2 * Copyright © 2010 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 * This file drives the GLSL IR -> LIR translation, contains the
27 * optimizations on the LIR, and drives the generation of native code
31 #include <sys/types.h>
33 #include "util/hash_table.h"
34 #include "main/macros.h"
35 #include "main/shaderobj.h"
36 #include "main/fbobject.h"
37 #include "program/prog_parameter.h"
38 #include "program/prog_print.h"
39 #include "util/register_allocate.h"
40 #include "program/hash_table.h"
41 #include "brw_context.h"
46 #include "brw_vec4_gs_visitor.h"
48 #include "brw_dead_control_flow.h"
49 #include "main/uniforms.h"
50 #include "brw_fs_live_variables.h"
51 #include "glsl/nir/glsl_types.h"
52 #include "program/sampler.h"
57 fs_inst::init(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
58 const fs_reg
*src
, unsigned sources
)
60 memset(this, 0, sizeof(*this));
62 this->src
= new fs_reg
[MAX2(sources
, 3)];
63 for (unsigned i
= 0; i
< sources
; i
++)
64 this->src
[i
] = src
[i
];
66 this->opcode
= opcode
;
68 this->sources
= sources
;
69 this->exec_size
= exec_size
;
71 assert(dst
.file
!= IMM
&& dst
.file
!= UNIFORM
);
73 assert(this->exec_size
!= 0);
75 this->conditional_mod
= BRW_CONDITIONAL_NONE
;
77 /* This will be the case for almost all instructions. */
83 this->regs_written
= DIV_ROUND_UP(dst
.component_size(exec_size
),
87 this->regs_written
= 0;
91 unreachable("Invalid destination register file");
94 this->writes_accumulator
= false;
99 init(BRW_OPCODE_NOP
, 8, dst
, NULL
, 0);
102 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
)
104 init(opcode
, exec_size
, reg_undef
, NULL
, 0);
107 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
)
109 init(opcode
, exec_size
, dst
, NULL
, 0);
112 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
115 const fs_reg src
[1] = { src0
};
116 init(opcode
, exec_size
, dst
, src
, 1);
119 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
120 const fs_reg
&src0
, const fs_reg
&src1
)
122 const fs_reg src
[2] = { src0
, src1
};
123 init(opcode
, exec_size
, dst
, src
, 2);
126 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
127 const fs_reg
&src0
, const fs_reg
&src1
, const fs_reg
&src2
)
129 const fs_reg src
[3] = { src0
, src1
, src2
};
130 init(opcode
, exec_size
, dst
, src
, 3);
133 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_width
, const fs_reg
&dst
,
134 const fs_reg src
[], unsigned sources
)
136 init(opcode
, exec_width
, dst
, src
, sources
);
139 fs_inst::fs_inst(const fs_inst
&that
)
141 memcpy(this, &that
, sizeof(that
));
143 this->src
= new fs_reg
[MAX2(that
.sources
, 3)];
145 for (unsigned i
= 0; i
< that
.sources
; i
++)
146 this->src
[i
] = that
.src
[i
];
155 fs_inst::resize_sources(uint8_t num_sources
)
157 if (this->sources
!= num_sources
) {
158 fs_reg
*src
= new fs_reg
[MAX2(num_sources
, 3)];
160 for (unsigned i
= 0; i
< MIN2(this->sources
, num_sources
); ++i
)
161 src
[i
] = this->src
[i
];
165 this->sources
= num_sources
;
170 fs_visitor::VARYING_PULL_CONSTANT_LOAD(const fs_builder
&bld
,
172 const fs_reg
&surf_index
,
173 const fs_reg
&varying_offset
,
174 uint32_t const_offset
)
176 /* We have our constant surface use a pitch of 4 bytes, so our index can
177 * be any component of a vector, and then we load 4 contiguous
178 * components starting from that.
180 * We break down the const_offset to a portion added to the variable
181 * offset and a portion done using reg_offset, which means that if you
182 * have GLSL using something like "uniform vec4 a[20]; gl_FragColor =
183 * a[i]", we'll temporarily generate 4 vec4 loads from offset i * 4, and
184 * CSE can later notice that those loads are all the same and eliminate
185 * the redundant ones.
187 fs_reg vec4_offset
= vgrf(glsl_type::int_type
);
188 bld
.ADD(vec4_offset
, varying_offset
, fs_reg(const_offset
& ~3));
191 if (devinfo
->gen
== 4 && bld
.dispatch_width() == 8) {
192 /* Pre-gen5, we can either use a SIMD8 message that requires (header,
193 * u, v, r) as parameters, or we can just use the SIMD16 message
194 * consisting of (header, u). We choose the second, at the cost of a
195 * longer return length.
201 if (devinfo
->gen
>= 7)
202 op
= FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7
;
204 op
= FS_OPCODE_VARYING_PULL_CONSTANT_LOAD
;
206 int regs_written
= 4 * (bld
.dispatch_width() / 8) * scale
;
207 fs_reg vec4_result
= fs_reg(GRF
, alloc
.allocate(regs_written
), dst
.type
);
208 fs_inst
*inst
= bld
.emit(op
, vec4_result
, surf_index
, vec4_offset
);
209 inst
->regs_written
= regs_written
;
211 if (devinfo
->gen
< 7) {
212 inst
->base_mrf
= FIRST_PULL_LOAD_MRF(devinfo
->gen
);
213 inst
->header_size
= 1;
214 if (devinfo
->gen
== 4)
217 inst
->mlen
= 1 + bld
.dispatch_width() / 8;
220 bld
.MOV(dst
, offset(vec4_result
, bld
, (const_offset
& 3) * scale
));
224 * A helper for MOV generation for fixing up broken hardware SEND dependency
228 fs_visitor::DEP_RESOLVE_MOV(const fs_builder
&bld
, int grf
)
230 /* The caller always wants uncompressed to emit the minimal extra
231 * dependencies, and to avoid having to deal with aligning its regs to 2.
233 const fs_builder ubld
= bld
.annotate("send dependency resolve")
236 ubld
.MOV(ubld
.null_reg_f(), fs_reg(GRF
, grf
, BRW_REGISTER_TYPE_F
));
240 fs_inst::equals(fs_inst
*inst
) const
242 return (opcode
== inst
->opcode
&&
243 dst
.equals(inst
->dst
) &&
244 src
[0].equals(inst
->src
[0]) &&
245 src
[1].equals(inst
->src
[1]) &&
246 src
[2].equals(inst
->src
[2]) &&
247 saturate
== inst
->saturate
&&
248 predicate
== inst
->predicate
&&
249 conditional_mod
== inst
->conditional_mod
&&
250 mlen
== inst
->mlen
&&
251 base_mrf
== inst
->base_mrf
&&
252 target
== inst
->target
&&
254 header_size
== inst
->header_size
&&
255 shadow_compare
== inst
->shadow_compare
&&
256 exec_size
== inst
->exec_size
&&
257 offset
== inst
->offset
);
261 fs_inst::overwrites_reg(const fs_reg
®
) const
263 return reg
.in_range(dst
, regs_written
);
267 fs_inst::is_send_from_grf() const
270 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7
:
271 case SHADER_OPCODE_SHADER_TIME_ADD
:
272 case FS_OPCODE_INTERPOLATE_AT_CENTROID
:
273 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
274 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
275 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
276 case SHADER_OPCODE_UNTYPED_ATOMIC
:
277 case SHADER_OPCODE_UNTYPED_SURFACE_READ
:
278 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE
:
279 case SHADER_OPCODE_TYPED_ATOMIC
:
280 case SHADER_OPCODE_TYPED_SURFACE_READ
:
281 case SHADER_OPCODE_TYPED_SURFACE_WRITE
:
282 case SHADER_OPCODE_URB_WRITE_SIMD8
:
283 case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
284 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
:
285 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
286 case SHADER_OPCODE_URB_READ_SIMD8
:
288 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
289 return src
[1].file
== GRF
;
290 case FS_OPCODE_FB_WRITE
:
291 return src
[0].file
== GRF
;
294 return src
[0].file
== GRF
;
301 fs_inst::is_copy_payload(const brw::simple_allocator
&grf_alloc
) const
303 if (this->opcode
!= SHADER_OPCODE_LOAD_PAYLOAD
)
306 fs_reg reg
= this->src
[0];
307 if (reg
.file
!= GRF
|| reg
.reg_offset
!= 0 || reg
.stride
== 0)
310 if (grf_alloc
.sizes
[reg
.reg
] != this->regs_written
)
313 for (int i
= 0; i
< this->sources
; i
++) {
314 reg
.type
= this->src
[i
].type
;
315 if (!this->src
[i
].equals(reg
))
318 if (i
< this->header_size
) {
321 reg
.reg_offset
+= this->exec_size
/ 8;
329 fs_inst::can_do_source_mods(const struct brw_device_info
*devinfo
)
331 if (devinfo
->gen
== 6 && is_math())
334 if (is_send_from_grf())
337 if (!backend_instruction::can_do_source_mods())
344 fs_inst::can_change_types() const
346 return dst
.type
== src
[0].type
&&
347 !src
[0].abs
&& !src
[0].negate
&& !saturate
&&
348 (opcode
== BRW_OPCODE_MOV
||
349 (opcode
== BRW_OPCODE_SEL
&&
350 dst
.type
== src
[1].type
&&
351 predicate
!= BRW_PREDICATE_NONE
&&
352 !src
[1].abs
&& !src
[1].negate
));
356 fs_inst::has_side_effects() const
358 return this->eot
|| backend_instruction::has_side_effects();
364 memset(this, 0, sizeof(*this));
368 /** Generic unset register constructor. */
372 this->file
= BAD_FILE
;
375 /** Immediate value constructor. */
376 fs_reg::fs_reg(float f
)
380 this->type
= BRW_REGISTER_TYPE_F
;
382 this->fixed_hw_reg
.dw1
.f
= f
;
385 /** Immediate value constructor. */
386 fs_reg::fs_reg(int32_t i
)
390 this->type
= BRW_REGISTER_TYPE_D
;
392 this->fixed_hw_reg
.dw1
.d
= i
;
395 /** Immediate value constructor. */
396 fs_reg::fs_reg(uint32_t u
)
400 this->type
= BRW_REGISTER_TYPE_UD
;
402 this->fixed_hw_reg
.dw1
.ud
= u
;
405 /** Vector float immediate value constructor. */
406 fs_reg::fs_reg(uint8_t vf
[4])
410 this->type
= BRW_REGISTER_TYPE_VF
;
411 memcpy(&this->fixed_hw_reg
.dw1
.ud
, vf
, sizeof(unsigned));
414 /** Vector float immediate value constructor. */
415 fs_reg::fs_reg(uint8_t vf0
, uint8_t vf1
, uint8_t vf2
, uint8_t vf3
)
419 this->type
= BRW_REGISTER_TYPE_VF
;
420 this->fixed_hw_reg
.dw1
.ud
= (vf0
<< 0) |
426 /** Fixed brw_reg. */
427 fs_reg::fs_reg(struct brw_reg fixed_hw_reg
)
431 this->fixed_hw_reg
= fixed_hw_reg
;
432 this->type
= fixed_hw_reg
.type
;
436 fs_reg::equals(const fs_reg
&r
) const
438 return (file
== r
.file
&&
440 reg_offset
== r
.reg_offset
&&
441 subreg_offset
== r
.subreg_offset
&&
443 negate
== r
.negate
&&
445 !reladdr
&& !r
.reladdr
&&
446 ((file
!= HW_REG
&& file
!= IMM
) ||
447 memcmp(&fixed_hw_reg
, &r
.fixed_hw_reg
,
448 sizeof(fixed_hw_reg
)) == 0) &&
453 fs_reg::set_smear(unsigned subreg
)
455 assert(file
!= HW_REG
&& file
!= IMM
);
456 subreg_offset
= subreg
* type_sz(type
);
462 fs_reg::is_contiguous() const
468 fs_reg::component_size(unsigned width
) const
470 const unsigned stride
= (file
!= HW_REG
? this->stride
:
471 fixed_hw_reg
.hstride
== 0 ? 0 :
472 1 << (fixed_hw_reg
.hstride
- 1));
473 return MAX2(width
* stride
, 1) * type_sz(type
);
477 type_size_scalar(const struct glsl_type
*type
)
479 unsigned int size
, i
;
481 switch (type
->base_type
) {
484 case GLSL_TYPE_FLOAT
:
486 return type
->components();
487 case GLSL_TYPE_ARRAY
:
488 return type_size_scalar(type
->fields
.array
) * type
->length
;
489 case GLSL_TYPE_STRUCT
:
491 for (i
= 0; i
< type
->length
; i
++) {
492 size
+= type_size_scalar(type
->fields
.structure
[i
].type
);
495 case GLSL_TYPE_SAMPLER
:
496 /* Samplers take up no register space, since they're baked in at
500 case GLSL_TYPE_ATOMIC_UINT
:
502 case GLSL_TYPE_SUBROUTINE
:
504 case GLSL_TYPE_IMAGE
:
505 return BRW_IMAGE_PARAM_SIZE
;
507 case GLSL_TYPE_ERROR
:
508 case GLSL_TYPE_INTERFACE
:
509 case GLSL_TYPE_DOUBLE
:
510 unreachable("not reached");
517 * Create a MOV to read the timestamp register.
519 * The caller is responsible for emitting the MOV. The return value is
520 * the destination of the MOV, with extra parameters set.
523 fs_visitor::get_timestamp(const fs_builder
&bld
)
525 assert(devinfo
->gen
>= 7);
527 fs_reg ts
= fs_reg(retype(brw_vec4_reg(BRW_ARCHITECTURE_REGISTER_FILE
,
530 BRW_REGISTER_TYPE_UD
));
532 fs_reg dst
= fs_reg(GRF
, alloc
.allocate(1), BRW_REGISTER_TYPE_UD
);
534 /* We want to read the 3 fields we care about even if it's not enabled in
537 bld
.group(4, 0).exec_all().MOV(dst
, ts
);
543 fs_visitor::emit_shader_time_begin()
545 shader_start_time
= get_timestamp(bld
.annotate("shader time start"));
547 /* We want only the low 32 bits of the timestamp. Since it's running
548 * at the GPU clock rate of ~1.2ghz, it will roll over every ~3 seconds,
549 * which is plenty of time for our purposes. It is identical across the
550 * EUs, but since it's tracking GPU core speed it will increment at a
551 * varying rate as render P-states change.
553 shader_start_time
.set_smear(0);
557 fs_visitor::emit_shader_time_end()
559 /* Insert our code just before the final SEND with EOT. */
560 exec_node
*end
= this->instructions
.get_tail();
561 assert(end
&& ((fs_inst
*) end
)->eot
);
562 const fs_builder ibld
= bld
.annotate("shader time end")
563 .exec_all().at(NULL
, end
);
565 fs_reg shader_end_time
= get_timestamp(ibld
);
567 /* We only use the low 32 bits of the timestamp - see
568 * emit_shader_time_begin()).
570 * We could also check if render P-states have changed (or anything
571 * else that might disrupt timing) by setting smear to 2 and checking if
572 * that field is != 0.
574 shader_end_time
.set_smear(0);
576 /* Check that there weren't any timestamp reset events (assuming these
577 * were the only two timestamp reads that happened).
579 fs_reg reset
= shader_end_time
;
581 set_condmod(BRW_CONDITIONAL_Z
,
582 ibld
.AND(ibld
.null_reg_ud(), reset
, fs_reg(1u)));
583 ibld
.IF(BRW_PREDICATE_NORMAL
);
585 fs_reg start
= shader_start_time
;
587 fs_reg diff
= fs_reg(GRF
, alloc
.allocate(1), BRW_REGISTER_TYPE_UD
);
590 const fs_builder cbld
= ibld
.group(1, 0);
591 cbld
.group(1, 0).ADD(diff
, start
, shader_end_time
);
593 /* If there were no instructions between the two timestamp gets, the diff
594 * is 2 cycles. Remove that overhead, so I can forget about that when
595 * trying to determine the time taken for single instructions.
597 cbld
.ADD(diff
, diff
, fs_reg(-2u));
598 SHADER_TIME_ADD(cbld
, 0, diff
);
599 SHADER_TIME_ADD(cbld
, 1, fs_reg(1u));
600 ibld
.emit(BRW_OPCODE_ELSE
);
601 SHADER_TIME_ADD(cbld
, 2, fs_reg(1u));
602 ibld
.emit(BRW_OPCODE_ENDIF
);
606 fs_visitor::SHADER_TIME_ADD(const fs_builder
&bld
,
607 int shader_time_subindex
,
610 int index
= shader_time_index
* 3 + shader_time_subindex
;
611 fs_reg offset
= fs_reg(index
* SHADER_TIME_STRIDE
);
614 if (dispatch_width
== 8)
615 payload
= vgrf(glsl_type::uvec2_type
);
617 payload
= vgrf(glsl_type::uint_type
);
619 bld
.emit(SHADER_OPCODE_SHADER_TIME_ADD
, fs_reg(), payload
, offset
, value
);
623 fs_visitor::vfail(const char *format
, va_list va
)
632 msg
= ralloc_vasprintf(mem_ctx
, format
, va
);
633 msg
= ralloc_asprintf(mem_ctx
, "%s compile failed: %s\n", stage_abbrev
, msg
);
635 this->fail_msg
= msg
;
638 fprintf(stderr
, "%s", msg
);
643 fs_visitor::fail(const char *format
, ...)
647 va_start(va
, format
);
653 * Mark this program as impossible to compile in SIMD16 mode.
655 * During the SIMD8 compile (which happens first), we can detect and flag
656 * things that are unsupported in SIMD16 mode, so the compiler can skip
657 * the SIMD16 compile altogether.
659 * During a SIMD16 compile (if one happens anyway), this just calls fail().
662 fs_visitor::no16(const char *msg
)
664 if (dispatch_width
== 16) {
667 simd16_unsupported
= true;
669 compiler
->shader_perf_log(log_data
,
670 "SIMD16 shader failed to compile: %s", msg
);
675 * Returns true if the instruction has a flag that means it won't
676 * update an entire destination register.
678 * For example, dead code elimination and live variable analysis want to know
679 * when a write to a variable screens off any preceding values that were in
683 fs_inst::is_partial_write() const
685 return ((this->predicate
&& this->opcode
!= BRW_OPCODE_SEL
) ||
686 (this->exec_size
* type_sz(this->dst
.type
)) < 32 ||
687 !this->dst
.is_contiguous());
691 fs_inst::components_read(unsigned i
) const
694 case FS_OPCODE_LINTERP
:
700 case FS_OPCODE_PIXEL_X
:
701 case FS_OPCODE_PIXEL_Y
:
705 case FS_OPCODE_FB_WRITE_LOGICAL
:
706 assert(src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].file
== IMM
);
707 /* First/second FB write color. */
709 return src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].fixed_hw_reg
.dw1
.ud
;
713 case SHADER_OPCODE_TEX_LOGICAL
:
714 case SHADER_OPCODE_TXD_LOGICAL
:
715 case SHADER_OPCODE_TXF_LOGICAL
:
716 case SHADER_OPCODE_TXL_LOGICAL
:
717 case SHADER_OPCODE_TXS_LOGICAL
:
718 case FS_OPCODE_TXB_LOGICAL
:
719 case SHADER_OPCODE_TXF_CMS_LOGICAL
:
720 case SHADER_OPCODE_TXF_CMS_W_LOGICAL
:
721 case SHADER_OPCODE_TXF_UMS_LOGICAL
:
722 case SHADER_OPCODE_TXF_MCS_LOGICAL
:
723 case SHADER_OPCODE_LOD_LOGICAL
:
724 case SHADER_OPCODE_TG4_LOGICAL
:
725 case SHADER_OPCODE_TG4_OFFSET_LOGICAL
:
726 assert(src
[8].file
== IMM
&& src
[9].file
== IMM
);
727 /* Texture coordinates. */
729 return src
[8].fixed_hw_reg
.dw1
.ud
;
730 /* Texture derivatives. */
731 else if ((i
== 2 || i
== 3) && opcode
== SHADER_OPCODE_TXD_LOGICAL
)
732 return src
[9].fixed_hw_reg
.dw1
.ud
;
733 /* Texture offset. */
737 else if (i
== 5 && opcode
== SHADER_OPCODE_TXF_CMS_W_LOGICAL
)
742 case SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
:
743 case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
:
744 assert(src
[3].file
== IMM
);
745 /* Surface coordinates. */
747 return src
[3].fixed_hw_reg
.dw1
.ud
;
748 /* Surface operation source (ignored for reads). */
754 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
:
755 case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
:
756 assert(src
[3].file
== IMM
&&
758 /* Surface coordinates. */
760 return src
[3].fixed_hw_reg
.dw1
.ud
;
761 /* Surface operation source. */
763 return src
[4].fixed_hw_reg
.dw1
.ud
;
767 case SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL
:
768 case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
: {
769 assert(src
[3].file
== IMM
&&
771 const unsigned op
= src
[4].fixed_hw_reg
.dw1
.ud
;
772 /* Surface coordinates. */
774 return src
[3].fixed_hw_reg
.dw1
.ud
;
775 /* Surface operation source. */
776 else if (i
== 1 && op
== BRW_AOP_CMPWR
)
778 else if (i
== 1 && (op
== BRW_AOP_INC
|| op
== BRW_AOP_DEC
||
779 op
== BRW_AOP_PREDEC
))
791 fs_inst::regs_read(int arg
) const
794 case FS_OPCODE_FB_WRITE
:
795 case SHADER_OPCODE_URB_WRITE_SIMD8
:
796 case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
797 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
:
798 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
799 case SHADER_OPCODE_URB_READ_SIMD8
:
800 case SHADER_OPCODE_UNTYPED_ATOMIC
:
801 case SHADER_OPCODE_UNTYPED_SURFACE_READ
:
802 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE
:
803 case SHADER_OPCODE_TYPED_ATOMIC
:
804 case SHADER_OPCODE_TYPED_SURFACE_READ
:
805 case SHADER_OPCODE_TYPED_SURFACE_WRITE
:
806 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
811 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7
:
812 /* The payload is actually stored in src1 */
817 case FS_OPCODE_LINTERP
:
822 case SHADER_OPCODE_LOAD_PAYLOAD
:
823 if (arg
< this->header_size
)
827 case CS_OPCODE_CS_TERMINATE
:
828 case SHADER_OPCODE_BARRIER
:
832 if (is_tex() && arg
== 0 && src
[0].file
== GRF
)
837 switch (src
[arg
].file
) {
846 return DIV_ROUND_UP(components_read(arg
) *
847 src
[arg
].component_size(exec_size
),
850 unreachable("MRF registers are not allowed as sources");
856 fs_inst::reads_flag() const
862 fs_inst::writes_flag() const
864 return (conditional_mod
&& (opcode
!= BRW_OPCODE_SEL
&&
865 opcode
!= BRW_OPCODE_IF
&&
866 opcode
!= BRW_OPCODE_WHILE
)) ||
867 opcode
== FS_OPCODE_MOV_DISPATCH_TO_FLAGS
;
871 * Returns how many MRFs an FS opcode will write over.
873 * Note that this is not the 0 or 1 implied writes in an actual gen
874 * instruction -- the FS opcodes often generate MOVs in addition.
877 fs_visitor::implied_mrf_writes(fs_inst
*inst
)
882 if (inst
->base_mrf
== -1)
885 switch (inst
->opcode
) {
886 case SHADER_OPCODE_RCP
:
887 case SHADER_OPCODE_RSQ
:
888 case SHADER_OPCODE_SQRT
:
889 case SHADER_OPCODE_EXP2
:
890 case SHADER_OPCODE_LOG2
:
891 case SHADER_OPCODE_SIN
:
892 case SHADER_OPCODE_COS
:
893 return 1 * dispatch_width
/ 8;
894 case SHADER_OPCODE_POW
:
895 case SHADER_OPCODE_INT_QUOTIENT
:
896 case SHADER_OPCODE_INT_REMAINDER
:
897 return 2 * dispatch_width
/ 8;
898 case SHADER_OPCODE_TEX
:
900 case SHADER_OPCODE_TXD
:
901 case SHADER_OPCODE_TXF
:
902 case SHADER_OPCODE_TXF_CMS
:
903 case SHADER_OPCODE_TXF_CMS_W
:
904 case SHADER_OPCODE_TXF_MCS
:
905 case SHADER_OPCODE_TG4
:
906 case SHADER_OPCODE_TG4_OFFSET
:
907 case SHADER_OPCODE_TXL
:
908 case SHADER_OPCODE_TXS
:
909 case SHADER_OPCODE_LOD
:
910 case SHADER_OPCODE_SAMPLEINFO
:
912 case FS_OPCODE_FB_WRITE
:
914 case FS_OPCODE_GET_BUFFER_SIZE
:
915 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
916 case SHADER_OPCODE_GEN4_SCRATCH_READ
:
918 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD
:
920 case SHADER_OPCODE_GEN4_SCRATCH_WRITE
:
922 case SHADER_OPCODE_UNTYPED_ATOMIC
:
923 case SHADER_OPCODE_UNTYPED_SURFACE_READ
:
924 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE
:
925 case SHADER_OPCODE_TYPED_ATOMIC
:
926 case SHADER_OPCODE_TYPED_SURFACE_READ
:
927 case SHADER_OPCODE_TYPED_SURFACE_WRITE
:
928 case SHADER_OPCODE_URB_WRITE_SIMD8
:
929 case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
930 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
:
931 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
932 case FS_OPCODE_INTERPOLATE_AT_CENTROID
:
933 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
934 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
935 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
938 unreachable("not reached");
943 fs_visitor::vgrf(const glsl_type
*const type
)
945 int reg_width
= dispatch_width
/ 8;
946 return fs_reg(GRF
, alloc
.allocate(type_size_scalar(type
) * reg_width
),
947 brw_type_for_base_type(type
));
950 /** Fixed HW reg constructor. */
951 fs_reg::fs_reg(enum register_file file
, int reg
)
956 this->type
= BRW_REGISTER_TYPE_F
;
957 this->stride
= (file
== UNIFORM
? 0 : 1);
960 /** Fixed HW reg constructor. */
961 fs_reg::fs_reg(enum register_file file
, int reg
, enum brw_reg_type type
)
967 this->stride
= (file
== UNIFORM
? 0 : 1);
970 /* For SIMD16, we need to follow from the uniform setup of SIMD8 dispatch.
971 * This brings in those uniform definitions
974 fs_visitor::import_uniforms(fs_visitor
*v
)
976 this->push_constant_loc
= v
->push_constant_loc
;
977 this->pull_constant_loc
= v
->pull_constant_loc
;
978 this->uniforms
= v
->uniforms
;
979 this->param_size
= v
->param_size
;
983 fs_visitor::emit_fragcoord_interpolation(bool pixel_center_integer
,
984 bool origin_upper_left
)
986 assert(stage
== MESA_SHADER_FRAGMENT
);
987 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
988 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::vec4_type
));
990 bool flip
= !origin_upper_left
^ key
->render_to_fbo
;
993 if (pixel_center_integer
) {
994 bld
.MOV(wpos
, this->pixel_x
);
996 bld
.ADD(wpos
, this->pixel_x
, fs_reg(0.5f
));
998 wpos
= offset(wpos
, bld
, 1);
1000 /* gl_FragCoord.y */
1001 if (!flip
&& pixel_center_integer
) {
1002 bld
.MOV(wpos
, this->pixel_y
);
1004 fs_reg pixel_y
= this->pixel_y
;
1005 float offset
= (pixel_center_integer
? 0.0f
: 0.5f
);
1008 pixel_y
.negate
= true;
1009 offset
+= key
->drawable_height
- 1.0f
;
1012 bld
.ADD(wpos
, pixel_y
, fs_reg(offset
));
1014 wpos
= offset(wpos
, bld
, 1);
1016 /* gl_FragCoord.z */
1017 if (devinfo
->gen
>= 6) {
1018 bld
.MOV(wpos
, fs_reg(brw_vec8_grf(payload
.source_depth_reg
, 0)));
1020 bld
.emit(FS_OPCODE_LINTERP
, wpos
,
1021 this->delta_xy
[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
],
1022 interp_reg(VARYING_SLOT_POS
, 2));
1024 wpos
= offset(wpos
, bld
, 1);
1026 /* gl_FragCoord.w: Already set up in emit_interpolation */
1027 bld
.MOV(wpos
, this->wpos_w
);
1033 fs_visitor::emit_linterp(const fs_reg
&attr
, const fs_reg
&interp
,
1034 glsl_interp_qualifier interpolation_mode
,
1035 bool is_centroid
, bool is_sample
)
1037 brw_wm_barycentric_interp_mode barycoord_mode
;
1038 if (devinfo
->gen
>= 6) {
1040 if (interpolation_mode
== INTERP_QUALIFIER_SMOOTH
)
1041 barycoord_mode
= BRW_WM_PERSPECTIVE_CENTROID_BARYCENTRIC
;
1043 barycoord_mode
= BRW_WM_NONPERSPECTIVE_CENTROID_BARYCENTRIC
;
1044 } else if (is_sample
) {
1045 if (interpolation_mode
== INTERP_QUALIFIER_SMOOTH
)
1046 barycoord_mode
= BRW_WM_PERSPECTIVE_SAMPLE_BARYCENTRIC
;
1048 barycoord_mode
= BRW_WM_NONPERSPECTIVE_SAMPLE_BARYCENTRIC
;
1050 if (interpolation_mode
== INTERP_QUALIFIER_SMOOTH
)
1051 barycoord_mode
= BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
;
1053 barycoord_mode
= BRW_WM_NONPERSPECTIVE_PIXEL_BARYCENTRIC
;
1056 /* On Ironlake and below, there is only one interpolation mode.
1057 * Centroid interpolation doesn't mean anything on this hardware --
1058 * there is no multisampling.
1060 barycoord_mode
= BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
;
1062 return bld
.emit(FS_OPCODE_LINTERP
, attr
,
1063 this->delta_xy
[barycoord_mode
], interp
);
1067 fs_visitor::emit_general_interpolation(fs_reg attr
, const char *name
,
1068 const glsl_type
*type
,
1069 glsl_interp_qualifier interpolation_mode
,
1070 int location
, bool mod_centroid
,
1073 attr
.type
= brw_type_for_base_type(type
->get_scalar_type());
1075 assert(stage
== MESA_SHADER_FRAGMENT
);
1076 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
1077 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1079 unsigned int array_elements
;
1081 if (type
->is_array()) {
1082 array_elements
= type
->arrays_of_arrays_size();
1083 if (array_elements
== 0) {
1084 fail("dereferenced array '%s' has length 0\n", name
);
1086 type
= type
->without_array();
1091 if (interpolation_mode
== INTERP_QUALIFIER_NONE
) {
1093 location
== VARYING_SLOT_COL0
|| location
== VARYING_SLOT_COL1
;
1094 if (key
->flat_shade
&& is_gl_Color
) {
1095 interpolation_mode
= INTERP_QUALIFIER_FLAT
;
1097 interpolation_mode
= INTERP_QUALIFIER_SMOOTH
;
1101 for (unsigned int i
= 0; i
< array_elements
; i
++) {
1102 for (unsigned int j
= 0; j
< type
->matrix_columns
; j
++) {
1103 if (prog_data
->urb_setup
[location
] == -1) {
1104 /* If there's no incoming setup data for this slot, don't
1105 * emit interpolation for it.
1107 attr
= offset(attr
, bld
, type
->vector_elements
);
1112 if (interpolation_mode
== INTERP_QUALIFIER_FLAT
) {
1113 /* Constant interpolation (flat shading) case. The SF has
1114 * handed us defined values in only the constant offset
1115 * field of the setup reg.
1117 for (unsigned int k
= 0; k
< type
->vector_elements
; k
++) {
1118 struct brw_reg interp
= interp_reg(location
, k
);
1119 interp
= suboffset(interp
, 3);
1120 interp
.type
= attr
.type
;
1121 bld
.emit(FS_OPCODE_CINTERP
, attr
, fs_reg(interp
));
1122 attr
= offset(attr
, bld
, 1);
1125 /* Smooth/noperspective interpolation case. */
1126 for (unsigned int k
= 0; k
< type
->vector_elements
; k
++) {
1127 struct brw_reg interp
= interp_reg(location
, k
);
1128 if (devinfo
->needs_unlit_centroid_workaround
&& mod_centroid
) {
1129 /* Get the pixel/sample mask into f0 so that we know
1130 * which pixels are lit. Then, for each channel that is
1131 * unlit, replace the centroid data with non-centroid
1134 bld
.emit(FS_OPCODE_MOV_DISPATCH_TO_FLAGS
);
1137 inst
= emit_linterp(attr
, fs_reg(interp
), interpolation_mode
,
1139 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1140 inst
->predicate_inverse
= true;
1141 if (devinfo
->has_pln
)
1142 inst
->no_dd_clear
= true;
1144 inst
= emit_linterp(attr
, fs_reg(interp
), interpolation_mode
,
1145 mod_centroid
&& !key
->persample_shading
,
1146 mod_sample
|| key
->persample_shading
);
1147 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1148 inst
->predicate_inverse
= false;
1149 if (devinfo
->has_pln
)
1150 inst
->no_dd_check
= true;
1153 emit_linterp(attr
, fs_reg(interp
), interpolation_mode
,
1154 mod_centroid
&& !key
->persample_shading
,
1155 mod_sample
|| key
->persample_shading
);
1157 if (devinfo
->gen
< 6 && interpolation_mode
== INTERP_QUALIFIER_SMOOTH
) {
1158 bld
.MUL(attr
, attr
, this->pixel_w
);
1160 attr
= offset(attr
, bld
, 1);
1170 fs_visitor::emit_frontfacing_interpolation()
1172 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::bool_type
));
1174 if (devinfo
->gen
>= 6) {
1175 /* Bit 15 of g0.0 is 0 if the polygon is front facing. We want to create
1176 * a boolean result from this (~0/true or 0/false).
1178 * We can use the fact that bit 15 is the MSB of g0.0:W to accomplish
1179 * this task in only one instruction:
1180 * - a negation source modifier will flip the bit; and
1181 * - a W -> D type conversion will sign extend the bit into the high
1182 * word of the destination.
1184 * An ASR 15 fills the low word of the destination.
1186 fs_reg g0
= fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_W
));
1189 bld
.ASR(*reg
, g0
, fs_reg(15));
1191 /* Bit 31 of g1.6 is 0 if the polygon is front facing. We want to create
1192 * a boolean result from this (1/true or 0/false).
1194 * Like in the above case, since the bit is the MSB of g1.6:UD we can use
1195 * the negation source modifier to flip it. Unfortunately the SHR
1196 * instruction only operates on UD (or D with an abs source modifier)
1197 * sources without negation.
1199 * Instead, use ASR (which will give ~0/true or 0/false).
1201 fs_reg g1_6
= fs_reg(retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_D
));
1204 bld
.ASR(*reg
, g1_6
, fs_reg(31));
1211 fs_visitor::compute_sample_position(fs_reg dst
, fs_reg int_sample_pos
)
1213 assert(stage
== MESA_SHADER_FRAGMENT
);
1214 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1215 assert(dst
.type
== BRW_REGISTER_TYPE_F
);
1217 if (key
->compute_pos_offset
) {
1218 /* Convert int_sample_pos to floating point */
1219 bld
.MOV(dst
, int_sample_pos
);
1220 /* Scale to the range [0, 1] */
1221 bld
.MUL(dst
, dst
, fs_reg(1 / 16.0f
));
1224 /* From ARB_sample_shading specification:
1225 * "When rendering to a non-multisample buffer, or if multisample
1226 * rasterization is disabled, gl_SamplePosition will always be
1229 bld
.MOV(dst
, fs_reg(0.5f
));
1234 fs_visitor::emit_samplepos_setup()
1236 assert(devinfo
->gen
>= 6);
1238 const fs_builder abld
= bld
.annotate("compute sample position");
1239 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::vec2_type
));
1241 fs_reg int_sample_x
= vgrf(glsl_type::int_type
);
1242 fs_reg int_sample_y
= vgrf(glsl_type::int_type
);
1244 /* WM will be run in MSDISPMODE_PERSAMPLE. So, only one of SIMD8 or SIMD16
1245 * mode will be enabled.
1247 * From the Ivy Bridge PRM, volume 2 part 1, page 344:
1248 * R31.1:0 Position Offset X/Y for Slot[3:0]
1249 * R31.3:2 Position Offset X/Y for Slot[7:4]
1252 * The X, Y sample positions come in as bytes in thread payload. So, read
1253 * the positions using vstride=16, width=8, hstride=2.
1255 struct brw_reg sample_pos_reg
=
1256 stride(retype(brw_vec1_grf(payload
.sample_pos_reg
, 0),
1257 BRW_REGISTER_TYPE_B
), 16, 8, 2);
1259 if (dispatch_width
== 8) {
1260 abld
.MOV(int_sample_x
, fs_reg(sample_pos_reg
));
1262 abld
.half(0).MOV(half(int_sample_x
, 0), fs_reg(sample_pos_reg
));
1263 abld
.half(1).MOV(half(int_sample_x
, 1),
1264 fs_reg(suboffset(sample_pos_reg
, 16)));
1266 /* Compute gl_SamplePosition.x */
1267 compute_sample_position(pos
, int_sample_x
);
1268 pos
= offset(pos
, abld
, 1);
1269 if (dispatch_width
== 8) {
1270 abld
.MOV(int_sample_y
, fs_reg(suboffset(sample_pos_reg
, 1)));
1272 abld
.half(0).MOV(half(int_sample_y
, 0),
1273 fs_reg(suboffset(sample_pos_reg
, 1)));
1274 abld
.half(1).MOV(half(int_sample_y
, 1),
1275 fs_reg(suboffset(sample_pos_reg
, 17)));
1277 /* Compute gl_SamplePosition.y */
1278 compute_sample_position(pos
, int_sample_y
);
1283 fs_visitor::emit_sampleid_setup()
1285 assert(stage
== MESA_SHADER_FRAGMENT
);
1286 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1287 assert(devinfo
->gen
>= 6);
1289 const fs_builder abld
= bld
.annotate("compute sample id");
1290 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::int_type
));
1292 if (key
->compute_sample_id
) {
1293 fs_reg
t1(GRF
, alloc
.allocate(1), BRW_REGISTER_TYPE_D
);
1295 fs_reg
t2(GRF
, alloc
.allocate(1), BRW_REGISTER_TYPE_W
);
1297 /* The PS will be run in MSDISPMODE_PERSAMPLE. For example with
1298 * 8x multisampling, subspan 0 will represent sample N (where N
1299 * is 0, 2, 4 or 6), subspan 1 will represent sample 1, 3, 5 or
1300 * 7. We can find the value of N by looking at R0.0 bits 7:6
1301 * ("Starting Sample Pair Index (SSPI)") and multiplying by two
1302 * (since samples are always delivered in pairs). That is, we
1303 * compute 2*((R0.0 & 0xc0) >> 6) == (R0.0 & 0xc0) >> 5. Then
1304 * we need to add N to the sequence (0, 0, 0, 0, 1, 1, 1, 1) in
1305 * case of SIMD8 and sequence (0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2,
1306 * 2, 3, 3, 3, 3) in case of SIMD16. We compute this sequence by
1307 * populating a temporary variable with the sequence (0, 1, 2, 3),
1308 * and then reading from it using vstride=1, width=4, hstride=0.
1309 * These computations hold good for 4x multisampling as well.
1311 * For 2x MSAA and SIMD16, we want to use the sequence (0, 1, 0, 1):
1312 * the first four slots are sample 0 of subspan 0; the next four
1313 * are sample 1 of subspan 0; the third group is sample 0 of
1314 * subspan 1, and finally sample 1 of subspan 1.
1317 /* SKL+ has an extra bit for the Starting Sample Pair Index to
1318 * accomodate 16x MSAA.
1320 unsigned sspi_mask
= devinfo
->gen
>= 9 ? 0x1c0 : 0xc0;
1322 abld
.exec_all().group(1, 0)
1323 .AND(t1
, fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_D
)),
1325 abld
.exec_all().group(1, 0).SHR(t1
, t1
, fs_reg(5));
1327 /* This works for both SIMD8 and SIMD16 */
1328 abld
.exec_all().group(4, 0)
1329 .MOV(t2
, brw_imm_v(key
->persample_2x
? 0x1010 : 0x3210));
1331 /* This special instruction takes care of setting vstride=1,
1332 * width=4, hstride=0 of t2 during an ADD instruction.
1334 abld
.emit(FS_OPCODE_SET_SAMPLE_ID
, *reg
, t1
, t2
);
1336 /* As per GL_ARB_sample_shading specification:
1337 * "When rendering to a non-multisample buffer, or if multisample
1338 * rasterization is disabled, gl_SampleID will always be zero."
1340 abld
.MOV(*reg
, fs_reg(0));
1347 fs_visitor::resolve_source_modifiers(const fs_reg
&src
)
1349 if (!src
.abs
&& !src
.negate
)
1352 fs_reg temp
= bld
.vgrf(src
.type
);
1359 fs_visitor::emit_discard_jump()
1361 assert(((brw_wm_prog_data
*) this->prog_data
)->uses_kill
);
1363 /* For performance, after a discard, jump to the end of the
1364 * shader if all relevant channels have been discarded.
1366 fs_inst
*discard_jump
= bld
.emit(FS_OPCODE_DISCARD_JUMP
);
1367 discard_jump
->flag_subreg
= 1;
1369 discard_jump
->predicate
= (dispatch_width
== 8)
1370 ? BRW_PREDICATE_ALIGN1_ANY8H
1371 : BRW_PREDICATE_ALIGN1_ANY16H
;
1372 discard_jump
->predicate_inverse
= true;
1376 fs_visitor::emit_gs_thread_end()
1378 assert(stage
== MESA_SHADER_GEOMETRY
);
1380 struct brw_gs_prog_data
*gs_prog_data
=
1381 (struct brw_gs_prog_data
*) prog_data
;
1383 if (gs_compile
->control_data_header_size_bits
> 0) {
1384 emit_gs_control_data_bits(this->final_gs_vertex_count
);
1387 const fs_builder abld
= bld
.annotate("thread end");
1390 if (gs_prog_data
->static_vertex_count
!= -1) {
1391 foreach_in_list_reverse(fs_inst
, prev
, &this->instructions
) {
1392 if (prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8
||
1393 prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
||
1394 prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
||
1395 prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
) {
1398 /* Delete now dead instructions. */
1399 foreach_in_list_reverse_safe(exec_node
, dead
, &this->instructions
) {
1405 } else if (prev
->is_control_flow() || prev
->has_side_effects()) {
1409 fs_reg hdr
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1410 abld
.MOV(hdr
, fs_reg(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
)));
1411 inst
= abld
.emit(SHADER_OPCODE_URB_WRITE_SIMD8
, reg_undef
, hdr
);
1414 fs_reg payload
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
1415 fs_reg
*sources
= ralloc_array(mem_ctx
, fs_reg
, 2);
1416 sources
[0] = fs_reg(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
));
1417 sources
[1] = this->final_gs_vertex_count
;
1418 abld
.LOAD_PAYLOAD(payload
, sources
, 2, 2);
1419 inst
= abld
.emit(SHADER_OPCODE_URB_WRITE_SIMD8
, reg_undef
, payload
);
1427 fs_visitor::assign_curb_setup()
1429 if (dispatch_width
== 8) {
1430 prog_data
->dispatch_grf_start_reg
= payload
.num_regs
;
1432 if (stage
== MESA_SHADER_FRAGMENT
) {
1433 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
1434 prog_data
->dispatch_grf_start_reg_16
= payload
.num_regs
;
1435 } else if (stage
== MESA_SHADER_COMPUTE
) {
1436 brw_cs_prog_data
*prog_data
= (brw_cs_prog_data
*) this->prog_data
;
1437 prog_data
->dispatch_grf_start_reg_16
= payload
.num_regs
;
1439 unreachable("Unsupported shader type!");
1443 prog_data
->curb_read_length
= ALIGN(stage_prog_data
->nr_params
, 8) / 8;
1445 /* Map the offsets in the UNIFORM file to fixed HW regs. */
1446 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1447 for (unsigned int i
= 0; i
< inst
->sources
; i
++) {
1448 if (inst
->src
[i
].file
== UNIFORM
) {
1449 int uniform_nr
= inst
->src
[i
].reg
+ inst
->src
[i
].reg_offset
;
1451 if (uniform_nr
>= 0 && uniform_nr
< (int) uniforms
) {
1452 constant_nr
= push_constant_loc
[uniform_nr
];
1454 /* Section 5.11 of the OpenGL 4.1 spec says:
1455 * "Out-of-bounds reads return undefined values, which include
1456 * values from other variables of the active program or zero."
1457 * Just return the first push constant.
1462 struct brw_reg brw_reg
= brw_vec1_grf(payload
.num_regs
+
1466 assert(inst
->src
[i
].stride
== 0);
1467 inst
->src
[i
].file
= HW_REG
;
1468 inst
->src
[i
].fixed_hw_reg
= byte_offset(
1469 retype(brw_reg
, inst
->src
[i
].type
),
1470 inst
->src
[i
].subreg_offset
);
1475 /* This may be updated in assign_urb_setup or assign_vs_urb_setup. */
1476 this->first_non_payload_grf
= payload
.num_regs
+ prog_data
->curb_read_length
;
1480 fs_visitor::calculate_urb_setup()
1482 assert(stage
== MESA_SHADER_FRAGMENT
);
1483 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
1484 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1486 memset(prog_data
->urb_setup
, -1,
1487 sizeof(prog_data
->urb_setup
[0]) * VARYING_SLOT_MAX
);
1490 /* Figure out where each of the incoming setup attributes lands. */
1491 if (devinfo
->gen
>= 6) {
1492 if (_mesa_bitcount_64(nir
->info
.inputs_read
&
1493 BRW_FS_VARYING_INPUT_MASK
) <= 16) {
1494 /* The SF/SBE pipeline stage can do arbitrary rearrangement of the
1495 * first 16 varying inputs, so we can put them wherever we want.
1496 * Just put them in order.
1498 * This is useful because it means that (a) inputs not used by the
1499 * fragment shader won't take up valuable register space, and (b) we
1500 * won't have to recompile the fragment shader if it gets paired with
1501 * a different vertex (or geometry) shader.
1503 for (unsigned int i
= 0; i
< VARYING_SLOT_MAX
; i
++) {
1504 if (nir
->info
.inputs_read
& BRW_FS_VARYING_INPUT_MASK
&
1505 BITFIELD64_BIT(i
)) {
1506 prog_data
->urb_setup
[i
] = urb_next
++;
1510 bool include_vue_header
=
1511 nir
->info
.inputs_read
& (VARYING_BIT_LAYER
| VARYING_BIT_VIEWPORT
);
1513 /* We have enough input varyings that the SF/SBE pipeline stage can't
1514 * arbitrarily rearrange them to suit our whim; we have to put them
1515 * in an order that matches the output of the previous pipeline stage
1516 * (geometry or vertex shader).
1518 struct brw_vue_map prev_stage_vue_map
;
1519 brw_compute_vue_map(devinfo
, &prev_stage_vue_map
,
1520 key
->input_slots_valid
,
1521 nir
->info
.separate_shader
);
1523 include_vue_header
? 0 : 2 * BRW_SF_URB_ENTRY_READ_OFFSET
;
1525 assert(prev_stage_vue_map
.num_slots
<= first_slot
+ 32);
1526 for (int slot
= first_slot
; slot
< prev_stage_vue_map
.num_slots
;
1528 int varying
= prev_stage_vue_map
.slot_to_varying
[slot
];
1529 if (varying
!= BRW_VARYING_SLOT_PAD
&&
1530 (nir
->info
.inputs_read
& BRW_FS_VARYING_INPUT_MASK
&
1531 BITFIELD64_BIT(varying
))) {
1532 prog_data
->urb_setup
[varying
] = slot
- first_slot
;
1535 urb_next
= prev_stage_vue_map
.num_slots
- first_slot
;
1538 /* FINISHME: The sf doesn't map VS->FS inputs for us very well. */
1539 for (unsigned int i
= 0; i
< VARYING_SLOT_MAX
; i
++) {
1540 /* Point size is packed into the header, not as a general attribute */
1541 if (i
== VARYING_SLOT_PSIZ
)
1544 if (key
->input_slots_valid
& BITFIELD64_BIT(i
)) {
1545 /* The back color slot is skipped when the front color is
1546 * also written to. In addition, some slots can be
1547 * written in the vertex shader and not read in the
1548 * fragment shader. So the register number must always be
1549 * incremented, mapped or not.
1551 if (_mesa_varying_slot_in_fs((gl_varying_slot
) i
))
1552 prog_data
->urb_setup
[i
] = urb_next
;
1558 * It's a FS only attribute, and we did interpolation for this attribute
1559 * in SF thread. So, count it here, too.
1561 * See compile_sf_prog() for more info.
1563 if (nir
->info
.inputs_read
& BITFIELD64_BIT(VARYING_SLOT_PNTC
))
1564 prog_data
->urb_setup
[VARYING_SLOT_PNTC
] = urb_next
++;
1567 prog_data
->num_varying_inputs
= urb_next
;
1571 fs_visitor::assign_urb_setup()
1573 assert(stage
== MESA_SHADER_FRAGMENT
);
1574 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
1576 int urb_start
= payload
.num_regs
+ prog_data
->base
.curb_read_length
;
1578 /* Offset all the urb_setup[] index by the actual position of the
1579 * setup regs, now that the location of the constants has been chosen.
1581 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1582 if (inst
->opcode
== FS_OPCODE_LINTERP
) {
1583 assert(inst
->src
[1].file
== HW_REG
);
1584 inst
->src
[1].fixed_hw_reg
.nr
+= urb_start
;
1587 if (inst
->opcode
== FS_OPCODE_CINTERP
) {
1588 assert(inst
->src
[0].file
== HW_REG
);
1589 inst
->src
[0].fixed_hw_reg
.nr
+= urb_start
;
1593 /* Each attribute is 4 setup channels, each of which is half a reg. */
1594 this->first_non_payload_grf
+= prog_data
->num_varying_inputs
* 2;
1598 fs_visitor::convert_attr_sources_to_hw_regs(fs_inst
*inst
)
1600 for (int i
= 0; i
< inst
->sources
; i
++) {
1601 if (inst
->src
[i
].file
== ATTR
) {
1602 int grf
= payload
.num_regs
+
1603 prog_data
->curb_read_length
+
1605 inst
->src
[i
].reg_offset
;
1607 inst
->src
[i
].file
= HW_REG
;
1608 inst
->src
[i
].fixed_hw_reg
=
1609 stride(byte_offset(retype(brw_vec8_grf(grf
, 0), inst
->src
[i
].type
),
1610 inst
->src
[i
].subreg_offset
),
1611 inst
->exec_size
* inst
->src
[i
].stride
,
1612 inst
->exec_size
, inst
->src
[i
].stride
);
1618 fs_visitor::assign_vs_urb_setup()
1620 brw_vs_prog_data
*vs_prog_data
= (brw_vs_prog_data
*) prog_data
;
1622 assert(stage
== MESA_SHADER_VERTEX
);
1623 int count
= _mesa_bitcount_64(vs_prog_data
->inputs_read
);
1624 if (vs_prog_data
->uses_vertexid
|| vs_prog_data
->uses_instanceid
)
1627 /* Each attribute is 4 regs. */
1628 this->first_non_payload_grf
+= 4 * vs_prog_data
->nr_attributes
;
1630 assert(vs_prog_data
->base
.urb_read_length
<= 15);
1632 /* Rewrite all ATTR file references to the hw grf that they land in. */
1633 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1634 convert_attr_sources_to_hw_regs(inst
);
1639 fs_visitor::assign_gs_urb_setup()
1641 assert(stage
== MESA_SHADER_GEOMETRY
);
1643 brw_vue_prog_data
*vue_prog_data
= (brw_vue_prog_data
*) prog_data
;
1645 first_non_payload_grf
+=
1646 8 * vue_prog_data
->urb_read_length
* nir
->info
.gs
.vertices_in
;
1648 const unsigned first_icp_handle
= payload
.num_regs
-
1649 (vue_prog_data
->include_vue_handles
? nir
->info
.gs
.vertices_in
: 0);
1651 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1652 /* Lower URB_READ_SIMD8 opcodes into real messages. */
1653 if (inst
->opcode
== SHADER_OPCODE_URB_READ_SIMD8
) {
1654 assert(inst
->src
[0].file
== IMM
);
1655 inst
->src
[0] = retype(brw_vec8_grf(first_icp_handle
+
1656 inst
->src
[0].fixed_hw_reg
.dw1
.ud
,
1657 0), BRW_REGISTER_TYPE_UD
);
1658 /* for now, assume constant - we can do per-slot offsets later */
1659 assert(inst
->src
[1].file
== IMM
);
1660 inst
->offset
= inst
->src
[1].fixed_hw_reg
.dw1
.ud
;
1661 inst
->src
[1] = fs_reg();
1663 inst
->base_mrf
= -1;
1666 /* Rewrite all ATTR file references to HW_REGs. */
1667 convert_attr_sources_to_hw_regs(inst
);
1673 * Split large virtual GRFs into separate components if we can.
1675 * This is mostly duplicated with what brw_fs_vector_splitting does,
1676 * but that's really conservative because it's afraid of doing
1677 * splitting that doesn't result in real progress after the rest of
1678 * the optimization phases, which would cause infinite looping in
1679 * optimization. We can do it once here, safely. This also has the
1680 * opportunity to split interpolated values, or maybe even uniforms,
1681 * which we don't have at the IR level.
1683 * We want to split, because virtual GRFs are what we register
1684 * allocate and spill (due to contiguousness requirements for some
1685 * instructions), and they're what we naturally generate in the
1686 * codegen process, but most virtual GRFs don't actually need to be
1687 * contiguous sets of GRFs. If we split, we'll end up with reduced
1688 * live intervals and better dead code elimination and coalescing.
1691 fs_visitor::split_virtual_grfs()
1693 int num_vars
= this->alloc
.count
;
1695 /* Count the total number of registers */
1697 int vgrf_to_reg
[num_vars
];
1698 for (int i
= 0; i
< num_vars
; i
++) {
1699 vgrf_to_reg
[i
] = reg_count
;
1700 reg_count
+= alloc
.sizes
[i
];
1703 /* An array of "split points". For each register slot, this indicates
1704 * if this slot can be separated from the previous slot. Every time an
1705 * instruction uses multiple elements of a register (as a source or
1706 * destination), we mark the used slots as inseparable. Then we go
1707 * through and split the registers into the smallest pieces we can.
1709 bool split_points
[reg_count
];
1710 memset(split_points
, 0, sizeof(split_points
));
1712 /* Mark all used registers as fully splittable */
1713 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1714 if (inst
->dst
.file
== GRF
) {
1715 int reg
= vgrf_to_reg
[inst
->dst
.reg
];
1716 for (unsigned j
= 1; j
< this->alloc
.sizes
[inst
->dst
.reg
]; j
++)
1717 split_points
[reg
+ j
] = true;
1720 for (int i
= 0; i
< inst
->sources
; i
++) {
1721 if (inst
->src
[i
].file
== GRF
) {
1722 int reg
= vgrf_to_reg
[inst
->src
[i
].reg
];
1723 for (unsigned j
= 1; j
< this->alloc
.sizes
[inst
->src
[i
].reg
]; j
++)
1724 split_points
[reg
+ j
] = true;
1729 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1730 if (inst
->dst
.file
== GRF
) {
1731 int reg
= vgrf_to_reg
[inst
->dst
.reg
] + inst
->dst
.reg_offset
;
1732 for (int j
= 1; j
< inst
->regs_written
; j
++)
1733 split_points
[reg
+ j
] = false;
1735 for (int i
= 0; i
< inst
->sources
; i
++) {
1736 if (inst
->src
[i
].file
== GRF
) {
1737 int reg
= vgrf_to_reg
[inst
->src
[i
].reg
] + inst
->src
[i
].reg_offset
;
1738 for (int j
= 1; j
< inst
->regs_read(i
); j
++)
1739 split_points
[reg
+ j
] = false;
1744 int new_virtual_grf
[reg_count
];
1745 int new_reg_offset
[reg_count
];
1748 for (int i
= 0; i
< num_vars
; i
++) {
1749 /* The first one should always be 0 as a quick sanity check. */
1750 assert(split_points
[reg
] == false);
1753 new_reg_offset
[reg
] = 0;
1758 for (unsigned j
= 1; j
< alloc
.sizes
[i
]; j
++) {
1759 /* If this is a split point, reset the offset to 0 and allocate a
1760 * new virtual GRF for the previous offset many registers
1762 if (split_points
[reg
]) {
1763 assert(offset
<= MAX_VGRF_SIZE
);
1764 int grf
= alloc
.allocate(offset
);
1765 for (int k
= reg
- offset
; k
< reg
; k
++)
1766 new_virtual_grf
[k
] = grf
;
1769 new_reg_offset
[reg
] = offset
;
1774 /* The last one gets the original register number */
1775 assert(offset
<= MAX_VGRF_SIZE
);
1776 alloc
.sizes
[i
] = offset
;
1777 for (int k
= reg
- offset
; k
< reg
; k
++)
1778 new_virtual_grf
[k
] = i
;
1780 assert(reg
== reg_count
);
1782 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1783 if (inst
->dst
.file
== GRF
) {
1784 reg
= vgrf_to_reg
[inst
->dst
.reg
] + inst
->dst
.reg_offset
;
1785 inst
->dst
.reg
= new_virtual_grf
[reg
];
1786 inst
->dst
.reg_offset
= new_reg_offset
[reg
];
1787 assert((unsigned)new_reg_offset
[reg
] < alloc
.sizes
[new_virtual_grf
[reg
]]);
1789 for (int i
= 0; i
< inst
->sources
; i
++) {
1790 if (inst
->src
[i
].file
== GRF
) {
1791 reg
= vgrf_to_reg
[inst
->src
[i
].reg
] + inst
->src
[i
].reg_offset
;
1792 inst
->src
[i
].reg
= new_virtual_grf
[reg
];
1793 inst
->src
[i
].reg_offset
= new_reg_offset
[reg
];
1794 assert((unsigned)new_reg_offset
[reg
] < alloc
.sizes
[new_virtual_grf
[reg
]]);
1798 invalidate_live_intervals();
1802 * Remove unused virtual GRFs and compact the virtual_grf_* arrays.
1804 * During code generation, we create tons of temporary variables, many of
1805 * which get immediately killed and are never used again. Yet, in later
1806 * optimization and analysis passes, such as compute_live_intervals, we need
1807 * to loop over all the virtual GRFs. Compacting them can save a lot of
1811 fs_visitor::compact_virtual_grfs()
1813 bool progress
= false;
1814 int remap_table
[this->alloc
.count
];
1815 memset(remap_table
, -1, sizeof(remap_table
));
1817 /* Mark which virtual GRFs are used. */
1818 foreach_block_and_inst(block
, const fs_inst
, inst
, cfg
) {
1819 if (inst
->dst
.file
== GRF
)
1820 remap_table
[inst
->dst
.reg
] = 0;
1822 for (int i
= 0; i
< inst
->sources
; i
++) {
1823 if (inst
->src
[i
].file
== GRF
)
1824 remap_table
[inst
->src
[i
].reg
] = 0;
1828 /* Compact the GRF arrays. */
1830 for (unsigned i
= 0; i
< this->alloc
.count
; i
++) {
1831 if (remap_table
[i
] == -1) {
1832 /* We just found an unused register. This means that we are
1833 * actually going to compact something.
1837 remap_table
[i
] = new_index
;
1838 alloc
.sizes
[new_index
] = alloc
.sizes
[i
];
1839 invalidate_live_intervals();
1844 this->alloc
.count
= new_index
;
1846 /* Patch all the instructions to use the newly renumbered registers */
1847 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1848 if (inst
->dst
.file
== GRF
)
1849 inst
->dst
.reg
= remap_table
[inst
->dst
.reg
];
1851 for (int i
= 0; i
< inst
->sources
; i
++) {
1852 if (inst
->src
[i
].file
== GRF
)
1853 inst
->src
[i
].reg
= remap_table
[inst
->src
[i
].reg
];
1857 /* Patch all the references to delta_xy, since they're used in register
1858 * allocation. If they're unused, switch them to BAD_FILE so we don't
1859 * think some random VGRF is delta_xy.
1861 for (unsigned i
= 0; i
< ARRAY_SIZE(delta_xy
); i
++) {
1862 if (delta_xy
[i
].file
== GRF
) {
1863 if (remap_table
[delta_xy
[i
].reg
] != -1) {
1864 delta_xy
[i
].reg
= remap_table
[delta_xy
[i
].reg
];
1866 delta_xy
[i
].file
= BAD_FILE
;
1875 * Assign UNIFORM file registers to either push constants or pull constants.
1877 * We allow a fragment shader to have more than the specified minimum
1878 * maximum number of fragment shader uniform components (64). If
1879 * there are too many of these, they'd fill up all of register space.
1880 * So, this will push some of them out to the pull constant buffer and
1881 * update the program to load them. We also use pull constants for all
1882 * indirect constant loads because we don't support indirect accesses in
1886 fs_visitor::assign_constant_locations()
1888 /* Only the first compile (SIMD8 mode) gets to decide on locations. */
1889 if (dispatch_width
!= 8)
1892 unsigned int num_pull_constants
= 0;
1894 pull_constant_loc
= ralloc_array(mem_ctx
, int, uniforms
);
1895 memset(pull_constant_loc
, -1, sizeof(pull_constant_loc
[0]) * uniforms
);
1897 bool is_live
[uniforms
];
1898 memset(is_live
, 0, sizeof(is_live
));
1900 /* First, we walk through the instructions and do two things:
1902 * 1) Figure out which uniforms are live.
1904 * 2) Find all indirect access of uniform arrays and flag them as needing
1905 * to go into the pull constant buffer.
1907 * Note that we don't move constant-indexed accesses to arrays. No
1908 * testing has been done of the performance impact of this choice.
1910 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
1911 for (int i
= 0 ; i
< inst
->sources
; i
++) {
1912 if (inst
->src
[i
].file
!= UNIFORM
)
1915 if (inst
->src
[i
].reladdr
) {
1916 int uniform
= inst
->src
[i
].reg
;
1918 /* If this array isn't already present in the pull constant buffer,
1921 if (pull_constant_loc
[uniform
] == -1) {
1922 assert(param_size
[uniform
]);
1923 for (int j
= 0; j
< param_size
[uniform
]; j
++)
1924 pull_constant_loc
[uniform
+ j
] = num_pull_constants
++;
1927 /* Mark the the one accessed uniform as live */
1928 int constant_nr
= inst
->src
[i
].reg
+ inst
->src
[i
].reg_offset
;
1929 if (constant_nr
>= 0 && constant_nr
< (int) uniforms
)
1930 is_live
[constant_nr
] = true;
1935 /* Only allow 16 registers (128 uniform components) as push constants.
1937 * Just demote the end of the list. We could probably do better
1938 * here, demoting things that are rarely used in the program first.
1940 * If changing this value, note the limitation about total_regs in
1943 unsigned int max_push_components
= 16 * 8;
1944 unsigned int num_push_constants
= 0;
1946 push_constant_loc
= ralloc_array(mem_ctx
, int, uniforms
);
1948 for (unsigned int i
= 0; i
< uniforms
; i
++) {
1949 if (!is_live
[i
] || pull_constant_loc
[i
] != -1) {
1950 /* This UNIFORM register is either dead, or has already been demoted
1951 * to a pull const. Mark it as no longer living in the param[] array.
1953 push_constant_loc
[i
] = -1;
1957 if (num_push_constants
< max_push_components
) {
1958 /* Retain as a push constant. Record the location in the params[]
1961 push_constant_loc
[i
] = num_push_constants
++;
1963 /* Demote to a pull constant. */
1964 push_constant_loc
[i
] = -1;
1965 pull_constant_loc
[i
] = num_pull_constants
++;
1969 stage_prog_data
->nr_params
= num_push_constants
;
1970 stage_prog_data
->nr_pull_params
= num_pull_constants
;
1972 /* Up until now, the param[] array has been indexed by reg + reg_offset
1973 * of UNIFORM registers. Move pull constants into pull_param[] and
1974 * condense param[] to only contain the uniforms we chose to push.
1976 * NOTE: Because we are condensing the params[] array, we know that
1977 * push_constant_loc[i] <= i and we can do it in one smooth loop without
1978 * having to make a copy.
1980 for (unsigned int i
= 0; i
< uniforms
; i
++) {
1981 const gl_constant_value
*value
= stage_prog_data
->param
[i
];
1983 if (pull_constant_loc
[i
] != -1) {
1984 stage_prog_data
->pull_param
[pull_constant_loc
[i
]] = value
;
1985 } else if (push_constant_loc
[i
] != -1) {
1986 stage_prog_data
->param
[push_constant_loc
[i
]] = value
;
1992 * Replace UNIFORM register file access with either UNIFORM_PULL_CONSTANT_LOAD
1993 * or VARYING_PULL_CONSTANT_LOAD instructions which load values into VGRFs.
1996 fs_visitor::demote_pull_constants()
1998 foreach_block_and_inst (block
, fs_inst
, inst
, cfg
) {
1999 for (int i
= 0; i
< inst
->sources
; i
++) {
2000 if (inst
->src
[i
].file
!= UNIFORM
)
2004 unsigned location
= inst
->src
[i
].reg
+ inst
->src
[i
].reg_offset
;
2005 if (location
>= uniforms
) /* Out of bounds access */
2008 pull_index
= pull_constant_loc
[location
];
2010 if (pull_index
== -1)
2013 /* Set up the annotation tracking for new generated instructions. */
2014 const fs_builder
ibld(this, block
, inst
);
2015 fs_reg
surf_index(stage_prog_data
->binding_table
.pull_constants_start
);
2016 fs_reg dst
= vgrf(glsl_type::float_type
);
2018 assert(inst
->src
[i
].stride
== 0);
2020 /* Generate a pull load into dst. */
2021 if (inst
->src
[i
].reladdr
) {
2022 VARYING_PULL_CONSTANT_LOAD(ibld
, dst
,
2024 *inst
->src
[i
].reladdr
,
2026 inst
->src
[i
].reladdr
= NULL
;
2027 inst
->src
[i
].stride
= 1;
2029 const fs_builder ubld
= ibld
.exec_all().group(8, 0);
2030 fs_reg offset
= fs_reg((unsigned)(pull_index
* 4) & ~15);
2031 ubld
.emit(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
,
2032 dst
, surf_index
, offset
);
2033 inst
->src
[i
].set_smear(pull_index
& 3);
2036 /* Rewrite the instruction to use the temporary VGRF. */
2037 inst
->src
[i
].file
= GRF
;
2038 inst
->src
[i
].reg
= dst
.reg
;
2039 inst
->src
[i
].reg_offset
= 0;
2042 invalidate_live_intervals();
2046 fs_visitor::opt_algebraic()
2048 bool progress
= false;
2050 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2051 switch (inst
->opcode
) {
2052 case BRW_OPCODE_MOV
:
2053 if (inst
->src
[0].file
!= IMM
)
2056 if (inst
->saturate
) {
2057 if (inst
->dst
.type
!= inst
->src
[0].type
)
2058 assert(!"unimplemented: saturate mixed types");
2060 if (brw_saturate_immediate(inst
->dst
.type
,
2061 &inst
->src
[0].fixed_hw_reg
)) {
2062 inst
->saturate
= false;
2068 case BRW_OPCODE_MUL
:
2069 if (inst
->src
[1].file
!= IMM
)
2073 if (inst
->src
[1].is_one()) {
2074 inst
->opcode
= BRW_OPCODE_MOV
;
2075 inst
->src
[1] = reg_undef
;
2081 if (inst
->src
[1].is_negative_one()) {
2082 inst
->opcode
= BRW_OPCODE_MOV
;
2083 inst
->src
[0].negate
= !inst
->src
[0].negate
;
2084 inst
->src
[1] = reg_undef
;
2090 if (inst
->src
[1].is_zero()) {
2091 inst
->opcode
= BRW_OPCODE_MOV
;
2092 inst
->src
[0] = inst
->src
[1];
2093 inst
->src
[1] = reg_undef
;
2098 if (inst
->src
[0].file
== IMM
) {
2099 assert(inst
->src
[0].type
== BRW_REGISTER_TYPE_F
);
2100 inst
->opcode
= BRW_OPCODE_MOV
;
2101 inst
->src
[0].fixed_hw_reg
.dw1
.f
*= inst
->src
[1].fixed_hw_reg
.dw1
.f
;
2102 inst
->src
[1] = reg_undef
;
2107 case BRW_OPCODE_ADD
:
2108 if (inst
->src
[1].file
!= IMM
)
2112 if (inst
->src
[1].is_zero()) {
2113 inst
->opcode
= BRW_OPCODE_MOV
;
2114 inst
->src
[1] = reg_undef
;
2119 if (inst
->src
[0].file
== IMM
) {
2120 assert(inst
->src
[0].type
== BRW_REGISTER_TYPE_F
);
2121 inst
->opcode
= BRW_OPCODE_MOV
;
2122 inst
->src
[0].fixed_hw_reg
.dw1
.f
+= inst
->src
[1].fixed_hw_reg
.dw1
.f
;
2123 inst
->src
[1] = reg_undef
;
2129 if (inst
->src
[0].equals(inst
->src
[1])) {
2130 inst
->opcode
= BRW_OPCODE_MOV
;
2131 inst
->src
[1] = reg_undef
;
2136 case BRW_OPCODE_LRP
:
2137 if (inst
->src
[1].equals(inst
->src
[2])) {
2138 inst
->opcode
= BRW_OPCODE_MOV
;
2139 inst
->src
[0] = inst
->src
[1];
2140 inst
->src
[1] = reg_undef
;
2141 inst
->src
[2] = reg_undef
;
2146 case BRW_OPCODE_CMP
:
2147 if (inst
->conditional_mod
== BRW_CONDITIONAL_GE
&&
2149 inst
->src
[0].negate
&&
2150 inst
->src
[1].is_zero()) {
2151 inst
->src
[0].abs
= false;
2152 inst
->src
[0].negate
= false;
2153 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
2158 case BRW_OPCODE_SEL
:
2159 if (inst
->src
[0].equals(inst
->src
[1])) {
2160 inst
->opcode
= BRW_OPCODE_MOV
;
2161 inst
->src
[1] = reg_undef
;
2162 inst
->predicate
= BRW_PREDICATE_NONE
;
2163 inst
->predicate_inverse
= false;
2165 } else if (inst
->saturate
&& inst
->src
[1].file
== IMM
) {
2166 switch (inst
->conditional_mod
) {
2167 case BRW_CONDITIONAL_LE
:
2168 case BRW_CONDITIONAL_L
:
2169 switch (inst
->src
[1].type
) {
2170 case BRW_REGISTER_TYPE_F
:
2171 if (inst
->src
[1].fixed_hw_reg
.dw1
.f
>= 1.0f
) {
2172 inst
->opcode
= BRW_OPCODE_MOV
;
2173 inst
->src
[1] = reg_undef
;
2174 inst
->conditional_mod
= BRW_CONDITIONAL_NONE
;
2182 case BRW_CONDITIONAL_GE
:
2183 case BRW_CONDITIONAL_G
:
2184 switch (inst
->src
[1].type
) {
2185 case BRW_REGISTER_TYPE_F
:
2186 if (inst
->src
[1].fixed_hw_reg
.dw1
.f
<= 0.0f
) {
2187 inst
->opcode
= BRW_OPCODE_MOV
;
2188 inst
->src
[1] = reg_undef
;
2189 inst
->conditional_mod
= BRW_CONDITIONAL_NONE
;
2201 case BRW_OPCODE_MAD
:
2202 if (inst
->src
[1].is_zero() || inst
->src
[2].is_zero()) {
2203 inst
->opcode
= BRW_OPCODE_MOV
;
2204 inst
->src
[1] = reg_undef
;
2205 inst
->src
[2] = reg_undef
;
2207 } else if (inst
->src
[0].is_zero()) {
2208 inst
->opcode
= BRW_OPCODE_MUL
;
2209 inst
->src
[0] = inst
->src
[2];
2210 inst
->src
[2] = reg_undef
;
2212 } else if (inst
->src
[1].is_one()) {
2213 inst
->opcode
= BRW_OPCODE_ADD
;
2214 inst
->src
[1] = inst
->src
[2];
2215 inst
->src
[2] = reg_undef
;
2217 } else if (inst
->src
[2].is_one()) {
2218 inst
->opcode
= BRW_OPCODE_ADD
;
2219 inst
->src
[2] = reg_undef
;
2221 } else if (inst
->src
[1].file
== IMM
&& inst
->src
[2].file
== IMM
) {
2222 inst
->opcode
= BRW_OPCODE_ADD
;
2223 inst
->src
[1].fixed_hw_reg
.dw1
.f
*= inst
->src
[2].fixed_hw_reg
.dw1
.f
;
2224 inst
->src
[2] = reg_undef
;
2228 case SHADER_OPCODE_RCP
: {
2229 fs_inst
*prev
= (fs_inst
*)inst
->prev
;
2230 if (prev
->opcode
== SHADER_OPCODE_SQRT
) {
2231 if (inst
->src
[0].equals(prev
->dst
)) {
2232 inst
->opcode
= SHADER_OPCODE_RSQ
;
2233 inst
->src
[0] = prev
->src
[0];
2239 case SHADER_OPCODE_BROADCAST
:
2240 if (is_uniform(inst
->src
[0])) {
2241 inst
->opcode
= BRW_OPCODE_MOV
;
2243 inst
->force_writemask_all
= true;
2245 } else if (inst
->src
[1].file
== IMM
) {
2246 inst
->opcode
= BRW_OPCODE_MOV
;
2247 inst
->src
[0] = component(inst
->src
[0],
2248 inst
->src
[1].fixed_hw_reg
.dw1
.ud
);
2250 inst
->force_writemask_all
= true;
2259 /* Swap if src[0] is immediate. */
2260 if (progress
&& inst
->is_commutative()) {
2261 if (inst
->src
[0].file
== IMM
) {
2262 fs_reg tmp
= inst
->src
[1];
2263 inst
->src
[1] = inst
->src
[0];
2272 * Optimize sample messages that have constant zero values for the trailing
2273 * texture coordinates. We can just reduce the message length for these
2274 * instructions instead of reserving a register for it. Trailing parameters
2275 * that aren't sent default to zero anyway. This will cause the dead code
2276 * eliminator to remove the MOV instruction that would otherwise be emitted to
2277 * set up the zero value.
2280 fs_visitor::opt_zero_samples()
2282 /* Gen4 infers the texturing opcode based on the message length so we can't
2285 if (devinfo
->gen
< 5)
2288 bool progress
= false;
2290 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2291 if (!inst
->is_tex())
2294 fs_inst
*load_payload
= (fs_inst
*) inst
->prev
;
2296 if (load_payload
->is_head_sentinel() ||
2297 load_payload
->opcode
!= SHADER_OPCODE_LOAD_PAYLOAD
)
2300 /* We don't want to remove the message header or the first parameter.
2301 * Removing the first parameter is not allowed, see the Haswell PRM
2302 * volume 7, page 149:
2304 * "Parameter 0 is required except for the sampleinfo message, which
2305 * has no parameter 0"
2307 while (inst
->mlen
> inst
->header_size
+ inst
->exec_size
/ 8 &&
2308 load_payload
->src
[(inst
->mlen
- inst
->header_size
) /
2309 (inst
->exec_size
/ 8) +
2310 inst
->header_size
- 1].is_zero()) {
2311 inst
->mlen
-= inst
->exec_size
/ 8;
2317 invalidate_live_intervals();
2323 * Optimize sample messages which are followed by the final RT write.
2325 * CHV, and GEN9+ can mark a texturing SEND instruction with EOT to have its
2326 * results sent directly to the framebuffer, bypassing the EU. Recognize the
2327 * final texturing results copied to the framebuffer write payload and modify
2328 * them to write to the framebuffer directly.
2331 fs_visitor::opt_sampler_eot()
2333 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
2335 if (stage
!= MESA_SHADER_FRAGMENT
)
2338 if (devinfo
->gen
< 9 && !devinfo
->is_cherryview
)
2341 /* FINISHME: It should be possible to implement this optimization when there
2342 * are multiple drawbuffers.
2344 if (key
->nr_color_regions
!= 1)
2347 /* Look for a texturing instruction immediately before the final FB_WRITE. */
2348 bblock_t
*block
= cfg
->blocks
[cfg
->num_blocks
- 1];
2349 fs_inst
*fb_write
= (fs_inst
*)block
->end();
2350 assert(fb_write
->eot
);
2351 assert(fb_write
->opcode
== FS_OPCODE_FB_WRITE
);
2353 fs_inst
*tex_inst
= (fs_inst
*) fb_write
->prev
;
2355 /* There wasn't one; nothing to do. */
2356 if (unlikely(tex_inst
->is_head_sentinel()) || !tex_inst
->is_tex())
2359 /* 3D Sampler » Messages » Message Format
2361 * “Response Length of zero is allowed on all SIMD8* and SIMD16* sampler
2362 * messages except sample+killpix, resinfo, sampleinfo, LOD, and gather4*”
2364 if (tex_inst
->opcode
== SHADER_OPCODE_TXS
||
2365 tex_inst
->opcode
== SHADER_OPCODE_SAMPLEINFO
||
2366 tex_inst
->opcode
== SHADER_OPCODE_LOD
||
2367 tex_inst
->opcode
== SHADER_OPCODE_TG4
||
2368 tex_inst
->opcode
== SHADER_OPCODE_TG4_OFFSET
)
2371 /* If there's no header present, we need to munge the LOAD_PAYLOAD as well.
2372 * It's very likely to be the previous instruction.
2374 fs_inst
*load_payload
= (fs_inst
*) tex_inst
->prev
;
2375 if (load_payload
->is_head_sentinel() ||
2376 load_payload
->opcode
!= SHADER_OPCODE_LOAD_PAYLOAD
)
2379 assert(!tex_inst
->eot
); /* We can't get here twice */
2380 assert((tex_inst
->offset
& (0xff << 24)) == 0);
2382 const fs_builder
ibld(this, block
, tex_inst
);
2384 tex_inst
->offset
|= fb_write
->target
<< 24;
2385 tex_inst
->eot
= true;
2386 tex_inst
->dst
= ibld
.null_reg_ud();
2387 fb_write
->remove(cfg
->blocks
[cfg
->num_blocks
- 1]);
2389 /* If a header is present, marking the eot is sufficient. Otherwise, we need
2390 * to create a new LOAD_PAYLOAD command with the same sources and a space
2391 * saved for the header. Using a new destination register not only makes sure
2392 * we have enough space, but it will make sure the dead code eliminator kills
2393 * the instruction that this will replace.
2395 if (tex_inst
->header_size
!= 0)
2398 fs_reg send_header
= ibld
.vgrf(BRW_REGISTER_TYPE_F
,
2399 load_payload
->sources
+ 1);
2400 fs_reg
*new_sources
=
2401 ralloc_array(mem_ctx
, fs_reg
, load_payload
->sources
+ 1);
2403 new_sources
[0] = fs_reg();
2404 for (int i
= 0; i
< load_payload
->sources
; i
++)
2405 new_sources
[i
+1] = load_payload
->src
[i
];
2407 /* The LOAD_PAYLOAD helper seems like the obvious choice here. However, it
2408 * requires a lot of information about the sources to appropriately figure
2409 * out the number of registers needed to be used. Given this stage in our
2410 * optimization, we may not have the appropriate GRFs required by
2411 * LOAD_PAYLOAD at this point (copy propagation). Therefore, we need to
2412 * manually emit the instruction.
2414 fs_inst
*new_load_payload
= new(mem_ctx
) fs_inst(SHADER_OPCODE_LOAD_PAYLOAD
,
2415 load_payload
->exec_size
,
2418 load_payload
->sources
+ 1);
2420 new_load_payload
->regs_written
= load_payload
->regs_written
+ 1;
2421 new_load_payload
->header_size
= 1;
2423 tex_inst
->header_size
= 1;
2424 tex_inst
->insert_before(cfg
->blocks
[cfg
->num_blocks
- 1], new_load_payload
);
2425 tex_inst
->src
[0] = send_header
;
2431 fs_visitor::opt_register_renaming()
2433 bool progress
= false;
2436 int remap
[alloc
.count
];
2437 memset(remap
, -1, sizeof(int) * alloc
.count
);
2439 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2440 if (inst
->opcode
== BRW_OPCODE_IF
|| inst
->opcode
== BRW_OPCODE_DO
) {
2442 } else if (inst
->opcode
== BRW_OPCODE_ENDIF
||
2443 inst
->opcode
== BRW_OPCODE_WHILE
) {
2447 /* Rewrite instruction sources. */
2448 for (int i
= 0; i
< inst
->sources
; i
++) {
2449 if (inst
->src
[i
].file
== GRF
&&
2450 remap
[inst
->src
[i
].reg
] != -1 &&
2451 remap
[inst
->src
[i
].reg
] != inst
->src
[i
].reg
) {
2452 inst
->src
[i
].reg
= remap
[inst
->src
[i
].reg
];
2457 const int dst
= inst
->dst
.reg
;
2460 inst
->dst
.file
== GRF
&&
2461 alloc
.sizes
[inst
->dst
.reg
] == inst
->exec_size
/ 8 &&
2462 !inst
->is_partial_write()) {
2463 if (remap
[dst
] == -1) {
2466 remap
[dst
] = alloc
.allocate(inst
->exec_size
/ 8);
2467 inst
->dst
.reg
= remap
[dst
];
2470 } else if (inst
->dst
.file
== GRF
&&
2472 remap
[dst
] != dst
) {
2473 inst
->dst
.reg
= remap
[dst
];
2479 invalidate_live_intervals();
2481 for (unsigned i
= 0; i
< ARRAY_SIZE(delta_xy
); i
++) {
2482 if (delta_xy
[i
].file
== GRF
&& remap
[delta_xy
[i
].reg
] != -1) {
2483 delta_xy
[i
].reg
= remap
[delta_xy
[i
].reg
];
2492 * Remove redundant or useless discard jumps.
2494 * For example, we can eliminate jumps in the following sequence:
2496 * discard-jump (redundant with the next jump)
2497 * discard-jump (useless; jumps to the next instruction)
2501 fs_visitor::opt_redundant_discard_jumps()
2503 bool progress
= false;
2505 bblock_t
*last_bblock
= cfg
->blocks
[cfg
->num_blocks
- 1];
2507 fs_inst
*placeholder_halt
= NULL
;
2508 foreach_inst_in_block_reverse(fs_inst
, inst
, last_bblock
) {
2509 if (inst
->opcode
== FS_OPCODE_PLACEHOLDER_HALT
) {
2510 placeholder_halt
= inst
;
2515 if (!placeholder_halt
)
2518 /* Delete any HALTs immediately before the placeholder halt. */
2519 for (fs_inst
*prev
= (fs_inst
*) placeholder_halt
->prev
;
2520 !prev
->is_head_sentinel() && prev
->opcode
== FS_OPCODE_DISCARD_JUMP
;
2521 prev
= (fs_inst
*) placeholder_halt
->prev
) {
2522 prev
->remove(last_bblock
);
2527 invalidate_live_intervals();
2533 fs_visitor::compute_to_mrf()
2535 bool progress
= false;
2538 /* No MRFs on Gen >= 7. */
2539 if (devinfo
->gen
>= 7)
2542 calculate_live_intervals();
2544 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
2548 if (inst
->opcode
!= BRW_OPCODE_MOV
||
2549 inst
->is_partial_write() ||
2550 inst
->dst
.file
!= MRF
|| inst
->src
[0].file
!= GRF
||
2551 inst
->dst
.type
!= inst
->src
[0].type
||
2552 inst
->src
[0].abs
|| inst
->src
[0].negate
||
2553 !inst
->src
[0].is_contiguous() ||
2554 inst
->src
[0].subreg_offset
)
2557 /* Work out which hardware MRF registers are written by this
2560 int mrf_low
= inst
->dst
.reg
& ~BRW_MRF_COMPR4
;
2562 if (inst
->dst
.reg
& BRW_MRF_COMPR4
) {
2563 mrf_high
= mrf_low
+ 4;
2564 } else if (inst
->exec_size
== 16) {
2565 mrf_high
= mrf_low
+ 1;
2570 /* Can't compute-to-MRF this GRF if someone else was going to
2573 if (this->virtual_grf_end
[inst
->src
[0].reg
] > ip
)
2576 /* Found a move of a GRF to a MRF. Let's see if we can go
2577 * rewrite the thing that made this GRF to write into the MRF.
2579 foreach_inst_in_block_reverse_starting_from(fs_inst
, scan_inst
, inst
) {
2580 if (scan_inst
->dst
.file
== GRF
&&
2581 scan_inst
->dst
.reg
== inst
->src
[0].reg
) {
2582 /* Found the last thing to write our reg we want to turn
2583 * into a compute-to-MRF.
2586 /* If this one instruction didn't populate all the
2587 * channels, bail. We might be able to rewrite everything
2588 * that writes that reg, but it would require smarter
2589 * tracking to delay the rewriting until complete success.
2591 if (scan_inst
->is_partial_write())
2594 /* Things returning more than one register would need us to
2595 * understand coalescing out more than one MOV at a time.
2597 if (scan_inst
->regs_written
> scan_inst
->exec_size
/ 8)
2600 /* SEND instructions can't have MRF as a destination. */
2601 if (scan_inst
->mlen
)
2604 if (devinfo
->gen
== 6) {
2605 /* gen6 math instructions must have the destination be
2606 * GRF, so no compute-to-MRF for them.
2608 if (scan_inst
->is_math()) {
2613 if (scan_inst
->dst
.reg_offset
== inst
->src
[0].reg_offset
) {
2614 /* Found the creator of our MRF's source value. */
2615 scan_inst
->dst
.file
= MRF
;
2616 scan_inst
->dst
.reg
= inst
->dst
.reg
;
2617 scan_inst
->saturate
|= inst
->saturate
;
2618 inst
->remove(block
);
2624 /* We don't handle control flow here. Most computation of
2625 * values that end up in MRFs are shortly before the MRF
2628 if (block
->start() == scan_inst
)
2631 /* You can't read from an MRF, so if someone else reads our
2632 * MRF's source GRF that we wanted to rewrite, that stops us.
2634 bool interfered
= false;
2635 for (int i
= 0; i
< scan_inst
->sources
; i
++) {
2636 if (scan_inst
->src
[i
].file
== GRF
&&
2637 scan_inst
->src
[i
].reg
== inst
->src
[0].reg
&&
2638 scan_inst
->src
[i
].reg_offset
== inst
->src
[0].reg_offset
) {
2645 if (scan_inst
->dst
.file
== MRF
) {
2646 /* If somebody else writes our MRF here, we can't
2647 * compute-to-MRF before that.
2649 int scan_mrf_low
= scan_inst
->dst
.reg
& ~BRW_MRF_COMPR4
;
2652 if (scan_inst
->dst
.reg
& BRW_MRF_COMPR4
) {
2653 scan_mrf_high
= scan_mrf_low
+ 4;
2654 } else if (scan_inst
->exec_size
== 16) {
2655 scan_mrf_high
= scan_mrf_low
+ 1;
2657 scan_mrf_high
= scan_mrf_low
;
2660 if (mrf_low
== scan_mrf_low
||
2661 mrf_low
== scan_mrf_high
||
2662 mrf_high
== scan_mrf_low
||
2663 mrf_high
== scan_mrf_high
) {
2668 if (scan_inst
->mlen
> 0 && scan_inst
->base_mrf
!= -1) {
2669 /* Found a SEND instruction, which means that there are
2670 * live values in MRFs from base_mrf to base_mrf +
2671 * scan_inst->mlen - 1. Don't go pushing our MRF write up
2674 if (mrf_low
>= scan_inst
->base_mrf
&&
2675 mrf_low
< scan_inst
->base_mrf
+ scan_inst
->mlen
) {
2678 if (mrf_high
>= scan_inst
->base_mrf
&&
2679 mrf_high
< scan_inst
->base_mrf
+ scan_inst
->mlen
) {
2687 invalidate_live_intervals();
2693 * Eliminate FIND_LIVE_CHANNEL instructions occurring outside any control
2694 * flow. We could probably do better here with some form of divergence
2698 fs_visitor::eliminate_find_live_channel()
2700 bool progress
= false;
2703 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
2704 switch (inst
->opcode
) {
2710 case BRW_OPCODE_ENDIF
:
2711 case BRW_OPCODE_WHILE
:
2715 case FS_OPCODE_DISCARD_JUMP
:
2716 /* This can potentially make control flow non-uniform until the end
2721 case SHADER_OPCODE_FIND_LIVE_CHANNEL
:
2723 inst
->opcode
= BRW_OPCODE_MOV
;
2724 inst
->src
[0] = fs_reg(0u);
2726 inst
->force_writemask_all
= true;
2740 * Once we've generated code, try to convert normal FS_OPCODE_FB_WRITE
2741 * instructions to FS_OPCODE_REP_FB_WRITE.
2744 fs_visitor::emit_repclear_shader()
2746 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
2748 int color_mrf
= base_mrf
+ 2;
2750 fs_inst
*mov
= bld
.exec_all().group(4, 0)
2751 .MOV(brw_message_reg(color_mrf
),
2752 fs_reg(UNIFORM
, 0, BRW_REGISTER_TYPE_F
));
2755 if (key
->nr_color_regions
== 1) {
2756 write
= bld
.emit(FS_OPCODE_REP_FB_WRITE
);
2757 write
->saturate
= key
->clamp_fragment_color
;
2758 write
->base_mrf
= color_mrf
;
2760 write
->header_size
= 0;
2763 assume(key
->nr_color_regions
> 0);
2764 for (int i
= 0; i
< key
->nr_color_regions
; ++i
) {
2765 write
= bld
.emit(FS_OPCODE_REP_FB_WRITE
);
2766 write
->saturate
= key
->clamp_fragment_color
;
2767 write
->base_mrf
= base_mrf
;
2769 write
->header_size
= 2;
2777 assign_constant_locations();
2778 assign_curb_setup();
2780 /* Now that we have the uniform assigned, go ahead and force it to a vec4. */
2781 assert(mov
->src
[0].file
== HW_REG
);
2782 mov
->src
[0] = brw_vec4_grf(mov
->src
[0].fixed_hw_reg
.nr
, 0);
2786 * Walks through basic blocks, looking for repeated MRF writes and
2787 * removing the later ones.
2790 fs_visitor::remove_duplicate_mrf_writes()
2792 fs_inst
*last_mrf_move
[BRW_MAX_MRF(devinfo
->gen
)];
2793 bool progress
= false;
2795 /* Need to update the MRF tracking for compressed instructions. */
2796 if (dispatch_width
== 16)
2799 memset(last_mrf_move
, 0, sizeof(last_mrf_move
));
2801 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
2802 if (inst
->is_control_flow()) {
2803 memset(last_mrf_move
, 0, sizeof(last_mrf_move
));
2806 if (inst
->opcode
== BRW_OPCODE_MOV
&&
2807 inst
->dst
.file
== MRF
) {
2808 fs_inst
*prev_inst
= last_mrf_move
[inst
->dst
.reg
];
2809 if (prev_inst
&& inst
->equals(prev_inst
)) {
2810 inst
->remove(block
);
2816 /* Clear out the last-write records for MRFs that were overwritten. */
2817 if (inst
->dst
.file
== MRF
) {
2818 last_mrf_move
[inst
->dst
.reg
] = NULL
;
2821 if (inst
->mlen
> 0 && inst
->base_mrf
!= -1) {
2822 /* Found a SEND instruction, which will include two or fewer
2823 * implied MRF writes. We could do better here.
2825 for (int i
= 0; i
< implied_mrf_writes(inst
); i
++) {
2826 last_mrf_move
[inst
->base_mrf
+ i
] = NULL
;
2830 /* Clear out any MRF move records whose sources got overwritten. */
2831 if (inst
->dst
.file
== GRF
) {
2832 for (unsigned int i
= 0; i
< ARRAY_SIZE(last_mrf_move
); i
++) {
2833 if (last_mrf_move
[i
] &&
2834 last_mrf_move
[i
]->src
[0].reg
== inst
->dst
.reg
) {
2835 last_mrf_move
[i
] = NULL
;
2840 if (inst
->opcode
== BRW_OPCODE_MOV
&&
2841 inst
->dst
.file
== MRF
&&
2842 inst
->src
[0].file
== GRF
&&
2843 !inst
->is_partial_write()) {
2844 last_mrf_move
[inst
->dst
.reg
] = inst
;
2849 invalidate_live_intervals();
2855 clear_deps_for_inst_src(fs_inst
*inst
, bool *deps
, int first_grf
, int grf_len
)
2857 /* Clear the flag for registers that actually got read (as expected). */
2858 for (int i
= 0; i
< inst
->sources
; i
++) {
2860 if (inst
->src
[i
].file
== GRF
) {
2861 grf
= inst
->src
[i
].reg
;
2862 } else if (inst
->src
[i
].file
== HW_REG
&&
2863 inst
->src
[i
].fixed_hw_reg
.file
== BRW_GENERAL_REGISTER_FILE
) {
2864 grf
= inst
->src
[i
].fixed_hw_reg
.nr
;
2869 if (grf
>= first_grf
&&
2870 grf
< first_grf
+ grf_len
) {
2871 deps
[grf
- first_grf
] = false;
2872 if (inst
->exec_size
== 16)
2873 deps
[grf
- first_grf
+ 1] = false;
2879 * Implements this workaround for the original 965:
2881 * "[DevBW, DevCL] Implementation Restrictions: As the hardware does not
2882 * check for post destination dependencies on this instruction, software
2883 * must ensure that there is no destination hazard for the case of ‘write
2884 * followed by a posted write’ shown in the following example.
2887 * 2. send r3.xy <rest of send instruction>
2890 * Due to no post-destination dependency check on the ‘send’, the above
2891 * code sequence could have two instructions (1 and 2) in flight at the
2892 * same time that both consider ‘r3’ as the target of their final writes.
2895 fs_visitor::insert_gen4_pre_send_dependency_workarounds(bblock_t
*block
,
2898 int write_len
= inst
->regs_written
;
2899 int first_write_grf
= inst
->dst
.reg
;
2900 bool needs_dep
[BRW_MAX_MRF(devinfo
->gen
)];
2901 assert(write_len
< (int)sizeof(needs_dep
) - 1);
2903 memset(needs_dep
, false, sizeof(needs_dep
));
2904 memset(needs_dep
, true, write_len
);
2906 clear_deps_for_inst_src(inst
, needs_dep
, first_write_grf
, write_len
);
2908 /* Walk backwards looking for writes to registers we're writing which
2909 * aren't read since being written. If we hit the start of the program,
2910 * we assume that there are no outstanding dependencies on entry to the
2913 foreach_inst_in_block_reverse_starting_from(fs_inst
, scan_inst
, inst
) {
2914 /* If we hit control flow, assume that there *are* outstanding
2915 * dependencies, and force their cleanup before our instruction.
2917 if (block
->start() == scan_inst
) {
2918 for (int i
= 0; i
< write_len
; i
++) {
2920 DEP_RESOLVE_MOV(fs_builder(this, block
, inst
),
2921 first_write_grf
+ i
);
2926 /* We insert our reads as late as possible on the assumption that any
2927 * instruction but a MOV that might have left us an outstanding
2928 * dependency has more latency than a MOV.
2930 if (scan_inst
->dst
.file
== GRF
) {
2931 for (int i
= 0; i
< scan_inst
->regs_written
; i
++) {
2932 int reg
= scan_inst
->dst
.reg
+ i
;
2934 if (reg
>= first_write_grf
&&
2935 reg
< first_write_grf
+ write_len
&&
2936 needs_dep
[reg
- first_write_grf
]) {
2937 DEP_RESOLVE_MOV(fs_builder(this, block
, inst
), reg
);
2938 needs_dep
[reg
- first_write_grf
] = false;
2939 if (scan_inst
->exec_size
== 16)
2940 needs_dep
[reg
- first_write_grf
+ 1] = false;
2945 /* Clear the flag for registers that actually got read (as expected). */
2946 clear_deps_for_inst_src(scan_inst
, needs_dep
, first_write_grf
, write_len
);
2948 /* Continue the loop only if we haven't resolved all the dependencies */
2950 for (i
= 0; i
< write_len
; i
++) {
2960 * Implements this workaround for the original 965:
2962 * "[DevBW, DevCL] Errata: A destination register from a send can not be
2963 * used as a destination register until after it has been sourced by an
2964 * instruction with a different destination register.
2967 fs_visitor::insert_gen4_post_send_dependency_workarounds(bblock_t
*block
, fs_inst
*inst
)
2969 int write_len
= inst
->regs_written
;
2970 int first_write_grf
= inst
->dst
.reg
;
2971 bool needs_dep
[BRW_MAX_MRF(devinfo
->gen
)];
2972 assert(write_len
< (int)sizeof(needs_dep
) - 1);
2974 memset(needs_dep
, false, sizeof(needs_dep
));
2975 memset(needs_dep
, true, write_len
);
2976 /* Walk forwards looking for writes to registers we're writing which aren't
2977 * read before being written.
2979 foreach_inst_in_block_starting_from(fs_inst
, scan_inst
, inst
) {
2980 /* If we hit control flow, force resolve all remaining dependencies. */
2981 if (block
->end() == scan_inst
) {
2982 for (int i
= 0; i
< write_len
; i
++) {
2984 DEP_RESOLVE_MOV(fs_builder(this, block
, scan_inst
),
2985 first_write_grf
+ i
);
2990 /* Clear the flag for registers that actually got read (as expected). */
2991 clear_deps_for_inst_src(scan_inst
, needs_dep
, first_write_grf
, write_len
);
2993 /* We insert our reads as late as possible since they're reading the
2994 * result of a SEND, which has massive latency.
2996 if (scan_inst
->dst
.file
== GRF
&&
2997 scan_inst
->dst
.reg
>= first_write_grf
&&
2998 scan_inst
->dst
.reg
< first_write_grf
+ write_len
&&
2999 needs_dep
[scan_inst
->dst
.reg
- first_write_grf
]) {
3000 DEP_RESOLVE_MOV(fs_builder(this, block
, scan_inst
),
3001 scan_inst
->dst
.reg
);
3002 needs_dep
[scan_inst
->dst
.reg
- first_write_grf
] = false;
3005 /* Continue the loop only if we haven't resolved all the dependencies */
3007 for (i
= 0; i
< write_len
; i
++) {
3017 fs_visitor::insert_gen4_send_dependency_workarounds()
3019 if (devinfo
->gen
!= 4 || devinfo
->is_g4x
)
3022 bool progress
= false;
3024 /* Note that we're done with register allocation, so GRF fs_regs always
3025 * have a .reg_offset of 0.
3028 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
3029 if (inst
->mlen
!= 0 && inst
->dst
.file
== GRF
) {
3030 insert_gen4_pre_send_dependency_workarounds(block
, inst
);
3031 insert_gen4_post_send_dependency_workarounds(block
, inst
);
3037 invalidate_live_intervals();
3041 * Turns the generic expression-style uniform pull constant load instruction
3042 * into a hardware-specific series of instructions for loading a pull
3045 * The expression style allows the CSE pass before this to optimize out
3046 * repeated loads from the same offset, and gives the pre-register-allocation
3047 * scheduling full flexibility, while the conversion to native instructions
3048 * allows the post-register-allocation scheduler the best information
3051 * Note that execution masking for setting up pull constant loads is special:
3052 * the channels that need to be written are unrelated to the current execution
3053 * mask, since a later instruction will use one of the result channels as a
3054 * source operand for all 8 or 16 of its channels.
3057 fs_visitor::lower_uniform_pull_constant_loads()
3059 foreach_block_and_inst (block
, fs_inst
, inst
, cfg
) {
3060 if (inst
->opcode
!= FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
)
3063 if (devinfo
->gen
>= 7) {
3064 /* The offset arg before was a vec4-aligned byte offset. We need to
3065 * turn it into a dword offset.
3067 fs_reg const_offset_reg
= inst
->src
[1];
3068 assert(const_offset_reg
.file
== IMM
&&
3069 const_offset_reg
.type
== BRW_REGISTER_TYPE_UD
);
3070 const_offset_reg
.fixed_hw_reg
.dw1
.ud
/= 4;
3072 fs_reg payload
, offset
;
3073 if (devinfo
->gen
>= 9) {
3074 /* We have to use a message header on Skylake to get SIMD4x2
3075 * mode. Reserve space for the register.
3077 offset
= payload
= fs_reg(GRF
, alloc
.allocate(2));
3078 offset
.reg_offset
++;
3081 offset
= payload
= fs_reg(GRF
, alloc
.allocate(1));
3085 /* This is actually going to be a MOV, but since only the first dword
3086 * is accessed, we have a special opcode to do just that one. Note
3087 * that this needs to be an operation that will be considered a def
3088 * by live variable analysis, or register allocation will explode.
3090 fs_inst
*setup
= new(mem_ctx
) fs_inst(FS_OPCODE_SET_SIMD4X2_OFFSET
,
3091 8, offset
, const_offset_reg
);
3092 setup
->force_writemask_all
= true;
3094 setup
->ir
= inst
->ir
;
3095 setup
->annotation
= inst
->annotation
;
3096 inst
->insert_before(block
, setup
);
3098 /* Similarly, this will only populate the first 4 channels of the
3099 * result register (since we only use smear values from 0-3), but we
3100 * don't tell the optimizer.
3102 inst
->opcode
= FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7
;
3103 inst
->src
[1] = payload
;
3104 inst
->base_mrf
= -1;
3106 invalidate_live_intervals();
3108 /* Before register allocation, we didn't tell the scheduler about the
3109 * MRF we use. We know it's safe to use this MRF because nothing
3110 * else does except for register spill/unspill, which generates and
3111 * uses its MRF within a single IR instruction.
3113 inst
->base_mrf
= FIRST_PULL_LOAD_MRF(devinfo
->gen
) + 1;
3120 fs_visitor::lower_load_payload()
3122 bool progress
= false;
3124 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
3125 if (inst
->opcode
!= SHADER_OPCODE_LOAD_PAYLOAD
)
3128 assert(inst
->dst
.file
== MRF
|| inst
->dst
.file
== GRF
);
3129 assert(inst
->saturate
== false);
3130 fs_reg dst
= inst
->dst
;
3132 /* Get rid of COMPR4. We'll add it back in if we need it */
3133 if (dst
.file
== MRF
)
3134 dst
.reg
= dst
.reg
& ~BRW_MRF_COMPR4
;
3136 const fs_builder
ibld(this, block
, inst
);
3137 const fs_builder hbld
= ibld
.exec_all().group(8, 0);
3139 for (uint8_t i
= 0; i
< inst
->header_size
; i
++) {
3140 if (inst
->src
[i
].file
!= BAD_FILE
) {
3141 fs_reg mov_dst
= retype(dst
, BRW_REGISTER_TYPE_UD
);
3142 fs_reg mov_src
= retype(inst
->src
[i
], BRW_REGISTER_TYPE_UD
);
3143 hbld
.MOV(mov_dst
, mov_src
);
3145 dst
= offset(dst
, hbld
, 1);
3148 if (inst
->dst
.file
== MRF
&& (inst
->dst
.reg
& BRW_MRF_COMPR4
) &&
3149 inst
->exec_size
> 8) {
3150 /* In this case, the payload portion of the LOAD_PAYLOAD isn't
3151 * a straightforward copy. Instead, the result of the
3152 * LOAD_PAYLOAD is treated as interleaved and the first four
3153 * non-header sources are unpacked as:
3164 * This is used for gen <= 5 fb writes.
3166 assert(inst
->exec_size
== 16);
3167 assert(inst
->header_size
+ 4 <= inst
->sources
);
3168 for (uint8_t i
= inst
->header_size
; i
< inst
->header_size
+ 4; i
++) {
3169 if (inst
->src
[i
].file
!= BAD_FILE
) {
3170 if (devinfo
->has_compr4
) {
3171 fs_reg compr4_dst
= retype(dst
, inst
->src
[i
].type
);
3172 compr4_dst
.reg
|= BRW_MRF_COMPR4
;
3173 ibld
.MOV(compr4_dst
, inst
->src
[i
]);
3175 /* Platform doesn't have COMPR4. We have to fake it */
3176 fs_reg mov_dst
= retype(dst
, inst
->src
[i
].type
);
3177 ibld
.half(0).MOV(mov_dst
, half(inst
->src
[i
], 0));
3179 ibld
.half(1).MOV(mov_dst
, half(inst
->src
[i
], 1));
3186 /* The loop above only ever incremented us through the first set
3187 * of 4 registers. However, thanks to the magic of COMPR4, we
3188 * actually wrote to the first 8 registers, so we need to take
3189 * that into account now.
3193 /* The COMPR4 code took care of the first 4 sources. We'll let
3194 * the regular path handle any remaining sources. Yes, we are
3195 * modifying the instruction but we're about to delete it so
3196 * this really doesn't hurt anything.
3198 inst
->header_size
+= 4;
3201 for (uint8_t i
= inst
->header_size
; i
< inst
->sources
; i
++) {
3202 if (inst
->src
[i
].file
!= BAD_FILE
)
3203 ibld
.MOV(retype(dst
, inst
->src
[i
].type
), inst
->src
[i
]);
3204 dst
= offset(dst
, ibld
, 1);
3207 inst
->remove(block
);
3212 invalidate_live_intervals();
3218 fs_visitor::lower_integer_multiplication()
3220 bool progress
= false;
3222 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
3223 const fs_builder
ibld(this, block
, inst
);
3225 if (inst
->opcode
== BRW_OPCODE_MUL
) {
3226 if (inst
->dst
.is_accumulator() ||
3227 (inst
->dst
.type
!= BRW_REGISTER_TYPE_D
&&
3228 inst
->dst
.type
!= BRW_REGISTER_TYPE_UD
))
3231 /* Gen8's MUL instruction can do a 32-bit x 32-bit -> 32-bit
3232 * operation directly, but CHV/BXT cannot.
3234 if (devinfo
->gen
>= 8 &&
3235 !devinfo
->is_cherryview
&& !devinfo
->is_broxton
)
3238 if (inst
->src
[1].file
== IMM
&&
3239 inst
->src
[1].fixed_hw_reg
.dw1
.ud
< (1 << 16)) {
3240 /* The MUL instruction isn't commutative. On Gen <= 6, only the low
3241 * 16-bits of src0 are read, and on Gen >= 7 only the low 16-bits of
3244 * If multiplying by an immediate value that fits in 16-bits, do a
3245 * single MUL instruction with that value in the proper location.
3247 if (devinfo
->gen
< 7) {
3248 fs_reg
imm(GRF
, alloc
.allocate(dispatch_width
/ 8),
3250 ibld
.MOV(imm
, inst
->src
[1]);
3251 ibld
.MUL(inst
->dst
, imm
, inst
->src
[0]);
3253 ibld
.MUL(inst
->dst
, inst
->src
[0], inst
->src
[1]);
3256 /* Gen < 8 (and some Gen8+ low-power parts like Cherryview) cannot
3257 * do 32-bit integer multiplication in one instruction, but instead
3258 * must do a sequence (which actually calculates a 64-bit result):
3260 * mul(8) acc0<1>D g3<8,8,1>D g4<8,8,1>D
3261 * mach(8) null g3<8,8,1>D g4<8,8,1>D
3262 * mov(8) g2<1>D acc0<8,8,1>D
3264 * But on Gen > 6, the ability to use second accumulator register
3265 * (acc1) for non-float data types was removed, preventing a simple
3266 * implementation in SIMD16. A 16-channel result can be calculated by
3267 * executing the three instructions twice in SIMD8, once with quarter
3268 * control of 1Q for the first eight channels and again with 2Q for
3269 * the second eight channels.
3271 * Which accumulator register is implicitly accessed (by AccWrEnable
3272 * for instance) is determined by the quarter control. Unfortunately
3273 * Ivybridge (and presumably Baytrail) has a hardware bug in which an
3274 * implicit accumulator access by an instruction with 2Q will access
3275 * acc1 regardless of whether the data type is usable in acc1.
3277 * Specifically, the 2Q mach(8) writes acc1 which does not exist for
3278 * integer data types.
3280 * Since we only want the low 32-bits of the result, we can do two
3281 * 32-bit x 16-bit multiplies (like the mul and mach are doing), and
3282 * adjust the high result and add them (like the mach is doing):
3284 * mul(8) g7<1>D g3<8,8,1>D g4.0<8,8,1>UW
3285 * mul(8) g8<1>D g3<8,8,1>D g4.1<8,8,1>UW
3286 * shl(8) g9<1>D g8<8,8,1>D 16D
3287 * add(8) g2<1>D g7<8,8,1>D g8<8,8,1>D
3289 * We avoid the shl instruction by realizing that we only want to add
3290 * the low 16-bits of the "high" result to the high 16-bits of the
3291 * "low" result and using proper regioning on the add:
3293 * mul(8) g7<1>D g3<8,8,1>D g4.0<16,8,2>UW
3294 * mul(8) g8<1>D g3<8,8,1>D g4.1<16,8,2>UW
3295 * add(8) g7.1<2>UW g7.1<16,8,2>UW g8<16,8,2>UW
3297 * Since it does not use the (single) accumulator register, we can
3298 * schedule multi-component multiplications much better.
3301 fs_reg orig_dst
= inst
->dst
;
3302 if (orig_dst
.is_null() || orig_dst
.file
== MRF
) {
3303 inst
->dst
= fs_reg(GRF
, alloc
.allocate(dispatch_width
/ 8),
3306 fs_reg low
= inst
->dst
;
3307 fs_reg
high(GRF
, alloc
.allocate(dispatch_width
/ 8),
3310 if (devinfo
->gen
>= 7) {
3311 fs_reg src1_0_w
= inst
->src
[1];
3312 fs_reg src1_1_w
= inst
->src
[1];
3314 if (inst
->src
[1].file
== IMM
) {
3315 src1_0_w
.fixed_hw_reg
.dw1
.ud
&= 0xffff;
3316 src1_1_w
.fixed_hw_reg
.dw1
.ud
>>= 16;
3318 src1_0_w
.type
= BRW_REGISTER_TYPE_UW
;
3319 if (src1_0_w
.stride
!= 0) {
3320 assert(src1_0_w
.stride
== 1);
3321 src1_0_w
.stride
= 2;
3324 src1_1_w
.type
= BRW_REGISTER_TYPE_UW
;
3325 if (src1_1_w
.stride
!= 0) {
3326 assert(src1_1_w
.stride
== 1);
3327 src1_1_w
.stride
= 2;
3329 src1_1_w
.subreg_offset
+= type_sz(BRW_REGISTER_TYPE_UW
);
3331 ibld
.MUL(low
, inst
->src
[0], src1_0_w
);
3332 ibld
.MUL(high
, inst
->src
[0], src1_1_w
);
3334 fs_reg src0_0_w
= inst
->src
[0];
3335 fs_reg src0_1_w
= inst
->src
[0];
3337 src0_0_w
.type
= BRW_REGISTER_TYPE_UW
;
3338 if (src0_0_w
.stride
!= 0) {
3339 assert(src0_0_w
.stride
== 1);
3340 src0_0_w
.stride
= 2;
3343 src0_1_w
.type
= BRW_REGISTER_TYPE_UW
;
3344 if (src0_1_w
.stride
!= 0) {
3345 assert(src0_1_w
.stride
== 1);
3346 src0_1_w
.stride
= 2;
3348 src0_1_w
.subreg_offset
+= type_sz(BRW_REGISTER_TYPE_UW
);
3350 ibld
.MUL(low
, src0_0_w
, inst
->src
[1]);
3351 ibld
.MUL(high
, src0_1_w
, inst
->src
[1]);
3354 fs_reg dst
= inst
->dst
;
3355 dst
.type
= BRW_REGISTER_TYPE_UW
;
3356 dst
.subreg_offset
= 2;
3359 high
.type
= BRW_REGISTER_TYPE_UW
;
3362 low
.type
= BRW_REGISTER_TYPE_UW
;
3363 low
.subreg_offset
= 2;
3366 ibld
.ADD(dst
, low
, high
);
3368 if (inst
->conditional_mod
|| orig_dst
.file
== MRF
) {
3369 set_condmod(inst
->conditional_mod
,
3370 ibld
.MOV(orig_dst
, inst
->dst
));
3374 } else if (inst
->opcode
== SHADER_OPCODE_MULH
) {
3375 /* Should have been lowered to 8-wide. */
3376 assert(inst
->exec_size
<= 8);
3377 const fs_reg acc
= retype(brw_acc_reg(inst
->exec_size
),
3379 fs_inst
*mul
= ibld
.MUL(acc
, inst
->src
[0], inst
->src
[1]);
3380 fs_inst
*mach
= ibld
.MACH(inst
->dst
, inst
->src
[0], inst
->src
[1]);
3382 if (devinfo
->gen
>= 8) {
3383 /* Until Gen8, integer multiplies read 32-bits from one source,
3384 * and 16-bits from the other, and relying on the MACH instruction
3385 * to generate the high bits of the result.
3387 * On Gen8, the multiply instruction does a full 32x32-bit
3388 * multiply, but in order to do a 64-bit multiply we can simulate
3389 * the previous behavior and then use a MACH instruction.
3391 * FINISHME: Don't use source modifiers on src1.
3393 assert(mul
->src
[1].type
== BRW_REGISTER_TYPE_D
||
3394 mul
->src
[1].type
== BRW_REGISTER_TYPE_UD
);
3395 mul
->src
[1].type
= (type_is_signed(mul
->src
[1].type
) ?
3396 BRW_REGISTER_TYPE_W
: BRW_REGISTER_TYPE_UW
);
3397 mul
->src
[1].stride
*= 2;
3399 } else if (devinfo
->gen
== 7 && !devinfo
->is_haswell
&&
3400 inst
->force_sechalf
) {
3401 /* Among other things the quarter control bits influence which
3402 * accumulator register is used by the hardware for instructions
3403 * that access the accumulator implicitly (e.g. MACH). A
3404 * second-half instruction would normally map to acc1, which
3405 * doesn't exist on Gen7 and up (the hardware does emulate it for
3406 * floating-point instructions *only* by taking advantage of the
3407 * extra precision of acc0 not normally used for floating point
3410 * HSW and up are careful enough not to try to access an
3411 * accumulator register that doesn't exist, but on earlier Gen7
3412 * hardware we need to make sure that the quarter control bits are
3413 * zero to avoid non-deterministic behaviour and emit an extra MOV
3414 * to get the result masked correctly according to the current
3417 mach
->force_sechalf
= false;
3418 mach
->force_writemask_all
= true;
3419 mach
->dst
= ibld
.vgrf(inst
->dst
.type
);
3420 ibld
.MOV(inst
->dst
, mach
->dst
);
3426 inst
->remove(block
);
3431 invalidate_live_intervals();
3437 setup_color_payload(const fs_builder
&bld
, const brw_wm_prog_key
*key
,
3438 fs_reg
*dst
, fs_reg color
, unsigned components
)
3440 if (key
->clamp_fragment_color
) {
3441 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 4);
3442 assert(color
.type
== BRW_REGISTER_TYPE_F
);
3444 for (unsigned i
= 0; i
< components
; i
++)
3446 bld
.MOV(offset(tmp
, bld
, i
), offset(color
, bld
, i
)));
3451 for (unsigned i
= 0; i
< components
; i
++)
3452 dst
[i
] = offset(color
, bld
, i
);
3456 lower_fb_write_logical_send(const fs_builder
&bld
, fs_inst
*inst
,
3457 const brw_wm_prog_data
*prog_data
,
3458 const brw_wm_prog_key
*key
,
3459 const fs_visitor::thread_payload
&payload
)
3461 assert(inst
->src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].file
== IMM
);
3462 const brw_device_info
*devinfo
= bld
.shader
->devinfo
;
3463 const fs_reg
&color0
= inst
->src
[FB_WRITE_LOGICAL_SRC_COLOR0
];
3464 const fs_reg
&color1
= inst
->src
[FB_WRITE_LOGICAL_SRC_COLOR1
];
3465 const fs_reg
&src0_alpha
= inst
->src
[FB_WRITE_LOGICAL_SRC_SRC0_ALPHA
];
3466 const fs_reg
&src_depth
= inst
->src
[FB_WRITE_LOGICAL_SRC_SRC_DEPTH
];
3467 const fs_reg
&dst_depth
= inst
->src
[FB_WRITE_LOGICAL_SRC_DST_DEPTH
];
3468 const fs_reg
&src_stencil
= inst
->src
[FB_WRITE_LOGICAL_SRC_SRC_STENCIL
];
3469 fs_reg sample_mask
= inst
->src
[FB_WRITE_LOGICAL_SRC_OMASK
];
3470 const unsigned components
=
3471 inst
->src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].fixed_hw_reg
.dw1
.ud
;
3473 /* We can potentially have a message length of up to 15, so we have to set
3474 * base_mrf to either 0 or 1 in order to fit in m0..m15.
3477 int header_size
= 2, payload_header_size
;
3478 unsigned length
= 0;
3480 /* From the Sandy Bridge PRM, volume 4, page 198:
3482 * "Dispatched Pixel Enables. One bit per pixel indicating
3483 * which pixels were originally enabled when the thread was
3484 * dispatched. This field is only required for the end-of-
3485 * thread message and on all dual-source messages."
3487 if (devinfo
->gen
>= 6 &&
3488 (devinfo
->is_haswell
|| devinfo
->gen
>= 8 || !prog_data
->uses_kill
) &&
3489 color1
.file
== BAD_FILE
&&
3490 key
->nr_color_regions
== 1) {
3494 if (header_size
!= 0) {
3495 assert(header_size
== 2);
3496 /* Allocate 2 registers for a header */
3500 if (payload
.aa_dest_stencil_reg
) {
3501 sources
[length
] = fs_reg(GRF
, bld
.shader
->alloc
.allocate(1));
3502 bld
.group(8, 0).exec_all().annotate("FB write stencil/AA alpha")
3503 .MOV(sources
[length
],
3504 fs_reg(brw_vec8_grf(payload
.aa_dest_stencil_reg
, 0)));
3508 if (prog_data
->uses_omask
) {
3509 sources
[length
] = fs_reg(GRF
, bld
.shader
->alloc
.allocate(1),
3510 BRW_REGISTER_TYPE_UD
);
3512 /* Hand over gl_SampleMask. Only the lower 16 bits of each channel are
3513 * relevant. Since it's unsigned single words one vgrf is always
3514 * 16-wide, but only the lower or higher 8 channels will be used by the
3515 * hardware when doing a SIMD8 write depending on whether we have
3516 * selected the subspans for the first or second half respectively.
3518 assert(sample_mask
.file
!= BAD_FILE
&& type_sz(sample_mask
.type
) == 4);
3519 sample_mask
.type
= BRW_REGISTER_TYPE_UW
;
3520 sample_mask
.stride
*= 2;
3522 bld
.exec_all().annotate("FB write oMask")
3523 .MOV(half(retype(sources
[length
], BRW_REGISTER_TYPE_UW
),
3524 inst
->force_sechalf
),
3529 payload_header_size
= length
;
3531 if (src0_alpha
.file
!= BAD_FILE
) {
3532 /* FIXME: This is being passed at the wrong location in the payload and
3533 * doesn't work when gl_SampleMask and MRTs are used simultaneously.
3534 * It's supposed to be immediately before oMask but there seems to be no
3535 * reasonable way to pass them in the correct order because LOAD_PAYLOAD
3536 * requires header sources to form a contiguous segment at the beginning
3537 * of the message and src0_alpha has per-channel semantics.
3539 setup_color_payload(bld
, key
, &sources
[length
], src0_alpha
, 1);
3543 setup_color_payload(bld
, key
, &sources
[length
], color0
, components
);
3546 if (color1
.file
!= BAD_FILE
) {
3547 setup_color_payload(bld
, key
, &sources
[length
], color1
, components
);
3551 if (src_depth
.file
!= BAD_FILE
) {
3552 sources
[length
] = src_depth
;
3556 if (dst_depth
.file
!= BAD_FILE
) {
3557 sources
[length
] = dst_depth
;
3561 if (src_stencil
.file
!= BAD_FILE
) {
3562 assert(devinfo
->gen
>= 9);
3563 assert(bld
.dispatch_width() != 16);
3565 sources
[length
] = bld
.vgrf(BRW_REGISTER_TYPE_UD
);
3566 bld
.exec_all().annotate("FB write OS")
3567 .emit(FS_OPCODE_PACK_STENCIL_REF
, sources
[length
],
3568 retype(src_stencil
, BRW_REGISTER_TYPE_UB
));
3573 if (devinfo
->gen
>= 7) {
3574 /* Send from the GRF */
3575 fs_reg payload
= fs_reg(GRF
, -1, BRW_REGISTER_TYPE_F
);
3576 load
= bld
.LOAD_PAYLOAD(payload
, sources
, length
, payload_header_size
);
3577 payload
.reg
= bld
.shader
->alloc
.allocate(load
->regs_written
);
3578 load
->dst
= payload
;
3580 inst
->src
[0] = payload
;
3581 inst
->resize_sources(1);
3582 inst
->base_mrf
= -1;
3584 /* Send from the MRF */
3585 load
= bld
.LOAD_PAYLOAD(fs_reg(MRF
, 1, BRW_REGISTER_TYPE_F
),
3586 sources
, length
, payload_header_size
);
3588 /* On pre-SNB, we have to interlace the color values. LOAD_PAYLOAD
3589 * will do this for us if we just give it a COMPR4 destination.
3591 if (devinfo
->gen
< 6 && bld
.dispatch_width() == 16)
3592 load
->dst
.reg
|= BRW_MRF_COMPR4
;
3594 inst
->resize_sources(0);
3598 inst
->opcode
= FS_OPCODE_FB_WRITE
;
3599 inst
->mlen
= load
->regs_written
;
3600 inst
->header_size
= header_size
;
3604 lower_sampler_logical_send_gen4(const fs_builder
&bld
, fs_inst
*inst
, opcode op
,
3605 const fs_reg
&coordinate
,
3606 const fs_reg
&shadow_c
,
3607 const fs_reg
&lod
, const fs_reg
&lod2
,
3608 const fs_reg
&sampler
,
3609 unsigned coord_components
,
3610 unsigned grad_components
)
3612 const bool has_lod
= (op
== SHADER_OPCODE_TXL
|| op
== FS_OPCODE_TXB
||
3613 op
== SHADER_OPCODE_TXF
|| op
== SHADER_OPCODE_TXS
);
3614 fs_reg
msg_begin(MRF
, 1, BRW_REGISTER_TYPE_F
);
3615 fs_reg msg_end
= msg_begin
;
3618 msg_end
= offset(msg_end
, bld
.group(8, 0), 1);
3620 for (unsigned i
= 0; i
< coord_components
; i
++)
3621 bld
.MOV(retype(offset(msg_end
, bld
, i
), coordinate
.type
),
3622 offset(coordinate
, bld
, i
));
3624 msg_end
= offset(msg_end
, bld
, coord_components
);
3626 /* Messages other than SAMPLE and RESINFO in SIMD16 and TXD in SIMD8
3627 * require all three components to be present and zero if they are unused.
3629 if (coord_components
> 0 &&
3630 (has_lod
|| shadow_c
.file
!= BAD_FILE
||
3631 (op
== SHADER_OPCODE_TEX
&& bld
.dispatch_width() == 8))) {
3632 for (unsigned i
= coord_components
; i
< 3; i
++)
3633 bld
.MOV(offset(msg_end
, bld
, i
), fs_reg(0.0f
));
3635 msg_end
= offset(msg_end
, bld
, 3 - coord_components
);
3638 if (op
== SHADER_OPCODE_TXD
) {
3639 /* TXD unsupported in SIMD16 mode. */
3640 assert(bld
.dispatch_width() == 8);
3642 /* the slots for u and v are always present, but r is optional */
3643 if (coord_components
< 2)
3644 msg_end
= offset(msg_end
, bld
, 2 - coord_components
);
3647 * dPdx = dudx, dvdx, drdx
3648 * dPdy = dudy, dvdy, drdy
3650 * 1-arg: Does not exist.
3652 * 2-arg: dudx dvdx dudy dvdy
3653 * dPdx.x dPdx.y dPdy.x dPdy.y
3656 * 3-arg: dudx dvdx drdx dudy dvdy drdy
3657 * dPdx.x dPdx.y dPdx.z dPdy.x dPdy.y dPdy.z
3658 * m5 m6 m7 m8 m9 m10
3660 for (unsigned i
= 0; i
< grad_components
; i
++)
3661 bld
.MOV(offset(msg_end
, bld
, i
), offset(lod
, bld
, i
));
3663 msg_end
= offset(msg_end
, bld
, MAX2(grad_components
, 2));
3665 for (unsigned i
= 0; i
< grad_components
; i
++)
3666 bld
.MOV(offset(msg_end
, bld
, i
), offset(lod2
, bld
, i
));
3668 msg_end
= offset(msg_end
, bld
, MAX2(grad_components
, 2));
3672 /* Bias/LOD with shadow comparitor is unsupported in SIMD16 -- *Without*
3673 * shadow comparitor (including RESINFO) it's unsupported in SIMD8 mode.
3675 assert(shadow_c
.file
!= BAD_FILE
? bld
.dispatch_width() == 8 :
3676 bld
.dispatch_width() == 16);
3678 const brw_reg_type type
=
3679 (op
== SHADER_OPCODE_TXF
|| op
== SHADER_OPCODE_TXS
?
3680 BRW_REGISTER_TYPE_UD
: BRW_REGISTER_TYPE_F
);
3681 bld
.MOV(retype(msg_end
, type
), lod
);
3682 msg_end
= offset(msg_end
, bld
, 1);
3685 if (shadow_c
.file
!= BAD_FILE
) {
3686 if (op
== SHADER_OPCODE_TEX
&& bld
.dispatch_width() == 8) {
3687 /* There's no plain shadow compare message, so we use shadow
3688 * compare with a bias of 0.0.
3690 bld
.MOV(msg_end
, fs_reg(0.0f
));
3691 msg_end
= offset(msg_end
, bld
, 1);
3694 bld
.MOV(msg_end
, shadow_c
);
3695 msg_end
= offset(msg_end
, bld
, 1);
3699 inst
->src
[0] = reg_undef
;
3700 inst
->src
[1] = sampler
;
3701 inst
->resize_sources(2);
3702 inst
->base_mrf
= msg_begin
.reg
;
3703 inst
->mlen
= msg_end
.reg
- msg_begin
.reg
;
3704 inst
->header_size
= 1;
3708 lower_sampler_logical_send_gen5(const fs_builder
&bld
, fs_inst
*inst
, opcode op
,
3710 const fs_reg
&shadow_c
,
3711 fs_reg lod
, fs_reg lod2
,
3712 const fs_reg
&sample_index
,
3713 const fs_reg
&sampler
,
3714 const fs_reg
&offset_value
,
3715 unsigned coord_components
,
3716 unsigned grad_components
)
3718 fs_reg
message(MRF
, 2, BRW_REGISTER_TYPE_F
);
3719 fs_reg msg_coords
= message
;
3720 unsigned header_size
= 0;
3722 if (offset_value
.file
!= BAD_FILE
) {
3723 /* The offsets set up by the visitor are in the m1 header, so we can't
3730 for (unsigned i
= 0; i
< coord_components
; i
++) {
3731 bld
.MOV(retype(offset(msg_coords
, bld
, i
), coordinate
.type
), coordinate
);
3732 coordinate
= offset(coordinate
, bld
, 1);
3734 fs_reg msg_end
= offset(msg_coords
, bld
, coord_components
);
3735 fs_reg msg_lod
= offset(msg_coords
, bld
, 4);
3737 if (shadow_c
.file
!= BAD_FILE
) {
3738 fs_reg msg_shadow
= msg_lod
;
3739 bld
.MOV(msg_shadow
, shadow_c
);
3740 msg_lod
= offset(msg_shadow
, bld
, 1);
3745 case SHADER_OPCODE_TXL
:
3747 bld
.MOV(msg_lod
, lod
);
3748 msg_end
= offset(msg_lod
, bld
, 1);
3750 case SHADER_OPCODE_TXD
:
3753 * dPdx = dudx, dvdx, drdx
3754 * dPdy = dudy, dvdy, drdy
3756 * Load up these values:
3757 * - dudx dudy dvdx dvdy drdx drdy
3758 * - dPdx.x dPdy.x dPdx.y dPdy.y dPdx.z dPdy.z
3761 for (unsigned i
= 0; i
< grad_components
; i
++) {
3762 bld
.MOV(msg_end
, lod
);
3763 lod
= offset(lod
, bld
, 1);
3764 msg_end
= offset(msg_end
, bld
, 1);
3766 bld
.MOV(msg_end
, lod2
);
3767 lod2
= offset(lod2
, bld
, 1);
3768 msg_end
= offset(msg_end
, bld
, 1);
3771 case SHADER_OPCODE_TXS
:
3772 msg_lod
= retype(msg_end
, BRW_REGISTER_TYPE_UD
);
3773 bld
.MOV(msg_lod
, lod
);
3774 msg_end
= offset(msg_lod
, bld
, 1);
3776 case SHADER_OPCODE_TXF
:
3777 msg_lod
= offset(msg_coords
, bld
, 3);
3778 bld
.MOV(retype(msg_lod
, BRW_REGISTER_TYPE_UD
), lod
);
3779 msg_end
= offset(msg_lod
, bld
, 1);
3781 case SHADER_OPCODE_TXF_CMS
:
3782 msg_lod
= offset(msg_coords
, bld
, 3);
3784 bld
.MOV(retype(msg_lod
, BRW_REGISTER_TYPE_UD
), fs_reg(0u));
3786 bld
.MOV(retype(offset(msg_lod
, bld
, 1), BRW_REGISTER_TYPE_UD
), sample_index
);
3787 msg_end
= offset(msg_lod
, bld
, 2);
3794 inst
->src
[0] = reg_undef
;
3795 inst
->src
[1] = sampler
;
3796 inst
->resize_sources(2);
3797 inst
->base_mrf
= message
.reg
;
3798 inst
->mlen
= msg_end
.reg
- message
.reg
;
3799 inst
->header_size
= header_size
;
3801 /* Message length > MAX_SAMPLER_MESSAGE_SIZE disallowed by hardware. */
3802 assert(inst
->mlen
<= MAX_SAMPLER_MESSAGE_SIZE
);
3806 is_high_sampler(const struct brw_device_info
*devinfo
, const fs_reg
&sampler
)
3808 if (devinfo
->gen
< 8 && !devinfo
->is_haswell
)
3811 return sampler
.file
!= IMM
|| sampler
.fixed_hw_reg
.dw1
.ud
>= 16;
3815 lower_sampler_logical_send_gen7(const fs_builder
&bld
, fs_inst
*inst
, opcode op
,
3817 const fs_reg
&shadow_c
,
3818 fs_reg lod
, fs_reg lod2
,
3819 const fs_reg
&sample_index
,
3820 const fs_reg
&mcs
, const fs_reg
&sampler
,
3821 fs_reg offset_value
,
3822 unsigned coord_components
,
3823 unsigned grad_components
)
3825 const brw_device_info
*devinfo
= bld
.shader
->devinfo
;
3826 int reg_width
= bld
.dispatch_width() / 8;
3827 unsigned header_size
= 0, length
= 0;
3828 fs_reg sources
[MAX_SAMPLER_MESSAGE_SIZE
];
3829 for (unsigned i
= 0; i
< ARRAY_SIZE(sources
); i
++)
3830 sources
[i
] = bld
.vgrf(BRW_REGISTER_TYPE_F
);
3832 if (op
== SHADER_OPCODE_TG4
|| op
== SHADER_OPCODE_TG4_OFFSET
||
3833 offset_value
.file
!= BAD_FILE
||
3834 is_high_sampler(devinfo
, sampler
)) {
3835 /* For general texture offsets (no txf workaround), we need a header to
3836 * put them in. Note that we're only reserving space for it in the
3837 * message payload as it will be initialized implicitly by the
3840 * TG4 needs to place its channel select in the header, for interaction
3841 * with ARB_texture_swizzle. The sampler index is only 4-bits, so for
3842 * larger sampler numbers we need to offset the Sampler State Pointer in
3846 sources
[0] = fs_reg();
3850 if (shadow_c
.file
!= BAD_FILE
) {
3851 bld
.MOV(sources
[length
], shadow_c
);
3855 bool coordinate_done
= false;
3857 /* The sampler can only meaningfully compute LOD for fragment shader
3858 * messages. For all other stages, we change the opcode to TXL and
3859 * hardcode the LOD to 0.
3861 if (bld
.shader
->stage
!= MESA_SHADER_FRAGMENT
&&
3862 op
== SHADER_OPCODE_TEX
) {
3863 op
= SHADER_OPCODE_TXL
;
3867 /* Set up the LOD info */
3870 case SHADER_OPCODE_TXL
:
3871 bld
.MOV(sources
[length
], lod
);
3874 case SHADER_OPCODE_TXD
:
3875 /* TXD should have been lowered in SIMD16 mode. */
3876 assert(bld
.dispatch_width() == 8);
3878 /* Load dPdx and the coordinate together:
3879 * [hdr], [ref], x, dPdx.x, dPdy.x, y, dPdx.y, dPdy.y, z, dPdx.z, dPdy.z
3881 for (unsigned i
= 0; i
< coord_components
; i
++) {
3882 bld
.MOV(sources
[length
], coordinate
);
3883 coordinate
= offset(coordinate
, bld
, 1);
3886 /* For cube map array, the coordinate is (u,v,r,ai) but there are
3887 * only derivatives for (u, v, r).
3889 if (i
< grad_components
) {
3890 bld
.MOV(sources
[length
], lod
);
3891 lod
= offset(lod
, bld
, 1);
3894 bld
.MOV(sources
[length
], lod2
);
3895 lod2
= offset(lod2
, bld
, 1);
3900 coordinate_done
= true;
3902 case SHADER_OPCODE_TXS
:
3903 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
), lod
);
3906 case SHADER_OPCODE_TXF
:
3907 /* Unfortunately, the parameters for LD are intermixed: u, lod, v, r.
3908 * On Gen9 they are u, v, lod, r
3910 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
), coordinate
);
3911 coordinate
= offset(coordinate
, bld
, 1);
3914 if (devinfo
->gen
>= 9) {
3915 if (coord_components
>= 2) {
3916 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
), coordinate
);
3917 coordinate
= offset(coordinate
, bld
, 1);
3922 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
), lod
);
3925 for (unsigned i
= devinfo
->gen
>= 9 ? 2 : 1; i
< coord_components
; i
++) {
3926 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
), coordinate
);
3927 coordinate
= offset(coordinate
, bld
, 1);
3931 coordinate_done
= true;
3933 case SHADER_OPCODE_TXF_CMS
:
3934 case SHADER_OPCODE_TXF_CMS_W
:
3935 case SHADER_OPCODE_TXF_UMS
:
3936 case SHADER_OPCODE_TXF_MCS
:
3937 if (op
== SHADER_OPCODE_TXF_UMS
||
3938 op
== SHADER_OPCODE_TXF_CMS
||
3939 op
== SHADER_OPCODE_TXF_CMS_W
) {
3940 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
), sample_index
);
3944 if (op
== SHADER_OPCODE_TXF_CMS
|| op
== SHADER_OPCODE_TXF_CMS_W
) {
3945 /* Data from the multisample control surface. */
3946 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
), mcs
);
3949 /* On Gen9+ we'll use ld2dms_w instead which has two registers for
3952 if (op
== SHADER_OPCODE_TXF_CMS_W
) {
3953 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
),
3956 offset(mcs
, bld
, 1));
3961 /* There is no offsetting for this message; just copy in the integer
3962 * texture coordinates.
3964 for (unsigned i
= 0; i
< coord_components
; i
++) {
3965 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
), coordinate
);
3966 coordinate
= offset(coordinate
, bld
, 1);
3970 coordinate_done
= true;
3972 case SHADER_OPCODE_TG4_OFFSET
:
3973 /* gather4_po_c should have been lowered in SIMD16 mode. */
3974 assert(bld
.dispatch_width() == 8 || shadow_c
.file
== BAD_FILE
);
3976 /* More crazy intermixing */
3977 for (unsigned i
= 0; i
< 2; i
++) { /* u, v */
3978 bld
.MOV(sources
[length
], coordinate
);
3979 coordinate
= offset(coordinate
, bld
, 1);
3983 for (unsigned i
= 0; i
< 2; i
++) { /* offu, offv */
3984 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
), offset_value
);
3985 offset_value
= offset(offset_value
, bld
, 1);
3989 if (coord_components
== 3) { /* r if present */
3990 bld
.MOV(sources
[length
], coordinate
);
3991 coordinate
= offset(coordinate
, bld
, 1);
3995 coordinate_done
= true;
4001 /* Set up the coordinate (except for cases where it was done above) */
4002 if (!coordinate_done
) {
4003 for (unsigned i
= 0; i
< coord_components
; i
++) {
4004 bld
.MOV(sources
[length
], coordinate
);
4005 coordinate
= offset(coordinate
, bld
, 1);
4012 mlen
= length
* reg_width
- header_size
;
4014 mlen
= length
* reg_width
;
4016 const fs_reg src_payload
= fs_reg(GRF
, bld
.shader
->alloc
.allocate(mlen
),
4017 BRW_REGISTER_TYPE_F
);
4018 bld
.LOAD_PAYLOAD(src_payload
, sources
, length
, header_size
);
4020 /* Generate the SEND. */
4022 inst
->src
[0] = src_payload
;
4023 inst
->src
[1] = sampler
;
4024 inst
->resize_sources(2);
4025 inst
->base_mrf
= -1;
4027 inst
->header_size
= header_size
;
4029 /* Message length > MAX_SAMPLER_MESSAGE_SIZE disallowed by hardware. */
4030 assert(inst
->mlen
<= MAX_SAMPLER_MESSAGE_SIZE
);
4034 lower_sampler_logical_send(const fs_builder
&bld
, fs_inst
*inst
, opcode op
)
4036 const brw_device_info
*devinfo
= bld
.shader
->devinfo
;
4037 const fs_reg
&coordinate
= inst
->src
[0];
4038 const fs_reg
&shadow_c
= inst
->src
[1];
4039 const fs_reg
&lod
= inst
->src
[2];
4040 const fs_reg
&lod2
= inst
->src
[3];
4041 const fs_reg
&sample_index
= inst
->src
[4];
4042 const fs_reg
&mcs
= inst
->src
[5];
4043 const fs_reg
&sampler
= inst
->src
[6];
4044 const fs_reg
&offset_value
= inst
->src
[7];
4045 assert(inst
->src
[8].file
== IMM
&& inst
->src
[9].file
== IMM
);
4046 const unsigned coord_components
= inst
->src
[8].fixed_hw_reg
.dw1
.ud
;
4047 const unsigned grad_components
= inst
->src
[9].fixed_hw_reg
.dw1
.ud
;
4049 if (devinfo
->gen
>= 7) {
4050 lower_sampler_logical_send_gen7(bld
, inst
, op
, coordinate
,
4051 shadow_c
, lod
, lod2
, sample_index
,
4052 mcs
, sampler
, offset_value
,
4053 coord_components
, grad_components
);
4054 } else if (devinfo
->gen
>= 5) {
4055 lower_sampler_logical_send_gen5(bld
, inst
, op
, coordinate
,
4056 shadow_c
, lod
, lod2
, sample_index
,
4057 sampler
, offset_value
,
4058 coord_components
, grad_components
);
4060 lower_sampler_logical_send_gen4(bld
, inst
, op
, coordinate
,
4061 shadow_c
, lod
, lod2
, sampler
,
4062 coord_components
, grad_components
);
4067 * Initialize the header present in some typed and untyped surface
4071 emit_surface_header(const fs_builder
&bld
, const fs_reg
&sample_mask
)
4073 fs_builder ubld
= bld
.exec_all().group(8, 0);
4074 const fs_reg dst
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
4075 ubld
.MOV(dst
, fs_reg(0));
4076 ubld
.MOV(component(dst
, 7), sample_mask
);
4081 lower_surface_logical_send(const fs_builder
&bld
, fs_inst
*inst
, opcode op
,
4082 const fs_reg
&sample_mask
)
4084 /* Get the logical send arguments. */
4085 const fs_reg
&addr
= inst
->src
[0];
4086 const fs_reg
&src
= inst
->src
[1];
4087 const fs_reg
&surface
= inst
->src
[2];
4088 const UNUSED fs_reg
&dims
= inst
->src
[3];
4089 const fs_reg
&arg
= inst
->src
[4];
4091 /* Calculate the total number of components of the payload. */
4092 const unsigned addr_sz
= inst
->components_read(0);
4093 const unsigned src_sz
= inst
->components_read(1);
4094 const unsigned header_sz
= (sample_mask
.file
== BAD_FILE
? 0 : 1);
4095 const unsigned sz
= header_sz
+ addr_sz
+ src_sz
;
4097 /* Allocate space for the payload. */
4098 fs_reg
*const components
= new fs_reg
[sz
];
4099 const fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, sz
);
4102 /* Construct the payload. */
4104 components
[n
++] = emit_surface_header(bld
, sample_mask
);
4106 for (unsigned i
= 0; i
< addr_sz
; i
++)
4107 components
[n
++] = offset(addr
, bld
, i
);
4109 for (unsigned i
= 0; i
< src_sz
; i
++)
4110 components
[n
++] = offset(src
, bld
, i
);
4112 bld
.LOAD_PAYLOAD(payload
, components
, sz
, header_sz
);
4114 /* Update the original instruction. */
4116 inst
->mlen
= header_sz
+ (addr_sz
+ src_sz
) * inst
->exec_size
/ 8;
4117 inst
->header_size
= header_sz
;
4119 inst
->src
[0] = payload
;
4120 inst
->src
[1] = surface
;
4122 inst
->resize_sources(3);
4124 delete[] components
;
4128 fs_visitor::lower_logical_sends()
4130 bool progress
= false;
4132 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
4133 const fs_builder
ibld(this, block
, inst
);
4135 switch (inst
->opcode
) {
4136 case FS_OPCODE_FB_WRITE_LOGICAL
:
4137 assert(stage
== MESA_SHADER_FRAGMENT
);
4138 lower_fb_write_logical_send(ibld
, inst
,
4139 (const brw_wm_prog_data
*)prog_data
,
4140 (const brw_wm_prog_key
*)key
,
4144 case SHADER_OPCODE_TEX_LOGICAL
:
4145 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TEX
);
4148 case SHADER_OPCODE_TXD_LOGICAL
:
4149 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXD
);
4152 case SHADER_OPCODE_TXF_LOGICAL
:
4153 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF
);
4156 case SHADER_OPCODE_TXL_LOGICAL
:
4157 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXL
);
4160 case SHADER_OPCODE_TXS_LOGICAL
:
4161 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXS
);
4164 case FS_OPCODE_TXB_LOGICAL
:
4165 lower_sampler_logical_send(ibld
, inst
, FS_OPCODE_TXB
);
4168 case SHADER_OPCODE_TXF_CMS_LOGICAL
:
4169 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_CMS
);
4172 case SHADER_OPCODE_TXF_CMS_W_LOGICAL
:
4173 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_CMS_W
);
4176 case SHADER_OPCODE_TXF_UMS_LOGICAL
:
4177 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_UMS
);
4180 case SHADER_OPCODE_TXF_MCS_LOGICAL
:
4181 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_MCS
);
4184 case SHADER_OPCODE_LOD_LOGICAL
:
4185 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_LOD
);
4188 case SHADER_OPCODE_TG4_LOGICAL
:
4189 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TG4
);
4192 case SHADER_OPCODE_TG4_OFFSET_LOGICAL
:
4193 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TG4_OFFSET
);
4196 case SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
:
4197 lower_surface_logical_send(ibld
, inst
,
4198 SHADER_OPCODE_UNTYPED_SURFACE_READ
,
4202 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
:
4203 lower_surface_logical_send(ibld
, inst
,
4204 SHADER_OPCODE_UNTYPED_SURFACE_WRITE
,
4205 ibld
.sample_mask_reg());
4208 case SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL
:
4209 lower_surface_logical_send(ibld
, inst
,
4210 SHADER_OPCODE_UNTYPED_ATOMIC
,
4211 ibld
.sample_mask_reg());
4214 case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
:
4215 lower_surface_logical_send(ibld
, inst
,
4216 SHADER_OPCODE_TYPED_SURFACE_READ
,
4220 case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
:
4221 lower_surface_logical_send(ibld
, inst
,
4222 SHADER_OPCODE_TYPED_SURFACE_WRITE
,
4223 ibld
.sample_mask_reg());
4226 case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
:
4227 lower_surface_logical_send(ibld
, inst
,
4228 SHADER_OPCODE_TYPED_ATOMIC
,
4229 ibld
.sample_mask_reg());
4240 invalidate_live_intervals();
4246 * Get the closest native SIMD width supported by the hardware for instruction
4247 * \p inst. The instruction will be left untouched by
4248 * fs_visitor::lower_simd_width() if the returned value is equal to the
4249 * original execution size.
4252 get_lowered_simd_width(const struct brw_device_info
*devinfo
,
4253 const fs_inst
*inst
)
4255 switch (inst
->opcode
) {
4256 case BRW_OPCODE_MOV
:
4257 case BRW_OPCODE_SEL
:
4258 case BRW_OPCODE_NOT
:
4259 case BRW_OPCODE_AND
:
4261 case BRW_OPCODE_XOR
:
4262 case BRW_OPCODE_SHR
:
4263 case BRW_OPCODE_SHL
:
4264 case BRW_OPCODE_ASR
:
4265 case BRW_OPCODE_CMP
:
4266 case BRW_OPCODE_CMPN
:
4267 case BRW_OPCODE_CSEL
:
4268 case BRW_OPCODE_F32TO16
:
4269 case BRW_OPCODE_F16TO32
:
4270 case BRW_OPCODE_BFREV
:
4271 case BRW_OPCODE_BFE
:
4272 case BRW_OPCODE_BFI1
:
4273 case BRW_OPCODE_BFI2
:
4274 case BRW_OPCODE_ADD
:
4275 case BRW_OPCODE_MUL
:
4276 case BRW_OPCODE_AVG
:
4277 case BRW_OPCODE_FRC
:
4278 case BRW_OPCODE_RNDU
:
4279 case BRW_OPCODE_RNDD
:
4280 case BRW_OPCODE_RNDE
:
4281 case BRW_OPCODE_RNDZ
:
4282 case BRW_OPCODE_LZD
:
4283 case BRW_OPCODE_FBH
:
4284 case BRW_OPCODE_FBL
:
4285 case BRW_OPCODE_CBIT
:
4286 case BRW_OPCODE_SAD2
:
4287 case BRW_OPCODE_MAD
:
4288 case BRW_OPCODE_LRP
:
4289 case SHADER_OPCODE_RCP
:
4290 case SHADER_OPCODE_RSQ
:
4291 case SHADER_OPCODE_SQRT
:
4292 case SHADER_OPCODE_EXP2
:
4293 case SHADER_OPCODE_LOG2
:
4294 case SHADER_OPCODE_POW
:
4295 case SHADER_OPCODE_INT_QUOTIENT
:
4296 case SHADER_OPCODE_INT_REMAINDER
:
4297 case SHADER_OPCODE_SIN
:
4298 case SHADER_OPCODE_COS
: {
4299 /* According to the PRMs:
4300 * "A. In Direct Addressing mode, a source cannot span more than 2
4301 * adjacent GRF registers.
4302 * B. A destination cannot span more than 2 adjacent GRF registers."
4304 * Look for the source or destination with the largest register region
4305 * which is the one that is going to limit the overal execution size of
4306 * the instruction due to this rule.
4308 unsigned reg_count
= inst
->regs_written
;
4310 for (unsigned i
= 0; i
< inst
->sources
; i
++)
4311 reg_count
= MAX2(reg_count
, (unsigned)inst
->regs_read(i
));
4313 /* Calculate the maximum execution size of the instruction based on the
4314 * factor by which it goes over the hardware limit of 2 GRFs.
4316 return inst
->exec_size
/ DIV_ROUND_UP(reg_count
, 2);
4318 case SHADER_OPCODE_MULH
:
4319 /* MULH is lowered to the MUL/MACH sequence using the accumulator, which
4320 * is 8-wide on Gen7+.
4322 return (devinfo
->gen
>= 7 ? 8 : inst
->exec_size
);
4324 case FS_OPCODE_FB_WRITE_LOGICAL
:
4325 /* Gen6 doesn't support SIMD16 depth writes but we cannot handle them
4328 assert(devinfo
->gen
!= 6 ||
4329 inst
->src
[FB_WRITE_LOGICAL_SRC_SRC_DEPTH
].file
== BAD_FILE
||
4330 inst
->exec_size
== 8);
4331 /* Dual-source FB writes are unsupported in SIMD16 mode. */
4332 return (inst
->src
[FB_WRITE_LOGICAL_SRC_COLOR1
].file
!= BAD_FILE
?
4333 8 : inst
->exec_size
);
4335 case SHADER_OPCODE_TXD_LOGICAL
:
4336 /* TXD is unsupported in SIMD16 mode. */
4339 case SHADER_OPCODE_TG4_OFFSET_LOGICAL
: {
4340 /* gather4_po_c is unsupported in SIMD16 mode. */
4341 const fs_reg
&shadow_c
= inst
->src
[1];
4342 return (shadow_c
.file
!= BAD_FILE
? 8 : inst
->exec_size
);
4344 case SHADER_OPCODE_TXL_LOGICAL
:
4345 case FS_OPCODE_TXB_LOGICAL
: {
4346 /* Gen4 doesn't have SIMD8 non-shadow-compare bias/LOD instructions, and
4347 * Gen4-6 can't support TXL and TXB with shadow comparison in SIMD16
4348 * mode because the message exceeds the maximum length of 11.
4350 const fs_reg
&shadow_c
= inst
->src
[1];
4351 if (devinfo
->gen
== 4 && shadow_c
.file
== BAD_FILE
)
4353 else if (devinfo
->gen
< 7 && shadow_c
.file
!= BAD_FILE
)
4356 return inst
->exec_size
;
4358 case SHADER_OPCODE_TXF_LOGICAL
:
4359 case SHADER_OPCODE_TXS_LOGICAL
:
4360 /* Gen4 doesn't have SIMD8 variants for the RESINFO and LD-with-LOD
4361 * messages. Use SIMD16 instead.
4363 if (devinfo
->gen
== 4)
4366 return inst
->exec_size
;
4368 case SHADER_OPCODE_TXF_CMS_W_LOGICAL
: {
4369 /* This opcode can take up to 6 arguments which means that in some
4370 * circumstances it can end up with a message that is too long in SIMD16
4373 const unsigned coord_components
= inst
->src
[8].fixed_hw_reg
.dw1
.ud
;
4374 /* First three arguments are the sample index and the two arguments for
4377 if ((coord_components
+ 3) * 2 > MAX_SAMPLER_MESSAGE_SIZE
)
4380 return inst
->exec_size
;
4383 case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
:
4384 case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
:
4385 case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
:
4389 return inst
->exec_size
;
4394 * The \p rows array of registers represents a \p num_rows by \p num_columns
4395 * matrix in row-major order, write it in column-major order into the register
4396 * passed as destination. \p stride gives the separation between matrix
4397 * elements in the input in fs_builder::dispatch_width() units.
4400 emit_transpose(const fs_builder
&bld
,
4401 const fs_reg
&dst
, const fs_reg
*rows
,
4402 unsigned num_rows
, unsigned num_columns
, unsigned stride
)
4404 fs_reg
*const components
= new fs_reg
[num_rows
* num_columns
];
4406 for (unsigned i
= 0; i
< num_columns
; ++i
) {
4407 for (unsigned j
= 0; j
< num_rows
; ++j
)
4408 components
[num_rows
* i
+ j
] = offset(rows
[j
], bld
, stride
* i
);
4411 bld
.LOAD_PAYLOAD(dst
, components
, num_rows
* num_columns
, 0);
4413 delete[] components
;
4417 fs_visitor::lower_simd_width()
4419 bool progress
= false;
4421 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
4422 const unsigned lower_width
= get_lowered_simd_width(devinfo
, inst
);
4424 if (lower_width
!= inst
->exec_size
) {
4425 /* Builder matching the original instruction. We may also need to
4426 * emit an instruction of width larger than the original, set the
4427 * execution size of the builder to the highest of both for now so
4428 * we're sure that both cases can be handled.
4430 const fs_builder ibld
= bld
.at(block
, inst
)
4431 .exec_all(inst
->force_writemask_all
)
4432 .group(MAX2(inst
->exec_size
, lower_width
),
4433 inst
->force_sechalf
);
4435 /* Split the copies in chunks of the execution width of either the
4436 * original or the lowered instruction, whichever is lower.
4438 const unsigned copy_width
= MIN2(lower_width
, inst
->exec_size
);
4439 const unsigned n
= inst
->exec_size
/ copy_width
;
4440 const unsigned dst_size
= inst
->regs_written
* REG_SIZE
/
4441 inst
->dst
.component_size(inst
->exec_size
);
4444 assert(n
> 0 && n
<= ARRAY_SIZE(dsts
) &&
4445 !inst
->writes_accumulator
&& !inst
->mlen
);
4447 for (unsigned i
= 0; i
< n
; i
++) {
4448 /* Emit a copy of the original instruction with the lowered width.
4449 * If the EOT flag was set throw it away except for the last
4450 * instruction to avoid killing the thread prematurely.
4452 fs_inst split_inst
= *inst
;
4453 split_inst
.exec_size
= lower_width
;
4454 split_inst
.eot
= inst
->eot
&& i
== n
- 1;
4456 /* Select the correct channel enables for the i-th group, then
4457 * transform the sources and destination and emit the lowered
4460 const fs_builder lbld
= ibld
.group(lower_width
, i
);
4462 for (unsigned j
= 0; j
< inst
->sources
; j
++) {
4463 if (inst
->src
[j
].file
!= BAD_FILE
&&
4464 !is_uniform(inst
->src
[j
])) {
4465 /* Get the i-th copy_width-wide chunk of the source. */
4466 const fs_reg src
= horiz_offset(inst
->src
[j
], copy_width
* i
);
4467 const unsigned src_size
= inst
->components_read(j
);
4469 /* Use a trivial transposition to copy one every n
4470 * copy_width-wide components of the register into a
4471 * temporary passed as source to the lowered instruction.
4473 split_inst
.src
[j
] = lbld
.vgrf(inst
->src
[j
].type
, src_size
);
4474 emit_transpose(lbld
.group(copy_width
, 0),
4475 split_inst
.src
[j
], &src
, 1, src_size
, n
);
4479 if (inst
->regs_written
) {
4480 /* Allocate enough space to hold the result of the lowered
4481 * instruction and fix up the number of registers written.
4483 split_inst
.dst
= dsts
[i
] =
4484 lbld
.vgrf(inst
->dst
.type
, dst_size
);
4485 split_inst
.regs_written
=
4486 DIV_ROUND_UP(inst
->regs_written
* lower_width
,
4490 lbld
.emit(split_inst
);
4493 if (inst
->regs_written
) {
4494 /* Distance between useful channels in the temporaries, skipping
4495 * garbage if the lowered instruction is wider than the original.
4497 const unsigned m
= lower_width
/ copy_width
;
4499 /* Interleave the components of the result from the lowered
4500 * instructions. We need to set exec_all() when copying more than
4501 * one half per component, because LOAD_PAYLOAD (in terms of which
4502 * emit_transpose is implemented) can only use the same channel
4503 * enable signals for all of its non-header sources.
4505 emit_transpose(ibld
.exec_all(inst
->exec_size
> copy_width
)
4506 .group(copy_width
, 0),
4507 inst
->dst
, dsts
, n
, dst_size
, m
);
4510 inst
->remove(block
);
4516 invalidate_live_intervals();
4522 fs_visitor::dump_instructions()
4524 dump_instructions(NULL
);
4528 fs_visitor::dump_instructions(const char *name
)
4530 FILE *file
= stderr
;
4531 if (name
&& geteuid() != 0) {
4532 file
= fopen(name
, "w");
4538 calculate_register_pressure();
4539 int ip
= 0, max_pressure
= 0;
4540 foreach_block_and_inst(block
, backend_instruction
, inst
, cfg
) {
4541 max_pressure
= MAX2(max_pressure
, regs_live_at_ip
[ip
]);
4542 fprintf(file
, "{%3d} %4d: ", regs_live_at_ip
[ip
], ip
);
4543 dump_instruction(inst
, file
);
4546 fprintf(file
, "Maximum %3d registers live at once.\n", max_pressure
);
4549 foreach_in_list(backend_instruction
, inst
, &instructions
) {
4550 fprintf(file
, "%4d: ", ip
++);
4551 dump_instruction(inst
, file
);
4555 if (file
!= stderr
) {
4561 fs_visitor::dump_instruction(backend_instruction
*be_inst
)
4563 dump_instruction(be_inst
, stderr
);
4567 fs_visitor::dump_instruction(backend_instruction
*be_inst
, FILE *file
)
4569 fs_inst
*inst
= (fs_inst
*)be_inst
;
4571 if (inst
->predicate
) {
4572 fprintf(file
, "(%cf0.%d) ",
4573 inst
->predicate_inverse
? '-' : '+',
4577 fprintf(file
, "%s", brw_instruction_name(inst
->opcode
));
4579 fprintf(file
, ".sat");
4580 if (inst
->conditional_mod
) {
4581 fprintf(file
, "%s", conditional_modifier
[inst
->conditional_mod
]);
4582 if (!inst
->predicate
&&
4583 (devinfo
->gen
< 5 || (inst
->opcode
!= BRW_OPCODE_SEL
&&
4584 inst
->opcode
!= BRW_OPCODE_IF
&&
4585 inst
->opcode
!= BRW_OPCODE_WHILE
))) {
4586 fprintf(file
, ".f0.%d", inst
->flag_subreg
);
4589 fprintf(file
, "(%d) ", inst
->exec_size
);
4592 fprintf(file
, "(mlen: %d) ", inst
->mlen
);
4595 switch (inst
->dst
.file
) {
4597 fprintf(file
, "vgrf%d", inst
->dst
.reg
);
4598 if (alloc
.sizes
[inst
->dst
.reg
] != inst
->regs_written
||
4599 inst
->dst
.subreg_offset
)
4600 fprintf(file
, "+%d.%d",
4601 inst
->dst
.reg_offset
, inst
->dst
.subreg_offset
);
4604 fprintf(file
, "m%d", inst
->dst
.reg
);
4607 fprintf(file
, "(null)");
4610 fprintf(file
, "***u%d***", inst
->dst
.reg
+ inst
->dst
.reg_offset
);
4613 fprintf(file
, "***attr%d***", inst
->dst
.reg
+ inst
->dst
.reg_offset
);
4616 if (inst
->dst
.fixed_hw_reg
.file
== BRW_ARCHITECTURE_REGISTER_FILE
) {
4617 switch (inst
->dst
.fixed_hw_reg
.nr
) {
4619 fprintf(file
, "null");
4621 case BRW_ARF_ADDRESS
:
4622 fprintf(file
, "a0.%d", inst
->dst
.fixed_hw_reg
.subnr
);
4624 case BRW_ARF_ACCUMULATOR
:
4625 fprintf(file
, "acc%d", inst
->dst
.fixed_hw_reg
.subnr
);
4628 fprintf(file
, "f%d.%d", inst
->dst
.fixed_hw_reg
.nr
& 0xf,
4629 inst
->dst
.fixed_hw_reg
.subnr
);
4632 fprintf(file
, "arf%d.%d", inst
->dst
.fixed_hw_reg
.nr
& 0xf,
4633 inst
->dst
.fixed_hw_reg
.subnr
);
4637 fprintf(file
, "hw_reg%d", inst
->dst
.fixed_hw_reg
.nr
);
4639 if (inst
->dst
.fixed_hw_reg
.subnr
)
4640 fprintf(file
, "+%d", inst
->dst
.fixed_hw_reg
.subnr
);
4643 unreachable("not reached");
4645 fprintf(file
, ":%s, ", brw_reg_type_letters(inst
->dst
.type
));
4647 for (int i
= 0; i
< inst
->sources
; i
++) {
4648 if (inst
->src
[i
].negate
)
4650 if (inst
->src
[i
].abs
)
4652 switch (inst
->src
[i
].file
) {
4654 fprintf(file
, "vgrf%d", inst
->src
[i
].reg
);
4655 if (alloc
.sizes
[inst
->src
[i
].reg
] != (unsigned)inst
->regs_read(i
) ||
4656 inst
->src
[i
].subreg_offset
)
4657 fprintf(file
, "+%d.%d", inst
->src
[i
].reg_offset
,
4658 inst
->src
[i
].subreg_offset
);
4661 fprintf(file
, "***m%d***", inst
->src
[i
].reg
);
4664 fprintf(file
, "attr%d+%d", inst
->src
[i
].reg
, inst
->src
[i
].reg_offset
);
4667 fprintf(file
, "u%d", inst
->src
[i
].reg
+ inst
->src
[i
].reg_offset
);
4668 if (inst
->src
[i
].reladdr
) {
4669 fprintf(file
, "+reladdr");
4670 } else if (inst
->src
[i
].subreg_offset
) {
4671 fprintf(file
, "+%d.%d", inst
->src
[i
].reg_offset
,
4672 inst
->src
[i
].subreg_offset
);
4676 fprintf(file
, "(null)");
4679 switch (inst
->src
[i
].type
) {
4680 case BRW_REGISTER_TYPE_F
:
4681 fprintf(file
, "%ff", inst
->src
[i
].fixed_hw_reg
.dw1
.f
);
4683 case BRW_REGISTER_TYPE_W
:
4684 case BRW_REGISTER_TYPE_D
:
4685 fprintf(file
, "%dd", inst
->src
[i
].fixed_hw_reg
.dw1
.d
);
4687 case BRW_REGISTER_TYPE_UW
:
4688 case BRW_REGISTER_TYPE_UD
:
4689 fprintf(file
, "%uu", inst
->src
[i
].fixed_hw_reg
.dw1
.ud
);
4691 case BRW_REGISTER_TYPE_VF
:
4692 fprintf(file
, "[%-gF, %-gF, %-gF, %-gF]",
4693 brw_vf_to_float((inst
->src
[i
].fixed_hw_reg
.dw1
.ud
>> 0) & 0xff),
4694 brw_vf_to_float((inst
->src
[i
].fixed_hw_reg
.dw1
.ud
>> 8) & 0xff),
4695 brw_vf_to_float((inst
->src
[i
].fixed_hw_reg
.dw1
.ud
>> 16) & 0xff),
4696 brw_vf_to_float((inst
->src
[i
].fixed_hw_reg
.dw1
.ud
>> 24) & 0xff));
4699 fprintf(file
, "???");
4704 if (inst
->src
[i
].fixed_hw_reg
.negate
)
4706 if (inst
->src
[i
].fixed_hw_reg
.abs
)
4708 if (inst
->src
[i
].fixed_hw_reg
.file
== BRW_ARCHITECTURE_REGISTER_FILE
) {
4709 switch (inst
->src
[i
].fixed_hw_reg
.nr
) {
4711 fprintf(file
, "null");
4713 case BRW_ARF_ADDRESS
:
4714 fprintf(file
, "a0.%d", inst
->src
[i
].fixed_hw_reg
.subnr
);
4716 case BRW_ARF_ACCUMULATOR
:
4717 fprintf(file
, "acc%d", inst
->src
[i
].fixed_hw_reg
.subnr
);
4720 fprintf(file
, "f%d.%d", inst
->src
[i
].fixed_hw_reg
.nr
& 0xf,
4721 inst
->src
[i
].fixed_hw_reg
.subnr
);
4724 fprintf(file
, "arf%d.%d", inst
->src
[i
].fixed_hw_reg
.nr
& 0xf,
4725 inst
->src
[i
].fixed_hw_reg
.subnr
);
4729 fprintf(file
, "hw_reg%d", inst
->src
[i
].fixed_hw_reg
.nr
);
4731 if (inst
->src
[i
].fixed_hw_reg
.subnr
)
4732 fprintf(file
, "+%d", inst
->src
[i
].fixed_hw_reg
.subnr
);
4733 if (inst
->src
[i
].fixed_hw_reg
.abs
)
4737 if (inst
->src
[i
].abs
)
4740 if (inst
->src
[i
].file
!= IMM
) {
4741 fprintf(file
, ":%s", brw_reg_type_letters(inst
->src
[i
].type
));
4744 if (i
< inst
->sources
- 1 && inst
->src
[i
+ 1].file
!= BAD_FILE
)
4745 fprintf(file
, ", ");
4750 if (dispatch_width
== 16 && inst
->exec_size
== 8) {
4751 if (inst
->force_sechalf
)
4752 fprintf(file
, "2ndhalf ");
4754 fprintf(file
, "1sthalf ");
4757 fprintf(file
, "\n");
4761 * Possibly returns an instruction that set up @param reg.
4763 * Sometimes we want to take the result of some expression/variable
4764 * dereference tree and rewrite the instruction generating the result
4765 * of the tree. When processing the tree, we know that the
4766 * instructions generated are all writing temporaries that are dead
4767 * outside of this tree. So, if we have some instructions that write
4768 * a temporary, we're free to point that temp write somewhere else.
4770 * Note that this doesn't guarantee that the instruction generated
4771 * only reg -- it might be the size=4 destination of a texture instruction.
4774 fs_visitor::get_instruction_generating_reg(fs_inst
*start
,
4779 end
->is_partial_write() ||
4781 !reg
.equals(end
->dst
)) {
4789 fs_visitor::setup_payload_gen6()
4792 (nir
->info
.inputs_read
& (1 << VARYING_SLOT_POS
)) != 0;
4793 unsigned barycentric_interp_modes
=
4794 (stage
== MESA_SHADER_FRAGMENT
) ?
4795 ((brw_wm_prog_data
*) this->prog_data
)->barycentric_interp_modes
: 0;
4797 assert(devinfo
->gen
>= 6);
4799 /* R0-1: masks, pixel X/Y coordinates. */
4800 payload
.num_regs
= 2;
4801 /* R2: only for 32-pixel dispatch.*/
4803 /* R3-26: barycentric interpolation coordinates. These appear in the
4804 * same order that they appear in the brw_wm_barycentric_interp_mode
4805 * enum. Each set of coordinates occupies 2 registers if dispatch width
4806 * == 8 and 4 registers if dispatch width == 16. Coordinates only
4807 * appear if they were enabled using the "Barycentric Interpolation
4808 * Mode" bits in WM_STATE.
4810 for (int i
= 0; i
< BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT
; ++i
) {
4811 if (barycentric_interp_modes
& (1 << i
)) {
4812 payload
.barycentric_coord_reg
[i
] = payload
.num_regs
;
4813 payload
.num_regs
+= 2;
4814 if (dispatch_width
== 16) {
4815 payload
.num_regs
+= 2;
4820 /* R27: interpolated depth if uses source depth */
4822 payload
.source_depth_reg
= payload
.num_regs
;
4824 if (dispatch_width
== 16) {
4825 /* R28: interpolated depth if not SIMD8. */
4829 /* R29: interpolated W set if GEN6_WM_USES_SOURCE_W. */
4831 payload
.source_w_reg
= payload
.num_regs
;
4833 if (dispatch_width
== 16) {
4834 /* R30: interpolated W if not SIMD8. */
4839 if (stage
== MESA_SHADER_FRAGMENT
) {
4840 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
4841 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
4842 prog_data
->uses_pos_offset
= key
->compute_pos_offset
;
4843 /* R31: MSAA position offsets. */
4844 if (prog_data
->uses_pos_offset
) {
4845 payload
.sample_pos_reg
= payload
.num_regs
;
4850 /* R32: MSAA input coverage mask */
4851 if (nir
->info
.system_values_read
& SYSTEM_BIT_SAMPLE_MASK_IN
) {
4852 assert(devinfo
->gen
>= 7);
4853 payload
.sample_mask_in_reg
= payload
.num_regs
;
4855 if (dispatch_width
== 16) {
4856 /* R33: input coverage mask if not SIMD8. */
4861 /* R34-: bary for 32-pixel. */
4862 /* R58-59: interp W for 32-pixel. */
4864 if (nir
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_DEPTH
)) {
4865 source_depth_to_render_target
= true;
4870 fs_visitor::setup_vs_payload()
4872 /* R0: thread header, R1: urb handles */
4873 payload
.num_regs
= 2;
4877 * We are building the local ID push constant data using the simplest possible
4878 * method. We simply push the local IDs directly as they should appear in the
4879 * registers for the uvec3 gl_LocalInvocationID variable.
4881 * Therefore, for SIMD8, we use 3 full registers, and for SIMD16 we use 6
4882 * registers worth of push constant space.
4884 * Note: Any updates to brw_cs_prog_local_id_payload_dwords,
4885 * fill_local_id_payload or fs_visitor::emit_cs_local_invocation_id_setup need
4888 * FINISHME: There are a few easy optimizations to consider.
4890 * 1. If gl_WorkGroupSize x, y or z is 1, we can just use zero, and there is
4891 * no need for using push constant space for that dimension.
4893 * 2. Since GL_MAX_COMPUTE_WORK_GROUP_SIZE is currently 1024 or less, we can
4894 * easily use 16-bit words rather than 32-bit dwords in the push constant
4897 * 3. If gl_WorkGroupSize x, y or z is small, then we can use bytes for
4898 * conveying the data, and thereby reduce push constant usage.
4902 fs_visitor::setup_gs_payload()
4904 assert(stage
== MESA_SHADER_GEOMETRY
);
4906 struct brw_gs_prog_data
*gs_prog_data
=
4907 (struct brw_gs_prog_data
*) prog_data
;
4908 struct brw_vue_prog_data
*vue_prog_data
=
4909 (struct brw_vue_prog_data
*) prog_data
;
4911 /* R0: thread header, R1: output URB handles */
4912 payload
.num_regs
= 2;
4914 if (gs_prog_data
->include_primitive_id
) {
4915 /* R2: Primitive ID 0..7 */
4919 /* Use a maximum of 32 registers for push-model inputs. */
4920 const unsigned max_push_components
= 32;
4922 /* If pushing our inputs would take too many registers, reduce the URB read
4923 * length (which is in HWords, or 8 registers), and resort to pulling.
4925 * Note that the GS reads <URB Read Length> HWords for every vertex - so we
4926 * have to multiply by VerticesIn to obtain the total storage requirement.
4928 if (8 * vue_prog_data
->urb_read_length
* nir
->info
.gs
.vertices_in
>
4929 max_push_components
) {
4930 gs_prog_data
->base
.include_vue_handles
= true;
4932 /* R3..RN: ICP Handles for each incoming vertex (when using pull model) */
4933 payload
.num_regs
+= nir
->info
.gs
.vertices_in
;
4935 vue_prog_data
->urb_read_length
=
4936 ROUND_DOWN_TO(max_push_components
/ nir
->info
.gs
.vertices_in
, 8) / 8;
4941 fs_visitor::setup_cs_payload()
4943 assert(devinfo
->gen
>= 7);
4944 brw_cs_prog_data
*prog_data
= (brw_cs_prog_data
*) this->prog_data
;
4946 payload
.num_regs
= 1;
4948 if (nir
->info
.system_values_read
& SYSTEM_BIT_LOCAL_INVOCATION_ID
) {
4949 prog_data
->local_invocation_id_regs
= dispatch_width
* 3 / 8;
4950 payload
.local_invocation_id_reg
= payload
.num_regs
;
4951 payload
.num_regs
+= prog_data
->local_invocation_id_regs
;
4956 fs_visitor::calculate_register_pressure()
4958 invalidate_live_intervals();
4959 calculate_live_intervals();
4961 unsigned num_instructions
= 0;
4962 foreach_block(block
, cfg
)
4963 num_instructions
+= block
->instructions
.length();
4965 regs_live_at_ip
= rzalloc_array(mem_ctx
, int, num_instructions
);
4967 for (unsigned reg
= 0; reg
< alloc
.count
; reg
++) {
4968 for (int ip
= virtual_grf_start
[reg
]; ip
<= virtual_grf_end
[reg
]; ip
++)
4969 regs_live_at_ip
[ip
] += alloc
.sizes
[reg
];
4974 fs_visitor::optimize()
4976 /* Start by validating the shader we currently have. */
4979 /* bld is the common builder object pointing at the end of the program we
4980 * used to translate it into i965 IR. For the optimization and lowering
4981 * passes coming next, any code added after the end of the program without
4982 * having explicitly called fs_builder::at() clearly points at a mistake.
4983 * Ideally optimization passes wouldn't be part of the visitor so they
4984 * wouldn't have access to bld at all, but they do, so just in case some
4985 * pass forgets to ask for a location explicitly set it to NULL here to
4986 * make it trip. The dispatch width is initialized to a bogus value to
4987 * make sure that optimizations set the execution controls explicitly to
4988 * match the code they are manipulating instead of relying on the defaults.
4990 bld
= fs_builder(this, 64);
4992 assign_constant_locations();
4993 demote_pull_constants();
4997 split_virtual_grfs();
5000 #define OPT(pass, args...) ({ \
5002 bool this_progress = pass(args); \
5004 if (unlikely(INTEL_DEBUG & DEBUG_OPTIMIZER) && this_progress) { \
5005 char filename[64]; \
5006 snprintf(filename, 64, "%s%d-%s-%02d-%02d-" #pass, \
5007 stage_abbrev, dispatch_width, nir->info.name, iteration, pass_num); \
5009 backend_shader::dump_instructions(filename); \
5014 progress = progress || this_progress; \
5018 if (unlikely(INTEL_DEBUG
& DEBUG_OPTIMIZER
)) {
5020 snprintf(filename
, 64, "%s%d-%s-00-start",
5021 stage_abbrev
, dispatch_width
, nir
->info
.name
);
5023 backend_shader::dump_instructions(filename
);
5026 bool progress
= false;
5030 OPT(lower_simd_width
);
5031 OPT(lower_logical_sends
);
5038 OPT(remove_duplicate_mrf_writes
);
5042 OPT(opt_copy_propagate
);
5043 OPT(opt_predicated_break
, this);
5044 OPT(opt_cmod_propagation
);
5045 OPT(dead_code_eliminate
);
5046 OPT(opt_peephole_sel
);
5047 OPT(dead_control_flow_eliminate
, this);
5048 OPT(opt_register_renaming
);
5049 OPT(opt_redundant_discard_jumps
);
5050 OPT(opt_saturate_propagation
);
5051 OPT(opt_zero_samples
);
5052 OPT(register_coalesce
);
5053 OPT(compute_to_mrf
);
5054 OPT(eliminate_find_live_channel
);
5056 OPT(compact_virtual_grfs
);
5061 OPT(opt_sampler_eot
);
5063 if (OPT(lower_load_payload
)) {
5064 split_virtual_grfs();
5065 OPT(register_coalesce
);
5066 OPT(compute_to_mrf
);
5067 OPT(dead_code_eliminate
);
5070 OPT(opt_combine_constants
);
5071 OPT(lower_integer_multiplication
);
5073 lower_uniform_pull_constant_loads();
5079 * Three source instruction must have a GRF/MRF destination register.
5080 * ARF NULL is not allowed. Fix that up by allocating a temporary GRF.
5083 fs_visitor::fixup_3src_null_dest()
5085 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
5086 if (inst
->is_3src() && inst
->dst
.is_null()) {
5087 inst
->dst
= fs_reg(GRF
, alloc
.allocate(dispatch_width
/ 8),
5094 fs_visitor::allocate_registers()
5096 bool allocated_without_spills
;
5098 static const enum instruction_scheduler_mode pre_modes
[] = {
5100 SCHEDULE_PRE_NON_LIFO
,
5104 /* Try each scheduling heuristic to see if it can successfully register
5105 * allocate without spilling. They should be ordered by decreasing
5106 * performance but increasing likelihood of allocating.
5108 for (unsigned i
= 0; i
< ARRAY_SIZE(pre_modes
); i
++) {
5109 schedule_instructions(pre_modes
[i
]);
5112 assign_regs_trivial();
5113 allocated_without_spills
= true;
5115 allocated_without_spills
= assign_regs(false);
5117 if (allocated_without_spills
)
5121 if (!allocated_without_spills
) {
5122 /* We assume that any spilling is worse than just dropping back to
5123 * SIMD8. There's probably actually some intermediate point where
5124 * SIMD16 with a couple of spills is still better.
5126 if (dispatch_width
== 16) {
5127 fail("Failure to register allocate. Reduce number of "
5128 "live scalar values to avoid this.");
5130 compiler
->shader_perf_log(log_data
,
5131 "%s shader triggered register spilling. "
5132 "Try reducing the number of live scalar "
5133 "values to improve performance.\n",
5137 /* Since we're out of heuristics, just go spill registers until we
5138 * get an allocation.
5140 while (!assign_regs(true)) {
5146 /* This must come after all optimization and register allocation, since
5147 * it inserts dead code that happens to have side effects, and it does
5148 * so based on the actual physical registers in use.
5150 insert_gen4_send_dependency_workarounds();
5155 schedule_instructions(SCHEDULE_POST
);
5157 if (last_scratch
> 0)
5158 prog_data
->total_scratch
= brw_get_scratch_size(last_scratch
);
5162 fs_visitor::run_vs(gl_clip_plane
*clip_planes
)
5164 assert(stage
== MESA_SHADER_VERTEX
);
5168 if (shader_time_index
>= 0)
5169 emit_shader_time_begin();
5176 compute_clip_distance(clip_planes
);
5180 if (shader_time_index
>= 0)
5181 emit_shader_time_end();
5187 assign_curb_setup();
5188 assign_vs_urb_setup();
5190 fixup_3src_null_dest();
5191 allocate_registers();
5197 fs_visitor::run_gs()
5199 assert(stage
== MESA_SHADER_GEOMETRY
);
5203 this->final_gs_vertex_count
= vgrf(glsl_type::uint_type
);
5205 if (gs_compile
->control_data_header_size_bits
> 0) {
5206 /* Create a VGRF to store accumulated control data bits. */
5207 this->control_data_bits
= vgrf(glsl_type::uint_type
);
5209 /* If we're outputting more than 32 control data bits, then EmitVertex()
5210 * will set control_data_bits to 0 after emitting the first vertex.
5211 * Otherwise, we need to initialize it to 0 here.
5213 if (gs_compile
->control_data_header_size_bits
<= 32) {
5214 const fs_builder abld
= bld
.annotate("initialize control data bits");
5215 abld
.MOV(this->control_data_bits
, fs_reg(0u));
5219 if (shader_time_index
>= 0)
5220 emit_shader_time_begin();
5224 emit_gs_thread_end();
5226 if (shader_time_index
>= 0)
5227 emit_shader_time_end();
5236 assign_curb_setup();
5237 assign_gs_urb_setup();
5239 fixup_3src_null_dest();
5240 allocate_registers();
5246 fs_visitor::run_fs(bool do_rep_send
)
5248 brw_wm_prog_data
*wm_prog_data
= (brw_wm_prog_data
*) this->prog_data
;
5249 brw_wm_prog_key
*wm_key
= (brw_wm_prog_key
*) this->key
;
5251 assert(stage
== MESA_SHADER_FRAGMENT
);
5253 if (devinfo
->gen
>= 6)
5254 setup_payload_gen6();
5256 setup_payload_gen4();
5260 } else if (do_rep_send
) {
5261 assert(dispatch_width
== 16);
5262 emit_repclear_shader();
5264 if (shader_time_index
>= 0)
5265 emit_shader_time_begin();
5267 calculate_urb_setup();
5268 if (nir
->info
.inputs_read
> 0) {
5269 if (devinfo
->gen
< 6)
5270 emit_interpolation_setup_gen4();
5272 emit_interpolation_setup_gen6();
5275 /* We handle discards by keeping track of the still-live pixels in f0.1.
5276 * Initialize it with the dispatched pixels.
5278 if (wm_prog_data
->uses_kill
) {
5279 fs_inst
*discard_init
= bld
.emit(FS_OPCODE_MOV_DISPATCH_TO_FLAGS
);
5280 discard_init
->flag_subreg
= 1;
5283 /* Generate FS IR for main(). (the visitor only descends into
5284 * functions called "main").
5291 if (wm_prog_data
->uses_kill
)
5292 bld
.emit(FS_OPCODE_PLACEHOLDER_HALT
);
5294 if (wm_key
->alpha_test_func
)
5299 if (shader_time_index
>= 0)
5300 emit_shader_time_end();
5306 assign_curb_setup();
5309 fixup_3src_null_dest();
5310 allocate_registers();
5316 if (dispatch_width
== 8)
5317 wm_prog_data
->reg_blocks
= brw_register_blocks(grf_used
);
5319 wm_prog_data
->reg_blocks_16
= brw_register_blocks(grf_used
);
5325 fs_visitor::run_cs()
5327 assert(stage
== MESA_SHADER_COMPUTE
);
5331 if (shader_time_index
>= 0)
5332 emit_shader_time_begin();
5339 emit_cs_terminate();
5341 if (shader_time_index
>= 0)
5342 emit_shader_time_end();
5348 assign_curb_setup();
5350 fixup_3src_null_dest();
5351 allocate_registers();
5360 * Return a bitfield where bit n is set if barycentric interpolation mode n
5361 * (see enum brw_wm_barycentric_interp_mode) is needed by the fragment shader.
5364 brw_compute_barycentric_interp_modes(const struct brw_device_info
*devinfo
,
5365 bool shade_model_flat
,
5366 bool persample_shading
,
5367 const nir_shader
*shader
)
5369 unsigned barycentric_interp_modes
= 0;
5371 nir_foreach_variable(var
, &shader
->inputs
) {
5372 enum glsl_interp_qualifier interp_qualifier
=
5373 (enum glsl_interp_qualifier
)var
->data
.interpolation
;
5374 bool is_centroid
= var
->data
.centroid
&& !persample_shading
;
5375 bool is_sample
= var
->data
.sample
|| persample_shading
;
5376 bool is_gl_Color
= (var
->data
.location
== VARYING_SLOT_COL0
) ||
5377 (var
->data
.location
== VARYING_SLOT_COL1
);
5379 /* Ignore WPOS and FACE, because they don't require interpolation. */
5380 if (var
->data
.location
== VARYING_SLOT_POS
||
5381 var
->data
.location
== VARYING_SLOT_FACE
)
5384 /* Determine the set (or sets) of barycentric coordinates needed to
5385 * interpolate this variable. Note that when
5386 * brw->needs_unlit_centroid_workaround is set, centroid interpolation
5387 * uses PIXEL interpolation for unlit pixels and CENTROID interpolation
5388 * for lit pixels, so we need both sets of barycentric coordinates.
5390 if (interp_qualifier
== INTERP_QUALIFIER_NOPERSPECTIVE
) {
5392 barycentric_interp_modes
|=
5393 1 << BRW_WM_NONPERSPECTIVE_CENTROID_BARYCENTRIC
;
5394 } else if (is_sample
) {
5395 barycentric_interp_modes
|=
5396 1 << BRW_WM_NONPERSPECTIVE_SAMPLE_BARYCENTRIC
;
5398 if ((!is_centroid
&& !is_sample
) ||
5399 devinfo
->needs_unlit_centroid_workaround
) {
5400 barycentric_interp_modes
|=
5401 1 << BRW_WM_NONPERSPECTIVE_PIXEL_BARYCENTRIC
;
5403 } else if (interp_qualifier
== INTERP_QUALIFIER_SMOOTH
||
5404 (!(shade_model_flat
&& is_gl_Color
) &&
5405 interp_qualifier
== INTERP_QUALIFIER_NONE
)) {
5407 barycentric_interp_modes
|=
5408 1 << BRW_WM_PERSPECTIVE_CENTROID_BARYCENTRIC
;
5409 } else if (is_sample
) {
5410 barycentric_interp_modes
|=
5411 1 << BRW_WM_PERSPECTIVE_SAMPLE_BARYCENTRIC
;
5413 if ((!is_centroid
&& !is_sample
) ||
5414 devinfo
->needs_unlit_centroid_workaround
) {
5415 barycentric_interp_modes
|=
5416 1 << BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
;
5421 return barycentric_interp_modes
;
5425 computed_depth_mode(const nir_shader
*shader
)
5427 if (shader
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_DEPTH
)) {
5428 switch (shader
->info
.fs
.depth_layout
) {
5429 case FRAG_DEPTH_LAYOUT_NONE
:
5430 case FRAG_DEPTH_LAYOUT_ANY
:
5431 return BRW_PSCDEPTH_ON
;
5432 case FRAG_DEPTH_LAYOUT_GREATER
:
5433 return BRW_PSCDEPTH_ON_GE
;
5434 case FRAG_DEPTH_LAYOUT_LESS
:
5435 return BRW_PSCDEPTH_ON_LE
;
5436 case FRAG_DEPTH_LAYOUT_UNCHANGED
:
5437 return BRW_PSCDEPTH_OFF
;
5440 return BRW_PSCDEPTH_OFF
;
5444 brw_compile_fs(const struct brw_compiler
*compiler
, void *log_data
,
5446 const struct brw_wm_prog_key
*key
,
5447 struct brw_wm_prog_data
*prog_data
,
5448 const nir_shader
*shader
,
5449 struct gl_program
*prog
,
5450 int shader_time_index8
, int shader_time_index16
,
5452 unsigned *final_assembly_size
,
5455 /* key->alpha_test_func means simulating alpha testing via discards,
5456 * so the shader definitely kills pixels.
5458 prog_data
->uses_kill
= shader
->info
.fs
.uses_discard
|| key
->alpha_test_func
;
5459 prog_data
->uses_omask
=
5460 shader
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_SAMPLE_MASK
);
5461 prog_data
->computed_depth_mode
= computed_depth_mode(shader
);
5462 prog_data
->computed_stencil
=
5463 shader
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_STENCIL
);
5465 prog_data
->early_fragment_tests
= shader
->info
.fs
.early_fragment_tests
;
5467 prog_data
->barycentric_interp_modes
=
5468 brw_compute_barycentric_interp_modes(compiler
->devinfo
,
5470 key
->persample_shading
,
5473 fs_visitor
v(compiler
, log_data
, mem_ctx
, key
,
5474 &prog_data
->base
, prog
, shader
, 8,
5475 shader_time_index8
);
5476 if (!v
.run_fs(false /* do_rep_send */)) {
5478 *error_str
= ralloc_strdup(mem_ctx
, v
.fail_msg
);
5483 cfg_t
*simd16_cfg
= NULL
;
5484 fs_visitor
v2(compiler
, log_data
, mem_ctx
, key
,
5485 &prog_data
->base
, prog
, shader
, 16,
5486 shader_time_index16
);
5487 if (likely(!(INTEL_DEBUG
& DEBUG_NO16
) || use_rep_send
)) {
5488 if (!v
.simd16_unsupported
) {
5489 /* Try a SIMD16 compile */
5490 v2
.import_uniforms(&v
);
5491 if (!v2
.run_fs(use_rep_send
)) {
5492 compiler
->shader_perf_log(log_data
,
5493 "SIMD16 shader failed to compile: %s",
5496 simd16_cfg
= v2
.cfg
;
5502 int no_simd8
= (INTEL_DEBUG
& DEBUG_NO8
) || use_rep_send
;
5503 if ((no_simd8
|| compiler
->devinfo
->gen
< 5) && simd16_cfg
) {
5505 prog_data
->no_8
= true;
5508 prog_data
->no_8
= false;
5511 fs_generator
g(compiler
, log_data
, mem_ctx
, (void *) key
, &prog_data
->base
,
5512 v
.promoted_constants
, v
.runtime_check_aads_emit
, "FS");
5514 if (unlikely(INTEL_DEBUG
& DEBUG_WM
)) {
5515 g
.enable_debug(ralloc_asprintf(mem_ctx
, "%s fragment shader %s",
5516 shader
->info
.label
? shader
->info
.label
:
5518 shader
->info
.name
));
5522 g
.generate_code(simd8_cfg
, 8);
5524 prog_data
->prog_offset_16
= g
.generate_code(simd16_cfg
, 16);
5526 return g
.get_assembly(final_assembly_size
);
5530 brw_cs_fill_local_id_payload(const struct brw_cs_prog_data
*prog_data
,
5531 void *buffer
, uint32_t threads
, uint32_t stride
)
5533 if (prog_data
->local_invocation_id_regs
== 0)
5536 /* 'stride' should be an integer number of registers, that is, a multiple
5539 assert(stride
% 32 == 0);
5541 unsigned x
= 0, y
= 0, z
= 0;
5542 for (unsigned t
= 0; t
< threads
; t
++) {
5543 uint32_t *param
= (uint32_t *) buffer
+ stride
* t
/ 4;
5545 for (unsigned i
= 0; i
< prog_data
->simd_size
; i
++) {
5546 param
[0 * prog_data
->simd_size
+ i
] = x
;
5547 param
[1 * prog_data
->simd_size
+ i
] = y
;
5548 param
[2 * prog_data
->simd_size
+ i
] = z
;
5551 if (x
== prog_data
->local_size
[0]) {
5554 if (y
== prog_data
->local_size
[1]) {
5557 if (z
== prog_data
->local_size
[2])
5566 fs_visitor::emit_cs_local_invocation_id_setup()
5568 assert(stage
== MESA_SHADER_COMPUTE
);
5570 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::uvec3_type
));
5572 struct brw_reg src
=
5573 brw_vec8_grf(payload
.local_invocation_id_reg
, 0);
5574 src
= retype(src
, BRW_REGISTER_TYPE_UD
);
5576 src
.nr
+= dispatch_width
/ 8;
5577 bld
.MOV(offset(*reg
, bld
, 1), src
);
5578 src
.nr
+= dispatch_width
/ 8;
5579 bld
.MOV(offset(*reg
, bld
, 2), src
);
5585 fs_visitor::emit_cs_work_group_id_setup()
5587 assert(stage
== MESA_SHADER_COMPUTE
);
5589 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::uvec3_type
));
5591 struct brw_reg
r0_1(retype(brw_vec1_grf(0, 1), BRW_REGISTER_TYPE_UD
));
5592 struct brw_reg
r0_6(retype(brw_vec1_grf(0, 6), BRW_REGISTER_TYPE_UD
));
5593 struct brw_reg
r0_7(retype(brw_vec1_grf(0, 7), BRW_REGISTER_TYPE_UD
));
5595 bld
.MOV(*reg
, r0_1
);
5596 bld
.MOV(offset(*reg
, bld
, 1), r0_6
);
5597 bld
.MOV(offset(*reg
, bld
, 2), r0_7
);
5603 brw_compile_cs(const struct brw_compiler
*compiler
, void *log_data
,
5605 const struct brw_cs_prog_key
*key
,
5606 struct brw_cs_prog_data
*prog_data
,
5607 const nir_shader
*shader
,
5608 int shader_time_index
,
5609 unsigned *final_assembly_size
,
5612 prog_data
->local_size
[0] = shader
->info
.cs
.local_size
[0];
5613 prog_data
->local_size
[1] = shader
->info
.cs
.local_size
[1];
5614 prog_data
->local_size
[2] = shader
->info
.cs
.local_size
[2];
5615 unsigned local_workgroup_size
=
5616 shader
->info
.cs
.local_size
[0] * shader
->info
.cs
.local_size
[1] *
5617 shader
->info
.cs
.local_size
[2];
5619 unsigned max_cs_threads
= compiler
->devinfo
->max_cs_threads
;
5622 const char *fail_msg
= NULL
;
5624 /* Now the main event: Visit the shader IR and generate our CS IR for it.
5626 fs_visitor
v8(compiler
, log_data
, mem_ctx
, key
, &prog_data
->base
,
5627 NULL
, /* Never used in core profile */
5628 shader
, 8, shader_time_index
);
5630 fail_msg
= v8
.fail_msg
;
5631 } else if (local_workgroup_size
<= 8 * max_cs_threads
) {
5633 prog_data
->simd_size
= 8;
5636 fs_visitor
v16(compiler
, log_data
, mem_ctx
, key
, &prog_data
->base
,
5637 NULL
, /* Never used in core profile */
5638 shader
, 16, shader_time_index
);
5639 if (likely(!(INTEL_DEBUG
& DEBUG_NO16
)) &&
5640 !fail_msg
&& !v8
.simd16_unsupported
&&
5641 local_workgroup_size
<= 16 * max_cs_threads
) {
5642 /* Try a SIMD16 compile */
5643 v16
.import_uniforms(&v8
);
5644 if (!v16
.run_cs()) {
5645 compiler
->shader_perf_log(log_data
,
5646 "SIMD16 shader failed to compile: %s",
5650 "Couldn't generate SIMD16 program and not "
5651 "enough threads for SIMD8";
5655 prog_data
->simd_size
= 16;
5659 if (unlikely(cfg
== NULL
)) {
5662 *error_str
= ralloc_strdup(mem_ctx
, fail_msg
);
5667 fs_generator
g(compiler
, log_data
, mem_ctx
, (void*) key
, &prog_data
->base
,
5668 v8
.promoted_constants
, v8
.runtime_check_aads_emit
, "CS");
5669 if (INTEL_DEBUG
& DEBUG_CS
) {
5670 char *name
= ralloc_asprintf(mem_ctx
, "%s compute shader %s",
5671 shader
->info
.label
? shader
->info
.label
:
5674 g
.enable_debug(name
);
5677 g
.generate_code(cfg
, prog_data
->simd_size
);
5679 return g
.get_assembly(final_assembly_size
);