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 "util/u_format.h"
  38 #include "gallivm/lp_bld_init.h"
  39 #include "gallivm/lp_bld_flow.h"
  40 #include "gallivm/lp_bld_struct.h"
  41 #include "gallivm/lp_bld_tgsi.h"
  42
  43 #include "swr_context.h"
  44 #include "swr_context_llvm.h"
  45 #include "swr_resource.h"
  46 #include "swr_state.h"
  47 #include "swr_screen.h"
  48
  49 using namespace SwrJit;
  50
  51 static unsigned
  52 locate_linkage(ubyte name, ubyte index, struct tgsi_shader_info *info);
  53
  54 bool operator==(const swr_jit_fs_key &lhs, const swr_jit_fs_key &rhs)
  55 {
  56    return !memcmp(&lhs, &rhs, sizeof(lhs));
  57 }
  58
  59 bool operator==(const swr_jit_vs_key &lhs, const swr_jit_vs_key &rhs)
  60 {
  61    return !memcmp(&lhs, &rhs, sizeof(lhs));
  62 }
  63
  64 static void
  65 swr_generate_sampler_key(const struct lp_tgsi_info &info,
  66                          struct swr_context *ctx,
  67                          unsigned shader_type,
  68                          struct swr_jit_sampler_key &key)
  69 {
  70    key.nr_samplers = info.base.file_max[TGSI_FILE_SAMPLER] + 1;
  71
  72    for (unsigned i = 0; i < key.nr_samplers; i++) {
  73       if (info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
  74          lp_sampler_static_sampler_state(
  75             &key.sampler[i].sampler_state,
  76             ctx->samplers[shader_type][i]);
  77       }
  78    }
  79
  80    /*
  81     * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
  82     * are dx10-style? Can't really have mixed opcodes, at least not
  83     * if we want to skip the holes here (without rescanning tgsi).
  84     */
  85    if (info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
  86       key.nr_sampler_views =
  87          info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
  88       for (unsigned i = 0; i < key.nr_sampler_views; i++) {
  89          if (info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1 << i)) {
  90             const struct pipe_sampler_view *view =
  91                ctx->sampler_views[shader_type][i];
  92             lp_sampler_static_texture_state(
  93                &key.sampler[i].texture_state, view);
  94             if (view) {
  95                struct swr_resource *swr_res = swr_resource(view->texture);
  96                const struct util_format_description *desc =
  97                   util_format_description(view->format);
  98                if (swr_res->has_depth && swr_res->has_stencil &&
  99                    !util_format_has_depth(desc))
 100                   key.sampler[i].texture_state.format = PIPE_FORMAT_S8_UINT;
 101             }
 102          }
 103       }
 104    } else {
 105       key.nr_sampler_views = key.nr_samplers;
 106       for (unsigned i = 0; i < key.nr_sampler_views; i++) {
 107          if (info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
 108             const struct pipe_sampler_view *view =
 109                ctx->sampler_views[shader_type][i];
 110             lp_sampler_static_texture_state(
 111                &key.sampler[i].texture_state, view);
 112             if (view) {
 113                struct swr_resource *swr_res = swr_resource(view->texture);
 114                const struct util_format_description *desc =
 115                   util_format_description(view->format);
 116                if (swr_res->has_depth && swr_res->has_stencil &&
 117                    !util_format_has_depth(desc))
 118                   key.sampler[i].texture_state.format = PIPE_FORMAT_S8_UINT;
 119             }
 120          }
 121       }
 122    }
 123 }
 124
 125 void
 126 swr_generate_fs_key(struct swr_jit_fs_key &key,
 127                     struct swr_context *ctx,
 128                     swr_fragment_shader *swr_fs)
 129 {
 130    memset(&key, 0, sizeof(key));
 131
 132    key.nr_cbufs = ctx->framebuffer.nr_cbufs;
 133    key.light_twoside = ctx->rasterizer->light_twoside;
 134    memcpy(&key.vs_output_semantic_name,
 135           &ctx->vs->info.base.output_semantic_name,
 136           sizeof(key.vs_output_semantic_name));
 137    memcpy(&key.vs_output_semantic_idx,
 138           &ctx->vs->info.base.output_semantic_index,
 139           sizeof(key.vs_output_semantic_idx));
 140
 141    swr_generate_sampler_key(swr_fs->info, ctx, PIPE_SHADER_FRAGMENT, key);
 142 }
 143
 144 void
 145 swr_generate_vs_key(struct swr_jit_vs_key &key,
 146                     struct swr_context *ctx,
 147                     swr_vertex_shader *swr_vs)
 148 {
 149    memset(&key, 0, sizeof(key));
 150
 151    key.clip_plane_mask =
 152       swr_vs->info.base.clipdist_writemask ?
 153       swr_vs->info.base.clipdist_writemask & ctx->rasterizer->clip_plane_enable :
 154       ctx->rasterizer->clip_plane_enable;
 155
 156    swr_generate_sampler_key(swr_vs->info, ctx, PIPE_SHADER_VERTEX, key);
 157 }
 158
 159 struct BuilderSWR : public Builder {
 160    BuilderSWR(JitManager *pJitMgr, const char *pName)
 161       : Builder(pJitMgr)
 162    {
 163       pJitMgr->SetupNewModule();
 164       gallivm = gallivm_create(pName, wrap(&JM()->mContext));
 165       pJitMgr->mpCurrentModule = unwrap(gallivm->module);
 166    }
 167
 168    ~BuilderSWR() {
 169       gallivm_free_ir(gallivm);
 170    }
 171
 172    struct gallivm_state *gallivm;
 173    PFN_VERTEX_FUNC CompileVS(struct swr_context *ctx, swr_jit_vs_key &key);
 174    PFN_PIXEL_KERNEL CompileFS(struct swr_context *ctx, swr_jit_fs_key &key);
 175 };
 176
 177 PFN_VERTEX_FUNC
 178 BuilderSWR::CompileVS(struct swr_context *ctx, swr_jit_vs_key &key)
 179 {
 180    struct swr_vertex_shader *swr_vs = ctx->vs;
 181
 182    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 183    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 184
 185    memset(outputs, 0, sizeof(outputs));
 186
 187    AttrBuilder attrBuilder;
 188    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 189    AttributeSet attrSet = AttributeSet::get(
 190       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 191
 192    std::vector<Type *> vsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 193                               PointerType::get(Gen_SWR_VS_CONTEXT(JM()), 0)};
 194    FunctionType *vsFuncType =
 195       FunctionType::get(Type::getVoidTy(JM()->mContext), vsArgs, false);
 196
 197    // create new vertex shader function
 198    auto pFunction = Function::Create(vsFuncType,
 199                                      GlobalValue::ExternalLinkage,
 200                                      "VS",
 201                                      JM()->mpCurrentModule);
 202    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 203
 204    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 205    IRB()->SetInsertPoint(block);
 206    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 207
 208    auto argitr = pFunction->arg_begin();
 209    Value *hPrivateData = &*argitr++;
 210    hPrivateData->setName("hPrivateData");
 211    Value *pVsCtx = &*argitr++;
 212    pVsCtx->setName("vsCtx");
 213
 214    Value *consts_ptr = GEP(hPrivateData, {C(0), C(swr_draw_context_constantVS)});
 215
 216    consts_ptr->setName("vs_constants");
 217    Value *const_sizes_ptr =
 218       GEP(hPrivateData, {0, swr_draw_context_num_constantsVS});
 219    const_sizes_ptr->setName("num_vs_constants");
 220
 221    Value *vtxInput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVin});
 222
 223    for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
 224       const unsigned mask = swr_vs->info.base.input_usage_mask[attrib];
 225       for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 226          if (mask & (1 << channel)) {
 227             inputs[attrib][channel] =
 228                wrap(LOAD(vtxInput, {0, 0, attrib, channel}));
 229          }
 230       }
 231    }
 232
 233    struct lp_build_sampler_soa *sampler =
 234       swr_sampler_soa_create(key.sampler, PIPE_SHADER_VERTEX);
 235
 236    struct lp_bld_tgsi_system_values system_values;
 237    memset(&system_values, 0, sizeof(system_values));
 238    system_values.instance_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_InstanceID}));
 239    system_values.vertex_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_VertexID}));
 240
 241    lp_build_tgsi_soa(gallivm,
 242                      swr_vs->pipe.tokens,
 243                      lp_type_float_vec(32, 32 * 8),
 244                      NULL, // mask
 245                      wrap(consts_ptr),
 246                      wrap(const_sizes_ptr),
 247                      &system_values,
 248                      inputs,
 249                      outputs,
 250                      wrap(hPrivateData), // (sampler context)
 251                      NULL, // thread data
 252                      sampler, // sampler
 253                      &swr_vs->info.base,
 254                      NULL); // geometry shader face
 255
 256    sampler->destroy(sampler);
 257
 258    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 259
 260    Value *vtxOutput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVout});
 261
 262    for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 263       for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_OUTPUTS; attrib++) {
 264          if (!outputs[attrib][channel])
 265             continue;
 266
 267          Value *val = LOAD(unwrap(outputs[attrib][channel]));
 268
 269          uint32_t outSlot = attrib;
 270          if (swr_vs->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_PSIZE)
 271             outSlot = VERTEX_POINT_SIZE_SLOT;
 272          STORE(val, vtxOutput, {0, 0, outSlot, channel});
 273       }
 274    }
 275
 276    if (ctx->rasterizer->clip_plane_enable ||
 277        swr_vs->info.base.culldist_writemask) {
 278       unsigned clip_mask = ctx->rasterizer->clip_plane_enable;
 279
 280       unsigned cv = 0;
 281       if (swr_vs->info.base.writes_clipvertex) {
 282          cv = 1 + locate_linkage(TGSI_SEMANTIC_CLIPVERTEX, 0,
 283                                  &swr_vs->info.base);
 284       } else {
 285          for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 286             if (swr_vs->info.base.output_semantic_name[i] == TGSI_SEMANTIC_POSITION &&
 287                 swr_vs->info.base.output_semantic_index[i] == 0) {
 288                cv = i;
 289                break;
 290             }
 291          }
 292       }
 293       LLVMValueRef cx = LLVMBuildLoad(gallivm->builder, outputs[cv][0], "");
 294       LLVMValueRef cy = LLVMBuildLoad(gallivm->builder, outputs[cv][1], "");
 295       LLVMValueRef cz = LLVMBuildLoad(gallivm->builder, outputs[cv][2], "");
 296       LLVMValueRef cw = LLVMBuildLoad(gallivm->builder, outputs[cv][3], "");
 297
 298       for (unsigned val = 0; val < PIPE_MAX_CLIP_PLANES; val++) {
 299          // clip distance overrides user clip planes
 300          if ((swr_vs->info.base.clipdist_writemask & clip_mask & (1 << val)) ||
 301              ((swr_vs->info.base.culldist_writemask << swr_vs->info.base.num_written_clipdistance) & (1 << val))) {
 302             unsigned cv = 1 + locate_linkage(TGSI_SEMANTIC_CLIPDIST, val < 4 ? 0 : 1,
 303                                              &swr_vs->info.base);
 304             if (val < 4) {
 305                LLVMValueRef dist = LLVMBuildLoad(gallivm->builder, outputs[cv][val], "");
 306                STORE(unwrap(dist), vtxOutput, {0, 0, VERTEX_CLIPCULL_DIST_LO_SLOT, val});
 307             } else {
 308                LLVMValueRef dist = LLVMBuildLoad(gallivm->builder, outputs[cv][val - 4], "");
 309                STORE(unwrap(dist), vtxOutput, {0, 0, VERTEX_CLIPCULL_DIST_HI_SLOT, val - 4});
 310             }
 311             continue;
 312          }
 313
 314          if (!(clip_mask & (1 << val)))
 315             continue;
 316
 317          Value *px = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 0}));
 318          Value *py = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 1}));
 319          Value *pz = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 2}));
 320          Value *pw = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 3}));
 321          Value *dist = FADD(FMUL(unwrap(cx), VBROADCAST(px)),
 322                             FADD(FMUL(unwrap(cy), VBROADCAST(py)),
 323                                  FADD(FMUL(unwrap(cz), VBROADCAST(pz)),
 324                                       FMUL(unwrap(cw), VBROADCAST(pw)))));
 325
 326          if (val < 4)
 327             STORE(dist, vtxOutput, {0, 0, VERTEX_CLIPCULL_DIST_LO_SLOT, val});
 328          else
 329             STORE(dist, vtxOutput, {0, 0, VERTEX_CLIPCULL_DIST_HI_SLOT, val - 4});
 330       }
 331    }
 332
 333    RET_VOID();
 334
 335    gallivm_verify_function(gallivm, wrap(pFunction));
 336    gallivm_compile_module(gallivm);
 337
 338    //   lp_debug_dump_value(func);
 339
 340    PFN_VERTEX_FUNC pFunc =
 341       (PFN_VERTEX_FUNC)gallivm_jit_function(gallivm, wrap(pFunction));
 342
 343    debug_printf("vert shader  %p\n", pFunc);
 344    assert(pFunc && "Error: VertShader = NULL");
 345
 346 #if (LLVM_VERSION_MAJOR == 3) && (LLVM_VERSION_MINOR >= 5)
 347    JM()->mIsModuleFinalized = true;
 348 #endif
 349
 350    return pFunc;
 351 }
 352
 353 PFN_VERTEX_FUNC
 354 swr_compile_vs(struct swr_context *ctx, swr_jit_vs_key &key)
 355 {
 356    BuilderSWR builder(
 357       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
 358       "VS");
 359    PFN_VERTEX_FUNC func = builder.CompileVS(ctx, key);
 360
 361    ctx->vs->map.insert(std::make_pair(key, make_unique<VariantVS>(builder.gallivm, func)));
 362    return func;
 363 }
 364
 365 static unsigned
 366 locate_linkage(ubyte name, ubyte index, struct tgsi_shader_info *info)
 367 {
 368    for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 369       if ((info->output_semantic_name[i] == name)
 370           && (info->output_semantic_index[i] == index)) {
 371          return i - 1; // position is not part of the linkage
 372       }
 373    }
 374
 375    if (name == TGSI_SEMANTIC_COLOR) { // BCOLOR fallback
 376       for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 377          if ((info->output_semantic_name[i] == TGSI_SEMANTIC_BCOLOR)
 378              && (info->output_semantic_index[i] == index)) {
 379             return i - 1; // position is not part of the linkage
 380          }
 381       }
 382    }
 383
 384    return 0xFFFFFFFF;
 385 }
 386
 387 PFN_PIXEL_KERNEL
 388 BuilderSWR::CompileFS(struct swr_context *ctx, swr_jit_fs_key &key)
 389 {
 390    struct swr_fragment_shader *swr_fs = ctx->fs;
 391
 392    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 393    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 394
 395    memset(inputs, 0, sizeof(inputs));
 396    memset(outputs, 0, sizeof(outputs));
 397
 398    struct lp_build_sampler_soa *sampler = NULL;
 399
 400    AttrBuilder attrBuilder;
 401    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 402    AttributeSet attrSet = AttributeSet::get(
 403       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 404
 405    std::vector<Type *> fsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 406                               PointerType::get(Gen_SWR_PS_CONTEXT(JM()), 0)};
 407    FunctionType *funcType =
 408       FunctionType::get(Type::getVoidTy(JM()->mContext), fsArgs, false);
 409
 410    auto pFunction = Function::Create(funcType,
 411                                      GlobalValue::ExternalLinkage,
 412                                      "FS",
 413                                      JM()->mpCurrentModule);
 414    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 415
 416    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 417    IRB()->SetInsertPoint(block);
 418    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 419
 420    auto args = pFunction->arg_begin();
 421    Value *hPrivateData = &*args++;
 422    hPrivateData->setName("hPrivateData");
 423    Value *pPS = &*args++;
 424    pPS->setName("psCtx");
 425
 426    Value *consts_ptr = GEP(hPrivateData, {0, swr_draw_context_constantFS});
 427    consts_ptr->setName("fs_constants");
 428    Value *const_sizes_ptr =
 429       GEP(hPrivateData, {0, swr_draw_context_num_constantsFS});
 430    const_sizes_ptr->setName("num_fs_constants");
 431
 432    // load *pAttribs, *pPerspAttribs
 433    Value *pRawAttribs = LOAD(pPS, {0, SWR_PS_CONTEXT_pAttribs}, "pRawAttribs");
 434    Value *pPerspAttribs =
 435       LOAD(pPS, {0, SWR_PS_CONTEXT_pPerspAttribs}, "pPerspAttribs");
 436
 437    swr_fs->constantMask = 0;
 438    swr_fs->flatConstantMask = 0;
 439    swr_fs->pointSpriteMask = 0;
 440
 441    for (int attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
 442       const unsigned mask = swr_fs->info.base.input_usage_mask[attrib];
 443       const unsigned interpMode = swr_fs->info.base.input_interpolate[attrib];
 444       const unsigned interpLoc = swr_fs->info.base.input_interpolate_loc[attrib];
 445
 446       if (!mask)
 447          continue;
 448
 449       // load i,j
 450       Value *vi = nullptr, *vj = nullptr;
 451       switch (interpLoc) {
 452       case TGSI_INTERPOLATE_LOC_CENTER:
 453          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_center}, "i");
 454          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_center}, "j");
 455          break;
 456       case TGSI_INTERPOLATE_LOC_CENTROID:
 457          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_centroid}, "i");
 458          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_centroid}, "j");
 459          break;
 460       case TGSI_INTERPOLATE_LOC_SAMPLE:
 461          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_sample}, "i");
 462          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_sample}, "j");
 463          break;
 464       }
 465
 466       // load/compute w
 467       Value *vw = nullptr, *pAttribs;
 468       if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE) {
 469          pAttribs = pPerspAttribs;
 470          switch (interpLoc) {
 471          case TGSI_INTERPOLATE_LOC_CENTER:
 472             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}));
 473             break;
 474          case TGSI_INTERPOLATE_LOC_CENTROID:
 475             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_centroid}));
 476             break;
 477          case TGSI_INTERPOLATE_LOC_SAMPLE:
 478             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_sample}));
 479             break;
 480          }
 481       } else {
 482          pAttribs = pRawAttribs;
 483          vw = VIMMED1(1.f);
 484       }
 485
 486       vw->setName("w");
 487
 488       ubyte semantic_name = swr_fs->info.base.input_semantic_name[attrib];
 489       ubyte semantic_idx = swr_fs->info.base.input_semantic_index[attrib];
 490
 491       if (semantic_name == TGSI_SEMANTIC_FACE) {
 492          Value *ff =
 493             UI_TO_FP(LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), mFP32Ty);
 494          ff = FSUB(FMUL(ff, C(2.0f)), C(1.0f));
 495          ff = VECTOR_SPLAT(JM()->mVWidth, ff, "vFrontFace");
 496
 497          inputs[attrib][0] = wrap(ff);
 498          inputs[attrib][1] = wrap(VIMMED1(0.0f));
 499          inputs[attrib][2] = wrap(VIMMED1(0.0f));
 500          inputs[attrib][3] = wrap(VIMMED1(1.0f));
 501          continue;
 502       } else if (semantic_name == TGSI_SEMANTIC_POSITION) { // gl_FragCoord
 503          if (swr_fs->info.base.properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER] ==
 504              TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER) {
 505             inputs[attrib][0] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_center}, "vX"));
 506             inputs[attrib][1] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_center}, "vY"));
 507          } else {
 508             inputs[attrib][0] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_UL}, "vX"));
 509             inputs[attrib][1] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_UL}, "vY"));
 510          }
 511          inputs[attrib][2] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vZ}, "vZ"));
 512          inputs[attrib][3] =
 513             wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}, "vOneOverW"));
 514          continue;
 515       } else if (semantic_name == TGSI_SEMANTIC_PRIMID) {
 516          Value *primID = LOAD(pPS, {0, SWR_PS_CONTEXT_primID}, "primID");
 517          inputs[attrib][0] = wrap(VECTOR_SPLAT(JM()->mVWidth, primID));
 518          inputs[attrib][1] = wrap(VIMMED1(0));
 519          inputs[attrib][2] = wrap(VIMMED1(0));
 520          inputs[attrib][3] = wrap(VIMMED1(0));
 521          continue;
 522       }
 523
 524       unsigned linkedAttrib =
 525          locate_linkage(semantic_name, semantic_idx, &ctx->vs->info.base);
 526       if (linkedAttrib == 0xFFFFFFFF) {
 527          // not found - check for point sprite
 528          if (ctx->rasterizer->sprite_coord_enable) {
 529             linkedAttrib = ctx->vs->info.base.num_outputs - 1;
 530             swr_fs->pointSpriteMask |= (1 << linkedAttrib);
 531          } else {
 532             fprintf(stderr,
 533                     "Missing %s[%d]\n",
 534                     tgsi_semantic_names[semantic_name],
 535                     semantic_idx);
 536             assert(0 && "attribute linkage not found");
 537          }
 538       }
 539
 540       if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
 541          swr_fs->constantMask |= 1 << linkedAttrib;
 542       } else if (interpMode == TGSI_INTERPOLATE_COLOR) {
 543          swr_fs->flatConstantMask |= 1 << linkedAttrib;
 544       }
 545
 546       for (int channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 547          if (mask & (1 << channel)) {
 548             Value *indexA = C(linkedAttrib * 12 + channel);
 549             Value *indexB = C(linkedAttrib * 12 + channel + 4);
 550             Value *indexC = C(linkedAttrib * 12 + channel + 8);
 551
 552             if ((semantic_name == TGSI_SEMANTIC_COLOR)
 553                 && ctx->rasterizer->light_twoside) {
 554                unsigned bcolorAttrib = locate_linkage(
 555                   TGSI_SEMANTIC_BCOLOR, semantic_idx, &ctx->vs->info.base);
 556
 557                unsigned diff = 12 * (bcolorAttrib - linkedAttrib);
 558
 559                Value *back =
 560                   XOR(C(1), LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), "backFace");
 561
 562                Value *offset = MUL(back, C(diff));
 563                offset->setName("offset");
 564
 565                indexA = ADD(indexA, offset);
 566                indexB = ADD(indexB, offset);
 567                indexC = ADD(indexC, offset);
 568
 569                if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
 570                   swr_fs->constantMask |= 1 << bcolorAttrib;
 571                } else if (interpMode == TGSI_INTERPOLATE_COLOR) {
 572                   swr_fs->flatConstantMask |= 1 << bcolorAttrib;
 573                }
 574             }
 575
 576             Value *va = VBROADCAST(LOAD(GEP(pAttribs, indexA)));
 577             Value *vb = VBROADCAST(LOAD(GEP(pAttribs, indexB)));
 578             Value *vc = VBROADCAST(LOAD(GEP(pAttribs, indexC)));
 579
 580             if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
 581                inputs[attrib][channel] = wrap(va);
 582             } else {
 583                Value *vk = FSUB(FSUB(VIMMED1(1.0f), vi), vj);
 584
 585                vc = FMUL(vk, vc);
 586
 587                Value *interp = FMUL(va, vi);
 588                Value *interp1 = FMUL(vb, vj);
 589                interp = FADD(interp, interp1);
 590                interp = FADD(interp, vc);
 591                if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE)
 592                   interp = FMUL(interp, vw);
 593                inputs[attrib][channel] = wrap(interp);
 594             }
 595          }
 596       }
 597    }
 598
 599    sampler = swr_sampler_soa_create(key.sampler, PIPE_SHADER_FRAGMENT);
 600
 601    struct lp_bld_tgsi_system_values system_values;
 602    memset(&system_values, 0, sizeof(system_values));
 603
 604    struct lp_build_mask_context mask;
 605
 606    if (swr_fs->info.base.uses_kill) {
 607       Value *mask_val = LOAD(pPS, {0, SWR_PS_CONTEXT_activeMask}, "activeMask");
 608       lp_build_mask_begin(
 609          &mask, gallivm, lp_type_float_vec(32, 32 * 8), wrap(mask_val));
 610    }
 611
 612    lp_build_tgsi_soa(gallivm,
 613                      swr_fs->pipe.tokens,
 614                      lp_type_float_vec(32, 32 * 8),
 615                      swr_fs->info.base.uses_kill ? &mask : NULL, // mask
 616                      wrap(consts_ptr),
 617                      wrap(const_sizes_ptr),
 618                      &system_values,
 619                      inputs,
 620                      outputs,
 621                      wrap(hPrivateData),
 622                      NULL, // thread data
 623                      sampler, // sampler
 624                      &swr_fs->info.base,
 625                      NULL); // geometry shader face
 626
 627    sampler->destroy(sampler);
 628
 629    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 630
 631    for (uint32_t attrib = 0; attrib < swr_fs->info.base.num_outputs;
 632         attrib++) {
 633       switch (swr_fs->info.base.output_semantic_name[attrib]) {
 634       case TGSI_SEMANTIC_POSITION: {
 635          // write z
 636          LLVMValueRef outZ =
 637             LLVMBuildLoad(gallivm->builder, outputs[attrib][2], "");
 638          STORE(unwrap(outZ), pPS, {0, SWR_PS_CONTEXT_vZ});
 639          break;
 640       }
 641       case TGSI_SEMANTIC_COLOR: {
 642          for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 643             if (!outputs[attrib][channel])
 644                continue;
 645
 646             LLVMValueRef out =
 647                LLVMBuildLoad(gallivm->builder, outputs[attrib][channel], "");
 648             if (swr_fs->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS]) {
 649                for (uint32_t rt = 0; rt < key.nr_cbufs; rt++) {
 650                   STORE(unwrap(out),
 651                         pPS,
 652                         {0, SWR_PS_CONTEXT_shaded, rt, channel});
 653                }
 654             } else {
 655                STORE(unwrap(out),
 656                      pPS,
 657                      {0,
 658                            SWR_PS_CONTEXT_shaded,
 659                            swr_fs->info.base.output_semantic_index[attrib],
 660                            channel});
 661             }
 662          }
 663          break;
 664       }
 665       default: {
 666          fprintf(stderr,
 667                  "unknown output from FS %s[%d]\n",
 668                  tgsi_semantic_names[swr_fs->info.base
 669                                         .output_semantic_name[attrib]],
 670                  swr_fs->info.base.output_semantic_index[attrib]);
 671          break;
 672       }
 673       }
 674    }
 675
 676    LLVMValueRef mask_result = 0;
 677    if (swr_fs->info.base.uses_kill) {
 678       mask_result = lp_build_mask_end(&mask);
 679    }
 680
 681    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 682
 683    if (swr_fs->info.base.uses_kill) {
 684       STORE(unwrap(mask_result), pPS, {0, SWR_PS_CONTEXT_activeMask});
 685    }
 686
 687    RET_VOID();
 688
 689    gallivm_verify_function(gallivm, wrap(pFunction));
 690
 691    gallivm_compile_module(gallivm);
 692
 693    PFN_PIXEL_KERNEL kernel =
 694       (PFN_PIXEL_KERNEL)gallivm_jit_function(gallivm, wrap(pFunction));
 695    debug_printf("frag shader  %p\n", kernel);
 696    assert(kernel && "Error: FragShader = NULL");
 697
 698 #if (LLVM_VERSION_MAJOR == 3) && (LLVM_VERSION_MINOR >= 5)
 699    JM()->mIsModuleFinalized = true;
 700 #endif
 701
 702    return kernel;
 703 }
 704
 705 PFN_PIXEL_KERNEL
 706 swr_compile_fs(struct swr_context *ctx, swr_jit_fs_key &key)
 707 {
 708    BuilderSWR builder(
 709       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
 710       "FS");
 711    PFN_PIXEL_KERNEL func = builder.CompileFS(ctx, key);
 712
 713    ctx->fs->map.insert(std::make_pair(key, make_unique<VariantFS>(builder.gallivm, func)));
 714    return func;
 715 }