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;
85 } else if (ir->location == FRAG_RESULT_SAMPLE_MASK) {
86 this->sample_mask = *reg;
88 /* gl_FragData or a user-defined FS output */
89 assert(ir->location >= FRAG_RESULT_DATA0 &&
90 ir->location < FRAG_RESULT_DATA0 + BRW_MAX_DRAW_BUFFERS);
93 ir->type->is_array() ? ir->type->fields.array->vector_elements
94 : ir->type->vector_elements;
96 /* General color output. */
97 for (unsigned int i = 0; i < MAX2(1, ir->type->length); i++) {
98 int output = ir->location - FRAG_RESULT_DATA0 + i;
99 this->outputs[output] = *reg;
100 this->outputs[output].reg_offset += vector_elements * i;
101 this->output_components[output] = vector_elements;
104 } else if (ir->mode == ir_var_uniform) {
105 int param_index = c->prog_data.nr_params;
107 /* Thanks to the lower_ubo_reference pass, we will see only
108 * ir_binop_ubo_load expressions and not ir_dereference_variable for UBO
109 * variables, so no need for them to be in variable_ht.
111 if (ir->is_in_uniform_block())
114 if (dispatch_width == 16) {
115 if (!variable_storage(ir)) {
116 fail("Failed to find uniform '%s' in 16-wide\n", ir->name);
121 param_size[param_index] = type_size(ir->type);
122 if (!strncmp(ir->name, "gl_", 3)) {
123 setup_builtin_uniform_values(ir);
125 setup_uniform_values(ir);
128 reg = new(this->mem_ctx) fs_reg(UNIFORM, param_index);
129 reg->type = brw_type_for_base_type(ir->type);
131 } else if (ir->mode == ir_var_system_value) {
132 if (ir->location == SYSTEM_VALUE_SAMPLE_POS) {
133 reg = emit_samplepos_setup(ir);
134 } else if (ir->location == SYSTEM_VALUE_SAMPLE_ID) {
135 reg = emit_sampleid_setup(ir);
140 reg = new(this->mem_ctx) fs_reg(this, ir->type);
142 hash_table_insert(this->variable_ht, reg, ir);
146 fs_visitor::visit(ir_dereference_variable *ir)
148 fs_reg *reg = variable_storage(ir->var);
153 fs_visitor::visit(ir_dereference_record *ir)
155 const glsl_type *struct_type = ir->record->type;
157 ir->record->accept(this);
159 unsigned int offset = 0;
160 for (unsigned int i = 0; i < struct_type->length; i++) {
161 if (strcmp(struct_type->fields.structure[i].name, ir->field) == 0)
163 offset += type_size(struct_type->fields.structure[i].type);
165 this->result.reg_offset += offset;
166 this->result.type = brw_type_for_base_type(ir->type);
170 fs_visitor::visit(ir_dereference_array *ir)
172 ir_constant *constant_index;
174 int element_size = type_size(ir->type);
176 constant_index = ir->array_index->as_constant();
178 ir->array->accept(this);
180 src.type = brw_type_for_base_type(ir->type);
182 if (constant_index) {
183 assert(src.file == UNIFORM || src.file == GRF);
184 src.reg_offset += constant_index->value.i[0] * element_size;
186 /* Variable index array dereference. We attach the variable index
187 * component to the reg as a pointer to a register containing the
188 * offset. Currently only uniform arrays are supported in this patch,
189 * and that reladdr pointer is resolved by
190 * move_uniform_array_access_to_pull_constants(). All other array types
191 * are lowered by lower_variable_index_to_cond_assign().
193 ir->array_index->accept(this);
196 index_reg = fs_reg(this, glsl_type::int_type);
197 emit(BRW_OPCODE_MUL, index_reg, this->result, fs_reg(element_size));
200 emit(BRW_OPCODE_ADD, index_reg, *src.reladdr, index_reg);
203 src.reladdr = ralloc(mem_ctx, fs_reg);
204 memcpy(src.reladdr, &index_reg, sizeof(index_reg));
210 fs_visitor::emit_lrp(fs_reg dst, fs_reg x, fs_reg y, fs_reg a)
213 !x.is_valid_3src() ||
214 !y.is_valid_3src() ||
215 !a.is_valid_3src()) {
216 /* We can't use the LRP instruction. Emit x*(1-a) + y*a. */
217 fs_reg y_times_a = fs_reg(this, glsl_type::float_type);
218 fs_reg one_minus_a = fs_reg(this, glsl_type::float_type);
219 fs_reg x_times_one_minus_a = fs_reg(this, glsl_type::float_type);
221 emit(MUL(y_times_a, y, a));
223 a.negate = !a.negate;
224 emit(ADD(one_minus_a, a, fs_reg(1.0f)));
225 emit(MUL(x_times_one_minus_a, x, one_minus_a));
227 emit(ADD(dst, x_times_one_minus_a, y_times_a));
229 /* The LRP instruction actually does op1 * op0 + op2 * (1 - op0), so
230 * we need to reorder the operands.
232 emit(LRP(dst, a, y, x));
237 fs_visitor::emit_minmax(uint32_t conditionalmod, fs_reg dst,
238 fs_reg src0, fs_reg src1)
243 inst = emit(BRW_OPCODE_SEL, dst, src0, src1);
244 inst->conditional_mod = conditionalmod;
246 emit(CMP(reg_null_d, src0, src1, conditionalmod));
248 inst = emit(BRW_OPCODE_SEL, dst, src0, src1);
249 inst->predicate = BRW_PREDICATE_NORMAL;
253 /* Instruction selection: Produce a MOV.sat instead of
254 * MIN(MAX(val, 0), 1) when possible.
257 fs_visitor::try_emit_saturate(ir_expression *ir)
259 ir_rvalue *sat_val = ir->as_rvalue_to_saturate();
264 fs_inst *pre_inst = (fs_inst *) this->instructions.get_tail();
266 sat_val->accept(this);
267 fs_reg src = this->result;
269 fs_inst *last_inst = (fs_inst *) this->instructions.get_tail();
271 /* If the last instruction from our accept() didn't generate our
272 * src, generate a saturated MOV
274 fs_inst *modify = get_instruction_generating_reg(pre_inst, last_inst, src);
275 if (!modify || modify->regs_written != 1) {
276 this->result = fs_reg(this, ir->type);
277 fs_inst *inst = emit(MOV(this->result, src));
278 inst->saturate = true;
280 modify->saturate = true;
289 fs_visitor::try_emit_mad(ir_expression *ir, int mul_arg)
291 /* 3-src instructions were introduced in gen6. */
295 /* MAD can only handle floating-point data. */
296 if (ir->type != glsl_type::float_type)
299 ir_rvalue *nonmul = ir->operands[1 - mul_arg];
300 ir_expression *mul = ir->operands[mul_arg]->as_expression();
302 if (!mul || mul->operation != ir_binop_mul)
305 if (nonmul->as_constant() ||
306 mul->operands[0]->as_constant() ||
307 mul->operands[1]->as_constant())
310 nonmul->accept(this);
311 fs_reg src0 = this->result;
313 mul->operands[0]->accept(this);
314 fs_reg src1 = this->result;
316 mul->operands[1]->accept(this);
317 fs_reg src2 = this->result;
319 this->result = fs_reg(this, ir->type);
320 emit(BRW_OPCODE_MAD, this->result, src0, src1, src2);
326 fs_visitor::visit(ir_expression *ir)
328 unsigned int operand;
332 assert(ir->get_num_operands() <= 3);
334 if (try_emit_saturate(ir))
336 if (ir->operation == ir_binop_add) {
337 if (try_emit_mad(ir, 0) || try_emit_mad(ir, 1))
341 for (operand = 0; operand < ir->get_num_operands(); operand++) {
342 ir->operands[operand]->accept(this);
343 if (this->result.file == BAD_FILE) {
344 fail("Failed to get tree for expression operand:\n");
345 ir->operands[operand]->print();
348 assert(this->result.is_valid_3src());
349 op[operand] = this->result;
351 /* Matrix expression operands should have been broken down to vector
352 * operations already.
354 assert(!ir->operands[operand]->type->is_matrix());
355 /* And then those vector operands should have been broken down to scalar.
357 assert(!ir->operands[operand]->type->is_vector());
360 /* Storage for our result. If our result goes into an assignment, it will
361 * just get copy-propagated out, so no worries.
363 this->result = fs_reg(this, ir->type);
365 switch (ir->operation) {
366 case ir_unop_logic_not:
367 /* Note that BRW_OPCODE_NOT is not appropriate here, since it is
368 * ones complement of the whole register, not just bit 0.
370 emit(XOR(this->result, op[0], fs_reg(1)));
373 op[0].negate = !op[0].negate;
374 emit(MOV(this->result, op[0]));
378 op[0].negate = false;
379 emit(MOV(this->result, op[0]));
382 temp = fs_reg(this, ir->type);
384 emit(MOV(this->result, fs_reg(0.0f)));
386 emit(CMP(reg_null_f, op[0], fs_reg(0.0f), BRW_CONDITIONAL_G));
387 inst = emit(MOV(this->result, fs_reg(1.0f)));
388 inst->predicate = BRW_PREDICATE_NORMAL;
390 emit(CMP(reg_null_f, op[0], fs_reg(0.0f), BRW_CONDITIONAL_L));
391 inst = emit(MOV(this->result, fs_reg(-1.0f)));
392 inst->predicate = BRW_PREDICATE_NORMAL;
396 emit_math(SHADER_OPCODE_RCP, this->result, op[0]);
400 emit_math(SHADER_OPCODE_EXP2, this->result, op[0]);
403 emit_math(SHADER_OPCODE_LOG2, this->result, op[0]);
407 assert(!"not reached: should be handled by ir_explog_to_explog2");
410 case ir_unop_sin_reduced:
411 emit_math(SHADER_OPCODE_SIN, this->result, op[0]);
414 case ir_unop_cos_reduced:
415 emit_math(SHADER_OPCODE_COS, this->result, op[0]);
419 emit(FS_OPCODE_DDX, this->result, op[0]);
422 emit(FS_OPCODE_DDY, this->result, op[0]);
426 emit(ADD(this->result, op[0], op[1]));
429 assert(!"not reached: should be handled by ir_sub_to_add_neg");
433 if (ir->type->is_integer()) {
434 /* For integer multiplication, the MUL uses the low 16 bits
435 * of one of the operands (src0 on gen6, src1 on gen7). The
436 * MACH accumulates in the contribution of the upper 16 bits
439 * FINISHME: Emit just the MUL if we know an operand is small
442 if (brw->gen >= 7 && dispatch_width == 16)
443 fail("16-wide explicit accumulator operands unsupported\n");
445 struct brw_reg acc = retype(brw_acc_reg(), this->result.type);
447 emit(MUL(acc, op[0], op[1]));
448 emit(MACH(reg_null_d, op[0], op[1]));
449 emit(MOV(this->result, fs_reg(acc)));
451 emit(MUL(this->result, op[0], op[1]));
454 case ir_binop_imul_high: {
455 if (brw->gen >= 7 && dispatch_width == 16)
456 fail("16-wide explicit accumulator operands unsupported\n");
458 struct brw_reg acc = retype(brw_acc_reg(), this->result.type);
460 emit(MUL(acc, op[0], op[1]));
461 emit(MACH(this->result, op[0], op[1]));
465 /* Floating point should be lowered by DIV_TO_MUL_RCP in the compiler. */
466 assert(ir->type->is_integer());
467 emit_math(SHADER_OPCODE_INT_QUOTIENT, this->result, op[0], op[1]);
469 case ir_binop_carry: {
470 if (brw->gen >= 7 && dispatch_width == 16)
471 fail("16-wide explicit accumulator operands unsupported\n");
473 struct brw_reg acc = retype(brw_acc_reg(), BRW_REGISTER_TYPE_UD);
475 emit(ADDC(reg_null_ud, op[0], op[1]));
476 emit(MOV(this->result, fs_reg(acc)));
479 case ir_binop_borrow: {
480 if (brw->gen >= 7 && dispatch_width == 16)
481 fail("16-wide explicit accumulator operands unsupported\n");
483 struct brw_reg acc = retype(brw_acc_reg(), BRW_REGISTER_TYPE_UD);
485 emit(SUBB(reg_null_ud, op[0], op[1]));
486 emit(MOV(this->result, fs_reg(acc)));
490 /* Floating point should be lowered by MOD_TO_FRACT in the compiler. */
491 assert(ir->type->is_integer());
492 emit_math(SHADER_OPCODE_INT_REMAINDER, this->result, op[0], op[1]);
496 case ir_binop_greater:
497 case ir_binop_lequal:
498 case ir_binop_gequal:
500 case ir_binop_all_equal:
501 case ir_binop_nequal:
502 case ir_binop_any_nequal:
503 resolve_bool_comparison(ir->operands[0], &op[0]);
504 resolve_bool_comparison(ir->operands[1], &op[1]);
506 emit(CMP(this->result, op[0], op[1],
507 brw_conditional_for_comparison(ir->operation)));
510 case ir_binop_logic_xor:
511 emit(XOR(this->result, op[0], op[1]));
514 case ir_binop_logic_or:
515 emit(OR(this->result, op[0], op[1]));
518 case ir_binop_logic_and:
519 emit(AND(this->result, op[0], op[1]));
524 assert(!"not reached: should be handled by brw_fs_channel_expressions");
528 assert(!"not reached: should be handled by lower_noise");
531 case ir_quadop_vector:
532 assert(!"not reached: should be handled by lower_quadop_vector");
535 case ir_binop_vector_extract:
536 assert(!"not reached: should be handled by lower_vec_index_to_cond_assign()");
539 case ir_triop_vector_insert:
540 assert(!"not reached: should be handled by lower_vector_insert()");
544 assert(!"not reached: should be handled by ldexp_to_arith()");
548 emit_math(SHADER_OPCODE_SQRT, this->result, op[0]);
552 emit_math(SHADER_OPCODE_RSQ, this->result, op[0]);
555 case ir_unop_bitcast_i2f:
556 case ir_unop_bitcast_u2f:
557 op[0].type = BRW_REGISTER_TYPE_F;
558 this->result = op[0];
561 case ir_unop_bitcast_f2u:
562 op[0].type = BRW_REGISTER_TYPE_UD;
563 this->result = op[0];
566 case ir_unop_bitcast_f2i:
567 op[0].type = BRW_REGISTER_TYPE_D;
568 this->result = op[0];
574 emit(MOV(this->result, op[0]));
578 emit(AND(this->result, op[0], fs_reg(1)));
581 temp = fs_reg(this, glsl_type::int_type);
582 emit(AND(temp, op[0], fs_reg(1)));
583 emit(MOV(this->result, temp));
587 emit(CMP(this->result, op[0], fs_reg(0.0f), BRW_CONDITIONAL_NZ));
590 emit(CMP(this->result, op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
594 emit(RNDZ(this->result, op[0]));
597 op[0].negate = !op[0].negate;
598 emit(RNDD(this->result, op[0]));
599 this->result.negate = true;
602 emit(RNDD(this->result, op[0]));
605 emit(FRC(this->result, op[0]));
607 case ir_unop_round_even:
608 emit(RNDE(this->result, op[0]));
613 resolve_ud_negate(&op[0]);
614 resolve_ud_negate(&op[1]);
615 emit_minmax(ir->operation == ir_binop_min ?
616 BRW_CONDITIONAL_L : BRW_CONDITIONAL_GE,
617 this->result, op[0], op[1]);
619 case ir_unop_pack_snorm_2x16:
620 case ir_unop_pack_snorm_4x8:
621 case ir_unop_pack_unorm_2x16:
622 case ir_unop_pack_unorm_4x8:
623 case ir_unop_unpack_snorm_2x16:
624 case ir_unop_unpack_snorm_4x8:
625 case ir_unop_unpack_unorm_2x16:
626 case ir_unop_unpack_unorm_4x8:
627 case ir_unop_unpack_half_2x16:
628 case ir_unop_pack_half_2x16:
629 assert(!"not reached: should be handled by lower_packing_builtins");
631 case ir_unop_unpack_half_2x16_split_x:
632 emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X, this->result, op[0]);
634 case ir_unop_unpack_half_2x16_split_y:
635 emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y, this->result, op[0]);
638 emit_math(SHADER_OPCODE_POW, this->result, op[0], op[1]);
641 case ir_unop_bitfield_reverse:
642 emit(BFREV(this->result, op[0]));
644 case ir_unop_bit_count:
645 emit(CBIT(this->result, op[0]));
647 case ir_unop_find_msb:
648 temp = fs_reg(this, glsl_type::uint_type);
649 emit(FBH(temp, op[0]));
651 /* FBH counts from the MSB side, while GLSL's findMSB() wants the count
652 * from the LSB side. If FBH didn't return an error (0xFFFFFFFF), then
653 * subtract the result from 31 to convert the MSB count into an LSB count.
656 /* FBH only supports UD type for dst, so use a MOV to convert UD to D. */
657 emit(MOV(this->result, temp));
658 emit(CMP(reg_null_d, this->result, fs_reg(-1), BRW_CONDITIONAL_NZ));
661 inst = emit(ADD(this->result, temp, fs_reg(31)));
662 inst->predicate = BRW_PREDICATE_NORMAL;
664 case ir_unop_find_lsb:
665 emit(FBL(this->result, op[0]));
667 case ir_triop_bitfield_extract:
668 /* Note that the instruction's argument order is reversed from GLSL
671 emit(BFE(this->result, op[2], op[1], op[0]));
674 emit(BFI1(this->result, op[0], op[1]));
677 emit(BFI2(this->result, op[0], op[1], op[2]));
679 case ir_quadop_bitfield_insert:
680 assert(!"not reached: should be handled by "
681 "lower_instructions::bitfield_insert_to_bfm_bfi");
684 case ir_unop_bit_not:
685 emit(NOT(this->result, op[0]));
687 case ir_binop_bit_and:
688 emit(AND(this->result, op[0], op[1]));
690 case ir_binop_bit_xor:
691 emit(XOR(this->result, op[0], op[1]));
693 case ir_binop_bit_or:
694 emit(OR(this->result, op[0], op[1]));
697 case ir_binop_lshift:
698 emit(SHL(this->result, op[0], op[1]));
701 case ir_binop_rshift:
702 if (ir->type->base_type == GLSL_TYPE_INT)
703 emit(ASR(this->result, op[0], op[1]));
705 emit(SHR(this->result, op[0], op[1]));
707 case ir_binop_pack_half_2x16_split:
708 emit(FS_OPCODE_PACK_HALF_2x16_SPLIT, this->result, op[0], op[1]);
710 case ir_binop_ubo_load: {
711 /* This IR node takes a constant uniform block and a constant or
712 * variable byte offset within the block and loads a vector from that.
714 ir_constant *uniform_block = ir->operands[0]->as_constant();
715 ir_constant *const_offset = ir->operands[1]->as_constant();
716 fs_reg surf_index = fs_reg(c->prog_data.base.binding_table.ubo_start +
717 uniform_block->value.u[0]);
719 fs_reg packed_consts = fs_reg(this, glsl_type::float_type);
720 packed_consts.type = result.type;
722 fs_reg const_offset_reg = fs_reg(const_offset->value.u[0] & ~15);
723 emit(fs_inst(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD,
724 packed_consts, surf_index, const_offset_reg));
726 packed_consts.smear = const_offset->value.u[0] % 16 / 4;
727 for (int i = 0; i < ir->type->vector_elements; i++) {
728 /* UBO bools are any nonzero value. We consider bools to be
729 * values with the low bit set to 1. Convert them using CMP.
731 if (ir->type->base_type == GLSL_TYPE_BOOL) {
732 emit(CMP(result, packed_consts, fs_reg(0u), BRW_CONDITIONAL_NZ));
734 emit(MOV(result, packed_consts));
737 packed_consts.smear++;
740 /* The std140 packing rules don't allow vectors to cross 16-byte
741 * boundaries, and a reg is 32 bytes.
743 assert(packed_consts.smear < 8);
746 /* Turn the byte offset into a dword offset. */
747 fs_reg base_offset = fs_reg(this, glsl_type::int_type);
748 emit(SHR(base_offset, op[1], fs_reg(2)));
750 for (int i = 0; i < ir->type->vector_elements; i++) {
751 emit(VARYING_PULL_CONSTANT_LOAD(result, surf_index,
754 if (ir->type->base_type == GLSL_TYPE_BOOL)
755 emit(CMP(result, result, fs_reg(0), BRW_CONDITIONAL_NZ));
761 result.reg_offset = 0;
766 /* Note that the instruction's argument order is reversed from GLSL
769 emit(MAD(this->result, op[2], op[1], op[0]));
773 emit_lrp(this->result, op[0], op[1], op[2]);
777 emit(CMP(reg_null_d, op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
778 inst = emit(BRW_OPCODE_SEL, this->result, op[1], op[2]);
779 inst->predicate = BRW_PREDICATE_NORMAL;
785 fs_visitor::emit_assignment_writes(fs_reg &l, fs_reg &r,
786 const glsl_type *type, bool predicated)
788 switch (type->base_type) {
789 case GLSL_TYPE_FLOAT:
793 for (unsigned int i = 0; i < type->components(); i++) {
794 l.type = brw_type_for_base_type(type);
795 r.type = brw_type_for_base_type(type);
797 if (predicated || !l.equals(r)) {
798 fs_inst *inst = emit(MOV(l, r));
799 inst->predicate = predicated ? BRW_PREDICATE_NORMAL : BRW_PREDICATE_NONE;
806 case GLSL_TYPE_ARRAY:
807 for (unsigned int i = 0; i < type->length; i++) {
808 emit_assignment_writes(l, r, type->fields.array, predicated);
812 case GLSL_TYPE_STRUCT:
813 for (unsigned int i = 0; i < type->length; i++) {
814 emit_assignment_writes(l, r, type->fields.structure[i].type,
819 case GLSL_TYPE_SAMPLER:
820 case GLSL_TYPE_ATOMIC_UINT:
824 case GLSL_TYPE_ERROR:
825 case GLSL_TYPE_INTERFACE:
826 assert(!"not reached");
831 /* If the RHS processing resulted in an instruction generating a
832 * temporary value, and it would be easy to rewrite the instruction to
833 * generate its result right into the LHS instead, do so. This ends
834 * up reliably removing instructions where it can be tricky to do so
835 * later without real UD chain information.
838 fs_visitor::try_rewrite_rhs_to_dst(ir_assignment *ir,
841 fs_inst *pre_rhs_inst,
842 fs_inst *last_rhs_inst)
844 /* Only attempt if we're doing a direct assignment. */
846 !(ir->lhs->type->is_scalar() ||
847 (ir->lhs->type->is_vector() &&
848 ir->write_mask == (1 << ir->lhs->type->vector_elements) - 1)))
851 /* Make sure the last instruction generated our source reg. */
852 fs_inst *modify = get_instruction_generating_reg(pre_rhs_inst,
858 /* If last_rhs_inst wrote a different number of components than our LHS,
859 * we can't safely rewrite it.
861 if (virtual_grf_sizes[dst.reg] != modify->regs_written)
864 /* Success! Rewrite the instruction. */
871 fs_visitor::visit(ir_assignment *ir)
876 /* FINISHME: arrays on the lhs */
877 ir->lhs->accept(this);
880 fs_inst *pre_rhs_inst = (fs_inst *) this->instructions.get_tail();
882 ir->rhs->accept(this);
885 fs_inst *last_rhs_inst = (fs_inst *) this->instructions.get_tail();
887 assert(l.file != BAD_FILE);
888 assert(r.file != BAD_FILE);
890 if (try_rewrite_rhs_to_dst(ir, l, r, pre_rhs_inst, last_rhs_inst))
894 emit_bool_to_cond_code(ir->condition);
897 if (ir->lhs->type->is_scalar() ||
898 ir->lhs->type->is_vector()) {
899 for (int i = 0; i < ir->lhs->type->vector_elements; i++) {
900 if (ir->write_mask & (1 << i)) {
901 inst = emit(MOV(l, r));
903 inst->predicate = BRW_PREDICATE_NORMAL;
909 emit_assignment_writes(l, r, ir->lhs->type, ir->condition != NULL);
914 fs_visitor::emit_texture_gen4(ir_texture *ir, fs_reg dst, fs_reg coordinate,
915 fs_reg shadow_c, fs_reg lod, fs_reg dPdy)
925 if (ir->shadow_comparitor) {
926 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
927 emit(MOV(fs_reg(MRF, base_mrf + mlen + i), coordinate));
928 coordinate.reg_offset++;
931 /* gen4's SIMD8 sampler always has the slots for u,v,r present.
932 * the unused slots must be zeroed.
934 for (int i = ir->coordinate->type->vector_elements; i < 3; i++) {
935 emit(MOV(fs_reg(MRF, base_mrf + mlen + i), fs_reg(0.0f)));
939 if (ir->op == ir_tex) {
940 /* There's no plain shadow compare message, so we use shadow
941 * compare with a bias of 0.0.
943 emit(MOV(fs_reg(MRF, base_mrf + mlen), fs_reg(0.0f)));
945 } else if (ir->op == ir_txb || ir->op == ir_txl) {
946 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
949 assert(!"Should not get here.");
952 emit(MOV(fs_reg(MRF, base_mrf + mlen), shadow_c));
954 } else if (ir->op == ir_tex) {
955 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
956 emit(MOV(fs_reg(MRF, base_mrf + mlen + i), coordinate));
957 coordinate.reg_offset++;
959 /* zero the others. */
960 for (int i = ir->coordinate->type->vector_elements; i<3; i++) {
961 emit(MOV(fs_reg(MRF, base_mrf + mlen + i), fs_reg(0.0f)));
963 /* gen4's SIMD8 sampler always has the slots for u,v,r present. */
965 } else if (ir->op == ir_txd) {
968 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
969 emit(MOV(fs_reg(MRF, base_mrf + mlen + i), coordinate));
970 coordinate.reg_offset++;
972 /* the slots for u and v are always present, but r is optional */
973 mlen += MAX2(ir->coordinate->type->vector_elements, 2);
976 * dPdx = dudx, dvdx, drdx
977 * dPdy = dudy, dvdy, drdy
979 * 1-arg: Does not exist.
981 * 2-arg: dudx dvdx dudy dvdy
982 * dPdx.x dPdx.y dPdy.x dPdy.y
985 * 3-arg: dudx dvdx drdx dudy dvdy drdy
986 * dPdx.x dPdx.y dPdx.z dPdy.x dPdy.y dPdy.z
989 for (int i = 0; i < ir->lod_info.grad.dPdx->type->vector_elements; i++) {
990 emit(MOV(fs_reg(MRF, base_mrf + mlen), dPdx));
993 mlen += MAX2(ir->lod_info.grad.dPdx->type->vector_elements, 2);
995 for (int i = 0; i < ir->lod_info.grad.dPdy->type->vector_elements; i++) {
996 emit(MOV(fs_reg(MRF, base_mrf + mlen), dPdy));
999 mlen += MAX2(ir->lod_info.grad.dPdy->type->vector_elements, 2);
1000 } else if (ir->op == ir_txs) {
1001 /* There's no SIMD8 resinfo message on Gen4. Use SIMD16 instead. */
1003 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), lod));
1006 /* Oh joy. gen4 doesn't have SIMD8 non-shadow-compare bias/lod
1007 * instructions. We'll need to do SIMD16 here.
1010 assert(ir->op == ir_txb || ir->op == ir_txl || ir->op == ir_txf);
1012 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
1013 emit(MOV(fs_reg(MRF, base_mrf + mlen + i * 2, coordinate.type),
1015 coordinate.reg_offset++;
1018 /* Initialize the rest of u/v/r with 0.0. Empirically, this seems to
1019 * be necessary for TXF (ld), but seems wise to do for all messages.
1021 for (int i = ir->coordinate->type->vector_elements; i < 3; i++) {
1022 emit(MOV(fs_reg(MRF, base_mrf + mlen + i * 2), fs_reg(0.0f)));
1025 /* lod/bias appears after u/v/r. */
1028 emit(MOV(fs_reg(MRF, base_mrf + mlen, lod.type), lod));
1031 /* The unused upper half. */
1036 /* Now, since we're doing simd16, the return is 2 interleaved
1037 * vec4s where the odd-indexed ones are junk. We'll need to move
1038 * this weirdness around to the expected layout.
1041 dst = fs_reg(GRF, virtual_grf_alloc(8),
1043 brw_type_for_base_type(ir->type) :
1044 BRW_REGISTER_TYPE_F));
1047 fs_inst *inst = NULL;
1050 inst = emit(SHADER_OPCODE_TEX, dst);
1053 inst = emit(FS_OPCODE_TXB, dst);
1056 inst = emit(SHADER_OPCODE_TXL, dst);
1059 inst = emit(SHADER_OPCODE_TXD, dst);
1062 inst = emit(SHADER_OPCODE_TXS, dst);
1065 inst = emit(SHADER_OPCODE_TXF, dst);
1068 fail("unrecognized texture opcode");
1070 inst->base_mrf = base_mrf;
1072 inst->header_present = true;
1073 inst->regs_written = simd16 ? 8 : 4;
1076 for (int i = 0; i < 4; i++) {
1077 emit(MOV(orig_dst, dst));
1078 orig_dst.reg_offset++;
1079 dst.reg_offset += 2;
1086 /* gen5's sampler has slots for u, v, r, array index, then optional
1087 * parameters like shadow comparitor or LOD bias. If optional
1088 * parameters aren't present, those base slots are optional and don't
1089 * need to be included in the message.
1091 * We don't fill in the unnecessary slots regardless, which may look
1092 * surprising in the disassembly.
1095 fs_visitor::emit_texture_gen5(ir_texture *ir, fs_reg dst, fs_reg coordinate,
1096 fs_reg shadow_c, fs_reg lod, fs_reg lod2,
1097 fs_reg sample_index)
1101 int reg_width = dispatch_width / 8;
1102 bool header_present = false;
1103 const int vector_elements =
1104 ir->coordinate ? ir->coordinate->type->vector_elements : 0;
1107 /* The offsets set up by the ir_texture visitor are in the
1108 * m1 header, so we can't go headerless.
1110 header_present = true;
1115 for (int i = 0; i < vector_elements; i++) {
1116 emit(MOV(fs_reg(MRF, base_mrf + mlen + i * reg_width, coordinate.type),
1118 coordinate.reg_offset++;
1120 mlen += vector_elements * reg_width;
1122 if (ir->shadow_comparitor) {
1123 mlen = MAX2(mlen, header_present + 4 * reg_width);
1125 emit(MOV(fs_reg(MRF, base_mrf + mlen), shadow_c));
1129 fs_inst *inst = NULL;
1132 inst = emit(SHADER_OPCODE_TEX, dst);
1135 mlen = MAX2(mlen, header_present + 4 * reg_width);
1136 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
1139 inst = emit(FS_OPCODE_TXB, dst);
1142 mlen = MAX2(mlen, header_present + 4 * reg_width);
1143 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
1146 inst = emit(SHADER_OPCODE_TXL, dst);
1149 mlen = MAX2(mlen, header_present + 4 * reg_width); /* skip over 'ai' */
1153 * dPdx = dudx, dvdx, drdx
1154 * dPdy = dudy, dvdy, drdy
1156 * Load up these values:
1157 * - dudx dudy dvdx dvdy drdx drdy
1158 * - dPdx.x dPdy.x dPdx.y dPdy.y dPdx.z dPdy.z
1160 for (int i = 0; i < ir->lod_info.grad.dPdx->type->vector_elements; i++) {
1161 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
1165 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod2));
1170 inst = emit(SHADER_OPCODE_TXD, dst);
1174 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), lod));
1176 inst = emit(SHADER_OPCODE_TXS, dst);
1178 case ir_query_levels:
1179 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), fs_reg(0u)));
1181 inst = emit(SHADER_OPCODE_TXS, dst);
1184 mlen = header_present + 4 * reg_width;
1185 emit(MOV(fs_reg(MRF, base_mrf + mlen - reg_width, BRW_REGISTER_TYPE_UD), lod));
1186 inst = emit(SHADER_OPCODE_TXF, dst);
1189 mlen = header_present + 4 * reg_width;
1192 emit(MOV(fs_reg(MRF, base_mrf + mlen - reg_width, BRW_REGISTER_TYPE_UD), fs_reg(0)));
1194 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), sample_index));
1196 inst = emit(SHADER_OPCODE_TXF_MS, dst);
1199 inst = emit(SHADER_OPCODE_LOD, dst);
1202 inst = emit(SHADER_OPCODE_TG4, dst);
1205 fail("unrecognized texture opcode");
1208 inst->base_mrf = base_mrf;
1210 inst->header_present = header_present;
1211 inst->regs_written = 4;
1214 fail("Message length >11 disallowed by hardware\n");
1221 fs_visitor::emit_texture_gen7(ir_texture *ir, fs_reg dst, fs_reg coordinate,
1222 fs_reg shadow_c, fs_reg lod, fs_reg lod2,
1223 fs_reg sample_index)
1225 int reg_width = dispatch_width / 8;
1226 bool header_present = false;
1228 fs_reg payload = fs_reg(this, glsl_type::float_type);
1229 fs_reg next = payload;
1231 if (ir->op == ir_tg4 || (ir->offset && ir->op != ir_txf)) {
1232 /* For general texture offsets (no txf workaround), we need a header to
1233 * put them in. Note that for 16-wide we're making space for two actual
1234 * hardware registers here, so the emit will have to fix up for this.
1236 * * ir4_tg4 needs to place its channel select in the header,
1237 * for interaction with ARB_texture_swizzle
1239 header_present = true;
1243 if (ir->shadow_comparitor) {
1244 emit(MOV(next, shadow_c));
1248 bool has_nonconstant_offset = ir->offset && !ir->offset->as_constant();
1249 bool coordinate_done = false;
1251 /* Set up the LOD info */
1257 emit(MOV(next, lod));
1261 emit(MOV(next, lod));
1265 if (dispatch_width == 16)
1266 fail("Gen7 does not support sample_d/sample_d_c in SIMD16 mode.");
1268 /* Load dPdx and the coordinate together:
1269 * [hdr], [ref], x, dPdx.x, dPdy.x, y, dPdx.y, dPdy.y, z, dPdx.z, dPdy.z
1271 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
1272 emit(MOV(next, coordinate));
1273 coordinate.reg_offset++;
1276 /* For cube map array, the coordinate is (u,v,r,ai) but there are
1277 * only derivatives for (u, v, r).
1279 if (i < ir->lod_info.grad.dPdx->type->vector_elements) {
1280 emit(MOV(next, lod));
1284 emit(MOV(next, lod2));
1290 coordinate_done = true;
1294 emit(MOV(next.retype(BRW_REGISTER_TYPE_UD), lod));
1297 case ir_query_levels:
1298 emit(MOV(next.retype(BRW_REGISTER_TYPE_UD), fs_reg(0u)));
1302 /* Unfortunately, the parameters for LD are intermixed: u, lod, v, r. */
1303 emit(MOV(next.retype(BRW_REGISTER_TYPE_D), coordinate));
1304 coordinate.reg_offset++;
1307 emit(MOV(next.retype(BRW_REGISTER_TYPE_D), lod));
1310 for (int i = 1; i < ir->coordinate->type->vector_elements; i++) {
1311 emit(MOV(next.retype(BRW_REGISTER_TYPE_D), coordinate));
1312 coordinate.reg_offset++;
1316 coordinate_done = true;
1319 emit(MOV(next.retype(BRW_REGISTER_TYPE_UD), sample_index));
1322 /* constant zero MCS; we arrange to never actually have a compressed
1323 * multisample surface here for now. TODO: issue ld_mcs to get this first,
1324 * if we ever support texturing from compressed multisample surfaces
1326 emit(MOV(next.retype(BRW_REGISTER_TYPE_UD), fs_reg(0u)));
1329 /* there is no offsetting for this message; just copy in the integer
1330 * texture coordinates
1332 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
1333 emit(MOV(next.retype(BRW_REGISTER_TYPE_D), coordinate));
1334 coordinate.reg_offset++;
1338 coordinate_done = true;
1341 if (has_nonconstant_offset) {
1342 if (ir->shadow_comparitor && dispatch_width == 16)
1343 fail("Gen7 does not support gather4_po_c in SIMD16 mode.");
1345 /* More crazy intermixing */
1346 ir->offset->accept(this);
1347 fs_reg offset_value = this->result;
1349 for (int i = 0; i < 2; i++) { /* u, v */
1350 emit(MOV(next, coordinate));
1351 coordinate.reg_offset++;
1355 for (int i = 0; i < 2; i++) { /* offu, offv */
1356 emit(MOV(next.retype(BRW_REGISTER_TYPE_D), offset_value));
1357 offset_value.reg_offset++;
1361 if (ir->coordinate->type->vector_elements == 3) { /* r if present */
1362 emit(MOV(next, coordinate));
1363 coordinate.reg_offset++;
1367 coordinate_done = true;
1372 /* Set up the coordinate (except for cases where it was done above) */
1373 if (ir->coordinate && !coordinate_done) {
1374 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
1375 emit(MOV(next, coordinate));
1376 coordinate.reg_offset++;
1381 /* Generate the SEND */
1382 fs_inst *inst = NULL;
1384 case ir_tex: inst = emit(SHADER_OPCODE_TEX, dst, payload); break;
1385 case ir_txb: inst = emit(FS_OPCODE_TXB, dst, payload); break;
1386 case ir_txl: inst = emit(SHADER_OPCODE_TXL, dst, payload); break;
1387 case ir_txd: inst = emit(SHADER_OPCODE_TXD, dst, payload); break;
1388 case ir_txf: inst = emit(SHADER_OPCODE_TXF, dst, payload); break;
1389 case ir_txf_ms: inst = emit(SHADER_OPCODE_TXF_MS, dst, payload); break;
1390 case ir_txs: inst = emit(SHADER_OPCODE_TXS, dst, payload); break;
1391 case ir_query_levels: inst = emit(SHADER_OPCODE_TXS, dst, payload); break;
1392 case ir_lod: inst = emit(SHADER_OPCODE_LOD, dst, payload); break;
1394 if (has_nonconstant_offset)
1395 inst = emit(SHADER_OPCODE_TG4_OFFSET, dst, payload);
1397 inst = emit(SHADER_OPCODE_TG4, dst, payload);
1400 inst->base_mrf = -1;
1402 inst->mlen = next.reg_offset * reg_width - header_present;
1404 inst->mlen = next.reg_offset * reg_width;
1405 inst->header_present = header_present;
1406 inst->regs_written = 4;
1408 virtual_grf_sizes[payload.reg] = next.reg_offset;
1409 if (inst->mlen > 11) {
1410 fail("Message length >11 disallowed by hardware\n");
1417 fs_visitor::rescale_texcoord(ir_texture *ir, fs_reg coordinate,
1418 bool is_rect, int sampler, int texunit)
1420 fs_inst *inst = NULL;
1421 bool needs_gl_clamp = true;
1422 fs_reg scale_x, scale_y;
1424 /* The 965 requires the EU to do the normalization of GL rectangle
1425 * texture coordinates. We use the program parameter state
1426 * tracking to get the scaling factor.
1430 (brw->gen >= 6 && (c->key.tex.gl_clamp_mask[0] & (1 << sampler) ||
1431 c->key.tex.gl_clamp_mask[1] & (1 << sampler))))) {
1432 struct gl_program_parameter_list *params = prog->Parameters;
1433 int tokens[STATE_LENGTH] = {
1435 STATE_TEXRECT_SCALE,
1441 if (dispatch_width == 16) {
1442 fail("rectangle scale uniform setup not supported on 16-wide\n");
1446 scale_x = fs_reg(UNIFORM, c->prog_data.nr_params);
1447 scale_y = fs_reg(UNIFORM, c->prog_data.nr_params + 1);
1449 GLuint index = _mesa_add_state_reference(params,
1450 (gl_state_index *)tokens);
1451 c->prog_data.param[c->prog_data.nr_params++] =
1452 &prog->Parameters->ParameterValues[index][0].f;
1453 c->prog_data.param[c->prog_data.nr_params++] =
1454 &prog->Parameters->ParameterValues[index][1].f;
1457 /* The 965 requires the EU to do the normalization of GL rectangle
1458 * texture coordinates. We use the program parameter state
1459 * tracking to get the scaling factor.
1461 if (brw->gen < 6 && is_rect) {
1462 fs_reg dst = fs_reg(this, ir->coordinate->type);
1463 fs_reg src = coordinate;
1466 emit(MUL(dst, src, scale_x));
1469 emit(MUL(dst, src, scale_y));
1470 } else if (is_rect) {
1471 /* On gen6+, the sampler handles the rectangle coordinates
1472 * natively, without needing rescaling. But that means we have
1473 * to do GL_CLAMP clamping at the [0, width], [0, height] scale,
1474 * not [0, 1] like the default case below.
1476 needs_gl_clamp = false;
1478 for (int i = 0; i < 2; i++) {
1479 if (c->key.tex.gl_clamp_mask[i] & (1 << sampler)) {
1480 fs_reg chan = coordinate;
1481 chan.reg_offset += i;
1483 inst = emit(BRW_OPCODE_SEL, chan, chan, brw_imm_f(0.0));
1484 inst->conditional_mod = BRW_CONDITIONAL_G;
1486 /* Our parameter comes in as 1.0/width or 1.0/height,
1487 * because that's what people normally want for doing
1488 * texture rectangle handling. We need width or height
1489 * for clamping, but we don't care enough to make a new
1490 * parameter type, so just invert back.
1492 fs_reg limit = fs_reg(this, glsl_type::float_type);
1493 emit(MOV(limit, i == 0 ? scale_x : scale_y));
1494 emit(SHADER_OPCODE_RCP, limit, limit);
1496 inst = emit(BRW_OPCODE_SEL, chan, chan, limit);
1497 inst->conditional_mod = BRW_CONDITIONAL_L;
1502 if (ir->coordinate && needs_gl_clamp) {
1503 for (unsigned int i = 0;
1504 i < MIN2(ir->coordinate->type->vector_elements, 3); i++) {
1505 if (c->key.tex.gl_clamp_mask[i] & (1 << sampler)) {
1506 fs_reg chan = coordinate;
1507 chan.reg_offset += i;
1509 fs_inst *inst = emit(MOV(chan, chan));
1510 inst->saturate = true;
1518 fs_visitor::visit(ir_texture *ir)
1520 fs_inst *inst = NULL;
1523 _mesa_get_sampler_uniform_value(ir->sampler, shader_prog, prog);
1524 /* FINISHME: We're failing to recompile our programs when the sampler is
1525 * updated. This only matters for the texture rectangle scale parameters
1526 * (pre-gen6, or gen6+ with GL_CLAMP).
1528 int texunit = prog->SamplerUnits[sampler];
1530 if (ir->op == ir_tg4) {
1531 /* When tg4 is used with the degenerate ZERO/ONE swizzles, don't bother
1532 * emitting anything other than setting up the constant result.
1534 ir_constant *chan = ir->lod_info.component->as_constant();
1535 int swiz = GET_SWZ(c->key.tex.swizzles[sampler], chan->value.i[0]);
1536 if (swiz == SWIZZLE_ZERO || swiz == SWIZZLE_ONE) {
1538 fs_reg res = fs_reg(this, glsl_type::vec4_type);
1541 for (int i=0; i<4; i++) {
1542 emit(MOV(res, fs_reg(swiz == SWIZZLE_ZERO ? 0.0f : 1.0f)));
1549 /* Should be lowered by do_lower_texture_projection */
1550 assert(!ir->projector);
1552 /* Should be lowered */
1553 assert(!ir->offset || !ir->offset->type->is_array());
1555 /* Generate code to compute all the subexpression trees. This has to be
1556 * done before loading any values into MRFs for the sampler message since
1557 * generating these values may involve SEND messages that need the MRFs.
1560 if (ir->coordinate) {
1561 ir->coordinate->accept(this);
1563 coordinate = rescale_texcoord(ir, this->result,
1564 ir->sampler->type->sampler_dimensionality ==
1565 GLSL_SAMPLER_DIM_RECT,
1569 fs_reg shadow_comparitor;
1570 if (ir->shadow_comparitor) {
1571 ir->shadow_comparitor->accept(this);
1572 shadow_comparitor = this->result;
1575 fs_reg lod, lod2, sample_index;
1580 case ir_query_levels:
1583 ir->lod_info.bias->accept(this);
1587 ir->lod_info.grad.dPdx->accept(this);
1590 ir->lod_info.grad.dPdy->accept(this);
1591 lod2 = this->result;
1596 ir->lod_info.lod->accept(this);
1600 ir->lod_info.sample_index->accept(this);
1601 sample_index = this->result;
1604 assert(!"Unrecognized texture opcode");
1607 /* Writemasking doesn't eliminate channels on SIMD8 texture
1608 * samples, so don't worry about them.
1610 fs_reg dst = fs_reg(this, glsl_type::get_instance(ir->type->base_type, 4, 1));
1612 if (brw->gen >= 7) {
1613 inst = emit_texture_gen7(ir, dst, coordinate, shadow_comparitor,
1614 lod, lod2, sample_index);
1615 } else if (brw->gen >= 5) {
1616 inst = emit_texture_gen5(ir, dst, coordinate, shadow_comparitor,
1617 lod, lod2, sample_index);
1619 inst = emit_texture_gen4(ir, dst, coordinate, shadow_comparitor,
1623 if (ir->offset != NULL && ir->op != ir_txf)
1624 inst->texture_offset = brw_texture_offset(ctx, ir->offset->as_constant());
1626 if (ir->op == ir_tg4)
1627 inst->texture_offset |= gather_channel(ir, sampler) << 16; // M0.2:16-17
1629 inst->sampler = sampler;
1631 if (ir->shadow_comparitor)
1632 inst->shadow_compare = true;
1634 /* fixup #layers for cube map arrays */
1635 if (ir->op == ir_txs) {
1636 glsl_type const *type = ir->sampler->type;
1637 if (type->sampler_dimensionality == GLSL_SAMPLER_DIM_CUBE &&
1638 type->sampler_array) {
1640 depth.reg_offset = 2;
1641 emit_math(SHADER_OPCODE_INT_QUOTIENT, depth, depth, fs_reg(6));
1645 swizzle_result(ir, dst, sampler);
1649 * Set up the gather channel based on the swizzle, for gather4.
1652 fs_visitor::gather_channel(ir_texture *ir, int sampler)
1654 ir_constant *chan = ir->lod_info.component->as_constant();
1655 int swiz = GET_SWZ(c->key.tex.swizzles[sampler], chan->value.i[0]);
1657 case SWIZZLE_X: return 0;
1659 /* gather4 sampler is broken for green channel on RG32F --
1660 * we must ask for blue instead.
1662 if (c->key.tex.gather_channel_quirk_mask & (1<<sampler))
1665 case SWIZZLE_Z: return 2;
1666 case SWIZZLE_W: return 3;
1668 assert(!"Not reached"); /* zero, one swizzles handled already */
1674 * Swizzle the result of a texture result. This is necessary for
1675 * EXT_texture_swizzle as well as DEPTH_TEXTURE_MODE for shadow comparisons.
1678 fs_visitor::swizzle_result(ir_texture *ir, fs_reg orig_val, int sampler)
1680 if (ir->op == ir_query_levels) {
1681 /* # levels is in .w */
1682 orig_val.reg_offset += 3;
1683 this->result = orig_val;
1687 this->result = orig_val;
1689 /* txs,lod don't actually sample the texture, so swizzling the result
1692 if (ir->op == ir_txs || ir->op == ir_lod || ir->op == ir_tg4)
1695 if (ir->type == glsl_type::float_type) {
1696 /* Ignore DEPTH_TEXTURE_MODE swizzling. */
1697 assert(ir->sampler->type->sampler_shadow);
1698 } else if (c->key.tex.swizzles[sampler] != SWIZZLE_NOOP) {
1699 fs_reg swizzled_result = fs_reg(this, glsl_type::vec4_type);
1701 for (int i = 0; i < 4; i++) {
1702 int swiz = GET_SWZ(c->key.tex.swizzles[sampler], i);
1703 fs_reg l = swizzled_result;
1706 if (swiz == SWIZZLE_ZERO) {
1707 emit(MOV(l, fs_reg(0.0f)));
1708 } else if (swiz == SWIZZLE_ONE) {
1709 emit(MOV(l, fs_reg(1.0f)));
1711 fs_reg r = orig_val;
1712 r.reg_offset += GET_SWZ(c->key.tex.swizzles[sampler], i);
1716 this->result = swizzled_result;
1721 fs_visitor::visit(ir_swizzle *ir)
1723 ir->val->accept(this);
1724 fs_reg val = this->result;
1726 if (ir->type->vector_elements == 1) {
1727 this->result.reg_offset += ir->mask.x;
1731 fs_reg result = fs_reg(this, ir->type);
1732 this->result = result;
1734 for (unsigned int i = 0; i < ir->type->vector_elements; i++) {
1735 fs_reg channel = val;
1753 channel.reg_offset += swiz;
1754 emit(MOV(result, channel));
1755 result.reg_offset++;
1760 fs_visitor::visit(ir_discard *ir)
1762 assert(ir->condition == NULL); /* FINISHME */
1764 /* We track our discarded pixels in f0.1. By predicating on it, we can
1765 * update just the flag bits that aren't yet discarded. By emitting a
1766 * CMP of g0 != g0, all our currently executing channels will get turned
1769 fs_reg some_reg = fs_reg(retype(brw_vec8_grf(0, 0),
1770 BRW_REGISTER_TYPE_UW));
1771 fs_inst *cmp = emit(CMP(reg_null_f, some_reg, some_reg,
1772 BRW_CONDITIONAL_NZ));
1773 cmp->predicate = BRW_PREDICATE_NORMAL;
1774 cmp->flag_subreg = 1;
1776 if (brw->gen >= 6) {
1777 /* For performance, after a discard, jump to the end of the shader.
1778 * However, many people will do foliage by discarding based on a
1779 * texture's alpha mask, and then continue on to texture with the
1780 * remaining pixels. To avoid trashing the derivatives for those
1781 * texture samples, we'll only jump if all of the pixels in the subspan
1782 * have been discarded.
1784 fs_inst *discard_jump = emit(FS_OPCODE_DISCARD_JUMP);
1785 discard_jump->flag_subreg = 1;
1786 discard_jump->predicate = BRW_PREDICATE_ALIGN1_ANY4H;
1787 discard_jump->predicate_inverse = true;
1792 fs_visitor::visit(ir_constant *ir)
1794 /* Set this->result to reg at the bottom of the function because some code
1795 * paths will cause this visitor to be applied to other fields. This will
1796 * cause the value stored in this->result to be modified.
1798 * Make reg constant so that it doesn't get accidentally modified along the
1799 * way. Yes, I actually had this problem. :(
1801 const fs_reg reg(this, ir->type);
1802 fs_reg dst_reg = reg;
1804 if (ir->type->is_array()) {
1805 const unsigned size = type_size(ir->type->fields.array);
1807 for (unsigned i = 0; i < ir->type->length; i++) {
1808 ir->array_elements[i]->accept(this);
1809 fs_reg src_reg = this->result;
1811 dst_reg.type = src_reg.type;
1812 for (unsigned j = 0; j < size; j++) {
1813 emit(MOV(dst_reg, src_reg));
1814 src_reg.reg_offset++;
1815 dst_reg.reg_offset++;
1818 } else if (ir->type->is_record()) {
1819 foreach_list(node, &ir->components) {
1820 ir_constant *const field = (ir_constant *) node;
1821 const unsigned size = type_size(field->type);
1823 field->accept(this);
1824 fs_reg src_reg = this->result;
1826 dst_reg.type = src_reg.type;
1827 for (unsigned j = 0; j < size; j++) {
1828 emit(MOV(dst_reg, src_reg));
1829 src_reg.reg_offset++;
1830 dst_reg.reg_offset++;
1834 const unsigned size = type_size(ir->type);
1836 for (unsigned i = 0; i < size; i++) {
1837 switch (ir->type->base_type) {
1838 case GLSL_TYPE_FLOAT:
1839 emit(MOV(dst_reg, fs_reg(ir->value.f[i])));
1841 case GLSL_TYPE_UINT:
1842 emit(MOV(dst_reg, fs_reg(ir->value.u[i])));
1845 emit(MOV(dst_reg, fs_reg(ir->value.i[i])));
1847 case GLSL_TYPE_BOOL:
1848 emit(MOV(dst_reg, fs_reg((int)ir->value.b[i])));
1851 assert(!"Non-float/uint/int/bool constant");
1853 dst_reg.reg_offset++;
1861 fs_visitor::emit_bool_to_cond_code(ir_rvalue *ir)
1863 ir_expression *expr = ir->as_expression();
1866 expr->operation != ir_binop_logic_and &&
1867 expr->operation != ir_binop_logic_or &&
1868 expr->operation != ir_binop_logic_xor) {
1872 assert(expr->get_num_operands() <= 2);
1873 for (unsigned int i = 0; i < expr->get_num_operands(); i++) {
1874 assert(expr->operands[i]->type->is_scalar());
1876 expr->operands[i]->accept(this);
1877 op[i] = this->result;
1879 resolve_ud_negate(&op[i]);
1882 switch (expr->operation) {
1883 case ir_unop_logic_not:
1884 inst = emit(AND(reg_null_d, op[0], fs_reg(1)));
1885 inst->conditional_mod = BRW_CONDITIONAL_Z;
1889 if (brw->gen >= 6) {
1890 emit(CMP(reg_null_d, op[0], fs_reg(0.0f), BRW_CONDITIONAL_NZ));
1892 inst = emit(MOV(reg_null_f, op[0]));
1893 inst->conditional_mod = BRW_CONDITIONAL_NZ;
1898 if (brw->gen >= 6) {
1899 emit(CMP(reg_null_d, op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
1901 inst = emit(MOV(reg_null_d, op[0]));
1902 inst->conditional_mod = BRW_CONDITIONAL_NZ;
1906 case ir_binop_greater:
1907 case ir_binop_gequal:
1909 case ir_binop_lequal:
1910 case ir_binop_equal:
1911 case ir_binop_all_equal:
1912 case ir_binop_nequal:
1913 case ir_binop_any_nequal:
1914 resolve_bool_comparison(expr->operands[0], &op[0]);
1915 resolve_bool_comparison(expr->operands[1], &op[1]);
1917 emit(CMP(reg_null_d, op[0], op[1],
1918 brw_conditional_for_comparison(expr->operation)));
1922 assert(!"not reached");
1923 fail("bad cond code\n");
1931 fs_inst *inst = emit(AND(reg_null_d, this->result, fs_reg(1)));
1932 inst->conditional_mod = BRW_CONDITIONAL_NZ;
1936 * Emit a gen6 IF statement with the comparison folded into the IF
1940 fs_visitor::emit_if_gen6(ir_if *ir)
1942 ir_expression *expr = ir->condition->as_expression();
1949 assert(expr->get_num_operands() <= 2);
1950 for (unsigned int i = 0; i < expr->get_num_operands(); i++) {
1951 assert(expr->operands[i]->type->is_scalar());
1953 expr->operands[i]->accept(this);
1954 op[i] = this->result;
1957 switch (expr->operation) {
1958 case ir_unop_logic_not:
1959 case ir_binop_logic_xor:
1960 case ir_binop_logic_or:
1961 case ir_binop_logic_and:
1962 /* For operations on bool arguments, only the low bit of the bool is
1963 * valid, and the others are undefined. Fall back to the condition
1969 inst = emit(BRW_OPCODE_IF, reg_null_f, op[0], fs_reg(0));
1970 inst->conditional_mod = BRW_CONDITIONAL_NZ;
1974 emit(IF(op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
1977 case ir_binop_greater:
1978 case ir_binop_gequal:
1980 case ir_binop_lequal:
1981 case ir_binop_equal:
1982 case ir_binop_all_equal:
1983 case ir_binop_nequal:
1984 case ir_binop_any_nequal:
1985 resolve_bool_comparison(expr->operands[0], &op[0]);
1986 resolve_bool_comparison(expr->operands[1], &op[1]);
1988 emit(IF(op[0], op[1],
1989 brw_conditional_for_comparison(expr->operation)));
1992 assert(!"not reached");
1993 emit(IF(op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
1994 fail("bad condition\n");
1999 emit_bool_to_cond_code(ir->condition);
2000 fs_inst *inst = emit(BRW_OPCODE_IF);
2001 inst->predicate = BRW_PREDICATE_NORMAL;
2005 * Try to replace IF/MOV/ELSE/MOV/ENDIF with SEL.
2007 * Many GLSL shaders contain the following pattern:
2009 * x = condition ? foo : bar
2011 * The compiler emits an ir_if tree for this, since each subexpression might be
2012 * a complex tree that could have side-effects or short-circuit logic.
2014 * However, the common case is to simply select one of two constants or
2015 * variable values---which is exactly what SEL is for. In this case, the
2016 * assembly looks like:
2024 * which can be easily translated into:
2026 * (+f0) SEL dst src0 src1
2028 * If src0 is an immediate value, we promote it to a temporary GRF.
2031 fs_visitor::try_replace_with_sel()
2033 fs_inst *endif_inst = (fs_inst *) instructions.get_tail();
2034 assert(endif_inst->opcode == BRW_OPCODE_ENDIF);
2036 /* Pattern match in reverse: IF, MOV, ELSE, MOV, ENDIF. */
2038 BRW_OPCODE_IF, BRW_OPCODE_MOV, BRW_OPCODE_ELSE, BRW_OPCODE_MOV,
2041 fs_inst *match = (fs_inst *) endif_inst->prev;
2042 for (int i = 0; i < 4; i++) {
2043 if (match->is_head_sentinel() || match->opcode != opcodes[4-i-1])
2045 match = (fs_inst *) match->prev;
2048 /* The opcodes match; it looks like the right sequence of instructions. */
2049 fs_inst *else_mov = (fs_inst *) endif_inst->prev;
2050 fs_inst *then_mov = (fs_inst *) else_mov->prev->prev;
2051 fs_inst *if_inst = (fs_inst *) then_mov->prev;
2053 /* Check that the MOVs are the right form. */
2054 if (then_mov->dst.equals(else_mov->dst) &&
2055 !then_mov->is_partial_write() &&
2056 !else_mov->is_partial_write()) {
2058 /* Remove the matched instructions; we'll emit a SEL to replace them. */
2059 while (!if_inst->next->is_tail_sentinel())
2060 if_inst->next->remove();
2063 /* Only the last source register can be a constant, so if the MOV in
2064 * the "then" clause uses a constant, we need to put it in a temporary.
2066 fs_reg src0(then_mov->src[0]);
2067 if (src0.file == IMM) {
2068 src0 = fs_reg(this, glsl_type::float_type);
2069 src0.type = then_mov->src[0].type;
2070 emit(MOV(src0, then_mov->src[0]));
2074 if (if_inst->conditional_mod) {
2075 /* Sandybridge-specific IF with embedded comparison */
2076 emit(CMP(reg_null_d, if_inst->src[0], if_inst->src[1],
2077 if_inst->conditional_mod));
2078 sel = emit(BRW_OPCODE_SEL, then_mov->dst, src0, else_mov->src[0]);
2079 sel->predicate = BRW_PREDICATE_NORMAL;
2081 /* Separate CMP and IF instructions */
2082 sel = emit(BRW_OPCODE_SEL, then_mov->dst, src0, else_mov->src[0]);
2083 sel->predicate = if_inst->predicate;
2084 sel->predicate_inverse = if_inst->predicate_inverse;
2090 fs_visitor::visit(ir_if *ir)
2092 if (brw->gen < 6 && dispatch_width == 16) {
2093 fail("Can't support (non-uniform) control flow on 16-wide\n");
2096 /* Don't point the annotation at the if statement, because then it plus
2097 * the then and else blocks get printed.
2099 this->base_ir = ir->condition;
2101 if (brw->gen == 6) {
2104 emit_bool_to_cond_code(ir->condition);
2106 emit(IF(BRW_PREDICATE_NORMAL));
2109 foreach_list(node, &ir->then_instructions) {
2110 ir_instruction *ir = (ir_instruction *)node;
2116 if (!ir->else_instructions.is_empty()) {
2117 emit(BRW_OPCODE_ELSE);
2119 foreach_list(node, &ir->else_instructions) {
2120 ir_instruction *ir = (ir_instruction *)node;
2127 emit(BRW_OPCODE_ENDIF);
2129 try_replace_with_sel();
2133 fs_visitor::visit(ir_loop *ir)
2135 fs_reg counter = reg_undef;
2137 if (brw->gen < 6 && dispatch_width == 16) {
2138 fail("Can't support (non-uniform) control flow on 16-wide\n");
2142 this->base_ir = ir->counter;
2143 ir->counter->accept(this);
2144 counter = *(variable_storage(ir->counter));
2147 this->base_ir = ir->from;
2148 ir->from->accept(this);
2150 emit(MOV(counter, this->result));
2154 this->base_ir = NULL;
2155 emit(BRW_OPCODE_DO);
2158 this->base_ir = ir->to;
2159 ir->to->accept(this);
2161 emit(CMP(reg_null_d, counter, this->result,
2162 brw_conditional_for_comparison(ir->cmp)));
2164 fs_inst *inst = emit(BRW_OPCODE_BREAK);
2165 inst->predicate = BRW_PREDICATE_NORMAL;
2168 foreach_list(node, &ir->body_instructions) {
2169 ir_instruction *ir = (ir_instruction *)node;
2175 if (ir->increment) {
2176 this->base_ir = ir->increment;
2177 ir->increment->accept(this);
2178 emit(ADD(counter, counter, this->result));
2181 this->base_ir = NULL;
2182 emit(BRW_OPCODE_WHILE);
2186 fs_visitor::visit(ir_loop_jump *ir)
2189 case ir_loop_jump::jump_break:
2190 emit(BRW_OPCODE_BREAK);
2192 case ir_loop_jump::jump_continue:
2193 emit(BRW_OPCODE_CONTINUE);
2199 fs_visitor::visit(ir_call *ir)
2201 assert(!"FINISHME");
2205 fs_visitor::visit(ir_return *ir)
2207 assert(!"FINISHME");
2211 fs_visitor::visit(ir_function *ir)
2213 /* Ignore function bodies other than main() -- we shouldn't see calls to
2214 * them since they should all be inlined before we get to ir_to_mesa.
2216 if (strcmp(ir->name, "main") == 0) {
2217 const ir_function_signature *sig;
2220 sig = ir->matching_signature(NULL, &empty);
2224 foreach_list(node, &sig->body) {
2225 ir_instruction *ir = (ir_instruction *)node;
2234 fs_visitor::visit(ir_function_signature *ir)
2236 assert(!"not reached");
2241 fs_visitor::visit(ir_emit_vertex *)
2243 assert(!"not reached");
2247 fs_visitor::visit(ir_end_primitive *)
2249 assert(!"not reached");
2253 fs_visitor::emit(fs_inst inst)
2255 fs_inst *list_inst = new(mem_ctx) fs_inst;
2262 fs_visitor::emit(fs_inst *inst)
2264 if (force_uncompressed_stack > 0)
2265 inst->force_uncompressed = true;
2266 else if (force_sechalf_stack > 0)
2267 inst->force_sechalf = true;
2269 inst->annotation = this->current_annotation;
2270 inst->ir = this->base_ir;
2272 this->instructions.push_tail(inst);
2278 fs_visitor::emit(exec_list list)
2280 foreach_list_safe(node, &list) {
2281 fs_inst *inst = (fs_inst *)node;
2287 /** Emits a dummy fragment shader consisting of magenta for bringup purposes. */
2289 fs_visitor::emit_dummy_fs()
2291 int reg_width = dispatch_width / 8;
2293 /* Everyone's favorite color. */
2294 emit(MOV(fs_reg(MRF, 2 + 0 * reg_width), fs_reg(1.0f)));
2295 emit(MOV(fs_reg(MRF, 2 + 1 * reg_width), fs_reg(0.0f)));
2296 emit(MOV(fs_reg(MRF, 2 + 2 * reg_width), fs_reg(1.0f)));
2297 emit(MOV(fs_reg(MRF, 2 + 3 * reg_width), fs_reg(0.0f)));
2300 write = emit(FS_OPCODE_FB_WRITE, fs_reg(0), fs_reg(0));
2301 write->base_mrf = 2;
2302 write->mlen = 4 * reg_width;
2306 /* The register location here is relative to the start of the URB
2307 * data. It will get adjusted to be a real location before
2308 * generate_code() time.
2311 fs_visitor::interp_reg(int location, int channel)
2313 int regnr = c->prog_data.urb_setup[location] * 2 + channel / 2;
2314 int stride = (channel & 1) * 4;
2316 assert(c->prog_data.urb_setup[location] != -1);
2318 return brw_vec1_grf(regnr, stride);
2321 /** Emits the interpolation for the varying inputs. */
2323 fs_visitor::emit_interpolation_setup_gen4()
2325 this->current_annotation = "compute pixel centers";
2326 this->pixel_x = fs_reg(this, glsl_type::uint_type);
2327 this->pixel_y = fs_reg(this, glsl_type::uint_type);
2328 this->pixel_x.type = BRW_REGISTER_TYPE_UW;
2329 this->pixel_y.type = BRW_REGISTER_TYPE_UW;
2331 emit(FS_OPCODE_PIXEL_X, this->pixel_x);
2332 emit(FS_OPCODE_PIXEL_Y, this->pixel_y);
2334 this->current_annotation = "compute pixel deltas from v0";
2336 this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC] =
2337 fs_reg(this, glsl_type::vec2_type);
2338 this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC] =
2339 this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC];
2340 this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC].reg_offset++;
2342 this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC] =
2343 fs_reg(this, glsl_type::float_type);
2344 this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC] =
2345 fs_reg(this, glsl_type::float_type);
2347 emit(ADD(this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
2348 this->pixel_x, fs_reg(negate(brw_vec1_grf(1, 0)))));
2349 emit(ADD(this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
2350 this->pixel_y, fs_reg(negate(brw_vec1_grf(1, 1)))));
2352 this->current_annotation = "compute pos.w and 1/pos.w";
2353 /* Compute wpos.w. It's always in our setup, since it's needed to
2354 * interpolate the other attributes.
2356 this->wpos_w = fs_reg(this, glsl_type::float_type);
2357 emit(FS_OPCODE_LINTERP, wpos_w,
2358 this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
2359 this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
2360 interp_reg(VARYING_SLOT_POS, 3));
2361 /* Compute the pixel 1/W value from wpos.w. */
2362 this->pixel_w = fs_reg(this, glsl_type::float_type);
2363 emit_math(SHADER_OPCODE_RCP, this->pixel_w, wpos_w);
2364 this->current_annotation = NULL;
2367 /** Emits the interpolation for the varying inputs. */
2369 fs_visitor::emit_interpolation_setup_gen6()
2371 struct brw_reg g1_uw = retype(brw_vec1_grf(1, 0), BRW_REGISTER_TYPE_UW);
2373 /* If the pixel centers end up used, the setup is the same as for gen4. */
2374 this->current_annotation = "compute pixel centers";
2375 fs_reg int_pixel_x = fs_reg(this, glsl_type::uint_type);
2376 fs_reg int_pixel_y = fs_reg(this, glsl_type::uint_type);
2377 int_pixel_x.type = BRW_REGISTER_TYPE_UW;
2378 int_pixel_y.type = BRW_REGISTER_TYPE_UW;
2379 emit(ADD(int_pixel_x,
2380 fs_reg(stride(suboffset(g1_uw, 4), 2, 4, 0)),
2381 fs_reg(brw_imm_v(0x10101010))));
2382 emit(ADD(int_pixel_y,
2383 fs_reg(stride(suboffset(g1_uw, 5), 2, 4, 0)),
2384 fs_reg(brw_imm_v(0x11001100))));
2386 /* As of gen6, we can no longer mix float and int sources. We have
2387 * to turn the integer pixel centers into floats for their actual
2390 this->pixel_x = fs_reg(this, glsl_type::float_type);
2391 this->pixel_y = fs_reg(this, glsl_type::float_type);
2392 emit(MOV(this->pixel_x, int_pixel_x));
2393 emit(MOV(this->pixel_y, int_pixel_y));
2395 this->current_annotation = "compute pos.w";
2396 this->pixel_w = fs_reg(brw_vec8_grf(c->source_w_reg, 0));
2397 this->wpos_w = fs_reg(this, glsl_type::float_type);
2398 emit_math(SHADER_OPCODE_RCP, this->wpos_w, this->pixel_w);
2400 for (int i = 0; i < BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT; ++i) {
2401 uint8_t reg = c->barycentric_coord_reg[i];
2402 this->delta_x[i] = fs_reg(brw_vec8_grf(reg, 0));
2403 this->delta_y[i] = fs_reg(brw_vec8_grf(reg + 1, 0));
2406 this->current_annotation = NULL;
2410 fs_visitor::emit_color_write(int target, int index, int first_color_mrf)
2412 int reg_width = dispatch_width / 8;
2414 fs_reg color = outputs[target];
2417 /* If there's no color data to be written, skip it. */
2418 if (color.file == BAD_FILE)
2421 color.reg_offset += index;
2423 if (dispatch_width == 8 || brw->gen >= 6) {
2424 /* SIMD8 write looks like:
2430 * gen6 SIMD16 DP write looks like:
2440 inst = emit(MOV(fs_reg(MRF, first_color_mrf + index * reg_width,
2443 inst->saturate = c->key.clamp_fragment_color;
2445 /* pre-gen6 SIMD16 single source DP write looks like:
2455 if (brw->has_compr4) {
2456 /* By setting the high bit of the MRF register number, we
2457 * indicate that we want COMPR4 mode - instead of doing the
2458 * usual destination + 1 for the second half we get
2461 inst = emit(MOV(fs_reg(MRF, BRW_MRF_COMPR4 + first_color_mrf + index,
2464 inst->saturate = c->key.clamp_fragment_color;
2466 push_force_uncompressed();
2467 inst = emit(MOV(fs_reg(MRF, first_color_mrf + index, color.type),
2469 inst->saturate = c->key.clamp_fragment_color;
2470 pop_force_uncompressed();
2472 push_force_sechalf();
2473 color.sechalf = true;
2474 inst = emit(MOV(fs_reg(MRF, first_color_mrf + index + 4, color.type),
2476 inst->saturate = c->key.clamp_fragment_color;
2477 pop_force_sechalf();
2478 color.sechalf = false;
2484 fs_visitor::emit_fb_writes()
2486 this->current_annotation = "FB write header";
2487 bool header_present = true;
2488 /* We can potentially have a message length of up to 15, so we have to set
2489 * base_mrf to either 0 or 1 in order to fit in m0..m15.
2493 int reg_width = dispatch_width / 8;
2494 bool do_dual_src = this->dual_src_output.file != BAD_FILE;
2495 bool src0_alpha_to_render_target = false;
2497 if (dispatch_width == 16 && do_dual_src) {
2498 fail("GL_ARB_blend_func_extended not yet supported in 16-wide.");
2499 do_dual_src = false;
2502 /* From the Sandy Bridge PRM, volume 4, page 198:
2504 * "Dispatched Pixel Enables. One bit per pixel indicating
2505 * which pixels were originally enabled when the thread was
2506 * dispatched. This field is only required for the end-of-
2507 * thread message and on all dual-source messages."
2509 if (brw->gen >= 6 &&
2510 !this->fp->UsesKill &&
2512 c->key.nr_color_regions == 1) {
2513 header_present = false;
2516 if (header_present) {
2517 src0_alpha_to_render_target = brw->gen >= 6 &&
2519 c->key.replicate_alpha;
2524 if (c->aa_dest_stencil_reg) {
2525 push_force_uncompressed();
2526 emit(MOV(fs_reg(MRF, nr++),
2527 fs_reg(brw_vec8_grf(c->aa_dest_stencil_reg, 0))));
2528 pop_force_uncompressed();
2531 c->prog_data.uses_omask =
2532 fp->Base.OutputsWritten & BITFIELD64_BIT(FRAG_RESULT_SAMPLE_MASK);
2533 if(c->prog_data.uses_omask) {
2534 this->current_annotation = "FB write oMask";
2535 assert(this->sample_mask.file != BAD_FILE);
2536 /* Hand over gl_SampleMask. Only lower 16 bits are relevant. */
2537 emit(FS_OPCODE_SET_OMASK, fs_reg(MRF, nr, BRW_REGISTER_TYPE_UW), this->sample_mask);
2541 /* Reserve space for color. It'll be filled in per MRT below. */
2543 nr += 4 * reg_width;
2546 if (src0_alpha_to_render_target)
2549 if (c->source_depth_to_render_target) {
2550 if (brw->gen == 6 && dispatch_width == 16) {
2551 /* For outputting oDepth on gen6, SIMD8 writes have to be
2552 * used. This would require 8-wide moves of each half to
2553 * message regs, kind of like pre-gen5 SIMD16 FB writes.
2554 * Just bail on doing so for now.
2556 fail("Missing support for simd16 depth writes on gen6\n");
2559 if (prog->OutputsWritten & BITFIELD64_BIT(FRAG_RESULT_DEPTH)) {
2560 /* Hand over gl_FragDepth. */
2561 assert(this->frag_depth.file != BAD_FILE);
2562 emit(MOV(fs_reg(MRF, nr), this->frag_depth));
2564 /* Pass through the payload depth. */
2565 emit(MOV(fs_reg(MRF, nr),
2566 fs_reg(brw_vec8_grf(c->source_depth_reg, 0))));
2571 if (c->dest_depth_reg) {
2572 emit(MOV(fs_reg(MRF, nr),
2573 fs_reg(brw_vec8_grf(c->dest_depth_reg, 0))));
2578 fs_reg src0 = this->outputs[0];
2579 fs_reg src1 = this->dual_src_output;
2581 this->current_annotation = ralloc_asprintf(this->mem_ctx,
2583 for (int i = 0; i < 4; i++) {
2584 fs_inst *inst = emit(MOV(fs_reg(MRF, color_mrf + i, src0.type), src0));
2586 inst->saturate = c->key.clamp_fragment_color;
2589 this->current_annotation = ralloc_asprintf(this->mem_ctx,
2591 for (int i = 0; i < 4; i++) {
2592 fs_inst *inst = emit(MOV(fs_reg(MRF, color_mrf + 4 + i, src1.type),
2595 inst->saturate = c->key.clamp_fragment_color;
2598 if (INTEL_DEBUG & DEBUG_SHADER_TIME)
2599 emit_shader_time_end();
2601 fs_inst *inst = emit(FS_OPCODE_FB_WRITE);
2603 inst->base_mrf = base_mrf;
2604 inst->mlen = nr - base_mrf;
2606 inst->header_present = header_present;
2608 c->prog_data.dual_src_blend = true;
2609 this->current_annotation = NULL;
2613 for (int target = 0; target < c->key.nr_color_regions; target++) {
2614 this->current_annotation = ralloc_asprintf(this->mem_ctx,
2615 "FB write target %d",
2617 /* If src0_alpha_to_render_target is true, include source zero alpha
2618 * data in RenderTargetWrite message for targets > 0.
2620 int write_color_mrf = color_mrf;
2621 if (src0_alpha_to_render_target && target != 0) {
2623 fs_reg color = outputs[0];
2624 color.reg_offset += 3;
2626 inst = emit(MOV(fs_reg(MRF, write_color_mrf, color.type),
2628 inst->saturate = c->key.clamp_fragment_color;
2629 write_color_mrf = color_mrf + reg_width;
2632 for (unsigned i = 0; i < this->output_components[target]; i++)
2633 emit_color_write(target, i, write_color_mrf);
2636 if (target == c->key.nr_color_regions - 1) {
2639 if (INTEL_DEBUG & DEBUG_SHADER_TIME)
2640 emit_shader_time_end();
2643 fs_inst *inst = emit(FS_OPCODE_FB_WRITE);
2644 inst->target = target;
2645 inst->base_mrf = base_mrf;
2646 if (src0_alpha_to_render_target && target == 0)
2647 inst->mlen = nr - base_mrf - reg_width;
2649 inst->mlen = nr - base_mrf;
2651 inst->header_present = header_present;
2654 if (c->key.nr_color_regions == 0) {
2655 /* Even if there's no color buffers enabled, we still need to send
2656 * alpha out the pipeline to our null renderbuffer to support
2657 * alpha-testing, alpha-to-coverage, and so on.
2659 emit_color_write(0, 3, color_mrf);
2661 if (INTEL_DEBUG & DEBUG_SHADER_TIME)
2662 emit_shader_time_end();
2664 fs_inst *inst = emit(FS_OPCODE_FB_WRITE);
2665 inst->base_mrf = base_mrf;
2666 inst->mlen = nr - base_mrf;
2668 inst->header_present = header_present;
2671 this->current_annotation = NULL;
2675 fs_visitor::resolve_ud_negate(fs_reg *reg)
2677 if (reg->type != BRW_REGISTER_TYPE_UD ||
2681 fs_reg temp = fs_reg(this, glsl_type::uint_type);
2682 emit(MOV(temp, *reg));
2687 fs_visitor::resolve_bool_comparison(ir_rvalue *rvalue, fs_reg *reg)
2689 if (rvalue->type != glsl_type::bool_type)
2692 fs_reg temp = fs_reg(this, glsl_type::bool_type);
2693 emit(AND(temp, *reg, fs_reg(1)));
2697 fs_visitor::fs_visitor(struct brw_context *brw,
2698 struct brw_wm_compile *c,
2699 struct gl_shader_program *shader_prog,
2700 struct gl_fragment_program *fp,
2701 unsigned dispatch_width)
2702 : dispatch_width(dispatch_width)
2707 this->prog = &fp->Base;
2708 this->shader_prog = shader_prog;
2709 this->prog = &fp->Base;
2710 this->stage_prog_data = &c->prog_data.base;
2711 this->ctx = &brw->ctx;
2712 this->mem_ctx = ralloc_context(NULL);
2714 shader = (struct brw_shader *)
2715 shader_prog->_LinkedShaders[MESA_SHADER_FRAGMENT];
2718 this->failed = false;
2719 this->variable_ht = hash_table_ctor(0,
2720 hash_table_pointer_hash,
2721 hash_table_pointer_compare);
2723 memset(this->outputs, 0, sizeof(this->outputs));
2724 memset(this->output_components, 0, sizeof(this->output_components));
2725 this->first_non_payload_grf = 0;
2726 this->max_grf = brw->gen >= 7 ? GEN7_MRF_HACK_START : BRW_MAX_GRF;
2728 this->current_annotation = NULL;
2729 this->base_ir = NULL;
2731 this->virtual_grf_sizes = NULL;
2732 this->virtual_grf_count = 0;
2733 this->virtual_grf_array_size = 0;
2734 this->virtual_grf_start = NULL;
2735 this->virtual_grf_end = NULL;
2736 this->live_intervals = NULL;
2738 this->params_remap = NULL;
2739 this->nr_params_remap = 0;
2741 this->force_uncompressed_stack = 0;
2742 this->force_sechalf_stack = 0;
2744 this->spilled_any_registers = false;
2746 memset(&this->param_size, 0, sizeof(this->param_size));
2749 fs_visitor::~fs_visitor()
2751 ralloc_free(this->mem_ctx);
2752 hash_table_dtor(this->variable_ht);