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