src/Shader/PixelRoutine.cpp

   1 // Copyright 2016 The SwiftShader Authors. All Rights Reserved.
   2 //
   3 // Licensed under the Apache License, Version 2.0 (the "License");
   4 // you may not use this file except in compliance with the License.
   5 // You may obtain a copy of the License at
   6 //
   7 //    http://www.apache.org/licenses/LICENSE-2.0
   8 //
   9 // Unless required by applicable law or agreed to in writing, software
  10 // distributed under the License is distributed on an "AS IS" BASIS,
  11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  12 // See the License for the specific language governing permissions and
  13 // limitations under the License.
  14
  15 #include "PixelRoutine.hpp"
  16
  17 #include "Renderer.hpp"
  18 #include "QuadRasterizer.hpp"
  19 #include "Surface.hpp"
  20 #include "Primitive.hpp"
  21 #include "CPUID.hpp"
  22 #include "SamplerCore.hpp"
  23 #include "Constants.hpp"
  24 #include "Debug.hpp"
  25
  26 namespace sw
  27 {
  28         extern bool complementaryDepthBuffer;
  29         extern bool postBlendSRGB;
  30         extern bool exactColorRounding;
  31         extern bool forceClearRegisters;
  32
  33         PixelRoutine::PixelRoutine(const PixelProcessor::State &state, const PixelShader *shader) : QuadRasterizer(state, shader), v(shader && shader->dynamicallyIndexedInput)
  34         {
  35                 if(!shader || shader->getVersion() < 0x0200 || forceClearRegisters)
  36                 {
  37                         for(int i = 0; i < MAX_FRAGMENT_INPUTS; i++)
  38                         {
  39                                 v[i].x = Float4(0.0f);
  40                                 v[i].y = Float4(0.0f);
  41                                 v[i].z = Float4(0.0f);
  42                                 v[i].w = Float4(0.0f);
  43                         }
  44                 }
  45         }
  46
  47         PixelRoutine::~PixelRoutine()
  48         {
  49                 for(int i = 0; i < TEXTURE_IMAGE_UNITS; i++)
  50                 {
  51                         delete sampler[i];
  52                 }
  53         }
  54
  55         void PixelRoutine::quad(Pointer<Byte> cBuffer[RENDERTARGETS], Pointer<Byte> &zBuffer, Pointer<Byte> &sBuffer, Int cMask[4], Int &x, Int &y)
  56         {
  57                 #if PERF_PROFILE
  58                         Long pipeTime = Ticks();
  59                 #endif
  60
  61                 for(int i = 0; i < TEXTURE_IMAGE_UNITS; i++)
  62                 {
  63                         sampler[i] = new SamplerCore(constants, state.sampler[i]);
  64                 }
  65
  66                 const bool earlyDepthTest = !state.depthOverride && !state.alphaTestActive();
  67
  68                 Int zMask[4];   // Depth mask
  69                 Int sMask[4];   // Stencil mask
  70
  71                 for(unsigned int q = 0; q < state.multiSample; q++)
  72                 {
  73                         zMask[q] = cMask[q];
  74                         sMask[q] = cMask[q];
  75                 }
  76
  77                 for(unsigned int q = 0; q < state.multiSample; q++)
  78                 {
  79                         stencilTest(sBuffer, q, x, sMask[q], cMask[q]);
  80                 }
  81
  82                 Float4 f;
  83                 Float4 rhwCentroid;
  84
  85                 Float4 xxxx = Float4(Float(x)) + *Pointer<Float4>(primitive + OFFSET(Primitive,xQuad), 16);
  86
  87                 if(interpolateZ())
  88                 {
  89                         for(unsigned int q = 0; q < state.multiSample; q++)
  90                         {
  91                                 Float4 x = xxxx;
  92
  93                                 if(state.multiSample > 1)
  94                                 {
  95                                         x -= *Pointer<Float4>(constants + OFFSET(Constants,X) + q * sizeof(float4));
  96                                 }
  97
  98                                 z[q] = interpolate(x, Dz[q], z[q], primitive + OFFSET(Primitive,z), false, false);
  99                         }
 100                 }
 101
 102                 Bool depthPass = false;
 103
 104                 if(earlyDepthTest)
 105                 {
 106                         for(unsigned int q = 0; q < state.multiSample; q++)
 107                         {
 108                                 depthPass = depthPass || depthTest(zBuffer, q, x, z[q], sMask[q], zMask[q], cMask[q]);
 109                         }
 110                 }
 111
 112                 If(depthPass || Bool(!earlyDepthTest))
 113                 {
 114                         #if PERF_PROFILE
 115                                 Long interpTime = Ticks();
 116                         #endif
 117
 118                         Float4 yyyy = Float4(Float(y)) + *Pointer<Float4>(primitive + OFFSET(Primitive,yQuad), 16);
 119
 120                         // Centroid locations
 121                         Float4 XXXX = Float4(0.0f);
 122                         Float4 YYYY = Float4(0.0f);
 123
 124                         if(state.centroid)
 125                         {
 126                                 Float4 WWWW(1.0e-9f);
 127
 128                                 for(unsigned int q = 0; q < state.multiSample; q++)
 129                                 {
 130                                         XXXX += *Pointer<Float4>(constants + OFFSET(Constants,sampleX[q]) + 16 * cMask[q]);
 131                                         YYYY += *Pointer<Float4>(constants + OFFSET(Constants,sampleY[q]) + 16 * cMask[q]);
 132                                         WWWW += *Pointer<Float4>(constants + OFFSET(Constants,weight) + 16 * cMask[q]);
 133                                 }
 134
 135                                 WWWW = Rcp_pp(WWWW);
 136                                 XXXX *= WWWW;
 137                                 YYYY *= WWWW;
 138
 139                                 XXXX += xxxx;
 140                                 YYYY += yyyy;
 141                         }
 142
 143                         if(interpolateW())
 144                         {
 145                                 w = interpolate(xxxx, Dw, rhw, primitive + OFFSET(Primitive,w), false, false);
 146                                 rhw = reciprocal(w, false, false, true);
 147
 148                                 if(state.centroid)
 149                                 {
 150                                         rhwCentroid = reciprocal(interpolateCentroid(XXXX, YYYY, rhwCentroid, primitive + OFFSET(Primitive,w), false, false));
 151                                 }
 152                         }
 153
 154                         for(int interpolant = 0; interpolant < MAX_FRAGMENT_INPUTS; interpolant++)
 155                         {
 156                                 for(int component = 0; component < 4; component++)
 157                                 {
 158                                         if(state.interpolant[interpolant].component & (1 << component))
 159                                         {
 160                                                 if(!state.interpolant[interpolant].centroid)
 161                                                 {
 162                                                         v[interpolant][component] = interpolate(xxxx, Dv[interpolant][component], rhw, primitive + OFFSET(Primitive, V[interpolant][component]), (state.interpolant[interpolant].flat & (1 << component)) != 0, state.perspective);
 163                                                 }
 164                                                 else
 165                                                 {
 166                                                         v[interpolant][component] = interpolateCentroid(XXXX, YYYY, rhwCentroid, primitive + OFFSET(Primitive, V[interpolant][component]), (state.interpolant[interpolant].flat & (1 << component)) != 0, state.perspective);
 167                                                 }
 168                                         }
 169                                 }
 170
 171                                 Float4 rcp;
 172
 173                                 switch(state.interpolant[interpolant].project)
 174                                 {
 175                                 case 0:
 176                                         break;
 177                                 case 1:
 178                                         rcp = reciprocal(v[interpolant].y);
 179                                         v[interpolant].x = v[interpolant].x * rcp;
 180                                         break;
 181                                 case 2:
 182                                         rcp = reciprocal(v[interpolant].z);
 183                                         v[interpolant].x = v[interpolant].x * rcp;
 184                                         v[interpolant].y = v[interpolant].y * rcp;
 185                                         break;
 186                                 case 3:
 187                                         rcp = reciprocal(v[interpolant].w);
 188                                         v[interpolant].x = v[interpolant].x * rcp;
 189                                         v[interpolant].y = v[interpolant].y * rcp;
 190                                         v[interpolant].z = v[interpolant].z * rcp;
 191                                         break;
 192                                 }
 193                         }
 194
 195                         if(state.fog.component)
 196                         {
 197                                 f = interpolate(xxxx, Df, rhw, primitive + OFFSET(Primitive,f), state.fog.flat & 0x01, state.perspective);
 198                         }
 199
 200                         setBuiltins(x, y, z, w);
 201
 202                         #if PERF_PROFILE
 203                                 cycles[PERF_INTERP] += Ticks() - interpTime;
 204                         #endif
 205
 206                         Bool alphaPass = true;
 207
 208                         if(colorUsed())
 209                         {
 210                                 #if PERF_PROFILE
 211                                         Long shaderTime = Ticks();
 212                                 #endif
 213
 214                                 applyShader(cMask);
 215
 216                                 #if PERF_PROFILE
 217                                         cycles[PERF_SHADER] += Ticks() - shaderTime;
 218                                 #endif
 219
 220                                 alphaPass = alphaTest(cMask);
 221
 222                                 if((shader && shader->containsKill()) || state.alphaTestActive())
 223                                 {
 224                                         for(unsigned int q = 0; q < state.multiSample; q++)
 225                                         {
 226                                                 zMask[q] &= cMask[q];
 227                                                 sMask[q] &= cMask[q];
 228                                         }
 229                                 }
 230                         }
 231
 232                         If(alphaPass)
 233                         {
 234                                 if(!earlyDepthTest)
 235                                 {
 236                                         for(unsigned int q = 0; q < state.multiSample; q++)
 237                                         {
 238                                                 depthPass = depthPass || depthTest(zBuffer, q, x, z[q], sMask[q], zMask[q], cMask[q]);
 239                                         }
 240                                 }
 241
 242                                 #if PERF_PROFILE
 243                                         Long ropTime = Ticks();
 244                                 #endif
 245
 246                                 If(depthPass || Bool(earlyDepthTest))
 247                                 {
 248                                         for(unsigned int q = 0; q < state.multiSample; q++)
 249                                         {
 250                                                 if(state.multiSampleMask & (1 << q))
 251                                                 {
 252                                                         writeDepth(zBuffer, q, x, z[q], zMask[q]);
 253
 254                                                         if(state.occlusionEnabled)
 255                                                         {
 256                                                                 occlusion += *Pointer<UInt>(constants + OFFSET(Constants,occlusionCount) + 4 * (zMask[q] & sMask[q]));
 257                                                         }
 258                                                 }
 259                                         }
 260
 261                                         if(colorUsed())
 262                                         {
 263                                                 #if PERF_PROFILE
 264                                                         AddAtomic(Pointer<Long>(&profiler.ropOperations), 4);
 265                                                 #endif
 266
 267                                                 rasterOperation(f, cBuffer, x, sMask, zMask, cMask);
 268                                         }
 269                                 }
 270
 271                                 #if PERF_PROFILE
 272                                         cycles[PERF_ROP] += Ticks() - ropTime;
 273                                 #endif
 274                         }
 275                 }
 276
 277                 for(unsigned int q = 0; q < state.multiSample; q++)
 278                 {
 279                         if(state.multiSampleMask & (1 << q))
 280                         {
 281                                 writeStencil(sBuffer, q, x, sMask[q], zMask[q], cMask[q]);
 282                         }
 283                 }
 284
 285                 #if PERF_PROFILE
 286                         cycles[PERF_PIPE] += Ticks() - pipeTime;
 287                 #endif
 288         }
 289
 290         Float4 PixelRoutine::interpolateCentroid(Float4 &x, Float4 &y, Float4 &rhw, Pointer<Byte> planeEquation, bool flat, bool perspective)
 291         {
 292                 Float4 interpolant = *Pointer<Float4>(planeEquation + OFFSET(PlaneEquation,C), 16);
 293
 294                 if(!flat)
 295                 {
 296                         interpolant += x * *Pointer<Float4>(planeEquation + OFFSET(PlaneEquation,A), 16) +
 297                                        y * *Pointer<Float4>(planeEquation + OFFSET(PlaneEquation,B), 16);
 298
 299                         if(perspective)
 300                         {
 301                                 interpolant *= rhw;
 302                         }
 303                 }
 304
 305                 return interpolant;
 306         }
 307
 308         void PixelRoutine::stencilTest(Pointer<Byte> &sBuffer, int q, Int &x, Int &sMask, Int &cMask)
 309         {
 310                 if(!state.stencilActive)
 311                 {
 312                         return;
 313                 }
 314
 315                 // (StencilRef & StencilMask) CompFunc (StencilBufferValue & StencilMask)
 316
 317                 Pointer<Byte> buffer = sBuffer + 2 * x;
 318
 319                 if(q > 0)
 320                 {
 321                         buffer += q * *Pointer<Int>(data + OFFSET(DrawData,stencilSliceB));
 322                 }
 323
 324                 Byte8 value = As<Byte8>(Long1(*Pointer<UInt>(buffer)));
 325                 Byte8 valueCCW = value;
 326
 327                 if(!state.noStencilMask)
 328                 {
 329                         value &= *Pointer<Byte8>(data + OFFSET(DrawData,stencil[0].testMaskQ));
 330                 }
 331
 332                 stencilTest(value, state.stencilCompareMode, false);
 333
 334                 if(state.twoSidedStencil)
 335                 {
 336                         if(!state.noStencilMaskCCW)
 337                         {
 338                                 valueCCW &= *Pointer<Byte8>(data + OFFSET(DrawData,stencil[1].testMaskQ));
 339                         }
 340
 341                         stencilTest(valueCCW, state.stencilCompareModeCCW, true);
 342
 343                         value &= *Pointer<Byte8>(primitive + OFFSET(Primitive,clockwiseMask));
 344                         valueCCW &= *Pointer<Byte8>(primitive + OFFSET(Primitive,invClockwiseMask));
 345                         value |= valueCCW;
 346                 }
 347
 348                 sMask = SignMask(value) & cMask;
 349         }
 350
 351         void PixelRoutine::stencilTest(Byte8 &value, StencilCompareMode stencilCompareMode, bool CCW)
 352         {
 353                 Byte8 equal;
 354
 355                 switch(stencilCompareMode)
 356                 {
 357                 case STENCIL_ALWAYS:
 358                         value = Byte8(0xFFFFFFFFFFFFFFFF);
 359                         break;
 360                 case STENCIL_NEVER:
 361                         value = Byte8(0x0000000000000000);
 362                         break;
 363                 case STENCIL_LESS:                      // a < b ~ b > a
 364                         value += Byte8(0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80);
 365                         value = CmpGT(As<SByte8>(value), *Pointer<SByte8>(data + OFFSET(DrawData,stencil[CCW].referenceMaskedSignedQ)));
 366                         break;
 367                 case STENCIL_EQUAL:
 368                         value = CmpEQ(value, *Pointer<Byte8>(data + OFFSET(DrawData,stencil[CCW].referenceMaskedQ)));
 369                         break;
 370                 case STENCIL_NOTEQUAL:          // a != b ~ !(a == b)
 371                         value = CmpEQ(value, *Pointer<Byte8>(data + OFFSET(DrawData,stencil[CCW].referenceMaskedQ)));
 372                         value ^= Byte8(0xFFFFFFFFFFFFFFFF);
 373                         break;
 374                 case STENCIL_LESSEQUAL: // a <= b ~ (b > a) || (a == b)
 375                         equal = value;
 376                         equal = CmpEQ(equal, *Pointer<Byte8>(data + OFFSET(DrawData,stencil[CCW].referenceMaskedQ)));
 377                         value += Byte8(0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80);
 378                         value = CmpGT(As<SByte8>(value), *Pointer<SByte8>(data + OFFSET(DrawData,stencil[CCW].referenceMaskedSignedQ)));
 379                         value |= equal;
 380                         break;
 381                 case STENCIL_GREATER:           // a > b
 382                         equal = *Pointer<Byte8>(data + OFFSET(DrawData,stencil[CCW].referenceMaskedSignedQ));
 383                         value += Byte8(0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80);
 384                         equal = CmpGT(As<SByte8>(equal), As<SByte8>(value));
 385                         value = equal;
 386                         break;
 387                 case STENCIL_GREATEREQUAL:      // a >= b ~ !(a < b) ~ !(b > a)
 388                         value += Byte8(0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80);
 389                         value = CmpGT(As<SByte8>(value), *Pointer<SByte8>(data + OFFSET(DrawData,stencil[CCW].referenceMaskedSignedQ)));
 390                         value ^= Byte8(0xFFFFFFFFFFFFFFFF);
 391                         break;
 392                 default:
 393                         ASSERT(false);
 394                 }
 395         }
 396
 397         Bool PixelRoutine::depthTest(Pointer<Byte> &zBuffer, int q, Int &x, Float4 &z, Int &sMask, Int &zMask, Int &cMask)
 398         {
 399                 if(!state.depthTestActive)
 400                 {
 401                         return true;
 402                 }
 403
 404                 Float4 Z = z;
 405
 406                 if(shader && shader->depthOverride())
 407                 {
 408                         if(complementaryDepthBuffer)
 409                         {
 410                                 Z = Float4(1.0f) - oDepth;
 411                         }
 412                         else
 413                         {
 414                                 Z = oDepth;
 415                         }
 416                 }
 417
 418                 Pointer<Byte> buffer;
 419                 Int pitch;
 420
 421                 if(!state.quadLayoutDepthBuffer)
 422                 {
 423                         buffer = zBuffer + 4 * x;
 424                         pitch = *Pointer<Int>(data + OFFSET(DrawData,depthPitchB));
 425                 }
 426                 else
 427                 {
 428                         buffer = zBuffer + 8 * x;
 429                 }
 430
 431                 if(q > 0)
 432                 {
 433                         buffer += q * *Pointer<Int>(data + OFFSET(DrawData,depthSliceB));
 434                 }
 435
 436                 Float4 zValue;
 437
 438                 if(state.depthCompareMode != DEPTH_NEVER || (state.depthCompareMode != DEPTH_ALWAYS && !state.depthWriteEnable))
 439                 {
 440                         if(!state.quadLayoutDepthBuffer)
 441                         {
 442                                 // FIXME: Properly optimizes?
 443                                 zValue.xy = *Pointer<Float4>(buffer);
 444                                 zValue.zw = *Pointer<Float4>(buffer + pitch - 8);
 445                         }
 446                         else
 447                         {
 448                                 zValue = *Pointer<Float4>(buffer, 16);
 449                         }
 450                 }
 451
 452                 Int4 zTest;
 453
 454                 switch(state.depthCompareMode)
 455                 {
 456                 case DEPTH_ALWAYS:
 457                         // Optimized
 458                         break;
 459                 case DEPTH_NEVER:
 460                         // Optimized
 461                         break;
 462                 case DEPTH_EQUAL:
 463                         zTest = CmpEQ(zValue, Z);
 464                         break;
 465                 case DEPTH_NOTEQUAL:
 466                         zTest = CmpNEQ(zValue, Z);
 467                         break;
 468                 case DEPTH_LESS:
 469                         if(complementaryDepthBuffer)
 470                         {
 471                                 zTest = CmpLT(zValue, Z);
 472                         }
 473                         else
 474                         {
 475                                 zTest = CmpNLE(zValue, Z);
 476                         }
 477                         break;
 478                 case DEPTH_GREATEREQUAL:
 479                         if(complementaryDepthBuffer)
 480                         {
 481                                 zTest = CmpNLT(zValue, Z);
 482                         }
 483                         else
 484                         {
 485                                 zTest = CmpLE(zValue, Z);
 486                         }
 487                         break;
 488                 case DEPTH_LESSEQUAL:
 489                         if(complementaryDepthBuffer)
 490                         {
 491                                 zTest = CmpLE(zValue, Z);
 492                         }
 493                         else
 494                         {
 495                                 zTest = CmpNLT(zValue, Z);
 496                         }
 497                         break;
 498                 case DEPTH_GREATER:
 499                         if(complementaryDepthBuffer)
 500                         {
 501                                 zTest = CmpNLE(zValue, Z);
 502                         }
 503                         else
 504                         {
 505                                 zTest = CmpLT(zValue, Z);
 506                         }
 507                         break;
 508                 default:
 509                         ASSERT(false);
 510                 }
 511
 512                 switch(state.depthCompareMode)
 513                 {
 514                 case DEPTH_ALWAYS:
 515                         zMask = cMask;
 516                         break;
 517                 case DEPTH_NEVER:
 518                         zMask = 0x0;
 519                         break;
 520                 default:
 521                         zMask = SignMask(zTest) & cMask;
 522                         break;
 523                 }
 524
 525                 if(state.stencilActive)
 526                 {
 527                         zMask &= sMask;
 528                 }
 529
 530                 return zMask != 0;
 531         }
 532
 533         void PixelRoutine::alphaTest(Int &aMask, Short4 &alpha)
 534         {
 535                 Short4 cmp;
 536                 Short4 equal;
 537
 538                 switch(state.alphaCompareMode)
 539                 {
 540                 case ALPHA_ALWAYS:
 541                         aMask = 0xF;
 542                         break;
 543                 case ALPHA_NEVER:
 544                         aMask = 0x0;
 545                         break;
 546                 case ALPHA_EQUAL:
 547                         cmp = CmpEQ(alpha, *Pointer<Short4>(data + OFFSET(DrawData,factor.alphaReference4)));
 548                         aMask = SignMask(Pack(cmp, Short4(0x0000)));
 549                         break;
 550                 case ALPHA_NOTEQUAL:       // a != b ~ !(a == b)
 551                         cmp = CmpEQ(alpha, *Pointer<Short4>(data + OFFSET(DrawData,factor.alphaReference4))) ^ Short4(0xFFFFu);   // FIXME
 552                         aMask = SignMask(Pack(cmp, Short4(0x0000)));
 553                         break;
 554                 case ALPHA_LESS:           // a < b ~ b > a
 555                         cmp = CmpGT(*Pointer<Short4>(data + OFFSET(DrawData,factor.alphaReference4)), alpha);
 556                         aMask = SignMask(Pack(cmp, Short4(0x0000)));
 557                         break;
 558                 case ALPHA_GREATEREQUAL:   // a >= b ~ (a > b) || (a == b) ~ !(b > a)   // TODO: Approximate
 559                         equal = CmpEQ(alpha, *Pointer<Short4>(data + OFFSET(DrawData,factor.alphaReference4)));
 560                         cmp = CmpGT(alpha, *Pointer<Short4>(data + OFFSET(DrawData,factor.alphaReference4)));
 561                         cmp |= equal;
 562                         aMask = SignMask(Pack(cmp, Short4(0x0000)));
 563                         break;
 564                 case ALPHA_LESSEQUAL:      // a <= b ~ !(a > b)
 565                         cmp = CmpGT(alpha, *Pointer<Short4>(data + OFFSET(DrawData,factor.alphaReference4))) ^ Short4(0xFFFFu);   // FIXME
 566                         aMask = SignMask(Pack(cmp, Short4(0x0000)));
 567                         break;
 568                 case ALPHA_GREATER:        // a > b
 569                         cmp = CmpGT(alpha, *Pointer<Short4>(data + OFFSET(DrawData,factor.alphaReference4)));
 570                         aMask = SignMask(Pack(cmp, Short4(0x0000)));
 571                         break;
 572                 default:
 573                         ASSERT(false);
 574                 }
 575         }
 576
 577         void PixelRoutine::alphaToCoverage(Int cMask[4], Float4 &alpha)
 578         {
 579                 Int4 coverage0 = CmpNLT(alpha, *Pointer<Float4>(data + OFFSET(DrawData,a2c0)));
 580                 Int4 coverage1 = CmpNLT(alpha, *Pointer<Float4>(data + OFFSET(DrawData,a2c1)));
 581                 Int4 coverage2 = CmpNLT(alpha, *Pointer<Float4>(data + OFFSET(DrawData,a2c2)));
 582                 Int4 coverage3 = CmpNLT(alpha, *Pointer<Float4>(data + OFFSET(DrawData,a2c3)));
 583
 584                 Int aMask0 = SignMask(coverage0);
 585                 Int aMask1 = SignMask(coverage1);
 586                 Int aMask2 = SignMask(coverage2);
 587                 Int aMask3 = SignMask(coverage3);
 588
 589                 cMask[0] &= aMask0;
 590                 cMask[1] &= aMask1;
 591                 cMask[2] &= aMask2;
 592                 cMask[3] &= aMask3;
 593         }
 594
 595         void PixelRoutine::fogBlend(Vector4f &c0, Float4 &fog)
 596         {
 597                 if(!state.fogActive)
 598                 {
 599                         return;
 600                 }
 601
 602                 if(state.pixelFogMode != FOG_NONE)
 603                 {
 604                         pixelFog(fog);
 605
 606                         fog = Min(fog, Float4(1.0f));
 607                         fog = Max(fog, Float4(0.0f));
 608                 }
 609
 610                 c0.x -= *Pointer<Float4>(data + OFFSET(DrawData,fog.colorF[0]));
 611                 c0.y -= *Pointer<Float4>(data + OFFSET(DrawData,fog.colorF[1]));
 612                 c0.z -= *Pointer<Float4>(data + OFFSET(DrawData,fog.colorF[2]));
 613
 614                 c0.x *= fog;
 615                 c0.y *= fog;
 616                 c0.z *= fog;
 617
 618                 c0.x += *Pointer<Float4>(data + OFFSET(DrawData,fog.colorF[0]));
 619                 c0.y += *Pointer<Float4>(data + OFFSET(DrawData,fog.colorF[1]));
 620                 c0.z += *Pointer<Float4>(data + OFFSET(DrawData,fog.colorF[2]));
 621         }
 622
 623         void PixelRoutine::pixelFog(Float4 &visibility)
 624         {
 625                 Float4 &zw = visibility;
 626
 627                 if(state.pixelFogMode != FOG_NONE)
 628                 {
 629                         if(state.wBasedFog)
 630                         {
 631                                 zw = rhw;
 632                         }
 633                         else
 634                         {
 635                                 if(complementaryDepthBuffer)
 636                                 {
 637                                         zw = Float4(1.0f) - z[0];
 638                                 }
 639                                 else
 640                                 {
 641                                         zw = z[0];
 642                                 }
 643                         }
 644                 }
 645
 646                 switch(state.pixelFogMode)
 647                 {
 648                 case FOG_NONE:
 649                         break;
 650                 case FOG_LINEAR:
 651                         zw *= *Pointer<Float4>(data + OFFSET(DrawData,fog.scale));
 652                         zw += *Pointer<Float4>(data + OFFSET(DrawData,fog.offset));
 653                         break;
 654                 case FOG_EXP:
 655                         zw *= *Pointer<Float4>(data + OFFSET(DrawData,fog.densityE));
 656                         zw = exponential2(zw, true);
 657                         break;
 658                 case FOG_EXP2:
 659                         zw *= zw;
 660                         zw *= *Pointer<Float4>(data + OFFSET(DrawData,fog.density2E));
 661                         zw = exponential2(zw, true);
 662                         break;
 663                 default:
 664                         ASSERT(false);
 665                 }
 666         }
 667
 668         void PixelRoutine::writeDepth(Pointer<Byte> &zBuffer, int q, Int &x, Float4 &z, Int &zMask)
 669         {
 670                 if(!state.depthWriteEnable)
 671                 {
 672                         return;
 673                 }
 674
 675                 Float4 Z = z;
 676
 677                 if(shader && shader->depthOverride())
 678                 {
 679                         if(complementaryDepthBuffer)
 680                         {
 681                                 Z = Float4(1.0f) - oDepth;
 682                         }
 683                         else
 684                         {
 685                                 Z = oDepth;
 686                         }
 687                 }
 688
 689                 Pointer<Byte> buffer;
 690                 Int pitch;
 691
 692                 if(!state.quadLayoutDepthBuffer)
 693                 {
 694                         buffer = zBuffer + 4 * x;
 695                         pitch = *Pointer<Int>(data + OFFSET(DrawData,depthPitchB));
 696                 }
 697                 else
 698                 {
 699                         buffer = zBuffer + 8 * x;
 700                 }
 701
 702                 if(q > 0)
 703                 {
 704                         buffer += q * *Pointer<Int>(data + OFFSET(DrawData,depthSliceB));
 705                 }
 706
 707                 Float4 zValue;
 708
 709                 if(state.depthCompareMode != DEPTH_NEVER || (state.depthCompareMode != DEPTH_ALWAYS && !state.depthWriteEnable))
 710                 {
 711                         if(!state.quadLayoutDepthBuffer)
 712                         {
 713                                 // FIXME: Properly optimizes?
 714                                 zValue.xy = *Pointer<Float4>(buffer);
 715                                 zValue.zw = *Pointer<Float4>(buffer + pitch - 8);
 716                         }
 717                         else
 718                         {
 719                                 zValue = *Pointer<Float4>(buffer, 16);
 720                         }
 721                 }
 722
 723                 Z = As<Float4>(As<Int4>(Z) & *Pointer<Int4>(constants + OFFSET(Constants,maskD4X) + zMask * 16, 16));
 724                 zValue = As<Float4>(As<Int4>(zValue) & *Pointer<Int4>(constants + OFFSET(Constants,invMaskD4X) + zMask * 16, 16));
 725                 Z = As<Float4>(As<Int4>(Z) | As<Int4>(zValue));
 726
 727                 if(!state.quadLayoutDepthBuffer)
 728                 {
 729                         // FIXME: Properly optimizes?
 730                         *Pointer<Float2>(buffer) = Float2(Z.xy);
 731                         *Pointer<Float2>(buffer + pitch) = Float2(Z.zw);
 732                 }
 733                 else
 734                 {
 735                         *Pointer<Float4>(buffer, 16) = Z;
 736                 }
 737         }
 738
 739         void PixelRoutine::writeStencil(Pointer<Byte> &sBuffer, int q, Int &x, Int &sMask, Int &zMask, Int &cMask)
 740         {
 741                 if(!state.stencilActive)
 742                 {
 743                         return;
 744                 }
 745
 746                 if(state.stencilPassOperation == OPERATION_KEEP && state.stencilZFailOperation == OPERATION_KEEP && state.stencilFailOperation == OPERATION_KEEP)
 747                 {
 748                         if(!state.twoSidedStencil || (state.stencilPassOperationCCW == OPERATION_KEEP && state.stencilZFailOperationCCW == OPERATION_KEEP && state.stencilFailOperationCCW == OPERATION_KEEP))
 749                         {
 750                                 return;
 751                         }
 752                 }
 753
 754                 if(state.stencilWriteMasked && (!state.twoSidedStencil || state.stencilWriteMaskedCCW))
 755                 {
 756                         return;
 757                 }
 758
 759                 Pointer<Byte> buffer = sBuffer + 2 * x;
 760
 761                 if(q > 0)
 762                 {
 763                         buffer += q * *Pointer<Int>(data + OFFSET(DrawData,stencilSliceB));
 764                 }
 765
 766                 Byte8 bufferValue = As<Byte8>(Long1(*Pointer<UInt>(buffer)));
 767
 768                 Byte8 newValue;
 769                 stencilOperation(newValue, bufferValue, state.stencilPassOperation, state.stencilZFailOperation, state.stencilFailOperation, false, zMask, sMask);
 770
 771                 if(!state.noStencilWriteMask)
 772                 {
 773                         Byte8 maskedValue = bufferValue;
 774                         newValue &= *Pointer<Byte8>(data + OFFSET(DrawData,stencil[0].writeMaskQ));
 775                         maskedValue &= *Pointer<Byte8>(data + OFFSET(DrawData,stencil[0].invWriteMaskQ));
 776                         newValue |= maskedValue;
 777                 }
 778
 779                 if(state.twoSidedStencil)
 780                 {
 781                         Byte8 newValueCCW;
 782
 783                         stencilOperation(newValueCCW, bufferValue, state.stencilPassOperationCCW, state.stencilZFailOperationCCW, state.stencilFailOperationCCW, true, zMask, sMask);
 784
 785                         if(!state.noStencilWriteMaskCCW)
 786                         {
 787                                 Byte8 maskedValue = bufferValue;
 788                                 newValueCCW &= *Pointer<Byte8>(data + OFFSET(DrawData,stencil[1].writeMaskQ));
 789                                 maskedValue &= *Pointer<Byte8>(data + OFFSET(DrawData,stencil[1].invWriteMaskQ));
 790                                 newValueCCW |= maskedValue;
 791                         }
 792
 793                         newValue &= *Pointer<Byte8>(primitive + OFFSET(Primitive,clockwiseMask));
 794                         newValueCCW &= *Pointer<Byte8>(primitive + OFFSET(Primitive,invClockwiseMask));
 795                         newValue |= newValueCCW;
 796                 }
 797
 798                 newValue &= *Pointer<Byte8>(constants + OFFSET(Constants,maskB4Q) + 8 * cMask);
 799                 bufferValue &= *Pointer<Byte8>(constants + OFFSET(Constants,invMaskB4Q) + 8 * cMask);
 800                 newValue |= bufferValue;
 801
 802                 *Pointer<UInt>(buffer) = UInt(As<Long>(newValue));
 803         }
 804
 805         void PixelRoutine::stencilOperation(Byte8 &newValue, Byte8 &bufferValue, StencilOperation stencilPassOperation, StencilOperation stencilZFailOperation, StencilOperation stencilFailOperation, bool CCW, Int &zMask, Int &sMask)
 806         {
 807                 Byte8 &pass = newValue;
 808                 Byte8 fail;
 809                 Byte8 zFail;
 810
 811                 stencilOperation(pass, bufferValue, stencilPassOperation, CCW);
 812
 813                 if(stencilZFailOperation != stencilPassOperation)
 814                 {
 815                         stencilOperation(zFail, bufferValue, stencilZFailOperation, CCW);
 816                 }
 817
 818                 if(stencilFailOperation != stencilPassOperation || stencilFailOperation != stencilZFailOperation)
 819                 {
 820                         stencilOperation(fail, bufferValue, stencilFailOperation, CCW);
 821                 }
 822
 823                 if(stencilFailOperation != stencilPassOperation || stencilFailOperation != stencilZFailOperation)
 824                 {
 825                         if(state.depthTestActive && stencilZFailOperation != stencilPassOperation)   // zMask valid and values not the same
 826                         {
 827                                 pass &= *Pointer<Byte8>(constants + OFFSET(Constants,maskB4Q) + 8 * zMask);
 828                                 zFail &= *Pointer<Byte8>(constants + OFFSET(Constants,invMaskB4Q) + 8 * zMask);
 829                                 pass |= zFail;
 830                         }
 831
 832                         pass &= *Pointer<Byte8>(constants + OFFSET(Constants,maskB4Q) + 8 * sMask);
 833                         fail &= *Pointer<Byte8>(constants + OFFSET(Constants,invMaskB4Q) + 8 * sMask);
 834                         pass |= fail;
 835                 }
 836         }
 837
 838         void PixelRoutine::stencilOperation(Byte8 &output, Byte8 &bufferValue, StencilOperation operation, bool CCW)
 839         {
 840                 switch(operation)
 841                 {
 842                 case OPERATION_KEEP:
 843                         output = bufferValue;
 844                         break;
 845                 case OPERATION_ZERO:
 846                         output = Byte8(0x0000000000000000);
 847                         break;
 848                 case OPERATION_REPLACE:
 849                         output = *Pointer<Byte8>(data + OFFSET(DrawData,stencil[CCW].referenceQ));
 850                         break;
 851                 case OPERATION_INCRSAT:
 852                         output = AddSat(bufferValue, Byte8(1, 1, 1, 1, 1, 1, 1, 1));
 853                         break;
 854                 case OPERATION_DECRSAT:
 855                         output = SubSat(bufferValue, Byte8(1, 1, 1, 1, 1, 1, 1, 1));
 856                         break;
 857                 case OPERATION_INVERT:
 858                         output = bufferValue ^ Byte8(0xFFFFFFFFFFFFFFFF);
 859                         break;
 860                 case OPERATION_INCR:
 861                         output = bufferValue + Byte8(1, 1, 1, 1, 1, 1, 1, 1);
 862                         break;
 863                 case OPERATION_DECR:
 864                         output = bufferValue - Byte8(1, 1, 1, 1, 1, 1, 1, 1);
 865                         break;
 866                 default:
 867                         ASSERT(false);
 868                 }
 869         }
 870
 871         void PixelRoutine::blendFactor(const Vector4s &blendFactor, const Vector4s &current, const Vector4s &pixel, BlendFactor blendFactorActive)
 872         {
 873                 switch(blendFactorActive)
 874                 {
 875                 case BLEND_ZERO:
 876                         // Optimized
 877                         break;
 878                 case BLEND_ONE:
 879                         // Optimized
 880                         break;
 881                 case BLEND_SOURCE:
 882                         blendFactor.x = current.x;
 883                         blendFactor.y = current.y;
 884                         blendFactor.z = current.z;
 885                         break;
 886                 case BLEND_INVSOURCE:
 887                         blendFactor.x = Short4(0xFFFFu) - current.x;
 888                         blendFactor.y = Short4(0xFFFFu) - current.y;
 889                         blendFactor.z = Short4(0xFFFFu) - current.z;
 890                         break;
 891                 case BLEND_DEST:
 892                         blendFactor.x = pixel.x;
 893                         blendFactor.y = pixel.y;
 894                         blendFactor.z = pixel.z;
 895                         break;
 896                 case BLEND_INVDEST:
 897                         blendFactor.x = Short4(0xFFFFu) - pixel.x;
 898                         blendFactor.y = Short4(0xFFFFu) - pixel.y;
 899                         blendFactor.z = Short4(0xFFFFu) - pixel.z;
 900                         break;
 901                 case BLEND_SOURCEALPHA:
 902                         blendFactor.x = current.w;
 903                         blendFactor.y = current.w;
 904                         blendFactor.z = current.w;
 905                         break;
 906                 case BLEND_INVSOURCEALPHA:
 907                         blendFactor.x = Short4(0xFFFFu) - current.w;
 908                         blendFactor.y = Short4(0xFFFFu) - current.w;
 909                         blendFactor.z = Short4(0xFFFFu) - current.w;
 910                         break;
 911                 case BLEND_DESTALPHA:
 912                         blendFactor.x = pixel.w;
 913                         blendFactor.y = pixel.w;
 914                         blendFactor.z = pixel.w;
 915                         break;
 916                 case BLEND_INVDESTALPHA:
 917                         blendFactor.x = Short4(0xFFFFu) - pixel.w;
 918                         blendFactor.y = Short4(0xFFFFu) - pixel.w;
 919                         blendFactor.z = Short4(0xFFFFu) - pixel.w;
 920                         break;
 921                 case BLEND_SRCALPHASAT:
 922                         blendFactor.x = Short4(0xFFFFu) - pixel.w;
 923                         blendFactor.x = Min(As<UShort4>(blendFactor.x), As<UShort4>(current.w));
 924                         blendFactor.y = blendFactor.x;
 925                         blendFactor.z = blendFactor.x;
 926                         break;
 927                 case BLEND_CONSTANT:
 928                         blendFactor.x = *Pointer<Short4>(data + OFFSET(DrawData,factor.blendConstant4W[0]));
 929                         blendFactor.y = *Pointer<Short4>(data + OFFSET(DrawData,factor.blendConstant4W[1]));
 930                         blendFactor.z = *Pointer<Short4>(data + OFFSET(DrawData,factor.blendConstant4W[2]));
 931                         break;
 932                 case BLEND_INVCONSTANT:
 933                         blendFactor.x = *Pointer<Short4>(data + OFFSET(DrawData,factor.invBlendConstant4W[0]));
 934                         blendFactor.y = *Pointer<Short4>(data + OFFSET(DrawData,factor.invBlendConstant4W[1]));
 935                         blendFactor.z = *Pointer<Short4>(data + OFFSET(DrawData,factor.invBlendConstant4W[2]));
 936                         break;
 937                 case BLEND_CONSTANTALPHA:
 938                         blendFactor.x = *Pointer<Short4>(data + OFFSET(DrawData,factor.blendConstant4W[3]));
 939                         blendFactor.y = *Pointer<Short4>(data + OFFSET(DrawData,factor.blendConstant4W[3]));
 940                         blendFactor.z = *Pointer<Short4>(data + OFFSET(DrawData,factor.blendConstant4W[3]));
 941                         break;
 942                 case BLEND_INVCONSTANTALPHA:
 943                         blendFactor.x = *Pointer<Short4>(data + OFFSET(DrawData,factor.invBlendConstant4W[3]));
 944                         blendFactor.y = *Pointer<Short4>(data + OFFSET(DrawData,factor.invBlendConstant4W[3]));
 945                         blendFactor.z = *Pointer<Short4>(data + OFFSET(DrawData,factor.invBlendConstant4W[3]));
 946                         break;
 947                 default:
 948                         ASSERT(false);
 949                 }
 950         }
 951
 952         void PixelRoutine::blendFactorAlpha(const Vector4s &blendFactor, const Vector4s &current, const Vector4s &pixel, BlendFactor blendFactorAlphaActive)
 953         {
 954                 switch(blendFactorAlphaActive)
 955                 {
 956                 case BLEND_ZERO:
 957                         // Optimized
 958                         break;
 959                 case BLEND_ONE:
 960                         // Optimized
 961                         break;
 962                 case BLEND_SOURCE:
 963                         blendFactor.w = current.w;
 964                         break;
 965                 case BLEND_INVSOURCE:
 966                         blendFactor.w = Short4(0xFFFFu) - current.w;
 967                         break;
 968                 case BLEND_DEST:
 969                         blendFactor.w = pixel.w;
 970                         break;
 971                 case BLEND_INVDEST:
 972                         blendFactor.w = Short4(0xFFFFu) - pixel.w;
 973                         break;
 974                 case BLEND_SOURCEALPHA:
 975                         blendFactor.w = current.w;
 976                         break;
 977                 case BLEND_INVSOURCEALPHA:
 978                         blendFactor.w = Short4(0xFFFFu) - current.w;
 979                         break;
 980                 case BLEND_DESTALPHA:
 981                         blendFactor.w = pixel.w;
 982                         break;
 983                 case BLEND_INVDESTALPHA:
 984                         blendFactor.w = Short4(0xFFFFu) - pixel.w;
 985                         break;
 986                 case BLEND_SRCALPHASAT:
 987                         blendFactor.w = Short4(0xFFFFu);
 988                         break;
 989                 case BLEND_CONSTANT:
 990                 case BLEND_CONSTANTALPHA:
 991                         blendFactor.w = *Pointer<Short4>(data + OFFSET(DrawData,factor.blendConstant4W[3]));
 992                         break;
 993                 case BLEND_INVCONSTANT:
 994                 case BLEND_INVCONSTANTALPHA:
 995                         blendFactor.w = *Pointer<Short4>(data + OFFSET(DrawData,factor.invBlendConstant4W[3]));
 996                         break;
 997                 default:
 998                         ASSERT(false);
 999                 }
1000         }
1001
1002         bool PixelRoutine::isSRGB(int index) const
1003         {
1004                 return state.targetFormat[index] == FORMAT_SRGB8_A8 || state.targetFormat[index] == FORMAT_SRGB8_X8;
1005         }
1006
1007         void PixelRoutine::readPixel(int index, Pointer<Byte> &cBuffer, Int &x, Vector4s &pixel)
1008         {
1009                 Short4 c01;
1010                 Short4 c23;
1011                 Pointer<Byte> buffer;
1012                 Pointer<Byte> buffer2;
1013
1014                 switch(state.targetFormat[index])
1015                 {
1016                 case FORMAT_R5G6B5:
1017                         buffer = cBuffer + 2 * x;
1018                         buffer2 = buffer + *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
1019                         c01 = As<Short4>(Int2(*Pointer<Int>(buffer), *Pointer<Int>(buffer2)));
1020
1021                         pixel.x = c01 & Short4(0xF800u);
1022                         pixel.y = (c01 & Short4(0x07E0u)) << 5;
1023                         pixel.z = (c01 & Short4(0x001Fu)) << 11;
1024                         pixel.w = Short4(0xFFFFu);
1025                         break;
1026                 case FORMAT_A8R8G8B8:
1027                         buffer = cBuffer + 4 * x;
1028                         c01 = *Pointer<Short4>(buffer);
1029                         buffer += *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
1030                         c23 = *Pointer<Short4>(buffer);
1031                         pixel.z = c01;
1032                         pixel.y = c01;
1033                         pixel.z = UnpackLow(As<Byte8>(pixel.z), As<Byte8>(c23));
1034                         pixel.y = UnpackHigh(As<Byte8>(pixel.y), As<Byte8>(c23));
1035                         pixel.x = pixel.z;
1036                         pixel.z = UnpackLow(As<Byte8>(pixel.z), As<Byte8>(pixel.y));
1037                         pixel.x = UnpackHigh(As<Byte8>(pixel.x), As<Byte8>(pixel.y));
1038                         pixel.y = pixel.z;
1039                         pixel.w = pixel.x;
1040                         pixel.x = UnpackLow(As<Byte8>(pixel.x), As<Byte8>(pixel.x));
1041                         pixel.y = UnpackHigh(As<Byte8>(pixel.y), As<Byte8>(pixel.y));
1042                         pixel.z = UnpackLow(As<Byte8>(pixel.z), As<Byte8>(pixel.z));
1043                         pixel.w = UnpackHigh(As<Byte8>(pixel.w), As<Byte8>(pixel.w));
1044                         break;
1045                 case FORMAT_A8B8G8R8:
1046                 case FORMAT_SRGB8_A8:
1047                         buffer = cBuffer + 4 * x;
1048                         c01 = *Pointer<Short4>(buffer);
1049                         buffer += *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
1050                         c23 = *Pointer<Short4>(buffer);
1051                         pixel.z = c01;
1052                         pixel.y = c01;
1053                         pixel.z = UnpackLow(As<Byte8>(pixel.z), As<Byte8>(c23));
1054                         pixel.y = UnpackHigh(As<Byte8>(pixel.y), As<Byte8>(c23));
1055                         pixel.x = pixel.z;
1056                         pixel.z = UnpackLow(As<Byte8>(pixel.z), As<Byte8>(pixel.y));
1057                         pixel.x = UnpackHigh(As<Byte8>(pixel.x), As<Byte8>(pixel.y));
1058                         pixel.y = pixel.z;
1059                         pixel.w = pixel.x;
1060                         pixel.x = UnpackLow(As<Byte8>(pixel.z), As<Byte8>(pixel.z));
1061                         pixel.y = UnpackHigh(As<Byte8>(pixel.y), As<Byte8>(pixel.y));
1062                         pixel.z = UnpackLow(As<Byte8>(pixel.w), As<Byte8>(pixel.w));
1063                         pixel.w = UnpackHigh(As<Byte8>(pixel.w), As<Byte8>(pixel.w));
1064                         break;
1065                 case FORMAT_A8:
1066                         buffer = cBuffer + 1 * x;
1067                         pixel.w = Insert(pixel.w, *Pointer<Short>(buffer), 0);
1068                         buffer += *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
1069                         pixel.w = Insert(pixel.w, *Pointer<Short>(buffer), 1);
1070                         pixel.w = UnpackLow(As<Byte8>(pixel.w), As<Byte8>(pixel.w));
1071                         pixel.x = Short4(0x0000);
1072                         pixel.y = Short4(0x0000);
1073                         pixel.z = Short4(0x0000);
1074                         break;
1075                 case FORMAT_X8R8G8B8:
1076                         buffer = cBuffer + 4 * x;
1077                         c01 = *Pointer<Short4>(buffer);
1078                         buffer += *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
1079                         c23 = *Pointer<Short4>(buffer);
1080                         pixel.z = c01;
1081                         pixel.y = c01;
1082                         pixel.z = UnpackLow(As<Byte8>(pixel.z), As<Byte8>(c23));
1083                         pixel.y = UnpackHigh(As<Byte8>(pixel.y), As<Byte8>(c23));
1084                         pixel.x = pixel.z;
1085                         pixel.z = UnpackLow(As<Byte8>(pixel.z), As<Byte8>(pixel.y));
1086                         pixel.x = UnpackHigh(As<Byte8>(pixel.x), As<Byte8>(pixel.y));
1087                         pixel.y = pixel.z;
1088                         pixel.x = UnpackLow(As<Byte8>(pixel.x), As<Byte8>(pixel.x));
1089                         pixel.y = UnpackHigh(As<Byte8>(pixel.y), As<Byte8>(pixel.y));
1090                         pixel.z = UnpackLow(As<Byte8>(pixel.z), As<Byte8>(pixel.z));
1091                         pixel.w = Short4(0xFFFFu);
1092                         break;
1093                 case FORMAT_X8B8G8R8:
1094                 case FORMAT_SRGB8_X8:
1095                         buffer = cBuffer + 4 * x;
1096                         c01 = *Pointer<Short4>(buffer);
1097                         buffer += *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
1098                         c23 = *Pointer<Short4>(buffer);
1099                         pixel.z = c01;
1100                         pixel.y = c01;
1101                         pixel.z = UnpackLow(As<Byte8>(pixel.z), As<Byte8>(c23));
1102                         pixel.y = UnpackHigh(As<Byte8>(pixel.y), As<Byte8>(c23));
1103                         pixel.x = pixel.z;
1104                         pixel.z = UnpackLow(As<Byte8>(pixel.z), As<Byte8>(pixel.y));
1105                         pixel.x = UnpackHigh(As<Byte8>(pixel.x), As<Byte8>(pixel.y));
1106                         pixel.y = pixel.z;
1107                         pixel.w = pixel.x;
1108                         pixel.x = UnpackLow(As<Byte8>(pixel.z), As<Byte8>(pixel.z));
1109                         pixel.y = UnpackHigh(As<Byte8>(pixel.y), As<Byte8>(pixel.y));
1110                         pixel.z = UnpackLow(As<Byte8>(pixel.w), As<Byte8>(pixel.w));
1111                         pixel.w = Short4(0xFFFFu);
1112                         break;
1113                 case FORMAT_A8G8R8B8Q:
1114                         UNIMPLEMENTED();
1115                 //      pixel.z = UnpackLow(As<Byte8>(pixel.z), *Pointer<Byte8>(cBuffer + 8 * x + 0));
1116                 //      pixel.x = UnpackHigh(As<Byte8>(pixel.x), *Pointer<Byte8>(cBuffer + 8 * x + 0));
1117                 //      pixel.y = UnpackLow(As<Byte8>(pixel.y), *Pointer<Byte8>(cBuffer + 8 * x + 8));
1118                 //      pixel.w = UnpackHigh(As<Byte8>(pixel.w), *Pointer<Byte8>(cBuffer + 8 * x + 8));
1119                         break;
1120                 case FORMAT_X8G8R8B8Q:
1121                         UNIMPLEMENTED();
1122                 //      pixel.z = UnpackLow(As<Byte8>(pixel.z), *Pointer<Byte8>(cBuffer + 8 * x + 0));
1123                 //      pixel.x = UnpackHigh(As<Byte8>(pixel.x), *Pointer<Byte8>(cBuffer + 8 * x + 0));
1124                 //      pixel.y = UnpackLow(As<Byte8>(pixel.y), *Pointer<Byte8>(cBuffer + 8 * x + 8));
1125                 //      pixel.w = Short4(0xFFFFu);
1126                         break;
1127                 case FORMAT_A16B16G16R16:
1128                         buffer = cBuffer;
1129                         pixel.x = *Pointer<Short4>(buffer + 8 * x);
1130                         pixel.y = *Pointer<Short4>(buffer + 8 * x + 8);
1131                         buffer += *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
1132                         pixel.z = *Pointer<Short4>(buffer + 8 * x);
1133                         pixel.w = *Pointer<Short4>(buffer + 8 * x + 8);
1134                         transpose4x4(pixel.x, pixel.y, pixel.z, pixel.w);
1135                         break;
1136                 case FORMAT_G16R16:
1137                         buffer = cBuffer;
1138                         pixel.x = *Pointer<Short4>(buffer + 4 * x);
1139                         buffer += *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
1140                         pixel.y = *Pointer<Short4>(buffer + 4 * x);
1141                         pixel.z = pixel.x;
1142                         pixel.x = As<Short4>(UnpackLow(pixel.x, pixel.y));
1143                         pixel.z = As<Short4>(UnpackHigh(pixel.z, pixel.y));
1144                         pixel.y = pixel.z;
1145                         pixel.x = As<Short4>(UnpackLow(pixel.x, pixel.z));
1146                         pixel.y = As<Short4>(UnpackHigh(pixel.y, pixel.z));
1147                         pixel.z = Short4(0xFFFFu);
1148                         pixel.w = Short4(0xFFFFu);
1149                         break;
1150                 default:
1151                         ASSERT(false);
1152                 }
1153
1154                 if((postBlendSRGB && state.writeSRGB) || isSRGB(index))
1155                 {
1156                         sRGBtoLinear16_12_16(pixel);
1157                 }
1158         }
1159
1160         void PixelRoutine::alphaBlend(int index, Pointer<Byte> &cBuffer, Vector4s &current, Int &x)
1161         {
1162                 if(!state.alphaBlendActive)
1163                 {
1164                         return;
1165                 }
1166
1167                 Vector4s pixel;
1168                 readPixel(index, cBuffer, x, pixel);
1169
1170                 // Final Color = ObjectColor * SourceBlendFactor + PixelColor * DestinationBlendFactor
1171                 Vector4s sourceFactor;
1172                 Vector4s destFactor;
1173
1174                 blendFactor(sourceFactor, current, pixel, state.sourceBlendFactor);
1175                 blendFactor(destFactor, current, pixel, state.destBlendFactor);
1176
1177                 if(state.sourceBlendFactor != BLEND_ONE && state.sourceBlendFactor != BLEND_ZERO)
1178                 {
1179                         current.x = MulHigh(As<UShort4>(current.x), As<UShort4>(sourceFactor.x));
1180                         current.y = MulHigh(As<UShort4>(current.y), As<UShort4>(sourceFactor.y));
1181                         current.z = MulHigh(As<UShort4>(current.z), As<UShort4>(sourceFactor.z));
1182                 }
1183
1184                 if(state.destBlendFactor != BLEND_ONE && state.destBlendFactor != BLEND_ZERO)
1185                 {
1186                         pixel.x = MulHigh(As<UShort4>(pixel.x), As<UShort4>(destFactor.x));
1187                         pixel.y = MulHigh(As<UShort4>(pixel.y), As<UShort4>(destFactor.y));
1188                         pixel.z = MulHigh(As<UShort4>(pixel.z), As<UShort4>(destFactor.z));
1189                 }
1190
1191                 switch(state.blendOperation)
1192                 {
1193                 case BLENDOP_ADD:
1194                         current.x = AddSat(As<UShort4>(current.x), As<UShort4>(pixel.x));
1195                         current.y = AddSat(As<UShort4>(current.y), As<UShort4>(pixel.y));
1196                         current.z = AddSat(As<UShort4>(current.z), As<UShort4>(pixel.z));
1197                         break;
1198                 case BLENDOP_SUB:
1199                         current.x = SubSat(As<UShort4>(current.x), As<UShort4>(pixel.x));
1200                         current.y = SubSat(As<UShort4>(current.y), As<UShort4>(pixel.y));
1201                         current.z = SubSat(As<UShort4>(current.z), As<UShort4>(pixel.z));
1202                         break;
1203                 case BLENDOP_INVSUB:
1204                         current.x = SubSat(As<UShort4>(pixel.x), As<UShort4>(current.x));
1205                         current.y = SubSat(As<UShort4>(pixel.y), As<UShort4>(current.y));
1206                         current.z = SubSat(As<UShort4>(pixel.z), As<UShort4>(current.z));
1207                         break;
1208                 case BLENDOP_MIN:
1209                         current.x = Min(As<UShort4>(current.x), As<UShort4>(pixel.x));
1210                         current.y = Min(As<UShort4>(current.y), As<UShort4>(pixel.y));
1211                         current.z = Min(As<UShort4>(current.z), As<UShort4>(pixel.z));
1212                         break;
1213                 case BLENDOP_MAX:
1214                         current.x = Max(As<UShort4>(current.x), As<UShort4>(pixel.x));
1215                         current.y = Max(As<UShort4>(current.y), As<UShort4>(pixel.y));
1216                         current.z = Max(As<UShort4>(current.z), As<UShort4>(pixel.z));
1217                         break;
1218                 case BLENDOP_SOURCE:
1219                         // No operation
1220                         break;
1221                 case BLENDOP_DEST:
1222                         current.x = pixel.x;
1223                         current.y = pixel.y;
1224                         current.z = pixel.z;
1225                         break;
1226                 case BLENDOP_NULL:
1227                         current.x = Short4(0x0000);
1228                         current.y = Short4(0x0000);
1229                         current.z = Short4(0x0000);
1230                         break;
1231                 default:
1232                         ASSERT(false);
1233                 }
1234
1235                 blendFactorAlpha(sourceFactor, current, pixel, state.sourceBlendFactorAlpha);
1236                 blendFactorAlpha(destFactor, current, pixel, state.destBlendFactorAlpha);
1237
1238                 if(state.sourceBlendFactorAlpha != BLEND_ONE && state.sourceBlendFactorAlpha != BLEND_ZERO)
1239                 {
1240                         current.w = MulHigh(As<UShort4>(current.w), As<UShort4>(sourceFactor.w));
1241                 }
1242
1243                 if(state.destBlendFactorAlpha != BLEND_ONE && state.destBlendFactorAlpha != BLEND_ZERO)
1244                 {
1245                         pixel.w = MulHigh(As<UShort4>(pixel.w), As<UShort4>(destFactor.w));
1246                 }
1247
1248                 switch(state.blendOperationAlpha)
1249                 {
1250                 case BLENDOP_ADD:
1251                         current.w = AddSat(As<UShort4>(current.w), As<UShort4>(pixel.w));
1252                         break;
1253                 case BLENDOP_SUB:
1254                         current.w = SubSat(As<UShort4>(current.w), As<UShort4>(pixel.w));
1255                         break;
1256                 case BLENDOP_INVSUB:
1257                         current.w = SubSat(As<UShort4>(pixel.w), As<UShort4>(current.w));
1258                         break;
1259                 case BLENDOP_MIN:
1260                         current.w = Min(As<UShort4>(current.w), As<UShort4>(pixel.w));
1261                         break;
1262                 case BLENDOP_MAX:
1263                         current.w = Max(As<UShort4>(current.w), As<UShort4>(pixel.w));
1264                         break;
1265                 case BLENDOP_SOURCE:
1266                         // No operation
1267                         break;
1268                 case BLENDOP_DEST:
1269                         current.w = pixel.w;
1270                         break;
1271                 case BLENDOP_NULL:
1272                         current.w = Short4(0x0000);
1273                         break;
1274                 default:
1275                         ASSERT(false);
1276                 }
1277         }
1278
1279         void PixelRoutine::logicOperation(int index, Pointer<Byte> &cBuffer, Vector4s &current, Int &x)
1280         {
1281                 if(state.logicalOperation == LOGICALOP_COPY)
1282                 {
1283                         return;
1284                 }
1285
1286                 Vector4s pixel;
1287                 readPixel(index, cBuffer, x, pixel);
1288
1289                 switch(state.logicalOperation)
1290                 {
1291                 case LOGICALOP_CLEAR:
1292                         current.x = UShort4(0);
1293                         current.y = UShort4(0);
1294                         current.z = UShort4(0);
1295                         break;
1296                 case LOGICALOP_SET:
1297                         current.x = UShort4(0xFFFFu);
1298                         current.y = UShort4(0xFFFFu);
1299                         current.z = UShort4(0xFFFFu);
1300                         break;
1301                 case LOGICALOP_COPY:
1302                         ASSERT(false);   // Optimized out
1303                         break;
1304                 case LOGICALOP_COPY_INVERTED:
1305                         current.x = ~current.x;
1306                         current.y = ~current.y;
1307                         current.z = ~current.z;
1308                         break;
1309                 case LOGICALOP_NOOP:
1310                         current.x = pixel.x;
1311                         current.y = pixel.y;
1312                         current.z = pixel.z;
1313                         break;
1314                 case LOGICALOP_INVERT:
1315                         current.x = ~pixel.x;
1316                         current.y = ~pixel.y;
1317                         current.z = ~pixel.z;
1318                         break;
1319                 case LOGICALOP_AND:
1320                         current.x = pixel.x & current.x;
1321                         current.y = pixel.y & current.y;
1322                         current.z = pixel.z & current.z;
1323                         break;
1324                 case LOGICALOP_NAND:
1325                         current.x = ~(pixel.x & current.x);
1326                         current.y = ~(pixel.y & current.y);
1327                         current.z = ~(pixel.z & current.z);
1328                         break;
1329                 case LOGICALOP_OR:
1330                         current.x = pixel.x | current.x;
1331                         current.y = pixel.y | current.y;
1332                         current.z = pixel.z | current.z;
1333                         break;
1334                 case LOGICALOP_NOR:
1335                         current.x = ~(pixel.x | current.x);
1336                         current.y = ~(pixel.y | current.y);
1337                         current.z = ~(pixel.z | current.z);
1338                         break;
1339                 case LOGICALOP_XOR:
1340                         current.x = pixel.x ^ current.x;
1341                         current.y = pixel.y ^ current.y;
1342                         current.z = pixel.z ^ current.z;
1343                         break;
1344                 case LOGICALOP_EQUIV:
1345                         current.x = ~(pixel.x ^ current.x);
1346                         current.y = ~(pixel.y ^ current.y);
1347                         current.z = ~(pixel.z ^ current.z);
1348                         break;
1349                 case LOGICALOP_AND_REVERSE:
1350                         current.x = ~pixel.x & current.x;
1351                         current.y = ~pixel.y & current.y;
1352                         current.z = ~pixel.z & current.z;
1353                         break;
1354                 case LOGICALOP_AND_INVERTED:
1355                         current.x = pixel.x & ~current.x;
1356                         current.y = pixel.y & ~current.y;
1357                         current.z = pixel.z & ~current.z;
1358                         break;
1359                 case LOGICALOP_OR_REVERSE:
1360                         current.x = ~pixel.x | current.x;
1361                         current.y = ~pixel.y | current.y;
1362                         current.z = ~pixel.z | current.z;
1363                         break;
1364                 case LOGICALOP_OR_INVERTED:
1365                         current.x = pixel.x | ~current.x;
1366                         current.y = pixel.y | ~current.y;
1367                         current.z = pixel.z | ~current.z;
1368                         break;
1369                 default:
1370                         ASSERT(false);
1371                 }
1372         }
1373
1374         void PixelRoutine::writeColor(int index, Pointer<Byte> &cBuffer, Int &x, Vector4s &current, Int &sMask, Int &zMask, Int &cMask)
1375         {
1376                 if((postBlendSRGB && state.writeSRGB) || isSRGB(index))
1377                 {
1378                         linearToSRGB16_12_16(current);
1379                 }
1380
1381                 if(exactColorRounding)
1382                 {
1383                         switch(state.targetFormat[index])
1384                         {
1385                         case FORMAT_R5G6B5:
1386                                 current.x = AddSat(As<UShort4>(current.x), UShort4(0x0400));
1387                                 current.y = AddSat(As<UShort4>(current.y), UShort4(0x0200));
1388                                 current.z = AddSat(As<UShort4>(current.z), UShort4(0x0400));
1389                                 break;
1390                         case FORMAT_X8G8R8B8Q:
1391                         case FORMAT_A8G8R8B8Q:
1392                         case FORMAT_X8R8G8B8:
1393                         case FORMAT_X8B8G8R8:
1394                         case FORMAT_A8R8G8B8:
1395                         case FORMAT_A8B8G8R8:
1396                         case FORMAT_SRGB8_X8:
1397                         case FORMAT_SRGB8_A8:
1398                         case FORMAT_R8:
1399                                 current.x = current.x - As<Short4>(As<UShort4>(current.x) >> 8) + Short4(0x0080);
1400                                 current.y = current.y - As<Short4>(As<UShort4>(current.y) >> 8) + Short4(0x0080);
1401                                 current.z = current.z - As<Short4>(As<UShort4>(current.z) >> 8) + Short4(0x0080);
1402                                 current.w = current.w - As<Short4>(As<UShort4>(current.w) >> 8) + Short4(0x0080);
1403                                 break;
1404                         default:
1405                                 break;
1406                         }
1407                 }
1408
1409                 int rgbaWriteMask = state.colorWriteActive(index);
1410                 int bgraWriteMask = (rgbaWriteMask & 0x0000000A) | (rgbaWriteMask & 0x00000001) << 2 | (rgbaWriteMask & 0x00000004) >> 2;
1411
1412                 switch(state.targetFormat[index])
1413                 {
1414                 case FORMAT_R5G6B5:
1415                         {
1416                                 current.x = current.x & Short4(0xF800u);
1417                                 current.y = As<UShort4>(current.y & Short4(0xFC00u)) >> 5;
1418                                 current.z = As<UShort4>(current.z) >> 11;
1419
1420                                 current.x = current.x | current.y | current.z;
1421                         }
1422                         break;
1423                 case FORMAT_X8G8R8B8Q:
1424                         UNIMPLEMENTED();
1425                 //      current.x = As<Short4>(As<UShort4>(current.x) >> 8);
1426                 //      current.y = As<Short4>(As<UShort4>(current.y) >> 8);
1427                 //      current.z = As<Short4>(As<UShort4>(current.z) >> 8);
1428
1429                 //      current.z = As<Short4>(Pack(As<UShort4>(current.z), As<UShort4>(current.x)));
1430                 //      current.y = As<Short4>(Pack(As<UShort4>(current.y), As<UShort4>(current.y)));
1431                         break;
1432                 case FORMAT_A8G8R8B8Q:
1433                         UNIMPLEMENTED();
1434                 //      current.x = As<Short4>(As<UShort4>(current.x) >> 8);
1435                 //      current.y = As<Short4>(As<UShort4>(current.y) >> 8);
1436                 //      current.z = As<Short4>(As<UShort4>(current.z) >> 8);
1437                 //      current.w = As<Short4>(As<UShort4>(current.w) >> 8);
1438
1439                 //      current.z = As<Short4>(Pack(As<UShort4>(current.z), As<UShort4>(current.x)));
1440                 //      current.y = As<Short4>(Pack(As<UShort4>(current.y), As<UShort4>(current.w)));
1441                         break;
1442                 case FORMAT_X8R8G8B8:
1443                 case FORMAT_A8R8G8B8:
1444                         if(state.targetFormat[index] == FORMAT_X8R8G8B8 || rgbaWriteMask == 0x7)
1445                         {
1446                                 current.x = As<Short4>(As<UShort4>(current.x) >> 8);
1447                                 current.y = As<Short4>(As<UShort4>(current.y) >> 8);
1448                                 current.z = As<Short4>(As<UShort4>(current.z) >> 8);
1449
1450                                 current.z = As<Short4>(Pack(As<UShort4>(current.z), As<UShort4>(current.x)));
1451                                 current.y = As<Short4>(Pack(As<UShort4>(current.y), As<UShort4>(current.y)));
1452
1453                                 current.x = current.z;
1454                                 current.z = UnpackLow(As<Byte8>(current.z), As<Byte8>(current.y));
1455                                 current.x = UnpackHigh(As<Byte8>(current.x), As<Byte8>(current.y));
1456                                 current.y = current.z;
1457                                 current.z = As<Short4>(UnpackLow(current.z, current.x));
1458                                 current.y = As<Short4>(UnpackHigh(current.y, current.x));
1459                         }
1460                         else
1461                         {
1462                                 current.x = As<Short4>(As<UShort4>(current.x) >> 8);
1463                                 current.y = As<Short4>(As<UShort4>(current.y) >> 8);
1464                                 current.z = As<Short4>(As<UShort4>(current.z) >> 8);
1465                                 current.w = As<Short4>(As<UShort4>(current.w) >> 8);
1466
1467                                 current.z = As<Short4>(Pack(As<UShort4>(current.z), As<UShort4>(current.x)));
1468                                 current.y = As<Short4>(Pack(As<UShort4>(current.y), As<UShort4>(current.w)));
1469
1470                                 current.x = current.z;
1471                                 current.z = UnpackLow(As<Byte8>(current.z), As<Byte8>(current.y));
1472                                 current.x = UnpackHigh(As<Byte8>(current.x), As<Byte8>(current.y));
1473                                 current.y = current.z;
1474                                 current.z = As<Short4>(UnpackLow(current.z, current.x));
1475                                 current.y = As<Short4>(UnpackHigh(current.y, current.x));
1476                         }
1477                         break;
1478                 case FORMAT_X8B8G8R8:
1479                 case FORMAT_A8B8G8R8:
1480                 case FORMAT_SRGB8_X8:
1481                 case FORMAT_SRGB8_A8:
1482                         if(state.targetFormat[index] == FORMAT_X8B8G8R8 || state.targetFormat[index] == FORMAT_SRGB8_X8 || rgbaWriteMask == 0x7)
1483                         {
1484                                 current.x = As<Short4>(As<UShort4>(current.x) >> 8);
1485                                 current.y = As<Short4>(As<UShort4>(current.y) >> 8);
1486                                 current.z = As<Short4>(As<UShort4>(current.z) >> 8);
1487
1488                                 current.z = As<Short4>(Pack(As<UShort4>(current.x), As<UShort4>(current.z)));
1489                                 current.y = As<Short4>(Pack(As<UShort4>(current.y), As<UShort4>(current.y)));
1490
1491                                 current.x = current.z;
1492                                 current.z = UnpackLow(As<Byte8>(current.z), As<Byte8>(current.y));
1493                                 current.x = UnpackHigh(As<Byte8>(current.x), As<Byte8>(current.y));
1494                                 current.y = current.z;
1495                                 current.z = As<Short4>(UnpackLow(current.z, current.x));
1496                                 current.y = As<Short4>(UnpackHigh(current.y, current.x));
1497                         }
1498                         else
1499                         {
1500                                 current.x = As<Short4>(As<UShort4>(current.x) >> 8);
1501                                 current.y = As<Short4>(As<UShort4>(current.y) >> 8);
1502                                 current.z = As<Short4>(As<UShort4>(current.z) >> 8);
1503                                 current.w = As<Short4>(As<UShort4>(current.w) >> 8);
1504
1505                                 current.z = As<Short4>(Pack(As<UShort4>(current.x), As<UShort4>(current.z)));
1506                                 current.y = As<Short4>(Pack(As<UShort4>(current.y), As<UShort4>(current.w)));
1507
1508                                 current.x = current.z;
1509                                 current.z = UnpackLow(As<Byte8>(current.z), As<Byte8>(current.y));
1510                                 current.x = UnpackHigh(As<Byte8>(current.x), As<Byte8>(current.y));
1511                                 current.y = current.z;
1512                                 current.z = As<Short4>(UnpackLow(current.z, current.x));
1513                                 current.y = As<Short4>(UnpackHigh(current.y, current.x));
1514                         }
1515                         break;
1516                 case FORMAT_R8:
1517                         current.x = As<Short4>(As<UShort4>(current.x) >> 8);
1518                         current.x = As<Short4>(Pack(As<UShort4>(current.x), As<UShort4>(current.x)));
1519                         break;
1520                 case FORMAT_A8:
1521                         current.w = As<Short4>(As<UShort4>(current.w) >> 8);
1522                         current.w = As<Short4>(Pack(As<UShort4>(current.w), As<UShort4>(current.w)));
1523                         break;
1524                 case FORMAT_G16R16:
1525                         current.z = current.x;
1526                         current.x = As<Short4>(UnpackLow(current.x, current.y));
1527                         current.z = As<Short4>(UnpackHigh(current.z, current.y));
1528                         current.y = current.z;
1529                         break;
1530                 case FORMAT_A16B16G16R16:
1531                         transpose4x4(current.x, current.y, current.z, current.w);
1532                         break;
1533                 default:
1534                         ASSERT(false);
1535                 }
1536
1537                 Short4 c01 = current.z;
1538                 Short4 c23 = current.y;
1539
1540                 Int xMask;   // Combination of all masks
1541
1542                 if(state.depthTestActive)
1543                 {
1544                         xMask = zMask;
1545                 }
1546                 else
1547                 {
1548                         xMask = cMask;
1549                 }
1550
1551                 if(state.stencilActive)
1552                 {
1553                         xMask &= sMask;
1554                 }
1555
1556                 switch(state.targetFormat[index])
1557                 {
1558                 case FORMAT_R5G6B5:
1559                         {
1560                                 Pointer<Byte> buffer = cBuffer + 2 * x;
1561                                 Int value = *Pointer<Int>(buffer);
1562
1563                                 Int c01 = Extract(As<Int2>(current.x), 0);
1564
1565                                 if((bgraWriteMask & 0x00000007) != 0x00000007)
1566                                 {
1567                                         Int masked = value;
1568                                         c01 &= *Pointer<Int>(constants + OFFSET(Constants,mask565Q[bgraWriteMask & 0x7][0]));
1569                                         masked &= *Pointer<Int>(constants + OFFSET(Constants,invMask565Q[bgraWriteMask & 0x7][0]));
1570                                         c01 |= masked;
1571                                 }
1572
1573                                 c01 &= *Pointer<Int>(constants + OFFSET(Constants,maskW4Q[0][0]) + xMask * 8);
1574                                 value &= *Pointer<Int>(constants + OFFSET(Constants,invMaskW4Q[0][0]) + xMask * 8);
1575                                 c01 |= value;
1576                                 *Pointer<Int>(buffer) = c01;
1577
1578                                 buffer += *Pointer<Int>(data + OFFSET(DrawData,colorPitchB[index]));
1579                                 value = *Pointer<Int>(buffer);
1580
1581                                 Int c23 = Extract(As<Int2>(current.x), 1);
1582
1583                                 if((bgraWriteMask & 0x00000007) != 0x00000007)
1584                                 {
1585                                         Int masked = value;
1586                                         c23 &= *Pointer<Int>(constants + OFFSET(Constants,mask565Q[bgraWriteMask & 0x7][0]));
1587                                         masked &= *Pointer<Int>(constants + OFFSET(Constants,invMask565Q[bgraWriteMask & 0x7][0]));
1588                                         c23 |= masked;
1589                                 }
1590
1591                                 c23 &= *Pointer<Int>(constants + OFFSET(Constants,maskW4Q[0][2]) + xMask * 8);
1592                                 value &= *Pointer<Int>(constants + OFFSET(Constants,invMaskW4Q[0][2]) + xMask * 8);
1593                                 c23 |= value;
1594                                 *Pointer<Int>(buffer) = c23;
1595                         }
1596                         break;
1597                 case FORMAT_A8G8R8B8Q:
1598                 case FORMAT_X8G8R8B8Q:   // FIXME: Don't touch alpha?
1599                         UNIMPLEMENTED();
1600                 //      value = *Pointer<Short4>(cBuffer + 8 * x + 0);
1601
1602                 //      if((state.targetFormat[index] == FORMAT_A8G8R8B8Q && bgraWriteMask != 0x0000000F) ||
1603                 //         ((state.targetFormat[index] == FORMAT_X8G8R8B8Q && bgraWriteMask != 0x00000007) &&
1604                 //          (state.targetFormat[index] == FORMAT_X8G8R8B8Q && bgraWriteMask != 0x0000000F)))   // FIXME: Need for masking when XRGB && Fh?
1605                 //      {
1606                 //              Short4 masked = value;
1607                 //              c01 &= *Pointer<Short4>(constants + OFFSET(Constants,maskB4Q[bgraWriteMask][0]));
1608                 //              masked &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskB4Q[bgraWriteMask][0]));
1609                 //              c01 |= masked;
1610                 //      }
1611
1612                 //      c01 &= *Pointer<Short4>(constants + OFFSET(Constants,maskD01Q) + xMask * 8);
1613                 //      value &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskD01Q) + xMask * 8);
1614                 //      c01 |= value;
1615                 //      *Pointer<Short4>(cBuffer + 8 * x + 0) = c01;
1616
1617                 //      value = *Pointer<Short4>(cBuffer + 8 * x + 8);
1618
1619                 //      if((state.targetFormat[index] == FORMAT_A8G8R8B8Q && bgraWriteMask != 0x0000000F) ||
1620                 //         ((state.targetFormat[index] == FORMAT_X8G8R8B8Q && bgraWriteMask != 0x00000007) &&
1621                 //          (state.targetFormat[index] == FORMAT_X8G8R8B8Q && bgraWriteMask != 0x0000000F)))   // FIXME: Need for masking when XRGB && Fh?
1622                 //      {
1623                 //              Short4 masked = value;
1624                 //              c23 &= *Pointer<Short4>(constants + OFFSET(Constants,maskB4Q[bgraWriteMask][0]));
1625                 //              masked &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskB4Q[bgraWriteMask][0]));
1626                 //              c23 |= masked;
1627                 //      }
1628
1629                 //      c23 &= *Pointer<Short4>(constants + OFFSET(Constants,maskD23Q) + xMask * 8);
1630                 //      value &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskD23Q) + xMask * 8);
1631                 //      c23 |= value;
1632                 //      *Pointer<Short4>(cBuffer + 8 * x + 8) = c23;
1633                         break;
1634                 case FORMAT_A8R8G8B8:
1635                 case FORMAT_X8R8G8B8:   // FIXME: Don't touch alpha?
1636                         {
1637                                 Pointer<Byte> buffer = cBuffer + x * 4;
1638                                 Short4 value = *Pointer<Short4>(buffer);
1639
1640                                 if((state.targetFormat[index] == FORMAT_A8R8G8B8 && bgraWriteMask != 0x0000000F) ||
1641                                    ((state.targetFormat[index] == FORMAT_X8R8G8B8 && bgraWriteMask != 0x00000007) &&
1642                                         (state.targetFormat[index] == FORMAT_X8R8G8B8 && bgraWriteMask != 0x0000000F)))   // FIXME: Need for masking when XRGB && Fh?
1643                                 {
1644                                         Short4 masked = value;
1645                                         c01 &= *Pointer<Short4>(constants + OFFSET(Constants,maskB4Q[bgraWriteMask][0]));
1646                                         masked &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskB4Q[bgraWriteMask][0]));
1647                                         c01 |= masked;
1648                                 }
1649
1650                                 c01 &= *Pointer<Short4>(constants + OFFSET(Constants,maskD01Q) + xMask * 8);
1651                                 value &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskD01Q) + xMask * 8);
1652                                 c01 |= value;
1653                                 *Pointer<Short4>(buffer) = c01;
1654
1655                                 buffer += *Pointer<Int>(data + OFFSET(DrawData,colorPitchB[index]));
1656                                 value = *Pointer<Short4>(buffer);
1657
1658                                 if((state.targetFormat[index] == FORMAT_A8R8G8B8 && bgraWriteMask != 0x0000000F) ||
1659                                    ((state.targetFormat[index] == FORMAT_X8R8G8B8 && bgraWriteMask != 0x00000007) &&
1660                                         (state.targetFormat[index] == FORMAT_X8R8G8B8 && bgraWriteMask != 0x0000000F)))   // FIXME: Need for masking when XRGB && Fh?
1661                                 {
1662                                         Short4 masked = value;
1663                                         c23 &= *Pointer<Short4>(constants + OFFSET(Constants,maskB4Q[bgraWriteMask][0]));
1664                                         masked &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskB4Q[bgraWriteMask][0]));
1665                                         c23 |= masked;
1666                                 }
1667
1668                                 c23 &= *Pointer<Short4>(constants + OFFSET(Constants,maskD23Q) + xMask * 8);
1669                                 value &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskD23Q) + xMask * 8);
1670                                 c23 |= value;
1671                                 *Pointer<Short4>(buffer) = c23;
1672                         }
1673                         break;
1674                 case FORMAT_A8B8G8R8:
1675                 case FORMAT_X8B8G8R8:   // FIXME: Don't touch alpha?
1676                 case FORMAT_SRGB8_X8:
1677                 case FORMAT_SRGB8_A8:
1678                         {
1679                                 Pointer<Byte> buffer = cBuffer + x * 4;
1680                                 Short4 value = *Pointer<Short4>(buffer);
1681
1682                                 bool masked = (((state.targetFormat[index] == FORMAT_A8B8G8R8 || state.targetFormat[index] == FORMAT_SRGB8_A8) && rgbaWriteMask != 0x0000000F) ||
1683                                               (((state.targetFormat[index] == FORMAT_X8B8G8R8 || state.targetFormat[index] == FORMAT_SRGB8_X8) && rgbaWriteMask != 0x00000007) &&
1684                                                ((state.targetFormat[index] == FORMAT_X8B8G8R8 || state.targetFormat[index] == FORMAT_SRGB8_X8) && rgbaWriteMask != 0x0000000F))); // FIXME: Need for masking when XBGR && Fh?
1685
1686                                 if(masked)
1687                                 {
1688                                         Short4 masked = value;
1689                                         c01 &= *Pointer<Short4>(constants + OFFSET(Constants,maskB4Q[rgbaWriteMask][0]));
1690                                         masked &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskB4Q[rgbaWriteMask][0]));
1691                                         c01 |= masked;
1692                                 }
1693
1694                                 c01 &= *Pointer<Short4>(constants + OFFSET(Constants,maskD01Q) + xMask * 8);
1695                                 value &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskD01Q) + xMask * 8);
1696                                 c01 |= value;
1697                                 *Pointer<Short4>(buffer) = c01;
1698
1699                                 buffer += *Pointer<Int>(data + OFFSET(DrawData,colorPitchB[index]));
1700                                 value = *Pointer<Short4>(buffer);
1701
1702                                 if(masked)
1703                                 {
1704                                         Short4 masked = value;
1705                                         c23 &= *Pointer<Short4>(constants + OFFSET(Constants,maskB4Q[rgbaWriteMask][0]));
1706                                         masked &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskB4Q[rgbaWriteMask][0]));
1707                                         c23 |= masked;
1708                                 }
1709
1710                                 c23 &= *Pointer<Short4>(constants + OFFSET(Constants,maskD23Q) + xMask * 8);
1711                                 value &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskD23Q) + xMask * 8);
1712                                 c23 |= value;
1713                                 *Pointer<Short4>(buffer) = c23;
1714                         }
1715                         break;
1716                 case FORMAT_R8:
1717                         if(rgbaWriteMask & 0x00000001)
1718                         {
1719                                 Pointer<Byte> buffer = cBuffer + 1 * x;
1720                                 Short4 value;
1721                                 value = Insert(value, *Pointer<Short>(buffer), 0);
1722                                 Int pitch = *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
1723                                 value = Insert(value, *Pointer<Short>(buffer + pitch), 1);
1724                                 value = UnpackLow(As<Byte8>(value), As<Byte8>(value));
1725
1726                                 current.x &= *Pointer<Short4>(constants + OFFSET(Constants, maskB4Q) + 8 * xMask);
1727                                 value &= *Pointer<Short4>(constants + OFFSET(Constants, invMaskB4Q) + 8 * xMask);
1728                                 current.x |= value;
1729
1730                                 *Pointer<Short>(buffer) = Extract(current.x, 0);
1731                                 *Pointer<Short>(buffer + pitch) = Extract(current.x, 1);
1732                         }
1733                         break;
1734                 case FORMAT_A8:
1735                         if(rgbaWriteMask & 0x00000008)
1736                         {
1737                                 Pointer<Byte> buffer = cBuffer + 1 * x;
1738                                 Short4 value;
1739                                 value = Insert(value, *Pointer<Short>(buffer), 0);
1740                                 Int pitch = *Pointer<Int>(data + OFFSET(DrawData,colorPitchB[index]));
1741                                 value = Insert(value, *Pointer<Short>(buffer + pitch), 1);
1742                                 value = UnpackLow(As<Byte8>(value), As<Byte8>(value));
1743
1744                                 current.w &= *Pointer<Short4>(constants + OFFSET(Constants,maskB4Q) + 8 * xMask);
1745                                 value &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskB4Q) + 8 * xMask);
1746                                 current.w |= value;
1747
1748                                 *Pointer<Short>(buffer) = Extract(current.w, 0);
1749                                 *Pointer<Short>(buffer + pitch) = Extract(current.w, 1);
1750                         }
1751                         break;
1752                 case FORMAT_G16R16:
1753                         {
1754                                 Pointer<Byte> buffer = cBuffer + 4 * x;
1755
1756                                 Short4 value = *Pointer<Short4>(buffer);
1757
1758                                 if((rgbaWriteMask & 0x00000003) != 0x00000003)
1759                                 {
1760                                         Short4 masked = value;
1761                                         current.x &= *Pointer<Short4>(constants + OFFSET(Constants,maskW01Q[rgbaWriteMask & 0x3][0]));
1762                                         masked &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskW01Q[rgbaWriteMask & 0x3][0]));
1763                                         current.x |= masked;
1764                                 }
1765
1766                                 current.x &= *Pointer<Short4>(constants + OFFSET(Constants,maskD01Q) + xMask * 8);
1767                                 value &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskD01Q) + xMask * 8);
1768                                 current.x |= value;
1769                                 *Pointer<Short4>(buffer) = current.x;
1770
1771                                 buffer += *Pointer<Int>(data + OFFSET(DrawData,colorPitchB[index]));
1772
1773                                 value = *Pointer<Short4>(buffer);
1774
1775                                 if((rgbaWriteMask & 0x00000003) != 0x00000003)
1776                                 {
1777                                         Short4 masked = value;
1778                                         current.y &= *Pointer<Short4>(constants + OFFSET(Constants,maskW01Q[rgbaWriteMask & 0x3][0]));
1779                                         masked &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskW01Q[rgbaWriteMask & 0x3][0]));
1780                                         current.y |= masked;
1781                                 }
1782
1783                                 current.y &= *Pointer<Short4>(constants + OFFSET(Constants,maskD23Q) + xMask * 8);
1784                                 value &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskD23Q) + xMask * 8);
1785                                 current.y |= value;
1786                                 *Pointer<Short4>(buffer) = current.y;
1787                         }
1788                         break;
1789                 case FORMAT_A16B16G16R16:
1790                         {
1791                                 Pointer<Byte> buffer = cBuffer + 8 * x;
1792
1793                                 {
1794                                         Short4 value = *Pointer<Short4>(buffer);
1795
1796                                         if(rgbaWriteMask != 0x0000000F)
1797                                         {
1798                                                 Short4 masked = value;
1799                                                 current.x &= *Pointer<Short4>(constants + OFFSET(Constants,maskW4Q[rgbaWriteMask][0]));
1800                                                 masked &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskW4Q[rgbaWriteMask][0]));
1801                                                 current.x |= masked;
1802                                         }
1803
1804                                         current.x &= *Pointer<Short4>(constants + OFFSET(Constants,maskQ0Q) + xMask * 8);
1805                                         value &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskQ0Q) + xMask * 8);
1806                                         current.x |= value;
1807                                         *Pointer<Short4>(buffer) = current.x;
1808                                 }
1809
1810                                 {
1811                                         Short4 value = *Pointer<Short4>(buffer + 8);
1812
1813                                         if(rgbaWriteMask != 0x0000000F)
1814                                         {
1815                                                 Short4 masked = value;
1816                                                 current.y &= *Pointer<Short4>(constants + OFFSET(Constants,maskW4Q[rgbaWriteMask][0]));
1817                                                 masked &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskW4Q[rgbaWriteMask][0]));
1818                                                 current.y |= masked;
1819                                         }
1820
1821                                         current.y &= *Pointer<Short4>(constants + OFFSET(Constants,maskQ1Q) + xMask * 8);
1822                                         value &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskQ1Q) + xMask * 8);
1823                                         current.y |= value;
1824                                         *Pointer<Short4>(buffer + 8) = current.y;
1825                                 }
1826
1827                                 buffer += *Pointer<Int>(data + OFFSET(DrawData,colorPitchB[index]));
1828
1829                                 {
1830                                         Short4 value = *Pointer<Short4>(buffer);
1831
1832                                         if(rgbaWriteMask != 0x0000000F)
1833                                         {
1834                                                 Short4 masked = value;
1835                                                 current.z &= *Pointer<Short4>(constants + OFFSET(Constants,maskW4Q[rgbaWriteMask][0]));
1836                                                 masked &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskW4Q[rgbaWriteMask][0]));
1837                                                 current.z |= masked;
1838                                         }
1839
1840                                         current.z &= *Pointer<Short4>(constants + OFFSET(Constants,maskQ2Q) + xMask * 8);
1841                                         value &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskQ2Q) + xMask * 8);
1842                                         current.z |= value;
1843                                         *Pointer<Short4>(buffer) = current.z;
1844                                 }
1845
1846                                 {
1847                                         Short4 value = *Pointer<Short4>(buffer + 8);
1848
1849                                         if(rgbaWriteMask != 0x0000000F)
1850                                         {
1851                                                 Short4 masked = value;
1852                                                 current.w &= *Pointer<Short4>(constants + OFFSET(Constants,maskW4Q[rgbaWriteMask][0]));
1853                                                 masked &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskW4Q[rgbaWriteMask][0]));
1854                                                 current.w |= masked;
1855                                         }
1856
1857                                         current.w &= *Pointer<Short4>(constants + OFFSET(Constants,maskQ3Q) + xMask * 8);
1858                                         value &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskQ3Q) + xMask * 8);
1859                                         current.w |= value;
1860                                         *Pointer<Short4>(buffer + 8) = current.w;
1861                                 }
1862                         }
1863                         break;
1864                 default:
1865                         ASSERT(false);
1866                 }
1867         }
1868
1869         void PixelRoutine::blendFactor(const Vector4f &blendFactor, const Vector4f &oC, const Vector4f &pixel, BlendFactor blendFactorActive)
1870         {
1871                 switch(blendFactorActive)
1872                 {
1873                 case BLEND_ZERO:
1874                         // Optimized
1875                         break;
1876                 case BLEND_ONE:
1877                         // Optimized
1878                         break;
1879                 case BLEND_SOURCE:
1880                         blendFactor.x = oC.x;
1881                         blendFactor.y = oC.y;
1882                         blendFactor.z = oC.z;
1883                         break;
1884                 case BLEND_INVSOURCE:
1885                         blendFactor.x = Float4(1.0f) - oC.x;
1886                         blendFactor.y = Float4(1.0f) - oC.y;
1887                         blendFactor.z = Float4(1.0f) - oC.z;
1888                         break;
1889                 case BLEND_DEST:
1890                         blendFactor.x = pixel.x;
1891                         blendFactor.y = pixel.y;
1892                         blendFactor.z = pixel.z;
1893                         break;
1894                 case BLEND_INVDEST:
1895                         blendFactor.x = Float4(1.0f) - pixel.x;
1896                         blendFactor.y = Float4(1.0f) - pixel.y;
1897                         blendFactor.z = Float4(1.0f) - pixel.z;
1898                         break;
1899                 case BLEND_SOURCEALPHA:
1900                         blendFactor.x = oC.w;
1901                         blendFactor.y = oC.w;
1902                         blendFactor.z = oC.w;
1903                         break;
1904                 case BLEND_INVSOURCEALPHA:
1905                         blendFactor.x = Float4(1.0f) - oC.w;
1906                         blendFactor.y = Float4(1.0f) - oC.w;
1907                         blendFactor.z = Float4(1.0f) - oC.w;
1908                         break;
1909                 case BLEND_DESTALPHA:
1910                         blendFactor.x = pixel.w;
1911                         blendFactor.y = pixel.w;
1912                         blendFactor.z = pixel.w;
1913                         break;
1914                 case BLEND_INVDESTALPHA:
1915                         blendFactor.x = Float4(1.0f) - pixel.w;
1916                         blendFactor.y = Float4(1.0f) - pixel.w;
1917                         blendFactor.z = Float4(1.0f) - pixel.w;
1918                         break;
1919                 case BLEND_SRCALPHASAT:
1920                         blendFactor.x = Float4(1.0f) - pixel.w;
1921                         blendFactor.x = Min(blendFactor.x, oC.w);
1922                         blendFactor.y = blendFactor.x;
1923                         blendFactor.z = blendFactor.x;
1924                         break;
1925                 case BLEND_CONSTANT:
1926                         blendFactor.x = *Pointer<Float4>(data + OFFSET(DrawData,factor.blendConstant4F[0]));
1927                         blendFactor.y = *Pointer<Float4>(data + OFFSET(DrawData,factor.blendConstant4F[1]));
1928                         blendFactor.z = *Pointer<Float4>(data + OFFSET(DrawData,factor.blendConstant4F[2]));
1929                         break;
1930                 case BLEND_INVCONSTANT:
1931                         blendFactor.x = *Pointer<Float4>(data + OFFSET(DrawData,factor.invBlendConstant4F[0]));
1932                         blendFactor.y = *Pointer<Float4>(data + OFFSET(DrawData,factor.invBlendConstant4F[1]));
1933                         blendFactor.z = *Pointer<Float4>(data + OFFSET(DrawData,factor.invBlendConstant4F[2]));
1934                         break;
1935                 default:
1936                         ASSERT(false);
1937                 }
1938         }
1939
1940         void PixelRoutine::blendFactorAlpha(const Vector4f &blendFactor, const Vector4f &oC, const Vector4f &pixel, BlendFactor blendFactorAlphaActive)
1941         {
1942                 switch(blendFactorAlphaActive)
1943                 {
1944                 case BLEND_ZERO:
1945                         // Optimized
1946                         break;
1947                 case BLEND_ONE:
1948                         // Optimized
1949                         break;
1950                 case BLEND_SOURCE:
1951                         blendFactor.w = oC.w;
1952                         break;
1953                 case BLEND_INVSOURCE:
1954                         blendFactor.w = Float4(1.0f) - oC.w;
1955                         break;
1956                 case BLEND_DEST:
1957                         blendFactor.w = pixel.w;
1958                         break;
1959                 case BLEND_INVDEST:
1960                         blendFactor.w = Float4(1.0f) - pixel.w;
1961                         break;
1962                 case BLEND_SOURCEALPHA:
1963                         blendFactor.w = oC.w;
1964                         break;
1965                 case BLEND_INVSOURCEALPHA:
1966                         blendFactor.w = Float4(1.0f) - oC.w;
1967                         break;
1968                 case BLEND_DESTALPHA:
1969                         blendFactor.w = pixel.w;
1970                         break;
1971                 case BLEND_INVDESTALPHA:
1972                         blendFactor.w = Float4(1.0f) - pixel.w;
1973                         break;
1974                 case BLEND_SRCALPHASAT:
1975                         blendFactor.w = Float4(1.0f);
1976                         break;
1977                 case BLEND_CONSTANT:
1978                         blendFactor.w = *Pointer<Float4>(data + OFFSET(DrawData,factor.blendConstant4F[3]));
1979                         break;
1980                 case BLEND_INVCONSTANT:
1981                         blendFactor.w = *Pointer<Float4>(data + OFFSET(DrawData,factor.invBlendConstant4F[3]));
1982                         break;
1983                 default:
1984                         ASSERT(false);
1985                 }
1986         }
1987
1988         void PixelRoutine::alphaBlend(int index, Pointer<Byte> &cBuffer, Vector4f &oC, Int &x)
1989         {
1990                 if(!state.alphaBlendActive)
1991                 {
1992                         return;
1993                 }
1994
1995                 Pointer<Byte> buffer;
1996                 Vector4f pixel;
1997
1998                 Vector4s color;
1999                 Short4 c01;
2000                 Short4 c23;
2001
2002                 Float4 one;
2003                 if(Surface::isFloatFormat(state.targetFormat[index]))
2004                 {
2005                         one = Float4(1.0f);
2006                 }
2007                 else if(Surface::isNonNormalizedInteger(state.targetFormat[index]))
2008                 {
2009                         one = As<Float4>(Surface::isUnsignedComponent(state.targetFormat[index], 0) ? Int4(0xFFFFFFFF) : Int4(0x7FFFFFFF));
2010                 }
2011
2012                 switch(state.targetFormat[index])
2013                 {
2014                 case FORMAT_R32I:
2015                 case FORMAT_R32UI:
2016                 case FORMAT_R32F:
2017                         buffer = cBuffer;
2018                         // FIXME: movlps
2019                         pixel.x.x = *Pointer<Float>(buffer + 4 * x + 0);
2020                         pixel.x.y = *Pointer<Float>(buffer + 4 * x + 4);
2021                         buffer += *Pointer<Int>(data + OFFSET(DrawData,colorPitchB[index]));
2022                         // FIXME: movhps
2023                         pixel.x.z = *Pointer<Float>(buffer + 4 * x + 0);
2024                         pixel.x.w = *Pointer<Float>(buffer + 4 * x + 4);
2025                         pixel.y = pixel.z = pixel.w = one;
2026                         break;
2027                 case FORMAT_G32R32I:
2028                 case FORMAT_G32R32UI:
2029                 case FORMAT_G32R32F:
2030                         buffer = cBuffer;
2031                         pixel.x = *Pointer<Float4>(buffer + 8 * x, 16);
2032                         buffer += *Pointer<Int>(data + OFFSET(DrawData,colorPitchB[index]));
2033                         pixel.y = *Pointer<Float4>(buffer + 8 * x, 16);
2034                         pixel.z = pixel.x;
2035                         pixel.x = ShuffleLowHigh(pixel.x, pixel.y, 0x88);
2036                         pixel.z = ShuffleLowHigh(pixel.z, pixel.y, 0xDD);
2037                         pixel.y = pixel.z;
2038                         pixel.z = pixel.w = one;
2039                         break;
2040                 case FORMAT_X32B32G32R32F:
2041                 case FORMAT_A32B32G32R32F:
2042                 case FORMAT_A32B32G32R32I:
2043                 case FORMAT_A32B32G32R32UI:
2044                         buffer = cBuffer;
2045                         pixel.x = *Pointer<Float4>(buffer + 16 * x, 16);
2046                         pixel.y = *Pointer<Float4>(buffer + 16 * x + 16, 16);
2047                         buffer += *Pointer<Int>(data + OFFSET(DrawData,colorPitchB[index]));
2048                         pixel.z = *Pointer<Float4>(buffer + 16 * x, 16);
2049                         pixel.w = *Pointer<Float4>(buffer + 16 * x + 16, 16);
2050                         transpose4x4(pixel.x, pixel.y, pixel.z, pixel.w);
2051                         if(state.targetFormat[index] == FORMAT_X32B32G32R32F)
2052                         {
2053                                 pixel.w = Float4(1.0f);
2054                         }
2055                         break;
2056                 default:
2057                         ASSERT(false);
2058                 }
2059
2060                 if((postBlendSRGB && state.writeSRGB) || isSRGB(index))
2061                 {
2062                         sRGBtoLinear(pixel.x);
2063                         sRGBtoLinear(pixel.y);
2064                         sRGBtoLinear(pixel.z);
2065                 }
2066
2067                 // Final Color = ObjectColor * SourceBlendFactor + PixelColor * DestinationBlendFactor
2068                 Vector4f sourceFactor;
2069                 Vector4f destFactor;
2070
2071                 blendFactor(sourceFactor, oC, pixel, state.sourceBlendFactor);
2072                 blendFactor(destFactor, oC, pixel, state.destBlendFactor);
2073
2074                 if(state.sourceBlendFactor != BLEND_ONE && state.sourceBlendFactor != BLEND_ZERO)
2075                 {
2076                         oC.x *= sourceFactor.x;
2077                         oC.y *= sourceFactor.y;
2078                         oC.z *= sourceFactor.z;
2079                 }
2080
2081                 if(state.destBlendFactor != BLEND_ONE && state.destBlendFactor != BLEND_ZERO)
2082                 {
2083                         pixel.x *= destFactor.x;
2084                         pixel.y *= destFactor.y;
2085                         pixel.z *= destFactor.z;
2086                 }
2087
2088                 switch(state.blendOperation)
2089                 {
2090                 case BLENDOP_ADD:
2091                         oC.x += pixel.x;
2092                         oC.y += pixel.y;
2093                         oC.z += pixel.z;
2094                         break;
2095                 case BLENDOP_SUB:
2096                         oC.x -= pixel.x;
2097                         oC.y -= pixel.y;
2098                         oC.z -= pixel.z;
2099                         break;
2100                 case BLENDOP_INVSUB:
2101                         oC.x = pixel.x - oC.x;
2102                         oC.y = pixel.y - oC.y;
2103                         oC.z = pixel.z - oC.z;
2104                         break;
2105                 case BLENDOP_MIN:
2106                         oC.x = Min(oC.x, pixel.x);
2107                         oC.y = Min(oC.y, pixel.y);
2108                         oC.z = Min(oC.z, pixel.z);
2109                         break;
2110                 case BLENDOP_MAX:
2111                         oC.x = Max(oC.x, pixel.x);
2112                         oC.y = Max(oC.y, pixel.y);
2113                         oC.z = Max(oC.z, pixel.z);
2114                         break;
2115                 case BLENDOP_SOURCE:
2116                         // No operation
2117                         break;
2118                 case BLENDOP_DEST:
2119                         oC.x = pixel.x;
2120                         oC.y = pixel.y;
2121                         oC.z = pixel.z;
2122                         break;
2123                 case BLENDOP_NULL:
2124                         oC.x = Float4(0.0f);
2125                         oC.y = Float4(0.0f);
2126                         oC.z = Float4(0.0f);
2127                         break;
2128                 default:
2129                         ASSERT(false);
2130                 }
2131
2132                 blendFactorAlpha(sourceFactor, oC, pixel, state.sourceBlendFactorAlpha);
2133                 blendFactorAlpha(destFactor, oC, pixel, state.destBlendFactorAlpha);
2134
2135                 if(state.sourceBlendFactorAlpha != BLEND_ONE && state.sourceBlendFactorAlpha != BLEND_ZERO)
2136                 {
2137                         oC.w *= sourceFactor.w;
2138                 }
2139
2140                 if(state.destBlendFactorAlpha != BLEND_ONE && state.destBlendFactorAlpha != BLEND_ZERO)
2141                 {
2142                         pixel.w *= destFactor.w;
2143                 }
2144
2145                 switch(state.blendOperationAlpha)
2146                 {
2147                 case BLENDOP_ADD:
2148                         oC.w += pixel.w;
2149                         break;
2150                 case BLENDOP_SUB:
2151                         oC.w -= pixel.w;
2152                         break;
2153                 case BLENDOP_INVSUB:
2154                         pixel.w -= oC.w;
2155                         oC.w = pixel.w;
2156                         break;
2157                 case BLENDOP_MIN:
2158                         oC.w = Min(oC.w, pixel.w);
2159                         break;
2160                 case BLENDOP_MAX:
2161                         oC.w = Max(oC.w, pixel.w);
2162                         break;
2163                 case BLENDOP_SOURCE:
2164                         // No operation
2165                         break;
2166                 case BLENDOP_DEST:
2167                         oC.w = pixel.w;
2168                         break;
2169                 case BLENDOP_NULL:
2170                         oC.w = Float4(0.0f);
2171                         break;
2172                 default:
2173                         ASSERT(false);
2174                 }
2175         }
2176
2177         void PixelRoutine::writeColor(int index, Pointer<Byte> &cBuffer, Int &x, Vector4f &oC, Int &sMask, Int &zMask, Int &cMask)
2178         {
2179                 switch(state.targetFormat[index])
2180                 {
2181                 case FORMAT_R32F:
2182                 case FORMAT_R32I:
2183                 case FORMAT_R32UI:
2184                 case FORMAT_R16I:
2185                 case FORMAT_R16UI:
2186                 case FORMAT_R8I:
2187                 case FORMAT_R8UI:
2188                         break;
2189                 case FORMAT_G32R32F:
2190                 case FORMAT_G32R32I:
2191                 case FORMAT_G32R32UI:
2192                 case FORMAT_G16R16I:
2193                 case FORMAT_G16R16UI:
2194                 case FORMAT_G8R8I:
2195                 case FORMAT_G8R8UI:
2196                         oC.z = oC.x;
2197                         oC.x = UnpackLow(oC.x, oC.y);
2198                         oC.z = UnpackHigh(oC.z, oC.y);
2199                         oC.y = oC.z;
2200                         break;
2201                 case FORMAT_X32B32G32R32F:
2202                 case FORMAT_A32B32G32R32F:
2203                 case FORMAT_A32B32G32R32I:
2204                 case FORMAT_A32B32G32R32UI:
2205                 case FORMAT_A16B16G16R16I:
2206                 case FORMAT_A16B16G16R16UI:
2207                 case FORMAT_A8B8G8R8I:
2208                 case FORMAT_A8B8G8R8UI:
2209                         transpose4x4(oC.x, oC.y, oC.z, oC.w);
2210                         break;
2211                 default:
2212                         ASSERT(false);
2213                 }
2214
2215                 int rgbaWriteMask = state.colorWriteActive(index);
2216
2217                 Int xMask;   // Combination of all masks
2218
2219                 if(state.depthTestActive)
2220                 {
2221                         xMask = zMask;
2222                 }
2223                 else
2224                 {
2225                         xMask = cMask;
2226                 }
2227
2228                 if(state.stencilActive)
2229                 {
2230                         xMask &= sMask;
2231                 }
2232
2233                 Pointer<Byte> buffer;
2234                 Float4 value;
2235
2236                 switch(state.targetFormat[index])
2237                 {
2238                 case FORMAT_R32F:
2239                 case FORMAT_R32I:
2240                 case FORMAT_R32UI:
2241                         if(rgbaWriteMask & 0x00000001)
2242                         {
2243                                 buffer = cBuffer + 4 * x;
2244
2245                                 // FIXME: movlps
2246                                 value.x = *Pointer<Float>(buffer + 0);
2247                                 value.y = *Pointer<Float>(buffer + 4);
2248
2249                                 buffer += *Pointer<Int>(data + OFFSET(DrawData,colorPitchB[index]));
2250
2251                                 // FIXME: movhps
2252                                 value.z = *Pointer<Float>(buffer + 0);
2253                                 value.w = *Pointer<Float>(buffer + 4);
2254
2255                                 oC.x = As<Float4>(As<Int4>(oC.x) & *Pointer<Int4>(constants + OFFSET(Constants,maskD4X) + xMask * 16, 16));
2256                                 value = As<Float4>(As<Int4>(value) & *Pointer<Int4>(constants + OFFSET(Constants,invMaskD4X) + xMask * 16, 16));
2257                                 oC.x = As<Float4>(As<Int4>(oC.x) | As<Int4>(value));
2258
2259                                 // FIXME: movhps
2260                                 *Pointer<Float>(buffer + 0) = oC.x.z;
2261                                 *Pointer<Float>(buffer + 4) = oC.x.w;
2262
2263                                 buffer -= *Pointer<Int>(data + OFFSET(DrawData,colorPitchB[index]));
2264
2265                                 // FIXME: movlps
2266                                 *Pointer<Float>(buffer + 0) = oC.x.x;
2267                                 *Pointer<Float>(buffer + 4) = oC.x.y;
2268                         }
2269                         break;
2270                 case FORMAT_R16I:
2271                 case FORMAT_R16UI:
2272                         if(rgbaWriteMask & 0x00000001)
2273                         {
2274                                 buffer = cBuffer + 2 * x;
2275
2276                                 UShort4 xyzw;
2277                                 xyzw = As<UShort4>(Insert(As<Int2>(xyzw), *Pointer<Int>(buffer), 0));
2278
2279                                 buffer += *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
2280
2281                                 xyzw = As<UShort4>(Insert(As<Int2>(xyzw), *Pointer<Int>(buffer), 1));
2282                                 value = As<Float4>(Int4(xyzw));
2283
2284                                 oC.x = As<Float4>(As<Int4>(oC.x) & *Pointer<Int4>(constants + OFFSET(Constants, maskD4X) + xMask * 16, 16));
2285                                 value = As<Float4>(As<Int4>(value) & *Pointer<Int4>(constants + OFFSET(Constants, invMaskD4X) + xMask * 16, 16));
2286                                 oC.x = As<Float4>(As<Int4>(oC.x) | As<Int4>(value));
2287
2288                                 if(state.targetFormat[index] == FORMAT_R16I)
2289                                 {
2290                                         Float component = oC.x.z;
2291                                         *Pointer<Short>(buffer + 0) = Short(As<Int>(component));
2292                                         component = oC.x.w;
2293                                         *Pointer<Short>(buffer + 2) = Short(As<Int>(component));
2294
2295                                         buffer -= *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
2296
2297                                         component = oC.x.x;
2298                                         *Pointer<Short>(buffer + 0) = Short(As<Int>(component));
2299                                         component = oC.x.y;
2300                                         *Pointer<Short>(buffer + 2) = Short(As<Int>(component));
2301                                 }
2302                                 else // FORMAT_R16UI
2303                                 {
2304                                         Float component = oC.x.z;
2305                                         *Pointer<UShort>(buffer + 0) = UShort(As<Int>(component));
2306                                         component = oC.x.w;
2307                                         *Pointer<UShort>(buffer + 2) = UShort(As<Int>(component));
2308
2309                                         buffer -= *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
2310
2311                                         component = oC.x.x;
2312                                         *Pointer<UShort>(buffer + 0) = UShort(As<Int>(component));
2313                                         component = oC.x.y;
2314                                         *Pointer<UShort>(buffer + 2) = UShort(As<Int>(component));
2315                                 }
2316                         }
2317                         break;
2318                 case FORMAT_R8I:
2319                 case FORMAT_R8UI:
2320                         if(rgbaWriteMask & 0x00000001)
2321                         {
2322                                 buffer = cBuffer + x;
2323
2324                                 UInt xyzw, packedCol;
2325
2326                                 xyzw = UInt(*Pointer<UShort>(buffer)) << 16;
2327                                 buffer += *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
2328                                 xyzw |= UInt(*Pointer<UShort>(buffer));
2329
2330                                 Short4 tmpCol = Short4(As<Int4>(oC.x));
2331                                 if(state.targetFormat[index] == FORMAT_R8I)
2332                                 {
2333                                         tmpCol = As<Short4>(Pack(tmpCol, tmpCol));
2334                                 }
2335                                 else
2336                                 {
2337                                         tmpCol = As<Short4>(Pack(As<UShort4>(tmpCol), As<UShort4>(tmpCol)));
2338                                 }
2339                                 packedCol = Extract(As<Int2>(tmpCol), 0);
2340
2341                                 packedCol = (packedCol & *Pointer<UInt>(constants + OFFSET(Constants, maskB4Q) + 8 * xMask)) |
2342                                             (xyzw & *Pointer<UInt>(constants + OFFSET(Constants, invMaskB4Q) + 8 * xMask));
2343
2344                                 *Pointer<UShort>(buffer) = UShort(packedCol >> 16);
2345                                 buffer -= *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
2346                                 *Pointer<UShort>(buffer) = UShort(packedCol);
2347                         }
2348                         break;
2349                 case FORMAT_G32R32F:
2350                 case FORMAT_G32R32I:
2351                 case FORMAT_G32R32UI:
2352                         buffer = cBuffer + 8 * x;
2353
2354                         value = *Pointer<Float4>(buffer);
2355
2356                         if((rgbaWriteMask & 0x00000003) != 0x00000003)
2357                         {
2358                                 Float4 masked = value;
2359                                 oC.x = As<Float4>(As<Int4>(oC.x) & *Pointer<Int4>(constants + OFFSET(Constants,maskD01X[rgbaWriteMask & 0x3][0])));
2360                                 masked = As<Float4>(As<Int4>(masked) & *Pointer<Int4>(constants + OFFSET(Constants,invMaskD01X[rgbaWriteMask & 0x3][0])));
2361                                 oC.x = As<Float4>(As<Int4>(oC.x) | As<Int4>(masked));
2362                         }
2363
2364                         oC.x = As<Float4>(As<Int4>(oC.x) & *Pointer<Int4>(constants + OFFSET(Constants,maskQ01X) + xMask * 16, 16));
2365                         value = As<Float4>(As<Int4>(value) & *Pointer<Int4>(constants + OFFSET(Constants,invMaskQ01X) + xMask * 16, 16));
2366                         oC.x = As<Float4>(As<Int4>(oC.x) | As<Int4>(value));
2367                         *Pointer<Float4>(buffer) = oC.x;
2368
2369                         buffer += *Pointer<Int>(data + OFFSET(DrawData,colorPitchB[index]));
2370
2371                         value = *Pointer<Float4>(buffer);
2372
2373                         if((rgbaWriteMask & 0x00000003) != 0x00000003)
2374                         {
2375                                 Float4 masked;
2376
2377                                 masked = value;
2378                                 oC.y = As<Float4>(As<Int4>(oC.y) & *Pointer<Int4>(constants + OFFSET(Constants,maskD01X[rgbaWriteMask & 0x3][0])));
2379                                 masked = As<Float4>(As<Int4>(masked) & *Pointer<Int4>(constants + OFFSET(Constants,invMaskD01X[rgbaWriteMask & 0x3][0])));
2380                                 oC.y = As<Float4>(As<Int4>(oC.y) | As<Int4>(masked));
2381                         }
2382
2383                         oC.y = As<Float4>(As<Int4>(oC.y) & *Pointer<Int4>(constants + OFFSET(Constants,maskQ23X) + xMask * 16, 16));
2384                         value = As<Float4>(As<Int4>(value) & *Pointer<Int4>(constants + OFFSET(Constants,invMaskQ23X) + xMask * 16, 16));
2385                         oC.y = As<Float4>(As<Int4>(oC.y) | As<Int4>(value));
2386                         *Pointer<Float4>(buffer) = oC.y;
2387                         break;
2388                 case FORMAT_G16R16I:
2389                 case FORMAT_G16R16UI:
2390                         if((rgbaWriteMask & 0x00000003) != 0x0)
2391                         {
2392                                 buffer = cBuffer + 4 * x;
2393
2394                                 UInt2 rgbaMask;
2395                                 UShort4 packedCol = UShort4(As<Int4>(oC.x));
2396                                 UShort4 value = *Pointer<UShort4>(buffer);
2397                                 UInt2 mergedMask = *Pointer<UInt2>(constants + OFFSET(Constants, maskD01Q) + xMask * 8);
2398                                 if((rgbaWriteMask & 0x3) != 0x3)
2399                                 {
2400                                         Int tmpMask = *Pointer<Int>(constants + OFFSET(Constants, maskW4Q[rgbaWriteMask & 0x3][0]));
2401                                         rgbaMask = As<UInt2>(Int2(tmpMask, tmpMask));
2402                                         mergedMask &= rgbaMask;
2403                                 }
2404                                 *Pointer<UInt2>(buffer) = (As<UInt2>(packedCol) & mergedMask) | (As<UInt2>(value) & ~mergedMask);
2405
2406                                 buffer += *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
2407
2408                                 packedCol = UShort4(As<Int4>(oC.y));
2409                                 value = *Pointer<UShort4>(buffer);
2410                                 mergedMask = *Pointer<UInt2>(constants + OFFSET(Constants, maskD23Q) + xMask * 8);
2411                                 if((rgbaWriteMask & 0x3) != 0x3)
2412                                 {
2413                                         mergedMask &= rgbaMask;
2414                                 }
2415                                 *Pointer<UInt2>(buffer) = (As<UInt2>(packedCol) & mergedMask) | (As<UInt2>(value) & ~mergedMask);
2416                         }
2417                         break;
2418                 case FORMAT_G8R8I:
2419                 case FORMAT_G8R8UI:
2420                         if((rgbaWriteMask & 0x00000003) != 0x0)
2421                         {
2422                                 buffer = cBuffer + 2 * x;
2423
2424                                 Int2 xyzw, packedCol;
2425
2426                                 xyzw = Insert(xyzw, *Pointer<Int>(buffer), 0);
2427                                 buffer += *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
2428                                 xyzw = Insert(xyzw, *Pointer<Int>(buffer), 1);
2429
2430                                 if(state.targetFormat[index] == FORMAT_G8R8I)
2431                                 {
2432                                         packedCol = As<Int2>(Pack(Short4(As<Int4>(oC.x)), Short4(As<Int4>(oC.y))));
2433                                 }
2434                                 else
2435                                 {
2436                                         packedCol = As<Int2>(Pack(UShort4(As<Int4>(oC.x)), UShort4(As<Int4>(oC.y))));
2437                                 }
2438
2439                                 UInt2 mergedMask = *Pointer<UInt2>(constants + OFFSET(Constants, maskW4Q) + xMask * 8);
2440                                 if((rgbaWriteMask & 0x3) != 0x3)
2441                                 {
2442                                         Int tmpMask = *Pointer<Int>(constants + OFFSET(Constants, maskB4Q[5 * (rgbaWriteMask & 0x3)][0]));
2443                                         UInt2 rgbaMask = As<UInt2>(Int2(tmpMask, tmpMask));
2444                                         mergedMask &= rgbaMask;
2445                                 }
2446
2447                                 packedCol = As<Int2>((As<UInt2>(packedCol) & mergedMask) | (As<UInt2>(xyzw) & ~mergedMask));
2448
2449                                 *Pointer<UInt>(buffer) = As<UInt>(Extract(packedCol, 1));
2450                                 buffer -= *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
2451                                 *Pointer<UInt>(buffer) = As<UInt>(Extract(packedCol, 0));
2452                         }
2453                         break;
2454                 case FORMAT_X32B32G32R32F:
2455                 case FORMAT_A32B32G32R32F:
2456                 case FORMAT_A32B32G32R32I:
2457                 case FORMAT_A32B32G32R32UI:
2458                         buffer = cBuffer + 16 * x;
2459
2460                         {
2461                                 value = *Pointer<Float4>(buffer, 16);
2462
2463                                 if(rgbaWriteMask != 0x0000000F)
2464                                 {
2465                                         Float4 masked = value;
2466                                         oC.x = As<Float4>(As<Int4>(oC.x) & *Pointer<Int4>(constants + OFFSET(Constants,maskD4X[rgbaWriteMask][0])));
2467                                         masked = As<Float4>(As<Int4>(masked) & *Pointer<Int4>(constants + OFFSET(Constants,invMaskD4X[rgbaWriteMask][0])));
2468                                         oC.x = As<Float4>(As<Int4>(oC.x) | As<Int4>(masked));
2469                                 }
2470
2471                                 oC.x = As<Float4>(As<Int4>(oC.x) & *Pointer<Int4>(constants + OFFSET(Constants,maskX0X) + xMask * 16, 16));
2472                                 value = As<Float4>(As<Int4>(value) & *Pointer<Int4>(constants + OFFSET(Constants,invMaskX0X) + xMask * 16, 16));
2473                                 oC.x = As<Float4>(As<Int4>(oC.x) | As<Int4>(value));
2474                                 *Pointer<Float4>(buffer, 16) = oC.x;
2475                         }
2476
2477                         {
2478                                 value = *Pointer<Float4>(buffer + 16, 16);
2479
2480                                 if(rgbaWriteMask != 0x0000000F)
2481                                 {
2482                                         Float4 masked = value;
2483                                         oC.y = As<Float4>(As<Int4>(oC.y) & *Pointer<Int4>(constants + OFFSET(Constants,maskD4X[rgbaWriteMask][0])));
2484                                         masked = As<Float4>(As<Int4>(masked) & *Pointer<Int4>(constants + OFFSET(Constants,invMaskD4X[rgbaWriteMask][0])));
2485                                         oC.y = As<Float4>(As<Int4>(oC.y) | As<Int4>(masked));
2486                                 }
2487
2488                                 oC.y = As<Float4>(As<Int4>(oC.y) & *Pointer<Int4>(constants + OFFSET(Constants,maskX1X) + xMask * 16, 16));
2489                                 value = As<Float4>(As<Int4>(value) & *Pointer<Int4>(constants + OFFSET(Constants,invMaskX1X) + xMask * 16, 16));
2490                                 oC.y = As<Float4>(As<Int4>(oC.y) | As<Int4>(value));
2491                                 *Pointer<Float4>(buffer + 16, 16) = oC.y;
2492                         }
2493
2494                         buffer += *Pointer<Int>(data + OFFSET(DrawData,colorPitchB[index]));
2495
2496                         {
2497                                 value = *Pointer<Float4>(buffer, 16);
2498
2499                                 if(rgbaWriteMask != 0x0000000F)
2500                                 {
2501                                         Float4 masked = value;
2502                                         oC.z = As<Float4>(As<Int4>(oC.z) & *Pointer<Int4>(constants + OFFSET(Constants,maskD4X[rgbaWriteMask][0])));
2503                                         masked = As<Float4>(As<Int4>(masked) & *Pointer<Int4>(constants + OFFSET(Constants,invMaskD4X[rgbaWriteMask][0])));
2504                                         oC.z = As<Float4>(As<Int4>(oC.z) | As<Int4>(masked));
2505                                 }
2506
2507                                 oC.z = As<Float4>(As<Int4>(oC.z) & *Pointer<Int4>(constants + OFFSET(Constants,maskX2X) + xMask * 16, 16));
2508                                 value = As<Float4>(As<Int4>(value) & *Pointer<Int4>(constants + OFFSET(Constants,invMaskX2X) + xMask * 16, 16));
2509                                 oC.z = As<Float4>(As<Int4>(oC.z) | As<Int4>(value));
2510                                 *Pointer<Float4>(buffer, 16) = oC.z;
2511                         }
2512
2513                         {
2514                                 value = (state.targetFormat[index] == FORMAT_X32B32G32R32F) ? Float4(1.0f) : *Pointer<Float4>(buffer + 16, 16);
2515
2516                                 if(rgbaWriteMask != 0x0000000F)
2517                                 {
2518                                         Float4 masked = value;
2519                                         oC.w = As<Float4>(As<Int4>(oC.w) & *Pointer<Int4>(constants + OFFSET(Constants,maskD4X[rgbaWriteMask][0])));
2520                                         masked = As<Float4>(As<Int4>(masked) & *Pointer<Int4>(constants + OFFSET(Constants,invMaskD4X[rgbaWriteMask][0])));
2521                                         oC.w = As<Float4>(As<Int4>(oC.w) | As<Int4>(masked));
2522                                 }
2523
2524                                 oC.w = As<Float4>(As<Int4>(oC.w) & *Pointer<Int4>(constants + OFFSET(Constants,maskX3X) + xMask * 16, 16));
2525                                 value = As<Float4>(As<Int4>(value) & *Pointer<Int4>(constants + OFFSET(Constants,invMaskX3X) + xMask * 16, 16));
2526                                 oC.w = As<Float4>(As<Int4>(oC.w) | As<Int4>(value));
2527                                 *Pointer<Float4>(buffer + 16, 16) = oC.w;
2528                         }
2529                         break;
2530                 case FORMAT_A16B16G16R16I:
2531                 case FORMAT_A16B16G16R16UI:
2532                         if((rgbaWriteMask & 0x0000000F) != 0x0)
2533                         {
2534                                 buffer = cBuffer + 8 * x;
2535
2536                                 UInt4 rgbaMask;
2537                                 UShort8 value = *Pointer<UShort8>(buffer);
2538                                 UShort8 packedCol = UShort8(UShort4(As<Int4>(oC.x)), UShort4(As<Int4>(oC.y)));
2539                                 UInt4 mergedMask = *Pointer<UInt4>(constants + OFFSET(Constants, maskQ01X) + xMask * 16);
2540                                 if((rgbaWriteMask & 0xF) != 0xF)
2541                                 {
2542                                         UInt2 tmpMask = *Pointer<UInt2>(constants + OFFSET(Constants, maskW4Q[rgbaWriteMask][0]));
2543                                         rgbaMask = UInt4(tmpMask, tmpMask);
2544                                         mergedMask &= rgbaMask;
2545                                 }
2546                                 *Pointer<UInt4>(buffer) = (As<UInt4>(packedCol) & mergedMask) | (As<UInt4>(value) & ~mergedMask);
2547
2548                                 buffer += *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
2549
2550                                 value = *Pointer<UShort8>(buffer);
2551                                 packedCol = UShort8(UShort4(As<Int4>(oC.z)), UShort4(As<Int4>(oC.w)));
2552                                 mergedMask = *Pointer<UInt4>(constants + OFFSET(Constants, maskQ23X) + xMask * 16);
2553                                 if((rgbaWriteMask & 0xF) != 0xF)
2554                                 {
2555                                         mergedMask &= rgbaMask;
2556                                 }
2557                                 *Pointer<UInt4>(buffer) = (As<UInt4>(packedCol) & mergedMask) | (As<UInt4>(value) & ~mergedMask);
2558                         }
2559                         break;
2560                 case FORMAT_A8B8G8R8I:
2561                 case FORMAT_A8B8G8R8UI:
2562                         if((rgbaWriteMask & 0x0000000F) != 0x0)
2563                         {
2564                                 UInt2 value, packedCol, mergedMask;
2565
2566                                 buffer = cBuffer + 4 * x;
2567
2568                                 if(state.targetFormat[index] == FORMAT_A8B8G8R8I)
2569                                 {
2570                                         packedCol = As<UInt2>(Pack(Short4(As<Int4>(oC.x)), Short4(As<Int4>(oC.y))));
2571                                 }
2572                                 else
2573                                 {
2574                                         packedCol = As<UInt2>(Pack(UShort4(As<Int4>(oC.x)), UShort4(As<Int4>(oC.y))));
2575                                 }
2576                                 value = *Pointer<UInt2>(buffer, 16);
2577                                 mergedMask = *Pointer<UInt2>(constants + OFFSET(Constants, maskD01Q) + xMask * 8);
2578                                 if(rgbaWriteMask != 0xF)
2579                                 {
2580                                         mergedMask &= *Pointer<UInt2>(constants + OFFSET(Constants, maskB4Q[rgbaWriteMask][0]));
2581                                 }
2582                                 *Pointer<UInt2>(buffer) = (packedCol & mergedMask) | (value & ~mergedMask);
2583
2584                                 buffer += *Pointer<Int>(data + OFFSET(DrawData, colorPitchB[index]));
2585
2586                                 if(state.targetFormat[index] == FORMAT_A8B8G8R8I)
2587                                 {
2588                                         packedCol = As<UInt2>(Pack(Short4(As<Int4>(oC.z)), Short4(As<Int4>(oC.w))));
2589                                 }
2590                                 else
2591                                 {
2592                                         packedCol = As<UInt2>(Pack(UShort4(As<Int4>(oC.z)), UShort4(As<Int4>(oC.w))));
2593                                 }
2594                                 value = *Pointer<UInt2>(buffer, 16);
2595                                 mergedMask = *Pointer<UInt2>(constants + OFFSET(Constants, maskD23Q) + xMask * 8);
2596                                 if(rgbaWriteMask != 0xF)
2597                                 {
2598                                         mergedMask &= *Pointer<UInt2>(constants + OFFSET(Constants, maskB4Q[rgbaWriteMask][0]));
2599                                 }
2600                                 *Pointer<UInt2>(buffer) = (packedCol & mergedMask) | (value & ~mergedMask);
2601                         }
2602                         break;
2603                 default:
2604                         ASSERT(false);
2605                 }
2606         }
2607
2608         UShort4 PixelRoutine::convertFixed16(Float4 &cf, bool saturate)
2609         {
2610                 return UShort4(cf * Float4(0xFFFF), saturate);
2611         }
2612
2613         void PixelRoutine::sRGBtoLinear16_12_16(Vector4s &c)
2614         {
2615                 c.x = As<UShort4>(c.x) >> 4;
2616                 c.y = As<UShort4>(c.y) >> 4;
2617                 c.z = As<UShort4>(c.z) >> 4;
2618
2619                 sRGBtoLinear12_16(c);
2620         }
2621
2622         void PixelRoutine::sRGBtoLinear12_16(Vector4s &c)
2623         {
2624                 Pointer<Byte> LUT = constants + OFFSET(Constants,sRGBtoLinear12_16);
2625
2626                 c.x = Insert(c.x, *Pointer<Short>(LUT + 2 * Int(Extract(c.x, 0))), 0);
2627                 c.x = Insert(c.x, *Pointer<Short>(LUT + 2 * Int(Extract(c.x, 1))), 1);
2628                 c.x = Insert(c.x, *Pointer<Short>(LUT + 2 * Int(Extract(c.x, 2))), 2);
2629                 c.x = Insert(c.x, *Pointer<Short>(LUT + 2 * Int(Extract(c.x, 3))), 3);
2630
2631                 c.y = Insert(c.y, *Pointer<Short>(LUT + 2 * Int(Extract(c.y, 0))), 0);
2632                 c.y = Insert(c.y, *Pointer<Short>(LUT + 2 * Int(Extract(c.y, 1))), 1);
2633                 c.y = Insert(c.y, *Pointer<Short>(LUT + 2 * Int(Extract(c.y, 2))), 2);
2634                 c.y = Insert(c.y, *Pointer<Short>(LUT + 2 * Int(Extract(c.y, 3))), 3);
2635
2636                 c.z = Insert(c.z, *Pointer<Short>(LUT + 2 * Int(Extract(c.z, 0))), 0);
2637                 c.z = Insert(c.z, *Pointer<Short>(LUT + 2 * Int(Extract(c.z, 1))), 1);
2638                 c.z = Insert(c.z, *Pointer<Short>(LUT + 2 * Int(Extract(c.z, 2))), 2);
2639                 c.z = Insert(c.z, *Pointer<Short>(LUT + 2 * Int(Extract(c.z, 3))), 3);
2640         }
2641
2642         void PixelRoutine::linearToSRGB16_12_16(Vector4s &c)
2643         {
2644                 c.x = As<UShort4>(c.x) >> 4;
2645                 c.y = As<UShort4>(c.y) >> 4;
2646                 c.z = As<UShort4>(c.z) >> 4;
2647
2648                 linearToSRGB12_16(c);
2649         }
2650
2651         void PixelRoutine::linearToSRGB12_16(Vector4s &c)
2652         {
2653                 Pointer<Byte> LUT = constants + OFFSET(Constants,linearToSRGB12_16);
2654
2655                 c.x = Insert(c.x, *Pointer<Short>(LUT + 2 * Int(Extract(c.x, 0))), 0);
2656                 c.x = Insert(c.x, *Pointer<Short>(LUT + 2 * Int(Extract(c.x, 1))), 1);
2657                 c.x = Insert(c.x, *Pointer<Short>(LUT + 2 * Int(Extract(c.x, 2))), 2);
2658                 c.x = Insert(c.x, *Pointer<Short>(LUT + 2 * Int(Extract(c.x, 3))), 3);
2659
2660                 c.y = Insert(c.y, *Pointer<Short>(LUT + 2 * Int(Extract(c.y, 0))), 0);
2661                 c.y = Insert(c.y, *Pointer<Short>(LUT + 2 * Int(Extract(c.y, 1))), 1);
2662                 c.y = Insert(c.y, *Pointer<Short>(LUT + 2 * Int(Extract(c.y, 2))), 2);
2663                 c.y = Insert(c.y, *Pointer<Short>(LUT + 2 * Int(Extract(c.y, 3))), 3);
2664
2665                 c.z = Insert(c.z, *Pointer<Short>(LUT + 2 * Int(Extract(c.z, 0))), 0);
2666                 c.z = Insert(c.z, *Pointer<Short>(LUT + 2 * Int(Extract(c.z, 1))), 1);
2667                 c.z = Insert(c.z, *Pointer<Short>(LUT + 2 * Int(Extract(c.z, 2))), 2);
2668                 c.z = Insert(c.z, *Pointer<Short>(LUT + 2 * Int(Extract(c.z, 3))), 3);
2669         }
2670
2671         Float4 PixelRoutine::sRGBtoLinear(const Float4 &x)   // Approximates x^2.2
2672         {
2673                 Float4 linear = x * x;
2674                 linear = linear * Float4(0.73f) + linear * x * Float4(0.27f);
2675
2676                 return Min(Max(linear, Float4(0.0f)), Float4(1.0f));
2677         }
2678
2679         bool PixelRoutine::colorUsed()
2680         {
2681                 return state.colorWriteMask || state.alphaTestActive() || state.shaderContainsKill;
2682         }
2683 }