brw_vec4_visitor.cpp revision 9195191e50429d9cf25e6498f9fb108758ac2be6
1/* 2 * Copyright © 2011 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#include "brw_vec4.h" 25extern "C" { 26#include "main/macros.h" 27#include "program/prog_parameter.h" 28#include "program/sampler.h" 29} 30 31namespace brw { 32 33src_reg::src_reg(dst_reg reg) 34{ 35 init(); 36 37 this->file = reg.file; 38 this->reg = reg.reg; 39 this->reg_offset = reg.reg_offset; 40 this->type = reg.type; 41 this->reladdr = reg.reladdr; 42 this->fixed_hw_reg = reg.fixed_hw_reg; 43 44 int swizzles[4]; 45 int next_chan = 0; 46 int last = 0; 47 48 for (int i = 0; i < 4; i++) { 49 if (!(reg.writemask & (1 << i))) 50 continue; 51 52 swizzles[next_chan++] = last = i; 53 } 54 55 for (; next_chan < 4; next_chan++) { 56 swizzles[next_chan] = last; 57 } 58 59 this->swizzle = BRW_SWIZZLE4(swizzles[0], swizzles[1], 60 swizzles[2], swizzles[3]); 61} 62 63dst_reg::dst_reg(src_reg reg) 64{ 65 init(); 66 67 this->file = reg.file; 68 this->reg = reg.reg; 69 this->reg_offset = reg.reg_offset; 70 this->type = reg.type; 71 this->writemask = WRITEMASK_XYZW; 72 this->reladdr = reg.reladdr; 73 this->fixed_hw_reg = reg.fixed_hw_reg; 74} 75 76vec4_instruction::vec4_instruction(vec4_visitor *v, 77 enum opcode opcode, dst_reg dst, 78 src_reg src0, src_reg src1, src_reg src2) 79{ 80 this->opcode = opcode; 81 this->dst = dst; 82 this->src[0] = src0; 83 this->src[1] = src1; 84 this->src[2] = src2; 85 this->ir = v->base_ir; 86 this->annotation = v->current_annotation; 87} 88 89vec4_instruction * 90vec4_visitor::emit(vec4_instruction *inst) 91{ 92 this->instructions.push_tail(inst); 93 94 return inst; 95} 96 97vec4_instruction * 98vec4_visitor::emit_before(vec4_instruction *inst, vec4_instruction *new_inst) 99{ 100 new_inst->ir = inst->ir; 101 new_inst->annotation = inst->annotation; 102 103 inst->insert_before(new_inst); 104 105 return inst; 106} 107 108vec4_instruction * 109vec4_visitor::emit(enum opcode opcode, dst_reg dst, 110 src_reg src0, src_reg src1, src_reg src2) 111{ 112 return emit(new(mem_ctx) vec4_instruction(this, opcode, dst, 113 src0, src1, src2)); 114} 115 116 117vec4_instruction * 118vec4_visitor::emit(enum opcode opcode, dst_reg dst, src_reg src0, src_reg src1) 119{ 120 return emit(new(mem_ctx) vec4_instruction(this, opcode, dst, src0, src1)); 121} 122 123vec4_instruction * 124vec4_visitor::emit(enum opcode opcode, dst_reg dst, src_reg src0) 125{ 126 return emit(new(mem_ctx) vec4_instruction(this, opcode, dst, src0)); 127} 128 129vec4_instruction * 130vec4_visitor::emit(enum opcode opcode) 131{ 132 return emit(new(mem_ctx) vec4_instruction(this, opcode, dst_reg())); 133} 134 135#define ALU1(op) \ 136 vec4_instruction * \ 137 vec4_visitor::op(dst_reg dst, src_reg src0) \ 138 { \ 139 return new(mem_ctx) vec4_instruction(this, BRW_OPCODE_##op, dst, \ 140 src0); \ 141 } 142 143#define ALU2(op) \ 144 vec4_instruction * \ 145 vec4_visitor::op(dst_reg dst, src_reg src0, src_reg src1) \ 146 { \ 147 return new(mem_ctx) vec4_instruction(this, BRW_OPCODE_##op, dst, \ 148 src0, src1); \ 149 } 150 151ALU1(NOT) 152ALU1(MOV) 153ALU1(FRC) 154ALU1(RNDD) 155ALU1(RNDE) 156ALU1(RNDZ) 157ALU2(ADD) 158ALU2(MUL) 159ALU2(MACH) 160ALU2(AND) 161ALU2(OR) 162ALU2(XOR) 163ALU2(DP3) 164ALU2(DP4) 165 166/** Gen4 predicated IF. */ 167vec4_instruction * 168vec4_visitor::IF(uint32_t predicate) 169{ 170 vec4_instruction *inst; 171 172 inst = new(mem_ctx) vec4_instruction(this, BRW_OPCODE_IF); 173 inst->predicate = predicate; 174 175 return inst; 176} 177 178/** Gen6+ IF with embedded comparison. */ 179vec4_instruction * 180vec4_visitor::IF(src_reg src0, src_reg src1, uint32_t condition) 181{ 182 assert(intel->gen >= 6); 183 184 vec4_instruction *inst; 185 186 resolve_ud_negate(&src0); 187 resolve_ud_negate(&src1); 188 189 inst = new(mem_ctx) vec4_instruction(this, BRW_OPCODE_IF, dst_null_d(), 190 src0, src1); 191 inst->conditional_mod = condition; 192 193 return inst; 194} 195 196/** 197 * CMP: Sets the low bit of the destination channels with the result 198 * of the comparison, while the upper bits are undefined, and updates 199 * the flag register with the packed 16 bits of the result. 200 */ 201vec4_instruction * 202vec4_visitor::CMP(dst_reg dst, src_reg src0, src_reg src1, uint32_t condition) 203{ 204 vec4_instruction *inst; 205 206 /* original gen4 does type conversion to the destination type 207 * before before comparison, producing garbage results for floating 208 * point comparisons. 209 */ 210 if (intel->gen == 4) { 211 dst.type = src0.type; 212 if (dst.file == HW_REG) 213 dst.fixed_hw_reg.type = dst.type; 214 } 215 216 resolve_ud_negate(&src0); 217 resolve_ud_negate(&src1); 218 219 inst = new(mem_ctx) vec4_instruction(this, BRW_OPCODE_CMP, dst, src0, src1); 220 inst->conditional_mod = condition; 221 222 return inst; 223} 224 225vec4_instruction * 226vec4_visitor::SCRATCH_READ(dst_reg dst, src_reg index) 227{ 228 vec4_instruction *inst; 229 230 inst = new(mem_ctx) vec4_instruction(this, VS_OPCODE_SCRATCH_READ, 231 dst, index); 232 inst->base_mrf = 14; 233 inst->mlen = 1; 234 235 return inst; 236} 237 238vec4_instruction * 239vec4_visitor::SCRATCH_WRITE(dst_reg dst, src_reg src, src_reg index) 240{ 241 vec4_instruction *inst; 242 243 inst = new(mem_ctx) vec4_instruction(this, VS_OPCODE_SCRATCH_WRITE, 244 dst, src, index); 245 inst->base_mrf = 13; 246 inst->mlen = 2; 247 248 return inst; 249} 250 251void 252vec4_visitor::emit_dp(dst_reg dst, src_reg src0, src_reg src1, unsigned elements) 253{ 254 static enum opcode dot_opcodes[] = { 255 BRW_OPCODE_DP2, BRW_OPCODE_DP3, BRW_OPCODE_DP4 256 }; 257 258 emit(dot_opcodes[elements - 2], dst, src0, src1); 259} 260 261void 262vec4_visitor::emit_math1_gen6(enum opcode opcode, dst_reg dst, src_reg src) 263{ 264 /* The gen6 math instruction ignores the source modifiers -- 265 * swizzle, abs, negate, and at least some parts of the register 266 * region description. 267 * 268 * While it would seem that this MOV could be avoided at this point 269 * in the case that the swizzle is matched up with the destination 270 * writemask, note that uniform packing and register allocation 271 * could rearrange our swizzle, so let's leave this matter up to 272 * copy propagation later. 273 */ 274 src_reg temp_src = src_reg(this, glsl_type::vec4_type); 275 emit(MOV(dst_reg(temp_src), src)); 276 277 if (dst.writemask != WRITEMASK_XYZW) { 278 /* The gen6 math instruction must be align1, so we can't do 279 * writemasks. 280 */ 281 dst_reg temp_dst = dst_reg(this, glsl_type::vec4_type); 282 283 emit(opcode, temp_dst, temp_src); 284 285 emit(MOV(dst, src_reg(temp_dst))); 286 } else { 287 emit(opcode, dst, temp_src); 288 } 289} 290 291void 292vec4_visitor::emit_math1_gen4(enum opcode opcode, dst_reg dst, src_reg src) 293{ 294 vec4_instruction *inst = emit(opcode, dst, src); 295 inst->base_mrf = 1; 296 inst->mlen = 1; 297} 298 299void 300vec4_visitor::emit_math(opcode opcode, dst_reg dst, src_reg src) 301{ 302 switch (opcode) { 303 case SHADER_OPCODE_RCP: 304 case SHADER_OPCODE_RSQ: 305 case SHADER_OPCODE_SQRT: 306 case SHADER_OPCODE_EXP2: 307 case SHADER_OPCODE_LOG2: 308 case SHADER_OPCODE_SIN: 309 case SHADER_OPCODE_COS: 310 break; 311 default: 312 assert(!"not reached: bad math opcode"); 313 return; 314 } 315 316 if (intel->gen >= 7) { 317 emit(opcode, dst, src); 318 } else if (intel->gen == 6) { 319 return emit_math1_gen6(opcode, dst, src); 320 } else { 321 return emit_math1_gen4(opcode, dst, src); 322 } 323} 324 325void 326vec4_visitor::emit_math2_gen6(enum opcode opcode, 327 dst_reg dst, src_reg src0, src_reg src1) 328{ 329 src_reg expanded; 330 331 /* The gen6 math instruction ignores the source modifiers -- 332 * swizzle, abs, negate, and at least some parts of the register 333 * region description. Move the sources to temporaries to make it 334 * generally work. 335 */ 336 337 expanded = src_reg(this, glsl_type::vec4_type); 338 expanded.type = src0.type; 339 emit(MOV(dst_reg(expanded), src0)); 340 src0 = expanded; 341 342 expanded = src_reg(this, glsl_type::vec4_type); 343 expanded.type = src1.type; 344 emit(MOV(dst_reg(expanded), src1)); 345 src1 = expanded; 346 347 if (dst.writemask != WRITEMASK_XYZW) { 348 /* The gen6 math instruction must be align1, so we can't do 349 * writemasks. 350 */ 351 dst_reg temp_dst = dst_reg(this, glsl_type::vec4_type); 352 temp_dst.type = dst.type; 353 354 emit(opcode, temp_dst, src0, src1); 355 356 emit(MOV(dst, src_reg(temp_dst))); 357 } else { 358 emit(opcode, dst, src0, src1); 359 } 360} 361 362void 363vec4_visitor::emit_math2_gen4(enum opcode opcode, 364 dst_reg dst, src_reg src0, src_reg src1) 365{ 366 vec4_instruction *inst = emit(opcode, dst, src0, src1); 367 inst->base_mrf = 1; 368 inst->mlen = 2; 369} 370 371void 372vec4_visitor::emit_math(enum opcode opcode, 373 dst_reg dst, src_reg src0, src_reg src1) 374{ 375 switch (opcode) { 376 case SHADER_OPCODE_POW: 377 case SHADER_OPCODE_INT_QUOTIENT: 378 case SHADER_OPCODE_INT_REMAINDER: 379 break; 380 default: 381 assert(!"not reached: unsupported binary math opcode"); 382 return; 383 } 384 385 if (intel->gen >= 7) { 386 emit(opcode, dst, src0, src1); 387 } else if (intel->gen == 6) { 388 return emit_math2_gen6(opcode, dst, src0, src1); 389 } else { 390 return emit_math2_gen4(opcode, dst, src0, src1); 391 } 392} 393 394void 395vec4_visitor::visit_instructions(const exec_list *list) 396{ 397 foreach_list(node, list) { 398 ir_instruction *ir = (ir_instruction *)node; 399 400 base_ir = ir; 401 ir->accept(this); 402 } 403} 404 405 406static int 407type_size(const struct glsl_type *type) 408{ 409 unsigned int i; 410 int size; 411 412 switch (type->base_type) { 413 case GLSL_TYPE_UINT: 414 case GLSL_TYPE_INT: 415 case GLSL_TYPE_FLOAT: 416 case GLSL_TYPE_BOOL: 417 if (type->is_matrix()) { 418 return type->matrix_columns; 419 } else { 420 /* Regardless of size of vector, it gets a vec4. This is bad 421 * packing for things like floats, but otherwise arrays become a 422 * mess. Hopefully a later pass over the code can pack scalars 423 * down if appropriate. 424 */ 425 return 1; 426 } 427 case GLSL_TYPE_ARRAY: 428 assert(type->length > 0); 429 return type_size(type->fields.array) * type->length; 430 case GLSL_TYPE_STRUCT: 431 size = 0; 432 for (i = 0; i < type->length; i++) { 433 size += type_size(type->fields.structure[i].type); 434 } 435 return size; 436 case GLSL_TYPE_SAMPLER: 437 /* Samplers take up one slot in UNIFORMS[], but they're baked in 438 * at link time. 439 */ 440 return 1; 441 default: 442 assert(0); 443 return 0; 444 } 445} 446 447int 448vec4_visitor::virtual_grf_alloc(int size) 449{ 450 if (virtual_grf_array_size <= virtual_grf_count) { 451 if (virtual_grf_array_size == 0) 452 virtual_grf_array_size = 16; 453 else 454 virtual_grf_array_size *= 2; 455 virtual_grf_sizes = reralloc(mem_ctx, virtual_grf_sizes, int, 456 virtual_grf_array_size); 457 virtual_grf_reg_map = reralloc(mem_ctx, virtual_grf_reg_map, int, 458 virtual_grf_array_size); 459 } 460 virtual_grf_reg_map[virtual_grf_count] = virtual_grf_reg_count; 461 virtual_grf_reg_count += size; 462 virtual_grf_sizes[virtual_grf_count] = size; 463 return virtual_grf_count++; 464} 465 466src_reg::src_reg(class vec4_visitor *v, const struct glsl_type *type) 467{ 468 init(); 469 470 this->file = GRF; 471 this->reg = v->virtual_grf_alloc(type_size(type)); 472 473 if (type->is_array() || type->is_record()) { 474 this->swizzle = BRW_SWIZZLE_NOOP; 475 } else { 476 this->swizzle = swizzle_for_size(type->vector_elements); 477 } 478 479 this->type = brw_type_for_base_type(type); 480} 481 482dst_reg::dst_reg(class vec4_visitor *v, const struct glsl_type *type) 483{ 484 init(); 485 486 this->file = GRF; 487 this->reg = v->virtual_grf_alloc(type_size(type)); 488 489 if (type->is_array() || type->is_record()) { 490 this->writemask = WRITEMASK_XYZW; 491 } else { 492 this->writemask = (1 << type->vector_elements) - 1; 493 } 494 495 this->type = brw_type_for_base_type(type); 496} 497 498/* Our support for uniforms is piggy-backed on the struct 499 * gl_fragment_program, because that's where the values actually 500 * get stored, rather than in some global gl_shader_program uniform 501 * store. 502 */ 503int 504vec4_visitor::setup_uniform_values(int loc, const glsl_type *type) 505{ 506 unsigned int offset = 0; 507 float *values = &this->vp->Base.Parameters->ParameterValues[loc][0].f; 508 509 if (type->is_matrix()) { 510 const glsl_type *column = type->column_type(); 511 512 for (unsigned int i = 0; i < type->matrix_columns; i++) { 513 offset += setup_uniform_values(loc + offset, column); 514 } 515 516 return offset; 517 } 518 519 switch (type->base_type) { 520 case GLSL_TYPE_FLOAT: 521 case GLSL_TYPE_UINT: 522 case GLSL_TYPE_INT: 523 case GLSL_TYPE_BOOL: 524 for (unsigned int i = 0; i < type->vector_elements; i++) { 525 c->prog_data.param[this->uniforms * 4 + i] = &values[i]; 526 } 527 528 /* Set up pad elements to get things aligned to a vec4 boundary. */ 529 for (unsigned int i = type->vector_elements; i < 4; i++) { 530 static float zero = 0; 531 532 c->prog_data.param[this->uniforms * 4 + i] = &zero; 533 } 534 535 /* Track the size of this uniform vector, for future packing of 536 * uniforms. 537 */ 538 this->uniform_vector_size[this->uniforms] = type->vector_elements; 539 this->uniforms++; 540 541 return 1; 542 543 case GLSL_TYPE_STRUCT: 544 for (unsigned int i = 0; i < type->length; i++) { 545 offset += setup_uniform_values(loc + offset, 546 type->fields.structure[i].type); 547 } 548 return offset; 549 550 case GLSL_TYPE_ARRAY: 551 for (unsigned int i = 0; i < type->length; i++) { 552 offset += setup_uniform_values(loc + offset, type->fields.array); 553 } 554 return offset; 555 556 case GLSL_TYPE_SAMPLER: 557 /* The sampler takes up a slot, but we don't use any values from it. */ 558 return 1; 559 560 default: 561 assert(!"not reached"); 562 return 0; 563 } 564} 565 566void 567vec4_visitor::setup_uniform_clipplane_values() 568{ 569 gl_clip_plane *clip_planes = brw_select_clip_planes(ctx); 570 571 /* Pre-Gen6, we compact clip planes. For example, if the user 572 * enables just clip planes 0, 1, and 3, we will enable clip planes 573 * 0, 1, and 2 in the hardware, and we'll move clip plane 3 to clip 574 * plane 2. This simplifies the implementation of the Gen6 clip 575 * thread. 576 * 577 * In Gen6 and later, we don't compact clip planes, because this 578 * simplifies the implementation of gl_ClipDistance. 579 */ 580 int compacted_clipplane_index = 0; 581 for (int i = 0; i < c->key.nr_userclip_plane_consts; ++i) { 582 if (intel->gen < 6 && 583 !(c->key.userclip_planes_enabled_gen_4_5 & (1 << i))) { 584 continue; 585 } 586 this->uniform_vector_size[this->uniforms] = 4; 587 this->userplane[compacted_clipplane_index] = dst_reg(UNIFORM, this->uniforms); 588 this->userplane[compacted_clipplane_index].type = BRW_REGISTER_TYPE_F; 589 for (int j = 0; j < 4; ++j) { 590 c->prog_data.param[this->uniforms * 4 + j] = &clip_planes[i][j]; 591 } 592 ++compacted_clipplane_index; 593 ++this->uniforms; 594 } 595} 596 597/* Our support for builtin uniforms is even scarier than non-builtin. 598 * It sits on top of the PROG_STATE_VAR parameters that are 599 * automatically updated from GL context state. 600 */ 601void 602vec4_visitor::setup_builtin_uniform_values(ir_variable *ir) 603{ 604 const ir_state_slot *const slots = ir->state_slots; 605 assert(ir->state_slots != NULL); 606 607 for (unsigned int i = 0; i < ir->num_state_slots; i++) { 608 /* This state reference has already been setup by ir_to_mesa, 609 * but we'll get the same index back here. We can reference 610 * ParameterValues directly, since unlike brw_fs.cpp, we never 611 * add new state references during compile. 612 */ 613 int index = _mesa_add_state_reference(this->vp->Base.Parameters, 614 (gl_state_index *)slots[i].tokens); 615 float *values = &this->vp->Base.Parameters->ParameterValues[index][0].f; 616 617 this->uniform_vector_size[this->uniforms] = 0; 618 /* Add each of the unique swizzled channels of the element. 619 * This will end up matching the size of the glsl_type of this field. 620 */ 621 int last_swiz = -1; 622 for (unsigned int j = 0; j < 4; j++) { 623 int swiz = GET_SWZ(slots[i].swizzle, j); 624 last_swiz = swiz; 625 626 c->prog_data.param[this->uniforms * 4 + j] = &values[swiz]; 627 if (swiz <= last_swiz) 628 this->uniform_vector_size[this->uniforms]++; 629 } 630 this->uniforms++; 631 } 632} 633 634dst_reg * 635vec4_visitor::variable_storage(ir_variable *var) 636{ 637 return (dst_reg *)hash_table_find(this->variable_ht, var); 638} 639 640void 641vec4_visitor::emit_bool_to_cond_code(ir_rvalue *ir, uint32_t *predicate) 642{ 643 ir_expression *expr = ir->as_expression(); 644 645 *predicate = BRW_PREDICATE_NORMAL; 646 647 if (expr) { 648 src_reg op[2]; 649 vec4_instruction *inst; 650 651 assert(expr->get_num_operands() <= 2); 652 for (unsigned int i = 0; i < expr->get_num_operands(); i++) { 653 expr->operands[i]->accept(this); 654 op[i] = this->result; 655 656 resolve_ud_negate(&op[i]); 657 } 658 659 switch (expr->operation) { 660 case ir_unop_logic_not: 661 inst = emit(AND(dst_null_d(), op[0], src_reg(1))); 662 inst->conditional_mod = BRW_CONDITIONAL_Z; 663 break; 664 665 case ir_binop_logic_xor: 666 inst = emit(XOR(dst_null_d(), op[0], op[1])); 667 inst->conditional_mod = BRW_CONDITIONAL_NZ; 668 break; 669 670 case ir_binop_logic_or: 671 inst = emit(OR(dst_null_d(), op[0], op[1])); 672 inst->conditional_mod = BRW_CONDITIONAL_NZ; 673 break; 674 675 case ir_binop_logic_and: 676 inst = emit(AND(dst_null_d(), op[0], op[1])); 677 inst->conditional_mod = BRW_CONDITIONAL_NZ; 678 break; 679 680 case ir_unop_f2b: 681 if (intel->gen >= 6) { 682 emit(CMP(dst_null_d(), op[0], src_reg(0.0f), BRW_CONDITIONAL_NZ)); 683 } else { 684 inst = emit(MOV(dst_null_f(), op[0])); 685 inst->conditional_mod = BRW_CONDITIONAL_NZ; 686 } 687 break; 688 689 case ir_unop_i2b: 690 if (intel->gen >= 6) { 691 emit(CMP(dst_null_d(), op[0], src_reg(0), BRW_CONDITIONAL_NZ)); 692 } else { 693 inst = emit(MOV(dst_null_d(), op[0])); 694 inst->conditional_mod = BRW_CONDITIONAL_NZ; 695 } 696 break; 697 698 case ir_binop_all_equal: 699 inst = emit(CMP(dst_null_d(), op[0], op[1], BRW_CONDITIONAL_Z)); 700 *predicate = BRW_PREDICATE_ALIGN16_ALL4H; 701 break; 702 703 case ir_binop_any_nequal: 704 inst = emit(CMP(dst_null_d(), op[0], op[1], BRW_CONDITIONAL_NZ)); 705 *predicate = BRW_PREDICATE_ALIGN16_ANY4H; 706 break; 707 708 case ir_unop_any: 709 inst = emit(CMP(dst_null_d(), op[0], src_reg(0), BRW_CONDITIONAL_NZ)); 710 *predicate = BRW_PREDICATE_ALIGN16_ANY4H; 711 break; 712 713 case ir_binop_greater: 714 case ir_binop_gequal: 715 case ir_binop_less: 716 case ir_binop_lequal: 717 case ir_binop_equal: 718 case ir_binop_nequal: 719 emit(CMP(dst_null_d(), op[0], op[1], 720 brw_conditional_for_comparison(expr->operation))); 721 break; 722 723 default: 724 assert(!"not reached"); 725 break; 726 } 727 return; 728 } 729 730 ir->accept(this); 731 732 resolve_ud_negate(&this->result); 733 734 if (intel->gen >= 6) { 735 vec4_instruction *inst = emit(AND(dst_null_d(), 736 this->result, src_reg(1))); 737 inst->conditional_mod = BRW_CONDITIONAL_NZ; 738 } else { 739 vec4_instruction *inst = emit(MOV(dst_null_d(), this->result)); 740 inst->conditional_mod = BRW_CONDITIONAL_NZ; 741 } 742} 743 744/** 745 * Emit a gen6 IF statement with the comparison folded into the IF 746 * instruction. 747 */ 748void 749vec4_visitor::emit_if_gen6(ir_if *ir) 750{ 751 ir_expression *expr = ir->condition->as_expression(); 752 753 if (expr) { 754 src_reg op[2]; 755 dst_reg temp; 756 757 assert(expr->get_num_operands() <= 2); 758 for (unsigned int i = 0; i < expr->get_num_operands(); i++) { 759 expr->operands[i]->accept(this); 760 op[i] = this->result; 761 } 762 763 switch (expr->operation) { 764 case ir_unop_logic_not: 765 emit(IF(op[0], src_reg(0), BRW_CONDITIONAL_Z)); 766 return; 767 768 case ir_binop_logic_xor: 769 emit(IF(op[0], op[1], BRW_CONDITIONAL_NZ)); 770 return; 771 772 case ir_binop_logic_or: 773 temp = dst_reg(this, glsl_type::bool_type); 774 emit(OR(temp, op[0], op[1])); 775 emit(IF(src_reg(temp), src_reg(0), BRW_CONDITIONAL_NZ)); 776 return; 777 778 case ir_binop_logic_and: 779 temp = dst_reg(this, glsl_type::bool_type); 780 emit(AND(temp, op[0], op[1])); 781 emit(IF(src_reg(temp), src_reg(0), BRW_CONDITIONAL_NZ)); 782 return; 783 784 case ir_unop_f2b: 785 emit(IF(op[0], src_reg(0), BRW_CONDITIONAL_NZ)); 786 return; 787 788 case ir_unop_i2b: 789 emit(IF(op[0], src_reg(0), BRW_CONDITIONAL_NZ)); 790 return; 791 792 case ir_binop_greater: 793 case ir_binop_gequal: 794 case ir_binop_less: 795 case ir_binop_lequal: 796 case ir_binop_equal: 797 case ir_binop_nequal: 798 emit(IF(op[0], op[1], 799 brw_conditional_for_comparison(expr->operation))); 800 return; 801 802 case ir_binop_all_equal: 803 emit(CMP(dst_null_d(), op[0], op[1], BRW_CONDITIONAL_Z)); 804 emit(IF(BRW_PREDICATE_ALIGN16_ALL4H)); 805 return; 806 807 case ir_binop_any_nequal: 808 emit(CMP(dst_null_d(), op[0], op[1], BRW_CONDITIONAL_NZ)); 809 emit(IF(BRW_PREDICATE_ALIGN16_ANY4H)); 810 return; 811 812 case ir_unop_any: 813 emit(CMP(dst_null_d(), op[0], src_reg(0), BRW_CONDITIONAL_NZ)); 814 emit(IF(BRW_PREDICATE_ALIGN16_ANY4H)); 815 return; 816 817 default: 818 assert(!"not reached"); 819 emit(IF(op[0], src_reg(0), BRW_CONDITIONAL_NZ)); 820 return; 821 } 822 return; 823 } 824 825 ir->condition->accept(this); 826 827 emit(IF(this->result, src_reg(0), BRW_CONDITIONAL_NZ)); 828} 829 830void 831vec4_visitor::visit(ir_variable *ir) 832{ 833 dst_reg *reg = NULL; 834 835 if (variable_storage(ir)) 836 return; 837 838 switch (ir->mode) { 839 case ir_var_in: 840 reg = new(mem_ctx) dst_reg(ATTR, ir->location); 841 842 /* Do GL_FIXED rescaling for GLES2.0. Our GL_FIXED attributes 843 * come in as floating point conversions of the integer values. 844 */ 845 for (int i = ir->location; i < ir->location + type_size(ir->type); i++) { 846 if (!c->key.gl_fixed_input_size[i]) 847 continue; 848 849 dst_reg dst = *reg; 850 dst.type = brw_type_for_base_type(ir->type); 851 dst.writemask = (1 << c->key.gl_fixed_input_size[i]) - 1; 852 emit(MUL(dst, src_reg(dst), src_reg(1.0f / 65536.0f))); 853 } 854 break; 855 856 case ir_var_out: 857 reg = new(mem_ctx) dst_reg(this, ir->type); 858 859 for (int i = 0; i < type_size(ir->type); i++) { 860 output_reg[ir->location + i] = *reg; 861 output_reg[ir->location + i].reg_offset = i; 862 output_reg[ir->location + i].type = 863 brw_type_for_base_type(ir->type->get_scalar_type()); 864 output_reg_annotation[ir->location + i] = ir->name; 865 } 866 break; 867 868 case ir_var_auto: 869 case ir_var_temporary: 870 reg = new(mem_ctx) dst_reg(this, ir->type); 871 break; 872 873 case ir_var_uniform: 874 reg = new(this->mem_ctx) dst_reg(UNIFORM, this->uniforms); 875 876 /* Track how big the whole uniform variable is, in case we need to put a 877 * copy of its data into pull constants for array access. 878 */ 879 this->uniform_size[this->uniforms] = type_size(ir->type); 880 881 if (!strncmp(ir->name, "gl_", 3)) { 882 setup_builtin_uniform_values(ir); 883 } else { 884 setup_uniform_values(ir->location, ir->type); 885 } 886 break; 887 888 case ir_var_system_value: 889 /* VertexID is stored by the VF as the last vertex element, but 890 * we don't represent it with a flag in inputs_read, so we call 891 * it VERT_ATTRIB_MAX, which setup_attributes() picks up on. 892 */ 893 reg = new(mem_ctx) dst_reg(ATTR, VERT_ATTRIB_MAX); 894 prog_data->uses_vertexid = true; 895 896 switch (ir->location) { 897 case SYSTEM_VALUE_VERTEX_ID: 898 reg->writemask = WRITEMASK_X; 899 break; 900 case SYSTEM_VALUE_INSTANCE_ID: 901 reg->writemask = WRITEMASK_Y; 902 break; 903 default: 904 assert(!"not reached"); 905 break; 906 } 907 break; 908 909 default: 910 assert(!"not reached"); 911 } 912 913 reg->type = brw_type_for_base_type(ir->type); 914 hash_table_insert(this->variable_ht, reg, ir); 915} 916 917void 918vec4_visitor::visit(ir_loop *ir) 919{ 920 dst_reg counter; 921 922 /* We don't want debugging output to print the whole body of the 923 * loop as the annotation. 924 */ 925 this->base_ir = NULL; 926 927 if (ir->counter != NULL) { 928 this->base_ir = ir->counter; 929 ir->counter->accept(this); 930 counter = *(variable_storage(ir->counter)); 931 932 if (ir->from != NULL) { 933 this->base_ir = ir->from; 934 ir->from->accept(this); 935 936 emit(MOV(counter, this->result)); 937 } 938 } 939 940 emit(BRW_OPCODE_DO); 941 942 if (ir->to) { 943 this->base_ir = ir->to; 944 ir->to->accept(this); 945 946 emit(CMP(dst_null_d(), src_reg(counter), this->result, 947 brw_conditional_for_comparison(ir->cmp))); 948 949 vec4_instruction *inst = emit(BRW_OPCODE_BREAK); 950 inst->predicate = BRW_PREDICATE_NORMAL; 951 } 952 953 visit_instructions(&ir->body_instructions); 954 955 956 if (ir->increment) { 957 this->base_ir = ir->increment; 958 ir->increment->accept(this); 959 emit(ADD(counter, src_reg(counter), this->result)); 960 } 961 962 emit(BRW_OPCODE_WHILE); 963} 964 965void 966vec4_visitor::visit(ir_loop_jump *ir) 967{ 968 switch (ir->mode) { 969 case ir_loop_jump::jump_break: 970 emit(BRW_OPCODE_BREAK); 971 break; 972 case ir_loop_jump::jump_continue: 973 emit(BRW_OPCODE_CONTINUE); 974 break; 975 } 976} 977 978 979void 980vec4_visitor::visit(ir_function_signature *ir) 981{ 982 assert(0); 983 (void)ir; 984} 985 986void 987vec4_visitor::visit(ir_function *ir) 988{ 989 /* Ignore function bodies other than main() -- we shouldn't see calls to 990 * them since they should all be inlined. 991 */ 992 if (strcmp(ir->name, "main") == 0) { 993 const ir_function_signature *sig; 994 exec_list empty; 995 996 sig = ir->matching_signature(&empty); 997 998 assert(sig); 999 1000 visit_instructions(&sig->body); 1001 } 1002} 1003 1004bool 1005vec4_visitor::try_emit_sat(ir_expression *ir) 1006{ 1007 ir_rvalue *sat_src = ir->as_rvalue_to_saturate(); 1008 if (!sat_src) 1009 return false; 1010 1011 sat_src->accept(this); 1012 src_reg src = this->result; 1013 1014 this->result = src_reg(this, ir->type); 1015 vec4_instruction *inst; 1016 inst = emit(MOV(dst_reg(this->result), src)); 1017 inst->saturate = true; 1018 1019 return true; 1020} 1021 1022void 1023vec4_visitor::emit_bool_comparison(unsigned int op, 1024 dst_reg dst, src_reg src0, src_reg src1) 1025{ 1026 /* original gen4 does destination conversion before comparison. */ 1027 if (intel->gen < 5) 1028 dst.type = src0.type; 1029 1030 emit(CMP(dst, src0, src1, brw_conditional_for_comparison(op))); 1031 1032 dst.type = BRW_REGISTER_TYPE_D; 1033 emit(AND(dst, src_reg(dst), src_reg(0x1))); 1034} 1035 1036void 1037vec4_visitor::visit(ir_expression *ir) 1038{ 1039 unsigned int operand; 1040 src_reg op[Elements(ir->operands)]; 1041 src_reg result_src; 1042 dst_reg result_dst; 1043 vec4_instruction *inst; 1044 1045 if (try_emit_sat(ir)) 1046 return; 1047 1048 for (operand = 0; operand < ir->get_num_operands(); operand++) { 1049 this->result.file = BAD_FILE; 1050 ir->operands[operand]->accept(this); 1051 if (this->result.file == BAD_FILE) { 1052 printf("Failed to get tree for expression operand:\n"); 1053 ir->operands[operand]->print(); 1054 exit(1); 1055 } 1056 op[operand] = this->result; 1057 1058 /* Matrix expression operands should have been broken down to vector 1059 * operations already. 1060 */ 1061 assert(!ir->operands[operand]->type->is_matrix()); 1062 } 1063 1064 int vector_elements = ir->operands[0]->type->vector_elements; 1065 if (ir->operands[1]) { 1066 vector_elements = MAX2(vector_elements, 1067 ir->operands[1]->type->vector_elements); 1068 } 1069 1070 this->result.file = BAD_FILE; 1071 1072 /* Storage for our result. Ideally for an assignment we'd be using 1073 * the actual storage for the result here, instead. 1074 */ 1075 result_src = src_reg(this, ir->type); 1076 /* convenience for the emit functions below. */ 1077 result_dst = dst_reg(result_src); 1078 /* If nothing special happens, this is the result. */ 1079 this->result = result_src; 1080 /* Limit writes to the channels that will be used by result_src later. 1081 * This does limit this temp's use as a temporary for multi-instruction 1082 * sequences. 1083 */ 1084 result_dst.writemask = (1 << ir->type->vector_elements) - 1; 1085 1086 switch (ir->operation) { 1087 case ir_unop_logic_not: 1088 /* Note that BRW_OPCODE_NOT is not appropriate here, since it is 1089 * ones complement of the whole register, not just bit 0. 1090 */ 1091 emit(XOR(result_dst, op[0], src_reg(1))); 1092 break; 1093 case ir_unop_neg: 1094 op[0].negate = !op[0].negate; 1095 this->result = op[0]; 1096 break; 1097 case ir_unop_abs: 1098 op[0].abs = true; 1099 op[0].negate = false; 1100 this->result = op[0]; 1101 break; 1102 1103 case ir_unop_sign: 1104 emit(MOV(result_dst, src_reg(0.0f))); 1105 1106 emit(CMP(dst_null_d(), op[0], src_reg(0.0f), BRW_CONDITIONAL_G)); 1107 inst = emit(MOV(result_dst, src_reg(1.0f))); 1108 inst->predicate = BRW_PREDICATE_NORMAL; 1109 1110 emit(CMP(dst_null_d(), op[0], src_reg(0.0f), BRW_CONDITIONAL_L)); 1111 inst = emit(MOV(result_dst, src_reg(-1.0f))); 1112 inst->predicate = BRW_PREDICATE_NORMAL; 1113 1114 break; 1115 1116 case ir_unop_rcp: 1117 emit_math(SHADER_OPCODE_RCP, result_dst, op[0]); 1118 break; 1119 1120 case ir_unop_exp2: 1121 emit_math(SHADER_OPCODE_EXP2, result_dst, op[0]); 1122 break; 1123 case ir_unop_log2: 1124 emit_math(SHADER_OPCODE_LOG2, result_dst, op[0]); 1125 break; 1126 case ir_unop_exp: 1127 case ir_unop_log: 1128 assert(!"not reached: should be handled by ir_explog_to_explog2"); 1129 break; 1130 case ir_unop_sin: 1131 case ir_unop_sin_reduced: 1132 emit_math(SHADER_OPCODE_SIN, result_dst, op[0]); 1133 break; 1134 case ir_unop_cos: 1135 case ir_unop_cos_reduced: 1136 emit_math(SHADER_OPCODE_COS, result_dst, op[0]); 1137 break; 1138 1139 case ir_unop_dFdx: 1140 case ir_unop_dFdy: 1141 assert(!"derivatives not valid in vertex shader"); 1142 break; 1143 1144 case ir_unop_noise: 1145 assert(!"not reached: should be handled by lower_noise"); 1146 break; 1147 1148 case ir_binop_add: 1149 emit(ADD(result_dst, op[0], op[1])); 1150 break; 1151 case ir_binop_sub: 1152 assert(!"not reached: should be handled by ir_sub_to_add_neg"); 1153 break; 1154 1155 case ir_binop_mul: 1156 if (ir->type->is_integer()) { 1157 /* For integer multiplication, the MUL uses the low 16 bits 1158 * of one of the operands (src0 on gen6, src1 on gen7). The 1159 * MACH accumulates in the contribution of the upper 16 bits 1160 * of that operand. 1161 * 1162 * FINISHME: Emit just the MUL if we know an operand is small 1163 * enough. 1164 */ 1165 struct brw_reg acc = retype(brw_acc_reg(), BRW_REGISTER_TYPE_D); 1166 1167 emit(MUL(acc, op[0], op[1])); 1168 emit(MACH(dst_null_d(), op[0], op[1])); 1169 emit(MOV(result_dst, src_reg(acc))); 1170 } else { 1171 emit(MUL(result_dst, op[0], op[1])); 1172 } 1173 break; 1174 case ir_binop_div: 1175 /* Floating point should be lowered by DIV_TO_MUL_RCP in the compiler. */ 1176 assert(ir->type->is_integer()); 1177 emit_math(SHADER_OPCODE_INT_QUOTIENT, result_dst, op[0], op[1]); 1178 break; 1179 case ir_binop_mod: 1180 /* Floating point should be lowered by MOD_TO_FRACT in the compiler. */ 1181 assert(ir->type->is_integer()); 1182 emit_math(SHADER_OPCODE_INT_REMAINDER, result_dst, op[0], op[1]); 1183 break; 1184 1185 case ir_binop_less: 1186 case ir_binop_greater: 1187 case ir_binop_lequal: 1188 case ir_binop_gequal: 1189 case ir_binop_equal: 1190 case ir_binop_nequal: { 1191 emit(CMP(result_dst, op[0], op[1], 1192 brw_conditional_for_comparison(ir->operation))); 1193 emit(AND(result_dst, result_src, src_reg(0x1))); 1194 break; 1195 } 1196 1197 case ir_binop_all_equal: 1198 /* "==" operator producing a scalar boolean. */ 1199 if (ir->operands[0]->type->is_vector() || 1200 ir->operands[1]->type->is_vector()) { 1201 emit(CMP(dst_null_d(), op[0], op[1], BRW_CONDITIONAL_Z)); 1202 emit(MOV(result_dst, src_reg(0))); 1203 inst = emit(MOV(result_dst, src_reg(1))); 1204 inst->predicate = BRW_PREDICATE_ALIGN16_ALL4H; 1205 } else { 1206 emit(CMP(result_dst, op[0], op[1], BRW_CONDITIONAL_Z)); 1207 emit(AND(result_dst, result_src, src_reg(0x1))); 1208 } 1209 break; 1210 case ir_binop_any_nequal: 1211 /* "!=" operator producing a scalar boolean. */ 1212 if (ir->operands[0]->type->is_vector() || 1213 ir->operands[1]->type->is_vector()) { 1214 emit(CMP(dst_null_d(), op[0], op[1], BRW_CONDITIONAL_NZ)); 1215 1216 emit(MOV(result_dst, src_reg(0))); 1217 inst = emit(MOV(result_dst, src_reg(1))); 1218 inst->predicate = BRW_PREDICATE_ALIGN16_ANY4H; 1219 } else { 1220 emit(CMP(result_dst, op[0], op[1], BRW_CONDITIONAL_NZ)); 1221 emit(AND(result_dst, result_src, src_reg(0x1))); 1222 } 1223 break; 1224 1225 case ir_unop_any: 1226 emit(CMP(dst_null_d(), op[0], src_reg(0), BRW_CONDITIONAL_NZ)); 1227 emit(MOV(result_dst, src_reg(0))); 1228 1229 inst = emit(MOV(result_dst, src_reg(1))); 1230 inst->predicate = BRW_PREDICATE_ALIGN16_ANY4H; 1231 break; 1232 1233 case ir_binop_logic_xor: 1234 emit(XOR(result_dst, op[0], op[1])); 1235 break; 1236 1237 case ir_binop_logic_or: 1238 emit(OR(result_dst, op[0], op[1])); 1239 break; 1240 1241 case ir_binop_logic_and: 1242 emit(AND(result_dst, op[0], op[1])); 1243 break; 1244 1245 case ir_binop_dot: 1246 assert(ir->operands[0]->type->is_vector()); 1247 assert(ir->operands[0]->type == ir->operands[1]->type); 1248 emit_dp(result_dst, op[0], op[1], ir->operands[0]->type->vector_elements); 1249 break; 1250 1251 case ir_unop_sqrt: 1252 emit_math(SHADER_OPCODE_SQRT, result_dst, op[0]); 1253 break; 1254 case ir_unop_rsq: 1255 emit_math(SHADER_OPCODE_RSQ, result_dst, op[0]); 1256 break; 1257 case ir_unop_i2f: 1258 case ir_unop_i2u: 1259 case ir_unop_u2i: 1260 case ir_unop_u2f: 1261 case ir_unop_b2f: 1262 case ir_unop_b2i: 1263 case ir_unop_f2i: 1264 emit(MOV(result_dst, op[0])); 1265 break; 1266 case ir_unop_f2b: 1267 case ir_unop_i2b: { 1268 emit(CMP(result_dst, op[0], src_reg(0.0f), BRW_CONDITIONAL_NZ)); 1269 emit(AND(result_dst, result_src, src_reg(1))); 1270 break; 1271 } 1272 1273 case ir_unop_trunc: 1274 emit(RNDZ(result_dst, op[0])); 1275 break; 1276 case ir_unop_ceil: 1277 op[0].negate = !op[0].negate; 1278 inst = emit(RNDD(result_dst, op[0])); 1279 this->result.negate = true; 1280 break; 1281 case ir_unop_floor: 1282 inst = emit(RNDD(result_dst, op[0])); 1283 break; 1284 case ir_unop_fract: 1285 inst = emit(FRC(result_dst, op[0])); 1286 break; 1287 case ir_unop_round_even: 1288 emit(RNDE(result_dst, op[0])); 1289 break; 1290 1291 case ir_binop_min: 1292 if (intel->gen >= 6) { 1293 inst = emit(BRW_OPCODE_SEL, result_dst, op[0], op[1]); 1294 inst->conditional_mod = BRW_CONDITIONAL_L; 1295 } else { 1296 emit(CMP(result_dst, op[0], op[1], BRW_CONDITIONAL_L)); 1297 1298 inst = emit(BRW_OPCODE_SEL, result_dst, op[0], op[1]); 1299 inst->predicate = BRW_PREDICATE_NORMAL; 1300 } 1301 break; 1302 case ir_binop_max: 1303 if (intel->gen >= 6) { 1304 inst = emit(BRW_OPCODE_SEL, result_dst, op[0], op[1]); 1305 inst->conditional_mod = BRW_CONDITIONAL_G; 1306 } else { 1307 emit(CMP(result_dst, op[0], op[1], BRW_CONDITIONAL_G)); 1308 1309 inst = emit(BRW_OPCODE_SEL, result_dst, op[0], op[1]); 1310 inst->predicate = BRW_PREDICATE_NORMAL; 1311 } 1312 break; 1313 1314 case ir_binop_pow: 1315 emit_math(SHADER_OPCODE_POW, result_dst, op[0], op[1]); 1316 break; 1317 1318 case ir_unop_bit_not: 1319 inst = emit(NOT(result_dst, op[0])); 1320 break; 1321 case ir_binop_bit_and: 1322 inst = emit(AND(result_dst, op[0], op[1])); 1323 break; 1324 case ir_binop_bit_xor: 1325 inst = emit(XOR(result_dst, op[0], op[1])); 1326 break; 1327 case ir_binop_bit_or: 1328 inst = emit(OR(result_dst, op[0], op[1])); 1329 break; 1330 1331 case ir_binop_lshift: 1332 inst = emit(BRW_OPCODE_SHL, result_dst, op[0], op[1]); 1333 break; 1334 1335 case ir_binop_rshift: 1336 if (ir->type->base_type == GLSL_TYPE_INT) 1337 inst = emit(BRW_OPCODE_ASR, result_dst, op[0], op[1]); 1338 else 1339 inst = emit(BRW_OPCODE_SHR, result_dst, op[0], op[1]); 1340 break; 1341 1342 case ir_quadop_vector: 1343 assert(!"not reached: should be handled by lower_quadop_vector"); 1344 break; 1345 } 1346} 1347 1348 1349void 1350vec4_visitor::visit(ir_swizzle *ir) 1351{ 1352 src_reg src; 1353 int i = 0; 1354 int swizzle[4]; 1355 1356 /* Note that this is only swizzles in expressions, not those on the left 1357 * hand side of an assignment, which do write masking. See ir_assignment 1358 * for that. 1359 */ 1360 1361 ir->val->accept(this); 1362 src = this->result; 1363 assert(src.file != BAD_FILE); 1364 1365 for (i = 0; i < ir->type->vector_elements; i++) { 1366 switch (i) { 1367 case 0: 1368 swizzle[i] = BRW_GET_SWZ(src.swizzle, ir->mask.x); 1369 break; 1370 case 1: 1371 swizzle[i] = BRW_GET_SWZ(src.swizzle, ir->mask.y); 1372 break; 1373 case 2: 1374 swizzle[i] = BRW_GET_SWZ(src.swizzle, ir->mask.z); 1375 break; 1376 case 3: 1377 swizzle[i] = BRW_GET_SWZ(src.swizzle, ir->mask.w); 1378 break; 1379 } 1380 } 1381 for (; i < 4; i++) { 1382 /* Replicate the last channel out. */ 1383 swizzle[i] = swizzle[ir->type->vector_elements - 1]; 1384 } 1385 1386 src.swizzle = BRW_SWIZZLE4(swizzle[0], swizzle[1], swizzle[2], swizzle[3]); 1387 1388 this->result = src; 1389} 1390 1391void 1392vec4_visitor::visit(ir_dereference_variable *ir) 1393{ 1394 const struct glsl_type *type = ir->type; 1395 dst_reg *reg = variable_storage(ir->var); 1396 1397 if (!reg) { 1398 fail("Failed to find variable storage for %s\n", ir->var->name); 1399 this->result = src_reg(brw_null_reg()); 1400 return; 1401 } 1402 1403 this->result = src_reg(*reg); 1404 1405 if (type->is_scalar() || type->is_vector() || type->is_matrix()) 1406 this->result.swizzle = swizzle_for_size(type->vector_elements); 1407} 1408 1409void 1410vec4_visitor::visit(ir_dereference_array *ir) 1411{ 1412 ir_constant *constant_index; 1413 src_reg src; 1414 int element_size = type_size(ir->type); 1415 1416 constant_index = ir->array_index->constant_expression_value(); 1417 1418 ir->array->accept(this); 1419 src = this->result; 1420 1421 if (constant_index) { 1422 src.reg_offset += constant_index->value.i[0] * element_size; 1423 } else { 1424 /* Variable index array dereference. It eats the "vec4" of the 1425 * base of the array and an index that offsets the Mesa register 1426 * index. 1427 */ 1428 ir->array_index->accept(this); 1429 1430 src_reg index_reg; 1431 1432 if (element_size == 1) { 1433 index_reg = this->result; 1434 } else { 1435 index_reg = src_reg(this, glsl_type::int_type); 1436 1437 emit(MUL(dst_reg(index_reg), this->result, src_reg(element_size))); 1438 } 1439 1440 if (src.reladdr) { 1441 src_reg temp = src_reg(this, glsl_type::int_type); 1442 1443 emit(ADD(dst_reg(temp), *src.reladdr, index_reg)); 1444 1445 index_reg = temp; 1446 } 1447 1448 src.reladdr = ralloc(mem_ctx, src_reg); 1449 memcpy(src.reladdr, &index_reg, sizeof(index_reg)); 1450 } 1451 1452 /* If the type is smaller than a vec4, replicate the last channel out. */ 1453 if (ir->type->is_scalar() || ir->type->is_vector()) 1454 src.swizzle = swizzle_for_size(ir->type->vector_elements); 1455 else 1456 src.swizzle = BRW_SWIZZLE_NOOP; 1457 src.type = brw_type_for_base_type(ir->type); 1458 1459 this->result = src; 1460} 1461 1462void 1463vec4_visitor::visit(ir_dereference_record *ir) 1464{ 1465 unsigned int i; 1466 const glsl_type *struct_type = ir->record->type; 1467 int offset = 0; 1468 1469 ir->record->accept(this); 1470 1471 for (i = 0; i < struct_type->length; i++) { 1472 if (strcmp(struct_type->fields.structure[i].name, ir->field) == 0) 1473 break; 1474 offset += type_size(struct_type->fields.structure[i].type); 1475 } 1476 1477 /* If the type is smaller than a vec4, replicate the last channel out. */ 1478 if (ir->type->is_scalar() || ir->type->is_vector()) 1479 this->result.swizzle = swizzle_for_size(ir->type->vector_elements); 1480 else 1481 this->result.swizzle = BRW_SWIZZLE_NOOP; 1482 this->result.type = brw_type_for_base_type(ir->type); 1483 1484 this->result.reg_offset += offset; 1485} 1486 1487/** 1488 * We want to be careful in assignment setup to hit the actual storage 1489 * instead of potentially using a temporary like we might with the 1490 * ir_dereference handler. 1491 */ 1492static dst_reg 1493get_assignment_lhs(ir_dereference *ir, vec4_visitor *v) 1494{ 1495 /* The LHS must be a dereference. If the LHS is a variable indexed array 1496 * access of a vector, it must be separated into a series conditional moves 1497 * before reaching this point (see ir_vec_index_to_cond_assign). 1498 */ 1499 assert(ir->as_dereference()); 1500 ir_dereference_array *deref_array = ir->as_dereference_array(); 1501 if (deref_array) { 1502 assert(!deref_array->array->type->is_vector()); 1503 } 1504 1505 /* Use the rvalue deref handler for the most part. We'll ignore 1506 * swizzles in it and write swizzles using writemask, though. 1507 */ 1508 ir->accept(v); 1509 return dst_reg(v->result); 1510} 1511 1512void 1513vec4_visitor::emit_block_move(dst_reg *dst, src_reg *src, 1514 const struct glsl_type *type, uint32_t predicate) 1515{ 1516 if (type->base_type == GLSL_TYPE_STRUCT) { 1517 for (unsigned int i = 0; i < type->length; i++) { 1518 emit_block_move(dst, src, type->fields.structure[i].type, predicate); 1519 } 1520 return; 1521 } 1522 1523 if (type->is_array()) { 1524 for (unsigned int i = 0; i < type->length; i++) { 1525 emit_block_move(dst, src, type->fields.array, predicate); 1526 } 1527 return; 1528 } 1529 1530 if (type->is_matrix()) { 1531 const struct glsl_type *vec_type; 1532 1533 vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT, 1534 type->vector_elements, 1); 1535 1536 for (int i = 0; i < type->matrix_columns; i++) { 1537 emit_block_move(dst, src, vec_type, predicate); 1538 } 1539 return; 1540 } 1541 1542 assert(type->is_scalar() || type->is_vector()); 1543 1544 dst->type = brw_type_for_base_type(type); 1545 src->type = dst->type; 1546 1547 dst->writemask = (1 << type->vector_elements) - 1; 1548 1549 src->swizzle = swizzle_for_size(type->vector_elements); 1550 1551 vec4_instruction *inst = emit(MOV(*dst, *src)); 1552 inst->predicate = predicate; 1553 1554 dst->reg_offset++; 1555 src->reg_offset++; 1556} 1557 1558 1559/* If the RHS processing resulted in an instruction generating a 1560 * temporary value, and it would be easy to rewrite the instruction to 1561 * generate its result right into the LHS instead, do so. This ends 1562 * up reliably removing instructions where it can be tricky to do so 1563 * later without real UD chain information. 1564 */ 1565bool 1566vec4_visitor::try_rewrite_rhs_to_dst(ir_assignment *ir, 1567 dst_reg dst, 1568 src_reg src, 1569 vec4_instruction *pre_rhs_inst, 1570 vec4_instruction *last_rhs_inst) 1571{ 1572 /* This could be supported, but it would take more smarts. */ 1573 if (ir->condition) 1574 return false; 1575 1576 if (pre_rhs_inst == last_rhs_inst) 1577 return false; /* No instructions generated to work with. */ 1578 1579 /* Make sure the last instruction generated our source reg. */ 1580 if (src.file != GRF || 1581 src.file != last_rhs_inst->dst.file || 1582 src.reg != last_rhs_inst->dst.reg || 1583 src.reg_offset != last_rhs_inst->dst.reg_offset || 1584 src.reladdr || 1585 src.abs || 1586 src.negate || 1587 last_rhs_inst->predicate != BRW_PREDICATE_NONE) 1588 return false; 1589 1590 /* Check that that last instruction fully initialized the channels 1591 * we want to use, in the order we want to use them. We could 1592 * potentially reswizzle the operands of many instructions so that 1593 * we could handle out of order channels, but don't yet. 1594 */ 1595 1596 for (unsigned i = 0; i < 4; i++) { 1597 if (dst.writemask & (1 << i)) { 1598 if (!(last_rhs_inst->dst.writemask & (1 << i))) 1599 return false; 1600 1601 if (BRW_GET_SWZ(src.swizzle, i) != i) 1602 return false; 1603 } 1604 } 1605 1606 /* Success! Rewrite the instruction. */ 1607 last_rhs_inst->dst.file = dst.file; 1608 last_rhs_inst->dst.reg = dst.reg; 1609 last_rhs_inst->dst.reg_offset = dst.reg_offset; 1610 last_rhs_inst->dst.reladdr = dst.reladdr; 1611 last_rhs_inst->dst.writemask &= dst.writemask; 1612 1613 return true; 1614} 1615 1616void 1617vec4_visitor::visit(ir_assignment *ir) 1618{ 1619 dst_reg dst = get_assignment_lhs(ir->lhs, this); 1620 uint32_t predicate = BRW_PREDICATE_NONE; 1621 1622 if (!ir->lhs->type->is_scalar() && 1623 !ir->lhs->type->is_vector()) { 1624 ir->rhs->accept(this); 1625 src_reg src = this->result; 1626 1627 if (ir->condition) { 1628 emit_bool_to_cond_code(ir->condition, &predicate); 1629 } 1630 1631 /* emit_block_move doesn't account for swizzles in the source register. 1632 * This should be ok, since the source register is a structure or an 1633 * array, and those can't be swizzled. But double-check to be sure. 1634 */ 1635 assert(src.swizzle == 1636 (ir->rhs->type->is_matrix() 1637 ? swizzle_for_size(ir->rhs->type->vector_elements) 1638 : BRW_SWIZZLE_NOOP)); 1639 1640 emit_block_move(&dst, &src, ir->rhs->type, predicate); 1641 return; 1642 } 1643 1644 /* Now we're down to just a scalar/vector with writemasks. */ 1645 int i; 1646 1647 vec4_instruction *pre_rhs_inst, *last_rhs_inst; 1648 pre_rhs_inst = (vec4_instruction *)this->instructions.get_tail(); 1649 1650 ir->rhs->accept(this); 1651 1652 last_rhs_inst = (vec4_instruction *)this->instructions.get_tail(); 1653 1654 src_reg src = this->result; 1655 1656 int swizzles[4]; 1657 int first_enabled_chan = 0; 1658 int src_chan = 0; 1659 1660 assert(ir->lhs->type->is_vector() || 1661 ir->lhs->type->is_scalar()); 1662 dst.writemask = ir->write_mask; 1663 1664 for (int i = 0; i < 4; i++) { 1665 if (dst.writemask & (1 << i)) { 1666 first_enabled_chan = BRW_GET_SWZ(src.swizzle, i); 1667 break; 1668 } 1669 } 1670 1671 /* Swizzle a small RHS vector into the channels being written. 1672 * 1673 * glsl ir treats write_mask as dictating how many channels are 1674 * present on the RHS while in our instructions we need to make 1675 * those channels appear in the slots of the vec4 they're written to. 1676 */ 1677 for (int i = 0; i < 4; i++) { 1678 if (dst.writemask & (1 << i)) 1679 swizzles[i] = BRW_GET_SWZ(src.swizzle, src_chan++); 1680 else 1681 swizzles[i] = first_enabled_chan; 1682 } 1683 src.swizzle = BRW_SWIZZLE4(swizzles[0], swizzles[1], 1684 swizzles[2], swizzles[3]); 1685 1686 if (try_rewrite_rhs_to_dst(ir, dst, src, pre_rhs_inst, last_rhs_inst)) { 1687 return; 1688 } 1689 1690 if (ir->condition) { 1691 emit_bool_to_cond_code(ir->condition, &predicate); 1692 } 1693 1694 for (i = 0; i < type_size(ir->lhs->type); i++) { 1695 vec4_instruction *inst = emit(MOV(dst, src)); 1696 inst->predicate = predicate; 1697 1698 dst.reg_offset++; 1699 src.reg_offset++; 1700 } 1701} 1702 1703void 1704vec4_visitor::emit_constant_values(dst_reg *dst, ir_constant *ir) 1705{ 1706 if (ir->type->base_type == GLSL_TYPE_STRUCT) { 1707 foreach_list(node, &ir->components) { 1708 ir_constant *field_value = (ir_constant *)node; 1709 1710 emit_constant_values(dst, field_value); 1711 } 1712 return; 1713 } 1714 1715 if (ir->type->is_array()) { 1716 for (unsigned int i = 0; i < ir->type->length; i++) { 1717 emit_constant_values(dst, ir->array_elements[i]); 1718 } 1719 return; 1720 } 1721 1722 if (ir->type->is_matrix()) { 1723 for (int i = 0; i < ir->type->matrix_columns; i++) { 1724 float *vec = &ir->value.f[i * ir->type->vector_elements]; 1725 1726 for (int j = 0; j < ir->type->vector_elements; j++) { 1727 dst->writemask = 1 << j; 1728 dst->type = BRW_REGISTER_TYPE_F; 1729 1730 emit(MOV(*dst, src_reg(vec[j]))); 1731 } 1732 dst->reg_offset++; 1733 } 1734 return; 1735 } 1736 1737 int remaining_writemask = (1 << ir->type->vector_elements) - 1; 1738 1739 for (int i = 0; i < ir->type->vector_elements; i++) { 1740 if (!(remaining_writemask & (1 << i))) 1741 continue; 1742 1743 dst->writemask = 1 << i; 1744 dst->type = brw_type_for_base_type(ir->type); 1745 1746 /* Find other components that match the one we're about to 1747 * write. Emits fewer instructions for things like vec4(0.5, 1748 * 1.5, 1.5, 1.5). 1749 */ 1750 for (int j = i + 1; j < ir->type->vector_elements; j++) { 1751 if (ir->type->base_type == GLSL_TYPE_BOOL) { 1752 if (ir->value.b[i] == ir->value.b[j]) 1753 dst->writemask |= (1 << j); 1754 } else { 1755 /* u, i, and f storage all line up, so no need for a 1756 * switch case for comparing each type. 1757 */ 1758 if (ir->value.u[i] == ir->value.u[j]) 1759 dst->writemask |= (1 << j); 1760 } 1761 } 1762 1763 switch (ir->type->base_type) { 1764 case GLSL_TYPE_FLOAT: 1765 emit(MOV(*dst, src_reg(ir->value.f[i]))); 1766 break; 1767 case GLSL_TYPE_INT: 1768 emit(MOV(*dst, src_reg(ir->value.i[i]))); 1769 break; 1770 case GLSL_TYPE_UINT: 1771 emit(MOV(*dst, src_reg(ir->value.u[i]))); 1772 break; 1773 case GLSL_TYPE_BOOL: 1774 emit(MOV(*dst, src_reg(ir->value.b[i]))); 1775 break; 1776 default: 1777 assert(!"Non-float/uint/int/bool constant"); 1778 break; 1779 } 1780 1781 remaining_writemask &= ~dst->writemask; 1782 } 1783 dst->reg_offset++; 1784} 1785 1786void 1787vec4_visitor::visit(ir_constant *ir) 1788{ 1789 dst_reg dst = dst_reg(this, ir->type); 1790 this->result = src_reg(dst); 1791 1792 emit_constant_values(&dst, ir); 1793} 1794 1795void 1796vec4_visitor::visit(ir_call *ir) 1797{ 1798 assert(!"not reached"); 1799} 1800 1801void 1802vec4_visitor::visit(ir_texture *ir) 1803{ 1804 int sampler = _mesa_get_sampler_uniform_value(ir->sampler, prog, &vp->Base); 1805 sampler = vp->Base.SamplerUnits[sampler]; 1806 1807 /* Should be lowered by do_lower_texture_projection */ 1808 assert(!ir->projector); 1809 1810 vec4_instruction *inst = NULL; 1811 switch (ir->op) { 1812 case ir_tex: 1813 case ir_txl: 1814 inst = new(mem_ctx) vec4_instruction(this, SHADER_OPCODE_TXL); 1815 break; 1816 case ir_txd: 1817 inst = new(mem_ctx) vec4_instruction(this, SHADER_OPCODE_TXD); 1818 break; 1819 case ir_txf: 1820 inst = new(mem_ctx) vec4_instruction(this, SHADER_OPCODE_TXF); 1821 break; 1822 case ir_txs: 1823 inst = new(mem_ctx) vec4_instruction(this, SHADER_OPCODE_TXS); 1824 break; 1825 case ir_txb: 1826 assert(!"TXB is not valid for vertex shaders."); 1827 } 1828 1829 /* Texel offsets go in the message header; Gen4 also requires headers. */ 1830 inst->header_present = ir->offset || intel->gen < 5; 1831 inst->base_mrf = 2; 1832 inst->mlen = inst->header_present + 1; /* always at least one */ 1833 inst->sampler = sampler; 1834 inst->dst = dst_reg(this, ir->type); 1835 inst->shadow_compare = ir->shadow_comparitor != NULL; 1836 1837 if (ir->offset != NULL) 1838 inst->texture_offset = brw_texture_offset(ir->offset->as_constant()); 1839 1840 /* MRF for the first parameter */ 1841 int param_base = inst->base_mrf + inst->header_present; 1842 1843 if (ir->op == ir_txs) { 1844 ir->lod_info.lod->accept(this); 1845 int writemask = intel->gen == 4 ? WRITEMASK_W : WRITEMASK_X; 1846 emit(MOV(dst_reg(MRF, param_base, ir->lod_info.lod->type, writemask), 1847 this->result)); 1848 } else { 1849 int i, coord_mask = 0, zero_mask = 0; 1850 /* Load the coordinate */ 1851 /* FINISHME: gl_clamp_mask and saturate */ 1852 for (i = 0; i < ir->coordinate->type->vector_elements; i++) 1853 coord_mask |= (1 << i); 1854 for (; i < 4; i++) 1855 zero_mask |= (1 << i); 1856 1857 ir->coordinate->accept(this); 1858 emit(MOV(dst_reg(MRF, param_base, ir->coordinate->type, coord_mask), 1859 this->result)); 1860 emit(MOV(dst_reg(MRF, param_base, ir->coordinate->type, zero_mask), 1861 src_reg(0))); 1862 /* Load the shadow comparitor */ 1863 if (ir->shadow_comparitor) { 1864 ir->shadow_comparitor->accept(this); 1865 emit(MOV(dst_reg(MRF, param_base + 1, ir->shadow_comparitor->type, 1866 WRITEMASK_X), 1867 this->result)); 1868 inst->mlen++; 1869 } 1870 1871 /* Load the LOD info */ 1872 if (ir->op == ir_txl) { 1873 int mrf, writemask; 1874 if (intel->gen >= 5) { 1875 mrf = param_base + 1; 1876 if (ir->shadow_comparitor) { 1877 writemask = WRITEMASK_Y; 1878 /* mlen already incremented */ 1879 } else { 1880 writemask = WRITEMASK_X; 1881 inst->mlen++; 1882 } 1883 } else /* intel->gen == 4 */ { 1884 mrf = param_base; 1885 writemask = WRITEMASK_Z; 1886 } 1887 ir->lod_info.lod->accept(this); 1888 emit(MOV(dst_reg(MRF, mrf, ir->lod_info.lod->type, writemask), 1889 this->result)); 1890 } else if (ir->op == ir_txf) { 1891 ir->lod_info.lod->accept(this); 1892 emit(MOV(dst_reg(MRF, param_base, ir->lod_info.lod->type, WRITEMASK_W), 1893 this->result)); 1894 } else if (ir->op == ir_txd) { 1895 const glsl_type *type = ir->lod_info.grad.dPdx->type; 1896 1897 ir->lod_info.grad.dPdx->accept(this); 1898 src_reg dPdx = this->result; 1899 ir->lod_info.grad.dPdy->accept(this); 1900 src_reg dPdy = this->result; 1901 1902 if (intel->gen >= 5) { 1903 dPdx.swizzle = BRW_SWIZZLE4(SWIZZLE_X,SWIZZLE_X,SWIZZLE_Y,SWIZZLE_Y); 1904 dPdy.swizzle = BRW_SWIZZLE4(SWIZZLE_X,SWIZZLE_X,SWIZZLE_Y,SWIZZLE_Y); 1905 emit(MOV(dst_reg(MRF, param_base + 1, type, WRITEMASK_XZ), dPdx)); 1906 emit(MOV(dst_reg(MRF, param_base + 1, type, WRITEMASK_YW), dPdy)); 1907 inst->mlen++; 1908 1909 if (ir->type->vector_elements == 3) { 1910 dPdx.swizzle = BRW_SWIZZLE_ZZZZ; 1911 dPdy.swizzle = BRW_SWIZZLE_ZZZZ; 1912 emit(MOV(dst_reg(MRF, param_base + 2, type, WRITEMASK_X), dPdx)); 1913 emit(MOV(dst_reg(MRF, param_base + 2, type, WRITEMASK_Y), dPdy)); 1914 inst->mlen++; 1915 } 1916 } else /* intel->gen == 4 */ { 1917 emit(MOV(dst_reg(MRF, param_base + 1, type, WRITEMASK_XYZ), dPdx)); 1918 emit(MOV(dst_reg(MRF, param_base + 2, type, WRITEMASK_XYZ), dPdy)); 1919 inst->mlen += 2; 1920 } 1921 } 1922 } 1923 1924 emit(inst); 1925 1926 swizzle_result(ir, src_reg(inst->dst), sampler); 1927} 1928 1929void 1930vec4_visitor::swizzle_result(ir_texture *ir, src_reg orig_val, int sampler) 1931{ 1932 this->result = orig_val; 1933 1934 int s = c->key.tex.swizzles[sampler]; 1935 1936 if (ir->op == ir_txs || ir->type == glsl_type::float_type 1937 || s == SWIZZLE_NOOP) 1938 return; 1939 1940 int zero_mask = 0, one_mask = 0, copy_mask = 0; 1941 int swizzle[4]; 1942 1943 for (int i = 0; i < 4; i++) { 1944 switch (GET_SWZ(s, i)) { 1945 case SWIZZLE_ZERO: 1946 zero_mask |= (1 << i); 1947 break; 1948 case SWIZZLE_ONE: 1949 one_mask |= (1 << i); 1950 break; 1951 default: 1952 copy_mask |= (1 << i); 1953 swizzle[i] = GET_SWZ(s, i); 1954 break; 1955 } 1956 } 1957 1958 this->result = src_reg(this, ir->type); 1959 dst_reg swizzled_result(this->result); 1960 1961 if (copy_mask) { 1962 orig_val.swizzle = BRW_SWIZZLE4(swizzle[0], swizzle[1], swizzle[2], swizzle[3]); 1963 swizzled_result.writemask = copy_mask; 1964 emit(MOV(swizzled_result, orig_val)); 1965 } 1966 1967 if (zero_mask) { 1968 swizzled_result.writemask = zero_mask; 1969 emit(MOV(swizzled_result, src_reg(0.0f))); 1970 } 1971 1972 if (one_mask) { 1973 swizzled_result.writemask = one_mask; 1974 emit(MOV(swizzled_result, src_reg(1.0f))); 1975 } 1976} 1977 1978void 1979vec4_visitor::visit(ir_return *ir) 1980{ 1981 assert(!"not reached"); 1982} 1983 1984void 1985vec4_visitor::visit(ir_discard *ir) 1986{ 1987 assert(!"not reached"); 1988} 1989 1990void 1991vec4_visitor::visit(ir_if *ir) 1992{ 1993 /* Don't point the annotation at the if statement, because then it plus 1994 * the then and else blocks get printed. 1995 */ 1996 this->base_ir = ir->condition; 1997 1998 if (intel->gen == 6) { 1999 emit_if_gen6(ir); 2000 } else { 2001 uint32_t predicate; 2002 emit_bool_to_cond_code(ir->condition, &predicate); 2003 emit(IF(predicate)); 2004 } 2005 2006 visit_instructions(&ir->then_instructions); 2007 2008 if (!ir->else_instructions.is_empty()) { 2009 this->base_ir = ir->condition; 2010 emit(BRW_OPCODE_ELSE); 2011 2012 visit_instructions(&ir->else_instructions); 2013 } 2014 2015 this->base_ir = ir->condition; 2016 emit(BRW_OPCODE_ENDIF); 2017} 2018 2019void 2020vec4_visitor::emit_ndc_computation() 2021{ 2022 /* Get the position */ 2023 src_reg pos = src_reg(output_reg[VERT_RESULT_HPOS]); 2024 2025 /* Build ndc coords, which are (x/w, y/w, z/w, 1/w) */ 2026 dst_reg ndc = dst_reg(this, glsl_type::vec4_type); 2027 output_reg[BRW_VERT_RESULT_NDC] = ndc; 2028 2029 current_annotation = "NDC"; 2030 dst_reg ndc_w = ndc; 2031 ndc_w.writemask = WRITEMASK_W; 2032 src_reg pos_w = pos; 2033 pos_w.swizzle = BRW_SWIZZLE4(SWIZZLE_W, SWIZZLE_W, SWIZZLE_W, SWIZZLE_W); 2034 emit_math(SHADER_OPCODE_RCP, ndc_w, pos_w); 2035 2036 dst_reg ndc_xyz = ndc; 2037 ndc_xyz.writemask = WRITEMASK_XYZ; 2038 2039 emit(MUL(ndc_xyz, pos, src_reg(ndc_w))); 2040} 2041 2042void 2043vec4_visitor::emit_psiz_and_flags(struct brw_reg reg) 2044{ 2045 if (intel->gen < 6 && 2046 ((c->prog_data.outputs_written & BITFIELD64_BIT(VERT_RESULT_PSIZ)) || 2047 c->key.userclip_active || brw->has_negative_rhw_bug)) { 2048 dst_reg header1 = dst_reg(this, glsl_type::uvec4_type); 2049 dst_reg header1_w = header1; 2050 header1_w.writemask = WRITEMASK_W; 2051 GLuint i; 2052 2053 emit(MOV(header1, 0u)); 2054 2055 if (c->prog_data.outputs_written & BITFIELD64_BIT(VERT_RESULT_PSIZ)) { 2056 src_reg psiz = src_reg(output_reg[VERT_RESULT_PSIZ]); 2057 2058 current_annotation = "Point size"; 2059 emit(MUL(header1_w, psiz, src_reg((float)(1 << 11)))); 2060 emit(AND(header1_w, src_reg(header1_w), 0x7ff << 8)); 2061 } 2062 2063 current_annotation = "Clipping flags"; 2064 for (i = 0; i < c->key.nr_userclip_plane_consts; i++) { 2065 vec4_instruction *inst; 2066 2067 inst = emit(DP4(dst_null_f(), src_reg(output_reg[VERT_RESULT_HPOS]), 2068 src_reg(this->userplane[i]))); 2069 inst->conditional_mod = BRW_CONDITIONAL_L; 2070 2071 inst = emit(OR(header1_w, src_reg(header1_w), 1u << i)); 2072 inst->predicate = BRW_PREDICATE_NORMAL; 2073 } 2074 2075 /* i965 clipping workaround: 2076 * 1) Test for -ve rhw 2077 * 2) If set, 2078 * set ndc = (0,0,0,0) 2079 * set ucp[6] = 1 2080 * 2081 * Later, clipping will detect ucp[6] and ensure the primitive is 2082 * clipped against all fixed planes. 2083 */ 2084 if (brw->has_negative_rhw_bug) { 2085#if 0 2086 /* FINISHME */ 2087 brw_CMP(p, 2088 vec8(brw_null_reg()), 2089 BRW_CONDITIONAL_L, 2090 brw_swizzle1(output_reg[BRW_VERT_RESULT_NDC], 3), 2091 brw_imm_f(0)); 2092 2093 brw_OR(p, brw_writemask(header1, WRITEMASK_W), header1, brw_imm_ud(1<<6)); 2094 brw_MOV(p, output_reg[BRW_VERT_RESULT_NDC], brw_imm_f(0)); 2095 brw_set_predicate_control(p, BRW_PREDICATE_NONE); 2096#endif 2097 } 2098 2099 emit(MOV(retype(reg, BRW_REGISTER_TYPE_UD), src_reg(header1))); 2100 } else if (intel->gen < 6) { 2101 emit(MOV(retype(reg, BRW_REGISTER_TYPE_UD), 0u)); 2102 } else { 2103 emit(MOV(retype(reg, BRW_REGISTER_TYPE_D), src_reg(0))); 2104 if (c->prog_data.outputs_written & BITFIELD64_BIT(VERT_RESULT_PSIZ)) { 2105 emit(MOV(brw_writemask(reg, WRITEMASK_W), 2106 src_reg(output_reg[VERT_RESULT_PSIZ]))); 2107 } 2108 } 2109} 2110 2111void 2112vec4_visitor::emit_clip_distances(struct brw_reg reg, int offset) 2113{ 2114 if (intel->gen < 6) { 2115 /* Clip distance slots are set aside in gen5, but they are not used. It 2116 * is not clear whether we actually need to set aside space for them, 2117 * but the performance cost is negligible. 2118 */ 2119 return; 2120 } 2121 2122 /* From the GLSL 1.30 spec, section 7.1 (Vertex Shader Special Variables): 2123 * 2124 * "If a linked set of shaders forming the vertex stage contains no 2125 * static write to gl_ClipVertex or gl_ClipDistance, but the 2126 * application has requested clipping against user clip planes through 2127 * the API, then the coordinate written to gl_Position is used for 2128 * comparison against the user clip planes." 2129 * 2130 * This function is only called if the shader didn't write to 2131 * gl_ClipDistance. Accordingly, we use gl_ClipVertex to perform clipping 2132 * if the user wrote to it; otherwise we use gl_Position. 2133 */ 2134 gl_vert_result clip_vertex = VERT_RESULT_CLIP_VERTEX; 2135 if (!(c->prog_data.outputs_written 2136 & BITFIELD64_BIT(VERT_RESULT_CLIP_VERTEX))) { 2137 clip_vertex = VERT_RESULT_HPOS; 2138 } 2139 2140 for (int i = 0; i + offset < c->key.nr_userclip_plane_consts && i < 4; 2141 ++i) { 2142 emit(DP4(dst_reg(brw_writemask(reg, 1 << i)), 2143 src_reg(output_reg[clip_vertex]), 2144 src_reg(this->userplane[i + offset]))); 2145 } 2146} 2147 2148void 2149vec4_visitor::emit_generic_urb_slot(dst_reg reg, int vert_result) 2150{ 2151 assert (vert_result < VERT_RESULT_MAX); 2152 reg.type = output_reg[vert_result].type; 2153 current_annotation = output_reg_annotation[vert_result]; 2154 /* Copy the register, saturating if necessary */ 2155 vec4_instruction *inst = emit(MOV(reg, 2156 src_reg(output_reg[vert_result]))); 2157 if ((vert_result == VERT_RESULT_COL0 || 2158 vert_result == VERT_RESULT_COL1 || 2159 vert_result == VERT_RESULT_BFC0 || 2160 vert_result == VERT_RESULT_BFC1) && 2161 c->key.clamp_vertex_color) { 2162 inst->saturate = true; 2163 } 2164} 2165 2166void 2167vec4_visitor::emit_urb_slot(int mrf, int vert_result) 2168{ 2169 struct brw_reg hw_reg = brw_message_reg(mrf); 2170 dst_reg reg = dst_reg(MRF, mrf); 2171 reg.type = BRW_REGISTER_TYPE_F; 2172 2173 switch (vert_result) { 2174 case VERT_RESULT_PSIZ: 2175 /* PSIZ is always in slot 0, and is coupled with other flags. */ 2176 current_annotation = "indices, point width, clip flags"; 2177 emit_psiz_and_flags(hw_reg); 2178 break; 2179 case BRW_VERT_RESULT_NDC: 2180 current_annotation = "NDC"; 2181 emit(MOV(reg, src_reg(output_reg[BRW_VERT_RESULT_NDC]))); 2182 break; 2183 case BRW_VERT_RESULT_HPOS_DUPLICATE: 2184 case VERT_RESULT_HPOS: 2185 current_annotation = "gl_Position"; 2186 emit(MOV(reg, src_reg(output_reg[VERT_RESULT_HPOS]))); 2187 break; 2188 case VERT_RESULT_CLIP_DIST0: 2189 case VERT_RESULT_CLIP_DIST1: 2190 if (this->c->key.uses_clip_distance) { 2191 emit_generic_urb_slot(reg, vert_result); 2192 } else { 2193 current_annotation = "user clip distances"; 2194 emit_clip_distances(hw_reg, (vert_result - VERT_RESULT_CLIP_DIST0) * 4); 2195 } 2196 break; 2197 case BRW_VERT_RESULT_PAD: 2198 /* No need to write to this slot */ 2199 break; 2200 default: 2201 emit_generic_urb_slot(reg, vert_result); 2202 break; 2203 } 2204} 2205 2206static int 2207align_interleaved_urb_mlen(struct brw_context *brw, int mlen) 2208{ 2209 struct intel_context *intel = &brw->intel; 2210 2211 if (intel->gen >= 6) { 2212 /* URB data written (does not include the message header reg) must 2213 * be a multiple of 256 bits, or 2 VS registers. See vol5c.5, 2214 * section 5.4.3.2.2: URB_INTERLEAVED. 2215 * 2216 * URB entries are allocated on a multiple of 1024 bits, so an 2217 * extra 128 bits written here to make the end align to 256 is 2218 * no problem. 2219 */ 2220 if ((mlen % 2) != 1) 2221 mlen++; 2222 } 2223 2224 return mlen; 2225} 2226 2227/** 2228 * Generates the VUE payload plus the 1 or 2 URB write instructions to 2229 * complete the VS thread. 2230 * 2231 * The VUE layout is documented in Volume 2a. 2232 */ 2233void 2234vec4_visitor::emit_urb_writes() 2235{ 2236 /* MRF 0 is reserved for the debugger, so start with message header 2237 * in MRF 1. 2238 */ 2239 int base_mrf = 1; 2240 int mrf = base_mrf; 2241 /* In the process of generating our URB write message contents, we 2242 * may need to unspill a register or load from an array. Those 2243 * reads would use MRFs 14-15. 2244 */ 2245 int max_usable_mrf = 13; 2246 2247 /* The following assertion verifies that max_usable_mrf causes an 2248 * even-numbered amount of URB write data, which will meet gen6's 2249 * requirements for length alignment. 2250 */ 2251 assert ((max_usable_mrf - base_mrf) % 2 == 0); 2252 2253 /* FINISHME: edgeflag */ 2254 2255 brw_compute_vue_map(&c->vue_map, intel, c->key.userclip_active, 2256 c->prog_data.outputs_written); 2257 2258 /* First mrf is the g0-based message header containing URB handles and such, 2259 * which is implied in VS_OPCODE_URB_WRITE. 2260 */ 2261 mrf++; 2262 2263 if (intel->gen < 6) { 2264 emit_ndc_computation(); 2265 } 2266 2267 /* Set up the VUE data for the first URB write */ 2268 int slot; 2269 for (slot = 0; slot < c->vue_map.num_slots; ++slot) { 2270 emit_urb_slot(mrf++, c->vue_map.slot_to_vert_result[slot]); 2271 2272 /* If this was max_usable_mrf, we can't fit anything more into this URB 2273 * WRITE. 2274 */ 2275 if (mrf > max_usable_mrf) { 2276 slot++; 2277 break; 2278 } 2279 } 2280 2281 current_annotation = "URB write"; 2282 vec4_instruction *inst = emit(VS_OPCODE_URB_WRITE); 2283 inst->base_mrf = base_mrf; 2284 inst->mlen = align_interleaved_urb_mlen(brw, mrf - base_mrf); 2285 inst->eot = (slot >= c->vue_map.num_slots); 2286 2287 /* Optional second URB write */ 2288 if (!inst->eot) { 2289 mrf = base_mrf + 1; 2290 2291 for (; slot < c->vue_map.num_slots; ++slot) { 2292 assert(mrf < max_usable_mrf); 2293 2294 emit_urb_slot(mrf++, c->vue_map.slot_to_vert_result[slot]); 2295 } 2296 2297 current_annotation = "URB write"; 2298 inst = emit(VS_OPCODE_URB_WRITE); 2299 inst->base_mrf = base_mrf; 2300 inst->mlen = align_interleaved_urb_mlen(brw, mrf - base_mrf); 2301 inst->eot = true; 2302 /* URB destination offset. In the previous write, we got MRFs 2303 * 2-13 minus the one header MRF, so 12 regs. URB offset is in 2304 * URB row increments, and each of our MRFs is half of one of 2305 * those, since we're doing interleaved writes. 2306 */ 2307 inst->offset = (max_usable_mrf - base_mrf) / 2; 2308 } 2309} 2310 2311src_reg 2312vec4_visitor::get_scratch_offset(vec4_instruction *inst, 2313 src_reg *reladdr, int reg_offset) 2314{ 2315 /* Because we store the values to scratch interleaved like our 2316 * vertex data, we need to scale the vec4 index by 2. 2317 */ 2318 int message_header_scale = 2; 2319 2320 /* Pre-gen6, the message header uses byte offsets instead of vec4 2321 * (16-byte) offset units. 2322 */ 2323 if (intel->gen < 6) 2324 message_header_scale *= 16; 2325 2326 if (reladdr) { 2327 src_reg index = src_reg(this, glsl_type::int_type); 2328 2329 emit_before(inst, ADD(dst_reg(index), *reladdr, src_reg(reg_offset))); 2330 emit_before(inst, MUL(dst_reg(index), 2331 index, src_reg(message_header_scale))); 2332 2333 return index; 2334 } else { 2335 return src_reg(reg_offset * message_header_scale); 2336 } 2337} 2338 2339src_reg 2340vec4_visitor::get_pull_constant_offset(vec4_instruction *inst, 2341 src_reg *reladdr, int reg_offset) 2342{ 2343 if (reladdr) { 2344 src_reg index = src_reg(this, glsl_type::int_type); 2345 2346 emit_before(inst, ADD(dst_reg(index), *reladdr, src_reg(reg_offset))); 2347 2348 /* Pre-gen6, the message header uses byte offsets instead of vec4 2349 * (16-byte) offset units. 2350 */ 2351 if (intel->gen < 6) { 2352 emit_before(inst, MUL(dst_reg(index), index, src_reg(16))); 2353 } 2354 2355 return index; 2356 } else { 2357 int message_header_scale = intel->gen < 6 ? 16 : 1; 2358 return src_reg(reg_offset * message_header_scale); 2359 } 2360} 2361 2362/** 2363 * Emits an instruction before @inst to load the value named by @orig_src 2364 * from scratch space at @base_offset to @temp. 2365 */ 2366void 2367vec4_visitor::emit_scratch_read(vec4_instruction *inst, 2368 dst_reg temp, src_reg orig_src, 2369 int base_offset) 2370{ 2371 int reg_offset = base_offset + orig_src.reg_offset; 2372 src_reg index = get_scratch_offset(inst, orig_src.reladdr, reg_offset); 2373 2374 emit_before(inst, SCRATCH_READ(temp, index)); 2375} 2376 2377/** 2378 * Emits an instruction after @inst to store the value to be written 2379 * to @orig_dst to scratch space at @base_offset, from @temp. 2380 */ 2381void 2382vec4_visitor::emit_scratch_write(vec4_instruction *inst, 2383 src_reg temp, dst_reg orig_dst, 2384 int base_offset) 2385{ 2386 int reg_offset = base_offset + orig_dst.reg_offset; 2387 src_reg index = get_scratch_offset(inst, orig_dst.reladdr, reg_offset); 2388 2389 dst_reg dst = dst_reg(brw_writemask(brw_vec8_grf(0, 0), 2390 orig_dst.writemask)); 2391 vec4_instruction *write = SCRATCH_WRITE(dst, temp, index); 2392 write->predicate = inst->predicate; 2393 write->ir = inst->ir; 2394 write->annotation = inst->annotation; 2395 inst->insert_after(write); 2396} 2397 2398/** 2399 * We can't generally support array access in GRF space, because a 2400 * single instruction's destination can only span 2 contiguous 2401 * registers. So, we send all GRF arrays that get variable index 2402 * access to scratch space. 2403 */ 2404void 2405vec4_visitor::move_grf_array_access_to_scratch() 2406{ 2407 int scratch_loc[this->virtual_grf_count]; 2408 2409 for (int i = 0; i < this->virtual_grf_count; i++) { 2410 scratch_loc[i] = -1; 2411 } 2412 2413 /* First, calculate the set of virtual GRFs that need to be punted 2414 * to scratch due to having any array access on them, and where in 2415 * scratch. 2416 */ 2417 foreach_list(node, &this->instructions) { 2418 vec4_instruction *inst = (vec4_instruction *)node; 2419 2420 if (inst->dst.file == GRF && inst->dst.reladdr && 2421 scratch_loc[inst->dst.reg] == -1) { 2422 scratch_loc[inst->dst.reg] = c->last_scratch; 2423 c->last_scratch += this->virtual_grf_sizes[inst->dst.reg] * 8 * 4; 2424 } 2425 2426 for (int i = 0 ; i < 3; i++) { 2427 src_reg *src = &inst->src[i]; 2428 2429 if (src->file == GRF && src->reladdr && 2430 scratch_loc[src->reg] == -1) { 2431 scratch_loc[src->reg] = c->last_scratch; 2432 c->last_scratch += this->virtual_grf_sizes[src->reg] * 8 * 4; 2433 } 2434 } 2435 } 2436 2437 /* Now, for anything that will be accessed through scratch, rewrite 2438 * it to load/store. Note that this is a _safe list walk, because 2439 * we may generate a new scratch_write instruction after the one 2440 * we're processing. 2441 */ 2442 foreach_list_safe(node, &this->instructions) { 2443 vec4_instruction *inst = (vec4_instruction *)node; 2444 2445 /* Set up the annotation tracking for new generated instructions. */ 2446 base_ir = inst->ir; 2447 current_annotation = inst->annotation; 2448 2449 if (inst->dst.file == GRF && scratch_loc[inst->dst.reg] != -1) { 2450 src_reg temp = src_reg(this, glsl_type::vec4_type); 2451 2452 emit_scratch_write(inst, temp, inst->dst, scratch_loc[inst->dst.reg]); 2453 2454 inst->dst.file = temp.file; 2455 inst->dst.reg = temp.reg; 2456 inst->dst.reg_offset = temp.reg_offset; 2457 inst->dst.reladdr = NULL; 2458 } 2459 2460 for (int i = 0 ; i < 3; i++) { 2461 if (inst->src[i].file != GRF || scratch_loc[inst->src[i].reg] == -1) 2462 continue; 2463 2464 dst_reg temp = dst_reg(this, glsl_type::vec4_type); 2465 2466 emit_scratch_read(inst, temp, inst->src[i], 2467 scratch_loc[inst->src[i].reg]); 2468 2469 inst->src[i].file = temp.file; 2470 inst->src[i].reg = temp.reg; 2471 inst->src[i].reg_offset = temp.reg_offset; 2472 inst->src[i].reladdr = NULL; 2473 } 2474 } 2475} 2476 2477/** 2478 * Emits an instruction before @inst to load the value named by @orig_src 2479 * from the pull constant buffer (surface) at @base_offset to @temp. 2480 */ 2481void 2482vec4_visitor::emit_pull_constant_load(vec4_instruction *inst, 2483 dst_reg temp, src_reg orig_src, 2484 int base_offset) 2485{ 2486 int reg_offset = base_offset + orig_src.reg_offset; 2487 src_reg index = get_pull_constant_offset(inst, orig_src.reladdr, reg_offset); 2488 vec4_instruction *load; 2489 2490 load = new(mem_ctx) vec4_instruction(this, VS_OPCODE_PULL_CONSTANT_LOAD, 2491 temp, index); 2492 load->base_mrf = 14; 2493 load->mlen = 1; 2494 emit_before(inst, load); 2495} 2496 2497/** 2498 * Implements array access of uniforms by inserting a 2499 * PULL_CONSTANT_LOAD instruction. 2500 * 2501 * Unlike temporary GRF array access (where we don't support it due to 2502 * the difficulty of doing relative addressing on instruction 2503 * destinations), we could potentially do array access of uniforms 2504 * that were loaded in GRF space as push constants. In real-world 2505 * usage we've seen, though, the arrays being used are always larger 2506 * than we could load as push constants, so just always move all 2507 * uniform array access out to a pull constant buffer. 2508 */ 2509void 2510vec4_visitor::move_uniform_array_access_to_pull_constants() 2511{ 2512 int pull_constant_loc[this->uniforms]; 2513 2514 for (int i = 0; i < this->uniforms; i++) { 2515 pull_constant_loc[i] = -1; 2516 } 2517 2518 /* Walk through and find array access of uniforms. Put a copy of that 2519 * uniform in the pull constant buffer. 2520 * 2521 * Note that we don't move constant-indexed accesses to arrays. No 2522 * testing has been done of the performance impact of this choice. 2523 */ 2524 foreach_list_safe(node, &this->instructions) { 2525 vec4_instruction *inst = (vec4_instruction *)node; 2526 2527 for (int i = 0 ; i < 3; i++) { 2528 if (inst->src[i].file != UNIFORM || !inst->src[i].reladdr) 2529 continue; 2530 2531 int uniform = inst->src[i].reg; 2532 2533 /* If this array isn't already present in the pull constant buffer, 2534 * add it. 2535 */ 2536 if (pull_constant_loc[uniform] == -1) { 2537 const float **values = &prog_data->param[uniform * 4]; 2538 2539 pull_constant_loc[uniform] = prog_data->nr_pull_params / 4; 2540 2541 for (int j = 0; j < uniform_size[uniform] * 4; j++) { 2542 prog_data->pull_param[prog_data->nr_pull_params++] = values[j]; 2543 } 2544 } 2545 2546 /* Set up the annotation tracking for new generated instructions. */ 2547 base_ir = inst->ir; 2548 current_annotation = inst->annotation; 2549 2550 dst_reg temp = dst_reg(this, glsl_type::vec4_type); 2551 2552 emit_pull_constant_load(inst, temp, inst->src[i], 2553 pull_constant_loc[uniform]); 2554 2555 inst->src[i].file = temp.file; 2556 inst->src[i].reg = temp.reg; 2557 inst->src[i].reg_offset = temp.reg_offset; 2558 inst->src[i].reladdr = NULL; 2559 } 2560 } 2561 2562 /* Now there are no accesses of the UNIFORM file with a reladdr, so 2563 * no need to track them as larger-than-vec4 objects. This will be 2564 * relied on in cutting out unused uniform vectors from push 2565 * constants. 2566 */ 2567 split_uniform_registers(); 2568} 2569 2570void 2571vec4_visitor::resolve_ud_negate(src_reg *reg) 2572{ 2573 if (reg->type != BRW_REGISTER_TYPE_UD || 2574 !reg->negate) 2575 return; 2576 2577 src_reg temp = src_reg(this, glsl_type::uvec4_type); 2578 emit(BRW_OPCODE_MOV, dst_reg(temp), *reg); 2579 *reg = temp; 2580} 2581 2582vec4_visitor::vec4_visitor(struct brw_vs_compile *c, 2583 struct gl_shader_program *prog, 2584 struct brw_shader *shader) 2585{ 2586 this->c = c; 2587 this->p = &c->func; 2588 this->brw = p->brw; 2589 this->intel = &brw->intel; 2590 this->ctx = &intel->ctx; 2591 this->prog = prog; 2592 this->shader = shader; 2593 2594 this->mem_ctx = ralloc_context(NULL); 2595 this->failed = false; 2596 2597 this->base_ir = NULL; 2598 this->current_annotation = NULL; 2599 2600 this->c = c; 2601 this->vp = (struct gl_vertex_program *) 2602 prog->_LinkedShaders[MESA_SHADER_VERTEX]->Program; 2603 this->prog_data = &c->prog_data; 2604 2605 this->variable_ht = hash_table_ctor(0, 2606 hash_table_pointer_hash, 2607 hash_table_pointer_compare); 2608 2609 this->virtual_grf_def = NULL; 2610 this->virtual_grf_use = NULL; 2611 this->virtual_grf_sizes = NULL; 2612 this->virtual_grf_count = 0; 2613 this->virtual_grf_reg_map = NULL; 2614 this->virtual_grf_reg_count = 0; 2615 this->virtual_grf_array_size = 0; 2616 this->live_intervals_valid = false; 2617 2618 this->max_grf = intel->gen >= 7 ? GEN7_MRF_HACK_START : BRW_MAX_GRF; 2619 2620 this->uniforms = 0; 2621} 2622 2623vec4_visitor::~vec4_visitor() 2624{ 2625 ralloc_free(this->mem_ctx); 2626 hash_table_dtor(this->variable_ht); 2627} 2628 2629 2630void 2631vec4_visitor::fail(const char *format, ...) 2632{ 2633 va_list va; 2634 char *msg; 2635 2636 if (failed) 2637 return; 2638 2639 failed = true; 2640 2641 va_start(va, format); 2642 msg = ralloc_vasprintf(mem_ctx, format, va); 2643 va_end(va); 2644 msg = ralloc_asprintf(mem_ctx, "VS compile failed: %s\n", msg); 2645 2646 this->fail_msg = msg; 2647 2648 if (INTEL_DEBUG & DEBUG_VS) { 2649 fprintf(stderr, "%s", msg); 2650 } 2651} 2652 2653} /* namespace brw */ 2654