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