brw_vec4_visitor.cpp revision 2e712e41db0c0676e9f30fc73172c0e8de8d84d4
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 >= 6) { 317 return emit_math1_gen6(opcode, dst, src); 318 } else { 319 return emit_math1_gen4(opcode, dst, src); 320 } 321} 322 323void 324vec4_visitor::emit_math2_gen6(enum opcode opcode, 325 dst_reg dst, src_reg src0, src_reg src1) 326{ 327 src_reg expanded; 328 329 /* The gen6 math instruction ignores the source modifiers -- 330 * swizzle, abs, negate, and at least some parts of the register 331 * region description. Move the sources to temporaries to make it 332 * generally work. 333 */ 334 335 expanded = src_reg(this, glsl_type::vec4_type); 336 expanded.type = src0.type; 337 emit(MOV(dst_reg(expanded), src0)); 338 src0 = expanded; 339 340 expanded = src_reg(this, glsl_type::vec4_type); 341 expanded.type = src1.type; 342 emit(MOV(dst_reg(expanded), src1)); 343 src1 = expanded; 344 345 if (dst.writemask != WRITEMASK_XYZW) { 346 /* The gen6 math instruction must be align1, so we can't do 347 * writemasks. 348 */ 349 dst_reg temp_dst = dst_reg(this, glsl_type::vec4_type); 350 temp_dst.type = dst.type; 351 352 emit(opcode, temp_dst, src0, src1); 353 354 emit(MOV(dst, src_reg(temp_dst))); 355 } else { 356 emit(opcode, dst, src0, src1); 357 } 358} 359 360void 361vec4_visitor::emit_math2_gen4(enum opcode opcode, 362 dst_reg dst, src_reg src0, src_reg src1) 363{ 364 vec4_instruction *inst = emit(opcode, dst, src0, src1); 365 inst->base_mrf = 1; 366 inst->mlen = 2; 367} 368 369void 370vec4_visitor::emit_math(enum opcode opcode, 371 dst_reg dst, src_reg src0, src_reg src1) 372{ 373 switch (opcode) { 374 case SHADER_OPCODE_POW: 375 case SHADER_OPCODE_INT_QUOTIENT: 376 case SHADER_OPCODE_INT_REMAINDER: 377 break; 378 default: 379 assert(!"not reached: unsupported binary math opcode"); 380 return; 381 } 382 383 if (intel->gen >= 6) { 384 return emit_math2_gen6(opcode, dst, src0, src1); 385 } else { 386 return emit_math2_gen4(opcode, dst, src0, src1); 387 } 388} 389 390void 391vec4_visitor::visit_instructions(const exec_list *list) 392{ 393 foreach_list(node, list) { 394 ir_instruction *ir = (ir_instruction *)node; 395 396 base_ir = ir; 397 ir->accept(this); 398 } 399} 400 401 402static int 403type_size(const struct glsl_type *type) 404{ 405 unsigned int i; 406 int size; 407 408 switch (type->base_type) { 409 case GLSL_TYPE_UINT: 410 case GLSL_TYPE_INT: 411 case GLSL_TYPE_FLOAT: 412 case GLSL_TYPE_BOOL: 413 if (type->is_matrix()) { 414 return type->matrix_columns; 415 } else { 416 /* Regardless of size of vector, it gets a vec4. This is bad 417 * packing for things like floats, but otherwise arrays become a 418 * mess. Hopefully a later pass over the code can pack scalars 419 * down if appropriate. 420 */ 421 return 1; 422 } 423 case GLSL_TYPE_ARRAY: 424 assert(type->length > 0); 425 return type_size(type->fields.array) * type->length; 426 case GLSL_TYPE_STRUCT: 427 size = 0; 428 for (i = 0; i < type->length; i++) { 429 size += type_size(type->fields.structure[i].type); 430 } 431 return size; 432 case GLSL_TYPE_SAMPLER: 433 /* Samplers take up one slot in UNIFORMS[], but they're baked in 434 * at link time. 435 */ 436 return 1; 437 default: 438 assert(0); 439 return 0; 440 } 441} 442 443int 444vec4_visitor::virtual_grf_alloc(int size) 445{ 446 if (virtual_grf_array_size <= virtual_grf_count) { 447 if (virtual_grf_array_size == 0) 448 virtual_grf_array_size = 16; 449 else 450 virtual_grf_array_size *= 2; 451 virtual_grf_sizes = reralloc(mem_ctx, virtual_grf_sizes, int, 452 virtual_grf_array_size); 453 virtual_grf_reg_map = reralloc(mem_ctx, virtual_grf_reg_map, int, 454 virtual_grf_array_size); 455 } 456 virtual_grf_reg_map[virtual_grf_count] = virtual_grf_reg_count; 457 virtual_grf_reg_count += size; 458 virtual_grf_sizes[virtual_grf_count] = size; 459 return virtual_grf_count++; 460} 461 462src_reg::src_reg(class vec4_visitor *v, const struct glsl_type *type) 463{ 464 init(); 465 466 this->file = GRF; 467 this->reg = v->virtual_grf_alloc(type_size(type)); 468 469 if (type->is_array() || type->is_record()) { 470 this->swizzle = BRW_SWIZZLE_NOOP; 471 } else { 472 this->swizzle = swizzle_for_size(type->vector_elements); 473 } 474 475 this->type = brw_type_for_base_type(type); 476} 477 478dst_reg::dst_reg(class vec4_visitor *v, const struct glsl_type *type) 479{ 480 init(); 481 482 this->file = GRF; 483 this->reg = v->virtual_grf_alloc(type_size(type)); 484 485 if (type->is_array() || type->is_record()) { 486 this->writemask = WRITEMASK_XYZW; 487 } else { 488 this->writemask = (1 << type->vector_elements) - 1; 489 } 490 491 this->type = brw_type_for_base_type(type); 492} 493 494/* Our support for uniforms is piggy-backed on the struct 495 * gl_fragment_program, because that's where the values actually 496 * get stored, rather than in some global gl_shader_program uniform 497 * store. 498 */ 499int 500vec4_visitor::setup_uniform_values(int loc, const glsl_type *type) 501{ 502 unsigned int offset = 0; 503 float *values = &this->vp->Base.Parameters->ParameterValues[loc][0].f; 504 505 if (type->is_matrix()) { 506 const glsl_type *column = type->column_type(); 507 508 for (unsigned int i = 0; i < type->matrix_columns; i++) { 509 offset += setup_uniform_values(loc + offset, column); 510 } 511 512 return offset; 513 } 514 515 switch (type->base_type) { 516 case GLSL_TYPE_FLOAT: 517 case GLSL_TYPE_UINT: 518 case GLSL_TYPE_INT: 519 case GLSL_TYPE_BOOL: 520 for (unsigned int i = 0; i < type->vector_elements; i++) { 521 c->prog_data.param[this->uniforms * 4 + i] = &values[i]; 522 } 523 524 /* Set up pad elements to get things aligned to a vec4 boundary. */ 525 for (unsigned int i = type->vector_elements; i < 4; i++) { 526 static float zero = 0; 527 528 c->prog_data.param[this->uniforms * 4 + i] = &zero; 529 } 530 531 /* Track the size of this uniform vector, for future packing of 532 * uniforms. 533 */ 534 this->uniform_vector_size[this->uniforms] = type->vector_elements; 535 this->uniforms++; 536 537 return 1; 538 539 case GLSL_TYPE_STRUCT: 540 for (unsigned int i = 0; i < type->length; i++) { 541 offset += setup_uniform_values(loc + offset, 542 type->fields.structure[i].type); 543 } 544 return offset; 545 546 case GLSL_TYPE_ARRAY: 547 for (unsigned int i = 0; i < type->length; i++) { 548 offset += setup_uniform_values(loc + offset, type->fields.array); 549 } 550 return offset; 551 552 case GLSL_TYPE_SAMPLER: 553 /* The sampler takes up a slot, but we don't use any values from it. */ 554 return 1; 555 556 default: 557 assert(!"not reached"); 558 return 0; 559 } 560} 561 562void 563vec4_visitor::setup_uniform_clipplane_values() 564{ 565 gl_clip_plane *clip_planes = brw_select_clip_planes(ctx); 566 567 /* Pre-Gen6, we compact clip planes. For example, if the user 568 * enables just clip planes 0, 1, and 3, we will enable clip planes 569 * 0, 1, and 2 in the hardware, and we'll move clip plane 3 to clip 570 * plane 2. This simplifies the implementation of the Gen6 clip 571 * thread. 572 * 573 * In Gen6 and later, we don't compact clip planes, because this 574 * simplifies the implementation of gl_ClipDistance. 575 */ 576 int compacted_clipplane_index = 0; 577 for (int i = 0; i < c->key.nr_userclip_plane_consts; ++i) { 578 if (intel->gen < 6 && 579 !(c->key.userclip_planes_enabled_gen_4_5 & (1 << i))) { 580 continue; 581 } 582 this->uniform_vector_size[this->uniforms] = 4; 583 this->userplane[compacted_clipplane_index] = dst_reg(UNIFORM, this->uniforms); 584 this->userplane[compacted_clipplane_index].type = BRW_REGISTER_TYPE_F; 585 for (int j = 0; j < 4; ++j) { 586 c->prog_data.param[this->uniforms * 4 + j] = &clip_planes[i][j]; 587 } 588 ++compacted_clipplane_index; 589 ++this->uniforms; 590 } 591} 592 593/* Our support for builtin uniforms is even scarier than non-builtin. 594 * It sits on top of the PROG_STATE_VAR parameters that are 595 * automatically updated from GL context state. 596 */ 597void 598vec4_visitor::setup_builtin_uniform_values(ir_variable *ir) 599{ 600 const ir_state_slot *const slots = ir->state_slots; 601 assert(ir->state_slots != NULL); 602 603 for (unsigned int i = 0; i < ir->num_state_slots; i++) { 604 /* This state reference has already been setup by ir_to_mesa, 605 * but we'll get the same index back here. We can reference 606 * ParameterValues directly, since unlike brw_fs.cpp, we never 607 * add new state references during compile. 608 */ 609 int index = _mesa_add_state_reference(this->vp->Base.Parameters, 610 (gl_state_index *)slots[i].tokens); 611 float *values = &this->vp->Base.Parameters->ParameterValues[index][0].f; 612 613 this->uniform_vector_size[this->uniforms] = 0; 614 /* Add each of the unique swizzled channels of the element. 615 * This will end up matching the size of the glsl_type of this field. 616 */ 617 int last_swiz = -1; 618 for (unsigned int j = 0; j < 4; j++) { 619 int swiz = GET_SWZ(slots[i].swizzle, j); 620 last_swiz = swiz; 621 622 c->prog_data.param[this->uniforms * 4 + j] = &values[swiz]; 623 if (swiz <= last_swiz) 624 this->uniform_vector_size[this->uniforms]++; 625 } 626 this->uniforms++; 627 } 628} 629 630dst_reg * 631vec4_visitor::variable_storage(ir_variable *var) 632{ 633 return (dst_reg *)hash_table_find(this->variable_ht, var); 634} 635 636void 637vec4_visitor::emit_bool_to_cond_code(ir_rvalue *ir, uint32_t *predicate) 638{ 639 ir_expression *expr = ir->as_expression(); 640 641 *predicate = BRW_PREDICATE_NORMAL; 642 643 if (expr) { 644 src_reg op[2]; 645 vec4_instruction *inst; 646 647 assert(expr->get_num_operands() <= 2); 648 for (unsigned int i = 0; i < expr->get_num_operands(); i++) { 649 expr->operands[i]->accept(this); 650 op[i] = this->result; 651 652 resolve_ud_negate(&op[i]); 653 } 654 655 switch (expr->operation) { 656 case ir_unop_logic_not: 657 inst = emit(AND(dst_null_d(), op[0], src_reg(1))); 658 inst->conditional_mod = BRW_CONDITIONAL_Z; 659 break; 660 661 case ir_binop_logic_xor: 662 inst = emit(XOR(dst_null_d(), op[0], op[1])); 663 inst->conditional_mod = BRW_CONDITIONAL_NZ; 664 break; 665 666 case ir_binop_logic_or: 667 inst = emit(OR(dst_null_d(), op[0], op[1])); 668 inst->conditional_mod = BRW_CONDITIONAL_NZ; 669 break; 670 671 case ir_binop_logic_and: 672 inst = emit(AND(dst_null_d(), op[0], op[1])); 673 inst->conditional_mod = BRW_CONDITIONAL_NZ; 674 break; 675 676 case ir_unop_f2b: 677 if (intel->gen >= 6) { 678 emit(CMP(dst_null_d(), op[0], src_reg(0.0f), BRW_CONDITIONAL_NZ)); 679 } else { 680 inst = emit(MOV(dst_null_f(), op[0])); 681 inst->conditional_mod = BRW_CONDITIONAL_NZ; 682 } 683 break; 684 685 case ir_unop_i2b: 686 if (intel->gen >= 6) { 687 emit(CMP(dst_null_d(), op[0], src_reg(0), BRW_CONDITIONAL_NZ)); 688 } else { 689 inst = emit(MOV(dst_null_d(), op[0])); 690 inst->conditional_mod = BRW_CONDITIONAL_NZ; 691 } 692 break; 693 694 case ir_binop_all_equal: 695 inst = emit(CMP(dst_null_d(), op[0], op[1], BRW_CONDITIONAL_Z)); 696 *predicate = BRW_PREDICATE_ALIGN16_ALL4H; 697 break; 698 699 case ir_binop_any_nequal: 700 inst = emit(CMP(dst_null_d(), op[0], op[1], BRW_CONDITIONAL_NZ)); 701 *predicate = BRW_PREDICATE_ALIGN16_ANY4H; 702 break; 703 704 case ir_unop_any: 705 inst = emit(CMP(dst_null_d(), op[0], src_reg(0), BRW_CONDITIONAL_NZ)); 706 *predicate = BRW_PREDICATE_ALIGN16_ANY4H; 707 break; 708 709 case ir_binop_greater: 710 case ir_binop_gequal: 711 case ir_binop_less: 712 case ir_binop_lequal: 713 case ir_binop_equal: 714 case ir_binop_nequal: 715 emit(CMP(dst_null_d(), op[0], op[1], 716 brw_conditional_for_comparison(expr->operation))); 717 break; 718 719 default: 720 assert(!"not reached"); 721 break; 722 } 723 return; 724 } 725 726 ir->accept(this); 727 728 resolve_ud_negate(&this->result); 729 730 if (intel->gen >= 6) { 731 vec4_instruction *inst = emit(AND(dst_null_d(), 732 this->result, src_reg(1))); 733 inst->conditional_mod = BRW_CONDITIONAL_NZ; 734 } else { 735 vec4_instruction *inst = emit(MOV(dst_null_d(), this->result)); 736 inst->conditional_mod = BRW_CONDITIONAL_NZ; 737 } 738} 739 740/** 741 * Emit a gen6 IF statement with the comparison folded into the IF 742 * instruction. 743 */ 744void 745vec4_visitor::emit_if_gen6(ir_if *ir) 746{ 747 ir_expression *expr = ir->condition->as_expression(); 748 749 if (expr) { 750 src_reg op[2]; 751 dst_reg temp; 752 753 assert(expr->get_num_operands() <= 2); 754 for (unsigned int i = 0; i < expr->get_num_operands(); i++) { 755 expr->operands[i]->accept(this); 756 op[i] = this->result; 757 } 758 759 switch (expr->operation) { 760 case ir_unop_logic_not: 761 emit(IF(op[0], src_reg(0), BRW_CONDITIONAL_Z)); 762 return; 763 764 case ir_binop_logic_xor: 765 emit(IF(op[0], op[1], BRW_CONDITIONAL_NZ)); 766 return; 767 768 case ir_binop_logic_or: 769 temp = dst_reg(this, glsl_type::bool_type); 770 emit(OR(temp, op[0], op[1])); 771 emit(IF(src_reg(temp), src_reg(0), BRW_CONDITIONAL_NZ)); 772 return; 773 774 case ir_binop_logic_and: 775 temp = dst_reg(this, glsl_type::bool_type); 776 emit(AND(temp, op[0], op[1])); 777 emit(IF(src_reg(temp), src_reg(0), BRW_CONDITIONAL_NZ)); 778 return; 779 780 case ir_unop_f2b: 781 emit(IF(op[0], src_reg(0), BRW_CONDITIONAL_NZ)); 782 return; 783 784 case ir_unop_i2b: 785 emit(IF(op[0], src_reg(0), BRW_CONDITIONAL_NZ)); 786 return; 787 788 case ir_binop_greater: 789 case ir_binop_gequal: 790 case ir_binop_less: 791 case ir_binop_lequal: 792 case ir_binop_equal: 793 case ir_binop_nequal: 794 emit(IF(op[0], op[1], 795 brw_conditional_for_comparison(expr->operation))); 796 return; 797 798 case ir_binop_all_equal: 799 emit(CMP(dst_null_d(), op[0], op[1], BRW_CONDITIONAL_Z)); 800 emit(IF(BRW_PREDICATE_ALIGN16_ALL4H)); 801 return; 802 803 case ir_binop_any_nequal: 804 emit(CMP(dst_null_d(), op[0], op[1], BRW_CONDITIONAL_NZ)); 805 emit(IF(BRW_PREDICATE_ALIGN16_ANY4H)); 806 return; 807 808 case ir_unop_any: 809 emit(CMP(dst_null_d(), op[0], src_reg(0), BRW_CONDITIONAL_NZ)); 810 emit(IF(BRW_PREDICATE_ALIGN16_ANY4H)); 811 return; 812 813 default: 814 assert(!"not reached"); 815 emit(IF(op[0], src_reg(0), BRW_CONDITIONAL_NZ)); 816 return; 817 } 818 return; 819 } 820 821 ir->condition->accept(this); 822 823 emit(IF(this->result, src_reg(0), BRW_CONDITIONAL_NZ)); 824} 825 826void 827vec4_visitor::visit(ir_variable *ir) 828{ 829 dst_reg *reg = NULL; 830 831 if (variable_storage(ir)) 832 return; 833 834 switch (ir->mode) { 835 case ir_var_in: 836 reg = new(mem_ctx) dst_reg(ATTR, ir->location); 837 838 /* Do GL_FIXED rescaling for GLES2.0. Our GL_FIXED attributes 839 * come in as floating point conversions of the integer values. 840 */ 841 for (int i = ir->location; i < ir->location + type_size(ir->type); i++) { 842 if (!c->key.gl_fixed_input_size[i]) 843 continue; 844 845 dst_reg dst = *reg; 846 dst.type = brw_type_for_base_type(ir->type); 847 dst.writemask = (1 << c->key.gl_fixed_input_size[i]) - 1; 848 emit(MUL(dst, src_reg(dst), src_reg(1.0f / 65536.0f))); 849 } 850 break; 851 852 case ir_var_out: 853 reg = new(mem_ctx) dst_reg(this, ir->type); 854 855 for (int i = 0; i < type_size(ir->type); i++) { 856 output_reg[ir->location + i] = *reg; 857 output_reg[ir->location + i].reg_offset = i; 858 output_reg[ir->location + i].type = 859 brw_type_for_base_type(ir->type->get_scalar_type()); 860 output_reg_annotation[ir->location + i] = ir->name; 861 } 862 break; 863 864 case ir_var_auto: 865 case ir_var_temporary: 866 reg = new(mem_ctx) dst_reg(this, ir->type); 867 break; 868 869 case ir_var_uniform: 870 reg = new(this->mem_ctx) dst_reg(UNIFORM, this->uniforms); 871 872 /* Track how big the whole uniform variable is, in case we need to put a 873 * copy of its data into pull constants for array access. 874 */ 875 this->uniform_size[this->uniforms] = type_size(ir->type); 876 877 if (!strncmp(ir->name, "gl_", 3)) { 878 setup_builtin_uniform_values(ir); 879 } else { 880 setup_uniform_values(ir->location, ir->type); 881 } 882 break; 883 884 case ir_var_system_value: 885 /* VertexID is stored by the VF as the last vertex element, but 886 * we don't represent it with a flag in inputs_read, so we call 887 * it VERT_ATTRIB_MAX, which setup_attributes() picks up on. 888 */ 889 reg = new(mem_ctx) dst_reg(ATTR, VERT_ATTRIB_MAX); 890 prog_data->uses_vertexid = true; 891 892 switch (ir->location) { 893 case SYSTEM_VALUE_VERTEX_ID: 894 reg->writemask = WRITEMASK_X; 895 break; 896 case SYSTEM_VALUE_INSTANCE_ID: 897 reg->writemask = WRITEMASK_Y; 898 break; 899 default: 900 assert(!"not reached"); 901 break; 902 } 903 break; 904 905 default: 906 assert(!"not reached"); 907 } 908 909 reg->type = brw_type_for_base_type(ir->type); 910 hash_table_insert(this->variable_ht, reg, ir); 911} 912 913void 914vec4_visitor::visit(ir_loop *ir) 915{ 916 dst_reg counter; 917 918 /* We don't want debugging output to print the whole body of the 919 * loop as the annotation. 920 */ 921 this->base_ir = NULL; 922 923 if (ir->counter != NULL) { 924 this->base_ir = ir->counter; 925 ir->counter->accept(this); 926 counter = *(variable_storage(ir->counter)); 927 928 if (ir->from != NULL) { 929 this->base_ir = ir->from; 930 ir->from->accept(this); 931 932 emit(MOV(counter, this->result)); 933 } 934 } 935 936 emit(BRW_OPCODE_DO); 937 938 if (ir->to) { 939 this->base_ir = ir->to; 940 ir->to->accept(this); 941 942 emit(CMP(dst_null_d(), src_reg(counter), this->result, 943 brw_conditional_for_comparison(ir->cmp))); 944 945 vec4_instruction *inst = emit(BRW_OPCODE_BREAK); 946 inst->predicate = BRW_PREDICATE_NORMAL; 947 } 948 949 visit_instructions(&ir->body_instructions); 950 951 952 if (ir->increment) { 953 this->base_ir = ir->increment; 954 ir->increment->accept(this); 955 emit(ADD(counter, src_reg(counter), this->result)); 956 } 957 958 emit(BRW_OPCODE_WHILE); 959} 960 961void 962vec4_visitor::visit(ir_loop_jump *ir) 963{ 964 switch (ir->mode) { 965 case ir_loop_jump::jump_break: 966 emit(BRW_OPCODE_BREAK); 967 break; 968 case ir_loop_jump::jump_continue: 969 emit(BRW_OPCODE_CONTINUE); 970 break; 971 } 972} 973 974 975void 976vec4_visitor::visit(ir_function_signature *ir) 977{ 978 assert(0); 979 (void)ir; 980} 981 982void 983vec4_visitor::visit(ir_function *ir) 984{ 985 /* Ignore function bodies other than main() -- we shouldn't see calls to 986 * them since they should all be inlined. 987 */ 988 if (strcmp(ir->name, "main") == 0) { 989 const ir_function_signature *sig; 990 exec_list empty; 991 992 sig = ir->matching_signature(&empty); 993 994 assert(sig); 995 996 visit_instructions(&sig->body); 997 } 998} 999 1000bool 1001vec4_visitor::try_emit_sat(ir_expression *ir) 1002{ 1003 ir_rvalue *sat_src = ir->as_rvalue_to_saturate(); 1004 if (!sat_src) 1005 return false; 1006 1007 sat_src->accept(this); 1008 src_reg src = this->result; 1009 1010 this->result = src_reg(this, ir->type); 1011 vec4_instruction *inst; 1012 inst = emit(MOV(dst_reg(this->result), src)); 1013 inst->saturate = true; 1014 1015 return true; 1016} 1017 1018void 1019vec4_visitor::emit_bool_comparison(unsigned int op, 1020 dst_reg dst, src_reg src0, src_reg src1) 1021{ 1022 /* original gen4 does destination conversion before comparison. */ 1023 if (intel->gen < 5) 1024 dst.type = src0.type; 1025 1026 emit(CMP(dst, src0, src1, brw_conditional_for_comparison(op))); 1027 1028 dst.type = BRW_REGISTER_TYPE_D; 1029 emit(AND(dst, src_reg(dst), src_reg(0x1))); 1030} 1031 1032void 1033vec4_visitor::visit(ir_expression *ir) 1034{ 1035 unsigned int operand; 1036 src_reg op[Elements(ir->operands)]; 1037 src_reg result_src; 1038 dst_reg result_dst; 1039 vec4_instruction *inst; 1040 1041 if (try_emit_sat(ir)) 1042 return; 1043 1044 for (operand = 0; operand < ir->get_num_operands(); operand++) { 1045 this->result.file = BAD_FILE; 1046 ir->operands[operand]->accept(this); 1047 if (this->result.file == BAD_FILE) { 1048 printf("Failed to get tree for expression operand:\n"); 1049 ir->operands[operand]->print(); 1050 exit(1); 1051 } 1052 op[operand] = this->result; 1053 1054 /* Matrix expression operands should have been broken down to vector 1055 * operations already. 1056 */ 1057 assert(!ir->operands[operand]->type->is_matrix()); 1058 } 1059 1060 int vector_elements = ir->operands[0]->type->vector_elements; 1061 if (ir->operands[1]) { 1062 vector_elements = MAX2(vector_elements, 1063 ir->operands[1]->type->vector_elements); 1064 } 1065 1066 this->result.file = BAD_FILE; 1067 1068 /* Storage for our result. Ideally for an assignment we'd be using 1069 * the actual storage for the result here, instead. 1070 */ 1071 result_src = src_reg(this, ir->type); 1072 /* convenience for the emit functions below. */ 1073 result_dst = dst_reg(result_src); 1074 /* If nothing special happens, this is the result. */ 1075 this->result = result_src; 1076 /* Limit writes to the channels that will be used by result_src later. 1077 * This does limit this temp's use as a temporary for multi-instruction 1078 * sequences. 1079 */ 1080 result_dst.writemask = (1 << ir->type->vector_elements) - 1; 1081 1082 switch (ir->operation) { 1083 case ir_unop_logic_not: 1084 /* Note that BRW_OPCODE_NOT is not appropriate here, since it is 1085 * ones complement of the whole register, not just bit 0. 1086 */ 1087 emit(XOR(result_dst, op[0], src_reg(1))); 1088 break; 1089 case ir_unop_neg: 1090 op[0].negate = !op[0].negate; 1091 this->result = op[0]; 1092 break; 1093 case ir_unop_abs: 1094 op[0].abs = true; 1095 op[0].negate = false; 1096 this->result = op[0]; 1097 break; 1098 1099 case ir_unop_sign: 1100 emit(MOV(result_dst, src_reg(0.0f))); 1101 1102 emit(CMP(dst_null_d(), op[0], src_reg(0.0f), BRW_CONDITIONAL_G)); 1103 inst = emit(MOV(result_dst, src_reg(1.0f))); 1104 inst->predicate = BRW_PREDICATE_NORMAL; 1105 1106 emit(CMP(dst_null_d(), op[0], src_reg(0.0f), BRW_CONDITIONAL_L)); 1107 inst = emit(MOV(result_dst, src_reg(-1.0f))); 1108 inst->predicate = BRW_PREDICATE_NORMAL; 1109 1110 break; 1111 1112 case ir_unop_rcp: 1113 emit_math(SHADER_OPCODE_RCP, result_dst, op[0]); 1114 break; 1115 1116 case ir_unop_exp2: 1117 emit_math(SHADER_OPCODE_EXP2, result_dst, op[0]); 1118 break; 1119 case ir_unop_log2: 1120 emit_math(SHADER_OPCODE_LOG2, result_dst, op[0]); 1121 break; 1122 case ir_unop_exp: 1123 case ir_unop_log: 1124 assert(!"not reached: should be handled by ir_explog_to_explog2"); 1125 break; 1126 case ir_unop_sin: 1127 case ir_unop_sin_reduced: 1128 emit_math(SHADER_OPCODE_SIN, result_dst, op[0]); 1129 break; 1130 case ir_unop_cos: 1131 case ir_unop_cos_reduced: 1132 emit_math(SHADER_OPCODE_COS, result_dst, op[0]); 1133 break; 1134 1135 case ir_unop_dFdx: 1136 case ir_unop_dFdy: 1137 assert(!"derivatives not valid in vertex shader"); 1138 break; 1139 1140 case ir_unop_noise: 1141 assert(!"not reached: should be handled by lower_noise"); 1142 break; 1143 1144 case ir_binop_add: 1145 emit(ADD(result_dst, op[0], op[1])); 1146 break; 1147 case ir_binop_sub: 1148 assert(!"not reached: should be handled by ir_sub_to_add_neg"); 1149 break; 1150 1151 case ir_binop_mul: 1152 if (ir->type->is_integer()) { 1153 /* For integer multiplication, the MUL uses the low 16 bits 1154 * of one of the operands (src0 on gen6, src1 on gen7). The 1155 * MACH accumulates in the contribution of the upper 16 bits 1156 * of that operand. 1157 * 1158 * FINISHME: Emit just the MUL if we know an operand is small 1159 * enough. 1160 */ 1161 struct brw_reg acc = retype(brw_acc_reg(), BRW_REGISTER_TYPE_D); 1162 1163 emit(MUL(acc, op[0], op[1])); 1164 emit(MACH(dst_null_d(), op[0], op[1])); 1165 emit(MOV(result_dst, src_reg(acc))); 1166 } else { 1167 emit(MUL(result_dst, op[0], op[1])); 1168 } 1169 break; 1170 case ir_binop_div: 1171 /* Floating point should be lowered by DIV_TO_MUL_RCP in the compiler. */ 1172 assert(ir->type->is_integer()); 1173 emit_math(SHADER_OPCODE_INT_QUOTIENT, result_dst, op[0], op[1]); 1174 break; 1175 case ir_binop_mod: 1176 /* Floating point should be lowered by MOD_TO_FRACT in the compiler. */ 1177 assert(ir->type->is_integer()); 1178 emit_math(SHADER_OPCODE_INT_REMAINDER, result_dst, op[0], op[1]); 1179 break; 1180 1181 case ir_binop_less: 1182 case ir_binop_greater: 1183 case ir_binop_lequal: 1184 case ir_binop_gequal: 1185 case ir_binop_equal: 1186 case ir_binop_nequal: { 1187 emit(CMP(result_dst, op[0], op[1], 1188 brw_conditional_for_comparison(ir->operation))); 1189 emit(AND(result_dst, result_src, src_reg(0x1))); 1190 break; 1191 } 1192 1193 case ir_binop_all_equal: 1194 /* "==" operator producing a scalar boolean. */ 1195 if (ir->operands[0]->type->is_vector() || 1196 ir->operands[1]->type->is_vector()) { 1197 emit(CMP(dst_null_d(), op[0], op[1], BRW_CONDITIONAL_Z)); 1198 emit(MOV(result_dst, src_reg(0))); 1199 inst = emit(MOV(result_dst, src_reg(1))); 1200 inst->predicate = BRW_PREDICATE_ALIGN16_ALL4H; 1201 } else { 1202 emit(CMP(result_dst, op[0], op[1], BRW_CONDITIONAL_Z)); 1203 emit(AND(result_dst, result_src, src_reg(0x1))); 1204 } 1205 break; 1206 case ir_binop_any_nequal: 1207 /* "!=" operator producing a scalar boolean. */ 1208 if (ir->operands[0]->type->is_vector() || 1209 ir->operands[1]->type->is_vector()) { 1210 emit(CMP(dst_null_d(), op[0], op[1], BRW_CONDITIONAL_NZ)); 1211 1212 emit(MOV(result_dst, src_reg(0))); 1213 inst = emit(MOV(result_dst, src_reg(1))); 1214 inst->predicate = BRW_PREDICATE_ALIGN16_ANY4H; 1215 } else { 1216 emit(CMP(result_dst, op[0], op[1], BRW_CONDITIONAL_NZ)); 1217 emit(AND(result_dst, result_src, src_reg(0x1))); 1218 } 1219 break; 1220 1221 case ir_unop_any: 1222 emit(CMP(dst_null_d(), op[0], src_reg(0), BRW_CONDITIONAL_NZ)); 1223 emit(MOV(result_dst, src_reg(0))); 1224 1225 inst = emit(MOV(result_dst, src_reg(1))); 1226 inst->predicate = BRW_PREDICATE_ALIGN16_ANY4H; 1227 break; 1228 1229 case ir_binop_logic_xor: 1230 emit(XOR(result_dst, op[0], op[1])); 1231 break; 1232 1233 case ir_binop_logic_or: 1234 emit(OR(result_dst, op[0], op[1])); 1235 break; 1236 1237 case ir_binop_logic_and: 1238 emit(AND(result_dst, op[0], op[1])); 1239 break; 1240 1241 case ir_binop_dot: 1242 assert(ir->operands[0]->type->is_vector()); 1243 assert(ir->operands[0]->type == ir->operands[1]->type); 1244 emit_dp(result_dst, op[0], op[1], ir->operands[0]->type->vector_elements); 1245 break; 1246 1247 case ir_unop_sqrt: 1248 emit_math(SHADER_OPCODE_SQRT, result_dst, op[0]); 1249 break; 1250 case ir_unop_rsq: 1251 emit_math(SHADER_OPCODE_RSQ, result_dst, op[0]); 1252 break; 1253 case ir_unop_i2f: 1254 case ir_unop_i2u: 1255 case ir_unop_u2i: 1256 case ir_unop_u2f: 1257 case ir_unop_b2f: 1258 case ir_unop_b2i: 1259 case ir_unop_f2i: 1260 emit(MOV(result_dst, op[0])); 1261 break; 1262 case ir_unop_f2b: 1263 case ir_unop_i2b: { 1264 emit(CMP(result_dst, op[0], src_reg(0.0f), BRW_CONDITIONAL_NZ)); 1265 emit(AND(result_dst, result_src, src_reg(1))); 1266 break; 1267 } 1268 1269 case ir_unop_trunc: 1270 emit(RNDZ(result_dst, op[0])); 1271 break; 1272 case ir_unop_ceil: 1273 op[0].negate = !op[0].negate; 1274 inst = emit(RNDD(result_dst, op[0])); 1275 this->result.negate = true; 1276 break; 1277 case ir_unop_floor: 1278 inst = emit(RNDD(result_dst, op[0])); 1279 break; 1280 case ir_unop_fract: 1281 inst = emit(FRC(result_dst, op[0])); 1282 break; 1283 case ir_unop_round_even: 1284 emit(RNDE(result_dst, op[0])); 1285 break; 1286 1287 case ir_binop_min: 1288 emit(CMP(result_dst, op[0], op[1], BRW_CONDITIONAL_L)); 1289 1290 inst = emit(BRW_OPCODE_SEL, result_dst, op[0], op[1]); 1291 inst->predicate = BRW_PREDICATE_NORMAL; 1292 break; 1293 case ir_binop_max: 1294 emit(CMP(result_dst, op[0], op[1], BRW_CONDITIONAL_G)); 1295 1296 inst = emit(BRW_OPCODE_SEL, result_dst, op[0], op[1]); 1297 inst->predicate = BRW_PREDICATE_NORMAL; 1298 break; 1299 1300 case ir_binop_pow: 1301 emit_math(SHADER_OPCODE_POW, result_dst, op[0], op[1]); 1302 break; 1303 1304 case ir_unop_bit_not: 1305 inst = emit(NOT(result_dst, op[0])); 1306 break; 1307 case ir_binop_bit_and: 1308 inst = emit(AND(result_dst, op[0], op[1])); 1309 break; 1310 case ir_binop_bit_xor: 1311 inst = emit(XOR(result_dst, op[0], op[1])); 1312 break; 1313 case ir_binop_bit_or: 1314 inst = emit(OR(result_dst, op[0], op[1])); 1315 break; 1316 1317 case ir_binop_lshift: 1318 inst = emit(BRW_OPCODE_SHL, result_dst, op[0], op[1]); 1319 break; 1320 1321 case ir_binop_rshift: 1322 if (ir->type->base_type == GLSL_TYPE_INT) 1323 inst = emit(BRW_OPCODE_ASR, result_dst, op[0], op[1]); 1324 else 1325 inst = emit(BRW_OPCODE_SHR, result_dst, op[0], op[1]); 1326 break; 1327 1328 case ir_quadop_vector: 1329 assert(!"not reached: should be handled by lower_quadop_vector"); 1330 break; 1331 } 1332} 1333 1334 1335void 1336vec4_visitor::visit(ir_swizzle *ir) 1337{ 1338 src_reg src; 1339 int i = 0; 1340 int swizzle[4]; 1341 1342 /* Note that this is only swizzles in expressions, not those on the left 1343 * hand side of an assignment, which do write masking. See ir_assignment 1344 * for that. 1345 */ 1346 1347 ir->val->accept(this); 1348 src = this->result; 1349 assert(src.file != BAD_FILE); 1350 1351 for (i = 0; i < ir->type->vector_elements; i++) { 1352 switch (i) { 1353 case 0: 1354 swizzle[i] = BRW_GET_SWZ(src.swizzle, ir->mask.x); 1355 break; 1356 case 1: 1357 swizzle[i] = BRW_GET_SWZ(src.swizzle, ir->mask.y); 1358 break; 1359 case 2: 1360 swizzle[i] = BRW_GET_SWZ(src.swizzle, ir->mask.z); 1361 break; 1362 case 3: 1363 swizzle[i] = BRW_GET_SWZ(src.swizzle, ir->mask.w); 1364 break; 1365 } 1366 } 1367 for (; i < 4; i++) { 1368 /* Replicate the last channel out. */ 1369 swizzle[i] = swizzle[ir->type->vector_elements - 1]; 1370 } 1371 1372 src.swizzle = BRW_SWIZZLE4(swizzle[0], swizzle[1], swizzle[2], swizzle[3]); 1373 1374 this->result = src; 1375} 1376 1377void 1378vec4_visitor::visit(ir_dereference_variable *ir) 1379{ 1380 const struct glsl_type *type = ir->type; 1381 dst_reg *reg = variable_storage(ir->var); 1382 1383 if (!reg) { 1384 fail("Failed to find variable storage for %s\n", ir->var->name); 1385 this->result = src_reg(brw_null_reg()); 1386 return; 1387 } 1388 1389 this->result = src_reg(*reg); 1390 1391 if (type->is_scalar() || type->is_vector() || type->is_matrix()) 1392 this->result.swizzle = swizzle_for_size(type->vector_elements); 1393} 1394 1395void 1396vec4_visitor::visit(ir_dereference_array *ir) 1397{ 1398 ir_constant *constant_index; 1399 src_reg src; 1400 int element_size = type_size(ir->type); 1401 1402 constant_index = ir->array_index->constant_expression_value(); 1403 1404 ir->array->accept(this); 1405 src = this->result; 1406 1407 if (constant_index) { 1408 src.reg_offset += constant_index->value.i[0] * element_size; 1409 } else { 1410 /* Variable index array dereference. It eats the "vec4" of the 1411 * base of the array and an index that offsets the Mesa register 1412 * index. 1413 */ 1414 ir->array_index->accept(this); 1415 1416 src_reg index_reg; 1417 1418 if (element_size == 1) { 1419 index_reg = this->result; 1420 } else { 1421 index_reg = src_reg(this, glsl_type::int_type); 1422 1423 emit(MUL(dst_reg(index_reg), this->result, src_reg(element_size))); 1424 } 1425 1426 if (src.reladdr) { 1427 src_reg temp = src_reg(this, glsl_type::int_type); 1428 1429 emit(ADD(dst_reg(temp), *src.reladdr, index_reg)); 1430 1431 index_reg = temp; 1432 } 1433 1434 src.reladdr = ralloc(mem_ctx, src_reg); 1435 memcpy(src.reladdr, &index_reg, sizeof(index_reg)); 1436 } 1437 1438 /* If the type is smaller than a vec4, replicate the last channel out. */ 1439 if (ir->type->is_scalar() || ir->type->is_vector()) 1440 src.swizzle = swizzle_for_size(ir->type->vector_elements); 1441 else 1442 src.swizzle = BRW_SWIZZLE_NOOP; 1443 src.type = brw_type_for_base_type(ir->type); 1444 1445 this->result = src; 1446} 1447 1448void 1449vec4_visitor::visit(ir_dereference_record *ir) 1450{ 1451 unsigned int i; 1452 const glsl_type *struct_type = ir->record->type; 1453 int offset = 0; 1454 1455 ir->record->accept(this); 1456 1457 for (i = 0; i < struct_type->length; i++) { 1458 if (strcmp(struct_type->fields.structure[i].name, ir->field) == 0) 1459 break; 1460 offset += type_size(struct_type->fields.structure[i].type); 1461 } 1462 1463 /* If the type is smaller than a vec4, replicate the last channel out. */ 1464 if (ir->type->is_scalar() || ir->type->is_vector()) 1465 this->result.swizzle = swizzle_for_size(ir->type->vector_elements); 1466 else 1467 this->result.swizzle = BRW_SWIZZLE_NOOP; 1468 this->result.type = brw_type_for_base_type(ir->type); 1469 1470 this->result.reg_offset += offset; 1471} 1472 1473/** 1474 * We want to be careful in assignment setup to hit the actual storage 1475 * instead of potentially using a temporary like we might with the 1476 * ir_dereference handler. 1477 */ 1478static dst_reg 1479get_assignment_lhs(ir_dereference *ir, vec4_visitor *v) 1480{ 1481 /* The LHS must be a dereference. If the LHS is a variable indexed array 1482 * access of a vector, it must be separated into a series conditional moves 1483 * before reaching this point (see ir_vec_index_to_cond_assign). 1484 */ 1485 assert(ir->as_dereference()); 1486 ir_dereference_array *deref_array = ir->as_dereference_array(); 1487 if (deref_array) { 1488 assert(!deref_array->array->type->is_vector()); 1489 } 1490 1491 /* Use the rvalue deref handler for the most part. We'll ignore 1492 * swizzles in it and write swizzles using writemask, though. 1493 */ 1494 ir->accept(v); 1495 return dst_reg(v->result); 1496} 1497 1498void 1499vec4_visitor::emit_block_move(dst_reg *dst, src_reg *src, 1500 const struct glsl_type *type, uint32_t predicate) 1501{ 1502 if (type->base_type == GLSL_TYPE_STRUCT) { 1503 for (unsigned int i = 0; i < type->length; i++) { 1504 emit_block_move(dst, src, type->fields.structure[i].type, predicate); 1505 } 1506 return; 1507 } 1508 1509 if (type->is_array()) { 1510 for (unsigned int i = 0; i < type->length; i++) { 1511 emit_block_move(dst, src, type->fields.array, predicate); 1512 } 1513 return; 1514 } 1515 1516 if (type->is_matrix()) { 1517 const struct glsl_type *vec_type; 1518 1519 vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT, 1520 type->vector_elements, 1); 1521 1522 for (int i = 0; i < type->matrix_columns; i++) { 1523 emit_block_move(dst, src, vec_type, predicate); 1524 } 1525 return; 1526 } 1527 1528 assert(type->is_scalar() || type->is_vector()); 1529 1530 dst->type = brw_type_for_base_type(type); 1531 src->type = dst->type; 1532 1533 dst->writemask = (1 << type->vector_elements) - 1; 1534 1535 /* Do we need to worry about swizzling a swizzle? */ 1536 assert(src->swizzle == BRW_SWIZZLE_NOOP 1537 || src->swizzle == swizzle_for_size(type->vector_elements)); 1538 src->swizzle = swizzle_for_size(type->vector_elements); 1539 1540 vec4_instruction *inst = emit(MOV(*dst, *src)); 1541 inst->predicate = predicate; 1542 1543 dst->reg_offset++; 1544 src->reg_offset++; 1545} 1546 1547 1548/* If the RHS processing resulted in an instruction generating a 1549 * temporary value, and it would be easy to rewrite the instruction to 1550 * generate its result right into the LHS instead, do so. This ends 1551 * up reliably removing instructions where it can be tricky to do so 1552 * later without real UD chain information. 1553 */ 1554bool 1555vec4_visitor::try_rewrite_rhs_to_dst(ir_assignment *ir, 1556 dst_reg dst, 1557 src_reg src, 1558 vec4_instruction *pre_rhs_inst, 1559 vec4_instruction *last_rhs_inst) 1560{ 1561 /* This could be supported, but it would take more smarts. */ 1562 if (ir->condition) 1563 return false; 1564 1565 if (pre_rhs_inst == last_rhs_inst) 1566 return false; /* No instructions generated to work with. */ 1567 1568 /* Make sure the last instruction generated our source reg. */ 1569 if (src.file != GRF || 1570 src.file != last_rhs_inst->dst.file || 1571 src.reg != last_rhs_inst->dst.reg || 1572 src.reg_offset != last_rhs_inst->dst.reg_offset || 1573 src.reladdr || 1574 src.abs || 1575 src.negate || 1576 last_rhs_inst->predicate != BRW_PREDICATE_NONE) 1577 return false; 1578 1579 /* Check that that last instruction fully initialized the channels 1580 * we want to use, in the order we want to use them. We could 1581 * potentially reswizzle the operands of many instructions so that 1582 * we could handle out of order channels, but don't yet. 1583 */ 1584 1585 for (unsigned i = 0; i < 4; i++) { 1586 if (dst.writemask & (1 << i)) { 1587 if (!(last_rhs_inst->dst.writemask & (1 << i))) 1588 return false; 1589 1590 if (BRW_GET_SWZ(src.swizzle, i) != i) 1591 return false; 1592 } 1593 } 1594 1595 /* Success! Rewrite the instruction. */ 1596 last_rhs_inst->dst.file = dst.file; 1597 last_rhs_inst->dst.reg = dst.reg; 1598 last_rhs_inst->dst.reg_offset = dst.reg_offset; 1599 last_rhs_inst->dst.reladdr = dst.reladdr; 1600 last_rhs_inst->dst.writemask &= dst.writemask; 1601 1602 return true; 1603} 1604 1605void 1606vec4_visitor::visit(ir_assignment *ir) 1607{ 1608 dst_reg dst = get_assignment_lhs(ir->lhs, this); 1609 uint32_t predicate = BRW_PREDICATE_NONE; 1610 1611 if (!ir->lhs->type->is_scalar() && 1612 !ir->lhs->type->is_vector()) { 1613 ir->rhs->accept(this); 1614 src_reg src = this->result; 1615 1616 if (ir->condition) { 1617 emit_bool_to_cond_code(ir->condition, &predicate); 1618 } 1619 1620 emit_block_move(&dst, &src, ir->rhs->type, predicate); 1621 return; 1622 } 1623 1624 /* Now we're down to just a scalar/vector with writemasks. */ 1625 int i; 1626 1627 vec4_instruction *pre_rhs_inst, *last_rhs_inst; 1628 pre_rhs_inst = (vec4_instruction *)this->instructions.get_tail(); 1629 1630 ir->rhs->accept(this); 1631 1632 last_rhs_inst = (vec4_instruction *)this->instructions.get_tail(); 1633 1634 src_reg src = this->result; 1635 1636 int swizzles[4]; 1637 int first_enabled_chan = 0; 1638 int src_chan = 0; 1639 1640 assert(ir->lhs->type->is_vector() || 1641 ir->lhs->type->is_scalar()); 1642 dst.writemask = ir->write_mask; 1643 1644 for (int i = 0; i < 4; i++) { 1645 if (dst.writemask & (1 << i)) { 1646 first_enabled_chan = BRW_GET_SWZ(src.swizzle, i); 1647 break; 1648 } 1649 } 1650 1651 /* Swizzle a small RHS vector into the channels being written. 1652 * 1653 * glsl ir treats write_mask as dictating how many channels are 1654 * present on the RHS while in our instructions we need to make 1655 * those channels appear in the slots of the vec4 they're written to. 1656 */ 1657 for (int i = 0; i < 4; i++) { 1658 if (dst.writemask & (1 << i)) 1659 swizzles[i] = BRW_GET_SWZ(src.swizzle, src_chan++); 1660 else 1661 swizzles[i] = first_enabled_chan; 1662 } 1663 src.swizzle = BRW_SWIZZLE4(swizzles[0], swizzles[1], 1664 swizzles[2], swizzles[3]); 1665 1666 if (try_rewrite_rhs_to_dst(ir, dst, src, pre_rhs_inst, last_rhs_inst)) { 1667 return; 1668 } 1669 1670 if (ir->condition) { 1671 emit_bool_to_cond_code(ir->condition, &predicate); 1672 } 1673 1674 for (i = 0; i < type_size(ir->lhs->type); i++) { 1675 vec4_instruction *inst = emit(MOV(dst, src)); 1676 inst->predicate = predicate; 1677 1678 dst.reg_offset++; 1679 src.reg_offset++; 1680 } 1681} 1682 1683void 1684vec4_visitor::emit_constant_values(dst_reg *dst, ir_constant *ir) 1685{ 1686 if (ir->type->base_type == GLSL_TYPE_STRUCT) { 1687 foreach_list(node, &ir->components) { 1688 ir_constant *field_value = (ir_constant *)node; 1689 1690 emit_constant_values(dst, field_value); 1691 } 1692 return; 1693 } 1694 1695 if (ir->type->is_array()) { 1696 for (unsigned int i = 0; i < ir->type->length; i++) { 1697 emit_constant_values(dst, ir->array_elements[i]); 1698 } 1699 return; 1700 } 1701 1702 if (ir->type->is_matrix()) { 1703 for (int i = 0; i < ir->type->matrix_columns; i++) { 1704 float *vec = &ir->value.f[i * ir->type->vector_elements]; 1705 1706 for (int j = 0; j < ir->type->vector_elements; j++) { 1707 dst->writemask = 1 << j; 1708 dst->type = BRW_REGISTER_TYPE_F; 1709 1710 emit(MOV(*dst, src_reg(vec[j]))); 1711 } 1712 dst->reg_offset++; 1713 } 1714 return; 1715 } 1716 1717 int remaining_writemask = (1 << ir->type->vector_elements) - 1; 1718 1719 for (int i = 0; i < ir->type->vector_elements; i++) { 1720 if (!(remaining_writemask & (1 << i))) 1721 continue; 1722 1723 dst->writemask = 1 << i; 1724 dst->type = brw_type_for_base_type(ir->type); 1725 1726 /* Find other components that match the one we're about to 1727 * write. Emits fewer instructions for things like vec4(0.5, 1728 * 1.5, 1.5, 1.5). 1729 */ 1730 for (int j = i + 1; j < ir->type->vector_elements; j++) { 1731 if (ir->type->base_type == GLSL_TYPE_BOOL) { 1732 if (ir->value.b[i] == ir->value.b[j]) 1733 dst->writemask |= (1 << j); 1734 } else { 1735 /* u, i, and f storage all line up, so no need for a 1736 * switch case for comparing each type. 1737 */ 1738 if (ir->value.u[i] == ir->value.u[j]) 1739 dst->writemask |= (1 << j); 1740 } 1741 } 1742 1743 switch (ir->type->base_type) { 1744 case GLSL_TYPE_FLOAT: 1745 emit(MOV(*dst, src_reg(ir->value.f[i]))); 1746 break; 1747 case GLSL_TYPE_INT: 1748 emit(MOV(*dst, src_reg(ir->value.i[i]))); 1749 break; 1750 case GLSL_TYPE_UINT: 1751 emit(MOV(*dst, src_reg(ir->value.u[i]))); 1752 break; 1753 case GLSL_TYPE_BOOL: 1754 emit(MOV(*dst, src_reg(ir->value.b[i]))); 1755 break; 1756 default: 1757 assert(!"Non-float/uint/int/bool constant"); 1758 break; 1759 } 1760 1761 remaining_writemask &= ~dst->writemask; 1762 } 1763 dst->reg_offset++; 1764} 1765 1766void 1767vec4_visitor::visit(ir_constant *ir) 1768{ 1769 dst_reg dst = dst_reg(this, ir->type); 1770 this->result = src_reg(dst); 1771 1772 emit_constant_values(&dst, ir); 1773} 1774 1775void 1776vec4_visitor::visit(ir_call *ir) 1777{ 1778 assert(!"not reached"); 1779} 1780 1781void 1782vec4_visitor::visit(ir_texture *ir) 1783{ 1784 int sampler = _mesa_get_sampler_uniform_value(ir->sampler, prog, &vp->Base); 1785 sampler = vp->Base.SamplerUnits[sampler]; 1786 1787 /* Should be lowered by do_lower_texture_projection */ 1788 assert(!ir->projector); 1789 1790 vec4_instruction *inst = NULL; 1791 switch (ir->op) { 1792 case ir_tex: 1793 case ir_txl: 1794 inst = new(mem_ctx) vec4_instruction(this, SHADER_OPCODE_TXL); 1795 break; 1796 case ir_txd: 1797 inst = new(mem_ctx) vec4_instruction(this, SHADER_OPCODE_TXD); 1798 break; 1799 case ir_txf: 1800 inst = new(mem_ctx) vec4_instruction(this, SHADER_OPCODE_TXF); 1801 break; 1802 case ir_txs: 1803 inst = new(mem_ctx) vec4_instruction(this, SHADER_OPCODE_TXS); 1804 break; 1805 case ir_txb: 1806 assert(!"TXB is not valid for vertex shaders."); 1807 } 1808 1809 /* Texel offsets go in the message header; Gen4 also requires headers. */ 1810 inst->header_present = ir->offset || intel->gen < 5; 1811 inst->base_mrf = 2; 1812 inst->mlen = inst->header_present + 1; /* always at least one */ 1813 inst->sampler = sampler; 1814 inst->dst = dst_reg(this, ir->type); 1815 inst->shadow_compare = ir->shadow_comparitor != NULL; 1816 1817 if (ir->offset != NULL) 1818 inst->texture_offset = brw_texture_offset(ir->offset->as_constant()); 1819 1820 /* MRF for the first parameter */ 1821 int param_base = inst->base_mrf + inst->header_present; 1822 1823 if (ir->op == ir_txs) { 1824 ir->lod_info.lod->accept(this); 1825 int writemask = intel->gen == 4 ? WRITEMASK_W : WRITEMASK_X; 1826 emit(MOV(dst_reg(MRF, param_base, ir->lod_info.lod->type, writemask), 1827 this->result)); 1828 } else { 1829 int i, coord_mask = 0, zero_mask = 0; 1830 /* Load the coordinate */ 1831 /* FINISHME: gl_clamp_mask and saturate */ 1832 for (i = 0; i < ir->coordinate->type->vector_elements; i++) 1833 coord_mask |= (1 << i); 1834 for (; i < 4; i++) 1835 zero_mask |= (1 << i); 1836 1837 ir->coordinate->accept(this); 1838 emit(MOV(dst_reg(MRF, param_base, ir->coordinate->type, coord_mask), 1839 this->result)); 1840 emit(MOV(dst_reg(MRF, param_base, ir->coordinate->type, zero_mask), 1841 src_reg(0))); 1842 /* Load the shadow comparitor */ 1843 if (ir->shadow_comparitor) { 1844 ir->shadow_comparitor->accept(this); 1845 emit(MOV(dst_reg(MRF, param_base + 1, ir->shadow_comparitor->type, 1846 WRITEMASK_X), 1847 this->result)); 1848 inst->mlen++; 1849 } 1850 1851 /* Load the LOD info */ 1852 if (ir->op == ir_txl) { 1853 int mrf, writemask; 1854 if (intel->gen >= 5) { 1855 mrf = param_base + 1; 1856 if (ir->shadow_comparitor) { 1857 writemask = WRITEMASK_Y; 1858 /* mlen already incremented */ 1859 } else { 1860 writemask = WRITEMASK_X; 1861 inst->mlen++; 1862 } 1863 } else /* intel->gen == 4 */ { 1864 mrf = param_base; 1865 writemask = WRITEMASK_Z; 1866 } 1867 ir->lod_info.lod->accept(this); 1868 emit(MOV(dst_reg(MRF, mrf, ir->lod_info.lod->type, writemask), 1869 this->result)); 1870 } else if (ir->op == ir_txf) { 1871 ir->lod_info.lod->accept(this); 1872 emit(MOV(dst_reg(MRF, param_base, ir->lod_info.lod->type, WRITEMASK_W), 1873 this->result)); 1874 } else if (ir->op == ir_txd) { 1875 const glsl_type *type = ir->lod_info.grad.dPdx->type; 1876 1877 ir->lod_info.grad.dPdx->accept(this); 1878 src_reg dPdx = this->result; 1879 ir->lod_info.grad.dPdy->accept(this); 1880 src_reg dPdy = this->result; 1881 1882 if (intel->gen >= 5) { 1883 dPdx.swizzle = BRW_SWIZZLE4(SWIZZLE_X,SWIZZLE_X,SWIZZLE_Y,SWIZZLE_Y); 1884 dPdy.swizzle = BRW_SWIZZLE4(SWIZZLE_X,SWIZZLE_X,SWIZZLE_Y,SWIZZLE_Y); 1885 emit(MOV(dst_reg(MRF, param_base + 1, type, WRITEMASK_XZ), dPdx)); 1886 emit(MOV(dst_reg(MRF, param_base + 1, type, WRITEMASK_YW), dPdy)); 1887 inst->mlen++; 1888 1889 if (ir->type->vector_elements == 3) { 1890 dPdx.swizzle = BRW_SWIZZLE_ZZZZ; 1891 dPdy.swizzle = BRW_SWIZZLE_ZZZZ; 1892 emit(MOV(dst_reg(MRF, param_base + 2, type, WRITEMASK_X), dPdx)); 1893 emit(MOV(dst_reg(MRF, param_base + 2, type, WRITEMASK_Y), dPdy)); 1894 inst->mlen++; 1895 } 1896 } else /* intel->gen == 4 */ { 1897 emit(MOV(dst_reg(MRF, param_base + 1, type, WRITEMASK_XYZ), dPdx)); 1898 emit(MOV(dst_reg(MRF, param_base + 2, type, WRITEMASK_XYZ), dPdy)); 1899 inst->mlen += 2; 1900 } 1901 } 1902 } 1903 1904 emit(inst); 1905 1906 swizzle_result(ir, src_reg(inst->dst), sampler); 1907} 1908 1909void 1910vec4_visitor::swizzle_result(ir_texture *ir, src_reg orig_val, int sampler) 1911{ 1912 this->result = orig_val; 1913 1914 int s = c->key.tex.swizzles[sampler]; 1915 1916 if (ir->op == ir_txs || ir->type == glsl_type::float_type 1917 || s == SWIZZLE_NOOP) 1918 return; 1919 1920 int zero_mask = 0, one_mask = 0, copy_mask = 0; 1921 int swizzle[4]; 1922 1923 for (int i = 0; i < 4; i++) { 1924 switch (GET_SWZ(s, i)) { 1925 case SWIZZLE_ZERO: 1926 zero_mask |= (1 << i); 1927 break; 1928 case SWIZZLE_ONE: 1929 one_mask |= (1 << i); 1930 break; 1931 default: 1932 copy_mask |= (1 << i); 1933 swizzle[i] = GET_SWZ(s, i); 1934 break; 1935 } 1936 } 1937 1938 this->result = src_reg(this, ir->type); 1939 dst_reg swizzled_result(this->result); 1940 1941 if (copy_mask) { 1942 orig_val.swizzle = BRW_SWIZZLE4(swizzle[0], swizzle[1], swizzle[2], swizzle[3]); 1943 swizzled_result.writemask = copy_mask; 1944 emit(MOV(swizzled_result, orig_val)); 1945 } 1946 1947 if (zero_mask) { 1948 swizzled_result.writemask = zero_mask; 1949 emit(MOV(swizzled_result, src_reg(0.0f))); 1950 } 1951 1952 if (one_mask) { 1953 swizzled_result.writemask = one_mask; 1954 emit(MOV(swizzled_result, src_reg(1.0f))); 1955 } 1956} 1957 1958void 1959vec4_visitor::visit(ir_return *ir) 1960{ 1961 assert(!"not reached"); 1962} 1963 1964void 1965vec4_visitor::visit(ir_discard *ir) 1966{ 1967 assert(!"not reached"); 1968} 1969 1970void 1971vec4_visitor::visit(ir_if *ir) 1972{ 1973 /* Don't point the annotation at the if statement, because then it plus 1974 * the then and else blocks get printed. 1975 */ 1976 this->base_ir = ir->condition; 1977 1978 if (intel->gen == 6) { 1979 emit_if_gen6(ir); 1980 } else { 1981 uint32_t predicate; 1982 emit_bool_to_cond_code(ir->condition, &predicate); 1983 emit(IF(predicate)); 1984 } 1985 1986 visit_instructions(&ir->then_instructions); 1987 1988 if (!ir->else_instructions.is_empty()) { 1989 this->base_ir = ir->condition; 1990 emit(BRW_OPCODE_ELSE); 1991 1992 visit_instructions(&ir->else_instructions); 1993 } 1994 1995 this->base_ir = ir->condition; 1996 emit(BRW_OPCODE_ENDIF); 1997} 1998 1999void 2000vec4_visitor::emit_ndc_computation() 2001{ 2002 /* Get the position */ 2003 src_reg pos = src_reg(output_reg[VERT_RESULT_HPOS]); 2004 2005 /* Build ndc coords, which are (x/w, y/w, z/w, 1/w) */ 2006 dst_reg ndc = dst_reg(this, glsl_type::vec4_type); 2007 output_reg[BRW_VERT_RESULT_NDC] = ndc; 2008 2009 current_annotation = "NDC"; 2010 dst_reg ndc_w = ndc; 2011 ndc_w.writemask = WRITEMASK_W; 2012 src_reg pos_w = pos; 2013 pos_w.swizzle = BRW_SWIZZLE4(SWIZZLE_W, SWIZZLE_W, SWIZZLE_W, SWIZZLE_W); 2014 emit_math(SHADER_OPCODE_RCP, ndc_w, pos_w); 2015 2016 dst_reg ndc_xyz = ndc; 2017 ndc_xyz.writemask = WRITEMASK_XYZ; 2018 2019 emit(MUL(ndc_xyz, pos, src_reg(ndc_w))); 2020} 2021 2022void 2023vec4_visitor::emit_psiz_and_flags(struct brw_reg reg) 2024{ 2025 if (intel->gen < 6 && 2026 ((c->prog_data.outputs_written & BITFIELD64_BIT(VERT_RESULT_PSIZ)) || 2027 c->key.userclip_active || brw->has_negative_rhw_bug)) { 2028 dst_reg header1 = dst_reg(this, glsl_type::uvec4_type); 2029 dst_reg header1_w = header1; 2030 header1_w.writemask = WRITEMASK_W; 2031 GLuint i; 2032 2033 emit(MOV(header1, 0u)); 2034 2035 if (c->prog_data.outputs_written & BITFIELD64_BIT(VERT_RESULT_PSIZ)) { 2036 src_reg psiz = src_reg(output_reg[VERT_RESULT_PSIZ]); 2037 2038 current_annotation = "Point size"; 2039 emit(MUL(header1_w, psiz, src_reg((float)(1 << 11)))); 2040 emit(AND(header1_w, src_reg(header1_w), 0x7ff << 8)); 2041 } 2042 2043 current_annotation = "Clipping flags"; 2044 for (i = 0; i < c->key.nr_userclip_plane_consts; i++) { 2045 vec4_instruction *inst; 2046 2047 inst = emit(DP4(dst_null_f(), src_reg(output_reg[VERT_RESULT_HPOS]), 2048 src_reg(this->userplane[i]))); 2049 inst->conditional_mod = BRW_CONDITIONAL_L; 2050 2051 inst = emit(OR(header1_w, src_reg(header1_w), 1u << i)); 2052 inst->predicate = BRW_PREDICATE_NORMAL; 2053 } 2054 2055 /* i965 clipping workaround: 2056 * 1) Test for -ve rhw 2057 * 2) If set, 2058 * set ndc = (0,0,0,0) 2059 * set ucp[6] = 1 2060 * 2061 * Later, clipping will detect ucp[6] and ensure the primitive is 2062 * clipped against all fixed planes. 2063 */ 2064 if (brw->has_negative_rhw_bug) { 2065#if 0 2066 /* FINISHME */ 2067 brw_CMP(p, 2068 vec8(brw_null_reg()), 2069 BRW_CONDITIONAL_L, 2070 brw_swizzle1(output_reg[BRW_VERT_RESULT_NDC], 3), 2071 brw_imm_f(0)); 2072 2073 brw_OR(p, brw_writemask(header1, WRITEMASK_W), header1, brw_imm_ud(1<<6)); 2074 brw_MOV(p, output_reg[BRW_VERT_RESULT_NDC], brw_imm_f(0)); 2075 brw_set_predicate_control(p, BRW_PREDICATE_NONE); 2076#endif 2077 } 2078 2079 emit(MOV(retype(reg, BRW_REGISTER_TYPE_UD), src_reg(header1))); 2080 } else if (intel->gen < 6) { 2081 emit(MOV(retype(reg, BRW_REGISTER_TYPE_UD), 0u)); 2082 } else { 2083 emit(MOV(retype(reg, BRW_REGISTER_TYPE_D), src_reg(0))); 2084 if (c->prog_data.outputs_written & BITFIELD64_BIT(VERT_RESULT_PSIZ)) { 2085 emit(MOV(brw_writemask(reg, WRITEMASK_W), 2086 src_reg(output_reg[VERT_RESULT_PSIZ]))); 2087 } 2088 } 2089} 2090 2091void 2092vec4_visitor::emit_clip_distances(struct brw_reg reg, int offset) 2093{ 2094 if (intel->gen < 6) { 2095 /* Clip distance slots are set aside in gen5, but they are not used. It 2096 * is not clear whether we actually need to set aside space for them, 2097 * but the performance cost is negligible. 2098 */ 2099 return; 2100 } 2101 2102 /* From the GLSL 1.30 spec, section 7.1 (Vertex Shader Special Variables): 2103 * 2104 * "If a linked set of shaders forming the vertex stage contains no 2105 * static write to gl_ClipVertex or gl_ClipDistance, but the 2106 * application has requested clipping against user clip planes through 2107 * the API, then the coordinate written to gl_Position is used for 2108 * comparison against the user clip planes." 2109 * 2110 * This function is only called if the shader didn't write to 2111 * gl_ClipDistance. Accordingly, we use gl_ClipVertex to perform clipping 2112 * if the user wrote to it; otherwise we use gl_Position. 2113 */ 2114 gl_vert_result clip_vertex = VERT_RESULT_CLIP_VERTEX; 2115 if (!(c->prog_data.outputs_written 2116 & BITFIELD64_BIT(VERT_RESULT_CLIP_VERTEX))) { 2117 clip_vertex = VERT_RESULT_HPOS; 2118 } 2119 2120 for (int i = 0; i + offset < c->key.nr_userclip_plane_consts && i < 4; 2121 ++i) { 2122 emit(DP4(dst_reg(brw_writemask(reg, 1 << i)), 2123 src_reg(output_reg[clip_vertex]), 2124 src_reg(this->userplane[i + offset]))); 2125 } 2126} 2127 2128void 2129vec4_visitor::emit_generic_urb_slot(dst_reg reg, int vert_result) 2130{ 2131 assert (vert_result < VERT_RESULT_MAX); 2132 reg.type = output_reg[vert_result].type; 2133 current_annotation = output_reg_annotation[vert_result]; 2134 /* Copy the register, saturating if necessary */ 2135 vec4_instruction *inst = emit(MOV(reg, 2136 src_reg(output_reg[vert_result]))); 2137 if ((vert_result == VERT_RESULT_COL0 || 2138 vert_result == VERT_RESULT_COL1 || 2139 vert_result == VERT_RESULT_BFC0 || 2140 vert_result == VERT_RESULT_BFC1) && 2141 c->key.clamp_vertex_color) { 2142 inst->saturate = true; 2143 } 2144} 2145 2146void 2147vec4_visitor::emit_urb_slot(int mrf, int vert_result) 2148{ 2149 struct brw_reg hw_reg = brw_message_reg(mrf); 2150 dst_reg reg = dst_reg(MRF, mrf); 2151 reg.type = BRW_REGISTER_TYPE_F; 2152 2153 switch (vert_result) { 2154 case VERT_RESULT_PSIZ: 2155 /* PSIZ is always in slot 0, and is coupled with other flags. */ 2156 current_annotation = "indices, point width, clip flags"; 2157 emit_psiz_and_flags(hw_reg); 2158 break; 2159 case BRW_VERT_RESULT_NDC: 2160 current_annotation = "NDC"; 2161 emit(MOV(reg, src_reg(output_reg[BRW_VERT_RESULT_NDC]))); 2162 break; 2163 case BRW_VERT_RESULT_HPOS_DUPLICATE: 2164 case VERT_RESULT_HPOS: 2165 current_annotation = "gl_Position"; 2166 emit(MOV(reg, src_reg(output_reg[VERT_RESULT_HPOS]))); 2167 break; 2168 case VERT_RESULT_CLIP_DIST0: 2169 case VERT_RESULT_CLIP_DIST1: 2170 if (this->c->key.uses_clip_distance) { 2171 emit_generic_urb_slot(reg, vert_result); 2172 } else { 2173 current_annotation = "user clip distances"; 2174 emit_clip_distances(hw_reg, (vert_result - VERT_RESULT_CLIP_DIST0) * 4); 2175 } 2176 break; 2177 case BRW_VERT_RESULT_PAD: 2178 /* No need to write to this slot */ 2179 break; 2180 default: 2181 emit_generic_urb_slot(reg, vert_result); 2182 break; 2183 } 2184} 2185 2186static int 2187align_interleaved_urb_mlen(struct brw_context *brw, int mlen) 2188{ 2189 struct intel_context *intel = &brw->intel; 2190 2191 if (intel->gen >= 6) { 2192 /* URB data written (does not include the message header reg) must 2193 * be a multiple of 256 bits, or 2 VS registers. See vol5c.5, 2194 * section 5.4.3.2.2: URB_INTERLEAVED. 2195 * 2196 * URB entries are allocated on a multiple of 1024 bits, so an 2197 * extra 128 bits written here to make the end align to 256 is 2198 * no problem. 2199 */ 2200 if ((mlen % 2) != 1) 2201 mlen++; 2202 } 2203 2204 return mlen; 2205} 2206 2207/** 2208 * Generates the VUE payload plus the 1 or 2 URB write instructions to 2209 * complete the VS thread. 2210 * 2211 * The VUE layout is documented in Volume 2a. 2212 */ 2213void 2214vec4_visitor::emit_urb_writes() 2215{ 2216 /* MRF 0 is reserved for the debugger, so start with message header 2217 * in MRF 1. 2218 */ 2219 int base_mrf = 1; 2220 int mrf = base_mrf; 2221 /* In the process of generating our URB write message contents, we 2222 * may need to unspill a register or load from an array. Those 2223 * reads would use MRFs 14-15. 2224 */ 2225 int max_usable_mrf = 13; 2226 2227 /* The following assertion verifies that max_usable_mrf causes an 2228 * even-numbered amount of URB write data, which will meet gen6's 2229 * requirements for length alignment. 2230 */ 2231 assert ((max_usable_mrf - base_mrf) % 2 == 0); 2232 2233 /* FINISHME: edgeflag */ 2234 2235 brw_compute_vue_map(&c->vue_map, intel, c->key.userclip_active, 2236 c->prog_data.outputs_written); 2237 2238 /* First mrf is the g0-based message header containing URB handles and such, 2239 * which is implied in VS_OPCODE_URB_WRITE. 2240 */ 2241 mrf++; 2242 2243 if (intel->gen < 6) { 2244 emit_ndc_computation(); 2245 } 2246 2247 /* Set up the VUE data for the first URB write */ 2248 int slot; 2249 for (slot = 0; slot < c->vue_map.num_slots; ++slot) { 2250 emit_urb_slot(mrf++, c->vue_map.slot_to_vert_result[slot]); 2251 2252 /* If this was max_usable_mrf, we can't fit anything more into this URB 2253 * WRITE. 2254 */ 2255 if (mrf > max_usable_mrf) { 2256 slot++; 2257 break; 2258 } 2259 } 2260 2261 current_annotation = "URB write"; 2262 vec4_instruction *inst = emit(VS_OPCODE_URB_WRITE); 2263 inst->base_mrf = base_mrf; 2264 inst->mlen = align_interleaved_urb_mlen(brw, mrf - base_mrf); 2265 inst->eot = (slot >= c->vue_map.num_slots); 2266 2267 /* Optional second URB write */ 2268 if (!inst->eot) { 2269 mrf = base_mrf + 1; 2270 2271 for (; slot < c->vue_map.num_slots; ++slot) { 2272 assert(mrf < max_usable_mrf); 2273 2274 emit_urb_slot(mrf++, c->vue_map.slot_to_vert_result[slot]); 2275 } 2276 2277 current_annotation = "URB write"; 2278 inst = emit(VS_OPCODE_URB_WRITE); 2279 inst->base_mrf = base_mrf; 2280 inst->mlen = align_interleaved_urb_mlen(brw, mrf - base_mrf); 2281 inst->eot = true; 2282 /* URB destination offset. In the previous write, we got MRFs 2283 * 2-13 minus the one header MRF, so 12 regs. URB offset is in 2284 * URB row increments, and each of our MRFs is half of one of 2285 * those, since we're doing interleaved writes. 2286 */ 2287 inst->offset = (max_usable_mrf - base_mrf) / 2; 2288 } 2289} 2290 2291src_reg 2292vec4_visitor::get_scratch_offset(vec4_instruction *inst, 2293 src_reg *reladdr, int reg_offset) 2294{ 2295 /* Because we store the values to scratch interleaved like our 2296 * vertex data, we need to scale the vec4 index by 2. 2297 */ 2298 int message_header_scale = 2; 2299 2300 /* Pre-gen6, the message header uses byte offsets instead of vec4 2301 * (16-byte) offset units. 2302 */ 2303 if (intel->gen < 6) 2304 message_header_scale *= 16; 2305 2306 if (reladdr) { 2307 src_reg index = src_reg(this, glsl_type::int_type); 2308 2309 emit_before(inst, ADD(dst_reg(index), *reladdr, src_reg(reg_offset))); 2310 emit_before(inst, MUL(dst_reg(index), 2311 index, src_reg(message_header_scale))); 2312 2313 return index; 2314 } else { 2315 return src_reg(reg_offset * message_header_scale); 2316 } 2317} 2318 2319src_reg 2320vec4_visitor::get_pull_constant_offset(vec4_instruction *inst, 2321 src_reg *reladdr, int reg_offset) 2322{ 2323 if (reladdr) { 2324 src_reg index = src_reg(this, glsl_type::int_type); 2325 2326 emit_before(inst, ADD(dst_reg(index), *reladdr, src_reg(reg_offset))); 2327 2328 /* Pre-gen6, the message header uses byte offsets instead of vec4 2329 * (16-byte) offset units. 2330 */ 2331 if (intel->gen < 6) { 2332 emit_before(inst, MUL(dst_reg(index), index, src_reg(16))); 2333 } 2334 2335 return index; 2336 } else { 2337 int message_header_scale = intel->gen < 6 ? 16 : 1; 2338 return src_reg(reg_offset * message_header_scale); 2339 } 2340} 2341 2342/** 2343 * Emits an instruction before @inst to load the value named by @orig_src 2344 * from scratch space at @base_offset to @temp. 2345 */ 2346void 2347vec4_visitor::emit_scratch_read(vec4_instruction *inst, 2348 dst_reg temp, src_reg orig_src, 2349 int base_offset) 2350{ 2351 int reg_offset = base_offset + orig_src.reg_offset; 2352 src_reg index = get_scratch_offset(inst, orig_src.reladdr, reg_offset); 2353 2354 emit_before(inst, SCRATCH_READ(temp, index)); 2355} 2356 2357/** 2358 * Emits an instruction after @inst to store the value to be written 2359 * to @orig_dst to scratch space at @base_offset, from @temp. 2360 */ 2361void 2362vec4_visitor::emit_scratch_write(vec4_instruction *inst, 2363 src_reg temp, dst_reg orig_dst, 2364 int base_offset) 2365{ 2366 int reg_offset = base_offset + orig_dst.reg_offset; 2367 src_reg index = get_scratch_offset(inst, orig_dst.reladdr, reg_offset); 2368 2369 dst_reg dst = dst_reg(brw_writemask(brw_vec8_grf(0, 0), 2370 orig_dst.writemask)); 2371 vec4_instruction *write = SCRATCH_WRITE(dst, temp, index); 2372 write->predicate = inst->predicate; 2373 write->ir = inst->ir; 2374 write->annotation = inst->annotation; 2375 inst->insert_after(write); 2376} 2377 2378/** 2379 * We can't generally support array access in GRF space, because a 2380 * single instruction's destination can only span 2 contiguous 2381 * registers. So, we send all GRF arrays that get variable index 2382 * access to scratch space. 2383 */ 2384void 2385vec4_visitor::move_grf_array_access_to_scratch() 2386{ 2387 int scratch_loc[this->virtual_grf_count]; 2388 2389 for (int i = 0; i < this->virtual_grf_count; i++) { 2390 scratch_loc[i] = -1; 2391 } 2392 2393 /* First, calculate the set of virtual GRFs that need to be punted 2394 * to scratch due to having any array access on them, and where in 2395 * scratch. 2396 */ 2397 foreach_list(node, &this->instructions) { 2398 vec4_instruction *inst = (vec4_instruction *)node; 2399 2400 if (inst->dst.file == GRF && inst->dst.reladdr && 2401 scratch_loc[inst->dst.reg] == -1) { 2402 scratch_loc[inst->dst.reg] = c->last_scratch; 2403 c->last_scratch += this->virtual_grf_sizes[inst->dst.reg] * 8 * 4; 2404 } 2405 2406 for (int i = 0 ; i < 3; i++) { 2407 src_reg *src = &inst->src[i]; 2408 2409 if (src->file == GRF && src->reladdr && 2410 scratch_loc[src->reg] == -1) { 2411 scratch_loc[src->reg] = c->last_scratch; 2412 c->last_scratch += this->virtual_grf_sizes[src->reg] * 8 * 4; 2413 } 2414 } 2415 } 2416 2417 /* Now, for anything that will be accessed through scratch, rewrite 2418 * it to load/store. Note that this is a _safe list walk, because 2419 * we may generate a new scratch_write instruction after the one 2420 * we're processing. 2421 */ 2422 foreach_list_safe(node, &this->instructions) { 2423 vec4_instruction *inst = (vec4_instruction *)node; 2424 2425 /* Set up the annotation tracking for new generated instructions. */ 2426 base_ir = inst->ir; 2427 current_annotation = inst->annotation; 2428 2429 if (inst->dst.file == GRF && scratch_loc[inst->dst.reg] != -1) { 2430 src_reg temp = src_reg(this, glsl_type::vec4_type); 2431 2432 emit_scratch_write(inst, temp, inst->dst, scratch_loc[inst->dst.reg]); 2433 2434 inst->dst.file = temp.file; 2435 inst->dst.reg = temp.reg; 2436 inst->dst.reg_offset = temp.reg_offset; 2437 inst->dst.reladdr = NULL; 2438 } 2439 2440 for (int i = 0 ; i < 3; i++) { 2441 if (inst->src[i].file != GRF || scratch_loc[inst->src[i].reg] == -1) 2442 continue; 2443 2444 dst_reg temp = dst_reg(this, glsl_type::vec4_type); 2445 2446 emit_scratch_read(inst, temp, inst->src[i], 2447 scratch_loc[inst->src[i].reg]); 2448 2449 inst->src[i].file = temp.file; 2450 inst->src[i].reg = temp.reg; 2451 inst->src[i].reg_offset = temp.reg_offset; 2452 inst->src[i].reladdr = NULL; 2453 } 2454 } 2455} 2456 2457/** 2458 * Emits an instruction before @inst to load the value named by @orig_src 2459 * from the pull constant buffer (surface) at @base_offset to @temp. 2460 */ 2461void 2462vec4_visitor::emit_pull_constant_load(vec4_instruction *inst, 2463 dst_reg temp, src_reg orig_src, 2464 int base_offset) 2465{ 2466 int reg_offset = base_offset + orig_src.reg_offset; 2467 src_reg index = get_pull_constant_offset(inst, orig_src.reladdr, reg_offset); 2468 vec4_instruction *load; 2469 2470 load = new(mem_ctx) vec4_instruction(this, VS_OPCODE_PULL_CONSTANT_LOAD, 2471 temp, index); 2472 load->base_mrf = 14; 2473 load->mlen = 1; 2474 emit_before(inst, load); 2475} 2476 2477/** 2478 * Implements array access of uniforms by inserting a 2479 * PULL_CONSTANT_LOAD instruction. 2480 * 2481 * Unlike temporary GRF array access (where we don't support it due to 2482 * the difficulty of doing relative addressing on instruction 2483 * destinations), we could potentially do array access of uniforms 2484 * that were loaded in GRF space as push constants. In real-world 2485 * usage we've seen, though, the arrays being used are always larger 2486 * than we could load as push constants, so just always move all 2487 * uniform array access out to a pull constant buffer. 2488 */ 2489void 2490vec4_visitor::move_uniform_array_access_to_pull_constants() 2491{ 2492 int pull_constant_loc[this->uniforms]; 2493 2494 for (int i = 0; i < this->uniforms; i++) { 2495 pull_constant_loc[i] = -1; 2496 } 2497 2498 /* Walk through and find array access of uniforms. Put a copy of that 2499 * uniform in the pull constant buffer. 2500 * 2501 * Note that we don't move constant-indexed accesses to arrays. No 2502 * testing has been done of the performance impact of this choice. 2503 */ 2504 foreach_list_safe(node, &this->instructions) { 2505 vec4_instruction *inst = (vec4_instruction *)node; 2506 2507 for (int i = 0 ; i < 3; i++) { 2508 if (inst->src[i].file != UNIFORM || !inst->src[i].reladdr) 2509 continue; 2510 2511 int uniform = inst->src[i].reg; 2512 2513 /* If this array isn't already present in the pull constant buffer, 2514 * add it. 2515 */ 2516 if (pull_constant_loc[uniform] == -1) { 2517 const float **values = &prog_data->param[uniform * 4]; 2518 2519 pull_constant_loc[uniform] = prog_data->nr_pull_params / 4; 2520 2521 for (int j = 0; j < uniform_size[uniform] * 4; j++) { 2522 prog_data->pull_param[prog_data->nr_pull_params++] = values[j]; 2523 } 2524 } 2525 2526 /* Set up the annotation tracking for new generated instructions. */ 2527 base_ir = inst->ir; 2528 current_annotation = inst->annotation; 2529 2530 dst_reg temp = dst_reg(this, glsl_type::vec4_type); 2531 2532 emit_pull_constant_load(inst, temp, inst->src[i], 2533 pull_constant_loc[uniform]); 2534 2535 inst->src[i].file = temp.file; 2536 inst->src[i].reg = temp.reg; 2537 inst->src[i].reg_offset = temp.reg_offset; 2538 inst->src[i].reladdr = NULL; 2539 } 2540 } 2541 2542 /* Now there are no accesses of the UNIFORM file with a reladdr, so 2543 * no need to track them as larger-than-vec4 objects. This will be 2544 * relied on in cutting out unused uniform vectors from push 2545 * constants. 2546 */ 2547 split_uniform_registers(); 2548} 2549 2550void 2551vec4_visitor::resolve_ud_negate(src_reg *reg) 2552{ 2553 if (reg->type != BRW_REGISTER_TYPE_UD || 2554 !reg->negate) 2555 return; 2556 2557 src_reg temp = src_reg(this, glsl_type::uvec4_type); 2558 emit(BRW_OPCODE_MOV, dst_reg(temp), *reg); 2559 *reg = temp; 2560} 2561 2562vec4_visitor::vec4_visitor(struct brw_vs_compile *c, 2563 struct gl_shader_program *prog, 2564 struct brw_shader *shader) 2565{ 2566 this->c = c; 2567 this->p = &c->func; 2568 this->brw = p->brw; 2569 this->intel = &brw->intel; 2570 this->ctx = &intel->ctx; 2571 this->prog = prog; 2572 this->shader = shader; 2573 2574 this->mem_ctx = ralloc_context(NULL); 2575 this->failed = false; 2576 2577 this->base_ir = NULL; 2578 this->current_annotation = NULL; 2579 2580 this->c = c; 2581 this->vp = (struct gl_vertex_program *) 2582 prog->_LinkedShaders[MESA_SHADER_VERTEX]->Program; 2583 this->prog_data = &c->prog_data; 2584 2585 this->variable_ht = hash_table_ctor(0, 2586 hash_table_pointer_hash, 2587 hash_table_pointer_compare); 2588 2589 this->virtual_grf_def = NULL; 2590 this->virtual_grf_use = NULL; 2591 this->virtual_grf_sizes = NULL; 2592 this->virtual_grf_count = 0; 2593 this->virtual_grf_reg_map = NULL; 2594 this->virtual_grf_reg_count = 0; 2595 this->virtual_grf_array_size = 0; 2596 this->live_intervals_valid = false; 2597 2598 this->uniforms = 0; 2599} 2600 2601vec4_visitor::~vec4_visitor() 2602{ 2603 ralloc_free(this->mem_ctx); 2604 hash_table_dtor(this->variable_ht); 2605} 2606 2607 2608void 2609vec4_visitor::fail(const char *format, ...) 2610{ 2611 va_list va; 2612 char *msg; 2613 2614 if (failed) 2615 return; 2616 2617 failed = true; 2618 2619 va_start(va, format); 2620 msg = ralloc_vasprintf(mem_ctx, format, va); 2621 va_end(va); 2622 msg = ralloc_asprintf(mem_ctx, "VS compile failed: %s\n", msg); 2623 2624 this->fail_msg = msg; 2625 2626 if (INTEL_DEBUG & DEBUG_VS) { 2627 fprintf(stderr, "%s", msg); 2628 } 2629} 2630 2631} /* namespace brw */ 2632