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 "gen_state_llvm.h"
  34 #include "builder.h"
  35
  36 #include "tgsi/tgsi_strings.h"
  37 #include "util/u_format.h"
  38 #include "util/u_prim.h"
  39 #include "gallivm/lp_bld_init.h"
  40 #include "gallivm/lp_bld_flow.h"
  41 #include "gallivm/lp_bld_struct.h"
  42 #include "gallivm/lp_bld_tgsi.h"
  43
  44 #include "swr_context.h"
  45 #include "gen_swr_context_llvm.h"
  46 #include "swr_resource.h"
  47 #include "swr_state.h"
  48 #include "swr_screen.h"
  49
  50 using namespace SwrJit;
  51 using namespace llvm;
  52
  53 static unsigned
  54 locate_linkage(ubyte name, ubyte index, struct tgsi_shader_info *info);
  55
  56 bool operator==(const swr_jit_fs_key &lhs, const swr_jit_fs_key &rhs)
  57 {
  58    return !memcmp(&lhs, &rhs, sizeof(lhs));
  59 }
  60
  61 bool operator==(const swr_jit_vs_key &lhs, const swr_jit_vs_key &rhs)
  62 {
  63    return !memcmp(&lhs, &rhs, sizeof(lhs));
  64 }
  65
  66 bool operator==(const swr_jit_fetch_key &lhs, const swr_jit_fetch_key &rhs)
  67 {
  68    return !memcmp(&lhs, &rhs, sizeof(lhs));
  69 }
  70
  71 bool operator==(const swr_jit_gs_key &lhs, const swr_jit_gs_key &rhs)
  72 {
  73    return !memcmp(&lhs, &rhs, sizeof(lhs));
  74 }
  75
  76 static void
  77 swr_generate_sampler_key(const struct lp_tgsi_info &info,
  78                          struct swr_context *ctx,
  79                          enum pipe_shader_type shader_type,
  80                          struct swr_jit_sampler_key &key)
  81 {
  82    key.nr_samplers = info.base.file_max[TGSI_FILE_SAMPLER] + 1;
  83
  84    for (unsigned i = 0; i < key.nr_samplers; i++) {
  85       if (info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
  86          lp_sampler_static_sampler_state(
  87             &key.sampler[i].sampler_state,
  88             ctx->samplers[shader_type][i]);
  89       }
  90    }
  91
  92    /*
  93     * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
  94     * are dx10-style? Can't really have mixed opcodes, at least not
  95     * if we want to skip the holes here (without rescanning tgsi).
  96     */
  97    if (info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
  98       key.nr_sampler_views =
  99          info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
 100       for (unsigned i = 0; i < key.nr_sampler_views; i++) {
 101          if (info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1 << i)) {
 102             const struct pipe_sampler_view *view =
 103                ctx->sampler_views[shader_type][i];
 104             lp_sampler_static_texture_state(
 105                &key.sampler[i].texture_state, view);
 106             if (view) {
 107                struct swr_resource *swr_res = swr_resource(view->texture);
 108                const struct util_format_description *desc =
 109                   util_format_description(view->format);
 110                if (swr_res->has_depth && swr_res->has_stencil &&
 111                    !util_format_has_depth(desc))
 112                   key.sampler[i].texture_state.format = PIPE_FORMAT_S8_UINT;
 113             }
 114          }
 115       }
 116    } else {
 117       key.nr_sampler_views = key.nr_samplers;
 118       for (unsigned i = 0; i < key.nr_sampler_views; i++) {
 119          if (info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
 120             const struct pipe_sampler_view *view =
 121                ctx->sampler_views[shader_type][i];
 122             lp_sampler_static_texture_state(
 123                &key.sampler[i].texture_state, view);
 124             if (view) {
 125                struct swr_resource *swr_res = swr_resource(view->texture);
 126                const struct util_format_description *desc =
 127                   util_format_description(view->format);
 128                if (swr_res->has_depth && swr_res->has_stencil &&
 129                    !util_format_has_depth(desc))
 130                   key.sampler[i].texture_state.format = PIPE_FORMAT_S8_UINT;
 131             }
 132          }
 133       }
 134    }
 135 }
 136
 137 void
 138 swr_generate_fs_key(struct swr_jit_fs_key &key,
 139                     struct swr_context *ctx,
 140                     swr_fragment_shader *swr_fs)
 141 {
 142    memset(&key, 0, sizeof(key));
 143
 144    key.nr_cbufs = ctx->framebuffer.nr_cbufs;
 145    key.light_twoside = ctx->rasterizer->light_twoside;
 146    key.sprite_coord_enable = ctx->rasterizer->sprite_coord_enable;
 147
 148    struct tgsi_shader_info *pPrevShader;
 149    if (ctx->gs)
 150       pPrevShader = &ctx->gs->info.base;
 151    else
 152       pPrevShader = &ctx->vs->info.base;
 153
 154    memcpy(&key.vs_output_semantic_name,
 155           &pPrevShader->output_semantic_name,
 156           sizeof(key.vs_output_semantic_name));
 157    memcpy(&key.vs_output_semantic_idx,
 158           &pPrevShader->output_semantic_index,
 159           sizeof(key.vs_output_semantic_idx));
 160
 161    swr_generate_sampler_key(swr_fs->info, ctx, PIPE_SHADER_FRAGMENT, key);
 162
 163    key.poly_stipple_enable = ctx->rasterizer->poly_stipple_enable &&
 164       ctx->poly_stipple.prim_is_poly;
 165 }
 166
 167 void
 168 swr_generate_vs_key(struct swr_jit_vs_key &key,
 169                     struct swr_context *ctx,
 170                     swr_vertex_shader *swr_vs)
 171 {
 172    memset(&key, 0, sizeof(key));
 173
 174    key.clip_plane_mask =
 175       swr_vs->info.base.clipdist_writemask ?
 176       swr_vs->info.base.clipdist_writemask & ctx->rasterizer->clip_plane_enable :
 177       ctx->rasterizer->clip_plane_enable;
 178
 179    swr_generate_sampler_key(swr_vs->info, ctx, PIPE_SHADER_VERTEX, key);
 180 }
 181
 182 void
 183 swr_generate_fetch_key(struct swr_jit_fetch_key &key,
 184                        struct swr_vertex_element_state *velems)
 185 {
 186    memset(&key, 0, sizeof(key));
 187
 188    key.fsState = velems->fsState;
 189 }
 190
 191 void
 192 swr_generate_gs_key(struct swr_jit_gs_key &key,
 193                     struct swr_context *ctx,
 194                     swr_geometry_shader *swr_gs)
 195 {
 196    memset(&key, 0, sizeof(key));
 197
 198    struct tgsi_shader_info *pPrevShader = &ctx->vs->info.base;
 199
 200    memcpy(&key.vs_output_semantic_name,
 201           &pPrevShader->output_semantic_name,
 202           sizeof(key.vs_output_semantic_name));
 203    memcpy(&key.vs_output_semantic_idx,
 204           &pPrevShader->output_semantic_index,
 205           sizeof(key.vs_output_semantic_idx));
 206
 207    swr_generate_sampler_key(swr_gs->info, ctx, PIPE_SHADER_GEOMETRY, key);
 208 }
 209
 210 struct BuilderSWR : public Builder {
 211    BuilderSWR(JitManager *pJitMgr, const char *pName)
 212       : Builder(pJitMgr)
 213    {
 214       pJitMgr->SetupNewModule();
 215       gallivm = gallivm_create(pName, wrap(&JM()->mContext));
 216       pJitMgr->mpCurrentModule = unwrap(gallivm->module);
 217    }
 218
 219    ~BuilderSWR() {
 220       gallivm_free_ir(gallivm);
 221    }
 222
 223    void WriteVS(Value *pVal, Value *pVsContext, Value *pVtxOutput,
 224                 unsigned slot, unsigned channel);
 225
 226    struct gallivm_state *gallivm;
 227    PFN_VERTEX_FUNC CompileVS(struct swr_context *ctx, swr_jit_vs_key &key);
 228    PFN_PIXEL_KERNEL CompileFS(struct swr_context *ctx, swr_jit_fs_key &key);
 229    PFN_GS_FUNC CompileGS(struct swr_context *ctx, swr_jit_gs_key &key);
 230
 231    LLVMValueRef
 232    swr_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
 233                            struct lp_build_tgsi_context * bld_base,
 234                            boolean is_vindex_indirect,
 235                            LLVMValueRef vertex_index,
 236                            boolean is_aindex_indirect,
 237                            LLVMValueRef attrib_index,
 238                            LLVMValueRef swizzle_index);
 239    void
 240    swr_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
 241                            struct lp_build_tgsi_context * bld_base,
 242                            LLVMValueRef (*outputs)[4],
 243                            LLVMValueRef emitted_vertices_vec);
 244
 245    void
 246    swr_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
 247                              struct lp_build_tgsi_context * bld_base,
 248                              LLVMValueRef verts_per_prim_vec,
 249                              LLVMValueRef emitted_prims_vec);
 250
 251    void
 252    swr_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
 253                         struct lp_build_tgsi_context * bld_base,
 254                         LLVMValueRef total_emitted_vertices_vec,
 255                         LLVMValueRef emitted_prims_vec);
 256
 257 };
 258
 259 struct swr_gs_llvm_iface {
 260    struct lp_build_tgsi_gs_iface base;
 261    struct tgsi_shader_info *info;
 262
 263    BuilderSWR *pBuilder;
 264
 265    Value *pGsCtx;
 266    SWR_GS_STATE *pGsState;
 267    uint32_t num_outputs;
 268    uint32_t num_verts_per_prim;
 269
 270    Value *pVtxAttribMap;
 271 };
 272
 273 // trampoline functions so we can use the builder llvm construction methods
 274 static LLVMValueRef
 275 swr_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
 276                            struct lp_build_tgsi_context * bld_base,
 277                            boolean is_vindex_indirect,
 278                            LLVMValueRef vertex_index,
 279                            boolean is_aindex_indirect,
 280                            LLVMValueRef attrib_index,
 281                            LLVMValueRef swizzle_index)
 282 {
 283     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_iface;
 284
 285     return iface->pBuilder->swr_gs_llvm_fetch_input(gs_iface, bld_base,
 286                                                    is_vindex_indirect,
 287                                                    vertex_index,
 288                                                    is_aindex_indirect,
 289                                                    attrib_index,
 290                                                    swizzle_index);
 291 }
 292
 293 static void
 294 swr_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
 295                            struct lp_build_tgsi_context * bld_base,
 296                            LLVMValueRef (*outputs)[4],
 297                            LLVMValueRef emitted_vertices_vec)
 298 {
 299     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 300
 301     iface->pBuilder->swr_gs_llvm_emit_vertex(gs_base, bld_base,
 302                                             outputs,
 303                                             emitted_vertices_vec);
 304 }
 305
 306 static void
 307 swr_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
 308                              struct lp_build_tgsi_context * bld_base,
 309                              LLVMValueRef verts_per_prim_vec,
 310                              LLVMValueRef emitted_prims_vec)
 311 {
 312     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 313
 314     iface->pBuilder->swr_gs_llvm_end_primitive(gs_base, bld_base,
 315                                               verts_per_prim_vec,
 316                                               emitted_prims_vec);
 317 }
 318
 319 static void
 320 swr_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
 321                         struct lp_build_tgsi_context * bld_base,
 322                         LLVMValueRef total_emitted_vertices_vec,
 323                         LLVMValueRef emitted_prims_vec)
 324 {
 325     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 326
 327     iface->pBuilder->swr_gs_llvm_epilogue(gs_base, bld_base,
 328                                          total_emitted_vertices_vec,
 329                                          emitted_prims_vec);
 330 }
 331
 332 LLVMValueRef
 333 BuilderSWR::swr_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
 334                            struct lp_build_tgsi_context * bld_base,
 335                            boolean is_vindex_indirect,
 336                            LLVMValueRef vertex_index,
 337                            boolean is_aindex_indirect,
 338                            LLVMValueRef attrib_index,
 339                            LLVMValueRef swizzle_index)
 340 {
 341     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_iface;
 342
 343     IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 344
 345     assert(is_vindex_indirect == false && is_aindex_indirect == false);
 346
 347     Value *attrib =
 348        LOAD(GEP(iface->pVtxAttribMap, {C(0), unwrap(attrib_index)}));
 349
 350     Value *pInput =
 351        LOAD(GEP(iface->pGsCtx,
 352                 {C(0),
 353                  C(SWR_GS_CONTEXT_vert),
 354                  unwrap(vertex_index),
 355                  C(0),
 356                  attrib,
 357                  unwrap(swizzle_index)}));
 358
 359     return wrap(pInput);
 360 }
 361
 362 void
 363 BuilderSWR::swr_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
 364                            struct lp_build_tgsi_context * bld_base,
 365                            LLVMValueRef (*outputs)[4],
 366                            LLVMValueRef emitted_vertices_vec)
 367 {
 368     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 369     SWR_GS_STATE *pGS = iface->pGsState;
 370
 371     IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 372
 373 #if USE_SIMD16_FRONTEND
 374     const uint32_t simdVertexStride = sizeof(simdvertex) * 2;
 375     const uint32_t numSimdBatches = (pGS->maxNumVerts + (mVWidth * 2) - 1) / (mVWidth * 2);
 376 #else
 377     const uint32_t simdVertexStride = sizeof(simdvertex);
 378     const uint32_t numSimdBatches = (pGS->maxNumVerts + mVWidth - 1) / mVWidth;
 379 #endif
 380     const uint32_t inputPrimStride = numSimdBatches * simdVertexStride;
 381
 382     Value *pStream = LOAD(iface->pGsCtx, { 0, SWR_GS_CONTEXT_pStream });
 383     Value *vMask = LOAD(iface->pGsCtx, { 0, SWR_GS_CONTEXT_mask });
 384     Value *vMask1 = TRUNC(vMask, VectorType::get(mInt1Ty, 8));
 385
 386     Value *vOffsets = C({
 387           inputPrimStride * 0,
 388           inputPrimStride * 1,
 389           inputPrimStride * 2,
 390           inputPrimStride * 3,
 391           inputPrimStride * 4,
 392           inputPrimStride * 5,
 393           inputPrimStride * 6,
 394           inputPrimStride * 7 } );
 395
 396 #if USE_SIMD16_FRONTEND
 397     const uint32_t simdShift = log2(mVWidth * 2);
 398     Value *vSimdSlot = AND(unwrap(emitted_vertices_vec), (mVWidth * 2) - 1);
 399 #else
 400     const uint32_t simdShift = log2(mVWidth);
 401     Value *vSimdSlot = AND(unwrap(emitted_vertices_vec), mVWidth - 1);
 402 #endif
 403     Value *vVertexSlot = ASHR(unwrap(emitted_vertices_vec), simdShift);
 404
 405     for (uint32_t attrib = 0; attrib < iface->num_outputs; ++attrib) {
 406        uint32_t attribSlot = attrib;
 407        if (iface->info->output_semantic_name[attrib] == TGSI_SEMANTIC_PSIZE)
 408           attribSlot = VERTEX_POINT_SIZE_SLOT;
 409        else if (iface->info->output_semantic_name[attrib] == TGSI_SEMANTIC_PRIMID)
 410           attribSlot = VERTEX_PRIMID_SLOT;
 411        else if (iface->info->output_semantic_name[attrib] == TGSI_SEMANTIC_LAYER)
 412           attribSlot = VERTEX_RTAI_SLOT;
 413
 414 #if USE_SIMD16_FRONTEND
 415        Value *vOffsetsAttrib =
 416           ADD(vOffsets, MUL(vVertexSlot, VIMMED1((uint32_t)sizeof(simdvertex) * 2)));
 417        vOffsetsAttrib =
 418           ADD(vOffsetsAttrib, VIMMED1((uint32_t)(attribSlot*sizeof(simdvector) * 2)));
 419 #else
 420        Value *vOffsetsAttrib =
 421           ADD(vOffsets, MUL(vVertexSlot, VIMMED1((uint32_t)sizeof(simdvertex))));
 422        vOffsetsAttrib =
 423           ADD(vOffsetsAttrib, VIMMED1((uint32_t)(attribSlot*sizeof(simdvector))));
 424 #endif
 425        vOffsetsAttrib =
 426           ADD(vOffsetsAttrib, MUL(vSimdSlot, VIMMED1((uint32_t)sizeof(float))));
 427
 428        for (uint32_t channel = 0; channel < 4; ++channel) {
 429           Value *vData = LOAD(unwrap(outputs[attrib][channel]));
 430           Value *vPtrs = GEP(pStream, vOffsetsAttrib);
 431
 432           vPtrs = BITCAST(vPtrs,
 433                           VectorType::get(PointerType::get(mFP32Ty, 0), 8));
 434
 435           MASKED_SCATTER(vData, vPtrs, 32, vMask1);
 436
 437 #if USE_SIMD16_FRONTEND
 438           vOffsetsAttrib =
 439              ADD(vOffsetsAttrib, VIMMED1((uint32_t)sizeof(simdscalar) * 2));
 440 #else
 441           vOffsetsAttrib =
 442              ADD(vOffsetsAttrib, VIMMED1((uint32_t)sizeof(simdscalar)));
 443 #endif
 444        }
 445     }
 446 }
 447
 448 void
 449 BuilderSWR::swr_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
 450                              struct lp_build_tgsi_context * bld_base,
 451                              LLVMValueRef verts_per_prim_vec,
 452                              LLVMValueRef emitted_prims_vec)
 453 {
 454     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 455     SWR_GS_STATE *pGS = iface->pGsState;
 456
 457     IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 458
 459     Value *pCutBuffer =
 460        LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_pCutOrStreamIdBuffer});
 461     Value *vMask = LOAD(iface->pGsCtx, { 0, SWR_GS_CONTEXT_mask });
 462     Value *vMask1 = TRUNC(vMask, VectorType::get(mInt1Ty, 8));
 463
 464     uint32_t vertsPerPrim = iface->num_verts_per_prim;
 465
 466     Value *vCount =
 467        ADD(MUL(unwrap(emitted_prims_vec), VIMMED1(vertsPerPrim)),
 468            unwrap(verts_per_prim_vec));
 469
 470     struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
 471     vCount = LOAD(unwrap(bld->total_emitted_vertices_vec_ptr));
 472
 473     struct lp_exec_mask *exec_mask = &bld->exec_mask;
 474     Value *mask = unwrap(lp_build_mask_value(bld->mask));
 475     if (exec_mask->has_mask)
 476        mask = AND(mask, unwrap(exec_mask->exec_mask));
 477
 478     Value *cmpMask = VMASK(ICMP_NE(unwrap(verts_per_prim_vec), VIMMED1(0)));
 479     mask = AND(mask, cmpMask);
 480     vMask1 = TRUNC(mask, VectorType::get(mInt1Ty, 8));
 481
 482     const uint32_t cutPrimStride =
 483        (pGS->maxNumVerts + JM()->mVWidth - 1) / JM()->mVWidth;
 484     Value *vOffsets = C({
 485           (uint32_t)(cutPrimStride * 0),
 486           (uint32_t)(cutPrimStride * 1),
 487           (uint32_t)(cutPrimStride * 2),
 488           (uint32_t)(cutPrimStride * 3),
 489           (uint32_t)(cutPrimStride * 4),
 490           (uint32_t)(cutPrimStride * 5),
 491           (uint32_t)(cutPrimStride * 6),
 492           (uint32_t)(cutPrimStride * 7) } );
 493
 494     vCount = SUB(vCount, VIMMED1(1));
 495     Value *vOffset = ADD(UDIV(vCount, VIMMED1(8)), vOffsets);
 496     Value *vValue = SHL(VIMMED1(1), UREM(vCount, VIMMED1(8)));
 497
 498     vValue = TRUNC(vValue, VectorType::get(mInt8Ty, 8));
 499
 500     Value *vPtrs = GEP(pCutBuffer, vOffset);
 501     vPtrs =
 502        BITCAST(vPtrs, VectorType::get(PointerType::get(mInt8Ty, 0), JM()->mVWidth));
 503
 504     Value *vGather = MASKED_GATHER(vPtrs, 32, vMask1);
 505     vValue = OR(vGather, vValue);
 506     MASKED_SCATTER(vValue, vPtrs, 32, vMask1);
 507 }
 508
 509 void
 510 BuilderSWR::swr_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
 511                         struct lp_build_tgsi_context * bld_base,
 512                         LLVMValueRef total_emitted_vertices_vec,
 513                         LLVMValueRef emitted_prims_vec)
 514 {
 515    swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 516
 517    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 518
 519    STORE(unwrap(total_emitted_vertices_vec), iface->pGsCtx, {0, SWR_GS_CONTEXT_vertexCount});
 520 }
 521
 522 PFN_GS_FUNC
 523 BuilderSWR::CompileGS(struct swr_context *ctx, swr_jit_gs_key &key)
 524 {
 525    SWR_GS_STATE *pGS = &ctx->gs->gsState;
 526    struct tgsi_shader_info *info = &ctx->gs->info.base;
 527
 528    pGS->gsEnable = true;
 529
 530    pGS->numInputAttribs = info->num_inputs;
 531    pGS->outputTopology =
 532       swr_convert_prim_topology(info->properties[TGSI_PROPERTY_GS_OUTPUT_PRIM]);
 533    pGS->maxNumVerts = info->properties[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES];
 534    pGS->instanceCount = info->properties[TGSI_PROPERTY_GS_INVOCATIONS];
 535
 536    pGS->emitsRenderTargetArrayIndex = info->writes_layer;
 537    pGS->emitsPrimitiveID = info->writes_primid;
 538    pGS->emitsViewportArrayIndex = info->writes_viewport_index;
 539
 540    // XXX: single stream for now...
 541    pGS->isSingleStream = true;
 542    pGS->singleStreamID = 0;
 543
 544    struct swr_geometry_shader *gs = ctx->gs;
 545
 546    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 547    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 548
 549    memset(outputs, 0, sizeof(outputs));
 550
 551    AttrBuilder attrBuilder;
 552    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 553
 554    std::vector<Type *> gsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 555                               PointerType::get(Gen_SWR_GS_CONTEXT(JM()), 0)};
 556    FunctionType *vsFuncType =
 557       FunctionType::get(Type::getVoidTy(JM()->mContext), gsArgs, false);
 558
 559    // create new vertex shader function
 560    auto pFunction = Function::Create(vsFuncType,
 561                                      GlobalValue::ExternalLinkage,
 562                                      "GS",
 563                                      JM()->mpCurrentModule);
 564 #if HAVE_LLVM < 0x0500
 565    AttributeSet attrSet = AttributeSet::get(
 566       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 567    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 568 #else
 569    pFunction->addAttributes(AttributeList::FunctionIndex, attrBuilder);
 570 #endif
 571
 572    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 573    IRB()->SetInsertPoint(block);
 574    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 575
 576    auto argitr = pFunction->arg_begin();
 577    Value *hPrivateData = &*argitr++;
 578    hPrivateData->setName("hPrivateData");
 579    Value *pGsCtx = &*argitr++;
 580    pGsCtx->setName("gsCtx");
 581
 582    Value *consts_ptr =
 583       GEP(hPrivateData, {C(0), C(swr_draw_context_constantGS)});
 584    consts_ptr->setName("gs_constants");
 585    Value *const_sizes_ptr =
 586       GEP(hPrivateData, {0, swr_draw_context_num_constantsGS});
 587    const_sizes_ptr->setName("num_gs_constants");
 588
 589    struct lp_build_sampler_soa *sampler =
 590       swr_sampler_soa_create(key.sampler, PIPE_SHADER_GEOMETRY);
 591
 592    struct lp_bld_tgsi_system_values system_values;
 593    memset(&system_values, 0, sizeof(system_values));
 594    system_values.prim_id = wrap(LOAD(pGsCtx, {0, SWR_GS_CONTEXT_PrimitiveID}));
 595    system_values.instance_id = wrap(LOAD(pGsCtx, {0, SWR_GS_CONTEXT_InstanceID}));
 596
 597    std::vector<Constant*> mapConstants;
 598    Value *vtxAttribMap = ALLOCA(ArrayType::get(mInt32Ty, PIPE_MAX_SHADER_INPUTS));
 599    for (unsigned slot = 0; slot < info->num_inputs; slot++) {
 600       ubyte semantic_name = info->input_semantic_name[slot];
 601       ubyte semantic_idx = info->input_semantic_index[slot];
 602
 603       unsigned vs_slot =
 604          locate_linkage(semantic_name, semantic_idx, &ctx->vs->info.base) + 1;
 605
 606       STORE(C(vs_slot), vtxAttribMap, {0, slot});
 607       mapConstants.push_back(C(vs_slot));
 608    }
 609
 610    struct lp_build_mask_context mask;
 611    Value *mask_val = LOAD(pGsCtx, {0, SWR_GS_CONTEXT_mask}, "gsMask");
 612    lp_build_mask_begin(&mask, gallivm,
 613                        lp_type_float_vec(32, 32 * 8), wrap(mask_val));
 614
 615    // zero out cut buffer so we can load/modify/store bits
 616    MEMSET(LOAD(pGsCtx, {0, SWR_GS_CONTEXT_pCutOrStreamIdBuffer}),
 617           C((char)0),
 618           pGS->instanceCount * ((pGS->maxNumVerts + 7) / 8) * JM()->mVWidth,
 619           sizeof(float) * KNOB_SIMD_WIDTH);
 620
 621    struct swr_gs_llvm_iface gs_iface;
 622    gs_iface.base.fetch_input = ::swr_gs_llvm_fetch_input;
 623    gs_iface.base.emit_vertex = ::swr_gs_llvm_emit_vertex;
 624    gs_iface.base.end_primitive = ::swr_gs_llvm_end_primitive;
 625    gs_iface.base.gs_epilogue = ::swr_gs_llvm_epilogue;
 626    gs_iface.pBuilder = this;
 627    gs_iface.pGsCtx = pGsCtx;
 628    gs_iface.pGsState = pGS;
 629    gs_iface.num_outputs = gs->info.base.num_outputs;
 630    gs_iface.num_verts_per_prim =
 631       u_vertices_per_prim((pipe_prim_type)info->properties[TGSI_PROPERTY_GS_OUTPUT_PRIM]);
 632    gs_iface.info = info;
 633    gs_iface.pVtxAttribMap = vtxAttribMap;
 634
 635    lp_build_tgsi_soa(gallivm,
 636                      gs->pipe.tokens,
 637                      lp_type_float_vec(32, 32 * 8),
 638                      &mask,
 639                      wrap(consts_ptr),
 640                      wrap(const_sizes_ptr),
 641                      &system_values,
 642                      inputs,
 643                      outputs,
 644                      wrap(hPrivateData), // (sampler context)
 645                      NULL, // thread data
 646                      sampler,
 647                      &gs->info.base,
 648                      &gs_iface.base);
 649
 650    lp_build_mask_end(&mask);
 651
 652    sampler->destroy(sampler);
 653
 654    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 655
 656    RET_VOID();
 657
 658    gallivm_verify_function(gallivm, wrap(pFunction));
 659    gallivm_compile_module(gallivm);
 660
 661    PFN_GS_FUNC pFunc =
 662       (PFN_GS_FUNC)gallivm_jit_function(gallivm, wrap(pFunction));
 663
 664    debug_printf("geom shader  %p\n", pFunc);
 665    assert(pFunc && "Error: GeomShader = NULL");
 666
 667    JM()->mIsModuleFinalized = true;
 668
 669    return pFunc;
 670 }
 671
 672 PFN_GS_FUNC
 673 swr_compile_gs(struct swr_context *ctx, swr_jit_gs_key &key)
 674 {
 675    BuilderSWR builder(
 676       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
 677       "GS");
 678    PFN_GS_FUNC func = builder.CompileGS(ctx, key);
 679
 680    ctx->gs->map.insert(std::make_pair(key, make_unique<VariantGS>(builder.gallivm, func)));
 681    return func;
 682 }
 683
 684 void
 685 BuilderSWR::WriteVS(Value *pVal, Value *pVsContext, Value *pVtxOutput, unsigned slot, unsigned channel)
 686 {
 687 #if USE_SIMD16_FRONTEND
 688    // interleave the simdvertex components into the dest simd16vertex
 689    //   slot16offset = slot8offset * 2
 690    //   comp16offset = comp8offset * 2 + alternateOffset
 691
 692    Value *offset = LOAD(pVsContext, { 0, SWR_VS_CONTEXT_AlternateOffset });
 693    Value *pOut = GEP(pVtxOutput, { C(0), C(0), C(slot * 2), offset } );
 694    STORE(pVal, pOut, {channel * 2});
 695 #else
 696    Value *pOut = GEP(pVtxOutput, {0, 0, slot});
 697    STORE(pVal, pOut, {0, channel});
 698 #endif
 699 }
 700
 701 PFN_VERTEX_FUNC
 702 BuilderSWR::CompileVS(struct swr_context *ctx, swr_jit_vs_key &key)
 703 {
 704    struct swr_vertex_shader *swr_vs = ctx->vs;
 705
 706    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 707    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 708
 709    memset(outputs, 0, sizeof(outputs));
 710
 711    AttrBuilder attrBuilder;
 712    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 713
 714    std::vector<Type *> vsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 715                               PointerType::get(Gen_SWR_VS_CONTEXT(JM()), 0)};
 716    FunctionType *vsFuncType =
 717       FunctionType::get(Type::getVoidTy(JM()->mContext), vsArgs, false);
 718
 719    // create new vertex shader function
 720    auto pFunction = Function::Create(vsFuncType,
 721                                      GlobalValue::ExternalLinkage,
 722                                      "VS",
 723                                      JM()->mpCurrentModule);
 724 #if HAVE_LLVM < 0x0500
 725    AttributeSet attrSet = AttributeSet::get(
 726       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 727    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 728 #else
 729    pFunction->addAttributes(AttributeList::FunctionIndex, attrBuilder);
 730 #endif
 731
 732    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 733    IRB()->SetInsertPoint(block);
 734    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 735
 736    auto argitr = pFunction->arg_begin();
 737    Value *hPrivateData = &*argitr++;
 738    hPrivateData->setName("hPrivateData");
 739    Value *pVsCtx = &*argitr++;
 740    pVsCtx->setName("vsCtx");
 741
 742    Value *consts_ptr = GEP(hPrivateData, {C(0), C(swr_draw_context_constantVS)});
 743
 744    consts_ptr->setName("vs_constants");
 745    Value *const_sizes_ptr =
 746       GEP(hPrivateData, {0, swr_draw_context_num_constantsVS});
 747    const_sizes_ptr->setName("num_vs_constants");
 748
 749    Value *vtxInput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVin});
 750
 751    for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
 752       const unsigned mask = swr_vs->info.base.input_usage_mask[attrib];
 753       for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 754          if (mask & (1 << channel)) {
 755             inputs[attrib][channel] =
 756                wrap(LOAD(vtxInput, {0, 0, attrib, channel}));
 757          }
 758       }
 759    }
 760
 761    struct lp_build_sampler_soa *sampler =
 762       swr_sampler_soa_create(key.sampler, PIPE_SHADER_VERTEX);
 763
 764    struct lp_bld_tgsi_system_values system_values;
 765    memset(&system_values, 0, sizeof(system_values));
 766    system_values.instance_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_InstanceID}));
 767    system_values.vertex_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_VertexID}));
 768
 769    lp_build_tgsi_soa(gallivm,
 770                      swr_vs->pipe.tokens,
 771                      lp_type_float_vec(32, 32 * 8),
 772                      NULL, // mask
 773                      wrap(consts_ptr),
 774                      wrap(const_sizes_ptr),
 775                      &system_values,
 776                      inputs,
 777                      outputs,
 778                      wrap(hPrivateData), // (sampler context)
 779                      NULL, // thread data
 780                      sampler, // sampler
 781                      &swr_vs->info.base,
 782                      NULL); // geometry shader face
 783
 784    sampler->destroy(sampler);
 785
 786    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 787
 788    Value *vtxOutput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVout});
 789
 790    for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 791       for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_OUTPUTS; attrib++) {
 792          if (!outputs[attrib][channel])
 793             continue;
 794
 795          Value *val = LOAD(unwrap(outputs[attrib][channel]));
 796
 797          uint32_t outSlot = attrib;
 798          if (swr_vs->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_PSIZE)
 799             outSlot = VERTEX_POINT_SIZE_SLOT;
 800          WriteVS(val, pVsCtx, vtxOutput, outSlot, channel);
 801       }
 802    }
 803
 804    if (ctx->rasterizer->clip_plane_enable ||
 805        swr_vs->info.base.culldist_writemask) {
 806       unsigned clip_mask = ctx->rasterizer->clip_plane_enable;
 807
 808       unsigned cv = 0;
 809       if (swr_vs->info.base.writes_clipvertex) {
 810          cv = 1 + locate_linkage(TGSI_SEMANTIC_CLIPVERTEX, 0,
 811                                  &swr_vs->info.base);
 812       } else {
 813          for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 814             if (swr_vs->info.base.output_semantic_name[i] == TGSI_SEMANTIC_POSITION &&
 815                 swr_vs->info.base.output_semantic_index[i] == 0) {
 816                cv = i;
 817                break;
 818             }
 819          }
 820       }
 821       LLVMValueRef cx = LLVMBuildLoad(gallivm->builder, outputs[cv][0], "");
 822       LLVMValueRef cy = LLVMBuildLoad(gallivm->builder, outputs[cv][1], "");
 823       LLVMValueRef cz = LLVMBuildLoad(gallivm->builder, outputs[cv][2], "");
 824       LLVMValueRef cw = LLVMBuildLoad(gallivm->builder, outputs[cv][3], "");
 825
 826       for (unsigned val = 0; val < PIPE_MAX_CLIP_PLANES; val++) {
 827          // clip distance overrides user clip planes
 828          if ((swr_vs->info.base.clipdist_writemask & clip_mask & (1 << val)) ||
 829              ((swr_vs->info.base.culldist_writemask << swr_vs->info.base.num_written_clipdistance) & (1 << val))) {
 830             unsigned cv = 1 + locate_linkage(TGSI_SEMANTIC_CLIPDIST, val < 4 ? 0 : 1,
 831                                              &swr_vs->info.base);
 832             if (val < 4) {
 833                LLVMValueRef dist = LLVMBuildLoad(gallivm->builder, outputs[cv][val], "");
 834                WriteVS(unwrap(dist), pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_LO_SLOT, val);
 835             } else {
 836                LLVMValueRef dist = LLVMBuildLoad(gallivm->builder, outputs[cv][val - 4], "");
 837                WriteVS(unwrap(dist), pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_HI_SLOT, val - 4);
 838             }
 839             continue;
 840          }
 841
 842          if (!(clip_mask & (1 << val)))
 843             continue;
 844
 845          Value *px = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 0}));
 846          Value *py = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 1}));
 847          Value *pz = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 2}));
 848          Value *pw = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 3}));
 849          Value *dist = FADD(FMUL(unwrap(cx), VBROADCAST(px)),
 850                             FADD(FMUL(unwrap(cy), VBROADCAST(py)),
 851                                  FADD(FMUL(unwrap(cz), VBROADCAST(pz)),
 852                                       FMUL(unwrap(cw), VBROADCAST(pw)))));
 853
 854          if (val < 4)
 855             WriteVS(dist, pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_LO_SLOT, val);
 856          else
 857             WriteVS(dist, pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_HI_SLOT, val - 4);
 858       }
 859    }
 860
 861    RET_VOID();
 862
 863    gallivm_verify_function(gallivm, wrap(pFunction));
 864    gallivm_compile_module(gallivm);
 865
 866    //   lp_debug_dump_value(func);
 867
 868    PFN_VERTEX_FUNC pFunc =
 869       (PFN_VERTEX_FUNC)gallivm_jit_function(gallivm, wrap(pFunction));
 870
 871    debug_printf("vert shader  %p\n", pFunc);
 872    assert(pFunc && "Error: VertShader = NULL");
 873
 874    JM()->mIsModuleFinalized = true;
 875
 876    return pFunc;
 877 }
 878
 879 PFN_VERTEX_FUNC
 880 swr_compile_vs(struct swr_context *ctx, swr_jit_vs_key &key)
 881 {
 882    if (!ctx->vs->pipe.tokens)
 883       return NULL;
 884
 885    BuilderSWR builder(
 886       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
 887       "VS");
 888    PFN_VERTEX_FUNC func = builder.CompileVS(ctx, key);
 889
 890    ctx->vs->map.insert(std::make_pair(key, make_unique<VariantVS>(builder.gallivm, func)));
 891    return func;
 892 }
 893
 894 static unsigned
 895 locate_linkage(ubyte name, ubyte index, struct tgsi_shader_info *info)
 896 {
 897    for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 898       if ((info->output_semantic_name[i] == name)
 899           && (info->output_semantic_index[i] == index)) {
 900          return i - 1; // position is not part of the linkage
 901       }
 902    }
 903
 904    return 0xFFFFFFFF;
 905 }
 906
 907 PFN_PIXEL_KERNEL
 908 BuilderSWR::CompileFS(struct swr_context *ctx, swr_jit_fs_key &key)
 909 {
 910    struct swr_fragment_shader *swr_fs = ctx->fs;
 911
 912    struct tgsi_shader_info *pPrevShader;
 913    if (ctx->gs)
 914       pPrevShader = &ctx->gs->info.base;
 915    else
 916       pPrevShader = &ctx->vs->info.base;
 917
 918    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 919    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 920
 921    memset(inputs, 0, sizeof(inputs));
 922    memset(outputs, 0, sizeof(outputs));
 923
 924    struct lp_build_sampler_soa *sampler = NULL;
 925
 926    AttrBuilder attrBuilder;
 927    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 928
 929    std::vector<Type *> fsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 930                               PointerType::get(Gen_SWR_PS_CONTEXT(JM()), 0)};
 931    FunctionType *funcType =
 932       FunctionType::get(Type::getVoidTy(JM()->mContext), fsArgs, false);
 933
 934    auto pFunction = Function::Create(funcType,
 935                                      GlobalValue::ExternalLinkage,
 936                                      "FS",
 937                                      JM()->mpCurrentModule);
 938 #if HAVE_LLVM < 0x0500
 939    AttributeSet attrSet = AttributeSet::get(
 940       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 941    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 942 #else
 943    pFunction->addAttributes(AttributeList::FunctionIndex, attrBuilder);
 944 #endif
 945
 946    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 947    IRB()->SetInsertPoint(block);
 948    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 949
 950    auto args = pFunction->arg_begin();
 951    Value *hPrivateData = &*args++;
 952    hPrivateData->setName("hPrivateData");
 953    Value *pPS = &*args++;
 954    pPS->setName("psCtx");
 955
 956    Value *consts_ptr = GEP(hPrivateData, {0, swr_draw_context_constantFS});
 957    consts_ptr->setName("fs_constants");
 958    Value *const_sizes_ptr =
 959       GEP(hPrivateData, {0, swr_draw_context_num_constantsFS});
 960    const_sizes_ptr->setName("num_fs_constants");
 961
 962    // load *pAttribs, *pPerspAttribs
 963    Value *pRawAttribs = LOAD(pPS, {0, SWR_PS_CONTEXT_pAttribs}, "pRawAttribs");
 964    Value *pPerspAttribs =
 965       LOAD(pPS, {0, SWR_PS_CONTEXT_pPerspAttribs}, "pPerspAttribs");
 966
 967    swr_fs->constantMask = 0;
 968    swr_fs->flatConstantMask = 0;
 969    swr_fs->pointSpriteMask = 0;
 970
 971    for (int attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
 972       const unsigned mask = swr_fs->info.base.input_usage_mask[attrib];
 973       const unsigned interpMode = swr_fs->info.base.input_interpolate[attrib];
 974       const unsigned interpLoc = swr_fs->info.base.input_interpolate_loc[attrib];
 975
 976       if (!mask)
 977          continue;
 978
 979       // load i,j
 980       Value *vi = nullptr, *vj = nullptr;
 981       switch (interpLoc) {
 982       case TGSI_INTERPOLATE_LOC_CENTER:
 983          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_center}, "i");
 984          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_center}, "j");
 985          break;
 986       case TGSI_INTERPOLATE_LOC_CENTROID:
 987          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_centroid}, "i");
 988          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_centroid}, "j");
 989          break;
 990       case TGSI_INTERPOLATE_LOC_SAMPLE:
 991          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_sample}, "i");
 992          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_sample}, "j");
 993          break;
 994       }
 995
 996       // load/compute w
 997       Value *vw = nullptr, *pAttribs;
 998       if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE ||
 999           interpMode == TGSI_INTERPOLATE_COLOR) {
1000          pAttribs = pPerspAttribs;
1001          switch (interpLoc) {
1002          case TGSI_INTERPOLATE_LOC_CENTER:
1003             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}));
1004             break;
1005          case TGSI_INTERPOLATE_LOC_CENTROID:
1006             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_centroid}));
1007             break;
1008          case TGSI_INTERPOLATE_LOC_SAMPLE:
1009             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_sample}));
1010             break;
1011          }
1012       } else {
1013          pAttribs = pRawAttribs;
1014          vw = VIMMED1(1.f);
1015       }
1016
1017       vw->setName("w");
1018
1019       ubyte semantic_name = swr_fs->info.base.input_semantic_name[attrib];
1020       ubyte semantic_idx = swr_fs->info.base.input_semantic_index[attrib];
1021
1022       if (semantic_name == TGSI_SEMANTIC_FACE) {
1023          Value *ff =
1024             UI_TO_FP(LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), mFP32Ty);
1025          ff = FSUB(FMUL(ff, C(2.0f)), C(1.0f));
1026          ff = VECTOR_SPLAT(JM()->mVWidth, ff, "vFrontFace");
1027
1028          inputs[attrib][0] = wrap(ff);
1029          inputs[attrib][1] = wrap(VIMMED1(0.0f));
1030          inputs[attrib][2] = wrap(VIMMED1(0.0f));
1031          inputs[attrib][3] = wrap(VIMMED1(1.0f));
1032          continue;
1033       } else if (semantic_name == TGSI_SEMANTIC_POSITION) { // gl_FragCoord
1034          if (swr_fs->info.base.properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER] ==
1035              TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER) {
1036             inputs[attrib][0] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_center}, "vX"));
1037             inputs[attrib][1] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_center}, "vY"));
1038          } else {
1039             inputs[attrib][0] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_UL}, "vX"));
1040             inputs[attrib][1] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_UL}, "vY"));
1041          }
1042          inputs[attrib][2] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vZ}, "vZ"));
1043          inputs[attrib][3] =
1044             wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}, "vOneOverW"));
1045          continue;
1046       } else if (semantic_name == TGSI_SEMANTIC_PRIMID) {
1047          Value *primID = LOAD(pPS, {0, SWR_PS_CONTEXT_primID}, "primID");
1048          inputs[attrib][0] = wrap(VECTOR_SPLAT(JM()->mVWidth, primID));
1049          inputs[attrib][1] = wrap(VIMMED1(0));
1050          inputs[attrib][2] = wrap(VIMMED1(0));
1051          inputs[attrib][3] = wrap(VIMMED1(0));
1052          continue;
1053       }
1054
1055       unsigned linkedAttrib =
1056          locate_linkage(semantic_name, semantic_idx, pPrevShader);
1057
1058       if (semantic_name == TGSI_SEMANTIC_GENERIC &&
1059           key.sprite_coord_enable & (1 << semantic_idx)) {
1060          /* we add an extra attrib to the backendState in swr_update_derived. */
1061          linkedAttrib = pPrevShader->num_outputs - 1;
1062          swr_fs->pointSpriteMask |= (1 << linkedAttrib);
1063       } else if (linkedAttrib == 0xFFFFFFFF) {
1064          inputs[attrib][0] = wrap(VIMMED1(0.0f));
1065          inputs[attrib][1] = wrap(VIMMED1(0.0f));
1066          inputs[attrib][2] = wrap(VIMMED1(0.0f));
1067          inputs[attrib][3] = wrap(VIMMED1(1.0f));
1068          /* If we're reading in color and 2-sided lighting is enabled, we have
1069           * to keep going.
1070           */
1071          if (semantic_name != TGSI_SEMANTIC_COLOR || !key.light_twoside)
1072             continue;
1073       } else {
1074          if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
1075             swr_fs->constantMask |= 1 << linkedAttrib;
1076          } else if (interpMode == TGSI_INTERPOLATE_COLOR) {
1077             swr_fs->flatConstantMask |= 1 << linkedAttrib;
1078          }
1079       }
1080
1081       unsigned bcolorAttrib = 0xFFFFFFFF;
1082       Value *offset = NULL;
1083       if (semantic_name == TGSI_SEMANTIC_COLOR && key.light_twoside) {
1084          bcolorAttrib = locate_linkage(
1085                TGSI_SEMANTIC_BCOLOR, semantic_idx, pPrevShader);
1086          /* Neither front nor back colors were available. Nothing to load. */
1087          if (bcolorAttrib == 0xFFFFFFFF && linkedAttrib == 0xFFFFFFFF)
1088             continue;
1089          /* If there is no front color, just always use the back color. */
1090          if (linkedAttrib == 0xFFFFFFFF)
1091             linkedAttrib = bcolorAttrib;
1092
1093          if (bcolorAttrib != 0xFFFFFFFF) {
1094             if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
1095                swr_fs->constantMask |= 1 << bcolorAttrib;
1096             } else if (interpMode == TGSI_INTERPOLATE_COLOR) {
1097                swr_fs->flatConstantMask |= 1 << bcolorAttrib;
1098             }
1099
1100             unsigned diff = 12 * (bcolorAttrib - linkedAttrib);
1101
1102             if (diff) {
1103                Value *back =
1104                   XOR(C(1), LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), "backFace");
1105
1106                offset = MUL(back, C(diff));
1107                offset->setName("offset");
1108             }
1109          }
1110       }
1111
1112       for (int channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
1113          if (mask & (1 << channel)) {
1114             Value *indexA = C(linkedAttrib * 12 + channel);
1115             Value *indexB = C(linkedAttrib * 12 + channel + 4);
1116             Value *indexC = C(linkedAttrib * 12 + channel + 8);
1117
1118             if (offset) {
1119                indexA = ADD(indexA, offset);
1120                indexB = ADD(indexB, offset);
1121                indexC = ADD(indexC, offset);
1122             }
1123
1124             Value *va = VBROADCAST(LOAD(GEP(pAttribs, indexA)));
1125             Value *vb = VBROADCAST(LOAD(GEP(pAttribs, indexB)));
1126             Value *vc = VBROADCAST(LOAD(GEP(pAttribs, indexC)));
1127
1128             if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
1129                inputs[attrib][channel] = wrap(va);
1130             } else {
1131                Value *vk = FSUB(FSUB(VIMMED1(1.0f), vi), vj);
1132
1133                vc = FMUL(vk, vc);
1134
1135                Value *interp = FMUL(va, vi);
1136                Value *interp1 = FMUL(vb, vj);
1137                interp = FADD(interp, interp1);
1138                interp = FADD(interp, vc);
1139                if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE ||
1140                    interpMode == TGSI_INTERPOLATE_COLOR)
1141                   interp = FMUL(interp, vw);
1142                inputs[attrib][channel] = wrap(interp);
1143             }
1144          }
1145       }
1146    }
1147
1148    sampler = swr_sampler_soa_create(key.sampler, PIPE_SHADER_FRAGMENT);
1149
1150    struct lp_bld_tgsi_system_values system_values;
1151    memset(&system_values, 0, sizeof(system_values));
1152
1153    struct lp_build_mask_context mask;
1154    bool uses_mask = false;
1155
1156    if (swr_fs->info.base.uses_kill ||
1157        key.poly_stipple_enable) {
1158       Value *vActiveMask = NULL;
1159       if (swr_fs->info.base.uses_kill) {
1160          vActiveMask = LOAD(pPS, {0, SWR_PS_CONTEXT_activeMask}, "activeMask");
1161       }
1162       if (key.poly_stipple_enable) {
1163          // first get fragment xy coords and clip to stipple bounds
1164          Value *vXf = LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_UL});
1165          Value *vYf = LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_UL});
1166          Value *vXu = FP_TO_UI(vXf, mSimdInt32Ty);
1167          Value *vYu = FP_TO_UI(vYf, mSimdInt32Ty);
1168
1169          // stipple pattern is 32x32, which means that one line of stipple
1170          // is stored in one word:
1171          // vXstipple is bit offset inside 32-bit stipple word
1172          // vYstipple is word index is stipple array
1173          Value *vXstipple = AND(vXu, VIMMED1(0x1f)); // & (32-1)
1174          Value *vYstipple = AND(vYu, VIMMED1(0x1f)); // & (32-1)
1175
1176          // grab stipple pattern base address
1177          Value *stipplePtr = GEP(hPrivateData, {0, swr_draw_context_polyStipple, 0});
1178          stipplePtr = BITCAST(stipplePtr, mInt8PtrTy);
1179
1180          // peform a gather to grab stipple words for each lane
1181          Value *vStipple = GATHERDD(VUNDEF_I(), stipplePtr, vYstipple,
1182                                     VIMMED1(0xffffffff), C((char)4));
1183
1184          // create a mask with one bit corresponding to the x stipple
1185          // and AND it with the pattern, to see if we have a bit
1186          Value *vBitMask = LSHR(VIMMED1(0x80000000), vXstipple);
1187          Value *vStippleMask = AND(vStipple, vBitMask);
1188          vStippleMask = ICMP_NE(vStippleMask, VIMMED1(0));
1189          vStippleMask = VMASK(vStippleMask);
1190
1191          if (swr_fs->info.base.uses_kill) {
1192             vActiveMask = AND(vActiveMask, vStippleMask);
1193          } else {
1194             vActiveMask = vStippleMask;
1195          }
1196       }
1197       lp_build_mask_begin(
1198          &mask, gallivm, lp_type_float_vec(32, 32 * 8), wrap(vActiveMask));
1199       uses_mask = true;
1200    }
1201
1202    lp_build_tgsi_soa(gallivm,
1203                      swr_fs->pipe.tokens,
1204                      lp_type_float_vec(32, 32 * 8),
1205                      uses_mask ? &mask : NULL, // mask
1206                      wrap(consts_ptr),
1207                      wrap(const_sizes_ptr),
1208                      &system_values,
1209                      inputs,
1210                      outputs,
1211                      wrap(hPrivateData),
1212                      NULL, // thread data
1213                      sampler, // sampler
1214                      &swr_fs->info.base,
1215                      NULL); // geometry shader face
1216
1217    sampler->destroy(sampler);
1218
1219    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
1220
1221    for (uint32_t attrib = 0; attrib < swr_fs->info.base.num_outputs;
1222         attrib++) {
1223       switch (swr_fs->info.base.output_semantic_name[attrib]) {
1224       case TGSI_SEMANTIC_POSITION: {
1225          // write z
1226          LLVMValueRef outZ =
1227             LLVMBuildLoad(gallivm->builder, outputs[attrib][2], "");
1228          STORE(unwrap(outZ), pPS, {0, SWR_PS_CONTEXT_vZ});
1229          break;
1230       }
1231       case TGSI_SEMANTIC_COLOR: {
1232          for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
1233             if (!outputs[attrib][channel])
1234                continue;
1235
1236             LLVMValueRef out =
1237                LLVMBuildLoad(gallivm->builder, outputs[attrib][channel], "");
1238             if (swr_fs->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS] &&
1239                 swr_fs->info.base.output_semantic_index[attrib] == 0) {
1240                for (uint32_t rt = 0; rt < key.nr_cbufs; rt++) {
1241                   STORE(unwrap(out),
1242                         pPS,
1243                         {0, SWR_PS_CONTEXT_shaded, rt, channel});
1244                }
1245             } else {
1246                STORE(unwrap(out),
1247                      pPS,
1248                      {0,
1249                            SWR_PS_CONTEXT_shaded,
1250                            swr_fs->info.base.output_semantic_index[attrib],
1251                            channel});
1252             }
1253          }
1254          break;
1255       }
1256       default: {
1257          fprintf(stderr,
1258                  "unknown output from FS %s[%d]\n",
1259                  tgsi_semantic_names[swr_fs->info.base
1260                                         .output_semantic_name[attrib]],
1261                  swr_fs->info.base.output_semantic_index[attrib]);
1262          break;
1263       }
1264       }
1265    }
1266
1267    LLVMValueRef mask_result = 0;
1268    if (uses_mask) {
1269       mask_result = lp_build_mask_end(&mask);
1270    }
1271
1272    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
1273
1274    if (uses_mask) {
1275       STORE(unwrap(mask_result), pPS, {0, SWR_PS_CONTEXT_activeMask});
1276    }
1277
1278    RET_VOID();
1279
1280    gallivm_verify_function(gallivm, wrap(pFunction));
1281
1282    gallivm_compile_module(gallivm);
1283
1284    PFN_PIXEL_KERNEL kernel =
1285       (PFN_PIXEL_KERNEL)gallivm_jit_function(gallivm, wrap(pFunction));
1286    debug_printf("frag shader  %p\n", kernel);
1287    assert(kernel && "Error: FragShader = NULL");
1288
1289    JM()->mIsModuleFinalized = true;
1290
1291    return kernel;
1292 }
1293
1294 PFN_PIXEL_KERNEL
1295 swr_compile_fs(struct swr_context *ctx, swr_jit_fs_key &key)
1296 {
1297    if (!ctx->fs->pipe.tokens)
1298       return NULL;
1299
1300    BuilderSWR builder(
1301       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
1302       "FS");
1303    PFN_PIXEL_KERNEL func = builder.CompileFS(ctx, key);
1304
1305    ctx->fs->map.insert(std::make_pair(key, make_unique<VariantFS>(builder.gallivm, func)));
1306    return func;
1307 }