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