| File: | dev/pci/drm/amd/display/dc/dml/dcn21/display_mode_vba_21.c |
| Warning: | line 1053, column 3 Value stored to 'TimeForFetchingRowInVBlank' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | /* |
| 2 | * Copyright 2017 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 | |
| 27 | #include "../display_mode_lib.h" |
| 28 | #include "../dml_inline_defs.h" |
| 29 | #include "../display_mode_vba.h" |
| 30 | #include "display_mode_vba_21.h" |
| 31 | |
| 32 | |
| 33 | /* |
| 34 | * NOTE: |
| 35 | * This file is gcc-parsable HW gospel, coming straight from HW engineers. |
| 36 | * |
| 37 | * It doesn't adhere to Linux kernel style and sometimes will do things in odd |
| 38 | * ways. Unless there is something clearly wrong with it the code should |
| 39 | * remain as-is as it provides us with a guarantee from HW that it is correct. |
| 40 | */ |
| 41 | typedef struct { |
| 42 | double DPPCLK; |
| 43 | double DISPCLK; |
| 44 | double PixelClock; |
| 45 | double DCFCLKDeepSleep; |
| 46 | unsigned int DPPPerPlane; |
| 47 | bool_Bool ScalerEnabled; |
| 48 | enum scan_direction_class SourceScan; |
| 49 | unsigned int BlockWidth256BytesY; |
| 50 | unsigned int BlockHeight256BytesY; |
| 51 | unsigned int BlockWidth256BytesC; |
| 52 | unsigned int BlockHeight256BytesC; |
| 53 | unsigned int InterlaceEnable; |
| 54 | unsigned int NumberOfCursors; |
| 55 | unsigned int VBlank; |
| 56 | unsigned int HTotal; |
| 57 | } Pipe; |
| 58 | |
| 59 | typedef struct { |
| 60 | bool_Bool Enable; |
| 61 | unsigned int MaxPageTableLevels; |
| 62 | unsigned int CachedPageTableLevels; |
| 63 | } HostVM; |
| 64 | |
| 65 | #define BPP_INVALID0 0 |
| 66 | #define BPP_BLENDED_PIPE0xffffffff 0xffffffff |
| 67 | #define DCN21_MAX_DSC_IMAGE_WIDTH5184 5184 |
| 68 | #define DCN21_MAX_420_IMAGE_WIDTH4096 4096 |
| 69 | |
| 70 | static void DisplayPipeConfiguration(struct display_mode_lib *mode_lib); |
| 71 | static void DISPCLKDPPCLKDCFCLKDeepSleepPrefetchParametersWatermarksAndPerformanceCalculation( |
| 72 | struct display_mode_lib *mode_lib); |
| 73 | static unsigned int dscceComputeDelay( |
| 74 | unsigned int bpc, |
| 75 | double bpp, |
| 76 | unsigned int sliceWidth, |
| 77 | unsigned int numSlices, |
| 78 | enum output_format_class pixelFormat); |
| 79 | static unsigned int dscComputeDelay(enum output_format_class pixelFormat); |
| 80 | // Super monster function with some 45 argument |
| 81 | static bool_Bool CalculatePrefetchSchedule( |
| 82 | struct display_mode_lib *mode_lib, |
| 83 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
| 84 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
| 85 | Pipe *myPipe, |
| 86 | unsigned int DSCDelay, |
| 87 | double DPPCLKDelaySubtotal, |
| 88 | double DPPCLKDelaySCL, |
| 89 | double DPPCLKDelaySCLLBOnly, |
| 90 | double DPPCLKDelayCNVCFormater, |
| 91 | double DPPCLKDelayCNVCCursor, |
| 92 | double DISPCLKDelaySubtotal, |
| 93 | unsigned int ScalerRecoutWidth, |
| 94 | enum output_format_class OutputFormat, |
| 95 | unsigned int MaxInterDCNTileRepeaters, |
| 96 | unsigned int VStartup, |
| 97 | unsigned int MaxVStartup, |
| 98 | unsigned int GPUVMPageTableLevels, |
| 99 | bool_Bool GPUVMEnable, |
| 100 | HostVM *myHostVM, |
| 101 | bool_Bool DynamicMetadataEnable, |
| 102 | int DynamicMetadataLinesBeforeActiveRequired, |
| 103 | unsigned int DynamicMetadataTransmittedBytes, |
| 104 | bool_Bool DCCEnable, |
| 105 | double UrgentLatency, |
| 106 | double UrgentExtraLatency, |
| 107 | double TCalc, |
| 108 | unsigned int PDEAndMetaPTEBytesFrame, |
| 109 | unsigned int MetaRowByte, |
| 110 | unsigned int PixelPTEBytesPerRow, |
| 111 | double PrefetchSourceLinesY, |
| 112 | unsigned int SwathWidthY, |
| 113 | double BytePerPixelDETY, |
| 114 | double VInitPreFillY, |
| 115 | unsigned int MaxNumSwathY, |
| 116 | double PrefetchSourceLinesC, |
| 117 | double BytePerPixelDETC, |
| 118 | double VInitPreFillC, |
| 119 | unsigned int MaxNumSwathC, |
| 120 | unsigned int SwathHeightY, |
| 121 | unsigned int SwathHeightC, |
| 122 | double TWait, |
| 123 | bool_Bool XFCEnabled, |
| 124 | double XFCRemoteSurfaceFlipDelay, |
| 125 | bool_Bool ProgressiveToInterlaceUnitInOPP, |
| 126 | double *DSTXAfterScaler, |
| 127 | double *DSTYAfterScaler, |
| 128 | double *DestinationLinesForPrefetch, |
| 129 | double *PrefetchBandwidth, |
| 130 | double *DestinationLinesToRequestVMInVBlank, |
| 131 | double *DestinationLinesToRequestRowInVBlank, |
| 132 | double *VRatioPrefetchY, |
| 133 | double *VRatioPrefetchC, |
| 134 | double *RequiredPrefetchPixDataBWLuma, |
| 135 | double *RequiredPrefetchPixDataBWChroma, |
| 136 | unsigned int *VStartupRequiredWhenNotEnoughTimeForDynamicMetadata, |
| 137 | double *Tno_bw, |
| 138 | double *prefetch_vmrow_bw, |
| 139 | unsigned int *swath_width_luma_ub, |
| 140 | unsigned int *swath_width_chroma_ub, |
| 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 CalculateDCCConfiguration( |
| 147 | bool_Bool DCCEnabled, |
| 148 | bool_Bool DCCProgrammingAssumesScanDirectionUnknown, |
| 149 | unsigned int ViewportWidth, |
| 150 | unsigned int ViewportHeight, |
| 151 | unsigned int DETBufferSize, |
| 152 | unsigned int RequestHeight256Byte, |
| 153 | unsigned int SwathHeight, |
| 154 | enum dm_swizzle_mode TilingFormat, |
| 155 | unsigned int BytePerPixel, |
| 156 | enum scan_direction_class ScanOrientation, |
| 157 | unsigned int *MaxUncompressedBlock, |
| 158 | unsigned int *MaxCompressedBlock, |
| 159 | unsigned int *Independent64ByteBlock); |
| 160 | static double CalculatePrefetchSourceLines( |
| 161 | struct display_mode_lib *mode_lib, |
| 162 | double VRatio, |
| 163 | double vtaps, |
| 164 | bool_Bool Interlace, |
| 165 | bool_Bool ProgressiveToInterlaceUnitInOPP, |
| 166 | unsigned int SwathHeight, |
| 167 | unsigned int ViewportYStart, |
| 168 | double *VInitPreFill, |
| 169 | unsigned int *MaxNumSwath); |
| 170 | static unsigned int CalculateVMAndRowBytes( |
| 171 | struct display_mode_lib *mode_lib, |
| 172 | bool_Bool DCCEnable, |
| 173 | unsigned int BlockHeight256Bytes, |
| 174 | unsigned int BlockWidth256Bytes, |
| 175 | enum source_format_class SourcePixelFormat, |
| 176 | unsigned int SurfaceTiling, |
| 177 | unsigned int BytePerPixel, |
| 178 | enum scan_direction_class ScanDirection, |
| 179 | unsigned int ViewportWidth, |
| 180 | unsigned int ViewportHeight, |
| 181 | unsigned int SwathWidthY, |
| 182 | bool_Bool GPUVMEnable, |
| 183 | bool_Bool HostVMEnable, |
| 184 | unsigned int HostVMMaxPageTableLevels, |
| 185 | unsigned int HostVMCachedPageTableLevels, |
| 186 | unsigned int VMMPageSize, |
| 187 | unsigned int PTEBufferSizeInRequests, |
| 188 | unsigned int Pitch, |
| 189 | unsigned int DCCMetaPitch, |
| 190 | unsigned int *MacroTileWidth, |
| 191 | unsigned int *MetaRowByte, |
| 192 | unsigned int *PixelPTEBytesPerRow, |
| 193 | bool_Bool *PTEBufferSizeNotExceeded, |
| 194 | unsigned int *dpte_row_width_ub, |
| 195 | unsigned int *dpte_row_height, |
| 196 | unsigned int *MetaRequestWidth, |
| 197 | unsigned int *MetaRequestHeight, |
| 198 | unsigned int *meta_row_width, |
| 199 | unsigned int *meta_row_height, |
| 200 | unsigned int *vm_group_bytes, |
| 201 | unsigned int *dpte_group_bytes, |
| 202 | unsigned int *PixelPTEReqWidth, |
| 203 | unsigned int *PixelPTEReqHeight, |
| 204 | unsigned int *PTERequestSize, |
| 205 | unsigned int *DPDE0BytesFrame, |
| 206 | unsigned int *MetaPTEBytesFrame); |
| 207 | |
| 208 | static double CalculateTWait( |
| 209 | unsigned int PrefetchMode, |
| 210 | double DRAMClockChangeLatency, |
| 211 | double UrgentLatency, |
| 212 | double SREnterPlusExitTime); |
| 213 | static double CalculateRemoteSurfaceFlipDelay( |
| 214 | struct display_mode_lib *mode_lib, |
| 215 | double VRatio, |
| 216 | double SwathWidth, |
| 217 | double Bpp, |
| 218 | double LineTime, |
| 219 | double XFCTSlvVupdateOffset, |
| 220 | double XFCTSlvVupdateWidth, |
| 221 | double XFCTSlvVreadyOffset, |
| 222 | double XFCXBUFLatencyTolerance, |
| 223 | double XFCFillBWOverhead, |
| 224 | double XFCSlvChunkSize, |
| 225 | double XFCBusTransportTime, |
| 226 | double TCalc, |
| 227 | double TWait, |
| 228 | double *SrcActiveDrainRate, |
| 229 | double *TInitXFill, |
| 230 | double *TslvChk); |
| 231 | static void CalculateActiveRowBandwidth( |
| 232 | bool_Bool GPUVMEnable, |
| 233 | enum source_format_class SourcePixelFormat, |
| 234 | double VRatio, |
| 235 | bool_Bool DCCEnable, |
| 236 | double LineTime, |
| 237 | unsigned int MetaRowByteLuma, |
| 238 | unsigned int MetaRowByteChroma, |
| 239 | unsigned int meta_row_height_luma, |
| 240 | unsigned int meta_row_height_chroma, |
| 241 | unsigned int PixelPTEBytesPerRowLuma, |
| 242 | unsigned int PixelPTEBytesPerRowChroma, |
| 243 | unsigned int dpte_row_height_luma, |
| 244 | unsigned int dpte_row_height_chroma, |
| 245 | double *meta_row_bw, |
| 246 | double *dpte_row_bw); |
| 247 | static void CalculateFlipSchedule( |
| 248 | struct display_mode_lib *mode_lib, |
| 249 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
| 250 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
| 251 | double UrgentExtraLatency, |
| 252 | double UrgentLatency, |
| 253 | unsigned int GPUVMMaxPageTableLevels, |
| 254 | bool_Bool HostVMEnable, |
| 255 | unsigned int HostVMMaxPageTableLevels, |
| 256 | unsigned int HostVMCachedPageTableLevels, |
| 257 | bool_Bool GPUVMEnable, |
| 258 | double PDEAndMetaPTEBytesPerFrame, |
| 259 | double MetaRowBytes, |
| 260 | double DPTEBytesPerRow, |
| 261 | double BandwidthAvailableForImmediateFlip, |
| 262 | unsigned int TotImmediateFlipBytes, |
| 263 | enum source_format_class SourcePixelFormat, |
| 264 | double LineTime, |
| 265 | double VRatio, |
| 266 | double Tno_bw, |
| 267 | bool_Bool DCCEnable, |
| 268 | unsigned int dpte_row_height, |
| 269 | unsigned int meta_row_height, |
| 270 | unsigned int dpte_row_height_chroma, |
| 271 | unsigned int meta_row_height_chroma, |
| 272 | double *DestinationLinesToRequestVMInImmediateFlip, |
| 273 | double *DestinationLinesToRequestRowInImmediateFlip, |
| 274 | double *final_flip_bw, |
| 275 | bool_Bool *ImmediateFlipSupportedForPipe); |
| 276 | static double CalculateWriteBackDelay( |
| 277 | enum source_format_class WritebackPixelFormat, |
| 278 | double WritebackHRatio, |
| 279 | double WritebackVRatio, |
| 280 | unsigned int WritebackLumaHTaps, |
| 281 | unsigned int WritebackLumaVTaps, |
| 282 | unsigned int WritebackChromaHTaps, |
| 283 | unsigned int WritebackChromaVTaps, |
| 284 | unsigned int WritebackDestinationWidth); |
| 285 | static void CalculateWatermarksAndDRAMSpeedChangeSupport( |
| 286 | struct display_mode_lib *mode_lib, |
| 287 | unsigned int PrefetchMode, |
| 288 | unsigned int NumberOfActivePlanes, |
| 289 | unsigned int MaxLineBufferLines, |
| 290 | unsigned int LineBufferSize, |
| 291 | unsigned int DPPOutputBufferPixels, |
| 292 | unsigned int DETBufferSizeInKByte, |
| 293 | unsigned int WritebackInterfaceLumaBufferSize, |
| 294 | unsigned int WritebackInterfaceChromaBufferSize, |
| 295 | double DCFCLK, |
| 296 | double UrgentOutOfOrderReturn, |
| 297 | double ReturnBW, |
| 298 | bool_Bool GPUVMEnable, |
| 299 | int dpte_group_bytes[], |
| 300 | unsigned int MetaChunkSize, |
| 301 | double UrgentLatency, |
| 302 | double ExtraLatency, |
| 303 | double WritebackLatency, |
| 304 | double WritebackChunkSize, |
| 305 | double SOCCLK, |
| 306 | double DRAMClockChangeLatency, |
| 307 | double SRExitTime, |
| 308 | double SREnterPlusExitTime, |
| 309 | double DCFCLKDeepSleep, |
| 310 | int DPPPerPlane[], |
| 311 | bool_Bool DCCEnable[], |
| 312 | double DPPCLK[], |
| 313 | double SwathWidthSingleDPPY[], |
| 314 | unsigned int SwathHeightY[], |
| 315 | double ReadBandwidthPlaneLuma[], |
| 316 | unsigned int SwathHeightC[], |
| 317 | double ReadBandwidthPlaneChroma[], |
| 318 | unsigned int LBBitPerPixel[], |
| 319 | double SwathWidthY[], |
| 320 | double HRatio[], |
| 321 | unsigned int vtaps[], |
| 322 | unsigned int VTAPsChroma[], |
| 323 | double VRatio[], |
| 324 | unsigned int HTotal[], |
| 325 | double PixelClock[], |
| 326 | unsigned int BlendingAndTiming[], |
| 327 | double BytePerPixelDETY[], |
| 328 | double BytePerPixelDETC[], |
| 329 | bool_Bool WritebackEnable[], |
| 330 | enum source_format_class WritebackPixelFormat[], |
| 331 | double WritebackDestinationWidth[], |
| 332 | double WritebackDestinationHeight[], |
| 333 | double WritebackSourceHeight[], |
| 334 | enum clock_change_support *DRAMClockChangeSupport, |
| 335 | double *UrgentWatermark, |
| 336 | double *WritebackUrgentWatermark, |
| 337 | double *DRAMClockChangeWatermark, |
| 338 | double *WritebackDRAMClockChangeWatermark, |
| 339 | double *StutterExitWatermark, |
| 340 | double *StutterEnterPlusExitWatermark, |
| 341 | double *MinActiveDRAMClockChangeLatencySupported); |
| 342 | static void CalculateDCFCLKDeepSleep( |
| 343 | struct display_mode_lib *mode_lib, |
| 344 | unsigned int NumberOfActivePlanes, |
| 345 | double BytePerPixelDETY[], |
| 346 | double BytePerPixelDETC[], |
| 347 | double VRatio[], |
| 348 | double SwathWidthY[], |
| 349 | int DPPPerPlane[], |
| 350 | double HRatio[], |
| 351 | double PixelClock[], |
| 352 | double PSCL_THROUGHPUT[], |
| 353 | double PSCL_THROUGHPUT_CHROMA[], |
| 354 | double DPPCLK[], |
| 355 | double *DCFCLKDeepSleep); |
| 356 | static void CalculateDETBufferSize( |
| 357 | unsigned int DETBufferSizeInKByte, |
| 358 | unsigned int SwathHeightY, |
| 359 | unsigned int SwathHeightC, |
| 360 | unsigned int *DETBufferSizeY, |
| 361 | unsigned int *DETBufferSizeC); |
| 362 | static void CalculateUrgentBurstFactor( |
| 363 | unsigned int DETBufferSizeInKByte, |
| 364 | unsigned int SwathHeightY, |
| 365 | unsigned int SwathHeightC, |
| 366 | unsigned int SwathWidthY, |
| 367 | double LineTime, |
| 368 | double UrgentLatency, |
| 369 | double CursorBufferSize, |
| 370 | unsigned int CursorWidth, |
| 371 | unsigned int CursorBPP, |
| 372 | double VRatio, |
| 373 | double VRatioPreY, |
| 374 | double VRatioPreC, |
| 375 | double BytePerPixelInDETY, |
| 376 | double BytePerPixelInDETC, |
| 377 | double *UrgentBurstFactorCursor, |
| 378 | double *UrgentBurstFactorCursorPre, |
| 379 | double *UrgentBurstFactorLuma, |
| 380 | double *UrgentBurstFactorLumaPre, |
| 381 | double *UrgentBurstFactorChroma, |
| 382 | double *UrgentBurstFactorChromaPre, |
| 383 | unsigned int *NotEnoughUrgentLatencyHiding, |
| 384 | unsigned int *NotEnoughUrgentLatencyHidingPre); |
| 385 | |
| 386 | static void CalculatePixelDeliveryTimes( |
| 387 | unsigned int NumberOfActivePlanes, |
| 388 | double VRatio[], |
| 389 | double VRatioPrefetchY[], |
| 390 | double VRatioPrefetchC[], |
| 391 | unsigned int swath_width_luma_ub[], |
| 392 | unsigned int swath_width_chroma_ub[], |
| 393 | int DPPPerPlane[], |
| 394 | double HRatio[], |
| 395 | double PixelClock[], |
| 396 | double PSCL_THROUGHPUT[], |
| 397 | double PSCL_THROUGHPUT_CHROMA[], |
| 398 | double DPPCLK[], |
| 399 | double BytePerPixelDETC[], |
| 400 | enum scan_direction_class SourceScan[], |
| 401 | unsigned int BlockWidth256BytesY[], |
| 402 | unsigned int BlockHeight256BytesY[], |
| 403 | unsigned int BlockWidth256BytesC[], |
| 404 | unsigned int BlockHeight256BytesC[], |
| 405 | double DisplayPipeLineDeliveryTimeLuma[], |
| 406 | double DisplayPipeLineDeliveryTimeChroma[], |
| 407 | double DisplayPipeLineDeliveryTimeLumaPrefetch[], |
| 408 | double DisplayPipeLineDeliveryTimeChromaPrefetch[], |
| 409 | double DisplayPipeRequestDeliveryTimeLuma[], |
| 410 | double DisplayPipeRequestDeliveryTimeChroma[], |
| 411 | double DisplayPipeRequestDeliveryTimeLumaPrefetch[], |
| 412 | double DisplayPipeRequestDeliveryTimeChromaPrefetch[]); |
| 413 | |
| 414 | static void CalculateMetaAndPTETimes( |
| 415 | unsigned int NumberOfActivePlanes, |
| 416 | bool_Bool GPUVMEnable, |
| 417 | unsigned int MetaChunkSize, |
| 418 | unsigned int MinMetaChunkSizeBytes, |
| 419 | unsigned int GPUVMMaxPageTableLevels, |
| 420 | unsigned int HTotal[], |
| 421 | double VRatio[], |
| 422 | double VRatioPrefetchY[], |
| 423 | double VRatioPrefetchC[], |
| 424 | double DestinationLinesToRequestRowInVBlank[], |
| 425 | double DestinationLinesToRequestRowInImmediateFlip[], |
| 426 | double DestinationLinesToRequestVMInVBlank[], |
| 427 | double DestinationLinesToRequestVMInImmediateFlip[], |
| 428 | bool_Bool DCCEnable[], |
| 429 | double PixelClock[], |
| 430 | double BytePerPixelDETY[], |
| 431 | double BytePerPixelDETC[], |
| 432 | enum scan_direction_class SourceScan[], |
| 433 | unsigned int dpte_row_height[], |
| 434 | unsigned int dpte_row_height_chroma[], |
| 435 | unsigned int meta_row_width[], |
| 436 | unsigned int meta_row_height[], |
| 437 | unsigned int meta_req_width[], |
| 438 | unsigned int meta_req_height[], |
| 439 | int dpte_group_bytes[], |
| 440 | unsigned int PTERequestSizeY[], |
| 441 | unsigned int PTERequestSizeC[], |
| 442 | unsigned int PixelPTEReqWidthY[], |
| 443 | unsigned int PixelPTEReqHeightY[], |
| 444 | unsigned int PixelPTEReqWidthC[], |
| 445 | unsigned int PixelPTEReqHeightC[], |
| 446 | unsigned int dpte_row_width_luma_ub[], |
| 447 | unsigned int dpte_row_width_chroma_ub[], |
| 448 | unsigned int vm_group_bytes[], |
| 449 | unsigned int dpde0_bytes_per_frame_ub_l[], |
| 450 | unsigned int dpde0_bytes_per_frame_ub_c[], |
| 451 | unsigned int meta_pte_bytes_per_frame_ub_l[], |
| 452 | unsigned int meta_pte_bytes_per_frame_ub_c[], |
| 453 | double DST_Y_PER_PTE_ROW_NOM_L[], |
| 454 | double DST_Y_PER_PTE_ROW_NOM_C[], |
| 455 | double DST_Y_PER_META_ROW_NOM_L[], |
| 456 | double TimePerMetaChunkNominal[], |
| 457 | double TimePerMetaChunkVBlank[], |
| 458 | double TimePerMetaChunkFlip[], |
| 459 | double time_per_pte_group_nom_luma[], |
| 460 | double time_per_pte_group_vblank_luma[], |
| 461 | double time_per_pte_group_flip_luma[], |
| 462 | double time_per_pte_group_nom_chroma[], |
| 463 | double time_per_pte_group_vblank_chroma[], |
| 464 | double time_per_pte_group_flip_chroma[], |
| 465 | double TimePerVMGroupVBlank[], |
| 466 | double TimePerVMGroupFlip[], |
| 467 | double TimePerVMRequestVBlank[], |
| 468 | double TimePerVMRequestFlip[]); |
| 469 | |
| 470 | static double CalculateExtraLatency( |
| 471 | double UrgentRoundTripAndOutOfOrderLatency, |
| 472 | int TotalNumberOfActiveDPP, |
| 473 | int PixelChunkSizeInKByte, |
| 474 | int TotalNumberOfDCCActiveDPP, |
| 475 | int MetaChunkSize, |
| 476 | double ReturnBW, |
| 477 | bool_Bool GPUVMEnable, |
| 478 | bool_Bool HostVMEnable, |
| 479 | int NumberOfActivePlanes, |
| 480 | int NumberOfDPP[], |
| 481 | int dpte_group_bytes[], |
| 482 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
| 483 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
| 484 | int HostVMMaxPageTableLevels, |
| 485 | int HostVMCachedPageTableLevels); |
| 486 | |
| 487 | void dml21_recalculate(struct display_mode_lib *mode_lib) |
| 488 | { |
| 489 | ModeSupportAndSystemConfiguration(mode_lib); |
| 490 | PixelClockAdjustmentForProgressiveToInterlaceUnit(mode_lib); |
| 491 | DisplayPipeConfiguration(mode_lib); |
| 492 | DISPCLKDPPCLKDCFCLKDeepSleepPrefetchParametersWatermarksAndPerformanceCalculation(mode_lib); |
| 493 | } |
| 494 | |
| 495 | static unsigned int dscceComputeDelay( |
| 496 | unsigned int bpc, |
| 497 | double bpp, |
| 498 | unsigned int sliceWidth, |
| 499 | unsigned int numSlices, |
| 500 | enum output_format_class pixelFormat) |
| 501 | { |
| 502 | // valid bpc = source bits per component in the set of {8, 10, 12} |
| 503 | // valid bpp = increments of 1/16 of a bit |
| 504 | // min = 6/7/8 in N420/N422/444, respectively |
| 505 | // max = such that compression is 1:1 |
| 506 | //valid sliceWidth = number of pixels per slice line, must be less than or equal to 5184/numSlices (or 4096/numSlices in 420 mode) |
| 507 | //valid numSlices = number of slices in the horiziontal direction per DSC engine in the set of {1, 2, 3, 4} |
| 508 | //valid pixelFormat = pixel/color format in the set of {:N444_RGB, :S422, :N422, :N420} |
| 509 | |
| 510 | // fixed value |
| 511 | unsigned int rcModelSize = 8192; |
| 512 | |
| 513 | // N422/N420 operate at 2 pixels per clock |
| 514 | unsigned int pixelsPerClock, lstall, D, initalXmitDelay, w, S, ix, wx, p, l0, a, ax, l, |
| 515 | Delay, pixels; |
| 516 | |
| 517 | if (pixelFormat == dm_n422 || pixelFormat == dm_420) |
| 518 | pixelsPerClock = 2; |
| 519 | // #all other modes operate at 1 pixel per clock |
| 520 | else |
| 521 | pixelsPerClock = 1; |
| 522 | |
| 523 | //initial transmit delay as per PPS |
| 524 | initalXmitDelay = dml_round(rcModelSize / 2.0 / bpp / pixelsPerClock); |
| 525 | |
| 526 | //compute ssm delay |
| 527 | if (bpc == 8) |
| 528 | D = 81; |
| 529 | else if (bpc == 10) |
| 530 | D = 89; |
| 531 | else |
| 532 | D = 113; |
| 533 | |
| 534 | //divide by pixel per cycle to compute slice width as seen by DSC |
| 535 | w = sliceWidth / pixelsPerClock; |
| 536 | |
| 537 | //422 mode has an additional cycle of delay |
| 538 | if (pixelFormat == dm_s422) |
| 539 | S = 1; |
| 540 | else |
| 541 | S = 0; |
| 542 | |
| 543 | //main calculation for the dscce |
| 544 | ix = initalXmitDelay + 45; |
| 545 | wx = (w + 2) / 3; |
| 546 | p = 3 * wx - w; |
| 547 | l0 = ix / w; |
| 548 | a = ix + p * l0; |
| 549 | ax = (a + 2) / 3 + D + 6 + 1; |
| 550 | l = (ax + wx - 1) / wx; |
| 551 | if ((ix % w) == 0 && p != 0) |
| 552 | lstall = 1; |
| 553 | else |
| 554 | lstall = 0; |
| 555 | Delay = l * wx * (numSlices - 1) + ax + S + lstall + 22; |
| 556 | |
| 557 | //dsc processes 3 pixel containers per cycle and a container can contain 1 or 2 pixels |
| 558 | pixels = Delay * 3 * pixelsPerClock; |
| 559 | return pixels; |
| 560 | } |
| 561 | |
| 562 | static unsigned int dscComputeDelay(enum output_format_class pixelFormat) |
| 563 | { |
| 564 | unsigned int Delay = 0; |
| 565 | |
| 566 | if (pixelFormat == dm_420) { |
| 567 | // sfr |
| 568 | Delay = Delay + 2; |
| 569 | // dsccif |
| 570 | Delay = Delay + 0; |
| 571 | // dscc - input deserializer |
| 572 | Delay = Delay + 3; |
| 573 | // dscc gets pixels every other cycle |
| 574 | Delay = Delay + 2; |
| 575 | // dscc - input cdc fifo |
| 576 | Delay = Delay + 12; |
| 577 | // dscc gets pixels every other cycle |
| 578 | Delay = Delay + 13; |
| 579 | // dscc - cdc uncertainty |
| 580 | Delay = Delay + 2; |
| 581 | // dscc - output cdc fifo |
| 582 | Delay = Delay + 7; |
| 583 | // dscc gets pixels every other cycle |
| 584 | Delay = Delay + 3; |
| 585 | // dscc - cdc uncertainty |
| 586 | Delay = Delay + 2; |
| 587 | // dscc - output serializer |
| 588 | Delay = Delay + 1; |
| 589 | // sft |
| 590 | Delay = Delay + 1; |
| 591 | } else if (pixelFormat == dm_n422) { |
| 592 | // sfr |
| 593 | Delay = Delay + 2; |
| 594 | // dsccif |
| 595 | Delay = Delay + 1; |
| 596 | // dscc - input deserializer |
| 597 | Delay = Delay + 5; |
| 598 | // dscc - input cdc fifo |
| 599 | Delay = Delay + 25; |
| 600 | // dscc - cdc uncertainty |
| 601 | Delay = Delay + 2; |
| 602 | // dscc - output cdc fifo |
| 603 | Delay = Delay + 10; |
| 604 | // dscc - cdc uncertainty |
| 605 | Delay = Delay + 2; |
| 606 | // dscc - output serializer |
| 607 | Delay = Delay + 1; |
| 608 | // sft |
| 609 | Delay = Delay + 1; |
| 610 | } else { |
| 611 | // sfr |
| 612 | Delay = Delay + 2; |
| 613 | // dsccif |
| 614 | Delay = Delay + 0; |
| 615 | // dscc - input deserializer |
| 616 | Delay = Delay + 3; |
| 617 | // dscc - input cdc fifo |
| 618 | Delay = Delay + 12; |
| 619 | // dscc - cdc uncertainty |
| 620 | Delay = Delay + 2; |
| 621 | // dscc - output cdc fifo |
| 622 | Delay = Delay + 7; |
| 623 | // dscc - output serializer |
| 624 | Delay = Delay + 1; |
| 625 | // dscc - cdc uncertainty |
| 626 | Delay = Delay + 2; |
| 627 | // sft |
| 628 | Delay = Delay + 1; |
| 629 | } |
| 630 | |
| 631 | return Delay; |
| 632 | } |
| 633 | |
| 634 | static bool_Bool CalculatePrefetchSchedule( |
| 635 | struct display_mode_lib *mode_lib, |
| 636 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
| 637 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
| 638 | Pipe *myPipe, |
| 639 | unsigned int DSCDelay, |
| 640 | double DPPCLKDelaySubtotal, |
| 641 | double DPPCLKDelaySCL, |
| 642 | double DPPCLKDelaySCLLBOnly, |
| 643 | double DPPCLKDelayCNVCFormater, |
| 644 | double DPPCLKDelayCNVCCursor, |
| 645 | double DISPCLKDelaySubtotal, |
| 646 | unsigned int ScalerRecoutWidth, |
| 647 | enum output_format_class OutputFormat, |
| 648 | unsigned int MaxInterDCNTileRepeaters, |
| 649 | unsigned int VStartup, |
| 650 | unsigned int MaxVStartup, |
| 651 | unsigned int GPUVMPageTableLevels, |
| 652 | bool_Bool GPUVMEnable, |
| 653 | HostVM *myHostVM, |
| 654 | bool_Bool DynamicMetadataEnable, |
| 655 | int DynamicMetadataLinesBeforeActiveRequired, |
| 656 | unsigned int DynamicMetadataTransmittedBytes, |
| 657 | bool_Bool DCCEnable, |
| 658 | double UrgentLatency, |
| 659 | double UrgentExtraLatency, |
| 660 | double TCalc, |
| 661 | unsigned int PDEAndMetaPTEBytesFrame, |
| 662 | unsigned int MetaRowByte, |
| 663 | unsigned int PixelPTEBytesPerRow, |
| 664 | double PrefetchSourceLinesY, |
| 665 | unsigned int SwathWidthY, |
| 666 | double BytePerPixelDETY, |
| 667 | double VInitPreFillY, |
| 668 | unsigned int MaxNumSwathY, |
| 669 | double PrefetchSourceLinesC, |
| 670 | double BytePerPixelDETC, |
| 671 | double VInitPreFillC, |
| 672 | unsigned int MaxNumSwathC, |
| 673 | unsigned int SwathHeightY, |
| 674 | unsigned int SwathHeightC, |
| 675 | double TWait, |
| 676 | bool_Bool XFCEnabled, |
| 677 | double XFCRemoteSurfaceFlipDelay, |
| 678 | bool_Bool ProgressiveToInterlaceUnitInOPP, |
| 679 | double *DSTXAfterScaler, |
| 680 | double *DSTYAfterScaler, |
| 681 | double *DestinationLinesForPrefetch, |
| 682 | double *PrefetchBandwidth, |
| 683 | double *DestinationLinesToRequestVMInVBlank, |
| 684 | double *DestinationLinesToRequestRowInVBlank, |
| 685 | double *VRatioPrefetchY, |
| 686 | double *VRatioPrefetchC, |
| 687 | double *RequiredPrefetchPixDataBWLuma, |
| 688 | double *RequiredPrefetchPixDataBWChroma, |
| 689 | unsigned int *VStartupRequiredWhenNotEnoughTimeForDynamicMetadata, |
| 690 | double *Tno_bw, |
| 691 | double *prefetch_vmrow_bw, |
| 692 | unsigned int *swath_width_luma_ub, |
| 693 | unsigned int *swath_width_chroma_ub, |
| 694 | unsigned int *VUpdateOffsetPix, |
| 695 | double *VUpdateWidthPix, |
| 696 | double *VReadyOffsetPix) |
| 697 | { |
| 698 | bool_Bool MyError = false0; |
| 699 | unsigned int DPPCycles, DISPCLKCycles; |
| 700 | double DSTTotalPixelsAfterScaler, TotalRepeaterDelayTime; |
| 701 | double Tdm, LineTime, Tsetup; |
| 702 | double dst_y_prefetch_equ; |
| 703 | double Tsw_oto; |
| 704 | double prefetch_bw_oto; |
| 705 | double Tvm_oto; |
| 706 | double Tr0_oto; |
| 707 | double Tvm_oto_lines; |
| 708 | double Tr0_oto_lines; |
| 709 | double Tsw_oto_lines; |
| 710 | double dst_y_prefetch_oto; |
| 711 | double TimeForFetchingMetaPTE = 0; |
| 712 | double TimeForFetchingRowInVBlank = 0; |
| 713 | double LinesToRequestPrefetchPixelData = 0; |
| 714 | double HostVMInefficiencyFactor; |
| 715 | unsigned int HostVMDynamicLevels; |
| 716 | |
| 717 | if (GPUVMEnable == true1 && myHostVM->Enable == true1) { |
| 718 | HostVMInefficiencyFactor = |
| 719 | PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData |
| 720 | / PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly; |
| 721 | HostVMDynamicLevels = myHostVM->MaxPageTableLevels |
| 722 | - myHostVM->CachedPageTableLevels; |
| 723 | } else { |
| 724 | HostVMInefficiencyFactor = 1; |
| 725 | HostVMDynamicLevels = 0; |
| 726 | } |
| 727 | |
| 728 | if (myPipe->ScalerEnabled) |
| 729 | DPPCycles = DPPCLKDelaySubtotal + DPPCLKDelaySCL; |
| 730 | else |
| 731 | DPPCycles = DPPCLKDelaySubtotal + DPPCLKDelaySCLLBOnly; |
| 732 | |
| 733 | DPPCycles = DPPCycles + DPPCLKDelayCNVCFormater + myPipe->NumberOfCursors * DPPCLKDelayCNVCCursor; |
| 734 | |
| 735 | DISPCLKCycles = DISPCLKDelaySubtotal; |
| 736 | |
| 737 | if (myPipe->DPPCLK == 0.0 || myPipe->DISPCLK == 0.0) |
| 738 | return true1; |
| 739 | |
| 740 | *DSTXAfterScaler = DPPCycles * myPipe->PixelClock / myPipe->DPPCLK |
| 741 | + DISPCLKCycles * myPipe->PixelClock / myPipe->DISPCLK + DSCDelay; |
| 742 | |
| 743 | if (myPipe->DPPPerPlane > 1) |
| 744 | *DSTXAfterScaler = *DSTXAfterScaler + ScalerRecoutWidth; |
| 745 | |
| 746 | if (OutputFormat == dm_420 || (myPipe->InterlaceEnable && ProgressiveToInterlaceUnitInOPP)) |
| 747 | *DSTYAfterScaler = 1; |
| 748 | else |
| 749 | *DSTYAfterScaler = 0; |
| 750 | |
| 751 | DSTTotalPixelsAfterScaler = ((double) (*DSTYAfterScaler * myPipe->HTotal)) + *DSTXAfterScaler; |
| 752 | *DSTYAfterScaler = dml_floor(DSTTotalPixelsAfterScaler / myPipe->HTotal, 1); |
| 753 | *DSTXAfterScaler = DSTTotalPixelsAfterScaler - ((double) (*DSTYAfterScaler * myPipe->HTotal)); |
| 754 | |
| 755 | *VUpdateOffsetPix = dml_ceil(myPipe->HTotal / 4.0, 1); |
| 756 | TotalRepeaterDelayTime = MaxInterDCNTileRepeaters * (2.0 / myPipe->DPPCLK + 3.0 / myPipe->DISPCLK); |
| 757 | *VUpdateWidthPix = (14.0 / myPipe->DCFCLKDeepSleep + 12.0 / myPipe->DPPCLK + TotalRepeaterDelayTime) |
| 758 | * myPipe->PixelClock; |
| 759 | |
| 760 | *VReadyOffsetPix = dml_max( |
| 761 | 150.0 / myPipe->DPPCLK, |
| 762 | TotalRepeaterDelayTime + 20.0 / myPipe->DCFCLKDeepSleep + 10.0 / myPipe->DPPCLK) |
| 763 | * myPipe->PixelClock; |
| 764 | |
| 765 | Tsetup = (double) (*VUpdateOffsetPix + *VUpdateWidthPix + *VReadyOffsetPix) / myPipe->PixelClock; |
| 766 | |
| 767 | LineTime = (double) myPipe->HTotal / myPipe->PixelClock; |
| 768 | |
| 769 | if (DynamicMetadataEnable) { |
| 770 | double Tdmbf, Tdmec, Tdmsks; |
| 771 | |
| 772 | Tdm = dml_max(0.0, UrgentExtraLatency - TCalc); |
| 773 | Tdmbf = DynamicMetadataTransmittedBytes / 4.0 / myPipe->DISPCLK; |
| 774 | Tdmec = LineTime; |
| 775 | if (DynamicMetadataLinesBeforeActiveRequired == -1) |
| 776 | Tdmsks = myPipe->VBlank * LineTime / 2.0; |
| 777 | else |
| 778 | Tdmsks = DynamicMetadataLinesBeforeActiveRequired * LineTime; |
| 779 | if (myPipe->InterlaceEnable && !ProgressiveToInterlaceUnitInOPP) |
| 780 | Tdmsks = Tdmsks / 2; |
| 781 | if (VStartup * LineTime |
| 782 | < Tsetup + TWait + UrgentExtraLatency + Tdmbf + Tdmec + Tdmsks) { |
| 783 | MyError = true1; |
| 784 | *VStartupRequiredWhenNotEnoughTimeForDynamicMetadata = (Tsetup + TWait |
| 785 | + UrgentExtraLatency + Tdmbf + Tdmec + Tdmsks) / LineTime; |
| 786 | } else |
| 787 | *VStartupRequiredWhenNotEnoughTimeForDynamicMetadata = 0.0; |
| 788 | } else |
| 789 | Tdm = 0; |
| 790 | |
| 791 | if (GPUVMEnable) { |
| 792 | if (GPUVMPageTableLevels >= 3) |
| 793 | *Tno_bw = UrgentExtraLatency + UrgentLatency * ((GPUVMPageTableLevels - 2) * (myHostVM->MaxPageTableLevels + 1) - 1); |
| 794 | else |
| 795 | *Tno_bw = 0; |
| 796 | } else if (!DCCEnable) |
| 797 | *Tno_bw = LineTime; |
| 798 | else |
| 799 | *Tno_bw = LineTime / 4; |
| 800 | |
| 801 | dst_y_prefetch_equ = VStartup - dml_max(TCalc + TWait, XFCRemoteSurfaceFlipDelay) / LineTime |
| 802 | - (Tsetup + Tdm) / LineTime |
| 803 | - (*DSTYAfterScaler + *DSTXAfterScaler / myPipe->HTotal); |
| 804 | |
| 805 | Tsw_oto = dml_max(PrefetchSourceLinesY, PrefetchSourceLinesC) * LineTime; |
| 806 | |
| 807 | if (myPipe->SourceScan == dm_horz) { |
| 808 | *swath_width_luma_ub = dml_ceil(SwathWidthY - 1, myPipe->BlockWidth256BytesY) + myPipe->BlockWidth256BytesY; |
| 809 | if (myPipe->BlockWidth256BytesC > 0) |
| 810 | *swath_width_chroma_ub = dml_ceil(SwathWidthY / 2 - 1, myPipe->BlockWidth256BytesC) + myPipe->BlockWidth256BytesC; |
| 811 | } else { |
| 812 | *swath_width_luma_ub = dml_ceil(SwathWidthY - 1, myPipe->BlockHeight256BytesY) + myPipe->BlockHeight256BytesY; |
| 813 | if (myPipe->BlockHeight256BytesC > 0) |
| 814 | *swath_width_chroma_ub = dml_ceil(SwathWidthY / 2 - 1, myPipe->BlockHeight256BytesC) + myPipe->BlockHeight256BytesC; |
| 815 | } |
| 816 | |
| 817 | prefetch_bw_oto = (PrefetchSourceLinesY * *swath_width_luma_ub * dml_ceil(BytePerPixelDETY, 1) + PrefetchSourceLinesC * *swath_width_chroma_ub * dml_ceil(BytePerPixelDETC, 2)) / Tsw_oto; |
| 818 | |
| 819 | |
| 820 | if (GPUVMEnable == true1) { |
| 821 | Tvm_oto = dml_max(*Tno_bw + PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor / prefetch_bw_oto, |
| 822 | dml_max(UrgentExtraLatency + UrgentLatency * (GPUVMPageTableLevels * (HostVMDynamicLevels + 1) - 1), |
| 823 | LineTime / 4.0)); |
| 824 | } else |
| 825 | Tvm_oto = LineTime / 4.0; |
| 826 | |
| 827 | if ((GPUVMEnable == true1 || DCCEnable == true1)) { |
| 828 | Tr0_oto = dml_max( |
| 829 | (MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor) / prefetch_bw_oto, |
| 830 | dml_max(UrgentLatency * (HostVMDynamicLevels + 1), dml_max(LineTime - Tvm_oto, LineTime / 4))); |
| 831 | } else |
| 832 | Tr0_oto = (LineTime - Tvm_oto) / 2.0; |
| 833 | |
| 834 | Tvm_oto_lines = dml_ceil(4 * Tvm_oto / LineTime, 1) / 4.0; |
| 835 | Tr0_oto_lines = dml_ceil(4 * Tr0_oto / LineTime, 1) / 4.0; |
| 836 | Tsw_oto_lines = dml_ceil(4 * Tsw_oto / LineTime, 1) / 4.0; |
| 837 | dst_y_prefetch_oto = Tvm_oto_lines + 2 * Tr0_oto_lines + Tsw_oto_lines + 0.75; |
| 838 | |
| 839 | dst_y_prefetch_equ = dml_floor(4.0 * (dst_y_prefetch_equ + 0.125), 1) / 4.0; |
| 840 | |
| 841 | if (dst_y_prefetch_oto < dst_y_prefetch_equ) |
| 842 | *DestinationLinesForPrefetch = dst_y_prefetch_oto; |
| 843 | else |
| 844 | *DestinationLinesForPrefetch = dst_y_prefetch_equ; |
| 845 | |
| 846 | // Limit to prevent overflow in DST_Y_PREFETCH register |
| 847 | *DestinationLinesForPrefetch = dml_min(*DestinationLinesForPrefetch, 63.75); |
| 848 | |
| 849 | dml_print("DML: VStartup: %d\n", VStartup){do { } while(0); }; |
| 850 | dml_print("DML: TCalc: %f\n", TCalc){do { } while(0); }; |
| 851 | dml_print("DML: TWait: %f\n", TWait){do { } while(0); }; |
| 852 | dml_print("DML: XFCRemoteSurfaceFlipDelay: %f\n", XFCRemoteSurfaceFlipDelay){do { } while(0); }; |
| 853 | dml_print("DML: LineTime: %f\n", LineTime){do { } while(0); }; |
| 854 | dml_print("DML: Tsetup: %f\n", Tsetup){do { } while(0); }; |
| 855 | dml_print("DML: Tdm: %f\n", Tdm){do { } while(0); }; |
| 856 | dml_print("DML: DSTYAfterScaler: %f\n", *DSTYAfterScaler){do { } while(0); }; |
| 857 | dml_print("DML: DSTXAfterScaler: %f\n", *DSTXAfterScaler){do { } while(0); }; |
| 858 | dml_print("DML: HTotal: %d\n", myPipe->HTotal){do { } while(0); }; |
| 859 | |
| 860 | *PrefetchBandwidth = 0; |
| 861 | *DestinationLinesToRequestVMInVBlank = 0; |
| 862 | *DestinationLinesToRequestRowInVBlank = 0; |
| 863 | *VRatioPrefetchY = 0; |
| 864 | *VRatioPrefetchC = 0; |
| 865 | *RequiredPrefetchPixDataBWLuma = 0; |
| 866 | if (*DestinationLinesForPrefetch > 1) { |
| 867 | double PrefetchBandwidth1 = (PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor + 2 * MetaRowByte |
| 868 | + 2 * PixelPTEBytesPerRow * HostVMInefficiencyFactor |
| 869 | + PrefetchSourceLinesY * *swath_width_luma_ub * dml_ceil(BytePerPixelDETY, 1) |
| 870 | + PrefetchSourceLinesC * *swath_width_chroma_ub * dml_ceil(BytePerPixelDETC, 2)) |
| 871 | / (*DestinationLinesForPrefetch * LineTime - *Tno_bw); |
| 872 | |
| 873 | double PrefetchBandwidth2 = (PDEAndMetaPTEBytesFrame * |
| 874 | HostVMInefficiencyFactor + PrefetchSourceLinesY * |
| 875 | *swath_width_luma_ub * dml_ceil(BytePerPixelDETY, 1) + |
| 876 | PrefetchSourceLinesC * *swath_width_chroma_ub * |
| 877 | dml_ceil(BytePerPixelDETC, 2)) / |
| 878 | (*DestinationLinesForPrefetch * LineTime - *Tno_bw - 2 * |
| 879 | UrgentLatency * (1 + HostVMDynamicLevels)); |
| 880 | |
| 881 | double PrefetchBandwidth3 = (2 * MetaRowByte + 2 * PixelPTEBytesPerRow |
| 882 | * HostVMInefficiencyFactor + PrefetchSourceLinesY * |
| 883 | *swath_width_luma_ub * dml_ceil(BytePerPixelDETY, 1) + |
| 884 | PrefetchSourceLinesC * *swath_width_chroma_ub * |
| 885 | dml_ceil(BytePerPixelDETC, 2)) / |
| 886 | (*DestinationLinesForPrefetch * LineTime - |
| 887 | UrgentExtraLatency - UrgentLatency * (GPUVMPageTableLevels |
| 888 | * (HostVMDynamicLevels + 1) - 1)); |
| 889 | |
| 890 | double PrefetchBandwidth4 = (PrefetchSourceLinesY * *swath_width_luma_ub * |
| 891 | dml_ceil(BytePerPixelDETY, 1) + PrefetchSourceLinesC * |
| 892 | *swath_width_chroma_ub * dml_ceil(BytePerPixelDETC, 2)) / |
| 893 | (*DestinationLinesForPrefetch * LineTime - |
| 894 | UrgentExtraLatency - UrgentLatency * (GPUVMPageTableLevels |
| 895 | * (HostVMDynamicLevels + 1) - 1) - 2 * UrgentLatency * |
| 896 | (1 + HostVMDynamicLevels)); |
| 897 | |
| 898 | if (VStartup == MaxVStartup && (PrefetchBandwidth1 > 4 * prefetch_bw_oto) && (*DestinationLinesForPrefetch - dml_ceil(Tsw_oto_lines, 1) / 4.0 - 0.75) * LineTime - *Tno_bw > 0) { |
| 899 | PrefetchBandwidth1 = (PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor + 2 * MetaRowByte + 2 * PixelPTEBytesPerRow * HostVMInefficiencyFactor) / ((*DestinationLinesForPrefetch - dml_ceil(Tsw_oto_lines, 1) / 4.0 - 0.75) * LineTime - *Tno_bw); |
| 900 | } |
| 901 | if (*Tno_bw + PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor / PrefetchBandwidth1 >= UrgentExtraLatency + UrgentLatency * (GPUVMPageTableLevels * (HostVMDynamicLevels + 1) - 1) && (MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor) / PrefetchBandwidth1 >= UrgentLatency * (1 + HostVMDynamicLevels)) { |
| 902 | *PrefetchBandwidth = PrefetchBandwidth1; |
| 903 | } else if (*Tno_bw + PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor / PrefetchBandwidth2 >= UrgentExtraLatency + UrgentLatency * (GPUVMPageTableLevels * (HostVMDynamicLevels + 1) - 1) && (MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor) / PrefetchBandwidth2 < UrgentLatency * (1 + HostVMDynamicLevels)) { |
| 904 | *PrefetchBandwidth = PrefetchBandwidth2; |
| 905 | } else if (*Tno_bw + PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor / PrefetchBandwidth3 < UrgentExtraLatency + UrgentLatency * (GPUVMPageTableLevels * (HostVMDynamicLevels + 1) - 1) && (MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor) / PrefetchBandwidth3 >= UrgentLatency * (1 + HostVMDynamicLevels)) { |
| 906 | *PrefetchBandwidth = PrefetchBandwidth3; |
| 907 | } else { |
| 908 | *PrefetchBandwidth = PrefetchBandwidth4; |
| 909 | } |
| 910 | |
| 911 | if (GPUVMEnable) { |
| 912 | TimeForFetchingMetaPTE = dml_max(*Tno_bw + (double) PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor / *PrefetchBandwidth, |
| 913 | dml_max(UrgentExtraLatency + UrgentLatency * (GPUVMPageTableLevels * (HostVMDynamicLevels + 1) - 1), LineTime / 4)); |
| 914 | } else { |
| 915 | // 5/30/2018 - This was an optimization requested from Sy but now NumberOfCursors is no longer a factor |
| 916 | // so if this needs to be reinstated, then it should be officially done in the VBA code as well. |
| 917 | // if (mode_lib->NumberOfCursors > 0 || XFCEnabled) |
| 918 | TimeForFetchingMetaPTE = LineTime / 4; |
| 919 | // else |
| 920 | // TimeForFetchingMetaPTE = 0.0; |
| 921 | } |
| 922 | |
| 923 | if ((GPUVMEnable == true1 || DCCEnable == true1)) { |
| 924 | TimeForFetchingRowInVBlank = |
| 925 | dml_max( |
| 926 | (MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor) |
| 927 | / *PrefetchBandwidth, |
| 928 | dml_max( |
| 929 | UrgentLatency * (1 + HostVMDynamicLevels), |
| 930 | dml_max( |
| 931 | (LineTime |
| 932 | - TimeForFetchingMetaPTE) / 2.0, |
| 933 | LineTime |
| 934 | / 4.0))); |
| 935 | } else { |
| 936 | // See note above dated 5/30/2018 |
| 937 | // if (NumberOfCursors > 0 || XFCEnabled) |
| 938 | TimeForFetchingRowInVBlank = (LineTime - TimeForFetchingMetaPTE) / 2.0; |
| 939 | // else // TODO: Did someone else add this?? |
| 940 | // TimeForFetchingRowInVBlank = 0.0; |
| 941 | } |
| 942 | |
| 943 | *DestinationLinesToRequestVMInVBlank = dml_ceil(4.0 * TimeForFetchingMetaPTE / LineTime, 1.0) / 4.0; |
| 944 | |
| 945 | *DestinationLinesToRequestRowInVBlank = dml_ceil(4.0 * TimeForFetchingRowInVBlank / LineTime, 1.0) / 4.0; |
| 946 | |
| 947 | LinesToRequestPrefetchPixelData = *DestinationLinesForPrefetch |
| 948 | // See note above dated 5/30/2018 |
| 949 | // - ((NumberOfCursors > 0 || GPUVMEnable || DCCEnable) ? |
| 950 | - ((GPUVMEnable || DCCEnable) ? |
| 951 | (*DestinationLinesToRequestVMInVBlank + 2 * *DestinationLinesToRequestRowInVBlank) : |
| 952 | 0.0); // TODO: Did someone else add this?? |
| 953 | |
| 954 | if (LinesToRequestPrefetchPixelData > 0) { |
| 955 | |
| 956 | *VRatioPrefetchY = (double) PrefetchSourceLinesY |
| 957 | / LinesToRequestPrefetchPixelData; |
| 958 | *VRatioPrefetchY = dml_max(*VRatioPrefetchY, 1.0); |
| 959 | if ((SwathHeightY > 4) && (VInitPreFillY > 3)) { |
| 960 | if (LinesToRequestPrefetchPixelData > (VInitPreFillY - 3.0) / 2.0) { |
| 961 | *VRatioPrefetchY = |
| 962 | dml_max( |
| 963 | (double) PrefetchSourceLinesY |
| 964 | / LinesToRequestPrefetchPixelData, |
| 965 | (double) MaxNumSwathY |
| 966 | * SwathHeightY |
| 967 | / (LinesToRequestPrefetchPixelData |
| 968 | - (VInitPreFillY |
| 969 | - 3.0) |
| 970 | / 2.0)); |
| 971 | *VRatioPrefetchY = dml_max(*VRatioPrefetchY, 1.0); |
| 972 | } else { |
| 973 | MyError = true1; |
| 974 | *VRatioPrefetchY = 0; |
| 975 | } |
| 976 | } |
| 977 | |
| 978 | *VRatioPrefetchC = (double) PrefetchSourceLinesC |
| 979 | / LinesToRequestPrefetchPixelData; |
| 980 | *VRatioPrefetchC = dml_max(*VRatioPrefetchC, 1.0); |
| 981 | |
| 982 | if ((SwathHeightC > 4)) { |
| 983 | if (LinesToRequestPrefetchPixelData > (VInitPreFillC - 3.0) / 2.0) { |
| 984 | *VRatioPrefetchC = |
| 985 | dml_max( |
| 986 | *VRatioPrefetchC, |
| 987 | (double) MaxNumSwathC |
| 988 | * SwathHeightC |
| 989 | / (LinesToRequestPrefetchPixelData |
| 990 | - (VInitPreFillC |
| 991 | - 3.0) |
| 992 | / 2.0)); |
| 993 | *VRatioPrefetchC = dml_max(*VRatioPrefetchC, 1.0); |
| 994 | } else { |
| 995 | MyError = true1; |
| 996 | *VRatioPrefetchC = 0; |
| 997 | } |
| 998 | } |
| 999 | |
| 1000 | *RequiredPrefetchPixDataBWLuma = myPipe->DPPPerPlane |
| 1001 | * (double) PrefetchSourceLinesY / LinesToRequestPrefetchPixelData |
| 1002 | * dml_ceil(BytePerPixelDETY, 1) |
| 1003 | * *swath_width_luma_ub / LineTime; |
| 1004 | *RequiredPrefetchPixDataBWChroma = myPipe->DPPPerPlane |
| 1005 | * (double) PrefetchSourceLinesC / LinesToRequestPrefetchPixelData |
| 1006 | * dml_ceil(BytePerPixelDETC, 2) |
| 1007 | * *swath_width_chroma_ub / LineTime; |
| 1008 | } else { |
| 1009 | MyError = true1; |
| 1010 | *VRatioPrefetchY = 0; |
| 1011 | *VRatioPrefetchC = 0; |
| 1012 | *RequiredPrefetchPixDataBWLuma = 0; |
| 1013 | *RequiredPrefetchPixDataBWChroma = 0; |
| 1014 | } |
| 1015 | |
| 1016 | dml_print("DML: Tvm: %fus\n", TimeForFetchingMetaPTE){do { } while(0); }; |
| 1017 | dml_print("DML: Tr0: %fus\n", TimeForFetchingRowInVBlank){do { } while(0); }; |
| 1018 | dml_print("DML: Tsw: %fus\n", (double)(*DestinationLinesForPrefetch) * LineTime - TimeForFetchingMetaPTE - TimeForFetchingRowInVBlank){do { } while(0); }; |
| 1019 | dml_print("DML: Tpre: %fus\n", (double)(*DestinationLinesForPrefetch) * LineTime){do { } while(0); }; |
| 1020 | dml_print("DML: row_bytes = dpte_row_bytes (per_pipe) = PixelPTEBytesPerRow = : %d\n", PixelPTEBytesPerRow){do { } while(0); }; |
| 1021 | |
| 1022 | } else { |
| 1023 | MyError = true1; |
| 1024 | } |
| 1025 | |
| 1026 | { |
| 1027 | double prefetch_vm_bw; |
| 1028 | double prefetch_row_bw; |
| 1029 | |
| 1030 | if (PDEAndMetaPTEBytesFrame == 0) { |
| 1031 | prefetch_vm_bw = 0; |
| 1032 | } else if (*DestinationLinesToRequestVMInVBlank > 0) { |
| 1033 | prefetch_vm_bw = PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor / (*DestinationLinesToRequestVMInVBlank * LineTime); |
| 1034 | } else { |
| 1035 | prefetch_vm_bw = 0; |
| 1036 | MyError = true1; |
| 1037 | } |
| 1038 | if (MetaRowByte + PixelPTEBytesPerRow == 0) { |
| 1039 | prefetch_row_bw = 0; |
| 1040 | } else if (*DestinationLinesToRequestRowInVBlank > 0) { |
| 1041 | prefetch_row_bw = (MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor) / (*DestinationLinesToRequestRowInVBlank * LineTime); |
| 1042 | } else { |
| 1043 | prefetch_row_bw = 0; |
| 1044 | MyError = true1; |
| 1045 | } |
| 1046 | |
| 1047 | *prefetch_vmrow_bw = dml_max(prefetch_vm_bw, prefetch_row_bw); |
| 1048 | } |
| 1049 | |
| 1050 | if (MyError) { |
| 1051 | *PrefetchBandwidth = 0; |
| 1052 | TimeForFetchingMetaPTE = 0; |
| 1053 | TimeForFetchingRowInVBlank = 0; |
Value stored to 'TimeForFetchingRowInVBlank' is never read | |
| 1054 | *DestinationLinesToRequestVMInVBlank = 0; |
| 1055 | *DestinationLinesToRequestRowInVBlank = 0; |
| 1056 | *DestinationLinesForPrefetch = 0; |
| 1057 | LinesToRequestPrefetchPixelData = 0; |
| 1058 | *VRatioPrefetchY = 0; |
| 1059 | *VRatioPrefetchC = 0; |
| 1060 | *RequiredPrefetchPixDataBWLuma = 0; |
| 1061 | *RequiredPrefetchPixDataBWChroma = 0; |
| 1062 | } |
| 1063 | |
| 1064 | return MyError; |
| 1065 | } |
| 1066 | |
| 1067 | static double RoundToDFSGranularityUp(double Clock, double VCOSpeed) |
| 1068 | { |
| 1069 | return VCOSpeed * 4 / dml_floor(VCOSpeed * 4 / Clock, 1); |
| 1070 | } |
| 1071 | |
| 1072 | static double RoundToDFSGranularityDown(double Clock, double VCOSpeed) |
| 1073 | { |
| 1074 | return VCOSpeed * 4 / dml_ceil(VCOSpeed * 4 / Clock, 1); |
| 1075 | } |
| 1076 | |
| 1077 | static double CalculateDCCConfiguration( |
| 1078 | bool_Bool DCCEnabled, |
| 1079 | bool_Bool DCCProgrammingAssumesScanDirectionUnknown, |
| 1080 | unsigned int ViewportWidth, |
| 1081 | unsigned int ViewportHeight, |
| 1082 | unsigned int DETBufferSize, |
| 1083 | unsigned int RequestHeight256Byte, |
| 1084 | unsigned int SwathHeight, |
| 1085 | enum dm_swizzle_mode TilingFormat, |
| 1086 | unsigned int BytePerPixel, |
| 1087 | enum scan_direction_class ScanOrientation, |
| 1088 | unsigned int *MaxUncompressedBlock, |
| 1089 | unsigned int *MaxCompressedBlock, |
| 1090 | unsigned int *Independent64ByteBlock) |
| 1091 | { |
| 1092 | double MaximumDCCCompressionSurface = 0.0; |
| 1093 | enum { |
| 1094 | REQ_256Bytes, |
| 1095 | REQ_128BytesNonContiguous, |
| 1096 | REQ_128BytesContiguous, |
| 1097 | REQ_NA |
| 1098 | } Request = REQ_NA; |
| 1099 | |
| 1100 | if (DCCEnabled == true1) { |
| 1101 | if (DCCProgrammingAssumesScanDirectionUnknown == true1) { |
| 1102 | if (DETBufferSize >= RequestHeight256Byte * ViewportWidth * BytePerPixel |
| 1103 | && DETBufferSize |
| 1104 | >= 256 / RequestHeight256Byte |
| 1105 | * ViewportHeight) { |
| 1106 | Request = REQ_256Bytes; |
| 1107 | } else if ((DETBufferSize |
| 1108 | < RequestHeight256Byte * ViewportWidth * BytePerPixel |
| 1109 | && (BytePerPixel == 2 || BytePerPixel == 4)) |
| 1110 | || (DETBufferSize |
| 1111 | < 256 / RequestHeight256Byte |
| 1112 | * ViewportHeight |
| 1113 | && BytePerPixel == 8 |
| 1114 | && (TilingFormat == dm_sw_4kb_d |
| 1115 | || TilingFormat |
| 1116 | == dm_sw_4kb_d_x |
| 1117 | || TilingFormat |
| 1118 | == dm_sw_var_d |
| 1119 | || TilingFormat |
| 1120 | == dm_sw_var_d_x |
| 1121 | || TilingFormat |
| 1122 | == dm_sw_64kb_d |
| 1123 | || TilingFormat |
| 1124 | == dm_sw_64kb_d_x |
| 1125 | || TilingFormat |
| 1126 | == dm_sw_64kb_d_t |
| 1127 | || TilingFormat |
| 1128 | == dm_sw_64kb_r_x))) { |
| 1129 | Request = REQ_128BytesNonContiguous; |
| 1130 | } else { |
| 1131 | Request = REQ_128BytesContiguous; |
| 1132 | } |
| 1133 | } else { |
| 1134 | if (BytePerPixel == 1) { |
| 1135 | if (ScanOrientation == dm_vert || SwathHeight == 16) { |
| 1136 | Request = REQ_256Bytes; |
| 1137 | } else { |
| 1138 | Request = REQ_128BytesContiguous; |
| 1139 | } |
| 1140 | } else if (BytePerPixel == 2) { |
| 1141 | if ((ScanOrientation == dm_vert && SwathHeight == 16) || (ScanOrientation != dm_vert && SwathHeight == 8)) { |
| 1142 | Request = REQ_256Bytes; |
| 1143 | } else if (ScanOrientation == dm_vert) { |
| 1144 | Request = REQ_128BytesContiguous; |
| 1145 | } else { |
| 1146 | Request = REQ_128BytesNonContiguous; |
| 1147 | } |
| 1148 | } else if (BytePerPixel == 4) { |
| 1149 | if (SwathHeight == 8) { |
| 1150 | Request = REQ_256Bytes; |
| 1151 | } else if (ScanOrientation == dm_vert) { |
| 1152 | Request = REQ_128BytesContiguous; |
| 1153 | } else { |
| 1154 | Request = REQ_128BytesNonContiguous; |
| 1155 | } |
| 1156 | } else if (BytePerPixel == 8) { |
| 1157 | if (TilingFormat == dm_sw_4kb_d || TilingFormat == dm_sw_4kb_d_x |
| 1158 | || TilingFormat == dm_sw_var_d |
| 1159 | || TilingFormat == dm_sw_var_d_x |
| 1160 | || TilingFormat == dm_sw_64kb_d |
| 1161 | || TilingFormat == dm_sw_64kb_d_x |
| 1162 | || TilingFormat == dm_sw_64kb_d_t |
| 1163 | || TilingFormat == dm_sw_64kb_r_x) { |
| 1164 | if ((ScanOrientation == dm_vert && SwathHeight == 8) |
| 1165 | || (ScanOrientation != dm_vert |
| 1166 | && SwathHeight == 4)) { |
| 1167 | Request = REQ_256Bytes; |
| 1168 | } else if (ScanOrientation != dm_vert) { |
| 1169 | Request = REQ_128BytesContiguous; |
| 1170 | } else { |
| 1171 | Request = REQ_128BytesNonContiguous; |
| 1172 | } |
| 1173 | } else { |
| 1174 | if (ScanOrientation != dm_vert || SwathHeight == 8) { |
| 1175 | Request = REQ_256Bytes; |
| 1176 | } else { |
| 1177 | Request = REQ_128BytesContiguous; |
| 1178 | } |
| 1179 | } |
| 1180 | } |
| 1181 | } |
| 1182 | } else { |
| 1183 | Request = REQ_NA; |
| 1184 | } |
| 1185 | |
| 1186 | if (Request == REQ_256Bytes) { |
| 1187 | *MaxUncompressedBlock = 256; |
| 1188 | *MaxCompressedBlock = 256; |
| 1189 | *Independent64ByteBlock = false0; |
| 1190 | MaximumDCCCompressionSurface = 4.0; |
| 1191 | } else if (Request == REQ_128BytesContiguous) { |
| 1192 | *MaxUncompressedBlock = 128; |
| 1193 | *MaxCompressedBlock = 128; |
| 1194 | *Independent64ByteBlock = false0; |
| 1195 | MaximumDCCCompressionSurface = 2.0; |
| 1196 | } else if (Request == REQ_128BytesNonContiguous) { |
| 1197 | *MaxUncompressedBlock = 256; |
| 1198 | *MaxCompressedBlock = 64; |
| 1199 | *Independent64ByteBlock = true1; |
| 1200 | MaximumDCCCompressionSurface = 4.0; |
| 1201 | } else { |
| 1202 | *MaxUncompressedBlock = 0; |
| 1203 | *MaxCompressedBlock = 0; |
| 1204 | *Independent64ByteBlock = 0; |
| 1205 | MaximumDCCCompressionSurface = 0.0; |
| 1206 | } |
| 1207 | |
| 1208 | return MaximumDCCCompressionSurface; |
| 1209 | } |
| 1210 | |
| 1211 | static double CalculatePrefetchSourceLines( |
| 1212 | struct display_mode_lib *mode_lib, |
| 1213 | double VRatio, |
| 1214 | double vtaps, |
| 1215 | bool_Bool Interlace, |
| 1216 | bool_Bool ProgressiveToInterlaceUnitInOPP, |
| 1217 | unsigned int SwathHeight, |
| 1218 | unsigned int ViewportYStart, |
| 1219 | double *VInitPreFill, |
| 1220 | unsigned int *MaxNumSwath) |
| 1221 | { |
| 1222 | unsigned int MaxPartialSwath; |
| 1223 | |
| 1224 | if (ProgressiveToInterlaceUnitInOPP) |
| 1225 | *VInitPreFill = dml_floor((VRatio + vtaps + 1) / 2.0, 1); |
| 1226 | else |
| 1227 | *VInitPreFill = dml_floor((VRatio + vtaps + 1 + Interlace * 0.5 * VRatio) / 2.0, 1); |
| 1228 | |
| 1229 | if (!mode_lib->vba.IgnoreViewportPositioning) { |
| 1230 | |
| 1231 | *MaxNumSwath = dml_ceil((*VInitPreFill - 1.0) / SwathHeight, 1) + 1.0; |
| 1232 | |
| 1233 | if (*VInitPreFill > 1.0) |
| 1234 | MaxPartialSwath = (unsigned int) (*VInitPreFill - 2) % SwathHeight; |
| 1235 | else |
| 1236 | MaxPartialSwath = (unsigned int) (*VInitPreFill + SwathHeight - 2) |
| 1237 | % SwathHeight; |
| 1238 | MaxPartialSwath = dml_max(1U, MaxPartialSwath); |
| 1239 | |
| 1240 | } else { |
| 1241 | |
| 1242 | if (ViewportYStart != 0) |
| 1243 | dml_print({do { } while(0); } |
| 1244 | "WARNING DML: using viewport y position of 0 even though actual viewport y position is non-zero in prefetch source lines calculation\n"){do { } while(0); }; |
| 1245 | |
| 1246 | *MaxNumSwath = dml_ceil(*VInitPreFill / SwathHeight, 1); |
| 1247 | |
| 1248 | if (*VInitPreFill > 1.0) |
| 1249 | MaxPartialSwath = (unsigned int) (*VInitPreFill - 1) % SwathHeight; |
| 1250 | else |
| 1251 | MaxPartialSwath = (unsigned int) (*VInitPreFill + SwathHeight - 1) |
| 1252 | % SwathHeight; |
| 1253 | } |
| 1254 | |
| 1255 | return *MaxNumSwath * SwathHeight + MaxPartialSwath; |
| 1256 | } |
| 1257 | |
| 1258 | static unsigned int CalculateVMAndRowBytes( |
| 1259 | struct display_mode_lib *mode_lib, |
| 1260 | bool_Bool DCCEnable, |
| 1261 | unsigned int BlockHeight256Bytes, |
| 1262 | unsigned int BlockWidth256Bytes, |
| 1263 | enum source_format_class SourcePixelFormat, |
| 1264 | unsigned int SurfaceTiling, |
| 1265 | unsigned int BytePerPixel, |
| 1266 | enum scan_direction_class ScanDirection, |
| 1267 | unsigned int ViewportWidth, |
| 1268 | unsigned int ViewportHeight, |
| 1269 | unsigned int SwathWidth, |
| 1270 | bool_Bool GPUVMEnable, |
| 1271 | bool_Bool HostVMEnable, |
| 1272 | unsigned int HostVMMaxPageTableLevels, |
| 1273 | unsigned int HostVMCachedPageTableLevels, |
| 1274 | unsigned int VMMPageSize, |
| 1275 | unsigned int PTEBufferSizeInRequests, |
| 1276 | unsigned int Pitch, |
| 1277 | unsigned int DCCMetaPitch, |
| 1278 | unsigned int *MacroTileWidth, |
| 1279 | unsigned int *MetaRowByte, |
| 1280 | unsigned int *PixelPTEBytesPerRow, |
| 1281 | bool_Bool *PTEBufferSizeNotExceeded, |
| 1282 | unsigned int *dpte_row_width_ub, |
| 1283 | unsigned int *dpte_row_height, |
| 1284 | unsigned int *MetaRequestWidth, |
| 1285 | unsigned int *MetaRequestHeight, |
| 1286 | unsigned int *meta_row_width, |
| 1287 | unsigned int *meta_row_height, |
| 1288 | unsigned int *vm_group_bytes, |
| 1289 | unsigned int *dpte_group_bytes, |
| 1290 | unsigned int *PixelPTEReqWidth, |
| 1291 | unsigned int *PixelPTEReqHeight, |
| 1292 | unsigned int *PTERequestSize, |
| 1293 | unsigned int *DPDE0BytesFrame, |
| 1294 | unsigned int *MetaPTEBytesFrame) |
| 1295 | { |
| 1296 | unsigned int MPDEBytesFrame; |
| 1297 | unsigned int DCCMetaSurfaceBytes; |
| 1298 | unsigned int MacroTileSizeBytes; |
| 1299 | unsigned int MacroTileHeight; |
| 1300 | unsigned int ExtraDPDEBytesFrame; |
| 1301 | unsigned int PDEAndMetaPTEBytesFrame; |
| 1302 | unsigned int PixelPTEReqHeightPTEs = 0; |
| 1303 | |
| 1304 | if (DCCEnable == true1) { |
| 1305 | *MetaRequestHeight = 8 * BlockHeight256Bytes; |
| 1306 | *MetaRequestWidth = 8 * BlockWidth256Bytes; |
| 1307 | if (ScanDirection == dm_horz) { |
| 1308 | *meta_row_height = *MetaRequestHeight; |
| 1309 | *meta_row_width = dml_ceil((double) SwathWidth - 1, *MetaRequestWidth) |
| 1310 | + *MetaRequestWidth; |
| 1311 | *MetaRowByte = *meta_row_width * *MetaRequestHeight * BytePerPixel / 256.0; |
| 1312 | } else { |
| 1313 | *meta_row_height = *MetaRequestWidth; |
| 1314 | *meta_row_width = dml_ceil((double) SwathWidth - 1, *MetaRequestHeight) |
| 1315 | + *MetaRequestHeight; |
| 1316 | *MetaRowByte = *meta_row_width * *MetaRequestWidth * BytePerPixel / 256.0; |
| 1317 | } |
| 1318 | if (ScanDirection == dm_horz) { |
| 1319 | DCCMetaSurfaceBytes = DCCMetaPitch |
| 1320 | * (dml_ceil(ViewportHeight - 1, 64 * BlockHeight256Bytes) |
| 1321 | + 64 * BlockHeight256Bytes) * BytePerPixel |
| 1322 | / 256; |
| 1323 | } else { |
| 1324 | DCCMetaSurfaceBytes = DCCMetaPitch |
| 1325 | * (dml_ceil( |
| 1326 | (double) ViewportHeight - 1, |
| 1327 | 64 * BlockHeight256Bytes) |
| 1328 | + 64 * BlockHeight256Bytes) * BytePerPixel |
| 1329 | / 256; |
| 1330 | } |
| 1331 | if (GPUVMEnable == true1) { |
| 1332 | *MetaPTEBytesFrame = (dml_ceil( |
| 1333 | (double) (DCCMetaSurfaceBytes - VMMPageSize) |
| 1334 | / (8 * VMMPageSize), |
| 1335 | 1) + 1) * 64; |
| 1336 | MPDEBytesFrame = 128 * ((mode_lib->vba.GPUVMMaxPageTableLevels + 1) * (mode_lib->vba.HostVMMaxPageTableLevels + 1) - 2); |
| 1337 | } else { |
| 1338 | *MetaPTEBytesFrame = 0; |
| 1339 | MPDEBytesFrame = 0; |
| 1340 | } |
| 1341 | } else { |
| 1342 | *MetaPTEBytesFrame = 0; |
| 1343 | MPDEBytesFrame = 0; |
| 1344 | *MetaRowByte = 0; |
| 1345 | } |
| 1346 | |
| 1347 | if (SurfaceTiling == dm_sw_linear || SurfaceTiling == dm_sw_gfx7_2d_thin_gl || SurfaceTiling == dm_sw_gfx7_2d_thin_l_vp) { |
| 1348 | MacroTileSizeBytes = 256; |
| 1349 | MacroTileHeight = BlockHeight256Bytes; |
| 1350 | } else if (SurfaceTiling == dm_sw_4kb_s || SurfaceTiling == dm_sw_4kb_s_x |
| 1351 | || SurfaceTiling == dm_sw_4kb_d || SurfaceTiling == dm_sw_4kb_d_x) { |
| 1352 | MacroTileSizeBytes = 4096; |
| 1353 | MacroTileHeight = 4 * BlockHeight256Bytes; |
| 1354 | } else if (SurfaceTiling == dm_sw_64kb_s || SurfaceTiling == dm_sw_64kb_s_t |
| 1355 | || SurfaceTiling == dm_sw_64kb_s_x || SurfaceTiling == dm_sw_64kb_d |
| 1356 | || SurfaceTiling == dm_sw_64kb_d_t || SurfaceTiling == dm_sw_64kb_d_x |
| 1357 | || SurfaceTiling == dm_sw_64kb_r_x) { |
| 1358 | MacroTileSizeBytes = 65536; |
| 1359 | MacroTileHeight = 16 * BlockHeight256Bytes; |
| 1360 | } else { |
| 1361 | MacroTileSizeBytes = 262144; |
| 1362 | MacroTileHeight = 32 * BlockHeight256Bytes; |
| 1363 | } |
| 1364 | *MacroTileWidth = MacroTileSizeBytes / BytePerPixel / MacroTileHeight; |
| 1365 | |
| 1366 | if (GPUVMEnable == true1 && (mode_lib->vba.GPUVMMaxPageTableLevels + 1) * (mode_lib->vba.HostVMMaxPageTableLevels + 1) > 2) { |
| 1367 | if (ScanDirection == dm_horz) { |
| 1368 | *DPDE0BytesFrame = 64 * (dml_ceil(((Pitch * (dml_ceil(ViewportHeight - 1, MacroTileHeight) + MacroTileHeight) * BytePerPixel) - MacroTileSizeBytes) / (8 * 2097152), 1) + 1); |
| 1369 | } else { |
| 1370 | *DPDE0BytesFrame = 64 * (dml_ceil(((Pitch * (dml_ceil((double) SwathWidth - 1, MacroTileHeight) + MacroTileHeight) * BytePerPixel) - MacroTileSizeBytes) / (8 * 2097152), 1) + 1); |
| 1371 | } |
| 1372 | ExtraDPDEBytesFrame = 128 * ((mode_lib->vba.GPUVMMaxPageTableLevels + 1) * (mode_lib->vba.HostVMMaxPageTableLevels + 1) - 3); |
| 1373 | } else { |
| 1374 | *DPDE0BytesFrame = 0; |
| 1375 | ExtraDPDEBytesFrame = 0; |
| 1376 | } |
| 1377 | |
| 1378 | PDEAndMetaPTEBytesFrame = *MetaPTEBytesFrame + MPDEBytesFrame + *DPDE0BytesFrame |
| 1379 | + ExtraDPDEBytesFrame; |
| 1380 | |
| 1381 | if (HostVMEnable == true1) { |
| 1382 | PDEAndMetaPTEBytesFrame = PDEAndMetaPTEBytesFrame * (1 + 8 * (HostVMMaxPageTableLevels - HostVMCachedPageTableLevels)); |
| 1383 | } |
| 1384 | |
| 1385 | if (GPUVMEnable == true1) { |
| 1386 | double FractionOfPTEReturnDrop; |
| 1387 | |
| 1388 | if (SurfaceTiling == dm_sw_linear) { |
| 1389 | PixelPTEReqHeightPTEs = 1; |
| 1390 | *PixelPTEReqHeight = 1; |
| 1391 | *PixelPTEReqWidth = 8.0 * VMMPageSize / BytePerPixel; |
| 1392 | *PTERequestSize = 64; |
| 1393 | FractionOfPTEReturnDrop = 0; |
| 1394 | } else if (MacroTileSizeBytes == 4096) { |
| 1395 | PixelPTEReqHeightPTEs = 1; |
| 1396 | *PixelPTEReqHeight = MacroTileHeight; |
| 1397 | *PixelPTEReqWidth = 8 * *MacroTileWidth; |
| 1398 | *PTERequestSize = 64; |
| 1399 | if (ScanDirection == dm_horz) |
| 1400 | FractionOfPTEReturnDrop = 0; |
| 1401 | else |
| 1402 | FractionOfPTEReturnDrop = 7 / 8; |
| 1403 | } else if (VMMPageSize == 4096 && MacroTileSizeBytes > 4096) { |
| 1404 | PixelPTEReqHeightPTEs = 16; |
| 1405 | *PixelPTEReqHeight = 16 * BlockHeight256Bytes; |
| 1406 | *PixelPTEReqWidth = 16 * BlockWidth256Bytes; |
| 1407 | *PTERequestSize = 128; |
| 1408 | FractionOfPTEReturnDrop = 0; |
| 1409 | } else { |
| 1410 | PixelPTEReqHeightPTEs = 1; |
| 1411 | *PixelPTEReqHeight = MacroTileHeight; |
| 1412 | *PixelPTEReqWidth = 8 * *MacroTileWidth; |
| 1413 | *PTERequestSize = 64; |
| 1414 | FractionOfPTEReturnDrop = 0; |
| 1415 | } |
| 1416 | |
| 1417 | if (SurfaceTiling == dm_sw_linear) { |
| 1418 | *dpte_row_height = dml_min(128, |
| 1419 | 1 << (unsigned int) dml_floor( |
| 1420 | dml_log2( |
| 1421 | (double) PTEBufferSizeInRequests * *PixelPTEReqWidth / Pitch), |
| 1422 | 1)); |
| 1423 | *dpte_row_width_ub = (dml_ceil((double) (Pitch * *dpte_row_height - 1) / *PixelPTEReqWidth, 1) + 1) * *PixelPTEReqWidth; |
| 1424 | *PixelPTEBytesPerRow = *dpte_row_width_ub / *PixelPTEReqWidth * *PTERequestSize; |
| 1425 | } else if (ScanDirection == dm_horz) { |
| 1426 | *dpte_row_height = *PixelPTEReqHeight; |
| 1427 | *dpte_row_width_ub = (dml_ceil((double) (SwathWidth - 1) / *PixelPTEReqWidth, 1) + 1) * *PixelPTEReqWidth; |
| 1428 | *PixelPTEBytesPerRow = *dpte_row_width_ub / *PixelPTEReqWidth * *PTERequestSize; |
| 1429 | } else { |
| 1430 | *dpte_row_height = dml_min(*PixelPTEReqWidth, *MacroTileWidth); |
| 1431 | *dpte_row_width_ub = (dml_ceil((double) (SwathWidth - 1) / *PixelPTEReqHeight, 1) + 1) * *PixelPTEReqHeight; |
| 1432 | *PixelPTEBytesPerRow = *dpte_row_width_ub / *PixelPTEReqHeight * *PTERequestSize; |
| 1433 | } |
| 1434 | if (*PixelPTEBytesPerRow * (1 - FractionOfPTEReturnDrop) |
| 1435 | <= 64 * PTEBufferSizeInRequests) { |
| 1436 | *PTEBufferSizeNotExceeded = true1; |
| 1437 | } else { |
| 1438 | *PTEBufferSizeNotExceeded = false0; |
| 1439 | } |
| 1440 | } else { |
| 1441 | *PixelPTEBytesPerRow = 0; |
| 1442 | *PTEBufferSizeNotExceeded = true1; |
| 1443 | } |
| 1444 | dml_print("DML: vm_bytes = meta_pte_bytes_per_frame (per_pipe) = MetaPTEBytesFrame = : %d\n", *MetaPTEBytesFrame){do { } while(0); }; |
| 1445 | |
| 1446 | if (HostVMEnable == true1) { |
| 1447 | *PixelPTEBytesPerRow = *PixelPTEBytesPerRow * (1 + 8 * (HostVMMaxPageTableLevels - HostVMCachedPageTableLevels)); |
| 1448 | } |
| 1449 | |
| 1450 | if (HostVMEnable == true1) { |
| 1451 | *vm_group_bytes = 512; |
| 1452 | *dpte_group_bytes = 512; |
| 1453 | } else if (GPUVMEnable == true1) { |
| 1454 | *vm_group_bytes = 2048; |
| 1455 | if (SurfaceTiling != dm_sw_linear && PixelPTEReqHeightPTEs == 1 && ScanDirection != dm_horz) { |
| 1456 | *dpte_group_bytes = 512; |
| 1457 | } else { |
| 1458 | *dpte_group_bytes = 2048; |
| 1459 | } |
| 1460 | } else { |
| 1461 | *vm_group_bytes = 0; |
| 1462 | *dpte_group_bytes = 0; |
| 1463 | } |
| 1464 | |
| 1465 | return PDEAndMetaPTEBytesFrame; |
| 1466 | } |
| 1467 | |
| 1468 | static void DISPCLKDPPCLKDCFCLKDeepSleepPrefetchParametersWatermarksAndPerformanceCalculation( |
| 1469 | struct display_mode_lib *mode_lib) |
| 1470 | { |
| 1471 | struct vba_vars_st *locals = &mode_lib->vba; |
| 1472 | unsigned int j, k; |
| 1473 | |
| 1474 | mode_lib->vba.WritebackDISPCLK = 0.0; |
| 1475 | mode_lib->vba.DISPCLKWithRamping = 0; |
| 1476 | mode_lib->vba.DISPCLKWithoutRamping = 0; |
| 1477 | mode_lib->vba.GlobalDPPCLK = 0.0; |
| 1478 | |
| 1479 | // DISPCLK and DPPCLK Calculation |
| 1480 | // |
| 1481 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 1482 | if (mode_lib->vba.WritebackEnable[k]) { |
| 1483 | mode_lib->vba.WritebackDISPCLK = |
| 1484 | dml_max( |
| 1485 | mode_lib->vba.WritebackDISPCLK, |
| 1486 | CalculateWriteBackDISPCLK( |
| 1487 | mode_lib->vba.WritebackPixelFormat[k], |
| 1488 | mode_lib->vba.PixelClock[k], |
| 1489 | mode_lib->vba.WritebackHRatio[k], |
| 1490 | mode_lib->vba.WritebackVRatio[k], |
| 1491 | mode_lib->vba.WritebackLumaHTaps[k], |
| 1492 | mode_lib->vba.WritebackLumaVTaps[k], |
| 1493 | mode_lib->vba.WritebackChromaHTaps[k], |
| 1494 | mode_lib->vba.WritebackChromaVTaps[k], |
| 1495 | mode_lib->vba.WritebackDestinationWidth[k], |
| 1496 | mode_lib->vba.HTotal[k], |
| 1497 | mode_lib->vba.WritebackChromaLineBufferWidth)); |
| 1498 | } |
| 1499 | } |
| 1500 | |
| 1501 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 1502 | if (mode_lib->vba.HRatio[k] > 1) { |
| 1503 | locals->PSCL_THROUGHPUT_LUMA[k] = dml_min( |
| 1504 | mode_lib->vba.MaxDCHUBToPSCLThroughput, |
| 1505 | mode_lib->vba.MaxPSCLToLBThroughput |
| 1506 | * mode_lib->vba.HRatio[k] |
| 1507 | / dml_ceil( |
| 1508 | mode_lib->vba.htaps[k] |
| 1509 | / 6.0, |
| 1510 | 1)); |
| 1511 | } else { |
| 1512 | locals->PSCL_THROUGHPUT_LUMA[k] = dml_min( |
| 1513 | mode_lib->vba.MaxDCHUBToPSCLThroughput, |
| 1514 | mode_lib->vba.MaxPSCLToLBThroughput); |
| 1515 | } |
| 1516 | |
| 1517 | mode_lib->vba.DPPCLKUsingSingleDPPLuma = |
| 1518 | mode_lib->vba.PixelClock[k] |
| 1519 | * dml_max( |
| 1520 | mode_lib->vba.vtaps[k] / 6.0 |
| 1521 | * dml_min( |
| 1522 | 1.0, |
| 1523 | mode_lib->vba.HRatio[k]), |
| 1524 | dml_max( |
| 1525 | mode_lib->vba.HRatio[k] |
| 1526 | * mode_lib->vba.VRatio[k] |
| 1527 | / locals->PSCL_THROUGHPUT_LUMA[k], |
| 1528 | 1.0)); |
| 1529 | |
| 1530 | if ((mode_lib->vba.htaps[k] > 6 || mode_lib->vba.vtaps[k] > 6) |
| 1531 | && mode_lib->vba.DPPCLKUsingSingleDPPLuma |
| 1532 | < 2 * mode_lib->vba.PixelClock[k]) { |
| 1533 | mode_lib->vba.DPPCLKUsingSingleDPPLuma = 2 * mode_lib->vba.PixelClock[k]; |
| 1534 | } |
| 1535 | |
| 1536 | if ((mode_lib->vba.SourcePixelFormat[k] != dm_420_8 |
| 1537 | && mode_lib->vba.SourcePixelFormat[k] != dm_420_10)) { |
| 1538 | locals->PSCL_THROUGHPUT_CHROMA[k] = 0.0; |
| 1539 | locals->DPPCLKUsingSingleDPP[k] = |
| 1540 | mode_lib->vba.DPPCLKUsingSingleDPPLuma; |
| 1541 | } else { |
| 1542 | if (mode_lib->vba.HRatio[k] > 1) { |
| 1543 | locals->PSCL_THROUGHPUT_CHROMA[k] = |
| 1544 | dml_min( |
| 1545 | mode_lib->vba.MaxDCHUBToPSCLThroughput, |
| 1546 | mode_lib->vba.MaxPSCLToLBThroughput |
| 1547 | * mode_lib->vba.HRatio[k] |
| 1548 | / 2 |
| 1549 | / dml_ceil( |
| 1550 | mode_lib->vba.HTAPsChroma[k] |
| 1551 | / 6.0, |
| 1552 | 1.0)); |
| 1553 | } else { |
| 1554 | locals->PSCL_THROUGHPUT_CHROMA[k] = dml_min( |
| 1555 | mode_lib->vba.MaxDCHUBToPSCLThroughput, |
| 1556 | mode_lib->vba.MaxPSCLToLBThroughput); |
| 1557 | } |
| 1558 | mode_lib->vba.DPPCLKUsingSingleDPPChroma = |
| 1559 | mode_lib->vba.PixelClock[k] |
| 1560 | * dml_max( |
| 1561 | mode_lib->vba.VTAPsChroma[k] |
| 1562 | / 6.0 |
| 1563 | * dml_min( |
| 1564 | 1.0, |
| 1565 | mode_lib->vba.HRatio[k] |
| 1566 | / 2), |
| 1567 | dml_max( |
| 1568 | mode_lib->vba.HRatio[k] |
| 1569 | * mode_lib->vba.VRatio[k] |
| 1570 | / 4 |
| 1571 | / locals->PSCL_THROUGHPUT_CHROMA[k], |
| 1572 | 1.0)); |
| 1573 | |
| 1574 | if ((mode_lib->vba.HTAPsChroma[k] > 6 || mode_lib->vba.VTAPsChroma[k] > 6) |
| 1575 | && mode_lib->vba.DPPCLKUsingSingleDPPChroma |
| 1576 | < 2 * mode_lib->vba.PixelClock[k]) { |
| 1577 | mode_lib->vba.DPPCLKUsingSingleDPPChroma = 2 |
| 1578 | * mode_lib->vba.PixelClock[k]; |
| 1579 | } |
| 1580 | |
| 1581 | locals->DPPCLKUsingSingleDPP[k] = dml_max( |
| 1582 | mode_lib->vba.DPPCLKUsingSingleDPPLuma, |
| 1583 | mode_lib->vba.DPPCLKUsingSingleDPPChroma); |
| 1584 | } |
| 1585 | } |
| 1586 | |
| 1587 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 1588 | if (mode_lib->vba.BlendingAndTiming[k] != k) |
| 1589 | continue; |
| 1590 | if (mode_lib->vba.ODMCombineEnabled[k]) { |
| 1591 | mode_lib->vba.DISPCLKWithRamping = |
| 1592 | dml_max( |
| 1593 | mode_lib->vba.DISPCLKWithRamping, |
| 1594 | mode_lib->vba.PixelClock[k] / 2 |
| 1595 | * (1 |
| 1596 | + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading |
| 1597 | / 100) |
| 1598 | * (1 |
| 1599 | + mode_lib->vba.DISPCLKRampingMargin |
| 1600 | / 100)); |
| 1601 | mode_lib->vba.DISPCLKWithoutRamping = |
| 1602 | dml_max( |
| 1603 | mode_lib->vba.DISPCLKWithoutRamping, |
| 1604 | mode_lib->vba.PixelClock[k] / 2 |
| 1605 | * (1 |
| 1606 | + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading |
| 1607 | / 100)); |
| 1608 | } else if (!mode_lib->vba.ODMCombineEnabled[k]) { |
| 1609 | mode_lib->vba.DISPCLKWithRamping = |
| 1610 | dml_max( |
| 1611 | mode_lib->vba.DISPCLKWithRamping, |
| 1612 | mode_lib->vba.PixelClock[k] |
| 1613 | * (1 |
| 1614 | + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading |
| 1615 | / 100) |
| 1616 | * (1 |
| 1617 | + mode_lib->vba.DISPCLKRampingMargin |
| 1618 | / 100)); |
| 1619 | mode_lib->vba.DISPCLKWithoutRamping = |
| 1620 | dml_max( |
| 1621 | mode_lib->vba.DISPCLKWithoutRamping, |
| 1622 | mode_lib->vba.PixelClock[k] |
| 1623 | * (1 |
| 1624 | + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading |
| 1625 | / 100)); |
| 1626 | } |
| 1627 | } |
| 1628 | |
| 1629 | mode_lib->vba.DISPCLKWithRamping = dml_max( |
| 1630 | mode_lib->vba.DISPCLKWithRamping, |
| 1631 | mode_lib->vba.WritebackDISPCLK); |
| 1632 | mode_lib->vba.DISPCLKWithoutRamping = dml_max( |
| 1633 | mode_lib->vba.DISPCLKWithoutRamping, |
| 1634 | mode_lib->vba.WritebackDISPCLK); |
| 1635 | |
| 1636 | ASSERT(mode_lib->vba.DISPCLKDPPCLKVCOSpeed != 0)do { if (({ static int __warned; int __ret = !!(!(mode_lib-> vba.DISPCLKDPPCLKVCOSpeed != 0)); if (__ret && !__warned ) { printf("WARNING %s failed at %s:%d\n", "!(mode_lib->vba.DISPCLKDPPCLKVCOSpeed != 0)" , "/usr/src/sys/dev/pci/drm/amd/display/dc/dml/dcn21/display_mode_vba_21.c" , 1636); __warned = 1; } __builtin_expect(!!(__ret), 0); })) do {} while (0); } while (0); |
| 1637 | mode_lib->vba.DISPCLKWithRampingRoundedToDFSGranularity = RoundToDFSGranularityUp( |
| 1638 | mode_lib->vba.DISPCLKWithRamping, |
| 1639 | mode_lib->vba.DISPCLKDPPCLKVCOSpeed); |
| 1640 | mode_lib->vba.DISPCLKWithoutRampingRoundedToDFSGranularity = RoundToDFSGranularityUp( |
| 1641 | mode_lib->vba.DISPCLKWithoutRamping, |
| 1642 | mode_lib->vba.DISPCLKDPPCLKVCOSpeed); |
| 1643 | mode_lib->vba.MaxDispclkRoundedToDFSGranularity = RoundToDFSGranularityDown( |
| 1644 | mode_lib->vba.soc.clock_limits[mode_lib->vba.soc.num_states - 1].dispclk_mhz, |
| 1645 | mode_lib->vba.DISPCLKDPPCLKVCOSpeed); |
| 1646 | if (mode_lib->vba.DISPCLKWithoutRampingRoundedToDFSGranularity |
| 1647 | > mode_lib->vba.MaxDispclkRoundedToDFSGranularity) { |
| 1648 | mode_lib->vba.DISPCLK_calculated = |
| 1649 | mode_lib->vba.DISPCLKWithoutRampingRoundedToDFSGranularity; |
| 1650 | } else if (mode_lib->vba.DISPCLKWithRampingRoundedToDFSGranularity |
| 1651 | > mode_lib->vba.MaxDispclkRoundedToDFSGranularity) { |
| 1652 | mode_lib->vba.DISPCLK_calculated = mode_lib->vba.MaxDispclkRoundedToDFSGranularity; |
| 1653 | } else { |
| 1654 | mode_lib->vba.DISPCLK_calculated = |
| 1655 | mode_lib->vba.DISPCLKWithRampingRoundedToDFSGranularity; |
| 1656 | } |
| 1657 | DTRACE(" dispclk_mhz (calculated) = %f", mode_lib->vba.DISPCLK_calculated){do { } while(0); }; |
| 1658 | |
| 1659 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 1660 | mode_lib->vba.DPPCLK_calculated[k] = locals->DPPCLKUsingSingleDPP[k] |
| 1661 | / mode_lib->vba.DPPPerPlane[k] |
| 1662 | * (1 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100); |
| 1663 | mode_lib->vba.GlobalDPPCLK = dml_max( |
| 1664 | mode_lib->vba.GlobalDPPCLK, |
| 1665 | mode_lib->vba.DPPCLK_calculated[k]); |
| 1666 | } |
| 1667 | mode_lib->vba.GlobalDPPCLK = RoundToDFSGranularityUp( |
| 1668 | mode_lib->vba.GlobalDPPCLK, |
| 1669 | mode_lib->vba.DISPCLKDPPCLKVCOSpeed); |
| 1670 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 1671 | mode_lib->vba.DPPCLK_calculated[k] = mode_lib->vba.GlobalDPPCLK / 255 |
| 1672 | * dml_ceil( |
| 1673 | mode_lib->vba.DPPCLK_calculated[k] * 255 |
| 1674 | / mode_lib->vba.GlobalDPPCLK, |
| 1675 | 1); |
| 1676 | DTRACE(" dppclk_mhz[%i] (calculated) = %f", k, mode_lib->vba.DPPCLK_calculated[k]){do { } while(0); }; |
| 1677 | } |
| 1678 | |
| 1679 | // Urgent and B P-State/DRAM Clock Change Watermark |
| 1680 | DTRACE(" dcfclk_mhz = %f", mode_lib->vba.DCFCLK){do { } while(0); }; |
| 1681 | DTRACE(" return_bw_to_dcn = %f", mode_lib->vba.ReturnBandwidthToDCN){do { } while(0); }; |
| 1682 | DTRACE(" return_bus_bw = %f", mode_lib->vba.ReturnBW){do { } while(0); }; |
| 1683 | |
| 1684 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 1685 | bool_Bool MainPlaneDoesODMCombine = false0; |
| 1686 | |
| 1687 | if (mode_lib->vba.SourceScan[k] == dm_horz) |
| 1688 | locals->SwathWidthSingleDPPY[k] = mode_lib->vba.ViewportWidth[k]; |
| 1689 | else |
| 1690 | locals->SwathWidthSingleDPPY[k] = mode_lib->vba.ViewportHeight[k]; |
| 1691 | |
| 1692 | if (mode_lib->vba.ODMCombineEnabled[k] == dm_odm_combine_mode_2to1) |
| 1693 | MainPlaneDoesODMCombine = true1; |
| 1694 | for (j = 0; j < mode_lib->vba.NumberOfActivePlanes; ++j) |
| 1695 | if (mode_lib->vba.BlendingAndTiming[k] == j |
| 1696 | && mode_lib->vba.ODMCombineEnabled[k] == dm_odm_combine_mode_2to1) |
| 1697 | MainPlaneDoesODMCombine = true1; |
| 1698 | |
| 1699 | if (MainPlaneDoesODMCombine == true1) |
| 1700 | locals->SwathWidthY[k] = dml_min( |
| 1701 | (double) locals->SwathWidthSingleDPPY[k], |
| 1702 | dml_round( |
| 1703 | mode_lib->vba.HActive[k] / 2.0 |
| 1704 | * mode_lib->vba.HRatio[k])); |
| 1705 | else |
| 1706 | locals->SwathWidthY[k] = locals->SwathWidthSingleDPPY[k] |
| 1707 | / mode_lib->vba.DPPPerPlane[k]; |
| 1708 | } |
| 1709 | |
| 1710 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 1711 | if (mode_lib->vba.SourcePixelFormat[k] == dm_444_64) { |
| 1712 | locals->BytePerPixelDETY[k] = 8; |
| 1713 | locals->BytePerPixelDETC[k] = 0; |
| 1714 | } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_32) { |
| 1715 | locals->BytePerPixelDETY[k] = 4; |
| 1716 | locals->BytePerPixelDETC[k] = 0; |
| 1717 | } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_16 || mode_lib->vba.SourcePixelFormat[k] == dm_mono_16) { |
| 1718 | locals->BytePerPixelDETY[k] = 2; |
| 1719 | locals->BytePerPixelDETC[k] = 0; |
| 1720 | } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_8 || mode_lib->vba.SourcePixelFormat[k] == dm_mono_8) { |
| 1721 | locals->BytePerPixelDETY[k] = 1; |
| 1722 | locals->BytePerPixelDETC[k] = 0; |
| 1723 | } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_8) { |
| 1724 | locals->BytePerPixelDETY[k] = 1; |
| 1725 | locals->BytePerPixelDETC[k] = 2; |
| 1726 | } else { // dm_420_10 |
| 1727 | locals->BytePerPixelDETY[k] = 4.0 / 3.0; |
| 1728 | locals->BytePerPixelDETC[k] = 8.0 / 3.0; |
| 1729 | } |
| 1730 | } |
| 1731 | |
| 1732 | mode_lib->vba.TotalDataReadBandwidth = 0.0; |
| 1733 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 1734 | locals->ReadBandwidthPlaneLuma[k] = locals->SwathWidthSingleDPPY[k] |
| 1735 | * dml_ceil(locals->BytePerPixelDETY[k], 1) |
| 1736 | / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]) |
| 1737 | * mode_lib->vba.VRatio[k]; |
| 1738 | locals->ReadBandwidthPlaneChroma[k] = locals->SwathWidthSingleDPPY[k] |
| 1739 | / 2 * dml_ceil(locals->BytePerPixelDETC[k], 2) |
| 1740 | / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]) |
| 1741 | * mode_lib->vba.VRatio[k] / 2; |
| 1742 | DTRACE({do { } while(0); } |
| 1743 | " read_bw[%i] = %fBps",{do { } while(0); } |
| 1744 | k,{do { } while(0); } |
| 1745 | locals->ReadBandwidthPlaneLuma[k]{do { } while(0); } |
| 1746 | + locals->ReadBandwidthPlaneChroma[k]){do { } while(0); }; |
| 1747 | mode_lib->vba.TotalDataReadBandwidth += locals->ReadBandwidthPlaneLuma[k] |
| 1748 | + locals->ReadBandwidthPlaneChroma[k]; |
| 1749 | } |
| 1750 | |
| 1751 | // DCFCLK Deep Sleep |
| 1752 | CalculateDCFCLKDeepSleep( |
| 1753 | mode_lib, |
| 1754 | mode_lib->vba.NumberOfActivePlanes, |
| 1755 | locals->BytePerPixelDETY, |
| 1756 | locals->BytePerPixelDETC, |
| 1757 | mode_lib->vba.VRatio, |
| 1758 | locals->SwathWidthY, |
| 1759 | mode_lib->vba.DPPPerPlane, |
| 1760 | mode_lib->vba.HRatio, |
| 1761 | mode_lib->vba.PixelClock, |
| 1762 | locals->PSCL_THROUGHPUT_LUMA, |
| 1763 | locals->PSCL_THROUGHPUT_CHROMA, |
| 1764 | locals->DPPCLK, |
| 1765 | &mode_lib->vba.DCFCLKDeepSleep); |
| 1766 | |
| 1767 | // DSCCLK |
| 1768 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 1769 | if ((mode_lib->vba.BlendingAndTiming[k] != k) || !mode_lib->vba.DSCEnabled[k]) { |
| 1770 | locals->DSCCLK_calculated[k] = 0.0; |
| 1771 | } else { |
| 1772 | if (mode_lib->vba.OutputFormat[k] == dm_420 |
| 1773 | || mode_lib->vba.OutputFormat[k] == dm_n422) |
| 1774 | mode_lib->vba.DSCFormatFactor = 2; |
| 1775 | else |
| 1776 | mode_lib->vba.DSCFormatFactor = 1; |
| 1777 | if (mode_lib->vba.ODMCombineEnabled[k]) |
| 1778 | locals->DSCCLK_calculated[k] = |
| 1779 | mode_lib->vba.PixelClockBackEnd[k] / 6 |
| 1780 | / mode_lib->vba.DSCFormatFactor |
| 1781 | / (1 |
| 1782 | - mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading |
| 1783 | / 100); |
| 1784 | else |
| 1785 | locals->DSCCLK_calculated[k] = |
| 1786 | mode_lib->vba.PixelClockBackEnd[k] / 3 |
| 1787 | / mode_lib->vba.DSCFormatFactor |
| 1788 | / (1 |
| 1789 | - mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading |
| 1790 | / 100); |
| 1791 | } |
| 1792 | } |
| 1793 | |
| 1794 | // DSC Delay |
| 1795 | // TODO |
| 1796 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 1797 | double bpp = mode_lib->vba.OutputBpp[k]; |
| 1798 | unsigned int slices = mode_lib->vba.NumberOfDSCSlices[k]; |
| 1799 | |
| 1800 | if (mode_lib->vba.DSCEnabled[k] && bpp != 0) { |
| 1801 | if (!mode_lib->vba.ODMCombineEnabled[k]) { |
| 1802 | locals->DSCDelay[k] = |
| 1803 | dscceComputeDelay( |
| 1804 | mode_lib->vba.DSCInputBitPerComponent[k], |
| 1805 | bpp, |
| 1806 | dml_ceil( |
| 1807 | (double) mode_lib->vba.HActive[k] |
| 1808 | / mode_lib->vba.NumberOfDSCSlices[k], |
| 1809 | 1), |
| 1810 | slices, |
| 1811 | mode_lib->vba.OutputFormat[k]) |
| 1812 | + dscComputeDelay( |
| 1813 | mode_lib->vba.OutputFormat[k]); |
| 1814 | } else { |
| 1815 | locals->DSCDelay[k] = |
| 1816 | 2 |
| 1817 | * (dscceComputeDelay( |
| 1818 | mode_lib->vba.DSCInputBitPerComponent[k], |
| 1819 | bpp, |
| 1820 | dml_ceil( |
| 1821 | (double) mode_lib->vba.HActive[k] |
| 1822 | / mode_lib->vba.NumberOfDSCSlices[k], |
| 1823 | 1), |
| 1824 | slices / 2.0, |
| 1825 | mode_lib->vba.OutputFormat[k]) |
| 1826 | + dscComputeDelay( |
| 1827 | mode_lib->vba.OutputFormat[k])); |
| 1828 | } |
| 1829 | locals->DSCDelay[k] = locals->DSCDelay[k] |
| 1830 | * mode_lib->vba.PixelClock[k] |
| 1831 | / mode_lib->vba.PixelClockBackEnd[k]; |
| 1832 | } else { |
| 1833 | locals->DSCDelay[k] = 0; |
| 1834 | } |
| 1835 | } |
| 1836 | |
| 1837 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) |
| 1838 | for (j = 0; j < mode_lib->vba.NumberOfActivePlanes; ++j) // NumberOfPlanes |
| 1839 | if (j != k && mode_lib->vba.BlendingAndTiming[k] == j |
| 1840 | && mode_lib->vba.DSCEnabled[j]) |
| 1841 | locals->DSCDelay[k] = locals->DSCDelay[j]; |
| 1842 | |
| 1843 | // Prefetch |
| 1844 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 1845 | unsigned int PDEAndMetaPTEBytesFrameY; |
| 1846 | unsigned int PixelPTEBytesPerRowY; |
| 1847 | unsigned int MetaRowByteY; |
| 1848 | unsigned int MetaRowByteC; |
| 1849 | unsigned int PDEAndMetaPTEBytesFrameC; |
| 1850 | unsigned int PixelPTEBytesPerRowC; |
| 1851 | bool_Bool PTEBufferSizeNotExceededY; |
| 1852 | bool_Bool PTEBufferSizeNotExceededC; |
| 1853 | |
| 1854 | Calculate256BBlockSizes( |
| 1855 | mode_lib->vba.SourcePixelFormat[k], |
| 1856 | mode_lib->vba.SurfaceTiling[k], |
| 1857 | dml_ceil(locals->BytePerPixelDETY[k], 1), |
| 1858 | dml_ceil(locals->BytePerPixelDETC[k], 2), |
| 1859 | &locals->BlockHeight256BytesY[k], |
| 1860 | &locals->BlockHeight256BytesC[k], |
| 1861 | &locals->BlockWidth256BytesY[k], |
| 1862 | &locals->BlockWidth256BytesC[k]); |
| 1863 | |
| 1864 | locals->PrefetchSourceLinesY[k] = CalculatePrefetchSourceLines( |
| 1865 | mode_lib, |
| 1866 | mode_lib->vba.VRatio[k], |
| 1867 | mode_lib->vba.vtaps[k], |
| 1868 | mode_lib->vba.Interlace[k], |
| 1869 | mode_lib->vba.ProgressiveToInterlaceUnitInOPP, |
| 1870 | mode_lib->vba.SwathHeightY[k], |
| 1871 | mode_lib->vba.ViewportYStartY[k], |
| 1872 | &locals->VInitPreFillY[k], |
| 1873 | &locals->MaxNumSwathY[k]); |
| 1874 | |
| 1875 | if ((mode_lib->vba.SourcePixelFormat[k] != dm_444_64 |
| 1876 | && mode_lib->vba.SourcePixelFormat[k] != dm_444_32 |
| 1877 | && mode_lib->vba.SourcePixelFormat[k] != dm_444_16 |
| 1878 | && mode_lib->vba.SourcePixelFormat[k] != dm_444_8)) { |
| 1879 | PDEAndMetaPTEBytesFrameC = |
| 1880 | CalculateVMAndRowBytes( |
| 1881 | mode_lib, |
| 1882 | mode_lib->vba.DCCEnable[k], |
| 1883 | locals->BlockHeight256BytesC[k], |
| 1884 | locals->BlockWidth256BytesC[k], |
| 1885 | mode_lib->vba.SourcePixelFormat[k], |
| 1886 | mode_lib->vba.SurfaceTiling[k], |
| 1887 | dml_ceil( |
| 1888 | locals->BytePerPixelDETC[k], |
| 1889 | 2), |
| 1890 | mode_lib->vba.SourceScan[k], |
| 1891 | mode_lib->vba.ViewportWidth[k] / 2, |
| 1892 | mode_lib->vba.ViewportHeight[k] / 2, |
| 1893 | locals->SwathWidthY[k] / 2, |
| 1894 | mode_lib->vba.GPUVMEnable, |
| 1895 | mode_lib->vba.HostVMEnable, |
| 1896 | mode_lib->vba.HostVMMaxPageTableLevels, |
| 1897 | mode_lib->vba.HostVMCachedPageTableLevels, |
| 1898 | mode_lib->vba.VMMPageSize, |
| 1899 | mode_lib->vba.PTEBufferSizeInRequestsChroma, |
| 1900 | mode_lib->vba.PitchC[k], |
| 1901 | mode_lib->vba.DCCMetaPitchC[k], |
| 1902 | &locals->MacroTileWidthC[k], |
| 1903 | &MetaRowByteC, |
| 1904 | &PixelPTEBytesPerRowC, |
| 1905 | &PTEBufferSizeNotExceededC, |
| 1906 | &locals->dpte_row_width_chroma_ub[k], |
| 1907 | &locals->dpte_row_height_chroma[k], |
| 1908 | &locals->meta_req_width_chroma[k], |
| 1909 | &locals->meta_req_height_chroma[k], |
| 1910 | &locals->meta_row_width_chroma[k], |
| 1911 | &locals->meta_row_height_chroma[k], |
| 1912 | &locals->vm_group_bytes_chroma, |
| 1913 | &locals->dpte_group_bytes_chroma, |
| 1914 | &locals->PixelPTEReqWidthC[k], |
| 1915 | &locals->PixelPTEReqHeightC[k], |
| 1916 | &locals->PTERequestSizeC[k], |
| 1917 | &locals->dpde0_bytes_per_frame_ub_c[k], |
| 1918 | &locals->meta_pte_bytes_per_frame_ub_c[k]); |
| 1919 | |
| 1920 | locals->PrefetchSourceLinesC[k] = CalculatePrefetchSourceLines( |
| 1921 | mode_lib, |
| 1922 | mode_lib->vba.VRatio[k] / 2, |
| 1923 | mode_lib->vba.VTAPsChroma[k], |
| 1924 | mode_lib->vba.Interlace[k], |
| 1925 | mode_lib->vba.ProgressiveToInterlaceUnitInOPP, |
| 1926 | mode_lib->vba.SwathHeightC[k], |
| 1927 | mode_lib->vba.ViewportYStartC[k], |
| 1928 | &locals->VInitPreFillC[k], |
| 1929 | &locals->MaxNumSwathC[k]); |
| 1930 | } else { |
| 1931 | PixelPTEBytesPerRowC = 0; |
| 1932 | PDEAndMetaPTEBytesFrameC = 0; |
| 1933 | MetaRowByteC = 0; |
| 1934 | locals->MaxNumSwathC[k] = 0; |
| 1935 | locals->PrefetchSourceLinesC[k] = 0; |
| 1936 | locals->PTEBufferSizeInRequestsForLuma = mode_lib->vba.PTEBufferSizeInRequestsLuma + mode_lib->vba.PTEBufferSizeInRequestsChroma; |
| 1937 | } |
| 1938 | |
| 1939 | PDEAndMetaPTEBytesFrameY = CalculateVMAndRowBytes( |
| 1940 | mode_lib, |
| 1941 | mode_lib->vba.DCCEnable[k], |
| 1942 | locals->BlockHeight256BytesY[k], |
| 1943 | locals->BlockWidth256BytesY[k], |
| 1944 | mode_lib->vba.SourcePixelFormat[k], |
| 1945 | mode_lib->vba.SurfaceTiling[k], |
| 1946 | dml_ceil(locals->BytePerPixelDETY[k], 1), |
| 1947 | mode_lib->vba.SourceScan[k], |
| 1948 | mode_lib->vba.ViewportWidth[k], |
| 1949 | mode_lib->vba.ViewportHeight[k], |
| 1950 | locals->SwathWidthY[k], |
| 1951 | mode_lib->vba.GPUVMEnable, |
| 1952 | mode_lib->vba.HostVMEnable, |
| 1953 | mode_lib->vba.HostVMMaxPageTableLevels, |
| 1954 | mode_lib->vba.HostVMCachedPageTableLevels, |
| 1955 | mode_lib->vba.VMMPageSize, |
| 1956 | locals->PTEBufferSizeInRequestsForLuma, |
| 1957 | mode_lib->vba.PitchY[k], |
| 1958 | mode_lib->vba.DCCMetaPitchY[k], |
| 1959 | &locals->MacroTileWidthY[k], |
| 1960 | &MetaRowByteY, |
| 1961 | &PixelPTEBytesPerRowY, |
| 1962 | &PTEBufferSizeNotExceededY, |
| 1963 | &locals->dpte_row_width_luma_ub[k], |
| 1964 | &locals->dpte_row_height[k], |
| 1965 | &locals->meta_req_width[k], |
| 1966 | &locals->meta_req_height[k], |
| 1967 | &locals->meta_row_width[k], |
| 1968 | &locals->meta_row_height[k], |
| 1969 | &locals->vm_group_bytes[k], |
| 1970 | &locals->dpte_group_bytes[k], |
| 1971 | &locals->PixelPTEReqWidthY[k], |
| 1972 | &locals->PixelPTEReqHeightY[k], |
| 1973 | &locals->PTERequestSizeY[k], |
| 1974 | &locals->dpde0_bytes_per_frame_ub_l[k], |
| 1975 | &locals->meta_pte_bytes_per_frame_ub_l[k]); |
| 1976 | |
| 1977 | locals->PixelPTEBytesPerRow[k] = PixelPTEBytesPerRowY + PixelPTEBytesPerRowC; |
| 1978 | locals->PDEAndMetaPTEBytesFrame[k] = PDEAndMetaPTEBytesFrameY |
| 1979 | + PDEAndMetaPTEBytesFrameC; |
| 1980 | locals->MetaRowByte[k] = MetaRowByteY + MetaRowByteC; |
| 1981 | |
| 1982 | CalculateActiveRowBandwidth( |
| 1983 | mode_lib->vba.GPUVMEnable, |
| 1984 | mode_lib->vba.SourcePixelFormat[k], |
| 1985 | mode_lib->vba.VRatio[k], |
| 1986 | mode_lib->vba.DCCEnable[k], |
| 1987 | mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k], |
| 1988 | MetaRowByteY, |
| 1989 | MetaRowByteC, |
| 1990 | locals->meta_row_height[k], |
| 1991 | locals->meta_row_height_chroma[k], |
| 1992 | PixelPTEBytesPerRowY, |
| 1993 | PixelPTEBytesPerRowC, |
| 1994 | locals->dpte_row_height[k], |
| 1995 | locals->dpte_row_height_chroma[k], |
| 1996 | &locals->meta_row_bw[k], |
| 1997 | &locals->dpte_row_bw[k]); |
| 1998 | } |
| 1999 | |
| 2000 | mode_lib->vba.TotalDCCActiveDPP = 0; |
| 2001 | mode_lib->vba.TotalActiveDPP = 0; |
| 2002 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 2003 | mode_lib->vba.TotalActiveDPP = mode_lib->vba.TotalActiveDPP |
| 2004 | + mode_lib->vba.DPPPerPlane[k]; |
| 2005 | if (mode_lib->vba.DCCEnable[k]) |
| 2006 | mode_lib->vba.TotalDCCActiveDPP = mode_lib->vba.TotalDCCActiveDPP |
| 2007 | + mode_lib->vba.DPPPerPlane[k]; |
| 2008 | } |
| 2009 | |
| 2010 | mode_lib->vba.UrgentOutOfOrderReturnPerChannel = dml_max3( |
| 2011 | mode_lib->vba.UrgentOutOfOrderReturnPerChannelPixelDataOnly, |
| 2012 | mode_lib->vba.UrgentOutOfOrderReturnPerChannelPixelMixedWithVMData, |
| 2013 | mode_lib->vba.UrgentOutOfOrderReturnPerChannelVMDataOnly); |
| 2014 | |
| 2015 | mode_lib->vba.UrgentRoundTripAndOutOfOrderLatency = |
| 2016 | (mode_lib->vba.RoundTripPingLatencyCycles + 32) / mode_lib->vba.DCFCLK |
| 2017 | + mode_lib->vba.UrgentOutOfOrderReturnPerChannel |
| 2018 | * mode_lib->vba.NumberOfChannels |
| 2019 | / mode_lib->vba.ReturnBW; |
| 2020 | |
| 2021 | mode_lib->vba.UrgentExtraLatency = CalculateExtraLatency( |
| 2022 | mode_lib->vba.UrgentRoundTripAndOutOfOrderLatency, |
| 2023 | mode_lib->vba.TotalActiveDPP, |
| 2024 | mode_lib->vba.PixelChunkSizeInKByte, |
| 2025 | mode_lib->vba.TotalDCCActiveDPP, |
| 2026 | mode_lib->vba.MetaChunkSize, |
| 2027 | mode_lib->vba.ReturnBW, |
| 2028 | mode_lib->vba.GPUVMEnable, |
| 2029 | mode_lib->vba.HostVMEnable, |
| 2030 | mode_lib->vba.NumberOfActivePlanes, |
| 2031 | mode_lib->vba.DPPPerPlane, |
| 2032 | locals->dpte_group_bytes, |
| 2033 | mode_lib->vba.PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
| 2034 | mode_lib->vba.PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
| 2035 | mode_lib->vba.HostVMMaxPageTableLevels, |
| 2036 | mode_lib->vba.HostVMCachedPageTableLevels); |
| 2037 | |
| 2038 | |
| 2039 | mode_lib->vba.TCalc = 24.0 / mode_lib->vba.DCFCLKDeepSleep; |
| 2040 | |
| 2041 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 2042 | if (mode_lib->vba.BlendingAndTiming[k] == k) { |
| 2043 | if (mode_lib->vba.WritebackEnable[k] == true1) { |
| 2044 | locals->WritebackDelay[mode_lib->vba.VoltageLevel][k] = |
| 2045 | mode_lib->vba.WritebackLatency |
| 2046 | + CalculateWriteBackDelay( |
| 2047 | mode_lib->vba.WritebackPixelFormat[k], |
| 2048 | mode_lib->vba.WritebackHRatio[k], |
| 2049 | mode_lib->vba.WritebackVRatio[k], |
| 2050 | mode_lib->vba.WritebackLumaHTaps[k], |
| 2051 | mode_lib->vba.WritebackLumaVTaps[k], |
| 2052 | mode_lib->vba.WritebackChromaHTaps[k], |
| 2053 | mode_lib->vba.WritebackChromaVTaps[k], |
| 2054 | mode_lib->vba.WritebackDestinationWidth[k]) |
| 2055 | / mode_lib->vba.DISPCLK; |
| 2056 | } else |
| 2057 | locals->WritebackDelay[mode_lib->vba.VoltageLevel][k] = 0; |
| 2058 | for (j = 0; j < mode_lib->vba.NumberOfActivePlanes; ++j) { |
| 2059 | if (mode_lib->vba.BlendingAndTiming[j] == k |
| 2060 | && mode_lib->vba.WritebackEnable[j] == true1) { |
| 2061 | locals->WritebackDelay[mode_lib->vba.VoltageLevel][k] = |
| 2062 | dml_max( |
| 2063 | locals->WritebackDelay[mode_lib->vba.VoltageLevel][k], |
| 2064 | mode_lib->vba.WritebackLatency |
| 2065 | + CalculateWriteBackDelay( |
| 2066 | mode_lib->vba.WritebackPixelFormat[j], |
| 2067 | mode_lib->vba.WritebackHRatio[j], |
| 2068 | mode_lib->vba.WritebackVRatio[j], |
| 2069 | mode_lib->vba.WritebackLumaHTaps[j], |
| 2070 | mode_lib->vba.WritebackLumaVTaps[j], |
| 2071 | mode_lib->vba.WritebackChromaHTaps[j], |
| 2072 | mode_lib->vba.WritebackChromaVTaps[j], |
| 2073 | mode_lib->vba.WritebackDestinationWidth[j]) |
| 2074 | / mode_lib->vba.DISPCLK); |
| 2075 | } |
| 2076 | } |
| 2077 | } |
| 2078 | } |
| 2079 | |
| 2080 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) |
| 2081 | for (j = 0; j < mode_lib->vba.NumberOfActivePlanes; ++j) |
| 2082 | if (mode_lib->vba.BlendingAndTiming[k] == j) |
| 2083 | locals->WritebackDelay[mode_lib->vba.VoltageLevel][k] = |
| 2084 | locals->WritebackDelay[mode_lib->vba.VoltageLevel][j]; |
| 2085 | |
| 2086 | mode_lib->vba.VStartupLines = 13; |
| 2087 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 2088 | locals->MaxVStartupLines[k] = mode_lib->vba.VTotal[k] - mode_lib->vba.VActive[k] - dml_max(1.0, dml_ceil(locals->WritebackDelay[mode_lib->vba.VoltageLevel][k] / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]), 1)); |
| 2089 | } |
| 2090 | |
| 2091 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) |
| 2092 | locals->MaximumMaxVStartupLines = dml_max(locals->MaximumMaxVStartupLines, locals->MaxVStartupLines[k]); |
| 2093 | |
| 2094 | // We don't really care to iterate between the various prefetch modes |
| 2095 | //mode_lib->vba.PrefetchERROR = CalculateMinAndMaxPrefetchMode(mode_lib->vba.AllowDRAMSelfRefreshOrDRAMClockChangeInVblank, &mode_lib->vba.MinPrefetchMode, &mode_lib->vba.MaxPrefetchMode); |
| 2096 | mode_lib->vba.UrgentLatency = dml_max3(mode_lib->vba.UrgentLatencyPixelDataOnly, mode_lib->vba.UrgentLatencyPixelMixedWithVMData, mode_lib->vba.UrgentLatencyVMDataOnly); |
| 2097 | |
| 2098 | do { |
| 2099 | double MaxTotalRDBandwidth = 0; |
| 2100 | double MaxTotalRDBandwidthNoUrgentBurst = 0; |
| 2101 | bool_Bool DestinationLineTimesForPrefetchLessThan2 = false0; |
| 2102 | bool_Bool VRatioPrefetchMoreThan4 = false0; |
| 2103 | double TWait = CalculateTWait( |
| 2104 | mode_lib->vba.PrefetchMode[mode_lib->vba.VoltageLevel][mode_lib->vba.maxMpcComb], |
| 2105 | mode_lib->vba.DRAMClockChangeLatency, |
| 2106 | mode_lib->vba.UrgentLatency, |
| 2107 | mode_lib->vba.SREnterPlusExitTime); |
| 2108 | |
| 2109 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 2110 | Pipe myPipe; |
| 2111 | HostVM myHostVM; |
| 2112 | |
| 2113 | if (mode_lib->vba.XFCEnabled[k] == true1) { |
| 2114 | mode_lib->vba.XFCRemoteSurfaceFlipDelay = |
| 2115 | CalculateRemoteSurfaceFlipDelay( |
| 2116 | mode_lib, |
| 2117 | mode_lib->vba.VRatio[k], |
| 2118 | locals->SwathWidthY[k], |
| 2119 | dml_ceil( |
| 2120 | locals->BytePerPixelDETY[k], |
| 2121 | 1), |
| 2122 | mode_lib->vba.HTotal[k] |
| 2123 | / mode_lib->vba.PixelClock[k], |
| 2124 | mode_lib->vba.XFCTSlvVupdateOffset, |
| 2125 | mode_lib->vba.XFCTSlvVupdateWidth, |
| 2126 | mode_lib->vba.XFCTSlvVreadyOffset, |
| 2127 | mode_lib->vba.XFCXBUFLatencyTolerance, |
| 2128 | mode_lib->vba.XFCFillBWOverhead, |
| 2129 | mode_lib->vba.XFCSlvChunkSize, |
| 2130 | mode_lib->vba.XFCBusTransportTime, |
| 2131 | mode_lib->vba.TCalc, |
| 2132 | TWait, |
| 2133 | &mode_lib->vba.SrcActiveDrainRate, |
| 2134 | &mode_lib->vba.TInitXFill, |
| 2135 | &mode_lib->vba.TslvChk); |
| 2136 | } else { |
| 2137 | mode_lib->vba.XFCRemoteSurfaceFlipDelay = 0; |
| 2138 | } |
| 2139 | |
| 2140 | myPipe.DPPCLK = locals->DPPCLK[k]; |
| 2141 | myPipe.DISPCLK = mode_lib->vba.DISPCLK; |
| 2142 | myPipe.PixelClock = mode_lib->vba.PixelClock[k]; |
| 2143 | myPipe.DCFCLKDeepSleep = mode_lib->vba.DCFCLKDeepSleep; |
| 2144 | myPipe.DPPPerPlane = mode_lib->vba.DPPPerPlane[k]; |
| 2145 | myPipe.ScalerEnabled = mode_lib->vba.ScalerEnabled[k]; |
| 2146 | myPipe.SourceScan = mode_lib->vba.SourceScan[k]; |
| 2147 | myPipe.BlockWidth256BytesY = locals->BlockWidth256BytesY[k]; |
| 2148 | myPipe.BlockHeight256BytesY = locals->BlockHeight256BytesY[k]; |
| 2149 | myPipe.BlockWidth256BytesC = locals->BlockWidth256BytesC[k]; |
| 2150 | myPipe.BlockHeight256BytesC = locals->BlockHeight256BytesC[k]; |
| 2151 | myPipe.InterlaceEnable = mode_lib->vba.Interlace[k]; |
| 2152 | myPipe.NumberOfCursors = mode_lib->vba.NumberOfCursors[k]; |
| 2153 | myPipe.VBlank = mode_lib->vba.VTotal[k] - mode_lib->vba.VActive[k]; |
| 2154 | myPipe.HTotal = mode_lib->vba.HTotal[k]; |
| 2155 | |
| 2156 | |
| 2157 | myHostVM.Enable = mode_lib->vba.HostVMEnable; |
| 2158 | myHostVM.MaxPageTableLevels = mode_lib->vba.HostVMMaxPageTableLevels; |
| 2159 | myHostVM.CachedPageTableLevels = mode_lib->vba.HostVMCachedPageTableLevels; |
| 2160 | |
| 2161 | mode_lib->vba.ErrorResult[k] = |
| 2162 | CalculatePrefetchSchedule( |
| 2163 | mode_lib, |
| 2164 | mode_lib->vba.PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
| 2165 | mode_lib->vba.PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
| 2166 | &myPipe, |
| 2167 | locals->DSCDelay[k], |
| 2168 | mode_lib->vba.DPPCLKDelaySubtotal, |
| 2169 | mode_lib->vba.DPPCLKDelaySCL, |
| 2170 | mode_lib->vba.DPPCLKDelaySCLLBOnly, |
| 2171 | mode_lib->vba.DPPCLKDelayCNVCFormater, |
| 2172 | mode_lib->vba.DPPCLKDelayCNVCCursor, |
| 2173 | mode_lib->vba.DISPCLKDelaySubtotal, |
| 2174 | (unsigned int) (locals->SwathWidthY[k] |
| 2175 | / mode_lib->vba.HRatio[k]), |
| 2176 | mode_lib->vba.OutputFormat[k], |
| 2177 | mode_lib->vba.MaxInterDCNTileRepeaters, |
| 2178 | dml_min(mode_lib->vba.VStartupLines, locals->MaxVStartupLines[k]), |
| 2179 | locals->MaxVStartupLines[k], |
| 2180 | mode_lib->vba.GPUVMMaxPageTableLevels, |
| 2181 | mode_lib->vba.GPUVMEnable, |
| 2182 | &myHostVM, |
| 2183 | mode_lib->vba.DynamicMetadataEnable[k], |
| 2184 | mode_lib->vba.DynamicMetadataLinesBeforeActiveRequired[k], |
| 2185 | mode_lib->vba.DynamicMetadataTransmittedBytes[k], |
| 2186 | mode_lib->vba.DCCEnable[k], |
| 2187 | mode_lib->vba.UrgentLatency, |
| 2188 | mode_lib->vba.UrgentExtraLatency, |
| 2189 | mode_lib->vba.TCalc, |
| 2190 | locals->PDEAndMetaPTEBytesFrame[k], |
| 2191 | locals->MetaRowByte[k], |
| 2192 | locals->PixelPTEBytesPerRow[k], |
| 2193 | locals->PrefetchSourceLinesY[k], |
| 2194 | locals->SwathWidthY[k], |
| 2195 | locals->BytePerPixelDETY[k], |
| 2196 | locals->VInitPreFillY[k], |
| 2197 | locals->MaxNumSwathY[k], |
| 2198 | locals->PrefetchSourceLinesC[k], |
| 2199 | locals->BytePerPixelDETC[k], |
| 2200 | locals->VInitPreFillC[k], |
| 2201 | locals->MaxNumSwathC[k], |
| 2202 | mode_lib->vba.SwathHeightY[k], |
| 2203 | mode_lib->vba.SwathHeightC[k], |
| 2204 | TWait, |
| 2205 | mode_lib->vba.XFCEnabled[k], |
| 2206 | mode_lib->vba.XFCRemoteSurfaceFlipDelay, |
| 2207 | mode_lib->vba.ProgressiveToInterlaceUnitInOPP, |
| 2208 | &locals->DSTXAfterScaler[k], |
| 2209 | &locals->DSTYAfterScaler[k], |
| 2210 | &locals->DestinationLinesForPrefetch[k], |
| 2211 | &locals->PrefetchBandwidth[k], |
| 2212 | &locals->DestinationLinesToRequestVMInVBlank[k], |
| 2213 | &locals->DestinationLinesToRequestRowInVBlank[k], |
| 2214 | &locals->VRatioPrefetchY[k], |
| 2215 | &locals->VRatioPrefetchC[k], |
| 2216 | &locals->RequiredPrefetchPixDataBWLuma[k], |
| 2217 | &locals->RequiredPrefetchPixDataBWChroma[k], |
| 2218 | &locals->VStartupRequiredWhenNotEnoughTimeForDynamicMetadata, |
| 2219 | &locals->Tno_bw[k], |
| 2220 | &locals->prefetch_vmrow_bw[k], |
| 2221 | &locals->swath_width_luma_ub[k], |
| 2222 | &locals->swath_width_chroma_ub[k], |
| 2223 | &mode_lib->vba.VUpdateOffsetPix[k], |
| 2224 | &mode_lib->vba.VUpdateWidthPix[k], |
| 2225 | &mode_lib->vba.VReadyOffsetPix[k]); |
| 2226 | if (mode_lib->vba.BlendingAndTiming[k] == k) { |
| 2227 | locals->VStartup[k] = dml_min( |
| 2228 | mode_lib->vba.VStartupLines, |
| 2229 | locals->MaxVStartupLines[k]); |
| 2230 | if (locals->VStartupRequiredWhenNotEnoughTimeForDynamicMetadata |
| 2231 | != 0) { |
| 2232 | locals->VStartup[k] = |
| 2233 | locals->VStartupRequiredWhenNotEnoughTimeForDynamicMetadata; |
| 2234 | } |
| 2235 | } else { |
| 2236 | locals->VStartup[k] = |
| 2237 | dml_min( |
| 2238 | mode_lib->vba.VStartupLines, |
| 2239 | locals->MaxVStartupLines[mode_lib->vba.BlendingAndTiming[k]]); |
| 2240 | } |
| 2241 | } |
| 2242 | |
| 2243 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 2244 | unsigned int m; |
| 2245 | |
| 2246 | locals->cursor_bw[k] = 0; |
| 2247 | locals->cursor_bw_pre[k] = 0; |
| 2248 | for (m = 0; m < mode_lib->vba.NumberOfCursors[k]; m++) { |
| 2249 | locals->cursor_bw[k] += mode_lib->vba.CursorWidth[k][m] * mode_lib->vba.CursorBPP[k][m] / 8.0 / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]) * mode_lib->vba.VRatio[k]; |
| 2250 | locals->cursor_bw_pre[k] += mode_lib->vba.CursorWidth[k][m] * mode_lib->vba.CursorBPP[k][m] / 8.0 / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]) * locals->VRatioPrefetchY[k]; |
| 2251 | } |
| 2252 | |
| 2253 | CalculateUrgentBurstFactor( |
| 2254 | mode_lib->vba.DETBufferSizeInKByte[0], |
| 2255 | mode_lib->vba.SwathHeightY[k], |
| 2256 | mode_lib->vba.SwathHeightC[k], |
| 2257 | locals->SwathWidthY[k], |
| 2258 | mode_lib->vba.HTotal[k] / |
| 2259 | mode_lib->vba.PixelClock[k], |
| 2260 | mode_lib->vba.UrgentLatency, |
| 2261 | mode_lib->vba.CursorBufferSize, |
| 2262 | mode_lib->vba.CursorWidth[k][0] + mode_lib->vba.CursorWidth[k][1], |
| 2263 | dml_max(mode_lib->vba.CursorBPP[k][0], mode_lib->vba.CursorBPP[k][1]), |
| 2264 | mode_lib->vba.VRatio[k], |
| 2265 | locals->VRatioPrefetchY[k], |
| 2266 | locals->VRatioPrefetchC[k], |
| 2267 | locals->BytePerPixelDETY[k], |
| 2268 | locals->BytePerPixelDETC[k], |
| 2269 | &locals->UrgentBurstFactorCursor[k], |
| 2270 | &locals->UrgentBurstFactorCursorPre[k], |
| 2271 | &locals->UrgentBurstFactorLuma[k], |
| 2272 | &locals->UrgentBurstFactorLumaPre[k], |
| 2273 | &locals->UrgentBurstFactorChroma[k], |
| 2274 | &locals->UrgentBurstFactorChromaPre[k], |
| 2275 | &locals->NotEnoughUrgentLatencyHiding[0][0], |
| 2276 | &locals->NotEnoughUrgentLatencyHidingPre); |
| 2277 | |
| 2278 | if (mode_lib->vba.UseUrgentBurstBandwidth == false0) { |
| 2279 | locals->UrgentBurstFactorLuma[k] = 1; |
| 2280 | locals->UrgentBurstFactorChroma[k] = 1; |
| 2281 | locals->UrgentBurstFactorCursor[k] = 1; |
| 2282 | locals->UrgentBurstFactorLumaPre[k] = 1; |
| 2283 | locals->UrgentBurstFactorChromaPre[k] = 1; |
| 2284 | locals->UrgentBurstFactorCursorPre[k] = 1; |
| 2285 | } |
| 2286 | |
| 2287 | MaxTotalRDBandwidth = MaxTotalRDBandwidth + |
| 2288 | dml_max3(locals->prefetch_vmrow_bw[k], |
| 2289 | locals->ReadBandwidthPlaneLuma[k] * locals->UrgentBurstFactorLuma[k] |
| 2290 | + locals->ReadBandwidthPlaneChroma[k] * locals->UrgentBurstFactorChroma[k] + locals->cursor_bw[k] |
| 2291 | * locals->UrgentBurstFactorCursor[k] + locals->meta_row_bw[k] + locals->dpte_row_bw[k], |
| 2292 | locals->RequiredPrefetchPixDataBWLuma[k] * locals->UrgentBurstFactorLumaPre[k] + locals->RequiredPrefetchPixDataBWChroma[k] |
| 2293 | * locals->UrgentBurstFactorChromaPre[k] + locals->cursor_bw_pre[k] * locals->UrgentBurstFactorCursorPre[k]); |
| 2294 | |
| 2295 | MaxTotalRDBandwidthNoUrgentBurst = MaxTotalRDBandwidthNoUrgentBurst + |
| 2296 | dml_max3(locals->prefetch_vmrow_bw[k], |
| 2297 | locals->ReadBandwidthPlaneLuma[k] + locals->ReadBandwidthPlaneChroma[k] + locals->cursor_bw[k] |
| 2298 | + locals->meta_row_bw[k] + locals->dpte_row_bw[k], |
| 2299 | locals->RequiredPrefetchPixDataBWLuma[k] + locals->RequiredPrefetchPixDataBWChroma[k] + locals->cursor_bw_pre[k]); |
| 2300 | |
| 2301 | if (locals->DestinationLinesForPrefetch[k] < 2) |
| 2302 | DestinationLineTimesForPrefetchLessThan2 = true1; |
| 2303 | if (locals->VRatioPrefetchY[k] > 4 || locals->VRatioPrefetchC[k] > 4) |
| 2304 | VRatioPrefetchMoreThan4 = true1; |
| 2305 | } |
| 2306 | mode_lib->vba.FractionOfUrgentBandwidth = MaxTotalRDBandwidthNoUrgentBurst / mode_lib->vba.ReturnBW; |
| 2307 | |
| 2308 | if (MaxTotalRDBandwidth <= mode_lib->vba.ReturnBW && locals->NotEnoughUrgentLatencyHiding[0][0] == 0 && |
| 2309 | locals->NotEnoughUrgentLatencyHidingPre == 0 && !VRatioPrefetchMoreThan4 |
| 2310 | && !DestinationLineTimesForPrefetchLessThan2) |
| 2311 | mode_lib->vba.PrefetchModeSupported = true1; |
| 2312 | else { |
| 2313 | mode_lib->vba.PrefetchModeSupported = false0; |
| 2314 | dml_print({do { } while(0); } |
| 2315 | "DML: CalculatePrefetchSchedule ***failed***. Bandwidth violation. Results are NOT valid\n"){do { } while(0); }; |
| 2316 | } |
| 2317 | |
| 2318 | if (mode_lib->vba.PrefetchModeSupported == true1) { |
| 2319 | mode_lib->vba.BandwidthAvailableForImmediateFlip = mode_lib->vba.ReturnBW; |
| 2320 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 2321 | mode_lib->vba.BandwidthAvailableForImmediateFlip = |
| 2322 | mode_lib->vba.BandwidthAvailableForImmediateFlip |
| 2323 | - dml_max( |
| 2324 | locals->ReadBandwidthPlaneLuma[k] * locals->UrgentBurstFactorLuma[k] |
| 2325 | + locals->ReadBandwidthPlaneChroma[k] * locals->UrgentBurstFactorChroma[k] |
| 2326 | + locals->cursor_bw[k] * locals->UrgentBurstFactorCursor[k], |
| 2327 | locals->RequiredPrefetchPixDataBWLuma[k] * locals->UrgentBurstFactorLumaPre[k] + |
| 2328 | locals->RequiredPrefetchPixDataBWChroma[k] * locals->UrgentBurstFactorChromaPre[k] + |
| 2329 | locals->cursor_bw_pre[k] * locals->UrgentBurstFactorCursorPre[k]); |
| 2330 | } |
| 2331 | |
| 2332 | mode_lib->vba.TotImmediateFlipBytes = 0; |
| 2333 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 2334 | mode_lib->vba.TotImmediateFlipBytes = mode_lib->vba.TotImmediateFlipBytes + locals->PDEAndMetaPTEBytesFrame[k] + locals->MetaRowByte[k] + locals->PixelPTEBytesPerRow[k]; |
| 2335 | } |
| 2336 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 2337 | CalculateFlipSchedule( |
| 2338 | mode_lib, |
| 2339 | mode_lib->vba.PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
| 2340 | mode_lib->vba.PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
| 2341 | mode_lib->vba.UrgentExtraLatency, |
| 2342 | mode_lib->vba.UrgentLatency, |
| 2343 | mode_lib->vba.GPUVMMaxPageTableLevels, |
| 2344 | mode_lib->vba.HostVMEnable, |
| 2345 | mode_lib->vba.HostVMMaxPageTableLevels, |
| 2346 | mode_lib->vba.HostVMCachedPageTableLevels, |
| 2347 | mode_lib->vba.GPUVMEnable, |
| 2348 | locals->PDEAndMetaPTEBytesFrame[k], |
| 2349 | locals->MetaRowByte[k], |
| 2350 | locals->PixelPTEBytesPerRow[k], |
| 2351 | mode_lib->vba.BandwidthAvailableForImmediateFlip, |
| 2352 | mode_lib->vba.TotImmediateFlipBytes, |
| 2353 | mode_lib->vba.SourcePixelFormat[k], |
| 2354 | mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k], |
| 2355 | mode_lib->vba.VRatio[k], |
| 2356 | locals->Tno_bw[k], |
| 2357 | mode_lib->vba.DCCEnable[k], |
| 2358 | locals->dpte_row_height[k], |
| 2359 | locals->meta_row_height[k], |
| 2360 | locals->dpte_row_height_chroma[k], |
| 2361 | locals->meta_row_height_chroma[k], |
| 2362 | &locals->DestinationLinesToRequestVMInImmediateFlip[k], |
| 2363 | &locals->DestinationLinesToRequestRowInImmediateFlip[k], |
| 2364 | &locals->final_flip_bw[k], |
| 2365 | &locals->ImmediateFlipSupportedForPipe[k]); |
| 2366 | } |
| 2367 | mode_lib->vba.total_dcn_read_bw_with_flip = 0.0; |
| 2368 | mode_lib->vba.total_dcn_read_bw_with_flip_no_urgent_burst = 0.0; |
| 2369 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 2370 | mode_lib->vba.total_dcn_read_bw_with_flip = |
| 2371 | mode_lib->vba.total_dcn_read_bw_with_flip + dml_max3( |
| 2372 | locals->prefetch_vmrow_bw[k], |
| 2373 | locals->final_flip_bw[k] + locals->ReadBandwidthLuma[k] * locals->UrgentBurstFactorLuma[k] |
| 2374 | + locals->ReadBandwidthChroma[k] * locals->UrgentBurstFactorChroma[k] + locals->cursor_bw[k] * locals->UrgentBurstFactorCursor[k], |
| 2375 | locals->final_flip_bw[k] + locals->RequiredPrefetchPixDataBWLuma[k] * locals->UrgentBurstFactorLumaPre[k] |
| 2376 | + locals->RequiredPrefetchPixDataBWChroma[k] * locals->UrgentBurstFactorChromaPre[k] |
| 2377 | + locals->cursor_bw_pre[k] * locals->UrgentBurstFactorCursorPre[k]); |
| 2378 | mode_lib->vba.total_dcn_read_bw_with_flip_no_urgent_burst = |
| 2379 | mode_lib->vba.total_dcn_read_bw_with_flip_no_urgent_burst + |
| 2380 | dml_max3(locals->prefetch_vmrow_bw[k], |
| 2381 | locals->final_flip_bw[k] + locals->ReadBandwidthPlaneLuma[k] + locals->ReadBandwidthPlaneChroma[k] + locals->cursor_bw[k], |
| 2382 | locals->final_flip_bw[k] + locals->RequiredPrefetchPixDataBWLuma[k] + locals->RequiredPrefetchPixDataBWChroma[k] + locals->cursor_bw_pre[k]); |
| 2383 | |
| 2384 | } |
| 2385 | mode_lib->vba.FractionOfUrgentBandwidthImmediateFlip = mode_lib->vba.total_dcn_read_bw_with_flip_no_urgent_burst / mode_lib->vba.ReturnBW; |
| 2386 | |
| 2387 | mode_lib->vba.ImmediateFlipSupported = true1; |
| 2388 | if (mode_lib->vba.total_dcn_read_bw_with_flip > mode_lib->vba.ReturnBW) { |
| 2389 | mode_lib->vba.ImmediateFlipSupported = false0; |
| 2390 | } |
| 2391 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 2392 | if (locals->ImmediateFlipSupportedForPipe[k] == false0) { |
| 2393 | mode_lib->vba.ImmediateFlipSupported = false0; |
| 2394 | } |
| 2395 | } |
| 2396 | } else { |
| 2397 | mode_lib->vba.ImmediateFlipSupported = false0; |
| 2398 | } |
| 2399 | |
| 2400 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 2401 | if (mode_lib->vba.ErrorResult[k]) { |
| 2402 | mode_lib->vba.PrefetchModeSupported = false0; |
| 2403 | dml_print({do { } while(0); } |
| 2404 | "DML: CalculatePrefetchSchedule ***failed***. Prefetch schedule violation. Results are NOT valid\n"){do { } while(0); }; |
| 2405 | } |
| 2406 | } |
| 2407 | |
| 2408 | mode_lib->vba.VStartupLines = mode_lib->vba.VStartupLines + 1; |
| 2409 | } while (!((mode_lib->vba.PrefetchModeSupported |
| 2410 | && ((!mode_lib->vba.ImmediateFlipSupport && !mode_lib->vba.HostVMEnable) |
| 2411 | || mode_lib->vba.ImmediateFlipSupported)) |
| 2412 | || locals->MaximumMaxVStartupLines < mode_lib->vba.VStartupLines)); |
| 2413 | |
| 2414 | //Watermarks and NB P-State/DRAM Clock Change Support |
| 2415 | { |
| 2416 | enum clock_change_support DRAMClockChangeSupport; // dummy |
| 2417 | CalculateWatermarksAndDRAMSpeedChangeSupport( |
| 2418 | mode_lib, |
| 2419 | mode_lib->vba.PrefetchMode[mode_lib->vba.VoltageLevel][mode_lib->vba.maxMpcComb], |
| 2420 | mode_lib->vba.NumberOfActivePlanes, |
| 2421 | mode_lib->vba.MaxLineBufferLines, |
| 2422 | mode_lib->vba.LineBufferSize, |
| 2423 | mode_lib->vba.DPPOutputBufferPixels, |
| 2424 | mode_lib->vba.DETBufferSizeInKByte[0], |
| 2425 | mode_lib->vba.WritebackInterfaceLumaBufferSize, |
| 2426 | mode_lib->vba.WritebackInterfaceChromaBufferSize, |
| 2427 | mode_lib->vba.DCFCLK, |
| 2428 | mode_lib->vba.UrgentOutOfOrderReturnPerChannel * mode_lib->vba.NumberOfChannels, |
| 2429 | mode_lib->vba.ReturnBW, |
| 2430 | mode_lib->vba.GPUVMEnable, |
| 2431 | locals->dpte_group_bytes, |
| 2432 | mode_lib->vba.MetaChunkSize, |
| 2433 | mode_lib->vba.UrgentLatency, |
| 2434 | mode_lib->vba.UrgentExtraLatency, |
| 2435 | mode_lib->vba.WritebackLatency, |
| 2436 | mode_lib->vba.WritebackChunkSize, |
| 2437 | mode_lib->vba.SOCCLK, |
| 2438 | mode_lib->vba.DRAMClockChangeLatency, |
| 2439 | mode_lib->vba.SRExitTime, |
| 2440 | mode_lib->vba.SREnterPlusExitTime, |
| 2441 | mode_lib->vba.DCFCLKDeepSleep, |
| 2442 | mode_lib->vba.DPPPerPlane, |
| 2443 | mode_lib->vba.DCCEnable, |
| 2444 | locals->DPPCLK, |
| 2445 | locals->SwathWidthSingleDPPY, |
| 2446 | mode_lib->vba.SwathHeightY, |
| 2447 | locals->ReadBandwidthPlaneLuma, |
| 2448 | mode_lib->vba.SwathHeightC, |
| 2449 | locals->ReadBandwidthPlaneChroma, |
| 2450 | mode_lib->vba.LBBitPerPixel, |
| 2451 | locals->SwathWidthY, |
| 2452 | mode_lib->vba.HRatio, |
| 2453 | mode_lib->vba.vtaps, |
| 2454 | mode_lib->vba.VTAPsChroma, |
| 2455 | mode_lib->vba.VRatio, |
| 2456 | mode_lib->vba.HTotal, |
| 2457 | mode_lib->vba.PixelClock, |
| 2458 | mode_lib->vba.BlendingAndTiming, |
| 2459 | locals->BytePerPixelDETY, |
| 2460 | locals->BytePerPixelDETC, |
| 2461 | mode_lib->vba.WritebackEnable, |
| 2462 | mode_lib->vba.WritebackPixelFormat, |
| 2463 | mode_lib->vba.WritebackDestinationWidth, |
| 2464 | mode_lib->vba.WritebackDestinationHeight, |
| 2465 | mode_lib->vba.WritebackSourceHeight, |
| 2466 | &DRAMClockChangeSupport, |
| 2467 | &mode_lib->vba.UrgentWatermark, |
| 2468 | &mode_lib->vba.WritebackUrgentWatermark, |
| 2469 | &mode_lib->vba.DRAMClockChangeWatermark, |
| 2470 | &mode_lib->vba.WritebackDRAMClockChangeWatermark, |
| 2471 | &mode_lib->vba.StutterExitWatermark, |
| 2472 | &mode_lib->vba.StutterEnterPlusExitWatermark, |
| 2473 | &mode_lib->vba.MinActiveDRAMClockChangeLatencySupported); |
| 2474 | } |
| 2475 | |
| 2476 | |
| 2477 | //Display Pipeline Delivery Time in Prefetch, Groups |
| 2478 | CalculatePixelDeliveryTimes( |
| 2479 | mode_lib->vba.NumberOfActivePlanes, |
| 2480 | mode_lib->vba.VRatio, |
| 2481 | locals->VRatioPrefetchY, |
| 2482 | locals->VRatioPrefetchC, |
| 2483 | locals->swath_width_luma_ub, |
| 2484 | locals->swath_width_chroma_ub, |
| 2485 | mode_lib->vba.DPPPerPlane, |
| 2486 | mode_lib->vba.HRatio, |
| 2487 | mode_lib->vba.PixelClock, |
| 2488 | locals->PSCL_THROUGHPUT_LUMA, |
| 2489 | locals->PSCL_THROUGHPUT_CHROMA, |
| 2490 | locals->DPPCLK, |
| 2491 | locals->BytePerPixelDETC, |
| 2492 | mode_lib->vba.SourceScan, |
| 2493 | locals->BlockWidth256BytesY, |
| 2494 | locals->BlockHeight256BytesY, |
| 2495 | locals->BlockWidth256BytesC, |
| 2496 | locals->BlockHeight256BytesC, |
| 2497 | locals->DisplayPipeLineDeliveryTimeLuma, |
| 2498 | locals->DisplayPipeLineDeliveryTimeChroma, |
| 2499 | locals->DisplayPipeLineDeliveryTimeLumaPrefetch, |
| 2500 | locals->DisplayPipeLineDeliveryTimeChromaPrefetch, |
| 2501 | locals->DisplayPipeRequestDeliveryTimeLuma, |
| 2502 | locals->DisplayPipeRequestDeliveryTimeChroma, |
| 2503 | locals->DisplayPipeRequestDeliveryTimeLumaPrefetch, |
| 2504 | locals->DisplayPipeRequestDeliveryTimeChromaPrefetch); |
| 2505 | |
| 2506 | CalculateMetaAndPTETimes( |
| 2507 | mode_lib->vba.NumberOfActivePlanes, |
| 2508 | mode_lib->vba.GPUVMEnable, |
| 2509 | mode_lib->vba.MetaChunkSize, |
| 2510 | mode_lib->vba.MinMetaChunkSizeBytes, |
| 2511 | mode_lib->vba.GPUVMMaxPageTableLevels, |
| 2512 | mode_lib->vba.HTotal, |
| 2513 | mode_lib->vba.VRatio, |
| 2514 | locals->VRatioPrefetchY, |
| 2515 | locals->VRatioPrefetchC, |
| 2516 | locals->DestinationLinesToRequestRowInVBlank, |
| 2517 | locals->DestinationLinesToRequestRowInImmediateFlip, |
| 2518 | locals->DestinationLinesToRequestVMInVBlank, |
| 2519 | locals->DestinationLinesToRequestVMInImmediateFlip, |
| 2520 | mode_lib->vba.DCCEnable, |
| 2521 | mode_lib->vba.PixelClock, |
| 2522 | locals->BytePerPixelDETY, |
| 2523 | locals->BytePerPixelDETC, |
| 2524 | mode_lib->vba.SourceScan, |
| 2525 | locals->dpte_row_height, |
| 2526 | locals->dpte_row_height_chroma, |
| 2527 | locals->meta_row_width, |
| 2528 | locals->meta_row_height, |
| 2529 | locals->meta_req_width, |
| 2530 | locals->meta_req_height, |
| 2531 | locals->dpte_group_bytes, |
| 2532 | locals->PTERequestSizeY, |
| 2533 | locals->PTERequestSizeC, |
| 2534 | locals->PixelPTEReqWidthY, |
| 2535 | locals->PixelPTEReqHeightY, |
| 2536 | locals->PixelPTEReqWidthC, |
| 2537 | locals->PixelPTEReqHeightC, |
| 2538 | locals->dpte_row_width_luma_ub, |
| 2539 | locals->dpte_row_width_chroma_ub, |
| 2540 | locals->vm_group_bytes, |
| 2541 | locals->dpde0_bytes_per_frame_ub_l, |
| 2542 | locals->dpde0_bytes_per_frame_ub_c, |
| 2543 | locals->meta_pte_bytes_per_frame_ub_l, |
| 2544 | locals->meta_pte_bytes_per_frame_ub_c, |
| 2545 | locals->DST_Y_PER_PTE_ROW_NOM_L, |
| 2546 | locals->DST_Y_PER_PTE_ROW_NOM_C, |
| 2547 | locals->DST_Y_PER_META_ROW_NOM_L, |
| 2548 | locals->TimePerMetaChunkNominal, |
| 2549 | locals->TimePerMetaChunkVBlank, |
| 2550 | locals->TimePerMetaChunkFlip, |
| 2551 | locals->time_per_pte_group_nom_luma, |
| 2552 | locals->time_per_pte_group_vblank_luma, |
| 2553 | locals->time_per_pte_group_flip_luma, |
| 2554 | locals->time_per_pte_group_nom_chroma, |
| 2555 | locals->time_per_pte_group_vblank_chroma, |
| 2556 | locals->time_per_pte_group_flip_chroma, |
| 2557 | locals->TimePerVMGroupVBlank, |
| 2558 | locals->TimePerVMGroupFlip, |
| 2559 | locals->TimePerVMRequestVBlank, |
| 2560 | locals->TimePerVMRequestFlip); |
| 2561 | |
| 2562 | |
| 2563 | // Min TTUVBlank |
| 2564 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 2565 | if (mode_lib->vba.PrefetchMode[mode_lib->vba.VoltageLevel][mode_lib->vba.maxMpcComb] == 0) { |
| 2566 | locals->AllowDRAMClockChangeDuringVBlank[k] = true1; |
| 2567 | locals->AllowDRAMSelfRefreshDuringVBlank[k] = true1; |
| 2568 | locals->MinTTUVBlank[k] = dml_max( |
| 2569 | mode_lib->vba.DRAMClockChangeWatermark, |
| 2570 | dml_max( |
| 2571 | mode_lib->vba.StutterEnterPlusExitWatermark, |
| 2572 | mode_lib->vba.UrgentWatermark)); |
| 2573 | } else if (mode_lib->vba.PrefetchMode[mode_lib->vba.VoltageLevel][mode_lib->vba.maxMpcComb] == 1) { |
| 2574 | locals->AllowDRAMClockChangeDuringVBlank[k] = false0; |
| 2575 | locals->AllowDRAMSelfRefreshDuringVBlank[k] = true1; |
| 2576 | locals->MinTTUVBlank[k] = dml_max( |
| 2577 | mode_lib->vba.StutterEnterPlusExitWatermark, |
| 2578 | mode_lib->vba.UrgentWatermark); |
| 2579 | } else { |
| 2580 | locals->AllowDRAMClockChangeDuringVBlank[k] = false0; |
| 2581 | locals->AllowDRAMSelfRefreshDuringVBlank[k] = false0; |
| 2582 | locals->MinTTUVBlank[k] = mode_lib->vba.UrgentWatermark; |
| 2583 | } |
| 2584 | if (!mode_lib->vba.DynamicMetadataEnable[k]) |
| 2585 | locals->MinTTUVBlank[k] = mode_lib->vba.TCalc |
| 2586 | + locals->MinTTUVBlank[k]; |
| 2587 | } |
| 2588 | |
| 2589 | // DCC Configuration |
| 2590 | mode_lib->vba.ActiveDPPs = 0; |
| 2591 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 2592 | locals->MaximumDCCCompressionYSurface[k] = CalculateDCCConfiguration( |
| 2593 | mode_lib->vba.DCCEnable[k], |
| 2594 | false0, // We should always know the direction DCCProgrammingAssumesScanDirectionUnknown, |
| 2595 | mode_lib->vba.ViewportWidth[k], |
| 2596 | mode_lib->vba.ViewportHeight[k], |
| 2597 | mode_lib->vba.DETBufferSizeInKByte[0] * 1024, |
| 2598 | locals->BlockHeight256BytesY[k], |
| 2599 | mode_lib->vba.SwathHeightY[k], |
| 2600 | mode_lib->vba.SurfaceTiling[k], |
| 2601 | locals->BytePerPixelDETY[k], |
| 2602 | mode_lib->vba.SourceScan[k], |
| 2603 | &locals->DCCYMaxUncompressedBlock[k], |
| 2604 | &locals->DCCYMaxCompressedBlock[k], |
| 2605 | &locals->DCCYIndependent64ByteBlock[k]); |
| 2606 | } |
| 2607 | |
| 2608 | //XFC Parameters: |
| 2609 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 2610 | if (mode_lib->vba.XFCEnabled[k] == true1) { |
| 2611 | double TWait; |
| 2612 | |
| 2613 | locals->XFCSlaveVUpdateOffset[k] = mode_lib->vba.XFCTSlvVupdateOffset; |
| 2614 | locals->XFCSlaveVupdateWidth[k] = mode_lib->vba.XFCTSlvVupdateWidth; |
| 2615 | locals->XFCSlaveVReadyOffset[k] = mode_lib->vba.XFCTSlvVreadyOffset; |
| 2616 | TWait = CalculateTWait( |
| 2617 | mode_lib->vba.PrefetchMode[mode_lib->vba.VoltageLevel][mode_lib->vba.maxMpcComb], |
| 2618 | mode_lib->vba.DRAMClockChangeLatency, |
| 2619 | mode_lib->vba.UrgentLatency, |
| 2620 | mode_lib->vba.SREnterPlusExitTime); |
| 2621 | mode_lib->vba.XFCRemoteSurfaceFlipDelay = CalculateRemoteSurfaceFlipDelay( |
| 2622 | mode_lib, |
| 2623 | mode_lib->vba.VRatio[k], |
| 2624 | locals->SwathWidthY[k], |
| 2625 | dml_ceil(locals->BytePerPixelDETY[k], 1), |
| 2626 | mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k], |
| 2627 | mode_lib->vba.XFCTSlvVupdateOffset, |
| 2628 | mode_lib->vba.XFCTSlvVupdateWidth, |
| 2629 | mode_lib->vba.XFCTSlvVreadyOffset, |
| 2630 | mode_lib->vba.XFCXBUFLatencyTolerance, |
| 2631 | mode_lib->vba.XFCFillBWOverhead, |
| 2632 | mode_lib->vba.XFCSlvChunkSize, |
| 2633 | mode_lib->vba.XFCBusTransportTime, |
| 2634 | mode_lib->vba.TCalc, |
| 2635 | TWait, |
| 2636 | &mode_lib->vba.SrcActiveDrainRate, |
| 2637 | &mode_lib->vba.TInitXFill, |
| 2638 | &mode_lib->vba.TslvChk); |
| 2639 | locals->XFCRemoteSurfaceFlipLatency[k] = |
| 2640 | dml_floor( |
| 2641 | mode_lib->vba.XFCRemoteSurfaceFlipDelay |
| 2642 | / (mode_lib->vba.HTotal[k] |
| 2643 | / mode_lib->vba.PixelClock[k]), |
| 2644 | 1); |
| 2645 | locals->XFCTransferDelay[k] = |
| 2646 | dml_ceil( |
| 2647 | mode_lib->vba.XFCBusTransportTime |
| 2648 | / (mode_lib->vba.HTotal[k] |
| 2649 | / mode_lib->vba.PixelClock[k]), |
| 2650 | 1); |
| 2651 | locals->XFCPrechargeDelay[k] = |
| 2652 | dml_ceil( |
| 2653 | (mode_lib->vba.XFCBusTransportTime |
| 2654 | + mode_lib->vba.TInitXFill |
| 2655 | + mode_lib->vba.TslvChk) |
| 2656 | / (mode_lib->vba.HTotal[k] |
| 2657 | / mode_lib->vba.PixelClock[k]), |
| 2658 | 1); |
| 2659 | mode_lib->vba.InitFillLevel = mode_lib->vba.XFCXBUFLatencyTolerance |
| 2660 | * mode_lib->vba.SrcActiveDrainRate; |
| 2661 | mode_lib->vba.FinalFillMargin = |
| 2662 | (locals->DestinationLinesToRequestVMInVBlank[k] |
| 2663 | + locals->DestinationLinesToRequestRowInVBlank[k]) |
| 2664 | * mode_lib->vba.HTotal[k] |
| 2665 | / mode_lib->vba.PixelClock[k] |
| 2666 | * mode_lib->vba.SrcActiveDrainRate |
| 2667 | + mode_lib->vba.XFCFillConstant; |
| 2668 | mode_lib->vba.FinalFillLevel = mode_lib->vba.XFCRemoteSurfaceFlipDelay |
| 2669 | * mode_lib->vba.SrcActiveDrainRate |
| 2670 | + mode_lib->vba.FinalFillMargin; |
| 2671 | mode_lib->vba.RemainingFillLevel = dml_max( |
| 2672 | 0.0, |
| 2673 | mode_lib->vba.FinalFillLevel - mode_lib->vba.InitFillLevel); |
| 2674 | mode_lib->vba.TFinalxFill = mode_lib->vba.RemainingFillLevel |
| 2675 | / (mode_lib->vba.SrcActiveDrainRate |
| 2676 | * mode_lib->vba.XFCFillBWOverhead / 100); |
| 2677 | locals->XFCPrefetchMargin[k] = |
| 2678 | mode_lib->vba.XFCRemoteSurfaceFlipDelay |
| 2679 | + mode_lib->vba.TFinalxFill |
| 2680 | + (locals->DestinationLinesToRequestVMInVBlank[k] |
| 2681 | + locals->DestinationLinesToRequestRowInVBlank[k]) |
| 2682 | * mode_lib->vba.HTotal[k] |
| 2683 | / mode_lib->vba.PixelClock[k]; |
| 2684 | } else { |
| 2685 | locals->XFCSlaveVUpdateOffset[k] = 0; |
| 2686 | locals->XFCSlaveVupdateWidth[k] = 0; |
| 2687 | locals->XFCSlaveVReadyOffset[k] = 0; |
| 2688 | locals->XFCRemoteSurfaceFlipLatency[k] = 0; |
| 2689 | locals->XFCPrechargeDelay[k] = 0; |
| 2690 | locals->XFCTransferDelay[k] = 0; |
| 2691 | locals->XFCPrefetchMargin[k] = 0; |
| 2692 | } |
| 2693 | } |
| 2694 | |
| 2695 | // Stutter Efficiency |
| 2696 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 2697 | CalculateDETBufferSize( |
| 2698 | mode_lib->vba.DETBufferSizeInKByte[0], |
| 2699 | mode_lib->vba.SwathHeightY[k], |
| 2700 | mode_lib->vba.SwathHeightC[k], |
| 2701 | &locals->DETBufferSizeY[k], |
| 2702 | &locals->DETBufferSizeC[k]); |
| 2703 | |
| 2704 | locals->LinesInDETY[k] = (double)locals->DETBufferSizeY[k] |
| 2705 | / locals->BytePerPixelDETY[k] / locals->SwathWidthY[k]; |
| 2706 | locals->LinesInDETYRoundedDownToSwath[k] = dml_floor( |
| 2707 | locals->LinesInDETY[k], |
| 2708 | mode_lib->vba.SwathHeightY[k]); |
| 2709 | locals->FullDETBufferingTimeY[k] = |
| 2710 | locals->LinesInDETYRoundedDownToSwath[k] |
| 2711 | * (mode_lib->vba.HTotal[k] |
| 2712 | / mode_lib->vba.PixelClock[k]) |
| 2713 | / mode_lib->vba.VRatio[k]; |
| 2714 | } |
| 2715 | |
| 2716 | mode_lib->vba.StutterPeriod = 999999.0; |
| 2717 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 2718 | if (locals->FullDETBufferingTimeY[k] < mode_lib->vba.StutterPeriod) { |
| 2719 | mode_lib->vba.StutterPeriod = locals->FullDETBufferingTimeY[k]; |
| 2720 | mode_lib->vba.FrameTimeForMinFullDETBufferingTime = |
| 2721 | (double) mode_lib->vba.VTotal[k] * mode_lib->vba.HTotal[k] |
| 2722 | / mode_lib->vba.PixelClock[k]; |
| 2723 | locals->BytePerPixelYCriticalPlane = dml_ceil(locals->BytePerPixelDETY[k], 1); |
| 2724 | locals->SwathWidthYCriticalPlane = locals->SwathWidthY[k]; |
| 2725 | locals->LinesToFinishSwathTransferStutterCriticalPlane = |
| 2726 | mode_lib->vba.SwathHeightY[k] - (locals->LinesInDETY[k] - locals->LinesInDETYRoundedDownToSwath[k]); |
| 2727 | } |
| 2728 | } |
| 2729 | |
| 2730 | mode_lib->vba.AverageReadBandwidth = 0.0; |
| 2731 | mode_lib->vba.TotalRowReadBandwidth = 0.0; |
| 2732 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 2733 | unsigned int DCCRateLimit; |
| 2734 | |
| 2735 | if (mode_lib->vba.DCCEnable[k]) { |
| 2736 | if (locals->DCCYMaxCompressedBlock[k] == 256) |
| 2737 | DCCRateLimit = 4; |
| 2738 | else |
| 2739 | DCCRateLimit = 2; |
| 2740 | |
| 2741 | mode_lib->vba.AverageReadBandwidth = |
| 2742 | mode_lib->vba.AverageReadBandwidth |
| 2743 | + (locals->ReadBandwidthPlaneLuma[k] + locals->ReadBandwidthPlaneChroma[k]) / |
| 2744 | dml_min(mode_lib->vba.DCCRate[k], DCCRateLimit); |
| 2745 | } else { |
| 2746 | mode_lib->vba.AverageReadBandwidth = |
| 2747 | mode_lib->vba.AverageReadBandwidth |
| 2748 | + locals->ReadBandwidthPlaneLuma[k] |
| 2749 | + locals->ReadBandwidthPlaneChroma[k]; |
| 2750 | } |
| 2751 | mode_lib->vba.TotalRowReadBandwidth = mode_lib->vba.TotalRowReadBandwidth + |
| 2752 | locals->meta_row_bw[k] + locals->dpte_row_bw[k]; |
| 2753 | } |
| 2754 | |
| 2755 | mode_lib->vba.AverageDCCCompressionRate = mode_lib->vba.TotalDataReadBandwidth / mode_lib->vba.AverageReadBandwidth; |
| 2756 | |
| 2757 | mode_lib->vba.PartOfBurstThatFitsInROB = |
| 2758 | dml_min( |
| 2759 | mode_lib->vba.StutterPeriod |
| 2760 | * mode_lib->vba.TotalDataReadBandwidth, |
| 2761 | mode_lib->vba.ROBBufferSizeInKByte * 1024 |
| 2762 | * mode_lib->vba.AverageDCCCompressionRate); |
| 2763 | mode_lib->vba.StutterBurstTime = mode_lib->vba.PartOfBurstThatFitsInROB |
| 2764 | / mode_lib->vba.AverageDCCCompressionRate / mode_lib->vba.ReturnBW |
| 2765 | + (mode_lib->vba.StutterPeriod * mode_lib->vba.TotalDataReadBandwidth |
| 2766 | - mode_lib->vba.PartOfBurstThatFitsInROB) |
| 2767 | / (mode_lib->vba.DCFCLK * 64) |
| 2768 | + mode_lib->vba.StutterPeriod * mode_lib->vba.TotalRowReadBandwidth / mode_lib->vba.ReturnBW; |
| 2769 | mode_lib->vba.StutterBurstTime = dml_max( |
| 2770 | mode_lib->vba.StutterBurstTime, |
| 2771 | (locals->LinesToFinishSwathTransferStutterCriticalPlane * locals->BytePerPixelYCriticalPlane * |
| 2772 | locals->SwathWidthYCriticalPlane / mode_lib->vba.ReturnBW) |
| 2773 | ); |
| 2774 | |
| 2775 | mode_lib->vba.TotalActiveWriteback = 0; |
| 2776 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 2777 | if (mode_lib->vba.WritebackEnable[k] == true1) { |
| 2778 | mode_lib->vba.TotalActiveWriteback = mode_lib->vba.TotalActiveWriteback + 1; |
| 2779 | } |
| 2780 | } |
| 2781 | |
| 2782 | if (mode_lib->vba.TotalActiveWriteback == 0) { |
| 2783 | mode_lib->vba.StutterEfficiencyNotIncludingVBlank = (1 |
| 2784 | - (mode_lib->vba.SRExitTime + mode_lib->vba.StutterBurstTime) |
| 2785 | / mode_lib->vba.StutterPeriod) * 100; |
| 2786 | } else { |
| 2787 | mode_lib->vba.StutterEfficiencyNotIncludingVBlank = 0; |
| 2788 | } |
| 2789 | |
| 2790 | mode_lib->vba.SmallestVBlank = 999999; |
| 2791 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 2792 | if (mode_lib->vba.SynchronizedVBlank || mode_lib->vba.NumberOfActivePlanes == 1) { |
| 2793 | mode_lib->vba.VBlankTime = (double) (mode_lib->vba.VTotal[k] |
| 2794 | - mode_lib->vba.VActive[k]) * mode_lib->vba.HTotal[k] |
| 2795 | / mode_lib->vba.PixelClock[k]; |
| 2796 | } else { |
| 2797 | mode_lib->vba.VBlankTime = 0; |
| 2798 | } |
| 2799 | mode_lib->vba.SmallestVBlank = dml_min( |
| 2800 | mode_lib->vba.SmallestVBlank, |
| 2801 | mode_lib->vba.VBlankTime); |
| 2802 | } |
| 2803 | |
| 2804 | mode_lib->vba.StutterEfficiency = (mode_lib->vba.StutterEfficiencyNotIncludingVBlank / 100 |
| 2805 | * (mode_lib->vba.FrameTimeForMinFullDETBufferingTime |
| 2806 | - mode_lib->vba.SmallestVBlank) |
| 2807 | + mode_lib->vba.SmallestVBlank) |
| 2808 | / mode_lib->vba.FrameTimeForMinFullDETBufferingTime * 100; |
| 2809 | } |
| 2810 | |
| 2811 | static void DisplayPipeConfiguration(struct display_mode_lib *mode_lib) |
| 2812 | { |
| 2813 | // Display Pipe Configuration |
| 2814 | double BytePerPixDETY; |
| 2815 | double BytePerPixDETC; |
| 2816 | double Read256BytesBlockHeightY; |
| 2817 | double Read256BytesBlockHeightC; |
| 2818 | double Read256BytesBlockWidthY; |
| 2819 | double Read256BytesBlockWidthC; |
| 2820 | double MaximumSwathHeightY; |
| 2821 | double MaximumSwathHeightC; |
| 2822 | double MinimumSwathHeightY; |
| 2823 | double MinimumSwathHeightC; |
| 2824 | double SwathWidth; |
| 2825 | double SwathWidthGranularityY; |
| 2826 | double SwathWidthGranularityC; |
| 2827 | double RoundedUpMaxSwathSizeBytesY; |
| 2828 | double RoundedUpMaxSwathSizeBytesC; |
| 2829 | unsigned int j, k; |
| 2830 | |
| 2831 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 2832 | bool_Bool MainPlaneDoesODMCombine = false0; |
| 2833 | |
| 2834 | if (mode_lib->vba.SourcePixelFormat[k] == dm_444_64) { |
| 2835 | BytePerPixDETY = 8; |
| 2836 | BytePerPixDETC = 0; |
| 2837 | } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_32) { |
| 2838 | BytePerPixDETY = 4; |
| 2839 | BytePerPixDETC = 0; |
| 2840 | } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_16) { |
| 2841 | BytePerPixDETY = 2; |
| 2842 | BytePerPixDETC = 0; |
| 2843 | } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_8) { |
| 2844 | BytePerPixDETY = 1; |
| 2845 | BytePerPixDETC = 0; |
| 2846 | } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_8) { |
| 2847 | BytePerPixDETY = 1; |
| 2848 | BytePerPixDETC = 2; |
| 2849 | } else { |
| 2850 | BytePerPixDETY = 4.0 / 3.0; |
| 2851 | BytePerPixDETC = 8.0 / 3.0; |
| 2852 | } |
| 2853 | |
| 2854 | if ((mode_lib->vba.SourcePixelFormat[k] == dm_444_64 |
| 2855 | || mode_lib->vba.SourcePixelFormat[k] == dm_444_32 |
| 2856 | || mode_lib->vba.SourcePixelFormat[k] == dm_444_16 |
| 2857 | || mode_lib->vba.SourcePixelFormat[k] == dm_444_8)) { |
| 2858 | if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear) { |
| 2859 | Read256BytesBlockHeightY = 1; |
| 2860 | } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_64) { |
| 2861 | Read256BytesBlockHeightY = 4; |
| 2862 | } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_32 |
| 2863 | || mode_lib->vba.SourcePixelFormat[k] == dm_444_16) { |
| 2864 | Read256BytesBlockHeightY = 8; |
| 2865 | } else { |
| 2866 | Read256BytesBlockHeightY = 16; |
| 2867 | } |
| 2868 | Read256BytesBlockWidthY = 256 / dml_ceil(BytePerPixDETY, 1) |
| 2869 | / Read256BytesBlockHeightY; |
| 2870 | Read256BytesBlockHeightC = 0; |
| 2871 | Read256BytesBlockWidthC = 0; |
| 2872 | } else { |
| 2873 | if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear) { |
| 2874 | Read256BytesBlockHeightY = 1; |
| 2875 | Read256BytesBlockHeightC = 1; |
| 2876 | } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_8) { |
| 2877 | Read256BytesBlockHeightY = 16; |
| 2878 | Read256BytesBlockHeightC = 8; |
| 2879 | } else { |
| 2880 | Read256BytesBlockHeightY = 8; |
| 2881 | Read256BytesBlockHeightC = 8; |
| 2882 | } |
| 2883 | Read256BytesBlockWidthY = 256 / dml_ceil(BytePerPixDETY, 1) |
| 2884 | / Read256BytesBlockHeightY; |
| 2885 | Read256BytesBlockWidthC = 256 / dml_ceil(BytePerPixDETC, 2) |
| 2886 | / Read256BytesBlockHeightC; |
| 2887 | } |
| 2888 | |
| 2889 | if (mode_lib->vba.SourceScan[k] == dm_horz) { |
| 2890 | MaximumSwathHeightY = Read256BytesBlockHeightY; |
| 2891 | MaximumSwathHeightC = Read256BytesBlockHeightC; |
| 2892 | } else { |
| 2893 | MaximumSwathHeightY = Read256BytesBlockWidthY; |
| 2894 | MaximumSwathHeightC = Read256BytesBlockWidthC; |
| 2895 | } |
| 2896 | |
| 2897 | if ((mode_lib->vba.SourcePixelFormat[k] == dm_444_64 |
| 2898 | || mode_lib->vba.SourcePixelFormat[k] == dm_444_32 |
| 2899 | || mode_lib->vba.SourcePixelFormat[k] == dm_444_16 |
| 2900 | || mode_lib->vba.SourcePixelFormat[k] == dm_444_8)) { |
| 2901 | if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear |
| 2902 | || (mode_lib->vba.SourcePixelFormat[k] == dm_444_64 |
| 2903 | && (mode_lib->vba.SurfaceTiling[k] |
| 2904 | == dm_sw_4kb_s |
| 2905 | || mode_lib->vba.SurfaceTiling[k] |
| 2906 | == dm_sw_4kb_s_x |
| 2907 | || mode_lib->vba.SurfaceTiling[k] |
| 2908 | == dm_sw_64kb_s |
| 2909 | || mode_lib->vba.SurfaceTiling[k] |
| 2910 | == dm_sw_64kb_s_t |
| 2911 | || mode_lib->vba.SurfaceTiling[k] |
| 2912 | == dm_sw_64kb_s_x |
| 2913 | || mode_lib->vba.SurfaceTiling[k] |
| 2914 | == dm_sw_var_s |
| 2915 | || mode_lib->vba.SurfaceTiling[k] |
| 2916 | == dm_sw_var_s_x) |
| 2917 | && mode_lib->vba.SourceScan[k] == dm_horz)) { |
| 2918 | MinimumSwathHeightY = MaximumSwathHeightY; |
| 2919 | } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_8 |
| 2920 | && mode_lib->vba.SourceScan[k] != dm_horz) { |
| 2921 | MinimumSwathHeightY = MaximumSwathHeightY; |
| 2922 | } else { |
| 2923 | MinimumSwathHeightY = MaximumSwathHeightY / 2.0; |
| 2924 | } |
| 2925 | MinimumSwathHeightC = MaximumSwathHeightC; |
| 2926 | } else { |
| 2927 | if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear) { |
| 2928 | MinimumSwathHeightY = MaximumSwathHeightY; |
| 2929 | MinimumSwathHeightC = MaximumSwathHeightC; |
| 2930 | } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_8 |
| 2931 | && mode_lib->vba.SourceScan[k] == dm_horz) { |
| 2932 | MinimumSwathHeightY = MaximumSwathHeightY / 2.0; |
| 2933 | MinimumSwathHeightC = MaximumSwathHeightC; |
| 2934 | } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_10 |
| 2935 | && mode_lib->vba.SourceScan[k] == dm_horz) { |
| 2936 | MinimumSwathHeightC = MaximumSwathHeightC / 2.0; |
| 2937 | MinimumSwathHeightY = MaximumSwathHeightY; |
| 2938 | } else { |
| 2939 | MinimumSwathHeightY = MaximumSwathHeightY; |
| 2940 | MinimumSwathHeightC = MaximumSwathHeightC; |
| 2941 | } |
| 2942 | } |
| 2943 | |
| 2944 | if (mode_lib->vba.SourceScan[k] == dm_horz) { |
| 2945 | SwathWidth = mode_lib->vba.ViewportWidth[k]; |
| 2946 | } else { |
| 2947 | SwathWidth = mode_lib->vba.ViewportHeight[k]; |
| 2948 | } |
| 2949 | |
| 2950 | if (mode_lib->vba.ODMCombineEnabled[k] == dm_odm_combine_mode_2to1) { |
| 2951 | MainPlaneDoesODMCombine = true1; |
| 2952 | } |
| 2953 | for (j = 0; j < mode_lib->vba.NumberOfActivePlanes; ++j) { |
| 2954 | if (mode_lib->vba.BlendingAndTiming[k] == j |
| 2955 | && mode_lib->vba.ODMCombineEnabled[k] == dm_odm_combine_mode_2to1) { |
| 2956 | MainPlaneDoesODMCombine = true1; |
| 2957 | } |
| 2958 | } |
| 2959 | |
| 2960 | if (MainPlaneDoesODMCombine == true1) { |
| 2961 | SwathWidth = dml_min( |
| 2962 | SwathWidth, |
| 2963 | mode_lib->vba.HActive[k] / 2.0 * mode_lib->vba.HRatio[k]); |
| 2964 | } else { |
| 2965 | SwathWidth = SwathWidth / mode_lib->vba.DPPPerPlane[k]; |
| 2966 | } |
| 2967 | |
| 2968 | SwathWidthGranularityY = 256 / dml_ceil(BytePerPixDETY, 1) / MaximumSwathHeightY; |
| 2969 | RoundedUpMaxSwathSizeBytesY = (dml_ceil( |
| 2970 | (double) (SwathWidth - 1), |
| 2971 | SwathWidthGranularityY) + SwathWidthGranularityY) * BytePerPixDETY |
| 2972 | * MaximumSwathHeightY; |
| 2973 | if (mode_lib->vba.SourcePixelFormat[k] == dm_420_10) { |
| 2974 | RoundedUpMaxSwathSizeBytesY = dml_ceil(RoundedUpMaxSwathSizeBytesY, 256) |
| 2975 | + 256; |
| 2976 | } |
| 2977 | if (MaximumSwathHeightC > 0) { |
| 2978 | SwathWidthGranularityC = 256.0 / dml_ceil(BytePerPixDETC, 2) |
| 2979 | / MaximumSwathHeightC; |
| 2980 | RoundedUpMaxSwathSizeBytesC = (dml_ceil( |
| 2981 | (double) (SwathWidth / 2.0 - 1), |
| 2982 | SwathWidthGranularityC) + SwathWidthGranularityC) |
| 2983 | * BytePerPixDETC * MaximumSwathHeightC; |
| 2984 | if (mode_lib->vba.SourcePixelFormat[k] == dm_420_10) { |
| 2985 | RoundedUpMaxSwathSizeBytesC = dml_ceil( |
| 2986 | RoundedUpMaxSwathSizeBytesC, |
| 2987 | 256) + 256; |
| 2988 | } |
| 2989 | } else |
| 2990 | RoundedUpMaxSwathSizeBytesC = 0.0; |
| 2991 | |
| 2992 | if (RoundedUpMaxSwathSizeBytesY + RoundedUpMaxSwathSizeBytesC |
| 2993 | <= mode_lib->vba.DETBufferSizeInKByte[0] * 1024.0 / 2.0) { |
| 2994 | mode_lib->vba.SwathHeightY[k] = MaximumSwathHeightY; |
| 2995 | mode_lib->vba.SwathHeightC[k] = MaximumSwathHeightC; |
| 2996 | } else { |
| 2997 | mode_lib->vba.SwathHeightY[k] = MinimumSwathHeightY; |
| 2998 | mode_lib->vba.SwathHeightC[k] = MinimumSwathHeightC; |
| 2999 | } |
| 3000 | |
| 3001 | CalculateDETBufferSize( |
| 3002 | mode_lib->vba.DETBufferSizeInKByte[0], |
| 3003 | mode_lib->vba.SwathHeightY[k], |
| 3004 | mode_lib->vba.SwathHeightC[k], |
| 3005 | &mode_lib->vba.DETBufferSizeY[k], |
| 3006 | &mode_lib->vba.DETBufferSizeC[k]); |
| 3007 | } |
| 3008 | } |
| 3009 | |
| 3010 | static double CalculateTWait( |
| 3011 | unsigned int PrefetchMode, |
| 3012 | double DRAMClockChangeLatency, |
| 3013 | double UrgentLatency, |
| 3014 | double SREnterPlusExitTime) |
| 3015 | { |
| 3016 | if (PrefetchMode == 0) { |
| 3017 | return dml_max( |
| 3018 | DRAMClockChangeLatency + UrgentLatency, |
| 3019 | dml_max(SREnterPlusExitTime, UrgentLatency)); |
| 3020 | } else if (PrefetchMode == 1) { |
| 3021 | return dml_max(SREnterPlusExitTime, UrgentLatency); |
| 3022 | } else { |
| 3023 | return UrgentLatency; |
| 3024 | } |
| 3025 | } |
| 3026 | |
| 3027 | static double CalculateRemoteSurfaceFlipDelay( |
| 3028 | struct display_mode_lib *mode_lib, |
| 3029 | double VRatio, |
| 3030 | double SwathWidth, |
| 3031 | double Bpp, |
| 3032 | double LineTime, |
| 3033 | double XFCTSlvVupdateOffset, |
| 3034 | double XFCTSlvVupdateWidth, |
| 3035 | double XFCTSlvVreadyOffset, |
| 3036 | double XFCXBUFLatencyTolerance, |
| 3037 | double XFCFillBWOverhead, |
| 3038 | double XFCSlvChunkSize, |
| 3039 | double XFCBusTransportTime, |
| 3040 | double TCalc, |
| 3041 | double TWait, |
| 3042 | double *SrcActiveDrainRate, |
| 3043 | double *TInitXFill, |
| 3044 | double *TslvChk) |
| 3045 | { |
| 3046 | double TSlvSetup, AvgfillRate, result; |
| 3047 | |
| 3048 | *SrcActiveDrainRate = VRatio * SwathWidth * Bpp / LineTime; |
| 3049 | TSlvSetup = XFCTSlvVupdateOffset + XFCTSlvVupdateWidth + XFCTSlvVreadyOffset; |
| 3050 | *TInitXFill = XFCXBUFLatencyTolerance / (1 + XFCFillBWOverhead / 100); |
| 3051 | AvgfillRate = *SrcActiveDrainRate * (1 + XFCFillBWOverhead / 100); |
| 3052 | *TslvChk = XFCSlvChunkSize / AvgfillRate; |
| 3053 | dml_print({do { } while(0); } |
| 3054 | "DML::CalculateRemoteSurfaceFlipDelay: SrcActiveDrainRate: %f\n",{do { } while(0); } |
| 3055 | *SrcActiveDrainRate){do { } while(0); }; |
| 3056 | dml_print("DML::CalculateRemoteSurfaceFlipDelay: TSlvSetup: %f\n", TSlvSetup){do { } while(0); }; |
| 3057 | dml_print("DML::CalculateRemoteSurfaceFlipDelay: TInitXFill: %f\n", *TInitXFill){do { } while(0); }; |
| 3058 | dml_print("DML::CalculateRemoteSurfaceFlipDelay: AvgfillRate: %f\n", AvgfillRate){do { } while(0); }; |
| 3059 | dml_print("DML::CalculateRemoteSurfaceFlipDelay: TslvChk: %f\n", *TslvChk){do { } while(0); }; |
| 3060 | result = 2 * XFCBusTransportTime + TSlvSetup + TCalc + TWait + *TslvChk + *TInitXFill; // TODO: This doesn't seem to match programming guide |
| 3061 | dml_print("DML::CalculateRemoteSurfaceFlipDelay: RemoteSurfaceFlipDelay: %f\n", result){do { } while(0); }; |
| 3062 | return result; |
| 3063 | } |
| 3064 | |
| 3065 | static double CalculateWriteBackDelay( |
| 3066 | enum source_format_class WritebackPixelFormat, |
| 3067 | double WritebackHRatio, |
| 3068 | double WritebackVRatio, |
| 3069 | unsigned int WritebackLumaHTaps, |
| 3070 | unsigned int WritebackLumaVTaps, |
| 3071 | unsigned int WritebackChromaHTaps, |
| 3072 | unsigned int WritebackChromaVTaps, |
| 3073 | unsigned int WritebackDestinationWidth) |
| 3074 | { |
| 3075 | double CalculateWriteBackDelay = |
| 3076 | dml_max( |
| 3077 | dml_ceil(WritebackLumaHTaps / 4.0, 1) / WritebackHRatio, |
| 3078 | WritebackLumaVTaps * dml_ceil(1.0 / WritebackVRatio, 1) |
| 3079 | * dml_ceil( |
| 3080 | WritebackDestinationWidth |
| 3081 | / 4.0, |
| 3082 | 1) |
| 3083 | + dml_ceil(1.0 / WritebackVRatio, 1) |
| 3084 | * (dml_ceil( |
| 3085 | WritebackLumaVTaps |
| 3086 | / 4.0, |
| 3087 | 1) + 4)); |
| 3088 | |
| 3089 | if (WritebackPixelFormat != dm_444_32) { |
| 3090 | CalculateWriteBackDelay = |
| 3091 | dml_max( |
| 3092 | CalculateWriteBackDelay, |
| 3093 | dml_max( |
| 3094 | dml_ceil( |
| 3095 | WritebackChromaHTaps |
| 3096 | / 2.0, |
| 3097 | 1) |
| 3098 | / (2 |
| 3099 | * WritebackHRatio), |
| 3100 | WritebackChromaVTaps |
| 3101 | * dml_ceil( |
| 3102 | 1 |
| 3103 | / (2 |
| 3104 | * WritebackVRatio), |
| 3105 | 1) |
| 3106 | * dml_ceil( |
| 3107 | WritebackDestinationWidth |
| 3108 | / 2.0 |
| 3109 | / 2.0, |
| 3110 | 1) |
| 3111 | + dml_ceil( |
| 3112 | 1 |
| 3113 | / (2 |
| 3114 | * WritebackVRatio), |
| 3115 | 1) |
| 3116 | * (dml_ceil( |
| 3117 | WritebackChromaVTaps |
| 3118 | / 4.0, |
| 3119 | 1) |
| 3120 | + 4))); |
| 3121 | } |
| 3122 | return CalculateWriteBackDelay; |
| 3123 | } |
| 3124 | |
| 3125 | static void CalculateActiveRowBandwidth( |
| 3126 | bool_Bool GPUVMEnable, |
| 3127 | enum source_format_class SourcePixelFormat, |
| 3128 | double VRatio, |
| 3129 | bool_Bool DCCEnable, |
| 3130 | double LineTime, |
| 3131 | unsigned int MetaRowByteLuma, |
| 3132 | unsigned int MetaRowByteChroma, |
| 3133 | unsigned int meta_row_height_luma, |
| 3134 | unsigned int meta_row_height_chroma, |
| 3135 | unsigned int PixelPTEBytesPerRowLuma, |
| 3136 | unsigned int PixelPTEBytesPerRowChroma, |
| 3137 | unsigned int dpte_row_height_luma, |
| 3138 | unsigned int dpte_row_height_chroma, |
| 3139 | double *meta_row_bw, |
| 3140 | double *dpte_row_bw) |
| 3141 | { |
| 3142 | if (DCCEnable != true1) { |
| 3143 | *meta_row_bw = 0; |
| 3144 | } else if (SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10) { |
| 3145 | *meta_row_bw = VRatio * MetaRowByteLuma / (meta_row_height_luma * LineTime) |
| 3146 | + VRatio / 2 * MetaRowByteChroma |
| 3147 | / (meta_row_height_chroma * LineTime); |
| 3148 | } else { |
| 3149 | *meta_row_bw = VRatio * MetaRowByteLuma / (meta_row_height_luma * LineTime); |
| 3150 | } |
| 3151 | |
| 3152 | if (GPUVMEnable != true1) { |
| 3153 | *dpte_row_bw = 0; |
| 3154 | } else if (SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10) { |
| 3155 | *dpte_row_bw = VRatio * PixelPTEBytesPerRowLuma / (dpte_row_height_luma * LineTime) |
| 3156 | + VRatio / 2 * PixelPTEBytesPerRowChroma |
| 3157 | / (dpte_row_height_chroma * LineTime); |
| 3158 | } else { |
| 3159 | *dpte_row_bw = VRatio * PixelPTEBytesPerRowLuma / (dpte_row_height_luma * LineTime); |
| 3160 | } |
| 3161 | } |
| 3162 | |
| 3163 | static void CalculateFlipSchedule( |
| 3164 | struct display_mode_lib *mode_lib, |
| 3165 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
| 3166 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
| 3167 | double UrgentExtraLatency, |
| 3168 | double UrgentLatency, |
| 3169 | unsigned int GPUVMMaxPageTableLevels, |
| 3170 | bool_Bool HostVMEnable, |
| 3171 | unsigned int HostVMMaxPageTableLevels, |
| 3172 | unsigned int HostVMCachedPageTableLevels, |
| 3173 | bool_Bool GPUVMEnable, |
| 3174 | double PDEAndMetaPTEBytesPerFrame, |
| 3175 | double MetaRowBytes, |
| 3176 | double DPTEBytesPerRow, |
| 3177 | double BandwidthAvailableForImmediateFlip, |
| 3178 | unsigned int TotImmediateFlipBytes, |
| 3179 | enum source_format_class SourcePixelFormat, |
| 3180 | double LineTime, |
| 3181 | double VRatio, |
| 3182 | double Tno_bw, |
| 3183 | bool_Bool DCCEnable, |
| 3184 | unsigned int dpte_row_height, |
| 3185 | unsigned int meta_row_height, |
| 3186 | unsigned int dpte_row_height_chroma, |
| 3187 | unsigned int meta_row_height_chroma, |
| 3188 | double *DestinationLinesToRequestVMInImmediateFlip, |
| 3189 | double *DestinationLinesToRequestRowInImmediateFlip, |
| 3190 | double *final_flip_bw, |
| 3191 | bool_Bool *ImmediateFlipSupportedForPipe) |
| 3192 | { |
| 3193 | double min_row_time = 0.0; |
| 3194 | unsigned int HostVMDynamicLevels; |
| 3195 | double TimeForFetchingMetaPTEImmediateFlip; |
| 3196 | double TimeForFetchingRowInVBlankImmediateFlip; |
| 3197 | double ImmediateFlipBW; |
| 3198 | double HostVMInefficiencyFactor; |
| 3199 | double VRatioClamped; |
| 3200 | |
| 3201 | if (GPUVMEnable == true1 && HostVMEnable == true1) { |
| 3202 | HostVMInefficiencyFactor = |
| 3203 | PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData |
| 3204 | / PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly; |
| 3205 | HostVMDynamicLevels = HostVMMaxPageTableLevels - HostVMCachedPageTableLevels; |
| 3206 | } else { |
| 3207 | HostVMInefficiencyFactor = 1; |
| 3208 | HostVMDynamicLevels = 0; |
| 3209 | } |
| 3210 | |
| 3211 | ImmediateFlipBW = (PDEAndMetaPTEBytesPerFrame + MetaRowBytes + DPTEBytesPerRow) |
| 3212 | * BandwidthAvailableForImmediateFlip / TotImmediateFlipBytes; |
| 3213 | |
| 3214 | if (GPUVMEnable == true1) { |
| 3215 | TimeForFetchingMetaPTEImmediateFlip = dml_max3( |
| 3216 | Tno_bw + PDEAndMetaPTEBytesPerFrame * HostVMInefficiencyFactor / ImmediateFlipBW, |
| 3217 | UrgentExtraLatency + UrgentLatency * (GPUVMMaxPageTableLevels * (HostVMDynamicLevels + 1) - 1), |
| 3218 | LineTime / 4.0); |
| 3219 | } else { |
| 3220 | TimeForFetchingMetaPTEImmediateFlip = 0; |
| 3221 | } |
| 3222 | |
| 3223 | *DestinationLinesToRequestVMInImmediateFlip = dml_ceil(4.0 * (TimeForFetchingMetaPTEImmediateFlip / LineTime), 1) / 4.0; |
| 3224 | if ((GPUVMEnable == true1 || DCCEnable == true1)) { |
| 3225 | TimeForFetchingRowInVBlankImmediateFlip = dml_max3((MetaRowBytes + DPTEBytesPerRow) * HostVMInefficiencyFactor / ImmediateFlipBW, UrgentLatency * (HostVMDynamicLevels + 1), LineTime / 4); |
| 3226 | } else { |
| 3227 | TimeForFetchingRowInVBlankImmediateFlip = 0; |
| 3228 | } |
| 3229 | |
| 3230 | *DestinationLinesToRequestRowInImmediateFlip = dml_ceil(4.0 * (TimeForFetchingRowInVBlankImmediateFlip / LineTime), 1) / 4.0; |
| 3231 | *final_flip_bw = dml_max(PDEAndMetaPTEBytesPerFrame * HostVMInefficiencyFactor / (*DestinationLinesToRequestVMInImmediateFlip * LineTime), (MetaRowBytes + DPTEBytesPerRow) * HostVMInefficiencyFactor / (*DestinationLinesToRequestRowInImmediateFlip * LineTime)); |
| 3232 | VRatioClamped = (VRatio < 1.0) ? 1.0 : VRatio; |
| 3233 | if (SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10) { |
| 3234 | if (GPUVMEnable == true1 && DCCEnable != true1) { |
| 3235 | min_row_time = dml_min( |
| 3236 | dpte_row_height * LineTime / VRatioClamped, |
| 3237 | dpte_row_height_chroma * LineTime / (VRatioClamped / 2)); |
| 3238 | } else if (GPUVMEnable != true1 && DCCEnable == true1) { |
| 3239 | min_row_time = dml_min( |
| 3240 | meta_row_height * LineTime / VRatioClamped, |
| 3241 | meta_row_height_chroma * LineTime / (VRatioClamped / 2)); |
| 3242 | } else { |
| 3243 | min_row_time = dml_min4( |
| 3244 | dpte_row_height * LineTime / VRatioClamped, |
| 3245 | meta_row_height * LineTime / VRatioClamped, |
| 3246 | dpte_row_height_chroma * LineTime / (VRatioClamped / 2), |
| 3247 | meta_row_height_chroma * LineTime / (VRatioClamped / 2)); |
| 3248 | } |
| 3249 | } else { |
| 3250 | if (GPUVMEnable == true1 && DCCEnable != true1) { |
| 3251 | min_row_time = dpte_row_height * LineTime / VRatioClamped; |
| 3252 | } else if (GPUVMEnable != true1 && DCCEnable == true1) { |
| 3253 | min_row_time = meta_row_height * LineTime / VRatioClamped; |
| 3254 | } else { |
| 3255 | min_row_time = dml_min( |
| 3256 | dpte_row_height * LineTime / VRatioClamped, |
| 3257 | meta_row_height * LineTime / VRatioClamped); |
| 3258 | } |
| 3259 | } |
| 3260 | |
| 3261 | if (*DestinationLinesToRequestVMInImmediateFlip >= 32 |
| 3262 | || *DestinationLinesToRequestRowInImmediateFlip >= 16 |
| 3263 | || TimeForFetchingMetaPTEImmediateFlip + 2 * TimeForFetchingRowInVBlankImmediateFlip > min_row_time) { |
| 3264 | *ImmediateFlipSupportedForPipe = false0; |
| 3265 | } else { |
| 3266 | *ImmediateFlipSupportedForPipe = true1; |
| 3267 | } |
| 3268 | } |
| 3269 | |
| 3270 | static unsigned int TruncToValidBPP( |
| 3271 | double DecimalBPP, |
| 3272 | double DesiredBPP, |
| 3273 | bool_Bool DSCEnabled, |
| 3274 | enum output_encoder_class Output, |
| 3275 | enum output_format_class Format, |
| 3276 | unsigned int DSCInputBitPerComponent) |
| 3277 | { |
| 3278 | if (Output == dm_hdmi) { |
| 3279 | if (Format == dm_420) { |
| 3280 | if (DecimalBPP >= 18 && (DesiredBPP == 0 || DesiredBPP == 18)) |
| 3281 | return 18; |
| 3282 | else if (DecimalBPP >= 15 && (DesiredBPP == 0 || DesiredBPP == 15)) |
| 3283 | return 15; |
| 3284 | else if (DecimalBPP >= 12 && (DesiredBPP == 0 || DesiredBPP == 12)) |
| 3285 | return 12; |
| 3286 | else |
| 3287 | return BPP_INVALID0; |
| 3288 | } else if (Format == dm_444) { |
| 3289 | if (DecimalBPP >= 36 && (DesiredBPP == 0 || DesiredBPP == 36)) |
| 3290 | return 36; |
| 3291 | else if (DecimalBPP >= 30 && (DesiredBPP == 0 || DesiredBPP == 30)) |
| 3292 | return 30; |
| 3293 | else if (DecimalBPP >= 24 && (DesiredBPP == 0 || DesiredBPP == 24)) |
| 3294 | return 24; |
| 3295 | else if (DecimalBPP >= 18 && (DesiredBPP == 0 || DesiredBPP == 18)) |
| 3296 | return 18; |
| 3297 | else |
| 3298 | return BPP_INVALID0; |
| 3299 | } else { |
| 3300 | if (DecimalBPP / 1.5 >= 24 && (DesiredBPP == 0 || DesiredBPP == 24)) |
| 3301 | return 24; |
| 3302 | else if (DecimalBPP / 1.5 >= 20 && (DesiredBPP == 0 || DesiredBPP == 20)) |
| 3303 | return 20; |
| 3304 | else if (DecimalBPP / 1.5 >= 16 && (DesiredBPP == 0 || DesiredBPP == 16)) |
| 3305 | return 16; |
| 3306 | else |
| 3307 | return BPP_INVALID0; |
| 3308 | } |
| 3309 | } else { |
| 3310 | if (DSCEnabled) { |
| 3311 | if (Format == dm_420) { |
| 3312 | if (DesiredBPP == 0) { |
| 3313 | if (DecimalBPP < 6) |
| 3314 | return BPP_INVALID0; |
| 3315 | else if (DecimalBPP >= 1.5 * DSCInputBitPerComponent - 1.0 / 16.0) |
| 3316 | return 1.5 * DSCInputBitPerComponent - 1.0 / 16.0; |
| 3317 | else |
| 3318 | return dml_floor(16 * DecimalBPP, 1) / 16.0; |
| 3319 | } else { |
| 3320 | if (DecimalBPP < 6 |
| 3321 | || DesiredBPP < 6 |
| 3322 | || DesiredBPP > 1.5 * DSCInputBitPerComponent - 1.0 / 16.0 |
| 3323 | || DecimalBPP < DesiredBPP) { |
| 3324 | return BPP_INVALID0; |
| 3325 | } else { |
| 3326 | return DesiredBPP; |
| 3327 | } |
| 3328 | } |
| 3329 | } else if (Format == dm_n422) { |
| 3330 | if (DesiredBPP == 0) { |
| 3331 | if (DecimalBPP < 7) |
| 3332 | return BPP_INVALID0; |
| 3333 | else if (DecimalBPP >= 2 * DSCInputBitPerComponent - 1.0 / 16.0) |
| 3334 | return 2 * DSCInputBitPerComponent - 1.0 / 16.0; |
| 3335 | else |
| 3336 | return dml_floor(16 * DecimalBPP, 1) / 16.0; |
| 3337 | } else { |
| 3338 | if (DecimalBPP < 7 |
| 3339 | || DesiredBPP < 7 |
| 3340 | || DesiredBPP > 2 * DSCInputBitPerComponent - 1.0 / 16.0 |
| 3341 | || DecimalBPP < DesiredBPP) { |
| 3342 | return BPP_INVALID0; |
| 3343 | } else { |
| 3344 | return DesiredBPP; |
| 3345 | } |
| 3346 | } |
| 3347 | } else { |
| 3348 | if (DesiredBPP == 0) { |
| 3349 | if (DecimalBPP < 8) |
| 3350 | return BPP_INVALID0; |
| 3351 | else if (DecimalBPP >= 3 * DSCInputBitPerComponent - 1.0 / 16.0) |
| 3352 | return 3 * DSCInputBitPerComponent - 1.0 / 16.0; |
| 3353 | else |
| 3354 | return dml_floor(16 * DecimalBPP, 1) / 16.0; |
| 3355 | } else { |
| 3356 | if (DecimalBPP < 8 |
| 3357 | || DesiredBPP < 8 |
| 3358 | || DesiredBPP > 3 * DSCInputBitPerComponent - 1.0 / 16.0 |
| 3359 | || DecimalBPP < DesiredBPP) { |
| 3360 | return BPP_INVALID0; |
| 3361 | } else { |
| 3362 | return DesiredBPP; |
| 3363 | } |
| 3364 | } |
| 3365 | } |
| 3366 | } else if (Format == dm_420) { |
| 3367 | if (DecimalBPP >= 18 && (DesiredBPP == 0 || DesiredBPP == 18)) |
| 3368 | return 18; |
| 3369 | else if (DecimalBPP >= 15 && (DesiredBPP == 0 || DesiredBPP == 15)) |
| 3370 | return 15; |
| 3371 | else if (DecimalBPP >= 12 && (DesiredBPP == 0 || DesiredBPP == 12)) |
| 3372 | return 12; |
| 3373 | else |
| 3374 | return BPP_INVALID0; |
| 3375 | } else if (Format == dm_s422 || Format == dm_n422) { |
| 3376 | if (DecimalBPP >= 24 && (DesiredBPP == 0 || DesiredBPP == 24)) |
| 3377 | return 24; |
| 3378 | else if (DecimalBPP >= 20 && (DesiredBPP == 0 || DesiredBPP == 20)) |
| 3379 | return 20; |
| 3380 | else if (DecimalBPP >= 16 && (DesiredBPP == 0 || DesiredBPP == 16)) |
| 3381 | return 16; |
| 3382 | else |
| 3383 | return BPP_INVALID0; |
| 3384 | } else { |
| 3385 | if (DecimalBPP >= 36 && (DesiredBPP == 0 || DesiredBPP == 36)) |
| 3386 | return 36; |
| 3387 | else if (DecimalBPP >= 30 && (DesiredBPP == 0 || DesiredBPP == 30)) |
| 3388 | return 30; |
| 3389 | else if (DecimalBPP >= 24 && (DesiredBPP == 0 || DesiredBPP == 24)) |
| 3390 | return 24; |
| 3391 | else if (DecimalBPP >= 18 && (DesiredBPP == 0 || DesiredBPP == 18)) |
| 3392 | return 18; |
| 3393 | else |
| 3394 | return BPP_INVALID0; |
| 3395 | } |
| 3396 | } |
| 3397 | } |
| 3398 | |
| 3399 | |
| 3400 | static noinline__attribute__((__noinline__)) void CalculatePrefetchSchedulePerPlane( |
| 3401 | struct display_mode_lib *mode_lib, |
| 3402 | int i, |
| 3403 | unsigned j, |
| 3404 | unsigned k) |
| 3405 | { |
| 3406 | struct vba_vars_st *locals = &mode_lib->vba; |
| 3407 | Pipe myPipe; |
| 3408 | HostVM myHostVM; |
| 3409 | |
| 3410 | if (mode_lib->vba.XFCEnabled[k] == true1) { |
| 3411 | mode_lib->vba.XFCRemoteSurfaceFlipDelay = |
| 3412 | CalculateRemoteSurfaceFlipDelay( |
| 3413 | mode_lib, |
| 3414 | mode_lib->vba.VRatio[k], |
| 3415 | locals->SwathWidthYThisState[k], |
| 3416 | dml_ceil(locals->BytePerPixelInDETY[k], 1.0), |
| 3417 | mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k], |
| 3418 | mode_lib->vba.XFCTSlvVupdateOffset, |
| 3419 | mode_lib->vba.XFCTSlvVupdateWidth, |
| 3420 | mode_lib->vba.XFCTSlvVreadyOffset, |
| 3421 | mode_lib->vba.XFCXBUFLatencyTolerance, |
| 3422 | mode_lib->vba.XFCFillBWOverhead, |
| 3423 | mode_lib->vba.XFCSlvChunkSize, |
| 3424 | mode_lib->vba.XFCBusTransportTime, |
| 3425 | mode_lib->vba.TimeCalc, |
| 3426 | mode_lib->vba.TWait, |
| 3427 | &mode_lib->vba.SrcActiveDrainRate, |
| 3428 | &mode_lib->vba.TInitXFill, |
| 3429 | &mode_lib->vba.TslvChk); |
| 3430 | } else { |
| 3431 | mode_lib->vba.XFCRemoteSurfaceFlipDelay = 0.0; |
| 3432 | } |
| 3433 | |
| 3434 | myPipe.DPPCLK = locals->RequiredDPPCLK[i][j][k]; |
| 3435 | myPipe.DISPCLK = locals->RequiredDISPCLK[i][j]; |
| 3436 | myPipe.PixelClock = mode_lib->vba.PixelClock[k]; |
| 3437 | myPipe.DCFCLKDeepSleep = mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0]; |
| 3438 | myPipe.DPPPerPlane = locals->NoOfDPP[i][j][k]; |
| 3439 | myPipe.ScalerEnabled = mode_lib->vba.ScalerEnabled[k]; |
| 3440 | myPipe.SourceScan = mode_lib->vba.SourceScan[k]; |
| 3441 | myPipe.BlockWidth256BytesY = locals->Read256BlockWidthY[k]; |
| 3442 | myPipe.BlockHeight256BytesY = locals->Read256BlockHeightY[k]; |
| 3443 | myPipe.BlockWidth256BytesC = locals->Read256BlockWidthC[k]; |
| 3444 | myPipe.BlockHeight256BytesC = locals->Read256BlockHeightC[k]; |
| 3445 | myPipe.InterlaceEnable = mode_lib->vba.Interlace[k]; |
| 3446 | myPipe.NumberOfCursors = mode_lib->vba.NumberOfCursors[k]; |
| 3447 | myPipe.VBlank = mode_lib->vba.VTotal[k] - mode_lib->vba.VActive[k]; |
| 3448 | myPipe.HTotal = mode_lib->vba.HTotal[k]; |
| 3449 | |
| 3450 | |
| 3451 | myHostVM.Enable = mode_lib->vba.HostVMEnable; |
| 3452 | myHostVM.MaxPageTableLevels = mode_lib->vba.HostVMMaxPageTableLevels; |
| 3453 | myHostVM.CachedPageTableLevels = mode_lib->vba.HostVMCachedPageTableLevels; |
| 3454 | |
| 3455 | |
| 3456 | mode_lib->vba.IsErrorResult[i][j][k] = CalculatePrefetchSchedule( |
| 3457 | mode_lib, |
| 3458 | mode_lib->vba.PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
| 3459 | mode_lib->vba.PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
| 3460 | &myPipe, |
| 3461 | locals->DSCDelayPerState[i][k], |
| 3462 | mode_lib->vba.DPPCLKDelaySubtotal, |
| 3463 | mode_lib->vba.DPPCLKDelaySCL, |
| 3464 | mode_lib->vba.DPPCLKDelaySCLLBOnly, |
| 3465 | mode_lib->vba.DPPCLKDelayCNVCFormater, |
| 3466 | mode_lib->vba.DPPCLKDelayCNVCCursor, |
| 3467 | mode_lib->vba.DISPCLKDelaySubtotal, |
| 3468 | locals->SwathWidthYThisState[k] / mode_lib->vba.HRatio[k], |
| 3469 | mode_lib->vba.OutputFormat[k], |
| 3470 | mode_lib->vba.MaxInterDCNTileRepeaters, |
| 3471 | dml_min(mode_lib->vba.MaxVStartup, locals->MaximumVStartup[0][0][k]), |
| 3472 | locals->MaximumVStartup[0][0][k], |
| 3473 | mode_lib->vba.GPUVMMaxPageTableLevels, |
| 3474 | mode_lib->vba.GPUVMEnable, |
| 3475 | &myHostVM, |
| 3476 | mode_lib->vba.DynamicMetadataEnable[k], |
| 3477 | mode_lib->vba.DynamicMetadataLinesBeforeActiveRequired[k], |
| 3478 | mode_lib->vba.DynamicMetadataTransmittedBytes[k], |
| 3479 | mode_lib->vba.DCCEnable[k], |
| 3480 | mode_lib->vba.UrgentLatency, |
| 3481 | mode_lib->vba.ExtraLatency, |
| 3482 | mode_lib->vba.TimeCalc, |
| 3483 | locals->PDEAndMetaPTEBytesPerFrame[0][0][k], |
| 3484 | locals->MetaRowBytes[0][0][k], |
| 3485 | locals->DPTEBytesPerRow[0][0][k], |
| 3486 | locals->PrefetchLinesY[0][0][k], |
| 3487 | locals->SwathWidthYThisState[k], |
| 3488 | locals->BytePerPixelInDETY[k], |
| 3489 | locals->PrefillY[k], |
| 3490 | locals->MaxNumSwY[k], |
| 3491 | locals->PrefetchLinesC[0][0][k], |
| 3492 | locals->BytePerPixelInDETC[k], |
| 3493 | locals->PrefillC[k], |
| 3494 | locals->MaxNumSwC[k], |
| 3495 | locals->SwathHeightYThisState[k], |
| 3496 | locals->SwathHeightCThisState[k], |
| 3497 | mode_lib->vba.TWait, |
| 3498 | mode_lib->vba.XFCEnabled[k], |
| 3499 | mode_lib->vba.XFCRemoteSurfaceFlipDelay, |
| 3500 | mode_lib->vba.ProgressiveToInterlaceUnitInOPP, |
| 3501 | &locals->dst_x_after_scaler, |
| 3502 | &locals->dst_y_after_scaler, |
| 3503 | &locals->LineTimesForPrefetch[k], |
| 3504 | &locals->PrefetchBW[k], |
| 3505 | &locals->LinesForMetaPTE[k], |
| 3506 | &locals->LinesForMetaAndDPTERow[k], |
| 3507 | &locals->VRatioPreY[i][j][k], |
| 3508 | &locals->VRatioPreC[i][j][k], |
| 3509 | &locals->RequiredPrefetchPixelDataBWLuma[i][j][k], |
| 3510 | &locals->RequiredPrefetchPixelDataBWChroma[i][j][k], |
| 3511 | &locals->VStartupRequiredWhenNotEnoughTimeForDynamicMetadata, |
| 3512 | &locals->Tno_bw[k], |
| 3513 | &locals->prefetch_vmrow_bw[k], |
| 3514 | locals->swath_width_luma_ub, |
| 3515 | locals->swath_width_chroma_ub, |
| 3516 | &mode_lib->vba.VUpdateOffsetPix[k], |
| 3517 | &mode_lib->vba.VUpdateWidthPix[k], |
| 3518 | &mode_lib->vba.VReadyOffsetPix[k]); |
| 3519 | } |
| 3520 | void dml21_ModeSupportAndSystemConfigurationFull(struct display_mode_lib *mode_lib) |
| 3521 | { |
| 3522 | struct vba_vars_st *locals = &mode_lib->vba; |
| 3523 | |
| 3524 | int i; |
| 3525 | unsigned int j, k, m; |
| 3526 | |
| 3527 | /*MODE SUPPORT, VOLTAGE STATE AND SOC CONFIGURATION*/ |
| 3528 | |
| 3529 | /*Scale Ratio, taps Support Check*/ |
| 3530 | |
| 3531 | mode_lib->vba.ScaleRatioAndTapsSupport = true1; |
| 3532 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 3533 | if (mode_lib->vba.ScalerEnabled[k] == false0 |
| 3534 | && ((mode_lib->vba.SourcePixelFormat[k] != dm_444_64 |
| 3535 | && mode_lib->vba.SourcePixelFormat[k] != dm_444_32 |
| 3536 | && mode_lib->vba.SourcePixelFormat[k] != dm_444_16 |
| 3537 | && mode_lib->vba.SourcePixelFormat[k] != dm_mono_16 |
| 3538 | && mode_lib->vba.SourcePixelFormat[k] != dm_mono_8) |
| 3539 | || mode_lib->vba.HRatio[k] != 1.0 |
| 3540 | || mode_lib->vba.htaps[k] != 1.0 |
| 3541 | || mode_lib->vba.VRatio[k] != 1.0 |
| 3542 | || mode_lib->vba.vtaps[k] != 1.0)) { |
| 3543 | mode_lib->vba.ScaleRatioAndTapsSupport = false0; |
| 3544 | } else if (mode_lib->vba.vtaps[k] < 1.0 || mode_lib->vba.vtaps[k] > 8.0 |
| 3545 | || mode_lib->vba.htaps[k] < 1.0 || mode_lib->vba.htaps[k] > 8.0 |
| 3546 | || (mode_lib->vba.htaps[k] > 1.0 |
| 3547 | && (mode_lib->vba.htaps[k] % 2) == 1) |
| 3548 | || mode_lib->vba.HRatio[k] > mode_lib->vba.MaxHSCLRatio |
| 3549 | || mode_lib->vba.VRatio[k] > mode_lib->vba.MaxVSCLRatio |
| 3550 | || mode_lib->vba.HRatio[k] > mode_lib->vba.htaps[k] |
| 3551 | || mode_lib->vba.VRatio[k] > mode_lib->vba.vtaps[k] |
| 3552 | || (mode_lib->vba.SourcePixelFormat[k] != dm_444_64 |
| 3553 | && mode_lib->vba.SourcePixelFormat[k] != dm_444_32 |
| 3554 | && mode_lib->vba.SourcePixelFormat[k] != dm_444_16 |
| 3555 | && mode_lib->vba.SourcePixelFormat[k] != dm_mono_16 |
| 3556 | && mode_lib->vba.SourcePixelFormat[k] != dm_mono_8 |
| 3557 | && (mode_lib->vba.HRatio[k] / 2.0 |
| 3558 | > mode_lib->vba.HTAPsChroma[k] |
| 3559 | || mode_lib->vba.VRatio[k] / 2.0 |
| 3560 | > mode_lib->vba.VTAPsChroma[k]))) { |
| 3561 | mode_lib->vba.ScaleRatioAndTapsSupport = false0; |
| 3562 | } |
| 3563 | } |
| 3564 | /*Source Format, Pixel Format and Scan Support Check*/ |
| 3565 | |
| 3566 | mode_lib->vba.SourceFormatPixelAndScanSupport = true1; |
| 3567 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 3568 | if ((mode_lib->vba.SurfaceTiling[k] == dm_sw_linear |
| 3569 | && mode_lib->vba.SourceScan[k] != dm_horz) |
| 3570 | || ((mode_lib->vba.SurfaceTiling[k] == dm_sw_4kb_d |
| 3571 | || mode_lib->vba.SurfaceTiling[k] == dm_sw_4kb_d_x |
| 3572 | || mode_lib->vba.SurfaceTiling[k] == dm_sw_64kb_d |
| 3573 | || mode_lib->vba.SurfaceTiling[k] == dm_sw_64kb_d_t |
| 3574 | || mode_lib->vba.SurfaceTiling[k] == dm_sw_64kb_d_x |
| 3575 | || mode_lib->vba.SurfaceTiling[k] == dm_sw_var_d |
| 3576 | || mode_lib->vba.SurfaceTiling[k] == dm_sw_var_d_x) |
| 3577 | && mode_lib->vba.SourcePixelFormat[k] != dm_444_64) |
| 3578 | || (mode_lib->vba.SurfaceTiling[k] == dm_sw_64kb_r_x |
| 3579 | && (mode_lib->vba.SourcePixelFormat[k] == dm_mono_8 |
| 3580 | || mode_lib->vba.SourcePixelFormat[k] |
| 3581 | == dm_420_8 |
| 3582 | || mode_lib->vba.SourcePixelFormat[k] |
| 3583 | == dm_420_10)) |
| 3584 | || (((mode_lib->vba.SurfaceTiling[k] == dm_sw_gfx7_2d_thin_gl |
| 3585 | || mode_lib->vba.SurfaceTiling[k] |
| 3586 | == dm_sw_gfx7_2d_thin_l_vp) |
| 3587 | && !((mode_lib->vba.SourcePixelFormat[k] |
| 3588 | == dm_444_64 |
| 3589 | || mode_lib->vba.SourcePixelFormat[k] |
| 3590 | == dm_444_32) |
| 3591 | && mode_lib->vba.SourceScan[k] |
| 3592 | == dm_horz |
| 3593 | && mode_lib->vba.SupportGFX7CompatibleTilingIn32bppAnd64bpp |
| 3594 | == true1 |
| 3595 | && mode_lib->vba.DCCEnable[k] |
| 3596 | == false0)) |
| 3597 | || (mode_lib->vba.DCCEnable[k] == true1 |
| 3598 | && (mode_lib->vba.SurfaceTiling[k] |
| 3599 | == dm_sw_linear |
| 3600 | || mode_lib->vba.SourcePixelFormat[k] |
| 3601 | == dm_420_8 |
| 3602 | || mode_lib->vba.SourcePixelFormat[k] |
| 3603 | == dm_420_10)))) { |
| 3604 | mode_lib->vba.SourceFormatPixelAndScanSupport = false0; |
| 3605 | } |
| 3606 | } |
| 3607 | /*Bandwidth Support Check*/ |
| 3608 | |
| 3609 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 3610 | if (mode_lib->vba.SourcePixelFormat[k] == dm_444_64) { |
| 3611 | locals->BytePerPixelInDETY[k] = 8.0; |
| 3612 | locals->BytePerPixelInDETC[k] = 0.0; |
| 3613 | } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_32) { |
| 3614 | locals->BytePerPixelInDETY[k] = 4.0; |
| 3615 | locals->BytePerPixelInDETC[k] = 0.0; |
| 3616 | } else if (mode_lib->vba.SourcePixelFormat[k] == dm_444_16 |
| 3617 | || mode_lib->vba.SourcePixelFormat[k] == dm_mono_16) { |
| 3618 | locals->BytePerPixelInDETY[k] = 2.0; |
| 3619 | locals->BytePerPixelInDETC[k] = 0.0; |
| 3620 | } else if (mode_lib->vba.SourcePixelFormat[k] == dm_mono_8) { |
| 3621 | locals->BytePerPixelInDETY[k] = 1.0; |
| 3622 | locals->BytePerPixelInDETC[k] = 0.0; |
| 3623 | } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_8) { |
| 3624 | locals->BytePerPixelInDETY[k] = 1.0; |
| 3625 | locals->BytePerPixelInDETC[k] = 2.0; |
| 3626 | } else { |
| 3627 | locals->BytePerPixelInDETY[k] = 4.0 / 3; |
| 3628 | locals->BytePerPixelInDETC[k] = 8.0 / 3; |
| 3629 | } |
| 3630 | if (mode_lib->vba.SourceScan[k] == dm_horz) { |
| 3631 | locals->SwathWidthYSingleDPP[k] = mode_lib->vba.ViewportWidth[k]; |
| 3632 | } else { |
| 3633 | locals->SwathWidthYSingleDPP[k] = mode_lib->vba.ViewportHeight[k]; |
| 3634 | } |
| 3635 | } |
| 3636 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 3637 | locals->ReadBandwidthLuma[k] = locals->SwathWidthYSingleDPP[k] * dml_ceil(locals->BytePerPixelInDETY[k], 1.0) |
| 3638 | / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]) * mode_lib->vba.VRatio[k]; |
| 3639 | locals->ReadBandwidthChroma[k] = locals->SwathWidthYSingleDPP[k] / 2 * dml_ceil(locals->BytePerPixelInDETC[k], 2.0) |
| 3640 | / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]) * mode_lib->vba.VRatio[k] / 2.0; |
| 3641 | locals->ReadBandwidth[k] = locals->ReadBandwidthLuma[k] + locals->ReadBandwidthChroma[k]; |
| 3642 | } |
| 3643 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 3644 | if (mode_lib->vba.WritebackEnable[k] == true1 |
| 3645 | && mode_lib->vba.WritebackPixelFormat[k] == dm_444_32) { |
| 3646 | locals->WriteBandwidth[k] = mode_lib->vba.WritebackDestinationWidth[k] |
| 3647 | * mode_lib->vba.WritebackDestinationHeight[k] |
| 3648 | / (mode_lib->vba.WritebackSourceHeight[k] |
| 3649 | * mode_lib->vba.HTotal[k] |
| 3650 | / mode_lib->vba.PixelClock[k]) * 4.0; |
| 3651 | } else if (mode_lib->vba.WritebackEnable[k] == true1 |
| 3652 | && mode_lib->vba.WritebackPixelFormat[k] == dm_420_10) { |
| 3653 | locals->WriteBandwidth[k] = mode_lib->vba.WritebackDestinationWidth[k] |
| 3654 | * mode_lib->vba.WritebackDestinationHeight[k] |
| 3655 | / (mode_lib->vba.WritebackSourceHeight[k] |
| 3656 | * mode_lib->vba.HTotal[k] |
| 3657 | / mode_lib->vba.PixelClock[k]) * 3.0; |
| 3658 | } else if (mode_lib->vba.WritebackEnable[k] == true1) { |
| 3659 | locals->WriteBandwidth[k] = mode_lib->vba.WritebackDestinationWidth[k] |
| 3660 | * mode_lib->vba.WritebackDestinationHeight[k] |
| 3661 | / (mode_lib->vba.WritebackSourceHeight[k] |
| 3662 | * mode_lib->vba.HTotal[k] |
| 3663 | / mode_lib->vba.PixelClock[k]) * 1.5; |
| 3664 | } else { |
| 3665 | locals->WriteBandwidth[k] = 0.0; |
| 3666 | } |
| 3667 | } |
| 3668 | mode_lib->vba.DCCEnabledInAnyPlane = false0; |
| 3669 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 3670 | if (mode_lib->vba.DCCEnable[k] == true1) { |
| 3671 | mode_lib->vba.DCCEnabledInAnyPlane = true1; |
| 3672 | } |
| 3673 | } |
| 3674 | for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { |
| 3675 | locals->IdealSDPPortBandwidthPerState[i][0] = dml_min3( |
| 3676 | mode_lib->vba.ReturnBusWidth * mode_lib->vba.DCFCLKPerState[i], |
| 3677 | mode_lib->vba.DRAMSpeedPerState[i] * mode_lib->vba.NumberOfChannels |
| 3678 | * mode_lib->vba.DRAMChannelWidth, |
| 3679 | mode_lib->vba.FabricClockPerState[i] |
| 3680 | * mode_lib->vba.FabricDatapathToDCNDataReturn); |
| 3681 | if (mode_lib->vba.HostVMEnable == false0) { |
| 3682 | locals->ReturnBWPerState[i][0] = locals->IdealSDPPortBandwidthPerState[i][0] |
| 3683 | * mode_lib->vba.PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelDataOnly / 100.0; |
| 3684 | } else { |
| 3685 | locals->ReturnBWPerState[i][0] = locals->IdealSDPPortBandwidthPerState[i][0] |
| 3686 | * mode_lib->vba.PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData / 100.0; |
| 3687 | } |
| 3688 | } |
| 3689 | /*Writeback Latency support check*/ |
| 3690 | |
| 3691 | mode_lib->vba.WritebackLatencySupport = true1; |
| 3692 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 3693 | if (mode_lib->vba.WritebackEnable[k] == true1) { |
| 3694 | if (mode_lib->vba.WritebackPixelFormat[k] == dm_444_32) { |
| 3695 | if (locals->WriteBandwidth[k] |
| 3696 | > (mode_lib->vba.WritebackInterfaceLumaBufferSize |
| 3697 | + mode_lib->vba.WritebackInterfaceChromaBufferSize) |
| 3698 | / mode_lib->vba.WritebackLatency) { |
| 3699 | mode_lib->vba.WritebackLatencySupport = false0; |
| 3700 | } |
| 3701 | } else { |
| 3702 | if (locals->WriteBandwidth[k] |
| 3703 | > 1.5 |
| 3704 | * dml_min( |
| 3705 | mode_lib->vba.WritebackInterfaceLumaBufferSize, |
| 3706 | 2.0 |
| 3707 | * mode_lib->vba.WritebackInterfaceChromaBufferSize) |
| 3708 | / mode_lib->vba.WritebackLatency) { |
| 3709 | mode_lib->vba.WritebackLatencySupport = false0; |
| 3710 | } |
| 3711 | } |
| 3712 | } |
| 3713 | } |
| 3714 | /*Re-ordering Buffer Support Check*/ |
| 3715 | |
| 3716 | for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { |
| 3717 | locals->UrgentRoundTripAndOutOfOrderLatencyPerState[i] = |
| 3718 | (mode_lib->vba.RoundTripPingLatencyCycles + 32.0) / mode_lib->vba.DCFCLKPerState[i] |
| 3719 | + dml_max3(mode_lib->vba.UrgentOutOfOrderReturnPerChannelPixelDataOnly, |
| 3720 | mode_lib->vba.UrgentOutOfOrderReturnPerChannelPixelMixedWithVMData, |
| 3721 | mode_lib->vba.UrgentOutOfOrderReturnPerChannelVMDataOnly) |
| 3722 | * mode_lib->vba.NumberOfChannels / locals->ReturnBWPerState[i][0]; |
| 3723 | if ((mode_lib->vba.ROBBufferSizeInKByte - mode_lib->vba.PixelChunkSizeInKByte) * 1024.0 / locals->ReturnBWPerState[i][0] |
| 3724 | > locals->UrgentRoundTripAndOutOfOrderLatencyPerState[i]) { |
| 3725 | locals->ROBSupport[i][0] = true1; |
| 3726 | } else { |
| 3727 | locals->ROBSupport[i][0] = false0; |
| 3728 | } |
| 3729 | } |
| 3730 | /*Writeback Mode Support Check*/ |
| 3731 | |
| 3732 | mode_lib->vba.TotalNumberOfActiveWriteback = 0; |
| 3733 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 3734 | if (mode_lib->vba.WritebackEnable[k] == true1) { |
| 3735 | if (mode_lib->vba.ActiveWritebacksPerPlane[k] == 0) |
| 3736 | mode_lib->vba.ActiveWritebacksPerPlane[k] = 1; |
| 3737 | mode_lib->vba.TotalNumberOfActiveWriteback = |
| 3738 | mode_lib->vba.TotalNumberOfActiveWriteback |
| 3739 | + mode_lib->vba.ActiveWritebacksPerPlane[k]; |
| 3740 | } |
| 3741 | } |
| 3742 | mode_lib->vba.WritebackModeSupport = true1; |
| 3743 | if (mode_lib->vba.TotalNumberOfActiveWriteback > mode_lib->vba.MaxNumWriteback) { |
| 3744 | mode_lib->vba.WritebackModeSupport = false0; |
| 3745 | } |
| 3746 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 3747 | if (mode_lib->vba.WritebackEnable[k] == true1 |
| 3748 | && mode_lib->vba.Writeback10bpc420Supported != true1 |
| 3749 | && mode_lib->vba.WritebackPixelFormat[k] == dm_420_10) { |
| 3750 | mode_lib->vba.WritebackModeSupport = false0; |
| 3751 | } |
| 3752 | } |
| 3753 | /*Writeback Scale Ratio and Taps Support Check*/ |
| 3754 | |
| 3755 | mode_lib->vba.WritebackScaleRatioAndTapsSupport = true1; |
| 3756 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 3757 | if (mode_lib->vba.WritebackEnable[k] == true1) { |
| 3758 | if (mode_lib->vba.WritebackLumaAndChromaScalingSupported == false0 |
| 3759 | && (mode_lib->vba.WritebackHRatio[k] != 1.0 |
| 3760 | || mode_lib->vba.WritebackVRatio[k] != 1.0)) { |
| 3761 | mode_lib->vba.WritebackScaleRatioAndTapsSupport = false0; |
| 3762 | } |
| 3763 | if (mode_lib->vba.WritebackHRatio[k] > mode_lib->vba.WritebackMaxHSCLRatio |
| 3764 | || mode_lib->vba.WritebackVRatio[k] |
| 3765 | > mode_lib->vba.WritebackMaxVSCLRatio |
| 3766 | || mode_lib->vba.WritebackHRatio[k] |
| 3767 | < mode_lib->vba.WritebackMinHSCLRatio |
| 3768 | || mode_lib->vba.WritebackVRatio[k] |
| 3769 | < mode_lib->vba.WritebackMinVSCLRatio |
| 3770 | || mode_lib->vba.WritebackLumaHTaps[k] |
| 3771 | > mode_lib->vba.WritebackMaxHSCLTaps |
| 3772 | || mode_lib->vba.WritebackLumaVTaps[k] |
| 3773 | > mode_lib->vba.WritebackMaxVSCLTaps |
| 3774 | || mode_lib->vba.WritebackHRatio[k] |
| 3775 | > mode_lib->vba.WritebackLumaHTaps[k] |
| 3776 | || mode_lib->vba.WritebackVRatio[k] |
| 3777 | > mode_lib->vba.WritebackLumaVTaps[k] |
| 3778 | || (mode_lib->vba.WritebackLumaHTaps[k] > 2.0 |
| 3779 | && ((mode_lib->vba.WritebackLumaHTaps[k] % 2) |
| 3780 | == 1)) |
| 3781 | || (mode_lib->vba.WritebackPixelFormat[k] != dm_444_32 |
| 3782 | && (mode_lib->vba.WritebackChromaHTaps[k] |
| 3783 | > mode_lib->vba.WritebackMaxHSCLTaps |
| 3784 | || mode_lib->vba.WritebackChromaVTaps[k] |
| 3785 | > mode_lib->vba.WritebackMaxVSCLTaps |
| 3786 | || 2.0 |
| 3787 | * mode_lib->vba.WritebackHRatio[k] |
| 3788 | > mode_lib->vba.WritebackChromaHTaps[k] |
| 3789 | || 2.0 |
| 3790 | * mode_lib->vba.WritebackVRatio[k] |
| 3791 | > mode_lib->vba.WritebackChromaVTaps[k] |
| 3792 | || (mode_lib->vba.WritebackChromaHTaps[k] > 2.0 |
| 3793 | && ((mode_lib->vba.WritebackChromaHTaps[k] % 2) == 1))))) { |
| 3794 | mode_lib->vba.WritebackScaleRatioAndTapsSupport = false0; |
| 3795 | } |
| 3796 | if (mode_lib->vba.WritebackVRatio[k] < 1.0) { |
| 3797 | mode_lib->vba.WritebackLumaVExtra = |
| 3798 | dml_max(1.0 - 2.0 / dml_ceil(1.0 / mode_lib->vba.WritebackVRatio[k], 1.0), 0.0); |
| 3799 | } else { |
| 3800 | mode_lib->vba.WritebackLumaVExtra = -1; |
| 3801 | } |
| 3802 | if ((mode_lib->vba.WritebackPixelFormat[k] == dm_444_32 |
| 3803 | && mode_lib->vba.WritebackLumaVTaps[k] |
| 3804 | > (mode_lib->vba.WritebackLineBufferLumaBufferSize |
| 3805 | + mode_lib->vba.WritebackLineBufferChromaBufferSize) |
| 3806 | / 3.0 |
| 3807 | / mode_lib->vba.WritebackDestinationWidth[k] |
| 3808 | - mode_lib->vba.WritebackLumaVExtra) |
| 3809 | || (mode_lib->vba.WritebackPixelFormat[k] == dm_420_8 |
| 3810 | && mode_lib->vba.WritebackLumaVTaps[k] |
| 3811 | > mode_lib->vba.WritebackLineBufferLumaBufferSize |
| 3812 | * 8.0 / 10.0 / mode_lib->vba.WritebackDestinationWidth[k] |
| 3813 | - mode_lib->vba.WritebackLumaVExtra) |
| 3814 | || (mode_lib->vba.WritebackPixelFormat[k] == dm_420_10 |
| 3815 | && mode_lib->vba.WritebackLumaVTaps[k] |
| 3816 | > mode_lib->vba.WritebackLineBufferLumaBufferSize |
| 3817 | * 8.0 / 10.0 |
| 3818 | / mode_lib->vba.WritebackDestinationWidth[k] |
| 3819 | - mode_lib->vba.WritebackLumaVExtra)) { |
| 3820 | mode_lib->vba.WritebackScaleRatioAndTapsSupport = false0; |
| 3821 | } |
| 3822 | if (2.0 * mode_lib->vba.WritebackVRatio[k] < 1) { |
| 3823 | mode_lib->vba.WritebackChromaVExtra = 0.0; |
| 3824 | } else { |
| 3825 | mode_lib->vba.WritebackChromaVExtra = -1; |
| 3826 | } |
| 3827 | if ((mode_lib->vba.WritebackPixelFormat[k] == dm_420_8 |
| 3828 | && mode_lib->vba.WritebackChromaVTaps[k] |
| 3829 | > mode_lib->vba.WritebackLineBufferChromaBufferSize |
| 3830 | * 8.0 / 10.0 / mode_lib->vba.WritebackDestinationWidth[k] |
| 3831 | - mode_lib->vba.WritebackChromaVExtra) |
| 3832 | || (mode_lib->vba.WritebackPixelFormat[k] == dm_420_10 |
| 3833 | && mode_lib->vba.WritebackChromaVTaps[k] |
| 3834 | > mode_lib->vba.WritebackLineBufferChromaBufferSize |
| 3835 | * 8.0 / 10.0 |
| 3836 | / mode_lib->vba.WritebackDestinationWidth[k] |
| 3837 | - mode_lib->vba.WritebackChromaVExtra)) { |
| 3838 | mode_lib->vba.WritebackScaleRatioAndTapsSupport = false0; |
| 3839 | } |
| 3840 | } |
| 3841 | } |
| 3842 | /*Maximum DISPCLK/DPPCLK Support check*/ |
| 3843 | |
| 3844 | mode_lib->vba.WritebackRequiredDISPCLK = 0.0; |
| 3845 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 3846 | if (mode_lib->vba.WritebackEnable[k] == true1) { |
| 3847 | mode_lib->vba.WritebackRequiredDISPCLK = |
| 3848 | dml_max( |
| 3849 | mode_lib->vba.WritebackRequiredDISPCLK, |
| 3850 | CalculateWriteBackDISPCLK( |
| 3851 | mode_lib->vba.WritebackPixelFormat[k], |
| 3852 | mode_lib->vba.PixelClock[k], |
| 3853 | mode_lib->vba.WritebackHRatio[k], |
| 3854 | mode_lib->vba.WritebackVRatio[k], |
| 3855 | mode_lib->vba.WritebackLumaHTaps[k], |
| 3856 | mode_lib->vba.WritebackLumaVTaps[k], |
| 3857 | mode_lib->vba.WritebackChromaHTaps[k], |
| 3858 | mode_lib->vba.WritebackChromaVTaps[k], |
| 3859 | mode_lib->vba.WritebackDestinationWidth[k], |
| 3860 | mode_lib->vba.HTotal[k], |
| 3861 | mode_lib->vba.WritebackChromaLineBufferWidth)); |
| 3862 | } |
| 3863 | } |
| 3864 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 3865 | if (mode_lib->vba.HRatio[k] > 1.0) { |
| 3866 | locals->PSCL_FACTOR[k] = dml_min( |
| 3867 | mode_lib->vba.MaxDCHUBToPSCLThroughput, |
| 3868 | mode_lib->vba.MaxPSCLToLBThroughput |
| 3869 | * mode_lib->vba.HRatio[k] |
| 3870 | / dml_ceil( |
| 3871 | mode_lib->vba.htaps[k] |
| 3872 | / 6.0, |
| 3873 | 1.0)); |
| 3874 | } else { |
| 3875 | locals->PSCL_FACTOR[k] = dml_min( |
| 3876 | mode_lib->vba.MaxDCHUBToPSCLThroughput, |
| 3877 | mode_lib->vba.MaxPSCLToLBThroughput); |
| 3878 | } |
| 3879 | if (locals->BytePerPixelInDETC[k] == 0.0) { |
| 3880 | locals->PSCL_FACTOR_CHROMA[k] = 0.0; |
| 3881 | locals->MinDPPCLKUsingSingleDPP[k] = |
| 3882 | mode_lib->vba.PixelClock[k] |
| 3883 | * dml_max3( |
| 3884 | mode_lib->vba.vtaps[k] / 6.0 |
| 3885 | * dml_min( |
| 3886 | 1.0, |
| 3887 | mode_lib->vba.HRatio[k]), |
| 3888 | mode_lib->vba.HRatio[k] |
| 3889 | * mode_lib->vba.VRatio[k] |
| 3890 | / locals->PSCL_FACTOR[k], |
| 3891 | 1.0); |
| 3892 | if ((mode_lib->vba.htaps[k] > 6.0 || mode_lib->vba.vtaps[k] > 6.0) |
| 3893 | && locals->MinDPPCLKUsingSingleDPP[k] |
| 3894 | < 2.0 * mode_lib->vba.PixelClock[k]) { |
| 3895 | locals->MinDPPCLKUsingSingleDPP[k] = 2.0 |
| 3896 | * mode_lib->vba.PixelClock[k]; |
| 3897 | } |
| 3898 | } else { |
| 3899 | if (mode_lib->vba.HRatio[k] / 2.0 > 1.0) { |
| 3900 | locals->PSCL_FACTOR_CHROMA[k] = |
| 3901 | dml_min( |
| 3902 | mode_lib->vba.MaxDCHUBToPSCLThroughput, |
| 3903 | mode_lib->vba.MaxPSCLToLBThroughput |
| 3904 | * mode_lib->vba.HRatio[k] |
| 3905 | / 2.0 |
| 3906 | / dml_ceil( |
| 3907 | mode_lib->vba.HTAPsChroma[k] |
| 3908 | / 6.0, |
| 3909 | 1.0)); |
| 3910 | } else { |
| 3911 | locals->PSCL_FACTOR_CHROMA[k] = dml_min( |
| 3912 | mode_lib->vba.MaxDCHUBToPSCLThroughput, |
| 3913 | mode_lib->vba.MaxPSCLToLBThroughput); |
| 3914 | } |
| 3915 | locals->MinDPPCLKUsingSingleDPP[k] = |
| 3916 | mode_lib->vba.PixelClock[k] |
| 3917 | * dml_max5( |
| 3918 | mode_lib->vba.vtaps[k] / 6.0 |
| 3919 | * dml_min( |
| 3920 | 1.0, |
| 3921 | mode_lib->vba.HRatio[k]), |
| 3922 | mode_lib->vba.HRatio[k] |
| 3923 | * mode_lib->vba.VRatio[k] |
| 3924 | / locals->PSCL_FACTOR[k], |
| 3925 | mode_lib->vba.VTAPsChroma[k] |
| 3926 | / 6.0 |
| 3927 | * dml_min( |
| 3928 | 1.0, |
| 3929 | mode_lib->vba.HRatio[k] |
| 3930 | / 2.0), |
| 3931 | mode_lib->vba.HRatio[k] |
| 3932 | * mode_lib->vba.VRatio[k] |
| 3933 | / 4.0 |
| 3934 | / locals->PSCL_FACTOR_CHROMA[k], |
| 3935 | 1.0); |
| 3936 | if ((mode_lib->vba.htaps[k] > 6.0 || mode_lib->vba.vtaps[k] > 6.0 |
| 3937 | || mode_lib->vba.HTAPsChroma[k] > 6.0 |
| 3938 | || mode_lib->vba.VTAPsChroma[k] > 6.0) |
| 3939 | && locals->MinDPPCLKUsingSingleDPP[k] |
| 3940 | < 2.0 * mode_lib->vba.PixelClock[k]) { |
| 3941 | locals->MinDPPCLKUsingSingleDPP[k] = 2.0 |
| 3942 | * mode_lib->vba.PixelClock[k]; |
| 3943 | } |
| 3944 | } |
| 3945 | } |
| 3946 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 3947 | Calculate256BBlockSizes( |
| 3948 | mode_lib->vba.SourcePixelFormat[k], |
| 3949 | mode_lib->vba.SurfaceTiling[k], |
| 3950 | dml_ceil(locals->BytePerPixelInDETY[k], 1.0), |
| 3951 | dml_ceil(locals->BytePerPixelInDETC[k], 2.0), |
| 3952 | &locals->Read256BlockHeightY[k], |
| 3953 | &locals->Read256BlockHeightC[k], |
| 3954 | &locals->Read256BlockWidthY[k], |
| 3955 | &locals->Read256BlockWidthC[k]); |
| 3956 | if (mode_lib->vba.SourceScan[k] == dm_horz) { |
| 3957 | locals->MaxSwathHeightY[k] = locals->Read256BlockHeightY[k]; |
| 3958 | locals->MaxSwathHeightC[k] = locals->Read256BlockHeightC[k]; |
| 3959 | } else { |
| 3960 | locals->MaxSwathHeightY[k] = locals->Read256BlockWidthY[k]; |
| 3961 | locals->MaxSwathHeightC[k] = locals->Read256BlockWidthC[k]; |
| 3962 | } |
| 3963 | if ((mode_lib->vba.SourcePixelFormat[k] == dm_444_64 |
| 3964 | || mode_lib->vba.SourcePixelFormat[k] == dm_444_32 |
| 3965 | || mode_lib->vba.SourcePixelFormat[k] == dm_444_16 |
| 3966 | || mode_lib->vba.SourcePixelFormat[k] == dm_mono_16 |
| 3967 | || mode_lib->vba.SourcePixelFormat[k] == dm_mono_8)) { |
| 3968 | if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear |
| 3969 | || (mode_lib->vba.SourcePixelFormat[k] == dm_444_64 |
| 3970 | && (mode_lib->vba.SurfaceTiling[k] |
| 3971 | == dm_sw_4kb_s |
| 3972 | || mode_lib->vba.SurfaceTiling[k] |
| 3973 | == dm_sw_4kb_s_x |
| 3974 | || mode_lib->vba.SurfaceTiling[k] |
| 3975 | == dm_sw_64kb_s |
| 3976 | || mode_lib->vba.SurfaceTiling[k] |
| 3977 | == dm_sw_64kb_s_t |
| 3978 | || mode_lib->vba.SurfaceTiling[k] |
| 3979 | == dm_sw_64kb_s_x |
| 3980 | || mode_lib->vba.SurfaceTiling[k] |
| 3981 | == dm_sw_var_s |
| 3982 | || mode_lib->vba.SurfaceTiling[k] |
| 3983 | == dm_sw_var_s_x) |
| 3984 | && mode_lib->vba.SourceScan[k] == dm_horz)) { |
| 3985 | locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k]; |
| 3986 | } else { |
| 3987 | locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k] |
| 3988 | / 2.0; |
| 3989 | } |
| 3990 | locals->MinSwathHeightC[k] = locals->MaxSwathHeightC[k]; |
| 3991 | } else { |
| 3992 | if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear) { |
| 3993 | locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k]; |
| 3994 | locals->MinSwathHeightC[k] = locals->MaxSwathHeightC[k]; |
| 3995 | } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_8 |
| 3996 | && mode_lib->vba.SourceScan[k] == dm_horz) { |
| 3997 | locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k] |
| 3998 | / 2.0; |
| 3999 | locals->MinSwathHeightC[k] = locals->MaxSwathHeightC[k]; |
| 4000 | } else if (mode_lib->vba.SourcePixelFormat[k] == dm_420_10 |
| 4001 | && mode_lib->vba.SourceScan[k] == dm_horz) { |
| 4002 | locals->MinSwathHeightC[k] = locals->MaxSwathHeightC[k] |
| 4003 | / 2.0; |
| 4004 | locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k]; |
| 4005 | } else { |
| 4006 | locals->MinSwathHeightY[k] = locals->MaxSwathHeightY[k]; |
| 4007 | locals->MinSwathHeightC[k] = locals->MaxSwathHeightC[k]; |
| 4008 | } |
| 4009 | } |
| 4010 | if (mode_lib->vba.SurfaceTiling[k] == dm_sw_linear) { |
| 4011 | mode_lib->vba.MaximumSwathWidthSupport = 8192.0; |
| 4012 | } else { |
| 4013 | mode_lib->vba.MaximumSwathWidthSupport = 5120.0; |
| 4014 | } |
| 4015 | mode_lib->vba.MaximumSwathWidthInDETBuffer = |
| 4016 | dml_min( |
| 4017 | mode_lib->vba.MaximumSwathWidthSupport, |
| 4018 | mode_lib->vba.DETBufferSizeInKByte[0] * 1024.0 / 2.0 |
| 4019 | / (locals->BytePerPixelInDETY[k] |
| 4020 | * locals->MinSwathHeightY[k] |
| 4021 | + locals->BytePerPixelInDETC[k] |
| 4022 | / 2.0 |
| 4023 | * locals->MinSwathHeightC[k])); |
| 4024 | if (locals->BytePerPixelInDETC[k] == 0.0) { |
| 4025 | mode_lib->vba.MaximumSwathWidthInLineBuffer = |
| 4026 | mode_lib->vba.LineBufferSize |
| 4027 | * dml_max(mode_lib->vba.HRatio[k], 1.0) |
| 4028 | / mode_lib->vba.LBBitPerPixel[k] |
| 4029 | / (mode_lib->vba.vtaps[k] |
| 4030 | + dml_max( |
| 4031 | dml_ceil( |
| 4032 | mode_lib->vba.VRatio[k], |
| 4033 | 1.0) |
| 4034 | - 2, |
| 4035 | 0.0)); |
| 4036 | } else { |
| 4037 | mode_lib->vba.MaximumSwathWidthInLineBuffer = |
| 4038 | dml_min( |
| 4039 | mode_lib->vba.LineBufferSize |
| 4040 | * dml_max( |
| 4041 | mode_lib->vba.HRatio[k], |
| 4042 | 1.0) |
| 4043 | / mode_lib->vba.LBBitPerPixel[k] |
| 4044 | / (mode_lib->vba.vtaps[k] |
| 4045 | + dml_max( |
| 4046 | dml_ceil( |
| 4047 | mode_lib->vba.VRatio[k], |
| 4048 | 1.0) |
| 4049 | - 2, |
| 4050 | 0.0)), |
| 4051 | 2.0 * mode_lib->vba.LineBufferSize |
| 4052 | * dml_max( |
| 4053 | mode_lib->vba.HRatio[k] |
| 4054 | / 2.0, |
| 4055 | 1.0) |
| 4056 | / mode_lib->vba.LBBitPerPixel[k] |
| 4057 | / (mode_lib->vba.VTAPsChroma[k] |
| 4058 | + dml_max( |
| 4059 | dml_ceil( |
| 4060 | mode_lib->vba.VRatio[k] |
| 4061 | / 2.0, |
| 4062 | 1.0) |
| 4063 | - 2, |
| 4064 | 0.0))); |
| 4065 | } |
| 4066 | locals->MaximumSwathWidth[k] = dml_min( |
| 4067 | mode_lib->vba.MaximumSwathWidthInDETBuffer, |
| 4068 | mode_lib->vba.MaximumSwathWidthInLineBuffer); |
| 4069 | } |
| 4070 | for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { |
| 4071 | double MaxMaxDispclkRoundedDown = RoundToDFSGranularityDown( |
| 4072 | mode_lib->vba.MaxDispclk[mode_lib->vba.soc.num_states], |
| 4073 | mode_lib->vba.DISPCLKDPPCLKVCOSpeed); |
| 4074 | |
| 4075 | for (j = 0; j < 2; j++) { |
| 4076 | mode_lib->vba.MaxDispclkRoundedDownToDFSGranularity = RoundToDFSGranularityDown( |
| 4077 | mode_lib->vba.MaxDispclk[i], |
| 4078 | mode_lib->vba.DISPCLKDPPCLKVCOSpeed); |
| 4079 | mode_lib->vba.MaxDppclkRoundedDownToDFSGranularity = RoundToDFSGranularityDown( |
| 4080 | mode_lib->vba.MaxDppclk[i], |
| 4081 | mode_lib->vba.DISPCLKDPPCLKVCOSpeed); |
| 4082 | locals->RequiredDISPCLK[i][j] = 0.0; |
| 4083 | locals->DISPCLK_DPPCLK_Support[i][j] = true1; |
| 4084 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4085 | mode_lib->vba.PlaneRequiredDISPCLKWithoutODMCombine = |
| 4086 | mode_lib->vba.PixelClock[k] |
| 4087 | * (1.0 |
| 4088 | + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading |
| 4089 | / 100.0) |
| 4090 | * (1.0 |
| 4091 | + mode_lib->vba.DISPCLKRampingMargin |
| 4092 | / 100.0); |
| 4093 | if (mode_lib->vba.PlaneRequiredDISPCLKWithoutODMCombine >= mode_lib->vba.MaxDispclk[i] |
| 4094 | && i == mode_lib->vba.soc.num_states) |
| 4095 | mode_lib->vba.PlaneRequiredDISPCLKWithoutODMCombine = mode_lib->vba.PixelClock[k] |
| 4096 | * (1 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); |
| 4097 | |
| 4098 | mode_lib->vba.PlaneRequiredDISPCLKWithODMCombine = mode_lib->vba.PixelClock[k] / 2 |
| 4099 | * (1 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) * (1 + mode_lib->vba.DISPCLKRampingMargin / 100.0); |
| 4100 | if (mode_lib->vba.PlaneRequiredDISPCLKWithODMCombine >= mode_lib->vba.MaxDispclk[i] |
| 4101 | && i == mode_lib->vba.soc.num_states) |
| 4102 | mode_lib->vba.PlaneRequiredDISPCLKWithODMCombine = mode_lib->vba.PixelClock[k] / 2 |
| 4103 | * (1 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); |
| 4104 | |
| 4105 | locals->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_disabled; |
| 4106 | mode_lib->vba.PlaneRequiredDISPCLK = mode_lib->vba.PlaneRequiredDISPCLKWithoutODMCombine; |
| 4107 | if (mode_lib->vba.ODMCapability) { |
| 4108 | if (locals->PlaneRequiredDISPCLKWithoutODMCombine > MaxMaxDispclkRoundedDown) { |
| 4109 | locals->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_2to1; |
| 4110 | mode_lib->vba.PlaneRequiredDISPCLK = mode_lib->vba.PlaneRequiredDISPCLKWithODMCombine; |
| 4111 | } else if (locals->DSCEnabled[k] && (locals->HActive[k] > DCN21_MAX_DSC_IMAGE_WIDTH5184)) { |
| 4112 | locals->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_2to1; |
| 4113 | mode_lib->vba.PlaneRequiredDISPCLK = mode_lib->vba.PlaneRequiredDISPCLKWithODMCombine; |
| 4114 | } else if (locals->HActive[k] > DCN21_MAX_420_IMAGE_WIDTH4096 && locals->OutputFormat[k] == dm_420) { |
| 4115 | locals->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_2to1; |
| 4116 | mode_lib->vba.PlaneRequiredDISPCLK = mode_lib->vba.PlaneRequiredDISPCLKWithODMCombine; |
| 4117 | } |
| 4118 | } |
| 4119 | |
| 4120 | if (locals->MinDPPCLKUsingSingleDPP[k] * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) <= mode_lib->vba.MaxDppclkRoundedDownToDFSGranularity |
| 4121 | && locals->SwathWidthYSingleDPP[k] <= locals->MaximumSwathWidth[k] |
| 4122 | && locals->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_disabled) { |
| 4123 | locals->NoOfDPP[i][j][k] = 1; |
| 4124 | locals->RequiredDPPCLK[i][j][k] = |
| 4125 | locals->MinDPPCLKUsingSingleDPP[k] * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); |
| 4126 | } else { |
| 4127 | locals->NoOfDPP[i][j][k] = 2; |
| 4128 | locals->RequiredDPPCLK[i][j][k] = |
| 4129 | locals->MinDPPCLKUsingSingleDPP[k] * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) / 2.0; |
| 4130 | } |
| 4131 | locals->RequiredDISPCLK[i][j] = dml_max( |
| 4132 | locals->RequiredDISPCLK[i][j], |
| 4133 | mode_lib->vba.PlaneRequiredDISPCLK); |
| 4134 | if ((locals->MinDPPCLKUsingSingleDPP[k] / locals->NoOfDPP[i][j][k] * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) |
| 4135 | > mode_lib->vba.MaxDppclkRoundedDownToDFSGranularity) |
| 4136 | || (mode_lib->vba.PlaneRequiredDISPCLK > mode_lib->vba.MaxDispclkRoundedDownToDFSGranularity)) { |
| 4137 | locals->DISPCLK_DPPCLK_Support[i][j] = false0; |
| 4138 | } |
| 4139 | } |
| 4140 | locals->TotalNumberOfActiveDPP[i][j] = 0.0; |
| 4141 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) |
| 4142 | locals->TotalNumberOfActiveDPP[i][j] = locals->TotalNumberOfActiveDPP[i][j] + locals->NoOfDPP[i][j][k]; |
| 4143 | if (j == 1) { |
| 4144 | while (locals->TotalNumberOfActiveDPP[i][j] < mode_lib->vba.MaxNumDPP |
| 4145 | && locals->TotalNumberOfActiveDPP[i][j] < 2 * mode_lib->vba.NumberOfActivePlanes) { |
| 4146 | double BWOfNonSplitPlaneOfMaximumBandwidth; |
| 4147 | unsigned int NumberOfNonSplitPlaneOfMaximumBandwidth; |
| 4148 | |
| 4149 | BWOfNonSplitPlaneOfMaximumBandwidth = 0; |
| 4150 | NumberOfNonSplitPlaneOfMaximumBandwidth = 0; |
| 4151 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 4152 | if (locals->ReadBandwidth[k] > BWOfNonSplitPlaneOfMaximumBandwidth && locals->NoOfDPP[i][j][k] == 1) { |
| 4153 | BWOfNonSplitPlaneOfMaximumBandwidth = locals->ReadBandwidth[k]; |
| 4154 | NumberOfNonSplitPlaneOfMaximumBandwidth = k; |
| 4155 | } |
| 4156 | } |
| 4157 | locals->NoOfDPP[i][j][NumberOfNonSplitPlaneOfMaximumBandwidth] = 2; |
| 4158 | locals->RequiredDPPCLK[i][j][NumberOfNonSplitPlaneOfMaximumBandwidth] = |
| 4159 | locals->MinDPPCLKUsingSingleDPP[NumberOfNonSplitPlaneOfMaximumBandwidth] |
| 4160 | * (1 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100) / 2; |
| 4161 | locals->TotalNumberOfActiveDPP[i][j] = locals->TotalNumberOfActiveDPP[i][j] + 1; |
| 4162 | } |
| 4163 | } |
| 4164 | if (locals->TotalNumberOfActiveDPP[i][j] > mode_lib->vba.MaxNumDPP) { |
| 4165 | locals->RequiredDISPCLK[i][j] = 0.0; |
| 4166 | locals->DISPCLK_DPPCLK_Support[i][j] = true1; |
| 4167 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4168 | locals->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_disabled; |
| 4169 | if (locals->SwathWidthYSingleDPP[k] <= locals->MaximumSwathWidth[k]) { |
| 4170 | locals->NoOfDPP[i][j][k] = 1; |
| 4171 | locals->RequiredDPPCLK[i][j][k] = locals->MinDPPCLKUsingSingleDPP[k] |
| 4172 | * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); |
| 4173 | } else { |
| 4174 | locals->NoOfDPP[i][j][k] = 2; |
| 4175 | locals->RequiredDPPCLK[i][j][k] = locals->MinDPPCLKUsingSingleDPP[k] |
| 4176 | * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) / 2.0; |
| 4177 | } |
| 4178 | if (i != mode_lib->vba.soc.num_states) { |
| 4179 | mode_lib->vba.PlaneRequiredDISPCLK = |
| 4180 | mode_lib->vba.PixelClock[k] |
| 4181 | * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) |
| 4182 | * (1.0 + mode_lib->vba.DISPCLKRampingMargin / 100.0); |
| 4183 | } else { |
| 4184 | mode_lib->vba.PlaneRequiredDISPCLK = mode_lib->vba.PixelClock[k] |
| 4185 | * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); |
| 4186 | } |
| 4187 | locals->RequiredDISPCLK[i][j] = dml_max( |
| 4188 | locals->RequiredDISPCLK[i][j], |
| 4189 | mode_lib->vba.PlaneRequiredDISPCLK); |
| 4190 | if (locals->MinDPPCLKUsingSingleDPP[k] / locals->NoOfDPP[i][j][k] * (1.0 + mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) |
| 4191 | > mode_lib->vba.MaxDppclkRoundedDownToDFSGranularity |
| 4192 | || mode_lib->vba.PlaneRequiredDISPCLK > mode_lib->vba.MaxDispclkRoundedDownToDFSGranularity) |
| 4193 | locals->DISPCLK_DPPCLK_Support[i][j] = false0; |
| 4194 | } |
| 4195 | locals->TotalNumberOfActiveDPP[i][j] = 0.0; |
| 4196 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) |
| 4197 | locals->TotalNumberOfActiveDPP[i][j] = locals->TotalNumberOfActiveDPP[i][j] + locals->NoOfDPP[i][j][k]; |
| 4198 | } |
| 4199 | locals->RequiredDISPCLK[i][j] = dml_max( |
| 4200 | locals->RequiredDISPCLK[i][j], |
| 4201 | mode_lib->vba.WritebackRequiredDISPCLK); |
| 4202 | if (mode_lib->vba.MaxDispclkRoundedDownToDFSGranularity |
| 4203 | < mode_lib->vba.WritebackRequiredDISPCLK) { |
| 4204 | locals->DISPCLK_DPPCLK_Support[i][j] = false0; |
| 4205 | } |
| 4206 | } |
| 4207 | } |
| 4208 | /*Viewport Size Check*/ |
| 4209 | |
| 4210 | for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { |
| 4211 | locals->ViewportSizeSupport[i][0] = true1; |
| 4212 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4213 | if (locals->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_2to1) { |
| 4214 | if (dml_min(locals->SwathWidthYSingleDPP[k], dml_round(mode_lib->vba.HActive[k] / 2.0 * mode_lib->vba.HRatio[k])) |
| 4215 | > locals->MaximumSwathWidth[k]) { |
| 4216 | locals->ViewportSizeSupport[i][0] = false0; |
| 4217 | } |
| 4218 | } else { |
| 4219 | if (locals->SwathWidthYSingleDPP[k] / 2.0 > locals->MaximumSwathWidth[k]) { |
| 4220 | locals->ViewportSizeSupport[i][0] = false0; |
| 4221 | } |
| 4222 | } |
| 4223 | } |
| 4224 | } |
| 4225 | /*Total Available Pipes Support Check*/ |
| 4226 | |
| 4227 | for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { |
| 4228 | for (j = 0; j < 2; j++) { |
| 4229 | if (locals->TotalNumberOfActiveDPP[i][j] <= mode_lib->vba.MaxNumDPP) |
| 4230 | locals->TotalAvailablePipesSupport[i][j] = true1; |
| 4231 | else |
| 4232 | locals->TotalAvailablePipesSupport[i][j] = false0; |
| 4233 | } |
| 4234 | } |
| 4235 | /*Total Available OTG Support Check*/ |
| 4236 | |
| 4237 | mode_lib->vba.TotalNumberOfActiveOTG = 0.0; |
| 4238 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4239 | if (mode_lib->vba.BlendingAndTiming[k] == k) { |
| 4240 | mode_lib->vba.TotalNumberOfActiveOTG = mode_lib->vba.TotalNumberOfActiveOTG |
| 4241 | + 1.0; |
| 4242 | } |
| 4243 | } |
| 4244 | if (mode_lib->vba.TotalNumberOfActiveOTG <= mode_lib->vba.MaxNumOTG) { |
| 4245 | mode_lib->vba.NumberOfOTGSupport = true1; |
| 4246 | } else { |
| 4247 | mode_lib->vba.NumberOfOTGSupport = false0; |
| 4248 | } |
| 4249 | /*Display IO and DSC Support Check*/ |
| 4250 | |
| 4251 | mode_lib->vba.NonsupportedDSCInputBPC = false0; |
| 4252 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4253 | if (!(mode_lib->vba.DSCInputBitPerComponent[k] == 12.0 |
| 4254 | || mode_lib->vba.DSCInputBitPerComponent[k] == 10.0 |
| 4255 | || mode_lib->vba.DSCInputBitPerComponent[k] == 8.0)) { |
| 4256 | mode_lib->vba.NonsupportedDSCInputBPC = true1; |
| 4257 | } |
| 4258 | } |
| 4259 | for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { |
| 4260 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4261 | locals->RequiresDSC[i][k] = false0; |
| 4262 | locals->RequiresFEC[i][k] = 0; |
| 4263 | if (mode_lib->vba.BlendingAndTiming[k] == k) { |
| 4264 | if (mode_lib->vba.Output[k] == dm_hdmi) { |
| 4265 | locals->RequiresDSC[i][k] = false0; |
| 4266 | locals->RequiresFEC[i][k] = 0; |
| 4267 | locals->OutputBppPerState[i][k] = TruncToValidBPP( |
| 4268 | dml_min(600.0, mode_lib->vba.PHYCLKPerState[i]) / mode_lib->vba.PixelClockBackEnd[k] * 24, |
| 4269 | mode_lib->vba.ForcedOutputLinkBPP[k], |
| 4270 | false0, |
| 4271 | mode_lib->vba.Output[k], |
| 4272 | mode_lib->vba.OutputFormat[k], |
| 4273 | mode_lib->vba.DSCInputBitPerComponent[k]); |
| 4274 | } else if (mode_lib->vba.Output[k] == dm_dp |
| 4275 | || mode_lib->vba.Output[k] == dm_edp) { |
| 4276 | if (mode_lib->vba.Output[k] == dm_edp) { |
| 4277 | mode_lib->vba.EffectiveFECOverhead = 0.0; |
| 4278 | } else { |
| 4279 | mode_lib->vba.EffectiveFECOverhead = |
| 4280 | mode_lib->vba.FECOverhead; |
| 4281 | } |
| 4282 | if (mode_lib->vba.PHYCLKPerState[i] >= 270.0) { |
| 4283 | mode_lib->vba.Outbpp = TruncToValidBPP( |
| 4284 | (1.0 - mode_lib->vba.Downspreading / 100.0) * 270.0 |
| 4285 | * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0, |
| 4286 | mode_lib->vba.ForcedOutputLinkBPP[k], |
| 4287 | false0, |
| 4288 | mode_lib->vba.Output[k], |
| 4289 | mode_lib->vba.OutputFormat[k], |
| 4290 | mode_lib->vba.DSCInputBitPerComponent[k]); |
| 4291 | mode_lib->vba.OutbppDSC = TruncToValidBPP( |
| 4292 | (1.0 - mode_lib->vba.Downspreading / 100.0) * (1.0 - mode_lib->vba.EffectiveFECOverhead / 100.0) * 270.0 |
| 4293 | * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0, |
| 4294 | mode_lib->vba.ForcedOutputLinkBPP[k], |
| 4295 | true1, |
| 4296 | mode_lib->vba.Output[k], |
| 4297 | mode_lib->vba.OutputFormat[k], |
| 4298 | mode_lib->vba.DSCInputBitPerComponent[k]); |
| 4299 | if (mode_lib->vba.DSCEnabled[k] == true1) { |
| 4300 | locals->RequiresDSC[i][k] = true1; |
| 4301 | if (mode_lib->vba.Output[k] == dm_dp) { |
| 4302 | locals->RequiresFEC[i][k] = true1; |
| 4303 | } else { |
| 4304 | locals->RequiresFEC[i][k] = false0; |
| 4305 | } |
| 4306 | mode_lib->vba.Outbpp = mode_lib->vba.OutbppDSC; |
| 4307 | } else { |
| 4308 | locals->RequiresDSC[i][k] = false0; |
| 4309 | locals->RequiresFEC[i][k] = false0; |
| 4310 | } |
| 4311 | locals->OutputBppPerState[i][k] = mode_lib->vba.Outbpp; |
| 4312 | } |
| 4313 | if (mode_lib->vba.Outbpp == BPP_INVALID0 && mode_lib->vba.PHYCLKPerState[i] >= 540.0) { |
| 4314 | mode_lib->vba.Outbpp = TruncToValidBPP( |
| 4315 | (1.0 - mode_lib->vba.Downspreading / 100.0) * 540.0 |
| 4316 | * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0, |
| 4317 | mode_lib->vba.ForcedOutputLinkBPP[k], |
| 4318 | false0, |
| 4319 | mode_lib->vba.Output[k], |
| 4320 | mode_lib->vba.OutputFormat[k], |
| 4321 | mode_lib->vba.DSCInputBitPerComponent[k]); |
| 4322 | mode_lib->vba.OutbppDSC = TruncToValidBPP( |
| 4323 | (1.0 - mode_lib->vba.Downspreading / 100.0) * (1.0 - mode_lib->vba.EffectiveFECOverhead / 100.0) * 540.0 |
| 4324 | * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0, |
| 4325 | mode_lib->vba.ForcedOutputLinkBPP[k], |
| 4326 | true1, |
| 4327 | mode_lib->vba.Output[k], |
| 4328 | mode_lib->vba.OutputFormat[k], |
| 4329 | mode_lib->vba.DSCInputBitPerComponent[k]); |
| 4330 | if (mode_lib->vba.DSCEnabled[k] == true1) { |
| 4331 | locals->RequiresDSC[i][k] = true1; |
| 4332 | if (mode_lib->vba.Output[k] == dm_dp) { |
| 4333 | locals->RequiresFEC[i][k] = true1; |
| 4334 | } else { |
| 4335 | locals->RequiresFEC[i][k] = false0; |
| 4336 | } |
| 4337 | mode_lib->vba.Outbpp = mode_lib->vba.OutbppDSC; |
| 4338 | } else { |
| 4339 | locals->RequiresDSC[i][k] = false0; |
| 4340 | locals->RequiresFEC[i][k] = false0; |
| 4341 | } |
| 4342 | locals->OutputBppPerState[i][k] = mode_lib->vba.Outbpp; |
| 4343 | } |
| 4344 | if (mode_lib->vba.Outbpp == BPP_INVALID0 |
| 4345 | && mode_lib->vba.PHYCLKPerState[i] |
| 4346 | >= 810.0) { |
| 4347 | mode_lib->vba.Outbpp = TruncToValidBPP( |
| 4348 | (1.0 - mode_lib->vba.Downspreading / 100.0) * 810.0 |
| 4349 | * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0, |
| 4350 | mode_lib->vba.ForcedOutputLinkBPP[k], |
| 4351 | false0, |
| 4352 | mode_lib->vba.Output[k], |
| 4353 | mode_lib->vba.OutputFormat[k], |
| 4354 | mode_lib->vba.DSCInputBitPerComponent[k]); |
| 4355 | mode_lib->vba.OutbppDSC = TruncToValidBPP( |
| 4356 | (1.0 - mode_lib->vba.Downspreading / 100.0) * (1.0 - mode_lib->vba.EffectiveFECOverhead / 100.0) * 810.0 |
| 4357 | * mode_lib->vba.OutputLinkDPLanes[k] / mode_lib->vba.PixelClockBackEnd[k] * 8.0, |
| 4358 | mode_lib->vba.ForcedOutputLinkBPP[k], |
| 4359 | true1, |
| 4360 | mode_lib->vba.Output[k], |
| 4361 | mode_lib->vba.OutputFormat[k], |
| 4362 | mode_lib->vba.DSCInputBitPerComponent[k]); |
| 4363 | if (mode_lib->vba.DSCEnabled[k] == true1 || mode_lib->vba.Outbpp == BPP_INVALID0) { |
| 4364 | locals->RequiresDSC[i][k] = true1; |
| 4365 | if (mode_lib->vba.Output[k] == dm_dp) { |
| 4366 | locals->RequiresFEC[i][k] = true1; |
| 4367 | } else { |
| 4368 | locals->RequiresFEC[i][k] = false0; |
| 4369 | } |
| 4370 | mode_lib->vba.Outbpp = mode_lib->vba.OutbppDSC; |
| 4371 | } else { |
| 4372 | locals->RequiresDSC[i][k] = false0; |
| 4373 | locals->RequiresFEC[i][k] = false0; |
| 4374 | } |
| 4375 | locals->OutputBppPerState[i][k] = |
| 4376 | mode_lib->vba.Outbpp; |
| 4377 | } |
| 4378 | } |
| 4379 | } else { |
| 4380 | locals->OutputBppPerState[i][k] = BPP_BLENDED_PIPE0xffffffff; |
| 4381 | } |
| 4382 | } |
| 4383 | } |
| 4384 | for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { |
| 4385 | locals->DIOSupport[i] = true1; |
| 4386 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4387 | if (!mode_lib->vba.skip_dio_check[k] |
| 4388 | && (locals->OutputBppPerState[i][k] == BPP_INVALID0 |
| 4389 | || (mode_lib->vba.OutputFormat[k] == dm_420 |
| 4390 | && mode_lib->vba.Interlace[k] == true1 |
| 4391 | && mode_lib->vba.ProgressiveToInterlaceUnitInOPP == true1))) { |
| 4392 | locals->DIOSupport[i] = false0; |
| 4393 | } |
| 4394 | } |
| 4395 | } |
| 4396 | for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { |
| 4397 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4398 | locals->DSCCLKRequiredMoreThanSupported[i] = false0; |
| 4399 | if (mode_lib->vba.BlendingAndTiming[k] == k) { |
| 4400 | if ((mode_lib->vba.Output[k] == dm_dp |
| 4401 | || mode_lib->vba.Output[k] == dm_edp)) { |
| 4402 | if (mode_lib->vba.OutputFormat[k] == dm_420 |
| 4403 | || mode_lib->vba.OutputFormat[k] |
| 4404 | == dm_n422) { |
| 4405 | mode_lib->vba.DSCFormatFactor = 2; |
| 4406 | } else { |
| 4407 | mode_lib->vba.DSCFormatFactor = 1; |
| 4408 | } |
| 4409 | if (locals->RequiresDSC[i][k] == true1) { |
| 4410 | if (locals->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_2to1) { |
| 4411 | if (mode_lib->vba.PixelClockBackEnd[k] / 6.0 / mode_lib->vba.DSCFormatFactor |
| 4412 | > (1.0 - mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) * mode_lib->vba.MaxDSCCLK[i]) { |
| 4413 | locals->DSCCLKRequiredMoreThanSupported[i] = |
| 4414 | true1; |
| 4415 | } |
| 4416 | } else { |
| 4417 | if (mode_lib->vba.PixelClockBackEnd[k] / 3.0 / mode_lib->vba.DSCFormatFactor |
| 4418 | > (1.0 - mode_lib->vba.DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) * mode_lib->vba.MaxDSCCLK[i]) { |
| 4419 | locals->DSCCLKRequiredMoreThanSupported[i] = |
| 4420 | true1; |
| 4421 | } |
| 4422 | } |
| 4423 | } |
| 4424 | } |
| 4425 | } |
| 4426 | } |
| 4427 | } |
| 4428 | for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { |
| 4429 | locals->NotEnoughDSCUnits[i] = false0; |
| 4430 | mode_lib->vba.TotalDSCUnitsRequired = 0.0; |
| 4431 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4432 | if (locals->RequiresDSC[i][k] == true1) { |
| 4433 | if (locals->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_2to1) { |
| 4434 | mode_lib->vba.TotalDSCUnitsRequired = |
| 4435 | mode_lib->vba.TotalDSCUnitsRequired + 2.0; |
| 4436 | } else { |
| 4437 | mode_lib->vba.TotalDSCUnitsRequired = |
| 4438 | mode_lib->vba.TotalDSCUnitsRequired + 1.0; |
| 4439 | } |
| 4440 | } |
| 4441 | } |
| 4442 | if (mode_lib->vba.TotalDSCUnitsRequired > mode_lib->vba.NumberOfDSC) { |
| 4443 | locals->NotEnoughDSCUnits[i] = true1; |
| 4444 | } |
| 4445 | } |
| 4446 | /*DSC Delay per state*/ |
| 4447 | |
| 4448 | for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { |
| 4449 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4450 | if (mode_lib->vba.BlendingAndTiming[k] != k) { |
| 4451 | mode_lib->vba.slices = 0; |
| 4452 | } else if (locals->RequiresDSC[i][k] == 0 |
| 4453 | || locals->RequiresDSC[i][k] == false0) { |
| 4454 | mode_lib->vba.slices = 0; |
| 4455 | } else if (mode_lib->vba.PixelClockBackEnd[k] > 3200.0) { |
| 4456 | mode_lib->vba.slices = dml_ceil( |
| 4457 | mode_lib->vba.PixelClockBackEnd[k] / 400.0, |
| 4458 | 4.0); |
| 4459 | } else if (mode_lib->vba.PixelClockBackEnd[k] > 1360.0) { |
| 4460 | mode_lib->vba.slices = 8.0; |
| 4461 | } else if (mode_lib->vba.PixelClockBackEnd[k] > 680.0) { |
| 4462 | mode_lib->vba.slices = 4.0; |
| 4463 | } else if (mode_lib->vba.PixelClockBackEnd[k] > 340.0) { |
| 4464 | mode_lib->vba.slices = 2.0; |
| 4465 | } else { |
| 4466 | mode_lib->vba.slices = 1.0; |
| 4467 | } |
| 4468 | if (locals->OutputBppPerState[i][k] == BPP_BLENDED_PIPE0xffffffff |
| 4469 | || locals->OutputBppPerState[i][k] == BPP_INVALID0) { |
| 4470 | mode_lib->vba.bpp = 0.0; |
| 4471 | } else { |
| 4472 | mode_lib->vba.bpp = locals->OutputBppPerState[i][k]; |
| 4473 | } |
| 4474 | if (locals->RequiresDSC[i][k] == true1 && mode_lib->vba.bpp != 0.0) { |
| 4475 | if (locals->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_disabled) { |
| 4476 | locals->DSCDelayPerState[i][k] = |
| 4477 | dscceComputeDelay( |
| 4478 | mode_lib->vba.DSCInputBitPerComponent[k], |
| 4479 | mode_lib->vba.bpp, |
| 4480 | dml_ceil( |
| 4481 | mode_lib->vba.HActive[k] |
| 4482 | / mode_lib->vba.slices, |
| 4483 | 1.0), |
| 4484 | mode_lib->vba.slices, |
| 4485 | mode_lib->vba.OutputFormat[k]) |
| 4486 | + dscComputeDelay( |
| 4487 | mode_lib->vba.OutputFormat[k]); |
| 4488 | } else { |
| 4489 | locals->DSCDelayPerState[i][k] = |
| 4490 | 2.0 * (dscceComputeDelay( |
| 4491 | mode_lib->vba.DSCInputBitPerComponent[k], |
| 4492 | mode_lib->vba.bpp, |
| 4493 | dml_ceil(mode_lib->vba.HActive[k] / mode_lib->vba.slices, 1.0), |
| 4494 | mode_lib->vba.slices / 2, |
| 4495 | mode_lib->vba.OutputFormat[k]) |
| 4496 | + dscComputeDelay(mode_lib->vba.OutputFormat[k])); |
| 4497 | } |
| 4498 | locals->DSCDelayPerState[i][k] = |
| 4499 | locals->DSCDelayPerState[i][k] * mode_lib->vba.PixelClock[k] / mode_lib->vba.PixelClockBackEnd[k]; |
| 4500 | } else { |
| 4501 | locals->DSCDelayPerState[i][k] = 0.0; |
| 4502 | } |
| 4503 | } |
| 4504 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4505 | for (m = 0; m <= mode_lib->vba.NumberOfActivePlanes - 1; m++) { |
| 4506 | for (j = 0; j <= mode_lib->vba.NumberOfActivePlanes - 1; j++) { |
| 4507 | if (mode_lib->vba.BlendingAndTiming[k] == m && locals->RequiresDSC[i][m] == true1) |
| 4508 | locals->DSCDelayPerState[i][k] = locals->DSCDelayPerState[i][m]; |
| 4509 | } |
| 4510 | } |
| 4511 | } |
| 4512 | } |
| 4513 | |
| 4514 | //Prefetch Check |
| 4515 | for (i = 0; i <= mode_lib->vba.soc.num_states; ++i) { |
| 4516 | for (j = 0; j <= 1; ++j) { |
| 4517 | locals->TotalNumberOfDCCActiveDPP[i][j] = 0; |
| 4518 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 4519 | if (mode_lib->vba.DCCEnable[k] == true1) |
| 4520 | locals->TotalNumberOfDCCActiveDPP[i][j] = locals->TotalNumberOfDCCActiveDPP[i][j] + locals->NoOfDPP[i][j][k]; |
| 4521 | } |
| 4522 | } |
| 4523 | } |
| 4524 | |
| 4525 | mode_lib->vba.UrgentLatency = dml_max3( |
| 4526 | mode_lib->vba.UrgentLatencyPixelDataOnly, |
| 4527 | mode_lib->vba.UrgentLatencyPixelMixedWithVMData, |
| 4528 | mode_lib->vba.UrgentLatencyVMDataOnly); |
| 4529 | mode_lib->vba.PrefetchERROR = CalculateMinAndMaxPrefetchMode( |
| 4530 | mode_lib->vba.AllowDRAMSelfRefreshOrDRAMClockChangeInVblank, |
| 4531 | &mode_lib->vba.MinPrefetchMode, |
| 4532 | &mode_lib->vba.MaxPrefetchMode); |
| 4533 | |
| 4534 | for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { |
| 4535 | for (j = 0; j < 2; j++) { |
| 4536 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4537 | locals->RequiredDPPCLKThisState[k] = locals->RequiredDPPCLK[i][j][k]; |
| 4538 | locals->NoOfDPPThisState[k] = locals->NoOfDPP[i][j][k]; |
| 4539 | if (locals->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_2to1) { |
| 4540 | locals->SwathWidthYThisState[k] = |
| 4541 | dml_min(locals->SwathWidthYSingleDPP[k], dml_round(mode_lib->vba.HActive[k] / 2.0 * mode_lib->vba.HRatio[k])); |
| 4542 | } else { |
| 4543 | locals->SwathWidthYThisState[k] = locals->SwathWidthYSingleDPP[k] / locals->NoOfDPP[i][j][k]; |
| 4544 | } |
| 4545 | mode_lib->vba.SwathWidthGranularityY = 256.0 |
| 4546 | / dml_ceil(locals->BytePerPixelInDETY[k], 1.0) |
| 4547 | / locals->MaxSwathHeightY[k]; |
| 4548 | mode_lib->vba.RoundedUpMaxSwathSizeBytesY = |
| 4549 | (dml_ceil(locals->SwathWidthYThisState[k] - 1.0, mode_lib->vba.SwathWidthGranularityY) |
| 4550 | + mode_lib->vba.SwathWidthGranularityY) * locals->BytePerPixelInDETY[k] * locals->MaxSwathHeightY[k]; |
| 4551 | if (mode_lib->vba.SourcePixelFormat[k] == dm_420_10) { |
| 4552 | mode_lib->vba.RoundedUpMaxSwathSizeBytesY = dml_ceil( |
| 4553 | mode_lib->vba.RoundedUpMaxSwathSizeBytesY, |
| 4554 | 256.0) + 256; |
| 4555 | } |
| 4556 | if (locals->MaxSwathHeightC[k] > 0.0) { |
| 4557 | mode_lib->vba.SwathWidthGranularityC = 256.0 / dml_ceil(locals->BytePerPixelInDETC[k], 2.0) / locals->MaxSwathHeightC[k]; |
| 4558 | mode_lib->vba.RoundedUpMaxSwathSizeBytesC = (dml_ceil(locals->SwathWidthYThisState[k] / 2.0 - 1.0, mode_lib->vba.SwathWidthGranularityC) |
| 4559 | + mode_lib->vba.SwathWidthGranularityC) * locals->BytePerPixelInDETC[k] * locals->MaxSwathHeightC[k]; |
| 4560 | if (mode_lib->vba.SourcePixelFormat[k] == dm_420_10) { |
| 4561 | mode_lib->vba.RoundedUpMaxSwathSizeBytesC = dml_ceil(mode_lib->vba.RoundedUpMaxSwathSizeBytesC, 256.0) + 256; |
| 4562 | } |
| 4563 | } else { |
| 4564 | mode_lib->vba.RoundedUpMaxSwathSizeBytesC = 0.0; |
| 4565 | } |
| 4566 | if (mode_lib->vba.RoundedUpMaxSwathSizeBytesY + mode_lib->vba.RoundedUpMaxSwathSizeBytesC |
| 4567 | <= mode_lib->vba.DETBufferSizeInKByte[0] * 1024.0 / 2.0) { |
| 4568 | locals->SwathHeightYThisState[k] = locals->MaxSwathHeightY[k]; |
| 4569 | locals->SwathHeightCThisState[k] = locals->MaxSwathHeightC[k]; |
| 4570 | } else { |
| 4571 | locals->SwathHeightYThisState[k] = |
| 4572 | locals->MinSwathHeightY[k]; |
| 4573 | locals->SwathHeightCThisState[k] = |
| 4574 | locals->MinSwathHeightC[k]; |
| 4575 | } |
| 4576 | } |
| 4577 | |
| 4578 | CalculateDCFCLKDeepSleep( |
| 4579 | mode_lib, |
| 4580 | mode_lib->vba.NumberOfActivePlanes, |
| 4581 | locals->BytePerPixelInDETY, |
| 4582 | locals->BytePerPixelInDETC, |
| 4583 | mode_lib->vba.VRatio, |
| 4584 | locals->SwathWidthYThisState, |
| 4585 | locals->NoOfDPPThisState, |
| 4586 | mode_lib->vba.HRatio, |
| 4587 | mode_lib->vba.PixelClock, |
| 4588 | locals->PSCL_FACTOR, |
| 4589 | locals->PSCL_FACTOR_CHROMA, |
| 4590 | locals->RequiredDPPCLKThisState, |
| 4591 | &mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0]); |
| 4592 | |
| 4593 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4594 | if ((mode_lib->vba.SourcePixelFormat[k] != dm_444_64 |
| 4595 | && mode_lib->vba.SourcePixelFormat[k] != dm_444_32 |
| 4596 | && mode_lib->vba.SourcePixelFormat[k] != dm_444_16 |
| 4597 | && mode_lib->vba.SourcePixelFormat[k] != dm_mono_16 |
| 4598 | && mode_lib->vba.SourcePixelFormat[k] != dm_mono_8)) { |
| 4599 | mode_lib->vba.PDEAndMetaPTEBytesPerFrameC = CalculateVMAndRowBytes( |
| 4600 | mode_lib, |
| 4601 | mode_lib->vba.DCCEnable[k], |
| 4602 | locals->Read256BlockHeightC[k], |
| 4603 | locals->Read256BlockWidthC[k], |
| 4604 | mode_lib->vba.SourcePixelFormat[k], |
| 4605 | mode_lib->vba.SurfaceTiling[k], |
| 4606 | dml_ceil(locals->BytePerPixelInDETC[k], 2.0), |
| 4607 | mode_lib->vba.SourceScan[k], |
| 4608 | mode_lib->vba.ViewportWidth[k] / 2.0, |
| 4609 | mode_lib->vba.ViewportHeight[k] / 2.0, |
| 4610 | locals->SwathWidthYThisState[k] / 2.0, |
| 4611 | mode_lib->vba.GPUVMEnable, |
| 4612 | mode_lib->vba.HostVMEnable, |
| 4613 | mode_lib->vba.HostVMMaxPageTableLevels, |
| 4614 | mode_lib->vba.HostVMCachedPageTableLevels, |
| 4615 | mode_lib->vba.VMMPageSize, |
| 4616 | mode_lib->vba.PTEBufferSizeInRequestsChroma, |
| 4617 | mode_lib->vba.PitchC[k], |
| 4618 | 0.0, |
| 4619 | &locals->MacroTileWidthC[k], |
| 4620 | &mode_lib->vba.MetaRowBytesC, |
| 4621 | &mode_lib->vba.DPTEBytesPerRowC, |
| 4622 | &locals->PTEBufferSizeNotExceededC[i][j][k], |
| 4623 | locals->dpte_row_width_chroma_ub, |
| 4624 | &locals->dpte_row_height_chroma[k], |
| 4625 | &locals->meta_req_width_chroma[k], |
| 4626 | &locals->meta_req_height_chroma[k], |
| 4627 | &locals->meta_row_width_chroma[k], |
| 4628 | &locals->meta_row_height_chroma[k], |
| 4629 | &locals->vm_group_bytes_chroma, |
| 4630 | &locals->dpte_group_bytes_chroma, |
| 4631 | locals->PixelPTEReqWidthC, |
| 4632 | locals->PixelPTEReqHeightC, |
| 4633 | locals->PTERequestSizeC, |
| 4634 | locals->dpde0_bytes_per_frame_ub_c, |
| 4635 | locals->meta_pte_bytes_per_frame_ub_c); |
| 4636 | locals->PrefetchLinesC[0][0][k] = CalculatePrefetchSourceLines( |
| 4637 | mode_lib, |
| 4638 | mode_lib->vba.VRatio[k]/2, |
| 4639 | mode_lib->vba.VTAPsChroma[k], |
| 4640 | mode_lib->vba.Interlace[k], |
| 4641 | mode_lib->vba.ProgressiveToInterlaceUnitInOPP, |
| 4642 | locals->SwathHeightCThisState[k], |
| 4643 | mode_lib->vba.ViewportYStartC[k], |
| 4644 | &locals->PrefillC[k], |
| 4645 | &locals->MaxNumSwC[k]); |
| 4646 | locals->PTEBufferSizeInRequestsForLuma = mode_lib->vba.PTEBufferSizeInRequestsLuma; |
| 4647 | } else { |
| 4648 | mode_lib->vba.PDEAndMetaPTEBytesPerFrameC = 0.0; |
| 4649 | mode_lib->vba.MetaRowBytesC = 0.0; |
| 4650 | mode_lib->vba.DPTEBytesPerRowC = 0.0; |
| 4651 | locals->PrefetchLinesC[0][0][k] = 0.0; |
| 4652 | locals->PTEBufferSizeNotExceededC[i][j][k] = true1; |
| 4653 | locals->PTEBufferSizeInRequestsForLuma = mode_lib->vba.PTEBufferSizeInRequestsLuma + mode_lib->vba.PTEBufferSizeInRequestsChroma; |
| 4654 | } |
| 4655 | mode_lib->vba.PDEAndMetaPTEBytesPerFrameY = CalculateVMAndRowBytes( |
| 4656 | mode_lib, |
| 4657 | mode_lib->vba.DCCEnable[k], |
| 4658 | locals->Read256BlockHeightY[k], |
| 4659 | locals->Read256BlockWidthY[k], |
| 4660 | mode_lib->vba.SourcePixelFormat[k], |
| 4661 | mode_lib->vba.SurfaceTiling[k], |
| 4662 | dml_ceil(locals->BytePerPixelInDETY[k], 1.0), |
| 4663 | mode_lib->vba.SourceScan[k], |
| 4664 | mode_lib->vba.ViewportWidth[k], |
| 4665 | mode_lib->vba.ViewportHeight[k], |
| 4666 | locals->SwathWidthYThisState[k], |
| 4667 | mode_lib->vba.GPUVMEnable, |
| 4668 | mode_lib->vba.HostVMEnable, |
| 4669 | mode_lib->vba.HostVMMaxPageTableLevels, |
| 4670 | mode_lib->vba.HostVMCachedPageTableLevels, |
| 4671 | mode_lib->vba.VMMPageSize, |
| 4672 | locals->PTEBufferSizeInRequestsForLuma, |
| 4673 | mode_lib->vba.PitchY[k], |
| 4674 | mode_lib->vba.DCCMetaPitchY[k], |
| 4675 | &locals->MacroTileWidthY[k], |
| 4676 | &mode_lib->vba.MetaRowBytesY, |
| 4677 | &mode_lib->vba.DPTEBytesPerRowY, |
| 4678 | &locals->PTEBufferSizeNotExceededY[i][j][k], |
| 4679 | locals->dpte_row_width_luma_ub, |
| 4680 | &locals->dpte_row_height[k], |
| 4681 | &locals->meta_req_width[k], |
| 4682 | &locals->meta_req_height[k], |
| 4683 | &locals->meta_row_width[k], |
| 4684 | &locals->meta_row_height[k], |
| 4685 | &locals->vm_group_bytes[k], |
| 4686 | &locals->dpte_group_bytes[k], |
| 4687 | locals->PixelPTEReqWidthY, |
| 4688 | locals->PixelPTEReqHeightY, |
| 4689 | locals->PTERequestSizeY, |
| 4690 | locals->dpde0_bytes_per_frame_ub_l, |
| 4691 | locals->meta_pte_bytes_per_frame_ub_l); |
| 4692 | locals->PrefetchLinesY[0][0][k] = CalculatePrefetchSourceLines( |
| 4693 | mode_lib, |
| 4694 | mode_lib->vba.VRatio[k], |
| 4695 | mode_lib->vba.vtaps[k], |
| 4696 | mode_lib->vba.Interlace[k], |
| 4697 | mode_lib->vba.ProgressiveToInterlaceUnitInOPP, |
| 4698 | locals->SwathHeightYThisState[k], |
| 4699 | mode_lib->vba.ViewportYStartY[k], |
| 4700 | &locals->PrefillY[k], |
| 4701 | &locals->MaxNumSwY[k]); |
| 4702 | locals->PDEAndMetaPTEBytesPerFrame[0][0][k] = |
| 4703 | mode_lib->vba.PDEAndMetaPTEBytesPerFrameY + mode_lib->vba.PDEAndMetaPTEBytesPerFrameC; |
| 4704 | locals->MetaRowBytes[0][0][k] = mode_lib->vba.MetaRowBytesY + mode_lib->vba.MetaRowBytesC; |
| 4705 | locals->DPTEBytesPerRow[0][0][k] = mode_lib->vba.DPTEBytesPerRowY + mode_lib->vba.DPTEBytesPerRowC; |
| 4706 | |
| 4707 | CalculateActiveRowBandwidth( |
| 4708 | mode_lib->vba.GPUVMEnable, |
| 4709 | mode_lib->vba.SourcePixelFormat[k], |
| 4710 | mode_lib->vba.VRatio[k], |
| 4711 | mode_lib->vba.DCCEnable[k], |
| 4712 | mode_lib->vba.HTotal[k] / |
| 4713 | mode_lib->vba.PixelClock[k], |
| 4714 | mode_lib->vba.MetaRowBytesY, |
| 4715 | mode_lib->vba.MetaRowBytesC, |
| 4716 | locals->meta_row_height[k], |
| 4717 | locals->meta_row_height_chroma[k], |
| 4718 | mode_lib->vba.DPTEBytesPerRowY, |
| 4719 | mode_lib->vba.DPTEBytesPerRowC, |
| 4720 | locals->dpte_row_height[k], |
| 4721 | locals->dpte_row_height_chroma[k], |
| 4722 | &locals->meta_row_bw[k], |
| 4723 | &locals->dpte_row_bw[k]); |
| 4724 | } |
| 4725 | mode_lib->vba.ExtraLatency = CalculateExtraLatency( |
| 4726 | locals->UrgentRoundTripAndOutOfOrderLatencyPerState[i], |
| 4727 | locals->TotalNumberOfActiveDPP[i][j], |
| 4728 | mode_lib->vba.PixelChunkSizeInKByte, |
| 4729 | locals->TotalNumberOfDCCActiveDPP[i][j], |
| 4730 | mode_lib->vba.MetaChunkSize, |
| 4731 | locals->ReturnBWPerState[i][0], |
| 4732 | mode_lib->vba.GPUVMEnable, |
| 4733 | mode_lib->vba.HostVMEnable, |
| 4734 | mode_lib->vba.NumberOfActivePlanes, |
| 4735 | locals->NoOfDPPThisState, |
| 4736 | locals->dpte_group_bytes, |
| 4737 | mode_lib->vba.PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
| 4738 | mode_lib->vba.PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
| 4739 | mode_lib->vba.HostVMMaxPageTableLevels, |
| 4740 | mode_lib->vba.HostVMCachedPageTableLevels); |
| 4741 | |
| 4742 | mode_lib->vba.TimeCalc = 24.0 / mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0]; |
| 4743 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4744 | if (mode_lib->vba.BlendingAndTiming[k] == k) { |
| 4745 | if (mode_lib->vba.WritebackEnable[k] == true1) { |
| 4746 | locals->WritebackDelay[i][k] = mode_lib->vba.WritebackLatency |
| 4747 | + CalculateWriteBackDelay( |
| 4748 | mode_lib->vba.WritebackPixelFormat[k], |
| 4749 | mode_lib->vba.WritebackHRatio[k], |
| 4750 | mode_lib->vba.WritebackVRatio[k], |
| 4751 | mode_lib->vba.WritebackLumaHTaps[k], |
| 4752 | mode_lib->vba.WritebackLumaVTaps[k], |
| 4753 | mode_lib->vba.WritebackChromaHTaps[k], |
| 4754 | mode_lib->vba.WritebackChromaVTaps[k], |
| 4755 | mode_lib->vba.WritebackDestinationWidth[k]) / locals->RequiredDISPCLK[i][j]; |
| 4756 | } else { |
| 4757 | locals->WritebackDelay[i][k] = 0.0; |
| 4758 | } |
| 4759 | for (m = 0; m <= mode_lib->vba.NumberOfActivePlanes - 1; m++) { |
| 4760 | if (mode_lib->vba.BlendingAndTiming[m] == k |
| 4761 | && mode_lib->vba.WritebackEnable[m] |
| 4762 | == true1) { |
| 4763 | locals->WritebackDelay[i][k] = dml_max(locals->WritebackDelay[i][k], |
| 4764 | mode_lib->vba.WritebackLatency + CalculateWriteBackDelay( |
| 4765 | mode_lib->vba.WritebackPixelFormat[m], |
| 4766 | mode_lib->vba.WritebackHRatio[m], |
| 4767 | mode_lib->vba.WritebackVRatio[m], |
| 4768 | mode_lib->vba.WritebackLumaHTaps[m], |
| 4769 | mode_lib->vba.WritebackLumaVTaps[m], |
| 4770 | mode_lib->vba.WritebackChromaHTaps[m], |
| 4771 | mode_lib->vba.WritebackChromaVTaps[m], |
| 4772 | mode_lib->vba.WritebackDestinationWidth[m]) / locals->RequiredDISPCLK[i][j]); |
| 4773 | } |
| 4774 | } |
| 4775 | } |
| 4776 | } |
| 4777 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4778 | for (m = 0; m <= mode_lib->vba.NumberOfActivePlanes - 1; m++) { |
| 4779 | if (mode_lib->vba.BlendingAndTiming[k] == m) { |
| 4780 | locals->WritebackDelay[i][k] = locals->WritebackDelay[i][m]; |
| 4781 | } |
| 4782 | } |
| 4783 | } |
| 4784 | mode_lib->vba.MaxMaxVStartup[0][0] = 0; |
| 4785 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4786 | locals->MaximumVStartup[0][0][k] = mode_lib->vba.VTotal[k] - mode_lib->vba.VActive[k] |
| 4787 | - dml_max(1.0, dml_ceil(locals->WritebackDelay[i][k] / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]), 1.0)); |
| 4788 | mode_lib->vba.MaxMaxVStartup[0][0] = dml_max(mode_lib->vba.MaxMaxVStartup[0][0], locals->MaximumVStartup[0][0][k]); |
| 4789 | } |
| 4790 | |
| 4791 | mode_lib->vba.NextPrefetchMode = mode_lib->vba.MinPrefetchMode; |
| 4792 | mode_lib->vba.NextMaxVStartup = mode_lib->vba.MaxMaxVStartup[0][0]; |
| 4793 | do { |
| 4794 | mode_lib->vba.PrefetchMode[i][j] = mode_lib->vba.NextPrefetchMode; |
| 4795 | mode_lib->vba.MaxVStartup = mode_lib->vba.NextMaxVStartup; |
| 4796 | |
| 4797 | mode_lib->vba.TWait = CalculateTWait( |
| 4798 | mode_lib->vba.PrefetchMode[i][j], |
| 4799 | mode_lib->vba.DRAMClockChangeLatency, |
| 4800 | mode_lib->vba.UrgentLatency, |
| 4801 | mode_lib->vba.SREnterPlusExitTime); |
| 4802 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) |
| 4803 | CalculatePrefetchSchedulePerPlane(mode_lib, i, j, k); |
| 4804 | |
| 4805 | mode_lib->vba.MaximumReadBandwidthWithoutPrefetch = 0.0; |
| 4806 | mode_lib->vba.MaximumReadBandwidthWithPrefetch = 0.0; |
| 4807 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4808 | unsigned int m; |
| 4809 | |
| 4810 | locals->cursor_bw[k] = 0; |
| 4811 | locals->cursor_bw_pre[k] = 0; |
| 4812 | for (m = 0; m < mode_lib->vba.NumberOfCursors[k]; m++) { |
| 4813 | locals->cursor_bw[k] = mode_lib->vba.CursorWidth[k][m] * mode_lib->vba.CursorBPP[k][m] |
| 4814 | / 8.0 / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]) * mode_lib->vba.VRatio[k]; |
| 4815 | locals->cursor_bw_pre[k] = mode_lib->vba.CursorWidth[k][m] * mode_lib->vba.CursorBPP[k][m] |
| 4816 | / 8.0 / (mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k]) * locals->VRatioPreY[i][j][k]; |
| 4817 | } |
| 4818 | |
| 4819 | CalculateUrgentBurstFactor( |
| 4820 | mode_lib->vba.DETBufferSizeInKByte[0], |
| 4821 | locals->SwathHeightYThisState[k], |
| 4822 | locals->SwathHeightCThisState[k], |
| 4823 | locals->SwathWidthYThisState[k], |
| 4824 | mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k], |
| 4825 | mode_lib->vba.UrgentLatency, |
| 4826 | mode_lib->vba.CursorBufferSize, |
| 4827 | mode_lib->vba.CursorWidth[k][0] + mode_lib->vba.CursorWidth[k][1], |
| 4828 | dml_max(mode_lib->vba.CursorBPP[k][0], mode_lib->vba.CursorBPP[k][1]), |
| 4829 | mode_lib->vba.VRatio[k], |
| 4830 | locals->VRatioPreY[i][j][k], |
| 4831 | locals->VRatioPreC[i][j][k], |
| 4832 | locals->BytePerPixelInDETY[k], |
| 4833 | locals->BytePerPixelInDETC[k], |
| 4834 | &locals->UrgentBurstFactorCursor[k], |
| 4835 | &locals->UrgentBurstFactorCursorPre[k], |
| 4836 | &locals->UrgentBurstFactorLuma[k], |
| 4837 | &locals->UrgentBurstFactorLumaPre[k], |
| 4838 | &locals->UrgentBurstFactorChroma[k], |
| 4839 | &locals->UrgentBurstFactorChromaPre[k], |
| 4840 | &locals->NotEnoughUrgentLatencyHiding[0][0], |
| 4841 | &locals->NotEnoughUrgentLatencyHidingPre); |
| 4842 | |
| 4843 | if (mode_lib->vba.UseUrgentBurstBandwidth == false0) { |
| 4844 | locals->UrgentBurstFactorCursor[k] = 1; |
| 4845 | locals->UrgentBurstFactorCursorPre[k] = 1; |
| 4846 | locals->UrgentBurstFactorLuma[k] = 1; |
| 4847 | locals->UrgentBurstFactorLumaPre[k] = 1; |
| 4848 | locals->UrgentBurstFactorChroma[k] = 1; |
| 4849 | locals->UrgentBurstFactorChromaPre[k] = 1; |
| 4850 | } |
| 4851 | |
| 4852 | mode_lib->vba.MaximumReadBandwidthWithoutPrefetch = mode_lib->vba.MaximumReadBandwidthWithoutPrefetch |
| 4853 | + locals->cursor_bw[k] * locals->UrgentBurstFactorCursor[k] + locals->ReadBandwidthLuma[k] |
| 4854 | * locals->UrgentBurstFactorLuma[k] + locals->ReadBandwidthChroma[k] |
| 4855 | * locals->UrgentBurstFactorChroma[k] + locals->meta_row_bw[k] + locals->dpte_row_bw[k]; |
| 4856 | mode_lib->vba.MaximumReadBandwidthWithPrefetch = mode_lib->vba.MaximumReadBandwidthWithPrefetch |
| 4857 | + dml_max3(locals->prefetch_vmrow_bw[k], |
| 4858 | locals->ReadBandwidthLuma[k] * locals->UrgentBurstFactorLuma[k] + locals->ReadBandwidthChroma[k] |
| 4859 | * locals->UrgentBurstFactorChroma[k] + locals->cursor_bw[k] * locals->UrgentBurstFactorCursor[k] |
| 4860 | + locals->meta_row_bw[k] + locals->dpte_row_bw[k], |
| 4861 | locals->RequiredPrefetchPixelDataBWLuma[i][j][k] * locals->UrgentBurstFactorLumaPre[k] |
| 4862 | + locals->RequiredPrefetchPixelDataBWChroma[i][j][k] * locals->UrgentBurstFactorChromaPre[k] |
| 4863 | + locals->cursor_bw_pre[k] * locals->UrgentBurstFactorCursorPre[k]); |
| 4864 | } |
| 4865 | locals->BandwidthWithoutPrefetchSupported[i][0] = true1; |
| 4866 | if (mode_lib->vba.MaximumReadBandwidthWithoutPrefetch > locals->ReturnBWPerState[i][0] |
| 4867 | || locals->NotEnoughUrgentLatencyHiding[0][0] == 1) { |
| 4868 | locals->BandwidthWithoutPrefetchSupported[i][0] = false0; |
| 4869 | } |
| 4870 | |
| 4871 | locals->PrefetchSupported[i][j] = true1; |
| 4872 | if (mode_lib->vba.MaximumReadBandwidthWithPrefetch > locals->ReturnBWPerState[i][0] |
| 4873 | || locals->NotEnoughUrgentLatencyHiding[0][0] == 1 |
| 4874 | || locals->NotEnoughUrgentLatencyHidingPre == 1) { |
| 4875 | locals->PrefetchSupported[i][j] = false0; |
| 4876 | } |
| 4877 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4878 | if (locals->LineTimesForPrefetch[k] < 2.0 |
| 4879 | || locals->LinesForMetaPTE[k] >= 32.0 |
| 4880 | || locals->LinesForMetaAndDPTERow[k] >= 16.0 |
| 4881 | || mode_lib->vba.IsErrorResult[i][j][k] == true1) { |
| 4882 | locals->PrefetchSupported[i][j] = false0; |
| 4883 | } |
| 4884 | } |
| 4885 | locals->VRatioInPrefetchSupported[i][j] = true1; |
| 4886 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4887 | if (locals->VRatioPreY[i][j][k] > 4.0 |
| 4888 | || locals->VRatioPreC[i][j][k] > 4.0 |
| 4889 | || mode_lib->vba.IsErrorResult[i][j][k] == true1) { |
| 4890 | locals->VRatioInPrefetchSupported[i][j] = false0; |
| 4891 | } |
| 4892 | } |
| 4893 | mode_lib->vba.AnyLinesForVMOrRowTooLarge = false0; |
| 4894 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 4895 | if (locals->LinesForMetaAndDPTERow[k] >= 16 || locals->LinesForMetaPTE[k] >= 32) { |
| 4896 | mode_lib->vba.AnyLinesForVMOrRowTooLarge = true1; |
| 4897 | } |
| 4898 | } |
| 4899 | |
| 4900 | if (mode_lib->vba.MaxVStartup <= 13 || mode_lib->vba.AnyLinesForVMOrRowTooLarge == false0) { |
| 4901 | mode_lib->vba.NextMaxVStartup = mode_lib->vba.MaxMaxVStartup[0][0]; |
| 4902 | mode_lib->vba.NextPrefetchMode = mode_lib->vba.NextPrefetchMode + 1; |
| 4903 | } else { |
| 4904 | mode_lib->vba.NextMaxVStartup = mode_lib->vba.NextMaxVStartup - 1; |
| 4905 | } |
| 4906 | } while ((locals->PrefetchSupported[i][j] != true1 || locals->VRatioInPrefetchSupported[i][j] != true1) |
| 4907 | && (mode_lib->vba.NextMaxVStartup != mode_lib->vba.MaxMaxVStartup[0][0] |
| 4908 | || mode_lib->vba.NextPrefetchMode <= mode_lib->vba.MaxPrefetchMode)); |
| 4909 | |
| 4910 | if (locals->PrefetchSupported[i][j] == true1 && locals->VRatioInPrefetchSupported[i][j] == true1) { |
| 4911 | mode_lib->vba.BandwidthAvailableForImmediateFlip = locals->ReturnBWPerState[i][0]; |
| 4912 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4913 | mode_lib->vba.BandwidthAvailableForImmediateFlip = mode_lib->vba.BandwidthAvailableForImmediateFlip |
| 4914 | - dml_max(locals->ReadBandwidthLuma[k] * locals->UrgentBurstFactorLuma[k] |
| 4915 | + locals->ReadBandwidthChroma[k] * locals->UrgentBurstFactorChroma[k] |
| 4916 | + locals->cursor_bw[k] * locals->UrgentBurstFactorCursor[k], |
| 4917 | locals->RequiredPrefetchPixelDataBWLuma[i][j][k] * locals->UrgentBurstFactorLumaPre[k] |
| 4918 | + locals->RequiredPrefetchPixelDataBWChroma[i][j][k] * locals->UrgentBurstFactorChromaPre[k] |
| 4919 | + locals->cursor_bw_pre[k] * locals->UrgentBurstFactorCursorPre[k]); |
| 4920 | } |
| 4921 | mode_lib->vba.TotImmediateFlipBytes = 0.0; |
| 4922 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4923 | mode_lib->vba.TotImmediateFlipBytes = mode_lib->vba.TotImmediateFlipBytes |
| 4924 | + locals->PDEAndMetaPTEBytesPerFrame[0][0][k] + locals->MetaRowBytes[0][0][k] + locals->DPTEBytesPerRow[0][0][k]; |
| 4925 | } |
| 4926 | |
| 4927 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4928 | CalculateFlipSchedule( |
| 4929 | mode_lib, |
| 4930 | mode_lib->vba.PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
| 4931 | mode_lib->vba.PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
| 4932 | mode_lib->vba.ExtraLatency, |
| 4933 | mode_lib->vba.UrgentLatency, |
| 4934 | mode_lib->vba.GPUVMMaxPageTableLevels, |
| 4935 | mode_lib->vba.HostVMEnable, |
| 4936 | mode_lib->vba.HostVMMaxPageTableLevels, |
| 4937 | mode_lib->vba.HostVMCachedPageTableLevels, |
| 4938 | mode_lib->vba.GPUVMEnable, |
| 4939 | locals->PDEAndMetaPTEBytesPerFrame[0][0][k], |
| 4940 | locals->MetaRowBytes[0][0][k], |
| 4941 | locals->DPTEBytesPerRow[0][0][k], |
| 4942 | mode_lib->vba.BandwidthAvailableForImmediateFlip, |
| 4943 | mode_lib->vba.TotImmediateFlipBytes, |
| 4944 | mode_lib->vba.SourcePixelFormat[k], |
| 4945 | mode_lib->vba.HTotal[k] / mode_lib->vba.PixelClock[k], |
| 4946 | mode_lib->vba.VRatio[k], |
| 4947 | locals->Tno_bw[k], |
| 4948 | mode_lib->vba.DCCEnable[k], |
| 4949 | locals->dpte_row_height[k], |
| 4950 | locals->meta_row_height[k], |
| 4951 | locals->dpte_row_height_chroma[k], |
| 4952 | locals->meta_row_height_chroma[k], |
| 4953 | &locals->DestinationLinesToRequestVMInImmediateFlip[k], |
| 4954 | &locals->DestinationLinesToRequestRowInImmediateFlip[k], |
| 4955 | &locals->final_flip_bw[k], |
| 4956 | &locals->ImmediateFlipSupportedForPipe[k]); |
| 4957 | } |
| 4958 | mode_lib->vba.total_dcn_read_bw_with_flip = 0.0; |
| 4959 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4960 | mode_lib->vba.total_dcn_read_bw_with_flip = mode_lib->vba.total_dcn_read_bw_with_flip + dml_max3( |
| 4961 | locals->prefetch_vmrow_bw[k], |
| 4962 | locals->final_flip_bw[k] + locals->ReadBandwidthLuma[k] * locals->UrgentBurstFactorLuma[k] |
| 4963 | + locals->ReadBandwidthChroma[k] * locals->UrgentBurstFactorChroma[k] |
| 4964 | + locals->cursor_bw[k] * locals->UrgentBurstFactorCursor[k], |
| 4965 | locals->final_flip_bw[k] + locals->RequiredPrefetchPixelDataBWLuma[i][j][k] |
| 4966 | * locals->UrgentBurstFactorLumaPre[k] + locals->RequiredPrefetchPixelDataBWChroma[i][j][k] |
| 4967 | * locals->UrgentBurstFactorChromaPre[k] + locals->cursor_bw_pre[k] |
| 4968 | * locals->UrgentBurstFactorCursorPre[k]); |
| 4969 | } |
| 4970 | locals->ImmediateFlipSupportedForState[i][j] = true1; |
| 4971 | if (mode_lib->vba.total_dcn_read_bw_with_flip |
| 4972 | > locals->ReturnBWPerState[i][0]) { |
| 4973 | locals->ImmediateFlipSupportedForState[i][j] = false0; |
| 4974 | } |
| 4975 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 4976 | if (locals->ImmediateFlipSupportedForPipe[k] == false0) { |
| 4977 | locals->ImmediateFlipSupportedForState[i][j] = false0; |
| 4978 | } |
| 4979 | } |
| 4980 | } else { |
| 4981 | locals->ImmediateFlipSupportedForState[i][j] = false0; |
| 4982 | } |
| 4983 | mode_lib->vba.UrgentOutOfOrderReturnPerChannel = dml_max3( |
| 4984 | mode_lib->vba.UrgentOutOfOrderReturnPerChannelPixelDataOnly, |
| 4985 | mode_lib->vba.UrgentOutOfOrderReturnPerChannelPixelMixedWithVMData, |
| 4986 | mode_lib->vba.UrgentOutOfOrderReturnPerChannelVMDataOnly); |
| 4987 | CalculateWatermarksAndDRAMSpeedChangeSupport( |
| 4988 | mode_lib, |
| 4989 | mode_lib->vba.PrefetchMode[i][j], |
| 4990 | mode_lib->vba.NumberOfActivePlanes, |
| 4991 | mode_lib->vba.MaxLineBufferLines, |
| 4992 | mode_lib->vba.LineBufferSize, |
| 4993 | mode_lib->vba.DPPOutputBufferPixels, |
| 4994 | mode_lib->vba.DETBufferSizeInKByte[0], |
| 4995 | mode_lib->vba.WritebackInterfaceLumaBufferSize, |
| 4996 | mode_lib->vba.WritebackInterfaceChromaBufferSize, |
| 4997 | mode_lib->vba.DCFCLKPerState[i], |
| 4998 | mode_lib->vba.UrgentOutOfOrderReturnPerChannel * mode_lib->vba.NumberOfChannels, |
| 4999 | locals->ReturnBWPerState[i][0], |
| 5000 | mode_lib->vba.GPUVMEnable, |
| 5001 | locals->dpte_group_bytes, |
| 5002 | mode_lib->vba.MetaChunkSize, |
| 5003 | mode_lib->vba.UrgentLatency, |
| 5004 | mode_lib->vba.ExtraLatency, |
| 5005 | mode_lib->vba.WritebackLatency, |
| 5006 | mode_lib->vba.WritebackChunkSize, |
| 5007 | mode_lib->vba.SOCCLKPerState[i], |
| 5008 | mode_lib->vba.DRAMClockChangeLatency, |
| 5009 | mode_lib->vba.SRExitTime, |
| 5010 | mode_lib->vba.SREnterPlusExitTime, |
| 5011 | mode_lib->vba.ProjectedDCFCLKDeepSleep[0][0], |
| 5012 | locals->NoOfDPPThisState, |
| 5013 | mode_lib->vba.DCCEnable, |
| 5014 | locals->RequiredDPPCLKThisState, |
| 5015 | locals->SwathWidthYSingleDPP, |
| 5016 | locals->SwathHeightYThisState, |
| 5017 | locals->ReadBandwidthLuma, |
| 5018 | locals->SwathHeightCThisState, |
| 5019 | locals->ReadBandwidthChroma, |
| 5020 | mode_lib->vba.LBBitPerPixel, |
| 5021 | locals->SwathWidthYThisState, |
| 5022 | mode_lib->vba.HRatio, |
| 5023 | mode_lib->vba.vtaps, |
| 5024 | mode_lib->vba.VTAPsChroma, |
| 5025 | mode_lib->vba.VRatio, |
| 5026 | mode_lib->vba.HTotal, |
| 5027 | mode_lib->vba.PixelClock, |
| 5028 | mode_lib->vba.BlendingAndTiming, |
| 5029 | locals->BytePerPixelInDETY, |
| 5030 | locals->BytePerPixelInDETC, |
| 5031 | mode_lib->vba.WritebackEnable, |
| 5032 | mode_lib->vba.WritebackPixelFormat, |
| 5033 | mode_lib->vba.WritebackDestinationWidth, |
| 5034 | mode_lib->vba.WritebackDestinationHeight, |
| 5035 | mode_lib->vba.WritebackSourceHeight, |
| 5036 | &locals->DRAMClockChangeSupport[i][j], |
| 5037 | &mode_lib->vba.UrgentWatermark, |
| 5038 | &mode_lib->vba.WritebackUrgentWatermark, |
| 5039 | &mode_lib->vba.DRAMClockChangeWatermark, |
| 5040 | &mode_lib->vba.WritebackDRAMClockChangeWatermark, |
| 5041 | &mode_lib->vba.StutterExitWatermark, |
| 5042 | &mode_lib->vba.StutterEnterPlusExitWatermark, |
| 5043 | &mode_lib->vba.MinActiveDRAMClockChangeLatencySupported); |
| 5044 | } |
| 5045 | } |
| 5046 | |
| 5047 | /*Vertical Active BW support*/ |
| 5048 | { |
| 5049 | double MaxTotalVActiveRDBandwidth = 0.0; |
| 5050 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
| 5051 | MaxTotalVActiveRDBandwidth = MaxTotalVActiveRDBandwidth + locals->ReadBandwidth[k]; |
| 5052 | } |
| 5053 | for (i = 0; i <= mode_lib->vba.soc.num_states; ++i) { |
| 5054 | locals->MaxTotalVerticalActiveAvailableBandwidth[i][0] = dml_min( |
| 5055 | locals->IdealSDPPortBandwidthPerState[i][0] * |
| 5056 | mode_lib->vba.MaxAveragePercentOfIdealSDPPortBWDisplayCanUseInNormalSystemOperation |
| 5057 | / 100.0, mode_lib->vba.DRAMSpeedPerState[i] * |
| 5058 | mode_lib->vba.NumberOfChannels * |
| 5059 | mode_lib->vba.DRAMChannelWidth * |
| 5060 | mode_lib->vba.MaxAveragePercentOfIdealDRAMBWDisplayCanUseInNormalSystemOperation |
| 5061 | / 100.0); |
| 5062 | |
| 5063 | if (MaxTotalVActiveRDBandwidth <= locals->MaxTotalVerticalActiveAvailableBandwidth[i][0]) { |
| 5064 | locals->TotalVerticalActiveBandwidthSupport[i][0] = true1; |
| 5065 | } else { |
| 5066 | locals->TotalVerticalActiveBandwidthSupport[i][0] = false0; |
| 5067 | } |
| 5068 | } |
| 5069 | } |
| 5070 | |
| 5071 | /*PTE Buffer Size Check*/ |
| 5072 | |
| 5073 | for (i = 0; i <= mode_lib->vba.soc.num_states; i++) { |
| 5074 | for (j = 0; j < 2; j++) { |
| 5075 | locals->PTEBufferSizeNotExceeded[i][j] = true1; |
| 5076 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 5077 | if (locals->PTEBufferSizeNotExceededY[i][j][k] == false0 |
| 5078 | || locals->PTEBufferSizeNotExceededC[i][j][k] == false0) { |
| 5079 | locals->PTEBufferSizeNotExceeded[i][j] = false0; |
| 5080 | } |
| 5081 | } |
| 5082 | } |
| 5083 | } |
| 5084 | /*Cursor Support Check*/ |
| 5085 | |
| 5086 | mode_lib->vba.CursorSupport = true1; |
| 5087 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 5088 | if (mode_lib->vba.CursorWidth[k][0] > 0.0) { |
| 5089 | for (m = 0; m < mode_lib->vba.NumberOfCursors[k]; m++) { |
| 5090 | if (mode_lib->vba.CursorBPP[k][m] == 64 && mode_lib->vba.Cursor64BppSupport == false0) { |
| 5091 | mode_lib->vba.CursorSupport = false0; |
| 5092 | } |
| 5093 | } |
| 5094 | } |
| 5095 | } |
| 5096 | /*Valid Pitch Check*/ |
| 5097 | |
| 5098 | mode_lib->vba.PitchSupport = true1; |
| 5099 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 5100 | locals->AlignedYPitch[k] = dml_ceil( |
| 5101 | dml_max(mode_lib->vba.PitchY[k], mode_lib->vba.ViewportWidth[k]), |
| 5102 | locals->MacroTileWidthY[k]); |
| 5103 | if (locals->AlignedYPitch[k] > mode_lib->vba.PitchY[k]) { |
| 5104 | mode_lib->vba.PitchSupport = false0; |
| 5105 | } |
| 5106 | if (mode_lib->vba.DCCEnable[k] == true1) { |
| 5107 | locals->AlignedDCCMetaPitch[k] = dml_ceil( |
| 5108 | dml_max( |
| 5109 | mode_lib->vba.DCCMetaPitchY[k], |
| 5110 | mode_lib->vba.ViewportWidth[k]), |
| 5111 | 64.0 * locals->Read256BlockWidthY[k]); |
| 5112 | } else { |
| 5113 | locals->AlignedDCCMetaPitch[k] = mode_lib->vba.DCCMetaPitchY[k]; |
| 5114 | } |
| 5115 | if (locals->AlignedDCCMetaPitch[k] > mode_lib->vba.DCCMetaPitchY[k]) { |
| 5116 | mode_lib->vba.PitchSupport = false0; |
| 5117 | } |
| 5118 | if (mode_lib->vba.SourcePixelFormat[k] != dm_444_64 |
| 5119 | && mode_lib->vba.SourcePixelFormat[k] != dm_444_32 |
| 5120 | && mode_lib->vba.SourcePixelFormat[k] != dm_444_16 |
| 5121 | && mode_lib->vba.SourcePixelFormat[k] != dm_mono_16 |
| 5122 | && mode_lib->vba.SourcePixelFormat[k] != dm_mono_8) { |
| 5123 | locals->AlignedCPitch[k] = dml_ceil( |
| 5124 | dml_max( |
| 5125 | mode_lib->vba.PitchC[k], |
| 5126 | mode_lib->vba.ViewportWidth[k] / 2.0), |
| 5127 | locals->MacroTileWidthC[k]); |
| 5128 | } else { |
| 5129 | locals->AlignedCPitch[k] = mode_lib->vba.PitchC[k]; |
| 5130 | } |
| 5131 | if (locals->AlignedCPitch[k] > mode_lib->vba.PitchC[k]) { |
| 5132 | mode_lib->vba.PitchSupport = false0; |
| 5133 | } |
| 5134 | } |
| 5135 | /*Mode Support, Voltage State and SOC Configuration*/ |
| 5136 | |
| 5137 | for (i = mode_lib->vba.soc.num_states; i >= 0; i--) { |
| 5138 | for (j = 0; j < 2; j++) { |
| 5139 | enum dm_validation_status status = DML_VALIDATION_OK; |
| 5140 | |
| 5141 | if (!mode_lib->vba.ScaleRatioAndTapsSupport) { |
| 5142 | status = DML_FAIL_SCALE_RATIO_TAP; |
| 5143 | } else if (!mode_lib->vba.SourceFormatPixelAndScanSupport) { |
| 5144 | status = DML_FAIL_SOURCE_PIXEL_FORMAT; |
| 5145 | } else if (!locals->ViewportSizeSupport[i][0]) { |
| 5146 | status = DML_FAIL_VIEWPORT_SIZE; |
| 5147 | } else if (!locals->DIOSupport[i]) { |
| 5148 | status = DML_FAIL_DIO_SUPPORT; |
| 5149 | } else if (locals->NotEnoughDSCUnits[i]) { |
| 5150 | status = DML_FAIL_NOT_ENOUGH_DSC; |
| 5151 | } else if (locals->DSCCLKRequiredMoreThanSupported[i]) { |
| 5152 | status = DML_FAIL_DSC_CLK_REQUIRED; |
| 5153 | } else if (!locals->ROBSupport[i][0]) { |
| 5154 | status = DML_FAIL_REORDERING_BUFFER; |
| 5155 | } else if (!locals->DISPCLK_DPPCLK_Support[i][j]) { |
| 5156 | status = DML_FAIL_DISPCLK_DPPCLK; |
| 5157 | } else if (!locals->TotalAvailablePipesSupport[i][j]) { |
| 5158 | status = DML_FAIL_TOTAL_AVAILABLE_PIPES; |
| 5159 | } else if (!mode_lib->vba.NumberOfOTGSupport) { |
| 5160 | status = DML_FAIL_NUM_OTG; |
| 5161 | } else if (!mode_lib->vba.WritebackModeSupport) { |
| 5162 | status = DML_FAIL_WRITEBACK_MODE; |
| 5163 | } else if (!mode_lib->vba.WritebackLatencySupport) { |
| 5164 | status = DML_FAIL_WRITEBACK_LATENCY; |
| 5165 | } else if (!mode_lib->vba.WritebackScaleRatioAndTapsSupport) { |
| 5166 | status = DML_FAIL_WRITEBACK_SCALE_RATIO_TAP; |
| 5167 | } else if (!mode_lib->vba.CursorSupport) { |
| 5168 | status = DML_FAIL_CURSOR_SUPPORT; |
| 5169 | } else if (!mode_lib->vba.PitchSupport) { |
| 5170 | status = DML_FAIL_PITCH_SUPPORT; |
| 5171 | } else if (!locals->TotalVerticalActiveBandwidthSupport[i][0]) { |
| 5172 | status = DML_FAIL_TOTAL_V_ACTIVE_BW; |
| 5173 | } else if (!locals->PTEBufferSizeNotExceeded[i][j]) { |
| 5174 | status = DML_FAIL_PTE_BUFFER_SIZE; |
| 5175 | } else if (mode_lib->vba.NonsupportedDSCInputBPC) { |
| 5176 | status = DML_FAIL_DSC_INPUT_BPC; |
| 5177 | } else if ((mode_lib->vba.HostVMEnable |
| 5178 | && !locals->ImmediateFlipSupportedForState[i][j])) { |
| 5179 | status = DML_FAIL_HOST_VM_IMMEDIATE_FLIP; |
| 5180 | } else if (!locals->PrefetchSupported[i][j]) { |
| 5181 | status = DML_FAIL_PREFETCH_SUPPORT; |
| 5182 | } else if (!locals->VRatioInPrefetchSupported[i][j]) { |
| 5183 | status = DML_FAIL_V_RATIO_PREFETCH; |
| 5184 | } |
| 5185 | |
| 5186 | if (status == DML_VALIDATION_OK) { |
| 5187 | locals->ModeSupport[i][j] = true1; |
| 5188 | } else { |
| 5189 | locals->ModeSupport[i][j] = false0; |
| 5190 | } |
| 5191 | locals->ValidationStatus[i] = status; |
| 5192 | } |
| 5193 | } |
| 5194 | { |
| 5195 | unsigned int MaximumMPCCombine = 0; |
| 5196 | mode_lib->vba.VoltageLevel = mode_lib->vba.soc.num_states + 1; |
| 5197 | for (i = mode_lib->vba.VoltageOverrideLevel; i <= mode_lib->vba.soc.num_states; i++) { |
| 5198 | if (locals->ModeSupport[i][0] == true1 || locals->ModeSupport[i][1] == true1) { |
| 5199 | mode_lib->vba.VoltageLevel = i; |
| 5200 | if (locals->ModeSupport[i][1] == true1 && (locals->ModeSupport[i][0] == false0 |
| 5201 | || mode_lib->vba.WhenToDoMPCCombine == dm_mpc_always_when_possible |
| 5202 | || (mode_lib->vba.WhenToDoMPCCombine == dm_mpc_reduce_voltage_and_clocks |
| 5203 | && ((locals->DRAMClockChangeSupport[i][1] == dm_dram_clock_change_vactive |
| 5204 | && locals->DRAMClockChangeSupport[i][0] != dm_dram_clock_change_vactive) |
| 5205 | || (locals->DRAMClockChangeSupport[i][1] == dm_dram_clock_change_vblank |
| 5206 | && locals->DRAMClockChangeSupport[i][0] == dm_dram_clock_change_unsupported))))) { |
| 5207 | MaximumMPCCombine = 1; |
| 5208 | } else { |
| 5209 | MaximumMPCCombine = 0; |
| 5210 | } |
| 5211 | break; |
| 5212 | } |
| 5213 | } |
| 5214 | mode_lib->vba.ImmediateFlipSupport = |
| 5215 | locals->ImmediateFlipSupportedForState[mode_lib->vba.VoltageLevel][MaximumMPCCombine]; |
| 5216 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 5217 | mode_lib->vba.DPPPerPlane[k] = locals->NoOfDPP[mode_lib->vba.VoltageLevel][MaximumMPCCombine][k]; |
| 5218 | locals->DPPCLK[k] = locals->RequiredDPPCLK[mode_lib->vba.VoltageLevel][MaximumMPCCombine][k]; |
| 5219 | } |
| 5220 | mode_lib->vba.DISPCLK = locals->RequiredDISPCLK[mode_lib->vba.VoltageLevel][MaximumMPCCombine]; |
| 5221 | mode_lib->vba.maxMpcComb = MaximumMPCCombine; |
| 5222 | } |
| 5223 | mode_lib->vba.DCFCLK = mode_lib->vba.DCFCLKPerState[mode_lib->vba.VoltageLevel]; |
| 5224 | mode_lib->vba.DRAMSpeed = mode_lib->vba.DRAMSpeedPerState[mode_lib->vba.VoltageLevel]; |
| 5225 | mode_lib->vba.FabricClock = mode_lib->vba.FabricClockPerState[mode_lib->vba.VoltageLevel]; |
| 5226 | mode_lib->vba.SOCCLK = mode_lib->vba.SOCCLKPerState[mode_lib->vba.VoltageLevel]; |
| 5227 | mode_lib->vba.ReturnBW = locals->ReturnBWPerState[mode_lib->vba.VoltageLevel][0]; |
| 5228 | for (k = 0; k <= mode_lib->vba.NumberOfActivePlanes - 1; k++) { |
| 5229 | if (mode_lib->vba.BlendingAndTiming[k] == k) { |
| 5230 | mode_lib->vba.ODMCombineEnabled[k] = |
| 5231 | locals->ODMCombineEnablePerState[mode_lib->vba.VoltageLevel][k]; |
| 5232 | } else { |
| 5233 | mode_lib->vba.ODMCombineEnabled[k] = dm_odm_combine_mode_disabled; |
| 5234 | } |
| 5235 | mode_lib->vba.DSCEnabled[k] = |
| 5236 | locals->RequiresDSC[mode_lib->vba.VoltageLevel][k]; |
| 5237 | mode_lib->vba.OutputBpp[k] = |
| 5238 | locals->OutputBppPerState[mode_lib->vba.VoltageLevel][k]; |
| 5239 | } |
| 5240 | } |
| 5241 | |
| 5242 | static void CalculateWatermarksAndDRAMSpeedChangeSupport( |
| 5243 | struct display_mode_lib *mode_lib, |
| 5244 | unsigned int PrefetchMode, |
| 5245 | unsigned int NumberOfActivePlanes, |
| 5246 | unsigned int MaxLineBufferLines, |
| 5247 | unsigned int LineBufferSize, |
| 5248 | unsigned int DPPOutputBufferPixels, |
| 5249 | unsigned int DETBufferSizeInKByte, |
| 5250 | unsigned int WritebackInterfaceLumaBufferSize, |
| 5251 | unsigned int WritebackInterfaceChromaBufferSize, |
| 5252 | double DCFCLK, |
| 5253 | double UrgentOutOfOrderReturn, |
| 5254 | double ReturnBW, |
| 5255 | bool_Bool GPUVMEnable, |
| 5256 | int dpte_group_bytes[], |
| 5257 | unsigned int MetaChunkSize, |
| 5258 | double UrgentLatency, |
| 5259 | double ExtraLatency, |
| 5260 | double WritebackLatency, |
| 5261 | double WritebackChunkSize, |
| 5262 | double SOCCLK, |
| 5263 | double DRAMClockChangeLatency, |
| 5264 | double SRExitTime, |
| 5265 | double SREnterPlusExitTime, |
| 5266 | double DCFCLKDeepSleep, |
| 5267 | int DPPPerPlane[], |
| 5268 | bool_Bool DCCEnable[], |
| 5269 | double DPPCLK[], |
| 5270 | double SwathWidthSingleDPPY[], |
| 5271 | unsigned int SwathHeightY[], |
| 5272 | double ReadBandwidthPlaneLuma[], |
| 5273 | unsigned int SwathHeightC[], |
| 5274 | double ReadBandwidthPlaneChroma[], |
| 5275 | unsigned int LBBitPerPixel[], |
| 5276 | double SwathWidthY[], |
| 5277 | double HRatio[], |
| 5278 | unsigned int vtaps[], |
| 5279 | unsigned int VTAPsChroma[], |
| 5280 | double VRatio[], |
| 5281 | unsigned int HTotal[], |
| 5282 | double PixelClock[], |
| 5283 | unsigned int BlendingAndTiming[], |
| 5284 | double BytePerPixelDETY[], |
| 5285 | double BytePerPixelDETC[], |
| 5286 | bool_Bool WritebackEnable[], |
| 5287 | enum source_format_class WritebackPixelFormat[], |
| 5288 | double WritebackDestinationWidth[], |
| 5289 | double WritebackDestinationHeight[], |
| 5290 | double WritebackSourceHeight[], |
| 5291 | enum clock_change_support *DRAMClockChangeSupport, |
| 5292 | double *UrgentWatermark, |
| 5293 | double *WritebackUrgentWatermark, |
| 5294 | double *DRAMClockChangeWatermark, |
| 5295 | double *WritebackDRAMClockChangeWatermark, |
| 5296 | double *StutterExitWatermark, |
| 5297 | double *StutterEnterPlusExitWatermark, |
| 5298 | double *MinActiveDRAMClockChangeLatencySupported) |
| 5299 | { |
| 5300 | double EffectiveLBLatencyHidingY; |
| 5301 | double EffectiveLBLatencyHidingC; |
| 5302 | double DPPOutputBufferLinesY; |
| 5303 | double DPPOutputBufferLinesC; |
| 5304 | unsigned int DETBufferSizeY; |
| 5305 | unsigned int DETBufferSizeC; |
| 5306 | double LinesInDETY[DC__NUM_DPP__MAX8]; |
| 5307 | double LinesInDETC; |
| 5308 | unsigned int LinesInDETYRoundedDownToSwath[DC__NUM_DPP__MAX8]; |
| 5309 | unsigned int LinesInDETCRoundedDownToSwath; |
| 5310 | double FullDETBufferingTimeY[DC__NUM_DPP__MAX8]; |
| 5311 | double FullDETBufferingTimeC; |
| 5312 | double ActiveDRAMClockChangeLatencyMarginY; |
| 5313 | double ActiveDRAMClockChangeLatencyMarginC; |
| 5314 | double WritebackDRAMClockChangeLatencyMargin; |
| 5315 | double PlaneWithMinActiveDRAMClockChangeMargin; |
| 5316 | double SecondMinActiveDRAMClockChangeMarginOneDisplayInVBLank; |
| 5317 | double FullDETBufferingTimeYStutterCriticalPlane = 0; |
| 5318 | double TimeToFinishSwathTransferStutterCriticalPlane = 0; |
| 5319 | unsigned int k, j; |
| 5320 | |
| 5321 | mode_lib->vba.TotalActiveDPP = 0; |
| 5322 | mode_lib->vba.TotalDCCActiveDPP = 0; |
| 5323 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 5324 | mode_lib->vba.TotalActiveDPP = mode_lib->vba.TotalActiveDPP + DPPPerPlane[k]; |
| 5325 | if (DCCEnable[k] == true1) { |
| 5326 | mode_lib->vba.TotalDCCActiveDPP = mode_lib->vba.TotalDCCActiveDPP + DPPPerPlane[k]; |
| 5327 | } |
| 5328 | } |
| 5329 | |
| 5330 | mode_lib->vba.TotalDataReadBandwidth = 0; |
| 5331 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 5332 | mode_lib->vba.TotalDataReadBandwidth = mode_lib->vba.TotalDataReadBandwidth |
| 5333 | + ReadBandwidthPlaneLuma[k] + ReadBandwidthPlaneChroma[k]; |
| 5334 | } |
| 5335 | |
| 5336 | *UrgentWatermark = UrgentLatency + ExtraLatency; |
| 5337 | |
| 5338 | *DRAMClockChangeWatermark = DRAMClockChangeLatency + *UrgentWatermark; |
| 5339 | |
| 5340 | mode_lib->vba.TotalActiveWriteback = 0; |
| 5341 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 5342 | if (WritebackEnable[k] == true1) { |
| 5343 | mode_lib->vba.TotalActiveWriteback = mode_lib->vba.TotalActiveWriteback + 1; |
| 5344 | } |
| 5345 | } |
| 5346 | |
| 5347 | if (mode_lib->vba.TotalActiveWriteback <= 1) { |
| 5348 | *WritebackUrgentWatermark = WritebackLatency; |
| 5349 | } else { |
| 5350 | *WritebackUrgentWatermark = WritebackLatency |
| 5351 | + WritebackChunkSize * 1024.0 / 32.0 / SOCCLK; |
| 5352 | } |
| 5353 | |
| 5354 | if (mode_lib->vba.TotalActiveWriteback <= 1) { |
| 5355 | *WritebackDRAMClockChangeWatermark = DRAMClockChangeLatency + WritebackLatency; |
| 5356 | } else { |
| 5357 | *WritebackDRAMClockChangeWatermark = DRAMClockChangeLatency + WritebackLatency |
| 5358 | + WritebackChunkSize * 1024.0 / 32.0 / SOCCLK; |
| 5359 | } |
| 5360 | |
| 5361 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 5362 | |
| 5363 | mode_lib->vba.LBLatencyHidingSourceLinesY = dml_min((double) MaxLineBufferLines, |
| 5364 | dml_floor(LineBufferSize / LBBitPerPixel[k] / (SwathWidthY[k] / dml_max(HRatio[k], 1.0)), 1)) |
| 5365 | - (vtaps[k] - 1); |
| 5366 | |
| 5367 | mode_lib->vba.LBLatencyHidingSourceLinesC = dml_min((double) MaxLineBufferLines, |
| 5368 | dml_floor(LineBufferSize / LBBitPerPixel[k] / (SwathWidthY[k] / 2 / dml_max(HRatio[k] / 2, 1.0)), 1)) |
| 5369 | - (VTAPsChroma[k] - 1); |
| 5370 | |
| 5371 | EffectiveLBLatencyHidingY = mode_lib->vba.LBLatencyHidingSourceLinesY / VRatio[k] |
| 5372 | * (HTotal[k] / PixelClock[k]); |
| 5373 | |
| 5374 | EffectiveLBLatencyHidingC = mode_lib->vba.LBLatencyHidingSourceLinesC |
| 5375 | / (VRatio[k] / 2) * (HTotal[k] / PixelClock[k]); |
| 5376 | |
| 5377 | if (SwathWidthY[k] > 2 * DPPOutputBufferPixels) { |
| 5378 | DPPOutputBufferLinesY = (double) DPPOutputBufferPixels / SwathWidthY[k]; |
| 5379 | } else if (SwathWidthY[k] > DPPOutputBufferPixels) { |
| 5380 | DPPOutputBufferLinesY = 0.5; |
| 5381 | } else { |
| 5382 | DPPOutputBufferLinesY = 1; |
| 5383 | } |
| 5384 | |
| 5385 | if (SwathWidthY[k] / 2.0 > 2 * DPPOutputBufferPixels) { |
| 5386 | DPPOutputBufferLinesC = (double) DPPOutputBufferPixels |
| 5387 | / (SwathWidthY[k] / 2.0); |
| 5388 | } else if (SwathWidthY[k] / 2.0 > DPPOutputBufferPixels) { |
| 5389 | DPPOutputBufferLinesC = 0.5; |
| 5390 | } else { |
| 5391 | DPPOutputBufferLinesC = 1; |
| 5392 | } |
| 5393 | |
| 5394 | CalculateDETBufferSize( |
| 5395 | DETBufferSizeInKByte, |
| 5396 | SwathHeightY[k], |
| 5397 | SwathHeightC[k], |
| 5398 | &DETBufferSizeY, |
| 5399 | &DETBufferSizeC); |
| 5400 | |
| 5401 | LinesInDETY[k] = (double)DETBufferSizeY / BytePerPixelDETY[k] / SwathWidthY[k]; |
| 5402 | LinesInDETYRoundedDownToSwath[k] = dml_floor(LinesInDETY[k], SwathHeightY[k]); |
| 5403 | FullDETBufferingTimeY[k] = LinesInDETYRoundedDownToSwath[k] |
| 5404 | * (HTotal[k] / PixelClock[k]) / VRatio[k]; |
| 5405 | if (BytePerPixelDETC[k] > 0) { |
| 5406 | LinesInDETC = (double)DETBufferSizeC / BytePerPixelDETC[k] / (SwathWidthY[k] / 2.0); |
| 5407 | LinesInDETCRoundedDownToSwath = dml_floor(LinesInDETC, SwathHeightC[k]); |
| 5408 | FullDETBufferingTimeC = LinesInDETCRoundedDownToSwath |
| 5409 | * (HTotal[k] / PixelClock[k]) / (VRatio[k] / 2); |
| 5410 | } else { |
| 5411 | LinesInDETC = 0; |
| 5412 | FullDETBufferingTimeC = 999999; |
| 5413 | } |
| 5414 | |
| 5415 | ActiveDRAMClockChangeLatencyMarginY = HTotal[k] / PixelClock[k] |
| 5416 | * DPPOutputBufferLinesY + EffectiveLBLatencyHidingY |
| 5417 | + FullDETBufferingTimeY[k] - *DRAMClockChangeWatermark; |
| 5418 | |
| 5419 | if (NumberOfActivePlanes > 1) { |
| 5420 | ActiveDRAMClockChangeLatencyMarginY = ActiveDRAMClockChangeLatencyMarginY |
| 5421 | - (1 - 1.0 / NumberOfActivePlanes) * SwathHeightY[k] * HTotal[k] / PixelClock[k] / VRatio[k]; |
| 5422 | } |
| 5423 | |
| 5424 | if (BytePerPixelDETC[k] > 0) { |
| 5425 | ActiveDRAMClockChangeLatencyMarginC = HTotal[k] / PixelClock[k] |
| 5426 | * DPPOutputBufferLinesC + EffectiveLBLatencyHidingC |
| 5427 | + FullDETBufferingTimeC - *DRAMClockChangeWatermark; |
| 5428 | if (NumberOfActivePlanes > 1) { |
| 5429 | ActiveDRAMClockChangeLatencyMarginC = ActiveDRAMClockChangeLatencyMarginC |
| 5430 | - (1 - 1.0 / NumberOfActivePlanes) * SwathHeightC[k] * HTotal[k] / PixelClock[k] / (VRatio[k] / 2); |
| 5431 | } |
| 5432 | mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k] = dml_min( |
| 5433 | ActiveDRAMClockChangeLatencyMarginY, |
| 5434 | ActiveDRAMClockChangeLatencyMarginC); |
| 5435 | } else { |
| 5436 | mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k] = ActiveDRAMClockChangeLatencyMarginY; |
| 5437 | } |
| 5438 | |
| 5439 | if (WritebackEnable[k] == true1) { |
| 5440 | if (WritebackPixelFormat[k] == dm_444_32) { |
| 5441 | WritebackDRAMClockChangeLatencyMargin = (WritebackInterfaceLumaBufferSize |
| 5442 | + WritebackInterfaceChromaBufferSize) / (WritebackDestinationWidth[k] |
| 5443 | * WritebackDestinationHeight[k] / (WritebackSourceHeight[k] * HTotal[k] |
| 5444 | / PixelClock[k]) * 4) - *WritebackDRAMClockChangeWatermark; |
| 5445 | } else { |
| 5446 | WritebackDRAMClockChangeLatencyMargin = dml_min( |
| 5447 | WritebackInterfaceLumaBufferSize * 8.0 / 10, |
| 5448 | 2 * WritebackInterfaceChromaBufferSize * 8.0 / 10) / (WritebackDestinationWidth[k] |
| 5449 | * WritebackDestinationHeight[k] / (WritebackSourceHeight[k] * HTotal[k] / PixelClock[k])) |
| 5450 | - *WritebackDRAMClockChangeWatermark; |
| 5451 | } |
| 5452 | mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k] = dml_min( |
| 5453 | mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k], |
| 5454 | WritebackDRAMClockChangeLatencyMargin); |
| 5455 | } |
| 5456 | } |
| 5457 | |
| 5458 | mode_lib->vba.MinActiveDRAMClockChangeMargin = 999999; |
| 5459 | PlaneWithMinActiveDRAMClockChangeMargin = 0; |
| 5460 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 5461 | if (mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k] |
| 5462 | < mode_lib->vba.MinActiveDRAMClockChangeMargin) { |
| 5463 | mode_lib->vba.MinActiveDRAMClockChangeMargin = |
| 5464 | mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k]; |
| 5465 | if (BlendingAndTiming[k] == k) { |
| 5466 | PlaneWithMinActiveDRAMClockChangeMargin = k; |
| 5467 | } else { |
| 5468 | for (j = 0; j < NumberOfActivePlanes; ++j) { |
| 5469 | if (BlendingAndTiming[k] == j) { |
| 5470 | PlaneWithMinActiveDRAMClockChangeMargin = j; |
| 5471 | } |
| 5472 | } |
| 5473 | } |
| 5474 | } |
| 5475 | } |
| 5476 | |
| 5477 | *MinActiveDRAMClockChangeLatencySupported = mode_lib->vba.MinActiveDRAMClockChangeMargin + DRAMClockChangeLatency; |
| 5478 | |
| 5479 | SecondMinActiveDRAMClockChangeMarginOneDisplayInVBLank = 999999; |
| 5480 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 5481 | if (!((k == PlaneWithMinActiveDRAMClockChangeMargin) && (BlendingAndTiming[k] == k)) |
| 5482 | && !(BlendingAndTiming[k] == PlaneWithMinActiveDRAMClockChangeMargin) |
| 5483 | && mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k] |
| 5484 | < SecondMinActiveDRAMClockChangeMarginOneDisplayInVBLank) { |
| 5485 | SecondMinActiveDRAMClockChangeMarginOneDisplayInVBLank = |
| 5486 | mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k]; |
| 5487 | } |
| 5488 | } |
| 5489 | |
| 5490 | mode_lib->vba.TotalNumberOfActiveOTG = 0; |
| 5491 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 5492 | if (BlendingAndTiming[k] == k) { |
| 5493 | mode_lib->vba.TotalNumberOfActiveOTG = mode_lib->vba.TotalNumberOfActiveOTG + 1; |
| 5494 | } |
| 5495 | } |
| 5496 | |
| 5497 | if (mode_lib->vba.MinActiveDRAMClockChangeMargin > 0 && PrefetchMode == 0) { |
| 5498 | *DRAMClockChangeSupport = dm_dram_clock_change_vactive; |
| 5499 | } else if (((mode_lib->vba.SynchronizedVBlank == true1 |
| 5500 | || mode_lib->vba.TotalNumberOfActiveOTG == 1 |
| 5501 | || SecondMinActiveDRAMClockChangeMarginOneDisplayInVBLank > 0) |
| 5502 | && PrefetchMode == 0)) { |
| 5503 | *DRAMClockChangeSupport = dm_dram_clock_change_vblank; |
| 5504 | } else { |
| 5505 | *DRAMClockChangeSupport = dm_dram_clock_change_unsupported; |
| 5506 | } |
| 5507 | |
| 5508 | FullDETBufferingTimeYStutterCriticalPlane = FullDETBufferingTimeY[0]; |
| 5509 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 5510 | if (FullDETBufferingTimeY[k] <= FullDETBufferingTimeYStutterCriticalPlane) { |
| 5511 | TimeToFinishSwathTransferStutterCriticalPlane = (SwathHeightY[k] |
| 5512 | - (LinesInDETY[k] - LinesInDETYRoundedDownToSwath[k])) |
| 5513 | * (HTotal[k] / PixelClock[k]) / VRatio[k]; |
| 5514 | } |
| 5515 | } |
| 5516 | |
| 5517 | *StutterExitWatermark = SRExitTime + mode_lib->vba.LastPixelOfLineExtraWatermark |
| 5518 | + ExtraLatency + 10 / DCFCLKDeepSleep; |
| 5519 | *StutterEnterPlusExitWatermark = dml_max( |
| 5520 | SREnterPlusExitTime + mode_lib->vba.LastPixelOfLineExtraWatermark |
| 5521 | + ExtraLatency + 10 / DCFCLKDeepSleep, |
| 5522 | TimeToFinishSwathTransferStutterCriticalPlane); |
| 5523 | |
| 5524 | } |
| 5525 | |
| 5526 | static void CalculateDCFCLKDeepSleep( |
| 5527 | struct display_mode_lib *mode_lib, |
| 5528 | unsigned int NumberOfActivePlanes, |
| 5529 | double BytePerPixelDETY[], |
| 5530 | double BytePerPixelDETC[], |
| 5531 | double VRatio[], |
| 5532 | double SwathWidthY[], |
| 5533 | int DPPPerPlane[], |
| 5534 | double HRatio[], |
| 5535 | double PixelClock[], |
| 5536 | double PSCL_THROUGHPUT[], |
| 5537 | double PSCL_THROUGHPUT_CHROMA[], |
| 5538 | double DPPCLK[], |
| 5539 | double *DCFCLKDeepSleep) |
| 5540 | { |
| 5541 | unsigned int k; |
| 5542 | double DisplayPipeLineDeliveryTimeLuma; |
| 5543 | double DisplayPipeLineDeliveryTimeChroma; |
| 5544 | //double DCFCLKDeepSleepPerPlane[DC__NUM_DPP__MAX]; |
| 5545 | |
| 5546 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 5547 | if (VRatio[k] <= 1) { |
| 5548 | DisplayPipeLineDeliveryTimeLuma = SwathWidthY[k] * DPPPerPlane[k] |
| 5549 | / HRatio[k] / PixelClock[k]; |
| 5550 | } else { |
| 5551 | DisplayPipeLineDeliveryTimeLuma = SwathWidthY[k] / PSCL_THROUGHPUT[k] |
| 5552 | / DPPCLK[k]; |
| 5553 | } |
| 5554 | if (BytePerPixelDETC[k] == 0) { |
| 5555 | DisplayPipeLineDeliveryTimeChroma = 0; |
| 5556 | } else { |
| 5557 | if (VRatio[k] / 2 <= 1) { |
| 5558 | DisplayPipeLineDeliveryTimeChroma = SwathWidthY[k] / 2.0 |
| 5559 | * DPPPerPlane[k] / (HRatio[k] / 2) / PixelClock[k]; |
| 5560 | } else { |
| 5561 | DisplayPipeLineDeliveryTimeChroma = SwathWidthY[k] / 2.0 |
| 5562 | / PSCL_THROUGHPUT_CHROMA[k] / DPPCLK[k]; |
| 5563 | } |
| 5564 | } |
| 5565 | |
| 5566 | if (BytePerPixelDETC[k] > 0) { |
| 5567 | mode_lib->vba.DCFCLKDeepSleepPerPlane[k] = dml_max( |
| 5568 | 1.1 * SwathWidthY[k] * dml_ceil(BytePerPixelDETY[k], 1) |
| 5569 | / 32.0 / DisplayPipeLineDeliveryTimeLuma, |
| 5570 | 1.1 * SwathWidthY[k] / 2.0 |
| 5571 | * dml_ceil(BytePerPixelDETC[k], 2) / 32.0 |
| 5572 | / DisplayPipeLineDeliveryTimeChroma); |
| 5573 | } else { |
| 5574 | mode_lib->vba.DCFCLKDeepSleepPerPlane[k] = 1.1 * SwathWidthY[k] |
| 5575 | * dml_ceil(BytePerPixelDETY[k], 1) / 64.0 |
| 5576 | / DisplayPipeLineDeliveryTimeLuma; |
| 5577 | } |
| 5578 | mode_lib->vba.DCFCLKDeepSleepPerPlane[k] = dml_max( |
| 5579 | mode_lib->vba.DCFCLKDeepSleepPerPlane[k], |
| 5580 | PixelClock[k] / 16); |
| 5581 | |
| 5582 | } |
| 5583 | |
| 5584 | *DCFCLKDeepSleep = 8; |
| 5585 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 5586 | *DCFCLKDeepSleep = dml_max( |
| 5587 | *DCFCLKDeepSleep, |
| 5588 | mode_lib->vba.DCFCLKDeepSleepPerPlane[k]); |
| 5589 | } |
| 5590 | } |
| 5591 | |
| 5592 | static void CalculateDETBufferSize( |
| 5593 | unsigned int DETBufferSizeInKByte, |
| 5594 | unsigned int SwathHeightY, |
| 5595 | unsigned int SwathHeightC, |
| 5596 | unsigned int *DETBufferSizeY, |
| 5597 | unsigned int *DETBufferSizeC) |
| 5598 | { |
| 5599 | if (SwathHeightC == 0) { |
| 5600 | *DETBufferSizeY = DETBufferSizeInKByte * 1024; |
| 5601 | *DETBufferSizeC = 0; |
| 5602 | } else if (SwathHeightY <= SwathHeightC) { |
| 5603 | *DETBufferSizeY = DETBufferSizeInKByte * 1024 / 2; |
| 5604 | *DETBufferSizeC = DETBufferSizeInKByte * 1024 / 2; |
| 5605 | } else { |
| 5606 | *DETBufferSizeY = DETBufferSizeInKByte * 1024 * 2 / 3; |
| 5607 | *DETBufferSizeC = DETBufferSizeInKByte * 1024 / 3; |
| 5608 | } |
| 5609 | } |
| 5610 | |
| 5611 | static void CalculateUrgentBurstFactor( |
| 5612 | unsigned int DETBufferSizeInKByte, |
| 5613 | unsigned int SwathHeightY, |
| 5614 | unsigned int SwathHeightC, |
| 5615 | unsigned int SwathWidthY, |
| 5616 | double LineTime, |
| 5617 | double UrgentLatency, |
| 5618 | double CursorBufferSize, |
| 5619 | unsigned int CursorWidth, |
| 5620 | unsigned int CursorBPP, |
| 5621 | double VRatio, |
| 5622 | double VRatioPreY, |
| 5623 | double VRatioPreC, |
| 5624 | double BytePerPixelInDETY, |
| 5625 | double BytePerPixelInDETC, |
| 5626 | double *UrgentBurstFactorCursor, |
| 5627 | double *UrgentBurstFactorCursorPre, |
| 5628 | double *UrgentBurstFactorLuma, |
| 5629 | double *UrgentBurstFactorLumaPre, |
| 5630 | double *UrgentBurstFactorChroma, |
| 5631 | double *UrgentBurstFactorChromaPre, |
| 5632 | unsigned int *NotEnoughUrgentLatencyHiding, |
| 5633 | unsigned int *NotEnoughUrgentLatencyHidingPre) |
| 5634 | { |
| 5635 | double LinesInDETLuma; |
| 5636 | double LinesInDETChroma; |
| 5637 | unsigned int LinesInCursorBuffer; |
| 5638 | double CursorBufferSizeInTime; |
| 5639 | double CursorBufferSizeInTimePre; |
| 5640 | double DETBufferSizeInTimeLuma; |
| 5641 | double DETBufferSizeInTimeLumaPre; |
| 5642 | double DETBufferSizeInTimeChroma; |
| 5643 | double DETBufferSizeInTimeChromaPre; |
| 5644 | unsigned int DETBufferSizeY; |
| 5645 | unsigned int DETBufferSizeC; |
| 5646 | |
| 5647 | *NotEnoughUrgentLatencyHiding = 0; |
| 5648 | *NotEnoughUrgentLatencyHidingPre = 0; |
| 5649 | |
| 5650 | if (CursorWidth > 0) { |
| 5651 | LinesInCursorBuffer = 1 << (unsigned int) dml_floor( |
| 5652 | dml_log2(CursorBufferSize * 1024.0 / (CursorWidth * CursorBPP / 8.0)), 1.0); |
| 5653 | CursorBufferSizeInTime = LinesInCursorBuffer * LineTime / VRatio; |
| 5654 | if (CursorBufferSizeInTime - UrgentLatency <= 0) { |
| 5655 | *NotEnoughUrgentLatencyHiding = 1; |
| 5656 | *UrgentBurstFactorCursor = 0; |
| 5657 | } else { |
| 5658 | *UrgentBurstFactorCursor = CursorBufferSizeInTime |
| 5659 | / (CursorBufferSizeInTime - UrgentLatency); |
| 5660 | } |
| 5661 | if (VRatioPreY > 0) { |
| 5662 | CursorBufferSizeInTimePre = LinesInCursorBuffer * LineTime / VRatioPreY; |
| 5663 | if (CursorBufferSizeInTimePre - UrgentLatency <= 0) { |
| 5664 | *NotEnoughUrgentLatencyHidingPre = 1; |
| 5665 | *UrgentBurstFactorCursorPre = 0; |
| 5666 | } else { |
| 5667 | *UrgentBurstFactorCursorPre = CursorBufferSizeInTimePre |
| 5668 | / (CursorBufferSizeInTimePre - UrgentLatency); |
| 5669 | } |
| 5670 | } else { |
| 5671 | *UrgentBurstFactorCursorPre = 1; |
| 5672 | } |
| 5673 | } |
| 5674 | |
| 5675 | CalculateDETBufferSize( |
| 5676 | DETBufferSizeInKByte, |
| 5677 | SwathHeightY, |
| 5678 | SwathHeightC, |
| 5679 | &DETBufferSizeY, |
| 5680 | &DETBufferSizeC); |
| 5681 | |
| 5682 | LinesInDETLuma = (double)DETBufferSizeY / BytePerPixelInDETY / SwathWidthY; |
| 5683 | DETBufferSizeInTimeLuma = dml_floor(LinesInDETLuma, SwathHeightY) * LineTime / VRatio; |
| 5684 | if (DETBufferSizeInTimeLuma - UrgentLatency <= 0) { |
| 5685 | *NotEnoughUrgentLatencyHiding = 1; |
| 5686 | *UrgentBurstFactorLuma = 0; |
| 5687 | } else { |
| 5688 | *UrgentBurstFactorLuma = DETBufferSizeInTimeLuma |
| 5689 | / (DETBufferSizeInTimeLuma - UrgentLatency); |
| 5690 | } |
| 5691 | if (VRatioPreY > 0) { |
| 5692 | DETBufferSizeInTimeLumaPre = dml_floor(LinesInDETLuma, SwathHeightY) * LineTime |
| 5693 | / VRatioPreY; |
| 5694 | if (DETBufferSizeInTimeLumaPre - UrgentLatency <= 0) { |
| 5695 | *NotEnoughUrgentLatencyHidingPre = 1; |
| 5696 | *UrgentBurstFactorLumaPre = 0; |
| 5697 | } else { |
| 5698 | *UrgentBurstFactorLumaPre = DETBufferSizeInTimeLumaPre |
| 5699 | / (DETBufferSizeInTimeLumaPre - UrgentLatency); |
| 5700 | } |
| 5701 | } else { |
| 5702 | *UrgentBurstFactorLumaPre = 1; |
| 5703 | } |
| 5704 | |
| 5705 | if (BytePerPixelInDETC > 0) { |
| 5706 | LinesInDETChroma = (double)DETBufferSizeC / BytePerPixelInDETC / (SwathWidthY / 2); |
| 5707 | DETBufferSizeInTimeChroma = dml_floor(LinesInDETChroma, SwathHeightC) * LineTime |
| 5708 | / (VRatio / 2); |
| 5709 | if (DETBufferSizeInTimeChroma - UrgentLatency <= 0) { |
| 5710 | *NotEnoughUrgentLatencyHiding = 1; |
| 5711 | *UrgentBurstFactorChroma = 0; |
| 5712 | } else { |
| 5713 | *UrgentBurstFactorChroma = DETBufferSizeInTimeChroma |
| 5714 | / (DETBufferSizeInTimeChroma - UrgentLatency); |
| 5715 | } |
| 5716 | if (VRatioPreC > 0) { |
| 5717 | DETBufferSizeInTimeChromaPre = dml_floor(LinesInDETChroma, SwathHeightC) |
| 5718 | * LineTime / VRatioPreC; |
| 5719 | if (DETBufferSizeInTimeChromaPre - UrgentLatency <= 0) { |
| 5720 | *NotEnoughUrgentLatencyHidingPre = 1; |
| 5721 | *UrgentBurstFactorChromaPre = 0; |
| 5722 | } else { |
| 5723 | *UrgentBurstFactorChromaPre = DETBufferSizeInTimeChromaPre |
| 5724 | / (DETBufferSizeInTimeChromaPre - UrgentLatency); |
| 5725 | } |
| 5726 | } else { |
| 5727 | *UrgentBurstFactorChromaPre = 1; |
| 5728 | } |
| 5729 | } |
| 5730 | } |
| 5731 | |
| 5732 | static void CalculatePixelDeliveryTimes( |
| 5733 | unsigned int NumberOfActivePlanes, |
| 5734 | double VRatio[], |
| 5735 | double VRatioPrefetchY[], |
| 5736 | double VRatioPrefetchC[], |
| 5737 | unsigned int swath_width_luma_ub[], |
| 5738 | unsigned int swath_width_chroma_ub[], |
| 5739 | int DPPPerPlane[], |
| 5740 | double HRatio[], |
| 5741 | double PixelClock[], |
| 5742 | double PSCL_THROUGHPUT[], |
| 5743 | double PSCL_THROUGHPUT_CHROMA[], |
| 5744 | double DPPCLK[], |
| 5745 | double BytePerPixelDETC[], |
| 5746 | enum scan_direction_class SourceScan[], |
| 5747 | unsigned int BlockWidth256BytesY[], |
| 5748 | unsigned int BlockHeight256BytesY[], |
| 5749 | unsigned int BlockWidth256BytesC[], |
| 5750 | unsigned int BlockHeight256BytesC[], |
| 5751 | double DisplayPipeLineDeliveryTimeLuma[], |
| 5752 | double DisplayPipeLineDeliveryTimeChroma[], |
| 5753 | double DisplayPipeLineDeliveryTimeLumaPrefetch[], |
| 5754 | double DisplayPipeLineDeliveryTimeChromaPrefetch[], |
| 5755 | double DisplayPipeRequestDeliveryTimeLuma[], |
| 5756 | double DisplayPipeRequestDeliveryTimeChroma[], |
| 5757 | double DisplayPipeRequestDeliveryTimeLumaPrefetch[], |
| 5758 | double DisplayPipeRequestDeliveryTimeChromaPrefetch[]) |
| 5759 | { |
| 5760 | double req_per_swath_ub; |
| 5761 | unsigned int k; |
| 5762 | |
| 5763 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 5764 | if (VRatio[k] <= 1) { |
| 5765 | DisplayPipeLineDeliveryTimeLuma[k] = swath_width_luma_ub[k] * DPPPerPlane[k] |
| 5766 | / HRatio[k] / PixelClock[k]; |
| 5767 | } else { |
| 5768 | DisplayPipeLineDeliveryTimeLuma[k] = swath_width_luma_ub[k] |
| 5769 | / PSCL_THROUGHPUT[k] / DPPCLK[k]; |
| 5770 | } |
| 5771 | |
| 5772 | if (BytePerPixelDETC[k] == 0) { |
| 5773 | DisplayPipeLineDeliveryTimeChroma[k] = 0; |
| 5774 | } else { |
| 5775 | if (VRatio[k] / 2 <= 1) { |
| 5776 | DisplayPipeLineDeliveryTimeChroma[k] = swath_width_chroma_ub[k] |
| 5777 | * DPPPerPlane[k] / (HRatio[k] / 2) / PixelClock[k]; |
| 5778 | } else { |
| 5779 | DisplayPipeLineDeliveryTimeChroma[k] = swath_width_chroma_ub[k] |
| 5780 | / PSCL_THROUGHPUT_CHROMA[k] / DPPCLK[k]; |
| 5781 | } |
| 5782 | } |
| 5783 | |
| 5784 | if (VRatioPrefetchY[k] <= 1) { |
| 5785 | DisplayPipeLineDeliveryTimeLumaPrefetch[k] = swath_width_luma_ub[k] |
| 5786 | * DPPPerPlane[k] / HRatio[k] / PixelClock[k]; |
| 5787 | } else { |
| 5788 | DisplayPipeLineDeliveryTimeLumaPrefetch[k] = swath_width_luma_ub[k] |
| 5789 | / PSCL_THROUGHPUT[k] / DPPCLK[k]; |
| 5790 | } |
| 5791 | |
| 5792 | if (BytePerPixelDETC[k] == 0) { |
| 5793 | DisplayPipeLineDeliveryTimeChromaPrefetch[k] = 0; |
| 5794 | } else { |
| 5795 | if (VRatioPrefetchC[k] <= 1) { |
| 5796 | DisplayPipeLineDeliveryTimeChromaPrefetch[k] = |
| 5797 | swath_width_chroma_ub[k] * DPPPerPlane[k] |
| 5798 | / (HRatio[k] / 2) / PixelClock[k]; |
| 5799 | } else { |
| 5800 | DisplayPipeLineDeliveryTimeChromaPrefetch[k] = |
| 5801 | swath_width_chroma_ub[k] / PSCL_THROUGHPUT_CHROMA[k] / DPPCLK[k]; |
| 5802 | } |
| 5803 | } |
| 5804 | } |
| 5805 | |
| 5806 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 5807 | if (SourceScan[k] == dm_horz) { |
| 5808 | req_per_swath_ub = swath_width_luma_ub[k] / BlockWidth256BytesY[k]; |
| 5809 | } else { |
| 5810 | req_per_swath_ub = swath_width_luma_ub[k] / BlockHeight256BytesY[k]; |
| 5811 | } |
| 5812 | DisplayPipeRequestDeliveryTimeLuma[k] = DisplayPipeLineDeliveryTimeLuma[k] |
| 5813 | / req_per_swath_ub; |
| 5814 | DisplayPipeRequestDeliveryTimeLumaPrefetch[k] = |
| 5815 | DisplayPipeLineDeliveryTimeLumaPrefetch[k] / req_per_swath_ub; |
| 5816 | if (BytePerPixelDETC[k] == 0) { |
| 5817 | DisplayPipeRequestDeliveryTimeChroma[k] = 0; |
| 5818 | DisplayPipeRequestDeliveryTimeChromaPrefetch[k] = 0; |
| 5819 | } else { |
| 5820 | if (SourceScan[k] == dm_horz) { |
| 5821 | req_per_swath_ub = swath_width_chroma_ub[k] |
| 5822 | / BlockWidth256BytesC[k]; |
| 5823 | } else { |
| 5824 | req_per_swath_ub = swath_width_chroma_ub[k] |
| 5825 | / BlockHeight256BytesC[k]; |
| 5826 | } |
| 5827 | DisplayPipeRequestDeliveryTimeChroma[k] = |
| 5828 | DisplayPipeLineDeliveryTimeChroma[k] / req_per_swath_ub; |
| 5829 | DisplayPipeRequestDeliveryTimeChromaPrefetch[k] = |
| 5830 | DisplayPipeLineDeliveryTimeChromaPrefetch[k] / req_per_swath_ub; |
| 5831 | } |
| 5832 | } |
| 5833 | } |
| 5834 | |
| 5835 | static void CalculateMetaAndPTETimes( |
| 5836 | unsigned int NumberOfActivePlanes, |
| 5837 | bool_Bool GPUVMEnable, |
| 5838 | unsigned int MetaChunkSize, |
| 5839 | unsigned int MinMetaChunkSizeBytes, |
| 5840 | unsigned int GPUVMMaxPageTableLevels, |
| 5841 | unsigned int HTotal[], |
| 5842 | double VRatio[], |
| 5843 | double VRatioPrefetchY[], |
| 5844 | double VRatioPrefetchC[], |
| 5845 | double DestinationLinesToRequestRowInVBlank[], |
| 5846 | double DestinationLinesToRequestRowInImmediateFlip[], |
| 5847 | double DestinationLinesToRequestVMInVBlank[], |
| 5848 | double DestinationLinesToRequestVMInImmediateFlip[], |
| 5849 | bool_Bool DCCEnable[], |
| 5850 | double PixelClock[], |
| 5851 | double BytePerPixelDETY[], |
| 5852 | double BytePerPixelDETC[], |
| 5853 | enum scan_direction_class SourceScan[], |
| 5854 | unsigned int dpte_row_height[], |
| 5855 | unsigned int dpte_row_height_chroma[], |
| 5856 | unsigned int meta_row_width[], |
| 5857 | unsigned int meta_row_height[], |
| 5858 | unsigned int meta_req_width[], |
| 5859 | unsigned int meta_req_height[], |
| 5860 | int dpte_group_bytes[], |
| 5861 | unsigned int PTERequestSizeY[], |
| 5862 | unsigned int PTERequestSizeC[], |
| 5863 | unsigned int PixelPTEReqWidthY[], |
| 5864 | unsigned int PixelPTEReqHeightY[], |
| 5865 | unsigned int PixelPTEReqWidthC[], |
| 5866 | unsigned int PixelPTEReqHeightC[], |
| 5867 | unsigned int dpte_row_width_luma_ub[], |
| 5868 | unsigned int dpte_row_width_chroma_ub[], |
| 5869 | unsigned int vm_group_bytes[], |
| 5870 | unsigned int dpde0_bytes_per_frame_ub_l[], |
| 5871 | unsigned int dpde0_bytes_per_frame_ub_c[], |
| 5872 | unsigned int meta_pte_bytes_per_frame_ub_l[], |
| 5873 | unsigned int meta_pte_bytes_per_frame_ub_c[], |
| 5874 | double DST_Y_PER_PTE_ROW_NOM_L[], |
| 5875 | double DST_Y_PER_PTE_ROW_NOM_C[], |
| 5876 | double DST_Y_PER_META_ROW_NOM_L[], |
| 5877 | double TimePerMetaChunkNominal[], |
| 5878 | double TimePerMetaChunkVBlank[], |
| 5879 | double TimePerMetaChunkFlip[], |
| 5880 | double time_per_pte_group_nom_luma[], |
| 5881 | double time_per_pte_group_vblank_luma[], |
| 5882 | double time_per_pte_group_flip_luma[], |
| 5883 | double time_per_pte_group_nom_chroma[], |
| 5884 | double time_per_pte_group_vblank_chroma[], |
| 5885 | double time_per_pte_group_flip_chroma[], |
| 5886 | double TimePerVMGroupVBlank[], |
| 5887 | double TimePerVMGroupFlip[], |
| 5888 | double TimePerVMRequestVBlank[], |
| 5889 | double TimePerVMRequestFlip[]) |
| 5890 | { |
| 5891 | unsigned int meta_chunk_width; |
| 5892 | unsigned int min_meta_chunk_width; |
| 5893 | unsigned int meta_chunk_per_row_int; |
| 5894 | unsigned int meta_row_remainder; |
| 5895 | unsigned int meta_chunk_threshold; |
| 5896 | unsigned int meta_chunks_per_row_ub; |
| 5897 | unsigned int dpte_group_width_luma; |
| 5898 | unsigned int dpte_group_width_chroma; |
| 5899 | unsigned int dpte_groups_per_row_luma_ub; |
| 5900 | unsigned int dpte_groups_per_row_chroma_ub; |
| 5901 | unsigned int num_group_per_lower_vm_stage; |
| 5902 | unsigned int num_req_per_lower_vm_stage; |
| 5903 | unsigned int k; |
| 5904 | |
| 5905 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 5906 | if (GPUVMEnable == true1) { |
| 5907 | DST_Y_PER_PTE_ROW_NOM_L[k] = dpte_row_height[k] / VRatio[k]; |
| 5908 | if (BytePerPixelDETC[k] == 0) { |
| 5909 | DST_Y_PER_PTE_ROW_NOM_C[k] = 0; |
| 5910 | } else { |
| 5911 | DST_Y_PER_PTE_ROW_NOM_C[k] = dpte_row_height_chroma[k] / (VRatio[k] / 2); |
| 5912 | } |
| 5913 | } else { |
| 5914 | DST_Y_PER_PTE_ROW_NOM_L[k] = 0; |
| 5915 | DST_Y_PER_PTE_ROW_NOM_C[k] = 0; |
| 5916 | } |
| 5917 | if (DCCEnable[k] == true1) { |
| 5918 | DST_Y_PER_META_ROW_NOM_L[k] = meta_row_height[k] / VRatio[k]; |
| 5919 | } else { |
| 5920 | DST_Y_PER_META_ROW_NOM_L[k] = 0; |
| 5921 | } |
| 5922 | } |
| 5923 | |
| 5924 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 5925 | if (DCCEnable[k] == true1) { |
| 5926 | meta_chunk_width = MetaChunkSize * 1024 * 256 |
| 5927 | / dml_ceil(BytePerPixelDETY[k], 1) / meta_row_height[k]; |
| 5928 | min_meta_chunk_width = MinMetaChunkSizeBytes * 256 |
| 5929 | / dml_ceil(BytePerPixelDETY[k], 1) / meta_row_height[k]; |
| 5930 | meta_chunk_per_row_int = meta_row_width[k] / meta_chunk_width; |
| 5931 | meta_row_remainder = meta_row_width[k] % meta_chunk_width; |
| 5932 | if (SourceScan[k] == dm_horz) { |
| 5933 | meta_chunk_threshold = 2 * min_meta_chunk_width - meta_req_width[k]; |
| 5934 | } else { |
| 5935 | meta_chunk_threshold = 2 * min_meta_chunk_width |
| 5936 | - meta_req_height[k]; |
| 5937 | } |
| 5938 | if (meta_row_remainder <= meta_chunk_threshold) { |
| 5939 | meta_chunks_per_row_ub = meta_chunk_per_row_int + 1; |
| 5940 | } else { |
| 5941 | meta_chunks_per_row_ub = meta_chunk_per_row_int + 2; |
| 5942 | } |
| 5943 | TimePerMetaChunkNominal[k] = meta_row_height[k] / VRatio[k] * HTotal[k] |
| 5944 | / PixelClock[k] / meta_chunks_per_row_ub; |
| 5945 | TimePerMetaChunkVBlank[k] = DestinationLinesToRequestRowInVBlank[k] |
| 5946 | * HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub; |
| 5947 | TimePerMetaChunkFlip[k] = DestinationLinesToRequestRowInImmediateFlip[k] |
| 5948 | * HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub; |
| 5949 | } else { |
| 5950 | TimePerMetaChunkNominal[k] = 0; |
| 5951 | TimePerMetaChunkVBlank[k] = 0; |
| 5952 | TimePerMetaChunkFlip[k] = 0; |
| 5953 | } |
| 5954 | } |
| 5955 | |
| 5956 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 5957 | if (GPUVMEnable == true1) { |
| 5958 | if (SourceScan[k] == dm_horz) { |
| 5959 | dpte_group_width_luma = dpte_group_bytes[k] / PTERequestSizeY[k] |
| 5960 | * PixelPTEReqWidthY[k]; |
| 5961 | } else { |
| 5962 | dpte_group_width_luma = dpte_group_bytes[k] / PTERequestSizeY[k] |
| 5963 | * PixelPTEReqHeightY[k]; |
| 5964 | } |
| 5965 | dpte_groups_per_row_luma_ub = dml_ceil( |
| 5966 | (float) dpte_row_width_luma_ub[k] / dpte_group_width_luma, |
| 5967 | 1); |
| 5968 | time_per_pte_group_nom_luma[k] = DST_Y_PER_PTE_ROW_NOM_L[k] * HTotal[k] |
| 5969 | / PixelClock[k] / dpte_groups_per_row_luma_ub; |
| 5970 | time_per_pte_group_vblank_luma[k] = DestinationLinesToRequestRowInVBlank[k] |
| 5971 | * HTotal[k] / PixelClock[k] / dpte_groups_per_row_luma_ub; |
| 5972 | time_per_pte_group_flip_luma[k] = |
| 5973 | DestinationLinesToRequestRowInImmediateFlip[k] * HTotal[k] |
| 5974 | / PixelClock[k] |
| 5975 | / dpte_groups_per_row_luma_ub; |
| 5976 | if (BytePerPixelDETC[k] == 0) { |
| 5977 | time_per_pte_group_nom_chroma[k] = 0; |
| 5978 | time_per_pte_group_vblank_chroma[k] = 0; |
| 5979 | time_per_pte_group_flip_chroma[k] = 0; |
| 5980 | } else { |
| 5981 | if (SourceScan[k] == dm_horz) { |
| 5982 | dpte_group_width_chroma = dpte_group_bytes[k] |
| 5983 | / PTERequestSizeC[k] * PixelPTEReqWidthC[k]; |
| 5984 | } else { |
| 5985 | dpte_group_width_chroma = dpte_group_bytes[k] |
| 5986 | / PTERequestSizeC[k] |
| 5987 | * PixelPTEReqHeightC[k]; |
| 5988 | } |
| 5989 | dpte_groups_per_row_chroma_ub = dml_ceil( |
| 5990 | (float) dpte_row_width_chroma_ub[k] |
| 5991 | / dpte_group_width_chroma, |
| 5992 | 1); |
| 5993 | time_per_pte_group_nom_chroma[k] = DST_Y_PER_PTE_ROW_NOM_C[k] |
| 5994 | * HTotal[k] / PixelClock[k] |
| 5995 | / dpte_groups_per_row_chroma_ub; |
| 5996 | time_per_pte_group_vblank_chroma[k] = |
| 5997 | DestinationLinesToRequestRowInVBlank[k] * HTotal[k] |
| 5998 | / PixelClock[k] |
| 5999 | / dpte_groups_per_row_chroma_ub; |
| 6000 | time_per_pte_group_flip_chroma[k] = |
| 6001 | DestinationLinesToRequestRowInImmediateFlip[k] |
| 6002 | * HTotal[k] / PixelClock[k] |
| 6003 | / dpte_groups_per_row_chroma_ub; |
| 6004 | } |
| 6005 | } else { |
| 6006 | time_per_pte_group_nom_luma[k] = 0; |
| 6007 | time_per_pte_group_vblank_luma[k] = 0; |
| 6008 | time_per_pte_group_flip_luma[k] = 0; |
| 6009 | time_per_pte_group_nom_chroma[k] = 0; |
| 6010 | time_per_pte_group_vblank_chroma[k] = 0; |
| 6011 | time_per_pte_group_flip_chroma[k] = 0; |
| 6012 | } |
| 6013 | } |
| 6014 | |
| 6015 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 6016 | if (GPUVMEnable == true1 && (DCCEnable[k] == true1 || GPUVMMaxPageTableLevels > 1)) { |
| 6017 | if (DCCEnable[k] == false0) { |
| 6018 | if (BytePerPixelDETC[k] > 0) { |
| 6019 | num_group_per_lower_vm_stage = |
| 6020 | dml_ceil((double) (dpde0_bytes_per_frame_ub_l[k]) / (double) (vm_group_bytes[k]), 1) |
| 6021 | + dml_ceil((double) (dpde0_bytes_per_frame_ub_c[k]) / (double) (vm_group_bytes[k]), 1); |
| 6022 | } else { |
| 6023 | num_group_per_lower_vm_stage = |
| 6024 | dml_ceil((double) (dpde0_bytes_per_frame_ub_l[k]) / (double) (vm_group_bytes[k]), 1); |
| 6025 | } |
| 6026 | } else { |
| 6027 | if (GPUVMMaxPageTableLevels == 1) { |
| 6028 | if (BytePerPixelDETC[k] > 0) { |
| 6029 | num_group_per_lower_vm_stage = |
| 6030 | dml_ceil((double) (meta_pte_bytes_per_frame_ub_l[k]) / (double) (vm_group_bytes[k]), 1) |
| 6031 | + dml_ceil((double) (meta_pte_bytes_per_frame_ub_c[k]) / (double) (vm_group_bytes[k]), 1); |
| 6032 | } else { |
| 6033 | num_group_per_lower_vm_stage = |
| 6034 | dml_ceil((double) (meta_pte_bytes_per_frame_ub_l[k]) / (double) (vm_group_bytes[k]), 1); |
| 6035 | } |
| 6036 | } else { |
| 6037 | if (BytePerPixelDETC[k] > 0) { |
| 6038 | num_group_per_lower_vm_stage = |
| 6039 | dml_ceil((double) (dpde0_bytes_per_frame_ub_l[k]) / (double) (vm_group_bytes[k]), 1) |
| 6040 | + dml_ceil((double) (dpde0_bytes_per_frame_ub_c[k]) / (double) (vm_group_bytes[k]), 1) |
| 6041 | + dml_ceil((double) (meta_pte_bytes_per_frame_ub_l[k]) / (double) (vm_group_bytes[k]), 1) |
| 6042 | + dml_ceil((double) (meta_pte_bytes_per_frame_ub_c[k]) / (double) (vm_group_bytes[k]), 1); |
| 6043 | } else { |
| 6044 | num_group_per_lower_vm_stage = |
| 6045 | dml_ceil((double) (dpde0_bytes_per_frame_ub_l[k]) / (double) (vm_group_bytes[k]), 1) |
| 6046 | + dml_ceil((double) (meta_pte_bytes_per_frame_ub_l[k]) / (double) (vm_group_bytes[k]), 1); |
| 6047 | } |
| 6048 | } |
| 6049 | } |
| 6050 | |
| 6051 | if (DCCEnable[k] == false0) { |
| 6052 | if (BytePerPixelDETC[k] > 0) { |
| 6053 | num_req_per_lower_vm_stage = dpde0_bytes_per_frame_ub_l[k] |
| 6054 | / 64 + dpde0_bytes_per_frame_ub_c[k] / 64; |
| 6055 | } else { |
| 6056 | num_req_per_lower_vm_stage = dpde0_bytes_per_frame_ub_l[k] |
| 6057 | / 64; |
| 6058 | } |
| 6059 | } else { |
| 6060 | if (GPUVMMaxPageTableLevels == 1) { |
| 6061 | if (BytePerPixelDETC[k] > 0) { |
| 6062 | num_req_per_lower_vm_stage = meta_pte_bytes_per_frame_ub_l[k] / 64 |
| 6063 | + meta_pte_bytes_per_frame_ub_c[k] / 64; |
| 6064 | } else { |
| 6065 | num_req_per_lower_vm_stage = meta_pte_bytes_per_frame_ub_l[k] / 64; |
| 6066 | } |
| 6067 | } else { |
| 6068 | if (BytePerPixelDETC[k] > 0) { |
| 6069 | num_req_per_lower_vm_stage = dpde0_bytes_per_frame_ub_l[k] / 64 |
| 6070 | + dpde0_bytes_per_frame_ub_c[k] / 64 |
| 6071 | + meta_pte_bytes_per_frame_ub_l[k] / 64 |
| 6072 | + meta_pte_bytes_per_frame_ub_c[k] / 64; |
| 6073 | } else { |
| 6074 | num_req_per_lower_vm_stage = dpde0_bytes_per_frame_ub_l[k] / 64 |
| 6075 | + meta_pte_bytes_per_frame_ub_l[k] / 64; |
| 6076 | } |
| 6077 | } |
| 6078 | } |
| 6079 | |
| 6080 | TimePerVMGroupVBlank[k] = DestinationLinesToRequestVMInVBlank[k] * HTotal[k] |
| 6081 | / PixelClock[k] / num_group_per_lower_vm_stage; |
| 6082 | TimePerVMGroupFlip[k] = DestinationLinesToRequestVMInImmediateFlip[k] |
| 6083 | * HTotal[k] / PixelClock[k] / num_group_per_lower_vm_stage; |
| 6084 | TimePerVMRequestVBlank[k] = DestinationLinesToRequestVMInVBlank[k] |
| 6085 | * HTotal[k] / PixelClock[k] / num_req_per_lower_vm_stage; |
| 6086 | TimePerVMRequestFlip[k] = DestinationLinesToRequestVMInImmediateFlip[k] |
| 6087 | * HTotal[k] / PixelClock[k] / num_req_per_lower_vm_stage; |
| 6088 | |
| 6089 | if (GPUVMMaxPageTableLevels > 2) { |
| 6090 | TimePerVMGroupVBlank[k] = TimePerVMGroupVBlank[k] / 2; |
| 6091 | TimePerVMGroupFlip[k] = TimePerVMGroupFlip[k] / 2; |
| 6092 | TimePerVMRequestVBlank[k] = TimePerVMRequestVBlank[k] / 2; |
| 6093 | TimePerVMRequestFlip[k] = TimePerVMRequestFlip[k] / 2; |
| 6094 | } |
| 6095 | |
| 6096 | } else { |
| 6097 | TimePerVMGroupVBlank[k] = 0; |
| 6098 | TimePerVMGroupFlip[k] = 0; |
| 6099 | TimePerVMRequestVBlank[k] = 0; |
| 6100 | TimePerVMRequestFlip[k] = 0; |
| 6101 | } |
| 6102 | } |
| 6103 | } |
| 6104 | |
| 6105 | static double CalculateExtraLatency( |
| 6106 | double UrgentRoundTripAndOutOfOrderLatency, |
| 6107 | int TotalNumberOfActiveDPP, |
| 6108 | int PixelChunkSizeInKByte, |
| 6109 | int TotalNumberOfDCCActiveDPP, |
| 6110 | int MetaChunkSize, |
| 6111 | double ReturnBW, |
| 6112 | bool_Bool GPUVMEnable, |
| 6113 | bool_Bool HostVMEnable, |
| 6114 | int NumberOfActivePlanes, |
| 6115 | int NumberOfDPP[], |
| 6116 | int dpte_group_bytes[], |
| 6117 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
| 6118 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
| 6119 | int HostVMMaxPageTableLevels, |
| 6120 | int HostVMCachedPageTableLevels) |
| 6121 | { |
| 6122 | double CalculateExtraLatency; |
| 6123 | double HostVMInefficiencyFactor; |
| 6124 | int HostVMDynamicLevels; |
| 6125 | |
| 6126 | if (GPUVMEnable && HostVMEnable) { |
| 6127 | HostVMInefficiencyFactor = |
| 6128 | PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData |
| 6129 | / PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly; |
| 6130 | HostVMDynamicLevels = HostVMMaxPageTableLevels - HostVMCachedPageTableLevels; |
| 6131 | } else { |
| 6132 | HostVMInefficiencyFactor = 1; |
| 6133 | HostVMDynamicLevels = 0; |
| 6134 | } |
| 6135 | |
| 6136 | CalculateExtraLatency = UrgentRoundTripAndOutOfOrderLatency |
| 6137 | + (TotalNumberOfActiveDPP * PixelChunkSizeInKByte |
| 6138 | + TotalNumberOfDCCActiveDPP * MetaChunkSize) * 1024.0 |
| 6139 | / ReturnBW; |
| 6140 | |
| 6141 | if (GPUVMEnable) { |
| 6142 | int k; |
| 6143 | |
| 6144 | for (k = 0; k < NumberOfActivePlanes; k++) { |
| 6145 | CalculateExtraLatency = CalculateExtraLatency |
| 6146 | + NumberOfDPP[k] * dpte_group_bytes[k] |
| 6147 | * (1 + 8 * HostVMDynamicLevels) |
| 6148 | * HostVMInefficiencyFactor / ReturnBW; |
| 6149 | } |
| 6150 | } |
| 6151 | return CalculateExtraLatency; |
| 6152 | } |
| 6153 |