src/gallium/drivers/swr/swr_shader.cpp

   1 /****************************************************************************
   2  * Copyright (C) 2015 Intel Corporation.   All Rights Reserved.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  ***************************************************************************/
  23
  24 #include "JitManager.h"
  25 #include "state.h"
  26 #include "state_llvm.h"
  27 #include "builder.h"
  28
  29 #include "llvm-c/Core.h"
  30 #include "llvm/Support/CBindingWrapping.h"
  31
  32 #include "tgsi/tgsi_strings.h"
  33 #include "gallivm/lp_bld_init.h"
  34 #include "gallivm/lp_bld_flow.h"
  35 #include "gallivm/lp_bld_struct.h"
  36 #include "gallivm/lp_bld_tgsi.h"
  37
  38 #include "swr_context.h"
  39 #include "swr_context_llvm.h"
  40 #include "swr_state.h"
  41 #include "swr_screen.h"
  42
  43 bool operator==(const swr_jit_key &lhs, const swr_jit_key &rhs)
  44 {
  45    return !memcmp(&lhs, &rhs, sizeof(lhs));
  46 }
  47
  48 void
  49 swr_generate_fs_key(struct swr_jit_key &key,
  50                     struct swr_context *ctx,
  51                     swr_fragment_shader *swr_fs)
  52 {
  53    key.nr_cbufs = ctx->framebuffer.nr_cbufs;
  54    key.light_twoside = ctx->rasterizer->light_twoside;
  55    memcpy(&key.vs_output_semantic_name,
  56           &ctx->vs->info.base.output_semantic_name,
  57           sizeof(key.vs_output_semantic_name));
  58    memcpy(&key.vs_output_semantic_idx,
  59           &ctx->vs->info.base.output_semantic_index,
  60           sizeof(key.vs_output_semantic_idx));
  61
  62    key.nr_samplers = swr_fs->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
  63
  64    for (unsigned i = 0; i < key.nr_samplers; i++) {
  65       if (swr_fs->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
  66          lp_sampler_static_sampler_state(
  67             &key.sampler[i].sampler_state,
  68             ctx->samplers[PIPE_SHADER_FRAGMENT][i]);
  69       }
  70    }
  71
  72    /*
  73     * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
  74     * are dx10-style? Can't really have mixed opcodes, at least not
  75     * if we want to skip the holes here (without rescanning tgsi).
  76     */
  77    if (swr_fs->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
  78       key.nr_sampler_views =
  79          swr_fs->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
  80       for (unsigned i = 0; i < key.nr_sampler_views; i++) {
  81          if (swr_fs->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1 << i)) {
  82             lp_sampler_static_texture_state(
  83                &key.sampler[i].texture_state,
  84                ctx->sampler_views[PIPE_SHADER_FRAGMENT][i]);
  85          }
  86       }
  87    } else {
  88       key.nr_sampler_views = key.nr_samplers;
  89       for (unsigned i = 0; i < key.nr_sampler_views; i++) {
  90          if (swr_fs->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
  91             lp_sampler_static_texture_state(
  92                &key.sampler[i].texture_state,
  93                ctx->sampler_views[PIPE_SHADER_FRAGMENT][i]);
  94          }
  95       }
  96    }
  97 }
  98
  99 struct BuilderSWR : public Builder {
 100    BuilderSWR(JitManager *pJitMgr)
 101       : Builder(pJitMgr)
 102    {
 103       pJitMgr->SetupNewModule();
 104    }
 105
 106    PFN_VERTEX_FUNC
 107    CompileVS(struct pipe_context *ctx, swr_vertex_shader *swr_vs);
 108    PFN_PIXEL_KERNEL CompileFS(struct swr_context *ctx, swr_jit_key &key);
 109 };
 110
 111 PFN_VERTEX_FUNC
 112 BuilderSWR::CompileVS(struct pipe_context *ctx, swr_vertex_shader *swr_vs)
 113 {
 114    swr_vs->linkageMask = 0;
 115
 116    for (unsigned i = 0; i < swr_vs->info.base.num_outputs; i++) {
 117       switch (swr_vs->info.base.output_semantic_name[i]) {
 118       case TGSI_SEMANTIC_POSITION:
 119          break;
 120       default:
 121          swr_vs->linkageMask |= (1 << i);
 122          break;
 123       }
 124    }
 125
 126    //   tgsi_dump(swr_vs->pipe.tokens, 0);
 127
 128    struct gallivm_state *gallivm =
 129       gallivm_create("VS", wrap(&JM()->mContext));
 130    gallivm->module = wrap(JM()->mpCurrentModule);
 131
 132    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 133    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 134
 135    memset(outputs, 0, sizeof(outputs));
 136
 137    AttrBuilder attrBuilder;
 138    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 139    AttributeSet attrSet = AttributeSet::get(
 140       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 141
 142    std::vector<Type *> vsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 143                               PointerType::get(Gen_SWR_VS_CONTEXT(JM()), 0)};
 144    FunctionType *vsFuncType =
 145       FunctionType::get(Type::getVoidTy(JM()->mContext), vsArgs, false);
 146
 147    // create new vertex shader function
 148    auto pFunction = Function::Create(vsFuncType,
 149                                      GlobalValue::ExternalLinkage,
 150                                      "VS",
 151                                      JM()->mpCurrentModule);
 152    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 153
 154    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 155    IRB()->SetInsertPoint(block);
 156    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 157
 158    auto argitr = pFunction->arg_begin();
 159    Value *hPrivateData = &*argitr++;
 160    hPrivateData->setName("hPrivateData");
 161    Value *pVsCtx = &*argitr++;
 162    pVsCtx->setName("vsCtx");
 163
 164    Value *consts_ptr = GEP(hPrivateData, {C(0), C(swr_draw_context_constantVS)});
 165
 166    consts_ptr->setName("vs_constants");
 167    Value *const_sizes_ptr =
 168       GEP(hPrivateData, {0, swr_draw_context_num_constantsVS});
 169    const_sizes_ptr->setName("num_vs_constants");
 170
 171    Value *vtxInput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVin});
 172
 173    for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
 174       const unsigned mask = swr_vs->info.base.input_usage_mask[attrib];
 175       for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 176          if (mask & (1 << channel)) {
 177             inputs[attrib][channel] =
 178                wrap(LOAD(vtxInput, {0, 0, attrib, channel}));
 179          }
 180       }
 181    }
 182
 183    struct lp_bld_tgsi_system_values system_values;
 184    memset(&system_values, 0, sizeof(system_values));
 185    system_values.instance_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_InstanceID}));
 186    system_values.vertex_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_VertexID}));
 187
 188    lp_build_tgsi_soa(gallivm,
 189                      swr_vs->pipe.tokens,
 190                      lp_type_float_vec(32, 32 * 8),
 191                      NULL, // mask
 192                      wrap(consts_ptr),
 193                      wrap(const_sizes_ptr),
 194                      &system_values,
 195                      inputs,
 196                      outputs,
 197                      NULL, // wrap(hPrivateData), (sampler context)
 198                      NULL, // thread data
 199                      NULL, // sampler
 200                      &swr_vs->info.base,
 201                      NULL); // geometry shader face
 202
 203    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 204
 205    Value *vtxOutput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVout});
 206
 207    for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 208       for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_OUTPUTS; attrib++) {
 209          if (!outputs[attrib][channel])
 210             continue;
 211
 212          Value *val = LOAD(unwrap(outputs[attrib][channel]));
 213
 214          uint32_t outSlot = attrib;
 215          if (swr_vs->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_PSIZE)
 216             outSlot = VERTEX_POINT_SIZE_SLOT;
 217          STORE(val, vtxOutput, {0, 0, outSlot, channel});
 218       }
 219    }
 220
 221    RET_VOID();
 222
 223    gallivm_verify_function(gallivm, wrap(pFunction));
 224    gallivm_compile_module(gallivm);
 225
 226    //   lp_debug_dump_value(func);
 227
 228    PFN_VERTEX_FUNC pFunc =
 229       (PFN_VERTEX_FUNC)gallivm_jit_function(gallivm, wrap(pFunction));
 230
 231    debug_printf("vert shader  %p\n", pFunc);
 232    assert(pFunc && "Error: VertShader = NULL");
 233
 234 #if (LLVM_VERSION_MAJOR == 3) && (LLVM_VERSION_MINOR >= 5)
 235    JM()->mIsModuleFinalized = true;
 236 #endif
 237
 238    return pFunc;
 239 }
 240
 241 PFN_VERTEX_FUNC
 242 swr_compile_vs(struct pipe_context *ctx, swr_vertex_shader *swr_vs)
 243 {
 244    BuilderSWR builder(
 245       reinterpret_cast<JitManager *>(swr_screen(ctx->screen)->hJitMgr));
 246    return builder.CompileVS(ctx, swr_vs);
 247 }
 248
 249 static unsigned
 250 locate_linkage(ubyte name, ubyte index, struct tgsi_shader_info *info)
 251 {
 252    for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 253       if ((info->output_semantic_name[i] == name)
 254           && (info->output_semantic_index[i] == index)) {
 255          return i - 1; // position is not part of the linkage
 256       }
 257    }
 258
 259    if (name == TGSI_SEMANTIC_COLOR) { // BCOLOR fallback
 260       for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 261          if ((info->output_semantic_name[i] == TGSI_SEMANTIC_BCOLOR)
 262              && (info->output_semantic_index[i] == index)) {
 263             return i - 1; // position is not part of the linkage
 264          }
 265       }
 266    }
 267
 268    return 0xFFFFFFFF;
 269 }
 270
 271 PFN_PIXEL_KERNEL
 272 BuilderSWR::CompileFS(struct swr_context *ctx, swr_jit_key &key)
 273 {
 274    struct swr_fragment_shader *swr_fs = ctx->fs;
 275
 276    //   tgsi_dump(swr_fs->pipe.tokens, 0);
 277
 278    struct gallivm_state *gallivm =
 279       gallivm_create("FS", wrap(&JM()->mContext));
 280    gallivm->module = wrap(JM()->mpCurrentModule);
 281
 282    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 283    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 284
 285    memset(inputs, 0, sizeof(inputs));
 286    memset(outputs, 0, sizeof(outputs));
 287
 288    struct lp_build_sampler_soa *sampler = NULL;
 289
 290    AttrBuilder attrBuilder;
 291    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 292    AttributeSet attrSet = AttributeSet::get(
 293       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 294
 295    std::vector<Type *> fsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 296                               PointerType::get(Gen_SWR_PS_CONTEXT(JM()), 0)};
 297    FunctionType *funcType =
 298       FunctionType::get(Type::getVoidTy(JM()->mContext), fsArgs, false);
 299
 300    auto pFunction = Function::Create(funcType,
 301                                      GlobalValue::ExternalLinkage,
 302                                      "FS",
 303                                      JM()->mpCurrentModule);
 304    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 305
 306    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 307    IRB()->SetInsertPoint(block);
 308    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 309
 310    auto args = pFunction->arg_begin();
 311    Value *hPrivateData = &*args++;
 312    hPrivateData->setName("hPrivateData");
 313    Value *pPS = &*args++;
 314    pPS->setName("psCtx");
 315
 316    Value *consts_ptr = GEP(hPrivateData, {0, swr_draw_context_constantFS});
 317    consts_ptr->setName("fs_constants");
 318    Value *const_sizes_ptr =
 319       GEP(hPrivateData, {0, swr_draw_context_num_constantsFS});
 320    const_sizes_ptr->setName("num_fs_constants");
 321
 322    // xxx should check for flat shading versus interpolation
 323
 324
 325    // load *pAttribs, *pPerspAttribs
 326    Value *pRawAttribs = LOAD(pPS, {0, SWR_PS_CONTEXT_pAttribs}, "pRawAttribs");
 327    Value *pPerspAttribs =
 328       LOAD(pPS, {0, SWR_PS_CONTEXT_pPerspAttribs}, "pPerspAttribs");
 329
 330    swr_fs->constantMask = 0;
 331    swr_fs->pointSpriteMask = 0;
 332
 333    for (int attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
 334       const unsigned mask = swr_fs->info.base.input_usage_mask[attrib];
 335       const unsigned interpMode = swr_fs->info.base.input_interpolate[attrib];
 336       const unsigned interpLoc = swr_fs->info.base.input_interpolate_loc[attrib];
 337
 338       if (!mask)
 339          continue;
 340
 341       // load i,j
 342       Value *vi = nullptr, *vj = nullptr;
 343       switch (interpLoc) {
 344       case TGSI_INTERPOLATE_LOC_CENTER:
 345          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_center}, "i");
 346          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_center}, "j");
 347          break;
 348       case TGSI_INTERPOLATE_LOC_CENTROID:
 349          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_centroid}, "i");
 350          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_centroid}, "j");
 351          break;
 352       case TGSI_INTERPOLATE_LOC_SAMPLE:
 353          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_sample}, "i");
 354          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_sample}, "j");
 355          break;
 356       }
 357
 358       // load/compute w
 359       Value *vw = nullptr, *pAttribs;
 360       if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE) {
 361          pAttribs = pPerspAttribs;
 362          switch (interpLoc) {
 363          case TGSI_INTERPOLATE_LOC_CENTER:
 364             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}));
 365             break;
 366          case TGSI_INTERPOLATE_LOC_CENTROID:
 367             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_centroid}));
 368             break;
 369          case TGSI_INTERPOLATE_LOC_SAMPLE:
 370             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_sample}));
 371             break;
 372          }
 373       } else {
 374          pAttribs = pRawAttribs;
 375          vw = VIMMED1(1.f);
 376       }
 377
 378       vw->setName("w");
 379
 380       ubyte semantic_name = swr_fs->info.base.input_semantic_name[attrib];
 381       ubyte semantic_idx = swr_fs->info.base.input_semantic_index[attrib];
 382
 383       if (semantic_name == TGSI_SEMANTIC_FACE) {
 384          Value *ff =
 385             UI_TO_FP(LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), mFP32Ty);
 386          ff = FSUB(FMUL(ff, C(2.0f)), C(1.0f));
 387          ff = VECTOR_SPLAT(JM()->mVWidth, ff, "vFrontFace");
 388
 389          inputs[attrib][0] = wrap(ff);
 390          inputs[attrib][1] = wrap(VIMMED1(0.0f));
 391          inputs[attrib][2] = wrap(VIMMED1(0.0f));
 392          inputs[attrib][3] = wrap(VIMMED1(1.0f));
 393          continue;
 394       } else if (semantic_name == TGSI_SEMANTIC_POSITION) { // gl_FragCoord
 395          inputs[attrib][0] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_center}, "vX"));
 396          inputs[attrib][1] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_center}, "vY"));
 397          inputs[attrib][2] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vZ}, "vZ"));
 398          inputs[attrib][3] =
 399             wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}, "vOneOverW"));
 400          continue;
 401       } else if (semantic_name == TGSI_SEMANTIC_PRIMID) {
 402          Value *primID = LOAD(pPS, {0, SWR_PS_CONTEXT_primID}, "primID");
 403          inputs[attrib][0] = wrap(VECTOR_SPLAT(JM()->mVWidth, primID));
 404          inputs[attrib][1] = wrap(VIMMED1(0));
 405          inputs[attrib][2] = wrap(VIMMED1(0));
 406          inputs[attrib][3] = wrap(VIMMED1(0));
 407          continue;
 408       }
 409
 410       unsigned linkedAttrib =
 411          locate_linkage(semantic_name, semantic_idx, &ctx->vs->info.base);
 412       if (linkedAttrib == 0xFFFFFFFF) {
 413          // not found - check for point sprite
 414          if (ctx->rasterizer->sprite_coord_enable) {
 415             linkedAttrib = ctx->vs->info.base.num_outputs - 1;
 416             swr_fs->pointSpriteMask |= (1 << linkedAttrib);
 417          } else {
 418             fprintf(stderr,
 419                     "Missing %s[%d]\n",
 420                     tgsi_semantic_names[semantic_name],
 421                     semantic_idx);
 422             assert(0 && "attribute linkage not found");
 423          }
 424       }
 425
 426       if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
 427          swr_fs->constantMask |= 1 << linkedAttrib;
 428       }
 429
 430       for (int channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 431          if (mask & (1 << channel)) {
 432             Value *indexA = C(linkedAttrib * 12 + channel);
 433             Value *indexB = C(linkedAttrib * 12 + channel + 4);
 434             Value *indexC = C(linkedAttrib * 12 + channel + 8);
 435
 436             if ((semantic_name == TGSI_SEMANTIC_COLOR)
 437                 && ctx->rasterizer->light_twoside) {
 438                unsigned bcolorAttrib = locate_linkage(
 439                   TGSI_SEMANTIC_BCOLOR, semantic_idx, &ctx->vs->info.base);
 440
 441                unsigned diff = 12 * (bcolorAttrib - linkedAttrib);
 442
 443                Value *back =
 444                   XOR(C(1), LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), "backFace");
 445
 446                Value *offset = MUL(back, C(diff));
 447                offset->setName("offset");
 448
 449                indexA = ADD(indexA, offset);
 450                indexB = ADD(indexB, offset);
 451                indexC = ADD(indexC, offset);
 452
 453                if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
 454                   swr_fs->constantMask |= 1 << bcolorAttrib;
 455                }
 456             }
 457
 458             Value *va = VBROADCAST(LOAD(GEP(pAttribs, indexA)));
 459             Value *vb = VBROADCAST(LOAD(GEP(pAttribs, indexB)));
 460             Value *vc = VBROADCAST(LOAD(GEP(pAttribs, indexC)));
 461
 462             if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
 463                inputs[attrib][channel] = wrap(va);
 464             } else {
 465                Value *vk = FSUB(FSUB(VIMMED1(1.0f), vi), vj);
 466
 467                vc = FMUL(vk, vc);
 468
 469                Value *interp = FMUL(va, vi);
 470                Value *interp1 = FMUL(vb, vj);
 471                interp = FADD(interp, interp1);
 472                interp = FADD(interp, vc);
 473                if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE)
 474                   interp = FMUL(interp, vw);
 475                inputs[attrib][channel] = wrap(interp);
 476             }
 477          }
 478       }
 479    }
 480
 481    sampler = swr_sampler_soa_create(key.sampler);
 482
 483    struct lp_bld_tgsi_system_values system_values;
 484    memset(&system_values, 0, sizeof(system_values));
 485
 486    struct lp_build_mask_context mask;
 487
 488    if (swr_fs->info.base.uses_kill) {
 489       Value *mask_val = LOAD(pPS, {0, SWR_PS_CONTEXT_activeMask}, "activeMask");
 490       lp_build_mask_begin(
 491          &mask, gallivm, lp_type_float_vec(32, 32 * 8), wrap(mask_val));
 492    }
 493
 494    lp_build_tgsi_soa(gallivm,
 495                      swr_fs->pipe.tokens,
 496                      lp_type_float_vec(32, 32 * 8),
 497                      swr_fs->info.base.uses_kill ? &mask : NULL, // mask
 498                      wrap(consts_ptr),
 499                      wrap(const_sizes_ptr),
 500                      &system_values,
 501                      inputs,
 502                      outputs,
 503                      wrap(hPrivateData),
 504                      NULL, // thread data
 505                      sampler, // sampler
 506                      &swr_fs->info.base,
 507                      NULL); // geometry shader face
 508
 509    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 510
 511    for (uint32_t attrib = 0; attrib < swr_fs->info.base.num_outputs;
 512         attrib++) {
 513       switch (swr_fs->info.base.output_semantic_name[attrib]) {
 514       case TGSI_SEMANTIC_POSITION: {
 515          // write z
 516          LLVMValueRef outZ =
 517             LLVMBuildLoad(gallivm->builder, outputs[attrib][2], "");
 518          STORE(unwrap(outZ), pPS, {0, SWR_PS_CONTEXT_vZ});
 519          break;
 520       }
 521       case TGSI_SEMANTIC_COLOR: {
 522          for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 523             if (!outputs[attrib][channel])
 524                continue;
 525
 526             LLVMValueRef out =
 527                LLVMBuildLoad(gallivm->builder, outputs[attrib][channel], "");
 528             if (swr_fs->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS]) {
 529                for (uint32_t rt = 0; rt < key.nr_cbufs; rt++) {
 530                   STORE(unwrap(out),
 531                         pPS,
 532                         {0, SWR_PS_CONTEXT_shaded, rt, channel});
 533                }
 534             } else {
 535                STORE(unwrap(out),
 536                      pPS,
 537                      {0,
 538                            SWR_PS_CONTEXT_shaded,
 539                            swr_fs->info.base.output_semantic_index[attrib],
 540                            channel});
 541             }
 542          }
 543          break;
 544       }
 545       default: {
 546          fprintf(stderr,
 547                  "unknown output from FS %s[%d]\n",
 548                  tgsi_semantic_names[swr_fs->info.base
 549                                         .output_semantic_name[attrib]],
 550                  swr_fs->info.base.output_semantic_index[attrib]);
 551          break;
 552       }
 553       }
 554    }
 555
 556    LLVMValueRef mask_result = 0;
 557    if (swr_fs->info.base.uses_kill) {
 558       mask_result = lp_build_mask_end(&mask);
 559    }
 560
 561    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 562
 563    if (swr_fs->info.base.uses_kill) {
 564       STORE(unwrap(mask_result), pPS, {0, SWR_PS_CONTEXT_activeMask});
 565    }
 566
 567    RET_VOID();
 568
 569    gallivm_verify_function(gallivm, wrap(pFunction));
 570
 571    gallivm_compile_module(gallivm);
 572
 573    PFN_PIXEL_KERNEL kernel =
 574       (PFN_PIXEL_KERNEL)gallivm_jit_function(gallivm, wrap(pFunction));
 575    debug_printf("frag shader  %p\n", kernel);
 576    assert(kernel && "Error: FragShader = NULL");
 577
 578 #if (LLVM_VERSION_MAJOR == 3) && (LLVM_VERSION_MINOR >= 5)
 579    JM()->mIsModuleFinalized = true;
 580 #endif
 581
 582    return kernel;
 583 }
 584
 585 PFN_PIXEL_KERNEL
 586 swr_compile_fs(struct swr_context *ctx, swr_jit_key &key)
 587 {
 588    BuilderSWR builder(
 589       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr));
 590    return builder.CompileFS(ctx, key);
 591 }