2 * Copyright © 2011 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
25 #include "glsl/ir_uniform.h"
27 #include "main/context.h"
28 #include "main/macros.h"
29 #include "program/prog_parameter.h"
30 #include "program/sampler.h"
35 vec4_instruction::vec4_instruction(vec4_visitor
*v
,
36 enum opcode opcode
, dst_reg dst
,
37 src_reg src0
, src_reg src1
, src_reg src2
)
39 this->opcode
= opcode
;
44 this->ir
= v
->base_ir
;
45 this->annotation
= v
->current_annotation
;
49 vec4_visitor::emit(vec4_instruction
*inst
)
51 this->instructions
.push_tail(inst
);
57 vec4_visitor::emit_before(vec4_instruction
*inst
, vec4_instruction
*new_inst
)
59 new_inst
->ir
= inst
->ir
;
60 new_inst
->annotation
= inst
->annotation
;
62 inst
->insert_before(new_inst
);
68 vec4_visitor::emit(enum opcode opcode
, dst_reg dst
,
69 src_reg src0
, src_reg src1
, src_reg src2
)
71 return emit(new(mem_ctx
) vec4_instruction(this, opcode
, dst
,
77 vec4_visitor::emit(enum opcode opcode
, dst_reg dst
, src_reg src0
, src_reg src1
)
79 return emit(new(mem_ctx
) vec4_instruction(this, opcode
, dst
, src0
, src1
));
83 vec4_visitor::emit(enum opcode opcode
, dst_reg dst
, src_reg src0
)
85 return emit(new(mem_ctx
) vec4_instruction(this, opcode
, dst
, src0
));
89 vec4_visitor::emit(enum opcode opcode
)
91 return emit(new(mem_ctx
) vec4_instruction(this, opcode
, dst_reg()));
96 vec4_visitor::op(dst_reg dst, src_reg src0) \
98 return new(mem_ctx) vec4_instruction(this, BRW_OPCODE_##op, dst, \
104 vec4_visitor::op(dst_reg dst, src_reg src0, src_reg src1) \
106 return new(mem_ctx) vec4_instruction(this, BRW_OPCODE_##op, dst, \
131 /** Gen4 predicated IF. */
133 vec4_visitor::IF(uint32_t predicate
)
135 vec4_instruction
*inst
;
137 inst
= new(mem_ctx
) vec4_instruction(this, BRW_OPCODE_IF
);
138 inst
->predicate
= predicate
;
143 /** Gen6+ IF with embedded comparison. */
145 vec4_visitor::IF(src_reg src0
, src_reg src1
, uint32_t condition
)
147 assert(intel
->gen
>= 6);
149 vec4_instruction
*inst
;
151 resolve_ud_negate(&src0
);
152 resolve_ud_negate(&src1
);
154 inst
= new(mem_ctx
) vec4_instruction(this, BRW_OPCODE_IF
, dst_null_d(),
156 inst
->conditional_mod
= condition
;
162 * CMP: Sets the low bit of the destination channels with the result
163 * of the comparison, while the upper bits are undefined, and updates
164 * the flag register with the packed 16 bits of the result.
167 vec4_visitor::CMP(dst_reg dst
, src_reg src0
, src_reg src1
, uint32_t condition
)
169 vec4_instruction
*inst
;
171 /* original gen4 does type conversion to the destination type
172 * before before comparison, producing garbage results for floating
175 if (intel
->gen
== 4) {
176 dst
.type
= src0
.type
;
177 if (dst
.file
== HW_REG
)
178 dst
.fixed_hw_reg
.type
= dst
.type
;
181 resolve_ud_negate(&src0
);
182 resolve_ud_negate(&src1
);
184 inst
= new(mem_ctx
) vec4_instruction(this, BRW_OPCODE_CMP
, dst
, src0
, src1
);
185 inst
->conditional_mod
= condition
;
191 vec4_visitor::SCRATCH_READ(dst_reg dst
, src_reg index
)
193 vec4_instruction
*inst
;
195 inst
= new(mem_ctx
) vec4_instruction(this, VS_OPCODE_SCRATCH_READ
,
204 vec4_visitor::SCRATCH_WRITE(dst_reg dst
, src_reg src
, src_reg index
)
206 vec4_instruction
*inst
;
208 inst
= new(mem_ctx
) vec4_instruction(this, VS_OPCODE_SCRATCH_WRITE
,
217 vec4_visitor::emit_dp(dst_reg dst
, src_reg src0
, src_reg src1
, unsigned elements
)
219 static enum opcode dot_opcodes
[] = {
220 BRW_OPCODE_DP2
, BRW_OPCODE_DP3
, BRW_OPCODE_DP4
223 emit(dot_opcodes
[elements
- 2], dst
, src0
, src1
);
227 vec4_visitor::fix_math_operand(src_reg src
)
229 /* The gen6 math instruction ignores the source modifiers --
230 * swizzle, abs, negate, and at least some parts of the register
231 * region description.
233 * Rather than trying to enumerate all these cases, *always* expand the
234 * operand to a temp GRF for gen6.
236 * For gen7, keep the operand as-is, except if immediate, which gen7 still
240 if (intel
->gen
== 7 && src
.file
!= IMM
)
243 dst_reg expanded
= dst_reg(this, glsl_type::vec4_type
);
244 expanded
.type
= src
.type
;
245 emit(MOV(expanded
, src
));
246 return src_reg(expanded
);
250 vec4_visitor::emit_math1_gen6(enum opcode opcode
, dst_reg dst
, src_reg src
)
252 src
= fix_math_operand(src
);
254 if (dst
.writemask
!= WRITEMASK_XYZW
) {
255 /* The gen6 math instruction must be align1, so we can't do
258 dst_reg temp_dst
= dst_reg(this, glsl_type::vec4_type
);
260 emit(opcode
, temp_dst
, src
);
262 emit(MOV(dst
, src_reg(temp_dst
)));
264 emit(opcode
, dst
, src
);
269 vec4_visitor::emit_math1_gen4(enum opcode opcode
, dst_reg dst
, src_reg src
)
271 vec4_instruction
*inst
= emit(opcode
, dst
, src
);
277 vec4_visitor::emit_math(opcode opcode
, dst_reg dst
, src_reg src
)
280 case SHADER_OPCODE_RCP
:
281 case SHADER_OPCODE_RSQ
:
282 case SHADER_OPCODE_SQRT
:
283 case SHADER_OPCODE_EXP2
:
284 case SHADER_OPCODE_LOG2
:
285 case SHADER_OPCODE_SIN
:
286 case SHADER_OPCODE_COS
:
289 assert(!"not reached: bad math opcode");
293 if (intel
->gen
>= 6) {
294 return emit_math1_gen6(opcode
, dst
, src
);
296 return emit_math1_gen4(opcode
, dst
, src
);
301 vec4_visitor::emit_math2_gen6(enum opcode opcode
,
302 dst_reg dst
, src_reg src0
, src_reg src1
)
304 src0
= fix_math_operand(src0
);
305 src1
= fix_math_operand(src1
);
307 if (dst
.writemask
!= WRITEMASK_XYZW
) {
308 /* The gen6 math instruction must be align1, so we can't do
311 dst_reg temp_dst
= dst_reg(this, glsl_type::vec4_type
);
312 temp_dst
.type
= dst
.type
;
314 emit(opcode
, temp_dst
, src0
, src1
);
316 emit(MOV(dst
, src_reg(temp_dst
)));
318 emit(opcode
, dst
, src0
, src1
);
323 vec4_visitor::emit_math2_gen4(enum opcode opcode
,
324 dst_reg dst
, src_reg src0
, src_reg src1
)
326 vec4_instruction
*inst
= emit(opcode
, dst
, src0
, src1
);
332 vec4_visitor::emit_math(enum opcode opcode
,
333 dst_reg dst
, src_reg src0
, src_reg src1
)
336 case SHADER_OPCODE_POW
:
337 case SHADER_OPCODE_INT_QUOTIENT
:
338 case SHADER_OPCODE_INT_REMAINDER
:
341 assert(!"not reached: unsupported binary math opcode");
345 if (intel
->gen
>= 6) {
346 return emit_math2_gen6(opcode
, dst
, src0
, src1
);
348 return emit_math2_gen4(opcode
, dst
, src0
, src1
);
353 vec4_visitor::emit_pack_half_2x16(dst_reg dst
, src_reg src0
)
356 assert(!"ir_unop_pack_half_2x16 should be lowered");
358 assert(dst
.type
== BRW_REGISTER_TYPE_UD
);
359 assert(src0
.type
== BRW_REGISTER_TYPE_F
);
361 /* From the Ivybridge PRM, Vol4, Part3, Section 6.27 f32to16:
363 * Because this instruction does not have a 16-bit floating-point type,
364 * the destination data type must be Word (W).
366 * The destination must be DWord-aligned and specify a horizontal stride
367 * (HorzStride) of 2. The 16-bit result is stored in the lower word of
368 * each destination channel and the upper word is not modified.
370 * The above restriction implies that the f32to16 instruction must use
371 * align1 mode, because only in align1 mode is it possible to specify
372 * horizontal stride. We choose here to defy the hardware docs and emit
373 * align16 instructions.
375 * (I [chadv] did attempt to emit align1 instructions for VS f32to16
376 * instructions. I was partially successful in that the code passed all
377 * tests. However, the code was dubiously correct and fragile, and the
378 * tests were not harsh enough to probe that frailty. Not trusting the
379 * code, I chose instead to remain in align16 mode in defiance of the hw
382 * I've [chadv] experimentally confirmed that, on gen7 hardware and the
383 * simulator, emitting a f32to16 in align16 mode with UD as destination
384 * data type is safe. The behavior differs from that specified in the PRM
385 * in that the upper word of each destination channel is cleared to 0.
388 dst_reg
tmp_dst(this, glsl_type::uvec2_type
);
389 src_reg
tmp_src(tmp_dst
);
392 /* Verify the undocumented behavior on which the following instructions
393 * rely. If f32to16 fails to clear the upper word of the X and Y channels,
394 * then the result of the bit-or instruction below will be incorrect.
396 * You should inspect the disasm output in order to verify that the MOV is
397 * not optimized away.
399 emit(MOV(tmp_dst
, src_reg(0x12345678u
)));
402 /* Give tmp the form below, where "." means untouched.
405 * |.|.|0x0000hhhh|0x0000llll|.|.|0x0000hhhh|0x0000llll|
407 * That the upper word of each write-channel be 0 is required for the
408 * following bit-shift and bit-or instructions to work. Note that this
409 * relies on the undocumented hardware behavior mentioned above.
411 tmp_dst
.writemask
= WRITEMASK_XY
;
412 emit(F32TO16(tmp_dst
, src0
));
414 /* Give the write-channels of dst the form:
417 tmp_src
.swizzle
= SWIZZLE_Y
;
418 emit(SHL(dst
, tmp_src
, src_reg(16u)));
420 /* Finally, give the write-channels of dst the form of packHalf2x16's
424 tmp_src
.swizzle
= SWIZZLE_X
;
425 emit(OR(dst
, src_reg(dst
), tmp_src
));
429 vec4_visitor::emit_unpack_half_2x16(dst_reg dst
, src_reg src0
)
432 assert(!"ir_unop_unpack_half_2x16 should be lowered");
434 assert(dst
.type
== BRW_REGISTER_TYPE_F
);
435 assert(src0
.type
== BRW_REGISTER_TYPE_UD
);
437 /* From the Ivybridge PRM, Vol4, Part3, Section 6.26 f16to32:
439 * Because this instruction does not have a 16-bit floating-point type,
440 * the source data type must be Word (W). The destination type must be
443 * To use W as the source data type, we must adjust horizontal strides,
444 * which is only possible in align1 mode. All my [chadv] attempts at
445 * emitting align1 instructions for unpackHalf2x16 failed to pass the
446 * Piglit tests, so I gave up.
448 * I've verified that, on gen7 hardware and the simulator, it is safe to
449 * emit f16to32 in align16 mode with UD as source data type.
452 dst_reg
tmp_dst(this, glsl_type::uvec2_type
);
453 src_reg
tmp_src(tmp_dst
);
455 tmp_dst
.writemask
= WRITEMASK_X
;
456 emit(AND(tmp_dst
, src0
, src_reg(0xffffu
)));
458 tmp_dst
.writemask
= WRITEMASK_Y
;
459 emit(SHR(tmp_dst
, src0
, src_reg(16u)));
461 dst
.writemask
= WRITEMASK_XY
;
462 emit(F16TO32(dst
, tmp_src
));
466 vec4_visitor::visit_instructions(const exec_list
*list
)
468 foreach_list(node
, list
) {
469 ir_instruction
*ir
= (ir_instruction
*)node
;
478 type_size(const struct glsl_type
*type
)
483 switch (type
->base_type
) {
486 case GLSL_TYPE_FLOAT
:
488 if (type
->is_matrix()) {
489 return type
->matrix_columns
;
491 /* Regardless of size of vector, it gets a vec4. This is bad
492 * packing for things like floats, but otherwise arrays become a
493 * mess. Hopefully a later pass over the code can pack scalars
494 * down if appropriate.
498 case GLSL_TYPE_ARRAY
:
499 assert(type
->length
> 0);
500 return type_size(type
->fields
.array
) * type
->length
;
501 case GLSL_TYPE_STRUCT
:
503 for (i
= 0; i
< type
->length
; i
++) {
504 size
+= type_size(type
->fields
.structure
[i
].type
);
507 case GLSL_TYPE_SAMPLER
:
508 /* Samplers take up one slot in UNIFORMS[], but they're baked in
513 case GLSL_TYPE_ERROR
:
522 vec4_visitor::virtual_grf_alloc(int size
)
524 if (virtual_grf_array_size
<= virtual_grf_count
) {
525 if (virtual_grf_array_size
== 0)
526 virtual_grf_array_size
= 16;
528 virtual_grf_array_size
*= 2;
529 virtual_grf_sizes
= reralloc(mem_ctx
, virtual_grf_sizes
, int,
530 virtual_grf_array_size
);
531 virtual_grf_reg_map
= reralloc(mem_ctx
, virtual_grf_reg_map
, int,
532 virtual_grf_array_size
);
534 virtual_grf_reg_map
[virtual_grf_count
] = virtual_grf_reg_count
;
535 virtual_grf_reg_count
+= size
;
536 virtual_grf_sizes
[virtual_grf_count
] = size
;
537 return virtual_grf_count
++;
540 src_reg::src_reg(class vec4_visitor
*v
, const struct glsl_type
*type
)
545 this->reg
= v
->virtual_grf_alloc(type_size(type
));
547 if (type
->is_array() || type
->is_record()) {
548 this->swizzle
= BRW_SWIZZLE_NOOP
;
550 this->swizzle
= swizzle_for_size(type
->vector_elements
);
553 this->type
= brw_type_for_base_type(type
);
556 dst_reg::dst_reg(class vec4_visitor
*v
, const struct glsl_type
*type
)
561 this->reg
= v
->virtual_grf_alloc(type_size(type
));
563 if (type
->is_array() || type
->is_record()) {
564 this->writemask
= WRITEMASK_XYZW
;
566 this->writemask
= (1 << type
->vector_elements
) - 1;
569 this->type
= brw_type_for_base_type(type
);
572 /* Our support for uniforms is piggy-backed on the struct
573 * gl_fragment_program, because that's where the values actually
574 * get stored, rather than in some global gl_shader_program uniform
578 vec4_visitor::setup_uniform_values(ir_variable
*ir
)
580 int namelen
= strlen(ir
->name
);
582 /* The data for our (non-builtin) uniforms is stored in a series of
583 * gl_uniform_driver_storage structs for each subcomponent that
584 * glGetUniformLocation() could name. We know it's been set up in the same
585 * order we'd walk the type, so walk the list of storage and find anything
586 * with our name, or the prefix of a component that starts with our name.
588 for (unsigned u
= 0; u
< prog
->NumUserUniformStorage
; u
++) {
589 struct gl_uniform_storage
*storage
= &prog
->UniformStorage
[u
];
591 if (strncmp(ir
->name
, storage
->name
, namelen
) != 0 ||
592 (storage
->name
[namelen
] != 0 &&
593 storage
->name
[namelen
] != '.' &&
594 storage
->name
[namelen
] != '[')) {
598 gl_constant_value
*components
= storage
->storage
;
599 unsigned vector_count
= (MAX2(storage
->array_elements
, 1) *
600 storage
->type
->matrix_columns
);
602 for (unsigned s
= 0; s
< vector_count
; s
++) {
603 uniform_vector_size
[uniforms
] = storage
->type
->vector_elements
;
606 for (i
= 0; i
< uniform_vector_size
[uniforms
]; i
++) {
607 c
->prog_data
.param
[uniforms
* 4 + i
] = &components
->f
;
611 static float zero
= 0;
612 c
->prog_data
.param
[uniforms
* 4 + i
] = &zero
;
621 vec4_visitor::setup_uniform_clipplane_values()
623 gl_clip_plane
*clip_planes
= brw_select_clip_planes(ctx
);
625 if (intel
->gen
< 6) {
626 /* Pre-Gen6, we compact clip planes. For example, if the user
627 * enables just clip planes 0, 1, and 3, we will enable clip planes
628 * 0, 1, and 2 in the hardware, and we'll move clip plane 3 to clip
629 * plane 2. This simplifies the implementation of the Gen6 clip
632 int compacted_clipplane_index
= 0;
633 for (int i
= 0; i
< MAX_CLIP_PLANES
; ++i
) {
634 if (!(c
->key
.userclip_planes_enabled_gen_4_5
& (1 << i
)))
637 this->uniform_vector_size
[this->uniforms
] = 4;
638 this->userplane
[compacted_clipplane_index
] = dst_reg(UNIFORM
, this->uniforms
);
639 this->userplane
[compacted_clipplane_index
].type
= BRW_REGISTER_TYPE_F
;
640 for (int j
= 0; j
< 4; ++j
) {
641 c
->prog_data
.param
[this->uniforms
* 4 + j
] = &clip_planes
[i
][j
];
643 ++compacted_clipplane_index
;
647 /* In Gen6 and later, we don't compact clip planes, because this
648 * simplifies the implementation of gl_ClipDistance.
650 for (int i
= 0; i
< c
->key
.nr_userclip_plane_consts
; ++i
) {
651 this->uniform_vector_size
[this->uniforms
] = 4;
652 this->userplane
[i
] = dst_reg(UNIFORM
, this->uniforms
);
653 this->userplane
[i
].type
= BRW_REGISTER_TYPE_F
;
654 for (int j
= 0; j
< 4; ++j
) {
655 c
->prog_data
.param
[this->uniforms
* 4 + j
] = &clip_planes
[i
][j
];
662 /* Our support for builtin uniforms is even scarier than non-builtin.
663 * It sits on top of the PROG_STATE_VAR parameters that are
664 * automatically updated from GL context state.
667 vec4_visitor::setup_builtin_uniform_values(ir_variable
*ir
)
669 const ir_state_slot
*const slots
= ir
->state_slots
;
670 assert(ir
->state_slots
!= NULL
);
672 for (unsigned int i
= 0; i
< ir
->num_state_slots
; i
++) {
673 /* This state reference has already been setup by ir_to_mesa,
674 * but we'll get the same index back here. We can reference
675 * ParameterValues directly, since unlike brw_fs.cpp, we never
676 * add new state references during compile.
678 int index
= _mesa_add_state_reference(this->vp
->Base
.Parameters
,
679 (gl_state_index
*)slots
[i
].tokens
);
680 float *values
= &this->vp
->Base
.Parameters
->ParameterValues
[index
][0].f
;
682 this->uniform_vector_size
[this->uniforms
] = 0;
683 /* Add each of the unique swizzled channels of the element.
684 * This will end up matching the size of the glsl_type of this field.
687 for (unsigned int j
= 0; j
< 4; j
++) {
688 int swiz
= GET_SWZ(slots
[i
].swizzle
, j
);
691 c
->prog_data
.param
[this->uniforms
* 4 + j
] = &values
[swiz
];
692 if (swiz
<= last_swiz
)
693 this->uniform_vector_size
[this->uniforms
]++;
700 vec4_visitor::variable_storage(ir_variable
*var
)
702 return (dst_reg
*)hash_table_find(this->variable_ht
, var
);
706 vec4_visitor::emit_bool_to_cond_code(ir_rvalue
*ir
, uint32_t *predicate
)
708 ir_expression
*expr
= ir
->as_expression();
710 *predicate
= BRW_PREDICATE_NORMAL
;
714 vec4_instruction
*inst
;
716 assert(expr
->get_num_operands() <= 2);
717 for (unsigned int i
= 0; i
< expr
->get_num_operands(); i
++) {
718 expr
->operands
[i
]->accept(this);
719 op
[i
] = this->result
;
721 resolve_ud_negate(&op
[i
]);
724 switch (expr
->operation
) {
725 case ir_unop_logic_not
:
726 inst
= emit(AND(dst_null_d(), op
[0], src_reg(1)));
727 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
730 case ir_binop_logic_xor
:
731 inst
= emit(XOR(dst_null_d(), op
[0], op
[1]));
732 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
735 case ir_binop_logic_or
:
736 inst
= emit(OR(dst_null_d(), op
[0], op
[1]));
737 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
740 case ir_binop_logic_and
:
741 inst
= emit(AND(dst_null_d(), op
[0], op
[1]));
742 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
746 if (intel
->gen
>= 6) {
747 emit(CMP(dst_null_d(), op
[0], src_reg(0.0f
), BRW_CONDITIONAL_NZ
));
749 inst
= emit(MOV(dst_null_f(), op
[0]));
750 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
755 if (intel
->gen
>= 6) {
756 emit(CMP(dst_null_d(), op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
758 inst
= emit(MOV(dst_null_d(), op
[0]));
759 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
763 case ir_binop_all_equal
:
764 inst
= emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_Z
));
765 *predicate
= BRW_PREDICATE_ALIGN16_ALL4H
;
768 case ir_binop_any_nequal
:
769 inst
= emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_NZ
));
770 *predicate
= BRW_PREDICATE_ALIGN16_ANY4H
;
774 inst
= emit(CMP(dst_null_d(), op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
775 *predicate
= BRW_PREDICATE_ALIGN16_ANY4H
;
778 case ir_binop_greater
:
779 case ir_binop_gequal
:
781 case ir_binop_lequal
:
783 case ir_binop_nequal
:
784 emit(CMP(dst_null_d(), op
[0], op
[1],
785 brw_conditional_for_comparison(expr
->operation
)));
789 assert(!"not reached");
797 resolve_ud_negate(&this->result
);
799 if (intel
->gen
>= 6) {
800 vec4_instruction
*inst
= emit(AND(dst_null_d(),
801 this->result
, src_reg(1)));
802 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
804 vec4_instruction
*inst
= emit(MOV(dst_null_d(), this->result
));
805 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
810 * Emit a gen6 IF statement with the comparison folded into the IF
814 vec4_visitor::emit_if_gen6(ir_if
*ir
)
816 ir_expression
*expr
= ir
->condition
->as_expression();
822 assert(expr
->get_num_operands() <= 2);
823 for (unsigned int i
= 0; i
< expr
->get_num_operands(); i
++) {
824 expr
->operands
[i
]->accept(this);
825 op
[i
] = this->result
;
828 switch (expr
->operation
) {
829 case ir_unop_logic_not
:
830 emit(IF(op
[0], src_reg(0), BRW_CONDITIONAL_Z
));
833 case ir_binop_logic_xor
:
834 emit(IF(op
[0], op
[1], BRW_CONDITIONAL_NZ
));
837 case ir_binop_logic_or
:
838 temp
= dst_reg(this, glsl_type::bool_type
);
839 emit(OR(temp
, op
[0], op
[1]));
840 emit(IF(src_reg(temp
), src_reg(0), BRW_CONDITIONAL_NZ
));
843 case ir_binop_logic_and
:
844 temp
= dst_reg(this, glsl_type::bool_type
);
845 emit(AND(temp
, op
[0], op
[1]));
846 emit(IF(src_reg(temp
), src_reg(0), BRW_CONDITIONAL_NZ
));
850 emit(IF(op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
854 emit(IF(op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
857 case ir_binop_greater
:
858 case ir_binop_gequal
:
860 case ir_binop_lequal
:
862 case ir_binop_nequal
:
863 emit(IF(op
[0], op
[1],
864 brw_conditional_for_comparison(expr
->operation
)));
867 case ir_binop_all_equal
:
868 emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_Z
));
869 emit(IF(BRW_PREDICATE_ALIGN16_ALL4H
));
872 case ir_binop_any_nequal
:
873 emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_NZ
));
874 emit(IF(BRW_PREDICATE_ALIGN16_ANY4H
));
878 emit(CMP(dst_null_d(), op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
879 emit(IF(BRW_PREDICATE_ALIGN16_ANY4H
));
883 assert(!"not reached");
884 emit(IF(op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
890 ir
->condition
->accept(this);
892 emit(IF(this->result
, src_reg(0), BRW_CONDITIONAL_NZ
));
896 with_writemask(dst_reg
const & r
, int mask
)
899 result
.writemask
= mask
;
904 vec4_visitor::emit_attribute_fixups()
906 dst_reg sign_recovery_shift
;
907 dst_reg normalize_factor
;
908 dst_reg es3_normalize_factor
;
910 for (int i
= 0; i
< VERT_ATTRIB_MAX
; i
++) {
911 if (prog_data
->inputs_read
& BITFIELD64_BIT(i
)) {
912 uint8_t wa_flags
= c
->key
.gl_attrib_wa_flags
[i
];
913 dst_reg
reg(ATTR
, i
);
915 reg_d
.type
= BRW_REGISTER_TYPE_D
;
916 dst_reg reg_ud
= reg
;
917 reg_ud
.type
= BRW_REGISTER_TYPE_UD
;
919 /* Do GL_FIXED rescaling for GLES2.0. Our GL_FIXED attributes
920 * come in as floating point conversions of the integer values.
922 if (wa_flags
& BRW_ATTRIB_WA_COMPONENT_MASK
) {
924 dst
.type
= brw_type_for_base_type(glsl_type::vec4_type
);
925 dst
.writemask
= (1 << (wa_flags
& BRW_ATTRIB_WA_COMPONENT_MASK
)) - 1;
926 emit(MUL(dst
, src_reg(dst
), src_reg(1.0f
/ 65536.0f
)));
929 /* Do sign recovery for 2101010 formats if required. */
930 if (wa_flags
& BRW_ATTRIB_WA_SIGN
) {
931 if (sign_recovery_shift
.file
== BAD_FILE
) {
932 /* shift constant: <22,22,22,30> */
933 sign_recovery_shift
= dst_reg(this, glsl_type::uvec4_type
);
934 emit(MOV(with_writemask(sign_recovery_shift
, WRITEMASK_XYZ
), src_reg(22u)));
935 emit(MOV(with_writemask(sign_recovery_shift
, WRITEMASK_W
), src_reg(30u)));
938 emit(SHL(reg_ud
, src_reg(reg_ud
), src_reg(sign_recovery_shift
)));
939 emit(ASR(reg_d
, src_reg(reg_d
), src_reg(sign_recovery_shift
)));
942 /* Apply BGRA swizzle if required. */
943 if (wa_flags
& BRW_ATTRIB_WA_BGRA
) {
944 src_reg temp
= src_reg(reg
);
945 temp
.swizzle
= BRW_SWIZZLE4(2,1,0,3);
946 emit(MOV(reg
, temp
));
949 if (wa_flags
& BRW_ATTRIB_WA_NORMALIZE
) {
950 /* ES 3.0 has different rules for converting signed normalized
951 * fixed-point numbers than desktop GL.
953 if (_mesa_is_gles3(ctx
) && (wa_flags
& BRW_ATTRIB_WA_SIGN
)) {
954 /* According to equation 2.2 of the ES 3.0 specification,
955 * signed normalization conversion is done by:
957 * f = c / (2^(b-1)-1)
959 if (es3_normalize_factor
.file
== BAD_FILE
) {
960 /* mul constant: 1 / (2^(b-1) - 1) */
961 es3_normalize_factor
= dst_reg(this, glsl_type::vec4_type
);
962 emit(MOV(with_writemask(es3_normalize_factor
, WRITEMASK_XYZ
),
963 src_reg(1.0f
/ ((1<<9) - 1))));
964 emit(MOV(with_writemask(es3_normalize_factor
, WRITEMASK_W
),
965 src_reg(1.0f
/ ((1<<1) - 1))));
969 dst
.type
= brw_type_for_base_type(glsl_type::vec4_type
);
970 emit(MOV(dst
, src_reg(reg_d
)));
971 emit(MUL(dst
, src_reg(dst
), src_reg(es3_normalize_factor
)));
972 emit_minmax(BRW_CONDITIONAL_G
, dst
, src_reg(dst
), src_reg(-1.0f
));
974 /* The following equations are from the OpenGL 3.2 specification:
976 * 2.1 unsigned normalization
979 * 2.2 signed normalization
982 * Both of these share a common divisor, which is represented by
983 * "normalize_factor" in the code below.
985 if (normalize_factor
.file
== BAD_FILE
) {
986 /* 1 / (2^b - 1) for b=<10,10,10,2> */
987 normalize_factor
= dst_reg(this, glsl_type::vec4_type
);
988 emit(MOV(with_writemask(normalize_factor
, WRITEMASK_XYZ
),
989 src_reg(1.0f
/ ((1<<10) - 1))));
990 emit(MOV(with_writemask(normalize_factor
, WRITEMASK_W
),
991 src_reg(1.0f
/ ((1<<2) - 1))));
995 dst
.type
= brw_type_for_base_type(glsl_type::vec4_type
);
996 emit(MOV(dst
, src_reg((wa_flags
& BRW_ATTRIB_WA_SIGN
) ? reg_d
: reg_ud
)));
998 /* For signed normalization, we want the numerator to be 2c+1. */
999 if (wa_flags
& BRW_ATTRIB_WA_SIGN
) {
1000 emit(MUL(dst
, src_reg(dst
), src_reg(2.0f
)));
1001 emit(ADD(dst
, src_reg(dst
), src_reg(1.0f
)));
1004 emit(MUL(dst
, src_reg(dst
), src_reg(normalize_factor
)));
1008 if (wa_flags
& BRW_ATTRIB_WA_SCALE
) {
1010 dst
.type
= brw_type_for_base_type(glsl_type::vec4_type
);
1011 emit(MOV(dst
, src_reg((wa_flags
& BRW_ATTRIB_WA_SIGN
) ? reg_d
: reg_ud
)));
1018 vec4_visitor::visit(ir_variable
*ir
)
1020 dst_reg
*reg
= NULL
;
1022 if (variable_storage(ir
))
1026 case ir_var_shader_in
:
1027 reg
= new(mem_ctx
) dst_reg(ATTR
, ir
->location
);
1030 case ir_var_shader_out
:
1031 reg
= new(mem_ctx
) dst_reg(this, ir
->type
);
1033 for (int i
= 0; i
< type_size(ir
->type
); i
++) {
1034 output_reg
[ir
->location
+ i
] = *reg
;
1035 output_reg
[ir
->location
+ i
].reg_offset
= i
;
1036 output_reg
[ir
->location
+ i
].type
=
1037 brw_type_for_base_type(ir
->type
->get_scalar_type());
1038 output_reg_annotation
[ir
->location
+ i
] = ir
->name
;
1043 case ir_var_temporary
:
1044 reg
= new(mem_ctx
) dst_reg(this, ir
->type
);
1047 case ir_var_uniform
:
1048 reg
= new(this->mem_ctx
) dst_reg(UNIFORM
, this->uniforms
);
1050 /* Thanks to the lower_ubo_reference pass, we will see only
1051 * ir_binop_ubo_load expressions and not ir_dereference_variable for UBO
1052 * variables, so no need for them to be in variable_ht.
1054 if (ir
->uniform_block
!= -1)
1057 /* Track how big the whole uniform variable is, in case we need to put a
1058 * copy of its data into pull constants for array access.
1060 this->uniform_size
[this->uniforms
] = type_size(ir
->type
);
1062 if (!strncmp(ir
->name
, "gl_", 3)) {
1063 setup_builtin_uniform_values(ir
);
1065 setup_uniform_values(ir
);
1069 case ir_var_system_value
:
1070 /* VertexID is stored by the VF as the last vertex element, but
1071 * we don't represent it with a flag in inputs_read, so we call
1072 * it VERT_ATTRIB_MAX, which setup_attributes() picks up on.
1074 reg
= new(mem_ctx
) dst_reg(ATTR
, VERT_ATTRIB_MAX
);
1075 prog_data
->uses_vertexid
= true;
1077 switch (ir
->location
) {
1078 case SYSTEM_VALUE_VERTEX_ID
:
1079 reg
->writemask
= WRITEMASK_X
;
1081 case SYSTEM_VALUE_INSTANCE_ID
:
1082 reg
->writemask
= WRITEMASK_Y
;
1085 assert(!"not reached");
1091 assert(!"not reached");
1094 reg
->type
= brw_type_for_base_type(ir
->type
);
1095 hash_table_insert(this->variable_ht
, reg
, ir
);
1099 vec4_visitor::visit(ir_loop
*ir
)
1103 /* We don't want debugging output to print the whole body of the
1104 * loop as the annotation.
1106 this->base_ir
= NULL
;
1108 if (ir
->counter
!= NULL
) {
1109 this->base_ir
= ir
->counter
;
1110 ir
->counter
->accept(this);
1111 counter
= *(variable_storage(ir
->counter
));
1113 if (ir
->from
!= NULL
) {
1114 this->base_ir
= ir
->from
;
1115 ir
->from
->accept(this);
1117 emit(MOV(counter
, this->result
));
1121 emit(BRW_OPCODE_DO
);
1124 this->base_ir
= ir
->to
;
1125 ir
->to
->accept(this);
1127 emit(CMP(dst_null_d(), src_reg(counter
), this->result
,
1128 brw_conditional_for_comparison(ir
->cmp
)));
1130 vec4_instruction
*inst
= emit(BRW_OPCODE_BREAK
);
1131 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1134 visit_instructions(&ir
->body_instructions
);
1137 if (ir
->increment
) {
1138 this->base_ir
= ir
->increment
;
1139 ir
->increment
->accept(this);
1140 emit(ADD(counter
, src_reg(counter
), this->result
));
1143 emit(BRW_OPCODE_WHILE
);
1147 vec4_visitor::visit(ir_loop_jump
*ir
)
1150 case ir_loop_jump::jump_break
:
1151 emit(BRW_OPCODE_BREAK
);
1153 case ir_loop_jump::jump_continue
:
1154 emit(BRW_OPCODE_CONTINUE
);
1161 vec4_visitor::visit(ir_function_signature
*ir
)
1168 vec4_visitor::visit(ir_function
*ir
)
1170 /* Ignore function bodies other than main() -- we shouldn't see calls to
1171 * them since they should all be inlined.
1173 if (strcmp(ir
->name
, "main") == 0) {
1174 const ir_function_signature
*sig
;
1177 sig
= ir
->matching_signature(&empty
);
1181 visit_instructions(&sig
->body
);
1186 vec4_visitor::try_emit_sat(ir_expression
*ir
)
1188 ir_rvalue
*sat_src
= ir
->as_rvalue_to_saturate();
1192 sat_src
->accept(this);
1193 src_reg src
= this->result
;
1195 this->result
= src_reg(this, ir
->type
);
1196 vec4_instruction
*inst
;
1197 inst
= emit(MOV(dst_reg(this->result
), src
));
1198 inst
->saturate
= true;
1204 vec4_visitor::emit_bool_comparison(unsigned int op
,
1205 dst_reg dst
, src_reg src0
, src_reg src1
)
1207 /* original gen4 does destination conversion before comparison. */
1209 dst
.type
= src0
.type
;
1211 emit(CMP(dst
, src0
, src1
, brw_conditional_for_comparison(op
)));
1213 dst
.type
= BRW_REGISTER_TYPE_D
;
1214 emit(AND(dst
, src_reg(dst
), src_reg(0x1)));
1218 vec4_visitor::emit_minmax(uint32_t conditionalmod
, dst_reg dst
,
1219 src_reg src0
, src_reg src1
)
1221 vec4_instruction
*inst
;
1223 if (intel
->gen
>= 6) {
1224 inst
= emit(BRW_OPCODE_SEL
, dst
, src0
, src1
);
1225 inst
->conditional_mod
= conditionalmod
;
1227 emit(CMP(dst
, src0
, src1
, conditionalmod
));
1229 inst
= emit(BRW_OPCODE_SEL
, dst
, src0
, src1
);
1230 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1235 vec4_visitor::visit(ir_expression
*ir
)
1237 unsigned int operand
;
1238 src_reg op
[Elements(ir
->operands
)];
1241 vec4_instruction
*inst
;
1243 if (try_emit_sat(ir
))
1246 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1247 this->result
.file
= BAD_FILE
;
1248 ir
->operands
[operand
]->accept(this);
1249 if (this->result
.file
== BAD_FILE
) {
1250 printf("Failed to get tree for expression operand:\n");
1251 ir
->operands
[operand
]->print();
1254 op
[operand
] = this->result
;
1256 /* Matrix expression operands should have been broken down to vector
1257 * operations already.
1259 assert(!ir
->operands
[operand
]->type
->is_matrix());
1262 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
1263 if (ir
->operands
[1]) {
1264 vector_elements
= MAX2(vector_elements
,
1265 ir
->operands
[1]->type
->vector_elements
);
1268 this->result
.file
= BAD_FILE
;
1270 /* Storage for our result. Ideally for an assignment we'd be using
1271 * the actual storage for the result here, instead.
1273 result_src
= src_reg(this, ir
->type
);
1274 /* convenience for the emit functions below. */
1275 result_dst
= dst_reg(result_src
);
1276 /* If nothing special happens, this is the result. */
1277 this->result
= result_src
;
1278 /* Limit writes to the channels that will be used by result_src later.
1279 * This does limit this temp's use as a temporary for multi-instruction
1282 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1284 switch (ir
->operation
) {
1285 case ir_unop_logic_not
:
1286 /* Note that BRW_OPCODE_NOT is not appropriate here, since it is
1287 * ones complement of the whole register, not just bit 0.
1289 emit(XOR(result_dst
, op
[0], src_reg(1)));
1292 op
[0].negate
= !op
[0].negate
;
1293 this->result
= op
[0];
1297 op
[0].negate
= false;
1298 this->result
= op
[0];
1302 emit(MOV(result_dst
, src_reg(0.0f
)));
1304 emit(CMP(dst_null_d(), op
[0], src_reg(0.0f
), BRW_CONDITIONAL_G
));
1305 inst
= emit(MOV(result_dst
, src_reg(1.0f
)));
1306 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1308 emit(CMP(dst_null_d(), op
[0], src_reg(0.0f
), BRW_CONDITIONAL_L
));
1309 inst
= emit(MOV(result_dst
, src_reg(-1.0f
)));
1310 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1315 emit_math(SHADER_OPCODE_RCP
, result_dst
, op
[0]);
1319 emit_math(SHADER_OPCODE_EXP2
, result_dst
, op
[0]);
1322 emit_math(SHADER_OPCODE_LOG2
, result_dst
, op
[0]);
1326 assert(!"not reached: should be handled by ir_explog_to_explog2");
1329 case ir_unop_sin_reduced
:
1330 emit_math(SHADER_OPCODE_SIN
, result_dst
, op
[0]);
1333 case ir_unop_cos_reduced
:
1334 emit_math(SHADER_OPCODE_COS
, result_dst
, op
[0]);
1339 assert(!"derivatives not valid in vertex shader");
1343 assert(!"not reached: should be handled by lower_noise");
1347 emit(ADD(result_dst
, op
[0], op
[1]));
1350 assert(!"not reached: should be handled by ir_sub_to_add_neg");
1354 if (ir
->type
->is_integer()) {
1355 /* For integer multiplication, the MUL uses the low 16 bits
1356 * of one of the operands (src0 on gen6, src1 on gen7). The
1357 * MACH accumulates in the contribution of the upper 16 bits
1360 * FINISHME: Emit just the MUL if we know an operand is small
1363 struct brw_reg acc
= retype(brw_acc_reg(), BRW_REGISTER_TYPE_D
);
1365 emit(MUL(acc
, op
[0], op
[1]));
1366 emit(MACH(dst_null_d(), op
[0], op
[1]));
1367 emit(MOV(result_dst
, src_reg(acc
)));
1369 emit(MUL(result_dst
, op
[0], op
[1]));
1373 /* Floating point should be lowered by DIV_TO_MUL_RCP in the compiler. */
1374 assert(ir
->type
->is_integer());
1375 emit_math(SHADER_OPCODE_INT_QUOTIENT
, result_dst
, op
[0], op
[1]);
1378 /* Floating point should be lowered by MOD_TO_FRACT in the compiler. */
1379 assert(ir
->type
->is_integer());
1380 emit_math(SHADER_OPCODE_INT_REMAINDER
, result_dst
, op
[0], op
[1]);
1384 case ir_binop_greater
:
1385 case ir_binop_lequal
:
1386 case ir_binop_gequal
:
1387 case ir_binop_equal
:
1388 case ir_binop_nequal
: {
1389 emit(CMP(result_dst
, op
[0], op
[1],
1390 brw_conditional_for_comparison(ir
->operation
)));
1391 emit(AND(result_dst
, result_src
, src_reg(0x1)));
1395 case ir_binop_all_equal
:
1396 /* "==" operator producing a scalar boolean. */
1397 if (ir
->operands
[0]->type
->is_vector() ||
1398 ir
->operands
[1]->type
->is_vector()) {
1399 emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_Z
));
1400 emit(MOV(result_dst
, src_reg(0)));
1401 inst
= emit(MOV(result_dst
, src_reg(1)));
1402 inst
->predicate
= BRW_PREDICATE_ALIGN16_ALL4H
;
1404 emit(CMP(result_dst
, op
[0], op
[1], BRW_CONDITIONAL_Z
));
1405 emit(AND(result_dst
, result_src
, src_reg(0x1)));
1408 case ir_binop_any_nequal
:
1409 /* "!=" operator producing a scalar boolean. */
1410 if (ir
->operands
[0]->type
->is_vector() ||
1411 ir
->operands
[1]->type
->is_vector()) {
1412 emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_NZ
));
1414 emit(MOV(result_dst
, src_reg(0)));
1415 inst
= emit(MOV(result_dst
, src_reg(1)));
1416 inst
->predicate
= BRW_PREDICATE_ALIGN16_ANY4H
;
1418 emit(CMP(result_dst
, op
[0], op
[1], BRW_CONDITIONAL_NZ
));
1419 emit(AND(result_dst
, result_src
, src_reg(0x1)));
1424 emit(CMP(dst_null_d(), op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
1425 emit(MOV(result_dst
, src_reg(0)));
1427 inst
= emit(MOV(result_dst
, src_reg(1)));
1428 inst
->predicate
= BRW_PREDICATE_ALIGN16_ANY4H
;
1431 case ir_binop_logic_xor
:
1432 emit(XOR(result_dst
, op
[0], op
[1]));
1435 case ir_binop_logic_or
:
1436 emit(OR(result_dst
, op
[0], op
[1]));
1439 case ir_binop_logic_and
:
1440 emit(AND(result_dst
, op
[0], op
[1]));
1444 assert(ir
->operands
[0]->type
->is_vector());
1445 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1446 emit_dp(result_dst
, op
[0], op
[1], ir
->operands
[0]->type
->vector_elements
);
1450 emit_math(SHADER_OPCODE_SQRT
, result_dst
, op
[0]);
1453 emit_math(SHADER_OPCODE_RSQ
, result_dst
, op
[0]);
1456 case ir_unop_bitcast_i2f
:
1457 case ir_unop_bitcast_u2f
:
1458 this->result
= op
[0];
1459 this->result
.type
= BRW_REGISTER_TYPE_F
;
1462 case ir_unop_bitcast_f2i
:
1463 this->result
= op
[0];
1464 this->result
.type
= BRW_REGISTER_TYPE_D
;
1467 case ir_unop_bitcast_f2u
:
1468 this->result
= op
[0];
1469 this->result
.type
= BRW_REGISTER_TYPE_UD
;
1480 emit(MOV(result_dst
, op
[0]));
1484 emit(CMP(result_dst
, op
[0], src_reg(0.0f
), BRW_CONDITIONAL_NZ
));
1485 emit(AND(result_dst
, result_src
, src_reg(1)));
1490 emit(RNDZ(result_dst
, op
[0]));
1493 op
[0].negate
= !op
[0].negate
;
1494 inst
= emit(RNDD(result_dst
, op
[0]));
1495 this->result
.negate
= true;
1498 inst
= emit(RNDD(result_dst
, op
[0]));
1501 inst
= emit(FRC(result_dst
, op
[0]));
1503 case ir_unop_round_even
:
1504 emit(RNDE(result_dst
, op
[0]));
1508 emit_minmax(BRW_CONDITIONAL_L
, result_dst
, op
[0], op
[1]);
1511 emit_minmax(BRW_CONDITIONAL_G
, result_dst
, op
[0], op
[1]);
1515 emit_math(SHADER_OPCODE_POW
, result_dst
, op
[0], op
[1]);
1518 case ir_unop_bit_not
:
1519 inst
= emit(NOT(result_dst
, op
[0]));
1521 case ir_binop_bit_and
:
1522 inst
= emit(AND(result_dst
, op
[0], op
[1]));
1524 case ir_binop_bit_xor
:
1525 inst
= emit(XOR(result_dst
, op
[0], op
[1]));
1527 case ir_binop_bit_or
:
1528 inst
= emit(OR(result_dst
, op
[0], op
[1]));
1531 case ir_binop_lshift
:
1532 inst
= emit(SHL(result_dst
, op
[0], op
[1]));
1535 case ir_binop_rshift
:
1536 if (ir
->type
->base_type
== GLSL_TYPE_INT
)
1537 inst
= emit(ASR(result_dst
, op
[0], op
[1]));
1539 inst
= emit(SHR(result_dst
, op
[0], op
[1]));
1542 case ir_binop_ubo_load
: {
1543 ir_constant
*uniform_block
= ir
->operands
[0]->as_constant();
1544 ir_constant
*const_offset_ir
= ir
->operands
[1]->as_constant();
1545 unsigned const_offset
= const_offset_ir
? const_offset_ir
->value
.u
[0] : 0;
1546 src_reg offset
= op
[1];
1548 /* Now, load the vector from that offset. */
1549 assert(ir
->type
->is_vector() || ir
->type
->is_scalar());
1551 src_reg packed_consts
= src_reg(this, glsl_type::vec4_type
);
1552 packed_consts
.type
= result
.type
;
1553 src_reg surf_index
=
1554 src_reg(SURF_INDEX_VS_UBO(uniform_block
->value
.u
[0]));
1555 if (const_offset_ir
) {
1556 offset
= src_reg(const_offset
/ 16);
1558 emit(SHR(dst_reg(offset
), offset
, src_reg(4)));
1561 vec4_instruction
*pull
=
1562 emit(new(mem_ctx
) vec4_instruction(this,
1563 VS_OPCODE_PULL_CONSTANT_LOAD
,
1564 dst_reg(packed_consts
),
1567 pull
->base_mrf
= 14;
1570 packed_consts
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1571 packed_consts
.swizzle
+= BRW_SWIZZLE4(const_offset
% 16 / 4,
1572 const_offset
% 16 / 4,
1573 const_offset
% 16 / 4,
1574 const_offset
% 16 / 4);
1576 /* UBO bools are any nonzero int. We store bools as either 0 or 1. */
1577 if (ir
->type
->base_type
== GLSL_TYPE_BOOL
) {
1578 emit(CMP(result_dst
, packed_consts
, src_reg(0u),
1579 BRW_CONDITIONAL_NZ
));
1580 emit(AND(result_dst
, result
, src_reg(0x1)));
1582 emit(MOV(result_dst
, packed_consts
));
1587 case ir_quadop_vector
:
1588 assert(!"not reached: should be handled by lower_quadop_vector");
1591 case ir_unop_pack_half_2x16
:
1592 emit_pack_half_2x16(result_dst
, op
[0]);
1594 case ir_unop_unpack_half_2x16
:
1595 emit_unpack_half_2x16(result_dst
, op
[0]);
1597 case ir_unop_pack_snorm_2x16
:
1598 case ir_unop_pack_unorm_2x16
:
1599 case ir_unop_unpack_snorm_2x16
:
1600 case ir_unop_unpack_unorm_2x16
:
1601 assert(!"not reached: should be handled by lower_packing_builtins");
1603 case ir_unop_unpack_half_2x16_split_x
:
1604 case ir_unop_unpack_half_2x16_split_y
:
1605 case ir_binop_pack_half_2x16_split
:
1606 assert(!"not reached: should not occur in vertex shader");
1613 vec4_visitor::visit(ir_swizzle
*ir
)
1619 /* Note that this is only swizzles in expressions, not those on the left
1620 * hand side of an assignment, which do write masking. See ir_assignment
1624 ir
->val
->accept(this);
1626 assert(src
.file
!= BAD_FILE
);
1628 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1631 swizzle
[i
] = BRW_GET_SWZ(src
.swizzle
, ir
->mask
.x
);
1634 swizzle
[i
] = BRW_GET_SWZ(src
.swizzle
, ir
->mask
.y
);
1637 swizzle
[i
] = BRW_GET_SWZ(src
.swizzle
, ir
->mask
.z
);
1640 swizzle
[i
] = BRW_GET_SWZ(src
.swizzle
, ir
->mask
.w
);
1644 for (; i
< 4; i
++) {
1645 /* Replicate the last channel out. */
1646 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
1649 src
.swizzle
= BRW_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
1655 vec4_visitor::visit(ir_dereference_variable
*ir
)
1657 const struct glsl_type
*type
= ir
->type
;
1658 dst_reg
*reg
= variable_storage(ir
->var
);
1661 fail("Failed to find variable storage for %s\n", ir
->var
->name
);
1662 this->result
= src_reg(brw_null_reg());
1666 this->result
= src_reg(*reg
);
1668 /* System values get their swizzle from the dst_reg writemask */
1669 if (ir
->var
->mode
== ir_var_system_value
)
1672 if (type
->is_scalar() || type
->is_vector() || type
->is_matrix())
1673 this->result
.swizzle
= swizzle_for_size(type
->vector_elements
);
1677 vec4_visitor::visit(ir_dereference_array
*ir
)
1679 ir_constant
*constant_index
;
1681 int element_size
= type_size(ir
->type
);
1683 constant_index
= ir
->array_index
->constant_expression_value();
1685 ir
->array
->accept(this);
1688 if (constant_index
) {
1689 src
.reg_offset
+= constant_index
->value
.i
[0] * element_size
;
1691 /* Variable index array dereference. It eats the "vec4" of the
1692 * base of the array and an index that offsets the Mesa register
1695 ir
->array_index
->accept(this);
1699 if (element_size
== 1) {
1700 index_reg
= this->result
;
1702 index_reg
= src_reg(this, glsl_type::int_type
);
1704 emit(MUL(dst_reg(index_reg
), this->result
, src_reg(element_size
)));
1708 src_reg temp
= src_reg(this, glsl_type::int_type
);
1710 emit(ADD(dst_reg(temp
), *src
.reladdr
, index_reg
));
1715 src
.reladdr
= ralloc(mem_ctx
, src_reg
);
1716 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
1719 /* If the type is smaller than a vec4, replicate the last channel out. */
1720 if (ir
->type
->is_scalar() || ir
->type
->is_vector() || ir
->type
->is_matrix())
1721 src
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1723 src
.swizzle
= BRW_SWIZZLE_NOOP
;
1724 src
.type
= brw_type_for_base_type(ir
->type
);
1730 vec4_visitor::visit(ir_dereference_record
*ir
)
1733 const glsl_type
*struct_type
= ir
->record
->type
;
1736 ir
->record
->accept(this);
1738 for (i
= 0; i
< struct_type
->length
; i
++) {
1739 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
1741 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
1744 /* If the type is smaller than a vec4, replicate the last channel out. */
1745 if (ir
->type
->is_scalar() || ir
->type
->is_vector() || ir
->type
->is_matrix())
1746 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1748 this->result
.swizzle
= BRW_SWIZZLE_NOOP
;
1749 this->result
.type
= brw_type_for_base_type(ir
->type
);
1751 this->result
.reg_offset
+= offset
;
1755 * We want to be careful in assignment setup to hit the actual storage
1756 * instead of potentially using a temporary like we might with the
1757 * ir_dereference handler.
1760 get_assignment_lhs(ir_dereference
*ir
, vec4_visitor
*v
)
1762 /* The LHS must be a dereference. If the LHS is a variable indexed array
1763 * access of a vector, it must be separated into a series conditional moves
1764 * before reaching this point (see ir_vec_index_to_cond_assign).
1766 assert(ir
->as_dereference());
1767 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
1769 assert(!deref_array
->array
->type
->is_vector());
1772 /* Use the rvalue deref handler for the most part. We'll ignore
1773 * swizzles in it and write swizzles using writemask, though.
1776 return dst_reg(v
->result
);
1780 vec4_visitor::emit_block_move(dst_reg
*dst
, src_reg
*src
,
1781 const struct glsl_type
*type
, uint32_t predicate
)
1783 if (type
->base_type
== GLSL_TYPE_STRUCT
) {
1784 for (unsigned int i
= 0; i
< type
->length
; i
++) {
1785 emit_block_move(dst
, src
, type
->fields
.structure
[i
].type
, predicate
);
1790 if (type
->is_array()) {
1791 for (unsigned int i
= 0; i
< type
->length
; i
++) {
1792 emit_block_move(dst
, src
, type
->fields
.array
, predicate
);
1797 if (type
->is_matrix()) {
1798 const struct glsl_type
*vec_type
;
1800 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
1801 type
->vector_elements
, 1);
1803 for (int i
= 0; i
< type
->matrix_columns
; i
++) {
1804 emit_block_move(dst
, src
, vec_type
, predicate
);
1809 assert(type
->is_scalar() || type
->is_vector());
1811 dst
->type
= brw_type_for_base_type(type
);
1812 src
->type
= dst
->type
;
1814 dst
->writemask
= (1 << type
->vector_elements
) - 1;
1816 src
->swizzle
= swizzle_for_size(type
->vector_elements
);
1818 vec4_instruction
*inst
= emit(MOV(*dst
, *src
));
1819 inst
->predicate
= predicate
;
1826 /* If the RHS processing resulted in an instruction generating a
1827 * temporary value, and it would be easy to rewrite the instruction to
1828 * generate its result right into the LHS instead, do so. This ends
1829 * up reliably removing instructions where it can be tricky to do so
1830 * later without real UD chain information.
1833 vec4_visitor::try_rewrite_rhs_to_dst(ir_assignment
*ir
,
1836 vec4_instruction
*pre_rhs_inst
,
1837 vec4_instruction
*last_rhs_inst
)
1839 /* This could be supported, but it would take more smarts. */
1843 if (pre_rhs_inst
== last_rhs_inst
)
1844 return false; /* No instructions generated to work with. */
1846 /* Make sure the last instruction generated our source reg. */
1847 if (src
.file
!= GRF
||
1848 src
.file
!= last_rhs_inst
->dst
.file
||
1849 src
.reg
!= last_rhs_inst
->dst
.reg
||
1850 src
.reg_offset
!= last_rhs_inst
->dst
.reg_offset
||
1854 last_rhs_inst
->predicate
!= BRW_PREDICATE_NONE
)
1857 /* Check that that last instruction fully initialized the channels
1858 * we want to use, in the order we want to use them. We could
1859 * potentially reswizzle the operands of many instructions so that
1860 * we could handle out of order channels, but don't yet.
1863 for (unsigned i
= 0; i
< 4; i
++) {
1864 if (dst
.writemask
& (1 << i
)) {
1865 if (!(last_rhs_inst
->dst
.writemask
& (1 << i
)))
1868 if (BRW_GET_SWZ(src
.swizzle
, i
) != i
)
1873 /* Success! Rewrite the instruction. */
1874 last_rhs_inst
->dst
.file
= dst
.file
;
1875 last_rhs_inst
->dst
.reg
= dst
.reg
;
1876 last_rhs_inst
->dst
.reg_offset
= dst
.reg_offset
;
1877 last_rhs_inst
->dst
.reladdr
= dst
.reladdr
;
1878 last_rhs_inst
->dst
.writemask
&= dst
.writemask
;
1884 vec4_visitor::visit(ir_assignment
*ir
)
1886 dst_reg dst
= get_assignment_lhs(ir
->lhs
, this);
1887 uint32_t predicate
= BRW_PREDICATE_NONE
;
1889 if (!ir
->lhs
->type
->is_scalar() &&
1890 !ir
->lhs
->type
->is_vector()) {
1891 ir
->rhs
->accept(this);
1892 src_reg src
= this->result
;
1894 if (ir
->condition
) {
1895 emit_bool_to_cond_code(ir
->condition
, &predicate
);
1898 /* emit_block_move doesn't account for swizzles in the source register.
1899 * This should be ok, since the source register is a structure or an
1900 * array, and those can't be swizzled. But double-check to be sure.
1902 assert(src
.swizzle
==
1903 (ir
->rhs
->type
->is_matrix()
1904 ? swizzle_for_size(ir
->rhs
->type
->vector_elements
)
1905 : BRW_SWIZZLE_NOOP
));
1907 emit_block_move(&dst
, &src
, ir
->rhs
->type
, predicate
);
1911 /* Now we're down to just a scalar/vector with writemasks. */
1914 vec4_instruction
*pre_rhs_inst
, *last_rhs_inst
;
1915 pre_rhs_inst
= (vec4_instruction
*)this->instructions
.get_tail();
1917 ir
->rhs
->accept(this);
1919 last_rhs_inst
= (vec4_instruction
*)this->instructions
.get_tail();
1921 src_reg src
= this->result
;
1924 int first_enabled_chan
= 0;
1927 assert(ir
->lhs
->type
->is_vector() ||
1928 ir
->lhs
->type
->is_scalar());
1929 dst
.writemask
= ir
->write_mask
;
1931 for (int i
= 0; i
< 4; i
++) {
1932 if (dst
.writemask
& (1 << i
)) {
1933 first_enabled_chan
= BRW_GET_SWZ(src
.swizzle
, i
);
1938 /* Swizzle a small RHS vector into the channels being written.
1940 * glsl ir treats write_mask as dictating how many channels are
1941 * present on the RHS while in our instructions we need to make
1942 * those channels appear in the slots of the vec4 they're written to.
1944 for (int i
= 0; i
< 4; i
++) {
1945 if (dst
.writemask
& (1 << i
))
1946 swizzles
[i
] = BRW_GET_SWZ(src
.swizzle
, src_chan
++);
1948 swizzles
[i
] = first_enabled_chan
;
1950 src
.swizzle
= BRW_SWIZZLE4(swizzles
[0], swizzles
[1],
1951 swizzles
[2], swizzles
[3]);
1953 if (try_rewrite_rhs_to_dst(ir
, dst
, src
, pre_rhs_inst
, last_rhs_inst
)) {
1957 if (ir
->condition
) {
1958 emit_bool_to_cond_code(ir
->condition
, &predicate
);
1961 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
1962 vec4_instruction
*inst
= emit(MOV(dst
, src
));
1963 inst
->predicate
= predicate
;
1971 vec4_visitor::emit_constant_values(dst_reg
*dst
, ir_constant
*ir
)
1973 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
1974 foreach_list(node
, &ir
->components
) {
1975 ir_constant
*field_value
= (ir_constant
*)node
;
1977 emit_constant_values(dst
, field_value
);
1982 if (ir
->type
->is_array()) {
1983 for (unsigned int i
= 0; i
< ir
->type
->length
; i
++) {
1984 emit_constant_values(dst
, ir
->array_elements
[i
]);
1989 if (ir
->type
->is_matrix()) {
1990 for (int i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
1991 float *vec
= &ir
->value
.f
[i
* ir
->type
->vector_elements
];
1993 for (int j
= 0; j
< ir
->type
->vector_elements
; j
++) {
1994 dst
->writemask
= 1 << j
;
1995 dst
->type
= BRW_REGISTER_TYPE_F
;
1997 emit(MOV(*dst
, src_reg(vec
[j
])));
2004 int remaining_writemask
= (1 << ir
->type
->vector_elements
) - 1;
2006 for (int i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2007 if (!(remaining_writemask
& (1 << i
)))
2010 dst
->writemask
= 1 << i
;
2011 dst
->type
= brw_type_for_base_type(ir
->type
);
2013 /* Find other components that match the one we're about to
2014 * write. Emits fewer instructions for things like vec4(0.5,
2017 for (int j
= i
+ 1; j
< ir
->type
->vector_elements
; j
++) {
2018 if (ir
->type
->base_type
== GLSL_TYPE_BOOL
) {
2019 if (ir
->value
.b
[i
] == ir
->value
.b
[j
])
2020 dst
->writemask
|= (1 << j
);
2022 /* u, i, and f storage all line up, so no need for a
2023 * switch case for comparing each type.
2025 if (ir
->value
.u
[i
] == ir
->value
.u
[j
])
2026 dst
->writemask
|= (1 << j
);
2030 switch (ir
->type
->base_type
) {
2031 case GLSL_TYPE_FLOAT
:
2032 emit(MOV(*dst
, src_reg(ir
->value
.f
[i
])));
2035 emit(MOV(*dst
, src_reg(ir
->value
.i
[i
])));
2037 case GLSL_TYPE_UINT
:
2038 emit(MOV(*dst
, src_reg(ir
->value
.u
[i
])));
2040 case GLSL_TYPE_BOOL
:
2041 emit(MOV(*dst
, src_reg(ir
->value
.b
[i
])));
2044 assert(!"Non-float/uint/int/bool constant");
2048 remaining_writemask
&= ~dst
->writemask
;
2054 vec4_visitor::visit(ir_constant
*ir
)
2056 dst_reg dst
= dst_reg(this, ir
->type
);
2057 this->result
= src_reg(dst
);
2059 emit_constant_values(&dst
, ir
);
2063 vec4_visitor::visit(ir_call
*ir
)
2065 assert(!"not reached");
2069 vec4_visitor::visit(ir_texture
*ir
)
2071 int sampler
= _mesa_get_sampler_uniform_value(ir
->sampler
, prog
, &vp
->Base
);
2073 /* Should be lowered by do_lower_texture_projection */
2074 assert(!ir
->projector
);
2076 /* Generate code to compute all the subexpression trees. This has to be
2077 * done before loading any values into MRFs for the sampler message since
2078 * generating these values may involve SEND messages that need the MRFs.
2081 if (ir
->coordinate
) {
2082 ir
->coordinate
->accept(this);
2083 coordinate
= this->result
;
2086 src_reg shadow_comparitor
;
2087 if (ir
->shadow_comparitor
) {
2088 ir
->shadow_comparitor
->accept(this);
2089 shadow_comparitor
= this->result
;
2092 const glsl_type
*lod_type
;
2093 src_reg lod
, dPdx
, dPdy
;
2096 lod
= src_reg(0.0f
);
2097 lod_type
= glsl_type::float_type
;
2102 ir
->lod_info
.lod
->accept(this);
2104 lod_type
= ir
->lod_info
.lod
->type
;
2107 ir
->lod_info
.grad
.dPdx
->accept(this);
2108 dPdx
= this->result
;
2110 ir
->lod_info
.grad
.dPdy
->accept(this);
2111 dPdy
= this->result
;
2113 lod_type
= ir
->lod_info
.grad
.dPdx
->type
;
2119 vec4_instruction
*inst
= NULL
;
2123 inst
= new(mem_ctx
) vec4_instruction(this, SHADER_OPCODE_TXL
);
2126 inst
= new(mem_ctx
) vec4_instruction(this, SHADER_OPCODE_TXD
);
2129 inst
= new(mem_ctx
) vec4_instruction(this, SHADER_OPCODE_TXF
);
2132 inst
= new(mem_ctx
) vec4_instruction(this, SHADER_OPCODE_TXS
);
2135 assert(!"TXB is not valid for vertex shaders.");
2138 bool use_texture_offset
= ir
->offset
!= NULL
&& ir
->op
!= ir_txf
;
2140 /* Texel offsets go in the message header; Gen4 also requires headers. */
2141 inst
->header_present
= use_texture_offset
|| intel
->gen
< 5;
2143 inst
->mlen
= inst
->header_present
+ 1; /* always at least one */
2144 inst
->sampler
= sampler
;
2145 inst
->dst
= dst_reg(this, ir
->type
);
2146 inst
->dst
.writemask
= WRITEMASK_XYZW
;
2147 inst
->shadow_compare
= ir
->shadow_comparitor
!= NULL
;
2149 if (use_texture_offset
)
2150 inst
->texture_offset
= brw_texture_offset(ir
->offset
->as_constant());
2152 /* MRF for the first parameter */
2153 int param_base
= inst
->base_mrf
+ inst
->header_present
;
2155 if (ir
->op
== ir_txs
) {
2156 int writemask
= intel
->gen
== 4 ? WRITEMASK_W
: WRITEMASK_X
;
2157 emit(MOV(dst_reg(MRF
, param_base
, lod_type
, writemask
), lod
));
2159 int i
, coord_mask
= 0, zero_mask
= 0;
2160 /* Load the coordinate */
2161 /* FINISHME: gl_clamp_mask and saturate */
2162 for (i
= 0; i
< ir
->coordinate
->type
->vector_elements
; i
++)
2163 coord_mask
|= (1 << i
);
2165 zero_mask
|= (1 << i
);
2167 if (ir
->offset
&& ir
->op
== ir_txf
) {
2168 /* It appears that the ld instruction used for txf does its
2169 * address bounds check before adding in the offset. To work
2170 * around this, just add the integer offset to the integer
2171 * texel coordinate, and don't put the offset in the header.
2173 ir_constant
*offset
= ir
->offset
->as_constant();
2176 for (int j
= 0; j
< ir
->coordinate
->type
->vector_elements
; j
++) {
2177 src_reg src
= coordinate
;
2178 src
.swizzle
= BRW_SWIZZLE4(BRW_GET_SWZ(src
.swizzle
, j
),
2179 BRW_GET_SWZ(src
.swizzle
, j
),
2180 BRW_GET_SWZ(src
.swizzle
, j
),
2181 BRW_GET_SWZ(src
.swizzle
, j
));
2182 emit(ADD(dst_reg(MRF
, param_base
, ir
->coordinate
->type
, 1 << j
),
2183 src
, offset
->value
.i
[j
]));
2186 emit(MOV(dst_reg(MRF
, param_base
, ir
->coordinate
->type
, coord_mask
),
2189 emit(MOV(dst_reg(MRF
, param_base
, ir
->coordinate
->type
, zero_mask
),
2191 /* Load the shadow comparitor */
2192 if (ir
->shadow_comparitor
) {
2193 emit(MOV(dst_reg(MRF
, param_base
+ 1, ir
->shadow_comparitor
->type
,
2195 shadow_comparitor
));
2199 /* Load the LOD info */
2200 if (ir
->op
== ir_tex
|| ir
->op
== ir_txl
) {
2202 if (intel
->gen
>= 5) {
2203 mrf
= param_base
+ 1;
2204 if (ir
->shadow_comparitor
) {
2205 writemask
= WRITEMASK_Y
;
2206 /* mlen already incremented */
2208 writemask
= WRITEMASK_X
;
2211 } else /* intel->gen == 4 */ {
2213 writemask
= WRITEMASK_Z
;
2215 emit(MOV(dst_reg(MRF
, mrf
, lod_type
, writemask
), lod
));
2216 } else if (ir
->op
== ir_txf
) {
2217 emit(MOV(dst_reg(MRF
, param_base
, lod_type
, WRITEMASK_W
),
2219 } else if (ir
->op
== ir_txd
) {
2220 const glsl_type
*type
= lod_type
;
2222 if (intel
->gen
>= 5) {
2223 dPdx
.swizzle
= BRW_SWIZZLE4(SWIZZLE_X
,SWIZZLE_X
,SWIZZLE_Y
,SWIZZLE_Y
);
2224 dPdy
.swizzle
= BRW_SWIZZLE4(SWIZZLE_X
,SWIZZLE_X
,SWIZZLE_Y
,SWIZZLE_Y
);
2225 emit(MOV(dst_reg(MRF
, param_base
+ 1, type
, WRITEMASK_XZ
), dPdx
));
2226 emit(MOV(dst_reg(MRF
, param_base
+ 1, type
, WRITEMASK_YW
), dPdy
));
2229 if (ir
->type
->vector_elements
== 3) {
2230 dPdx
.swizzle
= BRW_SWIZZLE_ZZZZ
;
2231 dPdy
.swizzle
= BRW_SWIZZLE_ZZZZ
;
2232 emit(MOV(dst_reg(MRF
, param_base
+ 2, type
, WRITEMASK_X
), dPdx
));
2233 emit(MOV(dst_reg(MRF
, param_base
+ 2, type
, WRITEMASK_Y
), dPdy
));
2236 } else /* intel->gen == 4 */ {
2237 emit(MOV(dst_reg(MRF
, param_base
+ 1, type
, WRITEMASK_XYZ
), dPdx
));
2238 emit(MOV(dst_reg(MRF
, param_base
+ 2, type
, WRITEMASK_XYZ
), dPdy
));
2246 /* fixup num layers (z) for cube arrays: hardware returns faces * layers;
2247 * spec requires layers.
2249 if (ir
->op
== ir_txs
) {
2250 glsl_type
const *type
= ir
->sampler
->type
;
2251 if (type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_CUBE
&&
2252 type
->sampler_array
) {
2253 emit_math(SHADER_OPCODE_INT_QUOTIENT
,
2254 with_writemask(inst
->dst
, WRITEMASK_Z
),
2255 src_reg(inst
->dst
), src_reg(6));
2259 swizzle_result(ir
, src_reg(inst
->dst
), sampler
);
2263 vec4_visitor::swizzle_result(ir_texture
*ir
, src_reg orig_val
, int sampler
)
2265 int s
= c
->key
.tex
.swizzles
[sampler
];
2267 this->result
= src_reg(this, ir
->type
);
2268 dst_reg
swizzled_result(this->result
);
2270 if (ir
->op
== ir_txs
|| ir
->type
== glsl_type::float_type
2271 || s
== SWIZZLE_NOOP
) {
2272 emit(MOV(swizzled_result
, orig_val
));
2276 int zero_mask
= 0, one_mask
= 0, copy_mask
= 0;
2279 for (int i
= 0; i
< 4; i
++) {
2280 switch (GET_SWZ(s
, i
)) {
2282 zero_mask
|= (1 << i
);
2285 one_mask
|= (1 << i
);
2288 copy_mask
|= (1 << i
);
2289 swizzle
[i
] = GET_SWZ(s
, i
);
2295 orig_val
.swizzle
= BRW_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
2296 swizzled_result
.writemask
= copy_mask
;
2297 emit(MOV(swizzled_result
, orig_val
));
2301 swizzled_result
.writemask
= zero_mask
;
2302 emit(MOV(swizzled_result
, src_reg(0.0f
)));
2306 swizzled_result
.writemask
= one_mask
;
2307 emit(MOV(swizzled_result
, src_reg(1.0f
)));
2312 vec4_visitor::visit(ir_return
*ir
)
2314 assert(!"not reached");
2318 vec4_visitor::visit(ir_discard
*ir
)
2320 assert(!"not reached");
2324 vec4_visitor::visit(ir_if
*ir
)
2326 /* Don't point the annotation at the if statement, because then it plus
2327 * the then and else blocks get printed.
2329 this->base_ir
= ir
->condition
;
2331 if (intel
->gen
== 6) {
2335 emit_bool_to_cond_code(ir
->condition
, &predicate
);
2336 emit(IF(predicate
));
2339 visit_instructions(&ir
->then_instructions
);
2341 if (!ir
->else_instructions
.is_empty()) {
2342 this->base_ir
= ir
->condition
;
2343 emit(BRW_OPCODE_ELSE
);
2345 visit_instructions(&ir
->else_instructions
);
2348 this->base_ir
= ir
->condition
;
2349 emit(BRW_OPCODE_ENDIF
);
2353 vec4_visitor::emit_ndc_computation()
2355 /* Get the position */
2356 src_reg pos
= src_reg(output_reg
[VERT_RESULT_HPOS
]);
2358 /* Build ndc coords, which are (x/w, y/w, z/w, 1/w) */
2359 dst_reg ndc
= dst_reg(this, glsl_type::vec4_type
);
2360 output_reg
[BRW_VERT_RESULT_NDC
] = ndc
;
2362 current_annotation
= "NDC";
2363 dst_reg ndc_w
= ndc
;
2364 ndc_w
.writemask
= WRITEMASK_W
;
2365 src_reg pos_w
= pos
;
2366 pos_w
.swizzle
= BRW_SWIZZLE4(SWIZZLE_W
, SWIZZLE_W
, SWIZZLE_W
, SWIZZLE_W
);
2367 emit_math(SHADER_OPCODE_RCP
, ndc_w
, pos_w
);
2369 dst_reg ndc_xyz
= ndc
;
2370 ndc_xyz
.writemask
= WRITEMASK_XYZ
;
2372 emit(MUL(ndc_xyz
, pos
, src_reg(ndc_w
)));
2376 vec4_visitor::emit_psiz_and_flags(struct brw_reg reg
)
2378 if (intel
->gen
< 6 &&
2379 ((c
->prog_data
.outputs_written
& BITFIELD64_BIT(VERT_RESULT_PSIZ
)) ||
2380 c
->key
.userclip_active
|| brw
->has_negative_rhw_bug
)) {
2381 dst_reg header1
= dst_reg(this, glsl_type::uvec4_type
);
2382 dst_reg header1_w
= header1
;
2383 header1_w
.writemask
= WRITEMASK_W
;
2386 emit(MOV(header1
, 0u));
2388 if (c
->prog_data
.outputs_written
& BITFIELD64_BIT(VERT_RESULT_PSIZ
)) {
2389 src_reg psiz
= src_reg(output_reg
[VERT_RESULT_PSIZ
]);
2391 current_annotation
= "Point size";
2392 emit(MUL(header1_w
, psiz
, src_reg((float)(1 << 11))));
2393 emit(AND(header1_w
, src_reg(header1_w
), 0x7ff << 8));
2396 current_annotation
= "Clipping flags";
2397 for (i
= 0; i
< c
->key
.nr_userclip_plane_consts
; i
++) {
2398 vec4_instruction
*inst
;
2400 inst
= emit(DP4(dst_null_f(), src_reg(output_reg
[VERT_RESULT_HPOS
]),
2401 src_reg(this->userplane
[i
])));
2402 inst
->conditional_mod
= BRW_CONDITIONAL_L
;
2404 inst
= emit(OR(header1_w
, src_reg(header1_w
), 1u << i
));
2405 inst
->predicate
= BRW_PREDICATE_NORMAL
;
2408 /* i965 clipping workaround:
2409 * 1) Test for -ve rhw
2411 * set ndc = (0,0,0,0)
2414 * Later, clipping will detect ucp[6] and ensure the primitive is
2415 * clipped against all fixed planes.
2417 if (brw
->has_negative_rhw_bug
) {
2421 vec8(brw_null_reg()),
2423 brw_swizzle1(output_reg
[BRW_VERT_RESULT_NDC
], 3),
2426 brw_OR(p
, brw_writemask(header1
, WRITEMASK_W
), header1
, brw_imm_ud(1<<6));
2427 brw_MOV(p
, output_reg
[BRW_VERT_RESULT_NDC
], brw_imm_f(0));
2428 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
2432 emit(MOV(retype(reg
, BRW_REGISTER_TYPE_UD
), src_reg(header1
)));
2433 } else if (intel
->gen
< 6) {
2434 emit(MOV(retype(reg
, BRW_REGISTER_TYPE_UD
), 0u));
2436 emit(MOV(retype(reg
, BRW_REGISTER_TYPE_D
), src_reg(0)));
2437 if (c
->prog_data
.outputs_written
& BITFIELD64_BIT(VERT_RESULT_PSIZ
)) {
2438 emit(MOV(brw_writemask(reg
, WRITEMASK_W
),
2439 src_reg(output_reg
[VERT_RESULT_PSIZ
])));
2445 vec4_visitor::emit_clip_distances(struct brw_reg reg
, int offset
)
2447 if (intel
->gen
< 6) {
2448 /* Clip distance slots are set aside in gen5, but they are not used. It
2449 * is not clear whether we actually need to set aside space for them,
2450 * but the performance cost is negligible.
2455 /* From the GLSL 1.30 spec, section 7.1 (Vertex Shader Special Variables):
2457 * "If a linked set of shaders forming the vertex stage contains no
2458 * static write to gl_ClipVertex or gl_ClipDistance, but the
2459 * application has requested clipping against user clip planes through
2460 * the API, then the coordinate written to gl_Position is used for
2461 * comparison against the user clip planes."
2463 * This function is only called if the shader didn't write to
2464 * gl_ClipDistance. Accordingly, we use gl_ClipVertex to perform clipping
2465 * if the user wrote to it; otherwise we use gl_Position.
2467 gl_vert_result clip_vertex
= VERT_RESULT_CLIP_VERTEX
;
2468 if (!(c
->prog_data
.outputs_written
2469 & BITFIELD64_BIT(VERT_RESULT_CLIP_VERTEX
))) {
2470 clip_vertex
= VERT_RESULT_HPOS
;
2473 for (int i
= 0; i
+ offset
< c
->key
.nr_userclip_plane_consts
&& i
< 4;
2475 emit(DP4(dst_reg(brw_writemask(reg
, 1 << i
)),
2476 src_reg(output_reg
[clip_vertex
]),
2477 src_reg(this->userplane
[i
+ offset
])));
2482 vec4_visitor::emit_generic_urb_slot(dst_reg reg
, int vert_result
)
2484 assert (vert_result
< VERT_RESULT_MAX
);
2485 reg
.type
= output_reg
[vert_result
].type
;
2486 current_annotation
= output_reg_annotation
[vert_result
];
2487 /* Copy the register, saturating if necessary */
2488 vec4_instruction
*inst
= emit(MOV(reg
,
2489 src_reg(output_reg
[vert_result
])));
2490 if ((vert_result
== VERT_RESULT_COL0
||
2491 vert_result
== VERT_RESULT_COL1
||
2492 vert_result
== VERT_RESULT_BFC0
||
2493 vert_result
== VERT_RESULT_BFC1
) &&
2494 c
->key
.clamp_vertex_color
) {
2495 inst
->saturate
= true;
2500 vec4_visitor::emit_urb_slot(int mrf
, int vert_result
)
2502 struct brw_reg hw_reg
= brw_message_reg(mrf
);
2503 dst_reg reg
= dst_reg(MRF
, mrf
);
2504 reg
.type
= BRW_REGISTER_TYPE_F
;
2506 switch (vert_result
) {
2507 case VERT_RESULT_PSIZ
:
2508 /* PSIZ is always in slot 0, and is coupled with other flags. */
2509 current_annotation
= "indices, point width, clip flags";
2510 emit_psiz_and_flags(hw_reg
);
2512 case BRW_VERT_RESULT_NDC
:
2513 current_annotation
= "NDC";
2514 emit(MOV(reg
, src_reg(output_reg
[BRW_VERT_RESULT_NDC
])));
2516 case BRW_VERT_RESULT_HPOS_DUPLICATE
:
2517 case VERT_RESULT_HPOS
:
2518 current_annotation
= "gl_Position";
2519 emit(MOV(reg
, src_reg(output_reg
[VERT_RESULT_HPOS
])));
2521 case VERT_RESULT_CLIP_DIST0
:
2522 case VERT_RESULT_CLIP_DIST1
:
2523 if (this->c
->key
.uses_clip_distance
) {
2524 emit_generic_urb_slot(reg
, vert_result
);
2526 current_annotation
= "user clip distances";
2527 emit_clip_distances(hw_reg
, (vert_result
- VERT_RESULT_CLIP_DIST0
) * 4);
2530 case VERT_RESULT_EDGE
:
2531 /* This is present when doing unfilled polygons. We're supposed to copy
2532 * the edge flag from the user-provided vertex array
2533 * (glEdgeFlagPointer), or otherwise we'll copy from the current value
2534 * of that attribute (starts as 1.0f). This is then used in clipping to
2535 * determine which edges should be drawn as wireframe.
2537 current_annotation
= "edge flag";
2538 emit(MOV(reg
, src_reg(dst_reg(ATTR
, VERT_ATTRIB_EDGEFLAG
,
2539 glsl_type::float_type
, WRITEMASK_XYZW
))));
2541 case BRW_VERT_RESULT_PAD
:
2542 /* No need to write to this slot */
2545 emit_generic_urb_slot(reg
, vert_result
);
2551 align_interleaved_urb_mlen(struct brw_context
*brw
, int mlen
)
2553 struct intel_context
*intel
= &brw
->intel
;
2555 if (intel
->gen
>= 6) {
2556 /* URB data written (does not include the message header reg) must
2557 * be a multiple of 256 bits, or 2 VS registers. See vol5c.5,
2558 * section 5.4.3.2.2: URB_INTERLEAVED.
2560 * URB entries are allocated on a multiple of 1024 bits, so an
2561 * extra 128 bits written here to make the end align to 256 is
2564 if ((mlen
% 2) != 1)
2572 * Generates the VUE payload plus the 1 or 2 URB write instructions to
2573 * complete the VS thread.
2575 * The VUE layout is documented in Volume 2a.
2578 vec4_visitor::emit_urb_writes()
2580 /* MRF 0 is reserved for the debugger, so start with message header
2585 /* In the process of generating our URB write message contents, we
2586 * may need to unspill a register or load from an array. Those
2587 * reads would use MRFs 14-15.
2589 int max_usable_mrf
= 13;
2591 /* The following assertion verifies that max_usable_mrf causes an
2592 * even-numbered amount of URB write data, which will meet gen6's
2593 * requirements for length alignment.
2595 assert ((max_usable_mrf
- base_mrf
) % 2 == 0);
2597 /* First mrf is the g0-based message header containing URB handles and such,
2598 * which is implied in VS_OPCODE_URB_WRITE.
2602 if (intel
->gen
< 6) {
2603 emit_ndc_computation();
2606 /* Set up the VUE data for the first URB write */
2608 for (slot
= 0; slot
< c
->prog_data
.vue_map
.num_slots
; ++slot
) {
2609 emit_urb_slot(mrf
++, c
->prog_data
.vue_map
.slot_to_vert_result
[slot
]);
2611 /* If this was max_usable_mrf, we can't fit anything more into this URB
2614 if (mrf
> max_usable_mrf
) {
2620 current_annotation
= "URB write";
2621 vec4_instruction
*inst
= emit(VS_OPCODE_URB_WRITE
);
2622 inst
->base_mrf
= base_mrf
;
2623 inst
->mlen
= align_interleaved_urb_mlen(brw
, mrf
- base_mrf
);
2624 inst
->eot
= (slot
>= c
->prog_data
.vue_map
.num_slots
);
2626 /* Optional second URB write */
2630 for (; slot
< c
->prog_data
.vue_map
.num_slots
; ++slot
) {
2631 assert(mrf
< max_usable_mrf
);
2633 emit_urb_slot(mrf
++, c
->prog_data
.vue_map
.slot_to_vert_result
[slot
]);
2636 current_annotation
= "URB write";
2637 inst
= emit(VS_OPCODE_URB_WRITE
);
2638 inst
->base_mrf
= base_mrf
;
2639 inst
->mlen
= align_interleaved_urb_mlen(brw
, mrf
- base_mrf
);
2641 /* URB destination offset. In the previous write, we got MRFs
2642 * 2-13 minus the one header MRF, so 12 regs. URB offset is in
2643 * URB row increments, and each of our MRFs is half of one of
2644 * those, since we're doing interleaved writes.
2646 inst
->offset
= (max_usable_mrf
- base_mrf
) / 2;
2651 vec4_visitor::get_scratch_offset(vec4_instruction
*inst
,
2652 src_reg
*reladdr
, int reg_offset
)
2654 /* Because we store the values to scratch interleaved like our
2655 * vertex data, we need to scale the vec4 index by 2.
2657 int message_header_scale
= 2;
2659 /* Pre-gen6, the message header uses byte offsets instead of vec4
2660 * (16-byte) offset units.
2663 message_header_scale
*= 16;
2666 src_reg index
= src_reg(this, glsl_type::int_type
);
2668 emit_before(inst
, ADD(dst_reg(index
), *reladdr
, src_reg(reg_offset
)));
2669 emit_before(inst
, MUL(dst_reg(index
),
2670 index
, src_reg(message_header_scale
)));
2674 return src_reg(reg_offset
* message_header_scale
);
2679 vec4_visitor::get_pull_constant_offset(vec4_instruction
*inst
,
2680 src_reg
*reladdr
, int reg_offset
)
2683 src_reg index
= src_reg(this, glsl_type::int_type
);
2685 emit_before(inst
, ADD(dst_reg(index
), *reladdr
, src_reg(reg_offset
)));
2687 /* Pre-gen6, the message header uses byte offsets instead of vec4
2688 * (16-byte) offset units.
2690 if (intel
->gen
< 6) {
2691 emit_before(inst
, MUL(dst_reg(index
), index
, src_reg(16)));
2696 int message_header_scale
= intel
->gen
< 6 ? 16 : 1;
2697 return src_reg(reg_offset
* message_header_scale
);
2702 * Emits an instruction before @inst to load the value named by @orig_src
2703 * from scratch space at @base_offset to @temp.
2705 * @base_offset is measured in 32-byte units (the size of a register).
2708 vec4_visitor::emit_scratch_read(vec4_instruction
*inst
,
2709 dst_reg temp
, src_reg orig_src
,
2712 int reg_offset
= base_offset
+ orig_src
.reg_offset
;
2713 src_reg index
= get_scratch_offset(inst
, orig_src
.reladdr
, reg_offset
);
2715 emit_before(inst
, SCRATCH_READ(temp
, index
));
2719 * Emits an instruction after @inst to store the value to be written
2720 * to @orig_dst to scratch space at @base_offset, from @temp.
2722 * @base_offset is measured in 32-byte units (the size of a register).
2725 vec4_visitor::emit_scratch_write(vec4_instruction
*inst
, int base_offset
)
2727 int reg_offset
= base_offset
+ inst
->dst
.reg_offset
;
2728 src_reg index
= get_scratch_offset(inst
, inst
->dst
.reladdr
, reg_offset
);
2730 /* Create a temporary register to store *inst's result in.
2732 * We have to be careful in MOVing from our temporary result register in
2733 * the scratch write. If we swizzle from channels of the temporary that
2734 * weren't initialized, it will confuse live interval analysis, which will
2735 * make spilling fail to make progress.
2737 src_reg temp
= src_reg(this, glsl_type::vec4_type
);
2738 temp
.type
= inst
->dst
.type
;
2739 int first_writemask_chan
= ffs(inst
->dst
.writemask
) - 1;
2741 for (int i
= 0; i
< 4; i
++)
2742 if (inst
->dst
.writemask
& (1 << i
))
2745 swizzles
[i
] = first_writemask_chan
;
2746 temp
.swizzle
= BRW_SWIZZLE4(swizzles
[0], swizzles
[1],
2747 swizzles
[2], swizzles
[3]);
2749 dst_reg dst
= dst_reg(brw_writemask(brw_vec8_grf(0, 0),
2750 inst
->dst
.writemask
));
2751 vec4_instruction
*write
= SCRATCH_WRITE(dst
, temp
, index
);
2752 write
->predicate
= inst
->predicate
;
2753 write
->ir
= inst
->ir
;
2754 write
->annotation
= inst
->annotation
;
2755 inst
->insert_after(write
);
2757 inst
->dst
.file
= temp
.file
;
2758 inst
->dst
.reg
= temp
.reg
;
2759 inst
->dst
.reg_offset
= temp
.reg_offset
;
2760 inst
->dst
.reladdr
= NULL
;
2764 * We can't generally support array access in GRF space, because a
2765 * single instruction's destination can only span 2 contiguous
2766 * registers. So, we send all GRF arrays that get variable index
2767 * access to scratch space.
2770 vec4_visitor::move_grf_array_access_to_scratch()
2772 int scratch_loc
[this->virtual_grf_count
];
2774 for (int i
= 0; i
< this->virtual_grf_count
; i
++) {
2775 scratch_loc
[i
] = -1;
2778 /* First, calculate the set of virtual GRFs that need to be punted
2779 * to scratch due to having any array access on them, and where in
2782 foreach_list(node
, &this->instructions
) {
2783 vec4_instruction
*inst
= (vec4_instruction
*)node
;
2785 if (inst
->dst
.file
== GRF
&& inst
->dst
.reladdr
&&
2786 scratch_loc
[inst
->dst
.reg
] == -1) {
2787 scratch_loc
[inst
->dst
.reg
] = c
->last_scratch
;
2788 c
->last_scratch
+= this->virtual_grf_sizes
[inst
->dst
.reg
];
2791 for (int i
= 0 ; i
< 3; i
++) {
2792 src_reg
*src
= &inst
->src
[i
];
2794 if (src
->file
== GRF
&& src
->reladdr
&&
2795 scratch_loc
[src
->reg
] == -1) {
2796 scratch_loc
[src
->reg
] = c
->last_scratch
;
2797 c
->last_scratch
+= this->virtual_grf_sizes
[src
->reg
];
2802 /* Now, for anything that will be accessed through scratch, rewrite
2803 * it to load/store. Note that this is a _safe list walk, because
2804 * we may generate a new scratch_write instruction after the one
2807 foreach_list_safe(node
, &this->instructions
) {
2808 vec4_instruction
*inst
= (vec4_instruction
*)node
;
2810 /* Set up the annotation tracking for new generated instructions. */
2812 current_annotation
= inst
->annotation
;
2814 if (inst
->dst
.file
== GRF
&& scratch_loc
[inst
->dst
.reg
] != -1) {
2815 emit_scratch_write(inst
, scratch_loc
[inst
->dst
.reg
]);
2818 for (int i
= 0 ; i
< 3; i
++) {
2819 if (inst
->src
[i
].file
!= GRF
|| scratch_loc
[inst
->src
[i
].reg
] == -1)
2822 dst_reg temp
= dst_reg(this, glsl_type::vec4_type
);
2824 emit_scratch_read(inst
, temp
, inst
->src
[i
],
2825 scratch_loc
[inst
->src
[i
].reg
]);
2827 inst
->src
[i
].file
= temp
.file
;
2828 inst
->src
[i
].reg
= temp
.reg
;
2829 inst
->src
[i
].reg_offset
= temp
.reg_offset
;
2830 inst
->src
[i
].reladdr
= NULL
;
2836 * Emits an instruction before @inst to load the value named by @orig_src
2837 * from the pull constant buffer (surface) at @base_offset to @temp.
2840 vec4_visitor::emit_pull_constant_load(vec4_instruction
*inst
,
2841 dst_reg temp
, src_reg orig_src
,
2844 int reg_offset
= base_offset
+ orig_src
.reg_offset
;
2845 src_reg index
= src_reg((unsigned)SURF_INDEX_VERT_CONST_BUFFER
);
2846 src_reg offset
= get_pull_constant_offset(inst
, orig_src
.reladdr
, reg_offset
);
2847 vec4_instruction
*load
;
2849 load
= new(mem_ctx
) vec4_instruction(this, VS_OPCODE_PULL_CONSTANT_LOAD
,
2850 temp
, index
, offset
);
2851 load
->base_mrf
= 14;
2853 emit_before(inst
, load
);
2857 * Implements array access of uniforms by inserting a
2858 * PULL_CONSTANT_LOAD instruction.
2860 * Unlike temporary GRF array access (where we don't support it due to
2861 * the difficulty of doing relative addressing on instruction
2862 * destinations), we could potentially do array access of uniforms
2863 * that were loaded in GRF space as push constants. In real-world
2864 * usage we've seen, though, the arrays being used are always larger
2865 * than we could load as push constants, so just always move all
2866 * uniform array access out to a pull constant buffer.
2869 vec4_visitor::move_uniform_array_access_to_pull_constants()
2871 int pull_constant_loc
[this->uniforms
];
2873 for (int i
= 0; i
< this->uniforms
; i
++) {
2874 pull_constant_loc
[i
] = -1;
2877 /* Walk through and find array access of uniforms. Put a copy of that
2878 * uniform in the pull constant buffer.
2880 * Note that we don't move constant-indexed accesses to arrays. No
2881 * testing has been done of the performance impact of this choice.
2883 foreach_list_safe(node
, &this->instructions
) {
2884 vec4_instruction
*inst
= (vec4_instruction
*)node
;
2886 for (int i
= 0 ; i
< 3; i
++) {
2887 if (inst
->src
[i
].file
!= UNIFORM
|| !inst
->src
[i
].reladdr
)
2890 int uniform
= inst
->src
[i
].reg
;
2892 /* If this array isn't already present in the pull constant buffer,
2895 if (pull_constant_loc
[uniform
] == -1) {
2896 const float **values
= &prog_data
->param
[uniform
* 4];
2898 pull_constant_loc
[uniform
] = prog_data
->nr_pull_params
/ 4;
2900 for (int j
= 0; j
< uniform_size
[uniform
] * 4; j
++) {
2901 prog_data
->pull_param
[prog_data
->nr_pull_params
++] = values
[j
];
2905 /* Set up the annotation tracking for new generated instructions. */
2907 current_annotation
= inst
->annotation
;
2909 dst_reg temp
= dst_reg(this, glsl_type::vec4_type
);
2911 emit_pull_constant_load(inst
, temp
, inst
->src
[i
],
2912 pull_constant_loc
[uniform
]);
2914 inst
->src
[i
].file
= temp
.file
;
2915 inst
->src
[i
].reg
= temp
.reg
;
2916 inst
->src
[i
].reg_offset
= temp
.reg_offset
;
2917 inst
->src
[i
].reladdr
= NULL
;
2921 /* Now there are no accesses of the UNIFORM file with a reladdr, so
2922 * no need to track them as larger-than-vec4 objects. This will be
2923 * relied on in cutting out unused uniform vectors from push
2926 split_uniform_registers();
2930 vec4_visitor::resolve_ud_negate(src_reg
*reg
)
2932 if (reg
->type
!= BRW_REGISTER_TYPE_UD
||
2936 src_reg temp
= src_reg(this, glsl_type::uvec4_type
);
2937 emit(BRW_OPCODE_MOV
, dst_reg(temp
), *reg
);
2941 vec4_visitor::vec4_visitor(struct brw_context
*brw
,
2942 struct brw_vs_compile
*c
,
2943 struct gl_shader_program
*prog
,
2944 struct brw_shader
*shader
,
2949 this->intel
= &brw
->intel
;
2950 this->ctx
= &intel
->ctx
;
2952 this->shader
= shader
;
2954 this->mem_ctx
= mem_ctx
;
2955 this->failed
= false;
2957 this->base_ir
= NULL
;
2958 this->current_annotation
= NULL
;
2959 memset(this->output_reg_annotation
, 0, sizeof(this->output_reg_annotation
));
2962 this->vp
= &c
->vp
->program
;
2963 this->prog_data
= &c
->prog_data
;
2965 this->variable_ht
= hash_table_ctor(0,
2966 hash_table_pointer_hash
,
2967 hash_table_pointer_compare
);
2969 this->virtual_grf_def
= NULL
;
2970 this->virtual_grf_use
= NULL
;
2971 this->virtual_grf_sizes
= NULL
;
2972 this->virtual_grf_count
= 0;
2973 this->virtual_grf_reg_map
= NULL
;
2974 this->virtual_grf_reg_count
= 0;
2975 this->virtual_grf_array_size
= 0;
2976 this->live_intervals_valid
= false;
2978 this->max_grf
= intel
->gen
>= 7 ? GEN7_MRF_HACK_START
: BRW_MAX_GRF
;
2983 vec4_visitor::~vec4_visitor()
2985 hash_table_dtor(this->variable_ht
);
2990 vec4_visitor::fail(const char *format
, ...)
3000 va_start(va
, format
);
3001 msg
= ralloc_vasprintf(mem_ctx
, format
, va
);
3003 msg
= ralloc_asprintf(mem_ctx
, "VS compile failed: %s\n", msg
);
3005 this->fail_msg
= msg
;
3007 if (INTEL_DEBUG
& DEBUG_VS
) {
3008 fprintf(stderr
, "%s", msg
);
3012 } /* namespace brw */