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_size
= 0;
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
->NumUniformStorage
; u
++) {
687 struct gl_uniform_storage
*storage
= &shader_prog
->UniformStorage
[u
];
689 if (storage
->builtin
)
692 if (strncmp(ir
->name
, storage
->name
, namelen
) != 0 ||
693 (storage
->name
[namelen
] != 0 &&
694 storage
->name
[namelen
] != '.' &&
695 storage
->name
[namelen
] != '[')) {
699 gl_constant_value
*components
= storage
->storage
;
700 unsigned vector_count
= (MAX2(storage
->array_elements
, 1) *
701 storage
->type
->matrix_columns
);
703 for (unsigned s
= 0; s
< vector_count
; s
++) {
704 assert(uniforms
< uniform_array_size
);
705 uniform_vector_size
[uniforms
] = storage
->type
->vector_elements
;
708 for (i
= 0; i
< uniform_vector_size
[uniforms
]; i
++) {
709 stage_prog_data
->param
[uniforms
* 4 + i
] = components
;
713 static gl_constant_value zero
= { 0.0 };
714 stage_prog_data
->param
[uniforms
* 4 + i
] = &zero
;
723 vec4_visitor::setup_uniform_clipplane_values(gl_clip_plane
*clip_planes
)
725 for (int i
= 0; i
< key
->nr_userclip_plane_consts
; ++i
) {
726 assert(this->uniforms
< uniform_array_size
);
727 this->uniform_vector_size
[this->uniforms
] = 4;
728 this->userplane
[i
] = dst_reg(UNIFORM
, this->uniforms
);
729 this->userplane
[i
].type
= BRW_REGISTER_TYPE_F
;
730 for (int j
= 0; j
< 4; ++j
) {
731 stage_prog_data
->param
[this->uniforms
* 4 + j
] =
732 (gl_constant_value
*) &clip_planes
[i
][j
];
738 /* Our support for builtin uniforms is even scarier than non-builtin.
739 * It sits on top of the PROG_STATE_VAR parameters that are
740 * automatically updated from GL context state.
743 vec4_visitor::setup_builtin_uniform_values(ir_variable
*ir
)
745 const ir_state_slot
*const slots
= ir
->get_state_slots();
746 assert(slots
!= NULL
);
748 for (unsigned int i
= 0; i
< ir
->get_num_state_slots(); i
++) {
749 /* This state reference has already been setup by ir_to_mesa,
750 * but we'll get the same index back here. We can reference
751 * ParameterValues directly, since unlike brw_fs.cpp, we never
752 * add new state references during compile.
754 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
755 (gl_state_index
*)slots
[i
].tokens
);
756 gl_constant_value
*values
=
757 &this->prog
->Parameters
->ParameterValues
[index
][0];
759 assert(this->uniforms
< uniform_array_size
);
761 for (unsigned j
= 0; j
< 4; j
++)
762 stage_prog_data
->param
[this->uniforms
* 4 + j
] =
763 &values
[GET_SWZ(slots
[i
].swizzle
, j
)];
765 this->uniform_vector_size
[this->uniforms
] =
766 (ir
->type
->is_scalar() || ir
->type
->is_vector() ||
767 ir
->type
->is_matrix() ? ir
->type
->vector_elements
: 4);
774 vec4_visitor::variable_storage(ir_variable
*var
)
776 return (dst_reg
*)hash_table_find(this->variable_ht
, var
);
780 vec4_visitor::emit_bool_to_cond_code(ir_rvalue
*ir
,
781 enum brw_predicate
*predicate
)
783 ir_expression
*expr
= ir
->as_expression();
785 *predicate
= BRW_PREDICATE_NORMAL
;
787 if (expr
&& expr
->operation
!= ir_binop_ubo_load
) {
789 vec4_instruction
*inst
;
791 assert(expr
->get_num_operands() <= 3);
792 for (unsigned int i
= 0; i
< expr
->get_num_operands(); i
++) {
793 expr
->operands
[i
]->accept(this);
794 op
[i
] = this->result
;
796 resolve_ud_negate(&op
[i
]);
799 switch (expr
->operation
) {
800 case ir_unop_logic_not
:
801 inst
= emit(AND(dst_null_d(), op
[0], src_reg(1)));
802 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
805 case ir_binop_logic_xor
:
806 if (devinfo
->gen
<= 5) {
807 src_reg temp
= src_reg(this, ir
->type
);
808 emit(XOR(dst_reg(temp
), op
[0], op
[1]));
809 inst
= emit(AND(dst_null_d(), temp
, src_reg(1)));
811 inst
= emit(XOR(dst_null_d(), op
[0], op
[1]));
813 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
816 case ir_binop_logic_or
:
817 if (devinfo
->gen
<= 5) {
818 src_reg temp
= src_reg(this, ir
->type
);
819 emit(OR(dst_reg(temp
), op
[0], op
[1]));
820 inst
= emit(AND(dst_null_d(), temp
, src_reg(1)));
822 inst
= emit(OR(dst_null_d(), op
[0], op
[1]));
824 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
827 case ir_binop_logic_and
:
828 if (devinfo
->gen
<= 5) {
829 src_reg temp
= src_reg(this, ir
->type
);
830 emit(AND(dst_reg(temp
), op
[0], op
[1]));
831 inst
= emit(AND(dst_null_d(), temp
, src_reg(1)));
833 inst
= emit(AND(dst_null_d(), op
[0], op
[1]));
835 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
839 if (devinfo
->gen
>= 6) {
840 emit(CMP(dst_null_d(), op
[0], src_reg(0.0f
), BRW_CONDITIONAL_NZ
));
842 inst
= emit(MOV(dst_null_f(), op
[0]));
843 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
848 if (devinfo
->gen
>= 6) {
849 emit(CMP(dst_null_d(), op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
851 inst
= emit(MOV(dst_null_d(), op
[0]));
852 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
856 case ir_binop_all_equal
:
857 if (devinfo
->gen
<= 5) {
858 resolve_bool_comparison(expr
->operands
[0], &op
[0]);
859 resolve_bool_comparison(expr
->operands
[1], &op
[1]);
861 inst
= emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_Z
));
862 *predicate
= BRW_PREDICATE_ALIGN16_ALL4H
;
865 case ir_binop_any_nequal
:
866 if (devinfo
->gen
<= 5) {
867 resolve_bool_comparison(expr
->operands
[0], &op
[0]);
868 resolve_bool_comparison(expr
->operands
[1], &op
[1]);
870 inst
= emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_NZ
));
871 *predicate
= BRW_PREDICATE_ALIGN16_ANY4H
;
875 if (devinfo
->gen
<= 5) {
876 resolve_bool_comparison(expr
->operands
[0], &op
[0]);
878 inst
= emit(CMP(dst_null_d(), op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
879 *predicate
= BRW_PREDICATE_ALIGN16_ANY4H
;
882 case ir_binop_greater
:
883 case ir_binop_gequal
:
885 case ir_binop_lequal
:
887 case ir_binop_nequal
:
888 if (devinfo
->gen
<= 5) {
889 resolve_bool_comparison(expr
->operands
[0], &op
[0]);
890 resolve_bool_comparison(expr
->operands
[1], &op
[1]);
892 emit(CMP(dst_null_d(), op
[0], op
[1],
893 brw_conditional_for_comparison(expr
->operation
)));
896 case ir_triop_csel
: {
897 /* Expand the boolean condition into the flag register. */
898 inst
= emit(MOV(dst_null_d(), op
[0]));
899 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
901 /* Select which boolean to return. */
902 dst_reg
temp(this, expr
->operands
[1]->type
);
903 inst
= emit(BRW_OPCODE_SEL
, temp
, op
[1], op
[2]);
904 inst
->predicate
= BRW_PREDICATE_NORMAL
;
906 /* Expand the result to a condition code. */
907 inst
= emit(MOV(dst_null_d(), src_reg(temp
)));
908 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
913 unreachable("not reached");
920 resolve_ud_negate(&this->result
);
922 vec4_instruction
*inst
= emit(AND(dst_null_d(), this->result
, src_reg(1)));
923 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
927 * Emit a gen6 IF statement with the comparison folded into the IF
931 vec4_visitor::emit_if_gen6(ir_if
*ir
)
933 ir_expression
*expr
= ir
->condition
->as_expression();
935 if (expr
&& expr
->operation
!= ir_binop_ubo_load
) {
939 assert(expr
->get_num_operands() <= 3);
940 for (unsigned int i
= 0; i
< expr
->get_num_operands(); i
++) {
941 expr
->operands
[i
]->accept(this);
942 op
[i
] = this->result
;
945 switch (expr
->operation
) {
946 case ir_unop_logic_not
:
947 emit(IF(op
[0], src_reg(0), BRW_CONDITIONAL_Z
));
950 case ir_binop_logic_xor
:
951 emit(IF(op
[0], op
[1], BRW_CONDITIONAL_NZ
));
954 case ir_binop_logic_or
:
955 temp
= dst_reg(this, glsl_type::bool_type
);
956 emit(OR(temp
, op
[0], op
[1]));
957 emit(IF(src_reg(temp
), src_reg(0), BRW_CONDITIONAL_NZ
));
960 case ir_binop_logic_and
:
961 temp
= dst_reg(this, glsl_type::bool_type
);
962 emit(AND(temp
, op
[0], op
[1]));
963 emit(IF(src_reg(temp
), src_reg(0), BRW_CONDITIONAL_NZ
));
967 emit(IF(op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
971 emit(IF(op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
974 case ir_binop_greater
:
975 case ir_binop_gequal
:
977 case ir_binop_lequal
:
979 case ir_binop_nequal
:
980 emit(IF(op
[0], op
[1],
981 brw_conditional_for_comparison(expr
->operation
)));
984 case ir_binop_all_equal
:
985 emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_Z
));
986 emit(IF(BRW_PREDICATE_ALIGN16_ALL4H
));
989 case ir_binop_any_nequal
:
990 emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_NZ
));
991 emit(IF(BRW_PREDICATE_ALIGN16_ANY4H
));
995 emit(CMP(dst_null_d(), op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
996 emit(IF(BRW_PREDICATE_ALIGN16_ANY4H
));
999 case ir_triop_csel
: {
1000 /* Expand the boolean condition into the flag register. */
1001 vec4_instruction
*inst
= emit(MOV(dst_null_d(), op
[0]));
1002 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
1004 /* Select which boolean to return. */
1005 dst_reg
temp(this, expr
->operands
[1]->type
);
1006 inst
= emit(BRW_OPCODE_SEL
, temp
, op
[1], op
[2]);
1007 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1009 emit(IF(src_reg(temp
), src_reg(0), BRW_CONDITIONAL_NZ
));
1014 unreachable("not reached");
1019 ir
->condition
->accept(this);
1021 emit(IF(this->result
, src_reg(0), BRW_CONDITIONAL_NZ
));
1025 vec4_visitor::visit(ir_variable
*ir
)
1027 dst_reg
*reg
= NULL
;
1029 if (variable_storage(ir
))
1032 switch (ir
->data
.mode
) {
1033 case ir_var_shader_in
:
1034 assert(ir
->data
.location
!= -1);
1035 reg
= new(mem_ctx
) dst_reg(ATTR
, ir
->data
.location
);
1038 case ir_var_shader_out
:
1039 assert(ir
->data
.location
!= -1);
1040 reg
= new(mem_ctx
) dst_reg(this, ir
->type
);
1042 for (int i
= 0; i
< type_size(ir
->type
); i
++) {
1043 output_reg
[ir
->data
.location
+ i
] = *reg
;
1044 output_reg
[ir
->data
.location
+ i
].reg_offset
= i
;
1045 output_reg
[ir
->data
.location
+ i
].type
=
1046 brw_type_for_base_type(ir
->type
->get_scalar_type());
1047 output_reg_annotation
[ir
->data
.location
+ i
] = ir
->name
;
1052 case ir_var_temporary
:
1053 reg
= new(mem_ctx
) dst_reg(this, ir
->type
);
1056 case ir_var_uniform
:
1057 reg
= new(this->mem_ctx
) dst_reg(UNIFORM
, this->uniforms
);
1059 /* Thanks to the lower_ubo_reference pass, we will see only
1060 * ir_binop_ubo_load expressions and not ir_dereference_variable for UBO
1061 * variables, so no need for them to be in variable_ht.
1063 * Some uniforms, such as samplers and atomic counters, have no actual
1064 * storage, so we should ignore them.
1066 if (ir
->is_in_buffer_block() || type_size(ir
->type
) == 0)
1069 /* Track how big the whole uniform variable is, in case we need to put a
1070 * copy of its data into pull constants for array access.
1072 assert(this->uniforms
< uniform_array_size
);
1073 this->uniform_size
[this->uniforms
] = type_size(ir
->type
);
1075 if (!strncmp(ir
->name
, "gl_", 3)) {
1076 setup_builtin_uniform_values(ir
);
1078 setup_uniform_values(ir
);
1082 case ir_var_system_value
:
1083 reg
= make_reg_for_system_value(ir
);
1087 unreachable("not reached");
1090 reg
->type
= brw_type_for_base_type(ir
->type
);
1091 hash_table_insert(this->variable_ht
, reg
, ir
);
1095 vec4_visitor::visit(ir_loop
*ir
)
1097 /* We don't want debugging output to print the whole body of the
1098 * loop as the annotation.
1100 this->base_ir
= NULL
;
1102 emit(BRW_OPCODE_DO
);
1104 visit_instructions(&ir
->body_instructions
);
1106 emit(BRW_OPCODE_WHILE
);
1110 vec4_visitor::visit(ir_loop_jump
*ir
)
1113 case ir_loop_jump::jump_break
:
1114 emit(BRW_OPCODE_BREAK
);
1116 case ir_loop_jump::jump_continue
:
1117 emit(BRW_OPCODE_CONTINUE
);
1124 vec4_visitor::visit(ir_function_signature
*)
1126 unreachable("not reached");
1130 vec4_visitor::visit(ir_function
*ir
)
1132 /* Ignore function bodies other than main() -- we shouldn't see calls to
1133 * them since they should all be inlined.
1135 if (strcmp(ir
->name
, "main") == 0) {
1136 const ir_function_signature
*sig
;
1139 sig
= ir
->matching_signature(NULL
, &empty
, false);
1143 visit_instructions(&sig
->body
);
1148 vec4_visitor::try_emit_mad(ir_expression
*ir
)
1150 /* 3-src instructions were introduced in gen6. */
1151 if (devinfo
->gen
< 6)
1154 /* MAD can only handle floating-point data. */
1155 if (ir
->type
->base_type
!= GLSL_TYPE_FLOAT
)
1160 bool mul_negate
, mul_abs
;
1162 for (int i
= 0; i
< 2; i
++) {
1166 mul
= ir
->operands
[i
]->as_expression();
1167 nonmul
= ir
->operands
[1 - i
];
1169 if (mul
&& mul
->operation
== ir_unop_abs
) {
1170 mul
= mul
->operands
[0]->as_expression();
1172 } else if (mul
&& mul
->operation
== ir_unop_neg
) {
1173 mul
= mul
->operands
[0]->as_expression();
1177 if (mul
&& mul
->operation
== ir_binop_mul
)
1181 if (!mul
|| mul
->operation
!= ir_binop_mul
)
1184 nonmul
->accept(this);
1185 src_reg src0
= fix_3src_operand(this->result
);
1187 mul
->operands
[0]->accept(this);
1188 src_reg src1
= fix_3src_operand(this->result
);
1189 src1
.negate
^= mul_negate
;
1192 src1
.negate
= false;
1194 mul
->operands
[1]->accept(this);
1195 src_reg src2
= fix_3src_operand(this->result
);
1198 src2
.negate
= false;
1200 this->result
= src_reg(this, ir
->type
);
1201 emit(BRW_OPCODE_MAD
, dst_reg(this->result
), src0
, src1
, src2
);
1207 vec4_visitor::try_emit_b2f_of_compare(ir_expression
*ir
)
1209 /* This optimization relies on CMP setting the destination to 0 when
1210 * false. Early hardware only sets the least significant bit, and
1211 * leaves the other bits undefined. So we can't use it.
1213 if (devinfo
->gen
< 6)
1216 ir_expression
*const cmp
= ir
->operands
[0]->as_expression();
1221 switch (cmp
->operation
) {
1223 case ir_binop_greater
:
1224 case ir_binop_lequal
:
1225 case ir_binop_gequal
:
1226 case ir_binop_equal
:
1227 case ir_binop_nequal
:
1234 cmp
->operands
[0]->accept(this);
1235 const src_reg cmp_src0
= this->result
;
1237 cmp
->operands
[1]->accept(this);
1238 const src_reg cmp_src1
= this->result
;
1240 this->result
= src_reg(this, ir
->type
);
1242 emit(CMP(dst_reg(this->result
), cmp_src0
, cmp_src1
,
1243 brw_conditional_for_comparison(cmp
->operation
)));
1245 /* If the comparison is false, this->result will just happen to be zero.
1247 vec4_instruction
*const inst
= emit(BRW_OPCODE_SEL
, dst_reg(this->result
),
1248 this->result
, src_reg(1.0f
));
1249 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1250 inst
->predicate_inverse
= true;
1256 vec4_visitor::emit_minmax(enum brw_conditional_mod conditionalmod
, dst_reg dst
,
1257 src_reg src0
, src_reg src1
)
1259 vec4_instruction
*inst
;
1261 if (devinfo
->gen
>= 6) {
1262 inst
= emit(BRW_OPCODE_SEL
, dst
, src0
, src1
);
1263 inst
->conditional_mod
= conditionalmod
;
1265 emit(CMP(dst
, src0
, src1
, conditionalmod
));
1267 inst
= emit(BRW_OPCODE_SEL
, dst
, src0
, src1
);
1268 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1273 vec4_visitor::emit_lrp(const dst_reg
&dst
,
1274 const src_reg
&x
, const src_reg
&y
, const src_reg
&a
)
1276 if (devinfo
->gen
>= 6) {
1277 /* Note that the instruction's argument order is reversed from GLSL
1281 fix_3src_operand(a
), fix_3src_operand(y
), fix_3src_operand(x
)));
1283 /* Earlier generations don't support three source operations, so we
1284 * need to emit x*(1-a) + y*a.
1286 dst_reg y_times_a
= dst_reg(this, glsl_type::vec4_type
);
1287 dst_reg one_minus_a
= dst_reg(this, glsl_type::vec4_type
);
1288 dst_reg x_times_one_minus_a
= dst_reg(this, glsl_type::vec4_type
);
1289 y_times_a
.writemask
= dst
.writemask
;
1290 one_minus_a
.writemask
= dst
.writemask
;
1291 x_times_one_minus_a
.writemask
= dst
.writemask
;
1293 emit(MUL(y_times_a
, y
, a
));
1294 emit(ADD(one_minus_a
, negate(a
), src_reg(1.0f
)));
1295 emit(MUL(x_times_one_minus_a
, x
, src_reg(one_minus_a
)));
1296 emit(ADD(dst
, src_reg(x_times_one_minus_a
), src_reg(y_times_a
)));
1301 * Emits the instructions needed to perform a pull constant load. before_block
1302 * and before_inst can be NULL in which case the instruction will be appended
1303 * to the end of the instruction list.
1306 vec4_visitor::emit_pull_constant_load_reg(dst_reg dst
,
1309 bblock_t
*before_block
,
1310 vec4_instruction
*before_inst
)
1312 assert((before_inst
== NULL
&& before_block
== NULL
) ||
1313 (before_inst
&& before_block
));
1315 vec4_instruction
*pull
;
1317 if (devinfo
->gen
>= 9) {
1318 /* Gen9+ needs a message header in order to use SIMD4x2 mode */
1319 src_reg
header(this, glsl_type::uvec4_type
, 2);
1322 vec4_instruction(VS_OPCODE_SET_SIMD4X2_HEADER_GEN9
,
1326 emit_before(before_block
, before_inst
, pull
);
1330 dst_reg index_reg
= retype(offset(dst_reg(header
), 1),
1332 pull
= MOV(writemask(index_reg
, WRITEMASK_X
), offset_reg
);
1335 emit_before(before_block
, before_inst
, pull
);
1339 pull
= new(mem_ctx
) vec4_instruction(VS_OPCODE_PULL_CONSTANT_LOAD_GEN7
,
1344 pull
->header_size
= 1;
1345 } else if (devinfo
->gen
>= 7) {
1346 dst_reg grf_offset
= dst_reg(this, glsl_type::int_type
);
1348 grf_offset
.type
= offset_reg
.type
;
1350 pull
= MOV(grf_offset
, offset_reg
);
1353 emit_before(before_block
, before_inst
, pull
);
1357 pull
= new(mem_ctx
) vec4_instruction(VS_OPCODE_PULL_CONSTANT_LOAD_GEN7
,
1360 src_reg(grf_offset
));
1363 pull
= new(mem_ctx
) vec4_instruction(VS_OPCODE_PULL_CONSTANT_LOAD
,
1367 pull
->base_mrf
= 14;
1372 emit_before(before_block
, before_inst
, pull
);
1378 vec4_visitor::emit_uniformize(const dst_reg
&dst
, const src_reg
&src
)
1380 const src_reg
chan_index(this, glsl_type::uint_type
);
1382 emit(SHADER_OPCODE_FIND_LIVE_CHANNEL
, dst_reg(chan_index
))
1383 ->force_writemask_all
= true;
1384 emit(SHADER_OPCODE_BROADCAST
, dst
, src
, chan_index
)
1385 ->force_writemask_all
= true;
1389 vec4_visitor::visit(ir_expression
*ir
)
1391 unsigned int operand
;
1392 src_reg op
[ARRAY_SIZE(ir
->operands
)];
1393 vec4_instruction
*inst
;
1395 if (ir
->operation
== ir_binop_add
) {
1396 if (try_emit_mad(ir
))
1400 if (ir
->operation
== ir_unop_b2f
) {
1401 if (try_emit_b2f_of_compare(ir
))
1405 /* Storage for our result. Ideally for an assignment we'd be using
1406 * the actual storage for the result here, instead.
1408 dst_reg
result_dst(this, ir
->type
);
1409 src_reg
result_src(result_dst
);
1411 if (ir
->operation
== ir_triop_csel
) {
1412 ir
->operands
[1]->accept(this);
1413 op
[1] = this->result
;
1414 ir
->operands
[2]->accept(this);
1415 op
[2] = this->result
;
1417 enum brw_predicate predicate
;
1418 emit_bool_to_cond_code(ir
->operands
[0], &predicate
);
1419 inst
= emit(BRW_OPCODE_SEL
, result_dst
, op
[1], op
[2]);
1420 inst
->predicate
= predicate
;
1421 this->result
= result_src
;
1425 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1426 this->result
.file
= BAD_FILE
;
1427 ir
->operands
[operand
]->accept(this);
1428 if (this->result
.file
== BAD_FILE
) {
1429 fprintf(stderr
, "Failed to get tree for expression operand:\n");
1430 ir
->operands
[operand
]->fprint(stderr
);
1433 op
[operand
] = this->result
;
1435 /* Matrix expression operands should have been broken down to vector
1436 * operations already.
1438 assert(!ir
->operands
[operand
]->type
->is_matrix());
1441 /* If nothing special happens, this is the result. */
1442 this->result
= result_src
;
1444 switch (ir
->operation
) {
1445 case ir_unop_logic_not
:
1446 emit(NOT(result_dst
, op
[0]));
1449 op
[0].negate
= !op
[0].negate
;
1450 emit(MOV(result_dst
, op
[0]));
1454 op
[0].negate
= false;
1455 emit(MOV(result_dst
, op
[0]));
1459 if (ir
->type
->is_float()) {
1460 /* AND(val, 0x80000000) gives the sign bit.
1462 * Predicated OR ORs 1.0 (0x3f800000) with the sign bit if val is not
1465 emit(CMP(dst_null_f(), op
[0], src_reg(0.0f
), BRW_CONDITIONAL_NZ
));
1467 op
[0].type
= BRW_REGISTER_TYPE_UD
;
1468 result_dst
.type
= BRW_REGISTER_TYPE_UD
;
1469 emit(AND(result_dst
, op
[0], src_reg(0x80000000u
)));
1471 inst
= emit(OR(result_dst
, src_reg(result_dst
), src_reg(0x3f800000u
)));
1472 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1474 this->result
.type
= BRW_REGISTER_TYPE_F
;
1476 /* ASR(val, 31) -> negative val generates 0xffffffff (signed -1).
1477 * -> non-negative val generates 0x00000000.
1478 * Predicated OR sets 1 if val is positive.
1480 emit(CMP(dst_null_d(), op
[0], src_reg(0), BRW_CONDITIONAL_G
));
1482 emit(ASR(result_dst
, op
[0], src_reg(31)));
1484 inst
= emit(OR(result_dst
, src_reg(result_dst
), src_reg(1)));
1485 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1490 emit_math(SHADER_OPCODE_RCP
, result_dst
, op
[0]);
1494 emit_math(SHADER_OPCODE_EXP2
, result_dst
, op
[0]);
1497 emit_math(SHADER_OPCODE_LOG2
, result_dst
, op
[0]);
1501 unreachable("not reached: should be handled by ir_explog_to_explog2");
1503 emit_math(SHADER_OPCODE_SIN
, result_dst
, op
[0]);
1506 emit_math(SHADER_OPCODE_COS
, result_dst
, op
[0]);
1510 case ir_unop_dFdx_coarse
:
1511 case ir_unop_dFdx_fine
:
1513 case ir_unop_dFdy_coarse
:
1514 case ir_unop_dFdy_fine
:
1515 unreachable("derivatives not valid in vertex shader");
1517 case ir_unop_bitfield_reverse
:
1518 emit(BFREV(result_dst
, op
[0]));
1520 case ir_unop_bit_count
:
1521 emit(CBIT(result_dst
, op
[0]));
1523 case ir_unop_find_msb
: {
1524 src_reg temp
= src_reg(this, glsl_type::uint_type
);
1526 inst
= emit(FBH(dst_reg(temp
), op
[0]));
1527 inst
->dst
.writemask
= WRITEMASK_XYZW
;
1529 /* FBH counts from the MSB side, while GLSL's findMSB() wants the count
1530 * from the LSB side. If FBH didn't return an error (0xFFFFFFFF), then
1531 * subtract the result from 31 to convert the MSB count into an LSB count.
1534 /* FBH only supports UD type for dst, so use a MOV to convert UD to D. */
1535 temp
.swizzle
= BRW_SWIZZLE_NOOP
;
1536 emit(MOV(result_dst
, temp
));
1538 src_reg src_tmp
= src_reg(result_dst
);
1539 emit(CMP(dst_null_d(), src_tmp
, src_reg(-1), BRW_CONDITIONAL_NZ
));
1541 src_tmp
.negate
= true;
1542 inst
= emit(ADD(result_dst
, src_tmp
, src_reg(31)));
1543 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1546 case ir_unop_find_lsb
:
1547 emit(FBL(result_dst
, op
[0]));
1549 case ir_unop_saturate
:
1550 inst
= emit(MOV(result_dst
, op
[0]));
1551 inst
->saturate
= true;
1555 unreachable("not reached: should be handled by lower_noise");
1558 emit(ADD(result_dst
, op
[0], op
[1]));
1561 unreachable("not reached: should be handled by ir_sub_to_add_neg");
1564 if (devinfo
->gen
< 8 && ir
->type
->is_integer()) {
1565 /* For integer multiplication, the MUL uses the low 16 bits of one of
1566 * the operands (src0 through SNB, src1 on IVB and later). The MACH
1567 * accumulates in the contribution of the upper 16 bits of that
1568 * operand. If we can determine that one of the args is in the low
1569 * 16 bits, though, we can just emit a single MUL.
1571 if (ir
->operands
[0]->is_uint16_constant()) {
1572 if (devinfo
->gen
< 7)
1573 emit(MUL(result_dst
, op
[0], op
[1]));
1575 emit(MUL(result_dst
, op
[1], op
[0]));
1576 } else if (ir
->operands
[1]->is_uint16_constant()) {
1577 if (devinfo
->gen
< 7)
1578 emit(MUL(result_dst
, op
[1], op
[0]));
1580 emit(MUL(result_dst
, op
[0], op
[1]));
1582 struct brw_reg acc
= retype(brw_acc_reg(8), result_dst
.type
);
1584 emit(MUL(acc
, op
[0], op
[1]));
1585 emit(MACH(dst_null_d(), op
[0], op
[1]));
1586 emit(MOV(result_dst
, src_reg(acc
)));
1589 emit(MUL(result_dst
, op
[0], op
[1]));
1592 case ir_binop_imul_high
: {
1593 struct brw_reg acc
= retype(brw_acc_reg(8), result_dst
.type
);
1595 emit(MUL(acc
, op
[0], op
[1]));
1596 emit(MACH(result_dst
, op
[0], op
[1]));
1600 /* Floating point should be lowered by DIV_TO_MUL_RCP in the compiler. */
1601 assert(ir
->type
->is_integer());
1602 emit_math(SHADER_OPCODE_INT_QUOTIENT
, result_dst
, op
[0], op
[1]);
1605 case ir_binop_carry
:
1606 unreachable("Should have been lowered by carry_to_arith().");
1608 case ir_binop_borrow
:
1609 unreachable("Should have been lowered by borrow_to_arith().");
1612 /* Floating point should be lowered by MOD_TO_FLOOR in the compiler. */
1613 assert(ir
->type
->is_integer());
1614 emit_math(SHADER_OPCODE_INT_REMAINDER
, result_dst
, op
[0], op
[1]);
1618 case ir_binop_greater
:
1619 case ir_binop_lequal
:
1620 case ir_binop_gequal
:
1621 case ir_binop_equal
:
1622 case ir_binop_nequal
: {
1623 if (devinfo
->gen
<= 5) {
1624 resolve_bool_comparison(ir
->operands
[0], &op
[0]);
1625 resolve_bool_comparison(ir
->operands
[1], &op
[1]);
1627 emit(CMP(result_dst
, op
[0], op
[1],
1628 brw_conditional_for_comparison(ir
->operation
)));
1632 case ir_binop_all_equal
:
1633 if (devinfo
->gen
<= 5) {
1634 resolve_bool_comparison(ir
->operands
[0], &op
[0]);
1635 resolve_bool_comparison(ir
->operands
[1], &op
[1]);
1638 /* "==" operator producing a scalar boolean. */
1639 if (ir
->operands
[0]->type
->is_vector() ||
1640 ir
->operands
[1]->type
->is_vector()) {
1641 emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_Z
));
1642 emit(MOV(result_dst
, src_reg(0)));
1643 inst
= emit(MOV(result_dst
, src_reg(~0)));
1644 inst
->predicate
= BRW_PREDICATE_ALIGN16_ALL4H
;
1646 emit(CMP(result_dst
, op
[0], op
[1], BRW_CONDITIONAL_Z
));
1649 case ir_binop_any_nequal
:
1650 if (devinfo
->gen
<= 5) {
1651 resolve_bool_comparison(ir
->operands
[0], &op
[0]);
1652 resolve_bool_comparison(ir
->operands
[1], &op
[1]);
1655 /* "!=" operator producing a scalar boolean. */
1656 if (ir
->operands
[0]->type
->is_vector() ||
1657 ir
->operands
[1]->type
->is_vector()) {
1658 emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_NZ
));
1660 emit(MOV(result_dst
, src_reg(0)));
1661 inst
= emit(MOV(result_dst
, src_reg(~0)));
1662 inst
->predicate
= BRW_PREDICATE_ALIGN16_ANY4H
;
1664 emit(CMP(result_dst
, op
[0], op
[1], BRW_CONDITIONAL_NZ
));
1669 if (devinfo
->gen
<= 5) {
1670 resolve_bool_comparison(ir
->operands
[0], &op
[0]);
1672 emit(CMP(dst_null_d(), op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
1673 emit(MOV(result_dst
, src_reg(0)));
1675 inst
= emit(MOV(result_dst
, src_reg(~0)));
1676 inst
->predicate
= BRW_PREDICATE_ALIGN16_ANY4H
;
1679 case ir_binop_logic_xor
:
1680 emit(XOR(result_dst
, op
[0], op
[1]));
1683 case ir_binop_logic_or
:
1684 emit(OR(result_dst
, op
[0], op
[1]));
1687 case ir_binop_logic_and
:
1688 emit(AND(result_dst
, op
[0], op
[1]));
1692 assert(ir
->operands
[0]->type
->is_vector());
1693 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1694 emit_dp(result_dst
, op
[0], op
[1], ir
->operands
[0]->type
->vector_elements
);
1698 emit_math(SHADER_OPCODE_SQRT
, result_dst
, op
[0]);
1701 emit_math(SHADER_OPCODE_RSQ
, result_dst
, op
[0]);
1704 case ir_unop_bitcast_i2f
:
1705 case ir_unop_bitcast_u2f
:
1706 this->result
= op
[0];
1707 this->result
.type
= BRW_REGISTER_TYPE_F
;
1710 case ir_unop_bitcast_f2i
:
1711 this->result
= op
[0];
1712 this->result
.type
= BRW_REGISTER_TYPE_D
;
1715 case ir_unop_bitcast_f2u
:
1716 this->result
= op
[0];
1717 this->result
.type
= BRW_REGISTER_TYPE_UD
;
1726 emit(MOV(result_dst
, op
[0]));
1730 if (devinfo
->gen
<= 5) {
1731 resolve_bool_comparison(ir
->operands
[0], &op
[0]);
1733 emit(MOV(result_dst
, negate(op
[0])));
1736 emit(CMP(result_dst
, op
[0], src_reg(0.0f
), BRW_CONDITIONAL_NZ
));
1739 emit(CMP(result_dst
, op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
1743 emit(RNDZ(result_dst
, op
[0]));
1745 case ir_unop_ceil
: {
1746 src_reg tmp
= src_reg(this, ir
->type
);
1747 op
[0].negate
= !op
[0].negate
;
1748 emit(RNDD(dst_reg(tmp
), op
[0]));
1750 emit(MOV(result_dst
, tmp
));
1754 inst
= emit(RNDD(result_dst
, op
[0]));
1757 inst
= emit(FRC(result_dst
, op
[0]));
1759 case ir_unop_round_even
:
1760 emit(RNDE(result_dst
, op
[0]));
1764 emit_minmax(BRW_CONDITIONAL_L
, result_dst
, op
[0], op
[1]);
1767 emit_minmax(BRW_CONDITIONAL_GE
, result_dst
, op
[0], op
[1]);
1771 emit_math(SHADER_OPCODE_POW
, result_dst
, op
[0], op
[1]);
1774 case ir_unop_bit_not
:
1775 inst
= emit(NOT(result_dst
, op
[0]));
1777 case ir_binop_bit_and
:
1778 inst
= emit(AND(result_dst
, op
[0], op
[1]));
1780 case ir_binop_bit_xor
:
1781 inst
= emit(XOR(result_dst
, op
[0], op
[1]));
1783 case ir_binop_bit_or
:
1784 inst
= emit(OR(result_dst
, op
[0], op
[1]));
1787 case ir_binop_lshift
:
1788 inst
= emit(SHL(result_dst
, op
[0], op
[1]));
1791 case ir_binop_rshift
:
1792 if (ir
->type
->base_type
== GLSL_TYPE_INT
)
1793 inst
= emit(ASR(result_dst
, op
[0], op
[1]));
1795 inst
= emit(SHR(result_dst
, op
[0], op
[1]));
1799 emit(BFI1(result_dst
, op
[0], op
[1]));
1802 case ir_binop_ubo_load
: {
1803 ir_constant
*const_uniform_block
= ir
->operands
[0]->as_constant();
1804 ir_constant
*const_offset_ir
= ir
->operands
[1]->as_constant();
1805 unsigned const_offset
= const_offset_ir
? const_offset_ir
->value
.u
[0] : 0;
1808 /* Now, load the vector from that offset. */
1809 assert(ir
->type
->is_vector() || ir
->type
->is_scalar());
1811 src_reg packed_consts
= src_reg(this, glsl_type::vec4_type
);
1812 packed_consts
.type
= result
.type
;
1815 if (const_uniform_block
) {
1816 /* The block index is a constant, so just emit the binding table entry
1819 surf_index
= src_reg(prog_data
->base
.binding_table
.ubo_start
+
1820 const_uniform_block
->value
.u
[0]);
1822 /* The block index is not a constant. Evaluate the index expression
1823 * per-channel and add the base UBO index; we have to select a value
1824 * from any live channel.
1826 surf_index
= src_reg(this, glsl_type::uint_type
);
1827 emit(ADD(dst_reg(surf_index
), op
[0],
1828 src_reg(prog_data
->base
.binding_table
.ubo_start
)));
1829 emit_uniformize(dst_reg(surf_index
), surf_index
);
1831 /* Assume this may touch any UBO. It would be nice to provide
1832 * a tighter bound, but the array information is already lowered away.
1834 brw_mark_surface_used(&prog_data
->base
,
1835 prog_data
->base
.binding_table
.ubo_start
+
1836 shader_prog
->NumUniformBlocks
- 1);
1839 if (const_offset_ir
) {
1840 if (devinfo
->gen
>= 8) {
1841 /* Store the offset in a GRF so we can send-from-GRF. */
1842 offset
= src_reg(this, glsl_type::int_type
);
1843 emit(MOV(dst_reg(offset
), src_reg(const_offset
/ 16)));
1845 /* Immediates are fine on older generations since they'll be moved
1846 * to a (potentially fake) MRF at the generator level.
1848 offset
= src_reg(const_offset
/ 16);
1851 offset
= src_reg(this, glsl_type::uint_type
);
1852 emit(SHR(dst_reg(offset
), op
[1], src_reg(4u)));
1855 emit_pull_constant_load_reg(dst_reg(packed_consts
),
1858 NULL
, NULL
/* before_block/inst */);
1860 packed_consts
.swizzle
= brw_swizzle_for_size(ir
->type
->vector_elements
);
1861 packed_consts
.swizzle
+= BRW_SWIZZLE4(const_offset
% 16 / 4,
1862 const_offset
% 16 / 4,
1863 const_offset
% 16 / 4,
1864 const_offset
% 16 / 4);
1866 /* UBO bools are any nonzero int. We need to convert them to 0/~0. */
1867 if (ir
->type
->base_type
== GLSL_TYPE_BOOL
) {
1868 emit(CMP(result_dst
, packed_consts
, src_reg(0u),
1869 BRW_CONDITIONAL_NZ
));
1871 emit(MOV(result_dst
, packed_consts
));
1876 case ir_binop_vector_extract
:
1877 unreachable("should have been lowered by vec_index_to_cond_assign");
1880 op
[0] = fix_3src_operand(op
[0]);
1881 op
[1] = fix_3src_operand(op
[1]);
1882 op
[2] = fix_3src_operand(op
[2]);
1883 /* Note that the instruction's argument order is reversed from GLSL
1886 emit(MAD(result_dst
, op
[2], op
[1], op
[0]));
1890 emit_lrp(result_dst
, op
[0], op
[1], op
[2]);
1894 unreachable("already handled above");
1898 op
[0] = fix_3src_operand(op
[0]);
1899 op
[1] = fix_3src_operand(op
[1]);
1900 op
[2] = fix_3src_operand(op
[2]);
1901 emit(BFI2(result_dst
, op
[0], op
[1], op
[2]));
1904 case ir_triop_bitfield_extract
:
1905 op
[0] = fix_3src_operand(op
[0]);
1906 op
[1] = fix_3src_operand(op
[1]);
1907 op
[2] = fix_3src_operand(op
[2]);
1908 /* Note that the instruction's argument order is reversed from GLSL
1911 emit(BFE(result_dst
, op
[2], op
[1], op
[0]));
1914 case ir_triop_vector_insert
:
1915 unreachable("should have been lowered by lower_vector_insert");
1917 case ir_quadop_bitfield_insert
:
1918 unreachable("not reached: should be handled by "
1919 "bitfield_insert_to_bfm_bfi\n");
1921 case ir_quadop_vector
:
1922 unreachable("not reached: should be handled by lower_quadop_vector");
1924 case ir_unop_pack_half_2x16
:
1925 emit_pack_half_2x16(result_dst
, op
[0]);
1927 case ir_unop_unpack_half_2x16
:
1928 emit_unpack_half_2x16(result_dst
, op
[0]);
1930 case ir_unop_unpack_unorm_4x8
:
1931 emit_unpack_unorm_4x8(result_dst
, op
[0]);
1933 case ir_unop_unpack_snorm_4x8
:
1934 emit_unpack_snorm_4x8(result_dst
, op
[0]);
1936 case ir_unop_pack_unorm_4x8
:
1937 emit_pack_unorm_4x8(result_dst
, op
[0]);
1939 case ir_unop_pack_snorm_4x8
:
1940 emit_pack_snorm_4x8(result_dst
, op
[0]);
1942 case ir_unop_pack_snorm_2x16
:
1943 case ir_unop_pack_unorm_2x16
:
1944 case ir_unop_unpack_snorm_2x16
:
1945 case ir_unop_unpack_unorm_2x16
:
1946 unreachable("not reached: should be handled by lower_packing_builtins");
1947 case ir_unop_unpack_half_2x16_split_x
:
1948 case ir_unop_unpack_half_2x16_split_y
:
1949 case ir_binop_pack_half_2x16_split
:
1950 case ir_unop_interpolate_at_centroid
:
1951 case ir_binop_interpolate_at_sample
:
1952 case ir_binop_interpolate_at_offset
:
1953 unreachable("not reached: should not occur in vertex shader");
1954 case ir_binop_ldexp
:
1955 unreachable("not reached: should be handled by ldexp_to_arith()");
1963 case ir_unop_pack_double_2x32
:
1964 case ir_unop_unpack_double_2x32
:
1965 case ir_unop_frexp_sig
:
1966 case ir_unop_frexp_exp
:
1967 unreachable("fp64 todo");
1973 vec4_visitor::visit(ir_swizzle
*ir
)
1975 /* Note that this is only swizzles in expressions, not those on the left
1976 * hand side of an assignment, which do write masking. See ir_assignment
1979 const unsigned swz
= brw_compose_swizzle(
1980 brw_swizzle_for_size(ir
->type
->vector_elements
),
1981 BRW_SWIZZLE4(ir
->mask
.x
, ir
->mask
.y
, ir
->mask
.z
, ir
->mask
.w
));
1983 ir
->val
->accept(this);
1984 this->result
= swizzle(this->result
, swz
);
1988 vec4_visitor::visit(ir_dereference_variable
*ir
)
1990 const struct glsl_type
*type
= ir
->type
;
1991 dst_reg
*reg
= variable_storage(ir
->var
);
1994 fail("Failed to find variable storage for %s\n", ir
->var
->name
);
1995 this->result
= src_reg(brw_null_reg());
1999 this->result
= src_reg(*reg
);
2001 /* System values get their swizzle from the dst_reg writemask */
2002 if (ir
->var
->data
.mode
== ir_var_system_value
)
2005 if (type
->is_scalar() || type
->is_vector() || type
->is_matrix())
2006 this->result
.swizzle
= brw_swizzle_for_size(type
->vector_elements
);
2011 vec4_visitor::compute_array_stride(ir_dereference_array
*ir
)
2013 /* Under normal circumstances array elements are stored consecutively, so
2014 * the stride is equal to the size of the array element.
2016 return type_size(ir
->type
);
2021 vec4_visitor::visit(ir_dereference_array
*ir
)
2023 ir_constant
*constant_index
;
2025 int array_stride
= compute_array_stride(ir
);
2027 constant_index
= ir
->array_index
->constant_expression_value();
2029 ir
->array
->accept(this);
2032 if (constant_index
) {
2033 src
.reg_offset
+= constant_index
->value
.i
[0] * array_stride
;
2035 /* Variable index array dereference. It eats the "vec4" of the
2036 * base of the array and an index that offsets the Mesa register
2039 ir
->array_index
->accept(this);
2043 if (array_stride
== 1) {
2044 index_reg
= this->result
;
2046 index_reg
= src_reg(this, glsl_type::int_type
);
2048 emit(MUL(dst_reg(index_reg
), this->result
, src_reg(array_stride
)));
2052 src_reg temp
= src_reg(this, glsl_type::int_type
);
2054 emit(ADD(dst_reg(temp
), *src
.reladdr
, index_reg
));
2059 src
.reladdr
= ralloc(mem_ctx
, src_reg
);
2060 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
2063 /* If the type is smaller than a vec4, replicate the last channel out. */
2064 if (ir
->type
->is_scalar() || ir
->type
->is_vector() || ir
->type
->is_matrix())
2065 src
.swizzle
= brw_swizzle_for_size(ir
->type
->vector_elements
);
2067 src
.swizzle
= BRW_SWIZZLE_NOOP
;
2068 src
.type
= brw_type_for_base_type(ir
->type
);
2074 vec4_visitor::visit(ir_dereference_record
*ir
)
2077 const glsl_type
*struct_type
= ir
->record
->type
;
2080 ir
->record
->accept(this);
2082 for (i
= 0; i
< struct_type
->length
; i
++) {
2083 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
2085 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
2088 /* If the type is smaller than a vec4, replicate the last channel out. */
2089 if (ir
->type
->is_scalar() || ir
->type
->is_vector() || ir
->type
->is_matrix())
2090 this->result
.swizzle
= brw_swizzle_for_size(ir
->type
->vector_elements
);
2092 this->result
.swizzle
= BRW_SWIZZLE_NOOP
;
2093 this->result
.type
= brw_type_for_base_type(ir
->type
);
2095 this->result
.reg_offset
+= offset
;
2099 * We want to be careful in assignment setup to hit the actual storage
2100 * instead of potentially using a temporary like we might with the
2101 * ir_dereference handler.
2104 get_assignment_lhs(ir_dereference
*ir
, vec4_visitor
*v
)
2106 /* The LHS must be a dereference. If the LHS is a variable indexed array
2107 * access of a vector, it must be separated into a series conditional moves
2108 * before reaching this point (see ir_vec_index_to_cond_assign).
2110 assert(ir
->as_dereference());
2111 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
2113 assert(!deref_array
->array
->type
->is_vector());
2116 /* Use the rvalue deref handler for the most part. We'll ignore
2117 * swizzles in it and write swizzles using writemask, though.
2120 return dst_reg(v
->result
);
2124 vec4_visitor::emit_block_move(dst_reg
*dst
, src_reg
*src
,
2125 const struct glsl_type
*type
,
2126 enum brw_predicate predicate
)
2128 if (type
->base_type
== GLSL_TYPE_STRUCT
) {
2129 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2130 emit_block_move(dst
, src
, type
->fields
.structure
[i
].type
, predicate
);
2135 if (type
->is_array()) {
2136 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2137 emit_block_move(dst
, src
, type
->fields
.array
, predicate
);
2142 if (type
->is_matrix()) {
2143 const struct glsl_type
*vec_type
;
2145 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
2146 type
->vector_elements
, 1);
2148 for (int i
= 0; i
< type
->matrix_columns
; i
++) {
2149 emit_block_move(dst
, src
, vec_type
, predicate
);
2154 assert(type
->is_scalar() || type
->is_vector());
2156 dst
->type
= brw_type_for_base_type(type
);
2157 src
->type
= dst
->type
;
2159 dst
->writemask
= (1 << type
->vector_elements
) - 1;
2161 src
->swizzle
= brw_swizzle_for_size(type
->vector_elements
);
2163 vec4_instruction
*inst
= emit(MOV(*dst
, *src
));
2164 inst
->predicate
= predicate
;
2171 /* If the RHS processing resulted in an instruction generating a
2172 * temporary value, and it would be easy to rewrite the instruction to
2173 * generate its result right into the LHS instead, do so. This ends
2174 * up reliably removing instructions where it can be tricky to do so
2175 * later without real UD chain information.
2178 vec4_visitor::try_rewrite_rhs_to_dst(ir_assignment
*ir
,
2181 vec4_instruction
*pre_rhs_inst
,
2182 vec4_instruction
*last_rhs_inst
)
2184 /* This could be supported, but it would take more smarts. */
2188 if (pre_rhs_inst
== last_rhs_inst
)
2189 return false; /* No instructions generated to work with. */
2191 /* Make sure the last instruction generated our source reg. */
2192 if (src
.file
!= GRF
||
2193 src
.file
!= last_rhs_inst
->dst
.file
||
2194 src
.reg
!= last_rhs_inst
->dst
.reg
||
2195 src
.reg_offset
!= last_rhs_inst
->dst
.reg_offset
||
2199 last_rhs_inst
->predicate
!= BRW_PREDICATE_NONE
)
2202 /* Check that that last instruction fully initialized the channels
2203 * we want to use, in the order we want to use them. We could
2204 * potentially reswizzle the operands of many instructions so that
2205 * we could handle out of order channels, but don't yet.
2208 for (unsigned i
= 0; i
< 4; i
++) {
2209 if (dst
.writemask
& (1 << i
)) {
2210 if (!(last_rhs_inst
->dst
.writemask
& (1 << i
)))
2213 if (BRW_GET_SWZ(src
.swizzle
, i
) != i
)
2218 /* Success! Rewrite the instruction. */
2219 last_rhs_inst
->dst
.file
= dst
.file
;
2220 last_rhs_inst
->dst
.reg
= dst
.reg
;
2221 last_rhs_inst
->dst
.reg_offset
= dst
.reg_offset
;
2222 last_rhs_inst
->dst
.reladdr
= dst
.reladdr
;
2223 last_rhs_inst
->dst
.writemask
&= dst
.writemask
;
2229 vec4_visitor::visit(ir_assignment
*ir
)
2231 dst_reg dst
= get_assignment_lhs(ir
->lhs
, this);
2232 enum brw_predicate predicate
= BRW_PREDICATE_NONE
;
2234 if (!ir
->lhs
->type
->is_scalar() &&
2235 !ir
->lhs
->type
->is_vector()) {
2236 ir
->rhs
->accept(this);
2237 src_reg src
= this->result
;
2239 if (ir
->condition
) {
2240 emit_bool_to_cond_code(ir
->condition
, &predicate
);
2243 /* emit_block_move doesn't account for swizzles in the source register.
2244 * This should be ok, since the source register is a structure or an
2245 * array, and those can't be swizzled. But double-check to be sure.
2247 assert(src
.swizzle
==
2248 (ir
->rhs
->type
->is_matrix()
2249 ? brw_swizzle_for_size(ir
->rhs
->type
->vector_elements
)
2250 : BRW_SWIZZLE_NOOP
));
2252 emit_block_move(&dst
, &src
, ir
->rhs
->type
, predicate
);
2256 /* Now we're down to just a scalar/vector with writemasks. */
2259 vec4_instruction
*pre_rhs_inst
, *last_rhs_inst
;
2260 pre_rhs_inst
= (vec4_instruction
*)this->instructions
.get_tail();
2262 ir
->rhs
->accept(this);
2264 last_rhs_inst
= (vec4_instruction
*)this->instructions
.get_tail();
2269 assert(ir
->lhs
->type
->is_vector() ||
2270 ir
->lhs
->type
->is_scalar());
2271 dst
.writemask
= ir
->write_mask
;
2273 /* Swizzle a small RHS vector into the channels being written.
2275 * glsl ir treats write_mask as dictating how many channels are
2276 * present on the RHS while in our instructions we need to make
2277 * those channels appear in the slots of the vec4 they're written to.
2279 for (int i
= 0; i
< 4; i
++)
2280 swizzles
[i
] = (ir
->write_mask
& (1 << i
) ? src_chan
++ : 0);
2282 src_reg src
= swizzle(this->result
,
2283 BRW_SWIZZLE4(swizzles
[0], swizzles
[1],
2284 swizzles
[2], swizzles
[3]));
2286 if (try_rewrite_rhs_to_dst(ir
, dst
, src
, pre_rhs_inst
, last_rhs_inst
)) {
2290 if (ir
->condition
) {
2291 emit_bool_to_cond_code(ir
->condition
, &predicate
);
2294 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2295 vec4_instruction
*inst
= emit(MOV(dst
, src
));
2296 inst
->predicate
= predicate
;
2304 vec4_visitor::emit_constant_values(dst_reg
*dst
, ir_constant
*ir
)
2306 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
2307 foreach_in_list(ir_constant
, field_value
, &ir
->components
) {
2308 emit_constant_values(dst
, field_value
);
2313 if (ir
->type
->is_array()) {
2314 for (unsigned int i
= 0; i
< ir
->type
->length
; i
++) {
2315 emit_constant_values(dst
, ir
->array_elements
[i
]);
2320 if (ir
->type
->is_matrix()) {
2321 for (int i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
2322 float *vec
= &ir
->value
.f
[i
* ir
->type
->vector_elements
];
2324 for (int j
= 0; j
< ir
->type
->vector_elements
; j
++) {
2325 dst
->writemask
= 1 << j
;
2326 dst
->type
= BRW_REGISTER_TYPE_F
;
2328 emit(MOV(*dst
, src_reg(vec
[j
])));
2335 int remaining_writemask
= (1 << ir
->type
->vector_elements
) - 1;
2337 for (int i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2338 if (!(remaining_writemask
& (1 << i
)))
2341 dst
->writemask
= 1 << i
;
2342 dst
->type
= brw_type_for_base_type(ir
->type
);
2344 /* Find other components that match the one we're about to
2345 * write. Emits fewer instructions for things like vec4(0.5,
2348 for (int j
= i
+ 1; j
< ir
->type
->vector_elements
; j
++) {
2349 if (ir
->type
->base_type
== GLSL_TYPE_BOOL
) {
2350 if (ir
->value
.b
[i
] == ir
->value
.b
[j
])
2351 dst
->writemask
|= (1 << j
);
2353 /* u, i, and f storage all line up, so no need for a
2354 * switch case for comparing each type.
2356 if (ir
->value
.u
[i
] == ir
->value
.u
[j
])
2357 dst
->writemask
|= (1 << j
);
2361 switch (ir
->type
->base_type
) {
2362 case GLSL_TYPE_FLOAT
:
2363 emit(MOV(*dst
, src_reg(ir
->value
.f
[i
])));
2366 emit(MOV(*dst
, src_reg(ir
->value
.i
[i
])));
2368 case GLSL_TYPE_UINT
:
2369 emit(MOV(*dst
, src_reg(ir
->value
.u
[i
])));
2371 case GLSL_TYPE_BOOL
:
2372 emit(MOV(*dst
, src_reg(ir
->value
.b
[i
] != 0 ? ~0 : 0)));
2375 unreachable("Non-float/uint/int/bool constant");
2378 remaining_writemask
&= ~dst
->writemask
;
2384 vec4_visitor::visit(ir_constant
*ir
)
2386 dst_reg dst
= dst_reg(this, ir
->type
);
2387 this->result
= src_reg(dst
);
2389 emit_constant_values(&dst
, ir
);
2393 vec4_visitor::visit_atomic_counter_intrinsic(ir_call
*ir
)
2395 ir_dereference
*deref
= static_cast<ir_dereference
*>(
2396 ir
->actual_parameters
.get_head());
2397 ir_variable
*location
= deref
->variable_referenced();
2398 unsigned surf_index
= (prog_data
->base
.binding_table
.abo_start
+
2399 location
->data
.binding
);
2401 /* Calculate the surface offset */
2402 src_reg
offset(this, glsl_type::uint_type
);
2403 ir_dereference_array
*deref_array
= deref
->as_dereference_array();
2405 deref_array
->array_index
->accept(this);
2407 src_reg
tmp(this, glsl_type::uint_type
);
2408 emit(MUL(dst_reg(tmp
), this->result
, ATOMIC_COUNTER_SIZE
));
2409 emit(ADD(dst_reg(offset
), tmp
, location
->data
.atomic
.offset
));
2411 offset
= location
->data
.atomic
.offset
;
2414 /* Emit the appropriate machine instruction */
2415 const char *callee
= ir
->callee
->function_name();
2416 dst_reg dst
= get_assignment_lhs(ir
->return_deref
, this);
2418 if (!strcmp("__intrinsic_atomic_read", callee
)) {
2419 emit_untyped_surface_read(surf_index
, dst
, offset
);
2421 } else if (!strcmp("__intrinsic_atomic_increment", callee
)) {
2422 emit_untyped_atomic(BRW_AOP_INC
, surf_index
, dst
, offset
,
2423 src_reg(), src_reg());
2425 } else if (!strcmp("__intrinsic_atomic_predecrement", callee
)) {
2426 emit_untyped_atomic(BRW_AOP_PREDEC
, surf_index
, dst
, offset
,
2427 src_reg(), src_reg());
2432 vec4_visitor::visit(ir_call
*ir
)
2434 const char *callee
= ir
->callee
->function_name();
2436 if (!strcmp("__intrinsic_atomic_read", callee
) ||
2437 !strcmp("__intrinsic_atomic_increment", callee
) ||
2438 !strcmp("__intrinsic_atomic_predecrement", callee
)) {
2439 visit_atomic_counter_intrinsic(ir
);
2441 unreachable("Unsupported intrinsic.");
2446 vec4_visitor::emit_mcs_fetch(ir_texture
*ir
, src_reg coordinate
, src_reg sampler
)
2448 vec4_instruction
*inst
=
2449 new(mem_ctx
) vec4_instruction(SHADER_OPCODE_TXF_MCS
,
2450 dst_reg(this, glsl_type::uvec4_type
));
2452 inst
->src
[1] = sampler
;
2456 if (devinfo
->gen
>= 9) {
2457 /* Gen9+ needs a message header in order to use SIMD4x2 mode */
2458 vec4_instruction
*header_inst
= new(mem_ctx
)
2459 vec4_instruction(VS_OPCODE_SET_SIMD4X2_HEADER_GEN9
,
2460 dst_reg(MRF
, inst
->base_mrf
));
2465 inst
->header_size
= 1;
2466 param_base
= inst
->base_mrf
+ 1;
2469 param_base
= inst
->base_mrf
;
2472 /* parameters are: u, v, r, lod; lod will always be zero due to api restrictions */
2473 int coord_mask
= (1 << ir
->coordinate
->type
->vector_elements
) - 1;
2474 int zero_mask
= 0xf & ~coord_mask
;
2476 emit(MOV(dst_reg(MRF
, param_base
, ir
->coordinate
->type
, coord_mask
),
2479 emit(MOV(dst_reg(MRF
, param_base
, ir
->coordinate
->type
, zero_mask
),
2483 return src_reg(inst
->dst
);
2487 is_high_sampler(const struct brw_device_info
*devinfo
, src_reg sampler
)
2489 if (devinfo
->gen
< 8 && !devinfo
->is_haswell
)
2492 return sampler
.file
!= IMM
|| sampler
.fixed_hw_reg
.dw1
.ud
>= 16;
2496 vec4_visitor::visit(ir_texture
*ir
)
2499 _mesa_get_sampler_uniform_value(ir
->sampler
, shader_prog
, prog
);
2501 ir_rvalue
*nonconst_sampler_index
=
2502 _mesa_get_sampler_array_nonconst_index(ir
->sampler
);
2504 /* Handle non-constant sampler array indexing */
2505 src_reg sampler_reg
;
2506 if (nonconst_sampler_index
) {
2507 /* The highest sampler which may be used by this operation is
2508 * the last element of the array. Mark it here, because the generator
2509 * doesn't have enough information to determine the bound.
2511 uint32_t array_size
= ir
->sampler
->as_dereference_array()
2512 ->array
->type
->array_size();
2514 uint32_t max_used
= sampler
+ array_size
- 1;
2515 if (ir
->op
== ir_tg4
&& devinfo
->gen
< 8) {
2516 max_used
+= prog_data
->base
.binding_table
.gather_texture_start
;
2518 max_used
+= prog_data
->base
.binding_table
.texture_start
;
2521 brw_mark_surface_used(&prog_data
->base
, max_used
);
2523 /* Emit code to evaluate the actual indexing expression */
2524 nonconst_sampler_index
->accept(this);
2525 dst_reg
temp(this, glsl_type::uint_type
);
2526 emit(ADD(temp
, this->result
, src_reg(sampler
)));
2527 emit_uniformize(temp
, src_reg(temp
));
2529 sampler_reg
= src_reg(temp
);
2531 /* Single sampler, or constant array index; the indexing expression
2532 * is just an immediate.
2534 sampler_reg
= src_reg(sampler
);
2537 /* When tg4 is used with the degenerate ZERO/ONE swizzles, don't bother
2538 * emitting anything other than setting up the constant result.
2540 if (ir
->op
== ir_tg4
) {
2541 ir_constant
*chan
= ir
->lod_info
.component
->as_constant();
2542 int swiz
= GET_SWZ(key
->tex
.swizzles
[sampler
], chan
->value
.i
[0]);
2543 if (swiz
== SWIZZLE_ZERO
|| swiz
== SWIZZLE_ONE
) {
2544 dst_reg
result(this, ir
->type
);
2545 this->result
= src_reg(result
);
2546 emit(MOV(result
, src_reg(swiz
== SWIZZLE_ONE
? 1.0f
: 0.0f
)));
2551 /* Should be lowered by do_lower_texture_projection */
2552 assert(!ir
->projector
);
2554 /* Should be lowered */
2555 assert(!ir
->offset
|| !ir
->offset
->type
->is_array());
2557 /* Generate code to compute all the subexpression trees. This has to be
2558 * done before loading any values into MRFs for the sampler message since
2559 * generating these values may involve SEND messages that need the MRFs.
2562 if (ir
->coordinate
) {
2563 ir
->coordinate
->accept(this);
2564 coordinate
= this->result
;
2567 src_reg shadow_comparitor
;
2568 if (ir
->shadow_comparitor
) {
2569 ir
->shadow_comparitor
->accept(this);
2570 shadow_comparitor
= this->result
;
2573 bool has_nonconstant_offset
= ir
->offset
&& !ir
->offset
->as_constant();
2574 src_reg offset_value
;
2575 if (has_nonconstant_offset
) {
2576 ir
->offset
->accept(this);
2577 offset_value
= src_reg(this->result
);
2580 const glsl_type
*lod_type
= NULL
, *sample_index_type
= NULL
;
2581 src_reg lod
, dPdx
, dPdy
, sample_index
, mcs
;
2584 lod
= src_reg(0.0f
);
2585 lod_type
= glsl_type::float_type
;
2590 ir
->lod_info
.lod
->accept(this);
2592 lod_type
= ir
->lod_info
.lod
->type
;
2594 case ir_query_levels
:
2596 lod_type
= glsl_type::int_type
;
2599 ir
->lod_info
.sample_index
->accept(this);
2600 sample_index
= this->result
;
2601 sample_index_type
= ir
->lod_info
.sample_index
->type
;
2603 if (devinfo
->gen
>= 7 && key
->tex
.compressed_multisample_layout_mask
& (1<<sampler
))
2604 mcs
= emit_mcs_fetch(ir
, coordinate
, sampler_reg
);
2609 ir
->lod_info
.grad
.dPdx
->accept(this);
2610 dPdx
= this->result
;
2612 ir
->lod_info
.grad
.dPdy
->accept(this);
2613 dPdy
= this->result
;
2615 lod_type
= ir
->lod_info
.grad
.dPdx
->type
;
2625 case ir_tex
: opcode
= SHADER_OPCODE_TXL
; break;
2626 case ir_txl
: opcode
= SHADER_OPCODE_TXL
; break;
2627 case ir_txd
: opcode
= SHADER_OPCODE_TXD
; break;
2628 case ir_txf
: opcode
= SHADER_OPCODE_TXF
; break;
2629 case ir_txf_ms
: opcode
= SHADER_OPCODE_TXF_CMS
; break;
2630 case ir_txs
: opcode
= SHADER_OPCODE_TXS
; break;
2631 case ir_tg4
: opcode
= has_nonconstant_offset
2632 ? SHADER_OPCODE_TG4_OFFSET
: SHADER_OPCODE_TG4
; break;
2633 case ir_query_levels
: opcode
= SHADER_OPCODE_TXS
; break;
2635 unreachable("TXB is not valid for vertex shaders.");
2637 unreachable("LOD is not valid for vertex shaders.");
2639 unreachable("Unrecognized tex op");
2642 vec4_instruction
*inst
= new(mem_ctx
) vec4_instruction(
2643 opcode
, dst_reg(this, ir
->type
));
2645 if (ir
->offset
!= NULL
&& !has_nonconstant_offset
) {
2647 brw_texture_offset(ir
->offset
->as_constant()->value
.i
,
2648 ir
->offset
->type
->vector_elements
);
2651 /* Stuff the channel select bits in the top of the texture offset */
2652 if (ir
->op
== ir_tg4
)
2653 inst
->offset
|= gather_channel(ir
, sampler
) << 16;
2655 /* The message header is necessary for:
2657 * - Gen9+ for selecting SIMD4x2
2659 * - Gather channel selection
2660 * - Sampler indices too large to fit in a 4-bit value.
2663 (devinfo
->gen
< 5 || devinfo
->gen
>= 9 ||
2664 inst
->offset
!= 0 || ir
->op
== ir_tg4
||
2665 is_high_sampler(devinfo
, sampler_reg
)) ? 1 : 0;
2667 inst
->mlen
= inst
->header_size
+ 1; /* always at least one */
2668 inst
->dst
.writemask
= WRITEMASK_XYZW
;
2669 inst
->shadow_compare
= ir
->shadow_comparitor
!= NULL
;
2671 inst
->src
[1] = sampler_reg
;
2673 /* MRF for the first parameter */
2674 int param_base
= inst
->base_mrf
+ inst
->header_size
;
2676 if (ir
->op
== ir_txs
|| ir
->op
== ir_query_levels
) {
2677 int writemask
= devinfo
->gen
== 4 ? WRITEMASK_W
: WRITEMASK_X
;
2678 emit(MOV(dst_reg(MRF
, param_base
, lod_type
, writemask
), lod
));
2680 /* Load the coordinate */
2681 /* FINISHME: gl_clamp_mask and saturate */
2682 int coord_mask
= (1 << ir
->coordinate
->type
->vector_elements
) - 1;
2683 int zero_mask
= 0xf & ~coord_mask
;
2685 emit(MOV(dst_reg(MRF
, param_base
, ir
->coordinate
->type
, coord_mask
),
2688 if (zero_mask
!= 0) {
2689 emit(MOV(dst_reg(MRF
, param_base
, ir
->coordinate
->type
, zero_mask
),
2692 /* Load the shadow comparitor */
2693 if (ir
->shadow_comparitor
&& ir
->op
!= ir_txd
&& (ir
->op
!= ir_tg4
|| !has_nonconstant_offset
)) {
2694 emit(MOV(dst_reg(MRF
, param_base
+ 1, ir
->shadow_comparitor
->type
,
2696 shadow_comparitor
));
2700 /* Load the LOD info */
2701 if (ir
->op
== ir_tex
|| ir
->op
== ir_txl
) {
2703 if (devinfo
->gen
>= 5) {
2704 mrf
= param_base
+ 1;
2705 if (ir
->shadow_comparitor
) {
2706 writemask
= WRITEMASK_Y
;
2707 /* mlen already incremented */
2709 writemask
= WRITEMASK_X
;
2712 } else /* devinfo->gen == 4 */ {
2714 writemask
= WRITEMASK_W
;
2716 emit(MOV(dst_reg(MRF
, mrf
, lod_type
, writemask
), lod
));
2717 } else if (ir
->op
== ir_txf
) {
2718 emit(MOV(dst_reg(MRF
, param_base
, lod_type
, WRITEMASK_W
), lod
));
2719 } else if (ir
->op
== ir_txf_ms
) {
2720 emit(MOV(dst_reg(MRF
, param_base
+ 1, sample_index_type
, WRITEMASK_X
),
2722 if (devinfo
->gen
>= 7) {
2723 /* MCS data is in the first channel of `mcs`, but we need to get it into
2724 * the .y channel of the second vec4 of params, so replicate .x across
2725 * the whole vec4 and then mask off everything except .y
2727 mcs
.swizzle
= BRW_SWIZZLE_XXXX
;
2728 emit(MOV(dst_reg(MRF
, param_base
+ 1, glsl_type::uint_type
, WRITEMASK_Y
),
2732 } else if (ir
->op
== ir_txd
) {
2733 const glsl_type
*type
= lod_type
;
2735 if (devinfo
->gen
>= 5) {
2736 dPdx
.swizzle
= BRW_SWIZZLE4(SWIZZLE_X
,SWIZZLE_X
,SWIZZLE_Y
,SWIZZLE_Y
);
2737 dPdy
.swizzle
= BRW_SWIZZLE4(SWIZZLE_X
,SWIZZLE_X
,SWIZZLE_Y
,SWIZZLE_Y
);
2738 emit(MOV(dst_reg(MRF
, param_base
+ 1, type
, WRITEMASK_XZ
), dPdx
));
2739 emit(MOV(dst_reg(MRF
, param_base
+ 1, type
, WRITEMASK_YW
), dPdy
));
2742 if (ir
->type
->vector_elements
== 3 || ir
->shadow_comparitor
) {
2743 dPdx
.swizzle
= BRW_SWIZZLE_ZZZZ
;
2744 dPdy
.swizzle
= BRW_SWIZZLE_ZZZZ
;
2745 emit(MOV(dst_reg(MRF
, param_base
+ 2, type
, WRITEMASK_X
), dPdx
));
2746 emit(MOV(dst_reg(MRF
, param_base
+ 2, type
, WRITEMASK_Y
), dPdy
));
2749 if (ir
->shadow_comparitor
) {
2750 emit(MOV(dst_reg(MRF
, param_base
+ 2,
2751 ir
->shadow_comparitor
->type
, WRITEMASK_Z
),
2752 shadow_comparitor
));
2755 } else /* devinfo->gen == 4 */ {
2756 emit(MOV(dst_reg(MRF
, param_base
+ 1, type
, WRITEMASK_XYZ
), dPdx
));
2757 emit(MOV(dst_reg(MRF
, param_base
+ 2, type
, WRITEMASK_XYZ
), dPdy
));
2760 } else if (ir
->op
== ir_tg4
&& has_nonconstant_offset
) {
2761 if (ir
->shadow_comparitor
) {
2762 emit(MOV(dst_reg(MRF
, param_base
, ir
->shadow_comparitor
->type
, WRITEMASK_W
),
2763 shadow_comparitor
));
2766 emit(MOV(dst_reg(MRF
, param_base
+ 1, glsl_type::ivec2_type
, WRITEMASK_XY
),
2774 /* fixup num layers (z) for cube arrays: hardware returns faces * layers;
2775 * spec requires layers.
2777 if (ir
->op
== ir_txs
) {
2778 glsl_type
const *type
= ir
->sampler
->type
;
2779 if (type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_CUBE
&&
2780 type
->sampler_array
) {
2781 emit_math(SHADER_OPCODE_INT_QUOTIENT
,
2782 writemask(inst
->dst
, WRITEMASK_Z
),
2783 src_reg(inst
->dst
), src_reg(6));
2787 if (devinfo
->gen
== 6 && ir
->op
== ir_tg4
) {
2788 emit_gen6_gather_wa(key
->tex
.gen6_gather_wa
[sampler
], inst
->dst
);
2791 swizzle_result(ir
, src_reg(inst
->dst
), sampler
);
2795 * Apply workarounds for Gen6 gather with UINT/SINT
2798 vec4_visitor::emit_gen6_gather_wa(uint8_t wa
, dst_reg dst
)
2803 int width
= (wa
& WA_8BIT
) ? 8 : 16;
2804 dst_reg dst_f
= dst
;
2805 dst_f
.type
= BRW_REGISTER_TYPE_F
;
2807 /* Convert from UNORM to UINT */
2808 emit(MUL(dst_f
, src_reg(dst_f
), src_reg((float)((1 << width
) - 1))));
2809 emit(MOV(dst
, src_reg(dst_f
)));
2812 /* Reinterpret the UINT value as a signed INT value by
2813 * shifting the sign bit into place, then shifting back
2816 emit(SHL(dst
, src_reg(dst
), src_reg(32 - width
)));
2817 emit(ASR(dst
, src_reg(dst
), src_reg(32 - width
)));
2822 * Set up the gather channel based on the swizzle, for gather4.
2825 vec4_visitor::gather_channel(ir_texture
*ir
, uint32_t sampler
)
2827 ir_constant
*chan
= ir
->lod_info
.component
->as_constant();
2828 int swiz
= GET_SWZ(key
->tex
.swizzles
[sampler
], chan
->value
.i
[0]);
2830 case SWIZZLE_X
: return 0;
2832 /* gather4 sampler is broken for green channel on RG32F --
2833 * we must ask for blue instead.
2835 if (key
->tex
.gather_channel_quirk_mask
& (1<<sampler
))
2838 case SWIZZLE_Z
: return 2;
2839 case SWIZZLE_W
: return 3;
2841 unreachable("Not reached"); /* zero, one swizzles handled already */
2846 vec4_visitor::swizzle_result(ir_texture
*ir
, src_reg orig_val
, uint32_t sampler
)
2848 int s
= key
->tex
.swizzles
[sampler
];
2850 this->result
= src_reg(this, ir
->type
);
2851 dst_reg
swizzled_result(this->result
);
2853 if (ir
->op
== ir_query_levels
) {
2854 /* # levels is in .w */
2855 orig_val
.swizzle
= BRW_SWIZZLE4(SWIZZLE_W
, SWIZZLE_W
, SWIZZLE_W
, SWIZZLE_W
);
2856 emit(MOV(swizzled_result
, orig_val
));
2860 if (ir
->op
== ir_txs
|| ir
->type
== glsl_type::float_type
2861 || s
== SWIZZLE_NOOP
|| ir
->op
== ir_tg4
) {
2862 emit(MOV(swizzled_result
, orig_val
));
2867 int zero_mask
= 0, one_mask
= 0, copy_mask
= 0;
2868 int swizzle
[4] = {0};
2870 for (int i
= 0; i
< 4; i
++) {
2871 switch (GET_SWZ(s
, i
)) {
2873 zero_mask
|= (1 << i
);
2876 one_mask
|= (1 << i
);
2879 copy_mask
|= (1 << i
);
2880 swizzle
[i
] = GET_SWZ(s
, i
);
2886 orig_val
.swizzle
= BRW_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
2887 swizzled_result
.writemask
= copy_mask
;
2888 emit(MOV(swizzled_result
, orig_val
));
2892 swizzled_result
.writemask
= zero_mask
;
2893 emit(MOV(swizzled_result
, src_reg(0.0f
)));
2897 swizzled_result
.writemask
= one_mask
;
2898 emit(MOV(swizzled_result
, src_reg(1.0f
)));
2903 vec4_visitor::visit(ir_return
*)
2905 unreachable("not reached");
2909 vec4_visitor::visit(ir_discard
*)
2911 unreachable("not reached");
2915 vec4_visitor::visit(ir_if
*ir
)
2917 /* Don't point the annotation at the if statement, because then it plus
2918 * the then and else blocks get printed.
2920 this->base_ir
= ir
->condition
;
2922 if (devinfo
->gen
== 6) {
2925 enum brw_predicate predicate
;
2926 emit_bool_to_cond_code(ir
->condition
, &predicate
);
2927 emit(IF(predicate
));
2930 visit_instructions(&ir
->then_instructions
);
2932 if (!ir
->else_instructions
.is_empty()) {
2933 this->base_ir
= ir
->condition
;
2934 emit(BRW_OPCODE_ELSE
);
2936 visit_instructions(&ir
->else_instructions
);
2939 this->base_ir
= ir
->condition
;
2940 emit(BRW_OPCODE_ENDIF
);
2944 vec4_visitor::visit(ir_emit_vertex
*)
2946 unreachable("not reached");
2950 vec4_visitor::visit(ir_end_primitive
*)
2952 unreachable("not reached");
2956 vec4_visitor::visit(ir_barrier
*)
2958 unreachable("not reached");
2962 vec4_visitor::emit_untyped_atomic(unsigned atomic_op
, unsigned surf_index
,
2963 dst_reg dst
, src_reg offset
,
2964 src_reg src0
, src_reg src1
)
2968 /* Set the atomic operation offset. */
2969 emit(MOV(brw_writemask(brw_uvec_mrf(8, mlen
, 0), WRITEMASK_X
), offset
));
2972 /* Set the atomic operation arguments. */
2973 if (src0
.file
!= BAD_FILE
) {
2974 emit(MOV(brw_writemask(brw_uvec_mrf(8, mlen
, 0), WRITEMASK_X
), src0
));
2978 if (src1
.file
!= BAD_FILE
) {
2979 emit(MOV(brw_writemask(brw_uvec_mrf(8, mlen
, 0), WRITEMASK_X
), src1
));
2983 /* Emit the instruction. Note that this maps to the normal SIMD8
2984 * untyped atomic message on Ivy Bridge, but that's OK because
2985 * unused channels will be masked out.
2987 vec4_instruction
*inst
= emit(SHADER_OPCODE_UNTYPED_ATOMIC
, dst
,
2989 src_reg(surf_index
), src_reg(atomic_op
));
2994 vec4_visitor::emit_untyped_surface_read(unsigned surf_index
, dst_reg dst
,
2997 /* Set the surface read offset. */
2998 emit(MOV(brw_writemask(brw_uvec_mrf(8, 0, 0), WRITEMASK_X
), offset
));
3000 /* Emit the instruction. Note that this maps to the normal SIMD8
3001 * untyped surface read message, but that's OK because unused
3002 * channels will be masked out.
3004 vec4_instruction
*inst
= emit(SHADER_OPCODE_UNTYPED_SURFACE_READ
, dst
,
3006 src_reg(surf_index
), src_reg(1));
3011 vec4_visitor::emit_ndc_computation()
3013 /* Get the position */
3014 src_reg pos
= src_reg(output_reg
[VARYING_SLOT_POS
]);
3016 /* Build ndc coords, which are (x/w, y/w, z/w, 1/w) */
3017 dst_reg ndc
= dst_reg(this, glsl_type::vec4_type
);
3018 output_reg
[BRW_VARYING_SLOT_NDC
] = ndc
;
3020 current_annotation
= "NDC";
3021 dst_reg ndc_w
= ndc
;
3022 ndc_w
.writemask
= WRITEMASK_W
;
3023 src_reg pos_w
= pos
;
3024 pos_w
.swizzle
= BRW_SWIZZLE4(SWIZZLE_W
, SWIZZLE_W
, SWIZZLE_W
, SWIZZLE_W
);
3025 emit_math(SHADER_OPCODE_RCP
, ndc_w
, pos_w
);
3027 dst_reg ndc_xyz
= ndc
;
3028 ndc_xyz
.writemask
= WRITEMASK_XYZ
;
3030 emit(MUL(ndc_xyz
, pos
, src_reg(ndc_w
)));
3034 vec4_visitor::emit_psiz_and_flags(dst_reg reg
)
3036 if (devinfo
->gen
< 6 &&
3037 ((prog_data
->vue_map
.slots_valid
& VARYING_BIT_PSIZ
) ||
3038 key
->userclip_active
|| devinfo
->has_negative_rhw_bug
)) {
3039 dst_reg header1
= dst_reg(this, glsl_type::uvec4_type
);
3040 dst_reg header1_w
= header1
;
3041 header1_w
.writemask
= WRITEMASK_W
;
3043 emit(MOV(header1
, 0u));
3045 if (prog_data
->vue_map
.slots_valid
& VARYING_BIT_PSIZ
) {
3046 src_reg psiz
= src_reg(output_reg
[VARYING_SLOT_PSIZ
]);
3048 current_annotation
= "Point size";
3049 emit(MUL(header1_w
, psiz
, src_reg((float)(1 << 11))));
3050 emit(AND(header1_w
, src_reg(header1_w
), 0x7ff << 8));
3053 if (key
->userclip_active
) {
3054 current_annotation
= "Clipping flags";
3055 dst_reg flags0
= dst_reg(this, glsl_type::uint_type
);
3056 dst_reg flags1
= dst_reg(this, glsl_type::uint_type
);
3058 emit(CMP(dst_null_f(), src_reg(output_reg
[VARYING_SLOT_CLIP_DIST0
]), src_reg(0.0f
), BRW_CONDITIONAL_L
));
3059 emit(VS_OPCODE_UNPACK_FLAGS_SIMD4X2
, flags0
, src_reg(0));
3060 emit(OR(header1_w
, src_reg(header1_w
), src_reg(flags0
)));
3062 emit(CMP(dst_null_f(), src_reg(output_reg
[VARYING_SLOT_CLIP_DIST1
]), src_reg(0.0f
), BRW_CONDITIONAL_L
));
3063 emit(VS_OPCODE_UNPACK_FLAGS_SIMD4X2
, flags1
, src_reg(0));
3064 emit(SHL(flags1
, src_reg(flags1
), src_reg(4)));
3065 emit(OR(header1_w
, src_reg(header1_w
), src_reg(flags1
)));
3068 /* i965 clipping workaround:
3069 * 1) Test for -ve rhw
3071 * set ndc = (0,0,0,0)
3074 * Later, clipping will detect ucp[6] and ensure the primitive is
3075 * clipped against all fixed planes.
3077 if (devinfo
->has_negative_rhw_bug
) {
3078 src_reg ndc_w
= src_reg(output_reg
[BRW_VARYING_SLOT_NDC
]);
3079 ndc_w
.swizzle
= BRW_SWIZZLE_WWWW
;
3080 emit(CMP(dst_null_f(), ndc_w
, src_reg(0.0f
), BRW_CONDITIONAL_L
));
3081 vec4_instruction
*inst
;
3082 inst
= emit(OR(header1_w
, src_reg(header1_w
), src_reg(1u << 6)));
3083 inst
->predicate
= BRW_PREDICATE_NORMAL
;
3084 inst
= emit(MOV(output_reg
[BRW_VARYING_SLOT_NDC
], src_reg(0.0f
)));
3085 inst
->predicate
= BRW_PREDICATE_NORMAL
;
3088 emit(MOV(retype(reg
, BRW_REGISTER_TYPE_UD
), src_reg(header1
)));
3089 } else if (devinfo
->gen
< 6) {
3090 emit(MOV(retype(reg
, BRW_REGISTER_TYPE_UD
), 0u));
3092 emit(MOV(retype(reg
, BRW_REGISTER_TYPE_D
), src_reg(0)));
3093 if (prog_data
->vue_map
.slots_valid
& VARYING_BIT_PSIZ
) {
3094 dst_reg reg_w
= reg
;
3095 reg_w
.writemask
= WRITEMASK_W
;
3096 emit(MOV(reg_w
, src_reg(output_reg
[VARYING_SLOT_PSIZ
])));
3098 if (prog_data
->vue_map
.slots_valid
& VARYING_BIT_LAYER
) {
3099 dst_reg reg_y
= reg
;
3100 reg_y
.writemask
= WRITEMASK_Y
;
3101 reg_y
.type
= BRW_REGISTER_TYPE_D
;
3102 emit(MOV(reg_y
, src_reg(output_reg
[VARYING_SLOT_LAYER
])));
3104 if (prog_data
->vue_map
.slots_valid
& VARYING_BIT_VIEWPORT
) {
3105 dst_reg reg_z
= reg
;
3106 reg_z
.writemask
= WRITEMASK_Z
;
3107 reg_z
.type
= BRW_REGISTER_TYPE_D
;
3108 emit(MOV(reg_z
, src_reg(output_reg
[VARYING_SLOT_VIEWPORT
])));
3114 vec4_visitor::emit_clip_distances(dst_reg reg
, int offset
)
3116 /* From the GLSL 1.30 spec, section 7.1 (Vertex Shader Special Variables):
3118 * "If a linked set of shaders forming the vertex stage contains no
3119 * static write to gl_ClipVertex or gl_ClipDistance, but the
3120 * application has requested clipping against user clip planes through
3121 * the API, then the coordinate written to gl_Position is used for
3122 * comparison against the user clip planes."
3124 * This function is only called if the shader didn't write to
3125 * gl_ClipDistance. Accordingly, we use gl_ClipVertex to perform clipping
3126 * if the user wrote to it; otherwise we use gl_Position.
3128 gl_varying_slot clip_vertex
= VARYING_SLOT_CLIP_VERTEX
;
3129 if (!(prog_data
->vue_map
.slots_valid
& VARYING_BIT_CLIP_VERTEX
)) {
3130 clip_vertex
= VARYING_SLOT_POS
;
3133 for (int i
= 0; i
+ offset
< key
->nr_userclip_plane_consts
&& i
< 4;
3135 reg
.writemask
= 1 << i
;
3137 src_reg(output_reg
[clip_vertex
]),
3138 src_reg(this->userplane
[i
+ offset
])));
3143 vec4_visitor::emit_generic_urb_slot(dst_reg reg
, int varying
)
3145 assert (varying
< VARYING_SLOT_MAX
);
3146 reg
.type
= output_reg
[varying
].type
;
3147 current_annotation
= output_reg_annotation
[varying
];
3148 /* Copy the register, saturating if necessary */
3149 return emit(MOV(reg
, src_reg(output_reg
[varying
])));
3153 vec4_visitor::emit_urb_slot(dst_reg reg
, int varying
)
3155 reg
.type
= BRW_REGISTER_TYPE_F
;
3158 case VARYING_SLOT_PSIZ
:
3160 /* PSIZ is always in slot 0, and is coupled with other flags. */
3161 current_annotation
= "indices, point width, clip flags";
3162 emit_psiz_and_flags(reg
);
3165 case BRW_VARYING_SLOT_NDC
:
3166 current_annotation
= "NDC";
3167 emit(MOV(reg
, src_reg(output_reg
[BRW_VARYING_SLOT_NDC
])));
3169 case VARYING_SLOT_POS
:
3170 current_annotation
= "gl_Position";
3171 emit(MOV(reg
, src_reg(output_reg
[VARYING_SLOT_POS
])));
3173 case VARYING_SLOT_EDGE
:
3174 /* This is present when doing unfilled polygons. We're supposed to copy
3175 * the edge flag from the user-provided vertex array
3176 * (glEdgeFlagPointer), or otherwise we'll copy from the current value
3177 * of that attribute (starts as 1.0f). This is then used in clipping to
3178 * determine which edges should be drawn as wireframe.
3180 current_annotation
= "edge flag";
3181 emit(MOV(reg
, src_reg(dst_reg(ATTR
, VERT_ATTRIB_EDGEFLAG
,
3182 glsl_type::float_type
, WRITEMASK_XYZW
))));
3184 case BRW_VARYING_SLOT_PAD
:
3185 /* No need to write to this slot */
3187 case VARYING_SLOT_COL0
:
3188 case VARYING_SLOT_COL1
:
3189 case VARYING_SLOT_BFC0
:
3190 case VARYING_SLOT_BFC1
: {
3191 /* These built-in varyings are only supported in compatibility mode,
3192 * and we only support GS in core profile. So, this must be a vertex
3195 assert(stage
== MESA_SHADER_VERTEX
);
3196 vec4_instruction
*inst
= emit_generic_urb_slot(reg
, varying
);
3197 if (((struct brw_vs_prog_key
*) key
)->clamp_vertex_color
)
3198 inst
->saturate
= true;
3203 emit_generic_urb_slot(reg
, varying
);
3209 align_interleaved_urb_mlen(const struct brw_device_info
*devinfo
, int mlen
)
3211 if (devinfo
->gen
>= 6) {
3212 /* URB data written (does not include the message header reg) must
3213 * be a multiple of 256 bits, or 2 VS registers. See vol5c.5,
3214 * section 5.4.3.2.2: URB_INTERLEAVED.
3216 * URB entries are allocated on a multiple of 1024 bits, so an
3217 * extra 128 bits written here to make the end align to 256 is
3220 if ((mlen
% 2) != 1)
3229 * Generates the VUE payload plus the necessary URB write instructions to
3232 * The VUE layout is documented in Volume 2a.
3235 vec4_visitor::emit_vertex()
3237 /* MRF 0 is reserved for the debugger, so start with message header
3242 /* In the process of generating our URB write message contents, we
3243 * may need to unspill a register or load from an array. Those
3244 * reads would use MRFs 14-15.
3246 int max_usable_mrf
= 13;
3248 /* The following assertion verifies that max_usable_mrf causes an
3249 * even-numbered amount of URB write data, which will meet gen6's
3250 * requirements for length alignment.
3252 assert ((max_usable_mrf
- base_mrf
) % 2 == 0);
3254 /* First mrf is the g0-based message header containing URB handles and
3257 emit_urb_write_header(mrf
++);
3259 if (devinfo
->gen
< 6) {
3260 emit_ndc_computation();
3263 /* Lower legacy ff and ClipVertex clipping to clip distances */
3264 if (key
->userclip_active
&& !prog
->UsesClipDistanceOut
) {
3265 current_annotation
= "user clip distances";
3267 output_reg
[VARYING_SLOT_CLIP_DIST0
] = dst_reg(this, glsl_type::vec4_type
);
3268 output_reg
[VARYING_SLOT_CLIP_DIST1
] = dst_reg(this, glsl_type::vec4_type
);
3270 emit_clip_distances(output_reg
[VARYING_SLOT_CLIP_DIST0
], 0);
3271 emit_clip_distances(output_reg
[VARYING_SLOT_CLIP_DIST1
], 4);
3274 /* We may need to split this up into several URB writes, so do them in a
3278 bool complete
= false;
3280 /* URB offset is in URB row increments, and each of our MRFs is half of
3281 * one of those, since we're doing interleaved writes.
3283 int offset
= slot
/ 2;
3286 for (; slot
< prog_data
->vue_map
.num_slots
; ++slot
) {
3287 emit_urb_slot(dst_reg(MRF
, mrf
++),
3288 prog_data
->vue_map
.slot_to_varying
[slot
]);
3290 /* If this was max_usable_mrf, we can't fit anything more into this
3293 if (mrf
> max_usable_mrf
) {
3299 complete
= slot
>= prog_data
->vue_map
.num_slots
;
3300 current_annotation
= "URB write";
3301 vec4_instruction
*inst
= emit_urb_write_opcode(complete
);
3302 inst
->base_mrf
= base_mrf
;
3303 inst
->mlen
= align_interleaved_urb_mlen(devinfo
, mrf
- base_mrf
);
3304 inst
->offset
+= offset
;
3310 vec4_visitor::get_scratch_offset(bblock_t
*block
, vec4_instruction
*inst
,
3311 src_reg
*reladdr
, int reg_offset
)
3313 /* Because we store the values to scratch interleaved like our
3314 * vertex data, we need to scale the vec4 index by 2.
3316 int message_header_scale
= 2;
3318 /* Pre-gen6, the message header uses byte offsets instead of vec4
3319 * (16-byte) offset units.
3321 if (devinfo
->gen
< 6)
3322 message_header_scale
*= 16;
3325 src_reg index
= src_reg(this, glsl_type::int_type
);
3327 emit_before(block
, inst
, ADD(dst_reg(index
), *reladdr
,
3328 src_reg(reg_offset
)));
3329 emit_before(block
, inst
, MUL(dst_reg(index
), index
,
3330 src_reg(message_header_scale
)));
3334 return src_reg(reg_offset
* message_header_scale
);
3339 vec4_visitor::get_pull_constant_offset(bblock_t
* block
, vec4_instruction
*inst
,
3340 src_reg
*reladdr
, int reg_offset
)
3343 src_reg index
= src_reg(this, glsl_type::int_type
);
3345 emit_before(block
, inst
, ADD(dst_reg(index
), *reladdr
,
3346 src_reg(reg_offset
)));
3348 /* Pre-gen6, the message header uses byte offsets instead of vec4
3349 * (16-byte) offset units.
3351 if (devinfo
->gen
< 6) {
3352 emit_before(block
, inst
, MUL(dst_reg(index
), index
, src_reg(16)));
3356 } else if (devinfo
->gen
>= 8) {
3357 /* Store the offset in a GRF so we can send-from-GRF. */
3358 src_reg offset
= src_reg(this, glsl_type::int_type
);
3359 emit_before(block
, inst
, MOV(dst_reg(offset
), src_reg(reg_offset
)));
3362 int message_header_scale
= devinfo
->gen
< 6 ? 16 : 1;
3363 return src_reg(reg_offset
* message_header_scale
);
3368 * Emits an instruction before @inst to load the value named by @orig_src
3369 * from scratch space at @base_offset to @temp.
3371 * @base_offset is measured in 32-byte units (the size of a register).
3374 vec4_visitor::emit_scratch_read(bblock_t
*block
, vec4_instruction
*inst
,
3375 dst_reg temp
, src_reg orig_src
,
3378 int reg_offset
= base_offset
+ orig_src
.reg_offset
;
3379 src_reg index
= get_scratch_offset(block
, inst
, orig_src
.reladdr
,
3382 emit_before(block
, inst
, SCRATCH_READ(temp
, index
));
3386 * Emits an instruction after @inst to store the value to be written
3387 * to @orig_dst to scratch space at @base_offset, from @temp.
3389 * @base_offset is measured in 32-byte units (the size of a register).
3392 vec4_visitor::emit_scratch_write(bblock_t
*block
, vec4_instruction
*inst
,
3395 int reg_offset
= base_offset
+ inst
->dst
.reg_offset
;
3396 src_reg index
= get_scratch_offset(block
, inst
, inst
->dst
.reladdr
,
3399 /* Create a temporary register to store *inst's result in.
3401 * We have to be careful in MOVing from our temporary result register in
3402 * the scratch write. If we swizzle from channels of the temporary that
3403 * weren't initialized, it will confuse live interval analysis, which will
3404 * make spilling fail to make progress.
3406 const src_reg temp
= swizzle(retype(src_reg(this, glsl_type::vec4_type
),
3408 brw_swizzle_for_mask(inst
->dst
.writemask
));
3409 dst_reg dst
= dst_reg(brw_writemask(brw_vec8_grf(0, 0),
3410 inst
->dst
.writemask
));
3411 vec4_instruction
*write
= SCRATCH_WRITE(dst
, temp
, index
);
3412 write
->predicate
= inst
->predicate
;
3413 write
->ir
= inst
->ir
;
3414 write
->annotation
= inst
->annotation
;
3415 inst
->insert_after(block
, write
);
3417 inst
->dst
.file
= temp
.file
;
3418 inst
->dst
.reg
= temp
.reg
;
3419 inst
->dst
.reg_offset
= temp
.reg_offset
;
3420 inst
->dst
.reladdr
= NULL
;
3424 * Checks if \p src and/or \p src.reladdr require a scratch read, and if so,
3425 * adds the scratch read(s) before \p inst. The function also checks for
3426 * recursive reladdr scratch accesses, issuing the corresponding scratch
3427 * loads and rewriting reladdr references accordingly.
3429 * \return \p src if it did not require a scratch load, otherwise, the
3430 * register holding the result of the scratch load that the caller should
3431 * use to rewrite src.
3434 vec4_visitor::emit_resolve_reladdr(int scratch_loc
[], bblock_t
*block
,
3435 vec4_instruction
*inst
, src_reg src
)
3437 /* Resolve recursive reladdr scratch access by calling ourselves
3441 *src
.reladdr
= emit_resolve_reladdr(scratch_loc
, block
, inst
,
3444 /* Now handle scratch access on src */
3445 if (src
.file
== GRF
&& scratch_loc
[src
.reg
] != -1) {
3446 dst_reg temp
= dst_reg(this, glsl_type::vec4_type
);
3447 emit_scratch_read(block
, inst
, temp
, src
, scratch_loc
[src
.reg
]);
3449 src
.reg_offset
= temp
.reg_offset
;
3457 * We can't generally support array access in GRF space, because a
3458 * single instruction's destination can only span 2 contiguous
3459 * registers. So, we send all GRF arrays that get variable index
3460 * access to scratch space.
3463 vec4_visitor::move_grf_array_access_to_scratch()
3465 int scratch_loc
[this->alloc
.count
];
3466 memset(scratch_loc
, -1, sizeof(scratch_loc
));
3468 /* First, calculate the set of virtual GRFs that need to be punted
3469 * to scratch due to having any array access on them, and where in
3472 foreach_block_and_inst(block
, vec4_instruction
, inst
, cfg
) {
3473 if (inst
->dst
.file
== GRF
&& inst
->dst
.reladdr
) {
3474 if (scratch_loc
[inst
->dst
.reg
] == -1) {
3475 scratch_loc
[inst
->dst
.reg
] = last_scratch
;
3476 last_scratch
+= this->alloc
.sizes
[inst
->dst
.reg
];
3479 for (src_reg
*iter
= inst
->dst
.reladdr
;
3481 iter
= iter
->reladdr
) {
3482 if (iter
->file
== GRF
&& scratch_loc
[iter
->reg
] == -1) {
3483 scratch_loc
[iter
->reg
] = last_scratch
;
3484 last_scratch
+= this->alloc
.sizes
[iter
->reg
];
3489 for (int i
= 0 ; i
< 3; i
++) {
3490 for (src_reg
*iter
= &inst
->src
[i
];
3492 iter
= iter
->reladdr
) {
3493 if (iter
->file
== GRF
&& scratch_loc
[iter
->reg
] == -1) {
3494 scratch_loc
[iter
->reg
] = last_scratch
;
3495 last_scratch
+= this->alloc
.sizes
[iter
->reg
];
3501 /* Now, for anything that will be accessed through scratch, rewrite
3502 * it to load/store. Note that this is a _safe list walk, because
3503 * we may generate a new scratch_write instruction after the one
3506 foreach_block_and_inst_safe(block
, vec4_instruction
, inst
, cfg
) {
3507 /* Set up the annotation tracking for new generated instructions. */
3509 current_annotation
= inst
->annotation
;
3511 /* First handle scratch access on the dst. Notice we have to handle
3512 * the case where the dst's reladdr also points to scratch space.
3514 if (inst
->dst
.reladdr
)
3515 *inst
->dst
.reladdr
= emit_resolve_reladdr(scratch_loc
, block
, inst
,
3516 *inst
->dst
.reladdr
);
3518 /* Now that we have handled any (possibly recursive) reladdr scratch
3519 * accesses for dst we can safely do the scratch write for dst itself
3521 if (inst
->dst
.file
== GRF
&& scratch_loc
[inst
->dst
.reg
] != -1)
3522 emit_scratch_write(block
, inst
, scratch_loc
[inst
->dst
.reg
]);
3524 /* Now handle scratch access on any src. In this case, since inst->src[i]
3525 * already is a src_reg, we can just call emit_resolve_reladdr with
3526 * inst->src[i] and it will take care of handling scratch loads for
3527 * both src and src.reladdr (recursively).
3529 for (int i
= 0 ; i
< 3; i
++) {
3530 inst
->src
[i
] = emit_resolve_reladdr(scratch_loc
, block
, inst
,
3537 * Emits an instruction before @inst to load the value named by @orig_src
3538 * from the pull constant buffer (surface) at @base_offset to @temp.
3541 vec4_visitor::emit_pull_constant_load(bblock_t
*block
, vec4_instruction
*inst
,
3542 dst_reg temp
, src_reg orig_src
,
3545 int reg_offset
= base_offset
+ orig_src
.reg_offset
;
3546 src_reg index
= src_reg(prog_data
->base
.binding_table
.pull_constants_start
);
3547 src_reg offset
= get_pull_constant_offset(block
, inst
, orig_src
.reladdr
,
3550 emit_pull_constant_load_reg(temp
,
3557 * Implements array access of uniforms by inserting a
3558 * PULL_CONSTANT_LOAD instruction.
3560 * Unlike temporary GRF array access (where we don't support it due to
3561 * the difficulty of doing relative addressing on instruction
3562 * destinations), we could potentially do array access of uniforms
3563 * that were loaded in GRF space as push constants. In real-world
3564 * usage we've seen, though, the arrays being used are always larger
3565 * than we could load as push constants, so just always move all
3566 * uniform array access out to a pull constant buffer.
3569 vec4_visitor::move_uniform_array_access_to_pull_constants()
3571 int pull_constant_loc
[this->uniforms
];
3572 memset(pull_constant_loc
, -1, sizeof(pull_constant_loc
));
3573 bool nested_reladdr
;
3575 /* Walk through and find array access of uniforms. Put a copy of that
3576 * uniform in the pull constant buffer.
3578 * Note that we don't move constant-indexed accesses to arrays. No
3579 * testing has been done of the performance impact of this choice.
3582 nested_reladdr
= false;
3584 foreach_block_and_inst_safe(block
, vec4_instruction
, inst
, cfg
) {
3585 for (int i
= 0 ; i
< 3; i
++) {
3586 if (inst
->src
[i
].file
!= UNIFORM
|| !inst
->src
[i
].reladdr
)
3589 int uniform
= inst
->src
[i
].reg
;
3591 if (inst
->src
[i
].reladdr
->reladdr
)
3592 nested_reladdr
= true; /* will need another pass */
3594 /* If this array isn't already present in the pull constant buffer,
3597 if (pull_constant_loc
[uniform
] == -1) {
3598 const gl_constant_value
**values
=
3599 &stage_prog_data
->param
[uniform
* 4];
3601 pull_constant_loc
[uniform
] = stage_prog_data
->nr_pull_params
/ 4;
3603 assert(uniform
< uniform_array_size
);
3604 for (int j
= 0; j
< uniform_size
[uniform
] * 4; j
++) {
3605 stage_prog_data
->pull_param
[stage_prog_data
->nr_pull_params
++]
3610 /* Set up the annotation tracking for new generated instructions. */
3612 current_annotation
= inst
->annotation
;
3614 dst_reg temp
= dst_reg(this, glsl_type::vec4_type
);
3616 emit_pull_constant_load(block
, inst
, temp
, inst
->src
[i
],
3617 pull_constant_loc
[uniform
]);
3619 inst
->src
[i
].file
= temp
.file
;
3620 inst
->src
[i
].reg
= temp
.reg
;
3621 inst
->src
[i
].reg_offset
= temp
.reg_offset
;
3622 inst
->src
[i
].reladdr
= NULL
;
3625 } while (nested_reladdr
);
3627 /* Now there are no accesses of the UNIFORM file with a reladdr, so
3628 * no need to track them as larger-than-vec4 objects. This will be
3629 * relied on in cutting out unused uniform vectors from push
3632 split_uniform_registers();
3636 vec4_visitor::resolve_ud_negate(src_reg
*reg
)
3638 if (reg
->type
!= BRW_REGISTER_TYPE_UD
||
3642 src_reg temp
= src_reg(this, glsl_type::uvec4_type
);
3643 emit(BRW_OPCODE_MOV
, dst_reg(temp
), *reg
);
3648 * Resolve the result of a Gen4-5 CMP instruction to a proper boolean.
3650 * CMP on Gen4-5 only sets the LSB of the result; the rest are undefined.
3651 * If we need a proper boolean value, we have to fix it up to be 0 or ~0.
3654 vec4_visitor::resolve_bool_comparison(ir_rvalue
*rvalue
, src_reg
*reg
)
3656 assert(devinfo
->gen
<= 5);
3658 if (!rvalue
->type
->is_boolean())
3661 src_reg and_result
= src_reg(this, rvalue
->type
);
3662 src_reg neg_result
= src_reg(this, rvalue
->type
);
3663 emit(AND(dst_reg(and_result
), *reg
, src_reg(1)));
3664 emit(MOV(dst_reg(neg_result
), negate(and_result
)));
3668 vec4_visitor::vec4_visitor(const struct brw_compiler
*compiler
,
3670 struct gl_program
*prog
,
3671 const struct brw_vue_prog_key
*key
,
3672 struct brw_vue_prog_data
*prog_data
,
3673 struct gl_shader_program
*shader_prog
,
3674 gl_shader_stage stage
,
3677 int shader_time_index
)
3678 : backend_shader(compiler
, log_data
, mem_ctx
,
3679 shader_prog
, prog
, &prog_data
->base
, stage
),
3681 prog_data(prog_data
),
3682 sanity_param_count(0),
3684 first_non_payload_grf(0),
3685 need_all_constants_in_pull_buffer(false),
3686 no_spills(no_spills
),
3687 shader_time_index(shader_time_index
),
3690 this->failed
= false;
3692 this->base_ir
= NULL
;
3693 this->current_annotation
= NULL
;
3694 memset(this->output_reg_annotation
, 0, sizeof(this->output_reg_annotation
));
3696 this->variable_ht
= hash_table_ctor(0,
3697 hash_table_pointer_hash
,
3698 hash_table_pointer_compare
);
3700 this->virtual_grf_start
= NULL
;
3701 this->virtual_grf_end
= NULL
;
3702 this->live_intervals
= NULL
;
3704 this->max_grf
= devinfo
->gen
>= 7 ? GEN7_MRF_HACK_START
: BRW_MAX_GRF
;
3708 /* Initialize uniform_array_size to at least 1 because pre-gen6 VS requires
3709 * at least one. See setup_uniforms() in brw_vec4.cpp.
3711 this->uniform_array_size
= 1;
3713 this->uniform_array_size
=
3714 MAX2(DIV_ROUND_UP(stage_prog_data
->nr_params
, 4), 1);
3717 this->uniform_size
= rzalloc_array(mem_ctx
, int, this->uniform_array_size
);
3718 this->uniform_vector_size
= rzalloc_array(mem_ctx
, int, this->uniform_array_size
);
3721 vec4_visitor::~vec4_visitor()
3723 hash_table_dtor(this->variable_ht
);
3728 vec4_visitor::fail(const char *format
, ...)
3738 va_start(va
, format
);
3739 msg
= ralloc_vasprintf(mem_ctx
, format
, va
);
3741 msg
= ralloc_asprintf(mem_ctx
, "%s compile failed: %s\n", stage_abbrev
, msg
);
3743 this->fail_msg
= msg
;
3745 if (debug_enabled
) {
3746 fprintf(stderr
, "%s", msg
);
3750 } /* namespace brw */