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
26 #include "glsl/ir_uniform.h"
27 #include "program/sampler.h"
31 vec4_instruction::vec4_instruction(enum opcode opcode
, const dst_reg
&dst
,
32 const src_reg
&src0
, const src_reg
&src1
,
35 this->opcode
= opcode
;
40 this->saturate
= false;
41 this->force_writemask_all
= false;
42 this->no_dd_clear
= false;
43 this->no_dd_check
= false;
44 this->writes_accumulator
= false;
45 this->conditional_mod
= BRW_CONDITIONAL_NONE
;
46 this->predicate
= BRW_PREDICATE_NONE
;
47 this->predicate_inverse
= false;
49 this->regs_written
= (dst
.file
== BAD_FILE
? 0 : 1);
50 this->shadow_compare
= false;
52 this->urb_write_flags
= BRW_URB_WRITE_NO_FLAGS
;
53 this->header_present
= false;
54 this->flag_subreg
= 0;
58 this->annotation
= NULL
;
62 vec4_visitor::emit(vec4_instruction
*inst
)
64 inst
->ir
= this->base_ir
;
65 inst
->annotation
= this->current_annotation
;
67 this->instructions
.push_tail(inst
);
73 vec4_visitor::emit_before(bblock_t
*block
, vec4_instruction
*inst
,
74 vec4_instruction
*new_inst
)
76 new_inst
->ir
= inst
->ir
;
77 new_inst
->annotation
= inst
->annotation
;
79 inst
->insert_before(block
, new_inst
);
85 vec4_visitor::emit(enum opcode opcode
, const dst_reg
&dst
, const src_reg
&src0
,
86 const src_reg
&src1
, const src_reg
&src2
)
88 return emit(new(mem_ctx
) vec4_instruction(opcode
, dst
, src0
, src1
, src2
));
93 vec4_visitor::emit(enum opcode opcode
, const dst_reg
&dst
, const src_reg
&src0
,
96 return emit(new(mem_ctx
) vec4_instruction(opcode
, dst
, src0
, src1
));
100 vec4_visitor::emit(enum opcode opcode
, const dst_reg
&dst
, const src_reg
&src0
)
102 return emit(new(mem_ctx
) vec4_instruction(opcode
, dst
, src0
));
106 vec4_visitor::emit(enum opcode opcode
, const dst_reg
&dst
)
108 return emit(new(mem_ctx
) vec4_instruction(opcode
, dst
));
112 vec4_visitor::emit(enum opcode opcode
)
114 return emit(new(mem_ctx
) vec4_instruction(opcode
, dst_reg()));
119 vec4_visitor::op(const dst_reg &dst, const src_reg &src0) \
121 return new(mem_ctx) vec4_instruction(BRW_OPCODE_##op, dst, src0); \
126 vec4_visitor::op(const dst_reg &dst, const src_reg &src0, \
127 const src_reg &src1) \
129 return new(mem_ctx) vec4_instruction(BRW_OPCODE_##op, dst, \
133 #define ALU2_ACC(op) \
135 vec4_visitor::op(const dst_reg &dst, const src_reg &src0, \
136 const src_reg &src1) \
138 vec4_instruction *inst = new(mem_ctx) vec4_instruction( \
139 BRW_OPCODE_##op, dst, src0, src1); \
140 inst->writes_accumulator = true; \
146 vec4_visitor::op(const dst_reg &dst, const src_reg &src0, \
147 const src_reg &src1, const src_reg &src2) \
149 assert(devinfo->gen >= 6); \
150 return new(mem_ctx) vec4_instruction(BRW_OPCODE_##op, dst, \
187 /** Gen4 predicated IF. */
189 vec4_visitor::IF(enum brw_predicate predicate
)
191 vec4_instruction
*inst
;
193 inst
= new(mem_ctx
) vec4_instruction(BRW_OPCODE_IF
);
194 inst
->predicate
= predicate
;
199 /** Gen6 IF with embedded comparison. */
201 vec4_visitor::IF(src_reg src0
, src_reg src1
,
202 enum brw_conditional_mod condition
)
204 assert(devinfo
->gen
== 6);
206 vec4_instruction
*inst
;
208 resolve_ud_negate(&src0
);
209 resolve_ud_negate(&src1
);
211 inst
= new(mem_ctx
) vec4_instruction(BRW_OPCODE_IF
, dst_null_d(),
213 inst
->conditional_mod
= condition
;
219 * CMP: Sets the low bit of the destination channels with the result
220 * of the comparison, while the upper bits are undefined, and updates
221 * the flag register with the packed 16 bits of the result.
224 vec4_visitor::CMP(dst_reg dst
, src_reg src0
, src_reg src1
,
225 enum brw_conditional_mod condition
)
227 vec4_instruction
*inst
;
229 /* Take the instruction:
231 * CMP null<d> src0<f> src1<f>
233 * Original gen4 does type conversion to the destination type before
234 * comparison, producing garbage results for floating point comparisons.
236 * The destination type doesn't matter on newer generations, so we set the
237 * type to match src0 so we can compact the instruction.
239 dst
.type
= src0
.type
;
240 if (dst
.file
== HW_REG
)
241 dst
.fixed_hw_reg
.type
= dst
.type
;
243 resolve_ud_negate(&src0
);
244 resolve_ud_negate(&src1
);
246 inst
= new(mem_ctx
) vec4_instruction(BRW_OPCODE_CMP
, dst
, src0
, src1
);
247 inst
->conditional_mod
= condition
;
253 vec4_visitor::SCRATCH_READ(const dst_reg
&dst
, const src_reg
&index
)
255 vec4_instruction
*inst
;
257 inst
= new(mem_ctx
) vec4_instruction(SHADER_OPCODE_GEN4_SCRATCH_READ
,
266 vec4_visitor::SCRATCH_WRITE(const dst_reg
&dst
, const src_reg
&src
,
267 const src_reg
&index
)
269 vec4_instruction
*inst
;
271 inst
= new(mem_ctx
) vec4_instruction(SHADER_OPCODE_GEN4_SCRATCH_WRITE
,
280 vec4_visitor::emit_dp(dst_reg dst
, src_reg src0
, src_reg src1
, unsigned elements
)
282 static enum opcode dot_opcodes
[] = {
283 BRW_OPCODE_DP2
, BRW_OPCODE_DP3
, BRW_OPCODE_DP4
286 emit(dot_opcodes
[elements
- 2], dst
, src0
, src1
);
290 vec4_visitor::fix_3src_operand(src_reg src
)
292 /* Using vec4 uniforms in SIMD4x2 programs is difficult. You'd like to be
293 * able to use vertical stride of zero to replicate the vec4 uniform, like
295 * g3<0;4,1>:f - [0, 4][1, 5][2, 6][3, 7]
297 * But you can't, since vertical stride is always four in three-source
298 * instructions. Instead, insert a MOV instruction to do the replication so
299 * that the three-source instruction can consume it.
302 /* The MOV is only needed if the source is a uniform or immediate. */
303 if (src
.file
!= UNIFORM
&& src
.file
!= IMM
)
306 if (src
.file
== UNIFORM
&& brw_is_single_value_swizzle(src
.swizzle
))
309 dst_reg expanded
= dst_reg(this, glsl_type::vec4_type
);
310 expanded
.type
= src
.type
;
311 emit(VEC4_OPCODE_UNPACK_UNIFORM
, expanded
, src
);
312 return src_reg(expanded
);
316 vec4_visitor::fix_math_operand(src_reg src
)
318 if (devinfo
->gen
< 6 || devinfo
->gen
>= 8 || src
.file
== BAD_FILE
)
321 /* The gen6 math instruction ignores the source modifiers --
322 * swizzle, abs, negate, and at least some parts of the register
323 * region description.
325 * Rather than trying to enumerate all these cases, *always* expand the
326 * operand to a temp GRF for gen6.
328 * For gen7, keep the operand as-is, except if immediate, which gen7 still
332 if (devinfo
->gen
== 7 && src
.file
!= IMM
)
335 dst_reg expanded
= dst_reg(this, glsl_type::vec4_type
);
336 expanded
.type
= src
.type
;
337 emit(MOV(expanded
, src
));
338 return src_reg(expanded
);
342 vec4_visitor::emit_math(enum opcode opcode
,
344 const src_reg
&src0
, const src_reg
&src1
)
346 vec4_instruction
*math
=
347 emit(opcode
, dst
, fix_math_operand(src0
), fix_math_operand(src1
));
349 if (devinfo
->gen
== 6 && dst
.writemask
!= WRITEMASK_XYZW
) {
350 /* MATH on Gen6 must be align1, so we can't do writemasks. */
351 math
->dst
= dst_reg(this, glsl_type::vec4_type
);
352 math
->dst
.type
= dst
.type
;
353 emit(MOV(dst
, src_reg(math
->dst
)));
354 } else if (devinfo
->gen
< 6) {
356 math
->mlen
= src1
.file
== BAD_FILE
? 1 : 2;
361 vec4_visitor::emit_pack_half_2x16(dst_reg dst
, src_reg src0
)
363 if (devinfo
->gen
< 7) {
364 unreachable("ir_unop_pack_half_2x16 should be lowered");
367 assert(dst
.type
== BRW_REGISTER_TYPE_UD
);
368 assert(src0
.type
== BRW_REGISTER_TYPE_F
);
370 /* From the Ivybridge PRM, Vol4, Part3, Section 6.27 f32to16:
372 * Because this instruction does not have a 16-bit floating-point type,
373 * the destination data type must be Word (W).
375 * The destination must be DWord-aligned and specify a horizontal stride
376 * (HorzStride) of 2. The 16-bit result is stored in the lower word of
377 * each destination channel and the upper word is not modified.
379 * The above restriction implies that the f32to16 instruction must use
380 * align1 mode, because only in align1 mode is it possible to specify
381 * horizontal stride. We choose here to defy the hardware docs and emit
382 * align16 instructions.
384 * (I [chadv] did attempt to emit align1 instructions for VS f32to16
385 * instructions. I was partially successful in that the code passed all
386 * tests. However, the code was dubiously correct and fragile, and the
387 * tests were not harsh enough to probe that frailty. Not trusting the
388 * code, I chose instead to remain in align16 mode in defiance of the hw
391 * I've [chadv] experimentally confirmed that, on gen7 hardware and the
392 * simulator, emitting a f32to16 in align16 mode with UD as destination
393 * data type is safe. The behavior differs from that specified in the PRM
394 * in that the upper word of each destination channel is cleared to 0.
397 dst_reg
tmp_dst(this, glsl_type::uvec2_type
);
398 src_reg
tmp_src(tmp_dst
);
401 /* Verify the undocumented behavior on which the following instructions
402 * rely. If f32to16 fails to clear the upper word of the X and Y channels,
403 * then the result of the bit-or instruction below will be incorrect.
405 * You should inspect the disasm output in order to verify that the MOV is
406 * not optimized away.
408 emit(MOV(tmp_dst
, src_reg(0x12345678u
)));
411 /* Give tmp the form below, where "." means untouched.
414 * |.|.|0x0000hhhh|0x0000llll|.|.|0x0000hhhh|0x0000llll|
416 * That the upper word of each write-channel be 0 is required for the
417 * following bit-shift and bit-or instructions to work. Note that this
418 * relies on the undocumented hardware behavior mentioned above.
420 tmp_dst
.writemask
= WRITEMASK_XY
;
421 emit(F32TO16(tmp_dst
, src0
));
423 /* Give the write-channels of dst the form:
426 tmp_src
.swizzle
= BRW_SWIZZLE_YYYY
;
427 emit(SHL(dst
, tmp_src
, src_reg(16u)));
429 /* Finally, give the write-channels of dst the form of packHalf2x16's
433 tmp_src
.swizzle
= BRW_SWIZZLE_XXXX
;
434 emit(OR(dst
, src_reg(dst
), tmp_src
));
438 vec4_visitor::emit_unpack_half_2x16(dst_reg dst
, src_reg src0
)
440 if (devinfo
->gen
< 7) {
441 unreachable("ir_unop_unpack_half_2x16 should be lowered");
444 assert(dst
.type
== BRW_REGISTER_TYPE_F
);
445 assert(src0
.type
== BRW_REGISTER_TYPE_UD
);
447 /* From the Ivybridge PRM, Vol4, Part3, Section 6.26 f16to32:
449 * Because this instruction does not have a 16-bit floating-point type,
450 * the source data type must be Word (W). The destination type must be
453 * To use W as the source data type, we must adjust horizontal strides,
454 * which is only possible in align1 mode. All my [chadv] attempts at
455 * emitting align1 instructions for unpackHalf2x16 failed to pass the
456 * Piglit tests, so I gave up.
458 * I've verified that, on gen7 hardware and the simulator, it is safe to
459 * emit f16to32 in align16 mode with UD as source data type.
462 dst_reg
tmp_dst(this, glsl_type::uvec2_type
);
463 src_reg
tmp_src(tmp_dst
);
465 tmp_dst
.writemask
= WRITEMASK_X
;
466 emit(AND(tmp_dst
, src0
, src_reg(0xffffu
)));
468 tmp_dst
.writemask
= WRITEMASK_Y
;
469 emit(SHR(tmp_dst
, src0
, src_reg(16u)));
471 dst
.writemask
= WRITEMASK_XY
;
472 emit(F16TO32(dst
, tmp_src
));
476 vec4_visitor::emit_unpack_unorm_4x8(const dst_reg
&dst
, src_reg src0
)
478 /* Instead of splitting the 32-bit integer, shifting, and ORing it back
479 * together, we can shift it by <0, 8, 16, 24>. The packed integer immediate
480 * is not suitable to generate the shift values, but we can use the packed
481 * vector float and a type-converting MOV.
483 dst_reg
shift(this, glsl_type::uvec4_type
);
484 emit(MOV(shift
, src_reg(0x00, 0x60, 0x70, 0x78)));
486 dst_reg
shifted(this, glsl_type::uvec4_type
);
487 src0
.swizzle
= BRW_SWIZZLE_XXXX
;
488 emit(SHR(shifted
, src0
, src_reg(shift
)));
490 shifted
.type
= BRW_REGISTER_TYPE_UB
;
491 dst_reg
f(this, glsl_type::vec4_type
);
492 emit(VEC4_OPCODE_MOV_BYTES
, f
, src_reg(shifted
));
494 emit(MUL(dst
, src_reg(f
), src_reg(1.0f
/ 255.0f
)));
498 vec4_visitor::emit_unpack_snorm_4x8(const dst_reg
&dst
, src_reg src0
)
500 /* Instead of splitting the 32-bit integer, shifting, and ORing it back
501 * together, we can shift it by <0, 8, 16, 24>. The packed integer immediate
502 * is not suitable to generate the shift values, but we can use the packed
503 * vector float and a type-converting MOV.
505 dst_reg
shift(this, glsl_type::uvec4_type
);
506 emit(MOV(shift
, src_reg(0x00, 0x60, 0x70, 0x78)));
508 dst_reg
shifted(this, glsl_type::uvec4_type
);
509 src0
.swizzle
= BRW_SWIZZLE_XXXX
;
510 emit(SHR(shifted
, src0
, src_reg(shift
)));
512 shifted
.type
= BRW_REGISTER_TYPE_B
;
513 dst_reg
f(this, glsl_type::vec4_type
);
514 emit(VEC4_OPCODE_MOV_BYTES
, f
, src_reg(shifted
));
516 dst_reg
scaled(this, glsl_type::vec4_type
);
517 emit(MUL(scaled
, src_reg(f
), src_reg(1.0f
/ 127.0f
)));
519 dst_reg
max(this, glsl_type::vec4_type
);
520 emit_minmax(BRW_CONDITIONAL_GE
, max
, src_reg(scaled
), src_reg(-1.0f
));
521 emit_minmax(BRW_CONDITIONAL_L
, dst
, src_reg(max
), src_reg(1.0f
));
525 vec4_visitor::emit_pack_unorm_4x8(const dst_reg
&dst
, const src_reg
&src0
)
527 dst_reg
saturated(this, glsl_type::vec4_type
);
528 vec4_instruction
*inst
= emit(MOV(saturated
, src0
));
529 inst
->saturate
= true;
531 dst_reg
scaled(this, glsl_type::vec4_type
);
532 emit(MUL(scaled
, src_reg(saturated
), src_reg(255.0f
)));
534 dst_reg
rounded(this, glsl_type::vec4_type
);
535 emit(RNDE(rounded
, src_reg(scaled
)));
537 dst_reg
u(this, glsl_type::uvec4_type
);
538 emit(MOV(u
, src_reg(rounded
)));
541 emit(VEC4_OPCODE_PACK_BYTES
, dst
, bytes
);
545 vec4_visitor::emit_pack_snorm_4x8(const dst_reg
&dst
, const src_reg
&src0
)
547 dst_reg
max(this, glsl_type::vec4_type
);
548 emit_minmax(BRW_CONDITIONAL_GE
, max
, src0
, src_reg(-1.0f
));
550 dst_reg
min(this, glsl_type::vec4_type
);
551 emit_minmax(BRW_CONDITIONAL_L
, min
, src_reg(max
), src_reg(1.0f
));
553 dst_reg
scaled(this, glsl_type::vec4_type
);
554 emit(MUL(scaled
, src_reg(min
), src_reg(127.0f
)));
556 dst_reg
rounded(this, glsl_type::vec4_type
);
557 emit(RNDE(rounded
, src_reg(scaled
)));
559 dst_reg
i(this, glsl_type::ivec4_type
);
560 emit(MOV(i
, src_reg(rounded
)));
563 emit(VEC4_OPCODE_PACK_BYTES
, dst
, bytes
);
567 vec4_visitor::visit_instructions(const exec_list
*list
)
569 foreach_in_list(ir_instruction
, ir
, list
) {
577 type_size(const struct glsl_type
*type
)
582 switch (type
->base_type
) {
585 case GLSL_TYPE_FLOAT
:
587 if (type
->is_matrix()) {
588 return type
->matrix_columns
;
590 /* Regardless of size of vector, it gets a vec4. This is bad
591 * packing for things like floats, but otherwise arrays become a
592 * mess. Hopefully a later pass over the code can pack scalars
593 * down if appropriate.
597 case GLSL_TYPE_ARRAY
:
598 assert(type
->length
> 0);
599 return type_size(type
->fields
.array
) * type
->length
;
600 case GLSL_TYPE_STRUCT
:
602 for (i
= 0; i
< type
->length
; i
++) {
603 size
+= type_size(type
->fields
.structure
[i
].type
);
606 case GLSL_TYPE_SAMPLER
:
607 /* Samplers take up no register space, since they're baked in at
611 case GLSL_TYPE_ATOMIC_UINT
:
613 case GLSL_TYPE_IMAGE
:
615 case GLSL_TYPE_DOUBLE
:
616 case GLSL_TYPE_ERROR
:
617 case GLSL_TYPE_INTERFACE
:
618 unreachable("not reached");
624 src_reg::src_reg(class vec4_visitor
*v
, const struct glsl_type
*type
)
629 this->reg
= v
->alloc
.allocate(type_size(type
));
631 if (type
->is_array() || type
->is_record()) {
632 this->swizzle
= BRW_SWIZZLE_NOOP
;
634 this->swizzle
= brw_swizzle_for_size(type
->vector_elements
);
637 this->type
= brw_type_for_base_type(type
);
640 src_reg::src_reg(class vec4_visitor
*v
, const struct glsl_type
*type
, int size
)
647 this->reg
= v
->alloc
.allocate(type_size(type
) * size
);
649 this->swizzle
= BRW_SWIZZLE_NOOP
;
651 this->type
= brw_type_for_base_type(type
);
654 dst_reg::dst_reg(class vec4_visitor
*v
, const struct glsl_type
*type
)
659 this->reg
= v
->alloc
.allocate(type_size(type
));
661 if (type
->is_array() || type
->is_record()) {
662 this->writemask
= WRITEMASK_XYZW
;
664 this->writemask
= (1 << type
->vector_elements
) - 1;
667 this->type
= brw_type_for_base_type(type
);
670 /* Our support for uniforms is piggy-backed on the struct
671 * gl_fragment_program, because that's where the values actually
672 * get stored, rather than in some global gl_shader_program uniform
676 vec4_visitor::setup_uniform_values(ir_variable
*ir
)
678 int namelen
= strlen(ir
->name
);
680 /* The data for our (non-builtin) uniforms is stored in a series of
681 * gl_uniform_driver_storage structs for each subcomponent that
682 * glGetUniformLocation() could name. We know it's been set up in the same
683 * order we'd walk the type, so walk the list of storage and find anything
684 * with our name, or the prefix of a component that starts with our name.
686 for (unsigned u
= 0; u
< shader_prog
->NumUserUniformStorage
; u
++) {
687 struct gl_uniform_storage
*storage
= &shader_prog
->UniformStorage
[u
];
689 if (strncmp(ir
->name
, storage
->name
, namelen
) != 0 ||
690 (storage
->name
[namelen
] != 0 &&
691 storage
->name
[namelen
] != '.' &&
692 storage
->name
[namelen
] != '[')) {
696 gl_constant_value
*components
= storage
->storage
;
697 unsigned vector_count
= (MAX2(storage
->array_elements
, 1) *
698 storage
->type
->matrix_columns
);
700 for (unsigned s
= 0; s
< vector_count
; s
++) {
701 assert(uniforms
< uniform_array_size
);
702 uniform_vector_size
[uniforms
] = storage
->type
->vector_elements
;
705 for (i
= 0; i
< uniform_vector_size
[uniforms
]; i
++) {
706 stage_prog_data
->param
[uniforms
* 4 + i
] = components
;
710 static gl_constant_value zero
= { 0.0 };
711 stage_prog_data
->param
[uniforms
* 4 + i
] = &zero
;
720 vec4_visitor::setup_uniform_clipplane_values()
722 gl_clip_plane
*clip_planes
= brw_select_clip_planes(ctx
);
724 for (int i
= 0; i
< key
->nr_userclip_plane_consts
; ++i
) {
725 assert(this->uniforms
< uniform_array_size
);
726 this->uniform_vector_size
[this->uniforms
] = 4;
727 this->userplane
[i
] = dst_reg(UNIFORM
, this->uniforms
);
728 this->userplane
[i
].type
= BRW_REGISTER_TYPE_F
;
729 for (int j
= 0; j
< 4; ++j
) {
730 stage_prog_data
->param
[this->uniforms
* 4 + j
] =
731 (gl_constant_value
*) &clip_planes
[i
][j
];
737 /* Our support for builtin uniforms is even scarier than non-builtin.
738 * It sits on top of the PROG_STATE_VAR parameters that are
739 * automatically updated from GL context state.
742 vec4_visitor::setup_builtin_uniform_values(ir_variable
*ir
)
744 const ir_state_slot
*const slots
= ir
->get_state_slots();
745 assert(slots
!= NULL
);
747 for (unsigned int i
= 0; i
< ir
->get_num_state_slots(); i
++) {
748 /* This state reference has already been setup by ir_to_mesa,
749 * but we'll get the same index back here. We can reference
750 * ParameterValues directly, since unlike brw_fs.cpp, we never
751 * add new state references during compile.
753 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
754 (gl_state_index
*)slots
[i
].tokens
);
755 gl_constant_value
*values
=
756 &this->prog
->Parameters
->ParameterValues
[index
][0];
758 assert(this->uniforms
< uniform_array_size
);
760 for (unsigned j
= 0; j
< 4; j
++)
761 stage_prog_data
->param
[this->uniforms
* 4 + j
] =
762 &values
[GET_SWZ(slots
[i
].swizzle
, j
)];
764 this->uniform_vector_size
[this->uniforms
] =
765 (ir
->type
->is_scalar() || ir
->type
->is_vector() ||
766 ir
->type
->is_matrix() ? ir
->type
->vector_elements
: 4);
773 vec4_visitor::variable_storage(ir_variable
*var
)
775 return (dst_reg
*)hash_table_find(this->variable_ht
, var
);
779 vec4_visitor::emit_bool_to_cond_code(ir_rvalue
*ir
,
780 enum brw_predicate
*predicate
)
782 ir_expression
*expr
= ir
->as_expression();
784 *predicate
= BRW_PREDICATE_NORMAL
;
786 if (expr
&& expr
->operation
!= ir_binop_ubo_load
) {
788 vec4_instruction
*inst
;
790 assert(expr
->get_num_operands() <= 3);
791 for (unsigned int i
= 0; i
< expr
->get_num_operands(); i
++) {
792 expr
->operands
[i
]->accept(this);
793 op
[i
] = this->result
;
795 resolve_ud_negate(&op
[i
]);
798 switch (expr
->operation
) {
799 case ir_unop_logic_not
:
800 inst
= emit(AND(dst_null_d(), op
[0], src_reg(1)));
801 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
804 case ir_binop_logic_xor
:
805 if (devinfo
->gen
<= 5) {
806 src_reg temp
= src_reg(this, ir
->type
);
807 emit(XOR(dst_reg(temp
), op
[0], op
[1]));
808 inst
= emit(AND(dst_null_d(), temp
, src_reg(1)));
810 inst
= emit(XOR(dst_null_d(), op
[0], op
[1]));
812 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
815 case ir_binop_logic_or
:
816 if (devinfo
->gen
<= 5) {
817 src_reg temp
= src_reg(this, ir
->type
);
818 emit(OR(dst_reg(temp
), op
[0], op
[1]));
819 inst
= emit(AND(dst_null_d(), temp
, src_reg(1)));
821 inst
= emit(OR(dst_null_d(), op
[0], op
[1]));
823 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
826 case ir_binop_logic_and
:
827 if (devinfo
->gen
<= 5) {
828 src_reg temp
= src_reg(this, ir
->type
);
829 emit(AND(dst_reg(temp
), op
[0], op
[1]));
830 inst
= emit(AND(dst_null_d(), temp
, src_reg(1)));
832 inst
= emit(AND(dst_null_d(), op
[0], op
[1]));
834 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
838 if (devinfo
->gen
>= 6) {
839 emit(CMP(dst_null_d(), op
[0], src_reg(0.0f
), BRW_CONDITIONAL_NZ
));
841 inst
= emit(MOV(dst_null_f(), op
[0]));
842 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
847 if (devinfo
->gen
>= 6) {
848 emit(CMP(dst_null_d(), op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
850 inst
= emit(MOV(dst_null_d(), op
[0]));
851 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
855 case ir_binop_all_equal
:
856 if (devinfo
->gen
<= 5) {
857 resolve_bool_comparison(expr
->operands
[0], &op
[0]);
858 resolve_bool_comparison(expr
->operands
[1], &op
[1]);
860 inst
= emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_Z
));
861 *predicate
= BRW_PREDICATE_ALIGN16_ALL4H
;
864 case ir_binop_any_nequal
:
865 if (devinfo
->gen
<= 5) {
866 resolve_bool_comparison(expr
->operands
[0], &op
[0]);
867 resolve_bool_comparison(expr
->operands
[1], &op
[1]);
869 inst
= emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_NZ
));
870 *predicate
= BRW_PREDICATE_ALIGN16_ANY4H
;
874 if (devinfo
->gen
<= 5) {
875 resolve_bool_comparison(expr
->operands
[0], &op
[0]);
877 inst
= emit(CMP(dst_null_d(), op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
878 *predicate
= BRW_PREDICATE_ALIGN16_ANY4H
;
881 case ir_binop_greater
:
882 case ir_binop_gequal
:
884 case ir_binop_lequal
:
886 case ir_binop_nequal
:
887 if (devinfo
->gen
<= 5) {
888 resolve_bool_comparison(expr
->operands
[0], &op
[0]);
889 resolve_bool_comparison(expr
->operands
[1], &op
[1]);
891 emit(CMP(dst_null_d(), op
[0], op
[1],
892 brw_conditional_for_comparison(expr
->operation
)));
895 case ir_triop_csel
: {
896 /* Expand the boolean condition into the flag register. */
897 inst
= emit(MOV(dst_null_d(), op
[0]));
898 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
900 /* Select which boolean to return. */
901 dst_reg
temp(this, expr
->operands
[1]->type
);
902 inst
= emit(BRW_OPCODE_SEL
, temp
, op
[1], op
[2]);
903 inst
->predicate
= BRW_PREDICATE_NORMAL
;
905 /* Expand the result to a condition code. */
906 inst
= emit(MOV(dst_null_d(), src_reg(temp
)));
907 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
912 unreachable("not reached");
919 resolve_ud_negate(&this->result
);
921 vec4_instruction
*inst
= emit(AND(dst_null_d(), this->result
, src_reg(1)));
922 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
926 * Emit a gen6 IF statement with the comparison folded into the IF
930 vec4_visitor::emit_if_gen6(ir_if
*ir
)
932 ir_expression
*expr
= ir
->condition
->as_expression();
934 if (expr
&& expr
->operation
!= ir_binop_ubo_load
) {
938 assert(expr
->get_num_operands() <= 3);
939 for (unsigned int i
= 0; i
< expr
->get_num_operands(); i
++) {
940 expr
->operands
[i
]->accept(this);
941 op
[i
] = this->result
;
944 switch (expr
->operation
) {
945 case ir_unop_logic_not
:
946 emit(IF(op
[0], src_reg(0), BRW_CONDITIONAL_Z
));
949 case ir_binop_logic_xor
:
950 emit(IF(op
[0], op
[1], BRW_CONDITIONAL_NZ
));
953 case ir_binop_logic_or
:
954 temp
= dst_reg(this, glsl_type::bool_type
);
955 emit(OR(temp
, op
[0], op
[1]));
956 emit(IF(src_reg(temp
), src_reg(0), BRW_CONDITIONAL_NZ
));
959 case ir_binop_logic_and
:
960 temp
= dst_reg(this, glsl_type::bool_type
);
961 emit(AND(temp
, op
[0], op
[1]));
962 emit(IF(src_reg(temp
), src_reg(0), BRW_CONDITIONAL_NZ
));
966 emit(IF(op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
970 emit(IF(op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
973 case ir_binop_greater
:
974 case ir_binop_gequal
:
976 case ir_binop_lequal
:
978 case ir_binop_nequal
:
979 emit(IF(op
[0], op
[1],
980 brw_conditional_for_comparison(expr
->operation
)));
983 case ir_binop_all_equal
:
984 emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_Z
));
985 emit(IF(BRW_PREDICATE_ALIGN16_ALL4H
));
988 case ir_binop_any_nequal
:
989 emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_NZ
));
990 emit(IF(BRW_PREDICATE_ALIGN16_ANY4H
));
994 emit(CMP(dst_null_d(), op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
995 emit(IF(BRW_PREDICATE_ALIGN16_ANY4H
));
998 case ir_triop_csel
: {
999 /* Expand the boolean condition into the flag register. */
1000 vec4_instruction
*inst
= emit(MOV(dst_null_d(), op
[0]));
1001 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
1003 /* Select which boolean to return. */
1004 dst_reg
temp(this, expr
->operands
[1]->type
);
1005 inst
= emit(BRW_OPCODE_SEL
, temp
, op
[1], op
[2]);
1006 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1008 emit(IF(src_reg(temp
), src_reg(0), BRW_CONDITIONAL_NZ
));
1013 unreachable("not reached");
1018 ir
->condition
->accept(this);
1020 emit(IF(this->result
, src_reg(0), BRW_CONDITIONAL_NZ
));
1024 vec4_visitor::visit(ir_variable
*ir
)
1026 dst_reg
*reg
= NULL
;
1028 if (variable_storage(ir
))
1031 switch (ir
->data
.mode
) {
1032 case ir_var_shader_in
:
1033 assert(ir
->data
.location
!= -1);
1034 reg
= new(mem_ctx
) dst_reg(ATTR
, ir
->data
.location
);
1037 case ir_var_shader_out
:
1038 assert(ir
->data
.location
!= -1);
1039 reg
= new(mem_ctx
) dst_reg(this, ir
->type
);
1041 for (int i
= 0; i
< type_size(ir
->type
); i
++) {
1042 output_reg
[ir
->data
.location
+ i
] = *reg
;
1043 output_reg
[ir
->data
.location
+ i
].reg_offset
= i
;
1044 output_reg
[ir
->data
.location
+ i
].type
=
1045 brw_type_for_base_type(ir
->type
->get_scalar_type());
1046 output_reg_annotation
[ir
->data
.location
+ i
] = ir
->name
;
1051 case ir_var_temporary
:
1052 reg
= new(mem_ctx
) dst_reg(this, ir
->type
);
1055 case ir_var_uniform
:
1056 reg
= new(this->mem_ctx
) dst_reg(UNIFORM
, this->uniforms
);
1058 /* Thanks to the lower_ubo_reference pass, we will see only
1059 * ir_binop_ubo_load expressions and not ir_dereference_variable for UBO
1060 * variables, so no need for them to be in variable_ht.
1062 * Some uniforms, such as samplers and atomic counters, have no actual
1063 * storage, so we should ignore them.
1065 if (ir
->is_in_uniform_block() || type_size(ir
->type
) == 0)
1068 /* Track how big the whole uniform variable is, in case we need to put a
1069 * copy of its data into pull constants for array access.
1071 assert(this->uniforms
< uniform_array_size
);
1072 this->uniform_size
[this->uniforms
] = type_size(ir
->type
);
1074 if (!strncmp(ir
->name
, "gl_", 3)) {
1075 setup_builtin_uniform_values(ir
);
1077 setup_uniform_values(ir
);
1081 case ir_var_system_value
:
1082 reg
= make_reg_for_system_value(ir
);
1086 unreachable("not reached");
1089 reg
->type
= brw_type_for_base_type(ir
->type
);
1090 hash_table_insert(this->variable_ht
, reg
, ir
);
1094 vec4_visitor::visit(ir_loop
*ir
)
1096 /* We don't want debugging output to print the whole body of the
1097 * loop as the annotation.
1099 this->base_ir
= NULL
;
1101 emit(BRW_OPCODE_DO
);
1103 visit_instructions(&ir
->body_instructions
);
1105 emit(BRW_OPCODE_WHILE
);
1109 vec4_visitor::visit(ir_loop_jump
*ir
)
1112 case ir_loop_jump::jump_break
:
1113 emit(BRW_OPCODE_BREAK
);
1115 case ir_loop_jump::jump_continue
:
1116 emit(BRW_OPCODE_CONTINUE
);
1123 vec4_visitor::visit(ir_function_signature
*)
1125 unreachable("not reached");
1129 vec4_visitor::visit(ir_function
*ir
)
1131 /* Ignore function bodies other than main() -- we shouldn't see calls to
1132 * them since they should all be inlined.
1134 if (strcmp(ir
->name
, "main") == 0) {
1135 const ir_function_signature
*sig
;
1138 sig
= ir
->matching_signature(NULL
, &empty
, false);
1142 visit_instructions(&sig
->body
);
1147 vec4_visitor::try_emit_mad(ir_expression
*ir
)
1149 /* 3-src instructions were introduced in gen6. */
1150 if (devinfo
->gen
< 6)
1153 /* MAD can only handle floating-point data. */
1154 if (ir
->type
->base_type
!= GLSL_TYPE_FLOAT
)
1159 bool mul_negate
, mul_abs
;
1161 for (int i
= 0; i
< 2; i
++) {
1165 mul
= ir
->operands
[i
]->as_expression();
1166 nonmul
= ir
->operands
[1 - i
];
1168 if (mul
&& mul
->operation
== ir_unop_abs
) {
1169 mul
= mul
->operands
[0]->as_expression();
1171 } else if (mul
&& mul
->operation
== ir_unop_neg
) {
1172 mul
= mul
->operands
[0]->as_expression();
1176 if (mul
&& mul
->operation
== ir_binop_mul
)
1180 if (!mul
|| mul
->operation
!= ir_binop_mul
)
1183 nonmul
->accept(this);
1184 src_reg src0
= fix_3src_operand(this->result
);
1186 mul
->operands
[0]->accept(this);
1187 src_reg src1
= fix_3src_operand(this->result
);
1188 src1
.negate
^= mul_negate
;
1191 src1
.negate
= false;
1193 mul
->operands
[1]->accept(this);
1194 src_reg src2
= fix_3src_operand(this->result
);
1197 src2
.negate
= false;
1199 this->result
= src_reg(this, ir
->type
);
1200 emit(BRW_OPCODE_MAD
, dst_reg(this->result
), src0
, src1
, src2
);
1206 vec4_visitor::try_emit_b2f_of_compare(ir_expression
*ir
)
1208 /* This optimization relies on CMP setting the destination to 0 when
1209 * false. Early hardware only sets the least significant bit, and
1210 * leaves the other bits undefined. So we can't use it.
1212 if (devinfo
->gen
< 6)
1215 ir_expression
*const cmp
= ir
->operands
[0]->as_expression();
1220 switch (cmp
->operation
) {
1222 case ir_binop_greater
:
1223 case ir_binop_lequal
:
1224 case ir_binop_gequal
:
1225 case ir_binop_equal
:
1226 case ir_binop_nequal
:
1233 cmp
->operands
[0]->accept(this);
1234 const src_reg cmp_src0
= this->result
;
1236 cmp
->operands
[1]->accept(this);
1237 const src_reg cmp_src1
= this->result
;
1239 this->result
= src_reg(this, ir
->type
);
1241 emit(CMP(dst_reg(this->result
), cmp_src0
, cmp_src1
,
1242 brw_conditional_for_comparison(cmp
->operation
)));
1244 /* If the comparison is false, this->result will just happen to be zero.
1246 vec4_instruction
*const inst
= emit(BRW_OPCODE_SEL
, dst_reg(this->result
),
1247 this->result
, src_reg(1.0f
));
1248 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1249 inst
->predicate_inverse
= true;
1255 vec4_visitor::emit_minmax(enum brw_conditional_mod conditionalmod
, dst_reg dst
,
1256 src_reg src0
, src_reg src1
)
1258 vec4_instruction
*inst
;
1260 if (devinfo
->gen
>= 6) {
1261 inst
= emit(BRW_OPCODE_SEL
, dst
, src0
, src1
);
1262 inst
->conditional_mod
= conditionalmod
;
1264 emit(CMP(dst
, src0
, src1
, conditionalmod
));
1266 inst
= emit(BRW_OPCODE_SEL
, dst
, src0
, src1
);
1267 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1272 vec4_visitor::emit_lrp(const dst_reg
&dst
,
1273 const src_reg
&x
, const src_reg
&y
, const src_reg
&a
)
1275 if (devinfo
->gen
>= 6) {
1276 /* Note that the instruction's argument order is reversed from GLSL
1280 fix_3src_operand(a
), fix_3src_operand(y
), fix_3src_operand(x
)));
1282 /* Earlier generations don't support three source operations, so we
1283 * need to emit x*(1-a) + y*a.
1285 dst_reg y_times_a
= dst_reg(this, glsl_type::vec4_type
);
1286 dst_reg one_minus_a
= dst_reg(this, glsl_type::vec4_type
);
1287 dst_reg x_times_one_minus_a
= dst_reg(this, glsl_type::vec4_type
);
1288 y_times_a
.writemask
= dst
.writemask
;
1289 one_minus_a
.writemask
= dst
.writemask
;
1290 x_times_one_minus_a
.writemask
= dst
.writemask
;
1292 emit(MUL(y_times_a
, y
, a
));
1293 emit(ADD(one_minus_a
, negate(a
), src_reg(1.0f
)));
1294 emit(MUL(x_times_one_minus_a
, x
, src_reg(one_minus_a
)));
1295 emit(ADD(dst
, src_reg(x_times_one_minus_a
), src_reg(y_times_a
)));
1300 * Emits the instructions needed to perform a pull constant load. before_block
1301 * and before_inst can be NULL in which case the instruction will be appended
1302 * to the end of the instruction list.
1305 vec4_visitor::emit_pull_constant_load_reg(dst_reg dst
,
1308 bblock_t
*before_block
,
1309 vec4_instruction
*before_inst
)
1311 assert((before_inst
== NULL
&& before_block
== NULL
) ||
1312 (before_inst
&& before_block
));
1314 vec4_instruction
*pull
;
1316 if (devinfo
->gen
>= 9) {
1317 /* Gen9+ needs a message header in order to use SIMD4x2 mode */
1318 src_reg
header(this, glsl_type::uvec4_type
, 2);
1321 vec4_instruction(VS_OPCODE_SET_SIMD4X2_HEADER_GEN9
,
1325 emit_before(before_block
, before_inst
, pull
);
1329 dst_reg index_reg
= retype(offset(dst_reg(header
), 1),
1331 pull
= MOV(writemask(index_reg
, WRITEMASK_X
), offset_reg
);
1334 emit_before(before_block
, before_inst
, pull
);
1338 pull
= new(mem_ctx
) vec4_instruction(VS_OPCODE_PULL_CONSTANT_LOAD_GEN7
,
1343 pull
->header_present
= true;
1344 } else if (devinfo
->gen
>= 7) {
1345 dst_reg grf_offset
= dst_reg(this, glsl_type::int_type
);
1347 grf_offset
.type
= offset_reg
.type
;
1349 pull
= MOV(grf_offset
, offset_reg
);
1352 emit_before(before_block
, before_inst
, pull
);
1356 pull
= new(mem_ctx
) vec4_instruction(VS_OPCODE_PULL_CONSTANT_LOAD_GEN7
,
1359 src_reg(grf_offset
));
1362 pull
= new(mem_ctx
) vec4_instruction(VS_OPCODE_PULL_CONSTANT_LOAD
,
1366 pull
->base_mrf
= 14;
1371 emit_before(before_block
, before_inst
, pull
);
1377 vec4_visitor::visit(ir_expression
*ir
)
1379 unsigned int operand
;
1380 src_reg op
[ARRAY_SIZE(ir
->operands
)];
1381 vec4_instruction
*inst
;
1383 if (ir
->operation
== ir_binop_add
) {
1384 if (try_emit_mad(ir
))
1388 if (ir
->operation
== ir_unop_b2f
) {
1389 if (try_emit_b2f_of_compare(ir
))
1393 /* Storage for our result. Ideally for an assignment we'd be using
1394 * the actual storage for the result here, instead.
1396 dst_reg
result_dst(this, ir
->type
);
1397 src_reg
result_src(result_dst
);
1399 if (ir
->operation
== ir_triop_csel
) {
1400 ir
->operands
[1]->accept(this);
1401 op
[1] = this->result
;
1402 ir
->operands
[2]->accept(this);
1403 op
[2] = this->result
;
1405 enum brw_predicate predicate
;
1406 emit_bool_to_cond_code(ir
->operands
[0], &predicate
);
1407 inst
= emit(BRW_OPCODE_SEL
, result_dst
, op
[1], op
[2]);
1408 inst
->predicate
= predicate
;
1409 this->result
= result_src
;
1413 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1414 this->result
.file
= BAD_FILE
;
1415 ir
->operands
[operand
]->accept(this);
1416 if (this->result
.file
== BAD_FILE
) {
1417 fprintf(stderr
, "Failed to get tree for expression operand:\n");
1418 ir
->operands
[operand
]->fprint(stderr
);
1421 op
[operand
] = this->result
;
1423 /* Matrix expression operands should have been broken down to vector
1424 * operations already.
1426 assert(!ir
->operands
[operand
]->type
->is_matrix());
1429 /* If nothing special happens, this is the result. */
1430 this->result
= result_src
;
1432 switch (ir
->operation
) {
1433 case ir_unop_logic_not
:
1434 emit(NOT(result_dst
, op
[0]));
1437 op
[0].negate
= !op
[0].negate
;
1438 emit(MOV(result_dst
, op
[0]));
1442 op
[0].negate
= false;
1443 emit(MOV(result_dst
, op
[0]));
1447 if (ir
->type
->is_float()) {
1448 /* AND(val, 0x80000000) gives the sign bit.
1450 * Predicated OR ORs 1.0 (0x3f800000) with the sign bit if val is not
1453 emit(CMP(dst_null_f(), op
[0], src_reg(0.0f
), BRW_CONDITIONAL_NZ
));
1455 op
[0].type
= BRW_REGISTER_TYPE_UD
;
1456 result_dst
.type
= BRW_REGISTER_TYPE_UD
;
1457 emit(AND(result_dst
, op
[0], src_reg(0x80000000u
)));
1459 inst
= emit(OR(result_dst
, src_reg(result_dst
), src_reg(0x3f800000u
)));
1460 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1462 this->result
.type
= BRW_REGISTER_TYPE_F
;
1464 /* ASR(val, 31) -> negative val generates 0xffffffff (signed -1).
1465 * -> non-negative val generates 0x00000000.
1466 * Predicated OR sets 1 if val is positive.
1468 emit(CMP(dst_null_d(), op
[0], src_reg(0), BRW_CONDITIONAL_G
));
1470 emit(ASR(result_dst
, op
[0], src_reg(31)));
1472 inst
= emit(OR(result_dst
, src_reg(result_dst
), src_reg(1)));
1473 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1478 emit_math(SHADER_OPCODE_RCP
, result_dst
, op
[0]);
1482 emit_math(SHADER_OPCODE_EXP2
, result_dst
, op
[0]);
1485 emit_math(SHADER_OPCODE_LOG2
, result_dst
, op
[0]);
1489 unreachable("not reached: should be handled by ir_explog_to_explog2");
1491 emit_math(SHADER_OPCODE_SIN
, result_dst
, op
[0]);
1494 emit_math(SHADER_OPCODE_COS
, result_dst
, op
[0]);
1498 case ir_unop_dFdx_coarse
:
1499 case ir_unop_dFdx_fine
:
1501 case ir_unop_dFdy_coarse
:
1502 case ir_unop_dFdy_fine
:
1503 unreachable("derivatives not valid in vertex shader");
1505 case ir_unop_bitfield_reverse
:
1506 emit(BFREV(result_dst
, op
[0]));
1508 case ir_unop_bit_count
:
1509 emit(CBIT(result_dst
, op
[0]));
1511 case ir_unop_find_msb
: {
1512 src_reg temp
= src_reg(this, glsl_type::uint_type
);
1514 inst
= emit(FBH(dst_reg(temp
), op
[0]));
1515 inst
->dst
.writemask
= WRITEMASK_XYZW
;
1517 /* FBH counts from the MSB side, while GLSL's findMSB() wants the count
1518 * from the LSB side. If FBH didn't return an error (0xFFFFFFFF), then
1519 * subtract the result from 31 to convert the MSB count into an LSB count.
1522 /* FBH only supports UD type for dst, so use a MOV to convert UD to D. */
1523 temp
.swizzle
= BRW_SWIZZLE_NOOP
;
1524 emit(MOV(result_dst
, temp
));
1526 src_reg src_tmp
= src_reg(result_dst
);
1527 emit(CMP(dst_null_d(), src_tmp
, src_reg(-1), BRW_CONDITIONAL_NZ
));
1529 src_tmp
.negate
= true;
1530 inst
= emit(ADD(result_dst
, src_tmp
, src_reg(31)));
1531 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1534 case ir_unop_find_lsb
:
1535 emit(FBL(result_dst
, op
[0]));
1537 case ir_unop_saturate
:
1538 inst
= emit(MOV(result_dst
, op
[0]));
1539 inst
->saturate
= true;
1543 unreachable("not reached: should be handled by lower_noise");
1546 emit(ADD(result_dst
, op
[0], op
[1]));
1549 unreachable("not reached: should be handled by ir_sub_to_add_neg");
1552 if (devinfo
->gen
< 8 && ir
->type
->is_integer()) {
1553 /* For integer multiplication, the MUL uses the low 16 bits of one of
1554 * the operands (src0 through SNB, src1 on IVB and later). The MACH
1555 * accumulates in the contribution of the upper 16 bits of that
1556 * operand. If we can determine that one of the args is in the low
1557 * 16 bits, though, we can just emit a single MUL.
1559 if (ir
->operands
[0]->is_uint16_constant()) {
1560 if (devinfo
->gen
< 7)
1561 emit(MUL(result_dst
, op
[0], op
[1]));
1563 emit(MUL(result_dst
, op
[1], op
[0]));
1564 } else if (ir
->operands
[1]->is_uint16_constant()) {
1565 if (devinfo
->gen
< 7)
1566 emit(MUL(result_dst
, op
[1], op
[0]));
1568 emit(MUL(result_dst
, op
[0], op
[1]));
1570 struct brw_reg acc
= retype(brw_acc_reg(8), result_dst
.type
);
1572 emit(MUL(acc
, op
[0], op
[1]));
1573 emit(MACH(dst_null_d(), op
[0], op
[1]));
1574 emit(MOV(result_dst
, src_reg(acc
)));
1577 emit(MUL(result_dst
, op
[0], op
[1]));
1580 case ir_binop_imul_high
: {
1581 struct brw_reg acc
= retype(brw_acc_reg(8), result_dst
.type
);
1583 emit(MUL(acc
, op
[0], op
[1]));
1584 emit(MACH(result_dst
, op
[0], op
[1]));
1588 /* Floating point should be lowered by DIV_TO_MUL_RCP in the compiler. */
1589 assert(ir
->type
->is_integer());
1590 emit_math(SHADER_OPCODE_INT_QUOTIENT
, result_dst
, op
[0], op
[1]);
1592 case ir_binop_carry
: {
1593 struct brw_reg acc
= retype(brw_acc_reg(8), BRW_REGISTER_TYPE_UD
);
1595 emit(ADDC(dst_null_ud(), op
[0], op
[1]));
1596 emit(MOV(result_dst
, src_reg(acc
)));
1599 case ir_binop_borrow
: {
1600 struct brw_reg acc
= retype(brw_acc_reg(8), BRW_REGISTER_TYPE_UD
);
1602 emit(SUBB(dst_null_ud(), op
[0], op
[1]));
1603 emit(MOV(result_dst
, src_reg(acc
)));
1607 /* Floating point should be lowered by MOD_TO_FLOOR in the compiler. */
1608 assert(ir
->type
->is_integer());
1609 emit_math(SHADER_OPCODE_INT_REMAINDER
, result_dst
, op
[0], op
[1]);
1613 case ir_binop_greater
:
1614 case ir_binop_lequal
:
1615 case ir_binop_gequal
:
1616 case ir_binop_equal
:
1617 case ir_binop_nequal
: {
1618 if (devinfo
->gen
<= 5) {
1619 resolve_bool_comparison(ir
->operands
[0], &op
[0]);
1620 resolve_bool_comparison(ir
->operands
[1], &op
[1]);
1622 emit(CMP(result_dst
, op
[0], op
[1],
1623 brw_conditional_for_comparison(ir
->operation
)));
1627 case ir_binop_all_equal
:
1628 if (devinfo
->gen
<= 5) {
1629 resolve_bool_comparison(ir
->operands
[0], &op
[0]);
1630 resolve_bool_comparison(ir
->operands
[1], &op
[1]);
1633 /* "==" operator producing a scalar boolean. */
1634 if (ir
->operands
[0]->type
->is_vector() ||
1635 ir
->operands
[1]->type
->is_vector()) {
1636 emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_Z
));
1637 emit(MOV(result_dst
, src_reg(0)));
1638 inst
= emit(MOV(result_dst
, src_reg(~0)));
1639 inst
->predicate
= BRW_PREDICATE_ALIGN16_ALL4H
;
1641 emit(CMP(result_dst
, op
[0], op
[1], BRW_CONDITIONAL_Z
));
1644 case ir_binop_any_nequal
:
1645 if (devinfo
->gen
<= 5) {
1646 resolve_bool_comparison(ir
->operands
[0], &op
[0]);
1647 resolve_bool_comparison(ir
->operands
[1], &op
[1]);
1650 /* "!=" operator producing a scalar boolean. */
1651 if (ir
->operands
[0]->type
->is_vector() ||
1652 ir
->operands
[1]->type
->is_vector()) {
1653 emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_NZ
));
1655 emit(MOV(result_dst
, src_reg(0)));
1656 inst
= emit(MOV(result_dst
, src_reg(~0)));
1657 inst
->predicate
= BRW_PREDICATE_ALIGN16_ANY4H
;
1659 emit(CMP(result_dst
, op
[0], op
[1], BRW_CONDITIONAL_NZ
));
1664 if (devinfo
->gen
<= 5) {
1665 resolve_bool_comparison(ir
->operands
[0], &op
[0]);
1667 emit(CMP(dst_null_d(), op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
1668 emit(MOV(result_dst
, src_reg(0)));
1670 inst
= emit(MOV(result_dst
, src_reg(~0)));
1671 inst
->predicate
= BRW_PREDICATE_ALIGN16_ANY4H
;
1674 case ir_binop_logic_xor
:
1675 emit(XOR(result_dst
, op
[0], op
[1]));
1678 case ir_binop_logic_or
:
1679 emit(OR(result_dst
, op
[0], op
[1]));
1682 case ir_binop_logic_and
:
1683 emit(AND(result_dst
, op
[0], op
[1]));
1687 assert(ir
->operands
[0]->type
->is_vector());
1688 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1689 emit_dp(result_dst
, op
[0], op
[1], ir
->operands
[0]->type
->vector_elements
);
1693 emit_math(SHADER_OPCODE_SQRT
, result_dst
, op
[0]);
1696 emit_math(SHADER_OPCODE_RSQ
, result_dst
, op
[0]);
1699 case ir_unop_bitcast_i2f
:
1700 case ir_unop_bitcast_u2f
:
1701 this->result
= op
[0];
1702 this->result
.type
= BRW_REGISTER_TYPE_F
;
1705 case ir_unop_bitcast_f2i
:
1706 this->result
= op
[0];
1707 this->result
.type
= BRW_REGISTER_TYPE_D
;
1710 case ir_unop_bitcast_f2u
:
1711 this->result
= op
[0];
1712 this->result
.type
= BRW_REGISTER_TYPE_UD
;
1721 emit(MOV(result_dst
, op
[0]));
1724 emit(AND(result_dst
, op
[0], src_reg(1)));
1727 if (devinfo
->gen
<= 5) {
1728 resolve_bool_comparison(ir
->operands
[0], &op
[0]);
1730 op
[0].type
= BRW_REGISTER_TYPE_D
;
1731 result_dst
.type
= BRW_REGISTER_TYPE_D
;
1732 emit(AND(result_dst
, op
[0], src_reg(0x3f800000u
)));
1733 result_dst
.type
= BRW_REGISTER_TYPE_F
;
1736 emit(CMP(result_dst
, op
[0], src_reg(0.0f
), BRW_CONDITIONAL_NZ
));
1739 emit(CMP(result_dst
, op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
1743 emit(RNDZ(result_dst
, op
[0]));
1745 case ir_unop_ceil
: {
1746 src_reg tmp
= src_reg(this, ir
->type
);
1747 op
[0].negate
= !op
[0].negate
;
1748 emit(RNDD(dst_reg(tmp
), op
[0]));
1750 emit(MOV(result_dst
, tmp
));
1754 inst
= emit(RNDD(result_dst
, op
[0]));
1757 inst
= emit(FRC(result_dst
, op
[0]));
1759 case ir_unop_round_even
:
1760 emit(RNDE(result_dst
, op
[0]));
1764 emit_minmax(BRW_CONDITIONAL_L
, result_dst
, op
[0], op
[1]);
1767 emit_minmax(BRW_CONDITIONAL_GE
, result_dst
, op
[0], op
[1]);
1771 emit_math(SHADER_OPCODE_POW
, result_dst
, op
[0], op
[1]);
1774 case ir_unop_bit_not
:
1775 inst
= emit(NOT(result_dst
, op
[0]));
1777 case ir_binop_bit_and
:
1778 inst
= emit(AND(result_dst
, op
[0], op
[1]));
1780 case ir_binop_bit_xor
:
1781 inst
= emit(XOR(result_dst
, op
[0], op
[1]));
1783 case ir_binop_bit_or
:
1784 inst
= emit(OR(result_dst
, op
[0], op
[1]));
1787 case ir_binop_lshift
:
1788 inst
= emit(SHL(result_dst
, op
[0], op
[1]));
1791 case ir_binop_rshift
:
1792 if (ir
->type
->base_type
== GLSL_TYPE_INT
)
1793 inst
= emit(ASR(result_dst
, op
[0], op
[1]));
1795 inst
= emit(SHR(result_dst
, op
[0], op
[1]));
1799 emit(BFI1(result_dst
, op
[0], op
[1]));
1802 case ir_binop_ubo_load
: {
1803 ir_constant
*const_uniform_block
= ir
->operands
[0]->as_constant();
1804 ir_constant
*const_offset_ir
= ir
->operands
[1]->as_constant();
1805 unsigned const_offset
= const_offset_ir
? const_offset_ir
->value
.u
[0] : 0;
1808 /* Now, load the vector from that offset. */
1809 assert(ir
->type
->is_vector() || ir
->type
->is_scalar());
1811 src_reg packed_consts
= src_reg(this, glsl_type::vec4_type
);
1812 packed_consts
.type
= result
.type
;
1815 if (const_uniform_block
) {
1816 /* The block index is a constant, so just emit the binding table entry
1819 surf_index
= src_reg(prog_data
->base
.binding_table
.ubo_start
+
1820 const_uniform_block
->value
.u
[0]);
1822 /* The block index is not a constant. Evaluate the index expression
1823 * per-channel and add the base UBO index; the generator will select
1824 * a value from any live channel.
1826 surf_index
= src_reg(this, glsl_type::uint_type
);
1827 emit(ADD(dst_reg(surf_index
), op
[0],
1828 src_reg(prog_data
->base
.binding_table
.ubo_start
)));
1830 /* Assume this may touch any UBO. It would be nice to provide
1831 * a tighter bound, but the array information is already lowered away.
1833 brw_mark_surface_used(&prog_data
->base
,
1834 prog_data
->base
.binding_table
.ubo_start
+
1835 shader_prog
->NumUniformBlocks
- 1);
1838 if (const_offset_ir
) {
1839 if (devinfo
->gen
>= 8) {
1840 /* Store the offset in a GRF so we can send-from-GRF. */
1841 offset
= src_reg(this, glsl_type::int_type
);
1842 emit(MOV(dst_reg(offset
), src_reg(const_offset
/ 16)));
1844 /* Immediates are fine on older generations since they'll be moved
1845 * to a (potentially fake) MRF at the generator level.
1847 offset
= src_reg(const_offset
/ 16);
1850 offset
= src_reg(this, glsl_type::uint_type
);
1851 emit(SHR(dst_reg(offset
), op
[1], src_reg(4)));
1854 emit_pull_constant_load_reg(dst_reg(packed_consts
),
1857 NULL
, NULL
/* before_block/inst */);
1859 packed_consts
.swizzle
= brw_swizzle_for_size(ir
->type
->vector_elements
);
1860 packed_consts
.swizzle
+= BRW_SWIZZLE4(const_offset
% 16 / 4,
1861 const_offset
% 16 / 4,
1862 const_offset
% 16 / 4,
1863 const_offset
% 16 / 4);
1865 /* UBO bools are any nonzero int. We need to convert them to 0/~0. */
1866 if (ir
->type
->base_type
== GLSL_TYPE_BOOL
) {
1867 emit(CMP(result_dst
, packed_consts
, src_reg(0u),
1868 BRW_CONDITIONAL_NZ
));
1870 emit(MOV(result_dst
, packed_consts
));
1875 case ir_binop_vector_extract
:
1876 unreachable("should have been lowered by vec_index_to_cond_assign");
1879 op
[0] = fix_3src_operand(op
[0]);
1880 op
[1] = fix_3src_operand(op
[1]);
1881 op
[2] = fix_3src_operand(op
[2]);
1882 /* Note that the instruction's argument order is reversed from GLSL
1885 emit(MAD(result_dst
, op
[2], op
[1], op
[0]));
1889 emit_lrp(result_dst
, op
[0], op
[1], op
[2]);
1893 unreachable("already handled above");
1897 op
[0] = fix_3src_operand(op
[0]);
1898 op
[1] = fix_3src_operand(op
[1]);
1899 op
[2] = fix_3src_operand(op
[2]);
1900 emit(BFI2(result_dst
, op
[0], op
[1], op
[2]));
1903 case ir_triop_bitfield_extract
:
1904 op
[0] = fix_3src_operand(op
[0]);
1905 op
[1] = fix_3src_operand(op
[1]);
1906 op
[2] = fix_3src_operand(op
[2]);
1907 /* Note that the instruction's argument order is reversed from GLSL
1910 emit(BFE(result_dst
, op
[2], op
[1], op
[0]));
1913 case ir_triop_vector_insert
:
1914 unreachable("should have been lowered by lower_vector_insert");
1916 case ir_quadop_bitfield_insert
:
1917 unreachable("not reached: should be handled by "
1918 "bitfield_insert_to_bfm_bfi\n");
1920 case ir_quadop_vector
:
1921 unreachable("not reached: should be handled by lower_quadop_vector");
1923 case ir_unop_pack_half_2x16
:
1924 emit_pack_half_2x16(result_dst
, op
[0]);
1926 case ir_unop_unpack_half_2x16
:
1927 emit_unpack_half_2x16(result_dst
, op
[0]);
1929 case ir_unop_unpack_unorm_4x8
:
1930 emit_unpack_unorm_4x8(result_dst
, op
[0]);
1932 case ir_unop_unpack_snorm_4x8
:
1933 emit_unpack_snorm_4x8(result_dst
, op
[0]);
1935 case ir_unop_pack_unorm_4x8
:
1936 emit_pack_unorm_4x8(result_dst
, op
[0]);
1938 case ir_unop_pack_snorm_4x8
:
1939 emit_pack_snorm_4x8(result_dst
, op
[0]);
1941 case ir_unop_pack_snorm_2x16
:
1942 case ir_unop_pack_unorm_2x16
:
1943 case ir_unop_unpack_snorm_2x16
:
1944 case ir_unop_unpack_unorm_2x16
:
1945 unreachable("not reached: should be handled by lower_packing_builtins");
1946 case ir_unop_unpack_half_2x16_split_x
:
1947 case ir_unop_unpack_half_2x16_split_y
:
1948 case ir_binop_pack_half_2x16_split
:
1949 case ir_unop_interpolate_at_centroid
:
1950 case ir_binop_interpolate_at_sample
:
1951 case ir_binop_interpolate_at_offset
:
1952 unreachable("not reached: should not occur in vertex shader");
1953 case ir_binop_ldexp
:
1954 unreachable("not reached: should be handled by ldexp_to_arith()");
1962 case ir_unop_pack_double_2x32
:
1963 case ir_unop_unpack_double_2x32
:
1964 case ir_unop_frexp_sig
:
1965 case ir_unop_frexp_exp
:
1966 unreachable("fp64 todo");
1972 vec4_visitor::visit(ir_swizzle
*ir
)
1974 /* Note that this is only swizzles in expressions, not those on the left
1975 * hand side of an assignment, which do write masking. See ir_assignment
1978 const unsigned swz
= brw_compose_swizzle(
1979 brw_swizzle_for_size(ir
->type
->vector_elements
),
1980 BRW_SWIZZLE4(ir
->mask
.x
, ir
->mask
.y
, ir
->mask
.z
, ir
->mask
.w
));
1982 ir
->val
->accept(this);
1983 this->result
= swizzle(this->result
, swz
);
1987 vec4_visitor::visit(ir_dereference_variable
*ir
)
1989 const struct glsl_type
*type
= ir
->type
;
1990 dst_reg
*reg
= variable_storage(ir
->var
);
1993 fail("Failed to find variable storage for %s\n", ir
->var
->name
);
1994 this->result
= src_reg(brw_null_reg());
1998 this->result
= src_reg(*reg
);
2000 /* System values get their swizzle from the dst_reg writemask */
2001 if (ir
->var
->data
.mode
== ir_var_system_value
)
2004 if (type
->is_scalar() || type
->is_vector() || type
->is_matrix())
2005 this->result
.swizzle
= brw_swizzle_for_size(type
->vector_elements
);
2010 vec4_visitor::compute_array_stride(ir_dereference_array
*ir
)
2012 /* Under normal circumstances array elements are stored consecutively, so
2013 * the stride is equal to the size of the array element.
2015 return type_size(ir
->type
);
2020 vec4_visitor::visit(ir_dereference_array
*ir
)
2022 ir_constant
*constant_index
;
2024 int array_stride
= compute_array_stride(ir
);
2026 constant_index
= ir
->array_index
->constant_expression_value();
2028 ir
->array
->accept(this);
2031 if (constant_index
) {
2032 src
.reg_offset
+= constant_index
->value
.i
[0] * array_stride
;
2034 /* Variable index array dereference. It eats the "vec4" of the
2035 * base of the array and an index that offsets the Mesa register
2038 ir
->array_index
->accept(this);
2042 if (array_stride
== 1) {
2043 index_reg
= this->result
;
2045 index_reg
= src_reg(this, glsl_type::int_type
);
2047 emit(MUL(dst_reg(index_reg
), this->result
, src_reg(array_stride
)));
2051 src_reg temp
= src_reg(this, glsl_type::int_type
);
2053 emit(ADD(dst_reg(temp
), *src
.reladdr
, index_reg
));
2058 src
.reladdr
= ralloc(mem_ctx
, src_reg
);
2059 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
2062 /* If the type is smaller than a vec4, replicate the last channel out. */
2063 if (ir
->type
->is_scalar() || ir
->type
->is_vector() || ir
->type
->is_matrix())
2064 src
.swizzle
= brw_swizzle_for_size(ir
->type
->vector_elements
);
2066 src
.swizzle
= BRW_SWIZZLE_NOOP
;
2067 src
.type
= brw_type_for_base_type(ir
->type
);
2073 vec4_visitor::visit(ir_dereference_record
*ir
)
2076 const glsl_type
*struct_type
= ir
->record
->type
;
2079 ir
->record
->accept(this);
2081 for (i
= 0; i
< struct_type
->length
; i
++) {
2082 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
2084 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
2087 /* If the type is smaller than a vec4, replicate the last channel out. */
2088 if (ir
->type
->is_scalar() || ir
->type
->is_vector() || ir
->type
->is_matrix())
2089 this->result
.swizzle
= brw_swizzle_for_size(ir
->type
->vector_elements
);
2091 this->result
.swizzle
= BRW_SWIZZLE_NOOP
;
2092 this->result
.type
= brw_type_for_base_type(ir
->type
);
2094 this->result
.reg_offset
+= offset
;
2098 * We want to be careful in assignment setup to hit the actual storage
2099 * instead of potentially using a temporary like we might with the
2100 * ir_dereference handler.
2103 get_assignment_lhs(ir_dereference
*ir
, vec4_visitor
*v
)
2105 /* The LHS must be a dereference. If the LHS is a variable indexed array
2106 * access of a vector, it must be separated into a series conditional moves
2107 * before reaching this point (see ir_vec_index_to_cond_assign).
2109 assert(ir
->as_dereference());
2110 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
2112 assert(!deref_array
->array
->type
->is_vector());
2115 /* Use the rvalue deref handler for the most part. We'll ignore
2116 * swizzles in it and write swizzles using writemask, though.
2119 return dst_reg(v
->result
);
2123 vec4_visitor::emit_block_move(dst_reg
*dst
, src_reg
*src
,
2124 const struct glsl_type
*type
,
2125 enum brw_predicate predicate
)
2127 if (type
->base_type
== GLSL_TYPE_STRUCT
) {
2128 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2129 emit_block_move(dst
, src
, type
->fields
.structure
[i
].type
, predicate
);
2134 if (type
->is_array()) {
2135 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2136 emit_block_move(dst
, src
, type
->fields
.array
, predicate
);
2141 if (type
->is_matrix()) {
2142 const struct glsl_type
*vec_type
;
2144 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
2145 type
->vector_elements
, 1);
2147 for (int i
= 0; i
< type
->matrix_columns
; i
++) {
2148 emit_block_move(dst
, src
, vec_type
, predicate
);
2153 assert(type
->is_scalar() || type
->is_vector());
2155 dst
->type
= brw_type_for_base_type(type
);
2156 src
->type
= dst
->type
;
2158 dst
->writemask
= (1 << type
->vector_elements
) - 1;
2160 src
->swizzle
= brw_swizzle_for_size(type
->vector_elements
);
2162 vec4_instruction
*inst
= emit(MOV(*dst
, *src
));
2163 inst
->predicate
= predicate
;
2170 /* If the RHS processing resulted in an instruction generating a
2171 * temporary value, and it would be easy to rewrite the instruction to
2172 * generate its result right into the LHS instead, do so. This ends
2173 * up reliably removing instructions where it can be tricky to do so
2174 * later without real UD chain information.
2177 vec4_visitor::try_rewrite_rhs_to_dst(ir_assignment
*ir
,
2180 vec4_instruction
*pre_rhs_inst
,
2181 vec4_instruction
*last_rhs_inst
)
2183 /* This could be supported, but it would take more smarts. */
2187 if (pre_rhs_inst
== last_rhs_inst
)
2188 return false; /* No instructions generated to work with. */
2190 /* Make sure the last instruction generated our source reg. */
2191 if (src
.file
!= GRF
||
2192 src
.file
!= last_rhs_inst
->dst
.file
||
2193 src
.reg
!= last_rhs_inst
->dst
.reg
||
2194 src
.reg_offset
!= last_rhs_inst
->dst
.reg_offset
||
2198 last_rhs_inst
->predicate
!= BRW_PREDICATE_NONE
)
2201 /* Check that that last instruction fully initialized the channels
2202 * we want to use, in the order we want to use them. We could
2203 * potentially reswizzle the operands of many instructions so that
2204 * we could handle out of order channels, but don't yet.
2207 for (unsigned i
= 0; i
< 4; i
++) {
2208 if (dst
.writemask
& (1 << i
)) {
2209 if (!(last_rhs_inst
->dst
.writemask
& (1 << i
)))
2212 if (BRW_GET_SWZ(src
.swizzle
, i
) != i
)
2217 /* Success! Rewrite the instruction. */
2218 last_rhs_inst
->dst
.file
= dst
.file
;
2219 last_rhs_inst
->dst
.reg
= dst
.reg
;
2220 last_rhs_inst
->dst
.reg_offset
= dst
.reg_offset
;
2221 last_rhs_inst
->dst
.reladdr
= dst
.reladdr
;
2222 last_rhs_inst
->dst
.writemask
&= dst
.writemask
;
2228 vec4_visitor::visit(ir_assignment
*ir
)
2230 dst_reg dst
= get_assignment_lhs(ir
->lhs
, this);
2231 enum brw_predicate predicate
= BRW_PREDICATE_NONE
;
2233 if (!ir
->lhs
->type
->is_scalar() &&
2234 !ir
->lhs
->type
->is_vector()) {
2235 ir
->rhs
->accept(this);
2236 src_reg src
= this->result
;
2238 if (ir
->condition
) {
2239 emit_bool_to_cond_code(ir
->condition
, &predicate
);
2242 /* emit_block_move doesn't account for swizzles in the source register.
2243 * This should be ok, since the source register is a structure or an
2244 * array, and those can't be swizzled. But double-check to be sure.
2246 assert(src
.swizzle
==
2247 (ir
->rhs
->type
->is_matrix()
2248 ? brw_swizzle_for_size(ir
->rhs
->type
->vector_elements
)
2249 : BRW_SWIZZLE_NOOP
));
2251 emit_block_move(&dst
, &src
, ir
->rhs
->type
, predicate
);
2255 /* Now we're down to just a scalar/vector with writemasks. */
2258 vec4_instruction
*pre_rhs_inst
, *last_rhs_inst
;
2259 pre_rhs_inst
= (vec4_instruction
*)this->instructions
.get_tail();
2261 ir
->rhs
->accept(this);
2263 last_rhs_inst
= (vec4_instruction
*)this->instructions
.get_tail();
2268 assert(ir
->lhs
->type
->is_vector() ||
2269 ir
->lhs
->type
->is_scalar());
2270 dst
.writemask
= ir
->write_mask
;
2272 /* Swizzle a small RHS vector into the channels being written.
2274 * glsl ir treats write_mask as dictating how many channels are
2275 * present on the RHS while in our instructions we need to make
2276 * those channels appear in the slots of the vec4 they're written to.
2278 for (int i
= 0; i
< 4; i
++)
2279 swizzles
[i
] = (ir
->write_mask
& (1 << i
) ? src_chan
++ : 0);
2281 src_reg src
= swizzle(this->result
,
2282 BRW_SWIZZLE4(swizzles
[0], swizzles
[1],
2283 swizzles
[2], swizzles
[3]));
2285 if (try_rewrite_rhs_to_dst(ir
, dst
, src
, pre_rhs_inst
, last_rhs_inst
)) {
2289 if (ir
->condition
) {
2290 emit_bool_to_cond_code(ir
->condition
, &predicate
);
2293 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2294 vec4_instruction
*inst
= emit(MOV(dst
, src
));
2295 inst
->predicate
= predicate
;
2303 vec4_visitor::emit_constant_values(dst_reg
*dst
, ir_constant
*ir
)
2305 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
2306 foreach_in_list(ir_constant
, field_value
, &ir
->components
) {
2307 emit_constant_values(dst
, field_value
);
2312 if (ir
->type
->is_array()) {
2313 for (unsigned int i
= 0; i
< ir
->type
->length
; i
++) {
2314 emit_constant_values(dst
, ir
->array_elements
[i
]);
2319 if (ir
->type
->is_matrix()) {
2320 for (int i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
2321 float *vec
= &ir
->value
.f
[i
* ir
->type
->vector_elements
];
2323 for (int j
= 0; j
< ir
->type
->vector_elements
; j
++) {
2324 dst
->writemask
= 1 << j
;
2325 dst
->type
= BRW_REGISTER_TYPE_F
;
2327 emit(MOV(*dst
, src_reg(vec
[j
])));
2334 int remaining_writemask
= (1 << ir
->type
->vector_elements
) - 1;
2336 for (int i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2337 if (!(remaining_writemask
& (1 << i
)))
2340 dst
->writemask
= 1 << i
;
2341 dst
->type
= brw_type_for_base_type(ir
->type
);
2343 /* Find other components that match the one we're about to
2344 * write. Emits fewer instructions for things like vec4(0.5,
2347 for (int j
= i
+ 1; j
< ir
->type
->vector_elements
; j
++) {
2348 if (ir
->type
->base_type
== GLSL_TYPE_BOOL
) {
2349 if (ir
->value
.b
[i
] == ir
->value
.b
[j
])
2350 dst
->writemask
|= (1 << j
);
2352 /* u, i, and f storage all line up, so no need for a
2353 * switch case for comparing each type.
2355 if (ir
->value
.u
[i
] == ir
->value
.u
[j
])
2356 dst
->writemask
|= (1 << j
);
2360 switch (ir
->type
->base_type
) {
2361 case GLSL_TYPE_FLOAT
:
2362 emit(MOV(*dst
, src_reg(ir
->value
.f
[i
])));
2365 emit(MOV(*dst
, src_reg(ir
->value
.i
[i
])));
2367 case GLSL_TYPE_UINT
:
2368 emit(MOV(*dst
, src_reg(ir
->value
.u
[i
])));
2370 case GLSL_TYPE_BOOL
:
2371 emit(MOV(*dst
, src_reg(ir
->value
.b
[i
] != 0 ? ~0 : 0)));
2374 unreachable("Non-float/uint/int/bool constant");
2377 remaining_writemask
&= ~dst
->writemask
;
2383 vec4_visitor::visit(ir_constant
*ir
)
2385 dst_reg dst
= dst_reg(this, ir
->type
);
2386 this->result
= src_reg(dst
);
2388 emit_constant_values(&dst
, ir
);
2392 vec4_visitor::visit_atomic_counter_intrinsic(ir_call
*ir
)
2394 ir_dereference
*deref
= static_cast<ir_dereference
*>(
2395 ir
->actual_parameters
.get_head());
2396 ir_variable
*location
= deref
->variable_referenced();
2397 unsigned surf_index
= (prog_data
->base
.binding_table
.abo_start
+
2398 location
->data
.binding
);
2400 /* Calculate the surface offset */
2401 src_reg
offset(this, glsl_type::uint_type
);
2402 ir_dereference_array
*deref_array
= deref
->as_dereference_array();
2404 deref_array
->array_index
->accept(this);
2406 src_reg
tmp(this, glsl_type::uint_type
);
2407 emit(MUL(dst_reg(tmp
), this->result
, ATOMIC_COUNTER_SIZE
));
2408 emit(ADD(dst_reg(offset
), tmp
, location
->data
.atomic
.offset
));
2410 offset
= location
->data
.atomic
.offset
;
2413 /* Emit the appropriate machine instruction */
2414 const char *callee
= ir
->callee
->function_name();
2415 dst_reg dst
= get_assignment_lhs(ir
->return_deref
, this);
2417 if (!strcmp("__intrinsic_atomic_read", callee
)) {
2418 emit_untyped_surface_read(surf_index
, dst
, offset
);
2420 } else if (!strcmp("__intrinsic_atomic_increment", callee
)) {
2421 emit_untyped_atomic(BRW_AOP_INC
, surf_index
, dst
, offset
,
2422 src_reg(), src_reg());
2424 } else if (!strcmp("__intrinsic_atomic_predecrement", callee
)) {
2425 emit_untyped_atomic(BRW_AOP_PREDEC
, surf_index
, dst
, offset
,
2426 src_reg(), src_reg());
2431 vec4_visitor::visit(ir_call
*ir
)
2433 const char *callee
= ir
->callee
->function_name();
2435 if (!strcmp("__intrinsic_atomic_read", callee
) ||
2436 !strcmp("__intrinsic_atomic_increment", callee
) ||
2437 !strcmp("__intrinsic_atomic_predecrement", callee
)) {
2438 visit_atomic_counter_intrinsic(ir
);
2440 unreachable("Unsupported intrinsic.");
2445 vec4_visitor::emit_mcs_fetch(ir_texture
*ir
, src_reg coordinate
, src_reg sampler
)
2447 vec4_instruction
*inst
=
2448 new(mem_ctx
) vec4_instruction(SHADER_OPCODE_TXF_MCS
,
2449 dst_reg(this, glsl_type::uvec4_type
));
2452 inst
->src
[1] = sampler
;
2454 /* parameters are: u, v, r, lod; lod will always be zero due to api restrictions */
2455 int param_base
= inst
->base_mrf
;
2456 int coord_mask
= (1 << ir
->coordinate
->type
->vector_elements
) - 1;
2457 int zero_mask
= 0xf & ~coord_mask
;
2459 emit(MOV(dst_reg(MRF
, param_base
, ir
->coordinate
->type
, coord_mask
),
2462 emit(MOV(dst_reg(MRF
, param_base
, ir
->coordinate
->type
, zero_mask
),
2466 return src_reg(inst
->dst
);
2470 is_high_sampler(const struct brw_device_info
*devinfo
, src_reg sampler
)
2472 if (devinfo
->gen
< 8 && !devinfo
->is_haswell
)
2475 return sampler
.file
!= IMM
|| sampler
.fixed_hw_reg
.dw1
.ud
>= 16;
2479 vec4_visitor::visit(ir_texture
*ir
)
2482 _mesa_get_sampler_uniform_value(ir
->sampler
, shader_prog
, prog
);
2484 ir_rvalue
*nonconst_sampler_index
=
2485 _mesa_get_sampler_array_nonconst_index(ir
->sampler
);
2487 /* Handle non-constant sampler array indexing */
2488 src_reg sampler_reg
;
2489 if (nonconst_sampler_index
) {
2490 /* The highest sampler which may be used by this operation is
2491 * the last element of the array. Mark it here, because the generator
2492 * doesn't have enough information to determine the bound.
2494 uint32_t array_size
= ir
->sampler
->as_dereference_array()
2495 ->array
->type
->array_size();
2497 uint32_t max_used
= sampler
+ array_size
- 1;
2498 if (ir
->op
== ir_tg4
&& devinfo
->gen
< 8) {
2499 max_used
+= prog_data
->base
.binding_table
.gather_texture_start
;
2501 max_used
+= prog_data
->base
.binding_table
.texture_start
;
2504 brw_mark_surface_used(&prog_data
->base
, max_used
);
2506 /* Emit code to evaluate the actual indexing expression */
2507 nonconst_sampler_index
->accept(this);
2508 dst_reg
temp(this, glsl_type::uint_type
);
2509 emit(ADD(temp
, this->result
, src_reg(sampler
)))
2510 ->force_writemask_all
= true;
2511 sampler_reg
= src_reg(temp
);
2513 /* Single sampler, or constant array index; the indexing expression
2514 * is just an immediate.
2516 sampler_reg
= src_reg(sampler
);
2519 /* When tg4 is used with the degenerate ZERO/ONE swizzles, don't bother
2520 * emitting anything other than setting up the constant result.
2522 if (ir
->op
== ir_tg4
) {
2523 ir_constant
*chan
= ir
->lod_info
.component
->as_constant();
2524 int swiz
= GET_SWZ(key
->tex
.swizzles
[sampler
], chan
->value
.i
[0]);
2525 if (swiz
== SWIZZLE_ZERO
|| swiz
== SWIZZLE_ONE
) {
2526 dst_reg
result(this, ir
->type
);
2527 this->result
= src_reg(result
);
2528 emit(MOV(result
, src_reg(swiz
== SWIZZLE_ONE
? 1.0f
: 0.0f
)));
2533 /* Should be lowered by do_lower_texture_projection */
2534 assert(!ir
->projector
);
2536 /* Should be lowered */
2537 assert(!ir
->offset
|| !ir
->offset
->type
->is_array());
2539 /* Generate code to compute all the subexpression trees. This has to be
2540 * done before loading any values into MRFs for the sampler message since
2541 * generating these values may involve SEND messages that need the MRFs.
2544 if (ir
->coordinate
) {
2545 ir
->coordinate
->accept(this);
2546 coordinate
= this->result
;
2549 src_reg shadow_comparitor
;
2550 if (ir
->shadow_comparitor
) {
2551 ir
->shadow_comparitor
->accept(this);
2552 shadow_comparitor
= this->result
;
2555 bool has_nonconstant_offset
= ir
->offset
&& !ir
->offset
->as_constant();
2556 src_reg offset_value
;
2557 if (has_nonconstant_offset
) {
2558 ir
->offset
->accept(this);
2559 offset_value
= src_reg(this->result
);
2562 const glsl_type
*lod_type
= NULL
, *sample_index_type
= NULL
;
2563 src_reg lod
, dPdx
, dPdy
, sample_index
, mcs
;
2566 lod
= src_reg(0.0f
);
2567 lod_type
= glsl_type::float_type
;
2572 ir
->lod_info
.lod
->accept(this);
2574 lod_type
= ir
->lod_info
.lod
->type
;
2576 case ir_query_levels
:
2578 lod_type
= glsl_type::int_type
;
2581 ir
->lod_info
.sample_index
->accept(this);
2582 sample_index
= this->result
;
2583 sample_index_type
= ir
->lod_info
.sample_index
->type
;
2585 if (devinfo
->gen
>= 7 && key
->tex
.compressed_multisample_layout_mask
& (1<<sampler
))
2586 mcs
= emit_mcs_fetch(ir
, coordinate
, sampler_reg
);
2591 ir
->lod_info
.grad
.dPdx
->accept(this);
2592 dPdx
= this->result
;
2594 ir
->lod_info
.grad
.dPdy
->accept(this);
2595 dPdy
= this->result
;
2597 lod_type
= ir
->lod_info
.grad
.dPdx
->type
;
2607 case ir_tex
: opcode
= SHADER_OPCODE_TXL
; break;
2608 case ir_txl
: opcode
= SHADER_OPCODE_TXL
; break;
2609 case ir_txd
: opcode
= SHADER_OPCODE_TXD
; break;
2610 case ir_txf
: opcode
= SHADER_OPCODE_TXF
; break;
2611 case ir_txf_ms
: opcode
= SHADER_OPCODE_TXF_CMS
; break;
2612 case ir_txs
: opcode
= SHADER_OPCODE_TXS
; break;
2613 case ir_tg4
: opcode
= has_nonconstant_offset
2614 ? SHADER_OPCODE_TG4_OFFSET
: SHADER_OPCODE_TG4
; break;
2615 case ir_query_levels
: opcode
= SHADER_OPCODE_TXS
; break;
2617 unreachable("TXB is not valid for vertex shaders.");
2619 unreachable("LOD is not valid for vertex shaders.");
2621 unreachable("Unrecognized tex op");
2624 vec4_instruction
*inst
= new(mem_ctx
) vec4_instruction(
2625 opcode
, dst_reg(this, ir
->type
));
2627 if (ir
->offset
!= NULL
&& !has_nonconstant_offset
) {
2629 brw_texture_offset(ir
->offset
->as_constant()->value
.i
,
2630 ir
->offset
->type
->vector_elements
);
2633 /* Stuff the channel select bits in the top of the texture offset */
2634 if (ir
->op
== ir_tg4
)
2635 inst
->offset
|= gather_channel(ir
, sampler
) << 16;
2637 /* The message header is necessary for:
2639 * - Gen9+ for selecting SIMD4x2
2641 * - Gather channel selection
2642 * - Sampler indices too large to fit in a 4-bit value.
2644 inst
->header_present
=
2645 devinfo
->gen
< 5 || devinfo
->gen
>= 9 ||
2646 inst
->offset
!= 0 || ir
->op
== ir_tg4
||
2647 is_high_sampler(devinfo
, sampler_reg
);
2649 inst
->mlen
= inst
->header_present
+ 1; /* always at least one */
2650 inst
->dst
.writemask
= WRITEMASK_XYZW
;
2651 inst
->shadow_compare
= ir
->shadow_comparitor
!= NULL
;
2653 inst
->src
[1] = sampler_reg
;
2655 /* MRF for the first parameter */
2656 int param_base
= inst
->base_mrf
+ inst
->header_present
;
2658 if (ir
->op
== ir_txs
|| ir
->op
== ir_query_levels
) {
2659 int writemask
= devinfo
->gen
== 4 ? WRITEMASK_W
: WRITEMASK_X
;
2660 emit(MOV(dst_reg(MRF
, param_base
, lod_type
, writemask
), lod
));
2662 /* Load the coordinate */
2663 /* FINISHME: gl_clamp_mask and saturate */
2664 int coord_mask
= (1 << ir
->coordinate
->type
->vector_elements
) - 1;
2665 int zero_mask
= 0xf & ~coord_mask
;
2667 emit(MOV(dst_reg(MRF
, param_base
, ir
->coordinate
->type
, coord_mask
),
2670 if (zero_mask
!= 0) {
2671 emit(MOV(dst_reg(MRF
, param_base
, ir
->coordinate
->type
, zero_mask
),
2674 /* Load the shadow comparitor */
2675 if (ir
->shadow_comparitor
&& ir
->op
!= ir_txd
&& (ir
->op
!= ir_tg4
|| !has_nonconstant_offset
)) {
2676 emit(MOV(dst_reg(MRF
, param_base
+ 1, ir
->shadow_comparitor
->type
,
2678 shadow_comparitor
));
2682 /* Load the LOD info */
2683 if (ir
->op
== ir_tex
|| ir
->op
== ir_txl
) {
2685 if (devinfo
->gen
>= 5) {
2686 mrf
= param_base
+ 1;
2687 if (ir
->shadow_comparitor
) {
2688 writemask
= WRITEMASK_Y
;
2689 /* mlen already incremented */
2691 writemask
= WRITEMASK_X
;
2694 } else /* devinfo->gen == 4 */ {
2696 writemask
= WRITEMASK_W
;
2698 emit(MOV(dst_reg(MRF
, mrf
, lod_type
, writemask
), lod
));
2699 } else if (ir
->op
== ir_txf
) {
2700 emit(MOV(dst_reg(MRF
, param_base
, lod_type
, WRITEMASK_W
), lod
));
2701 } else if (ir
->op
== ir_txf_ms
) {
2702 emit(MOV(dst_reg(MRF
, param_base
+ 1, sample_index_type
, WRITEMASK_X
),
2704 if (devinfo
->gen
>= 7) {
2705 /* MCS data is in the first channel of `mcs`, but we need to get it into
2706 * the .y channel of the second vec4 of params, so replicate .x across
2707 * the whole vec4 and then mask off everything except .y
2709 mcs
.swizzle
= BRW_SWIZZLE_XXXX
;
2710 emit(MOV(dst_reg(MRF
, param_base
+ 1, glsl_type::uint_type
, WRITEMASK_Y
),
2714 } else if (ir
->op
== ir_txd
) {
2715 const glsl_type
*type
= lod_type
;
2717 if (devinfo
->gen
>= 5) {
2718 dPdx
.swizzle
= BRW_SWIZZLE4(SWIZZLE_X
,SWIZZLE_X
,SWIZZLE_Y
,SWIZZLE_Y
);
2719 dPdy
.swizzle
= BRW_SWIZZLE4(SWIZZLE_X
,SWIZZLE_X
,SWIZZLE_Y
,SWIZZLE_Y
);
2720 emit(MOV(dst_reg(MRF
, param_base
+ 1, type
, WRITEMASK_XZ
), dPdx
));
2721 emit(MOV(dst_reg(MRF
, param_base
+ 1, type
, WRITEMASK_YW
), dPdy
));
2724 if (ir
->type
->vector_elements
== 3 || ir
->shadow_comparitor
) {
2725 dPdx
.swizzle
= BRW_SWIZZLE_ZZZZ
;
2726 dPdy
.swizzle
= BRW_SWIZZLE_ZZZZ
;
2727 emit(MOV(dst_reg(MRF
, param_base
+ 2, type
, WRITEMASK_X
), dPdx
));
2728 emit(MOV(dst_reg(MRF
, param_base
+ 2, type
, WRITEMASK_Y
), dPdy
));
2731 if (ir
->shadow_comparitor
) {
2732 emit(MOV(dst_reg(MRF
, param_base
+ 2,
2733 ir
->shadow_comparitor
->type
, WRITEMASK_Z
),
2734 shadow_comparitor
));
2737 } else /* devinfo->gen == 4 */ {
2738 emit(MOV(dst_reg(MRF
, param_base
+ 1, type
, WRITEMASK_XYZ
), dPdx
));
2739 emit(MOV(dst_reg(MRF
, param_base
+ 2, type
, WRITEMASK_XYZ
), dPdy
));
2742 } else if (ir
->op
== ir_tg4
&& has_nonconstant_offset
) {
2743 if (ir
->shadow_comparitor
) {
2744 emit(MOV(dst_reg(MRF
, param_base
, ir
->shadow_comparitor
->type
, WRITEMASK_W
),
2745 shadow_comparitor
));
2748 emit(MOV(dst_reg(MRF
, param_base
+ 1, glsl_type::ivec2_type
, WRITEMASK_XY
),
2756 /* fixup num layers (z) for cube arrays: hardware returns faces * layers;
2757 * spec requires layers.
2759 if (ir
->op
== ir_txs
) {
2760 glsl_type
const *type
= ir
->sampler
->type
;
2761 if (type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_CUBE
&&
2762 type
->sampler_array
) {
2763 emit_math(SHADER_OPCODE_INT_QUOTIENT
,
2764 writemask(inst
->dst
, WRITEMASK_Z
),
2765 src_reg(inst
->dst
), src_reg(6));
2769 if (devinfo
->gen
== 6 && ir
->op
== ir_tg4
) {
2770 emit_gen6_gather_wa(key
->tex
.gen6_gather_wa
[sampler
], inst
->dst
);
2773 swizzle_result(ir
, src_reg(inst
->dst
), sampler
);
2777 * Apply workarounds for Gen6 gather with UINT/SINT
2780 vec4_visitor::emit_gen6_gather_wa(uint8_t wa
, dst_reg dst
)
2785 int width
= (wa
& WA_8BIT
) ? 8 : 16;
2786 dst_reg dst_f
= dst
;
2787 dst_f
.type
= BRW_REGISTER_TYPE_F
;
2789 /* Convert from UNORM to UINT */
2790 emit(MUL(dst_f
, src_reg(dst_f
), src_reg((float)((1 << width
) - 1))));
2791 emit(MOV(dst
, src_reg(dst_f
)));
2794 /* Reinterpret the UINT value as a signed INT value by
2795 * shifting the sign bit into place, then shifting back
2798 emit(SHL(dst
, src_reg(dst
), src_reg(32 - width
)));
2799 emit(ASR(dst
, src_reg(dst
), src_reg(32 - width
)));
2804 * Set up the gather channel based on the swizzle, for gather4.
2807 vec4_visitor::gather_channel(ir_texture
*ir
, uint32_t sampler
)
2809 ir_constant
*chan
= ir
->lod_info
.component
->as_constant();
2810 int swiz
= GET_SWZ(key
->tex
.swizzles
[sampler
], chan
->value
.i
[0]);
2812 case SWIZZLE_X
: return 0;
2814 /* gather4 sampler is broken for green channel on RG32F --
2815 * we must ask for blue instead.
2817 if (key
->tex
.gather_channel_quirk_mask
& (1<<sampler
))
2820 case SWIZZLE_Z
: return 2;
2821 case SWIZZLE_W
: return 3;
2823 unreachable("Not reached"); /* zero, one swizzles handled already */
2828 vec4_visitor::swizzle_result(ir_texture
*ir
, src_reg orig_val
, uint32_t sampler
)
2830 int s
= key
->tex
.swizzles
[sampler
];
2832 this->result
= src_reg(this, ir
->type
);
2833 dst_reg
swizzled_result(this->result
);
2835 if (ir
->op
== ir_query_levels
) {
2836 /* # levels is in .w */
2837 orig_val
.swizzle
= BRW_SWIZZLE4(SWIZZLE_W
, SWIZZLE_W
, SWIZZLE_W
, SWIZZLE_W
);
2838 emit(MOV(swizzled_result
, orig_val
));
2842 if (ir
->op
== ir_txs
|| ir
->type
== glsl_type::float_type
2843 || s
== SWIZZLE_NOOP
|| ir
->op
== ir_tg4
) {
2844 emit(MOV(swizzled_result
, orig_val
));
2849 int zero_mask
= 0, one_mask
= 0, copy_mask
= 0;
2850 int swizzle
[4] = {0};
2852 for (int i
= 0; i
< 4; i
++) {
2853 switch (GET_SWZ(s
, i
)) {
2855 zero_mask
|= (1 << i
);
2858 one_mask
|= (1 << i
);
2861 copy_mask
|= (1 << i
);
2862 swizzle
[i
] = GET_SWZ(s
, i
);
2868 orig_val
.swizzle
= BRW_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
2869 swizzled_result
.writemask
= copy_mask
;
2870 emit(MOV(swizzled_result
, orig_val
));
2874 swizzled_result
.writemask
= zero_mask
;
2875 emit(MOV(swizzled_result
, src_reg(0.0f
)));
2879 swizzled_result
.writemask
= one_mask
;
2880 emit(MOV(swizzled_result
, src_reg(1.0f
)));
2885 vec4_visitor::visit(ir_return
*)
2887 unreachable("not reached");
2891 vec4_visitor::visit(ir_discard
*)
2893 unreachable("not reached");
2897 vec4_visitor::visit(ir_if
*ir
)
2899 /* Don't point the annotation at the if statement, because then it plus
2900 * the then and else blocks get printed.
2902 this->base_ir
= ir
->condition
;
2904 if (devinfo
->gen
== 6) {
2907 enum brw_predicate predicate
;
2908 emit_bool_to_cond_code(ir
->condition
, &predicate
);
2909 emit(IF(predicate
));
2912 visit_instructions(&ir
->then_instructions
);
2914 if (!ir
->else_instructions
.is_empty()) {
2915 this->base_ir
= ir
->condition
;
2916 emit(BRW_OPCODE_ELSE
);
2918 visit_instructions(&ir
->else_instructions
);
2921 this->base_ir
= ir
->condition
;
2922 emit(BRW_OPCODE_ENDIF
);
2926 vec4_visitor::visit(ir_emit_vertex
*)
2928 unreachable("not reached");
2932 vec4_visitor::visit(ir_end_primitive
*)
2934 unreachable("not reached");
2938 vec4_visitor::emit_untyped_atomic(unsigned atomic_op
, unsigned surf_index
,
2939 dst_reg dst
, src_reg offset
,
2940 src_reg src0
, src_reg src1
)
2944 /* Set the atomic operation offset. */
2945 emit(MOV(brw_writemask(brw_uvec_mrf(8, mlen
, 0), WRITEMASK_X
), offset
));
2948 /* Set the atomic operation arguments. */
2949 if (src0
.file
!= BAD_FILE
) {
2950 emit(MOV(brw_writemask(brw_uvec_mrf(8, mlen
, 0), WRITEMASK_X
), src0
));
2954 if (src1
.file
!= BAD_FILE
) {
2955 emit(MOV(brw_writemask(brw_uvec_mrf(8, mlen
, 0), WRITEMASK_X
), src1
));
2959 /* Emit the instruction. Note that this maps to the normal SIMD8
2960 * untyped atomic message on Ivy Bridge, but that's OK because
2961 * unused channels will be masked out.
2963 vec4_instruction
*inst
= emit(SHADER_OPCODE_UNTYPED_ATOMIC
, dst
,
2964 src_reg(atomic_op
), src_reg(surf_index
));
2970 vec4_visitor::emit_untyped_surface_read(unsigned surf_index
, dst_reg dst
,
2973 /* Set the surface read offset. */
2974 emit(MOV(brw_writemask(brw_uvec_mrf(8, 0, 0), WRITEMASK_X
), offset
));
2976 /* Emit the instruction. Note that this maps to the normal SIMD8
2977 * untyped surface read message, but that's OK because unused
2978 * channels will be masked out.
2980 vec4_instruction
*inst
= emit(SHADER_OPCODE_UNTYPED_SURFACE_READ
,
2981 dst
, src_reg(surf_index
));
2987 vec4_visitor::emit_ndc_computation()
2989 /* Get the position */
2990 src_reg pos
= src_reg(output_reg
[VARYING_SLOT_POS
]);
2992 /* Build ndc coords, which are (x/w, y/w, z/w, 1/w) */
2993 dst_reg ndc
= dst_reg(this, glsl_type::vec4_type
);
2994 output_reg
[BRW_VARYING_SLOT_NDC
] = ndc
;
2996 current_annotation
= "NDC";
2997 dst_reg ndc_w
= ndc
;
2998 ndc_w
.writemask
= WRITEMASK_W
;
2999 src_reg pos_w
= pos
;
3000 pos_w
.swizzle
= BRW_SWIZZLE4(SWIZZLE_W
, SWIZZLE_W
, SWIZZLE_W
, SWIZZLE_W
);
3001 emit_math(SHADER_OPCODE_RCP
, ndc_w
, pos_w
);
3003 dst_reg ndc_xyz
= ndc
;
3004 ndc_xyz
.writemask
= WRITEMASK_XYZ
;
3006 emit(MUL(ndc_xyz
, pos
, src_reg(ndc_w
)));
3010 vec4_visitor::emit_psiz_and_flags(dst_reg reg
)
3012 if (devinfo
->gen
< 6 &&
3013 ((prog_data
->vue_map
.slots_valid
& VARYING_BIT_PSIZ
) ||
3014 key
->userclip_active
|| devinfo
->has_negative_rhw_bug
)) {
3015 dst_reg header1
= dst_reg(this, glsl_type::uvec4_type
);
3016 dst_reg header1_w
= header1
;
3017 header1_w
.writemask
= WRITEMASK_W
;
3019 emit(MOV(header1
, 0u));
3021 if (prog_data
->vue_map
.slots_valid
& VARYING_BIT_PSIZ
) {
3022 src_reg psiz
= src_reg(output_reg
[VARYING_SLOT_PSIZ
]);
3024 current_annotation
= "Point size";
3025 emit(MUL(header1_w
, psiz
, src_reg((float)(1 << 11))));
3026 emit(AND(header1_w
, src_reg(header1_w
), 0x7ff << 8));
3029 if (key
->userclip_active
) {
3030 current_annotation
= "Clipping flags";
3031 dst_reg flags0
= dst_reg(this, glsl_type::uint_type
);
3032 dst_reg flags1
= dst_reg(this, glsl_type::uint_type
);
3034 emit(CMP(dst_null_f(), src_reg(output_reg
[VARYING_SLOT_CLIP_DIST0
]), src_reg(0.0f
), BRW_CONDITIONAL_L
));
3035 emit(VS_OPCODE_UNPACK_FLAGS_SIMD4X2
, flags0
, src_reg(0));
3036 emit(OR(header1_w
, src_reg(header1_w
), src_reg(flags0
)));
3038 emit(CMP(dst_null_f(), src_reg(output_reg
[VARYING_SLOT_CLIP_DIST1
]), src_reg(0.0f
), BRW_CONDITIONAL_L
));
3039 emit(VS_OPCODE_UNPACK_FLAGS_SIMD4X2
, flags1
, src_reg(0));
3040 emit(SHL(flags1
, src_reg(flags1
), src_reg(4)));
3041 emit(OR(header1_w
, src_reg(header1_w
), src_reg(flags1
)));
3044 /* i965 clipping workaround:
3045 * 1) Test for -ve rhw
3047 * set ndc = (0,0,0,0)
3050 * Later, clipping will detect ucp[6] and ensure the primitive is
3051 * clipped against all fixed planes.
3053 if (devinfo
->has_negative_rhw_bug
) {
3054 src_reg ndc_w
= src_reg(output_reg
[BRW_VARYING_SLOT_NDC
]);
3055 ndc_w
.swizzle
= BRW_SWIZZLE_WWWW
;
3056 emit(CMP(dst_null_f(), ndc_w
, src_reg(0.0f
), BRW_CONDITIONAL_L
));
3057 vec4_instruction
*inst
;
3058 inst
= emit(OR(header1_w
, src_reg(header1_w
), src_reg(1u << 6)));
3059 inst
->predicate
= BRW_PREDICATE_NORMAL
;
3060 inst
= emit(MOV(output_reg
[BRW_VARYING_SLOT_NDC
], src_reg(0.0f
)));
3061 inst
->predicate
= BRW_PREDICATE_NORMAL
;
3064 emit(MOV(retype(reg
, BRW_REGISTER_TYPE_UD
), src_reg(header1
)));
3065 } else if (devinfo
->gen
< 6) {
3066 emit(MOV(retype(reg
, BRW_REGISTER_TYPE_UD
), 0u));
3068 emit(MOV(retype(reg
, BRW_REGISTER_TYPE_D
), src_reg(0)));
3069 if (prog_data
->vue_map
.slots_valid
& VARYING_BIT_PSIZ
) {
3070 dst_reg reg_w
= reg
;
3071 reg_w
.writemask
= WRITEMASK_W
;
3072 emit(MOV(reg_w
, src_reg(output_reg
[VARYING_SLOT_PSIZ
])));
3074 if (prog_data
->vue_map
.slots_valid
& VARYING_BIT_LAYER
) {
3075 dst_reg reg_y
= reg
;
3076 reg_y
.writemask
= WRITEMASK_Y
;
3077 reg_y
.type
= BRW_REGISTER_TYPE_D
;
3078 emit(MOV(reg_y
, src_reg(output_reg
[VARYING_SLOT_LAYER
])));
3080 if (prog_data
->vue_map
.slots_valid
& VARYING_BIT_VIEWPORT
) {
3081 dst_reg reg_z
= reg
;
3082 reg_z
.writemask
= WRITEMASK_Z
;
3083 reg_z
.type
= BRW_REGISTER_TYPE_D
;
3084 emit(MOV(reg_z
, src_reg(output_reg
[VARYING_SLOT_VIEWPORT
])));
3090 vec4_visitor::emit_clip_distances(dst_reg reg
, int offset
)
3092 /* From the GLSL 1.30 spec, section 7.1 (Vertex Shader Special Variables):
3094 * "If a linked set of shaders forming the vertex stage contains no
3095 * static write to gl_ClipVertex or gl_ClipDistance, but the
3096 * application has requested clipping against user clip planes through
3097 * the API, then the coordinate written to gl_Position is used for
3098 * comparison against the user clip planes."
3100 * This function is only called if the shader didn't write to
3101 * gl_ClipDistance. Accordingly, we use gl_ClipVertex to perform clipping
3102 * if the user wrote to it; otherwise we use gl_Position.
3104 gl_varying_slot clip_vertex
= VARYING_SLOT_CLIP_VERTEX
;
3105 if (!(prog_data
->vue_map
.slots_valid
& VARYING_BIT_CLIP_VERTEX
)) {
3106 clip_vertex
= VARYING_SLOT_POS
;
3109 for (int i
= 0; i
+ offset
< key
->nr_userclip_plane_consts
&& i
< 4;
3111 reg
.writemask
= 1 << i
;
3113 src_reg(output_reg
[clip_vertex
]),
3114 src_reg(this->userplane
[i
+ offset
])));
3119 vec4_visitor::emit_generic_urb_slot(dst_reg reg
, int varying
)
3121 assert (varying
< VARYING_SLOT_MAX
);
3122 reg
.type
= output_reg
[varying
].type
;
3123 current_annotation
= output_reg_annotation
[varying
];
3124 /* Copy the register, saturating if necessary */
3125 return emit(MOV(reg
, src_reg(output_reg
[varying
])));
3129 vec4_visitor::emit_urb_slot(dst_reg reg
, int varying
)
3131 reg
.type
= BRW_REGISTER_TYPE_F
;
3134 case VARYING_SLOT_PSIZ
:
3136 /* PSIZ is always in slot 0, and is coupled with other flags. */
3137 current_annotation
= "indices, point width, clip flags";
3138 emit_psiz_and_flags(reg
);
3141 case BRW_VARYING_SLOT_NDC
:
3142 current_annotation
= "NDC";
3143 emit(MOV(reg
, src_reg(output_reg
[BRW_VARYING_SLOT_NDC
])));
3145 case VARYING_SLOT_POS
:
3146 current_annotation
= "gl_Position";
3147 emit(MOV(reg
, src_reg(output_reg
[VARYING_SLOT_POS
])));
3149 case VARYING_SLOT_EDGE
:
3150 /* This is present when doing unfilled polygons. We're supposed to copy
3151 * the edge flag from the user-provided vertex array
3152 * (glEdgeFlagPointer), or otherwise we'll copy from the current value
3153 * of that attribute (starts as 1.0f). This is then used in clipping to
3154 * determine which edges should be drawn as wireframe.
3156 current_annotation
= "edge flag";
3157 emit(MOV(reg
, src_reg(dst_reg(ATTR
, VERT_ATTRIB_EDGEFLAG
,
3158 glsl_type::float_type
, WRITEMASK_XYZW
))));
3160 case BRW_VARYING_SLOT_PAD
:
3161 /* No need to write to this slot */
3163 case VARYING_SLOT_COL0
:
3164 case VARYING_SLOT_COL1
:
3165 case VARYING_SLOT_BFC0
:
3166 case VARYING_SLOT_BFC1
: {
3167 /* These built-in varyings are only supported in compatibility mode,
3168 * and we only support GS in core profile. So, this must be a vertex
3171 assert(stage
== MESA_SHADER_VERTEX
);
3172 vec4_instruction
*inst
= emit_generic_urb_slot(reg
, varying
);
3173 if (((struct brw_vs_prog_key
*) key
)->clamp_vertex_color
)
3174 inst
->saturate
= true;
3179 emit_generic_urb_slot(reg
, varying
);
3185 align_interleaved_urb_mlen(const struct brw_device_info
*devinfo
, int mlen
)
3187 if (devinfo
->gen
>= 6) {
3188 /* URB data written (does not include the message header reg) must
3189 * be a multiple of 256 bits, or 2 VS registers. See vol5c.5,
3190 * section 5.4.3.2.2: URB_INTERLEAVED.
3192 * URB entries are allocated on a multiple of 1024 bits, so an
3193 * extra 128 bits written here to make the end align to 256 is
3196 if ((mlen
% 2) != 1)
3205 * Generates the VUE payload plus the necessary URB write instructions to
3208 * The VUE layout is documented in Volume 2a.
3211 vec4_visitor::emit_vertex()
3213 /* MRF 0 is reserved for the debugger, so start with message header
3218 /* In the process of generating our URB write message contents, we
3219 * may need to unspill a register or load from an array. Those
3220 * reads would use MRFs 14-15.
3222 int max_usable_mrf
= 13;
3224 /* The following assertion verifies that max_usable_mrf causes an
3225 * even-numbered amount of URB write data, which will meet gen6's
3226 * requirements for length alignment.
3228 assert ((max_usable_mrf
- base_mrf
) % 2 == 0);
3230 /* First mrf is the g0-based message header containing URB handles and
3233 emit_urb_write_header(mrf
++);
3235 if (devinfo
->gen
< 6) {
3236 emit_ndc_computation();
3239 /* Lower legacy ff and ClipVertex clipping to clip distances */
3240 if (key
->userclip_active
&& !prog
->UsesClipDistanceOut
) {
3241 current_annotation
= "user clip distances";
3243 output_reg
[VARYING_SLOT_CLIP_DIST0
] = dst_reg(this, glsl_type::vec4_type
);
3244 output_reg
[VARYING_SLOT_CLIP_DIST1
] = dst_reg(this, glsl_type::vec4_type
);
3246 emit_clip_distances(output_reg
[VARYING_SLOT_CLIP_DIST0
], 0);
3247 emit_clip_distances(output_reg
[VARYING_SLOT_CLIP_DIST1
], 4);
3250 /* We may need to split this up into several URB writes, so do them in a
3254 bool complete
= false;
3256 /* URB offset is in URB row increments, and each of our MRFs is half of
3257 * one of those, since we're doing interleaved writes.
3259 int offset
= slot
/ 2;
3262 for (; slot
< prog_data
->vue_map
.num_slots
; ++slot
) {
3263 emit_urb_slot(dst_reg(MRF
, mrf
++),
3264 prog_data
->vue_map
.slot_to_varying
[slot
]);
3266 /* If this was max_usable_mrf, we can't fit anything more into this
3269 if (mrf
> max_usable_mrf
) {
3275 complete
= slot
>= prog_data
->vue_map
.num_slots
;
3276 current_annotation
= "URB write";
3277 vec4_instruction
*inst
= emit_urb_write_opcode(complete
);
3278 inst
->base_mrf
= base_mrf
;
3279 inst
->mlen
= align_interleaved_urb_mlen(devinfo
, mrf
- base_mrf
);
3280 inst
->offset
+= offset
;
3286 vec4_visitor::get_scratch_offset(bblock_t
*block
, vec4_instruction
*inst
,
3287 src_reg
*reladdr
, int reg_offset
)
3289 /* Because we store the values to scratch interleaved like our
3290 * vertex data, we need to scale the vec4 index by 2.
3292 int message_header_scale
= 2;
3294 /* Pre-gen6, the message header uses byte offsets instead of vec4
3295 * (16-byte) offset units.
3297 if (devinfo
->gen
< 6)
3298 message_header_scale
*= 16;
3301 src_reg index
= src_reg(this, glsl_type::int_type
);
3303 emit_before(block
, inst
, ADD(dst_reg(index
), *reladdr
,
3304 src_reg(reg_offset
)));
3305 emit_before(block
, inst
, MUL(dst_reg(index
), index
,
3306 src_reg(message_header_scale
)));
3310 return src_reg(reg_offset
* message_header_scale
);
3315 vec4_visitor::get_pull_constant_offset(bblock_t
* block
, vec4_instruction
*inst
,
3316 src_reg
*reladdr
, int reg_offset
)
3319 src_reg index
= src_reg(this, glsl_type::int_type
);
3321 emit_before(block
, inst
, ADD(dst_reg(index
), *reladdr
,
3322 src_reg(reg_offset
)));
3324 /* Pre-gen6, the message header uses byte offsets instead of vec4
3325 * (16-byte) offset units.
3327 if (devinfo
->gen
< 6) {
3328 emit_before(block
, inst
, MUL(dst_reg(index
), index
, src_reg(16)));
3332 } else if (devinfo
->gen
>= 8) {
3333 /* Store the offset in a GRF so we can send-from-GRF. */
3334 src_reg offset
= src_reg(this, glsl_type::int_type
);
3335 emit_before(block
, inst
, MOV(dst_reg(offset
), src_reg(reg_offset
)));
3338 int message_header_scale
= devinfo
->gen
< 6 ? 16 : 1;
3339 return src_reg(reg_offset
* message_header_scale
);
3344 * Emits an instruction before @inst to load the value named by @orig_src
3345 * from scratch space at @base_offset to @temp.
3347 * @base_offset is measured in 32-byte units (the size of a register).
3350 vec4_visitor::emit_scratch_read(bblock_t
*block
, vec4_instruction
*inst
,
3351 dst_reg temp
, src_reg orig_src
,
3354 int reg_offset
= base_offset
+ orig_src
.reg_offset
;
3355 src_reg index
= get_scratch_offset(block
, inst
, orig_src
.reladdr
,
3358 emit_before(block
, inst
, SCRATCH_READ(temp
, index
));
3362 * Emits an instruction after @inst to store the value to be written
3363 * to @orig_dst to scratch space at @base_offset, from @temp.
3365 * @base_offset is measured in 32-byte units (the size of a register).
3368 vec4_visitor::emit_scratch_write(bblock_t
*block
, vec4_instruction
*inst
,
3371 int reg_offset
= base_offset
+ inst
->dst
.reg_offset
;
3372 src_reg index
= get_scratch_offset(block
, inst
, inst
->dst
.reladdr
,
3375 /* Create a temporary register to store *inst's result in.
3377 * We have to be careful in MOVing from our temporary result register in
3378 * the scratch write. If we swizzle from channels of the temporary that
3379 * weren't initialized, it will confuse live interval analysis, which will
3380 * make spilling fail to make progress.
3382 const src_reg temp
= swizzle(retype(src_reg(this, glsl_type::vec4_type
),
3384 brw_swizzle_for_mask(inst
->dst
.writemask
));
3385 dst_reg dst
= dst_reg(brw_writemask(brw_vec8_grf(0, 0),
3386 inst
->dst
.writemask
));
3387 vec4_instruction
*write
= SCRATCH_WRITE(dst
, temp
, index
);
3388 write
->predicate
= inst
->predicate
;
3389 write
->ir
= inst
->ir
;
3390 write
->annotation
= inst
->annotation
;
3391 inst
->insert_after(block
, write
);
3393 inst
->dst
.file
= temp
.file
;
3394 inst
->dst
.reg
= temp
.reg
;
3395 inst
->dst
.reg_offset
= temp
.reg_offset
;
3396 inst
->dst
.reladdr
= NULL
;
3400 * Checks if \p src and/or \p src.reladdr require a scratch read, and if so,
3401 * adds the scratch read(s) before \p inst. The function also checks for
3402 * recursive reladdr scratch accesses, issuing the corresponding scratch
3403 * loads and rewriting reladdr references accordingly.
3405 * \return \p src if it did not require a scratch load, otherwise, the
3406 * register holding the result of the scratch load that the caller should
3407 * use to rewrite src.
3410 vec4_visitor::emit_resolve_reladdr(int scratch_loc
[], bblock_t
*block
,
3411 vec4_instruction
*inst
, src_reg src
)
3413 /* Resolve recursive reladdr scratch access by calling ourselves
3417 *src
.reladdr
= emit_resolve_reladdr(scratch_loc
, block
, inst
,
3420 /* Now handle scratch access on src */
3421 if (src
.file
== GRF
&& scratch_loc
[src
.reg
] != -1) {
3422 dst_reg temp
= dst_reg(this, glsl_type::vec4_type
);
3423 emit_scratch_read(block
, inst
, temp
, src
, scratch_loc
[src
.reg
]);
3425 src
.reg_offset
= temp
.reg_offset
;
3433 * We can't generally support array access in GRF space, because a
3434 * single instruction's destination can only span 2 contiguous
3435 * registers. So, we send all GRF arrays that get variable index
3436 * access to scratch space.
3439 vec4_visitor::move_grf_array_access_to_scratch()
3441 int scratch_loc
[this->alloc
.count
];
3442 memset(scratch_loc
, -1, sizeof(scratch_loc
));
3444 /* First, calculate the set of virtual GRFs that need to be punted
3445 * to scratch due to having any array access on them, and where in
3448 foreach_block_and_inst(block
, vec4_instruction
, inst
, cfg
) {
3449 if (inst
->dst
.file
== GRF
&& inst
->dst
.reladdr
) {
3450 if (scratch_loc
[inst
->dst
.reg
] == -1) {
3451 scratch_loc
[inst
->dst
.reg
] = c
->last_scratch
;
3452 c
->last_scratch
+= this->alloc
.sizes
[inst
->dst
.reg
];
3455 for (src_reg
*iter
= inst
->dst
.reladdr
;
3457 iter
= iter
->reladdr
) {
3458 if (iter
->file
== GRF
&& scratch_loc
[iter
->reg
] == -1) {
3459 scratch_loc
[iter
->reg
] = c
->last_scratch
;
3460 c
->last_scratch
+= this->alloc
.sizes
[iter
->reg
];
3465 for (int i
= 0 ; i
< 3; i
++) {
3466 for (src_reg
*iter
= &inst
->src
[i
];
3468 iter
= iter
->reladdr
) {
3469 if (iter
->file
== GRF
&& scratch_loc
[iter
->reg
] == -1) {
3470 scratch_loc
[iter
->reg
] = c
->last_scratch
;
3471 c
->last_scratch
+= this->alloc
.sizes
[iter
->reg
];
3477 /* Now, for anything that will be accessed through scratch, rewrite
3478 * it to load/store. Note that this is a _safe list walk, because
3479 * we may generate a new scratch_write instruction after the one
3482 foreach_block_and_inst_safe(block
, vec4_instruction
, inst
, cfg
) {
3483 /* Set up the annotation tracking for new generated instructions. */
3485 current_annotation
= inst
->annotation
;
3487 /* First handle scratch access on the dst. Notice we have to handle
3488 * the case where the dst's reladdr also points to scratch space.
3490 if (inst
->dst
.reladdr
)
3491 *inst
->dst
.reladdr
= emit_resolve_reladdr(scratch_loc
, block
, inst
,
3492 *inst
->dst
.reladdr
);
3494 /* Now that we have handled any (possibly recursive) reladdr scratch
3495 * accesses for dst we can safely do the scratch write for dst itself
3497 if (inst
->dst
.file
== GRF
&& scratch_loc
[inst
->dst
.reg
] != -1)
3498 emit_scratch_write(block
, inst
, scratch_loc
[inst
->dst
.reg
]);
3500 /* Now handle scratch access on any src. In this case, since inst->src[i]
3501 * already is a src_reg, we can just call emit_resolve_reladdr with
3502 * inst->src[i] and it will take care of handling scratch loads for
3503 * both src and src.reladdr (recursively).
3505 for (int i
= 0 ; i
< 3; i
++) {
3506 inst
->src
[i
] = emit_resolve_reladdr(scratch_loc
, block
, inst
,
3513 * Emits an instruction before @inst to load the value named by @orig_src
3514 * from the pull constant buffer (surface) at @base_offset to @temp.
3517 vec4_visitor::emit_pull_constant_load(bblock_t
*block
, vec4_instruction
*inst
,
3518 dst_reg temp
, src_reg orig_src
,
3521 int reg_offset
= base_offset
+ orig_src
.reg_offset
;
3522 src_reg index
= src_reg(prog_data
->base
.binding_table
.pull_constants_start
);
3523 src_reg offset
= get_pull_constant_offset(block
, inst
, orig_src
.reladdr
,
3526 emit_pull_constant_load_reg(temp
,
3533 * Implements array access of uniforms by inserting a
3534 * PULL_CONSTANT_LOAD instruction.
3536 * Unlike temporary GRF array access (where we don't support it due to
3537 * the difficulty of doing relative addressing on instruction
3538 * destinations), we could potentially do array access of uniforms
3539 * that were loaded in GRF space as push constants. In real-world
3540 * usage we've seen, though, the arrays being used are always larger
3541 * than we could load as push constants, so just always move all
3542 * uniform array access out to a pull constant buffer.
3545 vec4_visitor::move_uniform_array_access_to_pull_constants()
3547 int pull_constant_loc
[this->uniforms
];
3548 memset(pull_constant_loc
, -1, sizeof(pull_constant_loc
));
3549 bool nested_reladdr
;
3551 /* Walk through and find array access of uniforms. Put a copy of that
3552 * uniform in the pull constant buffer.
3554 * Note that we don't move constant-indexed accesses to arrays. No
3555 * testing has been done of the performance impact of this choice.
3558 nested_reladdr
= false;
3560 foreach_block_and_inst_safe(block
, vec4_instruction
, inst
, cfg
) {
3561 for (int i
= 0 ; i
< 3; i
++) {
3562 if (inst
->src
[i
].file
!= UNIFORM
|| !inst
->src
[i
].reladdr
)
3565 int uniform
= inst
->src
[i
].reg
;
3567 if (inst
->src
[i
].reladdr
->reladdr
)
3568 nested_reladdr
= true; /* will need another pass */
3570 /* If this array isn't already present in the pull constant buffer,
3573 if (pull_constant_loc
[uniform
] == -1) {
3574 const gl_constant_value
**values
=
3575 &stage_prog_data
->param
[uniform
* 4];
3577 pull_constant_loc
[uniform
] = stage_prog_data
->nr_pull_params
/ 4;
3579 assert(uniform
< uniform_array_size
);
3580 for (int j
= 0; j
< uniform_size
[uniform
] * 4; j
++) {
3581 stage_prog_data
->pull_param
[stage_prog_data
->nr_pull_params
++]
3586 /* Set up the annotation tracking for new generated instructions. */
3588 current_annotation
= inst
->annotation
;
3590 dst_reg temp
= dst_reg(this, glsl_type::vec4_type
);
3592 emit_pull_constant_load(block
, inst
, temp
, inst
->src
[i
],
3593 pull_constant_loc
[uniform
]);
3595 inst
->src
[i
].file
= temp
.file
;
3596 inst
->src
[i
].reg
= temp
.reg
;
3597 inst
->src
[i
].reg_offset
= temp
.reg_offset
;
3598 inst
->src
[i
].reladdr
= NULL
;
3601 } while (nested_reladdr
);
3603 /* Now there are no accesses of the UNIFORM file with a reladdr, so
3604 * no need to track them as larger-than-vec4 objects. This will be
3605 * relied on in cutting out unused uniform vectors from push
3608 split_uniform_registers();
3612 vec4_visitor::resolve_ud_negate(src_reg
*reg
)
3614 if (reg
->type
!= BRW_REGISTER_TYPE_UD
||
3618 src_reg temp
= src_reg(this, glsl_type::uvec4_type
);
3619 emit(BRW_OPCODE_MOV
, dst_reg(temp
), *reg
);
3624 * Resolve the result of a Gen4-5 CMP instruction to a proper boolean.
3626 * CMP on Gen4-5 only sets the LSB of the result; the rest are undefined.
3627 * If we need a proper boolean value, we have to fix it up to be 0 or ~0.
3630 vec4_visitor::resolve_bool_comparison(ir_rvalue
*rvalue
, src_reg
*reg
)
3632 assert(devinfo
->gen
<= 5);
3634 if (!rvalue
->type
->is_boolean())
3637 src_reg and_result
= src_reg(this, rvalue
->type
);
3638 src_reg neg_result
= src_reg(this, rvalue
->type
);
3639 emit(AND(dst_reg(and_result
), *reg
, src_reg(1)));
3640 emit(MOV(dst_reg(neg_result
), negate(and_result
)));
3644 vec4_visitor::vec4_visitor(struct brw_context
*brw
,
3645 struct brw_vec4_compile
*c
,
3646 struct gl_program
*prog
,
3647 const struct brw_vue_prog_key
*key
,
3648 struct brw_vue_prog_data
*prog_data
,
3649 struct gl_shader_program
*shader_prog
,
3650 gl_shader_stage stage
,
3653 shader_time_shader_type st_base
,
3654 shader_time_shader_type st_written
,
3655 shader_time_shader_type st_reset
)
3656 : backend_visitor(brw
, shader_prog
, prog
, &prog_data
->base
, stage
),
3659 prog_data(prog_data
),
3660 sanity_param_count(0),
3662 first_non_payload_grf(0),
3663 need_all_constants_in_pull_buffer(false),
3664 no_spills(no_spills
),
3666 st_written(st_written
),
3669 this->mem_ctx
= mem_ctx
;
3670 this->failed
= false;
3672 this->base_ir
= NULL
;
3673 this->current_annotation
= NULL
;
3674 memset(this->output_reg_annotation
, 0, sizeof(this->output_reg_annotation
));
3676 this->variable_ht
= hash_table_ctor(0,
3677 hash_table_pointer_hash
,
3678 hash_table_pointer_compare
);
3680 this->virtual_grf_start
= NULL
;
3681 this->virtual_grf_end
= NULL
;
3682 this->live_intervals
= NULL
;
3684 this->max_grf
= devinfo
->gen
>= 7 ? GEN7_MRF_HACK_START
: BRW_MAX_GRF
;
3688 /* Initialize uniform_array_size to at least 1 because pre-gen6 VS requires
3689 * at least one. See setup_uniforms() in brw_vec4.cpp.
3691 this->uniform_array_size
= 1;
3693 this->uniform_array_size
=
3694 MAX2(DIV_ROUND_UP(stage_prog_data
->nr_params
, 4), 1);
3697 this->uniform_size
= rzalloc_array(mem_ctx
, int, this->uniform_array_size
);
3698 this->uniform_vector_size
= rzalloc_array(mem_ctx
, int, this->uniform_array_size
);
3701 vec4_visitor::~vec4_visitor()
3703 hash_table_dtor(this->variable_ht
);
3708 vec4_visitor::fail(const char *format
, ...)
3718 va_start(va
, format
);
3719 msg
= ralloc_vasprintf(mem_ctx
, format
, va
);
3721 msg
= ralloc_asprintf(mem_ctx
, "%s compile failed: %s\n", stage_abbrev
, msg
);
3723 this->fail_msg
= msg
;
3725 if (debug_enabled
) {
3726 fprintf(stderr
, "%s", msg
);
3730 } /* namespace brw */