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 bool operator==(const swr_jit_fetch_key &lhs, const swr_jit_fetch_key &rhs)
  65 {
  66    return !memcmp(&lhs, &rhs, sizeof(lhs));
  67 }
  68
  69 static void
  70 swr_generate_sampler_key(const struct lp_tgsi_info &info,
  71                          struct swr_context *ctx,
  72                          unsigned shader_type,
  73                          struct swr_jit_sampler_key &key)
  74 {
  75    key.nr_samplers = info.base.file_max[TGSI_FILE_SAMPLER] + 1;
  76
  77    for (unsigned i = 0; i < key.nr_samplers; i++) {
  78       if (info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
  79          lp_sampler_static_sampler_state(
  80             &key.sampler[i].sampler_state,
  81             ctx->samplers[shader_type][i]);
  82       }
  83    }
  84
  85    /*
  86     * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
  87     * are dx10-style? Can't really have mixed opcodes, at least not
  88     * if we want to skip the holes here (without rescanning tgsi).
  89     */
  90    if (info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
  91       key.nr_sampler_views =
  92          info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
  93       for (unsigned i = 0; i < key.nr_sampler_views; i++) {
  94          if (info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1 << i)) {
  95             const struct pipe_sampler_view *view =
  96                ctx->sampler_views[shader_type][i];
  97             lp_sampler_static_texture_state(
  98                &key.sampler[i].texture_state, view);
  99             if (view) {
 100                struct swr_resource *swr_res = swr_resource(view->texture);
 101                const struct util_format_description *desc =
 102                   util_format_description(view->format);
 103                if (swr_res->has_depth && swr_res->has_stencil &&
 104                    !util_format_has_depth(desc))
 105                   key.sampler[i].texture_state.format = PIPE_FORMAT_S8_UINT;
 106             }
 107          }
 108       }
 109    } else {
 110       key.nr_sampler_views = key.nr_samplers;
 111       for (unsigned i = 0; i < key.nr_sampler_views; i++) {
 112          if (info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
 113             const struct pipe_sampler_view *view =
 114                ctx->sampler_views[shader_type][i];
 115             lp_sampler_static_texture_state(
 116                &key.sampler[i].texture_state, view);
 117             if (view) {
 118                struct swr_resource *swr_res = swr_resource(view->texture);
 119                const struct util_format_description *desc =
 120                   util_format_description(view->format);
 121                if (swr_res->has_depth && swr_res->has_stencil &&
 122                    !util_format_has_depth(desc))
 123                   key.sampler[i].texture_state.format = PIPE_FORMAT_S8_UINT;
 124             }
 125          }
 126       }
 127    }
 128 }
 129
 130 void
 131 swr_generate_fs_key(struct swr_jit_fs_key &key,
 132                     struct swr_context *ctx,
 133                     swr_fragment_shader *swr_fs)
 134 {
 135    memset(&key, 0, sizeof(key));
 136
 137    key.nr_cbufs = ctx->framebuffer.nr_cbufs;
 138    key.light_twoside = ctx->rasterizer->light_twoside;
 139    key.sprite_coord_enable = ctx->rasterizer->sprite_coord_enable;
 140    memcpy(&key.vs_output_semantic_name,
 141           &ctx->vs->info.base.output_semantic_name,
 142           sizeof(key.vs_output_semantic_name));
 143    memcpy(&key.vs_output_semantic_idx,
 144           &ctx->vs->info.base.output_semantic_index,
 145           sizeof(key.vs_output_semantic_idx));
 146
 147    swr_generate_sampler_key(swr_fs->info, ctx, PIPE_SHADER_FRAGMENT, key);
 148 }
 149
 150 void
 151 swr_generate_vs_key(struct swr_jit_vs_key &key,
 152                     struct swr_context *ctx,
 153                     swr_vertex_shader *swr_vs)
 154 {
 155    memset(&key, 0, sizeof(key));
 156
 157    key.clip_plane_mask =
 158       swr_vs->info.base.clipdist_writemask ?
 159       swr_vs->info.base.clipdist_writemask & ctx->rasterizer->clip_plane_enable :
 160       ctx->rasterizer->clip_plane_enable;
 161
 162    swr_generate_sampler_key(swr_vs->info, ctx, PIPE_SHADER_VERTEX, key);
 163 }
 164
 165 void
 166 swr_generate_fetch_key(struct swr_jit_fetch_key &key,
 167                        struct swr_vertex_element_state *velems)
 168 {
 169    memset(&key, 0, sizeof(key));
 170
 171    key.fsState = velems->fsState;
 172 }
 173
 174 struct BuilderSWR : public Builder {
 175    BuilderSWR(JitManager *pJitMgr, const char *pName)
 176       : Builder(pJitMgr)
 177    {
 178       pJitMgr->SetupNewModule();
 179       gallivm = gallivm_create(pName, wrap(&JM()->mContext));
 180       pJitMgr->mpCurrentModule = unwrap(gallivm->module);
 181    }
 182
 183    ~BuilderSWR() {
 184       gallivm_free_ir(gallivm);
 185    }
 186
 187    struct gallivm_state *gallivm;
 188    PFN_VERTEX_FUNC CompileVS(struct swr_context *ctx, swr_jit_vs_key &key);
 189    PFN_PIXEL_KERNEL CompileFS(struct swr_context *ctx, swr_jit_fs_key &key);
 190 };
 191
 192 PFN_VERTEX_FUNC
 193 BuilderSWR::CompileVS(struct swr_context *ctx, swr_jit_vs_key &key)
 194 {
 195    struct swr_vertex_shader *swr_vs = ctx->vs;
 196
 197    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 198    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 199
 200    memset(outputs, 0, sizeof(outputs));
 201
 202    AttrBuilder attrBuilder;
 203    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 204    AttributeSet attrSet = AttributeSet::get(
 205       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 206
 207    std::vector<Type *> vsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 208                               PointerType::get(Gen_SWR_VS_CONTEXT(JM()), 0)};
 209    FunctionType *vsFuncType =
 210       FunctionType::get(Type::getVoidTy(JM()->mContext), vsArgs, false);
 211
 212    // create new vertex shader function
 213    auto pFunction = Function::Create(vsFuncType,
 214                                      GlobalValue::ExternalLinkage,
 215                                      "VS",
 216                                      JM()->mpCurrentModule);
 217    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 218
 219    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 220    IRB()->SetInsertPoint(block);
 221    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 222
 223    auto argitr = pFunction->arg_begin();
 224    Value *hPrivateData = &*argitr++;
 225    hPrivateData->setName("hPrivateData");
 226    Value *pVsCtx = &*argitr++;
 227    pVsCtx->setName("vsCtx");
 228
 229    Value *consts_ptr = GEP(hPrivateData, {C(0), C(swr_draw_context_constantVS)});
 230
 231    consts_ptr->setName("vs_constants");
 232    Value *const_sizes_ptr =
 233       GEP(hPrivateData, {0, swr_draw_context_num_constantsVS});
 234    const_sizes_ptr->setName("num_vs_constants");
 235
 236    Value *vtxInput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVin});
 237
 238    for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
 239       const unsigned mask = swr_vs->info.base.input_usage_mask[attrib];
 240       for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 241          if (mask & (1 << channel)) {
 242             inputs[attrib][channel] =
 243                wrap(LOAD(vtxInput, {0, 0, attrib, channel}));
 244          }
 245       }
 246    }
 247
 248    struct lp_build_sampler_soa *sampler =
 249       swr_sampler_soa_create(key.sampler, PIPE_SHADER_VERTEX);
 250
 251    struct lp_bld_tgsi_system_values system_values;
 252    memset(&system_values, 0, sizeof(system_values));
 253    system_values.instance_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_InstanceID}));
 254    system_values.vertex_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_VertexID}));
 255
 256    lp_build_tgsi_soa(gallivm,
 257                      swr_vs->pipe.tokens,
 258                      lp_type_float_vec(32, 32 * 8),
 259                      NULL, // mask
 260                      wrap(consts_ptr),
 261                      wrap(const_sizes_ptr),
 262                      &system_values,
 263                      inputs,
 264                      outputs,
 265                      wrap(hPrivateData), // (sampler context)
 266                      NULL, // thread data
 267                      sampler, // sampler
 268                      &swr_vs->info.base,
 269                      NULL); // geometry shader face
 270
 271    sampler->destroy(sampler);
 272
 273    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 274
 275    Value *vtxOutput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVout});
 276
 277    for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 278       for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_OUTPUTS; attrib++) {
 279          if (!outputs[attrib][channel])
 280             continue;
 281
 282          Value *val = LOAD(unwrap(outputs[attrib][channel]));
 283
 284          uint32_t outSlot = attrib;
 285          if (swr_vs->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_PSIZE)
 286             outSlot = VERTEX_POINT_SIZE_SLOT;
 287          STORE(val, vtxOutput, {0, 0, outSlot, channel});
 288       }
 289    }
 290
 291    if (ctx->rasterizer->clip_plane_enable ||
 292        swr_vs->info.base.culldist_writemask) {
 293       unsigned clip_mask = ctx->rasterizer->clip_plane_enable;
 294
 295       unsigned cv = 0;
 296       if (swr_vs->info.base.writes_clipvertex) {
 297          cv = 1 + locate_linkage(TGSI_SEMANTIC_CLIPVERTEX, 0,
 298                                  &swr_vs->info.base);
 299       } else {
 300          for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 301             if (swr_vs->info.base.output_semantic_name[i] == TGSI_SEMANTIC_POSITION &&
 302                 swr_vs->info.base.output_semantic_index[i] == 0) {
 303                cv = i;
 304                break;
 305             }
 306          }
 307       }
 308       LLVMValueRef cx = LLVMBuildLoad(gallivm->builder, outputs[cv][0], "");
 309       LLVMValueRef cy = LLVMBuildLoad(gallivm->builder, outputs[cv][1], "");
 310       LLVMValueRef cz = LLVMBuildLoad(gallivm->builder, outputs[cv][2], "");
 311       LLVMValueRef cw = LLVMBuildLoad(gallivm->builder, outputs[cv][3], "");
 312
 313       for (unsigned val = 0; val < PIPE_MAX_CLIP_PLANES; val++) {
 314          // clip distance overrides user clip planes
 315          if ((swr_vs->info.base.clipdist_writemask & clip_mask & (1 << val)) ||
 316              ((swr_vs->info.base.culldist_writemask << swr_vs->info.base.num_written_clipdistance) & (1 << val))) {
 317             unsigned cv = 1 + locate_linkage(TGSI_SEMANTIC_CLIPDIST, val < 4 ? 0 : 1,
 318                                              &swr_vs->info.base);
 319             if (val < 4) {
 320                LLVMValueRef dist = LLVMBuildLoad(gallivm->builder, outputs[cv][val], "");
 321                STORE(unwrap(dist), vtxOutput, {0, 0, VERTEX_CLIPCULL_DIST_LO_SLOT, val});
 322             } else {
 323                LLVMValueRef dist = LLVMBuildLoad(gallivm->builder, outputs[cv][val - 4], "");
 324                STORE(unwrap(dist), vtxOutput, {0, 0, VERTEX_CLIPCULL_DIST_HI_SLOT, val - 4});
 325             }
 326             continue;
 327          }
 328
 329          if (!(clip_mask & (1 << val)))
 330             continue;
 331
 332          Value *px = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 0}));
 333          Value *py = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 1}));
 334          Value *pz = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 2}));
 335          Value *pw = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 3}));
 336          Value *dist = FADD(FMUL(unwrap(cx), VBROADCAST(px)),
 337                             FADD(FMUL(unwrap(cy), VBROADCAST(py)),
 338                                  FADD(FMUL(unwrap(cz), VBROADCAST(pz)),
 339                                       FMUL(unwrap(cw), VBROADCAST(pw)))));
 340
 341          if (val < 4)
 342             STORE(dist, vtxOutput, {0, 0, VERTEX_CLIPCULL_DIST_LO_SLOT, val});
 343          else
 344             STORE(dist, vtxOutput, {0, 0, VERTEX_CLIPCULL_DIST_HI_SLOT, val - 4});
 345       }
 346    }
 347
 348    RET_VOID();
 349
 350    gallivm_verify_function(gallivm, wrap(pFunction));
 351    gallivm_compile_module(gallivm);
 352
 353    //   lp_debug_dump_value(func);
 354
 355    PFN_VERTEX_FUNC pFunc =
 356       (PFN_VERTEX_FUNC)gallivm_jit_function(gallivm, wrap(pFunction));
 357
 358    debug_printf("vert shader  %p\n", pFunc);
 359    assert(pFunc && "Error: VertShader = NULL");
 360
 361    JM()->mIsModuleFinalized = true;
 362
 363    return pFunc;
 364 }
 365
 366 PFN_VERTEX_FUNC
 367 swr_compile_vs(struct swr_context *ctx, swr_jit_vs_key &key)
 368 {
 369    BuilderSWR builder(
 370       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
 371       "VS");
 372    PFN_VERTEX_FUNC func = builder.CompileVS(ctx, key);
 373
 374    ctx->vs->map.insert(std::make_pair(key, make_unique<VariantVS>(builder.gallivm, func)));
 375    return func;
 376 }
 377
 378 static unsigned
 379 locate_linkage(ubyte name, ubyte index, struct tgsi_shader_info *info)
 380 {
 381    for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 382       if ((info->output_semantic_name[i] == name)
 383           && (info->output_semantic_index[i] == index)) {
 384          return i - 1; // position is not part of the linkage
 385       }
 386    }
 387
 388    return 0xFFFFFFFF;
 389 }
 390
 391 PFN_PIXEL_KERNEL
 392 BuilderSWR::CompileFS(struct swr_context *ctx, swr_jit_fs_key &key)
 393 {
 394    struct swr_fragment_shader *swr_fs = ctx->fs;
 395
 396    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 397    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 398
 399    memset(inputs, 0, sizeof(inputs));
 400    memset(outputs, 0, sizeof(outputs));
 401
 402    struct lp_build_sampler_soa *sampler = NULL;
 403
 404    AttrBuilder attrBuilder;
 405    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 406    AttributeSet attrSet = AttributeSet::get(
 407       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 408
 409    std::vector<Type *> fsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 410                               PointerType::get(Gen_SWR_PS_CONTEXT(JM()), 0)};
 411    FunctionType *funcType =
 412       FunctionType::get(Type::getVoidTy(JM()->mContext), fsArgs, false);
 413
 414    auto pFunction = Function::Create(funcType,
 415                                      GlobalValue::ExternalLinkage,
 416                                      "FS",
 417                                      JM()->mpCurrentModule);
 418    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 419
 420    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 421    IRB()->SetInsertPoint(block);
 422    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 423
 424    auto args = pFunction->arg_begin();
 425    Value *hPrivateData = &*args++;
 426    hPrivateData->setName("hPrivateData");
 427    Value *pPS = &*args++;
 428    pPS->setName("psCtx");
 429
 430    Value *consts_ptr = GEP(hPrivateData, {0, swr_draw_context_constantFS});
 431    consts_ptr->setName("fs_constants");
 432    Value *const_sizes_ptr =
 433       GEP(hPrivateData, {0, swr_draw_context_num_constantsFS});
 434    const_sizes_ptr->setName("num_fs_constants");
 435
 436    // load *pAttribs, *pPerspAttribs
 437    Value *pRawAttribs = LOAD(pPS, {0, SWR_PS_CONTEXT_pAttribs}, "pRawAttribs");
 438    Value *pPerspAttribs =
 439       LOAD(pPS, {0, SWR_PS_CONTEXT_pPerspAttribs}, "pPerspAttribs");
 440
 441    swr_fs->constantMask = 0;
 442    swr_fs->flatConstantMask = 0;
 443    swr_fs->pointSpriteMask = 0;
 444
 445    for (int attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
 446       const unsigned mask = swr_fs->info.base.input_usage_mask[attrib];
 447       const unsigned interpMode = swr_fs->info.base.input_interpolate[attrib];
 448       const unsigned interpLoc = swr_fs->info.base.input_interpolate_loc[attrib];
 449
 450       if (!mask)
 451          continue;
 452
 453       // load i,j
 454       Value *vi = nullptr, *vj = nullptr;
 455       switch (interpLoc) {
 456       case TGSI_INTERPOLATE_LOC_CENTER:
 457          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_center}, "i");
 458          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_center}, "j");
 459          break;
 460       case TGSI_INTERPOLATE_LOC_CENTROID:
 461          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_centroid}, "i");
 462          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_centroid}, "j");
 463          break;
 464       case TGSI_INTERPOLATE_LOC_SAMPLE:
 465          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_sample}, "i");
 466          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_sample}, "j");
 467          break;
 468       }
 469
 470       // load/compute w
 471       Value *vw = nullptr, *pAttribs;
 472       if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE ||
 473           interpMode == TGSI_INTERPOLATE_COLOR) {
 474          pAttribs = pPerspAttribs;
 475          switch (interpLoc) {
 476          case TGSI_INTERPOLATE_LOC_CENTER:
 477             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}));
 478             break;
 479          case TGSI_INTERPOLATE_LOC_CENTROID:
 480             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_centroid}));
 481             break;
 482          case TGSI_INTERPOLATE_LOC_SAMPLE:
 483             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_sample}));
 484             break;
 485          }
 486       } else {
 487          pAttribs = pRawAttribs;
 488          vw = VIMMED1(1.f);
 489       }
 490
 491       vw->setName("w");
 492
 493       ubyte semantic_name = swr_fs->info.base.input_semantic_name[attrib];
 494       ubyte semantic_idx = swr_fs->info.base.input_semantic_index[attrib];
 495
 496       if (semantic_name == TGSI_SEMANTIC_FACE) {
 497          Value *ff =
 498             UI_TO_FP(LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), mFP32Ty);
 499          ff = FSUB(FMUL(ff, C(2.0f)), C(1.0f));
 500          ff = VECTOR_SPLAT(JM()->mVWidth, ff, "vFrontFace");
 501
 502          inputs[attrib][0] = wrap(ff);
 503          inputs[attrib][1] = wrap(VIMMED1(0.0f));
 504          inputs[attrib][2] = wrap(VIMMED1(0.0f));
 505          inputs[attrib][3] = wrap(VIMMED1(1.0f));
 506          continue;
 507       } else if (semantic_name == TGSI_SEMANTIC_POSITION) { // gl_FragCoord
 508          if (swr_fs->info.base.properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER] ==
 509              TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER) {
 510             inputs[attrib][0] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_center}, "vX"));
 511             inputs[attrib][1] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_center}, "vY"));
 512          } else {
 513             inputs[attrib][0] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_UL}, "vX"));
 514             inputs[attrib][1] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_UL}, "vY"));
 515          }
 516          inputs[attrib][2] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vZ}, "vZ"));
 517          inputs[attrib][3] =
 518             wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}, "vOneOverW"));
 519          continue;
 520       } else if (semantic_name == TGSI_SEMANTIC_PRIMID) {
 521          Value *primID = LOAD(pPS, {0, SWR_PS_CONTEXT_primID}, "primID");
 522          inputs[attrib][0] = wrap(VECTOR_SPLAT(JM()->mVWidth, primID));
 523          inputs[attrib][1] = wrap(VIMMED1(0));
 524          inputs[attrib][2] = wrap(VIMMED1(0));
 525          inputs[attrib][3] = wrap(VIMMED1(0));
 526          continue;
 527       }
 528
 529       unsigned linkedAttrib =
 530          locate_linkage(semantic_name, semantic_idx, &ctx->vs->info.base);
 531       if (semantic_name == TGSI_SEMANTIC_GENERIC &&
 532           key.sprite_coord_enable & (1 << semantic_idx)) {
 533          /* we add an extra attrib to the backendState in swr_update_derived. */
 534          linkedAttrib = ctx->vs->info.base.num_outputs - 1;
 535          swr_fs->pointSpriteMask |= (1 << linkedAttrib);
 536       } else if (linkedAttrib == 0xFFFFFFFF) {
 537          inputs[attrib][0] = wrap(VIMMED1(0.0f));
 538          inputs[attrib][1] = wrap(VIMMED1(0.0f));
 539          inputs[attrib][2] = wrap(VIMMED1(0.0f));
 540          inputs[attrib][3] = wrap(VIMMED1(1.0f));
 541          /* If we're reading in color and 2-sided lighting is enabled, we have
 542           * to keep going.
 543           */
 544          if (semantic_name != TGSI_SEMANTIC_COLOR || !key.light_twoside)
 545             continue;
 546       } else {
 547          if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
 548             swr_fs->constantMask |= 1 << linkedAttrib;
 549          } else if (interpMode == TGSI_INTERPOLATE_COLOR) {
 550             swr_fs->flatConstantMask |= 1 << linkedAttrib;
 551          }
 552       }
 553
 554       unsigned bcolorAttrib = 0xFFFFFFFF;
 555       Value *offset = NULL;
 556       if (semantic_name == TGSI_SEMANTIC_COLOR && key.light_twoside) {
 557          bcolorAttrib = locate_linkage(
 558                TGSI_SEMANTIC_BCOLOR, semantic_idx, &ctx->vs->info.base);
 559          /* Neither front nor back colors were available. Nothing to load. */
 560          if (bcolorAttrib == 0xFFFFFFFF && linkedAttrib == 0xFFFFFFFF)
 561             continue;
 562          /* If there is no front color, just always use the back color. */
 563          if (linkedAttrib == 0xFFFFFFFF)
 564             linkedAttrib = bcolorAttrib;
 565
 566          if (bcolorAttrib != 0xFFFFFFFF) {
 567             if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
 568                swr_fs->constantMask |= 1 << bcolorAttrib;
 569             } else if (interpMode == TGSI_INTERPOLATE_COLOR) {
 570                swr_fs->flatConstantMask |= 1 << bcolorAttrib;
 571             }
 572
 573             unsigned diff = 12 * (bcolorAttrib - linkedAttrib);
 574
 575             if (diff) {
 576                Value *back =
 577                   XOR(C(1), LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), "backFace");
 578
 579                offset = MUL(back, C(diff));
 580                offset->setName("offset");
 581             }
 582          }
 583       }
 584
 585       for (int channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 586          if (mask & (1 << channel)) {
 587             Value *indexA = C(linkedAttrib * 12 + channel);
 588             Value *indexB = C(linkedAttrib * 12 + channel + 4);
 589             Value *indexC = C(linkedAttrib * 12 + channel + 8);
 590
 591             if (offset) {
 592                indexA = ADD(indexA, offset);
 593                indexB = ADD(indexB, offset);
 594                indexC = ADD(indexC, offset);
 595             }
 596
 597             Value *va = VBROADCAST(LOAD(GEP(pAttribs, indexA)));
 598             Value *vb = VBROADCAST(LOAD(GEP(pAttribs, indexB)));
 599             Value *vc = VBROADCAST(LOAD(GEP(pAttribs, indexC)));
 600
 601             if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
 602                inputs[attrib][channel] = wrap(va);
 603             } else {
 604                Value *vk = FSUB(FSUB(VIMMED1(1.0f), vi), vj);
 605
 606                vc = FMUL(vk, vc);
 607
 608                Value *interp = FMUL(va, vi);
 609                Value *interp1 = FMUL(vb, vj);
 610                interp = FADD(interp, interp1);
 611                interp = FADD(interp, vc);
 612                if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE ||
 613                    interpMode == TGSI_INTERPOLATE_COLOR)
 614                   interp = FMUL(interp, vw);
 615                inputs[attrib][channel] = wrap(interp);
 616             }
 617          }
 618       }
 619    }
 620
 621    sampler = swr_sampler_soa_create(key.sampler, PIPE_SHADER_FRAGMENT);
 622
 623    struct lp_bld_tgsi_system_values system_values;
 624    memset(&system_values, 0, sizeof(system_values));
 625
 626    struct lp_build_mask_context mask;
 627
 628    if (swr_fs->info.base.uses_kill) {
 629       Value *mask_val = LOAD(pPS, {0, SWR_PS_CONTEXT_activeMask}, "activeMask");
 630       lp_build_mask_begin(
 631          &mask, gallivm, lp_type_float_vec(32, 32 * 8), wrap(mask_val));
 632    }
 633
 634    lp_build_tgsi_soa(gallivm,
 635                      swr_fs->pipe.tokens,
 636                      lp_type_float_vec(32, 32 * 8),
 637                      swr_fs->info.base.uses_kill ? &mask : NULL, // mask
 638                      wrap(consts_ptr),
 639                      wrap(const_sizes_ptr),
 640                      &system_values,
 641                      inputs,
 642                      outputs,
 643                      wrap(hPrivateData),
 644                      NULL, // thread data
 645                      sampler, // sampler
 646                      &swr_fs->info.base,
 647                      NULL); // geometry shader face
 648
 649    sampler->destroy(sampler);
 650
 651    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 652
 653    for (uint32_t attrib = 0; attrib < swr_fs->info.base.num_outputs;
 654         attrib++) {
 655       switch (swr_fs->info.base.output_semantic_name[attrib]) {
 656       case TGSI_SEMANTIC_POSITION: {
 657          // write z
 658          LLVMValueRef outZ =
 659             LLVMBuildLoad(gallivm->builder, outputs[attrib][2], "");
 660          STORE(unwrap(outZ), pPS, {0, SWR_PS_CONTEXT_vZ});
 661          break;
 662       }
 663       case TGSI_SEMANTIC_COLOR: {
 664          for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 665             if (!outputs[attrib][channel])
 666                continue;
 667
 668             LLVMValueRef out =
 669                LLVMBuildLoad(gallivm->builder, outputs[attrib][channel], "");
 670             if (swr_fs->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS] &&
 671                 swr_fs->info.base.output_semantic_index[attrib] == 0) {
 672                for (uint32_t rt = 0; rt < key.nr_cbufs; rt++) {
 673                   STORE(unwrap(out),
 674                         pPS,
 675                         {0, SWR_PS_CONTEXT_shaded, rt, channel});
 676                }
 677             } else {
 678                STORE(unwrap(out),
 679                      pPS,
 680                      {0,
 681                            SWR_PS_CONTEXT_shaded,
 682                            swr_fs->info.base.output_semantic_index[attrib],
 683                            channel});
 684             }
 685          }
 686          break;
 687       }
 688       default: {
 689          fprintf(stderr,
 690                  "unknown output from FS %s[%d]\n",
 691                  tgsi_semantic_names[swr_fs->info.base
 692                                         .output_semantic_name[attrib]],
 693                  swr_fs->info.base.output_semantic_index[attrib]);
 694          break;
 695       }
 696       }
 697    }
 698
 699    LLVMValueRef mask_result = 0;
 700    if (swr_fs->info.base.uses_kill) {
 701       mask_result = lp_build_mask_end(&mask);
 702    }
 703
 704    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 705
 706    if (swr_fs->info.base.uses_kill) {
 707       STORE(unwrap(mask_result), pPS, {0, SWR_PS_CONTEXT_activeMask});
 708    }
 709
 710    RET_VOID();
 711
 712    gallivm_verify_function(gallivm, wrap(pFunction));
 713
 714    gallivm_compile_module(gallivm);
 715
 716    PFN_PIXEL_KERNEL kernel =
 717       (PFN_PIXEL_KERNEL)gallivm_jit_function(gallivm, wrap(pFunction));
 718    debug_printf("frag shader  %p\n", kernel);
 719    assert(kernel && "Error: FragShader = NULL");
 720
 721    JM()->mIsModuleFinalized = true;
 722
 723    return kernel;
 724 }
 725
 726 PFN_PIXEL_KERNEL
 727 swr_compile_fs(struct swr_context *ctx, swr_jit_fs_key &key)
 728 {
 729    BuilderSWR builder(
 730       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
 731       "FS");
 732    PFN_PIXEL_KERNEL func = builder.CompileFS(ctx, key);
 733
 734    ctx->fs->map.insert(std::make_pair(key, make_unique<VariantFS>(builder.gallivm, func)));
 735    return func;
 736 }