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