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 "main/uniforms.h"
37 #include "program/prog_parameter.h"
38 #include "program/prog_print.h"
39 #include "program/prog_optimize.h"
40 #include "program/register_allocate.h"
41 #include "program/sampler.h"
42 #include "program/hash_table.h"
43 #include "brw_context.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(), BRW_REGISTER_TYPE_D
);
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]));
446 /* Floating point should be lowered by DIV_TO_MUL_RCP in the compiler. */
447 assert(ir
->type
->is_integer());
448 emit_math(SHADER_OPCODE_INT_QUOTIENT
, this->result
, op
[0], op
[1]);
451 /* Floating point should be lowered by MOD_TO_FRACT in the compiler. */
452 assert(ir
->type
->is_integer());
453 emit_math(SHADER_OPCODE_INT_REMAINDER
, this->result
, op
[0], op
[1]);
457 case ir_binop_greater
:
458 case ir_binop_lequal
:
459 case ir_binop_gequal
:
461 case ir_binop_all_equal
:
462 case ir_binop_nequal
:
463 case ir_binop_any_nequal
:
464 resolve_bool_comparison(ir
->operands
[0], &op
[0]);
465 resolve_bool_comparison(ir
->operands
[1], &op
[1]);
467 emit(CMP(this->result
, op
[0], op
[1],
468 brw_conditional_for_comparison(ir
->operation
)));
471 case ir_binop_logic_xor
:
472 emit(XOR(this->result
, op
[0], op
[1]));
475 case ir_binop_logic_or
:
476 emit(OR(this->result
, op
[0], op
[1]));
479 case ir_binop_logic_and
:
480 emit(AND(this->result
, op
[0], op
[1]));
485 assert(!"not reached: should be handled by brw_fs_channel_expressions");
489 assert(!"not reached: should be handled by lower_noise");
492 case ir_quadop_vector
:
493 assert(!"not reached: should be handled by lower_quadop_vector");
496 case ir_binop_vector_extract
:
497 assert(!"not reached: should be handled by lower_vec_index_to_cond_assign()");
500 case ir_triop_vector_insert
:
501 assert(!"not reached: should be handled by lower_vector_insert()");
505 emit_math(SHADER_OPCODE_SQRT
, this->result
, op
[0]);
509 emit_math(SHADER_OPCODE_RSQ
, this->result
, op
[0]);
512 case ir_unop_bitcast_i2f
:
513 case ir_unop_bitcast_u2f
:
514 op
[0].type
= BRW_REGISTER_TYPE_F
;
515 this->result
= op
[0];
518 case ir_unop_bitcast_f2u
:
519 op
[0].type
= BRW_REGISTER_TYPE_UD
;
520 this->result
= op
[0];
523 case ir_unop_bitcast_f2i
:
524 op
[0].type
= BRW_REGISTER_TYPE_D
;
525 this->result
= op
[0];
531 emit(MOV(this->result
, op
[0]));
535 emit(AND(this->result
, op
[0], fs_reg(1)));
538 temp
= fs_reg(this, glsl_type::int_type
);
539 emit(AND(temp
, op
[0], fs_reg(1)));
540 emit(MOV(this->result
, temp
));
544 emit(CMP(this->result
, op
[0], fs_reg(0.0f
), BRW_CONDITIONAL_NZ
));
547 emit(CMP(this->result
, op
[0], fs_reg(0), BRW_CONDITIONAL_NZ
));
551 emit(RNDZ(this->result
, op
[0]));
554 op
[0].negate
= !op
[0].negate
;
555 emit(RNDD(this->result
, op
[0]));
556 this->result
.negate
= true;
559 emit(RNDD(this->result
, op
[0]));
562 emit(FRC(this->result
, op
[0]));
564 case ir_unop_round_even
:
565 emit(RNDE(this->result
, op
[0]));
570 resolve_ud_negate(&op
[0]);
571 resolve_ud_negate(&op
[1]);
572 emit_minmax(ir
->operation
== ir_binop_min
?
573 BRW_CONDITIONAL_L
: BRW_CONDITIONAL_GE
,
574 this->result
, op
[0], op
[1]);
576 case ir_unop_pack_snorm_2x16
:
577 case ir_unop_pack_snorm_4x8
:
578 case ir_unop_pack_unorm_2x16
:
579 case ir_unop_pack_unorm_4x8
:
580 case ir_unop_unpack_snorm_2x16
:
581 case ir_unop_unpack_snorm_4x8
:
582 case ir_unop_unpack_unorm_2x16
:
583 case ir_unop_unpack_unorm_4x8
:
584 case ir_unop_unpack_half_2x16
:
585 case ir_unop_pack_half_2x16
:
586 assert(!"not reached: should be handled by lower_packing_builtins");
588 case ir_unop_unpack_half_2x16_split_x
:
589 emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X
, this->result
, op
[0]);
591 case ir_unop_unpack_half_2x16_split_y
:
592 emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y
, this->result
, op
[0]);
595 emit_math(SHADER_OPCODE_POW
, this->result
, op
[0], op
[1]);
598 case ir_unop_bitfield_reverse
:
599 emit(BFREV(this->result
, op
[0]));
601 case ir_unop_bit_count
:
602 emit(CBIT(this->result
, op
[0]));
604 case ir_unop_find_msb
:
605 temp
= fs_reg(this, glsl_type::uint_type
);
606 emit(FBH(temp
, op
[0]));
608 /* FBH counts from the MSB side, while GLSL's findMSB() wants the count
609 * from the LSB side. If FBH didn't return an error (0xFFFFFFFF), then
610 * subtract the result from 31 to convert the MSB count into an LSB count.
613 /* FBH only supports UD type for dst, so use a MOV to convert UD to D. */
614 emit(MOV(this->result
, temp
));
615 emit(CMP(reg_null_d
, this->result
, fs_reg(-1), BRW_CONDITIONAL_NZ
));
618 inst
= emit(ADD(this->result
, temp
, fs_reg(31)));
619 inst
->predicate
= BRW_PREDICATE_NORMAL
;
621 case ir_unop_find_lsb
:
622 emit(FBL(this->result
, op
[0]));
624 case ir_triop_bitfield_extract
:
625 /* Note that the instruction's argument order is reversed from GLSL
628 emit(BFE(this->result
, op
[2], op
[1], op
[0]));
631 emit(BFI1(this->result
, op
[0], op
[1]));
634 emit(BFI2(this->result
, op
[0], op
[1], op
[2]));
636 case ir_quadop_bitfield_insert
:
637 assert(!"not reached: should be handled by "
638 "lower_instructions::bitfield_insert_to_bfm_bfi");
641 case ir_unop_bit_not
:
642 emit(NOT(this->result
, op
[0]));
644 case ir_binop_bit_and
:
645 emit(AND(this->result
, op
[0], op
[1]));
647 case ir_binop_bit_xor
:
648 emit(XOR(this->result
, op
[0], op
[1]));
650 case ir_binop_bit_or
:
651 emit(OR(this->result
, op
[0], op
[1]));
654 case ir_binop_lshift
:
655 emit(SHL(this->result
, op
[0], op
[1]));
658 case ir_binop_rshift
:
659 if (ir
->type
->base_type
== GLSL_TYPE_INT
)
660 emit(ASR(this->result
, op
[0], op
[1]));
662 emit(SHR(this->result
, op
[0], op
[1]));
664 case ir_binop_pack_half_2x16_split
:
665 emit(FS_OPCODE_PACK_HALF_2x16_SPLIT
, this->result
, op
[0], op
[1]);
667 case ir_binop_ubo_load
: {
668 /* This IR node takes a constant uniform block and a constant or
669 * variable byte offset within the block and loads a vector from that.
671 ir_constant
*uniform_block
= ir
->operands
[0]->as_constant();
672 ir_constant
*const_offset
= ir
->operands
[1]->as_constant();
673 fs_reg surf_index
= fs_reg((unsigned)SURF_INDEX_WM_UBO(uniform_block
->value
.u
[0]));
675 fs_reg packed_consts
= fs_reg(this, glsl_type::float_type
);
676 packed_consts
.type
= result
.type
;
678 fs_reg const_offset_reg
= fs_reg(const_offset
->value
.u
[0] & ~15);
679 emit(fs_inst(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
,
680 packed_consts
, surf_index
, const_offset_reg
));
682 packed_consts
.smear
= const_offset
->value
.u
[0] % 16 / 4;
683 for (int i
= 0; i
< ir
->type
->vector_elements
; i
++) {
684 /* UBO bools are any nonzero value. We consider bools to be
685 * values with the low bit set to 1. Convert them using CMP.
687 if (ir
->type
->base_type
== GLSL_TYPE_BOOL
) {
688 emit(CMP(result
, packed_consts
, fs_reg(0u), BRW_CONDITIONAL_NZ
));
690 emit(MOV(result
, packed_consts
));
693 packed_consts
.smear
++;
696 /* The std140 packing rules don't allow vectors to cross 16-byte
697 * boundaries, and a reg is 32 bytes.
699 assert(packed_consts
.smear
< 8);
702 /* Turn the byte offset into a dword offset. */
703 fs_reg base_offset
= fs_reg(this, glsl_type::int_type
);
704 emit(SHR(base_offset
, op
[1], fs_reg(2)));
706 for (int i
= 0; i
< ir
->type
->vector_elements
; i
++) {
707 emit(VARYING_PULL_CONSTANT_LOAD(result
, surf_index
,
710 if (ir
->type
->base_type
== GLSL_TYPE_BOOL
)
711 emit(CMP(result
, result
, fs_reg(0), BRW_CONDITIONAL_NZ
));
717 result
.reg_offset
= 0;
722 /* Note that the instruction's argument order is reversed from GLSL
725 emit(MAD(this->result
, op
[2], op
[1], op
[0]));
729 emit_lrp(this->result
, op
[0], op
[1], op
[2]);
735 fs_visitor::emit_assignment_writes(fs_reg
&l
, fs_reg
&r
,
736 const glsl_type
*type
, bool predicated
)
738 switch (type
->base_type
) {
739 case GLSL_TYPE_FLOAT
:
743 for (unsigned int i
= 0; i
< type
->components(); i
++) {
744 l
.type
= brw_type_for_base_type(type
);
745 r
.type
= brw_type_for_base_type(type
);
747 if (predicated
|| !l
.equals(r
)) {
748 fs_inst
*inst
= emit(MOV(l
, r
));
749 inst
->predicate
= predicated
? BRW_PREDICATE_NORMAL
: BRW_PREDICATE_NONE
;
756 case GLSL_TYPE_ARRAY
:
757 for (unsigned int i
= 0; i
< type
->length
; i
++) {
758 emit_assignment_writes(l
, r
, type
->fields
.array
, predicated
);
762 case GLSL_TYPE_STRUCT
:
763 for (unsigned int i
= 0; i
< type
->length
; i
++) {
764 emit_assignment_writes(l
, r
, type
->fields
.structure
[i
].type
,
769 case GLSL_TYPE_SAMPLER
:
773 case GLSL_TYPE_ERROR
:
774 case GLSL_TYPE_INTERFACE
:
775 assert(!"not reached");
780 /* If the RHS processing resulted in an instruction generating a
781 * temporary value, and it would be easy to rewrite the instruction to
782 * generate its result right into the LHS instead, do so. This ends
783 * up reliably removing instructions where it can be tricky to do so
784 * later without real UD chain information.
787 fs_visitor::try_rewrite_rhs_to_dst(ir_assignment
*ir
,
790 fs_inst
*pre_rhs_inst
,
791 fs_inst
*last_rhs_inst
)
793 /* Only attempt if we're doing a direct assignment. */
795 !(ir
->lhs
->type
->is_scalar() ||
796 (ir
->lhs
->type
->is_vector() &&
797 ir
->write_mask
== (1 << ir
->lhs
->type
->vector_elements
) - 1)))
800 /* Make sure the last instruction generated our source reg. */
801 fs_inst
*modify
= get_instruction_generating_reg(pre_rhs_inst
,
807 /* If last_rhs_inst wrote a different number of components than our LHS,
808 * we can't safely rewrite it.
810 if (virtual_grf_sizes
[dst
.reg
] != modify
->regs_written
)
813 /* Success! Rewrite the instruction. */
820 fs_visitor::visit(ir_assignment
*ir
)
825 /* FINISHME: arrays on the lhs */
826 ir
->lhs
->accept(this);
829 fs_inst
*pre_rhs_inst
= (fs_inst
*) this->instructions
.get_tail();
831 ir
->rhs
->accept(this);
834 fs_inst
*last_rhs_inst
= (fs_inst
*) this->instructions
.get_tail();
836 assert(l
.file
!= BAD_FILE
);
837 assert(r
.file
!= BAD_FILE
);
839 if (try_rewrite_rhs_to_dst(ir
, l
, r
, pre_rhs_inst
, last_rhs_inst
))
843 emit_bool_to_cond_code(ir
->condition
);
846 if (ir
->lhs
->type
->is_scalar() ||
847 ir
->lhs
->type
->is_vector()) {
848 for (int i
= 0; i
< ir
->lhs
->type
->vector_elements
; i
++) {
849 if (ir
->write_mask
& (1 << i
)) {
850 inst
= emit(MOV(l
, r
));
852 inst
->predicate
= BRW_PREDICATE_NORMAL
;
858 emit_assignment_writes(l
, r
, ir
->lhs
->type
, ir
->condition
!= NULL
);
863 fs_visitor::emit_texture_gen4(ir_texture
*ir
, fs_reg dst
, fs_reg coordinate
,
864 fs_reg shadow_c
, fs_reg lod
, fs_reg dPdy
)
874 if (ir
->shadow_comparitor
) {
875 for (int i
= 0; i
< ir
->coordinate
->type
->vector_elements
; i
++) {
876 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
+ i
), coordinate
));
877 coordinate
.reg_offset
++;
880 /* gen4's SIMD8 sampler always has the slots for u,v,r present.
881 * the unused slots must be zeroed.
883 for (int i
= ir
->coordinate
->type
->vector_elements
; i
< 3; i
++) {
884 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
+ i
), fs_reg(0.0f
)));
888 if (ir
->op
== ir_tex
) {
889 /* There's no plain shadow compare message, so we use shadow
890 * compare with a bias of 0.0.
892 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
), fs_reg(0.0f
)));
894 } else if (ir
->op
== ir_txb
|| ir
->op
== ir_txl
) {
895 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
), lod
));
898 assert(!"Should not get here.");
901 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
), shadow_c
));
903 } else if (ir
->op
== ir_tex
) {
904 for (int i
= 0; i
< ir
->coordinate
->type
->vector_elements
; i
++) {
905 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
+ i
), coordinate
));
906 coordinate
.reg_offset
++;
908 /* zero the others. */
909 for (int i
= ir
->coordinate
->type
->vector_elements
; i
<3; i
++) {
910 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
+ i
), fs_reg(0.0f
)));
912 /* gen4's SIMD8 sampler always has the slots for u,v,r present. */
914 } else if (ir
->op
== ir_txd
) {
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
++;
921 /* the slots for u and v are always present, but r is optional */
922 mlen
+= MAX2(ir
->coordinate
->type
->vector_elements
, 2);
925 * dPdx = dudx, dvdx, drdx
926 * dPdy = dudy, dvdy, drdy
928 * 1-arg: Does not exist.
930 * 2-arg: dudx dvdx dudy dvdy
931 * dPdx.x dPdx.y dPdy.x dPdy.y
934 * 3-arg: dudx dvdx drdx dudy dvdy drdy
935 * dPdx.x dPdx.y dPdx.z dPdy.x dPdy.y dPdy.z
938 for (int i
= 0; i
< ir
->lod_info
.grad
.dPdx
->type
->vector_elements
; i
++) {
939 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
), dPdx
));
942 mlen
+= MAX2(ir
->lod_info
.grad
.dPdx
->type
->vector_elements
, 2);
944 for (int i
= 0; i
< ir
->lod_info
.grad
.dPdy
->type
->vector_elements
; i
++) {
945 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
), dPdy
));
948 mlen
+= MAX2(ir
->lod_info
.grad
.dPdy
->type
->vector_elements
, 2);
949 } else if (ir
->op
== ir_txs
) {
950 /* There's no SIMD8 resinfo message on Gen4. Use SIMD16 instead. */
952 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
, BRW_REGISTER_TYPE_UD
), lod
));
955 /* Oh joy. gen4 doesn't have SIMD8 non-shadow-compare bias/lod
956 * instructions. We'll need to do SIMD16 here.
959 assert(ir
->op
== ir_txb
|| ir
->op
== ir_txl
|| ir
->op
== ir_txf
);
961 for (int i
= 0; i
< ir
->coordinate
->type
->vector_elements
; i
++) {
962 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
+ i
* 2, coordinate
.type
),
964 coordinate
.reg_offset
++;
967 /* Initialize the rest of u/v/r with 0.0. Empirically, this seems to
968 * be necessary for TXF (ld), but seems wise to do for all messages.
970 for (int i
= ir
->coordinate
->type
->vector_elements
; i
< 3; i
++) {
971 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
+ i
* 2), fs_reg(0.0f
)));
974 /* lod/bias appears after u/v/r. */
977 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
, lod
.type
), lod
));
980 /* The unused upper half. */
985 /* Now, since we're doing simd16, the return is 2 interleaved
986 * vec4s where the odd-indexed ones are junk. We'll need to move
987 * this weirdness around to the expected layout.
990 dst
= fs_reg(GRF
, virtual_grf_alloc(8),
992 brw_type_for_base_type(ir
->type
) :
993 BRW_REGISTER_TYPE_F
));
996 fs_inst
*inst
= NULL
;
999 inst
= emit(SHADER_OPCODE_TEX
, dst
);
1002 inst
= emit(FS_OPCODE_TXB
, dst
);
1005 inst
= emit(SHADER_OPCODE_TXL
, dst
);
1008 inst
= emit(SHADER_OPCODE_TXD
, dst
);
1011 inst
= emit(SHADER_OPCODE_TXS
, dst
);
1014 inst
= emit(SHADER_OPCODE_TXF
, dst
);
1017 fail("unrecognized texture opcode");
1019 inst
->base_mrf
= base_mrf
;
1021 inst
->header_present
= true;
1022 inst
->regs_written
= simd16
? 8 : 4;
1025 for (int i
= 0; i
< 4; i
++) {
1026 emit(MOV(orig_dst
, dst
));
1027 orig_dst
.reg_offset
++;
1028 dst
.reg_offset
+= 2;
1035 /* gen5's sampler has slots for u, v, r, array index, then optional
1036 * parameters like shadow comparitor or LOD bias. If optional
1037 * parameters aren't present, those base slots are optional and don't
1038 * need to be included in the message.
1040 * We don't fill in the unnecessary slots regardless, which may look
1041 * surprising in the disassembly.
1044 fs_visitor::emit_texture_gen5(ir_texture
*ir
, fs_reg dst
, fs_reg coordinate
,
1045 fs_reg shadow_c
, fs_reg lod
, fs_reg lod2
,
1046 fs_reg sample_index
)
1050 int reg_width
= dispatch_width
/ 8;
1051 bool header_present
= false;
1052 const int vector_elements
=
1053 ir
->coordinate
? ir
->coordinate
->type
->vector_elements
: 0;
1055 if (ir
->offset
!= NULL
&& ir
->op
== ir_txf
) {
1056 /* It appears that the ld instruction used for txf does its
1057 * address bounds check before adding in the offset. To work
1058 * around this, just add the integer offset to the integer texel
1059 * coordinate, and don't put the offset in the header.
1061 ir_constant
*offset
= ir
->offset
->as_constant();
1062 for (int i
= 0; i
< vector_elements
; i
++) {
1063 emit(ADD(fs_reg(MRF
, base_mrf
+ mlen
+ i
* reg_width
, coordinate
.type
),
1065 offset
->value
.i
[i
]));
1066 coordinate
.reg_offset
++;
1070 /* The offsets set up by the ir_texture visitor are in the
1071 * m1 header, so we can't go headerless.
1073 header_present
= true;
1078 for (int i
= 0; i
< vector_elements
; i
++) {
1079 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
+ i
* reg_width
, coordinate
.type
),
1081 coordinate
.reg_offset
++;
1084 mlen
+= vector_elements
* reg_width
;
1086 if (ir
->shadow_comparitor
) {
1087 mlen
= MAX2(mlen
, header_present
+ 4 * reg_width
);
1089 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
), shadow_c
));
1093 fs_inst
*inst
= NULL
;
1096 inst
= emit(SHADER_OPCODE_TEX
, dst
);
1099 mlen
= MAX2(mlen
, header_present
+ 4 * reg_width
);
1100 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
), lod
));
1103 inst
= emit(FS_OPCODE_TXB
, dst
);
1106 mlen
= MAX2(mlen
, header_present
+ 4 * reg_width
);
1107 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
), lod
));
1110 inst
= emit(SHADER_OPCODE_TXL
, dst
);
1113 mlen
= MAX2(mlen
, header_present
+ 4 * reg_width
); /* skip over 'ai' */
1117 * dPdx = dudx, dvdx, drdx
1118 * dPdy = dudy, dvdy, drdy
1120 * Load up these values:
1121 * - dudx dudy dvdx dvdy drdx drdy
1122 * - dPdx.x dPdy.x dPdx.y dPdy.y dPdx.z dPdy.z
1124 for (int i
= 0; i
< ir
->lod_info
.grad
.dPdx
->type
->vector_elements
; i
++) {
1125 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
), lod
));
1129 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
), lod2
));
1134 inst
= emit(SHADER_OPCODE_TXD
, dst
);
1138 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
, BRW_REGISTER_TYPE_UD
), lod
));
1140 inst
= emit(SHADER_OPCODE_TXS
, dst
);
1143 mlen
= header_present
+ 4 * reg_width
;
1144 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
- reg_width
, BRW_REGISTER_TYPE_UD
), lod
));
1145 inst
= emit(SHADER_OPCODE_TXF
, dst
);
1148 mlen
= header_present
+ 4 * reg_width
;
1151 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
- reg_width
, BRW_REGISTER_TYPE_UD
), fs_reg(0)));
1153 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
, BRW_REGISTER_TYPE_UD
), sample_index
));
1155 inst
= emit(SHADER_OPCODE_TXF_MS
, dst
);
1158 inst
= emit(SHADER_OPCODE_LOD
, dst
);
1161 inst
->base_mrf
= base_mrf
;
1163 inst
->header_present
= header_present
;
1164 inst
->regs_written
= 4;
1167 fail("Message length >11 disallowed by hardware\n");
1174 fs_visitor::emit_texture_gen7(ir_texture
*ir
, fs_reg dst
, fs_reg coordinate
,
1175 fs_reg shadow_c
, fs_reg lod
, fs_reg lod2
,
1176 fs_reg sample_index
)
1180 int reg_width
= dispatch_width
/ 8;
1181 bool header_present
= false;
1184 if (ir
->offset
&& ir
->op
!= ir_txf
) {
1185 /* The offsets set up by the ir_texture visitor are in the
1186 * m1 header, so we can't go headerless.
1188 header_present
= true;
1193 if (ir
->shadow_comparitor
) {
1194 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
), shadow_c
));
1198 /* Set up the LOD info */
1204 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
), lod
));
1208 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
), lod
));
1212 if (dispatch_width
== 16)
1213 fail("Gen7 does not support sample_d/sample_d_c in SIMD16 mode.");
1215 /* Load dPdx and the coordinate together:
1216 * [hdr], [ref], x, dPdx.x, dPdy.x, y, dPdx.y, dPdy.y, z, dPdx.z, dPdy.z
1218 for (int i
= 0; i
< ir
->coordinate
->type
->vector_elements
; i
++) {
1219 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
), coordinate
));
1220 coordinate
.reg_offset
++;
1223 /* For cube map array, the coordinate is (u,v,r,ai) but there are
1224 * only derivatives for (u, v, r).
1226 if (i
< ir
->lod_info
.grad
.dPdx
->type
->vector_elements
) {
1227 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
), lod
));
1231 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
), lod2
));
1239 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
, BRW_REGISTER_TYPE_UD
), lod
));
1243 /* It appears that the ld instruction used for txf does its
1244 * address bounds check before adding in the offset. To work
1245 * around this, just add the integer offset to the integer texel
1246 * coordinate, and don't put the offset in the header.
1249 ir_constant
*offset
= ir
->offset
->as_constant();
1250 offsets
[0] = offset
->value
.i
[0];
1251 offsets
[1] = offset
->value
.i
[1];
1252 offsets
[2] = offset
->value
.i
[2];
1254 memset(offsets
, 0, sizeof(offsets
));
1257 /* Unfortunately, the parameters for LD are intermixed: u, lod, v, r. */
1258 emit(ADD(fs_reg(MRF
, base_mrf
+ mlen
, BRW_REGISTER_TYPE_D
),
1259 coordinate
, offsets
[0]));
1260 coordinate
.reg_offset
++;
1263 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
, BRW_REGISTER_TYPE_D
), lod
));
1266 for (int i
= 1; i
< ir
->coordinate
->type
->vector_elements
; i
++) {
1267 emit(ADD(fs_reg(MRF
, base_mrf
+ mlen
, BRW_REGISTER_TYPE_D
),
1268 coordinate
, offsets
[i
]));
1269 coordinate
.reg_offset
++;
1274 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
, BRW_REGISTER_TYPE_UD
), sample_index
));
1277 /* constant zero MCS; we arrange to never actually have a compressed
1278 * multisample surface here for now. TODO: issue ld_mcs to get this first,
1279 * if we ever support texturing from compressed multisample surfaces
1281 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
, BRW_REGISTER_TYPE_UD
), fs_reg(0u)));
1284 /* there is no offsetting for this message; just copy in the integer
1285 * texture coordinates
1287 for (int i
= 0; i
< ir
->coordinate
->type
->vector_elements
; i
++) {
1288 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
, BRW_REGISTER_TYPE_D
),
1290 coordinate
.reg_offset
++;
1296 /* Set up the coordinate (except for cases where it was done above) */
1297 if (ir
->op
!= ir_txd
&& ir
->op
!= ir_txs
&& ir
->op
!= ir_txf
&& ir
->op
!= ir_txf_ms
) {
1298 for (int i
= 0; i
< ir
->coordinate
->type
->vector_elements
; i
++) {
1299 emit(MOV(fs_reg(MRF
, base_mrf
+ mlen
), coordinate
));
1300 coordinate
.reg_offset
++;
1305 /* Generate the SEND */
1306 fs_inst
*inst
= NULL
;
1308 case ir_tex
: inst
= emit(SHADER_OPCODE_TEX
, dst
); break;
1309 case ir_txb
: inst
= emit(FS_OPCODE_TXB
, dst
); break;
1310 case ir_txl
: inst
= emit(SHADER_OPCODE_TXL
, dst
); break;
1311 case ir_txd
: inst
= emit(SHADER_OPCODE_TXD
, dst
); break;
1312 case ir_txf
: inst
= emit(SHADER_OPCODE_TXF
, dst
); break;
1313 case ir_txf_ms
: inst
= emit(SHADER_OPCODE_TXF_MS
, dst
); break;
1314 case ir_txs
: inst
= emit(SHADER_OPCODE_TXS
, dst
); break;
1315 case ir_lod
: inst
= emit(SHADER_OPCODE_LOD
, dst
); break;
1317 inst
->base_mrf
= base_mrf
;
1319 inst
->header_present
= header_present
;
1320 inst
->regs_written
= 4;
1323 fail("Message length >11 disallowed by hardware\n");
1330 fs_visitor::rescale_texcoord(ir_texture
*ir
, fs_reg coordinate
,
1331 bool is_rect
, int sampler
, int texunit
)
1333 fs_inst
*inst
= NULL
;
1334 bool needs_gl_clamp
= true;
1335 fs_reg scale_x
, scale_y
;
1337 /* The 965 requires the EU to do the normalization of GL rectangle
1338 * texture coordinates. We use the program parameter state
1339 * tracking to get the scaling factor.
1343 (brw
->gen
>= 6 && (c
->key
.tex
.gl_clamp_mask
[0] & (1 << sampler
) ||
1344 c
->key
.tex
.gl_clamp_mask
[1] & (1 << sampler
))))) {
1345 struct gl_program_parameter_list
*params
= fp
->Base
.Parameters
;
1346 int tokens
[STATE_LENGTH
] = {
1348 STATE_TEXRECT_SCALE
,
1354 if (dispatch_width
== 16) {
1355 fail("rectangle scale uniform setup not supported on 16-wide\n");
1359 scale_x
= fs_reg(UNIFORM
, c
->prog_data
.nr_params
);
1360 scale_y
= fs_reg(UNIFORM
, c
->prog_data
.nr_params
+ 1);
1362 GLuint index
= _mesa_add_state_reference(params
,
1363 (gl_state_index
*)tokens
);
1364 c
->prog_data
.param
[c
->prog_data
.nr_params
++] =
1365 &fp
->Base
.Parameters
->ParameterValues
[index
][0].f
;
1366 c
->prog_data
.param
[c
->prog_data
.nr_params
++] =
1367 &fp
->Base
.Parameters
->ParameterValues
[index
][1].f
;
1370 /* The 965 requires the EU to do the normalization of GL rectangle
1371 * texture coordinates. We use the program parameter state
1372 * tracking to get the scaling factor.
1374 if (brw
->gen
< 6 && is_rect
) {
1375 fs_reg dst
= fs_reg(this, ir
->coordinate
->type
);
1376 fs_reg src
= coordinate
;
1379 emit(MUL(dst
, src
, scale_x
));
1382 emit(MUL(dst
, src
, scale_y
));
1383 } else if (is_rect
) {
1384 /* On gen6+, the sampler handles the rectangle coordinates
1385 * natively, without needing rescaling. But that means we have
1386 * to do GL_CLAMP clamping at the [0, width], [0, height] scale,
1387 * not [0, 1] like the default case below.
1389 needs_gl_clamp
= false;
1391 for (int i
= 0; i
< 2; i
++) {
1392 if (c
->key
.tex
.gl_clamp_mask
[i
] & (1 << sampler
)) {
1393 fs_reg chan
= coordinate
;
1394 chan
.reg_offset
+= i
;
1396 inst
= emit(BRW_OPCODE_SEL
, chan
, chan
, brw_imm_f(0.0));
1397 inst
->conditional_mod
= BRW_CONDITIONAL_G
;
1399 /* Our parameter comes in as 1.0/width or 1.0/height,
1400 * because that's what people normally want for doing
1401 * texture rectangle handling. We need width or height
1402 * for clamping, but we don't care enough to make a new
1403 * parameter type, so just invert back.
1405 fs_reg limit
= fs_reg(this, glsl_type::float_type
);
1406 emit(MOV(limit
, i
== 0 ? scale_x
: scale_y
));
1407 emit(SHADER_OPCODE_RCP
, limit
, limit
);
1409 inst
= emit(BRW_OPCODE_SEL
, chan
, chan
, limit
);
1410 inst
->conditional_mod
= BRW_CONDITIONAL_L
;
1415 if (ir
->coordinate
&& needs_gl_clamp
) {
1416 for (unsigned int i
= 0;
1417 i
< MIN2(ir
->coordinate
->type
->vector_elements
, 3); i
++) {
1418 if (c
->key
.tex
.gl_clamp_mask
[i
] & (1 << sampler
)) {
1419 fs_reg chan
= coordinate
;
1420 chan
.reg_offset
+= i
;
1422 fs_inst
*inst
= emit(MOV(chan
, chan
));
1423 inst
->saturate
= true;
1431 fs_visitor::visit(ir_texture
*ir
)
1433 fs_inst
*inst
= NULL
;
1436 _mesa_get_sampler_uniform_value(ir
->sampler
, shader_prog
, &fp
->Base
);
1437 /* FINISHME: We're failing to recompile our programs when the sampler is
1438 * updated. This only matters for the texture rectangle scale parameters
1439 * (pre-gen6, or gen6+ with GL_CLAMP).
1441 int texunit
= fp
->Base
.SamplerUnits
[sampler
];
1443 /* Should be lowered by do_lower_texture_projection */
1444 assert(!ir
->projector
);
1446 /* Generate code to compute all the subexpression trees. This has to be
1447 * done before loading any values into MRFs for the sampler message since
1448 * generating these values may involve SEND messages that need the MRFs.
1451 if (ir
->coordinate
) {
1452 ir
->coordinate
->accept(this);
1454 coordinate
= rescale_texcoord(ir
, this->result
,
1455 ir
->sampler
->type
->sampler_dimensionality
==
1456 GLSL_SAMPLER_DIM_RECT
,
1460 fs_reg shadow_comparitor
;
1461 if (ir
->shadow_comparitor
) {
1462 ir
->shadow_comparitor
->accept(this);
1463 shadow_comparitor
= this->result
;
1466 fs_reg lod
, lod2
, sample_index
;
1472 ir
->lod_info
.bias
->accept(this);
1476 ir
->lod_info
.grad
.dPdx
->accept(this);
1479 ir
->lod_info
.grad
.dPdy
->accept(this);
1480 lod2
= this->result
;
1485 ir
->lod_info
.lod
->accept(this);
1489 ir
->lod_info
.sample_index
->accept(this);
1490 sample_index
= this->result
;
1494 /* Writemasking doesn't eliminate channels on SIMD8 texture
1495 * samples, so don't worry about them.
1497 fs_reg dst
= fs_reg(this, glsl_type::get_instance(ir
->type
->base_type
, 4, 1));
1499 if (brw
->gen
>= 7) {
1500 inst
= emit_texture_gen7(ir
, dst
, coordinate
, shadow_comparitor
,
1501 lod
, lod2
, sample_index
);
1502 } else if (brw
->gen
>= 5) {
1503 inst
= emit_texture_gen5(ir
, dst
, coordinate
, shadow_comparitor
,
1504 lod
, lod2
, sample_index
);
1506 inst
= emit_texture_gen4(ir
, dst
, coordinate
, shadow_comparitor
,
1510 /* The header is set up by generate_tex() when necessary. */
1511 inst
->src
[0] = reg_undef
;
1513 if (ir
->offset
!= NULL
&& ir
->op
!= ir_txf
)
1514 inst
->texture_offset
= brw_texture_offset(ir
->offset
->as_constant());
1516 inst
->sampler
= sampler
;
1518 if (ir
->shadow_comparitor
)
1519 inst
->shadow_compare
= true;
1521 /* fixup #layers for cube map arrays */
1522 if (ir
->op
== ir_txs
) {
1523 glsl_type
const *type
= ir
->sampler
->type
;
1524 if (type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_CUBE
&&
1525 type
->sampler_array
) {
1527 depth
.reg_offset
= 2;
1528 emit_math(SHADER_OPCODE_INT_QUOTIENT
, depth
, depth
, fs_reg(6));
1532 swizzle_result(ir
, dst
, sampler
);
1536 * Swizzle the result of a texture result. This is necessary for
1537 * EXT_texture_swizzle as well as DEPTH_TEXTURE_MODE for shadow comparisons.
1540 fs_visitor::swizzle_result(ir_texture
*ir
, fs_reg orig_val
, int sampler
)
1542 this->result
= orig_val
;
1544 if (ir
->op
== ir_txs
|| ir
->op
== ir_lod
)
1547 if (ir
->type
== glsl_type::float_type
) {
1548 /* Ignore DEPTH_TEXTURE_MODE swizzling. */
1549 assert(ir
->sampler
->type
->sampler_shadow
);
1550 } else if (c
->key
.tex
.swizzles
[sampler
] != SWIZZLE_NOOP
) {
1551 fs_reg swizzled_result
= fs_reg(this, glsl_type::vec4_type
);
1553 for (int i
= 0; i
< 4; i
++) {
1554 int swiz
= GET_SWZ(c
->key
.tex
.swizzles
[sampler
], i
);
1555 fs_reg l
= swizzled_result
;
1558 if (swiz
== SWIZZLE_ZERO
) {
1559 emit(MOV(l
, fs_reg(0.0f
)));
1560 } else if (swiz
== SWIZZLE_ONE
) {
1561 emit(MOV(l
, fs_reg(1.0f
)));
1563 fs_reg r
= orig_val
;
1564 r
.reg_offset
+= GET_SWZ(c
->key
.tex
.swizzles
[sampler
], i
);
1568 this->result
= swizzled_result
;
1573 fs_visitor::visit(ir_swizzle
*ir
)
1575 ir
->val
->accept(this);
1576 fs_reg val
= this->result
;
1578 if (ir
->type
->vector_elements
== 1) {
1579 this->result
.reg_offset
+= ir
->mask
.x
;
1583 fs_reg result
= fs_reg(this, ir
->type
);
1584 this->result
= result
;
1586 for (unsigned int i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1587 fs_reg channel
= val
;
1605 channel
.reg_offset
+= swiz
;
1606 emit(MOV(result
, channel
));
1607 result
.reg_offset
++;
1612 fs_visitor::visit(ir_discard
*ir
)
1614 assert(ir
->condition
== NULL
); /* FINISHME */
1616 /* We track our discarded pixels in f0.1. By predicating on it, we can
1617 * update just the flag bits that aren't yet discarded. By emitting a
1618 * CMP of g0 != g0, all our currently executing channels will get turned
1621 fs_reg some_reg
= fs_reg(retype(brw_vec8_grf(0, 0),
1622 BRW_REGISTER_TYPE_UW
));
1623 fs_inst
*cmp
= emit(CMP(reg_null_f
, some_reg
, some_reg
,
1624 BRW_CONDITIONAL_NZ
));
1625 cmp
->predicate
= BRW_PREDICATE_NORMAL
;
1626 cmp
->flag_subreg
= 1;
1628 if (brw
->gen
>= 6) {
1629 /* For performance, after a discard, jump to the end of the shader.
1630 * However, many people will do foliage by discarding based on a
1631 * texture's alpha mask, and then continue on to texture with the
1632 * remaining pixels. To avoid trashing the derivatives for those
1633 * texture samples, we'll only jump if all of the pixels in the subspan
1634 * have been discarded.
1636 fs_inst
*discard_jump
= emit(FS_OPCODE_DISCARD_JUMP
);
1637 discard_jump
->flag_subreg
= 1;
1638 discard_jump
->predicate
= BRW_PREDICATE_ALIGN1_ANY4H
;
1639 discard_jump
->predicate_inverse
= true;
1644 fs_visitor::visit(ir_constant
*ir
)
1646 /* Set this->result to reg at the bottom of the function because some code
1647 * paths will cause this visitor to be applied to other fields. This will
1648 * cause the value stored in this->result to be modified.
1650 * Make reg constant so that it doesn't get accidentally modified along the
1651 * way. Yes, I actually had this problem. :(
1653 const fs_reg
reg(this, ir
->type
);
1654 fs_reg dst_reg
= reg
;
1656 if (ir
->type
->is_array()) {
1657 const unsigned size
= type_size(ir
->type
->fields
.array
);
1659 for (unsigned i
= 0; i
< ir
->type
->length
; i
++) {
1660 ir
->array_elements
[i
]->accept(this);
1661 fs_reg src_reg
= this->result
;
1663 dst_reg
.type
= src_reg
.type
;
1664 for (unsigned j
= 0; j
< size
; j
++) {
1665 emit(MOV(dst_reg
, src_reg
));
1666 src_reg
.reg_offset
++;
1667 dst_reg
.reg_offset
++;
1670 } else if (ir
->type
->is_record()) {
1671 foreach_list(node
, &ir
->components
) {
1672 ir_constant
*const field
= (ir_constant
*) node
;
1673 const unsigned size
= type_size(field
->type
);
1675 field
->accept(this);
1676 fs_reg src_reg
= this->result
;
1678 dst_reg
.type
= src_reg
.type
;
1679 for (unsigned j
= 0; j
< size
; j
++) {
1680 emit(MOV(dst_reg
, src_reg
));
1681 src_reg
.reg_offset
++;
1682 dst_reg
.reg_offset
++;
1686 const unsigned size
= type_size(ir
->type
);
1688 for (unsigned i
= 0; i
< size
; i
++) {
1689 switch (ir
->type
->base_type
) {
1690 case GLSL_TYPE_FLOAT
:
1691 emit(MOV(dst_reg
, fs_reg(ir
->value
.f
[i
])));
1693 case GLSL_TYPE_UINT
:
1694 emit(MOV(dst_reg
, fs_reg(ir
->value
.u
[i
])));
1697 emit(MOV(dst_reg
, fs_reg(ir
->value
.i
[i
])));
1699 case GLSL_TYPE_BOOL
:
1700 emit(MOV(dst_reg
, fs_reg((int)ir
->value
.b
[i
])));
1703 assert(!"Non-float/uint/int/bool constant");
1705 dst_reg
.reg_offset
++;
1713 fs_visitor::emit_bool_to_cond_code(ir_rvalue
*ir
)
1715 ir_expression
*expr
= ir
->as_expression();
1721 assert(expr
->get_num_operands() <= 2);
1722 for (unsigned int i
= 0; i
< expr
->get_num_operands(); i
++) {
1723 assert(expr
->operands
[i
]->type
->is_scalar());
1725 expr
->operands
[i
]->accept(this);
1726 op
[i
] = this->result
;
1728 resolve_ud_negate(&op
[i
]);
1731 switch (expr
->operation
) {
1732 case ir_unop_logic_not
:
1733 inst
= emit(AND(reg_null_d
, op
[0], fs_reg(1)));
1734 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
1737 case ir_binop_logic_xor
:
1738 case ir_binop_logic_or
:
1739 case ir_binop_logic_and
:
1743 if (brw
->gen
>= 6) {
1744 emit(CMP(reg_null_d
, op
[0], fs_reg(0.0f
), BRW_CONDITIONAL_NZ
));
1746 inst
= emit(MOV(reg_null_f
, op
[0]));
1747 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
1752 if (brw
->gen
>= 6) {
1753 emit(CMP(reg_null_d
, op
[0], fs_reg(0), BRW_CONDITIONAL_NZ
));
1755 inst
= emit(MOV(reg_null_d
, op
[0]));
1756 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
1760 case ir_binop_greater
:
1761 case ir_binop_gequal
:
1763 case ir_binop_lequal
:
1764 case ir_binop_equal
:
1765 case ir_binop_all_equal
:
1766 case ir_binop_nequal
:
1767 case ir_binop_any_nequal
:
1768 resolve_bool_comparison(expr
->operands
[0], &op
[0]);
1769 resolve_bool_comparison(expr
->operands
[1], &op
[1]);
1771 emit(CMP(reg_null_d
, op
[0], op
[1],
1772 brw_conditional_for_comparison(expr
->operation
)));
1776 assert(!"not reached");
1777 fail("bad cond code\n");
1786 fs_inst
*inst
= emit(AND(reg_null_d
, this->result
, fs_reg(1)));
1787 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
1791 * Emit a gen6 IF statement with the comparison folded into the IF
1795 fs_visitor::emit_if_gen6(ir_if
*ir
)
1797 ir_expression
*expr
= ir
->condition
->as_expression();
1804 assert(expr
->get_num_operands() <= 2);
1805 for (unsigned int i
= 0; i
< expr
->get_num_operands(); i
++) {
1806 assert(expr
->operands
[i
]->type
->is_scalar());
1808 expr
->operands
[i
]->accept(this);
1809 op
[i
] = this->result
;
1812 switch (expr
->operation
) {
1813 case ir_unop_logic_not
:
1814 case ir_binop_logic_xor
:
1815 case ir_binop_logic_or
:
1816 case ir_binop_logic_and
:
1817 /* For operations on bool arguments, only the low bit of the bool is
1818 * valid, and the others are undefined. Fall back to the condition
1824 inst
= emit(BRW_OPCODE_IF
, reg_null_f
, op
[0], fs_reg(0));
1825 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
1829 emit(IF(op
[0], fs_reg(0), BRW_CONDITIONAL_NZ
));
1832 case ir_binop_greater
:
1833 case ir_binop_gequal
:
1835 case ir_binop_lequal
:
1836 case ir_binop_equal
:
1837 case ir_binop_all_equal
:
1838 case ir_binop_nequal
:
1839 case ir_binop_any_nequal
:
1840 resolve_bool_comparison(expr
->operands
[0], &op
[0]);
1841 resolve_bool_comparison(expr
->operands
[1], &op
[1]);
1843 emit(IF(op
[0], op
[1],
1844 brw_conditional_for_comparison(expr
->operation
)));
1847 assert(!"not reached");
1848 emit(IF(op
[0], fs_reg(0), BRW_CONDITIONAL_NZ
));
1849 fail("bad condition\n");
1854 emit_bool_to_cond_code(ir
->condition
);
1855 fs_inst
*inst
= emit(BRW_OPCODE_IF
);
1856 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1860 * Try to replace IF/MOV/ELSE/MOV/ENDIF with SEL.
1862 * Many GLSL shaders contain the following pattern:
1864 * x = condition ? foo : bar
1866 * The compiler emits an ir_if tree for this, since each subexpression might be
1867 * a complex tree that could have side-effects or short-circuit logic.
1869 * However, the common case is to simply select one of two constants or
1870 * variable values---which is exactly what SEL is for. In this case, the
1871 * assembly looks like:
1879 * which can be easily translated into:
1881 * (+f0) SEL dst src0 src1
1883 * If src0 is an immediate value, we promote it to a temporary GRF.
1886 fs_visitor::try_replace_with_sel()
1888 fs_inst
*endif_inst
= (fs_inst
*) instructions
.get_tail();
1889 assert(endif_inst
->opcode
== BRW_OPCODE_ENDIF
);
1891 /* Pattern match in reverse: IF, MOV, ELSE, MOV, ENDIF. */
1893 BRW_OPCODE_IF
, BRW_OPCODE_MOV
, BRW_OPCODE_ELSE
, BRW_OPCODE_MOV
,
1896 fs_inst
*match
= (fs_inst
*) endif_inst
->prev
;
1897 for (int i
= 0; i
< 4; i
++) {
1898 if (match
->is_head_sentinel() || match
->opcode
!= opcodes
[4-i
-1])
1900 match
= (fs_inst
*) match
->prev
;
1903 /* The opcodes match; it looks like the right sequence of instructions. */
1904 fs_inst
*else_mov
= (fs_inst
*) endif_inst
->prev
;
1905 fs_inst
*then_mov
= (fs_inst
*) else_mov
->prev
->prev
;
1906 fs_inst
*if_inst
= (fs_inst
*) then_mov
->prev
;
1908 /* Check that the MOVs are the right form. */
1909 if (then_mov
->dst
.equals(else_mov
->dst
) &&
1910 !then_mov
->is_partial_write() &&
1911 !else_mov
->is_partial_write()) {
1913 /* Remove the matched instructions; we'll emit a SEL to replace them. */
1914 while (!if_inst
->next
->is_tail_sentinel())
1915 if_inst
->next
->remove();
1918 /* Only the last source register can be a constant, so if the MOV in
1919 * the "then" clause uses a constant, we need to put it in a temporary.
1921 fs_reg
src0(then_mov
->src
[0]);
1922 if (src0
.file
== IMM
) {
1923 src0
= fs_reg(this, glsl_type::float_type
);
1924 src0
.type
= then_mov
->src
[0].type
;
1925 emit(MOV(src0
, then_mov
->src
[0]));
1929 if (if_inst
->conditional_mod
) {
1930 /* Sandybridge-specific IF with embedded comparison */
1931 emit(CMP(reg_null_d
, if_inst
->src
[0], if_inst
->src
[1],
1932 if_inst
->conditional_mod
));
1933 sel
= emit(BRW_OPCODE_SEL
, then_mov
->dst
, src0
, else_mov
->src
[0]);
1934 sel
->predicate
= BRW_PREDICATE_NORMAL
;
1936 /* Separate CMP and IF instructions */
1937 sel
= emit(BRW_OPCODE_SEL
, then_mov
->dst
, src0
, else_mov
->src
[0]);
1938 sel
->predicate
= if_inst
->predicate
;
1939 sel
->predicate_inverse
= if_inst
->predicate_inverse
;
1945 fs_visitor::visit(ir_if
*ir
)
1947 if (brw
->gen
< 6 && dispatch_width
== 16) {
1948 fail("Can't support (non-uniform) control flow on 16-wide\n");
1951 /* Don't point the annotation at the if statement, because then it plus
1952 * the then and else blocks get printed.
1954 this->base_ir
= ir
->condition
;
1956 if (brw
->gen
== 6) {
1959 emit_bool_to_cond_code(ir
->condition
);
1961 emit(IF(BRW_PREDICATE_NORMAL
));
1964 foreach_list(node
, &ir
->then_instructions
) {
1965 ir_instruction
*ir
= (ir_instruction
*)node
;
1971 if (!ir
->else_instructions
.is_empty()) {
1972 emit(BRW_OPCODE_ELSE
);
1974 foreach_list(node
, &ir
->else_instructions
) {
1975 ir_instruction
*ir
= (ir_instruction
*)node
;
1982 emit(BRW_OPCODE_ENDIF
);
1984 try_replace_with_sel();
1988 fs_visitor::visit(ir_loop
*ir
)
1990 fs_reg counter
= reg_undef
;
1992 if (brw
->gen
< 6 && dispatch_width
== 16) {
1993 fail("Can't support (non-uniform) control flow on 16-wide\n");
1997 this->base_ir
= ir
->counter
;
1998 ir
->counter
->accept(this);
1999 counter
= *(variable_storage(ir
->counter
));
2002 this->base_ir
= ir
->from
;
2003 ir
->from
->accept(this);
2005 emit(MOV(counter
, this->result
));
2009 this->base_ir
= NULL
;
2010 emit(BRW_OPCODE_DO
);
2013 this->base_ir
= ir
->to
;
2014 ir
->to
->accept(this);
2016 emit(CMP(reg_null_d
, counter
, this->result
,
2017 brw_conditional_for_comparison(ir
->cmp
)));
2019 fs_inst
*inst
= emit(BRW_OPCODE_BREAK
);
2020 inst
->predicate
= BRW_PREDICATE_NORMAL
;
2023 foreach_list(node
, &ir
->body_instructions
) {
2024 ir_instruction
*ir
= (ir_instruction
*)node
;
2030 if (ir
->increment
) {
2031 this->base_ir
= ir
->increment
;
2032 ir
->increment
->accept(this);
2033 emit(ADD(counter
, counter
, this->result
));
2036 this->base_ir
= NULL
;
2037 emit(BRW_OPCODE_WHILE
);
2041 fs_visitor::visit(ir_loop_jump
*ir
)
2044 case ir_loop_jump::jump_break
:
2045 emit(BRW_OPCODE_BREAK
);
2047 case ir_loop_jump::jump_continue
:
2048 emit(BRW_OPCODE_CONTINUE
);
2054 fs_visitor::visit(ir_call
*ir
)
2056 assert(!"FINISHME");
2060 fs_visitor::visit(ir_return
*ir
)
2062 assert(!"FINISHME");
2066 fs_visitor::visit(ir_function
*ir
)
2068 /* Ignore function bodies other than main() -- we shouldn't see calls to
2069 * them since they should all be inlined before we get to ir_to_mesa.
2071 if (strcmp(ir
->name
, "main") == 0) {
2072 const ir_function_signature
*sig
;
2075 sig
= ir
->matching_signature(&empty
);
2079 foreach_list(node
, &sig
->body
) {
2080 ir_instruction
*ir
= (ir_instruction
*)node
;
2089 fs_visitor::visit(ir_function_signature
*ir
)
2091 assert(!"not reached");
2096 fs_visitor::visit(ir_emit_vertex
*)
2098 assert(!"not reached");
2102 fs_visitor::visit(ir_end_primitive
*)
2104 assert(!"not reached");
2108 fs_visitor::emit(fs_inst inst
)
2110 fs_inst
*list_inst
= new(mem_ctx
) fs_inst
;
2117 fs_visitor::emit(fs_inst
*inst
)
2119 if (force_uncompressed_stack
> 0)
2120 inst
->force_uncompressed
= true;
2121 else if (force_sechalf_stack
> 0)
2122 inst
->force_sechalf
= true;
2124 inst
->annotation
= this->current_annotation
;
2125 inst
->ir
= this->base_ir
;
2127 this->instructions
.push_tail(inst
);
2133 fs_visitor::emit(exec_list list
)
2135 foreach_list_safe(node
, &list
) {
2136 fs_inst
*inst
= (fs_inst
*)node
;
2142 /** Emits a dummy fragment shader consisting of magenta for bringup purposes. */
2144 fs_visitor::emit_dummy_fs()
2146 int reg_width
= dispatch_width
/ 8;
2148 /* Everyone's favorite color. */
2149 emit(MOV(fs_reg(MRF
, 2 + 0 * reg_width
), fs_reg(1.0f
)));
2150 emit(MOV(fs_reg(MRF
, 2 + 1 * reg_width
), fs_reg(0.0f
)));
2151 emit(MOV(fs_reg(MRF
, 2 + 2 * reg_width
), fs_reg(1.0f
)));
2152 emit(MOV(fs_reg(MRF
, 2 + 3 * reg_width
), fs_reg(0.0f
)));
2155 write
= emit(FS_OPCODE_FB_WRITE
, fs_reg(0), fs_reg(0));
2156 write
->base_mrf
= 2;
2157 write
->mlen
= 4 * reg_width
;
2161 /* The register location here is relative to the start of the URB
2162 * data. It will get adjusted to be a real location before
2163 * generate_code() time.
2166 fs_visitor::interp_reg(int location
, int channel
)
2168 int regnr
= urb_setup
[location
] * 2 + channel
/ 2;
2169 int stride
= (channel
& 1) * 4;
2171 assert(urb_setup
[location
] != -1);
2173 return brw_vec1_grf(regnr
, stride
);
2176 /** Emits the interpolation for the varying inputs. */
2178 fs_visitor::emit_interpolation_setup_gen4()
2180 this->current_annotation
= "compute pixel centers";
2181 this->pixel_x
= fs_reg(this, glsl_type::uint_type
);
2182 this->pixel_y
= fs_reg(this, glsl_type::uint_type
);
2183 this->pixel_x
.type
= BRW_REGISTER_TYPE_UW
;
2184 this->pixel_y
.type
= BRW_REGISTER_TYPE_UW
;
2186 emit(FS_OPCODE_PIXEL_X
, this->pixel_x
);
2187 emit(FS_OPCODE_PIXEL_Y
, this->pixel_y
);
2189 this->current_annotation
= "compute pixel deltas from v0";
2191 this->delta_x
[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
] =
2192 fs_reg(this, glsl_type::vec2_type
);
2193 this->delta_y
[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
] =
2194 this->delta_x
[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
];
2195 this->delta_y
[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
].reg_offset
++;
2197 this->delta_x
[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
] =
2198 fs_reg(this, glsl_type::float_type
);
2199 this->delta_y
[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
] =
2200 fs_reg(this, glsl_type::float_type
);
2202 emit(ADD(this->delta_x
[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
],
2203 this->pixel_x
, fs_reg(negate(brw_vec1_grf(1, 0)))));
2204 emit(ADD(this->delta_y
[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
],
2205 this->pixel_y
, fs_reg(negate(brw_vec1_grf(1, 1)))));
2207 this->current_annotation
= "compute pos.w and 1/pos.w";
2208 /* Compute wpos.w. It's always in our setup, since it's needed to
2209 * interpolate the other attributes.
2211 this->wpos_w
= fs_reg(this, glsl_type::float_type
);
2212 emit(FS_OPCODE_LINTERP
, wpos_w
,
2213 this->delta_x
[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
],
2214 this->delta_y
[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
],
2215 interp_reg(VARYING_SLOT_POS
, 3));
2216 /* Compute the pixel 1/W value from wpos.w. */
2217 this->pixel_w
= fs_reg(this, glsl_type::float_type
);
2218 emit_math(SHADER_OPCODE_RCP
, this->pixel_w
, wpos_w
);
2219 this->current_annotation
= NULL
;
2222 /** Emits the interpolation for the varying inputs. */
2224 fs_visitor::emit_interpolation_setup_gen6()
2226 struct brw_reg g1_uw
= retype(brw_vec1_grf(1, 0), BRW_REGISTER_TYPE_UW
);
2228 /* If the pixel centers end up used, the setup is the same as for gen4. */
2229 this->current_annotation
= "compute pixel centers";
2230 fs_reg int_pixel_x
= fs_reg(this, glsl_type::uint_type
);
2231 fs_reg int_pixel_y
= fs_reg(this, glsl_type::uint_type
);
2232 int_pixel_x
.type
= BRW_REGISTER_TYPE_UW
;
2233 int_pixel_y
.type
= BRW_REGISTER_TYPE_UW
;
2234 emit(ADD(int_pixel_x
,
2235 fs_reg(stride(suboffset(g1_uw
, 4), 2, 4, 0)),
2236 fs_reg(brw_imm_v(0x10101010))));
2237 emit(ADD(int_pixel_y
,
2238 fs_reg(stride(suboffset(g1_uw
, 5), 2, 4, 0)),
2239 fs_reg(brw_imm_v(0x11001100))));
2241 /* As of gen6, we can no longer mix float and int sources. We have
2242 * to turn the integer pixel centers into floats for their actual
2245 this->pixel_x
= fs_reg(this, glsl_type::float_type
);
2246 this->pixel_y
= fs_reg(this, glsl_type::float_type
);
2247 emit(MOV(this->pixel_x
, int_pixel_x
));
2248 emit(MOV(this->pixel_y
, int_pixel_y
));
2250 this->current_annotation
= "compute pos.w";
2251 this->pixel_w
= fs_reg(brw_vec8_grf(c
->source_w_reg
, 0));
2252 this->wpos_w
= fs_reg(this, glsl_type::float_type
);
2253 emit_math(SHADER_OPCODE_RCP
, this->wpos_w
, this->pixel_w
);
2255 for (int i
= 0; i
< BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT
; ++i
) {
2256 uint8_t reg
= c
->barycentric_coord_reg
[i
];
2257 this->delta_x
[i
] = fs_reg(brw_vec8_grf(reg
, 0));
2258 this->delta_y
[i
] = fs_reg(brw_vec8_grf(reg
+ 1, 0));
2261 this->current_annotation
= NULL
;
2265 fs_visitor::emit_color_write(int target
, int index
, int first_color_mrf
)
2267 int reg_width
= dispatch_width
/ 8;
2269 fs_reg color
= outputs
[target
];
2272 /* If there's no color data to be written, skip it. */
2273 if (color
.file
== BAD_FILE
)
2276 color
.reg_offset
+= index
;
2278 if (dispatch_width
== 8 || brw
->gen
>= 6) {
2279 /* SIMD8 write looks like:
2285 * gen6 SIMD16 DP write looks like:
2295 inst
= emit(MOV(fs_reg(MRF
, first_color_mrf
+ index
* reg_width
,
2298 inst
->saturate
= c
->key
.clamp_fragment_color
;
2300 /* pre-gen6 SIMD16 single source DP write looks like:
2310 if (brw
->has_compr4
) {
2311 /* By setting the high bit of the MRF register number, we
2312 * indicate that we want COMPR4 mode - instead of doing the
2313 * usual destination + 1 for the second half we get
2316 inst
= emit(MOV(fs_reg(MRF
, BRW_MRF_COMPR4
+ first_color_mrf
+ index
,
2319 inst
->saturate
= c
->key
.clamp_fragment_color
;
2321 push_force_uncompressed();
2322 inst
= emit(MOV(fs_reg(MRF
, first_color_mrf
+ index
, color
.type
),
2324 inst
->saturate
= c
->key
.clamp_fragment_color
;
2325 pop_force_uncompressed();
2327 push_force_sechalf();
2328 color
.sechalf
= true;
2329 inst
= emit(MOV(fs_reg(MRF
, first_color_mrf
+ index
+ 4, color
.type
),
2331 inst
->saturate
= c
->key
.clamp_fragment_color
;
2332 pop_force_sechalf();
2333 color
.sechalf
= false;
2339 fs_visitor::emit_fb_writes()
2341 this->current_annotation
= "FB write header";
2342 bool header_present
= true;
2343 /* We can potentially have a message length of up to 15, so we have to set
2344 * base_mrf to either 0 or 1 in order to fit in m0..m15.
2348 int reg_width
= dispatch_width
/ 8;
2349 bool do_dual_src
= this->dual_src_output
.file
!= BAD_FILE
;
2350 bool src0_alpha_to_render_target
= false;
2352 if (dispatch_width
== 16 && do_dual_src
) {
2353 fail("GL_ARB_blend_func_extended not yet supported in 16-wide.");
2354 do_dual_src
= false;
2357 /* From the Sandy Bridge PRM, volume 4, page 198:
2359 * "Dispatched Pixel Enables. One bit per pixel indicating
2360 * which pixels were originally enabled when the thread was
2361 * dispatched. This field is only required for the end-of-
2362 * thread message and on all dual-source messages."
2364 if (brw
->gen
>= 6 &&
2365 !this->fp
->UsesKill
&&
2367 c
->key
.nr_color_regions
== 1) {
2368 header_present
= false;
2371 if (header_present
) {
2372 src0_alpha_to_render_target
= brw
->gen
>= 6 &&
2374 c
->key
.replicate_alpha
;
2379 if (c
->aa_dest_stencil_reg
) {
2380 push_force_uncompressed();
2381 emit(MOV(fs_reg(MRF
, nr
++),
2382 fs_reg(brw_vec8_grf(c
->aa_dest_stencil_reg
, 0))));
2383 pop_force_uncompressed();
2386 /* Reserve space for color. It'll be filled in per MRT below. */
2388 nr
+= 4 * reg_width
;
2391 if (src0_alpha_to_render_target
)
2394 if (c
->source_depth_to_render_target
) {
2395 if (brw
->gen
== 6 && dispatch_width
== 16) {
2396 /* For outputting oDepth on gen6, SIMD8 writes have to be
2397 * used. This would require 8-wide moves of each half to
2398 * message regs, kind of like pre-gen5 SIMD16 FB writes.
2399 * Just bail on doing so for now.
2401 fail("Missing support for simd16 depth writes on gen6\n");
2404 if (fp
->Base
.OutputsWritten
& BITFIELD64_BIT(FRAG_RESULT_DEPTH
)) {
2405 /* Hand over gl_FragDepth. */
2406 assert(this->frag_depth
.file
!= BAD_FILE
);
2407 emit(MOV(fs_reg(MRF
, nr
), this->frag_depth
));
2409 /* Pass through the payload depth. */
2410 emit(MOV(fs_reg(MRF
, nr
),
2411 fs_reg(brw_vec8_grf(c
->source_depth_reg
, 0))));
2416 if (c
->dest_depth_reg
) {
2417 emit(MOV(fs_reg(MRF
, nr
),
2418 fs_reg(brw_vec8_grf(c
->dest_depth_reg
, 0))));
2423 fs_reg src0
= this->outputs
[0];
2424 fs_reg src1
= this->dual_src_output
;
2426 this->current_annotation
= ralloc_asprintf(this->mem_ctx
,
2428 for (int i
= 0; i
< 4; i
++) {
2429 fs_inst
*inst
= emit(MOV(fs_reg(MRF
, color_mrf
+ i
, src0
.type
), src0
));
2431 inst
->saturate
= c
->key
.clamp_fragment_color
;
2434 this->current_annotation
= ralloc_asprintf(this->mem_ctx
,
2436 for (int i
= 0; i
< 4; i
++) {
2437 fs_inst
*inst
= emit(MOV(fs_reg(MRF
, color_mrf
+ 4 + i
, src1
.type
),
2440 inst
->saturate
= c
->key
.clamp_fragment_color
;
2443 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
)
2444 emit_shader_time_end();
2446 fs_inst
*inst
= emit(FS_OPCODE_FB_WRITE
);
2448 inst
->base_mrf
= base_mrf
;
2449 inst
->mlen
= nr
- base_mrf
;
2451 inst
->header_present
= header_present
;
2453 c
->prog_data
.dual_src_blend
= true;
2454 this->current_annotation
= NULL
;
2458 for (int target
= 0; target
< c
->key
.nr_color_regions
; target
++) {
2459 this->current_annotation
= ralloc_asprintf(this->mem_ctx
,
2460 "FB write target %d",
2462 /* If src0_alpha_to_render_target is true, include source zero alpha
2463 * data in RenderTargetWrite message for targets > 0.
2465 int write_color_mrf
= color_mrf
;
2466 if (src0_alpha_to_render_target
&& target
!= 0) {
2468 fs_reg color
= outputs
[0];
2469 color
.reg_offset
+= 3;
2471 inst
= emit(MOV(fs_reg(MRF
, write_color_mrf
, color
.type
),
2473 inst
->saturate
= c
->key
.clamp_fragment_color
;
2474 write_color_mrf
= color_mrf
+ reg_width
;
2477 for (unsigned i
= 0; i
< this->output_components
[target
]; i
++)
2478 emit_color_write(target
, i
, write_color_mrf
);
2481 if (target
== c
->key
.nr_color_regions
- 1) {
2484 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
)
2485 emit_shader_time_end();
2488 fs_inst
*inst
= emit(FS_OPCODE_FB_WRITE
);
2489 inst
->target
= target
;
2490 inst
->base_mrf
= base_mrf
;
2491 if (src0_alpha_to_render_target
&& target
== 0)
2492 inst
->mlen
= nr
- base_mrf
- reg_width
;
2494 inst
->mlen
= nr
- base_mrf
;
2496 inst
->header_present
= header_present
;
2499 if (c
->key
.nr_color_regions
== 0) {
2500 /* Even if there's no color buffers enabled, we still need to send
2501 * alpha out the pipeline to our null renderbuffer to support
2502 * alpha-testing, alpha-to-coverage, and so on.
2504 emit_color_write(0, 3, color_mrf
);
2506 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
)
2507 emit_shader_time_end();
2509 fs_inst
*inst
= emit(FS_OPCODE_FB_WRITE
);
2510 inst
->base_mrf
= base_mrf
;
2511 inst
->mlen
= nr
- base_mrf
;
2513 inst
->header_present
= header_present
;
2516 this->current_annotation
= NULL
;
2520 fs_visitor::resolve_ud_negate(fs_reg
*reg
)
2522 if (reg
->type
!= BRW_REGISTER_TYPE_UD
||
2526 fs_reg temp
= fs_reg(this, glsl_type::uint_type
);
2527 emit(MOV(temp
, *reg
));
2532 fs_visitor::resolve_bool_comparison(ir_rvalue
*rvalue
, fs_reg
*reg
)
2534 if (rvalue
->type
!= glsl_type::bool_type
)
2537 fs_reg temp
= fs_reg(this, glsl_type::bool_type
);
2538 emit(AND(temp
, *reg
, fs_reg(1)));
2542 fs_visitor::fs_visitor(struct brw_context
*brw
,
2543 struct brw_wm_compile
*c
,
2544 struct gl_shader_program
*shader_prog
,
2545 struct gl_fragment_program
*fp
,
2546 unsigned dispatch_width
)
2547 : dispatch_width(dispatch_width
)
2552 this->shader_prog
= shader_prog
;
2553 this->ctx
= &brw
->ctx
;
2554 this->mem_ctx
= ralloc_context(NULL
);
2556 shader
= (struct brw_shader
*)
2557 shader_prog
->_LinkedShaders
[MESA_SHADER_FRAGMENT
];
2560 this->failed
= false;
2561 this->variable_ht
= hash_table_ctor(0,
2562 hash_table_pointer_hash
,
2563 hash_table_pointer_compare
);
2565 memset(this->outputs
, 0, sizeof(this->outputs
));
2566 memset(this->output_components
, 0, sizeof(this->output_components
));
2567 this->first_non_payload_grf
= 0;
2568 this->max_grf
= brw
->gen
>= 7 ? GEN7_MRF_HACK_START
: BRW_MAX_GRF
;
2570 this->current_annotation
= NULL
;
2571 this->base_ir
= NULL
;
2573 this->virtual_grf_sizes
= NULL
;
2574 this->virtual_grf_count
= 0;
2575 this->virtual_grf_array_size
= 0;
2576 this->virtual_grf_start
= NULL
;
2577 this->virtual_grf_end
= NULL
;
2578 this->live_intervals_valid
= false;
2580 this->params_remap
= NULL
;
2581 this->nr_params_remap
= 0;
2583 this->force_uncompressed_stack
= 0;
2584 this->force_sechalf_stack
= 0;
2586 memset(&this->param_size
, 0, sizeof(this->param_size
));
2589 fs_visitor::~fs_visitor()
2591 ralloc_free(this->mem_ctx
);
2592 hash_table_dtor(this->variable_ht
);