| File: | dev/pci/drm/amd/display/dc/dml/dcn31/display_mode_vba_31.c |
| Warning: | line 5613, column 4 Value stored to 'LinesInDETC' 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 | #include "dc.h" |
| 27 | #include "dc_link.h" |
| 28 | #include "../display_mode_lib.h" |
| 29 | #include "../dcn30/display_mode_vba_30.h" |
| 30 | #include "display_mode_vba_31.h" |
| 31 | #include "../dml_inline_defs.h" |
| 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 | |
| 42 | #define BPP_INVALID0 0 |
| 43 | #define BPP_BLENDED_PIPE0xffffffff 0xffffffff |
| 44 | #define DCN31_MAX_DSC_IMAGE_WIDTH5184 5184 |
| 45 | #define DCN31_MAX_FMT_420_BUFFER_WIDTH4096 4096 |
| 46 | #define DCN3_15_MIN_COMPBUF_SIZE_KB128 128 |
| 47 | #define DCN3_15_MAX_DET_SIZE384 384 |
| 48 | |
| 49 | // For DML-C changes that hasn't been propagated to VBA yet |
| 50 | //#define __DML_VBA_ALLOW_DELTA__ |
| 51 | |
| 52 | // Move these to ip paramaters/constant |
| 53 | |
| 54 | // At which vstartup the DML start to try if the mode can be supported |
| 55 | #define __DML_VBA_MIN_VSTARTUP__9 9 |
| 56 | |
| 57 | // Delay in DCFCLK from ARB to DET (1st num is ARB to SDPIF, 2nd number is SDPIF to DET) |
| 58 | #define __DML_ARB_TO_RET_DELAY__(7 + 95) (7 + 95) |
| 59 | |
| 60 | // fudge factor for min dcfclk calclation |
| 61 | #define __DML_MIN_DCFCLK_FACTOR__1.15 1.15 |
| 62 | |
| 63 | typedef struct { |
| 64 | double DPPCLK; |
| 65 | double DISPCLK; |
| 66 | double PixelClock; |
| 67 | double DCFCLKDeepSleep; |
| 68 | unsigned int DPPPerPlane; |
| 69 | bool_Bool ScalerEnabled; |
| 70 | double VRatio; |
| 71 | double VRatioChroma; |
| 72 | enum scan_direction_class SourceScan; |
| 73 | unsigned int BlockWidth256BytesY; |
| 74 | unsigned int BlockHeight256BytesY; |
| 75 | unsigned int BlockWidth256BytesC; |
| 76 | unsigned int BlockHeight256BytesC; |
| 77 | unsigned int InterlaceEnable; |
| 78 | unsigned int NumberOfCursors; |
| 79 | unsigned int VBlank; |
| 80 | unsigned int HTotal; |
| 81 | unsigned int DCCEnable; |
| 82 | bool_Bool ODMCombineIsEnabled; |
| 83 | enum source_format_class SourcePixelFormat; |
| 84 | int BytePerPixelY; |
| 85 | int BytePerPixelC; |
| 86 | bool_Bool ProgressiveToInterlaceUnitInOPP; |
| 87 | } Pipe; |
| 88 | |
| 89 | #define BPP_INVALID0 0 |
| 90 | #define BPP_BLENDED_PIPE0xffffffff 0xffffffff |
| 91 | |
| 92 | static void DisplayPipeConfiguration(struct display_mode_lib *mode_lib); |
| 93 | static void DISPCLKDPPCLKDCFCLKDeepSleepPrefetchParametersWatermarksAndPerformanceCalculation(struct display_mode_lib *mode_lib); |
| 94 | static unsigned int dscceComputeDelay( |
| 95 | unsigned int bpc, |
| 96 | double BPP, |
| 97 | unsigned int sliceWidth, |
| 98 | unsigned int numSlices, |
| 99 | enum output_format_class pixelFormat, |
| 100 | enum output_encoder_class Output); |
| 101 | static unsigned int dscComputeDelay(enum output_format_class pixelFormat, enum output_encoder_class Output); |
| 102 | static bool_Bool CalculatePrefetchSchedule( |
| 103 | struct display_mode_lib *mode_lib, |
| 104 | double HostVMInefficiencyFactor, |
| 105 | Pipe *myPipe, |
| 106 | unsigned int DSCDelay, |
| 107 | double DPPCLKDelaySubtotalPlusCNVCFormater, |
| 108 | double DPPCLKDelaySCL, |
| 109 | double DPPCLKDelaySCLLBOnly, |
| 110 | double DPPCLKDelayCNVCCursor, |
| 111 | double DISPCLKDelaySubtotal, |
| 112 | unsigned int DPP_RECOUT_WIDTH, |
| 113 | enum output_format_class OutputFormat, |
| 114 | unsigned int MaxInterDCNTileRepeaters, |
| 115 | unsigned int VStartup, |
| 116 | unsigned int MaxVStartup, |
| 117 | unsigned int GPUVMPageTableLevels, |
| 118 | bool_Bool GPUVMEnable, |
| 119 | bool_Bool HostVMEnable, |
| 120 | unsigned int HostVMMaxNonCachedPageTableLevels, |
| 121 | double HostVMMinPageSize, |
| 122 | bool_Bool DynamicMetadataEnable, |
| 123 | bool_Bool DynamicMetadataVMEnabled, |
| 124 | int DynamicMetadataLinesBeforeActiveRequired, |
| 125 | unsigned int DynamicMetadataTransmittedBytes, |
| 126 | double UrgentLatency, |
| 127 | double UrgentExtraLatency, |
| 128 | double TCalc, |
| 129 | unsigned int PDEAndMetaPTEBytesFrame, |
| 130 | unsigned int MetaRowByte, |
| 131 | unsigned int PixelPTEBytesPerRow, |
| 132 | double PrefetchSourceLinesY, |
| 133 | unsigned int SwathWidthY, |
| 134 | double VInitPreFillY, |
| 135 | unsigned int MaxNumSwathY, |
| 136 | double PrefetchSourceLinesC, |
| 137 | unsigned int SwathWidthC, |
| 138 | double VInitPreFillC, |
| 139 | unsigned int MaxNumSwathC, |
| 140 | int swath_width_luma_ub, |
| 141 | int swath_width_chroma_ub, |
| 142 | unsigned int SwathHeightY, |
| 143 | unsigned int SwathHeightC, |
| 144 | double TWait, |
| 145 | double *DSTXAfterScaler, |
| 146 | double *DSTYAfterScaler, |
| 147 | double *DestinationLinesForPrefetch, |
| 148 | double *PrefetchBandwidth, |
| 149 | double *DestinationLinesToRequestVMInVBlank, |
| 150 | double *DestinationLinesToRequestRowInVBlank, |
| 151 | double *VRatioPrefetchY, |
| 152 | double *VRatioPrefetchC, |
| 153 | double *RequiredPrefetchPixDataBWLuma, |
| 154 | double *RequiredPrefetchPixDataBWChroma, |
| 155 | bool_Bool *NotEnoughTimeForDynamicMetadata, |
| 156 | double *Tno_bw, |
| 157 | double *prefetch_vmrow_bw, |
| 158 | double *Tdmdl_vm, |
| 159 | double *Tdmdl, |
| 160 | double *TSetup, |
| 161 | int *VUpdateOffsetPix, |
| 162 | double *VUpdateWidthPix, |
| 163 | double *VReadyOffsetPix); |
| 164 | static double RoundToDFSGranularityUp(double Clock, double VCOSpeed); |
| 165 | static double RoundToDFSGranularityDown(double Clock, double VCOSpeed); |
| 166 | static void CalculateDCCConfiguration( |
| 167 | bool_Bool DCCEnabled, |
| 168 | bool_Bool DCCProgrammingAssumesScanDirectionUnknown, |
| 169 | enum source_format_class SourcePixelFormat, |
| 170 | unsigned int SurfaceWidthLuma, |
| 171 | unsigned int SurfaceWidthChroma, |
| 172 | unsigned int SurfaceHeightLuma, |
| 173 | unsigned int SurfaceHeightChroma, |
| 174 | double DETBufferSize, |
| 175 | unsigned int RequestHeight256ByteLuma, |
| 176 | unsigned int RequestHeight256ByteChroma, |
| 177 | enum dm_swizzle_mode TilingFormat, |
| 178 | unsigned int BytePerPixelY, |
| 179 | unsigned int BytePerPixelC, |
| 180 | double BytePerPixelDETY, |
| 181 | double BytePerPixelDETC, |
| 182 | enum scan_direction_class ScanOrientation, |
| 183 | unsigned int *MaxUncompressedBlockLuma, |
| 184 | unsigned int *MaxUncompressedBlockChroma, |
| 185 | unsigned int *MaxCompressedBlockLuma, |
| 186 | unsigned int *MaxCompressedBlockChroma, |
| 187 | unsigned int *IndependentBlockLuma, |
| 188 | unsigned int *IndependentBlockChroma); |
| 189 | static double CalculatePrefetchSourceLines( |
| 190 | struct display_mode_lib *mode_lib, |
| 191 | double VRatio, |
| 192 | double vtaps, |
| 193 | bool_Bool Interlace, |
| 194 | bool_Bool ProgressiveToInterlaceUnitInOPP, |
| 195 | unsigned int SwathHeight, |
| 196 | unsigned int ViewportYStart, |
| 197 | double *VInitPreFill, |
| 198 | unsigned int *MaxNumSwath); |
| 199 | static unsigned int CalculateVMAndRowBytes( |
| 200 | struct display_mode_lib *mode_lib, |
| 201 | bool_Bool DCCEnable, |
| 202 | unsigned int BlockHeight256Bytes, |
| 203 | unsigned int BlockWidth256Bytes, |
| 204 | enum source_format_class SourcePixelFormat, |
| 205 | unsigned int SurfaceTiling, |
| 206 | unsigned int BytePerPixel, |
| 207 | enum scan_direction_class ScanDirection, |
| 208 | unsigned int SwathWidth, |
| 209 | unsigned int ViewportHeight, |
| 210 | bool_Bool GPUVMEnable, |
| 211 | bool_Bool HostVMEnable, |
| 212 | unsigned int HostVMMaxNonCachedPageTableLevels, |
| 213 | unsigned int GPUVMMinPageSize, |
| 214 | unsigned int HostVMMinPageSize, |
| 215 | unsigned int PTEBufferSizeInRequests, |
| 216 | unsigned int Pitch, |
| 217 | unsigned int DCCMetaPitch, |
| 218 | unsigned int *MacroTileWidth, |
| 219 | unsigned int *MetaRowByte, |
| 220 | unsigned int *PixelPTEBytesPerRow, |
| 221 | bool_Bool *PTEBufferSizeNotExceeded, |
| 222 | int *dpte_row_width_ub, |
| 223 | unsigned int *dpte_row_height, |
| 224 | unsigned int *MetaRequestWidth, |
| 225 | unsigned int *MetaRequestHeight, |
| 226 | unsigned int *meta_row_width, |
| 227 | unsigned int *meta_row_height, |
| 228 | int *vm_group_bytes, |
| 229 | unsigned int *dpte_group_bytes, |
| 230 | unsigned int *PixelPTEReqWidth, |
| 231 | unsigned int *PixelPTEReqHeight, |
| 232 | unsigned int *PTERequestSize, |
| 233 | int *DPDE0BytesFrame, |
| 234 | int *MetaPTEBytesFrame); |
| 235 | static double CalculateTWait(unsigned int PrefetchMode, double DRAMClockChangeLatency, double UrgentLatency, double SREnterPlusExitTime); |
| 236 | static void CalculateRowBandwidth( |
| 237 | bool_Bool GPUVMEnable, |
| 238 | enum source_format_class SourcePixelFormat, |
| 239 | double VRatio, |
| 240 | double VRatioChroma, |
| 241 | bool_Bool DCCEnable, |
| 242 | double LineTime, |
| 243 | unsigned int MetaRowByteLuma, |
| 244 | unsigned int MetaRowByteChroma, |
| 245 | unsigned int meta_row_height_luma, |
| 246 | unsigned int meta_row_height_chroma, |
| 247 | unsigned int PixelPTEBytesPerRowLuma, |
| 248 | unsigned int PixelPTEBytesPerRowChroma, |
| 249 | unsigned int dpte_row_height_luma, |
| 250 | unsigned int dpte_row_height_chroma, |
| 251 | double *meta_row_bw, |
| 252 | double *dpte_row_bw); |
| 253 | |
| 254 | static void CalculateFlipSchedule( |
| 255 | struct display_mode_lib *mode_lib, |
| 256 | unsigned int k, |
| 257 | double HostVMInefficiencyFactor, |
| 258 | double UrgentExtraLatency, |
| 259 | double UrgentLatency, |
| 260 | double PDEAndMetaPTEBytesPerFrame, |
| 261 | double MetaRowBytes, |
| 262 | double DPTEBytesPerRow); |
| 263 | static double CalculateWriteBackDelay( |
| 264 | enum source_format_class WritebackPixelFormat, |
| 265 | double WritebackHRatio, |
| 266 | double WritebackVRatio, |
| 267 | unsigned int WritebackVTaps, |
| 268 | int WritebackDestinationWidth, |
| 269 | int WritebackDestinationHeight, |
| 270 | int WritebackSourceHeight, |
| 271 | unsigned int HTotal); |
| 272 | |
| 273 | static void CalculateVupdateAndDynamicMetadataParameters( |
| 274 | int MaxInterDCNTileRepeaters, |
| 275 | double DPPCLK, |
| 276 | double DISPCLK, |
| 277 | double DCFClkDeepSleep, |
| 278 | double PixelClock, |
| 279 | int HTotal, |
| 280 | int VBlank, |
| 281 | int DynamicMetadataTransmittedBytes, |
| 282 | int DynamicMetadataLinesBeforeActiveRequired, |
| 283 | int InterlaceEnable, |
| 284 | bool_Bool ProgressiveToInterlaceUnitInOPP, |
| 285 | double *TSetup, |
| 286 | double *Tdmbf, |
| 287 | double *Tdmec, |
| 288 | double *Tdmsks, |
| 289 | int *VUpdateOffsetPix, |
| 290 | double *VUpdateWidthPix, |
| 291 | double *VReadyOffsetPix); |
| 292 | |
| 293 | static void CalculateWatermarksAndDRAMSpeedChangeSupport( |
| 294 | struct display_mode_lib *mode_lib, |
| 295 | unsigned int PrefetchMode, |
| 296 | double DCFCLK, |
| 297 | double ReturnBW, |
| 298 | double UrgentLatency, |
| 299 | double ExtraLatency, |
| 300 | double SOCCLK, |
| 301 | double DCFCLKDeepSleep, |
| 302 | unsigned int DETBufferSizeY[], |
| 303 | unsigned int DETBufferSizeC[], |
| 304 | unsigned int SwathHeightY[], |
| 305 | unsigned int SwathHeightC[], |
| 306 | double SwathWidthY[], |
| 307 | double SwathWidthC[], |
| 308 | unsigned int DPPPerPlane[], |
| 309 | double BytePerPixelDETY[], |
| 310 | double BytePerPixelDETC[], |
| 311 | bool_Bool UnboundedRequestEnabled, |
| 312 | int unsigned CompressedBufferSizeInkByte, |
| 313 | enum clock_change_support *DRAMClockChangeSupport, |
| 314 | double *StutterExitWatermark, |
| 315 | double *StutterEnterPlusExitWatermark, |
| 316 | double *Z8StutterExitWatermark, |
| 317 | double *Z8StutterEnterPlusExitWatermark); |
| 318 | |
| 319 | static void CalculateDCFCLKDeepSleep( |
| 320 | struct display_mode_lib *mode_lib, |
| 321 | unsigned int NumberOfActivePlanes, |
| 322 | int BytePerPixelY[], |
| 323 | int BytePerPixelC[], |
| 324 | double VRatio[], |
| 325 | double VRatioChroma[], |
| 326 | double SwathWidthY[], |
| 327 | double SwathWidthC[], |
| 328 | unsigned int DPPPerPlane[], |
| 329 | double HRatio[], |
| 330 | double HRatioChroma[], |
| 331 | double PixelClock[], |
| 332 | double PSCL_THROUGHPUT[], |
| 333 | double PSCL_THROUGHPUT_CHROMA[], |
| 334 | double DPPCLK[], |
| 335 | double ReadBandwidthLuma[], |
| 336 | double ReadBandwidthChroma[], |
| 337 | int ReturnBusWidth, |
| 338 | double *DCFCLKDeepSleep); |
| 339 | |
| 340 | static void CalculateUrgentBurstFactor( |
| 341 | int swath_width_luma_ub, |
| 342 | int swath_width_chroma_ub, |
| 343 | unsigned int SwathHeightY, |
| 344 | unsigned int SwathHeightC, |
| 345 | double LineTime, |
| 346 | double UrgentLatency, |
| 347 | double CursorBufferSize, |
| 348 | unsigned int CursorWidth, |
| 349 | unsigned int CursorBPP, |
| 350 | double VRatio, |
| 351 | double VRatioC, |
| 352 | double BytePerPixelInDETY, |
| 353 | double BytePerPixelInDETC, |
| 354 | double DETBufferSizeY, |
| 355 | double DETBufferSizeC, |
| 356 | double *UrgentBurstFactorCursor, |
| 357 | double *UrgentBurstFactorLuma, |
| 358 | double *UrgentBurstFactorChroma, |
| 359 | bool_Bool *NotEnoughUrgentLatencyHiding); |
| 360 | |
| 361 | static void UseMinimumDCFCLK( |
| 362 | struct display_mode_lib *mode_lib, |
| 363 | int MaxPrefetchMode, |
| 364 | int ReorderingBytes); |
| 365 | |
| 366 | static void CalculatePixelDeliveryTimes( |
| 367 | unsigned int NumberOfActivePlanes, |
| 368 | double VRatio[], |
| 369 | double VRatioChroma[], |
| 370 | double VRatioPrefetchY[], |
| 371 | double VRatioPrefetchC[], |
| 372 | unsigned int swath_width_luma_ub[], |
| 373 | unsigned int swath_width_chroma_ub[], |
| 374 | unsigned int DPPPerPlane[], |
| 375 | double HRatio[], |
| 376 | double HRatioChroma[], |
| 377 | double PixelClock[], |
| 378 | double PSCL_THROUGHPUT[], |
| 379 | double PSCL_THROUGHPUT_CHROMA[], |
| 380 | double DPPCLK[], |
| 381 | int BytePerPixelC[], |
| 382 | enum scan_direction_class SourceScan[], |
| 383 | unsigned int NumberOfCursors[], |
| 384 | unsigned int CursorWidth[][DC__NUM_CURSOR__MAX2], |
| 385 | unsigned int CursorBPP[][DC__NUM_CURSOR__MAX2], |
| 386 | unsigned int BlockWidth256BytesY[], |
| 387 | unsigned int BlockHeight256BytesY[], |
| 388 | unsigned int BlockWidth256BytesC[], |
| 389 | unsigned int BlockHeight256BytesC[], |
| 390 | double DisplayPipeLineDeliveryTimeLuma[], |
| 391 | double DisplayPipeLineDeliveryTimeChroma[], |
| 392 | double DisplayPipeLineDeliveryTimeLumaPrefetch[], |
| 393 | double DisplayPipeLineDeliveryTimeChromaPrefetch[], |
| 394 | double DisplayPipeRequestDeliveryTimeLuma[], |
| 395 | double DisplayPipeRequestDeliveryTimeChroma[], |
| 396 | double DisplayPipeRequestDeliveryTimeLumaPrefetch[], |
| 397 | double DisplayPipeRequestDeliveryTimeChromaPrefetch[], |
| 398 | double CursorRequestDeliveryTime[], |
| 399 | double CursorRequestDeliveryTimePrefetch[]); |
| 400 | |
| 401 | static void CalculateMetaAndPTETimes( |
| 402 | int NumberOfActivePlanes, |
| 403 | bool_Bool GPUVMEnable, |
| 404 | int MetaChunkSize, |
| 405 | int MinMetaChunkSizeBytes, |
| 406 | int HTotal[], |
| 407 | double VRatio[], |
| 408 | double VRatioChroma[], |
| 409 | double DestinationLinesToRequestRowInVBlank[], |
| 410 | double DestinationLinesToRequestRowInImmediateFlip[], |
| 411 | bool_Bool DCCEnable[], |
| 412 | double PixelClock[], |
| 413 | int BytePerPixelY[], |
| 414 | int BytePerPixelC[], |
| 415 | enum scan_direction_class SourceScan[], |
| 416 | int dpte_row_height[], |
| 417 | int dpte_row_height_chroma[], |
| 418 | int meta_row_width[], |
| 419 | int meta_row_width_chroma[], |
| 420 | int meta_row_height[], |
| 421 | int meta_row_height_chroma[], |
| 422 | int meta_req_width[], |
| 423 | int meta_req_width_chroma[], |
| 424 | int meta_req_height[], |
| 425 | int meta_req_height_chroma[], |
| 426 | int dpte_group_bytes[], |
| 427 | int PTERequestSizeY[], |
| 428 | int PTERequestSizeC[], |
| 429 | int PixelPTEReqWidthY[], |
| 430 | int PixelPTEReqHeightY[], |
| 431 | int PixelPTEReqWidthC[], |
| 432 | int PixelPTEReqHeightC[], |
| 433 | int dpte_row_width_luma_ub[], |
| 434 | int dpte_row_width_chroma_ub[], |
| 435 | double DST_Y_PER_PTE_ROW_NOM_L[], |
| 436 | double DST_Y_PER_PTE_ROW_NOM_C[], |
| 437 | double DST_Y_PER_META_ROW_NOM_L[], |
| 438 | double DST_Y_PER_META_ROW_NOM_C[], |
| 439 | double TimePerMetaChunkNominal[], |
| 440 | double TimePerChromaMetaChunkNominal[], |
| 441 | double TimePerMetaChunkVBlank[], |
| 442 | double TimePerChromaMetaChunkVBlank[], |
| 443 | double TimePerMetaChunkFlip[], |
| 444 | double TimePerChromaMetaChunkFlip[], |
| 445 | double time_per_pte_group_nom_luma[], |
| 446 | double time_per_pte_group_vblank_luma[], |
| 447 | double time_per_pte_group_flip_luma[], |
| 448 | double time_per_pte_group_nom_chroma[], |
| 449 | double time_per_pte_group_vblank_chroma[], |
| 450 | double time_per_pte_group_flip_chroma[]); |
| 451 | |
| 452 | static void CalculateVMGroupAndRequestTimes( |
| 453 | unsigned int NumberOfActivePlanes, |
| 454 | bool_Bool GPUVMEnable, |
| 455 | unsigned int GPUVMMaxPageTableLevels, |
| 456 | unsigned int HTotal[], |
| 457 | int BytePerPixelC[], |
| 458 | double DestinationLinesToRequestVMInVBlank[], |
| 459 | double DestinationLinesToRequestVMInImmediateFlip[], |
| 460 | bool_Bool DCCEnable[], |
| 461 | double PixelClock[], |
| 462 | int dpte_row_width_luma_ub[], |
| 463 | int dpte_row_width_chroma_ub[], |
| 464 | int vm_group_bytes[], |
| 465 | unsigned int dpde0_bytes_per_frame_ub_l[], |
| 466 | unsigned int dpde0_bytes_per_frame_ub_c[], |
| 467 | int meta_pte_bytes_per_frame_ub_l[], |
| 468 | int meta_pte_bytes_per_frame_ub_c[], |
| 469 | double TimePerVMGroupVBlank[], |
| 470 | double TimePerVMGroupFlip[], |
| 471 | double TimePerVMRequestVBlank[], |
| 472 | double TimePerVMRequestFlip[]); |
| 473 | |
| 474 | static void CalculateStutterEfficiency( |
| 475 | struct display_mode_lib *mode_lib, |
| 476 | int CompressedBufferSizeInkByte, |
| 477 | bool_Bool UnboundedRequestEnabled, |
| 478 | int ConfigReturnBufferSizeInKByte, |
| 479 | int MetaFIFOSizeInKEntries, |
| 480 | int ZeroSizeBufferEntries, |
| 481 | int NumberOfActivePlanes, |
| 482 | int ROBBufferSizeInKByte, |
| 483 | double TotalDataReadBandwidth, |
| 484 | double DCFCLK, |
| 485 | double ReturnBW, |
| 486 | double COMPBUF_RESERVED_SPACE_64B, |
| 487 | double COMPBUF_RESERVED_SPACE_ZS, |
| 488 | double SRExitTime, |
| 489 | double SRExitZ8Time, |
| 490 | bool_Bool SynchronizedVBlank, |
| 491 | double Z8StutterEnterPlusExitWatermark, |
| 492 | double StutterEnterPlusExitWatermark, |
| 493 | bool_Bool ProgressiveToInterlaceUnitInOPP, |
| 494 | bool_Bool Interlace[], |
| 495 | double MinTTUVBlank[], |
| 496 | int DPPPerPlane[], |
| 497 | unsigned int DETBufferSizeY[], |
| 498 | int BytePerPixelY[], |
| 499 | double BytePerPixelDETY[], |
| 500 | double SwathWidthY[], |
| 501 | int SwathHeightY[], |
| 502 | int SwathHeightC[], |
| 503 | double NetDCCRateLuma[], |
| 504 | double NetDCCRateChroma[], |
| 505 | double DCCFractionOfZeroSizeRequestsLuma[], |
| 506 | double DCCFractionOfZeroSizeRequestsChroma[], |
| 507 | int HTotal[], |
| 508 | int VTotal[], |
| 509 | double PixelClock[], |
| 510 | double VRatio[], |
| 511 | enum scan_direction_class SourceScan[], |
| 512 | int BlockHeight256BytesY[], |
| 513 | int BlockWidth256BytesY[], |
| 514 | int BlockHeight256BytesC[], |
| 515 | int BlockWidth256BytesC[], |
| 516 | int DCCYMaxUncompressedBlock[], |
| 517 | int DCCCMaxUncompressedBlock[], |
| 518 | int VActive[], |
| 519 | bool_Bool DCCEnable[], |
| 520 | bool_Bool WritebackEnable[], |
| 521 | double ReadBandwidthPlaneLuma[], |
| 522 | double ReadBandwidthPlaneChroma[], |
| 523 | double meta_row_bw[], |
| 524 | double dpte_row_bw[], |
| 525 | double *StutterEfficiencyNotIncludingVBlank, |
| 526 | double *StutterEfficiency, |
| 527 | int *NumberOfStutterBurstsPerFrame, |
| 528 | double *Z8StutterEfficiencyNotIncludingVBlank, |
| 529 | double *Z8StutterEfficiency, |
| 530 | int *Z8NumberOfStutterBurstsPerFrame, |
| 531 | double *StutterPeriod); |
| 532 | |
| 533 | static void CalculateSwathAndDETConfiguration( |
| 534 | bool_Bool ForceSingleDPP, |
| 535 | int NumberOfActivePlanes, |
| 536 | bool_Bool DETSharedByAllDPP, |
| 537 | unsigned int DETBufferSizeInKByte[], |
| 538 | double MaximumSwathWidthLuma[], |
| 539 | double MaximumSwathWidthChroma[], |
| 540 | enum scan_direction_class SourceScan[], |
| 541 | enum source_format_class SourcePixelFormat[], |
| 542 | enum dm_swizzle_mode SurfaceTiling[], |
| 543 | int ViewportWidth[], |
| 544 | int ViewportHeight[], |
| 545 | int SurfaceWidthY[], |
| 546 | int SurfaceWidthC[], |
| 547 | int SurfaceHeightY[], |
| 548 | int SurfaceHeightC[], |
| 549 | int Read256BytesBlockHeightY[], |
| 550 | int Read256BytesBlockHeightC[], |
| 551 | int Read256BytesBlockWidthY[], |
| 552 | int Read256BytesBlockWidthC[], |
| 553 | enum odm_combine_mode ODMCombineEnabled[], |
| 554 | int BlendingAndTiming[], |
| 555 | int BytePerPixY[], |
| 556 | int BytePerPixC[], |
| 557 | double BytePerPixDETY[], |
| 558 | double BytePerPixDETC[], |
| 559 | int HActive[], |
| 560 | double HRatio[], |
| 561 | double HRatioChroma[], |
| 562 | int DPPPerPlane[], |
| 563 | int swath_width_luma_ub[], |
| 564 | int swath_width_chroma_ub[], |
| 565 | double SwathWidth[], |
| 566 | double SwathWidthChroma[], |
| 567 | int SwathHeightY[], |
| 568 | int SwathHeightC[], |
| 569 | unsigned int DETBufferSizeY[], |
| 570 | unsigned int DETBufferSizeC[], |
| 571 | bool_Bool ViewportSizeSupportPerPlane[], |
| 572 | bool_Bool *ViewportSizeSupport); |
| 573 | static void CalculateSwathWidth( |
| 574 | bool_Bool ForceSingleDPP, |
| 575 | int NumberOfActivePlanes, |
| 576 | enum source_format_class SourcePixelFormat[], |
| 577 | enum scan_direction_class SourceScan[], |
| 578 | int ViewportWidth[], |
| 579 | int ViewportHeight[], |
| 580 | int SurfaceWidthY[], |
| 581 | int SurfaceWidthC[], |
| 582 | int SurfaceHeightY[], |
| 583 | int SurfaceHeightC[], |
| 584 | enum odm_combine_mode ODMCombineEnabled[], |
| 585 | int BytePerPixY[], |
| 586 | int BytePerPixC[], |
| 587 | int Read256BytesBlockHeightY[], |
| 588 | int Read256BytesBlockHeightC[], |
| 589 | int Read256BytesBlockWidthY[], |
| 590 | int Read256BytesBlockWidthC[], |
| 591 | int BlendingAndTiming[], |
| 592 | int HActive[], |
| 593 | double HRatio[], |
| 594 | int DPPPerPlane[], |
| 595 | double SwathWidthSingleDPPY[], |
| 596 | double SwathWidthSingleDPPC[], |
| 597 | double SwathWidthY[], |
| 598 | double SwathWidthC[], |
| 599 | int MaximumSwathHeightY[], |
| 600 | int MaximumSwathHeightC[], |
| 601 | int swath_width_luma_ub[], |
| 602 | int swath_width_chroma_ub[]); |
| 603 | |
| 604 | static double CalculateExtraLatency( |
| 605 | int RoundTripPingLatencyCycles, |
| 606 | int ReorderingBytes, |
| 607 | double DCFCLK, |
| 608 | int TotalNumberOfActiveDPP, |
| 609 | int PixelChunkSizeInKByte, |
| 610 | int TotalNumberOfDCCActiveDPP, |
| 611 | int MetaChunkSize, |
| 612 | double ReturnBW, |
| 613 | bool_Bool GPUVMEnable, |
| 614 | bool_Bool HostVMEnable, |
| 615 | int NumberOfActivePlanes, |
| 616 | int NumberOfDPP[], |
| 617 | int dpte_group_bytes[], |
| 618 | double HostVMInefficiencyFactor, |
| 619 | double HostVMMinPageSize, |
| 620 | int HostVMMaxNonCachedPageTableLevels); |
| 621 | |
| 622 | static double CalculateExtraLatencyBytes( |
| 623 | int ReorderingBytes, |
| 624 | int TotalNumberOfActiveDPP, |
| 625 | int PixelChunkSizeInKByte, |
| 626 | int TotalNumberOfDCCActiveDPP, |
| 627 | int MetaChunkSize, |
| 628 | bool_Bool GPUVMEnable, |
| 629 | bool_Bool HostVMEnable, |
| 630 | int NumberOfActivePlanes, |
| 631 | int NumberOfDPP[], |
| 632 | int dpte_group_bytes[], |
| 633 | double HostVMInefficiencyFactor, |
| 634 | double HostVMMinPageSize, |
| 635 | int HostVMMaxNonCachedPageTableLevels); |
| 636 | |
| 637 | static double CalculateUrgentLatency( |
| 638 | double UrgentLatencyPixelDataOnly, |
| 639 | double UrgentLatencyPixelMixedWithVMData, |
| 640 | double UrgentLatencyVMDataOnly, |
| 641 | bool_Bool DoUrgentLatencyAdjustment, |
| 642 | double UrgentLatencyAdjustmentFabricClockComponent, |
| 643 | double UrgentLatencyAdjustmentFabricClockReference, |
| 644 | double FabricClockSingle); |
| 645 | |
| 646 | static void CalculateUnboundedRequestAndCompressedBufferSize( |
| 647 | unsigned int DETBufferSizeInKByte, |
| 648 | int ConfigReturnBufferSizeInKByte, |
| 649 | enum unbounded_requesting_policy UseUnboundedRequestingFinal, |
| 650 | int TotalActiveDPP, |
| 651 | bool_Bool NoChromaPlanes, |
| 652 | int MaxNumDPP, |
| 653 | int CompressedBufferSegmentSizeInkByteFinal, |
| 654 | enum output_encoder_class *Output, |
| 655 | bool_Bool *UnboundedRequestEnabled, |
| 656 | int *CompressedBufferSizeInkByte); |
| 657 | |
| 658 | static bool_Bool UnboundedRequest(enum unbounded_requesting_policy UseUnboundedRequestingFinal, int TotalNumberOfActiveDPP, bool_Bool NoChroma, enum output_encoder_class Output); |
| 659 | |
| 660 | void dml31_recalculate(struct display_mode_lib *mode_lib) |
| 661 | { |
| 662 | ModeSupportAndSystemConfiguration(mode_lib); |
| 663 | PixelClockAdjustmentForProgressiveToInterlaceUnit(mode_lib); |
| 664 | DisplayPipeConfiguration(mode_lib); |
| 665 | #ifdef __DML_VBA_DEBUG__ |
| 666 | dml_print("DML::%s: Calling DISPCLKDPPCLKDCFCLKDeepSleepPrefetchParametersWatermarksAndPerformanceCalculation\n", __func__){do { } while(0); }; |
| 667 | #endif |
| 668 | DISPCLKDPPCLKDCFCLKDeepSleepPrefetchParametersWatermarksAndPerformanceCalculation(mode_lib); |
| 669 | } |
| 670 | |
| 671 | static unsigned int dscceComputeDelay( |
| 672 | unsigned int bpc, |
| 673 | double BPP, |
| 674 | unsigned int sliceWidth, |
| 675 | unsigned int numSlices, |
| 676 | enum output_format_class pixelFormat, |
| 677 | enum output_encoder_class Output) |
| 678 | { |
| 679 | // valid bpc = source bits per component in the set of {8, 10, 12} |
| 680 | // valid bpp = increments of 1/16 of a bit |
| 681 | // min = 6/7/8 in N420/N422/444, respectively |
| 682 | // max = such that compression is 1:1 |
| 683 | //valid sliceWidth = number of pixels per slice line, must be less than or equal to 5184/numSlices (or 4096/numSlices in 420 mode) |
| 684 | //valid numSlices = number of slices in the horiziontal direction per DSC engine in the set of {1, 2, 3, 4} |
| 685 | //valid pixelFormat = pixel/color format in the set of {:N444_RGB, :S422, :N422, :N420} |
| 686 | |
| 687 | // fixed value |
| 688 | unsigned int rcModelSize = 8192; |
| 689 | |
| 690 | // N422/N420 operate at 2 pixels per clock |
| 691 | unsigned int pixelsPerClock = 0, lstall, D, initalXmitDelay, w, s, ix, wx, P, l0, a, ax, L, Delay, pixels; |
| 692 | |
| 693 | if (pixelFormat == dm_420) |
| 694 | pixelsPerClock = 2; |
| 695 | else if (pixelFormat == dm_444) |
| 696 | pixelsPerClock = 1; |
| 697 | else if (pixelFormat == dm_n422) |
| 698 | pixelsPerClock = 2; |
| 699 | // #all other modes operate at 1 pixel per clock |
| 700 | else |
| 701 | pixelsPerClock = 1; |
| 702 | |
| 703 | //initial transmit delay as per PPS |
| 704 | initalXmitDelay = dml_round(rcModelSize / 2.0 / BPP / pixelsPerClock); |
| 705 | |
| 706 | //compute ssm delay |
| 707 | if (bpc == 8) |
| 708 | D = 81; |
| 709 | else if (bpc == 10) |
| 710 | D = 89; |
| 711 | else |
| 712 | D = 113; |
| 713 | |
| 714 | //divide by pixel per cycle to compute slice width as seen by DSC |
| 715 | w = sliceWidth / pixelsPerClock; |
| 716 | |
| 717 | //422 mode has an additional cycle of delay |
| 718 | if (pixelFormat == dm_420 || pixelFormat == dm_444 || pixelFormat == dm_n422) |
| 719 | s = 0; |
| 720 | else |
| 721 | s = 1; |
| 722 | |
| 723 | //main calculation for the dscce |
| 724 | ix = initalXmitDelay + 45; |
| 725 | wx = (w + 2) / 3; |
| 726 | P = 3 * wx - w; |
| 727 | l0 = ix / w; |
| 728 | a = ix + P * l0; |
| 729 | ax = (a + 2) / 3 + D + 6 + 1; |
| 730 | L = (ax + wx - 1) / wx; |
| 731 | if ((ix % w) == 0 && P != 0) |
| 732 | lstall = 1; |
| 733 | else |
| 734 | lstall = 0; |
| 735 | Delay = L * wx * (numSlices - 1) + ax + s + lstall + 22; |
| 736 | |
| 737 | //dsc processes 3 pixel containers per cycle and a container can contain 1 or 2 pixels |
| 738 | pixels = Delay * 3 * pixelsPerClock; |
| 739 | return pixels; |
| 740 | } |
| 741 | |
| 742 | static unsigned int dscComputeDelay(enum output_format_class pixelFormat, enum output_encoder_class Output) |
| 743 | { |
| 744 | unsigned int Delay = 0; |
| 745 | |
| 746 | if (pixelFormat == dm_420) { |
| 747 | // sfr |
| 748 | Delay = Delay + 2; |
| 749 | // dsccif |
| 750 | Delay = Delay + 0; |
| 751 | // dscc - input deserializer |
| 752 | Delay = Delay + 3; |
| 753 | // dscc gets pixels every other cycle |
| 754 | Delay = Delay + 2; |
| 755 | // dscc - input cdc fifo |
| 756 | Delay = Delay + 12; |
| 757 | // dscc gets pixels every other cycle |
| 758 | Delay = Delay + 13; |
| 759 | // dscc - cdc uncertainty |
| 760 | Delay = Delay + 2; |
| 761 | // dscc - output cdc fifo |
| 762 | Delay = Delay + 7; |
| 763 | // dscc gets pixels every other cycle |
| 764 | Delay = Delay + 3; |
| 765 | // dscc - cdc uncertainty |
| 766 | Delay = Delay + 2; |
| 767 | // dscc - output serializer |
| 768 | Delay = Delay + 1; |
| 769 | // sft |
| 770 | Delay = Delay + 1; |
| 771 | } else if (pixelFormat == dm_n422) { |
| 772 | // sfr |
| 773 | Delay = Delay + 2; |
| 774 | // dsccif |
| 775 | Delay = Delay + 1; |
| 776 | // dscc - input deserializer |
| 777 | Delay = Delay + 5; |
| 778 | // dscc - input cdc fifo |
| 779 | Delay = Delay + 25; |
| 780 | // dscc - cdc uncertainty |
| 781 | Delay = Delay + 2; |
| 782 | // dscc - output cdc fifo |
| 783 | Delay = Delay + 10; |
| 784 | // dscc - cdc uncertainty |
| 785 | Delay = Delay + 2; |
| 786 | // dscc - output serializer |
| 787 | Delay = Delay + 1; |
| 788 | // sft |
| 789 | Delay = Delay + 1; |
| 790 | } else { |
| 791 | // sfr |
| 792 | Delay = Delay + 2; |
| 793 | // dsccif |
| 794 | Delay = Delay + 0; |
| 795 | // dscc - input deserializer |
| 796 | Delay = Delay + 3; |
| 797 | // dscc - input cdc fifo |
| 798 | Delay = Delay + 12; |
| 799 | // dscc - cdc uncertainty |
| 800 | Delay = Delay + 2; |
| 801 | // dscc - output cdc fifo |
| 802 | Delay = Delay + 7; |
| 803 | // dscc - output serializer |
| 804 | Delay = Delay + 1; |
| 805 | // dscc - cdc uncertainty |
| 806 | Delay = Delay + 2; |
| 807 | // sft |
| 808 | Delay = Delay + 1; |
| 809 | } |
| 810 | |
| 811 | return Delay; |
| 812 | } |
| 813 | |
| 814 | static bool_Bool CalculatePrefetchSchedule( |
| 815 | struct display_mode_lib *mode_lib, |
| 816 | double HostVMInefficiencyFactor, |
| 817 | Pipe *myPipe, |
| 818 | unsigned int DSCDelay, |
| 819 | double DPPCLKDelaySubtotalPlusCNVCFormater, |
| 820 | double DPPCLKDelaySCL, |
| 821 | double DPPCLKDelaySCLLBOnly, |
| 822 | double DPPCLKDelayCNVCCursor, |
| 823 | double DISPCLKDelaySubtotal, |
| 824 | unsigned int DPP_RECOUT_WIDTH, |
| 825 | enum output_format_class OutputFormat, |
| 826 | unsigned int MaxInterDCNTileRepeaters, |
| 827 | unsigned int VStartup, |
| 828 | unsigned int MaxVStartup, |
| 829 | unsigned int GPUVMPageTableLevels, |
| 830 | bool_Bool GPUVMEnable, |
| 831 | bool_Bool HostVMEnable, |
| 832 | unsigned int HostVMMaxNonCachedPageTableLevels, |
| 833 | double HostVMMinPageSize, |
| 834 | bool_Bool DynamicMetadataEnable, |
| 835 | bool_Bool DynamicMetadataVMEnabled, |
| 836 | int DynamicMetadataLinesBeforeActiveRequired, |
| 837 | unsigned int DynamicMetadataTransmittedBytes, |
| 838 | double UrgentLatency, |
| 839 | double UrgentExtraLatency, |
| 840 | double TCalc, |
| 841 | unsigned int PDEAndMetaPTEBytesFrame, |
| 842 | unsigned int MetaRowByte, |
| 843 | unsigned int PixelPTEBytesPerRow, |
| 844 | double PrefetchSourceLinesY, |
| 845 | unsigned int SwathWidthY, |
| 846 | double VInitPreFillY, |
| 847 | unsigned int MaxNumSwathY, |
| 848 | double PrefetchSourceLinesC, |
| 849 | unsigned int SwathWidthC, |
| 850 | double VInitPreFillC, |
| 851 | unsigned int MaxNumSwathC, |
| 852 | int swath_width_luma_ub, |
| 853 | int swath_width_chroma_ub, |
| 854 | unsigned int SwathHeightY, |
| 855 | unsigned int SwathHeightC, |
| 856 | double TWait, |
| 857 | double *DSTXAfterScaler, |
| 858 | double *DSTYAfterScaler, |
| 859 | double *DestinationLinesForPrefetch, |
| 860 | double *PrefetchBandwidth, |
| 861 | double *DestinationLinesToRequestVMInVBlank, |
| 862 | double *DestinationLinesToRequestRowInVBlank, |
| 863 | double *VRatioPrefetchY, |
| 864 | double *VRatioPrefetchC, |
| 865 | double *RequiredPrefetchPixDataBWLuma, |
| 866 | double *RequiredPrefetchPixDataBWChroma, |
| 867 | bool_Bool *NotEnoughTimeForDynamicMetadata, |
| 868 | double *Tno_bw, |
| 869 | double *prefetch_vmrow_bw, |
| 870 | double *Tdmdl_vm, |
| 871 | double *Tdmdl, |
| 872 | double *TSetup, |
| 873 | int *VUpdateOffsetPix, |
| 874 | double *VUpdateWidthPix, |
| 875 | double *VReadyOffsetPix) |
| 876 | { |
| 877 | bool_Bool MyError = false0; |
| 878 | unsigned int DPPCycles, DISPCLKCycles; |
| 879 | double DSTTotalPixelsAfterScaler; |
| 880 | double LineTime; |
| 881 | double dst_y_prefetch_equ; |
| 882 | double Tsw_oto; |
| 883 | double prefetch_bw_oto; |
| 884 | double prefetch_bw_pr; |
| 885 | double Tvm_oto; |
| 886 | double Tr0_oto; |
| 887 | double Tvm_oto_lines; |
| 888 | double Tr0_oto_lines; |
| 889 | double dst_y_prefetch_oto; |
| 890 | double TimeForFetchingMetaPTE = 0; |
| 891 | double TimeForFetchingRowInVBlank = 0; |
| 892 | double LinesToRequestPrefetchPixelData = 0; |
| 893 | unsigned int HostVMDynamicLevelsTrips; |
| 894 | double trip_to_mem; |
| 895 | double Tvm_trips; |
| 896 | double Tr0_trips; |
| 897 | double Tvm_trips_rounded; |
| 898 | double Tr0_trips_rounded; |
| 899 | double Lsw_oto; |
| 900 | double Tpre_rounded; |
| 901 | double prefetch_bw_equ; |
| 902 | double Tvm_equ; |
| 903 | double Tr0_equ; |
| 904 | double Tdmbf; |
| 905 | double Tdmec; |
| 906 | double Tdmsks; |
| 907 | double prefetch_sw_bytes; |
| 908 | double bytes_pp; |
| 909 | double dep_bytes; |
| 910 | int max_vratio_pre = 4; |
| 911 | double min_Lsw; |
| 912 | double Tsw_est1 = 0; |
| 913 | double Tsw_est3 = 0; |
| 914 | double max_Tsw = 0; |
| 915 | |
| 916 | if (GPUVMEnable == true1 && HostVMEnable == true1) { |
| 917 | HostVMDynamicLevelsTrips = HostVMMaxNonCachedPageTableLevels; |
| 918 | } else { |
| 919 | HostVMDynamicLevelsTrips = 0; |
| 920 | } |
| 921 | #ifdef __DML_VBA_DEBUG__ |
| 922 | dml_print("DML::%s: GPUVMEnable=%d HostVMEnable=%d HostVMInefficiencyFactor=%f\n", __func__, GPUVMEnable, HostVMEnable, HostVMInefficiencyFactor){do { } while(0); }; |
| 923 | #endif |
| 924 | CalculateVupdateAndDynamicMetadataParameters( |
| 925 | MaxInterDCNTileRepeaters, |
| 926 | myPipe->DPPCLK, |
| 927 | myPipe->DISPCLK, |
| 928 | myPipe->DCFCLKDeepSleep, |
| 929 | myPipe->PixelClock, |
| 930 | myPipe->HTotal, |
| 931 | myPipe->VBlank, |
| 932 | DynamicMetadataTransmittedBytes, |
| 933 | DynamicMetadataLinesBeforeActiveRequired, |
| 934 | myPipe->InterlaceEnable, |
| 935 | myPipe->ProgressiveToInterlaceUnitInOPP, |
| 936 | TSetup, |
| 937 | &Tdmbf, |
| 938 | &Tdmec, |
| 939 | &Tdmsks, |
| 940 | VUpdateOffsetPix, |
| 941 | VUpdateWidthPix, |
| 942 | VReadyOffsetPix); |
| 943 | |
| 944 | LineTime = myPipe->HTotal / myPipe->PixelClock; |
| 945 | trip_to_mem = UrgentLatency; |
| 946 | Tvm_trips = UrgentExtraLatency + trip_to_mem * (GPUVMPageTableLevels * (HostVMDynamicLevelsTrips + 1) - 1); |
| 947 | |
| 948 | #ifdef __DML_VBA_ALLOW_DELTA__ |
| 949 | if (DynamicMetadataVMEnabled == true1 && GPUVMEnable == true1) { |
| 950 | #else |
| 951 | if (DynamicMetadataVMEnabled == true1) { |
| 952 | #endif |
| 953 | *Tdmdl = TWait + Tvm_trips + trip_to_mem; |
| 954 | } else { |
| 955 | *Tdmdl = TWait + UrgentExtraLatency; |
| 956 | } |
| 957 | |
| 958 | #ifdef __DML_VBA_ALLOW_DELTA__ |
| 959 | if (DynamicMetadataEnable == false0) { |
| 960 | *Tdmdl = 0.0; |
| 961 | } |
| 962 | #endif |
| 963 | |
| 964 | if (DynamicMetadataEnable == true1) { |
| 965 | if (VStartup * LineTime < *TSetup + *Tdmdl + Tdmbf + Tdmec + Tdmsks) { |
| 966 | *NotEnoughTimeForDynamicMetadata = true1; |
| 967 | dml_print("DML::%s: Not Enough Time for Dynamic Meta!\n", __func__){do { } while(0); }; |
| 968 | dml_print("DML::%s: Tdmbf: %fus - time for dmd transfer from dchub to dio output buffer\n", __func__, Tdmbf){do { } while(0); }; |
| 969 | dml_print("DML::%s: Tdmec: %fus - time dio takes to transfer dmd\n", __func__, Tdmec){do { } while(0); }; |
| 970 | dml_print("DML::%s: Tdmsks: %fus - time before active dmd must complete transmission at dio\n", __func__, Tdmsks){do { } while(0); }; |
| 971 | dml_print("DML::%s: Tdmdl: %fus - time for fabric to become ready and fetch dmd \n", __func__, *Tdmdl){do { } while(0); }; |
| 972 | } else { |
| 973 | *NotEnoughTimeForDynamicMetadata = false0; |
| 974 | } |
| 975 | } else { |
| 976 | *NotEnoughTimeForDynamicMetadata = false0; |
| 977 | } |
| 978 | |
| 979 | *Tdmdl_vm = (DynamicMetadataEnable == true1 && DynamicMetadataVMEnabled == true1 && GPUVMEnable == true1 ? TWait + Tvm_trips : 0); |
| 980 | |
| 981 | if (myPipe->ScalerEnabled) |
| 982 | DPPCycles = DPPCLKDelaySubtotalPlusCNVCFormater + DPPCLKDelaySCL; |
| 983 | else |
| 984 | DPPCycles = DPPCLKDelaySubtotalPlusCNVCFormater + DPPCLKDelaySCLLBOnly; |
| 985 | |
| 986 | DPPCycles = DPPCycles + myPipe->NumberOfCursors * DPPCLKDelayCNVCCursor; |
| 987 | |
| 988 | DISPCLKCycles = DISPCLKDelaySubtotal; |
| 989 | |
| 990 | if (myPipe->DPPCLK == 0.0 || myPipe->DISPCLK == 0.0) |
| 991 | return true1; |
| 992 | |
| 993 | *DSTXAfterScaler = DPPCycles * myPipe->PixelClock / myPipe->DPPCLK + DISPCLKCycles * myPipe->PixelClock / myPipe->DISPCLK + DSCDelay; |
| 994 | |
| 995 | #ifdef __DML_VBA_DEBUG__ |
| 996 | dml_print("DML::%s: DPPCycles: %d\n", __func__, DPPCycles){do { } while(0); }; |
| 997 | dml_print("DML::%s: PixelClock: %f\n", __func__, myPipe->PixelClock){do { } while(0); }; |
| 998 | dml_print("DML::%s: DPPCLK: %f\n", __func__, myPipe->DPPCLK){do { } while(0); }; |
| 999 | dml_print("DML::%s: DISPCLKCycles: %d\n", __func__, DISPCLKCycles){do { } while(0); }; |
| 1000 | dml_print("DML::%s: DISPCLK: %f\n", __func__, myPipe->DISPCLK){do { } while(0); }; |
| 1001 | dml_print("DML::%s: DSCDelay: %d\n", __func__, DSCDelay){do { } while(0); }; |
| 1002 | dml_print("DML::%s: DSTXAfterScaler: %d\n", __func__, *DSTXAfterScaler){do { } while(0); }; |
| 1003 | dml_print("DML::%s: ODMCombineIsEnabled: %d\n", __func__, myPipe->ODMCombineIsEnabled){do { } while(0); }; |
| 1004 | #endif |
| 1005 | |
| 1006 | *DSTXAfterScaler = *DSTXAfterScaler + ((myPipe->ODMCombineIsEnabled) ? 18 : 0) + (myPipe->DPPPerPlane - 1) * DPP_RECOUT_WIDTH; |
| 1007 | |
| 1008 | if (OutputFormat == dm_420 || (myPipe->InterlaceEnable && myPipe->ProgressiveToInterlaceUnitInOPP)) |
| 1009 | *DSTYAfterScaler = 1; |
| 1010 | else |
| 1011 | *DSTYAfterScaler = 0; |
| 1012 | |
| 1013 | DSTTotalPixelsAfterScaler = *DSTYAfterScaler * myPipe->HTotal + *DSTXAfterScaler; |
| 1014 | *DSTYAfterScaler = dml_floor(DSTTotalPixelsAfterScaler / myPipe->HTotal, 1); |
| 1015 | *DSTXAfterScaler = DSTTotalPixelsAfterScaler - ((double) (*DSTYAfterScaler * myPipe->HTotal)); |
| 1016 | |
| 1017 | #ifdef __DML_VBA_DEBUG__ |
| 1018 | dml_print("DML::%s: DSTXAfterScaler: %d (final)\n", __func__, *DSTXAfterScaler){do { } while(0); }; |
| 1019 | #endif |
| 1020 | |
| 1021 | MyError = false0; |
| 1022 | |
| 1023 | Tr0_trips = trip_to_mem * (HostVMDynamicLevelsTrips + 1); |
| 1024 | Tvm_trips_rounded = dml_ceil(4.0 * Tvm_trips / LineTime, 1) / 4 * LineTime; |
| 1025 | Tr0_trips_rounded = dml_ceil(4.0 * Tr0_trips / LineTime, 1) / 4 * LineTime; |
| 1026 | |
| 1027 | #ifdef __DML_VBA_ALLOW_DELTA__ |
| 1028 | if (!myPipe->DCCEnable) { |
| 1029 | Tr0_trips = 0.0; |
| 1030 | Tr0_trips_rounded = 0.0; |
| 1031 | } |
| 1032 | #endif |
| 1033 | |
| 1034 | if (!GPUVMEnable) { |
| 1035 | Tvm_trips = 0.0; |
| 1036 | Tvm_trips_rounded = 0.0; |
| 1037 | } |
| 1038 | |
| 1039 | if (GPUVMEnable) { |
| 1040 | if (GPUVMPageTableLevels >= 3) { |
| 1041 | *Tno_bw = UrgentExtraLatency + trip_to_mem * ((GPUVMPageTableLevels - 2) - 1); |
| 1042 | } else { |
| 1043 | *Tno_bw = 0; |
| 1044 | } |
| 1045 | } else if (!myPipe->DCCEnable) { |
| 1046 | *Tno_bw = LineTime; |
| 1047 | } else { |
| 1048 | *Tno_bw = LineTime / 4; |
| 1049 | } |
| 1050 | |
| 1051 | if (myPipe->SourcePixelFormat == dm_420_8 || myPipe->SourcePixelFormat == dm_420_10 || myPipe->SourcePixelFormat == dm_420_12) |
| 1052 | bytes_pp = myPipe->BytePerPixelY + myPipe->BytePerPixelC / 4; |
| 1053 | else |
| 1054 | bytes_pp = myPipe->BytePerPixelY + myPipe->BytePerPixelC; |
| 1055 | /*rev 99*/ |
| 1056 | prefetch_bw_pr = dml_min(1, bytes_pp * myPipe->PixelClock / (double) myPipe->DPPPerPlane); |
| 1057 | max_Tsw = dml_max(PrefetchSourceLinesY, PrefetchSourceLinesC) * LineTime; |
| 1058 | prefetch_sw_bytes = PrefetchSourceLinesY * swath_width_luma_ub * myPipe->BytePerPixelY + PrefetchSourceLinesC * swath_width_chroma_ub * myPipe->BytePerPixelC; |
| 1059 | prefetch_bw_oto = dml_max(bytes_pp * myPipe->PixelClock / myPipe->DPPPerPlane, prefetch_sw_bytes / (dml_max(PrefetchSourceLinesY, PrefetchSourceLinesC) * LineTime)); |
| 1060 | prefetch_bw_oto = dml_max(prefetch_bw_pr, prefetch_sw_bytes / max_Tsw); |
| 1061 | |
| 1062 | min_Lsw = dml_max(1, dml_max(PrefetchSourceLinesY, PrefetchSourceLinesC) / max_vratio_pre); |
| 1063 | Lsw_oto = dml_ceil(4 * dml_max(prefetch_sw_bytes / prefetch_bw_oto / LineTime, min_Lsw), 1) / 4; |
| 1064 | Tsw_oto = Lsw_oto * LineTime; |
| 1065 | |
| 1066 | prefetch_bw_oto = (PrefetchSourceLinesY * swath_width_luma_ub * myPipe->BytePerPixelY + PrefetchSourceLinesC * swath_width_chroma_ub * myPipe->BytePerPixelC) / Tsw_oto; |
| 1067 | |
| 1068 | #ifdef __DML_VBA_DEBUG__ |
| 1069 | dml_print("DML: HTotal: %d\n", myPipe->HTotal){do { } while(0); }; |
| 1070 | dml_print("DML: prefetch_bw_oto: %f\n", prefetch_bw_oto){do { } while(0); }; |
| 1071 | dml_print("DML: PrefetchSourceLinesY: %f\n", PrefetchSourceLinesY){do { } while(0); }; |
| 1072 | dml_print("DML: swath_width_luma_ub: %d\n", swath_width_luma_ub){do { } while(0); }; |
| 1073 | dml_print("DML: BytePerPixelY: %d\n", myPipe->BytePerPixelY){do { } while(0); }; |
| 1074 | dml_print("DML: Tsw_oto: %f\n", Tsw_oto){do { } while(0); }; |
| 1075 | #endif |
| 1076 | |
| 1077 | if (GPUVMEnable == true1) |
| 1078 | Tvm_oto = dml_max3(*Tno_bw + PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor / prefetch_bw_oto, Tvm_trips, LineTime / 4.0); |
| 1079 | else |
| 1080 | Tvm_oto = LineTime / 4.0; |
| 1081 | |
| 1082 | if ((GPUVMEnable == true1 || myPipe->DCCEnable == true1)) { |
| 1083 | Tr0_oto = dml_max4((MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor) / prefetch_bw_oto, Tr0_trips, // PREVIOUS_ERROR (missing this term) |
| 1084 | LineTime - Tvm_oto, |
| 1085 | LineTime / 4); |
| 1086 | } else { |
| 1087 | Tr0_oto = (LineTime - Tvm_oto) / 2.0; |
| 1088 | } |
| 1089 | |
| 1090 | #ifdef __DML_VBA_DEBUG__ |
| 1091 | dml_print("DML::%s: Tvm_trips = %f\n", __func__, Tvm_trips){do { } while(0); }; |
| 1092 | dml_print("DML::%s: Tr0_trips = %f\n", __func__, Tr0_trips){do { } while(0); }; |
| 1093 | dml_print("DML::%s: PDEAndMetaPTEBytesFrame = %d\n", __func__, MetaRowByte){do { } while(0); }; |
| 1094 | dml_print("DML::%s: MetaRowByte = %d\n", __func__, MetaRowByte){do { } while(0); }; |
| 1095 | dml_print("DML::%s: PixelPTEBytesPerRow = %d\n", __func__, PixelPTEBytesPerRow){do { } while(0); }; |
| 1096 | dml_print("DML::%s: HostVMInefficiencyFactor = %f\n", __func__, HostVMInefficiencyFactor){do { } while(0); }; |
| 1097 | dml_print("DML::%s: prefetch_bw_oto = %f\n", __func__, prefetch_bw_oto){do { } while(0); }; |
| 1098 | dml_print("DML::%s: Tr0_oto = %f\n", __func__, Tr0_oto){do { } while(0); }; |
| 1099 | dml_print("DML::%s: Tvm_oto = %f\n", __func__, Tvm_oto){do { } while(0); }; |
| 1100 | #endif |
| 1101 | |
| 1102 | Tvm_oto_lines = dml_ceil(4.0 * Tvm_oto / LineTime, 1) / 4.0; |
| 1103 | Tr0_oto_lines = dml_ceil(4.0 * Tr0_oto / LineTime, 1) / 4.0; |
| 1104 | dst_y_prefetch_oto = Tvm_oto_lines + 2 * Tr0_oto_lines + Lsw_oto; |
| 1105 | dst_y_prefetch_equ = VStartup - (*TSetup + dml_max(TWait + TCalc, *Tdmdl)) / LineTime - (*DSTYAfterScaler + *DSTXAfterScaler / myPipe->HTotal); |
| 1106 | dst_y_prefetch_equ = dml_floor(4.0 * (dst_y_prefetch_equ + 0.125), 1) / 4.0; |
| 1107 | Tpre_rounded = dst_y_prefetch_equ * LineTime; |
| 1108 | |
| 1109 | dep_bytes = dml_max(PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor, MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor); |
| 1110 | |
| 1111 | if (prefetch_sw_bytes < dep_bytes) |
| 1112 | prefetch_sw_bytes = 2 * dep_bytes; |
| 1113 | |
| 1114 | dml_print("DML: dst_y_prefetch_oto: %f\n", dst_y_prefetch_oto){do { } while(0); }; |
| 1115 | dml_print("DML: Tvm_oto_lines: %f\n", Tvm_oto_lines){do { } while(0); }; |
| 1116 | dml_print("DML: Tr0_oto_lines: %f\n", Tr0_oto_lines){do { } while(0); }; |
| 1117 | dml_print("DML: Lsw_oto: %f\n", Lsw_oto){do { } while(0); }; |
| 1118 | dml_print("DML: LineTime: %f\n", LineTime){do { } while(0); }; |
| 1119 | dml_print("DML: dst_y_prefetch_equ: %f (after round)\n", dst_y_prefetch_equ){do { } while(0); }; |
| 1120 | |
| 1121 | dml_print("DML: LineTime: %f\n", LineTime){do { } while(0); }; |
| 1122 | dml_print("DML: VStartup: %d\n", VStartup){do { } while(0); }; |
| 1123 | dml_print("DML: Tvstartup: %fus - time between vstartup and first pixel of active\n", VStartup * LineTime){do { } while(0); }; |
| 1124 | dml_print("DML: TSetup: %fus - time from vstartup to vready\n", *TSetup){do { } while(0); }; |
| 1125 | dml_print("DML: TCalc: %fus - time for calculations in dchub starting at vready\n", TCalc){do { } while(0); }; |
| 1126 | dml_print("DML: TWait: %fus - time for fabric to become ready max(pstate exit,cstate enter/exit, urgent latency) after TCalc\n", TWait){do { } while(0); }; |
| 1127 | dml_print("DML: Tdmbf: %fus - time for dmd transfer from dchub to dio output buffer\n", Tdmbf){do { } while(0); }; |
| 1128 | dml_print("DML: Tdmec: %fus - time dio takes to transfer dmd\n", Tdmec){do { } while(0); }; |
| 1129 | dml_print("DML: Tdmsks: %fus - time before active dmd must complete transmission at dio\n", Tdmsks){do { } while(0); }; |
| 1130 | dml_print("DML: Tdmdl_vm: %fus - time for vm stages of dmd \n", *Tdmdl_vm){do { } while(0); }; |
| 1131 | dml_print("DML: Tdmdl: %fus - time for fabric to become ready and fetch dmd \n", *Tdmdl){do { } while(0); }; |
| 1132 | dml_print("DML: DSTXAfterScaler: %f pixels - number of pixel clocks pipeline and buffer delay after scaler \n", *DSTXAfterScaler){do { } while(0); }; |
| 1133 | dml_print("DML: DSTYAfterScaler: %f lines - number of lines of pipeline and buffer delay after scaler \n", *DSTYAfterScaler){do { } while(0); }; |
| 1134 | |
| 1135 | *PrefetchBandwidth = 0; |
| 1136 | *DestinationLinesToRequestVMInVBlank = 0; |
| 1137 | *DestinationLinesToRequestRowInVBlank = 0; |
| 1138 | *VRatioPrefetchY = 0; |
| 1139 | *VRatioPrefetchC = 0; |
| 1140 | *RequiredPrefetchPixDataBWLuma = 0; |
| 1141 | if (dst_y_prefetch_equ > 1) { |
| 1142 | double PrefetchBandwidth1; |
| 1143 | double PrefetchBandwidth2; |
| 1144 | double PrefetchBandwidth3; |
| 1145 | double PrefetchBandwidth4; |
| 1146 | |
| 1147 | if (Tpre_rounded - *Tno_bw > 0) { |
| 1148 | PrefetchBandwidth1 = (PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor + 2 * MetaRowByte + 2 * PixelPTEBytesPerRow * HostVMInefficiencyFactor |
| 1149 | + prefetch_sw_bytes) / (Tpre_rounded - *Tno_bw); |
| 1150 | Tsw_est1 = prefetch_sw_bytes / PrefetchBandwidth1; |
| 1151 | } else { |
| 1152 | PrefetchBandwidth1 = 0; |
| 1153 | } |
| 1154 | |
| 1155 | if (VStartup == MaxVStartup && Tsw_est1 / LineTime < min_Lsw && Tpre_rounded - min_Lsw * LineTime - 0.75 * LineTime - *Tno_bw > 0) { |
| 1156 | PrefetchBandwidth1 = (PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor + 2 * MetaRowByte + 2 * PixelPTEBytesPerRow * HostVMInefficiencyFactor) |
| 1157 | / (Tpre_rounded - min_Lsw * LineTime - 0.75 * LineTime - *Tno_bw); |
| 1158 | } |
| 1159 | |
| 1160 | if (Tpre_rounded - *Tno_bw - 2 * Tr0_trips_rounded > 0) |
| 1161 | PrefetchBandwidth2 = (PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor + prefetch_sw_bytes) / (Tpre_rounded - *Tno_bw - 2 * Tr0_trips_rounded); |
| 1162 | else |
| 1163 | PrefetchBandwidth2 = 0; |
| 1164 | |
| 1165 | if (Tpre_rounded - Tvm_trips_rounded > 0) { |
| 1166 | PrefetchBandwidth3 = (2 * MetaRowByte + 2 * PixelPTEBytesPerRow * HostVMInefficiencyFactor |
| 1167 | + prefetch_sw_bytes) / (Tpre_rounded - Tvm_trips_rounded); |
| 1168 | Tsw_est3 = prefetch_sw_bytes / PrefetchBandwidth3; |
| 1169 | } else { |
| 1170 | PrefetchBandwidth3 = 0; |
| 1171 | } |
| 1172 | |
| 1173 | #ifdef __DML_VBA_DEBUG__ |
| 1174 | dml_print("DML::%s: Tpre_rounded: %f\n", __func__, Tpre_rounded){do { } while(0); }; |
| 1175 | dml_print("DML::%s: Tvm_trips_rounded: %f\n", __func__, Tvm_trips_rounded){do { } while(0); }; |
| 1176 | dml_print("DML::%s: PrefetchBandwidth3: %f\n", __func__, PrefetchBandwidth3){do { } while(0); }; |
| 1177 | #endif |
| 1178 | if (VStartup == MaxVStartup && Tsw_est3 / LineTime < min_Lsw && Tpre_rounded - min_Lsw * LineTime - 0.75 * LineTime - Tvm_trips_rounded > 0) { |
| 1179 | PrefetchBandwidth3 = (2 * MetaRowByte + 2 * PixelPTEBytesPerRow * HostVMInefficiencyFactor) |
| 1180 | / (Tpre_rounded - min_Lsw * LineTime - 0.75 * LineTime - Tvm_trips_rounded); |
| 1181 | } |
| 1182 | |
| 1183 | if (Tpre_rounded - Tvm_trips_rounded - 2 * Tr0_trips_rounded > 0) |
| 1184 | PrefetchBandwidth4 = prefetch_sw_bytes / (Tpre_rounded - Tvm_trips_rounded - 2 * Tr0_trips_rounded); |
| 1185 | else |
| 1186 | PrefetchBandwidth4 = 0; |
| 1187 | |
| 1188 | { |
| 1189 | bool_Bool Case1OK; |
| 1190 | bool_Bool Case2OK; |
| 1191 | bool_Bool Case3OK; |
| 1192 | |
| 1193 | if (PrefetchBandwidth1 > 0) { |
| 1194 | if (*Tno_bw + PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor / PrefetchBandwidth1 >= Tvm_trips_rounded |
| 1195 | && (MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor) / PrefetchBandwidth1 >= Tr0_trips_rounded) { |
| 1196 | Case1OK = true1; |
| 1197 | } else { |
| 1198 | Case1OK = false0; |
| 1199 | } |
| 1200 | } else { |
| 1201 | Case1OK = false0; |
| 1202 | } |
| 1203 | |
| 1204 | if (PrefetchBandwidth2 > 0) { |
| 1205 | if (*Tno_bw + PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor / PrefetchBandwidth2 >= Tvm_trips_rounded |
| 1206 | && (MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor) / PrefetchBandwidth2 < Tr0_trips_rounded) { |
| 1207 | Case2OK = true1; |
| 1208 | } else { |
| 1209 | Case2OK = false0; |
| 1210 | } |
| 1211 | } else { |
| 1212 | Case2OK = false0; |
| 1213 | } |
| 1214 | |
| 1215 | if (PrefetchBandwidth3 > 0) { |
| 1216 | if (*Tno_bw + PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor / PrefetchBandwidth3 < Tvm_trips_rounded |
| 1217 | && (MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor) / PrefetchBandwidth3 >= Tr0_trips_rounded) { |
| 1218 | Case3OK = true1; |
| 1219 | } else { |
| 1220 | Case3OK = false0; |
| 1221 | } |
| 1222 | } else { |
| 1223 | Case3OK = false0; |
| 1224 | } |
| 1225 | |
| 1226 | if (Case1OK) { |
| 1227 | prefetch_bw_equ = PrefetchBandwidth1; |
| 1228 | } else if (Case2OK) { |
| 1229 | prefetch_bw_equ = PrefetchBandwidth2; |
| 1230 | } else if (Case3OK) { |
| 1231 | prefetch_bw_equ = PrefetchBandwidth3; |
| 1232 | } else { |
| 1233 | prefetch_bw_equ = PrefetchBandwidth4; |
| 1234 | } |
| 1235 | |
| 1236 | #ifdef __DML_VBA_DEBUG__ |
| 1237 | dml_print("DML::%s: Case1OK: %d\n", __func__, Case1OK){do { } while(0); }; |
| 1238 | dml_print("DML::%s: Case2OK: %d\n", __func__, Case2OK){do { } while(0); }; |
| 1239 | dml_print("DML::%s: Case3OK: %d\n", __func__, Case3OK){do { } while(0); }; |
| 1240 | dml_print("DML::%s: prefetch_bw_equ: %f\n", __func__, prefetch_bw_equ){do { } while(0); }; |
| 1241 | #endif |
| 1242 | |
| 1243 | if (prefetch_bw_equ > 0) { |
| 1244 | if (GPUVMEnable == true1) { |
| 1245 | Tvm_equ = dml_max3(*Tno_bw + PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor / prefetch_bw_equ, Tvm_trips, LineTime / 4); |
| 1246 | } else { |
| 1247 | Tvm_equ = LineTime / 4; |
| 1248 | } |
| 1249 | |
| 1250 | if ((GPUVMEnable == true1 || myPipe->DCCEnable == true1)) { |
| 1251 | Tr0_equ = dml_max4( |
| 1252 | (MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor) / prefetch_bw_equ, |
| 1253 | Tr0_trips, |
| 1254 | (LineTime - Tvm_equ) / 2, |
| 1255 | LineTime / 4); |
| 1256 | } else { |
| 1257 | Tr0_equ = (LineTime - Tvm_equ) / 2; |
| 1258 | } |
| 1259 | } else { |
| 1260 | Tvm_equ = 0; |
| 1261 | Tr0_equ = 0; |
| 1262 | dml_print("DML: prefetch_bw_equ equals 0! %s:%d\n", __FILE__, __LINE__){do { } while(0); }; |
| 1263 | } |
| 1264 | } |
| 1265 | |
| 1266 | if (dst_y_prefetch_oto < dst_y_prefetch_equ) { |
| 1267 | *DestinationLinesForPrefetch = dst_y_prefetch_oto; |
| 1268 | TimeForFetchingMetaPTE = Tvm_oto; |
| 1269 | TimeForFetchingRowInVBlank = Tr0_oto; |
| 1270 | *PrefetchBandwidth = prefetch_bw_oto; |
| 1271 | } else { |
| 1272 | *DestinationLinesForPrefetch = dst_y_prefetch_equ; |
| 1273 | TimeForFetchingMetaPTE = Tvm_equ; |
| 1274 | TimeForFetchingRowInVBlank = Tr0_equ; |
| 1275 | *PrefetchBandwidth = prefetch_bw_equ; |
| 1276 | } |
| 1277 | |
| 1278 | *DestinationLinesToRequestVMInVBlank = dml_ceil(4.0 * TimeForFetchingMetaPTE / LineTime, 1.0) / 4.0; |
| 1279 | |
| 1280 | *DestinationLinesToRequestRowInVBlank = dml_ceil(4.0 * TimeForFetchingRowInVBlank / LineTime, 1.0) / 4.0; |
| 1281 | |
| 1282 | #ifdef __DML_VBA_ALLOW_DELTA__ |
| 1283 | LinesToRequestPrefetchPixelData = *DestinationLinesForPrefetch |
| 1284 | // See note above dated 5/30/2018 |
| 1285 | // - ((NumberOfCursors > 0 || GPUVMEnable || DCCEnable) ? |
| 1286 | - ((GPUVMEnable || myPipe->DCCEnable) ? (*DestinationLinesToRequestVMInVBlank + 2 * *DestinationLinesToRequestRowInVBlank) : 0.0); // TODO: Did someone else add this?? |
| 1287 | #else |
| 1288 | LinesToRequestPrefetchPixelData = *DestinationLinesForPrefetch - *DestinationLinesToRequestVMInVBlank - 2 * *DestinationLinesToRequestRowInVBlank; |
| 1289 | #endif |
| 1290 | |
| 1291 | #ifdef __DML_VBA_DEBUG__ |
| 1292 | dml_print("DML::%s: DestinationLinesForPrefetch = %f\n", __func__, *DestinationLinesForPrefetch){do { } while(0); }; |
| 1293 | dml_print("DML::%s: DestinationLinesToRequestVMInVBlank = %f\n", __func__, *DestinationLinesToRequestVMInVBlank){do { } while(0); }; |
| 1294 | dml_print("DML::%s: TimeForFetchingRowInVBlank = %f\n", __func__, TimeForFetchingRowInVBlank){do { } while(0); }; |
| 1295 | dml_print("DML::%s: LineTime = %f\n", __func__, LineTime){do { } while(0); }; |
| 1296 | dml_print("DML::%s: DestinationLinesToRequestRowInVBlank = %f\n", __func__, *DestinationLinesToRequestRowInVBlank){do { } while(0); }; |
| 1297 | dml_print("DML::%s: PrefetchSourceLinesY = %f\n", __func__, PrefetchSourceLinesY){do { } while(0); }; |
| 1298 | dml_print("DML::%s: LinesToRequestPrefetchPixelData = %f\n", __func__, LinesToRequestPrefetchPixelData){do { } while(0); }; |
| 1299 | #endif |
| 1300 | |
| 1301 | if (LinesToRequestPrefetchPixelData > 0 && prefetch_bw_equ > 0) { |
| 1302 | |
| 1303 | *VRatioPrefetchY = (double) PrefetchSourceLinesY / LinesToRequestPrefetchPixelData; |
| 1304 | *VRatioPrefetchY = dml_max(*VRatioPrefetchY, 1.0); |
| 1305 | #ifdef __DML_VBA_DEBUG__ |
| 1306 | dml_print("DML::%s: VRatioPrefetchY = %f\n", __func__, *VRatioPrefetchY){do { } while(0); }; |
| 1307 | dml_print("DML::%s: SwathHeightY = %d\n", __func__, SwathHeightY){do { } while(0); }; |
| 1308 | dml_print("DML::%s: VInitPreFillY = %f\n", __func__, VInitPreFillY){do { } while(0); }; |
| 1309 | #endif |
| 1310 | if ((SwathHeightY > 4) && (VInitPreFillY > 3)) { |
| 1311 | if (LinesToRequestPrefetchPixelData > (VInitPreFillY - 3.0) / 2.0) { |
| 1312 | *VRatioPrefetchY = dml_max( |
| 1313 | (double) PrefetchSourceLinesY / LinesToRequestPrefetchPixelData, |
| 1314 | (double) MaxNumSwathY * SwathHeightY / (LinesToRequestPrefetchPixelData - (VInitPreFillY - 3.0) / 2.0)); |
| 1315 | *VRatioPrefetchY = dml_max(*VRatioPrefetchY, 1.0); |
| 1316 | } else { |
| 1317 | MyError = true1; |
| 1318 | dml_print("DML: MyErr set %s:%d\n", __FILE__, __LINE__){do { } while(0); }; |
| 1319 | *VRatioPrefetchY = 0; |
| 1320 | } |
| 1321 | #ifdef __DML_VBA_DEBUG__ |
| 1322 | dml_print("DML::%s: VRatioPrefetchY = %f\n", __func__, *VRatioPrefetchY){do { } while(0); }; |
| 1323 | dml_print("DML::%s: PrefetchSourceLinesY = %f\n", __func__, PrefetchSourceLinesY){do { } while(0); }; |
| 1324 | dml_print("DML::%s: MaxNumSwathY = %d\n", __func__, MaxNumSwathY){do { } while(0); }; |
| 1325 | #endif |
| 1326 | } |
| 1327 | |
| 1328 | *VRatioPrefetchC = (double) PrefetchSourceLinesC / LinesToRequestPrefetchPixelData; |
| 1329 | *VRatioPrefetchC = dml_max(*VRatioPrefetchC, 1.0); |
| 1330 | |
| 1331 | #ifdef __DML_VBA_DEBUG__ |
| 1332 | dml_print("DML::%s: VRatioPrefetchC = %f\n", __func__, *VRatioPrefetchC){do { } while(0); }; |
| 1333 | dml_print("DML::%s: SwathHeightC = %d\n", __func__, SwathHeightC){do { } while(0); }; |
| 1334 | dml_print("DML::%s: VInitPreFillC = %f\n", __func__, VInitPreFillC){do { } while(0); }; |
| 1335 | #endif |
| 1336 | if ((SwathHeightC > 4) || VInitPreFillC > 3) { |
| 1337 | if (LinesToRequestPrefetchPixelData > (VInitPreFillC - 3.0) / 2.0) { |
| 1338 | *VRatioPrefetchC = dml_max( |
| 1339 | *VRatioPrefetchC, |
| 1340 | (double) MaxNumSwathC * SwathHeightC / (LinesToRequestPrefetchPixelData - (VInitPreFillC - 3.0) / 2.0)); |
| 1341 | *VRatioPrefetchC = dml_max(*VRatioPrefetchC, 1.0); |
| 1342 | } else { |
| 1343 | MyError = true1; |
| 1344 | dml_print("DML: MyErr set %s:%d\n", __FILE__, __LINE__){do { } while(0); }; |
| 1345 | *VRatioPrefetchC = 0; |
| 1346 | } |
| 1347 | #ifdef __DML_VBA_DEBUG__ |
| 1348 | dml_print("DML::%s: VRatioPrefetchC = %f\n", __func__, *VRatioPrefetchC){do { } while(0); }; |
| 1349 | dml_print("DML::%s: PrefetchSourceLinesC = %f\n", __func__, PrefetchSourceLinesC){do { } while(0); }; |
| 1350 | dml_print("DML::%s: MaxNumSwathC = %d\n", __func__, MaxNumSwathC){do { } while(0); }; |
| 1351 | #endif |
| 1352 | } |
| 1353 | |
| 1354 | #ifdef __DML_VBA_DEBUG__ |
| 1355 | dml_print("DML::%s: BytePerPixelY = %d\n", __func__, myPipe->BytePerPixelY){do { } while(0); }; |
| 1356 | dml_print("DML::%s: swath_width_luma_ub = %d\n", __func__, swath_width_luma_ub){do { } while(0); }; |
| 1357 | dml_print("DML::%s: LineTime = %f\n", __func__, LineTime){do { } while(0); }; |
| 1358 | #endif |
| 1359 | |
| 1360 | *RequiredPrefetchPixDataBWLuma = (double) PrefetchSourceLinesY / LinesToRequestPrefetchPixelData * myPipe->BytePerPixelY * swath_width_luma_ub / LineTime; |
| 1361 | |
| 1362 | #ifdef __DML_VBA_DEBUG__ |
| 1363 | dml_print("DML::%s: RequiredPrefetchPixDataBWLuma = %f\n", __func__, *RequiredPrefetchPixDataBWLuma){do { } while(0); }; |
| 1364 | #endif |
| 1365 | |
| 1366 | *RequiredPrefetchPixDataBWChroma = (double) PrefetchSourceLinesC / LinesToRequestPrefetchPixelData * myPipe->BytePerPixelC * swath_width_chroma_ub |
| 1367 | / LineTime; |
| 1368 | } else { |
| 1369 | MyError = true1; |
| 1370 | dml_print("DML: MyErr set %s:%d\n", __FILE__, __LINE__){do { } while(0); }; |
| 1371 | dml_print("DML: LinesToRequestPrefetchPixelData: %f, should be > 0\n", LinesToRequestPrefetchPixelData){do { } while(0); }; |
| 1372 | *VRatioPrefetchY = 0; |
| 1373 | *VRatioPrefetchC = 0; |
| 1374 | *RequiredPrefetchPixDataBWLuma = 0; |
| 1375 | *RequiredPrefetchPixDataBWChroma = 0; |
| 1376 | } |
| 1377 | |
| 1378 | dml_print({do { } while(0); } |
| 1379 | "DML: Tpre: %fus - sum of time to request meta pte, 2 x data pte + meta data, swaths\n",{do { } while(0); } |
| 1380 | (double) LinesToRequestPrefetchPixelData * LineTime + 2.0 * TimeForFetchingRowInVBlank + TimeForFetchingMetaPTE){do { } while(0); }; |
| 1381 | dml_print("DML: Tvm: %fus - time to fetch page tables for meta surface\n", TimeForFetchingMetaPTE){do { } while(0); }; |
| 1382 | dml_print("DML: Tr0: %fus - time to fetch first row of data pagetables and first row of meta data (done in parallel)\n", TimeForFetchingRowInVBlank){do { } while(0); }; |
| 1383 | dml_print({do { } while(0); } |
| 1384 | "DML: Tsw: %fus = time to fetch enough pixel data and cursor data to feed the scalers init position and detile\n",{do { } while(0); } |
| 1385 | (double) LinesToRequestPrefetchPixelData * LineTime){do { } while(0); }; |
| 1386 | dml_print("DML: To: %fus - time for propagation from scaler to optc\n",{do { } while(0); } |
| 1387 | (*DSTYAfterScaler + ((double) (*DSTXAfterScaler) /{do { } while(0); } |
| 1388 | (double) myPipe->HTotal)) * LineTime){do { } while(0); }; |
| 1389 | dml_print("DML: Tvstartup - TSetup - Tcalc - Twait - Tpre - To > 0\n"){do { } while(0); }; |
| 1390 | dml_print("DML: Tslack(pre): %fus - time left over in schedule\n",{do { } while(0); } |
| 1391 | VStartup * LineTime - TimeForFetchingMetaPTE - 2 * TimeForFetchingRowInVBlank{do { } while(0); } |
| 1392 | - (*DSTYAfterScaler + ((double) (*DSTXAfterScaler) / (double) myPipe->HTotal)) * LineTime - TWait - TCalc - *TSetup){do { } while(0); }; |
| 1393 | dml_print("DML: row_bytes = dpte_row_bytes (per_pipe) = PixelPTEBytesPerRow = : %d\n", PixelPTEBytesPerRow){do { } while(0); }; |
| 1394 | |
| 1395 | } else { |
| 1396 | MyError = true1; |
| 1397 | dml_print("DML: MyErr set %s:%d\n", __FILE__, __LINE__){do { } while(0); }; |
| 1398 | } |
| 1399 | |
| 1400 | { |
| 1401 | double prefetch_vm_bw; |
| 1402 | double prefetch_row_bw; |
| 1403 | |
| 1404 | if (PDEAndMetaPTEBytesFrame == 0) { |
| 1405 | prefetch_vm_bw = 0; |
| 1406 | } else if (*DestinationLinesToRequestVMInVBlank > 0) { |
| 1407 | #ifdef __DML_VBA_DEBUG__ |
| 1408 | dml_print("DML::%s: PDEAndMetaPTEBytesFrame = %d\n", __func__, PDEAndMetaPTEBytesFrame){do { } while(0); }; |
| 1409 | dml_print("DML::%s: HostVMInefficiencyFactor = %f\n", __func__, HostVMInefficiencyFactor){do { } while(0); }; |
| 1410 | dml_print("DML::%s: DestinationLinesToRequestVMInVBlank = %f\n", __func__, *DestinationLinesToRequestVMInVBlank){do { } while(0); }; |
| 1411 | dml_print("DML::%s: LineTime = %f\n", __func__, LineTime){do { } while(0); }; |
| 1412 | #endif |
| 1413 | prefetch_vm_bw = PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor / (*DestinationLinesToRequestVMInVBlank * LineTime); |
| 1414 | #ifdef __DML_VBA_DEBUG__ |
| 1415 | dml_print("DML::%s: prefetch_vm_bw = %f\n", __func__, prefetch_vm_bw){do { } while(0); }; |
| 1416 | #endif |
| 1417 | } else { |
| 1418 | prefetch_vm_bw = 0; |
| 1419 | MyError = true1; |
| 1420 | dml_print("DML: MyErr set %s:%d\n", __FILE__, __LINE__){do { } while(0); }; |
| 1421 | } |
| 1422 | |
| 1423 | if (MetaRowByte + PixelPTEBytesPerRow == 0) { |
| 1424 | prefetch_row_bw = 0; |
| 1425 | } else if (*DestinationLinesToRequestRowInVBlank > 0) { |
| 1426 | prefetch_row_bw = (MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor) / (*DestinationLinesToRequestRowInVBlank * LineTime); |
| 1427 | |
| 1428 | #ifdef __DML_VBA_DEBUG__ |
| 1429 | dml_print("DML::%s: MetaRowByte = %d\n", __func__, MetaRowByte){do { } while(0); }; |
| 1430 | dml_print("DML::%s: PixelPTEBytesPerRow = %d\n", __func__, PixelPTEBytesPerRow){do { } while(0); }; |
| 1431 | dml_print("DML::%s: DestinationLinesToRequestRowInVBlank = %f\n", __func__, *DestinationLinesToRequestRowInVBlank){do { } while(0); }; |
| 1432 | dml_print("DML::%s: prefetch_row_bw = %f\n", __func__, prefetch_row_bw){do { } while(0); }; |
| 1433 | #endif |
| 1434 | } else { |
| 1435 | prefetch_row_bw = 0; |
| 1436 | MyError = true1; |
| 1437 | dml_print("DML: MyErr set %s:%d\n", __FILE__, __LINE__){do { } while(0); }; |
| 1438 | } |
| 1439 | |
| 1440 | *prefetch_vmrow_bw = dml_max(prefetch_vm_bw, prefetch_row_bw); |
| 1441 | } |
| 1442 | |
| 1443 | if (MyError) { |
| 1444 | *PrefetchBandwidth = 0; |
| 1445 | TimeForFetchingMetaPTE = 0; |
| 1446 | TimeForFetchingRowInVBlank = 0; |
| 1447 | *DestinationLinesToRequestVMInVBlank = 0; |
| 1448 | *DestinationLinesToRequestRowInVBlank = 0; |
| 1449 | *DestinationLinesForPrefetch = 0; |
| 1450 | LinesToRequestPrefetchPixelData = 0; |
| 1451 | *VRatioPrefetchY = 0; |
| 1452 | *VRatioPrefetchC = 0; |
| 1453 | *RequiredPrefetchPixDataBWLuma = 0; |
| 1454 | *RequiredPrefetchPixDataBWChroma = 0; |
| 1455 | } |
| 1456 | |
| 1457 | return MyError; |
| 1458 | } |
| 1459 | |
| 1460 | static double RoundToDFSGranularityUp(double Clock, double VCOSpeed) |
| 1461 | { |
| 1462 | return VCOSpeed * 4 / dml_floor(VCOSpeed * 4 / Clock, 1); |
| 1463 | } |
| 1464 | |
| 1465 | static double RoundToDFSGranularityDown(double Clock, double VCOSpeed) |
| 1466 | { |
| 1467 | return VCOSpeed * 4 / dml_ceil(VCOSpeed * 4.0 / Clock, 1); |
| 1468 | } |
| 1469 | |
| 1470 | static void CalculateDCCConfiguration( |
| 1471 | bool_Bool DCCEnabled, |
| 1472 | bool_Bool DCCProgrammingAssumesScanDirectionUnknown, |
| 1473 | enum source_format_class SourcePixelFormat, |
| 1474 | unsigned int SurfaceWidthLuma, |
| 1475 | unsigned int SurfaceWidthChroma, |
| 1476 | unsigned int SurfaceHeightLuma, |
| 1477 | unsigned int SurfaceHeightChroma, |
| 1478 | double DETBufferSize, |
| 1479 | unsigned int RequestHeight256ByteLuma, |
| 1480 | unsigned int RequestHeight256ByteChroma, |
| 1481 | enum dm_swizzle_mode TilingFormat, |
| 1482 | unsigned int BytePerPixelY, |
| 1483 | unsigned int BytePerPixelC, |
| 1484 | double BytePerPixelDETY, |
| 1485 | double BytePerPixelDETC, |
| 1486 | enum scan_direction_class ScanOrientation, |
| 1487 | unsigned int *MaxUncompressedBlockLuma, |
| 1488 | unsigned int *MaxUncompressedBlockChroma, |
| 1489 | unsigned int *MaxCompressedBlockLuma, |
| 1490 | unsigned int *MaxCompressedBlockChroma, |
| 1491 | unsigned int *IndependentBlockLuma, |
| 1492 | unsigned int *IndependentBlockChroma) |
| 1493 | { |
| 1494 | int yuv420; |
| 1495 | int horz_div_l; |
| 1496 | int horz_div_c; |
| 1497 | int vert_div_l; |
| 1498 | int vert_div_c; |
| 1499 | |
| 1500 | int swath_buf_size; |
| 1501 | double detile_buf_vp_horz_limit; |
| 1502 | double detile_buf_vp_vert_limit; |
| 1503 | |
| 1504 | int MAS_vp_horz_limit; |
| 1505 | int MAS_vp_vert_limit; |
| 1506 | int max_vp_horz_width; |
| 1507 | int max_vp_vert_height; |
| 1508 | int eff_surf_width_l; |
| 1509 | int eff_surf_width_c; |
| 1510 | int eff_surf_height_l; |
| 1511 | int eff_surf_height_c; |
| 1512 | |
| 1513 | int full_swath_bytes_horz_wc_l; |
| 1514 | int full_swath_bytes_horz_wc_c; |
| 1515 | int full_swath_bytes_vert_wc_l; |
| 1516 | int full_swath_bytes_vert_wc_c; |
| 1517 | int req128_horz_wc_l; |
| 1518 | int req128_horz_wc_c; |
| 1519 | int req128_vert_wc_l; |
| 1520 | int req128_vert_wc_c; |
| 1521 | int segment_order_horz_contiguous_luma; |
| 1522 | int segment_order_horz_contiguous_chroma; |
| 1523 | int segment_order_vert_contiguous_luma; |
| 1524 | int segment_order_vert_contiguous_chroma; |
| 1525 | |
| 1526 | typedef enum { |
| 1527 | REQ_256Bytes, REQ_128BytesNonContiguous, REQ_128BytesContiguous, REQ_NA |
| 1528 | } RequestType; |
| 1529 | RequestType RequestLuma; |
| 1530 | RequestType RequestChroma; |
| 1531 | |
| 1532 | yuv420 = ((SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10 || SourcePixelFormat == dm_420_12) ? 1 : 0); |
| 1533 | horz_div_l = 1; |
| 1534 | horz_div_c = 1; |
| 1535 | vert_div_l = 1; |
| 1536 | vert_div_c = 1; |
| 1537 | |
| 1538 | if (BytePerPixelY == 1) |
| 1539 | vert_div_l = 0; |
| 1540 | if (BytePerPixelC == 1) |
| 1541 | vert_div_c = 0; |
| 1542 | if (BytePerPixelY == 8 && (TilingFormat == dm_sw_64kb_s || TilingFormat == dm_sw_64kb_s_t || TilingFormat == dm_sw_64kb_s_x)) |
| 1543 | horz_div_l = 0; |
| 1544 | if (BytePerPixelC == 8 && (TilingFormat == dm_sw_64kb_s || TilingFormat == dm_sw_64kb_s_t || TilingFormat == dm_sw_64kb_s_x)) |
| 1545 | horz_div_c = 0; |
| 1546 | |
| 1547 | if (BytePerPixelC == 0) { |
| 1548 | swath_buf_size = DETBufferSize / 2 - 2 * 256; |
| 1549 | detile_buf_vp_horz_limit = (double) swath_buf_size / ((double) RequestHeight256ByteLuma * BytePerPixelY / (1 + horz_div_l)); |
| 1550 | detile_buf_vp_vert_limit = (double) swath_buf_size / (256.0 / RequestHeight256ByteLuma / (1 + vert_div_l)); |
| 1551 | } else { |
| 1552 | swath_buf_size = DETBufferSize / 2 - 2 * 2 * 256; |
| 1553 | detile_buf_vp_horz_limit = (double) swath_buf_size |
| 1554 | / ((double) RequestHeight256ByteLuma * BytePerPixelY / (1 + horz_div_l) |
| 1555 | + (double) RequestHeight256ByteChroma * BytePerPixelC / (1 + horz_div_c) / (1 + yuv420)); |
| 1556 | detile_buf_vp_vert_limit = (double) swath_buf_size |
| 1557 | / (256.0 / RequestHeight256ByteLuma / (1 + vert_div_l) + 256.0 / RequestHeight256ByteChroma / (1 + vert_div_c) / (1 + yuv420)); |
| 1558 | } |
| 1559 | |
| 1560 | if (SourcePixelFormat == dm_420_10) { |
| 1561 | detile_buf_vp_horz_limit = 1.5 * detile_buf_vp_horz_limit; |
| 1562 | detile_buf_vp_vert_limit = 1.5 * detile_buf_vp_vert_limit; |
| 1563 | } |
| 1564 | |
| 1565 | detile_buf_vp_horz_limit = dml_floor(detile_buf_vp_horz_limit - 1, 16); |
| 1566 | detile_buf_vp_vert_limit = dml_floor(detile_buf_vp_vert_limit - 1, 16); |
| 1567 | |
| 1568 | MAS_vp_horz_limit = SourcePixelFormat == dm_rgbe_alpha ? 3840 : 5760; |
| 1569 | MAS_vp_vert_limit = (BytePerPixelC > 0 ? 2880 : 5760); |
| 1570 | max_vp_horz_width = dml_min((double) MAS_vp_horz_limit, detile_buf_vp_horz_limit); |
| 1571 | max_vp_vert_height = dml_min((double) MAS_vp_vert_limit, detile_buf_vp_vert_limit); |
| 1572 | eff_surf_width_l = (SurfaceWidthLuma > max_vp_horz_width ? max_vp_horz_width : SurfaceWidthLuma); |
| 1573 | eff_surf_width_c = eff_surf_width_l / (1 + yuv420); |
| 1574 | eff_surf_height_l = (SurfaceHeightLuma > max_vp_vert_height ? max_vp_vert_height : SurfaceHeightLuma); |
| 1575 | eff_surf_height_c = eff_surf_height_l / (1 + yuv420); |
| 1576 | |
| 1577 | full_swath_bytes_horz_wc_l = eff_surf_width_l * RequestHeight256ByteLuma * BytePerPixelY; |
| 1578 | full_swath_bytes_vert_wc_l = eff_surf_height_l * 256 / RequestHeight256ByteLuma; |
| 1579 | if (BytePerPixelC > 0) { |
| 1580 | full_swath_bytes_horz_wc_c = eff_surf_width_c * RequestHeight256ByteChroma * BytePerPixelC; |
| 1581 | full_swath_bytes_vert_wc_c = eff_surf_height_c * 256 / RequestHeight256ByteChroma; |
| 1582 | } else { |
| 1583 | full_swath_bytes_horz_wc_c = 0; |
| 1584 | full_swath_bytes_vert_wc_c = 0; |
| 1585 | } |
| 1586 | |
| 1587 | if (SourcePixelFormat == dm_420_10) { |
| 1588 | full_swath_bytes_horz_wc_l = dml_ceil(full_swath_bytes_horz_wc_l * 2 / 3, 256); |
| 1589 | full_swath_bytes_horz_wc_c = dml_ceil(full_swath_bytes_horz_wc_c * 2 / 3, 256); |
| 1590 | full_swath_bytes_vert_wc_l = dml_ceil(full_swath_bytes_vert_wc_l * 2 / 3, 256); |
| 1591 | full_swath_bytes_vert_wc_c = dml_ceil(full_swath_bytes_vert_wc_c * 2 / 3, 256); |
| 1592 | } |
| 1593 | |
| 1594 | if (2 * full_swath_bytes_horz_wc_l + 2 * full_swath_bytes_horz_wc_c <= DETBufferSize) { |
| 1595 | req128_horz_wc_l = 0; |
| 1596 | req128_horz_wc_c = 0; |
| 1597 | } else if (full_swath_bytes_horz_wc_l < 1.5 * full_swath_bytes_horz_wc_c && 2 * full_swath_bytes_horz_wc_l + full_swath_bytes_horz_wc_c <= DETBufferSize) { |
| 1598 | req128_horz_wc_l = 0; |
| 1599 | req128_horz_wc_c = 1; |
| 1600 | } else if (full_swath_bytes_horz_wc_l >= 1.5 * full_swath_bytes_horz_wc_c && full_swath_bytes_horz_wc_l + 2 * full_swath_bytes_horz_wc_c <= DETBufferSize) { |
| 1601 | req128_horz_wc_l = 1; |
| 1602 | req128_horz_wc_c = 0; |
| 1603 | } else { |
| 1604 | req128_horz_wc_l = 1; |
| 1605 | req128_horz_wc_c = 1; |
| 1606 | } |
| 1607 | |
| 1608 | if (2 * full_swath_bytes_vert_wc_l + 2 * full_swath_bytes_vert_wc_c <= DETBufferSize) { |
| 1609 | req128_vert_wc_l = 0; |
| 1610 | req128_vert_wc_c = 0; |
| 1611 | } else if (full_swath_bytes_vert_wc_l < 1.5 * full_swath_bytes_vert_wc_c && 2 * full_swath_bytes_vert_wc_l + full_swath_bytes_vert_wc_c <= DETBufferSize) { |
| 1612 | req128_vert_wc_l = 0; |
| 1613 | req128_vert_wc_c = 1; |
| 1614 | } else if (full_swath_bytes_vert_wc_l >= 1.5 * full_swath_bytes_vert_wc_c && full_swath_bytes_vert_wc_l + 2 * full_swath_bytes_vert_wc_c <= DETBufferSize) { |
| 1615 | req128_vert_wc_l = 1; |
| 1616 | req128_vert_wc_c = 0; |
| 1617 | } else { |
| 1618 | req128_vert_wc_l = 1; |
| 1619 | req128_vert_wc_c = 1; |
| 1620 | } |
| 1621 | |
| 1622 | if (BytePerPixelY == 2 || (BytePerPixelY == 4 && TilingFormat != dm_sw_64kb_r_x)) { |
| 1623 | segment_order_horz_contiguous_luma = 0; |
| 1624 | } else { |
| 1625 | segment_order_horz_contiguous_luma = 1; |
| 1626 | } |
| 1627 | if ((BytePerPixelY == 8 && (TilingFormat == dm_sw_64kb_d || TilingFormat == dm_sw_64kb_d_x || TilingFormat == dm_sw_64kb_d_t || TilingFormat == dm_sw_64kb_r_x)) |
| 1628 | || (BytePerPixelY == 4 && TilingFormat == dm_sw_64kb_r_x)) { |
| 1629 | segment_order_vert_contiguous_luma = 0; |
| 1630 | } else { |
| 1631 | segment_order_vert_contiguous_luma = 1; |
| 1632 | } |
| 1633 | if (BytePerPixelC == 2 || (BytePerPixelC == 4 && TilingFormat != dm_sw_64kb_r_x)) { |
| 1634 | segment_order_horz_contiguous_chroma = 0; |
| 1635 | } else { |
| 1636 | segment_order_horz_contiguous_chroma = 1; |
| 1637 | } |
| 1638 | if ((BytePerPixelC == 8 && (TilingFormat == dm_sw_64kb_d || TilingFormat == dm_sw_64kb_d_x || TilingFormat == dm_sw_64kb_d_t || TilingFormat == dm_sw_64kb_r_x)) |
| 1639 | || (BytePerPixelC == 4 && TilingFormat == dm_sw_64kb_r_x)) { |
| 1640 | segment_order_vert_contiguous_chroma = 0; |
| 1641 | } else { |
| 1642 | segment_order_vert_contiguous_chroma = 1; |
| 1643 | } |
| 1644 | |
| 1645 | if (DCCProgrammingAssumesScanDirectionUnknown == true1) { |
| 1646 | if (req128_horz_wc_l == 0 && req128_vert_wc_l == 0) { |
| 1647 | RequestLuma = REQ_256Bytes; |
| 1648 | } else if ((req128_horz_wc_l == 1 && segment_order_horz_contiguous_luma == 0) || (req128_vert_wc_l == 1 && segment_order_vert_contiguous_luma == 0)) { |
| 1649 | RequestLuma = REQ_128BytesNonContiguous; |
| 1650 | } else { |
| 1651 | RequestLuma = REQ_128BytesContiguous; |
| 1652 | } |
| 1653 | if (req128_horz_wc_c == 0 && req128_vert_wc_c == 0) { |
| 1654 | RequestChroma = REQ_256Bytes; |
| 1655 | } else if ((req128_horz_wc_c == 1 && segment_order_horz_contiguous_chroma == 0) || (req128_vert_wc_c == 1 && segment_order_vert_contiguous_chroma == 0)) { |
| 1656 | RequestChroma = REQ_128BytesNonContiguous; |
| 1657 | } else { |
| 1658 | RequestChroma = REQ_128BytesContiguous; |
| 1659 | } |
| 1660 | } else if (ScanOrientation != dm_vert) { |
| 1661 | if (req128_horz_wc_l == 0) { |
| 1662 | RequestLuma = REQ_256Bytes; |
| 1663 | } else if (segment_order_horz_contiguous_luma == 0) { |
| 1664 | RequestLuma = REQ_128BytesNonContiguous; |
| 1665 | } else { |
| 1666 | RequestLuma = REQ_128BytesContiguous; |
| 1667 | } |
| 1668 | if (req128_horz_wc_c == 0) { |
| 1669 | RequestChroma = REQ_256Bytes; |
| 1670 | } else if (segment_order_horz_contiguous_chroma == 0) { |
| 1671 | RequestChroma = REQ_128BytesNonContiguous; |
| 1672 | } else { |
| 1673 | RequestChroma = REQ_128BytesContiguous; |
| 1674 | } |
| 1675 | } else { |
| 1676 | if (req128_vert_wc_l == 0) { |
| 1677 | RequestLuma = REQ_256Bytes; |
| 1678 | } else if (segment_order_vert_contiguous_luma == 0) { |
| 1679 | RequestLuma = REQ_128BytesNonContiguous; |
| 1680 | } else { |
| 1681 | RequestLuma = REQ_128BytesContiguous; |
| 1682 | } |
| 1683 | if (req128_vert_wc_c == 0) { |
| 1684 | RequestChroma = REQ_256Bytes; |
| 1685 | } else if (segment_order_vert_contiguous_chroma == 0) { |
| 1686 | RequestChroma = REQ_128BytesNonContiguous; |
| 1687 | } else { |
| 1688 | RequestChroma = REQ_128BytesContiguous; |
| 1689 | } |
| 1690 | } |
| 1691 | |
| 1692 | if (RequestLuma == REQ_256Bytes) { |
| 1693 | *MaxUncompressedBlockLuma = 256; |
| 1694 | *MaxCompressedBlockLuma = 256; |
| 1695 | *IndependentBlockLuma = 0; |
| 1696 | } else if (RequestLuma == REQ_128BytesContiguous) { |
| 1697 | *MaxUncompressedBlockLuma = 256; |
| 1698 | *MaxCompressedBlockLuma = 128; |
| 1699 | *IndependentBlockLuma = 128; |
| 1700 | } else { |
| 1701 | *MaxUncompressedBlockLuma = 256; |
| 1702 | *MaxCompressedBlockLuma = 64; |
| 1703 | *IndependentBlockLuma = 64; |
| 1704 | } |
| 1705 | |
| 1706 | if (RequestChroma == REQ_256Bytes) { |
| 1707 | *MaxUncompressedBlockChroma = 256; |
| 1708 | *MaxCompressedBlockChroma = 256; |
| 1709 | *IndependentBlockChroma = 0; |
| 1710 | } else if (RequestChroma == REQ_128BytesContiguous) { |
| 1711 | *MaxUncompressedBlockChroma = 256; |
| 1712 | *MaxCompressedBlockChroma = 128; |
| 1713 | *IndependentBlockChroma = 128; |
| 1714 | } else { |
| 1715 | *MaxUncompressedBlockChroma = 256; |
| 1716 | *MaxCompressedBlockChroma = 64; |
| 1717 | *IndependentBlockChroma = 64; |
| 1718 | } |
| 1719 | |
| 1720 | if (DCCEnabled != true1 || BytePerPixelC == 0) { |
| 1721 | *MaxUncompressedBlockChroma = 0; |
| 1722 | *MaxCompressedBlockChroma = 0; |
| 1723 | *IndependentBlockChroma = 0; |
| 1724 | } |
| 1725 | |
| 1726 | if (DCCEnabled != true1) { |
| 1727 | *MaxUncompressedBlockLuma = 0; |
| 1728 | *MaxCompressedBlockLuma = 0; |
| 1729 | *IndependentBlockLuma = 0; |
| 1730 | } |
| 1731 | } |
| 1732 | |
| 1733 | static double CalculatePrefetchSourceLines( |
| 1734 | struct display_mode_lib *mode_lib, |
| 1735 | double VRatio, |
| 1736 | double vtaps, |
| 1737 | bool_Bool Interlace, |
| 1738 | bool_Bool ProgressiveToInterlaceUnitInOPP, |
| 1739 | unsigned int SwathHeight, |
| 1740 | unsigned int ViewportYStart, |
| 1741 | double *VInitPreFill, |
| 1742 | unsigned int *MaxNumSwath) |
| 1743 | { |
| 1744 | struct vba_vars_st *v = &mode_lib->vba; |
| 1745 | unsigned int MaxPartialSwath; |
| 1746 | |
| 1747 | if (ProgressiveToInterlaceUnitInOPP) |
| 1748 | *VInitPreFill = dml_floor((VRatio + vtaps + 1) / 2.0, 1); |
| 1749 | else |
| 1750 | *VInitPreFill = dml_floor((VRatio + vtaps + 1 + Interlace * 0.5 * VRatio) / 2.0, 1); |
| 1751 | |
| 1752 | if (!v->IgnoreViewportPositioning) { |
| 1753 | |
| 1754 | *MaxNumSwath = dml_ceil((*VInitPreFill - 1.0) / SwathHeight, 1) + 1.0; |
| 1755 | |
| 1756 | if (*VInitPreFill > 1.0) |
| 1757 | MaxPartialSwath = (unsigned int) (*VInitPreFill - 2) % SwathHeight; |
| 1758 | else |
| 1759 | MaxPartialSwath = (unsigned int) (*VInitPreFill + SwathHeight - 2) % SwathHeight; |
| 1760 | MaxPartialSwath = dml_max(1U, MaxPartialSwath); |
| 1761 | |
| 1762 | } else { |
| 1763 | |
| 1764 | if (ViewportYStart != 0) |
| 1765 | dml_print("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); }; |
| 1766 | |
| 1767 | *MaxNumSwath = dml_ceil(*VInitPreFill / SwathHeight, 1); |
| 1768 | |
| 1769 | if (*VInitPreFill > 1.0) |
| 1770 | MaxPartialSwath = (unsigned int) (*VInitPreFill - 1) % SwathHeight; |
| 1771 | else |
| 1772 | MaxPartialSwath = (unsigned int) (*VInitPreFill + SwathHeight - 1) % SwathHeight; |
| 1773 | } |
| 1774 | |
| 1775 | #ifdef __DML_VBA_DEBUG__ |
| 1776 | dml_print("DML::%s: VRatio = %f\n", __func__, VRatio){do { } while(0); }; |
| 1777 | dml_print("DML::%s: vtaps = %f\n", __func__, vtaps){do { } while(0); }; |
| 1778 | dml_print("DML::%s: VInitPreFill = %f\n", __func__, *VInitPreFill){do { } while(0); }; |
| 1779 | dml_print("DML::%s: ProgressiveToInterlaceUnitInOPP = %d\n", __func__, ProgressiveToInterlaceUnitInOPP){do { } while(0); }; |
| 1780 | dml_print("DML::%s: IgnoreViewportPositioning = %d\n", __func__, v->IgnoreViewportPositioning){do { } while(0); }; |
| 1781 | dml_print("DML::%s: SwathHeight = %d\n", __func__, SwathHeight){do { } while(0); }; |
| 1782 | dml_print("DML::%s: MaxPartialSwath = %d\n", __func__, MaxPartialSwath){do { } while(0); }; |
| 1783 | dml_print("DML::%s: MaxNumSwath = %d\n", __func__, *MaxNumSwath){do { } while(0); }; |
| 1784 | dml_print("DML::%s: Prefetch source lines = %d\n", __func__, *MaxNumSwath * SwathHeight + MaxPartialSwath){do { } while(0); }; |
| 1785 | #endif |
| 1786 | return *MaxNumSwath * SwathHeight + MaxPartialSwath; |
| 1787 | } |
| 1788 | |
| 1789 | static unsigned int CalculateVMAndRowBytes( |
| 1790 | struct display_mode_lib *mode_lib, |
| 1791 | bool_Bool DCCEnable, |
| 1792 | unsigned int BlockHeight256Bytes, |
| 1793 | unsigned int BlockWidth256Bytes, |
| 1794 | enum source_format_class SourcePixelFormat, |
| 1795 | unsigned int SurfaceTiling, |
| 1796 | unsigned int BytePerPixel, |
| 1797 | enum scan_direction_class ScanDirection, |
| 1798 | unsigned int SwathWidth, |
| 1799 | unsigned int ViewportHeight, |
| 1800 | bool_Bool GPUVMEnable, |
| 1801 | bool_Bool HostVMEnable, |
| 1802 | unsigned int HostVMMaxNonCachedPageTableLevels, |
| 1803 | unsigned int GPUVMMinPageSize, |
| 1804 | unsigned int HostVMMinPageSize, |
| 1805 | unsigned int PTEBufferSizeInRequests, |
| 1806 | unsigned int Pitch, |
| 1807 | unsigned int DCCMetaPitch, |
| 1808 | unsigned int *MacroTileWidth, |
| 1809 | unsigned int *MetaRowByte, |
| 1810 | unsigned int *PixelPTEBytesPerRow, |
| 1811 | bool_Bool *PTEBufferSizeNotExceeded, |
| 1812 | int *dpte_row_width_ub, |
| 1813 | unsigned int *dpte_row_height, |
| 1814 | unsigned int *MetaRequestWidth, |
| 1815 | unsigned int *MetaRequestHeight, |
| 1816 | unsigned int *meta_row_width, |
| 1817 | unsigned int *meta_row_height, |
| 1818 | int *vm_group_bytes, |
| 1819 | unsigned int *dpte_group_bytes, |
| 1820 | unsigned int *PixelPTEReqWidth, |
| 1821 | unsigned int *PixelPTEReqHeight, |
| 1822 | unsigned int *PTERequestSize, |
| 1823 | int *DPDE0BytesFrame, |
| 1824 | int *MetaPTEBytesFrame) |
| 1825 | { |
| 1826 | struct vba_vars_st *v = &mode_lib->vba; |
| 1827 | unsigned int MPDEBytesFrame; |
| 1828 | unsigned int DCCMetaSurfaceBytes; |
| 1829 | unsigned int MacroTileSizeBytes; |
| 1830 | unsigned int MacroTileHeight; |
| 1831 | unsigned int ExtraDPDEBytesFrame; |
| 1832 | unsigned int PDEAndMetaPTEBytesFrame; |
| 1833 | unsigned int PixelPTEReqHeightPTEs = 0; |
| 1834 | unsigned int HostVMDynamicLevels = 0; |
| 1835 | double FractionOfPTEReturnDrop; |
| 1836 | |
| 1837 | if (GPUVMEnable == true1 && HostVMEnable == true1) { |
| 1838 | if (HostVMMinPageSize < 2048) { |
| 1839 | HostVMDynamicLevels = HostVMMaxNonCachedPageTableLevels; |
| 1840 | } else if (HostVMMinPageSize >= 2048 && HostVMMinPageSize < 1048576) { |
| 1841 | HostVMDynamicLevels = dml_max(0, (int) HostVMMaxNonCachedPageTableLevels - 1); |
| 1842 | } else { |
| 1843 | HostVMDynamicLevels = dml_max(0, (int) HostVMMaxNonCachedPageTableLevels - 2); |
| 1844 | } |
| 1845 | } |
| 1846 | |
| 1847 | *MetaRequestHeight = 8 * BlockHeight256Bytes; |
| 1848 | *MetaRequestWidth = 8 * BlockWidth256Bytes; |
| 1849 | if (ScanDirection != dm_vert) { |
| 1850 | *meta_row_height = *MetaRequestHeight; |
| 1851 | *meta_row_width = dml_ceil((double) SwathWidth - 1, *MetaRequestWidth) + *MetaRequestWidth; |
| 1852 | *MetaRowByte = *meta_row_width * *MetaRequestHeight * BytePerPixel / 256.0; |
| 1853 | } else { |
| 1854 | *meta_row_height = *MetaRequestWidth; |
| 1855 | *meta_row_width = dml_ceil((double) SwathWidth - 1, *MetaRequestHeight) + *MetaRequestHeight; |
| 1856 | *MetaRowByte = *meta_row_width * *MetaRequestWidth * BytePerPixel / 256.0; |
| 1857 | } |
| 1858 | DCCMetaSurfaceBytes = DCCMetaPitch * (dml_ceil(ViewportHeight - 1, 64 * BlockHeight256Bytes) + 64 * BlockHeight256Bytes) * BytePerPixel / 256; |
| 1859 | if (GPUVMEnable == true1) { |
| 1860 | *MetaPTEBytesFrame = (dml_ceil((double) (DCCMetaSurfaceBytes - 4.0 * 1024.0) / (8 * 4.0 * 1024), 1) + 1) * 64; |
| 1861 | MPDEBytesFrame = 128 * (v->GPUVMMaxPageTableLevels - 1); |
| 1862 | } else { |
| 1863 | *MetaPTEBytesFrame = 0; |
| 1864 | MPDEBytesFrame = 0; |
| 1865 | } |
| 1866 | |
| 1867 | if (DCCEnable != true1) { |
| 1868 | *MetaPTEBytesFrame = 0; |
| 1869 | MPDEBytesFrame = 0; |
| 1870 | *MetaRowByte = 0; |
| 1871 | } |
| 1872 | |
| 1873 | if (SurfaceTiling == dm_sw_linear) { |
| 1874 | MacroTileSizeBytes = 256; |
| 1875 | MacroTileHeight = BlockHeight256Bytes; |
| 1876 | } else { |
| 1877 | MacroTileSizeBytes = 65536; |
| 1878 | MacroTileHeight = 16 * BlockHeight256Bytes; |
| 1879 | } |
| 1880 | *MacroTileWidth = MacroTileSizeBytes / BytePerPixel / MacroTileHeight; |
| 1881 | |
| 1882 | if (GPUVMEnable == true1 && v->GPUVMMaxPageTableLevels > 1) { |
| 1883 | if (ScanDirection != dm_vert) { |
| 1884 | *DPDE0BytesFrame = 64 |
| 1885 | * (dml_ceil( |
| 1886 | ((Pitch * (dml_ceil(ViewportHeight - 1, MacroTileHeight) + MacroTileHeight) * BytePerPixel) - MacroTileSizeBytes) |
| 1887 | / (8 * 2097152), |
| 1888 | 1) + 1); |
| 1889 | } else { |
| 1890 | *DPDE0BytesFrame = 64 |
| 1891 | * (dml_ceil( |
| 1892 | ((Pitch * (dml_ceil((double) SwathWidth - 1, MacroTileHeight) + MacroTileHeight) * BytePerPixel) - MacroTileSizeBytes) |
| 1893 | / (8 * 2097152), |
| 1894 | 1) + 1); |
| 1895 | } |
| 1896 | ExtraDPDEBytesFrame = 128 * (v->GPUVMMaxPageTableLevels - 2); |
| 1897 | } else { |
| 1898 | *DPDE0BytesFrame = 0; |
| 1899 | ExtraDPDEBytesFrame = 0; |
| 1900 | } |
| 1901 | |
| 1902 | PDEAndMetaPTEBytesFrame = *MetaPTEBytesFrame + MPDEBytesFrame + *DPDE0BytesFrame + ExtraDPDEBytesFrame; |
| 1903 | |
| 1904 | #ifdef __DML_VBA_DEBUG__ |
| 1905 | dml_print("DML::%s: MetaPTEBytesFrame = %d\n", __func__, *MetaPTEBytesFrame){do { } while(0); }; |
| 1906 | dml_print("DML::%s: MPDEBytesFrame = %d\n", __func__, MPDEBytesFrame){do { } while(0); }; |
| 1907 | dml_print("DML::%s: DPDE0BytesFrame = %d\n", __func__, *DPDE0BytesFrame){do { } while(0); }; |
| 1908 | dml_print("DML::%s: ExtraDPDEBytesFrame= %d\n", __func__, ExtraDPDEBytesFrame){do { } while(0); }; |
| 1909 | dml_print("DML::%s: PDEAndMetaPTEBytesFrame = %d\n", __func__, PDEAndMetaPTEBytesFrame){do { } while(0); }; |
| 1910 | #endif |
| 1911 | |
| 1912 | if (HostVMEnable == true1) { |
| 1913 | PDEAndMetaPTEBytesFrame = PDEAndMetaPTEBytesFrame * (1 + 8 * HostVMDynamicLevels); |
| 1914 | } |
| 1915 | #ifdef __DML_VBA_DEBUG__ |
| 1916 | dml_print("DML::%s: PDEAndMetaPTEBytesFrame = %d\n", __func__, PDEAndMetaPTEBytesFrame){do { } while(0); }; |
| 1917 | #endif |
| 1918 | |
| 1919 | if (SurfaceTiling == dm_sw_linear) { |
| 1920 | PixelPTEReqHeightPTEs = 1; |
| 1921 | *PixelPTEReqHeight = 1; |
| 1922 | *PixelPTEReqWidth = 32768.0 / BytePerPixel; |
| 1923 | *PTERequestSize = 64; |
| 1924 | FractionOfPTEReturnDrop = 0; |
| 1925 | } else if (MacroTileSizeBytes == 4096) { |
| 1926 | PixelPTEReqHeightPTEs = 1; |
| 1927 | *PixelPTEReqHeight = MacroTileHeight; |
| 1928 | *PixelPTEReqWidth = 8 * *MacroTileWidth; |
| 1929 | *PTERequestSize = 64; |
| 1930 | if (ScanDirection != dm_vert) |
| 1931 | FractionOfPTEReturnDrop = 0; |
| 1932 | else |
| 1933 | FractionOfPTEReturnDrop = 7 / 8; |
| 1934 | } else if (GPUVMMinPageSize == 4 && MacroTileSizeBytes > 4096) { |
| 1935 | PixelPTEReqHeightPTEs = 16; |
| 1936 | *PixelPTEReqHeight = 16 * BlockHeight256Bytes; |
| 1937 | *PixelPTEReqWidth = 16 * BlockWidth256Bytes; |
| 1938 | *PTERequestSize = 128; |
| 1939 | FractionOfPTEReturnDrop = 0; |
| 1940 | } else { |
| 1941 | PixelPTEReqHeightPTEs = 1; |
| 1942 | *PixelPTEReqHeight = MacroTileHeight; |
| 1943 | *PixelPTEReqWidth = 8 * *MacroTileWidth; |
| 1944 | *PTERequestSize = 64; |
| 1945 | FractionOfPTEReturnDrop = 0; |
| 1946 | } |
| 1947 | |
| 1948 | if (SurfaceTiling == dm_sw_linear) { |
| 1949 | *dpte_row_height = dml_min(128, 1 << (unsigned int) dml_floor(dml_log2(PTEBufferSizeInRequests * *PixelPTEReqWidth / Pitch), 1)); |
| 1950 | *dpte_row_width_ub = (dml_ceil((double)(Pitch * *dpte_row_height - 1) / *PixelPTEReqWidth, 1) + 1) * *PixelPTEReqWidth; |
| 1951 | *PixelPTEBytesPerRow = *dpte_row_width_ub / *PixelPTEReqWidth * *PTERequestSize; |
| 1952 | } else if (ScanDirection != dm_vert) { |
| 1953 | *dpte_row_height = *PixelPTEReqHeight; |
| 1954 | *dpte_row_width_ub = (dml_ceil((double) (SwathWidth - 1) / *PixelPTEReqWidth, 1) + 1) * *PixelPTEReqWidth; |
| 1955 | *PixelPTEBytesPerRow = *dpte_row_width_ub / *PixelPTEReqWidth * *PTERequestSize; |
| 1956 | } else { |
| 1957 | *dpte_row_height = dml_min(*PixelPTEReqWidth, *MacroTileWidth); |
| 1958 | *dpte_row_width_ub = (dml_ceil((double) (SwathWidth - 1) / *PixelPTEReqHeight, 1) + 1) * *PixelPTEReqHeight; |
| 1959 | *PixelPTEBytesPerRow = *dpte_row_width_ub / *PixelPTEReqHeight * *PTERequestSize; |
| 1960 | } |
| 1961 | |
| 1962 | if (*PixelPTEBytesPerRow * (1 - FractionOfPTEReturnDrop) <= 64 * PTEBufferSizeInRequests) { |
| 1963 | *PTEBufferSizeNotExceeded = true1; |
| 1964 | } else { |
| 1965 | *PTEBufferSizeNotExceeded = false0; |
| 1966 | } |
| 1967 | |
| 1968 | if (GPUVMEnable != true1) { |
| 1969 | *PixelPTEBytesPerRow = 0; |
| 1970 | *PTEBufferSizeNotExceeded = true1; |
| 1971 | } |
| 1972 | |
| 1973 | dml_print("DML: vm_bytes = meta_pte_bytes_per_frame (per_pipe) = MetaPTEBytesFrame = : %i\n", *MetaPTEBytesFrame){do { } while(0); }; |
| 1974 | |
| 1975 | if (HostVMEnable == true1) { |
| 1976 | *PixelPTEBytesPerRow = *PixelPTEBytesPerRow * (1 + 8 * HostVMDynamicLevels); |
| 1977 | } |
| 1978 | |
| 1979 | if (HostVMEnable == true1) { |
| 1980 | *vm_group_bytes = 512; |
| 1981 | *dpte_group_bytes = 512; |
| 1982 | } else if (GPUVMEnable == true1) { |
| 1983 | *vm_group_bytes = 2048; |
| 1984 | if (SurfaceTiling != dm_sw_linear && PixelPTEReqHeightPTEs == 1 && ScanDirection == dm_vert) { |
| 1985 | *dpte_group_bytes = 512; |
| 1986 | } else { |
| 1987 | *dpte_group_bytes = 2048; |
| 1988 | } |
| 1989 | } else { |
| 1990 | *vm_group_bytes = 0; |
| 1991 | *dpte_group_bytes = 0; |
| 1992 | } |
| 1993 | return PDEAndMetaPTEBytesFrame; |
| 1994 | } |
| 1995 | |
| 1996 | static void DISPCLKDPPCLKDCFCLKDeepSleepPrefetchParametersWatermarksAndPerformanceCalculation(struct display_mode_lib *mode_lib) |
| 1997 | { |
| 1998 | struct vba_vars_st *v = &mode_lib->vba; |
| 1999 | unsigned int j, k; |
| 2000 | double HostVMInefficiencyFactor = 1.0; |
| 2001 | bool_Bool NoChromaPlanes = true1; |
| 2002 | int ReorderBytes; |
| 2003 | double VMDataOnlyReturnBW; |
| 2004 | double MaxTotalRDBandwidth = 0; |
| 2005 | int PrefetchMode = v->PrefetchModePerState[v->VoltageLevel][v->maxMpcComb]; |
| 2006 | |
| 2007 | v->WritebackDISPCLK = 0.0; |
| 2008 | v->DISPCLKWithRamping = 0; |
| 2009 | v->DISPCLKWithoutRamping = 0; |
| 2010 | v->GlobalDPPCLK = 0.0; |
| 2011 | /* DAL custom code: need to update ReturnBW in case min dcfclk is overriden */ |
| 2012 | { |
| 2013 | double IdealFabricAndSDPPortBandwidthPerState = dml_min( |
| 2014 | v->ReturnBusWidth * v->DCFCLKState[v->VoltageLevel][v->maxMpcComb], |
| 2015 | v->FabricClockPerState[v->VoltageLevel] * v->FabricDatapathToDCNDataReturn); |
| 2016 | double IdealDRAMBandwidthPerState = v->DRAMSpeedPerState[v->VoltageLevel] * v->NumberOfChannels * v->DRAMChannelWidth; |
| 2017 | if (v->HostVMEnable != true1) { |
| 2018 | v->ReturnBW = dml_min( |
| 2019 | IdealFabricAndSDPPortBandwidthPerState * v->PercentOfIdealFabricAndSDPPortBWReceivedAfterUrgLatency / 100.0, |
| 2020 | IdealDRAMBandwidthPerState * v->PercentOfIdealDRAMBWReceivedAfterUrgLatencyPixelDataOnly / 100.0); |
| 2021 | } else { |
| 2022 | v->ReturnBW = dml_min( |
| 2023 | IdealFabricAndSDPPortBandwidthPerState * v->PercentOfIdealFabricAndSDPPortBWReceivedAfterUrgLatency / 100.0, |
| 2024 | IdealDRAMBandwidthPerState * v->PercentOfIdealDRAMBWReceivedAfterUrgLatencyPixelMixedWithVMData / 100.0); |
| 2025 | } |
| 2026 | } |
| 2027 | /* End DAL custom code */ |
| 2028 | |
| 2029 | // DISPCLK and DPPCLK Calculation |
| 2030 | // |
| 2031 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 2032 | if (v->WritebackEnable[k]) { |
| 2033 | v->WritebackDISPCLK = dml_max( |
| 2034 | v->WritebackDISPCLK, |
| 2035 | dml31_CalculateWriteBackDISPCLK( |
| 2036 | v->WritebackPixelFormat[k], |
| 2037 | v->PixelClock[k], |
| 2038 | v->WritebackHRatio[k], |
| 2039 | v->WritebackVRatio[k], |
| 2040 | v->WritebackHTaps[k], |
| 2041 | v->WritebackVTaps[k], |
| 2042 | v->WritebackSourceWidth[k], |
| 2043 | v->WritebackDestinationWidth[k], |
| 2044 | v->HTotal[k], |
| 2045 | v->WritebackLineBufferSize)); |
| 2046 | } |
| 2047 | } |
| 2048 | |
| 2049 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 2050 | if (v->HRatio[k] > 1) { |
| 2051 | v->PSCL_THROUGHPUT_LUMA[k] = dml_min( |
| 2052 | v->MaxDCHUBToPSCLThroughput, |
| 2053 | v->MaxPSCLToLBThroughput * v->HRatio[k] / dml_ceil(v->htaps[k] / 6.0, 1)); |
| 2054 | } else { |
| 2055 | v->PSCL_THROUGHPUT_LUMA[k] = dml_min(v->MaxDCHUBToPSCLThroughput, v->MaxPSCLToLBThroughput); |
| 2056 | } |
| 2057 | |
| 2058 | v->DPPCLKUsingSingleDPPLuma = v->PixelClock[k] |
| 2059 | * dml_max( |
| 2060 | v->vtaps[k] / 6.0 * dml_min(1.0, v->HRatio[k]), |
| 2061 | dml_max(v->HRatio[k] * v->VRatio[k] / v->PSCL_THROUGHPUT_LUMA[k], 1.0)); |
| 2062 | |
| 2063 | if ((v->htaps[k] > 6 || v->vtaps[k] > 6) && v->DPPCLKUsingSingleDPPLuma < 2 * v->PixelClock[k]) { |
| 2064 | v->DPPCLKUsingSingleDPPLuma = 2 * v->PixelClock[k]; |
| 2065 | } |
| 2066 | |
| 2067 | if ((v->SourcePixelFormat[k] != dm_420_8 && v->SourcePixelFormat[k] != dm_420_10 && v->SourcePixelFormat[k] != dm_420_12 |
| 2068 | && v->SourcePixelFormat[k] != dm_rgbe_alpha)) { |
| 2069 | v->PSCL_THROUGHPUT_CHROMA[k] = 0.0; |
| 2070 | v->DPPCLKUsingSingleDPP[k] = v->DPPCLKUsingSingleDPPLuma; |
| 2071 | } else { |
| 2072 | if (v->HRatioChroma[k] > 1) { |
| 2073 | v->PSCL_THROUGHPUT_CHROMA[k] = dml_min( |
| 2074 | v->MaxDCHUBToPSCLThroughput, |
| 2075 | v->MaxPSCLToLBThroughput * v->HRatioChroma[k] / dml_ceil(v->HTAPsChroma[k] / 6.0, 1.0)); |
| 2076 | } else { |
| 2077 | v->PSCL_THROUGHPUT_CHROMA[k] = dml_min(v->MaxDCHUBToPSCLThroughput, v->MaxPSCLToLBThroughput); |
| 2078 | } |
| 2079 | v->DPPCLKUsingSingleDPPChroma = v->PixelClock[k] |
| 2080 | * dml_max3( |
| 2081 | v->VTAPsChroma[k] / 6.0 * dml_min(1.0, v->HRatioChroma[k]), |
| 2082 | v->HRatioChroma[k] * v->VRatioChroma[k] / v->PSCL_THROUGHPUT_CHROMA[k], |
| 2083 | 1.0); |
| 2084 | |
| 2085 | if ((v->HTAPsChroma[k] > 6 || v->VTAPsChroma[k] > 6) && v->DPPCLKUsingSingleDPPChroma < 2 * v->PixelClock[k]) { |
| 2086 | v->DPPCLKUsingSingleDPPChroma = 2 * v->PixelClock[k]; |
| 2087 | } |
| 2088 | |
| 2089 | v->DPPCLKUsingSingleDPP[k] = dml_max(v->DPPCLKUsingSingleDPPLuma, v->DPPCLKUsingSingleDPPChroma); |
| 2090 | } |
| 2091 | } |
| 2092 | |
| 2093 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 2094 | if (v->BlendingAndTiming[k] != k) |
| 2095 | continue; |
| 2096 | if (v->ODMCombineEnabled[k] == dm_odm_combine_mode_4to1) { |
| 2097 | v->DISPCLKWithRamping = dml_max( |
| 2098 | v->DISPCLKWithRamping, |
| 2099 | v->PixelClock[k] / 4 * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100) |
| 2100 | * (1 + v->DISPCLKRampingMargin / 100)); |
| 2101 | v->DISPCLKWithoutRamping = dml_max( |
| 2102 | v->DISPCLKWithoutRamping, |
| 2103 | v->PixelClock[k] / 4 * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100)); |
| 2104 | } else if (v->ODMCombineEnabled[k] == dm_odm_combine_mode_2to1) { |
| 2105 | v->DISPCLKWithRamping = dml_max( |
| 2106 | v->DISPCLKWithRamping, |
| 2107 | v->PixelClock[k] / 2 * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100) |
| 2108 | * (1 + v->DISPCLKRampingMargin / 100)); |
| 2109 | v->DISPCLKWithoutRamping = dml_max( |
| 2110 | v->DISPCLKWithoutRamping, |
| 2111 | v->PixelClock[k] / 2 * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100)); |
| 2112 | } else { |
| 2113 | v->DISPCLKWithRamping = dml_max( |
| 2114 | v->DISPCLKWithRamping, |
| 2115 | v->PixelClock[k] * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100) * (1 + v->DISPCLKRampingMargin / 100)); |
| 2116 | v->DISPCLKWithoutRamping = dml_max( |
| 2117 | v->DISPCLKWithoutRamping, |
| 2118 | v->PixelClock[k] * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100)); |
| 2119 | } |
| 2120 | } |
| 2121 | |
| 2122 | v->DISPCLKWithRamping = dml_max(v->DISPCLKWithRamping, v->WritebackDISPCLK); |
| 2123 | v->DISPCLKWithoutRamping = dml_max(v->DISPCLKWithoutRamping, v->WritebackDISPCLK); |
| 2124 | |
| 2125 | ASSERT(v->DISPCLKDPPCLKVCOSpeed != 0)do { if (({ static int __warned; int __ret = !!(!(v->DISPCLKDPPCLKVCOSpeed != 0)); if (__ret && !__warned) { printf("WARNING %s failed at %s:%d\n" , "!(v->DISPCLKDPPCLKVCOSpeed != 0)", "/usr/src/sys/dev/pci/drm/amd/display/dc/dml/dcn31/display_mode_vba_31.c" , 2125); __warned = 1; } __builtin_expect(!!(__ret), 0); })) do {} while (0); } while (0); |
| 2126 | v->DISPCLKWithRampingRoundedToDFSGranularity = RoundToDFSGranularityUp(v->DISPCLKWithRamping, v->DISPCLKDPPCLKVCOSpeed); |
| 2127 | v->DISPCLKWithoutRampingRoundedToDFSGranularity = RoundToDFSGranularityUp(v->DISPCLKWithoutRamping, v->DISPCLKDPPCLKVCOSpeed); |
| 2128 | v->MaxDispclkRoundedToDFSGranularity = RoundToDFSGranularityDown( |
| 2129 | v->soc.clock_limits[v->soc.num_states - 1].dispclk_mhz, |
| 2130 | v->DISPCLKDPPCLKVCOSpeed); |
| 2131 | if (v->DISPCLKWithoutRampingRoundedToDFSGranularity > v->MaxDispclkRoundedToDFSGranularity) { |
| 2132 | v->DISPCLK_calculated = v->DISPCLKWithoutRampingRoundedToDFSGranularity; |
| 2133 | } else if (v->DISPCLKWithRampingRoundedToDFSGranularity > v->MaxDispclkRoundedToDFSGranularity) { |
| 2134 | v->DISPCLK_calculated = v->MaxDispclkRoundedToDFSGranularity; |
| 2135 | } else { |
| 2136 | v->DISPCLK_calculated = v->DISPCLKWithRampingRoundedToDFSGranularity; |
| 2137 | } |
| 2138 | v->DISPCLK = v->DISPCLK_calculated; |
| 2139 | DTRACE(" dispclk_mhz (calculated) = %f", v->DISPCLK_calculated){do { } while(0); }; |
| 2140 | |
| 2141 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 2142 | v->DPPCLK_calculated[k] = v->DPPCLKUsingSingleDPP[k] / v->DPPPerPlane[k] * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100); |
| 2143 | v->GlobalDPPCLK = dml_max(v->GlobalDPPCLK, v->DPPCLK_calculated[k]); |
| 2144 | } |
| 2145 | v->GlobalDPPCLK = RoundToDFSGranularityUp(v->GlobalDPPCLK, v->DISPCLKDPPCLKVCOSpeed); |
| 2146 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 2147 | v->DPPCLK_calculated[k] = v->GlobalDPPCLK / 255 * dml_ceil(v->DPPCLK_calculated[k] * 255.0 / v->GlobalDPPCLK, 1); |
| 2148 | DTRACE(" dppclk_mhz[%i] (calculated) = %f", k, v->DPPCLK_calculated[k]){do { } while(0); }; |
| 2149 | } |
| 2150 | |
| 2151 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 2152 | v->DPPCLK[k] = v->DPPCLK_calculated[k]; |
| 2153 | } |
| 2154 | |
| 2155 | // Urgent and B P-State/DRAM Clock Change Watermark |
| 2156 | DTRACE(" dcfclk_mhz = %f", v->DCFCLK){do { } while(0); }; |
| 2157 | DTRACE(" return_bus_bw = %f", v->ReturnBW){do { } while(0); }; |
| 2158 | |
| 2159 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 2160 | dml30_CalculateBytePerPixelAnd256BBlockSizes( |
| 2161 | v->SourcePixelFormat[k], |
| 2162 | v->SurfaceTiling[k], |
| 2163 | &v->BytePerPixelY[k], |
| 2164 | &v->BytePerPixelC[k], |
| 2165 | &v->BytePerPixelDETY[k], |
| 2166 | &v->BytePerPixelDETC[k], |
| 2167 | &v->BlockHeight256BytesY[k], |
| 2168 | &v->BlockHeight256BytesC[k], |
| 2169 | &v->BlockWidth256BytesY[k], |
| 2170 | &v->BlockWidth256BytesC[k]); |
| 2171 | } |
| 2172 | |
| 2173 | CalculateSwathWidth( |
| 2174 | false0, |
| 2175 | v->NumberOfActivePlanes, |
| 2176 | v->SourcePixelFormat, |
| 2177 | v->SourceScan, |
| 2178 | v->ViewportWidth, |
| 2179 | v->ViewportHeight, |
| 2180 | v->SurfaceWidthY, |
| 2181 | v->SurfaceWidthC, |
| 2182 | v->SurfaceHeightY, |
| 2183 | v->SurfaceHeightC, |
| 2184 | v->ODMCombineEnabled, |
| 2185 | v->BytePerPixelY, |
| 2186 | v->BytePerPixelC, |
| 2187 | v->BlockHeight256BytesY, |
| 2188 | v->BlockHeight256BytesC, |
| 2189 | v->BlockWidth256BytesY, |
| 2190 | v->BlockWidth256BytesC, |
| 2191 | v->BlendingAndTiming, |
| 2192 | v->HActive, |
| 2193 | v->HRatio, |
| 2194 | v->DPPPerPlane, |
| 2195 | v->SwathWidthSingleDPPY, |
| 2196 | v->SwathWidthSingleDPPC, |
| 2197 | v->SwathWidthY, |
| 2198 | v->SwathWidthC, |
| 2199 | v->dummyinteger3, |
| 2200 | v->dummyinteger4, |
| 2201 | v->swath_width_luma_ub, |
| 2202 | v->swath_width_chroma_ub); |
| 2203 | |
| 2204 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 2205 | v->ReadBandwidthPlaneLuma[k] = v->SwathWidthSingleDPPY[k] * v->BytePerPixelY[k] / (v->HTotal[k] / v->PixelClock[k]) |
| 2206 | * v->VRatio[k]; |
| 2207 | v->ReadBandwidthPlaneChroma[k] = v->SwathWidthSingleDPPC[k] * v->BytePerPixelC[k] / (v->HTotal[k] / v->PixelClock[k]) |
| 2208 | * v->VRatioChroma[k]; |
| 2209 | DTRACE(" read_bw[%i] = %fBps", k, v->ReadBandwidthPlaneLuma[k] + v->ReadBandwidthPlaneChroma[k]){do { } while(0); }; |
| 2210 | } |
| 2211 | |
| 2212 | // DCFCLK Deep Sleep |
| 2213 | CalculateDCFCLKDeepSleep( |
| 2214 | mode_lib, |
| 2215 | v->NumberOfActivePlanes, |
| 2216 | v->BytePerPixelY, |
| 2217 | v->BytePerPixelC, |
| 2218 | v->VRatio, |
| 2219 | v->VRatioChroma, |
| 2220 | v->SwathWidthY, |
| 2221 | v->SwathWidthC, |
| 2222 | v->DPPPerPlane, |
| 2223 | v->HRatio, |
| 2224 | v->HRatioChroma, |
| 2225 | v->PixelClock, |
| 2226 | v->PSCL_THROUGHPUT_LUMA, |
| 2227 | v->PSCL_THROUGHPUT_CHROMA, |
| 2228 | v->DPPCLK, |
| 2229 | v->ReadBandwidthPlaneLuma, |
| 2230 | v->ReadBandwidthPlaneChroma, |
| 2231 | v->ReturnBusWidth, |
| 2232 | &v->DCFCLKDeepSleep); |
| 2233 | |
| 2234 | // DSCCLK |
| 2235 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 2236 | if ((v->BlendingAndTiming[k] != k) || !v->DSCEnabled[k]) { |
| 2237 | v->DSCCLK_calculated[k] = 0.0; |
| 2238 | } else { |
| 2239 | if (v->OutputFormat[k] == dm_420) |
| 2240 | v->DSCFormatFactor = 2; |
| 2241 | else if (v->OutputFormat[k] == dm_444) |
| 2242 | v->DSCFormatFactor = 1; |
| 2243 | else if (v->OutputFormat[k] == dm_n422) |
| 2244 | v->DSCFormatFactor = 2; |
| 2245 | else |
| 2246 | v->DSCFormatFactor = 1; |
| 2247 | if (v->ODMCombineEnabled[k] == dm_odm_combine_mode_4to1) |
| 2248 | v->DSCCLK_calculated[k] = v->PixelClockBackEnd[k] / 12 / v->DSCFormatFactor |
| 2249 | / (1 - v->DISPCLKDPPCLKDSCCLKDownSpreading / 100); |
| 2250 | else if (v->ODMCombineEnabled[k] == dm_odm_combine_mode_2to1) |
| 2251 | v->DSCCLK_calculated[k] = v->PixelClockBackEnd[k] / 6 / v->DSCFormatFactor |
| 2252 | / (1 - v->DISPCLKDPPCLKDSCCLKDownSpreading / 100); |
| 2253 | else |
| 2254 | v->DSCCLK_calculated[k] = v->PixelClockBackEnd[k] / 3 / v->DSCFormatFactor |
| 2255 | / (1 - v->DISPCLKDPPCLKDSCCLKDownSpreading / 100); |
| 2256 | } |
| 2257 | } |
| 2258 | |
| 2259 | // DSC Delay |
| 2260 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 2261 | double BPP = v->OutputBpp[k]; |
| 2262 | |
| 2263 | if (v->DSCEnabled[k] && BPP != 0) { |
| 2264 | if (v->ODMCombineEnabled[k] == dm_odm_combine_mode_disabled) { |
| 2265 | v->DSCDelay[k] = dscceComputeDelay( |
| 2266 | v->DSCInputBitPerComponent[k], |
| 2267 | BPP, |
| 2268 | dml_ceil((double) v->HActive[k] / v->NumberOfDSCSlices[k], 1), |
| 2269 | v->NumberOfDSCSlices[k], |
| 2270 | v->OutputFormat[k], |
| 2271 | v->Output[k]) + dscComputeDelay(v->OutputFormat[k], v->Output[k]); |
| 2272 | } else if (v->ODMCombineEnabled[k] == dm_odm_combine_mode_2to1) { |
| 2273 | v->DSCDelay[k] = 2 |
| 2274 | * (dscceComputeDelay( |
| 2275 | v->DSCInputBitPerComponent[k], |
| 2276 | BPP, |
| 2277 | dml_ceil((double) v->HActive[k] / v->NumberOfDSCSlices[k], 1), |
| 2278 | v->NumberOfDSCSlices[k] / 2.0, |
| 2279 | v->OutputFormat[k], |
| 2280 | v->Output[k]) + dscComputeDelay(v->OutputFormat[k], v->Output[k])); |
| 2281 | } else { |
| 2282 | v->DSCDelay[k] = 4 |
| 2283 | * (dscceComputeDelay( |
| 2284 | v->DSCInputBitPerComponent[k], |
| 2285 | BPP, |
| 2286 | dml_ceil((double) v->HActive[k] / v->NumberOfDSCSlices[k], 1), |
| 2287 | v->NumberOfDSCSlices[k] / 4.0, |
| 2288 | v->OutputFormat[k], |
| 2289 | v->Output[k]) + dscComputeDelay(v->OutputFormat[k], v->Output[k])); |
| 2290 | } |
| 2291 | v->DSCDelay[k] = v->DSCDelay[k] * v->PixelClock[k] / v->PixelClockBackEnd[k]; |
| 2292 | } else { |
| 2293 | v->DSCDelay[k] = 0; |
| 2294 | } |
| 2295 | } |
| 2296 | |
| 2297 | for (k = 0; k < v->NumberOfActivePlanes; ++k) |
| 2298 | for (j = 0; j < v->NumberOfActivePlanes; ++j) // NumberOfPlanes |
| 2299 | if (j != k && v->BlendingAndTiming[k] == j && v->DSCEnabled[j]) |
| 2300 | v->DSCDelay[k] = v->DSCDelay[j]; |
| 2301 | |
| 2302 | // Prefetch |
| 2303 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 2304 | unsigned int PDEAndMetaPTEBytesFrameY; |
| 2305 | unsigned int PixelPTEBytesPerRowY; |
| 2306 | unsigned int MetaRowByteY; |
| 2307 | unsigned int MetaRowByteC; |
| 2308 | unsigned int PDEAndMetaPTEBytesFrameC; |
| 2309 | unsigned int PixelPTEBytesPerRowC; |
| 2310 | bool_Bool PTEBufferSizeNotExceededY; |
| 2311 | bool_Bool PTEBufferSizeNotExceededC; |
| 2312 | |
| 2313 | if (v->SourcePixelFormat[k] == dm_420_8 || v->SourcePixelFormat[k] == dm_420_10 || v->SourcePixelFormat[k] == dm_420_12 |
| 2314 | || v->SourcePixelFormat[k] == dm_rgbe_alpha) { |
| 2315 | if ((v->SourcePixelFormat[k] == dm_420_10 || v->SourcePixelFormat[k] == dm_420_12) && v->SourceScan[k] != dm_vert) { |
| 2316 | v->PTEBufferSizeInRequestsForLuma = (v->PTEBufferSizeInRequestsLuma + v->PTEBufferSizeInRequestsChroma) / 2; |
| 2317 | v->PTEBufferSizeInRequestsForChroma = v->PTEBufferSizeInRequestsForLuma; |
| 2318 | } else { |
| 2319 | v->PTEBufferSizeInRequestsForLuma = v->PTEBufferSizeInRequestsLuma; |
| 2320 | v->PTEBufferSizeInRequestsForChroma = v->PTEBufferSizeInRequestsChroma; |
| 2321 | } |
| 2322 | |
| 2323 | PDEAndMetaPTEBytesFrameC = CalculateVMAndRowBytes( |
| 2324 | mode_lib, |
| 2325 | v->DCCEnable[k], |
| 2326 | v->BlockHeight256BytesC[k], |
| 2327 | v->BlockWidth256BytesC[k], |
| 2328 | v->SourcePixelFormat[k], |
| 2329 | v->SurfaceTiling[k], |
| 2330 | v->BytePerPixelC[k], |
| 2331 | v->SourceScan[k], |
| 2332 | v->SwathWidthC[k], |
| 2333 | v->ViewportHeightChroma[k], |
| 2334 | v->GPUVMEnable, |
| 2335 | v->HostVMEnable, |
| 2336 | v->HostVMMaxNonCachedPageTableLevels, |
| 2337 | v->GPUVMMinPageSize, |
| 2338 | v->HostVMMinPageSize, |
| 2339 | v->PTEBufferSizeInRequestsForChroma, |
| 2340 | v->PitchC[k], |
| 2341 | v->DCCMetaPitchC[k], |
| 2342 | &v->MacroTileWidthC[k], |
| 2343 | &MetaRowByteC, |
| 2344 | &PixelPTEBytesPerRowC, |
| 2345 | &PTEBufferSizeNotExceededC, |
| 2346 | &v->dpte_row_width_chroma_ub[k], |
| 2347 | &v->dpte_row_height_chroma[k], |
| 2348 | &v->meta_req_width_chroma[k], |
| 2349 | &v->meta_req_height_chroma[k], |
| 2350 | &v->meta_row_width_chroma[k], |
| 2351 | &v->meta_row_height_chroma[k], |
| 2352 | &v->dummyinteger1, |
| 2353 | &v->dummyinteger2, |
| 2354 | &v->PixelPTEReqWidthC[k], |
| 2355 | &v->PixelPTEReqHeightC[k], |
| 2356 | &v->PTERequestSizeC[k], |
| 2357 | &v->dpde0_bytes_per_frame_ub_c[k], |
| 2358 | &v->meta_pte_bytes_per_frame_ub_c[k]); |
| 2359 | |
| 2360 | v->PrefetchSourceLinesC[k] = CalculatePrefetchSourceLines( |
| 2361 | mode_lib, |
| 2362 | v->VRatioChroma[k], |
| 2363 | v->VTAPsChroma[k], |
| 2364 | v->Interlace[k], |
| 2365 | v->ProgressiveToInterlaceUnitInOPP, |
| 2366 | v->SwathHeightC[k], |
| 2367 | v->ViewportYStartC[k], |
| 2368 | &v->VInitPreFillC[k], |
| 2369 | &v->MaxNumSwathC[k]); |
| 2370 | } else { |
| 2371 | v->PTEBufferSizeInRequestsForLuma = v->PTEBufferSizeInRequestsLuma + v->PTEBufferSizeInRequestsChroma; |
| 2372 | v->PTEBufferSizeInRequestsForChroma = 0; |
| 2373 | PixelPTEBytesPerRowC = 0; |
| 2374 | PDEAndMetaPTEBytesFrameC = 0; |
| 2375 | MetaRowByteC = 0; |
| 2376 | v->MaxNumSwathC[k] = 0; |
| 2377 | v->PrefetchSourceLinesC[k] = 0; |
| 2378 | } |
| 2379 | |
| 2380 | PDEAndMetaPTEBytesFrameY = CalculateVMAndRowBytes( |
| 2381 | mode_lib, |
| 2382 | v->DCCEnable[k], |
| 2383 | v->BlockHeight256BytesY[k], |
| 2384 | v->BlockWidth256BytesY[k], |
| 2385 | v->SourcePixelFormat[k], |
| 2386 | v->SurfaceTiling[k], |
| 2387 | v->BytePerPixelY[k], |
| 2388 | v->SourceScan[k], |
| 2389 | v->SwathWidthY[k], |
| 2390 | v->ViewportHeight[k], |
| 2391 | v->GPUVMEnable, |
| 2392 | v->HostVMEnable, |
| 2393 | v->HostVMMaxNonCachedPageTableLevels, |
| 2394 | v->GPUVMMinPageSize, |
| 2395 | v->HostVMMinPageSize, |
| 2396 | v->PTEBufferSizeInRequestsForLuma, |
| 2397 | v->PitchY[k], |
| 2398 | v->DCCMetaPitchY[k], |
| 2399 | &v->MacroTileWidthY[k], |
| 2400 | &MetaRowByteY, |
| 2401 | &PixelPTEBytesPerRowY, |
| 2402 | &PTEBufferSizeNotExceededY, |
| 2403 | &v->dpte_row_width_luma_ub[k], |
| 2404 | &v->dpte_row_height[k], |
| 2405 | &v->meta_req_width[k], |
| 2406 | &v->meta_req_height[k], |
| 2407 | &v->meta_row_width[k], |
| 2408 | &v->meta_row_height[k], |
| 2409 | &v->vm_group_bytes[k], |
| 2410 | &v->dpte_group_bytes[k], |
| 2411 | &v->PixelPTEReqWidthY[k], |
| 2412 | &v->PixelPTEReqHeightY[k], |
| 2413 | &v->PTERequestSizeY[k], |
| 2414 | &v->dpde0_bytes_per_frame_ub_l[k], |
| 2415 | &v->meta_pte_bytes_per_frame_ub_l[k]); |
| 2416 | |
| 2417 | v->PrefetchSourceLinesY[k] = CalculatePrefetchSourceLines( |
| 2418 | mode_lib, |
| 2419 | v->VRatio[k], |
| 2420 | v->vtaps[k], |
| 2421 | v->Interlace[k], |
| 2422 | v->ProgressiveToInterlaceUnitInOPP, |
| 2423 | v->SwathHeightY[k], |
| 2424 | v->ViewportYStartY[k], |
| 2425 | &v->VInitPreFillY[k], |
| 2426 | &v->MaxNumSwathY[k]); |
| 2427 | v->PixelPTEBytesPerRow[k] = PixelPTEBytesPerRowY + PixelPTEBytesPerRowC; |
| 2428 | v->PDEAndMetaPTEBytesFrame[k] = PDEAndMetaPTEBytesFrameY + PDEAndMetaPTEBytesFrameC; |
| 2429 | v->MetaRowByte[k] = MetaRowByteY + MetaRowByteC; |
| 2430 | |
| 2431 | CalculateRowBandwidth( |
| 2432 | v->GPUVMEnable, |
| 2433 | v->SourcePixelFormat[k], |
| 2434 | v->VRatio[k], |
| 2435 | v->VRatioChroma[k], |
| 2436 | v->DCCEnable[k], |
| 2437 | v->HTotal[k] / v->PixelClock[k], |
| 2438 | MetaRowByteY, |
| 2439 | MetaRowByteC, |
| 2440 | v->meta_row_height[k], |
| 2441 | v->meta_row_height_chroma[k], |
| 2442 | PixelPTEBytesPerRowY, |
| 2443 | PixelPTEBytesPerRowC, |
| 2444 | v->dpte_row_height[k], |
| 2445 | v->dpte_row_height_chroma[k], |
| 2446 | &v->meta_row_bw[k], |
| 2447 | &v->dpte_row_bw[k]); |
| 2448 | } |
| 2449 | |
| 2450 | v->TotalDCCActiveDPP = 0; |
| 2451 | v->TotalActiveDPP = 0; |
| 2452 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 2453 | v->TotalActiveDPP = v->TotalActiveDPP + v->DPPPerPlane[k]; |
| 2454 | if (v->DCCEnable[k]) |
| 2455 | v->TotalDCCActiveDPP = v->TotalDCCActiveDPP + v->DPPPerPlane[k]; |
| 2456 | if (v->SourcePixelFormat[k] == dm_420_8 || v->SourcePixelFormat[k] == dm_420_10 || v->SourcePixelFormat[k] == dm_420_12 |
| 2457 | || v->SourcePixelFormat[k] == dm_rgbe_alpha) |
| 2458 | NoChromaPlanes = false0; |
| 2459 | } |
| 2460 | |
| 2461 | ReorderBytes = v->NumberOfChannels |
| 2462 | * dml_max3( |
| 2463 | v->UrgentOutOfOrderReturnPerChannelPixelDataOnly, |
| 2464 | v->UrgentOutOfOrderReturnPerChannelPixelMixedWithVMData, |
| 2465 | v->UrgentOutOfOrderReturnPerChannelVMDataOnly); |
| 2466 | |
| 2467 | VMDataOnlyReturnBW = dml_min( |
| 2468 | dml_min(v->ReturnBusWidth * v->DCFCLK, v->FabricClock * v->FabricDatapathToDCNDataReturn) |
| 2469 | * v->PercentOfIdealFabricAndSDPPortBWReceivedAfterUrgLatency / 100.0, |
| 2470 | v->DRAMSpeed * v->NumberOfChannels * v->DRAMChannelWidth |
| 2471 | * v->PercentOfIdealDRAMBWReceivedAfterUrgLatencyVMDataOnly / 100.0); |
| 2472 | |
| 2473 | #ifdef __DML_VBA_DEBUG__ |
| 2474 | dml_print("DML::%s: v->ReturnBusWidth = %f\n", __func__, v->ReturnBusWidth){do { } while(0); }; |
| 2475 | dml_print("DML::%s: v->DCFCLK = %f\n", __func__, v->DCFCLK){do { } while(0); }; |
| 2476 | dml_print("DML::%s: v->FabricClock = %f\n", __func__, v->FabricClock){do { } while(0); }; |
| 2477 | dml_print("DML::%s: v->FabricDatapathToDCNDataReturn = %f\n", __func__, v->FabricDatapathToDCNDataReturn){do { } while(0); }; |
| 2478 | dml_print("DML::%s: v->PercentOfIdealFabricAndSDPPortBWReceivedAfterUrgLatency = %f\n", __func__, v->PercentOfIdealFabricAndSDPPortBWReceivedAfterUrgLatency){do { } while(0); }; |
| 2479 | dml_print("DML::%s: v->DRAMSpeed = %f\n", __func__, v->DRAMSpeed){do { } while(0); }; |
| 2480 | dml_print("DML::%s: v->NumberOfChannels = %f\n", __func__, v->NumberOfChannels){do { } while(0); }; |
| 2481 | dml_print("DML::%s: v->DRAMChannelWidth = %f\n", __func__, v->DRAMChannelWidth){do { } while(0); }; |
| 2482 | dml_print("DML::%s: v->PercentOfIdealDRAMBWReceivedAfterUrgLatencyVMDataOnly = %f\n", __func__, v->PercentOfIdealDRAMBWReceivedAfterUrgLatencyVMDataOnly){do { } while(0); }; |
| 2483 | dml_print("DML::%s: VMDataOnlyReturnBW = %f\n", __func__, VMDataOnlyReturnBW){do { } while(0); }; |
| 2484 | dml_print("DML::%s: ReturnBW = %f\n", __func__, v->ReturnBW){do { } while(0); }; |
| 2485 | #endif |
| 2486 | |
| 2487 | if (v->GPUVMEnable && v->HostVMEnable) |
| 2488 | HostVMInefficiencyFactor = v->ReturnBW / VMDataOnlyReturnBW; |
| 2489 | |
| 2490 | v->UrgentExtraLatency = CalculateExtraLatency( |
| 2491 | v->RoundTripPingLatencyCycles, |
| 2492 | ReorderBytes, |
| 2493 | v->DCFCLK, |
| 2494 | v->TotalActiveDPP, |
| 2495 | v->PixelChunkSizeInKByte, |
| 2496 | v->TotalDCCActiveDPP, |
| 2497 | v->MetaChunkSize, |
| 2498 | v->ReturnBW, |
| 2499 | v->GPUVMEnable, |
| 2500 | v->HostVMEnable, |
| 2501 | v->NumberOfActivePlanes, |
| 2502 | v->DPPPerPlane, |
| 2503 | v->dpte_group_bytes, |
| 2504 | HostVMInefficiencyFactor, |
| 2505 | v->HostVMMinPageSize, |
| 2506 | v->HostVMMaxNonCachedPageTableLevels); |
| 2507 | |
| 2508 | v->TCalc = 24.0 / v->DCFCLKDeepSleep; |
| 2509 | |
| 2510 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 2511 | if (v->BlendingAndTiming[k] == k) { |
| 2512 | if (v->WritebackEnable[k] == true1) { |
| 2513 | v->WritebackDelay[v->VoltageLevel][k] = v->WritebackLatency |
| 2514 | + CalculateWriteBackDelay( |
| 2515 | v->WritebackPixelFormat[k], |
| 2516 | v->WritebackHRatio[k], |
| 2517 | v->WritebackVRatio[k], |
| 2518 | v->WritebackVTaps[k], |
| 2519 | v->WritebackDestinationWidth[k], |
| 2520 | v->WritebackDestinationHeight[k], |
| 2521 | v->WritebackSourceHeight[k], |
| 2522 | v->HTotal[k]) / v->DISPCLK; |
| 2523 | } else |
| 2524 | v->WritebackDelay[v->VoltageLevel][k] = 0; |
| 2525 | for (j = 0; j < v->NumberOfActivePlanes; ++j) { |
| 2526 | if (v->BlendingAndTiming[j] == k && v->WritebackEnable[j] == true1) { |
| 2527 | v->WritebackDelay[v->VoltageLevel][k] = dml_max( |
| 2528 | v->WritebackDelay[v->VoltageLevel][k], |
| 2529 | v->WritebackLatency |
| 2530 | + CalculateWriteBackDelay( |
| 2531 | v->WritebackPixelFormat[j], |
| 2532 | v->WritebackHRatio[j], |
| 2533 | v->WritebackVRatio[j], |
| 2534 | v->WritebackVTaps[j], |
| 2535 | v->WritebackDestinationWidth[j], |
| 2536 | v->WritebackDestinationHeight[j], |
| 2537 | v->WritebackSourceHeight[j], |
| 2538 | v->HTotal[k]) / v->DISPCLK); |
| 2539 | } |
| 2540 | } |
| 2541 | } |
| 2542 | } |
| 2543 | |
| 2544 | for (k = 0; k < v->NumberOfActivePlanes; ++k) |
| 2545 | for (j = 0; j < v->NumberOfActivePlanes; ++j) |
| 2546 | if (v->BlendingAndTiming[k] == j) |
| 2547 | v->WritebackDelay[v->VoltageLevel][k] = v->WritebackDelay[v->VoltageLevel][j]; |
| 2548 | |
| 2549 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 2550 | v->MaxVStartupLines[k] = |
| 2551 | (v->Interlace[k] && !v->ProgressiveToInterlaceUnitInOPP) ? |
| 2552 | dml_floor((v->VTotal[k] - v->VActive[k]) / 2.0, 1.0) : |
| 2553 | v->VTotal[k] - v->VActive[k] |
| 2554 | - dml_max( |
| 2555 | 1.0, |
| 2556 | dml_ceil( |
| 2557 | (double) v->WritebackDelay[v->VoltageLevel][k] |
| 2558 | / (v->HTotal[k] / v->PixelClock[k]), |
| 2559 | 1)); |
| 2560 | if (v->MaxVStartupLines[k] > 1023) |
| 2561 | v->MaxVStartupLines[k] = 1023; |
| 2562 | |
| 2563 | #ifdef __DML_VBA_DEBUG__ |
| 2564 | dml_print("DML::%s: k=%d MaxVStartupLines = %d\n", __func__, k, v->MaxVStartupLines[k]){do { } while(0); }; |
| 2565 | dml_print("DML::%s: k=%d VoltageLevel = %d\n", __func__, k, v->VoltageLevel){do { } while(0); }; |
| 2566 | dml_print("DML::%s: k=%d WritebackDelay = %f\n", __func__, k, v->WritebackDelay[v->VoltageLevel][k]){do { } while(0); }; |
| 2567 | #endif |
| 2568 | } |
| 2569 | |
| 2570 | v->MaximumMaxVStartupLines = 0; |
| 2571 | for (k = 0; k < v->NumberOfActivePlanes; ++k) |
| 2572 | v->MaximumMaxVStartupLines = dml_max(v->MaximumMaxVStartupLines, v->MaxVStartupLines[k]); |
| 2573 | |
| 2574 | // VBA_DELTA |
| 2575 | // We don't really care to iterate between the various prefetch modes |
| 2576 | //v->PrefetchERROR = CalculateMinAndMaxPrefetchMode(v->AllowDRAMSelfRefreshOrDRAMClockChangeInVblank, &v->MinPrefetchMode, &v->MaxPrefetchMode); |
| 2577 | |
| 2578 | v->UrgentLatency = CalculateUrgentLatency( |
| 2579 | v->UrgentLatencyPixelDataOnly, |
| 2580 | v->UrgentLatencyPixelMixedWithVMData, |
| 2581 | v->UrgentLatencyVMDataOnly, |
| 2582 | v->DoUrgentLatencyAdjustment, |
| 2583 | v->UrgentLatencyAdjustmentFabricClockComponent, |
| 2584 | v->UrgentLatencyAdjustmentFabricClockReference, |
| 2585 | v->FabricClock); |
| 2586 | |
| 2587 | v->FractionOfUrgentBandwidth = 0.0; |
| 2588 | v->FractionOfUrgentBandwidthImmediateFlip = 0.0; |
| 2589 | |
| 2590 | v->VStartupLines = __DML_VBA_MIN_VSTARTUP__9; |
| 2591 | |
| 2592 | do { |
| 2593 | double MaxTotalRDBandwidthNoUrgentBurst = 0.0; |
| 2594 | bool_Bool DestinationLineTimesForPrefetchLessThan2 = false0; |
| 2595 | bool_Bool VRatioPrefetchMoreThan4 = false0; |
| 2596 | double TWait = CalculateTWait(PrefetchMode, v->DRAMClockChangeLatency, v->UrgentLatency, v->SREnterPlusExitTime); |
| 2597 | MaxTotalRDBandwidth = 0; |
| 2598 | |
| 2599 | dml_print("DML::%s: Start loop: VStartup = %d\n", __func__, v->VStartupLines){do { } while(0); }; |
| 2600 | |
| 2601 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 2602 | Pipe myPipe; |
| 2603 | |
| 2604 | myPipe.DPPCLK = v->DPPCLK[k]; |
| 2605 | myPipe.DISPCLK = v->DISPCLK; |
| 2606 | myPipe.PixelClock = v->PixelClock[k]; |
| 2607 | myPipe.DCFCLKDeepSleep = v->DCFCLKDeepSleep; |
| 2608 | myPipe.DPPPerPlane = v->DPPPerPlane[k]; |
| 2609 | myPipe.ScalerEnabled = v->ScalerEnabled[k]; |
| 2610 | myPipe.VRatio = v->VRatio[k]; |
| 2611 | myPipe.VRatioChroma = v->VRatioChroma[k]; |
| 2612 | myPipe.SourceScan = v->SourceScan[k]; |
| 2613 | myPipe.BlockWidth256BytesY = v->BlockWidth256BytesY[k]; |
| 2614 | myPipe.BlockHeight256BytesY = v->BlockHeight256BytesY[k]; |
| 2615 | myPipe.BlockWidth256BytesC = v->BlockWidth256BytesC[k]; |
| 2616 | myPipe.BlockHeight256BytesC = v->BlockHeight256BytesC[k]; |
| 2617 | myPipe.InterlaceEnable = v->Interlace[k]; |
| 2618 | myPipe.NumberOfCursors = v->NumberOfCursors[k]; |
| 2619 | myPipe.VBlank = v->VTotal[k] - v->VActive[k]; |
| 2620 | myPipe.HTotal = v->HTotal[k]; |
| 2621 | myPipe.DCCEnable = v->DCCEnable[k]; |
| 2622 | myPipe.ODMCombineIsEnabled = v->ODMCombineEnabled[k] == dm_odm_combine_mode_4to1 |
| 2623 | || v->ODMCombineEnabled[k] == dm_odm_combine_mode_2to1; |
| 2624 | myPipe.SourcePixelFormat = v->SourcePixelFormat[k]; |
| 2625 | myPipe.BytePerPixelY = v->BytePerPixelY[k]; |
| 2626 | myPipe.BytePerPixelC = v->BytePerPixelC[k]; |
| 2627 | myPipe.ProgressiveToInterlaceUnitInOPP = v->ProgressiveToInterlaceUnitInOPP; |
| 2628 | v->ErrorResult[k] = CalculatePrefetchSchedule( |
| 2629 | mode_lib, |
| 2630 | HostVMInefficiencyFactor, |
| 2631 | &myPipe, |
| 2632 | v->DSCDelay[k], |
| 2633 | v->DPPCLKDelaySubtotal + v->DPPCLKDelayCNVCFormater, |
| 2634 | v->DPPCLKDelaySCL, |
| 2635 | v->DPPCLKDelaySCLLBOnly, |
| 2636 | v->DPPCLKDelayCNVCCursor, |
| 2637 | v->DISPCLKDelaySubtotal, |
| 2638 | (unsigned int) (v->SwathWidthY[k] / v->HRatio[k]), |
| 2639 | v->OutputFormat[k], |
| 2640 | v->MaxInterDCNTileRepeaters, |
| 2641 | dml_min(v->VStartupLines, v->MaxVStartupLines[k]), |
| 2642 | v->MaxVStartupLines[k], |
| 2643 | v->GPUVMMaxPageTableLevels, |
| 2644 | v->GPUVMEnable, |
| 2645 | v->HostVMEnable, |
| 2646 | v->HostVMMaxNonCachedPageTableLevels, |
| 2647 | v->HostVMMinPageSize, |
| 2648 | v->DynamicMetadataEnable[k], |
| 2649 | v->DynamicMetadataVMEnabled, |
| 2650 | v->DynamicMetadataLinesBeforeActiveRequired[k], |
| 2651 | v->DynamicMetadataTransmittedBytes[k], |
| 2652 | v->UrgentLatency, |
| 2653 | v->UrgentExtraLatency, |
| 2654 | v->TCalc, |
| 2655 | v->PDEAndMetaPTEBytesFrame[k], |
| 2656 | v->MetaRowByte[k], |
| 2657 | v->PixelPTEBytesPerRow[k], |
| 2658 | v->PrefetchSourceLinesY[k], |
| 2659 | v->SwathWidthY[k], |
| 2660 | v->VInitPreFillY[k], |
| 2661 | v->MaxNumSwathY[k], |
| 2662 | v->PrefetchSourceLinesC[k], |
| 2663 | v->SwathWidthC[k], |
| 2664 | v->VInitPreFillC[k], |
| 2665 | v->MaxNumSwathC[k], |
| 2666 | v->swath_width_luma_ub[k], |
| 2667 | v->swath_width_chroma_ub[k], |
| 2668 | v->SwathHeightY[k], |
| 2669 | v->SwathHeightC[k], |
| 2670 | TWait, |
| 2671 | &v->DSTXAfterScaler[k], |
| 2672 | &v->DSTYAfterScaler[k], |
| 2673 | &v->DestinationLinesForPrefetch[k], |
| 2674 | &v->PrefetchBandwidth[k], |
| 2675 | &v->DestinationLinesToRequestVMInVBlank[k], |
| 2676 | &v->DestinationLinesToRequestRowInVBlank[k], |
| 2677 | &v->VRatioPrefetchY[k], |
| 2678 | &v->VRatioPrefetchC[k], |
| 2679 | &v->RequiredPrefetchPixDataBWLuma[k], |
| 2680 | &v->RequiredPrefetchPixDataBWChroma[k], |
| 2681 | &v->NotEnoughTimeForDynamicMetadata[k], |
| 2682 | &v->Tno_bw[k], |
| 2683 | &v->prefetch_vmrow_bw[k], |
| 2684 | &v->Tdmdl_vm[k], |
| 2685 | &v->Tdmdl[k], |
| 2686 | &v->TSetup[k], |
| 2687 | &v->VUpdateOffsetPix[k], |
| 2688 | &v->VUpdateWidthPix[k], |
| 2689 | &v->VReadyOffsetPix[k]); |
| 2690 | |
| 2691 | #ifdef __DML_VBA_DEBUG__ |
| 2692 | dml_print("DML::%s: k=%0d Prefetch cal result=%0d\n", __func__, k, v->ErrorResult[k]){do { } while(0); }; |
| 2693 | #endif |
| 2694 | v->VStartup[k] = dml_min(v->VStartupLines, v->MaxVStartupLines[k]); |
| 2695 | } |
| 2696 | |
| 2697 | v->NoEnoughUrgentLatencyHiding = false0; |
| 2698 | v->NoEnoughUrgentLatencyHidingPre = false0; |
| 2699 | |
| 2700 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 2701 | v->cursor_bw[k] = v->NumberOfCursors[k] * v->CursorWidth[k][0] * v->CursorBPP[k][0] / 8.0 |
| 2702 | / (v->HTotal[k] / v->PixelClock[k]) * v->VRatio[k]; |
| 2703 | v->cursor_bw_pre[k] = v->NumberOfCursors[k] * v->CursorWidth[k][0] * v->CursorBPP[k][0] / 8.0 |
| 2704 | / (v->HTotal[k] / v->PixelClock[k]) * v->VRatioPrefetchY[k]; |
| 2705 | |
| 2706 | CalculateUrgentBurstFactor( |
| 2707 | v->swath_width_luma_ub[k], |
| 2708 | v->swath_width_chroma_ub[k], |
| 2709 | v->SwathHeightY[k], |
| 2710 | v->SwathHeightC[k], |
| 2711 | v->HTotal[k] / v->PixelClock[k], |
| 2712 | v->UrgentLatency, |
| 2713 | v->CursorBufferSize, |
| 2714 | v->CursorWidth[k][0], |
| 2715 | v->CursorBPP[k][0], |
| 2716 | v->VRatio[k], |
| 2717 | v->VRatioChroma[k], |
| 2718 | v->BytePerPixelDETY[k], |
| 2719 | v->BytePerPixelDETC[k], |
| 2720 | v->DETBufferSizeY[k], |
| 2721 | v->DETBufferSizeC[k], |
| 2722 | &v->UrgBurstFactorCursor[k], |
| 2723 | &v->UrgBurstFactorLuma[k], |
| 2724 | &v->UrgBurstFactorChroma[k], |
| 2725 | &v->NoUrgentLatencyHiding[k]); |
| 2726 | |
| 2727 | CalculateUrgentBurstFactor( |
| 2728 | v->swath_width_luma_ub[k], |
| 2729 | v->swath_width_chroma_ub[k], |
| 2730 | v->SwathHeightY[k], |
| 2731 | v->SwathHeightC[k], |
| 2732 | v->HTotal[k] / v->PixelClock[k], |
| 2733 | v->UrgentLatency, |
| 2734 | v->CursorBufferSize, |
| 2735 | v->CursorWidth[k][0], |
| 2736 | v->CursorBPP[k][0], |
| 2737 | v->VRatioPrefetchY[k], |
| 2738 | v->VRatioPrefetchC[k], |
| 2739 | v->BytePerPixelDETY[k], |
| 2740 | v->BytePerPixelDETC[k], |
| 2741 | v->DETBufferSizeY[k], |
| 2742 | v->DETBufferSizeC[k], |
| 2743 | &v->UrgBurstFactorCursorPre[k], |
| 2744 | &v->UrgBurstFactorLumaPre[k], |
| 2745 | &v->UrgBurstFactorChromaPre[k], |
| 2746 | &v->NoUrgentLatencyHidingPre[k]); |
| 2747 | |
| 2748 | MaxTotalRDBandwidth = MaxTotalRDBandwidth |
| 2749 | + dml_max3( |
| 2750 | v->DPPPerPlane[k] * v->prefetch_vmrow_bw[k], |
| 2751 | v->ReadBandwidthPlaneLuma[k] * v->UrgBurstFactorLuma[k] |
| 2752 | + v->ReadBandwidthPlaneChroma[k] * v->UrgBurstFactorChroma[k] |
| 2753 | + v->cursor_bw[k] * v->UrgBurstFactorCursor[k] |
| 2754 | + v->DPPPerPlane[k] * (v->meta_row_bw[k] + v->dpte_row_bw[k]), |
| 2755 | v->DPPPerPlane[k] |
| 2756 | * (v->RequiredPrefetchPixDataBWLuma[k] * v->UrgBurstFactorLumaPre[k] |
| 2757 | + v->RequiredPrefetchPixDataBWChroma[k] * v->UrgBurstFactorChromaPre[k]) |
| 2758 | + v->cursor_bw_pre[k] * v->UrgBurstFactorCursorPre[k]); |
| 2759 | |
| 2760 | MaxTotalRDBandwidthNoUrgentBurst = MaxTotalRDBandwidthNoUrgentBurst |
| 2761 | + dml_max3( |
| 2762 | v->DPPPerPlane[k] * v->prefetch_vmrow_bw[k], |
| 2763 | v->ReadBandwidthPlaneLuma[k] + v->ReadBandwidthPlaneChroma[k] + v->cursor_bw[k] |
| 2764 | + v->DPPPerPlane[k] * (v->meta_row_bw[k] + v->dpte_row_bw[k]), |
| 2765 | v->DPPPerPlane[k] * (v->RequiredPrefetchPixDataBWLuma[k] + v->RequiredPrefetchPixDataBWChroma[k]) |
| 2766 | + v->cursor_bw_pre[k]); |
| 2767 | |
| 2768 | #ifdef __DML_VBA_DEBUG__ |
| 2769 | dml_print("DML::%s: k=%0d DPPPerPlane=%d\n", __func__, k, v->DPPPerPlane[k]){do { } while(0); }; |
| 2770 | dml_print("DML::%s: k=%0d UrgBurstFactorLuma=%f\n", __func__, k, v->UrgBurstFactorLuma[k]){do { } while(0); }; |
| 2771 | dml_print("DML::%s: k=%0d UrgBurstFactorChroma=%f\n", __func__, k, v->UrgBurstFactorChroma[k]){do { } while(0); }; |
| 2772 | dml_print("DML::%s: k=%0d UrgBurstFactorLumaPre=%f\n", __func__, k, v->UrgBurstFactorLumaPre[k]){do { } while(0); }; |
| 2773 | dml_print("DML::%s: k=%0d UrgBurstFactorChromaPre=%f\n", __func__, k, v->UrgBurstFactorChromaPre[k]){do { } while(0); }; |
| 2774 | |
| 2775 | dml_print("DML::%s: k=%0d VRatioPrefetchY=%f\n", __func__, k, v->VRatioPrefetchY[k]){do { } while(0); }; |
| 2776 | dml_print("DML::%s: k=%0d VRatioY=%f\n", __func__, k, v->VRatio[k]){do { } while(0); }; |
| 2777 | |
| 2778 | dml_print("DML::%s: k=%0d prefetch_vmrow_bw=%f\n", __func__, k, v->prefetch_vmrow_bw[k]){do { } while(0); }; |
| 2779 | dml_print("DML::%s: k=%0d ReadBandwidthPlaneLuma=%f\n", __func__, k, v->ReadBandwidthPlaneLuma[k]){do { } while(0); }; |
| 2780 | dml_print("DML::%s: k=%0d ReadBandwidthPlaneChroma=%f\n", __func__, k, v->ReadBandwidthPlaneChroma[k]){do { } while(0); }; |
| 2781 | dml_print("DML::%s: k=%0d cursor_bw=%f\n", __func__, k, v->cursor_bw[k]){do { } while(0); }; |
| 2782 | dml_print("DML::%s: k=%0d meta_row_bw=%f\n", __func__, k, v->meta_row_bw[k]){do { } while(0); }; |
| 2783 | dml_print("DML::%s: k=%0d dpte_row_bw=%f\n", __func__, k, v->dpte_row_bw[k]){do { } while(0); }; |
| 2784 | dml_print("DML::%s: k=%0d RequiredPrefetchPixDataBWLuma=%f\n", __func__, k, v->RequiredPrefetchPixDataBWLuma[k]){do { } while(0); }; |
| 2785 | dml_print("DML::%s: k=%0d RequiredPrefetchPixDataBWChroma=%f\n", __func__, k, v->RequiredPrefetchPixDataBWChroma[k]){do { } while(0); }; |
| 2786 | dml_print("DML::%s: k=%0d cursor_bw_pre=%f\n", __func__, k, v->cursor_bw_pre[k]){do { } while(0); }; |
| 2787 | dml_print("DML::%s: k=%0d MaxTotalRDBandwidthNoUrgentBurst=%f\n", __func__, k, MaxTotalRDBandwidthNoUrgentBurst){do { } while(0); }; |
| 2788 | #endif |
| 2789 | |
| 2790 | if (v->DestinationLinesForPrefetch[k] < 2) |
| 2791 | DestinationLineTimesForPrefetchLessThan2 = true1; |
| 2792 | |
| 2793 | if (v->VRatioPrefetchY[k] > 4 || v->VRatioPrefetchC[k] > 4) |
| 2794 | VRatioPrefetchMoreThan4 = true1; |
| 2795 | |
| 2796 | if (v->NoUrgentLatencyHiding[k] == true1) |
| 2797 | v->NoEnoughUrgentLatencyHiding = true1; |
| 2798 | |
| 2799 | if (v->NoUrgentLatencyHidingPre[k] == true1) |
| 2800 | v->NoEnoughUrgentLatencyHidingPre = true1; |
| 2801 | } |
| 2802 | |
| 2803 | v->FractionOfUrgentBandwidth = MaxTotalRDBandwidthNoUrgentBurst / v->ReturnBW; |
| 2804 | |
| 2805 | #ifdef __DML_VBA_DEBUG__ |
| 2806 | dml_print("DML::%s: MaxTotalRDBandwidthNoUrgentBurst=%f \n", __func__, MaxTotalRDBandwidthNoUrgentBurst){do { } while(0); }; |
| 2807 | dml_print("DML::%s: ReturnBW=%f \n", __func__, v->ReturnBW){do { } while(0); }; |
| 2808 | dml_print("DML::%s: FractionOfUrgentBandwidth=%f \n", __func__, v->FractionOfUrgentBandwidth){do { } while(0); }; |
| 2809 | #endif |
| 2810 | |
| 2811 | if (MaxTotalRDBandwidth <= v->ReturnBW && v->NoEnoughUrgentLatencyHiding == 0 && v->NoEnoughUrgentLatencyHidingPre == 0 |
| 2812 | && !VRatioPrefetchMoreThan4 && !DestinationLineTimesForPrefetchLessThan2) |
| 2813 | v->PrefetchModeSupported = true1; |
| 2814 | else { |
| 2815 | v->PrefetchModeSupported = false0; |
| 2816 | dml_print("DML::%s: ***failed***. Bandwidth violation. Results are NOT valid\n", __func__){do { } while(0); }; |
| 2817 | dml_print("DML::%s: MaxTotalRDBandwidth:%f AvailReturnBandwidth:%f\n", __func__, MaxTotalRDBandwidth, v->ReturnBW){do { } while(0); }; |
| 2818 | dml_print("DML::%s: VRatioPrefetch %s more than 4\n", __func__, (VRatioPrefetchMoreThan4) ? "is" : "is not"){do { } while(0); }; |
| 2819 | dml_print("DML::%s: DestinationLines for Prefetch %s less than 2\n", __func__, (DestinationLineTimesForPrefetchLessThan2) ? "is" : "is not"){do { } while(0); }; |
| 2820 | } |
| 2821 | |
| 2822 | // PREVIOUS_ERROR |
| 2823 | // This error result check was done after the PrefetchModeSupported. So we will |
| 2824 | // still try to calculate flip schedule even prefetch mode not supported |
| 2825 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 2826 | if (v->ErrorResult[k] == true1 || v->NotEnoughTimeForDynamicMetadata[k] == true1) { |
| 2827 | v->PrefetchModeSupported = false0; |
| 2828 | dml_print("DML::%s: ***failed***. Prefetch schedule violation. Results are NOT valid\n", __func__){do { } while(0); }; |
| 2829 | } |
| 2830 | } |
| 2831 | |
| 2832 | if (v->PrefetchModeSupported == true1 && v->ImmediateFlipSupport == true1) { |
| 2833 | v->BandwidthAvailableForImmediateFlip = v->ReturnBW; |
| 2834 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 2835 | v->BandwidthAvailableForImmediateFlip = v->BandwidthAvailableForImmediateFlip |
| 2836 | - dml_max( |
| 2837 | v->ReadBandwidthPlaneLuma[k] * v->UrgBurstFactorLuma[k] |
| 2838 | + v->ReadBandwidthPlaneChroma[k] * v->UrgBurstFactorChroma[k] |
| 2839 | + v->cursor_bw[k] * v->UrgBurstFactorCursor[k], |
| 2840 | v->DPPPerPlane[k] |
| 2841 | * (v->RequiredPrefetchPixDataBWLuma[k] * v->UrgBurstFactorLumaPre[k] |
| 2842 | + v->RequiredPrefetchPixDataBWChroma[k] * v->UrgBurstFactorChromaPre[k]) |
| 2843 | + v->cursor_bw_pre[k] * v->UrgBurstFactorCursorPre[k]); |
| 2844 | } |
| 2845 | |
| 2846 | v->TotImmediateFlipBytes = 0; |
| 2847 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 2848 | v->TotImmediateFlipBytes = v->TotImmediateFlipBytes |
| 2849 | + v->DPPPerPlane[k] * (v->PDEAndMetaPTEBytesFrame[k] + v->MetaRowByte[k] + v->PixelPTEBytesPerRow[k]); |
| 2850 | } |
| 2851 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 2852 | CalculateFlipSchedule( |
| 2853 | mode_lib, |
| 2854 | k, |
| 2855 | HostVMInefficiencyFactor, |
| 2856 | v->UrgentExtraLatency, |
| 2857 | v->UrgentLatency, |
| 2858 | v->PDEAndMetaPTEBytesFrame[k], |
| 2859 | v->MetaRowByte[k], |
| 2860 | v->PixelPTEBytesPerRow[k]); |
| 2861 | } |
| 2862 | |
| 2863 | v->total_dcn_read_bw_with_flip = 0.0; |
| 2864 | v->total_dcn_read_bw_with_flip_no_urgent_burst = 0.0; |
| 2865 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 2866 | v->total_dcn_read_bw_with_flip = v->total_dcn_read_bw_with_flip |
| 2867 | + dml_max3( |
| 2868 | v->DPPPerPlane[k] * v->prefetch_vmrow_bw[k], |
| 2869 | v->DPPPerPlane[k] * v->final_flip_bw[k] |
| 2870 | + v->ReadBandwidthLuma[k] * v->UrgBurstFactorLuma[k] |
| 2871 | + v->ReadBandwidthChroma[k] * v->UrgBurstFactorChroma[k] |
| 2872 | + v->cursor_bw[k] * v->UrgBurstFactorCursor[k], |
| 2873 | v->DPPPerPlane[k] |
| 2874 | * (v->final_flip_bw[k] |
| 2875 | + v->RequiredPrefetchPixDataBWLuma[k] * v->UrgBurstFactorLumaPre[k] |
| 2876 | + v->RequiredPrefetchPixDataBWChroma[k] * v->UrgBurstFactorChromaPre[k]) |
| 2877 | + v->cursor_bw_pre[k] * v->UrgBurstFactorCursorPre[k]); |
| 2878 | v->total_dcn_read_bw_with_flip_no_urgent_burst = v->total_dcn_read_bw_with_flip_no_urgent_burst |
| 2879 | + dml_max3( |
| 2880 | v->DPPPerPlane[k] * v->prefetch_vmrow_bw[k], |
| 2881 | v->DPPPerPlane[k] * v->final_flip_bw[k] + v->ReadBandwidthPlaneLuma[k] |
| 2882 | + v->ReadBandwidthPlaneChroma[k] + v->cursor_bw[k], |
| 2883 | v->DPPPerPlane[k] |
| 2884 | * (v->final_flip_bw[k] + v->RequiredPrefetchPixDataBWLuma[k] |
| 2885 | + v->RequiredPrefetchPixDataBWChroma[k]) + v->cursor_bw_pre[k]); |
| 2886 | } |
| 2887 | v->FractionOfUrgentBandwidthImmediateFlip = v->total_dcn_read_bw_with_flip_no_urgent_burst / v->ReturnBW; |
| 2888 | |
| 2889 | v->ImmediateFlipSupported = true1; |
| 2890 | if (v->total_dcn_read_bw_with_flip > v->ReturnBW) { |
| 2891 | #ifdef __DML_VBA_DEBUG__ |
| 2892 | dml_print("DML::%s: total_dcn_read_bw_with_flip %f (bw w/ flip too high!)\n", __func__, v->total_dcn_read_bw_with_flip){do { } while(0); }; |
| 2893 | #endif |
| 2894 | v->ImmediateFlipSupported = false0; |
| 2895 | v->total_dcn_read_bw_with_flip = MaxTotalRDBandwidth; |
| 2896 | } |
| 2897 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 2898 | if (v->ImmediateFlipSupportedForPipe[k] == false0) { |
| 2899 | #ifdef __DML_VBA_DEBUG__ |
| 2900 | dml_print("DML::%s: Pipe %0d not supporting iflip\n",{do { } while(0); } |
| 2901 | __func__, k){do { } while(0); }; |
| 2902 | #endif |
| 2903 | v->ImmediateFlipSupported = false0; |
| 2904 | } |
| 2905 | } |
| 2906 | } else { |
| 2907 | v->ImmediateFlipSupported = false0; |
| 2908 | } |
| 2909 | |
| 2910 | v->PrefetchAndImmediateFlipSupported = |
| 2911 | (v->PrefetchModeSupported == true1 && ((!v->ImmediateFlipSupport && !v->HostVMEnable |
| 2912 | && v->ImmediateFlipRequirement[0] != dm_immediate_flip_required) || |
| 2913 | v->ImmediateFlipSupported)) ? true1 : false0; |
| 2914 | #ifdef __DML_VBA_DEBUG__ |
| 2915 | dml_print("DML::%s: PrefetchModeSupported %d\n", __func__, v->PrefetchModeSupported){do { } while(0); }; |
| 2916 | dml_print("DML::%s: ImmediateFlipRequirement[0] %d\n", __func__, v->ImmediateFlipRequirement[0] == dm_immediate_flip_required){do { } while(0); }; |
| 2917 | dml_print("DML::%s: ImmediateFlipSupported %d\n", __func__, v->ImmediateFlipSupported){do { } while(0); }; |
| 2918 | dml_print("DML::%s: ImmediateFlipSupport %d\n", __func__, v->ImmediateFlipSupport){do { } while(0); }; |
| 2919 | dml_print("DML::%s: HostVMEnable %d\n", __func__, v->HostVMEnable){do { } while(0); }; |
| 2920 | dml_print("DML::%s: PrefetchAndImmediateFlipSupported %d\n", __func__, v->PrefetchAndImmediateFlipSupported){do { } while(0); }; |
| 2921 | #endif |
| 2922 | dml_print("DML::%s: Done loop: Vstartup=%d, Max Vstartup is %d\n", __func__, v->VStartupLines, v->MaximumMaxVStartupLines){do { } while(0); }; |
| 2923 | |
| 2924 | v->VStartupLines = v->VStartupLines + 1; |
| 2925 | } while (!v->PrefetchAndImmediateFlipSupported && v->VStartupLines <= v->MaximumMaxVStartupLines); |
| 2926 | ASSERT(v->PrefetchAndImmediateFlipSupported)do { if (({ static int __warned; int __ret = !!(!(v->PrefetchAndImmediateFlipSupported )); if (__ret && !__warned) { printf("WARNING %s failed at %s:%d\n" , "!(v->PrefetchAndImmediateFlipSupported)", "/usr/src/sys/dev/pci/drm/amd/display/dc/dml/dcn31/display_mode_vba_31.c" , 2926); __warned = 1; } __builtin_expect(!!(__ret), 0); })) do {} while (0); } while (0); |
| 2927 | |
| 2928 | // Unbounded Request Enabled |
| 2929 | CalculateUnboundedRequestAndCompressedBufferSize( |
| 2930 | v->DETBufferSizeInKByte[0], |
| 2931 | v->ConfigReturnBufferSizeInKByte, |
| 2932 | v->UseUnboundedRequesting, |
| 2933 | v->TotalActiveDPP, |
| 2934 | NoChromaPlanes, |
| 2935 | v->MaxNumDPP, |
| 2936 | v->CompressedBufferSegmentSizeInkByte, |
| 2937 | v->Output, |
| 2938 | &v->UnboundedRequestEnabled, |
| 2939 | &v->CompressedBufferSizeInkByte); |
| 2940 | |
| 2941 | //Watermarks and NB P-State/DRAM Clock Change Support |
| 2942 | { |
| 2943 | enum clock_change_support DRAMClockChangeSupport; // dummy |
| 2944 | CalculateWatermarksAndDRAMSpeedChangeSupport( |
| 2945 | mode_lib, |
| 2946 | PrefetchMode, |
| 2947 | v->DCFCLK, |
| 2948 | v->ReturnBW, |
| 2949 | v->UrgentLatency, |
| 2950 | v->UrgentExtraLatency, |
| 2951 | v->SOCCLK, |
| 2952 | v->DCFCLKDeepSleep, |
| 2953 | v->DETBufferSizeY, |
| 2954 | v->DETBufferSizeC, |
| 2955 | v->SwathHeightY, |
| 2956 | v->SwathHeightC, |
| 2957 | v->SwathWidthY, |
| 2958 | v->SwathWidthC, |
| 2959 | v->DPPPerPlane, |
| 2960 | v->BytePerPixelDETY, |
| 2961 | v->BytePerPixelDETC, |
| 2962 | v->UnboundedRequestEnabled, |
| 2963 | v->CompressedBufferSizeInkByte, |
| 2964 | &DRAMClockChangeSupport, |
| 2965 | &v->StutterExitWatermark, |
| 2966 | &v->StutterEnterPlusExitWatermark, |
| 2967 | &v->Z8StutterExitWatermark, |
| 2968 | &v->Z8StutterEnterPlusExitWatermark); |
| 2969 | |
| 2970 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 2971 | if (v->WritebackEnable[k] == true1) { |
| 2972 | v->WritebackAllowDRAMClockChangeEndPosition[k] = dml_max( |
| 2973 | 0, |
| 2974 | v->VStartup[k] * v->HTotal[k] / v->PixelClock[k] - v->WritebackDRAMClockChangeWatermark); |
| 2975 | } else { |
| 2976 | v->WritebackAllowDRAMClockChangeEndPosition[k] = 0; |
| 2977 | } |
| 2978 | } |
| 2979 | } |
| 2980 | |
| 2981 | //Display Pipeline Delivery Time in Prefetch, Groups |
| 2982 | CalculatePixelDeliveryTimes( |
| 2983 | v->NumberOfActivePlanes, |
| 2984 | v->VRatio, |
| 2985 | v->VRatioChroma, |
| 2986 | v->VRatioPrefetchY, |
| 2987 | v->VRatioPrefetchC, |
| 2988 | v->swath_width_luma_ub, |
| 2989 | v->swath_width_chroma_ub, |
| 2990 | v->DPPPerPlane, |
| 2991 | v->HRatio, |
| 2992 | v->HRatioChroma, |
| 2993 | v->PixelClock, |
| 2994 | v->PSCL_THROUGHPUT_LUMA, |
| 2995 | v->PSCL_THROUGHPUT_CHROMA, |
| 2996 | v->DPPCLK, |
| 2997 | v->BytePerPixelC, |
| 2998 | v->SourceScan, |
| 2999 | v->NumberOfCursors, |
| 3000 | v->CursorWidth, |
| 3001 | v->CursorBPP, |
| 3002 | v->BlockWidth256BytesY, |
| 3003 | v->BlockHeight256BytesY, |
| 3004 | v->BlockWidth256BytesC, |
| 3005 | v->BlockHeight256BytesC, |
| 3006 | v->DisplayPipeLineDeliveryTimeLuma, |
| 3007 | v->DisplayPipeLineDeliveryTimeChroma, |
| 3008 | v->DisplayPipeLineDeliveryTimeLumaPrefetch, |
| 3009 | v->DisplayPipeLineDeliveryTimeChromaPrefetch, |
| 3010 | v->DisplayPipeRequestDeliveryTimeLuma, |
| 3011 | v->DisplayPipeRequestDeliveryTimeChroma, |
| 3012 | v->DisplayPipeRequestDeliveryTimeLumaPrefetch, |
| 3013 | v->DisplayPipeRequestDeliveryTimeChromaPrefetch, |
| 3014 | v->CursorRequestDeliveryTime, |
| 3015 | v->CursorRequestDeliveryTimePrefetch); |
| 3016 | |
| 3017 | CalculateMetaAndPTETimes( |
| 3018 | v->NumberOfActivePlanes, |
| 3019 | v->GPUVMEnable, |
| 3020 | v->MetaChunkSize, |
| 3021 | v->MinMetaChunkSizeBytes, |
| 3022 | v->HTotal, |
| 3023 | v->VRatio, |
| 3024 | v->VRatioChroma, |
| 3025 | v->DestinationLinesToRequestRowInVBlank, |
| 3026 | v->DestinationLinesToRequestRowInImmediateFlip, |
| 3027 | v->DCCEnable, |
| 3028 | v->PixelClock, |
| 3029 | v->BytePerPixelY, |
| 3030 | v->BytePerPixelC, |
| 3031 | v->SourceScan, |
| 3032 | v->dpte_row_height, |
| 3033 | v->dpte_row_height_chroma, |
| 3034 | v->meta_row_width, |
| 3035 | v->meta_row_width_chroma, |
| 3036 | v->meta_row_height, |
| 3037 | v->meta_row_height_chroma, |
| 3038 | v->meta_req_width, |
| 3039 | v->meta_req_width_chroma, |
| 3040 | v->meta_req_height, |
| 3041 | v->meta_req_height_chroma, |
| 3042 | v->dpte_group_bytes, |
| 3043 | v->PTERequestSizeY, |
| 3044 | v->PTERequestSizeC, |
| 3045 | v->PixelPTEReqWidthY, |
| 3046 | v->PixelPTEReqHeightY, |
| 3047 | v->PixelPTEReqWidthC, |
| 3048 | v->PixelPTEReqHeightC, |
| 3049 | v->dpte_row_width_luma_ub, |
| 3050 | v->dpte_row_width_chroma_ub, |
| 3051 | v->DST_Y_PER_PTE_ROW_NOM_L, |
| 3052 | v->DST_Y_PER_PTE_ROW_NOM_C, |
| 3053 | v->DST_Y_PER_META_ROW_NOM_L, |
| 3054 | v->DST_Y_PER_META_ROW_NOM_C, |
| 3055 | v->TimePerMetaChunkNominal, |
| 3056 | v->TimePerChromaMetaChunkNominal, |
| 3057 | v->TimePerMetaChunkVBlank, |
| 3058 | v->TimePerChromaMetaChunkVBlank, |
| 3059 | v->TimePerMetaChunkFlip, |
| 3060 | v->TimePerChromaMetaChunkFlip, |
| 3061 | v->time_per_pte_group_nom_luma, |
| 3062 | v->time_per_pte_group_vblank_luma, |
| 3063 | v->time_per_pte_group_flip_luma, |
| 3064 | v->time_per_pte_group_nom_chroma, |
| 3065 | v->time_per_pte_group_vblank_chroma, |
| 3066 | v->time_per_pte_group_flip_chroma); |
| 3067 | |
| 3068 | CalculateVMGroupAndRequestTimes( |
| 3069 | v->NumberOfActivePlanes, |
| 3070 | v->GPUVMEnable, |
| 3071 | v->GPUVMMaxPageTableLevels, |
| 3072 | v->HTotal, |
| 3073 | v->BytePerPixelC, |
| 3074 | v->DestinationLinesToRequestVMInVBlank, |
| 3075 | v->DestinationLinesToRequestVMInImmediateFlip, |
| 3076 | v->DCCEnable, |
| 3077 | v->PixelClock, |
| 3078 | v->dpte_row_width_luma_ub, |
| 3079 | v->dpte_row_width_chroma_ub, |
| 3080 | v->vm_group_bytes, |
| 3081 | v->dpde0_bytes_per_frame_ub_l, |
| 3082 | v->dpde0_bytes_per_frame_ub_c, |
| 3083 | v->meta_pte_bytes_per_frame_ub_l, |
| 3084 | v->meta_pte_bytes_per_frame_ub_c, |
| 3085 | v->TimePerVMGroupVBlank, |
| 3086 | v->TimePerVMGroupFlip, |
| 3087 | v->TimePerVMRequestVBlank, |
| 3088 | v->TimePerVMRequestFlip); |
| 3089 | |
| 3090 | // Min TTUVBlank |
| 3091 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 3092 | if (PrefetchMode == 0) { |
| 3093 | v->AllowDRAMClockChangeDuringVBlank[k] = true1; |
| 3094 | v->AllowDRAMSelfRefreshDuringVBlank[k] = true1; |
| 3095 | v->MinTTUVBlank[k] = dml_max( |
| 3096 | v->DRAMClockChangeWatermark, |
| 3097 | dml_max(v->StutterEnterPlusExitWatermark, v->UrgentWatermark)); |
| 3098 | } else if (PrefetchMode == 1) { |
| 3099 | v->AllowDRAMClockChangeDuringVBlank[k] = false0; |
| 3100 | v->AllowDRAMSelfRefreshDuringVBlank[k] = true1; |
| 3101 | v->MinTTUVBlank[k] = dml_max(v->StutterEnterPlusExitWatermark, v->UrgentWatermark); |
| 3102 | } else { |
| 3103 | v->AllowDRAMClockChangeDuringVBlank[k] = false0; |
| 3104 | v->AllowDRAMSelfRefreshDuringVBlank[k] = false0; |
| 3105 | v->MinTTUVBlank[k] = v->UrgentWatermark; |
| 3106 | } |
| 3107 | if (!v->DynamicMetadataEnable[k]) |
| 3108 | v->MinTTUVBlank[k] = v->TCalc + v->MinTTUVBlank[k]; |
| 3109 | } |
| 3110 | |
| 3111 | // DCC Configuration |
| 3112 | v->ActiveDPPs = 0; |
| 3113 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 3114 | CalculateDCCConfiguration(v->DCCEnable[k], false0, // We should always know the direction DCCProgrammingAssumesScanDirectionUnknown, |
| 3115 | v->SourcePixelFormat[k], |
| 3116 | v->SurfaceWidthY[k], |
| 3117 | v->SurfaceWidthC[k], |
| 3118 | v->SurfaceHeightY[k], |
| 3119 | v->SurfaceHeightC[k], |
| 3120 | v->DETBufferSizeInKByte[k] * 1024, |
| 3121 | v->BlockHeight256BytesY[k], |
| 3122 | v->BlockHeight256BytesC[k], |
| 3123 | v->SurfaceTiling[k], |
| 3124 | v->BytePerPixelY[k], |
| 3125 | v->BytePerPixelC[k], |
| 3126 | v->BytePerPixelDETY[k], |
| 3127 | v->BytePerPixelDETC[k], |
| 3128 | v->SourceScan[k], |
| 3129 | &v->DCCYMaxUncompressedBlock[k], |
| 3130 | &v->DCCCMaxUncompressedBlock[k], |
| 3131 | &v->DCCYMaxCompressedBlock[k], |
| 3132 | &v->DCCCMaxCompressedBlock[k], |
| 3133 | &v->DCCYIndependentBlock[k], |
| 3134 | &v->DCCCIndependentBlock[k]); |
| 3135 | } |
| 3136 | |
| 3137 | // VStartup Adjustment |
| 3138 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 3139 | bool_Bool isInterlaceTiming; |
| 3140 | double Tvstartup_margin = (v->MaxVStartupLines[k] - v->VStartup[k]) * v->HTotal[k] / v->PixelClock[k]; |
| 3141 | #ifdef __DML_VBA_DEBUG__ |
| 3142 | dml_print("DML::%s: k=%d, MinTTUVBlank = %f (before margin)\n", __func__, k, v->MinTTUVBlank[k]){do { } while(0); }; |
| 3143 | #endif |
| 3144 | |
| 3145 | v->MinTTUVBlank[k] = v->MinTTUVBlank[k] + Tvstartup_margin; |
| 3146 | |
| 3147 | #ifdef __DML_VBA_DEBUG__ |
| 3148 | dml_print("DML::%s: k=%d, Tvstartup_margin = %f\n", __func__, k, Tvstartup_margin){do { } while(0); }; |
| 3149 | dml_print("DML::%s: k=%d, MaxVStartupLines = %d\n", __func__, k, v->MaxVStartupLines[k]){do { } while(0); }; |
| 3150 | dml_print("DML::%s: k=%d, VStartup = %d\n", __func__, k, v->VStartup[k]){do { } while(0); }; |
| 3151 | dml_print("DML::%s: k=%d, MinTTUVBlank = %f\n", __func__, k, v->MinTTUVBlank[k]){do { } while(0); }; |
| 3152 | #endif |
| 3153 | |
| 3154 | v->Tdmdl[k] = v->Tdmdl[k] + Tvstartup_margin; |
| 3155 | if (v->DynamicMetadataEnable[k] && v->DynamicMetadataVMEnabled) { |
| 3156 | v->Tdmdl_vm[k] = v->Tdmdl_vm[k] + Tvstartup_margin; |
| 3157 | } |
| 3158 | |
| 3159 | isInterlaceTiming = (v->Interlace[k] && !v->ProgressiveToInterlaceUnitInOPP); |
| 3160 | |
| 3161 | v->MIN_DST_Y_NEXT_START[k] = ((isInterlaceTiming ? dml_floor((v->VTotal[k] - v->VFrontPorch[k]) / 2.0, 1.0) : v->VTotal[k]) |
| 3162 | - v->VFrontPorch[k]) |
| 3163 | + dml_max(1.0, dml_ceil(v->WritebackDelay[v->VoltageLevel][k] / (v->HTotal[k] / v->PixelClock[k]), 1.0)) |
| 3164 | + dml_floor(4.0 * v->TSetup[k] / (v->HTotal[k] / v->PixelClock[k]), 1.0) / 4.0; |
| 3165 | |
| 3166 | v->VStartup[k] = (isInterlaceTiming ? (2 * v->MaxVStartupLines[k]) : v->MaxVStartupLines[k]); |
| 3167 | |
| 3168 | if (((v->VUpdateOffsetPix[k] + v->VUpdateWidthPix[k] + v->VReadyOffsetPix[k]) / v->HTotal[k]) |
| 3169 | <= (isInterlaceTiming ? |
| 3170 | dml_floor((v->VTotal[k] - v->VActive[k] - v->VFrontPorch[k] - v->VStartup[k]) / 2.0, 1.0) : |
| 3171 | (int) (v->VTotal[k] - v->VActive[k] - v->VFrontPorch[k] - v->VStartup[k]))) { |
| 3172 | v->VREADY_AT_OR_AFTER_VSYNC[k] = true1; |
| 3173 | } else { |
| 3174 | v->VREADY_AT_OR_AFTER_VSYNC[k] = false0; |
| 3175 | } |
| 3176 | #ifdef __DML_VBA_DEBUG__ |
| 3177 | dml_print("DML::%s: k=%d, VStartup = %d (max)\n", __func__, k, v->VStartup[k]){do { } while(0); }; |
| 3178 | dml_print("DML::%s: k=%d, VUpdateOffsetPix = %d\n", __func__, k, v->VUpdateOffsetPix[k]){do { } while(0); }; |
| 3179 | dml_print("DML::%s: k=%d, VUpdateWidthPix = %d\n", __func__, k, v->VUpdateWidthPix[k]){do { } while(0); }; |
| 3180 | dml_print("DML::%s: k=%d, VReadyOffsetPix = %d\n", __func__, k, v->VReadyOffsetPix[k]){do { } while(0); }; |
| 3181 | dml_print("DML::%s: k=%d, HTotal = %d\n", __func__, k, v->HTotal[k]){do { } while(0); }; |
| 3182 | dml_print("DML::%s: k=%d, VTotal = %d\n", __func__, k, v->VTotal[k]){do { } while(0); }; |
| 3183 | dml_print("DML::%s: k=%d, VActive = %d\n", __func__, k, v->VActive[k]){do { } while(0); }; |
| 3184 | dml_print("DML::%s: k=%d, VFrontPorch = %d\n", __func__, k, v->VFrontPorch[k]){do { } while(0); }; |
| 3185 | dml_print("DML::%s: k=%d, VStartup = %d\n", __func__, k, v->VStartup[k]){do { } while(0); }; |
| 3186 | dml_print("DML::%s: k=%d, MIN_DST_Y_NEXT_START = %f\n", __func__, k, v->MIN_DST_Y_NEXT_START[k]){do { } while(0); }; |
| 3187 | dml_print("DML::%s: k=%d, VREADY_AT_OR_AFTER_VSYNC = %d\n", __func__, k, v->VREADY_AT_OR_AFTER_VSYNC[k]){do { } while(0); }; |
| 3188 | #endif |
| 3189 | } |
| 3190 | |
| 3191 | { |
| 3192 | //Maximum Bandwidth Used |
| 3193 | double TotalWRBandwidth = 0; |
| 3194 | double MaxPerPlaneVActiveWRBandwidth = 0; |
| 3195 | double WRBandwidth = 0; |
| 3196 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 3197 | if (v->WritebackEnable[k] == true1 && v->WritebackPixelFormat[k] == dm_444_32) { |
| 3198 | WRBandwidth = v->WritebackDestinationWidth[k] * v->WritebackDestinationHeight[k] |
| 3199 | / (v->HTotal[k] * v->WritebackSourceHeight[k] / v->PixelClock[k]) * 4; |
| 3200 | } else if (v->WritebackEnable[k] == true1) { |
| 3201 | WRBandwidth = v->WritebackDestinationWidth[k] * v->WritebackDestinationHeight[k] |
| 3202 | / (v->HTotal[k] * v->WritebackSourceHeight[k] / v->PixelClock[k]) * 8; |
| 3203 | } |
| 3204 | TotalWRBandwidth = TotalWRBandwidth + WRBandwidth; |
| 3205 | MaxPerPlaneVActiveWRBandwidth = dml_max(MaxPerPlaneVActiveWRBandwidth, WRBandwidth); |
| 3206 | } |
| 3207 | |
| 3208 | v->TotalDataReadBandwidth = 0; |
| 3209 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 3210 | v->TotalDataReadBandwidth = v->TotalDataReadBandwidth + v->ReadBandwidthPlaneLuma[k] + v->ReadBandwidthPlaneChroma[k]; |
| 3211 | } |
| 3212 | } |
| 3213 | // Stutter Efficiency |
| 3214 | CalculateStutterEfficiency( |
| 3215 | mode_lib, |
| 3216 | v->CompressedBufferSizeInkByte, |
| 3217 | v->UnboundedRequestEnabled, |
| 3218 | v->ConfigReturnBufferSizeInKByte, |
| 3219 | v->MetaFIFOSizeInKEntries, |
| 3220 | v->ZeroSizeBufferEntries, |
| 3221 | v->NumberOfActivePlanes, |
| 3222 | v->ROBBufferSizeInKByte, |
| 3223 | v->TotalDataReadBandwidth, |
| 3224 | v->DCFCLK, |
| 3225 | v->ReturnBW, |
| 3226 | v->COMPBUF_RESERVED_SPACE_64B, |
| 3227 | v->COMPBUF_RESERVED_SPACE_ZS, |
| 3228 | v->SRExitTime, |
| 3229 | v->SRExitZ8Time, |
| 3230 | v->SynchronizedVBlank, |
| 3231 | v->StutterEnterPlusExitWatermark, |
| 3232 | v->Z8StutterEnterPlusExitWatermark, |
| 3233 | v->ProgressiveToInterlaceUnitInOPP, |
| 3234 | v->Interlace, |
| 3235 | v->MinTTUVBlank, |
| 3236 | v->DPPPerPlane, |
| 3237 | v->DETBufferSizeY, |
| 3238 | v->BytePerPixelY, |
| 3239 | v->BytePerPixelDETY, |
| 3240 | v->SwathWidthY, |
| 3241 | v->SwathHeightY, |
| 3242 | v->SwathHeightC, |
| 3243 | v->DCCRateLuma, |
| 3244 | v->DCCRateChroma, |
| 3245 | v->DCCFractionOfZeroSizeRequestsLuma, |
| 3246 | v->DCCFractionOfZeroSizeRequestsChroma, |
| 3247 | v->HTotal, |
| 3248 | v->VTotal, |
| 3249 | v->PixelClock, |
| 3250 | v->VRatio, |
| 3251 | v->SourceScan, |
| 3252 | v->BlockHeight256BytesY, |
| 3253 | v->BlockWidth256BytesY, |
| 3254 | v->BlockHeight256BytesC, |
| 3255 | v->BlockWidth256BytesC, |
| 3256 | v->DCCYMaxUncompressedBlock, |
| 3257 | v->DCCCMaxUncompressedBlock, |
| 3258 | v->VActive, |
| 3259 | v->DCCEnable, |
| 3260 | v->WritebackEnable, |
| 3261 | v->ReadBandwidthPlaneLuma, |
| 3262 | v->ReadBandwidthPlaneChroma, |
| 3263 | v->meta_row_bw, |
| 3264 | v->dpte_row_bw, |
| 3265 | &v->StutterEfficiencyNotIncludingVBlank, |
| 3266 | &v->StutterEfficiency, |
| 3267 | &v->NumberOfStutterBurstsPerFrame, |
| 3268 | &v->Z8StutterEfficiencyNotIncludingVBlank, |
| 3269 | &v->Z8StutterEfficiency, |
| 3270 | &v->Z8NumberOfStutterBurstsPerFrame, |
| 3271 | &v->StutterPeriod); |
| 3272 | } |
| 3273 | |
| 3274 | static void DisplayPipeConfiguration(struct display_mode_lib *mode_lib) |
| 3275 | { |
| 3276 | struct vba_vars_st *v = &mode_lib->vba; |
| 3277 | // Display Pipe Configuration |
| 3278 | double BytePerPixDETY[DC__NUM_DPP__MAX8]; |
| 3279 | double BytePerPixDETC[DC__NUM_DPP__MAX8]; |
| 3280 | int BytePerPixY[DC__NUM_DPP__MAX8]; |
| 3281 | int BytePerPixC[DC__NUM_DPP__MAX8]; |
| 3282 | int Read256BytesBlockHeightY[DC__NUM_DPP__MAX8]; |
| 3283 | int Read256BytesBlockHeightC[DC__NUM_DPP__MAX8]; |
| 3284 | int Read256BytesBlockWidthY[DC__NUM_DPP__MAX8]; |
| 3285 | int Read256BytesBlockWidthC[DC__NUM_DPP__MAX8]; |
| 3286 | double dummy1[DC__NUM_DPP__MAX8]; |
| 3287 | double dummy2[DC__NUM_DPP__MAX8]; |
| 3288 | double dummy3[DC__NUM_DPP__MAX8]; |
| 3289 | double dummy4[DC__NUM_DPP__MAX8]; |
| 3290 | int dummy5[DC__NUM_DPP__MAX8]; |
| 3291 | int dummy6[DC__NUM_DPP__MAX8]; |
| 3292 | bool_Bool dummy7[DC__NUM_DPP__MAX8]; |
| 3293 | bool_Bool dummysinglestring; |
| 3294 | |
| 3295 | unsigned int k; |
| 3296 | |
| 3297 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 3298 | |
| 3299 | dml30_CalculateBytePerPixelAnd256BBlockSizes( |
| 3300 | v->SourcePixelFormat[k], |
| 3301 | v->SurfaceTiling[k], |
| 3302 | &BytePerPixY[k], |
| 3303 | &BytePerPixC[k], |
| 3304 | &BytePerPixDETY[k], |
| 3305 | &BytePerPixDETC[k], |
| 3306 | &Read256BytesBlockHeightY[k], |
| 3307 | &Read256BytesBlockHeightC[k], |
| 3308 | &Read256BytesBlockWidthY[k], |
| 3309 | &Read256BytesBlockWidthC[k]); |
| 3310 | } |
| 3311 | |
| 3312 | CalculateSwathAndDETConfiguration( |
| 3313 | false0, |
| 3314 | v->NumberOfActivePlanes, |
| 3315 | mode_lib->project == DML_PROJECT_DCN315 && v->DETSizeOverride[0], |
| 3316 | v->DETBufferSizeInKByte, |
| 3317 | dummy1, |
| 3318 | dummy2, |
| 3319 | v->SourceScan, |
| 3320 | v->SourcePixelFormat, |
| 3321 | v->SurfaceTiling, |
| 3322 | v->ViewportWidth, |
| 3323 | v->ViewportHeight, |
| 3324 | v->SurfaceWidthY, |
| 3325 | v->SurfaceWidthC, |
| 3326 | v->SurfaceHeightY, |
| 3327 | v->SurfaceHeightC, |
| 3328 | Read256BytesBlockHeightY, |
| 3329 | Read256BytesBlockHeightC, |
| 3330 | Read256BytesBlockWidthY, |
| 3331 | Read256BytesBlockWidthC, |
| 3332 | v->ODMCombineEnabled, |
| 3333 | v->BlendingAndTiming, |
| 3334 | BytePerPixY, |
| 3335 | BytePerPixC, |
| 3336 | BytePerPixDETY, |
| 3337 | BytePerPixDETC, |
| 3338 | v->HActive, |
| 3339 | v->HRatio, |
| 3340 | v->HRatioChroma, |
| 3341 | v->DPPPerPlane, |
| 3342 | dummy5, |
| 3343 | dummy6, |
| 3344 | dummy3, |
| 3345 | dummy4, |
| 3346 | v->SwathHeightY, |
| 3347 | v->SwathHeightC, |
| 3348 | v->DETBufferSizeY, |
| 3349 | v->DETBufferSizeC, |
| 3350 | dummy7, |
| 3351 | &dummysinglestring); |
| 3352 | } |
| 3353 | |
| 3354 | static double CalculateTWait(unsigned int PrefetchMode, double DRAMClockChangeLatency, double UrgentLatency, double SREnterPlusExitTime) |
| 3355 | { |
| 3356 | if (PrefetchMode == 0) { |
| 3357 | return dml_max(DRAMClockChangeLatency + UrgentLatency, dml_max(SREnterPlusExitTime, UrgentLatency)); |
| 3358 | } else if (PrefetchMode == 1) { |
| 3359 | return dml_max(SREnterPlusExitTime, UrgentLatency); |
| 3360 | } else { |
| 3361 | return UrgentLatency; |
| 3362 | } |
| 3363 | } |
| 3364 | |
| 3365 | double dml31_CalculateWriteBackDISPCLK( |
| 3366 | enum source_format_class WritebackPixelFormat, |
| 3367 | double PixelClock, |
| 3368 | double WritebackHRatio, |
| 3369 | double WritebackVRatio, |
| 3370 | unsigned int WritebackHTaps, |
| 3371 | unsigned int WritebackVTaps, |
| 3372 | long WritebackSourceWidth, |
| 3373 | long WritebackDestinationWidth, |
| 3374 | unsigned int HTotal, |
| 3375 | unsigned int WritebackLineBufferSize) |
| 3376 | { |
| 3377 | double DISPCLK_H, DISPCLK_V, DISPCLK_HB; |
| 3378 | |
| 3379 | DISPCLK_H = PixelClock * dml_ceil(WritebackHTaps / 8.0, 1) / WritebackHRatio; |
| 3380 | DISPCLK_V = PixelClock * (WritebackVTaps * dml_ceil(WritebackDestinationWidth / 6.0, 1) + 8.0) / HTotal; |
| 3381 | DISPCLK_HB = PixelClock * WritebackVTaps * (WritebackDestinationWidth * WritebackVTaps - WritebackLineBufferSize / 57.0) / 6.0 / WritebackSourceWidth; |
| 3382 | return dml_max3(DISPCLK_H, DISPCLK_V, DISPCLK_HB); |
| 3383 | } |
| 3384 | |
| 3385 | static double CalculateWriteBackDelay( |
| 3386 | enum source_format_class WritebackPixelFormat, |
| 3387 | double WritebackHRatio, |
| 3388 | double WritebackVRatio, |
| 3389 | unsigned int WritebackVTaps, |
| 3390 | int WritebackDestinationWidth, |
| 3391 | int WritebackDestinationHeight, |
| 3392 | int WritebackSourceHeight, |
| 3393 | unsigned int HTotal) |
| 3394 | { |
| 3395 | double CalculateWriteBackDelay; |
| 3396 | double Line_length; |
| 3397 | double Output_lines_last_notclamped; |
| 3398 | double WritebackVInit; |
| 3399 | |
| 3400 | WritebackVInit = (WritebackVRatio + WritebackVTaps + 1) / 2; |
| 3401 | Line_length = dml_max((double) WritebackDestinationWidth, dml_ceil(WritebackDestinationWidth / 6.0, 1) * WritebackVTaps); |
| 3402 | Output_lines_last_notclamped = WritebackDestinationHeight - 1 - dml_ceil((WritebackSourceHeight - WritebackVInit) / WritebackVRatio, 1); |
| 3403 | if (Output_lines_last_notclamped < 0) { |
| 3404 | CalculateWriteBackDelay = 0; |
| 3405 | } else { |
| 3406 | CalculateWriteBackDelay = Output_lines_last_notclamped * Line_length + (HTotal - WritebackDestinationWidth) + 80; |
| 3407 | } |
| 3408 | return CalculateWriteBackDelay; |
| 3409 | } |
| 3410 | |
| 3411 | static void CalculateVupdateAndDynamicMetadataParameters( |
| 3412 | int MaxInterDCNTileRepeaters, |
| 3413 | double DPPCLK, |
| 3414 | double DISPCLK, |
| 3415 | double DCFClkDeepSleep, |
| 3416 | double PixelClock, |
| 3417 | int HTotal, |
| 3418 | int VBlank, |
| 3419 | int DynamicMetadataTransmittedBytes, |
| 3420 | int DynamicMetadataLinesBeforeActiveRequired, |
| 3421 | int InterlaceEnable, |
| 3422 | bool_Bool ProgressiveToInterlaceUnitInOPP, |
| 3423 | double *TSetup, |
| 3424 | double *Tdmbf, |
| 3425 | double *Tdmec, |
| 3426 | double *Tdmsks, |
| 3427 | int *VUpdateOffsetPix, |
| 3428 | double *VUpdateWidthPix, |
| 3429 | double *VReadyOffsetPix) |
| 3430 | { |
| 3431 | double TotalRepeaterDelayTime; |
| 3432 | |
| 3433 | TotalRepeaterDelayTime = MaxInterDCNTileRepeaters * (2 / DPPCLK + 3 / DISPCLK); |
| 3434 | *VUpdateWidthPix = dml_ceil((14.0 / DCFClkDeepSleep + 12.0 / DPPCLK + TotalRepeaterDelayTime) * PixelClock, 1.0); |
| 3435 | *VReadyOffsetPix = dml_ceil(dml_max(150.0 / DPPCLK, TotalRepeaterDelayTime + 20.0 / DCFClkDeepSleep + 10.0 / DPPCLK) * PixelClock, 1.0); |
| 3436 | *VUpdateOffsetPix = dml_ceil(HTotal / 4.0, 1); |
| 3437 | *TSetup = (*VUpdateOffsetPix + *VUpdateWidthPix + *VReadyOffsetPix) / PixelClock; |
| 3438 | *Tdmbf = DynamicMetadataTransmittedBytes / 4.0 / DISPCLK; |
| 3439 | *Tdmec = HTotal / PixelClock; |
| 3440 | if (DynamicMetadataLinesBeforeActiveRequired == 0) { |
| 3441 | *Tdmsks = VBlank * HTotal / PixelClock / 2.0; |
| 3442 | } else { |
| 3443 | *Tdmsks = DynamicMetadataLinesBeforeActiveRequired * HTotal / PixelClock; |
| 3444 | } |
| 3445 | if (InterlaceEnable == 1 && ProgressiveToInterlaceUnitInOPP == false0) { |
| 3446 | *Tdmsks = *Tdmsks / 2; |
| 3447 | } |
| 3448 | #ifdef __DML_VBA_DEBUG__ |
| 3449 | dml_print("DML::%s: VUpdateWidthPix = %d\n", __func__, *VUpdateWidthPix){do { } while(0); }; |
| 3450 | dml_print("DML::%s: VReadyOffsetPix = %d\n", __func__, *VReadyOffsetPix){do { } while(0); }; |
| 3451 | dml_print("DML::%s: VUpdateOffsetPix = %d\n", __func__, *VUpdateOffsetPix){do { } while(0); }; |
| 3452 | #endif |
| 3453 | } |
| 3454 | |
| 3455 | static void CalculateRowBandwidth( |
| 3456 | bool_Bool GPUVMEnable, |
| 3457 | enum source_format_class SourcePixelFormat, |
| 3458 | double VRatio, |
| 3459 | double VRatioChroma, |
| 3460 | bool_Bool DCCEnable, |
| 3461 | double LineTime, |
| 3462 | unsigned int MetaRowByteLuma, |
| 3463 | unsigned int MetaRowByteChroma, |
| 3464 | unsigned int meta_row_height_luma, |
| 3465 | unsigned int meta_row_height_chroma, |
| 3466 | unsigned int PixelPTEBytesPerRowLuma, |
| 3467 | unsigned int PixelPTEBytesPerRowChroma, |
| 3468 | unsigned int dpte_row_height_luma, |
| 3469 | unsigned int dpte_row_height_chroma, |
| 3470 | double *meta_row_bw, |
| 3471 | double *dpte_row_bw) |
| 3472 | { |
| 3473 | if (DCCEnable != true1) { |
| 3474 | *meta_row_bw = 0; |
| 3475 | } else if (SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10 || SourcePixelFormat == dm_420_12 || SourcePixelFormat == dm_rgbe_alpha) { |
| 3476 | *meta_row_bw = VRatio * MetaRowByteLuma / (meta_row_height_luma * LineTime) + VRatioChroma * MetaRowByteChroma / (meta_row_height_chroma * LineTime); |
| 3477 | } else { |
| 3478 | *meta_row_bw = VRatio * MetaRowByteLuma / (meta_row_height_luma * LineTime); |
| 3479 | } |
| 3480 | |
| 3481 | if (GPUVMEnable != true1) { |
| 3482 | *dpte_row_bw = 0; |
| 3483 | } else if (SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10 || SourcePixelFormat == dm_420_12 || SourcePixelFormat == dm_rgbe_alpha) { |
| 3484 | *dpte_row_bw = VRatio * PixelPTEBytesPerRowLuma / (dpte_row_height_luma * LineTime) |
| 3485 | + VRatioChroma * PixelPTEBytesPerRowChroma / (dpte_row_height_chroma * LineTime); |
| 3486 | } else { |
| 3487 | *dpte_row_bw = VRatio * PixelPTEBytesPerRowLuma / (dpte_row_height_luma * LineTime); |
| 3488 | } |
| 3489 | } |
| 3490 | |
| 3491 | static void CalculateFlipSchedule( |
| 3492 | struct display_mode_lib *mode_lib, |
| 3493 | unsigned int k, |
| 3494 | double HostVMInefficiencyFactor, |
| 3495 | double UrgentExtraLatency, |
| 3496 | double UrgentLatency, |
| 3497 | double PDEAndMetaPTEBytesPerFrame, |
| 3498 | double MetaRowBytes, |
| 3499 | double DPTEBytesPerRow) |
| 3500 | { |
| 3501 | struct vba_vars_st *v = &mode_lib->vba; |
| 3502 | double min_row_time = 0.0; |
| 3503 | unsigned int HostVMDynamicLevelsTrips; |
| 3504 | double TimeForFetchingMetaPTEImmediateFlip; |
| 3505 | double TimeForFetchingRowInVBlankImmediateFlip; |
| 3506 | double ImmediateFlipBW; |
| 3507 | double LineTime = v->HTotal[k] / v->PixelClock[k]; |
| 3508 | |
| 3509 | if (v->GPUVMEnable == true1 && v->HostVMEnable == true1) { |
| 3510 | HostVMDynamicLevelsTrips = v->HostVMMaxNonCachedPageTableLevels; |
| 3511 | } else { |
| 3512 | HostVMDynamicLevelsTrips = 0; |
| 3513 | } |
| 3514 | |
| 3515 | if (v->GPUVMEnable == true1 || v->DCCEnable[k] == true1) { |
| 3516 | ImmediateFlipBW = (PDEAndMetaPTEBytesPerFrame + MetaRowBytes + DPTEBytesPerRow) * v->BandwidthAvailableForImmediateFlip / v->TotImmediateFlipBytes; |
| 3517 | } |
| 3518 | |
| 3519 | if (v->GPUVMEnable == true1) { |
| 3520 | TimeForFetchingMetaPTEImmediateFlip = dml_max3( |
| 3521 | v->Tno_bw[k] + PDEAndMetaPTEBytesPerFrame * HostVMInefficiencyFactor / ImmediateFlipBW, |
| 3522 | UrgentExtraLatency + UrgentLatency * (v->GPUVMMaxPageTableLevels * (HostVMDynamicLevelsTrips + 1) - 1), |
| 3523 | LineTime / 4.0); |
| 3524 | } else { |
| 3525 | TimeForFetchingMetaPTEImmediateFlip = 0; |
| 3526 | } |
| 3527 | |
| 3528 | v->DestinationLinesToRequestVMInImmediateFlip[k] = dml_ceil(4.0 * (TimeForFetchingMetaPTEImmediateFlip / LineTime), 1) / 4.0; |
| 3529 | if ((v->GPUVMEnable == true1 || v->DCCEnable[k] == true1)) { |
| 3530 | TimeForFetchingRowInVBlankImmediateFlip = dml_max3( |
| 3531 | (MetaRowBytes + DPTEBytesPerRow * HostVMInefficiencyFactor) / ImmediateFlipBW, |
| 3532 | UrgentLatency * (HostVMDynamicLevelsTrips + 1), |
| 3533 | LineTime / 4); |
| 3534 | } else { |
| 3535 | TimeForFetchingRowInVBlankImmediateFlip = 0; |
| 3536 | } |
| 3537 | |
| 3538 | v->DestinationLinesToRequestRowInImmediateFlip[k] = dml_ceil(4.0 * (TimeForFetchingRowInVBlankImmediateFlip / LineTime), 1) / 4.0; |
| 3539 | |
| 3540 | if (v->GPUVMEnable == true1) { |
| 3541 | v->final_flip_bw[k] = dml_max( |
| 3542 | PDEAndMetaPTEBytesPerFrame * HostVMInefficiencyFactor / (v->DestinationLinesToRequestVMInImmediateFlip[k] * LineTime), |
| 3543 | (MetaRowBytes + DPTEBytesPerRow * HostVMInefficiencyFactor) / (v->DestinationLinesToRequestRowInImmediateFlip[k] * LineTime)); |
| 3544 | } else if ((v->GPUVMEnable == true1 || v->DCCEnable[k] == true1)) { |
| 3545 | v->final_flip_bw[k] = (MetaRowBytes + DPTEBytesPerRow * HostVMInefficiencyFactor) / (v->DestinationLinesToRequestRowInImmediateFlip[k] * LineTime); |
| 3546 | } else { |
| 3547 | v->final_flip_bw[k] = 0; |
| 3548 | } |
| 3549 | |
| 3550 | if (v->SourcePixelFormat[k] == dm_420_8 || v->SourcePixelFormat[k] == dm_420_10 || v->SourcePixelFormat[k] == dm_rgbe_alpha) { |
| 3551 | if (v->GPUVMEnable == true1 && v->DCCEnable[k] != true1) { |
| 3552 | min_row_time = dml_min(v->dpte_row_height[k] * LineTime / v->VRatio[k], v->dpte_row_height_chroma[k] * LineTime / v->VRatioChroma[k]); |
| 3553 | } else if (v->GPUVMEnable != true1 && v->DCCEnable[k] == true1) { |
| 3554 | min_row_time = dml_min(v->meta_row_height[k] * LineTime / v->VRatio[k], v->meta_row_height_chroma[k] * LineTime / v->VRatioChroma[k]); |
| 3555 | } else { |
| 3556 | min_row_time = dml_min4( |
| 3557 | v->dpte_row_height[k] * LineTime / v->VRatio[k], |
| 3558 | v->meta_row_height[k] * LineTime / v->VRatio[k], |
| 3559 | v->dpte_row_height_chroma[k] * LineTime / v->VRatioChroma[k], |
| 3560 | v->meta_row_height_chroma[k] * LineTime / v->VRatioChroma[k]); |
| 3561 | } |
| 3562 | } else { |
| 3563 | if (v->GPUVMEnable == true1 && v->DCCEnable[k] != true1) { |
| 3564 | min_row_time = v->dpte_row_height[k] * LineTime / v->VRatio[k]; |
| 3565 | } else if (v->GPUVMEnable != true1 && v->DCCEnable[k] == true1) { |
| 3566 | min_row_time = v->meta_row_height[k] * LineTime / v->VRatio[k]; |
| 3567 | } else { |
| 3568 | min_row_time = dml_min(v->dpte_row_height[k] * LineTime / v->VRatio[k], v->meta_row_height[k] * LineTime / v->VRatio[k]); |
| 3569 | } |
| 3570 | } |
| 3571 | |
| 3572 | if (v->DestinationLinesToRequestVMInImmediateFlip[k] >= 32 || v->DestinationLinesToRequestRowInImmediateFlip[k] >= 16 |
| 3573 | || TimeForFetchingMetaPTEImmediateFlip + 2 * TimeForFetchingRowInVBlankImmediateFlip > min_row_time) { |
| 3574 | v->ImmediateFlipSupportedForPipe[k] = false0; |
| 3575 | } else { |
| 3576 | v->ImmediateFlipSupportedForPipe[k] = true1; |
| 3577 | } |
| 3578 | |
| 3579 | #ifdef __DML_VBA_DEBUG__ |
| 3580 | dml_print("DML::%s: DestinationLinesToRequestVMInImmediateFlip = %f\n", __func__, v->DestinationLinesToRequestVMInImmediateFlip[k]){do { } while(0); }; |
| 3581 | dml_print("DML::%s: DestinationLinesToRequestRowInImmediateFlip = %f\n", __func__, v->DestinationLinesToRequestRowInImmediateFlip[k]){do { } while(0); }; |
| 3582 | dml_print("DML::%s: TimeForFetchingMetaPTEImmediateFlip = %f\n", __func__, TimeForFetchingMetaPTEImmediateFlip){do { } while(0); }; |
| 3583 | dml_print("DML::%s: TimeForFetchingRowInVBlankImmediateFlip = %f\n", __func__, TimeForFetchingRowInVBlankImmediateFlip){do { } while(0); }; |
| 3584 | dml_print("DML::%s: min_row_time = %f\n", __func__, min_row_time){do { } while(0); }; |
| 3585 | dml_print("DML::%s: ImmediateFlipSupportedForPipe = %d\n", __func__, v->ImmediateFlipSupportedForPipe[k]){do { } while(0); }; |
| 3586 | #endif |
| 3587 | |
| 3588 | } |
| 3589 | |
| 3590 | static double TruncToValidBPP( |
| 3591 | double LinkBitRate, |
| 3592 | int Lanes, |
| 3593 | int HTotal, |
| 3594 | int HActive, |
| 3595 | double PixelClock, |
| 3596 | double DesiredBPP, |
| 3597 | bool_Bool DSCEnable, |
| 3598 | enum output_encoder_class Output, |
| 3599 | enum output_format_class Format, |
| 3600 | unsigned int DSCInputBitPerComponent, |
| 3601 | int DSCSlices, |
| 3602 | int AudioRate, |
| 3603 | int AudioLayout, |
| 3604 | enum odm_combine_mode ODMCombine) |
| 3605 | { |
| 3606 | double MaxLinkBPP; |
| 3607 | int MinDSCBPP; |
| 3608 | double MaxDSCBPP; |
| 3609 | int NonDSCBPP0; |
| 3610 | int NonDSCBPP1; |
| 3611 | int NonDSCBPP2; |
| 3612 | |
| 3613 | if (Format == dm_420) { |
| 3614 | NonDSCBPP0 = 12; |
| 3615 | NonDSCBPP1 = 15; |
| 3616 | NonDSCBPP2 = 18; |
| 3617 | MinDSCBPP = 6; |
| 3618 | MaxDSCBPP = 1.5 * DSCInputBitPerComponent - 1 / 16; |
| 3619 | } else if (Format == dm_444) { |
| 3620 | NonDSCBPP0 = 24; |
| 3621 | NonDSCBPP1 = 30; |
| 3622 | NonDSCBPP2 = 36; |
| 3623 | MinDSCBPP = 8; |
| 3624 | MaxDSCBPP = 3 * DSCInputBitPerComponent - 1.0 / 16; |
| 3625 | } else { |
| 3626 | |
| 3627 | NonDSCBPP0 = 16; |
| 3628 | NonDSCBPP1 = 20; |
| 3629 | NonDSCBPP2 = 24; |
| 3630 | |
| 3631 | if (Format == dm_n422) { |
| 3632 | MinDSCBPP = 7; |
| 3633 | MaxDSCBPP = 2 * DSCInputBitPerComponent - 1.0 / 16.0; |
| 3634 | } else { |
| 3635 | MinDSCBPP = 8; |
| 3636 | MaxDSCBPP = 3 * DSCInputBitPerComponent - 1.0 / 16.0; |
| 3637 | } |
| 3638 | } |
| 3639 | |
| 3640 | if (DSCEnable && Output == dm_dp) { |
| 3641 | MaxLinkBPP = LinkBitRate / 10 * 8 * Lanes / PixelClock * (1 - 2.4 / 100); |
| 3642 | } else { |
| 3643 | MaxLinkBPP = LinkBitRate / 10 * 8 * Lanes / PixelClock; |
| 3644 | } |
| 3645 | |
| 3646 | if (ODMCombine == dm_odm_combine_mode_4to1 && MaxLinkBPP > 16) { |
| 3647 | MaxLinkBPP = 16; |
| 3648 | } else if (ODMCombine == dm_odm_combine_mode_2to1 && MaxLinkBPP > 32) { |
| 3649 | MaxLinkBPP = 32; |
| 3650 | } |
| 3651 | |
| 3652 | if (DesiredBPP == 0) { |
| 3653 | if (DSCEnable) { |
| 3654 | if (MaxLinkBPP < MinDSCBPP) { |
| 3655 | return BPP_INVALID0; |
| 3656 | } else if (MaxLinkBPP >= MaxDSCBPP) { |
| 3657 | return MaxDSCBPP; |
| 3658 | } else { |
| 3659 | return dml_floor(16.0 * MaxLinkBPP, 1.0) / 16.0; |
| 3660 | } |
| 3661 | } else { |
| 3662 | if (MaxLinkBPP >= NonDSCBPP2) { |
| 3663 | return NonDSCBPP2; |
| 3664 | } else if (MaxLinkBPP >= NonDSCBPP1) { |
| 3665 | return NonDSCBPP1; |
| 3666 | } else if (MaxLinkBPP >= NonDSCBPP0) { |
| 3667 | return 16.0; |
| 3668 | } else { |
| 3669 | return BPP_INVALID0; |
| 3670 | } |
| 3671 | } |
| 3672 | } else { |
| 3673 | if (!((DSCEnable == false0 && (DesiredBPP == NonDSCBPP2 || DesiredBPP == NonDSCBPP1 || DesiredBPP <= NonDSCBPP0)) |
| 3674 | || (DSCEnable && DesiredBPP >= MinDSCBPP && DesiredBPP <= MaxDSCBPP))) { |
| 3675 | return BPP_INVALID0; |
| 3676 | } else { |
| 3677 | return DesiredBPP; |
| 3678 | } |
| 3679 | } |
| 3680 | return BPP_INVALID0; |
| 3681 | } |
| 3682 | |
| 3683 | static noinline__attribute__((__noinline__)) void CalculatePrefetchSchedulePerPlane( |
| 3684 | struct display_mode_lib *mode_lib, |
| 3685 | double HostVMInefficiencyFactor, |
| 3686 | int i, |
| 3687 | unsigned j, |
| 3688 | unsigned k) |
| 3689 | { |
| 3690 | struct vba_vars_st *v = &mode_lib->vba; |
| 3691 | Pipe myPipe; |
| 3692 | |
| 3693 | myPipe.DPPCLK = v->RequiredDPPCLK[i][j][k]; |
| 3694 | myPipe.DISPCLK = v->RequiredDISPCLK[i][j]; |
| 3695 | myPipe.PixelClock = v->PixelClock[k]; |
| 3696 | myPipe.DCFCLKDeepSleep = v->ProjectedDCFCLKDeepSleep[i][j]; |
| 3697 | myPipe.DPPPerPlane = v->NoOfDPP[i][j][k]; |
| 3698 | myPipe.ScalerEnabled = v->ScalerEnabled[k]; |
| 3699 | myPipe.VRatio = mode_lib->vba.VRatio[k]; |
| 3700 | myPipe.VRatioChroma = mode_lib->vba.VRatioChroma[k]; |
| 3701 | |
| 3702 | myPipe.SourceScan = v->SourceScan[k]; |
| 3703 | myPipe.BlockWidth256BytesY = v->Read256BlockWidthY[k]; |
| 3704 | myPipe.BlockHeight256BytesY = v->Read256BlockHeightY[k]; |
| 3705 | myPipe.BlockWidth256BytesC = v->Read256BlockWidthC[k]; |
| 3706 | myPipe.BlockHeight256BytesC = v->Read256BlockHeightC[k]; |
| 3707 | myPipe.InterlaceEnable = v->Interlace[k]; |
| 3708 | myPipe.NumberOfCursors = v->NumberOfCursors[k]; |
| 3709 | myPipe.VBlank = v->VTotal[k] - v->VActive[k]; |
| 3710 | myPipe.HTotal = v->HTotal[k]; |
| 3711 | myPipe.DCCEnable = v->DCCEnable[k]; |
| 3712 | myPipe.ODMCombineIsEnabled = v->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_4to1 |
| 3713 | || v->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_2to1; |
| 3714 | myPipe.SourcePixelFormat = v->SourcePixelFormat[k]; |
| 3715 | myPipe.BytePerPixelY = v->BytePerPixelY[k]; |
| 3716 | myPipe.BytePerPixelC = v->BytePerPixelC[k]; |
| 3717 | myPipe.ProgressiveToInterlaceUnitInOPP = v->ProgressiveToInterlaceUnitInOPP; |
| 3718 | v->NoTimeForPrefetch[i][j][k] = CalculatePrefetchSchedule( |
| 3719 | mode_lib, |
| 3720 | HostVMInefficiencyFactor, |
| 3721 | &myPipe, |
| 3722 | v->DSCDelayPerState[i][k], |
| 3723 | v->DPPCLKDelaySubtotal + v->DPPCLKDelayCNVCFormater, |
| 3724 | v->DPPCLKDelaySCL, |
| 3725 | v->DPPCLKDelaySCLLBOnly, |
| 3726 | v->DPPCLKDelayCNVCCursor, |
| 3727 | v->DISPCLKDelaySubtotal, |
| 3728 | v->SwathWidthYThisState[k] / v->HRatio[k], |
| 3729 | v->OutputFormat[k], |
| 3730 | v->MaxInterDCNTileRepeaters, |
| 3731 | dml_min(v->MaxVStartup, v->MaximumVStartup[i][j][k]), |
| 3732 | v->MaximumVStartup[i][j][k], |
| 3733 | v->GPUVMMaxPageTableLevels, |
| 3734 | v->GPUVMEnable, |
| 3735 | v->HostVMEnable, |
| 3736 | v->HostVMMaxNonCachedPageTableLevels, |
| 3737 | v->HostVMMinPageSize, |
| 3738 | v->DynamicMetadataEnable[k], |
| 3739 | v->DynamicMetadataVMEnabled, |
| 3740 | v->DynamicMetadataLinesBeforeActiveRequired[k], |
| 3741 | v->DynamicMetadataTransmittedBytes[k], |
| 3742 | v->UrgLatency[i], |
| 3743 | v->ExtraLatency, |
| 3744 | v->TimeCalc, |
| 3745 | v->PDEAndMetaPTEBytesPerFrame[i][j][k], |
| 3746 | v->MetaRowBytes[i][j][k], |
| 3747 | v->DPTEBytesPerRow[i][j][k], |
| 3748 | v->PrefetchLinesY[i][j][k], |
| 3749 | v->SwathWidthYThisState[k], |
| 3750 | v->PrefillY[k], |
| 3751 | v->MaxNumSwY[k], |
| 3752 | v->PrefetchLinesC[i][j][k], |
| 3753 | v->SwathWidthCThisState[k], |
| 3754 | v->PrefillC[k], |
| 3755 | v->MaxNumSwC[k], |
| 3756 | v->swath_width_luma_ub_this_state[k], |
| 3757 | v->swath_width_chroma_ub_this_state[k], |
| 3758 | v->SwathHeightYThisState[k], |
| 3759 | v->SwathHeightCThisState[k], |
| 3760 | v->TWait, |
| 3761 | &v->DSTXAfterScaler[k], |
| 3762 | &v->DSTYAfterScaler[k], |
| 3763 | &v->LineTimesForPrefetch[k], |
| 3764 | &v->PrefetchBW[k], |
| 3765 | &v->LinesForMetaPTE[k], |
| 3766 | &v->LinesForMetaAndDPTERow[k], |
| 3767 | &v->VRatioPreY[i][j][k], |
| 3768 | &v->VRatioPreC[i][j][k], |
| 3769 | &v->RequiredPrefetchPixelDataBWLuma[i][j][k], |
| 3770 | &v->RequiredPrefetchPixelDataBWChroma[i][j][k], |
| 3771 | &v->NoTimeForDynamicMetadata[i][j][k], |
| 3772 | &v->Tno_bw[k], |
| 3773 | &v->prefetch_vmrow_bw[k], |
| 3774 | &v->dummy7[k], |
| 3775 | &v->dummy8[k], |
| 3776 | &v->dummy13[k], |
| 3777 | &v->VUpdateOffsetPix[k], |
| 3778 | &v->VUpdateWidthPix[k], |
| 3779 | &v->VReadyOffsetPix[k]); |
| 3780 | } |
| 3781 | |
| 3782 | static void PatchDETBufferSizeInKByte(unsigned int NumberOfActivePlanes, int NoOfDPPThisState[], unsigned int config_return_buffer_size_in_kbytes, unsigned int DETBufferSizeInKByte[]) |
| 3783 | { |
| 3784 | int i, total_pipes = 0; |
| 3785 | for (i = 0; i < NumberOfActivePlanes; i++) |
| 3786 | total_pipes += NoOfDPPThisState[i]; |
| 3787 | DETBufferSizeInKByte[0] = ((config_return_buffer_size_in_kbytes - DCN3_15_MIN_COMPBUF_SIZE_KB128) / 64 / total_pipes) * 64; |
| 3788 | if (DETBufferSizeInKByte[0] > DCN3_15_MAX_DET_SIZE384) |
| 3789 | DETBufferSizeInKByte[0] = DCN3_15_MAX_DET_SIZE384; |
| 3790 | for (i = 1; i < NumberOfActivePlanes; i++) |
| 3791 | DETBufferSizeInKByte[i] = DETBufferSizeInKByte[0]; |
| 3792 | } |
| 3793 | |
| 3794 | |
| 3795 | void dml31_ModeSupportAndSystemConfigurationFull(struct display_mode_lib *mode_lib) |
| 3796 | { |
| 3797 | struct vba_vars_st *v = &mode_lib->vba; |
| 3798 | |
| 3799 | int i, j; |
| 3800 | unsigned int k, m; |
| 3801 | int ReorderingBytes; |
| 3802 | int MinPrefetchMode = 0, MaxPrefetchMode = 2; |
| 3803 | bool_Bool NoChroma = true1; |
| 3804 | bool_Bool EnoughWritebackUnits = true1; |
| 3805 | bool_Bool P2IWith420 = false0; |
| 3806 | bool_Bool DSCOnlyIfNecessaryWithBPP = false0; |
| 3807 | bool_Bool DSC422NativeNotSupported = false0; |
| 3808 | double MaxTotalVActiveRDBandwidth; |
| 3809 | bool_Bool ViewportExceedsSurface = false0; |
| 3810 | bool_Bool FMTBufferExceeded = false0; |
| 3811 | |
| 3812 | /*MODE SUPPORT, VOLTAGE STATE AND SOC CONFIGURATION*/ |
| 3813 | |
| 3814 | CalculateMinAndMaxPrefetchMode( |
| 3815 | mode_lib->vba.AllowDRAMSelfRefreshOrDRAMClockChangeInVblank, |
| 3816 | &MinPrefetchMode, &MaxPrefetchMode); |
| 3817 | |
| 3818 | /*Scale Ratio, taps Support Check*/ |
| 3819 | |
| 3820 | v->ScaleRatioAndTapsSupport = true1; |
| 3821 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 3822 | if (v->ScalerEnabled[k] == false0 |
| 3823 | && ((v->SourcePixelFormat[k] != dm_444_64 && v->SourcePixelFormat[k] != dm_444_32 |
| 3824 | && v->SourcePixelFormat[k] != dm_444_16 && v->SourcePixelFormat[k] != dm_mono_16 |
| 3825 | && v->SourcePixelFormat[k] != dm_mono_8 && v->SourcePixelFormat[k] != dm_rgbe |
| 3826 | && v->SourcePixelFormat[k] != dm_rgbe_alpha) || v->HRatio[k] != 1.0 || v->htaps[k] != 1.0 |
| 3827 | || v->VRatio[k] != 1.0 || v->vtaps[k] != 1.0)) { |
| 3828 | v->ScaleRatioAndTapsSupport = false0; |
| 3829 | } else if (v->vtaps[k] < 1.0 || v->vtaps[k] > 8.0 || v->htaps[k] < 1.0 || v->htaps[k] > 8.0 |
| 3830 | || (v->htaps[k] > 1.0 && (v->htaps[k] % 2) == 1) || v->HRatio[k] > v->MaxHSCLRatio |
| 3831 | || v->VRatio[k] > v->MaxVSCLRatio || v->HRatio[k] > v->htaps[k] |
| 3832 | || v->VRatio[k] > v->vtaps[k] |
| 3833 | || (v->SourcePixelFormat[k] != dm_444_64 && v->SourcePixelFormat[k] != dm_444_32 |
| 3834 | && v->SourcePixelFormat[k] != dm_444_16 && v->SourcePixelFormat[k] != dm_mono_16 |
| 3835 | && v->SourcePixelFormat[k] != dm_mono_8 && v->SourcePixelFormat[k] != dm_rgbe |
| 3836 | && (v->VTAPsChroma[k] < 1 || v->VTAPsChroma[k] > 8 || v->HTAPsChroma[k] < 1 |
| 3837 | || v->HTAPsChroma[k] > 8 || (v->HTAPsChroma[k] > 1 && v->HTAPsChroma[k] % 2 == 1) |
| 3838 | || v->HRatioChroma[k] > v->MaxHSCLRatio |
| 3839 | || v->VRatioChroma[k] > v->MaxVSCLRatio |
| 3840 | || v->HRatioChroma[k] > v->HTAPsChroma[k] |
| 3841 | || v->VRatioChroma[k] > v->VTAPsChroma[k]))) { |
| 3842 | v->ScaleRatioAndTapsSupport = false0; |
| 3843 | } |
| 3844 | } |
| 3845 | /*Source Format, Pixel Format and Scan Support Check*/ |
| 3846 | |
| 3847 | v->SourceFormatPixelAndScanSupport = true1; |
| 3848 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 3849 | if ((v->SurfaceTiling[k] == dm_sw_linear && (!(v->SourceScan[k] != dm_vert) || v->DCCEnable[k] == true1)) |
| 3850 | || ((v->SurfaceTiling[k] == dm_sw_64kb_d || v->SurfaceTiling[k] == dm_sw_64kb_d_t |
| 3851 | || v->SurfaceTiling[k] == dm_sw_64kb_d_x) && !(v->SourcePixelFormat[k] == dm_444_64))) { |
| 3852 | v->SourceFormatPixelAndScanSupport = false0; |
| 3853 | } |
| 3854 | } |
| 3855 | /*Bandwidth Support Check*/ |
| 3856 | |
| 3857 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 3858 | dml30_CalculateBytePerPixelAnd256BBlockSizes( |
| 3859 | v->SourcePixelFormat[k], |
| 3860 | v->SurfaceTiling[k], |
| 3861 | &v->BytePerPixelY[k], |
| 3862 | &v->BytePerPixelC[k], |
| 3863 | &v->BytePerPixelInDETY[k], |
| 3864 | &v->BytePerPixelInDETC[k], |
| 3865 | &v->Read256BlockHeightY[k], |
| 3866 | &v->Read256BlockHeightC[k], |
| 3867 | &v->Read256BlockWidthY[k], |
| 3868 | &v->Read256BlockWidthC[k]); |
| 3869 | } |
| 3870 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 3871 | if (v->SourceScan[k] != dm_vert) { |
| 3872 | v->SwathWidthYSingleDPP[k] = v->ViewportWidth[k]; |
| 3873 | v->SwathWidthCSingleDPP[k] = v->ViewportWidthChroma[k]; |
| 3874 | } else { |
| 3875 | v->SwathWidthYSingleDPP[k] = v->ViewportHeight[k]; |
| 3876 | v->SwathWidthCSingleDPP[k] = v->ViewportHeightChroma[k]; |
| 3877 | } |
| 3878 | } |
| 3879 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 3880 | v->ReadBandwidthLuma[k] = v->SwathWidthYSingleDPP[k] * dml_ceil(v->BytePerPixelInDETY[k], 1.0) |
| 3881 | / (v->HTotal[k] / v->PixelClock[k]) * v->VRatio[k]; |
| 3882 | v->ReadBandwidthChroma[k] = v->SwathWidthYSingleDPP[k] / 2 * dml_ceil(v->BytePerPixelInDETC[k], 2.0) |
| 3883 | / (v->HTotal[k] / v->PixelClock[k]) * v->VRatio[k] / 2.0; |
| 3884 | } |
| 3885 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 3886 | if (v->WritebackEnable[k] == true1 && v->WritebackPixelFormat[k] == dm_444_64) { |
| 3887 | v->WriteBandwidth[k] = v->WritebackDestinationWidth[k] * v->WritebackDestinationHeight[k] |
| 3888 | / (v->WritebackSourceHeight[k] * v->HTotal[k] / v->PixelClock[k]) * 8.0; |
| 3889 | } else if (v->WritebackEnable[k] == true1) { |
| 3890 | v->WriteBandwidth[k] = v->WritebackDestinationWidth[k] * v->WritebackDestinationHeight[k] |
| 3891 | / (v->WritebackSourceHeight[k] * v->HTotal[k] / v->PixelClock[k]) * 4.0; |
| 3892 | } else { |
| 3893 | v->WriteBandwidth[k] = 0.0; |
| 3894 | } |
| 3895 | } |
| 3896 | |
| 3897 | /*Writeback Latency support check*/ |
| 3898 | |
| 3899 | v->WritebackLatencySupport = true1; |
| 3900 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 3901 | if (v->WritebackEnable[k] == true1 && (v->WriteBandwidth[k] > v->WritebackInterfaceBufferSize * 1024 / v->WritebackLatency)) { |
| 3902 | v->WritebackLatencySupport = false0; |
| 3903 | } |
| 3904 | } |
| 3905 | |
| 3906 | /*Writeback Mode Support Check*/ |
| 3907 | |
| 3908 | v->TotalNumberOfActiveWriteback = 0; |
| 3909 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 3910 | if (v->WritebackEnable[k] == true1) { |
| 3911 | v->TotalNumberOfActiveWriteback = v->TotalNumberOfActiveWriteback + 1; |
| 3912 | } |
| 3913 | } |
| 3914 | |
| 3915 | if (v->TotalNumberOfActiveWriteback > v->MaxNumWriteback) { |
| 3916 | EnoughWritebackUnits = false0; |
| 3917 | } |
| 3918 | |
| 3919 | /*Writeback Scale Ratio and Taps Support Check*/ |
| 3920 | |
| 3921 | v->WritebackScaleRatioAndTapsSupport = true1; |
| 3922 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 3923 | if (v->WritebackEnable[k] == true1) { |
| 3924 | if (v->WritebackHRatio[k] > v->WritebackMaxHSCLRatio || v->WritebackVRatio[k] > v->WritebackMaxVSCLRatio |
| 3925 | || v->WritebackHRatio[k] < v->WritebackMinHSCLRatio |
| 3926 | || v->WritebackVRatio[k] < v->WritebackMinVSCLRatio |
| 3927 | || v->WritebackHTaps[k] > v->WritebackMaxHSCLTaps |
| 3928 | || v->WritebackVTaps[k] > v->WritebackMaxVSCLTaps |
| 3929 | || v->WritebackHRatio[k] > v->WritebackHTaps[k] || v->WritebackVRatio[k] > v->WritebackVTaps[k] |
| 3930 | || (v->WritebackHTaps[k] > 2.0 && ((v->WritebackHTaps[k] % 2) == 1))) { |
| 3931 | v->WritebackScaleRatioAndTapsSupport = false0; |
| 3932 | } |
| 3933 | if (2.0 * v->WritebackDestinationWidth[k] * (v->WritebackVTaps[k] - 1) * 57 > v->WritebackLineBufferSize) { |
| 3934 | v->WritebackScaleRatioAndTapsSupport = false0; |
| 3935 | } |
| 3936 | } |
| 3937 | } |
| 3938 | /*Maximum DISPCLK/DPPCLK Support check*/ |
| 3939 | |
| 3940 | v->WritebackRequiredDISPCLK = 0.0; |
| 3941 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 3942 | if (v->WritebackEnable[k] == true1) { |
| 3943 | v->WritebackRequiredDISPCLK = dml_max( |
| 3944 | v->WritebackRequiredDISPCLK, |
| 3945 | dml31_CalculateWriteBackDISPCLK( |
| 3946 | v->WritebackPixelFormat[k], |
| 3947 | v->PixelClock[k], |
| 3948 | v->WritebackHRatio[k], |
| 3949 | v->WritebackVRatio[k], |
| 3950 | v->WritebackHTaps[k], |
| 3951 | v->WritebackVTaps[k], |
| 3952 | v->WritebackSourceWidth[k], |
| 3953 | v->WritebackDestinationWidth[k], |
| 3954 | v->HTotal[k], |
| 3955 | v->WritebackLineBufferSize)); |
| 3956 | } |
| 3957 | } |
| 3958 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 3959 | if (v->HRatio[k] > 1.0) { |
| 3960 | v->PSCL_FACTOR[k] = dml_min( |
| 3961 | v->MaxDCHUBToPSCLThroughput, |
| 3962 | v->MaxPSCLToLBThroughput * v->HRatio[k] / dml_ceil(v->htaps[k] / 6.0, 1.0)); |
| 3963 | } else { |
| 3964 | v->PSCL_FACTOR[k] = dml_min(v->MaxDCHUBToPSCLThroughput, v->MaxPSCLToLBThroughput); |
| 3965 | } |
| 3966 | if (v->BytePerPixelC[k] == 0.0) { |
| 3967 | v->PSCL_FACTOR_CHROMA[k] = 0.0; |
| 3968 | v->MinDPPCLKUsingSingleDPP[k] = v->PixelClock[k] |
| 3969 | * dml_max3( |
| 3970 | v->vtaps[k] / 6.0 * dml_min(1.0, v->HRatio[k]), |
| 3971 | v->HRatio[k] * v->VRatio[k] / v->PSCL_FACTOR[k], |
| 3972 | 1.0); |
| 3973 | if ((v->htaps[k] > 6.0 || v->vtaps[k] > 6.0) && v->MinDPPCLKUsingSingleDPP[k] < 2.0 * v->PixelClock[k]) { |
| 3974 | v->MinDPPCLKUsingSingleDPP[k] = 2.0 * v->PixelClock[k]; |
| 3975 | } |
| 3976 | } else { |
| 3977 | if (v->HRatioChroma[k] > 1.0) { |
| 3978 | v->PSCL_FACTOR_CHROMA[k] = dml_min( |
| 3979 | v->MaxDCHUBToPSCLThroughput, |
| 3980 | v->MaxPSCLToLBThroughput * v->HRatioChroma[k] / dml_ceil(v->HTAPsChroma[k] / 6.0, 1.0)); |
| 3981 | } else { |
| 3982 | v->PSCL_FACTOR_CHROMA[k] = dml_min(v->MaxDCHUBToPSCLThroughput, v->MaxPSCLToLBThroughput); |
| 3983 | } |
| 3984 | v->MinDPPCLKUsingSingleDPP[k] = v->PixelClock[k] |
| 3985 | * dml_max5( |
| 3986 | v->vtaps[k] / 6.0 * dml_min(1.0, v->HRatio[k]), |
| 3987 | v->HRatio[k] * v->VRatio[k] / v->PSCL_FACTOR[k], |
| 3988 | v->VTAPsChroma[k] / 6.0 * dml_min(1.0, v->HRatioChroma[k]), |
| 3989 | v->HRatioChroma[k] * v->VRatioChroma[k] / v->PSCL_FACTOR_CHROMA[k], |
| 3990 | 1.0); |
| 3991 | if ((v->htaps[k] > 6.0 || v->vtaps[k] > 6.0 || v->HTAPsChroma[k] > 6.0 || v->VTAPsChroma[k] > 6.0) |
| 3992 | && v->MinDPPCLKUsingSingleDPP[k] < 2.0 * v->PixelClock[k]) { |
| 3993 | v->MinDPPCLKUsingSingleDPP[k] = 2.0 * v->PixelClock[k]; |
| 3994 | } |
| 3995 | } |
| 3996 | } |
| 3997 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 3998 | int MaximumSwathWidthSupportLuma; |
| 3999 | int MaximumSwathWidthSupportChroma; |
| 4000 | |
| 4001 | if (v->SurfaceTiling[k] == dm_sw_linear) { |
| 4002 | MaximumSwathWidthSupportLuma = 8192.0; |
| 4003 | } else if (v->SourceScan[k] == dm_vert && v->BytePerPixelC[k] > 0) { |
| 4004 | MaximumSwathWidthSupportLuma = 2880.0; |
| 4005 | } else if (v->SourcePixelFormat[k] == dm_rgbe_alpha) { |
| 4006 | MaximumSwathWidthSupportLuma = 3840.0; |
| 4007 | } else { |
| 4008 | MaximumSwathWidthSupportLuma = 5760.0; |
| 4009 | } |
| 4010 | |
| 4011 | if (v->SourcePixelFormat[k] == dm_420_8 || v->SourcePixelFormat[k] == dm_420_10 || v->SourcePixelFormat[k] == dm_420_12) { |
| 4012 | MaximumSwathWidthSupportChroma = MaximumSwathWidthSupportLuma / 2.0; |
| 4013 | } else { |
| 4014 | MaximumSwathWidthSupportChroma = MaximumSwathWidthSupportLuma; |
| 4015 | } |
| 4016 | v->MaximumSwathWidthInLineBufferLuma = v->LineBufferSize * dml_max(v->HRatio[k], 1.0) / v->LBBitPerPixel[k] |
| 4017 | / (v->vtaps[k] + dml_max(dml_ceil(v->VRatio[k], 1.0) - 2, 0.0)); |
| 4018 | if (v->BytePerPixelC[k] == 0.0) { |
| 4019 | v->MaximumSwathWidthInLineBufferChroma = 0; |
| 4020 | } else { |
| 4021 | v->MaximumSwathWidthInLineBufferChroma = v->LineBufferSize * dml_max(v->HRatioChroma[k], 1.0) / v->LBBitPerPixel[k] |
| 4022 | / (v->VTAPsChroma[k] + dml_max(dml_ceil(v->VRatioChroma[k], 1.0) - 2, 0.0)); |
| 4023 | } |
| 4024 | v->MaximumSwathWidthLuma[k] = dml_min(MaximumSwathWidthSupportLuma, v->MaximumSwathWidthInLineBufferLuma); |
| 4025 | v->MaximumSwathWidthChroma[k] = dml_min(MaximumSwathWidthSupportChroma, v->MaximumSwathWidthInLineBufferChroma); |
| 4026 | } |
| 4027 | |
| 4028 | CalculateSwathAndDETConfiguration( |
| 4029 | true1, |
| 4030 | v->NumberOfActivePlanes, |
| 4031 | mode_lib->project == DML_PROJECT_DCN315 && v->DETSizeOverride[0], |
| 4032 | v->DETBufferSizeInKByte, |
| 4033 | v->MaximumSwathWidthLuma, |
| 4034 | v->MaximumSwathWidthChroma, |
| 4035 | v->SourceScan, |
| 4036 | v->SourcePixelFormat, |
| 4037 | v->SurfaceTiling, |
| 4038 | v->ViewportWidth, |
| 4039 | v->ViewportHeight, |
| 4040 | v->SurfaceWidthY, |
| 4041 | v->SurfaceWidthC, |
| 4042 | v->SurfaceHeightY, |
| 4043 | v->SurfaceHeightC, |
| 4044 | v->Read256BlockHeightY, |
| 4045 | v->Read256BlockHeightC, |
| 4046 | v->Read256BlockWidthY, |
| 4047 | v->Read256BlockWidthC, |
| 4048 | v->odm_combine_dummy, |
| 4049 | v->BlendingAndTiming, |
| 4050 | v->BytePerPixelY, |
| 4051 | v->BytePerPixelC, |
| 4052 | v->BytePerPixelInDETY, |
| 4053 | v->BytePerPixelInDETC, |
| 4054 | v->HActive, |
| 4055 | v->HRatio, |
| 4056 | v->HRatioChroma, |
| 4057 | v->NoOfDPPThisState, |
| 4058 | v->swath_width_luma_ub_this_state, |
| 4059 | v->swath_width_chroma_ub_this_state, |
| 4060 | v->SwathWidthYThisState, |
| 4061 | v->SwathWidthCThisState, |
| 4062 | v->SwathHeightYThisState, |
| 4063 | v->SwathHeightCThisState, |
| 4064 | v->DETBufferSizeYThisState, |
| 4065 | v->DETBufferSizeCThisState, |
| 4066 | v->SingleDPPViewportSizeSupportPerPlane, |
| 4067 | &v->ViewportSizeSupport[0][0]); |
| 4068 | |
| 4069 | for (i = 0; i < v->soc.num_states; i++) { |
| 4070 | for (j = 0; j < 2; j++) { |
| 4071 | v->MaxDispclkRoundedDownToDFSGranularity = RoundToDFSGranularityDown(v->MaxDispclk[i], v->DISPCLKDPPCLKVCOSpeed); |
| 4072 | v->MaxDppclkRoundedDownToDFSGranularity = RoundToDFSGranularityDown(v->MaxDppclk[i], v->DISPCLKDPPCLKVCOSpeed); |
| 4073 | v->RequiredDISPCLK[i][j] = 0.0; |
| 4074 | v->DISPCLK_DPPCLK_Support[i][j] = true1; |
| 4075 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 4076 | v->PlaneRequiredDISPCLKWithoutODMCombine = v->PixelClock[k] * (1.0 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) |
| 4077 | * (1.0 + v->DISPCLKRampingMargin / 100.0); |
| 4078 | if ((v->PlaneRequiredDISPCLKWithoutODMCombine >= v->MaxDispclk[i] |
| 4079 | && v->MaxDispclk[i] == v->MaxDispclk[v->soc.num_states - 1] |
| 4080 | && v->MaxDppclk[i] == v->MaxDppclk[v->soc.num_states - 1])) { |
| 4081 | v->PlaneRequiredDISPCLKWithoutODMCombine = v->PixelClock[k] |
| 4082 | * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); |
| 4083 | } |
| 4084 | v->PlaneRequiredDISPCLKWithODMCombine2To1 = v->PixelClock[k] / 2 * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) |
| 4085 | * (1 + v->DISPCLKRampingMargin / 100.0); |
| 4086 | if ((v->PlaneRequiredDISPCLKWithODMCombine2To1 >= v->MaxDispclk[i] |
| 4087 | && v->MaxDispclk[i] == v->MaxDispclk[v->soc.num_states - 1] |
| 4088 | && v->MaxDppclk[i] == v->MaxDppclk[v->soc.num_states - 1])) { |
| 4089 | v->PlaneRequiredDISPCLKWithODMCombine2To1 = v->PixelClock[k] / 2 |
| 4090 | * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); |
| 4091 | } |
| 4092 | v->PlaneRequiredDISPCLKWithODMCombine4To1 = v->PixelClock[k] / 4 * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) |
| 4093 | * (1 + v->DISPCLKRampingMargin / 100.0); |
| 4094 | if ((v->PlaneRequiredDISPCLKWithODMCombine4To1 >= v->MaxDispclk[i] |
| 4095 | && v->MaxDispclk[i] == v->MaxDispclk[v->soc.num_states - 1] |
| 4096 | && v->MaxDppclk[i] == v->MaxDppclk[v->soc.num_states - 1])) { |
| 4097 | v->PlaneRequiredDISPCLKWithODMCombine4To1 = v->PixelClock[k] / 4 |
| 4098 | * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); |
| 4099 | } |
| 4100 | |
| 4101 | if (v->ODMCombinePolicy == dm_odm_combine_policy_none |
| 4102 | || !(v->Output[k] == dm_dp || |
| 4103 | v->Output[k] == dm_dp2p0 || |
| 4104 | v->Output[k] == dm_edp)) { |
| 4105 | v->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_disabled; |
| 4106 | v->PlaneRequiredDISPCLK = v->PlaneRequiredDISPCLKWithoutODMCombine; |
| 4107 | |
| 4108 | if (v->HActive[k] / 2 > DCN31_MAX_FMT_420_BUFFER_WIDTH4096) |
| 4109 | FMTBufferExceeded = true1; |
| 4110 | } else if (v->ODMCombinePolicy == dm_odm_combine_policy_2to1) { |
| 4111 | v->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_2to1; |
| 4112 | v->PlaneRequiredDISPCLK = v->PlaneRequiredDISPCLKWithODMCombine2To1; |
| 4113 | } else if (v->ODMCombinePolicy == dm_odm_combine_policy_4to1 |
| 4114 | || v->PlaneRequiredDISPCLKWithODMCombine2To1 > v->MaxDispclkRoundedDownToDFSGranularity) { |
| 4115 | v->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_4to1; |
| 4116 | v->PlaneRequiredDISPCLK = v->PlaneRequiredDISPCLKWithODMCombine4To1; |
| 4117 | } else if (v->PlaneRequiredDISPCLKWithoutODMCombine > v->MaxDispclkRoundedDownToDFSGranularity) { |
| 4118 | v->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_2to1; |
| 4119 | v->PlaneRequiredDISPCLK = v->PlaneRequiredDISPCLKWithODMCombine2To1; |
| 4120 | } else { |
| 4121 | v->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_disabled; |
| 4122 | v->PlaneRequiredDISPCLK = v->PlaneRequiredDISPCLKWithoutODMCombine; |
| 4123 | } |
| 4124 | if (v->DSCEnabled[k] && v->HActive[k] > DCN31_MAX_DSC_IMAGE_WIDTH5184 |
| 4125 | && v->ODMCombineEnablePerState[i][k] != dm_odm_combine_mode_4to1) { |
| 4126 | if (v->HActive[k] / 2 > DCN31_MAX_DSC_IMAGE_WIDTH5184) { |
| 4127 | v->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_4to1; |
| 4128 | v->PlaneRequiredDISPCLK = v->PlaneRequiredDISPCLKWithODMCombine4To1; |
| 4129 | } else { |
| 4130 | v->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_2to1; |
| 4131 | v->PlaneRequiredDISPCLK = v->PlaneRequiredDISPCLKWithODMCombine2To1; |
| 4132 | } |
| 4133 | } |
| 4134 | if (v->OutputFormat[k] == dm_420 && v->HActive[k] > DCN31_MAX_FMT_420_BUFFER_WIDTH4096 |
| 4135 | && v->ODMCombineEnablePerState[i][k] != dm_odm_combine_mode_4to1) { |
| 4136 | if (v->Output[k] == dm_hdmi) { |
| 4137 | FMTBufferExceeded = true1; |
| 4138 | } else if (v->HActive[k] / 2 > DCN31_MAX_FMT_420_BUFFER_WIDTH4096) { |
| 4139 | v->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_4to1; |
| 4140 | v->PlaneRequiredDISPCLK = v->PlaneRequiredDISPCLKWithODMCombine4To1; |
| 4141 | |
| 4142 | if (v->HActive[k] / 4 > DCN31_MAX_FMT_420_BUFFER_WIDTH4096) |
| 4143 | FMTBufferExceeded = true1; |
| 4144 | } else { |
| 4145 | v->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_2to1; |
| 4146 | v->PlaneRequiredDISPCLK = v->PlaneRequiredDISPCLKWithODMCombine2To1; |
| 4147 | } |
| 4148 | } |
| 4149 | if (v->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_4to1) { |
| 4150 | v->MPCCombine[i][j][k] = false0; |
| 4151 | v->NoOfDPP[i][j][k] = 4; |
| 4152 | v->RequiredDPPCLK[i][j][k] = v->MinDPPCLKUsingSingleDPP[k] * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) / 4; |
| 4153 | } else if (v->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_2to1) { |
| 4154 | v->MPCCombine[i][j][k] = false0; |
| 4155 | v->NoOfDPP[i][j][k] = 2; |
| 4156 | v->RequiredDPPCLK[i][j][k] = v->MinDPPCLKUsingSingleDPP[k] * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) / 2; |
| 4157 | } else if ((v->WhenToDoMPCCombine == dm_mpc_never |
| 4158 | || (v->MinDPPCLKUsingSingleDPP[k] * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) |
| 4159 | <= v->MaxDppclkRoundedDownToDFSGranularity && v->SingleDPPViewportSizeSupportPerPlane[k] == true1))) { |
| 4160 | v->MPCCombine[i][j][k] = false0; |
| 4161 | v->NoOfDPP[i][j][k] = 1; |
| 4162 | v->RequiredDPPCLK[i][j][k] = v->MinDPPCLKUsingSingleDPP[k] * (1.0 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); |
| 4163 | } else { |
| 4164 | v->MPCCombine[i][j][k] = true1; |
| 4165 | v->NoOfDPP[i][j][k] = 2; |
| 4166 | v->RequiredDPPCLK[i][j][k] = v->MinDPPCLKUsingSingleDPP[k] * (1.0 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) / 2.0; |
| 4167 | } |
| 4168 | v->RequiredDISPCLK[i][j] = dml_max(v->RequiredDISPCLK[i][j], v->PlaneRequiredDISPCLK); |
| 4169 | if ((v->MinDPPCLKUsingSingleDPP[k] / v->NoOfDPP[i][j][k] * (1.0 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) |
| 4170 | > v->MaxDppclkRoundedDownToDFSGranularity) |
| 4171 | || (v->PlaneRequiredDISPCLK > v->MaxDispclkRoundedDownToDFSGranularity)) { |
| 4172 | v->DISPCLK_DPPCLK_Support[i][j] = false0; |
| 4173 | } |
| 4174 | if (mode_lib->project == DML_PROJECT_DCN315 && v->DETSizeOverride[k] > DCN3_15_MAX_DET_SIZE384 && v->NoOfDPP[i][j][k] < 2) { |
| 4175 | v->MPCCombine[i][j][k] = true1; |
| 4176 | v->NoOfDPP[i][j][k] = 2; |
| 4177 | } |
| 4178 | } |
| 4179 | v->TotalNumberOfActiveDPP[i][j] = 0; |
| 4180 | v->TotalNumberOfSingleDPPPlanes[i][j] = 0; |
| 4181 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 4182 | v->TotalNumberOfActiveDPP[i][j] = v->TotalNumberOfActiveDPP[i][j] + v->NoOfDPP[i][j][k]; |
| 4183 | if (v->NoOfDPP[i][j][k] == 1) |
| 4184 | v->TotalNumberOfSingleDPPPlanes[i][j] = v->TotalNumberOfSingleDPPPlanes[i][j] + 1; |
| 4185 | if (v->SourcePixelFormat[k] == dm_420_8 || v->SourcePixelFormat[k] == dm_420_10 |
| 4186 | || v->SourcePixelFormat[k] == dm_420_12 || v->SourcePixelFormat[k] == dm_rgbe_alpha) |
| 4187 | NoChroma = false0; |
| 4188 | } |
| 4189 | |
| 4190 | // UPTO |
| 4191 | if (j == 1 && v->WhenToDoMPCCombine != dm_mpc_never |
| 4192 | && !UnboundedRequest(v->UseUnboundedRequesting, v->TotalNumberOfActiveDPP[i][j], NoChroma, v->Output[0])) { |
| 4193 | while (!(v->TotalNumberOfActiveDPP[i][j] >= v->MaxNumDPP || v->TotalNumberOfSingleDPPPlanes[i][j] == 0)) { |
| 4194 | double BWOfNonSplitPlaneOfMaximumBandwidth; |
| 4195 | unsigned int NumberOfNonSplitPlaneOfMaximumBandwidth; |
| 4196 | BWOfNonSplitPlaneOfMaximumBandwidth = 0; |
| 4197 | NumberOfNonSplitPlaneOfMaximumBandwidth = 0; |
| 4198 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 4199 | if (v->ReadBandwidthLuma[k] + v->ReadBandwidthChroma[k] > BWOfNonSplitPlaneOfMaximumBandwidth |
| 4200 | && v->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_disabled && v->MPCCombine[i][j][k] == false0) { |
| 4201 | BWOfNonSplitPlaneOfMaximumBandwidth = v->ReadBandwidthLuma[k] + v->ReadBandwidthChroma[k]; |
| 4202 | NumberOfNonSplitPlaneOfMaximumBandwidth = k; |
| 4203 | } |
| 4204 | } |
| 4205 | v->MPCCombine[i][j][NumberOfNonSplitPlaneOfMaximumBandwidth] = true1; |
| 4206 | v->NoOfDPP[i][j][NumberOfNonSplitPlaneOfMaximumBandwidth] = 2; |
| 4207 | v->RequiredDPPCLK[i][j][NumberOfNonSplitPlaneOfMaximumBandwidth] = |
| 4208 | v->MinDPPCLKUsingSingleDPP[NumberOfNonSplitPlaneOfMaximumBandwidth] |
| 4209 | * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100) / 2; |
| 4210 | v->TotalNumberOfActiveDPP[i][j] = v->TotalNumberOfActiveDPP[i][j] + 1; |
| 4211 | v->TotalNumberOfSingleDPPPlanes[i][j] = v->TotalNumberOfSingleDPPPlanes[i][j] - 1; |
| 4212 | } |
| 4213 | } |
| 4214 | if (v->TotalNumberOfActiveDPP[i][j] > v->MaxNumDPP) { |
| 4215 | v->RequiredDISPCLK[i][j] = 0.0; |
| 4216 | v->DISPCLK_DPPCLK_Support[i][j] = true1; |
| 4217 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 4218 | v->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_disabled; |
| 4219 | if (v->SingleDPPViewportSizeSupportPerPlane[k] == false0 && v->WhenToDoMPCCombine != dm_mpc_never) { |
| 4220 | v->MPCCombine[i][j][k] = true1; |
| 4221 | v->NoOfDPP[i][j][k] = 2; |
| 4222 | v->RequiredDPPCLK[i][j][k] = v->MinDPPCLKUsingSingleDPP[k] |
| 4223 | * (1.0 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) / 2.0; |
| 4224 | } else { |
| 4225 | v->MPCCombine[i][j][k] = false0; |
| 4226 | v->NoOfDPP[i][j][k] = 1; |
| 4227 | v->RequiredDPPCLK[i][j][k] = v->MinDPPCLKUsingSingleDPP[k] |
| 4228 | * (1.0 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); |
| 4229 | } |
| 4230 | if (!(v->MaxDispclk[i] == v->MaxDispclk[v->soc.num_states - 1] |
| 4231 | && v->MaxDppclk[i] == v->MaxDppclk[v->soc.num_states - 1])) { |
| 4232 | v->PlaneRequiredDISPCLK = v->PixelClock[k] * (1.0 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) |
| 4233 | * (1.0 + v->DISPCLKRampingMargin / 100.0); |
| 4234 | } else { |
| 4235 | v->PlaneRequiredDISPCLK = v->PixelClock[k] * (1.0 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); |
| 4236 | } |
| 4237 | v->RequiredDISPCLK[i][j] = dml_max(v->RequiredDISPCLK[i][j], v->PlaneRequiredDISPCLK); |
| 4238 | if ((v->MinDPPCLKUsingSingleDPP[k] / v->NoOfDPP[i][j][k] * (1.0 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) |
| 4239 | > v->MaxDppclkRoundedDownToDFSGranularity) |
| 4240 | || (v->PlaneRequiredDISPCLK > v->MaxDispclkRoundedDownToDFSGranularity)) { |
| 4241 | v->DISPCLK_DPPCLK_Support[i][j] = false0; |
| 4242 | } |
| 4243 | } |
| 4244 | v->TotalNumberOfActiveDPP[i][j] = 0.0; |
| 4245 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 4246 | v->TotalNumberOfActiveDPP[i][j] = v->TotalNumberOfActiveDPP[i][j] + v->NoOfDPP[i][j][k]; |
| 4247 | } |
| 4248 | } |
| 4249 | v->RequiredDISPCLK[i][j] = dml_max(v->RequiredDISPCLK[i][j], v->WritebackRequiredDISPCLK); |
| 4250 | if (v->MaxDispclkRoundedDownToDFSGranularity < v->WritebackRequiredDISPCLK) { |
| 4251 | v->DISPCLK_DPPCLK_Support[i][j] = false0; |
| 4252 | } |
| 4253 | } |
| 4254 | } |
| 4255 | |
| 4256 | /*Total Available Pipes Support Check*/ |
| 4257 | |
| 4258 | for (i = 0; i < v->soc.num_states; i++) { |
| 4259 | for (j = 0; j < 2; j++) { |
| 4260 | if (v->TotalNumberOfActiveDPP[i][j] <= v->MaxNumDPP) { |
| 4261 | v->TotalAvailablePipesSupport[i][j] = true1; |
| 4262 | } else { |
| 4263 | v->TotalAvailablePipesSupport[i][j] = false0; |
| 4264 | } |
| 4265 | } |
| 4266 | } |
| 4267 | /*Display IO and DSC Support Check*/ |
| 4268 | |
| 4269 | v->NonsupportedDSCInputBPC = false0; |
| 4270 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 4271 | if (!(v->DSCInputBitPerComponent[k] == 12.0 || v->DSCInputBitPerComponent[k] == 10.0 || v->DSCInputBitPerComponent[k] == 8.0) |
| 4272 | || v->DSCInputBitPerComponent[k] > v->MaximumDSCBitsPerComponent) { |
| 4273 | v->NonsupportedDSCInputBPC = true1; |
| 4274 | } |
| 4275 | } |
| 4276 | |
| 4277 | /*Number Of DSC Slices*/ |
| 4278 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 4279 | if (v->BlendingAndTiming[k] == k) { |
| 4280 | if (v->PixelClockBackEnd[k] > 3200) { |
| 4281 | v->NumberOfDSCSlices[k] = dml_ceil(v->PixelClockBackEnd[k] / 400.0, 4.0); |
| 4282 | } else if (v->PixelClockBackEnd[k] > 1360) { |
| 4283 | v->NumberOfDSCSlices[k] = 8; |
| 4284 | } else if (v->PixelClockBackEnd[k] > 680) { |
| 4285 | v->NumberOfDSCSlices[k] = 4; |
| 4286 | } else if (v->PixelClockBackEnd[k] > 340) { |
| 4287 | v->NumberOfDSCSlices[k] = 2; |
| 4288 | } else { |
| 4289 | v->NumberOfDSCSlices[k] = 1; |
| 4290 | } |
| 4291 | } else { |
| 4292 | v->NumberOfDSCSlices[k] = 0; |
| 4293 | } |
| 4294 | } |
| 4295 | |
| 4296 | for (i = 0; i < v->soc.num_states; i++) { |
| 4297 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 4298 | v->RequiresDSC[i][k] = false0; |
| 4299 | v->RequiresFEC[i][k] = false0; |
| 4300 | if (v->BlendingAndTiming[k] == k) { |
| 4301 | if (v->Output[k] == dm_hdmi) { |
| 4302 | v->RequiresDSC[i][k] = false0; |
| 4303 | v->RequiresFEC[i][k] = false0; |
| 4304 | v->OutputBppPerState[i][k] = TruncToValidBPP( |
| 4305 | dml_min(600.0, v->PHYCLKPerState[i]) * 10, |
| 4306 | 3, |
| 4307 | v->HTotal[k], |
| 4308 | v->HActive[k], |
| 4309 | v->PixelClockBackEnd[k], |
| 4310 | v->ForcedOutputLinkBPP[k], |
| 4311 | false0, |
| 4312 | v->Output[k], |
| 4313 | v->OutputFormat[k], |
| 4314 | v->DSCInputBitPerComponent[k], |
| 4315 | v->NumberOfDSCSlices[k], |
| 4316 | v->AudioSampleRate[k], |
| 4317 | v->AudioSampleLayout[k], |
| 4318 | v->ODMCombineEnablePerState[i][k]); |
| 4319 | } else if (v->Output[k] == dm_dp || v->Output[k] == dm_edp || v->Output[k] == dm_dp2p0) { |
| 4320 | if (v->DSCEnable[k] == true1) { |
| 4321 | v->RequiresDSC[i][k] = true1; |
| 4322 | v->LinkDSCEnable = true1; |
| 4323 | if (v->Output[k] == dm_dp || v->Output[k] == dm_dp2p0) { |
| 4324 | v->RequiresFEC[i][k] = true1; |
| 4325 | } else { |
| 4326 | v->RequiresFEC[i][k] = false0; |
| 4327 | } |
| 4328 | } else { |
| 4329 | v->RequiresDSC[i][k] = false0; |
| 4330 | v->LinkDSCEnable = false0; |
| 4331 | if (v->Output[k] == dm_dp2p0) { |
| 4332 | v->RequiresFEC[i][k] = true1; |
| 4333 | } else { |
| 4334 | v->RequiresFEC[i][k] = false0; |
| 4335 | } |
| 4336 | } |
| 4337 | if (v->Output[k] == dm_dp2p0) { |
| 4338 | v->Outbpp = BPP_INVALID0; |
| 4339 | if ((v->OutputLinkDPRate[k] == dm_dp_rate_na || v->OutputLinkDPRate[k] == dm_dp_rate_uhbr10) && |
| 4340 | v->PHYCLKD18PerState[k] >= 10000.0 / 18.0) { |
| 4341 | v->Outbpp = TruncToValidBPP( |
| 4342 | (1.0 - v->Downspreading / 100.0) * 10000, |
| 4343 | v->OutputLinkDPLanes[k], |
| 4344 | v->HTotal[k], |
| 4345 | v->HActive[k], |
| 4346 | v->PixelClockBackEnd[k], |
| 4347 | v->ForcedOutputLinkBPP[k], |
| 4348 | v->LinkDSCEnable, |
| 4349 | v->Output[k], |
| 4350 | v->OutputFormat[k], |
| 4351 | v->DSCInputBitPerComponent[k], |
| 4352 | v->NumberOfDSCSlices[k], |
| 4353 | v->AudioSampleRate[k], |
| 4354 | v->AudioSampleLayout[k], |
| 4355 | v->ODMCombineEnablePerState[i][k]); |
| 4356 | if (v->Outbpp == BPP_INVALID0 && v->PHYCLKD18PerState[k] < 13500.0 / 18.0 && |
| 4357 | v->DSCEnable[k] == true1 && v->ForcedOutputLinkBPP[k] == 0) { |
| 4358 | v->RequiresDSC[i][k] = true1; |
| 4359 | v->LinkDSCEnable = true1; |
| 4360 | v->Outbpp = TruncToValidBPP( |
| 4361 | (1.0 - v->Downspreading / 100.0) * 10000, |
| 4362 | v->OutputLinkDPLanes[k], |
| 4363 | v->HTotal[k], |
| 4364 | v->HActive[k], |
| 4365 | v->PixelClockBackEnd[k], |
| 4366 | v->ForcedOutputLinkBPP[k], |
| 4367 | v->LinkDSCEnable, |
| 4368 | v->Output[k], |
| 4369 | v->OutputFormat[k], |
| 4370 | v->DSCInputBitPerComponent[k], |
| 4371 | v->NumberOfDSCSlices[k], |
| 4372 | v->AudioSampleRate[k], |
| 4373 | v->AudioSampleLayout[k], |
| 4374 | v->ODMCombineEnablePerState[i][k]); |
| 4375 | } |
| 4376 | v->OutputBppPerState[i][k] = v->Outbpp; |
| 4377 | // TODO: Need some other way to handle this nonsense |
| 4378 | // v->OutputTypeAndRatePerState[i][k] = v->Output[k] & " UHBR10" |
| 4379 | } |
| 4380 | if (v->Outbpp == BPP_INVALID0 && |
| 4381 | (v->OutputLinkDPRate[k] == dm_dp_rate_na || v->OutputLinkDPRate[k] == dm_dp_rate_uhbr13p5) && |
| 4382 | v->PHYCLKD18PerState[k] >= 13500.0 / 18.0) { |
| 4383 | v->Outbpp = TruncToValidBPP( |
| 4384 | (1.0 - v->Downspreading / 100.0) * 13500, |
| 4385 | v->OutputLinkDPLanes[k], |
| 4386 | v->HTotal[k], |
| 4387 | v->HActive[k], |
| 4388 | v->PixelClockBackEnd[k], |
| 4389 | v->ForcedOutputLinkBPP[k], |
| 4390 | v->LinkDSCEnable, |
| 4391 | v->Output[k], |
| 4392 | v->OutputFormat[k], |
| 4393 | v->DSCInputBitPerComponent[k], |
| 4394 | v->NumberOfDSCSlices[k], |
| 4395 | v->AudioSampleRate[k], |
| 4396 | v->AudioSampleLayout[k], |
| 4397 | v->ODMCombineEnablePerState[i][k]); |
| 4398 | if (v->Outbpp == BPP_INVALID0 && v->PHYCLKD18PerState[k] < 20000.0 / 18.0 && |
| 4399 | v->DSCEnable[k] == true1 && v->ForcedOutputLinkBPP[k] == 0) { |
| 4400 | v->RequiresDSC[i][k] = true1; |
| 4401 | v->LinkDSCEnable = true1; |
| 4402 | v->Outbpp = TruncToValidBPP( |
| 4403 | (1.0 - v->Downspreading / 100.0) * 13500, |
| 4404 | v->OutputLinkDPLanes[k], |
| 4405 | v->HTotal[k], |
| 4406 | v->HActive[k], |
| 4407 | v->PixelClockBackEnd[k], |
| 4408 | v->ForcedOutputLinkBPP[k], |
| 4409 | v->LinkDSCEnable, |
| 4410 | v->Output[k], |
| 4411 | v->OutputFormat[k], |
| 4412 | v->DSCInputBitPerComponent[k], |
| 4413 | v->NumberOfDSCSlices[k], |
| 4414 | v->AudioSampleRate[k], |
| 4415 | v->AudioSampleLayout[k], |
| 4416 | v->ODMCombineEnablePerState[i][k]); |
| 4417 | } |
| 4418 | v->OutputBppPerState[i][k] = v->Outbpp; |
| 4419 | // TODO: Need some other way to handle this nonsense |
| 4420 | // v->OutputTypeAndRatePerState[i][k] = v->Output[k] & " UHBR13p5" |
| 4421 | } |
| 4422 | if (v->Outbpp == BPP_INVALID0 && |
| 4423 | (v->OutputLinkDPRate[k] == dm_dp_rate_na || v->OutputLinkDPRate[k] == dm_dp_rate_uhbr20) && |
| 4424 | v->PHYCLKD18PerState[k] >= 20000.0 / 18.0) { |
| 4425 | v->Outbpp = TruncToValidBPP( |
| 4426 | (1.0 - v->Downspreading / 100.0) * 20000, |
| 4427 | v->OutputLinkDPLanes[k], |
| 4428 | v->HTotal[k], |
| 4429 | v->HActive[k], |
| 4430 | v->PixelClockBackEnd[k], |
| 4431 | v->ForcedOutputLinkBPP[k], |
| 4432 | v->LinkDSCEnable, |
| 4433 | v->Output[k], |
| 4434 | v->OutputFormat[k], |
| 4435 | v->DSCInputBitPerComponent[k], |
| 4436 | v->NumberOfDSCSlices[k], |
| 4437 | v->AudioSampleRate[k], |
| 4438 | v->AudioSampleLayout[k], |
| 4439 | v->ODMCombineEnablePerState[i][k]); |
| 4440 | if (v->Outbpp == BPP_INVALID0 && v->DSCEnable[k] == true1 && |
| 4441 | v->ForcedOutputLinkBPP[k] == 0) { |
| 4442 | v->RequiresDSC[i][k] = true1; |
| 4443 | v->LinkDSCEnable = true1; |
| 4444 | v->Outbpp = TruncToValidBPP( |
| 4445 | (1.0 - v->Downspreading / 100.0) * 20000, |
| 4446 | v->OutputLinkDPLanes[k], |
| 4447 | v->HTotal[k], |
| 4448 | v->HActive[k], |
| 4449 | v->PixelClockBackEnd[k], |
| 4450 | v->ForcedOutputLinkBPP[k], |
| 4451 | v->LinkDSCEnable, |
| 4452 | v->Output[k], |
| 4453 | v->OutputFormat[k], |
| 4454 | v->DSCInputBitPerComponent[k], |
| 4455 | v->NumberOfDSCSlices[k], |
| 4456 | v->AudioSampleRate[k], |
| 4457 | v->AudioSampleLayout[k], |
| 4458 | v->ODMCombineEnablePerState[i][k]); |
| 4459 | } |
| 4460 | v->OutputBppPerState[i][k] = v->Outbpp; |
| 4461 | // TODO: Need some other way to handle this nonsense |
| 4462 | // v->OutputTypeAndRatePerState[i][k] = v->Output[k] & " UHBR20" |
| 4463 | } |
| 4464 | } else { |
| 4465 | v->Outbpp = BPP_INVALID0; |
| 4466 | if (v->PHYCLKPerState[i] >= 270.0) { |
| 4467 | v->Outbpp = TruncToValidBPP( |
| 4468 | (1.0 - v->Downspreading / 100.0) * 2700, |
| 4469 | v->OutputLinkDPLanes[k], |
| 4470 | v->HTotal[k], |
| 4471 | v->HActive[k], |
| 4472 | v->PixelClockBackEnd[k], |
| 4473 | v->ForcedOutputLinkBPP[k], |
| 4474 | v->LinkDSCEnable, |
| 4475 | v->Output[k], |
| 4476 | v->OutputFormat[k], |
| 4477 | v->DSCInputBitPerComponent[k], |
| 4478 | v->NumberOfDSCSlices[k], |
| 4479 | v->AudioSampleRate[k], |
| 4480 | v->AudioSampleLayout[k], |
| 4481 | v->ODMCombineEnablePerState[i][k]); |
| 4482 | v->OutputBppPerState[i][k] = v->Outbpp; |
| 4483 | // TODO: Need some other way to handle this nonsense |
| 4484 | // v->OutputTypeAndRatePerState[i][k] = v->Output[k] & " HBR" |
| 4485 | } |
| 4486 | if (v->Outbpp == BPP_INVALID0 && v->PHYCLKPerState[i] >= 540.0) { |
| 4487 | v->Outbpp = TruncToValidBPP( |
| 4488 | (1.0 - v->Downspreading / 100.0) * 5400, |
| 4489 | v->OutputLinkDPLanes[k], |
| 4490 | v->HTotal[k], |
| 4491 | v->HActive[k], |
| 4492 | v->PixelClockBackEnd[k], |
| 4493 | v->ForcedOutputLinkBPP[k], |
| 4494 | v->LinkDSCEnable, |
| 4495 | v->Output[k], |
| 4496 | v->OutputFormat[k], |
| 4497 | v->DSCInputBitPerComponent[k], |
| 4498 | v->NumberOfDSCSlices[k], |
| 4499 | v->AudioSampleRate[k], |
| 4500 | v->AudioSampleLayout[k], |
| 4501 | v->ODMCombineEnablePerState[i][k]); |
| 4502 | v->OutputBppPerState[i][k] = v->Outbpp; |
| 4503 | // TODO: Need some other way to handle this nonsense |
| 4504 | // v->OutputTypeAndRatePerState[i][k] = v->Output[k] & " HBR2" |
| 4505 | } |
| 4506 | if (v->Outbpp == BPP_INVALID0 && v->PHYCLKPerState[i] >= 810.0) { |
| 4507 | v->Outbpp = TruncToValidBPP( |
| 4508 | (1.0 - v->Downspreading / 100.0) * 8100, |
| 4509 | v->OutputLinkDPLanes[k], |
| 4510 | v->HTotal[k], |
| 4511 | v->HActive[k], |
| 4512 | v->PixelClockBackEnd[k], |
| 4513 | v->ForcedOutputLinkBPP[k], |
| 4514 | v->LinkDSCEnable, |
| 4515 | v->Output[k], |
| 4516 | v->OutputFormat[k], |
| 4517 | v->DSCInputBitPerComponent[k], |
| 4518 | v->NumberOfDSCSlices[k], |
| 4519 | v->AudioSampleRate[k], |
| 4520 | v->AudioSampleLayout[k], |
| 4521 | v->ODMCombineEnablePerState[i][k]); |
| 4522 | v->OutputBppPerState[i][k] = v->Outbpp; |
| 4523 | // TODO: Need some other way to handle this nonsense |
| 4524 | // v->OutputTypeAndRatePerState[i][k] = v->Output[k] & " HBR3" |
| 4525 | } |
| 4526 | } |
| 4527 | } |
| 4528 | } else { |
| 4529 | v->OutputBppPerState[i][k] = 0; |
| 4530 | } |
| 4531 | } |
| 4532 | } |
| 4533 | |
| 4534 | for (i = 0; i < v->soc.num_states; i++) { |
| 4535 | v->LinkCapacitySupport[i] = true1; |
| 4536 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 4537 | if (v->BlendingAndTiming[k] == k |
| 4538 | && (v->Output[k] == dm_dp || |
| 4539 | v->Output[k] == dm_edp || |
| 4540 | v->Output[k] == dm_hdmi) && v->OutputBppPerState[i][k] == 0) { |
| 4541 | v->LinkCapacitySupport[i] = false0; |
| 4542 | } |
| 4543 | } |
| 4544 | } |
| 4545 | |
| 4546 | // UPTO 2172 |
| 4547 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 4548 | if (v->BlendingAndTiming[k] == k |
| 4549 | && (v->Output[k] == dm_dp || |
| 4550 | v->Output[k] == dm_edp || |
| 4551 | v->Output[k] == dm_hdmi)) { |
| 4552 | if (v->OutputFormat[k] == dm_420 && v->Interlace[k] == 1 && v->ProgressiveToInterlaceUnitInOPP == true1) { |
| 4553 | P2IWith420 = true1; |
| 4554 | } |
| 4555 | if (v->DSCEnable[k] == true1 && v->OutputFormat[k] == dm_n422 |
| 4556 | && !v->DSC422NativeSupport) { |
| 4557 | DSC422NativeNotSupported = true1; |
| 4558 | } |
| 4559 | } |
| 4560 | } |
| 4561 | |
| 4562 | for (i = 0; i < v->soc.num_states; ++i) { |
| 4563 | v->ODMCombine4To1SupportCheckOK[i] = true1; |
| 4564 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 4565 | if (v->BlendingAndTiming[k] == k && v->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_4to1 |
| 4566 | && (v->ODMCombine4To1Supported == false0 || v->Output[k] == dm_dp || v->Output[k] == dm_edp |
| 4567 | || v->Output[k] == dm_hdmi)) { |
| 4568 | v->ODMCombine4To1SupportCheckOK[i] = false0; |
| 4569 | } |
| 4570 | } |
| 4571 | } |
| 4572 | |
| 4573 | /* Skip dscclk validation: as long as dispclk is supported, dscclk is also implicitly supported */ |
| 4574 | |
| 4575 | for (i = 0; i < v->soc.num_states; i++) { |
| 4576 | v->NotEnoughDSCUnits[i] = false0; |
| 4577 | v->TotalDSCUnitsRequired = 0.0; |
| 4578 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 4579 | if (v->RequiresDSC[i][k] == true1) { |
| 4580 | if (v->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_4to1) { |
| 4581 | v->TotalDSCUnitsRequired = v->TotalDSCUnitsRequired + 4.0; |
| 4582 | } else if (v->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_2to1) { |
| 4583 | v->TotalDSCUnitsRequired = v->TotalDSCUnitsRequired + 2.0; |
| 4584 | } else { |
| 4585 | v->TotalDSCUnitsRequired = v->TotalDSCUnitsRequired + 1.0; |
| 4586 | } |
| 4587 | } |
| 4588 | } |
| 4589 | if (v->TotalDSCUnitsRequired > v->NumberOfDSC) { |
| 4590 | v->NotEnoughDSCUnits[i] = true1; |
| 4591 | } |
| 4592 | } |
| 4593 | /*DSC Delay per state*/ |
| 4594 | |
| 4595 | for (i = 0; i < v->soc.num_states; i++) { |
| 4596 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 4597 | if (v->OutputBppPerState[i][k] == BPP_INVALID0) { |
| 4598 | v->BPP = 0.0; |
| 4599 | } else { |
| 4600 | v->BPP = v->OutputBppPerState[i][k]; |
| 4601 | } |
| 4602 | if (v->RequiresDSC[i][k] == true1 && v->BPP != 0.0) { |
| 4603 | if (v->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_disabled) { |
| 4604 | v->DSCDelayPerState[i][k] = dscceComputeDelay( |
| 4605 | v->DSCInputBitPerComponent[k], |
| 4606 | v->BPP, |
| 4607 | dml_ceil(1.0 * v->HActive[k] / v->NumberOfDSCSlices[k], 1.0), |
| 4608 | v->NumberOfDSCSlices[k], |
| 4609 | v->OutputFormat[k], |
| 4610 | v->Output[k]) + dscComputeDelay(v->OutputFormat[k], v->Output[k]); |
| 4611 | } else if (v->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_2to1) { |
| 4612 | v->DSCDelayPerState[i][k] = 2.0 |
| 4613 | * (dscceComputeDelay( |
| 4614 | v->DSCInputBitPerComponent[k], |
| 4615 | v->BPP, |
| 4616 | dml_ceil(1.0 * v->HActive[k] / v->NumberOfDSCSlices[k], 1.0), |
| 4617 | v->NumberOfDSCSlices[k] / 2, |
| 4618 | v->OutputFormat[k], |
| 4619 | v->Output[k]) + dscComputeDelay(v->OutputFormat[k], v->Output[k])); |
| 4620 | } else { |
| 4621 | v->DSCDelayPerState[i][k] = 4.0 |
| 4622 | * (dscceComputeDelay( |
| 4623 | v->DSCInputBitPerComponent[k], |
| 4624 | v->BPP, |
| 4625 | dml_ceil(1.0 * v->HActive[k] / v->NumberOfDSCSlices[k], 1.0), |
| 4626 | v->NumberOfDSCSlices[k] / 4, |
| 4627 | v->OutputFormat[k], |
| 4628 | v->Output[k]) + dscComputeDelay(v->OutputFormat[k], v->Output[k])); |
| 4629 | } |
| 4630 | v->DSCDelayPerState[i][k] = v->DSCDelayPerState[i][k] * v->PixelClock[k] / v->PixelClockBackEnd[k]; |
| 4631 | } else { |
| 4632 | v->DSCDelayPerState[i][k] = 0.0; |
| 4633 | } |
| 4634 | } |
| 4635 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 4636 | for (m = 0; m < v->NumberOfActivePlanes; m++) { |
| 4637 | if (v->BlendingAndTiming[k] == m && v->RequiresDSC[i][m] == true1) { |
| 4638 | v->DSCDelayPerState[i][k] = v->DSCDelayPerState[i][m]; |
| 4639 | } |
| 4640 | } |
| 4641 | } |
| 4642 | } |
| 4643 | |
| 4644 | //Calculate Swath, DET Configuration, DCFCLKDeepSleep |
| 4645 | // |
| 4646 | for (i = 0; i < v->soc.num_states; ++i) { |
| 4647 | for (j = 0; j <= 1; ++j) { |
| 4648 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 4649 | v->RequiredDPPCLKThisState[k] = v->RequiredDPPCLK[i][j][k]; |
| 4650 | v->NoOfDPPThisState[k] = v->NoOfDPP[i][j][k]; |
| 4651 | v->ODMCombineEnableThisState[k] = v->ODMCombineEnablePerState[i][k]; |
| 4652 | } |
| 4653 | |
| 4654 | if (v->NumberOfActivePlanes > 1 && mode_lib->project == DML_PROJECT_DCN315 && !v->DETSizeOverride[0]) |
| 4655 | PatchDETBufferSizeInKByte(v->NumberOfActivePlanes, v->NoOfDPPThisState, v->ip.config_return_buffer_size_in_kbytes, v->DETBufferSizeInKByte); |
| 4656 | CalculateSwathAndDETConfiguration( |
| 4657 | false0, |
| 4658 | v->NumberOfActivePlanes, |
| 4659 | mode_lib->project == DML_PROJECT_DCN315 && v->DETSizeOverride[0], |
| 4660 | v->DETBufferSizeInKByte, |
| 4661 | v->MaximumSwathWidthLuma, |
| 4662 | v->MaximumSwathWidthChroma, |
| 4663 | v->SourceScan, |
| 4664 | v->SourcePixelFormat, |
| 4665 | v->SurfaceTiling, |
| 4666 | v->ViewportWidth, |
| 4667 | v->ViewportHeight, |
| 4668 | v->SurfaceWidthY, |
| 4669 | v->SurfaceWidthC, |
| 4670 | v->SurfaceHeightY, |
| 4671 | v->SurfaceHeightC, |
| 4672 | v->Read256BlockHeightY, |
| 4673 | v->Read256BlockHeightC, |
| 4674 | v->Read256BlockWidthY, |
| 4675 | v->Read256BlockWidthC, |
| 4676 | v->ODMCombineEnableThisState, |
| 4677 | v->BlendingAndTiming, |
| 4678 | v->BytePerPixelY, |
| 4679 | v->BytePerPixelC, |
| 4680 | v->BytePerPixelInDETY, |
| 4681 | v->BytePerPixelInDETC, |
| 4682 | v->HActive, |
| 4683 | v->HRatio, |
| 4684 | v->HRatioChroma, |
| 4685 | v->NoOfDPPThisState, |
| 4686 | v->swath_width_luma_ub_this_state, |
| 4687 | v->swath_width_chroma_ub_this_state, |
| 4688 | v->SwathWidthYThisState, |
| 4689 | v->SwathWidthCThisState, |
| 4690 | v->SwathHeightYThisState, |
| 4691 | v->SwathHeightCThisState, |
| 4692 | v->DETBufferSizeYThisState, |
| 4693 | v->DETBufferSizeCThisState, |
| 4694 | v->dummystring, |
| 4695 | &v->ViewportSizeSupport[i][j]); |
| 4696 | |
| 4697 | CalculateDCFCLKDeepSleep( |
| 4698 | mode_lib, |
| 4699 | v->NumberOfActivePlanes, |
| 4700 | v->BytePerPixelY, |
| 4701 | v->BytePerPixelC, |
| 4702 | v->VRatio, |
| 4703 | v->VRatioChroma, |
| 4704 | v->SwathWidthYThisState, |
| 4705 | v->SwathWidthCThisState, |
| 4706 | v->NoOfDPPThisState, |
| 4707 | v->HRatio, |
| 4708 | v->HRatioChroma, |
| 4709 | v->PixelClock, |
| 4710 | v->PSCL_FACTOR, |
| 4711 | v->PSCL_FACTOR_CHROMA, |
| 4712 | v->RequiredDPPCLKThisState, |
| 4713 | v->ReadBandwidthLuma, |
| 4714 | v->ReadBandwidthChroma, |
| 4715 | v->ReturnBusWidth, |
| 4716 | &v->ProjectedDCFCLKDeepSleep[i][j]); |
| 4717 | |
| 4718 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 4719 | v->swath_width_luma_ub_all_states[i][j][k] = v->swath_width_luma_ub_this_state[k]; |
| 4720 | v->swath_width_chroma_ub_all_states[i][j][k] = v->swath_width_chroma_ub_this_state[k]; |
| 4721 | v->SwathWidthYAllStates[i][j][k] = v->SwathWidthYThisState[k]; |
| 4722 | v->SwathWidthCAllStates[i][j][k] = v->SwathWidthCThisState[k]; |
| 4723 | v->SwathHeightYAllStates[i][j][k] = v->SwathHeightYThisState[k]; |
| 4724 | v->SwathHeightCAllStates[i][j][k] = v->SwathHeightCThisState[k]; |
| 4725 | v->DETBufferSizeYAllStates[i][j][k] = v->DETBufferSizeYThisState[k]; |
| 4726 | v->DETBufferSizeCAllStates[i][j][k] = v->DETBufferSizeCThisState[k]; |
| 4727 | } |
| 4728 | } |
| 4729 | } |
| 4730 | |
| 4731 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 4732 | v->cursor_bw[k] = v->NumberOfCursors[k] * v->CursorWidth[k][0] * v->CursorBPP[k][0] / 8.0 |
| 4733 | / (v->HTotal[k] / v->PixelClock[k]) * v->VRatio[k]; |
| 4734 | } |
| 4735 | |
| 4736 | for (i = 0; i < v->soc.num_states; i++) { |
| 4737 | for (j = 0; j < 2; j++) { |
| 4738 | bool_Bool NotUrgentLatencyHiding[DC__NUM_DPP__MAX8]; |
| 4739 | |
| 4740 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 4741 | v->swath_width_luma_ub_this_state[k] = v->swath_width_luma_ub_all_states[i][j][k]; |
| 4742 | v->swath_width_chroma_ub_this_state[k] = v->swath_width_chroma_ub_all_states[i][j][k]; |
| 4743 | v->SwathWidthYThisState[k] = v->SwathWidthYAllStates[i][j][k]; |
| 4744 | v->SwathWidthCThisState[k] = v->SwathWidthCAllStates[i][j][k]; |
| 4745 | v->SwathHeightYThisState[k] = v->SwathHeightYAllStates[i][j][k]; |
| 4746 | v->SwathHeightCThisState[k] = v->SwathHeightCAllStates[i][j][k]; |
| 4747 | v->DETBufferSizeYThisState[k] = v->DETBufferSizeYAllStates[i][j][k]; |
| 4748 | v->DETBufferSizeCThisState[k] = v->DETBufferSizeCAllStates[i][j][k]; |
| 4749 | } |
| 4750 | |
| 4751 | v->TotalNumberOfDCCActiveDPP[i][j] = 0; |
| 4752 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 4753 | if (v->DCCEnable[k] == true1) { |
| 4754 | v->TotalNumberOfDCCActiveDPP[i][j] = v->TotalNumberOfDCCActiveDPP[i][j] + v->NoOfDPP[i][j][k]; |
| 4755 | } |
| 4756 | } |
| 4757 | |
| 4758 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 4759 | if (v->SourcePixelFormat[k] == dm_420_8 || v->SourcePixelFormat[k] == dm_420_10 |
| 4760 | || v->SourcePixelFormat[k] == dm_420_12 || v->SourcePixelFormat[k] == dm_rgbe_alpha) { |
| 4761 | |
| 4762 | if ((v->SourcePixelFormat[k] == dm_420_10 || v->SourcePixelFormat[k] == dm_420_12) |
| 4763 | && v->SourceScan[k] != dm_vert) { |
| 4764 | v->PTEBufferSizeInRequestsForLuma = (v->PTEBufferSizeInRequestsLuma + v->PTEBufferSizeInRequestsChroma) |
| 4765 | / 2; |
| 4766 | v->PTEBufferSizeInRequestsForChroma = v->PTEBufferSizeInRequestsForLuma; |
| 4767 | } else { |
| 4768 | v->PTEBufferSizeInRequestsForLuma = v->PTEBufferSizeInRequestsLuma; |
| 4769 | v->PTEBufferSizeInRequestsForChroma = v->PTEBufferSizeInRequestsChroma; |
| 4770 | } |
| 4771 | |
| 4772 | v->PDEAndMetaPTEBytesPerFrameC = CalculateVMAndRowBytes( |
| 4773 | mode_lib, |
| 4774 | v->DCCEnable[k], |
| 4775 | v->Read256BlockHeightC[k], |
| 4776 | v->Read256BlockWidthC[k], |
| 4777 | v->SourcePixelFormat[k], |
| 4778 | v->SurfaceTiling[k], |
| 4779 | v->BytePerPixelC[k], |
| 4780 | v->SourceScan[k], |
| 4781 | v->SwathWidthCThisState[k], |
| 4782 | v->ViewportHeightChroma[k], |
| 4783 | v->GPUVMEnable, |
| 4784 | v->HostVMEnable, |
| 4785 | v->HostVMMaxNonCachedPageTableLevels, |
| 4786 | v->GPUVMMinPageSize, |
| 4787 | v->HostVMMinPageSize, |
| 4788 | v->PTEBufferSizeInRequestsForChroma, |
| 4789 | v->PitchC[k], |
| 4790 | 0.0, |
| 4791 | &v->MacroTileWidthC[k], |
| 4792 | &v->MetaRowBytesC, |
| 4793 | &v->DPTEBytesPerRowC, |
| 4794 | &v->PTEBufferSizeNotExceededC[i][j][k], |
| 4795 | &v->dummyinteger7, |
| 4796 | &v->dpte_row_height_chroma[k], |
| 4797 | &v->dummyinteger28, |
| 4798 | &v->dummyinteger26, |
| 4799 | &v->dummyinteger23, |
| 4800 | &v->meta_row_height_chroma[k], |
| 4801 | &v->dummyinteger8, |
| 4802 | &v->dummyinteger9, |
| 4803 | &v->dummyinteger19, |
| 4804 | &v->dummyinteger20, |
| 4805 | &v->dummyinteger17, |
| 4806 | &v->dummyinteger10, |
| 4807 | &v->dummyinteger11); |
| 4808 | |
| 4809 | v->PrefetchLinesC[i][j][k] = CalculatePrefetchSourceLines( |
| 4810 | mode_lib, |
| 4811 | v->VRatioChroma[k], |
| 4812 | v->VTAPsChroma[k], |
| 4813 | v->Interlace[k], |
| 4814 | v->ProgressiveToInterlaceUnitInOPP, |
| 4815 | v->SwathHeightCThisState[k], |
| 4816 | v->ViewportYStartC[k], |
| 4817 | &v->PrefillC[k], |
| 4818 | &v->MaxNumSwC[k]); |
| 4819 | } else { |
| 4820 | v->PTEBufferSizeInRequestsForLuma = v->PTEBufferSizeInRequestsLuma + v->PTEBufferSizeInRequestsChroma; |
| 4821 | v->PTEBufferSizeInRequestsForChroma = 0; |
| 4822 | v->PDEAndMetaPTEBytesPerFrameC = 0.0; |
| 4823 | v->MetaRowBytesC = 0.0; |
| 4824 | v->DPTEBytesPerRowC = 0.0; |
| 4825 | v->PrefetchLinesC[i][j][k] = 0.0; |
| 4826 | v->PTEBufferSizeNotExceededC[i][j][k] = true1; |
| 4827 | } |
| 4828 | v->PDEAndMetaPTEBytesPerFrameY = CalculateVMAndRowBytes( |
| 4829 | mode_lib, |
| 4830 | v->DCCEnable[k], |
| 4831 | v->Read256BlockHeightY[k], |
| 4832 | v->Read256BlockWidthY[k], |
| 4833 | v->SourcePixelFormat[k], |
| 4834 | v->SurfaceTiling[k], |
| 4835 | v->BytePerPixelY[k], |
| 4836 | v->SourceScan[k], |
| 4837 | v->SwathWidthYThisState[k], |
| 4838 | v->ViewportHeight[k], |
| 4839 | v->GPUVMEnable, |
| 4840 | v->HostVMEnable, |
| 4841 | v->HostVMMaxNonCachedPageTableLevels, |
| 4842 | v->GPUVMMinPageSize, |
| 4843 | v->HostVMMinPageSize, |
| 4844 | v->PTEBufferSizeInRequestsForLuma, |
| 4845 | v->PitchY[k], |
| 4846 | v->DCCMetaPitchY[k], |
| 4847 | &v->MacroTileWidthY[k], |
| 4848 | &v->MetaRowBytesY, |
| 4849 | &v->DPTEBytesPerRowY, |
| 4850 | &v->PTEBufferSizeNotExceededY[i][j][k], |
| 4851 | &v->dummyinteger7, |
| 4852 | &v->dpte_row_height[k], |
| 4853 | &v->dummyinteger29, |
| 4854 | &v->dummyinteger27, |
| 4855 | &v->dummyinteger24, |
| 4856 | &v->meta_row_height[k], |
| 4857 | &v->dummyinteger25, |
| 4858 | &v->dpte_group_bytes[k], |
| 4859 | &v->dummyinteger21, |
| 4860 | &v->dummyinteger22, |
| 4861 | &v->dummyinteger18, |
| 4862 | &v->dummyinteger5, |
| 4863 | &v->dummyinteger6); |
| 4864 | v->PrefetchLinesY[i][j][k] = CalculatePrefetchSourceLines( |
| 4865 | mode_lib, |
| 4866 | v->VRatio[k], |
| 4867 | v->vtaps[k], |
| 4868 | v->Interlace[k], |
| 4869 | v->ProgressiveToInterlaceUnitInOPP, |
| 4870 | v->SwathHeightYThisState[k], |
| 4871 | v->ViewportYStartY[k], |
| 4872 | &v->PrefillY[k], |
| 4873 | &v->MaxNumSwY[k]); |
| 4874 | v->PDEAndMetaPTEBytesPerFrame[i][j][k] = v->PDEAndMetaPTEBytesPerFrameY + v->PDEAndMetaPTEBytesPerFrameC; |
| 4875 | v->MetaRowBytes[i][j][k] = v->MetaRowBytesY + v->MetaRowBytesC; |
| 4876 | v->DPTEBytesPerRow[i][j][k] = v->DPTEBytesPerRowY + v->DPTEBytesPerRowC; |
| 4877 | |
| 4878 | CalculateRowBandwidth( |
| 4879 | v->GPUVMEnable, |
| 4880 | v->SourcePixelFormat[k], |
| 4881 | v->VRatio[k], |
| 4882 | v->VRatioChroma[k], |
| 4883 | v->DCCEnable[k], |
| 4884 | v->HTotal[k] / v->PixelClock[k], |
| 4885 | v->MetaRowBytesY, |
| 4886 | v->MetaRowBytesC, |
| 4887 | v->meta_row_height[k], |
| 4888 | v->meta_row_height_chroma[k], |
| 4889 | v->DPTEBytesPerRowY, |
| 4890 | v->DPTEBytesPerRowC, |
| 4891 | v->dpte_row_height[k], |
| 4892 | v->dpte_row_height_chroma[k], |
| 4893 | &v->meta_row_bandwidth[i][j][k], |
| 4894 | &v->dpte_row_bandwidth[i][j][k]); |
| 4895 | } |
| 4896 | /*DCCMetaBufferSizeSupport(i, j) = True |
| 4897 | For k = 0 To NumberOfActivePlanes - 1 |
| 4898 | If MetaRowBytes(i, j, k) > 24064 Then |
| 4899 | DCCMetaBufferSizeSupport(i, j) = False |
| 4900 | End If |
| 4901 | Next k*/ |
| 4902 | v->DCCMetaBufferSizeSupport[i][j] = true1; |
| 4903 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 4904 | if (v->MetaRowBytes[i][j][k] > 24064) |
| 4905 | v->DCCMetaBufferSizeSupport[i][j] = false0; |
| 4906 | } |
| 4907 | v->UrgLatency[i] = CalculateUrgentLatency( |
| 4908 | v->UrgentLatencyPixelDataOnly, |
| 4909 | v->UrgentLatencyPixelMixedWithVMData, |
| 4910 | v->UrgentLatencyVMDataOnly, |
| 4911 | v->DoUrgentLatencyAdjustment, |
| 4912 | v->UrgentLatencyAdjustmentFabricClockComponent, |
| 4913 | v->UrgentLatencyAdjustmentFabricClockReference, |
| 4914 | v->FabricClockPerState[i]); |
| 4915 | |
| 4916 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 4917 | CalculateUrgentBurstFactor( |
| 4918 | v->swath_width_luma_ub_this_state[k], |
| 4919 | v->swath_width_chroma_ub_this_state[k], |
| 4920 | v->SwathHeightYThisState[k], |
| 4921 | v->SwathHeightCThisState[k], |
| 4922 | v->HTotal[k] / v->PixelClock[k], |
| 4923 | v->UrgLatency[i], |
| 4924 | v->CursorBufferSize, |
| 4925 | v->CursorWidth[k][0], |
| 4926 | v->CursorBPP[k][0], |
| 4927 | v->VRatio[k], |
| 4928 | v->VRatioChroma[k], |
| 4929 | v->BytePerPixelInDETY[k], |
| 4930 | v->BytePerPixelInDETC[k], |
| 4931 | v->DETBufferSizeYThisState[k], |
| 4932 | v->DETBufferSizeCThisState[k], |
| 4933 | &v->UrgentBurstFactorCursor[k], |
| 4934 | &v->UrgentBurstFactorLuma[k], |
| 4935 | &v->UrgentBurstFactorChroma[k], |
| 4936 | &NotUrgentLatencyHiding[k]); |
| 4937 | } |
| 4938 | |
| 4939 | v->NotEnoughUrgentLatencyHidingA[i][j] = false0; |
| 4940 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 4941 | if (NotUrgentLatencyHiding[k]) { |
| 4942 | v->NotEnoughUrgentLatencyHidingA[i][j] = true1; |
| 4943 | } |
| 4944 | } |
| 4945 | |
| 4946 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 4947 | v->VActivePixelBandwidth[i][j][k] = v->ReadBandwidthLuma[k] * v->UrgentBurstFactorLuma[k] |
| 4948 | + v->ReadBandwidthChroma[k] * v->UrgentBurstFactorChroma[k]; |
| 4949 | v->VActiveCursorBandwidth[i][j][k] = v->cursor_bw[k] * v->UrgentBurstFactorCursor[k]; |
| 4950 | } |
| 4951 | |
| 4952 | v->TotalVActivePixelBandwidth[i][j] = 0; |
| 4953 | v->TotalVActiveCursorBandwidth[i][j] = 0; |
| 4954 | v->TotalMetaRowBandwidth[i][j] = 0; |
| 4955 | v->TotalDPTERowBandwidth[i][j] = 0; |
| 4956 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 4957 | v->TotalVActivePixelBandwidth[i][j] = v->TotalVActivePixelBandwidth[i][j] + v->VActivePixelBandwidth[i][j][k]; |
| 4958 | v->TotalVActiveCursorBandwidth[i][j] = v->TotalVActiveCursorBandwidth[i][j] + v->VActiveCursorBandwidth[i][j][k]; |
| 4959 | v->TotalMetaRowBandwidth[i][j] = v->TotalMetaRowBandwidth[i][j] + v->NoOfDPP[i][j][k] * v->meta_row_bandwidth[i][j][k]; |
| 4960 | v->TotalDPTERowBandwidth[i][j] = v->TotalDPTERowBandwidth[i][j] + v->NoOfDPP[i][j][k] * v->dpte_row_bandwidth[i][j][k]; |
| 4961 | } |
| 4962 | } |
| 4963 | } |
| 4964 | |
| 4965 | //Calculate Return BW |
| 4966 | for (i = 0; i < v->soc.num_states; ++i) { |
| 4967 | for (j = 0; j <= 1; ++j) { |
| 4968 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 4969 | if (v->BlendingAndTiming[k] == k) { |
| 4970 | if (v->WritebackEnable[k] == true1) { |
| 4971 | v->WritebackDelayTime[k] = v->WritebackLatency |
| 4972 | + CalculateWriteBackDelay( |
| 4973 | v->WritebackPixelFormat[k], |
| 4974 | v->WritebackHRatio[k], |
| 4975 | v->WritebackVRatio[k], |
| 4976 | v->WritebackVTaps[k], |
| 4977 | v->WritebackDestinationWidth[k], |
| 4978 | v->WritebackDestinationHeight[k], |
| 4979 | v->WritebackSourceHeight[k], |
| 4980 | v->HTotal[k]) / v->RequiredDISPCLK[i][j]; |
| 4981 | } else { |
| 4982 | v->WritebackDelayTime[k] = 0.0; |
| 4983 | } |
| 4984 | for (m = 0; m < v->NumberOfActivePlanes; m++) { |
| 4985 | if (v->BlendingAndTiming[m] == k && v->WritebackEnable[m] == true1) { |
| 4986 | v->WritebackDelayTime[k] = dml_max( |
| 4987 | v->WritebackDelayTime[k], |
| 4988 | v->WritebackLatency |
| 4989 | + CalculateWriteBackDelay( |
| 4990 | v->WritebackPixelFormat[m], |
| 4991 | v->WritebackHRatio[m], |
| 4992 | v->WritebackVRatio[m], |
| 4993 | v->WritebackVTaps[m], |
| 4994 | v->WritebackDestinationWidth[m], |
| 4995 | v->WritebackDestinationHeight[m], |
| 4996 | v->WritebackSourceHeight[m], |
| 4997 | v->HTotal[m]) / v->RequiredDISPCLK[i][j]); |
| 4998 | } |
| 4999 | } |
| 5000 | } |
| 5001 | } |
| 5002 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 5003 | for (m = 0; m < v->NumberOfActivePlanes; m++) { |
| 5004 | if (v->BlendingAndTiming[k] == m) { |
| 5005 | v->WritebackDelayTime[k] = v->WritebackDelayTime[m]; |
| 5006 | } |
| 5007 | } |
| 5008 | } |
| 5009 | v->MaxMaxVStartup[i][j] = 0; |
| 5010 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 5011 | v->MaximumVStartup[i][j][k] = |
| 5012 | (v->Interlace[k] && !v->ProgressiveToInterlaceUnitInOPP) ? |
| 5013 | dml_floor((v->VTotal[k] - v->VActive[k]) / 2.0, 1.0) : |
| 5014 | v->VTotal[k] - v->VActive[k] |
| 5015 | - dml_max( |
| 5016 | 1.0, |
| 5017 | dml_ceil( |
| 5018 | 1.0 * v->WritebackDelayTime[k] |
| 5019 | / (v->HTotal[k] |
| 5020 | / v->PixelClock[k]), |
| 5021 | 1.0)); |
| 5022 | if (v->MaximumVStartup[i][j][k] > 1023) |
| 5023 | v->MaximumVStartup[i][j][k] = 1023; |
| 5024 | v->MaxMaxVStartup[i][j] = dml_max(v->MaxMaxVStartup[i][j], v->MaximumVStartup[i][j][k]); |
| 5025 | } |
| 5026 | } |
| 5027 | } |
| 5028 | |
| 5029 | ReorderingBytes = v->NumberOfChannels |
| 5030 | * dml_max3( |
| 5031 | v->UrgentOutOfOrderReturnPerChannelPixelDataOnly, |
| 5032 | v->UrgentOutOfOrderReturnPerChannelPixelMixedWithVMData, |
| 5033 | v->UrgentOutOfOrderReturnPerChannelVMDataOnly); |
| 5034 | |
| 5035 | for (i = 0; i < v->soc.num_states; ++i) { |
| 5036 | for (j = 0; j <= 1; ++j) { |
| 5037 | v->DCFCLKState[i][j] = v->DCFCLKPerState[i]; |
| 5038 | } |
| 5039 | } |
| 5040 | |
| 5041 | if (v->UseMinimumRequiredDCFCLK == true1) |
| 5042 | UseMinimumDCFCLK(mode_lib, MaxPrefetchMode, ReorderingBytes); |
| 5043 | |
| 5044 | for (i = 0; i < v->soc.num_states; ++i) { |
| 5045 | for (j = 0; j <= 1; ++j) { |
| 5046 | double IdealFabricAndSDPPortBandwidthPerState = dml_min( |
| 5047 | v->ReturnBusWidth * v->DCFCLKState[i][j], |
| 5048 | v->FabricClockPerState[i] * v->FabricDatapathToDCNDataReturn); |
| 5049 | double IdealDRAMBandwidthPerState = v->DRAMSpeedPerState[i] * v->NumberOfChannels * v->DRAMChannelWidth; |
| 5050 | double PixelDataOnlyReturnBWPerState = dml_min( |
| 5051 | IdealFabricAndSDPPortBandwidthPerState * v->PercentOfIdealFabricAndSDPPortBWReceivedAfterUrgLatency / 100.0, |
| 5052 | IdealDRAMBandwidthPerState * v->PercentOfIdealDRAMBWReceivedAfterUrgLatencyPixelDataOnly / 100.0); |
| 5053 | double PixelMixedWithVMDataReturnBWPerState = dml_min( |
| 5054 | IdealFabricAndSDPPortBandwidthPerState * v->PercentOfIdealFabricAndSDPPortBWReceivedAfterUrgLatency / 100.0, |
| 5055 | IdealDRAMBandwidthPerState * v->PercentOfIdealDRAMBWReceivedAfterUrgLatencyPixelMixedWithVMData / 100.0); |
| 5056 | |
| 5057 | if (v->HostVMEnable != true1) { |
| 5058 | v->ReturnBWPerState[i][j] = PixelDataOnlyReturnBWPerState; |
| 5059 | } else { |
| 5060 | v->ReturnBWPerState[i][j] = PixelMixedWithVMDataReturnBWPerState; |
| 5061 | } |
| 5062 | } |
| 5063 | } |
| 5064 | |
| 5065 | //Re-ordering Buffer Support Check |
| 5066 | for (i = 0; i < v->soc.num_states; ++i) { |
| 5067 | for (j = 0; j <= 1; ++j) { |
| 5068 | if ((v->ROBBufferSizeInKByte - v->PixelChunkSizeInKByte) * 1024 / v->ReturnBWPerState[i][j] |
| 5069 | > (v->RoundTripPingLatencyCycles + __DML_ARB_TO_RET_DELAY__(7 + 95)) / v->DCFCLKState[i][j] + ReorderingBytes / v->ReturnBWPerState[i][j]) { |
| 5070 | v->ROBSupport[i][j] = true1; |
| 5071 | } else { |
| 5072 | v->ROBSupport[i][j] = false0; |
| 5073 | } |
| 5074 | } |
| 5075 | } |
| 5076 | |
| 5077 | //Vertical Active BW support check |
| 5078 | |
| 5079 | MaxTotalVActiveRDBandwidth = 0; |
| 5080 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 5081 | MaxTotalVActiveRDBandwidth = MaxTotalVActiveRDBandwidth + v->ReadBandwidthLuma[k] + v->ReadBandwidthChroma[k]; |
| 5082 | } |
| 5083 | |
| 5084 | for (i = 0; i < v->soc.num_states; ++i) { |
| 5085 | for (j = 0; j <= 1; ++j) { |
| 5086 | v->MaxTotalVerticalActiveAvailableBandwidth[i][j] = dml_min( |
| 5087 | dml_min( |
| 5088 | v->ReturnBusWidth * v->DCFCLKState[i][j], |
| 5089 | v->FabricClockPerState[i] * v->FabricDatapathToDCNDataReturn) |
| 5090 | * v->MaxAveragePercentOfIdealFabricAndSDPPortBWDisplayCanUseInNormalSystemOperation / 100, |
| 5091 | v->DRAMSpeedPerState[i] * v->NumberOfChannels * v->DRAMChannelWidth |
| 5092 | * v->MaxAveragePercentOfIdealDRAMBWDisplayCanUseInNormalSystemOperation / 100); |
| 5093 | |
| 5094 | if (MaxTotalVActiveRDBandwidth <= v->MaxTotalVerticalActiveAvailableBandwidth[i][j]) { |
| 5095 | v->TotalVerticalActiveBandwidthSupport[i][j] = true1; |
| 5096 | } else { |
| 5097 | v->TotalVerticalActiveBandwidthSupport[i][j] = false0; |
| 5098 | } |
| 5099 | } |
| 5100 | } |
| 5101 | |
| 5102 | v->UrgentLatency = CalculateUrgentLatency( |
| 5103 | v->UrgentLatencyPixelDataOnly, |
| 5104 | v->UrgentLatencyPixelMixedWithVMData, |
| 5105 | v->UrgentLatencyVMDataOnly, |
| 5106 | v->DoUrgentLatencyAdjustment, |
| 5107 | v->UrgentLatencyAdjustmentFabricClockComponent, |
| 5108 | v->UrgentLatencyAdjustmentFabricClockReference, |
| 5109 | v->FabricClock); |
| 5110 | //Prefetch Check |
| 5111 | for (i = 0; i < v->soc.num_states; ++i) { |
| 5112 | for (j = 0; j <= 1; ++j) { |
| 5113 | double VMDataOnlyReturnBWPerState; |
| 5114 | double HostVMInefficiencyFactor = 1; |
| 5115 | int NextPrefetchModeState = MinPrefetchMode; |
| 5116 | bool_Bool UnboundedRequestEnabledThisState = false0; |
| 5117 | int CompressedBufferSizeInkByteThisState = 0; |
| 5118 | double dummy; |
| 5119 | |
| 5120 | v->TimeCalc = 24 / v->ProjectedDCFCLKDeepSleep[i][j]; |
| 5121 | |
| 5122 | v->BandwidthWithoutPrefetchSupported[i][j] = true1; |
| 5123 | if (v->TotalVActivePixelBandwidth[i][j] + v->TotalVActiveCursorBandwidth[i][j] + v->TotalMetaRowBandwidth[i][j] |
| 5124 | + v->TotalDPTERowBandwidth[i][j] > v->ReturnBWPerState[i][j] || v->NotEnoughUrgentLatencyHidingA[i][j]) { |
| 5125 | v->BandwidthWithoutPrefetchSupported[i][j] = false0; |
| 5126 | } |
| 5127 | |
| 5128 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 5129 | v->NoOfDPPThisState[k] = v->NoOfDPP[i][j][k]; |
| 5130 | v->swath_width_luma_ub_this_state[k] = v->swath_width_luma_ub_all_states[i][j][k]; |
| 5131 | v->swath_width_chroma_ub_this_state[k] = v->swath_width_chroma_ub_all_states[i][j][k]; |
| 5132 | v->SwathWidthYThisState[k] = v->SwathWidthYAllStates[i][j][k]; |
| 5133 | v->SwathWidthCThisState[k] = v->SwathWidthCAllStates[i][j][k]; |
| 5134 | v->SwathHeightYThisState[k] = v->SwathHeightYAllStates[i][j][k]; |
| 5135 | v->SwathHeightCThisState[k] = v->SwathHeightCAllStates[i][j][k]; |
| 5136 | v->DETBufferSizeYThisState[k] = v->DETBufferSizeYAllStates[i][j][k]; |
| 5137 | v->DETBufferSizeCThisState[k] = v->DETBufferSizeCAllStates[i][j][k]; |
| 5138 | } |
| 5139 | |
| 5140 | VMDataOnlyReturnBWPerState = dml_min( |
| 5141 | dml_min( |
| 5142 | v->ReturnBusWidth * v->DCFCLKState[i][j], |
| 5143 | v->FabricClockPerState[i] * v->FabricDatapathToDCNDataReturn) |
| 5144 | * v->PercentOfIdealFabricAndSDPPortBWReceivedAfterUrgLatency / 100.0, |
| 5145 | v->DRAMSpeedPerState[i] * v->NumberOfChannels * v->DRAMChannelWidth |
| 5146 | * v->PercentOfIdealDRAMBWReceivedAfterUrgLatencyVMDataOnly / 100.0); |
| 5147 | if (v->GPUVMEnable && v->HostVMEnable) |
| 5148 | HostVMInefficiencyFactor = v->ReturnBWPerState[i][j] / VMDataOnlyReturnBWPerState; |
| 5149 | |
| 5150 | v->ExtraLatency = CalculateExtraLatency( |
| 5151 | v->RoundTripPingLatencyCycles, |
| 5152 | ReorderingBytes, |
| 5153 | v->DCFCLKState[i][j], |
| 5154 | v->TotalNumberOfActiveDPP[i][j], |
| 5155 | v->PixelChunkSizeInKByte, |
| 5156 | v->TotalNumberOfDCCActiveDPP[i][j], |
| 5157 | v->MetaChunkSize, |
| 5158 | v->ReturnBWPerState[i][j], |
| 5159 | v->GPUVMEnable, |
| 5160 | v->HostVMEnable, |
| 5161 | v->NumberOfActivePlanes, |
| 5162 | v->NoOfDPPThisState, |
| 5163 | v->dpte_group_bytes, |
| 5164 | HostVMInefficiencyFactor, |
| 5165 | v->HostVMMinPageSize, |
| 5166 | v->HostVMMaxNonCachedPageTableLevels); |
| 5167 | |
| 5168 | v->NextMaxVStartup = v->MaxMaxVStartup[i][j]; |
| 5169 | do { |
| 5170 | v->PrefetchModePerState[i][j] = NextPrefetchModeState; |
| 5171 | v->MaxVStartup = v->NextMaxVStartup; |
| 5172 | |
| 5173 | v->TWait = CalculateTWait( |
| 5174 | v->PrefetchModePerState[i][j], |
| 5175 | v->DRAMClockChangeLatency, |
| 5176 | v->UrgLatency[i], |
| 5177 | v->SREnterPlusExitTime); |
| 5178 | |
| 5179 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 5180 | CalculatePrefetchSchedulePerPlane(mode_lib, |
| 5181 | HostVMInefficiencyFactor, |
| 5182 | i, j, k); |
| 5183 | } |
| 5184 | |
| 5185 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 5186 | CalculateUrgentBurstFactor( |
| 5187 | v->swath_width_luma_ub_this_state[k], |
| 5188 | v->swath_width_chroma_ub_this_state[k], |
| 5189 | v->SwathHeightYThisState[k], |
| 5190 | v->SwathHeightCThisState[k], |
| 5191 | v->HTotal[k] / v->PixelClock[k], |
| 5192 | v->UrgentLatency, |
| 5193 | v->CursorBufferSize, |
| 5194 | v->CursorWidth[k][0], |
| 5195 | v->CursorBPP[k][0], |
| 5196 | v->VRatioPreY[i][j][k], |
| 5197 | v->VRatioPreC[i][j][k], |
| 5198 | v->BytePerPixelInDETY[k], |
| 5199 | v->BytePerPixelInDETC[k], |
| 5200 | v->DETBufferSizeYThisState[k], |
| 5201 | v->DETBufferSizeCThisState[k], |
| 5202 | &v->UrgentBurstFactorCursorPre[k], |
| 5203 | &v->UrgentBurstFactorLumaPre[k], |
| 5204 | &v->UrgentBurstFactorChromaPre[k], |
| 5205 | &v->NotUrgentLatencyHidingPre[k]); |
| 5206 | } |
| 5207 | |
| 5208 | v->MaximumReadBandwidthWithPrefetch = 0.0; |
| 5209 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 5210 | v->cursor_bw_pre[k] = v->NumberOfCursors[k] * v->CursorWidth[k][0] * v->CursorBPP[k][0] / 8.0 |
| 5211 | / (v->HTotal[k] / v->PixelClock[k]) * v->VRatioPreY[i][j][k]; |
| 5212 | |
| 5213 | v->MaximumReadBandwidthWithPrefetch = |
| 5214 | v->MaximumReadBandwidthWithPrefetch |
| 5215 | + dml_max3( |
| 5216 | v->VActivePixelBandwidth[i][j][k] |
| 5217 | + v->VActiveCursorBandwidth[i][j][k] |
| 5218 | + v->NoOfDPP[i][j][k] |
| 5219 | * (v->meta_row_bandwidth[i][j][k] |
| 5220 | + v->dpte_row_bandwidth[i][j][k]), |
| 5221 | v->NoOfDPP[i][j][k] * v->prefetch_vmrow_bw[k], |
| 5222 | v->NoOfDPP[i][j][k] |
| 5223 | * (v->RequiredPrefetchPixelDataBWLuma[i][j][k] |
| 5224 | * v->UrgentBurstFactorLumaPre[k] |
| 5225 | + v->RequiredPrefetchPixelDataBWChroma[i][j][k] |
| 5226 | * v->UrgentBurstFactorChromaPre[k]) |
| 5227 | + v->cursor_bw_pre[k] * v->UrgentBurstFactorCursorPre[k]); |
| 5228 | } |
| 5229 | |
| 5230 | v->NotEnoughUrgentLatencyHidingPre = false0; |
| 5231 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 5232 | if (v->NotUrgentLatencyHidingPre[k] == true1) { |
| 5233 | v->NotEnoughUrgentLatencyHidingPre = true1; |
| 5234 | } |
| 5235 | } |
| 5236 | |
| 5237 | v->PrefetchSupported[i][j] = true1; |
| 5238 | if (v->BandwidthWithoutPrefetchSupported[i][j] == false0 || v->MaximumReadBandwidthWithPrefetch > v->ReturnBWPerState[i][j] |
| 5239 | || v->NotEnoughUrgentLatencyHidingPre == 1) { |
| 5240 | v->PrefetchSupported[i][j] = false0; |
| 5241 | } |
| 5242 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 5243 | if (v->LineTimesForPrefetch[k] < 2.0 || v->LinesForMetaPTE[k] >= 32.0 || v->LinesForMetaAndDPTERow[k] >= 16.0 |
| 5244 | || v->NoTimeForPrefetch[i][j][k] == true1) { |
| 5245 | v->PrefetchSupported[i][j] = false0; |
| 5246 | } |
| 5247 | } |
| 5248 | |
| 5249 | v->DynamicMetadataSupported[i][j] = true1; |
| 5250 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 5251 | if (v->NoTimeForDynamicMetadata[i][j][k] == true1) { |
| 5252 | v->DynamicMetadataSupported[i][j] = false0; |
| 5253 | } |
| 5254 | } |
| 5255 | |
| 5256 | v->VRatioInPrefetchSupported[i][j] = true1; |
| 5257 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 5258 | if (v->VRatioPreY[i][j][k] > 4.0 || v->VRatioPreC[i][j][k] > 4.0 || v->NoTimeForPrefetch[i][j][k] == true1) { |
| 5259 | v->VRatioInPrefetchSupported[i][j] = false0; |
| 5260 | } |
| 5261 | } |
| 5262 | v->AnyLinesForVMOrRowTooLarge = false0; |
| 5263 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 5264 | if (v->LinesForMetaAndDPTERow[k] >= 16 || v->LinesForMetaPTE[k] >= 32) { |
| 5265 | v->AnyLinesForVMOrRowTooLarge = true1; |
| 5266 | } |
| 5267 | } |
| 5268 | |
| 5269 | v->NextPrefetchMode = v->NextPrefetchMode + 1; |
| 5270 | |
| 5271 | if (v->PrefetchSupported[i][j] == true1 && v->VRatioInPrefetchSupported[i][j] == true1) { |
| 5272 | v->BandwidthAvailableForImmediateFlip = v->ReturnBWPerState[i][j]; |
| 5273 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 5274 | v->BandwidthAvailableForImmediateFlip = v->BandwidthAvailableForImmediateFlip |
| 5275 | - dml_max( |
| 5276 | v->VActivePixelBandwidth[i][j][k] + v->VActiveCursorBandwidth[i][j][k], |
| 5277 | v->NoOfDPP[i][j][k] |
| 5278 | * (v->RequiredPrefetchPixelDataBWLuma[i][j][k] |
| 5279 | * v->UrgentBurstFactorLumaPre[k] |
| 5280 | + v->RequiredPrefetchPixelDataBWChroma[i][j][k] |
| 5281 | * v->UrgentBurstFactorChromaPre[k]) |
| 5282 | + v->cursor_bw_pre[k] * v->UrgentBurstFactorCursorPre[k]); |
| 5283 | } |
| 5284 | v->TotImmediateFlipBytes = 0.0; |
| 5285 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 5286 | v->TotImmediateFlipBytes = v->TotImmediateFlipBytes |
| 5287 | + v->NoOfDPP[i][j][k] * v->PDEAndMetaPTEBytesPerFrame[i][j][k] + v->MetaRowBytes[i][j][k] |
| 5288 | + v->DPTEBytesPerRow[i][j][k]; |
| 5289 | } |
| 5290 | |
| 5291 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 5292 | CalculateFlipSchedule( |
| 5293 | mode_lib, |
| 5294 | k, |
| 5295 | HostVMInefficiencyFactor, |
| 5296 | v->ExtraLatency, |
| 5297 | v->UrgLatency[i], |
| 5298 | v->PDEAndMetaPTEBytesPerFrame[i][j][k], |
| 5299 | v->MetaRowBytes[i][j][k], |
| 5300 | v->DPTEBytesPerRow[i][j][k]); |
| 5301 | } |
| 5302 | v->total_dcn_read_bw_with_flip = 0.0; |
| 5303 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 5304 | v->total_dcn_read_bw_with_flip = v->total_dcn_read_bw_with_flip |
| 5305 | + dml_max3( |
| 5306 | v->NoOfDPP[i][j][k] * v->prefetch_vmrow_bw[k], |
| 5307 | v->NoOfDPP[i][j][k] * v->final_flip_bw[k] + v->VActivePixelBandwidth[i][j][k] |
| 5308 | + v->VActiveCursorBandwidth[i][j][k], |
| 5309 | v->NoOfDPP[i][j][k] |
| 5310 | * (v->final_flip_bw[k] |
| 5311 | + v->RequiredPrefetchPixelDataBWLuma[i][j][k] |
| 5312 | * v->UrgentBurstFactorLumaPre[k] |
| 5313 | + v->RequiredPrefetchPixelDataBWChroma[i][j][k] |
| 5314 | * v->UrgentBurstFactorChromaPre[k]) |
| 5315 | + v->cursor_bw_pre[k] * v->UrgentBurstFactorCursorPre[k]); |
| 5316 | } |
| 5317 | v->ImmediateFlipSupportedForState[i][j] = true1; |
| 5318 | if (v->total_dcn_read_bw_with_flip > v->ReturnBWPerState[i][j]) { |
| 5319 | v->ImmediateFlipSupportedForState[i][j] = false0; |
| 5320 | } |
| 5321 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 5322 | if (v->ImmediateFlipSupportedForPipe[k] == false0) { |
| 5323 | v->ImmediateFlipSupportedForState[i][j] = false0; |
| 5324 | } |
| 5325 | } |
| 5326 | } else { |
| 5327 | v->ImmediateFlipSupportedForState[i][j] = false0; |
| 5328 | } |
| 5329 | |
| 5330 | if (v->MaxVStartup <= __DML_VBA_MIN_VSTARTUP__9 || v->AnyLinesForVMOrRowTooLarge == false0) { |
| 5331 | v->NextMaxVStartup = v->MaxMaxVStartup[i][j]; |
| 5332 | NextPrefetchModeState = NextPrefetchModeState + 1; |
| 5333 | } else { |
| 5334 | v->NextMaxVStartup = v->NextMaxVStartup - 1; |
| 5335 | } |
| 5336 | v->NextPrefetchMode = v->NextPrefetchMode + 1; |
| 5337 | } while (!((v->PrefetchSupported[i][j] == true1 && v->DynamicMetadataSupported[i][j] == true1 && v->VRatioInPrefetchSupported[i][j] == true1 |
| 5338 | && ((v->HostVMEnable == false0 && |
| 5339 | v->ImmediateFlipRequirement[0] != dm_immediate_flip_required) |
| 5340 | || v->ImmediateFlipSupportedForState[i][j] == true1)) |
| 5341 | || (v->NextMaxVStartup == v->MaxMaxVStartup[i][j] && NextPrefetchModeState > MaxPrefetchMode))); |
| 5342 | |
| 5343 | CalculateUnboundedRequestAndCompressedBufferSize( |
| 5344 | v->DETBufferSizeInKByte[0], |
| 5345 | v->ConfigReturnBufferSizeInKByte, |
| 5346 | v->UseUnboundedRequesting, |
| 5347 | v->TotalNumberOfActiveDPP[i][j], |
| 5348 | NoChroma, |
| 5349 | v->MaxNumDPP, |
| 5350 | v->CompressedBufferSegmentSizeInkByte, |
| 5351 | v->Output, |
| 5352 | &UnboundedRequestEnabledThisState, |
| 5353 | &CompressedBufferSizeInkByteThisState); |
| 5354 | |
| 5355 | CalculateWatermarksAndDRAMSpeedChangeSupport( |
| 5356 | mode_lib, |
| 5357 | v->PrefetchModePerState[i][j], |
| 5358 | v->DCFCLKState[i][j], |
| 5359 | v->ReturnBWPerState[i][j], |
| 5360 | v->UrgLatency[i], |
| 5361 | v->ExtraLatency, |
| 5362 | v->SOCCLKPerState[i], |
| 5363 | v->ProjectedDCFCLKDeepSleep[i][j], |
| 5364 | v->DETBufferSizeYThisState, |
| 5365 | v->DETBufferSizeCThisState, |
| 5366 | v->SwathHeightYThisState, |
| 5367 | v->SwathHeightCThisState, |
| 5368 | v->SwathWidthYThisState, |
| 5369 | v->SwathWidthCThisState, |
| 5370 | v->NoOfDPPThisState, |
| 5371 | v->BytePerPixelInDETY, |
| 5372 | v->BytePerPixelInDETC, |
| 5373 | UnboundedRequestEnabledThisState, |
| 5374 | CompressedBufferSizeInkByteThisState, |
| 5375 | &v->DRAMClockChangeSupport[i][j], |
| 5376 | &dummy, |
| 5377 | &dummy, |
| 5378 | &dummy, |
| 5379 | &dummy); |
| 5380 | } |
| 5381 | } |
| 5382 | |
| 5383 | /*PTE Buffer Size Check*/ |
| 5384 | for (i = 0; i < v->soc.num_states; i++) { |
| 5385 | for (j = 0; j < 2; j++) { |
| 5386 | v->PTEBufferSizeNotExceeded[i][j] = true1; |
| 5387 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 5388 | if (v->PTEBufferSizeNotExceededY[i][j][k] == false0 || v->PTEBufferSizeNotExceededC[i][j][k] == false0) { |
| 5389 | v->PTEBufferSizeNotExceeded[i][j] = false0; |
| 5390 | } |
| 5391 | } |
| 5392 | } |
| 5393 | } |
| 5394 | |
| 5395 | /*Cursor Support Check*/ |
| 5396 | v->CursorSupport = true1; |
| 5397 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 5398 | if (v->CursorWidth[k][0] > 0.0) { |
| 5399 | if (v->CursorBPP[k][0] == 64 && v->Cursor64BppSupport == false0) { |
| 5400 | v->CursorSupport = false0; |
| 5401 | } |
| 5402 | } |
| 5403 | } |
| 5404 | |
| 5405 | /*Valid Pitch Check*/ |
| 5406 | v->PitchSupport = true1; |
| 5407 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 5408 | v->AlignedYPitch[k] = dml_ceil(dml_max(v->PitchY[k], v->SurfaceWidthY[k]), v->MacroTileWidthY[k]); |
| 5409 | if (v->DCCEnable[k] == true1) { |
| 5410 | v->AlignedDCCMetaPitchY[k] = dml_ceil(dml_max(v->DCCMetaPitchY[k], v->SurfaceWidthY[k]), 64.0 * v->Read256BlockWidthY[k]); |
| 5411 | } else { |
| 5412 | v->AlignedDCCMetaPitchY[k] = v->DCCMetaPitchY[k]; |
| 5413 | } |
| 5414 | if (v->SourcePixelFormat[k] != dm_444_64 && v->SourcePixelFormat[k] != dm_444_32 && v->SourcePixelFormat[k] != dm_444_16 |
| 5415 | && v->SourcePixelFormat[k] != dm_mono_16 && v->SourcePixelFormat[k] != dm_rgbe |
| 5416 | && v->SourcePixelFormat[k] != dm_mono_8) { |
| 5417 | v->AlignedCPitch[k] = dml_ceil(dml_max(v->PitchC[k], v->SurfaceWidthC[k]), v->MacroTileWidthC[k]); |
| 5418 | if (v->DCCEnable[k] == true1) { |
| 5419 | v->AlignedDCCMetaPitchC[k] = dml_ceil( |
| 5420 | dml_max(v->DCCMetaPitchC[k], v->SurfaceWidthC[k]), |
| 5421 | 64.0 * v->Read256BlockWidthC[k]); |
| 5422 | } else { |
| 5423 | v->AlignedDCCMetaPitchC[k] = v->DCCMetaPitchC[k]; |
| 5424 | } |
| 5425 | } else { |
| 5426 | v->AlignedCPitch[k] = v->PitchC[k]; |
| 5427 | v->AlignedDCCMetaPitchC[k] = v->DCCMetaPitchC[k]; |
| 5428 | } |
| 5429 | if (v->AlignedYPitch[k] > v->PitchY[k] || v->AlignedCPitch[k] > v->PitchC[k] |
| 5430 | || v->AlignedDCCMetaPitchY[k] > v->DCCMetaPitchY[k] || v->AlignedDCCMetaPitchC[k] > v->DCCMetaPitchC[k]) { |
| 5431 | v->PitchSupport = false0; |
| 5432 | } |
| 5433 | } |
| 5434 | |
| 5435 | for (k = 0; k < v->NumberOfActivePlanes; k++) { |
| 5436 | if (v->ViewportWidth[k] > v->SurfaceWidthY[k] || v->ViewportHeight[k] > v->SurfaceHeightY[k]) { |
| 5437 | ViewportExceedsSurface = true1; |
| 5438 | if (v->SourcePixelFormat[k] != dm_444_64 && v->SourcePixelFormat[k] != dm_444_32 |
| 5439 | && v->SourcePixelFormat[k] != dm_444_16 && v->SourcePixelFormat[k] != dm_444_8 |
| 5440 | && v->SourcePixelFormat[k] != dm_rgbe) { |
| 5441 | if (v->ViewportWidthChroma[k] > v->SurfaceWidthC[k] |
| 5442 | || v->ViewportHeightChroma[k] > v->SurfaceHeightC[k]) { |
| 5443 | ViewportExceedsSurface = true1; |
| 5444 | } |
| 5445 | } |
| 5446 | } |
| 5447 | } |
| 5448 | |
| 5449 | /*Mode Support, Voltage State and SOC Configuration*/ |
| 5450 | for (i = v->soc.num_states - 1; i >= 0; i--) { |
| 5451 | for (j = 0; j < 2; j++) { |
| 5452 | if (v->ScaleRatioAndTapsSupport == true1 && v->SourceFormatPixelAndScanSupport == true1 && v->ViewportSizeSupport[i][j] == true1 |
| 5453 | && v->LinkCapacitySupport[i] == true1 && !P2IWith420 && !DSCOnlyIfNecessaryWithBPP |
| 5454 | && !DSC422NativeNotSupported && v->ODMCombine4To1SupportCheckOK[i] == true1 && v->NotEnoughDSCUnits[i] == false0 |
| 5455 | && v->DTBCLKRequiredMoreThanSupported[i] == false0 |
| 5456 | && v->ROBSupport[i][j] == true1 && v->DISPCLK_DPPCLK_Support[i][j] == true1 |
| 5457 | && v->TotalAvailablePipesSupport[i][j] == true1 && EnoughWritebackUnits == true1 |
| 5458 | && v->WritebackLatencySupport == true1 && v->WritebackScaleRatioAndTapsSupport == true1 |
| 5459 | && v->CursorSupport == true1 && v->PitchSupport == true1 && ViewportExceedsSurface == false0 |
| 5460 | && v->PrefetchSupported[i][j] == true1 && v->DynamicMetadataSupported[i][j] == true1 |
| 5461 | && v->TotalVerticalActiveBandwidthSupport[i][j] == true1 && v->VRatioInPrefetchSupported[i][j] == true1 |
| 5462 | && v->PTEBufferSizeNotExceeded[i][j] == true1 && v->NonsupportedDSCInputBPC == false0 |
| 5463 | && ((v->HostVMEnable == false0 |
| 5464 | && v->ImmediateFlipRequirement[0] != dm_immediate_flip_required) |
| 5465 | || v->ImmediateFlipSupportedForState[i][j] == true1) |
| 5466 | && FMTBufferExceeded == false0) { |
| 5467 | v->ModeSupport[i][j] = true1; |
| 5468 | } else { |
| 5469 | v->ModeSupport[i][j] = false0; |
| 5470 | } |
| 5471 | } |
| 5472 | } |
| 5473 | |
| 5474 | { |
| 5475 | unsigned int MaximumMPCCombine = 0; |
| 5476 | for (i = v->soc.num_states; i >= 0; i--) { |
| 5477 | if (i == v->soc.num_states || v->ModeSupport[i][0] == true1 || v->ModeSupport[i][1] == true1) { |
| 5478 | v->VoltageLevel = i; |
| 5479 | v->ModeIsSupported = v->ModeSupport[i][0] == true1 || v->ModeSupport[i][1] == true1; |
| 5480 | if (v->ModeSupport[i][0] == true1) { |
| 5481 | MaximumMPCCombine = 0; |
| 5482 | } else { |
| 5483 | MaximumMPCCombine = 1; |
| 5484 | } |
| 5485 | } |
| 5486 | } |
| 5487 | v->ImmediateFlipSupport = v->ImmediateFlipSupportedForState[v->VoltageLevel][MaximumMPCCombine]; |
| 5488 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
| 5489 | v->MPCCombineEnable[k] = v->MPCCombine[v->VoltageLevel][MaximumMPCCombine][k]; |
| 5490 | v->DPPPerPlane[k] = v->NoOfDPP[v->VoltageLevel][MaximumMPCCombine][k]; |
| 5491 | } |
| 5492 | v->DCFCLK = v->DCFCLKState[v->VoltageLevel][MaximumMPCCombine]; |
| 5493 | v->DRAMSpeed = v->DRAMSpeedPerState[v->VoltageLevel]; |
| 5494 | v->FabricClock = v->FabricClockPerState[v->VoltageLevel]; |
| 5495 | v->SOCCLK = v->SOCCLKPerState[v->VoltageLevel]; |
| 5496 | v->ReturnBW = v->ReturnBWPerState[v->VoltageLevel][MaximumMPCCombine]; |
| 5497 | v->maxMpcComb = MaximumMPCCombine; |
| 5498 | } |
| 5499 | } |
| 5500 | |
| 5501 | static void CalculateWatermarksAndDRAMSpeedChangeSupport( |
| 5502 | struct display_mode_lib *mode_lib, |
| 5503 | unsigned int PrefetchMode, |
| 5504 | double DCFCLK, |
| 5505 | double ReturnBW, |
| 5506 | double UrgentLatency, |
| 5507 | double ExtraLatency, |
| 5508 | double SOCCLK, |
| 5509 | double DCFCLKDeepSleep, |
| 5510 | unsigned int DETBufferSizeY[], |
| 5511 | unsigned int DETBufferSizeC[], |
| 5512 | unsigned int SwathHeightY[], |
| 5513 | unsigned int SwathHeightC[], |
| 5514 | double SwathWidthY[], |
| 5515 | double SwathWidthC[], |
| 5516 | unsigned int DPPPerPlane[], |
| 5517 | double BytePerPixelDETY[], |
| 5518 | double BytePerPixelDETC[], |
| 5519 | bool_Bool UnboundedRequestEnabled, |
| 5520 | int unsigned CompressedBufferSizeInkByte, |
| 5521 | enum clock_change_support *DRAMClockChangeSupport, |
| 5522 | double *StutterExitWatermark, |
| 5523 | double *StutterEnterPlusExitWatermark, |
| 5524 | double *Z8StutterExitWatermark, |
| 5525 | double *Z8StutterEnterPlusExitWatermark) |
| 5526 | { |
| 5527 | struct vba_vars_st *v = &mode_lib->vba; |
| 5528 | double EffectiveLBLatencyHidingY; |
| 5529 | double EffectiveLBLatencyHidingC; |
| 5530 | double LinesInDETY[DC__NUM_DPP__MAX8]; |
| 5531 | double LinesInDETC; |
| 5532 | unsigned int LinesInDETYRoundedDownToSwath[DC__NUM_DPP__MAX8]; |
| 5533 | unsigned int LinesInDETCRoundedDownToSwath; |
| 5534 | double FullDETBufferingTimeY; |
| 5535 | double FullDETBufferingTimeC; |
| 5536 | double ActiveDRAMClockChangeLatencyMarginY; |
| 5537 | double ActiveDRAMClockChangeLatencyMarginC; |
| 5538 | double WritebackDRAMClockChangeLatencyMargin; |
| 5539 | double PlaneWithMinActiveDRAMClockChangeMargin; |
| 5540 | double SecondMinActiveDRAMClockChangeMarginOneDisplayInVBLank; |
| 5541 | double WritebackDRAMClockChangeLatencyHiding; |
| 5542 | double TotalPixelBW = 0.0; |
| 5543 | int k, j; |
| 5544 | |
| 5545 | v->UrgentWatermark = UrgentLatency + ExtraLatency; |
| 5546 | |
| 5547 | #ifdef __DML_VBA_DEBUG__ |
| 5548 | dml_print("DML::%s: UrgentLatency = %f\n", __func__, UrgentLatency){do { } while(0); }; |
| 5549 | dml_print("DML::%s: ExtraLatency = %f\n", __func__, ExtraLatency){do { } while(0); }; |
| 5550 | dml_print("DML::%s: UrgentWatermark = %f\n", __func__, v->UrgentWatermark){do { } while(0); }; |
| 5551 | #endif |
| 5552 | |
| 5553 | v->DRAMClockChangeWatermark = v->DRAMClockChangeLatency + v->UrgentWatermark; |
| 5554 | |
| 5555 | #ifdef __DML_VBA_DEBUG__ |
| 5556 | dml_print("DML::%s: v->DRAMClockChangeLatency = %f\n", __func__, v->DRAMClockChangeLatency){do { } while(0); }; |
| 5557 | dml_print("DML::%s: DRAMClockChangeWatermark = %f\n", __func__, v->DRAMClockChangeWatermark){do { } while(0); }; |
| 5558 | #endif |
| 5559 | |
| 5560 | v->TotalActiveWriteback = 0; |
| 5561 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 5562 | if (v->WritebackEnable[k] == true1) { |
| 5563 | v->TotalActiveWriteback = v->TotalActiveWriteback + 1; |
| 5564 | } |
| 5565 | } |
| 5566 | |
| 5567 | if (v->TotalActiveWriteback <= 1) { |
| 5568 | v->WritebackUrgentWatermark = v->WritebackLatency; |
| 5569 | } else { |
| 5570 | v->WritebackUrgentWatermark = v->WritebackLatency + v->WritebackChunkSize * 1024.0 / 32.0 / SOCCLK; |
| 5571 | } |
| 5572 | |
| 5573 | if (v->TotalActiveWriteback <= 1) { |
| 5574 | v->WritebackDRAMClockChangeWatermark = v->DRAMClockChangeLatency + v->WritebackLatency; |
| 5575 | } else { |
| 5576 | v->WritebackDRAMClockChangeWatermark = v->DRAMClockChangeLatency + v->WritebackLatency + v->WritebackChunkSize * 1024.0 / 32.0 / SOCCLK; |
| 5577 | } |
| 5578 | |
| 5579 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 5580 | TotalPixelBW = TotalPixelBW |
| 5581 | + DPPPerPlane[k] * (SwathWidthY[k] * BytePerPixelDETY[k] * v->VRatio[k] + SwathWidthC[k] * BytePerPixelDETC[k] * v->VRatioChroma[k]) |
| 5582 | / (v->HTotal[k] / v->PixelClock[k]); |
| 5583 | } |
| 5584 | |
| 5585 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 5586 | double EffectiveDETBufferSizeY = DETBufferSizeY[k]; |
| 5587 | |
| 5588 | v->LBLatencyHidingSourceLinesY = dml_min( |
| 5589 | (double) v->MaxLineBufferLines, |
| 5590 | dml_floor(v->LineBufferSize / v->LBBitPerPixel[k] / (SwathWidthY[k] / dml_max(v->HRatio[k], 1.0)), 1)) - (v->vtaps[k] - 1); |
| 5591 | |
| 5592 | v->LBLatencyHidingSourceLinesC = dml_min( |
| 5593 | (double) v->MaxLineBufferLines, |
| 5594 | dml_floor(v->LineBufferSize / v->LBBitPerPixel[k] / (SwathWidthC[k] / dml_max(v->HRatioChroma[k], 1.0)), 1)) - (v->VTAPsChroma[k] - 1); |
| 5595 | |
| 5596 | EffectiveLBLatencyHidingY = v->LBLatencyHidingSourceLinesY / v->VRatio[k] * (v->HTotal[k] / v->PixelClock[k]); |
| 5597 | |
| 5598 | EffectiveLBLatencyHidingC = v->LBLatencyHidingSourceLinesC / v->VRatioChroma[k] * (v->HTotal[k] / v->PixelClock[k]); |
| 5599 | |
| 5600 | if (UnboundedRequestEnabled) { |
| 5601 | EffectiveDETBufferSizeY = EffectiveDETBufferSizeY |
| 5602 | + CompressedBufferSizeInkByte * 1024 * SwathWidthY[k] * BytePerPixelDETY[k] * v->VRatio[k] / (v->HTotal[k] / v->PixelClock[k]) / TotalPixelBW; |
| 5603 | } |
| 5604 | |
| 5605 | LinesInDETY[k] = (double) EffectiveDETBufferSizeY / BytePerPixelDETY[k] / SwathWidthY[k]; |
| 5606 | LinesInDETYRoundedDownToSwath[k] = dml_floor(LinesInDETY[k], SwathHeightY[k]); |
| 5607 | FullDETBufferingTimeY = LinesInDETYRoundedDownToSwath[k] * (v->HTotal[k] / v->PixelClock[k]) / v->VRatio[k]; |
| 5608 | if (BytePerPixelDETC[k] > 0) { |
| 5609 | LinesInDETC = v->DETBufferSizeC[k] / BytePerPixelDETC[k] / SwathWidthC[k]; |
| 5610 | LinesInDETCRoundedDownToSwath = dml_floor(LinesInDETC, SwathHeightC[k]); |
| 5611 | FullDETBufferingTimeC = LinesInDETCRoundedDownToSwath * (v->HTotal[k] / v->PixelClock[k]) / v->VRatioChroma[k]; |
| 5612 | } else { |
| 5613 | LinesInDETC = 0; |
Value stored to 'LinesInDETC' is never read | |
| 5614 | FullDETBufferingTimeC = 999999; |
| 5615 | } |
| 5616 | |
| 5617 | ActiveDRAMClockChangeLatencyMarginY = EffectiveLBLatencyHidingY + FullDETBufferingTimeY |
| 5618 | - ((double) v->DSTXAfterScaler[k] / v->HTotal[k] + v->DSTYAfterScaler[k]) * v->HTotal[k] / v->PixelClock[k] - v->UrgentWatermark - v->DRAMClockChangeWatermark; |
| 5619 | |
| 5620 | if (v->NumberOfActivePlanes > 1) { |
| 5621 | ActiveDRAMClockChangeLatencyMarginY = ActiveDRAMClockChangeLatencyMarginY |
| 5622 | - (1 - 1.0 / v->NumberOfActivePlanes) * SwathHeightY[k] * v->HTotal[k] / v->PixelClock[k] / v->VRatio[k]; |
| 5623 | } |
| 5624 | |
| 5625 | if (BytePerPixelDETC[k] > 0) { |
| 5626 | ActiveDRAMClockChangeLatencyMarginC = EffectiveLBLatencyHidingC + FullDETBufferingTimeC |
| 5627 | - ((double) v->DSTXAfterScaler[k] / v->HTotal[k] + v->DSTYAfterScaler[k]) * v->HTotal[k] / v->PixelClock[k] - v->UrgentWatermark - v->DRAMClockChangeWatermark; |
| 5628 | |
| 5629 | if (v->NumberOfActivePlanes > 1) { |
| 5630 | ActiveDRAMClockChangeLatencyMarginC = ActiveDRAMClockChangeLatencyMarginC |
| 5631 | - (1 - 1.0 / v->NumberOfActivePlanes) * SwathHeightC[k] * v->HTotal[k] / v->PixelClock[k] / v->VRatioChroma[k]; |
| 5632 | } |
| 5633 | v->ActiveDRAMClockChangeLatencyMargin[k] = dml_min(ActiveDRAMClockChangeLatencyMarginY, ActiveDRAMClockChangeLatencyMarginC); |
| 5634 | } else { |
| 5635 | v->ActiveDRAMClockChangeLatencyMargin[k] = ActiveDRAMClockChangeLatencyMarginY; |
| 5636 | } |
| 5637 | |
| 5638 | if (v->WritebackEnable[k] == true1) { |
| 5639 | WritebackDRAMClockChangeLatencyHiding = v->WritebackInterfaceBufferSize * 1024 |
| 5640 | / (v->WritebackDestinationWidth[k] * v->WritebackDestinationHeight[k] / (v->WritebackSourceHeight[k] * v->HTotal[k] / v->PixelClock[k]) * 4); |
| 5641 | if (v->WritebackPixelFormat[k] == dm_444_64) { |
| 5642 | WritebackDRAMClockChangeLatencyHiding = WritebackDRAMClockChangeLatencyHiding / 2; |
| 5643 | } |
| 5644 | WritebackDRAMClockChangeLatencyMargin = WritebackDRAMClockChangeLatencyHiding - v->WritebackDRAMClockChangeWatermark; |
| 5645 | v->ActiveDRAMClockChangeLatencyMargin[k] = dml_min(v->ActiveDRAMClockChangeLatencyMargin[k], WritebackDRAMClockChangeLatencyMargin); |
| 5646 | } |
| 5647 | } |
| 5648 | |
| 5649 | v->MinActiveDRAMClockChangeMargin = 999999; |
| 5650 | PlaneWithMinActiveDRAMClockChangeMargin = 0; |
| 5651 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 5652 | if (v->ActiveDRAMClockChangeLatencyMargin[k] < v->MinActiveDRAMClockChangeMargin) { |
| 5653 | v->MinActiveDRAMClockChangeMargin = v->ActiveDRAMClockChangeLatencyMargin[k]; |
| 5654 | if (v->BlendingAndTiming[k] == k) { |
| 5655 | PlaneWithMinActiveDRAMClockChangeMargin = k; |
| 5656 | } else { |
| 5657 | for (j = 0; j < v->NumberOfActivePlanes; ++j) { |
| 5658 | if (v->BlendingAndTiming[k] == j) { |
| 5659 | PlaneWithMinActiveDRAMClockChangeMargin = j; |
| 5660 | } |
| 5661 | } |
| 5662 | } |
| 5663 | } |
| 5664 | } |
| 5665 | |
| 5666 | v->MinActiveDRAMClockChangeLatencySupported = v->MinActiveDRAMClockChangeMargin + v->DRAMClockChangeLatency ; |
| 5667 | |
| 5668 | SecondMinActiveDRAMClockChangeMarginOneDisplayInVBLank = 999999; |
| 5669 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 5670 | if (!((k == PlaneWithMinActiveDRAMClockChangeMargin) && (v->BlendingAndTiming[k] == k)) && !(v->BlendingAndTiming[k] == PlaneWithMinActiveDRAMClockChangeMargin) |
| 5671 | && v->ActiveDRAMClockChangeLatencyMargin[k] < SecondMinActiveDRAMClockChangeMarginOneDisplayInVBLank) { |
| 5672 | SecondMinActiveDRAMClockChangeMarginOneDisplayInVBLank = v->ActiveDRAMClockChangeLatencyMargin[k]; |
| 5673 | } |
| 5674 | } |
| 5675 | |
| 5676 | v->TotalNumberOfActiveOTG = 0; |
| 5677 | |
| 5678 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 5679 | if (v->BlendingAndTiming[k] == k) { |
| 5680 | v->TotalNumberOfActiveOTG = v->TotalNumberOfActiveOTG + 1; |
| 5681 | } |
| 5682 | } |
| 5683 | |
| 5684 | if (v->MinActiveDRAMClockChangeMargin > 0 && PrefetchMode == 0) { |
| 5685 | *DRAMClockChangeSupport = dm_dram_clock_change_vactive; |
| 5686 | } else if ((v->SynchronizedVBlank == true1 || v->TotalNumberOfActiveOTG == 1 |
| 5687 | || SecondMinActiveDRAMClockChangeMarginOneDisplayInVBLank > 0) && PrefetchMode == 0) { |
| 5688 | *DRAMClockChangeSupport = dm_dram_clock_change_vblank; |
| 5689 | } else { |
| 5690 | *DRAMClockChangeSupport = dm_dram_clock_change_unsupported; |
| 5691 | } |
| 5692 | |
| 5693 | *StutterExitWatermark = v->SRExitTime + ExtraLatency + 10 / DCFCLKDeepSleep; |
| 5694 | *StutterEnterPlusExitWatermark = (v->SREnterPlusExitTime + ExtraLatency + 10 / DCFCLKDeepSleep); |
| 5695 | *Z8StutterExitWatermark = v->SRExitZ8Time + ExtraLatency + 10 / DCFCLKDeepSleep; |
| 5696 | *Z8StutterEnterPlusExitWatermark = v->SREnterPlusExitZ8Time + ExtraLatency + 10 / DCFCLKDeepSleep; |
| 5697 | |
| 5698 | #ifdef __DML_VBA_DEBUG__ |
| 5699 | dml_print("DML::%s: StutterExitWatermark = %f\n", __func__, *StutterExitWatermark){do { } while(0); }; |
| 5700 | dml_print("DML::%s: StutterEnterPlusExitWatermark = %f\n", __func__, *StutterEnterPlusExitWatermark){do { } while(0); }; |
| 5701 | dml_print("DML::%s: Z8StutterExitWatermark = %f\n", __func__, *Z8StutterExitWatermark){do { } while(0); }; |
| 5702 | dml_print("DML::%s: Z8StutterEnterPlusExitWatermark = %f\n", __func__, *Z8StutterEnterPlusExitWatermark){do { } while(0); }; |
| 5703 | #endif |
| 5704 | } |
| 5705 | |
| 5706 | static void CalculateDCFCLKDeepSleep( |
| 5707 | struct display_mode_lib *mode_lib, |
| 5708 | unsigned int NumberOfActivePlanes, |
| 5709 | int BytePerPixelY[], |
| 5710 | int BytePerPixelC[], |
| 5711 | double VRatio[], |
| 5712 | double VRatioChroma[], |
| 5713 | double SwathWidthY[], |
| 5714 | double SwathWidthC[], |
| 5715 | unsigned int DPPPerPlane[], |
| 5716 | double HRatio[], |
| 5717 | double HRatioChroma[], |
| 5718 | double PixelClock[], |
| 5719 | double PSCL_THROUGHPUT[], |
| 5720 | double PSCL_THROUGHPUT_CHROMA[], |
| 5721 | double DPPCLK[], |
| 5722 | double ReadBandwidthLuma[], |
| 5723 | double ReadBandwidthChroma[], |
| 5724 | int ReturnBusWidth, |
| 5725 | double *DCFCLKDeepSleep) |
| 5726 | { |
| 5727 | struct vba_vars_st *v = &mode_lib->vba; |
| 5728 | double DisplayPipeLineDeliveryTimeLuma; |
| 5729 | double DisplayPipeLineDeliveryTimeChroma; |
| 5730 | double ReadBandwidth = 0.0; |
| 5731 | int k; |
| 5732 | |
| 5733 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 5734 | |
| 5735 | if (VRatio[k] <= 1) { |
| 5736 | DisplayPipeLineDeliveryTimeLuma = SwathWidthY[k] * DPPPerPlane[k] / HRatio[k] / PixelClock[k]; |
| 5737 | } else { |
| 5738 | DisplayPipeLineDeliveryTimeLuma = SwathWidthY[k] / PSCL_THROUGHPUT[k] / DPPCLK[k]; |
| 5739 | } |
| 5740 | if (BytePerPixelC[k] == 0) { |
| 5741 | DisplayPipeLineDeliveryTimeChroma = 0; |
| 5742 | } else { |
| 5743 | if (VRatioChroma[k] <= 1) { |
| 5744 | DisplayPipeLineDeliveryTimeChroma = SwathWidthC[k] * DPPPerPlane[k] / HRatioChroma[k] / PixelClock[k]; |
| 5745 | } else { |
| 5746 | DisplayPipeLineDeliveryTimeChroma = SwathWidthC[k] / PSCL_THROUGHPUT_CHROMA[k] / DPPCLK[k]; |
| 5747 | } |
| 5748 | } |
| 5749 | |
| 5750 | if (BytePerPixelC[k] > 0) { |
| 5751 | v->DCFCLKDeepSleepPerPlane[k] = dml_max(__DML_MIN_DCFCLK_FACTOR__1.15 * SwathWidthY[k] * BytePerPixelY[k] / 32.0 / DisplayPipeLineDeliveryTimeLuma, |
| 5752 | __DML_MIN_DCFCLK_FACTOR__1.15 * SwathWidthC[k] * BytePerPixelC[k] / 32.0 / DisplayPipeLineDeliveryTimeChroma); |
| 5753 | } else { |
| 5754 | v->DCFCLKDeepSleepPerPlane[k] = __DML_MIN_DCFCLK_FACTOR__1.15 * SwathWidthY[k] * BytePerPixelY[k] / 64.0 / DisplayPipeLineDeliveryTimeLuma; |
| 5755 | } |
| 5756 | v->DCFCLKDeepSleepPerPlane[k] = dml_max(v->DCFCLKDeepSleepPerPlane[k], PixelClock[k] / 16); |
| 5757 | |
| 5758 | } |
| 5759 | |
| 5760 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 5761 | ReadBandwidth = ReadBandwidth + ReadBandwidthLuma[k] + ReadBandwidthChroma[k]; |
| 5762 | } |
| 5763 | |
| 5764 | *DCFCLKDeepSleep = dml_max(8.0, __DML_MIN_DCFCLK_FACTOR__1.15 * ReadBandwidth / ReturnBusWidth); |
| 5765 | |
| 5766 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 5767 | *DCFCLKDeepSleep = dml_max(*DCFCLKDeepSleep, v->DCFCLKDeepSleepPerPlane[k]); |
| 5768 | } |
| 5769 | } |
| 5770 | |
| 5771 | static void CalculateUrgentBurstFactor( |
| 5772 | int swath_width_luma_ub, |
| 5773 | int swath_width_chroma_ub, |
| 5774 | unsigned int SwathHeightY, |
| 5775 | unsigned int SwathHeightC, |
| 5776 | double LineTime, |
| 5777 | double UrgentLatency, |
| 5778 | double CursorBufferSize, |
| 5779 | unsigned int CursorWidth, |
| 5780 | unsigned int CursorBPP, |
| 5781 | double VRatio, |
| 5782 | double VRatioC, |
| 5783 | double BytePerPixelInDETY, |
| 5784 | double BytePerPixelInDETC, |
| 5785 | double DETBufferSizeY, |
| 5786 | double DETBufferSizeC, |
| 5787 | double *UrgentBurstFactorCursor, |
| 5788 | double *UrgentBurstFactorLuma, |
| 5789 | double *UrgentBurstFactorChroma, |
| 5790 | bool_Bool *NotEnoughUrgentLatencyHiding) |
| 5791 | { |
| 5792 | double LinesInDETLuma; |
| 5793 | double LinesInDETChroma; |
| 5794 | unsigned int LinesInCursorBuffer; |
| 5795 | double CursorBufferSizeInTime; |
| 5796 | double DETBufferSizeInTimeLuma; |
| 5797 | double DETBufferSizeInTimeChroma; |
| 5798 | |
| 5799 | *NotEnoughUrgentLatencyHiding = 0; |
| 5800 | |
| 5801 | if (CursorWidth > 0) { |
| 5802 | LinesInCursorBuffer = 1 << (unsigned int) dml_floor(dml_log2(CursorBufferSize * 1024.0 / (CursorWidth * CursorBPP / 8.0)), 1.0); |
| 5803 | if (VRatio > 0) { |
| 5804 | CursorBufferSizeInTime = LinesInCursorBuffer * LineTime / VRatio; |
| 5805 | if (CursorBufferSizeInTime - UrgentLatency <= 0) { |
| 5806 | *NotEnoughUrgentLatencyHiding = 1; |
| 5807 | *UrgentBurstFactorCursor = 0; |
| 5808 | } else { |
| 5809 | *UrgentBurstFactorCursor = CursorBufferSizeInTime / (CursorBufferSizeInTime - UrgentLatency); |
| 5810 | } |
| 5811 | } else { |
| 5812 | *UrgentBurstFactorCursor = 1; |
| 5813 | } |
| 5814 | } |
| 5815 | |
| 5816 | LinesInDETLuma = DETBufferSizeY / BytePerPixelInDETY / swath_width_luma_ub; |
| 5817 | if (VRatio > 0) { |
| 5818 | DETBufferSizeInTimeLuma = dml_floor(LinesInDETLuma, SwathHeightY) * LineTime / VRatio; |
| 5819 | if (DETBufferSizeInTimeLuma - UrgentLatency <= 0) { |
| 5820 | *NotEnoughUrgentLatencyHiding = 1; |
| 5821 | *UrgentBurstFactorLuma = 0; |
| 5822 | } else { |
| 5823 | *UrgentBurstFactorLuma = DETBufferSizeInTimeLuma / (DETBufferSizeInTimeLuma - UrgentLatency); |
| 5824 | } |
| 5825 | } else { |
| 5826 | *UrgentBurstFactorLuma = 1; |
| 5827 | } |
| 5828 | |
| 5829 | if (BytePerPixelInDETC > 0) { |
| 5830 | LinesInDETChroma = DETBufferSizeC / BytePerPixelInDETC / swath_width_chroma_ub; |
| 5831 | if (VRatio > 0) { |
| 5832 | DETBufferSizeInTimeChroma = dml_floor(LinesInDETChroma, SwathHeightC) * LineTime / VRatio; |
| 5833 | if (DETBufferSizeInTimeChroma - UrgentLatency <= 0) { |
| 5834 | *NotEnoughUrgentLatencyHiding = 1; |
| 5835 | *UrgentBurstFactorChroma = 0; |
| 5836 | } else { |
| 5837 | *UrgentBurstFactorChroma = DETBufferSizeInTimeChroma / (DETBufferSizeInTimeChroma - UrgentLatency); |
| 5838 | } |
| 5839 | } else { |
| 5840 | *UrgentBurstFactorChroma = 1; |
| 5841 | } |
| 5842 | } |
| 5843 | } |
| 5844 | |
| 5845 | static void CalculatePixelDeliveryTimes( |
| 5846 | unsigned int NumberOfActivePlanes, |
| 5847 | double VRatio[], |
| 5848 | double VRatioChroma[], |
| 5849 | double VRatioPrefetchY[], |
| 5850 | double VRatioPrefetchC[], |
| 5851 | unsigned int swath_width_luma_ub[], |
| 5852 | unsigned int swath_width_chroma_ub[], |
| 5853 | unsigned int DPPPerPlane[], |
| 5854 | double HRatio[], |
| 5855 | double HRatioChroma[], |
| 5856 | double PixelClock[], |
| 5857 | double PSCL_THROUGHPUT[], |
| 5858 | double PSCL_THROUGHPUT_CHROMA[], |
| 5859 | double DPPCLK[], |
| 5860 | int BytePerPixelC[], |
| 5861 | enum scan_direction_class SourceScan[], |
| 5862 | unsigned int NumberOfCursors[], |
| 5863 | unsigned int CursorWidth[][DC__NUM_CURSOR__MAX2], |
| 5864 | unsigned int CursorBPP[][DC__NUM_CURSOR__MAX2], |
| 5865 | unsigned int BlockWidth256BytesY[], |
| 5866 | unsigned int BlockHeight256BytesY[], |
| 5867 | unsigned int BlockWidth256BytesC[], |
| 5868 | unsigned int BlockHeight256BytesC[], |
| 5869 | double DisplayPipeLineDeliveryTimeLuma[], |
| 5870 | double DisplayPipeLineDeliveryTimeChroma[], |
| 5871 | double DisplayPipeLineDeliveryTimeLumaPrefetch[], |
| 5872 | double DisplayPipeLineDeliveryTimeChromaPrefetch[], |
| 5873 | double DisplayPipeRequestDeliveryTimeLuma[], |
| 5874 | double DisplayPipeRequestDeliveryTimeChroma[], |
| 5875 | double DisplayPipeRequestDeliveryTimeLumaPrefetch[], |
| 5876 | double DisplayPipeRequestDeliveryTimeChromaPrefetch[], |
| 5877 | double CursorRequestDeliveryTime[], |
| 5878 | double CursorRequestDeliveryTimePrefetch[]) |
| 5879 | { |
| 5880 | double req_per_swath_ub; |
| 5881 | int k; |
| 5882 | |
| 5883 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 5884 | if (VRatio[k] <= 1) { |
| 5885 | DisplayPipeLineDeliveryTimeLuma[k] = swath_width_luma_ub[k] * DPPPerPlane[k] / HRatio[k] / PixelClock[k]; |
| 5886 | } else { |
| 5887 | DisplayPipeLineDeliveryTimeLuma[k] = swath_width_luma_ub[k] / PSCL_THROUGHPUT[k] / DPPCLK[k]; |
| 5888 | } |
| 5889 | |
| 5890 | if (BytePerPixelC[k] == 0) { |
| 5891 | DisplayPipeLineDeliveryTimeChroma[k] = 0; |
| 5892 | } else { |
| 5893 | if (VRatioChroma[k] <= 1) { |
| 5894 | DisplayPipeLineDeliveryTimeChroma[k] = swath_width_chroma_ub[k] * DPPPerPlane[k] / HRatioChroma[k] / PixelClock[k]; |
| 5895 | } else { |
| 5896 | DisplayPipeLineDeliveryTimeChroma[k] = swath_width_chroma_ub[k] / PSCL_THROUGHPUT_CHROMA[k] / DPPCLK[k]; |
| 5897 | } |
| 5898 | } |
| 5899 | |
| 5900 | if (VRatioPrefetchY[k] <= 1) { |
| 5901 | DisplayPipeLineDeliveryTimeLumaPrefetch[k] = swath_width_luma_ub[k] * DPPPerPlane[k] / HRatio[k] / PixelClock[k]; |
| 5902 | } else { |
| 5903 | DisplayPipeLineDeliveryTimeLumaPrefetch[k] = swath_width_luma_ub[k] / PSCL_THROUGHPUT[k] / DPPCLK[k]; |
| 5904 | } |
| 5905 | |
| 5906 | if (BytePerPixelC[k] == 0) { |
| 5907 | DisplayPipeLineDeliveryTimeChromaPrefetch[k] = 0; |
| 5908 | } else { |
| 5909 | if (VRatioPrefetchC[k] <= 1) { |
| 5910 | DisplayPipeLineDeliveryTimeChromaPrefetch[k] = swath_width_chroma_ub[k] * DPPPerPlane[k] / HRatioChroma[k] / PixelClock[k]; |
| 5911 | } else { |
| 5912 | DisplayPipeLineDeliveryTimeChromaPrefetch[k] = swath_width_chroma_ub[k] / PSCL_THROUGHPUT_CHROMA[k] / DPPCLK[k]; |
| 5913 | } |
| 5914 | } |
| 5915 | } |
| 5916 | |
| 5917 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 5918 | if (SourceScan[k] != dm_vert) { |
| 5919 | req_per_swath_ub = swath_width_luma_ub[k] / BlockWidth256BytesY[k]; |
| 5920 | } else { |
| 5921 | req_per_swath_ub = swath_width_luma_ub[k] / BlockHeight256BytesY[k]; |
| 5922 | } |
| 5923 | DisplayPipeRequestDeliveryTimeLuma[k] = DisplayPipeLineDeliveryTimeLuma[k] / req_per_swath_ub; |
| 5924 | DisplayPipeRequestDeliveryTimeLumaPrefetch[k] = DisplayPipeLineDeliveryTimeLumaPrefetch[k] / req_per_swath_ub; |
| 5925 | if (BytePerPixelC[k] == 0) { |
| 5926 | DisplayPipeRequestDeliveryTimeChroma[k] = 0; |
| 5927 | DisplayPipeRequestDeliveryTimeChromaPrefetch[k] = 0; |
| 5928 | } else { |
| 5929 | if (SourceScan[k] != dm_vert) { |
| 5930 | req_per_swath_ub = swath_width_chroma_ub[k] / BlockWidth256BytesC[k]; |
| 5931 | } else { |
| 5932 | req_per_swath_ub = swath_width_chroma_ub[k] / BlockHeight256BytesC[k]; |
| 5933 | } |
| 5934 | DisplayPipeRequestDeliveryTimeChroma[k] = DisplayPipeLineDeliveryTimeChroma[k] / req_per_swath_ub; |
| 5935 | DisplayPipeRequestDeliveryTimeChromaPrefetch[k] = DisplayPipeLineDeliveryTimeChromaPrefetch[k] / req_per_swath_ub; |
| 5936 | } |
| 5937 | #ifdef __DML_VBA_DEBUG__ |
| 5938 | dml_print("DML::%s: k=%d : HRatio = %f\n", __func__, k, HRatio[k]){do { } while(0); }; |
| 5939 | dml_print("DML::%s: k=%d : VRatio = %f\n", __func__, k, VRatio[k]){do { } while(0); }; |
| 5940 | dml_print("DML::%s: k=%d : HRatioChroma = %f\n", __func__, k, HRatioChroma[k]){do { } while(0); }; |
| 5941 | dml_print("DML::%s: k=%d : VRatioChroma = %f\n", __func__, k, VRatioChroma[k]){do { } while(0); }; |
| 5942 | dml_print("DML::%s: k=%d : DisplayPipeLineDeliveryTimeLuma = %f\n", __func__, k, DisplayPipeLineDeliveryTimeLuma[k]){do { } while(0); }; |
| 5943 | dml_print("DML::%s: k=%d : DisplayPipeLineDeliveryTimeLumaPrefetch = %f\n", __func__, k, DisplayPipeLineDeliveryTimeLumaPrefetch[k]){do { } while(0); }; |
| 5944 | dml_print("DML::%s: k=%d : DisplayPipeLineDeliveryTimeChroma = %f\n", __func__, k, DisplayPipeLineDeliveryTimeChroma[k]){do { } while(0); }; |
| 5945 | dml_print("DML::%s: k=%d : DisplayPipeLineDeliveryTimeChromaPrefetch = %f\n", __func__, k, DisplayPipeLineDeliveryTimeChromaPrefetch[k]){do { } while(0); }; |
| 5946 | dml_print("DML::%s: k=%d : DisplayPipeRequestDeliveryTimeLuma = %f\n", __func__, k, DisplayPipeRequestDeliveryTimeLuma[k]){do { } while(0); }; |
| 5947 | dml_print("DML::%s: k=%d : DisplayPipeRequestDeliveryTimeLumaPrefetch = %f\n", __func__, k, DisplayPipeRequestDeliveryTimeLumaPrefetch[k]){do { } while(0); }; |
| 5948 | dml_print("DML::%s: k=%d : DisplayPipeRequestDeliveryTimeChroma = %f\n", __func__, k, DisplayPipeRequestDeliveryTimeChroma[k]){do { } while(0); }; |
| 5949 | dml_print("DML::%s: k=%d : DisplayPipeRequestDeliveryTimeChromaPrefetch = %f\n", __func__, k, DisplayPipeRequestDeliveryTimeChromaPrefetch[k]){do { } while(0); }; |
| 5950 | #endif |
| 5951 | } |
| 5952 | |
| 5953 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 5954 | int cursor_req_per_width; |
| 5955 | cursor_req_per_width = dml_ceil(CursorWidth[k][0] * CursorBPP[k][0] / 256 / 8, 1); |
| 5956 | if (NumberOfCursors[k] > 0) { |
| 5957 | if (VRatio[k] <= 1) { |
| 5958 | CursorRequestDeliveryTime[k] = CursorWidth[k][0] / HRatio[k] / PixelClock[k] / cursor_req_per_width; |
| 5959 | } else { |
| 5960 | CursorRequestDeliveryTime[k] = CursorWidth[k][0] / PSCL_THROUGHPUT[k] / DPPCLK[k] / cursor_req_per_width; |
| 5961 | } |
| 5962 | if (VRatioPrefetchY[k] <= 1) { |
| 5963 | CursorRequestDeliveryTimePrefetch[k] = CursorWidth[k][0] / HRatio[k] / PixelClock[k] / cursor_req_per_width; |
| 5964 | } else { |
| 5965 | CursorRequestDeliveryTimePrefetch[k] = CursorWidth[k][0] / PSCL_THROUGHPUT[k] / DPPCLK[k] / cursor_req_per_width; |
| 5966 | } |
| 5967 | } else { |
| 5968 | CursorRequestDeliveryTime[k] = 0; |
| 5969 | CursorRequestDeliveryTimePrefetch[k] = 0; |
| 5970 | } |
| 5971 | #ifdef __DML_VBA_DEBUG__ |
| 5972 | dml_print("DML::%s: k=%d : NumberOfCursors = %d\n", __func__, k, NumberOfCursors[k]){do { } while(0); }; |
| 5973 | dml_print("DML::%s: k=%d : CursorRequestDeliveryTime = %f\n", __func__, k, CursorRequestDeliveryTime[k]){do { } while(0); }; |
| 5974 | dml_print("DML::%s: k=%d : CursorRequestDeliveryTimePrefetch = %f\n", __func__, k, CursorRequestDeliveryTimePrefetch[k]){do { } while(0); }; |
| 5975 | #endif |
| 5976 | } |
| 5977 | } |
| 5978 | |
| 5979 | static void CalculateMetaAndPTETimes( |
| 5980 | int NumberOfActivePlanes, |
| 5981 | bool_Bool GPUVMEnable, |
| 5982 | int MetaChunkSize, |
| 5983 | int MinMetaChunkSizeBytes, |
| 5984 | int HTotal[], |
| 5985 | double VRatio[], |
| 5986 | double VRatioChroma[], |
| 5987 | double DestinationLinesToRequestRowInVBlank[], |
| 5988 | double DestinationLinesToRequestRowInImmediateFlip[], |
| 5989 | bool_Bool DCCEnable[], |
| 5990 | double PixelClock[], |
| 5991 | int BytePerPixelY[], |
| 5992 | int BytePerPixelC[], |
| 5993 | enum scan_direction_class SourceScan[], |
| 5994 | int dpte_row_height[], |
| 5995 | int dpte_row_height_chroma[], |
| 5996 | int meta_row_width[], |
| 5997 | int meta_row_width_chroma[], |
| 5998 | int meta_row_height[], |
| 5999 | int meta_row_height_chroma[], |
| 6000 | int meta_req_width[], |
| 6001 | int meta_req_width_chroma[], |
| 6002 | int meta_req_height[], |
| 6003 | int meta_req_height_chroma[], |
| 6004 | int dpte_group_bytes[], |
| 6005 | int PTERequestSizeY[], |
| 6006 | int PTERequestSizeC[], |
| 6007 | int PixelPTEReqWidthY[], |
| 6008 | int PixelPTEReqHeightY[], |
| 6009 | int PixelPTEReqWidthC[], |
| 6010 | int PixelPTEReqHeightC[], |
| 6011 | int dpte_row_width_luma_ub[], |
| 6012 | int dpte_row_width_chroma_ub[], |
| 6013 | double DST_Y_PER_PTE_ROW_NOM_L[], |
| 6014 | double DST_Y_PER_PTE_ROW_NOM_C[], |
| 6015 | double DST_Y_PER_META_ROW_NOM_L[], |
| 6016 | double DST_Y_PER_META_ROW_NOM_C[], |
| 6017 | double TimePerMetaChunkNominal[], |
| 6018 | double TimePerChromaMetaChunkNominal[], |
| 6019 | double TimePerMetaChunkVBlank[], |
| 6020 | double TimePerChromaMetaChunkVBlank[], |
| 6021 | double TimePerMetaChunkFlip[], |
| 6022 | double TimePerChromaMetaChunkFlip[], |
| 6023 | double time_per_pte_group_nom_luma[], |
| 6024 | double time_per_pte_group_vblank_luma[], |
| 6025 | double time_per_pte_group_flip_luma[], |
| 6026 | double time_per_pte_group_nom_chroma[], |
| 6027 | double time_per_pte_group_vblank_chroma[], |
| 6028 | double time_per_pte_group_flip_chroma[]) |
| 6029 | { |
| 6030 | unsigned int meta_chunk_width; |
| 6031 | unsigned int min_meta_chunk_width; |
| 6032 | unsigned int meta_chunk_per_row_int; |
| 6033 | unsigned int meta_row_remainder; |
| 6034 | unsigned int meta_chunk_threshold; |
| 6035 | unsigned int meta_chunks_per_row_ub; |
| 6036 | unsigned int meta_chunk_width_chroma; |
| 6037 | unsigned int min_meta_chunk_width_chroma; |
| 6038 | unsigned int meta_chunk_per_row_int_chroma; |
| 6039 | unsigned int meta_row_remainder_chroma; |
| 6040 | unsigned int meta_chunk_threshold_chroma; |
| 6041 | unsigned int meta_chunks_per_row_ub_chroma; |
| 6042 | unsigned int dpte_group_width_luma; |
| 6043 | unsigned int dpte_groups_per_row_luma_ub; |
| 6044 | unsigned int dpte_group_width_chroma; |
| 6045 | unsigned int dpte_groups_per_row_chroma_ub; |
| 6046 | int k; |
| 6047 | |
| 6048 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 6049 | DST_Y_PER_PTE_ROW_NOM_L[k] = dpte_row_height[k] / VRatio[k]; |
| 6050 | if (BytePerPixelC[k] == 0) { |
| 6051 | DST_Y_PER_PTE_ROW_NOM_C[k] = 0; |
| 6052 | } else { |
| 6053 | DST_Y_PER_PTE_ROW_NOM_C[k] = dpte_row_height_chroma[k] / VRatioChroma[k]; |
| 6054 | } |
| 6055 | DST_Y_PER_META_ROW_NOM_L[k] = meta_row_height[k] / VRatio[k]; |
| 6056 | if (BytePerPixelC[k] == 0) { |
| 6057 | DST_Y_PER_META_ROW_NOM_C[k] = 0; |
| 6058 | } else { |
| 6059 | DST_Y_PER_META_ROW_NOM_C[k] = meta_row_height_chroma[k] / VRatioChroma[k]; |
| 6060 | } |
| 6061 | } |
| 6062 | |
| 6063 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 6064 | if (DCCEnable[k] == true1) { |
| 6065 | meta_chunk_width = MetaChunkSize * 1024 * 256 / BytePerPixelY[k] / meta_row_height[k]; |
| 6066 | min_meta_chunk_width = MinMetaChunkSizeBytes * 256 / BytePerPixelY[k] / meta_row_height[k]; |
| 6067 | meta_chunk_per_row_int = meta_row_width[k] / meta_chunk_width; |
| 6068 | meta_row_remainder = meta_row_width[k] % meta_chunk_width; |
| 6069 | if (SourceScan[k] != dm_vert) { |
| 6070 | meta_chunk_threshold = 2 * min_meta_chunk_width - meta_req_width[k]; |
| 6071 | } else { |
| 6072 | meta_chunk_threshold = 2 * min_meta_chunk_width - meta_req_height[k]; |
| 6073 | } |
| 6074 | if (meta_row_remainder <= meta_chunk_threshold) { |
| 6075 | meta_chunks_per_row_ub = meta_chunk_per_row_int + 1; |
| 6076 | } else { |
| 6077 | meta_chunks_per_row_ub = meta_chunk_per_row_int + 2; |
| 6078 | } |
| 6079 | TimePerMetaChunkNominal[k] = meta_row_height[k] / VRatio[k] * HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub; |
| 6080 | TimePerMetaChunkVBlank[k] = DestinationLinesToRequestRowInVBlank[k] * HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub; |
| 6081 | TimePerMetaChunkFlip[k] = DestinationLinesToRequestRowInImmediateFlip[k] * HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub; |
| 6082 | if (BytePerPixelC[k] == 0) { |
| 6083 | TimePerChromaMetaChunkNominal[k] = 0; |
| 6084 | TimePerChromaMetaChunkVBlank[k] = 0; |
| 6085 | TimePerChromaMetaChunkFlip[k] = 0; |
| 6086 | } else { |
| 6087 | meta_chunk_width_chroma = MetaChunkSize * 1024 * 256 / BytePerPixelC[k] / meta_row_height_chroma[k]; |
| 6088 | min_meta_chunk_width_chroma = MinMetaChunkSizeBytes * 256 / BytePerPixelC[k] / meta_row_height_chroma[k]; |
| 6089 | meta_chunk_per_row_int_chroma = (double) meta_row_width_chroma[k] / meta_chunk_width_chroma; |
| 6090 | meta_row_remainder_chroma = meta_row_width_chroma[k] % meta_chunk_width_chroma; |
| 6091 | if (SourceScan[k] != dm_vert) { |
| 6092 | meta_chunk_threshold_chroma = 2 * min_meta_chunk_width_chroma - meta_req_width_chroma[k]; |
| 6093 | } else { |
| 6094 | meta_chunk_threshold_chroma = 2 * min_meta_chunk_width_chroma - meta_req_height_chroma[k]; |
| 6095 | } |
| 6096 | if (meta_row_remainder_chroma <= meta_chunk_threshold_chroma) { |
| 6097 | meta_chunks_per_row_ub_chroma = meta_chunk_per_row_int_chroma + 1; |
| 6098 | } else { |
| 6099 | meta_chunks_per_row_ub_chroma = meta_chunk_per_row_int_chroma + 2; |
| 6100 | } |
| 6101 | TimePerChromaMetaChunkNominal[k] = meta_row_height_chroma[k] / VRatioChroma[k] * HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub_chroma; |
| 6102 | TimePerChromaMetaChunkVBlank[k] = DestinationLinesToRequestRowInVBlank[k] * HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub_chroma; |
| 6103 | TimePerChromaMetaChunkFlip[k] = DestinationLinesToRequestRowInImmediateFlip[k] * HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub_chroma; |
| 6104 | } |
| 6105 | } else { |
| 6106 | TimePerMetaChunkNominal[k] = 0; |
| 6107 | TimePerMetaChunkVBlank[k] = 0; |
| 6108 | TimePerMetaChunkFlip[k] = 0; |
| 6109 | TimePerChromaMetaChunkNominal[k] = 0; |
| 6110 | TimePerChromaMetaChunkVBlank[k] = 0; |
| 6111 | TimePerChromaMetaChunkFlip[k] = 0; |
| 6112 | } |
| 6113 | } |
| 6114 | |
| 6115 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 6116 | if (GPUVMEnable == true1) { |
| 6117 | if (SourceScan[k] != dm_vert) { |
| 6118 | dpte_group_width_luma = dpte_group_bytes[k] / PTERequestSizeY[k] * PixelPTEReqWidthY[k]; |
| 6119 | } else { |
| 6120 | dpte_group_width_luma = dpte_group_bytes[k] / PTERequestSizeY[k] * PixelPTEReqHeightY[k]; |
| 6121 | } |
| 6122 | dpte_groups_per_row_luma_ub = dml_ceil(1.0 * dpte_row_width_luma_ub[k] / dpte_group_width_luma, 1); |
| 6123 | time_per_pte_group_nom_luma[k] = DST_Y_PER_PTE_ROW_NOM_L[k] * HTotal[k] / PixelClock[k] / dpte_groups_per_row_luma_ub; |
| 6124 | time_per_pte_group_vblank_luma[k] = DestinationLinesToRequestRowInVBlank[k] * HTotal[k] / PixelClock[k] / dpte_groups_per_row_luma_ub; |
| 6125 | time_per_pte_group_flip_luma[k] = DestinationLinesToRequestRowInImmediateFlip[k] * HTotal[k] / PixelClock[k] / dpte_groups_per_row_luma_ub; |
| 6126 | if (BytePerPixelC[k] == 0) { |
| 6127 | time_per_pte_group_nom_chroma[k] = 0; |
| 6128 | time_per_pte_group_vblank_chroma[k] = 0; |
| 6129 | time_per_pte_group_flip_chroma[k] = 0; |
| 6130 | } else { |
| 6131 | if (SourceScan[k] != dm_vert) { |
| 6132 | dpte_group_width_chroma = dpte_group_bytes[k] / PTERequestSizeC[k] * PixelPTEReqWidthC[k]; |
| 6133 | } else { |
| 6134 | dpte_group_width_chroma = dpte_group_bytes[k] / PTERequestSizeC[k] * PixelPTEReqHeightC[k]; |
| 6135 | } |
| 6136 | dpte_groups_per_row_chroma_ub = dml_ceil(1.0 * dpte_row_width_chroma_ub[k] / dpte_group_width_chroma, 1); |
| 6137 | time_per_pte_group_nom_chroma[k] = DST_Y_PER_PTE_ROW_NOM_C[k] * HTotal[k] / PixelClock[k] / dpte_groups_per_row_chroma_ub; |
| 6138 | time_per_pte_group_vblank_chroma[k] = DestinationLinesToRequestRowInVBlank[k] * HTotal[k] / PixelClock[k] / dpte_groups_per_row_chroma_ub; |
| 6139 | time_per_pte_group_flip_chroma[k] = DestinationLinesToRequestRowInImmediateFlip[k] * HTotal[k] / PixelClock[k] / dpte_groups_per_row_chroma_ub; |
| 6140 | } |
| 6141 | } else { |
| 6142 | time_per_pte_group_nom_luma[k] = 0; |
| 6143 | time_per_pte_group_vblank_luma[k] = 0; |
| 6144 | time_per_pte_group_flip_luma[k] = 0; |
| 6145 | time_per_pte_group_nom_chroma[k] = 0; |
| 6146 | time_per_pte_group_vblank_chroma[k] = 0; |
| 6147 | time_per_pte_group_flip_chroma[k] = 0; |
| 6148 | } |
| 6149 | } |
| 6150 | } |
| 6151 | |
| 6152 | static void CalculateVMGroupAndRequestTimes( |
| 6153 | unsigned int NumberOfActivePlanes, |
| 6154 | bool_Bool GPUVMEnable, |
| 6155 | unsigned int GPUVMMaxPageTableLevels, |
| 6156 | unsigned int HTotal[], |
| 6157 | int BytePerPixelC[], |
| 6158 | double DestinationLinesToRequestVMInVBlank[], |
| 6159 | double DestinationLinesToRequestVMInImmediateFlip[], |
| 6160 | bool_Bool DCCEnable[], |
| 6161 | double PixelClock[], |
| 6162 | int dpte_row_width_luma_ub[], |
| 6163 | int dpte_row_width_chroma_ub[], |
| 6164 | int vm_group_bytes[], |
| 6165 | unsigned int dpde0_bytes_per_frame_ub_l[], |
| 6166 | unsigned int dpde0_bytes_per_frame_ub_c[], |
| 6167 | int meta_pte_bytes_per_frame_ub_l[], |
| 6168 | int meta_pte_bytes_per_frame_ub_c[], |
| 6169 | double TimePerVMGroupVBlank[], |
| 6170 | double TimePerVMGroupFlip[], |
| 6171 | double TimePerVMRequestVBlank[], |
| 6172 | double TimePerVMRequestFlip[]) |
| 6173 | { |
| 6174 | int num_group_per_lower_vm_stage; |
| 6175 | int num_req_per_lower_vm_stage; |
| 6176 | int k; |
| 6177 | |
| 6178 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 6179 | if (GPUVMEnable == true1 && (DCCEnable[k] == true1 || GPUVMMaxPageTableLevels > 1)) { |
| 6180 | if (DCCEnable[k] == false0) { |
| 6181 | if (BytePerPixelC[k] > 0) { |
| 6182 | num_group_per_lower_vm_stage = dml_ceil((double) (dpde0_bytes_per_frame_ub_l[k]) / (double) (vm_group_bytes[k]), 1) |
| 6183 | + dml_ceil((double) (dpde0_bytes_per_frame_ub_c[k]) / (double) (vm_group_bytes[k]), 1); |
| 6184 | } else { |
| 6185 | num_group_per_lower_vm_stage = dml_ceil((double) (dpde0_bytes_per_frame_ub_l[k]) / (double) (vm_group_bytes[k]), 1); |
| 6186 | } |
| 6187 | } else { |
| 6188 | if (GPUVMMaxPageTableLevels == 1) { |
| 6189 | if (BytePerPixelC[k] > 0) { |
| 6190 | num_group_per_lower_vm_stage = dml_ceil((double) (meta_pte_bytes_per_frame_ub_l[k]) / (double) (vm_group_bytes[k]), 1) |
| 6191 | + dml_ceil((double) (meta_pte_bytes_per_frame_ub_c[k]) / (double) (vm_group_bytes[k]), 1); |
| 6192 | } else { |
| 6193 | num_group_per_lower_vm_stage = dml_ceil((double) (meta_pte_bytes_per_frame_ub_l[k]) / (double) (vm_group_bytes[k]), 1); |
| 6194 | } |
| 6195 | } else { |
| 6196 | if (BytePerPixelC[k] > 0) { |
| 6197 | num_group_per_lower_vm_stage = 2 + dml_ceil((double) (dpde0_bytes_per_frame_ub_l[k]) / (double) (vm_group_bytes[k]), 1) |
| 6198 | + dml_ceil((double) (dpde0_bytes_per_frame_ub_c[k]) / (double) (vm_group_bytes[k]), 1) |
| 6199 | + dml_ceil((double) (meta_pte_bytes_per_frame_ub_l[k]) / (double) (vm_group_bytes[k]), 1) |
| 6200 | + dml_ceil((double) (meta_pte_bytes_per_frame_ub_c[k]) / (double) (vm_group_bytes[k]), 1); |
| 6201 | } else { |
| 6202 | num_group_per_lower_vm_stage = 1 + dml_ceil((double) (dpde0_bytes_per_frame_ub_l[k]) / (double) (vm_group_bytes[k]), 1) |
| 6203 | + dml_ceil((double) (meta_pte_bytes_per_frame_ub_l[k]) / (double) (vm_group_bytes[k]), 1); |
| 6204 | } |
| 6205 | } |
| 6206 | } |
| 6207 | |
| 6208 | if (DCCEnable[k] == false0) { |
| 6209 | if (BytePerPixelC[k] > 0) { |
| 6210 | num_req_per_lower_vm_stage = dpde0_bytes_per_frame_ub_l[k] / 64 + dpde0_bytes_per_frame_ub_c[k] / 64; |
| 6211 | } else { |
| 6212 | num_req_per_lower_vm_stage = dpde0_bytes_per_frame_ub_l[k] / 64; |
| 6213 | } |
| 6214 | } else { |
| 6215 | if (GPUVMMaxPageTableLevels == 1) { |
| 6216 | if (BytePerPixelC[k] > 0) { |
| 6217 | num_req_per_lower_vm_stage = meta_pte_bytes_per_frame_ub_l[k] / 64 + meta_pte_bytes_per_frame_ub_c[k] / 64; |
| 6218 | } else { |
| 6219 | num_req_per_lower_vm_stage = meta_pte_bytes_per_frame_ub_l[k] / 64; |
| 6220 | } |
| 6221 | } else { |
| 6222 | if (BytePerPixelC[k] > 0) { |
| 6223 | num_req_per_lower_vm_stage = dpde0_bytes_per_frame_ub_l[k] / 64 + dpde0_bytes_per_frame_ub_c[k] / 64 |
| 6224 | + meta_pte_bytes_per_frame_ub_l[k] / 64 + meta_pte_bytes_per_frame_ub_c[k] / 64; |
| 6225 | } else { |
| 6226 | num_req_per_lower_vm_stage = dpde0_bytes_per_frame_ub_l[k] / 64 + meta_pte_bytes_per_frame_ub_l[k] / 64; |
| 6227 | } |
| 6228 | } |
| 6229 | } |
| 6230 | |
| 6231 | TimePerVMGroupVBlank[k] = DestinationLinesToRequestVMInVBlank[k] * HTotal[k] / PixelClock[k] / num_group_per_lower_vm_stage; |
| 6232 | TimePerVMGroupFlip[k] = DestinationLinesToRequestVMInImmediateFlip[k] * HTotal[k] / PixelClock[k] / num_group_per_lower_vm_stage; |
| 6233 | TimePerVMRequestVBlank[k] = DestinationLinesToRequestVMInVBlank[k] * HTotal[k] / PixelClock[k] / num_req_per_lower_vm_stage; |
| 6234 | TimePerVMRequestFlip[k] = DestinationLinesToRequestVMInImmediateFlip[k] * HTotal[k] / PixelClock[k] / num_req_per_lower_vm_stage; |
| 6235 | |
| 6236 | if (GPUVMMaxPageTableLevels > 2) { |
| 6237 | TimePerVMGroupVBlank[k] = TimePerVMGroupVBlank[k] / 2; |
| 6238 | TimePerVMGroupFlip[k] = TimePerVMGroupFlip[k] / 2; |
| 6239 | TimePerVMRequestVBlank[k] = TimePerVMRequestVBlank[k] / 2; |
| 6240 | TimePerVMRequestFlip[k] = TimePerVMRequestFlip[k] / 2; |
| 6241 | } |
| 6242 | |
| 6243 | } else { |
| 6244 | TimePerVMGroupVBlank[k] = 0; |
| 6245 | TimePerVMGroupFlip[k] = 0; |
| 6246 | TimePerVMRequestVBlank[k] = 0; |
| 6247 | TimePerVMRequestFlip[k] = 0; |
| 6248 | } |
| 6249 | } |
| 6250 | } |
| 6251 | |
| 6252 | static void CalculateStutterEfficiency( |
| 6253 | struct display_mode_lib *mode_lib, |
| 6254 | int CompressedBufferSizeInkByte, |
| 6255 | bool_Bool UnboundedRequestEnabled, |
| 6256 | int ConfigReturnBufferSizeInKByte, |
| 6257 | int MetaFIFOSizeInKEntries, |
| 6258 | int ZeroSizeBufferEntries, |
| 6259 | int NumberOfActivePlanes, |
| 6260 | int ROBBufferSizeInKByte, |
| 6261 | double TotalDataReadBandwidth, |
| 6262 | double DCFCLK, |
| 6263 | double ReturnBW, |
| 6264 | double COMPBUF_RESERVED_SPACE_64B, |
| 6265 | double COMPBUF_RESERVED_SPACE_ZS, |
| 6266 | double SRExitTime, |
| 6267 | double SRExitZ8Time, |
| 6268 | bool_Bool SynchronizedVBlank, |
| 6269 | double Z8StutterEnterPlusExitWatermark, |
| 6270 | double StutterEnterPlusExitWatermark, |
| 6271 | bool_Bool ProgressiveToInterlaceUnitInOPP, |
| 6272 | bool_Bool Interlace[], |
| 6273 | double MinTTUVBlank[], |
| 6274 | int DPPPerPlane[], |
| 6275 | unsigned int DETBufferSizeY[], |
| 6276 | int BytePerPixelY[], |
| 6277 | double BytePerPixelDETY[], |
| 6278 | double SwathWidthY[], |
| 6279 | int SwathHeightY[], |
| 6280 | int SwathHeightC[], |
| 6281 | double NetDCCRateLuma[], |
| 6282 | double NetDCCRateChroma[], |
| 6283 | double DCCFractionOfZeroSizeRequestsLuma[], |
| 6284 | double DCCFractionOfZeroSizeRequestsChroma[], |
| 6285 | int HTotal[], |
| 6286 | int VTotal[], |
| 6287 | double PixelClock[], |
| 6288 | double VRatio[], |
| 6289 | enum scan_direction_class SourceScan[], |
| 6290 | int BlockHeight256BytesY[], |
| 6291 | int BlockWidth256BytesY[], |
| 6292 | int BlockHeight256BytesC[], |
| 6293 | int BlockWidth256BytesC[], |
| 6294 | int DCCYMaxUncompressedBlock[], |
| 6295 | int DCCCMaxUncompressedBlock[], |
| 6296 | int VActive[], |
| 6297 | bool_Bool DCCEnable[], |
| 6298 | bool_Bool WritebackEnable[], |
| 6299 | double ReadBandwidthPlaneLuma[], |
| 6300 | double ReadBandwidthPlaneChroma[], |
| 6301 | double meta_row_bw[], |
| 6302 | double dpte_row_bw[], |
| 6303 | double *StutterEfficiencyNotIncludingVBlank, |
| 6304 | double *StutterEfficiency, |
| 6305 | int *NumberOfStutterBurstsPerFrame, |
| 6306 | double *Z8StutterEfficiencyNotIncludingVBlank, |
| 6307 | double *Z8StutterEfficiency, |
| 6308 | int *Z8NumberOfStutterBurstsPerFrame, |
| 6309 | double *StutterPeriod) |
| 6310 | { |
| 6311 | struct vba_vars_st *v = &mode_lib->vba; |
| 6312 | |
| 6313 | double DETBufferingTimeY; |
| 6314 | double SwathWidthYCriticalPlane = 0; |
| 6315 | double VActiveTimeCriticalPlane = 0; |
| 6316 | double FrameTimeCriticalPlane = 0; |
| 6317 | int BytePerPixelYCriticalPlane = 0; |
| 6318 | double LinesToFinishSwathTransferStutterCriticalPlane = 0; |
| 6319 | double MinTTUVBlankCriticalPlane = 0; |
| 6320 | double TotalCompressedReadBandwidth; |
| 6321 | double TotalRowReadBandwidth; |
| 6322 | double AverageDCCCompressionRate; |
| 6323 | double EffectiveCompressedBufferSize; |
| 6324 | double PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer; |
| 6325 | double StutterBurstTime; |
| 6326 | int TotalActiveWriteback; |
| 6327 | double LinesInDETY; |
| 6328 | double LinesInDETYRoundedDownToSwath; |
| 6329 | double MaximumEffectiveCompressionLuma; |
| 6330 | double MaximumEffectiveCompressionChroma; |
| 6331 | double TotalZeroSizeRequestReadBandwidth; |
| 6332 | double TotalZeroSizeCompressedReadBandwidth; |
| 6333 | double AverageDCCZeroSizeFraction; |
| 6334 | double AverageZeroSizeCompressionRate; |
| 6335 | int TotalNumberOfActiveOTG = 0; |
| 6336 | double LastStutterPeriod = 0.0; |
| 6337 | double LastZ8StutterPeriod = 0.0; |
| 6338 | int k; |
| 6339 | |
| 6340 | TotalZeroSizeRequestReadBandwidth = 0; |
| 6341 | TotalZeroSizeCompressedReadBandwidth = 0; |
| 6342 | TotalRowReadBandwidth = 0; |
| 6343 | TotalCompressedReadBandwidth = 0; |
| 6344 | |
| 6345 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 6346 | if (DCCEnable[k] == true1) { |
| 6347 | if ((SourceScan[k] == dm_vert && BlockWidth256BytesY[k] > SwathHeightY[k]) || (SourceScan[k] != dm_vert && BlockHeight256BytesY[k] > SwathHeightY[k]) |
| 6348 | || DCCYMaxUncompressedBlock[k] < 256) { |
| 6349 | MaximumEffectiveCompressionLuma = 2; |
| 6350 | } else { |
| 6351 | MaximumEffectiveCompressionLuma = 4; |
| 6352 | } |
| 6353 | TotalCompressedReadBandwidth = TotalCompressedReadBandwidth + ReadBandwidthPlaneLuma[k] / dml_min(NetDCCRateLuma[k], MaximumEffectiveCompressionLuma); |
| 6354 | TotalZeroSizeRequestReadBandwidth = TotalZeroSizeRequestReadBandwidth + ReadBandwidthPlaneLuma[k] * DCCFractionOfZeroSizeRequestsLuma[k]; |
| 6355 | TotalZeroSizeCompressedReadBandwidth = TotalZeroSizeCompressedReadBandwidth |
| 6356 | + ReadBandwidthPlaneLuma[k] * DCCFractionOfZeroSizeRequestsLuma[k] / MaximumEffectiveCompressionLuma; |
| 6357 | if (ReadBandwidthPlaneChroma[k] > 0) { |
| 6358 | if ((SourceScan[k] == dm_vert && BlockWidth256BytesC[k] > SwathHeightC[k]) |
| 6359 | || (SourceScan[k] != dm_vert && BlockHeight256BytesC[k] > SwathHeightC[k]) || DCCCMaxUncompressedBlock[k] < 256) { |
| 6360 | MaximumEffectiveCompressionChroma = 2; |
| 6361 | } else { |
| 6362 | MaximumEffectiveCompressionChroma = 4; |
| 6363 | } |
| 6364 | TotalCompressedReadBandwidth = TotalCompressedReadBandwidth |
| 6365 | + ReadBandwidthPlaneChroma[k] / dml_min(NetDCCRateChroma[k], MaximumEffectiveCompressionChroma); |
| 6366 | TotalZeroSizeRequestReadBandwidth = TotalZeroSizeRequestReadBandwidth + ReadBandwidthPlaneChroma[k] * DCCFractionOfZeroSizeRequestsChroma[k]; |
| 6367 | TotalZeroSizeCompressedReadBandwidth = TotalZeroSizeCompressedReadBandwidth |
| 6368 | + ReadBandwidthPlaneChroma[k] * DCCFractionOfZeroSizeRequestsChroma[k] / MaximumEffectiveCompressionChroma; |
| 6369 | } |
| 6370 | } else { |
| 6371 | TotalCompressedReadBandwidth = TotalCompressedReadBandwidth + ReadBandwidthPlaneLuma[k] + ReadBandwidthPlaneChroma[k]; |
| 6372 | } |
| 6373 | TotalRowReadBandwidth = TotalRowReadBandwidth + DPPPerPlane[k] * (meta_row_bw[k] + dpte_row_bw[k]); |
| 6374 | } |
| 6375 | |
| 6376 | AverageDCCCompressionRate = TotalDataReadBandwidth / TotalCompressedReadBandwidth; |
| 6377 | AverageDCCZeroSizeFraction = TotalZeroSizeRequestReadBandwidth / TotalDataReadBandwidth; |
| 6378 | |
| 6379 | #ifdef __DML_VBA_DEBUG__ |
| 6380 | dml_print("DML::%s: TotalCompressedReadBandwidth = %f\n", __func__, TotalCompressedReadBandwidth){do { } while(0); }; |
| 6381 | dml_print("DML::%s: TotalZeroSizeRequestReadBandwidth = %f\n", __func__, TotalZeroSizeRequestReadBandwidth){do { } while(0); }; |
| 6382 | dml_print("DML::%s: TotalZeroSizeCompressedReadBandwidth = %f\n", __func__, TotalZeroSizeCompressedReadBandwidth){do { } while(0); }; |
| 6383 | dml_print("DML::%s: MaximumEffectiveCompressionLuma = %f\n", __func__, MaximumEffectiveCompressionLuma){do { } while(0); }; |
| 6384 | dml_print("DML::%s: MaximumEffectiveCompressionChroma = %f\n", __func__, MaximumEffectiveCompressionChroma){do { } while(0); }; |
| 6385 | dml_print("DML::%s: AverageDCCCompressionRate = %f\n", __func__, AverageDCCCompressionRate){do { } while(0); }; |
| 6386 | dml_print("DML::%s: AverageDCCZeroSizeFraction = %f\n", __func__, AverageDCCZeroSizeFraction){do { } while(0); }; |
| 6387 | dml_print("DML::%s: CompressedBufferSizeInkByte = %d\n", __func__, CompressedBufferSizeInkByte){do { } while(0); }; |
| 6388 | #endif |
| 6389 | |
| 6390 | if (AverageDCCZeroSizeFraction == 1) { |
| 6391 | AverageZeroSizeCompressionRate = TotalZeroSizeRequestReadBandwidth / TotalZeroSizeCompressedReadBandwidth; |
| 6392 | EffectiveCompressedBufferSize = MetaFIFOSizeInKEntries * 1024 * 64 * AverageZeroSizeCompressionRate + (ZeroSizeBufferEntries - COMPBUF_RESERVED_SPACE_ZS) * 64 * AverageZeroSizeCompressionRate; |
| 6393 | } else if (AverageDCCZeroSizeFraction > 0) { |
| 6394 | AverageZeroSizeCompressionRate = TotalZeroSizeRequestReadBandwidth / TotalZeroSizeCompressedReadBandwidth; |
| 6395 | EffectiveCompressedBufferSize = dml_min( |
| 6396 | CompressedBufferSizeInkByte * 1024 * AverageDCCCompressionRate, |
| 6397 | MetaFIFOSizeInKEntries * 1024 * 64 / (AverageDCCZeroSizeFraction / AverageZeroSizeCompressionRate + 1 / AverageDCCCompressionRate)) |
| 6398 | + dml_min((ROBBufferSizeInKByte * 1024 - COMPBUF_RESERVED_SPACE_64B * 64) * AverageDCCCompressionRate, |
| 6399 | (ZeroSizeBufferEntries - COMPBUF_RESERVED_SPACE_ZS) * 64 / (AverageDCCZeroSizeFraction / AverageZeroSizeCompressionRate)); |
| 6400 | dml_print("DML::%s: min 1 = %f\n", __func__, CompressedBufferSizeInkByte * 1024 * AverageDCCCompressionRate){do { } while(0); }; |
| 6401 | dml_print({do { } while(0); } |
| 6402 | "DML::%s: min 2 = %f\n",{do { } while(0); } |
| 6403 | __func__,{do { } while(0); } |
| 6404 | MetaFIFOSizeInKEntries * 1024 * 64 / (AverageDCCZeroSizeFraction / AverageZeroSizeCompressionRate + 1 / AverageDCCCompressionRate)){do { } while(0); }; |
| 6405 | dml_print("DML::%s: min 3 = %f\n", __func__, ROBBufferSizeInKByte * 1024 * AverageDCCCompressionRate){do { } while(0); }; |
| 6406 | dml_print("DML::%s: min 4 = %f\n", __func__, ZeroSizeBufferEntries * 64 / (AverageDCCZeroSizeFraction / AverageZeroSizeCompressionRate)){do { } while(0); }; |
| 6407 | } else { |
| 6408 | EffectiveCompressedBufferSize = dml_min( |
| 6409 | CompressedBufferSizeInkByte * 1024 * AverageDCCCompressionRate, |
| 6410 | MetaFIFOSizeInKEntries * 1024 * 64 * AverageDCCCompressionRate) + (ROBBufferSizeInKByte * 1024 - COMPBUF_RESERVED_SPACE_64B * 64) * AverageDCCCompressionRate; |
| 6411 | dml_print("DML::%s: min 1 = %f\n", __func__, CompressedBufferSizeInkByte * 1024 * AverageDCCCompressionRate){do { } while(0); }; |
| 6412 | dml_print("DML::%s: min 2 = %f\n", __func__, MetaFIFOSizeInKEntries * 1024 * 64 * AverageDCCCompressionRate){do { } while(0); }; |
| 6413 | } |
| 6414 | |
| 6415 | #ifdef __DML_VBA_DEBUG__ |
| 6416 | dml_print("DML::%s: MetaFIFOSizeInKEntries = %d\n", __func__, MetaFIFOSizeInKEntries){do { } while(0); }; |
| 6417 | dml_print("DML::%s: AverageZeroSizeCompressionRate = %f\n", __func__, AverageZeroSizeCompressionRate){do { } while(0); }; |
| 6418 | dml_print("DML::%s: EffectiveCompressedBufferSize = %f\n", __func__, EffectiveCompressedBufferSize){do { } while(0); }; |
| 6419 | #endif |
| 6420 | |
| 6421 | *StutterPeriod = 0; |
| 6422 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 6423 | LinesInDETY = (DETBufferSizeY[k] + (UnboundedRequestEnabled == true1 ? EffectiveCompressedBufferSize : 0) * ReadBandwidthPlaneLuma[k] / TotalDataReadBandwidth) |
| 6424 | / BytePerPixelDETY[k] / SwathWidthY[k]; |
| 6425 | LinesInDETYRoundedDownToSwath = dml_floor(LinesInDETY, SwathHeightY[k]); |
| 6426 | DETBufferingTimeY = LinesInDETYRoundedDownToSwath * (HTotal[k] / PixelClock[k]) / VRatio[k]; |
| 6427 | #ifdef __DML_VBA_DEBUG__ |
| 6428 | dml_print("DML::%s: k=%0d DETBufferSizeY = %f\n", __func__, k, DETBufferSizeY[k]){do { } while(0); }; |
| 6429 | dml_print("DML::%s: k=%0d BytePerPixelDETY = %f\n", __func__, k, BytePerPixelDETY[k]){do { } while(0); }; |
| 6430 | dml_print("DML::%s: k=%0d SwathWidthY = %f\n", __func__, k, SwathWidthY[k]){do { } while(0); }; |
| 6431 | dml_print("DML::%s: k=%0d ReadBandwidthPlaneLuma = %f\n", __func__, k, ReadBandwidthPlaneLuma[k]){do { } while(0); }; |
| 6432 | dml_print("DML::%s: k=%0d TotalDataReadBandwidth = %f\n", __func__, k, TotalDataReadBandwidth){do { } while(0); }; |
| 6433 | dml_print("DML::%s: k=%0d LinesInDETY = %f\n", __func__, k, LinesInDETY){do { } while(0); }; |
| 6434 | dml_print("DML::%s: k=%0d LinesInDETYRoundedDownToSwath = %f\n", __func__, k, LinesInDETYRoundedDownToSwath){do { } while(0); }; |
| 6435 | dml_print("DML::%s: k=%0d HTotal = %d\n", __func__, k, HTotal[k]){do { } while(0); }; |
| 6436 | dml_print("DML::%s: k=%0d PixelClock = %f\n", __func__, k, PixelClock[k]){do { } while(0); }; |
| 6437 | dml_print("DML::%s: k=%0d VRatio = %f\n", __func__, k, VRatio[k]){do { } while(0); }; |
| 6438 | dml_print("DML::%s: k=%0d DETBufferingTimeY = %f\n", __func__, k, DETBufferingTimeY){do { } while(0); }; |
| 6439 | dml_print("DML::%s: k=%0d PixelClock = %f\n", __func__, k, PixelClock[k]){do { } while(0); }; |
| 6440 | #endif |
| 6441 | |
| 6442 | if (k == 0 || DETBufferingTimeY < *StutterPeriod) { |
| 6443 | bool_Bool isInterlaceTiming = Interlace[k] && !ProgressiveToInterlaceUnitInOPP; |
| 6444 | |
| 6445 | *StutterPeriod = DETBufferingTimeY; |
| 6446 | FrameTimeCriticalPlane = (isInterlaceTiming ? dml_floor(VTotal[k] / 2.0, 1.0) : VTotal[k]) * HTotal[k] / PixelClock[k]; |
| 6447 | VActiveTimeCriticalPlane = (isInterlaceTiming ? dml_floor(VActive[k] / 2.0, 1.0) : VActive[k]) * HTotal[k] / PixelClock[k]; |
| 6448 | BytePerPixelYCriticalPlane = BytePerPixelY[k]; |
| 6449 | SwathWidthYCriticalPlane = SwathWidthY[k]; |
| 6450 | LinesToFinishSwathTransferStutterCriticalPlane = SwathHeightY[k] - (LinesInDETY - LinesInDETYRoundedDownToSwath); |
| 6451 | MinTTUVBlankCriticalPlane = MinTTUVBlank[k]; |
| 6452 | |
| 6453 | #ifdef __DML_VBA_DEBUG__ |
| 6454 | dml_print("DML::%s: StutterPeriod = %f\n", __func__, *StutterPeriod){do { } while(0); }; |
| 6455 | dml_print("DML::%s: MinTTUVBlankCriticalPlane = %f\n", __func__, MinTTUVBlankCriticalPlane){do { } while(0); }; |
| 6456 | dml_print("DML::%s: FrameTimeCriticalPlane = %f\n", __func__, FrameTimeCriticalPlane){do { } while(0); }; |
| 6457 | dml_print("DML::%s: VActiveTimeCriticalPlane = %f\n", __func__, VActiveTimeCriticalPlane){do { } while(0); }; |
| 6458 | dml_print("DML::%s: BytePerPixelYCriticalPlane = %d\n", __func__, BytePerPixelYCriticalPlane){do { } while(0); }; |
| 6459 | dml_print("DML::%s: SwathWidthYCriticalPlane = %f\n", __func__, SwathWidthYCriticalPlane){do { } while(0); }; |
| 6460 | dml_print("DML::%s: LinesToFinishSwathTransferStutterCriticalPlane = %f\n", __func__, LinesToFinishSwathTransferStutterCriticalPlane){do { } while(0); }; |
| 6461 | #endif |
| 6462 | } |
| 6463 | } |
| 6464 | |
| 6465 | PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer = dml_min(*StutterPeriod * TotalDataReadBandwidth, EffectiveCompressedBufferSize); |
| 6466 | #ifdef __DML_VBA_DEBUG__ |
| 6467 | dml_print("DML::%s: ROBBufferSizeInKByte = %d\n", __func__, ROBBufferSizeInKByte){do { } while(0); }; |
| 6468 | dml_print("DML::%s: AverageDCCCompressionRate = %f\n", __func__, AverageDCCCompressionRate){do { } while(0); }; |
| 6469 | dml_print("DML::%s: StutterPeriod * TotalDataReadBandwidth = %f\n", __func__, *StutterPeriod * TotalDataReadBandwidth){do { } while(0); }; |
| 6470 | dml_print("DML::%s: ROBBufferSizeInKByte * 1024 * AverageDCCCompressionRate + EffectiveCompressedBufferSize = %f\n", __func__, ROBBufferSizeInKByte * 1024 * AverageDCCCompressionRate + EffectiveCompressedBufferSize){do { } while(0); }; |
| 6471 | dml_print("DML::%s: EffectiveCompressedBufferSize = %f\n", __func__, EffectiveCompressedBufferSize){do { } while(0); }; |
| 6472 | dml_print("DML::%s: PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer = %f\n", __func__, PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer){do { } while(0); }; |
| 6473 | dml_print("DML::%s: ReturnBW = %f\n", __func__, ReturnBW){do { } while(0); }; |
| 6474 | dml_print("DML::%s: TotalDataReadBandwidth = %f\n", __func__, TotalDataReadBandwidth){do { } while(0); }; |
| 6475 | dml_print("DML::%s: TotalRowReadBandwidth = %f\n", __func__, TotalRowReadBandwidth){do { } while(0); }; |
| 6476 | dml_print("DML::%s: DCFCLK = %f\n", __func__, DCFCLK){do { } while(0); }; |
| 6477 | #endif |
| 6478 | |
| 6479 | StutterBurstTime = PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer / AverageDCCCompressionRate / ReturnBW |
| 6480 | + (*StutterPeriod * TotalDataReadBandwidth - PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer) / (DCFCLK * 64) |
| 6481 | + *StutterPeriod * TotalRowReadBandwidth / ReturnBW; |
| 6482 | #ifdef __DML_VBA_DEBUG__ |
| 6483 | dml_print("DML::%s: Part 1 = %f\n", __func__, PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer / AverageDCCCompressionRate / ReturnBW){do { } while(0); }; |
| 6484 | dml_print("DML::%s: StutterPeriod * TotalDataReadBandwidth = %f\n", __func__, (*StutterPeriod * TotalDataReadBandwidth)){do { } while(0); }; |
| 6485 | dml_print("DML::%s: Part 2 = %f\n", __func__, (*StutterPeriod * TotalDataReadBandwidth - PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer) / (DCFCLK * 64)){do { } while(0); }; |
| 6486 | dml_print("DML::%s: Part 3 = %f\n", __func__, *StutterPeriod * TotalRowReadBandwidth / ReturnBW){do { } while(0); }; |
| 6487 | dml_print("DML::%s: StutterBurstTime = %f\n", __func__, StutterBurstTime){do { } while(0); }; |
| 6488 | #endif |
| 6489 | StutterBurstTime = dml_max(StutterBurstTime, LinesToFinishSwathTransferStutterCriticalPlane * BytePerPixelYCriticalPlane * SwathWidthYCriticalPlane / ReturnBW); |
| 6490 | |
| 6491 | dml_print({do { } while(0); } |
| 6492 | "DML::%s: Time to finish residue swath=%f\n",{do { } while(0); } |
| 6493 | __func__,{do { } while(0); } |
| 6494 | LinesToFinishSwathTransferStutterCriticalPlane * BytePerPixelYCriticalPlane * SwathWidthYCriticalPlane / ReturnBW){do { } while(0); }; |
| 6495 | |
| 6496 | TotalActiveWriteback = 0; |
| 6497 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 6498 | if (WritebackEnable[k]) { |
| 6499 | TotalActiveWriteback = TotalActiveWriteback + 1; |
| 6500 | } |
| 6501 | } |
| 6502 | |
| 6503 | if (TotalActiveWriteback == 0) { |
| 6504 | #ifdef __DML_VBA_DEBUG__ |
| 6505 | dml_print("DML::%s: SRExitTime = %f\n", __func__, SRExitTime){do { } while(0); }; |
| 6506 | dml_print("DML::%s: SRExitZ8Time = %f\n", __func__, SRExitZ8Time){do { } while(0); }; |
| 6507 | dml_print("DML::%s: StutterBurstTime = %f (final)\n", __func__, StutterBurstTime){do { } while(0); }; |
| 6508 | dml_print("DML::%s: StutterPeriod = %f\n", __func__, *StutterPeriod){do { } while(0); }; |
| 6509 | #endif |
| 6510 | *StutterEfficiencyNotIncludingVBlank = dml_max(0., 1 - (SRExitTime + StutterBurstTime) / *StutterPeriod) * 100; |
| 6511 | *Z8StutterEfficiencyNotIncludingVBlank = dml_max(0., 1 - (SRExitZ8Time + StutterBurstTime) / *StutterPeriod) * 100; |
| 6512 | *NumberOfStutterBurstsPerFrame = (*StutterEfficiencyNotIncludingVBlank > 0 ? dml_ceil(VActiveTimeCriticalPlane / *StutterPeriod, 1) : 0); |
| 6513 | *Z8NumberOfStutterBurstsPerFrame = (*Z8StutterEfficiencyNotIncludingVBlank > 0 ? dml_ceil(VActiveTimeCriticalPlane / *StutterPeriod, 1) : 0); |
| 6514 | } else { |
| 6515 | *StutterEfficiencyNotIncludingVBlank = 0.; |
| 6516 | *Z8StutterEfficiencyNotIncludingVBlank = 0.; |
| 6517 | *NumberOfStutterBurstsPerFrame = 0; |
| 6518 | *Z8NumberOfStutterBurstsPerFrame = 0; |
| 6519 | } |
| 6520 | #ifdef __DML_VBA_DEBUG__ |
| 6521 | dml_print("DML::%s: VActiveTimeCriticalPlane = %f\n", __func__, VActiveTimeCriticalPlane){do { } while(0); }; |
| 6522 | dml_print("DML::%s: StutterEfficiencyNotIncludingVBlank = %f\n", __func__, *StutterEfficiencyNotIncludingVBlank){do { } while(0); }; |
| 6523 | dml_print("DML::%s: Z8StutterEfficiencyNotIncludingVBlank = %f\n", __func__, *Z8StutterEfficiencyNotIncludingVBlank){do { } while(0); }; |
| 6524 | dml_print("DML::%s: NumberOfStutterBurstsPerFrame = %d\n", __func__, *NumberOfStutterBurstsPerFrame){do { } while(0); }; |
| 6525 | dml_print("DML::%s: Z8NumberOfStutterBurstsPerFrame = %d\n", __func__, *Z8NumberOfStutterBurstsPerFrame){do { } while(0); }; |
| 6526 | #endif |
| 6527 | |
| 6528 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 6529 | if (v->BlendingAndTiming[k] == k) { |
| 6530 | TotalNumberOfActiveOTG = TotalNumberOfActiveOTG + 1; |
| 6531 | } |
| 6532 | } |
| 6533 | |
| 6534 | if (*StutterEfficiencyNotIncludingVBlank > 0) { |
| 6535 | LastStutterPeriod = VActiveTimeCriticalPlane - (*NumberOfStutterBurstsPerFrame - 1) * *StutterPeriod; |
| 6536 | |
| 6537 | if ((SynchronizedVBlank || TotalNumberOfActiveOTG == 1) && LastStutterPeriod + MinTTUVBlankCriticalPlane > StutterEnterPlusExitWatermark) { |
| 6538 | *StutterEfficiency = (1 - (*NumberOfStutterBurstsPerFrame * SRExitTime + StutterBurstTime * VActiveTimeCriticalPlane |
| 6539 | / *StutterPeriod) / FrameTimeCriticalPlane) * 100; |
| 6540 | } else { |
| 6541 | *StutterEfficiency = *StutterEfficiencyNotIncludingVBlank; |
| 6542 | } |
| 6543 | } else { |
| 6544 | *StutterEfficiency = 0; |
| 6545 | } |
| 6546 | |
| 6547 | if (*Z8StutterEfficiencyNotIncludingVBlank > 0) { |
| 6548 | LastZ8StutterPeriod = VActiveTimeCriticalPlane - (*NumberOfStutterBurstsPerFrame - 1) * *StutterPeriod; |
| 6549 | if ((SynchronizedVBlank || TotalNumberOfActiveOTG == 1) && LastZ8StutterPeriod + MinTTUVBlankCriticalPlane > Z8StutterEnterPlusExitWatermark) { |
| 6550 | *Z8StutterEfficiency = (1 - (*NumberOfStutterBurstsPerFrame * SRExitZ8Time + StutterBurstTime * VActiveTimeCriticalPlane |
| 6551 | / *StutterPeriod) / FrameTimeCriticalPlane) * 100; |
| 6552 | } else { |
| 6553 | *Z8StutterEfficiency = *Z8StutterEfficiencyNotIncludingVBlank; |
| 6554 | } |
| 6555 | } else { |
| 6556 | *Z8StutterEfficiency = 0.; |
| 6557 | } |
| 6558 | |
| 6559 | dml_print("DML::%s: LastZ8StutterPeriod = %f\n", __func__, LastZ8StutterPeriod){do { } while(0); }; |
| 6560 | dml_print("DML::%s: Z8StutterEnterPlusExitWatermark = %f\n", __func__, Z8StutterEnterPlusExitWatermark){do { } while(0); }; |
| 6561 | dml_print("DML::%s: StutterBurstTime = %f\n", __func__, StutterBurstTime){do { } while(0); }; |
| 6562 | dml_print("DML::%s: StutterPeriod = %f\n", __func__, *StutterPeriod){do { } while(0); }; |
| 6563 | dml_print("DML::%s: StutterEfficiency = %f\n", __func__, *StutterEfficiency){do { } while(0); }; |
| 6564 | dml_print("DML::%s: Z8StutterEfficiency = %f\n", __func__, *Z8StutterEfficiency){do { } while(0); }; |
| 6565 | dml_print("DML::%s: StutterEfficiencyNotIncludingVBlank = %f\n", __func__, *StutterEfficiencyNotIncludingVBlank){do { } while(0); }; |
| 6566 | dml_print("DML::%s: Z8NumberOfStutterBurstsPerFrame = %d\n", __func__, *Z8NumberOfStutterBurstsPerFrame){do { } while(0); }; |
| 6567 | } |
| 6568 | |
| 6569 | static void CalculateSwathAndDETConfiguration( |
| 6570 | bool_Bool ForceSingleDPP, |
| 6571 | int NumberOfActivePlanes, |
| 6572 | bool_Bool DETSharedByAllDPP, |
| 6573 | unsigned int DETBufferSizeInKByteA[], |
| 6574 | double MaximumSwathWidthLuma[], |
| 6575 | double MaximumSwathWidthChroma[], |
| 6576 | enum scan_direction_class SourceScan[], |
| 6577 | enum source_format_class SourcePixelFormat[], |
| 6578 | enum dm_swizzle_mode SurfaceTiling[], |
| 6579 | int ViewportWidth[], |
| 6580 | int ViewportHeight[], |
| 6581 | int SurfaceWidthY[], |
| 6582 | int SurfaceWidthC[], |
| 6583 | int SurfaceHeightY[], |
| 6584 | int SurfaceHeightC[], |
| 6585 | int Read256BytesBlockHeightY[], |
| 6586 | int Read256BytesBlockHeightC[], |
| 6587 | int Read256BytesBlockWidthY[], |
| 6588 | int Read256BytesBlockWidthC[], |
| 6589 | enum odm_combine_mode ODMCombineEnabled[], |
| 6590 | int BlendingAndTiming[], |
| 6591 | int BytePerPixY[], |
| 6592 | int BytePerPixC[], |
| 6593 | double BytePerPixDETY[], |
| 6594 | double BytePerPixDETC[], |
| 6595 | int HActive[], |
| 6596 | double HRatio[], |
| 6597 | double HRatioChroma[], |
| 6598 | int DPPPerPlane[], |
| 6599 | int swath_width_luma_ub[], |
| 6600 | int swath_width_chroma_ub[], |
| 6601 | double SwathWidth[], |
| 6602 | double SwathWidthChroma[], |
| 6603 | int SwathHeightY[], |
| 6604 | int SwathHeightC[], |
| 6605 | unsigned int DETBufferSizeY[], |
| 6606 | unsigned int DETBufferSizeC[], |
| 6607 | bool_Bool ViewportSizeSupportPerPlane[], |
| 6608 | bool_Bool *ViewportSizeSupport) |
| 6609 | { |
| 6610 | int MaximumSwathHeightY[DC__NUM_DPP__MAX8]; |
| 6611 | int MaximumSwathHeightC[DC__NUM_DPP__MAX8]; |
| 6612 | int MinimumSwathHeightY; |
| 6613 | int MinimumSwathHeightC; |
| 6614 | int RoundedUpMaxSwathSizeBytesY; |
| 6615 | int RoundedUpMaxSwathSizeBytesC; |
| 6616 | int RoundedUpMinSwathSizeBytesY; |
| 6617 | int RoundedUpMinSwathSizeBytesC; |
| 6618 | int RoundedUpSwathSizeBytesY; |
| 6619 | int RoundedUpSwathSizeBytesC; |
| 6620 | double SwathWidthSingleDPP[DC__NUM_DPP__MAX8]; |
| 6621 | double SwathWidthSingleDPPChroma[DC__NUM_DPP__MAX8]; |
| 6622 | int k; |
| 6623 | |
| 6624 | CalculateSwathWidth( |
| 6625 | ForceSingleDPP, |
| 6626 | NumberOfActivePlanes, |
| 6627 | SourcePixelFormat, |
| 6628 | SourceScan, |
| 6629 | ViewportWidth, |
| 6630 | ViewportHeight, |
| 6631 | SurfaceWidthY, |
| 6632 | SurfaceWidthC, |
| 6633 | SurfaceHeightY, |
| 6634 | SurfaceHeightC, |
| 6635 | ODMCombineEnabled, |
| 6636 | BytePerPixY, |
| 6637 | BytePerPixC, |
| 6638 | Read256BytesBlockHeightY, |
| 6639 | Read256BytesBlockHeightC, |
| 6640 | Read256BytesBlockWidthY, |
| 6641 | Read256BytesBlockWidthC, |
| 6642 | BlendingAndTiming, |
| 6643 | HActive, |
| 6644 | HRatio, |
| 6645 | DPPPerPlane, |
| 6646 | SwathWidthSingleDPP, |
| 6647 | SwathWidthSingleDPPChroma, |
| 6648 | SwathWidth, |
| 6649 | SwathWidthChroma, |
| 6650 | MaximumSwathHeightY, |
| 6651 | MaximumSwathHeightC, |
| 6652 | swath_width_luma_ub, |
| 6653 | swath_width_chroma_ub); |
| 6654 | |
| 6655 | *ViewportSizeSupport = true1; |
| 6656 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 6657 | unsigned int DETBufferSizeInKByte = DETBufferSizeInKByteA[k]; |
| 6658 | |
| 6659 | if (DETSharedByAllDPP && DPPPerPlane[k]) |
| 6660 | DETBufferSizeInKByte /= DPPPerPlane[k]; |
| 6661 | if ((SourcePixelFormat[k] == dm_444_64 || SourcePixelFormat[k] == dm_444_32 || SourcePixelFormat[k] == dm_444_16 || SourcePixelFormat[k] == dm_mono_16 |
| 6662 | || SourcePixelFormat[k] == dm_mono_8 || SourcePixelFormat[k] == dm_rgbe)) { |
| 6663 | if (SurfaceTiling[k] == dm_sw_linear |
| 6664 | || (SourcePixelFormat[k] == dm_444_64 |
| 6665 | && (SurfaceTiling[k] == dm_sw_64kb_s || SurfaceTiling[k] == dm_sw_64kb_s_t || SurfaceTiling[k] == dm_sw_64kb_s_x) |
| 6666 | && SourceScan[k] != dm_vert)) { |
| 6667 | MinimumSwathHeightY = MaximumSwathHeightY[k]; |
| 6668 | } else if (SourcePixelFormat[k] == dm_444_8 && SourceScan[k] == dm_vert) { |
| 6669 | MinimumSwathHeightY = MaximumSwathHeightY[k]; |
| 6670 | } else { |
| 6671 | MinimumSwathHeightY = MaximumSwathHeightY[k] / 2; |
| 6672 | } |
| 6673 | MinimumSwathHeightC = MaximumSwathHeightC[k]; |
| 6674 | } else { |
| 6675 | if (SurfaceTiling[k] == dm_sw_linear) { |
| 6676 | MinimumSwathHeightY = MaximumSwathHeightY[k]; |
| 6677 | MinimumSwathHeightC = MaximumSwathHeightC[k]; |
| 6678 | } else if (SourcePixelFormat[k] == dm_rgbe_alpha && SourceScan[k] == dm_vert) { |
| 6679 | MinimumSwathHeightY = MaximumSwathHeightY[k] / 2; |
| 6680 | MinimumSwathHeightC = MaximumSwathHeightC[k]; |
| 6681 | } else if (SourcePixelFormat[k] == dm_rgbe_alpha) { |
| 6682 | MinimumSwathHeightY = MaximumSwathHeightY[k] / 2; |
| 6683 | MinimumSwathHeightC = MaximumSwathHeightC[k] / 2; |
| 6684 | } else if (SourcePixelFormat[k] == dm_420_8 && SourceScan[k] == dm_vert) { |
| 6685 | MinimumSwathHeightY = MaximumSwathHeightY[k]; |
| 6686 | MinimumSwathHeightC = MaximumSwathHeightC[k] / 2; |
| 6687 | } else { |
| 6688 | MinimumSwathHeightC = MaximumSwathHeightC[k] / 2; |
| 6689 | MinimumSwathHeightY = MaximumSwathHeightY[k] / 2; |
| 6690 | } |
| 6691 | } |
| 6692 | |
| 6693 | RoundedUpMaxSwathSizeBytesY = swath_width_luma_ub[k] * BytePerPixDETY[k] * MaximumSwathHeightY[k]; |
| 6694 | RoundedUpMinSwathSizeBytesY = swath_width_luma_ub[k] * BytePerPixDETY[k] * MinimumSwathHeightY; |
| 6695 | if (SourcePixelFormat[k] == dm_420_10) { |
| 6696 | RoundedUpMaxSwathSizeBytesY = dml_ceil((double) RoundedUpMaxSwathSizeBytesY, 256); |
| 6697 | RoundedUpMinSwathSizeBytesY = dml_ceil((double) RoundedUpMinSwathSizeBytesY, 256); |
| 6698 | } |
| 6699 | RoundedUpMaxSwathSizeBytesC = swath_width_chroma_ub[k] * BytePerPixDETC[k] * MaximumSwathHeightC[k]; |
| 6700 | RoundedUpMinSwathSizeBytesC = swath_width_chroma_ub[k] * BytePerPixDETC[k] * MinimumSwathHeightC; |
| 6701 | if (SourcePixelFormat[k] == dm_420_10) { |
| 6702 | RoundedUpMaxSwathSizeBytesC = dml_ceil(RoundedUpMaxSwathSizeBytesC, 256); |
| 6703 | RoundedUpMinSwathSizeBytesC = dml_ceil(RoundedUpMinSwathSizeBytesC, 256); |
| 6704 | } |
| 6705 | |
| 6706 | if (RoundedUpMaxSwathSizeBytesY + RoundedUpMaxSwathSizeBytesC <= DETBufferSizeInKByte * 1024 / 2) { |
| 6707 | SwathHeightY[k] = MaximumSwathHeightY[k]; |
| 6708 | SwathHeightC[k] = MaximumSwathHeightC[k]; |
| 6709 | RoundedUpSwathSizeBytesY = RoundedUpMaxSwathSizeBytesY; |
| 6710 | RoundedUpSwathSizeBytesC = RoundedUpMaxSwathSizeBytesC; |
| 6711 | } else if (RoundedUpMaxSwathSizeBytesY >= 1.5 * RoundedUpMaxSwathSizeBytesC |
| 6712 | && RoundedUpMinSwathSizeBytesY + RoundedUpMaxSwathSizeBytesC <= DETBufferSizeInKByte * 1024 / 2) { |
| 6713 | SwathHeightY[k] = MinimumSwathHeightY; |
| 6714 | SwathHeightC[k] = MaximumSwathHeightC[k]; |
| 6715 | RoundedUpSwathSizeBytesY = RoundedUpMinSwathSizeBytesY; |
| 6716 | RoundedUpSwathSizeBytesC = RoundedUpMaxSwathSizeBytesC; |
| 6717 | } else if (RoundedUpMaxSwathSizeBytesY < 1.5 * RoundedUpMaxSwathSizeBytesC |
| 6718 | && RoundedUpMaxSwathSizeBytesY + RoundedUpMinSwathSizeBytesC <= DETBufferSizeInKByte * 1024 / 2) { |
| 6719 | SwathHeightY[k] = MaximumSwathHeightY[k]; |
| 6720 | SwathHeightC[k] = MinimumSwathHeightC; |
| 6721 | RoundedUpSwathSizeBytesY = RoundedUpMaxSwathSizeBytesY; |
| 6722 | RoundedUpSwathSizeBytesC = RoundedUpMinSwathSizeBytesC; |
| 6723 | } else { |
| 6724 | SwathHeightY[k] = MinimumSwathHeightY; |
| 6725 | SwathHeightC[k] = MinimumSwathHeightC; |
| 6726 | RoundedUpSwathSizeBytesY = RoundedUpMinSwathSizeBytesY; |
| 6727 | RoundedUpSwathSizeBytesC = RoundedUpMinSwathSizeBytesC; |
| 6728 | } |
| 6729 | { |
| 6730 | double actDETBufferSizeInKByte = dml_ceil(DETBufferSizeInKByte, 64); |
| 6731 | if (SwathHeightC[k] == 0) { |
| 6732 | DETBufferSizeY[k] = actDETBufferSizeInKByte * 1024; |
| 6733 | DETBufferSizeC[k] = 0; |
| 6734 | } else if (RoundedUpSwathSizeBytesY <= 1.5 * RoundedUpSwathSizeBytesC) { |
| 6735 | DETBufferSizeY[k] = actDETBufferSizeInKByte * 1024 / 2; |
| 6736 | DETBufferSizeC[k] = actDETBufferSizeInKByte * 1024 / 2; |
| 6737 | } else { |
| 6738 | DETBufferSizeY[k] = dml_floor(actDETBufferSizeInKByte * 1024 * 2 / 3, 1024); |
| 6739 | DETBufferSizeC[k] = actDETBufferSizeInKByte * 1024 / 3; |
| 6740 | } |
| 6741 | |
| 6742 | if (RoundedUpMinSwathSizeBytesY + RoundedUpMinSwathSizeBytesC > actDETBufferSizeInKByte * 1024 / 2 || SwathWidth[k] > MaximumSwathWidthLuma[k] |
| 6743 | || (SwathHeightC[k] > 0 && SwathWidthChroma[k] > MaximumSwathWidthChroma[k])) { |
| 6744 | *ViewportSizeSupport = false0; |
| 6745 | ViewportSizeSupportPerPlane[k] = false0; |
| 6746 | } else { |
| 6747 | ViewportSizeSupportPerPlane[k] = true1; |
| 6748 | } |
| 6749 | } |
| 6750 | } |
| 6751 | } |
| 6752 | |
| 6753 | static void CalculateSwathWidth( |
| 6754 | bool_Bool ForceSingleDPP, |
| 6755 | int NumberOfActivePlanes, |
| 6756 | enum source_format_class SourcePixelFormat[], |
| 6757 | enum scan_direction_class SourceScan[], |
| 6758 | int ViewportWidth[], |
| 6759 | int ViewportHeight[], |
| 6760 | int SurfaceWidthY[], |
| 6761 | int SurfaceWidthC[], |
| 6762 | int SurfaceHeightY[], |
| 6763 | int SurfaceHeightC[], |
| 6764 | enum odm_combine_mode ODMCombineEnabled[], |
| 6765 | int BytePerPixY[], |
| 6766 | int BytePerPixC[], |
| 6767 | int Read256BytesBlockHeightY[], |
| 6768 | int Read256BytesBlockHeightC[], |
| 6769 | int Read256BytesBlockWidthY[], |
| 6770 | int Read256BytesBlockWidthC[], |
| 6771 | int BlendingAndTiming[], |
| 6772 | int HActive[], |
| 6773 | double HRatio[], |
| 6774 | int DPPPerPlane[], |
| 6775 | double SwathWidthSingleDPPY[], |
| 6776 | double SwathWidthSingleDPPC[], |
| 6777 | double SwathWidthY[], |
| 6778 | double SwathWidthC[], |
| 6779 | int MaximumSwathHeightY[], |
| 6780 | int MaximumSwathHeightC[], |
| 6781 | int swath_width_luma_ub[], |
| 6782 | int swath_width_chroma_ub[]) |
| 6783 | { |
| 6784 | enum odm_combine_mode MainPlaneODMCombine; |
| 6785 | int j, k; |
| 6786 | |
| 6787 | #ifdef __DML_VBA_DEBUG__ |
| 6788 | dml_print("DML::%s: NumberOfActivePlanes = %d\n", __func__, NumberOfActivePlanes){do { } while(0); }; |
| 6789 | #endif |
| 6790 | |
| 6791 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 6792 | if (SourceScan[k] != dm_vert) { |
| 6793 | SwathWidthSingleDPPY[k] = ViewportWidth[k]; |
| 6794 | } else { |
| 6795 | SwathWidthSingleDPPY[k] = ViewportHeight[k]; |
| 6796 | } |
| 6797 | |
| 6798 | #ifdef __DML_VBA_DEBUG__ |
| 6799 | dml_print("DML::%s: k=%d ViewportWidth=%d\n", __func__, k, ViewportWidth[k]){do { } while(0); }; |
| 6800 | dml_print("DML::%s: k=%d ViewportHeight=%d\n", __func__, k, ViewportHeight[k]){do { } while(0); }; |
| 6801 | #endif |
| 6802 | |
| 6803 | MainPlaneODMCombine = ODMCombineEnabled[k]; |
| 6804 | for (j = 0; j < NumberOfActivePlanes; ++j) { |
| 6805 | if (BlendingAndTiming[k] == j) { |
| 6806 | MainPlaneODMCombine = ODMCombineEnabled[j]; |
| 6807 | } |
| 6808 | } |
| 6809 | |
| 6810 | if (MainPlaneODMCombine == dm_odm_combine_mode_4to1) { |
| 6811 | SwathWidthY[k] = dml_min(SwathWidthSingleDPPY[k], dml_round(HActive[k] / 4.0 * HRatio[k])); |
| 6812 | } else if (MainPlaneODMCombine == dm_odm_combine_mode_2to1) { |
| 6813 | SwathWidthY[k] = dml_min(SwathWidthSingleDPPY[k], dml_round(HActive[k] / 2.0 * HRatio[k])); |
| 6814 | } else if (DPPPerPlane[k] == 2) { |
| 6815 | SwathWidthY[k] = SwathWidthSingleDPPY[k] / 2; |
| 6816 | } else { |
| 6817 | SwathWidthY[k] = SwathWidthSingleDPPY[k]; |
| 6818 | } |
| 6819 | |
| 6820 | #ifdef __DML_VBA_DEBUG__ |
| 6821 | dml_print("DML::%s: k=%d SwathWidthSingleDPPY=%f\n", __func__, k, SwathWidthSingleDPPY[k]){do { } while(0); }; |
| 6822 | dml_print("DML::%s: k=%d SwathWidthY=%f\n", __func__, k, SwathWidthY[k]){do { } while(0); }; |
| 6823 | #endif |
| 6824 | |
| 6825 | if (SourcePixelFormat[k] == dm_420_8 || SourcePixelFormat[k] == dm_420_10 || SourcePixelFormat[k] == dm_420_12) { |
| 6826 | SwathWidthC[k] = SwathWidthY[k] / 2; |
| 6827 | SwathWidthSingleDPPC[k] = SwathWidthSingleDPPY[k] / 2; |
| 6828 | } else { |
| 6829 | SwathWidthC[k] = SwathWidthY[k]; |
| 6830 | SwathWidthSingleDPPC[k] = SwathWidthSingleDPPY[k]; |
| 6831 | } |
| 6832 | |
| 6833 | if (ForceSingleDPP == true1) { |
| 6834 | SwathWidthY[k] = SwathWidthSingleDPPY[k]; |
| 6835 | SwathWidthC[k] = SwathWidthSingleDPPC[k]; |
| 6836 | } |
| 6837 | { |
| 6838 | int surface_width_ub_l = dml_ceil(SurfaceWidthY[k], Read256BytesBlockWidthY[k]); |
| 6839 | int surface_height_ub_l = dml_ceil(SurfaceHeightY[k], Read256BytesBlockHeightY[k]); |
| 6840 | |
| 6841 | #ifdef __DML_VBA_DEBUG__ |
| 6842 | dml_print("DML::%s: k=%d surface_width_ub_l=%0d\n", __func__, k, surface_width_ub_l){do { } while(0); }; |
| 6843 | #endif |
| 6844 | |
| 6845 | if (SourceScan[k] != dm_vert) { |
| 6846 | MaximumSwathHeightY[k] = Read256BytesBlockHeightY[k]; |
| 6847 | MaximumSwathHeightC[k] = Read256BytesBlockHeightC[k]; |
| 6848 | swath_width_luma_ub[k] = dml_min(surface_width_ub_l, (int) dml_ceil(SwathWidthY[k] - 1, Read256BytesBlockWidthY[k]) + Read256BytesBlockWidthY[k]); |
| 6849 | if (BytePerPixC[k] > 0) { |
| 6850 | int surface_width_ub_c = dml_ceil(SurfaceWidthC[k], Read256BytesBlockWidthC[k]); |
| 6851 | |
| 6852 | swath_width_chroma_ub[k] = dml_min( |
| 6853 | surface_width_ub_c, |
| 6854 | (int) dml_ceil(SwathWidthC[k] - 1, Read256BytesBlockWidthC[k]) + Read256BytesBlockWidthC[k]); |
| 6855 | } else { |
| 6856 | swath_width_chroma_ub[k] = 0; |
| 6857 | } |
| 6858 | } else { |
| 6859 | MaximumSwathHeightY[k] = Read256BytesBlockWidthY[k]; |
| 6860 | MaximumSwathHeightC[k] = Read256BytesBlockWidthC[k]; |
| 6861 | swath_width_luma_ub[k] = dml_min(surface_height_ub_l, (int) dml_ceil(SwathWidthY[k] - 1, Read256BytesBlockHeightY[k]) + Read256BytesBlockHeightY[k]); |
| 6862 | if (BytePerPixC[k] > 0) { |
| 6863 | int surface_height_ub_c = dml_ceil(SurfaceHeightC[k], Read256BytesBlockHeightC[k]); |
| 6864 | |
| 6865 | swath_width_chroma_ub[k] = dml_min( |
| 6866 | surface_height_ub_c, |
| 6867 | (int) dml_ceil(SwathWidthC[k] - 1, Read256BytesBlockHeightC[k]) + Read256BytesBlockHeightC[k]); |
| 6868 | } else { |
| 6869 | swath_width_chroma_ub[k] = 0; |
| 6870 | } |
| 6871 | } |
| 6872 | } |
| 6873 | } |
| 6874 | } |
| 6875 | |
| 6876 | static double CalculateExtraLatency( |
| 6877 | int RoundTripPingLatencyCycles, |
| 6878 | int ReorderingBytes, |
| 6879 | double DCFCLK, |
| 6880 | int TotalNumberOfActiveDPP, |
| 6881 | int PixelChunkSizeInKByte, |
| 6882 | int TotalNumberOfDCCActiveDPP, |
| 6883 | int MetaChunkSize, |
| 6884 | double ReturnBW, |
| 6885 | bool_Bool GPUVMEnable, |
| 6886 | bool_Bool HostVMEnable, |
| 6887 | int NumberOfActivePlanes, |
| 6888 | int NumberOfDPP[], |
| 6889 | int dpte_group_bytes[], |
| 6890 | double HostVMInefficiencyFactor, |
| 6891 | double HostVMMinPageSize, |
| 6892 | int HostVMMaxNonCachedPageTableLevels) |
| 6893 | { |
| 6894 | double ExtraLatencyBytes; |
| 6895 | double ExtraLatency; |
| 6896 | |
| 6897 | ExtraLatencyBytes = CalculateExtraLatencyBytes( |
| 6898 | ReorderingBytes, |
| 6899 | TotalNumberOfActiveDPP, |
| 6900 | PixelChunkSizeInKByte, |
| 6901 | TotalNumberOfDCCActiveDPP, |
| 6902 | MetaChunkSize, |
| 6903 | GPUVMEnable, |
| 6904 | HostVMEnable, |
| 6905 | NumberOfActivePlanes, |
| 6906 | NumberOfDPP, |
| 6907 | dpte_group_bytes, |
| 6908 | HostVMInefficiencyFactor, |
| 6909 | HostVMMinPageSize, |
| 6910 | HostVMMaxNonCachedPageTableLevels); |
| 6911 | |
| 6912 | ExtraLatency = (RoundTripPingLatencyCycles + __DML_ARB_TO_RET_DELAY__(7 + 95)) / DCFCLK + ExtraLatencyBytes / ReturnBW; |
| 6913 | |
| 6914 | #ifdef __DML_VBA_DEBUG__ |
| 6915 | dml_print("DML::%s: RoundTripPingLatencyCycles=%d\n", __func__, RoundTripPingLatencyCycles){do { } while(0); }; |
| 6916 | dml_print("DML::%s: DCFCLK=%f\n", __func__, DCFCLK){do { } while(0); }; |
| 6917 | dml_print("DML::%s: ExtraLatencyBytes=%f\n", __func__, ExtraLatencyBytes){do { } while(0); }; |
| 6918 | dml_print("DML::%s: ReturnBW=%f\n", __func__, ReturnBW){do { } while(0); }; |
| 6919 | dml_print("DML::%s: ExtraLatency=%f\n", __func__, ExtraLatency){do { } while(0); }; |
| 6920 | #endif |
| 6921 | |
| 6922 | return ExtraLatency; |
| 6923 | } |
| 6924 | |
| 6925 | static double CalculateExtraLatencyBytes( |
| 6926 | int ReorderingBytes, |
| 6927 | int TotalNumberOfActiveDPP, |
| 6928 | int PixelChunkSizeInKByte, |
| 6929 | int TotalNumberOfDCCActiveDPP, |
| 6930 | int MetaChunkSize, |
| 6931 | bool_Bool GPUVMEnable, |
| 6932 | bool_Bool HostVMEnable, |
| 6933 | int NumberOfActivePlanes, |
| 6934 | int NumberOfDPP[], |
| 6935 | int dpte_group_bytes[], |
| 6936 | double HostVMInefficiencyFactor, |
| 6937 | double HostVMMinPageSize, |
| 6938 | int HostVMMaxNonCachedPageTableLevels) |
| 6939 | { |
| 6940 | double ret; |
| 6941 | int HostVMDynamicLevels = 0, k; |
| 6942 | |
| 6943 | if (GPUVMEnable == true1 && HostVMEnable == true1) { |
| 6944 | if (HostVMMinPageSize < 2048) { |
| 6945 | HostVMDynamicLevels = HostVMMaxNonCachedPageTableLevels; |
| 6946 | } else if (HostVMMinPageSize >= 2048 && HostVMMinPageSize < 1048576) { |
| 6947 | HostVMDynamicLevels = dml_max(0, (int) HostVMMaxNonCachedPageTableLevels - 1); |
| 6948 | } else { |
| 6949 | HostVMDynamicLevels = dml_max(0, (int) HostVMMaxNonCachedPageTableLevels - 2); |
| 6950 | } |
| 6951 | } else { |
| 6952 | HostVMDynamicLevels = 0; |
| 6953 | } |
| 6954 | |
| 6955 | ret = ReorderingBytes + (TotalNumberOfActiveDPP * PixelChunkSizeInKByte + TotalNumberOfDCCActiveDPP * MetaChunkSize) * 1024.0; |
| 6956 | |
| 6957 | if (GPUVMEnable == true1) { |
| 6958 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
| 6959 | ret = ret + NumberOfDPP[k] * dpte_group_bytes[k] * (1 + 8 * HostVMDynamicLevels) * HostVMInefficiencyFactor; |
| 6960 | } |
| 6961 | } |
| 6962 | return ret; |
| 6963 | } |
| 6964 | |
| 6965 | static double CalculateUrgentLatency( |
| 6966 | double UrgentLatencyPixelDataOnly, |
| 6967 | double UrgentLatencyPixelMixedWithVMData, |
| 6968 | double UrgentLatencyVMDataOnly, |
| 6969 | bool_Bool DoUrgentLatencyAdjustment, |
| 6970 | double UrgentLatencyAdjustmentFabricClockComponent, |
| 6971 | double UrgentLatencyAdjustmentFabricClockReference, |
| 6972 | double FabricClock) |
| 6973 | { |
| 6974 | double ret; |
| 6975 | |
| 6976 | ret = dml_max3(UrgentLatencyPixelDataOnly, UrgentLatencyPixelMixedWithVMData, UrgentLatencyVMDataOnly); |
| 6977 | if (DoUrgentLatencyAdjustment == true1) { |
| 6978 | ret = ret + UrgentLatencyAdjustmentFabricClockComponent * (UrgentLatencyAdjustmentFabricClockReference / FabricClock - 1); |
| 6979 | } |
| 6980 | return ret; |
| 6981 | } |
| 6982 | |
| 6983 | static void UseMinimumDCFCLK( |
| 6984 | struct display_mode_lib *mode_lib, |
| 6985 | int MaxPrefetchMode, |
| 6986 | int ReorderingBytes) |
| 6987 | { |
| 6988 | struct vba_vars_st *v = &mode_lib->vba; |
| 6989 | int dummy1, i, j, k; |
| 6990 | double NormalEfficiency, dummy2, dummy3; |
| 6991 | double TotalMaxPrefetchFlipDPTERowBandwidth[DC__VOLTAGE_STATES20][2]; |
| 6992 | |
| 6993 | NormalEfficiency = v->PercentOfIdealFabricAndSDPPortBWReceivedAfterUrgLatency / 100.0; |
| 6994 | for (i = 0; i < v->soc.num_states; ++i) { |
| 6995 | for (j = 0; j <= 1; ++j) { |
| 6996 | double PixelDCFCLKCyclesRequiredInPrefetch[DC__NUM_DPP__MAX8]; |
| 6997 | double PrefetchPixelLinesTime[DC__NUM_DPP__MAX8]; |
| 6998 | double DCFCLKRequiredForPeakBandwidthPerPlane[DC__NUM_DPP__MAX8]; |
| 6999 | double DynamicMetadataVMExtraLatency[DC__NUM_DPP__MAX8]; |
| 7000 | double MinimumTWait; |
| 7001 | double NonDPTEBandwidth; |
| 7002 | double DPTEBandwidth; |
| 7003 | double DCFCLKRequiredForAverageBandwidth; |
| 7004 | double ExtraLatencyBytes; |
| 7005 | double ExtraLatencyCycles; |
| 7006 | double DCFCLKRequiredForPeakBandwidth; |
| 7007 | int NoOfDPPState[DC__NUM_DPP__MAX8]; |
| 7008 | double MinimumTvmPlus2Tr0; |
| 7009 | |
| 7010 | TotalMaxPrefetchFlipDPTERowBandwidth[i][j] = 0; |
| 7011 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 7012 | TotalMaxPrefetchFlipDPTERowBandwidth[i][j] = TotalMaxPrefetchFlipDPTERowBandwidth[i][j] |
| 7013 | + v->NoOfDPP[i][j][k] * v->DPTEBytesPerRow[i][j][k] / (15.75 * v->HTotal[k] / v->PixelClock[k]); |
| 7014 | } |
| 7015 | |
| 7016 | for (k = 0; k <= v->NumberOfActivePlanes - 1; ++k) { |
| 7017 | NoOfDPPState[k] = v->NoOfDPP[i][j][k]; |
| 7018 | } |
| 7019 | |
| 7020 | MinimumTWait = CalculateTWait(MaxPrefetchMode, v->FinalDRAMClockChangeLatency, v->UrgLatency[i], v->SREnterPlusExitTime); |
| 7021 | NonDPTEBandwidth = v->TotalVActivePixelBandwidth[i][j] + v->TotalVActiveCursorBandwidth[i][j] + v->TotalMetaRowBandwidth[i][j]; |
| 7022 | DPTEBandwidth = (v->HostVMEnable == true1 || v->ImmediateFlipRequirement[0] == dm_immediate_flip_required) ? |
| 7023 | TotalMaxPrefetchFlipDPTERowBandwidth[i][j] : v->TotalDPTERowBandwidth[i][j]; |
| 7024 | DCFCLKRequiredForAverageBandwidth = dml_max3( |
| 7025 | v->ProjectedDCFCLKDeepSleep[i][j], |
| 7026 | (NonDPTEBandwidth + v->TotalDPTERowBandwidth[i][j]) / v->ReturnBusWidth |
| 7027 | / (v->MaxAveragePercentOfIdealFabricAndSDPPortBWDisplayCanUseInNormalSystemOperation / 100), |
| 7028 | (NonDPTEBandwidth + DPTEBandwidth / NormalEfficiency) / NormalEfficiency / v->ReturnBusWidth); |
| 7029 | |
| 7030 | ExtraLatencyBytes = CalculateExtraLatencyBytes( |
| 7031 | ReorderingBytes, |
| 7032 | v->TotalNumberOfActiveDPP[i][j], |
| 7033 | v->PixelChunkSizeInKByte, |
| 7034 | v->TotalNumberOfDCCActiveDPP[i][j], |
| 7035 | v->MetaChunkSize, |
| 7036 | v->GPUVMEnable, |
| 7037 | v->HostVMEnable, |
| 7038 | v->NumberOfActivePlanes, |
| 7039 | NoOfDPPState, |
| 7040 | v->dpte_group_bytes, |
| 7041 | 1, |
| 7042 | v->HostVMMinPageSize, |
| 7043 | v->HostVMMaxNonCachedPageTableLevels); |
| 7044 | ExtraLatencyCycles = v->RoundTripPingLatencyCycles + __DML_ARB_TO_RET_DELAY__(7 + 95) + ExtraLatencyBytes / NormalEfficiency / v->ReturnBusWidth; |
| 7045 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 7046 | double DCFCLKCyclesRequiredInPrefetch; |
| 7047 | double ExpectedPrefetchBWAcceleration; |
| 7048 | double PrefetchTime; |
| 7049 | |
| 7050 | PixelDCFCLKCyclesRequiredInPrefetch[k] = (v->PrefetchLinesY[i][j][k] * v->swath_width_luma_ub_all_states[i][j][k] * v->BytePerPixelY[k] |
| 7051 | + v->PrefetchLinesC[i][j][k] * v->swath_width_chroma_ub_all_states[i][j][k] * v->BytePerPixelC[k]) / NormalEfficiency / v->ReturnBusWidth; |
| 7052 | DCFCLKCyclesRequiredInPrefetch = 2 * ExtraLatencyCycles / NoOfDPPState[k] |
| 7053 | + v->PDEAndMetaPTEBytesPerFrame[i][j][k] / NormalEfficiency / NormalEfficiency / v->ReturnBusWidth * (v->GPUVMMaxPageTableLevels > 2 ? 1 : 0) |
| 7054 | + 2 * v->DPTEBytesPerRow[i][j][k] / NormalEfficiency / NormalEfficiency / v->ReturnBusWidth |
| 7055 | + 2 * v->MetaRowBytes[i][j][k] / NormalEfficiency / v->ReturnBusWidth + PixelDCFCLKCyclesRequiredInPrefetch[k]; |
| 7056 | PrefetchPixelLinesTime[k] = dml_max(v->PrefetchLinesY[i][j][k], v->PrefetchLinesC[i][j][k]) * v->HTotal[k] / v->PixelClock[k]; |
| 7057 | ExpectedPrefetchBWAcceleration = (v->VActivePixelBandwidth[i][j][k] + v->VActiveCursorBandwidth[i][j][k]) |
| 7058 | / (v->ReadBandwidthLuma[k] + v->ReadBandwidthChroma[k]); |
| 7059 | DynamicMetadataVMExtraLatency[k] = |
| 7060 | (v->GPUVMEnable == true1 && v->DynamicMetadataEnable[k] == true1 && v->DynamicMetadataVMEnabled == true1) ? |
| 7061 | v->UrgLatency[i] * v->GPUVMMaxPageTableLevels * (v->HostVMEnable == true1 ? v->HostVMMaxNonCachedPageTableLevels + 1 : 1) : 0; |
| 7062 | PrefetchTime = (v->MaximumVStartup[i][j][k] - 1) * v->HTotal[k] / v->PixelClock[k] - MinimumTWait |
| 7063 | - v->UrgLatency[i] |
| 7064 | * ((v->GPUVMMaxPageTableLevels <= 2 ? v->GPUVMMaxPageTableLevels : v->GPUVMMaxPageTableLevels - 2) |
| 7065 | * (v->HostVMEnable == true1 ? v->HostVMMaxNonCachedPageTableLevels + 1 : 1) - 1) |
| 7066 | - DynamicMetadataVMExtraLatency[k]; |
| 7067 | |
| 7068 | if (PrefetchTime > 0) { |
| 7069 | double ExpectedVRatioPrefetch; |
| 7070 | ExpectedVRatioPrefetch = PrefetchPixelLinesTime[k] |
| 7071 | / (PrefetchTime * PixelDCFCLKCyclesRequiredInPrefetch[k] / DCFCLKCyclesRequiredInPrefetch); |
| 7072 | DCFCLKRequiredForPeakBandwidthPerPlane[k] = NoOfDPPState[k] * PixelDCFCLKCyclesRequiredInPrefetch[k] / PrefetchPixelLinesTime[k] |
| 7073 | * dml_max(1.0, ExpectedVRatioPrefetch) * dml_max(1.0, ExpectedVRatioPrefetch / 4) * ExpectedPrefetchBWAcceleration; |
| 7074 | if (v->HostVMEnable == true1 || v->ImmediateFlipRequirement[0] == dm_immediate_flip_required) { |
| 7075 | DCFCLKRequiredForPeakBandwidthPerPlane[k] = DCFCLKRequiredForPeakBandwidthPerPlane[k] |
| 7076 | + NoOfDPPState[k] * DPTEBandwidth / NormalEfficiency / NormalEfficiency / v->ReturnBusWidth; |
| 7077 | } |
| 7078 | } else { |
| 7079 | DCFCLKRequiredForPeakBandwidthPerPlane[k] = v->DCFCLKPerState[i]; |
| 7080 | } |
| 7081 | if (v->DynamicMetadataEnable[k] == true1) { |
| 7082 | double TSetupPipe; |
| 7083 | double TdmbfPipe; |
| 7084 | double TdmsksPipe; |
| 7085 | double TdmecPipe; |
| 7086 | double AllowedTimeForUrgentExtraLatency; |
| 7087 | |
| 7088 | CalculateVupdateAndDynamicMetadataParameters( |
| 7089 | v->MaxInterDCNTileRepeaters, |
| 7090 | v->RequiredDPPCLK[i][j][k], |
| 7091 | v->RequiredDISPCLK[i][j], |
| 7092 | v->ProjectedDCFCLKDeepSleep[i][j], |
| 7093 | v->PixelClock[k], |
| 7094 | v->HTotal[k], |
| 7095 | v->VTotal[k] - v->VActive[k], |
| 7096 | v->DynamicMetadataTransmittedBytes[k], |
| 7097 | v->DynamicMetadataLinesBeforeActiveRequired[k], |
| 7098 | v->Interlace[k], |
| 7099 | v->ProgressiveToInterlaceUnitInOPP, |
| 7100 | &TSetupPipe, |
| 7101 | &TdmbfPipe, |
| 7102 | &TdmecPipe, |
| 7103 | &TdmsksPipe, |
| 7104 | &dummy1, |
| 7105 | &dummy2, |
| 7106 | &dummy3); |
| 7107 | AllowedTimeForUrgentExtraLatency = v->MaximumVStartup[i][j][k] * v->HTotal[k] / v->PixelClock[k] - MinimumTWait - TSetupPipe - TdmbfPipe - TdmecPipe |
| 7108 | - TdmsksPipe - DynamicMetadataVMExtraLatency[k]; |
| 7109 | if (AllowedTimeForUrgentExtraLatency > 0) { |
| 7110 | DCFCLKRequiredForPeakBandwidthPerPlane[k] = dml_max( |
| 7111 | DCFCLKRequiredForPeakBandwidthPerPlane[k], |
| 7112 | ExtraLatencyCycles / AllowedTimeForUrgentExtraLatency); |
| 7113 | } else { |
| 7114 | DCFCLKRequiredForPeakBandwidthPerPlane[k] = v->DCFCLKPerState[i]; |
| 7115 | } |
| 7116 | } |
| 7117 | } |
| 7118 | DCFCLKRequiredForPeakBandwidth = 0; |
| 7119 | for (k = 0; k <= v->NumberOfActivePlanes - 1; ++k) { |
| 7120 | DCFCLKRequiredForPeakBandwidth = DCFCLKRequiredForPeakBandwidth + DCFCLKRequiredForPeakBandwidthPerPlane[k]; |
| 7121 | } |
| 7122 | MinimumTvmPlus2Tr0 = v->UrgLatency[i] |
| 7123 | * (v->GPUVMEnable == true1 ? |
| 7124 | (v->HostVMEnable == true1 ? |
| 7125 | (v->GPUVMMaxPageTableLevels + 2) * (v->HostVMMaxNonCachedPageTableLevels + 1) - 1 : v->GPUVMMaxPageTableLevels + 1) : |
| 7126 | 0); |
| 7127 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
| 7128 | double MaximumTvmPlus2Tr0PlusTsw; |
| 7129 | MaximumTvmPlus2Tr0PlusTsw = (v->MaximumVStartup[i][j][k] - 2) * v->HTotal[k] / v->PixelClock[k] - MinimumTWait - DynamicMetadataVMExtraLatency[k]; |
| 7130 | if (MaximumTvmPlus2Tr0PlusTsw <= MinimumTvmPlus2Tr0 + PrefetchPixelLinesTime[k] / 4) { |
| 7131 | DCFCLKRequiredForPeakBandwidth = v->DCFCLKPerState[i]; |
| 7132 | } else { |
| 7133 | DCFCLKRequiredForPeakBandwidth = dml_max3( |
| 7134 | DCFCLKRequiredForPeakBandwidth, |
| 7135 | 2 * ExtraLatencyCycles / (MaximumTvmPlus2Tr0PlusTsw - MinimumTvmPlus2Tr0 - PrefetchPixelLinesTime[k] / 4), |
| 7136 | (2 * ExtraLatencyCycles + PixelDCFCLKCyclesRequiredInPrefetch[k]) / (MaximumTvmPlus2Tr0PlusTsw - MinimumTvmPlus2Tr0)); |
| 7137 | } |
| 7138 | } |
| 7139 | v->DCFCLKState[i][j] = dml_min(v->DCFCLKPerState[i], 1.05 * dml_max(DCFCLKRequiredForAverageBandwidth, DCFCLKRequiredForPeakBandwidth)); |
| 7140 | } |
| 7141 | } |
| 7142 | } |
| 7143 | |
| 7144 | static void CalculateUnboundedRequestAndCompressedBufferSize( |
| 7145 | unsigned int DETBufferSizeInKByte, |
| 7146 | int ConfigReturnBufferSizeInKByte, |
| 7147 | enum unbounded_requesting_policy UseUnboundedRequestingFinal, |
| 7148 | int TotalActiveDPP, |
| 7149 | bool_Bool NoChromaPlanes, |
| 7150 | int MaxNumDPP, |
| 7151 | int CompressedBufferSegmentSizeInkByteFinal, |
| 7152 | enum output_encoder_class *Output, |
| 7153 | bool_Bool *UnboundedRequestEnabled, |
| 7154 | int *CompressedBufferSizeInkByte) |
| 7155 | { |
| 7156 | double actDETBufferSizeInKByte = dml_ceil(DETBufferSizeInKByte, 64); |
| 7157 | |
| 7158 | *UnboundedRequestEnabled = UnboundedRequest(UseUnboundedRequestingFinal, TotalActiveDPP, NoChromaPlanes, Output[0]); |
| 7159 | *CompressedBufferSizeInkByte = ( |
| 7160 | *UnboundedRequestEnabled == true1 ? |
| 7161 | ConfigReturnBufferSizeInKByte - TotalActiveDPP * actDETBufferSizeInKByte : |
| 7162 | ConfigReturnBufferSizeInKByte - MaxNumDPP * actDETBufferSizeInKByte); |
| 7163 | *CompressedBufferSizeInkByte = *CompressedBufferSizeInkByte * CompressedBufferSegmentSizeInkByteFinal / 64; |
| 7164 | |
| 7165 | #ifdef __DML_VBA_DEBUG__ |
| 7166 | dml_print("DML::%s: TotalActiveDPP = %d\n", __func__, TotalActiveDPP){do { } while(0); }; |
| 7167 | dml_print("DML::%s: DETBufferSizeInKByte = %d\n", __func__, DETBufferSizeInKByte){do { } while(0); }; |
| 7168 | dml_print("DML::%s: ConfigReturnBufferSizeInKByte = %d\n", __func__, ConfigReturnBufferSizeInKByte){do { } while(0); }; |
| 7169 | dml_print("DML::%s: UseUnboundedRequestingFinal = %d\n", __func__, UseUnboundedRequestingFinal){do { } while(0); }; |
| 7170 | dml_print("DML::%s: actDETBufferSizeInKByte = %f\n", __func__, actDETBufferSizeInKByte){do { } while(0); }; |
| 7171 | dml_print("DML::%s: UnboundedRequestEnabled = %d\n", __func__, *UnboundedRequestEnabled){do { } while(0); }; |
| 7172 | dml_print("DML::%s: CompressedBufferSizeInkByte = %d\n", __func__, *CompressedBufferSizeInkByte){do { } while(0); }; |
| 7173 | #endif |
| 7174 | } |
| 7175 | |
| 7176 | static bool_Bool UnboundedRequest(enum unbounded_requesting_policy UseUnboundedRequestingFinal, int TotalNumberOfActiveDPP, bool_Bool NoChroma, enum output_encoder_class Output) |
| 7177 | { |
| 7178 | bool_Bool ret_val = false0; |
| 7179 | |
| 7180 | ret_val = (UseUnboundedRequestingFinal != dm_unbounded_requesting_disable && TotalNumberOfActiveDPP == 1 && NoChroma); |
| 7181 | if (UseUnboundedRequestingFinal == dm_unbounded_requesting_edp_only && Output != dm_edp) { |
| 7182 | ret_val = false0; |
| 7183 | } |
| 7184 | return (ret_val); |
| 7185 | } |
| 7186 |