src/gallium/drivers/swr/rasterizer/core/backend.h

   1 /****************************************************************************
   2 * Copyright (C) 2014-2015 Intel Corporation.   All Rights Reserved.
   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 (including the next
  12 * paragraph) shall be included in all copies or substantial portions of the
  13 * Software.
  14 *
  15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21 * IN THE SOFTWARE.
  22 *
  23 * @file backend.h
  24 *
  25 * @brief Backend handles rasterization, pixel shading and output merger
  26 *        operations.
  27 *
  28 ******************************************************************************/
  29 #pragma once
  30
  31 #include "common/os.h"
  32 #include "core/context.h"
  33 #include "core/multisample.h"
  34 #include "rdtsc_core.h"
  35
  36 void ProcessComputeBE(DRAW_CONTEXT* pDC, uint32_t workerId, uint32_t threadGroupId, void*& pSpillFillBuffer);
  37 void ProcessSyncBE(DRAW_CONTEXT *pDC, uint32_t workerId, uint32_t macroTile, void *pUserData);
  38 void ProcessQueryStatsBE(DRAW_CONTEXT *pDC, uint32_t workerId, uint32_t macroTile, void *pUserData);
  39 void ProcessClearBE(DRAW_CONTEXT *pDC, uint32_t workerId, uint32_t macroTile, void *pUserData);
  40 void ProcessStoreTileBE(DRAW_CONTEXT *pDC, uint32_t workerId, uint32_t macroTile, void *pData);
  41 void ProcessDiscardInvalidateTilesBE(DRAW_CONTEXT *pDC, uint32_t workerId, uint32_t macroTile, void *pData);
  42 void BackendNullPS(DRAW_CONTEXT *pDC, uint32_t workerId, uint32_t x, uint32_t y, SWR_TRIANGLE_DESC &work, RenderOutputBuffers &renderBuffers);
  43 void InitClearTilesTable();
  44 simdmask ComputeUserClipMask(uint8_t clipMask, float* pUserClipBuffer, simdscalar vI, simdscalar vJ);
  45 void InitBackendFuncTables();
  46 void InitCPSFuncTables();
  47 void CalcSampleBarycentrics(const BarycentricCoeffs& coeffs, SWR_PS_CONTEXT &psContext);
  48
  49 enum SWR_BACKEND_FUNCS
  50 {
  51     SWR_BACKEND_SINGLE_SAMPLE,
  52     SWR_BACKEND_MSAA_PIXEL_RATE,
  53     SWR_BACKEND_MSAA_SAMPLE_RATE,
  54     SWR_BACKEND_FUNCS_MAX,
  55 };
  56
  57 #if KNOB_SIMD_WIDTH == 8
  58 extern const __m256 vCenterOffsetsX;
  59 extern const __m256 vCenterOffsetsY;
  60 extern const __m256 vULOffsetsX;
  61 extern const __m256 vULOffsetsY;
  62 #define MASK 0xff
  63 #endif
  64
  65 INLINE static uint32_t RasterTileColorOffset(uint32_t sampleNum)
  66 {
  67     static const uint32_t RasterTileColorOffsets[16]
  68     { 0,
  69       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_COLOR_HOT_TILE_FORMAT>::bpp / 8),
  70       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_COLOR_HOT_TILE_FORMAT>::bpp / 8) * 2,
  71       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_COLOR_HOT_TILE_FORMAT>::bpp / 8) * 3,
  72       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_COLOR_HOT_TILE_FORMAT>::bpp / 8) * 4,
  73       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_COLOR_HOT_TILE_FORMAT>::bpp / 8) * 5,
  74       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_COLOR_HOT_TILE_FORMAT>::bpp / 8) * 6,
  75       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_COLOR_HOT_TILE_FORMAT>::bpp / 8) * 7,
  76       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_COLOR_HOT_TILE_FORMAT>::bpp / 8) * 8,
  77       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_COLOR_HOT_TILE_FORMAT>::bpp / 8) * 9,
  78       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_COLOR_HOT_TILE_FORMAT>::bpp / 8) * 10,
  79       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_COLOR_HOT_TILE_FORMAT>::bpp / 8) * 11,
  80       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_COLOR_HOT_TILE_FORMAT>::bpp / 8) * 12,
  81       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_COLOR_HOT_TILE_FORMAT>::bpp / 8) * 13,
  82       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_COLOR_HOT_TILE_FORMAT>::bpp / 8) * 14,
  83       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_COLOR_HOT_TILE_FORMAT>::bpp / 8) * 15,
  84     };
  85     assert(sampleNum < 16);
  86     return RasterTileColorOffsets[sampleNum];
  87 }
  88
  89 INLINE static uint32_t RasterTileDepthOffset(uint32_t sampleNum)
  90 {
  91     static const uint32_t RasterTileDepthOffsets[16]
  92     { 0,
  93       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_DEPTH_HOT_TILE_FORMAT>::bpp / 8),
  94       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_DEPTH_HOT_TILE_FORMAT>::bpp / 8) * 2,
  95       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_DEPTH_HOT_TILE_FORMAT>::bpp / 8) * 3,
  96       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_DEPTH_HOT_TILE_FORMAT>::bpp / 8) * 4,
  97       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_DEPTH_HOT_TILE_FORMAT>::bpp / 8) * 5,
  98       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_DEPTH_HOT_TILE_FORMAT>::bpp / 8) * 6,
  99       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_DEPTH_HOT_TILE_FORMAT>::bpp / 8) * 7,
 100       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_DEPTH_HOT_TILE_FORMAT>::bpp / 8) * 8,
 101       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_DEPTH_HOT_TILE_FORMAT>::bpp / 8) * 9,
 102       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_DEPTH_HOT_TILE_FORMAT>::bpp / 8) * 10,
 103       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_DEPTH_HOT_TILE_FORMAT>::bpp / 8) * 11,
 104       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_DEPTH_HOT_TILE_FORMAT>::bpp / 8) * 12,
 105       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_DEPTH_HOT_TILE_FORMAT>::bpp / 8) * 13,
 106       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_DEPTH_HOT_TILE_FORMAT>::bpp / 8) * 14,
 107       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_DEPTH_HOT_TILE_FORMAT>::bpp / 8) * 15,
 108     };
 109     assert(sampleNum < 16);
 110     return RasterTileDepthOffsets[sampleNum];
 111 }
 112
 113 INLINE static uint32_t RasterTileStencilOffset(uint32_t sampleNum)
 114 {
 115     static const uint32_t RasterTileStencilOffsets[16]
 116     { 0,
 117       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_STENCIL_HOT_TILE_FORMAT>::bpp / 8),
 118       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_STENCIL_HOT_TILE_FORMAT>::bpp / 8) * 2,
 119       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_STENCIL_HOT_TILE_FORMAT>::bpp / 8) * 3,
 120       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_STENCIL_HOT_TILE_FORMAT>::bpp / 8) * 4,
 121       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_STENCIL_HOT_TILE_FORMAT>::bpp / 8) * 5,
 122       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_STENCIL_HOT_TILE_FORMAT>::bpp / 8) * 6,
 123       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_STENCIL_HOT_TILE_FORMAT>::bpp / 8) * 7,
 124       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_STENCIL_HOT_TILE_FORMAT>::bpp / 8) * 8,
 125       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_STENCIL_HOT_TILE_FORMAT>::bpp / 8) * 9,
 126       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_STENCIL_HOT_TILE_FORMAT>::bpp / 8) * 10,
 127       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_STENCIL_HOT_TILE_FORMAT>::bpp / 8) * 11,
 128       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_STENCIL_HOT_TILE_FORMAT>::bpp / 8) * 12,
 129       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_STENCIL_HOT_TILE_FORMAT>::bpp / 8) * 13,
 130       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_STENCIL_HOT_TILE_FORMAT>::bpp / 8) * 14,
 131       (KNOB_TILE_X_DIM * KNOB_TILE_Y_DIM * FormatTraits<KNOB_STENCIL_HOT_TILE_FORMAT>::bpp / 8) * 15,
 132     };
 133     assert(sampleNum < 16);
 134     return RasterTileStencilOffsets[sampleNum];
 135 }
 136
 137 template<typename T, uint32_t InputCoverage>
 138 struct generateInputCoverage
 139 {
 140     INLINE generateInputCoverage(const uint64_t *const coverageMask, uint32_t (&inputMask)[KNOB_SIMD_WIDTH], const uint32_t sampleMask)
 141     {
 142         // will need to update for avx512
 143         assert(KNOB_SIMD_WIDTH == 8);
 144
 145         __m256i mask[2];
 146         __m256i sampleCoverage[2];
 147         if(T::bIsStandardPattern)
 148         {
 149             __m256i src = _mm256_set1_epi32(0);
 150             __m256i index0 = _mm256_set_epi32(7, 6, 5, 4, 3, 2, 1, 0), index1;
 151
 152             if(T::MultisampleT::numSamples == 1)
 153             {
 154                 mask[0] = _mm256_set_epi32(0, 0, 0, 0, 0, 0, 0, -1);
 155             }
 156             else if(T::MultisampleT::numSamples == 2)
 157             {
 158                 mask[0] = _mm256_set_epi32(0, 0, 0, 0, 0, 0, -1, -1);
 159             }
 160             else if(T::MultisampleT::numSamples == 4)
 161             {
 162                 mask[0] = _mm256_set_epi32(0, 0, 0, 0, -1, -1, -1, -1);
 163             }
 164             else if(T::MultisampleT::numSamples == 8)
 165             {
 166                 mask[0] = _mm256_set1_epi32(-1);
 167             }
 168             else if(T::MultisampleT::numSamples == 16)
 169             {
 170                 mask[0] = _mm256_set1_epi32(-1);
 171                 mask[1] = _mm256_set1_epi32(-1);
 172                 index1 = _mm256_set_epi32(15, 14, 13, 12, 11, 10, 9, 8);
 173             }
 174
 175             // gather coverage for samples 0-7
 176             sampleCoverage[0] = _mm256_castps_si256(_simd_mask_i32gather_ps(_mm256_castsi256_ps(src), (const float*)coverageMask, index0, _mm256_castsi256_ps(mask[0]), 8));
 177             if(T::MultisampleT::numSamples > 8)
 178             {
 179                 // gather coverage for samples 8-15
 180                 sampleCoverage[1] = _mm256_castps_si256(_simd_mask_i32gather_ps(_mm256_castsi256_ps(src), (const float*)coverageMask, index1, _mm256_castsi256_ps(mask[1]), 8));
 181             }
 182         }
 183         else
 184         {
 185             // center coverage is the same for all samples; just broadcast to the sample slots
 186             uint32_t centerCoverage = ((uint32_t)(*coverageMask) & MASK);
 187             if(T::MultisampleT::numSamples == 1)
 188             {
 189                 sampleCoverage[0] = _mm256_set_epi32(0, 0, 0, 0, 0, 0, 0, centerCoverage);
 190             }
 191             else if(T::MultisampleT::numSamples == 2)
 192             {
 193                 sampleCoverage[0] = _mm256_set_epi32(0, 0, 0, 0, 0, 0, centerCoverage, centerCoverage);
 194             }
 195             else if(T::MultisampleT::numSamples == 4)
 196             {
 197                 sampleCoverage[0] = _mm256_set_epi32(0, 0, 0, 0, centerCoverage, centerCoverage, centerCoverage, centerCoverage);
 198             }
 199             else if(T::MultisampleT::numSamples == 8)
 200             {
 201                 sampleCoverage[0] = _mm256_set1_epi32(centerCoverage);
 202             }
 203             else if(T::MultisampleT::numSamples == 16)
 204             {
 205                 sampleCoverage[0] = _mm256_set1_epi32(centerCoverage);
 206                 sampleCoverage[1] = _mm256_set1_epi32(centerCoverage);
 207             }
 208         }
 209
 210         mask[0] = _mm256_set_epi8(-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 0xC, 0x8, 0x4, 0x0,
 211                                   -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 0xC, 0x8, 0x4, 0x0);
 212         // pull out the 8bit 4x2 coverage for samples 0-7 into the lower 32 bits of each 128bit lane
 213         __m256i packedCoverage0 = _simd_shuffle_epi8(sampleCoverage[0], mask[0]);
 214
 215         __m256i packedCoverage1;
 216         if(T::MultisampleT::numSamples > 8)
 217         {
 218             // pull out the 8bit 4x2 coverage for samples 8-15 into the lower 32 bits of each 128bit lane
 219             packedCoverage1 = _simd_shuffle_epi8(sampleCoverage[1], mask[0]);
 220         }
 221
 222     #if (KNOB_ARCH == KNOB_ARCH_AVX)
 223         // pack lower 32 bits of each 128 bit lane into lower 64 bits of single 128 bit lane
 224         __m256i hiToLow = _mm256_permute2f128_si256(packedCoverage0, packedCoverage0, 0x83);
 225         __m256 shufRes = _mm256_shuffle_ps(_mm256_castsi256_ps(hiToLow), _mm256_castsi256_ps(hiToLow), _MM_SHUFFLE(1, 1, 0, 1));
 226         packedCoverage0 = _mm256_castps_si256(_mm256_blend_ps(_mm256_castsi256_ps(packedCoverage0), shufRes, 0xFE));
 227
 228         __m256i packedSampleCoverage;
 229         if(T::MultisampleT::numSamples > 8)
 230         {
 231             // pack lower 32 bits of each 128 bit lane into upper 64 bits of single 128 bit lane
 232             hiToLow = _mm256_permute2f128_si256(packedCoverage1, packedCoverage1, 0x83);
 233             shufRes = _mm256_shuffle_ps(_mm256_castsi256_ps(hiToLow), _mm256_castsi256_ps(hiToLow), _MM_SHUFFLE(1, 1, 0, 1));
 234             shufRes = _mm256_blend_ps(_mm256_castsi256_ps(packedCoverage1), shufRes, 0xFE);
 235             packedCoverage1 = _mm256_castps_si256(_mm256_castpd_ps(_mm256_shuffle_pd(_mm256_castps_pd(shufRes), _mm256_castps_pd(shufRes), 0x01)));
 236             packedSampleCoverage = _mm256_castps_si256(_mm256_blend_ps(_mm256_castsi256_ps(packedCoverage0), _mm256_castsi256_ps(packedCoverage1), 0xFC));
 237         }
 238         else
 239         {
 240             packedSampleCoverage = packedCoverage0;
 241         }
 242     #else
 243         __m256i permMask = _mm256_set_epi32(0x7, 0x7, 0x7, 0x7, 0x7, 0x7, 0x4, 0x0);
 244         // pack lower 32 bits of each 128 bit lane into lower 64 bits of single 128 bit lane
 245         packedCoverage0 = _mm256_permutevar8x32_epi32(packedCoverage0, permMask);
 246
 247         __m256i packedSampleCoverage;
 248         if(T::MultisampleT::numSamples > 8)
 249         {
 250             permMask = _mm256_set_epi32(0x7, 0x7, 0x7, 0x7, 0x4, 0x0, 0x7, 0x7);
 251             // pack lower 32 bits of each 128 bit lane into upper 64 bits of single 128 bit lane
 252             packedCoverage1 = _mm256_permutevar8x32_epi32(packedCoverage1, permMask);
 253
 254             // blend coverage masks for samples 0-7 and samples 8-15 into single 128 bit lane
 255             packedSampleCoverage = _mm256_blend_epi32(packedCoverage0, packedCoverage1, 0x0C);
 256         }
 257         else
 258         {
 259             packedSampleCoverage = packedCoverage0;
 260         }
 261     #endif
 262
 263         for(int32_t i = KNOB_SIMD_WIDTH - 1; i >= 0; i--)
 264         {
 265             // convert packed sample coverage masks into single coverage masks for all samples for each pixel in the 4x2
 266             inputMask[i] = _simd_movemask_epi8(packedSampleCoverage);
 267
 268             if(!T::bForcedSampleCount)
 269             {
 270                 // input coverage has to be anded with sample mask if MSAA isn't forced on
 271                 inputMask[i] &= sampleMask;
 272             }
 273
 274             // shift to the next pixel in the 4x2
 275             packedSampleCoverage = _simd_slli_epi32(packedSampleCoverage, 1);
 276         }
 277     }
 278
 279     INLINE generateInputCoverage(const uint64_t *const coverageMask, __m256 &inputCoverage, const uint32_t sampleMask)
 280     {
 281         uint32_t inputMask[KNOB_SIMD_WIDTH];
 282         generateInputCoverage<T, T::InputCoverage>(coverageMask, inputMask, sampleMask);
 283         inputCoverage = _simd_castsi_ps(_mm256_set_epi32(inputMask[7], inputMask[6], inputMask[5], inputMask[4], inputMask[3], inputMask[2], inputMask[1], inputMask[0]));
 284     }
 285
 286 };
 287
 288 template<typename T>
 289 struct generateInputCoverage<T, SWR_INPUT_COVERAGE_INNER_CONSERVATIVE>
 290 {
 291     INLINE generateInputCoverage(const uint64_t *const coverageMask, __m256 &inputCoverage, const uint32_t sampleMask)
 292     {
 293         // will need to update for avx512
 294         assert(KNOB_SIMD_WIDTH == 8);
 295         __m256i vec = _mm256_set1_epi32(coverageMask[0]);
 296         const __m256i bit = _mm256_set_epi32(0x80, 0x40, 0x20, 0x10, 0x08, 0x04, 0x02, 0x01);
 297         vec = _simd_and_si(vec, bit);
 298         vec = _simd_cmplt_epi32(_mm256_setzero_si256(), vec);
 299         vec = _simd_blendv_epi32(_simd_setzero_si(), _simd_set1_epi32(1), vec);
 300         inputCoverage = _simd_castsi_ps(vec);
 301     }
 302
 303     INLINE generateInputCoverage(const uint64_t *const coverageMask, uint32_t (&inputMask)[KNOB_SIMD_WIDTH], const uint32_t sampleMask)
 304     {
 305         unsigned long index;
 306         uint32_t simdCoverage = (coverageMask[0] & MASK);
 307         static const uint32_t FullCoverageMask = (1 << T::MultisampleT::numSamples) - 1;
 308         while(_BitScanForward(&index, simdCoverage))
 309         {
 310             // set all samples to covered
 311             inputMask[index] = FullCoverageMask;
 312         }
 313     }
 314 };
 315
 316 ////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
 317 // Centroid behaves exactly as follows :
 318 // (1) If all samples in the primitive are covered, the attribute is evaluated at the pixel center (even if the sample pattern does not happen to
 319 //     have a sample location there).
 320 // (2) Else the attribute is evaluated at the first covered sample, in increasing order of sample index, where sample coverage is after ANDing the
 321 //     coverage with the SampleMask Rasterizer State.
 322 // (3) If no samples are covered, such as on helper pixels executed off the bounds of a primitive to fill out 2x2 pixel stamps, the attribute is
 323 //     evaluated as follows : If the SampleMask Rasterizer state is a subset of the samples in the pixel, then the first sample covered by the
 324 //     SampleMask Rasterizer State is the evaluation point.Otherwise (full SampleMask), the pixel center is the evaluation point.
 325 ////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
 326 template<typename T>
 327 INLINE void CalcCentroidPos(SWR_PS_CONTEXT &psContext, const uint64_t *const coverageMask, const uint32_t sampleMask,
 328                             const simdscalar vXSamplePosUL, const simdscalar vYSamplePosUL)
 329 {
 330     uint32_t inputMask[KNOB_SIMD_WIDTH];
 331     generateInputCoverage<T, T::InputCoverage>(coverageMask, inputMask, sampleMask);
 332
 333     // Case (2) - partially covered pixel
 334
 335     // scan for first covered sample per pixel in the 4x2 span
 336     unsigned long sampleNum[KNOB_SIMD_WIDTH];
 337     (inputMask[0] > 0) ? (_BitScanForward(&sampleNum[0], inputMask[0])) : (sampleNum[0] = 0);
 338     (inputMask[1] > 0) ? (_BitScanForward(&sampleNum[1], inputMask[1])) : (sampleNum[1] = 0);
 339     (inputMask[2] > 0) ? (_BitScanForward(&sampleNum[2], inputMask[2])) : (sampleNum[2] = 0);
 340     (inputMask[3] > 0) ? (_BitScanForward(&sampleNum[3], inputMask[3])) : (sampleNum[3] = 0);
 341     (inputMask[4] > 0) ? (_BitScanForward(&sampleNum[4], inputMask[4])) : (sampleNum[4] = 0);
 342     (inputMask[5] > 0) ? (_BitScanForward(&sampleNum[5], inputMask[5])) : (sampleNum[5] = 0);
 343     (inputMask[6] > 0) ? (_BitScanForward(&sampleNum[6], inputMask[6])) : (sampleNum[6] = 0);
 344     (inputMask[7] > 0) ? (_BitScanForward(&sampleNum[7], inputMask[7])) : (sampleNum[7] = 0);
 345
 346     // look up and set the sample offsets from UL pixel corner for first covered sample
 347     __m256 vXSample = _mm256_set_ps(T::MultisampleT::X(sampleNum[7]),
 348                                     T::MultisampleT::X(sampleNum[6]),
 349                                     T::MultisampleT::X(sampleNum[5]),
 350                                     T::MultisampleT::X(sampleNum[4]),
 351                                     T::MultisampleT::X(sampleNum[3]),
 352                                     T::MultisampleT::X(sampleNum[2]),
 353                                     T::MultisampleT::X(sampleNum[1]),
 354                                     T::MultisampleT::X(sampleNum[0]));
 355
 356     __m256 vYSample = _mm256_set_ps(T::MultisampleT::Y(sampleNum[7]),
 357                                     T::MultisampleT::Y(sampleNum[6]),
 358                                     T::MultisampleT::Y(sampleNum[5]),
 359                                     T::MultisampleT::Y(sampleNum[4]),
 360                                     T::MultisampleT::Y(sampleNum[3]),
 361                                     T::MultisampleT::Y(sampleNum[2]),
 362                                     T::MultisampleT::Y(sampleNum[1]),
 363                                     T::MultisampleT::Y(sampleNum[0]));
 364     // add sample offset to UL pixel corner
 365     vXSample = _simd_add_ps(vXSamplePosUL, vXSample);
 366     vYSample = _simd_add_ps(vYSamplePosUL, vYSample);
 367
 368     // Case (1) and case (3b) - All samples covered or not covered with full SampleMask
 369     static const __m256i vFullyCoveredMask = T::MultisampleT::FullSampleMask();
 370     __m256i vInputCoveragei =  _mm256_set_epi32(inputMask[7], inputMask[6], inputMask[5], inputMask[4], inputMask[3], inputMask[2], inputMask[1], inputMask[0]);
 371     __m256i vAllSamplesCovered = _simd_cmpeq_epi32(vInputCoveragei, vFullyCoveredMask);
 372
 373     static const __m256i vZero = _simd_setzero_si();
 374     const __m256i vSampleMask = _simd_and_si(_simd_set1_epi32(sampleMask), vFullyCoveredMask);
 375     __m256i vNoSamplesCovered = _simd_cmpeq_epi32(vInputCoveragei, vZero);
 376     __m256i vIsFullSampleMask = _simd_cmpeq_epi32(vSampleMask, vFullyCoveredMask);
 377     __m256i vCase3b = _simd_and_si(vNoSamplesCovered, vIsFullSampleMask);
 378
 379     __m256i vEvalAtCenter = _simd_or_si(vAllSamplesCovered, vCase3b);
 380
 381     // set the centroid position based on results from above
 382     psContext.vX.centroid = _simd_blendv_ps(vXSample, psContext.vX.center, _simd_castsi_ps(vEvalAtCenter));
 383     psContext.vY.centroid = _simd_blendv_ps(vYSample, psContext.vY.center, _simd_castsi_ps(vEvalAtCenter));
 384
 385     // Case (3a) No samples covered and partial sample mask
 386     __m256i vSomeSampleMaskSamples = _simd_cmplt_epi32(vSampleMask, vFullyCoveredMask);
 387     // sample mask should never be all 0's for this case, but handle it anyways
 388     unsigned long firstCoveredSampleMaskSample = 0;
 389     (sampleMask > 0) ? (_BitScanForward(&firstCoveredSampleMaskSample, sampleMask)) : (firstCoveredSampleMaskSample = 0);
 390
 391     __m256i vCase3a = _simd_and_si(vNoSamplesCovered, vSomeSampleMaskSamples);
 392
 393     vXSample = _simd_set1_ps(T::MultisampleT::X(firstCoveredSampleMaskSample));
 394     vYSample = _simd_set1_ps(T::MultisampleT::Y(firstCoveredSampleMaskSample));
 395
 396     // blend in case 3a pixel locations
 397     psContext.vX.centroid = _simd_blendv_ps(psContext.vX.centroid, vXSample, _simd_castsi_ps(vCase3a));
 398     psContext.vY.centroid = _simd_blendv_ps(psContext.vY.centroid, vYSample, _simd_castsi_ps(vCase3a));
 399 }
 400
 401 INLINE void CalcCentroidBarycentrics(const BarycentricCoeffs& coeffs, SWR_PS_CONTEXT &psContext,
 402                                      const simdscalar vXSamplePosUL, const simdscalar vYSamplePosUL)
 403 {
 404     // evaluate I,J
 405     psContext.vI.centroid = vplaneps(coeffs.vIa, coeffs.vIb, coeffs.vIc, psContext.vX.centroid, psContext.vY.centroid);
 406     psContext.vJ.centroid = vplaneps(coeffs.vJa, coeffs.vJb, coeffs.vJc, psContext.vX.centroid, psContext.vY.centroid);
 407     psContext.vI.centroid = _simd_mul_ps(psContext.vI.centroid, coeffs.vRecipDet);
 408     psContext.vJ.centroid = _simd_mul_ps(psContext.vJ.centroid, coeffs.vRecipDet);
 409
 410     // interpolate 1/w
 411     psContext.vOneOverW.centroid = vplaneps(coeffs.vAOneOverW, coeffs.vBOneOverW, coeffs.vCOneOverW, psContext.vI.centroid, psContext.vJ.centroid);
 412 }
 413
 414 template<typename T>
 415 INLINE uint32_t GetNumOMSamples(SWR_MULTISAMPLE_COUNT blendSampleCount)
 416 {
 417     // RT has to be single sample if we're in forcedMSAA mode
 418     if(T::bForcedSampleCount && (T::MultisampleT::sampleCount > SWR_MULTISAMPLE_1X))
 419     {
 420         return 1;
 421     }
 422     // unless we're forced to single sample, in which case we run the OM at the sample count of the RT
 423     else if(T::bForcedSampleCount && (T::MultisampleT::sampleCount == SWR_MULTISAMPLE_1X))
 424     {
 425         return GetNumSamples(blendSampleCount);
 426     }
 427     // else we're in normal MSAA mode and rasterizer and OM are running at the same sample count
 428     else
 429     {
 430         return T::MultisampleT::numSamples;
 431     }
 432 }
 433
 434 template<typename T>
 435 struct PixelRateZTestLoop
 436 {
 437     PixelRateZTestLoop(DRAW_CONTEXT *DC, const SWR_TRIANGLE_DESC &Work, const BarycentricCoeffs& Coeffs, const API_STATE& apiState,
 438                        uint8_t*& depthBase, uint8_t*& stencilBase, const uint8_t ClipDistanceMask) :
 439                        work(Work), coeffs(Coeffs), state(apiState), psState(apiState.psState),
 440                        clipDistanceMask(ClipDistanceMask), pDepthBase(depthBase), pStencilBase(stencilBase) {};
 441
 442     INLINE
 443     uint32_t operator()(simdscalar& activeLanes, SWR_PS_CONTEXT& psContext,
 444                         const CORE_BUCKETS BEDepthBucket, uint32_t currentSimdIn8x8 = 0)
 445     {
 446         uint32_t statCount = 0;
 447         simdscalar anyDepthSamplePassed = _simd_setzero_ps();
 448         for(uint32_t sample = 0; sample < T::MultisampleT::numCoverageSamples; sample++)
 449         {
 450             const uint8_t *pCoverageMask = (uint8_t*)&work.coverageMask[sample];
 451             vCoverageMask[sample] = _simd_and_ps(activeLanes, vMask(pCoverageMask[currentSimdIn8x8] & MASK));
 452
 453             if(!_simd_movemask_ps(vCoverageMask[sample]))
 454             {
 455                 vCoverageMask[sample] = depthPassMask[sample] = stencilPassMask[sample] = _simd_setzero_ps();
 456                 continue;
 457             }
 458
 459             RDTSC_START(BEBarycentric);
 460             // calculate per sample positions
 461             psContext.vX.sample = _simd_add_ps(psContext.vX.UL, T::MultisampleT::vX(sample));
 462             psContext.vY.sample = _simd_add_ps(psContext.vY.UL, T::MultisampleT::vY(sample));
 463
 464             // calc I & J per sample
 465             CalcSampleBarycentrics(coeffs, psContext);
 466
 467             if(psState.writesODepth)
 468             {
 469                 // broadcast and test oDepth(psContext.vZ) written from the PS for each sample
 470                 vZ[sample] = psContext.vZ;
 471             }
 472             else
 473             {
 474                 vZ[sample] = vplaneps(coeffs.vZa, coeffs.vZb, coeffs.vZc, psContext.vI.sample, psContext.vJ.sample);
 475                 vZ[sample] = state.pfnQuantizeDepth(vZ[sample]);
 476             }
 477             RDTSC_STOP(BEBarycentric, 0, 0);
 478
 479             ///@todo: perspective correct vs non-perspective correct clipping?
 480             // if clip distances are enabled, we need to interpolate for each sample
 481             if(clipDistanceMask)
 482             {
 483                 uint8_t clipMask = ComputeUserClipMask(clipDistanceMask, work.pUserClipBuffer,
 484                                                        psContext.vI.sample, psContext.vJ.sample);
 485                 vCoverageMask[sample] = _simd_and_ps(vCoverageMask[sample], vMask(~clipMask));
 486             }
 487
 488             // offset depth/stencil buffers current sample
 489             uint8_t *pDepthSample = pDepthBase + RasterTileDepthOffset(sample);
 490             uint8_t * pStencilSample = pStencilBase + RasterTileStencilOffset(sample);
 491
 492             // ZTest for this sample
 493             RDTSC_START(BEDepthBucket);
 494             depthPassMask[sample] = vCoverageMask[sample];
 495             stencilPassMask[sample] = vCoverageMask[sample];
 496             depthPassMask[sample] = DepthStencilTest(&state, work.triFlags.frontFacing, vZ[sample], pDepthSample,
 497                                                      vCoverageMask[sample], pStencilSample, &stencilPassMask[sample]);
 498             RDTSC_STOP(BEDepthBucket, 0, 0);
 499
 500             // early-exit if no pixels passed depth or earlyZ is forced on
 501             if(psState.forceEarlyZ || !_simd_movemask_ps(depthPassMask[sample]))
 502             {
 503                 DepthStencilWrite(&state.vp[0], &state.depthStencilState, work.triFlags.frontFacing, vZ[sample],
 504                                   pDepthSample, depthPassMask[sample], vCoverageMask[sample], pStencilSample, stencilPassMask[sample]);
 505
 506                 if(!_simd_movemask_ps(depthPassMask[sample]))
 507                 {
 508                     continue;
 509                 }
 510             }
 511             anyDepthSamplePassed = _simd_or_ps(anyDepthSamplePassed, depthPassMask[sample]);
 512             uint32_t statMask = _simd_movemask_ps(depthPassMask[sample]);
 513             statCount += _mm_popcnt_u32(statMask);
 514         }
 515
 516         activeLanes = _simd_and_ps(anyDepthSamplePassed, activeLanes);
 517         // return number of samples that passed depth and coverage
 518         return statCount;
 519     }
 520
 521     // saved depth/stencil/coverage masks and interpolated Z used in OM and DepthWrite
 522     simdscalar vZ[T::MultisampleT::numCoverageSamples];
 523     simdscalar vCoverageMask[T::MultisampleT::numCoverageSamples];
 524     simdscalar depthPassMask[T::MultisampleT::numCoverageSamples];
 525     simdscalar stencilPassMask[T::MultisampleT::numCoverageSamples];
 526
 527 private:
 528     // functor inputs
 529     const SWR_TRIANGLE_DESC& work;
 530     const BarycentricCoeffs& coeffs;
 531     const API_STATE& state;
 532     const SWR_PS_STATE& psState;
 533     const uint8_t clipDistanceMask;
 534     uint8_t*& pDepthBase;
 535     uint8_t*& pStencilBase;
 536 };
 537
 538 INLINE void CalcPixelBarycentrics(const BarycentricCoeffs& coeffs, SWR_PS_CONTEXT &psContext)
 539 {
 540     // evaluate I,J
 541     psContext.vI.center = vplaneps(coeffs.vIa, coeffs.vIb, coeffs.vIc, psContext.vX.center, psContext.vY.center);
 542     psContext.vJ.center = vplaneps(coeffs.vJa, coeffs.vJb, coeffs.vJc, psContext.vX.center, psContext.vY.center);
 543     psContext.vI.center = _simd_mul_ps(psContext.vI.center, coeffs.vRecipDet);
 544     psContext.vJ.center = _simd_mul_ps(psContext.vJ.center, coeffs.vRecipDet);
 545
 546     // interpolate 1/w
 547     psContext.vOneOverW.center = vplaneps(coeffs.vAOneOverW, coeffs.vBOneOverW, coeffs.vCOneOverW, psContext.vI.center, psContext.vJ.center);
 548 }
 549
 550 INLINE void CalcSampleBarycentrics(const BarycentricCoeffs& coeffs, SWR_PS_CONTEXT &psContext)
 551 {
 552     // evaluate I,J
 553     psContext.vI.sample = vplaneps(coeffs.vIa, coeffs.vIb, coeffs.vIc, psContext.vX.sample, psContext.vY.sample);
 554     psContext.vJ.sample = vplaneps(coeffs.vJa, coeffs.vJb, coeffs.vJc, psContext.vX.sample, psContext.vY.sample);
 555     psContext.vI.sample = _simd_mul_ps(psContext.vI.sample, coeffs.vRecipDet);
 556     psContext.vJ.sample = _simd_mul_ps(psContext.vJ.sample, coeffs.vRecipDet);
 557
 558     // interpolate 1/w
 559     psContext.vOneOverW.sample = vplaneps(coeffs.vAOneOverW, coeffs.vBOneOverW, coeffs.vCOneOverW, psContext.vI.sample, psContext.vJ.sample);
 560 }
 561
 562 INLINE void OutputMerger(SWR_PS_CONTEXT &psContext, uint8_t* (&pColorBase)[SWR_NUM_RENDERTARGETS], uint32_t sample, const SWR_BLEND_STATE *pBlendState,
 563                          const PFN_BLEND_JIT_FUNC (&pfnBlendFunc)[SWR_NUM_RENDERTARGETS], simdscalar &coverageMask, simdscalar depthPassMask, const uint32_t NumRT)
 564 {
 565     // type safety guaranteed from template instantiation in BEChooser<>::GetFunc
 566     const uint32_t rasterTileColorOffset = RasterTileColorOffset(sample);
 567     simdvector blendOut;
 568
 569     for(uint32_t rt = 0; rt < NumRT; ++rt)
 570     {
 571         uint8_t *pColorSample = pColorBase[rt] + rasterTileColorOffset;
 572
 573         const SWR_RENDER_TARGET_BLEND_STATE *pRTBlend = &pBlendState->renderTarget[rt];
 574         // pfnBlendFunc may not update all channels.  Initialize with PS output.
 575         /// TODO: move this into the blend JIT.
 576         blendOut = psContext.shaded[rt];
 577
 578         // Blend outputs and update coverage mask for alpha test
 579         if(pfnBlendFunc[rt] != nullptr)
 580         {
 581             pfnBlendFunc[rt](
 582                 pBlendState,
 583                 psContext.shaded[rt],
 584                 psContext.shaded[1],
 585                 sample,
 586                 pColorSample,
 587                 blendOut,
 588                 &psContext.oMask,
 589                 (simdscalari*)&coverageMask);
 590         }
 591
 592         // final write mask
 593         simdscalari outputMask = _simd_castps_si(_simd_and_ps(coverageMask, depthPassMask));
 594
 595         ///@todo can only use maskstore fast path if bpc is 32. Assuming hot tile is RGBA32_FLOAT.
 596         static_assert(KNOB_COLOR_HOT_TILE_FORMAT == R32G32B32A32_FLOAT, "Unsupported hot tile format");
 597
 598         const uint32_t simd = KNOB_SIMD_WIDTH * sizeof(float);
 599
 600         // store with color mask
 601         if(!pRTBlend->writeDisableRed)
 602         {
 603             _simd_maskstore_ps((float*)pColorSample, outputMask, blendOut.x);
 604         }
 605         if(!pRTBlend->writeDisableGreen)
 606         {
 607             _simd_maskstore_ps((float*)(pColorSample + simd), outputMask, blendOut.y);
 608         }
 609         if(!pRTBlend->writeDisableBlue)
 610         {
 611             _simd_maskstore_ps((float*)(pColorSample + simd * 2), outputMask, blendOut.z);
 612         }
 613         if(!pRTBlend->writeDisableAlpha)
 614         {
 615             _simd_maskstore_ps((float*)(pColorSample + simd * 3), outputMask, blendOut.w);
 616         }
 617     }
 618 }
 619
 620 template<uint32_t sampleCountT = SWR_MULTISAMPLE_1X, uint32_t samplePattern = SWR_MSAA_STANDARD_PATTERN,
 621          uint32_t coverage = 0, uint32_t centroid = 0, uint32_t forced = 0, uint32_t canEarlyZ = 0>
 622 struct SwrBackendTraits
 623 {
 624     static const bool bIsStandardPattern = (samplePattern == SWR_MSAA_STANDARD_PATTERN);
 625     static const uint32_t InputCoverage = coverage;
 626     static const bool bCentroidPos = (centroid == 1);
 627     static const bool bForcedSampleCount = (forced == 1);
 628     static const bool bCanEarlyZ = (canEarlyZ == 1);
 629     typedef MultisampleTraits<(SWR_MULTISAMPLE_COUNT)sampleCountT, (bIsStandardPattern) ? SWR_MSAA_STANDARD_PATTERN : SWR_MSAA_CENTER_PATTERN> MultisampleT;
 630 };