src/Renderer/Renderer.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 "Renderer.hpp"
  16
  17 #include "Clipper.hpp"
  18 #include "Math.hpp"
  19 #include "FrameBuffer.hpp"
  20 #include "Timer.hpp"
  21 #include "Surface.hpp"
  22 #include "Half.hpp"
  23 #include "Primitive.hpp"
  24 #include "Polygon.hpp"
  25 #include "SwiftConfig.hpp"
  26 #include "MutexLock.hpp"
  27 #include "CPUID.hpp"
  28 #include "Memory.hpp"
  29 #include "Resource.hpp"
  30 #include "Constants.hpp"
  31 #include "Debug.hpp"
  32 #include "Reactor/Reactor.hpp"
  33
  34 #undef max
  35
  36 bool disableServer = true;
  37
  38 #ifndef NDEBUG
  39 unsigned int minPrimitives = 1;
  40 unsigned int maxPrimitives = 1 << 21;
  41 #endif
  42
  43 namespace sw
  44 {
  45         extern bool halfIntegerCoordinates;     // Pixel centers are not at integer coordinates
  46         extern bool symmetricNormalizedDepth;   // [-1, 1] instead of [0, 1]
  47         extern bool booleanFaceRegister;
  48         extern bool fullPixelPositionRegister;
  49         extern bool leadingVertexFirst;         // Flat shading uses first vertex, else last
  50         extern bool secondaryColor;             // Specular lighting is applied after texturing
  51
  52         extern bool forceWindowed;
  53         extern bool complementaryDepthBuffer;
  54         extern bool postBlendSRGB;
  55         extern bool exactColorRounding;
  56         extern TransparencyAntialiasing transparencyAntialiasing;
  57         extern bool forceClearRegisters;
  58
  59         extern bool precacheVertex;
  60         extern bool precacheSetup;
  61         extern bool precachePixel;
  62
  63         int batchSize = 128;
  64         int threadCount = 1;
  65         int unitCount = 1;
  66         int clusterCount = 1;
  67
  68         TranscendentalPrecision logPrecision = ACCURATE;
  69         TranscendentalPrecision expPrecision = ACCURATE;
  70         TranscendentalPrecision rcpPrecision = ACCURATE;
  71         TranscendentalPrecision rsqPrecision = ACCURATE;
  72         bool perspectiveCorrection = true;
  73
  74         struct Parameters
  75         {
  76                 Renderer *renderer;
  77                 int threadIndex;
  78         };
  79
  80         DrawCall::DrawCall()
  81         {
  82                 queries = 0;
  83
  84                 vsDirtyConstF = VERTEX_UNIFORM_VECTORS + 1;
  85                 vsDirtyConstI = 16;
  86                 vsDirtyConstB = 16;
  87
  88                 psDirtyConstF = FRAGMENT_UNIFORM_VECTORS;
  89                 psDirtyConstI = 16;
  90                 psDirtyConstB = 16;
  91
  92                 references = -1;
  93
  94                 data = (DrawData*)allocate(sizeof(DrawData));
  95                 data->constants = &constants;
  96         }
  97
  98         DrawCall::~DrawCall()
  99         {
 100                 delete queries;
 101
 102                 deallocate(data);
 103         }
 104
 105         Renderer::Renderer(Context *context, Conventions conventions, bool exactColorRounding) : VertexProcessor(context), PixelProcessor(context), SetupProcessor(context), context(context), viewport()
 106         {
 107                 sw::halfIntegerCoordinates = conventions.halfIntegerCoordinates;
 108                 sw::symmetricNormalizedDepth = conventions.symmetricNormalizedDepth;
 109                 sw::booleanFaceRegister = conventions.booleanFaceRegister;
 110                 sw::fullPixelPositionRegister = conventions.fullPixelPositionRegister;
 111                 sw::leadingVertexFirst = conventions.leadingVertexFirst;
 112                 sw::secondaryColor = conventions.secondaryColor;
 113                 sw::exactColorRounding = exactColorRounding;
 114
 115                 setRenderTarget(0, 0);
 116                 clipper = new Clipper();
 117
 118                 updateViewMatrix = true;
 119                 updateBaseMatrix = true;
 120                 updateProjectionMatrix = true;
 121                 updateClipPlanes = true;
 122
 123                 #if PERF_HUD
 124                         resetTimers();
 125                 #endif
 126
 127                 for(int i = 0; i < 16; i++)
 128                 {
 129                         vertexTask[i] = 0;
 130
 131                         worker[i] = 0;
 132                         resume[i] = 0;
 133                         suspend[i] = 0;
 134                 }
 135
 136                 threadsAwake = 0;
 137                 resumeApp = new Event();
 138
 139                 currentDraw = 0;
 140                 nextDraw = 0;
 141
 142                 qHead = 0;
 143                 qSize = 0;
 144
 145                 for(int i = 0; i < 16; i++)
 146                 {
 147                         triangleBatch[i] = 0;
 148                         primitiveBatch[i] = 0;
 149                 }
 150
 151                 for(int draw = 0; draw < DRAW_COUNT; draw++)
 152                 {
 153                         drawCall[draw] = new DrawCall();
 154                         drawList[draw] = drawCall[draw];
 155                 }
 156
 157                 for(int unit = 0; unit < 16; unit++)
 158                 {
 159                         primitiveProgress[unit].init();
 160                 }
 161
 162                 for(int cluster = 0; cluster < 16; cluster++)
 163                 {
 164                         pixelProgress[cluster].init();
 165                 }
 166
 167                 clipFlags = 0;
 168
 169                 swiftConfig = new SwiftConfig(disableServer);
 170                 updateConfiguration(true);
 171
 172                 sync = new Resource(0);
 173         }
 174
 175         Renderer::~Renderer()
 176         {
 177                 sync->destruct();
 178
 179                 delete clipper;
 180                 clipper = 0;
 181
 182                 terminateThreads();
 183                 delete resumeApp;
 184
 185                 for(int draw = 0; draw < DRAW_COUNT; draw++)
 186                 {
 187                         delete drawCall[draw];
 188                 }
 189
 190                 delete swiftConfig;
 191         }
 192
 193         void Renderer::clear(void *pixel, Format format, Surface *dest, const SliceRect &dRect, unsigned int rgbaMask)
 194         {
 195                 blitter.clear(pixel, format, dest, dRect, rgbaMask);
 196         }
 197
 198         void Renderer::blit(Surface *source, const SliceRect &sRect, Surface *dest, const SliceRect &dRect, bool filter)
 199         {
 200                 blitter.blit(source, sRect, dest, dRect, filter);
 201         }
 202
 203         void Renderer::blit3D(Surface *source, Surface *dest)
 204         {
 205                 blitter.blit3D(source, dest);
 206         }
 207
 208         void Renderer::draw(DrawType drawType, unsigned int indexOffset, unsigned int count, bool update)
 209         {
 210                 #ifndef NDEBUG
 211                         if(count < minPrimitives || count > maxPrimitives)
 212                         {
 213                                 return;
 214                         }
 215                 #endif
 216
 217                 context->drawType = drawType;
 218
 219                 updateConfiguration();
 220                 updateClipper();
 221
 222                 int ss = context->getSuperSampleCount();
 223                 int ms = context->getMultiSampleCount();
 224
 225                 for(int q = 0; q < ss; q++)
 226                 {
 227                         unsigned int oldMultiSampleMask = context->multiSampleMask;
 228                         context->multiSampleMask = (context->sampleMask >> (ms * q)) & ((unsigned)0xFFFFFFFF >> (32 - ms));
 229
 230                         if(!context->multiSampleMask)
 231                         {
 232                                 continue;
 233                         }
 234
 235                         sync->lock(sw::PRIVATE);
 236
 237                         Routine *vertexRoutine;
 238                         Routine *setupRoutine;
 239                         Routine *pixelRoutine;
 240
 241                         if(update || oldMultiSampleMask != context->multiSampleMask)
 242                         {
 243                                 vertexState = VertexProcessor::update(drawType);
 244                                 setupState = SetupProcessor::update();
 245                                 pixelState = PixelProcessor::update();
 246
 247                                 vertexRoutine = VertexProcessor::routine(vertexState);
 248                                 setupRoutine = SetupProcessor::routine(setupState);
 249                                 pixelRoutine = PixelProcessor::routine(pixelState);
 250                         }
 251
 252                         int batch = batchSize / ms;
 253
 254                         int (*setupPrimitives)(Renderer *renderer, int batch, int count);
 255
 256                         if(context->isDrawTriangle())
 257                         {
 258                                 switch(context->fillMode)
 259                                 {
 260                                 case FILL_SOLID:
 261                                         setupPrimitives = setupSolidTriangles;
 262                                         break;
 263                                 case FILL_WIREFRAME:
 264                                         setupPrimitives = setupWireframeTriangle;
 265                                         batch = 1;
 266                                         break;
 267                                 case FILL_VERTEX:
 268                                         setupPrimitives = setupVertexTriangle;
 269                                         batch = 1;
 270                                         break;
 271                                 default: ASSERT(false);
 272                                 }
 273                         }
 274                         else if(context->isDrawLine())
 275                         {
 276                                 setupPrimitives = setupLines;
 277                         }
 278                         else   // Point draw
 279                         {
 280                                 setupPrimitives = setupPoints;
 281                         }
 282
 283                         DrawCall *draw = 0;
 284
 285                         do
 286                         {
 287                                 for(int i = 0; i < DRAW_COUNT; i++)
 288                                 {
 289                                         if(drawCall[i]->references == -1)
 290                                         {
 291                                                 draw = drawCall[i];
 292                                                 drawList[nextDraw % DRAW_COUNT] = draw;
 293
 294                                                 break;
 295                                         }
 296                                 }
 297
 298                                 if(!draw)
 299                                 {
 300                                         resumeApp->wait();
 301                                 }
 302                         }
 303                         while(!draw);
 304
 305                         DrawData *data = draw->data;
 306
 307                         if(queries.size() != 0)
 308                         {
 309                                 draw->queries = new std::list<Query*>();
 310                                 bool includePrimitivesWrittenQueries = vertexState.transformFeedbackQueryEnabled && vertexState.transformFeedbackEnabled;
 311                                 for(std::list<Query*>::iterator query = queries.begin(); query != queries.end(); query++)
 312                                 {
 313                                         Query* q = *query;
 314                                         if(includePrimitivesWrittenQueries || (q->type != Query::TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN))
 315                                         {
 316                                                 atomicIncrement(&(q->reference));
 317                                                 draw->queries->push_back(q);
 318                                         }
 319                                 }
 320                         }
 321
 322                         draw->drawType = drawType;
 323                         draw->batchSize = batch;
 324
 325                         vertexRoutine->bind();
 326                         setupRoutine->bind();
 327                         pixelRoutine->bind();
 328
 329                         draw->vertexRoutine = vertexRoutine;
 330                         draw->setupRoutine = setupRoutine;
 331                         draw->pixelRoutine = pixelRoutine;
 332                         draw->vertexPointer = (VertexProcessor::RoutinePointer)vertexRoutine->getEntry();
 333                         draw->setupPointer = (SetupProcessor::RoutinePointer)setupRoutine->getEntry();
 334                         draw->pixelPointer = (PixelProcessor::RoutinePointer)pixelRoutine->getEntry();
 335                         draw->setupPrimitives = setupPrimitives;
 336                         draw->setupState = setupState;
 337
 338                         for(int i = 0; i < MAX_VERTEX_INPUTS; i++)
 339                         {
 340                                 draw->vertexStream[i] = context->input[i].resource;
 341                                 data->input[i] = context->input[i].buffer;
 342                                 data->stride[i] = context->input[i].stride;
 343
 344                                 if(draw->vertexStream[i])
 345                                 {
 346                                         draw->vertexStream[i]->lock(PUBLIC, PRIVATE);
 347                                 }
 348                         }
 349
 350                         if(context->indexBuffer)
 351                         {
 352                                 data->indices = (unsigned char*)context->indexBuffer->lock(PUBLIC, PRIVATE) + indexOffset;
 353                         }
 354
 355                         draw->indexBuffer = context->indexBuffer;
 356
 357                         for(int sampler = 0; sampler < TOTAL_IMAGE_UNITS; sampler++)
 358                         {
 359                                 draw->texture[sampler] = 0;
 360                         }
 361
 362                         for(int sampler = 0; sampler < TEXTURE_IMAGE_UNITS; sampler++)
 363                         {
 364                                 if(pixelState.sampler[sampler].textureType != TEXTURE_NULL)
 365                                 {
 366                                         draw->texture[sampler] = context->texture[sampler];
 367                                         draw->texture[sampler]->lock(PUBLIC, isReadWriteTexture(sampler) ? MANAGED : PRIVATE);   // If the texure is both read and written, use the same read/write lock as render targets
 368
 369                                         data->mipmap[sampler] = context->sampler[sampler].getTextureData();
 370                                 }
 371                         }
 372
 373                         if(context->pixelShader)
 374                         {
 375                                 if(draw->psDirtyConstF)
 376                                 {
 377                                         memcpy(&data->ps.cW, PixelProcessor::cW, sizeof(word4) * 4 * (draw->psDirtyConstF < 8 ? draw->psDirtyConstF : 8));
 378                                         memcpy(&data->ps.c, PixelProcessor::c, sizeof(float4) * draw->psDirtyConstF);
 379                                         draw->psDirtyConstF = 0;
 380                                 }
 381
 382                                 if(draw->psDirtyConstI)
 383                                 {
 384                                         memcpy(&data->ps.i, PixelProcessor::i, sizeof(int4) * draw->psDirtyConstI);
 385                                         draw->psDirtyConstI = 0;
 386                                 }
 387
 388                                 if(draw->psDirtyConstB)
 389                                 {
 390                                         memcpy(&data->ps.b, PixelProcessor::b, sizeof(bool) * draw->psDirtyConstB);
 391                                         draw->psDirtyConstB = 0;
 392                                 }
 393
 394                                 PixelProcessor::lockUniformBuffers(data->ps.u, draw->pUniformBuffers);
 395                         }
 396                         else
 397                         {
 398                                 for(int i = 0; i < MAX_UNIFORM_BUFFER_BINDINGS; i++)
 399                                 {
 400                                         draw->pUniformBuffers[i] = nullptr;
 401                                 }
 402                         }
 403
 404                         if(context->pixelShaderVersion() <= 0x0104)
 405                         {
 406                                 for(int stage = 0; stage < 8; stage++)
 407                                 {
 408                                         if(pixelState.textureStage[stage].stageOperation != TextureStage::STAGE_DISABLE || context->pixelShader)
 409                                         {
 410                                                 data->textureStage[stage] = context->textureStage[stage].uniforms;
 411                                         }
 412                                         else break;
 413                                 }
 414                         }
 415
 416                         if(context->vertexShader)
 417                         {
 418                                 if(context->vertexShader->getVersion() >= 0x0300)
 419                                 {
 420                                         for(int sampler = 0; sampler < VERTEX_TEXTURE_IMAGE_UNITS; sampler++)
 421                                         {
 422                                                 if(vertexState.samplerState[sampler].textureType != TEXTURE_NULL)
 423                                                 {
 424                                                         draw->texture[TEXTURE_IMAGE_UNITS + sampler] = context->texture[TEXTURE_IMAGE_UNITS + sampler];
 425                                                         draw->texture[TEXTURE_IMAGE_UNITS + sampler]->lock(PUBLIC, PRIVATE);
 426
 427                                                         data->mipmap[TEXTURE_IMAGE_UNITS + sampler] = context->sampler[TEXTURE_IMAGE_UNITS + sampler].getTextureData();
 428                                                 }
 429                                         }
 430                                 }
 431
 432                                 if(draw->vsDirtyConstF)
 433                                 {
 434                                         memcpy(&data->vs.c, VertexProcessor::c, sizeof(float4) * draw->vsDirtyConstF);
 435                                         draw->vsDirtyConstF = 0;
 436                                 }
 437
 438                                 if(draw->vsDirtyConstI)
 439                                 {
 440                                         memcpy(&data->vs.i, VertexProcessor::i, sizeof(int4) * draw->vsDirtyConstI);
 441                                         draw->vsDirtyConstI = 0;
 442                                 }
 443
 444                                 if(draw->vsDirtyConstB)
 445                                 {
 446                                         memcpy(&data->vs.b, VertexProcessor::b, sizeof(bool) * draw->vsDirtyConstB);
 447                                         draw->vsDirtyConstB = 0;
 448                                 }
 449
 450                                 if(context->vertexShader->instanceIdDeclared)
 451                                 {
 452                                         data->instanceID = context->instanceID;
 453                                 }
 454
 455                                 VertexProcessor::lockUniformBuffers(data->vs.u, draw->vUniformBuffers);
 456                                 VertexProcessor::lockTransformFeedbackBuffers(data->vs.t, data->vs.reg, data->vs.row, data->vs.col, data->vs.str, draw->transformFeedbackBuffers);
 457                         }
 458                         else
 459                         {
 460                                 data->ff = ff;
 461
 462                                 draw->vsDirtyConstF = VERTEX_UNIFORM_VECTORS + 1;
 463                                 draw->vsDirtyConstI = 16;
 464                                 draw->vsDirtyConstB = 16;
 465
 466                                 for(int i = 0; i < MAX_UNIFORM_BUFFER_BINDINGS; i++)
 467                                 {
 468                                         draw->vUniformBuffers[i] = nullptr;
 469                                 }
 470
 471                                 for(int i = 0; i < MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS; i++)
 472                                 {
 473                                         draw->transformFeedbackBuffers[i] = nullptr;
 474                                 }
 475                         }
 476
 477                         if(pixelState.stencilActive)
 478                         {
 479                                 data->stencil[0] = stencil;
 480                                 data->stencil[1] = stencilCCW;
 481                         }
 482
 483                         if(pixelState.fogActive)
 484                         {
 485                                 data->fog = fog;
 486                         }
 487
 488                         if(setupState.isDrawPoint)
 489                         {
 490                                 data->point = point;
 491                         }
 492
 493                         data->lineWidth = context->lineWidth;
 494
 495                         data->factor = factor;
 496
 497                         if(pixelState.transparencyAntialiasing == TRANSPARENCY_ALPHA_TO_COVERAGE)
 498                         {
 499                                 float ref = context->alphaReference * (1.0f / 255.0f);
 500                                 float margin = sw::min(ref, 1.0f - ref);
 501
 502                                 if(ms == 4)
 503                                 {
 504                                         data->a2c0 = replicate(ref - margin * 0.6f);
 505                                         data->a2c1 = replicate(ref - margin * 0.2f);
 506                                         data->a2c2 = replicate(ref + margin * 0.2f);
 507                                         data->a2c3 = replicate(ref + margin * 0.6f);
 508                                 }
 509                                 else if(ms == 2)
 510                                 {
 511                                         data->a2c0 = replicate(ref - margin * 0.3f);
 512                                         data->a2c1 = replicate(ref + margin * 0.3f);
 513                                 }
 514                                 else ASSERT(false);
 515                         }
 516
 517                         if(pixelState.occlusionEnabled)
 518                         {
 519                                 for(int cluster = 0; cluster < clusterCount; cluster++)
 520                                 {
 521                                         data->occlusion[cluster] = 0;
 522                                 }
 523                         }
 524
 525                         #if PERF_PROFILE
 526                                 for(int cluster = 0; cluster < clusterCount; cluster++)
 527                                 {
 528                                         for(int i = 0; i < PERF_TIMERS; i++)
 529                                         {
 530                                                 data->cycles[i][cluster] = 0;
 531                                         }
 532                                 }
 533                         #endif
 534
 535                         // Viewport
 536                         {
 537                                 float W = 0.5f * viewport.width;
 538                                 float H = 0.5f * viewport.height;
 539                                 float X0 = viewport.x0 + W;
 540                                 float Y0 = viewport.y0 + H;
 541                                 float N = viewport.minZ;
 542                                 float F = viewport.maxZ;
 543                                 float Z = F - N;
 544
 545                                 if(context->isDrawTriangle(false))
 546                                 {
 547                                         N += depthBias;
 548                                 }
 549
 550                                 if(complementaryDepthBuffer)
 551                                 {
 552                                         Z = -Z;
 553                                         N = 1 - N;
 554                                 }
 555
 556                                 static const float X[5][16] =   // Fragment offsets
 557                                 {
 558                                         {+0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f},   // 1 sample
 559                                         {-0.2500f, +0.2500f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f},   // 2 samples
 560                                         {-0.3000f, +0.1000f, +0.3000f, -0.1000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f},   // 4 samples
 561                                         {+0.1875f, -0.3125f, +0.3125f, -0.4375f, -0.0625f, +0.4375f, +0.0625f, -0.1875f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f},   // 8 samples
 562                                         {+0.2553f, -0.1155f, +0.1661f, -0.1828f, +0.2293f, -0.4132f, -0.1773f, -0.0577f, +0.3891f, -0.4656f, +0.4103f, +0.4248f, -0.2109f, +0.3966f, -0.2664f, -0.3872f}    // 16 samples
 563                                 };
 564
 565                                 static const float Y[5][16] =   // Fragment offsets
 566                                 {
 567                                         {+0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f},   // 1 sample
 568                                         {-0.2500f, +0.2500f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f},   // 2 samples
 569                                         {-0.1000f, -0.3000f, +0.1000f, +0.3000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f},   // 4 samples
 570                                         {-0.4375f, -0.3125f, -0.1875f, -0.0625f, +0.0625f, +0.1875f, +0.3125f, +0.4375f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f, +0.0000f},   // 8 samples
 571                                         {-0.4503f, +0.1883f, +0.3684f, -0.4668f, -0.0690f, -0.1315f, +0.4999f, +0.0728f, +0.1070f, -0.3086f, +0.3725f, -0.1547f, -0.1102f, -0.3588f, +0.1789f, +0.0269f}    // 16 samples
 572                                 };
 573
 574                                 int s = sw::log2(ss);
 575
 576                                 data->Wx16 = replicate(W * 16);
 577                                 data->Hx16 = replicate(H * 16);
 578                                 data->X0x16 = replicate(X0 * 16 - 8);
 579                                 data->Y0x16 = replicate(Y0 * 16 - 8);
 580                                 data->XXXX = replicate(X[s][q] / W);
 581                                 data->YYYY = replicate(Y[s][q] / H);
 582                                 data->halfPixelX = replicate(0.5f / W);
 583                                 data->halfPixelY = replicate(0.5f / H);
 584                                 data->viewportHeight = abs(viewport.height);
 585                                 data->slopeDepthBias = slopeDepthBias;
 586                                 data->depthRange = Z;
 587                                 data->depthNear = N;
 588                                 draw->clipFlags = clipFlags;
 589
 590                                 if(clipFlags)
 591                                 {
 592                                         if(clipFlags & Clipper::CLIP_PLANE0) data->clipPlane[0] = clipPlane[0];
 593                                         if(clipFlags & Clipper::CLIP_PLANE1) data->clipPlane[1] = clipPlane[1];
 594                                         if(clipFlags & Clipper::CLIP_PLANE2) data->clipPlane[2] = clipPlane[2];
 595                                         if(clipFlags & Clipper::CLIP_PLANE3) data->clipPlane[3] = clipPlane[3];
 596                                         if(clipFlags & Clipper::CLIP_PLANE4) data->clipPlane[4] = clipPlane[4];
 597                                         if(clipFlags & Clipper::CLIP_PLANE5) data->clipPlane[5] = clipPlane[5];
 598                                 }
 599                         }
 600
 601                         // Target
 602                         {
 603                                 for(int index = 0; index < RENDERTARGETS; index++)
 604                                 {
 605                                         draw->renderTarget[index] = context->renderTarget[index];
 606
 607                                         if(draw->renderTarget[index])
 608                                         {
 609                                                 data->colorBuffer[index] = (unsigned int*)context->renderTarget[index]->lockInternal(0, 0, q * ms, LOCK_READWRITE, MANAGED);
 610                                                 data->colorPitchB[index] = context->renderTarget[index]->getInternalPitchB();
 611                                                 data->colorSliceB[index] = context->renderTarget[index]->getInternalSliceB();
 612                                         }
 613                                 }
 614
 615                                 draw->depthBuffer = context->depthBuffer;
 616                                 draw->stencilBuffer = context->stencilBuffer;
 617
 618                                 if(draw->depthBuffer)
 619                                 {
 620                                         data->depthBuffer = (float*)context->depthBuffer->lockInternal(0, 0, q * ms, LOCK_READWRITE, MANAGED);
 621                                         data->depthPitchB = context->depthBuffer->getInternalPitchB();
 622                                         data->depthSliceB = context->depthBuffer->getInternalSliceB();
 623                                 }
 624
 625                                 if(draw->stencilBuffer)
 626                                 {
 627                                         data->stencilBuffer = (unsigned char*)context->stencilBuffer->lockStencil(q * ms, MANAGED);
 628                                         data->stencilPitchB = context->stencilBuffer->getStencilPitchB();
 629                                         data->stencilSliceB = context->stencilBuffer->getStencilSliceB();
 630                                 }
 631                         }
 632
 633                         // Scissor
 634                         {
 635                                 data->scissorX0 = scissor.x0;
 636                                 data->scissorX1 = scissor.x1;
 637                                 data->scissorY0 = scissor.y0;
 638                                 data->scissorY1 = scissor.y1;
 639                         }
 640
 641                         draw->primitive = 0;
 642                         draw->count = count;
 643
 644                         draw->references = (count + batch - 1) / batch;
 645
 646                         schedulerMutex.lock();
 647                         nextDraw++;
 648                         schedulerMutex.unlock();
 649
 650                         if(threadCount > 1)
 651                         {
 652                                 if(!threadsAwake)
 653                                 {
 654                                         suspend[0]->wait();
 655
 656                                         threadsAwake = 1;
 657                                         task[0].type = Task::RESUME;
 658
 659                                         resume[0]->signal();
 660                                 }
 661                         }
 662                         else   // Use main thread for draw execution
 663                         {
 664                                 threadsAwake = 1;
 665                                 task[0].type = Task::RESUME;
 666
 667                                 taskLoop(0);
 668                         }
 669                 }
 670         }
 671
 672         void Renderer::threadFunction(void *parameters)
 673         {
 674                 Renderer *renderer = static_cast<Parameters*>(parameters)->renderer;
 675                 int threadIndex = static_cast<Parameters*>(parameters)->threadIndex;
 676
 677                 if(logPrecision < IEEE)
 678                 {
 679                         CPUID::setFlushToZero(true);
 680                         CPUID::setDenormalsAreZero(true);
 681                 }
 682
 683                 renderer->threadLoop(threadIndex);
 684         }
 685
 686         void Renderer::threadLoop(int threadIndex)
 687         {
 688                 while(!exitThreads)
 689                 {
 690                         taskLoop(threadIndex);
 691
 692                         suspend[threadIndex]->signal();
 693                         resume[threadIndex]->wait();
 694                 }
 695         }
 696
 697         void Renderer::taskLoop(int threadIndex)
 698         {
 699                 while(task[threadIndex].type != Task::SUSPEND)
 700                 {
 701                         scheduleTask(threadIndex);
 702                         executeTask(threadIndex);
 703                 }
 704         }
 705
 706         void Renderer::findAvailableTasks()
 707         {
 708                 // Find pixel tasks
 709                 for(int cluster = 0; cluster < clusterCount; cluster++)
 710                 {
 711                         if(!pixelProgress[cluster].executing)
 712                         {
 713                                 for(int unit = 0; unit < unitCount; unit++)
 714                                 {
 715                                         if(primitiveProgress[unit].references > 0)   // Contains processed primitives
 716                                         {
 717                                                 if(pixelProgress[cluster].drawCall == primitiveProgress[unit].drawCall)
 718                                                 {
 719                                                         if(pixelProgress[cluster].processedPrimitives == primitiveProgress[unit].firstPrimitive)   // Previous primitives have been rendered
 720                                                         {
 721                                                                 Task &task = taskQueue[qHead];
 722                                                                 task.type = Task::PIXELS;
 723                                                                 task.primitiveUnit = unit;
 724                                                                 task.pixelCluster = cluster;
 725
 726                                                                 pixelProgress[cluster].executing = true;
 727
 728                                                                 // Commit to the task queue
 729                                                                 qHead = (qHead + 1) % 32;
 730                                                                 qSize++;
 731
 732                                                                 break;
 733                                                         }
 734                                                 }
 735                                         }
 736                                 }
 737                         }
 738                 }
 739
 740                 // Find primitive tasks
 741                 if(currentDraw == nextDraw)
 742                 {
 743                         return;   // No more primitives to process
 744                 }
 745
 746                 for(int unit = 0; unit < unitCount; unit++)
 747                 {
 748                         DrawCall *draw = drawList[currentDraw % DRAW_COUNT];
 749
 750                         if(draw->primitive >= draw->count)
 751                         {
 752                                 currentDraw++;
 753
 754                                 if(currentDraw == nextDraw)
 755                                 {
 756                                         return;   // No more primitives to process
 757                                 }
 758
 759                                 draw = drawList[currentDraw % DRAW_COUNT];
 760                         }
 761
 762                         if(!primitiveProgress[unit].references)   // Task not already being executed and not still in use by a pixel unit
 763                         {
 764                                 int primitive = draw->primitive;
 765                                 int count = draw->count;
 766                                 int batch = draw->batchSize;
 767
 768                                 primitiveProgress[unit].drawCall = currentDraw;
 769                                 primitiveProgress[unit].firstPrimitive = primitive;
 770                                 primitiveProgress[unit].primitiveCount = count - primitive >= batch ? batch : count - primitive;
 771
 772                                 draw->primitive += batch;
 773
 774                                 Task &task = taskQueue[qHead];
 775                                 task.type = Task::PRIMITIVES;
 776                                 task.primitiveUnit = unit;
 777
 778                                 primitiveProgress[unit].references = -1;
 779
 780                                 // Commit to the task queue
 781                                 qHead = (qHead + 1) % 32;
 782                                 qSize++;
 783                         }
 784                 }
 785         }
 786
 787         void Renderer::scheduleTask(int threadIndex)
 788         {
 789                 schedulerMutex.lock();
 790
 791                 if((int)qSize < threadCount - threadsAwake + 1)
 792                 {
 793                         findAvailableTasks();
 794                 }
 795
 796                 if(qSize != 0)
 797                 {
 798                         task[threadIndex] = taskQueue[(qHead - qSize) % 32];
 799                         qSize--;
 800
 801                         if(threadsAwake != threadCount)
 802                         {
 803                                 int wakeup = qSize - threadsAwake + 1;
 804
 805                                 for(int i = 0; i < threadCount && wakeup > 0; i++)
 806                                 {
 807                                         if(task[i].type == Task::SUSPEND)
 808                                         {
 809                                                 suspend[i]->wait();
 810                                                 task[i].type = Task::RESUME;
 811                                                 resume[i]->signal();
 812
 813                                                 threadsAwake++;
 814                                                 wakeup--;
 815                                         }
 816                                 }
 817                         }
 818                 }
 819                 else
 820                 {
 821                         task[threadIndex].type = Task::SUSPEND;
 822
 823                         threadsAwake--;
 824                 }
 825
 826                 schedulerMutex.unlock();
 827         }
 828
 829         void Renderer::executeTask(int threadIndex)
 830         {
 831                 #if PERF_HUD
 832                         int64_t startTick = Timer::ticks();
 833                 #endif
 834
 835                 switch(task[threadIndex].type)
 836                 {
 837                 case Task::PRIMITIVES:
 838                         {
 839                                 int unit = task[threadIndex].primitiveUnit;
 840
 841                                 int input = primitiveProgress[unit].firstPrimitive;
 842                                 int count = primitiveProgress[unit].primitiveCount;
 843                                 DrawCall *draw = drawList[primitiveProgress[unit].drawCall % DRAW_COUNT];
 844                                 int (*setupPrimitives)(Renderer *renderer, int batch, int count) = draw->setupPrimitives;
 845
 846                                 processPrimitiveVertices(unit, input, count, draw->count, threadIndex);
 847
 848                                 #if PERF_HUD
 849                                         int64_t time = Timer::ticks();
 850                                         vertexTime[threadIndex] += time - startTick;
 851                                         startTick = time;
 852                                 #endif
 853
 854                                 int visible = draw->setupState.rasterizerDiscard ? 0 : setupPrimitives(this, unit, count);
 855
 856                                 primitiveProgress[unit].visible = visible;
 857                                 primitiveProgress[unit].references = clusterCount;
 858
 859                                 #if PERF_HUD
 860                                         setupTime[threadIndex] += Timer::ticks() - startTick;
 861                                 #endif
 862                         }
 863                         break;
 864                 case Task::PIXELS:
 865                         {
 866                                 int unit = task[threadIndex].primitiveUnit;
 867                                 int visible = primitiveProgress[unit].visible;
 868
 869                                 if(visible > 0)
 870                                 {
 871                                         int cluster = task[threadIndex].pixelCluster;
 872                                         Primitive *primitive = primitiveBatch[unit];
 873                                         DrawCall *draw = drawList[pixelProgress[cluster].drawCall % DRAW_COUNT];
 874                                         DrawData *data = draw->data;
 875                                         PixelProcessor::RoutinePointer pixelRoutine = draw->pixelPointer;
 876
 877                                         pixelRoutine(primitive, visible, cluster, data);
 878                                 }
 879
 880                                 finishRendering(task[threadIndex]);
 881
 882                                 #if PERF_HUD
 883                                         pixelTime[threadIndex] += Timer::ticks() - startTick;
 884                                 #endif
 885                         }
 886                         break;
 887                 case Task::RESUME:
 888                         break;
 889                 case Task::SUSPEND:
 890                         break;
 891                 default:
 892                         ASSERT(false);
 893                 }
 894         }
 895
 896         void Renderer::synchronize()
 897         {
 898                 sync->lock(sw::PUBLIC);
 899                 sync->unlock();
 900         }
 901
 902         void Renderer::finishRendering(Task &pixelTask)
 903         {
 904                 int unit = pixelTask.primitiveUnit;
 905                 int cluster = pixelTask.pixelCluster;
 906
 907                 DrawCall &draw = *drawList[primitiveProgress[unit].drawCall % DRAW_COUNT];
 908                 DrawData &data = *draw.data;
 909                 int primitive = primitiveProgress[unit].firstPrimitive;
 910                 int count = primitiveProgress[unit].primitiveCount;
 911                 int processedPrimitives = primitive + count;
 912
 913                 pixelProgress[cluster].processedPrimitives = processedPrimitives;
 914
 915                 if(pixelProgress[cluster].processedPrimitives >= draw.count)
 916                 {
 917                         pixelProgress[cluster].drawCall++;
 918                         pixelProgress[cluster].processedPrimitives = 0;
 919                 }
 920
 921                 int ref = atomicDecrement(&primitiveProgress[unit].references);
 922
 923                 if(ref == 0)
 924                 {
 925                         ref = atomicDecrement(&draw.references);
 926
 927                         if(ref == 0)
 928                         {
 929                                 #if PERF_PROFILE
 930                                         for(int cluster = 0; cluster < clusterCount; cluster++)
 931                                         {
 932                                                 for(int i = 0; i < PERF_TIMERS; i++)
 933                                                 {
 934                                                         profiler.cycles[i] += data.cycles[i][cluster];
 935                                                 }
 936                                         }
 937                                 #endif
 938
 939                                 if(draw.queries)
 940                                 {
 941                                         for(std::list<Query*>::iterator q = draw.queries->begin(); q != draw.queries->end(); q++)
 942                                         {
 943                                                 Query *query = *q;
 944
 945                                                 switch(query->type)
 946                                                 {
 947                                                 case Query::FRAGMENTS_PASSED:
 948                                                         for(int cluster = 0; cluster < clusterCount; cluster++)
 949                                                         {
 950                                                                 atomicAdd((volatile int*)&query->data, data.occlusion[cluster]);
 951                                                         }
 952                                                         break;
 953                                                 case Query::TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
 954                                                         atomicAdd((volatile int*)&query->data, processedPrimitives);
 955                                                         break;
 956                                                 default:
 957                                                         break;
 958                                                 }
 959
 960                                                 atomicDecrement(&query->reference);
 961                                         }
 962
 963                                         delete draw.queries;
 964                                         draw.queries = 0;
 965                                 }
 966
 967                                 for(int i = 0; i < RENDERTARGETS; i++)
 968                                 {
 969                                         if(draw.renderTarget[i])
 970                                         {
 971                                                 draw.renderTarget[i]->unlockInternal();
 972                                         }
 973                                 }
 974
 975                                 if(draw.depthBuffer)
 976                                 {
 977                                         draw.depthBuffer->unlockInternal();
 978                                 }
 979
 980                                 if(draw.stencilBuffer)
 981                                 {
 982                                         draw.stencilBuffer->unlockStencil();
 983                                 }
 984
 985                                 for(int i = 0; i < TOTAL_IMAGE_UNITS; i++)
 986                                 {
 987                                         if(draw.texture[i])
 988                                         {
 989                                                 draw.texture[i]->unlock();
 990                                         }
 991                                 }
 992
 993                                 for(int i = 0; i < MAX_VERTEX_INPUTS; i++)
 994                                 {
 995                                         if(draw.vertexStream[i])
 996                                         {
 997                                                 draw.vertexStream[i]->unlock();
 998                                         }
 999                                 }
1000
1001                                 if(draw.indexBuffer)
1002                                 {
1003                                         draw.indexBuffer->unlock();
1004                                 }
1005
1006                                 for(int i = 0; i < MAX_UNIFORM_BUFFER_BINDINGS; i++)
1007                                 {
1008                                         if(draw.pUniformBuffers[i])
1009                                         {
1010                                                 draw.pUniformBuffers[i]->unlock();
1011                                         }
1012                                         if(draw.vUniformBuffers[i])
1013                                         {
1014                                                 draw.vUniformBuffers[i]->unlock();
1015                                         }
1016                                 }
1017
1018                                 for(int i = 0; i < MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS; i++)
1019                                 {
1020                                         if(draw.transformFeedbackBuffers[i])
1021                                         {
1022                                                 draw.transformFeedbackBuffers[i]->unlock();
1023                                         }
1024                                 }
1025
1026                                 draw.vertexRoutine->unbind();
1027                                 draw.setupRoutine->unbind();
1028                                 draw.pixelRoutine->unbind();
1029
1030                                 sync->unlock();
1031
1032                                 draw.references = -1;
1033                                 resumeApp->signal();
1034                         }
1035                 }
1036
1037                 pixelProgress[cluster].executing = false;
1038         }
1039
1040         void Renderer::processPrimitiveVertices(int unit, unsigned int start, unsigned int triangleCount, unsigned int loop, int thread)
1041         {
1042                 Triangle *triangle = triangleBatch[unit];
1043                 DrawCall *draw = drawList[primitiveProgress[unit].drawCall % DRAW_COUNT];
1044                 DrawData *data = draw->data;
1045                 VertexTask *task = vertexTask[thread];
1046
1047                 const void *indices = data->indices;
1048                 VertexProcessor::RoutinePointer vertexRoutine = draw->vertexPointer;
1049
1050                 if(task->vertexCache.drawCall != primitiveProgress[unit].drawCall)
1051                 {
1052                         task->vertexCache.clear();
1053                         task->vertexCache.drawCall = primitiveProgress[unit].drawCall;
1054                 }
1055
1056                 unsigned int batch[128][3];   // FIXME: Adjust to dynamic batch size
1057
1058                 switch(draw->drawType)
1059                 {
1060                 case DRAW_POINTLIST:
1061                         {
1062                                 unsigned int index = start;
1063
1064                                 for(unsigned int i = 0; i < triangleCount; i++)
1065                                 {
1066                                         batch[i][0] = index;
1067                                         batch[i][1] = index;
1068                                         batch[i][2] = index;
1069
1070                                         index += 1;
1071                                 }
1072                         }
1073                         break;
1074                 case DRAW_LINELIST:
1075                         {
1076                                 unsigned int index = 2 * start;
1077
1078                                 for(unsigned int i = 0; i < triangleCount; i++)
1079                                 {
1080                                         batch[i][0] = index + 0;
1081                                         batch[i][1] = index + 1;
1082                                         batch[i][2] = index + 1;
1083
1084                                         index += 2;
1085                                 }
1086                         }
1087                         break;
1088                 case DRAW_LINESTRIP:
1089                         {
1090                                 unsigned int index = start;
1091
1092                                 for(unsigned int i = 0; i < triangleCount; i++)
1093                                 {
1094                                         batch[i][0] = index + 0;
1095                                         batch[i][1] = index + 1;
1096                                         batch[i][2] = index + 1;
1097
1098                                         index += 1;
1099                                 }
1100                         }
1101                         break;
1102                 case DRAW_LINELOOP:
1103                         {
1104                                 unsigned int index = start;
1105
1106                                 for(unsigned int i = 0; i < triangleCount; i++)
1107                                 {
1108                                         batch[i][0] = (index + 0) % loop;
1109                                         batch[i][1] = (index + 1) % loop;
1110                                         batch[i][2] = (index + 1) % loop;
1111
1112                                         index += 1;
1113                                 }
1114                         }
1115                         break;
1116                 case DRAW_TRIANGLELIST:
1117                         {
1118                                 unsigned int index = 3 * start;
1119
1120                                 for(unsigned int i = 0; i < triangleCount; i++)
1121                                 {
1122                                         batch[i][0] = index + 0;
1123                                         batch[i][1] = index + 1;
1124                                         batch[i][2] = index + 2;
1125
1126                                         index += 3;
1127                                 }
1128                         }
1129                         break;
1130                 case DRAW_TRIANGLESTRIP:
1131                         {
1132                                 unsigned int index = start;
1133
1134                                 for(unsigned int i = 0; i < triangleCount; i++)
1135                                 {
1136                                         batch[i][0] = index + 0;
1137                                         batch[i][1] = index + (index & 1) + 1;
1138                                         batch[i][2] = index + (~index & 1) + 1;
1139
1140                                         index += 1;
1141                                 }
1142                         }
1143                         break;
1144                 case DRAW_TRIANGLEFAN:
1145                         {
1146                                 unsigned int index = start;
1147
1148                                 for(unsigned int i = 0; i < triangleCount; i++)
1149                                 {
1150                                         batch[i][0] = index + 1;
1151                                         batch[i][1] = index + 2;
1152                                         batch[i][2] = 0;
1153
1154                                         index += 1;
1155                                 }
1156                         }
1157                         break;
1158                 case DRAW_INDEXEDPOINTLIST8:
1159                         {
1160                                 const unsigned char *index = (const unsigned char*)indices + start;
1161
1162                                 for(unsigned int i = 0; i < triangleCount; i++)
1163                                 {
1164                                         batch[i][0] = *index;
1165                                         batch[i][1] = *index;
1166                                         batch[i][2] = *index;
1167
1168                                         index += 1;
1169                                 }
1170                         }
1171                         break;
1172                 case DRAW_INDEXEDPOINTLIST16:
1173                         {
1174                                 const unsigned short *index = (const unsigned short*)indices + start;
1175
1176                                 for(unsigned int i = 0; i < triangleCount; i++)
1177                                 {
1178                                         batch[i][0] = *index;
1179                                         batch[i][1] = *index;
1180                                         batch[i][2] = *index;
1181
1182                                         index += 1;
1183                                 }
1184                         }
1185                         break;
1186                 case DRAW_INDEXEDPOINTLIST32:
1187                         {
1188                                 const unsigned int *index = (const unsigned int*)indices + start;
1189
1190                                 for(unsigned int i = 0; i < triangleCount; i++)
1191                                 {
1192                                         batch[i][0] = *index;
1193                                         batch[i][1] = *index;
1194                                         batch[i][2] = *index;
1195
1196                                         index += 1;
1197                                 }
1198                         }
1199                         break;
1200                 case DRAW_INDEXEDLINELIST8:
1201                         {
1202                                 const unsigned char *index = (const unsigned char*)indices + 2 * start;
1203
1204                                 for(unsigned int i = 0; i < triangleCount; i++)
1205                                 {
1206                                         batch[i][0] = index[0];
1207                                         batch[i][1] = index[1];
1208                                         batch[i][2] = index[1];
1209
1210                                         index += 2;
1211                                 }
1212                         }
1213                         break;
1214                 case DRAW_INDEXEDLINELIST16:
1215                         {
1216                                 const unsigned short *index = (const unsigned short*)indices + 2 * start;
1217
1218                                 for(unsigned int i = 0; i < triangleCount; i++)
1219                                 {
1220                                         batch[i][0] = index[0];
1221                                         batch[i][1] = index[1];
1222                                         batch[i][2] = index[1];
1223
1224                                         index += 2;
1225                                 }
1226                         }
1227                         break;
1228                 case DRAW_INDEXEDLINELIST32:
1229                         {
1230                                 const unsigned int *index = (const unsigned int*)indices + 2 * start;
1231
1232                                 for(unsigned int i = 0; i < triangleCount; i++)
1233                                 {
1234                                         batch[i][0] = index[0];
1235                                         batch[i][1] = index[1];
1236                                         batch[i][2] = index[1];
1237
1238                                         index += 2;
1239                                 }
1240                         }
1241                         break;
1242                 case DRAW_INDEXEDLINESTRIP8:
1243                         {
1244                                 const unsigned char *index = (const unsigned char*)indices + start;
1245
1246                                 for(unsigned int i = 0; i < triangleCount; i++)
1247                                 {
1248                                         batch[i][0] = index[0];
1249                                         batch[i][1] = index[1];
1250                                         batch[i][2] = index[1];
1251
1252                                         index += 1;
1253                                 }
1254                         }
1255                         break;
1256                 case DRAW_INDEXEDLINESTRIP16:
1257                         {
1258                                 const unsigned short *index = (const unsigned short*)indices + start;
1259
1260                                 for(unsigned int i = 0; i < triangleCount; i++)
1261                                 {
1262                                         batch[i][0] = index[0];
1263                                         batch[i][1] = index[1];
1264                                         batch[i][2] = index[1];
1265
1266                                         index += 1;
1267                                 }
1268                         }
1269                         break;
1270                 case DRAW_INDEXEDLINESTRIP32:
1271                         {
1272                                 const unsigned int *index = (const unsigned int*)indices + start;
1273
1274                                 for(unsigned int i = 0; i < triangleCount; i++)
1275                                 {
1276                                         batch[i][0] = index[0];
1277                                         batch[i][1] = index[1];
1278                                         batch[i][2] = index[1];
1279
1280                                         index += 1;
1281                                 }
1282                         }
1283                         break;
1284                 case DRAW_INDEXEDLINELOOP8:
1285                         {
1286                                 const unsigned char *index = (const unsigned char*)indices;
1287
1288                                 for(unsigned int i = 0; i < triangleCount; i++)
1289                                 {
1290                                         batch[i][0] = index[(start + i + 0) % loop];
1291                                         batch[i][1] = index[(start + i + 1) % loop];
1292                                         batch[i][2] = index[(start + i + 1) % loop];
1293                                 }
1294                         }
1295                         break;
1296                 case DRAW_INDEXEDLINELOOP16:
1297                         {
1298                                 const unsigned short *index = (const unsigned short*)indices;
1299
1300                                 for(unsigned int i = 0; i < triangleCount; i++)
1301                                 {
1302                                         batch[i][0] = index[(start + i + 0) % loop];
1303                                         batch[i][1] = index[(start + i + 1) % loop];
1304                                         batch[i][2] = index[(start + i + 1) % loop];
1305                                 }
1306                         }
1307                         break;
1308                 case DRAW_INDEXEDLINELOOP32:
1309                         {
1310                                 const unsigned int *index = (const unsigned int*)indices;
1311
1312                                 for(unsigned int i = 0; i < triangleCount; i++)
1313                                 {
1314                                         batch[i][0] = index[(start + i + 0) % loop];
1315                                         batch[i][1] = index[(start + i + 1) % loop];
1316                                         batch[i][2] = index[(start + i + 1) % loop];
1317                                 }
1318                         }
1319                         break;
1320                 case DRAW_INDEXEDTRIANGLELIST8:
1321                         {
1322                                 const unsigned char *index = (const unsigned char*)indices + 3 * start;
1323
1324                                 for(unsigned int i = 0; i < triangleCount; i++)
1325                                 {
1326                                         batch[i][0] = index[0];
1327                                         batch[i][1] = index[1];
1328                                         batch[i][2] = index[2];
1329
1330                                         index += 3;
1331                                 }
1332                         }
1333                         break;
1334                 case DRAW_INDEXEDTRIANGLELIST16:
1335                         {
1336                                 const unsigned short *index = (const unsigned short*)indices + 3 * start;
1337
1338                                 for(unsigned int i = 0; i < triangleCount; i++)
1339                                 {
1340                                         batch[i][0] = index[0];
1341                                         batch[i][1] = index[1];
1342                                         batch[i][2] = index[2];
1343
1344                                         index += 3;
1345                                 }
1346                         }
1347                         break;
1348                 case DRAW_INDEXEDTRIANGLELIST32:
1349                         {
1350                                 const unsigned int *index = (const unsigned int*)indices + 3 * start;
1351
1352                                 for(unsigned int i = 0; i < triangleCount; i++)
1353                                 {
1354                                         batch[i][0] = index[0];
1355                                         batch[i][1] = index[1];
1356                                         batch[i][2] = index[2];
1357
1358                                         index += 3;
1359                                 }
1360                         }
1361                         break;
1362                 case DRAW_INDEXEDTRIANGLESTRIP8:
1363                         {
1364                                 const unsigned char *index = (const unsigned char*)indices + start;
1365
1366                                 for(unsigned int i = 0; i < triangleCount; i++)
1367                                 {
1368                                         batch[i][0] = index[0];
1369                                         batch[i][1] = index[((start + i) & 1) + 1];
1370                                         batch[i][2] = index[(~(start + i) & 1) + 1];
1371
1372                                         index += 1;
1373                                 }
1374                         }
1375                         break;
1376                 case DRAW_INDEXEDTRIANGLESTRIP16:
1377                         {
1378                                 const unsigned short *index = (const unsigned short*)indices + start;
1379
1380                                 for(unsigned int i = 0; i < triangleCount; i++)
1381                                 {
1382                                         batch[i][0] = index[0];
1383                                         batch[i][1] = index[((start + i) & 1) + 1];
1384                                         batch[i][2] = index[(~(start + i) & 1) + 1];
1385
1386                                         index += 1;
1387                                 }
1388                         }
1389                         break;
1390                 case DRAW_INDEXEDTRIANGLESTRIP32:
1391                         {
1392                                 const unsigned int *index = (const unsigned int*)indices + start;
1393
1394                                 for(unsigned int i = 0; i < triangleCount; i++)
1395                                 {
1396                                         batch[i][0] = index[0];
1397                                         batch[i][1] = index[((start + i) & 1) + 1];
1398                                         batch[i][2] = index[(~(start + i) & 1) + 1];
1399
1400                                         index += 1;
1401                                 }
1402                         }
1403                         break;
1404                 case DRAW_INDEXEDTRIANGLEFAN8:
1405                         {
1406                                 const unsigned char *index = (const unsigned char*)indices;
1407
1408                                 for(unsigned int i = 0; i < triangleCount; i++)
1409                                 {
1410                                         batch[i][0] = index[start + i + 1];
1411                                         batch[i][1] = index[start + i + 2];
1412                                         batch[i][2] = index[0];
1413                                 }
1414                         }
1415                         break;
1416                 case DRAW_INDEXEDTRIANGLEFAN16:
1417                         {
1418                                 const unsigned short *index = (const unsigned short*)indices;
1419
1420                                 for(unsigned int i = 0; i < triangleCount; i++)
1421                                 {
1422                                         batch[i][0] = index[start + i + 1];
1423                                         batch[i][1] = index[start + i + 2];
1424                                         batch[i][2] = index[0];
1425                                 }
1426                         }
1427                         break;
1428                 case DRAW_INDEXEDTRIANGLEFAN32:
1429                         {
1430                                 const unsigned int *index = (const unsigned int*)indices;
1431
1432                                 for(unsigned int i = 0; i < triangleCount; i++)
1433                                 {
1434                                         batch[i][0] = index[start + i + 1];
1435                                         batch[i][1] = index[start + i + 2];
1436                                         batch[i][2] = index[0];
1437                                 }
1438                         }
1439                         break;
1440                 case DRAW_QUADLIST:
1441                         {
1442                                 unsigned int index = 4 * start / 2;
1443
1444                                 for(unsigned int i = 0; i < triangleCount; i += 2)
1445                                 {
1446                                         batch[i+0][0] = index + 0;
1447                                         batch[i+0][1] = index + 1;
1448                                         batch[i+0][2] = index + 2;
1449
1450                                         batch[i+1][0] = index + 0;
1451                                         batch[i+1][1] = index + 2;
1452                                         batch[i+1][2] = index + 3;
1453
1454                                         index += 4;
1455                                 }
1456                         }
1457                         break;
1458                 default:
1459                         ASSERT(false);
1460                         return;
1461                 }
1462
1463                 task->primitiveStart = start;
1464                 task->vertexCount = triangleCount * 3;
1465                 vertexRoutine(&triangle->v0, (unsigned int*)&batch, task, data);
1466         }
1467
1468         int Renderer::setupSolidTriangles(Renderer *renderer, int unit, int count)
1469         {
1470                 Triangle *triangle = renderer->triangleBatch[unit];
1471                 Primitive *primitive = renderer->primitiveBatch[unit];
1472
1473                 DrawCall &draw = *renderer->drawList[renderer->primitiveProgress[unit].drawCall % DRAW_COUNT];
1474                 SetupProcessor::State &state = draw.setupState;
1475                 const SetupProcessor::RoutinePointer &setupRoutine = draw.setupPointer;
1476
1477                 int ms = state.multiSample;
1478                 int pos = state.positionRegister;
1479                 const DrawData *data = draw.data;
1480                 int visible = 0;
1481
1482                 for(int i = 0; i < count; i++, triangle++)
1483                 {
1484                         Vertex &v0 = triangle->v0;
1485                         Vertex &v1 = triangle->v1;
1486                         Vertex &v2 = triangle->v2;
1487
1488                         if((v0.clipFlags & v1.clipFlags & v2.clipFlags) == Clipper::CLIP_FINITE)
1489                         {
1490                                 Polygon polygon(&v0.v[pos], &v1.v[pos], &v2.v[pos]);
1491
1492                                 int clipFlagsOr = v0.clipFlags | v1.clipFlags | v2.clipFlags | draw.clipFlags;
1493
1494                                 if(clipFlagsOr != Clipper::CLIP_FINITE)
1495                                 {
1496                                         if(!renderer->clipper->clip(polygon, clipFlagsOr, draw))
1497                                         {
1498                                                 continue;
1499                                         }
1500                                 }
1501
1502                                 if(setupRoutine(primitive, triangle, &polygon, data))
1503                                 {
1504                                         primitive += ms;
1505                                         visible++;
1506                                 }
1507                         }
1508                 }
1509
1510                 return visible;
1511         }
1512
1513         int Renderer::setupWireframeTriangle(Renderer *renderer, int unit, int count)
1514         {
1515                 Triangle *triangle = renderer->triangleBatch[unit];
1516                 Primitive *primitive = renderer->primitiveBatch[unit];
1517                 int visible = 0;
1518
1519                 DrawCall &draw = *renderer->drawList[renderer->primitiveProgress[unit].drawCall % DRAW_COUNT];
1520                 SetupProcessor::State &state = draw.setupState;
1521                 SetupProcessor::RoutinePointer setupRoutine = draw.setupPointer;
1522
1523                 const Vertex &v0 = triangle[0].v0;
1524                 const Vertex &v1 = triangle[0].v1;
1525                 const Vertex &v2 = triangle[0].v2;
1526
1527                 float d = (v0.y * v1.x - v0.x * v1.y) * v2.w + (v0.x * v2.y - v0.y * v2.x) * v1.w + (v2.x * v1.y - v1.x * v2.y) * v0.w;
1528
1529                 if(state.cullMode == CULL_CLOCKWISE)
1530                 {
1531                         if(d >= 0) return 0;
1532                 }
1533                 else if(state.cullMode == CULL_COUNTERCLOCKWISE)
1534                 {
1535                         if(d <= 0) return 0;
1536                 }
1537
1538                 // Copy attributes
1539                 triangle[1].v0 = v1;
1540                 triangle[1].v1 = v2;
1541                 triangle[2].v0 = v2;
1542                 triangle[2].v1 = v0;
1543
1544                 if(state.color[0][0].flat)   // FIXME
1545                 {
1546                         for(int i = 0; i < 2; i++)
1547                         {
1548                                 triangle[1].v0.C[i] = triangle[0].v0.C[i];
1549                                 triangle[1].v1.C[i] = triangle[0].v0.C[i];
1550                                 triangle[2].v0.C[i] = triangle[0].v0.C[i];
1551                                 triangle[2].v1.C[i] = triangle[0].v0.C[i];
1552                         }
1553                 }
1554
1555                 for(int i = 0; i < 3; i++)
1556                 {
1557                         if(setupLine(renderer, *primitive, *triangle, draw))
1558                         {
1559                                 primitive->area = 0.5f * d;
1560
1561                                 primitive++;
1562                                 visible++;
1563                         }
1564
1565                         triangle++;
1566                 }
1567
1568                 return visible;
1569         }
1570
1571         int Renderer::setupVertexTriangle(Renderer *renderer, int unit, int count)
1572         {
1573                 Triangle *triangle = renderer->triangleBatch[unit];
1574                 Primitive *primitive = renderer->primitiveBatch[unit];
1575                 int visible = 0;
1576
1577                 DrawCall &draw = *renderer->drawList[renderer->primitiveProgress[unit].drawCall % DRAW_COUNT];
1578                 SetupProcessor::State &state = draw.setupState;
1579
1580                 const Vertex &v0 = triangle[0].v0;
1581                 const Vertex &v1 = triangle[0].v1;
1582                 const Vertex &v2 = triangle[0].v2;
1583
1584                 float d = (v0.y * v1.x - v0.x * v1.y) * v2.w + (v0.x * v2.y - v0.y * v2.x) * v1.w + (v2.x * v1.y - v1.x * v2.y) * v0.w;
1585
1586                 if(state.cullMode == CULL_CLOCKWISE)
1587                 {
1588                         if(d >= 0) return 0;
1589                 }
1590                 else if(state.cullMode == CULL_COUNTERCLOCKWISE)
1591                 {
1592                         if(d <= 0) return 0;
1593                 }
1594
1595                 // Copy attributes
1596                 triangle[1].v0 = v1;
1597                 triangle[2].v0 = v2;
1598
1599                 for(int i = 0; i < 3; i++)
1600                 {
1601                         if(setupPoint(renderer, *primitive, *triangle, draw))
1602                         {
1603                                 primitive->area = 0.5f * d;
1604
1605                                 primitive++;
1606                                 visible++;
1607                         }
1608
1609                         triangle++;
1610                 }
1611
1612                 return visible;
1613         }
1614
1615         int Renderer::setupLines(Renderer *renderer, int unit, int count)
1616         {
1617                 Triangle *triangle = renderer->triangleBatch[unit];
1618                 Primitive *primitive = renderer->primitiveBatch[unit];
1619                 int visible = 0;
1620
1621                 DrawCall &draw = *renderer->drawList[renderer->primitiveProgress[unit].drawCall % DRAW_COUNT];
1622                 SetupProcessor::State &state = draw.setupState;
1623
1624                 int ms = state.multiSample;
1625
1626                 for(int i = 0; i < count; i++)
1627                 {
1628                         if(setupLine(renderer, *primitive, *triangle, draw))
1629                         {
1630                                 primitive += ms;
1631                                 visible++;
1632                         }
1633
1634                         triangle++;
1635                 }
1636
1637                 return visible;
1638         }
1639
1640         int Renderer::setupPoints(Renderer *renderer, int unit, int count)
1641         {
1642                 Triangle *triangle = renderer->triangleBatch[unit];
1643                 Primitive *primitive = renderer->primitiveBatch[unit];
1644                 int visible = 0;
1645
1646                 DrawCall &draw = *renderer->drawList[renderer->primitiveProgress[unit].drawCall % DRAW_COUNT];
1647                 SetupProcessor::State &state = draw.setupState;
1648
1649                 int ms = state.multiSample;
1650
1651                 for(int i = 0; i < count; i++)
1652                 {
1653                         if(setupPoint(renderer, *primitive, *triangle, draw))
1654                         {
1655                                 primitive += ms;
1656                                 visible++;
1657                         }
1658
1659                         triangle++;
1660                 }
1661
1662                 return visible;
1663         }
1664
1665         bool Renderer::setupLine(Renderer *renderer, Primitive &primitive, Triangle &triangle, const DrawCall &draw)
1666         {
1667                 const SetupProcessor::RoutinePointer &setupRoutine = draw.setupPointer;
1668                 const SetupProcessor::State &state = draw.setupState;
1669                 const DrawData &data = *draw.data;
1670
1671                 float lineWidth = data.lineWidth;
1672
1673                 Vertex &v0 = triangle.v0;
1674                 Vertex &v1 = triangle.v1;
1675
1676                 int pos = state.positionRegister;
1677
1678                 const float4 &P0 = v0.v[pos];
1679                 const float4 &P1 = v1.v[pos];
1680
1681                 if(P0.w <= 0 && P1.w <= 0)
1682                 {
1683                         return false;
1684                 }
1685
1686                 const float W = data.Wx16[0] * (1.0f / 16.0f);
1687                 const float H = data.Hx16[0] * (1.0f / 16.0f);
1688
1689                 float dx = W * (P1.x / P1.w - P0.x / P0.w);
1690                 float dy = H * (P1.y / P1.w - P0.y / P0.w);
1691
1692                 if(dx == 0 && dy == 0)
1693                 {
1694                         return false;
1695                 }
1696
1697                 if(false)   // Rectangle
1698                 {
1699                         float4 P[4];
1700                         int C[4];
1701
1702                         P[0] = P0;
1703                         P[1] = P1;
1704                         P[2] = P1;
1705                         P[3] = P0;
1706
1707                         float scale = lineWidth * 0.5f / sqrt(dx*dx + dy*dy);
1708
1709                         dx *= scale;
1710                         dy *= scale;
1711
1712                         float dx0w = dx * P0.w / W;
1713                         float dy0h = dy * P0.w / H;
1714                         float dx0h = dx * P0.w / H;
1715                         float dy0w = dy * P0.w / W;
1716
1717                         float dx1w = dx * P1.w / W;
1718                         float dy1h = dy * P1.w / H;
1719                         float dx1h = dx * P1.w / H;
1720                         float dy1w = dy * P1.w / W;
1721
1722                         P[0].x += -dy0w + -dx0w;
1723                         P[0].y += -dx0h + +dy0h;
1724                         C[0] = computeClipFlags(P[0], data);
1725
1726                         P[1].x += -dy1w + +dx1w;
1727                         P[1].y += -dx1h + +dy1h;
1728                         C[1] = computeClipFlags(P[1], data);
1729
1730                         P[2].x += +dy1w + +dx1w;
1731                         P[2].y += +dx1h + -dy1h;
1732                         C[2] = computeClipFlags(P[2], data);
1733
1734                         P[3].x += +dy0w + -dx0w;
1735                         P[3].y += +dx0h + +dy0h;
1736                         C[3] = computeClipFlags(P[3], data);
1737
1738                         if((C[0] & C[1] & C[2] & C[3]) == Clipper::CLIP_FINITE)
1739                         {
1740                                 Polygon polygon(P, 4);
1741
1742                                 int clipFlagsOr = C[0] | C[1] | C[2] | C[3] | draw.clipFlags;
1743
1744                                 if(clipFlagsOr != Clipper::CLIP_FINITE)
1745                                 {
1746                                         if(!renderer->clipper->clip(polygon, clipFlagsOr, draw))
1747                                         {
1748                                                 return false;
1749                                         }
1750                                 }
1751
1752                                 return setupRoutine(&primitive, &triangle, &polygon, &data);
1753                         }
1754                 }
1755                 else   // Diamond test convention
1756                 {
1757                         float4 P[8];
1758                         int C[8];
1759
1760                         P[0] = P0;
1761                         P[1] = P0;
1762                         P[2] = P0;
1763                         P[3] = P0;
1764                         P[4] = P1;
1765                         P[5] = P1;
1766                         P[6] = P1;
1767                         P[7] = P1;
1768
1769                         float dx0 = lineWidth * 0.5f * P0.w / W;
1770                         float dy0 = lineWidth * 0.5f * P0.w / H;
1771
1772                         float dx1 = lineWidth * 0.5f * P1.w / W;
1773                         float dy1 = lineWidth * 0.5f * P1.w / H;
1774
1775                         P[0].x += -dx0;
1776                         C[0] = computeClipFlags(P[0], data);
1777
1778                         P[1].y += +dy0;
1779                         C[1] = computeClipFlags(P[1], data);
1780
1781                         P[2].x += +dx0;
1782                         C[2] = computeClipFlags(P[2], data);
1783
1784                         P[3].y += -dy0;
1785                         C[3] = computeClipFlags(P[3], data);
1786
1787                         P[4].x += -dx1;
1788                         C[4] = computeClipFlags(P[4], data);
1789
1790                         P[5].y += +dy1;
1791                         C[5] = computeClipFlags(P[5], data);
1792
1793                         P[6].x += +dx1;
1794                         C[6] = computeClipFlags(P[6], data);
1795
1796                         P[7].y += -dy1;
1797                         C[7] = computeClipFlags(P[7], data);
1798
1799                         if((C[0] & C[1] & C[2] & C[3] & C[4] & C[5] & C[6] & C[7]) == Clipper::CLIP_FINITE)
1800                         {
1801                                 float4 L[6];
1802
1803                                 if(dx > -dy)
1804                                 {
1805                                         if(dx > dy)   // Right
1806                                         {
1807                                                 L[0] = P[0];
1808                                                 L[1] = P[1];
1809                                                 L[2] = P[5];
1810                                                 L[3] = P[6];
1811                                                 L[4] = P[7];
1812                                                 L[5] = P[3];
1813                                         }
1814                                         else   // Down
1815                                         {
1816                                                 L[0] = P[0];
1817                                                 L[1] = P[4];
1818                                                 L[2] = P[5];
1819                                                 L[3] = P[6];
1820                                                 L[4] = P[2];
1821                                                 L[5] = P[3];
1822                                         }
1823                                 }
1824                                 else
1825                                 {
1826                                         if(dx > dy)   // Up
1827                                         {
1828                                                 L[0] = P[0];
1829                                                 L[1] = P[1];
1830                                                 L[2] = P[2];
1831                                                 L[3] = P[6];
1832                                                 L[4] = P[7];
1833                                                 L[5] = P[4];
1834                                         }
1835                                         else   // Left
1836                                         {
1837                                                 L[0] = P[1];
1838                                                 L[1] = P[2];
1839                                                 L[2] = P[3];
1840                                                 L[3] = P[7];
1841                                                 L[4] = P[4];
1842                                                 L[5] = P[5];
1843                                         }
1844                                 }
1845
1846                                 Polygon polygon(L, 6);
1847
1848                                 int clipFlagsOr = C[0] | C[1] | C[2] | C[3] | C[4] | C[5] | C[6] | C[7] | draw.clipFlags;
1849
1850                                 if(clipFlagsOr != Clipper::CLIP_FINITE)
1851                                 {
1852                                         if(!renderer->clipper->clip(polygon, clipFlagsOr, draw))
1853                                         {
1854                                                 return false;
1855                                         }
1856                                 }
1857
1858                                 return setupRoutine(&primitive, &triangle, &polygon, &data);
1859                         }
1860                 }
1861
1862                 return false;
1863         }
1864
1865         bool Renderer::setupPoint(Renderer *renderer, Primitive &primitive, Triangle &triangle, const DrawCall &draw)
1866         {
1867                 const SetupProcessor::RoutinePointer &setupRoutine = draw.setupPointer;
1868                 const SetupProcessor::State &state = draw.setupState;
1869                 const DrawData &data = *draw.data;
1870
1871                 Vertex &v = triangle.v0;
1872
1873                 float pSize;
1874
1875                 int pts = state.pointSizeRegister;
1876
1877                 if(state.pointSizeRegister != Unused)
1878                 {
1879                         pSize = v.v[pts].y;
1880                 }
1881                 else
1882                 {
1883                         pSize = data.point.pointSize[0];
1884                 }
1885
1886                 pSize = clamp(pSize, data.point.pointSizeMin, data.point.pointSizeMax);
1887
1888                 float4 P[4];
1889                 int C[4];
1890
1891                 int pos = state.positionRegister;
1892
1893                 P[0] = v.v[pos];
1894                 P[1] = v.v[pos];
1895                 P[2] = v.v[pos];
1896                 P[3] = v.v[pos];
1897
1898                 const float X = pSize * P[0].w * data.halfPixelX[0];
1899                 const float Y = pSize * P[0].w * data.halfPixelY[0];
1900
1901                 P[0].x -= X;
1902                 P[0].y += Y;
1903                 C[0] = computeClipFlags(P[0], data);
1904
1905                 P[1].x += X;
1906                 P[1].y += Y;
1907                 C[1] = computeClipFlags(P[1], data);
1908
1909                 P[2].x += X;
1910                 P[2].y -= Y;
1911                 C[2] = computeClipFlags(P[2], data);
1912
1913                 P[3].x -= X;
1914                 P[3].y -= Y;
1915                 C[3] = computeClipFlags(P[3], data);
1916
1917                 triangle.v1 = triangle.v0;
1918                 triangle.v2 = triangle.v0;
1919
1920                 triangle.v1.X += iround(16 * 0.5f * pSize);
1921                 triangle.v2.Y -= iround(16 * 0.5f * pSize) * (data.Hx16[0] > 0.0f ? 1 : -1);   // Both Direct3D and OpenGL expect (0, 0) in the top-left corner
1922
1923                 Polygon polygon(P, 4);
1924
1925                 if((C[0] & C[1] & C[2] & C[3]) == Clipper::CLIP_FINITE)
1926                 {
1927                         int clipFlagsOr = C[0] | C[1] | C[2] | C[3] | draw.clipFlags;
1928
1929                         if(clipFlagsOr != Clipper::CLIP_FINITE)
1930                         {
1931                                 if(!renderer->clipper->clip(polygon, clipFlagsOr, draw))
1932                                 {
1933                                         return false;
1934                                 }
1935                         }
1936
1937                         return setupRoutine(&primitive, &triangle, &polygon, &data);
1938                 }
1939
1940                 return false;
1941         }
1942
1943         unsigned int Renderer::computeClipFlags(const float4 &v, const DrawData &data)
1944         {
1945                 return ((v.x > v.w)  << 0) |
1946                        ((v.y > v.w)  << 1) |
1947                        ((v.z > v.w)  << 2) |
1948                        ((v.x < -v.w) << 3) |
1949                        ((v.y < -v.w) << 4) |
1950                        ((v.z < 0)    << 5) |
1951                        Clipper::CLIP_FINITE;   // FIXME: xyz finite
1952         }
1953
1954         void Renderer::initializeThreads()
1955         {
1956                 unitCount = ceilPow2(threadCount);
1957                 clusterCount = ceilPow2(threadCount);
1958
1959                 for(int i = 0; i < unitCount; i++)
1960                 {
1961                         triangleBatch[i] = (Triangle*)allocate(batchSize * sizeof(Triangle));
1962                         primitiveBatch[i] = (Primitive*)allocate(batchSize * sizeof(Primitive));
1963                 }
1964
1965                 for(int i = 0; i < threadCount; i++)
1966                 {
1967                         vertexTask[i] = (VertexTask*)allocate(sizeof(VertexTask));
1968                         vertexTask[i]->vertexCache.drawCall = -1;
1969
1970                         task[i].type = Task::SUSPEND;
1971
1972                         resume[i] = new Event();
1973                         suspend[i] = new Event();
1974
1975                         Parameters parameters;
1976                         parameters.threadIndex = i;
1977                         parameters.renderer = this;
1978
1979                         exitThreads = false;
1980                         worker[i] = new Thread(threadFunction, &parameters);
1981
1982                         suspend[i]->wait();
1983                         suspend[i]->signal();
1984                 }
1985         }
1986
1987         void Renderer::terminateThreads()
1988         {
1989                 while(threadsAwake != 0)
1990                 {
1991                         Thread::sleep(1);
1992                 }
1993
1994                 for(int thread = 0; thread < threadCount; thread++)
1995                 {
1996                         if(worker[thread])
1997                         {
1998                                 exitThreads = true;
1999                                 resume[thread]->signal();
2000                                 worker[thread]->join();
2001
2002                                 delete worker[thread];
2003                                 worker[thread] = 0;
2004                                 delete resume[thread];
2005                                 resume[thread] = 0;
2006                                 delete suspend[thread];
2007                                 suspend[thread] = 0;
2008                         }
2009
2010                         deallocate(vertexTask[thread]);
2011                         vertexTask[thread] = 0;
2012                 }
2013
2014                 for(int i = 0; i < 16; i++)
2015                 {
2016                         deallocate(triangleBatch[i]);
2017                         triangleBatch[i] = 0;
2018
2019                         deallocate(primitiveBatch[i]);
2020                         primitiveBatch[i] = 0;
2021                 }
2022         }
2023
2024         void Renderer::loadConstants(const VertexShader *vertexShader)
2025         {
2026                 if(!vertexShader) return;
2027
2028                 size_t count = vertexShader->getLength();
2029
2030                 for(size_t i = 0; i < count; i++)
2031                 {
2032                         const Shader::Instruction *instruction = vertexShader->getInstruction(i);
2033
2034                         if(instruction->opcode == Shader::OPCODE_DEF)
2035                         {
2036                                 int index = instruction->dst.index;
2037                                 float value[4];
2038
2039                                 value[0] = instruction->src[0].value[0];
2040                                 value[1] = instruction->src[0].value[1];
2041                                 value[2] = instruction->src[0].value[2];
2042                                 value[3] = instruction->src[0].value[3];
2043
2044                                 setVertexShaderConstantF(index, value);
2045                         }
2046                         else if(instruction->opcode == Shader::OPCODE_DEFI)
2047                         {
2048                                 int index = instruction->dst.index;
2049                                 int integer[4];
2050
2051                                 integer[0] = instruction->src[0].integer[0];
2052                                 integer[1] = instruction->src[0].integer[1];
2053                                 integer[2] = instruction->src[0].integer[2];
2054                                 integer[3] = instruction->src[0].integer[3];
2055
2056                                 setVertexShaderConstantI(index, integer);
2057                         }
2058                         else if(instruction->opcode == Shader::OPCODE_DEFB)
2059                         {
2060                                 int index = instruction->dst.index;
2061                                 int boolean = instruction->src[0].boolean[0];
2062
2063                                 setVertexShaderConstantB(index, &boolean);
2064                         }
2065                 }
2066         }
2067
2068         void Renderer::loadConstants(const PixelShader *pixelShader)
2069         {
2070                 if(!pixelShader) return;
2071
2072                 size_t count = pixelShader->getLength();
2073
2074                 for(size_t i = 0; i < count; i++)
2075                 {
2076                         const Shader::Instruction *instruction = pixelShader->getInstruction(i);
2077
2078                         if(instruction->opcode == Shader::OPCODE_DEF)
2079                         {
2080                                 int index = instruction->dst.index;
2081                                 float value[4];
2082
2083                                 value[0] = instruction->src[0].value[0];
2084                                 value[1] = instruction->src[0].value[1];
2085                                 value[2] = instruction->src[0].value[2];
2086                                 value[3] = instruction->src[0].value[3];
2087
2088                                 setPixelShaderConstantF(index, value);
2089                         }
2090                         else if(instruction->opcode == Shader::OPCODE_DEFI)
2091                         {
2092                                 int index = instruction->dst.index;
2093                                 int integer[4];
2094
2095                                 integer[0] = instruction->src[0].integer[0];
2096                                 integer[1] = instruction->src[0].integer[1];
2097                                 integer[2] = instruction->src[0].integer[2];
2098                                 integer[3] = instruction->src[0].integer[3];
2099
2100                                 setPixelShaderConstantI(index, integer);
2101                         }
2102                         else if(instruction->opcode == Shader::OPCODE_DEFB)
2103                         {
2104                                 int index = instruction->dst.index;
2105                                 int boolean = instruction->src[0].boolean[0];
2106
2107                                 setPixelShaderConstantB(index, &boolean);
2108                         }
2109                 }
2110         }
2111
2112         void Renderer::setIndexBuffer(Resource *indexBuffer)
2113         {
2114                 context->indexBuffer = indexBuffer;
2115         }
2116
2117         void Renderer::setMultiSampleMask(unsigned int mask)
2118         {
2119                 context->sampleMask = mask;
2120         }
2121
2122         void Renderer::setTransparencyAntialiasing(TransparencyAntialiasing transparencyAntialiasing)
2123         {
2124                 sw::transparencyAntialiasing = transparencyAntialiasing;
2125         }
2126
2127         bool Renderer::isReadWriteTexture(int sampler)
2128         {
2129                 for(int index = 0; index < RENDERTARGETS; index++)
2130                 {
2131                         if(context->renderTarget[index] && context->texture[sampler] == context->renderTarget[index]->getResource())
2132                         {
2133                                 return true;
2134                         }
2135                 }
2136
2137                 if(context->depthBuffer && context->texture[sampler] == context->depthBuffer->getResource())
2138                 {
2139                         return true;
2140                 }
2141
2142                 return false;
2143         }
2144
2145         void Renderer::updateClipper()
2146         {
2147                 if(updateClipPlanes)
2148                 {
2149                         if(VertexProcessor::isFixedFunction())   // User plane in world space
2150                         {
2151                                 const Matrix &scissorWorld = getViewTransform();
2152
2153                                 if(clipFlags & Clipper::CLIP_PLANE0) clipPlane[0] = scissorWorld * userPlane[0];
2154                                 if(clipFlags & Clipper::CLIP_PLANE1) clipPlane[1] = scissorWorld * userPlane[1];
2155                                 if(clipFlags & Clipper::CLIP_PLANE2) clipPlane[2] = scissorWorld * userPlane[2];
2156                                 if(clipFlags & Clipper::CLIP_PLANE3) clipPlane[3] = scissorWorld * userPlane[3];
2157                                 if(clipFlags & Clipper::CLIP_PLANE4) clipPlane[4] = scissorWorld * userPlane[4];
2158                                 if(clipFlags & Clipper::CLIP_PLANE5) clipPlane[5] = scissorWorld * userPlane[5];
2159                         }
2160                         else   // User plane in clip space
2161                         {
2162                                 if(clipFlags & Clipper::CLIP_PLANE0) clipPlane[0] = userPlane[0];
2163                                 if(clipFlags & Clipper::CLIP_PLANE1) clipPlane[1] = userPlane[1];
2164                                 if(clipFlags & Clipper::CLIP_PLANE2) clipPlane[2] = userPlane[2];
2165                                 if(clipFlags & Clipper::CLIP_PLANE3) clipPlane[3] = userPlane[3];
2166                                 if(clipFlags & Clipper::CLIP_PLANE4) clipPlane[4] = userPlane[4];
2167                                 if(clipFlags & Clipper::CLIP_PLANE5) clipPlane[5] = userPlane[5];
2168                         }
2169
2170                         updateClipPlanes = false;
2171                 }
2172         }
2173
2174         void Renderer::setTextureResource(unsigned int sampler, Resource *resource)
2175         {
2176                 ASSERT(sampler < TOTAL_IMAGE_UNITS);
2177
2178                 context->texture[sampler] = resource;
2179         }
2180
2181         void Renderer::setTextureLevel(unsigned int sampler, unsigned int face, unsigned int level, Surface *surface, TextureType type)
2182         {
2183                 ASSERT(sampler < TOTAL_IMAGE_UNITS && face < 6 && level < MIPMAP_LEVELS);
2184
2185                 context->sampler[sampler].setTextureLevel(face, level, surface, type);
2186         }
2187
2188         void Renderer::setTextureFilter(SamplerType type, int sampler, FilterType textureFilter)
2189         {
2190                 if(type == SAMPLER_PIXEL)
2191                 {
2192                         PixelProcessor::setTextureFilter(sampler, textureFilter);
2193                 }
2194                 else
2195                 {
2196                         VertexProcessor::setTextureFilter(sampler, textureFilter);
2197                 }
2198         }
2199
2200         void Renderer::setMipmapFilter(SamplerType type, int sampler, MipmapType mipmapFilter)
2201         {
2202                 if(type == SAMPLER_PIXEL)
2203                 {
2204                         PixelProcessor::setMipmapFilter(sampler, mipmapFilter);
2205                 }
2206                 else
2207                 {
2208                         VertexProcessor::setMipmapFilter(sampler, mipmapFilter);
2209                 }
2210         }
2211
2212         void Renderer::setGatherEnable(SamplerType type, int sampler, bool enable)
2213         {
2214                 if(type == SAMPLER_PIXEL)
2215                 {
2216                         PixelProcessor::setGatherEnable(sampler, enable);
2217                 }
2218                 else
2219                 {
2220                         VertexProcessor::setGatherEnable(sampler, enable);
2221                 }
2222         }
2223
2224         void Renderer::setAddressingModeU(SamplerType type, int sampler, AddressingMode addressMode)
2225         {
2226                 if(type == SAMPLER_PIXEL)
2227                 {
2228                         PixelProcessor::setAddressingModeU(sampler, addressMode);
2229                 }
2230                 else
2231                 {
2232                         VertexProcessor::setAddressingModeU(sampler, addressMode);
2233                 }
2234         }
2235
2236         void Renderer::setAddressingModeV(SamplerType type, int sampler, AddressingMode addressMode)
2237         {
2238                 if(type == SAMPLER_PIXEL)
2239                 {
2240                         PixelProcessor::setAddressingModeV(sampler, addressMode);
2241                 }
2242                 else
2243                 {
2244                         VertexProcessor::setAddressingModeV(sampler, addressMode);
2245                 }
2246         }
2247
2248         void Renderer::setAddressingModeW(SamplerType type, int sampler, AddressingMode addressMode)
2249         {
2250                 if(type == SAMPLER_PIXEL)
2251                 {
2252                         PixelProcessor::setAddressingModeW(sampler, addressMode);
2253                 }
2254                 else
2255                 {
2256                         VertexProcessor::setAddressingModeW(sampler, addressMode);
2257                 }
2258         }
2259
2260         void Renderer::setReadSRGB(SamplerType type, int sampler, bool sRGB)
2261         {
2262                 if(type == SAMPLER_PIXEL)
2263                 {
2264                         PixelProcessor::setReadSRGB(sampler, sRGB);
2265                 }
2266                 else
2267                 {
2268                         VertexProcessor::setReadSRGB(sampler, sRGB);
2269                 }
2270         }
2271
2272         void Renderer::setMipmapLOD(SamplerType type, int sampler, float bias)
2273         {
2274                 if(type == SAMPLER_PIXEL)
2275                 {
2276                         PixelProcessor::setMipmapLOD(sampler, bias);
2277                 }
2278                 else
2279                 {
2280                         VertexProcessor::setMipmapLOD(sampler, bias);
2281                 }
2282         }
2283
2284         void Renderer::setBorderColor(SamplerType type, int sampler, const Color<float> &borderColor)
2285         {
2286                 if(type == SAMPLER_PIXEL)
2287                 {
2288                         PixelProcessor::setBorderColor(sampler, borderColor);
2289                 }
2290                 else
2291                 {
2292                         VertexProcessor::setBorderColor(sampler, borderColor);
2293                 }
2294         }
2295
2296         void Renderer::setMaxAnisotropy(SamplerType type, int sampler, float maxAnisotropy)
2297         {
2298                 if(type == SAMPLER_PIXEL)
2299                 {
2300                         PixelProcessor::setMaxAnisotropy(sampler, maxAnisotropy);
2301                 }
2302                 else
2303                 {
2304                         VertexProcessor::setMaxAnisotropy(sampler, maxAnisotropy);
2305                 }
2306         }
2307
2308         void Renderer::setSwizzleR(SamplerType type, int sampler, SwizzleType swizzleR)
2309         {
2310                 if(type == SAMPLER_PIXEL)
2311                 {
2312                         PixelProcessor::setSwizzleR(sampler, swizzleR);
2313                 }
2314                 else
2315                 {
2316                         VertexProcessor::setSwizzleR(sampler, swizzleR);
2317                 }
2318         }
2319
2320         void Renderer::setSwizzleG(SamplerType type, int sampler, SwizzleType swizzleG)
2321         {
2322                 if(type == SAMPLER_PIXEL)
2323                 {
2324                         PixelProcessor::setSwizzleG(sampler, swizzleG);
2325                 }
2326                 else
2327                 {
2328                         VertexProcessor::setSwizzleG(sampler, swizzleG);
2329                 }
2330         }
2331
2332         void Renderer::setSwizzleB(SamplerType type, int sampler, SwizzleType swizzleB)
2333         {
2334                 if(type == SAMPLER_PIXEL)
2335                 {
2336                         PixelProcessor::setSwizzleB(sampler, swizzleB);
2337                 }
2338                 else
2339                 {
2340                         VertexProcessor::setSwizzleB(sampler, swizzleB);
2341                 }
2342         }
2343
2344         void Renderer::setSwizzleA(SamplerType type, int sampler, SwizzleType swizzleA)
2345         {
2346                 if(type == SAMPLER_PIXEL)
2347                 {
2348                         PixelProcessor::setSwizzleA(sampler, swizzleA);
2349                 }
2350                 else
2351                 {
2352                         VertexProcessor::setSwizzleA(sampler, swizzleA);
2353                 }
2354         }
2355
2356         void Renderer::setPointSpriteEnable(bool pointSpriteEnable)
2357         {
2358                 context->setPointSpriteEnable(pointSpriteEnable);
2359         }
2360
2361         void Renderer::setPointScaleEnable(bool pointScaleEnable)
2362         {
2363                 context->setPointScaleEnable(pointScaleEnable);
2364         }
2365
2366         void Renderer::setLineWidth(float width)
2367         {
2368                 context->lineWidth = width;
2369         }
2370
2371         void Renderer::setDepthBias(float bias)
2372         {
2373                 depthBias = bias;
2374         }
2375
2376         void Renderer::setSlopeDepthBias(float slopeBias)
2377         {
2378                 slopeDepthBias = slopeBias;
2379         }
2380
2381         void Renderer::setRasterizerDiscard(bool rasterizerDiscard)
2382         {
2383                 context->rasterizerDiscard = rasterizerDiscard;
2384         }
2385
2386         void Renderer::setPixelShader(const PixelShader *shader)
2387         {
2388                 context->pixelShader = shader;
2389
2390                 loadConstants(shader);
2391         }
2392
2393         void Renderer::setVertexShader(const VertexShader *shader)
2394         {
2395                 context->vertexShader = shader;
2396
2397                 loadConstants(shader);
2398         }
2399
2400         void Renderer::setPixelShaderConstantF(int index, const float value[4], int count)
2401         {
2402                 for(int i = 0; i < DRAW_COUNT; i++)
2403                 {
2404                         if(drawCall[i]->psDirtyConstF < index + count)
2405                         {
2406                                 drawCall[i]->psDirtyConstF = index + count;
2407                         }
2408                 }
2409
2410                 for(int i = 0; i < count; i++)
2411                 {
2412                         PixelProcessor::setFloatConstant(index + i, value);
2413                         value += 4;
2414                 }
2415         }
2416
2417         void Renderer::setPixelShaderConstantI(int index, const int value[4], int count)
2418         {
2419                 for(int i = 0; i < DRAW_COUNT; i++)
2420                 {
2421                         if(drawCall[i]->psDirtyConstI < index + count)
2422                         {
2423                                 drawCall[i]->psDirtyConstI = index + count;
2424                         }
2425                 }
2426
2427                 for(int i = 0; i < count; i++)
2428                 {
2429                         PixelProcessor::setIntegerConstant(index + i, value);
2430                         value += 4;
2431                 }
2432         }
2433
2434         void Renderer::setPixelShaderConstantB(int index, const int *boolean, int count)
2435         {
2436                 for(int i = 0; i < DRAW_COUNT; i++)
2437                 {
2438                         if(drawCall[i]->psDirtyConstB < index + count)
2439                         {
2440                                 drawCall[i]->psDirtyConstB = index + count;
2441                         }
2442                 }
2443
2444                 for(int i = 0; i < count; i++)
2445                 {
2446                         PixelProcessor::setBooleanConstant(index + i, *boolean);
2447                         boolean++;
2448                 }
2449         }
2450
2451         void Renderer::setVertexShaderConstantF(int index, const float value[4], int count)
2452         {
2453                 for(int i = 0; i < DRAW_COUNT; i++)
2454                 {
2455                         if(drawCall[i]->vsDirtyConstF < index + count)
2456                         {
2457                                 drawCall[i]->vsDirtyConstF = index + count;
2458                         }
2459                 }
2460
2461                 for(int i = 0; i < count; i++)
2462                 {
2463                         VertexProcessor::setFloatConstant(index + i, value);
2464                         value += 4;
2465                 }
2466         }
2467
2468         void Renderer::setVertexShaderConstantI(int index, const int value[4], int count)
2469         {
2470                 for(int i = 0; i < DRAW_COUNT; i++)
2471                 {
2472                         if(drawCall[i]->vsDirtyConstI < index + count)
2473                         {
2474                                 drawCall[i]->vsDirtyConstI = index + count;
2475                         }
2476                 }
2477
2478                 for(int i = 0; i < count; i++)
2479                 {
2480                         VertexProcessor::setIntegerConstant(index + i, value);
2481                         value += 4;
2482                 }
2483         }
2484
2485         void Renderer::setVertexShaderConstantB(int index, const int *boolean, int count)
2486         {
2487                 for(int i = 0; i < DRAW_COUNT; i++)
2488                 {
2489                         if(drawCall[i]->vsDirtyConstB < index + count)
2490                         {
2491                                 drawCall[i]->vsDirtyConstB = index + count;
2492                         }
2493                 }
2494
2495                 for(int i = 0; i < count; i++)
2496                 {
2497                         VertexProcessor::setBooleanConstant(index + i, *boolean);
2498                         boolean++;
2499                 }
2500         }
2501
2502         void Renderer::setModelMatrix(const Matrix &M, int i)
2503         {
2504                 VertexProcessor::setModelMatrix(M, i);
2505         }
2506
2507         void Renderer::setViewMatrix(const Matrix &V)
2508         {
2509                 VertexProcessor::setViewMatrix(V);
2510                 updateClipPlanes = true;
2511         }
2512
2513         void Renderer::setBaseMatrix(const Matrix &B)
2514         {
2515                 VertexProcessor::setBaseMatrix(B);
2516                 updateClipPlanes = true;
2517         }
2518
2519         void Renderer::setProjectionMatrix(const Matrix &P)
2520         {
2521                 VertexProcessor::setProjectionMatrix(P);
2522                 updateClipPlanes = true;
2523         }
2524
2525         void Renderer::addQuery(Query *query)
2526         {
2527                 queries.push_back(query);
2528         }
2529
2530         void Renderer::removeQuery(Query *query)
2531         {
2532                 queries.remove(query);
2533         }
2534
2535         #if PERF_HUD
2536                 int Renderer::getThreadCount()
2537                 {
2538                         return threadCount;
2539                 }
2540
2541                 int64_t Renderer::getVertexTime(int thread)
2542                 {
2543                         return vertexTime[thread];
2544                 }
2545
2546                 int64_t Renderer::getSetupTime(int thread)
2547                 {
2548                         return setupTime[thread];
2549                 }
2550
2551                 int64_t Renderer::getPixelTime(int thread)
2552                 {
2553                         return pixelTime[thread];
2554                 }
2555
2556                 void Renderer::resetTimers()
2557                 {
2558                         for(int thread = 0; thread < threadCount; thread++)
2559                         {
2560                                 vertexTime[thread] = 0;
2561                                 setupTime[thread] = 0;
2562                                 pixelTime[thread] = 0;
2563                         }
2564                 }
2565         #endif
2566
2567         void Renderer::setViewport(const Viewport &viewport)
2568         {
2569                 this->viewport = viewport;
2570         }
2571
2572         void Renderer::setScissor(const Rect &scissor)
2573         {
2574                 this->scissor = scissor;
2575         }
2576
2577         void Renderer::setClipFlags(int flags)
2578         {
2579                 clipFlags = flags << 8;   // Bottom 8 bits used by legacy frustum
2580         }
2581
2582         void Renderer::setClipPlane(unsigned int index, const float plane[4])
2583         {
2584                 if(index < MAX_CLIP_PLANES)
2585                 {
2586                         userPlane[index] = plane;
2587                 }
2588                 else ASSERT(false);
2589
2590                 updateClipPlanes = true;
2591         }
2592
2593         void Renderer::updateConfiguration(bool initialUpdate)
2594         {
2595                 bool newConfiguration = swiftConfig->hasNewConfiguration();
2596
2597                 if(newConfiguration || initialUpdate)
2598                 {
2599                         terminateThreads();
2600
2601                         SwiftConfig::Configuration configuration = {};
2602                         swiftConfig->getConfiguration(configuration);
2603
2604                         precacheVertex = !newConfiguration && configuration.precache;
2605                         precacheSetup = !newConfiguration && configuration.precache;
2606                         precachePixel = !newConfiguration && configuration.precache;
2607
2608                         VertexProcessor::setRoutineCacheSize(configuration.vertexRoutineCacheSize);
2609                         PixelProcessor::setRoutineCacheSize(configuration.pixelRoutineCacheSize);
2610                         SetupProcessor::setRoutineCacheSize(configuration.setupRoutineCacheSize);
2611
2612                         switch(configuration.textureSampleQuality)
2613                         {
2614                         case 0:  Sampler::setFilterQuality(FILTER_POINT);       break;
2615                         case 1:  Sampler::setFilterQuality(FILTER_LINEAR);      break;
2616                         case 2:  Sampler::setFilterQuality(FILTER_ANISOTROPIC); break;
2617                         default: Sampler::setFilterQuality(FILTER_ANISOTROPIC); break;
2618                         }
2619
2620                         switch(configuration.mipmapQuality)
2621                         {
2622                         case 0:  Sampler::setMipmapQuality(MIPMAP_POINT);  break;
2623                         case 1:  Sampler::setMipmapQuality(MIPMAP_LINEAR); break;
2624                         default: Sampler::setMipmapQuality(MIPMAP_LINEAR); break;
2625                         }
2626
2627                         setPerspectiveCorrection(configuration.perspectiveCorrection);
2628
2629                         switch(configuration.transcendentalPrecision)
2630                         {
2631                         case 0:
2632                                 logPrecision = APPROXIMATE;
2633                                 expPrecision = APPROXIMATE;
2634                                 rcpPrecision = APPROXIMATE;
2635                                 rsqPrecision = APPROXIMATE;
2636                                 break;
2637                         case 1:
2638                                 logPrecision = PARTIAL;
2639                                 expPrecision = PARTIAL;
2640                                 rcpPrecision = PARTIAL;
2641                                 rsqPrecision = PARTIAL;
2642                                 break;
2643                         case 2:
2644                                 logPrecision = ACCURATE;
2645                                 expPrecision = ACCURATE;
2646                                 rcpPrecision = ACCURATE;
2647                                 rsqPrecision = ACCURATE;
2648                                 break;
2649                         case 3:
2650                                 logPrecision = WHQL;
2651                                 expPrecision = WHQL;
2652                                 rcpPrecision = WHQL;
2653                                 rsqPrecision = WHQL;
2654                                 break;
2655                         case 4:
2656                                 logPrecision = IEEE;
2657                                 expPrecision = IEEE;
2658                                 rcpPrecision = IEEE;
2659                                 rsqPrecision = IEEE;
2660                                 break;
2661                         default:
2662                                 logPrecision = ACCURATE;
2663                                 expPrecision = ACCURATE;
2664                                 rcpPrecision = ACCURATE;
2665                                 rsqPrecision = ACCURATE;
2666                                 break;
2667                         }
2668
2669                         switch(configuration.transparencyAntialiasing)
2670                         {
2671                         case 0:  transparencyAntialiasing = TRANSPARENCY_NONE;              break;
2672                         case 1:  transparencyAntialiasing = TRANSPARENCY_ALPHA_TO_COVERAGE; break;
2673                         default: transparencyAntialiasing = TRANSPARENCY_NONE;              break;
2674                         }
2675
2676                         switch(configuration.threadCount)
2677                         {
2678                         case -1: threadCount = CPUID::coreCount();        break;
2679                         case 0:  threadCount = CPUID::processAffinity();  break;
2680                         default: threadCount = configuration.threadCount; break;
2681                         }
2682
2683                         CPUID::setEnableSSE4_1(configuration.enableSSE4_1);
2684                         CPUID::setEnableSSSE3(configuration.enableSSSE3);
2685                         CPUID::setEnableSSE3(configuration.enableSSE3);
2686                         CPUID::setEnableSSE2(configuration.enableSSE2);
2687                         CPUID::setEnableSSE(configuration.enableSSE);
2688
2689                         for(int pass = 0; pass < 10; pass++)
2690                         {
2691                                 optimization[pass] = configuration.optimization[pass];
2692                         }
2693
2694                         forceWindowed = configuration.forceWindowed;
2695                         complementaryDepthBuffer = configuration.complementaryDepthBuffer;
2696                         postBlendSRGB = configuration.postBlendSRGB;
2697                         exactColorRounding = configuration.exactColorRounding;
2698                         forceClearRegisters = configuration.forceClearRegisters;
2699
2700                 #ifndef NDEBUG
2701                         minPrimitives = configuration.minPrimitives;
2702                         maxPrimitives = configuration.maxPrimitives;
2703                 #endif
2704                 }
2705
2706                 if(!initialUpdate && !worker[0])
2707                 {
2708                         initializeThreads();
2709                 }
2710         }
2711 }