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 // llvm redefines DEBUG
  25 #pragma push_macro("DEBUG")
  26 #undef DEBUG
  27 #include "JitManager.h"
  28 #include "llvm-c/Core.h"
  29 #include "llvm/Support/CBindingWrapping.h"
  30 #pragma pop_macro("DEBUG")
  31
  32 #include "state.h"
  33 #include "state_llvm.h"
  34 #include "builder.h"
  35
  36 #include "tgsi/tgsi_strings.h"
  37 #include "gallivm/lp_bld_init.h"
  38 #include "gallivm/lp_bld_flow.h"
  39 #include "gallivm/lp_bld_struct.h"
  40 #include "gallivm/lp_bld_tgsi.h"
  41
  42 #include "swr_context.h"
  43 #include "swr_context_llvm.h"
  44 #include "swr_state.h"
  45 #include "swr_screen.h"
  46
  47 using namespace SwrJit;
  48
  49 static unsigned
  50 locate_linkage(ubyte name, ubyte index, struct tgsi_shader_info *info);
  51
  52 bool operator==(const swr_jit_fs_key &lhs, const swr_jit_fs_key &rhs)
  53 {
  54    return !memcmp(&lhs, &rhs, sizeof(lhs));
  55 }
  56
  57 bool operator==(const swr_jit_vs_key &lhs, const swr_jit_vs_key &rhs)
  58 {
  59    return !memcmp(&lhs, &rhs, sizeof(lhs));
  60 }
  61
  62 static void
  63 swr_generate_sampler_key(const struct lp_tgsi_info &info,
  64                          struct swr_context *ctx,
  65                          unsigned shader_type,
  66                          struct swr_jit_sampler_key &key)
  67 {
  68    key.nr_samplers = info.base.file_max[TGSI_FILE_SAMPLER] + 1;
  69
  70    for (unsigned i = 0; i < key.nr_samplers; i++) {
  71       if (info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
  72          lp_sampler_static_sampler_state(
  73             &key.sampler[i].sampler_state,
  74             ctx->samplers[shader_type][i]);
  75       }
  76    }
  77
  78    /*
  79     * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
  80     * are dx10-style? Can't really have mixed opcodes, at least not
  81     * if we want to skip the holes here (without rescanning tgsi).
  82     */
  83    if (info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
  84       key.nr_sampler_views =
  85          info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
  86       for (unsigned i = 0; i < key.nr_sampler_views; i++) {
  87          if (info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1 << i)) {
  88             lp_sampler_static_texture_state(
  89                &key.sampler[i].texture_state,
  90                ctx->sampler_views[shader_type][i]);
  91          }
  92       }
  93    } else {
  94       key.nr_sampler_views = key.nr_samplers;
  95       for (unsigned i = 0; i < key.nr_sampler_views; i++) {
  96          if (info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
  97             lp_sampler_static_texture_state(
  98                &key.sampler[i].texture_state,
  99                ctx->sampler_views[shader_type][i]);
 100          }
 101       }
 102    }
 103 }
 104
 105 void
 106 swr_generate_fs_key(struct swr_jit_fs_key &key,
 107                     struct swr_context *ctx,
 108                     swr_fragment_shader *swr_fs)
 109 {
 110    memset(&key, 0, sizeof(key));
 111
 112    key.nr_cbufs = ctx->framebuffer.nr_cbufs;
 113    key.light_twoside = ctx->rasterizer->light_twoside;
 114    memcpy(&key.vs_output_semantic_name,
 115           &ctx->vs->info.base.output_semantic_name,
 116           sizeof(key.vs_output_semantic_name));
 117    memcpy(&key.vs_output_semantic_idx,
 118           &ctx->vs->info.base.output_semantic_index,
 119           sizeof(key.vs_output_semantic_idx));
 120
 121    swr_generate_sampler_key(swr_fs->info, ctx, PIPE_SHADER_FRAGMENT, key);
 122 }
 123
 124 void
 125 swr_generate_vs_key(struct swr_jit_vs_key &key,
 126                     struct swr_context *ctx,
 127                     swr_vertex_shader *swr_vs)
 128 {
 129    memset(&key, 0, sizeof(key));
 130
 131    key.clip_plane_mask =
 132       swr_vs->info.base.clipdist_writemask ?
 133       swr_vs->info.base.clipdist_writemask & ctx->rasterizer->clip_plane_enable :
 134       ctx->rasterizer->clip_plane_enable;
 135
 136    swr_generate_sampler_key(swr_vs->info, ctx, PIPE_SHADER_VERTEX, key);
 137 }
 138
 139 struct BuilderSWR : public Builder {
 140    BuilderSWR(JitManager *pJitMgr, const char *pName)
 141       : Builder(pJitMgr)
 142    {
 143       pJitMgr->SetupNewModule();
 144       gallivm = gallivm_create(pName, wrap(&JM()->mContext));
 145       pJitMgr->mpCurrentModule = unwrap(gallivm->module);
 146    }
 147
 148    ~BuilderSWR() {
 149       gallivm_free_ir(gallivm);
 150    }
 151
 152    struct gallivm_state *gallivm;
 153    PFN_VERTEX_FUNC CompileVS(struct swr_context *ctx, swr_jit_vs_key &key);
 154    PFN_PIXEL_KERNEL CompileFS(struct swr_context *ctx, swr_jit_fs_key &key);
 155 };
 156
 157 PFN_VERTEX_FUNC
 158 BuilderSWR::CompileVS(struct swr_context *ctx, swr_jit_vs_key &key)
 159 {
 160    struct swr_vertex_shader *swr_vs = ctx->vs;
 161
 162    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 163    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 164
 165    memset(outputs, 0, sizeof(outputs));
 166
 167    AttrBuilder attrBuilder;
 168    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 169    AttributeSet attrSet = AttributeSet::get(
 170       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 171
 172    std::vector<Type *> vsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 173                               PointerType::get(Gen_SWR_VS_CONTEXT(JM()), 0)};
 174    FunctionType *vsFuncType =
 175       FunctionType::get(Type::getVoidTy(JM()->mContext), vsArgs, false);
 176
 177    // create new vertex shader function
 178    auto pFunction = Function::Create(vsFuncType,
 179                                      GlobalValue::ExternalLinkage,
 180                                      "VS",
 181                                      JM()->mpCurrentModule);
 182    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 183
 184    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 185    IRB()->SetInsertPoint(block);
 186    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 187
 188    auto argitr = pFunction->arg_begin();
 189    Value *hPrivateData = &*argitr++;
 190    hPrivateData->setName("hPrivateData");
 191    Value *pVsCtx = &*argitr++;
 192    pVsCtx->setName("vsCtx");
 193
 194    Value *consts_ptr = GEP(hPrivateData, {C(0), C(swr_draw_context_constantVS)});
 195
 196    consts_ptr->setName("vs_constants");
 197    Value *const_sizes_ptr =
 198       GEP(hPrivateData, {0, swr_draw_context_num_constantsVS});
 199    const_sizes_ptr->setName("num_vs_constants");
 200
 201    Value *vtxInput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVin});
 202
 203    for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
 204       const unsigned mask = swr_vs->info.base.input_usage_mask[attrib];
 205       for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 206          if (mask & (1 << channel)) {
 207             inputs[attrib][channel] =
 208                wrap(LOAD(vtxInput, {0, 0, attrib, channel}));
 209          }
 210       }
 211    }
 212
 213    struct lp_build_sampler_soa *sampler =
 214       swr_sampler_soa_create(key.sampler, PIPE_SHADER_VERTEX);
 215
 216    struct lp_bld_tgsi_system_values system_values;
 217    memset(&system_values, 0, sizeof(system_values));
 218    system_values.instance_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_InstanceID}));
 219    system_values.vertex_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_VertexID}));
 220
 221    lp_build_tgsi_soa(gallivm,
 222                      swr_vs->pipe.tokens,
 223                      lp_type_float_vec(32, 32 * 8),
 224                      NULL, // mask
 225                      wrap(consts_ptr),
 226                      wrap(const_sizes_ptr),
 227                      &system_values,
 228                      inputs,
 229                      outputs,
 230                      wrap(hPrivateData), // (sampler context)
 231                      NULL, // thread data
 232                      sampler, // sampler
 233                      &swr_vs->info.base,
 234                      NULL); // geometry shader face
 235
 236    sampler->destroy(sampler);
 237
 238    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 239
 240    Value *vtxOutput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVout});
 241
 242    for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 243       for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_OUTPUTS; attrib++) {
 244          if (!outputs[attrib][channel])
 245             continue;
 246
 247          Value *val = LOAD(unwrap(outputs[attrib][channel]));
 248
 249          uint32_t outSlot = attrib;
 250          if (swr_vs->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_PSIZE)
 251             outSlot = VERTEX_POINT_SIZE_SLOT;
 252          STORE(val, vtxOutput, {0, 0, outSlot, channel});
 253       }
 254    }
 255
 256    if (ctx->rasterizer->clip_plane_enable ||
 257        swr_vs->info.base.culldist_writemask) {
 258       unsigned clip_mask = ctx->rasterizer->clip_plane_enable;
 259
 260       unsigned cv = 0;
 261       if (swr_vs->info.base.writes_clipvertex) {
 262          cv = 1 + locate_linkage(TGSI_SEMANTIC_CLIPVERTEX, 0,
 263                                  &swr_vs->info.base);
 264       } else {
 265          for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 266             if (swr_vs->info.base.output_semantic_name[i] == TGSI_SEMANTIC_POSITION &&
 267                 swr_vs->info.base.output_semantic_index[i] == 0) {
 268                cv = i;
 269                break;
 270             }
 271          }
 272       }
 273       LLVMValueRef cx = LLVMBuildLoad(gallivm->builder, outputs[cv][0], "");
 274       LLVMValueRef cy = LLVMBuildLoad(gallivm->builder, outputs[cv][1], "");
 275       LLVMValueRef cz = LLVMBuildLoad(gallivm->builder, outputs[cv][2], "");
 276       LLVMValueRef cw = LLVMBuildLoad(gallivm->builder, outputs[cv][3], "");
 277
 278       for (unsigned val = 0; val < PIPE_MAX_CLIP_PLANES; val++) {
 279          // clip distance overrides user clip planes
 280          if ((swr_vs->info.base.clipdist_writemask & clip_mask & (1 << val)) ||
 281              ((swr_vs->info.base.culldist_writemask << swr_vs->info.base.num_written_clipdistance) & (1 << val))) {
 282             unsigned cv = 1 + locate_linkage(TGSI_SEMANTIC_CLIPDIST, val < 4 ? 0 : 1,
 283                                              &swr_vs->info.base);
 284             if (val < 4) {
 285                LLVMValueRef dist = LLVMBuildLoad(gallivm->builder, outputs[cv][val], "");
 286                STORE(unwrap(dist), vtxOutput, {0, 0, VERTEX_CLIPCULL_DIST_LO_SLOT, val});
 287             } else {
 288                LLVMValueRef dist = LLVMBuildLoad(gallivm->builder, outputs[cv][val - 4], "");
 289                STORE(unwrap(dist), vtxOutput, {0, 0, VERTEX_CLIPCULL_DIST_HI_SLOT, val - 4});
 290             }
 291             continue;
 292          }
 293
 294          if (!(clip_mask & (1 << val)))
 295             continue;
 296
 297          Value *px = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 0}));
 298          Value *py = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 1}));
 299          Value *pz = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 2}));
 300          Value *pw = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 3}));
 301          Value *dist = FADD(FMUL(unwrap(cx), VBROADCAST(px)),
 302                             FADD(FMUL(unwrap(cy), VBROADCAST(py)),
 303                                  FADD(FMUL(unwrap(cz), VBROADCAST(pz)),
 304                                       FMUL(unwrap(cw), VBROADCAST(pw)))));
 305
 306          if (val < 4)
 307             STORE(dist, vtxOutput, {0, 0, VERTEX_CLIPCULL_DIST_LO_SLOT, val});
 308          else
 309             STORE(dist, vtxOutput, {0, 0, VERTEX_CLIPCULL_DIST_HI_SLOT, val - 4});
 310       }
 311    }
 312
 313    RET_VOID();
 314
 315    gallivm_verify_function(gallivm, wrap(pFunction));
 316    gallivm_compile_module(gallivm);
 317
 318    //   lp_debug_dump_value(func);
 319
 320    PFN_VERTEX_FUNC pFunc =
 321       (PFN_VERTEX_FUNC)gallivm_jit_function(gallivm, wrap(pFunction));
 322
 323    debug_printf("vert shader  %p\n", pFunc);
 324    assert(pFunc && "Error: VertShader = NULL");
 325
 326 #if (LLVM_VERSION_MAJOR == 3) && (LLVM_VERSION_MINOR >= 5)
 327    JM()->mIsModuleFinalized = true;
 328 #endif
 329
 330    return pFunc;
 331 }
 332
 333 PFN_VERTEX_FUNC
 334 swr_compile_vs(struct swr_context *ctx, swr_jit_vs_key &key)
 335 {
 336    BuilderSWR builder(
 337       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
 338       "VS");
 339    PFN_VERTEX_FUNC func = builder.CompileVS(ctx, key);
 340
 341    ctx->vs->map.insert(std::make_pair(key, make_unique<VariantVS>(builder.gallivm, func)));
 342    return func;
 343 }
 344
 345 static unsigned
 346 locate_linkage(ubyte name, ubyte index, struct tgsi_shader_info *info)
 347 {
 348    for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 349       if ((info->output_semantic_name[i] == name)
 350           && (info->output_semantic_index[i] == index)) {
 351          return i - 1; // position is not part of the linkage
 352       }
 353    }
 354
 355    if (name == TGSI_SEMANTIC_COLOR) { // BCOLOR fallback
 356       for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 357          if ((info->output_semantic_name[i] == TGSI_SEMANTIC_BCOLOR)
 358              && (info->output_semantic_index[i] == index)) {
 359             return i - 1; // position is not part of the linkage
 360          }
 361       }
 362    }
 363
 364    return 0xFFFFFFFF;
 365 }
 366
 367 PFN_PIXEL_KERNEL
 368 BuilderSWR::CompileFS(struct swr_context *ctx, swr_jit_fs_key &key)
 369 {
 370    struct swr_fragment_shader *swr_fs = ctx->fs;
 371
 372    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 373    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 374
 375    memset(inputs, 0, sizeof(inputs));
 376    memset(outputs, 0, sizeof(outputs));
 377
 378    struct lp_build_sampler_soa *sampler = NULL;
 379
 380    AttrBuilder attrBuilder;
 381    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 382    AttributeSet attrSet = AttributeSet::get(
 383       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 384
 385    std::vector<Type *> fsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 386                               PointerType::get(Gen_SWR_PS_CONTEXT(JM()), 0)};
 387    FunctionType *funcType =
 388       FunctionType::get(Type::getVoidTy(JM()->mContext), fsArgs, false);
 389
 390    auto pFunction = Function::Create(funcType,
 391                                      GlobalValue::ExternalLinkage,
 392                                      "FS",
 393                                      JM()->mpCurrentModule);
 394    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 395
 396    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 397    IRB()->SetInsertPoint(block);
 398    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 399
 400    auto args = pFunction->arg_begin();
 401    Value *hPrivateData = &*args++;
 402    hPrivateData->setName("hPrivateData");
 403    Value *pPS = &*args++;
 404    pPS->setName("psCtx");
 405
 406    Value *consts_ptr = GEP(hPrivateData, {0, swr_draw_context_constantFS});
 407    consts_ptr->setName("fs_constants");
 408    Value *const_sizes_ptr =
 409       GEP(hPrivateData, {0, swr_draw_context_num_constantsFS});
 410    const_sizes_ptr->setName("num_fs_constants");
 411
 412    // load *pAttribs, *pPerspAttribs
 413    Value *pRawAttribs = LOAD(pPS, {0, SWR_PS_CONTEXT_pAttribs}, "pRawAttribs");
 414    Value *pPerspAttribs =
 415       LOAD(pPS, {0, SWR_PS_CONTEXT_pPerspAttribs}, "pPerspAttribs");
 416
 417    swr_fs->constantMask = 0;
 418    swr_fs->flatConstantMask = 0;
 419    swr_fs->pointSpriteMask = 0;
 420
 421    for (int attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
 422       const unsigned mask = swr_fs->info.base.input_usage_mask[attrib];
 423       const unsigned interpMode = swr_fs->info.base.input_interpolate[attrib];
 424       const unsigned interpLoc = swr_fs->info.base.input_interpolate_loc[attrib];
 425
 426       if (!mask)
 427          continue;
 428
 429       // load i,j
 430       Value *vi = nullptr, *vj = nullptr;
 431       switch (interpLoc) {
 432       case TGSI_INTERPOLATE_LOC_CENTER:
 433          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_center}, "i");
 434          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_center}, "j");
 435          break;
 436       case TGSI_INTERPOLATE_LOC_CENTROID:
 437          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_centroid}, "i");
 438          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_centroid}, "j");
 439          break;
 440       case TGSI_INTERPOLATE_LOC_SAMPLE:
 441          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_sample}, "i");
 442          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_sample}, "j");
 443          break;
 444       }
 445
 446       // load/compute w
 447       Value *vw = nullptr, *pAttribs;
 448       if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE) {
 449          pAttribs = pPerspAttribs;
 450          switch (interpLoc) {
 451          case TGSI_INTERPOLATE_LOC_CENTER:
 452             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}));
 453             break;
 454          case TGSI_INTERPOLATE_LOC_CENTROID:
 455             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_centroid}));
 456             break;
 457          case TGSI_INTERPOLATE_LOC_SAMPLE:
 458             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_sample}));
 459             break;
 460          }
 461       } else {
 462          pAttribs = pRawAttribs;
 463          vw = VIMMED1(1.f);
 464       }
 465
 466       vw->setName("w");
 467
 468       ubyte semantic_name = swr_fs->info.base.input_semantic_name[attrib];
 469       ubyte semantic_idx = swr_fs->info.base.input_semantic_index[attrib];
 470
 471       if (semantic_name == TGSI_SEMANTIC_FACE) {
 472          Value *ff =
 473             UI_TO_FP(LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), mFP32Ty);
 474          ff = FSUB(FMUL(ff, C(2.0f)), C(1.0f));
 475          ff = VECTOR_SPLAT(JM()->mVWidth, ff, "vFrontFace");
 476
 477          inputs[attrib][0] = wrap(ff);
 478          inputs[attrib][1] = wrap(VIMMED1(0.0f));
 479          inputs[attrib][2] = wrap(VIMMED1(0.0f));
 480          inputs[attrib][3] = wrap(VIMMED1(1.0f));
 481          continue;
 482       } else if (semantic_name == TGSI_SEMANTIC_POSITION) { // gl_FragCoord
 483          inputs[attrib][0] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_center}, "vX"));
 484          inputs[attrib][1] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_center}, "vY"));
 485          inputs[attrib][2] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vZ}, "vZ"));
 486          inputs[attrib][3] =
 487             wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}, "vOneOverW"));
 488          continue;
 489       } else if (semantic_name == TGSI_SEMANTIC_PRIMID) {
 490          Value *primID = LOAD(pPS, {0, SWR_PS_CONTEXT_primID}, "primID");
 491          inputs[attrib][0] = wrap(VECTOR_SPLAT(JM()->mVWidth, primID));
 492          inputs[attrib][1] = wrap(VIMMED1(0));
 493          inputs[attrib][2] = wrap(VIMMED1(0));
 494          inputs[attrib][3] = wrap(VIMMED1(0));
 495          continue;
 496       }
 497
 498       unsigned linkedAttrib =
 499          locate_linkage(semantic_name, semantic_idx, &ctx->vs->info.base);
 500       if (linkedAttrib == 0xFFFFFFFF) {
 501          // not found - check for point sprite
 502          if (ctx->rasterizer->sprite_coord_enable) {
 503             linkedAttrib = ctx->vs->info.base.num_outputs - 1;
 504             swr_fs->pointSpriteMask |= (1 << linkedAttrib);
 505          } else {
 506             fprintf(stderr,
 507                     "Missing %s[%d]\n",
 508                     tgsi_semantic_names[semantic_name],
 509                     semantic_idx);
 510             assert(0 && "attribute linkage not found");
 511          }
 512       }
 513
 514       if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
 515          swr_fs->constantMask |= 1 << linkedAttrib;
 516       } else if (interpMode == TGSI_INTERPOLATE_COLOR) {
 517          swr_fs->flatConstantMask |= 1 << linkedAttrib;
 518       }
 519
 520       for (int channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 521          if (mask & (1 << channel)) {
 522             Value *indexA = C(linkedAttrib * 12 + channel);
 523             Value *indexB = C(linkedAttrib * 12 + channel + 4);
 524             Value *indexC = C(linkedAttrib * 12 + channel + 8);
 525
 526             if ((semantic_name == TGSI_SEMANTIC_COLOR)
 527                 && ctx->rasterizer->light_twoside) {
 528                unsigned bcolorAttrib = locate_linkage(
 529                   TGSI_SEMANTIC_BCOLOR, semantic_idx, &ctx->vs->info.base);
 530
 531                unsigned diff = 12 * (bcolorAttrib - linkedAttrib);
 532
 533                Value *back =
 534                   XOR(C(1), LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), "backFace");
 535
 536                Value *offset = MUL(back, C(diff));
 537                offset->setName("offset");
 538
 539                indexA = ADD(indexA, offset);
 540                indexB = ADD(indexB, offset);
 541                indexC = ADD(indexC, offset);
 542
 543                if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
 544                   swr_fs->constantMask |= 1 << bcolorAttrib;
 545                } else if (interpMode == TGSI_INTERPOLATE_COLOR) {
 546                   swr_fs->flatConstantMask |= 1 << bcolorAttrib;
 547                }
 548             }
 549
 550             Value *va = VBROADCAST(LOAD(GEP(pAttribs, indexA)));
 551             Value *vb = VBROADCAST(LOAD(GEP(pAttribs, indexB)));
 552             Value *vc = VBROADCAST(LOAD(GEP(pAttribs, indexC)));
 553
 554             if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
 555                inputs[attrib][channel] = wrap(va);
 556             } else {
 557                Value *vk = FSUB(FSUB(VIMMED1(1.0f), vi), vj);
 558
 559                vc = FMUL(vk, vc);
 560
 561                Value *interp = FMUL(va, vi);
 562                Value *interp1 = FMUL(vb, vj);
 563                interp = FADD(interp, interp1);
 564                interp = FADD(interp, vc);
 565                if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE)
 566                   interp = FMUL(interp, vw);
 567                inputs[attrib][channel] = wrap(interp);
 568             }
 569          }
 570       }
 571    }
 572
 573    sampler = swr_sampler_soa_create(key.sampler, PIPE_SHADER_FRAGMENT);
 574
 575    struct lp_bld_tgsi_system_values system_values;
 576    memset(&system_values, 0, sizeof(system_values));
 577
 578    struct lp_build_mask_context mask;
 579
 580    if (swr_fs->info.base.uses_kill) {
 581       Value *mask_val = LOAD(pPS, {0, SWR_PS_CONTEXT_activeMask}, "activeMask");
 582       lp_build_mask_begin(
 583          &mask, gallivm, lp_type_float_vec(32, 32 * 8), wrap(mask_val));
 584    }
 585
 586    lp_build_tgsi_soa(gallivm,
 587                      swr_fs->pipe.tokens,
 588                      lp_type_float_vec(32, 32 * 8),
 589                      swr_fs->info.base.uses_kill ? &mask : NULL, // mask
 590                      wrap(consts_ptr),
 591                      wrap(const_sizes_ptr),
 592                      &system_values,
 593                      inputs,
 594                      outputs,
 595                      wrap(hPrivateData),
 596                      NULL, // thread data
 597                      sampler, // sampler
 598                      &swr_fs->info.base,
 599                      NULL); // geometry shader face
 600
 601    sampler->destroy(sampler);
 602
 603    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 604
 605    for (uint32_t attrib = 0; attrib < swr_fs->info.base.num_outputs;
 606         attrib++) {
 607       switch (swr_fs->info.base.output_semantic_name[attrib]) {
 608       case TGSI_SEMANTIC_POSITION: {
 609          // write z
 610          LLVMValueRef outZ =
 611             LLVMBuildLoad(gallivm->builder, outputs[attrib][2], "");
 612          STORE(unwrap(outZ), pPS, {0, SWR_PS_CONTEXT_vZ});
 613          break;
 614       }
 615       case TGSI_SEMANTIC_COLOR: {
 616          for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 617             if (!outputs[attrib][channel])
 618                continue;
 619
 620             LLVMValueRef out =
 621                LLVMBuildLoad(gallivm->builder, outputs[attrib][channel], "");
 622             if (swr_fs->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS]) {
 623                for (uint32_t rt = 0; rt < key.nr_cbufs; rt++) {
 624                   STORE(unwrap(out),
 625                         pPS,
 626                         {0, SWR_PS_CONTEXT_shaded, rt, channel});
 627                }
 628             } else {
 629                STORE(unwrap(out),
 630                      pPS,
 631                      {0,
 632                            SWR_PS_CONTEXT_shaded,
 633                            swr_fs->info.base.output_semantic_index[attrib],
 634                            channel});
 635             }
 636          }
 637          break;
 638       }
 639       default: {
 640          fprintf(stderr,
 641                  "unknown output from FS %s[%d]\n",
 642                  tgsi_semantic_names[swr_fs->info.base
 643                                         .output_semantic_name[attrib]],
 644                  swr_fs->info.base.output_semantic_index[attrib]);
 645          break;
 646       }
 647       }
 648    }
 649
 650    LLVMValueRef mask_result = 0;
 651    if (swr_fs->info.base.uses_kill) {
 652       mask_result = lp_build_mask_end(&mask);
 653    }
 654
 655    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 656
 657    if (swr_fs->info.base.uses_kill) {
 658       STORE(unwrap(mask_result), pPS, {0, SWR_PS_CONTEXT_activeMask});
 659    }
 660
 661    RET_VOID();
 662
 663    gallivm_verify_function(gallivm, wrap(pFunction));
 664
 665    gallivm_compile_module(gallivm);
 666
 667    PFN_PIXEL_KERNEL kernel =
 668       (PFN_PIXEL_KERNEL)gallivm_jit_function(gallivm, wrap(pFunction));
 669    debug_printf("frag shader  %p\n", kernel);
 670    assert(kernel && "Error: FragShader = NULL");
 671
 672 #if (LLVM_VERSION_MAJOR == 3) && (LLVM_VERSION_MINOR >= 5)
 673    JM()->mIsModuleFinalized = true;
 674 #endif
 675
 676    return kernel;
 677 }
 678
 679 PFN_PIXEL_KERNEL
 680 swr_compile_fs(struct swr_context *ctx, swr_jit_fs_key &key)
 681 {
 682    BuilderSWR builder(
 683       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
 684       "FS");
 685    PFN_PIXEL_KERNEL func = builder.CompileFS(ctx, key);
 686
 687    ctx->fs->map.insert(std::make_pair(key, make_unique<VariantFS>(builder.gallivm, func)));
 688    return func;
 689 }