| File: | dev/pci/drm/amd/display/dc/dml/dcn32/display_mode_vba_util_32.c |
| Warning: | line 4077, column 3 Value stored to 'TimeForFetchingRowInVBlank' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | /* |
| 2 | * Copyright 2022 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 | #include "display_mode_vba_util_32.h" |
| 26 | #include "../dml_inline_defs.h" |
| 27 | #include "display_mode_vba_32.h" |
| 28 | #include "../display_mode_lib.h" |
| 29 | |
| 30 | #define DCN32_MAX_FMT_420_BUFFER_WIDTH4096 4096 |
| 31 | |
| 32 | unsigned int dml32_dscceComputeDelay( |
| 33 | unsigned int bpc, |
| 34 | double BPP, |
| 35 | unsigned int sliceWidth, |
| 36 | unsigned int numSlices, |
| 37 | enum output_format_class pixelFormat, |
| 38 | enum output_encoder_class Output) |
| 39 | { |
| 40 | // valid bpc = source bits per component in the set of {8, 10, 12} |
| 41 | // valid bpp = increments of 1/16 of a bit |
| 42 | // min = 6/7/8 in N420/N422/444, respectively |
| 43 | // max = such that compression is 1:1 |
| 44 | //valid sliceWidth = number of pixels per slice line, |
| 45 | // must be less than or equal to 5184/numSlices (or 4096/numSlices in 420 mode) |
| 46 | //valid numSlices = number of slices in the horiziontal direction per DSC engine in the set of {1, 2, 3, 4} |
| 47 | //valid pixelFormat = pixel/color format in the set of {:N444_RGB, :S422, :N422, :N420} |
| 48 | |
| 49 | // fixed value |
| 50 | unsigned int rcModelSize = 8192; |
| 51 | |
| 52 | // N422/N420 operate at 2 pixels per clock |
| 53 | unsigned int pixelsPerClock, lstall, D, initalXmitDelay, w, s, ix, wx, p, l0, a, ax, L, |
| 54 | Delay, pixels; |
| 55 | |
| 56 | if (pixelFormat == dm_420) |
| 57 | pixelsPerClock = 2; |
| 58 | else if (pixelFormat == dm_n422) |
| 59 | pixelsPerClock = 2; |
| 60 | // #all other modes operate at 1 pixel per clock |
| 61 | else |
| 62 | pixelsPerClock = 1; |
| 63 | |
| 64 | //initial transmit delay as per PPS |
| 65 | initalXmitDelay = dml_round(rcModelSize / 2.0 / BPP / pixelsPerClock); |
| 66 | |
| 67 | //compute ssm delay |
| 68 | if (bpc == 8) |
| 69 | D = 81; |
| 70 | else if (bpc == 10) |
| 71 | D = 89; |
| 72 | else |
| 73 | D = 113; |
| 74 | |
| 75 | //divide by pixel per cycle to compute slice width as seen by DSC |
| 76 | w = sliceWidth / pixelsPerClock; |
| 77 | |
| 78 | //422 mode has an additional cycle of delay |
| 79 | if (pixelFormat == dm_420 || pixelFormat == dm_444 || pixelFormat == dm_n422) |
| 80 | s = 0; |
| 81 | else |
| 82 | s = 1; |
| 83 | |
| 84 | //main calculation for the dscce |
| 85 | ix = initalXmitDelay + 45; |
| 86 | wx = (w + 2) / 3; |
| 87 | p = 3 * wx - w; |
| 88 | l0 = ix / w; |
| 89 | a = ix + p * l0; |
| 90 | ax = (a + 2) / 3 + D + 6 + 1; |
| 91 | L = (ax + wx - 1) / wx; |
| 92 | if ((ix % w) == 0 && p != 0) |
| 93 | lstall = 1; |
| 94 | else |
| 95 | lstall = 0; |
| 96 | Delay = L * wx * (numSlices - 1) + ax + s + lstall + 22; |
| 97 | |
| 98 | //dsc processes 3 pixel containers per cycle and a container can contain 1 or 2 pixels |
| 99 | pixels = Delay * 3 * pixelsPerClock; |
| 100 | |
| 101 | #ifdef __DML_VBA_DEBUG__ |
| 102 | dml_print("DML::%s: bpc: %d\n", __func__, bpc){do { } while(0); }; |
| 103 | dml_print("DML::%s: BPP: %f\n", __func__, BPP){do { } while(0); }; |
| 104 | dml_print("DML::%s: sliceWidth: %d\n", __func__, sliceWidth){do { } while(0); }; |
| 105 | dml_print("DML::%s: numSlices: %d\n", __func__, numSlices){do { } while(0); }; |
| 106 | dml_print("DML::%s: pixelFormat: %d\n", __func__, pixelFormat){do { } while(0); }; |
| 107 | dml_print("DML::%s: Output: %d\n", __func__, Output){do { } while(0); }; |
| 108 | dml_print("DML::%s: pixels: %d\n", __func__, pixels){do { } while(0); }; |
| 109 | #endif |
| 110 | |
| 111 | return pixels; |
| 112 | } |
| 113 | |
| 114 | unsigned int dml32_dscComputeDelay(enum output_format_class pixelFormat, enum output_encoder_class Output) |
| 115 | { |
| 116 | unsigned int Delay = 0; |
| 117 | |
| 118 | if (pixelFormat == dm_420) { |
| 119 | // sfr |
| 120 | Delay = Delay + 2; |
| 121 | // dsccif |
| 122 | Delay = Delay + 0; |
| 123 | // dscc - input deserializer |
| 124 | Delay = Delay + 3; |
| 125 | // dscc gets pixels every other cycle |
| 126 | Delay = Delay + 2; |
| 127 | // dscc - input cdc fifo |
| 128 | Delay = Delay + 12; |
| 129 | // dscc gets pixels every other cycle |
| 130 | Delay = Delay + 13; |
| 131 | // dscc - cdc uncertainty |
| 132 | Delay = Delay + 2; |
| 133 | // dscc - output cdc fifo |
| 134 | Delay = Delay + 7; |
| 135 | // dscc gets pixels every other cycle |
| 136 | Delay = Delay + 3; |
| 137 | // dscc - cdc uncertainty |
| 138 | Delay = Delay + 2; |
| 139 | // dscc - output serializer |
| 140 | Delay = Delay + 1; |
| 141 | // sft |
| 142 | Delay = Delay + 1; |
| 143 | } else if (pixelFormat == dm_n422 || (pixelFormat != dm_444)) { |
| 144 | // sfr |
| 145 | Delay = Delay + 2; |
| 146 | // dsccif |
| 147 | Delay = Delay + 1; |
| 148 | // dscc - input deserializer |
| 149 | Delay = Delay + 5; |
| 150 | // dscc - input cdc fifo |
| 151 | Delay = Delay + 25; |
| 152 | // dscc - cdc uncertainty |
| 153 | Delay = Delay + 2; |
| 154 | // dscc - output cdc fifo |
| 155 | Delay = Delay + 10; |
| 156 | // dscc - cdc uncertainty |
| 157 | Delay = Delay + 2; |
| 158 | // dscc - output serializer |
| 159 | Delay = Delay + 1; |
| 160 | // sft |
| 161 | Delay = Delay + 1; |
| 162 | } else { |
| 163 | // sfr |
| 164 | Delay = Delay + 2; |
| 165 | // dsccif |
| 166 | Delay = Delay + 0; |
| 167 | // dscc - input deserializer |
| 168 | Delay = Delay + 3; |
| 169 | // dscc - input cdc fifo |
| 170 | Delay = Delay + 12; |
| 171 | // dscc - cdc uncertainty |
| 172 | Delay = Delay + 2; |
| 173 | // dscc - output cdc fifo |
| 174 | Delay = Delay + 7; |
| 175 | // dscc - output serializer |
| 176 | Delay = Delay + 1; |
| 177 | // dscc - cdc uncertainty |
| 178 | Delay = Delay + 2; |
| 179 | // sft |
| 180 | Delay = Delay + 1; |
| 181 | } |
| 182 | |
| 183 | return Delay; |
| 184 | } |
| 185 | |
| 186 | |
| 187 | bool_Bool IsVertical(enum dm_rotation_angle Scan) |
| 188 | { |
| 189 | bool_Bool is_vert = false0; |
| 190 | |
| 191 | if (Scan == dm_rotation_90 || Scan == dm_rotation_90m || Scan == dm_rotation_270 || Scan == dm_rotation_270m) |
| 192 | is_vert = true1; |
| 193 | else |
| 194 | is_vert = false0; |
| 195 | return is_vert; |
| 196 | } |
| 197 | |
| 198 | void dml32_CalculateSinglePipeDPPCLKAndSCLThroughput( |
| 199 | double HRatio, |
| 200 | double HRatioChroma, |
| 201 | double VRatio, |
| 202 | double VRatioChroma, |
| 203 | double MaxDCHUBToPSCLThroughput, |
| 204 | double MaxPSCLToLBThroughput, |
| 205 | double PixelClock, |
| 206 | enum source_format_class SourcePixelFormat, |
| 207 | unsigned int HTaps, |
| 208 | unsigned int HTapsChroma, |
| 209 | unsigned int VTaps, |
| 210 | unsigned int VTapsChroma, |
| 211 | |
| 212 | /* output */ |
| 213 | double *PSCL_THROUGHPUT, |
| 214 | double *PSCL_THROUGHPUT_CHROMA, |
| 215 | double *DPPCLKUsingSingleDPP) |
| 216 | { |
| 217 | double DPPCLKUsingSingleDPPLuma; |
| 218 | double DPPCLKUsingSingleDPPChroma; |
| 219 | |
| 220 | if (HRatio > 1) { |
| 221 | *PSCL_THROUGHPUT = dml_min(MaxDCHUBToPSCLThroughput, MaxPSCLToLBThroughput * HRatio / |
| 222 | dml_ceil((double) HTaps / 6.0, 1.0)); |
| 223 | } else { |
| 224 | *PSCL_THROUGHPUT = dml_min(MaxDCHUBToPSCLThroughput, MaxPSCLToLBThroughput); |
| 225 | } |
| 226 | |
| 227 | DPPCLKUsingSingleDPPLuma = PixelClock * dml_max3(VTaps / 6 * dml_min(1, HRatio), HRatio * VRatio / |
| 228 | *PSCL_THROUGHPUT, 1); |
| 229 | |
| 230 | if ((HTaps > 6 || VTaps > 6) && DPPCLKUsingSingleDPPLuma < 2 * PixelClock) |
| 231 | DPPCLKUsingSingleDPPLuma = 2 * PixelClock; |
| 232 | |
| 233 | if ((SourcePixelFormat != dm_420_8 && SourcePixelFormat != dm_420_10 && SourcePixelFormat != dm_420_12 && |
| 234 | SourcePixelFormat != dm_rgbe_alpha)) { |
| 235 | *PSCL_THROUGHPUT_CHROMA = 0; |
| 236 | *DPPCLKUsingSingleDPP = DPPCLKUsingSingleDPPLuma; |
| 237 | } else { |
| 238 | if (HRatioChroma > 1) { |
| 239 | *PSCL_THROUGHPUT_CHROMA = dml_min(MaxDCHUBToPSCLThroughput, MaxPSCLToLBThroughput * |
| 240 | HRatioChroma / dml_ceil((double) HTapsChroma / 6.0, 1.0)); |
| 241 | } else { |
| 242 | *PSCL_THROUGHPUT_CHROMA = dml_min(MaxDCHUBToPSCLThroughput, MaxPSCLToLBThroughput); |
| 243 | } |
| 244 | DPPCLKUsingSingleDPPChroma = PixelClock * dml_max3(VTapsChroma / 6 * dml_min(1, HRatioChroma), |
| 245 | HRatioChroma * VRatioChroma / *PSCL_THROUGHPUT_CHROMA, 1); |
| 246 | if ((HTapsChroma > 6 || VTapsChroma > 6) && DPPCLKUsingSingleDPPChroma < 2 * PixelClock) |
| 247 | DPPCLKUsingSingleDPPChroma = 2 * PixelClock; |
| 248 | *DPPCLKUsingSingleDPP = dml_max(DPPCLKUsingSingleDPPLuma, DPPCLKUsingSingleDPPChroma); |
| 249 | } |
| 250 | } |
| 251 | |
| 252 | void dml32_CalculateBytePerPixelAndBlockSizes( |
| 253 | enum source_format_class SourcePixelFormat, |
| 254 | enum dm_swizzle_mode SurfaceTiling, |
| 255 | |
| 256 | /* Output */ |
| 257 | unsigned int *BytePerPixelY, |
| 258 | unsigned int *BytePerPixelC, |
| 259 | double *BytePerPixelDETY, |
| 260 | double *BytePerPixelDETC, |
| 261 | unsigned int *BlockHeight256BytesY, |
| 262 | unsigned int *BlockHeight256BytesC, |
| 263 | unsigned int *BlockWidth256BytesY, |
| 264 | unsigned int *BlockWidth256BytesC, |
| 265 | unsigned int *MacroTileHeightY, |
| 266 | unsigned int *MacroTileHeightC, |
| 267 | unsigned int *MacroTileWidthY, |
| 268 | unsigned int *MacroTileWidthC) |
| 269 | { |
| 270 | if (SourcePixelFormat == dm_444_64) { |
| 271 | *BytePerPixelDETY = 8; |
| 272 | *BytePerPixelDETC = 0; |
| 273 | *BytePerPixelY = 8; |
| 274 | *BytePerPixelC = 0; |
| 275 | } else if (SourcePixelFormat == dm_444_32 || SourcePixelFormat == dm_rgbe) { |
| 276 | *BytePerPixelDETY = 4; |
| 277 | *BytePerPixelDETC = 0; |
| 278 | *BytePerPixelY = 4; |
| 279 | *BytePerPixelC = 0; |
| 280 | } else if (SourcePixelFormat == dm_444_16) { |
| 281 | *BytePerPixelDETY = 2; |
| 282 | *BytePerPixelDETC = 0; |
| 283 | *BytePerPixelY = 2; |
| 284 | *BytePerPixelC = 0; |
| 285 | } else if (SourcePixelFormat == dm_444_8) { |
| 286 | *BytePerPixelDETY = 1; |
| 287 | *BytePerPixelDETC = 0; |
| 288 | *BytePerPixelY = 1; |
| 289 | *BytePerPixelC = 0; |
| 290 | } else if (SourcePixelFormat == dm_rgbe_alpha) { |
| 291 | *BytePerPixelDETY = 4; |
| 292 | *BytePerPixelDETC = 1; |
| 293 | *BytePerPixelY = 4; |
| 294 | *BytePerPixelC = 1; |
| 295 | } else if (SourcePixelFormat == dm_420_8) { |
| 296 | *BytePerPixelDETY = 1; |
| 297 | *BytePerPixelDETC = 2; |
| 298 | *BytePerPixelY = 1; |
| 299 | *BytePerPixelC = 2; |
| 300 | } else if (SourcePixelFormat == dm_420_12) { |
| 301 | *BytePerPixelDETY = 2; |
| 302 | *BytePerPixelDETC = 4; |
| 303 | *BytePerPixelY = 2; |
| 304 | *BytePerPixelC = 4; |
| 305 | } else { |
| 306 | *BytePerPixelDETY = 4.0 / 3; |
| 307 | *BytePerPixelDETC = 8.0 / 3; |
| 308 | *BytePerPixelY = 2; |
| 309 | *BytePerPixelC = 4; |
| 310 | } |
| 311 | #ifdef __DML_VBA_DEBUG__ |
| 312 | dml_print("DML::%s: SourcePixelFormat = %d\n", __func__, SourcePixelFormat){do { } while(0); }; |
| 313 | dml_print("DML::%s: BytePerPixelDETY = %f\n", __func__, *BytePerPixelDETY){do { } while(0); }; |
| 314 | dml_print("DML::%s: BytePerPixelDETC = %f\n", __func__, *BytePerPixelDETC){do { } while(0); }; |
| 315 | dml_print("DML::%s: BytePerPixelY = %d\n", __func__, *BytePerPixelY){do { } while(0); }; |
| 316 | dml_print("DML::%s: BytePerPixelC = %d\n", __func__, *BytePerPixelC){do { } while(0); }; |
| 317 | #endif |
| 318 | if ((SourcePixelFormat == dm_444_64 || SourcePixelFormat == dm_444_32 |
| 319 | || SourcePixelFormat == dm_444_16 |
| 320 | || SourcePixelFormat == dm_444_8 |
| 321 | || SourcePixelFormat == dm_mono_16 |
| 322 | || SourcePixelFormat == dm_mono_8 |
| 323 | || SourcePixelFormat == dm_rgbe)) { |
| 324 | if (SurfaceTiling == dm_sw_linear) |
| 325 | *BlockHeight256BytesY = 1; |
| 326 | else if (SourcePixelFormat == dm_444_64) |
| 327 | *BlockHeight256BytesY = 4; |
| 328 | else if (SourcePixelFormat == dm_444_8) |
| 329 | *BlockHeight256BytesY = 16; |
| 330 | else |
| 331 | *BlockHeight256BytesY = 8; |
| 332 | |
| 333 | *BlockWidth256BytesY = 256U / *BytePerPixelY / *BlockHeight256BytesY; |
| 334 | *BlockHeight256BytesC = 0; |
| 335 | *BlockWidth256BytesC = 0; |
| 336 | } else { |
| 337 | if (SurfaceTiling == dm_sw_linear) { |
| 338 | *BlockHeight256BytesY = 1; |
| 339 | *BlockHeight256BytesC = 1; |
| 340 | } else if (SourcePixelFormat == dm_rgbe_alpha) { |
| 341 | *BlockHeight256BytesY = 8; |
| 342 | *BlockHeight256BytesC = 16; |
| 343 | } else if (SourcePixelFormat == dm_420_8) { |
| 344 | *BlockHeight256BytesY = 16; |
| 345 | *BlockHeight256BytesC = 8; |
| 346 | } else { |
| 347 | *BlockHeight256BytesY = 8; |
| 348 | *BlockHeight256BytesC = 8; |
| 349 | } |
| 350 | *BlockWidth256BytesY = 256U / *BytePerPixelY / *BlockHeight256BytesY; |
| 351 | *BlockWidth256BytesC = 256U / *BytePerPixelC / *BlockHeight256BytesC; |
| 352 | } |
| 353 | #ifdef __DML_VBA_DEBUG__ |
| 354 | dml_print("DML::%s: BlockWidth256BytesY = %d\n", __func__, *BlockWidth256BytesY){do { } while(0); }; |
| 355 | dml_print("DML::%s: BlockHeight256BytesY = %d\n", __func__, *BlockHeight256BytesY){do { } while(0); }; |
| 356 | dml_print("DML::%s: BlockWidth256BytesC = %d\n", __func__, *BlockWidth256BytesC){do { } while(0); }; |
| 357 | dml_print("DML::%s: BlockHeight256BytesC = %d\n", __func__, *BlockHeight256BytesC){do { } while(0); }; |
| 358 | #endif |
| 359 | |
| 360 | if (SurfaceTiling == dm_sw_linear) { |
| 361 | *MacroTileHeightY = *BlockHeight256BytesY; |
| 362 | *MacroTileWidthY = 256 / *BytePerPixelY / *MacroTileHeightY; |
| 363 | *MacroTileHeightC = *BlockHeight256BytesC; |
| 364 | if (*MacroTileHeightC == 0) |
| 365 | *MacroTileWidthC = 0; |
| 366 | else |
| 367 | *MacroTileWidthC = 256 / *BytePerPixelC / *MacroTileHeightC; |
| 368 | } else if (SurfaceTiling == dm_sw_64kb_d || SurfaceTiling == dm_sw_64kb_d_t || |
| 369 | SurfaceTiling == dm_sw_64kb_d_x || SurfaceTiling == dm_sw_64kb_r_x) { |
| 370 | *MacroTileHeightY = 16 * *BlockHeight256BytesY; |
| 371 | *MacroTileWidthY = 65536 / *BytePerPixelY / *MacroTileHeightY; |
| 372 | *MacroTileHeightC = 16 * *BlockHeight256BytesC; |
| 373 | if (*MacroTileHeightC == 0) |
| 374 | *MacroTileWidthC = 0; |
| 375 | else |
| 376 | *MacroTileWidthC = 65536 / *BytePerPixelC / *MacroTileHeightC; |
| 377 | } else { |
| 378 | *MacroTileHeightY = 32 * *BlockHeight256BytesY; |
| 379 | *MacroTileWidthY = 65536 * 4 / *BytePerPixelY / *MacroTileHeightY; |
| 380 | *MacroTileHeightC = 32 * *BlockHeight256BytesC; |
| 381 | if (*MacroTileHeightC == 0) |
| 382 | *MacroTileWidthC = 0; |
| 383 | else |
| 384 | *MacroTileWidthC = 65536 * 4 / *BytePerPixelC / *MacroTileHeightC; |
| 385 | } |
| 386 | |
| 387 | #ifdef __DML_VBA_DEBUG__ |
| 388 | dml_print("DML::%s: MacroTileWidthY = %d\n", __func__, *MacroTileWidthY){do { } while(0); }; |
| 389 | dml_print("DML::%s: MacroTileHeightY = %d\n", __func__, *MacroTileHeightY){do { } while(0); }; |
| 390 | dml_print("DML::%s: MacroTileWidthC = %d\n", __func__, *MacroTileWidthC){do { } while(0); }; |
| 391 | dml_print("DML::%s: MacroTileHeightC = %d\n", __func__, *MacroTileHeightC){do { } while(0); }; |
| 392 | #endif |
| 393 | } // CalculateBytePerPixelAndBlockSizes |
| 394 | |
| 395 | void dml32_CalculateSwathAndDETConfiguration( |
| 396 | unsigned int DETSizeOverride[], |
| 397 | enum dm_use_mall_for_pstate_change_mode UseMALLForPStateChange[], |
| 398 | unsigned int ConfigReturnBufferSizeInKByte, |
| 399 | unsigned int MaxTotalDETInKByte, |
| 400 | unsigned int MinCompressedBufferSizeInKByte, |
| 401 | double ForceSingleDPP, |
| 402 | unsigned int NumberOfActiveSurfaces, |
| 403 | unsigned int nomDETInKByte, |
| 404 | enum unbounded_requesting_policy UseUnboundedRequestingFinal, |
| 405 | bool_Bool DisableUnboundRequestIfCompBufReservedSpaceNeedAdjustment, |
| 406 | unsigned int PixelChunkSizeKBytes, |
| 407 | unsigned int ROBSizeKBytes, |
| 408 | unsigned int CompressedBufferSegmentSizeInkByteFinal, |
| 409 | enum output_encoder_class Output[], |
| 410 | double ReadBandwidthLuma[], |
| 411 | double ReadBandwidthChroma[], |
| 412 | double MaximumSwathWidthLuma[], |
| 413 | double MaximumSwathWidthChroma[], |
| 414 | enum dm_rotation_angle SourceRotation[], |
| 415 | bool_Bool ViewportStationary[], |
| 416 | enum source_format_class SourcePixelFormat[], |
| 417 | enum dm_swizzle_mode SurfaceTiling[], |
| 418 | unsigned int ViewportWidth[], |
| 419 | unsigned int ViewportHeight[], |
| 420 | unsigned int ViewportXStart[], |
| 421 | unsigned int ViewportYStart[], |
| 422 | unsigned int ViewportXStartC[], |
| 423 | unsigned int ViewportYStartC[], |
| 424 | unsigned int SurfaceWidthY[], |
| 425 | unsigned int SurfaceWidthC[], |
| 426 | unsigned int SurfaceHeightY[], |
| 427 | unsigned int SurfaceHeightC[], |
| 428 | unsigned int Read256BytesBlockHeightY[], |
| 429 | unsigned int Read256BytesBlockHeightC[], |
| 430 | unsigned int Read256BytesBlockWidthY[], |
| 431 | unsigned int Read256BytesBlockWidthC[], |
| 432 | enum odm_combine_mode ODMMode[], |
| 433 | unsigned int BlendingAndTiming[], |
| 434 | unsigned int BytePerPixY[], |
| 435 | unsigned int BytePerPixC[], |
| 436 | double BytePerPixDETY[], |
| 437 | double BytePerPixDETC[], |
| 438 | unsigned int HActive[], |
| 439 | double HRatio[], |
| 440 | double HRatioChroma[], |
| 441 | unsigned int DPPPerSurface[], |
| 442 | |
| 443 | /* Output */ |
| 444 | unsigned int swath_width_luma_ub[], |
| 445 | unsigned int swath_width_chroma_ub[], |
| 446 | double SwathWidth[], |
| 447 | double SwathWidthChroma[], |
| 448 | unsigned int SwathHeightY[], |
| 449 | unsigned int SwathHeightC[], |
| 450 | unsigned int DETBufferSizeInKByte[], |
| 451 | unsigned int DETBufferSizeY[], |
| 452 | unsigned int DETBufferSizeC[], |
| 453 | bool_Bool *UnboundedRequestEnabled, |
| 454 | unsigned int *CompressedBufferSizeInkByte, |
| 455 | unsigned int *CompBufReservedSpaceKBytes, |
| 456 | bool_Bool *CompBufReservedSpaceNeedAdjustment, |
| 457 | bool_Bool ViewportSizeSupportPerSurface[], |
| 458 | bool_Bool *ViewportSizeSupport) |
| 459 | { |
| 460 | unsigned int MaximumSwathHeightY[DC__NUM_DPP__MAX8]; |
| 461 | unsigned int MaximumSwathHeightC[DC__NUM_DPP__MAX8]; |
| 462 | unsigned int RoundedUpMaxSwathSizeBytesY[DC__NUM_DPP__MAX8]; |
| 463 | unsigned int RoundedUpMaxSwathSizeBytesC[DC__NUM_DPP__MAX8]; |
| 464 | unsigned int RoundedUpSwathSizeBytesY; |
| 465 | unsigned int RoundedUpSwathSizeBytesC; |
| 466 | double SwathWidthdoubleDPP[DC__NUM_DPP__MAX8]; |
| 467 | double SwathWidthdoubleDPPChroma[DC__NUM_DPP__MAX8]; |
| 468 | unsigned int k; |
| 469 | unsigned int TotalActiveDPP = 0; |
| 470 | bool_Bool NoChromaSurfaces = true1; |
| 471 | unsigned int DETBufferSizeInKByteForSwathCalculation; |
| 472 | |
| 473 | #ifdef __DML_VBA_DEBUG__ |
| 474 | dml_print("DML::%s: ForceSingleDPP = %d\n", __func__, ForceSingleDPP){do { } while(0); }; |
| 475 | dml_print("DML::%s: ROBSizeKBytes = %d\n", __func__, ROBSizeKBytes){do { } while(0); }; |
| 476 | dml_print("DML::%s: PixelChunkSizeKBytes = %d\n", __func__, PixelChunkSizeKBytes){do { } while(0); }; |
| 477 | #endif |
| 478 | dml32_CalculateSwathWidth(ForceSingleDPP, |
| 479 | NumberOfActiveSurfaces, |
| 480 | SourcePixelFormat, |
| 481 | SourceRotation, |
| 482 | ViewportStationary, |
| 483 | ViewportWidth, |
| 484 | ViewportHeight, |
| 485 | ViewportXStart, |
| 486 | ViewportYStart, |
| 487 | ViewportXStartC, |
| 488 | ViewportYStartC, |
| 489 | SurfaceWidthY, |
| 490 | SurfaceWidthC, |
| 491 | SurfaceHeightY, |
| 492 | SurfaceHeightC, |
| 493 | ODMMode, |
| 494 | BytePerPixY, |
| 495 | BytePerPixC, |
| 496 | Read256BytesBlockHeightY, |
| 497 | Read256BytesBlockHeightC, |
| 498 | Read256BytesBlockWidthY, |
| 499 | Read256BytesBlockWidthC, |
| 500 | BlendingAndTiming, |
| 501 | HActive, |
| 502 | HRatio, |
| 503 | DPPPerSurface, |
| 504 | |
| 505 | /* Output */ |
| 506 | SwathWidthdoubleDPP, |
| 507 | SwathWidthdoubleDPPChroma, |
| 508 | SwathWidth, |
| 509 | SwathWidthChroma, |
| 510 | MaximumSwathHeightY, |
| 511 | MaximumSwathHeightC, |
| 512 | swath_width_luma_ub, |
| 513 | swath_width_chroma_ub); |
| 514 | |
| 515 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 516 | RoundedUpMaxSwathSizeBytesY[k] = swath_width_luma_ub[k] * BytePerPixDETY[k] * MaximumSwathHeightY[k]; |
| 517 | RoundedUpMaxSwathSizeBytesC[k] = swath_width_chroma_ub[k] * BytePerPixDETC[k] * MaximumSwathHeightC[k]; |
| 518 | #ifdef __DML_VBA_DEBUG__ |
| 519 | dml_print("DML::%s: k=%0d DPPPerSurface = %d\n", __func__, k, DPPPerSurface[k]){do { } while(0); }; |
| 520 | dml_print("DML::%s: k=%0d swath_width_luma_ub = %d\n", __func__, k, swath_width_luma_ub[k]){do { } while(0); }; |
| 521 | dml_print("DML::%s: k=%0d BytePerPixDETY = %f\n", __func__, k, BytePerPixDETY[k]){do { } while(0); }; |
| 522 | dml_print("DML::%s: k=%0d MaximumSwathHeightY = %d\n", __func__, k, MaximumSwathHeightY[k]){do { } while(0); }; |
| 523 | dml_print("DML::%s: k=%0d RoundedUpMaxSwathSizeBytesY = %d\n", __func__, k,{do { } while(0); } |
| 524 | RoundedUpMaxSwathSizeBytesY[k]){do { } while(0); }; |
| 525 | dml_print("DML::%s: k=%0d swath_width_chroma_ub = %d\n", __func__, k, swath_width_chroma_ub[k]){do { } while(0); }; |
| 526 | dml_print("DML::%s: k=%0d BytePerPixDETC = %f\n", __func__, k, BytePerPixDETC[k]){do { } while(0); }; |
| 527 | dml_print("DML::%s: k=%0d MaximumSwathHeightC = %d\n", __func__, k, MaximumSwathHeightC[k]){do { } while(0); }; |
| 528 | dml_print("DML::%s: k=%0d RoundedUpMaxSwathSizeBytesC = %d\n", __func__, k,{do { } while(0); } |
| 529 | RoundedUpMaxSwathSizeBytesC[k]){do { } while(0); }; |
| 530 | #endif |
| 531 | |
| 532 | if (SourcePixelFormat[k] == dm_420_10) { |
| 533 | RoundedUpMaxSwathSizeBytesY[k] = dml_ceil((unsigned int) RoundedUpMaxSwathSizeBytesY[k], 256); |
| 534 | RoundedUpMaxSwathSizeBytesC[k] = dml_ceil((unsigned int) RoundedUpMaxSwathSizeBytesC[k], 256); |
| 535 | } |
| 536 | } |
| 537 | |
| 538 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 539 | TotalActiveDPP = TotalActiveDPP + (ForceSingleDPP ? 1 : DPPPerSurface[k]); |
| 540 | if (SourcePixelFormat[k] == dm_420_8 || SourcePixelFormat[k] == dm_420_10 || |
| 541 | SourcePixelFormat[k] == dm_420_12 || SourcePixelFormat[k] == dm_rgbe_alpha) { |
| 542 | NoChromaSurfaces = false0; |
| 543 | } |
| 544 | } |
| 545 | |
| 546 | // By default, just set the reserved space to 2 pixel chunks size |
| 547 | *CompBufReservedSpaceKBytes = PixelChunkSizeKBytes * 2; |
| 548 | |
| 549 | // if unbounded req is enabled, program reserved space such that the ROB will not hold more than 8 swaths worth of data |
| 550 | // - assume worst-case compression rate of 4. [ROB size - 8 * swath_size / max_compression ratio] |
| 551 | // - assume for "narrow" vp case in which the ROB can fit 8 swaths, the DET should be big enough to do full size req |
| 552 | *CompBufReservedSpaceNeedAdjustment = ((int) ROBSizeKBytes - (int) *CompBufReservedSpaceKBytes) > (int) (RoundedUpMaxSwathSizeBytesY[0]/512); |
| 553 | |
| 554 | if (*CompBufReservedSpaceNeedAdjustment == 1) { |
| 555 | *CompBufReservedSpaceKBytes = ROBSizeKBytes - RoundedUpMaxSwathSizeBytesY[0]/512; |
| 556 | } |
| 557 | |
| 558 | #ifdef __DML_VBA_DEBUG__ |
| 559 | dml_print("DML::%s: CompBufReservedSpaceKBytes = %d\n", __func__, *CompBufReservedSpaceKBytes){do { } while(0); }; |
| 560 | dml_print("DML::%s: CompBufReservedSpaceNeedAdjustment = %d\n", __func__, *CompBufReservedSpaceNeedAdjustment){do { } while(0); }; |
| 561 | #endif |
| 562 | |
| 563 | *UnboundedRequestEnabled = dml32_UnboundedRequest(UseUnboundedRequestingFinal, TotalActiveDPP, NoChromaSurfaces, Output[0], SurfaceTiling[0], *CompBufReservedSpaceNeedAdjustment, DisableUnboundRequestIfCompBufReservedSpaceNeedAdjustment); |
| 564 | |
| 565 | dml32_CalculateDETBufferSize(DETSizeOverride, |
| 566 | UseMALLForPStateChange, |
| 567 | ForceSingleDPP, |
| 568 | NumberOfActiveSurfaces, |
| 569 | *UnboundedRequestEnabled, |
| 570 | nomDETInKByte, |
| 571 | MaxTotalDETInKByte, |
| 572 | ConfigReturnBufferSizeInKByte, |
| 573 | MinCompressedBufferSizeInKByte, |
| 574 | CompressedBufferSegmentSizeInkByteFinal, |
| 575 | SourcePixelFormat, |
| 576 | ReadBandwidthLuma, |
| 577 | ReadBandwidthChroma, |
| 578 | RoundedUpMaxSwathSizeBytesY, |
| 579 | RoundedUpMaxSwathSizeBytesC, |
| 580 | DPPPerSurface, |
| 581 | |
| 582 | /* Output */ |
| 583 | DETBufferSizeInKByte, // per hubp pipe |
| 584 | CompressedBufferSizeInkByte); |
| 585 | |
| 586 | #ifdef __DML_VBA_DEBUG__ |
| 587 | dml_print("DML::%s: TotalActiveDPP = %d\n", __func__, TotalActiveDPP){do { } while(0); }; |
| 588 | dml_print("DML::%s: nomDETInKByte = %d\n", __func__, nomDETInKByte){do { } while(0); }; |
| 589 | dml_print("DML::%s: ConfigReturnBufferSizeInKByte = %d\n", __func__, ConfigReturnBufferSizeInKByte){do { } while(0); }; |
| 590 | dml_print("DML::%s: UseUnboundedRequestingFinal = %d\n", __func__, UseUnboundedRequestingFinal){do { } while(0); }; |
| 591 | dml_print("DML::%s: UnboundedRequestEnabled = %d\n", __func__, *UnboundedRequestEnabled){do { } while(0); }; |
| 592 | dml_print("DML::%s: CompressedBufferSizeInkByte = %d\n", __func__, *CompressedBufferSizeInkByte){do { } while(0); }; |
| 593 | #endif |
| 594 | |
| 595 | *ViewportSizeSupport = true1; |
| 596 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 597 | |
| 598 | DETBufferSizeInKByteForSwathCalculation = (UseMALLForPStateChange[k] == |
| 599 | dm_use_mall_pstate_change_phantom_pipe ? 1024 : DETBufferSizeInKByte[k]); |
| 600 | #ifdef __DML_VBA_DEBUG__ |
| 601 | dml_print("DML::%s: k=%0d DETBufferSizeInKByteForSwathCalculation = %d\n", __func__, k,{do { } while(0); } |
| 602 | DETBufferSizeInKByteForSwathCalculation){do { } while(0); }; |
| 603 | #endif |
| 604 | |
| 605 | if (RoundedUpMaxSwathSizeBytesY[k] + RoundedUpMaxSwathSizeBytesC[k] <= |
| 606 | DETBufferSizeInKByteForSwathCalculation * 1024 / 2) { |
| 607 | SwathHeightY[k] = MaximumSwathHeightY[k]; |
| 608 | SwathHeightC[k] = MaximumSwathHeightC[k]; |
| 609 | RoundedUpSwathSizeBytesY = RoundedUpMaxSwathSizeBytesY[k]; |
| 610 | RoundedUpSwathSizeBytesC = RoundedUpMaxSwathSizeBytesC[k]; |
| 611 | } else if (RoundedUpMaxSwathSizeBytesY[k] >= 1.5 * RoundedUpMaxSwathSizeBytesC[k] && |
| 612 | RoundedUpMaxSwathSizeBytesY[k] / 2 + RoundedUpMaxSwathSizeBytesC[k] <= |
| 613 | DETBufferSizeInKByteForSwathCalculation * 1024 / 2) { |
| 614 | SwathHeightY[k] = MaximumSwathHeightY[k] / 2; |
| 615 | SwathHeightC[k] = MaximumSwathHeightC[k]; |
| 616 | RoundedUpSwathSizeBytesY = RoundedUpMaxSwathSizeBytesY[k] / 2; |
| 617 | RoundedUpSwathSizeBytesC = RoundedUpMaxSwathSizeBytesC[k]; |
| 618 | } else if (RoundedUpMaxSwathSizeBytesY[k] < 1.5 * RoundedUpMaxSwathSizeBytesC[k] && |
| 619 | RoundedUpMaxSwathSizeBytesY[k] + RoundedUpMaxSwathSizeBytesC[k] / 2 <= |
| 620 | DETBufferSizeInKByteForSwathCalculation * 1024 / 2) { |
| 621 | SwathHeightY[k] = MaximumSwathHeightY[k]; |
| 622 | SwathHeightC[k] = MaximumSwathHeightC[k] / 2; |
| 623 | RoundedUpSwathSizeBytesY = RoundedUpMaxSwathSizeBytesY[k]; |
| 624 | RoundedUpSwathSizeBytesC = RoundedUpMaxSwathSizeBytesC[k] / 2; |
| 625 | } else { |
| 626 | SwathHeightY[k] = MaximumSwathHeightY[k] / 2; |
| 627 | SwathHeightC[k] = MaximumSwathHeightC[k] / 2; |
| 628 | RoundedUpSwathSizeBytesY = RoundedUpMaxSwathSizeBytesY[k] / 2; |
| 629 | RoundedUpSwathSizeBytesC = RoundedUpMaxSwathSizeBytesC[k] / 2; |
| 630 | } |
| 631 | |
| 632 | if ((RoundedUpMaxSwathSizeBytesY[k] / 2 + RoundedUpMaxSwathSizeBytesC[k] / 2 > |
| 633 | DETBufferSizeInKByteForSwathCalculation * 1024 / 2) |
| 634 | || SwathWidth[k] > MaximumSwathWidthLuma[k] || (SwathHeightC[k] > 0 && |
| 635 | SwathWidthChroma[k] > MaximumSwathWidthChroma[k])) { |
| 636 | *ViewportSizeSupport = false0; |
| 637 | ViewportSizeSupportPerSurface[k] = false0; |
| 638 | } else { |
| 639 | ViewportSizeSupportPerSurface[k] = true1; |
| 640 | } |
| 641 | |
| 642 | if (SwathHeightC[k] == 0) { |
| 643 | #ifdef __DML_VBA_DEBUG__ |
| 644 | dml_print("DML::%s: k=%0d All DET for plane0\n", __func__, k){do { } while(0); }; |
| 645 | #endif |
| 646 | DETBufferSizeY[k] = DETBufferSizeInKByte[k] * 1024; |
| 647 | DETBufferSizeC[k] = 0; |
| 648 | } else if (RoundedUpSwathSizeBytesY <= 1.5 * RoundedUpSwathSizeBytesC) { |
| 649 | #ifdef __DML_VBA_DEBUG__ |
| 650 | dml_print("DML::%s: k=%0d Half DET for plane0, half for plane1\n", __func__, k){do { } while(0); }; |
| 651 | #endif |
| 652 | DETBufferSizeY[k] = DETBufferSizeInKByte[k] * 1024 / 2; |
| 653 | DETBufferSizeC[k] = DETBufferSizeInKByte[k] * 1024 / 2; |
| 654 | } else { |
| 655 | #ifdef __DML_VBA_DEBUG__ |
| 656 | dml_print("DML::%s: k=%0d 2/3 DET for plane0, 1/3 for plane1\n", __func__, k){do { } while(0); }; |
| 657 | #endif |
| 658 | DETBufferSizeY[k] = dml_floor(DETBufferSizeInKByte[k] * 1024 * 2 / 3, 1024); |
| 659 | DETBufferSizeC[k] = DETBufferSizeInKByte[k] * 1024 - DETBufferSizeY[k]; |
| 660 | } |
| 661 | |
| 662 | #ifdef __DML_VBA_DEBUG__ |
| 663 | dml_print("DML::%s: k=%0d SwathHeightY = %d\n", __func__, k, SwathHeightY[k]){do { } while(0); }; |
| 664 | dml_print("DML::%s: k=%0d SwathHeightC = %d\n", __func__, k, SwathHeightC[k]){do { } while(0); }; |
| 665 | dml_print("DML::%s: k=%0d RoundedUpMaxSwathSizeBytesY = %d\n", __func__,{do { } while(0); } |
| 666 | k, RoundedUpMaxSwathSizeBytesY[k]){do { } while(0); }; |
| 667 | dml_print("DML::%s: k=%0d RoundedUpMaxSwathSizeBytesC = %d\n", __func__,{do { } while(0); } |
| 668 | k, RoundedUpMaxSwathSizeBytesC[k]){do { } while(0); }; |
| 669 | dml_print("DML::%s: k=%0d RoundedUpSwathSizeBytesY = %d\n", __func__, k, RoundedUpSwathSizeBytesY){do { } while(0); }; |
| 670 | dml_print("DML::%s: k=%0d RoundedUpSwathSizeBytesC = %d\n", __func__, k, RoundedUpSwathSizeBytesC){do { } while(0); }; |
| 671 | dml_print("DML::%s: k=%0d DETBufferSizeInKByte = %d\n", __func__, k, DETBufferSizeInKByte[k]){do { } while(0); }; |
| 672 | dml_print("DML::%s: k=%0d DETBufferSizeY = %d\n", __func__, k, DETBufferSizeY[k]){do { } while(0); }; |
| 673 | dml_print("DML::%s: k=%0d DETBufferSizeC = %d\n", __func__, k, DETBufferSizeC[k]){do { } while(0); }; |
| 674 | dml_print("DML::%s: k=%0d ViewportSizeSupportPerSurface = %d\n", __func__, k,{do { } while(0); } |
| 675 | ViewportSizeSupportPerSurface[k]){do { } while(0); }; |
| 676 | #endif |
| 677 | |
| 678 | } |
| 679 | } // CalculateSwathAndDETConfiguration |
| 680 | |
| 681 | void dml32_CalculateSwathWidth( |
| 682 | bool_Bool ForceSingleDPP, |
| 683 | unsigned int NumberOfActiveSurfaces, |
| 684 | enum source_format_class SourcePixelFormat[], |
| 685 | enum dm_rotation_angle SourceRotation[], |
| 686 | bool_Bool ViewportStationary[], |
| 687 | unsigned int ViewportWidth[], |
| 688 | unsigned int ViewportHeight[], |
| 689 | unsigned int ViewportXStart[], |
| 690 | unsigned int ViewportYStart[], |
| 691 | unsigned int ViewportXStartC[], |
| 692 | unsigned int ViewportYStartC[], |
| 693 | unsigned int SurfaceWidthY[], |
| 694 | unsigned int SurfaceWidthC[], |
| 695 | unsigned int SurfaceHeightY[], |
| 696 | unsigned int SurfaceHeightC[], |
| 697 | enum odm_combine_mode ODMMode[], |
| 698 | unsigned int BytePerPixY[], |
| 699 | unsigned int BytePerPixC[], |
| 700 | unsigned int Read256BytesBlockHeightY[], |
| 701 | unsigned int Read256BytesBlockHeightC[], |
| 702 | unsigned int Read256BytesBlockWidthY[], |
| 703 | unsigned int Read256BytesBlockWidthC[], |
| 704 | unsigned int BlendingAndTiming[], |
| 705 | unsigned int HActive[], |
| 706 | double HRatio[], |
| 707 | unsigned int DPPPerSurface[], |
| 708 | |
| 709 | /* Output */ |
| 710 | double SwathWidthdoubleDPPY[], |
| 711 | double SwathWidthdoubleDPPC[], |
| 712 | double SwathWidthY[], // per-pipe |
| 713 | double SwathWidthC[], // per-pipe |
| 714 | unsigned int MaximumSwathHeightY[], |
| 715 | unsigned int MaximumSwathHeightC[], |
| 716 | unsigned int swath_width_luma_ub[], // per-pipe |
| 717 | unsigned int swath_width_chroma_ub[]) // per-pipe |
| 718 | { |
| 719 | unsigned int k, j; |
| 720 | enum odm_combine_mode MainSurfaceODMMode; |
| 721 | |
| 722 | unsigned int surface_width_ub_l; |
| 723 | unsigned int surface_height_ub_l; |
| 724 | unsigned int surface_width_ub_c = 0; |
| 725 | unsigned int surface_height_ub_c = 0; |
| 726 | |
| 727 | #ifdef __DML_VBA_DEBUG__ |
| 728 | dml_print("DML::%s: ForceSingleDPP = %d\n", __func__, ForceSingleDPP){do { } while(0); }; |
| 729 | dml_print("DML::%s: NumberOfActiveSurfaces = %d\n", __func__, NumberOfActiveSurfaces){do { } while(0); }; |
| 730 | #endif |
| 731 | |
| 732 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 733 | if (!IsVertical(SourceRotation[k])) |
| 734 | SwathWidthdoubleDPPY[k] = ViewportWidth[k]; |
| 735 | else |
| 736 | SwathWidthdoubleDPPY[k] = ViewportHeight[k]; |
| 737 | |
| 738 | #ifdef __DML_VBA_DEBUG__ |
| 739 | dml_print("DML::%s: k=%d ViewportWidth=%d\n", __func__, k, ViewportWidth[k]){do { } while(0); }; |
| 740 | dml_print("DML::%s: k=%d ViewportHeight=%d\n", __func__, k, ViewportHeight[k]){do { } while(0); }; |
| 741 | #endif |
| 742 | |
| 743 | MainSurfaceODMMode = ODMMode[k]; |
| 744 | for (j = 0; j < NumberOfActiveSurfaces; ++j) { |
| 745 | if (BlendingAndTiming[k] == j) |
| 746 | MainSurfaceODMMode = ODMMode[j]; |
| 747 | } |
| 748 | |
| 749 | if (ForceSingleDPP) { |
| 750 | SwathWidthY[k] = SwathWidthdoubleDPPY[k]; |
| 751 | } else { |
| 752 | if (MainSurfaceODMMode == dm_odm_combine_mode_4to1) { |
| 753 | SwathWidthY[k] = dml_min(SwathWidthdoubleDPPY[k], |
| 754 | dml_round(HActive[k] / 4.0 * HRatio[k])); |
| 755 | } else if (MainSurfaceODMMode == dm_odm_combine_mode_2to1) { |
| 756 | SwathWidthY[k] = dml_min(SwathWidthdoubleDPPY[k], |
| 757 | dml_round(HActive[k] / 2.0 * HRatio[k])); |
| 758 | } else if (DPPPerSurface[k] == 2) { |
| 759 | SwathWidthY[k] = SwathWidthdoubleDPPY[k] / 2; |
| 760 | } else { |
| 761 | SwathWidthY[k] = SwathWidthdoubleDPPY[k]; |
| 762 | } |
| 763 | } |
| 764 | |
| 765 | #ifdef __DML_VBA_DEBUG__ |
| 766 | dml_print("DML::%s: k=%d HActive=%d\n", __func__, k, HActive[k]){do { } while(0); }; |
| 767 | dml_print("DML::%s: k=%d HRatio=%f\n", __func__, k, HRatio[k]){do { } while(0); }; |
| 768 | dml_print("DML::%s: k=%d MainSurfaceODMMode=%d\n", __func__, k, MainSurfaceODMMode){do { } while(0); }; |
| 769 | dml_print("DML::%s: k=%d SwathWidthdoubleDPPY=%d\n", __func__, k, SwathWidthdoubleDPPY[k]){do { } while(0); }; |
| 770 | dml_print("DML::%s: k=%d SwathWidthY=%d\n", __func__, k, SwathWidthY[k]){do { } while(0); }; |
| 771 | #endif |
| 772 | |
| 773 | if (SourcePixelFormat[k] == dm_420_8 || SourcePixelFormat[k] == dm_420_10 || |
| 774 | SourcePixelFormat[k] == dm_420_12) { |
| 775 | SwathWidthC[k] = SwathWidthY[k] / 2; |
| 776 | SwathWidthdoubleDPPC[k] = SwathWidthdoubleDPPY[k] / 2; |
| 777 | } else { |
| 778 | SwathWidthC[k] = SwathWidthY[k]; |
| 779 | SwathWidthdoubleDPPC[k] = SwathWidthdoubleDPPY[k]; |
| 780 | } |
| 781 | |
| 782 | if (ForceSingleDPP == true1) { |
| 783 | SwathWidthY[k] = SwathWidthdoubleDPPY[k]; |
| 784 | SwathWidthC[k] = SwathWidthdoubleDPPC[k]; |
| 785 | } |
| 786 | |
| 787 | surface_width_ub_l = dml_ceil(SurfaceWidthY[k], Read256BytesBlockWidthY[k]); |
| 788 | surface_height_ub_l = dml_ceil(SurfaceHeightY[k], Read256BytesBlockHeightY[k]); |
| 789 | |
| 790 | if (!IsVertical(SourceRotation[k])) { |
| 791 | MaximumSwathHeightY[k] = Read256BytesBlockHeightY[k]; |
| 792 | MaximumSwathHeightC[k] = Read256BytesBlockHeightC[k]; |
| 793 | if (ViewportStationary[k] && DPPPerSurface[k] == 1) { |
| 794 | swath_width_luma_ub[k] = dml_min(surface_width_ub_l, |
| 795 | dml_floor(ViewportXStart[k] + |
| 796 | SwathWidthY[k] + |
| 797 | Read256BytesBlockWidthY[k] - 1, |
| 798 | Read256BytesBlockWidthY[k]) - |
| 799 | dml_floor(ViewportXStart[k], |
| 800 | Read256BytesBlockWidthY[k])); |
| 801 | } else { |
| 802 | swath_width_luma_ub[k] = dml_min(surface_width_ub_l, |
| 803 | dml_ceil(SwathWidthY[k] - 1, |
| 804 | Read256BytesBlockWidthY[k]) + |
| 805 | Read256BytesBlockWidthY[k]); |
| 806 | } |
| 807 | if (BytePerPixC[k] > 0) { |
| 808 | surface_width_ub_c = dml_ceil(SurfaceWidthC[k], Read256BytesBlockWidthC[k]); |
| 809 | if (ViewportStationary[k] && DPPPerSurface[k] == 1) { |
| 810 | swath_width_chroma_ub[k] = dml_min(surface_width_ub_c, |
| 811 | dml_floor(ViewportXStartC[k] + SwathWidthC[k] + |
| 812 | Read256BytesBlockWidthC[k] - 1, |
| 813 | Read256BytesBlockWidthC[k]) - |
| 814 | dml_floor(ViewportXStartC[k], |
| 815 | Read256BytesBlockWidthC[k])); |
| 816 | } else { |
| 817 | swath_width_chroma_ub[k] = dml_min(surface_width_ub_c, |
| 818 | dml_ceil(SwathWidthC[k] - 1, |
| 819 | Read256BytesBlockWidthC[k]) + |
| 820 | Read256BytesBlockWidthC[k]); |
| 821 | } |
| 822 | } else { |
| 823 | swath_width_chroma_ub[k] = 0; |
| 824 | } |
| 825 | } else { |
| 826 | MaximumSwathHeightY[k] = Read256BytesBlockWidthY[k]; |
| 827 | MaximumSwathHeightC[k] = Read256BytesBlockWidthC[k]; |
| 828 | |
| 829 | if (ViewportStationary[k] && DPPPerSurface[k] == 1) { |
| 830 | swath_width_luma_ub[k] = dml_min(surface_height_ub_l, dml_floor(ViewportYStart[k] + |
| 831 | SwathWidthY[k] + Read256BytesBlockHeightY[k] - 1, |
| 832 | Read256BytesBlockHeightY[k]) - |
| 833 | dml_floor(ViewportYStart[k], Read256BytesBlockHeightY[k])); |
| 834 | } else { |
| 835 | swath_width_luma_ub[k] = dml_min(surface_height_ub_l, dml_ceil(SwathWidthY[k] - 1, |
| 836 | Read256BytesBlockHeightY[k]) + Read256BytesBlockHeightY[k]); |
| 837 | } |
| 838 | if (BytePerPixC[k] > 0) { |
| 839 | surface_height_ub_c = dml_ceil(SurfaceHeightC[k], Read256BytesBlockHeightC[k]); |
| 840 | if (ViewportStationary[k] && DPPPerSurface[k] == 1) { |
| 841 | swath_width_chroma_ub[k] = dml_min(surface_height_ub_c, |
| 842 | dml_floor(ViewportYStartC[k] + SwathWidthC[k] + |
| 843 | Read256BytesBlockHeightC[k] - 1, |
| 844 | Read256BytesBlockHeightC[k]) - |
| 845 | dml_floor(ViewportYStartC[k], |
| 846 | Read256BytesBlockHeightC[k])); |
| 847 | } else { |
| 848 | swath_width_chroma_ub[k] = dml_min(surface_height_ub_c, |
| 849 | dml_ceil(SwathWidthC[k] - 1, Read256BytesBlockHeightC[k]) + |
| 850 | Read256BytesBlockHeightC[k]); |
| 851 | } |
| 852 | } else { |
| 853 | swath_width_chroma_ub[k] = 0; |
| 854 | } |
| 855 | } |
| 856 | |
| 857 | #ifdef __DML_VBA_DEBUG__ |
| 858 | dml_print("DML::%s: k=%d surface_width_ub_l=%0d\n", __func__, k, surface_width_ub_l){do { } while(0); }; |
| 859 | dml_print("DML::%s: k=%d surface_height_ub_l=%0d\n", __func__, k, surface_height_ub_l){do { } while(0); }; |
| 860 | dml_print("DML::%s: k=%d surface_width_ub_c=%0d\n", __func__, k, surface_width_ub_c){do { } while(0); }; |
| 861 | dml_print("DML::%s: k=%d surface_height_ub_c=%0d\n", __func__, k, surface_height_ub_c){do { } while(0); }; |
| 862 | dml_print("DML::%s: k=%d Read256BytesBlockWidthY=%0d\n", __func__, k, Read256BytesBlockWidthY[k]){do { } while(0); }; |
| 863 | dml_print("DML::%s: k=%d Read256BytesBlockHeightY=%0d\n", __func__, k, Read256BytesBlockHeightY[k]){do { } while(0); }; |
| 864 | dml_print("DML::%s: k=%d Read256BytesBlockWidthC=%0d\n", __func__, k, Read256BytesBlockWidthC[k]){do { } while(0); }; |
| 865 | dml_print("DML::%s: k=%d Read256BytesBlockHeightC=%0d\n", __func__, k, Read256BytesBlockHeightC[k]){do { } while(0); }; |
| 866 | dml_print("DML::%s: k=%d ViewportStationary=%0d\n", __func__, k, ViewportStationary[k]){do { } while(0); }; |
| 867 | dml_print("DML::%s: k=%d DPPPerSurface=%0d\n", __func__, k, DPPPerSurface[k]){do { } while(0); }; |
| 868 | dml_print("DML::%s: k=%d swath_width_luma_ub=%0d\n", __func__, k, swath_width_luma_ub[k]){do { } while(0); }; |
| 869 | dml_print("DML::%s: k=%d swath_width_chroma_ub=%0d\n", __func__, k, swath_width_chroma_ub[k]){do { } while(0); }; |
| 870 | dml_print("DML::%s: k=%d MaximumSwathHeightY=%0d\n", __func__, k, MaximumSwathHeightY[k]){do { } while(0); }; |
| 871 | dml_print("DML::%s: k=%d MaximumSwathHeightC=%0d\n", __func__, k, MaximumSwathHeightC[k]){do { } while(0); }; |
| 872 | #endif |
| 873 | |
| 874 | } |
| 875 | } // CalculateSwathWidth |
| 876 | |
| 877 | bool_Bool dml32_UnboundedRequest(enum unbounded_requesting_policy UseUnboundedRequestingFinal, |
| 878 | unsigned int TotalNumberOfActiveDPP, |
| 879 | bool_Bool NoChroma, |
| 880 | enum output_encoder_class Output, |
| 881 | enum dm_swizzle_mode SurfaceTiling, |
| 882 | bool_Bool CompBufReservedSpaceNeedAdjustment, |
| 883 | bool_Bool DisableUnboundRequestIfCompBufReservedSpaceNeedAdjustment) |
| 884 | { |
| 885 | bool_Bool ret_val = false0; |
| 886 | |
| 887 | ret_val = (UseUnboundedRequestingFinal != dm_unbounded_requesting_disable && |
| 888 | TotalNumberOfActiveDPP == 1 && NoChroma); |
| 889 | if (UseUnboundedRequestingFinal == dm_unbounded_requesting_edp_only && Output != dm_edp) |
| 890 | ret_val = false0; |
| 891 | |
| 892 | if (SurfaceTiling == dm_sw_linear) |
| 893 | ret_val = false0; |
| 894 | |
| 895 | if (CompBufReservedSpaceNeedAdjustment == 1 && DisableUnboundRequestIfCompBufReservedSpaceNeedAdjustment) |
| 896 | ret_val = false0; |
| 897 | |
| 898 | #ifdef __DML_VBA_DEBUG__ |
| 899 | dml_print("DML::%s: CompBufReservedSpaceNeedAdjustment = %d\n", __func__, CompBufReservedSpaceNeedAdjustment){do { } while(0); }; |
| 900 | dml_print("DML::%s: DisableUnboundRequestIfCompBufReservedSpaceNeedAdjustment = %d\n", __func__, DisableUnboundRequestIfCompBufReservedSpaceNeedAdjustment){do { } while(0); }; |
| 901 | dml_print("DML::%s: ret_val = %d\n", __func__, ret_val){do { } while(0); }; |
| 902 | #endif |
| 903 | |
| 904 | return (ret_val); |
| 905 | } |
| 906 | |
| 907 | void dml32_CalculateDETBufferSize( |
| 908 | unsigned int DETSizeOverride[], |
| 909 | enum dm_use_mall_for_pstate_change_mode UseMALLForPStateChange[], |
| 910 | bool_Bool ForceSingleDPP, |
| 911 | unsigned int NumberOfActiveSurfaces, |
| 912 | bool_Bool UnboundedRequestEnabled, |
| 913 | unsigned int nomDETInKByte, |
| 914 | unsigned int MaxTotalDETInKByte, |
| 915 | unsigned int ConfigReturnBufferSizeInKByte, |
| 916 | unsigned int MinCompressedBufferSizeInKByte, |
| 917 | unsigned int CompressedBufferSegmentSizeInkByteFinal, |
| 918 | enum source_format_class SourcePixelFormat[], |
| 919 | double ReadBandwidthLuma[], |
| 920 | double ReadBandwidthChroma[], |
| 921 | unsigned int RoundedUpMaxSwathSizeBytesY[], |
| 922 | unsigned int RoundedUpMaxSwathSizeBytesC[], |
| 923 | unsigned int DPPPerSurface[], |
| 924 | /* Output */ |
| 925 | unsigned int DETBufferSizeInKByte[], |
| 926 | unsigned int *CompressedBufferSizeInkByte) |
| 927 | { |
| 928 | unsigned int DETBufferSizePoolInKByte; |
| 929 | unsigned int NextDETBufferPieceInKByte; |
| 930 | bool_Bool DETPieceAssignedToThisSurfaceAlready[DC__NUM_DPP__MAX8]; |
| 931 | bool_Bool NextPotentialSurfaceToAssignDETPieceFound; |
| 932 | unsigned int NextSurfaceToAssignDETPiece; |
| 933 | double TotalBandwidth; |
| 934 | double BandwidthOfSurfacesNotAssignedDETPiece; |
| 935 | unsigned int max_minDET; |
| 936 | unsigned int minDET; |
| 937 | unsigned int minDET_pipe; |
| 938 | unsigned int j, k; |
| 939 | |
| 940 | #ifdef __DML_VBA_DEBUG__ |
| 941 | dml_print("DML::%s: ForceSingleDPP = %d\n", __func__, ForceSingleDPP){do { } while(0); }; |
| 942 | dml_print("DML::%s: nomDETInKByte = %d\n", __func__, nomDETInKByte){do { } while(0); }; |
| 943 | dml_print("DML::%s: NumberOfActiveSurfaces = %d\n", __func__, NumberOfActiveSurfaces){do { } while(0); }; |
| 944 | dml_print("DML::%s: UnboundedRequestEnabled = %d\n", __func__, UnboundedRequestEnabled){do { } while(0); }; |
| 945 | dml_print("DML::%s: MaxTotalDETInKByte = %d\n", __func__, MaxTotalDETInKByte){do { } while(0); }; |
| 946 | dml_print("DML::%s: ConfigReturnBufferSizeInKByte = %d\n", __func__, ConfigReturnBufferSizeInKByte){do { } while(0); }; |
| 947 | dml_print("DML::%s: MinCompressedBufferSizeInKByte = %d\n", __func__, MinCompressedBufferSizeInKByte){do { } while(0); }; |
| 948 | dml_print("DML::%s: CompressedBufferSegmentSizeInkByteFinal = %d\n", __func__,{do { } while(0); } |
| 949 | CompressedBufferSegmentSizeInkByteFinal){do { } while(0); }; |
| 950 | #endif |
| 951 | |
| 952 | // Note: Will use default det size if that fits 2 swaths |
| 953 | if (UnboundedRequestEnabled) { |
| 954 | if (DETSizeOverride[0] > 0) { |
| 955 | DETBufferSizeInKByte[0] = DETSizeOverride[0]; |
| 956 | } else { |
| 957 | DETBufferSizeInKByte[0] = dml_max(nomDETInKByte, dml_ceil(2.0 * |
| 958 | ((double) RoundedUpMaxSwathSizeBytesY[0] + |
| 959 | (double) RoundedUpMaxSwathSizeBytesC[0]) / 1024.0, 64.0)); |
| 960 | } |
| 961 | *CompressedBufferSizeInkByte = ConfigReturnBufferSizeInKByte - DETBufferSizeInKByte[0]; |
| 962 | } else { |
| 963 | DETBufferSizePoolInKByte = MaxTotalDETInKByte; |
| 964 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 965 | DETBufferSizeInKByte[k] = nomDETInKByte; |
| 966 | if (SourcePixelFormat[k] == dm_420_8 || SourcePixelFormat[k] == dm_420_10 || |
| 967 | SourcePixelFormat[k] == dm_420_12) { |
| 968 | max_minDET = nomDETInKByte - 64; |
| 969 | } else { |
| 970 | max_minDET = nomDETInKByte; |
| 971 | } |
| 972 | minDET = 128; |
| 973 | minDET_pipe = 0; |
| 974 | |
| 975 | // add DET resource until can hold 2 full swaths |
| 976 | while (minDET <= max_minDET && minDET_pipe == 0) { |
| 977 | if (2.0 * ((double) RoundedUpMaxSwathSizeBytesY[k] + |
| 978 | (double) RoundedUpMaxSwathSizeBytesC[k]) / 1024.0 <= minDET) |
| 979 | minDET_pipe = minDET; |
| 980 | minDET = minDET + 64; |
| 981 | } |
| 982 | |
| 983 | #ifdef __DML_VBA_DEBUG__ |
| 984 | dml_print("DML::%s: k=%0d minDET = %d\n", __func__, k, minDET){do { } while(0); }; |
| 985 | dml_print("DML::%s: k=%0d max_minDET = %d\n", __func__, k, max_minDET){do { } while(0); }; |
| 986 | dml_print("DML::%s: k=%0d minDET_pipe = %d\n", __func__, k, minDET_pipe){do { } while(0); }; |
| 987 | dml_print("DML::%s: k=%0d RoundedUpMaxSwathSizeBytesY = %d\n", __func__, k,{do { } while(0); } |
| 988 | RoundedUpMaxSwathSizeBytesY[k]){do { } while(0); }; |
| 989 | dml_print("DML::%s: k=%0d RoundedUpMaxSwathSizeBytesC = %d\n", __func__, k,{do { } while(0); } |
| 990 | RoundedUpMaxSwathSizeBytesC[k]){do { } while(0); }; |
| 991 | #endif |
| 992 | |
| 993 | if (minDET_pipe == 0) { |
| 994 | minDET_pipe = dml_max(128, dml_ceil(((double)RoundedUpMaxSwathSizeBytesY[k] + |
| 995 | (double)RoundedUpMaxSwathSizeBytesC[k]) / 1024.0, 64)); |
| 996 | #ifdef __DML_VBA_DEBUG__ |
| 997 | dml_print("DML::%s: k=%0d minDET_pipe = %d (assume each plane take half DET)\n",{do { } while(0); } |
| 998 | __func__, k, minDET_pipe){do { } while(0); }; |
| 999 | #endif |
| 1000 | } |
| 1001 | |
| 1002 | if (UseMALLForPStateChange[k] == dm_use_mall_pstate_change_phantom_pipe) { |
| 1003 | DETBufferSizeInKByte[k] = 0; |
| 1004 | } else if (DETSizeOverride[k] > 0) { |
| 1005 | DETBufferSizeInKByte[k] = DETSizeOverride[k]; |
| 1006 | DETBufferSizePoolInKByte = DETBufferSizePoolInKByte - |
| 1007 | (ForceSingleDPP ? 1 : DPPPerSurface[k]) * DETSizeOverride[k]; |
| 1008 | } else if ((ForceSingleDPP ? 1 : DPPPerSurface[k]) * minDET_pipe <= DETBufferSizePoolInKByte) { |
| 1009 | DETBufferSizeInKByte[k] = minDET_pipe; |
| 1010 | DETBufferSizePoolInKByte = DETBufferSizePoolInKByte - |
| 1011 | (ForceSingleDPP ? 1 : DPPPerSurface[k]) * minDET_pipe; |
| 1012 | } |
| 1013 | |
| 1014 | #ifdef __DML_VBA_DEBUG__ |
| 1015 | dml_print("DML::%s: k=%d DPPPerSurface = %d\n", __func__, k, DPPPerSurface[k]){do { } while(0); }; |
| 1016 | dml_print("DML::%s: k=%d DETSizeOverride = %d\n", __func__, k, DETSizeOverride[k]){do { } while(0); }; |
| 1017 | dml_print("DML::%s: k=%d DETBufferSizeInKByte = %d\n", __func__, k, DETBufferSizeInKByte[k]){do { } while(0); }; |
| 1018 | dml_print("DML::%s: DETBufferSizePoolInKByte = %d\n", __func__, DETBufferSizePoolInKByte){do { } while(0); }; |
| 1019 | #endif |
| 1020 | } |
| 1021 | |
| 1022 | TotalBandwidth = 0; |
| 1023 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 1024 | if (UseMALLForPStateChange[k] != dm_use_mall_pstate_change_phantom_pipe) |
| 1025 | TotalBandwidth = TotalBandwidth + ReadBandwidthLuma[k] + ReadBandwidthChroma[k]; |
| 1026 | } |
| 1027 | #ifdef __DML_VBA_DEBUG__ |
| 1028 | dml_print("DML::%s: --- Before bandwidth adjustment ---\n", __func__){do { } while(0); }; |
| 1029 | for (uint k = 0; k < NumberOfActiveSurfaces; ++k) |
| 1030 | dml_print("DML::%s: k=%d DETBufferSizeInKByte = %d\n", __func__, k, DETBufferSizeInKByte[k]){do { } while(0); }; |
| 1031 | dml_print("DML::%s: --- DET allocation with bandwidth ---\n", __func__){do { } while(0); }; |
| 1032 | dml_print("DML::%s: TotalBandwidth = %f\n", __func__, TotalBandwidth){do { } while(0); }; |
| 1033 | #endif |
| 1034 | BandwidthOfSurfacesNotAssignedDETPiece = TotalBandwidth; |
| 1035 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 1036 | |
| 1037 | if (UseMALLForPStateChange[k] == dm_use_mall_pstate_change_phantom_pipe) { |
| 1038 | DETPieceAssignedToThisSurfaceAlready[k] = true1; |
| 1039 | } else if (DETSizeOverride[k] > 0 || (((double) (ForceSingleDPP ? 1 : DPPPerSurface[k]) * |
| 1040 | (double) DETBufferSizeInKByte[k] / (double) MaxTotalDETInKByte) >= |
| 1041 | ((ReadBandwidthLuma[k] + ReadBandwidthChroma[k]) / TotalBandwidth))) { |
| 1042 | DETPieceAssignedToThisSurfaceAlready[k] = true1; |
| 1043 | BandwidthOfSurfacesNotAssignedDETPiece = BandwidthOfSurfacesNotAssignedDETPiece - |
| 1044 | ReadBandwidthLuma[k] - ReadBandwidthChroma[k]; |
| 1045 | } else { |
| 1046 | DETPieceAssignedToThisSurfaceAlready[k] = false0; |
| 1047 | } |
| 1048 | #ifdef __DML_VBA_DEBUG__ |
| 1049 | dml_print("DML::%s: k=%d DETPieceAssignedToThisSurfaceAlready = %d\n", __func__, k,{do { } while(0); } |
| 1050 | DETPieceAssignedToThisSurfaceAlready[k]){do { } while(0); }; |
| 1051 | dml_print("DML::%s: k=%d BandwidthOfSurfacesNotAssignedDETPiece = %f\n", __func__, k,{do { } while(0); } |
| 1052 | BandwidthOfSurfacesNotAssignedDETPiece){do { } while(0); }; |
| 1053 | #endif |
| 1054 | } |
| 1055 | |
| 1056 | for (j = 0; j < NumberOfActiveSurfaces; ++j) { |
| 1057 | NextPotentialSurfaceToAssignDETPieceFound = false0; |
| 1058 | NextSurfaceToAssignDETPiece = 0; |
| 1059 | |
| 1060 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 1061 | #ifdef __DML_VBA_DEBUG__ |
| 1062 | dml_print("DML::%s: j=%d k=%d, ReadBandwidthLuma[k] = %f\n", __func__, j, k,{do { } while(0); } |
| 1063 | ReadBandwidthLuma[k]){do { } while(0); }; |
| 1064 | dml_print("DML::%s: j=%d k=%d, ReadBandwidthChroma[k] = %f\n", __func__, j, k,{do { } while(0); } |
| 1065 | ReadBandwidthChroma[k]){do { } while(0); }; |
| 1066 | dml_print("DML::%s: j=%d k=%d, ReadBandwidthLuma[Next] = %f\n", __func__, j, k,{do { } while(0); } |
| 1067 | ReadBandwidthLuma[NextSurfaceToAssignDETPiece]){do { } while(0); }; |
| 1068 | dml_print("DML::%s: j=%d k=%d, ReadBandwidthChroma[Next] = %f\n", __func__, j, k,{do { } while(0); } |
| 1069 | ReadBandwidthChroma[NextSurfaceToAssignDETPiece]){do { } while(0); }; |
| 1070 | dml_print("DML::%s: j=%d k=%d, NextSurfaceToAssignDETPiece = %d\n", __func__, j, k,{do { } while(0); } |
| 1071 | NextSurfaceToAssignDETPiece){do { } while(0); }; |
| 1072 | #endif |
| 1073 | if (!DETPieceAssignedToThisSurfaceAlready[k] && |
| 1074 | (!NextPotentialSurfaceToAssignDETPieceFound || |
| 1075 | ReadBandwidthLuma[k] + ReadBandwidthChroma[k] < |
| 1076 | ReadBandwidthLuma[NextSurfaceToAssignDETPiece] + |
| 1077 | ReadBandwidthChroma[NextSurfaceToAssignDETPiece])) { |
| 1078 | NextSurfaceToAssignDETPiece = k; |
| 1079 | NextPotentialSurfaceToAssignDETPieceFound = true1; |
| 1080 | } |
| 1081 | #ifdef __DML_VBA_DEBUG__ |
| 1082 | dml_print("DML::%s: j=%d k=%d, DETPieceAssignedToThisSurfaceAlready = %d\n",{do { } while(0); } |
| 1083 | __func__, j, k, DETPieceAssignedToThisSurfaceAlready[k]){do { } while(0); }; |
| 1084 | dml_print("DML::%s: j=%d k=%d, NextPotentialSurfaceToAssignDETPieceFound = %d\n",{do { } while(0); } |
| 1085 | __func__, j, k, NextPotentialSurfaceToAssignDETPieceFound){do { } while(0); }; |
| 1086 | #endif |
| 1087 | } |
| 1088 | |
| 1089 | if (NextPotentialSurfaceToAssignDETPieceFound) { |
| 1090 | // Note: To show the banker's rounding behavior in VBA and also the fact |
| 1091 | // that the DET buffer size varies due to precision issue |
| 1092 | // |
| 1093 | //double tmp1 = ((double) DETBufferSizePoolInKByte * |
| 1094 | // (ReadBandwidthLuma[NextSurfaceToAssignDETPiece] + |
| 1095 | // ReadBandwidthChroma[NextSurfaceToAssignDETPiece]) / |
| 1096 | // BandwidthOfSurfacesNotAssignedDETPiece / |
| 1097 | // ((ForceSingleDPP ? 1 : DPPPerSurface[NextSurfaceToAssignDETPiece]) * 64.0)); |
| 1098 | //double tmp2 = dml_round((double) DETBufferSizePoolInKByte * |
| 1099 | // (ReadBandwidthLuma[NextSurfaceToAssignDETPiece] + |
| 1100 | // ReadBandwidthChroma[NextSurfaceToAssignDETPiece]) / |
| 1101 | //BandwidthOfSurfacesNotAssignedDETPiece / |
| 1102 | // ((ForceSingleDPP ? 1 : DPPPerSurface[NextSurfaceToAssignDETPiece]) * 64.0)); |
| 1103 | // |
| 1104 | //dml_print("DML::%s: j=%d, tmp1 = %f\n", __func__, j, tmp1); |
| 1105 | //dml_print("DML::%s: j=%d, tmp2 = %f\n", __func__, j, tmp2); |
| 1106 | |
| 1107 | NextDETBufferPieceInKByte = dml_min( |
| 1108 | dml_round((double) DETBufferSizePoolInKByte * |
| 1109 | (ReadBandwidthLuma[NextSurfaceToAssignDETPiece] + |
| 1110 | ReadBandwidthChroma[NextSurfaceToAssignDETPiece]) / |
| 1111 | BandwidthOfSurfacesNotAssignedDETPiece / |
| 1112 | ((ForceSingleDPP ? 1 : |
| 1113 | DPPPerSurface[NextSurfaceToAssignDETPiece]) * 64.0)) * |
| 1114 | (ForceSingleDPP ? 1 : |
| 1115 | DPPPerSurface[NextSurfaceToAssignDETPiece]) * 64.0, |
| 1116 | dml_floor((double) DETBufferSizePoolInKByte, |
| 1117 | (ForceSingleDPP ? 1 : |
| 1118 | DPPPerSurface[NextSurfaceToAssignDETPiece]) * 64.0)); |
| 1119 | |
| 1120 | // Above calculation can assign the entire DET buffer allocation to a single pipe. |
| 1121 | // We should limit the per-pipe DET size to the nominal / max per pipe. |
| 1122 | if (NextDETBufferPieceInKByte > nomDETInKByte * (ForceSingleDPP ? 1 : DPPPerSurface[k])) { |
| 1123 | if (DETBufferSizeInKByte[NextSurfaceToAssignDETPiece] < |
| 1124 | nomDETInKByte * (ForceSingleDPP ? 1 : DPPPerSurface[k])) { |
| 1125 | NextDETBufferPieceInKByte = nomDETInKByte * (ForceSingleDPP ? 1 : DPPPerSurface[k]) - |
| 1126 | DETBufferSizeInKByte[NextSurfaceToAssignDETPiece]; |
| 1127 | } else { |
| 1128 | // Case where DETBufferSizeInKByte[NextSurfaceToAssignDETPiece] |
| 1129 | // already has the max per-pipe value |
| 1130 | NextDETBufferPieceInKByte = 0; |
| 1131 | } |
| 1132 | } |
| 1133 | |
| 1134 | #ifdef __DML_VBA_DEBUG__ |
| 1135 | dml_print("DML::%s: j=%0d, DETBufferSizePoolInKByte = %d\n", __func__, j,{do { } while(0); } |
| 1136 | DETBufferSizePoolInKByte){do { } while(0); }; |
| 1137 | dml_print("DML::%s: j=%0d, NextSurfaceToAssignDETPiece = %d\n", __func__, j,{do { } while(0); } |
| 1138 | NextSurfaceToAssignDETPiece){do { } while(0); }; |
| 1139 | dml_print("DML::%s: j=%0d, ReadBandwidthLuma[%0d] = %f\n", __func__, j,{do { } while(0); } |
| 1140 | NextSurfaceToAssignDETPiece, ReadBandwidthLuma[NextSurfaceToAssignDETPiece]){do { } while(0); }; |
| 1141 | dml_print("DML::%s: j=%0d, ReadBandwidthChroma[%0d] = %f\n", __func__, j,{do { } while(0); } |
| 1142 | NextSurfaceToAssignDETPiece, ReadBandwidthChroma[NextSurfaceToAssignDETPiece]){do { } while(0); }; |
| 1143 | dml_print("DML::%s: j=%0d, BandwidthOfSurfacesNotAssignedDETPiece = %f\n",{do { } while(0); } |
| 1144 | __func__, j, BandwidthOfSurfacesNotAssignedDETPiece){do { } while(0); }; |
| 1145 | dml_print("DML::%s: j=%0d, NextDETBufferPieceInKByte = %d\n", __func__, j,{do { } while(0); } |
| 1146 | NextDETBufferPieceInKByte){do { } while(0); }; |
| 1147 | dml_print("DML::%s: j=%0d, DETBufferSizeInKByte[%0d] increases from %0d ",{do { } while(0); } |
| 1148 | __func__, j, NextSurfaceToAssignDETPiece,{do { } while(0); } |
| 1149 | DETBufferSizeInKByte[NextSurfaceToAssignDETPiece]){do { } while(0); }; |
| 1150 | #endif |
| 1151 | |
| 1152 | DETBufferSizeInKByte[NextSurfaceToAssignDETPiece] = |
| 1153 | DETBufferSizeInKByte[NextSurfaceToAssignDETPiece] |
| 1154 | + NextDETBufferPieceInKByte |
| 1155 | / (ForceSingleDPP ? 1 : DPPPerSurface[NextSurfaceToAssignDETPiece]); |
| 1156 | #ifdef __DML_VBA_DEBUG__ |
| 1157 | dml_print("to %0d\n", DETBufferSizeInKByte[NextSurfaceToAssignDETPiece]){do { } while(0); }; |
| 1158 | #endif |
| 1159 | |
| 1160 | DETBufferSizePoolInKByte = DETBufferSizePoolInKByte - NextDETBufferPieceInKByte; |
| 1161 | DETPieceAssignedToThisSurfaceAlready[NextSurfaceToAssignDETPiece] = true1; |
| 1162 | BandwidthOfSurfacesNotAssignedDETPiece = BandwidthOfSurfacesNotAssignedDETPiece - |
| 1163 | (ReadBandwidthLuma[NextSurfaceToAssignDETPiece] + |
| 1164 | ReadBandwidthChroma[NextSurfaceToAssignDETPiece]); |
| 1165 | } |
| 1166 | } |
| 1167 | *CompressedBufferSizeInkByte = MinCompressedBufferSizeInKByte; |
| 1168 | } |
| 1169 | *CompressedBufferSizeInkByte = *CompressedBufferSizeInkByte * CompressedBufferSegmentSizeInkByteFinal / 64; |
| 1170 | |
| 1171 | #ifdef __DML_VBA_DEBUG__ |
| 1172 | dml_print("DML::%s: --- After bandwidth adjustment ---\n", __func__){do { } while(0); }; |
| 1173 | dml_print("DML::%s: CompressedBufferSizeInkByte = %d\n", __func__, *CompressedBufferSizeInkByte){do { } while(0); }; |
| 1174 | for (uint k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 1175 | dml_print("DML::%s: k=%d DETBufferSizeInKByte = %d (TotalReadBandWidth=%f)\n",{do { } while(0); } |
| 1176 | __func__, k, DETBufferSizeInKByte[k], ReadBandwidthLuma[k] + ReadBandwidthChroma[k]){do { } while(0); }; |
| 1177 | } |
| 1178 | #endif |
| 1179 | } // CalculateDETBufferSize |
| 1180 | |
| 1181 | void dml32_CalculateODMMode( |
| 1182 | unsigned int MaximumPixelsPerLinePerDSCUnit, |
| 1183 | unsigned int HActive, |
| 1184 | enum output_format_class OutFormat, |
| 1185 | enum output_encoder_class Output, |
| 1186 | enum odm_combine_policy ODMUse, |
| 1187 | double StateDispclk, |
| 1188 | double MaxDispclk, |
| 1189 | bool_Bool DSCEnable, |
| 1190 | unsigned int TotalNumberOfActiveDPP, |
| 1191 | unsigned int MaxNumDPP, |
| 1192 | double PixelClock, |
| 1193 | double DISPCLKDPPCLKDSCCLKDownSpreading, |
| 1194 | double DISPCLKRampingMargin, |
| 1195 | double DISPCLKDPPCLKVCOSpeed, |
| 1196 | unsigned int NumberOfDSCSlices, |
| 1197 | |
| 1198 | /* Output */ |
| 1199 | bool_Bool *TotalAvailablePipesSupport, |
| 1200 | unsigned int *NumberOfDPP, |
| 1201 | enum odm_combine_mode *ODMMode, |
| 1202 | double *RequiredDISPCLKPerSurface) |
| 1203 | { |
| 1204 | |
| 1205 | double SurfaceRequiredDISPCLKWithoutODMCombine; |
| 1206 | double SurfaceRequiredDISPCLKWithODMCombineTwoToOne; |
| 1207 | double SurfaceRequiredDISPCLKWithODMCombineFourToOne; |
| 1208 | |
| 1209 | SurfaceRequiredDISPCLKWithoutODMCombine = dml32_CalculateRequiredDispclk(dm_odm_combine_mode_disabled, |
| 1210 | PixelClock, DISPCLKDPPCLKDSCCLKDownSpreading, DISPCLKRampingMargin, DISPCLKDPPCLKVCOSpeed, |
| 1211 | MaxDispclk); |
| 1212 | SurfaceRequiredDISPCLKWithODMCombineTwoToOne = dml32_CalculateRequiredDispclk(dm_odm_combine_mode_2to1, |
| 1213 | PixelClock, DISPCLKDPPCLKDSCCLKDownSpreading, DISPCLKRampingMargin, DISPCLKDPPCLKVCOSpeed, |
| 1214 | MaxDispclk); |
| 1215 | SurfaceRequiredDISPCLKWithODMCombineFourToOne = dml32_CalculateRequiredDispclk(dm_odm_combine_mode_4to1, |
| 1216 | PixelClock, DISPCLKDPPCLKDSCCLKDownSpreading, DISPCLKRampingMargin, DISPCLKDPPCLKVCOSpeed, |
| 1217 | MaxDispclk); |
| 1218 | *TotalAvailablePipesSupport = true1; |
| 1219 | *ODMMode = dm_odm_combine_mode_disabled; // initialize as disable |
| 1220 | |
| 1221 | if (ODMUse == dm_odm_combine_policy_none) |
| 1222 | *ODMMode = dm_odm_combine_mode_disabled; |
| 1223 | |
| 1224 | *RequiredDISPCLKPerSurface = SurfaceRequiredDISPCLKWithoutODMCombine; |
| 1225 | *NumberOfDPP = 0; |
| 1226 | |
| 1227 | // FIXME check ODMUse == "" condition does it mean bypass or Gabriel means something like don't care?? |
| 1228 | // (ODMUse == "" || ODMUse == "CombineAsNeeded") |
| 1229 | |
| 1230 | if (!(Output == dm_hdmi || Output == dm_dp || Output == dm_edp) && (ODMUse == dm_odm_combine_policy_4to1 || |
| 1231 | ((SurfaceRequiredDISPCLKWithODMCombineTwoToOne > StateDispclk || |
| 1232 | (DSCEnable && (HActive > 2 * MaximumPixelsPerLinePerDSCUnit)) |
| 1233 | || NumberOfDSCSlices > 8)))) { |
| 1234 | if (TotalNumberOfActiveDPP + 4 <= MaxNumDPP) { |
| 1235 | *ODMMode = dm_odm_combine_mode_4to1; |
| 1236 | *RequiredDISPCLKPerSurface = SurfaceRequiredDISPCLKWithODMCombineFourToOne; |
| 1237 | *NumberOfDPP = 4; |
| 1238 | } else { |
| 1239 | *TotalAvailablePipesSupport = false0; |
| 1240 | } |
| 1241 | } else if (Output != dm_hdmi && (ODMUse == dm_odm_combine_policy_2to1 || |
| 1242 | (((SurfaceRequiredDISPCLKWithoutODMCombine > StateDispclk && |
| 1243 | SurfaceRequiredDISPCLKWithODMCombineTwoToOne <= StateDispclk) || |
| 1244 | (DSCEnable && (HActive > MaximumPixelsPerLinePerDSCUnit)) |
| 1245 | || (NumberOfDSCSlices <= 8 && NumberOfDSCSlices > 4))))) { |
| 1246 | if (TotalNumberOfActiveDPP + 2 <= MaxNumDPP) { |
| 1247 | *ODMMode = dm_odm_combine_mode_2to1; |
| 1248 | *RequiredDISPCLKPerSurface = SurfaceRequiredDISPCLKWithODMCombineTwoToOne; |
| 1249 | *NumberOfDPP = 2; |
| 1250 | } else { |
| 1251 | *TotalAvailablePipesSupport = false0; |
| 1252 | } |
| 1253 | } else { |
| 1254 | if (TotalNumberOfActiveDPP + 1 <= MaxNumDPP) |
| 1255 | *NumberOfDPP = 1; |
| 1256 | else |
| 1257 | *TotalAvailablePipesSupport = false0; |
| 1258 | } |
| 1259 | if (OutFormat == dm_420 && HActive > DCN32_MAX_FMT_420_BUFFER_WIDTH4096 && |
| 1260 | ODMUse != dm_odm_combine_policy_4to1) { |
| 1261 | if (HActive > DCN32_MAX_FMT_420_BUFFER_WIDTH4096 * 4) { |
| 1262 | *ODMMode = dm_odm_combine_mode_disabled; |
| 1263 | *NumberOfDPP = 0; |
| 1264 | *TotalAvailablePipesSupport = false0; |
| 1265 | } else if (HActive > DCN32_MAX_FMT_420_BUFFER_WIDTH4096 * 2 || |
| 1266 | *ODMMode == dm_odm_combine_mode_4to1) { |
| 1267 | *ODMMode = dm_odm_combine_mode_4to1; |
| 1268 | *RequiredDISPCLKPerSurface = SurfaceRequiredDISPCLKWithODMCombineFourToOne; |
| 1269 | *NumberOfDPP = 4; |
| 1270 | } else { |
| 1271 | *ODMMode = dm_odm_combine_mode_2to1; |
| 1272 | *RequiredDISPCLKPerSurface = SurfaceRequiredDISPCLKWithODMCombineTwoToOne; |
| 1273 | *NumberOfDPP = 2; |
| 1274 | } |
| 1275 | } |
| 1276 | if (Output == dm_hdmi && OutFormat == dm_420 && |
| 1277 | HActive > DCN32_MAX_FMT_420_BUFFER_WIDTH4096) { |
| 1278 | *ODMMode = dm_odm_combine_mode_disabled; |
| 1279 | *NumberOfDPP = 0; |
| 1280 | *TotalAvailablePipesSupport = false0; |
| 1281 | } |
| 1282 | } |
| 1283 | |
| 1284 | double dml32_CalculateRequiredDispclk( |
| 1285 | enum odm_combine_mode ODMMode, |
| 1286 | double PixelClock, |
| 1287 | double DISPCLKDPPCLKDSCCLKDownSpreading, |
| 1288 | double DISPCLKRampingMargin, |
| 1289 | double DISPCLKDPPCLKVCOSpeed, |
| 1290 | double MaxDispclk) |
| 1291 | { |
| 1292 | double RequiredDispclk = 0.; |
| 1293 | double PixelClockAfterODM; |
| 1294 | double DISPCLKWithRampingRoundedToDFSGranularity; |
| 1295 | double DISPCLKWithoutRampingRoundedToDFSGranularity; |
| 1296 | double MaxDispclkRoundedDownToDFSGranularity; |
| 1297 | |
| 1298 | if (ODMMode == dm_odm_combine_mode_4to1) |
| 1299 | PixelClockAfterODM = PixelClock / 4; |
| 1300 | else if (ODMMode == dm_odm_combine_mode_2to1) |
| 1301 | PixelClockAfterODM = PixelClock / 2; |
| 1302 | else |
| 1303 | PixelClockAfterODM = PixelClock; |
| 1304 | |
| 1305 | |
| 1306 | DISPCLKWithRampingRoundedToDFSGranularity = dml32_RoundToDFSGranularity( |
| 1307 | PixelClockAfterODM * (1 + DISPCLKDPPCLKDSCCLKDownSpreading / 100) |
| 1308 | * (1 + DISPCLKRampingMargin / 100), 1, DISPCLKDPPCLKVCOSpeed); |
| 1309 | |
| 1310 | DISPCLKWithoutRampingRoundedToDFSGranularity = dml32_RoundToDFSGranularity( |
| 1311 | PixelClockAfterODM * (1 + DISPCLKDPPCLKDSCCLKDownSpreading / 100), 1, DISPCLKDPPCLKVCOSpeed); |
| 1312 | |
| 1313 | MaxDispclkRoundedDownToDFSGranularity = dml32_RoundToDFSGranularity(MaxDispclk, 0, DISPCLKDPPCLKVCOSpeed); |
| 1314 | |
| 1315 | if (DISPCLKWithoutRampingRoundedToDFSGranularity > MaxDispclkRoundedDownToDFSGranularity) |
| 1316 | RequiredDispclk = DISPCLKWithoutRampingRoundedToDFSGranularity; |
| 1317 | else if (DISPCLKWithRampingRoundedToDFSGranularity > MaxDispclkRoundedDownToDFSGranularity) |
| 1318 | RequiredDispclk = MaxDispclkRoundedDownToDFSGranularity; |
| 1319 | else |
| 1320 | RequiredDispclk = DISPCLKWithRampingRoundedToDFSGranularity; |
| 1321 | |
| 1322 | return RequiredDispclk; |
| 1323 | } |
| 1324 | |
| 1325 | double dml32_RoundToDFSGranularity(double Clock, bool_Bool round_up, double VCOSpeed) |
| 1326 | { |
| 1327 | if (Clock <= 0.0) |
| 1328 | return 0.0; |
| 1329 | |
| 1330 | if (round_up) |
| 1331 | return VCOSpeed * 4.0 / dml_floor(VCOSpeed * 4.0 / Clock, 1.0); |
| 1332 | else |
| 1333 | return VCOSpeed * 4.0 / dml_ceil(VCOSpeed * 4.0 / Clock, 1.0); |
| 1334 | } |
| 1335 | |
| 1336 | void dml32_CalculateOutputLink( |
| 1337 | double PHYCLKPerState, |
| 1338 | double PHYCLKD18PerState, |
| 1339 | double PHYCLKD32PerState, |
| 1340 | double Downspreading, |
| 1341 | bool_Bool IsMainSurfaceUsingTheIndicatedTiming, |
| 1342 | enum output_encoder_class Output, |
| 1343 | enum output_format_class OutputFormat, |
| 1344 | unsigned int HTotal, |
| 1345 | unsigned int HActive, |
| 1346 | double PixelClockBackEnd, |
| 1347 | double ForcedOutputLinkBPP, |
| 1348 | unsigned int DSCInputBitPerComponent, |
| 1349 | unsigned int NumberOfDSCSlices, |
| 1350 | double AudioSampleRate, |
| 1351 | unsigned int AudioSampleLayout, |
| 1352 | enum odm_combine_mode ODMModeNoDSC, |
| 1353 | enum odm_combine_mode ODMModeDSC, |
| 1354 | bool_Bool DSCEnable, |
| 1355 | unsigned int OutputLinkDPLanes, |
| 1356 | enum dm_output_link_dp_rate OutputLinkDPRate, |
| 1357 | |
| 1358 | /* Output */ |
| 1359 | bool_Bool *RequiresDSC, |
| 1360 | double *RequiresFEC, |
| 1361 | double *OutBpp, |
| 1362 | enum dm_output_type *OutputType, |
| 1363 | enum dm_output_rate *OutputRate, |
| 1364 | unsigned int *RequiredSlots) |
| 1365 | { |
| 1366 | bool_Bool LinkDSCEnable; |
| 1367 | unsigned int dummy; |
| 1368 | *RequiresDSC = false0; |
| 1369 | *RequiresFEC = false0; |
| 1370 | *OutBpp = 0; |
| 1371 | *OutputType = dm_output_type_unknown; |
| 1372 | *OutputRate = dm_output_rate_unknown; |
| 1373 | |
| 1374 | if (IsMainSurfaceUsingTheIndicatedTiming) { |
| 1375 | if (Output == dm_hdmi) { |
| 1376 | *RequiresDSC = false0; |
| 1377 | *RequiresFEC = false0; |
| 1378 | *OutBpp = dml32_TruncToValidBPP(dml_min(600, PHYCLKPerState) * 10, 3, HTotal, HActive, |
| 1379 | PixelClockBackEnd, ForcedOutputLinkBPP, false0, Output, OutputFormat, |
| 1380 | DSCInputBitPerComponent, NumberOfDSCSlices, AudioSampleRate, AudioSampleLayout, |
| 1381 | ODMModeNoDSC, ODMModeDSC, &dummy); |
| 1382 | //OutputTypeAndRate = "HDMI"; |
| 1383 | *OutputType = dm_output_type_hdmi; |
| 1384 | |
| 1385 | } else if (Output == dm_dp || Output == dm_dp2p0 || Output == dm_edp) { |
| 1386 | if (DSCEnable == true1) { |
| 1387 | *RequiresDSC = true1; |
| 1388 | LinkDSCEnable = true1; |
| 1389 | if (Output == dm_dp || Output == dm_dp2p0) |
| 1390 | *RequiresFEC = true1; |
| 1391 | else |
| 1392 | *RequiresFEC = false0; |
| 1393 | } else { |
| 1394 | *RequiresDSC = false0; |
| 1395 | LinkDSCEnable = false0; |
| 1396 | if (Output == dm_dp2p0) |
| 1397 | *RequiresFEC = true1; |
| 1398 | else |
| 1399 | *RequiresFEC = false0; |
| 1400 | } |
| 1401 | if (Output == dm_dp2p0) { |
| 1402 | *OutBpp = 0; |
| 1403 | if ((OutputLinkDPRate == dm_dp_rate_na || OutputLinkDPRate == dm_dp_rate_uhbr10) && |
| 1404 | PHYCLKD32PerState >= 10000 / 32) { |
| 1405 | *OutBpp = dml32_TruncToValidBPP((1 - Downspreading / 100) * 10000, |
| 1406 | OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, |
| 1407 | ForcedOutputLinkBPP, LinkDSCEnable, Output, OutputFormat, |
| 1408 | DSCInputBitPerComponent, NumberOfDSCSlices, AudioSampleRate, |
| 1409 | AudioSampleLayout, ODMModeNoDSC, ODMModeDSC, RequiredSlots); |
| 1410 | if (*OutBpp == 0 && PHYCLKD32PerState < 13500 / 32 && DSCEnable == true1 && |
| 1411 | ForcedOutputLinkBPP == 0) { |
| 1412 | *RequiresDSC = true1; |
| 1413 | LinkDSCEnable = true1; |
| 1414 | *OutBpp = dml32_TruncToValidBPP((1 - Downspreading / 100) * 10000, |
| 1415 | OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, |
| 1416 | ForcedOutputLinkBPP, LinkDSCEnable, Output, |
| 1417 | OutputFormat, DSCInputBitPerComponent, |
| 1418 | NumberOfDSCSlices, AudioSampleRate, AudioSampleLayout, |
| 1419 | ODMModeNoDSC, ODMModeDSC, RequiredSlots); |
| 1420 | } |
| 1421 | //OutputTypeAndRate = Output & " UHBR10"; |
| 1422 | *OutputType = dm_output_type_dp2p0; |
| 1423 | *OutputRate = dm_output_rate_dp_rate_uhbr10; |
| 1424 | } |
| 1425 | if ((OutputLinkDPRate == dm_dp_rate_na || OutputLinkDPRate == dm_dp_rate_uhbr13p5) && |
| 1426 | *OutBpp == 0 && PHYCLKD32PerState >= 13500 / 32) { |
| 1427 | *OutBpp = dml32_TruncToValidBPP((1 - Downspreading / 100) * 13500, |
| 1428 | OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, |
| 1429 | ForcedOutputLinkBPP, LinkDSCEnable, Output, OutputFormat, |
| 1430 | DSCInputBitPerComponent, NumberOfDSCSlices, AudioSampleRate, |
| 1431 | AudioSampleLayout, ODMModeNoDSC, ODMModeDSC, RequiredSlots); |
| 1432 | |
| 1433 | if (*OutBpp == 0 && PHYCLKD32PerState < 20000 / 32 && DSCEnable == true1 && |
| 1434 | ForcedOutputLinkBPP == 0) { |
| 1435 | *RequiresDSC = true1; |
| 1436 | LinkDSCEnable = true1; |
| 1437 | *OutBpp = dml32_TruncToValidBPP((1 - Downspreading / 100) * 13500, |
| 1438 | OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, |
| 1439 | ForcedOutputLinkBPP, LinkDSCEnable, Output, |
| 1440 | OutputFormat, DSCInputBitPerComponent, |
| 1441 | NumberOfDSCSlices, AudioSampleRate, AudioSampleLayout, |
| 1442 | ODMModeNoDSC, ODMModeDSC, RequiredSlots); |
| 1443 | } |
| 1444 | //OutputTypeAndRate = Output & " UHBR13p5"; |
| 1445 | *OutputType = dm_output_type_dp2p0; |
| 1446 | *OutputRate = dm_output_rate_dp_rate_uhbr13p5; |
| 1447 | } |
| 1448 | if ((OutputLinkDPRate == dm_dp_rate_na || OutputLinkDPRate == dm_dp_rate_uhbr20) && |
| 1449 | *OutBpp == 0 && PHYCLKD32PerState >= 20000 / 32) { |
| 1450 | *OutBpp = dml32_TruncToValidBPP((1 - Downspreading / 100) * 20000, |
| 1451 | OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, |
| 1452 | ForcedOutputLinkBPP, LinkDSCEnable, Output, OutputFormat, |
| 1453 | DSCInputBitPerComponent, NumberOfDSCSlices, AudioSampleRate, |
| 1454 | AudioSampleLayout, ODMModeNoDSC, ODMModeDSC, RequiredSlots); |
| 1455 | if (*OutBpp == 0 && DSCEnable == true1 && ForcedOutputLinkBPP == 0) { |
| 1456 | *RequiresDSC = true1; |
| 1457 | LinkDSCEnable = true1; |
| 1458 | *OutBpp = dml32_TruncToValidBPP((1 - Downspreading / 100) * 20000, |
| 1459 | OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, |
| 1460 | ForcedOutputLinkBPP, LinkDSCEnable, Output, |
| 1461 | OutputFormat, DSCInputBitPerComponent, |
| 1462 | NumberOfDSCSlices, AudioSampleRate, AudioSampleLayout, |
| 1463 | ODMModeNoDSC, ODMModeDSC, RequiredSlots); |
| 1464 | } |
| 1465 | //OutputTypeAndRate = Output & " UHBR20"; |
| 1466 | *OutputType = dm_output_type_dp2p0; |
| 1467 | *OutputRate = dm_output_rate_dp_rate_uhbr20; |
| 1468 | } |
| 1469 | } else { |
| 1470 | *OutBpp = 0; |
| 1471 | if ((OutputLinkDPRate == dm_dp_rate_na || OutputLinkDPRate == dm_dp_rate_hbr) && |
| 1472 | PHYCLKPerState >= 270) { |
| 1473 | *OutBpp = dml32_TruncToValidBPP((1 - Downspreading / 100) * 2700, |
| 1474 | OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, |
| 1475 | ForcedOutputLinkBPP, LinkDSCEnable, Output, OutputFormat, |
| 1476 | DSCInputBitPerComponent, NumberOfDSCSlices, AudioSampleRate, |
| 1477 | AudioSampleLayout, ODMModeNoDSC, ODMModeDSC, RequiredSlots); |
| 1478 | if (*OutBpp == 0 && PHYCLKPerState < 540 && DSCEnable == true1 && |
| 1479 | ForcedOutputLinkBPP == 0) { |
| 1480 | *RequiresDSC = true1; |
| 1481 | LinkDSCEnable = true1; |
| 1482 | if (Output == dm_dp) |
| 1483 | *RequiresFEC = true1; |
| 1484 | *OutBpp = dml32_TruncToValidBPP((1 - Downspreading / 100) * 2700, |
| 1485 | OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, |
| 1486 | ForcedOutputLinkBPP, LinkDSCEnable, Output, |
| 1487 | OutputFormat, DSCInputBitPerComponent, |
| 1488 | NumberOfDSCSlices, AudioSampleRate, AudioSampleLayout, |
| 1489 | ODMModeNoDSC, ODMModeDSC, RequiredSlots); |
| 1490 | } |
| 1491 | //OutputTypeAndRate = Output & " HBR"; |
| 1492 | *OutputType = (Output == dm_dp) ? dm_output_type_dp : dm_output_type_edp; |
| 1493 | *OutputRate = dm_output_rate_dp_rate_hbr; |
| 1494 | } |
| 1495 | if ((OutputLinkDPRate == dm_dp_rate_na || OutputLinkDPRate == dm_dp_rate_hbr2) && |
| 1496 | *OutBpp == 0 && PHYCLKPerState >= 540) { |
| 1497 | *OutBpp = dml32_TruncToValidBPP((1 - Downspreading / 100) * 5400, |
| 1498 | OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, |
| 1499 | ForcedOutputLinkBPP, LinkDSCEnable, Output, OutputFormat, |
| 1500 | DSCInputBitPerComponent, NumberOfDSCSlices, AudioSampleRate, |
| 1501 | AudioSampleLayout, ODMModeNoDSC, ODMModeDSC, RequiredSlots); |
| 1502 | |
| 1503 | if (*OutBpp == 0 && PHYCLKPerState < 810 && DSCEnable == true1 && |
| 1504 | ForcedOutputLinkBPP == 0) { |
| 1505 | *RequiresDSC = true1; |
| 1506 | LinkDSCEnable = true1; |
| 1507 | if (Output == dm_dp) |
| 1508 | *RequiresFEC = true1; |
| 1509 | |
| 1510 | *OutBpp = dml32_TruncToValidBPP((1 - Downspreading / 100) * 5400, |
| 1511 | OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, |
| 1512 | ForcedOutputLinkBPP, LinkDSCEnable, Output, |
| 1513 | OutputFormat, DSCInputBitPerComponent, |
| 1514 | NumberOfDSCSlices, AudioSampleRate, AudioSampleLayout, |
| 1515 | ODMModeNoDSC, ODMModeDSC, RequiredSlots); |
| 1516 | } |
| 1517 | //OutputTypeAndRate = Output & " HBR2"; |
| 1518 | *OutputType = (Output == dm_dp) ? dm_output_type_dp : dm_output_type_edp; |
| 1519 | *OutputRate = dm_output_rate_dp_rate_hbr2; |
| 1520 | } |
| 1521 | if ((OutputLinkDPRate == dm_dp_rate_na || OutputLinkDPRate == dm_dp_rate_hbr3) && *OutBpp == 0 && PHYCLKPerState >= 810) { |
| 1522 | *OutBpp = dml32_TruncToValidBPP((1 - Downspreading / 100) * 8100, |
| 1523 | OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, |
| 1524 | ForcedOutputLinkBPP, LinkDSCEnable, Output, |
| 1525 | OutputFormat, DSCInputBitPerComponent, NumberOfDSCSlices, |
| 1526 | AudioSampleRate, AudioSampleLayout, ODMModeNoDSC, ODMModeDSC, |
| 1527 | RequiredSlots); |
| 1528 | |
| 1529 | if (*OutBpp == 0 && DSCEnable == true1 && ForcedOutputLinkBPP == 0) { |
| 1530 | *RequiresDSC = true1; |
| 1531 | LinkDSCEnable = true1; |
| 1532 | if (Output == dm_dp) |
| 1533 | *RequiresFEC = true1; |
| 1534 | |
| 1535 | *OutBpp = dml32_TruncToValidBPP((1 - Downspreading / 100) * 8100, |
| 1536 | OutputLinkDPLanes, HTotal, HActive, PixelClockBackEnd, |
| 1537 | ForcedOutputLinkBPP, LinkDSCEnable, Output, |
| 1538 | OutputFormat, DSCInputBitPerComponent, |
| 1539 | NumberOfDSCSlices, AudioSampleRate, AudioSampleLayout, |
| 1540 | ODMModeNoDSC, ODMModeDSC, RequiredSlots); |
| 1541 | } |
| 1542 | //OutputTypeAndRate = Output & " HBR3"; |
| 1543 | *OutputType = (Output == dm_dp) ? dm_output_type_dp : dm_output_type_edp; |
| 1544 | *OutputRate = dm_output_rate_dp_rate_hbr3; |
| 1545 | } |
| 1546 | } |
| 1547 | } |
| 1548 | } |
| 1549 | } |
| 1550 | |
| 1551 | void dml32_CalculateDPPCLK( |
| 1552 | unsigned int NumberOfActiveSurfaces, |
| 1553 | double DISPCLKDPPCLKDSCCLKDownSpreading, |
| 1554 | double DISPCLKDPPCLKVCOSpeed, |
| 1555 | double DPPCLKUsingSingleDPP[], |
| 1556 | unsigned int DPPPerSurface[], |
| 1557 | |
| 1558 | /* output */ |
| 1559 | double *GlobalDPPCLK, |
| 1560 | double Dppclk[]) |
| 1561 | { |
| 1562 | unsigned int k; |
| 1563 | *GlobalDPPCLK = 0; |
| 1564 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 1565 | Dppclk[k] = DPPCLKUsingSingleDPP[k] / DPPPerSurface[k] * (1 + DISPCLKDPPCLKDSCCLKDownSpreading / 100); |
| 1566 | *GlobalDPPCLK = dml_max(*GlobalDPPCLK, Dppclk[k]); |
| 1567 | } |
| 1568 | *GlobalDPPCLK = dml32_RoundToDFSGranularity(*GlobalDPPCLK, 1, DISPCLKDPPCLKVCOSpeed); |
| 1569 | for (k = 0; k < NumberOfActiveSurfaces; ++k) |
| 1570 | Dppclk[k] = *GlobalDPPCLK / 255 * dml_ceil(Dppclk[k] * 255.0 / *GlobalDPPCLK, 1.0); |
| 1571 | } |
| 1572 | |
| 1573 | double dml32_TruncToValidBPP( |
| 1574 | double LinkBitRate, |
| 1575 | unsigned int Lanes, |
| 1576 | unsigned int HTotal, |
| 1577 | unsigned int HActive, |
| 1578 | double PixelClock, |
| 1579 | double DesiredBPP, |
| 1580 | bool_Bool DSCEnable, |
| 1581 | enum output_encoder_class Output, |
| 1582 | enum output_format_class Format, |
| 1583 | unsigned int DSCInputBitPerComponent, |
| 1584 | unsigned int DSCSlices, |
| 1585 | unsigned int AudioRate, |
| 1586 | unsigned int AudioLayout, |
| 1587 | enum odm_combine_mode ODMModeNoDSC, |
| 1588 | enum odm_combine_mode ODMModeDSC, |
| 1589 | /* Output */ |
| 1590 | unsigned int *RequiredSlots) |
| 1591 | { |
| 1592 | double MaxLinkBPP; |
| 1593 | unsigned int MinDSCBPP; |
| 1594 | double MaxDSCBPP; |
| 1595 | unsigned int NonDSCBPP0; |
| 1596 | unsigned int NonDSCBPP1; |
| 1597 | unsigned int NonDSCBPP2; |
| 1598 | unsigned int NonDSCBPP3; |
| 1599 | |
| 1600 | if (Format == dm_420) { |
| 1601 | NonDSCBPP0 = 12; |
| 1602 | NonDSCBPP1 = 15; |
| 1603 | NonDSCBPP2 = 18; |
| 1604 | MinDSCBPP = 6; |
| 1605 | MaxDSCBPP = 1.5 * DSCInputBitPerComponent - 1 / 16; |
| 1606 | } else if (Format == dm_444) { |
| 1607 | NonDSCBPP0 = 18; |
| 1608 | NonDSCBPP1 = 24; |
| 1609 | NonDSCBPP2 = 30; |
| 1610 | NonDSCBPP3 = 36; |
| 1611 | MinDSCBPP = 8; |
| 1612 | MaxDSCBPP = 3 * DSCInputBitPerComponent - 1.0 / 16; |
| 1613 | } else { |
| 1614 | if (Output == dm_hdmi) { |
| 1615 | NonDSCBPP0 = 24; |
| 1616 | NonDSCBPP1 = 24; |
| 1617 | NonDSCBPP2 = 24; |
| 1618 | } else { |
| 1619 | NonDSCBPP0 = 16; |
| 1620 | NonDSCBPP1 = 20; |
| 1621 | NonDSCBPP2 = 24; |
| 1622 | } |
| 1623 | if (Format == dm_n422) { |
| 1624 | MinDSCBPP = 7; |
| 1625 | MaxDSCBPP = 2 * DSCInputBitPerComponent - 1.0 / 16.0; |
| 1626 | } else { |
| 1627 | MinDSCBPP = 8; |
| 1628 | MaxDSCBPP = 3 * DSCInputBitPerComponent - 1.0 / 16.0; |
| 1629 | } |
| 1630 | } |
| 1631 | if (Output == dm_dp2p0) { |
| 1632 | MaxLinkBPP = LinkBitRate * Lanes / PixelClock * 128 / 132 * 383 / 384 * 65536 / 65540; |
| 1633 | } else if (DSCEnable && Output == dm_dp) { |
| 1634 | MaxLinkBPP = LinkBitRate / 10 * 8 * Lanes / PixelClock * (1 - 2.4 / 100); |
| 1635 | } else { |
| 1636 | MaxLinkBPP = LinkBitRate / 10 * 8 * Lanes / PixelClock; |
| 1637 | } |
| 1638 | |
| 1639 | if (DSCEnable) { |
| 1640 | if (ODMModeDSC == dm_odm_combine_mode_4to1) |
| 1641 | MaxLinkBPP = dml_min(MaxLinkBPP, 16); |
| 1642 | else if (ODMModeDSC == dm_odm_combine_mode_2to1) |
| 1643 | MaxLinkBPP = dml_min(MaxLinkBPP, 32); |
| 1644 | else if (ODMModeDSC == dm_odm_split_mode_1to2) |
| 1645 | MaxLinkBPP = 2 * MaxLinkBPP; |
| 1646 | } else { |
| 1647 | if (ODMModeNoDSC == dm_odm_combine_mode_4to1) |
| 1648 | MaxLinkBPP = dml_min(MaxLinkBPP, 16); |
| 1649 | else if (ODMModeNoDSC == dm_odm_combine_mode_2to1) |
| 1650 | MaxLinkBPP = dml_min(MaxLinkBPP, 32); |
| 1651 | else if (ODMModeNoDSC == dm_odm_split_mode_1to2) |
| 1652 | MaxLinkBPP = 2 * MaxLinkBPP; |
| 1653 | } |
| 1654 | |
| 1655 | if (DesiredBPP == 0) { |
| 1656 | if (DSCEnable) { |
| 1657 | if (MaxLinkBPP < MinDSCBPP) |
| 1658 | return BPP_INVALID0; |
| 1659 | else if (MaxLinkBPP >= MaxDSCBPP) |
| 1660 | return MaxDSCBPP; |
| 1661 | else |
| 1662 | return dml_floor(16.0 * MaxLinkBPP, 1.0) / 16.0; |
| 1663 | } else { |
| 1664 | if (MaxLinkBPP >= NonDSCBPP3) |
| 1665 | return NonDSCBPP3; |
| 1666 | else if (MaxLinkBPP >= NonDSCBPP2) |
| 1667 | return NonDSCBPP2; |
| 1668 | else if (MaxLinkBPP >= NonDSCBPP1) |
| 1669 | return NonDSCBPP1; |
| 1670 | else if (MaxLinkBPP >= NonDSCBPP0) |
| 1671 | return 16.0; |
| 1672 | else |
| 1673 | return BPP_INVALID0; |
| 1674 | } |
| 1675 | } else { |
| 1676 | if (!((DSCEnable == false0 && (DesiredBPP == NonDSCBPP2 || DesiredBPP == NonDSCBPP1 || |
| 1677 | DesiredBPP == NonDSCBPP0 || DesiredBPP == NonDSCBPP3)) || |
| 1678 | (DSCEnable && DesiredBPP >= MinDSCBPP && DesiredBPP <= MaxDSCBPP))) |
| 1679 | return BPP_INVALID0; |
| 1680 | else |
| 1681 | return DesiredBPP; |
| 1682 | } |
| 1683 | |
| 1684 | *RequiredSlots = dml_ceil(DesiredBPP / MaxLinkBPP * 64, 1); |
| 1685 | |
| 1686 | return BPP_INVALID0; |
| 1687 | } // TruncToValidBPP |
| 1688 | |
| 1689 | double dml32_RequiredDTBCLK( |
| 1690 | bool_Bool DSCEnable, |
| 1691 | double PixelClock, |
| 1692 | enum output_format_class OutputFormat, |
| 1693 | double OutputBpp, |
| 1694 | unsigned int DSCSlices, |
| 1695 | unsigned int HTotal, |
| 1696 | unsigned int HActive, |
| 1697 | unsigned int AudioRate, |
| 1698 | unsigned int AudioLayout) |
| 1699 | { |
| 1700 | double PixelWordRate; |
| 1701 | double HCActive; |
| 1702 | double HCBlank; |
| 1703 | double AverageTribyteRate; |
| 1704 | double HActiveTribyteRate; |
| 1705 | |
| 1706 | if (DSCEnable != true1) |
| 1707 | return dml_max(PixelClock / 4.0 * OutputBpp / 24.0, 25.0); |
| 1708 | |
| 1709 | PixelWordRate = PixelClock / (OutputFormat == dm_444 ? 1 : 2); |
| 1710 | HCActive = dml_ceil(DSCSlices * dml_ceil(OutputBpp * |
| 1711 | dml_ceil(HActive / DSCSlices, 1) / 8.0, 1) / 3.0, 1); |
| 1712 | HCBlank = 64 + 32 * |
| 1713 | dml_ceil(AudioRate * (AudioLayout == 1 ? 1 : 0.25) * HTotal / (PixelClock * 1000), 1); |
| 1714 | AverageTribyteRate = PixelWordRate * (HCActive + HCBlank) / HTotal; |
| 1715 | HActiveTribyteRate = PixelWordRate * HCActive / HActive; |
| 1716 | return dml_max4(PixelWordRate / 4.0, AverageTribyteRate / 4.0, HActiveTribyteRate / 4.0, 25.0) * 1.002; |
| 1717 | } |
| 1718 | |
| 1719 | unsigned int dml32_DSCDelayRequirement(bool_Bool DSCEnabled, |
| 1720 | enum odm_combine_mode ODMMode, |
| 1721 | unsigned int DSCInputBitPerComponent, |
| 1722 | double OutputBpp, |
| 1723 | unsigned int HActive, |
| 1724 | unsigned int HTotal, |
| 1725 | unsigned int NumberOfDSCSlices, |
| 1726 | enum output_format_class OutputFormat, |
| 1727 | enum output_encoder_class Output, |
| 1728 | double PixelClock, |
| 1729 | double PixelClockBackEnd, |
| 1730 | double dsc_delay_factor_wa) |
| 1731 | { |
| 1732 | unsigned int DSCDelayRequirement_val; |
| 1733 | |
| 1734 | if (DSCEnabled == true1 && OutputBpp != 0) { |
| 1735 | if (ODMMode == dm_odm_combine_mode_4to1) { |
| 1736 | DSCDelayRequirement_val = 4 * (dml32_dscceComputeDelay(DSCInputBitPerComponent, OutputBpp, |
| 1737 | dml_ceil(HActive / NumberOfDSCSlices, 1), NumberOfDSCSlices / 4, |
| 1738 | OutputFormat, Output) + dml32_dscComputeDelay(OutputFormat, Output)); |
| 1739 | } else if (ODMMode == dm_odm_combine_mode_2to1) { |
| 1740 | DSCDelayRequirement_val = 2 * (dml32_dscceComputeDelay(DSCInputBitPerComponent, OutputBpp, |
| 1741 | dml_ceil(HActive / NumberOfDSCSlices, 1), NumberOfDSCSlices / 2, |
| 1742 | OutputFormat, Output) + dml32_dscComputeDelay(OutputFormat, Output)); |
| 1743 | } else { |
| 1744 | DSCDelayRequirement_val = dml32_dscceComputeDelay(DSCInputBitPerComponent, OutputBpp, |
| 1745 | dml_ceil(HActive / NumberOfDSCSlices, 1), NumberOfDSCSlices, |
| 1746 | OutputFormat, Output) + dml32_dscComputeDelay(OutputFormat, Output); |
| 1747 | } |
| 1748 | |
| 1749 | DSCDelayRequirement_val = DSCDelayRequirement_val + (HTotal - HActive) * |
| 1750 | dml_ceil((double)DSCDelayRequirement_val / HActive, 1); |
| 1751 | |
| 1752 | DSCDelayRequirement_val = DSCDelayRequirement_val * PixelClock / PixelClockBackEnd; |
| 1753 | |
| 1754 | } else { |
| 1755 | DSCDelayRequirement_val = 0; |
| 1756 | } |
| 1757 | |
| 1758 | #ifdef __DML_VBA_DEBUG__ |
| 1759 | dml_print("DML::%s: DSCEnabled = %d\n", __func__, DSCEnabled){do { } while(0); }; |
| 1760 | dml_print("DML::%s: OutputBpp = %f\n", __func__, OutputBpp){do { } while(0); }; |
| 1761 | dml_print("DML::%s: HActive = %d\n", __func__, HActive){do { } while(0); }; |
| 1762 | dml_print("DML::%s: OutputFormat = %d\n", __func__, OutputFormat){do { } while(0); }; |
| 1763 | dml_print("DML::%s: DSCInputBitPerComponent = %d\n", __func__, DSCInputBitPerComponent){do { } while(0); }; |
| 1764 | dml_print("DML::%s: NumberOfDSCSlices = %d\n", __func__, NumberOfDSCSlices){do { } while(0); }; |
| 1765 | dml_print("DML::%s: DSCDelayRequirement_val = %d\n", __func__, DSCDelayRequirement_val){do { } while(0); }; |
| 1766 | #endif |
| 1767 | |
| 1768 | return dml_ceil(DSCDelayRequirement_val * dsc_delay_factor_wa, 1); |
| 1769 | } |
| 1770 | |
| 1771 | void dml32_CalculateSurfaceSizeInMall( |
| 1772 | unsigned int NumberOfActiveSurfaces, |
| 1773 | unsigned int MALLAllocatedForDCN, |
| 1774 | enum dm_use_mall_for_static_screen_mode UseMALLForStaticScreen[], |
| 1775 | bool_Bool DCCEnable[], |
| 1776 | bool_Bool ViewportStationary[], |
| 1777 | unsigned int ViewportXStartY[], |
| 1778 | unsigned int ViewportYStartY[], |
| 1779 | unsigned int ViewportXStartC[], |
| 1780 | unsigned int ViewportYStartC[], |
| 1781 | unsigned int ViewportWidthY[], |
| 1782 | unsigned int ViewportHeightY[], |
| 1783 | unsigned int BytesPerPixelY[], |
| 1784 | unsigned int ViewportWidthC[], |
| 1785 | unsigned int ViewportHeightC[], |
| 1786 | unsigned int BytesPerPixelC[], |
| 1787 | unsigned int SurfaceWidthY[], |
| 1788 | unsigned int SurfaceWidthC[], |
| 1789 | unsigned int SurfaceHeightY[], |
| 1790 | unsigned int SurfaceHeightC[], |
| 1791 | unsigned int Read256BytesBlockWidthY[], |
| 1792 | unsigned int Read256BytesBlockWidthC[], |
| 1793 | unsigned int Read256BytesBlockHeightY[], |
| 1794 | unsigned int Read256BytesBlockHeightC[], |
| 1795 | unsigned int ReadBlockWidthY[], |
| 1796 | unsigned int ReadBlockWidthC[], |
| 1797 | unsigned int ReadBlockHeightY[], |
| 1798 | unsigned int ReadBlockHeightC[], |
| 1799 | |
| 1800 | /* Output */ |
| 1801 | unsigned int SurfaceSizeInMALL[], |
| 1802 | bool_Bool *ExceededMALLSize) |
| 1803 | { |
| 1804 | unsigned int TotalSurfaceSizeInMALL = 0; |
| 1805 | unsigned int k; |
| 1806 | |
| 1807 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 1808 | if (ViewportStationary[k]) { |
| 1809 | SurfaceSizeInMALL[k] = dml_min(dml_ceil(SurfaceWidthY[k], ReadBlockWidthY[k]), |
| 1810 | dml_floor(ViewportXStartY[k] + ViewportWidthY[k] + ReadBlockWidthY[k] - 1, |
| 1811 | ReadBlockWidthY[k]) - dml_floor(ViewportXStartY[k], |
| 1812 | ReadBlockWidthY[k])) * dml_min(dml_ceil(SurfaceHeightY[k], |
| 1813 | ReadBlockHeightY[k]), dml_floor(ViewportYStartY[k] + |
| 1814 | ViewportHeightY[k] + ReadBlockHeightY[k] - 1, ReadBlockHeightY[k]) - |
| 1815 | dml_floor(ViewportYStartY[k], ReadBlockHeightY[k])) * BytesPerPixelY[k]; |
| 1816 | |
| 1817 | if (ReadBlockWidthC[k] > 0) { |
| 1818 | SurfaceSizeInMALL[k] = SurfaceSizeInMALL[k] + |
| 1819 | dml_min(dml_ceil(SurfaceWidthC[k], ReadBlockWidthC[k]), |
| 1820 | dml_floor(ViewportXStartC[k] + ViewportWidthC[k] + |
| 1821 | ReadBlockWidthC[k] - 1, ReadBlockWidthC[k]) - |
| 1822 | dml_floor(ViewportXStartC[k], ReadBlockWidthC[k])) * |
| 1823 | dml_min(dml_ceil(SurfaceHeightC[k], ReadBlockHeightC[k]), |
| 1824 | dml_floor(ViewportYStartC[k] + ViewportHeightC[k] + |
| 1825 | ReadBlockHeightC[k] - 1, ReadBlockHeightC[k]) - |
| 1826 | dml_floor(ViewportYStartC[k], ReadBlockHeightC[k])) * |
| 1827 | BytesPerPixelC[k]; |
| 1828 | } |
| 1829 | if (DCCEnable[k] == true1) { |
| 1830 | SurfaceSizeInMALL[k] = SurfaceSizeInMALL[k] + |
| 1831 | dml_min(dml_ceil(SurfaceWidthY[k], 8 * Read256BytesBlockWidthY[k]), |
| 1832 | dml_floor(ViewportXStartY[k] + ViewportWidthY[k] + 8 * |
| 1833 | Read256BytesBlockWidthY[k] - 1, 8 * Read256BytesBlockWidthY[k]) |
| 1834 | - dml_floor(ViewportXStartY[k], 8 * Read256BytesBlockWidthY[k])) |
| 1835 | * dml_min(dml_ceil(SurfaceHeightY[k], 8 * |
| 1836 | Read256BytesBlockHeightY[k]), dml_floor(ViewportYStartY[k] + |
| 1837 | ViewportHeightY[k] + 8 * Read256BytesBlockHeightY[k] - 1, 8 * |
| 1838 | Read256BytesBlockHeightY[k]) - dml_floor(ViewportYStartY[k], 8 |
| 1839 | * Read256BytesBlockHeightY[k])) * BytesPerPixelY[k] / 256; |
| 1840 | if (Read256BytesBlockWidthC[k] > 0) { |
| 1841 | SurfaceSizeInMALL[k] = SurfaceSizeInMALL[k] + |
| 1842 | dml_min(dml_ceil(SurfaceWidthC[k], 8 * |
| 1843 | Read256BytesBlockWidthC[k]), |
| 1844 | dml_floor(ViewportXStartC[k] + ViewportWidthC[k] + 8 |
| 1845 | * Read256BytesBlockWidthC[k] - 1, 8 * |
| 1846 | Read256BytesBlockWidthC[k]) - |
| 1847 | dml_floor(ViewportXStartC[k], 8 * |
| 1848 | Read256BytesBlockWidthC[k])) * |
| 1849 | dml_min(dml_ceil(SurfaceHeightC[k], 8 * |
| 1850 | Read256BytesBlockHeightC[k]), |
| 1851 | dml_floor(ViewportYStartC[k] + ViewportHeightC[k] + |
| 1852 | 8 * Read256BytesBlockHeightC[k] - 1, 8 * |
| 1853 | Read256BytesBlockHeightC[k]) - |
| 1854 | dml_floor(ViewportYStartC[k], 8 * |
| 1855 | Read256BytesBlockHeightC[k])) * |
| 1856 | BytesPerPixelC[k] / 256; |
| 1857 | } |
| 1858 | } |
| 1859 | } else { |
| 1860 | SurfaceSizeInMALL[k] = dml_ceil(dml_min(SurfaceWidthY[k], ViewportWidthY[k] + |
| 1861 | ReadBlockWidthY[k] - 1), ReadBlockWidthY[k]) * |
| 1862 | dml_ceil(dml_min(SurfaceHeightY[k], ViewportHeightY[k] + |
| 1863 | ReadBlockHeightY[k] - 1), ReadBlockHeightY[k]) * |
| 1864 | BytesPerPixelY[k]; |
| 1865 | if (ReadBlockWidthC[k] > 0) { |
| 1866 | SurfaceSizeInMALL[k] = SurfaceSizeInMALL[k] + |
| 1867 | dml_ceil(dml_min(SurfaceWidthC[k], ViewportWidthC[k] + |
| 1868 | ReadBlockWidthC[k] - 1), ReadBlockWidthC[k]) * |
| 1869 | dml_ceil(dml_min(SurfaceHeightC[k], ViewportHeightC[k] + |
| 1870 | ReadBlockHeightC[k] - 1), ReadBlockHeightC[k]) * |
| 1871 | BytesPerPixelC[k]; |
| 1872 | } |
| 1873 | if (DCCEnable[k] == true1) { |
| 1874 | SurfaceSizeInMALL[k] = SurfaceSizeInMALL[k] + |
| 1875 | dml_ceil(dml_min(SurfaceWidthY[k], ViewportWidthY[k] + 8 * |
| 1876 | Read256BytesBlockWidthY[k] - 1), 8 * |
| 1877 | Read256BytesBlockWidthY[k]) * |
| 1878 | dml_ceil(dml_min(SurfaceHeightY[k], ViewportHeightY[k] + 8 * |
| 1879 | Read256BytesBlockHeightY[k] - 1), 8 * |
| 1880 | Read256BytesBlockHeightY[k]) * BytesPerPixelY[k] / 256; |
| 1881 | |
| 1882 | if (Read256BytesBlockWidthC[k] > 0) { |
| 1883 | SurfaceSizeInMALL[k] = SurfaceSizeInMALL[k] + |
| 1884 | dml_ceil(dml_min(SurfaceWidthC[k], ViewportWidthC[k] + 8 * |
| 1885 | Read256BytesBlockWidthC[k] - 1), 8 * |
| 1886 | Read256BytesBlockWidthC[k]) * |
| 1887 | dml_ceil(dml_min(SurfaceHeightC[k], ViewportHeightC[k] + 8 * |
| 1888 | Read256BytesBlockHeightC[k] - 1), 8 * |
| 1889 | Read256BytesBlockHeightC[k]) * |
| 1890 | BytesPerPixelC[k] / 256; |
| 1891 | } |
| 1892 | } |
| 1893 | } |
| 1894 | } |
| 1895 | |
| 1896 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 1897 | if (UseMALLForStaticScreen[k] == dm_use_mall_static_screen_enable) |
| 1898 | TotalSurfaceSizeInMALL = TotalSurfaceSizeInMALL + SurfaceSizeInMALL[k]; |
| 1899 | } |
| 1900 | *ExceededMALLSize = (TotalSurfaceSizeInMALL > MALLAllocatedForDCN * 1024 * 1024); |
| 1901 | } // CalculateSurfaceSizeInMall |
| 1902 | |
| 1903 | void dml32_CalculateVMRowAndSwath( |
| 1904 | unsigned int NumberOfActiveSurfaces, |
| 1905 | DmlPipe myPipe[], |
| 1906 | unsigned int SurfaceSizeInMALL[], |
| 1907 | unsigned int PTEBufferSizeInRequestsLuma, |
| 1908 | unsigned int PTEBufferSizeInRequestsChroma, |
| 1909 | unsigned int DCCMetaBufferSizeBytes, |
| 1910 | enum dm_use_mall_for_static_screen_mode UseMALLForStaticScreen[], |
| 1911 | enum dm_use_mall_for_pstate_change_mode UseMALLForPStateChange[], |
| 1912 | unsigned int MALLAllocatedForDCN, |
| 1913 | double SwathWidthY[], |
| 1914 | double SwathWidthC[], |
| 1915 | bool_Bool GPUVMEnable, |
| 1916 | bool_Bool HostVMEnable, |
| 1917 | unsigned int HostVMMaxNonCachedPageTableLevels, |
| 1918 | unsigned int GPUVMMaxPageTableLevels, |
| 1919 | unsigned int GPUVMMinPageSizeKBytes[], |
| 1920 | unsigned int HostVMMinPageSize, |
| 1921 | |
| 1922 | /* Output */ |
| 1923 | bool_Bool PTEBufferSizeNotExceeded[], |
| 1924 | bool_Bool DCCMetaBufferSizeNotExceeded[], |
| 1925 | unsigned int dpte_row_width_luma_ub[], |
| 1926 | unsigned int dpte_row_width_chroma_ub[], |
| 1927 | unsigned int dpte_row_height_luma[], |
| 1928 | unsigned int dpte_row_height_chroma[], |
| 1929 | unsigned int dpte_row_height_linear_luma[], // VBA_DELTA |
| 1930 | unsigned int dpte_row_height_linear_chroma[], // VBA_DELTA |
| 1931 | unsigned int meta_req_width[], |
| 1932 | unsigned int meta_req_width_chroma[], |
| 1933 | unsigned int meta_req_height[], |
| 1934 | unsigned int meta_req_height_chroma[], |
| 1935 | unsigned int meta_row_width[], |
| 1936 | unsigned int meta_row_width_chroma[], |
| 1937 | unsigned int meta_row_height[], |
| 1938 | unsigned int meta_row_height_chroma[], |
| 1939 | unsigned int vm_group_bytes[], |
| 1940 | unsigned int dpte_group_bytes[], |
| 1941 | unsigned int PixelPTEReqWidthY[], |
| 1942 | unsigned int PixelPTEReqHeightY[], |
| 1943 | unsigned int PTERequestSizeY[], |
| 1944 | unsigned int PixelPTEReqWidthC[], |
| 1945 | unsigned int PixelPTEReqHeightC[], |
| 1946 | unsigned int PTERequestSizeC[], |
| 1947 | unsigned int dpde0_bytes_per_frame_ub_l[], |
| 1948 | unsigned int meta_pte_bytes_per_frame_ub_l[], |
| 1949 | unsigned int dpde0_bytes_per_frame_ub_c[], |
| 1950 | unsigned int meta_pte_bytes_per_frame_ub_c[], |
| 1951 | double PrefetchSourceLinesY[], |
| 1952 | double PrefetchSourceLinesC[], |
| 1953 | double VInitPreFillY[], |
| 1954 | double VInitPreFillC[], |
| 1955 | unsigned int MaxNumSwathY[], |
| 1956 | unsigned int MaxNumSwathC[], |
| 1957 | double meta_row_bw[], |
| 1958 | double dpte_row_bw[], |
| 1959 | double PixelPTEBytesPerRow[], |
| 1960 | double PDEAndMetaPTEBytesFrame[], |
| 1961 | double MetaRowByte[], |
| 1962 | bool_Bool use_one_row_for_frame[], |
| 1963 | bool_Bool use_one_row_for_frame_flip[], |
| 1964 | bool_Bool UsesMALLForStaticScreen[], |
| 1965 | bool_Bool PTE_BUFFER_MODE[], |
| 1966 | unsigned int BIGK_FRAGMENT_SIZE[]) |
| 1967 | { |
| 1968 | unsigned int k; |
| 1969 | unsigned int PTEBufferSizeInRequestsForLuma[DC__NUM_DPP__MAX8]; |
| 1970 | unsigned int PTEBufferSizeInRequestsForChroma[DC__NUM_DPP__MAX8]; |
| 1971 | unsigned int PDEAndMetaPTEBytesFrameY; |
| 1972 | unsigned int PDEAndMetaPTEBytesFrameC; |
| 1973 | unsigned int MetaRowByteY[DC__NUM_DPP__MAX8]; |
| 1974 | unsigned int MetaRowByteC[DC__NUM_DPP__MAX8]; |
| 1975 | unsigned int PixelPTEBytesPerRowY[DC__NUM_DPP__MAX8]; |
| 1976 | unsigned int PixelPTEBytesPerRowC[DC__NUM_DPP__MAX8]; |
| 1977 | unsigned int PixelPTEBytesPerRowY_one_row_per_frame[DC__NUM_DPP__MAX8]; |
| 1978 | unsigned int PixelPTEBytesPerRowC_one_row_per_frame[DC__NUM_DPP__MAX8]; |
| 1979 | unsigned int dpte_row_width_luma_ub_one_row_per_frame[DC__NUM_DPP__MAX8]; |
| 1980 | unsigned int dpte_row_height_luma_one_row_per_frame[DC__NUM_DPP__MAX8]; |
| 1981 | unsigned int dpte_row_width_chroma_ub_one_row_per_frame[DC__NUM_DPP__MAX8]; |
| 1982 | unsigned int dpte_row_height_chroma_one_row_per_frame[DC__NUM_DPP__MAX8]; |
| 1983 | bool_Bool one_row_per_frame_fits_in_buffer[DC__NUM_DPP__MAX8]; |
| 1984 | |
| 1985 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 1986 | if (HostVMEnable == true1) { |
| 1987 | vm_group_bytes[k] = 512; |
| 1988 | dpte_group_bytes[k] = 512; |
| 1989 | } else if (GPUVMEnable == true1) { |
| 1990 | vm_group_bytes[k] = 2048; |
| 1991 | if (GPUVMMinPageSizeKBytes[k] >= 64 && IsVertical(myPipe[k].SourceRotation)) |
| 1992 | dpte_group_bytes[k] = 512; |
| 1993 | else |
| 1994 | dpte_group_bytes[k] = 2048; |
| 1995 | } else { |
| 1996 | vm_group_bytes[k] = 0; |
| 1997 | dpte_group_bytes[k] = 0; |
| 1998 | } |
| 1999 | |
| 2000 | if (myPipe[k].SourcePixelFormat == dm_420_8 || myPipe[k].SourcePixelFormat == dm_420_10 || |
| 2001 | myPipe[k].SourcePixelFormat == dm_420_12 || |
| 2002 | myPipe[k].SourcePixelFormat == dm_rgbe_alpha) { |
| 2003 | if ((myPipe[k].SourcePixelFormat == dm_420_10 || myPipe[k].SourcePixelFormat == dm_420_12) && |
| 2004 | !IsVertical(myPipe[k].SourceRotation)) { |
| 2005 | PTEBufferSizeInRequestsForLuma[k] = |
| 2006 | (PTEBufferSizeInRequestsLuma + PTEBufferSizeInRequestsChroma) / 2; |
| 2007 | PTEBufferSizeInRequestsForChroma[k] = PTEBufferSizeInRequestsForLuma[k]; |
| 2008 | } else { |
| 2009 | PTEBufferSizeInRequestsForLuma[k] = PTEBufferSizeInRequestsLuma; |
| 2010 | PTEBufferSizeInRequestsForChroma[k] = PTEBufferSizeInRequestsChroma; |
| 2011 | } |
| 2012 | |
| 2013 | PDEAndMetaPTEBytesFrameC = dml32_CalculateVMAndRowBytes( |
| 2014 | myPipe[k].ViewportStationary, |
| 2015 | myPipe[k].DCCEnable, |
| 2016 | myPipe[k].DPPPerSurface, |
| 2017 | myPipe[k].BlockHeight256BytesC, |
| 2018 | myPipe[k].BlockWidth256BytesC, |
| 2019 | myPipe[k].SourcePixelFormat, |
| 2020 | myPipe[k].SurfaceTiling, |
| 2021 | myPipe[k].BytePerPixelC, |
| 2022 | myPipe[k].SourceRotation, |
| 2023 | SwathWidthC[k], |
| 2024 | myPipe[k].ViewportHeightChroma, |
| 2025 | myPipe[k].ViewportXStartC, |
| 2026 | myPipe[k].ViewportYStartC, |
| 2027 | GPUVMEnable, |
| 2028 | HostVMEnable, |
| 2029 | HostVMMaxNonCachedPageTableLevels, |
| 2030 | GPUVMMaxPageTableLevels, |
| 2031 | GPUVMMinPageSizeKBytes[k], |
| 2032 | HostVMMinPageSize, |
| 2033 | PTEBufferSizeInRequestsForChroma[k], |
| 2034 | myPipe[k].PitchC, |
| 2035 | myPipe[k].DCCMetaPitchC, |
| 2036 | myPipe[k].BlockWidthC, |
| 2037 | myPipe[k].BlockHeightC, |
| 2038 | |
| 2039 | /* Output */ |
| 2040 | &MetaRowByteC[k], |
| 2041 | &PixelPTEBytesPerRowC[k], |
| 2042 | &dpte_row_width_chroma_ub[k], |
| 2043 | &dpte_row_height_chroma[k], |
| 2044 | &dpte_row_height_linear_chroma[k], |
| 2045 | &PixelPTEBytesPerRowC_one_row_per_frame[k], |
| 2046 | &dpte_row_width_chroma_ub_one_row_per_frame[k], |
| 2047 | &dpte_row_height_chroma_one_row_per_frame[k], |
| 2048 | &meta_req_width_chroma[k], |
| 2049 | &meta_req_height_chroma[k], |
| 2050 | &meta_row_width_chroma[k], |
| 2051 | &meta_row_height_chroma[k], |
| 2052 | &PixelPTEReqWidthC[k], |
| 2053 | &PixelPTEReqHeightC[k], |
| 2054 | &PTERequestSizeC[k], |
| 2055 | &dpde0_bytes_per_frame_ub_c[k], |
| 2056 | &meta_pte_bytes_per_frame_ub_c[k]); |
| 2057 | |
| 2058 | PrefetchSourceLinesC[k] = dml32_CalculatePrefetchSourceLines( |
| 2059 | myPipe[k].VRatioChroma, |
| 2060 | myPipe[k].VTapsChroma, |
| 2061 | myPipe[k].InterlaceEnable, |
| 2062 | myPipe[k].ProgressiveToInterlaceUnitInOPP, |
| 2063 | myPipe[k].SwathHeightC, |
| 2064 | myPipe[k].SourceRotation, |
| 2065 | myPipe[k].ViewportStationary, |
| 2066 | SwathWidthC[k], |
| 2067 | myPipe[k].ViewportHeightChroma, |
| 2068 | myPipe[k].ViewportXStartC, |
| 2069 | myPipe[k].ViewportYStartC, |
| 2070 | |
| 2071 | /* Output */ |
| 2072 | &VInitPreFillC[k], |
| 2073 | &MaxNumSwathC[k]); |
| 2074 | } else { |
| 2075 | PTEBufferSizeInRequestsForLuma[k] = PTEBufferSizeInRequestsLuma + PTEBufferSizeInRequestsChroma; |
| 2076 | PTEBufferSizeInRequestsForChroma[k] = 0; |
| 2077 | PixelPTEBytesPerRowC[k] = 0; |
| 2078 | PDEAndMetaPTEBytesFrameC = 0; |
| 2079 | MetaRowByteC[k] = 0; |
| 2080 | MaxNumSwathC[k] = 0; |
| 2081 | PrefetchSourceLinesC[k] = 0; |
| 2082 | dpte_row_height_chroma_one_row_per_frame[k] = 0; |
| 2083 | dpte_row_width_chroma_ub_one_row_per_frame[k] = 0; |
| 2084 | PixelPTEBytesPerRowC_one_row_per_frame[k] = 0; |
| 2085 | } |
| 2086 | |
| 2087 | PDEAndMetaPTEBytesFrameY = dml32_CalculateVMAndRowBytes( |
| 2088 | myPipe[k].ViewportStationary, |
| 2089 | myPipe[k].DCCEnable, |
| 2090 | myPipe[k].DPPPerSurface, |
| 2091 | myPipe[k].BlockHeight256BytesY, |
| 2092 | myPipe[k].BlockWidth256BytesY, |
| 2093 | myPipe[k].SourcePixelFormat, |
| 2094 | myPipe[k].SurfaceTiling, |
| 2095 | myPipe[k].BytePerPixelY, |
| 2096 | myPipe[k].SourceRotation, |
| 2097 | SwathWidthY[k], |
| 2098 | myPipe[k].ViewportHeight, |
| 2099 | myPipe[k].ViewportXStart, |
| 2100 | myPipe[k].ViewportYStart, |
| 2101 | GPUVMEnable, |
| 2102 | HostVMEnable, |
| 2103 | HostVMMaxNonCachedPageTableLevels, |
| 2104 | GPUVMMaxPageTableLevels, |
| 2105 | GPUVMMinPageSizeKBytes[k], |
| 2106 | HostVMMinPageSize, |
| 2107 | PTEBufferSizeInRequestsForLuma[k], |
| 2108 | myPipe[k].PitchY, |
| 2109 | myPipe[k].DCCMetaPitchY, |
| 2110 | myPipe[k].BlockWidthY, |
| 2111 | myPipe[k].BlockHeightY, |
| 2112 | |
| 2113 | /* Output */ |
| 2114 | &MetaRowByteY[k], |
| 2115 | &PixelPTEBytesPerRowY[k], |
| 2116 | &dpte_row_width_luma_ub[k], |
| 2117 | &dpte_row_height_luma[k], |
| 2118 | &dpte_row_height_linear_luma[k], |
| 2119 | &PixelPTEBytesPerRowY_one_row_per_frame[k], |
| 2120 | &dpte_row_width_luma_ub_one_row_per_frame[k], |
| 2121 | &dpte_row_height_luma_one_row_per_frame[k], |
| 2122 | &meta_req_width[k], |
| 2123 | &meta_req_height[k], |
| 2124 | &meta_row_width[k], |
| 2125 | &meta_row_height[k], |
| 2126 | &PixelPTEReqWidthY[k], |
| 2127 | &PixelPTEReqHeightY[k], |
| 2128 | &PTERequestSizeY[k], |
| 2129 | &dpde0_bytes_per_frame_ub_l[k], |
| 2130 | &meta_pte_bytes_per_frame_ub_l[k]); |
| 2131 | |
| 2132 | PrefetchSourceLinesY[k] = dml32_CalculatePrefetchSourceLines( |
| 2133 | myPipe[k].VRatio, |
| 2134 | myPipe[k].VTaps, |
| 2135 | myPipe[k].InterlaceEnable, |
| 2136 | myPipe[k].ProgressiveToInterlaceUnitInOPP, |
| 2137 | myPipe[k].SwathHeightY, |
| 2138 | myPipe[k].SourceRotation, |
| 2139 | myPipe[k].ViewportStationary, |
| 2140 | SwathWidthY[k], |
| 2141 | myPipe[k].ViewportHeight, |
| 2142 | myPipe[k].ViewportXStart, |
| 2143 | myPipe[k].ViewportYStart, |
| 2144 | |
| 2145 | /* Output */ |
| 2146 | &VInitPreFillY[k], |
| 2147 | &MaxNumSwathY[k]); |
| 2148 | |
| 2149 | PDEAndMetaPTEBytesFrame[k] = PDEAndMetaPTEBytesFrameY + PDEAndMetaPTEBytesFrameC; |
| 2150 | MetaRowByte[k] = MetaRowByteY[k] + MetaRowByteC[k]; |
| 2151 | |
| 2152 | if (PixelPTEBytesPerRowY[k] <= 64 * PTEBufferSizeInRequestsForLuma[k] && |
| 2153 | PixelPTEBytesPerRowC[k] <= 64 * PTEBufferSizeInRequestsForChroma[k]) { |
| 2154 | PTEBufferSizeNotExceeded[k] = true1; |
| 2155 | } else { |
| 2156 | PTEBufferSizeNotExceeded[k] = false0; |
| 2157 | } |
| 2158 | |
| 2159 | one_row_per_frame_fits_in_buffer[k] = (PixelPTEBytesPerRowY_one_row_per_frame[k] <= 64 * 2 * |
| 2160 | PTEBufferSizeInRequestsForLuma[k] && |
| 2161 | PixelPTEBytesPerRowC_one_row_per_frame[k] <= 64 * 2 * PTEBufferSizeInRequestsForChroma[k]); |
| 2162 | } |
| 2163 | |
| 2164 | dml32_CalculateMALLUseForStaticScreen( |
| 2165 | NumberOfActiveSurfaces, |
| 2166 | MALLAllocatedForDCN, |
| 2167 | UseMALLForStaticScreen, // mode |
| 2168 | SurfaceSizeInMALL, |
| 2169 | one_row_per_frame_fits_in_buffer, |
| 2170 | /* Output */ |
| 2171 | UsesMALLForStaticScreen); // boolen |
| 2172 | |
| 2173 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 2174 | PTE_BUFFER_MODE[k] = myPipe[k].FORCE_ONE_ROW_FOR_FRAME || UsesMALLForStaticScreen[k] || |
| 2175 | (UseMALLForPStateChange[k] == dm_use_mall_pstate_change_sub_viewport) || |
| 2176 | (UseMALLForPStateChange[k] == dm_use_mall_pstate_change_phantom_pipe) || |
| 2177 | (GPUVMMinPageSizeKBytes[k] > 64); |
| 2178 | BIGK_FRAGMENT_SIZE[k] = dml_log2(GPUVMMinPageSizeKBytes[k] * 1024) - 12; |
| 2179 | } |
| 2180 | |
| 2181 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 2182 | #ifdef __DML_VBA_DEBUG__ |
| 2183 | dml_print("DML::%s: k=%d, SurfaceSizeInMALL = %d\n", __func__, k, SurfaceSizeInMALL[k]){do { } while(0); }; |
| 2184 | dml_print("DML::%s: k=%d, UsesMALLForStaticScreen = %d\n", __func__, k, UsesMALLForStaticScreen[k]){do { } while(0); }; |
| 2185 | #endif |
| 2186 | use_one_row_for_frame[k] = myPipe[k].FORCE_ONE_ROW_FOR_FRAME || UsesMALLForStaticScreen[k] || |
| 2187 | (UseMALLForPStateChange[k] == dm_use_mall_pstate_change_sub_viewport) || |
| 2188 | (UseMALLForPStateChange[k] == dm_use_mall_pstate_change_phantom_pipe) || |
| 2189 | (GPUVMMinPageSizeKBytes[k] > 64 && IsVertical(myPipe[k].SourceRotation)); |
| 2190 | |
| 2191 | use_one_row_for_frame_flip[k] = use_one_row_for_frame[k] && |
| 2192 | !(UseMALLForPStateChange[k] == dm_use_mall_pstate_change_full_frame); |
| 2193 | |
| 2194 | if (use_one_row_for_frame[k]) { |
| 2195 | dpte_row_height_luma[k] = dpte_row_height_luma_one_row_per_frame[k]; |
| 2196 | dpte_row_width_luma_ub[k] = dpte_row_width_luma_ub_one_row_per_frame[k]; |
| 2197 | PixelPTEBytesPerRowY[k] = PixelPTEBytesPerRowY_one_row_per_frame[k]; |
| 2198 | dpte_row_height_chroma[k] = dpte_row_height_chroma_one_row_per_frame[k]; |
| 2199 | dpte_row_width_chroma_ub[k] = dpte_row_width_chroma_ub_one_row_per_frame[k]; |
| 2200 | PixelPTEBytesPerRowC[k] = PixelPTEBytesPerRowC_one_row_per_frame[k]; |
| 2201 | PTEBufferSizeNotExceeded[k] = one_row_per_frame_fits_in_buffer[k]; |
| 2202 | } |
| 2203 | |
| 2204 | if (MetaRowByte[k] <= DCCMetaBufferSizeBytes) |
| 2205 | DCCMetaBufferSizeNotExceeded[k] = true1; |
| 2206 | else |
| 2207 | DCCMetaBufferSizeNotExceeded[k] = false0; |
| 2208 | |
| 2209 | PixelPTEBytesPerRow[k] = PixelPTEBytesPerRowY[k] + PixelPTEBytesPerRowC[k]; |
| 2210 | if (use_one_row_for_frame[k]) |
| 2211 | PixelPTEBytesPerRow[k] = PixelPTEBytesPerRow[k] / 2; |
| 2212 | |
| 2213 | dml32_CalculateRowBandwidth( |
| 2214 | GPUVMEnable, |
| 2215 | myPipe[k].SourcePixelFormat, |
| 2216 | myPipe[k].VRatio, |
| 2217 | myPipe[k].VRatioChroma, |
| 2218 | myPipe[k].DCCEnable, |
| 2219 | myPipe[k].HTotal / myPipe[k].PixelClock, |
| 2220 | MetaRowByteY[k], MetaRowByteC[k], |
| 2221 | meta_row_height[k], |
| 2222 | meta_row_height_chroma[k], |
| 2223 | PixelPTEBytesPerRowY[k], |
| 2224 | PixelPTEBytesPerRowC[k], |
| 2225 | dpte_row_height_luma[k], |
| 2226 | dpte_row_height_chroma[k], |
| 2227 | |
| 2228 | /* Output */ |
| 2229 | &meta_row_bw[k], |
| 2230 | &dpte_row_bw[k]); |
| 2231 | #ifdef __DML_VBA_DEBUG__ |
| 2232 | dml_print("DML::%s: k=%d, use_one_row_for_frame = %d\n", __func__, k, use_one_row_for_frame[k]){do { } while(0); }; |
| 2233 | dml_print("DML::%s: k=%d, use_one_row_for_frame_flip = %d\n",{do { } while(0); } |
| 2234 | __func__, k, use_one_row_for_frame_flip[k]){do { } while(0); }; |
| 2235 | dml_print("DML::%s: k=%d, UseMALLForPStateChange = %d\n",{do { } while(0); } |
| 2236 | __func__, k, UseMALLForPStateChange[k]){do { } while(0); }; |
| 2237 | dml_print("DML::%s: k=%d, dpte_row_height_luma = %d\n", __func__, k, dpte_row_height_luma[k]){do { } while(0); }; |
| 2238 | dml_print("DML::%s: k=%d, dpte_row_width_luma_ub = %d\n",{do { } while(0); } |
| 2239 | __func__, k, dpte_row_width_luma_ub[k]){do { } while(0); }; |
| 2240 | dml_print("DML::%s: k=%d, PixelPTEBytesPerRowY = %d\n", __func__, k, PixelPTEBytesPerRowY[k]){do { } while(0); }; |
| 2241 | dml_print("DML::%s: k=%d, dpte_row_height_chroma = %d\n",{do { } while(0); } |
| 2242 | __func__, k, dpte_row_height_chroma[k]){do { } while(0); }; |
| 2243 | dml_print("DML::%s: k=%d, dpte_row_width_chroma_ub = %d\n",{do { } while(0); } |
| 2244 | __func__, k, dpte_row_width_chroma_ub[k]){do { } while(0); }; |
| 2245 | dml_print("DML::%s: k=%d, PixelPTEBytesPerRowC = %d\n", __func__, k, PixelPTEBytesPerRowC[k]){do { } while(0); }; |
| 2246 | dml_print("DML::%s: k=%d, PixelPTEBytesPerRow = %d\n", __func__, k, PixelPTEBytesPerRow[k]){do { } while(0); }; |
| 2247 | dml_print("DML::%s: k=%d, PTEBufferSizeNotExceeded = %d\n",{do { } while(0); } |
| 2248 | __func__, k, PTEBufferSizeNotExceeded[k]){do { } while(0); }; |
| 2249 | dml_print("DML::%s: k=%d, PTE_BUFFER_MODE = %d\n", __func__, k, PTE_BUFFER_MODE[k]){do { } while(0); }; |
| 2250 | dml_print("DML::%s: k=%d, BIGK_FRAGMENT_SIZE = %d\n", __func__, k, BIGK_FRAGMENT_SIZE[k]){do { } while(0); }; |
| 2251 | #endif |
| 2252 | } |
| 2253 | } // CalculateVMRowAndSwath |
| 2254 | |
| 2255 | unsigned int dml32_CalculateVMAndRowBytes( |
| 2256 | bool_Bool ViewportStationary, |
| 2257 | bool_Bool DCCEnable, |
| 2258 | unsigned int NumberOfDPPs, |
| 2259 | unsigned int BlockHeight256Bytes, |
| 2260 | unsigned int BlockWidth256Bytes, |
| 2261 | enum source_format_class SourcePixelFormat, |
| 2262 | unsigned int SurfaceTiling, |
| 2263 | unsigned int BytePerPixel, |
| 2264 | enum dm_rotation_angle SourceRotation, |
| 2265 | double SwathWidth, |
| 2266 | unsigned int ViewportHeight, |
| 2267 | unsigned int ViewportXStart, |
| 2268 | unsigned int ViewportYStart, |
| 2269 | bool_Bool GPUVMEnable, |
| 2270 | bool_Bool HostVMEnable, |
| 2271 | unsigned int HostVMMaxNonCachedPageTableLevels, |
| 2272 | unsigned int GPUVMMaxPageTableLevels, |
| 2273 | unsigned int GPUVMMinPageSizeKBytes, |
| 2274 | unsigned int HostVMMinPageSize, |
| 2275 | unsigned int PTEBufferSizeInRequests, |
| 2276 | unsigned int Pitch, |
| 2277 | unsigned int DCCMetaPitch, |
| 2278 | unsigned int MacroTileWidth, |
| 2279 | unsigned int MacroTileHeight, |
| 2280 | |
| 2281 | /* Output */ |
| 2282 | unsigned int *MetaRowByte, |
| 2283 | unsigned int *PixelPTEBytesPerRow, |
| 2284 | unsigned int *dpte_row_width_ub, |
| 2285 | unsigned int *dpte_row_height, |
| 2286 | unsigned int *dpte_row_height_linear, |
| 2287 | unsigned int *PixelPTEBytesPerRow_one_row_per_frame, |
| 2288 | unsigned int *dpte_row_width_ub_one_row_per_frame, |
| 2289 | unsigned int *dpte_row_height_one_row_per_frame, |
| 2290 | unsigned int *MetaRequestWidth, |
| 2291 | unsigned int *MetaRequestHeight, |
| 2292 | unsigned int *meta_row_width, |
| 2293 | unsigned int *meta_row_height, |
| 2294 | unsigned int *PixelPTEReqWidth, |
| 2295 | unsigned int *PixelPTEReqHeight, |
| 2296 | unsigned int *PTERequestSize, |
| 2297 | unsigned int *DPDE0BytesFrame, |
| 2298 | unsigned int *MetaPTEBytesFrame) |
| 2299 | { |
| 2300 | unsigned int MPDEBytesFrame; |
| 2301 | unsigned int DCCMetaSurfaceBytes; |
| 2302 | unsigned int ExtraDPDEBytesFrame; |
| 2303 | unsigned int PDEAndMetaPTEBytesFrame; |
| 2304 | unsigned int HostVMDynamicLevels = 0; |
| 2305 | unsigned int MacroTileSizeBytes; |
| 2306 | unsigned int vp_height_meta_ub; |
| 2307 | unsigned int vp_height_dpte_ub; |
| 2308 | unsigned int PixelPTEReqWidth_linear = 0; // VBA_DELTA. VBA doesn't calculate this |
| 2309 | |
| 2310 | if (GPUVMEnable == true1 && HostVMEnable == true1) { |
| 2311 | if (HostVMMinPageSize < 2048) |
| 2312 | HostVMDynamicLevels = HostVMMaxNonCachedPageTableLevels; |
| 2313 | else if (HostVMMinPageSize >= 2048 && HostVMMinPageSize < 1048576) |
| 2314 | HostVMDynamicLevels = dml_max(0, (int) HostVMMaxNonCachedPageTableLevels - 1); |
| 2315 | else |
| 2316 | HostVMDynamicLevels = dml_max(0, (int) HostVMMaxNonCachedPageTableLevels - 2); |
| 2317 | } |
| 2318 | |
| 2319 | *MetaRequestHeight = 8 * BlockHeight256Bytes; |
| 2320 | *MetaRequestWidth = 8 * BlockWidth256Bytes; |
| 2321 | if (SurfaceTiling == dm_sw_linear) { |
| 2322 | *meta_row_height = 32; |
| 2323 | *meta_row_width = dml_floor(ViewportXStart + SwathWidth + *MetaRequestWidth - 1, *MetaRequestWidth) |
| 2324 | - dml_floor(ViewportXStart, *MetaRequestWidth); |
| 2325 | } else if (!IsVertical(SourceRotation)) { |
| 2326 | *meta_row_height = *MetaRequestHeight; |
| 2327 | if (ViewportStationary && NumberOfDPPs == 1) { |
| 2328 | *meta_row_width = dml_floor(ViewportXStart + SwathWidth + *MetaRequestWidth - 1, |
| 2329 | *MetaRequestWidth) - dml_floor(ViewportXStart, *MetaRequestWidth); |
| 2330 | } else { |
| 2331 | *meta_row_width = dml_ceil(SwathWidth - 1, *MetaRequestWidth) + *MetaRequestWidth; |
| 2332 | } |
| 2333 | *MetaRowByte = *meta_row_width * *MetaRequestHeight * BytePerPixel / 256.0; |
| 2334 | } else { |
| 2335 | *meta_row_height = *MetaRequestWidth; |
| 2336 | if (ViewportStationary && NumberOfDPPs == 1) { |
| 2337 | *meta_row_width = dml_floor(ViewportYStart + ViewportHeight + *MetaRequestHeight - 1, |
| 2338 | *MetaRequestHeight) - dml_floor(ViewportYStart, *MetaRequestHeight); |
| 2339 | } else { |
| 2340 | *meta_row_width = dml_ceil(SwathWidth - 1, *MetaRequestHeight) + *MetaRequestHeight; |
| 2341 | } |
| 2342 | *MetaRowByte = *meta_row_width * *MetaRequestWidth * BytePerPixel / 256.0; |
| 2343 | } |
| 2344 | |
| 2345 | if (ViewportStationary && (NumberOfDPPs == 1 || !IsVertical(SourceRotation))) { |
| 2346 | vp_height_meta_ub = dml_floor(ViewportYStart + ViewportHeight + 64 * BlockHeight256Bytes - 1, |
| 2347 | 64 * BlockHeight256Bytes) - dml_floor(ViewportYStart, 64 * BlockHeight256Bytes); |
| 2348 | } else if (!IsVertical(SourceRotation)) { |
| 2349 | vp_height_meta_ub = dml_ceil(ViewportHeight - 1, 64 * BlockHeight256Bytes) + 64 * BlockHeight256Bytes; |
| 2350 | } else { |
| 2351 | vp_height_meta_ub = dml_ceil(SwathWidth - 1, 64 * BlockHeight256Bytes) + 64 * BlockHeight256Bytes; |
| 2352 | } |
| 2353 | |
| 2354 | DCCMetaSurfaceBytes = DCCMetaPitch * vp_height_meta_ub * BytePerPixel / 256.0; |
| 2355 | |
| 2356 | if (GPUVMEnable == true1) { |
| 2357 | *MetaPTEBytesFrame = (dml_ceil((double) (DCCMetaSurfaceBytes - 4.0 * 1024.0) / |
| 2358 | (8 * 4.0 * 1024), 1) + 1) * 64; |
| 2359 | MPDEBytesFrame = 128 * (GPUVMMaxPageTableLevels - 1); |
| 2360 | } else { |
| 2361 | *MetaPTEBytesFrame = 0; |
| 2362 | MPDEBytesFrame = 0; |
| 2363 | } |
| 2364 | |
| 2365 | if (DCCEnable != true1) { |
| 2366 | *MetaPTEBytesFrame = 0; |
| 2367 | MPDEBytesFrame = 0; |
| 2368 | *MetaRowByte = 0; |
| 2369 | } |
| 2370 | |
| 2371 | MacroTileSizeBytes = MacroTileWidth * BytePerPixel * MacroTileHeight; |
| 2372 | |
| 2373 | if (GPUVMEnable == true1 && GPUVMMaxPageTableLevels > 1) { |
| 2374 | if (ViewportStationary && (NumberOfDPPs == 1 || !IsVertical(SourceRotation))) { |
| 2375 | vp_height_dpte_ub = dml_floor(ViewportYStart + ViewportHeight + |
| 2376 | MacroTileHeight - 1, MacroTileHeight) - |
| 2377 | dml_floor(ViewportYStart, MacroTileHeight); |
| 2378 | } else if (!IsVertical(SourceRotation)) { |
| 2379 | vp_height_dpte_ub = dml_ceil(ViewportHeight - 1, MacroTileHeight) + MacroTileHeight; |
| 2380 | } else { |
| 2381 | vp_height_dpte_ub = dml_ceil(SwathWidth - 1, MacroTileHeight) + MacroTileHeight; |
| 2382 | } |
| 2383 | *DPDE0BytesFrame = 64 * (dml_ceil((Pitch * vp_height_dpte_ub * BytePerPixel - MacroTileSizeBytes) / |
| 2384 | (8 * 2097152), 1) + 1); |
| 2385 | ExtraDPDEBytesFrame = 128 * (GPUVMMaxPageTableLevels - 2); |
| 2386 | } else { |
| 2387 | *DPDE0BytesFrame = 0; |
| 2388 | ExtraDPDEBytesFrame = 0; |
| 2389 | vp_height_dpte_ub = 0; |
| 2390 | } |
| 2391 | |
| 2392 | PDEAndMetaPTEBytesFrame = *MetaPTEBytesFrame + MPDEBytesFrame + *DPDE0BytesFrame + ExtraDPDEBytesFrame; |
| 2393 | |
| 2394 | #ifdef __DML_VBA_DEBUG__ |
| 2395 | dml_print("DML::%s: DCCEnable = %d\n", __func__, DCCEnable){do { } while(0); }; |
| 2396 | dml_print("DML::%s: GPUVMEnable = %d\n", __func__, GPUVMEnable){do { } while(0); }; |
| 2397 | dml_print("DML::%s: SwModeLinear = %d\n", __func__, SurfaceTiling == dm_sw_linear){do { } while(0); }; |
| 2398 | dml_print("DML::%s: BytePerPixel = %d\n", __func__, BytePerPixel){do { } while(0); }; |
| 2399 | dml_print("DML::%s: GPUVMMaxPageTableLevels = %d\n", __func__, GPUVMMaxPageTableLevels){do { } while(0); }; |
| 2400 | dml_print("DML::%s: BlockHeight256Bytes = %d\n", __func__, BlockHeight256Bytes){do { } while(0); }; |
| 2401 | dml_print("DML::%s: BlockWidth256Bytes = %d\n", __func__, BlockWidth256Bytes){do { } while(0); }; |
| 2402 | dml_print("DML::%s: MacroTileHeight = %d\n", __func__, MacroTileHeight){do { } while(0); }; |
| 2403 | dml_print("DML::%s: MacroTileWidth = %d\n", __func__, MacroTileWidth){do { } while(0); }; |
| 2404 | dml_print("DML::%s: MetaPTEBytesFrame = %d\n", __func__, *MetaPTEBytesFrame){do { } while(0); }; |
| 2405 | dml_print("DML::%s: MPDEBytesFrame = %d\n", __func__, MPDEBytesFrame){do { } while(0); }; |
| 2406 | dml_print("DML::%s: DPDE0BytesFrame = %d\n", __func__, *DPDE0BytesFrame){do { } while(0); }; |
| 2407 | dml_print("DML::%s: ExtraDPDEBytesFrame= %d\n", __func__, ExtraDPDEBytesFrame){do { } while(0); }; |
| 2408 | dml_print("DML::%s: PDEAndMetaPTEBytesFrame = %d\n", __func__, PDEAndMetaPTEBytesFrame){do { } while(0); }; |
| 2409 | dml_print("DML::%s: ViewportHeight = %d\n", __func__, ViewportHeight){do { } while(0); }; |
| 2410 | dml_print("DML::%s: SwathWidth = %d\n", __func__, SwathWidth){do { } while(0); }; |
| 2411 | dml_print("DML::%s: vp_height_dpte_ub = %d\n", __func__, vp_height_dpte_ub){do { } while(0); }; |
| 2412 | #endif |
| 2413 | |
| 2414 | if (HostVMEnable == true1) |
| 2415 | PDEAndMetaPTEBytesFrame = PDEAndMetaPTEBytesFrame * (1 + 8 * HostVMDynamicLevels); |
| 2416 | |
| 2417 | if (SurfaceTiling == dm_sw_linear) { |
| 2418 | *PixelPTEReqHeight = 1; |
| 2419 | *PixelPTEReqWidth = GPUVMMinPageSizeKBytes * 1024 * 8 / BytePerPixel; |
| 2420 | PixelPTEReqWidth_linear = GPUVMMinPageSizeKBytes * 1024 * 8 / BytePerPixel; |
| 2421 | *PTERequestSize = 64; |
| 2422 | } else if (GPUVMMinPageSizeKBytes == 4) { |
| 2423 | *PixelPTEReqHeight = 16 * BlockHeight256Bytes; |
| 2424 | *PixelPTEReqWidth = 16 * BlockWidth256Bytes; |
| 2425 | *PTERequestSize = 128; |
| 2426 | } else { |
| 2427 | *PixelPTEReqHeight = MacroTileHeight; |
| 2428 | *PixelPTEReqWidth = 8 * 1024 * GPUVMMinPageSizeKBytes / (MacroTileHeight * BytePerPixel); |
| 2429 | *PTERequestSize = 64; |
| 2430 | } |
| 2431 | #ifdef __DML_VBA_DEBUG__ |
| 2432 | dml_print("DML::%s: GPUVMMinPageSizeKBytes = %d\n", __func__, GPUVMMinPageSizeKBytes){do { } while(0); }; |
| 2433 | dml_print("DML::%s: PDEAndMetaPTEBytesFrame = %d (after HostVM factor)\n", __func__, PDEAndMetaPTEBytesFrame){do { } while(0); }; |
| 2434 | dml_print("DML::%s: PixelPTEReqHeight = %d\n", __func__, *PixelPTEReqHeight){do { } while(0); }; |
| 2435 | dml_print("DML::%s: PixelPTEReqWidth = %d\n", __func__, *PixelPTEReqWidth){do { } while(0); }; |
| 2436 | dml_print("DML::%s: PixelPTEReqWidth_linear = %d\n", __func__, PixelPTEReqWidth_linear){do { } while(0); }; |
| 2437 | dml_print("DML::%s: PTERequestSize = %d\n", __func__, *PTERequestSize){do { } while(0); }; |
| 2438 | dml_print("DML::%s: Pitch = %d\n", __func__, Pitch){do { } while(0); }; |
| 2439 | #endif |
| 2440 | |
| 2441 | *dpte_row_height_one_row_per_frame = vp_height_dpte_ub; |
| 2442 | *dpte_row_width_ub_one_row_per_frame = (dml_ceil(((double)Pitch * (double)*dpte_row_height_one_row_per_frame / |
| 2443 | (double) *PixelPTEReqHeight - 1) / (double) *PixelPTEReqWidth, 1) + 1) * |
| 2444 | (double) *PixelPTEReqWidth; |
| 2445 | *PixelPTEBytesPerRow_one_row_per_frame = *dpte_row_width_ub_one_row_per_frame / *PixelPTEReqWidth * |
| 2446 | *PTERequestSize; |
| 2447 | |
| 2448 | if (SurfaceTiling == dm_sw_linear) { |
| 2449 | *dpte_row_height = dml_min(128, 1 << (unsigned int) dml_floor(dml_log2(PTEBufferSizeInRequests * |
| 2450 | *PixelPTEReqWidth / Pitch), 1)); |
| 2451 | #ifdef __DML_VBA_DEBUG__ |
| 2452 | dml_print("DML::%s: dpte_row_height = %d (1)\n", __func__,{do { } while(0); } |
| 2453 | PTEBufferSizeInRequests * *PixelPTEReqWidth / Pitch){do { } while(0); }; |
| 2454 | dml_print("DML::%s: dpte_row_height = %f (2)\n", __func__,{do { } while(0); } |
| 2455 | dml_log2(PTEBufferSizeInRequests * *PixelPTEReqWidth / Pitch)){do { } while(0); }; |
| 2456 | dml_print("DML::%s: dpte_row_height = %f (3)\n", __func__,{do { } while(0); } |
| 2457 | dml_floor(dml_log2(PTEBufferSizeInRequests * *PixelPTEReqWidth / Pitch), 1)){do { } while(0); }; |
| 2458 | dml_print("DML::%s: dpte_row_height = %d (4)\n", __func__,{do { } while(0); } |
| 2459 | 1 << (unsigned int) dml_floor(dml_log2(PTEBufferSizeInRequests *{do { } while(0); } |
| 2460 | *PixelPTEReqWidth / Pitch), 1)){do { } while(0); }; |
| 2461 | dml_print("DML::%s: dpte_row_height = %d\n", __func__, *dpte_row_height){do { } while(0); }; |
| 2462 | #endif |
| 2463 | *dpte_row_width_ub = dml_ceil(((double) Pitch * (double) *dpte_row_height - 1), |
| 2464 | (double) *PixelPTEReqWidth) + *PixelPTEReqWidth; |
| 2465 | *PixelPTEBytesPerRow = *dpte_row_width_ub / (double)*PixelPTEReqWidth * (double)*PTERequestSize; |
| 2466 | |
| 2467 | // VBA_DELTA, VBA doesn't have programming value for pte row height linear. |
| 2468 | *dpte_row_height_linear = 1 << (unsigned int) dml_floor(dml_log2(PTEBufferSizeInRequests * |
| 2469 | PixelPTEReqWidth_linear / Pitch), 1); |
| 2470 | if (*dpte_row_height_linear > 128) |
| 2471 | *dpte_row_height_linear = 128; |
| 2472 | |
| 2473 | } else if (!IsVertical(SourceRotation)) { |
| 2474 | *dpte_row_height = *PixelPTEReqHeight; |
| 2475 | |
| 2476 | if (GPUVMMinPageSizeKBytes > 64) { |
| 2477 | *dpte_row_width_ub = (dml_ceil((Pitch * *dpte_row_height / *PixelPTEReqHeight - 1) / |
| 2478 | *PixelPTEReqWidth, 1) + 1) * *PixelPTEReqWidth; |
| 2479 | } else if (ViewportStationary && (NumberOfDPPs == 1)) { |
| 2480 | *dpte_row_width_ub = dml_floor(ViewportXStart + SwathWidth + |
| 2481 | *PixelPTEReqWidth - 1, *PixelPTEReqWidth) - |
| 2482 | dml_floor(ViewportXStart, *PixelPTEReqWidth); |
| 2483 | } else { |
| 2484 | *dpte_row_width_ub = (dml_ceil((SwathWidth - 1) / *PixelPTEReqWidth, 1) + 1) * |
| 2485 | *PixelPTEReqWidth; |
| 2486 | } |
| 2487 | |
| 2488 | *PixelPTEBytesPerRow = *dpte_row_width_ub / *PixelPTEReqWidth * *PTERequestSize; |
| 2489 | } else { |
| 2490 | *dpte_row_height = dml_min(*PixelPTEReqWidth, MacroTileWidth); |
| 2491 | |
| 2492 | if (ViewportStationary && (NumberOfDPPs == 1)) { |
| 2493 | *dpte_row_width_ub = dml_floor(ViewportYStart + ViewportHeight + *PixelPTEReqHeight - 1, |
| 2494 | *PixelPTEReqHeight) - dml_floor(ViewportYStart, *PixelPTEReqHeight); |
| 2495 | } else { |
| 2496 | *dpte_row_width_ub = (dml_ceil((SwathWidth - 1) / *PixelPTEReqHeight, 1) + 1) |
| 2497 | * *PixelPTEReqHeight; |
| 2498 | } |
| 2499 | |
| 2500 | *PixelPTEBytesPerRow = *dpte_row_width_ub / *PixelPTEReqHeight * *PTERequestSize; |
| 2501 | } |
| 2502 | |
| 2503 | if (GPUVMEnable != true1) |
| 2504 | *PixelPTEBytesPerRow = 0; |
| 2505 | if (HostVMEnable == true1) |
| 2506 | *PixelPTEBytesPerRow = *PixelPTEBytesPerRow * (1 + 8 * HostVMDynamicLevels); |
| 2507 | |
| 2508 | #ifdef __DML_VBA_DEBUG__ |
| 2509 | dml_print("DML::%s: GPUVMMinPageSizeKBytes = %d\n", __func__, GPUVMMinPageSizeKBytes){do { } while(0); }; |
| 2510 | dml_print("DML::%s: dpte_row_height = %d\n", __func__, *dpte_row_height){do { } while(0); }; |
| 2511 | dml_print("DML::%s: dpte_row_height_linear = %d\n", __func__, *dpte_row_height_linear){do { } while(0); }; |
| 2512 | dml_print("DML::%s: dpte_row_width_ub = %d\n", __func__, *dpte_row_width_ub){do { } while(0); }; |
| 2513 | dml_print("DML::%s: PixelPTEBytesPerRow = %d\n", __func__, *PixelPTEBytesPerRow){do { } while(0); }; |
| 2514 | dml_print("DML::%s: PTEBufferSizeInRequests = %d\n", __func__, PTEBufferSizeInRequests){do { } while(0); }; |
| 2515 | dml_print("DML::%s: dpte_row_height_one_row_per_frame = %d\n", __func__, *dpte_row_height_one_row_per_frame){do { } while(0); }; |
| 2516 | dml_print("DML::%s: dpte_row_width_ub_one_row_per_frame = %d\n",{do { } while(0); } |
| 2517 | __func__, *dpte_row_width_ub_one_row_per_frame){do { } while(0); }; |
| 2518 | dml_print("DML::%s: PixelPTEBytesPerRow_one_row_per_frame = %d\n",{do { } while(0); } |
| 2519 | __func__, *PixelPTEBytesPerRow_one_row_per_frame){do { } while(0); }; |
| 2520 | dml_print("DML: vm_bytes = meta_pte_bytes_per_frame (per_pipe) = MetaPTEBytesFrame = : %i\n",{do { } while(0); } |
| 2521 | *MetaPTEBytesFrame){do { } while(0); }; |
| 2522 | #endif |
| 2523 | |
| 2524 | return PDEAndMetaPTEBytesFrame; |
| 2525 | } // CalculateVMAndRowBytes |
| 2526 | |
| 2527 | double dml32_CalculatePrefetchSourceLines( |
| 2528 | double VRatio, |
| 2529 | unsigned int VTaps, |
| 2530 | bool_Bool Interlace, |
| 2531 | bool_Bool ProgressiveToInterlaceUnitInOPP, |
| 2532 | unsigned int SwathHeight, |
| 2533 | enum dm_rotation_angle SourceRotation, |
| 2534 | bool_Bool ViewportStationary, |
| 2535 | double SwathWidth, |
| 2536 | unsigned int ViewportHeight, |
| 2537 | unsigned int ViewportXStart, |
| 2538 | unsigned int ViewportYStart, |
| 2539 | |
| 2540 | /* Output */ |
| 2541 | double *VInitPreFill, |
| 2542 | unsigned int *MaxNumSwath) |
| 2543 | { |
| 2544 | |
| 2545 | unsigned int vp_start_rot; |
| 2546 | unsigned int sw0_tmp; |
| 2547 | unsigned int MaxPartialSwath; |
| 2548 | double numLines; |
| 2549 | |
| 2550 | #ifdef __DML_VBA_DEBUG__ |
| 2551 | dml_print("DML::%s: VRatio = %f\n", __func__, VRatio){do { } while(0); }; |
| 2552 | dml_print("DML::%s: VTaps = %d\n", __func__, VTaps){do { } while(0); }; |
| 2553 | dml_print("DML::%s: ViewportXStart = %d\n", __func__, ViewportXStart){do { } while(0); }; |
| 2554 | dml_print("DML::%s: ViewportYStart = %d\n", __func__, ViewportYStart){do { } while(0); }; |
| 2555 | dml_print("DML::%s: ViewportStationary = %d\n", __func__, ViewportStationary){do { } while(0); }; |
| 2556 | dml_print("DML::%s: SwathHeight = %d\n", __func__, SwathHeight){do { } while(0); }; |
| 2557 | #endif |
| 2558 | if (ProgressiveToInterlaceUnitInOPP) |
| 2559 | *VInitPreFill = dml_floor((VRatio + (double) VTaps + 1) / 2.0, 1); |
| 2560 | else |
| 2561 | *VInitPreFill = dml_floor((VRatio + (double) VTaps + 1 + Interlace * 0.5 * VRatio) / 2.0, 1); |
| 2562 | |
| 2563 | if (ViewportStationary) { |
| 2564 | if (SourceRotation == dm_rotation_180 || SourceRotation == dm_rotation_180m) { |
| 2565 | vp_start_rot = SwathHeight - |
| 2566 | (((unsigned int) (ViewportYStart + ViewportHeight - 1) % SwathHeight) + 1); |
| 2567 | } else if (SourceRotation == dm_rotation_270 || SourceRotation == dm_rotation_90m) { |
| 2568 | vp_start_rot = ViewportXStart; |
| 2569 | } else if (SourceRotation == dm_rotation_90 || SourceRotation == dm_rotation_270m) { |
| 2570 | vp_start_rot = SwathHeight - |
| 2571 | (((unsigned int)(ViewportYStart + SwathWidth - 1) % SwathHeight) + 1); |
| 2572 | } else { |
| 2573 | vp_start_rot = ViewportYStart; |
| 2574 | } |
| 2575 | sw0_tmp = SwathHeight - (vp_start_rot % SwathHeight); |
| 2576 | if (sw0_tmp < *VInitPreFill) |
| 2577 | *MaxNumSwath = dml_ceil((*VInitPreFill - sw0_tmp) / SwathHeight, 1) + 1; |
| 2578 | else |
| 2579 | *MaxNumSwath = 1; |
| 2580 | MaxPartialSwath = dml_max(1, (unsigned int) (vp_start_rot + *VInitPreFill - 1) % SwathHeight); |
| 2581 | } else { |
| 2582 | *MaxNumSwath = dml_ceil((*VInitPreFill - 1.0) / SwathHeight, 1) + 1; |
| 2583 | if (*VInitPreFill > 1) |
| 2584 | MaxPartialSwath = dml_max(1, (unsigned int) (*VInitPreFill - 2) % SwathHeight); |
| 2585 | else |
| 2586 | MaxPartialSwath = dml_max(1, (unsigned int) (*VInitPreFill + SwathHeight - 2) % SwathHeight); |
| 2587 | } |
| 2588 | numLines = *MaxNumSwath * SwathHeight + MaxPartialSwath; |
| 2589 | |
| 2590 | #ifdef __DML_VBA_DEBUG__ |
| 2591 | dml_print("DML::%s: vp_start_rot = %d\n", __func__, vp_start_rot){do { } while(0); }; |
| 2592 | dml_print("DML::%s: VInitPreFill = %d\n", __func__, *VInitPreFill){do { } while(0); }; |
| 2593 | dml_print("DML::%s: MaxPartialSwath = %d\n", __func__, MaxPartialSwath){do { } while(0); }; |
| 2594 | dml_print("DML::%s: MaxNumSwath = %d\n", __func__, *MaxNumSwath){do { } while(0); }; |
| 2595 | dml_print("DML::%s: Prefetch source lines = %3.2f\n", __func__, numLines){do { } while(0); }; |
| 2596 | #endif |
| 2597 | return numLines; |
| 2598 | |
| 2599 | } // CalculatePrefetchSourceLines |
| 2600 | |
| 2601 | void dml32_CalculateMALLUseForStaticScreen( |
| 2602 | unsigned int NumberOfActiveSurfaces, |
| 2603 | unsigned int MALLAllocatedForDCNFinal, |
| 2604 | enum dm_use_mall_for_static_screen_mode *UseMALLForStaticScreen, |
| 2605 | unsigned int SurfaceSizeInMALL[], |
| 2606 | bool_Bool one_row_per_frame_fits_in_buffer[], |
| 2607 | |
| 2608 | /* output */ |
| 2609 | bool_Bool UsesMALLForStaticScreen[]) |
| 2610 | { |
| 2611 | unsigned int k; |
| 2612 | unsigned int SurfaceToAddToMALL; |
| 2613 | bool_Bool CanAddAnotherSurfaceToMALL; |
| 2614 | unsigned int TotalSurfaceSizeInMALL; |
| 2615 | |
| 2616 | TotalSurfaceSizeInMALL = 0; |
| 2617 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 2618 | UsesMALLForStaticScreen[k] = (UseMALLForStaticScreen[k] == dm_use_mall_static_screen_enable); |
| 2619 | if (UsesMALLForStaticScreen[k]) |
| 2620 | TotalSurfaceSizeInMALL = TotalSurfaceSizeInMALL + SurfaceSizeInMALL[k]; |
| 2621 | #ifdef __DML_VBA_DEBUG__ |
| 2622 | dml_print("DML::%s: k=%d, UsesMALLForStaticScreen = %d\n", __func__, k, UsesMALLForStaticScreen[k]){do { } while(0); }; |
| 2623 | dml_print("DML::%s: k=%d, TotalSurfaceSizeInMALL = %d\n", __func__, k, TotalSurfaceSizeInMALL){do { } while(0); }; |
| 2624 | #endif |
| 2625 | } |
| 2626 | |
| 2627 | SurfaceToAddToMALL = 0; |
| 2628 | CanAddAnotherSurfaceToMALL = true1; |
| 2629 | while (CanAddAnotherSurfaceToMALL) { |
| 2630 | CanAddAnotherSurfaceToMALL = false0; |
| 2631 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 2632 | if (TotalSurfaceSizeInMALL + SurfaceSizeInMALL[k] <= MALLAllocatedForDCNFinal * 1024 * 1024 && |
| 2633 | !UsesMALLForStaticScreen[k] && |
| 2634 | UseMALLForStaticScreen[k] != dm_use_mall_static_screen_disable && |
| 2635 | one_row_per_frame_fits_in_buffer[k] && |
| 2636 | (!CanAddAnotherSurfaceToMALL || |
| 2637 | SurfaceSizeInMALL[k] < SurfaceSizeInMALL[SurfaceToAddToMALL])) { |
| 2638 | CanAddAnotherSurfaceToMALL = true1; |
| 2639 | SurfaceToAddToMALL = k; |
| 2640 | #ifdef __DML_VBA_DEBUG__ |
| 2641 | dml_print("DML::%s: k=%d, UseMALLForStaticScreen = %d (dis, en, optimize)\n",{do { } while(0); } |
| 2642 | __func__, k, UseMALLForStaticScreen[k]){do { } while(0); }; |
| 2643 | #endif |
| 2644 | } |
| 2645 | } |
| 2646 | if (CanAddAnotherSurfaceToMALL) { |
| 2647 | UsesMALLForStaticScreen[SurfaceToAddToMALL] = true1; |
| 2648 | TotalSurfaceSizeInMALL = TotalSurfaceSizeInMALL + SurfaceSizeInMALL[SurfaceToAddToMALL]; |
| 2649 | |
| 2650 | #ifdef __DML_VBA_DEBUG__ |
| 2651 | dml_print("DML::%s: SurfaceToAddToMALL = %d\n", __func__, SurfaceToAddToMALL){do { } while(0); }; |
| 2652 | dml_print("DML::%s: TotalSurfaceSizeInMALL = %d\n", __func__, TotalSurfaceSizeInMALL){do { } while(0); }; |
| 2653 | #endif |
| 2654 | |
| 2655 | } |
| 2656 | } |
| 2657 | } |
| 2658 | |
| 2659 | void dml32_CalculateRowBandwidth( |
| 2660 | bool_Bool GPUVMEnable, |
| 2661 | enum source_format_class SourcePixelFormat, |
| 2662 | double VRatio, |
| 2663 | double VRatioChroma, |
| 2664 | bool_Bool DCCEnable, |
| 2665 | double LineTime, |
| 2666 | unsigned int MetaRowByteLuma, |
| 2667 | unsigned int MetaRowByteChroma, |
| 2668 | unsigned int meta_row_height_luma, |
| 2669 | unsigned int meta_row_height_chroma, |
| 2670 | unsigned int PixelPTEBytesPerRowLuma, |
| 2671 | unsigned int PixelPTEBytesPerRowChroma, |
| 2672 | unsigned int dpte_row_height_luma, |
| 2673 | unsigned int dpte_row_height_chroma, |
| 2674 | /* Output */ |
| 2675 | double *meta_row_bw, |
| 2676 | double *dpte_row_bw) |
| 2677 | { |
| 2678 | if (DCCEnable != true1) { |
| 2679 | *meta_row_bw = 0; |
| 2680 | } else if (SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10 || SourcePixelFormat == dm_420_12 || |
| 2681 | SourcePixelFormat == dm_rgbe_alpha) { |
| 2682 | *meta_row_bw = VRatio * MetaRowByteLuma / (meta_row_height_luma * LineTime) + VRatioChroma * |
| 2683 | MetaRowByteChroma / (meta_row_height_chroma * LineTime); |
| 2684 | } else { |
| 2685 | *meta_row_bw = VRatio * MetaRowByteLuma / (meta_row_height_luma * LineTime); |
| 2686 | } |
| 2687 | |
| 2688 | if (GPUVMEnable != true1) { |
| 2689 | *dpte_row_bw = 0; |
| 2690 | } else if (SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10 || SourcePixelFormat == dm_420_12 || |
| 2691 | SourcePixelFormat == dm_rgbe_alpha) { |
| 2692 | *dpte_row_bw = VRatio * PixelPTEBytesPerRowLuma / (dpte_row_height_luma * LineTime) + |
| 2693 | VRatioChroma * PixelPTEBytesPerRowChroma / (dpte_row_height_chroma * LineTime); |
| 2694 | } else { |
| 2695 | *dpte_row_bw = VRatio * PixelPTEBytesPerRowLuma / (dpte_row_height_luma * LineTime); |
| 2696 | } |
| 2697 | } |
| 2698 | |
| 2699 | double dml32_CalculateUrgentLatency( |
| 2700 | double UrgentLatencyPixelDataOnly, |
| 2701 | double UrgentLatencyPixelMixedWithVMData, |
| 2702 | double UrgentLatencyVMDataOnly, |
| 2703 | bool_Bool DoUrgentLatencyAdjustment, |
| 2704 | double UrgentLatencyAdjustmentFabricClockComponent, |
| 2705 | double UrgentLatencyAdjustmentFabricClockReference, |
| 2706 | double FabricClock) |
| 2707 | { |
| 2708 | double ret; |
| 2709 | |
| 2710 | ret = dml_max3(UrgentLatencyPixelDataOnly, UrgentLatencyPixelMixedWithVMData, UrgentLatencyVMDataOnly); |
| 2711 | if (DoUrgentLatencyAdjustment == true1) { |
| 2712 | ret = ret + UrgentLatencyAdjustmentFabricClockComponent * |
| 2713 | (UrgentLatencyAdjustmentFabricClockReference / FabricClock - 1); |
| 2714 | } |
| 2715 | return ret; |
| 2716 | } |
| 2717 | |
| 2718 | void dml32_CalculateUrgentBurstFactor( |
| 2719 | enum dm_use_mall_for_pstate_change_mode UseMALLForPStateChange, |
| 2720 | unsigned int swath_width_luma_ub, |
| 2721 | unsigned int swath_width_chroma_ub, |
| 2722 | unsigned int SwathHeightY, |
| 2723 | unsigned int SwathHeightC, |
| 2724 | double LineTime, |
| 2725 | double UrgentLatency, |
| 2726 | double CursorBufferSize, |
| 2727 | unsigned int CursorWidth, |
| 2728 | unsigned int CursorBPP, |
| 2729 | double VRatio, |
| 2730 | double VRatioC, |
| 2731 | double BytePerPixelInDETY, |
| 2732 | double BytePerPixelInDETC, |
| 2733 | unsigned int DETBufferSizeY, |
| 2734 | unsigned int DETBufferSizeC, |
| 2735 | /* Output */ |
| 2736 | double *UrgentBurstFactorCursor, |
| 2737 | double *UrgentBurstFactorLuma, |
| 2738 | double *UrgentBurstFactorChroma, |
| 2739 | bool_Bool *NotEnoughUrgentLatencyHiding) |
| 2740 | { |
| 2741 | double LinesInDETLuma; |
| 2742 | double LinesInDETChroma; |
| 2743 | unsigned int LinesInCursorBuffer; |
| 2744 | double CursorBufferSizeInTime; |
| 2745 | double DETBufferSizeInTimeLuma; |
| 2746 | double DETBufferSizeInTimeChroma; |
| 2747 | |
| 2748 | *NotEnoughUrgentLatencyHiding = 0; |
| 2749 | |
| 2750 | if (CursorWidth > 0) { |
| 2751 | LinesInCursorBuffer = 1 << (unsigned int) dml_floor(dml_log2(CursorBufferSize * 1024.0 / |
| 2752 | (CursorWidth * CursorBPP / 8.0)), 1.0); |
| 2753 | if (VRatio > 0) { |
| 2754 | CursorBufferSizeInTime = LinesInCursorBuffer * LineTime / VRatio; |
| 2755 | if (CursorBufferSizeInTime - UrgentLatency <= 0) { |
| 2756 | *NotEnoughUrgentLatencyHiding = 1; |
| 2757 | *UrgentBurstFactorCursor = 0; |
| 2758 | } else { |
| 2759 | *UrgentBurstFactorCursor = CursorBufferSizeInTime / |
| 2760 | (CursorBufferSizeInTime - UrgentLatency); |
| 2761 | } |
| 2762 | } else { |
| 2763 | *UrgentBurstFactorCursor = 1; |
| 2764 | } |
| 2765 | } |
| 2766 | |
| 2767 | LinesInDETLuma = (UseMALLForPStateChange == dm_use_mall_pstate_change_phantom_pipe ? 1024*1024 : |
| 2768 | DETBufferSizeY) / BytePerPixelInDETY / swath_width_luma_ub; |
| 2769 | |
| 2770 | if (VRatio > 0) { |
| 2771 | DETBufferSizeInTimeLuma = dml_floor(LinesInDETLuma, SwathHeightY) * LineTime / VRatio; |
| 2772 | if (DETBufferSizeInTimeLuma - UrgentLatency <= 0) { |
| 2773 | *NotEnoughUrgentLatencyHiding = 1; |
| 2774 | *UrgentBurstFactorLuma = 0; |
| 2775 | } else { |
| 2776 | *UrgentBurstFactorLuma = DETBufferSizeInTimeLuma / (DETBufferSizeInTimeLuma - UrgentLatency); |
| 2777 | } |
| 2778 | } else { |
| 2779 | *UrgentBurstFactorLuma = 1; |
| 2780 | } |
| 2781 | |
| 2782 | if (BytePerPixelInDETC > 0) { |
| 2783 | LinesInDETChroma = (UseMALLForPStateChange == dm_use_mall_pstate_change_phantom_pipe ? |
| 2784 | 1024 * 1024 : DETBufferSizeC) / BytePerPixelInDETC |
| 2785 | / swath_width_chroma_ub; |
| 2786 | |
| 2787 | if (VRatio > 0) { |
| 2788 | DETBufferSizeInTimeChroma = dml_floor(LinesInDETChroma, SwathHeightC) * LineTime / VRatio; |
| 2789 | if (DETBufferSizeInTimeChroma - UrgentLatency <= 0) { |
| 2790 | *NotEnoughUrgentLatencyHiding = 1; |
| 2791 | *UrgentBurstFactorChroma = 0; |
| 2792 | } else { |
| 2793 | *UrgentBurstFactorChroma = DETBufferSizeInTimeChroma |
| 2794 | / (DETBufferSizeInTimeChroma - UrgentLatency); |
| 2795 | } |
| 2796 | } else { |
| 2797 | *UrgentBurstFactorChroma = 1; |
| 2798 | } |
| 2799 | } |
| 2800 | } // CalculateUrgentBurstFactor |
| 2801 | |
| 2802 | void dml32_CalculateDCFCLKDeepSleep( |
| 2803 | unsigned int NumberOfActiveSurfaces, |
| 2804 | unsigned int BytePerPixelY[], |
| 2805 | unsigned int BytePerPixelC[], |
| 2806 | double VRatio[], |
| 2807 | double VRatioChroma[], |
| 2808 | double SwathWidthY[], |
| 2809 | double SwathWidthC[], |
| 2810 | unsigned int DPPPerSurface[], |
| 2811 | double HRatio[], |
| 2812 | double HRatioChroma[], |
| 2813 | double PixelClock[], |
| 2814 | double PSCL_THROUGHPUT[], |
| 2815 | double PSCL_THROUGHPUT_CHROMA[], |
| 2816 | double Dppclk[], |
| 2817 | double ReadBandwidthLuma[], |
| 2818 | double ReadBandwidthChroma[], |
| 2819 | unsigned int ReturnBusWidth, |
| 2820 | |
| 2821 | /* Output */ |
| 2822 | double *DCFClkDeepSleep) |
| 2823 | { |
| 2824 | unsigned int k; |
| 2825 | double DisplayPipeLineDeliveryTimeLuma; |
| 2826 | double DisplayPipeLineDeliveryTimeChroma; |
| 2827 | double DCFClkDeepSleepPerSurface[DC__NUM_DPP__MAX8]; |
| 2828 | double ReadBandwidth = 0.0; |
| 2829 | |
| 2830 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 2831 | |
| 2832 | if (VRatio[k] <= 1) { |
| 2833 | DisplayPipeLineDeliveryTimeLuma = SwathWidthY[k] * DPPPerSurface[k] / HRatio[k] |
| 2834 | / PixelClock[k]; |
| 2835 | } else { |
| 2836 | DisplayPipeLineDeliveryTimeLuma = SwathWidthY[k] / PSCL_THROUGHPUT[k] / Dppclk[k]; |
| 2837 | } |
| 2838 | if (BytePerPixelC[k] == 0) { |
| 2839 | DisplayPipeLineDeliveryTimeChroma = 0; |
| 2840 | } else { |
| 2841 | if (VRatioChroma[k] <= 1) { |
| 2842 | DisplayPipeLineDeliveryTimeChroma = SwathWidthC[k] * |
| 2843 | DPPPerSurface[k] / HRatioChroma[k] / PixelClock[k]; |
| 2844 | } else { |
| 2845 | DisplayPipeLineDeliveryTimeChroma = SwathWidthC[k] / PSCL_THROUGHPUT_CHROMA[k] |
| 2846 | / Dppclk[k]; |
| 2847 | } |
| 2848 | } |
| 2849 | |
| 2850 | if (BytePerPixelC[k] > 0) { |
| 2851 | DCFClkDeepSleepPerSurface[k] = dml_max(__DML_MIN_DCFCLK_FACTOR__1.15 * SwathWidthY[k] * |
| 2852 | BytePerPixelY[k] / 32.0 / DisplayPipeLineDeliveryTimeLuma, |
| 2853 | __DML_MIN_DCFCLK_FACTOR__1.15 * SwathWidthC[k] * BytePerPixelC[k] / |
| 2854 | 32.0 / DisplayPipeLineDeliveryTimeChroma); |
| 2855 | } else { |
| 2856 | DCFClkDeepSleepPerSurface[k] = __DML_MIN_DCFCLK_FACTOR__1.15 * SwathWidthY[k] * BytePerPixelY[k] / |
| 2857 | 64.0 / DisplayPipeLineDeliveryTimeLuma; |
| 2858 | } |
| 2859 | DCFClkDeepSleepPerSurface[k] = dml_max(DCFClkDeepSleepPerSurface[k], PixelClock[k] / 16); |
| 2860 | |
| 2861 | #ifdef __DML_VBA_DEBUG__ |
| 2862 | dml_print("DML::%s: k=%d, PixelClock = %f\n", __func__, k, PixelClock[k]){do { } while(0); }; |
| 2863 | dml_print("DML::%s: k=%d, DCFClkDeepSleepPerSurface = %f\n", __func__, k, DCFClkDeepSleepPerSurface[k]){do { } while(0); }; |
| 2864 | #endif |
| 2865 | } |
| 2866 | |
| 2867 | for (k = 0; k < NumberOfActiveSurfaces; ++k) |
| 2868 | ReadBandwidth = ReadBandwidth + ReadBandwidthLuma[k] + ReadBandwidthChroma[k]; |
| 2869 | |
| 2870 | *DCFClkDeepSleep = dml_max(8.0, __DML_MIN_DCFCLK_FACTOR__1.15 * ReadBandwidth / (double) ReturnBusWidth); |
| 2871 | |
| 2872 | #ifdef __DML_VBA_DEBUG__ |
| 2873 | dml_print("DML::%s: __DML_MIN_DCFCLK_FACTOR__ = %f\n", __func__, __DML_MIN_DCFCLK_FACTOR__){do { } while(0); }; |
| 2874 | dml_print("DML::%s: ReadBandwidth = %f\n", __func__, ReadBandwidth){do { } while(0); }; |
| 2875 | dml_print("DML::%s: ReturnBusWidth = %d\n", __func__, ReturnBusWidth){do { } while(0); }; |
| 2876 | dml_print("DML::%s: DCFClkDeepSleep = %f\n", __func__, *DCFClkDeepSleep){do { } while(0); }; |
| 2877 | #endif |
| 2878 | |
| 2879 | for (k = 0; k < NumberOfActiveSurfaces; ++k) |
| 2880 | *DCFClkDeepSleep = dml_max(*DCFClkDeepSleep, DCFClkDeepSleepPerSurface[k]); |
| 2881 | #ifdef __DML_VBA_DEBUG__ |
| 2882 | dml_print("DML::%s: DCFClkDeepSleep = %f (final)\n", __func__, *DCFClkDeepSleep){do { } while(0); }; |
| 2883 | #endif |
| 2884 | } // CalculateDCFCLKDeepSleep |
| 2885 | |
| 2886 | double dml32_CalculateWriteBackDelay( |
| 2887 | enum source_format_class WritebackPixelFormat, |
| 2888 | double WritebackHRatio, |
| 2889 | double WritebackVRatio, |
| 2890 | unsigned int WritebackVTaps, |
| 2891 | unsigned int WritebackDestinationWidth, |
| 2892 | unsigned int WritebackDestinationHeight, |
| 2893 | unsigned int WritebackSourceHeight, |
| 2894 | unsigned int HTotal) |
| 2895 | { |
| 2896 | double CalculateWriteBackDelay; |
| 2897 | double Line_length; |
| 2898 | double Output_lines_last_notclamped; |
| 2899 | double WritebackVInit; |
| 2900 | |
| 2901 | WritebackVInit = (WritebackVRatio + WritebackVTaps + 1) / 2; |
| 2902 | Line_length = dml_max((double) WritebackDestinationWidth, |
| 2903 | dml_ceil((double)WritebackDestinationWidth / 6.0, 1.0) * WritebackVTaps); |
| 2904 | Output_lines_last_notclamped = WritebackDestinationHeight - 1 - |
| 2905 | dml_ceil(((double)WritebackSourceHeight - |
| 2906 | (double) WritebackVInit) / (double)WritebackVRatio, 1.0); |
| 2907 | if (Output_lines_last_notclamped < 0) { |
| 2908 | CalculateWriteBackDelay = 0; |
| 2909 | } else { |
| 2910 | CalculateWriteBackDelay = Output_lines_last_notclamped * Line_length + |
| 2911 | (HTotal - WritebackDestinationWidth) + 80; |
| 2912 | } |
| 2913 | return CalculateWriteBackDelay; |
| 2914 | } |
| 2915 | |
| 2916 | void dml32_UseMinimumDCFCLK( |
| 2917 | enum dm_use_mall_for_pstate_change_mode UseMALLForPStateChange[], |
| 2918 | bool_Bool DRRDisplay[], |
| 2919 | bool_Bool SynchronizeDRRDisplaysForUCLKPStateChangeFinal, |
| 2920 | unsigned int MaxInterDCNTileRepeaters, |
| 2921 | unsigned int MaxPrefetchMode, |
| 2922 | double DRAMClockChangeLatencyFinal, |
| 2923 | double FCLKChangeLatency, |
| 2924 | double SREnterPlusExitTime, |
| 2925 | unsigned int ReturnBusWidth, |
| 2926 | unsigned int RoundTripPingLatencyCycles, |
| 2927 | unsigned int ReorderingBytes, |
| 2928 | unsigned int PixelChunkSizeInKByte, |
| 2929 | unsigned int MetaChunkSize, |
| 2930 | bool_Bool GPUVMEnable, |
| 2931 | unsigned int GPUVMMaxPageTableLevels, |
| 2932 | bool_Bool HostVMEnable, |
| 2933 | unsigned int NumberOfActiveSurfaces, |
| 2934 | double HostVMMinPageSize, |
| 2935 | unsigned int HostVMMaxNonCachedPageTableLevels, |
| 2936 | bool_Bool DynamicMetadataVMEnabled, |
| 2937 | bool_Bool ImmediateFlipRequirement, |
| 2938 | bool_Bool ProgressiveToInterlaceUnitInOPP, |
| 2939 | double MaxAveragePercentOfIdealSDPPortBWDisplayCanUseInNormalSystemOperation, |
| 2940 | double PercentOfIdealSDPPortBWReceivedAfterUrgLatency, |
| 2941 | unsigned int VTotal[], |
| 2942 | unsigned int VActive[], |
| 2943 | unsigned int DynamicMetadataTransmittedBytes[], |
| 2944 | unsigned int DynamicMetadataLinesBeforeActiveRequired[], |
| 2945 | bool_Bool Interlace[], |
| 2946 | double RequiredDPPCLKPerSurface[][2][DC__NUM_DPP__MAX8], |
| 2947 | double RequiredDISPCLK[][2], |
| 2948 | double UrgLatency[], |
| 2949 | unsigned int NoOfDPP[][2][DC__NUM_DPP__MAX8], |
| 2950 | double ProjectedDCFClkDeepSleep[][2], |
| 2951 | double MaximumVStartup[][2][DC__NUM_DPP__MAX8], |
| 2952 | unsigned int TotalNumberOfActiveDPP[][2], |
| 2953 | unsigned int TotalNumberOfDCCActiveDPP[][2], |
| 2954 | unsigned int dpte_group_bytes[], |
| 2955 | double PrefetchLinesY[][2][DC__NUM_DPP__MAX8], |
| 2956 | double PrefetchLinesC[][2][DC__NUM_DPP__MAX8], |
| 2957 | unsigned int swath_width_luma_ub_all_states[][2][DC__NUM_DPP__MAX8], |
| 2958 | unsigned int swath_width_chroma_ub_all_states[][2][DC__NUM_DPP__MAX8], |
| 2959 | unsigned int BytePerPixelY[], |
| 2960 | unsigned int BytePerPixelC[], |
| 2961 | unsigned int HTotal[], |
| 2962 | double PixelClock[], |
| 2963 | double PDEAndMetaPTEBytesPerFrame[][2][DC__NUM_DPP__MAX8], |
| 2964 | double DPTEBytesPerRow[][2][DC__NUM_DPP__MAX8], |
| 2965 | double MetaRowBytes[][2][DC__NUM_DPP__MAX8], |
| 2966 | bool_Bool DynamicMetadataEnable[], |
| 2967 | double ReadBandwidthLuma[], |
| 2968 | double ReadBandwidthChroma[], |
| 2969 | double DCFCLKPerState[], |
| 2970 | /* Output */ |
| 2971 | double DCFCLKState[][2]) |
| 2972 | { |
| 2973 | unsigned int i, j, k; |
| 2974 | unsigned int dummy1; |
| 2975 | double dummy2, dummy3; |
| 2976 | double NormalEfficiency; |
| 2977 | double TotalMaxPrefetchFlipDPTERowBandwidth[DC__VOLTAGE_STATES20][2]; |
| 2978 | |
| 2979 | NormalEfficiency = PercentOfIdealSDPPortBWReceivedAfterUrgLatency / 100.0; |
| 2980 | for (i = 0; i < DC__VOLTAGE_STATES20; ++i) { |
| 2981 | for (j = 0; j <= 1; ++j) { |
| 2982 | double PixelDCFCLKCyclesRequiredInPrefetch[DC__NUM_DPP__MAX8]; |
| 2983 | double PrefetchPixelLinesTime[DC__NUM_DPP__MAX8]; |
| 2984 | double DCFCLKRequiredForPeakBandwidthPerSurface[DC__NUM_DPP__MAX8]; |
| 2985 | double DynamicMetadataVMExtraLatency[DC__NUM_DPP__MAX8]; |
| 2986 | double MinimumTWait = 0.0; |
| 2987 | double DPTEBandwidth; |
| 2988 | double DCFCLKRequiredForAverageBandwidth; |
| 2989 | unsigned int ExtraLatencyBytes; |
| 2990 | double ExtraLatencyCycles; |
| 2991 | double DCFCLKRequiredForPeakBandwidth; |
| 2992 | unsigned int NoOfDPPState[DC__NUM_DPP__MAX8]; |
| 2993 | double MinimumTvmPlus2Tr0; |
| 2994 | |
| 2995 | TotalMaxPrefetchFlipDPTERowBandwidth[i][j] = 0; |
| 2996 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 2997 | TotalMaxPrefetchFlipDPTERowBandwidth[i][j] = TotalMaxPrefetchFlipDPTERowBandwidth[i][j] |
| 2998 | + NoOfDPP[i][j][k] * DPTEBytesPerRow[i][j][k] |
| 2999 | / (15.75 * HTotal[k] / PixelClock[k]); |
| 3000 | } |
| 3001 | |
| 3002 | for (k = 0; k <= NumberOfActiveSurfaces - 1; ++k) |
| 3003 | NoOfDPPState[k] = NoOfDPP[i][j][k]; |
| 3004 | |
| 3005 | DPTEBandwidth = TotalMaxPrefetchFlipDPTERowBandwidth[i][j]; |
| 3006 | DCFCLKRequiredForAverageBandwidth = dml_max(ProjectedDCFClkDeepSleep[i][j], DPTEBandwidth / NormalEfficiency / ReturnBusWidth); |
| 3007 | |
| 3008 | ExtraLatencyBytes = dml32_CalculateExtraLatencyBytes(ReorderingBytes, |
| 3009 | TotalNumberOfActiveDPP[i][j], PixelChunkSizeInKByte, |
| 3010 | TotalNumberOfDCCActiveDPP[i][j], MetaChunkSize, GPUVMEnable, HostVMEnable, |
| 3011 | NumberOfActiveSurfaces, NoOfDPPState, dpte_group_bytes, 1, HostVMMinPageSize, |
| 3012 | HostVMMaxNonCachedPageTableLevels); |
| 3013 | ExtraLatencyCycles = RoundTripPingLatencyCycles + __DML_ARB_TO_RET_DELAY__7 + 95 |
| 3014 | + ExtraLatencyBytes / NormalEfficiency / ReturnBusWidth; |
| 3015 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 3016 | double DCFCLKCyclesRequiredInPrefetch; |
| 3017 | double PrefetchTime; |
| 3018 | |
| 3019 | PixelDCFCLKCyclesRequiredInPrefetch[k] = (PrefetchLinesY[i][j][k] |
| 3020 | * swath_width_luma_ub_all_states[i][j][k] * BytePerPixelY[k] |
| 3021 | + PrefetchLinesC[i][j][k] * swath_width_chroma_ub_all_states[i][j][k] |
| 3022 | * BytePerPixelC[k]) / NormalEfficiency |
| 3023 | / ReturnBusWidth; |
| 3024 | DCFCLKCyclesRequiredInPrefetch = 2 * ExtraLatencyCycles / NoOfDPPState[k] |
| 3025 | + PDEAndMetaPTEBytesPerFrame[i][j][k] / NormalEfficiency |
| 3026 | / NormalEfficiency / ReturnBusWidth |
| 3027 | * (GPUVMMaxPageTableLevels > 2 ? 1 : 0) |
| 3028 | + 2 * DPTEBytesPerRow[i][j][k] / NormalEfficiency / NormalEfficiency |
| 3029 | / ReturnBusWidth |
| 3030 | + 2 * MetaRowBytes[i][j][k] / NormalEfficiency / ReturnBusWidth |
| 3031 | + PixelDCFCLKCyclesRequiredInPrefetch[k]; |
| 3032 | PrefetchPixelLinesTime[k] = dml_max(PrefetchLinesY[i][j][k], PrefetchLinesC[i][j][k]) |
| 3033 | * HTotal[k] / PixelClock[k]; |
| 3034 | DynamicMetadataVMExtraLatency[k] = (GPUVMEnable == true1 && |
| 3035 | DynamicMetadataEnable[k] == true1 && DynamicMetadataVMEnabled == true1) ? |
| 3036 | UrgLatency[i] * GPUVMMaxPageTableLevels * |
| 3037 | (HostVMEnable == true1 ? HostVMMaxNonCachedPageTableLevels + 1 : 1) : 0; |
| 3038 | |
| 3039 | MinimumTWait = dml32_CalculateTWait(MaxPrefetchMode, |
| 3040 | UseMALLForPStateChange[k], |
| 3041 | SynchronizeDRRDisplaysForUCLKPStateChangeFinal, |
| 3042 | DRRDisplay[k], |
| 3043 | DRAMClockChangeLatencyFinal, |
| 3044 | FCLKChangeLatency, |
| 3045 | UrgLatency[i], |
| 3046 | SREnterPlusExitTime); |
| 3047 | |
| 3048 | PrefetchTime = (MaximumVStartup[i][j][k] - 1) * HTotal[k] / PixelClock[k] - |
| 3049 | MinimumTWait - UrgLatency[i] * |
| 3050 | ((GPUVMMaxPageTableLevels <= 2 ? GPUVMMaxPageTableLevels : |
| 3051 | GPUVMMaxPageTableLevels - 2) * (HostVMEnable == true1 ? |
| 3052 | HostVMMaxNonCachedPageTableLevels + 1 : 1) - 1) - |
| 3053 | DynamicMetadataVMExtraLatency[k]; |
| 3054 | |
| 3055 | if (PrefetchTime > 0) { |
| 3056 | double ExpectedVRatioPrefetch; |
| 3057 | |
| 3058 | ExpectedVRatioPrefetch = PrefetchPixelLinesTime[k] / (PrefetchTime * |
| 3059 | PixelDCFCLKCyclesRequiredInPrefetch[k] / |
| 3060 | DCFCLKCyclesRequiredInPrefetch); |
| 3061 | DCFCLKRequiredForPeakBandwidthPerSurface[k] = NoOfDPPState[k] * |
| 3062 | PixelDCFCLKCyclesRequiredInPrefetch[k] / |
| 3063 | PrefetchPixelLinesTime[k] * |
| 3064 | dml_max(1.0, ExpectedVRatioPrefetch) * |
| 3065 | dml_max(1.0, ExpectedVRatioPrefetch / 4); |
| 3066 | if (HostVMEnable == true1 || ImmediateFlipRequirement == true1) { |
| 3067 | DCFCLKRequiredForPeakBandwidthPerSurface[k] = |
| 3068 | DCFCLKRequiredForPeakBandwidthPerSurface[k] + |
| 3069 | NoOfDPPState[k] * DPTEBandwidth / NormalEfficiency / |
| 3070 | NormalEfficiency / ReturnBusWidth; |
| 3071 | } |
| 3072 | } else { |
| 3073 | DCFCLKRequiredForPeakBandwidthPerSurface[k] = DCFCLKPerState[i]; |
| 3074 | } |
| 3075 | if (DynamicMetadataEnable[k] == true1) { |
| 3076 | double TSetupPipe; |
| 3077 | double TdmbfPipe; |
| 3078 | double TdmsksPipe; |
| 3079 | double TdmecPipe; |
| 3080 | double AllowedTimeForUrgentExtraLatency; |
| 3081 | |
| 3082 | dml32_CalculateVUpdateAndDynamicMetadataParameters( |
| 3083 | MaxInterDCNTileRepeaters, |
| 3084 | RequiredDPPCLKPerSurface[i][j][k], |
| 3085 | RequiredDISPCLK[i][j], |
| 3086 | ProjectedDCFClkDeepSleep[i][j], |
| 3087 | PixelClock[k], |
| 3088 | HTotal[k], |
| 3089 | VTotal[k] - VActive[k], |
| 3090 | DynamicMetadataTransmittedBytes[k], |
| 3091 | DynamicMetadataLinesBeforeActiveRequired[k], |
| 3092 | Interlace[k], |
| 3093 | ProgressiveToInterlaceUnitInOPP, |
| 3094 | |
| 3095 | /* output */ |
| 3096 | &TSetupPipe, |
| 3097 | &TdmbfPipe, |
| 3098 | &TdmecPipe, |
| 3099 | &TdmsksPipe, |
| 3100 | &dummy1, |
| 3101 | &dummy2, |
| 3102 | &dummy3); |
| 3103 | AllowedTimeForUrgentExtraLatency = MaximumVStartup[i][j][k] * HTotal[k] / |
| 3104 | PixelClock[k] - MinimumTWait - TSetupPipe - TdmbfPipe - |
| 3105 | TdmecPipe - TdmsksPipe - DynamicMetadataVMExtraLatency[k]; |
| 3106 | if (AllowedTimeForUrgentExtraLatency > 0) |
| 3107 | DCFCLKRequiredForPeakBandwidthPerSurface[k] = |
| 3108 | dml_max(DCFCLKRequiredForPeakBandwidthPerSurface[k], |
| 3109 | ExtraLatencyCycles / AllowedTimeForUrgentExtraLatency); |
| 3110 | else |
| 3111 | DCFCLKRequiredForPeakBandwidthPerSurface[k] = DCFCLKPerState[i]; |
| 3112 | } |
| 3113 | } |
| 3114 | DCFCLKRequiredForPeakBandwidth = 0; |
| 3115 | for (k = 0; k <= NumberOfActiveSurfaces - 1; ++k) { |
| 3116 | DCFCLKRequiredForPeakBandwidth = DCFCLKRequiredForPeakBandwidth + |
| 3117 | DCFCLKRequiredForPeakBandwidthPerSurface[k]; |
| 3118 | } |
| 3119 | MinimumTvmPlus2Tr0 = UrgLatency[i] * (GPUVMEnable == true1 ? |
| 3120 | (HostVMEnable == true1 ? (GPUVMMaxPageTableLevels + 2) * |
| 3121 | (HostVMMaxNonCachedPageTableLevels + 1) - 1 : GPUVMMaxPageTableLevels + 1) : 0); |
| 3122 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 3123 | double MaximumTvmPlus2Tr0PlusTsw; |
| 3124 | |
| 3125 | MaximumTvmPlus2Tr0PlusTsw = (MaximumVStartup[i][j][k] - 2) * HTotal[k] / |
| 3126 | PixelClock[k] - MinimumTWait - DynamicMetadataVMExtraLatency[k]; |
| 3127 | if (MaximumTvmPlus2Tr0PlusTsw <= MinimumTvmPlus2Tr0 + PrefetchPixelLinesTime[k] / 4) { |
| 3128 | DCFCLKRequiredForPeakBandwidth = DCFCLKPerState[i]; |
| 3129 | } else { |
| 3130 | DCFCLKRequiredForPeakBandwidth = dml_max3(DCFCLKRequiredForPeakBandwidth, |
| 3131 | 2 * ExtraLatencyCycles / (MaximumTvmPlus2Tr0PlusTsw - |
| 3132 | MinimumTvmPlus2Tr0 - |
| 3133 | PrefetchPixelLinesTime[k] / 4), |
| 3134 | (2 * ExtraLatencyCycles + |
| 3135 | PixelDCFCLKCyclesRequiredInPrefetch[k]) / |
| 3136 | (MaximumTvmPlus2Tr0PlusTsw - MinimumTvmPlus2Tr0)); |
| 3137 | } |
| 3138 | } |
| 3139 | DCFCLKState[i][j] = dml_min(DCFCLKPerState[i], 1.05 * |
| 3140 | dml_max(DCFCLKRequiredForAverageBandwidth, DCFCLKRequiredForPeakBandwidth)); |
| 3141 | } |
| 3142 | } |
| 3143 | } |
| 3144 | |
| 3145 | unsigned int dml32_CalculateExtraLatencyBytes(unsigned int ReorderingBytes, |
| 3146 | unsigned int TotalNumberOfActiveDPP, |
| 3147 | unsigned int PixelChunkSizeInKByte, |
| 3148 | unsigned int TotalNumberOfDCCActiveDPP, |
| 3149 | unsigned int MetaChunkSize, |
| 3150 | bool_Bool GPUVMEnable, |
| 3151 | bool_Bool HostVMEnable, |
| 3152 | unsigned int NumberOfActiveSurfaces, |
| 3153 | unsigned int NumberOfDPP[], |
| 3154 | unsigned int dpte_group_bytes[], |
| 3155 | double HostVMInefficiencyFactor, |
| 3156 | double HostVMMinPageSize, |
| 3157 | unsigned int HostVMMaxNonCachedPageTableLevels) |
| 3158 | { |
| 3159 | unsigned int k; |
| 3160 | double ret; |
| 3161 | unsigned int HostVMDynamicLevels; |
| 3162 | |
| 3163 | if (GPUVMEnable == true1 && HostVMEnable == true1) { |
| 3164 | if (HostVMMinPageSize < 2048) |
| 3165 | HostVMDynamicLevels = HostVMMaxNonCachedPageTableLevels; |
| 3166 | else if (HostVMMinPageSize >= 2048 && HostVMMinPageSize < 1048576) |
| 3167 | HostVMDynamicLevels = dml_max(0, (int) HostVMMaxNonCachedPageTableLevels - 1); |
| 3168 | else |
| 3169 | HostVMDynamicLevels = dml_max(0, (int) HostVMMaxNonCachedPageTableLevels - 2); |
| 3170 | } else { |
| 3171 | HostVMDynamicLevels = 0; |
| 3172 | } |
| 3173 | |
| 3174 | ret = ReorderingBytes + (TotalNumberOfActiveDPP * PixelChunkSizeInKByte + |
| 3175 | TotalNumberOfDCCActiveDPP * MetaChunkSize) * 1024.0; |
| 3176 | |
| 3177 | if (GPUVMEnable == true1) { |
| 3178 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 3179 | ret = ret + NumberOfDPP[k] * dpte_group_bytes[k] * |
| 3180 | (1 + 8 * HostVMDynamicLevels) * HostVMInefficiencyFactor; |
| 3181 | } |
| 3182 | } |
| 3183 | return ret; |
| 3184 | } |
| 3185 | |
| 3186 | void dml32_CalculateVUpdateAndDynamicMetadataParameters( |
| 3187 | unsigned int MaxInterDCNTileRepeaters, |
| 3188 | double Dppclk, |
| 3189 | double Dispclk, |
| 3190 | double DCFClkDeepSleep, |
| 3191 | double PixelClock, |
| 3192 | unsigned int HTotal, |
| 3193 | unsigned int VBlank, |
| 3194 | unsigned int DynamicMetadataTransmittedBytes, |
| 3195 | unsigned int DynamicMetadataLinesBeforeActiveRequired, |
| 3196 | unsigned int InterlaceEnable, |
| 3197 | bool_Bool ProgressiveToInterlaceUnitInOPP, |
| 3198 | |
| 3199 | /* output */ |
| 3200 | double *TSetup, |
| 3201 | double *Tdmbf, |
| 3202 | double *Tdmec, |
| 3203 | double *Tdmsks, |
| 3204 | unsigned int *VUpdateOffsetPix, |
| 3205 | double *VUpdateWidthPix, |
| 3206 | double *VReadyOffsetPix) |
| 3207 | { |
| 3208 | double TotalRepeaterDelayTime; |
| 3209 | |
| 3210 | TotalRepeaterDelayTime = MaxInterDCNTileRepeaters * (2 / Dppclk + 3 / Dispclk); |
| 3211 | *VUpdateWidthPix = |
| 3212 | dml_ceil((14.0 / DCFClkDeepSleep + 12.0 / Dppclk + TotalRepeaterDelayTime) * PixelClock, 1.0); |
| 3213 | *VReadyOffsetPix = dml_ceil(dml_max(150.0 / Dppclk, |
| 3214 | TotalRepeaterDelayTime + 20.0 / DCFClkDeepSleep + 10.0 / Dppclk) * PixelClock, 1.0); |
| 3215 | *VUpdateOffsetPix = dml_ceil(HTotal / 4.0, 1.0); |
| 3216 | *TSetup = (*VUpdateOffsetPix + *VUpdateWidthPix + *VReadyOffsetPix) / PixelClock; |
| 3217 | *Tdmbf = DynamicMetadataTransmittedBytes / 4.0 / Dispclk; |
| 3218 | *Tdmec = HTotal / PixelClock; |
| 3219 | |
| 3220 | if (DynamicMetadataLinesBeforeActiveRequired == 0) |
| 3221 | *Tdmsks = VBlank * HTotal / PixelClock / 2.0; |
| 3222 | else |
| 3223 | *Tdmsks = DynamicMetadataLinesBeforeActiveRequired * HTotal / PixelClock; |
| 3224 | |
| 3225 | if (InterlaceEnable == 1 && ProgressiveToInterlaceUnitInOPP == false0) |
| 3226 | *Tdmsks = *Tdmsks / 2; |
| 3227 | #ifdef __DML_VBA_DEBUG__ |
| 3228 | dml_print("DML::%s: VUpdateWidthPix = %d\n", __func__, *VUpdateWidthPix){do { } while(0); }; |
| 3229 | dml_print("DML::%s: VReadyOffsetPix = %d\n", __func__, *VReadyOffsetPix){do { } while(0); }; |
| 3230 | dml_print("DML::%s: VUpdateOffsetPix = %d\n", __func__, *VUpdateOffsetPix){do { } while(0); }; |
| 3231 | |
| 3232 | dml_print("DML::%s: DynamicMetadataLinesBeforeActiveRequired = %d\n",{do { } while(0); } |
| 3233 | __func__, DynamicMetadataLinesBeforeActiveRequired){do { } while(0); }; |
| 3234 | dml_print("DML::%s: VBlank = %d\n", __func__, VBlank){do { } while(0); }; |
| 3235 | dml_print("DML::%s: HTotal = %d\n", __func__, HTotal){do { } while(0); }; |
| 3236 | dml_print("DML::%s: PixelClock = %f\n", __func__, PixelClock){do { } while(0); }; |
| 3237 | dml_print("DML::%s: Tdmsks = %f\n", __func__, *Tdmsks){do { } while(0); }; |
| 3238 | #endif |
| 3239 | } |
| 3240 | |
| 3241 | double dml32_CalculateTWait( |
| 3242 | unsigned int PrefetchMode, |
| 3243 | enum dm_use_mall_for_pstate_change_mode UseMALLForPStateChange, |
| 3244 | bool_Bool SynchronizeDRRDisplaysForUCLKPStateChangeFinal, |
| 3245 | bool_Bool DRRDisplay, |
| 3246 | double DRAMClockChangeLatency, |
| 3247 | double FCLKChangeLatency, |
| 3248 | double UrgentLatency, |
| 3249 | double SREnterPlusExitTime) |
| 3250 | { |
| 3251 | double TWait = 0.0; |
| 3252 | |
| 3253 | if (PrefetchMode == 0 && |
| 3254 | !(UseMALLForPStateChange == dm_use_mall_pstate_change_full_frame) && |
| 3255 | !(UseMALLForPStateChange == dm_use_mall_pstate_change_sub_viewport) && |
| 3256 | !(UseMALLForPStateChange == dm_use_mall_pstate_change_phantom_pipe) && |
| 3257 | !(SynchronizeDRRDisplaysForUCLKPStateChangeFinal && DRRDisplay)) { |
| 3258 | TWait = dml_max3(DRAMClockChangeLatency + UrgentLatency, SREnterPlusExitTime, UrgentLatency); |
| 3259 | } else if (PrefetchMode <= 1 && !(UseMALLForPStateChange == dm_use_mall_pstate_change_phantom_pipe)) { |
| 3260 | TWait = dml_max3(FCLKChangeLatency + UrgentLatency, SREnterPlusExitTime, UrgentLatency); |
| 3261 | } else if (PrefetchMode <= 2 && !(UseMALLForPStateChange == dm_use_mall_pstate_change_phantom_pipe)) { |
| 3262 | TWait = dml_max(SREnterPlusExitTime, UrgentLatency); |
| 3263 | } else { |
| 3264 | TWait = UrgentLatency; |
| 3265 | } |
| 3266 | |
| 3267 | #ifdef __DML_VBA_DEBUG__ |
| 3268 | dml_print("DML::%s: PrefetchMode = %d\n", __func__, PrefetchMode){do { } while(0); }; |
| 3269 | dml_print("DML::%s: TWait = %f\n", __func__, TWait){do { } while(0); }; |
| 3270 | #endif |
| 3271 | return TWait; |
| 3272 | } // CalculateTWait |
| 3273 | |
| 3274 | // Function: get_return_bw_mbps |
| 3275 | // Megabyte per second |
| 3276 | double dml32_get_return_bw_mbps(const soc_bounding_box_st *soc, |
| 3277 | const int VoltageLevel, |
| 3278 | const bool_Bool HostVMEnable, |
| 3279 | const double DCFCLK, |
| 3280 | const double FabricClock, |
| 3281 | const double DRAMSpeed) |
| 3282 | { |
| 3283 | double ReturnBW = 0.; |
| 3284 | double IdealSDPPortBandwidth = soc->return_bus_width_bytes /*mode_lib->vba.ReturnBusWidth*/ * DCFCLK; |
| 3285 | double IdealFabricBandwidth = FabricClock * soc->fabric_datapath_to_dcn_data_return_bytes; |
| 3286 | double IdealDRAMBandwidth = DRAMSpeed * soc->num_chans * soc->dram_channel_width_bytes; |
| 3287 | double PixelDataOnlyReturnBW = dml_min3(IdealSDPPortBandwidth * soc->pct_ideal_sdp_bw_after_urgent / 100, |
| 3288 | IdealFabricBandwidth * soc->pct_ideal_fabric_bw_after_urgent / 100, |
| 3289 | IdealDRAMBandwidth * (VoltageLevel < 2 ? soc->pct_ideal_dram_bw_after_urgent_strobe : |
| 3290 | soc->pct_ideal_dram_sdp_bw_after_urgent_pixel_only) / 100); |
| 3291 | double PixelMixedWithVMDataReturnBW = dml_min3(IdealSDPPortBandwidth * soc->pct_ideal_sdp_bw_after_urgent / 100, |
| 3292 | IdealFabricBandwidth * soc->pct_ideal_fabric_bw_after_urgent / 100, |
| 3293 | IdealDRAMBandwidth * (VoltageLevel < 2 ? soc->pct_ideal_dram_bw_after_urgent_strobe : |
| 3294 | soc->pct_ideal_dram_sdp_bw_after_urgent_pixel_only) / 100); |
| 3295 | |
| 3296 | if (HostVMEnable != true1) |
| 3297 | ReturnBW = PixelDataOnlyReturnBW; |
| 3298 | else |
| 3299 | ReturnBW = PixelMixedWithVMDataReturnBW; |
| 3300 | |
| 3301 | #ifdef __DML_VBA_DEBUG__ |
| 3302 | dml_print("DML::%s: VoltageLevel = %d\n", __func__, VoltageLevel){do { } while(0); }; |
| 3303 | dml_print("DML::%s: HostVMEnable = %d\n", __func__, HostVMEnable){do { } while(0); }; |
| 3304 | dml_print("DML::%s: DCFCLK = %f\n", __func__, DCFCLK){do { } while(0); }; |
| 3305 | dml_print("DML::%s: FabricClock = %f\n", __func__, FabricClock){do { } while(0); }; |
| 3306 | dml_print("DML::%s: DRAMSpeed = %f\n", __func__, DRAMSpeed){do { } while(0); }; |
| 3307 | dml_print("DML::%s: IdealSDPPortBandwidth = %f\n", __func__, IdealSDPPortBandwidth){do { } while(0); }; |
| 3308 | dml_print("DML::%s: IdealFabricBandwidth = %f\n", __func__, IdealFabricBandwidth){do { } while(0); }; |
| 3309 | dml_print("DML::%s: IdealDRAMBandwidth = %f\n", __func__, IdealDRAMBandwidth){do { } while(0); }; |
| 3310 | dml_print("DML::%s: PixelDataOnlyReturnBW = %f\n", __func__, PixelDataOnlyReturnBW){do { } while(0); }; |
| 3311 | dml_print("DML::%s: PixelMixedWithVMDataReturnBW = %f\n", __func__, PixelMixedWithVMDataReturnBW){do { } while(0); }; |
| 3312 | dml_print("DML::%s: ReturnBW = %f MBps\n", __func__, ReturnBW){do { } while(0); }; |
| 3313 | #endif |
| 3314 | return ReturnBW; |
| 3315 | } |
| 3316 | |
| 3317 | // Function: get_return_bw_mbps_vm_only |
| 3318 | // Megabyte per second |
| 3319 | double dml32_get_return_bw_mbps_vm_only(const soc_bounding_box_st *soc, |
| 3320 | const int VoltageLevel, |
| 3321 | const double DCFCLK, |
| 3322 | const double FabricClock, |
| 3323 | const double DRAMSpeed) |
| 3324 | { |
| 3325 | double VMDataOnlyReturnBW = dml_min3( |
| 3326 | soc->return_bus_width_bytes * DCFCLK * soc->pct_ideal_sdp_bw_after_urgent / 100.0, |
| 3327 | FabricClock * soc->fabric_datapath_to_dcn_data_return_bytes |
| 3328 | * soc->pct_ideal_sdp_bw_after_urgent / 100.0, |
| 3329 | DRAMSpeed * soc->num_chans * soc->dram_channel_width_bytes |
| 3330 | * (VoltageLevel < 2 ? |
| 3331 | soc->pct_ideal_dram_bw_after_urgent_strobe : |
| 3332 | soc->pct_ideal_dram_sdp_bw_after_urgent_vm_only) / 100.0); |
| 3333 | #ifdef __DML_VBA_DEBUG__ |
| 3334 | dml_print("DML::%s: VoltageLevel = %d\n", __func__, VoltageLevel){do { } while(0); }; |
| 3335 | dml_print("DML::%s: DCFCLK = %f\n", __func__, DCFCLK){do { } while(0); }; |
| 3336 | dml_print("DML::%s: FabricClock = %f\n", __func__, FabricClock){do { } while(0); }; |
| 3337 | dml_print("DML::%s: DRAMSpeed = %f\n", __func__, DRAMSpeed){do { } while(0); }; |
| 3338 | dml_print("DML::%s: VMDataOnlyReturnBW = %f\n", __func__, VMDataOnlyReturnBW){do { } while(0); }; |
| 3339 | #endif |
| 3340 | return VMDataOnlyReturnBW; |
| 3341 | } |
| 3342 | |
| 3343 | double dml32_CalculateExtraLatency( |
| 3344 | unsigned int RoundTripPingLatencyCycles, |
| 3345 | unsigned int ReorderingBytes, |
| 3346 | double DCFCLK, |
| 3347 | unsigned int TotalNumberOfActiveDPP, |
| 3348 | unsigned int PixelChunkSizeInKByte, |
| 3349 | unsigned int TotalNumberOfDCCActiveDPP, |
| 3350 | unsigned int MetaChunkSize, |
| 3351 | double ReturnBW, |
| 3352 | bool_Bool GPUVMEnable, |
| 3353 | bool_Bool HostVMEnable, |
| 3354 | unsigned int NumberOfActiveSurfaces, |
| 3355 | unsigned int NumberOfDPP[], |
| 3356 | unsigned int dpte_group_bytes[], |
| 3357 | double HostVMInefficiencyFactor, |
| 3358 | double HostVMMinPageSize, |
| 3359 | unsigned int HostVMMaxNonCachedPageTableLevels) |
| 3360 | { |
| 3361 | double ExtraLatencyBytes; |
| 3362 | double ExtraLatency; |
| 3363 | |
| 3364 | ExtraLatencyBytes = dml32_CalculateExtraLatencyBytes( |
| 3365 | ReorderingBytes, |
| 3366 | TotalNumberOfActiveDPP, |
| 3367 | PixelChunkSizeInKByte, |
| 3368 | TotalNumberOfDCCActiveDPP, |
| 3369 | MetaChunkSize, |
| 3370 | GPUVMEnable, |
| 3371 | HostVMEnable, |
| 3372 | NumberOfActiveSurfaces, |
| 3373 | NumberOfDPP, |
| 3374 | dpte_group_bytes, |
| 3375 | HostVMInefficiencyFactor, |
| 3376 | HostVMMinPageSize, |
| 3377 | HostVMMaxNonCachedPageTableLevels); |
| 3378 | |
| 3379 | ExtraLatency = (RoundTripPingLatencyCycles + __DML_ARB_TO_RET_DELAY__7 + 95) / DCFCLK + ExtraLatencyBytes / ReturnBW; |
| 3380 | |
| 3381 | #ifdef __DML_VBA_DEBUG__ |
| 3382 | dml_print("DML::%s: RoundTripPingLatencyCycles=%d\n", __func__, RoundTripPingLatencyCycles){do { } while(0); }; |
| 3383 | dml_print("DML::%s: DCFCLK=%f\n", __func__, DCFCLK){do { } while(0); }; |
| 3384 | dml_print("DML::%s: ExtraLatencyBytes=%f\n", __func__, ExtraLatencyBytes){do { } while(0); }; |
| 3385 | dml_print("DML::%s: ReturnBW=%f\n", __func__, ReturnBW){do { } while(0); }; |
| 3386 | dml_print("DML::%s: ExtraLatency=%f\n", __func__, ExtraLatency){do { } while(0); }; |
| 3387 | #endif |
| 3388 | |
| 3389 | return ExtraLatency; |
| 3390 | } // CalculateExtraLatency |
| 3391 | |
| 3392 | bool_Bool dml32_CalculatePrefetchSchedule( |
| 3393 | struct vba_vars_st *v, |
| 3394 | unsigned int k, |
| 3395 | double HostVMInefficiencyFactor, |
| 3396 | DmlPipe *myPipe, |
| 3397 | unsigned int DSCDelay, |
| 3398 | unsigned int DPP_RECOUT_WIDTH, |
| 3399 | unsigned int VStartup, |
| 3400 | unsigned int MaxVStartup, |
| 3401 | double UrgentLatency, |
| 3402 | double UrgentExtraLatency, |
| 3403 | double TCalc, |
| 3404 | unsigned int PDEAndMetaPTEBytesFrame, |
| 3405 | unsigned int MetaRowByte, |
| 3406 | unsigned int PixelPTEBytesPerRow, |
| 3407 | double PrefetchSourceLinesY, |
| 3408 | unsigned int SwathWidthY, |
| 3409 | unsigned int VInitPreFillY, |
| 3410 | unsigned int MaxNumSwathY, |
| 3411 | double PrefetchSourceLinesC, |
| 3412 | unsigned int SwathWidthC, |
| 3413 | unsigned int VInitPreFillC, |
| 3414 | unsigned int MaxNumSwathC, |
| 3415 | unsigned int swath_width_luma_ub, |
| 3416 | unsigned int swath_width_chroma_ub, |
| 3417 | unsigned int SwathHeightY, |
| 3418 | unsigned int SwathHeightC, |
| 3419 | double TWait, |
| 3420 | double TPreReq, |
| 3421 | /* Output */ |
| 3422 | double *DSTXAfterScaler, |
| 3423 | double *DSTYAfterScaler, |
| 3424 | double *DestinationLinesForPrefetch, |
| 3425 | double *PrefetchBandwidth, |
| 3426 | double *DestinationLinesToRequestVMInVBlank, |
| 3427 | double *DestinationLinesToRequestRowInVBlank, |
| 3428 | double *VRatioPrefetchY, |
| 3429 | double *VRatioPrefetchC, |
| 3430 | double *RequiredPrefetchPixDataBWLuma, |
| 3431 | double *RequiredPrefetchPixDataBWChroma, |
| 3432 | bool_Bool *NotEnoughTimeForDynamicMetadata, |
| 3433 | double *Tno_bw, |
| 3434 | double *prefetch_vmrow_bw, |
| 3435 | double *Tdmdl_vm, |
| 3436 | double *Tdmdl, |
| 3437 | double *TSetup, |
| 3438 | unsigned int *VUpdateOffsetPix, |
| 3439 | double *VUpdateWidthPix, |
| 3440 | double *VReadyOffsetPix) |
| 3441 | { |
| 3442 | double DPPCLKDelaySubtotalPlusCNVCFormater = v->DPPCLKDelaySubtotal + v->DPPCLKDelayCNVCFormater; |
| 3443 | bool_Bool MyError = false0; |
| 3444 | unsigned int DPPCycles, DISPCLKCycles; |
| 3445 | double DSTTotalPixelsAfterScaler; |
| 3446 | double LineTime; |
| 3447 | double dst_y_prefetch_equ; |
| 3448 | double prefetch_bw_oto; |
| 3449 | double Tvm_oto; |
| 3450 | double Tr0_oto; |
| 3451 | double Tvm_oto_lines; |
| 3452 | double Tr0_oto_lines; |
| 3453 | double dst_y_prefetch_oto; |
| 3454 | double TimeForFetchingMetaPTE = 0; |
| 3455 | double TimeForFetchingRowInVBlank = 0; |
| 3456 | double LinesToRequestPrefetchPixelData = 0; |
| 3457 | double LinesForPrefetchBandwidth = 0; |
| 3458 | unsigned int HostVMDynamicLevelsTrips; |
| 3459 | double trip_to_mem; |
| 3460 | double Tvm_trips; |
| 3461 | double Tr0_trips; |
| 3462 | double Tvm_trips_rounded; |
| 3463 | double Tr0_trips_rounded; |
| 3464 | double Lsw_oto; |
| 3465 | double Tpre_rounded; |
| 3466 | double prefetch_bw_equ; |
| 3467 | double Tvm_equ; |
| 3468 | double Tr0_equ; |
| 3469 | double Tdmbf; |
| 3470 | double Tdmec; |
| 3471 | double Tdmsks; |
| 3472 | double prefetch_sw_bytes; |
| 3473 | double bytes_pp; |
| 3474 | double dep_bytes; |
| 3475 | unsigned int max_vratio_pre = __DML_MAX_VRATIO_PRE__4.0; |
| 3476 | double min_Lsw; |
| 3477 | double Tsw_est1 = 0; |
| 3478 | double Tsw_est3 = 0; |
| 3479 | |
| 3480 | if (v->GPUVMEnable == true1 && v->HostVMEnable == true1) |
| 3481 | HostVMDynamicLevelsTrips = v->HostVMMaxNonCachedPageTableLevels; |
| 3482 | else |
| 3483 | HostVMDynamicLevelsTrips = 0; |
| 3484 | #ifdef __DML_VBA_DEBUG__ |
| 3485 | dml_print("DML::%s: v->GPUVMEnable = %d\n", __func__, v->GPUVMEnable){do { } while(0); }; |
| 3486 | dml_print("DML::%s: v->GPUVMMaxPageTableLevels = %d\n", __func__, v->GPUVMMaxPageTableLevels){do { } while(0); }; |
| 3487 | dml_print("DML::%s: DCCEnable = %d\n", __func__, myPipe->DCCEnable){do { } while(0); }; |
| 3488 | dml_print("DML::%s: v->HostVMEnable=%d HostVMInefficiencyFactor=%f\n",{do { } while(0); } |
| 3489 | __func__, v->HostVMEnable, HostVMInefficiencyFactor){do { } while(0); }; |
| 3490 | #endif |
| 3491 | dml32_CalculateVUpdateAndDynamicMetadataParameters( |
| 3492 | v->MaxInterDCNTileRepeaters, |
| 3493 | myPipe->Dppclk, |
| 3494 | myPipe->Dispclk, |
| 3495 | myPipe->DCFClkDeepSleep, |
| 3496 | myPipe->PixelClock, |
| 3497 | myPipe->HTotal, |
| 3498 | myPipe->VBlank, |
| 3499 | v->DynamicMetadataTransmittedBytes[k], |
| 3500 | v->DynamicMetadataLinesBeforeActiveRequired[k], |
| 3501 | myPipe->InterlaceEnable, |
| 3502 | myPipe->ProgressiveToInterlaceUnitInOPP, |
| 3503 | TSetup, |
| 3504 | |
| 3505 | /* output */ |
| 3506 | &Tdmbf, |
| 3507 | &Tdmec, |
| 3508 | &Tdmsks, |
| 3509 | VUpdateOffsetPix, |
| 3510 | VUpdateWidthPix, |
| 3511 | VReadyOffsetPix); |
| 3512 | |
| 3513 | LineTime = myPipe->HTotal / myPipe->PixelClock; |
| 3514 | trip_to_mem = UrgentLatency; |
| 3515 | Tvm_trips = UrgentExtraLatency + trip_to_mem * (v->GPUVMMaxPageTableLevels * (HostVMDynamicLevelsTrips + 1) - 1); |
| 3516 | |
| 3517 | if (v->DynamicMetadataVMEnabled == true1) |
| 3518 | *Tdmdl = TWait + Tvm_trips + trip_to_mem; |
| 3519 | else |
| 3520 | *Tdmdl = TWait + UrgentExtraLatency; |
| 3521 | |
| 3522 | #ifdef __DML_VBA_ALLOW_DELTA__ |
| 3523 | if (v->DynamicMetadataEnable[k] == false0) |
| 3524 | *Tdmdl = 0.0; |
| 3525 | #endif |
| 3526 | |
| 3527 | if (v->DynamicMetadataEnable[k] == true1) { |
| 3528 | if (VStartup * LineTime < *TSetup + *Tdmdl + Tdmbf + Tdmec + Tdmsks) { |
| 3529 | *NotEnoughTimeForDynamicMetadata = true1; |
| 3530 | #ifdef __DML_VBA_DEBUG__ |
| 3531 | dml_print("DML::%s: Not Enough Time for Dynamic Meta!\n", __func__){do { } while(0); }; |
| 3532 | dml_print("DML::%s: Tdmbf: %fus - time for dmd transfer from dchub to dio output buffer\n",{do { } while(0); } |
| 3533 | __func__, Tdmbf){do { } while(0); }; |
| 3534 | dml_print("DML::%s: Tdmec: %fus - time dio takes to transfer dmd\n", __func__, Tdmec){do { } while(0); }; |
| 3535 | dml_print("DML::%s: Tdmsks: %fus - time before active dmd must complete transmission at dio\n",{do { } while(0); } |
| 3536 | __func__, Tdmsks){do { } while(0); }; |
| 3537 | dml_print("DML::%s: Tdmdl: %fus - time for fabric to become ready and fetch dmd\n",{do { } while(0); } |
| 3538 | __func__, *Tdmdl){do { } while(0); }; |
| 3539 | #endif |
| 3540 | } else { |
| 3541 | *NotEnoughTimeForDynamicMetadata = false0; |
| 3542 | } |
| 3543 | } else { |
| 3544 | *NotEnoughTimeForDynamicMetadata = false0; |
| 3545 | } |
| 3546 | |
| 3547 | *Tdmdl_vm = (v->DynamicMetadataEnable[k] == true1 && v->DynamicMetadataVMEnabled == true1 && |
| 3548 | v->GPUVMEnable == true1 ? TWait + Tvm_trips : 0); |
| 3549 | |
| 3550 | if (myPipe->ScalerEnabled) |
| 3551 | DPPCycles = DPPCLKDelaySubtotalPlusCNVCFormater + v->DPPCLKDelaySCL; |
| 3552 | else |
| 3553 | DPPCycles = DPPCLKDelaySubtotalPlusCNVCFormater + v->DPPCLKDelaySCLLBOnly; |
| 3554 | |
| 3555 | DPPCycles = DPPCycles + myPipe->NumberOfCursors * v->DPPCLKDelayCNVCCursor; |
| 3556 | |
| 3557 | DISPCLKCycles = v->DISPCLKDelaySubtotal; |
| 3558 | |
| 3559 | if (myPipe->Dppclk == 0.0 || myPipe->Dispclk == 0.0) |
| 3560 | return true1; |
| 3561 | |
| 3562 | *DSTXAfterScaler = DPPCycles * myPipe->PixelClock / myPipe->Dppclk + DISPCLKCycles * |
| 3563 | myPipe->PixelClock / myPipe->Dispclk + DSCDelay; |
| 3564 | |
| 3565 | *DSTXAfterScaler = *DSTXAfterScaler + (myPipe->ODMMode != dm_odm_combine_mode_disabled ? 18 : 0) |
| 3566 | + (myPipe->DPPPerSurface - 1) * DPP_RECOUT_WIDTH |
| 3567 | + ((myPipe->ODMMode == dm_odm_split_mode_1to2 || myPipe->ODMMode == dm_odm_mode_mso_1to2) ? |
| 3568 | myPipe->HActive / 2 : 0) |
| 3569 | + ((myPipe->ODMMode == dm_odm_mode_mso_1to4) ? myPipe->HActive * 3 / 4 : 0); |
| 3570 | |
| 3571 | #ifdef __DML_VBA_DEBUG__ |
| 3572 | dml_print("DML::%s: DPPCycles: %d\n", __func__, DPPCycles){do { } while(0); }; |
| 3573 | dml_print("DML::%s: PixelClock: %f\n", __func__, myPipe->PixelClock){do { } while(0); }; |
| 3574 | dml_print("DML::%s: Dppclk: %f\n", __func__, myPipe->Dppclk){do { } while(0); }; |
| 3575 | dml_print("DML::%s: DISPCLKCycles: %d\n", __func__, DISPCLKCycles){do { } while(0); }; |
| 3576 | dml_print("DML::%s: DISPCLK: %f\n", __func__, myPipe->Dispclk){do { } while(0); }; |
| 3577 | dml_print("DML::%s: DSCDelay: %d\n", __func__, DSCDelay){do { } while(0); }; |
| 3578 | dml_print("DML::%s: ODMMode: %d\n", __func__, myPipe->ODMMode){do { } while(0); }; |
| 3579 | dml_print("DML::%s: DPP_RECOUT_WIDTH: %d\n", __func__, DPP_RECOUT_WIDTH){do { } while(0); }; |
| 3580 | dml_print("DML::%s: DSTXAfterScaler: %d\n", __func__, *DSTXAfterScaler){do { } while(0); }; |
| 3581 | #endif |
| 3582 | |
| 3583 | if (v->OutputFormat[k] == dm_420 || (myPipe->InterlaceEnable && myPipe->ProgressiveToInterlaceUnitInOPP)) |
| 3584 | *DSTYAfterScaler = 1; |
| 3585 | else |
| 3586 | *DSTYAfterScaler = 0; |
| 3587 | |
| 3588 | DSTTotalPixelsAfterScaler = *DSTYAfterScaler * myPipe->HTotal + *DSTXAfterScaler; |
| 3589 | *DSTYAfterScaler = dml_floor(DSTTotalPixelsAfterScaler / myPipe->HTotal, 1); |
| 3590 | *DSTXAfterScaler = DSTTotalPixelsAfterScaler - ((double) (*DSTYAfterScaler * myPipe->HTotal)); |
| 3591 | #ifdef __DML_VBA_DEBUG__ |
| 3592 | dml_print("DML::%s: DSTXAfterScaler: %d (final)\n", __func__, *DSTXAfterScaler){do { } while(0); }; |
| 3593 | dml_print("DML::%s: DSTYAfterScaler: %d (final)\n", __func__, *DSTYAfterScaler){do { } while(0); }; |
| 3594 | #endif |
| 3595 | |
| 3596 | MyError = false0; |
| 3597 | |
| 3598 | Tr0_trips = trip_to_mem * (HostVMDynamicLevelsTrips + 1); |
| 3599 | |
| 3600 | if (v->GPUVMEnable == true1) { |
| 3601 | Tvm_trips_rounded = dml_ceil(4.0 * Tvm_trips / LineTime, 1.0) / 4.0 * LineTime; |
| 3602 | Tr0_trips_rounded = dml_ceil(4.0 * Tr0_trips / LineTime, 1.0) / 4.0 * LineTime; |
| 3603 | if (v->GPUVMMaxPageTableLevels >= 3) { |
| 3604 | *Tno_bw = UrgentExtraLatency + trip_to_mem * |
| 3605 | (double) ((v->GPUVMMaxPageTableLevels - 2) * (HostVMDynamicLevelsTrips + 1) - 1); |
| 3606 | } else if (v->GPUVMMaxPageTableLevels == 1 && myPipe->DCCEnable != true1) { |
| 3607 | Tr0_trips_rounded = dml_ceil(4.0 * UrgentExtraLatency / LineTime, 1.0) / |
| 3608 | 4.0 * LineTime; // VBA_ERROR |
| 3609 | *Tno_bw = UrgentExtraLatency; |
| 3610 | } else { |
| 3611 | *Tno_bw = 0; |
| 3612 | } |
| 3613 | } else if (myPipe->DCCEnable == true1) { |
| 3614 | Tvm_trips_rounded = LineTime / 4.0; |
| 3615 | Tr0_trips_rounded = dml_ceil(4.0 * Tr0_trips / LineTime, 1.0) / 4.0 * LineTime; |
| 3616 | *Tno_bw = 0; |
| 3617 | } else { |
| 3618 | Tvm_trips_rounded = LineTime / 4.0; |
| 3619 | Tr0_trips_rounded = LineTime / 2.0; |
| 3620 | *Tno_bw = 0; |
| 3621 | } |
| 3622 | Tvm_trips_rounded = dml_max(Tvm_trips_rounded, LineTime / 4.0); |
| 3623 | Tr0_trips_rounded = dml_max(Tr0_trips_rounded, LineTime / 4.0); |
| 3624 | |
| 3625 | if (myPipe->SourcePixelFormat == dm_420_8 || myPipe->SourcePixelFormat == dm_420_10 |
| 3626 | || myPipe->SourcePixelFormat == dm_420_12) { |
| 3627 | bytes_pp = myPipe->BytePerPixelY + myPipe->BytePerPixelC / 4; |
| 3628 | } else { |
| 3629 | bytes_pp = myPipe->BytePerPixelY + myPipe->BytePerPixelC; |
| 3630 | } |
| 3631 | |
| 3632 | prefetch_sw_bytes = PrefetchSourceLinesY * swath_width_luma_ub * myPipe->BytePerPixelY |
| 3633 | + PrefetchSourceLinesC * swath_width_chroma_ub * myPipe->BytePerPixelC; |
| 3634 | prefetch_bw_oto = dml_max(bytes_pp * myPipe->PixelClock / myPipe->DPPPerSurface, |
| 3635 | prefetch_sw_bytes / (dml_max(PrefetchSourceLinesY, PrefetchSourceLinesC) * LineTime)); |
| 3636 | |
| 3637 | min_Lsw = dml_max(PrefetchSourceLinesY, PrefetchSourceLinesC) / max_vratio_pre; |
| 3638 | min_Lsw = dml_max(min_Lsw, 1.0); |
| 3639 | Lsw_oto = dml_ceil(4.0 * dml_max(prefetch_sw_bytes / prefetch_bw_oto / LineTime, min_Lsw), 1.0) / 4.0; |
| 3640 | |
| 3641 | if (v->GPUVMEnable == true1) { |
| 3642 | Tvm_oto = dml_max3( |
| 3643 | Tvm_trips, |
| 3644 | *Tno_bw + PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor / prefetch_bw_oto, |
| 3645 | LineTime / 4.0); |
| 3646 | } else |
| 3647 | Tvm_oto = LineTime / 4.0; |
| 3648 | |
| 3649 | if ((v->GPUVMEnable == true1 || myPipe->DCCEnable == true1)) { |
| 3650 | Tr0_oto = dml_max4( |
| 3651 | Tr0_trips, |
| 3652 | (MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor) / prefetch_bw_oto, |
| 3653 | (LineTime - Tvm_oto)/2.0, |
| 3654 | LineTime / 4.0); |
| 3655 | #ifdef __DML_VBA_DEBUG__ |
| 3656 | dml_print("DML::%s: Tr0_oto max0 = %f\n", __func__,{do { } while(0); } |
| 3657 | (MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor) / prefetch_bw_oto){do { } while(0); }; |
| 3658 | dml_print("DML::%s: Tr0_oto max1 = %f\n", __func__, Tr0_trips){do { } while(0); }; |
| 3659 | dml_print("DML::%s: Tr0_oto max2 = %f\n", __func__, LineTime - Tvm_oto){do { } while(0); }; |
| 3660 | dml_print("DML::%s: Tr0_oto max3 = %f\n", __func__, LineTime / 4){do { } while(0); }; |
| 3661 | #endif |
| 3662 | } else |
| 3663 | Tr0_oto = (LineTime - Tvm_oto) / 2.0; |
| 3664 | |
| 3665 | Tvm_oto_lines = dml_ceil(4.0 * Tvm_oto / LineTime, 1) / 4.0; |
| 3666 | Tr0_oto_lines = dml_ceil(4.0 * Tr0_oto / LineTime, 1) / 4.0; |
| 3667 | dst_y_prefetch_oto = Tvm_oto_lines + 2 * Tr0_oto_lines + Lsw_oto; |
| 3668 | |
| 3669 | dst_y_prefetch_equ = VStartup - (*TSetup + dml_max(TWait + TCalc, *Tdmdl)) / LineTime - |
| 3670 | (*DSTYAfterScaler + (double) *DSTXAfterScaler / (double) myPipe->HTotal); |
| 3671 | |
| 3672 | dst_y_prefetch_equ = dml_min(dst_y_prefetch_equ, __DML_VBA_MAX_DST_Y_PRE__63.75); |
| 3673 | #ifdef __DML_VBA_DEBUG__ |
| 3674 | dml_print("DML::%s: HTotal = %d\n", __func__, myPipe->HTotal){do { } while(0); }; |
| 3675 | dml_print("DML::%s: min_Lsw = %f\n", __func__, min_Lsw){do { } while(0); }; |
| 3676 | dml_print("DML::%s: *Tno_bw = %f\n", __func__, *Tno_bw){do { } while(0); }; |
| 3677 | dml_print("DML::%s: UrgentExtraLatency = %f\n", __func__, UrgentExtraLatency){do { } while(0); }; |
| 3678 | dml_print("DML::%s: trip_to_mem = %f\n", __func__, trip_to_mem){do { } while(0); }; |
| 3679 | dml_print("DML::%s: BytePerPixelY = %d\n", __func__, myPipe->BytePerPixelY){do { } while(0); }; |
| 3680 | dml_print("DML::%s: PrefetchSourceLinesY = %f\n", __func__, PrefetchSourceLinesY){do { } while(0); }; |
| 3681 | dml_print("DML::%s: swath_width_luma_ub = %d\n", __func__, swath_width_luma_ub){do { } while(0); }; |
| 3682 | dml_print("DML::%s: BytePerPixelC = %d\n", __func__, myPipe->BytePerPixelC){do { } while(0); }; |
| 3683 | dml_print("DML::%s: PrefetchSourceLinesC = %f\n", __func__, PrefetchSourceLinesC){do { } while(0); }; |
| 3684 | dml_print("DML::%s: swath_width_chroma_ub = %d\n", __func__, swath_width_chroma_ub){do { } while(0); }; |
| 3685 | dml_print("DML::%s: prefetch_sw_bytes = %f\n", __func__, prefetch_sw_bytes){do { } while(0); }; |
| 3686 | dml_print("DML::%s: bytes_pp = %f\n", __func__, bytes_pp){do { } while(0); }; |
| 3687 | dml_print("DML::%s: PDEAndMetaPTEBytesFrame = %d\n", __func__, PDEAndMetaPTEBytesFrame){do { } while(0); }; |
| 3688 | dml_print("DML::%s: MetaRowByte = %d\n", __func__, MetaRowByte){do { } while(0); }; |
| 3689 | dml_print("DML::%s: PixelPTEBytesPerRow = %d\n", __func__, PixelPTEBytesPerRow){do { } while(0); }; |
| 3690 | dml_print("DML::%s: HostVMInefficiencyFactor = %f\n", __func__, HostVMInefficiencyFactor){do { } while(0); }; |
| 3691 | dml_print("DML::%s: Tvm_trips = %f\n", __func__, Tvm_trips){do { } while(0); }; |
| 3692 | dml_print("DML::%s: Tr0_trips = %f\n", __func__, Tr0_trips){do { } while(0); }; |
| 3693 | dml_print("DML::%s: prefetch_bw_oto = %f\n", __func__, prefetch_bw_oto){do { } while(0); }; |
| 3694 | dml_print("DML::%s: Tr0_oto = %f\n", __func__, Tr0_oto){do { } while(0); }; |
| 3695 | dml_print("DML::%s: Tvm_oto = %f\n", __func__, Tvm_oto){do { } while(0); }; |
| 3696 | dml_print("DML::%s: Tvm_oto_lines = %f\n", __func__, Tvm_oto_lines){do { } while(0); }; |
| 3697 | dml_print("DML::%s: Tr0_oto_lines = %f\n", __func__, Tr0_oto_lines){do { } while(0); }; |
| 3698 | dml_print("DML::%s: Lsw_oto = %f\n", __func__, Lsw_oto){do { } while(0); }; |
| 3699 | dml_print("DML::%s: dst_y_prefetch_oto = %f\n", __func__, dst_y_prefetch_oto){do { } while(0); }; |
| 3700 | dml_print("DML::%s: dst_y_prefetch_equ = %f\n", __func__, dst_y_prefetch_equ){do { } while(0); }; |
| 3701 | #endif |
| 3702 | |
| 3703 | dst_y_prefetch_equ = dml_floor(4.0 * (dst_y_prefetch_equ + 0.125), 1) / 4.0; |
| 3704 | Tpre_rounded = dst_y_prefetch_equ * LineTime; |
| 3705 | #ifdef __DML_VBA_DEBUG__ |
| 3706 | dml_print("DML::%s: dst_y_prefetch_equ: %f (after round)\n", __func__, dst_y_prefetch_equ){do { } while(0); }; |
| 3707 | dml_print("DML::%s: LineTime: %f\n", __func__, LineTime){do { } while(0); }; |
| 3708 | dml_print("DML::%s: VStartup: %d\n", __func__, VStartup){do { } while(0); }; |
| 3709 | dml_print("DML::%s: Tvstartup: %fus - time between vstartup and first pixel of active\n",{do { } while(0); } |
| 3710 | __func__, VStartup * LineTime){do { } while(0); }; |
| 3711 | dml_print("DML::%s: TSetup: %fus - time from vstartup to vready\n", __func__, *TSetup){do { } while(0); }; |
| 3712 | dml_print("DML::%s: TCalc: %fus - time for calculations in dchub starting at vready\n", __func__, TCalc){do { } while(0); }; |
| 3713 | dml_print("DML::%s: Tdmbf: %fus - time for dmd transfer from dchub to dio output buffer\n", __func__, Tdmbf){do { } while(0); }; |
| 3714 | dml_print("DML::%s: Tdmec: %fus - time dio takes to transfer dmd\n", __func__, Tdmec){do { } while(0); }; |
| 3715 | dml_print("DML::%s: Tdmdl_vm: %fus - time for vm stages of dmd\n", __func__, *Tdmdl_vm){do { } while(0); }; |
| 3716 | dml_print("DML::%s: Tdmdl: %fus - time for fabric to become ready and fetch dmd\n", __func__, *Tdmdl){do { } while(0); }; |
| 3717 | dml_print("DML::%s: DSTYAfterScaler: %d lines - number of lines of pipeline and buffer delay after scaler\n",{do { } while(0); } |
| 3718 | __func__, *DSTYAfterScaler){do { } while(0); }; |
| 3719 | #endif |
| 3720 | dep_bytes = dml_max(PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor, |
| 3721 | MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor); |
| 3722 | |
| 3723 | if (prefetch_sw_bytes < dep_bytes) |
| 3724 | prefetch_sw_bytes = 2 * dep_bytes; |
| 3725 | |
| 3726 | *PrefetchBandwidth = 0; |
| 3727 | *DestinationLinesToRequestVMInVBlank = 0; |
| 3728 | *DestinationLinesToRequestRowInVBlank = 0; |
| 3729 | *VRatioPrefetchY = 0; |
| 3730 | *VRatioPrefetchC = 0; |
| 3731 | *RequiredPrefetchPixDataBWLuma = 0; |
| 3732 | if (dst_y_prefetch_equ > 1 && |
| 3733 | (Tpre_rounded >= TPreReq || dst_y_prefetch_equ == __DML_VBA_MAX_DST_Y_PRE__63.75)) { |
| 3734 | double PrefetchBandwidth1; |
| 3735 | double PrefetchBandwidth2; |
| 3736 | double PrefetchBandwidth3; |
| 3737 | double PrefetchBandwidth4; |
| 3738 | |
| 3739 | if (Tpre_rounded - *Tno_bw > 0) { |
| 3740 | PrefetchBandwidth1 = (PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor + 2 * MetaRowByte |
| 3741 | + 2 * PixelPTEBytesPerRow * HostVMInefficiencyFactor |
| 3742 | + prefetch_sw_bytes) / (Tpre_rounded - *Tno_bw); |
| 3743 | Tsw_est1 = prefetch_sw_bytes / PrefetchBandwidth1; |
| 3744 | } else |
| 3745 | PrefetchBandwidth1 = 0; |
| 3746 | |
| 3747 | if (VStartup == MaxVStartup && (Tsw_est1 / LineTime < min_Lsw) |
| 3748 | && Tpre_rounded - min_Lsw * LineTime - 0.75 * LineTime - *Tno_bw > 0) { |
| 3749 | PrefetchBandwidth1 = (PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor + 2 * MetaRowByte |
| 3750 | + 2 * PixelPTEBytesPerRow * HostVMInefficiencyFactor) |
| 3751 | / (Tpre_rounded - min_Lsw * LineTime - 0.75 * LineTime - *Tno_bw); |
| 3752 | } |
| 3753 | |
| 3754 | if (Tpre_rounded - *Tno_bw - 2 * Tr0_trips_rounded > 0) |
| 3755 | PrefetchBandwidth2 = (PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor + prefetch_sw_bytes) / |
| 3756 | (Tpre_rounded - *Tno_bw - 2 * Tr0_trips_rounded); |
| 3757 | else |
| 3758 | PrefetchBandwidth2 = 0; |
| 3759 | |
| 3760 | if (Tpre_rounded - Tvm_trips_rounded > 0) { |
| 3761 | PrefetchBandwidth3 = (2 * MetaRowByte + 2 * PixelPTEBytesPerRow * HostVMInefficiencyFactor |
| 3762 | + prefetch_sw_bytes) / (Tpre_rounded - Tvm_trips_rounded); |
| 3763 | Tsw_est3 = prefetch_sw_bytes / PrefetchBandwidth3; |
| 3764 | } else |
| 3765 | PrefetchBandwidth3 = 0; |
| 3766 | |
| 3767 | |
| 3768 | if (VStartup == MaxVStartup && |
| 3769 | (Tsw_est3 / LineTime < min_Lsw) && Tpre_rounded - min_Lsw * LineTime - 0.75 * |
| 3770 | LineTime - Tvm_trips_rounded > 0) { |
| 3771 | PrefetchBandwidth3 = (2 * MetaRowByte + 2 * PixelPTEBytesPerRow * HostVMInefficiencyFactor) |
| 3772 | / (Tpre_rounded - min_Lsw * LineTime - 0.75 * LineTime - Tvm_trips_rounded); |
| 3773 | } |
| 3774 | |
| 3775 | if (Tpre_rounded - Tvm_trips_rounded - 2 * Tr0_trips_rounded > 0) { |
| 3776 | PrefetchBandwidth4 = prefetch_sw_bytes / |
| 3777 | (Tpre_rounded - Tvm_trips_rounded - 2 * Tr0_trips_rounded); |
| 3778 | } else { |
| 3779 | PrefetchBandwidth4 = 0; |
| 3780 | } |
| 3781 | |
| 3782 | #ifdef __DML_VBA_DEBUG__ |
| 3783 | dml_print("DML::%s: Tpre_rounded: %f\n", __func__, Tpre_rounded){do { } while(0); }; |
| 3784 | dml_print("DML::%s: Tno_bw: %f\n", __func__, *Tno_bw){do { } while(0); }; |
| 3785 | dml_print("DML::%s: Tvm_trips_rounded: %f\n", __func__, Tvm_trips_rounded){do { } while(0); }; |
| 3786 | dml_print("DML::%s: Tsw_est1: %f\n", __func__, Tsw_est1){do { } while(0); }; |
| 3787 | dml_print("DML::%s: Tsw_est3: %f\n", __func__, Tsw_est3){do { } while(0); }; |
| 3788 | dml_print("DML::%s: PrefetchBandwidth1: %f\n", __func__, PrefetchBandwidth1){do { } while(0); }; |
| 3789 | dml_print("DML::%s: PrefetchBandwidth2: %f\n", __func__, PrefetchBandwidth2){do { } while(0); }; |
| 3790 | dml_print("DML::%s: PrefetchBandwidth3: %f\n", __func__, PrefetchBandwidth3){do { } while(0); }; |
| 3791 | dml_print("DML::%s: PrefetchBandwidth4: %f\n", __func__, PrefetchBandwidth4){do { } while(0); }; |
| 3792 | #endif |
| 3793 | { |
| 3794 | bool_Bool Case1OK; |
| 3795 | bool_Bool Case2OK; |
| 3796 | bool_Bool Case3OK; |
| 3797 | |
| 3798 | if (PrefetchBandwidth1 > 0) { |
| 3799 | if (*Tno_bw + PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor / PrefetchBandwidth1 |
| 3800 | >= Tvm_trips_rounded |
| 3801 | && (MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor) |
| 3802 | / PrefetchBandwidth1 >= Tr0_trips_rounded) { |
| 3803 | Case1OK = true1; |
| 3804 | } else { |
| 3805 | Case1OK = false0; |
| 3806 | } |
| 3807 | } else { |
| 3808 | Case1OK = false0; |
| 3809 | } |
| 3810 | |
| 3811 | if (PrefetchBandwidth2 > 0) { |
| 3812 | if (*Tno_bw + PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor / PrefetchBandwidth2 |
| 3813 | >= Tvm_trips_rounded |
| 3814 | && (MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor) |
| 3815 | / PrefetchBandwidth2 < Tr0_trips_rounded) { |
| 3816 | Case2OK = true1; |
| 3817 | } else { |
| 3818 | Case2OK = false0; |
| 3819 | } |
| 3820 | } else { |
| 3821 | Case2OK = false0; |
| 3822 | } |
| 3823 | |
| 3824 | if (PrefetchBandwidth3 > 0) { |
| 3825 | if (*Tno_bw + PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor / PrefetchBandwidth3 < |
| 3826 | Tvm_trips_rounded && (MetaRowByte + PixelPTEBytesPerRow * |
| 3827 | HostVMInefficiencyFactor) / PrefetchBandwidth3 >= |
| 3828 | Tr0_trips_rounded) { |
| 3829 | Case3OK = true1; |
| 3830 | } else { |
| 3831 | Case3OK = false0; |
| 3832 | } |
| 3833 | } else { |
| 3834 | Case3OK = false0; |
| 3835 | } |
| 3836 | |
| 3837 | if (Case1OK) |
| 3838 | prefetch_bw_equ = PrefetchBandwidth1; |
| 3839 | else if (Case2OK) |
| 3840 | prefetch_bw_equ = PrefetchBandwidth2; |
| 3841 | else if (Case3OK) |
| 3842 | prefetch_bw_equ = PrefetchBandwidth3; |
| 3843 | else |
| 3844 | prefetch_bw_equ = PrefetchBandwidth4; |
| 3845 | |
| 3846 | #ifdef __DML_VBA_DEBUG__ |
| 3847 | dml_print("DML::%s: Case1OK: %d\n", __func__, Case1OK){do { } while(0); }; |
| 3848 | dml_print("DML::%s: Case2OK: %d\n", __func__, Case2OK){do { } while(0); }; |
| 3849 | dml_print("DML::%s: Case3OK: %d\n", __func__, Case3OK){do { } while(0); }; |
| 3850 | dml_print("DML::%s: prefetch_bw_equ: %f\n", __func__, prefetch_bw_equ){do { } while(0); }; |
| 3851 | #endif |
| 3852 | |
| 3853 | if (prefetch_bw_equ > 0) { |
| 3854 | if (v->GPUVMEnable == true1) { |
| 3855 | Tvm_equ = dml_max3(*Tno_bw + PDEAndMetaPTEBytesFrame * |
| 3856 | HostVMInefficiencyFactor / prefetch_bw_equ, |
| 3857 | Tvm_trips, LineTime / 4); |
| 3858 | } else { |
| 3859 | Tvm_equ = LineTime / 4; |
| 3860 | } |
| 3861 | |
| 3862 | if ((v->GPUVMEnable == true1 || myPipe->DCCEnable == true1)) { |
| 3863 | Tr0_equ = dml_max4((MetaRowByte + PixelPTEBytesPerRow * |
| 3864 | HostVMInefficiencyFactor) / prefetch_bw_equ, Tr0_trips, |
| 3865 | (LineTime - Tvm_equ) / 2, LineTime / 4); |
| 3866 | } else { |
| 3867 | Tr0_equ = (LineTime - Tvm_equ) / 2; |
| 3868 | } |
| 3869 | } else { |
| 3870 | Tvm_equ = 0; |
| 3871 | Tr0_equ = 0; |
| 3872 | #ifdef __DML_VBA_DEBUG__ |
| 3873 | dml_print("DML: prefetch_bw_equ equals 0! %s:%d\n", __FILE__, __LINE__){do { } while(0); }; |
| 3874 | #endif |
| 3875 | } |
| 3876 | } |
| 3877 | |
| 3878 | if (dst_y_prefetch_oto < dst_y_prefetch_equ) { |
| 3879 | if (dst_y_prefetch_oto * LineTime < TPreReq) { |
| 3880 | *DestinationLinesForPrefetch = dst_y_prefetch_equ; |
| 3881 | } else { |
| 3882 | *DestinationLinesForPrefetch = dst_y_prefetch_oto; |
| 3883 | } |
| 3884 | TimeForFetchingMetaPTE = Tvm_oto; |
| 3885 | TimeForFetchingRowInVBlank = Tr0_oto; |
| 3886 | *PrefetchBandwidth = prefetch_bw_oto; |
| 3887 | /* Clamp to oto for bandwidth calculation */ |
| 3888 | LinesForPrefetchBandwidth = dst_y_prefetch_oto; |
| 3889 | } else { |
| 3890 | *DestinationLinesForPrefetch = dst_y_prefetch_equ; |
| 3891 | TimeForFetchingMetaPTE = Tvm_equ; |
| 3892 | TimeForFetchingRowInVBlank = Tr0_equ; |
| 3893 | *PrefetchBandwidth = prefetch_bw_equ; |
| 3894 | /* Clamp to equ for bandwidth calculation */ |
| 3895 | LinesForPrefetchBandwidth = dst_y_prefetch_equ; |
| 3896 | } |
| 3897 | |
| 3898 | *DestinationLinesToRequestVMInVBlank = dml_ceil(4.0 * TimeForFetchingMetaPTE / LineTime, 1.0) / 4.0; |
| 3899 | |
| 3900 | *DestinationLinesToRequestRowInVBlank = |
| 3901 | dml_ceil(4.0 * TimeForFetchingRowInVBlank / LineTime, 1.0) / 4.0; |
| 3902 | |
| 3903 | LinesToRequestPrefetchPixelData = LinesForPrefetchBandwidth - |
| 3904 | *DestinationLinesToRequestVMInVBlank - 2 * *DestinationLinesToRequestRowInVBlank; |
| 3905 | |
| 3906 | #ifdef __DML_VBA_DEBUG__ |
| 3907 | dml_print("DML::%s: DestinationLinesForPrefetch = %f\n", __func__, *DestinationLinesForPrefetch){do { } while(0); }; |
| 3908 | dml_print("DML::%s: DestinationLinesToRequestVMInVBlank = %f\n",{do { } while(0); } |
| 3909 | __func__, *DestinationLinesToRequestVMInVBlank){do { } while(0); }; |
| 3910 | dml_print("DML::%s: TimeForFetchingRowInVBlank = %f\n", __func__, TimeForFetchingRowInVBlank){do { } while(0); }; |
| 3911 | dml_print("DML::%s: LineTime = %f\n", __func__, LineTime){do { } while(0); }; |
| 3912 | dml_print("DML::%s: DestinationLinesToRequestRowInVBlank = %f\n",{do { } while(0); } |
| 3913 | __func__, *DestinationLinesToRequestRowInVBlank){do { } while(0); }; |
| 3914 | dml_print("DML::%s: PrefetchSourceLinesY = %f\n", __func__, PrefetchSourceLinesY){do { } while(0); }; |
| 3915 | dml_print("DML::%s: LinesToRequestPrefetchPixelData = %f\n", __func__, LinesToRequestPrefetchPixelData){do { } while(0); }; |
| 3916 | #endif |
| 3917 | |
| 3918 | if (LinesToRequestPrefetchPixelData >= 1 && prefetch_bw_equ > 0) { |
| 3919 | *VRatioPrefetchY = (double) PrefetchSourceLinesY / LinesToRequestPrefetchPixelData; |
| 3920 | *VRatioPrefetchY = dml_max(*VRatioPrefetchY, 1.0); |
| 3921 | #ifdef __DML_VBA_DEBUG__ |
| 3922 | dml_print("DML::%s: VRatioPrefetchY = %f\n", __func__, *VRatioPrefetchY){do { } while(0); }; |
| 3923 | dml_print("DML::%s: SwathHeightY = %d\n", __func__, SwathHeightY){do { } while(0); }; |
| 3924 | dml_print("DML::%s: VInitPreFillY = %d\n", __func__, VInitPreFillY){do { } while(0); }; |
| 3925 | #endif |
| 3926 | if ((SwathHeightY > 4) && (VInitPreFillY > 3)) { |
| 3927 | if (LinesToRequestPrefetchPixelData > (VInitPreFillY - 3.0) / 2.0) { |
| 3928 | *VRatioPrefetchY = |
| 3929 | dml_max((double) PrefetchSourceLinesY / |
| 3930 | LinesToRequestPrefetchPixelData, |
| 3931 | (double) MaxNumSwathY * SwathHeightY / |
| 3932 | (LinesToRequestPrefetchPixelData - |
| 3933 | (VInitPreFillY - 3.0) / 2.0)); |
| 3934 | *VRatioPrefetchY = dml_max(*VRatioPrefetchY, 1.0); |
| 3935 | } else { |
| 3936 | MyError = true1; |
| 3937 | *VRatioPrefetchY = 0; |
| 3938 | } |
| 3939 | #ifdef __DML_VBA_DEBUG__ |
| 3940 | dml_print("DML::%s: VRatioPrefetchY = %f\n", __func__, *VRatioPrefetchY){do { } while(0); }; |
| 3941 | dml_print("DML::%s: PrefetchSourceLinesY = %f\n", __func__, PrefetchSourceLinesY){do { } while(0); }; |
| 3942 | dml_print("DML::%s: MaxNumSwathY = %d\n", __func__, MaxNumSwathY){do { } while(0); }; |
| 3943 | #endif |
| 3944 | } |
| 3945 | |
| 3946 | *VRatioPrefetchC = (double) PrefetchSourceLinesC / LinesToRequestPrefetchPixelData; |
| 3947 | *VRatioPrefetchC = dml_max(*VRatioPrefetchC, 1.0); |
| 3948 | |
| 3949 | #ifdef __DML_VBA_DEBUG__ |
| 3950 | dml_print("DML::%s: VRatioPrefetchC = %f\n", __func__, *VRatioPrefetchC){do { } while(0); }; |
| 3951 | dml_print("DML::%s: SwathHeightC = %d\n", __func__, SwathHeightC){do { } while(0); }; |
| 3952 | dml_print("DML::%s: VInitPreFillC = %d\n", __func__, VInitPreFillC){do { } while(0); }; |
| 3953 | #endif |
| 3954 | if ((SwathHeightC > 4)) { |
| 3955 | if (LinesToRequestPrefetchPixelData > (VInitPreFillC - 3.0) / 2.0) { |
| 3956 | *VRatioPrefetchC = |
| 3957 | dml_max(*VRatioPrefetchC, |
| 3958 | (double) MaxNumSwathC * SwathHeightC / |
| 3959 | (LinesToRequestPrefetchPixelData - |
| 3960 | (VInitPreFillC - 3.0) / 2.0)); |
| 3961 | *VRatioPrefetchC = dml_max(*VRatioPrefetchC, 1.0); |
| 3962 | } else { |
| 3963 | MyError = true1; |
| 3964 | *VRatioPrefetchC = 0; |
| 3965 | } |
| 3966 | #ifdef __DML_VBA_DEBUG__ |
| 3967 | dml_print("DML::%s: VRatioPrefetchC = %f\n", __func__, *VRatioPrefetchC){do { } while(0); }; |
| 3968 | dml_print("DML::%s: PrefetchSourceLinesC = %f\n", __func__, PrefetchSourceLinesC){do { } while(0); }; |
| 3969 | dml_print("DML::%s: MaxNumSwathC = %d\n", __func__, MaxNumSwathC){do { } while(0); }; |
| 3970 | #endif |
| 3971 | } |
| 3972 | |
| 3973 | *RequiredPrefetchPixDataBWLuma = (double) PrefetchSourceLinesY |
| 3974 | / LinesToRequestPrefetchPixelData * myPipe->BytePerPixelY * swath_width_luma_ub |
| 3975 | / LineTime; |
| 3976 | |
| 3977 | #ifdef __DML_VBA_DEBUG__ |
| 3978 | dml_print("DML::%s: BytePerPixelY = %d\n", __func__, myPipe->BytePerPixelY){do { } while(0); }; |
| 3979 | dml_print("DML::%s: swath_width_luma_ub = %d\n", __func__, swath_width_luma_ub){do { } while(0); }; |
| 3980 | dml_print("DML::%s: LineTime = %f\n", __func__, LineTime){do { } while(0); }; |
| 3981 | dml_print("DML::%s: RequiredPrefetchPixDataBWLuma = %f\n",{do { } while(0); } |
| 3982 | __func__, *RequiredPrefetchPixDataBWLuma){do { } while(0); }; |
| 3983 | #endif |
| 3984 | *RequiredPrefetchPixDataBWChroma = (double) PrefetchSourceLinesC / |
| 3985 | LinesToRequestPrefetchPixelData |
| 3986 | * myPipe->BytePerPixelC |
| 3987 | * swath_width_chroma_ub / LineTime; |
| 3988 | } else { |
| 3989 | MyError = true1; |
| 3990 | #ifdef __DML_VBA_DEBUG__ |
| 3991 | dml_print("DML:%s: MyErr set. LinesToRequestPrefetchPixelData: %f, should be > 0\n",{do { } while(0); } |
| 3992 | __func__, LinesToRequestPrefetchPixelData){do { } while(0); }; |
| 3993 | #endif |
| 3994 | *VRatioPrefetchY = 0; |
| 3995 | *VRatioPrefetchC = 0; |
| 3996 | *RequiredPrefetchPixDataBWLuma = 0; |
| 3997 | *RequiredPrefetchPixDataBWChroma = 0; |
| 3998 | } |
| 3999 | #ifdef __DML_VBA_DEBUG__ |
| 4000 | dml_print("DML: Tpre: %fus - sum of time to request meta pte, 2 x data pte + meta data, swaths\n",{do { } while(0); } |
| 4001 | (double)LinesToRequestPrefetchPixelData * LineTime +{do { } while(0); } |
| 4002 | 2.0*TimeForFetchingRowInVBlank + TimeForFetchingMetaPTE){do { } while(0); }; |
| 4003 | dml_print("DML: Tvm: %fus - time to fetch page tables for meta surface\n", TimeForFetchingMetaPTE){do { } while(0); }; |
| 4004 | dml_print("DML: To: %fus - time for propagation from scaler to optc\n",{do { } while(0); } |
| 4005 | (*DSTYAfterScaler + ((double) (*DSTXAfterScaler) / (double) myPipe->HTotal)) * LineTime){do { } while(0); }; |
| 4006 | dml_print("DML: Tvstartup - TSetup - Tcalc - Twait - Tpre - To > 0\n"){do { } while(0); }; |
| 4007 | dml_print("DML: Tslack(pre): %fus - time left over in schedule\n", VStartup * LineTime -{do { } while(0); } |
| 4008 | TimeForFetchingMetaPTE - 2*TimeForFetchingRowInVBlank - (*DSTYAfterScaler +{do { } while(0); } |
| 4009 | ((double) (*DSTXAfterScaler) / (double) myPipe->HTotal)) * LineTime - TWait - TCalc - *TSetup){do { } while(0); }; |
| 4010 | dml_print("DML: row_bytes = dpte_row_bytes (per_pipe) = PixelPTEBytesPerRow = : %d\n",{do { } while(0); } |
| 4011 | PixelPTEBytesPerRow){do { } while(0); }; |
| 4012 | #endif |
| 4013 | } else { |
| 4014 | MyError = true1; |
| 4015 | #ifdef __DML_VBA_DEBUG__ |
| 4016 | dml_print("DML::%s: MyErr set, dst_y_prefetch_equ = %f (should be > 1)\n",{do { } while(0); } |
| 4017 | __func__, dst_y_prefetch_equ){do { } while(0); }; |
| 4018 | #endif |
| 4019 | } |
| 4020 | |
| 4021 | { |
| 4022 | double prefetch_vm_bw; |
| 4023 | double prefetch_row_bw; |
| 4024 | |
| 4025 | if (PDEAndMetaPTEBytesFrame == 0) { |
| 4026 | prefetch_vm_bw = 0; |
| 4027 | } else if (*DestinationLinesToRequestVMInVBlank > 0) { |
| 4028 | #ifdef __DML_VBA_DEBUG__ |
| 4029 | dml_print("DML::%s: PDEAndMetaPTEBytesFrame = %d\n", __func__, PDEAndMetaPTEBytesFrame){do { } while(0); }; |
| 4030 | dml_print("DML::%s: HostVMInefficiencyFactor = %f\n", __func__, HostVMInefficiencyFactor){do { } while(0); }; |
| 4031 | dml_print("DML::%s: DestinationLinesToRequestVMInVBlank = %f\n",{do { } while(0); } |
| 4032 | __func__, *DestinationLinesToRequestVMInVBlank){do { } while(0); }; |
| 4033 | dml_print("DML::%s: LineTime = %f\n", __func__, LineTime){do { } while(0); }; |
| 4034 | #endif |
| 4035 | prefetch_vm_bw = PDEAndMetaPTEBytesFrame * HostVMInefficiencyFactor / |
| 4036 | (*DestinationLinesToRequestVMInVBlank * LineTime); |
| 4037 | #ifdef __DML_VBA_DEBUG__ |
| 4038 | dml_print("DML::%s: prefetch_vm_bw = %f\n", __func__, prefetch_vm_bw){do { } while(0); }; |
| 4039 | #endif |
| 4040 | } else { |
| 4041 | prefetch_vm_bw = 0; |
| 4042 | MyError = true1; |
| 4043 | #ifdef __DML_VBA_DEBUG__ |
| 4044 | dml_print("DML::%s: MyErr set. DestinationLinesToRequestVMInVBlank=%f (should be > 0)\n",{do { } while(0); } |
| 4045 | __func__, *DestinationLinesToRequestVMInVBlank){do { } while(0); }; |
| 4046 | #endif |
| 4047 | } |
| 4048 | |
| 4049 | if (MetaRowByte + PixelPTEBytesPerRow == 0) { |
| 4050 | prefetch_row_bw = 0; |
| 4051 | } else if (*DestinationLinesToRequestRowInVBlank > 0) { |
| 4052 | prefetch_row_bw = (MetaRowByte + PixelPTEBytesPerRow * HostVMInefficiencyFactor) / |
| 4053 | (*DestinationLinesToRequestRowInVBlank * LineTime); |
| 4054 | |
| 4055 | #ifdef __DML_VBA_DEBUG__ |
| 4056 | dml_print("DML::%s: MetaRowByte = %d\n", __func__, MetaRowByte){do { } while(0); }; |
| 4057 | dml_print("DML::%s: PixelPTEBytesPerRow = %d\n", __func__, PixelPTEBytesPerRow){do { } while(0); }; |
| 4058 | dml_print("DML::%s: DestinationLinesToRequestRowInVBlank = %f\n",{do { } while(0); } |
| 4059 | __func__, *DestinationLinesToRequestRowInVBlank){do { } while(0); }; |
| 4060 | dml_print("DML::%s: prefetch_row_bw = %f\n", __func__, prefetch_row_bw){do { } while(0); }; |
| 4061 | #endif |
| 4062 | } else { |
| 4063 | prefetch_row_bw = 0; |
| 4064 | MyError = true1; |
| 4065 | #ifdef __DML_VBA_DEBUG__ |
| 4066 | dml_print("DML::%s: MyErr set. DestinationLinesToRequestRowInVBlank=%f (should be > 0)\n",{do { } while(0); } |
| 4067 | __func__, *DestinationLinesToRequestRowInVBlank){do { } while(0); }; |
| 4068 | #endif |
| 4069 | } |
| 4070 | |
| 4071 | *prefetch_vmrow_bw = dml_max(prefetch_vm_bw, prefetch_row_bw); |
| 4072 | } |
| 4073 | |
| 4074 | if (MyError) { |
| 4075 | *PrefetchBandwidth = 0; |
| 4076 | TimeForFetchingMetaPTE = 0; |
| 4077 | TimeForFetchingRowInVBlank = 0; |
Value stored to 'TimeForFetchingRowInVBlank' is never read | |
| 4078 | *DestinationLinesToRequestVMInVBlank = 0; |
| 4079 | *DestinationLinesToRequestRowInVBlank = 0; |
| 4080 | *DestinationLinesForPrefetch = 0; |
| 4081 | LinesToRequestPrefetchPixelData = 0; |
| 4082 | *VRatioPrefetchY = 0; |
| 4083 | *VRatioPrefetchC = 0; |
| 4084 | *RequiredPrefetchPixDataBWLuma = 0; |
| 4085 | *RequiredPrefetchPixDataBWChroma = 0; |
| 4086 | } |
| 4087 | |
| 4088 | return MyError; |
| 4089 | } // CalculatePrefetchSchedule |
| 4090 | |
| 4091 | void dml32_CalculateFlipSchedule( |
| 4092 | double HostVMInefficiencyFactor, |
| 4093 | double UrgentExtraLatency, |
| 4094 | double UrgentLatency, |
| 4095 | unsigned int GPUVMMaxPageTableLevels, |
| 4096 | bool_Bool HostVMEnable, |
| 4097 | unsigned int HostVMMaxNonCachedPageTableLevels, |
| 4098 | bool_Bool GPUVMEnable, |
| 4099 | double HostVMMinPageSize, |
| 4100 | double PDEAndMetaPTEBytesPerFrame, |
| 4101 | double MetaRowBytes, |
| 4102 | double DPTEBytesPerRow, |
| 4103 | double BandwidthAvailableForImmediateFlip, |
| 4104 | unsigned int TotImmediateFlipBytes, |
| 4105 | enum source_format_class SourcePixelFormat, |
| 4106 | double LineTime, |
| 4107 | double VRatio, |
| 4108 | double VRatioChroma, |
| 4109 | double Tno_bw, |
| 4110 | bool_Bool DCCEnable, |
| 4111 | unsigned int dpte_row_height, |
| 4112 | unsigned int meta_row_height, |
| 4113 | unsigned int dpte_row_height_chroma, |
| 4114 | unsigned int meta_row_height_chroma, |
| 4115 | bool_Bool use_one_row_for_frame_flip, |
| 4116 | |
| 4117 | /* Output */ |
| 4118 | double *DestinationLinesToRequestVMInImmediateFlip, |
| 4119 | double *DestinationLinesToRequestRowInImmediateFlip, |
| 4120 | double *final_flip_bw, |
| 4121 | bool_Bool *ImmediateFlipSupportedForPipe) |
| 4122 | { |
| 4123 | double min_row_time = 0.0; |
| 4124 | unsigned int HostVMDynamicLevelsTrips; |
| 4125 | double TimeForFetchingMetaPTEImmediateFlip; |
| 4126 | double TimeForFetchingRowInVBlankImmediateFlip; |
| 4127 | double ImmediateFlipBW; |
| 4128 | |
| 4129 | if (GPUVMEnable == true1 && HostVMEnable == true1) |
| 4130 | HostVMDynamicLevelsTrips = HostVMMaxNonCachedPageTableLevels; |
| 4131 | else |
| 4132 | HostVMDynamicLevelsTrips = 0; |
| 4133 | |
| 4134 | #ifdef __DML_VBA_DEBUG__ |
| 4135 | dml_print("DML::%s: TotImmediateFlipBytes = %d\n", __func__, TotImmediateFlipBytes){do { } while(0); }; |
| 4136 | dml_print("DML::%s: BandwidthAvailableForImmediateFlip = %f\n", __func__, BandwidthAvailableForImmediateFlip){do { } while(0); }; |
| 4137 | #endif |
| 4138 | |
| 4139 | if (TotImmediateFlipBytes > 0) { |
| 4140 | if (use_one_row_for_frame_flip) { |
| 4141 | ImmediateFlipBW = (PDEAndMetaPTEBytesPerFrame + MetaRowBytes + 2 * DPTEBytesPerRow) * |
| 4142 | BandwidthAvailableForImmediateFlip / TotImmediateFlipBytes; |
| 4143 | } else { |
| 4144 | ImmediateFlipBW = (PDEAndMetaPTEBytesPerFrame + MetaRowBytes + DPTEBytesPerRow) * |
| 4145 | BandwidthAvailableForImmediateFlip / TotImmediateFlipBytes; |
| 4146 | } |
| 4147 | if (GPUVMEnable == true1) { |
| 4148 | TimeForFetchingMetaPTEImmediateFlip = dml_max3(Tno_bw + PDEAndMetaPTEBytesPerFrame * |
| 4149 | HostVMInefficiencyFactor / ImmediateFlipBW, |
| 4150 | UrgentExtraLatency + UrgentLatency * |
| 4151 | (GPUVMMaxPageTableLevels * (HostVMDynamicLevelsTrips + 1) - 1), |
| 4152 | LineTime / 4.0); |
| 4153 | } else { |
| 4154 | TimeForFetchingMetaPTEImmediateFlip = 0; |
| 4155 | } |
| 4156 | if ((GPUVMEnable == true1 || DCCEnable == true1)) { |
| 4157 | TimeForFetchingRowInVBlankImmediateFlip = dml_max3( |
| 4158 | (MetaRowBytes + DPTEBytesPerRow * HostVMInefficiencyFactor) / ImmediateFlipBW, |
| 4159 | UrgentLatency * (HostVMDynamicLevelsTrips + 1), LineTime / 4.0); |
| 4160 | } else { |
| 4161 | TimeForFetchingRowInVBlankImmediateFlip = 0; |
| 4162 | } |
| 4163 | |
| 4164 | *DestinationLinesToRequestVMInImmediateFlip = |
| 4165 | dml_ceil(4.0 * (TimeForFetchingMetaPTEImmediateFlip / LineTime), 1.0) / 4.0; |
| 4166 | *DestinationLinesToRequestRowInImmediateFlip = |
| 4167 | dml_ceil(4.0 * (TimeForFetchingRowInVBlankImmediateFlip / LineTime), 1.0) / 4.0; |
| 4168 | |
| 4169 | if (GPUVMEnable == true1) { |
| 4170 | *final_flip_bw = dml_max(PDEAndMetaPTEBytesPerFrame * HostVMInefficiencyFactor / |
| 4171 | (*DestinationLinesToRequestVMInImmediateFlip * LineTime), |
| 4172 | (MetaRowBytes + DPTEBytesPerRow * HostVMInefficiencyFactor) / |
| 4173 | (*DestinationLinesToRequestRowInImmediateFlip * LineTime)); |
| 4174 | } else if ((GPUVMEnable == true1 || DCCEnable == true1)) { |
| 4175 | *final_flip_bw = (MetaRowBytes + DPTEBytesPerRow * HostVMInefficiencyFactor) / |
| 4176 | (*DestinationLinesToRequestRowInImmediateFlip * LineTime); |
| 4177 | } else { |
| 4178 | *final_flip_bw = 0; |
| 4179 | } |
| 4180 | } else { |
| 4181 | TimeForFetchingMetaPTEImmediateFlip = 0; |
| 4182 | TimeForFetchingRowInVBlankImmediateFlip = 0; |
| 4183 | *DestinationLinesToRequestVMInImmediateFlip = 0; |
| 4184 | *DestinationLinesToRequestRowInImmediateFlip = 0; |
| 4185 | *final_flip_bw = 0; |
| 4186 | } |
| 4187 | |
| 4188 | if (SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10 || SourcePixelFormat == dm_rgbe_alpha) { |
| 4189 | if (GPUVMEnable == true1 && DCCEnable != true1) { |
| 4190 | min_row_time = dml_min(dpte_row_height * |
| 4191 | LineTime / VRatio, dpte_row_height_chroma * LineTime / VRatioChroma); |
| 4192 | } else if (GPUVMEnable != true1 && DCCEnable == true1) { |
| 4193 | min_row_time = dml_min(meta_row_height * |
| 4194 | LineTime / VRatio, meta_row_height_chroma * LineTime / VRatioChroma); |
| 4195 | } else { |
| 4196 | min_row_time = dml_min4(dpte_row_height * LineTime / VRatio, meta_row_height * |
| 4197 | LineTime / VRatio, dpte_row_height_chroma * LineTime / |
| 4198 | VRatioChroma, meta_row_height_chroma * LineTime / VRatioChroma); |
| 4199 | } |
| 4200 | } else { |
| 4201 | if (GPUVMEnable == true1 && DCCEnable != true1) { |
| 4202 | min_row_time = dpte_row_height * LineTime / VRatio; |
| 4203 | } else if (GPUVMEnable != true1 && DCCEnable == true1) { |
| 4204 | min_row_time = meta_row_height * LineTime / VRatio; |
| 4205 | } else { |
| 4206 | min_row_time = |
| 4207 | dml_min(dpte_row_height * LineTime / VRatio, meta_row_height * LineTime / VRatio); |
| 4208 | } |
| 4209 | } |
| 4210 | |
| 4211 | if (*DestinationLinesToRequestVMInImmediateFlip >= 32 || *DestinationLinesToRequestRowInImmediateFlip >= 16 |
| 4212 | || TimeForFetchingMetaPTEImmediateFlip + 2 * TimeForFetchingRowInVBlankImmediateFlip |
| 4213 | > min_row_time) { |
| 4214 | *ImmediateFlipSupportedForPipe = false0; |
| 4215 | } else { |
| 4216 | *ImmediateFlipSupportedForPipe = true1; |
| 4217 | } |
| 4218 | |
| 4219 | #ifdef __DML_VBA_DEBUG__ |
| 4220 | dml_print("DML::%s: GPUVMEnable = %d\n", __func__, GPUVMEnable){do { } while(0); }; |
| 4221 | dml_print("DML::%s: DCCEnable = %d\n", __func__, DCCEnable){do { } while(0); }; |
| 4222 | dml_print("DML::%s: DestinationLinesToRequestVMInImmediateFlip = %f\n",{do { } while(0); } |
| 4223 | __func__, *DestinationLinesToRequestVMInImmediateFlip){do { } while(0); }; |
| 4224 | dml_print("DML::%s: DestinationLinesToRequestRowInImmediateFlip = %f\n",{do { } while(0); } |
| 4225 | __func__, *DestinationLinesToRequestRowInImmediateFlip){do { } while(0); }; |
| 4226 | dml_print("DML::%s: TimeForFetchingMetaPTEImmediateFlip = %f\n", __func__, TimeForFetchingMetaPTEImmediateFlip){do { } while(0); }; |
| 4227 | dml_print("DML::%s: TimeForFetchingRowInVBlankImmediateFlip = %f\n",{do { } while(0); } |
| 4228 | __func__, TimeForFetchingRowInVBlankImmediateFlip){do { } while(0); }; |
| 4229 | dml_print("DML::%s: min_row_time = %f\n", __func__, min_row_time){do { } while(0); }; |
| 4230 | dml_print("DML::%s: ImmediateFlipSupportedForPipe = %d\n", __func__, *ImmediateFlipSupportedForPipe){do { } while(0); }; |
| 4231 | #endif |
| 4232 | } // CalculateFlipSchedule |
| 4233 | |
| 4234 | void dml32_CalculateWatermarksMALLUseAndDRAMSpeedChangeSupport( |
| 4235 | struct vba_vars_st *v, |
| 4236 | unsigned int PrefetchMode, |
| 4237 | double DCFCLK, |
| 4238 | double ReturnBW, |
| 4239 | SOCParametersList mmSOCParameters, |
| 4240 | double SOCCLK, |
| 4241 | double DCFClkDeepSleep, |
| 4242 | unsigned int DETBufferSizeY[], |
| 4243 | unsigned int DETBufferSizeC[], |
| 4244 | unsigned int SwathHeightY[], |
| 4245 | unsigned int SwathHeightC[], |
| 4246 | double SwathWidthY[], |
| 4247 | double SwathWidthC[], |
| 4248 | unsigned int DPPPerSurface[], |
| 4249 | double BytePerPixelDETY[], |
| 4250 | double BytePerPixelDETC[], |
| 4251 | double DSTXAfterScaler[], |
| 4252 | double DSTYAfterScaler[], |
| 4253 | bool_Bool UnboundedRequestEnabled, |
| 4254 | unsigned int CompressedBufferSizeInkByte, |
| 4255 | |
| 4256 | /* Output */ |
| 4257 | enum clock_change_support *DRAMClockChangeSupport, |
| 4258 | double MaxActiveDRAMClockChangeLatencySupported[], |
| 4259 | unsigned int SubViewportLinesNeededInMALL[], |
| 4260 | enum dm_fclock_change_support *FCLKChangeSupport, |
| 4261 | double *MinActiveFCLKChangeLatencySupported, |
| 4262 | bool_Bool *USRRetrainingSupport, |
| 4263 | double ActiveDRAMClockChangeLatencyMargin[]) |
| 4264 | { |
| 4265 | unsigned int i, j, k; |
| 4266 | unsigned int SurfaceWithMinActiveFCLKChangeMargin = 0; |
| 4267 | unsigned int DRAMClockChangeSupportNumber = 0; |
| 4268 | unsigned int LastSurfaceWithoutMargin; |
| 4269 | unsigned int DRAMClockChangeMethod = 0; |
| 4270 | bool_Bool FoundFirstSurfaceWithMinActiveFCLKChangeMargin = false0; |
| 4271 | double MinActiveFCLKChangeMargin = 0.; |
| 4272 | double SecondMinActiveFCLKChangeMarginOneDisplayInVBLank = 0.; |
| 4273 | double ActiveClockChangeLatencyHidingY; |
| 4274 | double ActiveClockChangeLatencyHidingC; |
| 4275 | double ActiveClockChangeLatencyHiding; |
| 4276 | double EffectiveDETBufferSizeY; |
| 4277 | double ActiveFCLKChangeLatencyMargin[DC__NUM_DPP__MAX8]; |
| 4278 | double USRRetrainingLatencyMargin[DC__NUM_DPP__MAX8]; |
| 4279 | double TotalPixelBW = 0.0; |
| 4280 | bool_Bool SynchronizedSurfaces[DC__NUM_DPP__MAX8][DC__NUM_DPP__MAX8]; |
| 4281 | double EffectiveLBLatencyHidingY; |
| 4282 | double EffectiveLBLatencyHidingC; |
| 4283 | double LinesInDETY[DC__NUM_DPP__MAX8]; |
| 4284 | double LinesInDETC[DC__NUM_DPP__MAX8]; |
| 4285 | unsigned int LinesInDETYRoundedDownToSwath[DC__NUM_DPP__MAX8]; |
| 4286 | unsigned int LinesInDETCRoundedDownToSwath[DC__NUM_DPP__MAX8]; |
| 4287 | double FullDETBufferingTimeY; |
| 4288 | double FullDETBufferingTimeC; |
| 4289 | double WritebackDRAMClockChangeLatencyMargin; |
| 4290 | double WritebackFCLKChangeLatencyMargin; |
| 4291 | double WritebackLatencyHiding; |
| 4292 | bool_Bool SameTimingForFCLKChange; |
| 4293 | |
| 4294 | unsigned int TotalActiveWriteback = 0; |
| 4295 | unsigned int LBLatencyHidingSourceLinesY[DC__NUM_DPP__MAX8]; |
| 4296 | unsigned int LBLatencyHidingSourceLinesC[DC__NUM_DPP__MAX8]; |
| 4297 | |
| 4298 | v->Watermark.UrgentWatermark = mmSOCParameters.UrgentLatency + mmSOCParameters.ExtraLatency; |
| 4299 | v->Watermark.USRRetrainingWatermark = mmSOCParameters.UrgentLatency + mmSOCParameters.ExtraLatency |
| 4300 | + mmSOCParameters.USRRetrainingLatency + mmSOCParameters.SMNLatency; |
| 4301 | v->Watermark.DRAMClockChangeWatermark = mmSOCParameters.DRAMClockChangeLatency + v->Watermark.UrgentWatermark; |
| 4302 | v->Watermark.FCLKChangeWatermark = mmSOCParameters.FCLKChangeLatency + v->Watermark.UrgentWatermark; |
| 4303 | v->Watermark.StutterExitWatermark = mmSOCParameters.SRExitTime + mmSOCParameters.ExtraLatency |
| 4304 | + 10 / DCFClkDeepSleep; |
| 4305 | v->Watermark.StutterEnterPlusExitWatermark = mmSOCParameters.SREnterPlusExitTime + mmSOCParameters.ExtraLatency |
| 4306 | + 10 / DCFClkDeepSleep; |
| 4307 | v->Watermark.Z8StutterExitWatermark = mmSOCParameters.SRExitZ8Time + mmSOCParameters.ExtraLatency |
| 4308 | + 10 / DCFClkDeepSleep; |
| 4309 | v->Watermark.Z8StutterEnterPlusExitWatermark = mmSOCParameters.SREnterPlusExitZ8Time |
| 4310 | + mmSOCParameters.ExtraLatency + 10 / DCFClkDeepSleep; |
| 4311 | |
| 4312 | #ifdef __DML_VBA_DEBUG__ |
| 4313 | dml_print("DML::%s: UrgentLatency = %f\n", __func__, mmSOCParameters.UrgentLatency){do { } while(0); }; |
| 4314 | dml_print("DML::%s: ExtraLatency = %f\n", __func__, mmSOCParameters.ExtraLatency){do { } while(0); }; |
| 4315 | dml_print("DML::%s: DRAMClockChangeLatency = %f\n", __func__, mmSOCParameters.DRAMClockChangeLatency){do { } while(0); }; |
| 4316 | dml_print("DML::%s: UrgentWatermark = %f\n", __func__, v->Watermark.UrgentWatermark){do { } while(0); }; |
| 4317 | dml_print("DML::%s: USRRetrainingWatermark = %f\n", __func__, v->Watermark.USRRetrainingWatermark){do { } while(0); }; |
| 4318 | dml_print("DML::%s: DRAMClockChangeWatermark = %f\n", __func__, v->Watermark.DRAMClockChangeWatermark){do { } while(0); }; |
| 4319 | dml_print("DML::%s: FCLKChangeWatermark = %f\n", __func__, v->Watermark.FCLKChangeWatermark){do { } while(0); }; |
| 4320 | dml_print("DML::%s: StutterExitWatermark = %f\n", __func__, v->Watermark.StutterExitWatermark){do { } while(0); }; |
| 4321 | dml_print("DML::%s: StutterEnterPlusExitWatermark = %f\n", __func__, v->Watermark.StutterEnterPlusExitWatermark){do { } while(0); }; |
| 4322 | dml_print("DML::%s: Z8StutterExitWatermark = %f\n", __func__, v->Watermark.Z8StutterExitWatermark){do { } while(0); }; |
| 4323 | dml_print("DML::%s: Z8StutterEnterPlusExitWatermark = %f\n",{do { } while(0); } |
| 4324 | __func__, v->Watermark.Z8StutterEnterPlusExitWatermark){do { } while(0); }; |
| 4325 | #endif |
| 4326 | |
| 4327 | |
| 4328 | TotalActiveWriteback = 0; |
| 4329 | for (k = 0; k < v->NumberOfActiveSurfaces; ++k) { |
| 4330 | if (v->WritebackEnable[k] == true1) |
| 4331 | TotalActiveWriteback = TotalActiveWriteback + 1; |
| 4332 | } |
| 4333 | |
| 4334 | if (TotalActiveWriteback <= 1) { |
| 4335 | v->Watermark.WritebackUrgentWatermark = mmSOCParameters.WritebackLatency; |
| 4336 | } else { |
| 4337 | v->Watermark.WritebackUrgentWatermark = mmSOCParameters.WritebackLatency |
| 4338 | + v->WritebackChunkSize * 1024.0 / 32.0 / SOCCLK; |
| 4339 | } |
| 4340 | if (v->USRRetrainingRequiredFinal) |
| 4341 | v->Watermark.WritebackUrgentWatermark = v->Watermark.WritebackUrgentWatermark |
| 4342 | + mmSOCParameters.USRRetrainingLatency; |
| 4343 | |
| 4344 | if (TotalActiveWriteback <= 1) { |
| 4345 | v->Watermark.WritebackDRAMClockChangeWatermark = mmSOCParameters.DRAMClockChangeLatency |
| 4346 | + mmSOCParameters.WritebackLatency; |
| 4347 | v->Watermark.WritebackFCLKChangeWatermark = mmSOCParameters.FCLKChangeLatency |
| 4348 | + mmSOCParameters.WritebackLatency; |
| 4349 | } else { |
| 4350 | v->Watermark.WritebackDRAMClockChangeWatermark = mmSOCParameters.DRAMClockChangeLatency |
| 4351 | + mmSOCParameters.WritebackLatency + v->WritebackChunkSize * 1024.0 / 32.0 / SOCCLK; |
| 4352 | v->Watermark.WritebackFCLKChangeWatermark = mmSOCParameters.FCLKChangeLatency |
| 4353 | + mmSOCParameters.WritebackLatency + v->WritebackChunkSize * 1024 / 32 / SOCCLK; |
| 4354 | } |
| 4355 | |
| 4356 | if (v->USRRetrainingRequiredFinal) |
| 4357 | v->Watermark.WritebackDRAMClockChangeWatermark = v->Watermark.WritebackDRAMClockChangeWatermark |
| 4358 | + mmSOCParameters.USRRetrainingLatency; |
| 4359 | |
| 4360 | if (v->USRRetrainingRequiredFinal) |
| 4361 | v->Watermark.WritebackFCLKChangeWatermark = v->Watermark.WritebackFCLKChangeWatermark |
| 4362 | + mmSOCParameters.USRRetrainingLatency; |
| 4363 | |
| 4364 | #ifdef __DML_VBA_DEBUG__ |
| 4365 | dml_print("DML::%s: WritebackDRAMClockChangeWatermark = %f\n",{do { } while(0); } |
| 4366 | __func__, v->Watermark.WritebackDRAMClockChangeWatermark){do { } while(0); }; |
| 4367 | dml_print("DML::%s: WritebackFCLKChangeWatermark = %f\n", __func__, v->Watermark.WritebackFCLKChangeWatermark){do { } while(0); }; |
| 4368 | dml_print("DML::%s: WritebackUrgentWatermark = %f\n", __func__, v->Watermark.WritebackUrgentWatermark){do { } while(0); }; |
| 4369 | dml_print("DML::%s: v->USRRetrainingRequiredFinal = %d\n", __func__, v->USRRetrainingRequiredFinal){do { } while(0); }; |
| 4370 | dml_print("DML::%s: USRRetrainingLatency = %f\n", __func__, mmSOCParameters.USRRetrainingLatency){do { } while(0); }; |
| 4371 | #endif |
| 4372 | |
| 4373 | for (k = 0; k < v->NumberOfActiveSurfaces; ++k) { |
| 4374 | TotalPixelBW = TotalPixelBW + DPPPerSurface[k] * (SwathWidthY[k] * BytePerPixelDETY[k] * v->VRatio[k] + |
| 4375 | SwathWidthC[k] * BytePerPixelDETC[k] * v->VRatioChroma[k]) / (v->HTotal[k] / v->PixelClock[k]); |
| 4376 | } |
| 4377 | |
| 4378 | for (k = 0; k < v->NumberOfActiveSurfaces; ++k) { |
| 4379 | |
| 4380 | LBLatencyHidingSourceLinesY[k] = dml_min((double) v->MaxLineBufferLines, dml_floor(v->LineBufferSizeFinal / v->LBBitPerPixel[k] / (SwathWidthY[k] / dml_max(v->HRatio[k], 1.0)), 1)) - (v->vtaps[k] - 1); |
| 4381 | LBLatencyHidingSourceLinesC[k] = dml_min((double) v->MaxLineBufferLines, dml_floor(v->LineBufferSizeFinal / v->LBBitPerPixel[k] / (SwathWidthC[k] / dml_max(v->HRatioChroma[k], 1.0)), 1)) - (v->VTAPsChroma[k] - 1); |
| 4382 | |
| 4383 | |
| 4384 | #ifdef __DML_VBA_DEBUG__ |
| 4385 | dml_print("DML::%s: k=%d, v->MaxLineBufferLines = %d\n", __func__, k, v->MaxLineBufferLines){do { } while(0); }; |
| 4386 | dml_print("DML::%s: k=%d, v->LineBufferSizeFinal = %d\n", __func__, k, v->LineBufferSizeFinal){do { } while(0); }; |
| 4387 | dml_print("DML::%s: k=%d, v->LBBitPerPixel = %d\n", __func__, k, v->LBBitPerPixel[k]){do { } while(0); }; |
| 4388 | dml_print("DML::%s: k=%d, v->HRatio = %f\n", __func__, k, v->HRatio[k]){do { } while(0); }; |
| 4389 | dml_print("DML::%s: k=%d, v->vtaps = %d\n", __func__, k, v->vtaps[k]){do { } while(0); }; |
| 4390 | #endif |
| 4391 | |
| 4392 | EffectiveLBLatencyHidingY = LBLatencyHidingSourceLinesY[k] / v->VRatio[k] * (v->HTotal[k] / v->PixelClock[k]); |
| 4393 | EffectiveLBLatencyHidingC = LBLatencyHidingSourceLinesC[k] / v->VRatioChroma[k] * (v->HTotal[k] / v->PixelClock[k]); |
| 4394 | EffectiveDETBufferSizeY = DETBufferSizeY[k]; |
| 4395 | |
| 4396 | if (UnboundedRequestEnabled) { |
| 4397 | EffectiveDETBufferSizeY = EffectiveDETBufferSizeY |
| 4398 | + CompressedBufferSizeInkByte * 1024 |
| 4399 | * (SwathWidthY[k] * BytePerPixelDETY[k] * v->VRatio[k]) |
| 4400 | / (v->HTotal[k] / v->PixelClock[k]) / TotalPixelBW; |
| 4401 | } |
| 4402 | |
| 4403 | LinesInDETY[k] = (double) EffectiveDETBufferSizeY / BytePerPixelDETY[k] / SwathWidthY[k]; |
| 4404 | LinesInDETYRoundedDownToSwath[k] = dml_floor(LinesInDETY[k], SwathHeightY[k]); |
| 4405 | FullDETBufferingTimeY = LinesInDETYRoundedDownToSwath[k] * (v->HTotal[k] / v->PixelClock[k]) / v->VRatio[k]; |
| 4406 | |
| 4407 | ActiveClockChangeLatencyHidingY = EffectiveLBLatencyHidingY + FullDETBufferingTimeY |
| 4408 | - (DSTXAfterScaler[k] / v->HTotal[k] + DSTYAfterScaler[k]) * v->HTotal[k] / v->PixelClock[k]; |
| 4409 | |
| 4410 | if (v->NumberOfActiveSurfaces > 1) { |
| 4411 | ActiveClockChangeLatencyHidingY = ActiveClockChangeLatencyHidingY |
| 4412 | - (1.0 - 1.0 / v->NumberOfActiveSurfaces) * SwathHeightY[k] * v->HTotal[k] |
| 4413 | / v->PixelClock[k] / v->VRatio[k]; |
| 4414 | } |
| 4415 | |
| 4416 | if (BytePerPixelDETC[k] > 0) { |
| 4417 | LinesInDETC[k] = DETBufferSizeC[k] / BytePerPixelDETC[k] / SwathWidthC[k]; |
| 4418 | LinesInDETCRoundedDownToSwath[k] = dml_floor(LinesInDETC[k], SwathHeightC[k]); |
| 4419 | FullDETBufferingTimeC = LinesInDETCRoundedDownToSwath[k] * (v->HTotal[k] / v->PixelClock[k]) |
| 4420 | / v->VRatioChroma[k]; |
| 4421 | ActiveClockChangeLatencyHidingC = EffectiveLBLatencyHidingC + FullDETBufferingTimeC |
| 4422 | - (DSTXAfterScaler[k] / v->HTotal[k] + DSTYAfterScaler[k]) * v->HTotal[k] |
| 4423 | / v->PixelClock[k]; |
| 4424 | if (v->NumberOfActiveSurfaces > 1) { |
| 4425 | ActiveClockChangeLatencyHidingC = ActiveClockChangeLatencyHidingC |
| 4426 | - (1 - 1 / v->NumberOfActiveSurfaces) * SwathHeightC[k] * v->HTotal[k] |
| 4427 | / v->PixelClock[k] / v->VRatioChroma[k]; |
| 4428 | } |
| 4429 | ActiveClockChangeLatencyHiding = dml_min(ActiveClockChangeLatencyHidingY, |
| 4430 | ActiveClockChangeLatencyHidingC); |
| 4431 | } else { |
| 4432 | ActiveClockChangeLatencyHiding = ActiveClockChangeLatencyHidingY; |
| 4433 | } |
| 4434 | |
| 4435 | ActiveDRAMClockChangeLatencyMargin[k] = ActiveClockChangeLatencyHiding - v->Watermark.UrgentWatermark |
| 4436 | - v->Watermark.DRAMClockChangeWatermark; |
| 4437 | ActiveFCLKChangeLatencyMargin[k] = ActiveClockChangeLatencyHiding - v->Watermark.UrgentWatermark |
| 4438 | - v->Watermark.FCLKChangeWatermark; |
| 4439 | USRRetrainingLatencyMargin[k] = ActiveClockChangeLatencyHiding - v->Watermark.USRRetrainingWatermark; |
| 4440 | |
| 4441 | if (v->WritebackEnable[k]) { |
| 4442 | WritebackLatencyHiding = v->WritebackInterfaceBufferSize * 1024 |
| 4443 | / (v->WritebackDestinationWidth[k] * v->WritebackDestinationHeight[k] |
| 4444 | / (v->WritebackSourceHeight[k] * v->HTotal[k] / v->PixelClock[k]) * 4); |
| 4445 | if (v->WritebackPixelFormat[k] == dm_444_64) |
| 4446 | WritebackLatencyHiding = WritebackLatencyHiding / 2; |
| 4447 | |
| 4448 | WritebackDRAMClockChangeLatencyMargin = WritebackLatencyHiding |
| 4449 | - v->Watermark.WritebackDRAMClockChangeWatermark; |
| 4450 | |
| 4451 | WritebackFCLKChangeLatencyMargin = WritebackLatencyHiding |
| 4452 | - v->Watermark.WritebackFCLKChangeWatermark; |
| 4453 | |
| 4454 | ActiveDRAMClockChangeLatencyMargin[k] = dml_min(ActiveDRAMClockChangeLatencyMargin[k], |
| 4455 | WritebackFCLKChangeLatencyMargin); |
| 4456 | ActiveFCLKChangeLatencyMargin[k] = dml_min(ActiveFCLKChangeLatencyMargin[k], |
| 4457 | WritebackDRAMClockChangeLatencyMargin); |
| 4458 | } |
| 4459 | MaxActiveDRAMClockChangeLatencySupported[k] = |
| 4460 | (v->UsesMALLForPStateChange[k] == dm_use_mall_pstate_change_phantom_pipe) ? |
| 4461 | 0 : |
| 4462 | (ActiveDRAMClockChangeLatencyMargin[k] |
| 4463 | + mmSOCParameters.DRAMClockChangeLatency); |
| 4464 | } |
| 4465 | |
| 4466 | for (i = 0; i < v->NumberOfActiveSurfaces; ++i) { |
| 4467 | for (j = 0; j < v->NumberOfActiveSurfaces; ++j) { |
| 4468 | if (i == j || |
| 4469 | (v->BlendingAndTiming[i] == i && v->BlendingAndTiming[j] == i) || |
| 4470 | (v->BlendingAndTiming[j] == j && v->BlendingAndTiming[i] == j) || |
| 4471 | (v->BlendingAndTiming[i] == v->BlendingAndTiming[j] && v->BlendingAndTiming[i] != i) || |
| 4472 | (v->SynchronizeTimingsFinal && v->PixelClock[i] == v->PixelClock[j] && |
| 4473 | v->HTotal[i] == v->HTotal[j] && v->VTotal[i] == v->VTotal[j] && |
| 4474 | v->VActive[i] == v->VActive[j]) || (v->SynchronizeDRRDisplaysForUCLKPStateChangeFinal && |
| 4475 | (v->DRRDisplay[i] || v->DRRDisplay[j]))) { |
| 4476 | SynchronizedSurfaces[i][j] = true1; |
| 4477 | } else { |
| 4478 | SynchronizedSurfaces[i][j] = false0; |
| 4479 | } |
| 4480 | } |
| 4481 | } |
| 4482 | |
| 4483 | for (k = 0; k < v->NumberOfActiveSurfaces; ++k) { |
| 4484 | if ((v->UsesMALLForPStateChange[k] != dm_use_mall_pstate_change_phantom_pipe) && |
| 4485 | (!FoundFirstSurfaceWithMinActiveFCLKChangeMargin || |
| 4486 | ActiveFCLKChangeLatencyMargin[k] < MinActiveFCLKChangeMargin)) { |
| 4487 | FoundFirstSurfaceWithMinActiveFCLKChangeMargin = true1; |
| 4488 | MinActiveFCLKChangeMargin = ActiveFCLKChangeLatencyMargin[k]; |
| 4489 | SurfaceWithMinActiveFCLKChangeMargin = k; |
| 4490 | } |
| 4491 | } |
| 4492 | |
| 4493 | *MinActiveFCLKChangeLatencySupported = MinActiveFCLKChangeMargin + mmSOCParameters.FCLKChangeLatency; |
| 4494 | |
| 4495 | SameTimingForFCLKChange = true1; |
| 4496 | for (k = 0; k < v->NumberOfActiveSurfaces; ++k) { |
| 4497 | if (!SynchronizedSurfaces[k][SurfaceWithMinActiveFCLKChangeMargin]) { |
| 4498 | if ((v->UsesMALLForPStateChange[k] != dm_use_mall_pstate_change_phantom_pipe) && |
| 4499 | (SameTimingForFCLKChange || |
| 4500 | ActiveFCLKChangeLatencyMargin[k] < |
| 4501 | SecondMinActiveFCLKChangeMarginOneDisplayInVBLank)) { |
| 4502 | SecondMinActiveFCLKChangeMarginOneDisplayInVBLank = ActiveFCLKChangeLatencyMargin[k]; |
| 4503 | } |
| 4504 | SameTimingForFCLKChange = false0; |
| 4505 | } |
| 4506 | } |
| 4507 | |
| 4508 | if (MinActiveFCLKChangeMargin > 0) { |
| 4509 | *FCLKChangeSupport = dm_fclock_change_vactive; |
| 4510 | } else if ((SameTimingForFCLKChange || SecondMinActiveFCLKChangeMarginOneDisplayInVBLank > 0) && |
| 4511 | (PrefetchMode <= 1)) { |
| 4512 | *FCLKChangeSupport = dm_fclock_change_vblank; |
| 4513 | } else { |
| 4514 | *FCLKChangeSupport = dm_fclock_change_unsupported; |
| 4515 | } |
| 4516 | |
| 4517 | *USRRetrainingSupport = true1; |
| 4518 | for (k = 0; k < v->NumberOfActiveSurfaces; ++k) { |
| 4519 | if ((v->UsesMALLForPStateChange[k] != dm_use_mall_pstate_change_phantom_pipe) && |
| 4520 | (USRRetrainingLatencyMargin[k] < 0)) { |
| 4521 | *USRRetrainingSupport = false0; |
| 4522 | } |
| 4523 | } |
| 4524 | |
| 4525 | for (k = 0; k < v->NumberOfActiveSurfaces; ++k) { |
| 4526 | if (v->UsesMALLForPStateChange[k] != dm_use_mall_pstate_change_full_frame && |
| 4527 | v->UsesMALLForPStateChange[k] != dm_use_mall_pstate_change_sub_viewport && |
| 4528 | v->UsesMALLForPStateChange[k] != dm_use_mall_pstate_change_phantom_pipe && |
| 4529 | ActiveDRAMClockChangeLatencyMargin[k] < 0) { |
| 4530 | if (PrefetchMode > 0) { |
| 4531 | DRAMClockChangeSupportNumber = 2; |
| 4532 | } else if (DRAMClockChangeSupportNumber == 0) { |
| 4533 | DRAMClockChangeSupportNumber = 1; |
| 4534 | LastSurfaceWithoutMargin = k; |
| 4535 | } else if (DRAMClockChangeSupportNumber == 1 && |
| 4536 | !SynchronizedSurfaces[LastSurfaceWithoutMargin][k]) { |
| 4537 | DRAMClockChangeSupportNumber = 2; |
| 4538 | } |
| 4539 | } |
| 4540 | } |
| 4541 | |
| 4542 | for (k = 0; k < v->NumberOfActiveSurfaces; ++k) { |
| 4543 | if (v->UsesMALLForPStateChange[k] == dm_use_mall_pstate_change_full_frame) |
| 4544 | DRAMClockChangeMethod = 1; |
| 4545 | else if (v->UsesMALLForPStateChange[k] == dm_use_mall_pstate_change_sub_viewport) |
| 4546 | DRAMClockChangeMethod = 2; |
| 4547 | } |
| 4548 | |
| 4549 | if (DRAMClockChangeMethod == 0) { |
| 4550 | if (DRAMClockChangeSupportNumber == 0) |
| 4551 | *DRAMClockChangeSupport = dm_dram_clock_change_vactive; |
| 4552 | else if (DRAMClockChangeSupportNumber == 1) |
| 4553 | *DRAMClockChangeSupport = dm_dram_clock_change_vblank; |
| 4554 | else |
| 4555 | *DRAMClockChangeSupport = dm_dram_clock_change_unsupported; |
| 4556 | } else if (DRAMClockChangeMethod == 1) { |
| 4557 | if (DRAMClockChangeSupportNumber == 0) |
| 4558 | *DRAMClockChangeSupport = dm_dram_clock_change_vactive_w_mall_full_frame; |
| 4559 | else if (DRAMClockChangeSupportNumber == 1) |
| 4560 | *DRAMClockChangeSupport = dm_dram_clock_change_vblank_w_mall_full_frame; |
| 4561 | else |
| 4562 | *DRAMClockChangeSupport = dm_dram_clock_change_unsupported; |
| 4563 | } else { |
| 4564 | if (DRAMClockChangeSupportNumber == 0) |
| 4565 | *DRAMClockChangeSupport = dm_dram_clock_change_vactive_w_mall_sub_vp; |
| 4566 | else if (DRAMClockChangeSupportNumber == 1) |
| 4567 | *DRAMClockChangeSupport = dm_dram_clock_change_vblank_w_mall_sub_vp; |
| 4568 | else |
| 4569 | *DRAMClockChangeSupport = dm_dram_clock_change_unsupported; |
| 4570 | } |
| 4571 | |
| 4572 | for (k = 0; k < v->NumberOfActiveSurfaces; ++k) { |
| 4573 | unsigned int dst_y_pstate; |
| 4574 | unsigned int src_y_pstate_l; |
| 4575 | unsigned int src_y_pstate_c; |
| 4576 | unsigned int src_y_ahead_l, src_y_ahead_c, sub_vp_lines_l, sub_vp_lines_c; |
| 4577 | |
| 4578 | dst_y_pstate = dml_ceil((mmSOCParameters.DRAMClockChangeLatency + mmSOCParameters.UrgentLatency) / (v->HTotal[k] / v->PixelClock[k]), 1); |
| 4579 | src_y_pstate_l = dml_ceil(dst_y_pstate * v->VRatio[k], SwathHeightY[k]); |
| 4580 | src_y_ahead_l = dml_floor(DETBufferSizeY[k] / BytePerPixelDETY[k] / SwathWidthY[k], SwathHeightY[k]) + LBLatencyHidingSourceLinesY[k]; |
| 4581 | sub_vp_lines_l = src_y_pstate_l + src_y_ahead_l + v->meta_row_height[k]; |
| 4582 | |
| 4583 | #ifdef __DML_VBA_DEBUG__ |
| 4584 | dml_print("DML::%s: k=%d, DETBufferSizeY = %d\n", __func__, k, DETBufferSizeY[k]){do { } while(0); }; |
| 4585 | dml_print("DML::%s: k=%d, BytePerPixelDETY = %f\n", __func__, k, BytePerPixelDETY[k]){do { } while(0); }; |
| 4586 | dml_print("DML::%s: k=%d, SwathWidthY = %d\n", __func__, k, SwathWidthY[k]){do { } while(0); }; |
| 4587 | dml_print("DML::%s: k=%d, SwathHeightY = %d\n", __func__, k, SwathHeightY[k]){do { } while(0); }; |
| 4588 | dml_print("DML::%s: k=%d, LBLatencyHidingSourceLinesY = %d\n", __func__, k, LBLatencyHidingSourceLinesY[k]){do { } while(0); }; |
| 4589 | dml_print("DML::%s: k=%d, dst_y_pstate = %d\n", __func__, k, dst_y_pstate){do { } while(0); }; |
| 4590 | dml_print("DML::%s: k=%d, src_y_pstate_l = %d\n", __func__, k, src_y_pstate_l){do { } while(0); }; |
| 4591 | dml_print("DML::%s: k=%d, src_y_ahead_l = %d\n", __func__, k, src_y_ahead_l){do { } while(0); }; |
| 4592 | dml_print("DML::%s: k=%d, v->meta_row_height = %d\n", __func__, k, v->meta_row_height[k]){do { } while(0); }; |
| 4593 | dml_print("DML::%s: k=%d, sub_vp_lines_l = %d\n", __func__, k, sub_vp_lines_l){do { } while(0); }; |
| 4594 | #endif |
| 4595 | SubViewportLinesNeededInMALL[k] = sub_vp_lines_l; |
| 4596 | |
| 4597 | if (BytePerPixelDETC[k] > 0) { |
| 4598 | src_y_pstate_c = dml_ceil(dst_y_pstate * v->VRatioChroma[k], SwathHeightC[k]); |
| 4599 | src_y_ahead_c = dml_floor(DETBufferSizeC[k] / BytePerPixelDETC[k] / SwathWidthC[k], SwathHeightC[k]) + LBLatencyHidingSourceLinesC[k]; |
| 4600 | sub_vp_lines_c = src_y_pstate_c + src_y_ahead_c + v->meta_row_height_chroma[k]; |
| 4601 | SubViewportLinesNeededInMALL[k] = dml_max(sub_vp_lines_l, sub_vp_lines_c); |
| 4602 | |
| 4603 | #ifdef __DML_VBA_DEBUG__ |
| 4604 | dml_print("DML::%s: k=%d, src_y_pstate_c = %d\n", __func__, k, src_y_pstate_c){do { } while(0); }; |
| 4605 | dml_print("DML::%s: k=%d, src_y_ahead_c = %d\n", __func__, k, src_y_ahead_c){do { } while(0); }; |
| 4606 | dml_print("DML::%s: k=%d, v->meta_row_height_chroma = %d\n", __func__, k, v->meta_row_height_chroma[k]){do { } while(0); }; |
| 4607 | dml_print("DML::%s: k=%d, sub_vp_lines_c = %d\n", __func__, k, sub_vp_lines_c){do { } while(0); }; |
| 4608 | #endif |
| 4609 | } |
| 4610 | } |
| 4611 | #ifdef __DML_VBA_DEBUG__ |
| 4612 | dml_print("DML::%s: DRAMClockChangeSupport = %d\n", __func__, *DRAMClockChangeSupport){do { } while(0); }; |
| 4613 | dml_print("DML::%s: FCLKChangeSupport = %d\n", __func__, *FCLKChangeSupport){do { } while(0); }; |
| 4614 | dml_print("DML::%s: MinActiveFCLKChangeLatencySupported = %f\n",{do { } while(0); } |
| 4615 | __func__, *MinActiveFCLKChangeLatencySupported){do { } while(0); }; |
| 4616 | dml_print("DML::%s: USRRetrainingSupport = %d\n", __func__, *USRRetrainingSupport){do { } while(0); }; |
| 4617 | #endif |
| 4618 | } // CalculateWatermarksMALLUseAndDRAMSpeedChangeSupport |
| 4619 | |
| 4620 | double dml32_CalculateWriteBackDISPCLK( |
| 4621 | enum source_format_class WritebackPixelFormat, |
| 4622 | double PixelClock, |
| 4623 | double WritebackHRatio, |
| 4624 | double WritebackVRatio, |
| 4625 | unsigned int WritebackHTaps, |
| 4626 | unsigned int WritebackVTaps, |
| 4627 | unsigned int WritebackSourceWidth, |
| 4628 | unsigned int WritebackDestinationWidth, |
| 4629 | unsigned int HTotal, |
| 4630 | unsigned int WritebackLineBufferSize, |
| 4631 | double DISPCLKDPPCLKVCOSpeed) |
| 4632 | { |
| 4633 | double DISPCLK_H, DISPCLK_V, DISPCLK_HB; |
| 4634 | |
| 4635 | DISPCLK_H = PixelClock * dml_ceil(WritebackHTaps / 8.0, 1) / WritebackHRatio; |
| 4636 | DISPCLK_V = PixelClock * (WritebackVTaps * dml_ceil(WritebackDestinationWidth / 6.0, 1) + 8.0) / HTotal; |
| 4637 | DISPCLK_HB = PixelClock * WritebackVTaps * (WritebackDestinationWidth * |
| 4638 | WritebackVTaps - WritebackLineBufferSize / 57.0) / 6.0 / WritebackSourceWidth; |
| 4639 | return dml32_RoundToDFSGranularity(dml_max3(DISPCLK_H, DISPCLK_V, DISPCLK_HB), 1, DISPCLKDPPCLKVCOSpeed); |
| 4640 | } |
| 4641 | |
| 4642 | void dml32_CalculateMinAndMaxPrefetchMode( |
| 4643 | enum dm_prefetch_modes AllowForPStateChangeOrStutterInVBlankFinal, |
| 4644 | unsigned int *MinPrefetchMode, |
| 4645 | unsigned int *MaxPrefetchMode) |
| 4646 | { |
| 4647 | if (AllowForPStateChangeOrStutterInVBlankFinal == dm_prefetch_support_none) { |
| 4648 | *MinPrefetchMode = 3; |
| 4649 | *MaxPrefetchMode = 3; |
| 4650 | } else if (AllowForPStateChangeOrStutterInVBlankFinal == dm_prefetch_support_stutter) { |
| 4651 | *MinPrefetchMode = 2; |
| 4652 | *MaxPrefetchMode = 2; |
| 4653 | } else if (AllowForPStateChangeOrStutterInVBlankFinal == dm_prefetch_support_fclk_and_stutter) { |
| 4654 | *MinPrefetchMode = 1; |
| 4655 | *MaxPrefetchMode = 1; |
| 4656 | } else if (AllowForPStateChangeOrStutterInVBlankFinal == dm_prefetch_support_uclk_fclk_and_stutter) { |
| 4657 | *MinPrefetchMode = 0; |
| 4658 | *MaxPrefetchMode = 0; |
| 4659 | } else { |
| 4660 | *MinPrefetchMode = 0; |
| 4661 | *MaxPrefetchMode = 3; |
| 4662 | } |
| 4663 | } // CalculateMinAndMaxPrefetchMode |
| 4664 | |
| 4665 | void dml32_CalculatePixelDeliveryTimes( |
| 4666 | unsigned int NumberOfActiveSurfaces, |
| 4667 | double VRatio[], |
| 4668 | double VRatioChroma[], |
| 4669 | double VRatioPrefetchY[], |
| 4670 | double VRatioPrefetchC[], |
| 4671 | unsigned int swath_width_luma_ub[], |
| 4672 | unsigned int swath_width_chroma_ub[], |
| 4673 | unsigned int DPPPerSurface[], |
| 4674 | double HRatio[], |
| 4675 | double HRatioChroma[], |
| 4676 | double PixelClock[], |
| 4677 | double PSCL_THROUGHPUT[], |
| 4678 | double PSCL_THROUGHPUT_CHROMA[], |
| 4679 | double Dppclk[], |
| 4680 | unsigned int BytePerPixelC[], |
| 4681 | enum dm_rotation_angle SourceRotation[], |
| 4682 | unsigned int NumberOfCursors[], |
| 4683 | unsigned int CursorWidth[][DC__NUM_CURSOR__MAX2], |
| 4684 | unsigned int CursorBPP[][DC__NUM_CURSOR__MAX2], |
| 4685 | unsigned int BlockWidth256BytesY[], |
| 4686 | unsigned int BlockHeight256BytesY[], |
| 4687 | unsigned int BlockWidth256BytesC[], |
| 4688 | unsigned int BlockHeight256BytesC[], |
| 4689 | |
| 4690 | /* Output */ |
| 4691 | double DisplayPipeLineDeliveryTimeLuma[], |
| 4692 | double DisplayPipeLineDeliveryTimeChroma[], |
| 4693 | double DisplayPipeLineDeliveryTimeLumaPrefetch[], |
| 4694 | double DisplayPipeLineDeliveryTimeChromaPrefetch[], |
| 4695 | double DisplayPipeRequestDeliveryTimeLuma[], |
| 4696 | double DisplayPipeRequestDeliveryTimeChroma[], |
| 4697 | double DisplayPipeRequestDeliveryTimeLumaPrefetch[], |
| 4698 | double DisplayPipeRequestDeliveryTimeChromaPrefetch[], |
| 4699 | double CursorRequestDeliveryTime[], |
| 4700 | double CursorRequestDeliveryTimePrefetch[]) |
| 4701 | { |
| 4702 | double req_per_swath_ub; |
| 4703 | unsigned int k; |
| 4704 | |
| 4705 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 4706 | |
| 4707 | #ifdef __DML_VBA_DEBUG__ |
| 4708 | dml_print("DML::%s: k=%d : HRatio = %f\n", __func__, k, HRatio[k]){do { } while(0); }; |
| 4709 | dml_print("DML::%s: k=%d : VRatio = %f\n", __func__, k, VRatio[k]){do { } while(0); }; |
| 4710 | dml_print("DML::%s: k=%d : HRatioChroma = %f\n", __func__, k, HRatioChroma[k]){do { } while(0); }; |
| 4711 | dml_print("DML::%s: k=%d : VRatioChroma = %f\n", __func__, k, VRatioChroma[k]){do { } while(0); }; |
| 4712 | dml_print("DML::%s: k=%d : swath_width_luma_ub = %d\n", __func__, k, swath_width_luma_ub[k]){do { } while(0); }; |
| 4713 | dml_print("DML::%s: k=%d : swath_width_chroma_ub = %d\n", __func__, k, swath_width_chroma_ub[k]){do { } while(0); }; |
| 4714 | dml_print("DML::%s: k=%d : PSCL_THROUGHPUT = %f\n", __func__, k, PSCL_THROUGHPUT[k]){do { } while(0); }; |
| 4715 | dml_print("DML::%s: k=%d : PSCL_THROUGHPUT_CHROMA = %f\n", __func__, k, PSCL_THROUGHPUT_CHROMA[k]){do { } while(0); }; |
| 4716 | dml_print("DML::%s: k=%d : DPPPerSurface = %d\n", __func__, k, DPPPerSurface[k]){do { } while(0); }; |
| 4717 | dml_print("DML::%s: k=%d : PixelClock = %f\n", __func__, k, PixelClock[k]){do { } while(0); }; |
| 4718 | dml_print("DML::%s: k=%d : Dppclk = %f\n", __func__, k, Dppclk[k]){do { } while(0); }; |
| 4719 | #endif |
| 4720 | |
| 4721 | if (VRatio[k] <= 1) { |
| 4722 | DisplayPipeLineDeliveryTimeLuma[k] = |
| 4723 | swath_width_luma_ub[k] * DPPPerSurface[k] / HRatio[k] / PixelClock[k]; |
| 4724 | } else { |
| 4725 | DisplayPipeLineDeliveryTimeLuma[k] = swath_width_luma_ub[k] / PSCL_THROUGHPUT[k] / Dppclk[k]; |
| 4726 | } |
| 4727 | |
| 4728 | if (BytePerPixelC[k] == 0) { |
| 4729 | DisplayPipeLineDeliveryTimeChroma[k] = 0; |
| 4730 | } else { |
| 4731 | if (VRatioChroma[k] <= 1) { |
| 4732 | DisplayPipeLineDeliveryTimeChroma[k] = |
| 4733 | swath_width_chroma_ub[k] * DPPPerSurface[k] / HRatioChroma[k] / PixelClock[k]; |
| 4734 | } else { |
| 4735 | DisplayPipeLineDeliveryTimeChroma[k] = |
| 4736 | swath_width_chroma_ub[k] / PSCL_THROUGHPUT_CHROMA[k] / Dppclk[k]; |
| 4737 | } |
| 4738 | } |
| 4739 | |
| 4740 | if (VRatioPrefetchY[k] <= 1) { |
| 4741 | DisplayPipeLineDeliveryTimeLumaPrefetch[k] = |
| 4742 | swath_width_luma_ub[k] * DPPPerSurface[k] / HRatio[k] / PixelClock[k]; |
| 4743 | } else { |
| 4744 | DisplayPipeLineDeliveryTimeLumaPrefetch[k] = |
| 4745 | swath_width_luma_ub[k] / PSCL_THROUGHPUT[k] / Dppclk[k]; |
| 4746 | } |
| 4747 | |
| 4748 | if (BytePerPixelC[k] == 0) { |
| 4749 | DisplayPipeLineDeliveryTimeChromaPrefetch[k] = 0; |
| 4750 | } else { |
| 4751 | if (VRatioPrefetchC[k] <= 1) { |
| 4752 | DisplayPipeLineDeliveryTimeChromaPrefetch[k] = swath_width_chroma_ub[k] * |
| 4753 | DPPPerSurface[k] / HRatioChroma[k] / PixelClock[k]; |
| 4754 | } else { |
| 4755 | DisplayPipeLineDeliveryTimeChromaPrefetch[k] = |
| 4756 | swath_width_chroma_ub[k] / PSCL_THROUGHPUT_CHROMA[k] / Dppclk[k]; |
| 4757 | } |
| 4758 | } |
| 4759 | #ifdef __DML_VBA_DEBUG__ |
| 4760 | dml_print("DML::%s: k=%d : DisplayPipeLineDeliveryTimeLuma = %f\n",{do { } while(0); } |
| 4761 | __func__, k, DisplayPipeLineDeliveryTimeLuma[k]){do { } while(0); }; |
| 4762 | dml_print("DML::%s: k=%d : DisplayPipeLineDeliveryTimeLumaPrefetch = %f\n",{do { } while(0); } |
| 4763 | __func__, k, DisplayPipeLineDeliveryTimeLumaPrefetch[k]){do { } while(0); }; |
| 4764 | dml_print("DML::%s: k=%d : DisplayPipeLineDeliveryTimeChroma = %f\n",{do { } while(0); } |
| 4765 | __func__, k, DisplayPipeLineDeliveryTimeChroma[k]){do { } while(0); }; |
| 4766 | dml_print("DML::%s: k=%d : DisplayPipeLineDeliveryTimeChromaPrefetch = %f\n",{do { } while(0); } |
| 4767 | __func__, k, DisplayPipeLineDeliveryTimeChromaPrefetch[k]){do { } while(0); }; |
| 4768 | #endif |
| 4769 | } |
| 4770 | |
| 4771 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 4772 | if (!IsVertical(SourceRotation[k])) |
| 4773 | req_per_swath_ub = swath_width_luma_ub[k] / BlockWidth256BytesY[k]; |
| 4774 | else |
| 4775 | req_per_swath_ub = swath_width_luma_ub[k] / BlockHeight256BytesY[k]; |
| 4776 | #ifdef __DML_VBA_DEBUG__ |
| 4777 | dml_print("DML::%s: k=%d : req_per_swath_ub = %f (Luma)\n", __func__, k, req_per_swath_ub){do { } while(0); }; |
| 4778 | #endif |
| 4779 | |
| 4780 | DisplayPipeRequestDeliveryTimeLuma[k] = DisplayPipeLineDeliveryTimeLuma[k] / req_per_swath_ub; |
| 4781 | DisplayPipeRequestDeliveryTimeLumaPrefetch[k] = |
| 4782 | DisplayPipeLineDeliveryTimeLumaPrefetch[k] / req_per_swath_ub; |
| 4783 | if (BytePerPixelC[k] == 0) { |
| 4784 | DisplayPipeRequestDeliveryTimeChroma[k] = 0; |
| 4785 | DisplayPipeRequestDeliveryTimeChromaPrefetch[k] = 0; |
| 4786 | } else { |
| 4787 | if (!IsVertical(SourceRotation[k])) |
| 4788 | req_per_swath_ub = swath_width_chroma_ub[k] / BlockWidth256BytesC[k]; |
| 4789 | else |
| 4790 | req_per_swath_ub = swath_width_chroma_ub[k] / BlockHeight256BytesC[k]; |
| 4791 | #ifdef __DML_VBA_DEBUG__ |
| 4792 | dml_print("DML::%s: k=%d : req_per_swath_ub = %f (Chroma)\n", __func__, k, req_per_swath_ub){do { } while(0); }; |
| 4793 | #endif |
| 4794 | DisplayPipeRequestDeliveryTimeChroma[k] = |
| 4795 | DisplayPipeLineDeliveryTimeChroma[k] / req_per_swath_ub; |
| 4796 | DisplayPipeRequestDeliveryTimeChromaPrefetch[k] = |
| 4797 | DisplayPipeLineDeliveryTimeChromaPrefetch[k] / req_per_swath_ub; |
| 4798 | } |
| 4799 | #ifdef __DML_VBA_DEBUG__ |
| 4800 | dml_print("DML::%s: k=%d : DisplayPipeRequestDeliveryTimeLuma = %f\n",{do { } while(0); } |
| 4801 | __func__, k, DisplayPipeRequestDeliveryTimeLuma[k]){do { } while(0); }; |
| 4802 | dml_print("DML::%s: k=%d : DisplayPipeRequestDeliveryTimeLumaPrefetch = %f\n",{do { } while(0); } |
| 4803 | __func__, k, DisplayPipeRequestDeliveryTimeLumaPrefetch[k]){do { } while(0); }; |
| 4804 | dml_print("DML::%s: k=%d : DisplayPipeRequestDeliveryTimeChroma = %f\n",{do { } while(0); } |
| 4805 | __func__, k, DisplayPipeRequestDeliveryTimeChroma[k]){do { } while(0); }; |
| 4806 | dml_print("DML::%s: k=%d : DisplayPipeRequestDeliveryTimeChromaPrefetch = %f\n",{do { } while(0); } |
| 4807 | __func__, k, DisplayPipeRequestDeliveryTimeChromaPrefetch[k]){do { } while(0); }; |
| 4808 | #endif |
| 4809 | } |
| 4810 | |
| 4811 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 4812 | unsigned int cursor_req_per_width; |
| 4813 | |
| 4814 | cursor_req_per_width = dml_ceil((double) CursorWidth[k][0] * (double) CursorBPP[k][0] / |
| 4815 | 256.0 / 8.0, 1.0); |
| 4816 | if (NumberOfCursors[k] > 0) { |
| 4817 | if (VRatio[k] <= 1) { |
| 4818 | CursorRequestDeliveryTime[k] = (double) CursorWidth[k][0] / |
| 4819 | HRatio[k] / PixelClock[k] / cursor_req_per_width; |
| 4820 | } else { |
| 4821 | CursorRequestDeliveryTime[k] = (double) CursorWidth[k][0] / |
| 4822 | PSCL_THROUGHPUT[k] / Dppclk[k] / cursor_req_per_width; |
| 4823 | } |
| 4824 | if (VRatioPrefetchY[k] <= 1) { |
| 4825 | CursorRequestDeliveryTimePrefetch[k] = (double) CursorWidth[k][0] / |
| 4826 | HRatio[k] / PixelClock[k] / cursor_req_per_width; |
| 4827 | } else { |
| 4828 | CursorRequestDeliveryTimePrefetch[k] = (double) CursorWidth[k][0] / |
| 4829 | PSCL_THROUGHPUT[k] / Dppclk[k] / cursor_req_per_width; |
| 4830 | } |
| 4831 | } else { |
| 4832 | CursorRequestDeliveryTime[k] = 0; |
| 4833 | CursorRequestDeliveryTimePrefetch[k] = 0; |
| 4834 | } |
| 4835 | #ifdef __DML_VBA_DEBUG__ |
| 4836 | dml_print("DML::%s: k=%d : NumberOfCursors = %d\n",{do { } while(0); } |
| 4837 | __func__, k, NumberOfCursors[k]){do { } while(0); }; |
| 4838 | dml_print("DML::%s: k=%d : CursorRequestDeliveryTime = %f\n",{do { } while(0); } |
| 4839 | __func__, k, CursorRequestDeliveryTime[k]){do { } while(0); }; |
| 4840 | dml_print("DML::%s: k=%d : CursorRequestDeliveryTimePrefetch = %f\n",{do { } while(0); } |
| 4841 | __func__, k, CursorRequestDeliveryTimePrefetch[k]){do { } while(0); }; |
| 4842 | #endif |
| 4843 | } |
| 4844 | } // CalculatePixelDeliveryTimes |
| 4845 | |
| 4846 | void dml32_CalculateMetaAndPTETimes( |
| 4847 | bool_Bool use_one_row_for_frame[], |
| 4848 | unsigned int NumberOfActiveSurfaces, |
| 4849 | bool_Bool GPUVMEnable, |
| 4850 | unsigned int MetaChunkSize, |
| 4851 | unsigned int MinMetaChunkSizeBytes, |
| 4852 | unsigned int HTotal[], |
| 4853 | double VRatio[], |
| 4854 | double VRatioChroma[], |
| 4855 | double DestinationLinesToRequestRowInVBlank[], |
| 4856 | double DestinationLinesToRequestRowInImmediateFlip[], |
| 4857 | bool_Bool DCCEnable[], |
| 4858 | double PixelClock[], |
| 4859 | unsigned int BytePerPixelY[], |
| 4860 | unsigned int BytePerPixelC[], |
| 4861 | enum dm_rotation_angle SourceRotation[], |
| 4862 | unsigned int dpte_row_height[], |
| 4863 | unsigned int dpte_row_height_chroma[], |
| 4864 | unsigned int meta_row_width[], |
| 4865 | unsigned int meta_row_width_chroma[], |
| 4866 | unsigned int meta_row_height[], |
| 4867 | unsigned int meta_row_height_chroma[], |
| 4868 | unsigned int meta_req_width[], |
| 4869 | unsigned int meta_req_width_chroma[], |
| 4870 | unsigned int meta_req_height[], |
| 4871 | unsigned int meta_req_height_chroma[], |
| 4872 | unsigned int dpte_group_bytes[], |
| 4873 | unsigned int PTERequestSizeY[], |
| 4874 | unsigned int PTERequestSizeC[], |
| 4875 | unsigned int PixelPTEReqWidthY[], |
| 4876 | unsigned int PixelPTEReqHeightY[], |
| 4877 | unsigned int PixelPTEReqWidthC[], |
| 4878 | unsigned int PixelPTEReqHeightC[], |
| 4879 | unsigned int dpte_row_width_luma_ub[], |
| 4880 | unsigned int dpte_row_width_chroma_ub[], |
| 4881 | |
| 4882 | /* Output */ |
| 4883 | double DST_Y_PER_PTE_ROW_NOM_L[], |
| 4884 | double DST_Y_PER_PTE_ROW_NOM_C[], |
| 4885 | double DST_Y_PER_META_ROW_NOM_L[], |
| 4886 | double DST_Y_PER_META_ROW_NOM_C[], |
| 4887 | double TimePerMetaChunkNominal[], |
| 4888 | double TimePerChromaMetaChunkNominal[], |
| 4889 | double TimePerMetaChunkVBlank[], |
| 4890 | double TimePerChromaMetaChunkVBlank[], |
| 4891 | double TimePerMetaChunkFlip[], |
| 4892 | double TimePerChromaMetaChunkFlip[], |
| 4893 | double time_per_pte_group_nom_luma[], |
| 4894 | double time_per_pte_group_vblank_luma[], |
| 4895 | double time_per_pte_group_flip_luma[], |
| 4896 | double time_per_pte_group_nom_chroma[], |
| 4897 | double time_per_pte_group_vblank_chroma[], |
| 4898 | double time_per_pte_group_flip_chroma[]) |
| 4899 | { |
| 4900 | unsigned int meta_chunk_width; |
| 4901 | unsigned int min_meta_chunk_width; |
| 4902 | unsigned int meta_chunk_per_row_int; |
| 4903 | unsigned int meta_row_remainder; |
| 4904 | unsigned int meta_chunk_threshold; |
| 4905 | unsigned int meta_chunks_per_row_ub; |
| 4906 | unsigned int meta_chunk_width_chroma; |
| 4907 | unsigned int min_meta_chunk_width_chroma; |
| 4908 | unsigned int meta_chunk_per_row_int_chroma; |
| 4909 | unsigned int meta_row_remainder_chroma; |
| 4910 | unsigned int meta_chunk_threshold_chroma; |
| 4911 | unsigned int meta_chunks_per_row_ub_chroma; |
| 4912 | unsigned int dpte_group_width_luma; |
| 4913 | unsigned int dpte_groups_per_row_luma_ub; |
| 4914 | unsigned int dpte_group_width_chroma; |
| 4915 | unsigned int dpte_groups_per_row_chroma_ub; |
| 4916 | unsigned int k; |
| 4917 | |
| 4918 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 4919 | DST_Y_PER_PTE_ROW_NOM_L[k] = dpte_row_height[k] / VRatio[k]; |
| 4920 | if (BytePerPixelC[k] == 0) |
| 4921 | DST_Y_PER_PTE_ROW_NOM_C[k] = 0; |
| 4922 | else |
| 4923 | DST_Y_PER_PTE_ROW_NOM_C[k] = dpte_row_height_chroma[k] / VRatioChroma[k]; |
| 4924 | DST_Y_PER_META_ROW_NOM_L[k] = meta_row_height[k] / VRatio[k]; |
| 4925 | if (BytePerPixelC[k] == 0) |
| 4926 | DST_Y_PER_META_ROW_NOM_C[k] = 0; |
| 4927 | else |
| 4928 | DST_Y_PER_META_ROW_NOM_C[k] = meta_row_height_chroma[k] / VRatioChroma[k]; |
| 4929 | } |
| 4930 | |
| 4931 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 4932 | if (DCCEnable[k] == true1) { |
| 4933 | meta_chunk_width = MetaChunkSize * 1024 * 256 / BytePerPixelY[k] / meta_row_height[k]; |
| 4934 | min_meta_chunk_width = MinMetaChunkSizeBytes * 256 / BytePerPixelY[k] / meta_row_height[k]; |
| 4935 | meta_chunk_per_row_int = meta_row_width[k] / meta_chunk_width; |
| 4936 | meta_row_remainder = meta_row_width[k] % meta_chunk_width; |
| 4937 | if (!IsVertical(SourceRotation[k])) |
| 4938 | meta_chunk_threshold = 2 * min_meta_chunk_width - meta_req_width[k]; |
| 4939 | else |
| 4940 | meta_chunk_threshold = 2 * min_meta_chunk_width - meta_req_height[k]; |
| 4941 | |
| 4942 | if (meta_row_remainder <= meta_chunk_threshold) |
| 4943 | meta_chunks_per_row_ub = meta_chunk_per_row_int + 1; |
| 4944 | else |
| 4945 | meta_chunks_per_row_ub = meta_chunk_per_row_int + 2; |
| 4946 | |
| 4947 | TimePerMetaChunkNominal[k] = meta_row_height[k] / VRatio[k] * |
| 4948 | HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub; |
| 4949 | TimePerMetaChunkVBlank[k] = DestinationLinesToRequestRowInVBlank[k] * |
| 4950 | HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub; |
| 4951 | TimePerMetaChunkFlip[k] = DestinationLinesToRequestRowInImmediateFlip[k] * |
| 4952 | HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub; |
| 4953 | if (BytePerPixelC[k] == 0) { |
| 4954 | TimePerChromaMetaChunkNominal[k] = 0; |
| 4955 | TimePerChromaMetaChunkVBlank[k] = 0; |
| 4956 | TimePerChromaMetaChunkFlip[k] = 0; |
| 4957 | } else { |
| 4958 | meta_chunk_width_chroma = MetaChunkSize * 1024 * 256 / BytePerPixelC[k] / |
| 4959 | meta_row_height_chroma[k]; |
| 4960 | min_meta_chunk_width_chroma = MinMetaChunkSizeBytes * 256 / BytePerPixelC[k] / |
| 4961 | meta_row_height_chroma[k]; |
| 4962 | meta_chunk_per_row_int_chroma = (double) meta_row_width_chroma[k] / |
| 4963 | meta_chunk_width_chroma; |
| 4964 | meta_row_remainder_chroma = meta_row_width_chroma[k] % meta_chunk_width_chroma; |
| 4965 | if (!IsVertical(SourceRotation[k])) { |
| 4966 | meta_chunk_threshold_chroma = 2 * min_meta_chunk_width_chroma - |
| 4967 | meta_req_width_chroma[k]; |
| 4968 | } else { |
| 4969 | meta_chunk_threshold_chroma = 2 * min_meta_chunk_width_chroma - |
| 4970 | meta_req_height_chroma[k]; |
| 4971 | } |
| 4972 | if (meta_row_remainder_chroma <= meta_chunk_threshold_chroma) |
| 4973 | meta_chunks_per_row_ub_chroma = meta_chunk_per_row_int_chroma + 1; |
| 4974 | else |
| 4975 | meta_chunks_per_row_ub_chroma = meta_chunk_per_row_int_chroma + 2; |
| 4976 | |
| 4977 | TimePerChromaMetaChunkNominal[k] = meta_row_height_chroma[k] / VRatioChroma[k] * |
| 4978 | HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub_chroma; |
| 4979 | TimePerChromaMetaChunkVBlank[k] = DestinationLinesToRequestRowInVBlank[k] * |
| 4980 | HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub_chroma; |
| 4981 | TimePerChromaMetaChunkFlip[k] = DestinationLinesToRequestRowInImmediateFlip[k] * |
| 4982 | HTotal[k] / PixelClock[k] / meta_chunks_per_row_ub_chroma; |
| 4983 | } |
| 4984 | } else { |
| 4985 | TimePerMetaChunkNominal[k] = 0; |
| 4986 | TimePerMetaChunkVBlank[k] = 0; |
| 4987 | TimePerMetaChunkFlip[k] = 0; |
| 4988 | TimePerChromaMetaChunkNominal[k] = 0; |
| 4989 | TimePerChromaMetaChunkVBlank[k] = 0; |
| 4990 | TimePerChromaMetaChunkFlip[k] = 0; |
| 4991 | } |
| 4992 | } |
| 4993 | |
| 4994 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 4995 | if (GPUVMEnable == true1) { |
| 4996 | if (!IsVertical(SourceRotation[k])) { |
| 4997 | dpte_group_width_luma = (double) dpte_group_bytes[k] / |
| 4998 | (double) PTERequestSizeY[k] * PixelPTEReqWidthY[k]; |
| 4999 | } else { |
| 5000 | dpte_group_width_luma = (double) dpte_group_bytes[k] / |
| 5001 | (double) PTERequestSizeY[k] * PixelPTEReqHeightY[k]; |
| 5002 | } |
| 5003 | |
| 5004 | if (use_one_row_for_frame[k]) { |
| 5005 | dpte_groups_per_row_luma_ub = dml_ceil((double) dpte_row_width_luma_ub[k] / |
| 5006 | (double) dpte_group_width_luma / 2.0, 1.0); |
| 5007 | } else { |
| 5008 | dpte_groups_per_row_luma_ub = dml_ceil((double) dpte_row_width_luma_ub[k] / |
| 5009 | (double) dpte_group_width_luma, 1.0); |
| 5010 | } |
| 5011 | #ifdef __DML_VBA_DEBUG__ |
| 5012 | dml_print("DML::%s: k=%0d, use_one_row_for_frame = %d\n",{do { } while(0); } |
| 5013 | __func__, k, use_one_row_for_frame[k]){do { } while(0); }; |
| 5014 | dml_print("DML::%s: k=%0d, dpte_group_bytes = %d\n",{do { } while(0); } |
| 5015 | __func__, k, dpte_group_bytes[k]){do { } while(0); }; |
| 5016 | dml_print("DML::%s: k=%0d, PTERequestSizeY = %d\n",{do { } while(0); } |
| 5017 | __func__, k, PTERequestSizeY[k]){do { } while(0); }; |
| 5018 | dml_print("DML::%s: k=%0d, PixelPTEReqWidthY = %d\n",{do { } while(0); } |
| 5019 | __func__, k, PixelPTEReqWidthY[k]){do { } while(0); }; |
| 5020 | dml_print("DML::%s: k=%0d, PixelPTEReqHeightY = %d\n",{do { } while(0); } |
| 5021 | __func__, k, PixelPTEReqHeightY[k]){do { } while(0); }; |
| 5022 | dml_print("DML::%s: k=%0d, dpte_row_width_luma_ub = %d\n",{do { } while(0); } |
| 5023 | __func__, k, dpte_row_width_luma_ub[k]){do { } while(0); }; |
| 5024 | dml_print("DML::%s: k=%0d, dpte_group_width_luma = %d\n",{do { } while(0); } |
| 5025 | __func__, k, dpte_group_width_luma){do { } while(0); }; |
| 5026 | dml_print("DML::%s: k=%0d, dpte_groups_per_row_luma_ub = %d\n",{do { } while(0); } |
| 5027 | __func__, k, dpte_groups_per_row_luma_ub){do { } while(0); }; |
| 5028 | #endif |
| 5029 | |
| 5030 | time_per_pte_group_nom_luma[k] = DST_Y_PER_PTE_ROW_NOM_L[k] * |
| 5031 | HTotal[k] / PixelClock[k] / dpte_groups_per_row_luma_ub; |
| 5032 | time_per_pte_group_vblank_luma[k] = DestinationLinesToRequestRowInVBlank[k] * |
| 5033 | HTotal[k] / PixelClock[k] / dpte_groups_per_row_luma_ub; |
| 5034 | time_per_pte_group_flip_luma[k] = DestinationLinesToRequestRowInImmediateFlip[k] * |
| 5035 | HTotal[k] / PixelClock[k] / dpte_groups_per_row_luma_ub; |
| 5036 | if (BytePerPixelC[k] == 0) { |
| 5037 | time_per_pte_group_nom_chroma[k] = 0; |
| 5038 | time_per_pte_group_vblank_chroma[k] = 0; |
| 5039 | time_per_pte_group_flip_chroma[k] = 0; |
| 5040 | } else { |
| 5041 | if (!IsVertical(SourceRotation[k])) { |
| 5042 | dpte_group_width_chroma = (double) dpte_group_bytes[k] / |
| 5043 | (double) PTERequestSizeC[k] * PixelPTEReqWidthC[k]; |
| 5044 | } else { |
| 5045 | dpte_group_width_chroma = (double) dpte_group_bytes[k] / |
| 5046 | (double) PTERequestSizeC[k] * PixelPTEReqHeightC[k]; |
| 5047 | } |
| 5048 | |
| 5049 | if (use_one_row_for_frame[k]) { |
| 5050 | dpte_groups_per_row_chroma_ub = dml_ceil((double) dpte_row_width_chroma_ub[k] / |
| 5051 | (double) dpte_group_width_chroma / 2.0, 1.0); |
| 5052 | } else { |
| 5053 | dpte_groups_per_row_chroma_ub = dml_ceil((double) dpte_row_width_chroma_ub[k] / |
| 5054 | (double) dpte_group_width_chroma, 1.0); |
| 5055 | } |
| 5056 | #ifdef __DML_VBA_DEBUG__ |
| 5057 | dml_print("DML::%s: k=%0d, dpte_row_width_chroma_ub = %d\n",{do { } while(0); } |
| 5058 | __func__, k, dpte_row_width_chroma_ub[k]){do { } while(0); }; |
| 5059 | dml_print("DML::%s: k=%0d, dpte_group_width_chroma = %d\n",{do { } while(0); } |
| 5060 | __func__, k, dpte_group_width_chroma){do { } while(0); }; |
| 5061 | dml_print("DML::%s: k=%0d, dpte_groups_per_row_chroma_ub = %d\n",{do { } while(0); } |
| 5062 | __func__, k, dpte_groups_per_row_chroma_ub){do { } while(0); }; |
| 5063 | #endif |
| 5064 | time_per_pte_group_nom_chroma[k] = DST_Y_PER_PTE_ROW_NOM_C[k] * |
| 5065 | HTotal[k] / PixelClock[k] / dpte_groups_per_row_chroma_ub; |
| 5066 | time_per_pte_group_vblank_chroma[k] = DestinationLinesToRequestRowInVBlank[k] * |
| 5067 | HTotal[k] / PixelClock[k] / dpte_groups_per_row_chroma_ub; |
| 5068 | time_per_pte_group_flip_chroma[k] = DestinationLinesToRequestRowInImmediateFlip[k] * |
| 5069 | HTotal[k] / PixelClock[k] / dpte_groups_per_row_chroma_ub; |
| 5070 | } |
| 5071 | } else { |
| 5072 | time_per_pte_group_nom_luma[k] = 0; |
| 5073 | time_per_pte_group_vblank_luma[k] = 0; |
| 5074 | time_per_pte_group_flip_luma[k] = 0; |
| 5075 | time_per_pte_group_nom_chroma[k] = 0; |
| 5076 | time_per_pte_group_vblank_chroma[k] = 0; |
| 5077 | time_per_pte_group_flip_chroma[k] = 0; |
| 5078 | } |
| 5079 | #ifdef __DML_VBA_DEBUG__ |
| 5080 | dml_print("DML::%s: k=%0d, DestinationLinesToRequestRowInVBlank = %f\n",{do { } while(0); } |
| 5081 | __func__, k, DestinationLinesToRequestRowInVBlank[k]){do { } while(0); }; |
| 5082 | dml_print("DML::%s: k=%0d, DestinationLinesToRequestRowInImmediateFlip = %f\n",{do { } while(0); } |
| 5083 | __func__, k, DestinationLinesToRequestRowInImmediateFlip[k]){do { } while(0); }; |
| 5084 | dml_print("DML::%s: k=%0d, DST_Y_PER_PTE_ROW_NOM_L = %f\n",{do { } while(0); } |
| 5085 | __func__, k, DST_Y_PER_PTE_ROW_NOM_L[k]){do { } while(0); }; |
| 5086 | dml_print("DML::%s: k=%0d, DST_Y_PER_PTE_ROW_NOM_C = %f\n",{do { } while(0); } |
| 5087 | __func__, k, DST_Y_PER_PTE_ROW_NOM_C[k]){do { } while(0); }; |
| 5088 | dml_print("DML::%s: k=%0d, DST_Y_PER_META_ROW_NOM_L = %f\n",{do { } while(0); } |
| 5089 | __func__, k, DST_Y_PER_META_ROW_NOM_L[k]){do { } while(0); }; |
| 5090 | dml_print("DML::%s: k=%0d, DST_Y_PER_META_ROW_NOM_C = %f\n",{do { } while(0); } |
| 5091 | __func__, k, DST_Y_PER_META_ROW_NOM_C[k]){do { } while(0); }; |
| 5092 | dml_print("DML::%s: k=%0d, TimePerMetaChunkNominal = %f\n",{do { } while(0); } |
| 5093 | __func__, k, TimePerMetaChunkNominal[k]){do { } while(0); }; |
| 5094 | dml_print("DML::%s: k=%0d, TimePerMetaChunkVBlank = %f\n",{do { } while(0); } |
| 5095 | __func__, k, TimePerMetaChunkVBlank[k]){do { } while(0); }; |
| 5096 | dml_print("DML::%s: k=%0d, TimePerMetaChunkFlip = %f\n",{do { } while(0); } |
| 5097 | __func__, k, TimePerMetaChunkFlip[k]){do { } while(0); }; |
| 5098 | dml_print("DML::%s: k=%0d, TimePerChromaMetaChunkNominal = %f\n",{do { } while(0); } |
| 5099 | __func__, k, TimePerChromaMetaChunkNominal[k]){do { } while(0); }; |
| 5100 | dml_print("DML::%s: k=%0d, TimePerChromaMetaChunkVBlank = %f\n",{do { } while(0); } |
| 5101 | __func__, k, TimePerChromaMetaChunkVBlank[k]){do { } while(0); }; |
| 5102 | dml_print("DML::%s: k=%0d, TimePerChromaMetaChunkFlip = %f\n",{do { } while(0); } |
| 5103 | __func__, k, TimePerChromaMetaChunkFlip[k]){do { } while(0); }; |
| 5104 | dml_print("DML::%s: k=%0d, time_per_pte_group_nom_luma = %f\n",{do { } while(0); } |
| 5105 | __func__, k, time_per_pte_group_nom_luma[k]){do { } while(0); }; |
| 5106 | dml_print("DML::%s: k=%0d, time_per_pte_group_vblank_luma = %f\n",{do { } while(0); } |
| 5107 | __func__, k, time_per_pte_group_vblank_luma[k]){do { } while(0); }; |
| 5108 | dml_print("DML::%s: k=%0d, time_per_pte_group_flip_luma = %f\n",{do { } while(0); } |
| 5109 | __func__, k, time_per_pte_group_flip_luma[k]){do { } while(0); }; |
| 5110 | dml_print("DML::%s: k=%0d, time_per_pte_group_nom_chroma = %f\n",{do { } while(0); } |
| 5111 | __func__, k, time_per_pte_group_nom_chroma[k]){do { } while(0); }; |
| 5112 | dml_print("DML::%s: k=%0d, time_per_pte_group_vblank_chroma = %f\n",{do { } while(0); } |
| 5113 | __func__, k, time_per_pte_group_vblank_chroma[k]){do { } while(0); }; |
| 5114 | dml_print("DML::%s: k=%0d, time_per_pte_group_flip_chroma = %f\n",{do { } while(0); } |
| 5115 | __func__, k, time_per_pte_group_flip_chroma[k]){do { } while(0); }; |
| 5116 | #endif |
| 5117 | } |
| 5118 | } // CalculateMetaAndPTETimes |
| 5119 | |
| 5120 | void dml32_CalculateVMGroupAndRequestTimes( |
| 5121 | unsigned int NumberOfActiveSurfaces, |
| 5122 | bool_Bool GPUVMEnable, |
| 5123 | unsigned int GPUVMMaxPageTableLevels, |
| 5124 | unsigned int HTotal[], |
| 5125 | unsigned int BytePerPixelC[], |
| 5126 | double DestinationLinesToRequestVMInVBlank[], |
| 5127 | double DestinationLinesToRequestVMInImmediateFlip[], |
| 5128 | bool_Bool DCCEnable[], |
| 5129 | double PixelClock[], |
| 5130 | unsigned int dpte_row_width_luma_ub[], |
| 5131 | unsigned int dpte_row_width_chroma_ub[], |
| 5132 | unsigned int vm_group_bytes[], |
| 5133 | unsigned int dpde0_bytes_per_frame_ub_l[], |
| 5134 | unsigned int dpde0_bytes_per_frame_ub_c[], |
| 5135 | unsigned int meta_pte_bytes_per_frame_ub_l[], |
| 5136 | unsigned int meta_pte_bytes_per_frame_ub_c[], |
| 5137 | |
| 5138 | /* Output */ |
| 5139 | double TimePerVMGroupVBlank[], |
| 5140 | double TimePerVMGroupFlip[], |
| 5141 | double TimePerVMRequestVBlank[], |
| 5142 | double TimePerVMRequestFlip[]) |
| 5143 | { |
| 5144 | unsigned int k; |
| 5145 | unsigned int num_group_per_lower_vm_stage; |
| 5146 | unsigned int num_req_per_lower_vm_stage; |
| 5147 | |
| 5148 | #ifdef __DML_VBA_DEBUG__ |
| 5149 | dml_print("DML::%s: NumberOfActiveSurfaces = %d\n", __func__, NumberOfActiveSurfaces){do { } while(0); }; |
| 5150 | dml_print("DML::%s: GPUVMEnable = %d\n", __func__, GPUVMEnable){do { } while(0); }; |
| 5151 | #endif |
| 5152 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 5153 | |
| 5154 | #ifdef __DML_VBA_DEBUG__ |
| 5155 | dml_print("DML::%s: k=%0d, DCCEnable = %d\n", __func__, k, DCCEnable[k]){do { } while(0); }; |
| 5156 | dml_print("DML::%s: k=%0d, vm_group_bytes = %d\n", __func__, k, vm_group_bytes[k]){do { } while(0); }; |
| 5157 | dml_print("DML::%s: k=%0d, dpde0_bytes_per_frame_ub_l = %d\n",{do { } while(0); } |
| 5158 | __func__, k, dpde0_bytes_per_frame_ub_l[k]){do { } while(0); }; |
| 5159 | dml_print("DML::%s: k=%0d, dpde0_bytes_per_frame_ub_c = %d\n",{do { } while(0); } |
| 5160 | __func__, k, dpde0_bytes_per_frame_ub_c[k]){do { } while(0); }; |
| 5161 | dml_print("DML::%s: k=%0d, meta_pte_bytes_per_frame_ub_l = %d\n",{do { } while(0); } |
| 5162 | __func__, k, meta_pte_bytes_per_frame_ub_l[k]){do { } while(0); }; |
| 5163 | dml_print("DML::%s: k=%0d, meta_pte_bytes_per_frame_ub_c = %d\n",{do { } while(0); } |
| 5164 | __func__, k, meta_pte_bytes_per_frame_ub_c[k]){do { } while(0); }; |
| 5165 | #endif |
| 5166 | |
| 5167 | if (GPUVMEnable == true1 && (DCCEnable[k] == true1 || GPUVMMaxPageTableLevels > 1)) { |
| 5168 | if (DCCEnable[k] == false0) { |
| 5169 | if (BytePerPixelC[k] > 0) { |
| 5170 | num_group_per_lower_vm_stage = dml_ceil( |
| 5171 | (double) (dpde0_bytes_per_frame_ub_l[k]) / |
| 5172 | (double) (vm_group_bytes[k]), 1.0) + |
| 5173 | dml_ceil((double) (dpde0_bytes_per_frame_ub_c[k]) / |
| 5174 | (double) (vm_group_bytes[k]), 1.0); |
| 5175 | } else { |
| 5176 | num_group_per_lower_vm_stage = dml_ceil( |
| 5177 | (double) (dpde0_bytes_per_frame_ub_l[k]) / |
| 5178 | (double) (vm_group_bytes[k]), 1.0); |
| 5179 | } |
| 5180 | } else { |
| 5181 | if (GPUVMMaxPageTableLevels == 1) { |
| 5182 | if (BytePerPixelC[k] > 0) { |
| 5183 | num_group_per_lower_vm_stage = dml_ceil( |
| 5184 | (double) (meta_pte_bytes_per_frame_ub_l[k]) / |
| 5185 | (double) (vm_group_bytes[k]), 1.0) + |
| 5186 | dml_ceil((double) (meta_pte_bytes_per_frame_ub_c[k]) / |
| 5187 | (double) (vm_group_bytes[k]), 1.0); |
| 5188 | } else { |
| 5189 | num_group_per_lower_vm_stage = dml_ceil( |
| 5190 | (double) (meta_pte_bytes_per_frame_ub_l[k]) / |
| 5191 | (double) (vm_group_bytes[k]), 1.0); |
| 5192 | } |
| 5193 | } else { |
| 5194 | if (BytePerPixelC[k] > 0) { |
| 5195 | num_group_per_lower_vm_stage = 2 + dml_ceil( |
| 5196 | (double) (dpde0_bytes_per_frame_ub_l[k]) / |
| 5197 | (double) (vm_group_bytes[k]), 1) + |
| 5198 | dml_ceil((double) (dpde0_bytes_per_frame_ub_c[k]) / |
| 5199 | (double) (vm_group_bytes[k]), 1) + |
| 5200 | dml_ceil((double) (meta_pte_bytes_per_frame_ub_l[k]) / |
| 5201 | (double) (vm_group_bytes[k]), 1) + |
| 5202 | dml_ceil((double) (meta_pte_bytes_per_frame_ub_c[k]) / |
| 5203 | (double) (vm_group_bytes[k]), 1); |
| 5204 | } else { |
| 5205 | num_group_per_lower_vm_stage = 1 + dml_ceil( |
| 5206 | (double) (dpde0_bytes_per_frame_ub_l[k]) / |
| 5207 | (double) (vm_group_bytes[k]), 1) + dml_ceil( |
| 5208 | (double) (meta_pte_bytes_per_frame_ub_l[k]) / |
| 5209 | (double) (vm_group_bytes[k]), 1); |
| 5210 | } |
| 5211 | } |
| 5212 | } |
| 5213 | |
| 5214 | if (DCCEnable[k] == false0) { |
| 5215 | if (BytePerPixelC[k] > 0) { |
| 5216 | num_req_per_lower_vm_stage = dpde0_bytes_per_frame_ub_l[k] / 64 + |
| 5217 | dpde0_bytes_per_frame_ub_c[k] / 64; |
| 5218 | } else { |
| 5219 | num_req_per_lower_vm_stage = dpde0_bytes_per_frame_ub_l[k] / 64; |
| 5220 | } |
| 5221 | } else { |
| 5222 | if (GPUVMMaxPageTableLevels == 1) { |
| 5223 | if (BytePerPixelC[k] > 0) { |
| 5224 | num_req_per_lower_vm_stage = meta_pte_bytes_per_frame_ub_l[k] / 64 + |
| 5225 | meta_pte_bytes_per_frame_ub_c[k] / 64; |
| 5226 | } else { |
| 5227 | num_req_per_lower_vm_stage = meta_pte_bytes_per_frame_ub_l[k] / 64; |
| 5228 | } |
| 5229 | } else { |
| 5230 | if (BytePerPixelC[k] > 0) { |
| 5231 | num_req_per_lower_vm_stage = dpde0_bytes_per_frame_ub_l[k] / |
| 5232 | 64 + dpde0_bytes_per_frame_ub_c[k] / 64 + |
| 5233 | meta_pte_bytes_per_frame_ub_l[k] / 64 + |
| 5234 | meta_pte_bytes_per_frame_ub_c[k] / 64; |
| 5235 | } else { |
| 5236 | num_req_per_lower_vm_stage = dpde0_bytes_per_frame_ub_l[k] / |
| 5237 | 64 + meta_pte_bytes_per_frame_ub_l[k] / 64; |
| 5238 | } |
| 5239 | } |
| 5240 | } |
| 5241 | |
| 5242 | TimePerVMGroupVBlank[k] = DestinationLinesToRequestVMInVBlank[k] * |
| 5243 | HTotal[k] / PixelClock[k] / num_group_per_lower_vm_stage; |
| 5244 | TimePerVMGroupFlip[k] = DestinationLinesToRequestVMInImmediateFlip[k] * |
| 5245 | HTotal[k] / PixelClock[k] / num_group_per_lower_vm_stage; |
| 5246 | TimePerVMRequestVBlank[k] = DestinationLinesToRequestVMInVBlank[k] * |
| 5247 | HTotal[k] / PixelClock[k] / num_req_per_lower_vm_stage; |
| 5248 | TimePerVMRequestFlip[k] = DestinationLinesToRequestVMInImmediateFlip[k] * |
| 5249 | HTotal[k] / PixelClock[k] / num_req_per_lower_vm_stage; |
| 5250 | |
| 5251 | if (GPUVMMaxPageTableLevels > 2) { |
| 5252 | TimePerVMGroupVBlank[k] = TimePerVMGroupVBlank[k] / 2; |
| 5253 | TimePerVMGroupFlip[k] = TimePerVMGroupFlip[k] / 2; |
| 5254 | TimePerVMRequestVBlank[k] = TimePerVMRequestVBlank[k] / 2; |
| 5255 | TimePerVMRequestFlip[k] = TimePerVMRequestFlip[k] / 2; |
| 5256 | } |
| 5257 | |
| 5258 | } else { |
| 5259 | TimePerVMGroupVBlank[k] = 0; |
| 5260 | TimePerVMGroupFlip[k] = 0; |
| 5261 | TimePerVMRequestVBlank[k] = 0; |
| 5262 | TimePerVMRequestFlip[k] = 0; |
| 5263 | } |
| 5264 | |
| 5265 | #ifdef __DML_VBA_DEBUG__ |
| 5266 | dml_print("DML::%s: k=%0d, TimePerVMGroupVBlank = %f\n", __func__, k, TimePerVMGroupVBlank[k]){do { } while(0); }; |
| 5267 | dml_print("DML::%s: k=%0d, TimePerVMGroupFlip = %f\n", __func__, k, TimePerVMGroupFlip[k]){do { } while(0); }; |
| 5268 | dml_print("DML::%s: k=%0d, TimePerVMRequestVBlank = %f\n", __func__, k, TimePerVMRequestVBlank[k]){do { } while(0); }; |
| 5269 | dml_print("DML::%s: k=%0d, TimePerVMRequestFlip = %f\n", __func__, k, TimePerVMRequestFlip[k]){do { } while(0); }; |
| 5270 | #endif |
| 5271 | } |
| 5272 | } // CalculateVMGroupAndRequestTimes |
| 5273 | |
| 5274 | void dml32_CalculateDCCConfiguration( |
| 5275 | bool_Bool DCCEnabled, |
| 5276 | bool_Bool DCCProgrammingAssumesScanDirectionUnknown, |
| 5277 | enum source_format_class SourcePixelFormat, |
| 5278 | unsigned int SurfaceWidthLuma, |
| 5279 | unsigned int SurfaceWidthChroma, |
| 5280 | unsigned int SurfaceHeightLuma, |
| 5281 | unsigned int SurfaceHeightChroma, |
| 5282 | unsigned int nomDETInKByte, |
| 5283 | unsigned int RequestHeight256ByteLuma, |
| 5284 | unsigned int RequestHeight256ByteChroma, |
| 5285 | enum dm_swizzle_mode TilingFormat, |
| 5286 | unsigned int BytePerPixelY, |
| 5287 | unsigned int BytePerPixelC, |
| 5288 | double BytePerPixelDETY, |
| 5289 | double BytePerPixelDETC, |
| 5290 | enum dm_rotation_angle SourceRotation, |
| 5291 | /* Output */ |
| 5292 | unsigned int *MaxUncompressedBlockLuma, |
| 5293 | unsigned int *MaxUncompressedBlockChroma, |
| 5294 | unsigned int *MaxCompressedBlockLuma, |
| 5295 | unsigned int *MaxCompressedBlockChroma, |
| 5296 | unsigned int *IndependentBlockLuma, |
| 5297 | unsigned int *IndependentBlockChroma) |
| 5298 | { |
| 5299 | typedef enum { |
| 5300 | REQ_256Bytes, |
| 5301 | REQ_128BytesNonContiguous, |
| 5302 | REQ_128BytesContiguous, |
| 5303 | REQ_NA |
| 5304 | } RequestType; |
| 5305 | |
| 5306 | RequestType RequestLuma; |
| 5307 | RequestType RequestChroma; |
| 5308 | |
| 5309 | unsigned int segment_order_horz_contiguous_luma; |
| 5310 | unsigned int segment_order_horz_contiguous_chroma; |
| 5311 | unsigned int segment_order_vert_contiguous_luma; |
| 5312 | unsigned int segment_order_vert_contiguous_chroma; |
| 5313 | unsigned int req128_horz_wc_l; |
| 5314 | unsigned int req128_horz_wc_c; |
| 5315 | unsigned int req128_vert_wc_l; |
| 5316 | unsigned int req128_vert_wc_c; |
| 5317 | unsigned int MAS_vp_horz_limit; |
| 5318 | unsigned int MAS_vp_vert_limit; |
| 5319 | unsigned int max_vp_horz_width; |
| 5320 | unsigned int max_vp_vert_height; |
| 5321 | unsigned int eff_surf_width_l; |
| 5322 | unsigned int eff_surf_width_c; |
| 5323 | unsigned int eff_surf_height_l; |
| 5324 | unsigned int eff_surf_height_c; |
| 5325 | unsigned int full_swath_bytes_horz_wc_l; |
| 5326 | unsigned int full_swath_bytes_horz_wc_c; |
| 5327 | unsigned int full_swath_bytes_vert_wc_l; |
| 5328 | unsigned int full_swath_bytes_vert_wc_c; |
| 5329 | unsigned int DETBufferSizeForDCC = nomDETInKByte * 1024; |
| 5330 | |
| 5331 | unsigned int yuv420; |
| 5332 | unsigned int horz_div_l; |
| 5333 | unsigned int horz_div_c; |
| 5334 | unsigned int vert_div_l; |
| 5335 | unsigned int vert_div_c; |
| 5336 | |
| 5337 | unsigned int swath_buf_size; |
| 5338 | double detile_buf_vp_horz_limit; |
| 5339 | double detile_buf_vp_vert_limit; |
| 5340 | |
| 5341 | yuv420 = ((SourcePixelFormat == dm_420_8 || SourcePixelFormat == dm_420_10 || |
| 5342 | SourcePixelFormat == dm_420_12) ? 1 : 0); |
| 5343 | horz_div_l = 1; |
| 5344 | horz_div_c = 1; |
| 5345 | vert_div_l = 1; |
| 5346 | vert_div_c = 1; |
| 5347 | |
| 5348 | if (BytePerPixelY == 1) |
| 5349 | vert_div_l = 0; |
| 5350 | if (BytePerPixelC == 1) |
| 5351 | vert_div_c = 0; |
| 5352 | |
| 5353 | if (BytePerPixelC == 0) { |
| 5354 | swath_buf_size = DETBufferSizeForDCC / 2 - 2 * 256; |
| 5355 | detile_buf_vp_horz_limit = (double) swath_buf_size / ((double) RequestHeight256ByteLuma * |
| 5356 | BytePerPixelY / (1 + horz_div_l)); |
| 5357 | detile_buf_vp_vert_limit = (double) swath_buf_size / (256.0 / RequestHeight256ByteLuma / |
| 5358 | (1 + vert_div_l)); |
| 5359 | } else { |
| 5360 | swath_buf_size = DETBufferSizeForDCC / 2 - 2 * 2 * 256; |
| 5361 | detile_buf_vp_horz_limit = (double) swath_buf_size / ((double) RequestHeight256ByteLuma * |
| 5362 | BytePerPixelY / (1 + horz_div_l) + (double) RequestHeight256ByteChroma * |
| 5363 | BytePerPixelC / (1 + horz_div_c) / (1 + yuv420)); |
| 5364 | detile_buf_vp_vert_limit = (double) swath_buf_size / (256.0 / RequestHeight256ByteLuma / |
| 5365 | (1 + vert_div_l) + 256.0 / RequestHeight256ByteChroma / |
| 5366 | (1 + vert_div_c) / (1 + yuv420)); |
| 5367 | } |
| 5368 | |
| 5369 | if (SourcePixelFormat == dm_420_10) { |
| 5370 | detile_buf_vp_horz_limit = 1.5 * detile_buf_vp_horz_limit; |
| 5371 | detile_buf_vp_vert_limit = 1.5 * detile_buf_vp_vert_limit; |
| 5372 | } |
| 5373 | |
| 5374 | detile_buf_vp_horz_limit = dml_floor(detile_buf_vp_horz_limit - 1, 16); |
| 5375 | detile_buf_vp_vert_limit = dml_floor(detile_buf_vp_vert_limit - 1, 16); |
| 5376 | |
| 5377 | MAS_vp_horz_limit = SourcePixelFormat == dm_rgbe_alpha ? 3840 : 6144; |
| 5378 | MAS_vp_vert_limit = SourcePixelFormat == dm_rgbe_alpha ? 3840 : (BytePerPixelY == 8 ? 3072 : 6144); |
| 5379 | max_vp_horz_width = dml_min((double) MAS_vp_horz_limit, detile_buf_vp_horz_limit); |
| 5380 | max_vp_vert_height = dml_min((double) MAS_vp_vert_limit, detile_buf_vp_vert_limit); |
| 5381 | eff_surf_width_l = (SurfaceWidthLuma > max_vp_horz_width ? max_vp_horz_width : SurfaceWidthLuma); |
| 5382 | eff_surf_width_c = eff_surf_width_l / (1 + yuv420); |
| 5383 | eff_surf_height_l = (SurfaceHeightLuma > max_vp_vert_height ? max_vp_vert_height : SurfaceHeightLuma); |
| 5384 | eff_surf_height_c = eff_surf_height_l / (1 + yuv420); |
| 5385 | |
| 5386 | full_swath_bytes_horz_wc_l = eff_surf_width_l * RequestHeight256ByteLuma * BytePerPixelY; |
| 5387 | full_swath_bytes_vert_wc_l = eff_surf_height_l * 256 / RequestHeight256ByteLuma; |
| 5388 | if (BytePerPixelC > 0) { |
| 5389 | full_swath_bytes_horz_wc_c = eff_surf_width_c * RequestHeight256ByteChroma * BytePerPixelC; |
| 5390 | full_swath_bytes_vert_wc_c = eff_surf_height_c * 256 / RequestHeight256ByteChroma; |
| 5391 | } else { |
| 5392 | full_swath_bytes_horz_wc_c = 0; |
| 5393 | full_swath_bytes_vert_wc_c = 0; |
| 5394 | } |
| 5395 | |
| 5396 | if (SourcePixelFormat == dm_420_10) { |
| 5397 | full_swath_bytes_horz_wc_l = dml_ceil((double) full_swath_bytes_horz_wc_l * 2.0 / 3.0, 256.0); |
| 5398 | full_swath_bytes_horz_wc_c = dml_ceil((double) full_swath_bytes_horz_wc_c * 2.0 / 3.0, 256.0); |
| 5399 | full_swath_bytes_vert_wc_l = dml_ceil((double) full_swath_bytes_vert_wc_l * 2.0 / 3.0, 256.0); |
| 5400 | full_swath_bytes_vert_wc_c = dml_ceil((double) full_swath_bytes_vert_wc_c * 2.0 / 3.0, 256.0); |
| 5401 | } |
| 5402 | |
| 5403 | if (2 * full_swath_bytes_horz_wc_l + 2 * full_swath_bytes_horz_wc_c <= DETBufferSizeForDCC) { |
| 5404 | req128_horz_wc_l = 0; |
| 5405 | req128_horz_wc_c = 0; |
| 5406 | } else if (full_swath_bytes_horz_wc_l < 1.5 * full_swath_bytes_horz_wc_c && 2 * full_swath_bytes_horz_wc_l + |
| 5407 | full_swath_bytes_horz_wc_c <= DETBufferSizeForDCC) { |
| 5408 | req128_horz_wc_l = 0; |
| 5409 | req128_horz_wc_c = 1; |
| 5410 | } else if (full_swath_bytes_horz_wc_l >= 1.5 * full_swath_bytes_horz_wc_c && full_swath_bytes_horz_wc_l + 2 * |
| 5411 | full_swath_bytes_horz_wc_c <= DETBufferSizeForDCC) { |
| 5412 | req128_horz_wc_l = 1; |
| 5413 | req128_horz_wc_c = 0; |
| 5414 | } else { |
| 5415 | req128_horz_wc_l = 1; |
| 5416 | req128_horz_wc_c = 1; |
| 5417 | } |
| 5418 | |
| 5419 | if (2 * full_swath_bytes_vert_wc_l + 2 * full_swath_bytes_vert_wc_c <= DETBufferSizeForDCC) { |
| 5420 | req128_vert_wc_l = 0; |
| 5421 | req128_vert_wc_c = 0; |
| 5422 | } else if (full_swath_bytes_vert_wc_l < 1.5 * full_swath_bytes_vert_wc_c && 2 * |
| 5423 | full_swath_bytes_vert_wc_l + full_swath_bytes_vert_wc_c <= DETBufferSizeForDCC) { |
| 5424 | req128_vert_wc_l = 0; |
| 5425 | req128_vert_wc_c = 1; |
| 5426 | } else if (full_swath_bytes_vert_wc_l >= 1.5 * full_swath_bytes_vert_wc_c && |
| 5427 | full_swath_bytes_vert_wc_l + 2 * full_swath_bytes_vert_wc_c <= DETBufferSizeForDCC) { |
| 5428 | req128_vert_wc_l = 1; |
| 5429 | req128_vert_wc_c = 0; |
| 5430 | } else { |
| 5431 | req128_vert_wc_l = 1; |
| 5432 | req128_vert_wc_c = 1; |
| 5433 | } |
| 5434 | |
| 5435 | if (BytePerPixelY == 2) { |
| 5436 | segment_order_horz_contiguous_luma = 0; |
| 5437 | segment_order_vert_contiguous_luma = 1; |
| 5438 | } else { |
| 5439 | segment_order_horz_contiguous_luma = 1; |
| 5440 | segment_order_vert_contiguous_luma = 0; |
| 5441 | } |
| 5442 | |
| 5443 | if (BytePerPixelC == 2) { |
| 5444 | segment_order_horz_contiguous_chroma = 0; |
| 5445 | segment_order_vert_contiguous_chroma = 1; |
| 5446 | } else { |
| 5447 | segment_order_horz_contiguous_chroma = 1; |
| 5448 | segment_order_vert_contiguous_chroma = 0; |
| 5449 | } |
| 5450 | #ifdef __DML_VBA_DEBUG__ |
| 5451 | dml_print("DML::%s: DCCEnabled = %d\n", __func__, DCCEnabled){do { } while(0); }; |
| 5452 | dml_print("DML::%s: nomDETInKByte = %d\n", __func__, nomDETInKByte){do { } while(0); }; |
| 5453 | dml_print("DML::%s: DETBufferSizeForDCC = %d\n", __func__, DETBufferSizeForDCC){do { } while(0); }; |
| 5454 | dml_print("DML::%s: req128_horz_wc_l = %d\n", __func__, req128_horz_wc_l){do { } while(0); }; |
| 5455 | dml_print("DML::%s: req128_horz_wc_c = %d\n", __func__, req128_horz_wc_c){do { } while(0); }; |
| 5456 | dml_print("DML::%s: full_swath_bytes_horz_wc_l = %d\n", __func__, full_swath_bytes_horz_wc_l){do { } while(0); }; |
| 5457 | dml_print("DML::%s: full_swath_bytes_vert_wc_c = %d\n", __func__, full_swath_bytes_vert_wc_c){do { } while(0); }; |
| 5458 | dml_print("DML::%s: segment_order_horz_contiguous_luma = %d\n", __func__, segment_order_horz_contiguous_luma){do { } while(0); }; |
| 5459 | dml_print("DML::%s: segment_order_horz_contiguous_chroma = %d\n",{do { } while(0); } |
| 5460 | __func__, segment_order_horz_contiguous_chroma){do { } while(0); }; |
| 5461 | #endif |
| 5462 | |
| 5463 | if (DCCProgrammingAssumesScanDirectionUnknown == true1) { |
| 5464 | if (req128_horz_wc_l == 0 && req128_vert_wc_l == 0) |
| 5465 | RequestLuma = REQ_256Bytes; |
| 5466 | else if ((req128_horz_wc_l == 1 && segment_order_horz_contiguous_luma == 0) || |
| 5467 | (req128_vert_wc_l == 1 && segment_order_vert_contiguous_luma == 0)) |
| 5468 | RequestLuma = REQ_128BytesNonContiguous; |
| 5469 | else |
| 5470 | RequestLuma = REQ_128BytesContiguous; |
| 5471 | |
| 5472 | if (req128_horz_wc_c == 0 && req128_vert_wc_c == 0) |
| 5473 | RequestChroma = REQ_256Bytes; |
| 5474 | else if ((req128_horz_wc_c == 1 && segment_order_horz_contiguous_chroma == 0) || |
| 5475 | (req128_vert_wc_c == 1 && segment_order_vert_contiguous_chroma == 0)) |
| 5476 | RequestChroma = REQ_128BytesNonContiguous; |
| 5477 | else |
| 5478 | RequestChroma = REQ_128BytesContiguous; |
| 5479 | |
| 5480 | } else if (!IsVertical(SourceRotation)) { |
| 5481 | if (req128_horz_wc_l == 0) |
| 5482 | RequestLuma = REQ_256Bytes; |
| 5483 | else if (segment_order_horz_contiguous_luma == 0) |
| 5484 | RequestLuma = REQ_128BytesNonContiguous; |
| 5485 | else |
| 5486 | RequestLuma = REQ_128BytesContiguous; |
| 5487 | |
| 5488 | if (req128_horz_wc_c == 0) |
| 5489 | RequestChroma = REQ_256Bytes; |
| 5490 | else if (segment_order_horz_contiguous_chroma == 0) |
| 5491 | RequestChroma = REQ_128BytesNonContiguous; |
| 5492 | else |
| 5493 | RequestChroma = REQ_128BytesContiguous; |
| 5494 | |
| 5495 | } else { |
| 5496 | if (req128_vert_wc_l == 0) |
| 5497 | RequestLuma = REQ_256Bytes; |
| 5498 | else if (segment_order_vert_contiguous_luma == 0) |
| 5499 | RequestLuma = REQ_128BytesNonContiguous; |
| 5500 | else |
| 5501 | RequestLuma = REQ_128BytesContiguous; |
| 5502 | |
| 5503 | if (req128_vert_wc_c == 0) |
| 5504 | RequestChroma = REQ_256Bytes; |
| 5505 | else if (segment_order_vert_contiguous_chroma == 0) |
| 5506 | RequestChroma = REQ_128BytesNonContiguous; |
| 5507 | else |
| 5508 | RequestChroma = REQ_128BytesContiguous; |
| 5509 | } |
| 5510 | |
| 5511 | if (RequestLuma == REQ_256Bytes) { |
| 5512 | *MaxUncompressedBlockLuma = 256; |
| 5513 | *MaxCompressedBlockLuma = 256; |
| 5514 | *IndependentBlockLuma = 0; |
| 5515 | } else if (RequestLuma == REQ_128BytesContiguous) { |
| 5516 | *MaxUncompressedBlockLuma = 256; |
| 5517 | *MaxCompressedBlockLuma = 128; |
| 5518 | *IndependentBlockLuma = 128; |
| 5519 | } else { |
| 5520 | *MaxUncompressedBlockLuma = 256; |
| 5521 | *MaxCompressedBlockLuma = 64; |
| 5522 | *IndependentBlockLuma = 64; |
| 5523 | } |
| 5524 | |
| 5525 | if (RequestChroma == REQ_256Bytes) { |
| 5526 | *MaxUncompressedBlockChroma = 256; |
| 5527 | *MaxCompressedBlockChroma = 256; |
| 5528 | *IndependentBlockChroma = 0; |
| 5529 | } else if (RequestChroma == REQ_128BytesContiguous) { |
| 5530 | *MaxUncompressedBlockChroma = 256; |
| 5531 | *MaxCompressedBlockChroma = 128; |
| 5532 | *IndependentBlockChroma = 128; |
| 5533 | } else { |
| 5534 | *MaxUncompressedBlockChroma = 256; |
| 5535 | *MaxCompressedBlockChroma = 64; |
| 5536 | *IndependentBlockChroma = 64; |
| 5537 | } |
| 5538 | |
| 5539 | if (DCCEnabled != true1 || BytePerPixelC == 0) { |
| 5540 | *MaxUncompressedBlockChroma = 0; |
| 5541 | *MaxCompressedBlockChroma = 0; |
| 5542 | *IndependentBlockChroma = 0; |
| 5543 | } |
| 5544 | |
| 5545 | if (DCCEnabled != true1) { |
| 5546 | *MaxUncompressedBlockLuma = 0; |
| 5547 | *MaxCompressedBlockLuma = 0; |
| 5548 | *IndependentBlockLuma = 0; |
| 5549 | } |
| 5550 | |
| 5551 | #ifdef __DML_VBA_DEBUG__ |
| 5552 | dml_print("DML::%s: MaxUncompressedBlockLuma = %d\n", __func__, *MaxUncompressedBlockLuma){do { } while(0); }; |
| 5553 | dml_print("DML::%s: MaxCompressedBlockLuma = %d\n", __func__, *MaxCompressedBlockLuma){do { } while(0); }; |
| 5554 | dml_print("DML::%s: IndependentBlockLuma = %d\n", __func__, *IndependentBlockLuma){do { } while(0); }; |
| 5555 | dml_print("DML::%s: MaxUncompressedBlockChroma = %d\n", __func__, *MaxUncompressedBlockChroma){do { } while(0); }; |
| 5556 | dml_print("DML::%s: MaxCompressedBlockChroma = %d\n", __func__, *MaxCompressedBlockChroma){do { } while(0); }; |
| 5557 | dml_print("DML::%s: IndependentBlockChroma = %d\n", __func__, *IndependentBlockChroma){do { } while(0); }; |
| 5558 | #endif |
| 5559 | |
| 5560 | } // CalculateDCCConfiguration |
| 5561 | |
| 5562 | void dml32_CalculateStutterEfficiency( |
| 5563 | unsigned int CompressedBufferSizeInkByte, |
| 5564 | enum dm_use_mall_for_pstate_change_mode UseMALLForPStateChange[], |
| 5565 | bool_Bool UnboundedRequestEnabled, |
| 5566 | unsigned int MetaFIFOSizeInKEntries, |
| 5567 | unsigned int ZeroSizeBufferEntries, |
| 5568 | unsigned int PixelChunkSizeInKByte, |
| 5569 | unsigned int NumberOfActiveSurfaces, |
| 5570 | unsigned int ROBBufferSizeInKByte, |
| 5571 | double TotalDataReadBandwidth, |
| 5572 | double DCFCLK, |
| 5573 | double ReturnBW, |
| 5574 | unsigned int CompbufReservedSpace64B, |
| 5575 | unsigned int CompbufReservedSpaceZs, |
| 5576 | double SRExitTime, |
| 5577 | double SRExitZ8Time, |
| 5578 | bool_Bool SynchronizeTimingsFinal, |
| 5579 | unsigned int BlendingAndTiming[], |
| 5580 | double StutterEnterPlusExitWatermark, |
| 5581 | double Z8StutterEnterPlusExitWatermark, |
| 5582 | bool_Bool ProgressiveToInterlaceUnitInOPP, |
| 5583 | bool_Bool Interlace[], |
| 5584 | double MinTTUVBlank[], |
| 5585 | unsigned int DPPPerSurface[], |
| 5586 | unsigned int DETBufferSizeY[], |
| 5587 | unsigned int BytePerPixelY[], |
| 5588 | double BytePerPixelDETY[], |
| 5589 | double SwathWidthY[], |
| 5590 | unsigned int SwathHeightY[], |
| 5591 | unsigned int SwathHeightC[], |
| 5592 | double NetDCCRateLuma[], |
| 5593 | double NetDCCRateChroma[], |
| 5594 | double DCCFractionOfZeroSizeRequestsLuma[], |
| 5595 | double DCCFractionOfZeroSizeRequestsChroma[], |
| 5596 | unsigned int HTotal[], |
| 5597 | unsigned int VTotal[], |
| 5598 | double PixelClock[], |
| 5599 | double VRatio[], |
| 5600 | enum dm_rotation_angle SourceRotation[], |
| 5601 | unsigned int BlockHeight256BytesY[], |
| 5602 | unsigned int BlockWidth256BytesY[], |
| 5603 | unsigned int BlockHeight256BytesC[], |
| 5604 | unsigned int BlockWidth256BytesC[], |
| 5605 | unsigned int DCCYMaxUncompressedBlock[], |
| 5606 | unsigned int DCCCMaxUncompressedBlock[], |
| 5607 | unsigned int VActive[], |
| 5608 | bool_Bool DCCEnable[], |
| 5609 | bool_Bool WritebackEnable[], |
| 5610 | double ReadBandwidthSurfaceLuma[], |
| 5611 | double ReadBandwidthSurfaceChroma[], |
| 5612 | double meta_row_bw[], |
| 5613 | double dpte_row_bw[], |
| 5614 | |
| 5615 | /* Output */ |
| 5616 | double *StutterEfficiencyNotIncludingVBlank, |
| 5617 | double *StutterEfficiency, |
| 5618 | unsigned int *NumberOfStutterBurstsPerFrame, |
| 5619 | double *Z8StutterEfficiencyNotIncludingVBlank, |
| 5620 | double *Z8StutterEfficiency, |
| 5621 | unsigned int *Z8NumberOfStutterBurstsPerFrame, |
| 5622 | double *StutterPeriod, |
| 5623 | bool_Bool *DCHUBBUB_ARB_CSTATE_MAX_CAP_MODE) |
| 5624 | { |
| 5625 | |
| 5626 | bool_Bool FoundCriticalSurface = false0; |
| 5627 | unsigned int SwathSizeCriticalSurface = 0; |
| 5628 | unsigned int LastChunkOfSwathSize; |
| 5629 | unsigned int MissingPartOfLastSwathOfDETSize; |
| 5630 | double LastZ8StutterPeriod = 0.0; |
| 5631 | double LastStutterPeriod = 0.0; |
| 5632 | unsigned int TotalNumberOfActiveOTG = 0; |
| 5633 | double doublePixelClock; |
| 5634 | unsigned int doubleHTotal; |
| 5635 | unsigned int doubleVTotal; |
| 5636 | bool_Bool SameTiming = true1; |
| 5637 | double DETBufferingTimeY; |
| 5638 | double SwathWidthYCriticalSurface = 0.0; |
| 5639 | double SwathHeightYCriticalSurface = 0.0; |
| 5640 | double VActiveTimeCriticalSurface = 0.0; |
| 5641 | double FrameTimeCriticalSurface = 0.0; |
| 5642 | unsigned int BytePerPixelYCriticalSurface = 0; |
| 5643 | double LinesToFinishSwathTransferStutterCriticalSurface = 0.0; |
| 5644 | unsigned int DETBufferSizeYCriticalSurface = 0; |
| 5645 | double MinTTUVBlankCriticalSurface = 0.0; |
| 5646 | unsigned int BlockWidth256BytesYCriticalSurface = 0; |
| 5647 | bool_Bool doublePlaneCriticalSurface = 0; |
| 5648 | bool_Bool doublePipeCriticalSurface = 0; |
| 5649 | double TotalCompressedReadBandwidth; |
| 5650 | double TotalRowReadBandwidth; |
| 5651 | double AverageDCCCompressionRate; |
| 5652 | double EffectiveCompressedBufferSize; |
| 5653 | double PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer; |
| 5654 | double StutterBurstTime; |
| 5655 | unsigned int TotalActiveWriteback; |
| 5656 | double LinesInDETY; |
| 5657 | double LinesInDETYRoundedDownToSwath; |
| 5658 | double MaximumEffectiveCompressionLuma; |
| 5659 | double MaximumEffectiveCompressionChroma; |
| 5660 | double TotalZeroSizeRequestReadBandwidth; |
| 5661 | double TotalZeroSizeCompressedReadBandwidth; |
| 5662 | double AverageDCCZeroSizeFraction; |
| 5663 | double AverageZeroSizeCompressionRate; |
| 5664 | unsigned int k; |
| 5665 | |
| 5666 | TotalZeroSizeRequestReadBandwidth = 0; |
| 5667 | TotalZeroSizeCompressedReadBandwidth = 0; |
| 5668 | TotalRowReadBandwidth = 0; |
| 5669 | TotalCompressedReadBandwidth = 0; |
| 5670 | |
| 5671 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 5672 | if (UseMALLForPStateChange[k] != dm_use_mall_pstate_change_phantom_pipe) { |
| 5673 | if (DCCEnable[k] == true1) { |
| 5674 | if ((IsVertical(SourceRotation[k]) && BlockWidth256BytesY[k] > SwathHeightY[k]) |
| 5675 | || (!IsVertical(SourceRotation[k]) |
| 5676 | && BlockHeight256BytesY[k] > SwathHeightY[k]) |
| 5677 | || DCCYMaxUncompressedBlock[k] < 256) { |
| 5678 | MaximumEffectiveCompressionLuma = 2; |
| 5679 | } else { |
| 5680 | MaximumEffectiveCompressionLuma = 4; |
| 5681 | } |
| 5682 | TotalCompressedReadBandwidth = TotalCompressedReadBandwidth |
| 5683 | + ReadBandwidthSurfaceLuma[k] |
| 5684 | / dml_min(NetDCCRateLuma[k], |
| 5685 | MaximumEffectiveCompressionLuma); |
| 5686 | #ifdef __DML_VBA_DEBUG__ |
| 5687 | dml_print("DML::%s: k=%0d, ReadBandwidthSurfaceLuma = %f\n",{do { } while(0); } |
| 5688 | __func__, k, ReadBandwidthSurfaceLuma[k]){do { } while(0); }; |
| 5689 | dml_print("DML::%s: k=%0d, NetDCCRateLuma = %f\n",{do { } while(0); } |
| 5690 | __func__, k, NetDCCRateLuma[k]){do { } while(0); }; |
| 5691 | dml_print("DML::%s: k=%0d, MaximumEffectiveCompressionLuma = %f\n",{do { } while(0); } |
| 5692 | __func__, k, MaximumEffectiveCompressionLuma){do { } while(0); }; |
| 5693 | #endif |
| 5694 | TotalZeroSizeRequestReadBandwidth = TotalZeroSizeRequestReadBandwidth |
| 5695 | + ReadBandwidthSurfaceLuma[k] * DCCFractionOfZeroSizeRequestsLuma[k]; |
| 5696 | TotalZeroSizeCompressedReadBandwidth = TotalZeroSizeCompressedReadBandwidth |
| 5697 | + ReadBandwidthSurfaceLuma[k] * DCCFractionOfZeroSizeRequestsLuma[k] |
| 5698 | / MaximumEffectiveCompressionLuma; |
| 5699 | |
| 5700 | if (ReadBandwidthSurfaceChroma[k] > 0) { |
| 5701 | if ((IsVertical(SourceRotation[k]) && BlockWidth256BytesC[k] > SwathHeightC[k]) |
| 5702 | || (!IsVertical(SourceRotation[k]) |
| 5703 | && BlockHeight256BytesC[k] > SwathHeightC[k]) |
| 5704 | || DCCCMaxUncompressedBlock[k] < 256) { |
| 5705 | MaximumEffectiveCompressionChroma = 2; |
| 5706 | } else { |
| 5707 | MaximumEffectiveCompressionChroma = 4; |
| 5708 | } |
| 5709 | TotalCompressedReadBandwidth = |
| 5710 | TotalCompressedReadBandwidth |
| 5711 | + ReadBandwidthSurfaceChroma[k] |
| 5712 | / dml_min(NetDCCRateChroma[k], |
| 5713 | MaximumEffectiveCompressionChroma); |
| 5714 | #ifdef __DML_VBA_DEBUG__ |
| 5715 | dml_print("DML::%s: k=%0d, ReadBandwidthSurfaceChroma = %f\n",{do { } while(0); } |
| 5716 | __func__, k, ReadBandwidthSurfaceChroma[k]){do { } while(0); }; |
| 5717 | dml_print("DML::%s: k=%0d, NetDCCRateChroma = %f\n",{do { } while(0); } |
| 5718 | __func__, k, NetDCCRateChroma[k]){do { } while(0); }; |
| 5719 | dml_print("DML::%s: k=%0d, MaximumEffectiveCompressionChroma = %f\n",{do { } while(0); } |
| 5720 | __func__, k, MaximumEffectiveCompressionChroma){do { } while(0); }; |
| 5721 | #endif |
| 5722 | TotalZeroSizeRequestReadBandwidth = TotalZeroSizeRequestReadBandwidth |
| 5723 | + ReadBandwidthSurfaceChroma[k] |
| 5724 | * DCCFractionOfZeroSizeRequestsChroma[k]; |
| 5725 | TotalZeroSizeCompressedReadBandwidth = TotalZeroSizeCompressedReadBandwidth |
| 5726 | + ReadBandwidthSurfaceChroma[k] |
| 5727 | * DCCFractionOfZeroSizeRequestsChroma[k] |
| 5728 | / MaximumEffectiveCompressionChroma; |
| 5729 | } |
| 5730 | } else { |
| 5731 | TotalCompressedReadBandwidth = TotalCompressedReadBandwidth |
| 5732 | + ReadBandwidthSurfaceLuma[k] + ReadBandwidthSurfaceChroma[k]; |
| 5733 | } |
| 5734 | TotalRowReadBandwidth = TotalRowReadBandwidth |
| 5735 | + DPPPerSurface[k] * (meta_row_bw[k] + dpte_row_bw[k]); |
| 5736 | } |
| 5737 | } |
| 5738 | |
| 5739 | AverageDCCCompressionRate = TotalDataReadBandwidth / TotalCompressedReadBandwidth; |
| 5740 | AverageDCCZeroSizeFraction = TotalZeroSizeRequestReadBandwidth / TotalDataReadBandwidth; |
| 5741 | |
| 5742 | #ifdef __DML_VBA_DEBUG__ |
| 5743 | dml_print("DML::%s: UnboundedRequestEnabled = %d\n", __func__, UnboundedRequestEnabled){do { } while(0); }; |
| 5744 | dml_print("DML::%s: TotalCompressedReadBandwidth = %f\n", __func__, TotalCompressedReadBandwidth){do { } while(0); }; |
| 5745 | dml_print("DML::%s: TotalZeroSizeRequestReadBandwidth = %f\n", __func__, TotalZeroSizeRequestReadBandwidth){do { } while(0); }; |
| 5746 | dml_print("DML::%s: TotalZeroSizeCompressedReadBandwidth = %f\n",{do { } while(0); } |
| 5747 | __func__, TotalZeroSizeCompressedReadBandwidth){do { } while(0); }; |
| 5748 | dml_print("DML::%s: MaximumEffectiveCompressionLuma = %f\n", __func__, MaximumEffectiveCompressionLuma){do { } while(0); }; |
| 5749 | dml_print("DML::%s: MaximumEffectiveCompressionChroma = %f\n", __func__, MaximumEffectiveCompressionChroma){do { } while(0); }; |
| 5750 | dml_print("DML::%s: AverageDCCCompressionRate = %f\n", __func__, AverageDCCCompressionRate){do { } while(0); }; |
| 5751 | dml_print("DML::%s: AverageDCCZeroSizeFraction = %f\n", __func__, AverageDCCZeroSizeFraction){do { } while(0); }; |
| 5752 | dml_print("DML::%s: CompbufReservedSpace64B = %d\n", __func__, CompbufReservedSpace64B){do { } while(0); }; |
| 5753 | dml_print("DML::%s: CompbufReservedSpaceZs = %d\n", __func__, CompbufReservedSpaceZs){do { } while(0); }; |
| 5754 | dml_print("DML::%s: CompressedBufferSizeInkByte = %d\n", __func__, CompressedBufferSizeInkByte){do { } while(0); }; |
| 5755 | #endif |
| 5756 | if (AverageDCCZeroSizeFraction == 1) { |
| 5757 | AverageZeroSizeCompressionRate = TotalZeroSizeRequestReadBandwidth |
| 5758 | / TotalZeroSizeCompressedReadBandwidth; |
| 5759 | EffectiveCompressedBufferSize = (double) MetaFIFOSizeInKEntries * 1024 * 64 |
| 5760 | * AverageZeroSizeCompressionRate |
| 5761 | + ((double) ZeroSizeBufferEntries - CompbufReservedSpaceZs) * 64 |
| 5762 | * AverageZeroSizeCompressionRate; |
| 5763 | } else if (AverageDCCZeroSizeFraction > 0) { |
| 5764 | AverageZeroSizeCompressionRate = TotalZeroSizeRequestReadBandwidth |
| 5765 | / TotalZeroSizeCompressedReadBandwidth; |
| 5766 | EffectiveCompressedBufferSize = dml_min( |
| 5767 | (double) CompressedBufferSizeInkByte * 1024 * AverageDCCCompressionRate, |
| 5768 | (double) MetaFIFOSizeInKEntries * 1024 * 64 |
| 5769 | / (AverageDCCZeroSizeFraction / AverageZeroSizeCompressionRate |
| 5770 | + 1 / AverageDCCCompressionRate)) |
| 5771 | + dml_min(((double) ROBBufferSizeInKByte * 1024 - CompbufReservedSpace64B * 64) |
| 5772 | * AverageDCCCompressionRate, |
| 5773 | ((double) ZeroSizeBufferEntries - CompbufReservedSpaceZs) * 64 |
| 5774 | / (AverageDCCZeroSizeFraction / AverageZeroSizeCompressionRate)); |
| 5775 | |
| 5776 | #ifdef __DML_VBA_DEBUG__ |
| 5777 | dml_print("DML::%s: min 1 = %f\n", __func__,{do { } while(0); } |
| 5778 | CompressedBufferSizeInkByte * 1024 * AverageDCCCompressionRate){do { } while(0); }; |
| 5779 | dml_print("DML::%s: min 2 = %f\n", __func__, MetaFIFOSizeInKEntries * 1024 * 64 /{do { } while(0); } |
| 5780 | (AverageDCCZeroSizeFraction / AverageZeroSizeCompressionRate + 1 /{do { } while(0); } |
| 5781 | AverageDCCCompressionRate)){do { } while(0); }; |
| 5782 | dml_print("DML::%s: min 3 = %f\n", __func__, (ROBBufferSizeInKByte * 1024 -{do { } while(0); } |
| 5783 | CompbufReservedSpace64B * 64) * AverageDCCCompressionRate){do { } while(0); }; |
| 5784 | dml_print("DML::%s: min 4 = %f\n", __func__, (ZeroSizeBufferEntries - CompbufReservedSpaceZs) * 64 /{do { } while(0); } |
| 5785 | (AverageDCCZeroSizeFraction / AverageZeroSizeCompressionRate)){do { } while(0); }; |
| 5786 | #endif |
| 5787 | } else { |
| 5788 | EffectiveCompressedBufferSize = dml_min( |
| 5789 | (double) CompressedBufferSizeInkByte * 1024 * AverageDCCCompressionRate, |
| 5790 | (double) MetaFIFOSizeInKEntries * 1024 * 64 * AverageDCCCompressionRate) |
| 5791 | + ((double) ROBBufferSizeInKByte * 1024 - CompbufReservedSpace64B * 64) |
| 5792 | * AverageDCCCompressionRate; |
| 5793 | |
| 5794 | #ifdef __DML_VBA_DEBUG__ |
| 5795 | dml_print("DML::%s: min 1 = %f\n", __func__,{do { } while(0); } |
| 5796 | CompressedBufferSizeInkByte * 1024 * AverageDCCCompressionRate){do { } while(0); }; |
| 5797 | dml_print("DML::%s: min 2 = %f\n", __func__,{do { } while(0); } |
| 5798 | MetaFIFOSizeInKEntries * 1024 * 64 * AverageDCCCompressionRate){do { } while(0); }; |
| 5799 | #endif |
| 5800 | } |
| 5801 | |
| 5802 | #ifdef __DML_VBA_DEBUG__ |
| 5803 | dml_print("DML::%s: MetaFIFOSizeInKEntries = %d\n", __func__, MetaFIFOSizeInKEntries){do { } while(0); }; |
| 5804 | dml_print("DML::%s: AverageZeroSizeCompressionRate = %f\n", __func__, AverageZeroSizeCompressionRate){do { } while(0); }; |
| 5805 | dml_print("DML::%s: EffectiveCompressedBufferSize = %f\n", __func__, EffectiveCompressedBufferSize){do { } while(0); }; |
| 5806 | #endif |
| 5807 | |
| 5808 | *StutterPeriod = 0; |
| 5809 | |
| 5810 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 5811 | if (UseMALLForPStateChange[k] != dm_use_mall_pstate_change_phantom_pipe) { |
| 5812 | LinesInDETY = ((double) DETBufferSizeY[k] |
| 5813 | + (UnboundedRequestEnabled == true1 ? EffectiveCompressedBufferSize : 0) |
| 5814 | * ReadBandwidthSurfaceLuma[k] / TotalDataReadBandwidth) |
| 5815 | / BytePerPixelDETY[k] / SwathWidthY[k]; |
| 5816 | LinesInDETYRoundedDownToSwath = dml_floor(LinesInDETY, SwathHeightY[k]); |
| 5817 | DETBufferingTimeY = LinesInDETYRoundedDownToSwath * ((double) HTotal[k] / PixelClock[k]) |
| 5818 | / VRatio[k]; |
| 5819 | #ifdef __DML_VBA_DEBUG__ |
| 5820 | dml_print("DML::%s: k=%0d, DETBufferSizeY = %d\n", __func__, k, DETBufferSizeY[k]){do { } while(0); }; |
| 5821 | dml_print("DML::%s: k=%0d, BytePerPixelDETY = %f\n", __func__, k, BytePerPixelDETY[k]){do { } while(0); }; |
| 5822 | dml_print("DML::%s: k=%0d, SwathWidthY = %d\n", __func__, k, SwathWidthY[k]){do { } while(0); }; |
| 5823 | dml_print("DML::%s: k=%0d, ReadBandwidthSurfaceLuma = %f\n",{do { } while(0); } |
| 5824 | __func__, k, ReadBandwidthSurfaceLuma[k]){do { } while(0); }; |
| 5825 | dml_print("DML::%s: k=%0d, TotalDataReadBandwidth = %f\n", __func__, k, TotalDataReadBandwidth){do { } while(0); }; |
| 5826 | dml_print("DML::%s: k=%0d, LinesInDETY = %f\n", __func__, k, LinesInDETY){do { } while(0); }; |
| 5827 | dml_print("DML::%s: k=%0d, LinesInDETYRoundedDownToSwath = %f\n",{do { } while(0); } |
| 5828 | __func__, k, LinesInDETYRoundedDownToSwath){do { } while(0); }; |
| 5829 | dml_print("DML::%s: k=%0d, HTotal = %d\n", __func__, k, HTotal[k]){do { } while(0); }; |
| 5830 | dml_print("DML::%s: k=%0d, PixelClock = %f\n", __func__, k, PixelClock[k]){do { } while(0); }; |
| 5831 | dml_print("DML::%s: k=%0d, VRatio = %f\n", __func__, k, VRatio[k]){do { } while(0); }; |
| 5832 | dml_print("DML::%s: k=%0d, DETBufferingTimeY = %f\n", __func__, k, DETBufferingTimeY){do { } while(0); }; |
| 5833 | dml_print("DML::%s: k=%0d, PixelClock = %f\n", __func__, k, PixelClock[k]){do { } while(0); }; |
| 5834 | #endif |
| 5835 | |
| 5836 | if (!FoundCriticalSurface || DETBufferingTimeY < *StutterPeriod) { |
| 5837 | bool_Bool isInterlaceTiming = Interlace[k] && !ProgressiveToInterlaceUnitInOPP; |
| 5838 | |
| 5839 | FoundCriticalSurface = true1; |
| 5840 | *StutterPeriod = DETBufferingTimeY; |
| 5841 | FrameTimeCriticalSurface = ( |
| 5842 | isInterlaceTiming ? |
| 5843 | dml_floor((double) VTotal[k] / 2.0, 1.0) : VTotal[k]) |
| 5844 | * (double) HTotal[k] / PixelClock[k]; |
| 5845 | VActiveTimeCriticalSurface = ( |
| 5846 | isInterlaceTiming ? |
| 5847 | dml_floor((double) VActive[k] / 2.0, 1.0) : VActive[k]) |
| 5848 | * (double) HTotal[k] / PixelClock[k]; |
| 5849 | BytePerPixelYCriticalSurface = BytePerPixelY[k]; |
| 5850 | SwathWidthYCriticalSurface = SwathWidthY[k]; |
| 5851 | SwathHeightYCriticalSurface = SwathHeightY[k]; |
| 5852 | BlockWidth256BytesYCriticalSurface = BlockWidth256BytesY[k]; |
| 5853 | LinesToFinishSwathTransferStutterCriticalSurface = SwathHeightY[k] |
| 5854 | - (LinesInDETY - LinesInDETYRoundedDownToSwath); |
| 5855 | DETBufferSizeYCriticalSurface = DETBufferSizeY[k]; |
| 5856 | MinTTUVBlankCriticalSurface = MinTTUVBlank[k]; |
| 5857 | doublePlaneCriticalSurface = (ReadBandwidthSurfaceChroma[k] == 0); |
| 5858 | doublePipeCriticalSurface = (DPPPerSurface[k] == 1); |
| 5859 | |
| 5860 | #ifdef __DML_VBA_DEBUG__ |
| 5861 | dml_print("DML::%s: k=%0d, FoundCriticalSurface = %d\n",{do { } while(0); } |
| 5862 | __func__, k, FoundCriticalSurface){do { } while(0); }; |
| 5863 | dml_print("DML::%s: k=%0d, StutterPeriod = %f\n",{do { } while(0); } |
| 5864 | __func__, k, *StutterPeriod){do { } while(0); }; |
| 5865 | dml_print("DML::%s: k=%0d, MinTTUVBlankCriticalSurface = %f\n",{do { } while(0); } |
| 5866 | __func__, k, MinTTUVBlankCriticalSurface){do { } while(0); }; |
| 5867 | dml_print("DML::%s: k=%0d, FrameTimeCriticalSurface = %f\n",{do { } while(0); } |
| 5868 | __func__, k, FrameTimeCriticalSurface){do { } while(0); }; |
| 5869 | dml_print("DML::%s: k=%0d, VActiveTimeCriticalSurface = %f\n",{do { } while(0); } |
| 5870 | __func__, k, VActiveTimeCriticalSurface){do { } while(0); }; |
| 5871 | dml_print("DML::%s: k=%0d, BytePerPixelYCriticalSurface = %d\n",{do { } while(0); } |
| 5872 | __func__, k, BytePerPixelYCriticalSurface){do { } while(0); }; |
| 5873 | dml_print("DML::%s: k=%0d, SwathWidthYCriticalSurface = %f\n",{do { } while(0); } |
| 5874 | __func__, k, SwathWidthYCriticalSurface){do { } while(0); }; |
| 5875 | dml_print("DML::%s: k=%0d, SwathHeightYCriticalSurface = %f\n",{do { } while(0); } |
| 5876 | __func__, k, SwathHeightYCriticalSurface){do { } while(0); }; |
| 5877 | dml_print("DML::%s: k=%0d, BlockWidth256BytesYCriticalSurface = %d\n",{do { } while(0); } |
| 5878 | __func__, k, BlockWidth256BytesYCriticalSurface){do { } while(0); }; |
| 5879 | dml_print("DML::%s: k=%0d, doublePlaneCriticalSurface = %d\n",{do { } while(0); } |
| 5880 | __func__, k, doublePlaneCriticalSurface){do { } while(0); }; |
| 5881 | dml_print("DML::%s: k=%0d, doublePipeCriticalSurface = %d\n",{do { } while(0); } |
| 5882 | __func__, k, doublePipeCriticalSurface){do { } while(0); }; |
| 5883 | dml_print("DML::%s: k=%0d, LinesToFinishSwathTransferStutterCriticalSurface = %f\n",{do { } while(0); } |
| 5884 | __func__, k, LinesToFinishSwathTransferStutterCriticalSurface){do { } while(0); }; |
| 5885 | #endif |
| 5886 | } |
| 5887 | } |
| 5888 | } |
| 5889 | |
| 5890 | PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer = dml_min(*StutterPeriod * TotalDataReadBandwidth, |
| 5891 | EffectiveCompressedBufferSize); |
| 5892 | #ifdef __DML_VBA_DEBUG__ |
| 5893 | dml_print("DML::%s: ROBBufferSizeInKByte = %d\n", __func__, ROBBufferSizeInKByte){do { } while(0); }; |
| 5894 | dml_print("DML::%s: AverageDCCCompressionRate = %f\n", __func__, AverageDCCCompressionRate){do { } while(0); }; |
| 5895 | dml_print("DML::%s: StutterPeriod * TotalDataReadBandwidth = %f\n",{do { } while(0); } |
| 5896 | __func__, *StutterPeriod * TotalDataReadBandwidth){do { } while(0); }; |
| 5897 | dml_print("DML::%s: EffectiveCompressedBufferSize = %f\n", __func__, EffectiveCompressedBufferSize){do { } while(0); }; |
| 5898 | dml_print("DML::%s: PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer = %f\n", __func__,{do { } while(0); } |
| 5899 | PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer){do { } while(0); }; |
| 5900 | dml_print("DML::%s: ReturnBW = %f\n", __func__, ReturnBW){do { } while(0); }; |
| 5901 | dml_print("DML::%s: TotalDataReadBandwidth = %f\n", __func__, TotalDataReadBandwidth){do { } while(0); }; |
| 5902 | dml_print("DML::%s: TotalRowReadBandwidth = %f\n", __func__, TotalRowReadBandwidth){do { } while(0); }; |
| 5903 | dml_print("DML::%s: DCFCLK = %f\n", __func__, DCFCLK){do { } while(0); }; |
| 5904 | #endif |
| 5905 | |
| 5906 | StutterBurstTime = PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer / AverageDCCCompressionRate |
| 5907 | / ReturnBW |
| 5908 | + (*StutterPeriod * TotalDataReadBandwidth |
| 5909 | - PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer) / (DCFCLK * 64) |
| 5910 | + *StutterPeriod * TotalRowReadBandwidth / ReturnBW; |
| 5911 | #ifdef __DML_VBA_DEBUG__ |
| 5912 | dml_print("DML::%s: Part 1 = %f\n", __func__, PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer /{do { } while(0); } |
| 5913 | AverageDCCCompressionRate / ReturnBW){do { } while(0); }; |
| 5914 | dml_print("DML::%s: StutterPeriod * TotalDataReadBandwidth = %f\n",{do { } while(0); } |
| 5915 | __func__, (*StutterPeriod * TotalDataReadBandwidth)){do { } while(0); }; |
| 5916 | dml_print("DML::%s: Part 2 = %f\n", __func__, (*StutterPeriod * TotalDataReadBandwidth -{do { } while(0); } |
| 5917 | PartOfUncompressedPixelBurstThatFitsInROBAndCompressedBuffer) / (DCFCLK * 64)){do { } while(0); }; |
| 5918 | dml_print("DML::%s: Part 3 = %f\n", __func__, *StutterPeriod * TotalRowReadBandwidth / ReturnBW){do { } while(0); }; |
| 5919 | dml_print("DML::%s: StutterBurstTime = %f\n", __func__, StutterBurstTime){do { } while(0); }; |
| 5920 | #endif |
| 5921 | StutterBurstTime = dml_max(StutterBurstTime, |
| 5922 | LinesToFinishSwathTransferStutterCriticalSurface * BytePerPixelYCriticalSurface |
| 5923 | * SwathWidthYCriticalSurface / ReturnBW); |
| 5924 | |
| 5925 | #ifdef __DML_VBA_DEBUG__ |
| 5926 | dml_print("DML::%s: Time to finish residue swath=%f\n",{do { } while(0); } |
| 5927 | __func__,{do { } while(0); } |
| 5928 | LinesToFinishSwathTransferStutterCriticalSurface *{do { } while(0); } |
| 5929 | BytePerPixelYCriticalSurface * SwathWidthYCriticalSurface / ReturnBW){do { } while(0); }; |
| 5930 | #endif |
| 5931 | |
| 5932 | TotalActiveWriteback = 0; |
| 5933 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 5934 | if (WritebackEnable[k]) |
| 5935 | TotalActiveWriteback = TotalActiveWriteback + 1; |
| 5936 | } |
| 5937 | |
| 5938 | if (TotalActiveWriteback == 0) { |
| 5939 | #ifdef __DML_VBA_DEBUG__ |
| 5940 | dml_print("DML::%s: SRExitTime = %f\n", __func__, SRExitTime){do { } while(0); }; |
| 5941 | dml_print("DML::%s: SRExitZ8Time = %f\n", __func__, SRExitZ8Time){do { } while(0); }; |
| 5942 | dml_print("DML::%s: StutterBurstTime = %f (final)\n", __func__, StutterBurstTime){do { } while(0); }; |
| 5943 | dml_print("DML::%s: StutterPeriod = %f\n", __func__, *StutterPeriod){do { } while(0); }; |
| 5944 | #endif |
| 5945 | *StutterEfficiencyNotIncludingVBlank = dml_max(0., |
| 5946 | 1 - (SRExitTime + StutterBurstTime) / *StutterPeriod) * 100; |
| 5947 | *Z8StutterEfficiencyNotIncludingVBlank = dml_max(0., |
| 5948 | 1 - (SRExitZ8Time + StutterBurstTime) / *StutterPeriod) * 100; |
| 5949 | *NumberOfStutterBurstsPerFrame = ( |
| 5950 | *StutterEfficiencyNotIncludingVBlank > 0 ? |
| 5951 | dml_ceil(VActiveTimeCriticalSurface / *StutterPeriod, 1) : 0); |
| 5952 | *Z8NumberOfStutterBurstsPerFrame = ( |
| 5953 | *Z8StutterEfficiencyNotIncludingVBlank > 0 ? |
| 5954 | dml_ceil(VActiveTimeCriticalSurface / *StutterPeriod, 1) : 0); |
| 5955 | } else { |
| 5956 | *StutterEfficiencyNotIncludingVBlank = 0.; |
| 5957 | *Z8StutterEfficiencyNotIncludingVBlank = 0.; |
| 5958 | *NumberOfStutterBurstsPerFrame = 0; |
| 5959 | *Z8NumberOfStutterBurstsPerFrame = 0; |
| 5960 | } |
| 5961 | #ifdef __DML_VBA_DEBUG__ |
| 5962 | dml_print("DML::%s: VActiveTimeCriticalSurface = %f\n", __func__, VActiveTimeCriticalSurface){do { } while(0); }; |
| 5963 | dml_print("DML::%s: StutterEfficiencyNotIncludingVBlank = %f\n",{do { } while(0); } |
| 5964 | __func__, *StutterEfficiencyNotIncludingVBlank){do { } while(0); }; |
| 5965 | dml_print("DML::%s: Z8StutterEfficiencyNotIncludingVBlank = %f\n",{do { } while(0); } |
| 5966 | __func__, *Z8StutterEfficiencyNotIncludingVBlank){do { } while(0); }; |
| 5967 | dml_print("DML::%s: NumberOfStutterBurstsPerFrame = %d\n", __func__, *NumberOfStutterBurstsPerFrame){do { } while(0); }; |
| 5968 | dml_print("DML::%s: Z8NumberOfStutterBurstsPerFrame = %d\n", __func__, *Z8NumberOfStutterBurstsPerFrame){do { } while(0); }; |
| 5969 | #endif |
| 5970 | |
| 5971 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 5972 | if (UseMALLForPStateChange[k] != dm_use_mall_pstate_change_phantom_pipe) { |
| 5973 | if (BlendingAndTiming[k] == k) { |
| 5974 | if (TotalNumberOfActiveOTG == 0) { |
| 5975 | doublePixelClock = PixelClock[k]; |
| 5976 | doubleHTotal = HTotal[k]; |
| 5977 | doubleVTotal = VTotal[k]; |
| 5978 | } else if (doublePixelClock != PixelClock[k] || doubleHTotal != HTotal[k] |
| 5979 | || doubleVTotal != VTotal[k]) { |
| 5980 | SameTiming = false0; |
| 5981 | } |
| 5982 | TotalNumberOfActiveOTG = TotalNumberOfActiveOTG + 1; |
| 5983 | } |
| 5984 | } |
| 5985 | } |
| 5986 | |
| 5987 | if (*StutterEfficiencyNotIncludingVBlank > 0) { |
| 5988 | LastStutterPeriod = VActiveTimeCriticalSurface - (*NumberOfStutterBurstsPerFrame - 1) * *StutterPeriod; |
| 5989 | |
| 5990 | if ((SynchronizeTimingsFinal || TotalNumberOfActiveOTG == 1) && SameTiming |
| 5991 | && LastStutterPeriod + MinTTUVBlankCriticalSurface > StutterEnterPlusExitWatermark) { |
| 5992 | *StutterEfficiency = (1 - (*NumberOfStutterBurstsPerFrame * SRExitTime |
| 5993 | + StutterBurstTime * VActiveTimeCriticalSurface |
| 5994 | / *StutterPeriod) / FrameTimeCriticalSurface) * 100; |
| 5995 | } else { |
| 5996 | *StutterEfficiency = *StutterEfficiencyNotIncludingVBlank; |
| 5997 | } |
| 5998 | } else { |
| 5999 | *StutterEfficiency = 0; |
| 6000 | } |
| 6001 | |
| 6002 | if (*Z8StutterEfficiencyNotIncludingVBlank > 0) { |
| 6003 | LastZ8StutterPeriod = VActiveTimeCriticalSurface |
| 6004 | - (*NumberOfStutterBurstsPerFrame - 1) * *StutterPeriod; |
| 6005 | if ((SynchronizeTimingsFinal || TotalNumberOfActiveOTG == 1) && SameTiming && LastZ8StutterPeriod + |
| 6006 | MinTTUVBlankCriticalSurface > Z8StutterEnterPlusExitWatermark) { |
| 6007 | *Z8StutterEfficiency = (1 - (*NumberOfStutterBurstsPerFrame * SRExitZ8Time + StutterBurstTime |
| 6008 | * VActiveTimeCriticalSurface / *StutterPeriod) / FrameTimeCriticalSurface) * 100; |
| 6009 | } else { |
| 6010 | *Z8StutterEfficiency = *Z8StutterEfficiencyNotIncludingVBlank; |
| 6011 | } |
| 6012 | } else { |
| 6013 | *Z8StutterEfficiency = 0.; |
| 6014 | } |
| 6015 | |
| 6016 | #ifdef __DML_VBA_DEBUG__ |
| 6017 | dml_print("DML::%s: LastZ8StutterPeriod = %f\n", __func__, LastZ8StutterPeriod){do { } while(0); }; |
| 6018 | dml_print("DML::%s: Z8StutterEnterPlusExitWatermark = %f\n", __func__, Z8StutterEnterPlusExitWatermark){do { } while(0); }; |
| 6019 | dml_print("DML::%s: StutterBurstTime = %f\n", __func__, StutterBurstTime){do { } while(0); }; |
| 6020 | dml_print("DML::%s: StutterPeriod = %f\n", __func__, *StutterPeriod){do { } while(0); }; |
| 6021 | dml_print("DML::%s: StutterEfficiency = %f\n", __func__, *StutterEfficiency){do { } while(0); }; |
| 6022 | dml_print("DML::%s: Z8StutterEfficiency = %f\n", __func__, *Z8StutterEfficiency){do { } while(0); }; |
| 6023 | dml_print("DML::%s: StutterEfficiencyNotIncludingVBlank = %f\n",{do { } while(0); } |
| 6024 | __func__, *StutterEfficiencyNotIncludingVBlank){do { } while(0); }; |
| 6025 | dml_print("DML::%s: Z8NumberOfStutterBurstsPerFrame = %d\n", __func__, *Z8NumberOfStutterBurstsPerFrame){do { } while(0); }; |
| 6026 | #endif |
| 6027 | |
| 6028 | SwathSizeCriticalSurface = BytePerPixelYCriticalSurface * SwathHeightYCriticalSurface |
| 6029 | * dml_ceil(SwathWidthYCriticalSurface, BlockWidth256BytesYCriticalSurface); |
| 6030 | LastChunkOfSwathSize = SwathSizeCriticalSurface % (PixelChunkSizeInKByte * 1024); |
| 6031 | MissingPartOfLastSwathOfDETSize = dml_ceil(DETBufferSizeYCriticalSurface, SwathSizeCriticalSurface) |
| 6032 | - DETBufferSizeYCriticalSurface; |
| 6033 | |
| 6034 | *DCHUBBUB_ARB_CSTATE_MAX_CAP_MODE = !(!UnboundedRequestEnabled && (NumberOfActiveSurfaces == 1) |
| 6035 | && doublePlaneCriticalSurface && doublePipeCriticalSurface && (LastChunkOfSwathSize > 0) |
| 6036 | && (LastChunkOfSwathSize <= 4096) && (MissingPartOfLastSwathOfDETSize > 0) |
| 6037 | && (MissingPartOfLastSwathOfDETSize <= LastChunkOfSwathSize)); |
| 6038 | |
| 6039 | #ifdef __DML_VBA_DEBUG__ |
| 6040 | dml_print("DML::%s: SwathSizeCriticalSurface = %d\n", __func__, SwathSizeCriticalSurface){do { } while(0); }; |
| 6041 | dml_print("DML::%s: LastChunkOfSwathSize = %d\n", __func__, LastChunkOfSwathSize){do { } while(0); }; |
| 6042 | dml_print("DML::%s: MissingPartOfLastSwathOfDETSize = %d\n", __func__, MissingPartOfLastSwathOfDETSize){do { } while(0); }; |
| 6043 | dml_print("DML::%s: DCHUBBUB_ARB_CSTATE_MAX_CAP_MODE = %d\n", __func__, *DCHUBBUB_ARB_CSTATE_MAX_CAP_MODE){do { } while(0); }; |
| 6044 | #endif |
| 6045 | } // CalculateStutterEfficiency |
| 6046 | |
| 6047 | void dml32_CalculateMaxDETAndMinCompressedBufferSize( |
| 6048 | unsigned int ConfigReturnBufferSizeInKByte, |
| 6049 | unsigned int ROBBufferSizeInKByte, |
| 6050 | unsigned int MaxNumDPP, |
| 6051 | bool_Bool nomDETInKByteOverrideEnable, // VBA_DELTA, allow DV to override default DET size |
| 6052 | unsigned int nomDETInKByteOverrideValue, // VBA_DELTA |
| 6053 | |
| 6054 | /* Output */ |
| 6055 | unsigned int *MaxTotalDETInKByte, |
| 6056 | unsigned int *nomDETInKByte, |
| 6057 | unsigned int *MinCompressedBufferSizeInKByte) |
| 6058 | { |
| 6059 | bool_Bool det_buff_size_override_en = nomDETInKByteOverrideEnable; |
| 6060 | unsigned int det_buff_size_override_val = nomDETInKByteOverrideValue; |
| 6061 | |
| 6062 | *MaxTotalDETInKByte = dml_ceil(((double)ConfigReturnBufferSizeInKByte + |
| 6063 | (double) ROBBufferSizeInKByte) * 4.0 / 5.0, 64); |
| 6064 | *nomDETInKByte = dml_floor((double) *MaxTotalDETInKByte / (double) MaxNumDPP, 64); |
| 6065 | *MinCompressedBufferSizeInKByte = ConfigReturnBufferSizeInKByte - *MaxTotalDETInKByte; |
| 6066 | |
| 6067 | #ifdef __DML_VBA_DEBUG__ |
| 6068 | dml_print("DML::%s: ConfigReturnBufferSizeInKByte = %0d\n", __func__, ConfigReturnBufferSizeInKByte){do { } while(0); }; |
| 6069 | dml_print("DML::%s: ROBBufferSizeInKByte = %0d\n", __func__, ROBBufferSizeInKByte){do { } while(0); }; |
| 6070 | dml_print("DML::%s: MaxNumDPP = %0d\n", __func__, MaxNumDPP){do { } while(0); }; |
| 6071 | dml_print("DML::%s: MaxTotalDETInKByte = %0d\n", __func__, *MaxTotalDETInKByte){do { } while(0); }; |
| 6072 | dml_print("DML::%s: nomDETInKByte = %0d\n", __func__, *nomDETInKByte){do { } while(0); }; |
| 6073 | dml_print("DML::%s: MinCompressedBufferSizeInKByte = %0d\n", __func__, *MinCompressedBufferSizeInKByte){do { } while(0); }; |
| 6074 | #endif |
| 6075 | |
| 6076 | if (det_buff_size_override_en) { |
| 6077 | *nomDETInKByte = det_buff_size_override_val; |
| 6078 | #ifdef __DML_VBA_DEBUG__ |
| 6079 | dml_print("DML::%s: nomDETInKByte = %0d (override)\n", __func__, *nomDETInKByte){do { } while(0); }; |
| 6080 | #endif |
| 6081 | } |
| 6082 | } // CalculateMaxDETAndMinCompressedBufferSize |
| 6083 | |
| 6084 | bool_Bool dml32_CalculateVActiveBandwithSupport(unsigned int NumberOfActiveSurfaces, |
| 6085 | double ReturnBW, |
| 6086 | bool_Bool NotUrgentLatencyHiding[], |
| 6087 | double ReadBandwidthLuma[], |
| 6088 | double ReadBandwidthChroma[], |
| 6089 | double cursor_bw[], |
| 6090 | double meta_row_bandwidth[], |
| 6091 | double dpte_row_bandwidth[], |
| 6092 | unsigned int NumberOfDPP[], |
| 6093 | double UrgentBurstFactorLuma[], |
| 6094 | double UrgentBurstFactorChroma[], |
| 6095 | double UrgentBurstFactorCursor[]) |
| 6096 | { |
| 6097 | unsigned int k; |
| 6098 | bool_Bool NotEnoughUrgentLatencyHiding = false0; |
| 6099 | bool_Bool CalculateVActiveBandwithSupport_val = false0; |
| 6100 | double VActiveBandwith = 0; |
| 6101 | |
| 6102 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 6103 | if (NotUrgentLatencyHiding[k]) { |
| 6104 | NotEnoughUrgentLatencyHiding = true1; |
| 6105 | } |
| 6106 | } |
| 6107 | |
| 6108 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 6109 | VActiveBandwith = VActiveBandwith + ReadBandwidthLuma[k] * UrgentBurstFactorLuma[k] + ReadBandwidthChroma[k] * UrgentBurstFactorChroma[k] + cursor_bw[k] * UrgentBurstFactorCursor[k] + NumberOfDPP[k] * meta_row_bandwidth[k] + NumberOfDPP[k] * dpte_row_bandwidth[k]; |
| 6110 | } |
| 6111 | |
| 6112 | CalculateVActiveBandwithSupport_val = (VActiveBandwith <= ReturnBW) && !NotEnoughUrgentLatencyHiding; |
| 6113 | |
| 6114 | #ifdef __DML_VBA_DEBUG__ |
| 6115 | dml_print("DML::%s: NotEnoughUrgentLatencyHiding = %d\n", __func__, NotEnoughUrgentLatencyHiding){do { } while(0); }; |
| 6116 | dml_print("DML::%s: VActiveBandwith = %f\n", __func__, VActiveBandwith){do { } while(0); }; |
| 6117 | dml_print("DML::%s: ReturnBW = %f\n", __func__, ReturnBW){do { } while(0); }; |
| 6118 | dml_print("DML::%s: CalculateVActiveBandwithSupport_val = %d\n", __func__, CalculateVActiveBandwithSupport_val){do { } while(0); }; |
| 6119 | #endif |
| 6120 | return CalculateVActiveBandwithSupport_val; |
| 6121 | } |
| 6122 | |
| 6123 | void dml32_CalculatePrefetchBandwithSupport(unsigned int NumberOfActiveSurfaces, |
| 6124 | double ReturnBW, |
| 6125 | bool_Bool NotUrgentLatencyHiding[], |
| 6126 | double ReadBandwidthLuma[], |
| 6127 | double ReadBandwidthChroma[], |
| 6128 | double PrefetchBandwidthLuma[], |
| 6129 | double PrefetchBandwidthChroma[], |
| 6130 | double cursor_bw[], |
| 6131 | double meta_row_bandwidth[], |
| 6132 | double dpte_row_bandwidth[], |
| 6133 | double cursor_bw_pre[], |
| 6134 | double prefetch_vmrow_bw[], |
| 6135 | unsigned int NumberOfDPP[], |
| 6136 | double UrgentBurstFactorLuma[], |
| 6137 | double UrgentBurstFactorChroma[], |
| 6138 | double UrgentBurstFactorCursor[], |
| 6139 | double UrgentBurstFactorLumaPre[], |
| 6140 | double UrgentBurstFactorChromaPre[], |
| 6141 | double UrgentBurstFactorCursorPre[], |
| 6142 | |
| 6143 | /* output */ |
| 6144 | double *PrefetchBandwidth, |
| 6145 | double *FractionOfUrgentBandwidth, |
| 6146 | bool_Bool *PrefetchBandwidthSupport) |
| 6147 | { |
| 6148 | unsigned int k; |
| 6149 | bool_Bool NotEnoughUrgentLatencyHiding = false0; |
| 6150 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 6151 | if (NotUrgentLatencyHiding[k]) { |
| 6152 | NotEnoughUrgentLatencyHiding = true1; |
| 6153 | } |
| 6154 | } |
| 6155 | |
| 6156 | *PrefetchBandwidth = 0; |
| 6157 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 6158 | *PrefetchBandwidth = *PrefetchBandwidth + dml_max3(NumberOfDPP[k] * prefetch_vmrow_bw[k], |
| 6159 | ReadBandwidthLuma[k] * UrgentBurstFactorLuma[k] + ReadBandwidthChroma[k] * UrgentBurstFactorChroma[k] + cursor_bw[k] * UrgentBurstFactorCursor[k] + NumberOfDPP[k] * (meta_row_bandwidth[k] + dpte_row_bandwidth[k]), |
| 6160 | NumberOfDPP[k] * (PrefetchBandwidthLuma[k] * UrgentBurstFactorLumaPre[k] + PrefetchBandwidthChroma[k] * UrgentBurstFactorChromaPre[k]) + cursor_bw_pre[k] * UrgentBurstFactorCursorPre[k]); |
| 6161 | } |
| 6162 | |
| 6163 | *PrefetchBandwidthSupport = (*PrefetchBandwidth <= ReturnBW) && !NotEnoughUrgentLatencyHiding; |
| 6164 | *FractionOfUrgentBandwidth = *PrefetchBandwidth / ReturnBW; |
| 6165 | } |
| 6166 | |
| 6167 | double dml32_CalculateBandwidthAvailableForImmediateFlip(unsigned int NumberOfActiveSurfaces, |
| 6168 | double ReturnBW, |
| 6169 | double ReadBandwidthLuma[], |
| 6170 | double ReadBandwidthChroma[], |
| 6171 | double PrefetchBandwidthLuma[], |
| 6172 | double PrefetchBandwidthChroma[], |
| 6173 | double cursor_bw[], |
| 6174 | double cursor_bw_pre[], |
| 6175 | unsigned int NumberOfDPP[], |
| 6176 | double UrgentBurstFactorLuma[], |
| 6177 | double UrgentBurstFactorChroma[], |
| 6178 | double UrgentBurstFactorCursor[], |
| 6179 | double UrgentBurstFactorLumaPre[], |
| 6180 | double UrgentBurstFactorChromaPre[], |
| 6181 | double UrgentBurstFactorCursorPre[]) |
| 6182 | { |
| 6183 | unsigned int k; |
| 6184 | double CalculateBandwidthAvailableForImmediateFlip_val = ReturnBW; |
| 6185 | |
| 6186 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 6187 | CalculateBandwidthAvailableForImmediateFlip_val = CalculateBandwidthAvailableForImmediateFlip_val - dml_max(ReadBandwidthLuma[k] * UrgentBurstFactorLuma[k] + ReadBandwidthChroma[k] * UrgentBurstFactorChroma[k] + cursor_bw[k] * UrgentBurstFactorCursor[k], |
| 6188 | NumberOfDPP[k] * (PrefetchBandwidthLuma[k] * UrgentBurstFactorLumaPre[k] + PrefetchBandwidthChroma[k] * UrgentBurstFactorChromaPre[k]) + cursor_bw_pre[k] * UrgentBurstFactorCursorPre[k]); |
| 6189 | } |
| 6190 | |
| 6191 | return CalculateBandwidthAvailableForImmediateFlip_val; |
| 6192 | } |
| 6193 | |
| 6194 | void dml32_CalculateImmediateFlipBandwithSupport(unsigned int NumberOfActiveSurfaces, |
| 6195 | double ReturnBW, |
| 6196 | enum immediate_flip_requirement ImmediateFlipRequirement[], |
| 6197 | double final_flip_bw[], |
| 6198 | double ReadBandwidthLuma[], |
| 6199 | double ReadBandwidthChroma[], |
| 6200 | double PrefetchBandwidthLuma[], |
| 6201 | double PrefetchBandwidthChroma[], |
| 6202 | double cursor_bw[], |
| 6203 | double meta_row_bandwidth[], |
| 6204 | double dpte_row_bandwidth[], |
| 6205 | double cursor_bw_pre[], |
| 6206 | double prefetch_vmrow_bw[], |
| 6207 | unsigned int NumberOfDPP[], |
| 6208 | double UrgentBurstFactorLuma[], |
| 6209 | double UrgentBurstFactorChroma[], |
| 6210 | double UrgentBurstFactorCursor[], |
| 6211 | double UrgentBurstFactorLumaPre[], |
| 6212 | double UrgentBurstFactorChromaPre[], |
| 6213 | double UrgentBurstFactorCursorPre[], |
| 6214 | |
| 6215 | /* output */ |
| 6216 | double *TotalBandwidth, |
| 6217 | double *FractionOfUrgentBandwidth, |
| 6218 | bool_Bool *ImmediateFlipBandwidthSupport) |
| 6219 | { |
| 6220 | unsigned int k; |
| 6221 | *TotalBandwidth = 0; |
| 6222 | for (k = 0; k < NumberOfActiveSurfaces; ++k) { |
| 6223 | if (ImmediateFlipRequirement[k] != dm_immediate_flip_not_required) { |
| 6224 | *TotalBandwidth = *TotalBandwidth + dml_max3(NumberOfDPP[k] * prefetch_vmrow_bw[k], |
| 6225 | NumberOfDPP[k] * final_flip_bw[k] + ReadBandwidthLuma[k] * UrgentBurstFactorLuma[k] + ReadBandwidthChroma[k] * UrgentBurstFactorChroma[k] + cursor_bw[k] * UrgentBurstFactorCursor[k], |
| 6226 | NumberOfDPP[k] * (final_flip_bw[k] + PrefetchBandwidthLuma[k] * UrgentBurstFactorLumaPre[k] + PrefetchBandwidthChroma[k] * UrgentBurstFactorChromaPre[k]) + cursor_bw_pre[k] * UrgentBurstFactorCursorPre[k]); |
| 6227 | } else { |
| 6228 | *TotalBandwidth = *TotalBandwidth + dml_max3(NumberOfDPP[k] * prefetch_vmrow_bw[k], |
| 6229 | NumberOfDPP[k] * (meta_row_bandwidth[k] + dpte_row_bandwidth[k]) + ReadBandwidthLuma[k] * UrgentBurstFactorLuma[k] + ReadBandwidthChroma[k] * UrgentBurstFactorChroma[k] + cursor_bw[k] * UrgentBurstFactorCursor[k], |
| 6230 | NumberOfDPP[k] * (PrefetchBandwidthLuma[k] * UrgentBurstFactorLumaPre[k] + PrefetchBandwidthChroma[k] * UrgentBurstFactorChromaPre[k]) + cursor_bw_pre[k] * UrgentBurstFactorCursorPre[k]); |
| 6231 | } |
| 6232 | } |
| 6233 | *ImmediateFlipBandwidthSupport = (*TotalBandwidth <= ReturnBW); |
| 6234 | *FractionOfUrgentBandwidth = *TotalBandwidth / ReturnBW; |
| 6235 | } |
| 6236 | |
| 6237 | bool_Bool dml32_CalculateDETSwathFillLatencyHiding(unsigned int NumberOfActiveSurfaces, |
| 6238 | double ReturnBW, |
| 6239 | double UrgentLatency, |
| 6240 | unsigned int SwathHeightY[], |
| 6241 | unsigned int SwathHeightC[], |
| 6242 | unsigned int SwathWidthY[], |
| 6243 | unsigned int SwathWidthC[], |
| 6244 | double BytePerPixelInDETY[], |
| 6245 | double BytePerPixelInDETC[], |
| 6246 | unsigned int DETBufferSizeY[], |
| 6247 | unsigned int DETBufferSizeC[], |
| 6248 | unsigned int NumOfDPP[], |
| 6249 | unsigned int HTotal[], |
| 6250 | double PixelClock[], |
| 6251 | double VRatioY[], |
| 6252 | double VRatioC[], |
| 6253 | enum dm_use_mall_for_pstate_change_mode UsesMALLForPStateChange[DC__NUM_DPP__MAX8]) |
| 6254 | { |
| 6255 | int k; |
| 6256 | double SwathSizeAllSurfaces = 0; |
| 6257 | double SwathSizeAllSurfacesInFetchTimeUs; |
| 6258 | double DETSwathLatencyHidingUs; |
| 6259 | double DETSwathLatencyHidingYUs; |
| 6260 | double DETSwathLatencyHidingCUs; |
| 6261 | double SwathSizePerSurfaceY[DC__NUM_DPP__MAX8]; |
| 6262 | double SwathSizePerSurfaceC[DC__NUM_DPP__MAX8]; |
| 6263 | bool_Bool NotEnoughDETSwathFillLatencyHiding = false0; |
| 6264 | |
| 6265 | /* calculate sum of single swath size for all pipes in bytes */ |
| 6266 | for (k = 0; k < NumberOfActiveSurfaces; k++) { |
| 6267 | SwathSizePerSurfaceY[k] = SwathHeightY[k] * SwathWidthY[k] * BytePerPixelInDETY[k] * NumOfDPP[k]; |
| 6268 | |
| 6269 | if (SwathHeightC[k] != 0) |
| 6270 | SwathSizePerSurfaceC[k] = SwathHeightC[k] * SwathWidthC[k] * BytePerPixelInDETC[k] * NumOfDPP[k]; |
| 6271 | else |
| 6272 | SwathSizePerSurfaceC[k] = 0; |
| 6273 | |
| 6274 | SwathSizeAllSurfaces += SwathSizePerSurfaceY[k] + SwathSizePerSurfaceC[k]; |
| 6275 | } |
| 6276 | |
| 6277 | SwathSizeAllSurfacesInFetchTimeUs = SwathSizeAllSurfaces / ReturnBW + UrgentLatency; |
| 6278 | |
| 6279 | /* ensure all DET - 1 swath can hide a fetch for all surfaces */ |
| 6280 | for (k = 0; k < NumberOfActiveSurfaces; k++) { |
| 6281 | double LineTime = HTotal[k] / PixelClock[k]; |
| 6282 | |
| 6283 | /* only care if surface is not phantom */ |
| 6284 | if (UsesMALLForPStateChange[k] != dm_use_mall_pstate_change_phantom_pipe) { |
| 6285 | DETSwathLatencyHidingYUs = (dml_floor(DETBufferSizeY[k] / BytePerPixelInDETY[k] / SwathWidthY[k], 1.0) - SwathHeightY[k]) / VRatioY[k] * LineTime; |
| 6286 | |
| 6287 | if (SwathHeightC[k] != 0) { |
| 6288 | DETSwathLatencyHidingCUs = (dml_floor(DETBufferSizeC[k] / BytePerPixelInDETC[k] / SwathWidthC[k], 1.0) - SwathHeightC[k]) / VRatioC[k] * LineTime; |
| 6289 | |
| 6290 | DETSwathLatencyHidingUs = dml_min(DETSwathLatencyHidingYUs, DETSwathLatencyHidingCUs); |
| 6291 | } else { |
| 6292 | DETSwathLatencyHidingUs = DETSwathLatencyHidingYUs; |
| 6293 | } |
| 6294 | |
| 6295 | /* DET must be able to hide time to fetch 1 swath for each surface */ |
| 6296 | if (DETSwathLatencyHidingUs < SwathSizeAllSurfacesInFetchTimeUs) { |
| 6297 | NotEnoughDETSwathFillLatencyHiding = true1; |
| 6298 | break; |
| 6299 | } |
| 6300 | } |
| 6301 | } |
| 6302 | |
| 6303 | return NotEnoughDETSwathFillLatencyHiding; |
| 6304 | } |