src/third_party/angle/src/libGLESv2/DynamicHLSL.cpp

   1 //
   2 // Copyright (c) 2014 The ANGLE Project Authors. All rights reserved.
   3 // Use of this source code is governed by a BSD-style license that can be
   4 // found in the LICENSE file.
   5 //
   6 // DynamicHLSL.cpp: Implementation for link and run-time HLSL generation
   7 //
   8
   9 #include "precompiled.h"
  10
  11 #include "libGLESv2/DynamicHLSL.h"
  12 #include "libGLESv2/Shader.h"
  13 #include "libGLESv2/Program.h"
  14 #include "libGLESv2/renderer/Renderer.h"
  15 #include "common/utilities.h"
  16 #include "libGLESv2/ProgramBinary.h"
  17 #include "libGLESv2/formatutils.h"
  18 #include "common/blocklayout.h"
  19
  20 // For use with ArrayString, see angleutils.h
  21 META_ASSERT(GL_INVALID_INDEX == UINT_MAX);
  22
  23 namespace gl_d3d
  24 {
  25
  26 std::string HLSLComponentTypeString(GLenum componentType)
  27 {
  28     switch (componentType)
  29     {
  30       case GL_UNSIGNED_INT:         return "uint";
  31       case GL_INT:                  return "int";
  32       case GL_UNSIGNED_NORMALIZED:
  33       case GL_SIGNED_NORMALIZED:
  34       case GL_FLOAT:                return "float";
  35       default: UNREACHABLE();       return "not-component-type";
  36     }
  37 }
  38
  39 std::string HLSLComponentTypeString(GLenum componentType, int componentCount)
  40 {
  41     return HLSLComponentTypeString(componentType) + (componentCount > 1 ? Str(componentCount) : "");
  42 }
  43
  44 std::string HLSLMatrixTypeString(GLenum type)
  45 {
  46     switch (type)
  47     {
  48       case GL_FLOAT_MAT2:     return "float2x2";
  49       case GL_FLOAT_MAT3:     return "float3x3";
  50       case GL_FLOAT_MAT4:     return "float4x4";
  51       case GL_FLOAT_MAT2x3:   return "float2x3";
  52       case GL_FLOAT_MAT3x2:   return "float3x2";
  53       case GL_FLOAT_MAT2x4:   return "float2x4";
  54       case GL_FLOAT_MAT4x2:   return "float4x2";
  55       case GL_FLOAT_MAT3x4:   return "float3x4";
  56       case GL_FLOAT_MAT4x3:   return "float4x3";
  57       default: UNREACHABLE(); return "not-matrix-type";
  58     }
  59 }
  60
  61 std::string HLSLTypeString(GLenum type)
  62 {
  63     if (gl::IsMatrixType(type))
  64     {
  65         return HLSLMatrixTypeString(type);
  66     }
  67
  68     return HLSLComponentTypeString(gl::VariableComponentType(type), gl::VariableComponentCount(type));
  69 }
  70
  71 }
  72
  73 namespace gl
  74 {
  75
  76 const std::string VERTEX_ATTRIBUTE_STUB_STRING = "@@ VERTEX ATTRIBUTES @@";
  77 const std::string PIXEL_OUTPUT_STUB_STRING = "@@ PIXEL OUTPUT @@";
  78
  79 DynamicHLSL::DynamicHLSL(rx::Renderer *const renderer)
  80     : mRenderer(renderer)
  81 {
  82 }
  83
  84 static bool packVarying(PackedVarying *varying, const int maxVaryingVectors, VaryingPacking packing)
  85 {
  86     GLenum transposedType = TransposeMatrixType(varying->type);
  87
  88     // matrices within varying structs are not transposed
  89     int registers = (varying->isStruct() ? HLSLVariableRegisterCount(*varying) : VariableRowCount(transposedType)) * varying->elementCount();
  90     int elements = (varying->isStruct() ? 4 : VariableColumnCount(transposedType));
  91
  92     if (elements >= 2 && elements <= 4)
  93     {
  94         for (int r = 0; r <= maxVaryingVectors - registers; r++)
  95         {
  96             bool available = true;
  97
  98             for (int y = 0; y < registers && available; y++)
  99             {
 100                 for (int x = 0; x < elements && available; x++)
 101                 {
 102                     if (packing[r + y][x])
 103                     {
 104                         available = false;
 105                     }
 106                 }
 107             }
 108
 109             if (available)
 110             {
 111                 varying->registerIndex = r;
 112
 113                 for (int y = 0; y < registers; y++)
 114                 {
 115                     for (int x = 0; x < elements; x++)
 116                     {
 117                         packing[r + y][x] = &*varying;
 118                     }
 119                 }
 120
 121                 return true;
 122             }
 123         }
 124
 125         if (elements == 2)
 126         {
 127             for (int r = maxVaryingVectors - registers; r >= 0; r--)
 128             {
 129                 bool available = true;
 130
 131                 for (int y = 0; y < registers && available; y++)
 132                 {
 133                     for (int x = 2; x < 4 && available; x++)
 134                     {
 135                         if (packing[r + y][x])
 136                         {
 137                             available = false;
 138                         }
 139                     }
 140                 }
 141
 142                 if (available)
 143                 {
 144                     varying->registerIndex = r;
 145
 146                     for (int y = 0; y < registers; y++)
 147                     {
 148                         for (int x = 2; x < 4; x++)
 149                         {
 150                             packing[r + y][x] = &*varying;
 151                         }
 152                     }
 153
 154                     return true;
 155                 }
 156             }
 157         }
 158     }
 159     else if (elements == 1)
 160     {
 161         int space[4] = { 0 };
 162
 163         for (int y = 0; y < maxVaryingVectors; y++)
 164         {
 165             for (int x = 0; x < 4; x++)
 166             {
 167                 space[x] += packing[y][x] ? 0 : 1;
 168             }
 169         }
 170
 171         int column = 0;
 172
 173         for (int x = 0; x < 4; x++)
 174         {
 175             if (space[x] >= registers && space[x] < space[column])
 176             {
 177                 column = x;
 178             }
 179         }
 180
 181         if (space[column] >= registers)
 182         {
 183             for (int r = 0; r < maxVaryingVectors; r++)
 184             {
 185                 if (!packing[r][column])
 186                 {
 187                     varying->registerIndex = r;
 188
 189                     for (int y = r; y < r + registers; y++)
 190                     {
 191                         packing[y][column] = &*varying;
 192                     }
 193
 194                     break;
 195                 }
 196             }
 197
 198             return true;
 199         }
 200     }
 201     else UNREACHABLE();
 202
 203     return false;
 204 }
 205
 206 // Packs varyings into generic varying registers, using the algorithm from [OpenGL ES Shading Language 1.00 rev. 17] appendix A section 7 page 111
 207 // Returns the number of used varying registers, or -1 if unsuccesful
 208 int DynamicHLSL::packVaryings(InfoLog &infoLog, VaryingPacking packing, FragmentShader *fragmentShader,
 209                               VertexShader *vertexShader, const std::vector<std::string>& transformFeedbackVaryings)
 210 {
 211     const int maxVaryingVectors = mRenderer->getMaxVaryingVectors();
 212
 213     vertexShader->resetVaryingsRegisterAssignment();
 214     fragmentShader->resetVaryingsRegisterAssignment();
 215
 216     std::set<std::string> packedVaryings;
 217
 218     for (unsigned int varyingIndex = 0; varyingIndex < fragmentShader->mVaryings.size(); varyingIndex++)
 219     {
 220         PackedVarying *varying = &fragmentShader->mVaryings[varyingIndex];
 221         if (packVarying(varying, maxVaryingVectors, packing))
 222         {
 223             packedVaryings.insert(varying->name);
 224         }
 225         else
 226         {
 227             infoLog.append("Could not pack varying %s", varying->name.c_str());
 228             return -1;
 229         }
 230     }
 231
 232     for (unsigned int feedbackVaryingIndex = 0; feedbackVaryingIndex < transformFeedbackVaryings.size(); feedbackVaryingIndex++)
 233     {
 234         const std::string &transformFeedbackVarying = transformFeedbackVaryings[feedbackVaryingIndex];
 235         if (packedVaryings.find(transformFeedbackVarying) == packedVaryings.end())
 236         {
 237             bool found = false;
 238             for (unsigned int varyingIndex = 0; varyingIndex < vertexShader->mVaryings.size(); varyingIndex++)
 239             {
 240                 PackedVarying *varying = &vertexShader->mVaryings[varyingIndex];
 241                 if (transformFeedbackVarying == varying->name)
 242                 {
 243                     if (!packVarying(varying, maxVaryingVectors, packing))
 244                     {
 245                         infoLog.append("Could not pack varying %s", varying->name.c_str());
 246                         return -1;
 247                     }
 248
 249                     found = true;
 250                     break;
 251                 }
 252             }
 253
 254             if (!found && transformFeedbackVarying != "gl_Position" && transformFeedbackVarying != "gl_PointSize")
 255             {
 256                 infoLog.append("Transform feedback varying %s does not exist in the vertex shader.", transformFeedbackVarying.c_str());
 257                 return -1;
 258             }
 259         }
 260     }
 261
 262     // Return the number of used registers
 263     int registers = 0;
 264
 265     for (int r = 0; r < maxVaryingVectors; r++)
 266     {
 267         if (packing[r][0] || packing[r][1] || packing[r][2] || packing[r][3])
 268         {
 269             registers++;
 270         }
 271     }
 272
 273     return registers;
 274 }
 275
 276 std::string DynamicHLSL::generateVaryingHLSL(VertexShader *shader) const
 277 {
 278     std::string varyingSemantic = getVaryingSemantic(shader->mUsesPointSize);
 279     std::string varyingHLSL;
 280
 281     for (unsigned int varyingIndex = 0; varyingIndex < shader->mVaryings.size(); varyingIndex++)
 282     {
 283         const PackedVarying &varying = shader->mVaryings[varyingIndex];
 284         if (varying.registerAssigned())
 285         {
 286             GLenum transposedType = TransposeMatrixType(varying.type);
 287             int variableRows = (varying.isStruct() ? 1 : VariableRowCount(transposedType));
 288
 289             for (unsigned int elementIndex = 0; elementIndex < varying.elementCount(); elementIndex++)
 290             {
 291                 for (int row = 0; row < variableRows; row++)
 292                 {
 293                     switch (varying.interpolation)
 294                     {
 295                       case sh::INTERPOLATION_SMOOTH:   varyingHLSL += "    ";                 break;
 296                       case sh::INTERPOLATION_FLAT:     varyingHLSL += "    nointerpolation "; break;
 297                       case sh::INTERPOLATION_CENTROID: varyingHLSL += "    centroid ";        break;
 298                       default:  UNREACHABLE();
 299                     }
 300
 301                     unsigned int semanticIndex = elementIndex * variableRows + varying.registerIndex + row;
 302                     std::string n = Str(semanticIndex);
 303
 304                     std::string typeString;
 305
 306                     if (varying.isStruct())
 307                     {
 308                         // matrices within structs are not transposed, so
 309                         // do not use the special struct prefix "rm"
 310                         typeString = decorateVariable(varying.structName);
 311                     }
 312                     else
 313                     {
 314                         GLenum componentType = VariableComponentType(transposedType);
 315                         int columnCount = VariableColumnCount(transposedType);
 316                         typeString = gl_d3d::HLSLComponentTypeString(componentType, columnCount);
 317                     }
 318                     varyingHLSL += typeString + " v" + n + " : " + varyingSemantic + n + ";\n";
 319                 }
 320             }
 321         }
 322     }
 323
 324     return varyingHLSL;
 325 }
 326
 327 std::string DynamicHLSL::generateVertexShaderForInputLayout(const std::string &sourceShader,
 328                                                             const VertexFormat inputLayout[],
 329                                                             const sh::Attribute shaderAttributes[]) const
 330 {
 331     std::string structHLSL, initHLSL;
 332
 333     int semanticIndex = 0;
 334     unsigned int inputIndex = 0;
 335
 336     for (unsigned int attributeIndex = 0; attributeIndex < MAX_VERTEX_ATTRIBS; attributeIndex++)
 337     {
 338         ASSERT(inputIndex < MAX_VERTEX_ATTRIBS);
 339
 340         const VertexFormat &vertexFormat = inputLayout[inputIndex];
 341         const sh::Attribute &shaderAttribute = shaderAttributes[attributeIndex];
 342
 343         if (!shaderAttribute.name.empty())
 344         {
 345             // HLSL code for input structure
 346             if (IsMatrixType(shaderAttribute.type))
 347             {
 348                 // Matrix types are always transposed
 349                 structHLSL += "    " + gl_d3d::HLSLMatrixTypeString(TransposeMatrixType(shaderAttribute.type));
 350             }
 351             else
 352             {
 353                 GLenum componentType = mRenderer->getVertexComponentType(vertexFormat);
 354                 structHLSL += "    " + gl_d3d::HLSLComponentTypeString(componentType, VariableComponentCount(shaderAttribute.type));
 355             }
 356
 357             structHLSL += " " + decorateVariable(shaderAttribute.name) + " : TEXCOORD" + Str(semanticIndex) + ";\n";
 358             semanticIndex += VariableRegisterCount(shaderAttribute.type);
 359
 360             // HLSL code for initialization
 361             initHLSL += "    " + decorateVariable(shaderAttribute.name) + " = ";
 362
 363             // Mismatched vertex attribute to vertex input may result in an undefined
 364             // data reinterpretation (eg for pure integer->float, float->pure integer)
 365             // TODO: issue warning with gl debug info extension, when supported
 366             if (IsMatrixType(shaderAttribute.type) ||
 367                 (mRenderer->getVertexConversionType(vertexFormat) & rx::VERTEX_CONVERT_GPU) != 0)
 368             {
 369                 initHLSL += generateAttributeConversionHLSL(vertexFormat, shaderAttribute);
 370             }
 371             else
 372             {
 373                 initHLSL += "input." + decorateVariable(shaderAttribute.name);
 374             }
 375
 376             initHLSL += ";\n";
 377
 378             inputIndex += VariableRowCount(TransposeMatrixType(shaderAttribute.type));
 379         }
 380     }
 381
 382     std::string replacementHLSL = "struct VS_INPUT\n"
 383                                   "{\n" +
 384                                   structHLSL +
 385                                   "};\n"
 386                                   "\n"
 387                                   "void initAttributes(VS_INPUT input)\n"
 388                                   "{\n" +
 389                                   initHLSL +
 390                                   "}\n";
 391
 392     std::string vertexHLSL(sourceShader);
 393
 394     size_t copyInsertionPos = vertexHLSL.find(VERTEX_ATTRIBUTE_STUB_STRING);
 395     vertexHLSL.replace(copyInsertionPos, VERTEX_ATTRIBUTE_STUB_STRING.length(), replacementHLSL);
 396
 397     return vertexHLSL;
 398 }
 399
 400 std::string DynamicHLSL::generatePixelShaderForOutputSignature(const std::string &sourceShader, const std::vector<PixelShaderOuputVariable> &outputVariables,
 401                                                                bool usesFragDepth, const std::vector<GLenum> &outputLayout) const
 402 {
 403     const int shaderModel = mRenderer->getMajorShaderModel();
 404     std::string targetSemantic = (shaderModel >= 4) ? "SV_TARGET" : "COLOR";
 405     std::string depthSemantic = (shaderModel >= 4) ? "SV_Depth" : "DEPTH";
 406
 407     std::string declarationHLSL;
 408     std::string copyHLSL;
 409     for (size_t i = 0; i < outputVariables.size(); i++)
 410     {
 411         const PixelShaderOuputVariable& outputVariable = outputVariables[i];
 412         ASSERT(outputLayout.size() > outputVariable.outputIndex);
 413
 414         // FIXME(geofflang): Work around NVIDIA driver bug by repacking buffers
 415         bool outputIndexEnabled = true; // outputLayout[outputVariable.outputIndex] != GL_NONE
 416         if (outputIndexEnabled)
 417         {
 418             declarationHLSL += "    " + gl_d3d::HLSLTypeString(outputVariable.type) + " " + outputVariable.name +
 419                                " : " + targetSemantic + Str(outputVariable.outputIndex) + ";\n";
 420
 421             copyHLSL += "    output." + outputVariable.name + " = " + outputVariable.source + ";\n";
 422         }
 423     }
 424
 425     if (usesFragDepth)
 426     {
 427         declarationHLSL += "    float gl_Depth : " + depthSemantic + ";\n";
 428         copyHLSL += "    output.gl_Depth = gl_Depth; \n";
 429     }
 430
 431     std::string replacementHLSL = "struct PS_OUTPUT\n"
 432                                   "{\n" +
 433                                   declarationHLSL +
 434                                   "};\n"
 435                                   "\n"
 436                                   "PS_OUTPUT generateOutput()\n"
 437                                   "{\n"
 438                                   "    PS_OUTPUT output;\n" +
 439                                   copyHLSL +
 440                                   "    return output;\n"
 441                                   "}\n";
 442
 443     std::string pixelHLSL(sourceShader);
 444
 445     size_t outputInsertionPos = pixelHLSL.find(PIXEL_OUTPUT_STUB_STRING);
 446     pixelHLSL.replace(outputInsertionPos, PIXEL_OUTPUT_STUB_STRING.length(), replacementHLSL);
 447
 448     return pixelHLSL;
 449 }
 450
 451 std::string DynamicHLSL::getVaryingSemantic(bool pointSize) const
 452 {
 453     // SM3 reserves the TEXCOORD semantic for point sprite texcoords (gl_PointCoord)
 454     // In D3D11 we manually compute gl_PointCoord in the GS.
 455     int shaderModel = mRenderer->getMajorShaderModel();
 456     return ((pointSize && shaderModel < 4) ? "COLOR" : "TEXCOORD");
 457 }
 458
 459 struct DynamicHLSL::SemanticInfo
 460 {
 461     struct BuiltinInfo
 462     {
 463         BuiltinInfo()
 464             : enabled(false),
 465               index(0),
 466               systemValue(false)
 467         {}
 468
 469         bool enabled;
 470         std::string semantic;
 471         unsigned int index;
 472         bool systemValue;
 473
 474         std::string str() const
 475         {
 476             return (systemValue ? semantic : (semantic + Str(index)));
 477         }
 478
 479         void enableSystem(const std::string &systemValueSemantic)
 480         {
 481             enabled = true;
 482             semantic = systemValueSemantic;
 483             systemValue = true;
 484         }
 485
 486         void enable(const std::string &semanticVal, unsigned int indexVal)
 487         {
 488             enabled = true;
 489             semantic = semanticVal;
 490             index = indexVal;
 491         }
 492     };
 493
 494     BuiltinInfo dxPosition;
 495     BuiltinInfo glPosition;
 496     BuiltinInfo glFragCoord;
 497     BuiltinInfo glPointCoord;
 498     BuiltinInfo glPointSize;
 499 };
 500
 501 DynamicHLSL::SemanticInfo DynamicHLSL::getSemanticInfo(int startRegisters, bool fragCoord, bool pointCoord,
 502                                                        bool pointSize, bool pixelShader) const
 503 {
 504     SemanticInfo info;
 505     bool hlsl4 = (mRenderer->getMajorShaderModel() >= 4);
 506     const std::string &varyingSemantic = getVaryingSemantic(pointSize);
 507
 508     int reservedRegisterIndex = startRegisters;
 509
 510     if (hlsl4)
 511     {
 512         info.dxPosition.enableSystem("SV_Position");
 513     }
 514     else if (pixelShader)
 515     {
 516         info.dxPosition.enableSystem("VPOS");
 517     }
 518     else
 519     {
 520         info.dxPosition.enableSystem("POSITION");
 521     }
 522
 523     info.glPosition.enable(varyingSemantic, reservedRegisterIndex++);
 524
 525     if (fragCoord)
 526     {
 527         info.glFragCoord.enable(varyingSemantic, reservedRegisterIndex++);
 528     }
 529
 530     if (pointCoord)
 531     {
 532         // SM3 reserves the TEXCOORD semantic for point sprite texcoords (gl_PointCoord)
 533         // In D3D11 we manually compute gl_PointCoord in the GS.
 534         if (hlsl4)
 535         {
 536             info.glPointCoord.enable(varyingSemantic, reservedRegisterIndex++);
 537         }
 538         else
 539         {
 540             info.glPointCoord.enable("TEXCOORD", 0);
 541         }
 542     }
 543
 544     // Special case: do not include PSIZE semantic in HLSL 3 pixel shaders
 545     if (pointSize && (!pixelShader || hlsl4))
 546     {
 547         info.glPointSize.enableSystem("PSIZE");
 548     }
 549
 550     return info;
 551 }
 552
 553 std::string DynamicHLSL::generateVaryingLinkHLSL(const SemanticInfo &info, const std::string &varyingHLSL) const
 554 {
 555     std::string linkHLSL = "{\n";
 556
 557     ASSERT(info.dxPosition.enabled && info.glPosition.enabled);
 558
 559     linkHLSL += "    float4 dx_Position : " + info.dxPosition.str() + ";\n";
 560     linkHLSL += "    float4 gl_Position : " + info.glPosition.str() + ";\n";
 561
 562     if (info.glFragCoord.enabled)
 563     {
 564         linkHLSL += "    float4 gl_FragCoord : " + info.glFragCoord.str() + ";\n";
 565     }
 566
 567     if (info.glPointCoord.enabled)
 568     {
 569         linkHLSL += "    float2 gl_PointCoord : " + info.glPointCoord.str() + ";\n";
 570     }
 571
 572     linkHLSL += varyingHLSL;
 573
 574     if (info.glPointSize.enabled)
 575     {
 576         linkHLSL += "    float gl_PointSize : " + info.glPointSize.str() + ";\n";
 577     }
 578
 579     linkHLSL += "};\n";
 580
 581     return linkHLSL;
 582 }
 583
 584 void DynamicHLSL::storeBuiltinLinkedVaryings(const SemanticInfo &info,
 585                                              std::vector<LinkedVarying> *linkedVaryings) const
 586 {
 587     ASSERT(info.glPosition.enabled);
 588
 589     linkedVaryings->push_back(LinkedVarying("gl_Position", GL_FLOAT_VEC4, 1, info.glPosition.semantic,
 590                                             info.glPosition.index, 1));
 591
 592     if (info.glFragCoord.enabled)
 593     {
 594         linkedVaryings->push_back(LinkedVarying("gl_FragCoord", GL_FLOAT_VEC4, 1, info.glFragCoord.semantic,
 595                                                 info.glFragCoord.index, 1));
 596     }
 597
 598     if (info.glPointSize.enabled)
 599     {
 600         linkedVaryings->push_back(LinkedVarying("gl_PointSize", GL_FLOAT, 1, "PSIZE", 0, 1));
 601     }
 602 }
 603
 604 void DynamicHLSL::storeUserLinkedVaryings(const VertexShader *vertexShader,
 605                                           std::vector<LinkedVarying> *linkedVaryings) const
 606 {
 607     const std::string &varyingSemantic = getVaryingSemantic(vertexShader->mUsesPointSize);
 608     const std::vector<PackedVarying> &varyings = vertexShader->mVaryings;
 609
 610     for (unsigned int varyingIndex = 0; varyingIndex < varyings.size(); varyingIndex++)
 611     {
 612         const PackedVarying &varying = varyings[varyingIndex];
 613         if (varying.registerAssigned())
 614         {
 615             GLenum transposedType = TransposeMatrixType(varying.type);
 616             int variableRows = (varying.isStruct() ? 1 : VariableRowCount(transposedType));
 617
 618             linkedVaryings->push_back(LinkedVarying(varying.name, varying.type, varying.elementCount(),
 619                                                     varyingSemantic, varying.registerIndex,
 620                                                     variableRows * varying.elementCount()));
 621         }
 622     }
 623 }
 624
 625 bool DynamicHLSL::generateShaderLinkHLSL(InfoLog &infoLog, int registers, const VaryingPacking packing,
 626                                          std::string& pixelHLSL, std::string& vertexHLSL,
 627                                          FragmentShader *fragmentShader, VertexShader *vertexShader,
 628                                          const std::vector<std::string>& transformFeedbackVaryings,
 629                                          std::vector<LinkedVarying> *linkedVaryings,
 630                                          std::map<int, VariableLocation> *programOutputVars,
 631                                          std::vector<PixelShaderOuputVariable> *outPixelShaderKey,
 632                                          bool *outUsesFragDepth) const
 633 {
 634     if (pixelHLSL.empty() || vertexHLSL.empty())
 635     {
 636         return false;
 637     }
 638
 639     bool usesMRT = fragmentShader->mUsesMultipleRenderTargets;
 640     bool usesFragColor = fragmentShader->mUsesFragColor;
 641     bool usesFragData = fragmentShader->mUsesFragData;
 642     bool usesFragCoord = fragmentShader->mUsesFragCoord;
 643     bool usesPointCoord = fragmentShader->mUsesPointCoord;
 644     bool usesPointSize = vertexShader->mUsesPointSize;
 645
 646     if (usesFragColor && usesFragData)
 647     {
 648         infoLog.append("Cannot use both gl_FragColor and gl_FragData in the same fragment shader.");
 649         return false;
 650     }
 651
 652     // Write the HLSL input/output declarations
 653     const int shaderModel = mRenderer->getMajorShaderModel();
 654     const int maxVaryingVectors = mRenderer->getMaxVaryingVectors();
 655
 656     const int registersNeeded = registers + (usesFragCoord ? 1 : 0) + (usesPointCoord ? 1 : 0);
 657
 658     // Two cases when writing to gl_FragColor and using ESSL 1.0:
 659     // - with a 3.0 context, the output color is copied to channel 0
 660     // - with a 2.0 context, the output color is broadcast to all channels
 661     const bool broadcast = (fragmentShader->mUsesFragColor && mRenderer->getCurrentClientVersion() < 3);
 662     const unsigned int numRenderTargets = (broadcast || usesMRT ? mRenderer->getRendererCaps().maxDrawBuffers : 1);
 663
 664     int shaderVersion = vertexShader->getShaderVersion();
 665
 666     if (registersNeeded > maxVaryingVectors)
 667     {
 668         infoLog.append("No varying registers left to support gl_FragCoord/gl_PointCoord");
 669         return false;
 670     }
 671
 672     const std::string &varyingHLSL = generateVaryingHLSL(vertexShader);
 673     const SemanticInfo &vertexSemantics = getSemanticInfo(registers, usesFragCoord,
 674                                                           false, usesPointSize, false);
 675
 676     storeUserLinkedVaryings(vertexShader, linkedVaryings);
 677     storeBuiltinLinkedVaryings(vertexSemantics, linkedVaryings);
 678
 679     // Add stub string to be replaced when shader is dynamically defined by its layout
 680     vertexHLSL += "\n" + VERTEX_ATTRIBUTE_STUB_STRING + "\n"
 681                   "struct VS_OUTPUT\n" + generateVaryingLinkHLSL(vertexSemantics, varyingHLSL) + "\n"
 682                   "VS_OUTPUT main(VS_INPUT input)\n"
 683                   "{\n"
 684                   "    initAttributes(input);\n";
 685
 686     if (shaderModel >= 4)
 687     {
 688         vertexHLSL += "\n"
 689                       "    gl_main();\n"
 690                       "\n"
 691                       "    VS_OUTPUT output;\n"
 692                       "    output.gl_Position = gl_Position;\n"
 693                       "    output.dx_Position.x = gl_Position.x;\n"
 694                       "    output.dx_Position.y = -gl_Position.y;\n"
 695                       "    output.dx_Position.z = (gl_Position.z + gl_Position.w) * 0.5;\n"
 696                       "    output.dx_Position.w = gl_Position.w;\n";
 697     }
 698     else
 699     {
 700         vertexHLSL += "\n"
 701                       "    gl_main();\n"
 702                       "\n"
 703                       "    VS_OUTPUT output;\n"
 704                       "    output.gl_Position = gl_Position;\n"
 705                       "    output.dx_Position.x = gl_Position.x * dx_ViewAdjust.z + dx_ViewAdjust.x * gl_Position.w;\n"
 706                       "    output.dx_Position.y = -(gl_Position.y * dx_ViewAdjust.w + dx_ViewAdjust.y * gl_Position.w);\n"
 707                       "    output.dx_Position.z = (gl_Position.z + gl_Position.w) * 0.5;\n"
 708                       "    output.dx_Position.w = gl_Position.w;\n";
 709     }
 710
 711     if (usesPointSize && shaderModel >= 3)
 712     {
 713         vertexHLSL += "    output.gl_PointSize = gl_PointSize;\n";
 714     }
 715
 716     if (usesFragCoord)
 717     {
 718         vertexHLSL += "    output.gl_FragCoord = gl_Position;\n";
 719     }
 720
 721     for (unsigned int vertVaryingIndex = 0; vertVaryingIndex < vertexShader->mVaryings.size(); vertVaryingIndex++)
 722     {
 723         const PackedVarying &varying = vertexShader->mVaryings[vertVaryingIndex];
 724         if (varying.registerAssigned())
 725         {
 726             for (unsigned int elementIndex = 0; elementIndex < varying.elementCount(); elementIndex++)
 727             {
 728                 int variableRows = (varying.isStruct() ? 1 : VariableRowCount(TransposeMatrixType(varying.type)));
 729
 730                 for (int row = 0; row < variableRows; row++)
 731                 {
 732                     int r = varying.registerIndex + elementIndex * variableRows + row;
 733                     vertexHLSL += "    output.v" + Str(r);
 734
 735                     bool sharedRegister = false;   // Register used by multiple varyings
 736
 737                     for (int x = 0; x < 4; x++)
 738                     {
 739                         if (packing[r][x] && packing[r][x] != packing[r][0])
 740                         {
 741                             sharedRegister = true;
 742                             break;
 743                         }
 744                     }
 745
 746                     if(sharedRegister)
 747                     {
 748                         vertexHLSL += ".";
 749
 750                         for (int x = 0; x < 4; x++)
 751                         {
 752                             if (packing[r][x] == &varying)
 753                             {
 754                                 switch(x)
 755                                 {
 756                                   case 0: vertexHLSL += "x"; break;
 757                                   case 1: vertexHLSL += "y"; break;
 758                                   case 2: vertexHLSL += "z"; break;
 759                                   case 3: vertexHLSL += "w"; break;
 760                                 }
 761                             }
 762                         }
 763                     }
 764
 765                     vertexHLSL += " = _" + varying.name;
 766
 767                     if (varying.isArray())
 768                     {
 769                         vertexHLSL += ArrayString(elementIndex);
 770                     }
 771
 772                     if (variableRows > 1)
 773                     {
 774                         vertexHLSL += ArrayString(row);
 775                     }
 776
 777                     vertexHLSL += ";\n";
 778                 }
 779             }
 780         }
 781     }
 782
 783     vertexHLSL += "\n"
 784                   "    return output;\n"
 785                   "}\n";
 786
 787     const SemanticInfo &pixelSemantics = getSemanticInfo(registers, usesFragCoord, usesPointCoord,
 788                                                          usesPointSize, true);
 789
 790     pixelHLSL += "struct PS_INPUT\n" + generateVaryingLinkHLSL(pixelSemantics, varyingHLSL) + "\n";
 791
 792     if (shaderVersion < 300)
 793     {
 794         for (unsigned int renderTargetIndex = 0; renderTargetIndex < numRenderTargets; renderTargetIndex++)
 795         {
 796             PixelShaderOuputVariable outputKeyVariable;
 797             outputKeyVariable.type = GL_FLOAT_VEC4;
 798             outputKeyVariable.name = "gl_Color" + Str(renderTargetIndex);
 799             outputKeyVariable.source = broadcast ? "gl_Color[0]" : "gl_Color[" + Str(renderTargetIndex) + "]";
 800             outputKeyVariable.outputIndex = renderTargetIndex;
 801
 802             outPixelShaderKey->push_back(outputKeyVariable);
 803         }
 804
 805         *outUsesFragDepth = fragmentShader->mUsesFragDepth;
 806     }
 807     else
 808     {
 809         defineOutputVariables(fragmentShader, programOutputVars);
 810
 811         const std::vector<sh::Attribute> &shaderOutputVars = fragmentShader->getOutputVariables();
 812         for (auto locationIt = programOutputVars->begin(); locationIt != programOutputVars->end(); locationIt++)
 813         {
 814             const VariableLocation &outputLocation = locationIt->second;
 815             const sh::ShaderVariable &outputVariable = shaderOutputVars[outputLocation.index];
 816             const std::string &variableName = "out_" + outputLocation.name;
 817             const std::string &elementString = (outputLocation.element == GL_INVALID_INDEX ? "" : Str(outputLocation.element));
 818
 819             PixelShaderOuputVariable outputKeyVariable;
 820             outputKeyVariable.type = outputVariable.type;
 821             outputKeyVariable.name = variableName + elementString;
 822             outputKeyVariable.source = variableName + ArrayString(outputLocation.element);
 823             outputKeyVariable.outputIndex = locationIt->first;
 824
 825             outPixelShaderKey->push_back(outputKeyVariable);
 826         }
 827
 828         *outUsesFragDepth = false;
 829     }
 830
 831     pixelHLSL += PIXEL_OUTPUT_STUB_STRING + "\n";
 832
 833     if (fragmentShader->mUsesFrontFacing)
 834     {
 835         if (shaderModel >= 4)
 836         {
 837             pixelHLSL += "PS_OUTPUT main(PS_INPUT input, bool isFrontFace : SV_IsFrontFace)\n"
 838                          "{\n";
 839         }
 840         else
 841         {
 842             pixelHLSL += "PS_OUTPUT main(PS_INPUT input, float vFace : VFACE)\n"
 843                          "{\n";
 844         }
 845     }
 846     else
 847     {
 848         pixelHLSL += "PS_OUTPUT main(PS_INPUT input)\n"
 849                      "{\n";
 850     }
 851
 852     if (usesFragCoord)
 853     {
 854         pixelHLSL += "    float rhw = 1.0 / input.gl_FragCoord.w;\n";
 855
 856         if (shaderModel >= 4)
 857         {
 858             pixelHLSL += "    gl_FragCoord.x = input.dx_Position.x;\n"
 859                          "    gl_FragCoord.y = input.dx_Position.y;\n";
 860         }
 861         else if (shaderModel >= 3)
 862         {
 863             pixelHLSL += "    gl_FragCoord.x = input.dx_Position.x + 0.5;\n"
 864                          "    gl_FragCoord.y = input.dx_Position.y + 0.5;\n";
 865         }
 866         else
 867         {
 868             // dx_ViewCoords contains the viewport width/2, height/2, center.x and center.y. See Renderer::setViewport()
 869             pixelHLSL += "    gl_FragCoord.x = (input.gl_FragCoord.x * rhw) * dx_ViewCoords.x + dx_ViewCoords.z;\n"
 870                          "    gl_FragCoord.y = (input.gl_FragCoord.y * rhw) * dx_ViewCoords.y + dx_ViewCoords.w;\n";
 871         }
 872
 873         pixelHLSL += "    gl_FragCoord.z = (input.gl_FragCoord.z * rhw) * dx_DepthFront.x + dx_DepthFront.y;\n"
 874                      "    gl_FragCoord.w = rhw;\n";
 875     }
 876
 877     if (usesPointCoord && shaderModel >= 3)
 878     {
 879         pixelHLSL += "    gl_PointCoord.x = input.gl_PointCoord.x;\n";
 880         pixelHLSL += "    gl_PointCoord.y = 1.0 - input.gl_PointCoord.y;\n";
 881     }
 882
 883     if (fragmentShader->mUsesFrontFacing)
 884     {
 885         if (shaderModel <= 3)
 886         {
 887             pixelHLSL += "    gl_FrontFacing = (vFace * dx_DepthFront.z >= 0.0);\n";
 888         }
 889         else
 890         {
 891             pixelHLSL += "    gl_FrontFacing = isFrontFace;\n";
 892         }
 893     }
 894
 895     for (unsigned int varyingIndex = 0; varyingIndex < fragmentShader->mVaryings.size(); varyingIndex++)
 896     {
 897         const PackedVarying &varying = fragmentShader->mVaryings[varyingIndex];
 898         if (varying.registerAssigned())
 899         {
 900             for (unsigned int elementIndex = 0; elementIndex < varying.elementCount(); elementIndex++)
 901             {
 902                 GLenum transposedType = TransposeMatrixType(varying.type);
 903                 int variableRows = (varying.isStruct() ? 1 : VariableRowCount(transposedType));
 904                 for (int row = 0; row < variableRows; row++)
 905                 {
 906                     std::string n = Str(varying.registerIndex + elementIndex * variableRows + row);
 907                     pixelHLSL += "    _" + varying.name;
 908
 909                     if (varying.isArray())
 910                     {
 911                         pixelHLSL += ArrayString(elementIndex);
 912                     }
 913
 914                     if (variableRows > 1)
 915                     {
 916                         pixelHLSL += ArrayString(row);
 917                     }
 918
 919                     if (varying.isStruct())
 920                     {
 921                         pixelHLSL += " = input.v" + n + ";\n";   break;
 922                     }
 923                     else
 924                     {
 925                         switch (VariableColumnCount(transposedType))
 926                         {
 927                           case 1: pixelHLSL += " = input.v" + n + ".x;\n";   break;
 928                           case 2: pixelHLSL += " = input.v" + n + ".xy;\n";  break;
 929                           case 3: pixelHLSL += " = input.v" + n + ".xyz;\n"; break;
 930                           case 4: pixelHLSL += " = input.v" + n + ";\n";     break;
 931                           default: UNREACHABLE();
 932                         }
 933                     }
 934                 }
 935             }
 936         }
 937         else UNREACHABLE();
 938     }
 939
 940     pixelHLSL += "\n"
 941                  "    gl_main();\n"
 942                  "\n"
 943                  "    return generateOutput();\n"
 944                  "}\n";
 945
 946     return true;
 947 }
 948
 949 void DynamicHLSL::defineOutputVariables(FragmentShader *fragmentShader, std::map<int, VariableLocation> *programOutputVars) const
 950 {
 951     const std::vector<sh::Attribute> &shaderOutputVars = fragmentShader->getOutputVariables();
 952
 953     for (unsigned int outputVariableIndex = 0; outputVariableIndex < shaderOutputVars.size(); outputVariableIndex++)
 954     {
 955         const sh::Attribute &outputVariable = shaderOutputVars[outputVariableIndex];
 956         const int baseLocation = outputVariable.location == -1 ? 0 : outputVariable.location;
 957
 958         if (outputVariable.arraySize > 0)
 959         {
 960             for (unsigned int elementIndex = 0; elementIndex < outputVariable.arraySize; elementIndex++)
 961             {
 962                 const int location = baseLocation + elementIndex;
 963                 ASSERT(programOutputVars->count(location) == 0);
 964                 (*programOutputVars)[location] = VariableLocation(outputVariable.name, elementIndex, outputVariableIndex);
 965             }
 966         }
 967         else
 968         {
 969             ASSERT(programOutputVars->count(baseLocation) == 0);
 970             (*programOutputVars)[baseLocation] = VariableLocation(outputVariable.name, GL_INVALID_INDEX, outputVariableIndex);
 971         }
 972     }
 973 }
 974
 975 std::string DynamicHLSL::generateGeometryShaderHLSL(int registers, FragmentShader *fragmentShader, VertexShader *vertexShader) const
 976 {
 977     // for now we only handle point sprite emulation
 978     ASSERT(vertexShader->mUsesPointSize && mRenderer->getMajorShaderModel() >= 4);
 979     return generatePointSpriteHLSL(registers, fragmentShader, vertexShader);
 980 }
 981
 982 std::string DynamicHLSL::generatePointSpriteHLSL(int registers, FragmentShader *fragmentShader, VertexShader *vertexShader) const
 983 {
 984     ASSERT(registers >= 0);
 985     ASSERT(vertexShader->mUsesPointSize);
 986     ASSERT(mRenderer->getMajorShaderModel() >= 4);
 987
 988     std::string geomHLSL;
 989
 990     const SemanticInfo &inSemantics = getSemanticInfo(registers, fragmentShader->mUsesFragCoord,
 991                                                       false, true, false);
 992     const SemanticInfo &outSemantics = getSemanticInfo(registers, fragmentShader->mUsesFragCoord,
 993                                                        fragmentShader->mUsesPointCoord, true, false);
 994
 995     std::string varyingHLSL = generateVaryingHLSL(vertexShader);
 996     std::string inLinkHLSL = generateVaryingLinkHLSL(inSemantics, varyingHLSL);
 997     std::string outLinkHLSL = generateVaryingLinkHLSL(outSemantics, varyingHLSL);
 998
 999     // TODO(geofflang): use context's caps
1000     geomHLSL += "uniform float4 dx_ViewCoords : register(c1);\n"
1001                 "\n"
1002                 "struct GS_INPUT\n" + inLinkHLSL + "\n" +
1003                 "struct GS_OUTPUT\n" + outLinkHLSL + "\n" +
1004                 "\n"
1005                   "static float2 pointSpriteCorners[] = \n"
1006                   "{\n"
1007                   "    float2( 0.5f, -0.5f),\n"
1008                   "    float2( 0.5f,  0.5f),\n"
1009                   "    float2(-0.5f, -0.5f),\n"
1010                   "    float2(-0.5f,  0.5f)\n"
1011                   "};\n"
1012                   "\n"
1013                   "static float2 pointSpriteTexcoords[] = \n"
1014                   "{\n"
1015                   "    float2(1.0f, 1.0f),\n"
1016                   "    float2(1.0f, 0.0f),\n"
1017                   "    float2(0.0f, 1.0f),\n"
1018                   "    float2(0.0f, 0.0f)\n"
1019                   "};\n"
1020                   "\n"
1021                   "static float minPointSize = " + Str(mRenderer->getRendererCaps().minAliasedPointSize) + ".0f;\n"
1022                   "static float maxPointSize = " + Str(mRenderer->getRendererCaps().maxAliasedPointSize) + ".0f;\n"
1023                   "\n"
1024                   "[maxvertexcount(4)]\n"
1025                   "void main(point GS_INPUT input[1], inout TriangleStream<GS_OUTPUT> outStream)\n"
1026                   "{\n"
1027                   "    GS_OUTPUT output = (GS_OUTPUT)0;\n"
1028                   "    output.gl_Position = input[0].gl_Position;\n";
1029                   "    output.gl_PointSize = input[0].gl_PointSize;\n";
1030
1031     for (int r = 0; r < registers; r++)
1032     {
1033         geomHLSL += "    output.v" + Str(r) + " = input[0].v" + Str(r) + ";\n";
1034     }
1035
1036     if (fragmentShader->mUsesFragCoord)
1037     {
1038         geomHLSL += "    output.gl_FragCoord = input[0].gl_FragCoord;\n";
1039     }
1040
1041     geomHLSL += "    \n"
1042                 "    float gl_PointSize = clamp(input[0].gl_PointSize, minPointSize, maxPointSize);\n"
1043                 "    float4 dx_Position = input[0].dx_Position;\n"
1044                 "    float2 viewportScale = float2(1.0f / dx_ViewCoords.x, 1.0f / dx_ViewCoords.y) * dx_Position.w;\n";
1045
1046     for (int corner = 0; corner < 4; corner++)
1047     {
1048         geomHLSL += "    \n"
1049                     "    output.dx_Position = dx_Position + float4(pointSpriteCorners[" + Str(corner) + "] * viewportScale * gl_PointSize, 0.0f, 0.0f);\n";
1050
1051         if (fragmentShader->mUsesPointCoord)
1052         {
1053             geomHLSL += "    output.gl_PointCoord = pointSpriteTexcoords[" + Str(corner) + "];\n";
1054         }
1055
1056         geomHLSL += "    outStream.Append(output);\n";
1057     }
1058
1059     geomHLSL += "    \n"
1060                 "    outStream.RestartStrip();\n"
1061                 "}\n";
1062
1063     return geomHLSL;
1064 }
1065
1066 // This method needs to match OutputHLSL::decorate
1067 std::string DynamicHLSL::decorateVariable(const std::string &name)
1068 {
1069     if (name.compare(0, 3, "gl_") != 0)
1070     {
1071         return "_" + name;
1072     }
1073
1074     return name;
1075 }
1076
1077 std::string DynamicHLSL::generateAttributeConversionHLSL(const VertexFormat &vertexFormat, const sh::ShaderVariable &shaderAttrib) const
1078 {
1079     std::string attribString = "input." + decorateVariable(shaderAttrib.name);
1080
1081     // Matrix
1082     if (IsMatrixType(shaderAttrib.type))
1083     {
1084         return "transpose(" + attribString + ")";
1085     }
1086
1087     GLenum shaderComponentType = VariableComponentType(shaderAttrib.type);
1088     int shaderComponentCount = VariableComponentCount(shaderAttrib.type);
1089
1090     // Perform integer to float conversion (if necessary)
1091     bool requiresTypeConversion = (shaderComponentType == GL_FLOAT && vertexFormat.mType != GL_FLOAT);
1092
1093     if (requiresTypeConversion)
1094     {
1095         // TODO: normalization for 32-bit integer formats
1096         ASSERT(!vertexFormat.mNormalized && !vertexFormat.mPureInteger);
1097         return "float" + Str(shaderComponentCount) + "(" + attribString + ")";
1098     }
1099
1100     // No conversion necessary
1101     return attribString;
1102 }
1103
1104 void DynamicHLSL::getInputLayoutSignature(const VertexFormat inputLayout[], GLenum signature[]) const
1105 {
1106     for (size_t inputIndex = 0; inputIndex < MAX_VERTEX_ATTRIBS; inputIndex++)
1107     {
1108         const VertexFormat &vertexFormat = inputLayout[inputIndex];
1109
1110         if (vertexFormat.mType == GL_NONE)
1111         {
1112             signature[inputIndex] = GL_NONE;
1113         }
1114         else
1115         {
1116             bool gpuConverted = ((mRenderer->getVertexConversionType(vertexFormat) & rx::VERTEX_CONVERT_GPU) != 0);
1117             signature[inputIndex] = (gpuConverted ? GL_TRUE : GL_FALSE);
1118         }
1119     }
1120 }
1121
1122 }