File: | dev/pci/drm/amd/display/dc/dml/dcn30/display_mode_vba_30.c |
Warning: | line 1288, column 3 Value stored to 'TimeForFetchingMetaPTE' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | /* |
2 | * Copyright 2020 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 <generated/autoconf.h> |
27 | |
28 | #ifdef CONFIG_DRM_AMD_DC_DCN1 |
29 | #include "dc.h" |
30 | #include "dc_link.h" |
31 | #include "../display_mode_lib.h" |
32 | #include "display_mode_vba_30.h" |
33 | #include "../dml_inline_defs.h" |
34 | |
35 | |
36 | /* |
37 | * NOTE: |
38 | * This file is gcc-parsable HW gospel, coming straight from HW engineers. |
39 | * |
40 | * It doesn't adhere to Linux kernel style and sometimes will do things in odd |
41 | * ways. Unless there is something clearly wrong with it the code should |
42 | * remain as-is as it provides us with a guarantee from HW that it is correct. |
43 | */ |
44 | |
45 | |
46 | typedef struct { |
47 | double DPPCLK; |
48 | double DISPCLK; |
49 | double PixelClock; |
50 | double DCFCLKDeepSleep; |
51 | unsigned int DPPPerPlane; |
52 | bool_Bool ScalerEnabled; |
53 | enum scan_direction_class SourceScan; |
54 | unsigned int BlockWidth256BytesY; |
55 | unsigned int BlockHeight256BytesY; |
56 | unsigned int BlockWidth256BytesC; |
57 | unsigned int BlockHeight256BytesC; |
58 | unsigned int InterlaceEnable; |
59 | unsigned int NumberOfCursors; |
60 | unsigned int VBlank; |
61 | unsigned int HTotal; |
62 | unsigned int DCCEnable; |
63 | bool_Bool ODMCombineEnabled; |
64 | } Pipe; |
65 | |
66 | #define BPP_INVALID0 0 |
67 | #define BPP_BLENDED_PIPE0xffffffff 0xffffffff |
68 | #define DCN30_MAX_DSC_IMAGE_WIDTH5184 5184 |
69 | #define DCN30_MAX_FMT_420_BUFFER_WIDTH4096 4096 |
70 | |
71 | static void DisplayPipeConfiguration(struct display_mode_lib *mode_lib); |
72 | static void DISPCLKDPPCLKDCFCLKDeepSleepPrefetchParametersWatermarksAndPerformanceCalculation( |
73 | struct display_mode_lib *mode_lib); |
74 | static unsigned int dscceComputeDelay( |
75 | unsigned int bpc, |
76 | double BPP, |
77 | unsigned int sliceWidth, |
78 | unsigned int numSlices, |
79 | enum output_format_class pixelFormat, |
80 | enum output_encoder_class Output); |
81 | static unsigned int dscComputeDelay( |
82 | enum output_format_class pixelFormat, |
83 | enum output_encoder_class Output); |
84 | // Super monster function with some 45 argument |
85 | static bool_Bool CalculatePrefetchSchedule( |
86 | struct display_mode_lib *mode_lib, |
87 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
88 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
89 | Pipe *myPipe, |
90 | unsigned int DSCDelay, |
91 | double DPPCLKDelaySubtotalPlusCNVCFormater, |
92 | double DPPCLKDelaySCL, |
93 | double DPPCLKDelaySCLLBOnly, |
94 | double DPPCLKDelayCNVCCursor, |
95 | double DISPCLKDelaySubtotal, |
96 | unsigned int DPP_RECOUT_WIDTH, |
97 | enum output_format_class OutputFormat, |
98 | unsigned int MaxInterDCNTileRepeaters, |
99 | unsigned int VStartup, |
100 | unsigned int MaxVStartup, |
101 | unsigned int GPUVMPageTableLevels, |
102 | bool_Bool GPUVMEnable, |
103 | bool_Bool HostVMEnable, |
104 | unsigned int HostVMMaxNonCachedPageTableLevels, |
105 | double HostVMMinPageSize, |
106 | bool_Bool DynamicMetadataEnable, |
107 | bool_Bool DynamicMetadataVMEnabled, |
108 | int DynamicMetadataLinesBeforeActiveRequired, |
109 | unsigned int DynamicMetadataTransmittedBytes, |
110 | double UrgentLatency, |
111 | double UrgentExtraLatency, |
112 | double TCalc, |
113 | unsigned int PDEAndMetaPTEBytesFrame, |
114 | unsigned int MetaRowByte, |
115 | unsigned int PixelPTEBytesPerRow, |
116 | double PrefetchSourceLinesY, |
117 | unsigned int SwathWidthY, |
118 | int BytePerPixelY, |
119 | double VInitPreFillY, |
120 | unsigned int MaxNumSwathY, |
121 | double PrefetchSourceLinesC, |
122 | unsigned int SwathWidthC, |
123 | int BytePerPixelC, |
124 | double VInitPreFillC, |
125 | unsigned int MaxNumSwathC, |
126 | long swath_width_luma_ub, |
127 | long swath_width_chroma_ub, |
128 | unsigned int SwathHeightY, |
129 | unsigned int SwathHeightC, |
130 | double TWait, |
131 | bool_Bool ProgressiveToInterlaceUnitInOPP, |
132 | double *DSTXAfterScaler, |
133 | double *DSTYAfterScaler, |
134 | double *DestinationLinesForPrefetch, |
135 | double *PrefetchBandwidth, |
136 | double *DestinationLinesToRequestVMInVBlank, |
137 | double *DestinationLinesToRequestRowInVBlank, |
138 | double *VRatioPrefetchY, |
139 | double *VRatioPrefetchC, |
140 | double *RequiredPrefetchPixDataBWLuma, |
141 | double *RequiredPrefetchPixDataBWChroma, |
142 | bool_Bool *NotEnoughTimeForDynamicMetadata, |
143 | double *Tno_bw, |
144 | double *prefetch_vmrow_bw, |
145 | double *Tdmdl_vm, |
146 | double *Tdmdl, |
147 | unsigned int *VUpdateOffsetPix, |
148 | double *VUpdateWidthPix, |
149 | double *VReadyOffsetPix); |
150 | static double RoundToDFSGranularityUp(double Clock, double VCOSpeed); |
151 | static double RoundToDFSGranularityDown(double Clock, double VCOSpeed); |
152 | static void CalculateDCCConfiguration( |
153 | bool_Bool DCCEnabled, |
154 | bool_Bool DCCProgrammingAssumesScanDirectionUnknown, |
155 | enum source_format_class SourcePixelFormat, |
156 | unsigned int ViewportWidthLuma, |
157 | unsigned int ViewportWidthChroma, |
158 | unsigned int ViewportHeightLuma, |
159 | unsigned int ViewportHeightChroma, |
160 | double DETBufferSize, |
161 | unsigned int RequestHeight256ByteLuma, |
162 | unsigned int RequestHeight256ByteChroma, |
163 | enum dm_swizzle_mode TilingFormat, |
164 | unsigned int BytePerPixelY, |
165 | unsigned int BytePerPixelC, |
166 | double BytePerPixelDETY, |
167 | double BytePerPixelDETC, |
168 | enum scan_direction_class ScanOrientation, |
169 | unsigned int *MaxUncompressedBlockLuma, |
170 | unsigned int *MaxUncompressedBlockChroma, |
171 | unsigned int *MaxCompressedBlockLuma, |
172 | unsigned int *MaxCompressedBlockChroma, |
173 | unsigned int *IndependentBlockLuma, |
174 | unsigned int *IndependentBlockChroma); |
175 | static double CalculatePrefetchSourceLines( |
176 | struct display_mode_lib *mode_lib, |
177 | double VRatio, |
178 | double vtaps, |
179 | bool_Bool Interlace, |
180 | bool_Bool ProgressiveToInterlaceUnitInOPP, |
181 | unsigned int SwathHeight, |
182 | unsigned int ViewportYStart, |
183 | double *VInitPreFill, |
184 | unsigned int *MaxNumSwath); |
185 | static unsigned int CalculateVMAndRowBytes( |
186 | struct display_mode_lib *mode_lib, |
187 | bool_Bool DCCEnable, |
188 | unsigned int BlockHeight256Bytes, |
189 | unsigned int BlockWidth256Bytes, |
190 | enum source_format_class SourcePixelFormat, |
191 | unsigned int SurfaceTiling, |
192 | unsigned int BytePerPixel, |
193 | enum scan_direction_class ScanDirection, |
194 | unsigned int SwathWidth, |
195 | unsigned int ViewportHeight, |
196 | bool_Bool GPUVMEnable, |
197 | bool_Bool HostVMEnable, |
198 | unsigned int HostVMMaxNonCachedPageTableLevels, |
199 | unsigned int GPUVMMinPageSize, |
200 | unsigned int HostVMMinPageSize, |
201 | unsigned int PTEBufferSizeInRequests, |
202 | unsigned int Pitch, |
203 | unsigned int DCCMetaPitch, |
204 | unsigned int *MacroTileWidth, |
205 | unsigned int *MetaRowByte, |
206 | unsigned int *PixelPTEBytesPerRow, |
207 | bool_Bool *PTEBufferSizeNotExceeded, |
208 | unsigned int *dpte_row_width_ub, |
209 | unsigned int *dpte_row_height, |
210 | unsigned int *MetaRequestWidth, |
211 | unsigned int *MetaRequestHeight, |
212 | unsigned int *meta_row_width, |
213 | unsigned int *meta_row_height, |
214 | unsigned int *vm_group_bytes, |
215 | unsigned int *dpte_group_bytes, |
216 | unsigned int *PixelPTEReqWidth, |
217 | unsigned int *PixelPTEReqHeight, |
218 | unsigned int *PTERequestSize, |
219 | unsigned int *DPDE0BytesFrame, |
220 | unsigned int *MetaPTEBytesFrame); |
221 | static double CalculateTWait( |
222 | unsigned int PrefetchMode, |
223 | double DRAMClockChangeLatency, |
224 | double UrgentLatency, |
225 | double SREnterPlusExitTime); |
226 | static void CalculateRowBandwidth( |
227 | bool_Bool GPUVMEnable, |
228 | enum source_format_class SourcePixelFormat, |
229 | double VRatio, |
230 | double VRatioChroma, |
231 | bool_Bool DCCEnable, |
232 | double LineTime, |
233 | unsigned int MetaRowByteLuma, |
234 | unsigned int MetaRowByteChroma, |
235 | unsigned int meta_row_height_luma, |
236 | unsigned int meta_row_height_chroma, |
237 | unsigned int PixelPTEBytesPerRowLuma, |
238 | unsigned int PixelPTEBytesPerRowChroma, |
239 | unsigned int dpte_row_height_luma, |
240 | unsigned int dpte_row_height_chroma, |
241 | double *meta_row_bw, |
242 | double *dpte_row_bw); |
243 | static void CalculateFlipSchedule( |
244 | struct display_mode_lib *mode_lib, |
245 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
246 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
247 | double UrgentExtraLatency, |
248 | double UrgentLatency, |
249 | unsigned int GPUVMMaxPageTableLevels, |
250 | bool_Bool HostVMEnable, |
251 | unsigned int HostVMMaxNonCachedPageTableLevels, |
252 | bool_Bool GPUVMEnable, |
253 | double HostVMMinPageSize, |
254 | double PDEAndMetaPTEBytesPerFrame, |
255 | double MetaRowBytes, |
256 | double DPTEBytesPerRow, |
257 | double BandwidthAvailableForImmediateFlip, |
258 | unsigned int TotImmediateFlipBytes, |
259 | enum source_format_class SourcePixelFormat, |
260 | double LineTime, |
261 | double VRatio, |
262 | double VRatioChroma, |
263 | double Tno_bw, |
264 | bool_Bool DCCEnable, |
265 | unsigned int dpte_row_height, |
266 | unsigned int meta_row_height, |
267 | unsigned int dpte_row_height_chroma, |
268 | unsigned int meta_row_height_chroma, |
269 | double *DestinationLinesToRequestVMInImmediateFlip, |
270 | double *DestinationLinesToRequestRowInImmediateFlip, |
271 | double *final_flip_bw, |
272 | bool_Bool *ImmediateFlipSupportedForPipe); |
273 | static double CalculateWriteBackDelay( |
274 | enum source_format_class WritebackPixelFormat, |
275 | double WritebackHRatio, |
276 | double WritebackVRatio, |
277 | unsigned int WritebackVTaps, |
278 | long WritebackDestinationWidth, |
279 | long WritebackDestinationHeight, |
280 | long WritebackSourceHeight, |
281 | unsigned int HTotal); |
282 | static void CalculateDynamicMetadataParameters( |
283 | int MaxInterDCNTileRepeaters, |
284 | double DPPCLK, |
285 | double DISPCLK, |
286 | double DCFClkDeepSleep, |
287 | double PixelClock, |
288 | long HTotal, |
289 | long VBlank, |
290 | long DynamicMetadataTransmittedBytes, |
291 | long DynamicMetadataLinesBeforeActiveRequired, |
292 | int InterlaceEnable, |
293 | bool_Bool ProgressiveToInterlaceUnitInOPP, |
294 | double *Tsetup, |
295 | double *Tdmbf, |
296 | double *Tdmec, |
297 | double *Tdmsks); |
298 | static void CalculateWatermarksAndDRAMSpeedChangeSupport( |
299 | struct display_mode_lib *mode_lib, |
300 | unsigned int PrefetchMode, |
301 | unsigned int NumberOfActivePlanes, |
302 | unsigned int MaxLineBufferLines, |
303 | unsigned int LineBufferSize, |
304 | unsigned int DPPOutputBufferPixels, |
305 | unsigned int DETBufferSizeInKByte, |
306 | unsigned int WritebackInterfaceBufferSize, |
307 | double DCFCLK, |
308 | double ReturnBW, |
309 | bool_Bool GPUVMEnable, |
310 | unsigned int dpte_group_bytes[], |
311 | unsigned int MetaChunkSize, |
312 | double UrgentLatency, |
313 | double ExtraLatency, |
314 | double WritebackLatency, |
315 | double WritebackChunkSize, |
316 | double SOCCLK, |
317 | double DRAMClockChangeLatency, |
318 | double SRExitTime, |
319 | double SREnterPlusExitTime, |
320 | double DCFCLKDeepSleep, |
321 | unsigned int DPPPerPlane[], |
322 | bool_Bool DCCEnable[], |
323 | double DPPCLK[], |
324 | unsigned int DETBufferSizeY[], |
325 | unsigned int DETBufferSizeC[], |
326 | unsigned int SwathHeightY[], |
327 | unsigned int SwathHeightC[], |
328 | unsigned int LBBitPerPixel[], |
329 | double SwathWidthY[], |
330 | double SwathWidthC[], |
331 | double HRatio[], |
332 | double HRatioChroma[], |
333 | unsigned int vtaps[], |
334 | unsigned int VTAPsChroma[], |
335 | double VRatio[], |
336 | double VRatioChroma[], |
337 | unsigned int HTotal[], |
338 | double PixelClock[], |
339 | unsigned int BlendingAndTiming[], |
340 | double BytePerPixelDETY[], |
341 | double BytePerPixelDETC[], |
342 | double DSTXAfterScaler[], |
343 | double DSTYAfterScaler[], |
344 | bool_Bool WritebackEnable[], |
345 | enum source_format_class WritebackPixelFormat[], |
346 | double WritebackDestinationWidth[], |
347 | double WritebackDestinationHeight[], |
348 | double WritebackSourceHeight[], |
349 | enum clock_change_support *DRAMClockChangeSupport, |
350 | double *UrgentWatermark, |
351 | double *WritebackUrgentWatermark, |
352 | double *DRAMClockChangeWatermark, |
353 | double *WritebackDRAMClockChangeWatermark, |
354 | double *StutterExitWatermark, |
355 | double *StutterEnterPlusExitWatermark, |
356 | double *MinActiveDRAMClockChangeLatencySupported); |
357 | static void CalculateDCFCLKDeepSleep( |
358 | struct display_mode_lib *mode_lib, |
359 | unsigned int NumberOfActivePlanes, |
360 | int BytePerPixelY[], |
361 | int BytePerPixelC[], |
362 | double VRatio[], |
363 | double VRatioChroma[], |
364 | double SwathWidthY[], |
365 | double SwathWidthC[], |
366 | unsigned int DPPPerPlane[], |
367 | double HRatio[], |
368 | double HRatioChroma[], |
369 | double PixelClock[], |
370 | double PSCL_THROUGHPUT[], |
371 | double PSCL_THROUGHPUT_CHROMA[], |
372 | double DPPCLK[], |
373 | double ReadBandwidthLuma[], |
374 | double ReadBandwidthChroma[], |
375 | int ReturnBusWidth, |
376 | double *DCFCLKDeepSleep); |
377 | static void CalculateUrgentBurstFactor( |
378 | long swath_width_luma_ub, |
379 | long swath_width_chroma_ub, |
380 | unsigned int DETBufferSizeInKByte, |
381 | unsigned int SwathHeightY, |
382 | unsigned int SwathHeightC, |
383 | double LineTime, |
384 | double UrgentLatency, |
385 | double CursorBufferSize, |
386 | unsigned int CursorWidth, |
387 | unsigned int CursorBPP, |
388 | double VRatio, |
389 | double VRatioC, |
390 | double BytePerPixelInDETY, |
391 | double BytePerPixelInDETC, |
392 | double DETBufferSizeY, |
393 | double DETBufferSizeC, |
394 | double *UrgentBurstFactorCursor, |
395 | double *UrgentBurstFactorLuma, |
396 | double *UrgentBurstFactorChroma, |
397 | bool_Bool *NotEnoughUrgentLatencyHiding); |
398 | |
399 | static void UseMinimumDCFCLK( |
400 | struct display_mode_lib *mode_lib, |
401 | struct vba_vars_st *v, |
402 | int MaxPrefetchMode, |
403 | int ReorderingBytes); |
404 | |
405 | static void CalculatePixelDeliveryTimes( |
406 | unsigned int NumberOfActivePlanes, |
407 | double VRatio[], |
408 | double VRatioChroma[], |
409 | double VRatioPrefetchY[], |
410 | double VRatioPrefetchC[], |
411 | unsigned int swath_width_luma_ub[], |
412 | unsigned int swath_width_chroma_ub[], |
413 | unsigned int DPPPerPlane[], |
414 | double HRatio[], |
415 | double HRatioChroma[], |
416 | double PixelClock[], |
417 | double PSCL_THROUGHPUT[], |
418 | double PSCL_THROUGHPUT_CHROMA[], |
419 | double DPPCLK[], |
420 | int BytePerPixelC[], |
421 | enum scan_direction_class SourceScan[], |
422 | unsigned int NumberOfCursors[], |
423 | unsigned int CursorWidth[][2], |
424 | unsigned int CursorBPP[][2], |
425 | unsigned int BlockWidth256BytesY[], |
426 | unsigned int BlockHeight256BytesY[], |
427 | unsigned int BlockWidth256BytesC[], |
428 | unsigned int BlockHeight256BytesC[], |
429 | double DisplayPipeLineDeliveryTimeLuma[], |
430 | double DisplayPipeLineDeliveryTimeChroma[], |
431 | double DisplayPipeLineDeliveryTimeLumaPrefetch[], |
432 | double DisplayPipeLineDeliveryTimeChromaPrefetch[], |
433 | double DisplayPipeRequestDeliveryTimeLuma[], |
434 | double DisplayPipeRequestDeliveryTimeChroma[], |
435 | double DisplayPipeRequestDeliveryTimeLumaPrefetch[], |
436 | double DisplayPipeRequestDeliveryTimeChromaPrefetch[], |
437 | double CursorRequestDeliveryTime[], |
438 | double CursorRequestDeliveryTimePrefetch[]); |
439 | |
440 | static void CalculateMetaAndPTETimes( |
441 | int NumberOfActivePlanes, |
442 | bool_Bool GPUVMEnable, |
443 | int MetaChunkSize, |
444 | int MinMetaChunkSizeBytes, |
445 | int HTotal[], |
446 | double VRatio[], |
447 | double VRatioChroma[], |
448 | double DestinationLinesToRequestRowInVBlank[], |
449 | double DestinationLinesToRequestRowInImmediateFlip[], |
450 | bool_Bool DCCEnable[], |
451 | double PixelClock[], |
452 | int BytePerPixelY[], |
453 | int BytePerPixelC[], |
454 | enum scan_direction_class SourceScan[], |
455 | int dpte_row_height[], |
456 | int dpte_row_height_chroma[], |
457 | int meta_row_width[], |
458 | int meta_row_width_chroma[], |
459 | int meta_row_height[], |
460 | int meta_row_height_chroma[], |
461 | int meta_req_width[], |
462 | int meta_req_width_chroma[], |
463 | int meta_req_height[], |
464 | int meta_req_height_chroma[], |
465 | int dpte_group_bytes[], |
466 | int PTERequestSizeY[], |
467 | int PTERequestSizeC[], |
468 | int PixelPTEReqWidthY[], |
469 | int PixelPTEReqHeightY[], |
470 | int PixelPTEReqWidthC[], |
471 | int PixelPTEReqHeightC[], |
472 | int dpte_row_width_luma_ub[], |
473 | int dpte_row_width_chroma_ub[], |
474 | double DST_Y_PER_PTE_ROW_NOM_L[], |
475 | double DST_Y_PER_PTE_ROW_NOM_C[], |
476 | double DST_Y_PER_META_ROW_NOM_L[], |
477 | double DST_Y_PER_META_ROW_NOM_C[], |
478 | double TimePerMetaChunkNominal[], |
479 | double TimePerChromaMetaChunkNominal[], |
480 | double TimePerMetaChunkVBlank[], |
481 | double TimePerChromaMetaChunkVBlank[], |
482 | double TimePerMetaChunkFlip[], |
483 | double TimePerChromaMetaChunkFlip[], |
484 | double time_per_pte_group_nom_luma[], |
485 | double time_per_pte_group_vblank_luma[], |
486 | double time_per_pte_group_flip_luma[], |
487 | double time_per_pte_group_nom_chroma[], |
488 | double time_per_pte_group_vblank_chroma[], |
489 | double time_per_pte_group_flip_chroma[]); |
490 | |
491 | static void CalculateVMGroupAndRequestTimes( |
492 | unsigned int NumberOfActivePlanes, |
493 | bool_Bool GPUVMEnable, |
494 | unsigned int GPUVMMaxPageTableLevels, |
495 | unsigned int HTotal[], |
496 | int BytePerPixelC[], |
497 | double DestinationLinesToRequestVMInVBlank[], |
498 | double DestinationLinesToRequestVMInImmediateFlip[], |
499 | bool_Bool DCCEnable[], |
500 | double PixelClock[], |
501 | int dpte_row_width_luma_ub[], |
502 | int dpte_row_width_chroma_ub[], |
503 | int vm_group_bytes[], |
504 | unsigned int dpde0_bytes_per_frame_ub_l[], |
505 | unsigned int dpde0_bytes_per_frame_ub_c[], |
506 | int meta_pte_bytes_per_frame_ub_l[], |
507 | int meta_pte_bytes_per_frame_ub_c[], |
508 | double TimePerVMGroupVBlank[], |
509 | double TimePerVMGroupFlip[], |
510 | double TimePerVMRequestVBlank[], |
511 | double TimePerVMRequestFlip[]); |
512 | |
513 | static void CalculateStutterEfficiency( |
514 | int NumberOfActivePlanes, |
515 | long ROBBufferSizeInKByte, |
516 | double TotalDataReadBandwidth, |
517 | double DCFCLK, |
518 | double ReturnBW, |
519 | double SRExitTime, |
520 | bool_Bool SynchronizedVBlank, |
521 | int DPPPerPlane[], |
522 | unsigned int DETBufferSizeY[], |
523 | int BytePerPixelY[], |
524 | double BytePerPixelDETY[], |
525 | double SwathWidthY[], |
526 | int SwathHeightY[], |
527 | int SwathHeightC[], |
528 | double DCCRateLuma[], |
529 | double DCCRateChroma[], |
530 | int HTotal[], |
531 | int VTotal[], |
532 | double PixelClock[], |
533 | double VRatio[], |
534 | enum scan_direction_class SourceScan[], |
535 | int BlockHeight256BytesY[], |
536 | int BlockWidth256BytesY[], |
537 | int BlockHeight256BytesC[], |
538 | int BlockWidth256BytesC[], |
539 | int DCCYMaxUncompressedBlock[], |
540 | int DCCCMaxUncompressedBlock[], |
541 | int VActive[], |
542 | bool_Bool DCCEnable[], |
543 | bool_Bool WritebackEnable[], |
544 | double ReadBandwidthPlaneLuma[], |
545 | double ReadBandwidthPlaneChroma[], |
546 | double meta_row_bw[], |
547 | double dpte_row_bw[], |
548 | double *StutterEfficiencyNotIncludingVBlank, |
549 | double *StutterEfficiency, |
550 | double *StutterPeriodOut); |
551 | |
552 | static void CalculateSwathAndDETConfiguration( |
553 | bool_Bool ForceSingleDPP, |
554 | int NumberOfActivePlanes, |
555 | unsigned int DETBufferSizeInKByte, |
556 | double MaximumSwathWidthLuma[], |
557 | double MaximumSwathWidthChroma[], |
558 | enum scan_direction_class SourceScan[], |
559 | enum source_format_class SourcePixelFormat[], |
560 | enum dm_swizzle_mode SurfaceTiling[], |
561 | int ViewportWidth[], |
562 | int ViewportHeight[], |
563 | int SurfaceWidthY[], |
564 | int SurfaceWidthC[], |
565 | int SurfaceHeightY[], |
566 | int SurfaceHeightC[], |
567 | int Read256BytesBlockHeightY[], |
568 | int Read256BytesBlockHeightC[], |
569 | int Read256BytesBlockWidthY[], |
570 | int Read256BytesBlockWidthC[], |
571 | enum odm_combine_mode ODMCombineEnabled[], |
572 | int BlendingAndTiming[], |
573 | int BytePerPixY[], |
574 | int BytePerPixC[], |
575 | double BytePerPixDETY[], |
576 | double BytePerPixDETC[], |
577 | int HActive[], |
578 | double HRatio[], |
579 | double HRatioChroma[], |
580 | int DPPPerPlane[], |
581 | int swath_width_luma_ub[], |
582 | int swath_width_chroma_ub[], |
583 | double SwathWidth[], |
584 | double SwathWidthChroma[], |
585 | int SwathHeightY[], |
586 | int SwathHeightC[], |
587 | unsigned int DETBufferSizeY[], |
588 | unsigned int DETBufferSizeC[], |
589 | bool_Bool ViewportSizeSupportPerPlane[], |
590 | bool_Bool *ViewportSizeSupport); |
591 | static void CalculateSwathWidth( |
592 | bool_Bool ForceSingleDPP, |
593 | int NumberOfActivePlanes, |
594 | enum source_format_class SourcePixelFormat[], |
595 | enum scan_direction_class SourceScan[], |
596 | unsigned int ViewportWidth[], |
597 | unsigned int ViewportHeight[], |
598 | unsigned int SurfaceWidthY[], |
599 | unsigned int SurfaceWidthC[], |
600 | unsigned int SurfaceHeightY[], |
601 | unsigned int SurfaceHeightC[], |
602 | enum odm_combine_mode ODMCombineEnabled[], |
603 | int BytePerPixY[], |
604 | int BytePerPixC[], |
605 | int Read256BytesBlockHeightY[], |
606 | int Read256BytesBlockHeightC[], |
607 | int Read256BytesBlockWidthY[], |
608 | int Read256BytesBlockWidthC[], |
609 | int BlendingAndTiming[], |
610 | unsigned int HActive[], |
611 | double HRatio[], |
612 | int DPPPerPlane[], |
613 | double SwathWidthSingleDPPY[], |
614 | double SwathWidthSingleDPPC[], |
615 | double SwathWidthY[], |
616 | double SwathWidthC[], |
617 | int MaximumSwathHeightY[], |
618 | int MaximumSwathHeightC[], |
619 | unsigned int swath_width_luma_ub[], |
620 | unsigned int swath_width_chroma_ub[]); |
621 | static double CalculateExtraLatency( |
622 | long RoundTripPingLatencyCycles, |
623 | long ReorderingBytes, |
624 | double DCFCLK, |
625 | int TotalNumberOfActiveDPP, |
626 | int PixelChunkSizeInKByte, |
627 | int TotalNumberOfDCCActiveDPP, |
628 | int MetaChunkSize, |
629 | double ReturnBW, |
630 | bool_Bool GPUVMEnable, |
631 | bool_Bool HostVMEnable, |
632 | int NumberOfActivePlanes, |
633 | int NumberOfDPP[], |
634 | int dpte_group_bytes[], |
635 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
636 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
637 | double HostVMMinPageSize, |
638 | int HostVMMaxNonCachedPageTableLevels); |
639 | static double CalculateExtraLatencyBytes( |
640 | long ReorderingBytes, |
641 | int TotalNumberOfActiveDPP, |
642 | int PixelChunkSizeInKByte, |
643 | int TotalNumberOfDCCActiveDPP, |
644 | int MetaChunkSize, |
645 | bool_Bool GPUVMEnable, |
646 | bool_Bool HostVMEnable, |
647 | int NumberOfActivePlanes, |
648 | int NumberOfDPP[], |
649 | int dpte_group_bytes[], |
650 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
651 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
652 | double HostVMMinPageSize, |
653 | int HostVMMaxNonCachedPageTableLevels); |
654 | static double CalculateUrgentLatency( |
655 | double UrgentLatencyPixelDataOnly, |
656 | double UrgentLatencyPixelMixedWithVMData, |
657 | double UrgentLatencyVMDataOnly, |
658 | bool_Bool DoUrgentLatencyAdjustment, |
659 | double UrgentLatencyAdjustmentFabricClockComponent, |
660 | double UrgentLatencyAdjustmentFabricClockReference, |
661 | double FabricClockSingle); |
662 | |
663 | void dml30_recalculate(struct display_mode_lib *mode_lib) |
664 | { |
665 | ModeSupportAndSystemConfiguration(mode_lib); |
666 | PixelClockAdjustmentForProgressiveToInterlaceUnit(mode_lib); |
667 | DisplayPipeConfiguration(mode_lib); |
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, lstall, D, initalXmitDelay, w, s, ix, wx, P, l0, a, ax, L, |
692 | Delay, pixels; |
693 | |
694 | if (pixelFormat == dm_420) |
695 | pixelsPerClock = 2; |
696 | // #all other modes operate at 1 pixel per clock |
697 | else if (pixelFormat == dm_444) |
698 | pixelsPerClock = 1; |
699 | else if (pixelFormat == dm_n422) |
700 | pixelsPerClock = 2; |
701 | else |
702 | pixelsPerClock = 1; |
703 | |
704 | //initial transmit delay as per PPS |
705 | initalXmitDelay = dml_round(rcModelSize / 2.0 / BPP / pixelsPerClock); |
706 | |
707 | //compute ssm delay |
708 | if (bpc == 8) |
709 | D = 81; |
710 | else if (bpc == 10) |
711 | D = 89; |
712 | else |
713 | D = 113; |
714 | |
715 | //divide by pixel per cycle to compute slice width as seen by DSC |
716 | w = sliceWidth / pixelsPerClock; |
717 | |
718 | //422 mode has an additional cycle of delay |
719 | if (pixelFormat == dm_420 || pixelFormat == dm_444 || pixelFormat == dm_n422) |
720 | s = 0; |
721 | else |
722 | s = 1; |
723 | |
724 | //main calculation for the dscce |
725 | ix = initalXmitDelay + 45; |
726 | wx = (w + 2) / 3; |
727 | P = 3 * wx - w; |
728 | l0 = ix / w; |
729 | a = ix + P * l0; |
730 | ax = (a + 2) / 3 + D + 6 + 1; |
731 | L = (ax + wx - 1) / wx; |
732 | if ((ix % w) == 0 && P != 0) |
733 | lstall = 1; |
734 | else |
735 | lstall = 0; |
736 | Delay = L * wx * (numSlices - 1) + ax + s + lstall + 22; |
737 | |
738 | //dsc processes 3 pixel containers per cycle and a container can contain 1 or 2 pixels |
739 | pixels = Delay * 3 * pixelsPerClock; |
740 | return pixels; |
741 | } |
742 | |
743 | static unsigned int dscComputeDelay(enum output_format_class pixelFormat, enum output_encoder_class Output) |
744 | { |
745 | unsigned int Delay = 0; |
746 | |
747 | if (pixelFormat == dm_420) { |
748 | // sfr |
749 | Delay = Delay + 2; |
750 | // dsccif |
751 | Delay = Delay + 0; |
752 | // dscc - input deserializer |
753 | Delay = Delay + 3; |
754 | // dscc gets pixels every other cycle |
755 | Delay = Delay + 2; |
756 | // dscc - input cdc fifo |
757 | Delay = Delay + 12; |
758 | // dscc gets pixels every other cycle |
759 | Delay = Delay + 13; |
760 | // dscc - cdc uncertainty |
761 | Delay = Delay + 2; |
762 | // dscc - output cdc fifo |
763 | Delay = Delay + 7; |
764 | // dscc gets pixels every other cycle |
765 | Delay = Delay + 3; |
766 | // dscc - cdc uncertainty |
767 | Delay = Delay + 2; |
768 | // dscc - output serializer |
769 | Delay = Delay + 1; |
770 | // sft |
771 | Delay = Delay + 1; |
772 | } else if (pixelFormat == dm_n422) { |
773 | // sfr |
774 | Delay = Delay + 2; |
775 | // dsccif |
776 | Delay = Delay + 1; |
777 | // dscc - input deserializer |
778 | Delay = Delay + 5; |
779 | // dscc - input cdc fifo |
780 | Delay = Delay + 25; |
781 | // dscc - cdc uncertainty |
782 | Delay = Delay + 2; |
783 | // dscc - output cdc fifo |
784 | Delay = Delay + 10; |
785 | // dscc - cdc uncertainty |
786 | Delay = Delay + 2; |
787 | // dscc - output serializer |
788 | Delay = Delay + 1; |
789 | // sft |
790 | Delay = Delay + 1; |
791 | } |
792 | else { |
793 | // sfr |
794 | Delay = Delay + 2; |
795 | // dsccif |
796 | Delay = Delay + 0; |
797 | // dscc - input deserializer |
798 | Delay = Delay + 3; |
799 | // dscc - input cdc fifo |
800 | Delay = Delay + 12; |
801 | // dscc - cdc uncertainty |
802 | Delay = Delay + 2; |
803 | // dscc - output cdc fifo |
804 | Delay = Delay + 7; |
805 | // dscc - output serializer |
806 | Delay = Delay + 1; |
807 | // dscc - cdc uncertainty |
808 | Delay = Delay + 2; |
809 | // sft |
810 | Delay = Delay + 1; |
811 | } |
812 | |
813 | return Delay; |
814 | } |
815 | |
816 | static bool_Bool CalculatePrefetchSchedule( |
817 | struct display_mode_lib *mode_lib, |
818 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
819 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
820 | Pipe *myPipe, |
821 | unsigned int DSCDelay, |
822 | double DPPCLKDelaySubtotalPlusCNVCFormater, |
823 | double DPPCLKDelaySCL, |
824 | double DPPCLKDelaySCLLBOnly, |
825 | double DPPCLKDelayCNVCCursor, |
826 | double DISPCLKDelaySubtotal, |
827 | unsigned int DPP_RECOUT_WIDTH, |
828 | enum output_format_class OutputFormat, |
829 | unsigned int MaxInterDCNTileRepeaters, |
830 | unsigned int VStartup, |
831 | unsigned int MaxVStartup, |
832 | unsigned int GPUVMPageTableLevels, |
833 | bool_Bool GPUVMEnable, |
834 | bool_Bool HostVMEnable, |
835 | unsigned int HostVMMaxNonCachedPageTableLevels, |
836 | double HostVMMinPageSize, |
837 | bool_Bool DynamicMetadataEnable, |
838 | bool_Bool DynamicMetadataVMEnabled, |
839 | int DynamicMetadataLinesBeforeActiveRequired, |
840 | unsigned int DynamicMetadataTransmittedBytes, |
841 | double UrgentLatency, |
842 | double UrgentExtraLatency, |
843 | double TCalc, |
844 | unsigned int PDEAndMetaPTEBytesFrame, |
845 | unsigned int MetaRowByte, |
846 | unsigned int PixelPTEBytesPerRow, |
847 | double PrefetchSourceLinesY, |
848 | unsigned int SwathWidthY, |
849 | int BytePerPixelY, |
850 | double VInitPreFillY, |
851 | unsigned int MaxNumSwathY, |
852 | double PrefetchSourceLinesC, |
853 | unsigned int SwathWidthC, |
854 | int BytePerPixelC, |
855 | double VInitPreFillC, |
856 | unsigned int MaxNumSwathC, |
857 | long swath_width_luma_ub, |
858 | long swath_width_chroma_ub, |
859 | unsigned int SwathHeightY, |
860 | unsigned int SwathHeightC, |
861 | double TWait, |
862 | bool_Bool ProgressiveToInterlaceUnitInOPP, |
863 | double *DSTXAfterScaler, |
864 | double *DSTYAfterScaler, |
865 | double *DestinationLinesForPrefetch, |
866 | double *PrefetchBandwidth, |
867 | double *DestinationLinesToRequestVMInVBlank, |
868 | double *DestinationLinesToRequestRowInVBlank, |
869 | double *VRatioPrefetchY, |
870 | double *VRatioPrefetchC, |
871 | double *RequiredPrefetchPixDataBWLuma, |
872 | double *RequiredPrefetchPixDataBWChroma, |
873 | bool_Bool *NotEnoughTimeForDynamicMetadata, |
874 | double *Tno_bw, |
875 | double *prefetch_vmrow_bw, |
876 | double *Tdmdl_vm, |
877 | double *Tdmdl, |
878 | unsigned int *VUpdateOffsetPix, |
879 | double *VUpdateWidthPix, |
880 | double *VReadyOffsetPix) |
881 | { |
882 | bool_Bool MyError = false0; |
883 | unsigned int DPPCycles = 0, DISPCLKCycles = 0; |
884 | double DSTTotalPixelsAfterScaler = 0; |
885 | double LineTime = 0, Tsetup = 0; |
886 | double dst_y_prefetch_equ = 0; |
887 | double Tsw_oto = 0; |
888 | double prefetch_bw_oto = 0; |
889 | double Tvm_oto = 0; |
890 | double Tr0_oto = 0; |
891 | double Tvm_oto_lines = 0; |
892 | double Tr0_oto_lines = 0; |
893 | double dst_y_prefetch_oto = 0; |
894 | double TimeForFetchingMetaPTE = 0; |
895 | double TimeForFetchingRowInVBlank = 0; |
896 | double LinesToRequestPrefetchPixelData = 0; |
897 | double HostVMInefficiencyFactor = 0; |
898 | unsigned int HostVMDynamicLevelsTrips = 0; |
899 | double trip_to_mem = 0; |
900 | double Tvm_trips = 0; |
901 | double Tr0_trips = 0; |
902 | double Tvm_trips_rounded = 0; |
903 | double Tr0_trips_rounded = 0; |
904 | double Lsw_oto = 0; |
905 | double Tpre_rounded = 0; |
906 | double prefetch_bw_equ = 0; |
907 | double Tvm_equ = 0; |
908 | double Tr0_equ = 0; |
909 | double Tdmbf = 0; |
910 | double Tdmec = 0; |
911 | double Tdmsks = 0; |
912 | |
913 | if (GPUVMEnable == true1 && HostVMEnable == true1) { |
914 | HostVMInefficiencyFactor = PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData / PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly; |
915 | HostVMDynamicLevelsTrips = HostVMMaxNonCachedPageTableLevels; |
916 | } else { |
917 | HostVMInefficiencyFactor = 1; |
918 | HostVMDynamicLevelsTrips = 0; |
919 | } |
920 | |
921 | CalculateDynamicMetadataParameters( |
922 | MaxInterDCNTileRepeaters, |
923 | myPipe->DPPCLK, |
924 | myPipe->DISPCLK, |
925 | myPipe->DCFCLKDeepSleep, |
926 | myPipe->PixelClock, |
927 | myPipe->HTotal, |
928 | myPipe->VBlank, |
929 | DynamicMetadataTransmittedBytes, |
930 | DynamicMetadataLinesBeforeActiveRequired, |
931 | myPipe->InterlaceEnable, |
932 | ProgressiveToInterlaceUnitInOPP, |
933 | &Tsetup, |
934 | &Tdmbf, |
935 | &Tdmec, |
936 | &Tdmsks); |
937 | |
938 | LineTime = myPipe->HTotal / myPipe->PixelClock; |
939 | trip_to_mem = UrgentLatency; |
940 | Tvm_trips = UrgentExtraLatency + trip_to_mem * (GPUVMPageTableLevels * (HostVMDynamicLevelsTrips + 1) - 1); |
941 | |
942 | if (DynamicMetadataVMEnabled == true1 && GPUVMEnable == true1) { |
943 | *Tdmdl = TWait + Tvm_trips + trip_to_mem; |
944 | } else { |
945 | *Tdmdl = TWait + UrgentExtraLatency; |
946 | } |
947 | |
948 | if (DynamicMetadataEnable == true1) { |
949 | if (VStartup * LineTime < Tsetup + *Tdmdl + Tdmbf + Tdmec + Tdmsks) { |
950 | *NotEnoughTimeForDynamicMetadata = true1; |
951 | } else { |
952 | *NotEnoughTimeForDynamicMetadata = false0; |
953 | dml_print("DML: Not Enough Time for Dynamic Meta!\n"){do { } while(0); }; |
954 | dml_print("DML: Tdmbf: %fus - time for dmd transfer from dchub to dio output buffer\n", Tdmbf){do { } while(0); }; |
955 | dml_print("DML: Tdmec: %fus - time dio takes to transfer dmd\n", Tdmec){do { } while(0); }; |
956 | dml_print("DML: Tdmsks: %fus - time before active dmd must complete transmission at dio\n", Tdmsks){do { } while(0); }; |
957 | dml_print("DML: Tdmdl: %fus - time for fabric to become ready and fetch dmd \n", *Tdmdl){do { } while(0); }; |
958 | } |
959 | } else { |
960 | *NotEnoughTimeForDynamicMetadata = false0; |
961 | } |
962 | |
963 | *Tdmdl_vm = (DynamicMetadataEnable == true1 && DynamicMetadataVMEnabled == true1 && GPUVMEnable == true1 ? TWait + Tvm_trips : 0); |
964 | |
965 | if (myPipe->ScalerEnabled) |
966 | DPPCycles = DPPCLKDelaySubtotalPlusCNVCFormater + DPPCLKDelaySCL; |
967 | else |
968 | DPPCycles = DPPCLKDelaySubtotalPlusCNVCFormater + DPPCLKDelaySCLLBOnly; |
969 | |
970 | DPPCycles = DPPCycles + myPipe->NumberOfCursors * DPPCLKDelayCNVCCursor; |
971 | |
972 | DISPCLKCycles = DISPCLKDelaySubtotal; |
973 | |
974 | if (myPipe->DPPCLK == 0.0 || myPipe->DISPCLK == 0.0) |
975 | return true1; |
976 | |
977 | *DSTXAfterScaler = DPPCycles * myPipe->PixelClock / myPipe->DPPCLK + DISPCLKCycles * myPipe->PixelClock / myPipe->DISPCLK |
978 | + DSCDelay; |
979 | |
980 | *DSTXAfterScaler = *DSTXAfterScaler + ((myPipe->ODMCombineEnabled)?18:0) + (myPipe->DPPPerPlane - 1) * DPP_RECOUT_WIDTH; |
981 | |
982 | if (OutputFormat == dm_420 || (myPipe->InterlaceEnable && ProgressiveToInterlaceUnitInOPP)) |
983 | *DSTYAfterScaler = 1; |
984 | else |
985 | *DSTYAfterScaler = 0; |
986 | |
987 | DSTTotalPixelsAfterScaler = *DSTYAfterScaler * myPipe->HTotal + *DSTXAfterScaler; |
988 | *DSTYAfterScaler = dml_floor(DSTTotalPixelsAfterScaler / myPipe->HTotal, 1); |
989 | *DSTXAfterScaler = DSTTotalPixelsAfterScaler - ((double) (*DSTYAfterScaler * myPipe->HTotal)); |
990 | |
991 | MyError = false0; |
992 | |
993 | |
994 | Tr0_trips = trip_to_mem * (HostVMDynamicLevelsTrips + 1); |
995 | Tvm_trips_rounded = dml_ceil(4.0 * Tvm_trips / LineTime, 1) / 4 * LineTime; |
996 | Tr0_trips_rounded = dml_ceil(4.0 * Tr0_trips / LineTime, 1) / 4 * LineTime; |
997 | |
998 | if (GPUVMEnable) { |
999 | if (GPUVMPageTableLevels >= 3) { |
1000 | *Tno_bw = UrgentExtraLatency + trip_to_mem * ((GPUVMPageTableLevels - 2) - 1); |
1001 | } else |
1002 | *Tno_bw = 0; |
1003 | } else if (!myPipe->DCCEnable) |
1004 | *Tno_bw = LineTime; |
1005 | else |
1006 | *Tno_bw = LineTime / 4; |
1007 | |
1008 | dst_y_prefetch_equ = VStartup - (Tsetup + dml_max(TWait + TCalc, *Tdmdl)) / LineTime |
1009 | - (*DSTYAfterScaler + *DSTXAfterScaler / myPipe->HTotal); |
1010 | |
1011 | Lsw_oto = dml_max(PrefetchSourceLinesY, PrefetchSourceLinesC); |
1012 | Tsw_oto = Lsw_oto * LineTime; |
1013 | |
1014 | prefetch_bw_oto = (PrefetchSourceLinesY * swath_width_luma_ub * BytePerPixelY + PrefetchSourceLinesC * swath_width_chroma_ub * BytePerPixelC) / Tsw_oto; |
1015 | |
1016 | if (GPUVMEnable == true1) { |
1017 | Tvm_oto = dml_max3(*Tno_bw + PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor / prefetch_bw_oto, |
1018 | Tvm_trips, |
1019 | LineTime / 4.0); |
1020 | } else |
1021 | Tvm_oto = LineTime / 4.0; |
1022 | |
1023 | if ((GPUVMEnable == true1 || myPipe->DCCEnable == true1)) { |
1024 | Tr0_oto = dml_max3( |
1025 | (MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor) / prefetch_bw_oto, |
1026 | LineTime - Tvm_oto, LineTime / 4); |
1027 | } else |
1028 | Tr0_oto = (LineTime - Tvm_oto) / 2.0; |
1029 | |
1030 | Tvm_oto_lines = dml_ceil(4.0 * Tvm_oto / LineTime, 1) / 4.0; |
1031 | Tr0_oto_lines = dml_ceil(4.0 * Tr0_oto / LineTime, 1) / 4.0; |
1032 | dst_y_prefetch_oto = Tvm_oto_lines + 2 * Tr0_oto_lines + Lsw_oto; |
1033 | |
1034 | dst_y_prefetch_equ = dml_floor(4.0 * (dst_y_prefetch_equ + 0.125), 1) / 4.0; |
1035 | Tpre_rounded = dst_y_prefetch_equ * LineTime; |
1036 | |
1037 | dml_print("DML: dst_y_prefetch_oto: %f\n", dst_y_prefetch_oto){do { } while(0); }; |
1038 | dml_print("DML: dst_y_prefetch_equ: %f\n", dst_y_prefetch_equ){do { } while(0); }; |
1039 | |
1040 | dml_print("DML: LineTime: %f\n", LineTime){do { } while(0); }; |
1041 | dml_print("DML: VStartup: %d\n", VStartup){do { } while(0); }; |
1042 | dml_print("DML: Tvstartup: %fus - time between vstartup and first pixel of active\n", VStartup * LineTime){do { } while(0); }; |
1043 | dml_print("DML: Tsetup: %fus - time from vstartup to vready\n", Tsetup){do { } while(0); }; |
1044 | dml_print("DML: TCalc: %fus - time for calculations in dchub starting at vready\n", TCalc){do { } while(0); }; |
1045 | 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); }; |
1046 | dml_print("DML: Tdmbf: %fus - time for dmd transfer from dchub to dio output buffer\n", Tdmbf){do { } while(0); }; |
1047 | dml_print("DML: Tdmec: %fus - time dio takes to transfer dmd\n", Tdmec){do { } while(0); }; |
1048 | dml_print("DML: Tdmsks: %fus - time before active dmd must complete transmission at dio\n", Tdmsks){do { } while(0); }; |
1049 | dml_print("DML: Tdmdl_vm: %fus - time for vm stages of dmd \n", *Tdmdl_vm){do { } while(0); }; |
1050 | dml_print("DML: Tdmdl: %fus - time for fabric to become ready and fetch dmd \n", *Tdmdl){do { } while(0); }; |
1051 | dml_print("DML: dst_x_after_scl: %f pixels - number of pixel clocks pipeline and buffer delay after scaler \n", *DSTXAfterScaler){do { } while(0); }; |
1052 | dml_print("DML: dst_y_after_scl: %d lines - number of lines of pipeline and buffer delay after scaler \n", (int)*DSTYAfterScaler){do { } while(0); }; |
1053 | |
1054 | *PrefetchBandwidth = 0; |
1055 | *DestinationLinesToRequestVMInVBlank = 0; |
1056 | *DestinationLinesToRequestRowInVBlank = 0; |
1057 | *VRatioPrefetchY = 0; |
1058 | *VRatioPrefetchC = 0; |
1059 | *RequiredPrefetchPixDataBWLuma = 0; |
1060 | if (dst_y_prefetch_equ > 1) { |
1061 | double PrefetchBandwidth1 = 0; |
1062 | double PrefetchBandwidth2 = 0; |
1063 | double PrefetchBandwidth3 = 0; |
1064 | double PrefetchBandwidth4 = 0; |
1065 | |
1066 | if (Tpre_rounded - *Tno_bw > 0) |
1067 | PrefetchBandwidth1 = (PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor + 2 * MetaRowByte |
1068 | + 2 * PixelPTEBytesPerRow * HostVMInefficiencyFactor |
1069 | + PrefetchSourceLinesY * swath_width_luma_ub * BytePerPixelY |
1070 | + PrefetchSourceLinesC * swath_width_chroma_ub * BytePerPixelC) |
1071 | / (Tpre_rounded - *Tno_bw); |
1072 | else |
1073 | PrefetchBandwidth1 = 0; |
1074 | |
1075 | if (VStartup == MaxVStartup && (PrefetchBandwidth1 > 4 * prefetch_bw_oto) && (Tpre_rounded - Tsw_oto / 4 - 0.75 * LineTime - *Tno_bw) > 0) { |
1076 | PrefetchBandwidth1 = (PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor + 2 * MetaRowByte + 2 * PixelPTEBytesPerRow * HostVMInefficiencyFactor) / (Tpre_rounded - Tsw_oto / 4 - 0.75 * LineTime - *Tno_bw); |
1077 | } |
1078 | |
1079 | if (Tpre_rounded - *Tno_bw - 2 * Tr0_trips_rounded > 0) |
1080 | PrefetchBandwidth2 = (PDEAndMetaPTEBytesFrame * |
1081 | HostVMInefficiencyFactor + PrefetchSourceLinesY * |
1082 | swath_width_luma_ub * BytePerPixelY + |
1083 | PrefetchSourceLinesC * swath_width_chroma_ub * |
1084 | BytePerPixelC) / |
1085 | (Tpre_rounded - *Tno_bw - 2 * Tr0_trips_rounded); |
1086 | else |
1087 | PrefetchBandwidth2 = 0; |
1088 | |
1089 | if (Tpre_rounded - Tvm_trips_rounded > 0) |
1090 | PrefetchBandwidth3 = (2 * MetaRowByte + 2 * PixelPTEBytesPerRow * |
1091 | HostVMInefficiencyFactor + PrefetchSourceLinesY * |
1092 | swath_width_luma_ub * BytePerPixelY + PrefetchSourceLinesC * |
1093 | swath_width_chroma_ub * BytePerPixelC) / (Tpre_rounded - |
1094 | Tvm_trips_rounded); |
1095 | else |
1096 | PrefetchBandwidth3 = 0; |
1097 | |
1098 | if (VStartup == MaxVStartup && (PrefetchBandwidth3 > 4 * prefetch_bw_oto) && Tpre_rounded - Tsw_oto / 4 - 0.75 * LineTime - Tvm_trips_rounded > 0) { |
1099 | PrefetchBandwidth3 = (2 * MetaRowByte + 2 * PixelPTEBytesPerRow * HostVMInefficiencyFactor) / (Tpre_rounded - Tsw_oto / 4 - 0.75 * LineTime - Tvm_trips_rounded); |
1100 | } |
1101 | |
1102 | if (Tpre_rounded - Tvm_trips_rounded - 2 * Tr0_trips_rounded > 0) |
1103 | PrefetchBandwidth4 = (PrefetchSourceLinesY * swath_width_luma_ub * BytePerPixelY + PrefetchSourceLinesC * swath_width_chroma_ub * BytePerPixelC) |
1104 | / (Tpre_rounded - Tvm_trips_rounded - 2 * Tr0_trips_rounded); |
1105 | else |
1106 | PrefetchBandwidth4 = 0; |
1107 | |
1108 | { |
1109 | bool_Bool Case1OK; |
1110 | bool_Bool Case2OK; |
1111 | bool_Bool Case3OK; |
1112 | |
1113 | if (PrefetchBandwidth1 > 0) { |
1114 | if (*Tno_bw + PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor / PrefetchBandwidth1 |
1115 | >= Tvm_trips_rounded && (MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor) / PrefetchBandwidth1 >= Tr0_trips_rounded) { |
1116 | Case1OK = true1; |
1117 | } else { |
1118 | Case1OK = false0; |
1119 | } |
1120 | } else { |
1121 | Case1OK = false0; |
1122 | } |
1123 | |
1124 | if (PrefetchBandwidth2 > 0) { |
1125 | if (*Tno_bw + PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor / PrefetchBandwidth2 |
1126 | >= Tvm_trips_rounded && (MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor) / PrefetchBandwidth2 < Tr0_trips_rounded) { |
1127 | Case2OK = true1; |
1128 | } else { |
1129 | Case2OK = false0; |
1130 | } |
1131 | } else { |
1132 | Case2OK = false0; |
1133 | } |
1134 | |
1135 | if (PrefetchBandwidth3 > 0) { |
1136 | if (*Tno_bw + PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor / PrefetchBandwidth3 |
1137 | < Tvm_trips_rounded && (MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor) / PrefetchBandwidth3 >= Tr0_trips_rounded) { |
1138 | Case3OK = true1; |
1139 | } else { |
1140 | Case3OK = false0; |
1141 | } |
1142 | } else { |
1143 | Case3OK = false0; |
1144 | } |
1145 | |
1146 | if (Case1OK) { |
1147 | prefetch_bw_equ = PrefetchBandwidth1; |
1148 | } else if (Case2OK) { |
1149 | prefetch_bw_equ = PrefetchBandwidth2; |
1150 | } else if (Case3OK) { |
1151 | prefetch_bw_equ = PrefetchBandwidth3; |
1152 | } else { |
1153 | prefetch_bw_equ = PrefetchBandwidth4; |
1154 | } |
1155 | |
1156 | dml_print("DML: prefetch_bw_equ: %f\n", prefetch_bw_equ){do { } while(0); }; |
1157 | |
1158 | if (prefetch_bw_equ > 0) { |
1159 | if (GPUVMEnable) { |
1160 | Tvm_equ = dml_max3(*Tno_bw + PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor / prefetch_bw_equ, Tvm_trips, LineTime / 4); |
1161 | } else { |
1162 | Tvm_equ = LineTime / 4; |
1163 | } |
1164 | |
1165 | if ((GPUVMEnable || myPipe->DCCEnable)) { |
1166 | Tr0_equ = dml_max4( |
1167 | (MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor) / prefetch_bw_equ, |
1168 | Tr0_trips, |
1169 | (LineTime - Tvm_equ) / 2, |
1170 | LineTime / 4); |
1171 | } else { |
1172 | Tr0_equ = (LineTime - Tvm_equ) / 2; |
1173 | } |
1174 | } else { |
1175 | Tvm_equ = 0; |
1176 | Tr0_equ = 0; |
1177 | dml_print("DML: prefetch_bw_equ equals 0! %s:%d\n", __FILE__, __LINE__){do { } while(0); }; |
1178 | } |
1179 | } |
1180 | |
1181 | if (dst_y_prefetch_oto < dst_y_prefetch_equ) { |
1182 | *DestinationLinesForPrefetch = dst_y_prefetch_oto; |
1183 | TimeForFetchingMetaPTE = Tvm_oto; |
1184 | TimeForFetchingRowInVBlank = Tr0_oto; |
1185 | *PrefetchBandwidth = prefetch_bw_oto; |
1186 | } else { |
1187 | *DestinationLinesForPrefetch = dst_y_prefetch_equ; |
1188 | TimeForFetchingMetaPTE = Tvm_equ; |
1189 | TimeForFetchingRowInVBlank = Tr0_equ; |
1190 | *PrefetchBandwidth = prefetch_bw_equ; |
1191 | } |
1192 | |
1193 | *DestinationLinesToRequestVMInVBlank = dml_ceil(4.0 * TimeForFetchingMetaPTE / LineTime, 1.0) / 4.0; |
1194 | |
1195 | *DestinationLinesToRequestRowInVBlank = dml_ceil(4.0 * TimeForFetchingRowInVBlank / LineTime, 1.0) / 4.0; |
1196 | |
1197 | |
1198 | LinesToRequestPrefetchPixelData = *DestinationLinesForPrefetch - *DestinationLinesToRequestVMInVBlank |
1199 | - 2 * *DestinationLinesToRequestRowInVBlank; |
1200 | |
1201 | if (LinesToRequestPrefetchPixelData > 0 && prefetch_bw_equ > 0) { |
1202 | |
1203 | *VRatioPrefetchY = (double) PrefetchSourceLinesY |
1204 | / LinesToRequestPrefetchPixelData; |
1205 | *VRatioPrefetchY = dml_max(*VRatioPrefetchY, 1.0); |
1206 | if ((SwathHeightY > 4) && (VInitPreFillY > 3)) { |
1207 | if (LinesToRequestPrefetchPixelData > (VInitPreFillY - 3.0) / 2.0) { |
1208 | *VRatioPrefetchY = dml_max((double) PrefetchSourceLinesY / LinesToRequestPrefetchPixelData, |
1209 | (double) MaxNumSwathY * SwathHeightY / (LinesToRequestPrefetchPixelData - (VInitPreFillY - 3.0) / 2.0)); |
1210 | *VRatioPrefetchY = dml_max(*VRatioPrefetchY, 1.0); |
1211 | } else { |
1212 | MyError = true1; |
1213 | dml_print("DML: MyErr set %s:%d\n", __FILE__, __LINE__){do { } while(0); }; |
1214 | *VRatioPrefetchY = 0; |
1215 | } |
1216 | } |
1217 | |
1218 | *VRatioPrefetchC = (double) PrefetchSourceLinesC / LinesToRequestPrefetchPixelData; |
1219 | *VRatioPrefetchC = dml_max(*VRatioPrefetchC, 1.0); |
1220 | |
1221 | if ((SwathHeightC > 4)) { |
1222 | if (LinesToRequestPrefetchPixelData > (VInitPreFillC - 3.0) / 2.0) { |
1223 | *VRatioPrefetchC = dml_max(*VRatioPrefetchC, |
1224 | (double) MaxNumSwathC * SwathHeightC / (LinesToRequestPrefetchPixelData - (VInitPreFillC - 3.0) / 2.0)); |
1225 | *VRatioPrefetchC = dml_max(*VRatioPrefetchC, 1.0); |
1226 | } else { |
1227 | MyError = true1; |
1228 | dml_print("DML: MyErr set %s:%d\n", __FILE__, __LINE__){do { } while(0); }; |
1229 | *VRatioPrefetchC = 0; |
1230 | } |
1231 | } |
1232 | |
1233 | *RequiredPrefetchPixDataBWLuma = (double) PrefetchSourceLinesY / LinesToRequestPrefetchPixelData * BytePerPixelY * swath_width_luma_ub / LineTime; |
1234 | *RequiredPrefetchPixDataBWChroma = (double) PrefetchSourceLinesC / LinesToRequestPrefetchPixelData * BytePerPixelC * swath_width_chroma_ub / LineTime; |
1235 | } else { |
1236 | MyError = true1; |
1237 | dml_print("DML: MyErr set %s:%d\n", __FILE__, __LINE__){do { } while(0); }; |
1238 | dml_print("DML: LinesToRequestPrefetchPixelData: %f, should be > 0\n", LinesToRequestPrefetchPixelData){do { } while(0); }; |
1239 | *VRatioPrefetchY = 0; |
1240 | *VRatioPrefetchC = 0; |
1241 | *RequiredPrefetchPixDataBWLuma = 0; |
1242 | *RequiredPrefetchPixDataBWChroma = 0; |
1243 | } |
1244 | |
1245 | dml_print("DML: Tpre: %fus - sum of tim to request meta pte, 2 x data pte + meta data, swaths\n", (double)LinesToRequestPrefetchPixelData * LineTime + 2.0*TimeForFetchingRowInVBlank + TimeForFetchingMetaPTE){do { } while(0); }; |
1246 | dml_print("DML: Tvm: %fus - time to fetch page tables for meta surface\n", TimeForFetchingMetaPTE){do { } while(0); }; |
1247 | 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); }; |
1248 | dml_print("DML: Tr1: %fus - time to fetch second row of data pagetables and second row of meta data (done in parallel)\n", TimeForFetchingRowInVBlank){do { } while(0); }; |
1249 | dml_print("DML: Tsw: %fus = time to fetch enough pixel data and cursor data to feed the scalers init position and detile\n", (double)LinesToRequestPrefetchPixelData * LineTime){do { } while(0); }; |
1250 | dml_print("DML: To: %fus - time for propagation from scaler to optc\n", (*DSTYAfterScaler + ((*DSTXAfterScaler) / (double) myPipe->HTotal)) * LineTime){do { } while(0); }; |
1251 | dml_print("DML: Tvstartup - Tsetup - Tcalc - Twait - Tpre - To > 0\n"){do { } while(0); }; |
1252 | dml_print("DML: Tslack(pre): %fus - time left over in schedule\n", VStartup * LineTime - TimeForFetchingMetaPTE - 2 * TimeForFetchingRowInVBlank - (*DSTYAfterScaler + ((*DSTXAfterScaler) / (double) myPipe->HTotal)) * LineTime - TWait - TCalc - Tsetup){do { } while(0); }; |
1253 | dml_print("DML: row_bytes = dpte_row_bytes (per_pipe) = PixelPTEBytesPerRow = : %d\n", PixelPTEBytesPerRow){do { } while(0); }; |
1254 | |
1255 | } else { |
1256 | MyError = true1; |
1257 | dml_print("DML: MyErr set %s:%d\n", __FILE__, __LINE__){do { } while(0); }; |
1258 | } |
1259 | |
1260 | { |
1261 | double prefetch_vm_bw = 0; |
1262 | double prefetch_row_bw = 0; |
1263 | |
1264 | if (PDEAndMetaPTEBytesFrame == 0) { |
1265 | prefetch_vm_bw = 0; |
1266 | } else if (*DestinationLinesToRequestVMInVBlank > 0) { |
1267 | prefetch_vm_bw = PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor / (*DestinationLinesToRequestVMInVBlank * LineTime); |
1268 | } else { |
1269 | prefetch_vm_bw = 0; |
1270 | MyError = true1; |
1271 | dml_print("DML: MyErr set %s:%d\n", __FILE__, __LINE__){do { } while(0); }; |
1272 | } |
1273 | if (MetaRowByte + PixelPTEBytesPerRow == 0) { |
1274 | prefetch_row_bw = 0; |
1275 | } else if (*DestinationLinesToRequestRowInVBlank > 0) { |
1276 | prefetch_row_bw = (MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor) / (*DestinationLinesToRequestRowInVBlank * LineTime); |
1277 | } else { |
1278 | prefetch_row_bw = 0; |
1279 | MyError = true1; |
1280 | dml_print("DML: MyErr set %s:%d\n", __FILE__, __LINE__){do { } while(0); }; |
1281 | } |
1282 | |
1283 | *prefetch_vmrow_bw = dml_max(prefetch_vm_bw, prefetch_row_bw); |
1284 | } |
1285 | |
1286 | if (MyError) { |
1287 | *PrefetchBandwidth = 0; |
1288 | TimeForFetchingMetaPTE = 0; |
Value stored to 'TimeForFetchingMetaPTE' is never read | |
1289 | TimeForFetchingRowInVBlank = 0; |
1290 | *DestinationLinesToRequestVMInVBlank = 0; |
1291 | *DestinationLinesToRequestRowInVBlank = 0; |
1292 | *DestinationLinesForPrefetch = 0; |
1293 | LinesToRequestPrefetchPixelData = 0; |
1294 | *VRatioPrefetchY = 0; |
1295 | *VRatioPrefetchC = 0; |
1296 | *RequiredPrefetchPixDataBWLuma = 0; |
1297 | *RequiredPrefetchPixDataBWChroma = 0; |
1298 | } |
1299 | |
1300 | return MyError; |
1301 | } |
1302 | |
1303 | static double RoundToDFSGranularityUp(double Clock, double VCOSpeed) |
1304 | { |
1305 | return VCOSpeed * 4 / dml_floor(VCOSpeed * 4 / Clock, 1); |
1306 | } |
1307 | |
1308 | static double RoundToDFSGranularityDown(double Clock, double VCOSpeed) |
1309 | { |
1310 | return VCOSpeed * 4 / dml_ceil(VCOSpeed * 4.0 / Clock, 1); |
1311 | } |
1312 | |
1313 | static void CalculateDCCConfiguration( |
1314 | bool_Bool DCCEnabled, |
1315 | bool_Bool DCCProgrammingAssumesScanDirectionUnknown, |
1316 | enum source_format_class SourcePixelFormat, |
1317 | unsigned int SurfaceWidthLuma, |
1318 | unsigned int SurfaceWidthChroma, |
1319 | unsigned int SurfaceHeightLuma, |
1320 | unsigned int SurfaceHeightChroma, |
1321 | double DETBufferSize, |
1322 | unsigned int RequestHeight256ByteLuma, |
1323 | unsigned int RequestHeight256ByteChroma, |
1324 | enum dm_swizzle_mode TilingFormat, |
1325 | unsigned int BytePerPixelY, |
1326 | unsigned int BytePerPixelC, |
1327 | double BytePerPixelDETY, |
1328 | double BytePerPixelDETC, |
1329 | enum scan_direction_class ScanOrientation, |
1330 | unsigned int *MaxUncompressedBlockLuma, |
1331 | unsigned int *MaxUncompressedBlockChroma, |
1332 | unsigned int *MaxCompressedBlockLuma, |
1333 | unsigned int *MaxCompressedBlockChroma, |
1334 | unsigned int *IndependentBlockLuma, |
1335 | unsigned int *IndependentBlockChroma) |
1336 | { |
1337 | int yuv420 = 0; |
1338 | int horz_div_l = 0; |
1339 | int horz_div_c = 0; |
1340 | int vert_div_l = 0; |
1341 | int vert_div_c = 0; |
1342 | |
1343 | int req128_horz_wc_l = 0; |
1344 | int req128_horz_wc_c = 0; |
1345 | int req128_vert_wc_l = 0; |
1346 | int req128_vert_wc_c = 0; |
1347 | int segment_order_horz_contiguous_luma = 0; |
1348 | int segment_order_horz_contiguous_chroma = 0; |
1349 | int segment_order_vert_contiguous_luma = 0; |
1350 | int segment_order_vert_contiguous_chroma = 0; |
1351 | |
1352 | long full_swath_bytes_horz_wc_l = 0; |
1353 | long full_swath_bytes_horz_wc_c = 0; |
1354 | long full_swath_bytes_vert_wc_l = 0; |
1355 | long full_swath_bytes_vert_wc_c = 0; |
1356 | |
1357 | long swath_buf_size = 0; |
1358 | double detile_buf_vp_horz_limit = 0; |
1359 | double detile_buf_vp_vert_limit = 0; |
1360 | |
1361 | long MAS_vp_horz_limit = 0; |
1362 | long MAS_vp_vert_limit = 0; |
1363 | long max_vp_horz_width = 0; |
1364 | long max_vp_vert_height = 0; |
1365 | long eff_surf_width_l = 0; |
1366 | long eff_surf_width_c = 0; |
1367 | long eff_surf_height_l = 0; |
1368 | long eff_surf_height_c = 0; |
1369 | |
1370 | typedef enum { |
1371 | REQ_256Bytes, |
1372 | REQ_128BytesNonContiguous, |
1373 | REQ_128BytesContiguous, |
1374 | REQ_NA |
1375 | } RequestType; |
1376 | |
1377 | RequestType RequestLuma; |
1378 | RequestType RequestChroma; |
1379 | |
1380 | yuv420 = ((SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10 || SourcePixelFormat == dm_420_12) ? 1 : 0); |
1381 | horz_div_l = 1; |
1382 | horz_div_c = 1; |
1383 | vert_div_l = 1; |
1384 | vert_div_c = 1; |
1385 | |
1386 | if (BytePerPixelY == 1) |
1387 | vert_div_l = 0; |
1388 | if (BytePerPixelC == 1) |
1389 | vert_div_c = 0; |
1390 | if (BytePerPixelY == 8 |
1391 | && (TilingFormat == dm_sw_64kb_s || TilingFormat == dm_sw_64kb_s_t |
1392 | || TilingFormat == dm_sw_64kb_s_x)) |
1393 | horz_div_l = 0; |
1394 | if (BytePerPixelC == 8 |
1395 | && (TilingFormat == dm_sw_64kb_s || TilingFormat == dm_sw_64kb_s_t |
1396 | || TilingFormat == dm_sw_64kb_s_x)) |
1397 | horz_div_c = 0; |
1398 | |
1399 | if (BytePerPixelC == 0) { |
1400 | swath_buf_size = DETBufferSize / 2 - 2 * 256; |
1401 | detile_buf_vp_horz_limit = (double) swath_buf_size |
1402 | / ((double) RequestHeight256ByteLuma * BytePerPixelY |
1403 | / (1 + horz_div_l)); |
1404 | detile_buf_vp_vert_limit = (double) swath_buf_size |
1405 | / (256.0 / RequestHeight256ByteLuma / (1 + vert_div_l)); |
1406 | } else { |
1407 | swath_buf_size = DETBufferSize / 2 - 2 * 2 * 256; |
1408 | detile_buf_vp_horz_limit = (double) swath_buf_size |
1409 | / ((double) RequestHeight256ByteLuma * BytePerPixelY |
1410 | / (1 + horz_div_l) |
1411 | + (double) RequestHeight256ByteChroma |
1412 | * BytePerPixelC / (1 + horz_div_c) |
1413 | / (1 + yuv420)); |
1414 | detile_buf_vp_vert_limit = (double) swath_buf_size |
1415 | / (256.0 / RequestHeight256ByteLuma / (1 + vert_div_l) |
1416 | + 256.0 / RequestHeight256ByteChroma |
1417 | / (1 + vert_div_c) / (1 + yuv420)); |
1418 | } |
1419 | |
1420 | if (SourcePixelFormat == dm_420_10) { |
1421 | detile_buf_vp_horz_limit = 1.5 * detile_buf_vp_horz_limit; |
1422 | detile_buf_vp_vert_limit = 1.5 * detile_buf_vp_vert_limit; |
1423 | } |
1424 | |
1425 | detile_buf_vp_horz_limit = dml_floor(detile_buf_vp_horz_limit - 1, 16); |
1426 | detile_buf_vp_vert_limit = dml_floor(detile_buf_vp_vert_limit - 1, 16); |
1427 | |
1428 | MAS_vp_horz_limit = 5760; |
1429 | MAS_vp_vert_limit = (BytePerPixelC > 0 ? 2880 : 5760); |
1430 | max_vp_horz_width = dml_min((double) MAS_vp_horz_limit, detile_buf_vp_horz_limit); |
1431 | max_vp_vert_height = dml_min((double) MAS_vp_vert_limit, detile_buf_vp_vert_limit); |
1432 | eff_surf_width_l = |
1433 | (SurfaceWidthLuma > max_vp_horz_width ? max_vp_horz_width : SurfaceWidthLuma); |
1434 | eff_surf_width_c = eff_surf_width_l / (1 + yuv420); |
1435 | eff_surf_height_l = ( |
1436 | SurfaceHeightLuma > max_vp_vert_height ? |
1437 | max_vp_vert_height : SurfaceHeightLuma); |
1438 | eff_surf_height_c = eff_surf_height_l / (1 + yuv420); |
1439 | |
1440 | full_swath_bytes_horz_wc_l = eff_surf_width_l * RequestHeight256ByteLuma * BytePerPixelY; |
1441 | full_swath_bytes_vert_wc_l = eff_surf_height_l * 256 / RequestHeight256ByteLuma; |
1442 | if (BytePerPixelC > 0) { |
1443 | full_swath_bytes_horz_wc_c = eff_surf_width_c * RequestHeight256ByteChroma |
1444 | * BytePerPixelC; |
1445 | full_swath_bytes_vert_wc_c = eff_surf_height_c * 256 / RequestHeight256ByteChroma; |
1446 | } else { |
1447 | full_swath_bytes_horz_wc_c = 0; |
1448 | full_swath_bytes_vert_wc_c = 0; |
1449 | } |
1450 | |
1451 | if (SourcePixelFormat == dm_420_10) { |
1452 | full_swath_bytes_horz_wc_l = dml_ceil(full_swath_bytes_horz_wc_l * 2 / 3, 256); |
1453 | full_swath_bytes_horz_wc_c = dml_ceil(full_swath_bytes_horz_wc_c * 2 / 3, 256); |
1454 | full_swath_bytes_vert_wc_l = dml_ceil(full_swath_bytes_vert_wc_l * 2 / 3, 256); |
1455 | full_swath_bytes_vert_wc_c = dml_ceil(full_swath_bytes_vert_wc_c * 2 / 3, 256); |
1456 | } |
1457 | |
1458 | if (2 * full_swath_bytes_horz_wc_l + 2 * full_swath_bytes_horz_wc_c <= DETBufferSize) { |
1459 | req128_horz_wc_l = 0; |
1460 | req128_horz_wc_c = 0; |
1461 | } else if (full_swath_bytes_horz_wc_l < 1.5 * full_swath_bytes_horz_wc_c |
1462 | && 2 * full_swath_bytes_horz_wc_l + full_swath_bytes_horz_wc_c |
1463 | <= DETBufferSize) { |
1464 | req128_horz_wc_l = 0; |
1465 | req128_horz_wc_c = 1; |
1466 | } else if (full_swath_bytes_horz_wc_l >= 1.5 * full_swath_bytes_horz_wc_c |
1467 | && full_swath_bytes_horz_wc_l + 2 * full_swath_bytes_horz_wc_c |
1468 | <= DETBufferSize) { |
1469 | req128_horz_wc_l = 1; |
1470 | req128_horz_wc_c = 0; |
1471 | } else { |
1472 | req128_horz_wc_l = 1; |
1473 | req128_horz_wc_c = 1; |
1474 | } |
1475 | |
1476 | if (2 * full_swath_bytes_vert_wc_l + 2 * full_swath_bytes_vert_wc_c <= DETBufferSize) { |
1477 | req128_vert_wc_l = 0; |
1478 | req128_vert_wc_c = 0; |
1479 | } else if (full_swath_bytes_vert_wc_l < 1.5 * full_swath_bytes_vert_wc_c |
1480 | && 2 * full_swath_bytes_vert_wc_l + full_swath_bytes_vert_wc_c |
1481 | <= DETBufferSize) { |
1482 | req128_vert_wc_l = 0; |
1483 | req128_vert_wc_c = 1; |
1484 | } else if (full_swath_bytes_vert_wc_l >= 1.5 * full_swath_bytes_vert_wc_c |
1485 | && full_swath_bytes_vert_wc_l + 2 * full_swath_bytes_vert_wc_c |
1486 | <= DETBufferSize) { |
1487 | req128_vert_wc_l = 1; |
1488 | req128_vert_wc_c = 0; |
1489 | } else { |
1490 | req128_vert_wc_l = 1; |
1491 | req128_vert_wc_c = 1; |
1492 | } |
1493 | |
1494 | if (BytePerPixelY == 2 || (BytePerPixelY == 4 && TilingFormat != dm_sw_64kb_r_x)) { |
1495 | segment_order_horz_contiguous_luma = 0; |
1496 | } else { |
1497 | segment_order_horz_contiguous_luma = 1; |
1498 | } |
1499 | if ((BytePerPixelY == 8 |
1500 | && (TilingFormat == dm_sw_64kb_d || TilingFormat == dm_sw_64kb_d_x |
1501 | || TilingFormat == dm_sw_64kb_d_t |
1502 | || TilingFormat == dm_sw_64kb_r_x)) |
1503 | || (BytePerPixelY == 4 && TilingFormat == dm_sw_64kb_r_x)) { |
1504 | segment_order_vert_contiguous_luma = 0; |
1505 | } else { |
1506 | segment_order_vert_contiguous_luma = 1; |
1507 | } |
1508 | if (BytePerPixelC == 2 || (BytePerPixelC == 4 && TilingFormat != dm_sw_64kb_r_x)) { |
1509 | segment_order_horz_contiguous_chroma = 0; |
1510 | } else { |
1511 | segment_order_horz_contiguous_chroma = 1; |
1512 | } |
1513 | if ((BytePerPixelC == 8 |
1514 | && (TilingFormat == dm_sw_64kb_d || TilingFormat == dm_sw_64kb_d_x |
1515 | || TilingFormat == dm_sw_64kb_d_t |
1516 | || TilingFormat == dm_sw_64kb_r_x)) |
1517 | || (BytePerPixelC == 4 && TilingFormat == dm_sw_64kb_r_x)) { |
1518 | segment_order_vert_contiguous_chroma = 0; |
1519 | } else { |
1520 | segment_order_vert_contiguous_chroma = 1; |
1521 | } |
1522 | |
1523 | if (DCCProgrammingAssumesScanDirectionUnknown == true1) { |
1524 | if (req128_horz_wc_l == 0 && req128_vert_wc_l == 0) { |
1525 | RequestLuma = REQ_256Bytes; |
1526 | } else if ((req128_horz_wc_l == 1 && segment_order_horz_contiguous_luma == 0) |
1527 | || (req128_vert_wc_l == 1 && segment_order_vert_contiguous_luma == 0)) { |
1528 | RequestLuma = REQ_128BytesNonContiguous; |
1529 | } else { |
1530 | RequestLuma = REQ_128BytesContiguous; |
1531 | } |
1532 | if (req128_horz_wc_c == 0 && req128_vert_wc_c == 0) { |
1533 | RequestChroma = REQ_256Bytes; |
1534 | } else if ((req128_horz_wc_c == 1 && segment_order_horz_contiguous_chroma == 0) |
1535 | || (req128_vert_wc_c == 1 |
1536 | && segment_order_vert_contiguous_chroma == 0)) { |
1537 | RequestChroma = REQ_128BytesNonContiguous; |
1538 | } else { |
1539 | RequestChroma = REQ_128BytesContiguous; |
1540 | } |
1541 | } else if (ScanOrientation != dm_vert) { |
1542 | if (req128_horz_wc_l == 0) { |
1543 | RequestLuma = REQ_256Bytes; |
1544 | } else if (segment_order_horz_contiguous_luma == 0) { |
1545 | RequestLuma = REQ_128BytesNonContiguous; |
1546 | } else { |
1547 | RequestLuma = REQ_128BytesContiguous; |
1548 | } |
1549 | if (req128_horz_wc_c == 0) { |
1550 | RequestChroma = REQ_256Bytes; |
1551 | } else if (segment_order_horz_contiguous_chroma == 0) { |
1552 | RequestChroma = REQ_128BytesNonContiguous; |
1553 | } else { |
1554 | RequestChroma = REQ_128BytesContiguous; |
1555 | } |
1556 | } else { |
1557 | if (req128_vert_wc_l == 0) { |
1558 | RequestLuma = REQ_256Bytes; |
1559 | } else if (segment_order_vert_contiguous_luma == 0) { |
1560 | RequestLuma = REQ_128BytesNonContiguous; |
1561 | } else { |
1562 | RequestLuma = REQ_128BytesContiguous; |
1563 | } |
1564 | if (req128_vert_wc_c == 0) { |
1565 | RequestChroma = REQ_256Bytes; |
1566 | } else if (segment_order_vert_contiguous_chroma == 0) { |
1567 | RequestChroma = REQ_128BytesNonContiguous; |
1568 | } else { |
1569 | RequestChroma = REQ_128BytesContiguous; |
1570 | } |
1571 | } |
1572 | |
1573 | if (RequestLuma == REQ_256Bytes) { |
1574 | *MaxUncompressedBlockLuma = 256; |
1575 | *MaxCompressedBlockLuma = 256; |
1576 | *IndependentBlockLuma = 0; |
1577 | } else if (RequestLuma == REQ_128BytesContiguous) { |
1578 | *MaxUncompressedBlockLuma = 256; |
1579 | *MaxCompressedBlockLuma = 128; |
1580 | *IndependentBlockLuma = 128; |
1581 | } else { |
1582 | *MaxUncompressedBlockLuma = 256; |
1583 | *MaxCompressedBlockLuma = 64; |
1584 | *IndependentBlockLuma = 64; |
1585 | } |
1586 | |
1587 | if (RequestChroma == REQ_256Bytes) { |
1588 | *MaxUncompressedBlockChroma = 256; |
1589 | *MaxCompressedBlockChroma = 256; |
1590 | *IndependentBlockChroma = 0; |
1591 | } else if (RequestChroma == REQ_128BytesContiguous) { |
1592 | *MaxUncompressedBlockChroma = 256; |
1593 | *MaxCompressedBlockChroma = 128; |
1594 | *IndependentBlockChroma = 128; |
1595 | } else { |
1596 | *MaxUncompressedBlockChroma = 256; |
1597 | *MaxCompressedBlockChroma = 64; |
1598 | *IndependentBlockChroma = 64; |
1599 | } |
1600 | |
1601 | if (DCCEnabled != true1 || BytePerPixelC == 0) { |
1602 | *MaxUncompressedBlockChroma = 0; |
1603 | *MaxCompressedBlockChroma = 0; |
1604 | *IndependentBlockChroma = 0; |
1605 | } |
1606 | |
1607 | if (DCCEnabled != true1) { |
1608 | *MaxUncompressedBlockLuma = 0; |
1609 | *MaxCompressedBlockLuma = 0; |
1610 | *IndependentBlockLuma = 0; |
1611 | } |
1612 | } |
1613 | |
1614 | |
1615 | static double CalculatePrefetchSourceLines( |
1616 | struct display_mode_lib *mode_lib, |
1617 | double VRatio, |
1618 | double vtaps, |
1619 | bool_Bool Interlace, |
1620 | bool_Bool ProgressiveToInterlaceUnitInOPP, |
1621 | unsigned int SwathHeight, |
1622 | unsigned int ViewportYStart, |
1623 | double *VInitPreFill, |
1624 | unsigned int *MaxNumSwath) |
1625 | { |
1626 | unsigned int MaxPartialSwath = 0; |
1627 | |
1628 | if (ProgressiveToInterlaceUnitInOPP) |
1629 | *VInitPreFill = dml_floor((VRatio + vtaps + 1) / 2.0, 1); |
1630 | else |
1631 | *VInitPreFill = dml_floor((VRatio + vtaps + 1 + Interlace * 0.5 * VRatio) / 2.0, 1); |
1632 | |
1633 | if (!mode_lib->vba.IgnoreViewportPositioning) { |
1634 | |
1635 | *MaxNumSwath = dml_ceil((*VInitPreFill - 1.0) / SwathHeight, 1) + 1.0; |
1636 | |
1637 | if (*VInitPreFill > 1.0) |
1638 | MaxPartialSwath = (unsigned int) (*VInitPreFill - 2) % SwathHeight; |
1639 | else |
1640 | MaxPartialSwath = (unsigned int) (*VInitPreFill + SwathHeight - 2) |
1641 | % SwathHeight; |
1642 | MaxPartialSwath = dml_max(1U, MaxPartialSwath); |
1643 | |
1644 | } else { |
1645 | |
1646 | if (ViewportYStart != 0) |
1647 | dml_print({do { } while(0); } |
1648 | "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); }; |
1649 | |
1650 | *MaxNumSwath = dml_ceil(*VInitPreFill / SwathHeight, 1); |
1651 | |
1652 | if (*VInitPreFill > 1.0) |
1653 | MaxPartialSwath = (unsigned int) (*VInitPreFill - 1) % SwathHeight; |
1654 | else |
1655 | MaxPartialSwath = (unsigned int) (*VInitPreFill + SwathHeight - 1) |
1656 | % SwathHeight; |
1657 | } |
1658 | |
1659 | return *MaxNumSwath * SwathHeight + MaxPartialSwath; |
1660 | } |
1661 | |
1662 | static unsigned int CalculateVMAndRowBytes( |
1663 | struct display_mode_lib *mode_lib, |
1664 | bool_Bool DCCEnable, |
1665 | unsigned int BlockHeight256Bytes, |
1666 | unsigned int BlockWidth256Bytes, |
1667 | enum source_format_class SourcePixelFormat, |
1668 | unsigned int SurfaceTiling, |
1669 | unsigned int BytePerPixel, |
1670 | enum scan_direction_class ScanDirection, |
1671 | unsigned int SwathWidth, |
1672 | unsigned int ViewportHeight, |
1673 | bool_Bool GPUVMEnable, |
1674 | bool_Bool HostVMEnable, |
1675 | unsigned int HostVMMaxNonCachedPageTableLevels, |
1676 | unsigned int GPUVMMinPageSize, |
1677 | unsigned int HostVMMinPageSize, |
1678 | unsigned int PTEBufferSizeInRequests, |
1679 | unsigned int Pitch, |
1680 | unsigned int DCCMetaPitch, |
1681 | unsigned int *MacroTileWidth, |
1682 | unsigned int *MetaRowByte, |
1683 | unsigned int *PixelPTEBytesPerRow, |
1684 | bool_Bool *PTEBufferSizeNotExceeded, |
1685 | unsigned int *dpte_row_width_ub, |
1686 | unsigned int *dpte_row_height, |
1687 | unsigned int *MetaRequestWidth, |
1688 | unsigned int *MetaRequestHeight, |
1689 | unsigned int *meta_row_width, |
1690 | unsigned int *meta_row_height, |
1691 | unsigned int *vm_group_bytes, |
1692 | unsigned int *dpte_group_bytes, |
1693 | unsigned int *PixelPTEReqWidth, |
1694 | unsigned int *PixelPTEReqHeight, |
1695 | unsigned int *PTERequestSize, |
1696 | unsigned int *DPDE0BytesFrame, |
1697 | unsigned int *MetaPTEBytesFrame) |
1698 | { |
1699 | unsigned int MPDEBytesFrame = 0; |
1700 | unsigned int DCCMetaSurfaceBytes = 0; |
1701 | unsigned int MacroTileSizeBytes = 0; |
1702 | unsigned int MacroTileHeight = 0; |
1703 | unsigned int ExtraDPDEBytesFrame = 0; |
1704 | unsigned int PDEAndMetaPTEBytesFrame = 0; |
1705 | unsigned int PixelPTEReqHeightPTEs = 0; |
1706 | unsigned int HostVMDynamicLevels = 0; |
1707 | |
1708 | double FractionOfPTEReturnDrop; |
1709 | |
1710 | if (GPUVMEnable == true1 && HostVMEnable == true1) { |
1711 | if (HostVMMinPageSize < 2048) { |
1712 | HostVMDynamicLevels = HostVMMaxNonCachedPageTableLevels; |
1713 | } else if (HostVMMinPageSize >= 2048 && HostVMMinPageSize < 1048576) { |
1714 | HostVMDynamicLevels = dml_max(0, (int) HostVMMaxNonCachedPageTableLevels - 1); |
1715 | } else { |
1716 | HostVMDynamicLevels = dml_max(0, (int) HostVMMaxNonCachedPageTableLevels - 2); |
1717 | } |
1718 | } |
1719 | |
1720 | *MetaRequestHeight = 8 * BlockHeight256Bytes; |
1721 | *MetaRequestWidth = 8 * BlockWidth256Bytes; |
1722 | if (ScanDirection != dm_vert) { |
1723 | *meta_row_height = *MetaRequestHeight; |
1724 | *meta_row_width = dml_ceil((double) SwathWidth - 1, *MetaRequestWidth) |
1725 | + *MetaRequestWidth; |
1726 | *MetaRowByte = *meta_row_width * *MetaRequestHeight * BytePerPixel / 256.0; |
1727 | } else { |
1728 | *meta_row_height = *MetaRequestWidth; |
1729 | *meta_row_width = dml_ceil((double) SwathWidth - 1, *MetaRequestHeight) |
1730 | + *MetaRequestHeight; |
1731 | *MetaRowByte = *meta_row_width * *MetaRequestWidth * BytePerPixel / 256.0; |
1732 | } |
1733 | DCCMetaSurfaceBytes = DCCMetaPitch * (dml_ceil(ViewportHeight - 1, 64 * BlockHeight256Bytes) |
1734 | + 64 * BlockHeight256Bytes) * BytePerPixel / 256; |
1735 | if (GPUVMEnable == true1) { |
1736 | *MetaPTEBytesFrame = (dml_ceil((double) (DCCMetaSurfaceBytes - 4.0 * 1024.0) / (8 * 4.0 * 1024), 1) + 1) * 64; |
1737 | MPDEBytesFrame = 128 * (mode_lib->vba.GPUVMMaxPageTableLevels - 1); |
1738 | } else { |
1739 | *MetaPTEBytesFrame = 0; |
1740 | MPDEBytesFrame = 0; |
1741 | } |
1742 | |
1743 | if (DCCEnable != true1) { |
1744 | *MetaPTEBytesFrame = 0; |
1745 | MPDEBytesFrame = 0; |
1746 | *MetaRowByte = 0; |
1747 | } |
1748 | |
1749 | if (SurfaceTiling == dm_sw_linear) { |
1750 | MacroTileSizeBytes = 256; |
1751 | MacroTileHeight = BlockHeight256Bytes; |
1752 | } else { |
1753 | MacroTileSizeBytes = 65536; |
1754 | MacroTileHeight = 16 * BlockHeight256Bytes; |
1755 | } |
1756 | *MacroTileWidth = MacroTileSizeBytes / BytePerPixel / MacroTileHeight; |
1757 | |
1758 | if (GPUVMEnable == true1 && mode_lib->vba.GPUVMMaxPageTableLevels > 1) { |
1759 | if (ScanDirection != dm_vert) { |
1760 | *DPDE0BytesFrame = 64 * (dml_ceil(((Pitch * (dml_ceil(ViewportHeight - 1, MacroTileHeight) + MacroTileHeight) * BytePerPixel) - MacroTileSizeBytes) / (8 * 2097152), 1) + 1); |
1761 | } else { |
1762 | *DPDE0BytesFrame = 64 * (dml_ceil(((Pitch * (dml_ceil((double) SwathWidth - 1, MacroTileHeight) + MacroTileHeight) * BytePerPixel) - MacroTileSizeBytes) / (8 * 2097152), 1) + 1); |
1763 | } |
1764 | ExtraDPDEBytesFrame = 128 * (mode_lib->vba.GPUVMMaxPageTableLevels - 2); |
1765 | } else { |
1766 | *DPDE0BytesFrame = 0; |
1767 | ExtraDPDEBytesFrame = 0; |
1768 | } |
1769 | |
1770 | PDEAndMetaPTEBytesFrame = *MetaPTEBytesFrame + MPDEBytesFrame + *DPDE0BytesFrame |
1771 | + ExtraDPDEBytesFrame; |
1772 | |
1773 | if (HostVMEnable == true1) { |
1774 | PDEAndMetaPTEBytesFrame = PDEAndMetaPTEBytesFrame * (1 + 8 * HostVMDynamicLevels); |
1775 | } |
1776 | |
1777 | if (SurfaceTiling == dm_sw_linear) { |
1778 | PixelPTEReqHeightPTEs = 1; |
1779 | *PixelPTEReqHeight = 1; |
1780 | *PixelPTEReqWidth = 32768.0 / BytePerPixel; |
1781 | *PTERequestSize = 64; |
1782 | FractionOfPTEReturnDrop = 0; |
1783 | } else if (MacroTileSizeBytes == 4096) { |
1784 | PixelPTEReqHeightPTEs = 1; |
1785 | *PixelPTEReqHeight = MacroTileHeight; |
1786 | *PixelPTEReqWidth = 8 * *MacroTileWidth; |
1787 | *PTERequestSize = 64; |
1788 | if (ScanDirection != dm_vert) |
1789 | FractionOfPTEReturnDrop = 0; |
1790 | else |
1791 | FractionOfPTEReturnDrop = 7 / 8; |
1792 | } else if (GPUVMMinPageSize == 4 && MacroTileSizeBytes > 4096) { |
1793 | PixelPTEReqHeightPTEs = 16; |
1794 | *PixelPTEReqHeight = 16 * BlockHeight256Bytes; |
1795 | *PixelPTEReqWidth = 16 * BlockWidth256Bytes; |
1796 | *PTERequestSize = 128; |
1797 | FractionOfPTEReturnDrop = 0; |
1798 | } else { |
1799 | PixelPTEReqHeightPTEs = 1; |
1800 | *PixelPTEReqHeight = MacroTileHeight; |
1801 | *PixelPTEReqWidth = 8 * *MacroTileWidth; |
1802 | *PTERequestSize = 64; |
1803 | FractionOfPTEReturnDrop = 0; |
1804 | } |
1805 | |
1806 | if (SurfaceTiling == dm_sw_linear) { |
1807 | if (PTEBufferSizeInRequests == 0) |
1808 | *dpte_row_height = 1; |
1809 | else |
1810 | *dpte_row_height = dml_min(128, 1 << (unsigned int) dml_floor(dml_log2(PTEBufferSizeInRequests * *PixelPTEReqWidth / Pitch), 1)); |
1811 | *dpte_row_width_ub = (dml_ceil(((double) SwathWidth - 1) / *PixelPTEReqWidth, 1) + 1) * *PixelPTEReqWidth; |
1812 | *PixelPTEBytesPerRow = *dpte_row_width_ub / *PixelPTEReqWidth * *PTERequestSize; |
1813 | } else if (ScanDirection != dm_vert) { |
1814 | *dpte_row_height = *PixelPTEReqHeight; |
1815 | *dpte_row_width_ub = (dml_ceil((double) (SwathWidth - 1) / *PixelPTEReqWidth, 1) + 1) * *PixelPTEReqWidth; |
1816 | *PixelPTEBytesPerRow = *dpte_row_width_ub / *PixelPTEReqWidth * *PTERequestSize; |
1817 | } else { |
1818 | *dpte_row_height = dml_min(*PixelPTEReqWidth, *MacroTileWidth); |
1819 | *dpte_row_width_ub = (dml_ceil((double) (SwathWidth - 1) / *PixelPTEReqHeight, 1) + 1) * *PixelPTEReqHeight; |
1820 | *PixelPTEBytesPerRow = *dpte_row_width_ub / *PixelPTEReqHeight * *PTERequestSize; |
1821 | } |
1822 | if (*PixelPTEBytesPerRow * (1 - FractionOfPTEReturnDrop) |
1823 | <= 64 * PTEBufferSizeInRequests) { |
1824 | *PTEBufferSizeNotExceeded = true1; |
1825 | } else { |
1826 | *PTEBufferSizeNotExceeded = false0; |
1827 | } |
1828 | |
1829 | if (GPUVMEnable != true1) { |
1830 | *PixelPTEBytesPerRow = 0; |
1831 | *PTEBufferSizeNotExceeded = true1; |
1832 | } |
1833 | dml_print("DML: vm_bytes = meta_pte_bytes_per_frame (per_pipe) = MetaPTEBytesFrame = : %i\n", *MetaPTEBytesFrame){do { } while(0); }; |
1834 | |
1835 | if (HostVMEnable == true1) { |
1836 | *PixelPTEBytesPerRow = *PixelPTEBytesPerRow * (1 + 8 * HostVMDynamicLevels); |
1837 | } |
1838 | |
1839 | if (HostVMEnable == true1) { |
1840 | *vm_group_bytes = 512; |
1841 | *dpte_group_bytes = 512; |
1842 | } else if (GPUVMEnable == true1) { |
1843 | *vm_group_bytes = 2048; |
1844 | if (SurfaceTiling != dm_sw_linear && PixelPTEReqHeightPTEs == 1 && ScanDirection == dm_vert) { |
1845 | *dpte_group_bytes = 512; |
1846 | } else { |
1847 | *dpte_group_bytes = 2048; |
1848 | } |
1849 | } else { |
1850 | *vm_group_bytes = 0; |
1851 | *dpte_group_bytes = 0; |
1852 | } |
1853 | |
1854 | return PDEAndMetaPTEBytesFrame; |
1855 | } |
1856 | |
1857 | static void DISPCLKDPPCLKDCFCLKDeepSleepPrefetchParametersWatermarksAndPerformanceCalculation( |
1858 | struct display_mode_lib *mode_lib) |
1859 | { |
1860 | struct vba_vars_st *v = &mode_lib->vba; |
1861 | unsigned int j, k; |
1862 | long ReorderBytes = 0; |
1863 | unsigned int PrefetchMode = v->PrefetchModePerState[v->VoltageLevel][v->maxMpcComb]; |
1864 | double MaxTotalRDBandwidth = 0; |
1865 | double MaxTotalRDBandwidthNoUrgentBurst = 0; |
1866 | bool_Bool DestinationLineTimesForPrefetchLessThan2 = false0; |
1867 | bool_Bool VRatioPrefetchMoreThan4 = false0; |
1868 | double TWait; |
1869 | |
1870 | v->WritebackDISPCLK = 0.0; |
1871 | v->DISPCLKWithRamping = 0; |
1872 | v->DISPCLKWithoutRamping = 0; |
1873 | v->GlobalDPPCLK = 0.0; |
1874 | /* DAL custom code: need to update ReturnBW in case min dcfclk is overriden */ |
1875 | v->IdealSDPPortBandwidthPerState[v->VoltageLevel][v->maxMpcComb] = dml_min3( |
1876 | v->ReturnBusWidth * v->DCFCLK, |
1877 | v->DRAMSpeedPerState[v->VoltageLevel] * v->NumberOfChannels * v->DRAMChannelWidth, |
1878 | v->FabricClockPerState[v->VoltageLevel] * v->FabricDatapathToDCNDataReturn); |
1879 | if (v->HostVMEnable != true1) { |
1880 | v->ReturnBW = v->IdealSDPPortBandwidthPerState[v->VoltageLevel][v->maxMpcComb] * v->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelDataOnly / 100; |
1881 | } else { |
1882 | v->ReturnBW = v->IdealSDPPortBandwidthPerState[v->VoltageLevel][v->maxMpcComb] * v->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData / 100; |
1883 | } |
1884 | /* End DAL custom code */ |
1885 | |
1886 | // DISPCLK and DPPCLK Calculation |
1887 | // |
1888 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
1889 | if (v->WritebackEnable[k]) { |
1890 | v->WritebackDISPCLK = dml_max(v->WritebackDISPCLK, |
1891 | dml30_CalculateWriteBackDISPCLK( |
1892 | v->WritebackPixelFormat[k], |
1893 | v->PixelClock[k], |
1894 | v->WritebackHRatio[k], |
1895 | v->WritebackVRatio[k], |
1896 | v->WritebackHTaps[k], |
1897 | v->WritebackVTaps[k], |
1898 | v->WritebackSourceWidth[k], |
1899 | v->WritebackDestinationWidth[k], |
1900 | v->HTotal[k], |
1901 | v->WritebackLineBufferSize)); |
1902 | } |
1903 | } |
1904 | |
1905 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
1906 | if (v->HRatio[k] > 1) { |
1907 | v->PSCL_THROUGHPUT_LUMA[k] = dml_min(v->MaxDCHUBToPSCLThroughput, |
1908 | v->MaxPSCLToLBThroughput * v->HRatio[k] / dml_ceil(v->htaps[k] / 6.0, 1)); |
1909 | } else { |
1910 | v->PSCL_THROUGHPUT_LUMA[k] = dml_min( |
1911 | v->MaxDCHUBToPSCLThroughput, |
1912 | v->MaxPSCLToLBThroughput); |
1913 | } |
1914 | |
1915 | v->DPPCLKUsingSingleDPPLuma = v->PixelClock[k] |
1916 | * dml_max(v->vtaps[k] / 6.0 * dml_min(1.0, v->HRatio[k]), |
1917 | dml_max(v->HRatio[k] * v->VRatio[k] / v->PSCL_THROUGHPUT_LUMA[k], 1.0)); |
1918 | |
1919 | if ((v->htaps[k] > 6 || v->vtaps[k] > 6) |
1920 | && v->DPPCLKUsingSingleDPPLuma < 2 * v->PixelClock[k]) { |
1921 | v->DPPCLKUsingSingleDPPLuma = 2 * v->PixelClock[k]; |
1922 | } |
1923 | |
1924 | if ((v->SourcePixelFormat[k] != dm_420_8 |
1925 | && v->SourcePixelFormat[k] != dm_420_10 |
1926 | && v->SourcePixelFormat[k] != dm_420_12 |
1927 | && v->SourcePixelFormat[k] != dm_rgbe_alpha)) { |
1928 | v->PSCL_THROUGHPUT_CHROMA[k] = 0.0; |
1929 | v->DPPCLKUsingSingleDPP[k] = v->DPPCLKUsingSingleDPPLuma; |
1930 | } else { |
1931 | if (v->HRatioChroma[k] > 1) { |
1932 | v->PSCL_THROUGHPUT_CHROMA[k] = dml_min(v->MaxDCHUBToPSCLThroughput, |
1933 | v->MaxPSCLToLBThroughput * v->HRatioChroma[k] / dml_ceil(v->HTAPsChroma[k] / 6.0, 1.0)); |
1934 | } else { |
1935 | v->PSCL_THROUGHPUT_CHROMA[k] = dml_min( |
1936 | v->MaxDCHUBToPSCLThroughput, |
1937 | v->MaxPSCLToLBThroughput); |
1938 | } |
1939 | v->DPPCLKUsingSingleDPPChroma = v->PixelClock[k] |
1940 | * dml_max3(v->VTAPsChroma[k] / 6.0 * dml_min(1.0, v->HRatioChroma[k]), |
1941 | v->HRatioChroma[k] * v->VRatioChroma[k] / v->PSCL_THROUGHPUT_CHROMA[k], 1.0); |
1942 | |
1943 | if ((v->HTAPsChroma[k] > 6 || v->VTAPsChroma[k] > 6) |
1944 | && v->DPPCLKUsingSingleDPPChroma |
1945 | < 2 * v->PixelClock[k]) { |
1946 | v->DPPCLKUsingSingleDPPChroma = 2 |
1947 | * v->PixelClock[k]; |
1948 | } |
1949 | |
1950 | v->DPPCLKUsingSingleDPP[k] = dml_max( |
1951 | v->DPPCLKUsingSingleDPPLuma, |
1952 | v->DPPCLKUsingSingleDPPChroma); |
1953 | } |
1954 | } |
1955 | |
1956 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
1957 | if (v->BlendingAndTiming[k] != k) |
1958 | continue; |
1959 | if (v->ODMCombineEnabled[k] == dm_odm_combine_mode_4to1) { |
1960 | v->DISPCLKWithRamping = dml_max(v->DISPCLKWithRamping, |
1961 | v->PixelClock[k] / 4 * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100) |
1962 | * (1 + v->DISPCLKRampingMargin / 100)); |
1963 | v->DISPCLKWithoutRamping = dml_max(v->DISPCLKWithoutRamping, |
1964 | v->PixelClock[k] / 4 * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100)); |
1965 | } else if (v->ODMCombineEnabled[k] == dm_odm_combine_mode_2to1) { |
1966 | v->DISPCLKWithRamping = dml_max(v->DISPCLKWithRamping, |
1967 | v->PixelClock[k] / 2 * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100) |
1968 | * (1 + v->DISPCLKRampingMargin / 100)); |
1969 | v->DISPCLKWithoutRamping = dml_max(v->DISPCLKWithoutRamping, |
1970 | v->PixelClock[k] / 2 * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100)); |
1971 | } else { |
1972 | v->DISPCLKWithRamping = dml_max(v->DISPCLKWithRamping, |
1973 | v->PixelClock[k] * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100) |
1974 | * (1 + v->DISPCLKRampingMargin / 100)); |
1975 | v->DISPCLKWithoutRamping = dml_max(v->DISPCLKWithoutRamping, |
1976 | v->PixelClock[k] * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100)); |
1977 | } |
1978 | } |
1979 | |
1980 | v->DISPCLKWithRamping = dml_max( |
1981 | v->DISPCLKWithRamping, |
1982 | v->WritebackDISPCLK); |
1983 | v->DISPCLKWithoutRamping = dml_max( |
1984 | v->DISPCLKWithoutRamping, |
1985 | v->WritebackDISPCLK); |
1986 | |
1987 | 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/dcn30/display_mode_vba_30.c" , 1987); __warned = 1; } __builtin_expect(!!(__ret), 0); })) do {} while (0); } while (0); |
1988 | v->DISPCLKWithRampingRoundedToDFSGranularity = RoundToDFSGranularityUp( |
1989 | v->DISPCLKWithRamping, |
1990 | v->DISPCLKDPPCLKVCOSpeed); |
1991 | v->DISPCLKWithoutRampingRoundedToDFSGranularity = RoundToDFSGranularityUp( |
1992 | v->DISPCLKWithoutRamping, |
1993 | v->DISPCLKDPPCLKVCOSpeed); |
1994 | v->MaxDispclkRoundedToDFSGranularity = RoundToDFSGranularityDown( |
1995 | v->soc.clock_limits[mode_lib->soc.num_states - 1].dispclk_mhz, |
1996 | v->DISPCLKDPPCLKVCOSpeed); |
1997 | if (v->DISPCLKWithoutRampingRoundedToDFSGranularity |
1998 | > v->MaxDispclkRoundedToDFSGranularity) { |
1999 | v->DISPCLK_calculated = |
2000 | v->DISPCLKWithoutRampingRoundedToDFSGranularity; |
2001 | } else if (v->DISPCLKWithRampingRoundedToDFSGranularity |
2002 | > v->MaxDispclkRoundedToDFSGranularity) { |
2003 | v->DISPCLK_calculated = v->MaxDispclkRoundedToDFSGranularity; |
2004 | } else { |
2005 | v->DISPCLK_calculated = |
2006 | v->DISPCLKWithRampingRoundedToDFSGranularity; |
2007 | } |
2008 | v->DISPCLK = v->DISPCLK_calculated; |
2009 | DTRACE(" dispclk_mhz (calculated) = %f", v->DISPCLK_calculated){do { } while(0); }; |
2010 | |
2011 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
2012 | v->DPPCLK_calculated[k] = v->DPPCLKUsingSingleDPP[k] |
2013 | / v->DPPPerPlane[k] |
2014 | * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100); |
2015 | v->GlobalDPPCLK = dml_max( |
2016 | v->GlobalDPPCLK, |
2017 | v->DPPCLK_calculated[k]); |
2018 | } |
2019 | v->GlobalDPPCLK = RoundToDFSGranularityUp( |
2020 | v->GlobalDPPCLK, |
2021 | v->DISPCLKDPPCLKVCOSpeed); |
2022 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
2023 | v->DPPCLK_calculated[k] = v->GlobalDPPCLK / 255 |
2024 | * dml_ceil( |
2025 | v->DPPCLK_calculated[k] * 255.0 |
2026 | / v->GlobalDPPCLK, |
2027 | 1); |
2028 | DTRACE(" dppclk_mhz[%i] (calculated) = %f", k, v->DPPCLK_calculated[k]){do { } while(0); }; |
2029 | v->DPPCLK[k] = v->DPPCLK_calculated[k]; |
2030 | } |
2031 | |
2032 | // Urgent and B P-State/DRAM Clock Change Watermark |
2033 | DTRACE(" dcfclk_mhz = %f", v->DCFCLK){do { } while(0); }; |
2034 | DTRACE(" return_bus_bw = %f", v->ReturnBW){do { } while(0); }; |
2035 | |
2036 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
2037 | dml30_CalculateBytePerPixelAnd256BBlockSizes( |
2038 | v->SourcePixelFormat[k], |
2039 | v->SurfaceTiling[k], |
2040 | &v->BytePerPixelY[k], |
2041 | &v->BytePerPixelC[k], |
2042 | &v->BytePerPixelDETY[k], |
2043 | &v->BytePerPixelDETC[k], |
2044 | &v->BlockHeight256BytesY[k], |
2045 | &v->BlockHeight256BytesC[k], |
2046 | &v->BlockWidth256BytesY[k], |
2047 | &v->BlockWidth256BytesC[k]); |
2048 | } |
2049 | |
2050 | CalculateSwathWidth( |
2051 | false0, |
2052 | v->NumberOfActivePlanes, |
2053 | v->SourcePixelFormat, |
2054 | v->SourceScan, |
2055 | v->ViewportWidth, |
2056 | v->ViewportHeight, |
2057 | v->SurfaceWidthY, |
2058 | v->SurfaceWidthC, |
2059 | v->SurfaceHeightY, |
2060 | v->SurfaceHeightC, |
2061 | v->ODMCombineEnabled, |
2062 | v->BytePerPixelY, |
2063 | v->BytePerPixelC, |
2064 | v->BlockHeight256BytesY, |
2065 | v->BlockHeight256BytesC, |
2066 | v->BlockWidth256BytesY, |
2067 | v->BlockWidth256BytesC, |
2068 | v->BlendingAndTiming, |
2069 | v->HActive, |
2070 | v->HRatio, |
2071 | v->DPPPerPlane, |
2072 | v->SwathWidthSingleDPPY, |
2073 | v->SwathWidthSingleDPPC, |
2074 | v->SwathWidthY, |
2075 | v->SwathWidthC, |
2076 | v->dummyinteger3, |
2077 | v->dummyinteger4, |
2078 | v->swath_width_luma_ub, |
2079 | v->swath_width_chroma_ub); |
2080 | |
2081 | |
2082 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
2083 | v->ReadBandwidthPlaneLuma[k] = v->SwathWidthSingleDPPY[k] * v->BytePerPixelY[k] / (v->HTotal[k] / v->PixelClock[k]) * v->VRatio[k]; |
2084 | v->ReadBandwidthPlaneChroma[k] = v->SwathWidthSingleDPPC[k] * v->BytePerPixelC[k] / (v->HTotal[k] / v->PixelClock[k]) * v->VRatioChroma[k]; |
2085 | DTRACE("read_bw[%i] = %fBps", k, v->ReadBandwidthPlaneLuma[k] + v->ReadBandwidthPlaneChroma[k]){do { } while(0); }; |
2086 | } |
2087 | |
2088 | |
2089 | // DCFCLK Deep Sleep |
2090 | CalculateDCFCLKDeepSleep( |
2091 | mode_lib, |
2092 | v->NumberOfActivePlanes, |
2093 | v->BytePerPixelY, |
2094 | v->BytePerPixelC, |
2095 | v->VRatio, |
2096 | v->VRatioChroma, |
2097 | v->SwathWidthY, |
2098 | v->SwathWidthC, |
2099 | v->DPPPerPlane, |
2100 | v->HRatio, |
2101 | v->HRatioChroma, |
2102 | v->PixelClock, |
2103 | v->PSCL_THROUGHPUT_LUMA, |
2104 | v->PSCL_THROUGHPUT_CHROMA, |
2105 | v->DPPCLK, |
2106 | v->ReadBandwidthPlaneLuma, |
2107 | v->ReadBandwidthPlaneChroma, |
2108 | v->ReturnBusWidth, |
2109 | &v->DCFCLKDeepSleep); |
2110 | |
2111 | // DSCCLK |
2112 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
2113 | if ((v->BlendingAndTiming[k] != k) || !v->DSCEnabled[k]) { |
2114 | v->DSCCLK_calculated[k] = 0.0; |
2115 | } else { |
2116 | if (v->OutputFormat[k] == dm_420) |
2117 | v->DSCFormatFactor = 2; |
2118 | else if (v->OutputFormat[k] == dm_444) |
2119 | v->DSCFormatFactor = 1; |
2120 | else if (v->OutputFormat[k] == dm_n422) |
2121 | v->DSCFormatFactor = 2; |
2122 | else |
2123 | v->DSCFormatFactor = 1; |
2124 | if (v->ODMCombineEnabled[k] == dm_odm_combine_mode_4to1) |
2125 | v->DSCCLK_calculated[k] = v->PixelClockBackEnd[k] / 12 |
2126 | / v->DSCFormatFactor / (1 - v->DISPCLKDPPCLKDSCCLKDownSpreading / 100); |
2127 | else if (v->ODMCombineEnabled[k] == dm_odm_combine_mode_2to1) |
2128 | v->DSCCLK_calculated[k] = v->PixelClockBackEnd[k] / 6 |
2129 | / v->DSCFormatFactor / (1 - v->DISPCLKDPPCLKDSCCLKDownSpreading / 100); |
2130 | else |
2131 | v->DSCCLK_calculated[k] = v->PixelClockBackEnd[k] / 3 |
2132 | / v->DSCFormatFactor / (1 - v->DISPCLKDPPCLKDSCCLKDownSpreading / 100); |
2133 | } |
2134 | } |
2135 | |
2136 | // DSC Delay |
2137 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
2138 | double BPP = v->OutputBppPerState[k][v->VoltageLevel]; |
2139 | |
2140 | if (v->DSCEnabled[k] && BPP != 0) { |
2141 | if (v->ODMCombineEnabled[k] == dm_odm_combine_mode_disabled) { |
2142 | v->DSCDelay[k] = dscceComputeDelay(v->DSCInputBitPerComponent[k], |
2143 | BPP, |
2144 | dml_ceil((double) v->HActive[k] / v->NumberOfDSCSlices[k], 1), |
2145 | v->NumberOfDSCSlices[k], |
2146 | v->OutputFormat[k], |
2147 | v->Output[k]) |
2148 | + dscComputeDelay(v->OutputFormat[k], v->Output[k]); |
2149 | } else if (v->ODMCombineEnabled[k] == dm_odm_combine_mode_2to1) { |
2150 | v->DSCDelay[k] = 2 * dscceComputeDelay(v->DSCInputBitPerComponent[k], |
2151 | BPP, |
2152 | dml_ceil((double) v->HActive[k] / v->NumberOfDSCSlices[k], 1), |
2153 | v->NumberOfDSCSlices[k] / 2.0, |
2154 | v->OutputFormat[k], |
2155 | v->Output[k]) |
2156 | + dscComputeDelay(v->OutputFormat[k], v->Output[k]); |
2157 | } else { |
2158 | v->DSCDelay[k] = 4 * dscceComputeDelay(v->DSCInputBitPerComponent[k], |
2159 | BPP, |
2160 | dml_ceil((double) v->HActive[k] / v->NumberOfDSCSlices[k], 1), |
2161 | v->NumberOfDSCSlices[k] / 4.0, |
2162 | v->OutputFormat[k], |
2163 | v->Output[k]) |
2164 | + dscComputeDelay(v->OutputFormat[k], v->Output[k]); |
2165 | } |
2166 | v->DSCDelay[k] = v->DSCDelay[k] * v->PixelClock[k] / v->PixelClockBackEnd[k]; |
2167 | } else { |
2168 | v->DSCDelay[k] = 0; |
2169 | } |
2170 | } |
2171 | |
2172 | for (k = 0; k < v->NumberOfActivePlanes; ++k) |
2173 | for (j = 0; j < v->NumberOfActivePlanes; ++j) // NumberOfPlanes |
2174 | if (j != k && v->BlendingAndTiming[k] == j |
2175 | && v->DSCEnabled[j]) |
2176 | v->DSCDelay[k] = v->DSCDelay[j]; |
2177 | |
2178 | // Prefetch |
2179 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
2180 | unsigned int PDEAndMetaPTEBytesFrameY = 0; |
2181 | unsigned int PixelPTEBytesPerRowY = 0; |
2182 | unsigned int MetaRowByteY = 0; |
2183 | unsigned int MetaRowByteC = 0; |
2184 | unsigned int PDEAndMetaPTEBytesFrameC = 0; |
2185 | unsigned int PixelPTEBytesPerRowC = 0; |
2186 | bool_Bool PTEBufferSizeNotExceededY = 0; |
2187 | bool_Bool PTEBufferSizeNotExceededC = 0; |
2188 | |
2189 | |
2190 | if (v->SourcePixelFormat[k] == dm_420_8 || v->SourcePixelFormat[k] == dm_420_10 || v->SourcePixelFormat[k] == dm_420_12 || v->SourcePixelFormat[k] == dm_rgbe_alpha) { |
2191 | if ((v->SourcePixelFormat[k] == dm_420_10 || v->SourcePixelFormat[k] == dm_420_12) && v->SourceScan[k] != dm_vert) { |
2192 | v->PTEBufferSizeInRequestsForLuma = (v->PTEBufferSizeInRequestsLuma + v->PTEBufferSizeInRequestsChroma) / 2; |
2193 | v->PTEBufferSizeInRequestsForChroma = v->PTEBufferSizeInRequestsForLuma; |
2194 | } else { |
2195 | v->PTEBufferSizeInRequestsForLuma = v->PTEBufferSizeInRequestsLuma; |
2196 | v->PTEBufferSizeInRequestsForChroma = v->PTEBufferSizeInRequestsChroma; |
2197 | |
2198 | } |
2199 | PDEAndMetaPTEBytesFrameC = CalculateVMAndRowBytes( |
2200 | mode_lib, |
2201 | v->DCCEnable[k], |
2202 | v->BlockHeight256BytesC[k], |
2203 | v->BlockWidth256BytesC[k], |
2204 | v->SourcePixelFormat[k], |
2205 | v->SurfaceTiling[k], |
2206 | v->BytePerPixelC[k], |
2207 | v->SourceScan[k], |
2208 | v->SwathWidthC[k], |
2209 | v->ViewportHeightChroma[k], |
2210 | v->GPUVMEnable, |
2211 | v->HostVMEnable, |
2212 | v->HostVMMaxNonCachedPageTableLevels, |
2213 | v->GPUVMMinPageSize, |
2214 | v->HostVMMinPageSize, |
2215 | v->PTEBufferSizeInRequestsForChroma, |
2216 | v->PitchC[k], |
2217 | v->DCCMetaPitchC[k], |
2218 | &v->MacroTileWidthC[k], |
2219 | &MetaRowByteC, |
2220 | &PixelPTEBytesPerRowC, |
2221 | &PTEBufferSizeNotExceededC, |
2222 | &v->dpte_row_width_chroma_ub[k], |
2223 | &v->dpte_row_height_chroma[k], |
2224 | &v->meta_req_width_chroma[k], |
2225 | &v->meta_req_height_chroma[k], |
2226 | &v->meta_row_width_chroma[k], |
2227 | &v->meta_row_height_chroma[k], |
2228 | &v->dummyinteger1, |
2229 | &v->dummyinteger2, |
2230 | &v->PixelPTEReqWidthC[k], |
2231 | &v->PixelPTEReqHeightC[k], |
2232 | &v->PTERequestSizeC[k], |
2233 | &v->dpde0_bytes_per_frame_ub_c[k], |
2234 | &v->meta_pte_bytes_per_frame_ub_c[k]); |
2235 | |
2236 | v->PrefetchSourceLinesC[k] = CalculatePrefetchSourceLines( |
2237 | mode_lib, |
2238 | v->VRatioChroma[k], |
2239 | v->VTAPsChroma[k], |
2240 | v->Interlace[k], |
2241 | v->ProgressiveToInterlaceUnitInOPP, |
2242 | v->SwathHeightC[k], |
2243 | v->ViewportYStartC[k], |
2244 | &v->VInitPreFillC[k], |
2245 | &v->MaxNumSwathC[k]); |
2246 | } else { |
2247 | v->PTEBufferSizeInRequestsForLuma = v->PTEBufferSizeInRequestsLuma + v->PTEBufferSizeInRequestsChroma; |
2248 | v->PTEBufferSizeInRequestsForChroma = 0; |
2249 | PixelPTEBytesPerRowC = 0; |
2250 | PDEAndMetaPTEBytesFrameC = 0; |
2251 | MetaRowByteC = 0; |
2252 | v->MaxNumSwathC[k] = 0; |
2253 | v->PrefetchSourceLinesC[k] = 0; |
2254 | } |
2255 | |
2256 | PDEAndMetaPTEBytesFrameY = CalculateVMAndRowBytes( |
2257 | mode_lib, |
2258 | v->DCCEnable[k], |
2259 | v->BlockHeight256BytesY[k], |
2260 | v->BlockWidth256BytesY[k], |
2261 | v->SourcePixelFormat[k], |
2262 | v->SurfaceTiling[k], |
2263 | v->BytePerPixelY[k], |
2264 | v->SourceScan[k], |
2265 | v->SwathWidthY[k], |
2266 | v->ViewportHeight[k], |
2267 | v->GPUVMEnable, |
2268 | v->HostVMEnable, |
2269 | v->HostVMMaxNonCachedPageTableLevels, |
2270 | v->GPUVMMinPageSize, |
2271 | v->HostVMMinPageSize, |
2272 | v->PTEBufferSizeInRequestsForLuma, |
2273 | v->PitchY[k], |
2274 | v->DCCMetaPitchY[k], |
2275 | &v->MacroTileWidthY[k], |
2276 | &MetaRowByteY, |
2277 | &PixelPTEBytesPerRowY, |
2278 | &PTEBufferSizeNotExceededY, |
2279 | &v->dpte_row_width_luma_ub[k], |
2280 | &v->dpte_row_height[k], |
2281 | &v->meta_req_width[k], |
2282 | &v->meta_req_height[k], |
2283 | &v->meta_row_width[k], |
2284 | &v->meta_row_height[k], |
2285 | &v->vm_group_bytes[k], |
2286 | &v->dpte_group_bytes[k], |
2287 | &v->PixelPTEReqWidthY[k], |
2288 | &v->PixelPTEReqHeightY[k], |
2289 | &v->PTERequestSizeY[k], |
2290 | &v->dpde0_bytes_per_frame_ub_l[k], |
2291 | &v->meta_pte_bytes_per_frame_ub_l[k]); |
2292 | |
2293 | v->PrefetchSourceLinesY[k] = CalculatePrefetchSourceLines( |
2294 | mode_lib, |
2295 | v->VRatio[k], |
2296 | v->vtaps[k], |
2297 | v->Interlace[k], |
2298 | v->ProgressiveToInterlaceUnitInOPP, |
2299 | v->SwathHeightY[k], |
2300 | v->ViewportYStartY[k], |
2301 | &v->VInitPreFillY[k], |
2302 | &v->MaxNumSwathY[k]); |
2303 | v->PixelPTEBytesPerRow[k] = PixelPTEBytesPerRowY + PixelPTEBytesPerRowC; |
2304 | v->PDEAndMetaPTEBytesFrame[k] = PDEAndMetaPTEBytesFrameY |
2305 | + PDEAndMetaPTEBytesFrameC; |
2306 | v->MetaRowByte[k] = MetaRowByteY + MetaRowByteC; |
2307 | |
2308 | CalculateRowBandwidth( |
2309 | v->GPUVMEnable, |
2310 | v->SourcePixelFormat[k], |
2311 | v->VRatio[k], |
2312 | v->VRatioChroma[k], |
2313 | v->DCCEnable[k], |
2314 | v->HTotal[k] / v->PixelClock[k], |
2315 | MetaRowByteY, |
2316 | MetaRowByteC, |
2317 | v->meta_row_height[k], |
2318 | v->meta_row_height_chroma[k], |
2319 | PixelPTEBytesPerRowY, |
2320 | PixelPTEBytesPerRowC, |
2321 | v->dpte_row_height[k], |
2322 | v->dpte_row_height_chroma[k], |
2323 | &v->meta_row_bw[k], |
2324 | &v->dpte_row_bw[k]); |
2325 | } |
2326 | |
2327 | v->TotalDCCActiveDPP = 0; |
2328 | v->TotalActiveDPP = 0; |
2329 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
2330 | v->TotalActiveDPP = v->TotalActiveDPP |
2331 | + v->DPPPerPlane[k]; |
2332 | if (v->DCCEnable[k]) |
2333 | v->TotalDCCActiveDPP = v->TotalDCCActiveDPP |
2334 | + v->DPPPerPlane[k]; |
2335 | } |
2336 | |
2337 | |
2338 | ReorderBytes = v->NumberOfChannels * dml_max3( |
2339 | v->UrgentOutOfOrderReturnPerChannelPixelDataOnly, |
2340 | v->UrgentOutOfOrderReturnPerChannelPixelMixedWithVMData, |
2341 | v->UrgentOutOfOrderReturnPerChannelVMDataOnly); |
2342 | |
2343 | v->UrgentExtraLatency = CalculateExtraLatency( |
2344 | v->RoundTripPingLatencyCycles, |
2345 | ReorderBytes, |
2346 | v->DCFCLK, |
2347 | v->TotalActiveDPP, |
2348 | v->PixelChunkSizeInKByte, |
2349 | v->TotalDCCActiveDPP, |
2350 | v->MetaChunkSize, |
2351 | v->ReturnBW, |
2352 | v->GPUVMEnable, |
2353 | v->HostVMEnable, |
2354 | v->NumberOfActivePlanes, |
2355 | v->DPPPerPlane, |
2356 | v->dpte_group_bytes, |
2357 | v->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
2358 | v->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
2359 | v->HostVMMinPageSize, |
2360 | v->HostVMMaxNonCachedPageTableLevels); |
2361 | |
2362 | v->TCalc = 24.0 / v->DCFCLKDeepSleep; |
2363 | |
2364 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
2365 | if (v->BlendingAndTiming[k] == k) { |
2366 | if (v->WritebackEnable[k] == true1) { |
2367 | v->WritebackDelay[v->VoltageLevel][k] = v->WritebackLatency + |
2368 | CalculateWriteBackDelay(v->WritebackPixelFormat[k], |
2369 | v->WritebackHRatio[k], |
2370 | v->WritebackVRatio[k], |
2371 | v->WritebackVTaps[k], |
2372 | v->WritebackDestinationWidth[k], |
2373 | v->WritebackDestinationHeight[k], |
2374 | v->WritebackSourceHeight[k], |
2375 | v->HTotal[k]) / v->DISPCLK; |
2376 | } else |
2377 | v->WritebackDelay[v->VoltageLevel][k] = 0; |
2378 | for (j = 0; j < v->NumberOfActivePlanes; ++j) { |
2379 | if (v->BlendingAndTiming[j] == k |
2380 | && v->WritebackEnable[j] == true1) { |
2381 | v->WritebackDelay[v->VoltageLevel][k] = dml_max(v->WritebackDelay[v->VoltageLevel][k], |
2382 | v->WritebackLatency + CalculateWriteBackDelay( |
2383 | v->WritebackPixelFormat[j], |
2384 | v->WritebackHRatio[j], |
2385 | v->WritebackVRatio[j], |
2386 | v->WritebackVTaps[j], |
2387 | v->WritebackDestinationWidth[j], |
2388 | v->WritebackDestinationHeight[j], |
2389 | v->WritebackSourceHeight[j], |
2390 | v->HTotal[k]) / v->DISPCLK); |
2391 | } |
2392 | } |
2393 | } |
2394 | } |
2395 | |
2396 | for (k = 0; k < v->NumberOfActivePlanes; ++k) |
2397 | for (j = 0; j < v->NumberOfActivePlanes; ++j) |
2398 | if (v->BlendingAndTiming[k] == j) |
2399 | v->WritebackDelay[v->VoltageLevel][k] = v->WritebackDelay[v->VoltageLevel][j]; |
2400 | |
2401 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
2402 | v->MaxVStartupLines[k] = v->VTotal[k] - v->VActive[k] - dml_max(1.0, dml_ceil((double) v->WritebackDelay[v->VoltageLevel][k] / (v->HTotal[k] / v->PixelClock[k]), 1)); |
2403 | } |
2404 | |
2405 | v->MaximumMaxVStartupLines = 0; |
2406 | for (k = 0; k < v->NumberOfActivePlanes; ++k) |
2407 | v->MaximumMaxVStartupLines = dml_max(v->MaximumMaxVStartupLines, v->MaxVStartupLines[k]); |
2408 | |
2409 | if (v->DRAMClockChangeLatencyOverride > 0.0) { |
2410 | v->FinalDRAMClockChangeLatency = v->DRAMClockChangeLatencyOverride; |
2411 | } else { |
2412 | v->FinalDRAMClockChangeLatency = v->DRAMClockChangeLatency; |
2413 | } |
2414 | v->UrgentLatency = CalculateUrgentLatency(v->UrgentLatencyPixelDataOnly, v->UrgentLatencyPixelMixedWithVMData, v->UrgentLatencyVMDataOnly, v->DoUrgentLatencyAdjustment, v->UrgentLatencyAdjustmentFabricClockComponent, v->UrgentLatencyAdjustmentFabricClockReference, v->FabricClock); |
2415 | |
2416 | |
2417 | v->FractionOfUrgentBandwidth = 0.0; |
2418 | v->FractionOfUrgentBandwidthImmediateFlip = 0.0; |
2419 | |
2420 | v->VStartupLines = 13; |
2421 | |
2422 | do { |
2423 | MaxTotalRDBandwidth = 0; |
2424 | MaxTotalRDBandwidthNoUrgentBurst = 0; |
2425 | DestinationLineTimesForPrefetchLessThan2 = false0; |
2426 | VRatioPrefetchMoreThan4 = false0; |
2427 | TWait = CalculateTWait( |
2428 | PrefetchMode, |
2429 | v->FinalDRAMClockChangeLatency, |
2430 | v->UrgentLatency, |
2431 | v->SREnterPlusExitTime); |
2432 | |
2433 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
2434 | Pipe myPipe = { 0 }; |
2435 | |
2436 | myPipe.DPPCLK = v->DPPCLK[k]; |
2437 | myPipe.DISPCLK = v->DISPCLK; |
2438 | myPipe.PixelClock = v->PixelClock[k]; |
2439 | myPipe.DCFCLKDeepSleep = v->DCFCLKDeepSleep; |
2440 | myPipe.DPPPerPlane = v->DPPPerPlane[k]; |
2441 | myPipe.ScalerEnabled = v->ScalerEnabled[k]; |
2442 | myPipe.SourceScan = v->SourceScan[k]; |
2443 | myPipe.BlockWidth256BytesY = v->BlockWidth256BytesY[k]; |
2444 | myPipe.BlockHeight256BytesY = v->BlockHeight256BytesY[k]; |
2445 | myPipe.BlockWidth256BytesC = v->BlockWidth256BytesC[k]; |
2446 | myPipe.BlockHeight256BytesC = v->BlockHeight256BytesC[k]; |
2447 | myPipe.InterlaceEnable = v->Interlace[k]; |
2448 | myPipe.NumberOfCursors = v->NumberOfCursors[k]; |
2449 | myPipe.VBlank = v->VTotal[k] - v->VActive[k]; |
2450 | myPipe.HTotal = v->HTotal[k]; |
2451 | myPipe.DCCEnable = v->DCCEnable[k]; |
2452 | myPipe.ODMCombineEnabled = !!v->ODMCombineEnabled[k]; |
2453 | |
2454 | v->ErrorResult[k] = CalculatePrefetchSchedule( |
2455 | mode_lib, |
2456 | v->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
2457 | v->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
2458 | &myPipe, |
2459 | v->DSCDelay[k], |
2460 | v->DPPCLKDelaySubtotal |
2461 | + v->DPPCLKDelayCNVCFormater, |
2462 | v->DPPCLKDelaySCL, |
2463 | v->DPPCLKDelaySCLLBOnly, |
2464 | v->DPPCLKDelayCNVCCursor, |
2465 | v->DISPCLKDelaySubtotal, |
2466 | (unsigned int) (v->SwathWidthY[k] / v->HRatio[k]), |
2467 | v->OutputFormat[k], |
2468 | v->MaxInterDCNTileRepeaters, |
2469 | dml_min(v->VStartupLines, v->MaxVStartupLines[k]), |
2470 | v->MaxVStartupLines[k], |
2471 | v->GPUVMMaxPageTableLevels, |
2472 | v->GPUVMEnable, |
2473 | v->HostVMEnable, |
2474 | v->HostVMMaxNonCachedPageTableLevels, |
2475 | v->HostVMMinPageSize, |
2476 | v->DynamicMetadataEnable[k], |
2477 | v->DynamicMetadataVMEnabled, |
2478 | v->DynamicMetadataLinesBeforeActiveRequired[k], |
2479 | v->DynamicMetadataTransmittedBytes[k], |
2480 | v->UrgentLatency, |
2481 | v->UrgentExtraLatency, |
2482 | v->TCalc, |
2483 | v->PDEAndMetaPTEBytesFrame[k], |
2484 | v->MetaRowByte[k], |
2485 | v->PixelPTEBytesPerRow[k], |
2486 | v->PrefetchSourceLinesY[k], |
2487 | v->SwathWidthY[k], |
2488 | v->BytePerPixelY[k], |
2489 | v->VInitPreFillY[k], |
2490 | v->MaxNumSwathY[k], |
2491 | v->PrefetchSourceLinesC[k], |
2492 | v->SwathWidthC[k], |
2493 | v->BytePerPixelC[k], |
2494 | v->VInitPreFillC[k], |
2495 | v->MaxNumSwathC[k], |
2496 | v->swath_width_luma_ub[k], |
2497 | v->swath_width_chroma_ub[k], |
2498 | v->SwathHeightY[k], |
2499 | v->SwathHeightC[k], |
2500 | TWait, |
2501 | v->ProgressiveToInterlaceUnitInOPP, |
2502 | &v->DSTXAfterScaler[k], |
2503 | &v->DSTYAfterScaler[k], |
2504 | &v->DestinationLinesForPrefetch[k], |
2505 | &v->PrefetchBandwidth[k], |
2506 | &v->DestinationLinesToRequestVMInVBlank[k], |
2507 | &v->DestinationLinesToRequestRowInVBlank[k], |
2508 | &v->VRatioPrefetchY[k], |
2509 | &v->VRatioPrefetchC[k], |
2510 | &v->RequiredPrefetchPixDataBWLuma[k], |
2511 | &v->RequiredPrefetchPixDataBWChroma[k], |
2512 | &v->NotEnoughTimeForDynamicMetadata[k], |
2513 | &v->Tno_bw[k], |
2514 | &v->prefetch_vmrow_bw[k], |
2515 | &v->Tdmdl_vm[k], |
2516 | &v->Tdmdl[k], |
2517 | &v->VUpdateOffsetPix[k], |
2518 | &v->VUpdateWidthPix[k], |
2519 | &v->VReadyOffsetPix[k]); |
2520 | if (v->BlendingAndTiming[k] == k) { |
2521 | double TotalRepeaterDelayTime = v->MaxInterDCNTileRepeaters * (2 / v->DPPCLK[k] + 3 / v->DISPCLK); |
2522 | v->VUpdateWidthPix[k] = (14 / v->DCFCLKDeepSleep + 12 / v->DPPCLK[k] + TotalRepeaterDelayTime) * v->PixelClock[k]; |
2523 | v->VReadyOffsetPix[k] = dml_max(150.0 / v->DPPCLK[k], TotalRepeaterDelayTime + 20 / v->DCFCLKDeepSleep + 10 / v->DPPCLK[k]) * v->PixelClock[k]; |
2524 | v->VUpdateOffsetPix[k] = dml_ceil(v->HTotal[k] / 4.0, 1); |
2525 | v->VStartup[k] = dml_min(v->VStartupLines, v->MaxVStartupLines[k]); |
2526 | } else { |
2527 | int x = v->BlendingAndTiming[k]; |
2528 | double TotalRepeaterDelayTime = v->MaxInterDCNTileRepeaters * (2 / v->DPPCLK[k] + 3 / v->DISPCLK); |
2529 | v->VUpdateWidthPix[k] = (14 / v->DCFCLKDeepSleep + 12 / v->DPPCLK[k] + TotalRepeaterDelayTime) * v->PixelClock[x]; |
2530 | v->VReadyOffsetPix[k] = dml_max(150.0 / v->DPPCLK[k], TotalRepeaterDelayTime + 20 / v->DCFCLKDeepSleep + 10 / v->DPPCLK[k]) * v->PixelClock[x]; |
2531 | v->VUpdateOffsetPix[k] = dml_ceil(v->HTotal[x] / 4.0, 1); |
2532 | if (!v->MaxVStartupLines[x]) |
2533 | v->MaxVStartupLines[x] = v->MaxVStartupLines[k]; |
2534 | v->VStartup[k] = dml_min(v->VStartupLines, v->MaxVStartupLines[x]); |
2535 | } |
2536 | } |
2537 | |
2538 | v->NotEnoughUrgentLatencyHiding[0][0] = false0; |
2539 | v->NotEnoughUrgentLatencyHidingPre = false0; |
2540 | |
2541 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
2542 | v->cursor_bw[k] = v->NumberOfCursors[k] |
2543 | * v->CursorWidth[k][0] * v->CursorBPP[k][0] |
2544 | / 8.0 |
2545 | / (v->HTotal[k] / v->PixelClock[k]) |
2546 | * v->VRatio[k]; |
2547 | v->cursor_bw_pre[k] = v->NumberOfCursors[k] |
2548 | * v->CursorWidth[k][0] * v->CursorBPP[k][0] |
2549 | / 8.0 |
2550 | / (v->HTotal[k] / v->PixelClock[k]) |
2551 | * v->VRatioPrefetchY[k]; |
2552 | |
2553 | CalculateUrgentBurstFactor( |
2554 | v->swath_width_luma_ub[k], |
2555 | v->swath_width_chroma_ub[k], |
2556 | v->DETBufferSizeInKByte[0], |
2557 | v->SwathHeightY[k], |
2558 | v->SwathHeightC[k], |
2559 | v->HTotal[k] / v->PixelClock[k], |
2560 | v->UrgentLatency, |
2561 | v->CursorBufferSize, |
2562 | v->CursorWidth[k][0], |
2563 | v->CursorBPP[k][0], |
2564 | v->VRatio[k], |
2565 | v->VRatioChroma[k], |
2566 | v->BytePerPixelDETY[k], |
2567 | v->BytePerPixelDETC[k], |
2568 | v->DETBufferSizeY[k], |
2569 | v->DETBufferSizeC[k], |
2570 | &v->UrgentBurstFactorCursor[k], |
2571 | &v->UrgentBurstFactorLuma[k], |
2572 | &v->UrgentBurstFactorChroma[k], |
2573 | &v->NoUrgentLatencyHiding[k]); |
2574 | |
2575 | CalculateUrgentBurstFactor( |
2576 | v->swath_width_luma_ub[k], |
2577 | v->swath_width_chroma_ub[k], |
2578 | v->DETBufferSizeInKByte[0], |
2579 | v->SwathHeightY[k], |
2580 | v->SwathHeightC[k], |
2581 | v->HTotal[k] / v->PixelClock[k], |
2582 | v->UrgentLatency, |
2583 | v->CursorBufferSize, |
2584 | v->CursorWidth[k][0], |
2585 | v->CursorBPP[k][0], |
2586 | v->VRatioPrefetchY[k], |
2587 | v->VRatioPrefetchC[k], |
2588 | v->BytePerPixelDETY[k], |
2589 | v->BytePerPixelDETC[k], |
2590 | v->DETBufferSizeY[k], |
2591 | v->DETBufferSizeC[k], |
2592 | &v->UrgentBurstFactorCursorPre[k], |
2593 | &v->UrgentBurstFactorLumaPre[k], |
2594 | &v->UrgentBurstFactorChromaPre[k], |
2595 | &v->NoUrgentLatencyHidingPre[k]); |
2596 | |
2597 | MaxTotalRDBandwidth = MaxTotalRDBandwidth + |
2598 | dml_max3(v->DPPPerPlane[k] * v->prefetch_vmrow_bw[k], |
2599 | v->ReadBandwidthPlaneLuma[k] * |
2600 | v->UrgentBurstFactorLuma[k] + |
2601 | v->ReadBandwidthPlaneChroma[k] * |
2602 | v->UrgentBurstFactorChroma[k] + |
2603 | v->cursor_bw[k] * |
2604 | v->UrgentBurstFactorCursor[k] + |
2605 | v->DPPPerPlane[k] * (v->meta_row_bw[k] + v->dpte_row_bw[k]), |
2606 | v->DPPPerPlane[k] * (v->RequiredPrefetchPixDataBWLuma[k] * v->UrgentBurstFactorLumaPre[k] + |
2607 | v->RequiredPrefetchPixDataBWChroma[k] * v->UrgentBurstFactorChromaPre[k]) + v->cursor_bw_pre[k] * |
2608 | v->UrgentBurstFactorCursorPre[k]); |
2609 | |
2610 | MaxTotalRDBandwidthNoUrgentBurst = MaxTotalRDBandwidthNoUrgentBurst + |
2611 | dml_max3(v->DPPPerPlane[k] * v->prefetch_vmrow_bw[k], |
2612 | v->ReadBandwidthPlaneLuma[k] + |
2613 | v->ReadBandwidthPlaneChroma[k] + |
2614 | v->cursor_bw[k] + |
2615 | v->DPPPerPlane[k] * (v->meta_row_bw[k] + v->dpte_row_bw[k]), |
2616 | v->DPPPerPlane[k] * (v->RequiredPrefetchPixDataBWLuma[k] + v->RequiredPrefetchPixDataBWChroma[k]) + v->cursor_bw_pre[k]); |
2617 | |
2618 | if (v->DestinationLinesForPrefetch[k] < 2) |
2619 | DestinationLineTimesForPrefetchLessThan2 = true1; |
2620 | if (v->VRatioPrefetchY[k] > 4 || v->VRatioPrefetchC[k] > 4) |
2621 | VRatioPrefetchMoreThan4 = true1; |
2622 | if (v->NoUrgentLatencyHiding[k] == true1) |
2623 | v->NotEnoughUrgentLatencyHiding[0][0] = true1; |
2624 | |
2625 | if (v->NoUrgentLatencyHidingPre[k] == true1) |
2626 | v->NotEnoughUrgentLatencyHidingPre = true1; |
2627 | } |
2628 | v->FractionOfUrgentBandwidth = MaxTotalRDBandwidthNoUrgentBurst / v->ReturnBW; |
2629 | |
2630 | |
2631 | if (MaxTotalRDBandwidth <= v->ReturnBW && v->NotEnoughUrgentLatencyHiding[0][0] == 0 |
2632 | && v->NotEnoughUrgentLatencyHidingPre == 0 && !VRatioPrefetchMoreThan4 |
2633 | && !DestinationLineTimesForPrefetchLessThan2) |
2634 | v->PrefetchModeSupported = true1; |
2635 | else { |
2636 | v->PrefetchModeSupported = false0; |
2637 | dml_print("DML: CalculatePrefetchSchedule ***failed***. Bandwidth violation. Results are NOT valid\n"){do { } while(0); }; |
2638 | dml_print("DML: MaxTotalRDBandwidth:%f AvailReturnBandwidth:%f\n", MaxTotalRDBandwidth, v->ReturnBW){do { } while(0); }; |
2639 | dml_print("DML: VRatioPrefetch %s more than 4\n", (VRatioPrefetchMoreThan4) ? "is" : "is not"){do { } while(0); }; |
2640 | dml_print("DML: DestinationLines for Prefetch %s less than 2\n", (DestinationLineTimesForPrefetchLessThan2) ? "is" : "is not"){do { } while(0); }; |
2641 | } |
2642 | |
2643 | if (v->PrefetchModeSupported == true1 && v->ImmediateFlipSupport == true1) { |
2644 | v->BandwidthAvailableForImmediateFlip = v->ReturnBW; |
2645 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
2646 | v->BandwidthAvailableForImmediateFlip = |
2647 | v->BandwidthAvailableForImmediateFlip |
2648 | - dml_max( |
2649 | v->ReadBandwidthPlaneLuma[k] * v->UrgentBurstFactorLuma[k] |
2650 | + v->ReadBandwidthPlaneChroma[k] * v->UrgentBurstFactorChroma[k] |
2651 | + v->cursor_bw[k] * v->UrgentBurstFactorCursor[k], |
2652 | v->DPPPerPlane[k] * (v->RequiredPrefetchPixDataBWLuma[k] * v->UrgentBurstFactorLumaPre[k] + |
2653 | v->RequiredPrefetchPixDataBWChroma[k] * v->UrgentBurstFactorChromaPre[k]) + |
2654 | v->cursor_bw_pre[k] * v->UrgentBurstFactorCursorPre[k]); |
2655 | } |
2656 | |
2657 | v->TotImmediateFlipBytes = 0; |
2658 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
2659 | v->TotImmediateFlipBytes = v->TotImmediateFlipBytes + v->DPPPerPlane[k] * (v->PDEAndMetaPTEBytesFrame[k] + v->MetaRowByte[k] + v->PixelPTEBytesPerRow[k]); |
2660 | } |
2661 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
2662 | CalculateFlipSchedule( |
2663 | mode_lib, |
2664 | v->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
2665 | v->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
2666 | v->UrgentExtraLatency, |
2667 | v->UrgentLatency, |
2668 | v->GPUVMMaxPageTableLevels, |
2669 | v->HostVMEnable, |
2670 | v->HostVMMaxNonCachedPageTableLevels, |
2671 | v->GPUVMEnable, |
2672 | v->HostVMMinPageSize, |
2673 | v->PDEAndMetaPTEBytesFrame[k], |
2674 | v->MetaRowByte[k], |
2675 | v->PixelPTEBytesPerRow[k], |
2676 | v->BandwidthAvailableForImmediateFlip, |
2677 | v->TotImmediateFlipBytes, |
2678 | v->SourcePixelFormat[k], |
2679 | v->HTotal[k] / v->PixelClock[k], |
2680 | v->VRatio[k], |
2681 | v->VRatioChroma[k], |
2682 | v->Tno_bw[k], |
2683 | v->DCCEnable[k], |
2684 | v->dpte_row_height[k], |
2685 | v->meta_row_height[k], |
2686 | v->dpte_row_height_chroma[k], |
2687 | v->meta_row_height_chroma[k], |
2688 | &v->DestinationLinesToRequestVMInImmediateFlip[k], |
2689 | &v->DestinationLinesToRequestRowInImmediateFlip[k], |
2690 | &v->final_flip_bw[k], |
2691 | &v->ImmediateFlipSupportedForPipe[k]); |
2692 | } |
2693 | v->total_dcn_read_bw_with_flip = 0.0; |
2694 | v->total_dcn_read_bw_with_flip_no_urgent_burst = 0.0; |
2695 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
2696 | v->total_dcn_read_bw_with_flip = v->total_dcn_read_bw_with_flip + dml_max3( |
2697 | v->DPPPerPlane[k] * v->prefetch_vmrow_bw[k], |
2698 | v->DPPPerPlane[k] * v->final_flip_bw[k] + |
2699 | v->ReadBandwidthLuma[k] * v->UrgentBurstFactorLuma[k] + |
2700 | v->ReadBandwidthChroma[k] * v->UrgentBurstFactorChroma[k] + |
2701 | v->cursor_bw[k] * v->UrgentBurstFactorCursor[k], |
2702 | v->DPPPerPlane[k] * (v->final_flip_bw[k] + |
2703 | v->RequiredPrefetchPixDataBWLuma[k] * v->UrgentBurstFactorLumaPre[k] + |
2704 | v->RequiredPrefetchPixDataBWChroma[k] * v->UrgentBurstFactorChromaPre[k]) + |
2705 | v->cursor_bw_pre[k] * v->UrgentBurstFactorCursorPre[k]); |
2706 | v->total_dcn_read_bw_with_flip_no_urgent_burst = |
2707 | v->total_dcn_read_bw_with_flip_no_urgent_burst + |
2708 | dml_max3(v->DPPPerPlane[k] * v->prefetch_vmrow_bw[k], |
2709 | v->DPPPerPlane[k] * v->final_flip_bw[k] + v->ReadBandwidthPlaneLuma[k] + v->ReadBandwidthPlaneChroma[k] + v->cursor_bw[k], |
2710 | v->DPPPerPlane[k] * (v->final_flip_bw[k] + v->RequiredPrefetchPixDataBWLuma[k] + v->RequiredPrefetchPixDataBWChroma[k]) + v->cursor_bw_pre[k]); |
2711 | |
2712 | } |
2713 | v->FractionOfUrgentBandwidthImmediateFlip = v->total_dcn_read_bw_with_flip_no_urgent_burst / v->ReturnBW; |
2714 | |
2715 | v->ImmediateFlipSupported = true1; |
2716 | if (v->total_dcn_read_bw_with_flip > v->ReturnBW) { |
2717 | v->ImmediateFlipSupported = false0; |
2718 | v->total_dcn_read_bw_with_flip = MaxTotalRDBandwidth; |
2719 | } |
2720 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
2721 | if (v->ImmediateFlipSupportedForPipe[k] == false0) { |
2722 | v->ImmediateFlipSupported = false0; |
2723 | } |
2724 | } |
2725 | } else { |
2726 | v->ImmediateFlipSupported = false0; |
2727 | } |
2728 | |
2729 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
2730 | if (v->ErrorResult[k] || v->NotEnoughTimeForDynamicMetadata[k]) { |
2731 | v->PrefetchModeSupported = false0; |
2732 | dml_print("DML: CalculatePrefetchSchedule ***failed***. Prefetch schedule violation. Results are NOT valid\n"){do { } while(0); }; |
2733 | } |
2734 | } |
2735 | |
2736 | v->VStartupLines = v->VStartupLines + 1; |
2737 | v->PrefetchModeSupported = (v->PrefetchModeSupported == true1 && ((!v->ImmediateFlipSupport && |
2738 | !v->HostVMEnable && v->ImmediateFlipRequirement[0] != dm_immediate_flip_required) || |
2739 | v->ImmediateFlipSupported)) ? true1 : false0; |
2740 | } while (!v->PrefetchModeSupported && v->VStartupLines <= v->MaximumMaxVStartupLines); |
2741 | ASSERT(v->PrefetchModeSupported)do { if (({ static int __warned; int __ret = !!(!(v->PrefetchModeSupported )); if (__ret && !__warned) { printf("WARNING %s failed at %s:%d\n" , "!(v->PrefetchModeSupported)", "/usr/src/sys/dev/pci/drm/amd/display/dc/dml/dcn30/display_mode_vba_30.c" , 2741); __warned = 1; } __builtin_expect(!!(__ret), 0); })) do {} while (0); } while (0); |
2742 | |
2743 | //Watermarks and NB P-State/DRAM Clock Change Support |
2744 | { |
2745 | enum clock_change_support DRAMClockChangeSupport = 0; // dummy |
2746 | CalculateWatermarksAndDRAMSpeedChangeSupport( |
2747 | mode_lib, |
2748 | PrefetchMode, |
2749 | v->NumberOfActivePlanes, |
2750 | v->MaxLineBufferLines, |
2751 | v->LineBufferSize, |
2752 | v->DPPOutputBufferPixels, |
2753 | v->DETBufferSizeInKByte[0], |
2754 | v->WritebackInterfaceBufferSize, |
2755 | v->DCFCLK, |
2756 | v->ReturnBW, |
2757 | v->GPUVMEnable, |
2758 | v->dpte_group_bytes, |
2759 | v->MetaChunkSize, |
2760 | v->UrgentLatency, |
2761 | v->UrgentExtraLatency, |
2762 | v->WritebackLatency, |
2763 | v->WritebackChunkSize, |
2764 | v->SOCCLK, |
2765 | v->FinalDRAMClockChangeLatency, |
2766 | v->SRExitTime, |
2767 | v->SREnterPlusExitTime, |
2768 | v->DCFCLKDeepSleep, |
2769 | v->DPPPerPlane, |
2770 | v->DCCEnable, |
2771 | v->DPPCLK, |
2772 | v->DETBufferSizeY, |
2773 | v->DETBufferSizeC, |
2774 | v->SwathHeightY, |
2775 | v->SwathHeightC, |
2776 | v->LBBitPerPixel, |
2777 | v->SwathWidthY, |
2778 | v->SwathWidthC, |
2779 | v->HRatio, |
2780 | v->HRatioChroma, |
2781 | v->vtaps, |
2782 | v->VTAPsChroma, |
2783 | v->VRatio, |
2784 | v->VRatioChroma, |
2785 | v->HTotal, |
2786 | v->PixelClock, |
2787 | v->BlendingAndTiming, |
2788 | v->BytePerPixelDETY, |
2789 | v->BytePerPixelDETC, |
2790 | v->DSTXAfterScaler, |
2791 | v->DSTYAfterScaler, |
2792 | v->WritebackEnable, |
2793 | v->WritebackPixelFormat, |
2794 | v->WritebackDestinationWidth, |
2795 | v->WritebackDestinationHeight, |
2796 | v->WritebackSourceHeight, |
2797 | &DRAMClockChangeSupport, |
2798 | &v->UrgentWatermark, |
2799 | &v->WritebackUrgentWatermark, |
2800 | &v->DRAMClockChangeWatermark, |
2801 | &v->WritebackDRAMClockChangeWatermark, |
2802 | &v->StutterExitWatermark, |
2803 | &v->StutterEnterPlusExitWatermark, |
2804 | &v->MinActiveDRAMClockChangeLatencySupported); |
2805 | |
2806 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
2807 | if (v->WritebackEnable[k] == true1) { |
2808 | if (v->BlendingAndTiming[k] == k) { |
2809 | v->ThisVStartup = v->VStartup[k]; |
2810 | } else { |
2811 | for (j = 0; j < v->NumberOfActivePlanes; ++j) { |
2812 | if (v->BlendingAndTiming[k] == j) { |
2813 | v->ThisVStartup = v->VStartup[j]; |
2814 | } |
2815 | } |
2816 | } |
2817 | v->WritebackAllowDRAMClockChangeEndPosition[k] = dml_max(0, |
2818 | v->ThisVStartup * v->HTotal[k] / v->PixelClock[k] - v->WritebackDRAMClockChangeWatermark); |
2819 | } else { |
2820 | v->WritebackAllowDRAMClockChangeEndPosition[k] = 0; |
2821 | } |
2822 | } |
2823 | |
2824 | } |
2825 | |
2826 | |
2827 | //Display Pipeline Delivery Time in Prefetch, Groups |
2828 | CalculatePixelDeliveryTimes( |
2829 | v->NumberOfActivePlanes, |
2830 | v->VRatio, |
2831 | v->VRatioChroma, |
2832 | v->VRatioPrefetchY, |
2833 | v->VRatioPrefetchC, |
2834 | v->swath_width_luma_ub, |
2835 | v->swath_width_chroma_ub, |
2836 | v->DPPPerPlane, |
2837 | v->HRatio, |
2838 | v->HRatioChroma, |
2839 | v->PixelClock, |
2840 | v->PSCL_THROUGHPUT_LUMA, |
2841 | v->PSCL_THROUGHPUT_CHROMA, |
2842 | v->DPPCLK, |
2843 | v->BytePerPixelC, |
2844 | v->SourceScan, |
2845 | v->NumberOfCursors, |
2846 | v->CursorWidth, |
2847 | v->CursorBPP, |
2848 | v->BlockWidth256BytesY, |
2849 | v->BlockHeight256BytesY, |
2850 | v->BlockWidth256BytesC, |
2851 | v->BlockHeight256BytesC, |
2852 | v->DisplayPipeLineDeliveryTimeLuma, |
2853 | v->DisplayPipeLineDeliveryTimeChroma, |
2854 | v->DisplayPipeLineDeliveryTimeLumaPrefetch, |
2855 | v->DisplayPipeLineDeliveryTimeChromaPrefetch, |
2856 | v->DisplayPipeRequestDeliveryTimeLuma, |
2857 | v->DisplayPipeRequestDeliveryTimeChroma, |
2858 | v->DisplayPipeRequestDeliveryTimeLumaPrefetch, |
2859 | v->DisplayPipeRequestDeliveryTimeChromaPrefetch, |
2860 | v->CursorRequestDeliveryTime, |
2861 | v->CursorRequestDeliveryTimePrefetch); |
2862 | |
2863 | CalculateMetaAndPTETimes( |
2864 | v->NumberOfActivePlanes, |
2865 | v->GPUVMEnable, |
2866 | v->MetaChunkSize, |
2867 | v->MinMetaChunkSizeBytes, |
2868 | v->HTotal, |
2869 | v->VRatio, |
2870 | v->VRatioChroma, |
2871 | v->DestinationLinesToRequestRowInVBlank, |
2872 | v->DestinationLinesToRequestRowInImmediateFlip, |
2873 | v->DCCEnable, |
2874 | v->PixelClock, |
2875 | v->BytePerPixelY, |
2876 | v->BytePerPixelC, |
2877 | v->SourceScan, |
2878 | v->dpte_row_height, |
2879 | v->dpte_row_height_chroma, |
2880 | v->meta_row_width, |
2881 | v->meta_row_width_chroma, |
2882 | v->meta_row_height, |
2883 | v->meta_row_height_chroma, |
2884 | v->meta_req_width, |
2885 | v->meta_req_width_chroma, |
2886 | v->meta_req_height, |
2887 | v->meta_req_height_chroma, |
2888 | v->dpte_group_bytes, |
2889 | v->PTERequestSizeY, |
2890 | v->PTERequestSizeC, |
2891 | v->PixelPTEReqWidthY, |
2892 | v->PixelPTEReqHeightY, |
2893 | v->PixelPTEReqWidthC, |
2894 | v->PixelPTEReqHeightC, |
2895 | v->dpte_row_width_luma_ub, |
2896 | v->dpte_row_width_chroma_ub, |
2897 | v->DST_Y_PER_PTE_ROW_NOM_L, |
2898 | v->DST_Y_PER_PTE_ROW_NOM_C, |
2899 | v->DST_Y_PER_META_ROW_NOM_L, |
2900 | v->DST_Y_PER_META_ROW_NOM_C, |
2901 | v->TimePerMetaChunkNominal, |
2902 | v->TimePerChromaMetaChunkNominal, |
2903 | v->TimePerMetaChunkVBlank, |
2904 | v->TimePerChromaMetaChunkVBlank, |
2905 | v->TimePerMetaChunkFlip, |
2906 | v->TimePerChromaMetaChunkFlip, |
2907 | v->time_per_pte_group_nom_luma, |
2908 | v->time_per_pte_group_vblank_luma, |
2909 | v->time_per_pte_group_flip_luma, |
2910 | v->time_per_pte_group_nom_chroma, |
2911 | v->time_per_pte_group_vblank_chroma, |
2912 | v->time_per_pte_group_flip_chroma); |
2913 | |
2914 | CalculateVMGroupAndRequestTimes( |
2915 | v->NumberOfActivePlanes, |
2916 | v->GPUVMEnable, |
2917 | v->GPUVMMaxPageTableLevels, |
2918 | v->HTotal, |
2919 | v->BytePerPixelC, |
2920 | v->DestinationLinesToRequestVMInVBlank, |
2921 | v->DestinationLinesToRequestVMInImmediateFlip, |
2922 | v->DCCEnable, |
2923 | v->PixelClock, |
2924 | v->dpte_row_width_luma_ub, |
2925 | v->dpte_row_width_chroma_ub, |
2926 | v->vm_group_bytes, |
2927 | v->dpde0_bytes_per_frame_ub_l, |
2928 | v->dpde0_bytes_per_frame_ub_c, |
2929 | v->meta_pte_bytes_per_frame_ub_l, |
2930 | v->meta_pte_bytes_per_frame_ub_c, |
2931 | v->TimePerVMGroupVBlank, |
2932 | v->TimePerVMGroupFlip, |
2933 | v->TimePerVMRequestVBlank, |
2934 | v->TimePerVMRequestFlip); |
2935 | |
2936 | |
2937 | // Min TTUVBlank |
2938 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
2939 | if (PrefetchMode == 0) { |
2940 | v->AllowDRAMClockChangeDuringVBlank[k] = true1; |
2941 | v->AllowDRAMSelfRefreshDuringVBlank[k] = true1; |
2942 | v->MinTTUVBlank[k] = dml_max( |
2943 | v->DRAMClockChangeWatermark, |
2944 | dml_max( |
2945 | v->StutterEnterPlusExitWatermark, |
2946 | v->UrgentWatermark)); |
2947 | } else if (PrefetchMode == 1) { |
2948 | v->AllowDRAMClockChangeDuringVBlank[k] = false0; |
2949 | v->AllowDRAMSelfRefreshDuringVBlank[k] = true1; |
2950 | v->MinTTUVBlank[k] = dml_max( |
2951 | v->StutterEnterPlusExitWatermark, |
2952 | v->UrgentWatermark); |
2953 | } else { |
2954 | v->AllowDRAMClockChangeDuringVBlank[k] = false0; |
2955 | v->AllowDRAMSelfRefreshDuringVBlank[k] = false0; |
2956 | v->MinTTUVBlank[k] = v->UrgentWatermark; |
2957 | } |
2958 | if (!v->DynamicMetadataEnable[k]) |
2959 | v->MinTTUVBlank[k] = v->TCalc |
2960 | + v->MinTTUVBlank[k]; |
2961 | } |
2962 | |
2963 | // DCC Configuration |
2964 | v->ActiveDPPs = 0; |
2965 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
2966 | CalculateDCCConfiguration(v->DCCEnable[k], false0, // We should always know the direction DCCProgrammingAssumesScanDirectionUnknown, |
2967 | v->SourcePixelFormat[k], |
2968 | v->SurfaceWidthY[k], |
2969 | v->SurfaceWidthC[k], |
2970 | v->SurfaceHeightY[k], |
2971 | v->SurfaceHeightC[k], |
2972 | v->DETBufferSizeInKByte[0] * 1024, |
2973 | v->BlockHeight256BytesY[k], |
2974 | v->BlockHeight256BytesC[k], |
2975 | v->SurfaceTiling[k], |
2976 | v->BytePerPixelY[k], |
2977 | v->BytePerPixelC[k], |
2978 | v->BytePerPixelDETY[k], |
2979 | v->BytePerPixelDETC[k], |
2980 | v->SourceScan[k], |
2981 | &v->DCCYMaxUncompressedBlock[k], |
2982 | &v->DCCCMaxUncompressedBlock[k], |
2983 | &v->DCCYMaxCompressedBlock[k], |
2984 | &v->DCCCMaxCompressedBlock[k], |
2985 | &v->DCCYIndependentBlock[k], |
2986 | &v->DCCCIndependentBlock[k]); |
2987 | } |
2988 | |
2989 | { |
2990 | //Maximum Bandwidth Used |
2991 | v->TotalDataReadBandwidth = 0; |
2992 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
2993 | v->TotalDataReadBandwidth = v->TotalDataReadBandwidth |
2994 | + v->ReadBandwidthPlaneLuma[k] |
2995 | + v->ReadBandwidthPlaneChroma[k]; |
2996 | } |
2997 | } |
2998 | |
2999 | // VStartup Margin |
3000 | v->VStartupMargin = 0; |
3001 | v->FirstMainPlane = true1; |
3002 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
3003 | if (v->BlendingAndTiming[k] == k) { |
3004 | double margin = (v->MaxVStartupLines[k] - v->VStartup[k]) * v->HTotal[k] |
3005 | / v->PixelClock[k]; |
3006 | if (v->FirstMainPlane == true1) { |
3007 | v->VStartupMargin = margin; |
3008 | v->FirstMainPlane = false0; |
3009 | } else { |
3010 | v->VStartupMargin = dml_min(v->VStartupMargin, margin); |
3011 | } |
3012 | } |
3013 | } |
3014 | |
3015 | // Stutter Efficiency |
3016 | CalculateStutterEfficiency( |
3017 | v->NumberOfActivePlanes, |
3018 | v->ROBBufferSizeInKByte, |
3019 | v->TotalDataReadBandwidth, |
3020 | v->DCFCLK, |
3021 | v->ReturnBW, |
3022 | v->SRExitTime, |
3023 | v->SynchronizedVBlank, |
3024 | v->DPPPerPlane, |
3025 | v->DETBufferSizeY, |
3026 | v->BytePerPixelY, |
3027 | v->BytePerPixelDETY, |
3028 | v->SwathWidthY, |
3029 | v->SwathHeightY, |
3030 | v->SwathHeightC, |
3031 | v->DCCRateLuma, |
3032 | v->DCCRateChroma, |
3033 | v->HTotal, |
3034 | v->VTotal, |
3035 | v->PixelClock, |
3036 | v->VRatio, |
3037 | v->SourceScan, |
3038 | v->BlockHeight256BytesY, |
3039 | v->BlockWidth256BytesY, |
3040 | v->BlockHeight256BytesC, |
3041 | v->BlockWidth256BytesC, |
3042 | v->DCCYMaxUncompressedBlock, |
3043 | v->DCCCMaxUncompressedBlock, |
3044 | v->VActive, |
3045 | v->DCCEnable, |
3046 | v->WritebackEnable, |
3047 | v->ReadBandwidthPlaneLuma, |
3048 | v->ReadBandwidthPlaneChroma, |
3049 | v->meta_row_bw, |
3050 | v->dpte_row_bw, |
3051 | &v->StutterEfficiencyNotIncludingVBlank, |
3052 | &v->StutterEfficiency, |
3053 | &v->StutterPeriod); |
3054 | } |
3055 | |
3056 | static void DisplayPipeConfiguration(struct display_mode_lib *mode_lib) |
3057 | { |
3058 | // Display Pipe Configuration |
3059 | double BytePerPixDETY[DC__NUM_DPP__MAX8] = { 0 }; |
3060 | double BytePerPixDETC[DC__NUM_DPP__MAX8] = { 0 }; |
3061 | int BytePerPixY[DC__NUM_DPP__MAX8] = { 0 }; |
3062 | int BytePerPixC[DC__NUM_DPP__MAX8] = { 0 }; |
3063 | int Read256BytesBlockHeightY[DC__NUM_DPP__MAX8] = { 0 }; |
3064 | int Read256BytesBlockHeightC[DC__NUM_DPP__MAX8] = { 0 }; |
3065 | int Read256BytesBlockWidthY[DC__NUM_DPP__MAX8] = { 0 }; |
3066 | int Read256BytesBlockWidthC[DC__NUM_DPP__MAX8] = { 0 }; |
3067 | double dummy1[DC__NUM_DPP__MAX8] = { 0 }; |
3068 | double dummy2[DC__NUM_DPP__MAX8] = { 0 }; |
3069 | double dummy3[DC__NUM_DPP__MAX8] = { 0 }; |
3070 | double dummy4[DC__NUM_DPP__MAX8] = { 0 }; |
3071 | int dummy5[DC__NUM_DPP__MAX8] = { 0 }; |
3072 | int dummy6[DC__NUM_DPP__MAX8] = { 0 }; |
3073 | bool_Bool dummy7[DC__NUM_DPP__MAX8] = { 0 }; |
3074 | bool_Bool dummysinglestring = 0; |
3075 | unsigned int k; |
3076 | |
3077 | for (k = 0; k < mode_lib->vba.NumberOfActivePlanes; ++k) { |
3078 | |
3079 | dml30_CalculateBytePerPixelAnd256BBlockSizes( |
3080 | mode_lib->vba.SourcePixelFormat[k], |
3081 | mode_lib->vba.SurfaceTiling[k], |
3082 | &BytePerPixY[k], |
3083 | &BytePerPixC[k], |
3084 | &BytePerPixDETY[k], |
3085 | &BytePerPixDETC[k], |
3086 | &Read256BytesBlockHeightY[k], |
3087 | &Read256BytesBlockHeightC[k], |
3088 | &Read256BytesBlockWidthY[k], |
3089 | &Read256BytesBlockWidthC[k]); |
3090 | } |
3091 | CalculateSwathAndDETConfiguration( |
3092 | false0, |
3093 | mode_lib->vba.NumberOfActivePlanes, |
3094 | mode_lib->vba.DETBufferSizeInKByte[0], |
3095 | dummy1, |
3096 | dummy2, |
3097 | mode_lib->vba.SourceScan, |
3098 | mode_lib->vba.SourcePixelFormat, |
3099 | mode_lib->vba.SurfaceTiling, |
3100 | mode_lib->vba.ViewportWidth, |
3101 | mode_lib->vba.ViewportHeight, |
3102 | mode_lib->vba.SurfaceWidthY, |
3103 | mode_lib->vba.SurfaceWidthC, |
3104 | mode_lib->vba.SurfaceHeightY, |
3105 | mode_lib->vba.SurfaceHeightC, |
3106 | Read256BytesBlockHeightY, |
3107 | Read256BytesBlockHeightC, |
3108 | Read256BytesBlockWidthY, |
3109 | Read256BytesBlockWidthC, |
3110 | mode_lib->vba.ODMCombineEnabled, |
3111 | mode_lib->vba.BlendingAndTiming, |
3112 | BytePerPixY, |
3113 | BytePerPixC, |
3114 | BytePerPixDETY, |
3115 | BytePerPixDETC, |
3116 | mode_lib->vba.HActive, |
3117 | mode_lib->vba.HRatio, |
3118 | mode_lib->vba.HRatioChroma, |
3119 | mode_lib->vba.DPPPerPlane, |
3120 | dummy5, |
3121 | dummy6, |
3122 | dummy3, |
3123 | dummy4, |
3124 | mode_lib->vba.SwathHeightY, |
3125 | mode_lib->vba.SwathHeightC, |
3126 | mode_lib->vba.DETBufferSizeY, |
3127 | mode_lib->vba.DETBufferSizeC, |
3128 | dummy7, |
3129 | &dummysinglestring); |
3130 | } |
3131 | |
3132 | void dml30_CalculateBytePerPixelAnd256BBlockSizes( |
3133 | enum source_format_class SourcePixelFormat, |
3134 | enum dm_swizzle_mode SurfaceTiling, |
3135 | unsigned int *BytePerPixelY, |
3136 | unsigned int *BytePerPixelC, |
3137 | double *BytePerPixelDETY, |
3138 | double *BytePerPixelDETC, |
3139 | unsigned int *BlockHeight256BytesY, |
3140 | unsigned int *BlockHeight256BytesC, |
3141 | unsigned int *BlockWidth256BytesY, |
3142 | unsigned int *BlockWidth256BytesC) |
3143 | { |
3144 | if (SourcePixelFormat == dm_444_64) { |
3145 | *BytePerPixelDETY = 8; |
3146 | *BytePerPixelDETC = 0; |
3147 | *BytePerPixelY = 8; |
3148 | *BytePerPixelC = 0; |
3149 | } else if (SourcePixelFormat == dm_444_32 || SourcePixelFormat == dm_rgbe) { |
3150 | *BytePerPixelDETY = 4; |
3151 | *BytePerPixelDETC = 0; |
3152 | *BytePerPixelY = 4; |
3153 | *BytePerPixelC = 0; |
3154 | } else if (SourcePixelFormat == dm_444_16) { |
3155 | *BytePerPixelDETY = 2; |
3156 | *BytePerPixelDETC = 0; |
3157 | *BytePerPixelY = 2; |
3158 | *BytePerPixelC = 0; |
3159 | } else if (SourcePixelFormat == dm_444_8) { |
3160 | *BytePerPixelDETY = 1; |
3161 | *BytePerPixelDETC = 0; |
3162 | *BytePerPixelY = 1; |
3163 | *BytePerPixelC = 0; |
3164 | } else if (SourcePixelFormat == dm_rgbe_alpha) { |
3165 | *BytePerPixelDETY = 4; |
3166 | *BytePerPixelDETC = 1; |
3167 | *BytePerPixelY = 4; |
3168 | *BytePerPixelC = 1; |
3169 | } else if (SourcePixelFormat == dm_420_8) { |
3170 | *BytePerPixelDETY = 1; |
3171 | *BytePerPixelDETC = 2; |
3172 | *BytePerPixelY = 1; |
3173 | *BytePerPixelC = 2; |
3174 | } else if (SourcePixelFormat == dm_420_12) { |
3175 | *BytePerPixelDETY = 2; |
3176 | *BytePerPixelDETC = 4; |
3177 | *BytePerPixelY = 2; |
3178 | *BytePerPixelC = 4; |
3179 | } else { |
3180 | *BytePerPixelDETY = 4.0 / 3; |
3181 | *BytePerPixelDETC = 8.0 / 3; |
3182 | *BytePerPixelY = 2; |
3183 | *BytePerPixelC = 4; |
3184 | } |
3185 | |
3186 | if ((SourcePixelFormat == dm_444_64 || SourcePixelFormat == dm_444_32 |
3187 | || SourcePixelFormat == dm_444_16 || SourcePixelFormat == dm_444_8 |
3188 | || SourcePixelFormat == dm_mono_16 || SourcePixelFormat == dm_mono_8 |
3189 | || SourcePixelFormat == dm_rgbe)) { |
3190 | if (SurfaceTiling == dm_sw_linear) { |
3191 | *BlockHeight256BytesY = 1; |
3192 | } else if (SourcePixelFormat == dm_444_64) { |
3193 | *BlockHeight256BytesY = 4; |
3194 | } else if (SourcePixelFormat == dm_444_8) { |
3195 | *BlockHeight256BytesY = 16; |
3196 | } else { |
3197 | *BlockHeight256BytesY = 8; |
3198 | } |
3199 | *BlockWidth256BytesY = 256U / *BytePerPixelY / *BlockHeight256BytesY; |
3200 | *BlockHeight256BytesC = 0; |
3201 | *BlockWidth256BytesC = 0; |
3202 | } else { |
3203 | if (SurfaceTiling == dm_sw_linear) { |
3204 | *BlockHeight256BytesY = 1; |
3205 | *BlockHeight256BytesC = 1; |
3206 | } else if (SourcePixelFormat == dm_rgbe_alpha) { |
3207 | *BlockHeight256BytesY = 8; |
3208 | *BlockHeight256BytesC = 16; |
3209 | } else if (SourcePixelFormat == dm_420_8) { |
3210 | *BlockHeight256BytesY = 16; |
3211 | *BlockHeight256BytesC = 8; |
3212 | } else { |
3213 | *BlockHeight256BytesY = 8; |
3214 | *BlockHeight256BytesC = 8; |
3215 | } |
3216 | *BlockWidth256BytesY = 256U / *BytePerPixelY / *BlockHeight256BytesY; |
3217 | *BlockWidth256BytesC = 256U / *BytePerPixelC / *BlockHeight256BytesC; |
3218 | } |
3219 | } |
3220 | |
3221 | static double CalculateTWait( |
3222 | unsigned int PrefetchMode, |
3223 | double DRAMClockChangeLatency, |
3224 | double UrgentLatency, |
3225 | double SREnterPlusExitTime) |
3226 | { |
3227 | if (PrefetchMode == 0) { |
3228 | return dml_max(DRAMClockChangeLatency + UrgentLatency, |
3229 | dml_max(SREnterPlusExitTime, UrgentLatency)); |
3230 | } else if (PrefetchMode == 1) { |
3231 | return dml_max(SREnterPlusExitTime, UrgentLatency); |
3232 | } else { |
3233 | return UrgentLatency; |
3234 | } |
3235 | } |
3236 | |
3237 | double dml30_CalculateWriteBackDISPCLK( |
3238 | enum source_format_class WritebackPixelFormat, |
3239 | double PixelClock, |
3240 | double WritebackHRatio, |
3241 | double WritebackVRatio, |
3242 | unsigned int WritebackHTaps, |
3243 | unsigned int WritebackVTaps, |
3244 | long WritebackSourceWidth, |
3245 | long WritebackDestinationWidth, |
3246 | unsigned int HTotal, |
3247 | unsigned int WritebackLineBufferSize) |
3248 | { |
3249 | double DISPCLK_H = 0, DISPCLK_V = 0, DISPCLK_HB = 0; |
3250 | |
3251 | DISPCLK_H = PixelClock * dml_ceil(WritebackHTaps / 8.0, 1) / WritebackHRatio; |
3252 | DISPCLK_V = PixelClock * (WritebackVTaps * dml_ceil(WritebackDestinationWidth / 6.0, 1) + 8.0) / HTotal; |
3253 | DISPCLK_HB = PixelClock * WritebackVTaps * (WritebackDestinationWidth * WritebackVTaps - WritebackLineBufferSize / 57.0) / 6.0 / WritebackSourceWidth; |
3254 | return dml_max3(DISPCLK_H, DISPCLK_V, DISPCLK_HB); |
3255 | } |
3256 | |
3257 | static double CalculateWriteBackDelay( |
3258 | enum source_format_class WritebackPixelFormat, |
3259 | double WritebackHRatio, |
3260 | double WritebackVRatio, |
3261 | unsigned int WritebackVTaps, |
3262 | long WritebackDestinationWidth, |
3263 | long WritebackDestinationHeight, |
3264 | long WritebackSourceHeight, |
3265 | unsigned int HTotal) |
3266 | { |
3267 | double CalculateWriteBackDelay = 0; |
3268 | double Line_length = 0; |
3269 | double Output_lines_last_notclamped = 0; |
3270 | double WritebackVInit = 0; |
3271 | |
3272 | WritebackVInit = (WritebackVRatio + WritebackVTaps + 1) / 2; |
3273 | Line_length = dml_max((double) WritebackDestinationWidth, dml_ceil(WritebackDestinationWidth / 6.0, 1) * WritebackVTaps); |
3274 | Output_lines_last_notclamped = WritebackDestinationHeight - 1 - dml_ceil((WritebackSourceHeight - WritebackVInit) / WritebackVRatio, 1); |
3275 | if (Output_lines_last_notclamped < 0) { |
3276 | CalculateWriteBackDelay = 0; |
3277 | } else { |
3278 | CalculateWriteBackDelay = Output_lines_last_notclamped * Line_length + (HTotal - WritebackDestinationWidth) + 80; |
3279 | } |
3280 | return CalculateWriteBackDelay; |
3281 | } |
3282 | |
3283 | |
3284 | static void CalculateDynamicMetadataParameters(int MaxInterDCNTileRepeaters, double DPPCLK, double DISPCLK, |
3285 | double DCFClkDeepSleep, double PixelClock, long HTotal, long VBlank, long DynamicMetadataTransmittedBytes, |
3286 | long DynamicMetadataLinesBeforeActiveRequired, int InterlaceEnable, bool_Bool ProgressiveToInterlaceUnitInOPP, |
3287 | double *Tsetup, double *Tdmbf, double *Tdmec, double *Tdmsks) |
3288 | { |
3289 | double TotalRepeaterDelayTime = 0; |
3290 | double VUpdateWidthPix = 0; |
3291 | double VReadyOffsetPix = 0; |
3292 | double VUpdateOffsetPix = 0; |
3293 | TotalRepeaterDelayTime = MaxInterDCNTileRepeaters * (2 / DPPCLK + 3 / DISPCLK); |
3294 | VUpdateWidthPix = (14 / DCFClkDeepSleep + 12 / DPPCLK + TotalRepeaterDelayTime) * PixelClock; |
3295 | VReadyOffsetPix = dml_max(150.0 / DPPCLK, TotalRepeaterDelayTime + 20 / DCFClkDeepSleep + 10 / DPPCLK) * PixelClock; |
3296 | VUpdateOffsetPix = dml_ceil(HTotal / 4.0, 1); |
3297 | *Tsetup = (VUpdateOffsetPix + VUpdateWidthPix + VReadyOffsetPix) / PixelClock; |
3298 | *Tdmbf = DynamicMetadataTransmittedBytes / 4.0 / DISPCLK; |
3299 | *Tdmec = HTotal / PixelClock; |
3300 | if (DynamicMetadataLinesBeforeActiveRequired == 0) { |
3301 | *Tdmsks = VBlank * HTotal / PixelClock / 2.0; |
3302 | } else { |
3303 | *Tdmsks = DynamicMetadataLinesBeforeActiveRequired * HTotal / PixelClock; |
3304 | } |
3305 | if (InterlaceEnable == 1 && ProgressiveToInterlaceUnitInOPP == false0) { |
3306 | *Tdmsks = *Tdmsks / 2; |
3307 | } |
3308 | } |
3309 | |
3310 | static void CalculateRowBandwidth( |
3311 | bool_Bool GPUVMEnable, |
3312 | enum source_format_class SourcePixelFormat, |
3313 | double VRatio, |
3314 | double VRatioChroma, |
3315 | bool_Bool DCCEnable, |
3316 | double LineTime, |
3317 | unsigned int MetaRowByteLuma, |
3318 | unsigned int MetaRowByteChroma, |
3319 | unsigned int meta_row_height_luma, |
3320 | unsigned int meta_row_height_chroma, |
3321 | unsigned int PixelPTEBytesPerRowLuma, |
3322 | unsigned int PixelPTEBytesPerRowChroma, |
3323 | unsigned int dpte_row_height_luma, |
3324 | unsigned int dpte_row_height_chroma, |
3325 | double *meta_row_bw, |
3326 | double *dpte_row_bw) |
3327 | { |
3328 | if (DCCEnable != true1) { |
3329 | *meta_row_bw = 0; |
3330 | } else if (SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10 || SourcePixelFormat == dm_420_12 || SourcePixelFormat == dm_rgbe_alpha) { |
3331 | *meta_row_bw = VRatio * MetaRowByteLuma / (meta_row_height_luma * LineTime) |
3332 | + VRatioChroma * MetaRowByteChroma |
3333 | / (meta_row_height_chroma * LineTime); |
3334 | } else { |
3335 | *meta_row_bw = VRatio * MetaRowByteLuma / (meta_row_height_luma * LineTime); |
3336 | } |
3337 | |
3338 | if (GPUVMEnable != true1) { |
3339 | *dpte_row_bw = 0; |
3340 | } else if (SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10 || SourcePixelFormat == dm_420_12 || SourcePixelFormat == dm_rgbe_alpha) { |
3341 | *dpte_row_bw = VRatio * PixelPTEBytesPerRowLuma / (dpte_row_height_luma * LineTime) |
3342 | + VRatioChroma * PixelPTEBytesPerRowChroma |
3343 | / (dpte_row_height_chroma * LineTime); |
3344 | } else { |
3345 | *dpte_row_bw = VRatio * PixelPTEBytesPerRowLuma / (dpte_row_height_luma * LineTime); |
3346 | } |
3347 | } |
3348 | |
3349 | static void CalculateFlipSchedule( |
3350 | struct display_mode_lib *mode_lib, |
3351 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
3352 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
3353 | double UrgentExtraLatency, |
3354 | double UrgentLatency, |
3355 | unsigned int GPUVMMaxPageTableLevels, |
3356 | bool_Bool HostVMEnable, |
3357 | unsigned int HostVMMaxNonCachedPageTableLevels, |
3358 | bool_Bool GPUVMEnable, |
3359 | double HostVMMinPageSize, |
3360 | double PDEAndMetaPTEBytesPerFrame, |
3361 | double MetaRowBytes, |
3362 | double DPTEBytesPerRow, |
3363 | double BandwidthAvailableForImmediateFlip, |
3364 | unsigned int TotImmediateFlipBytes, |
3365 | enum source_format_class SourcePixelFormat, |
3366 | double LineTime, |
3367 | double VRatio, |
3368 | double VRatioChroma, |
3369 | double Tno_bw, |
3370 | bool_Bool DCCEnable, |
3371 | unsigned int dpte_row_height, |
3372 | unsigned int meta_row_height, |
3373 | unsigned int dpte_row_height_chroma, |
3374 | unsigned int meta_row_height_chroma, |
3375 | double *DestinationLinesToRequestVMInImmediateFlip, |
3376 | double *DestinationLinesToRequestRowInImmediateFlip, |
3377 | double *final_flip_bw, |
3378 | bool_Bool *ImmediateFlipSupportedForPipe) |
3379 | { |
3380 | double min_row_time = 0.0; |
3381 | unsigned int HostVMDynamicLevelsTrips = 0; |
3382 | double TimeForFetchingMetaPTEImmediateFlip = 0; |
3383 | double TimeForFetchingRowInVBlankImmediateFlip = 0; |
3384 | double ImmediateFlipBW = 0; |
3385 | double HostVMInefficiencyFactor = 0; |
3386 | |
3387 | if (GPUVMEnable == true1 && HostVMEnable == true1) { |
3388 | HostVMInefficiencyFactor = PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData / PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly; |
3389 | HostVMDynamicLevelsTrips = HostVMMaxNonCachedPageTableLevels; |
3390 | } else { |
3391 | HostVMInefficiencyFactor = 1; |
3392 | HostVMDynamicLevelsTrips = 0; |
3393 | } |
3394 | |
3395 | if (GPUVMEnable == true1 || DCCEnable == true1) { |
3396 | ImmediateFlipBW = (PDEAndMetaPTEBytesPerFrame + MetaRowBytes + DPTEBytesPerRow) * BandwidthAvailableForImmediateFlip / TotImmediateFlipBytes; |
3397 | } |
3398 | |
3399 | if (GPUVMEnable == true1) { |
3400 | TimeForFetchingMetaPTEImmediateFlip = dml_max3(Tno_bw + PDEAndMetaPTEBytesPerFrame * HostVMInefficiencyFactor / ImmediateFlipBW, |
3401 | UrgentExtraLatency + UrgentLatency * (GPUVMMaxPageTableLevels * (HostVMDynamicLevelsTrips + 1) - 1), LineTime / 4.0); |
3402 | } else { |
3403 | TimeForFetchingMetaPTEImmediateFlip = 0; |
3404 | } |
3405 | |
3406 | *DestinationLinesToRequestVMInImmediateFlip = dml_ceil(4.0 * (TimeForFetchingMetaPTEImmediateFlip / LineTime), 1) / 4.0; |
3407 | if ((GPUVMEnable == true1 || DCCEnable == true1)) { |
3408 | TimeForFetchingRowInVBlankImmediateFlip = dml_max3((MetaRowBytes + DPTEBytesPerRow * HostVMInefficiencyFactor) / ImmediateFlipBW, |
3409 | UrgentLatency * (HostVMDynamicLevelsTrips + 1), LineTime / 4); |
3410 | } else { |
3411 | TimeForFetchingRowInVBlankImmediateFlip = 0; |
3412 | } |
3413 | |
3414 | *DestinationLinesToRequestRowInImmediateFlip = dml_ceil(4.0 * (TimeForFetchingRowInVBlankImmediateFlip / LineTime), 1) / 4.0; |
3415 | |
3416 | if (GPUVMEnable == true1) { |
3417 | *final_flip_bw = dml_max(PDEAndMetaPTEBytesPerFrame * HostVMInefficiencyFactor / (*DestinationLinesToRequestVMInImmediateFlip * LineTime), |
3418 | (MetaRowBytes + DPTEBytesPerRow * HostVMInefficiencyFactor) / (*DestinationLinesToRequestRowInImmediateFlip * LineTime)); |
3419 | } else if ((GPUVMEnable == true1 || DCCEnable == true1)) { |
3420 | *final_flip_bw = (MetaRowBytes + DPTEBytesPerRow * HostVMInefficiencyFactor) / (*DestinationLinesToRequestRowInImmediateFlip * LineTime); |
3421 | } else { |
3422 | *final_flip_bw = 0; |
3423 | } |
3424 | |
3425 | |
3426 | if (SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10 || SourcePixelFormat == dm_rgbe_alpha) { |
3427 | if (GPUVMEnable == true1 && DCCEnable != true1) { |
3428 | min_row_time = dml_min(dpte_row_height * LineTime / VRatio, dpte_row_height_chroma * LineTime / VRatioChroma); |
3429 | } else if (GPUVMEnable != true1 && DCCEnable == true1) { |
3430 | min_row_time = dml_min(meta_row_height * LineTime / VRatio, meta_row_height_chroma * LineTime / VRatioChroma); |
3431 | } else { |
3432 | min_row_time = dml_min4(dpte_row_height * LineTime / VRatio, meta_row_height * LineTime / VRatio, |
3433 | dpte_row_height_chroma * LineTime / VRatioChroma, meta_row_height_chroma * LineTime / VRatioChroma); |
3434 | } |
3435 | } else { |
3436 | if (GPUVMEnable == true1 && DCCEnable != true1) { |
3437 | min_row_time = dpte_row_height * LineTime / VRatio; |
3438 | } else if (GPUVMEnable != true1 && DCCEnable == true1) { |
3439 | min_row_time = meta_row_height * LineTime / VRatio; |
3440 | } else { |
3441 | min_row_time = dml_min(dpte_row_height * LineTime / VRatio, meta_row_height * LineTime / VRatio); |
3442 | } |
3443 | } |
3444 | |
3445 | if (*DestinationLinesToRequestVMInImmediateFlip >= 32 || *DestinationLinesToRequestRowInImmediateFlip >= 16 |
3446 | || TimeForFetchingMetaPTEImmediateFlip + 2 * TimeForFetchingRowInVBlankImmediateFlip > min_row_time) { |
3447 | *ImmediateFlipSupportedForPipe = false0; |
3448 | } else { |
3449 | *ImmediateFlipSupportedForPipe = true1; |
3450 | } |
3451 | } |
3452 | |
3453 | static double TruncToValidBPP( |
3454 | double LinkBitRate, |
3455 | int Lanes, |
3456 | long HTotal, |
3457 | long HActive, |
3458 | double PixelClock, |
3459 | double DesiredBPP, |
3460 | bool_Bool DSCEnable, |
3461 | enum output_encoder_class Output, |
3462 | enum output_format_class Format, |
3463 | unsigned int DSCInputBitPerComponent, |
3464 | int DSCSlices, |
3465 | int AudioRate, |
3466 | int AudioLayout, |
3467 | enum odm_combine_mode ODMCombine) |
3468 | { |
3469 | double MaxLinkBPP = 0; |
3470 | int MinDSCBPP = 0; |
3471 | double MaxDSCBPP = 0; |
3472 | int NonDSCBPP0 = 0; |
3473 | int NonDSCBPP1 = 0; |
3474 | int NonDSCBPP2 = 0; |
3475 | |
3476 | if (Format == dm_420) { |
3477 | NonDSCBPP0 = 12; |
3478 | NonDSCBPP1 = 15; |
3479 | NonDSCBPP2 = 18; |
3480 | MinDSCBPP = 6; |
3481 | MaxDSCBPP = 1.5 * DSCInputBitPerComponent - 1.0 / 16; |
3482 | } else if (Format == dm_444) { |
3483 | NonDSCBPP0 = 24; |
3484 | NonDSCBPP1 = 30; |
3485 | NonDSCBPP2 = 36; |
3486 | MinDSCBPP = 8; |
3487 | MaxDSCBPP = 3 * DSCInputBitPerComponent - 1.0 / 16; |
3488 | } else { |
3489 | NonDSCBPP0 = 16; |
3490 | NonDSCBPP1 = 20; |
3491 | NonDSCBPP2 = 24; |
3492 | |
3493 | if (Format == dm_n422) { |
3494 | MinDSCBPP = 7; |
3495 | MaxDSCBPP = 2 * DSCInputBitPerComponent - 1.0 / 16.0; |
3496 | } |
3497 | else { |
3498 | MinDSCBPP = 8; |
3499 | MaxDSCBPP = 3 * DSCInputBitPerComponent - 1.0 / 16.0; |
3500 | } |
3501 | } |
3502 | |
3503 | if (DSCEnable && Output == dm_dp) { |
3504 | MaxLinkBPP = LinkBitRate / 10 * 8 * Lanes / PixelClock * (1 - 2.4 / 100); |
3505 | } else { |
3506 | MaxLinkBPP = LinkBitRate / 10 * 8 * Lanes / PixelClock; |
3507 | } |
3508 | |
3509 | if (ODMCombine == dm_odm_combine_mode_4to1 && MaxLinkBPP > 16) { |
3510 | MaxLinkBPP = 16; |
3511 | } else if (ODMCombine == dm_odm_combine_mode_2to1 && MaxLinkBPP > 32) { |
3512 | MaxLinkBPP = 32; |
3513 | } |
3514 | |
3515 | |
3516 | if (DesiredBPP == 0) { |
3517 | if (DSCEnable) { |
3518 | if (MaxLinkBPP < MinDSCBPP) { |
3519 | return BPP_INVALID0; |
3520 | } else if (MaxLinkBPP >= MaxDSCBPP) { |
3521 | return MaxDSCBPP; |
3522 | } else { |
3523 | return dml_floor(16.0 * MaxLinkBPP, 1.0) / 16.0; |
3524 | } |
3525 | } else { |
3526 | if (MaxLinkBPP >= NonDSCBPP2) { |
3527 | return NonDSCBPP2; |
3528 | } else if (MaxLinkBPP >= NonDSCBPP1) { |
3529 | return NonDSCBPP1; |
3530 | } else if (MaxLinkBPP >= NonDSCBPP0) { |
3531 | return NonDSCBPP0; |
3532 | } else { |
3533 | return BPP_INVALID0; |
3534 | } |
3535 | } |
3536 | } else { |
3537 | if (!((DSCEnable == false0 && (DesiredBPP == NonDSCBPP2 || DesiredBPP == NonDSCBPP1 || DesiredBPP == NonDSCBPP0 || DesiredBPP == 18)) || |
3538 | (DSCEnable && DesiredBPP >= MinDSCBPP && DesiredBPP <= MaxDSCBPP))) { |
3539 | return BPP_INVALID0; |
3540 | } else { |
3541 | return DesiredBPP; |
3542 | } |
3543 | } |
3544 | return BPP_INVALID0; |
3545 | } |
3546 | |
3547 | void dml30_ModeSupportAndSystemConfigurationFull(struct display_mode_lib *mode_lib) |
3548 | { |
3549 | struct vba_vars_st *v = &mode_lib->vba; |
3550 | int MinPrefetchMode, MaxPrefetchMode; |
3551 | int i; |
3552 | unsigned int j, k, m; |
3553 | bool_Bool EnoughWritebackUnits = true1; |
3554 | bool_Bool WritebackModeSupport = true1; |
3555 | bool_Bool ViewportExceedsSurface = false0; |
3556 | double MaxTotalVActiveRDBandwidth = 0; |
3557 | long ReorderingBytes = 0; |
3558 | bool_Bool NotUrgentLatencyHiding[DC__NUM_DPP__MAX8] = { 0 }; |
3559 | |
3560 | /*MODE SUPPORT, VOLTAGE STATE AND SOC CONFIGURATION*/ |
3561 | |
3562 | CalculateMinAndMaxPrefetchMode( |
3563 | mode_lib->vba.AllowDRAMSelfRefreshOrDRAMClockChangeInVblank, |
3564 | &MinPrefetchMode, &MaxPrefetchMode); |
3565 | |
3566 | /*Scale Ratio, taps Support Check*/ |
3567 | |
3568 | v->ScaleRatioAndTapsSupport = true1; |
3569 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
3570 | if (v->ScalerEnabled[k] == false0 |
3571 | && ((v->SourcePixelFormat[k] != dm_444_64 |
3572 | && v->SourcePixelFormat[k] != dm_444_32 |
3573 | && v->SourcePixelFormat[k] != dm_444_16 |
3574 | && v->SourcePixelFormat[k] != dm_mono_16 |
3575 | && v->SourcePixelFormat[k] != dm_mono_8 |
3576 | && v->SourcePixelFormat[k] != dm_rgbe |
3577 | && v->SourcePixelFormat[k] != dm_rgbe_alpha) |
3578 | || v->HRatio[k] != 1.0 |
3579 | || v->htaps[k] != 1.0 |
3580 | || v->VRatio[k] != 1.0 |
3581 | || v->vtaps[k] != 1.0)) { |
3582 | v->ScaleRatioAndTapsSupport = false0; |
3583 | } else if (v->vtaps[k] < 1.0 || v->vtaps[k] > 8.0 |
3584 | || v->htaps[k] < 1.0 || v->htaps[k] > 8.0 |
3585 | || (v->htaps[k] > 1.0 |
3586 | && (v->htaps[k] % 2) == 1) |
3587 | || v->HRatio[k] > v->MaxHSCLRatio |
3588 | || v->VRatio[k] > v->MaxVSCLRatio |
3589 | || v->HRatio[k] > v->htaps[k] |
3590 | || v->VRatio[k] > v->vtaps[k] |
3591 | || (v->SourcePixelFormat[k] != dm_444_64 |
3592 | && v->SourcePixelFormat[k] != dm_444_32 |
3593 | && v->SourcePixelFormat[k] != dm_444_16 |
3594 | && v->SourcePixelFormat[k] != dm_mono_16 |
3595 | && v->SourcePixelFormat[k] != dm_mono_8 |
3596 | && v->SourcePixelFormat[k] != dm_rgbe |
3597 | && (v->VTAPsChroma[k] < 1 |
3598 | || v->VTAPsChroma[k] > 8 |
3599 | || v->HTAPsChroma[k] < 1 |
3600 | || v->HTAPsChroma[k] > 8 |
3601 | || (v->HTAPsChroma[k] > 1 && v->HTAPsChroma[k] % 2 == 1) |
3602 | || v->HRatioChroma[k] > v->MaxHSCLRatio |
3603 | || v->VRatioChroma[k] > v->MaxVSCLRatio |
3604 | || v->HRatioChroma[k] > v->HTAPsChroma[k] |
3605 | || v->VRatioChroma[k] > v->VTAPsChroma[k]))) { |
3606 | v->ScaleRatioAndTapsSupport = false0; |
3607 | } |
3608 | } |
3609 | /*Source Format, Pixel Format and Scan Support Check*/ |
3610 | |
3611 | v->SourceFormatPixelAndScanSupport = true1; |
3612 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
3613 | if ((v->SurfaceTiling[k] == dm_sw_linear && (!(v->SourceScan[k] != dm_vert) || v->DCCEnable[k] == true1)) |
3614 | || ((v->SurfaceTiling[k] == dm_sw_64kb_d || v->SurfaceTiling[k] == dm_sw_64kb_d_t || v->SurfaceTiling[k] == dm_sw_64kb_d_x) |
3615 | && !(v->SourcePixelFormat[k] == dm_444_64))) { |
3616 | v->SourceFormatPixelAndScanSupport = false0; |
3617 | } |
3618 | } |
3619 | /*Bandwidth Support Check*/ |
3620 | |
3621 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
3622 | dml30_CalculateBytePerPixelAnd256BBlockSizes( |
3623 | v->SourcePixelFormat[k], |
3624 | v->SurfaceTiling[k], |
3625 | &v->BytePerPixelY[k], |
3626 | &v->BytePerPixelC[k], |
3627 | &v->BytePerPixelInDETY[k], |
3628 | &v->BytePerPixelInDETC[k], |
3629 | &v->Read256BlockHeightY[k], |
3630 | &v->Read256BlockHeightC[k], |
3631 | &v->Read256BlockWidthY[k], |
3632 | &v->Read256BlockWidthC[k]); |
3633 | } |
3634 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
3635 | if (v->SourceScan[k] != dm_vert) { |
3636 | v->SwathWidthYSingleDPP[k] = v->ViewportWidth[k]; |
3637 | v->SwathWidthCSingleDPP[k] = v->ViewportWidthChroma[k]; |
3638 | } else { |
3639 | v->SwathWidthYSingleDPP[k] = v->ViewportHeight[k]; |
3640 | v->SwathWidthCSingleDPP[k] = v->ViewportHeightChroma[k]; |
3641 | } |
3642 | } |
3643 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
3644 | v->ReadBandwidthLuma[k] = v->SwathWidthYSingleDPP[k] * dml_ceil(v->BytePerPixelInDETY[k], 1.0) / (v->HTotal[k] / v->PixelClock[k]) * v->VRatio[k]; |
3645 | v->ReadBandwidthChroma[k] = v->SwathWidthYSingleDPP[k] / 2 * dml_ceil(v->BytePerPixelInDETC[k], 2.0) / (v->HTotal[k] / v->PixelClock[k]) * v->VRatio[k] / 2.0; |
3646 | } |
3647 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
3648 | if (v->WritebackEnable[k] == true1 |
3649 | && v->WritebackPixelFormat[k] == dm_444_64) { |
3650 | v->WriteBandwidth[k] = v->WritebackDestinationWidth[k] |
3651 | * v->WritebackDestinationHeight[k] |
3652 | / (v->WritebackSourceHeight[k] |
3653 | * v->HTotal[k] |
3654 | / v->PixelClock[k]) * 8.0; |
3655 | } else if (v->WritebackEnable[k] == true1) { |
3656 | v->WriteBandwidth[k] = v->WritebackDestinationWidth[k] |
3657 | * v->WritebackDestinationHeight[k] |
3658 | / (v->WritebackSourceHeight[k] |
3659 | * v->HTotal[k] |
3660 | / v->PixelClock[k]) * 4.0; |
3661 | } else { |
3662 | v->WriteBandwidth[k] = 0.0; |
3663 | } |
3664 | } |
3665 | |
3666 | /*Writeback Latency support check*/ |
3667 | |
3668 | v->WritebackLatencySupport = true1; |
3669 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
3670 | if (v->WritebackEnable[k] == true1) { |
3671 | if (v->WritebackConfiguration == dm_whole_buffer_for_single_stream_no_interleave || |
3672 | v->WritebackConfiguration == dm_whole_buffer_for_single_stream_interleave) { |
3673 | if (v->WriteBandwidth[k] |
3674 | > 2.0 * v->WritebackInterfaceBufferSize * 1024 |
3675 | / v->WritebackLatency) { |
3676 | v->WritebackLatencySupport = false0; |
3677 | } |
3678 | } else { |
3679 | if (v->WriteBandwidth[k] |
3680 | > v->WritebackInterfaceBufferSize * 1024 |
3681 | / v->WritebackLatency) { |
3682 | v->WritebackLatencySupport = false0; |
3683 | } |
3684 | } |
3685 | } |
3686 | } |
3687 | |
3688 | /*Writeback Mode Support Check*/ |
3689 | |
3690 | v->TotalNumberOfActiveWriteback = 0; |
3691 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
3692 | if (v->WritebackEnable[k] == true1) { |
3693 | v->TotalNumberOfActiveWriteback = |
3694 | v->TotalNumberOfActiveWriteback + 1; |
3695 | } |
3696 | } |
3697 | |
3698 | if (v->TotalNumberOfActiveWriteback > v->MaxNumWriteback) { |
3699 | EnoughWritebackUnits = false0; |
3700 | } |
3701 | if (!v->WritebackSupportInterleaveAndUsingWholeBufferForASingleStream |
3702 | && (v->WritebackConfiguration == dm_whole_buffer_for_single_stream_no_interleave |
3703 | || v->WritebackConfiguration == dm_whole_buffer_for_single_stream_interleave)) { |
3704 | |
3705 | WritebackModeSupport = false0; |
3706 | } |
3707 | if (v->WritebackConfiguration == dm_whole_buffer_for_single_stream_no_interleave && v->TotalNumberOfActiveWriteback > 1) { |
3708 | WritebackModeSupport = false0; |
3709 | } |
3710 | |
3711 | /*Writeback Scale Ratio and Taps Support Check*/ |
3712 | |
3713 | v->WritebackScaleRatioAndTapsSupport = true1; |
3714 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
3715 | if (v->WritebackEnable[k] == true1) { |
3716 | if (v->WritebackHRatio[k] > v->WritebackMaxHSCLRatio |
3717 | || v->WritebackVRatio[k] |
3718 | > v->WritebackMaxVSCLRatio |
3719 | || v->WritebackHRatio[k] |
3720 | < v->WritebackMinHSCLRatio |
3721 | || v->WritebackVRatio[k] |
3722 | < v->WritebackMinVSCLRatio |
3723 | || v->WritebackHTaps[k] |
3724 | > v->WritebackMaxHSCLTaps |
3725 | || v->WritebackVTaps[k] |
3726 | > v->WritebackMaxVSCLTaps |
3727 | || v->WritebackHRatio[k] |
3728 | > v->WritebackHTaps[k] |
3729 | || v->WritebackVRatio[k] |
3730 | > v->WritebackVTaps[k] |
3731 | || (v->WritebackHTaps[k] > 2.0 |
3732 | && ((v->WritebackHTaps[k] % 2) |
3733 | == 1))) { |
3734 | v->WritebackScaleRatioAndTapsSupport = false0; |
3735 | } |
3736 | if (2.0 * v->WritebackDestinationWidth[k] * (v->WritebackVTaps[k] - 1) * 57 > v->WritebackLineBufferSize) { |
3737 | v->WritebackScaleRatioAndTapsSupport = false0; |
3738 | } |
3739 | } |
3740 | } |
3741 | /*Maximum DISPCLK/DPPCLK Support check*/ |
3742 | |
3743 | v->WritebackRequiredDISPCLK = 0.0; |
3744 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
3745 | if (v->WritebackEnable[k] == true1) { |
3746 | v->WritebackRequiredDISPCLK = dml_max(v->WritebackRequiredDISPCLK, |
3747 | dml30_CalculateWriteBackDISPCLK( |
3748 | v->WritebackPixelFormat[k], |
3749 | v->PixelClock[k], |
3750 | v->WritebackHRatio[k], |
3751 | v->WritebackVRatio[k], |
3752 | v->WritebackHTaps[k], |
3753 | v->WritebackVTaps[k], |
3754 | v->WritebackSourceWidth[k], |
3755 | v->WritebackDestinationWidth[k], |
3756 | v->HTotal[k], |
3757 | v->WritebackLineBufferSize)); |
3758 | } |
3759 | } |
3760 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
3761 | if (v->HRatio[k] > 1.0) { |
3762 | v->PSCL_FACTOR[k] = dml_min(v->MaxDCHUBToPSCLThroughput, v->MaxPSCLToLBThroughput * v->HRatio[k] / dml_ceil(v->htaps[k] / 6.0, 1.0)); |
3763 | } else { |
3764 | v->PSCL_FACTOR[k] = dml_min(v->MaxDCHUBToPSCLThroughput, v->MaxPSCLToLBThroughput); |
3765 | } |
3766 | if (v->BytePerPixelC[k] == 0.0) { |
3767 | v->PSCL_FACTOR_CHROMA[k] = 0.0; |
3768 | v->MinDPPCLKUsingSingleDPP[k] = v->PixelClock[k] |
3769 | * dml_max3(v->vtaps[k] / 6.0 * dml_min(1.0, v->HRatio[k]), v->HRatio[k] * v->VRatio[k] / v->PSCL_FACTOR[k], 1.0); |
3770 | if ((v->htaps[k] > 6.0 || v->vtaps[k] > 6.0) && v->MinDPPCLKUsingSingleDPP[k] < 2.0 * v->PixelClock[k]) { |
3771 | v->MinDPPCLKUsingSingleDPP[k] = 2.0 * v->PixelClock[k]; |
3772 | } |
3773 | } else { |
3774 | if (v->HRatioChroma[k] > 1.0) { |
3775 | v->PSCL_FACTOR_CHROMA[k] = dml_min(v->MaxDCHUBToPSCLThroughput, |
3776 | v->MaxPSCLToLBThroughput * v->HRatioChroma[k] / dml_ceil(v->HTAPsChroma[k] / 6.0, 1.0)); |
3777 | } else { |
3778 | v->PSCL_FACTOR_CHROMA[k] = dml_min(v->MaxDCHUBToPSCLThroughput, v->MaxPSCLToLBThroughput); |
3779 | } |
3780 | v->MinDPPCLKUsingSingleDPP[k] = v->PixelClock[k] * dml_max5(v->vtaps[k] / 6.0 * dml_min(1.0, v->HRatio[k]), |
3781 | v->HRatio[k] * v->VRatio[k] / v->PSCL_FACTOR[k], |
3782 | v->VTAPsChroma[k] / 6.0 * dml_min(1.0, v->HRatioChroma[k]), |
3783 | v->HRatioChroma[k] * v->VRatioChroma[k] / v->PSCL_FACTOR_CHROMA[k], |
3784 | 1.0); |
3785 | if ((v->htaps[k] > 6.0 || v->vtaps[k] > 6.0 || v->HTAPsChroma[k] > 6.0 || v->VTAPsChroma[k] > 6.0) |
3786 | && v->MinDPPCLKUsingSingleDPP[k] < 2.0 * v->PixelClock[k]) { |
3787 | v->MinDPPCLKUsingSingleDPP[k] = 2.0 * v->PixelClock[k]; |
3788 | } |
3789 | } |
3790 | } |
3791 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
3792 | int MaximumSwathWidthSupportLuma = 0; |
3793 | int MaximumSwathWidthSupportChroma = 0; |
3794 | |
3795 | if (v->SurfaceTiling[k] == dm_sw_linear) { |
3796 | MaximumSwathWidthSupportLuma = 8192.0; |
3797 | } else if (v->SourceScan[k] == dm_vert && v->BytePerPixelC[k] > 0) { |
3798 | MaximumSwathWidthSupportLuma = 2880.0; |
3799 | } else { |
3800 | MaximumSwathWidthSupportLuma = 5760.0; |
3801 | } |
3802 | |
3803 | if (v->SourcePixelFormat[k] == dm_420_8 || v->SourcePixelFormat[k] == dm_420_10 || v->SourcePixelFormat[k] == dm_420_12) { |
3804 | MaximumSwathWidthSupportChroma = MaximumSwathWidthSupportLuma / 2.0; |
3805 | } else { |
3806 | MaximumSwathWidthSupportChroma = MaximumSwathWidthSupportLuma; |
3807 | } |
3808 | v->MaximumSwathWidthInLineBufferLuma = v->LineBufferSize * dml_max(v->HRatio[k], 1.0) / v->LBBitPerPixel[k] |
3809 | / (v->vtaps[k] + dml_max(dml_ceil(v->VRatio[k], 1.0) - 2, 0.0)); |
3810 | if (v->BytePerPixelC[k] == 0.0) { |
3811 | v->MaximumSwathWidthInLineBufferChroma = 0; |
3812 | } else { |
3813 | v->MaximumSwathWidthInLineBufferChroma = v->LineBufferSize * dml_max(v->HRatioChroma[k], 1.0) / v->LBBitPerPixel[k] |
3814 | / (v->VTAPsChroma[k] + dml_max(dml_ceil(v->VRatioChroma[k], 1.0) - 2, 0.0)); |
3815 | } |
3816 | v->MaximumSwathWidthLuma[k] = dml_min(MaximumSwathWidthSupportLuma, v->MaximumSwathWidthInLineBufferLuma); |
3817 | v->MaximumSwathWidthChroma[k] = dml_min(MaximumSwathWidthSupportChroma, v->MaximumSwathWidthInLineBufferChroma); |
3818 | } |
3819 | |
3820 | CalculateSwathAndDETConfiguration( |
3821 | true1, |
3822 | v->NumberOfActivePlanes, |
3823 | v->DETBufferSizeInKByte[0], |
3824 | v->MaximumSwathWidthLuma, |
3825 | v->MaximumSwathWidthChroma, |
3826 | v->SourceScan, |
3827 | v->SourcePixelFormat, |
3828 | v->SurfaceTiling, |
3829 | v->ViewportWidth, |
3830 | v->ViewportHeight, |
3831 | v->SurfaceWidthY, |
3832 | v->SurfaceWidthC, |
3833 | v->SurfaceHeightY, |
3834 | v->SurfaceHeightC, |
3835 | v->Read256BlockHeightY, |
3836 | v->Read256BlockHeightC, |
3837 | v->Read256BlockWidthY, |
3838 | v->Read256BlockWidthC, |
3839 | v->odm_combine_dummy, |
3840 | v->BlendingAndTiming, |
3841 | v->BytePerPixelY, |
3842 | v->BytePerPixelC, |
3843 | v->BytePerPixelInDETY, |
3844 | v->BytePerPixelInDETC, |
3845 | v->HActive, |
3846 | v->HRatio, |
3847 | v->HRatioChroma, |
3848 | v->DPPPerPlane, |
3849 | v->swath_width_luma_ub, |
3850 | v->swath_width_chroma_ub, |
3851 | v->SwathWidthY, |
3852 | v->SwathWidthC, |
3853 | v->SwathHeightY, |
3854 | v->SwathHeightC, |
3855 | v->DETBufferSizeY, |
3856 | v->DETBufferSizeC, |
3857 | v->SingleDPPViewportSizeSupportPerPlane, |
3858 | &v->ViewportSizeSupport[0][0]); |
3859 | |
3860 | for (i = 0; i < v->soc.num_states; i++) { |
3861 | for (j = 0; j < 2; j++) { |
3862 | v->MaxDispclkRoundedDownToDFSGranularity = RoundToDFSGranularityDown(v->MaxDispclk[i], v->DISPCLKDPPCLKVCOSpeed); |
3863 | v->MaxDppclkRoundedDownToDFSGranularity = RoundToDFSGranularityDown(v->MaxDppclk[i], v->DISPCLKDPPCLKVCOSpeed); |
3864 | v->RequiredDISPCLK[i][j] = 0.0; |
3865 | v->DISPCLK_DPPCLK_Support[i][j] = true1; |
3866 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
3867 | v->PlaneRequiredDISPCLKWithoutODMCombine = v->PixelClock[k] * (1.0 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) |
3868 | * (1.0 + v->DISPCLKRampingMargin / 100.0); |
3869 | if ((v->PlaneRequiredDISPCLKWithoutODMCombine >= v->MaxDispclk[i] && v->MaxDispclk[i] == v->MaxDispclk[mode_lib->soc.num_states - 1] |
3870 | && v->MaxDppclk[i] == v->MaxDppclk[mode_lib->soc.num_states - 1])) { |
3871 | v->PlaneRequiredDISPCLKWithoutODMCombine = v->PixelClock[k] * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); |
3872 | } |
3873 | v->PlaneRequiredDISPCLKWithODMCombine2To1 = v->PixelClock[k] / 2 * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) |
3874 | * (1 + v->DISPCLKRampingMargin / 100.0); |
3875 | if ((v->PlaneRequiredDISPCLKWithODMCombine2To1 >= v->MaxDispclk[i] && v->MaxDispclk[i] == v->MaxDispclk[mode_lib->soc.num_states - 1] |
3876 | && v->MaxDppclk[i] == v->MaxDppclk[mode_lib->soc.num_states - 1])) { |
3877 | v->PlaneRequiredDISPCLKWithODMCombine2To1 = v->PixelClock[k] / 2 * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); |
3878 | } |
3879 | v->PlaneRequiredDISPCLKWithODMCombine4To1 = v->PixelClock[k] / 4 * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) |
3880 | * (1 + v->DISPCLKRampingMargin / 100.0); |
3881 | if ((v->PlaneRequiredDISPCLKWithODMCombine4To1 >= v->MaxDispclk[i] && v->MaxDispclk[i] == v->MaxDispclk[mode_lib->soc.num_states - 1] |
3882 | && v->MaxDppclk[i] == v->MaxDppclk[mode_lib->soc.num_states - 1])) { |
3883 | v->PlaneRequiredDISPCLKWithODMCombine4To1 = v->PixelClock[k] / 4 * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); |
3884 | } |
3885 | |
3886 | if (v->ODMCombinePolicy == dm_odm_combine_policy_none) { |
3887 | v->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_disabled; |
3888 | v->PlaneRequiredDISPCLK = v->PlaneRequiredDISPCLKWithoutODMCombine; |
3889 | } else if (v->ODMCombinePolicy == dm_odm_combine_policy_2to1) { |
3890 | v->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_2to1; |
3891 | v->PlaneRequiredDISPCLK = v->PlaneRequiredDISPCLKWithODMCombine2To1; |
3892 | } else if (v->ODMCombinePolicy == dm_odm_combine_policy_4to1 |
3893 | || v->PlaneRequiredDISPCLKWithODMCombine2To1 > v->MaxDispclkRoundedDownToDFSGranularity) { |
3894 | v->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_4to1; |
3895 | v->PlaneRequiredDISPCLK = v->PlaneRequiredDISPCLKWithODMCombine4To1; |
3896 | } else if (v->PlaneRequiredDISPCLKWithoutODMCombine > v->MaxDispclkRoundedDownToDFSGranularity) { |
3897 | v->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_2to1; |
3898 | v->PlaneRequiredDISPCLK = v->PlaneRequiredDISPCLKWithODMCombine2To1; |
3899 | } else { |
3900 | v->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_disabled; |
3901 | v->PlaneRequiredDISPCLK = v->PlaneRequiredDISPCLKWithoutODMCombine; |
3902 | } |
3903 | if (v->DSCEnabled[k] && v->HActive[k] > DCN30_MAX_DSC_IMAGE_WIDTH5184 |
3904 | && v->ODMCombineEnablePerState[i][k] != dm_odm_combine_mode_4to1) { |
3905 | if (v->HActive[k] / 2 > DCN30_MAX_DSC_IMAGE_WIDTH5184) { |
3906 | v->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_4to1; |
3907 | v->PlaneRequiredDISPCLK = v->PlaneRequiredDISPCLKWithODMCombine4To1; |
3908 | } else { |
3909 | v->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_2to1; |
3910 | v->PlaneRequiredDISPCLK = v->PlaneRequiredDISPCLKWithODMCombine2To1; |
3911 | } |
3912 | } |
3913 | if (v->OutputFormat[k] == dm_420 && v->HActive[k] > DCN30_MAX_FMT_420_BUFFER_WIDTH4096 |
3914 | && v->ODMCombineEnablePerState[i][k] != dm_odm_combine_mode_4to1) { |
3915 | if (v->HActive[k] / 2 > DCN30_MAX_FMT_420_BUFFER_WIDTH4096) { |
3916 | v->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_4to1; |
3917 | v->PlaneRequiredDISPCLK = v->PlaneRequiredDISPCLKWithODMCombine4To1; |
3918 | } else { |
3919 | v->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_2to1; |
3920 | v->PlaneRequiredDISPCLK = v->PlaneRequiredDISPCLKWithODMCombine2To1; |
3921 | } |
3922 | } |
3923 | if (v->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_4to1) { |
3924 | v->MPCCombine[i][j][k] = false0; |
3925 | v->NoOfDPP[i][j][k] = 4; |
3926 | v->RequiredDPPCLK[i][j][k] = v->MinDPPCLKUsingSingleDPP[k] * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) / 4; |
3927 | } else if (v->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_2to1) { |
3928 | v->MPCCombine[i][j][k] = false0; |
3929 | v->NoOfDPP[i][j][k] = 2; |
3930 | v->RequiredDPPCLK[i][j][k] = v->MinDPPCLKUsingSingleDPP[k] * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) / 2; |
3931 | } else if ((v->WhenToDoMPCCombine == dm_mpc_never |
3932 | || (v->MinDPPCLKUsingSingleDPP[k] * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) <= v->MaxDppclkRoundedDownToDFSGranularity |
3933 | && v->SingleDPPViewportSizeSupportPerPlane[k] == true1))) { |
3934 | v->MPCCombine[i][j][k] = false0; |
3935 | v->NoOfDPP[i][j][k] = 1; |
3936 | v->RequiredDPPCLK[i][j][k] = v->MinDPPCLKUsingSingleDPP[k] * (1.0 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); |
3937 | } else { |
3938 | v->MPCCombine[i][j][k] = true1; |
3939 | v->NoOfDPP[i][j][k] = 2; |
3940 | v->RequiredDPPCLK[i][j][k] = v->MinDPPCLKUsingSingleDPP[k] * (1.0 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) / 2.0; |
3941 | } |
3942 | v->RequiredDISPCLK[i][j] = dml_max(v->RequiredDISPCLK[i][j], v->PlaneRequiredDISPCLK); |
3943 | if ((v->MinDPPCLKUsingSingleDPP[k] / v->NoOfDPP[i][j][k] * (1.0 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) |
3944 | > v->MaxDppclkRoundedDownToDFSGranularity) || (v->PlaneRequiredDISPCLK > v->MaxDispclkRoundedDownToDFSGranularity)) { |
3945 | v->DISPCLK_DPPCLK_Support[i][j] = false0; |
3946 | } |
3947 | } |
3948 | v->TotalNumberOfActiveDPP[i][j] = 0; |
3949 | v->TotalNumberOfSingleDPPPlanes[i][j] = 0; |
3950 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
3951 | v->TotalNumberOfActiveDPP[i][j] = v->TotalNumberOfActiveDPP[i][j] + v->NoOfDPP[i][j][k]; |
3952 | if (v->NoOfDPP[i][j][k] == 1) |
3953 | v->TotalNumberOfSingleDPPPlanes[i][j] = v->TotalNumberOfSingleDPPPlanes[i][j] + 1; |
3954 | } |
3955 | if (j == 1 && v->WhenToDoMPCCombine != dm_mpc_never) { |
3956 | while (!(v->TotalNumberOfActiveDPP[i][j] >= v->MaxNumDPP || v->TotalNumberOfSingleDPPPlanes[i][j] == 0)) { |
3957 | double BWOfNonSplitPlaneOfMaximumBandwidth = 0; |
3958 | unsigned int NumberOfNonSplitPlaneOfMaximumBandwidth = 0; |
3959 | BWOfNonSplitPlaneOfMaximumBandwidth = 0; |
3960 | NumberOfNonSplitPlaneOfMaximumBandwidth = 0; |
3961 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
3962 | if (v->ReadBandwidthLuma[k] + v->ReadBandwidthChroma[k] > BWOfNonSplitPlaneOfMaximumBandwidth |
3963 | && v->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_disabled && v->MPCCombine[i][j][k] == false0) { |
3964 | BWOfNonSplitPlaneOfMaximumBandwidth = v->ReadBandwidthLuma[k] + v->ReadBandwidthChroma[k]; |
3965 | NumberOfNonSplitPlaneOfMaximumBandwidth = k; |
3966 | } |
3967 | } |
3968 | v->MPCCombine[i][j][NumberOfNonSplitPlaneOfMaximumBandwidth] = true1; |
3969 | v->NoOfDPP[i][j][NumberOfNonSplitPlaneOfMaximumBandwidth] = 2; |
3970 | v->RequiredDPPCLK[i][j][NumberOfNonSplitPlaneOfMaximumBandwidth] = v->MinDPPCLKUsingSingleDPP[NumberOfNonSplitPlaneOfMaximumBandwidth] |
3971 | * (1 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100) / 2; |
3972 | v->TotalNumberOfActiveDPP[i][j] = v->TotalNumberOfActiveDPP[i][j] + 1; |
3973 | v->TotalNumberOfSingleDPPPlanes[i][j] = v->TotalNumberOfSingleDPPPlanes[i][j] - 1; |
3974 | } |
3975 | } |
3976 | if (v->TotalNumberOfActiveDPP[i][j] > v->MaxNumDPP) { |
3977 | v->RequiredDISPCLK[i][j] = 0.0; |
3978 | v->DISPCLK_DPPCLK_Support[i][j] = true1; |
3979 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
3980 | v->ODMCombineEnablePerState[i][k] = dm_odm_combine_mode_disabled; |
3981 | if (v->SingleDPPViewportSizeSupportPerPlane[k] == false0 && v->WhenToDoMPCCombine != dm_mpc_never) { |
3982 | v->MPCCombine[i][j][k] = true1; |
3983 | v->NoOfDPP[i][j][k] = 2; |
3984 | v->RequiredDPPCLK[i][j][k] = v->MinDPPCLKUsingSingleDPP[k] * (1.0 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) / 2.0; |
3985 | } else { |
3986 | v->MPCCombine[i][j][k] = false0; |
3987 | v->NoOfDPP[i][j][k] = 1; |
3988 | v->RequiredDPPCLK[i][j][k] = v->MinDPPCLKUsingSingleDPP[k] * (1.0 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); |
3989 | } |
3990 | if (!(v->MaxDispclk[i] == v->MaxDispclk[v->soc.num_states - 1] && v->MaxDppclk[i] == v->MaxDppclk[v->soc.num_states - 1])) { |
3991 | v->PlaneRequiredDISPCLK = v->PixelClock[k] * (1.0 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) |
3992 | * (1.0 + v->DISPCLKRampingMargin / 100.0); |
3993 | } else { |
3994 | v->PlaneRequiredDISPCLK = v->PixelClock[k] * (1.0 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0); |
3995 | } |
3996 | v->RequiredDISPCLK[i][j] = dml_max(v->RequiredDISPCLK[i][j], v->PlaneRequiredDISPCLK); |
3997 | if ((v->MinDPPCLKUsingSingleDPP[k] / v->NoOfDPP[i][j][k] * (1.0 + v->DISPCLKDPPCLKDSCCLKDownSpreading / 100.0) |
3998 | > v->MaxDppclkRoundedDownToDFSGranularity) || (v->PlaneRequiredDISPCLK > v->MaxDispclkRoundedDownToDFSGranularity)) { |
3999 | v->DISPCLK_DPPCLK_Support[i][j] = false0; |
4000 | } |
4001 | } |
4002 | v->TotalNumberOfActiveDPP[i][j] = 0.0; |
4003 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
4004 | v->TotalNumberOfActiveDPP[i][j] = v->TotalNumberOfActiveDPP[i][j] + v->NoOfDPP[i][j][k]; |
4005 | } |
4006 | } |
4007 | v->RequiredDISPCLK[i][j] = dml_max(v->RequiredDISPCLK[i][j], v->WritebackRequiredDISPCLK); |
4008 | if (v->MaxDispclkRoundedDownToDFSGranularity < v->WritebackRequiredDISPCLK) { |
4009 | v->DISPCLK_DPPCLK_Support[i][j] = false0; |
4010 | } |
4011 | } |
4012 | } |
4013 | |
4014 | /*Total Available Pipes Support Check*/ |
4015 | |
4016 | for (i = 0; i < v->soc.num_states; i++) { |
4017 | for (j = 0; j < 2; j++) { |
4018 | if (v->TotalNumberOfActiveDPP[i][j] <= v->MaxNumDPP) { |
4019 | v->TotalAvailablePipesSupport[i][j] = true1; |
4020 | } else { |
4021 | v->TotalAvailablePipesSupport[i][j] = false0; |
4022 | } |
4023 | } |
4024 | } |
4025 | /*Display IO and DSC Support Check*/ |
4026 | |
4027 | v->NonsupportedDSCInputBPC = false0; |
4028 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
4029 | if (!(v->DSCInputBitPerComponent[k] == 12.0 |
4030 | || v->DSCInputBitPerComponent[k] == 10.0 |
4031 | || v->DSCInputBitPerComponent[k] == 8.0)) { |
4032 | v->NonsupportedDSCInputBPC = true1; |
4033 | } |
4034 | } |
4035 | |
4036 | /*Number Of DSC Slices*/ |
4037 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
4038 | if (v->BlendingAndTiming[k] == k) { |
4039 | if (v->PixelClockBackEnd[k] > 3200) { |
4040 | v->NumberOfDSCSlices[k] = dml_ceil(v->PixelClockBackEnd[k] / 400.0, 4.0); |
4041 | } else if (v->PixelClockBackEnd[k] > 1360) { |
4042 | v->NumberOfDSCSlices[k] = 8; |
4043 | } else if (v->PixelClockBackEnd[k] > 680) { |
4044 | v->NumberOfDSCSlices[k] = 4; |
4045 | } else if (v->PixelClockBackEnd[k] > 340) { |
4046 | v->NumberOfDSCSlices[k] = 2; |
4047 | } else { |
4048 | v->NumberOfDSCSlices[k] = 1; |
4049 | } |
4050 | } else { |
4051 | v->NumberOfDSCSlices[k] = 0; |
4052 | } |
4053 | } |
4054 | |
4055 | for (i = 0; i < v->soc.num_states; i++) { |
4056 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
4057 | v->RequiresDSC[i][k] = false0; |
4058 | v->RequiresFEC[i][k] = false0; |
4059 | if (v->BlendingAndTiming[k] == k) { |
4060 | if (v->Output[k] == dm_hdmi) { |
4061 | v->RequiresDSC[i][k] = false0; |
4062 | v->RequiresFEC[i][k] = false0; |
4063 | v->OutputBppPerState[i][k] = TruncToValidBPP( |
4064 | dml_min(600.0, v->PHYCLKPerState[i]) * 10, |
4065 | 3, |
4066 | v->HTotal[k], |
4067 | v->HActive[k], |
4068 | v->PixelClockBackEnd[k], |
4069 | v->ForcedOutputLinkBPP[k], |
4070 | false0, |
4071 | v->Output[k], |
4072 | v->OutputFormat[k], |
4073 | v->DSCInputBitPerComponent[k], |
4074 | v->NumberOfDSCSlices[k], |
4075 | v->AudioSampleRate[k], |
4076 | v->AudioSampleLayout[k], |
4077 | v->ODMCombineEnablePerState[i][k]); |
4078 | } else if (v->Output[k] == dm_dp || v->Output[k] == dm_edp) { |
4079 | if (v->DSCEnable[k] == true1) { |
4080 | v->RequiresDSC[i][k] = true1; |
4081 | v->LinkDSCEnable = true1; |
4082 | if (v->Output[k] == dm_dp) { |
4083 | v->RequiresFEC[i][k] = true1; |
4084 | } else { |
4085 | v->RequiresFEC[i][k] = false0; |
4086 | } |
4087 | } else { |
4088 | v->RequiresDSC[i][k] = false0; |
4089 | v->LinkDSCEnable = false0; |
4090 | v->RequiresFEC[i][k] = false0; |
4091 | } |
4092 | |
4093 | v->Outbpp = BPP_INVALID0; |
4094 | if (v->PHYCLKPerState[i] >= 270.0) { |
4095 | v->Outbpp = TruncToValidBPP( |
4096 | (1.0 - v->Downspreading / 100.0) * 2700, |
4097 | v->OutputLinkDPLanes[k], |
4098 | v->HTotal[k], |
4099 | v->HActive[k], |
4100 | v->PixelClockBackEnd[k], |
4101 | v->ForcedOutputLinkBPP[k], |
4102 | v->LinkDSCEnable, |
4103 | v->Output[k], |
4104 | v->OutputFormat[k], |
4105 | v->DSCInputBitPerComponent[k], |
4106 | v->NumberOfDSCSlices[k], |
4107 | v->AudioSampleRate[k], |
4108 | v->AudioSampleLayout[k], |
4109 | v->ODMCombineEnablePerState[i][k]); |
4110 | v->OutputBppPerState[i][k] = v->Outbpp; |
4111 | // TODO: Need some other way to handle this nonsense |
4112 | // v->OutputTypeAndRatePerState[i][k] = v->Output[k] & " HBR" |
4113 | } |
4114 | if (v->Outbpp == BPP_INVALID0 && v->PHYCLKPerState[i] >= 540.0) { |
4115 | v->Outbpp = TruncToValidBPP( |
4116 | (1.0 - v->Downspreading / 100.0) * 5400, |
4117 | v->OutputLinkDPLanes[k], |
4118 | v->HTotal[k], |
4119 | v->HActive[k], |
4120 | v->PixelClockBackEnd[k], |
4121 | v->ForcedOutputLinkBPP[k], |
4122 | v->LinkDSCEnable, |
4123 | v->Output[k], |
4124 | v->OutputFormat[k], |
4125 | v->DSCInputBitPerComponent[k], |
4126 | v->NumberOfDSCSlices[k], |
4127 | v->AudioSampleRate[k], |
4128 | v->AudioSampleLayout[k], |
4129 | v->ODMCombineEnablePerState[i][k]); |
4130 | v->OutputBppPerState[i][k] = v->Outbpp; |
4131 | // TODO: Need some other way to handle this nonsense |
4132 | // v->OutputTypeAndRatePerState[i][k] = v->Output[k] & " HBR2" |
4133 | } |
4134 | if (v->Outbpp == BPP_INVALID0 && v->PHYCLKPerState[i] >= 810.0) { |
4135 | v->Outbpp = TruncToValidBPP( |
4136 | (1.0 - v->Downspreading / 100.0) * 8100, |
4137 | v->OutputLinkDPLanes[k], |
4138 | v->HTotal[k], |
4139 | v->HActive[k], |
4140 | v->PixelClockBackEnd[k], |
4141 | v->ForcedOutputLinkBPP[k], |
4142 | v->LinkDSCEnable, |
4143 | v->Output[k], |
4144 | v->OutputFormat[k], |
4145 | v->DSCInputBitPerComponent[k], |
4146 | v->NumberOfDSCSlices[k], |
4147 | v->AudioSampleRate[k], |
4148 | v->AudioSampleLayout[k], |
4149 | v->ODMCombineEnablePerState[i][k]); |
4150 | if (v->Outbpp == BPP_INVALID0 && v->ForcedOutputLinkBPP[k] == 0) { |
4151 | //if (v->Outbpp == BPP_INVALID && v->DSCEnabled[k] == dm_dsc_enable_only_if_necessary && v->ForcedOutputLinkBPP[k] == 0) { |
4152 | v->RequiresDSC[i][k] = true1; |
4153 | v->LinkDSCEnable = true1; |
4154 | if (v->Output[k] == dm_dp) { |
4155 | v->RequiresFEC[i][k] = true1; |
4156 | } |
4157 | v->Outbpp = TruncToValidBPP( |
4158 | (1.0 - v->Downspreading / 100.0) * 8100, |
4159 | v->OutputLinkDPLanes[k], |
4160 | v->HTotal[k], |
4161 | v->HActive[k], |
4162 | v->PixelClockBackEnd[k], |
4163 | v->ForcedOutputLinkBPP[k], |
4164 | v->LinkDSCEnable, |
4165 | v->Output[k], |
4166 | v->OutputFormat[k], |
4167 | v->DSCInputBitPerComponent[k], |
4168 | v->NumberOfDSCSlices[k], |
4169 | v->AudioSampleRate[k], |
4170 | v->AudioSampleLayout[k], |
4171 | v->ODMCombineEnablePerState[i][k]); |
4172 | } |
4173 | v->OutputBppPerState[i][k] = v->Outbpp; |
4174 | // TODO: Need some other way to handle this nonsense |
4175 | // v->OutputTypeAndRatePerState[i][k] = v->Output[k] & " HBR3" |
4176 | } |
4177 | } |
4178 | } else { |
4179 | v->OutputBppPerState[i][k] = 0; |
4180 | } |
4181 | } |
4182 | } |
4183 | for (i = 0; i < v->soc.num_states; i++) { |
4184 | v->DIOSupport[i] = true1; |
4185 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
4186 | if (!v->skip_dio_check[k] && v->BlendingAndTiming[k] == k && (v->Output[k] == dm_dp || v->Output[k] == dm_edp || v->Output[k] == dm_hdmi) |
4187 | && (v->OutputBppPerState[i][k] == 0 |
4188 | || (v->OutputFormat[k] == dm_420 && v->Interlace[k] == true1 && v->ProgressiveToInterlaceUnitInOPP == true1))) { |
4189 | v->DIOSupport[i] = false0; |
4190 | } |
4191 | } |
4192 | } |
4193 | |
4194 | for (i = 0; i < v->soc.num_states; ++i) { |
4195 | v->ODMCombine4To1SupportCheckOK[i] = true1; |
4196 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
4197 | if (v->BlendingAndTiming[k] == k && v->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_4to1 |
4198 | && (v->ODMCombine4To1Supported == false0 || v->Output[k] == dm_dp || v->Output[k] == dm_edp || v->Output[k] == dm_hdmi)) { |
4199 | v->ODMCombine4To1SupportCheckOK[i] = false0; |
4200 | } |
4201 | } |
4202 | } |
4203 | |
4204 | /* Skip dscclk validation: as long as dispclk is supported, dscclk is also implicitly supported */ |
4205 | |
4206 | for (i = 0; i < v->soc.num_states; i++) { |
4207 | v->NotEnoughDSCUnits[i] = false0; |
4208 | v->TotalDSCUnitsRequired = 0.0; |
4209 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
4210 | if (v->RequiresDSC[i][k] == true1) { |
4211 | if (v->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_4to1) { |
4212 | v->TotalDSCUnitsRequired = v->TotalDSCUnitsRequired + 4.0; |
4213 | } else if (v->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_2to1) { |
4214 | v->TotalDSCUnitsRequired = v->TotalDSCUnitsRequired + 2.0; |
4215 | } else { |
4216 | v->TotalDSCUnitsRequired = v->TotalDSCUnitsRequired + 1.0; |
4217 | } |
4218 | } |
4219 | } |
4220 | if (v->TotalDSCUnitsRequired > v->NumberOfDSC) { |
4221 | v->NotEnoughDSCUnits[i] = true1; |
4222 | } |
4223 | } |
4224 | /*DSC Delay per state*/ |
4225 | |
4226 | for (i = 0; i < v->soc.num_states; i++) { |
4227 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
4228 | if (v->OutputBppPerState[i][k] == BPP_INVALID0) { |
4229 | v->BPP = 0.0; |
4230 | } else { |
4231 | v->BPP = v->OutputBppPerState[i][k]; |
4232 | } |
4233 | if (v->RequiresDSC[i][k] == true1 && v->BPP != 0.0) { |
4234 | if (v->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_disabled) { |
4235 | v->DSCDelayPerState[i][k] = dscceComputeDelay( |
4236 | v->DSCInputBitPerComponent[k], |
4237 | v->BPP, |
4238 | dml_ceil(1.0 * v->HActive[k] / v->NumberOfDSCSlices[k], 1.0), |
4239 | v->NumberOfDSCSlices[k], |
4240 | v->OutputFormat[k], |
4241 | v->Output[k]) + dscComputeDelay(v->OutputFormat[k], v->Output[k]); |
4242 | } else if (v->ODMCombineEnablePerState[i][k] == dm_odm_combine_mode_2to1) { |
4243 | v->DSCDelayPerState[i][k] = 2.0 |
4244 | * dscceComputeDelay( |
4245 | v->DSCInputBitPerComponent[k], |
4246 | v->BPP, |
4247 | dml_ceil(1.0 * v->HActive[k] / v->NumberOfDSCSlices[k], 1.0), |
4248 | v->NumberOfDSCSlices[k] / 2, |
4249 | v->OutputFormat[k], |
4250 | v->Output[k]) + dscComputeDelay(v->OutputFormat[k], v->Output[k]); |
4251 | } else { |
4252 | v->DSCDelayPerState[i][k] = 4.0 |
4253 | * (dscceComputeDelay( |
4254 | v->DSCInputBitPerComponent[k], |
4255 | v->BPP, |
4256 | dml_ceil(1.0 * v->HActive[k] / v->NumberOfDSCSlices[k], 1.0), |
4257 | v->NumberOfDSCSlices[k] / 4, |
4258 | v->OutputFormat[k], |
4259 | v->Output[k]) + dscComputeDelay(v->OutputFormat[k], v->Output[k])); |
4260 | } |
4261 | v->DSCDelayPerState[i][k] = v->DSCDelayPerState[i][k] * v->PixelClock[k] / v->PixelClockBackEnd[k]; |
4262 | } else { |
4263 | v->DSCDelayPerState[i][k] = 0.0; |
4264 | } |
4265 | } |
4266 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
4267 | for (m = 0; m <= v->NumberOfActivePlanes - 1; m++) { |
4268 | if (v->BlendingAndTiming[k] == m && v->RequiresDSC[i][m] == true1) { |
4269 | v->DSCDelayPerState[i][k] = v->DSCDelayPerState[i][m]; |
4270 | } |
4271 | } |
4272 | } |
4273 | } |
4274 | |
4275 | //Calculate Swath, DET Configuration, DCFCLKDeepSleep |
4276 | // |
4277 | for (i = 0; i < mode_lib->soc.num_states; ++i) { |
4278 | for (j = 0; j <= 1; ++j) { |
4279 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
4280 | v->RequiredDPPCLKThisState[k] = v->RequiredDPPCLK[i][j][k]; |
4281 | v->NoOfDPPThisState[k] = v->NoOfDPP[i][j][k]; |
4282 | v->ODMCombineEnableThisState[k] = v->ODMCombineEnablePerState[i][k]; |
4283 | } |
4284 | |
4285 | CalculateSwathAndDETConfiguration( |
4286 | false0, |
4287 | v->NumberOfActivePlanes, |
4288 | v->DETBufferSizeInKByte[0], |
4289 | v->MaximumSwathWidthLuma, |
4290 | v->MaximumSwathWidthChroma, |
4291 | v->SourceScan, |
4292 | v->SourcePixelFormat, |
4293 | v->SurfaceTiling, |
4294 | v->ViewportWidth, |
4295 | v->ViewportHeight, |
4296 | v->SurfaceWidthY, |
4297 | v->SurfaceWidthC, |
4298 | v->SurfaceHeightY, |
4299 | v->SurfaceHeightC, |
4300 | v->Read256BlockHeightY, |
4301 | v->Read256BlockHeightC, |
4302 | v->Read256BlockWidthY, |
4303 | v->Read256BlockWidthC, |
4304 | v->ODMCombineEnableThisState, |
4305 | v->BlendingAndTiming, |
4306 | v->BytePerPixelY, |
4307 | v->BytePerPixelC, |
4308 | v->BytePerPixelInDETY, |
4309 | v->BytePerPixelInDETC, |
4310 | v->HActive, |
4311 | v->HRatio, |
4312 | v->HRatioChroma, |
4313 | v->NoOfDPPThisState, |
4314 | v->swath_width_luma_ub_this_state, |
4315 | v->swath_width_chroma_ub_this_state, |
4316 | v->SwathWidthYThisState, |
4317 | v->SwathWidthCThisState, |
4318 | v->SwathHeightYThisState, |
4319 | v->SwathHeightCThisState, |
4320 | v->DETBufferSizeYThisState, |
4321 | v->DETBufferSizeCThisState, |
4322 | v->dummystring, |
4323 | &v->ViewportSizeSupport[i][j]); |
4324 | |
4325 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
4326 | v->swath_width_luma_ub_all_states[i][j][k] = v->swath_width_luma_ub_this_state[k]; |
4327 | v->swath_width_chroma_ub_all_states[i][j][k] = v->swath_width_chroma_ub_this_state[k]; |
4328 | v->SwathWidthYAllStates[i][j][k] = v->SwathWidthYThisState[k]; |
4329 | v->SwathWidthCAllStates[i][j][k] = v->SwathWidthCThisState[k]; |
4330 | v->SwathHeightYAllStates[i][j][k] = v->SwathHeightYThisState[k]; |
4331 | v->SwathHeightCAllStates[i][j][k] = v->SwathHeightCThisState[k]; |
4332 | v->DETBufferSizeYAllStates[i][j][k] = v->DETBufferSizeYThisState[k]; |
4333 | v->DETBufferSizeCAllStates[i][j][k] = v->DETBufferSizeCThisState[k]; |
4334 | } |
4335 | |
4336 | } |
4337 | } |
4338 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
4339 | v->cursor_bw[k] = v->NumberOfCursors[k] * v->CursorWidth[k][0] * v->CursorBPP[k][0] / 8.0 / (v->HTotal[k] / v->PixelClock[k]) * v->VRatio[k]; |
4340 | } |
4341 | |
4342 | for (i = 0; i < v->soc.num_states; i++) { |
4343 | for (j = 0; j < 2; j++) { |
4344 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
4345 | v->swath_width_luma_ub_this_state[k] = v->swath_width_luma_ub_all_states[i][j][k]; |
4346 | v->swath_width_chroma_ub_this_state[k] = v->swath_width_chroma_ub_all_states[i][j][k]; |
4347 | v->SwathWidthYThisState[k] = v->SwathWidthYAllStates[i][j][k]; |
4348 | v->SwathWidthCThisState[k] = v->SwathWidthCAllStates[i][j][k]; |
4349 | v->SwathHeightYThisState[k] = v->SwathHeightYAllStates[i][j][k]; |
4350 | v->SwathHeightCThisState[k] = v->SwathHeightCAllStates[i][j][k]; |
4351 | v->DETBufferSizeYThisState[k] = v->DETBufferSizeYAllStates[i][j][k]; |
4352 | v->DETBufferSizeCThisState[k] = v->DETBufferSizeCAllStates[i][j][k]; |
4353 | } |
4354 | |
4355 | v->TotalNumberOfDCCActiveDPP[i][j] = 0; |
4356 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
4357 | if (v->DCCEnable[k] == true1) { |
4358 | v->TotalNumberOfDCCActiveDPP[i][j] = v->TotalNumberOfDCCActiveDPP[i][j] + v->NoOfDPP[i][j][k]; |
4359 | } |
4360 | } |
4361 | |
4362 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
4363 | if (v->SourcePixelFormat[k] == dm_420_8 || v->SourcePixelFormat[k] == dm_420_10 || v->SourcePixelFormat[k] == dm_420_12 |
4364 | || v->SourcePixelFormat[k] == dm_rgbe_alpha) { |
4365 | |
4366 | if ((v->SourcePixelFormat[k] == dm_420_10 || v->SourcePixelFormat[k] == dm_420_12) && v->SourceScan[k] != dm_vert) { |
4367 | v->PTEBufferSizeInRequestsForLuma = (v->PTEBufferSizeInRequestsLuma + v->PTEBufferSizeInRequestsChroma) / 2; |
4368 | v->PTEBufferSizeInRequestsForChroma = v->PTEBufferSizeInRequestsForLuma; |
4369 | } else { |
4370 | v->PTEBufferSizeInRequestsForLuma = v->PTEBufferSizeInRequestsLuma; |
4371 | v->PTEBufferSizeInRequestsForChroma = v->PTEBufferSizeInRequestsChroma; |
4372 | } |
4373 | |
4374 | v->PDEAndMetaPTEBytesPerFrameC = CalculateVMAndRowBytes( |
4375 | mode_lib, |
4376 | v->DCCEnable[k], |
4377 | v->Read256BlockHeightC[k], |
4378 | v->Read256BlockWidthY[k], |
4379 | v->SourcePixelFormat[k], |
4380 | v->SurfaceTiling[k], |
4381 | v->BytePerPixelC[k], |
4382 | v->SourceScan[k], |
4383 | v->SwathWidthCThisState[k], |
4384 | v->ViewportHeightChroma[k], |
4385 | v->GPUVMEnable, |
4386 | v->HostVMEnable, |
4387 | v->HostVMMaxNonCachedPageTableLevels, |
4388 | v->GPUVMMinPageSize, |
4389 | v->HostVMMinPageSize, |
4390 | v->PTEBufferSizeInRequestsForChroma, |
4391 | v->PitchC[k], |
4392 | 0.0, |
4393 | &v->MacroTileWidthC[k], |
4394 | &v->MetaRowBytesC, |
4395 | &v->DPTEBytesPerRowC, |
4396 | &v->PTEBufferSizeNotExceededC[i][j][k], |
4397 | &v->dummyinteger7, |
4398 | &v->dpte_row_height_chroma[k], |
4399 | &v->dummyinteger28, |
4400 | &v->dummyinteger26, |
4401 | &v->dummyinteger23, |
4402 | &v->meta_row_height_chroma[k], |
4403 | &v->dummyinteger8, |
4404 | &v->dummyinteger9, |
4405 | &v->dummyinteger19, |
4406 | &v->dummyinteger20, |
4407 | &v->dummyinteger17, |
4408 | &v->dummyinteger10, |
4409 | &v->dummyinteger11); |
4410 | |
4411 | v->PrefetchLinesC[i][j][k] = CalculatePrefetchSourceLines( |
4412 | mode_lib, |
4413 | v->VRatioChroma[k], |
4414 | v->VTAPsChroma[k], |
4415 | v->Interlace[k], |
4416 | v->ProgressiveToInterlaceUnitInOPP, |
4417 | v->SwathHeightCThisState[k], |
4418 | v->ViewportYStartC[k], |
4419 | &v->PrefillC[k], |
4420 | &v->MaxNumSwC[k]); |
4421 | } else { |
4422 | v->PTEBufferSizeInRequestsForLuma = v->PTEBufferSizeInRequestsLuma + v->PTEBufferSizeInRequestsChroma; |
4423 | v->PTEBufferSizeInRequestsForChroma = 0; |
4424 | v->PDEAndMetaPTEBytesPerFrameC = 0.0; |
4425 | v->MetaRowBytesC = 0.0; |
4426 | v->DPTEBytesPerRowC = 0.0; |
4427 | v->PrefetchLinesC[i][j][k] = 0.0; |
4428 | v->PTEBufferSizeNotExceededC[i][j][k] = true1; |
4429 | } |
4430 | v->PDEAndMetaPTEBytesPerFrameY = CalculateVMAndRowBytes( |
4431 | mode_lib, |
4432 | v->DCCEnable[k], |
4433 | v->Read256BlockHeightY[k], |
4434 | v->Read256BlockWidthY[k], |
4435 | v->SourcePixelFormat[k], |
4436 | v->SurfaceTiling[k], |
4437 | v->BytePerPixelY[k], |
4438 | v->SourceScan[k], |
4439 | v->SwathWidthYThisState[k], |
4440 | v->ViewportHeight[k], |
4441 | v->GPUVMEnable, |
4442 | v->HostVMEnable, |
4443 | v->HostVMMaxNonCachedPageTableLevels, |
4444 | v->GPUVMMinPageSize, |
4445 | v->HostVMMinPageSize, |
4446 | v->PTEBufferSizeInRequestsForLuma, |
4447 | v->PitchY[k], |
4448 | v->DCCMetaPitchY[k], |
4449 | &v->MacroTileWidthY[k], |
4450 | &v->MetaRowBytesY, |
4451 | &v->DPTEBytesPerRowY, |
4452 | &v->PTEBufferSizeNotExceededY[i][j][k], |
4453 | v->dummyinteger4, |
4454 | &v->dpte_row_height[k], |
4455 | &v->dummyinteger29, |
4456 | &v->dummyinteger27, |
4457 | &v->dummyinteger24, |
4458 | &v->meta_row_height[k], |
4459 | &v->dummyinteger25, |
4460 | &v->dpte_group_bytes[k], |
4461 | &v->dummyinteger21, |
4462 | &v->dummyinteger22, |
4463 | &v->dummyinteger18, |
4464 | &v->dummyinteger5, |
4465 | &v->dummyinteger6); |
4466 | v->PrefetchLinesY[i][j][k] = CalculatePrefetchSourceLines( |
4467 | mode_lib, |
4468 | v->VRatio[k], |
4469 | v->vtaps[k], |
4470 | v->Interlace[k], |
4471 | v->ProgressiveToInterlaceUnitInOPP, |
4472 | v->SwathHeightYThisState[k], |
4473 | v->ViewportYStartY[k], |
4474 | &v->PrefillY[k], |
4475 | &v->MaxNumSwY[k]); |
4476 | v->PDEAndMetaPTEBytesPerFrame[i][j][k] = v->PDEAndMetaPTEBytesPerFrameY + v->PDEAndMetaPTEBytesPerFrameC; |
4477 | v->MetaRowBytes[i][j][k] = v->MetaRowBytesY + v->MetaRowBytesC; |
4478 | v->DPTEBytesPerRow[i][j][k] = v->DPTEBytesPerRowY + v->DPTEBytesPerRowC; |
4479 | |
4480 | CalculateRowBandwidth( |
4481 | v->GPUVMEnable, |
4482 | v->SourcePixelFormat[k], |
4483 | v->VRatio[k], |
4484 | v->VRatioChroma[k], |
4485 | v->DCCEnable[k], |
4486 | v->HTotal[k] / v->PixelClock[k], |
4487 | v->MetaRowBytesY, |
4488 | v->MetaRowBytesC, |
4489 | v->meta_row_height[k], |
4490 | v->meta_row_height_chroma[k], |
4491 | v->DPTEBytesPerRowY, |
4492 | v->DPTEBytesPerRowC, |
4493 | v->dpte_row_height[k], |
4494 | v->dpte_row_height_chroma[k], |
4495 | &v->meta_row_bandwidth[i][j][k], |
4496 | &v->dpte_row_bandwidth[i][j][k]); |
4497 | } |
4498 | v->UrgLatency[i] = CalculateUrgentLatency( |
4499 | v->UrgentLatencyPixelDataOnly, |
4500 | v->UrgentLatencyPixelMixedWithVMData, |
4501 | v->UrgentLatencyVMDataOnly, |
4502 | v->DoUrgentLatencyAdjustment, |
4503 | v->UrgentLatencyAdjustmentFabricClockComponent, |
4504 | v->UrgentLatencyAdjustmentFabricClockReference, |
4505 | v->FabricClockPerState[i]); |
4506 | |
4507 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
4508 | CalculateUrgentBurstFactor( |
4509 | v->swath_width_luma_ub_this_state[k], |
4510 | v->swath_width_chroma_ub_this_state[k], |
4511 | v->DETBufferSizeInKByte[0], |
4512 | v->SwathHeightYThisState[k], |
4513 | v->SwathHeightCThisState[k], |
4514 | v->HTotal[k] / v->PixelClock[k], |
4515 | v->UrgLatency[i], |
4516 | v->CursorBufferSize, |
4517 | v->CursorWidth[k][0], |
4518 | v->CursorBPP[k][0], |
4519 | v->VRatio[k], |
4520 | v->VRatioChroma[k], |
4521 | v->BytePerPixelInDETY[k], |
4522 | v->BytePerPixelInDETC[k], |
4523 | v->DETBufferSizeYThisState[k], |
4524 | v->DETBufferSizeCThisState[k], |
4525 | &v->UrgentBurstFactorCursor[k], |
4526 | &v->UrgentBurstFactorLuma[k], |
4527 | &v->UrgentBurstFactorChroma[k], |
4528 | &NotUrgentLatencyHiding[k]); |
4529 | } |
4530 | |
4531 | v->NotUrgentLatencyHiding[i][j] = false0; |
4532 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
4533 | if (NotUrgentLatencyHiding[k]) { |
4534 | v->NotUrgentLatencyHiding[i][j] = true1; |
4535 | } |
4536 | } |
4537 | |
4538 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
4539 | v->VActivePixelBandwidth[i][j][k] = v->ReadBandwidthLuma[k] * v->UrgentBurstFactorLuma[k] |
4540 | + v->ReadBandwidthChroma[k] * v->UrgentBurstFactorChroma[k]; |
4541 | v->VActiveCursorBandwidth[i][j][k] = v->cursor_bw[k] * v->UrgentBurstFactorCursor[k]; |
4542 | } |
4543 | |
4544 | v->TotalVActivePixelBandwidth[i][j] = 0; |
4545 | v->TotalVActiveCursorBandwidth[i][j] = 0; |
4546 | v->TotalMetaRowBandwidth[i][j] = 0; |
4547 | v->TotalDPTERowBandwidth[i][j] = 0; |
4548 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
4549 | v->TotalVActivePixelBandwidth[i][j] = v->TotalVActivePixelBandwidth[i][j] + v->VActivePixelBandwidth[i][j][k]; |
4550 | v->TotalVActiveCursorBandwidth[i][j] = v->TotalVActiveCursorBandwidth[i][j] + v->VActiveCursorBandwidth[i][j][k]; |
4551 | v->TotalMetaRowBandwidth[i][j] = v->TotalMetaRowBandwidth[i][j] + v->NoOfDPP[i][j][k] * v->meta_row_bandwidth[i][j][k]; |
4552 | v->TotalDPTERowBandwidth[i][j] = v->TotalDPTERowBandwidth[i][j] + v->NoOfDPP[i][j][k] * v->dpte_row_bandwidth[i][j][k]; |
4553 | } |
4554 | |
4555 | CalculateDCFCLKDeepSleep( |
4556 | mode_lib, |
4557 | v->NumberOfActivePlanes, |
4558 | v->BytePerPixelY, |
4559 | v->BytePerPixelC, |
4560 | v->VRatio, |
4561 | v->VRatioChroma, |
4562 | v->SwathWidthYThisState, |
4563 | v->SwathWidthCThisState, |
4564 | v->NoOfDPPThisState, |
4565 | v->HRatio, |
4566 | v->HRatioChroma, |
4567 | v->PixelClock, |
4568 | v->PSCL_FACTOR, |
4569 | v->PSCL_FACTOR_CHROMA, |
4570 | v->RequiredDPPCLKThisState, |
4571 | v->ReadBandwidthLuma, |
4572 | v->ReadBandwidthChroma, |
4573 | v->ReturnBusWidth, |
4574 | &v->ProjectedDCFCLKDeepSleep[i][j]); |
4575 | } |
4576 | } |
4577 | |
4578 | //Calculate Return BW |
4579 | |
4580 | for (i = 0; i < mode_lib->soc.num_states; ++i) { |
4581 | for (j = 0; j <= 1; ++j) { |
4582 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
4583 | if (v->BlendingAndTiming[k] == k) { |
4584 | if (v->WritebackEnable[k] == true1) { |
4585 | v->WritebackDelayTime[k] = v->WritebackLatency |
4586 | + CalculateWriteBackDelay( |
4587 | v->WritebackPixelFormat[k], |
4588 | v->WritebackHRatio[k], |
4589 | v->WritebackVRatio[k], |
4590 | v->WritebackVTaps[k], |
4591 | v->WritebackDestinationWidth[k], |
4592 | v->WritebackDestinationHeight[k], |
4593 | v->WritebackSourceHeight[k], |
4594 | v->HTotal[k]) / v->RequiredDISPCLK[i][j]; |
4595 | } else { |
4596 | v->WritebackDelayTime[k] = 0.0; |
4597 | } |
4598 | for (m = 0; m <= v->NumberOfActivePlanes - 1; m++) { |
4599 | if (v->BlendingAndTiming[m] == k && v->WritebackEnable[m] == true1) { |
4600 | v->WritebackDelayTime[k] = dml_max( |
4601 | v->WritebackDelayTime[k], |
4602 | v->WritebackLatency |
4603 | + CalculateWriteBackDelay( |
4604 | v->WritebackPixelFormat[m], |
4605 | v->WritebackHRatio[m], |
4606 | v->WritebackVRatio[m], |
4607 | v->WritebackVTaps[m], |
4608 | v->WritebackDestinationWidth[m], |
4609 | v->WritebackDestinationHeight[m], |
4610 | v->WritebackSourceHeight[m], |
4611 | v->HTotal[m]) / v->RequiredDISPCLK[i][j]); |
4612 | } |
4613 | } |
4614 | } |
4615 | } |
4616 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
4617 | for (m = 0; m <= v->NumberOfActivePlanes - 1; m++) { |
4618 | if (v->BlendingAndTiming[k] == m) { |
4619 | v->WritebackDelayTime[k] = v->WritebackDelayTime[m]; |
4620 | } |
4621 | } |
4622 | } |
4623 | v->MaxMaxVStartup[i][j] = 0; |
4624 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
4625 | v->MaximumVStartup[i][j][k] = v->VTotal[k] - v->VActive[k] |
4626 | - dml_max(1.0, dml_ceil(1.0 * v->WritebackDelayTime[k] / (v->HTotal[k] / v->PixelClock[k]), 1.0)); |
4627 | v->MaxMaxVStartup[i][j] = dml_max(v->MaxMaxVStartup[i][j], v->MaximumVStartup[i][j][k]); |
4628 | } |
4629 | } |
4630 | } |
4631 | |
4632 | ReorderingBytes = v->NumberOfChannels |
4633 | * dml_max3( |
4634 | v->UrgentOutOfOrderReturnPerChannelPixelDataOnly, |
4635 | v->UrgentOutOfOrderReturnPerChannelPixelMixedWithVMData, |
4636 | v->UrgentOutOfOrderReturnPerChannelVMDataOnly); |
4637 | v->FinalDRAMClockChangeLatency = (v->DRAMClockChangeLatencyOverride > 0 ? v->DRAMClockChangeLatencyOverride : v->DRAMClockChangeLatency); |
4638 | |
4639 | for (i = 0; i < mode_lib->soc.num_states; ++i) { |
4640 | for (j = 0; j <= 1; ++j) { |
4641 | v->DCFCLKState[i][j] = v->DCFCLKPerState[i]; |
4642 | } |
4643 | } |
4644 | |
4645 | if (v->UseMinimumRequiredDCFCLK == true1) { |
4646 | UseMinimumDCFCLK(mode_lib, v, MaxPrefetchMode, ReorderingBytes); |
4647 | |
4648 | if (v->ClampMinDCFCLK) { |
4649 | /* Clamp calculated values to actual minimum */ |
4650 | for (i = 0; i < mode_lib->soc.num_states; ++i) { |
4651 | for (j = 0; j <= 1; ++j) { |
4652 | if (v->DCFCLKState[i][j] < mode_lib->soc.min_dcfclk) { |
4653 | v->DCFCLKState[i][j] = mode_lib->soc.min_dcfclk; |
4654 | } |
4655 | } |
4656 | } |
4657 | } |
4658 | } |
4659 | |
4660 | for (i = 0; i < mode_lib->soc.num_states; ++i) { |
4661 | for (j = 0; j <= 1; ++j) { |
4662 | v->IdealSDPPortBandwidthPerState[i][j] = dml_min3( |
4663 | v->ReturnBusWidth * v->DCFCLKState[i][j], |
4664 | v->DRAMSpeedPerState[i] * v->NumberOfChannels * v->DRAMChannelWidth, |
4665 | v->FabricClockPerState[i] * v->FabricDatapathToDCNDataReturn); |
4666 | if (v->HostVMEnable != true1) { |
4667 | v->ReturnBWPerState[i][j] = v->IdealSDPPortBandwidthPerState[i][j] * v->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelDataOnly |
4668 | / 100; |
4669 | } else { |
4670 | v->ReturnBWPerState[i][j] = v->IdealSDPPortBandwidthPerState[i][j] |
4671 | * v->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData / 100; |
4672 | } |
4673 | } |
4674 | } |
4675 | |
4676 | //Re-ordering Buffer Support Check |
4677 | |
4678 | for (i = 0; i < mode_lib->soc.num_states; ++i) { |
4679 | for (j = 0; j <= 1; ++j) { |
4680 | if ((v->ROBBufferSizeInKByte - v->PixelChunkSizeInKByte) * 1024 / v->ReturnBWPerState[i][j] |
4681 | > (v->RoundTripPingLatencyCycles + 32) / v->DCFCLKState[i][j] + ReorderingBytes / v->ReturnBWPerState[i][j]) { |
4682 | v->ROBSupport[i][j] = true1; |
4683 | } else { |
4684 | v->ROBSupport[i][j] = false0; |
4685 | } |
4686 | } |
4687 | } |
4688 | |
4689 | //Vertical Active BW support check |
4690 | |
4691 | MaxTotalVActiveRDBandwidth = 0; |
4692 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
4693 | MaxTotalVActiveRDBandwidth = MaxTotalVActiveRDBandwidth + v->ReadBandwidthLuma[k] + v->ReadBandwidthChroma[k]; |
4694 | } |
4695 | |
4696 | for (i = 0; i < mode_lib->soc.num_states; ++i) { |
4697 | for (j = 0; j <= 1; ++j) { |
4698 | v->MaxTotalVerticalActiveAvailableBandwidth[i][j] = dml_min( |
4699 | v->IdealSDPPortBandwidthPerState[i][j] * v->MaxAveragePercentOfIdealSDPPortBWDisplayCanUseInNormalSystemOperation / 100, |
4700 | v->DRAMSpeedPerState[i] * v->NumberOfChannels * v->DRAMChannelWidth * v->MaxAveragePercentOfIdealDRAMBWDisplayCanUseInNormalSystemOperation |
4701 | / 100); |
4702 | if (MaxTotalVActiveRDBandwidth <= v->MaxTotalVerticalActiveAvailableBandwidth[i][j]) { |
4703 | v->TotalVerticalActiveBandwidthSupport[i][j] = true1; |
4704 | } else { |
4705 | v->TotalVerticalActiveBandwidthSupport[i][j] = false0; |
4706 | } |
4707 | } |
4708 | } |
4709 | |
4710 | //Prefetch Check |
4711 | |
4712 | for (i = 0; i < mode_lib->soc.num_states; ++i) { |
4713 | for (j = 0; j <= 1; ++j) { |
4714 | int NextPrefetchModeState = MinPrefetchMode; |
4715 | |
4716 | v->TimeCalc = 24 / v->ProjectedDCFCLKDeepSleep[i][j]; |
4717 | |
4718 | v->BandwidthWithoutPrefetchSupported[i][j] = true1; |
4719 | if (v->TotalVActivePixelBandwidth[i][j] + v->TotalVActiveCursorBandwidth[i][j] + v->TotalMetaRowBandwidth[i][j] + v->TotalDPTERowBandwidth[i][j] |
4720 | > v->ReturnBWPerState[i][j] || v->NotUrgentLatencyHiding[i][j]) { |
4721 | v->BandwidthWithoutPrefetchSupported[i][j] = false0; |
4722 | } |
4723 | |
4724 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
4725 | v->NoOfDPPThisState[k] = v->NoOfDPP[i][j][k]; |
4726 | v->swath_width_luma_ub_this_state[k] = v->swath_width_luma_ub_all_states[i][j][k]; |
4727 | v->swath_width_chroma_ub_this_state[k] = v->swath_width_chroma_ub_all_states[i][j][k]; |
4728 | v->SwathWidthYThisState[k] = v->SwathWidthYAllStates[i][j][k]; |
4729 | v->SwathWidthCThisState[k] = v->SwathWidthCAllStates[i][j][k]; |
4730 | v->SwathHeightYThisState[k] = v->SwathHeightYAllStates[i][j][k]; |
4731 | v->SwathHeightCThisState[k] = v->SwathHeightCAllStates[i][j][k]; |
4732 | v->DETBufferSizeYThisState[k] = v->DETBufferSizeYAllStates[i][j][k]; |
4733 | v->DETBufferSizeCThisState[k] = v->DETBufferSizeCAllStates[i][j][k]; |
4734 | v->ODMCombineEnabled[k] = v->ODMCombineEnablePerState[i][k]; |
4735 | } |
4736 | |
4737 | v->ExtraLatency = CalculateExtraLatency( |
4738 | v->RoundTripPingLatencyCycles, |
4739 | ReorderingBytes, |
4740 | v->DCFCLKState[i][j], |
4741 | v->TotalNumberOfActiveDPP[i][j], |
4742 | v->PixelChunkSizeInKByte, |
4743 | v->TotalNumberOfDCCActiveDPP[i][j], |
4744 | v->MetaChunkSize, |
4745 | v->ReturnBWPerState[i][j], |
4746 | v->GPUVMEnable, |
4747 | v->HostVMEnable, |
4748 | v->NumberOfActivePlanes, |
4749 | v->NoOfDPPThisState, |
4750 | v->dpte_group_bytes, |
4751 | v->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
4752 | v->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
4753 | v->HostVMMinPageSize, |
4754 | v->HostVMMaxNonCachedPageTableLevels); |
4755 | |
4756 | v->NextMaxVStartup = v->MaxMaxVStartup[i][j]; |
4757 | do { |
4758 | v->PrefetchModePerState[i][j] = NextPrefetchModeState; |
4759 | v->MaxVStartup = v->NextMaxVStartup; |
4760 | |
4761 | v->TWait = CalculateTWait(v->PrefetchModePerState[i][j], v->FinalDRAMClockChangeLatency, v->UrgLatency[i], v->SREnterPlusExitTime); |
4762 | |
4763 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
4764 | Pipe myPipe = { 0 }; |
4765 | |
4766 | myPipe.DPPCLK = v->RequiredDPPCLK[i][j][k]; |
4767 | myPipe.DISPCLK = v->RequiredDISPCLK[i][j]; |
4768 | myPipe.PixelClock = v->PixelClock[k]; |
4769 | myPipe.DCFCLKDeepSleep = v->ProjectedDCFCLKDeepSleep[i][j]; |
4770 | myPipe.DPPPerPlane = v->NoOfDPP[i][j][k]; |
4771 | myPipe.ScalerEnabled = v->ScalerEnabled[k]; |
4772 | myPipe.SourceScan = v->SourceScan[k]; |
4773 | myPipe.BlockWidth256BytesY = v->Read256BlockWidthY[k]; |
4774 | myPipe.BlockHeight256BytesY = v->Read256BlockHeightY[k]; |
4775 | myPipe.BlockWidth256BytesC = v->Read256BlockWidthC[k]; |
4776 | myPipe.BlockHeight256BytesC = v->Read256BlockHeightC[k]; |
4777 | myPipe.InterlaceEnable = v->Interlace[k]; |
4778 | myPipe.NumberOfCursors = v->NumberOfCursors[k]; |
4779 | myPipe.VBlank = v->VTotal[k] - v->VActive[k]; |
4780 | myPipe.HTotal = v->HTotal[k]; |
4781 | myPipe.DCCEnable = v->DCCEnable[k]; |
4782 | myPipe.ODMCombineEnabled = !!v->ODMCombineEnabled[k]; |
4783 | |
4784 | v->NoTimeForPrefetch[i][j][k] = CalculatePrefetchSchedule( |
4785 | mode_lib, |
4786 | v->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
4787 | v->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
4788 | &myPipe, |
4789 | v->DSCDelayPerState[i][k], |
4790 | v->DPPCLKDelaySubtotal + v->DPPCLKDelayCNVCFormater, |
4791 | v->DPPCLKDelaySCL, |
4792 | v->DPPCLKDelaySCLLBOnly, |
4793 | v->DPPCLKDelayCNVCCursor, |
4794 | v->DISPCLKDelaySubtotal, |
4795 | v->SwathWidthYThisState[k] / v->HRatio[k], |
4796 | v->OutputFormat[k], |
4797 | v->MaxInterDCNTileRepeaters, |
4798 | dml_min(v->MaxVStartup, v->MaximumVStartup[i][j][k]), |
4799 | v->MaximumVStartup[i][j][k], |
4800 | v->GPUVMMaxPageTableLevels, |
4801 | v->GPUVMEnable, |
4802 | v->HostVMEnable, |
4803 | v->HostVMMaxNonCachedPageTableLevels, |
4804 | v->HostVMMinPageSize, |
4805 | v->DynamicMetadataEnable[k], |
4806 | v->DynamicMetadataVMEnabled, |
4807 | v->DynamicMetadataLinesBeforeActiveRequired[k], |
4808 | v->DynamicMetadataTransmittedBytes[k], |
4809 | v->UrgLatency[i], |
4810 | v->ExtraLatency, |
4811 | v->TimeCalc, |
4812 | v->PDEAndMetaPTEBytesPerFrame[i][j][k], |
4813 | v->MetaRowBytes[i][j][k], |
4814 | v->DPTEBytesPerRow[i][j][k], |
4815 | v->PrefetchLinesY[i][j][k], |
4816 | v->SwathWidthYThisState[k], |
4817 | v->BytePerPixelY[k], |
4818 | v->PrefillY[k], |
4819 | v->MaxNumSwY[k], |
4820 | v->PrefetchLinesC[i][j][k], |
4821 | v->SwathWidthCThisState[k], |
4822 | v->BytePerPixelC[k], |
4823 | v->PrefillC[k], |
4824 | v->MaxNumSwC[k], |
4825 | v->swath_width_luma_ub_this_state[k], |
4826 | v->swath_width_chroma_ub_this_state[k], |
4827 | v->SwathHeightYThisState[k], |
4828 | v->SwathHeightCThisState[k], |
4829 | v->TWait, |
4830 | v->ProgressiveToInterlaceUnitInOPP, |
4831 | &v->DSTXAfterScaler[k], |
4832 | &v->DSTYAfterScaler[k], |
4833 | &v->LineTimesForPrefetch[k], |
4834 | &v->PrefetchBW[k], |
4835 | &v->LinesForMetaPTE[k], |
4836 | &v->LinesForMetaAndDPTERow[k], |
4837 | &v->VRatioPreY[i][j][k], |
4838 | &v->VRatioPreC[i][j][k], |
4839 | &v->RequiredPrefetchPixelDataBWLuma[i][j][k], |
4840 | &v->RequiredPrefetchPixelDataBWChroma[i][j][k], |
4841 | &v->NoTimeForDynamicMetadata[i][j][k], |
4842 | &v->Tno_bw[k], |
4843 | &v->prefetch_vmrow_bw[k], |
4844 | &v->Tdmdl_vm[k], |
4845 | &v->Tdmdl[k], |
4846 | &v->VUpdateOffsetPix[k], |
4847 | &v->VUpdateWidthPix[k], |
4848 | &v->VReadyOffsetPix[k]); |
4849 | } |
4850 | |
4851 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
4852 | CalculateUrgentBurstFactor( |
4853 | v->swath_width_luma_ub_this_state[k], |
4854 | v->swath_width_chroma_ub_this_state[k], |
4855 | v->DETBufferSizeInKByte[0], |
4856 | v->SwathHeightYThisState[k], |
4857 | v->SwathHeightCThisState[k], |
4858 | v->HTotal[k] / v->PixelClock[k], |
4859 | v->UrgentLatency, |
4860 | v->CursorBufferSize, |
4861 | v->CursorWidth[k][0], |
4862 | v->CursorBPP[k][0], |
4863 | v->VRatioPreY[i][j][k], |
4864 | v->VRatioPreC[i][j][k], |
4865 | v->BytePerPixelInDETY[k], |
4866 | v->BytePerPixelInDETC[k], |
4867 | v->DETBufferSizeYThisState[k], |
4868 | v->DETBufferSizeCThisState[k], |
4869 | &v->UrgentBurstFactorCursorPre[k], |
4870 | &v->UrgentBurstFactorLumaPre[k], |
4871 | &v->UrgentBurstFactorChromaPre[k], |
4872 | &v->NoUrgentLatencyHidingPre[k]); |
4873 | } |
4874 | |
4875 | v->MaximumReadBandwidthWithPrefetch = 0.0; |
4876 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
4877 | v->cursor_bw_pre[k] = v->NumberOfCursors[k] * v->CursorWidth[k][0] * v->CursorBPP[k][0] / 8.0 / (v->HTotal[k] / v->PixelClock[k]) |
4878 | * v->VRatioPreY[i][j][k]; |
4879 | |
4880 | v->MaximumReadBandwidthWithPrefetch = v->MaximumReadBandwidthWithPrefetch |
4881 | + dml_max4( |
4882 | v->VActivePixelBandwidth[i][j][k], |
4883 | v->VActiveCursorBandwidth[i][j][k] |
4884 | + v->NoOfDPP[i][j][k] * (v->meta_row_bandwidth[i][j][k] + v->dpte_row_bandwidth[i][j][k]), |
4885 | v->NoOfDPP[i][j][k] * v->prefetch_vmrow_bw[k], |
4886 | v->NoOfDPP[i][j][k] |
4887 | * (v->RequiredPrefetchPixelDataBWLuma[i][j][k] * v->UrgentBurstFactorLumaPre[k] |
4888 | + v->RequiredPrefetchPixelDataBWChroma[i][j][k] |
4889 | * v->UrgentBurstFactorChromaPre[k]) |
4890 | + v->cursor_bw_pre[k] * v->UrgentBurstFactorCursorPre[k]); |
4891 | } |
4892 | |
4893 | v->NotEnoughUrgentLatencyHidingPre = false0; |
4894 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
4895 | if (v->NoUrgentLatencyHidingPre[k] == true1) { |
4896 | v->NotEnoughUrgentLatencyHidingPre = true1; |
4897 | } |
4898 | } |
4899 | |
4900 | v->PrefetchSupported[i][j] = true1; |
4901 | if (v->BandwidthWithoutPrefetchSupported[i][j] == false0 || v->MaximumReadBandwidthWithPrefetch > v->ReturnBWPerState[i][j] |
4902 | || v->NotEnoughUrgentLatencyHidingPre == 1) { |
4903 | v->PrefetchSupported[i][j] = false0; |
4904 | } |
4905 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
4906 | if (v->LineTimesForPrefetch[k] < 2.0 || v->LinesForMetaPTE[k] >= 32.0 || v->LinesForMetaAndDPTERow[k] >= 16.0 |
4907 | || v->NoTimeForPrefetch[i][j][k] == true1) { |
4908 | v->PrefetchSupported[i][j] = false0; |
4909 | } |
4910 | } |
4911 | |
4912 | v->DynamicMetadataSupported[i][j] = true1; |
4913 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
4914 | if (v->NoTimeForDynamicMetadata[i][j][k] == true1) { |
4915 | v->DynamicMetadataSupported[i][j] = false0; |
4916 | } |
4917 | } |
4918 | |
4919 | v->VRatioInPrefetchSupported[i][j] = true1; |
4920 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
4921 | if (v->VRatioPreY[i][j][k] > 4.0 || v->VRatioPreC[i][j][k] > 4.0 || v->NoTimeForPrefetch[i][j][k] == true1) { |
4922 | v->VRatioInPrefetchSupported[i][j] = false0; |
4923 | } |
4924 | } |
4925 | v->AnyLinesForVMOrRowTooLarge = false0; |
4926 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
4927 | if (v->LinesForMetaAndDPTERow[k] >= 16 || v->LinesForMetaPTE[k] >= 32) { |
4928 | v->AnyLinesForVMOrRowTooLarge = true1; |
4929 | } |
4930 | } |
4931 | |
4932 | if (v->PrefetchSupported[i][j] == true1 && v->VRatioInPrefetchSupported[i][j] == true1) { |
4933 | v->BandwidthAvailableForImmediateFlip = v->ReturnBWPerState[i][j]; |
4934 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
4935 | v->BandwidthAvailableForImmediateFlip = v->BandwidthAvailableForImmediateFlip |
4936 | - dml_max( |
4937 | v->VActivePixelBandwidth[i][j][k] + v->VActiveCursorBandwidth[i][j][k], |
4938 | v->NoOfDPP[i][j][k] |
4939 | * (v->RequiredPrefetchPixelDataBWLuma[i][j][k] * v->UrgentBurstFactorLumaPre[k] |
4940 | + v->RequiredPrefetchPixelDataBWChroma[i][j][k] |
4941 | * v->UrgentBurstFactorChromaPre[k]) |
4942 | + v->cursor_bw_pre[k] * v->UrgentBurstFactorCursorPre[k]); |
4943 | } |
4944 | v->TotImmediateFlipBytes = 0.0; |
4945 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
4946 | v->TotImmediateFlipBytes = v->TotImmediateFlipBytes + v->NoOfDPP[i][j][k] * v->PDEAndMetaPTEBytesPerFrame[i][j][k] |
4947 | + v->MetaRowBytes[i][j][k] + v->DPTEBytesPerRow[i][j][k]; |
4948 | } |
4949 | |
4950 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
4951 | CalculateFlipSchedule( |
4952 | mode_lib, |
4953 | v->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
4954 | v->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
4955 | v->ExtraLatency, |
4956 | v->UrgLatency[i], |
4957 | v->GPUVMMaxPageTableLevels, |
4958 | v->HostVMEnable, |
4959 | v->HostVMMaxNonCachedPageTableLevels, |
4960 | v->GPUVMEnable, |
4961 | v->HostVMMinPageSize, |
4962 | v->PDEAndMetaPTEBytesPerFrame[i][j][k], |
4963 | v->MetaRowBytes[i][j][k], |
4964 | v->DPTEBytesPerRow[i][j][k], |
4965 | v->BandwidthAvailableForImmediateFlip, |
4966 | v->TotImmediateFlipBytes, |
4967 | v->SourcePixelFormat[k], |
4968 | v->HTotal[k] / v->PixelClock[k], |
4969 | v->VRatio[k], |
4970 | v->VRatioChroma[k], |
4971 | v->Tno_bw[k], |
4972 | v->DCCEnable[k], |
4973 | v->dpte_row_height[k], |
4974 | v->meta_row_height[k], |
4975 | v->dpte_row_height_chroma[k], |
4976 | v->meta_row_height_chroma[k], |
4977 | &v->DestinationLinesToRequestVMInImmediateFlip[k], |
4978 | &v->DestinationLinesToRequestRowInImmediateFlip[k], |
4979 | &v->final_flip_bw[k], |
4980 | &v->ImmediateFlipSupportedForPipe[k]); |
4981 | } |
4982 | v->total_dcn_read_bw_with_flip = 0.0; |
4983 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
4984 | v->total_dcn_read_bw_with_flip = v->total_dcn_read_bw_with_flip |
4985 | + dml_max3( |
4986 | v->NoOfDPP[i][j][k] * v->prefetch_vmrow_bw[k], |
4987 | v->NoOfDPP[i][j][k] * v->final_flip_bw[k] + v->VActivePixelBandwidth[i][j][k] |
4988 | + v->VActiveCursorBandwidth[i][j][k], |
4989 | v->NoOfDPP[i][j][k] |
4990 | * (v->final_flip_bw[k] |
4991 | + v->RequiredPrefetchPixelDataBWLuma[i][j][k] |
4992 | * v->UrgentBurstFactorLumaPre[k] |
4993 | + v->RequiredPrefetchPixelDataBWChroma[i][j][k] |
4994 | * v->UrgentBurstFactorChromaPre[k]) |
4995 | + v->cursor_bw_pre[k] * v->UrgentBurstFactorCursorPre[k]); |
4996 | } |
4997 | v->ImmediateFlipSupportedForState[i][j] = true1; |
4998 | if (v->total_dcn_read_bw_with_flip > v->ReturnBWPerState[i][j]) { |
4999 | v->ImmediateFlipSupportedForState[i][j] = false0; |
5000 | } |
5001 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
5002 | if (v->ImmediateFlipSupportedForPipe[k] == false0) { |
5003 | v->ImmediateFlipSupportedForState[i][j] = false0; |
5004 | } |
5005 | } |
5006 | } else { |
5007 | v->ImmediateFlipSupportedForState[i][j] = false0; |
5008 | } |
5009 | if (v->MaxVStartup <= 13 || v->AnyLinesForVMOrRowTooLarge == false0) { |
5010 | v->NextMaxVStartup = v->MaxMaxVStartup[i][j]; |
5011 | NextPrefetchModeState = NextPrefetchModeState + 1; |
5012 | } else { |
5013 | v->NextMaxVStartup = v->NextMaxVStartup - 1; |
5014 | } |
5015 | } while (!((v->PrefetchSupported[i][j] == true1 && v->DynamicMetadataSupported[i][j] == true1 && v->VRatioInPrefetchSupported[i][j] == true1 |
5016 | && ((v->HostVMEnable == false0 && v->ImmediateFlipRequirement[0] != dm_immediate_flip_required) |
5017 | || v->ImmediateFlipSupportedForState[i][j] == true1)) |
5018 | || (v->NextMaxVStartup == v->MaxMaxVStartup[i][j] && NextPrefetchModeState > MaxPrefetchMode))); |
5019 | |
5020 | CalculateWatermarksAndDRAMSpeedChangeSupport( |
5021 | mode_lib, |
5022 | v->PrefetchModePerState[i][j], |
5023 | v->NumberOfActivePlanes, |
5024 | v->MaxLineBufferLines, |
5025 | v->LineBufferSize, |
5026 | v->DPPOutputBufferPixels, |
5027 | v->DETBufferSizeInKByte[0], |
5028 | v->WritebackInterfaceBufferSize, |
5029 | v->DCFCLKState[i][j], |
5030 | v->ReturnBWPerState[i][j], |
5031 | v->GPUVMEnable, |
5032 | v->dpte_group_bytes, |
5033 | v->MetaChunkSize, |
5034 | v->UrgLatency[i], |
5035 | v->ExtraLatency, |
5036 | v->WritebackLatency, |
5037 | v->WritebackChunkSize, |
5038 | v->SOCCLKPerState[i], |
5039 | v->FinalDRAMClockChangeLatency, |
5040 | v->SRExitTime, |
5041 | v->SREnterPlusExitTime, |
5042 | v->ProjectedDCFCLKDeepSleep[i][j], |
5043 | v->NoOfDPPThisState, |
5044 | v->DCCEnable, |
5045 | v->RequiredDPPCLKThisState, |
5046 | v->DETBufferSizeYThisState, |
5047 | v->DETBufferSizeCThisState, |
5048 | v->SwathHeightYThisState, |
5049 | v->SwathHeightCThisState, |
5050 | v->LBBitPerPixel, |
5051 | v->SwathWidthYThisState, |
5052 | v->SwathWidthCThisState, |
5053 | v->HRatio, |
5054 | v->HRatioChroma, |
5055 | v->vtaps, |
5056 | v->VTAPsChroma, |
5057 | v->VRatio, |
5058 | v->VRatioChroma, |
5059 | v->HTotal, |
5060 | v->PixelClock, |
5061 | v->BlendingAndTiming, |
5062 | v->BytePerPixelInDETY, |
5063 | v->BytePerPixelInDETC, |
5064 | v->DSTXAfterScaler, |
5065 | v->DSTYAfterScaler, |
5066 | v->WritebackEnable, |
5067 | v->WritebackPixelFormat, |
5068 | v->WritebackDestinationWidth, |
5069 | v->WritebackDestinationHeight, |
5070 | v->WritebackSourceHeight, |
5071 | &v->DRAMClockChangeSupport[i][j], |
5072 | &v->UrgentWatermark, |
5073 | &v->WritebackUrgentWatermark, |
5074 | &v->DRAMClockChangeWatermark, |
5075 | &v->WritebackDRAMClockChangeWatermark, |
5076 | &v->StutterExitWatermark, |
5077 | &v->StutterEnterPlusExitWatermark, |
5078 | &v->MinActiveDRAMClockChangeLatencySupported); |
5079 | } |
5080 | } |
5081 | |
5082 | /*PTE Buffer Size Check*/ |
5083 | |
5084 | for (i = 0; i < v->soc.num_states; i++) { |
5085 | for (j = 0; j < 2; j++) { |
5086 | v->PTEBufferSizeNotExceeded[i][j] = true1; |
5087 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
5088 | if (v->PTEBufferSizeNotExceededY[i][j][k] == false0 || v->PTEBufferSizeNotExceededC[i][j][k] == false0) { |
5089 | v->PTEBufferSizeNotExceeded[i][j] = false0; |
5090 | } |
5091 | } |
5092 | } |
5093 | } |
5094 | /*Cursor Support Check*/ |
5095 | |
5096 | v->CursorSupport = true1; |
5097 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
5098 | if (v->CursorWidth[k][0] > 0.0) { |
5099 | if (v->CursorBPP[k][0] == 64 && v->Cursor64BppSupport == false0) { |
5100 | v->CursorSupport = false0; |
5101 | } |
5102 | } |
5103 | } |
5104 | /*Valid Pitch Check*/ |
5105 | |
5106 | v->PitchSupport = true1; |
5107 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
5108 | v->AlignedYPitch[k] = dml_ceil(dml_max(v->PitchY[k], v->SurfaceWidthY[k]), v->MacroTileWidthY[k]); |
5109 | if (v->DCCEnable[k] == true1) { |
5110 | v->AlignedDCCMetaPitchY[k] = dml_ceil(dml_max(v->DCCMetaPitchY[k], v->SurfaceWidthY[k]), 64.0 * v->Read256BlockWidthY[k]); |
5111 | } else { |
5112 | v->AlignedDCCMetaPitchY[k] = v->DCCMetaPitchY[k]; |
5113 | } |
5114 | if (v->SourcePixelFormat[k] != dm_444_64 && v->SourcePixelFormat[k] != dm_444_32 && v->SourcePixelFormat[k] != dm_444_16 && v->SourcePixelFormat[k] != dm_mono_16 |
5115 | && v->SourcePixelFormat[k] != dm_rgbe && v->SourcePixelFormat[k] != dm_mono_8) { |
5116 | v->AlignedCPitch[k] = dml_ceil(dml_max(v->PitchC[k], v->SurfaceWidthC[k]), v->MacroTileWidthC[k]); |
5117 | if (v->DCCEnable[k] == true1) { |
5118 | v->AlignedDCCMetaPitchC[k] = dml_ceil(dml_max(v->DCCMetaPitchC[k], v->SurfaceWidthC[k]), 64.0 * v->Read256BlockWidthC[k]); |
5119 | } else { |
5120 | v->AlignedDCCMetaPitchC[k] = v->DCCMetaPitchC[k]; |
5121 | } |
5122 | } else { |
5123 | v->AlignedCPitch[k] = v->PitchC[k]; |
5124 | v->AlignedDCCMetaPitchC[k] = v->DCCMetaPitchC[k]; |
5125 | } |
5126 | if (v->AlignedYPitch[k] > v->PitchY[k] || v->AlignedCPitch[k] > v->PitchC[k] || v->AlignedDCCMetaPitchY[k] > v->DCCMetaPitchY[k] |
5127 | || v->AlignedDCCMetaPitchC[k] > v->DCCMetaPitchC[k]) { |
5128 | v->PitchSupport = false0; |
5129 | } |
5130 | } |
5131 | |
5132 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
5133 | if (v->ViewportWidth[k] > v->SurfaceWidthY[k] || v->ViewportHeight[k] > v->SurfaceHeightY[k]) |
5134 | ViewportExceedsSurface = true1; |
5135 | |
5136 | if (v->SourcePixelFormat[k] != dm_444_64 && v->SourcePixelFormat[k] != dm_444_32 && v->SourcePixelFormat[k] != dm_444_16 |
5137 | && v->SourcePixelFormat[k] != dm_444_8 && v->SourcePixelFormat[k] != dm_rgbe) { |
5138 | if (v->ViewportWidthChroma[k] > v->SurfaceWidthC[k] || v->ViewportHeightChroma[k] > v->SurfaceHeightC[k]) { |
5139 | ViewportExceedsSurface = true1; |
5140 | } |
5141 | } |
5142 | } |
5143 | /*Mode Support, Voltage State and SOC Configuration*/ |
5144 | |
5145 | for (i = v->soc.num_states - 1; i >= 0; i--) { |
5146 | for (j = 0; j < 2; j++) { |
5147 | if (v->ScaleRatioAndTapsSupport == 1 && v->SourceFormatPixelAndScanSupport == 1 && v->ViewportSizeSupport[i][j] == 1 |
5148 | && v->DIOSupport[i] == 1 && v->ODMCombine4To1SupportCheckOK[i] == 1 |
5149 | && v->NotEnoughDSCUnits[i] == 0 |
5150 | && v->DTBCLKRequiredMoreThanSupported[i] == 0 |
5151 | && v->ROBSupport[i][j] == 1 && v->DISPCLK_DPPCLK_Support[i][j] == 1 && v->TotalAvailablePipesSupport[i][j] == 1 |
5152 | && EnoughWritebackUnits == 1 && WritebackModeSupport == 1 |
5153 | && v->WritebackLatencySupport == 1 && v->WritebackScaleRatioAndTapsSupport == 1 && v->CursorSupport == 1 && v->PitchSupport == 1 |
5154 | && ViewportExceedsSurface == 0 && v->PrefetchSupported[i][j] == 1 && v->DynamicMetadataSupported[i][j] == 1 |
5155 | && v->TotalVerticalActiveBandwidthSupport[i][j] == 1 && v->VRatioInPrefetchSupported[i][j] == 1 |
5156 | && v->PTEBufferSizeNotExceeded[i][j] == 1 && v->NonsupportedDSCInputBPC == 0 |
5157 | && ((v->HostVMEnable == 0 && v->ImmediateFlipRequirement[0] != dm_immediate_flip_required) |
5158 | || v->ImmediateFlipSupportedForState[i][j] == true1)) { |
5159 | v->ModeSupport[i][j] = true1; |
5160 | } else { |
5161 | v->ModeSupport[i][j] = false0; |
5162 | } |
5163 | } |
5164 | } |
5165 | { |
5166 | unsigned int MaximumMPCCombine = 0; |
5167 | for (i = v->soc.num_states; i >= 0; i--) { |
5168 | if (i == v->soc.num_states || v->ModeSupport[i][0] == true1 || v->ModeSupport[i][1] == true1) { |
5169 | v->VoltageLevel = i; |
5170 | v->ModeIsSupported = v->ModeSupport[i][0] == true1 || v->ModeSupport[i][1] == true1; |
5171 | if (v->ModeSupport[i][1] == true1) { |
5172 | MaximumMPCCombine = 1; |
5173 | } else { |
5174 | MaximumMPCCombine = 0; |
5175 | } |
5176 | } |
5177 | } |
5178 | v->ImmediateFlipSupport = v->ImmediateFlipSupportedForState[v->VoltageLevel][MaximumMPCCombine]; |
5179 | for (k = 0; k <= v->NumberOfActivePlanes - 1; k++) { |
5180 | v->MPCCombineEnable[k] = v->MPCCombine[v->VoltageLevel][MaximumMPCCombine][k]; |
5181 | v->DPPPerPlane[k] = v->NoOfDPP[v->VoltageLevel][MaximumMPCCombine][k]; |
5182 | } |
5183 | v->DCFCLK = v->DCFCLKState[v->VoltageLevel][MaximumMPCCombine]; |
5184 | v->DRAMSpeed = v->DRAMSpeedPerState[v->VoltageLevel]; |
5185 | v->FabricClock = v->FabricClockPerState[v->VoltageLevel]; |
5186 | v->SOCCLK = v->SOCCLKPerState[v->VoltageLevel]; |
5187 | v->ReturnBW = v->ReturnBWPerState[v->VoltageLevel][MaximumMPCCombine]; |
5188 | v->maxMpcComb = MaximumMPCCombine; |
5189 | } |
5190 | } |
5191 | |
5192 | static void CalculateWatermarksAndDRAMSpeedChangeSupport( |
5193 | struct display_mode_lib *mode_lib, |
5194 | unsigned int PrefetchMode, |
5195 | unsigned int NumberOfActivePlanes, |
5196 | unsigned int MaxLineBufferLines, |
5197 | unsigned int LineBufferSize, |
5198 | unsigned int DPPOutputBufferPixels, |
5199 | unsigned int DETBufferSizeInKByte, |
5200 | unsigned int WritebackInterfaceBufferSize, |
5201 | double DCFCLK, |
5202 | double ReturnBW, |
5203 | bool_Bool GPUVMEnable, |
5204 | unsigned int dpte_group_bytes[], |
5205 | unsigned int MetaChunkSize, |
5206 | double UrgentLatency, |
5207 | double ExtraLatency, |
5208 | double WritebackLatency, |
5209 | double WritebackChunkSize, |
5210 | double SOCCLK, |
5211 | double DRAMClockChangeLatency, |
5212 | double SRExitTime, |
5213 | double SREnterPlusExitTime, |
5214 | double DCFCLKDeepSleep, |
5215 | unsigned int DPPPerPlane[], |
5216 | bool_Bool DCCEnable[], |
5217 | double DPPCLK[], |
5218 | unsigned int DETBufferSizeY[], |
5219 | unsigned int DETBufferSizeC[], |
5220 | unsigned int SwathHeightY[], |
5221 | unsigned int SwathHeightC[], |
5222 | unsigned int LBBitPerPixel[], |
5223 | double SwathWidthY[], |
5224 | double SwathWidthC[], |
5225 | double HRatio[], |
5226 | double HRatioChroma[], |
5227 | unsigned int vtaps[], |
5228 | unsigned int VTAPsChroma[], |
5229 | double VRatio[], |
5230 | double VRatioChroma[], |
5231 | unsigned int HTotal[], |
5232 | double PixelClock[], |
5233 | unsigned int BlendingAndTiming[], |
5234 | double BytePerPixelDETY[], |
5235 | double BytePerPixelDETC[], |
5236 | double DSTXAfterScaler[], |
5237 | double DSTYAfterScaler[], |
5238 | bool_Bool WritebackEnable[], |
5239 | enum source_format_class WritebackPixelFormat[], |
5240 | double WritebackDestinationWidth[], |
5241 | double WritebackDestinationHeight[], |
5242 | double WritebackSourceHeight[], |
5243 | enum clock_change_support *DRAMClockChangeSupport, |
5244 | double *UrgentWatermark, |
5245 | double *WritebackUrgentWatermark, |
5246 | double *DRAMClockChangeWatermark, |
5247 | double *WritebackDRAMClockChangeWatermark, |
5248 | double *StutterExitWatermark, |
5249 | double *StutterEnterPlusExitWatermark, |
5250 | double *MinActiveDRAMClockChangeLatencySupported) |
5251 | { |
5252 | double EffectiveLBLatencyHidingY = 0; |
5253 | double EffectiveLBLatencyHidingC = 0; |
5254 | double LinesInDETY[DC__NUM_DPP__MAX8] = { 0 }; |
5255 | double LinesInDETC = 0; |
5256 | unsigned int LinesInDETYRoundedDownToSwath[DC__NUM_DPP__MAX8] = { 0 }; |
5257 | unsigned int LinesInDETCRoundedDownToSwath = 0; |
5258 | double FullDETBufferingTimeY[DC__NUM_DPP__MAX8] = { 0 }; |
5259 | double FullDETBufferingTimeC = 0; |
5260 | double ActiveDRAMClockChangeLatencyMarginY = 0; |
5261 | double ActiveDRAMClockChangeLatencyMarginC = 0; |
5262 | double WritebackDRAMClockChangeLatencyMargin = 0; |
5263 | double PlaneWithMinActiveDRAMClockChangeMargin = 0; |
5264 | double SecondMinActiveDRAMClockChangeMarginOneDisplayInVBLank = 0; |
5265 | double FullDETBufferingTimeYStutterCriticalPlane = 0; |
5266 | double TimeToFinishSwathTransferStutterCriticalPlane = 0; |
5267 | double WritebackDRAMClockChangeLatencyHiding = 0; |
5268 | unsigned int k, j; |
5269 | |
5270 | mode_lib->vba.TotalActiveDPP = 0; |
5271 | mode_lib->vba.TotalDCCActiveDPP = 0; |
5272 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
5273 | mode_lib->vba.TotalActiveDPP = mode_lib->vba.TotalActiveDPP + DPPPerPlane[k]; |
5274 | if (DCCEnable[k] == true1) { |
5275 | mode_lib->vba.TotalDCCActiveDPP = mode_lib->vba.TotalDCCActiveDPP + DPPPerPlane[k]; |
5276 | } |
5277 | } |
5278 | |
5279 | *UrgentWatermark = UrgentLatency + ExtraLatency; |
5280 | |
5281 | *DRAMClockChangeWatermark = DRAMClockChangeLatency + *UrgentWatermark; |
5282 | |
5283 | mode_lib->vba.TotalActiveWriteback = 0; |
5284 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
5285 | if (WritebackEnable[k] == true1) { |
5286 | mode_lib->vba.TotalActiveWriteback = mode_lib->vba.TotalActiveWriteback + 1; |
5287 | } |
5288 | } |
5289 | |
5290 | if (mode_lib->vba.TotalActiveWriteback <= 1) { |
5291 | *WritebackUrgentWatermark = WritebackLatency; |
5292 | } else { |
5293 | *WritebackUrgentWatermark = WritebackLatency + WritebackChunkSize * 1024.0 / 32.0 / SOCCLK; |
5294 | } |
5295 | |
5296 | if (mode_lib->vba.TotalActiveWriteback <= 1) { |
5297 | *WritebackDRAMClockChangeWatermark = DRAMClockChangeLatency + WritebackLatency; |
5298 | } else { |
5299 | *WritebackDRAMClockChangeWatermark = DRAMClockChangeLatency + WritebackLatency + WritebackChunkSize * 1024.0 / 32.0 / SOCCLK; |
5300 | } |
5301 | |
5302 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
5303 | |
5304 | mode_lib->vba.LBLatencyHidingSourceLinesY = dml_min((double) MaxLineBufferLines, dml_floor(LineBufferSize / LBBitPerPixel[k] / (SwathWidthY[k] / dml_max(HRatio[k], 1.0)), 1)) - (vtaps[k] - 1); |
5305 | |
5306 | mode_lib->vba.LBLatencyHidingSourceLinesC = dml_min((double) MaxLineBufferLines, dml_floor(LineBufferSize / LBBitPerPixel[k] / (SwathWidthC[k] / dml_max(HRatioChroma[k], 1.0)), 1)) - (VTAPsChroma[k] - 1); |
5307 | |
5308 | EffectiveLBLatencyHidingY = mode_lib->vba.LBLatencyHidingSourceLinesY / VRatio[k] * (HTotal[k] / PixelClock[k]); |
5309 | |
5310 | EffectiveLBLatencyHidingC = mode_lib->vba.LBLatencyHidingSourceLinesC / VRatioChroma[k] * (HTotal[k] / PixelClock[k]); |
5311 | |
5312 | LinesInDETY[k] = (double) DETBufferSizeY[k] / BytePerPixelDETY[k] / SwathWidthY[k]; |
5313 | LinesInDETYRoundedDownToSwath[k] = dml_floor(LinesInDETY[k], SwathHeightY[k]); |
5314 | FullDETBufferingTimeY[k] = LinesInDETYRoundedDownToSwath[k] * (HTotal[k] / PixelClock[k]) / VRatio[k]; |
5315 | if (BytePerPixelDETC[k] > 0) { |
5316 | LinesInDETC = mode_lib->vba.DETBufferSizeC[k] / BytePerPixelDETC[k] / SwathWidthC[k]; |
5317 | LinesInDETCRoundedDownToSwath = dml_floor(LinesInDETC, SwathHeightC[k]); |
5318 | FullDETBufferingTimeC = LinesInDETCRoundedDownToSwath * (HTotal[k] / PixelClock[k]) / VRatioChroma[k]; |
5319 | } else { |
5320 | LinesInDETC = 0; |
5321 | FullDETBufferingTimeC = 999999; |
5322 | } |
5323 | |
5324 | ActiveDRAMClockChangeLatencyMarginY = EffectiveLBLatencyHidingY + FullDETBufferingTimeY[k] - *UrgentWatermark - (HTotal[k] / PixelClock[k]) * (DSTXAfterScaler[k] / HTotal[k] + DSTYAfterScaler[k]) - *DRAMClockChangeWatermark; |
5325 | |
5326 | if (NumberOfActivePlanes > 1) { |
5327 | ActiveDRAMClockChangeLatencyMarginY = ActiveDRAMClockChangeLatencyMarginY - (1 - 1.0 / NumberOfActivePlanes) * SwathHeightY[k] * HTotal[k] / PixelClock[k] / VRatio[k]; |
5328 | } |
5329 | |
5330 | if (BytePerPixelDETC[k] > 0) { |
5331 | ActiveDRAMClockChangeLatencyMarginC = EffectiveLBLatencyHidingC + FullDETBufferingTimeC - *UrgentWatermark - (HTotal[k] / PixelClock[k]) * (DSTXAfterScaler[k] / HTotal[k] + DSTYAfterScaler[k]) - *DRAMClockChangeWatermark; |
5332 | |
5333 | if (NumberOfActivePlanes > 1) { |
5334 | ActiveDRAMClockChangeLatencyMarginC = ActiveDRAMClockChangeLatencyMarginC - (1 - 1.0 / NumberOfActivePlanes) * SwathHeightC[k] * HTotal[k] / PixelClock[k] / VRatioChroma[k]; |
5335 | } |
5336 | mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k] = dml_min(ActiveDRAMClockChangeLatencyMarginY, ActiveDRAMClockChangeLatencyMarginC); |
5337 | } else { |
5338 | mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k] = ActiveDRAMClockChangeLatencyMarginY; |
5339 | } |
5340 | |
5341 | if (WritebackEnable[k] == true1) { |
5342 | |
5343 | WritebackDRAMClockChangeLatencyHiding = WritebackInterfaceBufferSize * 1024 / (WritebackDestinationWidth[k] * WritebackDestinationHeight[k] / (WritebackSourceHeight[k] * HTotal[k] / PixelClock[k]) * 4); |
5344 | if (WritebackPixelFormat[k] == dm_444_64) { |
5345 | WritebackDRAMClockChangeLatencyHiding = WritebackDRAMClockChangeLatencyHiding / 2; |
5346 | } |
5347 | if (mode_lib->vba.WritebackConfiguration == dm_whole_buffer_for_single_stream_interleave) { |
5348 | WritebackDRAMClockChangeLatencyHiding = WritebackDRAMClockChangeLatencyHiding * 2; |
5349 | } |
5350 | WritebackDRAMClockChangeLatencyMargin = WritebackDRAMClockChangeLatencyHiding - mode_lib->vba.WritebackDRAMClockChangeWatermark; |
5351 | mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k] = dml_min(mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k], WritebackDRAMClockChangeLatencyMargin); |
5352 | } |
5353 | } |
5354 | |
5355 | mode_lib->vba.MinActiveDRAMClockChangeMargin = 999999; |
5356 | PlaneWithMinActiveDRAMClockChangeMargin = 0; |
5357 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
5358 | if (mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k] < mode_lib->vba.MinActiveDRAMClockChangeMargin) { |
5359 | mode_lib->vba.MinActiveDRAMClockChangeMargin = mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k]; |
5360 | if (BlendingAndTiming[k] == k) { |
5361 | PlaneWithMinActiveDRAMClockChangeMargin = k; |
5362 | } else { |
5363 | for (j = 0; j < NumberOfActivePlanes; ++j) { |
5364 | if (BlendingAndTiming[k] == j) { |
5365 | PlaneWithMinActiveDRAMClockChangeMargin = j; |
5366 | } |
5367 | } |
5368 | } |
5369 | } |
5370 | } |
5371 | |
5372 | *MinActiveDRAMClockChangeLatencySupported = mode_lib->vba.MinActiveDRAMClockChangeMargin + DRAMClockChangeLatency; |
5373 | |
5374 | SecondMinActiveDRAMClockChangeMarginOneDisplayInVBLank = 999999; |
5375 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
5376 | if (!((k == PlaneWithMinActiveDRAMClockChangeMargin) && (BlendingAndTiming[k] == k)) && !(BlendingAndTiming[k] == PlaneWithMinActiveDRAMClockChangeMargin) && mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k] < SecondMinActiveDRAMClockChangeMarginOneDisplayInVBLank) { |
5377 | SecondMinActiveDRAMClockChangeMarginOneDisplayInVBLank = mode_lib->vba.ActiveDRAMClockChangeLatencyMargin[k]; |
5378 | } |
5379 | } |
5380 | |
5381 | mode_lib->vba.TotalNumberOfActiveOTG = 0; |
5382 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
5383 | if (BlendingAndTiming[k] == k) { |
5384 | mode_lib->vba.TotalNumberOfActiveOTG = mode_lib->vba.TotalNumberOfActiveOTG + 1; |
5385 | } |
5386 | } |
5387 | |
5388 | if (mode_lib->vba.MinActiveDRAMClockChangeMargin > 0) { |
5389 | *DRAMClockChangeSupport = dm_dram_clock_change_vactive; |
5390 | } else if (((mode_lib->vba.SynchronizedVBlank == true1 || mode_lib->vba.TotalNumberOfActiveOTG == 1 || SecondMinActiveDRAMClockChangeMarginOneDisplayInVBLank > 0) && PrefetchMode == 0)) { |
5391 | *DRAMClockChangeSupport = dm_dram_clock_change_vblank; |
5392 | } else { |
5393 | *DRAMClockChangeSupport = dm_dram_clock_change_unsupported; |
5394 | } |
5395 | |
5396 | FullDETBufferingTimeYStutterCriticalPlane = FullDETBufferingTimeY[0]; |
5397 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
5398 | if (FullDETBufferingTimeY[k] <= FullDETBufferingTimeYStutterCriticalPlane) { |
5399 | FullDETBufferingTimeYStutterCriticalPlane = FullDETBufferingTimeY[k]; |
5400 | TimeToFinishSwathTransferStutterCriticalPlane = (SwathHeightY[k] - (LinesInDETY[k] - LinesInDETYRoundedDownToSwath[k])) * (HTotal[k] / PixelClock[k]) / VRatio[k]; |
5401 | } |
5402 | } |
5403 | |
5404 | *StutterExitWatermark = SRExitTime + ExtraLatency + 10 / DCFCLKDeepSleep; |
5405 | *StutterEnterPlusExitWatermark = dml_max(SREnterPlusExitTime + ExtraLatency + 10 / DCFCLKDeepSleep, TimeToFinishSwathTransferStutterCriticalPlane); |
5406 | |
5407 | } |
5408 | |
5409 | static void CalculateDCFCLKDeepSleep( |
5410 | struct display_mode_lib *mode_lib, |
5411 | unsigned int NumberOfActivePlanes, |
5412 | int BytePerPixelY[], |
5413 | int BytePerPixelC[], |
5414 | double VRatio[], |
5415 | double VRatioChroma[], |
5416 | double SwathWidthY[], |
5417 | double SwathWidthC[], |
5418 | unsigned int DPPPerPlane[], |
5419 | double HRatio[], |
5420 | double HRatioChroma[], |
5421 | double PixelClock[], |
5422 | double PSCL_THROUGHPUT[], |
5423 | double PSCL_THROUGHPUT_CHROMA[], |
5424 | double DPPCLK[], |
5425 | double ReadBandwidthLuma[], |
5426 | double ReadBandwidthChroma[], |
5427 | int ReturnBusWidth, |
5428 | double *DCFCLKDeepSleep) |
5429 | { |
5430 | double DisplayPipeLineDeliveryTimeLuma = 0; |
5431 | double DisplayPipeLineDeliveryTimeChroma = 0; |
5432 | unsigned int k; |
5433 | double ReadBandwidth = 0.0; |
5434 | |
5435 | //double DCFCLKDeepSleepPerPlane[DC__NUM_DPP__MAX]; |
5436 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
5437 | |
5438 | if (VRatio[k] <= 1) { |
5439 | DisplayPipeLineDeliveryTimeLuma = SwathWidthY[k] * DPPPerPlane[k] / HRatio[k] / PixelClock[k]; |
5440 | } else { |
5441 | DisplayPipeLineDeliveryTimeLuma = SwathWidthY[k] / PSCL_THROUGHPUT[k] / DPPCLK[k]; |
5442 | } |
5443 | if (BytePerPixelC[k] == 0) { |
5444 | DisplayPipeLineDeliveryTimeChroma = 0; |
5445 | } else { |
5446 | if (VRatioChroma[k] <= 1) { |
5447 | DisplayPipeLineDeliveryTimeChroma = SwathWidthC[k] * DPPPerPlane[k] / HRatioChroma[k] / PixelClock[k]; |
5448 | } else { |
5449 | DisplayPipeLineDeliveryTimeChroma = SwathWidthC[k] / PSCL_THROUGHPUT_CHROMA[k] / DPPCLK[k]; |
5450 | } |
5451 | } |
5452 | |
5453 | if (BytePerPixelC[k] > 0) { |
5454 | mode_lib->vba.DCFCLKDeepSleepPerPlane[k] = dml_max(1.1 * SwathWidthY[k] * BytePerPixelY[k] / 32.0 / DisplayPipeLineDeliveryTimeLuma, 1.1 * SwathWidthC[k] * BytePerPixelC[k] / 32.0 / DisplayPipeLineDeliveryTimeChroma); |
5455 | } else { |
5456 | mode_lib->vba.DCFCLKDeepSleepPerPlane[k] = 1.1 * SwathWidthY[k] * BytePerPixelY[k] / 64.0 / DisplayPipeLineDeliveryTimeLuma; |
5457 | } |
5458 | mode_lib->vba.DCFCLKDeepSleepPerPlane[k] = dml_max(mode_lib->vba.DCFCLKDeepSleepPerPlane[k], PixelClock[k] / 16); |
5459 | |
5460 | } |
5461 | |
5462 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
5463 | ReadBandwidth = ReadBandwidth + ReadBandwidthLuma[k] + ReadBandwidthChroma[k]; |
5464 | } |
5465 | |
5466 | *DCFCLKDeepSleep = dml_max(8.0, ReadBandwidth / ReturnBusWidth); |
5467 | |
5468 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
5469 | *DCFCLKDeepSleep = dml_max(*DCFCLKDeepSleep, mode_lib->vba.DCFCLKDeepSleepPerPlane[k]); |
5470 | } |
5471 | } |
5472 | |
5473 | static void CalculateUrgentBurstFactor( |
5474 | long swath_width_luma_ub, |
5475 | long swath_width_chroma_ub, |
5476 | unsigned int DETBufferSizeInKByte, |
5477 | unsigned int SwathHeightY, |
5478 | unsigned int SwathHeightC, |
5479 | double LineTime, |
5480 | double UrgentLatency, |
5481 | double CursorBufferSize, |
5482 | unsigned int CursorWidth, |
5483 | unsigned int CursorBPP, |
5484 | double VRatio, |
5485 | double VRatioC, |
5486 | double BytePerPixelInDETY, |
5487 | double BytePerPixelInDETC, |
5488 | double DETBufferSizeY, |
5489 | double DETBufferSizeC, |
5490 | double *UrgentBurstFactorCursor, |
5491 | double *UrgentBurstFactorLuma, |
5492 | double *UrgentBurstFactorChroma, |
5493 | bool_Bool *NotEnoughUrgentLatencyHiding) |
5494 | { |
5495 | double LinesInDETLuma = 0; |
5496 | double LinesInDETChroma = 0; |
5497 | unsigned int LinesInCursorBuffer = 0; |
5498 | double CursorBufferSizeInTime = 0; |
5499 | double DETBufferSizeInTimeLuma = 0; |
5500 | double DETBufferSizeInTimeChroma = 0; |
5501 | |
5502 | *NotEnoughUrgentLatencyHiding = 0; |
5503 | |
5504 | if (CursorWidth > 0) { |
5505 | LinesInCursorBuffer = 1 << (unsigned int) dml_floor(dml_log2(CursorBufferSize * 1024.0 / (CursorWidth * CursorBPP / 8.0)), 1.0); |
5506 | if (VRatio > 0) { |
5507 | CursorBufferSizeInTime = LinesInCursorBuffer * LineTime / VRatio; |
5508 | if (CursorBufferSizeInTime - UrgentLatency <= 0) { |
5509 | *NotEnoughUrgentLatencyHiding = 1; |
5510 | *UrgentBurstFactorCursor = 0; |
5511 | } else { |
5512 | *UrgentBurstFactorCursor = CursorBufferSizeInTime / (CursorBufferSizeInTime - UrgentLatency); |
5513 | } |
5514 | } else { |
5515 | *UrgentBurstFactorCursor = 1; |
5516 | } |
5517 | } |
5518 | |
5519 | LinesInDETLuma = DETBufferSizeY / BytePerPixelInDETY / swath_width_luma_ub; |
5520 | if (VRatio > 0) { |
5521 | DETBufferSizeInTimeLuma = dml_floor(LinesInDETLuma, SwathHeightY) * LineTime / VRatio; |
5522 | if (DETBufferSizeInTimeLuma - UrgentLatency <= 0) { |
5523 | *NotEnoughUrgentLatencyHiding = 1; |
5524 | *UrgentBurstFactorLuma = 0; |
5525 | } else { |
5526 | *UrgentBurstFactorLuma = DETBufferSizeInTimeLuma / (DETBufferSizeInTimeLuma - UrgentLatency); |
5527 | } |
5528 | } else { |
5529 | *UrgentBurstFactorLuma = 1; |
5530 | } |
5531 | |
5532 | if (BytePerPixelInDETC > 0) { |
5533 | LinesInDETChroma = DETBufferSizeC / BytePerPixelInDETC / swath_width_chroma_ub; |
5534 | if (VRatio > 0) { |
5535 | DETBufferSizeInTimeChroma = dml_floor(LinesInDETChroma, SwathHeightC) * LineTime / VRatio; |
5536 | if (DETBufferSizeInTimeChroma - UrgentLatency <= 0) { |
5537 | *NotEnoughUrgentLatencyHiding = 1; |
5538 | *UrgentBurstFactorChroma = 0; |
5539 | } else { |
5540 | *UrgentBurstFactorChroma = DETBufferSizeInTimeChroma / (DETBufferSizeInTimeChroma - UrgentLatency); |
5541 | } |
5542 | } else { |
5543 | *UrgentBurstFactorChroma = 1; |
5544 | } |
5545 | } |
5546 | } |
5547 | |
5548 | static void CalculatePixelDeliveryTimes( |
5549 | unsigned int NumberOfActivePlanes, |
5550 | double VRatio[], |
5551 | double VRatioChroma[], |
5552 | double VRatioPrefetchY[], |
5553 | double VRatioPrefetchC[], |
5554 | unsigned int swath_width_luma_ub[], |
5555 | unsigned int swath_width_chroma_ub[], |
5556 | unsigned int DPPPerPlane[], |
5557 | double HRatio[], |
5558 | double HRatioChroma[], |
5559 | double PixelClock[], |
5560 | double PSCL_THROUGHPUT[], |
5561 | double PSCL_THROUGHPUT_CHROMA[], |
5562 | double DPPCLK[], |
5563 | int BytePerPixelC[], |
5564 | enum scan_direction_class SourceScan[], |
5565 | unsigned int NumberOfCursors[], |
5566 | unsigned int CursorWidth[][2], |
5567 | unsigned int CursorBPP[][2], |
5568 | unsigned int BlockWidth256BytesY[], |
5569 | unsigned int BlockHeight256BytesY[], |
5570 | unsigned int BlockWidth256BytesC[], |
5571 | unsigned int BlockHeight256BytesC[], |
5572 | double DisplayPipeLineDeliveryTimeLuma[], |
5573 | double DisplayPipeLineDeliveryTimeChroma[], |
5574 | double DisplayPipeLineDeliveryTimeLumaPrefetch[], |
5575 | double DisplayPipeLineDeliveryTimeChromaPrefetch[], |
5576 | double DisplayPipeRequestDeliveryTimeLuma[], |
5577 | double DisplayPipeRequestDeliveryTimeChroma[], |
5578 | double DisplayPipeRequestDeliveryTimeLumaPrefetch[], |
5579 | double DisplayPipeRequestDeliveryTimeChromaPrefetch[], |
5580 | double CursorRequestDeliveryTime[], |
5581 | double CursorRequestDeliveryTimePrefetch[]) |
5582 | { |
5583 | double req_per_swath_ub = 0; |
5584 | unsigned int k; |
5585 | |
5586 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
5587 | if (VRatio[k] <= 1) { |
5588 | DisplayPipeLineDeliveryTimeLuma[k] = swath_width_luma_ub[k] * DPPPerPlane[k] / HRatio[k] / PixelClock[k]; |
5589 | } else { |
5590 | DisplayPipeLineDeliveryTimeLuma[k] = swath_width_luma_ub[k] / PSCL_THROUGHPUT[k] / DPPCLK[k]; |
5591 | } |
5592 | |
5593 | if (BytePerPixelC[k] == 0) { |
5594 | DisplayPipeLineDeliveryTimeChroma[k] = 0; |
5595 | } else { |
5596 | if (VRatioChroma[k] <= 1) { |
5597 | DisplayPipeLineDeliveryTimeChroma[k] = swath_width_chroma_ub[k] * DPPPerPlane[k] / HRatioChroma[k] / PixelClock[k]; |
5598 | } else { |
5599 | DisplayPipeLineDeliveryTimeChroma[k] = swath_width_chroma_ub[k] / PSCL_THROUGHPUT_CHROMA[k] / DPPCLK[k]; |
5600 | } |
5601 | } |
5602 | |
5603 | if (VRatioPrefetchY[k] <= 1) { |
5604 | DisplayPipeLineDeliveryTimeLumaPrefetch[k] = swath_width_luma_ub[k] * DPPPerPlane[k] / HRatio[k] / PixelClock[k]; |
5605 | } else { |
5606 | DisplayPipeLineDeliveryTimeLumaPrefetch[k] = swath_width_luma_ub[k] / PSCL_THROUGHPUT[k] / DPPCLK[k]; |
5607 | } |
5608 | |
5609 | if (BytePerPixelC[k] == 0) { |
5610 | DisplayPipeLineDeliveryTimeChromaPrefetch[k] = 0; |
5611 | } else { |
5612 | if (VRatioPrefetchC[k] <= 1) { |
5613 | DisplayPipeLineDeliveryTimeChromaPrefetch[k] = swath_width_chroma_ub[k] * DPPPerPlane[k] / HRatioChroma[k] / PixelClock[k]; |
5614 | } else { |
5615 | DisplayPipeLineDeliveryTimeChromaPrefetch[k] = swath_width_chroma_ub[k] / PSCL_THROUGHPUT_CHROMA[k] / DPPCLK[k]; |
5616 | } |
5617 | } |
5618 | } |
5619 | |
5620 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
5621 | if (SourceScan[k] != dm_vert) { |
5622 | req_per_swath_ub = swath_width_luma_ub[k] / BlockWidth256BytesY[k]; |
5623 | } else { |
5624 | req_per_swath_ub = swath_width_luma_ub[k] / BlockHeight256BytesY[k]; |
5625 | } |
5626 | DisplayPipeRequestDeliveryTimeLuma[k] = DisplayPipeLineDeliveryTimeLuma[k] / req_per_swath_ub; |
5627 | DisplayPipeRequestDeliveryTimeLumaPrefetch[k] = DisplayPipeLineDeliveryTimeLumaPrefetch[k] / req_per_swath_ub; |
5628 | if (BytePerPixelC[k] == 0) { |
5629 | DisplayPipeRequestDeliveryTimeChroma[k] = 0; |
5630 | DisplayPipeRequestDeliveryTimeChromaPrefetch[k] = 0; |
5631 | } else { |
5632 | if (SourceScan[k] != dm_vert) { |
5633 | req_per_swath_ub = swath_width_chroma_ub[k] / BlockWidth256BytesC[k]; |
5634 | } else { |
5635 | req_per_swath_ub = swath_width_chroma_ub[k] / BlockHeight256BytesC[k]; |
5636 | } |
5637 | DisplayPipeRequestDeliveryTimeChroma[k] = DisplayPipeLineDeliveryTimeChroma[k] / req_per_swath_ub; |
5638 | DisplayPipeRequestDeliveryTimeChromaPrefetch[k] = DisplayPipeLineDeliveryTimeChromaPrefetch[k] / req_per_swath_ub; |
5639 | } |
5640 | } |
5641 | |
5642 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
5643 | int cursor_req_per_width = 0; |
5644 | cursor_req_per_width = dml_ceil(CursorWidth[k][0] * CursorBPP[k][0] / 256 / 8, 1); |
5645 | if (NumberOfCursors[k] > 0) { |
5646 | if (VRatio[k] <= 1) { |
5647 | CursorRequestDeliveryTime[k] = CursorWidth[k][0] / HRatio[k] / PixelClock[k] / cursor_req_per_width; |
5648 | } else { |
5649 | CursorRequestDeliveryTime[k] = CursorWidth[k][0] / PSCL_THROUGHPUT[k] / DPPCLK[k] / cursor_req_per_width; |
5650 | } |
5651 | if (VRatioPrefetchY[k] <= 1) { |
5652 | CursorRequestDeliveryTimePrefetch[k] = CursorWidth[k][0] / HRatio[k] / PixelClock[k] / cursor_req_per_width; |
5653 | } else { |
5654 | CursorRequestDeliveryTimePrefetch[k] = CursorWidth[k][0] / PSCL_THROUGHPUT[k] / DPPCLK[k] / cursor_req_per_width; |
5655 | } |
5656 | } else { |
5657 | CursorRequestDeliveryTime[k] = 0; |
5658 | CursorRequestDeliveryTimePrefetch[k] = 0; |
5659 | } |
5660 | } |
5661 | } |
5662 | |
5663 | static void CalculateMetaAndPTETimes( |
5664 | int NumberOfActivePlanes, |
5665 | bool_Bool GPUVMEnable, |
5666 | int MetaChunkSize, |
5667 | int MinMetaChunkSizeBytes, |
5668 | int HTotal[], |
5669 | double VRatio[], |
5670 | double VRatioChroma[], |
5671 | double DestinationLinesToRequestRowInVBlank[], |
5672 | double DestinationLinesToRequestRowInImmediateFlip[], |
5673 | bool_Bool DCCEnable[], |
5674 | double PixelClock[], |
5675 | int BytePerPixelY[], |
5676 | int BytePerPixelC[], |
5677 | enum scan_direction_class SourceScan[], |
5678 | int dpte_row_height[], |
5679 | int dpte_row_height_chroma[], |
5680 | int meta_row_width[], |
5681 | int meta_row_width_chroma[], |
5682 | int meta_row_height[], |
5683 | int meta_row_height_chroma[], |
5684 | int meta_req_width[], |
5685 | int meta_req_width_chroma[], |
5686 | int meta_req_height[], |
5687 | int meta_req_height_chroma[], |
5688 | int dpte_group_bytes[], |
5689 | int PTERequestSizeY[], |
5690 | int PTERequestSizeC[], |
5691 | int PixelPTEReqWidthY[], |
5692 | int PixelPTEReqHeightY[], |
5693 | int PixelPTEReqWidthC[], |
5694 | int PixelPTEReqHeightC[], |
5695 | int dpte_row_width_luma_ub[], |
5696 | int dpte_row_width_chroma_ub[], |
5697 | double DST_Y_PER_PTE_ROW_NOM_L[], |
5698 | double DST_Y_PER_PTE_ROW_NOM_C[], |
5699 | double DST_Y_PER_META_ROW_NOM_L[], |
5700 | double DST_Y_PER_META_ROW_NOM_C[], |
5701 | double TimePerMetaChunkNominal[], |
5702 | double TimePerChromaMetaChunkNominal[], |
5703 | double TimePerMetaChunkVBlank[], |
5704 | double TimePerChromaMetaChunkVBlank[], |
5705 | double TimePerMetaChunkFlip[], |
5706 | double TimePerChromaMetaChunkFlip[], |
5707 | double time_per_pte_group_nom_luma[], |
5708 | double time_per_pte_group_vblank_luma[], |
5709 | double time_per_pte_group_flip_luma[], |
5710 | double time_per_pte_group_nom_chroma[], |
5711 | double time_per_pte_group_vblank_chroma[], |
5712 | double time_per_pte_group_flip_chroma[]) |
5713 | { |
5714 | unsigned int meta_chunk_width = 0; |
5715 | unsigned int min_meta_chunk_width = 0; |
5716 | unsigned int meta_chunk_per_row_int = 0; |
5717 | unsigned int meta_row_remainder = 0; |
5718 | unsigned int meta_chunk_threshold = 0; |
5719 | unsigned int meta_chunks_per_row_ub = 0; |
5720 | unsigned int meta_chunk_width_chroma = 0; |
5721 | unsigned int min_meta_chunk_width_chroma = 0; |
5722 | unsigned int meta_chunk_per_row_int_chroma = 0; |
5723 | unsigned int meta_row_remainder_chroma = 0; |
5724 | unsigned int meta_chunk_threshold_chroma = 0; |
5725 | unsigned int meta_chunks_per_row_ub_chroma = 0; |
5726 | unsigned int dpte_group_width_luma = 0; |
5727 | unsigned int dpte_groups_per_row_luma_ub = 0; |
5728 | unsigned int dpte_group_width_chroma = 0; |
5729 | unsigned int dpte_groups_per_row_chroma_ub = 0; |
5730 | unsigned int k; |
5731 | |
5732 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
5733 | DST_Y_PER_PTE_ROW_NOM_L[k] = dpte_row_height[k] / VRatio[k]; |
5734 | if (BytePerPixelC[k] == 0) { |
5735 | DST_Y_PER_PTE_ROW_NOM_C[k] = 0; |
5736 | } else { |
5737 | DST_Y_PER_PTE_ROW_NOM_C[k] = dpte_row_height_chroma[k] / VRatioChroma[k]; |
5738 | } |
5739 | DST_Y_PER_META_ROW_NOM_L[k] = meta_row_height[k] / VRatio[k]; |
5740 | if (BytePerPixelC[k] == 0) { |
5741 | DST_Y_PER_META_ROW_NOM_C[k] = 0; |
5742 | } else { |
5743 | DST_Y_PER_META_ROW_NOM_C[k] = meta_row_height_chroma[k] / VRatioChroma[k]; |
5744 | } |
5745 | } |
5746 | |
5747 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
5748 | if (DCCEnable[k] == true1) { |
5749 | meta_chunk_width = MetaChunkSize * 1024 * 256 / BytePerPixelY[k] / meta_row_height[k]; |
5750 | min_meta_chunk_width = MinMetaChunkSizeBytes * 256 / BytePerPixelY[k] / meta_row_height[k]; |
5751 | meta_chunk_per_row_int = meta_row_width[k] / meta_chunk_width; |
5752 | meta_row_remainder = meta_row_width[k] % meta_chunk_width; |
5753 | if (SourceScan[k] != dm_vert) { |
5754 | meta_chunk_threshold = 2 * min_meta_chunk_width - meta_req_width[k]; |
5755 | } else { |
5756 | meta_chunk_threshold = 2 * min_meta_chunk_width - meta_req_height[k]; |
5757 | } |
5758 | if (meta_row_remainder <= meta_chunk_threshold) { |
5759 | meta_chunks_per_row_ub = meta_chunk_per_row_int + 1; |
5760 | } else { |
5761 | meta_chunks_per_row_ub = meta_chunk_per_row_int + 2; |
5762 | } |
5763 | TimePerMetaChunkNominal[k] = meta_row_height[k] / VRatio[k] * HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub; |
5764 | TimePerMetaChunkVBlank[k] = DestinationLinesToRequestRowInVBlank[k] * HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub; |
5765 | TimePerMetaChunkFlip[k] = DestinationLinesToRequestRowInImmediateFlip[k] * HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub; |
5766 | if (BytePerPixelC[k] == 0) { |
5767 | TimePerChromaMetaChunkNominal[k] = 0; |
5768 | TimePerChromaMetaChunkVBlank[k] = 0; |
5769 | TimePerChromaMetaChunkFlip[k] = 0; |
5770 | } else { |
5771 | meta_chunk_width_chroma = MetaChunkSize * 1024 * 256 / BytePerPixelC[k] / meta_row_height_chroma[k]; |
5772 | min_meta_chunk_width_chroma = MinMetaChunkSizeBytes * 256 / BytePerPixelC[k] / meta_row_height_chroma[k]; |
5773 | meta_chunk_per_row_int_chroma = (double) meta_row_width_chroma[k] / meta_chunk_width_chroma; |
5774 | meta_row_remainder_chroma = meta_row_width_chroma[k] % meta_chunk_width_chroma; |
5775 | if (SourceScan[k] != dm_vert) { |
5776 | meta_chunk_threshold_chroma = 2 * min_meta_chunk_width_chroma - meta_req_width_chroma[k]; |
5777 | } else { |
5778 | meta_chunk_threshold_chroma = 2 * min_meta_chunk_width_chroma - meta_req_height_chroma[k]; |
5779 | } |
5780 | if (meta_row_remainder_chroma <= meta_chunk_threshold_chroma) { |
5781 | meta_chunks_per_row_ub_chroma = meta_chunk_per_row_int_chroma + 1; |
5782 | } else { |
5783 | meta_chunks_per_row_ub_chroma = meta_chunk_per_row_int_chroma + 2; |
5784 | } |
5785 | TimePerChromaMetaChunkNominal[k] = meta_row_height_chroma[k] / VRatioChroma[k] * HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub_chroma; |
5786 | TimePerChromaMetaChunkVBlank[k] = DestinationLinesToRequestRowInVBlank[k] * HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub_chroma; |
5787 | TimePerChromaMetaChunkFlip[k] = DestinationLinesToRequestRowInImmediateFlip[k] * HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub_chroma; |
5788 | } |
5789 | } else { |
5790 | TimePerMetaChunkNominal[k] = 0; |
5791 | TimePerMetaChunkVBlank[k] = 0; |
5792 | TimePerMetaChunkFlip[k] = 0; |
5793 | TimePerChromaMetaChunkNominal[k] = 0; |
5794 | TimePerChromaMetaChunkVBlank[k] = 0; |
5795 | TimePerChromaMetaChunkFlip[k] = 0; |
5796 | } |
5797 | } |
5798 | |
5799 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
5800 | if (GPUVMEnable == true1) { |
5801 | if (SourceScan[k] != dm_vert) { |
5802 | dpte_group_width_luma = dpte_group_bytes[k] / PTERequestSizeY[k] * PixelPTEReqWidthY[k]; |
5803 | } else { |
5804 | dpte_group_width_luma = dpte_group_bytes[k] / PTERequestSizeY[k] * PixelPTEReqHeightY[k]; |
5805 | } |
5806 | dpte_groups_per_row_luma_ub = dml_ceil(1.0 * dpte_row_width_luma_ub[k] / dpte_group_width_luma, 1); |
5807 | 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; |
5808 | time_per_pte_group_vblank_luma[k] = DestinationLinesToRequestRowInVBlank[k] * HTotal[k] / PixelClock[k] / dpte_groups_per_row_luma_ub; |
5809 | time_per_pte_group_flip_luma[k] = DestinationLinesToRequestRowInImmediateFlip[k] * HTotal[k] / PixelClock[k] / dpte_groups_per_row_luma_ub; |
5810 | if (BytePerPixelC[k] == 0) { |
5811 | time_per_pte_group_nom_chroma[k] = 0; |
5812 | time_per_pte_group_vblank_chroma[k] = 0; |
5813 | time_per_pte_group_flip_chroma[k] = 0; |
5814 | } else { |
5815 | if (SourceScan[k] != dm_vert) { |
5816 | dpte_group_width_chroma = dpte_group_bytes[k] / PTERequestSizeC[k] * PixelPTEReqWidthC[k]; |
5817 | } else { |
5818 | dpte_group_width_chroma = dpte_group_bytes[k] / PTERequestSizeC[k] * PixelPTEReqHeightC[k]; |
5819 | } |
5820 | dpte_groups_per_row_chroma_ub = dml_ceil(1.0 * dpte_row_width_chroma_ub[k] / dpte_group_width_chroma, 1); |
5821 | 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; |
5822 | time_per_pte_group_vblank_chroma[k] = DestinationLinesToRequestRowInVBlank[k] * HTotal[k] / PixelClock[k] / dpte_groups_per_row_chroma_ub; |
5823 | time_per_pte_group_flip_chroma[k] = DestinationLinesToRequestRowInImmediateFlip[k] * HTotal[k] / PixelClock[k] / dpte_groups_per_row_chroma_ub; |
5824 | } |
5825 | } else { |
5826 | time_per_pte_group_nom_luma[k] = 0; |
5827 | time_per_pte_group_vblank_luma[k] = 0; |
5828 | time_per_pte_group_flip_luma[k] = 0; |
5829 | time_per_pte_group_nom_chroma[k] = 0; |
5830 | time_per_pte_group_vblank_chroma[k] = 0; |
5831 | time_per_pte_group_flip_chroma[k] = 0; |
5832 | } |
5833 | } |
5834 | } |
5835 | |
5836 | static void CalculateVMGroupAndRequestTimes( |
5837 | unsigned int NumberOfActivePlanes, |
5838 | bool_Bool GPUVMEnable, |
5839 | unsigned int GPUVMMaxPageTableLevels, |
5840 | unsigned int HTotal[], |
5841 | int BytePerPixelC[], |
5842 | double DestinationLinesToRequestVMInVBlank[], |
5843 | double DestinationLinesToRequestVMInImmediateFlip[], |
5844 | bool_Bool DCCEnable[], |
5845 | double PixelClock[], |
5846 | int dpte_row_width_luma_ub[], |
5847 | int dpte_row_width_chroma_ub[], |
5848 | int vm_group_bytes[], |
5849 | unsigned int dpde0_bytes_per_frame_ub_l[], |
5850 | unsigned int dpde0_bytes_per_frame_ub_c[], |
5851 | int meta_pte_bytes_per_frame_ub_l[], |
5852 | int meta_pte_bytes_per_frame_ub_c[], |
5853 | double TimePerVMGroupVBlank[], |
5854 | double TimePerVMGroupFlip[], |
5855 | double TimePerVMRequestVBlank[], |
5856 | double TimePerVMRequestFlip[]) |
5857 | { |
5858 | int num_group_per_lower_vm_stage = 0; |
5859 | int num_req_per_lower_vm_stage = 0; |
5860 | unsigned int k; |
5861 | |
5862 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
5863 | if (GPUVMEnable == true1 && (DCCEnable[k] == true1 || GPUVMMaxPageTableLevels > 1)) { |
5864 | if (DCCEnable[k] == false0) { |
5865 | if (BytePerPixelC[k] > 0) { |
5866 | num_group_per_lower_vm_stage = dml_ceil((double) (dpde0_bytes_per_frame_ub_l[k]) |
5867 | / (double) (vm_group_bytes[k]), 1) + dml_ceil((double) (dpde0_bytes_per_frame_ub_c[k]) |
5868 | / (double) (vm_group_bytes[k]), 1); |
5869 | } else { |
5870 | num_group_per_lower_vm_stage = dml_ceil((double) (dpde0_bytes_per_frame_ub_l[k]) |
5871 | / (double) (vm_group_bytes[k]), 1); |
5872 | } |
5873 | } else { |
5874 | if (GPUVMMaxPageTableLevels == 1) { |
5875 | if (BytePerPixelC[k] > 0) { |
5876 | num_group_per_lower_vm_stage = dml_ceil((double) (meta_pte_bytes_per_frame_ub_l[k]) |
5877 | / (double) (vm_group_bytes[k]), 1) + dml_ceil((double) (meta_pte_bytes_per_frame_ub_c[k]) |
5878 | / (double) (vm_group_bytes[k]), 1); |
5879 | } else { |
5880 | num_group_per_lower_vm_stage = dml_ceil((double) (meta_pte_bytes_per_frame_ub_l[k]) |
5881 | / (double) (vm_group_bytes[k]), 1); |
5882 | } |
5883 | } else { |
5884 | if (BytePerPixelC[k] > 0) { |
5885 | num_group_per_lower_vm_stage = 2 + dml_ceil((double) (dpde0_bytes_per_frame_ub_l[k]) / (double) (vm_group_bytes[k]), 1) |
5886 | + dml_ceil((double) (dpde0_bytes_per_frame_ub_c[k]) / (double) (vm_group_bytes[k]), 1) |
5887 | + dml_ceil((double) (meta_pte_bytes_per_frame_ub_l[k]) / (double) (vm_group_bytes[k]), 1) |
5888 | + dml_ceil((double) (meta_pte_bytes_per_frame_ub_c[k]) / (double) (vm_group_bytes[k]), 1); |
5889 | } else { |
5890 | num_group_per_lower_vm_stage = 1 + dml_ceil((double) (dpde0_bytes_per_frame_ub_l[k]) / (double) (vm_group_bytes[k]), 1) |
5891 | + dml_ceil((double) (meta_pte_bytes_per_frame_ub_l[k]) / (double) (vm_group_bytes[k]), 1); |
5892 | } |
5893 | } |
5894 | } |
5895 | |
5896 | if (DCCEnable[k] == false0) { |
5897 | if (BytePerPixelC[k] > 0) { |
5898 | num_req_per_lower_vm_stage = dpde0_bytes_per_frame_ub_l[k] / 64 + dpde0_bytes_per_frame_ub_c[k] / 64; |
5899 | } else { |
5900 | num_req_per_lower_vm_stage = dpde0_bytes_per_frame_ub_l[k] / 64; |
5901 | } |
5902 | } else { |
5903 | if (GPUVMMaxPageTableLevels == 1) { |
5904 | if (BytePerPixelC[k] > 0) { |
5905 | num_req_per_lower_vm_stage = meta_pte_bytes_per_frame_ub_l[k] / 64 |
5906 | + meta_pte_bytes_per_frame_ub_c[k] / 64; |
5907 | } else { |
5908 | num_req_per_lower_vm_stage = meta_pte_bytes_per_frame_ub_l[k] / 64; |
5909 | } |
5910 | } else { |
5911 | if (BytePerPixelC[k] > 0) { |
5912 | num_req_per_lower_vm_stage = dpde0_bytes_per_frame_ub_l[k] / 64 |
5913 | + dpde0_bytes_per_frame_ub_c[k] / 64 + meta_pte_bytes_per_frame_ub_l[k] |
5914 | / 64 + meta_pte_bytes_per_frame_ub_c[k] / 64; |
5915 | } else { |
5916 | num_req_per_lower_vm_stage = dpde0_bytes_per_frame_ub_l[k] / 64 |
5917 | + meta_pte_bytes_per_frame_ub_l[k] / 64; |
5918 | } |
5919 | } |
5920 | } |
5921 | |
5922 | TimePerVMGroupVBlank[k] = DestinationLinesToRequestVMInVBlank[k] * HTotal[k] / PixelClock[k] |
5923 | / num_group_per_lower_vm_stage; |
5924 | TimePerVMGroupFlip[k] = DestinationLinesToRequestVMInImmediateFlip[k] * HTotal[k] / PixelClock[k] |
5925 | / num_group_per_lower_vm_stage; |
5926 | TimePerVMRequestVBlank[k] = DestinationLinesToRequestVMInVBlank[k] * HTotal[k] / PixelClock[k] |
5927 | / num_req_per_lower_vm_stage; |
5928 | TimePerVMRequestFlip[k] = DestinationLinesToRequestVMInImmediateFlip[k] * HTotal[k] / PixelClock[k] |
5929 | / num_req_per_lower_vm_stage; |
5930 | |
5931 | if (GPUVMMaxPageTableLevels > 2) { |
5932 | TimePerVMGroupVBlank[k] = TimePerVMGroupVBlank[k] / 2; |
5933 | TimePerVMGroupFlip[k] = TimePerVMGroupFlip[k] / 2; |
5934 | TimePerVMRequestVBlank[k] = TimePerVMRequestVBlank[k] / 2; |
5935 | TimePerVMRequestFlip[k] = TimePerVMRequestFlip[k] / 2; |
5936 | } |
5937 | |
5938 | } else { |
5939 | TimePerVMGroupVBlank[k] = 0; |
5940 | TimePerVMGroupFlip[k] = 0; |
5941 | TimePerVMRequestVBlank[k] = 0; |
5942 | TimePerVMRequestFlip[k] = 0; |
5943 | } |
5944 | } |
5945 | } |
5946 | |
5947 | static void CalculateStutterEfficiency( |
5948 | int NumberOfActivePlanes, |
5949 | long ROBBufferSizeInKByte, |
5950 | double TotalDataReadBandwidth, |
5951 | double DCFCLK, |
5952 | double ReturnBW, |
5953 | double SRExitTime, |
5954 | bool_Bool SynchronizedVBlank, |
5955 | int DPPPerPlane[], |
5956 | unsigned int DETBufferSizeY[], |
5957 | int BytePerPixelY[], |
5958 | double BytePerPixelDETY[], |
5959 | double SwathWidthY[], |
5960 | int SwathHeightY[], |
5961 | int SwathHeightC[], |
5962 | double DCCRateLuma[], |
5963 | double DCCRateChroma[], |
5964 | int HTotal[], |
5965 | int VTotal[], |
5966 | double PixelClock[], |
5967 | double VRatio[], |
5968 | enum scan_direction_class SourceScan[], |
5969 | int BlockHeight256BytesY[], |
5970 | int BlockWidth256BytesY[], |
5971 | int BlockHeight256BytesC[], |
5972 | int BlockWidth256BytesC[], |
5973 | int DCCYMaxUncompressedBlock[], |
5974 | int DCCCMaxUncompressedBlock[], |
5975 | int VActive[], |
5976 | bool_Bool DCCEnable[], |
5977 | bool_Bool WritebackEnable[], |
5978 | double ReadBandwidthPlaneLuma[], |
5979 | double ReadBandwidthPlaneChroma[], |
5980 | double meta_row_bw[], |
5981 | double dpte_row_bw[], |
5982 | double *StutterEfficiencyNotIncludingVBlank, |
5983 | double *StutterEfficiency, |
5984 | double *StutterPeriodOut) |
5985 | { |
5986 | double FullDETBufferingTimeY[DC__NUM_DPP__MAX8] = { 0 }; |
5987 | double FrameTimeForMinFullDETBufferingTime = 0; |
5988 | double StutterPeriod = 0; |
5989 | double AverageReadBandwidth = 0; |
5990 | double TotalRowReadBandwidth = 0; |
5991 | double AverageDCCCompressionRate = 0; |
5992 | double PartOfBurstThatFitsInROB = 0; |
5993 | double StutterBurstTime = 0; |
5994 | int TotalActiveWriteback = 0; |
5995 | double VBlankTime = 0; |
5996 | double SmallestVBlank = 0; |
5997 | int BytePerPixelYCriticalPlane = 0; |
5998 | double SwathWidthYCriticalPlane = 0; |
5999 | double LinesInDETY[DC__NUM_DPP__MAX8] = { 0 }; |
6000 | double LinesInDETYRoundedDownToSwath[DC__NUM_DPP__MAX8] = { 0 }; |
6001 | double LinesToFinishSwathTransferStutterCriticalPlane = 0; |
6002 | double MaximumEffectiveCompressionLuma = 0; |
6003 | double MaximumEffectiveCompressionChroma = 0; |
6004 | unsigned int k; |
6005 | |
6006 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
6007 | LinesInDETY[k] = DETBufferSizeY[k] / BytePerPixelDETY[k] / SwathWidthY[k]; |
6008 | LinesInDETYRoundedDownToSwath[k] = dml_floor(LinesInDETY[k], SwathHeightY[k]); |
6009 | FullDETBufferingTimeY[k] = LinesInDETYRoundedDownToSwath[k] * (HTotal[k] / PixelClock[k]) / VRatio[k]; |
6010 | } |
6011 | |
6012 | StutterPeriod = FullDETBufferingTimeY[0]; |
6013 | FrameTimeForMinFullDETBufferingTime = VTotal[0] * HTotal[0] / PixelClock[0]; |
6014 | BytePerPixelYCriticalPlane = BytePerPixelY[0]; |
6015 | SwathWidthYCriticalPlane = SwathWidthY[0]; |
6016 | LinesToFinishSwathTransferStutterCriticalPlane = SwathHeightY[0] |
6017 | - (LinesInDETY[0] - LinesInDETYRoundedDownToSwath[0]); |
6018 | |
6019 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
6020 | if (FullDETBufferingTimeY[k] < StutterPeriod) { |
6021 | StutterPeriod = FullDETBufferingTimeY[k]; |
6022 | FrameTimeForMinFullDETBufferingTime = VTotal[k] * HTotal[k] / PixelClock[k]; |
6023 | BytePerPixelYCriticalPlane = BytePerPixelY[k]; |
6024 | SwathWidthYCriticalPlane = SwathWidthY[k]; |
6025 | LinesToFinishSwathTransferStutterCriticalPlane = SwathHeightY[k] |
6026 | - (LinesInDETY[k] - LinesInDETYRoundedDownToSwath[k]); |
6027 | } |
6028 | } |
6029 | |
6030 | AverageReadBandwidth = 0; |
6031 | TotalRowReadBandwidth = 0; |
6032 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
6033 | if (DCCEnable[k] == true1) { |
6034 | if ((SourceScan[k] == dm_vert && BlockWidth256BytesY[k] > SwathHeightY[k]) |
6035 | || (SourceScan[k] != dm_vert |
6036 | && BlockHeight256BytesY[k] > SwathHeightY[k]) |
6037 | || DCCYMaxUncompressedBlock[k] < 256) { |
6038 | MaximumEffectiveCompressionLuma = 2; |
6039 | } else { |
6040 | MaximumEffectiveCompressionLuma = 4; |
6041 | } |
6042 | AverageReadBandwidth = AverageReadBandwidth + ReadBandwidthPlaneLuma[k] / dml_min(DCCRateLuma[k], MaximumEffectiveCompressionLuma); |
6043 | |
6044 | if (ReadBandwidthPlaneChroma[k] > 0) { |
6045 | if ((SourceScan[k] == dm_vert && BlockWidth256BytesC[k] > SwathHeightC[k]) |
6046 | || (SourceScan[k] != dm_vert && BlockHeight256BytesC[k] > SwathHeightC[k]) |
6047 | || DCCCMaxUncompressedBlock[k] < 256) { |
6048 | MaximumEffectiveCompressionChroma = 2; |
6049 | } else { |
6050 | MaximumEffectiveCompressionChroma = 4; |
6051 | } |
6052 | AverageReadBandwidth = AverageReadBandwidth + ReadBandwidthPlaneChroma[k] / dml_min(DCCRateChroma[k], MaximumEffectiveCompressionChroma); |
6053 | } |
6054 | } else { |
6055 | AverageReadBandwidth = AverageReadBandwidth + ReadBandwidthPlaneLuma[k] + ReadBandwidthPlaneChroma[k]; |
6056 | } |
6057 | TotalRowReadBandwidth = TotalRowReadBandwidth + DPPPerPlane[k] * (meta_row_bw[k] + dpte_row_bw[k]); |
6058 | } |
6059 | |
6060 | AverageDCCCompressionRate = TotalDataReadBandwidth / AverageReadBandwidth; |
6061 | PartOfBurstThatFitsInROB = dml_min(StutterPeriod * TotalDataReadBandwidth, ROBBufferSizeInKByte * 1024 * AverageDCCCompressionRate); |
6062 | StutterBurstTime = PartOfBurstThatFitsInROB / AverageDCCCompressionRate / ReturnBW + (StutterPeriod * TotalDataReadBandwidth |
6063 | - PartOfBurstThatFitsInROB) / (DCFCLK * 64) + StutterPeriod * TotalRowReadBandwidth / ReturnBW; |
6064 | StutterBurstTime = dml_max(StutterBurstTime, LinesToFinishSwathTransferStutterCriticalPlane * BytePerPixelYCriticalPlane * SwathWidthYCriticalPlane / ReturnBW); |
6065 | |
6066 | TotalActiveWriteback = 0; |
6067 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
6068 | if (WritebackEnable[k] == true1) { |
6069 | TotalActiveWriteback = TotalActiveWriteback + 1; |
6070 | } |
6071 | } |
6072 | |
6073 | if (TotalActiveWriteback == 0) { |
6074 | *StutterEfficiencyNotIncludingVBlank = (1 |
6075 | - (SRExitTime + StutterBurstTime) / StutterPeriod) * 100; |
6076 | } else { |
6077 | *StutterEfficiencyNotIncludingVBlank = 0; |
6078 | } |
6079 | |
6080 | if (SynchronizedVBlank == true1 || NumberOfActivePlanes == 1) { |
6081 | SmallestVBlank = (VTotal[0] - VActive[0]) * HTotal[0] / PixelClock[0]; |
6082 | } else { |
6083 | SmallestVBlank = 0; |
6084 | } |
6085 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
6086 | if (SynchronizedVBlank == true1 || NumberOfActivePlanes == 1) { |
6087 | VBlankTime = (VTotal[k] - VActive[k]) * HTotal[k] / PixelClock[k]; |
6088 | } else { |
6089 | VBlankTime = 0; |
6090 | } |
6091 | SmallestVBlank = dml_min(SmallestVBlank, VBlankTime); |
6092 | } |
6093 | |
6094 | *StutterEfficiency = (*StutterEfficiencyNotIncludingVBlank / 100.0 * (FrameTimeForMinFullDETBufferingTime - SmallestVBlank) + SmallestVBlank) / FrameTimeForMinFullDETBufferingTime * 100; |
6095 | |
6096 | if (StutterPeriodOut) |
6097 | *StutterPeriodOut = StutterPeriod; |
6098 | } |
6099 | |
6100 | static void CalculateSwathAndDETConfiguration( |
6101 | bool_Bool ForceSingleDPP, |
6102 | int NumberOfActivePlanes, |
6103 | unsigned int DETBufferSizeInKByte, |
6104 | double MaximumSwathWidthLuma[], |
6105 | double MaximumSwathWidthChroma[], |
6106 | enum scan_direction_class SourceScan[], |
6107 | enum source_format_class SourcePixelFormat[], |
6108 | enum dm_swizzle_mode SurfaceTiling[], |
6109 | int ViewportWidth[], |
6110 | int ViewportHeight[], |
6111 | int SurfaceWidthY[], |
6112 | int SurfaceWidthC[], |
6113 | int SurfaceHeightY[], |
6114 | int SurfaceHeightC[], |
6115 | int Read256BytesBlockHeightY[], |
6116 | int Read256BytesBlockHeightC[], |
6117 | int Read256BytesBlockWidthY[], |
6118 | int Read256BytesBlockWidthC[], |
6119 | enum odm_combine_mode ODMCombineEnabled[], |
6120 | int BlendingAndTiming[], |
6121 | int BytePerPixY[], |
6122 | int BytePerPixC[], |
6123 | double BytePerPixDETY[], |
6124 | double BytePerPixDETC[], |
6125 | int HActive[], |
6126 | double HRatio[], |
6127 | double HRatioChroma[], |
6128 | int DPPPerPlane[], |
6129 | int swath_width_luma_ub[], |
6130 | int swath_width_chroma_ub[], |
6131 | double SwathWidth[], |
6132 | double SwathWidthChroma[], |
6133 | int SwathHeightY[], |
6134 | int SwathHeightC[], |
6135 | unsigned int DETBufferSizeY[], |
6136 | unsigned int DETBufferSizeC[], |
6137 | bool_Bool ViewportSizeSupportPerPlane[], |
6138 | bool_Bool *ViewportSizeSupport) |
6139 | { |
6140 | int MaximumSwathHeightY[DC__NUM_DPP__MAX8] = { 0 }; |
6141 | int MaximumSwathHeightC[DC__NUM_DPP__MAX8] = { 0 }; |
6142 | int MinimumSwathHeightY = 0; |
6143 | int MinimumSwathHeightC = 0; |
6144 | long RoundedUpMaxSwathSizeBytesY = 0; |
6145 | long RoundedUpMaxSwathSizeBytesC = 0; |
6146 | long RoundedUpMinSwathSizeBytesY = 0; |
6147 | long RoundedUpMinSwathSizeBytesC = 0; |
6148 | long RoundedUpSwathSizeBytesY = 0; |
6149 | long RoundedUpSwathSizeBytesC = 0; |
6150 | double SwathWidthSingleDPP[DC__NUM_DPP__MAX8] = { 0 }; |
6151 | double SwathWidthSingleDPPChroma[DC__NUM_DPP__MAX8] = { 0 }; |
6152 | int k; |
6153 | |
6154 | CalculateSwathWidth( |
6155 | ForceSingleDPP, |
6156 | NumberOfActivePlanes, |
6157 | SourcePixelFormat, |
6158 | SourceScan, |
6159 | ViewportWidth, |
6160 | ViewportHeight, |
6161 | SurfaceWidthY, |
6162 | SurfaceWidthC, |
6163 | SurfaceHeightY, |
6164 | SurfaceHeightC, |
6165 | ODMCombineEnabled, |
6166 | BytePerPixY, |
6167 | BytePerPixC, |
6168 | Read256BytesBlockHeightY, |
6169 | Read256BytesBlockHeightC, |
6170 | Read256BytesBlockWidthY, |
6171 | Read256BytesBlockWidthC, |
6172 | BlendingAndTiming, |
6173 | HActive, |
6174 | HRatio, |
6175 | DPPPerPlane, |
6176 | SwathWidthSingleDPP, |
6177 | SwathWidthSingleDPPChroma, |
6178 | SwathWidth, |
6179 | SwathWidthChroma, |
6180 | MaximumSwathHeightY, |
6181 | MaximumSwathHeightC, |
6182 | swath_width_luma_ub, |
6183 | swath_width_chroma_ub); |
6184 | |
6185 | *ViewportSizeSupport = true1; |
6186 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
6187 | if ((SourcePixelFormat[k] == dm_444_64 || SourcePixelFormat[k] == dm_444_32 |
6188 | || SourcePixelFormat[k] == dm_444_16 |
6189 | || SourcePixelFormat[k] == dm_mono_16 |
6190 | || SourcePixelFormat[k] == dm_mono_8 |
6191 | || SourcePixelFormat[k] == dm_rgbe)) { |
6192 | if (SurfaceTiling[k] == dm_sw_linear |
6193 | || (SourcePixelFormat[k] == dm_444_64 |
6194 | && (SurfaceTiling[k] == dm_sw_64kb_s || SurfaceTiling[k] == dm_sw_64kb_s_t || SurfaceTiling[k] == dm_sw_64kb_s_x) |
6195 | && SourceScan[k] != dm_vert)) { |
6196 | MinimumSwathHeightY = MaximumSwathHeightY[k]; |
6197 | } else if (SourcePixelFormat[k] == dm_444_8 && SourceScan[k] == dm_vert) { |
6198 | MinimumSwathHeightY = MaximumSwathHeightY[k]; |
6199 | } else { |
6200 | MinimumSwathHeightY = MaximumSwathHeightY[k] / 2; |
6201 | } |
6202 | MinimumSwathHeightC = MaximumSwathHeightC[k]; |
6203 | } else { |
6204 | if (SurfaceTiling[k] == dm_sw_linear) { |
6205 | MinimumSwathHeightY = MaximumSwathHeightY[k]; |
6206 | MinimumSwathHeightC = MaximumSwathHeightC[k]; |
6207 | } else if (SourcePixelFormat[k] == dm_rgbe_alpha |
6208 | && SourceScan[k] == dm_vert) { |
6209 | MinimumSwathHeightY = MaximumSwathHeightY[k] / 2; |
6210 | MinimumSwathHeightC = MaximumSwathHeightC[k]; |
6211 | } else if (SourcePixelFormat[k] == dm_rgbe_alpha) { |
6212 | MinimumSwathHeightY = MaximumSwathHeightY[k] / 2; |
6213 | MinimumSwathHeightC = MaximumSwathHeightC[k] / 2; |
6214 | } else if (SourcePixelFormat[k] == dm_420_8 && SourceScan[k] == dm_vert) { |
6215 | MinimumSwathHeightY = MaximumSwathHeightY[k]; |
6216 | MinimumSwathHeightC = MaximumSwathHeightC[k] / 2; |
6217 | } else { |
6218 | MinimumSwathHeightC = MaximumSwathHeightC[k] / 2; |
6219 | MinimumSwathHeightY = MaximumSwathHeightY[k] / 2; |
6220 | } |
6221 | } |
6222 | |
6223 | RoundedUpMaxSwathSizeBytesY = swath_width_luma_ub[k] * BytePerPixDETY[k] |
6224 | * MaximumSwathHeightY[k]; |
6225 | RoundedUpMinSwathSizeBytesY = swath_width_luma_ub[k] * BytePerPixDETY[k] |
6226 | * MinimumSwathHeightY; |
6227 | if (SourcePixelFormat[k] == dm_420_10) { |
6228 | RoundedUpMaxSwathSizeBytesY = dml_ceil((double) RoundedUpMaxSwathSizeBytesY, 256); |
6229 | RoundedUpMinSwathSizeBytesY = dml_ceil((double) RoundedUpMinSwathSizeBytesY, 256); |
6230 | } |
6231 | RoundedUpMaxSwathSizeBytesC = swath_width_chroma_ub[k] * BytePerPixDETC[k] |
6232 | * MaximumSwathHeightC[k]; |
6233 | RoundedUpMinSwathSizeBytesC = swath_width_chroma_ub[k] * BytePerPixDETC[k] |
6234 | * MinimumSwathHeightC; |
6235 | if (SourcePixelFormat[k] == dm_420_10) { |
6236 | RoundedUpMaxSwathSizeBytesC = dml_ceil(RoundedUpMaxSwathSizeBytesC, 256); |
6237 | RoundedUpMinSwathSizeBytesC = dml_ceil(RoundedUpMinSwathSizeBytesC, 256); |
6238 | } |
6239 | |
6240 | if (RoundedUpMaxSwathSizeBytesY + RoundedUpMaxSwathSizeBytesC |
6241 | <= DETBufferSizeInKByte * 1024 / 2) { |
6242 | SwathHeightY[k] = MaximumSwathHeightY[k]; |
6243 | SwathHeightC[k] = MaximumSwathHeightC[k]; |
6244 | RoundedUpSwathSizeBytesY = RoundedUpMaxSwathSizeBytesY; |
6245 | RoundedUpSwathSizeBytesC = RoundedUpMaxSwathSizeBytesC; |
6246 | } else if (RoundedUpMaxSwathSizeBytesY >= 1.5 * RoundedUpMaxSwathSizeBytesC |
6247 | && RoundedUpMinSwathSizeBytesY + RoundedUpMaxSwathSizeBytesC |
6248 | <= DETBufferSizeInKByte * 1024 / 2) { |
6249 | SwathHeightY[k] = MinimumSwathHeightY; |
6250 | SwathHeightC[k] = MaximumSwathHeightC[k]; |
6251 | RoundedUpSwathSizeBytesY = RoundedUpMinSwathSizeBytesY; |
6252 | RoundedUpSwathSizeBytesC = RoundedUpMaxSwathSizeBytesC; |
6253 | } else if (RoundedUpMaxSwathSizeBytesY < 1.5 * RoundedUpMaxSwathSizeBytesC |
6254 | && RoundedUpMaxSwathSizeBytesY + RoundedUpMinSwathSizeBytesC |
6255 | <= DETBufferSizeInKByte * 1024 / 2) { |
6256 | SwathHeightY[k] = MaximumSwathHeightY[k]; |
6257 | SwathHeightC[k] = MinimumSwathHeightC; |
6258 | RoundedUpSwathSizeBytesY = RoundedUpMaxSwathSizeBytesY; |
6259 | RoundedUpSwathSizeBytesC = RoundedUpMinSwathSizeBytesC; |
6260 | } else { |
6261 | SwathHeightY[k] = MinimumSwathHeightY; |
6262 | SwathHeightC[k] = MinimumSwathHeightC; |
6263 | RoundedUpSwathSizeBytesY = RoundedUpMinSwathSizeBytesY; |
6264 | RoundedUpSwathSizeBytesC = RoundedUpMinSwathSizeBytesC; |
6265 | } |
6266 | |
6267 | if (SwathHeightC[k] == 0) { |
6268 | DETBufferSizeY[k] = DETBufferSizeInKByte * 1024; |
6269 | DETBufferSizeC[k] = 0; |
6270 | } else if (RoundedUpSwathSizeBytesY <= 1.5 * RoundedUpSwathSizeBytesC) { |
6271 | DETBufferSizeY[k] = DETBufferSizeInKByte * 1024 / 2; |
6272 | DETBufferSizeC[k] = DETBufferSizeInKByte * 1024 / 2; |
6273 | } else { |
6274 | DETBufferSizeY[k] = DETBufferSizeInKByte * 1024 * 2 / 3; |
6275 | DETBufferSizeC[k] = DETBufferSizeInKByte * 1024 / 3; |
6276 | } |
6277 | |
6278 | if (RoundedUpMinSwathSizeBytesY + RoundedUpMinSwathSizeBytesC |
6279 | > DETBufferSizeInKByte * 1024 / 2 |
6280 | || SwathWidth[k] > MaximumSwathWidthLuma[k] |
6281 | || (SwathHeightC[k] > 0 |
6282 | && SwathWidthChroma[k] > MaximumSwathWidthChroma[k])) { |
6283 | *ViewportSizeSupport = false0; |
6284 | ViewportSizeSupportPerPlane[k] = false0; |
6285 | } else { |
6286 | ViewportSizeSupportPerPlane[k] = true1; |
6287 | } |
6288 | } |
6289 | } |
6290 | |
6291 | static void CalculateSwathWidth( |
6292 | bool_Bool ForceSingleDPP, |
6293 | int NumberOfActivePlanes, |
6294 | enum source_format_class SourcePixelFormat[], |
6295 | enum scan_direction_class SourceScan[], |
6296 | unsigned int ViewportWidth[], |
6297 | unsigned int ViewportHeight[], |
6298 | unsigned int SurfaceWidthY[], |
6299 | unsigned int SurfaceWidthC[], |
6300 | unsigned int SurfaceHeightY[], |
6301 | unsigned int SurfaceHeightC[], |
6302 | enum odm_combine_mode ODMCombineEnabled[], |
6303 | int BytePerPixY[], |
6304 | int BytePerPixC[], |
6305 | int Read256BytesBlockHeightY[], |
6306 | int Read256BytesBlockHeightC[], |
6307 | int Read256BytesBlockWidthY[], |
6308 | int Read256BytesBlockWidthC[], |
6309 | int BlendingAndTiming[], |
6310 | unsigned int HActive[], |
6311 | double HRatio[], |
6312 | int DPPPerPlane[], |
6313 | double SwathWidthSingleDPPY[], |
6314 | double SwathWidthSingleDPPC[], |
6315 | double SwathWidthY[], |
6316 | double SwathWidthC[], |
6317 | int MaximumSwathHeightY[], |
6318 | int MaximumSwathHeightC[], |
6319 | unsigned int swath_width_luma_ub[], |
6320 | unsigned int swath_width_chroma_ub[]) |
6321 | { |
6322 | unsigned int k, j; |
6323 | long surface_width_ub_l; |
6324 | long surface_height_ub_l; |
6325 | long surface_width_ub_c; |
6326 | long surface_height_ub_c; |
6327 | |
6328 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
6329 | enum odm_combine_mode MainPlaneODMCombine = 0; |
6330 | |
6331 | if (SourceScan[k] != dm_vert) { |
6332 | SwathWidthSingleDPPY[k] = ViewportWidth[k]; |
6333 | } else { |
6334 | SwathWidthSingleDPPY[k] = ViewportHeight[k]; |
6335 | } |
6336 | |
6337 | MainPlaneODMCombine = ODMCombineEnabled[k]; |
6338 | for (j = 0; j < NumberOfActivePlanes; ++j) { |
6339 | if (BlendingAndTiming[k] == j) { |
6340 | MainPlaneODMCombine = ODMCombineEnabled[j]; |
6341 | } |
6342 | } |
6343 | |
6344 | if (MainPlaneODMCombine == dm_odm_combine_mode_4to1) { |
6345 | SwathWidthY[k] = dml_min(SwathWidthSingleDPPY[k], dml_round(HActive[k] / 4.0 * HRatio[k])); |
6346 | } else if (MainPlaneODMCombine == dm_odm_combine_mode_2to1) { |
6347 | SwathWidthY[k] = dml_min(SwathWidthSingleDPPY[k], dml_round(HActive[k] / 2.0 * HRatio[k])); |
6348 | } else if (DPPPerPlane[k] == 2) { |
6349 | SwathWidthY[k] = SwathWidthSingleDPPY[k] / 2; |
6350 | } else { |
6351 | SwathWidthY[k] = SwathWidthSingleDPPY[k]; |
6352 | } |
6353 | |
6354 | if (SourcePixelFormat[k] == dm_420_8 || SourcePixelFormat[k] == dm_420_10 || SourcePixelFormat[k] == dm_420_12) { |
6355 | SwathWidthC[k] = SwathWidthY[k] / 2; |
6356 | SwathWidthSingleDPPC[k] = SwathWidthSingleDPPY[k] / 2; |
6357 | } else { |
6358 | SwathWidthC[k] = SwathWidthY[k]; |
6359 | SwathWidthSingleDPPC[k] = SwathWidthSingleDPPY[k]; |
6360 | } |
6361 | |
6362 | if (ForceSingleDPP == true1) { |
6363 | SwathWidthY[k] = SwathWidthSingleDPPY[k]; |
6364 | SwathWidthC[k] = SwathWidthSingleDPPC[k]; |
6365 | } |
6366 | |
6367 | surface_width_ub_l = dml_ceil(SurfaceWidthY[k], Read256BytesBlockWidthY[k]); |
6368 | surface_height_ub_l = dml_ceil(SurfaceHeightY[k], Read256BytesBlockHeightY[k]); |
6369 | |
6370 | if (SourceScan[k] != dm_vert) { |
6371 | MaximumSwathHeightY[k] = Read256BytesBlockHeightY[k]; |
6372 | MaximumSwathHeightC[k] = Read256BytesBlockHeightC[k]; |
6373 | swath_width_luma_ub[k] = dml_min(surface_width_ub_l, (long) dml_ceil(SwathWidthY[k] - 1, |
6374 | Read256BytesBlockWidthY[k]) + Read256BytesBlockWidthY[k]); |
6375 | if (BytePerPixC[k] > 0) { |
6376 | surface_width_ub_c = dml_ceil(SurfaceWidthC[k], Read256BytesBlockWidthC[k]); |
6377 | swath_width_chroma_ub[k] = dml_min(surface_width_ub_c, (long) dml_ceil(SwathWidthC[k] - 1, |
6378 | Read256BytesBlockWidthC[k]) + Read256BytesBlockWidthC[k]); |
6379 | } else { |
6380 | swath_width_chroma_ub[k] = 0; |
6381 | } |
6382 | } else { |
6383 | MaximumSwathHeightY[k] = Read256BytesBlockWidthY[k]; |
6384 | MaximumSwathHeightC[k] = Read256BytesBlockWidthC[k]; |
6385 | swath_width_luma_ub[k] = dml_min(surface_height_ub_l, (long) dml_ceil(SwathWidthY[k] - 1, |
6386 | Read256BytesBlockHeightY[k]) + Read256BytesBlockHeightY[k]); |
6387 | if (BytePerPixC[k] > 0) { |
6388 | surface_height_ub_c = dml_ceil(SurfaceHeightC[k], Read256BytesBlockHeightC[k]); |
6389 | swath_width_chroma_ub[k] = dml_min(surface_height_ub_c, (long) dml_ceil(SwathWidthC[k] - 1, |
6390 | Read256BytesBlockHeightC[k]) + Read256BytesBlockHeightC[k]); |
6391 | } else { |
6392 | swath_width_chroma_ub[k] = 0; |
6393 | } |
6394 | } |
6395 | } |
6396 | } |
6397 | |
6398 | static double CalculateExtraLatency( |
6399 | long RoundTripPingLatencyCycles, |
6400 | long ReorderingBytes, |
6401 | double DCFCLK, |
6402 | int TotalNumberOfActiveDPP, |
6403 | int PixelChunkSizeInKByte, |
6404 | int TotalNumberOfDCCActiveDPP, |
6405 | int MetaChunkSize, |
6406 | double ReturnBW, |
6407 | bool_Bool GPUVMEnable, |
6408 | bool_Bool HostVMEnable, |
6409 | int NumberOfActivePlanes, |
6410 | int NumberOfDPP[], |
6411 | int dpte_group_bytes[], |
6412 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
6413 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
6414 | double HostVMMinPageSize, |
6415 | int HostVMMaxNonCachedPageTableLevels) |
6416 | { |
6417 | double ExtraLatencyBytes = 0; |
6418 | ExtraLatencyBytes = CalculateExtraLatencyBytes( |
6419 | ReorderingBytes, |
6420 | TotalNumberOfActiveDPP, |
6421 | PixelChunkSizeInKByte, |
6422 | TotalNumberOfDCCActiveDPP, |
6423 | MetaChunkSize, |
6424 | GPUVMEnable, |
6425 | HostVMEnable, |
6426 | NumberOfActivePlanes, |
6427 | NumberOfDPP, |
6428 | dpte_group_bytes, |
6429 | PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
6430 | PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
6431 | HostVMMinPageSize, |
6432 | HostVMMaxNonCachedPageTableLevels); |
6433 | |
6434 | return (RoundTripPingLatencyCycles + 32) / DCFCLK + ExtraLatencyBytes / ReturnBW; |
6435 | } |
6436 | |
6437 | static double CalculateExtraLatencyBytes( |
6438 | long ReorderingBytes, |
6439 | int TotalNumberOfActiveDPP, |
6440 | int PixelChunkSizeInKByte, |
6441 | int TotalNumberOfDCCActiveDPP, |
6442 | int MetaChunkSize, |
6443 | bool_Bool GPUVMEnable, |
6444 | bool_Bool HostVMEnable, |
6445 | int NumberOfActivePlanes, |
6446 | int NumberOfDPP[], |
6447 | int dpte_group_bytes[], |
6448 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, |
6449 | double PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
6450 | double HostVMMinPageSize, |
6451 | int HostVMMaxNonCachedPageTableLevels) |
6452 | { |
6453 | double ret = 0; |
6454 | double HostVMInefficiencyFactor = 0; |
6455 | int HostVMDynamicLevels = 0; |
6456 | unsigned int k; |
6457 | |
6458 | if (GPUVMEnable == true1 && HostVMEnable == true1) { |
6459 | HostVMInefficiencyFactor = PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData / PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly; |
6460 | if (HostVMMinPageSize < 2048) { |
6461 | HostVMDynamicLevels = HostVMMaxNonCachedPageTableLevels; |
6462 | } else if (HostVMMinPageSize >= 2048 && HostVMMinPageSize < 1048576) { |
6463 | HostVMDynamicLevels = dml_max(0, (int) HostVMMaxNonCachedPageTableLevels - 1); |
6464 | } else { |
6465 | HostVMDynamicLevels = dml_max(0, (int) HostVMMaxNonCachedPageTableLevels - 2); |
6466 | } |
6467 | } else { |
6468 | HostVMInefficiencyFactor = 1; |
6469 | HostVMDynamicLevels = 0; |
6470 | } |
6471 | |
6472 | ret = ReorderingBytes + (TotalNumberOfActiveDPP * PixelChunkSizeInKByte + TotalNumberOfDCCActiveDPP * MetaChunkSize) * 1024.0; |
6473 | |
6474 | if (GPUVMEnable == true1) { |
6475 | for (k = 0; k < NumberOfActivePlanes; ++k) { |
6476 | ret = ret + NumberOfDPP[k] * dpte_group_bytes[k] * (1 + 8 * HostVMDynamicLevels) * HostVMInefficiencyFactor; |
6477 | } |
6478 | } |
6479 | return ret; |
6480 | } |
6481 | |
6482 | |
6483 | static double CalculateUrgentLatency( |
6484 | double UrgentLatencyPixelDataOnly, |
6485 | double UrgentLatencyPixelMixedWithVMData, |
6486 | double UrgentLatencyVMDataOnly, |
6487 | bool_Bool DoUrgentLatencyAdjustment, |
6488 | double UrgentLatencyAdjustmentFabricClockComponent, |
6489 | double UrgentLatencyAdjustmentFabricClockReference, |
6490 | double FabricClock) |
6491 | { |
6492 | double ret; |
6493 | |
6494 | ret = dml_max3(UrgentLatencyPixelDataOnly, UrgentLatencyPixelMixedWithVMData, UrgentLatencyVMDataOnly); |
6495 | if (DoUrgentLatencyAdjustment == true1) { |
6496 | ret = ret + UrgentLatencyAdjustmentFabricClockComponent * (UrgentLatencyAdjustmentFabricClockReference / FabricClock - 1); |
6497 | } |
6498 | return ret; |
6499 | } |
6500 | |
6501 | static noinline_for_stack__attribute__((__noinline__)) void UseMinimumDCFCLK( |
6502 | struct display_mode_lib *mode_lib, |
6503 | struct vba_vars_st *v, |
6504 | int MaxPrefetchMode, |
6505 | int ReorderingBytes) |
6506 | { |
6507 | double NormalEfficiency = 0; |
6508 | double PTEEfficiency = 0; |
6509 | double TotalMaxPrefetchFlipDPTERowBandwidth[DC__VOLTAGE_STATES20][2] = { { 0 } }; |
6510 | unsigned int i, j, k; |
6511 | |
6512 | NormalEfficiency = (v->HostVMEnable == true1 ? v->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData |
6513 | : v->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelDataOnly) / 100.0; |
6514 | PTEEfficiency = (v->HostVMEnable == true1 ? v->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly |
6515 | / v->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData : 1.0); |
6516 | for (i = 0; i < mode_lib->soc.num_states; ++i) { |
6517 | for (j = 0; j <= 1; ++j) { |
6518 | double PixelDCFCLKCyclesRequiredInPrefetch[DC__NUM_DPP__MAX8] = { 0 }; |
6519 | double PrefetchPixelLinesTime[DC__NUM_DPP__MAX8] = { 0 }; |
6520 | double DCFCLKRequiredForPeakBandwidthPerPlane[DC__NUM_DPP__MAX8] = { 0 }; |
6521 | double DynamicMetadataVMExtraLatency[DC__NUM_DPP__MAX8] = { 0 }; |
6522 | double MinimumTWait = 0; |
6523 | double NonDPTEBandwidth = 0; |
6524 | double DPTEBandwidth = 0; |
6525 | double DCFCLKRequiredForAverageBandwidth = 0; |
6526 | double ExtraLatencyBytes = 0; |
6527 | double ExtraLatencyCycles = 0; |
6528 | double DCFCLKRequiredForPeakBandwidth = 0; |
6529 | int NoOfDPPState[DC__NUM_DPP__MAX8] = { 0 }; |
6530 | double MinimumTvmPlus2Tr0 = 0; |
6531 | |
6532 | TotalMaxPrefetchFlipDPTERowBandwidth[i][j] = 0; |
6533 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
6534 | TotalMaxPrefetchFlipDPTERowBandwidth[i][j] = TotalMaxPrefetchFlipDPTERowBandwidth[i][j] |
6535 | + v->NoOfDPP[i][j][k] * v->DPTEBytesPerRow[i][j][k] / (15.75 * v->HTotal[k] / v->PixelClock[k]); |
6536 | } |
6537 | |
6538 | for (k = 0; k <= v->NumberOfActivePlanes - 1; ++k) { |
6539 | NoOfDPPState[k] = v->NoOfDPP[i][j][k]; |
6540 | } |
6541 | |
6542 | MinimumTWait = CalculateTWait(MaxPrefetchMode, v->FinalDRAMClockChangeLatency, v->UrgLatency[i], v->SREnterPlusExitTime); |
6543 | NonDPTEBandwidth = v->TotalVActivePixelBandwidth[i][j] + v->TotalVActiveCursorBandwidth[i][j] + v->TotalMetaRowBandwidth[i][j]; |
6544 | DPTEBandwidth = (v->HostVMEnable == true1 || v->ImmediateFlipRequirement[0] == dm_immediate_flip_required) ? |
6545 | TotalMaxPrefetchFlipDPTERowBandwidth[i][j] : v->TotalDPTERowBandwidth[i][j]; |
6546 | DCFCLKRequiredForAverageBandwidth = dml_max3(v->ProjectedDCFCLKDeepSleep[i][j], |
6547 | (NonDPTEBandwidth + v->TotalDPTERowBandwidth[i][j]) / v->ReturnBusWidth / (v->MaxAveragePercentOfIdealSDPPortBWDisplayCanUseInNormalSystemOperation / 100), |
6548 | (NonDPTEBandwidth + DPTEBandwidth / PTEEfficiency) / NormalEfficiency / v->ReturnBusWidth); |
6549 | |
6550 | ExtraLatencyBytes = CalculateExtraLatencyBytes(ReorderingBytes, v->TotalNumberOfActiveDPP[i][j], v->PixelChunkSizeInKByte, v->TotalNumberOfDCCActiveDPP[i][j], |
6551 | v->MetaChunkSize, v->GPUVMEnable, v->HostVMEnable, v->NumberOfActivePlanes, NoOfDPPState, v->dpte_group_bytes, |
6552 | v->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyPixelMixedWithVMData, v->PercentOfIdealDRAMFabricAndSDPPortBWReceivedAfterUrgLatencyVMDataOnly, |
6553 | v->HostVMMinPageSize, v->HostVMMaxNonCachedPageTableLevels); |
6554 | ExtraLatencyCycles = v->RoundTripPingLatencyCycles + 32 + ExtraLatencyBytes / NormalEfficiency / v->ReturnBusWidth; |
6555 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
6556 | double DCFCLKCyclesRequiredInPrefetch = { 0 }; |
6557 | double ExpectedPrefetchBWAcceleration = { 0 }; |
6558 | double PrefetchTime = { 0 }; |
6559 | |
6560 | PixelDCFCLKCyclesRequiredInPrefetch[k] = (v->PrefetchLinesY[i][j][k] * v->swath_width_luma_ub_all_states[i][j][k] * v->BytePerPixelY[k] |
6561 | + v->PrefetchLinesC[i][j][k] * v->swath_width_chroma_ub_all_states[i][j][k] * v->BytePerPixelC[k]) / NormalEfficiency / v->ReturnBusWidth; |
6562 | DCFCLKCyclesRequiredInPrefetch = 2 * ExtraLatencyCycles / NoOfDPPState[k] + v->PDEAndMetaPTEBytesPerFrame[i][j][k] / PTEEfficiency |
6563 | / NormalEfficiency / v->ReturnBusWidth * (v->GPUVMMaxPageTableLevels > 2 ? 1 : 0) + 2 * v->DPTEBytesPerRow[i][j][k] / PTEEfficiency |
6564 | / NormalEfficiency / v->ReturnBusWidth + 2 * v->MetaRowBytes[i][j][k] / NormalEfficiency / v->ReturnBusWidth + PixelDCFCLKCyclesRequiredInPrefetch[k]; |
6565 | PrefetchPixelLinesTime[k] = dml_max(v->PrefetchLinesY[i][j][k], v->PrefetchLinesC[i][j][k]) * v->HTotal[k] / v->PixelClock[k]; |
6566 | ExpectedPrefetchBWAcceleration = (v->VActivePixelBandwidth[i][j][k] + v->VActiveCursorBandwidth[i][j][k]) / (v->ReadBandwidthLuma[k] + v->ReadBandwidthChroma[k]); |
6567 | DynamicMetadataVMExtraLatency[k] = (v->GPUVMEnable == true1 && v->DynamicMetadataEnable[k] == true1 && v->DynamicMetadataVMEnabled == true1) ? |
6568 | v->UrgLatency[i] * v->GPUVMMaxPageTableLevels * (v->HostVMEnable == true1 ? v->HostVMMaxNonCachedPageTableLevels + 1 : 1) : 0; |
6569 | PrefetchTime = (v->MaximumVStartup[i][j][k] - 1) * v->HTotal[k] / v->PixelClock[k] - MinimumTWait - v->UrgLatency[i] * ((v->GPUVMMaxPageTableLevels <= 2 ? v->GPUVMMaxPageTableLevels |
6570 | : v->GPUVMMaxPageTableLevels - 2) * (v->HostVMEnable == true1 ? v->HostVMMaxNonCachedPageTableLevels + 1 : 1) - 1) - DynamicMetadataVMExtraLatency[k]; |
6571 | |
6572 | if (PrefetchTime > 0) { |
6573 | double ExpectedVRatioPrefetch = { 0 }; |
6574 | ExpectedVRatioPrefetch = PrefetchPixelLinesTime[k] / (PrefetchTime * PixelDCFCLKCyclesRequiredInPrefetch[k] / DCFCLKCyclesRequiredInPrefetch); |
6575 | DCFCLKRequiredForPeakBandwidthPerPlane[k] = NoOfDPPState[k] * PixelDCFCLKCyclesRequiredInPrefetch[k] / PrefetchPixelLinesTime[k] |
6576 | * dml_max(1.0, ExpectedVRatioPrefetch) * dml_max(1.0, ExpectedVRatioPrefetch / 4) * ExpectedPrefetchBWAcceleration; |
6577 | if (v->HostVMEnable == true1 || v->ImmediateFlipRequirement[0] == dm_immediate_flip_required) { |
6578 | DCFCLKRequiredForPeakBandwidthPerPlane[k] = DCFCLKRequiredForPeakBandwidthPerPlane[k] |
6579 | + NoOfDPPState[k] * DPTEBandwidth / PTEEfficiency / NormalEfficiency / v->ReturnBusWidth; |
6580 | } |
6581 | } else { |
6582 | DCFCLKRequiredForPeakBandwidthPerPlane[k] = v->DCFCLKPerState[i]; |
6583 | } |
6584 | if (v->DynamicMetadataEnable[k] == true1) { |
6585 | double TsetupPipe = { 0 }; |
6586 | double TdmbfPipe = { 0 }; |
6587 | double TdmsksPipe = { 0 }; |
6588 | double TdmecPipe = { 0 }; |
6589 | double AllowedTimeForUrgentExtraLatency = { 0 }; |
6590 | |
6591 | CalculateDynamicMetadataParameters( |
6592 | v->MaxInterDCNTileRepeaters, |
6593 | v->RequiredDPPCLK[i][j][k], |
6594 | v->RequiredDISPCLK[i][j], |
6595 | v->ProjectedDCFCLKDeepSleep[i][j], |
6596 | v->PixelClock[k], |
6597 | v->HTotal[k], |
6598 | v->VTotal[k] - v->VActive[k], |
6599 | v->DynamicMetadataTransmittedBytes[k], |
6600 | v->DynamicMetadataLinesBeforeActiveRequired[k], |
6601 | v->Interlace[k], |
6602 | v->ProgressiveToInterlaceUnitInOPP, |
6603 | &TsetupPipe, |
6604 | &TdmbfPipe, |
6605 | &TdmecPipe, |
6606 | &TdmsksPipe); |
6607 | AllowedTimeForUrgentExtraLatency = v->MaximumVStartup[i][j][k] * v->HTotal[k] / v->PixelClock[k] - MinimumTWait - TsetupPipe |
6608 | - TdmbfPipe - TdmecPipe - TdmsksPipe - DynamicMetadataVMExtraLatency[k]; |
6609 | if (AllowedTimeForUrgentExtraLatency > 0) { |
6610 | DCFCLKRequiredForPeakBandwidthPerPlane[k] = dml_max(DCFCLKRequiredForPeakBandwidthPerPlane[k], |
6611 | ExtraLatencyCycles / AllowedTimeForUrgentExtraLatency); |
6612 | } else { |
6613 | DCFCLKRequiredForPeakBandwidthPerPlane[k] = v->DCFCLKPerState[i]; |
6614 | } |
6615 | } |
6616 | } |
6617 | DCFCLKRequiredForPeakBandwidth = 0; |
6618 | for (k = 0; k <= v->NumberOfActivePlanes - 1; ++k) { |
6619 | DCFCLKRequiredForPeakBandwidth = DCFCLKRequiredForPeakBandwidth + DCFCLKRequiredForPeakBandwidthPerPlane[k]; |
6620 | } |
6621 | MinimumTvmPlus2Tr0 = v->UrgLatency[i] * (v->GPUVMEnable == true1 ? (v->HostVMEnable == true1 ? |
6622 | (v->GPUVMMaxPageTableLevels + 2) * (v->HostVMMaxNonCachedPageTableLevels + 1) - 1 : v->GPUVMMaxPageTableLevels + 1) : 0); |
6623 | for (k = 0; k < v->NumberOfActivePlanes; ++k) { |
6624 | double MaximumTvmPlus2Tr0PlusTsw = { 0 }; |
6625 | MaximumTvmPlus2Tr0PlusTsw = (v->MaximumVStartup[i][j][k] - 2) * v->HTotal[k] / v->PixelClock[k] - MinimumTWait - DynamicMetadataVMExtraLatency[k]; |
6626 | if (MaximumTvmPlus2Tr0PlusTsw <= MinimumTvmPlus2Tr0 + PrefetchPixelLinesTime[k] / 4) { |
6627 | DCFCLKRequiredForPeakBandwidth = v->DCFCLKPerState[i]; |
6628 | } else { |
6629 | DCFCLKRequiredForPeakBandwidth = dml_max3(DCFCLKRequiredForPeakBandwidth, 2 * ExtraLatencyCycles |
6630 | / (MaximumTvmPlus2Tr0PlusTsw - MinimumTvmPlus2Tr0 - PrefetchPixelLinesTime[k] / 4), |
6631 | (2 * ExtraLatencyCycles + PixelDCFCLKCyclesRequiredInPrefetch[k]) / (MaximumTvmPlus2Tr0PlusTsw - MinimumTvmPlus2Tr0)); |
6632 | } |
6633 | } |
6634 | v->DCFCLKState[i][j] = dml_min(v->DCFCLKPerState[i], 1.05 * (1 + mode_lib->vba.PercentMarginOverMinimumRequiredDCFCLK / 100) |
6635 | * dml_max(DCFCLKRequiredForAverageBandwidth, DCFCLKRequiredForPeakBandwidth)); |
6636 | } |
6637 | } |
6638 | } |
6639 | |
6640 | #endif /* CONFIG_DRM_AMD_DC_DCN */ |