brw_vec4_visitor.cpp revision d4444b8e5b914cde428e549c0db9418ddc1402f6
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} 29 30namespace brw { 31 32src_reg::src_reg(dst_reg reg) 33{ 34 init(); 35 36 this->file = reg.file; 37 this->reg = reg.reg; 38 this->reg_offset = reg.reg_offset; 39 this->type = reg.type; 40 this->reladdr = reg.reladdr; 41 this->fixed_hw_reg = reg.fixed_hw_reg; 42 43 int swizzles[4]; 44 int next_chan = 0; 45 int last = 0; 46 47 for (int i = 0; i < 4; i++) { 48 if (!(reg.writemask & (1 << i))) 49 continue; 50 51 swizzles[next_chan++] = last = i; 52 } 53 54 for (; next_chan < 4; next_chan++) { 55 swizzles[next_chan] = last; 56 } 57 58 this->swizzle = BRW_SWIZZLE4(swizzles[0], swizzles[1], 59 swizzles[2], swizzles[3]); 60} 61 62dst_reg::dst_reg(src_reg reg) 63{ 64 init(); 65 66 this->file = reg.file; 67 this->reg = reg.reg; 68 this->reg_offset = reg.reg_offset; 69 this->type = reg.type; 70 this->writemask = WRITEMASK_XYZW; 71 this->reladdr = reg.reladdr; 72 this->fixed_hw_reg = reg.fixed_hw_reg; 73} 74 75vec4_instruction::vec4_instruction(vec4_visitor *v, 76 enum opcode opcode, dst_reg dst, 77 src_reg src0, src_reg src1, src_reg src2) 78{ 79 this->opcode = opcode; 80 this->dst = dst; 81 this->src[0] = src0; 82 this->src[1] = src1; 83 this->src[2] = src2; 84 this->ir = v->base_ir; 85 this->annotation = v->current_annotation; 86} 87 88vec4_instruction * 89vec4_visitor::emit(vec4_instruction *inst) 90{ 91 this->instructions.push_tail(inst); 92 93 return inst; 94} 95 96vec4_instruction * 97vec4_visitor::emit_before(vec4_instruction *inst, vec4_instruction *new_inst) 98{ 99 new_inst->ir = inst->ir; 100 new_inst->annotation = inst->annotation; 101 102 inst->insert_before(new_inst); 103 104 return inst; 105} 106 107vec4_instruction * 108vec4_visitor::emit(enum opcode opcode, dst_reg dst, 109 src_reg src0, src_reg src1, src_reg src2) 110{ 111 return emit(new(mem_ctx) vec4_instruction(this, opcode, dst, 112 src0, src1, src2)); 113} 114 115 116vec4_instruction * 117vec4_visitor::emit(enum opcode opcode, dst_reg dst, src_reg src0, src_reg src1) 118{ 119 return emit(new(mem_ctx) vec4_instruction(this, opcode, dst, src0, src1)); 120} 121 122vec4_instruction * 123vec4_visitor::emit(enum opcode opcode, dst_reg dst, src_reg src0) 124{ 125 return emit(new(mem_ctx) vec4_instruction(this, opcode, dst, src0)); 126} 127 128vec4_instruction * 129vec4_visitor::emit(enum opcode opcode) 130{ 131 return emit(new(mem_ctx) vec4_instruction(this, opcode, dst_reg())); 132} 133 134#define ALU1(op) \ 135 vec4_instruction * \ 136 vec4_visitor::op(dst_reg dst, src_reg src0) \ 137 { \ 138 return new(mem_ctx) vec4_instruction(this, BRW_OPCODE_##op, dst, \ 139 src0); \ 140 } 141 142#define ALU2(op) \ 143 vec4_instruction * \ 144 vec4_visitor::op(dst_reg dst, src_reg src0, src_reg src1) \ 145 { \ 146 return new(mem_ctx) vec4_instruction(this, BRW_OPCODE_##op, dst, \ 147 src0, src1); \ 148 } 149 150ALU1(NOT) 151ALU1(MOV) 152ALU1(FRC) 153ALU1(RNDD) 154ALU1(RNDE) 155ALU1(RNDZ) 156ALU2(ADD) 157ALU2(MUL) 158ALU2(MACH) 159ALU2(AND) 160ALU2(OR) 161ALU2(XOR) 162ALU2(DP3) 163ALU2(DP4) 164 165/** Gen4 predicated IF. */ 166vec4_instruction * 167vec4_visitor::IF(uint32_t predicate) 168{ 169 vec4_instruction *inst; 170 171 inst = new(mem_ctx) vec4_instruction(this, BRW_OPCODE_IF); 172 inst->predicate = predicate; 173 174 return inst; 175} 176 177/** Gen6+ IF with embedded comparison. */ 178vec4_instruction * 179vec4_visitor::IF(src_reg src0, src_reg src1, uint32_t condition) 180{ 181 assert(intel->gen >= 6); 182 183 vec4_instruction *inst; 184 185 inst = new(mem_ctx) vec4_instruction(this, BRW_OPCODE_IF, dst_null_d(), 186 src0, src1); 187 inst->conditional_mod = condition; 188 189 return inst; 190} 191 192/** 193 * CMP: Sets the low bit of the destination channels with the result 194 * of the comparison, while the upper bits are undefined, and updates 195 * the flag register with the packed 16 bits of the result. 196 */ 197vec4_instruction * 198vec4_visitor::CMP(dst_reg dst, src_reg src0, src_reg src1, uint32_t condition) 199{ 200 vec4_instruction *inst; 201 202 /* original gen4 does type conversion to the destination type 203 * before before comparison, producing garbage results for floating 204 * point comparisons. 205 */ 206 if (intel->gen == 4) { 207 dst.type = src0.type; 208 if (dst.file == HW_REG) 209 dst.fixed_hw_reg.type = dst.type; 210 } 211 212 inst = new(mem_ctx) vec4_instruction(this, BRW_OPCODE_CMP, dst, src0, src1); 213 inst->conditional_mod = condition; 214 215 return inst; 216} 217 218vec4_instruction * 219vec4_visitor::SCRATCH_READ(dst_reg dst, src_reg index) 220{ 221 vec4_instruction *inst; 222 223 inst = new(mem_ctx) vec4_instruction(this, VS_OPCODE_SCRATCH_READ, 224 dst, index); 225 inst->base_mrf = 14; 226 inst->mlen = 1; 227 228 return inst; 229} 230 231vec4_instruction * 232vec4_visitor::SCRATCH_WRITE(dst_reg dst, src_reg src, src_reg index) 233{ 234 vec4_instruction *inst; 235 236 inst = new(mem_ctx) vec4_instruction(this, VS_OPCODE_SCRATCH_WRITE, 237 dst, src, index); 238 inst->base_mrf = 13; 239 inst->mlen = 2; 240 241 return inst; 242} 243 244void 245vec4_visitor::emit_dp(dst_reg dst, src_reg src0, src_reg src1, unsigned elements) 246{ 247 static enum opcode dot_opcodes[] = { 248 BRW_OPCODE_DP2, BRW_OPCODE_DP3, BRW_OPCODE_DP4 249 }; 250 251 emit(dot_opcodes[elements - 2], dst, src0, src1); 252} 253 254void 255vec4_visitor::emit_math1_gen6(enum opcode opcode, dst_reg dst, src_reg src) 256{ 257 /* The gen6 math instruction ignores the source modifiers -- 258 * swizzle, abs, negate, and at least some parts of the register 259 * region description. 260 * 261 * While it would seem that this MOV could be avoided at this point 262 * in the case that the swizzle is matched up with the destination 263 * writemask, note that uniform packing and register allocation 264 * could rearrange our swizzle, so let's leave this matter up to 265 * copy propagation later. 266 */ 267 src_reg temp_src = src_reg(this, glsl_type::vec4_type); 268 emit(MOV(dst_reg(temp_src), src)); 269 270 if (dst.writemask != WRITEMASK_XYZW) { 271 /* The gen6 math instruction must be align1, so we can't do 272 * writemasks. 273 */ 274 dst_reg temp_dst = dst_reg(this, glsl_type::vec4_type); 275 276 emit(opcode, temp_dst, temp_src); 277 278 emit(MOV(dst, src_reg(temp_dst))); 279 } else { 280 emit(opcode, dst, temp_src); 281 } 282} 283 284void 285vec4_visitor::emit_math1_gen4(enum opcode opcode, dst_reg dst, src_reg src) 286{ 287 vec4_instruction *inst = emit(opcode, dst, src); 288 inst->base_mrf = 1; 289 inst->mlen = 1; 290} 291 292void 293vec4_visitor::emit_math(opcode opcode, dst_reg dst, src_reg src) 294{ 295 switch (opcode) { 296 case SHADER_OPCODE_RCP: 297 case SHADER_OPCODE_RSQ: 298 case SHADER_OPCODE_SQRT: 299 case SHADER_OPCODE_EXP2: 300 case SHADER_OPCODE_LOG2: 301 case SHADER_OPCODE_SIN: 302 case SHADER_OPCODE_COS: 303 break; 304 default: 305 assert(!"not reached: bad math opcode"); 306 return; 307 } 308 309 if (intel->gen >= 6) { 310 return emit_math1_gen6(opcode, dst, src); 311 } else { 312 return emit_math1_gen4(opcode, dst, src); 313 } 314} 315 316void 317vec4_visitor::emit_math2_gen6(enum opcode opcode, 318 dst_reg dst, src_reg src0, src_reg src1) 319{ 320 src_reg expanded; 321 322 /* The gen6 math instruction ignores the source modifiers -- 323 * swizzle, abs, negate, and at least some parts of the register 324 * region description. Move the sources to temporaries to make it 325 * generally work. 326 */ 327 328 expanded = src_reg(this, glsl_type::vec4_type); 329 emit(MOV(dst_reg(expanded), src0)); 330 src0 = expanded; 331 332 expanded = src_reg(this, glsl_type::vec4_type); 333 emit(MOV(dst_reg(expanded), src1)); 334 src1 = expanded; 335 336 if (dst.writemask != WRITEMASK_XYZW) { 337 /* The gen6 math instruction must be align1, so we can't do 338 * writemasks. 339 */ 340 dst_reg temp_dst = dst_reg(this, glsl_type::vec4_type); 341 342 emit(opcode, temp_dst, src0, src1); 343 344 emit(MOV(dst, src_reg(temp_dst))); 345 } else { 346 emit(opcode, dst, src0, src1); 347 } 348} 349 350void 351vec4_visitor::emit_math2_gen4(enum opcode opcode, 352 dst_reg dst, src_reg src0, src_reg src1) 353{ 354 vec4_instruction *inst = emit(opcode, dst, src0, src1); 355 inst->base_mrf = 1; 356 inst->mlen = 2; 357} 358 359void 360vec4_visitor::emit_math(enum opcode opcode, 361 dst_reg dst, src_reg src0, src_reg src1) 362{ 363 assert(opcode == SHADER_OPCODE_POW); 364 365 if (intel->gen >= 6) { 366 return emit_math2_gen6(opcode, dst, src0, src1); 367 } else { 368 return emit_math2_gen4(opcode, dst, src0, src1); 369 } 370} 371 372void 373vec4_visitor::visit_instructions(const exec_list *list) 374{ 375 foreach_list(node, list) { 376 ir_instruction *ir = (ir_instruction *)node; 377 378 base_ir = ir; 379 ir->accept(this); 380 } 381} 382 383 384static int 385type_size(const struct glsl_type *type) 386{ 387 unsigned int i; 388 int size; 389 390 switch (type->base_type) { 391 case GLSL_TYPE_UINT: 392 case GLSL_TYPE_INT: 393 case GLSL_TYPE_FLOAT: 394 case GLSL_TYPE_BOOL: 395 if (type->is_matrix()) { 396 return type->matrix_columns; 397 } else { 398 /* Regardless of size of vector, it gets a vec4. This is bad 399 * packing for things like floats, but otherwise arrays become a 400 * mess. Hopefully a later pass over the code can pack scalars 401 * down if appropriate. 402 */ 403 return 1; 404 } 405 case GLSL_TYPE_ARRAY: 406 assert(type->length > 0); 407 return type_size(type->fields.array) * type->length; 408 case GLSL_TYPE_STRUCT: 409 size = 0; 410 for (i = 0; i < type->length; i++) { 411 size += type_size(type->fields.structure[i].type); 412 } 413 return size; 414 case GLSL_TYPE_SAMPLER: 415 /* Samplers take up one slot in UNIFORMS[], but they're baked in 416 * at link time. 417 */ 418 return 1; 419 default: 420 assert(0); 421 return 0; 422 } 423} 424 425int 426vec4_visitor::virtual_grf_alloc(int size) 427{ 428 if (virtual_grf_array_size <= virtual_grf_count) { 429 if (virtual_grf_array_size == 0) 430 virtual_grf_array_size = 16; 431 else 432 virtual_grf_array_size *= 2; 433 virtual_grf_sizes = reralloc(mem_ctx, virtual_grf_sizes, int, 434 virtual_grf_array_size); 435 virtual_grf_reg_map = reralloc(mem_ctx, virtual_grf_reg_map, int, 436 virtual_grf_array_size); 437 } 438 virtual_grf_reg_map[virtual_grf_count] = virtual_grf_reg_count; 439 virtual_grf_reg_count += size; 440 virtual_grf_sizes[virtual_grf_count] = size; 441 return virtual_grf_count++; 442} 443 444src_reg::src_reg(class vec4_visitor *v, const struct glsl_type *type) 445{ 446 init(); 447 448 this->file = GRF; 449 this->reg = v->virtual_grf_alloc(type_size(type)); 450 451 if (type->is_array() || type->is_record()) { 452 this->swizzle = BRW_SWIZZLE_NOOP; 453 } else { 454 this->swizzle = swizzle_for_size(type->vector_elements); 455 } 456 457 this->type = brw_type_for_base_type(type); 458} 459 460dst_reg::dst_reg(class vec4_visitor *v, const struct glsl_type *type) 461{ 462 init(); 463 464 this->file = GRF; 465 this->reg = v->virtual_grf_alloc(type_size(type)); 466 467 if (type->is_array() || type->is_record()) { 468 this->writemask = WRITEMASK_XYZW; 469 } else { 470 this->writemask = (1 << type->vector_elements) - 1; 471 } 472 473 this->type = brw_type_for_base_type(type); 474} 475 476/* Our support for uniforms is piggy-backed on the struct 477 * gl_fragment_program, because that's where the values actually 478 * get stored, rather than in some global gl_shader_program uniform 479 * store. 480 */ 481int 482vec4_visitor::setup_uniform_values(int loc, const glsl_type *type) 483{ 484 unsigned int offset = 0; 485 float *values = &this->vp->Base.Parameters->ParameterValues[loc][0].f; 486 487 if (type->is_matrix()) { 488 const glsl_type *column = glsl_type::get_instance(GLSL_TYPE_FLOAT, 489 type->vector_elements, 490 1); 491 492 for (unsigned int i = 0; i < type->matrix_columns; i++) { 493 offset += setup_uniform_values(loc + offset, column); 494 } 495 496 return offset; 497 } 498 499 switch (type->base_type) { 500 case GLSL_TYPE_FLOAT: 501 case GLSL_TYPE_UINT: 502 case GLSL_TYPE_INT: 503 case GLSL_TYPE_BOOL: 504 for (unsigned int i = 0; i < type->vector_elements; i++) { 505 c->prog_data.param[this->uniforms * 4 + i] = &values[i]; 506 } 507 508 /* Set up pad elements to get things aligned to a vec4 boundary. */ 509 for (unsigned int i = type->vector_elements; i < 4; i++) { 510 static float zero = 0; 511 512 c->prog_data.param[this->uniforms * 4 + i] = &zero; 513 } 514 515 /* Track the size of this uniform vector, for future packing of 516 * uniforms. 517 */ 518 this->uniform_vector_size[this->uniforms] = type->vector_elements; 519 this->uniforms++; 520 521 return 1; 522 523 case GLSL_TYPE_STRUCT: 524 for (unsigned int i = 0; i < type->length; i++) { 525 offset += setup_uniform_values(loc + offset, 526 type->fields.structure[i].type); 527 } 528 return offset; 529 530 case GLSL_TYPE_ARRAY: 531 for (unsigned int i = 0; i < type->length; i++) { 532 offset += setup_uniform_values(loc + offset, type->fields.array); 533 } 534 return offset; 535 536 case GLSL_TYPE_SAMPLER: 537 /* The sampler takes up a slot, but we don't use any values from it. */ 538 return 1; 539 540 default: 541 assert(!"not reached"); 542 return 0; 543 } 544} 545 546/* Our support for builtin uniforms is even scarier than non-builtin. 547 * It sits on top of the PROG_STATE_VAR parameters that are 548 * automatically updated from GL context state. 549 */ 550void 551vec4_visitor::setup_builtin_uniform_values(ir_variable *ir) 552{ 553 const ir_state_slot *const slots = ir->state_slots; 554 assert(ir->state_slots != NULL); 555 556 for (unsigned int i = 0; i < ir->num_state_slots; i++) { 557 /* This state reference has already been setup by ir_to_mesa, 558 * but we'll get the same index back here. We can reference 559 * ParameterValues directly, since unlike brw_fs.cpp, we never 560 * add new state references during compile. 561 */ 562 int index = _mesa_add_state_reference(this->vp->Base.Parameters, 563 (gl_state_index *)slots[i].tokens); 564 float *values = &this->vp->Base.Parameters->ParameterValues[index][0].f; 565 566 this->uniform_vector_size[this->uniforms] = 0; 567 /* Add each of the unique swizzled channels of the element. 568 * This will end up matching the size of the glsl_type of this field. 569 */ 570 int last_swiz = -1; 571 for (unsigned int j = 0; j < 4; j++) { 572 int swiz = GET_SWZ(slots[i].swizzle, j); 573 last_swiz = swiz; 574 575 c->prog_data.param[this->uniforms * 4 + j] = &values[swiz]; 576 if (swiz <= last_swiz) 577 this->uniform_vector_size[this->uniforms]++; 578 } 579 this->uniforms++; 580 } 581} 582 583dst_reg * 584vec4_visitor::variable_storage(ir_variable *var) 585{ 586 return (dst_reg *)hash_table_find(this->variable_ht, var); 587} 588 589void 590vec4_visitor::emit_bool_to_cond_code(ir_rvalue *ir, uint32_t *predicate) 591{ 592 ir_expression *expr = ir->as_expression(); 593 594 *predicate = BRW_PREDICATE_NORMAL; 595 596 if (expr) { 597 src_reg op[2]; 598 vec4_instruction *inst; 599 600 assert(expr->get_num_operands() <= 2); 601 for (unsigned int i = 0; i < expr->get_num_operands(); i++) { 602 expr->operands[i]->accept(this); 603 op[i] = this->result; 604 } 605 606 switch (expr->operation) { 607 case ir_unop_logic_not: 608 inst = emit(AND(dst_null_d(), op[0], src_reg(1))); 609 inst->conditional_mod = BRW_CONDITIONAL_Z; 610 break; 611 612 case ir_binop_logic_xor: 613 inst = emit(XOR(dst_null_d(), op[0], op[1])); 614 inst->conditional_mod = BRW_CONDITIONAL_NZ; 615 break; 616 617 case ir_binop_logic_or: 618 inst = emit(OR(dst_null_d(), op[0], op[1])); 619 inst->conditional_mod = BRW_CONDITIONAL_NZ; 620 break; 621 622 case ir_binop_logic_and: 623 inst = emit(AND(dst_null_d(), op[0], op[1])); 624 inst->conditional_mod = BRW_CONDITIONAL_NZ; 625 break; 626 627 case ir_unop_f2b: 628 if (intel->gen >= 6) { 629 emit(CMP(dst_null_d(), op[0], src_reg(0.0f), BRW_CONDITIONAL_NZ)); 630 } else { 631 inst = emit(MOV(dst_null_f(), op[0])); 632 inst->conditional_mod = BRW_CONDITIONAL_NZ; 633 } 634 break; 635 636 case ir_unop_i2b: 637 if (intel->gen >= 6) { 638 emit(CMP(dst_null_d(), op[0], src_reg(0), BRW_CONDITIONAL_NZ)); 639 } else { 640 inst = emit(MOV(dst_null_d(), op[0])); 641 inst->conditional_mod = BRW_CONDITIONAL_NZ; 642 } 643 break; 644 645 case ir_binop_all_equal: 646 inst = emit(CMP(dst_null_d(), op[0], op[1], BRW_CONDITIONAL_Z)); 647 *predicate = BRW_PREDICATE_ALIGN16_ALL4H; 648 break; 649 650 case ir_binop_any_nequal: 651 inst = emit(CMP(dst_null_d(), op[0], op[1], BRW_CONDITIONAL_NZ)); 652 *predicate = BRW_PREDICATE_ALIGN16_ANY4H; 653 break; 654 655 case ir_unop_any: 656 inst = emit(CMP(dst_null_d(), op[0], src_reg(0), BRW_CONDITIONAL_NZ)); 657 *predicate = BRW_PREDICATE_ALIGN16_ANY4H; 658 break; 659 660 case ir_binop_greater: 661 case ir_binop_gequal: 662 case ir_binop_less: 663 case ir_binop_lequal: 664 case ir_binop_equal: 665 case ir_binop_nequal: 666 emit(CMP(dst_null_d(), op[0], op[1], 667 brw_conditional_for_comparison(expr->operation))); 668 break; 669 670 default: 671 assert(!"not reached"); 672 break; 673 } 674 return; 675 } 676 677 ir->accept(this); 678 679 if (intel->gen >= 6) { 680 vec4_instruction *inst = emit(AND(dst_null_d(), 681 this->result, src_reg(1))); 682 inst->conditional_mod = BRW_CONDITIONAL_NZ; 683 } else { 684 vec4_instruction *inst = emit(MOV(dst_null_d(), this->result)); 685 inst->conditional_mod = BRW_CONDITIONAL_NZ; 686 } 687} 688 689/** 690 * Emit a gen6 IF statement with the comparison folded into the IF 691 * instruction. 692 */ 693void 694vec4_visitor::emit_if_gen6(ir_if *ir) 695{ 696 ir_expression *expr = ir->condition->as_expression(); 697 698 if (expr) { 699 src_reg op[2]; 700 dst_reg temp; 701 702 assert(expr->get_num_operands() <= 2); 703 for (unsigned int i = 0; i < expr->get_num_operands(); i++) { 704 expr->operands[i]->accept(this); 705 op[i] = this->result; 706 } 707 708 switch (expr->operation) { 709 case ir_unop_logic_not: 710 emit(IF(op[0], src_reg(0), BRW_CONDITIONAL_Z)); 711 return; 712 713 case ir_binop_logic_xor: 714 emit(IF(op[0], op[1], BRW_CONDITIONAL_NZ)); 715 return; 716 717 case ir_binop_logic_or: 718 temp = dst_reg(this, glsl_type::bool_type); 719 emit(OR(temp, op[0], op[1])); 720 emit(IF(src_reg(temp), src_reg(0), BRW_CONDITIONAL_NZ)); 721 return; 722 723 case ir_binop_logic_and: 724 temp = dst_reg(this, glsl_type::bool_type); 725 emit(AND(temp, op[0], op[1])); 726 emit(IF(src_reg(temp), src_reg(0), BRW_CONDITIONAL_NZ)); 727 return; 728 729 case ir_unop_f2b: 730 emit(IF(op[0], src_reg(0), BRW_CONDITIONAL_NZ)); 731 return; 732 733 case ir_unop_i2b: 734 emit(IF(op[0], src_reg(0), BRW_CONDITIONAL_NZ)); 735 return; 736 737 case ir_binop_greater: 738 case ir_binop_gequal: 739 case ir_binop_less: 740 case ir_binop_lequal: 741 case ir_binop_equal: 742 case ir_binop_nequal: 743 emit(IF(op[0], op[1], 744 brw_conditional_for_comparison(expr->operation))); 745 return; 746 747 case ir_binop_all_equal: 748 emit(CMP(dst_null_d(), op[0], op[1], BRW_CONDITIONAL_Z)); 749 emit(IF(BRW_PREDICATE_ALIGN16_ALL4H)); 750 return; 751 752 case ir_binop_any_nequal: 753 emit(CMP(dst_null_d(), op[0], op[1], BRW_CONDITIONAL_NZ)); 754 emit(IF(BRW_PREDICATE_ALIGN16_ANY4H)); 755 return; 756 757 case ir_unop_any: 758 emit(CMP(dst_null_d(), op[0], src_reg(0), BRW_CONDITIONAL_NZ)); 759 emit(IF(BRW_PREDICATE_ALIGN16_ANY4H)); 760 return; 761 762 default: 763 assert(!"not reached"); 764 emit(IF(op[0], src_reg(0), BRW_CONDITIONAL_NZ)); 765 return; 766 } 767 return; 768 } 769 770 ir->condition->accept(this); 771 772 emit(IF(this->result, src_reg(0), BRW_CONDITIONAL_NZ)); 773} 774 775void 776vec4_visitor::visit(ir_variable *ir) 777{ 778 dst_reg *reg = NULL; 779 780 if (variable_storage(ir)) 781 return; 782 783 switch (ir->mode) { 784 case ir_var_in: 785 reg = new(mem_ctx) dst_reg(ATTR, ir->location); 786 787 /* Do GL_FIXED rescaling for GLES2.0. Our GL_FIXED attributes 788 * come in as floating point conversions of the integer values. 789 */ 790 for (int i = ir->location; i < ir->location + type_size(ir->type); i++) { 791 if (!c->key.gl_fixed_input_size[i]) 792 continue; 793 794 dst_reg dst = *reg; 795 dst.writemask = (1 << c->key.gl_fixed_input_size[i]) - 1; 796 emit(MUL(dst, src_reg(dst), src_reg(1.0f / 65536.0f))); 797 } 798 break; 799 800 case ir_var_out: 801 reg = new(mem_ctx) dst_reg(this, ir->type); 802 803 for (int i = 0; i < type_size(ir->type); i++) { 804 output_reg[ir->location + i] = *reg; 805 output_reg[ir->location + i].reg_offset = i; 806 output_reg[ir->location + i].type = BRW_REGISTER_TYPE_F; 807 output_reg_annotation[ir->location + i] = ir->name; 808 } 809 break; 810 811 case ir_var_auto: 812 case ir_var_temporary: 813 reg = new(mem_ctx) dst_reg(this, ir->type); 814 break; 815 816 case ir_var_uniform: 817 reg = new(this->mem_ctx) dst_reg(UNIFORM, this->uniforms); 818 819 /* Track how big the whole uniform variable is, in case we need to put a 820 * copy of its data into pull constants for array access. 821 */ 822 this->uniform_size[this->uniforms] = type_size(ir->type); 823 824 if (!strncmp(ir->name, "gl_", 3)) { 825 setup_builtin_uniform_values(ir); 826 } else { 827 setup_uniform_values(ir->location, ir->type); 828 } 829 break; 830 831 default: 832 assert(!"not reached"); 833 } 834 835 reg->type = brw_type_for_base_type(ir->type); 836 hash_table_insert(this->variable_ht, reg, ir); 837} 838 839void 840vec4_visitor::visit(ir_loop *ir) 841{ 842 dst_reg counter; 843 844 /* We don't want debugging output to print the whole body of the 845 * loop as the annotation. 846 */ 847 this->base_ir = NULL; 848 849 if (ir->counter != NULL) { 850 this->base_ir = ir->counter; 851 ir->counter->accept(this); 852 counter = *(variable_storage(ir->counter)); 853 854 if (ir->from != NULL) { 855 this->base_ir = ir->from; 856 ir->from->accept(this); 857 858 emit(MOV(counter, this->result)); 859 } 860 } 861 862 emit(BRW_OPCODE_DO); 863 864 if (ir->to) { 865 this->base_ir = ir->to; 866 ir->to->accept(this); 867 868 emit(CMP(dst_null_d(), src_reg(counter), this->result, 869 brw_conditional_for_comparison(ir->cmp))); 870 871 vec4_instruction *inst = emit(BRW_OPCODE_BREAK); 872 inst->predicate = BRW_PREDICATE_NORMAL; 873 } 874 875 visit_instructions(&ir->body_instructions); 876 877 878 if (ir->increment) { 879 this->base_ir = ir->increment; 880 ir->increment->accept(this); 881 emit(ADD(counter, src_reg(counter), this->result)); 882 } 883 884 emit(BRW_OPCODE_WHILE); 885} 886 887void 888vec4_visitor::visit(ir_loop_jump *ir) 889{ 890 switch (ir->mode) { 891 case ir_loop_jump::jump_break: 892 emit(BRW_OPCODE_BREAK); 893 break; 894 case ir_loop_jump::jump_continue: 895 emit(BRW_OPCODE_CONTINUE); 896 break; 897 } 898} 899 900 901void 902vec4_visitor::visit(ir_function_signature *ir) 903{ 904 assert(0); 905 (void)ir; 906} 907 908void 909vec4_visitor::visit(ir_function *ir) 910{ 911 /* Ignore function bodies other than main() -- we shouldn't see calls to 912 * them since they should all be inlined. 913 */ 914 if (strcmp(ir->name, "main") == 0) { 915 const ir_function_signature *sig; 916 exec_list empty; 917 918 sig = ir->matching_signature(&empty); 919 920 assert(sig); 921 922 visit_instructions(&sig->body); 923 } 924} 925 926GLboolean 927vec4_visitor::try_emit_sat(ir_expression *ir) 928{ 929 ir_rvalue *sat_src = ir->as_rvalue_to_saturate(); 930 if (!sat_src) 931 return false; 932 933 sat_src->accept(this); 934 src_reg src = this->result; 935 936 this->result = src_reg(this, ir->type); 937 vec4_instruction *inst; 938 inst = emit(MOV(dst_reg(this->result), src)); 939 inst->saturate = true; 940 941 return true; 942} 943 944void 945vec4_visitor::emit_bool_comparison(unsigned int op, 946 dst_reg dst, src_reg src0, src_reg src1) 947{ 948 /* original gen4 does destination conversion before comparison. */ 949 if (intel->gen < 5) 950 dst.type = src0.type; 951 952 emit(CMP(dst, src0, src1, brw_conditional_for_comparison(op))); 953 954 dst.type = BRW_REGISTER_TYPE_D; 955 emit(AND(dst, src_reg(dst), src_reg(0x1))); 956} 957 958void 959vec4_visitor::visit(ir_expression *ir) 960{ 961 unsigned int operand; 962 src_reg op[Elements(ir->operands)]; 963 src_reg result_src; 964 dst_reg result_dst; 965 vec4_instruction *inst; 966 967 if (try_emit_sat(ir)) 968 return; 969 970 for (operand = 0; operand < ir->get_num_operands(); operand++) { 971 this->result.file = BAD_FILE; 972 ir->operands[operand]->accept(this); 973 if (this->result.file == BAD_FILE) { 974 printf("Failed to get tree for expression operand:\n"); 975 ir->operands[operand]->print(); 976 exit(1); 977 } 978 op[operand] = this->result; 979 980 /* Matrix expression operands should have been broken down to vector 981 * operations already. 982 */ 983 assert(!ir->operands[operand]->type->is_matrix()); 984 } 985 986 int vector_elements = ir->operands[0]->type->vector_elements; 987 if (ir->operands[1]) { 988 vector_elements = MAX2(vector_elements, 989 ir->operands[1]->type->vector_elements); 990 } 991 992 this->result.file = BAD_FILE; 993 994 /* Storage for our result. Ideally for an assignment we'd be using 995 * the actual storage for the result here, instead. 996 */ 997 result_src = src_reg(this, ir->type); 998 /* convenience for the emit functions below. */ 999 result_dst = dst_reg(result_src); 1000 /* If nothing special happens, this is the result. */ 1001 this->result = result_src; 1002 /* Limit writes to the channels that will be used by result_src later. 1003 * This does limit this temp's use as a temporary for multi-instruction 1004 * sequences. 1005 */ 1006 result_dst.writemask = (1 << ir->type->vector_elements) - 1; 1007 1008 switch (ir->operation) { 1009 case ir_unop_logic_not: 1010 /* Note that BRW_OPCODE_NOT is not appropriate here, since it is 1011 * ones complement of the whole register, not just bit 0. 1012 */ 1013 emit(XOR(result_dst, op[0], src_reg(1))); 1014 break; 1015 case ir_unop_neg: 1016 op[0].negate = !op[0].negate; 1017 this->result = op[0]; 1018 break; 1019 case ir_unop_abs: 1020 op[0].abs = true; 1021 op[0].negate = false; 1022 this->result = op[0]; 1023 break; 1024 1025 case ir_unop_sign: 1026 emit(MOV(result_dst, src_reg(0.0f))); 1027 1028 emit(CMP(dst_null_d(), op[0], src_reg(0.0f), BRW_CONDITIONAL_G)); 1029 inst = emit(MOV(result_dst, src_reg(1.0f))); 1030 inst->predicate = BRW_PREDICATE_NORMAL; 1031 1032 emit(CMP(dst_null_d(), op[0], src_reg(0.0f), BRW_CONDITIONAL_L)); 1033 inst = emit(MOV(result_dst, src_reg(-1.0f))); 1034 inst->predicate = BRW_PREDICATE_NORMAL; 1035 1036 break; 1037 1038 case ir_unop_rcp: 1039 emit_math(SHADER_OPCODE_RCP, result_dst, op[0]); 1040 break; 1041 1042 case ir_unop_exp2: 1043 emit_math(SHADER_OPCODE_EXP2, result_dst, op[0]); 1044 break; 1045 case ir_unop_log2: 1046 emit_math(SHADER_OPCODE_LOG2, result_dst, op[0]); 1047 break; 1048 case ir_unop_exp: 1049 case ir_unop_log: 1050 assert(!"not reached: should be handled by ir_explog_to_explog2"); 1051 break; 1052 case ir_unop_sin: 1053 case ir_unop_sin_reduced: 1054 emit_math(SHADER_OPCODE_SIN, result_dst, op[0]); 1055 break; 1056 case ir_unop_cos: 1057 case ir_unop_cos_reduced: 1058 emit_math(SHADER_OPCODE_COS, result_dst, op[0]); 1059 break; 1060 1061 case ir_unop_dFdx: 1062 case ir_unop_dFdy: 1063 assert(!"derivatives not valid in vertex shader"); 1064 break; 1065 1066 case ir_unop_noise: 1067 assert(!"not reached: should be handled by lower_noise"); 1068 break; 1069 1070 case ir_binop_add: 1071 emit(ADD(result_dst, op[0], op[1])); 1072 break; 1073 case ir_binop_sub: 1074 assert(!"not reached: should be handled by ir_sub_to_add_neg"); 1075 break; 1076 1077 case ir_binop_mul: 1078 if (ir->type->is_integer()) { 1079 /* For integer multiplication, the MUL uses the low 16 bits 1080 * of one of the operands (src0 on gen6, src1 on gen7). The 1081 * MACH accumulates in the contribution of the upper 16 bits 1082 * of that operand. 1083 * 1084 * FINISHME: Emit just the MUL if we know an operand is small 1085 * enough. 1086 */ 1087 struct brw_reg acc = retype(brw_acc_reg(), BRW_REGISTER_TYPE_D); 1088 1089 emit(MUL(acc, op[0], op[1])); 1090 emit(MACH(dst_null_d(), op[0], op[1])); 1091 emit(MOV(result_dst, src_reg(acc))); 1092 } else { 1093 emit(MUL(result_dst, op[0], op[1])); 1094 } 1095 break; 1096 case ir_binop_div: 1097 assert(!"not reached: should be handled by ir_div_to_mul_rcp"); 1098 case ir_binop_mod: 1099 assert(!"ir_binop_mod should have been converted to b * fract(a/b)"); 1100 break; 1101 1102 case ir_binop_less: 1103 case ir_binop_greater: 1104 case ir_binop_lequal: 1105 case ir_binop_gequal: 1106 case ir_binop_equal: 1107 case ir_binop_nequal: { 1108 emit(CMP(result_dst, op[0], op[1], 1109 brw_conditional_for_comparison(ir->operation))); 1110 emit(AND(result_dst, result_src, src_reg(0x1))); 1111 break; 1112 } 1113 1114 case ir_binop_all_equal: 1115 /* "==" operator producing a scalar boolean. */ 1116 if (ir->operands[0]->type->is_vector() || 1117 ir->operands[1]->type->is_vector()) { 1118 emit(CMP(dst_null_d(), op[0], op[1], BRW_CONDITIONAL_Z)); 1119 emit(MOV(result_dst, src_reg(0))); 1120 inst = emit(MOV(result_dst, src_reg(1))); 1121 inst->predicate = BRW_PREDICATE_ALIGN16_ALL4H; 1122 } else { 1123 emit(CMP(result_dst, op[0], op[1], BRW_CONDITIONAL_Z)); 1124 emit(AND(result_dst, result_src, src_reg(0x1))); 1125 } 1126 break; 1127 case ir_binop_any_nequal: 1128 /* "!=" operator producing a scalar boolean. */ 1129 if (ir->operands[0]->type->is_vector() || 1130 ir->operands[1]->type->is_vector()) { 1131 emit(CMP(dst_null_d(), op[0], op[1], BRW_CONDITIONAL_NZ)); 1132 1133 emit(MOV(result_dst, src_reg(0))); 1134 inst = emit(MOV(result_dst, src_reg(1))); 1135 inst->predicate = BRW_PREDICATE_ALIGN16_ANY4H; 1136 } else { 1137 emit(CMP(result_dst, op[0], op[1], BRW_CONDITIONAL_NZ)); 1138 emit(AND(result_dst, result_src, src_reg(0x1))); 1139 } 1140 break; 1141 1142 case ir_unop_any: 1143 emit(CMP(dst_null_d(), op[0], src_reg(0), BRW_CONDITIONAL_NZ)); 1144 emit(MOV(result_dst, src_reg(0))); 1145 1146 inst = emit(MOV(result_dst, src_reg(1))); 1147 inst->predicate = BRW_PREDICATE_ALIGN16_ANY4H; 1148 break; 1149 1150 case ir_binop_logic_xor: 1151 emit(XOR(result_dst, op[0], op[1])); 1152 break; 1153 1154 case ir_binop_logic_or: 1155 emit(OR(result_dst, op[0], op[1])); 1156 break; 1157 1158 case ir_binop_logic_and: 1159 emit(AND(result_dst, op[0], op[1])); 1160 break; 1161 1162 case ir_binop_dot: 1163 assert(ir->operands[0]->type->is_vector()); 1164 assert(ir->operands[0]->type == ir->operands[1]->type); 1165 emit_dp(result_dst, op[0], op[1], ir->operands[0]->type->vector_elements); 1166 break; 1167 1168 case ir_unop_sqrt: 1169 emit_math(SHADER_OPCODE_SQRT, result_dst, op[0]); 1170 break; 1171 case ir_unop_rsq: 1172 emit_math(SHADER_OPCODE_RSQ, result_dst, op[0]); 1173 break; 1174 case ir_unop_i2f: 1175 case ir_unop_i2u: 1176 case ir_unop_u2i: 1177 case ir_unop_u2f: 1178 case ir_unop_b2f: 1179 case ir_unop_b2i: 1180 case ir_unop_f2i: 1181 emit(MOV(result_dst, op[0])); 1182 break; 1183 case ir_unop_f2b: 1184 case ir_unop_i2b: { 1185 emit(CMP(result_dst, op[0], src_reg(0.0f), BRW_CONDITIONAL_NZ)); 1186 emit(AND(result_dst, result_src, src_reg(1))); 1187 break; 1188 } 1189 1190 case ir_unop_trunc: 1191 emit(RNDZ(result_dst, op[0])); 1192 break; 1193 case ir_unop_ceil: 1194 op[0].negate = !op[0].negate; 1195 inst = emit(RNDD(result_dst, op[0])); 1196 this->result.negate = true; 1197 break; 1198 case ir_unop_floor: 1199 inst = emit(RNDD(result_dst, op[0])); 1200 break; 1201 case ir_unop_fract: 1202 inst = emit(FRC(result_dst, op[0])); 1203 break; 1204 case ir_unop_round_even: 1205 emit(RNDE(result_dst, op[0])); 1206 break; 1207 1208 case ir_binop_min: 1209 emit(CMP(result_dst, op[0], op[1], BRW_CONDITIONAL_L)); 1210 1211 inst = emit(BRW_OPCODE_SEL, result_dst, op[0], op[1]); 1212 inst->predicate = BRW_PREDICATE_NORMAL; 1213 break; 1214 case ir_binop_max: 1215 emit(CMP(result_dst, op[0], op[1], BRW_CONDITIONAL_G)); 1216 1217 inst = emit(BRW_OPCODE_SEL, result_dst, op[0], op[1]); 1218 inst->predicate = BRW_PREDICATE_NORMAL; 1219 break; 1220 1221 case ir_binop_pow: 1222 emit_math(SHADER_OPCODE_POW, result_dst, op[0], op[1]); 1223 break; 1224 1225 case ir_unop_bit_not: 1226 inst = emit(NOT(result_dst, op[0])); 1227 break; 1228 case ir_binop_bit_and: 1229 inst = emit(AND(result_dst, op[0], op[1])); 1230 break; 1231 case ir_binop_bit_xor: 1232 inst = emit(XOR(result_dst, op[0], op[1])); 1233 break; 1234 case ir_binop_bit_or: 1235 inst = emit(OR(result_dst, op[0], op[1])); 1236 break; 1237 1238 case ir_binop_lshift: 1239 case ir_binop_rshift: 1240 assert(!"GLSL 1.30 features unsupported"); 1241 break; 1242 1243 case ir_quadop_vector: 1244 assert(!"not reached: should be handled by lower_quadop_vector"); 1245 break; 1246 } 1247} 1248 1249 1250void 1251vec4_visitor::visit(ir_swizzle *ir) 1252{ 1253 src_reg src; 1254 int i = 0; 1255 int swizzle[4]; 1256 1257 /* Note that this is only swizzles in expressions, not those on the left 1258 * hand side of an assignment, which do write masking. See ir_assignment 1259 * for that. 1260 */ 1261 1262 ir->val->accept(this); 1263 src = this->result; 1264 assert(src.file != BAD_FILE); 1265 1266 for (i = 0; i < ir->type->vector_elements; i++) { 1267 switch (i) { 1268 case 0: 1269 swizzle[i] = BRW_GET_SWZ(src.swizzle, ir->mask.x); 1270 break; 1271 case 1: 1272 swizzle[i] = BRW_GET_SWZ(src.swizzle, ir->mask.y); 1273 break; 1274 case 2: 1275 swizzle[i] = BRW_GET_SWZ(src.swizzle, ir->mask.z); 1276 break; 1277 case 3: 1278 swizzle[i] = BRW_GET_SWZ(src.swizzle, ir->mask.w); 1279 break; 1280 } 1281 } 1282 for (; i < 4; i++) { 1283 /* Replicate the last channel out. */ 1284 swizzle[i] = swizzle[ir->type->vector_elements - 1]; 1285 } 1286 1287 src.swizzle = BRW_SWIZZLE4(swizzle[0], swizzle[1], swizzle[2], swizzle[3]); 1288 1289 this->result = src; 1290} 1291 1292void 1293vec4_visitor::visit(ir_dereference_variable *ir) 1294{ 1295 const struct glsl_type *type = ir->type; 1296 dst_reg *reg = variable_storage(ir->var); 1297 1298 if (!reg) { 1299 fail("Failed to find variable storage for %s\n", ir->var->name); 1300 this->result = src_reg(brw_null_reg()); 1301 return; 1302 } 1303 1304 this->result = src_reg(*reg); 1305 1306 if (type->is_scalar() || type->is_vector() || type->is_matrix()) 1307 this->result.swizzle = swizzle_for_size(type->vector_elements); 1308} 1309 1310void 1311vec4_visitor::visit(ir_dereference_array *ir) 1312{ 1313 ir_constant *constant_index; 1314 src_reg src; 1315 int element_size = type_size(ir->type); 1316 1317 constant_index = ir->array_index->constant_expression_value(); 1318 1319 ir->array->accept(this); 1320 src = this->result; 1321 1322 if (constant_index) { 1323 src.reg_offset += constant_index->value.i[0] * element_size; 1324 } else { 1325 /* Variable index array dereference. It eats the "vec4" of the 1326 * base of the array and an index that offsets the Mesa register 1327 * index. 1328 */ 1329 ir->array_index->accept(this); 1330 1331 src_reg index_reg; 1332 1333 if (element_size == 1) { 1334 index_reg = this->result; 1335 } else { 1336 index_reg = src_reg(this, glsl_type::int_type); 1337 1338 emit(MUL(dst_reg(index_reg), this->result, src_reg(element_size))); 1339 } 1340 1341 if (src.reladdr) { 1342 src_reg temp = src_reg(this, glsl_type::int_type); 1343 1344 emit(ADD(dst_reg(temp), *src.reladdr, index_reg)); 1345 1346 index_reg = temp; 1347 } 1348 1349 src.reladdr = ralloc(mem_ctx, src_reg); 1350 memcpy(src.reladdr, &index_reg, sizeof(index_reg)); 1351 } 1352 1353 /* If the type is smaller than a vec4, replicate the last channel out. */ 1354 if (ir->type->is_scalar() || ir->type->is_vector()) 1355 src.swizzle = swizzle_for_size(ir->type->vector_elements); 1356 else 1357 src.swizzle = BRW_SWIZZLE_NOOP; 1358 src.type = brw_type_for_base_type(ir->type); 1359 1360 this->result = src; 1361} 1362 1363void 1364vec4_visitor::visit(ir_dereference_record *ir) 1365{ 1366 unsigned int i; 1367 const glsl_type *struct_type = ir->record->type; 1368 int offset = 0; 1369 1370 ir->record->accept(this); 1371 1372 for (i = 0; i < struct_type->length; i++) { 1373 if (strcmp(struct_type->fields.structure[i].name, ir->field) == 0) 1374 break; 1375 offset += type_size(struct_type->fields.structure[i].type); 1376 } 1377 1378 /* If the type is smaller than a vec4, replicate the last channel out. */ 1379 if (ir->type->is_scalar() || ir->type->is_vector()) 1380 this->result.swizzle = swizzle_for_size(ir->type->vector_elements); 1381 else 1382 this->result.swizzle = BRW_SWIZZLE_NOOP; 1383 this->result.type = brw_type_for_base_type(ir->type); 1384 1385 this->result.reg_offset += offset; 1386} 1387 1388/** 1389 * We want to be careful in assignment setup to hit the actual storage 1390 * instead of potentially using a temporary like we might with the 1391 * ir_dereference handler. 1392 */ 1393static dst_reg 1394get_assignment_lhs(ir_dereference *ir, vec4_visitor *v) 1395{ 1396 /* The LHS must be a dereference. If the LHS is a variable indexed array 1397 * access of a vector, it must be separated into a series conditional moves 1398 * before reaching this point (see ir_vec_index_to_cond_assign). 1399 */ 1400 assert(ir->as_dereference()); 1401 ir_dereference_array *deref_array = ir->as_dereference_array(); 1402 if (deref_array) { 1403 assert(!deref_array->array->type->is_vector()); 1404 } 1405 1406 /* Use the rvalue deref handler for the most part. We'll ignore 1407 * swizzles in it and write swizzles using writemask, though. 1408 */ 1409 ir->accept(v); 1410 return dst_reg(v->result); 1411} 1412 1413void 1414vec4_visitor::emit_block_move(dst_reg *dst, src_reg *src, 1415 const struct glsl_type *type, uint32_t predicate) 1416{ 1417 if (type->base_type == GLSL_TYPE_STRUCT) { 1418 for (unsigned int i = 0; i < type->length; i++) { 1419 emit_block_move(dst, src, type->fields.structure[i].type, predicate); 1420 } 1421 return; 1422 } 1423 1424 if (type->is_array()) { 1425 for (unsigned int i = 0; i < type->length; i++) { 1426 emit_block_move(dst, src, type->fields.array, predicate); 1427 } 1428 return; 1429 } 1430 1431 if (type->is_matrix()) { 1432 const struct glsl_type *vec_type; 1433 1434 vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT, 1435 type->vector_elements, 1); 1436 1437 for (int i = 0; i < type->matrix_columns; i++) { 1438 emit_block_move(dst, src, vec_type, predicate); 1439 } 1440 return; 1441 } 1442 1443 assert(type->is_scalar() || type->is_vector()); 1444 1445 dst->type = brw_type_for_base_type(type); 1446 src->type = dst->type; 1447 1448 dst->writemask = (1 << type->vector_elements) - 1; 1449 1450 /* Do we need to worry about swizzling a swizzle? */ 1451 assert(src->swizzle = BRW_SWIZZLE_NOOP); 1452 src->swizzle = swizzle_for_size(type->vector_elements); 1453 1454 vec4_instruction *inst = emit(MOV(*dst, *src)); 1455 inst->predicate = predicate; 1456 1457 dst->reg_offset++; 1458 src->reg_offset++; 1459} 1460 1461 1462/* If the RHS processing resulted in an instruction generating a 1463 * temporary value, and it would be easy to rewrite the instruction to 1464 * generate its result right into the LHS instead, do so. This ends 1465 * up reliably removing instructions where it can be tricky to do so 1466 * later without real UD chain information. 1467 */ 1468bool 1469vec4_visitor::try_rewrite_rhs_to_dst(ir_assignment *ir, 1470 dst_reg dst, 1471 src_reg src, 1472 vec4_instruction *pre_rhs_inst, 1473 vec4_instruction *last_rhs_inst) 1474{ 1475 /* This could be supported, but it would take more smarts. */ 1476 if (ir->condition) 1477 return false; 1478 1479 if (pre_rhs_inst == last_rhs_inst) 1480 return false; /* No instructions generated to work with. */ 1481 1482 /* Make sure the last instruction generated our source reg. */ 1483 if (src.file != GRF || 1484 src.file != last_rhs_inst->dst.file || 1485 src.reg != last_rhs_inst->dst.reg || 1486 src.reg_offset != last_rhs_inst->dst.reg_offset || 1487 src.reladdr || 1488 src.abs || 1489 src.negate || 1490 last_rhs_inst->predicate != BRW_PREDICATE_NONE) 1491 return false; 1492 1493 /* Check that that last instruction fully initialized the channels 1494 * we want to use, in the order we want to use them. We could 1495 * potentially reswizzle the operands of many instructions so that 1496 * we could handle out of order channels, but don't yet. 1497 */ 1498 for (int i = 0; i < 4; i++) { 1499 if (dst.writemask & (1 << i)) { 1500 if (!(last_rhs_inst->dst.writemask & (1 << i))) 1501 return false; 1502 1503 if (BRW_GET_SWZ(src.swizzle, i) != i) 1504 return false; 1505 } 1506 } 1507 1508 /* Success! Rewrite the instruction. */ 1509 last_rhs_inst->dst.file = dst.file; 1510 last_rhs_inst->dst.reg = dst.reg; 1511 last_rhs_inst->dst.reg_offset = dst.reg_offset; 1512 last_rhs_inst->dst.reladdr = dst.reladdr; 1513 last_rhs_inst->dst.writemask &= dst.writemask; 1514 1515 return true; 1516} 1517 1518void 1519vec4_visitor::visit(ir_assignment *ir) 1520{ 1521 dst_reg dst = get_assignment_lhs(ir->lhs, this); 1522 uint32_t predicate = BRW_PREDICATE_NONE; 1523 1524 if (!ir->lhs->type->is_scalar() && 1525 !ir->lhs->type->is_vector()) { 1526 ir->rhs->accept(this); 1527 src_reg src = this->result; 1528 1529 if (ir->condition) { 1530 emit_bool_to_cond_code(ir->condition, &predicate); 1531 } 1532 1533 emit_block_move(&dst, &src, ir->rhs->type, predicate); 1534 return; 1535 } 1536 1537 /* Now we're down to just a scalar/vector with writemasks. */ 1538 int i; 1539 1540 vec4_instruction *pre_rhs_inst, *last_rhs_inst; 1541 pre_rhs_inst = (vec4_instruction *)this->instructions.get_tail(); 1542 1543 ir->rhs->accept(this); 1544 1545 last_rhs_inst = (vec4_instruction *)this->instructions.get_tail(); 1546 1547 src_reg src = this->result; 1548 1549 int swizzles[4]; 1550 int first_enabled_chan = 0; 1551 int src_chan = 0; 1552 1553 assert(ir->lhs->type->is_vector() || 1554 ir->lhs->type->is_scalar()); 1555 dst.writemask = ir->write_mask; 1556 1557 for (int i = 0; i < 4; i++) { 1558 if (dst.writemask & (1 << i)) { 1559 first_enabled_chan = BRW_GET_SWZ(src.swizzle, i); 1560 break; 1561 } 1562 } 1563 1564 /* Swizzle a small RHS vector into the channels being written. 1565 * 1566 * glsl ir treats write_mask as dictating how many channels are 1567 * present on the RHS while in our instructions we need to make 1568 * those channels appear in the slots of the vec4 they're written to. 1569 */ 1570 for (int i = 0; i < 4; i++) { 1571 if (dst.writemask & (1 << i)) 1572 swizzles[i] = BRW_GET_SWZ(src.swizzle, src_chan++); 1573 else 1574 swizzles[i] = first_enabled_chan; 1575 } 1576 src.swizzle = BRW_SWIZZLE4(swizzles[0], swizzles[1], 1577 swizzles[2], swizzles[3]); 1578 1579 if (try_rewrite_rhs_to_dst(ir, dst, src, pre_rhs_inst, last_rhs_inst)) { 1580 return; 1581 } 1582 1583 if (ir->condition) { 1584 emit_bool_to_cond_code(ir->condition, &predicate); 1585 } 1586 1587 for (i = 0; i < type_size(ir->lhs->type); i++) { 1588 vec4_instruction *inst = emit(MOV(dst, src)); 1589 inst->predicate = predicate; 1590 1591 dst.reg_offset++; 1592 src.reg_offset++; 1593 } 1594} 1595 1596void 1597vec4_visitor::emit_constant_values(dst_reg *dst, ir_constant *ir) 1598{ 1599 if (ir->type->base_type == GLSL_TYPE_STRUCT) { 1600 foreach_list(node, &ir->components) { 1601 ir_constant *field_value = (ir_constant *)node; 1602 1603 emit_constant_values(dst, field_value); 1604 } 1605 return; 1606 } 1607 1608 if (ir->type->is_array()) { 1609 for (unsigned int i = 0; i < ir->type->length; i++) { 1610 emit_constant_values(dst, ir->array_elements[i]); 1611 } 1612 return; 1613 } 1614 1615 if (ir->type->is_matrix()) { 1616 for (int i = 0; i < ir->type->matrix_columns; i++) { 1617 for (int j = 0; j < ir->type->vector_elements; j++) { 1618 dst->writemask = 1 << j; 1619 dst->type = BRW_REGISTER_TYPE_F; 1620 1621 emit(MOV(*dst, 1622 src_reg(ir->value.f[i * ir->type->vector_elements + j]))); 1623 } 1624 dst->reg_offset++; 1625 } 1626 return; 1627 } 1628 1629 for (int i = 0; i < ir->type->vector_elements; i++) { 1630 dst->writemask = 1 << i; 1631 dst->type = brw_type_for_base_type(ir->type); 1632 1633 switch (ir->type->base_type) { 1634 case GLSL_TYPE_FLOAT: 1635 emit(MOV(*dst, src_reg(ir->value.f[i]))); 1636 break; 1637 case GLSL_TYPE_INT: 1638 emit(MOV(*dst, src_reg(ir->value.i[i]))); 1639 break; 1640 case GLSL_TYPE_UINT: 1641 emit(MOV(*dst, src_reg(ir->value.u[i]))); 1642 break; 1643 case GLSL_TYPE_BOOL: 1644 emit(MOV(*dst, src_reg(ir->value.b[i]))); 1645 break; 1646 default: 1647 assert(!"Non-float/uint/int/bool constant"); 1648 break; 1649 } 1650 } 1651 dst->reg_offset++; 1652} 1653 1654void 1655vec4_visitor::visit(ir_constant *ir) 1656{ 1657 dst_reg dst = dst_reg(this, ir->type); 1658 this->result = src_reg(dst); 1659 1660 emit_constant_values(&dst, ir); 1661} 1662 1663void 1664vec4_visitor::visit(ir_call *ir) 1665{ 1666 assert(!"not reached"); 1667} 1668 1669void 1670vec4_visitor::visit(ir_texture *ir) 1671{ 1672 /* FINISHME: Implement vertex texturing. 1673 * 1674 * With 0 vertex samplers available, the linker will reject 1675 * programs that do vertex texturing, but after our visitor has 1676 * run. 1677 */ 1678 this->result = src_reg(this, glsl_type::vec4_type); 1679} 1680 1681void 1682vec4_visitor::visit(ir_return *ir) 1683{ 1684 assert(!"not reached"); 1685} 1686 1687void 1688vec4_visitor::visit(ir_discard *ir) 1689{ 1690 assert(!"not reached"); 1691} 1692 1693void 1694vec4_visitor::visit(ir_if *ir) 1695{ 1696 /* Don't point the annotation at the if statement, because then it plus 1697 * the then and else blocks get printed. 1698 */ 1699 this->base_ir = ir->condition; 1700 1701 if (intel->gen == 6) { 1702 emit_if_gen6(ir); 1703 } else { 1704 uint32_t predicate; 1705 emit_bool_to_cond_code(ir->condition, &predicate); 1706 emit(IF(predicate)); 1707 } 1708 1709 visit_instructions(&ir->then_instructions); 1710 1711 if (!ir->else_instructions.is_empty()) { 1712 this->base_ir = ir->condition; 1713 emit(BRW_OPCODE_ELSE); 1714 1715 visit_instructions(&ir->else_instructions); 1716 } 1717 1718 this->base_ir = ir->condition; 1719 emit(BRW_OPCODE_ENDIF); 1720} 1721 1722void 1723vec4_visitor::emit_ndc_computation() 1724{ 1725 /* Get the position */ 1726 src_reg pos = src_reg(output_reg[VERT_RESULT_HPOS]); 1727 1728 /* Build ndc coords, which are (x/w, y/w, z/w, 1/w) */ 1729 dst_reg ndc = dst_reg(this, glsl_type::vec4_type); 1730 output_reg[BRW_VERT_RESULT_NDC] = ndc; 1731 1732 current_annotation = "NDC"; 1733 dst_reg ndc_w = ndc; 1734 ndc_w.writemask = WRITEMASK_W; 1735 src_reg pos_w = pos; 1736 pos_w.swizzle = BRW_SWIZZLE4(SWIZZLE_W, SWIZZLE_W, SWIZZLE_W, SWIZZLE_W); 1737 emit_math(SHADER_OPCODE_RCP, ndc_w, pos_w); 1738 1739 dst_reg ndc_xyz = ndc; 1740 ndc_xyz.writemask = WRITEMASK_XYZ; 1741 1742 emit(MUL(ndc_xyz, pos, src_reg(ndc_w))); 1743} 1744 1745void 1746vec4_visitor::emit_psiz_and_flags(struct brw_reg reg) 1747{ 1748 if (intel->gen < 6 && 1749 ((c->prog_data.outputs_written & BITFIELD64_BIT(VERT_RESULT_PSIZ)) || 1750 c->key.nr_userclip || brw->has_negative_rhw_bug)) { 1751 dst_reg header1 = dst_reg(this, glsl_type::uvec4_type); 1752 GLuint i; 1753 1754 emit(MOV(header1, 0u)); 1755 1756 if (c->prog_data.outputs_written & BITFIELD64_BIT(VERT_RESULT_PSIZ)) { 1757 src_reg psiz = src_reg(output_reg[VERT_RESULT_PSIZ]); 1758 1759 current_annotation = "Point size"; 1760 header1.writemask = WRITEMASK_W; 1761 emit(MUL(header1, psiz, src_reg((float)(1 << 11)))); 1762 emit(AND(header1, src_reg(header1), 0x7ff << 8)); 1763 } 1764 1765 current_annotation = "Clipping flags"; 1766 for (i = 0; i < c->key.nr_userclip; i++) { 1767 vec4_instruction *inst; 1768 1769 inst = emit(DP4(dst_null_f(), src_reg(output_reg[VERT_RESULT_HPOS]), 1770 src_reg(c->userplane[i]))); 1771 inst->conditional_mod = BRW_CONDITIONAL_L; 1772 1773 emit(OR(header1, src_reg(header1), 1u << i)); 1774 inst->predicate = BRW_PREDICATE_NORMAL; 1775 } 1776 1777 /* i965 clipping workaround: 1778 * 1) Test for -ve rhw 1779 * 2) If set, 1780 * set ndc = (0,0,0,0) 1781 * set ucp[6] = 1 1782 * 1783 * Later, clipping will detect ucp[6] and ensure the primitive is 1784 * clipped against all fixed planes. 1785 */ 1786 if (brw->has_negative_rhw_bug) { 1787#if 0 1788 /* FINISHME */ 1789 brw_CMP(p, 1790 vec8(brw_null_reg()), 1791 BRW_CONDITIONAL_L, 1792 brw_swizzle1(output_reg[BRW_VERT_RESULT_NDC], 3), 1793 brw_imm_f(0)); 1794 1795 brw_OR(p, brw_writemask(header1, WRITEMASK_W), header1, brw_imm_ud(1<<6)); 1796 brw_MOV(p, output_reg[BRW_VERT_RESULT_NDC], brw_imm_f(0)); 1797 brw_set_predicate_control(p, BRW_PREDICATE_NONE); 1798#endif 1799 } 1800 1801 header1.writemask = WRITEMASK_XYZW; 1802 emit(MOV(retype(reg, BRW_REGISTER_TYPE_UD), src_reg(header1))); 1803 } else if (intel->gen < 6) { 1804 emit(MOV(retype(reg, BRW_REGISTER_TYPE_UD), 0u)); 1805 } else { 1806 emit(MOV(retype(reg, BRW_REGISTER_TYPE_D), src_reg(0))); 1807 if (c->prog_data.outputs_written & BITFIELD64_BIT(VERT_RESULT_PSIZ)) { 1808 emit(MOV(brw_writemask(reg, WRITEMASK_W), 1809 src_reg(output_reg[VERT_RESULT_PSIZ]))); 1810 } 1811 } 1812} 1813 1814void 1815vec4_visitor::emit_clip_distances(struct brw_reg reg, int offset) 1816{ 1817 if (intel->gen < 6) { 1818 /* Clip distance slots are set aside in gen5, but they are not used. It 1819 * is not clear whether we actually need to set aside space for them, 1820 * but the performance cost is negligible. 1821 */ 1822 return; 1823 } 1824 1825 for (int i = 0; i + offset < c->key.nr_userclip && i < 4; ++i) { 1826 emit(DP4(dst_reg(brw_writemask(reg, 1 << i)), 1827 src_reg(output_reg[VERT_RESULT_HPOS]), 1828 src_reg(c->userplane[i + offset]))); 1829 } 1830} 1831 1832void 1833vec4_visitor::emit_urb_slot(int mrf, int vert_result) 1834{ 1835 struct brw_reg hw_reg = brw_message_reg(mrf); 1836 dst_reg reg = dst_reg(MRF, mrf); 1837 reg.type = BRW_REGISTER_TYPE_F; 1838 1839 switch (vert_result) { 1840 case VERT_RESULT_PSIZ: 1841 /* PSIZ is always in slot 0, and is coupled with other flags. */ 1842 current_annotation = "indices, point width, clip flags"; 1843 emit_psiz_and_flags(hw_reg); 1844 break; 1845 case BRW_VERT_RESULT_NDC: 1846 current_annotation = "NDC"; 1847 emit(MOV(reg, src_reg(output_reg[BRW_VERT_RESULT_NDC]))); 1848 break; 1849 case BRW_VERT_RESULT_HPOS_DUPLICATE: 1850 case VERT_RESULT_HPOS: 1851 current_annotation = "gl_Position"; 1852 emit(MOV(reg, src_reg(output_reg[VERT_RESULT_HPOS]))); 1853 break; 1854 case BRW_VERT_RESULT_CLIP0: 1855 current_annotation = "user clip distances"; 1856 emit_clip_distances(hw_reg, 0); 1857 break; 1858 case BRW_VERT_RESULT_CLIP1: 1859 current_annotation = "user clip distances"; 1860 emit_clip_distances(hw_reg, 4); 1861 break; 1862 case BRW_VERT_RESULT_PAD: 1863 /* No need to write to this slot */ 1864 break; 1865 default: { 1866 assert (vert_result < VERT_RESULT_MAX); 1867 current_annotation = output_reg_annotation[vert_result]; 1868 /* Copy the register, saturating if necessary */ 1869 vec4_instruction *inst = emit(MOV(reg, 1870 src_reg(output_reg[vert_result]))); 1871 if ((vert_result == VERT_RESULT_COL0 || 1872 vert_result == VERT_RESULT_COL1 || 1873 vert_result == VERT_RESULT_BFC0 || 1874 vert_result == VERT_RESULT_BFC1) && 1875 c->key.clamp_vertex_color) { 1876 inst->saturate = true; 1877 } 1878 } 1879 break; 1880 } 1881} 1882 1883static int 1884align_interleaved_urb_mlen(struct brw_context *brw, int mlen) 1885{ 1886 struct intel_context *intel = &brw->intel; 1887 1888 if (intel->gen >= 6) { 1889 /* URB data written (does not include the message header reg) must 1890 * be a multiple of 256 bits, or 2 VS registers. See vol5c.5, 1891 * section 5.4.3.2.2: URB_INTERLEAVED. 1892 * 1893 * URB entries are allocated on a multiple of 1024 bits, so an 1894 * extra 128 bits written here to make the end align to 256 is 1895 * no problem. 1896 */ 1897 if ((mlen % 2) != 1) 1898 mlen++; 1899 } 1900 1901 return mlen; 1902} 1903 1904/** 1905 * Generates the VUE payload plus the 1 or 2 URB write instructions to 1906 * complete the VS thread. 1907 * 1908 * The VUE layout is documented in Volume 2a. 1909 */ 1910void 1911vec4_visitor::emit_urb_writes() 1912{ 1913 /* MRF 0 is reserved for the debugger, so start with message header 1914 * in MRF 1. 1915 */ 1916 int base_mrf = 1; 1917 int mrf = base_mrf; 1918 /* In the process of generating our URB write message contents, we 1919 * may need to unspill a register or load from an array. Those 1920 * reads would use MRFs 14-15. 1921 */ 1922 int max_usable_mrf = 13; 1923 1924 /* The following assertion verifies that max_usable_mrf causes an 1925 * even-numbered amount of URB write data, which will meet gen6's 1926 * requirements for length alignment. 1927 */ 1928 assert ((max_usable_mrf - base_mrf) % 2 == 0); 1929 1930 /* FINISHME: edgeflag */ 1931 1932 brw_compute_vue_map(&c->vue_map, intel, c->key.nr_userclip, 1933 c->prog_data.outputs_written); 1934 1935 /* First mrf is the g0-based message header containing URB handles and such, 1936 * which is implied in VS_OPCODE_URB_WRITE. 1937 */ 1938 mrf++; 1939 1940 if (intel->gen < 6) { 1941 emit_ndc_computation(); 1942 } 1943 1944 /* Set up the VUE data for the first URB write */ 1945 int slot; 1946 for (slot = 0; slot < c->vue_map.num_slots; ++slot) { 1947 emit_urb_slot(mrf++, c->vue_map.slot_to_vert_result[slot]); 1948 1949 /* If this was max_usable_mrf, we can't fit anything more into this URB 1950 * WRITE. 1951 */ 1952 if (mrf > max_usable_mrf) { 1953 slot++; 1954 break; 1955 } 1956 } 1957 1958 current_annotation = "URB write"; 1959 vec4_instruction *inst = emit(VS_OPCODE_URB_WRITE); 1960 inst->base_mrf = base_mrf; 1961 inst->mlen = align_interleaved_urb_mlen(brw, mrf - base_mrf); 1962 inst->eot = (slot >= c->vue_map.num_slots); 1963 1964 /* Optional second URB write */ 1965 if (!inst->eot) { 1966 mrf = base_mrf + 1; 1967 1968 for (; slot < c->vue_map.num_slots; ++slot) { 1969 assert(mrf < max_usable_mrf); 1970 1971 emit_urb_slot(mrf++, c->vue_map.slot_to_vert_result[slot]); 1972 } 1973 1974 current_annotation = "URB write"; 1975 inst = emit(VS_OPCODE_URB_WRITE); 1976 inst->base_mrf = base_mrf; 1977 inst->mlen = align_interleaved_urb_mlen(brw, mrf - base_mrf); 1978 inst->eot = true; 1979 /* URB destination offset. In the previous write, we got MRFs 1980 * 2-13 minus the one header MRF, so 12 regs. URB offset is in 1981 * URB row increments, and each of our MRFs is half of one of 1982 * those, since we're doing interleaved writes. 1983 */ 1984 inst->offset = (max_usable_mrf - base_mrf) / 2; 1985 } 1986 1987 if (intel->gen == 6) 1988 c->prog_data.urb_entry_size = ALIGN(c->vue_map.num_slots, 8) / 8; 1989 else 1990 c->prog_data.urb_entry_size = ALIGN(c->vue_map.num_slots, 4) / 4; 1991} 1992 1993src_reg 1994vec4_visitor::get_scratch_offset(vec4_instruction *inst, 1995 src_reg *reladdr, int reg_offset) 1996{ 1997 /* Because we store the values to scratch interleaved like our 1998 * vertex data, we need to scale the vec4 index by 2. 1999 */ 2000 int message_header_scale = 2; 2001 2002 /* Pre-gen6, the message header uses byte offsets instead of vec4 2003 * (16-byte) offset units. 2004 */ 2005 if (intel->gen < 6) 2006 message_header_scale *= 16; 2007 2008 if (reladdr) { 2009 src_reg index = src_reg(this, glsl_type::int_type); 2010 2011 emit_before(inst, ADD(dst_reg(index), *reladdr, src_reg(reg_offset))); 2012 emit_before(inst, MUL(dst_reg(index), 2013 index, src_reg(message_header_scale))); 2014 2015 return index; 2016 } else { 2017 return src_reg(reg_offset * message_header_scale); 2018 } 2019} 2020 2021src_reg 2022vec4_visitor::get_pull_constant_offset(vec4_instruction *inst, 2023 src_reg *reladdr, int reg_offset) 2024{ 2025 if (reladdr) { 2026 src_reg index = src_reg(this, glsl_type::int_type); 2027 2028 emit_before(inst, ADD(dst_reg(index), *reladdr, src_reg(reg_offset))); 2029 2030 /* Pre-gen6, the message header uses byte offsets instead of vec4 2031 * (16-byte) offset units. 2032 */ 2033 if (intel->gen < 6) { 2034 emit_before(inst, MUL(dst_reg(index), index, src_reg(16))); 2035 } 2036 2037 return index; 2038 } else { 2039 int message_header_scale = intel->gen < 6 ? 16 : 1; 2040 return src_reg(reg_offset * message_header_scale); 2041 } 2042} 2043 2044/** 2045 * Emits an instruction before @inst to load the value named by @orig_src 2046 * from scratch space at @base_offset to @temp. 2047 */ 2048void 2049vec4_visitor::emit_scratch_read(vec4_instruction *inst, 2050 dst_reg temp, src_reg orig_src, 2051 int base_offset) 2052{ 2053 int reg_offset = base_offset + orig_src.reg_offset; 2054 src_reg index = get_scratch_offset(inst, orig_src.reladdr, reg_offset); 2055 2056 emit_before(inst, SCRATCH_READ(temp, index)); 2057} 2058 2059/** 2060 * Emits an instruction after @inst to store the value to be written 2061 * to @orig_dst to scratch space at @base_offset, from @temp. 2062 */ 2063void 2064vec4_visitor::emit_scratch_write(vec4_instruction *inst, 2065 src_reg temp, dst_reg orig_dst, 2066 int base_offset) 2067{ 2068 int reg_offset = base_offset + orig_dst.reg_offset; 2069 src_reg index = get_scratch_offset(inst, orig_dst.reladdr, reg_offset); 2070 2071 dst_reg dst = dst_reg(brw_writemask(brw_vec8_grf(0, 0), 2072 orig_dst.writemask)); 2073 vec4_instruction *write = SCRATCH_WRITE(dst, temp, index); 2074 write->predicate = inst->predicate; 2075 write->ir = inst->ir; 2076 write->annotation = inst->annotation; 2077 inst->insert_after(write); 2078} 2079 2080/** 2081 * We can't generally support array access in GRF space, because a 2082 * single instruction's destination can only span 2 contiguous 2083 * registers. So, we send all GRF arrays that get variable index 2084 * access to scratch space. 2085 */ 2086void 2087vec4_visitor::move_grf_array_access_to_scratch() 2088{ 2089 int scratch_loc[this->virtual_grf_count]; 2090 2091 for (int i = 0; i < this->virtual_grf_count; i++) { 2092 scratch_loc[i] = -1; 2093 } 2094 2095 /* First, calculate the set of virtual GRFs that need to be punted 2096 * to scratch due to having any array access on them, and where in 2097 * scratch. 2098 */ 2099 foreach_list(node, &this->instructions) { 2100 vec4_instruction *inst = (vec4_instruction *)node; 2101 2102 if (inst->dst.file == GRF && inst->dst.reladdr && 2103 scratch_loc[inst->dst.reg] == -1) { 2104 scratch_loc[inst->dst.reg] = c->last_scratch; 2105 c->last_scratch += this->virtual_grf_sizes[inst->dst.reg] * 8 * 4; 2106 } 2107 2108 for (int i = 0 ; i < 3; i++) { 2109 src_reg *src = &inst->src[i]; 2110 2111 if (src->file == GRF && src->reladdr && 2112 scratch_loc[src->reg] == -1) { 2113 scratch_loc[src->reg] = c->last_scratch; 2114 c->last_scratch += this->virtual_grf_sizes[src->reg] * 8 * 4; 2115 } 2116 } 2117 } 2118 2119 /* Now, for anything that will be accessed through scratch, rewrite 2120 * it to load/store. Note that this is a _safe list walk, because 2121 * we may generate a new scratch_write instruction after the one 2122 * we're processing. 2123 */ 2124 foreach_list_safe(node, &this->instructions) { 2125 vec4_instruction *inst = (vec4_instruction *)node; 2126 2127 /* Set up the annotation tracking for new generated instructions. */ 2128 base_ir = inst->ir; 2129 current_annotation = inst->annotation; 2130 2131 if (inst->dst.file == GRF && scratch_loc[inst->dst.reg] != -1) { 2132 src_reg temp = src_reg(this, glsl_type::vec4_type); 2133 2134 emit_scratch_write(inst, temp, inst->dst, scratch_loc[inst->dst.reg]); 2135 2136 inst->dst.file = temp.file; 2137 inst->dst.reg = temp.reg; 2138 inst->dst.reg_offset = temp.reg_offset; 2139 inst->dst.reladdr = NULL; 2140 } 2141 2142 for (int i = 0 ; i < 3; i++) { 2143 if (inst->src[i].file != GRF || scratch_loc[inst->src[i].reg] == -1) 2144 continue; 2145 2146 dst_reg temp = dst_reg(this, glsl_type::vec4_type); 2147 2148 emit_scratch_read(inst, temp, inst->src[i], 2149 scratch_loc[inst->src[i].reg]); 2150 2151 inst->src[i].file = temp.file; 2152 inst->src[i].reg = temp.reg; 2153 inst->src[i].reg_offset = temp.reg_offset; 2154 inst->src[i].reladdr = NULL; 2155 } 2156 } 2157} 2158 2159/** 2160 * Emits an instruction before @inst to load the value named by @orig_src 2161 * from the pull constant buffer (surface) at @base_offset to @temp. 2162 */ 2163void 2164vec4_visitor::emit_pull_constant_load(vec4_instruction *inst, 2165 dst_reg temp, src_reg orig_src, 2166 int base_offset) 2167{ 2168 int reg_offset = base_offset + orig_src.reg_offset; 2169 src_reg index = get_pull_constant_offset(inst, orig_src.reladdr, reg_offset); 2170 vec4_instruction *load; 2171 2172 load = new(mem_ctx) vec4_instruction(this, VS_OPCODE_PULL_CONSTANT_LOAD, 2173 temp, index); 2174 load->base_mrf = 14; 2175 load->mlen = 1; 2176 emit_before(inst, load); 2177} 2178 2179/** 2180 * Implements array access of uniforms by inserting a 2181 * PULL_CONSTANT_LOAD instruction. 2182 * 2183 * Unlike temporary GRF array access (where we don't support it due to 2184 * the difficulty of doing relative addressing on instruction 2185 * destinations), we could potentially do array access of uniforms 2186 * that were loaded in GRF space as push constants. In real-world 2187 * usage we've seen, though, the arrays being used are always larger 2188 * than we could load as push constants, so just always move all 2189 * uniform array access out to a pull constant buffer. 2190 */ 2191void 2192vec4_visitor::move_uniform_array_access_to_pull_constants() 2193{ 2194 int pull_constant_loc[this->uniforms]; 2195 2196 for (int i = 0; i < this->uniforms; i++) { 2197 pull_constant_loc[i] = -1; 2198 } 2199 2200 /* Walk through and find array access of uniforms. Put a copy of that 2201 * uniform in the pull constant buffer. 2202 * 2203 * Note that we don't move constant-indexed accesses to arrays. No 2204 * testing has been done of the performance impact of this choice. 2205 */ 2206 foreach_list_safe(node, &this->instructions) { 2207 vec4_instruction *inst = (vec4_instruction *)node; 2208 2209 for (int i = 0 ; i < 3; i++) { 2210 if (inst->src[i].file != UNIFORM || !inst->src[i].reladdr) 2211 continue; 2212 2213 int uniform = inst->src[i].reg; 2214 2215 /* If this array isn't already present in the pull constant buffer, 2216 * add it. 2217 */ 2218 if (pull_constant_loc[uniform] == -1) { 2219 const float **values = &prog_data->param[uniform * 4]; 2220 2221 pull_constant_loc[uniform] = prog_data->nr_pull_params / 4; 2222 2223 for (int j = 0; j < uniform_size[uniform] * 4; j++) { 2224 prog_data->pull_param[prog_data->nr_pull_params++] = values[j]; 2225 } 2226 } 2227 2228 /* Set up the annotation tracking for new generated instructions. */ 2229 base_ir = inst->ir; 2230 current_annotation = inst->annotation; 2231 2232 dst_reg temp = dst_reg(this, glsl_type::vec4_type); 2233 2234 emit_pull_constant_load(inst, temp, inst->src[i], 2235 pull_constant_loc[uniform]); 2236 2237 inst->src[i].file = temp.file; 2238 inst->src[i].reg = temp.reg; 2239 inst->src[i].reg_offset = temp.reg_offset; 2240 inst->src[i].reladdr = NULL; 2241 } 2242 } 2243 2244 /* Now there are no accesses of the UNIFORM file with a reladdr, so 2245 * no need to track them as larger-than-vec4 objects. This will be 2246 * relied on in cutting out unused uniform vectors from push 2247 * constants. 2248 */ 2249 split_uniform_registers(); 2250} 2251 2252vec4_visitor::vec4_visitor(struct brw_vs_compile *c, 2253 struct gl_shader_program *prog, 2254 struct brw_shader *shader) 2255{ 2256 this->c = c; 2257 this->p = &c->func; 2258 this->brw = p->brw; 2259 this->intel = &brw->intel; 2260 this->ctx = &intel->ctx; 2261 this->prog = prog; 2262 this->shader = shader; 2263 2264 this->mem_ctx = ralloc_context(NULL); 2265 this->failed = false; 2266 2267 this->base_ir = NULL; 2268 this->current_annotation = NULL; 2269 2270 this->c = c; 2271 this->vp = prog->VertexProgram; 2272 this->prog_data = &c->prog_data; 2273 2274 this->variable_ht = hash_table_ctor(0, 2275 hash_table_pointer_hash, 2276 hash_table_pointer_compare); 2277 2278 this->virtual_grf_def = NULL; 2279 this->virtual_grf_use = NULL; 2280 this->virtual_grf_sizes = NULL; 2281 this->virtual_grf_count = 0; 2282 this->virtual_grf_reg_map = NULL; 2283 this->virtual_grf_reg_count = 0; 2284 this->virtual_grf_array_size = 0; 2285 this->live_intervals_valid = false; 2286 2287 this->uniforms = 0; 2288 2289 this->variable_ht = hash_table_ctor(0, 2290 hash_table_pointer_hash, 2291 hash_table_pointer_compare); 2292} 2293 2294vec4_visitor::~vec4_visitor() 2295{ 2296 ralloc_free(this->mem_ctx); 2297 hash_table_dtor(this->variable_ht); 2298} 2299 2300 2301void 2302vec4_visitor::fail(const char *format, ...) 2303{ 2304 va_list va; 2305 char *msg; 2306 2307 if (failed) 2308 return; 2309 2310 failed = true; 2311 2312 va_start(va, format); 2313 msg = ralloc_vasprintf(mem_ctx, format, va); 2314 va_end(va); 2315 msg = ralloc_asprintf(mem_ctx, "VS compile failed: %s\n", msg); 2316 2317 this->fail_msg = msg; 2318 2319 if (INTEL_DEBUG & DEBUG_VS) { 2320 fprintf(stderr, "%s", msg); 2321 } 2322} 2323 2324} /* namespace brw */ 2325