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
,
259 inst
->base_mrf
= FIRST_SPILL_MRF(devinfo
->gen
) + 1;
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
,
273 inst
->base_mrf
= FIRST_SPILL_MRF(devinfo
->gen
);
280 vec4_visitor::fix_3src_operand(const src_reg
&src
)
282 /* Using vec4 uniforms in SIMD4x2 programs is difficult. You'd like to be
283 * able to use vertical stride of zero to replicate the vec4 uniform, like
285 * g3<0;4,1>:f - [0, 4][1, 5][2, 6][3, 7]
287 * But you can't, since vertical stride is always four in three-source
288 * instructions. Instead, insert a MOV instruction to do the replication so
289 * that the three-source instruction can consume it.
292 /* The MOV is only needed if the source is a uniform or immediate. */
293 if (src
.file
!= UNIFORM
&& src
.file
!= IMM
)
296 if (src
.file
== UNIFORM
&& brw_is_single_value_swizzle(src
.swizzle
))
299 dst_reg expanded
= dst_reg(this, glsl_type::vec4_type
);
300 expanded
.type
= src
.type
;
301 emit(VEC4_OPCODE_UNPACK_UNIFORM
, expanded
, src
);
302 return src_reg(expanded
);
306 vec4_visitor::resolve_source_modifiers(const src_reg
&src
)
308 if (!src
.abs
&& !src
.negate
)
311 dst_reg resolved
= dst_reg(this, glsl_type::ivec4_type
);
312 resolved
.type
= src
.type
;
313 emit(MOV(resolved
, src
));
315 return src_reg(resolved
);
319 vec4_visitor::fix_math_operand(const src_reg
&src
)
321 if (devinfo
->gen
< 6 || devinfo
->gen
>= 8 || src
.file
== BAD_FILE
)
324 /* The gen6 math instruction ignores the source modifiers --
325 * swizzle, abs, negate, and at least some parts of the register
326 * region description.
328 * Rather than trying to enumerate all these cases, *always* expand the
329 * operand to a temp GRF for gen6.
331 * For gen7, keep the operand as-is, except if immediate, which gen7 still
335 if (devinfo
->gen
== 7 && src
.file
!= IMM
)
338 dst_reg expanded
= dst_reg(this, glsl_type::vec4_type
);
339 expanded
.type
= src
.type
;
340 emit(MOV(expanded
, src
));
341 return src_reg(expanded
);
345 vec4_visitor::emit_math(enum opcode opcode
,
347 const src_reg
&src0
, const src_reg
&src1
)
349 vec4_instruction
*math
=
350 emit(opcode
, dst
, fix_math_operand(src0
), fix_math_operand(src1
));
352 if (devinfo
->gen
== 6 && dst
.writemask
!= WRITEMASK_XYZW
) {
353 /* MATH on Gen6 must be align1, so we can't do writemasks. */
354 math
->dst
= dst_reg(this, glsl_type::vec4_type
);
355 math
->dst
.type
= dst
.type
;
356 math
= emit(MOV(dst
, src_reg(math
->dst
)));
357 } else if (devinfo
->gen
< 6) {
359 math
->mlen
= src1
.file
== BAD_FILE
? 1 : 2;
366 vec4_visitor::emit_pack_half_2x16(dst_reg dst
, src_reg src0
)
368 if (devinfo
->gen
< 7) {
369 unreachable("ir_unop_pack_half_2x16 should be lowered");
372 assert(dst
.type
== BRW_REGISTER_TYPE_UD
);
373 assert(src0
.type
== BRW_REGISTER_TYPE_F
);
375 /* From the Ivybridge PRM, Vol4, Part3, Section 6.27 f32to16:
377 * Because this instruction does not have a 16-bit floating-point type,
378 * the destination data type must be Word (W).
380 * The destination must be DWord-aligned and specify a horizontal stride
381 * (HorzStride) of 2. The 16-bit result is stored in the lower word of
382 * each destination channel and the upper word is not modified.
384 * The above restriction implies that the f32to16 instruction must use
385 * align1 mode, because only in align1 mode is it possible to specify
386 * horizontal stride. We choose here to defy the hardware docs and emit
387 * align16 instructions.
389 * (I [chadv] did attempt to emit align1 instructions for VS f32to16
390 * instructions. I was partially successful in that the code passed all
391 * tests. However, the code was dubiously correct and fragile, and the
392 * tests were not harsh enough to probe that frailty. Not trusting the
393 * code, I chose instead to remain in align16 mode in defiance of the hw
396 * I've [chadv] experimentally confirmed that, on gen7 hardware and the
397 * simulator, emitting a f32to16 in align16 mode with UD as destination
398 * data type is safe. The behavior differs from that specified in the PRM
399 * in that the upper word of each destination channel is cleared to 0.
402 dst_reg
tmp_dst(this, glsl_type::uvec2_type
);
403 src_reg
tmp_src(tmp_dst
);
406 /* Verify the undocumented behavior on which the following instructions
407 * rely. If f32to16 fails to clear the upper word of the X and Y channels,
408 * then the result of the bit-or instruction below will be incorrect.
410 * You should inspect the disasm output in order to verify that the MOV is
411 * not optimized away.
413 emit(MOV(tmp_dst
, src_reg(0x12345678u
)));
416 /* Give tmp the form below, where "." means untouched.
419 * |.|.|0x0000hhhh|0x0000llll|.|.|0x0000hhhh|0x0000llll|
421 * That the upper word of each write-channel be 0 is required for the
422 * following bit-shift and bit-or instructions to work. Note that this
423 * relies on the undocumented hardware behavior mentioned above.
425 tmp_dst
.writemask
= WRITEMASK_XY
;
426 emit(F32TO16(tmp_dst
, src0
));
428 /* Give the write-channels of dst the form:
431 tmp_src
.swizzle
= BRW_SWIZZLE_YYYY
;
432 emit(SHL(dst
, tmp_src
, src_reg(16u)));
434 /* Finally, give the write-channels of dst the form of packHalf2x16's
438 tmp_src
.swizzle
= BRW_SWIZZLE_XXXX
;
439 emit(OR(dst
, src_reg(dst
), tmp_src
));
443 vec4_visitor::emit_unpack_half_2x16(dst_reg dst
, src_reg src0
)
445 if (devinfo
->gen
< 7) {
446 unreachable("ir_unop_unpack_half_2x16 should be lowered");
449 assert(dst
.type
== BRW_REGISTER_TYPE_F
);
450 assert(src0
.type
== BRW_REGISTER_TYPE_UD
);
452 /* From the Ivybridge PRM, Vol4, Part3, Section 6.26 f16to32:
454 * Because this instruction does not have a 16-bit floating-point type,
455 * the source data type must be Word (W). The destination type must be
458 * To use W as the source data type, we must adjust horizontal strides,
459 * which is only possible in align1 mode. All my [chadv] attempts at
460 * emitting align1 instructions for unpackHalf2x16 failed to pass the
461 * Piglit tests, so I gave up.
463 * I've verified that, on gen7 hardware and the simulator, it is safe to
464 * emit f16to32 in align16 mode with UD as source data type.
467 dst_reg
tmp_dst(this, glsl_type::uvec2_type
);
468 src_reg
tmp_src(tmp_dst
);
470 tmp_dst
.writemask
= WRITEMASK_X
;
471 emit(AND(tmp_dst
, src0
, src_reg(0xffffu
)));
473 tmp_dst
.writemask
= WRITEMASK_Y
;
474 emit(SHR(tmp_dst
, src0
, src_reg(16u)));
476 dst
.writemask
= WRITEMASK_XY
;
477 emit(F16TO32(dst
, tmp_src
));
481 vec4_visitor::emit_unpack_unorm_4x8(const dst_reg
&dst
, src_reg src0
)
483 /* Instead of splitting the 32-bit integer, shifting, and ORing it back
484 * together, we can shift it by <0, 8, 16, 24>. The packed integer immediate
485 * is not suitable to generate the shift values, but we can use the packed
486 * vector float and a type-converting MOV.
488 dst_reg
shift(this, glsl_type::uvec4_type
);
489 emit(MOV(shift
, src_reg(0x00, 0x60, 0x70, 0x78)));
491 dst_reg
shifted(this, glsl_type::uvec4_type
);
492 src0
.swizzle
= BRW_SWIZZLE_XXXX
;
493 emit(SHR(shifted
, src0
, src_reg(shift
)));
495 shifted
.type
= BRW_REGISTER_TYPE_UB
;
496 dst_reg
f(this, glsl_type::vec4_type
);
497 emit(VEC4_OPCODE_MOV_BYTES
, f
, src_reg(shifted
));
499 emit(MUL(dst
, src_reg(f
), src_reg(1.0f
/ 255.0f
)));
503 vec4_visitor::emit_unpack_snorm_4x8(const dst_reg
&dst
, src_reg src0
)
505 /* Instead of splitting the 32-bit integer, shifting, and ORing it back
506 * together, we can shift it by <0, 8, 16, 24>. The packed integer immediate
507 * is not suitable to generate the shift values, but we can use the packed
508 * vector float and a type-converting MOV.
510 dst_reg
shift(this, glsl_type::uvec4_type
);
511 emit(MOV(shift
, src_reg(0x00, 0x60, 0x70, 0x78)));
513 dst_reg
shifted(this, glsl_type::uvec4_type
);
514 src0
.swizzle
= BRW_SWIZZLE_XXXX
;
515 emit(SHR(shifted
, src0
, src_reg(shift
)));
517 shifted
.type
= BRW_REGISTER_TYPE_B
;
518 dst_reg
f(this, glsl_type::vec4_type
);
519 emit(VEC4_OPCODE_MOV_BYTES
, f
, src_reg(shifted
));
521 dst_reg
scaled(this, glsl_type::vec4_type
);
522 emit(MUL(scaled
, src_reg(f
), src_reg(1.0f
/ 127.0f
)));
524 dst_reg
max(this, glsl_type::vec4_type
);
525 emit_minmax(BRW_CONDITIONAL_GE
, max
, src_reg(scaled
), src_reg(-1.0f
));
526 emit_minmax(BRW_CONDITIONAL_L
, dst
, src_reg(max
), src_reg(1.0f
));
530 vec4_visitor::emit_pack_unorm_4x8(const dst_reg
&dst
, const src_reg
&src0
)
532 dst_reg
saturated(this, glsl_type::vec4_type
);
533 vec4_instruction
*inst
= emit(MOV(saturated
, src0
));
534 inst
->saturate
= true;
536 dst_reg
scaled(this, glsl_type::vec4_type
);
537 emit(MUL(scaled
, src_reg(saturated
), src_reg(255.0f
)));
539 dst_reg
rounded(this, glsl_type::vec4_type
);
540 emit(RNDE(rounded
, src_reg(scaled
)));
542 dst_reg
u(this, glsl_type::uvec4_type
);
543 emit(MOV(u
, src_reg(rounded
)));
546 emit(VEC4_OPCODE_PACK_BYTES
, dst
, bytes
);
550 vec4_visitor::emit_pack_snorm_4x8(const dst_reg
&dst
, const src_reg
&src0
)
552 dst_reg
max(this, glsl_type::vec4_type
);
553 emit_minmax(BRW_CONDITIONAL_GE
, max
, src0
, src_reg(-1.0f
));
555 dst_reg
min(this, glsl_type::vec4_type
);
556 emit_minmax(BRW_CONDITIONAL_L
, min
, src_reg(max
), src_reg(1.0f
));
558 dst_reg
scaled(this, glsl_type::vec4_type
);
559 emit(MUL(scaled
, src_reg(min
), src_reg(127.0f
)));
561 dst_reg
rounded(this, glsl_type::vec4_type
);
562 emit(RNDE(rounded
, src_reg(scaled
)));
564 dst_reg
i(this, glsl_type::ivec4_type
);
565 emit(MOV(i
, src_reg(rounded
)));
568 emit(VEC4_OPCODE_PACK_BYTES
, dst
, bytes
);
572 * Returns the minimum number of vec4 elements needed to pack a type.
574 * For simple types, it will return 1 (a single vec4); for matrices, the
575 * number of columns; for array and struct, the sum of the vec4_size of
576 * each of its elements; and for sampler and atomic, zero.
578 * This method is useful to calculate how much register space is needed to
579 * store a particular type.
582 type_size_vec4(const struct glsl_type
*type
)
587 switch (type
->base_type
) {
590 case GLSL_TYPE_FLOAT
:
592 if (type
->is_matrix()) {
593 return type
->matrix_columns
;
595 /* Regardless of size of vector, it gets a vec4. This is bad
596 * packing for things like floats, but otherwise arrays become a
597 * mess. Hopefully a later pass over the code can pack scalars
598 * down if appropriate.
602 case GLSL_TYPE_ARRAY
:
603 assert(type
->length
> 0);
604 return type_size_vec4(type
->fields
.array
) * type
->length
;
605 case GLSL_TYPE_STRUCT
:
607 for (i
= 0; i
< type
->length
; i
++) {
608 size
+= type_size_vec4(type
->fields
.structure
[i
].type
);
611 case GLSL_TYPE_SUBROUTINE
:
614 case GLSL_TYPE_SAMPLER
:
615 /* Samplers take up no register space, since they're baked in at
619 case GLSL_TYPE_ATOMIC_UINT
:
621 case GLSL_TYPE_IMAGE
:
622 return DIV_ROUND_UP(BRW_IMAGE_PARAM_SIZE
, 4);
624 case GLSL_TYPE_DOUBLE
:
625 case GLSL_TYPE_ERROR
:
626 case GLSL_TYPE_INTERFACE
:
627 case GLSL_TYPE_FUNCTION
:
628 unreachable("not reached");
634 src_reg::src_reg(class vec4_visitor
*v
, const struct glsl_type
*type
)
639 this->reg
= v
->alloc
.allocate(type_size_vec4(type
));
641 if (type
->is_array() || type
->is_record()) {
642 this->swizzle
= BRW_SWIZZLE_NOOP
;
644 this->swizzle
= brw_swizzle_for_size(type
->vector_elements
);
647 this->type
= brw_type_for_base_type(type
);
650 src_reg::src_reg(class vec4_visitor
*v
, const struct glsl_type
*type
, int size
)
657 this->reg
= v
->alloc
.allocate(type_size_vec4(type
) * size
);
659 this->swizzle
= BRW_SWIZZLE_NOOP
;
661 this->type
= brw_type_for_base_type(type
);
664 dst_reg::dst_reg(class vec4_visitor
*v
, const struct glsl_type
*type
)
669 this->reg
= v
->alloc
.allocate(type_size_vec4(type
));
671 if (type
->is_array() || type
->is_record()) {
672 this->writemask
= WRITEMASK_XYZW
;
674 this->writemask
= (1 << type
->vector_elements
) - 1;
677 this->type
= brw_type_for_base_type(type
);
681 vec4_visitor::emit_minmax(enum brw_conditional_mod conditionalmod
, dst_reg dst
,
682 src_reg src0
, src_reg src1
)
684 vec4_instruction
*inst
;
686 if (devinfo
->gen
>= 6) {
687 inst
= emit(BRW_OPCODE_SEL
, dst
, src0
, src1
);
688 inst
->conditional_mod
= conditionalmod
;
690 emit(CMP(dst
, src0
, src1
, conditionalmod
));
692 inst
= emit(BRW_OPCODE_SEL
, dst
, src0
, src1
);
693 inst
->predicate
= BRW_PREDICATE_NORMAL
;
700 vec4_visitor::emit_lrp(const dst_reg
&dst
,
701 const src_reg
&x
, const src_reg
&y
, const src_reg
&a
)
703 if (devinfo
->gen
>= 6) {
704 /* Note that the instruction's argument order is reversed from GLSL
707 return emit(LRP(dst
, fix_3src_operand(a
), fix_3src_operand(y
),
708 fix_3src_operand(x
)));
710 /* Earlier generations don't support three source operations, so we
711 * need to emit x*(1-a) + y*a.
713 dst_reg y_times_a
= dst_reg(this, glsl_type::vec4_type
);
714 dst_reg one_minus_a
= dst_reg(this, glsl_type::vec4_type
);
715 dst_reg x_times_one_minus_a
= dst_reg(this, glsl_type::vec4_type
);
716 y_times_a
.writemask
= dst
.writemask
;
717 one_minus_a
.writemask
= dst
.writemask
;
718 x_times_one_minus_a
.writemask
= dst
.writemask
;
720 emit(MUL(y_times_a
, y
, a
));
721 emit(ADD(one_minus_a
, negate(a
), src_reg(1.0f
)));
722 emit(MUL(x_times_one_minus_a
, x
, src_reg(one_minus_a
)));
723 return emit(ADD(dst
, src_reg(x_times_one_minus_a
), src_reg(y_times_a
)));
728 * Emits the instructions needed to perform a pull constant load. before_block
729 * and before_inst can be NULL in which case the instruction will be appended
730 * to the end of the instruction list.
733 vec4_visitor::emit_pull_constant_load_reg(dst_reg dst
,
736 bblock_t
*before_block
,
737 vec4_instruction
*before_inst
)
739 assert((before_inst
== NULL
&& before_block
== NULL
) ||
740 (before_inst
&& before_block
));
742 vec4_instruction
*pull
;
744 if (devinfo
->gen
>= 9) {
745 /* Gen9+ needs a message header in order to use SIMD4x2 mode */
746 src_reg
header(this, glsl_type::uvec4_type
, 2);
749 vec4_instruction(VS_OPCODE_SET_SIMD4X2_HEADER_GEN9
,
753 emit_before(before_block
, before_inst
, pull
);
757 dst_reg index_reg
= retype(offset(dst_reg(header
), 1),
759 pull
= MOV(writemask(index_reg
, WRITEMASK_X
), offset_reg
);
762 emit_before(before_block
, before_inst
, pull
);
766 pull
= new(mem_ctx
) vec4_instruction(VS_OPCODE_PULL_CONSTANT_LOAD_GEN7
,
771 pull
->header_size
= 1;
772 } else if (devinfo
->gen
>= 7) {
773 dst_reg grf_offset
= dst_reg(this, glsl_type::int_type
);
775 grf_offset
.type
= offset_reg
.type
;
777 pull
= MOV(grf_offset
, offset_reg
);
780 emit_before(before_block
, before_inst
, pull
);
784 pull
= new(mem_ctx
) vec4_instruction(VS_OPCODE_PULL_CONSTANT_LOAD_GEN7
,
787 src_reg(grf_offset
));
790 pull
= new(mem_ctx
) vec4_instruction(VS_OPCODE_PULL_CONSTANT_LOAD
,
794 pull
->base_mrf
= FIRST_PULL_LOAD_MRF(devinfo
->gen
) + 1;
799 emit_before(before_block
, before_inst
, pull
);
805 vec4_visitor::emit_uniformize(const src_reg
&src
)
807 const src_reg
chan_index(this, glsl_type::uint_type
);
808 const dst_reg dst
= retype(dst_reg(this, glsl_type::uint_type
),
811 emit(SHADER_OPCODE_FIND_LIVE_CHANNEL
, dst_reg(chan_index
))
812 ->force_writemask_all
= true;
813 emit(SHADER_OPCODE_BROADCAST
, dst
, src
, chan_index
)
814 ->force_writemask_all
= true;
820 vec4_visitor::emit_mcs_fetch(const glsl_type
*coordinate_type
,
821 src_reg coordinate
, src_reg sampler
)
823 vec4_instruction
*inst
=
824 new(mem_ctx
) vec4_instruction(SHADER_OPCODE_TXF_MCS
,
825 dst_reg(this, glsl_type::uvec4_type
));
827 inst
->src
[1] = sampler
;
831 if (devinfo
->gen
>= 9) {
832 /* Gen9+ needs a message header in order to use SIMD4x2 mode */
833 vec4_instruction
*header_inst
= new(mem_ctx
)
834 vec4_instruction(VS_OPCODE_SET_SIMD4X2_HEADER_GEN9
,
835 dst_reg(MRF
, inst
->base_mrf
));
840 inst
->header_size
= 1;
841 param_base
= inst
->base_mrf
+ 1;
844 param_base
= inst
->base_mrf
;
847 /* parameters are: u, v, r, lod; lod will always be zero due to api restrictions */
848 int coord_mask
= (1 << coordinate_type
->vector_elements
) - 1;
849 int zero_mask
= 0xf & ~coord_mask
;
851 emit(MOV(dst_reg(MRF
, param_base
, coordinate_type
, coord_mask
),
854 emit(MOV(dst_reg(MRF
, param_base
, coordinate_type
, zero_mask
),
858 return src_reg(inst
->dst
);
862 vec4_visitor::is_high_sampler(src_reg sampler
)
864 if (devinfo
->gen
< 8 && !devinfo
->is_haswell
)
867 return sampler
.file
!= IMM
|| sampler
.fixed_hw_reg
.dw1
.ud
>= 16;
871 vec4_visitor::emit_texture(ir_texture_opcode op
,
873 const glsl_type
*dest_type
,
875 int coord_components
,
876 src_reg shadow_comparitor
,
877 src_reg lod
, src_reg lod2
,
878 src_reg sample_index
,
879 uint32_t constant_offset
,
880 src_reg offset_value
,
888 case ir_tex
: opcode
= SHADER_OPCODE_TXL
; break;
889 case ir_txl
: opcode
= SHADER_OPCODE_TXL
; break;
890 case ir_txd
: opcode
= SHADER_OPCODE_TXD
; break;
891 case ir_txf
: opcode
= SHADER_OPCODE_TXF
; break;
892 case ir_txf_ms
: opcode
= SHADER_OPCODE_TXF_CMS
; break;
893 case ir_txs
: opcode
= SHADER_OPCODE_TXS
; break;
894 case ir_tg4
: opcode
= offset_value
.file
!= BAD_FILE
895 ? SHADER_OPCODE_TG4_OFFSET
: SHADER_OPCODE_TG4
; break;
896 case ir_query_levels
: opcode
= SHADER_OPCODE_TXS
; break;
897 case ir_texture_samples
: opcode
= SHADER_OPCODE_SAMPLEINFO
; break;
899 unreachable("TXB is not valid for vertex shaders.");
901 unreachable("LOD is not valid for vertex shaders.");
903 unreachable("Unrecognized tex op");
906 vec4_instruction
*inst
= new(mem_ctx
) vec4_instruction(
907 opcode
, dst_reg(this, dest_type
));
909 inst
->offset
= constant_offset
;
911 /* The message header is necessary for:
913 * - Gen9+ for selecting SIMD4x2
915 * - Gather channel selection
916 * - Sampler indices too large to fit in a 4-bit value.
917 * - Sampleinfo message - takes no parameters, but mlen = 0 is illegal
920 (devinfo
->gen
< 5 || devinfo
->gen
>= 9 ||
921 inst
->offset
!= 0 || op
== ir_tg4
||
922 op
== ir_texture_samples
||
923 is_high_sampler(sampler_reg
)) ? 1 : 0;
925 inst
->mlen
= inst
->header_size
;
926 inst
->dst
.writemask
= WRITEMASK_XYZW
;
927 inst
->shadow_compare
= shadow_comparitor
.file
!= BAD_FILE
;
929 inst
->src
[1] = sampler_reg
;
931 /* MRF for the first parameter */
932 int param_base
= inst
->base_mrf
+ inst
->header_size
;
934 if (op
== ir_txs
|| op
== ir_query_levels
) {
935 int writemask
= devinfo
->gen
== 4 ? WRITEMASK_W
: WRITEMASK_X
;
936 emit(MOV(dst_reg(MRF
, param_base
, lod
.type
, writemask
), lod
));
938 } else if (op
== ir_texture_samples
) {
939 inst
->dst
.writemask
= WRITEMASK_X
;
941 /* Load the coordinate */
942 /* FINISHME: gl_clamp_mask and saturate */
943 int coord_mask
= (1 << coord_components
) - 1;
944 int zero_mask
= 0xf & ~coord_mask
;
946 emit(MOV(dst_reg(MRF
, param_base
, coordinate
.type
, coord_mask
),
950 if (zero_mask
!= 0) {
951 emit(MOV(dst_reg(MRF
, param_base
, coordinate
.type
, zero_mask
),
954 /* Load the shadow comparitor */
955 if (shadow_comparitor
.file
!= BAD_FILE
&& op
!= ir_txd
&& (op
!= ir_tg4
|| offset_value
.file
== BAD_FILE
)) {
956 emit(MOV(dst_reg(MRF
, param_base
+ 1, shadow_comparitor
.type
,
962 /* Load the LOD info */
963 if (op
== ir_tex
|| op
== ir_txl
) {
965 if (devinfo
->gen
>= 5) {
966 mrf
= param_base
+ 1;
967 if (shadow_comparitor
.file
!= BAD_FILE
) {
968 writemask
= WRITEMASK_Y
;
969 /* mlen already incremented */
971 writemask
= WRITEMASK_X
;
974 } else /* devinfo->gen == 4 */ {
976 writemask
= WRITEMASK_W
;
978 emit(MOV(dst_reg(MRF
, mrf
, lod
.type
, writemask
), lod
));
979 } else if (op
== ir_txf
) {
980 emit(MOV(dst_reg(MRF
, param_base
, lod
.type
, WRITEMASK_W
), lod
));
981 } else if (op
== ir_txf_ms
) {
982 emit(MOV(dst_reg(MRF
, param_base
+ 1, sample_index
.type
, WRITEMASK_X
),
984 if (devinfo
->gen
>= 7) {
985 /* MCS data is in the first channel of `mcs`, but we need to get it into
986 * the .y channel of the second vec4 of params, so replicate .x across
987 * the whole vec4 and then mask off everything except .y
989 mcs
.swizzle
= BRW_SWIZZLE_XXXX
;
990 emit(MOV(dst_reg(MRF
, param_base
+ 1, glsl_type::uint_type
, WRITEMASK_Y
),
994 } else if (op
== ir_txd
) {
995 const brw_reg_type type
= lod
.type
;
997 if (devinfo
->gen
>= 5) {
998 lod
.swizzle
= BRW_SWIZZLE4(SWIZZLE_X
,SWIZZLE_X
,SWIZZLE_Y
,SWIZZLE_Y
);
999 lod2
.swizzle
= BRW_SWIZZLE4(SWIZZLE_X
,SWIZZLE_X
,SWIZZLE_Y
,SWIZZLE_Y
);
1000 emit(MOV(dst_reg(MRF
, param_base
+ 1, type
, WRITEMASK_XZ
), lod
));
1001 emit(MOV(dst_reg(MRF
, param_base
+ 1, type
, WRITEMASK_YW
), lod2
));
1004 if (dest_type
->vector_elements
== 3 || shadow_comparitor
.file
!= BAD_FILE
) {
1005 lod
.swizzle
= BRW_SWIZZLE_ZZZZ
;
1006 lod2
.swizzle
= BRW_SWIZZLE_ZZZZ
;
1007 emit(MOV(dst_reg(MRF
, param_base
+ 2, type
, WRITEMASK_X
), lod
));
1008 emit(MOV(dst_reg(MRF
, param_base
+ 2, type
, WRITEMASK_Y
), lod2
));
1011 if (shadow_comparitor
.file
!= BAD_FILE
) {
1012 emit(MOV(dst_reg(MRF
, param_base
+ 2,
1013 shadow_comparitor
.type
, WRITEMASK_Z
),
1014 shadow_comparitor
));
1017 } else /* devinfo->gen == 4 */ {
1018 emit(MOV(dst_reg(MRF
, param_base
+ 1, type
, WRITEMASK_XYZ
), lod
));
1019 emit(MOV(dst_reg(MRF
, param_base
+ 2, type
, WRITEMASK_XYZ
), lod2
));
1022 } else if (op
== ir_tg4
&& offset_value
.file
!= BAD_FILE
) {
1023 if (shadow_comparitor
.file
!= BAD_FILE
) {
1024 emit(MOV(dst_reg(MRF
, param_base
, shadow_comparitor
.type
, WRITEMASK_W
),
1025 shadow_comparitor
));
1028 emit(MOV(dst_reg(MRF
, param_base
+ 1, glsl_type::ivec2_type
, WRITEMASK_XY
),
1036 /* fixup num layers (z) for cube arrays: hardware returns faces * layers;
1037 * spec requires layers.
1039 if (op
== ir_txs
&& is_cube_array
) {
1040 emit_math(SHADER_OPCODE_INT_QUOTIENT
,
1041 writemask(inst
->dst
, WRITEMASK_Z
),
1042 src_reg(inst
->dst
), src_reg(6));
1045 if (devinfo
->gen
== 6 && op
== ir_tg4
) {
1046 emit_gen6_gather_wa(key_tex
->gen6_gather_wa
[sampler
], inst
->dst
);
1049 swizzle_result(op
, dest
,
1050 src_reg(inst
->dst
), sampler
, dest_type
);
1054 * Apply workarounds for Gen6 gather with UINT/SINT
1057 vec4_visitor::emit_gen6_gather_wa(uint8_t wa
, dst_reg dst
)
1062 int width
= (wa
& WA_8BIT
) ? 8 : 16;
1063 dst_reg dst_f
= dst
;
1064 dst_f
.type
= BRW_REGISTER_TYPE_F
;
1066 /* Convert from UNORM to UINT */
1067 emit(MUL(dst_f
, src_reg(dst_f
), src_reg((float)((1 << width
) - 1))));
1068 emit(MOV(dst
, src_reg(dst_f
)));
1071 /* Reinterpret the UINT value as a signed INT value by
1072 * shifting the sign bit into place, then shifting back
1075 emit(SHL(dst
, src_reg(dst
), src_reg(32 - width
)));
1076 emit(ASR(dst
, src_reg(dst
), src_reg(32 - width
)));
1081 * Set up the gather channel based on the swizzle, for gather4.
1084 vec4_visitor::gather_channel(unsigned gather_component
, uint32_t sampler
)
1086 int swiz
= GET_SWZ(key_tex
->swizzles
[sampler
], gather_component
);
1088 case SWIZZLE_X
: return 0;
1090 /* gather4 sampler is broken for green channel on RG32F --
1091 * we must ask for blue instead.
1093 if (key_tex
->gather_channel_quirk_mask
& (1 << sampler
))
1096 case SWIZZLE_Z
: return 2;
1097 case SWIZZLE_W
: return 3;
1099 unreachable("Not reached"); /* zero, one swizzles handled already */
1104 vec4_visitor::swizzle_result(ir_texture_opcode op
, dst_reg dest
,
1105 src_reg orig_val
, uint32_t sampler
,
1106 const glsl_type
*dest_type
)
1108 int s
= key_tex
->swizzles
[sampler
];
1110 dst_reg swizzled_result
= dest
;
1112 if (op
== ir_query_levels
) {
1113 /* # levels is in .w */
1114 orig_val
.swizzle
= BRW_SWIZZLE4(SWIZZLE_W
, SWIZZLE_W
, SWIZZLE_W
, SWIZZLE_W
);
1115 emit(MOV(swizzled_result
, orig_val
));
1119 if (op
== ir_txs
|| dest_type
== glsl_type::float_type
1120 || s
== SWIZZLE_NOOP
|| op
== ir_tg4
) {
1121 emit(MOV(swizzled_result
, orig_val
));
1126 int zero_mask
= 0, one_mask
= 0, copy_mask
= 0;
1127 int swizzle
[4] = {0};
1129 for (int i
= 0; i
< 4; i
++) {
1130 switch (GET_SWZ(s
, i
)) {
1132 zero_mask
|= (1 << i
);
1135 one_mask
|= (1 << i
);
1138 copy_mask
|= (1 << i
);
1139 swizzle
[i
] = GET_SWZ(s
, i
);
1145 orig_val
.swizzle
= BRW_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
1146 swizzled_result
.writemask
= copy_mask
;
1147 emit(MOV(swizzled_result
, orig_val
));
1151 swizzled_result
.writemask
= zero_mask
;
1152 emit(MOV(swizzled_result
, src_reg(0.0f
)));
1156 swizzled_result
.writemask
= one_mask
;
1157 emit(MOV(swizzled_result
, src_reg(1.0f
)));
1162 vec4_visitor::gs_emit_vertex(int stream_id
)
1164 unreachable("not reached");
1168 vec4_visitor::gs_end_primitive()
1170 unreachable("not reached");
1174 vec4_visitor::emit_untyped_atomic(unsigned atomic_op
, unsigned surf_index
,
1175 dst_reg dst
, src_reg offset
,
1176 src_reg src0
, src_reg src1
)
1180 /* Set the atomic operation offset. */
1181 emit(MOV(brw_writemask(brw_uvec_mrf(8, mlen
, 0), WRITEMASK_X
), offset
));
1184 /* Set the atomic operation arguments. */
1185 if (src0
.file
!= BAD_FILE
) {
1186 emit(MOV(brw_writemask(brw_uvec_mrf(8, mlen
, 0), WRITEMASK_X
), src0
));
1190 if (src1
.file
!= BAD_FILE
) {
1191 emit(MOV(brw_writemask(brw_uvec_mrf(8, mlen
, 0), WRITEMASK_X
), src1
));
1195 /* Emit the instruction. Note that this maps to the normal SIMD8
1196 * untyped atomic message on Ivy Bridge, but that's OK because
1197 * unused channels will be masked out.
1199 vec4_instruction
*inst
= emit(SHADER_OPCODE_UNTYPED_ATOMIC
, dst
,
1201 src_reg(surf_index
), src_reg(atomic_op
));
1206 vec4_visitor::emit_untyped_surface_read(unsigned surf_index
, dst_reg dst
,
1209 /* Set the surface read offset. */
1210 emit(MOV(brw_writemask(brw_uvec_mrf(8, 0, 0), WRITEMASK_X
), offset
));
1212 /* Emit the instruction. Note that this maps to the normal SIMD8
1213 * untyped surface read message, but that's OK because unused
1214 * channels will be masked out.
1216 vec4_instruction
*inst
= emit(SHADER_OPCODE_UNTYPED_SURFACE_READ
, dst
,
1218 src_reg(surf_index
), src_reg(1));
1223 vec4_visitor::emit_ndc_computation()
1225 /* Get the position */
1226 src_reg pos
= src_reg(output_reg
[VARYING_SLOT_POS
]);
1228 /* Build ndc coords, which are (x/w, y/w, z/w, 1/w) */
1229 dst_reg ndc
= dst_reg(this, glsl_type::vec4_type
);
1230 output_reg
[BRW_VARYING_SLOT_NDC
] = ndc
;
1232 current_annotation
= "NDC";
1233 dst_reg ndc_w
= ndc
;
1234 ndc_w
.writemask
= WRITEMASK_W
;
1235 src_reg pos_w
= pos
;
1236 pos_w
.swizzle
= BRW_SWIZZLE4(SWIZZLE_W
, SWIZZLE_W
, SWIZZLE_W
, SWIZZLE_W
);
1237 emit_math(SHADER_OPCODE_RCP
, ndc_w
, pos_w
);
1239 dst_reg ndc_xyz
= ndc
;
1240 ndc_xyz
.writemask
= WRITEMASK_XYZ
;
1242 emit(MUL(ndc_xyz
, pos
, src_reg(ndc_w
)));
1246 vec4_visitor::emit_psiz_and_flags(dst_reg reg
)
1248 if (devinfo
->gen
< 6 &&
1249 ((prog_data
->vue_map
.slots_valid
& VARYING_BIT_PSIZ
) ||
1250 output_reg
[VARYING_SLOT_CLIP_DIST0
].file
!= BAD_FILE
||
1251 devinfo
->has_negative_rhw_bug
)) {
1252 dst_reg header1
= dst_reg(this, glsl_type::uvec4_type
);
1253 dst_reg header1_w
= header1
;
1254 header1_w
.writemask
= WRITEMASK_W
;
1256 emit(MOV(header1
, 0u));
1258 if (prog_data
->vue_map
.slots_valid
& VARYING_BIT_PSIZ
) {
1259 src_reg psiz
= src_reg(output_reg
[VARYING_SLOT_PSIZ
]);
1261 current_annotation
= "Point size";
1262 emit(MUL(header1_w
, psiz
, src_reg((float)(1 << 11))));
1263 emit(AND(header1_w
, src_reg(header1_w
), 0x7ff << 8));
1266 if (output_reg
[VARYING_SLOT_CLIP_DIST0
].file
!= BAD_FILE
) {
1267 current_annotation
= "Clipping flags";
1268 dst_reg flags0
= dst_reg(this, glsl_type::uint_type
);
1269 dst_reg flags1
= dst_reg(this, glsl_type::uint_type
);
1271 emit(CMP(dst_null_f(), src_reg(output_reg
[VARYING_SLOT_CLIP_DIST0
]), src_reg(0.0f
), BRW_CONDITIONAL_L
));
1272 emit(VS_OPCODE_UNPACK_FLAGS_SIMD4X2
, flags0
, src_reg(0));
1273 emit(OR(header1_w
, src_reg(header1_w
), src_reg(flags0
)));
1275 emit(CMP(dst_null_f(), src_reg(output_reg
[VARYING_SLOT_CLIP_DIST1
]), src_reg(0.0f
), BRW_CONDITIONAL_L
));
1276 emit(VS_OPCODE_UNPACK_FLAGS_SIMD4X2
, flags1
, src_reg(0));
1277 emit(SHL(flags1
, src_reg(flags1
), src_reg(4)));
1278 emit(OR(header1_w
, src_reg(header1_w
), src_reg(flags1
)));
1281 /* i965 clipping workaround:
1282 * 1) Test for -ve rhw
1284 * set ndc = (0,0,0,0)
1287 * Later, clipping will detect ucp[6] and ensure the primitive is
1288 * clipped against all fixed planes.
1290 if (devinfo
->has_negative_rhw_bug
) {
1291 src_reg ndc_w
= src_reg(output_reg
[BRW_VARYING_SLOT_NDC
]);
1292 ndc_w
.swizzle
= BRW_SWIZZLE_WWWW
;
1293 emit(CMP(dst_null_f(), ndc_w
, src_reg(0.0f
), BRW_CONDITIONAL_L
));
1294 vec4_instruction
*inst
;
1295 inst
= emit(OR(header1_w
, src_reg(header1_w
), src_reg(1u << 6)));
1296 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1297 output_reg
[BRW_VARYING_SLOT_NDC
].type
= BRW_REGISTER_TYPE_F
;
1298 inst
= emit(MOV(output_reg
[BRW_VARYING_SLOT_NDC
], src_reg(0.0f
)));
1299 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1302 emit(MOV(retype(reg
, BRW_REGISTER_TYPE_UD
), src_reg(header1
)));
1303 } else if (devinfo
->gen
< 6) {
1304 emit(MOV(retype(reg
, BRW_REGISTER_TYPE_UD
), 0u));
1306 emit(MOV(retype(reg
, BRW_REGISTER_TYPE_D
), src_reg(0)));
1307 if (prog_data
->vue_map
.slots_valid
& VARYING_BIT_PSIZ
) {
1308 dst_reg reg_w
= reg
;
1309 reg_w
.writemask
= WRITEMASK_W
;
1310 src_reg reg_as_src
= src_reg(output_reg
[VARYING_SLOT_PSIZ
]);
1311 reg_as_src
.type
= reg_w
.type
;
1312 reg_as_src
.swizzle
= brw_swizzle_for_size(1);
1313 emit(MOV(reg_w
, reg_as_src
));
1315 if (prog_data
->vue_map
.slots_valid
& VARYING_BIT_LAYER
) {
1316 dst_reg reg_y
= reg
;
1317 reg_y
.writemask
= WRITEMASK_Y
;
1318 reg_y
.type
= BRW_REGISTER_TYPE_D
;
1319 output_reg
[VARYING_SLOT_LAYER
].type
= reg_y
.type
;
1320 emit(MOV(reg_y
, src_reg(output_reg
[VARYING_SLOT_LAYER
])));
1322 if (prog_data
->vue_map
.slots_valid
& VARYING_BIT_VIEWPORT
) {
1323 dst_reg reg_z
= reg
;
1324 reg_z
.writemask
= WRITEMASK_Z
;
1325 reg_z
.type
= BRW_REGISTER_TYPE_D
;
1326 output_reg
[VARYING_SLOT_VIEWPORT
].type
= reg_z
.type
;
1327 emit(MOV(reg_z
, src_reg(output_reg
[VARYING_SLOT_VIEWPORT
])));
1333 vec4_visitor::emit_generic_urb_slot(dst_reg reg
, int varying
)
1335 assert(varying
< VARYING_SLOT_MAX
);
1336 assert(output_reg
[varying
].type
== reg
.type
);
1337 current_annotation
= output_reg_annotation
[varying
];
1338 /* Copy the register, saturating if necessary */
1339 return emit(MOV(reg
, src_reg(output_reg
[varying
])));
1343 vec4_visitor::emit_urb_slot(dst_reg reg
, int varying
)
1345 reg
.type
= BRW_REGISTER_TYPE_F
;
1346 output_reg
[varying
].type
= reg
.type
;
1349 case VARYING_SLOT_PSIZ
:
1351 /* PSIZ is always in slot 0, and is coupled with other flags. */
1352 current_annotation
= "indices, point width, clip flags";
1353 emit_psiz_and_flags(reg
);
1356 case BRW_VARYING_SLOT_NDC
:
1357 current_annotation
= "NDC";
1358 emit(MOV(reg
, src_reg(output_reg
[BRW_VARYING_SLOT_NDC
])));
1360 case VARYING_SLOT_POS
:
1361 current_annotation
= "gl_Position";
1362 emit(MOV(reg
, src_reg(output_reg
[VARYING_SLOT_POS
])));
1364 case VARYING_SLOT_EDGE
:
1365 /* This is present when doing unfilled polygons. We're supposed to copy
1366 * the edge flag from the user-provided vertex array
1367 * (glEdgeFlagPointer), or otherwise we'll copy from the current value
1368 * of that attribute (starts as 1.0f). This is then used in clipping to
1369 * determine which edges should be drawn as wireframe.
1371 current_annotation
= "edge flag";
1372 emit(MOV(reg
, src_reg(dst_reg(ATTR
, VERT_ATTRIB_EDGEFLAG
,
1373 glsl_type::float_type
, WRITEMASK_XYZW
))));
1375 case BRW_VARYING_SLOT_PAD
:
1376 /* No need to write to this slot */
1379 emit_generic_urb_slot(reg
, varying
);
1385 align_interleaved_urb_mlen(const struct brw_device_info
*devinfo
, int mlen
)
1387 if (devinfo
->gen
>= 6) {
1388 /* URB data written (does not include the message header reg) must
1389 * be a multiple of 256 bits, or 2 VS registers. See vol5c.5,
1390 * section 5.4.3.2.2: URB_INTERLEAVED.
1392 * URB entries are allocated on a multiple of 1024 bits, so an
1393 * extra 128 bits written here to make the end align to 256 is
1396 if ((mlen
% 2) != 1)
1405 * Generates the VUE payload plus the necessary URB write instructions to
1408 * The VUE layout is documented in Volume 2a.
1411 vec4_visitor::emit_vertex()
1413 /* MRF 0 is reserved for the debugger, so start with message header
1418 /* In the process of generating our URB write message contents, we
1419 * may need to unspill a register or load from an array. Those
1420 * reads would use MRFs 14-15.
1422 int max_usable_mrf
= FIRST_SPILL_MRF(devinfo
->gen
);
1424 /* The following assertion verifies that max_usable_mrf causes an
1425 * even-numbered amount of URB write data, which will meet gen6's
1426 * requirements for length alignment.
1428 assert ((max_usable_mrf
- base_mrf
) % 2 == 0);
1430 /* First mrf is the g0-based message header containing URB handles and
1433 emit_urb_write_header(mrf
++);
1435 if (devinfo
->gen
< 6) {
1436 emit_ndc_computation();
1439 /* We may need to split this up into several URB writes, so do them in a
1443 bool complete
= false;
1445 /* URB offset is in URB row increments, and each of our MRFs is half of
1446 * one of those, since we're doing interleaved writes.
1448 int offset
= slot
/ 2;
1451 for (; slot
< prog_data
->vue_map
.num_slots
; ++slot
) {
1452 emit_urb_slot(dst_reg(MRF
, mrf
++),
1453 prog_data
->vue_map
.slot_to_varying
[slot
]);
1455 /* If this was max_usable_mrf, we can't fit anything more into this
1456 * URB WRITE. Same thing if we reached the maximum length available.
1458 if (mrf
> max_usable_mrf
||
1459 align_interleaved_urb_mlen(devinfo
, mrf
- base_mrf
+ 1) > BRW_MAX_MSG_LENGTH
) {
1465 complete
= slot
>= prog_data
->vue_map
.num_slots
;
1466 current_annotation
= "URB write";
1467 vec4_instruction
*inst
= emit_urb_write_opcode(complete
);
1468 inst
->base_mrf
= base_mrf
;
1469 inst
->mlen
= align_interleaved_urb_mlen(devinfo
, mrf
- base_mrf
);
1470 inst
->offset
+= offset
;
1476 vec4_visitor::get_scratch_offset(bblock_t
*block
, vec4_instruction
*inst
,
1477 src_reg
*reladdr
, int reg_offset
)
1479 /* Because we store the values to scratch interleaved like our
1480 * vertex data, we need to scale the vec4 index by 2.
1482 int message_header_scale
= 2;
1484 /* Pre-gen6, the message header uses byte offsets instead of vec4
1485 * (16-byte) offset units.
1487 if (devinfo
->gen
< 6)
1488 message_header_scale
*= 16;
1491 src_reg index
= src_reg(this, glsl_type::int_type
);
1493 emit_before(block
, inst
, ADD(dst_reg(index
), *reladdr
,
1494 src_reg(reg_offset
)));
1495 emit_before(block
, inst
, MUL(dst_reg(index
), index
,
1496 src_reg(message_header_scale
)));
1500 return src_reg(reg_offset
* message_header_scale
);
1505 vec4_visitor::get_pull_constant_offset(bblock_t
* block
, vec4_instruction
*inst
,
1506 src_reg
*reladdr
, int reg_offset
)
1509 src_reg index
= src_reg(this, glsl_type::int_type
);
1511 emit_before(block
, inst
, ADD(dst_reg(index
), *reladdr
,
1512 src_reg(reg_offset
)));
1514 /* Pre-gen6, the message header uses byte offsets instead of vec4
1515 * (16-byte) offset units.
1517 if (devinfo
->gen
< 6) {
1518 emit_before(block
, inst
, MUL(dst_reg(index
), index
, src_reg(16)));
1522 } else if (devinfo
->gen
>= 8) {
1523 /* Store the offset in a GRF so we can send-from-GRF. */
1524 src_reg offset
= src_reg(this, glsl_type::int_type
);
1525 emit_before(block
, inst
, MOV(dst_reg(offset
), src_reg(reg_offset
)));
1528 int message_header_scale
= devinfo
->gen
< 6 ? 16 : 1;
1529 return src_reg(reg_offset
* message_header_scale
);
1534 * Emits an instruction before @inst to load the value named by @orig_src
1535 * from scratch space at @base_offset to @temp.
1537 * @base_offset is measured in 32-byte units (the size of a register).
1540 vec4_visitor::emit_scratch_read(bblock_t
*block
, vec4_instruction
*inst
,
1541 dst_reg temp
, src_reg orig_src
,
1544 int reg_offset
= base_offset
+ orig_src
.reg_offset
;
1545 src_reg index
= get_scratch_offset(block
, inst
, orig_src
.reladdr
,
1548 emit_before(block
, inst
, SCRATCH_READ(temp
, index
));
1552 * Emits an instruction after @inst to store the value to be written
1553 * to @orig_dst to scratch space at @base_offset, from @temp.
1555 * @base_offset is measured in 32-byte units (the size of a register).
1558 vec4_visitor::emit_scratch_write(bblock_t
*block
, vec4_instruction
*inst
,
1561 int reg_offset
= base_offset
+ inst
->dst
.reg_offset
;
1562 src_reg index
= get_scratch_offset(block
, inst
, inst
->dst
.reladdr
,
1565 /* Create a temporary register to store *inst's result in.
1567 * We have to be careful in MOVing from our temporary result register in
1568 * the scratch write. If we swizzle from channels of the temporary that
1569 * weren't initialized, it will confuse live interval analysis, which will
1570 * make spilling fail to make progress.
1572 const src_reg temp
= swizzle(retype(src_reg(this, glsl_type::vec4_type
),
1574 brw_swizzle_for_mask(inst
->dst
.writemask
));
1575 dst_reg dst
= dst_reg(brw_writemask(brw_vec8_grf(0, 0),
1576 inst
->dst
.writemask
));
1577 vec4_instruction
*write
= SCRATCH_WRITE(dst
, temp
, index
);
1578 if (inst
->opcode
!= BRW_OPCODE_SEL
)
1579 write
->predicate
= inst
->predicate
;
1580 write
->ir
= inst
->ir
;
1581 write
->annotation
= inst
->annotation
;
1582 inst
->insert_after(block
, write
);
1584 inst
->dst
.file
= temp
.file
;
1585 inst
->dst
.reg
= temp
.reg
;
1586 inst
->dst
.reg_offset
= temp
.reg_offset
;
1587 inst
->dst
.reladdr
= NULL
;
1591 * Checks if \p src and/or \p src.reladdr require a scratch read, and if so,
1592 * adds the scratch read(s) before \p inst. The function also checks for
1593 * recursive reladdr scratch accesses, issuing the corresponding scratch
1594 * loads and rewriting reladdr references accordingly.
1596 * \return \p src if it did not require a scratch load, otherwise, the
1597 * register holding the result of the scratch load that the caller should
1598 * use to rewrite src.
1601 vec4_visitor::emit_resolve_reladdr(int scratch_loc
[], bblock_t
*block
,
1602 vec4_instruction
*inst
, src_reg src
)
1604 /* Resolve recursive reladdr scratch access by calling ourselves
1608 *src
.reladdr
= emit_resolve_reladdr(scratch_loc
, block
, inst
,
1611 /* Now handle scratch access on src */
1612 if (src
.file
== GRF
&& scratch_loc
[src
.reg
] != -1) {
1613 dst_reg temp
= dst_reg(this, glsl_type::vec4_type
);
1614 emit_scratch_read(block
, inst
, temp
, src
, scratch_loc
[src
.reg
]);
1616 src
.reg_offset
= temp
.reg_offset
;
1624 * We can't generally support array access in GRF space, because a
1625 * single instruction's destination can only span 2 contiguous
1626 * registers. So, we send all GRF arrays that get variable index
1627 * access to scratch space.
1630 vec4_visitor::move_grf_array_access_to_scratch()
1632 int scratch_loc
[this->alloc
.count
];
1633 memset(scratch_loc
, -1, sizeof(scratch_loc
));
1635 /* First, calculate the set of virtual GRFs that need to be punted
1636 * to scratch due to having any array access on them, and where in
1639 foreach_block_and_inst(block
, vec4_instruction
, inst
, cfg
) {
1640 if (inst
->dst
.file
== GRF
&& inst
->dst
.reladdr
) {
1641 if (scratch_loc
[inst
->dst
.reg
] == -1) {
1642 scratch_loc
[inst
->dst
.reg
] = last_scratch
;
1643 last_scratch
+= this->alloc
.sizes
[inst
->dst
.reg
];
1646 for (src_reg
*iter
= inst
->dst
.reladdr
;
1648 iter
= iter
->reladdr
) {
1649 if (iter
->file
== GRF
&& scratch_loc
[iter
->reg
] == -1) {
1650 scratch_loc
[iter
->reg
] = last_scratch
;
1651 last_scratch
+= this->alloc
.sizes
[iter
->reg
];
1656 for (int i
= 0 ; i
< 3; i
++) {
1657 for (src_reg
*iter
= &inst
->src
[i
];
1659 iter
= iter
->reladdr
) {
1660 if (iter
->file
== GRF
&& scratch_loc
[iter
->reg
] == -1) {
1661 scratch_loc
[iter
->reg
] = last_scratch
;
1662 last_scratch
+= this->alloc
.sizes
[iter
->reg
];
1668 /* Now, for anything that will be accessed through scratch, rewrite
1669 * it to load/store. Note that this is a _safe list walk, because
1670 * we may generate a new scratch_write instruction after the one
1673 foreach_block_and_inst_safe(block
, vec4_instruction
, inst
, cfg
) {
1674 /* Set up the annotation tracking for new generated instructions. */
1676 current_annotation
= inst
->annotation
;
1678 /* First handle scratch access on the dst. Notice we have to handle
1679 * the case where the dst's reladdr also points to scratch space.
1681 if (inst
->dst
.reladdr
)
1682 *inst
->dst
.reladdr
= emit_resolve_reladdr(scratch_loc
, block
, inst
,
1683 *inst
->dst
.reladdr
);
1685 /* Now that we have handled any (possibly recursive) reladdr scratch
1686 * accesses for dst we can safely do the scratch write for dst itself
1688 if (inst
->dst
.file
== GRF
&& scratch_loc
[inst
->dst
.reg
] != -1)
1689 emit_scratch_write(block
, inst
, scratch_loc
[inst
->dst
.reg
]);
1691 /* Now handle scratch access on any src. In this case, since inst->src[i]
1692 * already is a src_reg, we can just call emit_resolve_reladdr with
1693 * inst->src[i] and it will take care of handling scratch loads for
1694 * both src and src.reladdr (recursively).
1696 for (int i
= 0 ; i
< 3; i
++) {
1697 inst
->src
[i
] = emit_resolve_reladdr(scratch_loc
, block
, inst
,
1704 * Emits an instruction before @inst to load the value named by @orig_src
1705 * from the pull constant buffer (surface) at @base_offset to @temp.
1708 vec4_visitor::emit_pull_constant_load(bblock_t
*block
, vec4_instruction
*inst
,
1709 dst_reg temp
, src_reg orig_src
,
1712 int reg_offset
= base_offset
+ orig_src
.reg_offset
;
1713 src_reg index
= src_reg(prog_data
->base
.binding_table
.pull_constants_start
);
1714 src_reg offset
= get_pull_constant_offset(block
, inst
, orig_src
.reladdr
,
1717 emit_pull_constant_load_reg(temp
,
1724 * Implements array access of uniforms by inserting a
1725 * PULL_CONSTANT_LOAD instruction.
1727 * Unlike temporary GRF array access (where we don't support it due to
1728 * the difficulty of doing relative addressing on instruction
1729 * destinations), we could potentially do array access of uniforms
1730 * that were loaded in GRF space as push constants. In real-world
1731 * usage we've seen, though, the arrays being used are always larger
1732 * than we could load as push constants, so just always move all
1733 * uniform array access out to a pull constant buffer.
1736 vec4_visitor::move_uniform_array_access_to_pull_constants()
1738 int pull_constant_loc
[this->uniforms
];
1739 memset(pull_constant_loc
, -1, sizeof(pull_constant_loc
));
1740 bool nested_reladdr
;
1742 /* Walk through and find array access of uniforms. Put a copy of that
1743 * uniform in the pull constant buffer.
1745 * Note that we don't move constant-indexed accesses to arrays. No
1746 * testing has been done of the performance impact of this choice.
1749 nested_reladdr
= false;
1751 foreach_block_and_inst_safe(block
, vec4_instruction
, inst
, cfg
) {
1752 for (int i
= 0 ; i
< 3; i
++) {
1753 if (inst
->src
[i
].file
!= UNIFORM
|| !inst
->src
[i
].reladdr
)
1756 int uniform
= inst
->src
[i
].reg
;
1758 if (inst
->src
[i
].reladdr
->reladdr
)
1759 nested_reladdr
= true; /* will need another pass */
1761 /* If this array isn't already present in the pull constant buffer,
1764 if (pull_constant_loc
[uniform
] == -1) {
1765 const gl_constant_value
**values
=
1766 &stage_prog_data
->param
[uniform
* 4];
1768 pull_constant_loc
[uniform
] = stage_prog_data
->nr_pull_params
/ 4;
1770 assert(uniform
< uniform_array_size
);
1771 for (int j
= 0; j
< uniform_size
[uniform
] * 4; j
++) {
1772 stage_prog_data
->pull_param
[stage_prog_data
->nr_pull_params
++]
1777 /* Set up the annotation tracking for new generated instructions. */
1779 current_annotation
= inst
->annotation
;
1781 dst_reg temp
= dst_reg(this, glsl_type::vec4_type
);
1783 emit_pull_constant_load(block
, inst
, temp
, inst
->src
[i
],
1784 pull_constant_loc
[uniform
]);
1786 inst
->src
[i
].file
= temp
.file
;
1787 inst
->src
[i
].reg
= temp
.reg
;
1788 inst
->src
[i
].reg_offset
= temp
.reg_offset
;
1789 inst
->src
[i
].reladdr
= NULL
;
1792 } while (nested_reladdr
);
1794 /* Now there are no accesses of the UNIFORM file with a reladdr, so
1795 * no need to track them as larger-than-vec4 objects. This will be
1796 * relied on in cutting out unused uniform vectors from push
1799 split_uniform_registers();
1803 vec4_visitor::resolve_ud_negate(src_reg
*reg
)
1805 if (reg
->type
!= BRW_REGISTER_TYPE_UD
||
1809 src_reg temp
= src_reg(this, glsl_type::uvec4_type
);
1810 emit(BRW_OPCODE_MOV
, dst_reg(temp
), *reg
);
1814 vec4_visitor::vec4_visitor(const struct brw_compiler
*compiler
,
1816 const struct brw_sampler_prog_key_data
*key_tex
,
1817 struct brw_vue_prog_data
*prog_data
,
1818 const nir_shader
*shader
,
1821 int shader_time_index
)
1822 : backend_shader(compiler
, log_data
, mem_ctx
, shader
, &prog_data
->base
),
1824 prog_data(prog_data
),
1826 first_non_payload_grf(0),
1827 need_all_constants_in_pull_buffer(false),
1828 no_spills(no_spills
),
1829 shader_time_index(shader_time_index
),
1832 this->failed
= false;
1834 this->base_ir
= NULL
;
1835 this->current_annotation
= NULL
;
1836 memset(this->output_reg_annotation
, 0, sizeof(this->output_reg_annotation
));
1838 this->virtual_grf_start
= NULL
;
1839 this->virtual_grf_end
= NULL
;
1840 this->live_intervals
= NULL
;
1842 this->max_grf
= devinfo
->gen
>= 7 ? GEN7_MRF_HACK_START
: BRW_MAX_GRF
;
1846 /* Initialize uniform_array_size to at least 1 because pre-gen6 VS requires
1847 * at least one. See setup_uniforms() in brw_vec4.cpp.
1849 this->uniform_array_size
= 1;
1851 this->uniform_array_size
=
1852 MAX2(DIV_ROUND_UP(stage_prog_data
->nr_params
, 4), 1);
1855 this->uniform_size
= rzalloc_array(mem_ctx
, int, this->uniform_array_size
);
1858 vec4_visitor::~vec4_visitor()
1864 vec4_visitor::fail(const char *format
, ...)
1874 va_start(va
, format
);
1875 msg
= ralloc_vasprintf(mem_ctx
, format
, va
);
1877 msg
= ralloc_asprintf(mem_ctx
, "%s compile failed: %s\n", stage_abbrev
, msg
);
1879 this->fail_msg
= msg
;
1881 if (debug_enabled
) {
1882 fprintf(stderr
, "%s", msg
);
1886 } /* namespace brw */