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 mode_lib->vba.MinActiveDRAMClockChangeMargin = 999999; 2603 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2604 if (mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k] 2605 < mode_lib->vba.MinActiveDRAMClockChangeMargin) { 2606 mode_lib->vba.MinActiveDRAMClockChangeMargin = 2607 mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k]; 2608 } 2609 } 2610 2611 mode_lib->vba.MinActiveDRAMClockChangeLatencySupported = 2612 mode_lib->vba.MinActiveDRAMClockChangeMargin 2613 + mode_lib->vba.DRAMClockChangeLatency; 2614 2615 if (mode_lib->vba.DRAMClockChangeSupportsVActive && 2616 mode_lib->vba.MinActiveDRAMClockChangeMargin > 60) { 2617 mode_lib->vba.DRAMClockChangeWatermark += 25; 2618 mode_lib->vba.DRAMClockChangeSupport[0][0] = dm_dram_clock_change_vactive; 2619 } else if (mode_lib->vba.DummyPStateCheck && 2620 mode_lib->vba.MinActiveDRAMClockChangeMargin > 0) { 2621 mode_lib->vba.DRAMClockChangeSupport[0][0] = dm_dram_clock_change_vactive; 2622 } else { 2623 if (mode_lib->vba.SynchronizedVBlank || mode_lib->vba.NumberOfActivePlanes == 1) { 2624 mode_lib->vba.DRAMClockChangeSupport[0][0] = dm_dram_clock_change_vblank; 2625 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2626 if (!mode_lib->vba.AllowDRAMClockChangeDuringVBlank[k]) { 2627 mode_lib->vba.DRAMClockChangeSupport[0][0] = 2628 dm_dram_clock_change_unsupported; 2629 } 2630 } 2631 } else { 2632 mode_lib->vba.DRAMClockChangeSupport[0][0] = dm_dram_clock_change_unsupported; 2633 } 2634 } 2635 for (k = 0; k <= mode_lib->vba.soc.num_states; k++) 2636 for (j = 0; j < 2; j++) 2637 mode_lib->vba.DRAMClockChangeSupport[k][j] = mode_lib->vba.DRAMClockChangeSupport[0][0]; 2638 2639 //XFC Parameters: 2640 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2641 if (mode_lib->vba.XFCEnabled[k] == true) { 2642 double TWait; 2643 2644 mode_lib->vba.XFCSlaveVUpdateOffset[k] = mode_lib->vba.XFCTSlvVupdateOffset; 2645 mode_lib->vba.XFCSlaveVupdateWidth[k] = mode_lib->vba.XFCTSlvVupdateWidth; 2646 mode_lib->vba.XFCSlaveVReadyOffset[k] = mode_lib->vba.XFCTSlvVreadyOffset; 2647 TWait = CalculateTWait( 2648 mode_lib->vba.PrefetchMode[mode_lib->vba.VoltageLevel][mode_lib->vba.maxMpcComb], 2649 mode_lib->vba.DRAMClockChangeLatency, 2650 mode_lib->vba.UrgentLatencyPixelDataOnly, 2651 mode_lib->vba.SREnterPlusExitTime); 2652 mode_lib->vba.XFCRemoteSurfaceFlipDelay = CalculateRemoteSurfaceFlipDelay( 2653 mode_lib, 2654 mode_lib->vba.VRatio[k], 2655 mode_lib->vba.SwathWidthY[k], 2656 dml_ceil(mode_lib->vba.BytePerPixelDETY[k], 1), 2657 mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k], 2658 mode_lib->vba.XFCTSlvVupdateOffset, 2659 mode_lib->vba.XFCTSlvVupdateWidth, 2660 mode_lib->vba.XFCTSlvVreadyOffset, 2661 mode_lib->vba.XFCXBUFLatencyTolerance, 2662 mode_lib->vba.XFCFillBWOverhead, 2663 mode_lib->vba.XFCSlvChunkSize, 2664 mode_lib->vba.XFCBusTransportTime, 2665 mode_lib->vba.TCalc, 2666 TWait, 2667 &mode_lib->vba.SrcActiveDrainRate, 2668 &mode_lib->vba.TInitXFill, 2669 &mode_lib->vba.TslvChk); 2670 mode_lib->vba.XFCRemoteSurfaceFlipLatency[k] = 2671 dml_floor( 2672 mode_lib->vba.XFCRemoteSurfaceFlipDelay 2673 / (mode_lib->vba.HTotal[k] 2674 / mode_lib->vba.PixelClock[k]), 2675 1); 2676 mode_lib->vba.XFCTransferDelay[k] = 2677 dml_ceil( 2678 mode_lib->vba.XFCBusTransportTime 2679 / (mode_lib->vba.HTotal[k] 2680 / mode_lib->vba.PixelClock[k]), 2681 1); 2682 mode_lib->vba.XFCPrechargeDelay[k] = 2683 dml_ceil( 2684 (mode_lib->vba.XFCBusTransportTime 2685 + mode_lib->vba.TInitXFill 2686 + mode_lib->vba.TslvChk) 2687 / (mode_lib->vba.HTotal[k] 2688 / mode_lib->vba.PixelClock[k]), 2689 1); 2690 mode_lib->vba.InitFillLevel = mode_lib->vba.XFCXBUFLatencyTolerance 2691 * mode_lib->vba.SrcActiveDrainRate; 2692 mode_lib->vba.FinalFillMargin = 2693 (mode_lib->vba.DestinationLinesToRequestVMInVBlank[k] 2694 + mode_lib->vba.DestinationLinesToRequestRowInVBlank[k]) 2695 * mode_lib->vba.HTotal[k] 2696 / mode_lib->vba.PixelClock[k] 2697 * mode_lib->vba.SrcActiveDrainRate 2698 + mode_lib->vba.XFCFillConstant; 2699 mode_lib->vba.FinalFillLevel = mode_lib->vba.XFCRemoteSurfaceFlipDelay 2700 * mode_lib->vba.SrcActiveDrainRate 2701 + mode_lib->vba.FinalFillMargin; 2702 mode_lib->vba.RemainingFillLevel = dml_max( 2703 0.0, 2704 mode_lib->vba.FinalFillLevel - mode_lib->vba.InitFillLevel); 2705 mode_lib->vba.TFinalxFill = mode_lib->vba.RemainingFillLevel 2706 / (mode_lib->vba.SrcActiveDrainRate 2707 * mode_lib->vba.XFCFillBWOverhead / 100); 2708 mode_lib->vba.XFCPrefetchMargin[k] = 2709 mode_lib->vba.XFCRemoteSurfaceFlipDelay 2710 + mode_lib->vba.TFinalxFill 2711 + (mode_lib->vba.DestinationLinesToRequestVMInVBlank[k] 2712 + mode_lib->vba.DestinationLinesToRequestRowInVBlank[k]) 2713 * mode_lib->vba.HTotal[k] 2714 / mode_lib->vba.PixelClock[k]; 2715 } else { 2716 mode_lib->vba.XFCSlaveVUpdateOffset[k] = 0; 2717 mode_lib->vba.XFCSlaveVupdateWidth[k] = 0; 2718 mode_lib->vba.XFCSlaveVReadyOffset[k] = 0; 2719 mode_lib->vba.XFCRemoteSurfaceFlipLatency[k] = 0; 2720 mode_lib->vba.XFCPrechargeDelay[k] = 0; 2721 mode_lib->vba.XFCTransferDelay[k] = 0; 2722 mode_lib->vba.XFCPrefetchMargin[k] = 0; 2723 } 2724 } 2725 { 2726 unsigned int VStartupMargin = 0; 2727 bool FirstMainPlane = true; 2728 2729 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2730 if (mode_lib->vba.BlendingAndTiming[k] == k) { 2731 unsigned int Margin = (mode_lib->vba.MaxVStartupLines[k] - mode_lib->vba.VStartup[k]) 2732 * mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]; 2733 2734 if (FirstMainPlane) { 2735 VStartupMargin = Margin; 2736 FirstMainPlane = false; 2737 } else 2738 VStartupMargin = dml_min(VStartupMargin, Margin); 2739 } 2740 2741 if (mode_lib->vba.UseMaximumVStartup) { 2742 if (mode_lib->vba.VTotal_Max[k] == mode_lib->vba.VTotal[k]) { 2743 //only use max vstart if it is not drr or lateflip. 2744 mode_lib->vba.VStartup[k] = mode_lib->vba.MaxVStartupLines[mode_lib->vba.BlendingAndTiming[k]]; 2745 } 2746 } 2747 } 2748 } 2749 } 2750 2751 static void dml20v2_DisplayPipeConfiguration(struct display_mode_lib *mode_lib) 2752 { 2753 double BytePerPixDETY; 2754 double BytePerPixDETC; 2755 double Read256BytesBlockHeightY; 2756 double Read256BytesBlockHeightC; 2757 double Read256BytesBlockWidthY; 2758 double Read256BytesBlockWidthC; 2759 double MaximumSwathHeightY; 2760 double MaximumSwathHeightC; 2761 double MinimumSwathHeightY; 2762 double MinimumSwathHeightC; 2763 double SwathWidth; 2764 double SwathWidthGranularityY; 2765 double SwathWidthGranularityC; 2766 double RoundedUpMaxSwathSizeBytesY; 2767 double RoundedUpMaxSwathSizeBytesC; 2768 unsigned int j, k; 2769 2770 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 2771 bool MainPlaneDoesODMCombine = false; 2772 2773 if (mode_lib->vba.SourcePixelFormat[k] == dm_444_64) { 2774 BytePerPixDETY = 8; 2775 BytePerPixDETC = 0; 2776 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_32) { 2777 BytePerPixDETY = 4; 2778 BytePerPixDETC = 0; 2779 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_16) { 2780 BytePerPixDETY = 2; 2781 BytePerPixDETC = 0; 2782 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_8) { 2783 BytePerPixDETY = 1; 2784 BytePerPixDETC = 0; 2785 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_8) { 2786 BytePerPixDETY = 1; 2787 BytePerPixDETC = 2; 2788 } else { 2789 BytePerPixDETY = 4.0 / 3.0; 2790 BytePerPixDETC = 8.0 / 3.0; 2791 } 2792 2793 if ((mode_lib->vba.SourcePixelFormat[k] == dm_444_64 2794 || mode_lib->vba.SourcePixelFormat[k] == dm_444_32 2795 || mode_lib->vba.SourcePixelFormat[k] == dm_444_16 2796 || mode_lib->vba.SourcePixelFormat[k] == dm_444_8)) { 2797 if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear) { 2798 Read256BytesBlockHeightY = 1; 2799 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_64) { 2800 Read256BytesBlockHeightY = 4; 2801 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_32 2802 || mode_lib->vba.SourcePixelFormat[k] == dm_444_16) { 2803 Read256BytesBlockHeightY = 8; 2804 } else { 2805 Read256BytesBlockHeightY = 16; 2806 } 2807 Read256BytesBlockWidthY = 256 / dml_ceil(BytePerPixDETY, 1) 2808 / Read256BytesBlockHeightY; 2809 Read256BytesBlockHeightC = 0; 2810 Read256BytesBlockWidthC = 0; 2811 } else { 2812 if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear) { 2813 Read256BytesBlockHeightY = 1; 2814 Read256BytesBlockHeightC = 1; 2815 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_8) { 2816 Read256BytesBlockHeightY = 16; 2817 Read256BytesBlockHeightC = 8; 2818 } else { 2819 Read256BytesBlockHeightY = 8; 2820 Read256BytesBlockHeightC = 8; 2821 } 2822 Read256BytesBlockWidthY = 256 / dml_ceil(BytePerPixDETY, 1) 2823 / Read256BytesBlockHeightY; 2824 Read256BytesBlockWidthC = 256 / dml_ceil(BytePerPixDETC, 2) 2825 / Read256BytesBlockHeightC; 2826 } 2827 2828 if (mode_lib->vba.SourceScan[k] == dm_horz) { 2829 MaximumSwathHeightY = Read256BytesBlockHeightY; 2830 MaximumSwathHeightC = Read256BytesBlockHeightC; 2831 } else { 2832 MaximumSwathHeightY = Read256BytesBlockWidthY; 2833 MaximumSwathHeightC = Read256BytesBlockWidthC; 2834 } 2835 2836 if ((mode_lib->vba.SourcePixelFormat[k] == dm_444_64 2837 || mode_lib->vba.SourcePixelFormat[k] == dm_444_32 2838 || mode_lib->vba.SourcePixelFormat[k] == dm_444_16 2839 || mode_lib->vba.SourcePixelFormat[k] == dm_444_8)) { 2840 if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear 2841 || (mode_lib->vba.SourcePixelFormat[k] == dm_444_64 2842 && (mode_lib->vba.SurfaceTiling[k] 2843 == dm_sw_4kb_s 2844 || mode_lib->vba.SurfaceTiling[k] 2845 == dm_sw_4kb_s_x 2846 || mode_lib->vba.SurfaceTiling[k] 2847 == dm_sw_64kb_s 2848 || mode_lib->vba.SurfaceTiling[k] 2849 == dm_sw_64kb_s_t 2850 || mode_lib->vba.SurfaceTiling[k] 2851 == dm_sw_64kb_s_x 2852 || mode_lib->vba.SurfaceTiling[k] 2853 == dm_sw_var_s 2854 || mode_lib->vba.SurfaceTiling[k] 2855 == dm_sw_var_s_x) 2856 && mode_lib->vba.SourceScan[k] == dm_horz)) { 2857 MinimumSwathHeightY = MaximumSwathHeightY; 2858 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_8 2859 && mode_lib->vba.SourceScan[k] != dm_horz) { 2860 MinimumSwathHeightY = MaximumSwathHeightY; 2861 } else { 2862 MinimumSwathHeightY = MaximumSwathHeightY / 2.0; 2863 } 2864 MinimumSwathHeightC = MaximumSwathHeightC; 2865 } else { 2866 if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear) { 2867 MinimumSwathHeightY = MaximumSwathHeightY; 2868 MinimumSwathHeightC = MaximumSwathHeightC; 2869 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_8 2870 && mode_lib->vba.SourceScan[k] == dm_horz) { 2871 MinimumSwathHeightY = MaximumSwathHeightY / 2.0; 2872 MinimumSwathHeightC = MaximumSwathHeightC; 2873 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_10 2874 && mode_lib->vba.SourceScan[k] == dm_horz) { 2875 MinimumSwathHeightC = MaximumSwathHeightC / 2.0; 2876 MinimumSwathHeightY = MaximumSwathHeightY; 2877 } else { 2878 MinimumSwathHeightY = MaximumSwathHeightY; 2879 MinimumSwathHeightC = MaximumSwathHeightC; 2880 } 2881 } 2882 2883 if (mode_lib->vba.SourceScan[k] == dm_horz) { 2884 SwathWidth = mode_lib->vba.ViewportWidth[k]; 2885 } else { 2886 SwathWidth = mode_lib->vba.ViewportHeight[k]; 2887 } 2888 2889 if (mode_lib->vba.ODMCombineEnabled[k] == dm_odm_combine_mode_2to1) { 2890 MainPlaneDoesODMCombine = true; 2891 } 2892 for (j = 0; j < mode_lib->vba.NumberOfActivePlanes; ++j) { 2893 if (mode_lib->vba.BlendingAndTiming[k] == j 2894 && mode_lib->vba.ODMCombineEnabled[k] == dm_odm_combine_mode_2to1) { 2895 MainPlaneDoesODMCombine = true; 2896 } 2897 } 2898 2899 if (MainPlaneDoesODMCombine == true) { 2900 SwathWidth = dml_min( 2901 SwathWidth, 2902 mode_lib->vba.HActive[k] / 2.0 * mode_lib->vba.HRatio[k]); 2903 } else { 2904 if (mode_lib->vba.DPPPerPlane[k] == 0) 2905 SwathWidth = 0; 2906 else 2907 SwathWidth = SwathWidth / mode_lib->vba.DPPPerPlane[k]; 2908 } 2909 2910 SwathWidthGranularityY = 256 / dml_ceil(BytePerPixDETY, 1) / MaximumSwathHeightY; 2911 RoundedUpMaxSwathSizeBytesY = (dml_ceil( 2912 (double) (SwathWidth - 1), 2913 SwathWidthGranularityY) + SwathWidthGranularityY) * BytePerPixDETY 2914 * MaximumSwathHeightY; 2915 if (mode_lib->vba.SourcePixelFormat[k] == dm_420_10) { 2916 RoundedUpMaxSwathSizeBytesY = dml_ceil(RoundedUpMaxSwathSizeBytesY, 256) 2917 + 256; 2918 } 2919 if (MaximumSwathHeightC > 0) { 2920 SwathWidthGranularityC = 256.0 / dml_ceil(BytePerPixDETC, 2) 2921 / MaximumSwathHeightC; 2922 RoundedUpMaxSwathSizeBytesC = (dml_ceil( 2923 (double) (SwathWidth / 2.0 - 1), 2924 SwathWidthGranularityC) + SwathWidthGranularityC) 2925 * BytePerPixDETC * MaximumSwathHeightC; 2926 if (mode_lib->vba.SourcePixelFormat[k] == dm_420_10) { 2927 RoundedUpMaxSwathSizeBytesC = dml_ceil( 2928 RoundedUpMaxSwathSizeBytesC, 2929 256) + 256; 2930 } 2931 } else 2932 RoundedUpMaxSwathSizeBytesC = 0.0; 2933 2934 if (RoundedUpMaxSwathSizeBytesY + RoundedUpMaxSwathSizeBytesC 2935 <= mode_lib->vba.DETBufferSizeInKByte * 1024.0 / 2.0) { 2936 mode_lib->vba.SwathHeightY[k] = MaximumSwathHeightY; 2937 mode_lib->vba.SwathHeightC[k] = MaximumSwathHeightC; 2938 } else { 2939 mode_lib->vba.SwathHeightY[k] = MinimumSwathHeightY; 2940 mode_lib->vba.SwathHeightC[k] = MinimumSwathHeightC; 2941 } 2942 2943 if (mode_lib->vba.SwathHeightC[k] == 0) { 2944 mode_lib->vba.DETBufferSizeY[k] = mode_lib->vba.DETBufferSizeInKByte * 1024; 2945 mode_lib->vba.DETBufferSizeC[k] = 0; 2946 } else if (mode_lib->vba.SwathHeightY[k] <= mode_lib->vba.SwathHeightC[k]) { 2947 mode_lib->vba.DETBufferSizeY[k] = mode_lib->vba.DETBufferSizeInKByte 2948 * 1024.0 / 2; 2949 mode_lib->vba.DETBufferSizeC[k] = mode_lib->vba.DETBufferSizeInKByte 2950 * 1024.0 / 2; 2951 } else { 2952 mode_lib->vba.DETBufferSizeY[k] = mode_lib->vba.DETBufferSizeInKByte 2953 * 1024.0 * 2 / 3; 2954 mode_lib->vba.DETBufferSizeC[k] = mode_lib->vba.DETBufferSizeInKByte 2955 * 1024.0 / 3; 2956 } 2957 } 2958 } 2959 2960 static double CalculateTWait( 2961 unsigned int PrefetchMode, 2962 double DRAMClockChangeLatency, 2963 double UrgentLatencyPixelDataOnly, 2964 double SREnterPlusExitTime) 2965 { 2966 if (PrefetchMode == 0) { 2967 return dml_max( 2968 DRAMClockChangeLatency + UrgentLatencyPixelDataOnly, 2969 dml_max(SREnterPlusExitTime, UrgentLatencyPixelDataOnly)); 2970 } else if (PrefetchMode == 1) { 2971 return dml_max(SREnterPlusExitTime, UrgentLatencyPixelDataOnly); 2972 } else { 2973 return UrgentLatencyPixelDataOnly; 2974 } 2975 } 2976 2977 static double CalculateRemoteSurfaceFlipDelay( 2978 struct display_mode_lib *mode_lib, 2979 double VRatio, 2980 double SwathWidth, 2981 double Bpp, 2982 double LineTime, 2983 double XFCTSlvVupdateOffset, 2984 double XFCTSlvVupdateWidth, 2985 double XFCTSlvVreadyOffset, 2986 double XFCXBUFLatencyTolerance, 2987 double XFCFillBWOverhead, 2988 double XFCSlvChunkSize, 2989 double XFCBusTransportTime, 2990 double TCalc, 2991 double TWait, 2992 double *SrcActiveDrainRate, 2993 double *TInitXFill, 2994 double *TslvChk) 2995 { 2996 double TSlvSetup, AvgfillRate, result; 2997 2998 *SrcActiveDrainRate = VRatio * SwathWidth * Bpp / LineTime; 2999 TSlvSetup = XFCTSlvVupdateOffset + XFCTSlvVupdateWidth + XFCTSlvVreadyOffset; 3000 *TInitXFill = XFCXBUFLatencyTolerance / (1 + XFCFillBWOverhead / 100); 3001 AvgfillRate = *SrcActiveDrainRate * (1 + XFCFillBWOverhead / 100); 3002 *TslvChk = XFCSlvChunkSize / AvgfillRate; 3003 dml_print( 3004 "DML::CalculateRemoteSurfaceFlipDelay: SrcActiveDrainRate: %f\n", 3005 *SrcActiveDrainRate); 3006 dml_print("DML::CalculateRemoteSurfaceFlipDelay: TSlvSetup: %f\n", TSlvSetup); 3007 dml_print("DML::CalculateRemoteSurfaceFlipDelay: TInitXFill: %f\n", *TInitXFill); 3008 dml_print("DML::CalculateRemoteSurfaceFlipDelay: AvgfillRate: %f\n", AvgfillRate); 3009 dml_print("DML::CalculateRemoteSurfaceFlipDelay: TslvChk: %f\n", *TslvChk); 3010 result = 2 * XFCBusTransportTime + TSlvSetup + TCalc + TWait + *TslvChk + *TInitXFill; // TODO: This doesn't seem to match programming guide 3011 dml_print("DML::CalculateRemoteSurfaceFlipDelay: RemoteSurfaceFlipDelay: %f\n", result); 3012 return result; 3013 } 3014 3015 static double CalculateWriteBackDelay( 3016 enum source_format_class WritebackPixelFormat, 3017 double WritebackHRatio, 3018 double WritebackVRatio, 3019 unsigned int WritebackLumaHTaps, 3020 unsigned int WritebackLumaVTaps, 3021 unsigned int WritebackChromaHTaps, 3022 unsigned int WritebackChromaVTaps, 3023 unsigned int WritebackDestinationWidth) 3024 { 3025 double CalculateWriteBackDelay = 3026 dml_max( 3027 dml_ceil(WritebackLumaHTaps / 4.0, 1) / WritebackHRatio, 3028 WritebackLumaVTaps * dml_ceil(1.0 / WritebackVRatio, 1) 3029 * dml_ceil( 3030 WritebackDestinationWidth 3031 / 4.0, 3032 1) 3033 + dml_ceil(1.0 / WritebackVRatio, 1) 3034 * (dml_ceil( 3035 WritebackLumaVTaps 3036 / 4.0, 3037 1) + 4)); 3038 3039 if (WritebackPixelFormat != dm_444_32) { 3040 CalculateWriteBackDelay = 3041 dml_max( 3042 CalculateWriteBackDelay, 3043 dml_max( 3044 dml_ceil( 3045 WritebackChromaHTaps 3046 / 2.0, 3047 1) 3048 / (2 3049 * WritebackHRatio), 3050 WritebackChromaVTaps 3051 * dml_ceil( 3052 1 3053 / (2 3054 * WritebackVRatio), 3055 1) 3056 * dml_ceil( 3057 WritebackDestinationWidth 3058 / 2.0 3059 / 2.0, 3060 1) 3061 + dml_ceil( 3062 1 3063 / (2 3064 * WritebackVRatio), 3065 1) 3066 * (dml_ceil( 3067 WritebackChromaVTaps 3068 / 4.0, 3069 1) 3070 + 4))); 3071 } 3072 return CalculateWriteBackDelay; 3073 } 3074 3075 static void CalculateActiveRowBandwidth( 3076 bool GPUVMEnable, 3077 enum source_format_class SourcePixelFormat, 3078 double VRatio, 3079 bool DCCEnable, 3080 double LineTime, 3081 unsigned int MetaRowByteLuma, 3082 unsigned int MetaRowByteChroma, 3083 unsigned int meta_row_height_luma, 3084 unsigned int meta_row_height_chroma, 3085 unsigned int PixelPTEBytesPerRowLuma, 3086 unsigned int PixelPTEBytesPerRowChroma, 3087 unsigned int dpte_row_height_luma, 3088 unsigned int dpte_row_height_chroma, 3089 double *meta_row_bw, 3090 double *dpte_row_bw, 3091 double *qual_row_bw) 3092 { 3093 if (DCCEnable != true) { 3094 *meta_row_bw = 0; 3095 } else if (SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10) { 3096 *meta_row_bw = VRatio * MetaRowByteLuma / (meta_row_height_luma * LineTime) 3097 + VRatio / 2 * MetaRowByteChroma 3098 / (meta_row_height_chroma * LineTime); 3099 } else { 3100 *meta_row_bw = VRatio * MetaRowByteLuma / (meta_row_height_luma * LineTime); 3101 } 3102 3103 if (GPUVMEnable != true) { 3104 *dpte_row_bw = 0; 3105 } else if (SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10) { 3106 *dpte_row_bw = VRatio * PixelPTEBytesPerRowLuma / (dpte_row_height_luma * LineTime) 3107 + VRatio / 2 * PixelPTEBytesPerRowChroma 3108 / (dpte_row_height_chroma * LineTime); 3109 } else { 3110 *dpte_row_bw = VRatio * PixelPTEBytesPerRowLuma / (dpte_row_height_luma * LineTime); 3111 } 3112 3113 if ((SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10)) { 3114 *qual_row_bw = *meta_row_bw + *dpte_row_bw; 3115 } else { 3116 *qual_row_bw = 0; 3117 } 3118 } 3119 3120 static void CalculateFlipSchedule( 3121 struct display_mode_lib *mode_lib, 3122 double UrgentExtraLatency, 3123 double UrgentLatencyPixelDataOnly, 3124 unsigned int GPUVMMaxPageTableLevels, 3125 bool GPUVMEnable, 3126 double BandwidthAvailableForImmediateFlip, 3127 unsigned int TotImmediateFlipBytes, 3128 enum source_format_class SourcePixelFormat, 3129 unsigned int ImmediateFlipBytes, 3130 double LineTime, 3131 double VRatio, 3132 double Tno_bw, 3133 double PDEAndMetaPTEBytesFrame, 3134 unsigned int MetaRowByte, 3135 unsigned int PixelPTEBytesPerRow, 3136 bool DCCEnable, 3137 unsigned int dpte_row_height, 3138 unsigned int meta_row_height, 3139 double qual_row_bw, 3140 double *DestinationLinesToRequestVMInImmediateFlip, 3141 double *DestinationLinesToRequestRowInImmediateFlip, 3142 double *final_flip_bw, 3143 bool *ImmediateFlipSupportedForPipe) 3144 { 3145 double min_row_time = 0.0; 3146 3147 if (SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10) { 3148 *DestinationLinesToRequestVMInImmediateFlip = 0.0; 3149 *DestinationLinesToRequestRowInImmediateFlip = 0.0; 3150 *final_flip_bw = qual_row_bw; 3151 *ImmediateFlipSupportedForPipe = true; 3152 } else { 3153 double TimeForFetchingMetaPTEImmediateFlip; 3154 double TimeForFetchingRowInVBlankImmediateFlip; 3155 3156 if (GPUVMEnable == true) { 3157 mode_lib->vba.ImmediateFlipBW[0] = BandwidthAvailableForImmediateFlip 3158 * ImmediateFlipBytes / TotImmediateFlipBytes; 3159 TimeForFetchingMetaPTEImmediateFlip = 3160 dml_max( 3161 Tno_bw 3162 + PDEAndMetaPTEBytesFrame 3163 / mode_lib->vba.ImmediateFlipBW[0], 3164 dml_max( 3165 UrgentExtraLatency 3166 + UrgentLatencyPixelDataOnly 3167 * (GPUVMMaxPageTableLevels 3168 - 1), 3169 LineTime / 4.0)); 3170 } else { 3171 TimeForFetchingMetaPTEImmediateFlip = 0; 3172 } 3173 3174 *DestinationLinesToRequestVMInImmediateFlip = dml_floor( 3175 4.0 * (TimeForFetchingMetaPTEImmediateFlip / LineTime + 0.125), 3176 1) / 4.0; 3177 3178 if ((GPUVMEnable == true || DCCEnable == true)) { 3179 mode_lib->vba.ImmediateFlipBW[0] = BandwidthAvailableForImmediateFlip 3180 * ImmediateFlipBytes / TotImmediateFlipBytes; 3181 TimeForFetchingRowInVBlankImmediateFlip = dml_max( 3182 (MetaRowByte + PixelPTEBytesPerRow) 3183 / mode_lib->vba.ImmediateFlipBW[0], 3184 dml_max(UrgentLatencyPixelDataOnly, LineTime / 4.0)); 3185 } else { 3186 TimeForFetchingRowInVBlankImmediateFlip = 0; 3187 } 3188 3189 *DestinationLinesToRequestRowInImmediateFlip = dml_floor( 3190 4.0 * (TimeForFetchingRowInVBlankImmediateFlip / LineTime + 0.125), 3191 1) / 4.0; 3192 3193 if (GPUVMEnable == true) { 3194 *final_flip_bw = 3195 dml_max( 3196 PDEAndMetaPTEBytesFrame 3197 / (*DestinationLinesToRequestVMInImmediateFlip 3198 * LineTime), 3199 (MetaRowByte + PixelPTEBytesPerRow) 3200 / (TimeForFetchingRowInVBlankImmediateFlip 3201 * LineTime)); 3202 } else if (MetaRowByte + PixelPTEBytesPerRow > 0) { 3203 *final_flip_bw = (MetaRowByte + PixelPTEBytesPerRow) 3204 / (TimeForFetchingRowInVBlankImmediateFlip * LineTime); 3205 } else { 3206 *final_flip_bw = 0; 3207 } 3208 3209 if (GPUVMEnable && !DCCEnable) 3210 min_row_time = dpte_row_height * LineTime / VRatio; 3211 else if (!GPUVMEnable && DCCEnable) 3212 min_row_time = meta_row_height * LineTime / VRatio; 3213 else 3214 min_row_time = dml_min(dpte_row_height, meta_row_height) * LineTime 3215 / VRatio; 3216 3217 if (*DestinationLinesToRequestVMInImmediateFlip >= 8 3218 || *DestinationLinesToRequestRowInImmediateFlip >= 16 3219 || TimeForFetchingMetaPTEImmediateFlip 3220 + 2 * TimeForFetchingRowInVBlankImmediateFlip 3221 > min_row_time) 3222 *ImmediateFlipSupportedForPipe = false; 3223 else 3224 *ImmediateFlipSupportedForPipe = true; 3225 } 3226 } 3227 3228 static unsigned int TruncToValidBPP( 3229 double DecimalBPP, 3230 bool DSCEnabled, 3231 enum output_encoder_class Output, 3232 enum output_format_class Format, 3233 unsigned int DSCInputBitPerComponent) 3234 { 3235 if (Output == dm_hdmi) { 3236 if (Format == dm_420) { 3237 if (DecimalBPP >= 18) 3238 return 18; 3239 else if (DecimalBPP >= 15) 3240 return 15; 3241 else if (DecimalBPP >= 12) 3242 return 12; 3243 else 3244 return BPP_INVALID; 3245 } else if (Format == dm_444) { 3246 if (DecimalBPP >= 36) 3247 return 36; 3248 else if (DecimalBPP >= 30) 3249 return 30; 3250 else if (DecimalBPP >= 24) 3251 return 24; 3252 else if (DecimalBPP >= 18) 3253 return 18; 3254 else 3255 return BPP_INVALID; 3256 } else { 3257 if (DecimalBPP / 1.5 >= 24) 3258 return 24; 3259 else if (DecimalBPP / 1.5 >= 20) 3260 return 20; 3261 else if (DecimalBPP / 1.5 >= 16) 3262 return 16; 3263 else 3264 return BPP_INVALID; 3265 } 3266 } else { 3267 if (DSCEnabled) { 3268 if (Format == dm_420) { 3269 if (DecimalBPP < 6) 3270 return BPP_INVALID; 3271 else if (DecimalBPP >= 1.5 * DSCInputBitPerComponent - 1 / 16) 3272 return 1.5 * DSCInputBitPerComponent - 1 / 16; 3273 else 3274 return dml_floor(16 * DecimalBPP, 1) / 16; 3275 } else if (Format == dm_n422) { 3276 if (DecimalBPP < 7) 3277 return BPP_INVALID; 3278 else if (DecimalBPP >= 2 * DSCInputBitPerComponent - 1 / 16) 3279 return 2 * DSCInputBitPerComponent - 1 / 16; 3280 else 3281 return dml_floor(16 * DecimalBPP, 1) / 16; 3282 } else { 3283 if (DecimalBPP < 8) 3284 return BPP_INVALID; 3285 else if (DecimalBPP >= 3 * DSCInputBitPerComponent - 1 / 16) 3286 return 3 * DSCInputBitPerComponent - 1 / 16; 3287 else 3288 return dml_floor(16 * DecimalBPP, 1) / 16; 3289 } 3290 } else if (Format == dm_420) { 3291 if (DecimalBPP >= 18) 3292 return 18; 3293 else if (DecimalBPP >= 15) 3294 return 15; 3295 else if (DecimalBPP >= 12) 3296 return 12; 3297 else 3298 return BPP_INVALID; 3299 } else if (Format == dm_s422 || Format == dm_n422) { 3300 if (DecimalBPP >= 24) 3301 return 24; 3302 else if (DecimalBPP >= 20) 3303 return 20; 3304 else if (DecimalBPP >= 16) 3305 return 16; 3306 else 3307 return BPP_INVALID; 3308 } else { 3309 if (DecimalBPP >= 36) 3310 return 36; 3311 else if (DecimalBPP >= 30) 3312 return 30; 3313 else if (DecimalBPP >= 24) 3314 return 24; 3315 else if (DecimalBPP >= 18) 3316 return 18; 3317 else 3318 return BPP_INVALID; 3319 } 3320 } 3321 } 3322 3323 void dml20v2_ModeSupportAndSystemConfigurationFull(struct display_mode_lib *mode_lib) 3324 { 3325 struct vba_vars_st *locals = &mode_lib->vba; 3326 3327 int i; 3328 unsigned int j, k, m; 3329 3330 /*MODE SUPPORT, VOLTAGE STATE AND SOC CONFIGURATION*/ 3331 3332 /*Scale Ratio, taps Support Check*/ 3333 3334 mode_lib->vba.ScaleRatioAndTapsSupport = true; 3335 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3336 if (mode_lib->vba.ScalerEnabled[k] == false 3337 && ((mode_lib->vba.SourcePixelFormat[k] != dm_444_64 3338 && mode_lib->vba.SourcePixelFormat[k] != dm_444_32 3339 && mode_lib->vba.SourcePixelFormat[k] != dm_444_16 3340 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_16 3341 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_8) 3342 || mode_lib->vba.HRatio[k] != 1.0 3343 || mode_lib->vba.htaps[k] != 1.0 3344 || mode_lib->vba.VRatio[k] != 1.0 3345 || mode_lib->vba.vtaps[k] != 1.0)) { 3346 mode_lib->vba.ScaleRatioAndTapsSupport = false; 3347 } else if (mode_lib->vba.vtaps[k] < 1.0 || mode_lib->vba.vtaps[k] > 8.0 3348 || mode_lib->vba.htaps[k] < 1.0 || mode_lib->vba.htaps[k] > 8.0 3349 || (mode_lib->vba.htaps[k] > 1.0 3350 && (mode_lib->vba.htaps[k] % 2) == 1) 3351 || mode_lib->vba.HRatio[k] > mode_lib->vba.MaxHSCLRatio 3352 || mode_lib->vba.VRatio[k] > mode_lib->vba.MaxVSCLRatio 3353 || mode_lib->vba.HRatio[k] > mode_lib->vba.htaps[k] 3354 || mode_lib->vba.VRatio[k] > mode_lib->vba.vtaps[k] 3355 || (mode_lib->vba.SourcePixelFormat[k] != dm_444_64 3356 && mode_lib->vba.SourcePixelFormat[k] != dm_444_32 3357 && mode_lib->vba.SourcePixelFormat[k] != dm_444_16 3358 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_16 3359 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_8 3360 && (mode_lib->vba.HRatio[k] / 2.0 3361 > mode_lib->vba.HTAPsChroma[k] 3362 || mode_lib->vba.VRatio[k] / 2.0 3363 > mode_lib->vba.VTAPsChroma[k]))) { 3364 mode_lib->vba.ScaleRatioAndTapsSupport = false; 3365 } 3366 } 3367 /*Source Format, Pixel Format and Scan Support Check*/ 3368 3369 mode_lib->vba.SourceFormatPixelAndScanSupport = true; 3370 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3371 if ((mode_lib->vba.SurfaceTiling[k] == dm_sw_linear 3372 && mode_lib->vba.SourceScan[k] != dm_horz) 3373 || ((mode_lib->vba.SurfaceTiling[k] == dm_sw_4kb_d 3374 || mode_lib->vba.SurfaceTiling[k] == dm_sw_4kb_d_x 3375 || mode_lib->vba.SurfaceTiling[k] == dm_sw_64kb_d 3376 || mode_lib->vba.SurfaceTiling[k] == dm_sw_64kb_d_t 3377 || mode_lib->vba.SurfaceTiling[k] == dm_sw_64kb_d_x 3378 || mode_lib->vba.SurfaceTiling[k] == dm_sw_var_d 3379 || mode_lib->vba.SurfaceTiling[k] == dm_sw_var_d_x) 3380 && mode_lib->vba.SourcePixelFormat[k] != dm_444_64) 3381 || (mode_lib->vba.SurfaceTiling[k] == dm_sw_64kb_r_x 3382 && (mode_lib->vba.SourcePixelFormat[k] == dm_mono_8 3383 || mode_lib->vba.SourcePixelFormat[k] 3384 == dm_420_8 3385 || mode_lib->vba.SourcePixelFormat[k] 3386 == dm_420_10)) 3387 || (((mode_lib->vba.SurfaceTiling[k] == dm_sw_gfx7_2d_thin_gl 3388 || mode_lib->vba.SurfaceTiling[k] 3389 == dm_sw_gfx7_2d_thin_l_vp) 3390 && !((mode_lib->vba.SourcePixelFormat[k] 3391 == dm_444_64 3392 || mode_lib->vba.SourcePixelFormat[k] 3393 == dm_444_32) 3394 && mode_lib->vba.SourceScan[k] 3395 == dm_horz 3396 && mode_lib->vba.SupportGFX7CompatibleTilingIn32bppAnd64bpp 3397 == true 3398 && mode_lib->vba.DCCEnable[k] 3399 == false)) 3400 || (mode_lib->vba.DCCEnable[k] == true 3401 && (mode_lib->vba.SurfaceTiling[k] 3402 == dm_sw_linear 3403 || mode_lib->vba.SourcePixelFormat[k] 3404 == dm_420_8 3405 || mode_lib->vba.SourcePixelFormat[k] 3406 == dm_420_10)))) { 3407 mode_lib->vba.SourceFormatPixelAndScanSupport = false; 3408 } 3409 } 3410 /*Bandwidth Support Check*/ 3411 3412 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3413 if (mode_lib->vba.SourcePixelFormat[k] == dm_444_64) { 3414 locals->BytePerPixelInDETY[k] = 8.0; 3415 locals->BytePerPixelInDETC[k] = 0.0; 3416 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_32) { 3417 locals->BytePerPixelInDETY[k] = 4.0; 3418 locals->BytePerPixelInDETC[k] = 0.0; 3419 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_16 3420 || mode_lib->vba.SourcePixelFormat[k] == dm_mono_16) { 3421 locals->BytePerPixelInDETY[k] = 2.0; 3422 locals->BytePerPixelInDETC[k] = 0.0; 3423 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_mono_8) { 3424 locals->BytePerPixelInDETY[k] = 1.0; 3425 locals->BytePerPixelInDETC[k] = 0.0; 3426 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_8) { 3427 locals->BytePerPixelInDETY[k] = 1.0; 3428 locals->BytePerPixelInDETC[k] = 2.0; 3429 } else { 3430 locals->BytePerPixelInDETY[k] = 4.0 / 3; 3431 locals->BytePerPixelInDETC[k] = 8.0 / 3; 3432 } 3433 if (mode_lib->vba.SourceScan[k] == dm_horz) { 3434 locals->SwathWidthYSingleDPP[k] = mode_lib->vba.ViewportWidth[k]; 3435 } else { 3436 locals->SwathWidthYSingleDPP[k] = mode_lib->vba.ViewportHeight[k]; 3437 } 3438 } 3439 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3440 locals->ReadBandwidthLuma[k] = locals->SwathWidthYSingleDPP[k] * dml_ceil(locals->BytePerPixelInDETY[k], 1.0) 3441 / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]) * mode_lib->vba.VRatio[k]; 3442 locals->ReadBandwidthChroma[k] = locals->SwathWidthYSingleDPP[k] / 2 * dml_ceil(locals->BytePerPixelInDETC[k], 2.0) 3443 / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]) * mode_lib->vba.VRatio[k] / 2.0; 3444 locals->ReadBandwidth[k] = locals->ReadBandwidthLuma[k] + locals->ReadBandwidthChroma[k]; 3445 } 3446 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3447 if (mode_lib->vba.WritebackEnable[k] == true 3448 && mode_lib->vba.WritebackPixelFormat[k] == dm_444_32) { 3449 locals->WriteBandwidth[k] = mode_lib->vba.WritebackDestinationWidth[k] 3450 * mode_lib->vba.WritebackDestinationHeight[k] 3451 / (mode_lib->vba.WritebackSourceHeight[k] 3452 * mode_lib->vba.HTotal[k] 3453 / mode_lib->vba.PixelClock[k]) * 4.0; 3454 } else if (mode_lib->vba.WritebackEnable[k] == true 3455 && mode_lib->vba.WritebackPixelFormat[k] == dm_420_10) { 3456 locals->WriteBandwidth[k] = mode_lib->vba.WritebackDestinationWidth[k] 3457 * mode_lib->vba.WritebackDestinationHeight[k] 3458 / (mode_lib->vba.WritebackSourceHeight[k] 3459 * mode_lib->vba.HTotal[k] 3460 / mode_lib->vba.PixelClock[k]) * 3.0; 3461 } else if (mode_lib->vba.WritebackEnable[k] == true) { 3462 locals->WriteBandwidth[k] = mode_lib->vba.WritebackDestinationWidth[k] 3463 * mode_lib->vba.WritebackDestinationHeight[k] 3464 / (mode_lib->vba.WritebackSourceHeight[k] 3465 * mode_lib->vba.HTotal[k] 3466 / mode_lib->vba.PixelClock[k]) * 1.5; 3467 } else { 3468 locals->WriteBandwidth[k] = 0.0; 3469 } 3470 } 3471 mode_lib->vba.DCCEnabledInAnyPlane = false; 3472 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3473 if (mode_lib->vba.DCCEnable[k] == true) { 3474 mode_lib->vba.DCCEnabledInAnyPlane = true; 3475 } 3476 } 3477 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 3478 locals->FabricAndDRAMBandwidthPerState[i] = dml_min( 3479 mode_lib->vba.DRAMSpeedPerState[i] * mode_lib->vba.NumberOfChannels 3480 * mode_lib->vba.DRAMChannelWidth, 3481 mode_lib->vba.FabricClockPerState[i] 3482 * mode_lib->vba.FabricDatapathToDCNDataReturn) / 1000; 3483 locals->ReturnBWToDCNPerState = dml_min(locals->ReturnBusWidth * locals->DCFCLKPerState[i], 3484 locals->FabricAndDRAMBandwidthPerState[i] * 1000) 3485 * locals->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelDataOnly / 100; 3486 3487 locals->ReturnBWPerState[i][0] = locals->ReturnBWToDCNPerState; 3488 3489 if (locals->DCCEnabledInAnyPlane == true && locals->ReturnBWToDCNPerState > locals->DCFCLKPerState[i] * locals->ReturnBusWidth / 4) { 3490 locals->ReturnBWPerState[i][0] = dml_min(locals->ReturnBWPerState[i][0], 3491 locals->ReturnBWToDCNPerState * 4 * (1 - locals->UrgentLatency / 3492 ((locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024 3493 / (locals->ReturnBWToDCNPerState - locals->DCFCLKPerState[i] 3494 * locals->ReturnBusWidth / 4) + locals->UrgentLatency))); 3495 } 3496 locals->CriticalPoint = 2 * locals->ReturnBusWidth * locals->DCFCLKPerState[i] * 3497 locals->UrgentLatency / (locals->ReturnBWToDCNPerState * locals->UrgentLatency 3498 + (locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024); 3499 3500 if (locals->DCCEnabledInAnyPlane && locals->CriticalPoint > 1 && locals->CriticalPoint < 4) { 3501 locals->ReturnBWPerState[i][0] = dml_min(locals->ReturnBWPerState[i][0], 3502 4 * locals->ReturnBWToDCNPerState * 3503 (locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024 3504 * locals->ReturnBusWidth * locals->DCFCLKPerState[i] * locals->UrgentLatency / 3505 dml_pow((locals->ReturnBWToDCNPerState * locals->UrgentLatency 3506 + (locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024), 2)); 3507 } 3508 3509 locals->ReturnBWToDCNPerState = dml_min(locals->ReturnBusWidth * 3510 locals->DCFCLKPerState[i], locals->FabricAndDRAMBandwidthPerState[i] * 1000); 3511 3512 if (locals->DCCEnabledInAnyPlane == true && locals->ReturnBWToDCNPerState > locals->DCFCLKPerState[i] * locals->ReturnBusWidth / 4) { 3513 locals->ReturnBWPerState[i][0] = dml_min(locals->ReturnBWPerState[i][0], 3514 locals->ReturnBWToDCNPerState * 4 * (1 - locals->UrgentLatency / 3515 ((locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024 3516 / (locals->ReturnBWToDCNPerState - locals->DCFCLKPerState[i] 3517 * locals->ReturnBusWidth / 4) + locals->UrgentLatency))); 3518 } 3519 locals->CriticalPoint = 2 * locals->ReturnBusWidth * locals->DCFCLKPerState[i] * 3520 locals->UrgentLatency / (locals->ReturnBWToDCNPerState * locals->UrgentLatency 3521 + (locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024); 3522 3523 if (locals->DCCEnabledInAnyPlane && locals->CriticalPoint > 1 && locals->CriticalPoint < 4) { 3524 locals->ReturnBWPerState[i][0] = dml_min(locals->ReturnBWPerState[i][0], 3525 4 * locals->ReturnBWToDCNPerState * 3526 (locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024 3527 * locals->ReturnBusWidth * locals->DCFCLKPerState[i] * locals->UrgentLatency / 3528 dml_pow((locals->ReturnBWToDCNPerState * locals->UrgentLatency 3529 + (locals->ROBBufferSizeInKByte - locals->PixelChunkSizeInKByte) * 1024), 2)); 3530 } 3531 } 3532 /*Writeback Latency support check*/ 3533 3534 mode_lib->vba.WritebackLatencySupport = true; 3535 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3536 if (mode_lib->vba.WritebackEnable[k] == true) { 3537 if (mode_lib->vba.WritebackPixelFormat[k] == dm_444_32) { 3538 if (locals->WriteBandwidth[k] 3539 > (mode_lib->vba.WritebackInterfaceLumaBufferSize 3540 + mode_lib->vba.WritebackInterfaceChromaBufferSize) 3541 / mode_lib->vba.WritebackLatency) { 3542 mode_lib->vba.WritebackLatencySupport = false; 3543 } 3544 } else { 3545 if (locals->WriteBandwidth[k] 3546 > 1.5 3547 * dml_min( 3548 mode_lib->vba.WritebackInterfaceLumaBufferSize, 3549 2.0 3550 * mode_lib->vba.WritebackInterfaceChromaBufferSize) 3551 / mode_lib->vba.WritebackLatency) { 3552 mode_lib->vba.WritebackLatencySupport = false; 3553 } 3554 } 3555 } 3556 } 3557 /*Re-ordering Buffer Support Check*/ 3558 3559 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 3560 locals->UrgentRoundTripAndOutOfOrderLatencyPerState[i] = 3561 (mode_lib->vba.RoundTripPingLatencyCycles + 32.0) / mode_lib->vba.DCFCLKPerState[i] 3562 + locals->UrgentOutOfOrderReturnPerChannel * mode_lib->vba.NumberOfChannels / locals->ReturnBWPerState[i][0]; 3563 if ((mode_lib->vba.ROBBufferSizeInKByte - mode_lib->vba.PixelChunkSizeInKByte) * 1024.0 / locals->ReturnBWPerState[i][0] 3564 > locals->UrgentRoundTripAndOutOfOrderLatencyPerState[i]) { 3565 locals->ROBSupport[i][0] = true; 3566 } else { 3567 locals->ROBSupport[i][0] = false; 3568 } 3569 } 3570 /*Writeback Mode Support Check*/ 3571 3572 mode_lib->vba.TotalNumberOfActiveWriteback = 0; 3573 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3574 if (mode_lib->vba.WritebackEnable[k] == true) { 3575 if (mode_lib->vba.ActiveWritebacksPerPlane[k] == 0) 3576 mode_lib->vba.ActiveWritebacksPerPlane[k] = 1; 3577 mode_lib->vba.TotalNumberOfActiveWriteback = 3578 mode_lib->vba.TotalNumberOfActiveWriteback 3579 + mode_lib->vba.ActiveWritebacksPerPlane[k]; 3580 } 3581 } 3582 mode_lib->vba.WritebackModeSupport = true; 3583 if (mode_lib->vba.TotalNumberOfActiveWriteback > mode_lib->vba.MaxNumWriteback) { 3584 mode_lib->vba.WritebackModeSupport = false; 3585 } 3586 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3587 if (mode_lib->vba.WritebackEnable[k] == true 3588 && mode_lib->vba.Writeback10bpc420Supported != true 3589 && mode_lib->vba.WritebackPixelFormat[k] == dm_420_10) { 3590 mode_lib->vba.WritebackModeSupport = false; 3591 } 3592 } 3593 /*Writeback Scale Ratio and Taps Support Check*/ 3594 3595 mode_lib->vba.WritebackScaleRatioAndTapsSupport = true; 3596 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3597 if (mode_lib->vba.WritebackEnable[k] == true) { 3598 if (mode_lib->vba.WritebackLumaAndChromaScalingSupported == false 3599 && (mode_lib->vba.WritebackHRatio[k] != 1.0 3600 || mode_lib->vba.WritebackVRatio[k] != 1.0)) { 3601 mode_lib->vba.WritebackScaleRatioAndTapsSupport = false; 3602 } 3603 if (mode_lib->vba.WritebackHRatio[k] > mode_lib->vba.WritebackMaxHSCLRatio 3604 || mode_lib->vba.WritebackVRatio[k] 3605 > mode_lib->vba.WritebackMaxVSCLRatio 3606 || mode_lib->vba.WritebackHRatio[k] 3607 < mode_lib->vba.WritebackMinHSCLRatio 3608 || mode_lib->vba.WritebackVRatio[k] 3609 < mode_lib->vba.WritebackMinVSCLRatio 3610 || mode_lib->vba.WritebackLumaHTaps[k] 3611 > mode_lib->vba.WritebackMaxHSCLTaps 3612 || mode_lib->vba.WritebackLumaVTaps[k] 3613 > mode_lib->vba.WritebackMaxVSCLTaps 3614 || mode_lib->vba.WritebackHRatio[k] 3615 > mode_lib->vba.WritebackLumaHTaps[k] 3616 || mode_lib->vba.WritebackVRatio[k] 3617 > mode_lib->vba.WritebackLumaVTaps[k] 3618 || (mode_lib->vba.WritebackLumaHTaps[k] > 2.0 3619 && ((mode_lib->vba.WritebackLumaHTaps[k] % 2) 3620 == 1)) 3621 || (mode_lib->vba.WritebackPixelFormat[k] != dm_444_32 3622 && (mode_lib->vba.WritebackChromaHTaps[k] 3623 > mode_lib->vba.WritebackMaxHSCLTaps 3624 || mode_lib->vba.WritebackChromaVTaps[k] 3625 > mode_lib->vba.WritebackMaxVSCLTaps 3626 || 2.0 3627 * mode_lib->vba.WritebackHRatio[k] 3628 > mode_lib->vba.WritebackChromaHTaps[k] 3629 || 2.0 3630 * mode_lib->vba.WritebackVRatio[k] 3631 > mode_lib->vba.WritebackChromaVTaps[k] 3632 || (mode_lib->vba.WritebackChromaHTaps[k] > 2.0 3633 && ((mode_lib->vba.WritebackChromaHTaps[k] % 2) == 1))))) { 3634 mode_lib->vba.WritebackScaleRatioAndTapsSupport = false; 3635 } 3636 if (mode_lib->vba.WritebackVRatio[k] < 1.0) { 3637 mode_lib->vba.WritebackLumaVExtra = 3638 dml_max(1.0 - 2.0 / dml_ceil(1.0 / mode_lib->vba.WritebackVRatio[k], 1.0), 0.0); 3639 } else { 3640 mode_lib->vba.WritebackLumaVExtra = -1; 3641 } 3642 if ((mode_lib->vba.WritebackPixelFormat[k] == dm_444_32 3643 && mode_lib->vba.WritebackLumaVTaps[k] 3644 > (mode_lib->vba.WritebackLineBufferLumaBufferSize 3645 + mode_lib->vba.WritebackLineBufferChromaBufferSize) 3646 / 3.0 3647 / mode_lib->vba.WritebackDestinationWidth[k] 3648 - mode_lib->vba.WritebackLumaVExtra) 3649 || (mode_lib->vba.WritebackPixelFormat[k] == dm_420_8 3650 && mode_lib->vba.WritebackLumaVTaps[k] 3651 > mode_lib->vba.WritebackLineBufferLumaBufferSize 3652 * 8.0 / 10.0 / mode_lib->vba.WritebackDestinationWidth[k] 3653 - mode_lib->vba.WritebackLumaVExtra) 3654 || (mode_lib->vba.WritebackPixelFormat[k] == dm_420_10 3655 && mode_lib->vba.WritebackLumaVTaps[k] 3656 > mode_lib->vba.WritebackLineBufferLumaBufferSize 3657 * 8.0 / 10.0 3658 / mode_lib->vba.WritebackDestinationWidth[k] 3659 - mode_lib->vba.WritebackLumaVExtra)) { 3660 mode_lib->vba.WritebackScaleRatioAndTapsSupport = false; 3661 } 3662 if (2.0 * mode_lib->vba.WritebackVRatio[k] < 1) { 3663 mode_lib->vba.WritebackChromaVExtra = 0.0; 3664 } else { 3665 mode_lib->vba.WritebackChromaVExtra = -1; 3666 } 3667 if ((mode_lib->vba.WritebackPixelFormat[k] == dm_420_8 3668 && mode_lib->vba.WritebackChromaVTaps[k] 3669 > mode_lib->vba.WritebackLineBufferChromaBufferSize 3670 * 8.0 / 10.0 / mode_lib->vba.WritebackDestinationWidth[k] 3671 - mode_lib->vba.WritebackChromaVExtra) 3672 || (mode_lib->vba.WritebackPixelFormat[k] == dm_420_10 3673 && mode_lib->vba.WritebackChromaVTaps[k] 3674 > mode_lib->vba.WritebackLineBufferChromaBufferSize 3675 * 8.0 / 10.0 3676 / mode_lib->vba.WritebackDestinationWidth[k] 3677 - mode_lib->vba.WritebackChromaVExtra)) { 3678 mode_lib->vba.WritebackScaleRatioAndTapsSupport = false; 3679 } 3680 } 3681 } 3682 /*Maximum DISPCLK/DPPCLK Support check*/ 3683 3684 mode_lib->vba.WritebackRequiredDISPCLK = 0.0; 3685 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3686 if (mode_lib->vba.WritebackEnable[k] == true) { 3687 mode_lib->vba.WritebackRequiredDISPCLK = 3688 dml_max( 3689 mode_lib->vba.WritebackRequiredDISPCLK, 3690 CalculateWriteBackDISPCLK( 3691 mode_lib->vba.WritebackPixelFormat[k], 3692 mode_lib->vba.PixelClock[k], 3693 mode_lib->vba.WritebackHRatio[k], 3694 mode_lib->vba.WritebackVRatio[k], 3695 mode_lib->vba.WritebackLumaHTaps[k], 3696 mode_lib->vba.WritebackLumaVTaps[k], 3697 mode_lib->vba.WritebackChromaHTaps[k], 3698 mode_lib->vba.WritebackChromaVTaps[k], 3699 mode_lib->vba.WritebackDestinationWidth[k], 3700 mode_lib->vba.HTotal[k], 3701 mode_lib->vba.WritebackChromaLineBufferWidth)); 3702 } 3703 } 3704 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3705 if (mode_lib->vba.HRatio[k] > 1.0) { 3706 locals->PSCL_FACTOR[k] = dml_min( 3707 mode_lib->vba.MaxDCHUBToPSCLThroughput, 3708 mode_lib->vba.MaxPSCLToLBThroughput 3709 * mode_lib->vba.HRatio[k] 3710 / dml_ceil( 3711 mode_lib->vba.htaps[k] 3712 / 6.0, 3713 1.0)); 3714 } else { 3715 locals->PSCL_FACTOR[k] = dml_min( 3716 mode_lib->vba.MaxDCHUBToPSCLThroughput, 3717 mode_lib->vba.MaxPSCLToLBThroughput); 3718 } 3719 if (locals->BytePerPixelInDETC[k] == 0.0) { 3720 locals->PSCL_FACTOR_CHROMA[k] = 0.0; 3721 locals->MinDPPCLKUsingSingleDPP[k] = 3722 mode_lib->vba.PixelClock[k] 3723 * dml_max3( 3724 mode_lib->vba.vtaps[k] / 6.0 3725 * dml_min( 3726 1.0, 3727 mode_lib->vba.HRatio[k]), 3728 mode_lib->vba.HRatio[k] 3729 * mode_lib->vba.VRatio[k] 3730 / locals->PSCL_FACTOR[k], 3731 1.0); 3732 if ((mode_lib->vba.htaps[k] > 6.0 || mode_lib->vba.vtaps[k] > 6.0) 3733 && locals->MinDPPCLKUsingSingleDPP[k] 3734 < 2.0 * mode_lib->vba.PixelClock[k]) { 3735 locals->MinDPPCLKUsingSingleDPP[k] = 2.0 3736 * mode_lib->vba.PixelClock[k]; 3737 } 3738 } else { 3739 if (mode_lib->vba.HRatio[k] / 2.0 > 1.0) { 3740 locals->PSCL_FACTOR_CHROMA[k] = 3741 dml_min( 3742 mode_lib->vba.MaxDCHUBToPSCLThroughput, 3743 mode_lib->vba.MaxPSCLToLBThroughput 3744 * mode_lib->vba.HRatio[k] 3745 / 2.0 3746 / dml_ceil( 3747 mode_lib->vba.HTAPsChroma[k] 3748 / 6.0, 3749 1.0)); 3750 } else { 3751 locals->PSCL_FACTOR_CHROMA[k] = dml_min( 3752 mode_lib->vba.MaxDCHUBToPSCLThroughput, 3753 mode_lib->vba.MaxPSCLToLBThroughput); 3754 } 3755 locals->MinDPPCLKUsingSingleDPP[k] = 3756 mode_lib->vba.PixelClock[k] 3757 * dml_max5( 3758 mode_lib->vba.vtaps[k] / 6.0 3759 * dml_min( 3760 1.0, 3761 mode_lib->vba.HRatio[k]), 3762 mode_lib->vba.HRatio[k] 3763 * mode_lib->vba.VRatio[k] 3764 / locals->PSCL_FACTOR[k], 3765 mode_lib->vba.VTAPsChroma[k] 3766 / 6.0 3767 * dml_min( 3768 1.0, 3769 mode_lib->vba.HRatio[k] 3770 / 2.0), 3771 mode_lib->vba.HRatio[k] 3772 * mode_lib->vba.VRatio[k] 3773 / 4.0 3774 / locals->PSCL_FACTOR_CHROMA[k], 3775 1.0); 3776 if ((mode_lib->vba.htaps[k] > 6.0 || mode_lib->vba.vtaps[k] > 6.0 3777 || mode_lib->vba.HTAPsChroma[k] > 6.0 3778 || mode_lib->vba.VTAPsChroma[k] > 6.0) 3779 && locals->MinDPPCLKUsingSingleDPP[k] 3780 < 2.0 * mode_lib->vba.PixelClock[k]) { 3781 locals->MinDPPCLKUsingSingleDPP[k] = 2.0 3782 * mode_lib->vba.PixelClock[k]; 3783 } 3784 } 3785 } 3786 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3787 Calculate256BBlockSizes( 3788 mode_lib->vba.SourcePixelFormat[k], 3789 mode_lib->vba.SurfaceTiling[k], 3790 dml_ceil(locals->BytePerPixelInDETY[k], 1.0), 3791 dml_ceil(locals->BytePerPixelInDETC[k], 2.0), 3792 &locals->Read256BlockHeightY[k], 3793 &locals->Read256BlockHeightC[k], 3794 &locals->Read256BlockWidthY[k], 3795 &locals->Read256BlockWidthC[k]); 3796 if (mode_lib->vba.SourceScan[k] == dm_horz) { 3797 locals->MaxSwathHeightY[k] = locals->Read256BlockHeightY[k]; 3798 locals->MaxSwathHeightC[k] = locals->Read256BlockHeightC[k]; 3799 } else { 3800 locals->MaxSwathHeightY[k] = locals->Read256BlockWidthY[k]; 3801 locals->MaxSwathHeightC[k] = locals->Read256BlockWidthC[k]; 3802 } 3803 if ((mode_lib->vba.SourcePixelFormat[k] == dm_444_64 3804 || mode_lib->vba.SourcePixelFormat[k] == dm_444_32 3805 || mode_lib->vba.SourcePixelFormat[k] == dm_444_16 3806 || mode_lib->vba.SourcePixelFormat[k] == dm_mono_16 3807 || mode_lib->vba.SourcePixelFormat[k] == dm_mono_8)) { 3808 if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear 3809 || (mode_lib->vba.SourcePixelFormat[k] == dm_444_64 3810 && (mode_lib->vba.SurfaceTiling[k] 3811 == dm_sw_4kb_s 3812 || mode_lib->vba.SurfaceTiling[k] 3813 == dm_sw_4kb_s_x 3814 || mode_lib->vba.SurfaceTiling[k] 3815 == dm_sw_64kb_s 3816 || mode_lib->vba.SurfaceTiling[k] 3817 == dm_sw_64kb_s_t 3818 || mode_lib->vba.SurfaceTiling[k] 3819 == dm_sw_64kb_s_x 3820 || mode_lib->vba.SurfaceTiling[k] 3821 == dm_sw_var_s 3822 || mode_lib->vba.SurfaceTiling[k] 3823 == dm_sw_var_s_x) 3824 && mode_lib->vba.SourceScan[k] == dm_horz)) { 3825 locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k]; 3826 } else { 3827 locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k] 3828 / 2.0; 3829 } 3830 locals->MinSwathHeightC[k] = locals->MaxSwathHeightC[k]; 3831 } else { 3832 if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear) { 3833 locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k]; 3834 locals->MinSwathHeightC[k] = locals->MaxSwathHeightC[k]; 3835 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_8 3836 && mode_lib->vba.SourceScan[k] == dm_horz) { 3837 locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k] 3838 / 2.0; 3839 locals->MinSwathHeightC[k] = locals->MaxSwathHeightC[k]; 3840 } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_10 3841 && mode_lib->vba.SourceScan[k] == dm_horz) { 3842 locals->MinSwathHeightC[k] = locals->MaxSwathHeightC[k] 3843 / 2.0; 3844 locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k]; 3845 } else { 3846 locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k]; 3847 locals->MinSwathHeightC[k] = locals->MaxSwathHeightC[k]; 3848 } 3849 } 3850 if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear) { 3851 mode_lib->vba.MaximumSwathWidthSupport = 8192.0; 3852 } else { 3853 mode_lib->vba.MaximumSwathWidthSupport = 5120.0; 3854 } 3855 mode_lib->vba.MaximumSwathWidthInDETBuffer = 3856 dml_min( 3857 mode_lib->vba.MaximumSwathWidthSupport, 3858 mode_lib->vba.DETBufferSizeInKByte * 1024.0 / 2.0 3859 / (locals->BytePerPixelInDETY[k] 3860 * locals->MinSwathHeightY[k] 3861 + locals->BytePerPixelInDETC[k] 3862 / 2.0 3863 * locals->MinSwathHeightC[k])); 3864 if (locals->BytePerPixelInDETC[k] == 0.0) { 3865 mode_lib->vba.MaximumSwathWidthInLineBuffer = 3866 mode_lib->vba.LineBufferSize 3867 * dml_max(mode_lib->vba.HRatio[k], 1.0) 3868 / mode_lib->vba.LBBitPerPixel[k] 3869 / (mode_lib->vba.vtaps[k] 3870 + dml_max( 3871 dml_ceil( 3872 mode_lib->vba.VRatio[k], 3873 1.0) 3874 - 2, 3875 0.0)); 3876 } else { 3877 mode_lib->vba.MaximumSwathWidthInLineBuffer = 3878 dml_min( 3879 mode_lib->vba.LineBufferSize 3880 * dml_max( 3881 mode_lib->vba.HRatio[k], 3882 1.0) 3883 / mode_lib->vba.LBBitPerPixel[k] 3884 / (mode_lib->vba.vtaps[k] 3885 + dml_max( 3886 dml_ceil( 3887 mode_lib->vba.VRatio[k], 3888 1.0) 3889 - 2, 3890 0.0)), 3891 2.0 * mode_lib->vba.LineBufferSize 3892 * dml_max( 3893 mode_lib->vba.HRatio[k] 3894 / 2.0, 3895 1.0) 3896 / mode_lib->vba.LBBitPerPixel[k] 3897 / (mode_lib->vba.VTAPsChroma[k] 3898 + dml_max( 3899 dml_ceil( 3900 mode_lib->vba.VRatio[k] 3901 / 2.0, 3902 1.0) 3903 - 2, 3904 0.0))); 3905 } 3906 locals->MaximumSwathWidth[k] = dml_min( 3907 mode_lib->vba.MaximumSwathWidthInDETBuffer, 3908 mode_lib->vba.MaximumSwathWidthInLineBuffer); 3909 } 3910 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 3911 double MaxMaxDispclkRoundedDown = RoundToDFSGranularityDown( 3912 mode_lib->vba.MaxDispclk[mode_lib->vba.soc.num_states], 3913 mode_lib->vba.DISPCLKDPPCLKVCOSpeed); 3914 3915 for (j = 0; j < 2; j++) { 3916 mode_lib->vba.MaxDispclkRoundedDownToDFSGranularity = RoundToDFSGranularityDown( 3917 mode_lib->vba.MaxDispclk[i], 3918 mode_lib->vba.DISPCLKDPPCLKVCOSpeed); 3919 mode_lib->vba.MaxDppclkRoundedDownToDFSGranularity = RoundToDFSGranularityDown( 3920 mode_lib->vba.MaxDppclk[i], 3921 mode_lib->vba.DISPCLKDPPCLKVCOSpeed); 3922 locals->RequiredDISPCLK[i][j] = 0.0; 3923 locals->DISPCLK_DPPCLK_Support[i][j] = true; 3924 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 3925 mode_lib->vba.PlaneRequiredDISPCLKWithoutODMCombine = 3926 mode_lib->vba.PixelClock[k] * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) 3927 * (1.0 + mode_lib->vba.DISPCLKRampingMargin / 100.0); 3928 if (mode_lib->vba.PlaneRequiredDISPCLKWithoutODMCombine >= mode_lib->vba.MaxDispclk[i] 3929 && i == mode_lib->vba.soc.num_states) 3930 mode_lib->vba.PlaneRequiredDISPCLKWithoutODMCombine = mode_lib->vba.PixelClock[k] 3931 * (1 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); 3932 3933 mode_lib->vba.PlaneRequiredDISPCLKWithODMCombine = mode_lib->vba.PixelClock[k] / 2 3934 * (1 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) * (1 + mode_lib->vba.DISPCLKRampingMargin / 100.0); 3935 if (mode_lib->vba.PlaneRequiredDISPCLKWithODMCombine >= mode_lib->vba.MaxDispclk[i] 3936 && i == mode_lib->vba.soc.num_states) 3937 mode_lib->vba.PlaneRequiredDISPCLKWithODMCombine = mode_lib->vba.PixelClock[k] / 2 3938 * (1 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); 3939 3940 locals->ODMCombineEnablePerState[i][k] = false; 3941 mode_lib->vba.PlaneRequiredDISPCLK = mode_lib->vba.PlaneRequiredDISPCLKWithoutODMCombine; 3942 if (mode_lib->vba.ODMCapability) { 3943 if (locals->PlaneRequiredDISPCLKWithoutODMCombine > MaxMaxDispclkRoundedDown) { 3944 locals->ODMCombineEnablePerState[i][k] = true; 3945 mode_lib->vba.PlaneRequiredDISPCLK = mode_lib->vba.PlaneRequiredDISPCLKWithODMCombine; 3946 } else if (locals->DSCEnabled[k] && (locals->HActive[k] > DCN20_MAX_DSC_IMAGE_WIDTH)) { 3947 locals->ODMCombineEnablePerState[i][k] = true; 3948 mode_lib->vba.PlaneRequiredDISPCLK = mode_lib->vba.PlaneRequiredDISPCLKWithODMCombine; 3949 } else if (locals->HActive[k] > DCN20_MAX_420_IMAGE_WIDTH && locals->OutputFormat[k] == dm_420) { 3950 locals->ODMCombineEnablePerState[i][k] = true; 3951 mode_lib->vba.PlaneRequiredDISPCLK = mode_lib->vba.PlaneRequiredDISPCLKWithODMCombine; 3952 } 3953 } 3954 3955 if (locals->MinDPPCLKUsingSingleDPP[k] * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) <= mode_lib->vba.MaxDppclkRoundedDownToDFSGranularity 3956 && locals->SwathWidthYSingleDPP[k] <= locals->MaximumSwathWidth[k] 3957 && locals->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_disabled) { 3958 locals->NoOfDPP[i][j][k] = 1; 3959 locals->RequiredDPPCLK[i][j][k] = 3960 locals->MinDPPCLKUsingSingleDPP[k] * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); 3961 } else { 3962 locals->NoOfDPP[i][j][k] = 2; 3963 locals->RequiredDPPCLK[i][j][k] = 3964 locals->MinDPPCLKUsingSingleDPP[k] * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) / 2.0; 3965 } 3966 locals->RequiredDISPCLK[i][j] = dml_max( 3967 locals->RequiredDISPCLK[i][j], 3968 mode_lib->vba.PlaneRequiredDISPCLK); 3969 if ((locals->MinDPPCLKUsingSingleDPP[k] / locals->NoOfDPP[i][j][k] * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) 3970 > mode_lib->vba.MaxDppclkRoundedDownToDFSGranularity) 3971 || (mode_lib->vba.PlaneRequiredDISPCLK > mode_lib->vba.MaxDispclkRoundedDownToDFSGranularity)) { 3972 locals->DISPCLK_DPPCLK_Support[i][j] = false; 3973 } 3974 } 3975 locals->TotalNumberOfActiveDPP[i][j] = 0.0; 3976 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) 3977 locals->TotalNumberOfActiveDPP[i][j] = locals->TotalNumberOfActiveDPP[i][j] + locals->NoOfDPP[i][j][k]; 3978 if (j == 1) { 3979 while (locals->TotalNumberOfActiveDPP[i][j] < mode_lib->vba.MaxNumDPP 3980 && locals->TotalNumberOfActiveDPP[i][j] < 2 * mode_lib->vba.NumberOfActivePlanes) { 3981 double BWOfNonSplitPlaneOfMaximumBandwidth; 3982 unsigned int NumberOfNonSplitPlaneOfMaximumBandwidth; 3983 3984 BWOfNonSplitPlaneOfMaximumBandwidth = 0; 3985 NumberOfNonSplitPlaneOfMaximumBandwidth = 0; 3986 for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { 3987 if (locals->ReadBandwidth[k] > BWOfNonSplitPlaneOfMaximumBandwidth && locals->NoOfDPP[i][j][k] == 1) { 3988 BWOfNonSplitPlaneOfMaximumBandwidth = locals->ReadBandwidth[k]; 3989 NumberOfNonSplitPlaneOfMaximumBandwidth = k; 3990 } 3991 } 3992 locals->NoOfDPP[i][j][NumberOfNonSplitPlaneOfMaximumBandwidth] = 2; 3993 locals->RequiredDPPCLK[i][j][NumberOfNonSplitPlaneOfMaximumBandwidth] = 3994 locals->MinDPPCLKUsingSingleDPP[NumberOfNonSplitPlaneOfMaximumBandwidth] 3995 * (1 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100) / 2; 3996 locals->TotalNumberOfActiveDPP[i][j] = locals->TotalNumberOfActiveDPP[i][j] + 1; 3997 } 3998 } 3999 if (locals->TotalNumberOfActiveDPP[i][j] > mode_lib->vba.MaxNumDPP) { 4000 locals->RequiredDISPCLK[i][j] = 0.0; 4001 locals->DISPCLK_DPPCLK_Support[i][j] = true; 4002 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4003 locals->ODMCombineEnablePerState[i][k] = false; 4004 if (locals->SwathWidthYSingleDPP[k] <= locals->MaximumSwathWidth[k]) { 4005 locals->NoOfDPP[i][j][k] = 1; 4006 locals->RequiredDPPCLK[i][j][k] = locals->MinDPPCLKUsingSingleDPP[k] 4007 * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); 4008 } else { 4009 locals->NoOfDPP[i][j][k] = 2; 4010 locals->RequiredDPPCLK[i][j][k] = locals->MinDPPCLKUsingSingleDPP[k] 4011 * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) / 2.0; 4012 } 4013 if (i != mode_lib->vba.soc.num_states) { 4014 mode_lib->vba.PlaneRequiredDISPCLK = 4015 mode_lib->vba.PixelClock[k] 4016 * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) 4017 * (1.0 + mode_lib->vba.DISPCLKRampingMargin / 100.0); 4018 } else { 4019 mode_lib->vba.PlaneRequiredDISPCLK = mode_lib->vba.PixelClock[k] 4020 * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); 4021 } 4022 locals->RequiredDISPCLK[i][j] = dml_max( 4023 locals->RequiredDISPCLK[i][j], 4024 mode_lib->vba.PlaneRequiredDISPCLK); 4025 if (locals->MinDPPCLKUsingSingleDPP[k] / locals->NoOfDPP[i][j][k] * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) 4026 > mode_lib->vba.MaxDppclkRoundedDownToDFSGranularity 4027 || mode_lib->vba.PlaneRequiredDISPCLK > mode_lib->vba.MaxDispclkRoundedDownToDFSGranularity) 4028 locals->DISPCLK_DPPCLK_Support[i][j] = false; 4029 } 4030 locals->TotalNumberOfActiveDPP[i][j] = 0.0; 4031 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) 4032 locals->TotalNumberOfActiveDPP[i][j] = locals->TotalNumberOfActiveDPP[i][j] + locals->NoOfDPP[i][j][k]; 4033 } 4034 locals->RequiredDISPCLK[i][j] = dml_max( 4035 locals->RequiredDISPCLK[i][j], 4036 mode_lib->vba.WritebackRequiredDISPCLK); 4037 if (mode_lib->vba.MaxDispclkRoundedDownToDFSGranularity 4038 < mode_lib->vba.WritebackRequiredDISPCLK) { 4039 locals->DISPCLK_DPPCLK_Support[i][j] = false; 4040 } 4041 } 4042 } 4043 /*Viewport Size Check*/ 4044 4045 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 4046 locals->ViewportSizeSupport[i][0] = true; 4047 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4048 if (locals->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_2to1) { 4049 if (dml_min(locals->SwathWidthYSingleDPP[k], dml_round(mode_lib->vba.HActive[k] / 2.0 * mode_lib->vba.HRatio[k])) 4050 > locals->MaximumSwathWidth[k]) { 4051 locals->ViewportSizeSupport[i][0] = false; 4052 } 4053 } else { 4054 if (locals->SwathWidthYSingleDPP[k] / 2.0 > locals->MaximumSwathWidth[k]) { 4055 locals->ViewportSizeSupport[i][0] = false; 4056 } 4057 } 4058 } 4059 } 4060 /*Total Available Pipes Support Check*/ 4061 4062 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 4063 for (j = 0; j < 2; j++) { 4064 if (locals->TotalNumberOfActiveDPP[i][j] <= mode_lib->vba.MaxNumDPP) 4065 locals->TotalAvailablePipesSupport[i][j] = true; 4066 else 4067 locals->TotalAvailablePipesSupport[i][j] = false; 4068 } 4069 } 4070 /*Total Available OTG Support Check*/ 4071 4072 mode_lib->vba.TotalNumberOfActiveOTG = 0.0; 4073 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4074 if (mode_lib->vba.BlendingAndTiming[k] == k) { 4075 mode_lib->vba.TotalNumberOfActiveOTG = mode_lib->vba.TotalNumberOfActiveOTG 4076 + 1.0; 4077 } 4078 } 4079 if (mode_lib->vba.TotalNumberOfActiveOTG <= mode_lib->vba.MaxNumOTG) { 4080 mode_lib->vba.NumberOfOTGSupport = true; 4081 } else { 4082 mode_lib->vba.NumberOfOTGSupport = false; 4083 } 4084 /*Display IO and DSC Support Check*/ 4085 4086 mode_lib->vba.NonsupportedDSCInputBPC = false; 4087 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4088 if (!(mode_lib->vba.DSCInputBitPerComponent[k] == 12.0 4089 || mode_lib->vba.DSCInputBitPerComponent[k] == 10.0 4090 || mode_lib->vba.DSCInputBitPerComponent[k] == 8.0)) { 4091 mode_lib->vba.NonsupportedDSCInputBPC = true; 4092 } 4093 } 4094 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 4095 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4096 locals->RequiresDSC[i][k] = 0; 4097 locals->RequiresFEC[i][k] = 0; 4098 if (mode_lib->vba.BlendingAndTiming[k] == k) { 4099 if (mode_lib->vba.Output[k] == dm_hdmi) { 4100 locals->RequiresDSC[i][k] = 0; 4101 locals->RequiresFEC[i][k] = 0; 4102 locals->OutputBppPerState[i][k] = TruncToValidBPP( 4103 dml_min(600.0, mode_lib->vba.PHYCLKPerState[i]) / mode_lib->vba.PixelClockBackEnd[k] * 24, 4104 false, 4105 mode_lib->vba.Output[k], 4106 mode_lib->vba.OutputFormat[k], 4107 mode_lib->vba.DSCInputBitPerComponent[k]); 4108 } else if (mode_lib->vba.Output[k] == dm_dp 4109 || mode_lib->vba.Output[k] == dm_edp) { 4110 if (mode_lib->vba.Output[k] == dm_edp) { 4111 mode_lib->vba.EffectiveFECOverhead = 0.0; 4112 } else { 4113 mode_lib->vba.EffectiveFECOverhead = 4114 mode_lib->vba.FECOverhead; 4115 } 4116 if (mode_lib->vba.PHYCLKPerState[i] >= 270.0) { 4117 mode_lib->vba.Outbpp = TruncToValidBPP( 4118 (1.0 - mode_lib->vba.Downspreading / 100.0) * 270.0 4119 * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0, 4120 false, 4121 mode_lib->vba.Output[k], 4122 mode_lib->vba.OutputFormat[k], 4123 mode_lib->vba.DSCInputBitPerComponent[k]); 4124 mode_lib->vba.OutbppDSC = TruncToValidBPP( 4125 (1.0 - mode_lib->vba.Downspreading / 100.0) * (1.0 - mode_lib->vba.EffectiveFECOverhead / 100.0) * 270.0 4126 * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0, 4127 true, 4128 mode_lib->vba.Output[k], 4129 mode_lib->vba.OutputFormat[k], 4130 mode_lib->vba.DSCInputBitPerComponent[k]); 4131 if (mode_lib->vba.DSCEnabled[k] == true) { 4132 locals->RequiresDSC[i][k] = true; 4133 if (mode_lib->vba.Output[k] == dm_dp) { 4134 locals->RequiresFEC[i][k] = true; 4135 } else { 4136 locals->RequiresFEC[i][k] = false; 4137 } 4138 mode_lib->vba.Outbpp = mode_lib->vba.OutbppDSC; 4139 } else { 4140 locals->RequiresDSC[i][k] = false; 4141 locals->RequiresFEC[i][k] = false; 4142 } 4143 locals->OutputBppPerState[i][k] = mode_lib->vba.Outbpp; 4144 } 4145 if (mode_lib->vba.Outbpp == BPP_INVALID && mode_lib->vba.PHYCLKPerState[i] >= 540.0) { 4146 mode_lib->vba.Outbpp = TruncToValidBPP( 4147 (1.0 - mode_lib->vba.Downspreading / 100.0) * 540.0 4148 * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0, 4149 false, 4150 mode_lib->vba.Output[k], 4151 mode_lib->vba.OutputFormat[k], 4152 mode_lib->vba.DSCInputBitPerComponent[k]); 4153 mode_lib->vba.OutbppDSC = TruncToValidBPP( 4154 (1.0 - mode_lib->vba.Downspreading / 100.0) * (1.0 - mode_lib->vba.EffectiveFECOverhead / 100.0) * 540.0 4155 * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0, 4156 true, 4157 mode_lib->vba.Output[k], 4158 mode_lib->vba.OutputFormat[k], 4159 mode_lib->vba.DSCInputBitPerComponent[k]); 4160 if (mode_lib->vba.DSCEnabled[k] == true) { 4161 locals->RequiresDSC[i][k] = true; 4162 if (mode_lib->vba.Output[k] == dm_dp) { 4163 locals->RequiresFEC[i][k] = true; 4164 } else { 4165 locals->RequiresFEC[i][k] = false; 4166 } 4167 mode_lib->vba.Outbpp = mode_lib->vba.OutbppDSC; 4168 } else { 4169 locals->RequiresDSC[i][k] = false; 4170 locals->RequiresFEC[i][k] = false; 4171 } 4172 locals->OutputBppPerState[i][k] = mode_lib->vba.Outbpp; 4173 } 4174 if (mode_lib->vba.Outbpp == BPP_INVALID 4175 && mode_lib->vba.PHYCLKPerState[i] 4176 >= 810.0) { 4177 mode_lib->vba.Outbpp = TruncToValidBPP( 4178 (1.0 - mode_lib->vba.Downspreading / 100.0) * 810.0 4179 * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0, 4180 false, 4181 mode_lib->vba.Output[k], 4182 mode_lib->vba.OutputFormat[k], 4183 mode_lib->vba.DSCInputBitPerComponent[k]); 4184 mode_lib->vba.OutbppDSC = TruncToValidBPP( 4185 (1.0 - mode_lib->vba.Downspreading / 100.0) * (1.0 - mode_lib->vba.EffectiveFECOverhead / 100.0) * 810.0 4186 * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0, 4187 true, 4188 mode_lib->vba.Output[k], 4189 mode_lib->vba.OutputFormat[k], 4190 mode_lib->vba.DSCInputBitPerComponent[k]); 4191 if (mode_lib->vba.DSCEnabled[k] == true || mode_lib->vba.Outbpp == BPP_INVALID) { 4192 locals->RequiresDSC[i][k] = true; 4193 if (mode_lib->vba.Output[k] == dm_dp) { 4194 locals->RequiresFEC[i][k] = true; 4195 } else { 4196 locals->RequiresFEC[i][k] = false; 4197 } 4198 mode_lib->vba.Outbpp = mode_lib->vba.OutbppDSC; 4199 } else { 4200 locals->RequiresDSC[i][k] = false; 4201 locals->RequiresFEC[i][k] = false; 4202 } 4203 locals->OutputBppPerState[i][k] = 4204 mode_lib->vba.Outbpp; 4205 } 4206 } 4207 } else { 4208 locals->OutputBppPerState[i][k] = BPP_BLENDED_PIPE; 4209 } 4210 } 4211 } 4212 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 4213 locals->DIOSupport[i] = true; 4214 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4215 if (locals->OutputBppPerState[i][k] == BPP_INVALID 4216 || (mode_lib->vba.OutputFormat[k] == dm_420 4217 && mode_lib->vba.Interlace[k] == true 4218 && mode_lib->vba.ProgressiveToInterlaceUnitInOPP == true)) { 4219 locals->DIOSupport[i] = false; 4220 } 4221 } 4222 } 4223 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 4224 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4225 locals->DSCCLKRequiredMoreThanSupported[i] = false; 4226 if (mode_lib->vba.BlendingAndTiming[k] == k) { 4227 if ((mode_lib->vba.Output[k] == dm_dp 4228 || mode_lib->vba.Output[k] == dm_edp)) { 4229 if (mode_lib->vba.OutputFormat[k] == dm_420 4230 || mode_lib->vba.OutputFormat[k] 4231 == dm_n422) { 4232 mode_lib->vba.DSCFormatFactor = 2; 4233 } else { 4234 mode_lib->vba.DSCFormatFactor = 1; 4235 } 4236 if (locals->RequiresDSC[i][k] == true) { 4237 if (locals->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_2to1) { 4238 if (mode_lib->vba.PixelClockBackEnd[k] / 6.0 / mode_lib->vba.DSCFormatFactor 4239 > (1.0 - mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) * mode_lib->vba.MaxDSCCLK[i]) { 4240 locals->DSCCLKRequiredMoreThanSupported[i] = 4241 true; 4242 } 4243 } else { 4244 if (mode_lib->vba.PixelClockBackEnd[k] / 3.0 / mode_lib->vba.DSCFormatFactor 4245 > (1.0 - mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) * mode_lib->vba.MaxDSCCLK[i]) { 4246 locals->DSCCLKRequiredMoreThanSupported[i] = 4247 true; 4248 } 4249 } 4250 } 4251 } 4252 } 4253 } 4254 } 4255 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 4256 locals->NotEnoughDSCUnits[i] = false; 4257 mode_lib->vba.TotalDSCUnitsRequired = 0.0; 4258 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4259 if (locals->RequiresDSC[i][k] == true) { 4260 if (locals->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_2to1) { 4261 mode_lib->vba.TotalDSCUnitsRequired = 4262 mode_lib->vba.TotalDSCUnitsRequired + 2.0; 4263 } else { 4264 mode_lib->vba.TotalDSCUnitsRequired = 4265 mode_lib->vba.TotalDSCUnitsRequired + 1.0; 4266 } 4267 } 4268 } 4269 if (mode_lib->vba.TotalDSCUnitsRequired > mode_lib->vba.NumberOfDSC) { 4270 locals->NotEnoughDSCUnits[i] = true; 4271 } 4272 } 4273 /*DSC Delay per state*/ 4274 4275 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 4276 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4277 if (mode_lib->vba.BlendingAndTiming[k] != k) { 4278 mode_lib->vba.slices = 0; 4279 } else if (locals->RequiresDSC[i][k] == 0 4280 || locals->RequiresDSC[i][k] == false) { 4281 mode_lib->vba.slices = 0; 4282 } else if (mode_lib->vba.PixelClockBackEnd[k] > 3200.0) { 4283 mode_lib->vba.slices = dml_ceil( 4284 mode_lib->vba.PixelClockBackEnd[k] / 400.0, 4285 4.0); 4286 } else if (mode_lib->vba.PixelClockBackEnd[k] > 1360.0) { 4287 mode_lib->vba.slices = 8.0; 4288 } else if (mode_lib->vba.PixelClockBackEnd[k] > 680.0) { 4289 mode_lib->vba.slices = 4.0; 4290 } else if (mode_lib->vba.PixelClockBackEnd[k] > 340.0) { 4291 mode_lib->vba.slices = 2.0; 4292 } else { 4293 mode_lib->vba.slices = 1.0; 4294 } 4295 if (locals->OutputBppPerState[i][k] == BPP_BLENDED_PIPE 4296 || locals->OutputBppPerState[i][k] == BPP_INVALID) { 4297 mode_lib->vba.bpp = 0.0; 4298 } else { 4299 mode_lib->vba.bpp = locals->OutputBppPerState[i][k]; 4300 } 4301 if (locals->RequiresDSC[i][k] == true && mode_lib->vba.bpp != 0.0) { 4302 if (locals->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_disabled) { 4303 locals->DSCDelayPerState[i][k] = 4304 dscceComputeDelay( 4305 mode_lib->vba.DSCInputBitPerComponent[k], 4306 mode_lib->vba.bpp, 4307 dml_ceil( 4308 mode_lib->vba.HActive[k] 4309 / mode_lib->vba.slices, 4310 1.0), 4311 mode_lib->vba.slices, 4312 mode_lib->vba.OutputFormat[k]) 4313 + dscComputeDelay( 4314 mode_lib->vba.OutputFormat[k]); 4315 } else { 4316 locals->DSCDelayPerState[i][k] = 4317 2.0 * (dscceComputeDelay( 4318 mode_lib->vba.DSCInputBitPerComponent[k], 4319 mode_lib->vba.bpp, 4320 dml_ceil(mode_lib->vba.HActive[k] / mode_lib->vba.slices, 1.0), 4321 mode_lib->vba.slices / 2, 4322 mode_lib->vba.OutputFormat[k]) 4323 + dscComputeDelay(mode_lib->vba.OutputFormat[k])); 4324 } 4325 locals->DSCDelayPerState[i][k] = 4326 locals->DSCDelayPerState[i][k] * mode_lib->vba.PixelClock[k] / mode_lib->vba.PixelClockBackEnd[k]; 4327 } else { 4328 locals->DSCDelayPerState[i][k] = 0.0; 4329 } 4330 } 4331 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4332 for (m = 0; m <= mode_lib->vba.NumberOfActivePlanes - 1; m++) { 4333 for (j = 0; j <= mode_lib->vba.NumberOfActivePlanes - 1; j++) { 4334 if (mode_lib->vba.BlendingAndTiming[k] == m && locals->RequiresDSC[i][m] == true) 4335 locals->DSCDelayPerState[i][k] = locals->DSCDelayPerState[i][m]; 4336 } 4337 } 4338 } 4339 } 4340 4341 //Prefetch Check 4342 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 4343 for (j = 0; j < 2; j++) { 4344 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4345 if (locals->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_2to1) 4346 locals->SwathWidthYPerState[i][j][k] = dml_min(locals->SwathWidthYSingleDPP[k], dml_round(locals->HActive[k] / 2 * locals->HRatio[k])); 4347 else 4348 locals->SwathWidthYPerState[i][j][k] = locals->SwathWidthYSingleDPP[k] / locals->NoOfDPP[i][j][k]; 4349 locals->SwathWidthGranularityY = 256 / dml_ceil(locals->BytePerPixelInDETY[k], 1) / locals->MaxSwathHeightY[k]; 4350 locals->RoundedUpMaxSwathSizeBytesY = (dml_ceil(locals->SwathWidthYPerState[i][j][k] - 1, locals->SwathWidthGranularityY) 4351 + locals->SwathWidthGranularityY) * locals->BytePerPixelInDETY[k] * locals->MaxSwathHeightY[k]; 4352 if (locals->SourcePixelFormat[k] == dm_420_10) { 4353 locals->RoundedUpMaxSwathSizeBytesY = dml_ceil(locals->RoundedUpMaxSwathSizeBytesY, 256) + 256; 4354 } 4355 if (locals->MaxSwathHeightC[k] > 0) { 4356 locals->SwathWidthGranularityC = 256 / dml_ceil(locals->BytePerPixelInDETC[k], 2) / locals->MaxSwathHeightC[k]; 4357 4358 locals->RoundedUpMaxSwathSizeBytesC = (dml_ceil(locals->SwathWidthYPerState[i][j][k] / 2 - 1, locals->SwathWidthGranularityC) 4359 + locals->SwathWidthGranularityC) * locals->BytePerPixelInDETC[k] * locals->MaxSwathHeightC[k]; 4360 } 4361 if (locals->SourcePixelFormat[k] == dm_420_10) { 4362 locals->RoundedUpMaxSwathSizeBytesC = dml_ceil(locals->RoundedUpMaxSwathSizeBytesC, 256) + 256; 4363 } else { 4364 locals->RoundedUpMaxSwathSizeBytesC = 0; 4365 } 4366 4367 if (locals->RoundedUpMaxSwathSizeBytesY + locals->RoundedUpMaxSwathSizeBytesC <= locals->DETBufferSizeInKByte * 1024 / 2) { 4368 locals->SwathHeightYPerState[i][j][k] = locals->MaxSwathHeightY[k]; 4369 locals->SwathHeightCPerState[i][j][k] = locals->MaxSwathHeightC[k]; 4370 } else { 4371 locals->SwathHeightYPerState[i][j][k] = locals->MinSwathHeightY[k]; 4372 locals->SwathHeightCPerState[i][j][k] = locals->MinSwathHeightC[k]; 4373 } 4374 4375 if (locals->BytePerPixelInDETC[k] == 0) { 4376 locals->LinesInDETLuma = locals->DETBufferSizeInKByte * 1024 / locals->BytePerPixelInDETY[k] / locals->SwathWidthYPerState[i][j][k]; 4377 locals->LinesInDETChroma = 0; 4378 } else if (locals->SwathHeightYPerState[i][j][k] <= locals->SwathHeightCPerState[i][j][k]) { 4379 locals->LinesInDETLuma = locals->DETBufferSizeInKByte * 1024 / 2 / locals->BytePerPixelInDETY[k] / 4380 locals->SwathWidthYPerState[i][j][k]; 4381 locals->LinesInDETChroma = locals->DETBufferSizeInKByte * 1024 / 2 / locals->BytePerPixelInDETC[k] / (locals->SwathWidthYPerState[i][j][k] / 2); 4382 } else { 4383 locals->LinesInDETLuma = locals->DETBufferSizeInKByte * 1024 * 2 / 3 / locals->BytePerPixelInDETY[k] / locals->SwathWidthYPerState[i][j][k]; 4384 locals->LinesInDETChroma = locals->DETBufferSizeInKByte * 1024 / 3 / locals->BytePerPixelInDETY[k] / (locals->SwathWidthYPerState[i][j][k] / 2); 4385 } 4386 4387 locals->EffectiveLBLatencyHidingSourceLinesLuma = dml_min(locals->MaxLineBufferLines, 4388 dml_floor(locals->LineBufferSize / locals->LBBitPerPixel[k] / (locals->SwathWidthYPerState[i][j][k] 4389 / dml_max(locals->HRatio[k], 1)), 1)) - (locals->vtaps[k] - 1); 4390 4391 locals->EffectiveLBLatencyHidingSourceLinesChroma = dml_min(locals->MaxLineBufferLines, 4392 dml_floor(locals->LineBufferSize / locals->LBBitPerPixel[k] 4393 / (locals->SwathWidthYPerState[i][j][k] / 2 4394 / dml_max(locals->HRatio[k] / 2, 1)), 1)) - (locals->VTAPsChroma[k] - 1); 4395 4396 locals->EffectiveDETLBLinesLuma = dml_floor(locals->LinesInDETLuma + dml_min( 4397 locals->LinesInDETLuma * locals->RequiredDISPCLK[i][j] * locals->BytePerPixelInDETY[k] * 4398 locals->PSCL_FACTOR[k] / locals->ReturnBWPerState[i][0], 4399 locals->EffectiveLBLatencyHidingSourceLinesLuma), 4400 locals->SwathHeightYPerState[i][j][k]); 4401 4402 locals->EffectiveDETLBLinesChroma = dml_floor(locals->LinesInDETChroma + dml_min( 4403 locals->LinesInDETChroma * locals->RequiredDISPCLK[i][j] * locals->BytePerPixelInDETC[k] * 4404 locals->PSCL_FACTOR_CHROMA[k] / locals->ReturnBWPerState[i][0], 4405 locals->EffectiveLBLatencyHidingSourceLinesChroma), 4406 locals->SwathHeightCPerState[i][j][k]); 4407 4408 if (locals->BytePerPixelInDETC[k] == 0) { 4409 locals->UrgentLatencySupportUsPerState[i][j][k] = locals->EffectiveDETLBLinesLuma * (locals->HTotal[k] / locals->PixelClock[k]) 4410 / locals->VRatio[k] - locals->EffectiveDETLBLinesLuma * locals->SwathWidthYPerState[i][j][k] * 4411 dml_ceil(locals->BytePerPixelInDETY[k], 1) / (locals->ReturnBWPerState[i][0] / locals->NoOfDPP[i][j][k]); 4412 } else { 4413 locals->UrgentLatencySupportUsPerState[i][j][k] = dml_min( 4414 locals->EffectiveDETLBLinesLuma * (locals->HTotal[k] / locals->PixelClock[k]) 4415 / locals->VRatio[k] - locals->EffectiveDETLBLinesLuma * locals->SwathWidthYPerState[i][j][k] * 4416 dml_ceil(locals->BytePerPixelInDETY[k], 1) / (locals->ReturnBWPerState[i][0] / locals->NoOfDPP[i][j][k]), 4417 locals->EffectiveDETLBLinesChroma * (locals->HTotal[k] / locals->PixelClock[k]) / (locals->VRatio[k] / 2) - 4418 locals->EffectiveDETLBLinesChroma * locals->SwathWidthYPerState[i][j][k] / 2 * 4419 dml_ceil(locals->BytePerPixelInDETC[k], 2) / (locals->ReturnBWPerState[i][0] / locals->NoOfDPP[i][j][k])); 4420 } 4421 } 4422 } 4423 } 4424 4425 for (i = 0; i <= locals->soc.num_states; i++) { 4426 for (j = 0; j < 2; j++) { 4427 locals->UrgentLatencySupport[i][j] = true; 4428 for (k = 0; k < locals->NumberOfActivePlanes; k++) { 4429 if (locals->UrgentLatencySupportUsPerState[i][j][k] < locals->UrgentLatency) 4430 locals->UrgentLatencySupport[i][j] = false; 4431 } 4432 } 4433 } 4434 4435 4436 /*Prefetch Check*/ 4437 for (i = 0; i <= locals->soc.num_states; i++) { 4438 for (j = 0; j < 2; j++) { 4439 locals->TotalNumberOfDCCActiveDPP[i][j] = 0; 4440 for (k = 0; k < locals->NumberOfActivePlanes; k++) { 4441 if (locals->DCCEnable[k] == true) { 4442 locals->TotalNumberOfDCCActiveDPP[i][j] = 4443 locals->TotalNumberOfDCCActiveDPP[i][j] + locals->NoOfDPP[i][j][k]; 4444 } 4445 } 4446 } 4447 } 4448 4449 CalculateMinAndMaxPrefetchMode(locals->AllowDRAMSelfRefreshOrDRAMClockChangeInVblank, &locals->MinPrefetchMode, &locals->MaxPrefetchMode); 4450 4451 locals->MaxTotalVActiveRDBandwidth = 0; 4452 for (k = 0; k < locals->NumberOfActivePlanes; k++) { 4453 locals->MaxTotalVActiveRDBandwidth = locals->MaxTotalVActiveRDBandwidth + locals->ReadBandwidth[k]; 4454 } 4455 4456 for (i = 0; i <= locals->soc.num_states; i++) { 4457 for (j = 0; j < 2; j++) { 4458 for (k = 0; k < locals->NumberOfActivePlanes; k++) { 4459 locals->NoOfDPPThisState[k] = locals->NoOfDPP[i][j][k]; 4460 locals->RequiredDPPCLKThisState[k] = locals->RequiredDPPCLK[i][j][k]; 4461 locals->SwathHeightYThisState[k] = locals->SwathHeightYPerState[i][j][k]; 4462 locals->SwathHeightCThisState[k] = locals->SwathHeightCPerState[i][j][k]; 4463 locals->SwathWidthYThisState[k] = locals->SwathWidthYPerState[i][j][k]; 4464 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0] = dml_max( 4465 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0], 4466 mode_lib->vba.PixelClock[k] / 16.0); 4467 if (mode_lib->vba.BytePerPixelInDETC[k] == 0.0) { 4468 if (mode_lib->vba.VRatio[k] <= 1.0) { 4469 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0] = 4470 dml_max( 4471 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0], 4472 1.1 4473 * dml_ceil( 4474 mode_lib->vba.BytePerPixelInDETY[k], 4475 1.0) 4476 / 64.0 4477 * mode_lib->vba.HRatio[k] 4478 * mode_lib->vba.PixelClock[k] 4479 / mode_lib->vba.NoOfDPP[i][j][k]); 4480 } else { 4481 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0] = 4482 dml_max( 4483 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0], 4484 1.1 4485 * dml_ceil( 4486 mode_lib->vba.BytePerPixelInDETY[k], 4487 1.0) 4488 / 64.0 4489 * mode_lib->vba.PSCL_FACTOR[k] 4490 * mode_lib->vba.RequiredDPPCLK[i][j][k]); 4491 } 4492 } else { 4493 if (mode_lib->vba.VRatio[k] <= 1.0) { 4494 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0] = 4495 dml_max( 4496 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0], 4497 1.1 4498 * dml_ceil( 4499 mode_lib->vba.BytePerPixelInDETY[k], 4500 1.0) 4501 / 32.0 4502 * mode_lib->vba.HRatio[k] 4503 * mode_lib->vba.PixelClock[k] 4504 / mode_lib->vba.NoOfDPP[i][j][k]); 4505 } else { 4506 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0] = 4507 dml_max( 4508 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0], 4509 1.1 4510 * dml_ceil( 4511 mode_lib->vba.BytePerPixelInDETY[k], 4512 1.0) 4513 / 32.0 4514 * mode_lib->vba.PSCL_FACTOR[k] 4515 * mode_lib->vba.RequiredDPPCLK[i][j][k]); 4516 } 4517 if (mode_lib->vba.VRatio[k] / 2.0 <= 1.0) { 4518 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0] = 4519 dml_max( 4520 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0], 4521 1.1 4522 * dml_ceil( 4523 mode_lib->vba.BytePerPixelInDETC[k], 4524 2.0) 4525 / 32.0 4526 * mode_lib->vba.HRatio[k] 4527 / 2.0 4528 * mode_lib->vba.PixelClock[k] 4529 / mode_lib->vba.NoOfDPP[i][j][k]); 4530 } else { 4531 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0] = 4532 dml_max( 4533 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0], 4534 1.1 4535 * dml_ceil( 4536 mode_lib->vba.BytePerPixelInDETC[k], 4537 2.0) 4538 / 32.0 4539 * mode_lib->vba.PSCL_FACTOR_CHROMA[k] 4540 * mode_lib->vba.RequiredDPPCLK[i][j][k]); 4541 } 4542 } 4543 } 4544 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4545 mode_lib->vba.PDEAndMetaPTEBytesPerFrameY = CalculateVMAndRowBytes( 4546 mode_lib, 4547 mode_lib->vba.DCCEnable[k], 4548 mode_lib->vba.Read256BlockHeightY[k], 4549 mode_lib->vba.Read256BlockWidthY[k], 4550 mode_lib->vba.SourcePixelFormat[k], 4551 mode_lib->vba.SurfaceTiling[k], 4552 dml_ceil(mode_lib->vba.BytePerPixelInDETY[k], 1.0), 4553 mode_lib->vba.SourceScan[k], 4554 mode_lib->vba.ViewportWidth[k], 4555 mode_lib->vba.ViewportHeight[k], 4556 mode_lib->vba.SwathWidthYPerState[i][j][k], 4557 mode_lib->vba.GPUVMEnable, 4558 mode_lib->vba.VMMPageSize, 4559 mode_lib->vba.PTEBufferSizeInRequestsLuma, 4560 mode_lib->vba.PDEProcessingBufIn64KBReqs, 4561 mode_lib->vba.PitchY[k], 4562 mode_lib->vba.DCCMetaPitchY[k], 4563 &mode_lib->vba.MacroTileWidthY[k], 4564 &mode_lib->vba.MetaRowBytesY, 4565 &mode_lib->vba.DPTEBytesPerRowY, 4566 &mode_lib->vba.PTEBufferSizeNotExceededY[i][j][k], 4567 &mode_lib->vba.dpte_row_height[k], 4568 &mode_lib->vba.meta_row_height[k]); 4569 mode_lib->vba.PrefetchLinesY[0][0][k] = CalculatePrefetchSourceLines( 4570 mode_lib, 4571 mode_lib->vba.VRatio[k], 4572 mode_lib->vba.vtaps[k], 4573 mode_lib->vba.Interlace[k], 4574 mode_lib->vba.ProgressiveToInterlaceUnitInOPP, 4575 mode_lib->vba.SwathHeightYPerState[i][j][k], 4576 mode_lib->vba.ViewportYStartY[k], 4577 &mode_lib->vba.PrefillY[k], 4578 &mode_lib->vba.MaxNumSwY[k]); 4579 if ((mode_lib->vba.SourcePixelFormat[k] != dm_444_64 4580 && mode_lib->vba.SourcePixelFormat[k] != dm_444_32 4581 && mode_lib->vba.SourcePixelFormat[k] != dm_444_16 4582 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_16 4583 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_8)) { 4584 mode_lib->vba.PDEAndMetaPTEBytesPerFrameC = CalculateVMAndRowBytes( 4585 mode_lib, 4586 mode_lib->vba.DCCEnable[k], 4587 mode_lib->vba.Read256BlockHeightY[k], 4588 mode_lib->vba.Read256BlockWidthY[k], 4589 mode_lib->vba.SourcePixelFormat[k], 4590 mode_lib->vba.SurfaceTiling[k], 4591 dml_ceil(mode_lib->vba.BytePerPixelInDETC[k], 2.0), 4592 mode_lib->vba.SourceScan[k], 4593 mode_lib->vba.ViewportWidth[k] / 2.0, 4594 mode_lib->vba.ViewportHeight[k] / 2.0, 4595 mode_lib->vba.SwathWidthYPerState[i][j][k] / 2.0, 4596 mode_lib->vba.GPUVMEnable, 4597 mode_lib->vba.VMMPageSize, 4598 mode_lib->vba.PTEBufferSizeInRequestsLuma, 4599 mode_lib->vba.PDEProcessingBufIn64KBReqs, 4600 mode_lib->vba.PitchC[k], 4601 0.0, 4602 &mode_lib->vba.MacroTileWidthC[k], 4603 &mode_lib->vba.MetaRowBytesC, 4604 &mode_lib->vba.DPTEBytesPerRowC, 4605 &mode_lib->vba.PTEBufferSizeNotExceededC[i][j][k], 4606 &mode_lib->vba.dpte_row_height_chroma[k], 4607 &mode_lib->vba.meta_row_height_chroma[k]); 4608 mode_lib->vba.PrefetchLinesC[0][0][k] = CalculatePrefetchSourceLines( 4609 mode_lib, 4610 mode_lib->vba.VRatio[k] / 2.0, 4611 mode_lib->vba.VTAPsChroma[k], 4612 mode_lib->vba.Interlace[k], 4613 mode_lib->vba.ProgressiveToInterlaceUnitInOPP, 4614 mode_lib->vba.SwathHeightCPerState[i][j][k], 4615 mode_lib->vba.ViewportYStartC[k], 4616 &mode_lib->vba.PrefillC[k], 4617 &mode_lib->vba.MaxNumSwC[k]); 4618 } else { 4619 mode_lib->vba.PDEAndMetaPTEBytesPerFrameC = 0.0; 4620 mode_lib->vba.MetaRowBytesC = 0.0; 4621 mode_lib->vba.DPTEBytesPerRowC = 0.0; 4622 locals->PrefetchLinesC[0][0][k] = 0.0; 4623 locals->PTEBufferSizeNotExceededC[i][j][k] = true; 4624 locals->PTEBufferSizeInRequestsForLuma = mode_lib->vba.PTEBufferSizeInRequestsLuma + mode_lib->vba.PTEBufferSizeInRequestsChroma; 4625 } 4626 locals->PDEAndMetaPTEBytesPerFrame[0][0][k] = 4627 mode_lib->vba.PDEAndMetaPTEBytesPerFrameY + mode_lib->vba.PDEAndMetaPTEBytesPerFrameC; 4628 locals->MetaRowBytes[0][0][k] = mode_lib->vba.MetaRowBytesY + mode_lib->vba.MetaRowBytesC; 4629 locals->DPTEBytesPerRow[0][0][k] = mode_lib->vba.DPTEBytesPerRowY + mode_lib->vba.DPTEBytesPerRowC; 4630 4631 CalculateActiveRowBandwidth( 4632 mode_lib->vba.GPUVMEnable, 4633 mode_lib->vba.SourcePixelFormat[k], 4634 mode_lib->vba.VRatio[k], 4635 mode_lib->vba.DCCEnable[k], 4636 mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k], 4637 mode_lib->vba.MetaRowBytesY, 4638 mode_lib->vba.MetaRowBytesC, 4639 mode_lib->vba.meta_row_height[k], 4640 mode_lib->vba.meta_row_height_chroma[k], 4641 mode_lib->vba.DPTEBytesPerRowY, 4642 mode_lib->vba.DPTEBytesPerRowC, 4643 mode_lib->vba.dpte_row_height[k], 4644 mode_lib->vba.dpte_row_height_chroma[k], 4645 &mode_lib->vba.meta_row_bw[k], 4646 &mode_lib->vba.dpte_row_bw[k], 4647 &mode_lib->vba.qual_row_bw[k]); 4648 } 4649 mode_lib->vba.ExtraLatency = 4650 mode_lib->vba.UrgentRoundTripAndOutOfOrderLatencyPerState[i] 4651 + (mode_lib->vba.TotalNumberOfActiveDPP[i][j] 4652 * mode_lib->vba.PixelChunkSizeInKByte 4653 + mode_lib->vba.TotalNumberOfDCCActiveDPP[i][j] 4654 * mode_lib->vba.MetaChunkSize) 4655 * 1024.0 4656 / mode_lib->vba.ReturnBWPerState[i][0]; 4657 if (mode_lib->vba.GPUVMEnable == true) { 4658 mode_lib->vba.ExtraLatency = mode_lib->vba.ExtraLatency 4659 + mode_lib->vba.TotalNumberOfActiveDPP[i][j] 4660 * mode_lib->vba.PTEGroupSize 4661 / mode_lib->vba.ReturnBWPerState[i][0]; 4662 } 4663 mode_lib->vba.TimeCalc = 24.0 / mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0]; 4664 4665 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4666 if (mode_lib->vba.BlendingAndTiming[k] == k) { 4667 if (mode_lib->vba.WritebackEnable[k] == true) { 4668 locals->WritebackDelay[i][k] = mode_lib->vba.WritebackLatency 4669 + CalculateWriteBackDelay( 4670 mode_lib->vba.WritebackPixelFormat[k], 4671 mode_lib->vba.WritebackHRatio[k], 4672 mode_lib->vba.WritebackVRatio[k], 4673 mode_lib->vba.WritebackLumaHTaps[k], 4674 mode_lib->vba.WritebackLumaVTaps[k], 4675 mode_lib->vba.WritebackChromaHTaps[k], 4676 mode_lib->vba.WritebackChromaVTaps[k], 4677 mode_lib->vba.WritebackDestinationWidth[k]) / locals->RequiredDISPCLK[i][j]; 4678 } else { 4679 locals->WritebackDelay[i][k] = 0.0; 4680 } 4681 for (m = 0; m <= mode_lib->vba.NumberOfActivePlanes - 1; m++) { 4682 if (mode_lib->vba.BlendingAndTiming[m] == k 4683 && mode_lib->vba.WritebackEnable[m] 4684 == true) { 4685 locals->WritebackDelay[i][k] = dml_max(locals->WritebackDelay[i][k], 4686 mode_lib->vba.WritebackLatency + CalculateWriteBackDelay( 4687 mode_lib->vba.WritebackPixelFormat[m], 4688 mode_lib->vba.WritebackHRatio[m], 4689 mode_lib->vba.WritebackVRatio[m], 4690 mode_lib->vba.WritebackLumaHTaps[m], 4691 mode_lib->vba.WritebackLumaVTaps[m], 4692 mode_lib->vba.WritebackChromaHTaps[m], 4693 mode_lib->vba.WritebackChromaVTaps[m], 4694 mode_lib->vba.WritebackDestinationWidth[m]) / locals->RequiredDISPCLK[i][j]); 4695 } 4696 } 4697 } 4698 } 4699 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4700 for (m = 0; m <= mode_lib->vba.NumberOfActivePlanes - 1; m++) { 4701 if (mode_lib->vba.BlendingAndTiming[k] == m) { 4702 locals->WritebackDelay[i][k] = locals->WritebackDelay[i][m]; 4703 } 4704 } 4705 } 4706 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4707 for (m = 0; m < locals->NumberOfCursors[k]; m++) 4708 locals->cursor_bw[k] = locals->NumberOfCursors[k] * locals->CursorWidth[k][m] * locals->CursorBPP[k][m] 4709 / 8 / (locals->HTotal[k] / locals->PixelClock[k]) * locals->VRatio[k]; 4710 } 4711 4712 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4713 locals->MaximumVStartup[0][0][k] = mode_lib->vba.VTotal[k] - mode_lib->vba.VActive[k] 4714 - dml_max(1.0, dml_ceil(locals->WritebackDelay[i][k] / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]), 1.0)); 4715 } 4716 4717 mode_lib->vba.NextPrefetchMode = mode_lib->vba.MinPrefetchMode; 4718 do { 4719 mode_lib->vba.PrefetchMode[i][j] = mode_lib->vba.NextPrefetchMode; 4720 mode_lib->vba.NextPrefetchMode = mode_lib->vba.NextPrefetchMode + 1; 4721 4722 mode_lib->vba.TWait = CalculateTWait( 4723 mode_lib->vba.PrefetchMode[i][j], 4724 mode_lib->vba.DRAMClockChangeLatency, 4725 mode_lib->vba.UrgentLatency, 4726 mode_lib->vba.SREnterPlusExitTime); 4727 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4728 4729 if (mode_lib->vba.XFCEnabled[k] == true) { 4730 mode_lib->vba.XFCRemoteSurfaceFlipDelay = 4731 CalculateRemoteSurfaceFlipDelay( 4732 mode_lib, 4733 mode_lib->vba.VRatio[k], 4734 locals->SwathWidthYPerState[i][j][k], 4735 dml_ceil(locals->BytePerPixelInDETY[k], 1.0), 4736 mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k], 4737 mode_lib->vba.XFCTSlvVupdateOffset, 4738 mode_lib->vba.XFCTSlvVupdateWidth, 4739 mode_lib->vba.XFCTSlvVreadyOffset, 4740 mode_lib->vba.XFCXBUFLatencyTolerance, 4741 mode_lib->vba.XFCFillBWOverhead, 4742 mode_lib->vba.XFCSlvChunkSize, 4743 mode_lib->vba.XFCBusTransportTime, 4744 mode_lib->vba.TimeCalc, 4745 mode_lib->vba.TWait, 4746 &mode_lib->vba.SrcActiveDrainRate, 4747 &mode_lib->vba.TInitXFill, 4748 &mode_lib->vba.TslvChk); 4749 } else { 4750 mode_lib->vba.XFCRemoteSurfaceFlipDelay = 0.0; 4751 } 4752 4753 CalculateDelayAfterScaler(mode_lib, mode_lib->vba.ReturnBWPerState[i][0], mode_lib->vba.ReadBandwidthLuma[k], mode_lib->vba.ReadBandwidthChroma[k], mode_lib->vba.MaxTotalVActiveRDBandwidth, 4754 mode_lib->vba.DisplayPipeLineDeliveryTimeLuma[k], mode_lib->vba.DisplayPipeLineDeliveryTimeChroma[k], 4755 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], 4756 mode_lib->vba.DPPCLKDelaySubtotal, mode_lib->vba.DPPCLKDelaySCL, mode_lib->vba.DPPCLKDelaySCLLBOnly, mode_lib->vba.DPPCLKDelayCNVCFormater, mode_lib->vba.DPPCLKDelayCNVCCursor, mode_lib->vba.DISPCLKDelaySubtotal, 4757 mode_lib->vba.SwathWidthYPerState[i][j][k] / mode_lib->vba.HRatio[k], mode_lib->vba.OutputFormat[k], mode_lib->vba.HTotal[k], 4758 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, 4759 &mode_lib->vba.DSTXAfterScaler[k], &mode_lib->vba.DSTYAfterScaler[k]); 4760 4761 mode_lib->vba.IsErrorResult[i][j][k] = 4762 CalculatePrefetchSchedule( 4763 mode_lib, 4764 mode_lib->vba.RequiredDPPCLK[i][j][k], 4765 mode_lib->vba.RequiredDISPCLK[i][j], 4766 mode_lib->vba.PixelClock[k], 4767 mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0], 4768 mode_lib->vba.NoOfDPP[i][j][k], 4769 mode_lib->vba.NumberOfCursors[k], 4770 mode_lib->vba.VTotal[k] 4771 - mode_lib->vba.VActive[k], 4772 mode_lib->vba.HTotal[k], 4773 mode_lib->vba.MaxInterDCNTileRepeaters, 4774 mode_lib->vba.MaximumVStartup[0][0][k], 4775 mode_lib->vba.GPUVMMaxPageTableLevels, 4776 mode_lib->vba.GPUVMEnable, 4777 mode_lib->vba.DynamicMetadataEnable[k], 4778 mode_lib->vba.DynamicMetadataLinesBeforeActiveRequired[k], 4779 mode_lib->vba.DynamicMetadataTransmittedBytes[k], 4780 mode_lib->vba.DCCEnable[k], 4781 mode_lib->vba.UrgentLatencyPixelDataOnly, 4782 mode_lib->vba.ExtraLatency, 4783 mode_lib->vba.TimeCalc, 4784 mode_lib->vba.PDEAndMetaPTEBytesPerFrame[0][0][k], 4785 mode_lib->vba.MetaRowBytes[0][0][k], 4786 mode_lib->vba.DPTEBytesPerRow[0][0][k], 4787 mode_lib->vba.PrefetchLinesY[0][0][k], 4788 mode_lib->vba.SwathWidthYPerState[i][j][k], 4789 mode_lib->vba.BytePerPixelInDETY[k], 4790 mode_lib->vba.PrefillY[k], 4791 mode_lib->vba.MaxNumSwY[k], 4792 mode_lib->vba.PrefetchLinesC[0][0][k], 4793 mode_lib->vba.BytePerPixelInDETC[k], 4794 mode_lib->vba.PrefillC[k], 4795 mode_lib->vba.MaxNumSwC[k], 4796 mode_lib->vba.SwathHeightYPerState[i][j][k], 4797 mode_lib->vba.SwathHeightCPerState[i][j][k], 4798 mode_lib->vba.TWait, 4799 mode_lib->vba.XFCEnabled[k], 4800 mode_lib->vba.XFCRemoteSurfaceFlipDelay, 4801 mode_lib->vba.Interlace[k], 4802 mode_lib->vba.ProgressiveToInterlaceUnitInOPP, 4803 mode_lib->vba.DSTXAfterScaler[k], 4804 mode_lib->vba.DSTYAfterScaler[k], 4805 &mode_lib->vba.LineTimesForPrefetch[k], 4806 &mode_lib->vba.PrefetchBW[k], 4807 &mode_lib->vba.LinesForMetaPTE[k], 4808 &mode_lib->vba.LinesForMetaAndDPTERow[k], 4809 &mode_lib->vba.VRatioPreY[i][j][k], 4810 &mode_lib->vba.VRatioPreC[i][j][k], 4811 &mode_lib->vba.RequiredPrefetchPixelDataBWLuma[i][j][k], 4812 &mode_lib->vba.Tno_bw[k], 4813 &mode_lib->vba.VUpdateOffsetPix[k], 4814 &mode_lib->vba.VUpdateWidthPix[k], 4815 &mode_lib->vba.VReadyOffsetPix[k]); 4816 } 4817 mode_lib->vba.MaximumReadBandwidthWithoutPrefetch = 0.0; 4818 mode_lib->vba.MaximumReadBandwidthWithPrefetch = 0.0; 4819 locals->prefetch_vm_bw_valid = true; 4820 locals->prefetch_row_bw_valid = true; 4821 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4822 if (locals->PDEAndMetaPTEBytesPerFrame[0][0][k] == 0) 4823 locals->prefetch_vm_bw[k] = 0; 4824 else if (locals->LinesForMetaPTE[k] > 0) 4825 locals->prefetch_vm_bw[k] = locals->PDEAndMetaPTEBytesPerFrame[0][0][k] 4826 / (locals->LinesForMetaPTE[k] * locals->HTotal[k] / locals->PixelClock[k]); 4827 else { 4828 locals->prefetch_vm_bw[k] = 0; 4829 locals->prefetch_vm_bw_valid = false; 4830 } 4831 if (locals->MetaRowBytes[0][0][k] + locals->DPTEBytesPerRow[0][0][k] == 0) 4832 locals->prefetch_row_bw[k] = 0; 4833 else if (locals->LinesForMetaAndDPTERow[k] > 0) 4834 locals->prefetch_row_bw[k] = (locals->MetaRowBytes[0][0][k] + locals->DPTEBytesPerRow[0][0][k]) 4835 / (locals->LinesForMetaAndDPTERow[k] * locals->HTotal[k] / locals->PixelClock[k]); 4836 else { 4837 locals->prefetch_row_bw[k] = 0; 4838 locals->prefetch_row_bw_valid = false; 4839 } 4840 4841 mode_lib->vba.MaximumReadBandwidthWithoutPrefetch = mode_lib->vba.MaximumReadBandwidthWithPrefetch 4842 + 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]; 4843 mode_lib->vba.MaximumReadBandwidthWithPrefetch = 4844 mode_lib->vba.MaximumReadBandwidthWithPrefetch 4845 + mode_lib->vba.cursor_bw[k] 4846 + dml_max3( 4847 mode_lib->vba.prefetch_vm_bw[k], 4848 mode_lib->vba.prefetch_row_bw[k], 4849 dml_max(mode_lib->vba.ReadBandwidth[k], 4850 mode_lib->vba.RequiredPrefetchPixelDataBWLuma[i][j][k]) 4851 + mode_lib->vba.meta_row_bw[k] + mode_lib->vba.dpte_row_bw[k]); 4852 } 4853 locals->BandwidthWithoutPrefetchSupported[i][0] = true; 4854 if (mode_lib->vba.MaximumReadBandwidthWithoutPrefetch > locals->ReturnBWPerState[i][0]) { 4855 locals->BandwidthWithoutPrefetchSupported[i][0] = false; 4856 } 4857 4858 locals->PrefetchSupported[i][j] = true; 4859 if (mode_lib->vba.MaximumReadBandwidthWithPrefetch > locals->ReturnBWPerState[i][0]) { 4860 locals->PrefetchSupported[i][j] = false; 4861 } 4862 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4863 if (locals->LineTimesForPrefetch[k] < 2.0 4864 || locals->LinesForMetaPTE[k] >= 8.0 4865 || locals->LinesForMetaAndDPTERow[k] >= 16.0 4866 || mode_lib->vba.IsErrorResult[i][j][k] == true) { 4867 locals->PrefetchSupported[i][j] = false; 4868 } 4869 } 4870 locals->VRatioInPrefetchSupported[i][j] = true; 4871 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4872 if (locals->VRatioPreY[i][j][k] > 4.0 4873 || locals->VRatioPreC[i][j][k] > 4.0 4874 || mode_lib->vba.IsErrorResult[i][j][k] == true) { 4875 locals->VRatioInPrefetchSupported[i][j] = false; 4876 } 4877 } 4878 } while ((locals->PrefetchSupported[i][j] != true || locals->VRatioInPrefetchSupported[i][j] != true) 4879 && mode_lib->vba.NextPrefetchMode < mode_lib->vba.MaxPrefetchMode); 4880 4881 if (mode_lib->vba.PrefetchSupported[i][j] == true 4882 && mode_lib->vba.VRatioInPrefetchSupported[i][j] == true) { 4883 mode_lib->vba.BandwidthAvailableForImmediateFlip = 4884 mode_lib->vba.ReturnBWPerState[i][0]; 4885 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4886 mode_lib->vba.BandwidthAvailableForImmediateFlip = 4887 mode_lib->vba.BandwidthAvailableForImmediateFlip 4888 - mode_lib->vba.cursor_bw[k] 4889 - dml_max( 4890 mode_lib->vba.ReadBandwidth[k] + mode_lib->vba.qual_row_bw[k], 4891 mode_lib->vba.PrefetchBW[k]); 4892 } 4893 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4894 mode_lib->vba.ImmediateFlipBytes[k] = 0.0; 4895 if ((mode_lib->vba.SourcePixelFormat[k] != dm_420_8 4896 && mode_lib->vba.SourcePixelFormat[k] != dm_420_10)) { 4897 mode_lib->vba.ImmediateFlipBytes[k] = 4898 mode_lib->vba.PDEAndMetaPTEBytesPerFrame[0][0][k] 4899 + mode_lib->vba.MetaRowBytes[0][0][k] 4900 + mode_lib->vba.DPTEBytesPerRow[0][0][k]; 4901 } 4902 } 4903 mode_lib->vba.TotImmediateFlipBytes = 0.0; 4904 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4905 if ((mode_lib->vba.SourcePixelFormat[k] != dm_420_8 4906 && mode_lib->vba.SourcePixelFormat[k] != dm_420_10)) { 4907 mode_lib->vba.TotImmediateFlipBytes = 4908 mode_lib->vba.TotImmediateFlipBytes 4909 + mode_lib->vba.ImmediateFlipBytes[k]; 4910 } 4911 } 4912 4913 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4914 CalculateFlipSchedule( 4915 mode_lib, 4916 mode_lib->vba.ExtraLatency, 4917 mode_lib->vba.UrgentLatencyPixelDataOnly, 4918 mode_lib->vba.GPUVMMaxPageTableLevels, 4919 mode_lib->vba.GPUVMEnable, 4920 mode_lib->vba.BandwidthAvailableForImmediateFlip, 4921 mode_lib->vba.TotImmediateFlipBytes, 4922 mode_lib->vba.SourcePixelFormat[k], 4923 mode_lib->vba.ImmediateFlipBytes[k], 4924 mode_lib->vba.HTotal[k] 4925 / mode_lib->vba.PixelClock[k], 4926 mode_lib->vba.VRatio[k], 4927 mode_lib->vba.Tno_bw[k], 4928 mode_lib->vba.PDEAndMetaPTEBytesPerFrame[0][0][k], 4929 mode_lib->vba.MetaRowBytes[0][0][k], 4930 mode_lib->vba.DPTEBytesPerRow[0][0][k], 4931 mode_lib->vba.DCCEnable[k], 4932 mode_lib->vba.dpte_row_height[k], 4933 mode_lib->vba.meta_row_height[k], 4934 mode_lib->vba.qual_row_bw[k], 4935 &mode_lib->vba.DestinationLinesToRequestVMInImmediateFlip[k], 4936 &mode_lib->vba.DestinationLinesToRequestRowInImmediateFlip[k], 4937 &mode_lib->vba.final_flip_bw[k], 4938 &mode_lib->vba.ImmediateFlipSupportedForPipe[k]); 4939 } 4940 mode_lib->vba.total_dcn_read_bw_with_flip = 0.0; 4941 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4942 mode_lib->vba.total_dcn_read_bw_with_flip = 4943 mode_lib->vba.total_dcn_read_bw_with_flip 4944 + mode_lib->vba.cursor_bw[k] 4945 + dml_max3( 4946 mode_lib->vba.prefetch_vm_bw[k], 4947 mode_lib->vba.prefetch_row_bw[k], 4948 mode_lib->vba.final_flip_bw[k] 4949 + dml_max( 4950 mode_lib->vba.ReadBandwidth[k], 4951 mode_lib->vba.RequiredPrefetchPixelDataBWLuma[i][j][k])); 4952 } 4953 mode_lib->vba.ImmediateFlipSupportedForState[i][j] = true; 4954 if (mode_lib->vba.total_dcn_read_bw_with_flip 4955 > mode_lib->vba.ReturnBWPerState[i][0]) { 4956 mode_lib->vba.ImmediateFlipSupportedForState[i][j] = false; 4957 } 4958 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4959 if (mode_lib->vba.ImmediateFlipSupportedForPipe[k] == false) { 4960 mode_lib->vba.ImmediateFlipSupportedForState[i][j] = false; 4961 } 4962 } 4963 } else { 4964 mode_lib->vba.ImmediateFlipSupportedForState[i][j] = false; 4965 } 4966 } 4967 } 4968 4969 /*Vertical Active BW support*/ 4970 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 4971 mode_lib->vba.MaxTotalVerticalActiveAvailableBandwidth[i][0] = dml_min(mode_lib->vba.ReturnBusWidth * 4972 mode_lib->vba.DCFCLKPerState[i], mode_lib->vba.FabricAndDRAMBandwidthPerState[i] * 1000) * 4973 mode_lib->vba.MaxAveragePercentOfIdealDRAMBWDisplayCanUseInNormalSystemOperation / 100; 4974 if (mode_lib->vba.MaxTotalVActiveRDBandwidth <= mode_lib->vba.MaxTotalVerticalActiveAvailableBandwidth[i][0]) 4975 mode_lib->vba.TotalVerticalActiveBandwidthSupport[i][0] = true; 4976 else 4977 mode_lib->vba.TotalVerticalActiveBandwidthSupport[i][0] = false; 4978 } 4979 4980 /*PTE Buffer Size Check*/ 4981 4982 for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { 4983 for (j = 0; j < 2; j++) { 4984 locals->PTEBufferSizeNotExceeded[i][j] = true; 4985 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4986 if (locals->PTEBufferSizeNotExceededY[i][j][k] == false 4987 || locals->PTEBufferSizeNotExceededC[i][j][k] == false) { 4988 locals->PTEBufferSizeNotExceeded[i][j] = false; 4989 } 4990 } 4991 } 4992 } 4993 /*Cursor Support Check*/ 4994 mode_lib->vba.CursorSupport = true; 4995 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 4996 for (j = 0; j < 2; j++) { 4997 if (mode_lib->vba.CursorWidth[k][j] > 0.0) { 4998 if (dml_floor( 4999 dml_floor( 5000 mode_lib->vba.CursorBufferSize 5001 - mode_lib->vba.CursorChunkSize, 5002 mode_lib->vba.CursorChunkSize) * 1024.0 5003 / (mode_lib->vba.CursorWidth[k][j] 5004 * mode_lib->vba.CursorBPP[k][j] 5005 / 8.0), 5006 1.0) 5007 * (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]) 5008 / mode_lib->vba.VRatio[k] < mode_lib->vba.UrgentLatencyPixelDataOnly 5009 || (mode_lib->vba.CursorBPP[k][j] == 64.0 5010 && mode_lib->vba.Cursor64BppSupport == false)) { 5011 mode_lib->vba.CursorSupport = false; 5012 } 5013 } 5014 } 5015 } 5016 /*Valid Pitch Check*/ 5017 5018 mode_lib->vba.PitchSupport = true; 5019 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 5020 locals->AlignedYPitch[k] = dml_ceil( 5021 dml_max(mode_lib->vba.PitchY[k], mode_lib->vba.ViewportWidth[k]), 5022 locals->MacroTileWidthY[k]); 5023 if (locals->AlignedYPitch[k] > mode_lib->vba.PitchY[k]) { 5024 mode_lib->vba.PitchSupport = false; 5025 } 5026 if (mode_lib->vba.DCCEnable[k] == true) { 5027 locals->AlignedDCCMetaPitch[k] = dml_ceil( 5028 dml_max( 5029 mode_lib->vba.DCCMetaPitchY[k], 5030 mode_lib->vba.ViewportWidth[k]), 5031 64.0 * locals->Read256BlockWidthY[k]); 5032 } else { 5033 locals->AlignedDCCMetaPitch[k] = mode_lib->vba.DCCMetaPitchY[k]; 5034 } 5035 if (locals->AlignedDCCMetaPitch[k] > mode_lib->vba.DCCMetaPitchY[k]) { 5036 mode_lib->vba.PitchSupport = false; 5037 } 5038 if (mode_lib->vba.SourcePixelFormat[k] != dm_444_64 5039 && mode_lib->vba.SourcePixelFormat[k] != dm_444_32 5040 && mode_lib->vba.SourcePixelFormat[k] != dm_444_16 5041 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_16 5042 && mode_lib->vba.SourcePixelFormat[k] != dm_mono_8) { 5043 locals->AlignedCPitch[k] = dml_ceil( 5044 dml_max( 5045 mode_lib->vba.PitchC[k], 5046 mode_lib->vba.ViewportWidth[k] / 2.0), 5047 locals->MacroTileWidthC[k]); 5048 } else { 5049 locals->AlignedCPitch[k] = mode_lib->vba.PitchC[k]; 5050 } 5051 if (locals->AlignedCPitch[k] > mode_lib->vba.PitchC[k]) { 5052 mode_lib->vba.PitchSupport = false; 5053 } 5054 } 5055 /*Mode Support, Voltage State and SOC Configuration*/ 5056 5057 for (i = mode_lib->vba.soc.num_states; i >= 0; i--) { 5058 for (j = 0; j < 2; j++) { 5059 enum dm_validation_status status = DML_VALIDATION_OK; 5060 5061 if (mode_lib->vba.ScaleRatioAndTapsSupport != true) { 5062 status = DML_FAIL_SCALE_RATIO_TAP; 5063 } else if (mode_lib->vba.SourceFormatPixelAndScanSupport != true) { 5064 status = DML_FAIL_SOURCE_PIXEL_FORMAT; 5065 } else if (locals->ViewportSizeSupport[i][0] != true) { 5066 status = DML_FAIL_VIEWPORT_SIZE; 5067 } else if (locals->DIOSupport[i] != true) { 5068 status = DML_FAIL_DIO_SUPPORT; 5069 } else if (locals->NotEnoughDSCUnits[i] != false) { 5070 status = DML_FAIL_NOT_ENOUGH_DSC; 5071 } else if (locals->DSCCLKRequiredMoreThanSupported[i] != false) { 5072 status = DML_FAIL_DSC_CLK_REQUIRED; 5073 } else if (locals->UrgentLatencySupport[i][j] != true) { 5074 status = DML_FAIL_URGENT_LATENCY; 5075 } else if (locals->ROBSupport[i][0] != true) { 5076 status = DML_FAIL_REORDERING_BUFFER; 5077 } else if (locals->DISPCLK_DPPCLK_Support[i][j] != true) { 5078 status = DML_FAIL_DISPCLK_DPPCLK; 5079 } else if (locals->TotalAvailablePipesSupport[i][j] != true) { 5080 status = DML_FAIL_TOTAL_AVAILABLE_PIPES; 5081 } else if (mode_lib->vba.NumberOfOTGSupport != true) { 5082 status = DML_FAIL_NUM_OTG; 5083 } else if (mode_lib->vba.WritebackModeSupport != true) { 5084 status = DML_FAIL_WRITEBACK_MODE; 5085 } else if (mode_lib->vba.WritebackLatencySupport != true) { 5086 status = DML_FAIL_WRITEBACK_LATENCY; 5087 } else if (mode_lib->vba.WritebackScaleRatioAndTapsSupport != true) { 5088 status = DML_FAIL_WRITEBACK_SCALE_RATIO_TAP; 5089 } else if (mode_lib->vba.CursorSupport != true) { 5090 status = DML_FAIL_CURSOR_SUPPORT; 5091 } else if (mode_lib->vba.PitchSupport != true) { 5092 status = DML_FAIL_PITCH_SUPPORT; 5093 } else if (locals->PrefetchSupported[i][j] != true) { 5094 status = DML_FAIL_PREFETCH_SUPPORT; 5095 } else if (locals->TotalVerticalActiveBandwidthSupport[i][0] != true) { 5096 status = DML_FAIL_TOTAL_V_ACTIVE_BW; 5097 } else if (locals->VRatioInPrefetchSupported[i][j] != true) { 5098 status = DML_FAIL_V_RATIO_PREFETCH; 5099 } else if (locals->PTEBufferSizeNotExceeded[i][j] != true) { 5100 status = DML_FAIL_PTE_BUFFER_SIZE; 5101 } else if (mode_lib->vba.NonsupportedDSCInputBPC != false) { 5102 status = DML_FAIL_DSC_INPUT_BPC; 5103 } 5104 5105 if (status == DML_VALIDATION_OK) { 5106 locals->ModeSupport[i][j] = true; 5107 } else { 5108 locals->ModeSupport[i][j] = false; 5109 } 5110 locals->ValidationStatus[i] = status; 5111 } 5112 } 5113 { 5114 unsigned int MaximumMPCCombine = 0; 5115 mode_lib->vba.VoltageLevel = mode_lib->vba.soc.num_states + 1; 5116 for (i = mode_lib->vba.VoltageOverrideLevel; i <= mode_lib->vba.soc.num_states; i++) { 5117 if (locals->ModeSupport[i][0] == true || locals->ModeSupport[i][1] == true) { 5118 mode_lib->vba.VoltageLevel = i; 5119 if (locals->ModeSupport[i][1] == true && (locals->ModeSupport[i][0] == false 5120 || mode_lib->vba.WhenToDoMPCCombine == dm_mpc_always_when_possible)) { 5121 MaximumMPCCombine = 1; 5122 } else { 5123 MaximumMPCCombine = 0; 5124 } 5125 break; 5126 } 5127 } 5128 mode_lib->vba.ImmediateFlipSupport = 5129 locals->ImmediateFlipSupportedForState[mode_lib->vba.VoltageLevel][MaximumMPCCombine]; 5130 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 5131 mode_lib->vba.DPPPerPlane[k] = locals->NoOfDPP[mode_lib->vba.VoltageLevel][MaximumMPCCombine][k]; 5132 locals->DPPCLK[k] = locals->RequiredDPPCLK[mode_lib->vba.VoltageLevel][MaximumMPCCombine][k]; 5133 } 5134 mode_lib->vba.DISPCLK = locals->RequiredDISPCLK[mode_lib->vba.VoltageLevel][MaximumMPCCombine]; 5135 mode_lib->vba.maxMpcComb = MaximumMPCCombine; 5136 } 5137 mode_lib->vba.DCFCLK = mode_lib->vba.DCFCLKPerState[mode_lib->vba.VoltageLevel]; 5138 mode_lib->vba.DRAMSpeed = mode_lib->vba.DRAMSpeedPerState[mode_lib->vba.VoltageLevel]; 5139 mode_lib->vba.FabricClock = mode_lib->vba.FabricClockPerState[mode_lib->vba.VoltageLevel]; 5140 mode_lib->vba.SOCCLK = mode_lib->vba.SOCCLKPerState[mode_lib->vba.VoltageLevel]; 5141 mode_lib->vba.ReturnBW = locals->ReturnBWPerState[mode_lib->vba.VoltageLevel][0]; 5142 mode_lib->vba.FabricAndDRAMBandwidth = locals->FabricAndDRAMBandwidthPerState[mode_lib->vba.VoltageLevel]; 5143 for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { 5144 if (mode_lib->vba.BlendingAndTiming[k] == k) { 5145 mode_lib->vba.ODMCombineEnabled[k] = 5146 locals->ODMCombineEnablePerState[mode_lib->vba.VoltageLevel][k]; 5147 } else { 5148 mode_lib->vba.ODMCombineEnabled[k] = 0; 5149 } 5150 mode_lib->vba.DSCEnabled[k] = 5151 locals->RequiresDSC[mode_lib->vba.VoltageLevel][k]; 5152 mode_lib->vba.OutputBpp[k] = 5153 locals->OutputBppPerState[mode_lib->vba.VoltageLevel][k]; 5154 } 5155 } 5156