib->vba.qual_row_bw[k]); 4727 } 4728 mode_lib->vba.ExtraLatency = 4729 mode_lib->vba.UrgentRoundTripAndOutOfOrderLatencyPerState[i] 4730 + (mode_lib->vba.TotalNumberOfActiveDPP[i][j] 4731 * mode_lib->vba.PixelChunkSizeInKByte 4732 + mode_lib->vba.TotalNumberOfDCCActiveDPP[i][j] 4733 * mode_lib->vba.MetaChunkSize) 4734 * 1024.0 4735 / mode_lib->vba.ReturnBWPerState[i][0]; 4736 if (mode_lib->vba.GPUVMEnable == true) { 4737 mode_lib->vba.ExtraLatency = mode_lib->vba.ExtraLatency 4738 + mode_lib->vba.TotalNumberOfActiveDPP[i][j] 4739 * mode_lib->vba.PTEGroupSize 4740 / mode_lib->vba.ReturnBWPerState[i][0]; 4741 } 4742 mode_lib->vba.TimeCalc = 24.0 / mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0]; 4743 4744 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4745 if (mode_lib->vba.BlendingAndTiming[k] == k) { 4746 if (mode_lib->vba.WritebackEnable[k] == true) { 4747 locals->WritebackDelay[i][k] = mode_lib->vba.WritebackLatency 4748 + CalculateWriteBackDelay( 4749 mode_lib->vba.WritebackPixelFormat[k], 4750 mode_lib->vba.WritebackHRatio[k], 4751 mode_lib->vba.WritebackVRatio[k], 4752 mode_lib->vba.WritebackLumaHTaps[k], 4753 mode_lib->vba.WritebackLumaVTaps[k], 4754 mode_lib->vba.WritebackChromaHTaps[k], 4755 mode_lib->vba.WritebackChromaVTaps[k], 4756 mode_lib->vba.WritebackDestinationWidth[k]) / locals->RequiredDISPCLK[i][j]; 4757 } else { 4758 locals->WritebackDelay[i][k] = 0.0; 4759 } 4760 for (m = 0; m <= mode_lib->vba.NumberOfActivePlanes - 1; m++) { 4761 if (mode_lib->vba.BlendingAndTiming[m] == k 4762 && mode_lib->vba.WritebackEnable[m] 4763 == true) { 4764 locals->WritebackDelay[i][k] = dml_max(locals->WritebackDelay[i][k], 4765 mode_lib->vba.WritebackLatency + CalculateWriteBackDelay( 4766 mode_lib->vba.WritebackPixelFormat[m], 4767 mode_lib->vba.WritebackHRatio[m], 4768 mode_lib->vba.WritebackVRatio[m], 4769 mode_lib->vba.WritebackLumaHTaps[m], 4770 mode_lib->vba.WritebackLumaVTaps[m], 4771 mode_lib->vba.WritebackChromaHTaps[m], 4772 mode_lib->vba.WritebackChromaVTaps[m], 4773 mode_lib->vba.WritebackDestinationWidth[m]) / locals->RequiredDISPCLK[i][j]); 4774 } 4775 } 4776 } 4777 } 4778 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4779 for (m = 0; m <= mode_lib->vba.NumberOfActivePlanes - 1; m++) { 4780 if (mode_lib->vba.BlendingAndTiming[k] == m) { 4781 locals->WritebackDelay[i][k] = locals->WritebackDelay[i][m]; 4782 } 4783 } 4784 } 4785 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4786 for (m = 0; m < locals->NumberOfCursors[k]; m++) 4787 locals->cursor_bw[k] = locals->NumberOfCursors[k] * locals->CursorWidth[k][m] * locals->CursorBPP[k][m] 4788 / 8 / (locals->HTotal[k] / locals->PixelClock[k]) * locals->VRatio[k]; 4789 } 4790 4791 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4792 locals->MaximumVStartup[0][0][k] = mode_lib->vba.VTotal[k] - mode_lib->vba.VActive[k] 4793 - dml_max(1.0, dml_ceil(locals->WritebackDelay[i][k] / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]), 1.0)); 4794 } 4795 4796 mode_lib->vba.NextPrefetchMode = mode_lib->vba.MinPrefetchMode; 4797 do { 4798 mode_lib->vba.PrefetchMode[i][j] = mode_lib->vba.NextPrefetchMode; 4799 mode_lib->vba.NextPrefetchMode = mode_lib->vba.NextPrefetchMode + 1; 4800 4801 mode_lib->vba.TWait = CalculateTWait( 4802 mode_lib->vba.PrefetchMode[i][j], 4803 mode_lib->vba.DRAMClockChangeLatency, 4804 mode_lib->vba.UrgentLatency, 4805 mode_lib->vba.SREnterPlusExitTime); 4806 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4807 4808 if (mode_lib->vba.XFCEnabled[k] == true) { 4809 mode_lib->vba.XFCRemoteSurfaceFlipDelay = 4810 CalculateRemoteSurfaceFlipDelay( 4811 mode_lib, 4812 mode_lib->vba.VRatio[k], 4813 locals->SwathWidthYPerState[i][j][k], 4814 dml_ceil(locals->BytePerPixelInDETY[k], 1.0), 4815 mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k], 4816 mode_lib->vba.XFCTSlvVupdateOffset, 4817 mode_lib->vba.XFCTSlvVupdateWidth, 4818 mode_lib->vba.XFCTSlvVreadyOffset, 4819 mode_lib->vba.XFCXBUFLatencyTolerance, 4820 mode_lib->vba.XFCFillBWOverhead, 4821 mode_lib->vba.XFCSlvChunkSize, 4822 mode_lib->vba.XFCBusTransportTime, 4823 mode_lib->vba.TimeCalc, 4824 mode_lib->vba.TWait, 4825 &mode_lib->vba.SrcActiveDrainRate, 4826 &mode_lib->vba.TInitXFill, 4827 &mode_lib->vba.TslvChk); 4828 } else { 4829 mode_lib->vba.XFCRemoteSurfaceFlipDelay = 0.0; 4830 } 4831 4832 CalculateDelayAfterScaler(mode_lib, mode_lib->vba.ReturnBWPerState[i][0], mode_lib->vba.ReadBandwidthLuma[k], mode_lib->vba.ReadBandwidthChroma[k], mode_lib->vba.MaxTotalVActiveRDBandwidth, 4833 mode_lib->vba.DisplayPipeLineDeliveryTimeLuma[k], mode_lib->vba.DisplayPipeLineDeliveryTimeChroma[k], 4834 mode_lib->vba.RequiredDPPCLK[i][j][k], mode_lib->vba.RequiredDISPCLK[i][j], mode_lib->vba.PixelClock[k], mode_lib->vba.DSCDelayPerState[i][k], mode_lib->vba.NoOfDPP[i][j][k], mode_lib->vba.ScalerEnabled[k], mode_lib->vba.NumberOfCursors[k], 4835 mode_lib->vba.DPPCLKDelaySubtotal, mode_lib->vba.DPPCLKDelaySCL, mode_lib->vba.DPPCLKDelaySCLLBOnly, mode_lib->vba.DPPCLKDelayCNVCFormater, mode_lib->vba.DPPCLKDelayCNVCCursor, mode_lib->vba.DISPCLKDelaySubtotal, 4836 mode_lib->vba.SwathWidthYPerState[i][j][k] / mode_lib->vba.HRatio[k], mode_lib->vba.OutputFormat[k], mode_lib->vba.HTotal[k], 4837 mode_lib->vba.SwathWidthYSingleDPP[k], mode_lib->vba.BytePerPixelInDETY[k], mode_lib->vba.BytePerPixelInDETC[k], mode_lib->vba.SwathHeightYThisState[k], mode_lib->vba.SwathHeightCThisState[k], mode_lib->vba.Interlace[k], mode_lib->vba.ProgressiveToInterlaceUnitInOPP, 4838 &mode_lib->vba.DSTXAfterScaler[k], &mode_lib->vba.DSTYAfterScaler[k]); 4839 4840 mode_lib->vba.IsErrorResult[i][j][k] = 4841 CalculatePrefetchSchedule( 4842 mode_lib, 4843 mode_lib->vba.RequiredDPPCLK[i][j][k], 4844 mode_lib->vba.RequiredDISPCLK[i][j], 4845 mode_lib->vba.PixelClock[k], 4846 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0], 4847 mode_lib->vba.NoOfDPP[i][j][k], 4848 mode_lib->vba.NumberOfCursors[k], 4849 mode_lib->vba.VTotal[k] 4850 - mode_lib->vba.VActive[k], 4851 mode_lib->vba.HTotal[k], 4852 mode_lib->vba.MaxInterDCNTileRepeaters, 4853 mode_lib->vba.MaximumVStartup[0][0][k], 4854 mode_lib->vba.GPUVMMaxPageTableLevels, 4855 mode_lib->vba.GPUVMEnable, 4856 mode_lib->vba.DynamicMetadataEnable[k], 4857 mode_lib->vba.DynamicMetadataLinesBeforeActiveRequired[k], 4858 mode_lib->vba.DynamicMetadataTransmittedBytes[k], 4859 mode_lib->vba.DCCEnable[k], 4860 mode_lib->vba.UrgentLatencyPixelDataOnly, 4861 mode_lib->vba.ExtraLatency, 4862 mode_lib->vba.TimeCalc, 4863 mode_lib->vba.PDEAndMetaPTEBytesPerFrame[0][0][k], 4864 mode_lib->vba.MetaRowBytes[0][0][k], 4865 mode_lib->vba.DPTEBytesPerRow[0][0][k], 4866 mode_lib->vba.PrefetchLinesY[0][0][k], 4867 mode_lib->vba.SwathWidthYPerState[i][j][k], 4868 mode_lib->vba.BytePerPixelInDETY[k], 4869 mode_lib->vba.PrefillY[k], 4870 mode_lib->vba.MaxNumSwY[k], 4871 mode_lib->vba.PrefetchLinesC[0][0][k], 4872 mode_lib->vba.BytePerPixelInDETC[k], 4873 mode_lib->vba.PrefillC[k], 4874 mode_lib->vba.MaxNumSwC[k], 4875 mode_lib->vba.SwathHeightYPerState[i][j][k], 4876 mode_lib->vba.SwathHeightCPerState[i][j][k], 4877 mode_lib->vba.TWait, 4878 mode_lib->vba.XFCEnabled[k], 4879 mode_lib->vba.XFCRemoteSurfaceFlipDelay, 4880 mode_lib->vba.Interlace[k], 4881 mode_lib->vba.ProgressiveToInterlaceUnitInOPP, 4882 mode_lib->vba.DSTXAfterScaler[k], 4883 mode_lib->vba.DSTYAfterScaler[k], 4884 &mode_lib->vba.LineTimesForPrefetch[k], 4885 &mode_lib->vba.PrefetchBW[k], 4886 &mode_lib->vba.LinesForMetaPTE[k], 4887 &mode_lib->vba.LinesForMetaAndDPTERow[k], 4888 &mode_lib->vba.VRatioPreY[i][j][k], 4889 &mode_lib->vba.VRatioPreC[i][j][k], 4890 &mode_lib->vba.RequiredPrefetchPixelDataBWLuma[i][j][k], 4891 &mode_lib->vba.Tno_bw[k], 4892 &mode_lib->vba.VUpdateOffsetPix[k], 4893 &mode_lib->vba.VUpdateWidthPix[k], 4894 &mode_lib->vba.VReadyOffsetPix[k]); 4895 } 4896 mode_lib->vba.MaximumReadBandwidthWithoutPrefetch = 0.0; 4897 mode_lib->vba.MaximumReadBandwidthWithPrefetch = 0.0; 4898 locals->prefetch_vm_bw_valid = true; 4899 locals->prefetch_row_bw_valid = true; 4900 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4901 if (locals->PDEAndMetaPTEBytesPerFrame[0][0][k] == 0) 4902 locals->prefetch_vm_bw[k] = 0; 4903 else if (locals->LinesForMetaPTE[k] > 0) 4904 locals->prefetch_vm_bw[k] = locals->PDEAndMetaPTEBytesPerFrame[0][0][k] 4905 / (locals->LinesForMetaPTE[k] * locals->HTotal[k] / locals->PixelClock[k]); 4906 else { 4907 locals->prefetch_vm_bw[k] = 0; 4908 locals->prefetch_vm_bw_valid = false; 4909 } 4910 if (locals->MetaRowBytes[0][0][k] + locals->DPTEBytesPerRow[0][0][k] == 0) 4911 locals->prefetch_row_bw[k] = 0; 4912 else if (locals->LinesForMetaAndDPTERow[k] > 0) 4913 locals->prefetch_row_bw[k] = (locals->MetaRowBytes[0][0][k] + locals->DPTEBytesPerRow[0][0][k]) 4914 / (locals->LinesForMetaAndDPTERow[k] * locals->HTotal[k] / locals->PixelClock[k]); 4915 else { 4916 locals->prefetch_row_bw[k] = 0; 4917 locals->prefetch_row_bw_valid = false; 4918 } 4919 4920 mode_lib->vba.MaximumReadBandwidthWithoutPrefetch = mode_lib->vba.MaximumReadBandwidthWithPrefetch 4921 + 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]; 4922 mode_lib->vba.MaximumReadBandwidthWithPrefetch = 4923 mode_lib->vba.MaximumReadBandwidthWithPrefetch 4924 + mode_lib->vba.cursor_bw[k] 4925 + dml_max3( 4926 mode_lib->vba.prefetch_vm_bw[k], 4927 mode_lib->vba.prefetch_row_bw[k], 4928 dml_max(mode_lib->vba.ReadBandwidth[k], 4929 mode_lib->vba.RequiredPrefetchPixelDataBWLuma[i][j][k]) 4930 + mode_lib->vba.meta_row_bw[k] + mode_lib->vba.dpte_row_bw[k]); 4931 } 4932 locals->BandwidthWithoutPrefetchSupported[i][0] = true; 4933 if (mode_lib->vba.MaximumReadBandwidthWithoutPrefetch > locals->ReturnBWPerState[i][0]) { 4934 locals->BandwidthWithoutPrefetchSupported[i][0] = false; 4935 } 4936 4937 locals->PrefetchSupported[i][j] = true; 4938 if (mode_lib->vba.MaximumReadBandwidthWithPrefetch > locals->ReturnBWPerState[i][0]) { 4939 locals->PrefetchSupported[i][j] = false; 4940 } 4941 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4942 if (locals->LineTimesForPrefetch[k] < 2.0 4943 || locals->LinesForMetaPTE[k] >= 8.0 4944 || locals->LinesForMetaAndDPTERow[k] >= 16.0 4945 || mode_lib->vba.IsErrorResult[i][j][k] == true) { 4946 locals->PrefetchSupported[i][j] = false; 4947 } 4948 } 4949 locals->VRatioInPrefetchSupported[i][j] = true; 4950 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4951 if (locals->VRatioPreY[i][j][k] > 4.0 4952 || locals->VRatioPreC[i][j][k] > 4.0 4953 || mode_lib->vba.IsErrorResult[i][j][k] == true) { 4954 locals->VRatioInPrefetchSupported[i][j] = false; 4955 } 4956 } 4957 } while ((locals->PrefetchSupported[i][j] != true || locals->VRatioInPrefetchSupported[i][j] != true) 4958 && mode_lib->vba.NextPrefetchMode < mode_lib->vba.MaxPrefetchMode); 4959 4960 if (mode_lib->vba.PrefetchSupported[i][j] == true 4961 && mode_lib->vba.VRatioInPrefetchSupported[i][j] == true) { 4962 mode_lib->vba.BandwidthAvailableForImmediateFlip = 4963 mode_lib->vba.ReturnBWPerState[i][0]; 4964 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4965 mode_lib->vba.BandwidthAvailableForImmediateFlip = 4966 mode_lib->vba.BandwidthAvailableForImmediateFlip 4967 - mode_lib->vba.cursor_bw[k] 4968 - dml_max( 4969 mode_lib->vba.ReadBandwidth[k] + mode_lib->vba.qual_row_bw[k], 4970 mode_lib->vba.PrefetchBW[k]); 4971 } 4972 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4973 mode_lib->vba.ImmediateFlipBytes[k] = 0.0; 4974 if ((mode_lib->vba.SourcePixelFormat[k] != dm_420_8 4975 && mode_lib->vba.SourcePixelFormat[k] != dm_420_10)) { 4976 mode_lib->vba.ImmediateFlipBytes[k] = 4977 mode_lib->vba.PDEAndMetaPTEBytesPerFrame[0][0][k] 4978 + mode_lib->vba.MetaRowBytes[0][0][k] 4979 + mode_lib->vba.DPTEBytesPerRow[0][0][k]; 4980 } 4981 } 4982 mode_lib->vba.TotImmediateFlipBytes = 0.0; 4983 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4984 if ((mode_lib->vba.SourcePixelFormat[k] != dm_420_8 4985 && mode_lib->vba.SourcePixelFormat[k] != dm_420_10)) { 4986 mode_lib->vba.TotImmediateFlipBytes = 4987 mode_lib->vba.TotImmediateFlipBytes 4988 + mode_lib->vba.ImmediateFlipBytes[k]; 4989 } 4990 } 4991 4992 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4993 CalculateFlipSchedule( 4994 mode_lib, 4995 mode_lib->vba.ExtraLatency, 4996 mode_lib->vba.UrgentLatencyPixelDataOnly, 4997 mode_lib->vba.GPUVMMaxPageTableLevels, 4998 mode_lib->vba.GPUVMEnable, 4999 mode_lib->vba.BandwidthAvailableForImmediateFlip, 5000 mode_lib->vba.TotImmediateFlipBytes, 5001 mode_lib->vba.SourcePixelFormat[k], 5002 mode_lib->vba.ImmediateFlipBytes[k], 5003 mode_lib->vba.HTotal[k] 5004 / mode_lib->vba.PixelClock[k], 5005 mode_lib->vba.VRatio[k], 5006 mode_lib->vba.Tno_bw[k], 5007 mode_lib->vba.PDEAndMetaPTEBytesPerFrame[0][0][k], 5008 mode_lib->vba.MetaRowBytes[0][0][k], 5009 mode_lib->vba.DPTEBytesPerRow[0][0][k], 5010 mode_lib->vba.DCCEnable[k], 5011 mode_lib->vba.dpte_row_height[k], 5012 mode_lib->vba.meta_row_height[k], 5013 mode_lib->vba.qual_row_bw[k], 5014 &mode_lib->vba.DestinationLinesToRequestVMInImmediateFlip[k], 5015 &mode_lib->vba.DestinationLinesToRequestRowInImmediateFlip[k], 5016 &mode_lib->vba.final_flip_bw[k], 5017 &mode_lib->vba.ImmediateFlipSupportedForPipe[k]); 5018 } 5019 mode_lib->vba.total_dcn_read_bw_with_flip = 0.0; 5020 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 5021 mode_lib->vba.total_dcn_read_bw_with_flip = 5022 mode_lib->vba.total_dcn_read_bw_with_flip 5023 + mode_lib->vba.cursor_bw[k] 5024 + dml_max3( 5025 mode_lib->vba.prefetch_vm_bw[k], 5026 mode_lib->vba.prefetch_row_bw[k], 5027 mode_lib->vba.final_flip_bw[k] 5028 + dml_max( 5029 mode_lib->vba.ReadBandwidth[k], 5030 mode_lib->vba.RequiredPrefetchPixelDataBWLuma[i][j][k])); 5031 } 5032 mode_lib->vba.ImmediateFlipSupportedForState[i][j] = true; 5033 if (mode_lib->vba.total_dcn_read_bw_with_flip 5034 > mode_lib->vba.ReturnBWPerState[i][0]) { 5035 mode_lib->vba.ImmediateFlipSupportedForState[i][j] = false; 5036 } 5037 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 5038 if (mode_lib->vba.ImmediateFlipSupportedForPipe[k] == false) { 5039 mode_lib->vba.ImmediateFlipSupportedForState[i][j] = false; 5040 } 5041 } 5042 } else { 5043 mode_lib->vba.ImmediateFlipSupportedForState[i][j] = false; 5044 } 5045 } 5046 } 5047 5048 /*Vertical Active BW support*/ 5049 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 5050 mode_lib->vba.MaxTotalVerticalActiveAvailableBandwidth[i][0] = dml_min(mode_lib->vba.ReturnBusWidth * 5051 mode_lib->vba.DCFCLKPerState[i], mode_lib->vba.FabricAndDRAMBandwidthPerState[i] * 1000) * 5052 mode_lib->vba.MaxAveragePercentOfIdealDRAMBWDisplayCanUseInNormalSystemOperation / 100; 5053 if (mode_lib->vba.MaxTotalVActiveRDBandwidth <= mode_lib->vba.MaxTotalVerticalActiveAvailableBandwidth[i][0]) 5054 mode_lib->vba.TotalVerticalActiveBandwidthSupport[i][0] = true; 5055 else 5056 mode_lib->vba.TotalVerticalActiveBandwidthSupport[i][0] = false; 5057 } 5058 5059 /*PTE Buffer Size Check*/ 5060 5061 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 5062 for (j = 0; j < 2; j++) { 5063 locals->PTEBufferSizeNotExceeded[i][j] = true; 5064 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 5065 if (locals->PTEBufferSizeNotExceededY[i][j][k] == false 5066 || locals->PTEBufferSizeNotExceededC[i][j][k] == false) { 5067 locals->PTEBufferSizeNotExceeded[i][j] = false; 5068 } 5069 } 5070 } 5071 } 5072 /*Cursor Support Check*/ 5073 mode_lib->vba.CursorSupport = true; 5074 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 5075 for (j = 0; j < 2; j++) { 5076 if (mode_lib->vba.CursorWidth[k][j] > 0.0) { 5077 if (dml_floor( 5078 dml_floor( 5079 mode_lib->vba.CursorBufferSize 5080 - mode_lib->vba.CursorChunkSize, 5081 mode_lib->vba.CursorChunkSize) * 1024.0 5082 / (mode_lib->vba.CursorWidth[k][j] 5083 * mode_lib->vba.CursorBPP[k][j] 5084 / 8.0), 5085 1.0) 5086 * (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]) 5087 / mode_lib->vba.VRatio[k] < mode_lib->vba.UrgentLatencyPixelDataOnly 5088 || (mode_lib->vba.CursorBPP[k][j] == 64.0 5089 && mode_lib->vba.Cursor64BppSupport == false)) { 5090 mode_lib->vba.CursorSupport = false; 5091 } 5092 } 5093 } 5094 } 5095 /*Valid Pitch Check*/ 5096 5097 mode_lib->vba.PitchSupport = true; 5098 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 5099 locals->AlignedYPitch[k] = dml_ceil( 5100 dml_max(mode_lib->vba.PitchY[k], mode_lib->vba.ViewportWidth[k]), 5101 locals->MacroTileWidthY[k]); 5102 if (locals->AlignedYPitch[k] > mode_lib->vba.PitchY[k]) { 5103 mode_lib->vba.PitchSupport = false; 5104 } 5105 if (mode_lib->vba.DCCEnable[k] == true) { 5106 locals->AlignedDCCMetaPitch[k] = dml_ceil( 5107 dml_max( 5108 mode_lib->vba.DCCMetaPitchY[k], 5109 mode_lib->vba.ViewportWidth[k]), 5110 64.0 * locals->Read256BlockWidthY[k]); 5111 } else { 5112 locals->AlignedDCCMetaPitch[k] = mode_lib->vba.DCCMetaPitchY[k]; 5113 } 5114 if (locals->AlignedDCCMetaPitch[k] > mode_lib->vba.DCCMetaPitchY[k]) { 5115 mode_lib->vba.PitchSupport = false; 5116 } 5117 if (mode_lib->vba.SourcePixelFormat[k] != dm_444_64 5118 && mode_lib->vba.SourcePixelFormat[k] != dm_444_32 5119 && mode_lib->vba.SourcePixelFormat[k] != dm_444_16 5120 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_16 5121 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_8) { 5122 locals->AlignedCPitch[k] = dml_ceil( 5123 dml_max( 5124 mode_lib->vba.PitchC[k], 5125 mode_lib->vba.ViewportWidth[k] / 2.0), 5126 locals->MacroTileWidthC[k]); 5127 } else { 5128 locals->AlignedCPitch[k] = mode_lib->vba.PitchC[k]; 5129 } 5130 if (locals->AlignedCPitch[k] > mode_lib->vba.PitchC[k]) { 5131 mode_lib->vba.PitchSupport = false; 5132 } 5133 } 5134 /*Mode Support, Voltage State and SOC Configuration*/ 5135 5136 for (i = mode_lib->vba.soc.num_states; i >= 0; i--) { 5137 for (j = 0; j < 2; j++) { 5138 enum dm_validation_status status = DML_VALIDATION_OK; 5139 5140 if (mode_lib->vba.ScaleRatioAndTapsSupport != true) { 5141 status = DML_FAIL_SCALE_RATIO_TAP; 5142 } else if (mode_lib->vba.SourceFormatPixelAndScanSupport != true) { 5143 status = DML_FAIL_SOURCE_PIXEL_FORMAT; 5144 } else if (locals->ViewportSizeSupport[i][0] != true) { 5145 status = DML_FAIL_VIEWPORT_SIZE; 5146 } else if (locals->DIOSupport[i] != true) { 5147 status = DML_FAIL_DIO_SUPPORT; 5148 } else if (locals->NotEnoughDSCUnits[i] != false) { 5149 status = DML_FAIL_NOT_ENOUGH_DSC; 5150 } else if (locals->DSCCLKRequiredMoreThanSupported[i] != false) { 5151 status = DML_FAIL_DSC_CLK_REQUIRED; 5152 } else if (locals->UrgentLatencySupport[i][j] != true) { 5153 status = DML_FAIL_URGENT_LATENCY; 5154 } else if (locals->ROBSupport[i][0] != true) { 5155 status = DML_FAIL_REORDERING_BUFFER; 5156 } else if (locals->DISPCLK_DPPCLK_Support[i][j] != true) { 5157 status = DML_FAIL_DISPCLK_DPPCLK; 5158 } else if (locals->TotalAvailablePipesSupport[i][j] != true) { 5159 status = DML_FAIL_TOTAL_AVAILABLE_PIPES; 5160 } else if (mode_lib->vba.NumberOfOTGSupport != true) { 5161 status = DML_FAIL_NUM_OTG; 5162 } else if (mode_lib->vba.WritebackModeSupport != true) { 5163 status = DML_FAIL_WRITEBACK_MODE; 5164 } else if (mode_lib->vba.WritebackLatencySupport != true) { 5165 status = DML_FAIL_WRITEBACK_LATENCY; 5166 } else if (mode_lib->vba.WritebackScaleRatioAndTapsSupport != true) { 5167 status = DML_FAIL_WRITEBACK_SCALE_RATIO_TAP; 5168 } else if (mode_lib->vba.CursorSupport != true) { 5169 status = DML_FAIL_CURSOR_SUPPORT; 5170 } else if (mode_lib->vba.PitchSupport != true) { 5171 status = DML_FAIL_PITCH_SUPPORT; 5172 } else if (locals->PrefetchSupported[i][j] != true) { 5173 status = DML_FAIL_PREFETCH_SUPPORT; 5174 } else if (locals->TotalVerticalActiveBandwidthSupport[i][0] != true) { 5175 status = DML_FAIL_TOTAL_V_ACTIVE_BW; 5176 } else if (locals->VRatioInPrefetchSupported[i][j] != true) { 5177 status = DML_FAIL_V_RATIO_PREFETCH; 5178 } else if (locals->PTEBufferSizeNotExceeded[i][j] != true) { 5179 status = DML_FAIL_PTE_BUFFER_SIZE; 5180 } else if (mode_lib->vba.NonsupportedDSCInputBPC != false) { 5181 status = DML_FAIL_DSC_INPUT_BPC; 5182 } 5183 5184 if (status == DML_VALIDATION_OK) { 5185 locals->ModeSupport[i][j] = true; 5186 } else { 5187 locals->ModeSupport[i][j] = false; 5188 } 5189 locals->ValidationStatus[i] = status; 5190 } 5191 } 5192 { 5193 unsigned int MaximumMPCCombine = 0; 5194 mode_lib->vba.VoltageLevel = mode_lib->vba.soc.num_states + 1; 5195 for (i = mode_lib->vba.VoltageOverrideLevel; i <= mode_lib->vba.soc.num_states; i++) { 5196 if (locals->ModeSupport[i][0] == true || locals->ModeSupport[i][1] == true) { 5197 mode_lib->vba.VoltageLevel = i; 5198 if (locals->ModeSupport[i][1] == true && (locals->ModeSupport[i][0] == false 5199 || mode_lib->vba.WhenToDoMPCCombine == dm_mpc_always_when_possible)) { 5200 MaximumMPCCombine = 1; 5201 } else { 5202 MaximumMPCCombine = 0; 5203 } 5204 break; 5205 } 5206 } 5207 mode_lib->vba.ImmediateFlipSupport = 5208 locals->ImmediateFlipSupportedForState[mode_lib->vba.VoltageLevel][MaximumMPCCombine]; 5209 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 5210 mode_lib->vba.DPPPerPlane[k] = locals->NoOfDPP[mode_lib->vba.VoltageLevel][MaximumMPCCombine][k]; 5211 locals->DPPCLK[k] = locals->RequiredDPPCLK[mode_lib->vba.VoltageLevel][MaximumMPCCombine][k]; 5212 } 5213 mode_lib->vba.DISPCLK = locals->RequiredDISPCLK[mode_lib->vba.VoltageLevel][MaximumMPCCombine]; 5214 mode_lib->vba.maxMpcComb = MaximumMPCCombine; 5215 } 5216 mode_lib->vba.DCFCLK = mode_lib->vba.DCFCLKPerState[mode_lib->vba.VoltageLevel]; 5217 mode_lib->vba.DRAMSpeed = mode_lib->vba.DRAMSpeedPerState[mode_lib->vba.VoltageLevel]; 5218 mode_lib->vba.FabricClock = mode_lib->vba.FabricClockPerState[mode_lib->vba.VoltageLevel]; 5219 mode_lib->vba.SOCCLK = mode_lib->vba.SOCCLKPerState[mode_lib->vba.VoltageLevel]; 5220 mode_lib->vba.ReturnBW = locals->ReturnBWPerState[mode_lib->vba.VoltageLevel][0]; 5221 mode_lib->vba.FabricAndDRAMBandwidth = locals->FabricAndDRAMBandwidthPerState[mode_lib->vba.VoltageLevel]; 5222 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 5223 if (mode_lib->vba.BlendingAndTiming[k] == k) { 5224 mode_lib->vba.ODMCombineEnabled[k] = 5225 locals->ODMCombineEnablePerState[mode_lib->vba.VoltageLevel][k]; 5226 } else { 5227 mode_lib->vba.ODMCombineEnabled[k] = 0; 5228 } 5229 mode_lib->vba.DSCEnabled[k] = 5230 locals->RequiresDSC[mode_lib->vba.VoltageLevel][k]; 5231 mode_lib->vba.OutputBpp[k] = 5232 locals->OutputBppPerState[mode_lib->vba.VoltageLevel][k]; 5233 } 5234 } 5235