2 * Copyright © 2010 Intel Corporation
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:
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
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
24 /** @file brw_fs_visitor.cpp
26 * This file supports generating the FS LIR from the GLSL IR. The LIR
27 * makes it easier to do backend-specific optimizations than doing so
28 * in the GLSL IR or in the native code.
32 #include <sys/types.h>
34 #include "main/macros.h"
35 #include "main/shaderobj.h"
36 #include "program/prog_parameter.h"
37 #include "program/prog_print.h"
38 #include "program/prog_optimize.h"
39 #include "program/register_allocate.h"
40 #include "program/sampler.h"
41 #include "program/hash_table.h"
42 #include "brw_context.h"
47 #include "main/uniforms.h"
48 #include "glsl/glsl_types.h"
49 #include "glsl/ir_optimization.h"
52 fs_visitor::visit(ir_variable *ir)
56 if (variable_storage(ir))
59 if (ir->mode == ir_var_shader_in) {
60 if (!strcmp(ir->name, "gl_FragCoord")) {
61 reg = emit_fragcoord_interpolation(ir);
62 } else if (!strcmp(ir->name, "gl_FrontFacing")) {
63 reg = emit_frontfacing_interpolation(ir);
65 reg = emit_general_interpolation(ir);
68 hash_table_insert(this->variable_ht, reg, ir);
70 } else if (ir->mode == ir_var_shader_out) {
71 reg = new(this->mem_ctx) fs_reg(this, ir->type);
74 assert(ir->location == FRAG_RESULT_DATA0);
75 assert(ir->index == 1);
76 this->dual_src_output = *reg;
77 } else if (ir->location == FRAG_RESULT_COLOR) {
78 /* Writing gl_FragColor outputs to all color regions. */
79 for (unsigned int i = 0; i < MAX2(c->key.nr_color_regions, 1); i++) {
80 this->outputs[i] = *reg;
81 this->output_components[i] = 4;
83 } else if (ir->location == FRAG_RESULT_DEPTH) {
84 this->frag_depth = *reg;
86 /* gl_FragData or a user-defined FS output */
87 assert(ir->location >= FRAG_RESULT_DATA0 &&
88 ir->location < FRAG_RESULT_DATA0 + BRW_MAX_DRAW_BUFFERS);
91 ir->type->is_array() ? ir->type->fields.array->vector_elements
92 : ir->type->vector_elements;
94 /* General color output. */
95 for (unsigned int i = 0; i < MAX2(1, ir->type->length); i++) {
96 int output = ir->location - FRAG_RESULT_DATA0 + i;
97 this->outputs[output] = *reg;
98 this->outputs[output].reg_offset += vector_elements * i;
99 this->output_components[output] = vector_elements;
102 } else if (ir->mode == ir_var_uniform) {
103 int param_index = c->prog_data.nr_params;
105 /* Thanks to the lower_ubo_reference pass, we will see only
106 * ir_binop_ubo_load expressions and not ir_dereference_variable for UBO
107 * variables, so no need for them to be in variable_ht.
109 if (ir->is_in_uniform_block())
112 if (dispatch_width == 16) {
113 if (!variable_storage(ir)) {
114 fail("Failed to find uniform '%s' in 16-wide\n", ir->name);
119 param_size[param_index] = type_size(ir->type);
120 if (!strncmp(ir->name, "gl_", 3)) {
121 setup_builtin_uniform_values(ir);
123 setup_uniform_values(ir);
126 reg = new(this->mem_ctx) fs_reg(UNIFORM, param_index);
127 reg->type = brw_type_for_base_type(ir->type);
131 reg = new(this->mem_ctx) fs_reg(this, ir->type);
133 hash_table_insert(this->variable_ht, reg, ir);
137 fs_visitor::visit(ir_dereference_variable *ir)
139 fs_reg *reg = variable_storage(ir->var);
144 fs_visitor::visit(ir_dereference_record *ir)
146 const glsl_type *struct_type = ir->record->type;
148 ir->record->accept(this);
150 unsigned int offset = 0;
151 for (unsigned int i = 0; i < struct_type->length; i++) {
152 if (strcmp(struct_type->fields.structure[i].name, ir->field) == 0)
154 offset += type_size(struct_type->fields.structure[i].type);
156 this->result.reg_offset += offset;
157 this->result.type = brw_type_for_base_type(ir->type);
161 fs_visitor::visit(ir_dereference_array *ir)
163 ir_constant *constant_index;
165 int element_size = type_size(ir->type);
167 constant_index = ir->array_index->as_constant();
169 ir->array->accept(this);
171 src.type = brw_type_for_base_type(ir->type);
173 if (constant_index) {
174 assert(src.file == UNIFORM || src.file == GRF);
175 src.reg_offset += constant_index->value.i[0] * element_size;
177 /* Variable index array dereference. We attach the variable index
178 * component to the reg as a pointer to a register containing the
179 * offset. Currently only uniform arrays are supported in this patch,
180 * and that reladdr pointer is resolved by
181 * move_uniform_array_access_to_pull_constants(). All other array types
182 * are lowered by lower_variable_index_to_cond_assign().
184 ir->array_index->accept(this);
187 index_reg = fs_reg(this, glsl_type::int_type);
188 emit(BRW_OPCODE_MUL, index_reg, this->result, fs_reg(element_size));
191 emit(BRW_OPCODE_ADD, index_reg, *src.reladdr, index_reg);
194 src.reladdr = ralloc(mem_ctx, fs_reg);
195 memcpy(src.reladdr, &index_reg, sizeof(index_reg));
201 fs_visitor::emit_lrp(fs_reg dst, fs_reg x, fs_reg y, fs_reg a)
204 !x.is_valid_3src() ||
205 !y.is_valid_3src() ||
206 !a.is_valid_3src()) {
207 /* We can't use the LRP instruction. Emit x*(1-a) + y*a. */
208 fs_reg y_times_a = fs_reg(this, glsl_type::float_type);
209 fs_reg one_minus_a = fs_reg(this, glsl_type::float_type);
210 fs_reg x_times_one_minus_a = fs_reg(this, glsl_type::float_type);
212 emit(MUL(y_times_a, y, a));
214 a.negate = !a.negate;
215 emit(ADD(one_minus_a, a, fs_reg(1.0f)));
216 emit(MUL(x_times_one_minus_a, x, one_minus_a));
218 emit(ADD(dst, x_times_one_minus_a, y_times_a));
220 /* The LRP instruction actually does op1 * op0 + op2 * (1 - op0), so
221 * we need to reorder the operands.
223 emit(LRP(dst, a, y, x));
228 fs_visitor::emit_minmax(uint32_t conditionalmod, fs_reg dst,
229 fs_reg src0, fs_reg src1)
234 inst = emit(BRW_OPCODE_SEL, dst, src0, src1);
235 inst->conditional_mod = conditionalmod;
237 emit(CMP(reg_null_d, src0, src1, conditionalmod));
239 inst = emit(BRW_OPCODE_SEL, dst, src0, src1);
240 inst->predicate = BRW_PREDICATE_NORMAL;
244 /* Instruction selection: Produce a MOV.sat instead of
245 * MIN(MAX(val, 0), 1) when possible.
248 fs_visitor::try_emit_saturate(ir_expression *ir)
250 ir_rvalue *sat_val = ir->as_rvalue_to_saturate();
255 fs_inst *pre_inst = (fs_inst *) this->instructions.get_tail();
257 sat_val->accept(this);
258 fs_reg src = this->result;
260 fs_inst *last_inst = (fs_inst *) this->instructions.get_tail();
262 /* If the last instruction from our accept() didn't generate our
263 * src, generate a saturated MOV
265 fs_inst *modify = get_instruction_generating_reg(pre_inst, last_inst, src);
266 if (!modify || modify->regs_written != 1) {
267 this->result = fs_reg(this, ir->type);
268 fs_inst *inst = emit(MOV(this->result, src));
269 inst->saturate = true;
271 modify->saturate = true;
280 fs_visitor::try_emit_mad(ir_expression *ir, int mul_arg)
282 /* 3-src instructions were introduced in gen6. */
286 /* MAD can only handle floating-point data. */
287 if (ir->type != glsl_type::float_type)
290 ir_rvalue *nonmul = ir->operands[1 - mul_arg];
291 ir_expression *mul = ir->operands[mul_arg]->as_expression();
293 if (!mul || mul->operation != ir_binop_mul)
296 if (nonmul->as_constant() ||
297 mul->operands[0]->as_constant() ||
298 mul->operands[1]->as_constant())
301 nonmul->accept(this);
302 fs_reg src0 = this->result;
304 mul->operands[0]->accept(this);
305 fs_reg src1 = this->result;
307 mul->operands[1]->accept(this);
308 fs_reg src2 = this->result;
310 this->result = fs_reg(this, ir->type);
311 emit(BRW_OPCODE_MAD, this->result, src0, src1, src2);
317 fs_visitor::visit(ir_expression *ir)
319 unsigned int operand;
323 assert(ir->get_num_operands() <= 3);
325 if (try_emit_saturate(ir))
327 if (ir->operation == ir_binop_add) {
328 if (try_emit_mad(ir, 0) || try_emit_mad(ir, 1))
332 for (operand = 0; operand < ir->get_num_operands(); operand++) {
333 ir->operands[operand]->accept(this);
334 if (this->result.file == BAD_FILE) {
335 fail("Failed to get tree for expression operand:\n");
336 ir->operands[operand]->print();
339 assert(this->result.is_valid_3src());
340 op[operand] = this->result;
342 /* Matrix expression operands should have been broken down to vector
343 * operations already.
345 assert(!ir->operands[operand]->type->is_matrix());
346 /* And then those vector operands should have been broken down to scalar.
348 assert(!ir->operands[operand]->type->is_vector());
351 /* Storage for our result. If our result goes into an assignment, it will
352 * just get copy-propagated out, so no worries.
354 this->result = fs_reg(this, ir->type);
356 switch (ir->operation) {
357 case ir_unop_logic_not:
358 /* Note that BRW_OPCODE_NOT is not appropriate here, since it is
359 * ones complement of the whole register, not just bit 0.
361 emit(XOR(this->result, op[0], fs_reg(1)));
364 op[0].negate = !op[0].negate;
365 emit(MOV(this->result, op[0]));
369 op[0].negate = false;
370 emit(MOV(this->result, op[0]));
373 temp = fs_reg(this, ir->type);
375 emit(MOV(this->result, fs_reg(0.0f)));
377 emit(CMP(reg_null_f, op[0], fs_reg(0.0f), BRW_CONDITIONAL_G));
378 inst = emit(MOV(this->result, fs_reg(1.0f)));
379 inst->predicate = BRW_PREDICATE_NORMAL;
381 emit(CMP(reg_null_f, op[0], fs_reg(0.0f), BRW_CONDITIONAL_L));
382 inst = emit(MOV(this->result, fs_reg(-1.0f)));
383 inst->predicate = BRW_PREDICATE_NORMAL;
387 emit_math(SHADER_OPCODE_RCP, this->result, op[0]);
391 emit_math(SHADER_OPCODE_EXP2, this->result, op[0]);
394 emit_math(SHADER_OPCODE_LOG2, this->result, op[0]);
398 assert(!"not reached: should be handled by ir_explog_to_explog2");
401 case ir_unop_sin_reduced:
402 emit_math(SHADER_OPCODE_SIN, this->result, op[0]);
405 case ir_unop_cos_reduced:
406 emit_math(SHADER_OPCODE_COS, this->result, op[0]);
410 emit(FS_OPCODE_DDX, this->result, op[0]);
413 emit(FS_OPCODE_DDY, this->result, op[0]);
417 emit(ADD(this->result, op[0], op[1]));
420 assert(!"not reached: should be handled by ir_sub_to_add_neg");
424 if (ir->type->is_integer()) {
425 /* For integer multiplication, the MUL uses the low 16 bits
426 * of one of the operands (src0 on gen6, src1 on gen7). The
427 * MACH accumulates in the contribution of the upper 16 bits
430 * FINISHME: Emit just the MUL if we know an operand is small
433 if (brw->gen >= 7 && dispatch_width == 16)
434 fail("16-wide explicit accumulator operands unsupported\n");
436 struct brw_reg acc = retype(brw_acc_reg(), this->result.type);
438 emit(MUL(acc, op[0], op[1]));
439 emit(MACH(reg_null_d, op[0], op[1]));
440 emit(MOV(this->result, fs_reg(acc)));
442 emit(MUL(this->result, op[0], op[1]));
445 case ir_binop_imul_high: {
446 if (brw->gen >= 7 && dispatch_width == 16)
447 fail("16-wide explicit accumulator operands unsupported\n");
449 struct brw_reg acc = retype(brw_acc_reg(), this->result.type);
451 emit(MUL(acc, op[0], op[1]));
452 emit(MACH(this->result, op[0], op[1]));
456 /* Floating point should be lowered by DIV_TO_MUL_RCP in the compiler. */
457 assert(ir->type->is_integer());
458 emit_math(SHADER_OPCODE_INT_QUOTIENT, this->result, op[0], op[1]);
460 case ir_binop_carry: {
461 if (brw->gen >= 7 && dispatch_width == 16)
462 fail("16-wide explicit accumulator operands unsupported\n");
464 struct brw_reg acc = retype(brw_acc_reg(), BRW_REGISTER_TYPE_UD);
466 emit(ADDC(reg_null_ud, op[0], op[1]));
467 emit(MOV(this->result, fs_reg(acc)));
470 case ir_binop_borrow: {
471 if (brw->gen >= 7 && dispatch_width == 16)
472 fail("16-wide explicit accumulator operands unsupported\n");
474 struct brw_reg acc = retype(brw_acc_reg(), BRW_REGISTER_TYPE_UD);
476 emit(SUBB(reg_null_ud, op[0], op[1]));
477 emit(MOV(this->result, fs_reg(acc)));
481 /* Floating point should be lowered by MOD_TO_FRACT in the compiler. */
482 assert(ir->type->is_integer());
483 emit_math(SHADER_OPCODE_INT_REMAINDER, this->result, op[0], op[1]);
487 case ir_binop_greater:
488 case ir_binop_lequal:
489 case ir_binop_gequal:
491 case ir_binop_all_equal:
492 case ir_binop_nequal:
493 case ir_binop_any_nequal:
494 resolve_bool_comparison(ir->operands[0], &op[0]);
495 resolve_bool_comparison(ir->operands[1], &op[1]);
497 emit(CMP(this->result, op[0], op[1],
498 brw_conditional_for_comparison(ir->operation)));
501 case ir_binop_logic_xor:
502 emit(XOR(this->result, op[0], op[1]));
505 case ir_binop_logic_or:
506 emit(OR(this->result, op[0], op[1]));
509 case ir_binop_logic_and:
510 emit(AND(this->result, op[0], op[1]));
515 assert(!"not reached: should be handled by brw_fs_channel_expressions");
519 assert(!"not reached: should be handled by lower_noise");
522 case ir_quadop_vector:
523 assert(!"not reached: should be handled by lower_quadop_vector");
526 case ir_binop_vector_extract:
527 assert(!"not reached: should be handled by lower_vec_index_to_cond_assign()");
530 case ir_triop_vector_insert:
531 assert(!"not reached: should be handled by lower_vector_insert()");
535 assert(!"not reached: should be handled by ldexp_to_arith()");
539 emit_math(SHADER_OPCODE_SQRT, this->result, op[0]);
543 emit_math(SHADER_OPCODE_RSQ, this->result, op[0]);
546 case ir_unop_bitcast_i2f:
547 case ir_unop_bitcast_u2f:
548 op[0].type = BRW_REGISTER_TYPE_F;
549 this->result = op[0];
552 case ir_unop_bitcast_f2u:
553 op[0].type = BRW_REGISTER_TYPE_UD;
554 this->result = op[0];
557 case ir_unop_bitcast_f2i:
558 op[0].type = BRW_REGISTER_TYPE_D;
559 this->result = op[0];
565 emit(MOV(this->result, op[0]));
569 emit(AND(this->result, op[0], fs_reg(1)));
572 temp = fs_reg(this, glsl_type::int_type);
573 emit(AND(temp, op[0], fs_reg(1)));
574 emit(MOV(this->result, temp));
578 emit(CMP(this->result, op[0], fs_reg(0.0f), BRW_CONDITIONAL_NZ));
581 emit(CMP(this->result, op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
585 emit(RNDZ(this->result, op[0]));
588 op[0].negate = !op[0].negate;
589 emit(RNDD(this->result, op[0]));
590 this->result.negate = true;
593 emit(RNDD(this->result, op[0]));
596 emit(FRC(this->result, op[0]));
598 case ir_unop_round_even:
599 emit(RNDE(this->result, op[0]));
604 resolve_ud_negate(&op[0]);
605 resolve_ud_negate(&op[1]);
606 emit_minmax(ir->operation == ir_binop_min ?
607 BRW_CONDITIONAL_L : BRW_CONDITIONAL_GE,
608 this->result, op[0], op[1]);
610 case ir_unop_pack_snorm_2x16:
611 case ir_unop_pack_snorm_4x8:
612 case ir_unop_pack_unorm_2x16:
613 case ir_unop_pack_unorm_4x8:
614 case ir_unop_unpack_snorm_2x16:
615 case ir_unop_unpack_snorm_4x8:
616 case ir_unop_unpack_unorm_2x16:
617 case ir_unop_unpack_unorm_4x8:
618 case ir_unop_unpack_half_2x16:
619 case ir_unop_pack_half_2x16:
620 assert(!"not reached: should be handled by lower_packing_builtins");
622 case ir_unop_unpack_half_2x16_split_x:
623 emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X, this->result, op[0]);
625 case ir_unop_unpack_half_2x16_split_y:
626 emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y, this->result, op[0]);
629 emit_math(SHADER_OPCODE_POW, this->result, op[0], op[1]);
632 case ir_unop_bitfield_reverse:
633 emit(BFREV(this->result, op[0]));
635 case ir_unop_bit_count:
636 emit(CBIT(this->result, op[0]));
638 case ir_unop_find_msb:
639 temp = fs_reg(this, glsl_type::uint_type);
640 emit(FBH(temp, op[0]));
642 /* FBH counts from the MSB side, while GLSL's findMSB() wants the count
643 * from the LSB side. If FBH didn't return an error (0xFFFFFFFF), then
644 * subtract the result from 31 to convert the MSB count into an LSB count.
647 /* FBH only supports UD type for dst, so use a MOV to convert UD to D. */
648 emit(MOV(this->result, temp));
649 emit(CMP(reg_null_d, this->result, fs_reg(-1), BRW_CONDITIONAL_NZ));
652 inst = emit(ADD(this->result, temp, fs_reg(31)));
653 inst->predicate = BRW_PREDICATE_NORMAL;
655 case ir_unop_find_lsb:
656 emit(FBL(this->result, op[0]));
658 case ir_triop_bitfield_extract:
659 /* Note that the instruction's argument order is reversed from GLSL
662 emit(BFE(this->result, op[2], op[1], op[0]));
665 emit(BFI1(this->result, op[0], op[1]));
668 emit(BFI2(this->result, op[0], op[1], op[2]));
670 case ir_quadop_bitfield_insert:
671 assert(!"not reached: should be handled by "
672 "lower_instructions::bitfield_insert_to_bfm_bfi");
675 case ir_unop_bit_not:
676 emit(NOT(this->result, op[0]));
678 case ir_binop_bit_and:
679 emit(AND(this->result, op[0], op[1]));
681 case ir_binop_bit_xor:
682 emit(XOR(this->result, op[0], op[1]));
684 case ir_binop_bit_or:
685 emit(OR(this->result, op[0], op[1]));
688 case ir_binop_lshift:
689 emit(SHL(this->result, op[0], op[1]));
692 case ir_binop_rshift:
693 if (ir->type->base_type == GLSL_TYPE_INT)
694 emit(ASR(this->result, op[0], op[1]));
696 emit(SHR(this->result, op[0], op[1]));
698 case ir_binop_pack_half_2x16_split:
699 emit(FS_OPCODE_PACK_HALF_2x16_SPLIT, this->result, op[0], op[1]);
701 case ir_binop_ubo_load: {
702 /* This IR node takes a constant uniform block and a constant or
703 * variable byte offset within the block and loads a vector from that.
705 ir_constant *uniform_block = ir->operands[0]->as_constant();
706 ir_constant *const_offset = ir->operands[1]->as_constant();
707 fs_reg surf_index = fs_reg(c->prog_data.base.binding_table.ubo_start +
708 uniform_block->value.u[0]);
710 fs_reg packed_consts = fs_reg(this, glsl_type::float_type);
711 packed_consts.type = result.type;
713 fs_reg const_offset_reg = fs_reg(const_offset->value.u[0] & ~15);
714 emit(fs_inst(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD,
715 packed_consts, surf_index, const_offset_reg));
717 packed_consts.smear = const_offset->value.u[0] % 16 / 4;
718 for (int i = 0; i < ir->type->vector_elements; i++) {
719 /* UBO bools are any nonzero value. We consider bools to be
720 * values with the low bit set to 1. Convert them using CMP.
722 if (ir->type->base_type == GLSL_TYPE_BOOL) {
723 emit(CMP(result, packed_consts, fs_reg(0u), BRW_CONDITIONAL_NZ));
725 emit(MOV(result, packed_consts));
728 packed_consts.smear++;
731 /* The std140 packing rules don't allow vectors to cross 16-byte
732 * boundaries, and a reg is 32 bytes.
734 assert(packed_consts.smear < 8);
737 /* Turn the byte offset into a dword offset. */
738 fs_reg base_offset = fs_reg(this, glsl_type::int_type);
739 emit(SHR(base_offset, op[1], fs_reg(2)));
741 for (int i = 0; i < ir->type->vector_elements; i++) {
742 emit(VARYING_PULL_CONSTANT_LOAD(result, surf_index,
745 if (ir->type->base_type == GLSL_TYPE_BOOL)
746 emit(CMP(result, result, fs_reg(0), BRW_CONDITIONAL_NZ));
752 result.reg_offset = 0;
757 /* Note that the instruction's argument order is reversed from GLSL
760 emit(MAD(this->result, op[2], op[1], op[0]));
764 emit_lrp(this->result, op[0], op[1], op[2]);
768 emit(CMP(reg_null_d, op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
769 inst = emit(BRW_OPCODE_SEL, this->result, op[1], op[2]);
770 inst->predicate = BRW_PREDICATE_NORMAL;
776 fs_visitor::emit_assignment_writes(fs_reg &l, fs_reg &r,
777 const glsl_type *type, bool predicated)
779 switch (type->base_type) {
780 case GLSL_TYPE_FLOAT:
784 for (unsigned int i = 0; i < type->components(); i++) {
785 l.type = brw_type_for_base_type(type);
786 r.type = brw_type_for_base_type(type);
788 if (predicated || !l.equals(r)) {
789 fs_inst *inst = emit(MOV(l, r));
790 inst->predicate = predicated ? BRW_PREDICATE_NORMAL : BRW_PREDICATE_NONE;
797 case GLSL_TYPE_ARRAY:
798 for (unsigned int i = 0; i < type->length; i++) {
799 emit_assignment_writes(l, r, type->fields.array, predicated);
803 case GLSL_TYPE_STRUCT:
804 for (unsigned int i = 0; i < type->length; i++) {
805 emit_assignment_writes(l, r, type->fields.structure[i].type,
810 case GLSL_TYPE_SAMPLER:
811 case GLSL_TYPE_ATOMIC_UINT:
815 case GLSL_TYPE_ERROR:
816 case GLSL_TYPE_INTERFACE:
817 assert(!"not reached");
822 /* If the RHS processing resulted in an instruction generating a
823 * temporary value, and it would be easy to rewrite the instruction to
824 * generate its result right into the LHS instead, do so. This ends
825 * up reliably removing instructions where it can be tricky to do so
826 * later without real UD chain information.
829 fs_visitor::try_rewrite_rhs_to_dst(ir_assignment *ir,
832 fs_inst *pre_rhs_inst,
833 fs_inst *last_rhs_inst)
835 /* Only attempt if we're doing a direct assignment. */
837 !(ir->lhs->type->is_scalar() ||
838 (ir->lhs->type->is_vector() &&
839 ir->write_mask == (1 << ir->lhs->type->vector_elements) - 1)))
842 /* Make sure the last instruction generated our source reg. */
843 fs_inst *modify = get_instruction_generating_reg(pre_rhs_inst,
849 /* If last_rhs_inst wrote a different number of components than our LHS,
850 * we can't safely rewrite it.
852 if (virtual_grf_sizes[dst.reg] != modify->regs_written)
855 /* Success! Rewrite the instruction. */
862 fs_visitor::visit(ir_assignment *ir)
867 /* FINISHME: arrays on the lhs */
868 ir->lhs->accept(this);
871 fs_inst *pre_rhs_inst = (fs_inst *) this->instructions.get_tail();
873 ir->rhs->accept(this);
876 fs_inst *last_rhs_inst = (fs_inst *) this->instructions.get_tail();
878 assert(l.file != BAD_FILE);
879 assert(r.file != BAD_FILE);
881 if (try_rewrite_rhs_to_dst(ir, l, r, pre_rhs_inst, last_rhs_inst))
885 emit_bool_to_cond_code(ir->condition);
888 if (ir->lhs->type->is_scalar() ||
889 ir->lhs->type->is_vector()) {
890 for (int i = 0; i < ir->lhs->type->vector_elements; i++) {
891 if (ir->write_mask & (1 << i)) {
892 inst = emit(MOV(l, r));
894 inst->predicate = BRW_PREDICATE_NORMAL;
900 emit_assignment_writes(l, r, ir->lhs->type, ir->condition != NULL);
905 fs_visitor::emit_texture_gen4(ir_texture *ir, fs_reg dst, fs_reg coordinate,
906 fs_reg shadow_c, fs_reg lod, fs_reg dPdy)
916 if (ir->shadow_comparitor) {
917 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
918 emit(MOV(fs_reg(MRF, base_mrf + mlen + i), coordinate));
919 coordinate.reg_offset++;
922 /* gen4's SIMD8 sampler always has the slots for u,v,r present.
923 * the unused slots must be zeroed.
925 for (int i = ir->coordinate->type->vector_elements; i < 3; i++) {
926 emit(MOV(fs_reg(MRF, base_mrf + mlen + i), fs_reg(0.0f)));
930 if (ir->op == ir_tex) {
931 /* There's no plain shadow compare message, so we use shadow
932 * compare with a bias of 0.0.
934 emit(MOV(fs_reg(MRF, base_mrf + mlen), fs_reg(0.0f)));
936 } else if (ir->op == ir_txb || ir->op == ir_txl) {
937 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
940 assert(!"Should not get here.");
943 emit(MOV(fs_reg(MRF, base_mrf + mlen), shadow_c));
945 } else if (ir->op == ir_tex) {
946 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
947 emit(MOV(fs_reg(MRF, base_mrf + mlen + i), coordinate));
948 coordinate.reg_offset++;
950 /* zero the others. */
951 for (int i = ir->coordinate->type->vector_elements; i<3; i++) {
952 emit(MOV(fs_reg(MRF, base_mrf + mlen + i), fs_reg(0.0f)));
954 /* gen4's SIMD8 sampler always has the slots for u,v,r present. */
956 } else if (ir->op == ir_txd) {
959 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
960 emit(MOV(fs_reg(MRF, base_mrf + mlen + i), coordinate));
961 coordinate.reg_offset++;
963 /* the slots for u and v are always present, but r is optional */
964 mlen += MAX2(ir->coordinate->type->vector_elements, 2);
967 * dPdx = dudx, dvdx, drdx
968 * dPdy = dudy, dvdy, drdy
970 * 1-arg: Does not exist.
972 * 2-arg: dudx dvdx dudy dvdy
973 * dPdx.x dPdx.y dPdy.x dPdy.y
976 * 3-arg: dudx dvdx drdx dudy dvdy drdy
977 * dPdx.x dPdx.y dPdx.z dPdy.x dPdy.y dPdy.z
980 for (int i = 0; i < ir->lod_info.grad.dPdx->type->vector_elements; i++) {
981 emit(MOV(fs_reg(MRF, base_mrf + mlen), dPdx));
984 mlen += MAX2(ir->lod_info.grad.dPdx->type->vector_elements, 2);
986 for (int i = 0; i < ir->lod_info.grad.dPdy->type->vector_elements; i++) {
987 emit(MOV(fs_reg(MRF, base_mrf + mlen), dPdy));
990 mlen += MAX2(ir->lod_info.grad.dPdy->type->vector_elements, 2);
991 } else if (ir->op == ir_txs) {
992 /* There's no SIMD8 resinfo message on Gen4. Use SIMD16 instead. */
994 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), lod));
997 /* Oh joy. gen4 doesn't have SIMD8 non-shadow-compare bias/lod
998 * instructions. We'll need to do SIMD16 here.
1001 assert(ir->op == ir_txb || ir->op == ir_txl || ir->op == ir_txf);
1003 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
1004 emit(MOV(fs_reg(MRF, base_mrf + mlen + i * 2, coordinate.type),
1006 coordinate.reg_offset++;
1009 /* Initialize the rest of u/v/r with 0.0. Empirically, this seems to
1010 * be necessary for TXF (ld), but seems wise to do for all messages.
1012 for (int i = ir->coordinate->type->vector_elements; i < 3; i++) {
1013 emit(MOV(fs_reg(MRF, base_mrf + mlen + i * 2), fs_reg(0.0f)));
1016 /* lod/bias appears after u/v/r. */
1019 emit(MOV(fs_reg(MRF, base_mrf + mlen, lod.type), lod));
1022 /* The unused upper half. */
1027 /* Now, since we're doing simd16, the return is 2 interleaved
1028 * vec4s where the odd-indexed ones are junk. We'll need to move
1029 * this weirdness around to the expected layout.
1032 dst = fs_reg(GRF, virtual_grf_alloc(8),
1034 brw_type_for_base_type(ir->type) :
1035 BRW_REGISTER_TYPE_F));
1038 fs_inst *inst = NULL;
1041 inst = emit(SHADER_OPCODE_TEX, dst);
1044 inst = emit(FS_OPCODE_TXB, dst);
1047 inst = emit(SHADER_OPCODE_TXL, dst);
1050 inst = emit(SHADER_OPCODE_TXD, dst);
1053 inst = emit(SHADER_OPCODE_TXS, dst);
1056 inst = emit(SHADER_OPCODE_TXF, dst);
1059 fail("unrecognized texture opcode");
1061 inst->base_mrf = base_mrf;
1063 inst->header_present = true;
1064 inst->regs_written = simd16 ? 8 : 4;
1067 for (int i = 0; i < 4; i++) {
1068 emit(MOV(orig_dst, dst));
1069 orig_dst.reg_offset++;
1070 dst.reg_offset += 2;
1077 /* gen5's sampler has slots for u, v, r, array index, then optional
1078 * parameters like shadow comparitor or LOD bias. If optional
1079 * parameters aren't present, those base slots are optional and don't
1080 * need to be included in the message.
1082 * We don't fill in the unnecessary slots regardless, which may look
1083 * surprising in the disassembly.
1086 fs_visitor::emit_texture_gen5(ir_texture *ir, fs_reg dst, fs_reg coordinate,
1087 fs_reg shadow_c, fs_reg lod, fs_reg lod2,
1088 fs_reg sample_index)
1092 int reg_width = dispatch_width / 8;
1093 bool header_present = false;
1094 const int vector_elements =
1095 ir->coordinate ? ir->coordinate->type->vector_elements : 0;
1098 /* The offsets set up by the ir_texture visitor are in the
1099 * m1 header, so we can't go headerless.
1101 header_present = true;
1106 for (int i = 0; i < vector_elements; i++) {
1107 emit(MOV(fs_reg(MRF, base_mrf + mlen + i * reg_width, coordinate.type),
1109 coordinate.reg_offset++;
1111 mlen += vector_elements * reg_width;
1113 if (ir->shadow_comparitor) {
1114 mlen = MAX2(mlen, header_present + 4 * reg_width);
1116 emit(MOV(fs_reg(MRF, base_mrf + mlen), shadow_c));
1120 fs_inst *inst = NULL;
1123 inst = emit(SHADER_OPCODE_TEX, dst);
1126 mlen = MAX2(mlen, header_present + 4 * reg_width);
1127 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
1130 inst = emit(FS_OPCODE_TXB, dst);
1133 mlen = MAX2(mlen, header_present + 4 * reg_width);
1134 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
1137 inst = emit(SHADER_OPCODE_TXL, dst);
1140 mlen = MAX2(mlen, header_present + 4 * reg_width); /* skip over 'ai' */
1144 * dPdx = dudx, dvdx, drdx
1145 * dPdy = dudy, dvdy, drdy
1147 * Load up these values:
1148 * - dudx dudy dvdx dvdy drdx drdy
1149 * - dPdx.x dPdy.x dPdx.y dPdy.y dPdx.z dPdy.z
1151 for (int i = 0; i < ir->lod_info.grad.dPdx->type->vector_elements; i++) {
1152 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
1156 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod2));
1161 inst = emit(SHADER_OPCODE_TXD, dst);
1165 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), lod));
1167 inst = emit(SHADER_OPCODE_TXS, dst);
1169 case ir_query_levels:
1170 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), fs_reg(0u)));
1172 inst = emit(SHADER_OPCODE_TXS, dst);
1175 mlen = header_present + 4 * reg_width;
1176 emit(MOV(fs_reg(MRF, base_mrf + mlen - reg_width, BRW_REGISTER_TYPE_UD), lod));
1177 inst = emit(SHADER_OPCODE_TXF, dst);
1180 mlen = header_present + 4 * reg_width;
1183 emit(MOV(fs_reg(MRF, base_mrf + mlen - reg_width, BRW_REGISTER_TYPE_UD), fs_reg(0)));
1185 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), sample_index));
1187 inst = emit(SHADER_OPCODE_TXF_MS, dst);
1190 inst = emit(SHADER_OPCODE_LOD, dst);
1193 inst = emit(SHADER_OPCODE_TG4, dst);
1196 fail("unrecognized texture opcode");
1199 inst->base_mrf = base_mrf;
1201 inst->header_present = header_present;
1202 inst->regs_written = 4;
1205 fail("Message length >11 disallowed by hardware\n");
1212 fs_visitor::emit_texture_gen7(ir_texture *ir, fs_reg dst, fs_reg coordinate,
1213 fs_reg shadow_c, fs_reg lod, fs_reg lod2,
1214 fs_reg sample_index)
1216 int reg_width = dispatch_width / 8;
1217 bool header_present = false;
1219 fs_reg payload = fs_reg(this, glsl_type::float_type);
1220 fs_reg next = payload;
1222 if (ir->op == ir_tg4 || (ir->offset && ir->op != ir_txf)) {
1223 /* For general texture offsets (no txf workaround), we need a header to
1224 * put them in. Note that for 16-wide we're making space for two actual
1225 * hardware registers here, so the emit will have to fix up for this.
1227 * * ir4_tg4 needs to place its channel select in the header,
1228 * for interaction with ARB_texture_swizzle
1230 header_present = true;
1234 if (ir->shadow_comparitor) {
1235 emit(MOV(next, shadow_c));
1239 bool has_nonconstant_offset = ir->offset && !ir->offset->as_constant();
1240 bool coordinate_done = false;
1242 /* Set up the LOD info */
1248 emit(MOV(next, lod));
1252 emit(MOV(next, lod));
1256 if (dispatch_width == 16)
1257 fail("Gen7 does not support sample_d/sample_d_c in SIMD16 mode.");
1259 /* Load dPdx and the coordinate together:
1260 * [hdr], [ref], x, dPdx.x, dPdy.x, y, dPdx.y, dPdy.y, z, dPdx.z, dPdy.z
1262 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
1263 emit(MOV(next, coordinate));
1264 coordinate.reg_offset++;
1267 /* For cube map array, the coordinate is (u,v,r,ai) but there are
1268 * only derivatives for (u, v, r).
1270 if (i < ir->lod_info.grad.dPdx->type->vector_elements) {
1271 emit(MOV(next, lod));
1275 emit(MOV(next, lod2));
1281 coordinate_done = true;
1285 emit(MOV(next.retype(BRW_REGISTER_TYPE_UD), lod));
1288 case ir_query_levels:
1289 emit(MOV(next.retype(BRW_REGISTER_TYPE_UD), fs_reg(0u)));
1293 /* Unfortunately, the parameters for LD are intermixed: u, lod, v, r. */
1294 emit(MOV(next.retype(BRW_REGISTER_TYPE_D), coordinate));
1295 coordinate.reg_offset++;
1298 emit(MOV(next.retype(BRW_REGISTER_TYPE_D), lod));
1301 for (int i = 1; i < ir->coordinate->type->vector_elements; i++) {
1302 emit(MOV(next.retype(BRW_REGISTER_TYPE_D), coordinate));
1303 coordinate.reg_offset++;
1307 coordinate_done = true;
1310 emit(MOV(next.retype(BRW_REGISTER_TYPE_UD), sample_index));
1313 /* constant zero MCS; we arrange to never actually have a compressed
1314 * multisample surface here for now. TODO: issue ld_mcs to get this first,
1315 * if we ever support texturing from compressed multisample surfaces
1317 emit(MOV(next.retype(BRW_REGISTER_TYPE_UD), fs_reg(0u)));
1320 /* there is no offsetting for this message; just copy in the integer
1321 * texture coordinates
1323 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
1324 emit(MOV(next.retype(BRW_REGISTER_TYPE_D), coordinate));
1325 coordinate.reg_offset++;
1329 coordinate_done = true;
1332 if (has_nonconstant_offset) {
1333 if (ir->shadow_comparitor && dispatch_width == 16)
1334 fail("Gen7 does not support gather4_po_c in SIMD16 mode.");
1336 /* More crazy intermixing */
1337 ir->offset->accept(this);
1338 fs_reg offset_value = this->result;
1340 for (int i = 0; i < 2; i++) { /* u, v */
1341 emit(MOV(next, coordinate));
1342 coordinate.reg_offset++;
1346 for (int i = 0; i < 2; i++) { /* offu, offv */
1347 emit(MOV(next.retype(BRW_REGISTER_TYPE_D), offset_value));
1348 offset_value.reg_offset++;
1352 if (ir->coordinate->type->vector_elements == 3) { /* r if present */
1353 emit(MOV(next, coordinate));
1354 coordinate.reg_offset++;
1358 coordinate_done = true;
1363 /* Set up the coordinate (except for cases where it was done above) */
1364 if (ir->coordinate && !coordinate_done) {
1365 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
1366 emit(MOV(next, coordinate));
1367 coordinate.reg_offset++;
1372 /* Generate the SEND */
1373 fs_inst *inst = NULL;
1375 case ir_tex: inst = emit(SHADER_OPCODE_TEX, dst, payload); break;
1376 case ir_txb: inst = emit(FS_OPCODE_TXB, dst, payload); break;
1377 case ir_txl: inst = emit(SHADER_OPCODE_TXL, dst, payload); break;
1378 case ir_txd: inst = emit(SHADER_OPCODE_TXD, dst, payload); break;
1379 case ir_txf: inst = emit(SHADER_OPCODE_TXF, dst, payload); break;
1380 case ir_txf_ms: inst = emit(SHADER_OPCODE_TXF_MS, dst, payload); break;
1381 case ir_txs: inst = emit(SHADER_OPCODE_TXS, dst, payload); break;
1382 case ir_query_levels: inst = emit(SHADER_OPCODE_TXS, dst, payload); break;
1383 case ir_lod: inst = emit(SHADER_OPCODE_LOD, dst, payload); break;
1385 if (has_nonconstant_offset)
1386 inst = emit(SHADER_OPCODE_TG4_OFFSET, dst, payload);
1388 inst = emit(SHADER_OPCODE_TG4, dst, payload);
1391 inst->base_mrf = -1;
1393 inst->mlen = next.reg_offset * reg_width - header_present;
1395 inst->mlen = next.reg_offset * reg_width;
1396 inst->header_present = header_present;
1397 inst->regs_written = 4;
1399 virtual_grf_sizes[payload.reg] = next.reg_offset;
1400 if (inst->mlen > 11) {
1401 fail("Message length >11 disallowed by hardware\n");
1408 fs_visitor::rescale_texcoord(ir_texture *ir, fs_reg coordinate,
1409 bool is_rect, int sampler, int texunit)
1411 fs_inst *inst = NULL;
1412 bool needs_gl_clamp = true;
1413 fs_reg scale_x, scale_y;
1415 /* The 965 requires the EU to do the normalization of GL rectangle
1416 * texture coordinates. We use the program parameter state
1417 * tracking to get the scaling factor.
1421 (brw->gen >= 6 && (c->key.tex.gl_clamp_mask[0] & (1 << sampler) ||
1422 c->key.tex.gl_clamp_mask[1] & (1 << sampler))))) {
1423 struct gl_program_parameter_list *params = prog->Parameters;
1424 int tokens[STATE_LENGTH] = {
1426 STATE_TEXRECT_SCALE,
1432 if (dispatch_width == 16) {
1433 fail("rectangle scale uniform setup not supported on 16-wide\n");
1437 scale_x = fs_reg(UNIFORM, c->prog_data.nr_params);
1438 scale_y = fs_reg(UNIFORM, c->prog_data.nr_params + 1);
1440 GLuint index = _mesa_add_state_reference(params,
1441 (gl_state_index *)tokens);
1442 c->prog_data.param[c->prog_data.nr_params++] =
1443 &prog->Parameters->ParameterValues[index][0].f;
1444 c->prog_data.param[c->prog_data.nr_params++] =
1445 &prog->Parameters->ParameterValues[index][1].f;
1448 /* The 965 requires the EU to do the normalization of GL rectangle
1449 * texture coordinates. We use the program parameter state
1450 * tracking to get the scaling factor.
1452 if (brw->gen < 6 && is_rect) {
1453 fs_reg dst = fs_reg(this, ir->coordinate->type);
1454 fs_reg src = coordinate;
1457 emit(MUL(dst, src, scale_x));
1460 emit(MUL(dst, src, scale_y));
1461 } else if (is_rect) {
1462 /* On gen6+, the sampler handles the rectangle coordinates
1463 * natively, without needing rescaling. But that means we have
1464 * to do GL_CLAMP clamping at the [0, width], [0, height] scale,
1465 * not [0, 1] like the default case below.
1467 needs_gl_clamp = false;
1469 for (int i = 0; i < 2; i++) {
1470 if (c->key.tex.gl_clamp_mask[i] & (1 << sampler)) {
1471 fs_reg chan = coordinate;
1472 chan.reg_offset += i;
1474 inst = emit(BRW_OPCODE_SEL, chan, chan, brw_imm_f(0.0));
1475 inst->conditional_mod = BRW_CONDITIONAL_G;
1477 /* Our parameter comes in as 1.0/width or 1.0/height,
1478 * because that's what people normally want for doing
1479 * texture rectangle handling. We need width or height
1480 * for clamping, but we don't care enough to make a new
1481 * parameter type, so just invert back.
1483 fs_reg limit = fs_reg(this, glsl_type::float_type);
1484 emit(MOV(limit, i == 0 ? scale_x : scale_y));
1485 emit(SHADER_OPCODE_RCP, limit, limit);
1487 inst = emit(BRW_OPCODE_SEL, chan, chan, limit);
1488 inst->conditional_mod = BRW_CONDITIONAL_L;
1493 if (ir->coordinate && needs_gl_clamp) {
1494 for (unsigned int i = 0;
1495 i < MIN2(ir->coordinate->type->vector_elements, 3); i++) {
1496 if (c->key.tex.gl_clamp_mask[i] & (1 << sampler)) {
1497 fs_reg chan = coordinate;
1498 chan.reg_offset += i;
1500 fs_inst *inst = emit(MOV(chan, chan));
1501 inst->saturate = true;
1509 fs_visitor::visit(ir_texture *ir)
1511 fs_inst *inst = NULL;
1514 _mesa_get_sampler_uniform_value(ir->sampler, shader_prog, prog);
1515 /* FINISHME: We're failing to recompile our programs when the sampler is
1516 * updated. This only matters for the texture rectangle scale parameters
1517 * (pre-gen6, or gen6+ with GL_CLAMP).
1519 int texunit = prog->SamplerUnits[sampler];
1521 if (ir->op == ir_tg4) {
1522 /* When tg4 is used with the degenerate ZERO/ONE swizzles, don't bother
1523 * emitting anything other than setting up the constant result.
1525 ir_constant *chan = ir->lod_info.component->as_constant();
1526 int swiz = GET_SWZ(c->key.tex.swizzles[sampler], chan->value.i[0]);
1527 if (swiz == SWIZZLE_ZERO || swiz == SWIZZLE_ONE) {
1529 fs_reg res = fs_reg(this, glsl_type::vec4_type);
1532 for (int i=0; i<4; i++) {
1533 emit(MOV(res, fs_reg(swiz == SWIZZLE_ZERO ? 0.0f : 1.0f)));
1540 /* Should be lowered by do_lower_texture_projection */
1541 assert(!ir->projector);
1543 /* Should be lowered */
1544 assert(!ir->offset || !ir->offset->type->is_array());
1546 /* Generate code to compute all the subexpression trees. This has to be
1547 * done before loading any values into MRFs for the sampler message since
1548 * generating these values may involve SEND messages that need the MRFs.
1551 if (ir->coordinate) {
1552 ir->coordinate->accept(this);
1554 coordinate = rescale_texcoord(ir, this->result,
1555 ir->sampler->type->sampler_dimensionality ==
1556 GLSL_SAMPLER_DIM_RECT,
1560 fs_reg shadow_comparitor;
1561 if (ir->shadow_comparitor) {
1562 ir->shadow_comparitor->accept(this);
1563 shadow_comparitor = this->result;
1566 fs_reg lod, lod2, sample_index;
1571 case ir_query_levels:
1574 ir->lod_info.bias->accept(this);
1578 ir->lod_info.grad.dPdx->accept(this);
1581 ir->lod_info.grad.dPdy->accept(this);
1582 lod2 = this->result;
1587 ir->lod_info.lod->accept(this);
1591 ir->lod_info.sample_index->accept(this);
1592 sample_index = this->result;
1595 assert(!"Unrecognized texture opcode");
1598 /* Writemasking doesn't eliminate channels on SIMD8 texture
1599 * samples, so don't worry about them.
1601 fs_reg dst = fs_reg(this, glsl_type::get_instance(ir->type->base_type, 4, 1));
1603 if (brw->gen >= 7) {
1604 inst = emit_texture_gen7(ir, dst, coordinate, shadow_comparitor,
1605 lod, lod2, sample_index);
1606 } else if (brw->gen >= 5) {
1607 inst = emit_texture_gen5(ir, dst, coordinate, shadow_comparitor,
1608 lod, lod2, sample_index);
1610 inst = emit_texture_gen4(ir, dst, coordinate, shadow_comparitor,
1614 if (ir->offset != NULL && ir->op != ir_txf)
1615 inst->texture_offset = brw_texture_offset(ctx, ir->offset->as_constant());
1617 if (ir->op == ir_tg4)
1618 inst->texture_offset |= gather_channel(ir, sampler) << 16; // M0.2:16-17
1620 inst->sampler = sampler;
1622 if (ir->shadow_comparitor)
1623 inst->shadow_compare = true;
1625 /* fixup #layers for cube map arrays */
1626 if (ir->op == ir_txs) {
1627 glsl_type const *type = ir->sampler->type;
1628 if (type->sampler_dimensionality == GLSL_SAMPLER_DIM_CUBE &&
1629 type->sampler_array) {
1631 depth.reg_offset = 2;
1632 emit_math(SHADER_OPCODE_INT_QUOTIENT, depth, depth, fs_reg(6));
1636 swizzle_result(ir, dst, sampler);
1640 * Set up the gather channel based on the swizzle, for gather4.
1643 fs_visitor::gather_channel(ir_texture *ir, int sampler)
1645 ir_constant *chan = ir->lod_info.component->as_constant();
1646 int swiz = GET_SWZ(c->key.tex.swizzles[sampler], chan->value.i[0]);
1648 case SWIZZLE_X: return 0;
1650 /* gather4 sampler is broken for green channel on RG32F --
1651 * we must ask for blue instead.
1653 if (c->key.tex.gather_channel_quirk_mask & (1<<sampler))
1656 case SWIZZLE_Z: return 2;
1657 case SWIZZLE_W: return 3;
1659 assert(!"Not reached"); /* zero, one swizzles handled already */
1665 * Swizzle the result of a texture result. This is necessary for
1666 * EXT_texture_swizzle as well as DEPTH_TEXTURE_MODE for shadow comparisons.
1669 fs_visitor::swizzle_result(ir_texture *ir, fs_reg orig_val, int sampler)
1671 if (ir->op == ir_query_levels) {
1672 /* # levels is in .w */
1673 orig_val.reg_offset += 3;
1674 this->result = orig_val;
1678 this->result = orig_val;
1680 /* txs,lod don't actually sample the texture, so swizzling the result
1683 if (ir->op == ir_txs || ir->op == ir_lod || ir->op == ir_tg4)
1686 if (ir->type == glsl_type::float_type) {
1687 /* Ignore DEPTH_TEXTURE_MODE swizzling. */
1688 assert(ir->sampler->type->sampler_shadow);
1689 } else if (c->key.tex.swizzles[sampler] != SWIZZLE_NOOP) {
1690 fs_reg swizzled_result = fs_reg(this, glsl_type::vec4_type);
1692 for (int i = 0; i < 4; i++) {
1693 int swiz = GET_SWZ(c->key.tex.swizzles[sampler], i);
1694 fs_reg l = swizzled_result;
1697 if (swiz == SWIZZLE_ZERO) {
1698 emit(MOV(l, fs_reg(0.0f)));
1699 } else if (swiz == SWIZZLE_ONE) {
1700 emit(MOV(l, fs_reg(1.0f)));
1702 fs_reg r = orig_val;
1703 r.reg_offset += GET_SWZ(c->key.tex.swizzles[sampler], i);
1707 this->result = swizzled_result;
1712 fs_visitor::visit(ir_swizzle *ir)
1714 ir->val->accept(this);
1715 fs_reg val = this->result;
1717 if (ir->type->vector_elements == 1) {
1718 this->result.reg_offset += ir->mask.x;
1722 fs_reg result = fs_reg(this, ir->type);
1723 this->result = result;
1725 for (unsigned int i = 0; i < ir->type->vector_elements; i++) {
1726 fs_reg channel = val;
1744 channel.reg_offset += swiz;
1745 emit(MOV(result, channel));
1746 result.reg_offset++;
1751 fs_visitor::visit(ir_discard *ir)
1753 assert(ir->condition == NULL); /* FINISHME */
1755 /* We track our discarded pixels in f0.1. By predicating on it, we can
1756 * update just the flag bits that aren't yet discarded. By emitting a
1757 * CMP of g0 != g0, all our currently executing channels will get turned
1760 fs_reg some_reg = fs_reg(retype(brw_vec8_grf(0, 0),
1761 BRW_REGISTER_TYPE_UW));
1762 fs_inst *cmp = emit(CMP(reg_null_f, some_reg, some_reg,
1763 BRW_CONDITIONAL_NZ));
1764 cmp->predicate = BRW_PREDICATE_NORMAL;
1765 cmp->flag_subreg = 1;
1767 if (brw->gen >= 6) {
1768 /* For performance, after a discard, jump to the end of the shader.
1769 * However, many people will do foliage by discarding based on a
1770 * texture's alpha mask, and then continue on to texture with the
1771 * remaining pixels. To avoid trashing the derivatives for those
1772 * texture samples, we'll only jump if all of the pixels in the subspan
1773 * have been discarded.
1775 fs_inst *discard_jump = emit(FS_OPCODE_DISCARD_JUMP);
1776 discard_jump->flag_subreg = 1;
1777 discard_jump->predicate = BRW_PREDICATE_ALIGN1_ANY4H;
1778 discard_jump->predicate_inverse = true;
1783 fs_visitor::visit(ir_constant *ir)
1785 /* Set this->result to reg at the bottom of the function because some code
1786 * paths will cause this visitor to be applied to other fields. This will
1787 * cause the value stored in this->result to be modified.
1789 * Make reg constant so that it doesn't get accidentally modified along the
1790 * way. Yes, I actually had this problem. :(
1792 const fs_reg reg(this, ir->type);
1793 fs_reg dst_reg = reg;
1795 if (ir->type->is_array()) {
1796 const unsigned size = type_size(ir->type->fields.array);
1798 for (unsigned i = 0; i < ir->type->length; i++) {
1799 ir->array_elements[i]->accept(this);
1800 fs_reg src_reg = this->result;
1802 dst_reg.type = src_reg.type;
1803 for (unsigned j = 0; j < size; j++) {
1804 emit(MOV(dst_reg, src_reg));
1805 src_reg.reg_offset++;
1806 dst_reg.reg_offset++;
1809 } else if (ir->type->is_record()) {
1810 foreach_list(node, &ir->components) {
1811 ir_constant *const field = (ir_constant *) node;
1812 const unsigned size = type_size(field->type);
1814 field->accept(this);
1815 fs_reg src_reg = this->result;
1817 dst_reg.type = src_reg.type;
1818 for (unsigned j = 0; j < size; j++) {
1819 emit(MOV(dst_reg, src_reg));
1820 src_reg.reg_offset++;
1821 dst_reg.reg_offset++;
1825 const unsigned size = type_size(ir->type);
1827 for (unsigned i = 0; i < size; i++) {
1828 switch (ir->type->base_type) {
1829 case GLSL_TYPE_FLOAT:
1830 emit(MOV(dst_reg, fs_reg(ir->value.f[i])));
1832 case GLSL_TYPE_UINT:
1833 emit(MOV(dst_reg, fs_reg(ir->value.u[i])));
1836 emit(MOV(dst_reg, fs_reg(ir->value.i[i])));
1838 case GLSL_TYPE_BOOL:
1839 emit(MOV(dst_reg, fs_reg((int)ir->value.b[i])));
1842 assert(!"Non-float/uint/int/bool constant");
1844 dst_reg.reg_offset++;
1852 fs_visitor::emit_bool_to_cond_code(ir_rvalue *ir)
1854 ir_expression *expr = ir->as_expression();
1857 expr->operation != ir_binop_logic_and &&
1858 expr->operation != ir_binop_logic_or &&
1859 expr->operation != ir_binop_logic_xor) {
1863 assert(expr->get_num_operands() <= 2);
1864 for (unsigned int i = 0; i < expr->get_num_operands(); i++) {
1865 assert(expr->operands[i]->type->is_scalar());
1867 expr->operands[i]->accept(this);
1868 op[i] = this->result;
1870 resolve_ud_negate(&op[i]);
1873 switch (expr->operation) {
1874 case ir_unop_logic_not:
1875 inst = emit(AND(reg_null_d, op[0], fs_reg(1)));
1876 inst->conditional_mod = BRW_CONDITIONAL_Z;
1880 if (brw->gen >= 6) {
1881 emit(CMP(reg_null_d, op[0], fs_reg(0.0f), BRW_CONDITIONAL_NZ));
1883 inst = emit(MOV(reg_null_f, op[0]));
1884 inst->conditional_mod = BRW_CONDITIONAL_NZ;
1889 if (brw->gen >= 6) {
1890 emit(CMP(reg_null_d, op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
1892 inst = emit(MOV(reg_null_d, op[0]));
1893 inst->conditional_mod = BRW_CONDITIONAL_NZ;
1897 case ir_binop_greater:
1898 case ir_binop_gequal:
1900 case ir_binop_lequal:
1901 case ir_binop_equal:
1902 case ir_binop_all_equal:
1903 case ir_binop_nequal:
1904 case ir_binop_any_nequal:
1905 resolve_bool_comparison(expr->operands[0], &op[0]);
1906 resolve_bool_comparison(expr->operands[1], &op[1]);
1908 emit(CMP(reg_null_d, op[0], op[1],
1909 brw_conditional_for_comparison(expr->operation)));
1913 assert(!"not reached");
1914 fail("bad cond code\n");
1922 fs_inst *inst = emit(AND(reg_null_d, this->result, fs_reg(1)));
1923 inst->conditional_mod = BRW_CONDITIONAL_NZ;
1927 * Emit a gen6 IF statement with the comparison folded into the IF
1931 fs_visitor::emit_if_gen6(ir_if *ir)
1933 ir_expression *expr = ir->condition->as_expression();
1940 assert(expr->get_num_operands() <= 2);
1941 for (unsigned int i = 0; i < expr->get_num_operands(); i++) {
1942 assert(expr->operands[i]->type->is_scalar());
1944 expr->operands[i]->accept(this);
1945 op[i] = this->result;
1948 switch (expr->operation) {
1949 case ir_unop_logic_not:
1950 case ir_binop_logic_xor:
1951 case ir_binop_logic_or:
1952 case ir_binop_logic_and:
1953 /* For operations on bool arguments, only the low bit of the bool is
1954 * valid, and the others are undefined. Fall back to the condition
1960 inst = emit(BRW_OPCODE_IF, reg_null_f, op[0], fs_reg(0));
1961 inst->conditional_mod = BRW_CONDITIONAL_NZ;
1965 emit(IF(op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
1968 case ir_binop_greater:
1969 case ir_binop_gequal:
1971 case ir_binop_lequal:
1972 case ir_binop_equal:
1973 case ir_binop_all_equal:
1974 case ir_binop_nequal:
1975 case ir_binop_any_nequal:
1976 resolve_bool_comparison(expr->operands[0], &op[0]);
1977 resolve_bool_comparison(expr->operands[1], &op[1]);
1979 emit(IF(op[0], op[1],
1980 brw_conditional_for_comparison(expr->operation)));
1983 assert(!"not reached");
1984 emit(IF(op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
1985 fail("bad condition\n");
1990 emit_bool_to_cond_code(ir->condition);
1991 fs_inst *inst = emit(BRW_OPCODE_IF);
1992 inst->predicate = BRW_PREDICATE_NORMAL;
1996 * Try to replace IF/MOV/ELSE/MOV/ENDIF with SEL.
1998 * Many GLSL shaders contain the following pattern:
2000 * x = condition ? foo : bar
2002 * The compiler emits an ir_if tree for this, since each subexpression might be
2003 * a complex tree that could have side-effects or short-circuit logic.
2005 * However, the common case is to simply select one of two constants or
2006 * variable values---which is exactly what SEL is for. In this case, the
2007 * assembly looks like:
2015 * which can be easily translated into:
2017 * (+f0) SEL dst src0 src1
2019 * If src0 is an immediate value, we promote it to a temporary GRF.
2022 fs_visitor::try_replace_with_sel()
2024 fs_inst *endif_inst = (fs_inst *) instructions.get_tail();
2025 assert(endif_inst->opcode == BRW_OPCODE_ENDIF);
2027 /* Pattern match in reverse: IF, MOV, ELSE, MOV, ENDIF. */
2029 BRW_OPCODE_IF, BRW_OPCODE_MOV, BRW_OPCODE_ELSE, BRW_OPCODE_MOV,
2032 fs_inst *match = (fs_inst *) endif_inst->prev;
2033 for (int i = 0; i < 4; i++) {
2034 if (match->is_head_sentinel() || match->opcode != opcodes[4-i-1])
2036 match = (fs_inst *) match->prev;
2039 /* The opcodes match; it looks like the right sequence of instructions. */
2040 fs_inst *else_mov = (fs_inst *) endif_inst->prev;
2041 fs_inst *then_mov = (fs_inst *) else_mov->prev->prev;
2042 fs_inst *if_inst = (fs_inst *) then_mov->prev;
2044 /* Check that the MOVs are the right form. */
2045 if (then_mov->dst.equals(else_mov->dst) &&
2046 !then_mov->is_partial_write() &&
2047 !else_mov->is_partial_write()) {
2049 /* Remove the matched instructions; we'll emit a SEL to replace them. */
2050 while (!if_inst->next->is_tail_sentinel())
2051 if_inst->next->remove();
2054 /* Only the last source register can be a constant, so if the MOV in
2055 * the "then" clause uses a constant, we need to put it in a temporary.
2057 fs_reg src0(then_mov->src[0]);
2058 if (src0.file == IMM) {
2059 src0 = fs_reg(this, glsl_type::float_type);
2060 src0.type = then_mov->src[0].type;
2061 emit(MOV(src0, then_mov->src[0]));
2065 if (if_inst->conditional_mod) {
2066 /* Sandybridge-specific IF with embedded comparison */
2067 emit(CMP(reg_null_d, if_inst->src[0], if_inst->src[1],
2068 if_inst->conditional_mod));
2069 sel = emit(BRW_OPCODE_SEL, then_mov->dst, src0, else_mov->src[0]);
2070 sel->predicate = BRW_PREDICATE_NORMAL;
2072 /* Separate CMP and IF instructions */
2073 sel = emit(BRW_OPCODE_SEL, then_mov->dst, src0, else_mov->src[0]);
2074 sel->predicate = if_inst->predicate;
2075 sel->predicate_inverse = if_inst->predicate_inverse;
2081 fs_visitor::visit(ir_if *ir)
2083 if (brw->gen < 6 && dispatch_width == 16) {
2084 fail("Can't support (non-uniform) control flow on 16-wide\n");
2087 /* Don't point the annotation at the if statement, because then it plus
2088 * the then and else blocks get printed.
2090 this->base_ir = ir->condition;
2092 if (brw->gen == 6) {
2095 emit_bool_to_cond_code(ir->condition);
2097 emit(IF(BRW_PREDICATE_NORMAL));
2100 foreach_list(node, &ir->then_instructions) {
2101 ir_instruction *ir = (ir_instruction *)node;
2107 if (!ir->else_instructions.is_empty()) {
2108 emit(BRW_OPCODE_ELSE);
2110 foreach_list(node, &ir->else_instructions) {
2111 ir_instruction *ir = (ir_instruction *)node;
2118 emit(BRW_OPCODE_ENDIF);
2120 try_replace_with_sel();
2124 fs_visitor::visit(ir_loop *ir)
2126 fs_reg counter = reg_undef;
2128 if (brw->gen < 6 && dispatch_width == 16) {
2129 fail("Can't support (non-uniform) control flow on 16-wide\n");
2133 this->base_ir = ir->counter;
2134 ir->counter->accept(this);
2135 counter = *(variable_storage(ir->counter));
2138 this->base_ir = ir->from;
2139 ir->from->accept(this);
2141 emit(MOV(counter, this->result));
2145 this->base_ir = NULL;
2146 emit(BRW_OPCODE_DO);
2149 this->base_ir = ir->to;
2150 ir->to->accept(this);
2152 emit(CMP(reg_null_d, counter, this->result,
2153 brw_conditional_for_comparison(ir->cmp)));
2155 fs_inst *inst = emit(BRW_OPCODE_BREAK);
2156 inst->predicate = BRW_PREDICATE_NORMAL;
2159 foreach_list(node, &ir->body_instructions) {
2160 ir_instruction *ir = (ir_instruction *)node;
2166 if (ir->increment) {
2167 this->base_ir = ir->increment;
2168 ir->increment->accept(this);
2169 emit(ADD(counter, counter, this->result));
2172 this->base_ir = NULL;
2173 emit(BRW_OPCODE_WHILE);
2177 fs_visitor::visit(ir_loop_jump *ir)
2180 case ir_loop_jump::jump_break:
2181 emit(BRW_OPCODE_BREAK);
2183 case ir_loop_jump::jump_continue:
2184 emit(BRW_OPCODE_CONTINUE);
2190 fs_visitor::visit(ir_call *ir)
2192 assert(!"FINISHME");
2196 fs_visitor::visit(ir_return *ir)
2198 assert(!"FINISHME");
2202 fs_visitor::visit(ir_function *ir)
2204 /* Ignore function bodies other than main() -- we shouldn't see calls to
2205 * them since they should all be inlined before we get to ir_to_mesa.
2207 if (strcmp(ir->name, "main") == 0) {
2208 const ir_function_signature *sig;
2211 sig = ir->matching_signature(NULL, &empty);
2215 foreach_list(node, &sig->body) {
2216 ir_instruction *ir = (ir_instruction *)node;
2225 fs_visitor::visit(ir_function_signature *ir)
2227 assert(!"not reached");
2232 fs_visitor::visit(ir_emit_vertex *)
2234 assert(!"not reached");
2238 fs_visitor::visit(ir_end_primitive *)
2240 assert(!"not reached");
2244 fs_visitor::emit(fs_inst inst)
2246 fs_inst *list_inst = new(mem_ctx) fs_inst;
2253 fs_visitor::emit(fs_inst *inst)
2255 if (force_uncompressed_stack > 0)
2256 inst->force_uncompressed = true;
2257 else if (force_sechalf_stack > 0)
2258 inst->force_sechalf = true;
2260 inst->annotation = this->current_annotation;
2261 inst->ir = this->base_ir;
2263 this->instructions.push_tail(inst);
2269 fs_visitor::emit(exec_list list)
2271 foreach_list_safe(node, &list) {
2272 fs_inst *inst = (fs_inst *)node;
2278 /** Emits a dummy fragment shader consisting of magenta for bringup purposes. */
2280 fs_visitor::emit_dummy_fs()
2282 int reg_width = dispatch_width / 8;
2284 /* Everyone's favorite color. */
2285 emit(MOV(fs_reg(MRF, 2 + 0 * reg_width), fs_reg(1.0f)));
2286 emit(MOV(fs_reg(MRF, 2 + 1 * reg_width), fs_reg(0.0f)));
2287 emit(MOV(fs_reg(MRF, 2 + 2 * reg_width), fs_reg(1.0f)));
2288 emit(MOV(fs_reg(MRF, 2 + 3 * reg_width), fs_reg(0.0f)));
2291 write = emit(FS_OPCODE_FB_WRITE, fs_reg(0), fs_reg(0));
2292 write->base_mrf = 2;
2293 write->mlen = 4 * reg_width;
2297 /* The register location here is relative to the start of the URB
2298 * data. It will get adjusted to be a real location before
2299 * generate_code() time.
2302 fs_visitor::interp_reg(int location, int channel)
2304 int regnr = c->prog_data.urb_setup[location] * 2 + channel / 2;
2305 int stride = (channel & 1) * 4;
2307 assert(c->prog_data.urb_setup[location] != -1);
2309 return brw_vec1_grf(regnr, stride);
2312 /** Emits the interpolation for the varying inputs. */
2314 fs_visitor::emit_interpolation_setup_gen4()
2316 this->current_annotation = "compute pixel centers";
2317 this->pixel_x = fs_reg(this, glsl_type::uint_type);
2318 this->pixel_y = fs_reg(this, glsl_type::uint_type);
2319 this->pixel_x.type = BRW_REGISTER_TYPE_UW;
2320 this->pixel_y.type = BRW_REGISTER_TYPE_UW;
2322 emit(FS_OPCODE_PIXEL_X, this->pixel_x);
2323 emit(FS_OPCODE_PIXEL_Y, this->pixel_y);
2325 this->current_annotation = "compute pixel deltas from v0";
2327 this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC] =
2328 fs_reg(this, glsl_type::vec2_type);
2329 this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC] =
2330 this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC];
2331 this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC].reg_offset++;
2333 this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC] =
2334 fs_reg(this, glsl_type::float_type);
2335 this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC] =
2336 fs_reg(this, glsl_type::float_type);
2338 emit(ADD(this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
2339 this->pixel_x, fs_reg(negate(brw_vec1_grf(1, 0)))));
2340 emit(ADD(this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
2341 this->pixel_y, fs_reg(negate(brw_vec1_grf(1, 1)))));
2343 this->current_annotation = "compute pos.w and 1/pos.w";
2344 /* Compute wpos.w. It's always in our setup, since it's needed to
2345 * interpolate the other attributes.
2347 this->wpos_w = fs_reg(this, glsl_type::float_type);
2348 emit(FS_OPCODE_LINTERP, wpos_w,
2349 this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
2350 this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
2351 interp_reg(VARYING_SLOT_POS, 3));
2352 /* Compute the pixel 1/W value from wpos.w. */
2353 this->pixel_w = fs_reg(this, glsl_type::float_type);
2354 emit_math(SHADER_OPCODE_RCP, this->pixel_w, wpos_w);
2355 this->current_annotation = NULL;
2358 /** Emits the interpolation for the varying inputs. */
2360 fs_visitor::emit_interpolation_setup_gen6()
2362 struct brw_reg g1_uw = retype(brw_vec1_grf(1, 0), BRW_REGISTER_TYPE_UW);
2364 /* If the pixel centers end up used, the setup is the same as for gen4. */
2365 this->current_annotation = "compute pixel centers";
2366 fs_reg int_pixel_x = fs_reg(this, glsl_type::uint_type);
2367 fs_reg int_pixel_y = fs_reg(this, glsl_type::uint_type);
2368 int_pixel_x.type = BRW_REGISTER_TYPE_UW;
2369 int_pixel_y.type = BRW_REGISTER_TYPE_UW;
2370 emit(ADD(int_pixel_x,
2371 fs_reg(stride(suboffset(g1_uw, 4), 2, 4, 0)),
2372 fs_reg(brw_imm_v(0x10101010))));
2373 emit(ADD(int_pixel_y,
2374 fs_reg(stride(suboffset(g1_uw, 5), 2, 4, 0)),
2375 fs_reg(brw_imm_v(0x11001100))));
2377 /* As of gen6, we can no longer mix float and int sources. We have
2378 * to turn the integer pixel centers into floats for their actual
2381 this->pixel_x = fs_reg(this, glsl_type::float_type);
2382 this->pixel_y = fs_reg(this, glsl_type::float_type);
2383 emit(MOV(this->pixel_x, int_pixel_x));
2384 emit(MOV(this->pixel_y, int_pixel_y));
2386 this->current_annotation = "compute pos.w";
2387 this->pixel_w = fs_reg(brw_vec8_grf(c->source_w_reg, 0));
2388 this->wpos_w = fs_reg(this, glsl_type::float_type);
2389 emit_math(SHADER_OPCODE_RCP, this->wpos_w, this->pixel_w);
2391 for (int i = 0; i < BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT; ++i) {
2392 uint8_t reg = c->barycentric_coord_reg[i];
2393 this->delta_x[i] = fs_reg(brw_vec8_grf(reg, 0));
2394 this->delta_y[i] = fs_reg(brw_vec8_grf(reg + 1, 0));
2397 this->current_annotation = NULL;
2401 fs_visitor::emit_color_write(int target, int index, int first_color_mrf)
2403 int reg_width = dispatch_width / 8;
2405 fs_reg color = outputs[target];
2408 /* If there's no color data to be written, skip it. */
2409 if (color.file == BAD_FILE)
2412 color.reg_offset += index;
2414 if (dispatch_width == 8 || brw->gen >= 6) {
2415 /* SIMD8 write looks like:
2421 * gen6 SIMD16 DP write looks like:
2431 inst = emit(MOV(fs_reg(MRF, first_color_mrf + index * reg_width,
2434 inst->saturate = c->key.clamp_fragment_color;
2436 /* pre-gen6 SIMD16 single source DP write looks like:
2446 if (brw->has_compr4) {
2447 /* By setting the high bit of the MRF register number, we
2448 * indicate that we want COMPR4 mode - instead of doing the
2449 * usual destination + 1 for the second half we get
2452 inst = emit(MOV(fs_reg(MRF, BRW_MRF_COMPR4 + first_color_mrf + index,
2455 inst->saturate = c->key.clamp_fragment_color;
2457 push_force_uncompressed();
2458 inst = emit(MOV(fs_reg(MRF, first_color_mrf + index, color.type),
2460 inst->saturate = c->key.clamp_fragment_color;
2461 pop_force_uncompressed();
2463 push_force_sechalf();
2464 color.sechalf = true;
2465 inst = emit(MOV(fs_reg(MRF, first_color_mrf + index + 4, color.type),
2467 inst->saturate = c->key.clamp_fragment_color;
2468 pop_force_sechalf();
2469 color.sechalf = false;
2475 fs_visitor::emit_fb_writes()
2477 this->current_annotation = "FB write header";
2478 bool header_present = true;
2479 /* We can potentially have a message length of up to 15, so we have to set
2480 * base_mrf to either 0 or 1 in order to fit in m0..m15.
2484 int reg_width = dispatch_width / 8;
2485 bool do_dual_src = this->dual_src_output.file != BAD_FILE;
2486 bool src0_alpha_to_render_target = false;
2488 if (dispatch_width == 16 && do_dual_src) {
2489 fail("GL_ARB_blend_func_extended not yet supported in 16-wide.");
2490 do_dual_src = false;
2493 /* From the Sandy Bridge PRM, volume 4, page 198:
2495 * "Dispatched Pixel Enables. One bit per pixel indicating
2496 * which pixels were originally enabled when the thread was
2497 * dispatched. This field is only required for the end-of-
2498 * thread message and on all dual-source messages."
2500 if (brw->gen >= 6 &&
2501 !this->fp->UsesKill &&
2503 c->key.nr_color_regions == 1) {
2504 header_present = false;
2507 if (header_present) {
2508 src0_alpha_to_render_target = brw->gen >= 6 &&
2510 c->key.replicate_alpha;
2515 if (c->aa_dest_stencil_reg) {
2516 push_force_uncompressed();
2517 emit(MOV(fs_reg(MRF, nr++),
2518 fs_reg(brw_vec8_grf(c->aa_dest_stencil_reg, 0))));
2519 pop_force_uncompressed();
2522 /* Reserve space for color. It'll be filled in per MRT below. */
2524 nr += 4 * reg_width;
2527 if (src0_alpha_to_render_target)
2530 if (c->source_depth_to_render_target) {
2531 if (brw->gen == 6 && dispatch_width == 16) {
2532 /* For outputting oDepth on gen6, SIMD8 writes have to be
2533 * used. This would require 8-wide moves of each half to
2534 * message regs, kind of like pre-gen5 SIMD16 FB writes.
2535 * Just bail on doing so for now.
2537 fail("Missing support for simd16 depth writes on gen6\n");
2540 if (prog->OutputsWritten & BITFIELD64_BIT(FRAG_RESULT_DEPTH)) {
2541 /* Hand over gl_FragDepth. */
2542 assert(this->frag_depth.file != BAD_FILE);
2543 emit(MOV(fs_reg(MRF, nr), this->frag_depth));
2545 /* Pass through the payload depth. */
2546 emit(MOV(fs_reg(MRF, nr),
2547 fs_reg(brw_vec8_grf(c->source_depth_reg, 0))));
2552 if (c->dest_depth_reg) {
2553 emit(MOV(fs_reg(MRF, nr),
2554 fs_reg(brw_vec8_grf(c->dest_depth_reg, 0))));
2559 fs_reg src0 = this->outputs[0];
2560 fs_reg src1 = this->dual_src_output;
2562 this->current_annotation = ralloc_asprintf(this->mem_ctx,
2564 for (int i = 0; i < 4; i++) {
2565 fs_inst *inst = emit(MOV(fs_reg(MRF, color_mrf + i, src0.type), src0));
2567 inst->saturate = c->key.clamp_fragment_color;
2570 this->current_annotation = ralloc_asprintf(this->mem_ctx,
2572 for (int i = 0; i < 4; i++) {
2573 fs_inst *inst = emit(MOV(fs_reg(MRF, color_mrf + 4 + i, src1.type),
2576 inst->saturate = c->key.clamp_fragment_color;
2579 if (INTEL_DEBUG & DEBUG_SHADER_TIME)
2580 emit_shader_time_end();
2582 fs_inst *inst = emit(FS_OPCODE_FB_WRITE);
2584 inst->base_mrf = base_mrf;
2585 inst->mlen = nr - base_mrf;
2587 inst->header_present = header_present;
2589 c->prog_data.dual_src_blend = true;
2590 this->current_annotation = NULL;
2594 for (int target = 0; target < c->key.nr_color_regions; target++) {
2595 this->current_annotation = ralloc_asprintf(this->mem_ctx,
2596 "FB write target %d",
2598 /* If src0_alpha_to_render_target is true, include source zero alpha
2599 * data in RenderTargetWrite message for targets > 0.
2601 int write_color_mrf = color_mrf;
2602 if (src0_alpha_to_render_target && target != 0) {
2604 fs_reg color = outputs[0];
2605 color.reg_offset += 3;
2607 inst = emit(MOV(fs_reg(MRF, write_color_mrf, color.type),
2609 inst->saturate = c->key.clamp_fragment_color;
2610 write_color_mrf = color_mrf + reg_width;
2613 for (unsigned i = 0; i < this->output_components[target]; i++)
2614 emit_color_write(target, i, write_color_mrf);
2617 if (target == c->key.nr_color_regions - 1) {
2620 if (INTEL_DEBUG & DEBUG_SHADER_TIME)
2621 emit_shader_time_end();
2624 fs_inst *inst = emit(FS_OPCODE_FB_WRITE);
2625 inst->target = target;
2626 inst->base_mrf = base_mrf;
2627 if (src0_alpha_to_render_target && target == 0)
2628 inst->mlen = nr - base_mrf - reg_width;
2630 inst->mlen = nr - base_mrf;
2632 inst->header_present = header_present;
2635 if (c->key.nr_color_regions == 0) {
2636 /* Even if there's no color buffers enabled, we still need to send
2637 * alpha out the pipeline to our null renderbuffer to support
2638 * alpha-testing, alpha-to-coverage, and so on.
2640 emit_color_write(0, 3, color_mrf);
2642 if (INTEL_DEBUG & DEBUG_SHADER_TIME)
2643 emit_shader_time_end();
2645 fs_inst *inst = emit(FS_OPCODE_FB_WRITE);
2646 inst->base_mrf = base_mrf;
2647 inst->mlen = nr - base_mrf;
2649 inst->header_present = header_present;
2652 this->current_annotation = NULL;
2656 fs_visitor::resolve_ud_negate(fs_reg *reg)
2658 if (reg->type != BRW_REGISTER_TYPE_UD ||
2662 fs_reg temp = fs_reg(this, glsl_type::uint_type);
2663 emit(MOV(temp, *reg));
2668 fs_visitor::resolve_bool_comparison(ir_rvalue *rvalue, fs_reg *reg)
2670 if (rvalue->type != glsl_type::bool_type)
2673 fs_reg temp = fs_reg(this, glsl_type::bool_type);
2674 emit(AND(temp, *reg, fs_reg(1)));
2678 fs_visitor::fs_visitor(struct brw_context *brw,
2679 struct brw_wm_compile *c,
2680 struct gl_shader_program *shader_prog,
2681 struct gl_fragment_program *fp,
2682 unsigned dispatch_width)
2683 : dispatch_width(dispatch_width)
2688 this->prog = &fp->Base;
2689 this->shader_prog = shader_prog;
2690 this->prog = &fp->Base;
2691 this->stage_prog_data = &c->prog_data.base;
2692 this->ctx = &brw->ctx;
2693 this->mem_ctx = ralloc_context(NULL);
2695 shader = (struct brw_shader *)
2696 shader_prog->_LinkedShaders[MESA_SHADER_FRAGMENT];
2699 this->failed = false;
2700 this->variable_ht = hash_table_ctor(0,
2701 hash_table_pointer_hash,
2702 hash_table_pointer_compare);
2704 memset(this->outputs, 0, sizeof(this->outputs));
2705 memset(this->output_components, 0, sizeof(this->output_components));
2706 this->first_non_payload_grf = 0;
2707 this->max_grf = brw->gen >= 7 ? GEN7_MRF_HACK_START : BRW_MAX_GRF;
2709 this->current_annotation = NULL;
2710 this->base_ir = NULL;
2712 this->virtual_grf_sizes = NULL;
2713 this->virtual_grf_count = 0;
2714 this->virtual_grf_array_size = 0;
2715 this->virtual_grf_start = NULL;
2716 this->virtual_grf_end = NULL;
2717 this->live_intervals = NULL;
2719 this->params_remap = NULL;
2720 this->nr_params_remap = 0;
2722 this->force_uncompressed_stack = 0;
2723 this->force_sechalf_stack = 0;
2725 this->spilled_any_registers = false;
2727 memset(&this->param_size, 0, sizeof(this->param_size));
2730 fs_visitor::~fs_visitor()
2732 ralloc_free(this->mem_ctx);
2733 hash_table_dtor(this->variable_ht);