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::emit_uniformize(const dst_reg
&dst
, const src_reg
&src
)
1379 const src_reg
chan_index(this, glsl_type::uint_type
);
1381 emit(SHADER_OPCODE_FIND_LIVE_CHANNEL
, dst_reg(chan_index
))
1382 ->force_writemask_all
= true;
1383 emit(SHADER_OPCODE_BROADCAST
, dst
, src
, chan_index
)
1384 ->force_writemask_all
= true;
1388 vec4_visitor::visit(ir_expression
*ir
)
1390 unsigned int operand
;
1391 src_reg op
[ARRAY_SIZE(ir
->operands
)];
1392 vec4_instruction
*inst
;
1394 if (ir
->operation
== ir_binop_add
) {
1395 if (try_emit_mad(ir
))
1399 if (ir
->operation
== ir_unop_b2f
) {
1400 if (try_emit_b2f_of_compare(ir
))
1404 /* Storage for our result. Ideally for an assignment we'd be using
1405 * the actual storage for the result here, instead.
1407 dst_reg
result_dst(this, ir
->type
);
1408 src_reg
result_src(result_dst
);
1410 if (ir
->operation
== ir_triop_csel
) {
1411 ir
->operands
[1]->accept(this);
1412 op
[1] = this->result
;
1413 ir
->operands
[2]->accept(this);
1414 op
[2] = this->result
;
1416 enum brw_predicate predicate
;
1417 emit_bool_to_cond_code(ir
->operands
[0], &predicate
);
1418 inst
= emit(BRW_OPCODE_SEL
, result_dst
, op
[1], op
[2]);
1419 inst
->predicate
= predicate
;
1420 this->result
= result_src
;
1424 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1425 this->result
.file
= BAD_FILE
;
1426 ir
->operands
[operand
]->accept(this);
1427 if (this->result
.file
== BAD_FILE
) {
1428 fprintf(stderr
, "Failed to get tree for expression operand:\n");
1429 ir
->operands
[operand
]->fprint(stderr
);
1432 op
[operand
] = this->result
;
1434 /* Matrix expression operands should have been broken down to vector
1435 * operations already.
1437 assert(!ir
->operands
[operand
]->type
->is_matrix());
1440 /* If nothing special happens, this is the result. */
1441 this->result
= result_src
;
1443 switch (ir
->operation
) {
1444 case ir_unop_logic_not
:
1445 emit(NOT(result_dst
, op
[0]));
1448 op
[0].negate
= !op
[0].negate
;
1449 emit(MOV(result_dst
, op
[0]));
1453 op
[0].negate
= false;
1454 emit(MOV(result_dst
, op
[0]));
1458 if (ir
->type
->is_float()) {
1459 /* AND(val, 0x80000000) gives the sign bit.
1461 * Predicated OR ORs 1.0 (0x3f800000) with the sign bit if val is not
1464 emit(CMP(dst_null_f(), op
[0], src_reg(0.0f
), BRW_CONDITIONAL_NZ
));
1466 op
[0].type
= BRW_REGISTER_TYPE_UD
;
1467 result_dst
.type
= BRW_REGISTER_TYPE_UD
;
1468 emit(AND(result_dst
, op
[0], src_reg(0x80000000u
)));
1470 inst
= emit(OR(result_dst
, src_reg(result_dst
), src_reg(0x3f800000u
)));
1471 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1473 this->result
.type
= BRW_REGISTER_TYPE_F
;
1475 /* ASR(val, 31) -> negative val generates 0xffffffff (signed -1).
1476 * -> non-negative val generates 0x00000000.
1477 * Predicated OR sets 1 if val is positive.
1479 emit(CMP(dst_null_d(), op
[0], src_reg(0), BRW_CONDITIONAL_G
));
1481 emit(ASR(result_dst
, op
[0], src_reg(31)));
1483 inst
= emit(OR(result_dst
, src_reg(result_dst
), src_reg(1)));
1484 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1489 emit_math(SHADER_OPCODE_RCP
, result_dst
, op
[0]);
1493 emit_math(SHADER_OPCODE_EXP2
, result_dst
, op
[0]);
1496 emit_math(SHADER_OPCODE_LOG2
, result_dst
, op
[0]);
1500 unreachable("not reached: should be handled by ir_explog_to_explog2");
1502 emit_math(SHADER_OPCODE_SIN
, result_dst
, op
[0]);
1505 emit_math(SHADER_OPCODE_COS
, result_dst
, op
[0]);
1509 case ir_unop_dFdx_coarse
:
1510 case ir_unop_dFdx_fine
:
1512 case ir_unop_dFdy_coarse
:
1513 case ir_unop_dFdy_fine
:
1514 unreachable("derivatives not valid in vertex shader");
1516 case ir_unop_bitfield_reverse
:
1517 emit(BFREV(result_dst
, op
[0]));
1519 case ir_unop_bit_count
:
1520 emit(CBIT(result_dst
, op
[0]));
1522 case ir_unop_find_msb
: {
1523 src_reg temp
= src_reg(this, glsl_type::uint_type
);
1525 inst
= emit(FBH(dst_reg(temp
), op
[0]));
1526 inst
->dst
.writemask
= WRITEMASK_XYZW
;
1528 /* FBH counts from the MSB side, while GLSL's findMSB() wants the count
1529 * from the LSB side. If FBH didn't return an error (0xFFFFFFFF), then
1530 * subtract the result from 31 to convert the MSB count into an LSB count.
1533 /* FBH only supports UD type for dst, so use a MOV to convert UD to D. */
1534 temp
.swizzle
= BRW_SWIZZLE_NOOP
;
1535 emit(MOV(result_dst
, temp
));
1537 src_reg src_tmp
= src_reg(result_dst
);
1538 emit(CMP(dst_null_d(), src_tmp
, src_reg(-1), BRW_CONDITIONAL_NZ
));
1540 src_tmp
.negate
= true;
1541 inst
= emit(ADD(result_dst
, src_tmp
, src_reg(31)));
1542 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1545 case ir_unop_find_lsb
:
1546 emit(FBL(result_dst
, op
[0]));
1548 case ir_unop_saturate
:
1549 inst
= emit(MOV(result_dst
, op
[0]));
1550 inst
->saturate
= true;
1554 unreachable("not reached: should be handled by lower_noise");
1557 emit(ADD(result_dst
, op
[0], op
[1]));
1560 unreachable("not reached: should be handled by ir_sub_to_add_neg");
1563 if (devinfo
->gen
< 8 && ir
->type
->is_integer()) {
1564 /* For integer multiplication, the MUL uses the low 16 bits of one of
1565 * the operands (src0 through SNB, src1 on IVB and later). The MACH
1566 * accumulates in the contribution of the upper 16 bits of that
1567 * operand. If we can determine that one of the args is in the low
1568 * 16 bits, though, we can just emit a single MUL.
1570 if (ir
->operands
[0]->is_uint16_constant()) {
1571 if (devinfo
->gen
< 7)
1572 emit(MUL(result_dst
, op
[0], op
[1]));
1574 emit(MUL(result_dst
, op
[1], op
[0]));
1575 } else if (ir
->operands
[1]->is_uint16_constant()) {
1576 if (devinfo
->gen
< 7)
1577 emit(MUL(result_dst
, op
[1], op
[0]));
1579 emit(MUL(result_dst
, op
[0], op
[1]));
1581 struct brw_reg acc
= retype(brw_acc_reg(8), result_dst
.type
);
1583 emit(MUL(acc
, op
[0], op
[1]));
1584 emit(MACH(dst_null_d(), op
[0], op
[1]));
1585 emit(MOV(result_dst
, src_reg(acc
)));
1588 emit(MUL(result_dst
, op
[0], op
[1]));
1591 case ir_binop_imul_high
: {
1592 struct brw_reg acc
= retype(brw_acc_reg(8), result_dst
.type
);
1594 emit(MUL(acc
, op
[0], op
[1]));
1595 emit(MACH(result_dst
, op
[0], op
[1]));
1599 /* Floating point should be lowered by DIV_TO_MUL_RCP in the compiler. */
1600 assert(ir
->type
->is_integer());
1601 emit_math(SHADER_OPCODE_INT_QUOTIENT
, result_dst
, op
[0], op
[1]);
1603 case ir_binop_carry
: {
1604 struct brw_reg acc
= retype(brw_acc_reg(8), BRW_REGISTER_TYPE_UD
);
1606 emit(ADDC(dst_null_ud(), op
[0], op
[1]));
1607 emit(MOV(result_dst
, src_reg(acc
)));
1610 case ir_binop_borrow
: {
1611 struct brw_reg acc
= retype(brw_acc_reg(8), BRW_REGISTER_TYPE_UD
);
1613 emit(SUBB(dst_null_ud(), op
[0], op
[1]));
1614 emit(MOV(result_dst
, src_reg(acc
)));
1618 /* Floating point should be lowered by MOD_TO_FLOOR in the compiler. */
1619 assert(ir
->type
->is_integer());
1620 emit_math(SHADER_OPCODE_INT_REMAINDER
, result_dst
, op
[0], op
[1]);
1624 case ir_binop_greater
:
1625 case ir_binop_lequal
:
1626 case ir_binop_gequal
:
1627 case ir_binop_equal
:
1628 case ir_binop_nequal
: {
1629 if (devinfo
->gen
<= 5) {
1630 resolve_bool_comparison(ir
->operands
[0], &op
[0]);
1631 resolve_bool_comparison(ir
->operands
[1], &op
[1]);
1633 emit(CMP(result_dst
, op
[0], op
[1],
1634 brw_conditional_for_comparison(ir
->operation
)));
1638 case ir_binop_all_equal
:
1639 if (devinfo
->gen
<= 5) {
1640 resolve_bool_comparison(ir
->operands
[0], &op
[0]);
1641 resolve_bool_comparison(ir
->operands
[1], &op
[1]);
1644 /* "==" operator producing a scalar boolean. */
1645 if (ir
->operands
[0]->type
->is_vector() ||
1646 ir
->operands
[1]->type
->is_vector()) {
1647 emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_Z
));
1648 emit(MOV(result_dst
, src_reg(0)));
1649 inst
= emit(MOV(result_dst
, src_reg(~0)));
1650 inst
->predicate
= BRW_PREDICATE_ALIGN16_ALL4H
;
1652 emit(CMP(result_dst
, op
[0], op
[1], BRW_CONDITIONAL_Z
));
1655 case ir_binop_any_nequal
:
1656 if (devinfo
->gen
<= 5) {
1657 resolve_bool_comparison(ir
->operands
[0], &op
[0]);
1658 resolve_bool_comparison(ir
->operands
[1], &op
[1]);
1661 /* "!=" operator producing a scalar boolean. */
1662 if (ir
->operands
[0]->type
->is_vector() ||
1663 ir
->operands
[1]->type
->is_vector()) {
1664 emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_NZ
));
1666 emit(MOV(result_dst
, src_reg(0)));
1667 inst
= emit(MOV(result_dst
, src_reg(~0)));
1668 inst
->predicate
= BRW_PREDICATE_ALIGN16_ANY4H
;
1670 emit(CMP(result_dst
, op
[0], op
[1], BRW_CONDITIONAL_NZ
));
1675 if (devinfo
->gen
<= 5) {
1676 resolve_bool_comparison(ir
->operands
[0], &op
[0]);
1678 emit(CMP(dst_null_d(), op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
1679 emit(MOV(result_dst
, src_reg(0)));
1681 inst
= emit(MOV(result_dst
, src_reg(~0)));
1682 inst
->predicate
= BRW_PREDICATE_ALIGN16_ANY4H
;
1685 case ir_binop_logic_xor
:
1686 emit(XOR(result_dst
, op
[0], op
[1]));
1689 case ir_binop_logic_or
:
1690 emit(OR(result_dst
, op
[0], op
[1]));
1693 case ir_binop_logic_and
:
1694 emit(AND(result_dst
, op
[0], op
[1]));
1698 assert(ir
->operands
[0]->type
->is_vector());
1699 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1700 emit_dp(result_dst
, op
[0], op
[1], ir
->operands
[0]->type
->vector_elements
);
1704 emit_math(SHADER_OPCODE_SQRT
, result_dst
, op
[0]);
1707 emit_math(SHADER_OPCODE_RSQ
, result_dst
, op
[0]);
1710 case ir_unop_bitcast_i2f
:
1711 case ir_unop_bitcast_u2f
:
1712 this->result
= op
[0];
1713 this->result
.type
= BRW_REGISTER_TYPE_F
;
1716 case ir_unop_bitcast_f2i
:
1717 this->result
= op
[0];
1718 this->result
.type
= BRW_REGISTER_TYPE_D
;
1721 case ir_unop_bitcast_f2u
:
1722 this->result
= op
[0];
1723 this->result
.type
= BRW_REGISTER_TYPE_UD
;
1732 emit(MOV(result_dst
, op
[0]));
1735 emit(AND(result_dst
, op
[0], src_reg(1)));
1738 if (devinfo
->gen
<= 5) {
1739 resolve_bool_comparison(ir
->operands
[0], &op
[0]);
1741 op
[0].type
= BRW_REGISTER_TYPE_D
;
1742 result_dst
.type
= BRW_REGISTER_TYPE_D
;
1743 emit(AND(result_dst
, op
[0], src_reg(0x3f800000u
)));
1744 result_dst
.type
= BRW_REGISTER_TYPE_F
;
1747 emit(CMP(result_dst
, op
[0], src_reg(0.0f
), BRW_CONDITIONAL_NZ
));
1750 emit(CMP(result_dst
, op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
1754 emit(RNDZ(result_dst
, op
[0]));
1756 case ir_unop_ceil
: {
1757 src_reg tmp
= src_reg(this, ir
->type
);
1758 op
[0].negate
= !op
[0].negate
;
1759 emit(RNDD(dst_reg(tmp
), op
[0]));
1761 emit(MOV(result_dst
, tmp
));
1765 inst
= emit(RNDD(result_dst
, op
[0]));
1768 inst
= emit(FRC(result_dst
, op
[0]));
1770 case ir_unop_round_even
:
1771 emit(RNDE(result_dst
, op
[0]));
1775 emit_minmax(BRW_CONDITIONAL_L
, result_dst
, op
[0], op
[1]);
1778 emit_minmax(BRW_CONDITIONAL_GE
, result_dst
, op
[0], op
[1]);
1782 emit_math(SHADER_OPCODE_POW
, result_dst
, op
[0], op
[1]);
1785 case ir_unop_bit_not
:
1786 inst
= emit(NOT(result_dst
, op
[0]));
1788 case ir_binop_bit_and
:
1789 inst
= emit(AND(result_dst
, op
[0], op
[1]));
1791 case ir_binop_bit_xor
:
1792 inst
= emit(XOR(result_dst
, op
[0], op
[1]));
1794 case ir_binop_bit_or
:
1795 inst
= emit(OR(result_dst
, op
[0], op
[1]));
1798 case ir_binop_lshift
:
1799 inst
= emit(SHL(result_dst
, op
[0], op
[1]));
1802 case ir_binop_rshift
:
1803 if (ir
->type
->base_type
== GLSL_TYPE_INT
)
1804 inst
= emit(ASR(result_dst
, op
[0], op
[1]));
1806 inst
= emit(SHR(result_dst
, op
[0], op
[1]));
1810 emit(BFI1(result_dst
, op
[0], op
[1]));
1813 case ir_binop_ubo_load
: {
1814 ir_constant
*const_uniform_block
= ir
->operands
[0]->as_constant();
1815 ir_constant
*const_offset_ir
= ir
->operands
[1]->as_constant();
1816 unsigned const_offset
= const_offset_ir
? const_offset_ir
->value
.u
[0] : 0;
1819 /* Now, load the vector from that offset. */
1820 assert(ir
->type
->is_vector() || ir
->type
->is_scalar());
1822 src_reg packed_consts
= src_reg(this, glsl_type::vec4_type
);
1823 packed_consts
.type
= result
.type
;
1826 if (const_uniform_block
) {
1827 /* The block index is a constant, so just emit the binding table entry
1830 surf_index
= src_reg(prog_data
->base
.binding_table
.ubo_start
+
1831 const_uniform_block
->value
.u
[0]);
1833 /* The block index is not a constant. Evaluate the index expression
1834 * per-channel and add the base UBO index; we have to select a value
1835 * from any live channel.
1837 surf_index
= src_reg(this, glsl_type::uint_type
);
1838 emit(ADD(dst_reg(surf_index
), op
[0],
1839 src_reg(prog_data
->base
.binding_table
.ubo_start
)));
1840 emit_uniformize(dst_reg(surf_index
), surf_index
);
1842 /* Assume this may touch any UBO. It would be nice to provide
1843 * a tighter bound, but the array information is already lowered away.
1845 brw_mark_surface_used(&prog_data
->base
,
1846 prog_data
->base
.binding_table
.ubo_start
+
1847 shader_prog
->NumUniformBlocks
- 1);
1850 if (const_offset_ir
) {
1851 if (devinfo
->gen
>= 8) {
1852 /* Store the offset in a GRF so we can send-from-GRF. */
1853 offset
= src_reg(this, glsl_type::int_type
);
1854 emit(MOV(dst_reg(offset
), src_reg(const_offset
/ 16)));
1856 /* Immediates are fine on older generations since they'll be moved
1857 * to a (potentially fake) MRF at the generator level.
1859 offset
= src_reg(const_offset
/ 16);
1862 offset
= src_reg(this, glsl_type::uint_type
);
1863 emit(SHR(dst_reg(offset
), op
[1], src_reg(4)));
1866 emit_pull_constant_load_reg(dst_reg(packed_consts
),
1869 NULL
, NULL
/* before_block/inst */);
1871 packed_consts
.swizzle
= brw_swizzle_for_size(ir
->type
->vector_elements
);
1872 packed_consts
.swizzle
+= BRW_SWIZZLE4(const_offset
% 16 / 4,
1873 const_offset
% 16 / 4,
1874 const_offset
% 16 / 4,
1875 const_offset
% 16 / 4);
1877 /* UBO bools are any nonzero int. We need to convert them to 0/~0. */
1878 if (ir
->type
->base_type
== GLSL_TYPE_BOOL
) {
1879 emit(CMP(result_dst
, packed_consts
, src_reg(0u),
1880 BRW_CONDITIONAL_NZ
));
1882 emit(MOV(result_dst
, packed_consts
));
1887 case ir_binop_vector_extract
:
1888 unreachable("should have been lowered by vec_index_to_cond_assign");
1891 op
[0] = fix_3src_operand(op
[0]);
1892 op
[1] = fix_3src_operand(op
[1]);
1893 op
[2] = fix_3src_operand(op
[2]);
1894 /* Note that the instruction's argument order is reversed from GLSL
1897 emit(MAD(result_dst
, op
[2], op
[1], op
[0]));
1901 emit_lrp(result_dst
, op
[0], op
[1], op
[2]);
1905 unreachable("already handled above");
1909 op
[0] = fix_3src_operand(op
[0]);
1910 op
[1] = fix_3src_operand(op
[1]);
1911 op
[2] = fix_3src_operand(op
[2]);
1912 emit(BFI2(result_dst
, op
[0], op
[1], op
[2]));
1915 case ir_triop_bitfield_extract
:
1916 op
[0] = fix_3src_operand(op
[0]);
1917 op
[1] = fix_3src_operand(op
[1]);
1918 op
[2] = fix_3src_operand(op
[2]);
1919 /* Note that the instruction's argument order is reversed from GLSL
1922 emit(BFE(result_dst
, op
[2], op
[1], op
[0]));
1925 case ir_triop_vector_insert
:
1926 unreachable("should have been lowered by lower_vector_insert");
1928 case ir_quadop_bitfield_insert
:
1929 unreachable("not reached: should be handled by "
1930 "bitfield_insert_to_bfm_bfi\n");
1932 case ir_quadop_vector
:
1933 unreachable("not reached: should be handled by lower_quadop_vector");
1935 case ir_unop_pack_half_2x16
:
1936 emit_pack_half_2x16(result_dst
, op
[0]);
1938 case ir_unop_unpack_half_2x16
:
1939 emit_unpack_half_2x16(result_dst
, op
[0]);
1941 case ir_unop_unpack_unorm_4x8
:
1942 emit_unpack_unorm_4x8(result_dst
, op
[0]);
1944 case ir_unop_unpack_snorm_4x8
:
1945 emit_unpack_snorm_4x8(result_dst
, op
[0]);
1947 case ir_unop_pack_unorm_4x8
:
1948 emit_pack_unorm_4x8(result_dst
, op
[0]);
1950 case ir_unop_pack_snorm_4x8
:
1951 emit_pack_snorm_4x8(result_dst
, op
[0]);
1953 case ir_unop_pack_snorm_2x16
:
1954 case ir_unop_pack_unorm_2x16
:
1955 case ir_unop_unpack_snorm_2x16
:
1956 case ir_unop_unpack_unorm_2x16
:
1957 unreachable("not reached: should be handled by lower_packing_builtins");
1958 case ir_unop_unpack_half_2x16_split_x
:
1959 case ir_unop_unpack_half_2x16_split_y
:
1960 case ir_binop_pack_half_2x16_split
:
1961 case ir_unop_interpolate_at_centroid
:
1962 case ir_binop_interpolate_at_sample
:
1963 case ir_binop_interpolate_at_offset
:
1964 unreachable("not reached: should not occur in vertex shader");
1965 case ir_binop_ldexp
:
1966 unreachable("not reached: should be handled by ldexp_to_arith()");
1974 case ir_unop_pack_double_2x32
:
1975 case ir_unop_unpack_double_2x32
:
1976 case ir_unop_frexp_sig
:
1977 case ir_unop_frexp_exp
:
1978 unreachable("fp64 todo");
1984 vec4_visitor::visit(ir_swizzle
*ir
)
1986 /* Note that this is only swizzles in expressions, not those on the left
1987 * hand side of an assignment, which do write masking. See ir_assignment
1990 const unsigned swz
= brw_compose_swizzle(
1991 brw_swizzle_for_size(ir
->type
->vector_elements
),
1992 BRW_SWIZZLE4(ir
->mask
.x
, ir
->mask
.y
, ir
->mask
.z
, ir
->mask
.w
));
1994 ir
->val
->accept(this);
1995 this->result
= swizzle(this->result
, swz
);
1999 vec4_visitor::visit(ir_dereference_variable
*ir
)
2001 const struct glsl_type
*type
= ir
->type
;
2002 dst_reg
*reg
= variable_storage(ir
->var
);
2005 fail("Failed to find variable storage for %s\n", ir
->var
->name
);
2006 this->result
= src_reg(brw_null_reg());
2010 this->result
= src_reg(*reg
);
2012 /* System values get their swizzle from the dst_reg writemask */
2013 if (ir
->var
->data
.mode
== ir_var_system_value
)
2016 if (type
->is_scalar() || type
->is_vector() || type
->is_matrix())
2017 this->result
.swizzle
= brw_swizzle_for_size(type
->vector_elements
);
2022 vec4_visitor::compute_array_stride(ir_dereference_array
*ir
)
2024 /* Under normal circumstances array elements are stored consecutively, so
2025 * the stride is equal to the size of the array element.
2027 return type_size(ir
->type
);
2032 vec4_visitor::visit(ir_dereference_array
*ir
)
2034 ir_constant
*constant_index
;
2036 int array_stride
= compute_array_stride(ir
);
2038 constant_index
= ir
->array_index
->constant_expression_value();
2040 ir
->array
->accept(this);
2043 if (constant_index
) {
2044 src
.reg_offset
+= constant_index
->value
.i
[0] * array_stride
;
2046 /* Variable index array dereference. It eats the "vec4" of the
2047 * base of the array and an index that offsets the Mesa register
2050 ir
->array_index
->accept(this);
2054 if (array_stride
== 1) {
2055 index_reg
= this->result
;
2057 index_reg
= src_reg(this, glsl_type::int_type
);
2059 emit(MUL(dst_reg(index_reg
), this->result
, src_reg(array_stride
)));
2063 src_reg temp
= src_reg(this, glsl_type::int_type
);
2065 emit(ADD(dst_reg(temp
), *src
.reladdr
, index_reg
));
2070 src
.reladdr
= ralloc(mem_ctx
, src_reg
);
2071 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
2074 /* If the type is smaller than a vec4, replicate the last channel out. */
2075 if (ir
->type
->is_scalar() || ir
->type
->is_vector() || ir
->type
->is_matrix())
2076 src
.swizzle
= brw_swizzle_for_size(ir
->type
->vector_elements
);
2078 src
.swizzle
= BRW_SWIZZLE_NOOP
;
2079 src
.type
= brw_type_for_base_type(ir
->type
);
2085 vec4_visitor::visit(ir_dereference_record
*ir
)
2088 const glsl_type
*struct_type
= ir
->record
->type
;
2091 ir
->record
->accept(this);
2093 for (i
= 0; i
< struct_type
->length
; i
++) {
2094 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
2096 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
2099 /* If the type is smaller than a vec4, replicate the last channel out. */
2100 if (ir
->type
->is_scalar() || ir
->type
->is_vector() || ir
->type
->is_matrix())
2101 this->result
.swizzle
= brw_swizzle_for_size(ir
->type
->vector_elements
);
2103 this->result
.swizzle
= BRW_SWIZZLE_NOOP
;
2104 this->result
.type
= brw_type_for_base_type(ir
->type
);
2106 this->result
.reg_offset
+= offset
;
2110 * We want to be careful in assignment setup to hit the actual storage
2111 * instead of potentially using a temporary like we might with the
2112 * ir_dereference handler.
2115 get_assignment_lhs(ir_dereference
*ir
, vec4_visitor
*v
)
2117 /* The LHS must be a dereference. If the LHS is a variable indexed array
2118 * access of a vector, it must be separated into a series conditional moves
2119 * before reaching this point (see ir_vec_index_to_cond_assign).
2121 assert(ir
->as_dereference());
2122 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
2124 assert(!deref_array
->array
->type
->is_vector());
2127 /* Use the rvalue deref handler for the most part. We'll ignore
2128 * swizzles in it and write swizzles using writemask, though.
2131 return dst_reg(v
->result
);
2135 vec4_visitor::emit_block_move(dst_reg
*dst
, src_reg
*src
,
2136 const struct glsl_type
*type
,
2137 enum brw_predicate predicate
)
2139 if (type
->base_type
== GLSL_TYPE_STRUCT
) {
2140 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2141 emit_block_move(dst
, src
, type
->fields
.structure
[i
].type
, predicate
);
2146 if (type
->is_array()) {
2147 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2148 emit_block_move(dst
, src
, type
->fields
.array
, predicate
);
2153 if (type
->is_matrix()) {
2154 const struct glsl_type
*vec_type
;
2156 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
2157 type
->vector_elements
, 1);
2159 for (int i
= 0; i
< type
->matrix_columns
; i
++) {
2160 emit_block_move(dst
, src
, vec_type
, predicate
);
2165 assert(type
->is_scalar() || type
->is_vector());
2167 dst
->type
= brw_type_for_base_type(type
);
2168 src
->type
= dst
->type
;
2170 dst
->writemask
= (1 << type
->vector_elements
) - 1;
2172 src
->swizzle
= brw_swizzle_for_size(type
->vector_elements
);
2174 vec4_instruction
*inst
= emit(MOV(*dst
, *src
));
2175 inst
->predicate
= predicate
;
2182 /* If the RHS processing resulted in an instruction generating a
2183 * temporary value, and it would be easy to rewrite the instruction to
2184 * generate its result right into the LHS instead, do so. This ends
2185 * up reliably removing instructions where it can be tricky to do so
2186 * later without real UD chain information.
2189 vec4_visitor::try_rewrite_rhs_to_dst(ir_assignment
*ir
,
2192 vec4_instruction
*pre_rhs_inst
,
2193 vec4_instruction
*last_rhs_inst
)
2195 /* This could be supported, but it would take more smarts. */
2199 if (pre_rhs_inst
== last_rhs_inst
)
2200 return false; /* No instructions generated to work with. */
2202 /* Make sure the last instruction generated our source reg. */
2203 if (src
.file
!= GRF
||
2204 src
.file
!= last_rhs_inst
->dst
.file
||
2205 src
.reg
!= last_rhs_inst
->dst
.reg
||
2206 src
.reg_offset
!= last_rhs_inst
->dst
.reg_offset
||
2210 last_rhs_inst
->predicate
!= BRW_PREDICATE_NONE
)
2213 /* Check that that last instruction fully initialized the channels
2214 * we want to use, in the order we want to use them. We could
2215 * potentially reswizzle the operands of many instructions so that
2216 * we could handle out of order channels, but don't yet.
2219 for (unsigned i
= 0; i
< 4; i
++) {
2220 if (dst
.writemask
& (1 << i
)) {
2221 if (!(last_rhs_inst
->dst
.writemask
& (1 << i
)))
2224 if (BRW_GET_SWZ(src
.swizzle
, i
) != i
)
2229 /* Success! Rewrite the instruction. */
2230 last_rhs_inst
->dst
.file
= dst
.file
;
2231 last_rhs_inst
->dst
.reg
= dst
.reg
;
2232 last_rhs_inst
->dst
.reg_offset
= dst
.reg_offset
;
2233 last_rhs_inst
->dst
.reladdr
= dst
.reladdr
;
2234 last_rhs_inst
->dst
.writemask
&= dst
.writemask
;
2240 vec4_visitor::visit(ir_assignment
*ir
)
2242 dst_reg dst
= get_assignment_lhs(ir
->lhs
, this);
2243 enum brw_predicate predicate
= BRW_PREDICATE_NONE
;
2245 if (!ir
->lhs
->type
->is_scalar() &&
2246 !ir
->lhs
->type
->is_vector()) {
2247 ir
->rhs
->accept(this);
2248 src_reg src
= this->result
;
2250 if (ir
->condition
) {
2251 emit_bool_to_cond_code(ir
->condition
, &predicate
);
2254 /* emit_block_move doesn't account for swizzles in the source register.
2255 * This should be ok, since the source register is a structure or an
2256 * array, and those can't be swizzled. But double-check to be sure.
2258 assert(src
.swizzle
==
2259 (ir
->rhs
->type
->is_matrix()
2260 ? brw_swizzle_for_size(ir
->rhs
->type
->vector_elements
)
2261 : BRW_SWIZZLE_NOOP
));
2263 emit_block_move(&dst
, &src
, ir
->rhs
->type
, predicate
);
2267 /* Now we're down to just a scalar/vector with writemasks. */
2270 vec4_instruction
*pre_rhs_inst
, *last_rhs_inst
;
2271 pre_rhs_inst
= (vec4_instruction
*)this->instructions
.get_tail();
2273 ir
->rhs
->accept(this);
2275 last_rhs_inst
= (vec4_instruction
*)this->instructions
.get_tail();
2280 assert(ir
->lhs
->type
->is_vector() ||
2281 ir
->lhs
->type
->is_scalar());
2282 dst
.writemask
= ir
->write_mask
;
2284 /* Swizzle a small RHS vector into the channels being written.
2286 * glsl ir treats write_mask as dictating how many channels are
2287 * present on the RHS while in our instructions we need to make
2288 * those channels appear in the slots of the vec4 they're written to.
2290 for (int i
= 0; i
< 4; i
++)
2291 swizzles
[i
] = (ir
->write_mask
& (1 << i
) ? src_chan
++ : 0);
2293 src_reg src
= swizzle(this->result
,
2294 BRW_SWIZZLE4(swizzles
[0], swizzles
[1],
2295 swizzles
[2], swizzles
[3]));
2297 if (try_rewrite_rhs_to_dst(ir
, dst
, src
, pre_rhs_inst
, last_rhs_inst
)) {
2301 if (ir
->condition
) {
2302 emit_bool_to_cond_code(ir
->condition
, &predicate
);
2305 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2306 vec4_instruction
*inst
= emit(MOV(dst
, src
));
2307 inst
->predicate
= predicate
;
2315 vec4_visitor::emit_constant_values(dst_reg
*dst
, ir_constant
*ir
)
2317 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
2318 foreach_in_list(ir_constant
, field_value
, &ir
->components
) {
2319 emit_constant_values(dst
, field_value
);
2324 if (ir
->type
->is_array()) {
2325 for (unsigned int i
= 0; i
< ir
->type
->length
; i
++) {
2326 emit_constant_values(dst
, ir
->array_elements
[i
]);
2331 if (ir
->type
->is_matrix()) {
2332 for (int i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
2333 float *vec
= &ir
->value
.f
[i
* ir
->type
->vector_elements
];
2335 for (int j
= 0; j
< ir
->type
->vector_elements
; j
++) {
2336 dst
->writemask
= 1 << j
;
2337 dst
->type
= BRW_REGISTER_TYPE_F
;
2339 emit(MOV(*dst
, src_reg(vec
[j
])));
2346 int remaining_writemask
= (1 << ir
->type
->vector_elements
) - 1;
2348 for (int i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2349 if (!(remaining_writemask
& (1 << i
)))
2352 dst
->writemask
= 1 << i
;
2353 dst
->type
= brw_type_for_base_type(ir
->type
);
2355 /* Find other components that match the one we're about to
2356 * write. Emits fewer instructions for things like vec4(0.5,
2359 for (int j
= i
+ 1; j
< ir
->type
->vector_elements
; j
++) {
2360 if (ir
->type
->base_type
== GLSL_TYPE_BOOL
) {
2361 if (ir
->value
.b
[i
] == ir
->value
.b
[j
])
2362 dst
->writemask
|= (1 << j
);
2364 /* u, i, and f storage all line up, so no need for a
2365 * switch case for comparing each type.
2367 if (ir
->value
.u
[i
] == ir
->value
.u
[j
])
2368 dst
->writemask
|= (1 << j
);
2372 switch (ir
->type
->base_type
) {
2373 case GLSL_TYPE_FLOAT
:
2374 emit(MOV(*dst
, src_reg(ir
->value
.f
[i
])));
2377 emit(MOV(*dst
, src_reg(ir
->value
.i
[i
])));
2379 case GLSL_TYPE_UINT
:
2380 emit(MOV(*dst
, src_reg(ir
->value
.u
[i
])));
2382 case GLSL_TYPE_BOOL
:
2383 emit(MOV(*dst
, src_reg(ir
->value
.b
[i
] != 0 ? ~0 : 0)));
2386 unreachable("Non-float/uint/int/bool constant");
2389 remaining_writemask
&= ~dst
->writemask
;
2395 vec4_visitor::visit(ir_constant
*ir
)
2397 dst_reg dst
= dst_reg(this, ir
->type
);
2398 this->result
= src_reg(dst
);
2400 emit_constant_values(&dst
, ir
);
2404 vec4_visitor::visit_atomic_counter_intrinsic(ir_call
*ir
)
2406 ir_dereference
*deref
= static_cast<ir_dereference
*>(
2407 ir
->actual_parameters
.get_head());
2408 ir_variable
*location
= deref
->variable_referenced();
2409 unsigned surf_index
= (prog_data
->base
.binding_table
.abo_start
+
2410 location
->data
.binding
);
2412 /* Calculate the surface offset */
2413 src_reg
offset(this, glsl_type::uint_type
);
2414 ir_dereference_array
*deref_array
= deref
->as_dereference_array();
2416 deref_array
->array_index
->accept(this);
2418 src_reg
tmp(this, glsl_type::uint_type
);
2419 emit(MUL(dst_reg(tmp
), this->result
, ATOMIC_COUNTER_SIZE
));
2420 emit(ADD(dst_reg(offset
), tmp
, location
->data
.atomic
.offset
));
2422 offset
= location
->data
.atomic
.offset
;
2425 /* Emit the appropriate machine instruction */
2426 const char *callee
= ir
->callee
->function_name();
2427 dst_reg dst
= get_assignment_lhs(ir
->return_deref
, this);
2429 if (!strcmp("__intrinsic_atomic_read", callee
)) {
2430 emit_untyped_surface_read(surf_index
, dst
, offset
);
2432 } else if (!strcmp("__intrinsic_atomic_increment", callee
)) {
2433 emit_untyped_atomic(BRW_AOP_INC
, surf_index
, dst
, offset
,
2434 src_reg(), src_reg());
2436 } else if (!strcmp("__intrinsic_atomic_predecrement", callee
)) {
2437 emit_untyped_atomic(BRW_AOP_PREDEC
, surf_index
, dst
, offset
,
2438 src_reg(), src_reg());
2443 vec4_visitor::visit(ir_call
*ir
)
2445 const char *callee
= ir
->callee
->function_name();
2447 if (!strcmp("__intrinsic_atomic_read", callee
) ||
2448 !strcmp("__intrinsic_atomic_increment", callee
) ||
2449 !strcmp("__intrinsic_atomic_predecrement", callee
)) {
2450 visit_atomic_counter_intrinsic(ir
);
2452 unreachable("Unsupported intrinsic.");
2457 vec4_visitor::emit_mcs_fetch(ir_texture
*ir
, src_reg coordinate
, src_reg sampler
)
2459 vec4_instruction
*inst
=
2460 new(mem_ctx
) vec4_instruction(SHADER_OPCODE_TXF_MCS
,
2461 dst_reg(this, glsl_type::uvec4_type
));
2464 inst
->src
[1] = sampler
;
2466 /* parameters are: u, v, r, lod; lod will always be zero due to api restrictions */
2467 int param_base
= inst
->base_mrf
;
2468 int coord_mask
= (1 << ir
->coordinate
->type
->vector_elements
) - 1;
2469 int zero_mask
= 0xf & ~coord_mask
;
2471 emit(MOV(dst_reg(MRF
, param_base
, ir
->coordinate
->type
, coord_mask
),
2474 emit(MOV(dst_reg(MRF
, param_base
, ir
->coordinate
->type
, zero_mask
),
2478 return src_reg(inst
->dst
);
2482 is_high_sampler(const struct brw_device_info
*devinfo
, src_reg sampler
)
2484 if (devinfo
->gen
< 8 && !devinfo
->is_haswell
)
2487 return sampler
.file
!= IMM
|| sampler
.fixed_hw_reg
.dw1
.ud
>= 16;
2491 vec4_visitor::visit(ir_texture
*ir
)
2494 _mesa_get_sampler_uniform_value(ir
->sampler
, shader_prog
, prog
);
2496 ir_rvalue
*nonconst_sampler_index
=
2497 _mesa_get_sampler_array_nonconst_index(ir
->sampler
);
2499 /* Handle non-constant sampler array indexing */
2500 src_reg sampler_reg
;
2501 if (nonconst_sampler_index
) {
2502 /* The highest sampler which may be used by this operation is
2503 * the last element of the array. Mark it here, because the generator
2504 * doesn't have enough information to determine the bound.
2506 uint32_t array_size
= ir
->sampler
->as_dereference_array()
2507 ->array
->type
->array_size();
2509 uint32_t max_used
= sampler
+ array_size
- 1;
2510 if (ir
->op
== ir_tg4
&& devinfo
->gen
< 8) {
2511 max_used
+= prog_data
->base
.binding_table
.gather_texture_start
;
2513 max_used
+= prog_data
->base
.binding_table
.texture_start
;
2516 brw_mark_surface_used(&prog_data
->base
, max_used
);
2518 /* Emit code to evaluate the actual indexing expression */
2519 nonconst_sampler_index
->accept(this);
2520 dst_reg
temp(this, glsl_type::uint_type
);
2521 emit(ADD(temp
, this->result
, src_reg(sampler
)));
2522 emit_uniformize(temp
, src_reg(temp
));
2524 sampler_reg
= src_reg(temp
);
2526 /* Single sampler, or constant array index; the indexing expression
2527 * is just an immediate.
2529 sampler_reg
= src_reg(sampler
);
2532 /* When tg4 is used with the degenerate ZERO/ONE swizzles, don't bother
2533 * emitting anything other than setting up the constant result.
2535 if (ir
->op
== ir_tg4
) {
2536 ir_constant
*chan
= ir
->lod_info
.component
->as_constant();
2537 int swiz
= GET_SWZ(key
->tex
.swizzles
[sampler
], chan
->value
.i
[0]);
2538 if (swiz
== SWIZZLE_ZERO
|| swiz
== SWIZZLE_ONE
) {
2539 dst_reg
result(this, ir
->type
);
2540 this->result
= src_reg(result
);
2541 emit(MOV(result
, src_reg(swiz
== SWIZZLE_ONE
? 1.0f
: 0.0f
)));
2546 /* Should be lowered by do_lower_texture_projection */
2547 assert(!ir
->projector
);
2549 /* Should be lowered */
2550 assert(!ir
->offset
|| !ir
->offset
->type
->is_array());
2552 /* Generate code to compute all the subexpression trees. This has to be
2553 * done before loading any values into MRFs for the sampler message since
2554 * generating these values may involve SEND messages that need the MRFs.
2557 if (ir
->coordinate
) {
2558 ir
->coordinate
->accept(this);
2559 coordinate
= this->result
;
2562 src_reg shadow_comparitor
;
2563 if (ir
->shadow_comparitor
) {
2564 ir
->shadow_comparitor
->accept(this);
2565 shadow_comparitor
= this->result
;
2568 bool has_nonconstant_offset
= ir
->offset
&& !ir
->offset
->as_constant();
2569 src_reg offset_value
;
2570 if (has_nonconstant_offset
) {
2571 ir
->offset
->accept(this);
2572 offset_value
= src_reg(this->result
);
2575 const glsl_type
*lod_type
= NULL
, *sample_index_type
= NULL
;
2576 src_reg lod
, dPdx
, dPdy
, sample_index
, mcs
;
2579 lod
= src_reg(0.0f
);
2580 lod_type
= glsl_type::float_type
;
2585 ir
->lod_info
.lod
->accept(this);
2587 lod_type
= ir
->lod_info
.lod
->type
;
2589 case ir_query_levels
:
2591 lod_type
= glsl_type::int_type
;
2594 ir
->lod_info
.sample_index
->accept(this);
2595 sample_index
= this->result
;
2596 sample_index_type
= ir
->lod_info
.sample_index
->type
;
2598 if (devinfo
->gen
>= 7 && key
->tex
.compressed_multisample_layout_mask
& (1<<sampler
))
2599 mcs
= emit_mcs_fetch(ir
, coordinate
, sampler_reg
);
2604 ir
->lod_info
.grad
.dPdx
->accept(this);
2605 dPdx
= this->result
;
2607 ir
->lod_info
.grad
.dPdy
->accept(this);
2608 dPdy
= this->result
;
2610 lod_type
= ir
->lod_info
.grad
.dPdx
->type
;
2620 case ir_tex
: opcode
= SHADER_OPCODE_TXL
; break;
2621 case ir_txl
: opcode
= SHADER_OPCODE_TXL
; break;
2622 case ir_txd
: opcode
= SHADER_OPCODE_TXD
; break;
2623 case ir_txf
: opcode
= SHADER_OPCODE_TXF
; break;
2624 case ir_txf_ms
: opcode
= SHADER_OPCODE_TXF_CMS
; break;
2625 case ir_txs
: opcode
= SHADER_OPCODE_TXS
; break;
2626 case ir_tg4
: opcode
= has_nonconstant_offset
2627 ? SHADER_OPCODE_TG4_OFFSET
: SHADER_OPCODE_TG4
; break;
2628 case ir_query_levels
: opcode
= SHADER_OPCODE_TXS
; break;
2630 unreachable("TXB is not valid for vertex shaders.");
2632 unreachable("LOD is not valid for vertex shaders.");
2634 unreachable("Unrecognized tex op");
2637 vec4_instruction
*inst
= new(mem_ctx
) vec4_instruction(
2638 opcode
, dst_reg(this, ir
->type
));
2640 if (ir
->offset
!= NULL
&& !has_nonconstant_offset
) {
2642 brw_texture_offset(ir
->offset
->as_constant()->value
.i
,
2643 ir
->offset
->type
->vector_elements
);
2646 /* Stuff the channel select bits in the top of the texture offset */
2647 if (ir
->op
== ir_tg4
)
2648 inst
->offset
|= gather_channel(ir
, sampler
) << 16;
2650 /* The message header is necessary for:
2652 * - Gen9+ for selecting SIMD4x2
2654 * - Gather channel selection
2655 * - Sampler indices too large to fit in a 4-bit value.
2657 inst
->header_present
=
2658 devinfo
->gen
< 5 || devinfo
->gen
>= 9 ||
2659 inst
->offset
!= 0 || ir
->op
== ir_tg4
||
2660 is_high_sampler(devinfo
, sampler_reg
);
2662 inst
->mlen
= inst
->header_present
+ 1; /* always at least one */
2663 inst
->dst
.writemask
= WRITEMASK_XYZW
;
2664 inst
->shadow_compare
= ir
->shadow_comparitor
!= NULL
;
2666 inst
->src
[1] = sampler_reg
;
2668 /* MRF for the first parameter */
2669 int param_base
= inst
->base_mrf
+ inst
->header_present
;
2671 if (ir
->op
== ir_txs
|| ir
->op
== ir_query_levels
) {
2672 int writemask
= devinfo
->gen
== 4 ? WRITEMASK_W
: WRITEMASK_X
;
2673 emit(MOV(dst_reg(MRF
, param_base
, lod_type
, writemask
), lod
));
2675 /* Load the coordinate */
2676 /* FINISHME: gl_clamp_mask and saturate */
2677 int coord_mask
= (1 << ir
->coordinate
->type
->vector_elements
) - 1;
2678 int zero_mask
= 0xf & ~coord_mask
;
2680 emit(MOV(dst_reg(MRF
, param_base
, ir
->coordinate
->type
, coord_mask
),
2683 if (zero_mask
!= 0) {
2684 emit(MOV(dst_reg(MRF
, param_base
, ir
->coordinate
->type
, zero_mask
),
2687 /* Load the shadow comparitor */
2688 if (ir
->shadow_comparitor
&& ir
->op
!= ir_txd
&& (ir
->op
!= ir_tg4
|| !has_nonconstant_offset
)) {
2689 emit(MOV(dst_reg(MRF
, param_base
+ 1, ir
->shadow_comparitor
->type
,
2691 shadow_comparitor
));
2695 /* Load the LOD info */
2696 if (ir
->op
== ir_tex
|| ir
->op
== ir_txl
) {
2698 if (devinfo
->gen
>= 5) {
2699 mrf
= param_base
+ 1;
2700 if (ir
->shadow_comparitor
) {
2701 writemask
= WRITEMASK_Y
;
2702 /* mlen already incremented */
2704 writemask
= WRITEMASK_X
;
2707 } else /* devinfo->gen == 4 */ {
2709 writemask
= WRITEMASK_W
;
2711 emit(MOV(dst_reg(MRF
, mrf
, lod_type
, writemask
), lod
));
2712 } else if (ir
->op
== ir_txf
) {
2713 emit(MOV(dst_reg(MRF
, param_base
, lod_type
, WRITEMASK_W
), lod
));
2714 } else if (ir
->op
== ir_txf_ms
) {
2715 emit(MOV(dst_reg(MRF
, param_base
+ 1, sample_index_type
, WRITEMASK_X
),
2717 if (devinfo
->gen
>= 7) {
2718 /* MCS data is in the first channel of `mcs`, but we need to get it into
2719 * the .y channel of the second vec4 of params, so replicate .x across
2720 * the whole vec4 and then mask off everything except .y
2722 mcs
.swizzle
= BRW_SWIZZLE_XXXX
;
2723 emit(MOV(dst_reg(MRF
, param_base
+ 1, glsl_type::uint_type
, WRITEMASK_Y
),
2727 } else if (ir
->op
== ir_txd
) {
2728 const glsl_type
*type
= lod_type
;
2730 if (devinfo
->gen
>= 5) {
2731 dPdx
.swizzle
= BRW_SWIZZLE4(SWIZZLE_X
,SWIZZLE_X
,SWIZZLE_Y
,SWIZZLE_Y
);
2732 dPdy
.swizzle
= BRW_SWIZZLE4(SWIZZLE_X
,SWIZZLE_X
,SWIZZLE_Y
,SWIZZLE_Y
);
2733 emit(MOV(dst_reg(MRF
, param_base
+ 1, type
, WRITEMASK_XZ
), dPdx
));
2734 emit(MOV(dst_reg(MRF
, param_base
+ 1, type
, WRITEMASK_YW
), dPdy
));
2737 if (ir
->type
->vector_elements
== 3 || ir
->shadow_comparitor
) {
2738 dPdx
.swizzle
= BRW_SWIZZLE_ZZZZ
;
2739 dPdy
.swizzle
= BRW_SWIZZLE_ZZZZ
;
2740 emit(MOV(dst_reg(MRF
, param_base
+ 2, type
, WRITEMASK_X
), dPdx
));
2741 emit(MOV(dst_reg(MRF
, param_base
+ 2, type
, WRITEMASK_Y
), dPdy
));
2744 if (ir
->shadow_comparitor
) {
2745 emit(MOV(dst_reg(MRF
, param_base
+ 2,
2746 ir
->shadow_comparitor
->type
, WRITEMASK_Z
),
2747 shadow_comparitor
));
2750 } else /* devinfo->gen == 4 */ {
2751 emit(MOV(dst_reg(MRF
, param_base
+ 1, type
, WRITEMASK_XYZ
), dPdx
));
2752 emit(MOV(dst_reg(MRF
, param_base
+ 2, type
, WRITEMASK_XYZ
), dPdy
));
2755 } else if (ir
->op
== ir_tg4
&& has_nonconstant_offset
) {
2756 if (ir
->shadow_comparitor
) {
2757 emit(MOV(dst_reg(MRF
, param_base
, ir
->shadow_comparitor
->type
, WRITEMASK_W
),
2758 shadow_comparitor
));
2761 emit(MOV(dst_reg(MRF
, param_base
+ 1, glsl_type::ivec2_type
, WRITEMASK_XY
),
2769 /* fixup num layers (z) for cube arrays: hardware returns faces * layers;
2770 * spec requires layers.
2772 if (ir
->op
== ir_txs
) {
2773 glsl_type
const *type
= ir
->sampler
->type
;
2774 if (type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_CUBE
&&
2775 type
->sampler_array
) {
2776 emit_math(SHADER_OPCODE_INT_QUOTIENT
,
2777 writemask(inst
->dst
, WRITEMASK_Z
),
2778 src_reg(inst
->dst
), src_reg(6));
2782 if (devinfo
->gen
== 6 && ir
->op
== ir_tg4
) {
2783 emit_gen6_gather_wa(key
->tex
.gen6_gather_wa
[sampler
], inst
->dst
);
2786 swizzle_result(ir
, src_reg(inst
->dst
), sampler
);
2790 * Apply workarounds for Gen6 gather with UINT/SINT
2793 vec4_visitor::emit_gen6_gather_wa(uint8_t wa
, dst_reg dst
)
2798 int width
= (wa
& WA_8BIT
) ? 8 : 16;
2799 dst_reg dst_f
= dst
;
2800 dst_f
.type
= BRW_REGISTER_TYPE_F
;
2802 /* Convert from UNORM to UINT */
2803 emit(MUL(dst_f
, src_reg(dst_f
), src_reg((float)((1 << width
) - 1))));
2804 emit(MOV(dst
, src_reg(dst_f
)));
2807 /* Reinterpret the UINT value as a signed INT value by
2808 * shifting the sign bit into place, then shifting back
2811 emit(SHL(dst
, src_reg(dst
), src_reg(32 - width
)));
2812 emit(ASR(dst
, src_reg(dst
), src_reg(32 - width
)));
2817 * Set up the gather channel based on the swizzle, for gather4.
2820 vec4_visitor::gather_channel(ir_texture
*ir
, uint32_t sampler
)
2822 ir_constant
*chan
= ir
->lod_info
.component
->as_constant();
2823 int swiz
= GET_SWZ(key
->tex
.swizzles
[sampler
], chan
->value
.i
[0]);
2825 case SWIZZLE_X
: return 0;
2827 /* gather4 sampler is broken for green channel on RG32F --
2828 * we must ask for blue instead.
2830 if (key
->tex
.gather_channel_quirk_mask
& (1<<sampler
))
2833 case SWIZZLE_Z
: return 2;
2834 case SWIZZLE_W
: return 3;
2836 unreachable("Not reached"); /* zero, one swizzles handled already */
2841 vec4_visitor::swizzle_result(ir_texture
*ir
, src_reg orig_val
, uint32_t sampler
)
2843 int s
= key
->tex
.swizzles
[sampler
];
2845 this->result
= src_reg(this, ir
->type
);
2846 dst_reg
swizzled_result(this->result
);
2848 if (ir
->op
== ir_query_levels
) {
2849 /* # levels is in .w */
2850 orig_val
.swizzle
= BRW_SWIZZLE4(SWIZZLE_W
, SWIZZLE_W
, SWIZZLE_W
, SWIZZLE_W
);
2851 emit(MOV(swizzled_result
, orig_val
));
2855 if (ir
->op
== ir_txs
|| ir
->type
== glsl_type::float_type
2856 || s
== SWIZZLE_NOOP
|| ir
->op
== ir_tg4
) {
2857 emit(MOV(swizzled_result
, orig_val
));
2862 int zero_mask
= 0, one_mask
= 0, copy_mask
= 0;
2863 int swizzle
[4] = {0};
2865 for (int i
= 0; i
< 4; i
++) {
2866 switch (GET_SWZ(s
, i
)) {
2868 zero_mask
|= (1 << i
);
2871 one_mask
|= (1 << i
);
2874 copy_mask
|= (1 << i
);
2875 swizzle
[i
] = GET_SWZ(s
, i
);
2881 orig_val
.swizzle
= BRW_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
2882 swizzled_result
.writemask
= copy_mask
;
2883 emit(MOV(swizzled_result
, orig_val
));
2887 swizzled_result
.writemask
= zero_mask
;
2888 emit(MOV(swizzled_result
, src_reg(0.0f
)));
2892 swizzled_result
.writemask
= one_mask
;
2893 emit(MOV(swizzled_result
, src_reg(1.0f
)));
2898 vec4_visitor::visit(ir_return
*)
2900 unreachable("not reached");
2904 vec4_visitor::visit(ir_discard
*)
2906 unreachable("not reached");
2910 vec4_visitor::visit(ir_if
*ir
)
2912 /* Don't point the annotation at the if statement, because then it plus
2913 * the then and else blocks get printed.
2915 this->base_ir
= ir
->condition
;
2917 if (devinfo
->gen
== 6) {
2920 enum brw_predicate predicate
;
2921 emit_bool_to_cond_code(ir
->condition
, &predicate
);
2922 emit(IF(predicate
));
2925 visit_instructions(&ir
->then_instructions
);
2927 if (!ir
->else_instructions
.is_empty()) {
2928 this->base_ir
= ir
->condition
;
2929 emit(BRW_OPCODE_ELSE
);
2931 visit_instructions(&ir
->else_instructions
);
2934 this->base_ir
= ir
->condition
;
2935 emit(BRW_OPCODE_ENDIF
);
2939 vec4_visitor::visit(ir_emit_vertex
*)
2941 unreachable("not reached");
2945 vec4_visitor::visit(ir_end_primitive
*)
2947 unreachable("not reached");
2951 vec4_visitor::emit_untyped_atomic(unsigned atomic_op
, unsigned surf_index
,
2952 dst_reg dst
, src_reg offset
,
2953 src_reg src0
, src_reg src1
)
2957 /* Set the atomic operation offset. */
2958 emit(MOV(brw_writemask(brw_uvec_mrf(8, mlen
, 0), WRITEMASK_X
), offset
));
2961 /* Set the atomic operation arguments. */
2962 if (src0
.file
!= BAD_FILE
) {
2963 emit(MOV(brw_writemask(brw_uvec_mrf(8, mlen
, 0), WRITEMASK_X
), src0
));
2967 if (src1
.file
!= BAD_FILE
) {
2968 emit(MOV(brw_writemask(brw_uvec_mrf(8, mlen
, 0), WRITEMASK_X
), src1
));
2972 /* Emit the instruction. Note that this maps to the normal SIMD8
2973 * untyped atomic message on Ivy Bridge, but that's OK because
2974 * unused channels will be masked out.
2976 vec4_instruction
*inst
= emit(SHADER_OPCODE_UNTYPED_ATOMIC
, dst
,
2978 src_reg(surf_index
), src_reg(atomic_op
));
2983 vec4_visitor::emit_untyped_surface_read(unsigned surf_index
, dst_reg dst
,
2986 /* Set the surface read offset. */
2987 emit(MOV(brw_writemask(brw_uvec_mrf(8, 0, 0), WRITEMASK_X
), offset
));
2989 /* Emit the instruction. Note that this maps to the normal SIMD8
2990 * untyped surface read message, but that's OK because unused
2991 * channels will be masked out.
2993 vec4_instruction
*inst
= emit(SHADER_OPCODE_UNTYPED_SURFACE_READ
, dst
,
2995 src_reg(surf_index
), src_reg(1));
3000 vec4_visitor::emit_ndc_computation()
3002 /* Get the position */
3003 src_reg pos
= src_reg(output_reg
[VARYING_SLOT_POS
]);
3005 /* Build ndc coords, which are (x/w, y/w, z/w, 1/w) */
3006 dst_reg ndc
= dst_reg(this, glsl_type::vec4_type
);
3007 output_reg
[BRW_VARYING_SLOT_NDC
] = ndc
;
3009 current_annotation
= "NDC";
3010 dst_reg ndc_w
= ndc
;
3011 ndc_w
.writemask
= WRITEMASK_W
;
3012 src_reg pos_w
= pos
;
3013 pos_w
.swizzle
= BRW_SWIZZLE4(SWIZZLE_W
, SWIZZLE_W
, SWIZZLE_W
, SWIZZLE_W
);
3014 emit_math(SHADER_OPCODE_RCP
, ndc_w
, pos_w
);
3016 dst_reg ndc_xyz
= ndc
;
3017 ndc_xyz
.writemask
= WRITEMASK_XYZ
;
3019 emit(MUL(ndc_xyz
, pos
, src_reg(ndc_w
)));
3023 vec4_visitor::emit_psiz_and_flags(dst_reg reg
)
3025 if (devinfo
->gen
< 6 &&
3026 ((prog_data
->vue_map
.slots_valid
& VARYING_BIT_PSIZ
) ||
3027 key
->userclip_active
|| devinfo
->has_negative_rhw_bug
)) {
3028 dst_reg header1
= dst_reg(this, glsl_type::uvec4_type
);
3029 dst_reg header1_w
= header1
;
3030 header1_w
.writemask
= WRITEMASK_W
;
3032 emit(MOV(header1
, 0u));
3034 if (prog_data
->vue_map
.slots_valid
& VARYING_BIT_PSIZ
) {
3035 src_reg psiz
= src_reg(output_reg
[VARYING_SLOT_PSIZ
]);
3037 current_annotation
= "Point size";
3038 emit(MUL(header1_w
, psiz
, src_reg((float)(1 << 11))));
3039 emit(AND(header1_w
, src_reg(header1_w
), 0x7ff << 8));
3042 if (key
->userclip_active
) {
3043 current_annotation
= "Clipping flags";
3044 dst_reg flags0
= dst_reg(this, glsl_type::uint_type
);
3045 dst_reg flags1
= dst_reg(this, glsl_type::uint_type
);
3047 emit(CMP(dst_null_f(), src_reg(output_reg
[VARYING_SLOT_CLIP_DIST0
]), src_reg(0.0f
), BRW_CONDITIONAL_L
));
3048 emit(VS_OPCODE_UNPACK_FLAGS_SIMD4X2
, flags0
, src_reg(0));
3049 emit(OR(header1_w
, src_reg(header1_w
), src_reg(flags0
)));
3051 emit(CMP(dst_null_f(), src_reg(output_reg
[VARYING_SLOT_CLIP_DIST1
]), src_reg(0.0f
), BRW_CONDITIONAL_L
));
3052 emit(VS_OPCODE_UNPACK_FLAGS_SIMD4X2
, flags1
, src_reg(0));
3053 emit(SHL(flags1
, src_reg(flags1
), src_reg(4)));
3054 emit(OR(header1_w
, src_reg(header1_w
), src_reg(flags1
)));
3057 /* i965 clipping workaround:
3058 * 1) Test for -ve rhw
3060 * set ndc = (0,0,0,0)
3063 * Later, clipping will detect ucp[6] and ensure the primitive is
3064 * clipped against all fixed planes.
3066 if (devinfo
->has_negative_rhw_bug
) {
3067 src_reg ndc_w
= src_reg(output_reg
[BRW_VARYING_SLOT_NDC
]);
3068 ndc_w
.swizzle
= BRW_SWIZZLE_WWWW
;
3069 emit(CMP(dst_null_f(), ndc_w
, src_reg(0.0f
), BRW_CONDITIONAL_L
));
3070 vec4_instruction
*inst
;
3071 inst
= emit(OR(header1_w
, src_reg(header1_w
), src_reg(1u << 6)));
3072 inst
->predicate
= BRW_PREDICATE_NORMAL
;
3073 inst
= emit(MOV(output_reg
[BRW_VARYING_SLOT_NDC
], src_reg(0.0f
)));
3074 inst
->predicate
= BRW_PREDICATE_NORMAL
;
3077 emit(MOV(retype(reg
, BRW_REGISTER_TYPE_UD
), src_reg(header1
)));
3078 } else if (devinfo
->gen
< 6) {
3079 emit(MOV(retype(reg
, BRW_REGISTER_TYPE_UD
), 0u));
3081 emit(MOV(retype(reg
, BRW_REGISTER_TYPE_D
), src_reg(0)));
3082 if (prog_data
->vue_map
.slots_valid
& VARYING_BIT_PSIZ
) {
3083 dst_reg reg_w
= reg
;
3084 reg_w
.writemask
= WRITEMASK_W
;
3085 emit(MOV(reg_w
, src_reg(output_reg
[VARYING_SLOT_PSIZ
])));
3087 if (prog_data
->vue_map
.slots_valid
& VARYING_BIT_LAYER
) {
3088 dst_reg reg_y
= reg
;
3089 reg_y
.writemask
= WRITEMASK_Y
;
3090 reg_y
.type
= BRW_REGISTER_TYPE_D
;
3091 emit(MOV(reg_y
, src_reg(output_reg
[VARYING_SLOT_LAYER
])));
3093 if (prog_data
->vue_map
.slots_valid
& VARYING_BIT_VIEWPORT
) {
3094 dst_reg reg_z
= reg
;
3095 reg_z
.writemask
= WRITEMASK_Z
;
3096 reg_z
.type
= BRW_REGISTER_TYPE_D
;
3097 emit(MOV(reg_z
, src_reg(output_reg
[VARYING_SLOT_VIEWPORT
])));
3103 vec4_visitor::emit_clip_distances(dst_reg reg
, int offset
)
3105 /* From the GLSL 1.30 spec, section 7.1 (Vertex Shader Special Variables):
3107 * "If a linked set of shaders forming the vertex stage contains no
3108 * static write to gl_ClipVertex or gl_ClipDistance, but the
3109 * application has requested clipping against user clip planes through
3110 * the API, then the coordinate written to gl_Position is used for
3111 * comparison against the user clip planes."
3113 * This function is only called if the shader didn't write to
3114 * gl_ClipDistance. Accordingly, we use gl_ClipVertex to perform clipping
3115 * if the user wrote to it; otherwise we use gl_Position.
3117 gl_varying_slot clip_vertex
= VARYING_SLOT_CLIP_VERTEX
;
3118 if (!(prog_data
->vue_map
.slots_valid
& VARYING_BIT_CLIP_VERTEX
)) {
3119 clip_vertex
= VARYING_SLOT_POS
;
3122 for (int i
= 0; i
+ offset
< key
->nr_userclip_plane_consts
&& i
< 4;
3124 reg
.writemask
= 1 << i
;
3126 src_reg(output_reg
[clip_vertex
]),
3127 src_reg(this->userplane
[i
+ offset
])));
3132 vec4_visitor::emit_generic_urb_slot(dst_reg reg
, int varying
)
3134 assert (varying
< VARYING_SLOT_MAX
);
3135 reg
.type
= output_reg
[varying
].type
;
3136 current_annotation
= output_reg_annotation
[varying
];
3137 /* Copy the register, saturating if necessary */
3138 return emit(MOV(reg
, src_reg(output_reg
[varying
])));
3142 vec4_visitor::emit_urb_slot(dst_reg reg
, int varying
)
3144 reg
.type
= BRW_REGISTER_TYPE_F
;
3147 case VARYING_SLOT_PSIZ
:
3149 /* PSIZ is always in slot 0, and is coupled with other flags. */
3150 current_annotation
= "indices, point width, clip flags";
3151 emit_psiz_and_flags(reg
);
3154 case BRW_VARYING_SLOT_NDC
:
3155 current_annotation
= "NDC";
3156 emit(MOV(reg
, src_reg(output_reg
[BRW_VARYING_SLOT_NDC
])));
3158 case VARYING_SLOT_POS
:
3159 current_annotation
= "gl_Position";
3160 emit(MOV(reg
, src_reg(output_reg
[VARYING_SLOT_POS
])));
3162 case VARYING_SLOT_EDGE
:
3163 /* This is present when doing unfilled polygons. We're supposed to copy
3164 * the edge flag from the user-provided vertex array
3165 * (glEdgeFlagPointer), or otherwise we'll copy from the current value
3166 * of that attribute (starts as 1.0f). This is then used in clipping to
3167 * determine which edges should be drawn as wireframe.
3169 current_annotation
= "edge flag";
3170 emit(MOV(reg
, src_reg(dst_reg(ATTR
, VERT_ATTRIB_EDGEFLAG
,
3171 glsl_type::float_type
, WRITEMASK_XYZW
))));
3173 case BRW_VARYING_SLOT_PAD
:
3174 /* No need to write to this slot */
3176 case VARYING_SLOT_COL0
:
3177 case VARYING_SLOT_COL1
:
3178 case VARYING_SLOT_BFC0
:
3179 case VARYING_SLOT_BFC1
: {
3180 /* These built-in varyings are only supported in compatibility mode,
3181 * and we only support GS in core profile. So, this must be a vertex
3184 assert(stage
== MESA_SHADER_VERTEX
);
3185 vec4_instruction
*inst
= emit_generic_urb_slot(reg
, varying
);
3186 if (((struct brw_vs_prog_key
*) key
)->clamp_vertex_color
)
3187 inst
->saturate
= true;
3192 emit_generic_urb_slot(reg
, varying
);
3198 align_interleaved_urb_mlen(const struct brw_device_info
*devinfo
, int mlen
)
3200 if (devinfo
->gen
>= 6) {
3201 /* URB data written (does not include the message header reg) must
3202 * be a multiple of 256 bits, or 2 VS registers. See vol5c.5,
3203 * section 5.4.3.2.2: URB_INTERLEAVED.
3205 * URB entries are allocated on a multiple of 1024 bits, so an
3206 * extra 128 bits written here to make the end align to 256 is
3209 if ((mlen
% 2) != 1)
3218 * Generates the VUE payload plus the necessary URB write instructions to
3221 * The VUE layout is documented in Volume 2a.
3224 vec4_visitor::emit_vertex()
3226 /* MRF 0 is reserved for the debugger, so start with message header
3231 /* In the process of generating our URB write message contents, we
3232 * may need to unspill a register or load from an array. Those
3233 * reads would use MRFs 14-15.
3235 int max_usable_mrf
= 13;
3237 /* The following assertion verifies that max_usable_mrf causes an
3238 * even-numbered amount of URB write data, which will meet gen6's
3239 * requirements for length alignment.
3241 assert ((max_usable_mrf
- base_mrf
) % 2 == 0);
3243 /* First mrf is the g0-based message header containing URB handles and
3246 emit_urb_write_header(mrf
++);
3248 if (devinfo
->gen
< 6) {
3249 emit_ndc_computation();
3252 /* Lower legacy ff and ClipVertex clipping to clip distances */
3253 if (key
->userclip_active
&& !prog
->UsesClipDistanceOut
) {
3254 current_annotation
= "user clip distances";
3256 output_reg
[VARYING_SLOT_CLIP_DIST0
] = dst_reg(this, glsl_type::vec4_type
);
3257 output_reg
[VARYING_SLOT_CLIP_DIST1
] = dst_reg(this, glsl_type::vec4_type
);
3259 emit_clip_distances(output_reg
[VARYING_SLOT_CLIP_DIST0
], 0);
3260 emit_clip_distances(output_reg
[VARYING_SLOT_CLIP_DIST1
], 4);
3263 /* We may need to split this up into several URB writes, so do them in a
3267 bool complete
= false;
3269 /* URB offset is in URB row increments, and each of our MRFs is half of
3270 * one of those, since we're doing interleaved writes.
3272 int offset
= slot
/ 2;
3275 for (; slot
< prog_data
->vue_map
.num_slots
; ++slot
) {
3276 emit_urb_slot(dst_reg(MRF
, mrf
++),
3277 prog_data
->vue_map
.slot_to_varying
[slot
]);
3279 /* If this was max_usable_mrf, we can't fit anything more into this
3282 if (mrf
> max_usable_mrf
) {
3288 complete
= slot
>= prog_data
->vue_map
.num_slots
;
3289 current_annotation
= "URB write";
3290 vec4_instruction
*inst
= emit_urb_write_opcode(complete
);
3291 inst
->base_mrf
= base_mrf
;
3292 inst
->mlen
= align_interleaved_urb_mlen(devinfo
, mrf
- base_mrf
);
3293 inst
->offset
+= offset
;
3299 vec4_visitor::get_scratch_offset(bblock_t
*block
, vec4_instruction
*inst
,
3300 src_reg
*reladdr
, int reg_offset
)
3302 /* Because we store the values to scratch interleaved like our
3303 * vertex data, we need to scale the vec4 index by 2.
3305 int message_header_scale
= 2;
3307 /* Pre-gen6, the message header uses byte offsets instead of vec4
3308 * (16-byte) offset units.
3310 if (devinfo
->gen
< 6)
3311 message_header_scale
*= 16;
3314 src_reg index
= src_reg(this, glsl_type::int_type
);
3316 emit_before(block
, inst
, ADD(dst_reg(index
), *reladdr
,
3317 src_reg(reg_offset
)));
3318 emit_before(block
, inst
, MUL(dst_reg(index
), index
,
3319 src_reg(message_header_scale
)));
3323 return src_reg(reg_offset
* message_header_scale
);
3328 vec4_visitor::get_pull_constant_offset(bblock_t
* block
, vec4_instruction
*inst
,
3329 src_reg
*reladdr
, int reg_offset
)
3332 src_reg index
= src_reg(this, glsl_type::int_type
);
3334 emit_before(block
, inst
, ADD(dst_reg(index
), *reladdr
,
3335 src_reg(reg_offset
)));
3337 /* Pre-gen6, the message header uses byte offsets instead of vec4
3338 * (16-byte) offset units.
3340 if (devinfo
->gen
< 6) {
3341 emit_before(block
, inst
, MUL(dst_reg(index
), index
, src_reg(16)));
3345 } else if (devinfo
->gen
>= 8) {
3346 /* Store the offset in a GRF so we can send-from-GRF. */
3347 src_reg offset
= src_reg(this, glsl_type::int_type
);
3348 emit_before(block
, inst
, MOV(dst_reg(offset
), src_reg(reg_offset
)));
3351 int message_header_scale
= devinfo
->gen
< 6 ? 16 : 1;
3352 return src_reg(reg_offset
* message_header_scale
);
3357 * Emits an instruction before @inst to load the value named by @orig_src
3358 * from scratch space at @base_offset to @temp.
3360 * @base_offset is measured in 32-byte units (the size of a register).
3363 vec4_visitor::emit_scratch_read(bblock_t
*block
, vec4_instruction
*inst
,
3364 dst_reg temp
, src_reg orig_src
,
3367 int reg_offset
= base_offset
+ orig_src
.reg_offset
;
3368 src_reg index
= get_scratch_offset(block
, inst
, orig_src
.reladdr
,
3371 emit_before(block
, inst
, SCRATCH_READ(temp
, index
));
3375 * Emits an instruction after @inst to store the value to be written
3376 * to @orig_dst to scratch space at @base_offset, from @temp.
3378 * @base_offset is measured in 32-byte units (the size of a register).
3381 vec4_visitor::emit_scratch_write(bblock_t
*block
, vec4_instruction
*inst
,
3384 int reg_offset
= base_offset
+ inst
->dst
.reg_offset
;
3385 src_reg index
= get_scratch_offset(block
, inst
, inst
->dst
.reladdr
,
3388 /* Create a temporary register to store *inst's result in.
3390 * We have to be careful in MOVing from our temporary result register in
3391 * the scratch write. If we swizzle from channels of the temporary that
3392 * weren't initialized, it will confuse live interval analysis, which will
3393 * make spilling fail to make progress.
3395 const src_reg temp
= swizzle(retype(src_reg(this, glsl_type::vec4_type
),
3397 brw_swizzle_for_mask(inst
->dst
.writemask
));
3398 dst_reg dst
= dst_reg(brw_writemask(brw_vec8_grf(0, 0),
3399 inst
->dst
.writemask
));
3400 vec4_instruction
*write
= SCRATCH_WRITE(dst
, temp
, index
);
3401 write
->predicate
= inst
->predicate
;
3402 write
->ir
= inst
->ir
;
3403 write
->annotation
= inst
->annotation
;
3404 inst
->insert_after(block
, write
);
3406 inst
->dst
.file
= temp
.file
;
3407 inst
->dst
.reg
= temp
.reg
;
3408 inst
->dst
.reg_offset
= temp
.reg_offset
;
3409 inst
->dst
.reladdr
= NULL
;
3413 * Checks if \p src and/or \p src.reladdr require a scratch read, and if so,
3414 * adds the scratch read(s) before \p inst. The function also checks for
3415 * recursive reladdr scratch accesses, issuing the corresponding scratch
3416 * loads and rewriting reladdr references accordingly.
3418 * \return \p src if it did not require a scratch load, otherwise, the
3419 * register holding the result of the scratch load that the caller should
3420 * use to rewrite src.
3423 vec4_visitor::emit_resolve_reladdr(int scratch_loc
[], bblock_t
*block
,
3424 vec4_instruction
*inst
, src_reg src
)
3426 /* Resolve recursive reladdr scratch access by calling ourselves
3430 *src
.reladdr
= emit_resolve_reladdr(scratch_loc
, block
, inst
,
3433 /* Now handle scratch access on src */
3434 if (src
.file
== GRF
&& scratch_loc
[src
.reg
] != -1) {
3435 dst_reg temp
= dst_reg(this, glsl_type::vec4_type
);
3436 emit_scratch_read(block
, inst
, temp
, src
, scratch_loc
[src
.reg
]);
3438 src
.reg_offset
= temp
.reg_offset
;
3446 * We can't generally support array access in GRF space, because a
3447 * single instruction's destination can only span 2 contiguous
3448 * registers. So, we send all GRF arrays that get variable index
3449 * access to scratch space.
3452 vec4_visitor::move_grf_array_access_to_scratch()
3454 int scratch_loc
[this->alloc
.count
];
3455 memset(scratch_loc
, -1, sizeof(scratch_loc
));
3457 /* First, calculate the set of virtual GRFs that need to be punted
3458 * to scratch due to having any array access on them, and where in
3461 foreach_block_and_inst(block
, vec4_instruction
, inst
, cfg
) {
3462 if (inst
->dst
.file
== GRF
&& inst
->dst
.reladdr
) {
3463 if (scratch_loc
[inst
->dst
.reg
] == -1) {
3464 scratch_loc
[inst
->dst
.reg
] = c
->last_scratch
;
3465 c
->last_scratch
+= this->alloc
.sizes
[inst
->dst
.reg
];
3468 for (src_reg
*iter
= inst
->dst
.reladdr
;
3470 iter
= iter
->reladdr
) {
3471 if (iter
->file
== GRF
&& scratch_loc
[iter
->reg
] == -1) {
3472 scratch_loc
[iter
->reg
] = c
->last_scratch
;
3473 c
->last_scratch
+= this->alloc
.sizes
[iter
->reg
];
3478 for (int i
= 0 ; i
< 3; i
++) {
3479 for (src_reg
*iter
= &inst
->src
[i
];
3481 iter
= iter
->reladdr
) {
3482 if (iter
->file
== GRF
&& scratch_loc
[iter
->reg
] == -1) {
3483 scratch_loc
[iter
->reg
] = c
->last_scratch
;
3484 c
->last_scratch
+= this->alloc
.sizes
[iter
->reg
];
3490 /* Now, for anything that will be accessed through scratch, rewrite
3491 * it to load/store. Note that this is a _safe list walk, because
3492 * we may generate a new scratch_write instruction after the one
3495 foreach_block_and_inst_safe(block
, vec4_instruction
, inst
, cfg
) {
3496 /* Set up the annotation tracking for new generated instructions. */
3498 current_annotation
= inst
->annotation
;
3500 /* First handle scratch access on the dst. Notice we have to handle
3501 * the case where the dst's reladdr also points to scratch space.
3503 if (inst
->dst
.reladdr
)
3504 *inst
->dst
.reladdr
= emit_resolve_reladdr(scratch_loc
, block
, inst
,
3505 *inst
->dst
.reladdr
);
3507 /* Now that we have handled any (possibly recursive) reladdr scratch
3508 * accesses for dst we can safely do the scratch write for dst itself
3510 if (inst
->dst
.file
== GRF
&& scratch_loc
[inst
->dst
.reg
] != -1)
3511 emit_scratch_write(block
, inst
, scratch_loc
[inst
->dst
.reg
]);
3513 /* Now handle scratch access on any src. In this case, since inst->src[i]
3514 * already is a src_reg, we can just call emit_resolve_reladdr with
3515 * inst->src[i] and it will take care of handling scratch loads for
3516 * both src and src.reladdr (recursively).
3518 for (int i
= 0 ; i
< 3; i
++) {
3519 inst
->src
[i
] = emit_resolve_reladdr(scratch_loc
, block
, inst
,
3526 * Emits an instruction before @inst to load the value named by @orig_src
3527 * from the pull constant buffer (surface) at @base_offset to @temp.
3530 vec4_visitor::emit_pull_constant_load(bblock_t
*block
, vec4_instruction
*inst
,
3531 dst_reg temp
, src_reg orig_src
,
3534 int reg_offset
= base_offset
+ orig_src
.reg_offset
;
3535 src_reg index
= src_reg(prog_data
->base
.binding_table
.pull_constants_start
);
3536 src_reg offset
= get_pull_constant_offset(block
, inst
, orig_src
.reladdr
,
3539 emit_pull_constant_load_reg(temp
,
3546 * Implements array access of uniforms by inserting a
3547 * PULL_CONSTANT_LOAD instruction.
3549 * Unlike temporary GRF array access (where we don't support it due to
3550 * the difficulty of doing relative addressing on instruction
3551 * destinations), we could potentially do array access of uniforms
3552 * that were loaded in GRF space as push constants. In real-world
3553 * usage we've seen, though, the arrays being used are always larger
3554 * than we could load as push constants, so just always move all
3555 * uniform array access out to a pull constant buffer.
3558 vec4_visitor::move_uniform_array_access_to_pull_constants()
3560 int pull_constant_loc
[this->uniforms
];
3561 memset(pull_constant_loc
, -1, sizeof(pull_constant_loc
));
3562 bool nested_reladdr
;
3564 /* Walk through and find array access of uniforms. Put a copy of that
3565 * uniform in the pull constant buffer.
3567 * Note that we don't move constant-indexed accesses to arrays. No
3568 * testing has been done of the performance impact of this choice.
3571 nested_reladdr
= false;
3573 foreach_block_and_inst_safe(block
, vec4_instruction
, inst
, cfg
) {
3574 for (int i
= 0 ; i
< 3; i
++) {
3575 if (inst
->src
[i
].file
!= UNIFORM
|| !inst
->src
[i
].reladdr
)
3578 int uniform
= inst
->src
[i
].reg
;
3580 if (inst
->src
[i
].reladdr
->reladdr
)
3581 nested_reladdr
= true; /* will need another pass */
3583 /* If this array isn't already present in the pull constant buffer,
3586 if (pull_constant_loc
[uniform
] == -1) {
3587 const gl_constant_value
**values
=
3588 &stage_prog_data
->param
[uniform
* 4];
3590 pull_constant_loc
[uniform
] = stage_prog_data
->nr_pull_params
/ 4;
3592 assert(uniform
< uniform_array_size
);
3593 for (int j
= 0; j
< uniform_size
[uniform
] * 4; j
++) {
3594 stage_prog_data
->pull_param
[stage_prog_data
->nr_pull_params
++]
3599 /* Set up the annotation tracking for new generated instructions. */
3601 current_annotation
= inst
->annotation
;
3603 dst_reg temp
= dst_reg(this, glsl_type::vec4_type
);
3605 emit_pull_constant_load(block
, inst
, temp
, inst
->src
[i
],
3606 pull_constant_loc
[uniform
]);
3608 inst
->src
[i
].file
= temp
.file
;
3609 inst
->src
[i
].reg
= temp
.reg
;
3610 inst
->src
[i
].reg_offset
= temp
.reg_offset
;
3611 inst
->src
[i
].reladdr
= NULL
;
3614 } while (nested_reladdr
);
3616 /* Now there are no accesses of the UNIFORM file with a reladdr, so
3617 * no need to track them as larger-than-vec4 objects. This will be
3618 * relied on in cutting out unused uniform vectors from push
3621 split_uniform_registers();
3625 vec4_visitor::resolve_ud_negate(src_reg
*reg
)
3627 if (reg
->type
!= BRW_REGISTER_TYPE_UD
||
3631 src_reg temp
= src_reg(this, glsl_type::uvec4_type
);
3632 emit(BRW_OPCODE_MOV
, dst_reg(temp
), *reg
);
3637 * Resolve the result of a Gen4-5 CMP instruction to a proper boolean.
3639 * CMP on Gen4-5 only sets the LSB of the result; the rest are undefined.
3640 * If we need a proper boolean value, we have to fix it up to be 0 or ~0.
3643 vec4_visitor::resolve_bool_comparison(ir_rvalue
*rvalue
, src_reg
*reg
)
3645 assert(devinfo
->gen
<= 5);
3647 if (!rvalue
->type
->is_boolean())
3650 src_reg and_result
= src_reg(this, rvalue
->type
);
3651 src_reg neg_result
= src_reg(this, rvalue
->type
);
3652 emit(AND(dst_reg(and_result
), *reg
, src_reg(1)));
3653 emit(MOV(dst_reg(neg_result
), negate(and_result
)));
3657 vec4_visitor::vec4_visitor(struct brw_context
*brw
,
3658 struct brw_vec4_compile
*c
,
3659 struct gl_program
*prog
,
3660 const struct brw_vue_prog_key
*key
,
3661 struct brw_vue_prog_data
*prog_data
,
3662 struct gl_shader_program
*shader_prog
,
3663 gl_shader_stage stage
,
3666 shader_time_shader_type st_base
,
3667 shader_time_shader_type st_written
,
3668 shader_time_shader_type st_reset
)
3669 : backend_visitor(brw
, shader_prog
, prog
, &prog_data
->base
, stage
),
3672 prog_data(prog_data
),
3673 sanity_param_count(0),
3675 first_non_payload_grf(0),
3676 need_all_constants_in_pull_buffer(false),
3677 no_spills(no_spills
),
3679 st_written(st_written
),
3682 this->mem_ctx
= mem_ctx
;
3683 this->failed
= false;
3685 this->base_ir
= NULL
;
3686 this->current_annotation
= NULL
;
3687 memset(this->output_reg_annotation
, 0, sizeof(this->output_reg_annotation
));
3689 this->variable_ht
= hash_table_ctor(0,
3690 hash_table_pointer_hash
,
3691 hash_table_pointer_compare
);
3693 this->virtual_grf_start
= NULL
;
3694 this->virtual_grf_end
= NULL
;
3695 this->live_intervals
= NULL
;
3697 this->max_grf
= devinfo
->gen
>= 7 ? GEN7_MRF_HACK_START
: BRW_MAX_GRF
;
3701 /* Initialize uniform_array_size to at least 1 because pre-gen6 VS requires
3702 * at least one. See setup_uniforms() in brw_vec4.cpp.
3704 this->uniform_array_size
= 1;
3706 this->uniform_array_size
=
3707 MAX2(DIV_ROUND_UP(stage_prog_data
->nr_params
, 4), 1);
3710 this->uniform_size
= rzalloc_array(mem_ctx
, int, this->uniform_array_size
);
3711 this->uniform_vector_size
= rzalloc_array(mem_ctx
, int, this->uniform_array_size
);
3714 vec4_visitor::~vec4_visitor()
3716 hash_table_dtor(this->variable_ht
);
3721 vec4_visitor::fail(const char *format
, ...)
3731 va_start(va
, format
);
3732 msg
= ralloc_vasprintf(mem_ctx
, format
, va
);
3734 msg
= ralloc_asprintf(mem_ctx
, "%s compile failed: %s\n", stage_abbrev
, msg
);
3736 this->fail_msg
= msg
;
3738 if (debug_enabled
) {
3739 fprintf(stderr
, "%s", msg
);
3743 } /* namespace brw */