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_fs_key &lhs, const swr_jit_fs_key &rhs)
  44 {
  45    return !memcmp(&lhs, &rhs, sizeof(lhs));
  46 }
  47
  48 bool operator==(const swr_jit_vs_key &lhs, const swr_jit_vs_key &rhs)
  49 {
  50    return !memcmp(&lhs, &rhs, sizeof(lhs));
  51 }
  52
  53 static void
  54 swr_generate_sampler_key(const struct lp_tgsi_info &info,
  55                          struct swr_context *ctx,
  56                          unsigned shader_type,
  57                          struct swr_jit_sampler_key &key)
  58 {
  59    key.nr_samplers = info.base.file_max[TGSI_FILE_SAMPLER] + 1;
  60
  61    for (unsigned i = 0; i < key.nr_samplers; i++) {
  62       if (info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
  63          lp_sampler_static_sampler_state(
  64             &key.sampler[i].sampler_state,
  65             ctx->samplers[shader_type][i]);
  66       }
  67    }
  68
  69    /*
  70     * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
  71     * are dx10-style? Can't really have mixed opcodes, at least not
  72     * if we want to skip the holes here (without rescanning tgsi).
  73     */
  74    if (info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
  75       key.nr_sampler_views =
  76          info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
  77       for (unsigned i = 0; i < key.nr_sampler_views; i++) {
  78          if (info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1 << i)) {
  79             lp_sampler_static_texture_state(
  80                &key.sampler[i].texture_state,
  81                ctx->sampler_views[shader_type][i]);
  82          }
  83       }
  84    } else {
  85       key.nr_sampler_views = key.nr_samplers;
  86       for (unsigned i = 0; i < key.nr_sampler_views; i++) {
  87          if (info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
  88             lp_sampler_static_texture_state(
  89                &key.sampler[i].texture_state,
  90                ctx->sampler_views[shader_type][i]);
  91          }
  92       }
  93    }
  94 }
  95
  96 void
  97 swr_generate_fs_key(struct swr_jit_fs_key &key,
  98                     struct swr_context *ctx,
  99                     swr_fragment_shader *swr_fs)
 100 {
 101    memset(&key, 0, sizeof(key));
 102
 103    key.nr_cbufs = ctx->framebuffer.nr_cbufs;
 104    key.light_twoside = ctx->rasterizer->light_twoside;
 105    key.flatshade = ctx->rasterizer->flatshade;
 106    memcpy(&key.vs_output_semantic_name,
 107           &ctx->vs->info.base.output_semantic_name,
 108           sizeof(key.vs_output_semantic_name));
 109    memcpy(&key.vs_output_semantic_idx,
 110           &ctx->vs->info.base.output_semantic_index,
 111           sizeof(key.vs_output_semantic_idx));
 112
 113    swr_generate_sampler_key(swr_fs->info, ctx, PIPE_SHADER_FRAGMENT, key);
 114 }
 115
 116 void
 117 swr_generate_vs_key(struct swr_jit_vs_key &key,
 118                     struct swr_context *ctx,
 119                     swr_vertex_shader *swr_vs)
 120 {
 121    memset(&key, 0, sizeof(key));
 122
 123    swr_generate_sampler_key(swr_vs->info, ctx, PIPE_SHADER_VERTEX, key);
 124 }
 125
 126 struct BuilderSWR : public Builder {
 127    BuilderSWR(JitManager *pJitMgr)
 128       : Builder(pJitMgr)
 129    {
 130       pJitMgr->SetupNewModule();
 131    }
 132
 133    PFN_VERTEX_FUNC CompileVS(struct swr_context *ctx, swr_jit_vs_key &key);
 134    PFN_PIXEL_KERNEL CompileFS(struct swr_context *ctx, swr_jit_fs_key &key);
 135 };
 136
 137 PFN_VERTEX_FUNC
 138 BuilderSWR::CompileVS(struct swr_context *ctx, swr_jit_vs_key &key)
 139 {
 140    struct swr_vertex_shader *swr_vs = ctx->vs;
 141
 142    swr_vs->linkageMask = 0;
 143
 144    for (unsigned i = 0; i < swr_vs->info.base.num_outputs; i++) {
 145       switch (swr_vs->info.base.output_semantic_name[i]) {
 146       case TGSI_SEMANTIC_POSITION:
 147          break;
 148       default:
 149          swr_vs->linkageMask |= (1 << i);
 150          break;
 151       }
 152    }
 153
 154    //   tgsi_dump(swr_vs->pipe.tokens, 0);
 155
 156    struct gallivm_state *gallivm =
 157       gallivm_create("VS", wrap(&JM()->mContext));
 158    gallivm->module = wrap(JM()->mpCurrentModule);
 159
 160    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 161    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 162
 163    memset(outputs, 0, sizeof(outputs));
 164
 165    AttrBuilder attrBuilder;
 166    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 167    AttributeSet attrSet = AttributeSet::get(
 168       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 169
 170    std::vector<Type *> vsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 171                               PointerType::get(Gen_SWR_VS_CONTEXT(JM()), 0)};
 172    FunctionType *vsFuncType =
 173       FunctionType::get(Type::getVoidTy(JM()->mContext), vsArgs, false);
 174
 175    // create new vertex shader function
 176    auto pFunction = Function::Create(vsFuncType,
 177                                      GlobalValue::ExternalLinkage,
 178                                      "VS",
 179                                      JM()->mpCurrentModule);
 180    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 181
 182    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 183    IRB()->SetInsertPoint(block);
 184    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 185
 186    auto argitr = pFunction->arg_begin();
 187    Value *hPrivateData = &*argitr++;
 188    hPrivateData->setName("hPrivateData");
 189    Value *pVsCtx = &*argitr++;
 190    pVsCtx->setName("vsCtx");
 191
 192    Value *consts_ptr = GEP(hPrivateData, {C(0), C(swr_draw_context_constantVS)});
 193
 194    consts_ptr->setName("vs_constants");
 195    Value *const_sizes_ptr =
 196       GEP(hPrivateData, {0, swr_draw_context_num_constantsVS});
 197    const_sizes_ptr->setName("num_vs_constants");
 198
 199    Value *vtxInput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVin});
 200
 201    for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
 202       const unsigned mask = swr_vs->info.base.input_usage_mask[attrib];
 203       for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 204          if (mask & (1 << channel)) {
 205             inputs[attrib][channel] =
 206                wrap(LOAD(vtxInput, {0, 0, attrib, channel}));
 207          }
 208       }
 209    }
 210
 211    struct lp_build_sampler_soa *sampler =
 212       swr_sampler_soa_create(key.sampler, PIPE_SHADER_VERTEX);
 213
 214    struct lp_bld_tgsi_system_values system_values;
 215    memset(&system_values, 0, sizeof(system_values));
 216    system_values.instance_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_InstanceID}));
 217    system_values.vertex_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_VertexID}));
 218
 219    lp_build_tgsi_soa(gallivm,
 220                      swr_vs->pipe.tokens,
 221                      lp_type_float_vec(32, 32 * 8),
 222                      NULL, // mask
 223                      wrap(consts_ptr),
 224                      wrap(const_sizes_ptr),
 225                      &system_values,
 226                      inputs,
 227                      outputs,
 228                      wrap(hPrivateData), // (sampler context)
 229                      NULL, // thread data
 230                      sampler, // sampler
 231                      &swr_vs->info.base,
 232                      NULL); // geometry shader face
 233
 234    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 235
 236    Value *vtxOutput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVout});
 237
 238    for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 239       for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_OUTPUTS; attrib++) {
 240          if (!outputs[attrib][channel])
 241             continue;
 242
 243          Value *val = LOAD(unwrap(outputs[attrib][channel]));
 244
 245          uint32_t outSlot = attrib;
 246          if (swr_vs->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_PSIZE)
 247             outSlot = VERTEX_POINT_SIZE_SLOT;
 248          STORE(val, vtxOutput, {0, 0, outSlot, channel});
 249       }
 250    }
 251
 252    RET_VOID();
 253
 254    gallivm_verify_function(gallivm, wrap(pFunction));
 255    gallivm_compile_module(gallivm);
 256
 257    //   lp_debug_dump_value(func);
 258
 259    PFN_VERTEX_FUNC pFunc =
 260       (PFN_VERTEX_FUNC)gallivm_jit_function(gallivm, wrap(pFunction));
 261
 262    debug_printf("vert shader  %p\n", pFunc);
 263    assert(pFunc && "Error: VertShader = NULL");
 264
 265 #if (LLVM_VERSION_MAJOR == 3) && (LLVM_VERSION_MINOR >= 5)
 266    JM()->mIsModuleFinalized = true;
 267 #endif
 268
 269    return pFunc;
 270 }
 271
 272 PFN_VERTEX_FUNC
 273 swr_compile_vs(struct swr_context *ctx, swr_jit_vs_key &key)
 274 {
 275    BuilderSWR builder(
 276       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr));
 277    return builder.CompileVS(ctx, key);
 278 }
 279
 280 static unsigned
 281 locate_linkage(ubyte name, ubyte index, struct tgsi_shader_info *info)
 282 {
 283    for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 284       if ((info->output_semantic_name[i] == name)
 285           && (info->output_semantic_index[i] == index)) {
 286          return i - 1; // position is not part of the linkage
 287       }
 288    }
 289
 290    if (name == TGSI_SEMANTIC_COLOR) { // BCOLOR fallback
 291       for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 292          if ((info->output_semantic_name[i] == TGSI_SEMANTIC_BCOLOR)
 293              && (info->output_semantic_index[i] == index)) {
 294             return i - 1; // position is not part of the linkage
 295          }
 296       }
 297    }
 298
 299    return 0xFFFFFFFF;
 300 }
 301
 302 PFN_PIXEL_KERNEL
 303 BuilderSWR::CompileFS(struct swr_context *ctx, swr_jit_fs_key &key)
 304 {
 305    struct swr_fragment_shader *swr_fs = ctx->fs;
 306
 307    //   tgsi_dump(swr_fs->pipe.tokens, 0);
 308
 309    struct gallivm_state *gallivm =
 310       gallivm_create("FS", wrap(&JM()->mContext));
 311    gallivm->module = wrap(JM()->mpCurrentModule);
 312
 313    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 314    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 315
 316    memset(inputs, 0, sizeof(inputs));
 317    memset(outputs, 0, sizeof(outputs));
 318
 319    struct lp_build_sampler_soa *sampler = NULL;
 320
 321    AttrBuilder attrBuilder;
 322    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 323    AttributeSet attrSet = AttributeSet::get(
 324       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 325
 326    std::vector<Type *> fsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 327                               PointerType::get(Gen_SWR_PS_CONTEXT(JM()), 0)};
 328    FunctionType *funcType =
 329       FunctionType::get(Type::getVoidTy(JM()->mContext), fsArgs, false);
 330
 331    auto pFunction = Function::Create(funcType,
 332                                      GlobalValue::ExternalLinkage,
 333                                      "FS",
 334                                      JM()->mpCurrentModule);
 335    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 336
 337    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 338    IRB()->SetInsertPoint(block);
 339    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 340
 341    auto args = pFunction->arg_begin();
 342    Value *hPrivateData = &*args++;
 343    hPrivateData->setName("hPrivateData");
 344    Value *pPS = &*args++;
 345    pPS->setName("psCtx");
 346
 347    Value *consts_ptr = GEP(hPrivateData, {0, swr_draw_context_constantFS});
 348    consts_ptr->setName("fs_constants");
 349    Value *const_sizes_ptr =
 350       GEP(hPrivateData, {0, swr_draw_context_num_constantsFS});
 351    const_sizes_ptr->setName("num_fs_constants");
 352
 353    // xxx should check for flat shading versus interpolation
 354
 355
 356    // load *pAttribs, *pPerspAttribs
 357    Value *pRawAttribs = LOAD(pPS, {0, SWR_PS_CONTEXT_pAttribs}, "pRawAttribs");
 358    Value *pPerspAttribs =
 359       LOAD(pPS, {0, SWR_PS_CONTEXT_pPerspAttribs}, "pPerspAttribs");
 360
 361    swr_fs->constantMask = 0;
 362    swr_fs->pointSpriteMask = 0;
 363
 364    for (int attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
 365       const unsigned mask = swr_fs->info.base.input_usage_mask[attrib];
 366       const unsigned interpMode = swr_fs->info.base.input_interpolate[attrib];
 367       const unsigned interpLoc = swr_fs->info.base.input_interpolate_loc[attrib];
 368
 369       if (!mask)
 370          continue;
 371
 372       // load i,j
 373       Value *vi = nullptr, *vj = nullptr;
 374       switch (interpLoc) {
 375       case TGSI_INTERPOLATE_LOC_CENTER:
 376          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_center}, "i");
 377          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_center}, "j");
 378          break;
 379       case TGSI_INTERPOLATE_LOC_CENTROID:
 380          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_centroid}, "i");
 381          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_centroid}, "j");
 382          break;
 383       case TGSI_INTERPOLATE_LOC_SAMPLE:
 384          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_sample}, "i");
 385          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_sample}, "j");
 386          break;
 387       }
 388
 389       // load/compute w
 390       Value *vw = nullptr, *pAttribs;
 391       if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE) {
 392          pAttribs = pPerspAttribs;
 393          switch (interpLoc) {
 394          case TGSI_INTERPOLATE_LOC_CENTER:
 395             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}));
 396             break;
 397          case TGSI_INTERPOLATE_LOC_CENTROID:
 398             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_centroid}));
 399             break;
 400          case TGSI_INTERPOLATE_LOC_SAMPLE:
 401             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_sample}));
 402             break;
 403          }
 404       } else {
 405          pAttribs = pRawAttribs;
 406          vw = VIMMED1(1.f);
 407       }
 408
 409       vw->setName("w");
 410
 411       ubyte semantic_name = swr_fs->info.base.input_semantic_name[attrib];
 412       ubyte semantic_idx = swr_fs->info.base.input_semantic_index[attrib];
 413
 414       if (semantic_name == TGSI_SEMANTIC_FACE) {
 415          Value *ff =
 416             UI_TO_FP(LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), mFP32Ty);
 417          ff = FSUB(FMUL(ff, C(2.0f)), C(1.0f));
 418          ff = VECTOR_SPLAT(JM()->mVWidth, ff, "vFrontFace");
 419
 420          inputs[attrib][0] = wrap(ff);
 421          inputs[attrib][1] = wrap(VIMMED1(0.0f));
 422          inputs[attrib][2] = wrap(VIMMED1(0.0f));
 423          inputs[attrib][3] = wrap(VIMMED1(1.0f));
 424          continue;
 425       } else if (semantic_name == TGSI_SEMANTIC_POSITION) { // gl_FragCoord
 426          inputs[attrib][0] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_center}, "vX"));
 427          inputs[attrib][1] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_center}, "vY"));
 428          inputs[attrib][2] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vZ}, "vZ"));
 429          inputs[attrib][3] =
 430             wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}, "vOneOverW"));
 431          continue;
 432       } else if (semantic_name == TGSI_SEMANTIC_PRIMID) {
 433          Value *primID = LOAD(pPS, {0, SWR_PS_CONTEXT_primID}, "primID");
 434          inputs[attrib][0] = wrap(VECTOR_SPLAT(JM()->mVWidth, primID));
 435          inputs[attrib][1] = wrap(VIMMED1(0));
 436          inputs[attrib][2] = wrap(VIMMED1(0));
 437          inputs[attrib][3] = wrap(VIMMED1(0));
 438          continue;
 439       }
 440
 441       unsigned linkedAttrib =
 442          locate_linkage(semantic_name, semantic_idx, &ctx->vs->info.base);
 443       if (linkedAttrib == 0xFFFFFFFF) {
 444          // not found - check for point sprite
 445          if (ctx->rasterizer->sprite_coord_enable) {
 446             linkedAttrib = ctx->vs->info.base.num_outputs - 1;
 447             swr_fs->pointSpriteMask |= (1 << linkedAttrib);
 448          } else {
 449             fprintf(stderr,
 450                     "Missing %s[%d]\n",
 451                     tgsi_semantic_names[semantic_name],
 452                     semantic_idx);
 453             assert(0 && "attribute linkage not found");
 454          }
 455       }
 456
 457       if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
 458          swr_fs->constantMask |= 1 << linkedAttrib;
 459       }
 460
 461       for (int channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 462          if (mask & (1 << channel)) {
 463             Value *indexA = C(linkedAttrib * 12 + channel);
 464             Value *indexB = C(linkedAttrib * 12 + channel + 4);
 465             Value *indexC = C(linkedAttrib * 12 + channel + 8);
 466
 467             if ((semantic_name == TGSI_SEMANTIC_COLOR)
 468                 && ctx->rasterizer->light_twoside) {
 469                unsigned bcolorAttrib = locate_linkage(
 470                   TGSI_SEMANTIC_BCOLOR, semantic_idx, &ctx->vs->info.base);
 471
 472                unsigned diff = 12 * (bcolorAttrib - linkedAttrib);
 473
 474                Value *back =
 475                   XOR(C(1), LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), "backFace");
 476
 477                Value *offset = MUL(back, C(diff));
 478                offset->setName("offset");
 479
 480                indexA = ADD(indexA, offset);
 481                indexB = ADD(indexB, offset);
 482                indexC = ADD(indexC, offset);
 483
 484                if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
 485                   swr_fs->constantMask |= 1 << bcolorAttrib;
 486                }
 487             }
 488
 489             Value *va = VBROADCAST(LOAD(GEP(pAttribs, indexA)));
 490             Value *vb = VBROADCAST(LOAD(GEP(pAttribs, indexB)));
 491             Value *vc = VBROADCAST(LOAD(GEP(pAttribs, indexC)));
 492
 493             if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
 494                inputs[attrib][channel] = wrap(va);
 495             } else if ((interpMode == TGSI_INTERPOLATE_COLOR) &&
 496                        (key.flatshade == true)) {
 497                inputs[attrib][channel] = wrap(vc);
 498             } else {
 499                Value *vk = FSUB(FSUB(VIMMED1(1.0f), vi), vj);
 500
 501                vc = FMUL(vk, vc);
 502
 503                Value *interp = FMUL(va, vi);
 504                Value *interp1 = FMUL(vb, vj);
 505                interp = FADD(interp, interp1);
 506                interp = FADD(interp, vc);
 507                if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE)
 508                   interp = FMUL(interp, vw);
 509                inputs[attrib][channel] = wrap(interp);
 510             }
 511          }
 512       }
 513    }
 514
 515    sampler = swr_sampler_soa_create(key.sampler, PIPE_SHADER_FRAGMENT);
 516
 517    struct lp_bld_tgsi_system_values system_values;
 518    memset(&system_values, 0, sizeof(system_values));
 519
 520    struct lp_build_mask_context mask;
 521
 522    if (swr_fs->info.base.uses_kill) {
 523       Value *mask_val = LOAD(pPS, {0, SWR_PS_CONTEXT_activeMask}, "activeMask");
 524       lp_build_mask_begin(
 525          &mask, gallivm, lp_type_float_vec(32, 32 * 8), wrap(mask_val));
 526    }
 527
 528    lp_build_tgsi_soa(gallivm,
 529                      swr_fs->pipe.tokens,
 530                      lp_type_float_vec(32, 32 * 8),
 531                      swr_fs->info.base.uses_kill ? &mask : NULL, // mask
 532                      wrap(consts_ptr),
 533                      wrap(const_sizes_ptr),
 534                      &system_values,
 535                      inputs,
 536                      outputs,
 537                      wrap(hPrivateData),
 538                      NULL, // thread data
 539                      sampler, // sampler
 540                      &swr_fs->info.base,
 541                      NULL); // geometry shader face
 542
 543    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 544
 545    for (uint32_t attrib = 0; attrib < swr_fs->info.base.num_outputs;
 546         attrib++) {
 547       switch (swr_fs->info.base.output_semantic_name[attrib]) {
 548       case TGSI_SEMANTIC_POSITION: {
 549          // write z
 550          LLVMValueRef outZ =
 551             LLVMBuildLoad(gallivm->builder, outputs[attrib][2], "");
 552          STORE(unwrap(outZ), pPS, {0, SWR_PS_CONTEXT_vZ});
 553          break;
 554       }
 555       case TGSI_SEMANTIC_COLOR: {
 556          for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 557             if (!outputs[attrib][channel])
 558                continue;
 559
 560             LLVMValueRef out =
 561                LLVMBuildLoad(gallivm->builder, outputs[attrib][channel], "");
 562             if (swr_fs->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS]) {
 563                for (uint32_t rt = 0; rt < key.nr_cbufs; rt++) {
 564                   STORE(unwrap(out),
 565                         pPS,
 566                         {0, SWR_PS_CONTEXT_shaded, rt, channel});
 567                }
 568             } else {
 569                STORE(unwrap(out),
 570                      pPS,
 571                      {0,
 572                            SWR_PS_CONTEXT_shaded,
 573                            swr_fs->info.base.output_semantic_index[attrib],
 574                            channel});
 575             }
 576          }
 577          break;
 578       }
 579       default: {
 580          fprintf(stderr,
 581                  "unknown output from FS %s[%d]\n",
 582                  tgsi_semantic_names[swr_fs->info.base
 583                                         .output_semantic_name[attrib]],
 584                  swr_fs->info.base.output_semantic_index[attrib]);
 585          break;
 586       }
 587       }
 588    }
 589
 590    LLVMValueRef mask_result = 0;
 591    if (swr_fs->info.base.uses_kill) {
 592       mask_result = lp_build_mask_end(&mask);
 593    }
 594
 595    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 596
 597    if (swr_fs->info.base.uses_kill) {
 598       STORE(unwrap(mask_result), pPS, {0, SWR_PS_CONTEXT_activeMask});
 599    }
 600
 601    RET_VOID();
 602
 603    gallivm_verify_function(gallivm, wrap(pFunction));
 604
 605    gallivm_compile_module(gallivm);
 606
 607    PFN_PIXEL_KERNEL kernel =
 608       (PFN_PIXEL_KERNEL)gallivm_jit_function(gallivm, wrap(pFunction));
 609    debug_printf("frag shader  %p\n", kernel);
 610    assert(kernel && "Error: FragShader = NULL");
 611
 612 #if (LLVM_VERSION_MAJOR == 3) && (LLVM_VERSION_MINOR >= 5)
 613    JM()->mIsModuleFinalized = true;
 614 #endif
 615
 616    return kernel;
 617 }
 618
 619 PFN_PIXEL_KERNEL
 620 swr_compile_fs(struct swr_context *ctx, swr_jit_fs_key &key)
 621 {
 622    BuilderSWR builder(
 623       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr));
 624    return builder.CompileFS(ctx, key);
 625 }