src/mesa/drivers/dri/i965/brw_shader.cpp

   1 /*
   2  * Copyright © 2010 Intel Corporation
   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 extern "C" {
  25 #include "main/macros.h"
  26 #include "brw_context.h"
  27 }
  28 #include "brw_vs.h"
  29 #include "brw_fs.h"
  30 #include "glsl/ir_optimization.h"
  31 #include "glsl/glsl_parser_extras.h"
  32 #include "main/shaderapi.h"
  33
  34 struct gl_shader *
  35 brw_new_shader(struct gl_context *ctx, GLuint name, GLuint type)
  36 {
  37    struct brw_shader *shader;
  38
  39    shader = rzalloc(NULL, struct brw_shader);
  40    if (shader) {
  41       shader->base.Type = type;
  42       shader->base.Name = name;
  43       _mesa_init_shader(ctx, &shader->base);
  44    }
  45
  46    return &shader->base;
  47 }
  48
  49 struct gl_shader_program *
  50 brw_new_shader_program(struct gl_context *ctx, GLuint name)
  51 {
  52    struct gl_shader_program *prog = rzalloc(NULL, struct gl_shader_program);
  53    if (prog) {
  54       prog->Name = name;
  55       _mesa_init_shader_program(ctx, prog);
  56    }
  57    return prog;
  58 }
  59
  60 /**
  61  * Performs a compile of the shader stages even when we don't know
  62  * what non-orthogonal state will be set, in the hope that it reflects
  63  * the eventual NOS used, and thus allows us to produce link failures.
  64  */
  65 static bool
  66 brw_shader_precompile(struct gl_context *ctx, struct gl_shader_program *prog)
  67 {
  68    struct brw_context *brw = brw_context(ctx);
  69
  70    if (brw->precompile && !brw_fs_precompile(ctx, prog))
  71       return false;
  72
  73    if (brw->precompile && !brw_vs_precompile(ctx, prog))
  74       return false;
  75
  76    return true;
  77 }
  78
  79 static void
  80 brw_lower_packing_builtins(struct brw_context *brw,
  81                            gl_shader_type shader_type,
  82                            exec_list *ir)
  83 {
  84    int ops = LOWER_PACK_SNORM_2x16
  85            | LOWER_UNPACK_SNORM_2x16
  86            | LOWER_PACK_UNORM_2x16
  87            | LOWER_UNPACK_UNORM_2x16
  88            | LOWER_PACK_SNORM_4x8
  89            | LOWER_UNPACK_SNORM_4x8
  90            | LOWER_PACK_UNORM_4x8
  91            | LOWER_UNPACK_UNORM_4x8;
  92
  93    if (brw->gen >= 7) {
  94       /* Gen7 introduced the f32to16 and f16to32 instructions, which can be
  95        * used to execute packHalf2x16 and unpackHalf2x16. For AOS code, no
  96        * lowering is needed. For SOA code, the Half2x16 ops must be
  97        * scalarized.
  98        */
  99       if (shader_type == MESA_SHADER_FRAGMENT) {
 100          ops |= LOWER_PACK_HALF_2x16_TO_SPLIT
 101              |  LOWER_UNPACK_HALF_2x16_TO_SPLIT;
 102       }
 103    } else {
 104       ops |= LOWER_PACK_HALF_2x16
 105           |  LOWER_UNPACK_HALF_2x16;
 106    }
 107
 108    lower_packing_builtins(ir, ops);
 109 }
 110
 111 GLboolean
 112 brw_link_shader(struct gl_context *ctx, struct gl_shader_program *shProg)
 113 {
 114    struct brw_context *brw = brw_context(ctx);
 115    unsigned int stage;
 116
 117    for (stage = 0; stage < ARRAY_SIZE(shProg->_LinkedShaders); stage++) {
 118       struct brw_shader *shader =
 119          (struct brw_shader *)shProg->_LinkedShaders[stage];
 120
 121       if (!shader)
 122          continue;
 123
 124       struct gl_program *prog =
 125          ctx->Driver.NewProgram(ctx, _mesa_program_index_to_target(stage),
 126                                 shader->base.Name);
 127       if (!prog)
 128         return false;
 129       prog->Parameters = _mesa_new_parameter_list();
 130
 131       _mesa_copy_linked_program_data((gl_shader_type) stage, shProg, prog);
 132
 133       void *mem_ctx = ralloc_context(NULL);
 134       bool progress;
 135
 136       if (shader->ir)
 137          ralloc_free(shader->ir);
 138       shader->ir = new(shader) exec_list;
 139       clone_ir_list(mem_ctx, shader->ir, shader->base.ir);
 140
 141       /* lower_packing_builtins() inserts arithmetic instructions, so it
 142        * must precede lower_instructions().
 143        */
 144       brw_lower_packing_builtins(brw, (gl_shader_type) stage, shader->ir);
 145       do_mat_op_to_vec(shader->ir);
 146       const int bitfield_insert = brw->gen >= 7
 147                                   ? BITFIELD_INSERT_TO_BFM_BFI
 148                                   : 0;
 149       const int lrp_to_arith = brw->gen < 6 ? LRP_TO_ARITH : 0;
 150       lower_instructions(shader->ir,
 151                          MOD_TO_FRACT |
 152                          DIV_TO_MUL_RCP |
 153                          SUB_TO_ADD_NEG |
 154                          EXP_TO_EXP2 |
 155                          LOG_TO_LOG2 |
 156                          bitfield_insert |
 157                          lrp_to_arith);
 158
 159       /* Pre-gen6 HW can only nest if-statements 16 deep.  Beyond this,
 160        * if-statements need to be flattened.
 161        */
 162       if (brw->gen < 6)
 163          lower_if_to_cond_assign(shader->ir, 16);
 164
 165       do_lower_texture_projection(shader->ir);
 166       brw_lower_texture_gradients(brw, shader->ir);
 167       do_vec_index_to_cond_assign(shader->ir);
 168       lower_vector_insert(shader->ir, true);
 169       brw_do_cubemap_normalize(shader->ir);
 170       lower_noise(shader->ir);
 171       lower_quadop_vector(shader->ir, false);
 172
 173       bool input = true;
 174       bool output = stage == MESA_SHADER_FRAGMENT;
 175       bool temp = stage == MESA_SHADER_FRAGMENT;
 176       bool uniform = false;
 177
 178       bool lowered_variable_indexing =
 179          lower_variable_index_to_cond_assign(shader->ir,
 180                                              input, output, temp, uniform);
 181
 182       if (unlikely(brw->perf_debug && lowered_variable_indexing)) {
 183          perf_debug("Unsupported form of variable indexing in FS; falling "
 184                     "back to very inefficient code generation\n");
 185       }
 186
 187       /* FINISHME: Do this before the variable index lowering. */
 188       lower_ubo_reference(&shader->base, shader->ir);
 189
 190       do {
 191          progress = false;
 192
 193          if (stage == MESA_SHADER_FRAGMENT) {
 194             brw_do_channel_expressions(shader->ir);
 195             brw_do_vector_splitting(shader->ir);
 196          }
 197
 198          progress = do_lower_jumps(shader->ir, true, true,
 199                                    true, /* main return */
 200                                    false, /* continue */
 201                                    false /* loops */
 202                                    ) || progress;
 203
 204          progress = do_common_optimization(shader->ir, true, true, 32,
 205                                            &ctx->ShaderCompilerOptions[stage])
 206            || progress;
 207       } while (progress);
 208
 209       /* Make a pass over the IR to add state references for any built-in
 210        * uniforms that are used.  This has to be done now (during linking).
 211        * Code generation doesn't happen until the first time this shader is
 212        * used for rendering.  Waiting until then to generate the parameters is
 213        * too late.  At that point, the values for the built-in uniforms won't
 214        * get sent to the shader.
 215        */
 216       foreach_list(node, shader->ir) {
 217          ir_variable *var = ((ir_instruction *) node)->as_variable();
 218
 219          if ((var == NULL) || (var->mode != ir_var_uniform)
 220              || (strncmp(var->name, "gl_", 3) != 0))
 221             continue;
 222
 223          const ir_state_slot *const slots = var->state_slots;
 224          assert(var->state_slots != NULL);
 225
 226          for (unsigned int i = 0; i < var->num_state_slots; i++) {
 227             _mesa_add_state_reference(prog->Parameters,
 228                                       (gl_state_index *) slots[i].tokens);
 229          }
 230       }
 231
 232       validate_ir_tree(shader->ir);
 233
 234       reparent_ir(shader->ir, shader->ir);
 235       ralloc_free(mem_ctx);
 236
 237       do_set_program_inouts(shader->ir, prog, shader->base.Type);
 238
 239       prog->SamplersUsed = shader->base.active_samplers;
 240       _mesa_update_shader_textures_used(shProg, prog);
 241
 242       _mesa_reference_program(ctx, &shader->base.Program, prog);
 243
 244       brw_add_texrect_params(prog);
 245
 246       /* This has to be done last.  Any operation that can cause
 247        * prog->ParameterValues to get reallocated (e.g., anything that adds a
 248        * program constant) has to happen before creating this linkage.
 249        */
 250       _mesa_associate_uniform_storage(ctx, shProg, prog->Parameters);
 251
 252       _mesa_reference_program(ctx, &prog, NULL);
 253
 254       if (ctx->Shader.Flags & GLSL_DUMP) {
 255          printf("\n");
 256          printf("GLSL IR for linked %s program %d:\n",
 257                 _mesa_glsl_shader_target_name(shader->base.Type), shProg->Name);
 258          _mesa_print_ir(shader->base.ir, NULL);
 259          printf("\n");
 260       }
 261    }
 262
 263    if (ctx->Shader.Flags & GLSL_DUMP) {
 264       for (unsigned i = 0; i < shProg->NumShaders; i++) {
 265          const struct gl_shader *sh = shProg->Shaders[i];
 266          if (!sh)
 267             continue;
 268
 269          printf("GLSL %s shader %d source for linked program %d:\n",
 270                 _mesa_glsl_shader_target_name(sh->Type),
 271                 i,
 272                 shProg->Name);
 273          printf("%s", sh->Source);
 274          printf("\n");
 275       }
 276    }
 277
 278    if (!brw_shader_precompile(ctx, shProg))
 279       return false;
 280
 281    return true;
 282 }
 283
 284
 285 int
 286 brw_type_for_base_type(const struct glsl_type *type)
 287 {
 288    switch (type->base_type) {
 289    case GLSL_TYPE_FLOAT:
 290       return BRW_REGISTER_TYPE_F;
 291    case GLSL_TYPE_INT:
 292    case GLSL_TYPE_BOOL:
 293       return BRW_REGISTER_TYPE_D;
 294    case GLSL_TYPE_UINT:
 295       return BRW_REGISTER_TYPE_UD;
 296    case GLSL_TYPE_ARRAY:
 297       return brw_type_for_base_type(type->fields.array);
 298    case GLSL_TYPE_STRUCT:
 299    case GLSL_TYPE_SAMPLER:
 300       /* These should be overridden with the type of the member when
 301        * dereferenced into.  BRW_REGISTER_TYPE_UD seems like a likely
 302        * way to trip up if we don't.
 303        */
 304       return BRW_REGISTER_TYPE_UD;
 305    case GLSL_TYPE_VOID:
 306    case GLSL_TYPE_ERROR:
 307    case GLSL_TYPE_INTERFACE:
 308       assert(!"not reached");
 309       break;
 310    }
 311
 312    return BRW_REGISTER_TYPE_F;
 313 }
 314
 315 uint32_t
 316 brw_conditional_for_comparison(unsigned int op)
 317 {
 318    switch (op) {
 319    case ir_binop_less:
 320       return BRW_CONDITIONAL_L;
 321    case ir_binop_greater:
 322       return BRW_CONDITIONAL_G;
 323    case ir_binop_lequal:
 324       return BRW_CONDITIONAL_LE;
 325    case ir_binop_gequal:
 326       return BRW_CONDITIONAL_GE;
 327    case ir_binop_equal:
 328    case ir_binop_all_equal: /* same as equal for scalars */
 329       return BRW_CONDITIONAL_Z;
 330    case ir_binop_nequal:
 331    case ir_binop_any_nequal: /* same as nequal for scalars */
 332       return BRW_CONDITIONAL_NZ;
 333    default:
 334       assert(!"not reached: bad operation for comparison");
 335       return BRW_CONDITIONAL_NZ;
 336    }
 337 }
 338
 339 uint32_t
 340 brw_math_function(enum opcode op)
 341 {
 342    switch (op) {
 343    case SHADER_OPCODE_RCP:
 344       return BRW_MATH_FUNCTION_INV;
 345    case SHADER_OPCODE_RSQ:
 346       return BRW_MATH_FUNCTION_RSQ;
 347    case SHADER_OPCODE_SQRT:
 348       return BRW_MATH_FUNCTION_SQRT;
 349    case SHADER_OPCODE_EXP2:
 350       return BRW_MATH_FUNCTION_EXP;
 351    case SHADER_OPCODE_LOG2:
 352       return BRW_MATH_FUNCTION_LOG;
 353    case SHADER_OPCODE_POW:
 354       return BRW_MATH_FUNCTION_POW;
 355    case SHADER_OPCODE_SIN:
 356       return BRW_MATH_FUNCTION_SIN;
 357    case SHADER_OPCODE_COS:
 358       return BRW_MATH_FUNCTION_COS;
 359    case SHADER_OPCODE_INT_QUOTIENT:
 360       return BRW_MATH_FUNCTION_INT_DIV_QUOTIENT;
 361    case SHADER_OPCODE_INT_REMAINDER:
 362       return BRW_MATH_FUNCTION_INT_DIV_REMAINDER;
 363    default:
 364       assert(!"not reached: unknown math function");
 365       return 0;
 366    }
 367 }
 368
 369 uint32_t
 370 brw_texture_offset(ir_constant *offset)
 371 {
 372    assert(offset != NULL);
 373
 374    signed char offsets[3];
 375    for (unsigned i = 0; i < offset->type->vector_elements; i++)
 376       offsets[i] = (signed char) offset->value.i[i];
 377
 378    /* Combine all three offsets into a single unsigned dword:
 379     *
 380     *    bits 11:8 - U Offset (X component)
 381     *    bits  7:4 - V Offset (Y component)
 382     *    bits  3:0 - R Offset (Z component)
 383     */
 384    unsigned offset_bits = 0;
 385    for (unsigned i = 0; i < offset->type->vector_elements; i++) {
 386       const unsigned shift = 4 * (2 - i);
 387       offset_bits |= (offsets[i] << shift) & (0xF << shift);
 388    }
 389    return offset_bits;
 390 }
 391
 392 const char *
 393 brw_instruction_name(enum opcode op)
 394 {
 395    char *fallback;
 396
 397    if (op < ARRAY_SIZE(opcode_descs) && opcode_descs[op].name)
 398       return opcode_descs[op].name;
 399
 400    switch (op) {
 401    case FS_OPCODE_FB_WRITE:
 402       return "fb_write";
 403
 404    case SHADER_OPCODE_RCP:
 405       return "rcp";
 406    case SHADER_OPCODE_RSQ:
 407       return "rsq";
 408    case SHADER_OPCODE_SQRT:
 409       return "sqrt";
 410    case SHADER_OPCODE_EXP2:
 411       return "exp2";
 412    case SHADER_OPCODE_LOG2:
 413       return "log2";
 414    case SHADER_OPCODE_POW:
 415       return "pow";
 416    case SHADER_OPCODE_INT_QUOTIENT:
 417       return "int_quot";
 418    case SHADER_OPCODE_INT_REMAINDER:
 419       return "int_rem";
 420    case SHADER_OPCODE_SIN:
 421       return "sin";
 422    case SHADER_OPCODE_COS:
 423       return "cos";
 424
 425    case SHADER_OPCODE_TEX:
 426       return "tex";
 427    case SHADER_OPCODE_TXD:
 428       return "txd";
 429    case SHADER_OPCODE_TXF:
 430       return "txf";
 431    case SHADER_OPCODE_TXL:
 432       return "txl";
 433    case SHADER_OPCODE_TXS:
 434       return "txs";
 435    case FS_OPCODE_TXB:
 436       return "txb";
 437    case SHADER_OPCODE_TXF_MS:
 438       return "txf_ms";
 439
 440    case FS_OPCODE_DDX:
 441       return "ddx";
 442    case FS_OPCODE_DDY:
 443       return "ddy";
 444
 445    case FS_OPCODE_PIXEL_X:
 446       return "pixel_x";
 447    case FS_OPCODE_PIXEL_Y:
 448       return "pixel_y";
 449
 450    case FS_OPCODE_CINTERP:
 451       return "cinterp";
 452    case FS_OPCODE_LINTERP:
 453       return "linterp";
 454
 455    case FS_OPCODE_SPILL:
 456       return "spill";
 457    case FS_OPCODE_UNSPILL:
 458       return "unspill";
 459
 460    case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD:
 461       return "uniform_pull_const";
 462    case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7:
 463       return "uniform_pull_const_gen7";
 464    case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD:
 465       return "varying_pull_const";
 466    case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7:
 467       return "varying_pull_const_gen7";
 468
 469    case FS_OPCODE_MOV_DISPATCH_TO_FLAGS:
 470       return "mov_dispatch_to_flags";
 471    case FS_OPCODE_DISCARD_JUMP:
 472       return "discard_jump";
 473
 474    case FS_OPCODE_SET_SIMD4X2_OFFSET:
 475       return "set_simd4x2_offset";
 476
 477    case FS_OPCODE_PACK_HALF_2x16_SPLIT:
 478       return "pack_half_2x16_split";
 479    case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X:
 480       return "unpack_half_2x16_split_x";
 481    case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y:
 482       return "unpack_half_2x16_split_y";
 483
 484    case FS_OPCODE_PLACEHOLDER_HALT:
 485       return "placeholder_halt";
 486
 487    case VS_OPCODE_URB_WRITE:
 488       return "vs_urb_write";
 489    case VS_OPCODE_SCRATCH_READ:
 490       return "scratch_read";
 491    case VS_OPCODE_SCRATCH_WRITE:
 492       return "scratch_write";
 493    case VS_OPCODE_PULL_CONSTANT_LOAD:
 494       return "pull_constant_load";
 495    case VS_OPCODE_PULL_CONSTANT_LOAD_GEN7:
 496       return "pull_constant_load_gen7";
 497    case VS_OPCODE_UNPACK_FLAGS_SIMD4X2:
 498       return "unpack_flags_simd4x2";
 499
 500    case GS_OPCODE_URB_WRITE:
 501       return "gs_urb_write";
 502    case GS_OPCODE_THREAD_END:
 503       return "gs_thread_end";
 504    case GS_OPCODE_SET_WRITE_OFFSET:
 505       return "set_write_offset";
 506    case GS_OPCODE_SET_VERTEX_COUNT:
 507       return "set_vertex_count";
 508
 509    default:
 510       /* Yes, this leaks.  It's in debug code, it should never occur, and if
 511        * it does, you should just add the case to the list above.
 512        */
 513       asprintf(&fallback, "op%d", op);
 514       return fallback;
 515    }
 516 }
 517
 518 bool
 519 backend_instruction::is_tex()
 520 {
 521    return (opcode == SHADER_OPCODE_TEX ||
 522            opcode == FS_OPCODE_TXB ||
 523            opcode == SHADER_OPCODE_TXD ||
 524            opcode == SHADER_OPCODE_TXF ||
 525            opcode == SHADER_OPCODE_TXF_MS ||
 526            opcode == SHADER_OPCODE_TXL ||
 527            opcode == SHADER_OPCODE_TXS ||
 528            opcode == SHADER_OPCODE_LOD);
 529 }
 530
 531 bool
 532 backend_instruction::is_math()
 533 {
 534    return (opcode == SHADER_OPCODE_RCP ||
 535            opcode == SHADER_OPCODE_RSQ ||
 536            opcode == SHADER_OPCODE_SQRT ||
 537            opcode == SHADER_OPCODE_EXP2 ||
 538            opcode == SHADER_OPCODE_LOG2 ||
 539            opcode == SHADER_OPCODE_SIN ||
 540            opcode == SHADER_OPCODE_COS ||
 541            opcode == SHADER_OPCODE_INT_QUOTIENT ||
 542            opcode == SHADER_OPCODE_INT_REMAINDER ||
 543            opcode == SHADER_OPCODE_POW);
 544 }
 545
 546 bool
 547 backend_instruction::is_control_flow()
 548 {
 549    switch (opcode) {
 550    case BRW_OPCODE_DO:
 551    case BRW_OPCODE_WHILE:
 552    case BRW_OPCODE_IF:
 553    case BRW_OPCODE_ELSE:
 554    case BRW_OPCODE_ENDIF:
 555    case BRW_OPCODE_BREAK:
 556    case BRW_OPCODE_CONTINUE:
 557       return true;
 558    default:
 559       return false;
 560    }
 561 }
 562
 563 void
 564 backend_visitor::dump_instructions()
 565 {
 566    int ip = 0;
 567    foreach_list(node, &this->instructions) {
 568       backend_instruction *inst = (backend_instruction *)node;
 569       printf("%d: ", ip++);
 570       dump_instruction(inst);
 571    }
 572 }