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, const char *pName)
 128       : Builder(pJitMgr)
 129    {
 130       pJitMgr->SetupNewModule();
 131       gallivm = gallivm_create(pName, wrap(&JM()->mContext));
 132       pJitMgr->mpCurrentModule = unwrap(gallivm->module);
 133    }
 134
 135    ~BuilderSWR() {
 136       gallivm_free_ir(gallivm);
 137    }
 138
 139    struct gallivm_state *gallivm;
 140    PFN_VERTEX_FUNC CompileVS(struct swr_context *ctx, swr_jit_vs_key &key);
 141    PFN_PIXEL_KERNEL CompileFS(struct swr_context *ctx, swr_jit_fs_key &key);
 142 };
 143
 144 PFN_VERTEX_FUNC
 145 BuilderSWR::CompileVS(struct swr_context *ctx, swr_jit_vs_key &key)
 146 {
 147    struct swr_vertex_shader *swr_vs = ctx->vs;
 148
 149    swr_vs->linkageMask = 0;
 150
 151    for (unsigned i = 0; i < swr_vs->info.base.num_outputs; i++) {
 152       switch (swr_vs->info.base.output_semantic_name[i]) {
 153       case TGSI_SEMANTIC_POSITION:
 154          break;
 155       default:
 156          swr_vs->linkageMask |= (1 << i);
 157          break;
 158       }
 159    }
 160
 161    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 162    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 163
 164    memset(outputs, 0, sizeof(outputs));
 165
 166    AttrBuilder attrBuilder;
 167    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 168    AttributeSet attrSet = AttributeSet::get(
 169       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 170
 171    std::vector<Type *> vsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 172                               PointerType::get(Gen_SWR_VS_CONTEXT(JM()), 0)};
 173    FunctionType *vsFuncType =
 174       FunctionType::get(Type::getVoidTy(JM()->mContext), vsArgs, false);
 175
 176    // create new vertex shader function
 177    auto pFunction = Function::Create(vsFuncType,
 178                                      GlobalValue::ExternalLinkage,
 179                                      "VS",
 180                                      JM()->mpCurrentModule);
 181    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 182
 183    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 184    IRB()->SetInsertPoint(block);
 185    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 186
 187    auto argitr = pFunction->arg_begin();
 188    Value *hPrivateData = &*argitr++;
 189    hPrivateData->setName("hPrivateData");
 190    Value *pVsCtx = &*argitr++;
 191    pVsCtx->setName("vsCtx");
 192
 193    Value *consts_ptr = GEP(hPrivateData, {C(0), C(swr_draw_context_constantVS)});
 194
 195    consts_ptr->setName("vs_constants");
 196    Value *const_sizes_ptr =
 197       GEP(hPrivateData, {0, swr_draw_context_num_constantsVS});
 198    const_sizes_ptr->setName("num_vs_constants");
 199
 200    Value *vtxInput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVin});
 201
 202    for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
 203       const unsigned mask = swr_vs->info.base.input_usage_mask[attrib];
 204       for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 205          if (mask & (1 << channel)) {
 206             inputs[attrib][channel] =
 207                wrap(LOAD(vtxInput, {0, 0, attrib, channel}));
 208          }
 209       }
 210    }
 211
 212    struct lp_build_sampler_soa *sampler =
 213       swr_sampler_soa_create(key.sampler, PIPE_SHADER_VERTEX);
 214
 215    struct lp_bld_tgsi_system_values system_values;
 216    memset(&system_values, 0, sizeof(system_values));
 217    system_values.instance_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_InstanceID}));
 218    system_values.vertex_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_VertexID}));
 219
 220    lp_build_tgsi_soa(gallivm,
 221                      swr_vs->pipe.tokens,
 222                      lp_type_float_vec(32, 32 * 8),
 223                      NULL, // mask
 224                      wrap(consts_ptr),
 225                      wrap(const_sizes_ptr),
 226                      &system_values,
 227                      inputs,
 228                      outputs,
 229                      wrap(hPrivateData), // (sampler context)
 230                      NULL, // thread data
 231                      sampler, // sampler
 232                      &swr_vs->info.base,
 233                      NULL); // geometry shader face
 234
 235    sampler->destroy(sampler);
 236
 237    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 238
 239    Value *vtxOutput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVout});
 240
 241    for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 242       for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_OUTPUTS; attrib++) {
 243          if (!outputs[attrib][channel])
 244             continue;
 245
 246          Value *val = LOAD(unwrap(outputs[attrib][channel]));
 247
 248          uint32_t outSlot = attrib;
 249          if (swr_vs->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_PSIZE)
 250             outSlot = VERTEX_POINT_SIZE_SLOT;
 251          STORE(val, vtxOutput, {0, 0, outSlot, channel});
 252       }
 253    }
 254
 255    RET_VOID();
 256
 257    gallivm_verify_function(gallivm, wrap(pFunction));
 258    gallivm_compile_module(gallivm);
 259
 260    //   lp_debug_dump_value(func);
 261
 262    PFN_VERTEX_FUNC pFunc =
 263       (PFN_VERTEX_FUNC)gallivm_jit_function(gallivm, wrap(pFunction));
 264
 265    debug_printf("vert shader  %p\n", pFunc);
 266    assert(pFunc && "Error: VertShader = NULL");
 267
 268 #if (LLVM_VERSION_MAJOR == 3) && (LLVM_VERSION_MINOR >= 5)
 269    JM()->mIsModuleFinalized = true;
 270 #endif
 271
 272    return pFunc;
 273 }
 274
 275 PFN_VERTEX_FUNC
 276 swr_compile_vs(struct swr_context *ctx, swr_jit_vs_key &key)
 277 {
 278    BuilderSWR builder(
 279       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
 280       "VS");
 281    PFN_VERTEX_FUNC func = builder.CompileVS(ctx, key);
 282
 283    ctx->vs->map.insert(std::make_pair(key, make_unique<VariantVS>(builder.gallivm, func)));
 284    return func;
 285 }
 286
 287 static unsigned
 288 locate_linkage(ubyte name, ubyte index, struct tgsi_shader_info *info)
 289 {
 290    for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 291       if ((info->output_semantic_name[i] == name)
 292           && (info->output_semantic_index[i] == index)) {
 293          return i - 1; // position is not part of the linkage
 294       }
 295    }
 296
 297    if (name == TGSI_SEMANTIC_COLOR) { // BCOLOR fallback
 298       for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 299          if ((info->output_semantic_name[i] == TGSI_SEMANTIC_BCOLOR)
 300              && (info->output_semantic_index[i] == index)) {
 301             return i - 1; // position is not part of the linkage
 302          }
 303       }
 304    }
 305
 306    return 0xFFFFFFFF;
 307 }
 308
 309 PFN_PIXEL_KERNEL
 310 BuilderSWR::CompileFS(struct swr_context *ctx, swr_jit_fs_key &key)
 311 {
 312    struct swr_fragment_shader *swr_fs = ctx->fs;
 313
 314    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 315    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 316
 317    memset(inputs, 0, sizeof(inputs));
 318    memset(outputs, 0, sizeof(outputs));
 319
 320    struct lp_build_sampler_soa *sampler = NULL;
 321
 322    AttrBuilder attrBuilder;
 323    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 324    AttributeSet attrSet = AttributeSet::get(
 325       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 326
 327    std::vector<Type *> fsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 328                               PointerType::get(Gen_SWR_PS_CONTEXT(JM()), 0)};
 329    FunctionType *funcType =
 330       FunctionType::get(Type::getVoidTy(JM()->mContext), fsArgs, false);
 331
 332    auto pFunction = Function::Create(funcType,
 333                                      GlobalValue::ExternalLinkage,
 334                                      "FS",
 335                                      JM()->mpCurrentModule);
 336    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 337
 338    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 339    IRB()->SetInsertPoint(block);
 340    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 341
 342    auto args = pFunction->arg_begin();
 343    Value *hPrivateData = &*args++;
 344    hPrivateData->setName("hPrivateData");
 345    Value *pPS = &*args++;
 346    pPS->setName("psCtx");
 347
 348    Value *consts_ptr = GEP(hPrivateData, {0, swr_draw_context_constantFS});
 349    consts_ptr->setName("fs_constants");
 350    Value *const_sizes_ptr =
 351       GEP(hPrivateData, {0, swr_draw_context_num_constantsFS});
 352    const_sizes_ptr->setName("num_fs_constants");
 353
 354    // xxx should check for flat shading versus interpolation
 355
 356
 357    // load *pAttribs, *pPerspAttribs
 358    Value *pRawAttribs = LOAD(pPS, {0, SWR_PS_CONTEXT_pAttribs}, "pRawAttribs");
 359    Value *pPerspAttribs =
 360       LOAD(pPS, {0, SWR_PS_CONTEXT_pPerspAttribs}, "pPerspAttribs");
 361
 362    swr_fs->constantMask = 0;
 363    swr_fs->pointSpriteMask = 0;
 364
 365    for (int attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
 366       const unsigned mask = swr_fs->info.base.input_usage_mask[attrib];
 367       const unsigned interpMode = swr_fs->info.base.input_interpolate[attrib];
 368       const unsigned interpLoc = swr_fs->info.base.input_interpolate_loc[attrib];
 369
 370       if (!mask)
 371          continue;
 372
 373       // load i,j
 374       Value *vi = nullptr, *vj = nullptr;
 375       switch (interpLoc) {
 376       case TGSI_INTERPOLATE_LOC_CENTER:
 377          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_center}, "i");
 378          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_center}, "j");
 379          break;
 380       case TGSI_INTERPOLATE_LOC_CENTROID:
 381          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_centroid}, "i");
 382          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_centroid}, "j");
 383          break;
 384       case TGSI_INTERPOLATE_LOC_SAMPLE:
 385          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_sample}, "i");
 386          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_sample}, "j");
 387          break;
 388       }
 389
 390       // load/compute w
 391       Value *vw = nullptr, *pAttribs;
 392       if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE) {
 393          pAttribs = pPerspAttribs;
 394          switch (interpLoc) {
 395          case TGSI_INTERPOLATE_LOC_CENTER:
 396             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}));
 397             break;
 398          case TGSI_INTERPOLATE_LOC_CENTROID:
 399             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_centroid}));
 400             break;
 401          case TGSI_INTERPOLATE_LOC_SAMPLE:
 402             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_sample}));
 403             break;
 404          }
 405       } else {
 406          pAttribs = pRawAttribs;
 407          vw = VIMMED1(1.f);
 408       }
 409
 410       vw->setName("w");
 411
 412       ubyte semantic_name = swr_fs->info.base.input_semantic_name[attrib];
 413       ubyte semantic_idx = swr_fs->info.base.input_semantic_index[attrib];
 414
 415       if (semantic_name == TGSI_SEMANTIC_FACE) {
 416          Value *ff =
 417             UI_TO_FP(LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), mFP32Ty);
 418          ff = FSUB(FMUL(ff, C(2.0f)), C(1.0f));
 419          ff = VECTOR_SPLAT(JM()->mVWidth, ff, "vFrontFace");
 420
 421          inputs[attrib][0] = wrap(ff);
 422          inputs[attrib][1] = wrap(VIMMED1(0.0f));
 423          inputs[attrib][2] = wrap(VIMMED1(0.0f));
 424          inputs[attrib][3] = wrap(VIMMED1(1.0f));
 425          continue;
 426       } else if (semantic_name == TGSI_SEMANTIC_POSITION) { // gl_FragCoord
 427          inputs[attrib][0] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_center}, "vX"));
 428          inputs[attrib][1] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_center}, "vY"));
 429          inputs[attrib][2] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vZ}, "vZ"));
 430          inputs[attrib][3] =
 431             wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}, "vOneOverW"));
 432          continue;
 433       } else if (semantic_name == TGSI_SEMANTIC_PRIMID) {
 434          Value *primID = LOAD(pPS, {0, SWR_PS_CONTEXT_primID}, "primID");
 435          inputs[attrib][0] = wrap(VECTOR_SPLAT(JM()->mVWidth, primID));
 436          inputs[attrib][1] = wrap(VIMMED1(0));
 437          inputs[attrib][2] = wrap(VIMMED1(0));
 438          inputs[attrib][3] = wrap(VIMMED1(0));
 439          continue;
 440       }
 441
 442       unsigned linkedAttrib =
 443          locate_linkage(semantic_name, semantic_idx, &ctx->vs->info.base);
 444       if (linkedAttrib == 0xFFFFFFFF) {
 445          // not found - check for point sprite
 446          if (ctx->rasterizer->sprite_coord_enable) {
 447             linkedAttrib = ctx->vs->info.base.num_outputs - 1;
 448             swr_fs->pointSpriteMask |= (1 << linkedAttrib);
 449          } else {
 450             fprintf(stderr,
 451                     "Missing %s[%d]\n",
 452                     tgsi_semantic_names[semantic_name],
 453                     semantic_idx);
 454             assert(0 && "attribute linkage not found");
 455          }
 456       }
 457
 458       if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
 459          swr_fs->constantMask |= 1 << linkedAttrib;
 460       }
 461
 462       for (int channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 463          if (mask & (1 << channel)) {
 464             Value *indexA = C(linkedAttrib * 12 + channel);
 465             Value *indexB = C(linkedAttrib * 12 + channel + 4);
 466             Value *indexC = C(linkedAttrib * 12 + channel + 8);
 467
 468             if ((semantic_name == TGSI_SEMANTIC_COLOR)
 469                 && ctx->rasterizer->light_twoside) {
 470                unsigned bcolorAttrib = locate_linkage(
 471                   TGSI_SEMANTIC_BCOLOR, semantic_idx, &ctx->vs->info.base);
 472
 473                unsigned diff = 12 * (bcolorAttrib - linkedAttrib);
 474
 475                Value *back =
 476                   XOR(C(1), LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), "backFace");
 477
 478                Value *offset = MUL(back, C(diff));
 479                offset->setName("offset");
 480
 481                indexA = ADD(indexA, offset);
 482                indexB = ADD(indexB, offset);
 483                indexC = ADD(indexC, offset);
 484
 485                if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
 486                   swr_fs->constantMask |= 1 << bcolorAttrib;
 487                }
 488             }
 489
 490             Value *va = VBROADCAST(LOAD(GEP(pAttribs, indexA)));
 491             Value *vb = VBROADCAST(LOAD(GEP(pAttribs, indexB)));
 492             Value *vc = VBROADCAST(LOAD(GEP(pAttribs, indexC)));
 493
 494             if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
 495                inputs[attrib][channel] = wrap(va);
 496             } else if ((interpMode == TGSI_INTERPOLATE_COLOR) &&
 497                        (key.flatshade == true)) {
 498                inputs[attrib][channel] = wrap(vc);
 499             } else {
 500                Value *vk = FSUB(FSUB(VIMMED1(1.0f), vi), vj);
 501
 502                vc = FMUL(vk, vc);
 503
 504                Value *interp = FMUL(va, vi);
 505                Value *interp1 = FMUL(vb, vj);
 506                interp = FADD(interp, interp1);
 507                interp = FADD(interp, vc);
 508                if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE)
 509                   interp = FMUL(interp, vw);
 510                inputs[attrib][channel] = wrap(interp);
 511             }
 512          }
 513       }
 514    }
 515
 516    sampler = swr_sampler_soa_create(key.sampler, PIPE_SHADER_FRAGMENT);
 517
 518    struct lp_bld_tgsi_system_values system_values;
 519    memset(&system_values, 0, sizeof(system_values));
 520
 521    struct lp_build_mask_context mask;
 522
 523    if (swr_fs->info.base.uses_kill) {
 524       Value *mask_val = LOAD(pPS, {0, SWR_PS_CONTEXT_activeMask}, "activeMask");
 525       lp_build_mask_begin(
 526          &mask, gallivm, lp_type_float_vec(32, 32 * 8), wrap(mask_val));
 527    }
 528
 529    lp_build_tgsi_soa(gallivm,
 530                      swr_fs->pipe.tokens,
 531                      lp_type_float_vec(32, 32 * 8),
 532                      swr_fs->info.base.uses_kill ? &mask : NULL, // mask
 533                      wrap(consts_ptr),
 534                      wrap(const_sizes_ptr),
 535                      &system_values,
 536                      inputs,
 537                      outputs,
 538                      wrap(hPrivateData),
 539                      NULL, // thread data
 540                      sampler, // sampler
 541                      &swr_fs->info.base,
 542                      NULL); // geometry shader face
 543
 544    sampler->destroy(sampler);
 545
 546    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 547
 548    for (uint32_t attrib = 0; attrib < swr_fs->info.base.num_outputs;
 549         attrib++) {
 550       switch (swr_fs->info.base.output_semantic_name[attrib]) {
 551       case TGSI_SEMANTIC_POSITION: {
 552          // write z
 553          LLVMValueRef outZ =
 554             LLVMBuildLoad(gallivm->builder, outputs[attrib][2], "");
 555          STORE(unwrap(outZ), pPS, {0, SWR_PS_CONTEXT_vZ});
 556          break;
 557       }
 558       case TGSI_SEMANTIC_COLOR: {
 559          for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 560             if (!outputs[attrib][channel])
 561                continue;
 562
 563             LLVMValueRef out =
 564                LLVMBuildLoad(gallivm->builder, outputs[attrib][channel], "");
 565             if (swr_fs->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS]) {
 566                for (uint32_t rt = 0; rt < key.nr_cbufs; rt++) {
 567                   STORE(unwrap(out),
 568                         pPS,
 569                         {0, SWR_PS_CONTEXT_shaded, rt, channel});
 570                }
 571             } else {
 572                STORE(unwrap(out),
 573                      pPS,
 574                      {0,
 575                            SWR_PS_CONTEXT_shaded,
 576                            swr_fs->info.base.output_semantic_index[attrib],
 577                            channel});
 578             }
 579          }
 580          break;
 581       }
 582       default: {
 583          fprintf(stderr,
 584                  "unknown output from FS %s[%d]\n",
 585                  tgsi_semantic_names[swr_fs->info.base
 586                                         .output_semantic_name[attrib]],
 587                  swr_fs->info.base.output_semantic_index[attrib]);
 588          break;
 589       }
 590       }
 591    }
 592
 593    LLVMValueRef mask_result = 0;
 594    if (swr_fs->info.base.uses_kill) {
 595       mask_result = lp_build_mask_end(&mask);
 596    }
 597
 598    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 599
 600    if (swr_fs->info.base.uses_kill) {
 601       STORE(unwrap(mask_result), pPS, {0, SWR_PS_CONTEXT_activeMask});
 602    }
 603
 604    RET_VOID();
 605
 606    gallivm_verify_function(gallivm, wrap(pFunction));
 607
 608    gallivm_compile_module(gallivm);
 609
 610    PFN_PIXEL_KERNEL kernel =
 611       (PFN_PIXEL_KERNEL)gallivm_jit_function(gallivm, wrap(pFunction));
 612    debug_printf("frag shader  %p\n", kernel);
 613    assert(kernel && "Error: FragShader = NULL");
 614
 615 #if (LLVM_VERSION_MAJOR == 3) && (LLVM_VERSION_MINOR >= 5)
 616    JM()->mIsModuleFinalized = true;
 617 #endif
 618
 619    return kernel;
 620 }
 621
 622 PFN_PIXEL_KERNEL
 623 swr_compile_fs(struct swr_context *ctx, swr_jit_fs_key &key)
 624 {
 625    BuilderSWR builder(
 626       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
 627       "FS");
 628    PFN_PIXEL_KERNEL func = builder.CompileFS(ctx, key);
 629
 630    ctx->fs->map.insert(std::make_pair(key, make_unique<VariantFS>(builder.gallivm, func)));
 631    return func;
 632 }