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 "main/macros.h"
35 #include "brw_vec4_gs_visitor.h"
37 #include "brw_dead_control_flow.h"
38 #include "common/gen_debug.h"
39 #include "compiler/glsl_types.h"
40 #include "compiler/nir/nir_builder.h"
41 #include "program/prog_parameter.h"
45 static unsigned get_lowered_simd_width(const struct gen_device_info
*devinfo
,
49 fs_inst::init(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
50 const fs_reg
*src
, unsigned sources
)
52 memset(this, 0, sizeof(*this));
54 this->src
= new fs_reg
[MAX2(sources
, 3)];
55 for (unsigned i
= 0; i
< sources
; i
++)
56 this->src
[i
] = src
[i
];
58 this->opcode
= opcode
;
60 this->sources
= sources
;
61 this->exec_size
= exec_size
;
64 assert(dst
.file
!= IMM
&& dst
.file
!= UNIFORM
);
66 assert(this->exec_size
!= 0);
68 this->conditional_mod
= BRW_CONDITIONAL_NONE
;
70 /* This will be the case for almost all instructions. */
77 this->size_written
= dst
.component_size(exec_size
);
80 this->size_written
= 0;
84 unreachable("Invalid destination register file");
87 this->writes_accumulator
= false;
92 init(BRW_OPCODE_NOP
, 8, dst
, NULL
, 0);
95 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
)
97 init(opcode
, exec_size
, reg_undef
, NULL
, 0);
100 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
)
102 init(opcode
, exec_size
, dst
, NULL
, 0);
105 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
108 const fs_reg src
[1] = { src0
};
109 init(opcode
, exec_size
, dst
, src
, 1);
112 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
113 const fs_reg
&src0
, const fs_reg
&src1
)
115 const fs_reg src
[2] = { src0
, src1
};
116 init(opcode
, exec_size
, dst
, src
, 2);
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
, const fs_reg
&src2
)
122 const fs_reg src
[3] = { src0
, src1
, src2
};
123 init(opcode
, exec_size
, dst
, src
, 3);
126 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_width
, const fs_reg
&dst
,
127 const fs_reg src
[], unsigned sources
)
129 init(opcode
, exec_width
, dst
, src
, sources
);
132 fs_inst::fs_inst(const fs_inst
&that
)
134 memcpy(this, &that
, sizeof(that
));
136 this->src
= new fs_reg
[MAX2(that
.sources
, 3)];
138 for (unsigned i
= 0; i
< that
.sources
; i
++)
139 this->src
[i
] = that
.src
[i
];
148 fs_inst::resize_sources(uint8_t num_sources
)
150 if (this->sources
!= num_sources
) {
151 fs_reg
*src
= new fs_reg
[MAX2(num_sources
, 3)];
153 for (unsigned i
= 0; i
< MIN2(this->sources
, num_sources
); ++i
)
154 src
[i
] = this->src
[i
];
158 this->sources
= num_sources
;
163 fs_visitor::VARYING_PULL_CONSTANT_LOAD(const fs_builder
&bld
,
165 const fs_reg
&surf_index
,
166 const fs_reg
&varying_offset
,
167 uint32_t const_offset
)
169 /* We have our constant surface use a pitch of 4 bytes, so our index can
170 * be any component of a vector, and then we load 4 contiguous
171 * components starting from that.
173 * We break down the const_offset to a portion added to the variable offset
174 * and a portion done using fs_reg::offset, which means that if you have
175 * GLSL using something like "uniform vec4 a[20]; gl_FragColor = a[i]",
176 * we'll temporarily generate 4 vec4 loads from offset i * 4, and CSE can
177 * later notice that those loads are all the same and eliminate the
180 fs_reg vec4_offset
= vgrf(glsl_type::uint_type
);
181 bld
.ADD(vec4_offset
, varying_offset
, brw_imm_ud(const_offset
& ~0xf));
183 /* The pull load message will load a vec4 (16 bytes). If we are loading
184 * a double this means we are only loading 2 elements worth of data.
185 * We also want to use a 32-bit data type for the dst of the load operation
186 * so other parts of the driver don't get confused about the size of the
189 fs_reg vec4_result
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 4);
190 fs_inst
*inst
= bld
.emit(FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_LOGICAL
,
191 vec4_result
, surf_index
, vec4_offset
);
192 inst
->size_written
= 4 * vec4_result
.component_size(inst
->exec_size
);
194 if (type_sz(dst
.type
) == 8) {
195 shuffle_32bit_load_result_to_64bit_data(
196 bld
, retype(vec4_result
, dst
.type
), vec4_result
, 2);
199 vec4_result
.type
= dst
.type
;
200 bld
.MOV(dst
, offset(vec4_result
, bld
,
201 (const_offset
& 0xf) / type_sz(vec4_result
.type
)));
205 * A helper for MOV generation for fixing up broken hardware SEND dependency
209 fs_visitor::DEP_RESOLVE_MOV(const fs_builder
&bld
, int grf
)
211 /* The caller always wants uncompressed to emit the minimal extra
212 * dependencies, and to avoid having to deal with aligning its regs to 2.
214 const fs_builder ubld
= bld
.annotate("send dependency resolve")
217 ubld
.MOV(ubld
.null_reg_f(), fs_reg(VGRF
, grf
, BRW_REGISTER_TYPE_F
));
221 fs_inst::equals(fs_inst
*inst
) const
223 return (opcode
== inst
->opcode
&&
224 dst
.equals(inst
->dst
) &&
225 src
[0].equals(inst
->src
[0]) &&
226 src
[1].equals(inst
->src
[1]) &&
227 src
[2].equals(inst
->src
[2]) &&
228 saturate
== inst
->saturate
&&
229 predicate
== inst
->predicate
&&
230 conditional_mod
== inst
->conditional_mod
&&
231 mlen
== inst
->mlen
&&
232 base_mrf
== inst
->base_mrf
&&
233 target
== inst
->target
&&
235 header_size
== inst
->header_size
&&
236 shadow_compare
== inst
->shadow_compare
&&
237 exec_size
== inst
->exec_size
&&
238 offset
== inst
->offset
);
242 fs_inst::is_send_from_grf() const
245 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7
:
246 case SHADER_OPCODE_SHADER_TIME_ADD
:
247 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
248 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
249 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
250 case SHADER_OPCODE_UNTYPED_ATOMIC
:
251 case SHADER_OPCODE_UNTYPED_SURFACE_READ
:
252 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE
:
253 case SHADER_OPCODE_TYPED_ATOMIC
:
254 case SHADER_OPCODE_TYPED_SURFACE_READ
:
255 case SHADER_OPCODE_TYPED_SURFACE_WRITE
:
256 case SHADER_OPCODE_URB_WRITE_SIMD8
:
257 case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
258 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
:
259 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
260 case SHADER_OPCODE_URB_READ_SIMD8
:
261 case SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
:
263 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
264 return src
[1].file
== VGRF
;
265 case FS_OPCODE_FB_WRITE
:
266 case FS_OPCODE_FB_READ
:
267 return src
[0].file
== VGRF
;
270 return src
[0].file
== VGRF
;
277 * Returns true if this instruction's sources and destinations cannot
278 * safely be the same register.
280 * In most cases, a register can be written over safely by the same
281 * instruction that is its last use. For a single instruction, the
282 * sources are dereferenced before writing of the destination starts
285 * However, there are a few cases where this can be problematic:
287 * - Virtual opcodes that translate to multiple instructions in the
288 * code generator: if src == dst and one instruction writes the
289 * destination before a later instruction reads the source, then
290 * src will have been clobbered.
292 * - SIMD16 compressed instructions with certain regioning (see below).
294 * The register allocator uses this information to set up conflicts between
295 * GRF sources and the destination.
298 fs_inst::has_source_and_destination_hazard() const
301 case FS_OPCODE_PACK_HALF_2x16_SPLIT
:
302 /* Multiple partial writes to the destination */
305 /* The SIMD16 compressed instruction
307 * add(16) g4<1>F g4<8,8,1>F g6<8,8,1>F
309 * is actually decoded in hardware as:
311 * add(8) g4<1>F g4<8,8,1>F g6<8,8,1>F
312 * add(8) g5<1>F g5<8,8,1>F g7<8,8,1>F
314 * Which is safe. However, if we have uniform accesses
315 * happening, we get into trouble:
317 * add(8) g4<1>F g4<0,1,0>F g6<8,8,1>F
318 * add(8) g5<1>F g4<0,1,0>F g7<8,8,1>F
320 * Now our destination for the first instruction overwrote the
321 * second instruction's src0, and we get garbage for those 8
322 * pixels. There's a similar issue for the pre-gen6
323 * pixel_x/pixel_y, which are registers of 16-bit values and thus
324 * would get stomped by the first decode as well.
326 if (exec_size
== 16) {
327 for (int i
= 0; i
< sources
; i
++) {
328 if (src
[i
].file
== VGRF
&& (src
[i
].stride
== 0 ||
329 src
[i
].type
== BRW_REGISTER_TYPE_UW
||
330 src
[i
].type
== BRW_REGISTER_TYPE_W
||
331 src
[i
].type
== BRW_REGISTER_TYPE_UB
||
332 src
[i
].type
== BRW_REGISTER_TYPE_B
)) {
342 fs_inst::is_copy_payload(const brw::simple_allocator
&grf_alloc
) const
344 if (this->opcode
!= SHADER_OPCODE_LOAD_PAYLOAD
)
347 fs_reg reg
= this->src
[0];
348 if (reg
.file
!= VGRF
|| reg
.offset
!= 0 || reg
.stride
!= 1)
351 if (grf_alloc
.sizes
[reg
.nr
] * REG_SIZE
!= this->size_written
)
354 for (int i
= 0; i
< this->sources
; i
++) {
355 reg
.type
= this->src
[i
].type
;
356 if (!this->src
[i
].equals(reg
))
359 if (i
< this->header_size
) {
360 reg
.offset
+= REG_SIZE
;
362 reg
= horiz_offset(reg
, this->exec_size
);
370 fs_inst::can_do_source_mods(const struct gen_device_info
*devinfo
)
372 if (devinfo
->gen
== 6 && is_math())
375 if (is_send_from_grf())
378 if (!backend_instruction::can_do_source_mods())
385 fs_inst::can_change_types() const
387 return dst
.type
== src
[0].type
&&
388 !src
[0].abs
&& !src
[0].negate
&& !saturate
&&
389 (opcode
== BRW_OPCODE_MOV
||
390 (opcode
== BRW_OPCODE_SEL
&&
391 dst
.type
== src
[1].type
&&
392 predicate
!= BRW_PREDICATE_NONE
&&
393 !src
[1].abs
&& !src
[1].negate
));
397 fs_inst::has_side_effects() const
399 return this->eot
|| backend_instruction::has_side_effects();
405 memset(this, 0, sizeof(*this));
409 /** Generic unset register constructor. */
413 this->file
= BAD_FILE
;
416 fs_reg::fs_reg(struct ::brw_reg reg
) :
421 if (this->file
== IMM
&&
422 (this->type
!= BRW_REGISTER_TYPE_V
&&
423 this->type
!= BRW_REGISTER_TYPE_UV
&&
424 this->type
!= BRW_REGISTER_TYPE_VF
)) {
430 fs_reg::equals(const fs_reg
&r
) const
432 return (this->backend_reg::equals(r
) &&
437 fs_reg::is_contiguous() const
443 fs_reg::component_size(unsigned width
) const
445 const unsigned stride
= ((file
!= ARF
&& file
!= FIXED_GRF
) ? this->stride
:
448 return MAX2(width
* stride
, 1) * type_sz(type
);
452 type_size_scalar(const struct glsl_type
*type
)
454 unsigned int size
, i
;
456 switch (type
->base_type
) {
459 case GLSL_TYPE_FLOAT
:
461 return type
->components();
462 case GLSL_TYPE_DOUBLE
:
463 case GLSL_TYPE_UINT64
:
464 case GLSL_TYPE_INT64
:
465 return type
->components() * 2;
466 case GLSL_TYPE_ARRAY
:
467 return type_size_scalar(type
->fields
.array
) * type
->length
;
468 case GLSL_TYPE_STRUCT
:
470 for (i
= 0; i
< type
->length
; i
++) {
471 size
+= type_size_scalar(type
->fields
.structure
[i
].type
);
474 case GLSL_TYPE_SAMPLER
:
475 /* Samplers take up no register space, since they're baked in at
479 case GLSL_TYPE_ATOMIC_UINT
:
481 case GLSL_TYPE_SUBROUTINE
:
483 case GLSL_TYPE_IMAGE
:
484 return BRW_IMAGE_PARAM_SIZE
;
486 case GLSL_TYPE_ERROR
:
487 case GLSL_TYPE_INTERFACE
:
488 case GLSL_TYPE_FUNCTION
:
489 unreachable("not reached");
496 * Create a MOV to read the timestamp register.
498 * The caller is responsible for emitting the MOV. The return value is
499 * the destination of the MOV, with extra parameters set.
502 fs_visitor::get_timestamp(const fs_builder
&bld
)
504 assert(devinfo
->gen
>= 7);
506 fs_reg ts
= fs_reg(retype(brw_vec4_reg(BRW_ARCHITECTURE_REGISTER_FILE
,
509 BRW_REGISTER_TYPE_UD
));
511 fs_reg dst
= fs_reg(VGRF
, alloc
.allocate(1), BRW_REGISTER_TYPE_UD
);
513 /* We want to read the 3 fields we care about even if it's not enabled in
516 bld
.group(4, 0).exec_all().MOV(dst
, ts
);
522 fs_visitor::emit_shader_time_begin()
524 /* We want only the low 32 bits of the timestamp. Since it's running
525 * at the GPU clock rate of ~1.2ghz, it will roll over every ~3 seconds,
526 * which is plenty of time for our purposes. It is identical across the
527 * EUs, but since it's tracking GPU core speed it will increment at a
528 * varying rate as render P-states change.
530 shader_start_time
= component(
531 get_timestamp(bld
.annotate("shader time start")), 0);
535 fs_visitor::emit_shader_time_end()
537 /* Insert our code just before the final SEND with EOT. */
538 exec_node
*end
= this->instructions
.get_tail();
539 assert(end
&& ((fs_inst
*) end
)->eot
);
540 const fs_builder ibld
= bld
.annotate("shader time end")
541 .exec_all().at(NULL
, end
);
542 const fs_reg timestamp
= get_timestamp(ibld
);
544 /* We only use the low 32 bits of the timestamp - see
545 * emit_shader_time_begin()).
547 * We could also check if render P-states have changed (or anything
548 * else that might disrupt timing) by setting smear to 2 and checking if
549 * that field is != 0.
551 const fs_reg shader_end_time
= component(timestamp
, 0);
553 /* Check that there weren't any timestamp reset events (assuming these
554 * were the only two timestamp reads that happened).
556 const fs_reg reset
= component(timestamp
, 2);
557 set_condmod(BRW_CONDITIONAL_Z
,
558 ibld
.AND(ibld
.null_reg_ud(), reset
, brw_imm_ud(1u)));
559 ibld
.IF(BRW_PREDICATE_NORMAL
);
561 fs_reg start
= shader_start_time
;
563 const fs_reg diff
= component(fs_reg(VGRF
, alloc
.allocate(1),
564 BRW_REGISTER_TYPE_UD
),
566 const fs_builder cbld
= ibld
.group(1, 0);
567 cbld
.group(1, 0).ADD(diff
, start
, shader_end_time
);
569 /* If there were no instructions between the two timestamp gets, the diff
570 * is 2 cycles. Remove that overhead, so I can forget about that when
571 * trying to determine the time taken for single instructions.
573 cbld
.ADD(diff
, diff
, brw_imm_ud(-2u));
574 SHADER_TIME_ADD(cbld
, 0, diff
);
575 SHADER_TIME_ADD(cbld
, 1, brw_imm_ud(1u));
576 ibld
.emit(BRW_OPCODE_ELSE
);
577 SHADER_TIME_ADD(cbld
, 2, brw_imm_ud(1u));
578 ibld
.emit(BRW_OPCODE_ENDIF
);
582 fs_visitor::SHADER_TIME_ADD(const fs_builder
&bld
,
583 int shader_time_subindex
,
586 int index
= shader_time_index
* 3 + shader_time_subindex
;
587 struct brw_reg offset
= brw_imm_d(index
* BRW_SHADER_TIME_STRIDE
);
590 if (dispatch_width
== 8)
591 payload
= vgrf(glsl_type::uvec2_type
);
593 payload
= vgrf(glsl_type::uint_type
);
595 bld
.emit(SHADER_OPCODE_SHADER_TIME_ADD
, fs_reg(), payload
, offset
, value
);
599 fs_visitor::vfail(const char *format
, va_list va
)
608 msg
= ralloc_vasprintf(mem_ctx
, format
, va
);
609 msg
= ralloc_asprintf(mem_ctx
, "%s compile failed: %s\n", stage_abbrev
, msg
);
611 this->fail_msg
= msg
;
614 fprintf(stderr
, "%s", msg
);
619 fs_visitor::fail(const char *format
, ...)
623 va_start(va
, format
);
629 * Mark this program as impossible to compile with dispatch width greater
632 * During the SIMD8 compile (which happens first), we can detect and flag
633 * things that are unsupported in SIMD16+ mode, so the compiler can skip the
634 * SIMD16+ compile altogether.
636 * During a compile of dispatch width greater than n (if one happens anyway),
637 * this just calls fail().
640 fs_visitor::limit_dispatch_width(unsigned n
, const char *msg
)
642 if (dispatch_width
> n
) {
645 max_dispatch_width
= n
;
646 compiler
->shader_perf_log(log_data
,
647 "Shader dispatch width limited to SIMD%d: %s",
653 * Returns true if the instruction has a flag that means it won't
654 * update an entire destination register.
656 * For example, dead code elimination and live variable analysis want to know
657 * when a write to a variable screens off any preceding values that were in
661 fs_inst::is_partial_write() const
663 return ((this->predicate
&& this->opcode
!= BRW_OPCODE_SEL
) ||
664 (this->exec_size
* type_sz(this->dst
.type
)) < 32 ||
665 !this->dst
.is_contiguous() ||
666 this->dst
.offset
% REG_SIZE
!= 0);
670 fs_inst::components_read(unsigned i
) const
672 /* Return zero if the source is not present. */
673 if (src
[i
].file
== BAD_FILE
)
677 case FS_OPCODE_LINTERP
:
683 case FS_OPCODE_PIXEL_X
:
684 case FS_OPCODE_PIXEL_Y
:
688 case FS_OPCODE_FB_WRITE_LOGICAL
:
689 assert(src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].file
== IMM
);
690 /* First/second FB write color. */
692 return src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].ud
;
696 case SHADER_OPCODE_TEX_LOGICAL
:
697 case SHADER_OPCODE_TXD_LOGICAL
:
698 case SHADER_OPCODE_TXF_LOGICAL
:
699 case SHADER_OPCODE_TXL_LOGICAL
:
700 case SHADER_OPCODE_TXS_LOGICAL
:
701 case FS_OPCODE_TXB_LOGICAL
:
702 case SHADER_OPCODE_TXF_CMS_LOGICAL
:
703 case SHADER_OPCODE_TXF_CMS_W_LOGICAL
:
704 case SHADER_OPCODE_TXF_UMS_LOGICAL
:
705 case SHADER_OPCODE_TXF_MCS_LOGICAL
:
706 case SHADER_OPCODE_LOD_LOGICAL
:
707 case SHADER_OPCODE_TG4_LOGICAL
:
708 case SHADER_OPCODE_TG4_OFFSET_LOGICAL
:
709 case SHADER_OPCODE_SAMPLEINFO_LOGICAL
:
710 assert(src
[TEX_LOGICAL_SRC_COORD_COMPONENTS
].file
== IMM
&&
711 src
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
].file
== IMM
);
712 /* Texture coordinates. */
713 if (i
== TEX_LOGICAL_SRC_COORDINATE
)
714 return src
[TEX_LOGICAL_SRC_COORD_COMPONENTS
].ud
;
715 /* Texture derivatives. */
716 else if ((i
== TEX_LOGICAL_SRC_LOD
|| i
== TEX_LOGICAL_SRC_LOD2
) &&
717 opcode
== SHADER_OPCODE_TXD_LOGICAL
)
718 return src
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
].ud
;
719 /* Texture offset. */
720 else if (i
== TEX_LOGICAL_SRC_TG4_OFFSET
)
723 else if (i
== TEX_LOGICAL_SRC_MCS
&& opcode
== SHADER_OPCODE_TXF_CMS_W_LOGICAL
)
728 case SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
:
729 case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
:
730 assert(src
[3].file
== IMM
);
731 /* Surface coordinates. */
734 /* Surface operation source (ignored for reads). */
740 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
:
741 case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
:
742 assert(src
[3].file
== IMM
&&
744 /* Surface coordinates. */
747 /* Surface operation source. */
753 case SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL
:
754 case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
: {
755 assert(src
[3].file
== IMM
&&
757 const unsigned op
= src
[4].ud
;
758 /* Surface coordinates. */
761 /* Surface operation source. */
762 else if (i
== 1 && op
== BRW_AOP_CMPWR
)
764 else if (i
== 1 && (op
== BRW_AOP_INC
|| op
== BRW_AOP_DEC
||
765 op
== BRW_AOP_PREDEC
))
777 fs_inst::size_read(int arg
) const
780 case FS_OPCODE_FB_WRITE
:
781 case FS_OPCODE_FB_READ
:
782 case SHADER_OPCODE_URB_WRITE_SIMD8
:
783 case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
784 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
:
785 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
786 case SHADER_OPCODE_URB_READ_SIMD8
:
787 case SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
:
788 case SHADER_OPCODE_UNTYPED_ATOMIC
:
789 case SHADER_OPCODE_UNTYPED_SURFACE_READ
:
790 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE
:
791 case SHADER_OPCODE_TYPED_ATOMIC
:
792 case SHADER_OPCODE_TYPED_SURFACE_READ
:
793 case SHADER_OPCODE_TYPED_SURFACE_WRITE
:
794 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
796 return mlen
* REG_SIZE
;
799 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7
:
800 /* The payload is actually stored in src1 */
802 return mlen
* REG_SIZE
;
805 case FS_OPCODE_LINTERP
:
810 case SHADER_OPCODE_LOAD_PAYLOAD
:
811 if (arg
< this->header_size
)
815 case CS_OPCODE_CS_TERMINATE
:
816 case SHADER_OPCODE_BARRIER
:
819 case SHADER_OPCODE_MOV_INDIRECT
:
821 assert(src
[2].file
== IMM
);
827 if (is_tex() && arg
== 0 && src
[0].file
== VGRF
)
828 return mlen
* REG_SIZE
;
832 switch (src
[arg
].file
) {
835 return components_read(arg
) * type_sz(src
[arg
].type
);
841 return components_read(arg
) * src
[arg
].component_size(exec_size
);
843 unreachable("MRF registers are not allowed as sources");
849 /* Return the subset of flag registers that an instruction could
850 * potentially read or write based on the execution controls and flag
851 * subregister number of the instruction.
854 flag_mask(const fs_inst
*inst
)
856 const unsigned start
= inst
->flag_subreg
* 16 + inst
->group
;
857 const unsigned end
= start
+ inst
->exec_size
;
858 return ((1 << DIV_ROUND_UP(end
, 8)) - 1) & ~((1 << (start
/ 8)) - 1);
863 fs_inst::flags_read(const gen_device_info
*devinfo
) const
865 /* XXX - This doesn't consider explicit uses of the flag register as source
868 if (predicate
== BRW_PREDICATE_ALIGN1_ANYV
||
869 predicate
== BRW_PREDICATE_ALIGN1_ALLV
) {
870 /* The vertical predication modes combine corresponding bits from
871 * f0.0 and f1.0 on Gen7+, and f0.0 and f0.1 on older hardware.
873 const unsigned shift
= devinfo
->gen
>= 7 ? 4 : 2;
874 return flag_mask(this) << shift
| flag_mask(this);
875 } else if (predicate
) {
876 return flag_mask(this);
883 fs_inst::flags_written() const
885 /* XXX - This doesn't consider explicit uses of the flag register as
886 * destination region.
888 if ((conditional_mod
&& (opcode
!= BRW_OPCODE_SEL
&&
889 opcode
!= BRW_OPCODE_IF
&&
890 opcode
!= BRW_OPCODE_WHILE
)) ||
891 opcode
== FS_OPCODE_MOV_DISPATCH_TO_FLAGS
) {
892 return flag_mask(this);
899 * Returns how many MRFs an FS opcode will write over.
901 * Note that this is not the 0 or 1 implied writes in an actual gen
902 * instruction -- the FS opcodes often generate MOVs in addition.
905 fs_visitor::implied_mrf_writes(fs_inst
*inst
)
910 if (inst
->base_mrf
== -1)
913 switch (inst
->opcode
) {
914 case SHADER_OPCODE_RCP
:
915 case SHADER_OPCODE_RSQ
:
916 case SHADER_OPCODE_SQRT
:
917 case SHADER_OPCODE_EXP2
:
918 case SHADER_OPCODE_LOG2
:
919 case SHADER_OPCODE_SIN
:
920 case SHADER_OPCODE_COS
:
921 return 1 * dispatch_width
/ 8;
922 case SHADER_OPCODE_POW
:
923 case SHADER_OPCODE_INT_QUOTIENT
:
924 case SHADER_OPCODE_INT_REMAINDER
:
925 return 2 * dispatch_width
/ 8;
926 case SHADER_OPCODE_TEX
:
928 case SHADER_OPCODE_TXD
:
929 case SHADER_OPCODE_TXF
:
930 case SHADER_OPCODE_TXF_CMS
:
931 case SHADER_OPCODE_TXF_MCS
:
932 case SHADER_OPCODE_TG4
:
933 case SHADER_OPCODE_TG4_OFFSET
:
934 case SHADER_OPCODE_TXL
:
935 case SHADER_OPCODE_TXS
:
936 case SHADER_OPCODE_LOD
:
937 case SHADER_OPCODE_SAMPLEINFO
:
939 case FS_OPCODE_FB_WRITE
:
941 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
942 case SHADER_OPCODE_GEN4_SCRATCH_READ
:
944 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN4
:
946 case SHADER_OPCODE_GEN4_SCRATCH_WRITE
:
949 unreachable("not reached");
954 fs_visitor::vgrf(const glsl_type
*const type
)
956 int reg_width
= dispatch_width
/ 8;
957 return fs_reg(VGRF
, alloc
.allocate(type_size_scalar(type
) * reg_width
),
958 brw_type_for_base_type(type
));
961 fs_reg::fs_reg(enum brw_reg_file file
, int nr
)
966 this->type
= BRW_REGISTER_TYPE_F
;
967 this->stride
= (file
== UNIFORM
? 0 : 1);
970 fs_reg::fs_reg(enum brw_reg_file file
, int nr
, enum brw_reg_type type
)
976 this->stride
= (file
== UNIFORM
? 0 : 1);
979 /* For SIMD16, we need to follow from the uniform setup of SIMD8 dispatch.
980 * This brings in those uniform definitions
983 fs_visitor::import_uniforms(fs_visitor
*v
)
985 this->push_constant_loc
= v
->push_constant_loc
;
986 this->pull_constant_loc
= v
->pull_constant_loc
;
987 this->uniforms
= v
->uniforms
;
991 fs_visitor::emit_fragcoord_interpolation(fs_reg wpos
)
993 assert(stage
== MESA_SHADER_FRAGMENT
);
996 bld
.MOV(wpos
, this->pixel_x
);
997 wpos
= offset(wpos
, bld
, 1);
1000 bld
.MOV(wpos
, this->pixel_y
);
1001 wpos
= offset(wpos
, bld
, 1);
1003 /* gl_FragCoord.z */
1004 if (devinfo
->gen
>= 6) {
1005 bld
.MOV(wpos
, fs_reg(brw_vec8_grf(payload
.source_depth_reg
, 0)));
1007 bld
.emit(FS_OPCODE_LINTERP
, wpos
,
1008 this->delta_xy
[BRW_BARYCENTRIC_PERSPECTIVE_PIXEL
],
1009 interp_reg(VARYING_SLOT_POS
, 2));
1011 wpos
= offset(wpos
, bld
, 1);
1013 /* gl_FragCoord.w: Already set up in emit_interpolation */
1014 bld
.MOV(wpos
, this->wpos_w
);
1017 enum brw_barycentric_mode
1018 brw_barycentric_mode(enum glsl_interp_mode mode
, nir_intrinsic_op op
)
1020 /* Barycentric modes don't make sense for flat inputs. */
1021 assert(mode
!= INTERP_MODE_FLAT
);
1025 case nir_intrinsic_load_barycentric_pixel
:
1026 case nir_intrinsic_load_barycentric_at_offset
:
1027 bary
= BRW_BARYCENTRIC_PERSPECTIVE_PIXEL
;
1029 case nir_intrinsic_load_barycentric_centroid
:
1030 bary
= BRW_BARYCENTRIC_PERSPECTIVE_CENTROID
;
1032 case nir_intrinsic_load_barycentric_sample
:
1033 case nir_intrinsic_load_barycentric_at_sample
:
1034 bary
= BRW_BARYCENTRIC_PERSPECTIVE_SAMPLE
;
1037 unreachable("invalid intrinsic");
1040 if (mode
== INTERP_MODE_NOPERSPECTIVE
)
1043 return (enum brw_barycentric_mode
) bary
;
1047 * Turn one of the two CENTROID barycentric modes into PIXEL mode.
1049 static enum brw_barycentric_mode
1050 centroid_to_pixel(enum brw_barycentric_mode bary
)
1052 assert(bary
== BRW_BARYCENTRIC_PERSPECTIVE_CENTROID
||
1053 bary
== BRW_BARYCENTRIC_NONPERSPECTIVE_CENTROID
);
1054 return (enum brw_barycentric_mode
) ((unsigned) bary
- 1);
1058 fs_visitor::emit_frontfacing_interpolation()
1060 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::bool_type
));
1062 if (devinfo
->gen
>= 6) {
1063 /* Bit 15 of g0.0 is 0 if the polygon is front facing. We want to create
1064 * a boolean result from this (~0/true or 0/false).
1066 * We can use the fact that bit 15 is the MSB of g0.0:W to accomplish
1067 * this task in only one instruction:
1068 * - a negation source modifier will flip the bit; and
1069 * - a W -> D type conversion will sign extend the bit into the high
1070 * word of the destination.
1072 * An ASR 15 fills the low word of the destination.
1074 fs_reg g0
= fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_W
));
1077 bld
.ASR(*reg
, g0
, brw_imm_d(15));
1079 /* Bit 31 of g1.6 is 0 if the polygon is front facing. We want to create
1080 * a boolean result from this (1/true or 0/false).
1082 * Like in the above case, since the bit is the MSB of g1.6:UD we can use
1083 * the negation source modifier to flip it. Unfortunately the SHR
1084 * instruction only operates on UD (or D with an abs source modifier)
1085 * sources without negation.
1087 * Instead, use ASR (which will give ~0/true or 0/false).
1089 fs_reg g1_6
= fs_reg(retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_D
));
1092 bld
.ASR(*reg
, g1_6
, brw_imm_d(31));
1099 fs_visitor::compute_sample_position(fs_reg dst
, fs_reg int_sample_pos
)
1101 assert(stage
== MESA_SHADER_FRAGMENT
);
1102 struct brw_wm_prog_data
*wm_prog_data
= brw_wm_prog_data(this->prog_data
);
1103 assert(dst
.type
== BRW_REGISTER_TYPE_F
);
1105 if (wm_prog_data
->persample_dispatch
) {
1106 /* Convert int_sample_pos to floating point */
1107 bld
.MOV(dst
, int_sample_pos
);
1108 /* Scale to the range [0, 1] */
1109 bld
.MUL(dst
, dst
, brw_imm_f(1 / 16.0f
));
1112 /* From ARB_sample_shading specification:
1113 * "When rendering to a non-multisample buffer, or if multisample
1114 * rasterization is disabled, gl_SamplePosition will always be
1117 bld
.MOV(dst
, brw_imm_f(0.5f
));
1122 fs_visitor::emit_samplepos_setup()
1124 assert(devinfo
->gen
>= 6);
1126 const fs_builder abld
= bld
.annotate("compute sample position");
1127 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::vec2_type
));
1129 fs_reg int_sample_x
= vgrf(glsl_type::int_type
);
1130 fs_reg int_sample_y
= vgrf(glsl_type::int_type
);
1132 /* WM will be run in MSDISPMODE_PERSAMPLE. So, only one of SIMD8 or SIMD16
1133 * mode will be enabled.
1135 * From the Ivy Bridge PRM, volume 2 part 1, page 344:
1136 * R31.1:0 Position Offset X/Y for Slot[3:0]
1137 * R31.3:2 Position Offset X/Y for Slot[7:4]
1140 * The X, Y sample positions come in as bytes in thread payload. So, read
1141 * the positions using vstride=16, width=8, hstride=2.
1143 struct brw_reg sample_pos_reg
=
1144 stride(retype(brw_vec1_grf(payload
.sample_pos_reg
, 0),
1145 BRW_REGISTER_TYPE_B
), 16, 8, 2);
1147 if (dispatch_width
== 8) {
1148 abld
.MOV(int_sample_x
, fs_reg(sample_pos_reg
));
1150 abld
.half(0).MOV(half(int_sample_x
, 0), fs_reg(sample_pos_reg
));
1151 abld
.half(1).MOV(half(int_sample_x
, 1),
1152 fs_reg(suboffset(sample_pos_reg
, 16)));
1154 /* Compute gl_SamplePosition.x */
1155 compute_sample_position(pos
, int_sample_x
);
1156 pos
= offset(pos
, abld
, 1);
1157 if (dispatch_width
== 8) {
1158 abld
.MOV(int_sample_y
, fs_reg(suboffset(sample_pos_reg
, 1)));
1160 abld
.half(0).MOV(half(int_sample_y
, 0),
1161 fs_reg(suboffset(sample_pos_reg
, 1)));
1162 abld
.half(1).MOV(half(int_sample_y
, 1),
1163 fs_reg(suboffset(sample_pos_reg
, 17)));
1165 /* Compute gl_SamplePosition.y */
1166 compute_sample_position(pos
, int_sample_y
);
1171 fs_visitor::emit_sampleid_setup()
1173 assert(stage
== MESA_SHADER_FRAGMENT
);
1174 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1175 assert(devinfo
->gen
>= 6);
1177 const fs_builder abld
= bld
.annotate("compute sample id");
1178 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::int_type
));
1180 if (!key
->multisample_fbo
) {
1181 /* As per GL_ARB_sample_shading specification:
1182 * "When rendering to a non-multisample buffer, or if multisample
1183 * rasterization is disabled, gl_SampleID will always be zero."
1185 abld
.MOV(*reg
, brw_imm_d(0));
1186 } else if (devinfo
->gen
>= 8) {
1187 /* Sample ID comes in as 4-bit numbers in g1.0:
1189 * 15:12 Slot 3 SampleID (only used in SIMD16)
1190 * 11:8 Slot 2 SampleID (only used in SIMD16)
1191 * 7:4 Slot 1 SampleID
1192 * 3:0 Slot 0 SampleID
1194 * Each slot corresponds to four channels, so we want to replicate each
1195 * half-byte value to 4 channels in a row:
1197 * dst+0: .7 .6 .5 .4 .3 .2 .1 .0
1198 * 7:4 7:4 7:4 7:4 3:0 3:0 3:0 3:0
1200 * dst+1: .7 .6 .5 .4 .3 .2 .1 .0 (if SIMD16)
1201 * 15:12 15:12 15:12 15:12 11:8 11:8 11:8 11:8
1203 * First, we read g1.0 with a <1,8,0>UB region, causing the first 8
1204 * channels to read the first byte (7:0), and the second group of 8
1205 * channels to read the second byte (15:8). Then, we shift right by
1206 * a vector immediate of <4, 4, 4, 4, 0, 0, 0, 0>, moving the slot 1 / 3
1207 * values into place. Finally, we AND with 0xf to keep the low nibble.
1209 * shr(16) tmp<1>W g1.0<1,8,0>B 0x44440000:V
1210 * and(16) dst<1>D tmp<8,8,1>W 0xf:W
1212 * TODO: These payload bits exist on Gen7 too, but they appear to always
1213 * be zero, so this code fails to work. We should find out why.
1215 fs_reg
tmp(VGRF
, alloc
.allocate(1), BRW_REGISTER_TYPE_W
);
1217 abld
.SHR(tmp
, fs_reg(stride(retype(brw_vec1_grf(1, 0),
1218 BRW_REGISTER_TYPE_B
), 1, 8, 0)),
1219 brw_imm_v(0x44440000));
1220 abld
.AND(*reg
, tmp
, brw_imm_w(0xf));
1222 const fs_reg t1
= component(fs_reg(VGRF
, alloc
.allocate(1),
1223 BRW_REGISTER_TYPE_D
), 0);
1224 const fs_reg
t2(VGRF
, alloc
.allocate(1), BRW_REGISTER_TYPE_W
);
1226 /* The PS will be run in MSDISPMODE_PERSAMPLE. For example with
1227 * 8x multisampling, subspan 0 will represent sample N (where N
1228 * is 0, 2, 4 or 6), subspan 1 will represent sample 1, 3, 5 or
1229 * 7. We can find the value of N by looking at R0.0 bits 7:6
1230 * ("Starting Sample Pair Index (SSPI)") and multiplying by two
1231 * (since samples are always delivered in pairs). That is, we
1232 * compute 2*((R0.0 & 0xc0) >> 6) == (R0.0 & 0xc0) >> 5. Then
1233 * we need to add N to the sequence (0, 0, 0, 0, 1, 1, 1, 1) in
1234 * case of SIMD8 and sequence (0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2,
1235 * 2, 3, 3, 3, 3) in case of SIMD16. We compute this sequence by
1236 * populating a temporary variable with the sequence (0, 1, 2, 3),
1237 * and then reading from it using vstride=1, width=4, hstride=0.
1238 * These computations hold good for 4x multisampling as well.
1240 * For 2x MSAA and SIMD16, we want to use the sequence (0, 1, 0, 1):
1241 * the first four slots are sample 0 of subspan 0; the next four
1242 * are sample 1 of subspan 0; the third group is sample 0 of
1243 * subspan 1, and finally sample 1 of subspan 1.
1246 /* SKL+ has an extra bit for the Starting Sample Pair Index to
1247 * accomodate 16x MSAA.
1249 abld
.exec_all().group(1, 0)
1250 .AND(t1
, fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_D
)),
1252 abld
.exec_all().group(1, 0).SHR(t1
, t1
, brw_imm_d(5));
1254 /* This works for both SIMD8 and SIMD16 */
1255 abld
.exec_all().group(4, 0).MOV(t2
, brw_imm_v(0x3210));
1257 /* This special instruction takes care of setting vstride=1,
1258 * width=4, hstride=0 of t2 during an ADD instruction.
1260 abld
.emit(FS_OPCODE_SET_SAMPLE_ID
, *reg
, t1
, t2
);
1267 fs_visitor::emit_samplemaskin_setup()
1269 assert(stage
== MESA_SHADER_FRAGMENT
);
1270 struct brw_wm_prog_data
*wm_prog_data
= brw_wm_prog_data(this->prog_data
);
1271 assert(devinfo
->gen
>= 6);
1273 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::int_type
));
1275 fs_reg
coverage_mask(retype(brw_vec8_grf(payload
.sample_mask_in_reg
, 0),
1276 BRW_REGISTER_TYPE_D
));
1278 if (wm_prog_data
->persample_dispatch
) {
1279 /* gl_SampleMaskIn[] comes from two sources: the input coverage mask,
1280 * and a mask representing which sample is being processed by the
1281 * current shader invocation.
1283 * From the OES_sample_variables specification:
1284 * "When per-sample shading is active due to the use of a fragment input
1285 * qualified by "sample" or due to the use of the gl_SampleID or
1286 * gl_SamplePosition variables, only the bit for the current sample is
1287 * set in gl_SampleMaskIn."
1289 const fs_builder abld
= bld
.annotate("compute gl_SampleMaskIn");
1291 if (nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
].file
== BAD_FILE
)
1292 nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
] = *emit_sampleid_setup();
1294 fs_reg one
= vgrf(glsl_type::int_type
);
1295 fs_reg enabled_mask
= vgrf(glsl_type::int_type
);
1296 abld
.MOV(one
, brw_imm_d(1));
1297 abld
.SHL(enabled_mask
, one
, nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
]);
1298 abld
.AND(*reg
, enabled_mask
, coverage_mask
);
1300 /* In per-pixel mode, the coverage mask is sufficient. */
1301 *reg
= coverage_mask
;
1307 fs_visitor::resolve_source_modifiers(const fs_reg
&src
)
1309 if (!src
.abs
&& !src
.negate
)
1312 fs_reg temp
= bld
.vgrf(src
.type
);
1319 fs_visitor::emit_discard_jump()
1321 assert(brw_wm_prog_data(this->prog_data
)->uses_kill
);
1323 /* For performance, after a discard, jump to the end of the
1324 * shader if all relevant channels have been discarded.
1326 fs_inst
*discard_jump
= bld
.emit(FS_OPCODE_DISCARD_JUMP
);
1327 discard_jump
->flag_subreg
= 1;
1329 discard_jump
->predicate
= BRW_PREDICATE_ALIGN1_ANY4H
;
1330 discard_jump
->predicate_inverse
= true;
1334 fs_visitor::emit_gs_thread_end()
1336 assert(stage
== MESA_SHADER_GEOMETRY
);
1338 struct brw_gs_prog_data
*gs_prog_data
= brw_gs_prog_data(prog_data
);
1340 if (gs_compile
->control_data_header_size_bits
> 0) {
1341 emit_gs_control_data_bits(this->final_gs_vertex_count
);
1344 const fs_builder abld
= bld
.annotate("thread end");
1347 if (gs_prog_data
->static_vertex_count
!= -1) {
1348 foreach_in_list_reverse(fs_inst
, prev
, &this->instructions
) {
1349 if (prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8
||
1350 prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
||
1351 prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
||
1352 prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
) {
1355 /* Delete now dead instructions. */
1356 foreach_in_list_reverse_safe(exec_node
, dead
, &this->instructions
) {
1362 } else if (prev
->is_control_flow() || prev
->has_side_effects()) {
1366 fs_reg hdr
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1367 abld
.MOV(hdr
, fs_reg(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
)));
1368 inst
= abld
.emit(SHADER_OPCODE_URB_WRITE_SIMD8
, reg_undef
, hdr
);
1371 fs_reg payload
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
1372 fs_reg
*sources
= ralloc_array(mem_ctx
, fs_reg
, 2);
1373 sources
[0] = fs_reg(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
));
1374 sources
[1] = this->final_gs_vertex_count
;
1375 abld
.LOAD_PAYLOAD(payload
, sources
, 2, 2);
1376 inst
= abld
.emit(SHADER_OPCODE_URB_WRITE_SIMD8
, reg_undef
, payload
);
1384 fs_visitor::assign_curb_setup()
1386 prog_data
->curb_read_length
= ALIGN(stage_prog_data
->nr_params
, 8) / 8;
1388 /* Map the offsets in the UNIFORM file to fixed HW regs. */
1389 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1390 for (unsigned int i
= 0; i
< inst
->sources
; i
++) {
1391 if (inst
->src
[i
].file
== UNIFORM
) {
1392 int uniform_nr
= inst
->src
[i
].nr
+ inst
->src
[i
].offset
/ 4;
1394 if (uniform_nr
>= 0 && uniform_nr
< (int) uniforms
) {
1395 constant_nr
= push_constant_loc
[uniform_nr
];
1397 /* Section 5.11 of the OpenGL 4.1 spec says:
1398 * "Out-of-bounds reads return undefined values, which include
1399 * values from other variables of the active program or zero."
1400 * Just return the first push constant.
1405 struct brw_reg brw_reg
= brw_vec1_grf(payload
.num_regs
+
1408 brw_reg
.abs
= inst
->src
[i
].abs
;
1409 brw_reg
.negate
= inst
->src
[i
].negate
;
1411 assert(inst
->src
[i
].stride
== 0);
1412 inst
->src
[i
] = byte_offset(
1413 retype(brw_reg
, inst
->src
[i
].type
),
1414 inst
->src
[i
].offset
% 4);
1419 /* This may be updated in assign_urb_setup or assign_vs_urb_setup. */
1420 this->first_non_payload_grf
= payload
.num_regs
+ prog_data
->curb_read_length
;
1424 fs_visitor::calculate_urb_setup()
1426 assert(stage
== MESA_SHADER_FRAGMENT
);
1427 struct brw_wm_prog_data
*prog_data
= brw_wm_prog_data(this->prog_data
);
1428 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1430 memset(prog_data
->urb_setup
, -1,
1431 sizeof(prog_data
->urb_setup
[0]) * VARYING_SLOT_MAX
);
1434 /* Figure out where each of the incoming setup attributes lands. */
1435 if (devinfo
->gen
>= 6) {
1436 if (_mesa_bitcount_64(nir
->info
.inputs_read
&
1437 BRW_FS_VARYING_INPUT_MASK
) <= 16) {
1438 /* The SF/SBE pipeline stage can do arbitrary rearrangement of the
1439 * first 16 varying inputs, so we can put them wherever we want.
1440 * Just put them in order.
1442 * This is useful because it means that (a) inputs not used by the
1443 * fragment shader won't take up valuable register space, and (b) we
1444 * won't have to recompile the fragment shader if it gets paired with
1445 * a different vertex (or geometry) shader.
1447 for (unsigned int i
= 0; i
< VARYING_SLOT_MAX
; i
++) {
1448 if (nir
->info
.inputs_read
& BRW_FS_VARYING_INPUT_MASK
&
1449 BITFIELD64_BIT(i
)) {
1450 prog_data
->urb_setup
[i
] = urb_next
++;
1454 bool include_vue_header
=
1455 nir
->info
.inputs_read
& (VARYING_BIT_LAYER
| VARYING_BIT_VIEWPORT
);
1457 /* We have enough input varyings that the SF/SBE pipeline stage can't
1458 * arbitrarily rearrange them to suit our whim; we have to put them
1459 * in an order that matches the output of the previous pipeline stage
1460 * (geometry or vertex shader).
1462 struct brw_vue_map prev_stage_vue_map
;
1463 brw_compute_vue_map(devinfo
, &prev_stage_vue_map
,
1464 key
->input_slots_valid
,
1465 nir
->info
.separate_shader
);
1467 include_vue_header
? 0 : 2 * BRW_SF_URB_ENTRY_READ_OFFSET
;
1469 assert(prev_stage_vue_map
.num_slots
<= first_slot
+ 32);
1470 for (int slot
= first_slot
; slot
< prev_stage_vue_map
.num_slots
;
1472 int varying
= prev_stage_vue_map
.slot_to_varying
[slot
];
1473 if (varying
!= BRW_VARYING_SLOT_PAD
&&
1474 (nir
->info
.inputs_read
& BRW_FS_VARYING_INPUT_MASK
&
1475 BITFIELD64_BIT(varying
))) {
1476 prog_data
->urb_setup
[varying
] = slot
- first_slot
;
1479 urb_next
= prev_stage_vue_map
.num_slots
- first_slot
;
1482 /* FINISHME: The sf doesn't map VS->FS inputs for us very well. */
1483 for (unsigned int i
= 0; i
< VARYING_SLOT_MAX
; i
++) {
1484 /* Point size is packed into the header, not as a general attribute */
1485 if (i
== VARYING_SLOT_PSIZ
)
1488 if (key
->input_slots_valid
& BITFIELD64_BIT(i
)) {
1489 /* The back color slot is skipped when the front color is
1490 * also written to. In addition, some slots can be
1491 * written in the vertex shader and not read in the
1492 * fragment shader. So the register number must always be
1493 * incremented, mapped or not.
1495 if (_mesa_varying_slot_in_fs((gl_varying_slot
) i
))
1496 prog_data
->urb_setup
[i
] = urb_next
;
1502 * It's a FS only attribute, and we did interpolation for this attribute
1503 * in SF thread. So, count it here, too.
1505 * See compile_sf_prog() for more info.
1507 if (nir
->info
.inputs_read
& BITFIELD64_BIT(VARYING_SLOT_PNTC
))
1508 prog_data
->urb_setup
[VARYING_SLOT_PNTC
] = urb_next
++;
1511 prog_data
->num_varying_inputs
= urb_next
;
1515 fs_visitor::assign_urb_setup()
1517 assert(stage
== MESA_SHADER_FRAGMENT
);
1518 struct brw_wm_prog_data
*prog_data
= brw_wm_prog_data(this->prog_data
);
1520 int urb_start
= payload
.num_regs
+ prog_data
->base
.curb_read_length
;
1522 /* Offset all the urb_setup[] index by the actual position of the
1523 * setup regs, now that the location of the constants has been chosen.
1525 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1526 if (inst
->opcode
== FS_OPCODE_LINTERP
) {
1527 assert(inst
->src
[1].file
== FIXED_GRF
);
1528 inst
->src
[1].nr
+= urb_start
;
1531 if (inst
->opcode
== FS_OPCODE_CINTERP
) {
1532 assert(inst
->src
[0].file
== FIXED_GRF
);
1533 inst
->src
[0].nr
+= urb_start
;
1537 /* Each attribute is 4 setup channels, each of which is half a reg. */
1538 this->first_non_payload_grf
+= prog_data
->num_varying_inputs
* 2;
1542 fs_visitor::convert_attr_sources_to_hw_regs(fs_inst
*inst
)
1544 for (int i
= 0; i
< inst
->sources
; i
++) {
1545 if (inst
->src
[i
].file
== ATTR
) {
1546 int grf
= payload
.num_regs
+
1547 prog_data
->curb_read_length
+
1549 inst
->src
[i
].offset
/ REG_SIZE
;
1551 /* As explained at brw_reg_from_fs_reg, From the Haswell PRM:
1553 * VertStride must be used to cross GRF register boundaries. This
1554 * rule implies that elements within a 'Width' cannot cross GRF
1557 * So, for registers that are large enough, we have to split the exec
1558 * size in two and trust the compression state to sort it out.
1560 unsigned total_size
= inst
->exec_size
*
1561 inst
->src
[i
].stride
*
1562 type_sz(inst
->src
[i
].type
);
1564 assert(total_size
<= 2 * REG_SIZE
);
1565 const unsigned exec_size
=
1566 (total_size
<= REG_SIZE
) ? inst
->exec_size
: inst
->exec_size
/ 2;
1568 unsigned width
= inst
->src
[i
].stride
== 0 ? 1 : exec_size
;
1569 struct brw_reg reg
=
1570 stride(byte_offset(retype(brw_vec8_grf(grf
, 0), inst
->src
[i
].type
),
1571 inst
->src
[i
].offset
% REG_SIZE
),
1572 exec_size
* inst
->src
[i
].stride
,
1573 width
, inst
->src
[i
].stride
);
1574 reg
.abs
= inst
->src
[i
].abs
;
1575 reg
.negate
= inst
->src
[i
].negate
;
1583 fs_visitor::assign_vs_urb_setup()
1585 struct brw_vs_prog_data
*vs_prog_data
= brw_vs_prog_data(prog_data
);
1587 assert(stage
== MESA_SHADER_VERTEX
);
1589 /* Each attribute is 4 regs. */
1590 this->first_non_payload_grf
+= 4 * vs_prog_data
->nr_attribute_slots
;
1592 assert(vs_prog_data
->base
.urb_read_length
<= 15);
1594 /* Rewrite all ATTR file references to the hw grf that they land in. */
1595 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1596 convert_attr_sources_to_hw_regs(inst
);
1601 fs_visitor::assign_tcs_single_patch_urb_setup()
1603 assert(stage
== MESA_SHADER_TESS_CTRL
);
1605 /* Rewrite all ATTR file references to HW_REGs. */
1606 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1607 convert_attr_sources_to_hw_regs(inst
);
1612 fs_visitor::assign_tes_urb_setup()
1614 assert(stage
== MESA_SHADER_TESS_EVAL
);
1616 struct brw_vue_prog_data
*vue_prog_data
= brw_vue_prog_data(prog_data
);
1618 first_non_payload_grf
+= 8 * vue_prog_data
->urb_read_length
;
1620 /* Rewrite all ATTR file references to HW_REGs. */
1621 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1622 convert_attr_sources_to_hw_regs(inst
);
1627 fs_visitor::assign_gs_urb_setup()
1629 assert(stage
== MESA_SHADER_GEOMETRY
);
1631 struct brw_vue_prog_data
*vue_prog_data
= brw_vue_prog_data(prog_data
);
1633 first_non_payload_grf
+=
1634 8 * vue_prog_data
->urb_read_length
* nir
->info
.gs
.vertices_in
;
1636 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1637 /* Rewrite all ATTR file references to GRFs. */
1638 convert_attr_sources_to_hw_regs(inst
);
1644 * Split large virtual GRFs into separate components if we can.
1646 * This is mostly duplicated with what brw_fs_vector_splitting does,
1647 * but that's really conservative because it's afraid of doing
1648 * splitting that doesn't result in real progress after the rest of
1649 * the optimization phases, which would cause infinite looping in
1650 * optimization. We can do it once here, safely. This also has the
1651 * opportunity to split interpolated values, or maybe even uniforms,
1652 * which we don't have at the IR level.
1654 * We want to split, because virtual GRFs are what we register
1655 * allocate and spill (due to contiguousness requirements for some
1656 * instructions), and they're what we naturally generate in the
1657 * codegen process, but most virtual GRFs don't actually need to be
1658 * contiguous sets of GRFs. If we split, we'll end up with reduced
1659 * live intervals and better dead code elimination and coalescing.
1662 fs_visitor::split_virtual_grfs()
1664 /* Compact the register file so we eliminate dead vgrfs. This
1665 * only defines split points for live registers, so if we have
1666 * too large dead registers they will hit assertions later.
1668 compact_virtual_grfs();
1670 int num_vars
= this->alloc
.count
;
1672 /* Count the total number of registers */
1674 int vgrf_to_reg
[num_vars
];
1675 for (int i
= 0; i
< num_vars
; i
++) {
1676 vgrf_to_reg
[i
] = reg_count
;
1677 reg_count
+= alloc
.sizes
[i
];
1680 /* An array of "split points". For each register slot, this indicates
1681 * if this slot can be separated from the previous slot. Every time an
1682 * instruction uses multiple elements of a register (as a source or
1683 * destination), we mark the used slots as inseparable. Then we go
1684 * through and split the registers into the smallest pieces we can.
1686 bool split_points
[reg_count
];
1687 memset(split_points
, 0, sizeof(split_points
));
1689 /* Mark all used registers as fully splittable */
1690 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1691 if (inst
->dst
.file
== VGRF
) {
1692 int reg
= vgrf_to_reg
[inst
->dst
.nr
];
1693 for (unsigned j
= 1; j
< this->alloc
.sizes
[inst
->dst
.nr
]; j
++)
1694 split_points
[reg
+ j
] = true;
1697 for (int i
= 0; i
< inst
->sources
; i
++) {
1698 if (inst
->src
[i
].file
== VGRF
) {
1699 int reg
= vgrf_to_reg
[inst
->src
[i
].nr
];
1700 for (unsigned j
= 1; j
< this->alloc
.sizes
[inst
->src
[i
].nr
]; j
++)
1701 split_points
[reg
+ j
] = true;
1706 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1707 if (inst
->dst
.file
== VGRF
) {
1708 int reg
= vgrf_to_reg
[inst
->dst
.nr
] + inst
->dst
.offset
/ REG_SIZE
;
1709 for (unsigned j
= 1; j
< regs_written(inst
); j
++)
1710 split_points
[reg
+ j
] = false;
1712 for (int i
= 0; i
< inst
->sources
; i
++) {
1713 if (inst
->src
[i
].file
== VGRF
) {
1714 int reg
= vgrf_to_reg
[inst
->src
[i
].nr
] + inst
->src
[i
].offset
/ REG_SIZE
;
1715 for (unsigned j
= 1; j
< regs_read(inst
, i
); j
++)
1716 split_points
[reg
+ j
] = false;
1721 int new_virtual_grf
[reg_count
];
1722 int new_reg_offset
[reg_count
];
1725 for (int i
= 0; i
< num_vars
; i
++) {
1726 /* The first one should always be 0 as a quick sanity check. */
1727 assert(split_points
[reg
] == false);
1730 new_reg_offset
[reg
] = 0;
1735 for (unsigned j
= 1; j
< alloc
.sizes
[i
]; j
++) {
1736 /* If this is a split point, reset the offset to 0 and allocate a
1737 * new virtual GRF for the previous offset many registers
1739 if (split_points
[reg
]) {
1740 assert(offset
<= MAX_VGRF_SIZE
);
1741 int grf
= alloc
.allocate(offset
);
1742 for (int k
= reg
- offset
; k
< reg
; k
++)
1743 new_virtual_grf
[k
] = grf
;
1746 new_reg_offset
[reg
] = offset
;
1751 /* The last one gets the original register number */
1752 assert(offset
<= MAX_VGRF_SIZE
);
1753 alloc
.sizes
[i
] = offset
;
1754 for (int k
= reg
- offset
; k
< reg
; k
++)
1755 new_virtual_grf
[k
] = i
;
1757 assert(reg
== reg_count
);
1759 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1760 if (inst
->dst
.file
== VGRF
) {
1761 reg
= vgrf_to_reg
[inst
->dst
.nr
] + inst
->dst
.offset
/ REG_SIZE
;
1762 inst
->dst
.nr
= new_virtual_grf
[reg
];
1763 inst
->dst
.offset
= new_reg_offset
[reg
] * REG_SIZE
+
1764 inst
->dst
.offset
% REG_SIZE
;
1765 assert((unsigned)new_reg_offset
[reg
] < alloc
.sizes
[new_virtual_grf
[reg
]]);
1767 for (int i
= 0; i
< inst
->sources
; i
++) {
1768 if (inst
->src
[i
].file
== VGRF
) {
1769 reg
= vgrf_to_reg
[inst
->src
[i
].nr
] + inst
->src
[i
].offset
/ REG_SIZE
;
1770 inst
->src
[i
].nr
= new_virtual_grf
[reg
];
1771 inst
->src
[i
].offset
= new_reg_offset
[reg
] * REG_SIZE
+
1772 inst
->src
[i
].offset
% REG_SIZE
;
1773 assert((unsigned)new_reg_offset
[reg
] < alloc
.sizes
[new_virtual_grf
[reg
]]);
1777 invalidate_live_intervals();
1781 * Remove unused virtual GRFs and compact the virtual_grf_* arrays.
1783 * During code generation, we create tons of temporary variables, many of
1784 * which get immediately killed and are never used again. Yet, in later
1785 * optimization and analysis passes, such as compute_live_intervals, we need
1786 * to loop over all the virtual GRFs. Compacting them can save a lot of
1790 fs_visitor::compact_virtual_grfs()
1792 bool progress
= false;
1793 int remap_table
[this->alloc
.count
];
1794 memset(remap_table
, -1, sizeof(remap_table
));
1796 /* Mark which virtual GRFs are used. */
1797 foreach_block_and_inst(block
, const fs_inst
, inst
, cfg
) {
1798 if (inst
->dst
.file
== VGRF
)
1799 remap_table
[inst
->dst
.nr
] = 0;
1801 for (int i
= 0; i
< inst
->sources
; i
++) {
1802 if (inst
->src
[i
].file
== VGRF
)
1803 remap_table
[inst
->src
[i
].nr
] = 0;
1807 /* Compact the GRF arrays. */
1809 for (unsigned i
= 0; i
< this->alloc
.count
; i
++) {
1810 if (remap_table
[i
] == -1) {
1811 /* We just found an unused register. This means that we are
1812 * actually going to compact something.
1816 remap_table
[i
] = new_index
;
1817 alloc
.sizes
[new_index
] = alloc
.sizes
[i
];
1818 invalidate_live_intervals();
1823 this->alloc
.count
= new_index
;
1825 /* Patch all the instructions to use the newly renumbered registers */
1826 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1827 if (inst
->dst
.file
== VGRF
)
1828 inst
->dst
.nr
= remap_table
[inst
->dst
.nr
];
1830 for (int i
= 0; i
< inst
->sources
; i
++) {
1831 if (inst
->src
[i
].file
== VGRF
)
1832 inst
->src
[i
].nr
= remap_table
[inst
->src
[i
].nr
];
1836 /* Patch all the references to delta_xy, since they're used in register
1837 * allocation. If they're unused, switch them to BAD_FILE so we don't
1838 * think some random VGRF is delta_xy.
1840 for (unsigned i
= 0; i
< ARRAY_SIZE(delta_xy
); i
++) {
1841 if (delta_xy
[i
].file
== VGRF
) {
1842 if (remap_table
[delta_xy
[i
].nr
] != -1) {
1843 delta_xy
[i
].nr
= remap_table
[delta_xy
[i
].nr
];
1845 delta_xy
[i
].file
= BAD_FILE
;
1854 set_push_pull_constant_loc(unsigned uniform
, int *chunk_start
,
1855 unsigned *max_chunk_bitsize
,
1856 bool contiguous
, unsigned bitsize
,
1857 const unsigned target_bitsize
,
1858 int *push_constant_loc
, int *pull_constant_loc
,
1859 unsigned *num_push_constants
,
1860 unsigned *num_pull_constants
,
1861 const unsigned max_push_components
,
1862 const unsigned max_chunk_size
,
1863 struct brw_stage_prog_data
*stage_prog_data
)
1865 /* This is the first live uniform in the chunk */
1866 if (*chunk_start
< 0)
1867 *chunk_start
= uniform
;
1869 /* Keep track of the maximum bit size access in contiguous uniforms */
1870 *max_chunk_bitsize
= MAX2(*max_chunk_bitsize
, bitsize
);
1872 /* If this element does not need to be contiguous with the next, we
1873 * split at this point and everything between chunk_start and u forms a
1877 /* If bitsize doesn't match the target one, skip it */
1878 if (*max_chunk_bitsize
!= target_bitsize
) {
1879 /* FIXME: right now we only support 32 and 64-bit accesses */
1880 assert(*max_chunk_bitsize
== 4 || *max_chunk_bitsize
== 8);
1881 *max_chunk_bitsize
= 0;
1886 unsigned chunk_size
= uniform
- *chunk_start
+ 1;
1888 /* Decide whether we should push or pull this parameter. In the
1889 * Vulkan driver, push constants are explicitly exposed via the API
1890 * so we push everything. In GL, we only push small arrays.
1892 if (stage_prog_data
->pull_param
== NULL
||
1893 (*num_push_constants
+ chunk_size
<= max_push_components
&&
1894 chunk_size
<= max_chunk_size
)) {
1895 assert(*num_push_constants
+ chunk_size
<= max_push_components
);
1896 for (unsigned j
= *chunk_start
; j
<= uniform
; j
++)
1897 push_constant_loc
[j
] = (*num_push_constants
)++;
1899 for (unsigned j
= *chunk_start
; j
<= uniform
; j
++)
1900 pull_constant_loc
[j
] = (*num_pull_constants
)++;
1903 *max_chunk_bitsize
= 0;
1909 * Assign UNIFORM file registers to either push constants or pull constants.
1911 * We allow a fragment shader to have more than the specified minimum
1912 * maximum number of fragment shader uniform components (64). If
1913 * there are too many of these, they'd fill up all of register space.
1914 * So, this will push some of them out to the pull constant buffer and
1915 * update the program to load them.
1918 fs_visitor::assign_constant_locations()
1920 /* Only the first compile gets to decide on locations. */
1921 if (dispatch_width
!= min_dispatch_width
)
1924 bool is_live
[uniforms
];
1925 memset(is_live
, 0, sizeof(is_live
));
1926 unsigned bitsize_access
[uniforms
];
1927 memset(bitsize_access
, 0, sizeof(bitsize_access
));
1929 /* For each uniform slot, a value of true indicates that the given slot and
1930 * the next slot must remain contiguous. This is used to keep us from
1931 * splitting arrays apart.
1933 bool contiguous
[uniforms
];
1934 memset(contiguous
, 0, sizeof(contiguous
));
1936 int thread_local_id_index
=
1937 (stage
== MESA_SHADER_COMPUTE
) ?
1938 brw_cs_prog_data(stage_prog_data
)->thread_local_id_index
: -1;
1940 /* First, we walk through the instructions and do two things:
1942 * 1) Figure out which uniforms are live.
1944 * 2) Mark any indirectly used ranges of registers as contiguous.
1946 * Note that we don't move constant-indexed accesses to arrays. No
1947 * testing has been done of the performance impact of this choice.
1949 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
1950 for (int i
= 0 ; i
< inst
->sources
; i
++) {
1951 if (inst
->src
[i
].file
!= UNIFORM
)
1954 int constant_nr
= inst
->src
[i
].nr
+ inst
->src
[i
].offset
/ 4;
1956 if (inst
->opcode
== SHADER_OPCODE_MOV_INDIRECT
&& i
== 0) {
1957 assert(inst
->src
[2].ud
% 4 == 0);
1958 unsigned last
= constant_nr
+ (inst
->src
[2].ud
/ 4) - 1;
1959 assert(last
< uniforms
);
1961 for (unsigned j
= constant_nr
; j
< last
; j
++) {
1963 contiguous
[j
] = true;
1964 bitsize_access
[j
] = MAX2(bitsize_access
[j
], type_sz(inst
->src
[i
].type
));
1966 is_live
[last
] = true;
1967 bitsize_access
[last
] = MAX2(bitsize_access
[last
], type_sz(inst
->src
[i
].type
));
1969 if (constant_nr
>= 0 && constant_nr
< (int) uniforms
) {
1970 int regs_read
= inst
->components_read(i
) *
1971 type_sz(inst
->src
[i
].type
) / 4;
1972 for (int j
= 0; j
< regs_read
; j
++) {
1973 is_live
[constant_nr
+ j
] = true;
1974 bitsize_access
[constant_nr
+ j
] =
1975 MAX2(bitsize_access
[constant_nr
+ j
], type_sz(inst
->src
[i
].type
));
1982 if (thread_local_id_index
>= 0 && !is_live
[thread_local_id_index
])
1983 thread_local_id_index
= -1;
1985 /* Only allow 16 registers (128 uniform components) as push constants.
1987 * Just demote the end of the list. We could probably do better
1988 * here, demoting things that are rarely used in the program first.
1990 * If changing this value, note the limitation about total_regs in
1993 unsigned int max_push_components
= 16 * 8;
1994 if (thread_local_id_index
>= 0)
1995 max_push_components
--; /* Save a slot for the thread ID */
1997 /* We push small arrays, but no bigger than 16 floats. This is big enough
1998 * for a vec4 but hopefully not large enough to push out other stuff. We
1999 * should probably use a better heuristic at some point.
2001 const unsigned int max_chunk_size
= 16;
2003 unsigned int num_push_constants
= 0;
2004 unsigned int num_pull_constants
= 0;
2006 push_constant_loc
= ralloc_array(mem_ctx
, int, uniforms
);
2007 pull_constant_loc
= ralloc_array(mem_ctx
, int, uniforms
);
2009 /* Default to -1 meaning no location */
2010 memset(push_constant_loc
, -1, uniforms
* sizeof(*push_constant_loc
));
2011 memset(pull_constant_loc
, -1, uniforms
* sizeof(*pull_constant_loc
));
2013 int chunk_start
= -1;
2014 unsigned max_chunk_bitsize
= 0;
2016 /* First push 64-bit uniforms to ensure they are properly aligned */
2017 const unsigned uniform_64_bit_size
= type_sz(BRW_REGISTER_TYPE_DF
);
2018 for (unsigned u
= 0; u
< uniforms
; u
++) {
2022 set_push_pull_constant_loc(u
, &chunk_start
, &max_chunk_bitsize
,
2023 contiguous
[u
], bitsize_access
[u
],
2024 uniform_64_bit_size
,
2025 push_constant_loc
, pull_constant_loc
,
2026 &num_push_constants
, &num_pull_constants
,
2027 max_push_components
, max_chunk_size
,
2032 /* Then push the rest of uniforms */
2033 const unsigned uniform_32_bit_size
= type_sz(BRW_REGISTER_TYPE_F
);
2034 for (unsigned u
= 0; u
< uniforms
; u
++) {
2038 /* Skip thread_local_id_index to put it in the last push register. */
2039 if (thread_local_id_index
== (int)u
)
2042 set_push_pull_constant_loc(u
, &chunk_start
, &max_chunk_bitsize
,
2043 contiguous
[u
], bitsize_access
[u
],
2044 uniform_32_bit_size
,
2045 push_constant_loc
, pull_constant_loc
,
2046 &num_push_constants
, &num_pull_constants
,
2047 max_push_components
, max_chunk_size
,
2051 /* Add the CS local thread ID uniform at the end of the push constants */
2052 if (thread_local_id_index
>= 0)
2053 push_constant_loc
[thread_local_id_index
] = num_push_constants
++;
2055 /* As the uniforms are going to be reordered, take the data from a temporary
2056 * copy of the original param[].
2058 gl_constant_value
**param
= ralloc_array(NULL
, gl_constant_value
*,
2059 stage_prog_data
->nr_params
);
2060 memcpy(param
, stage_prog_data
->param
,
2061 sizeof(gl_constant_value
*) * stage_prog_data
->nr_params
);
2062 stage_prog_data
->nr_params
= num_push_constants
;
2063 stage_prog_data
->nr_pull_params
= num_pull_constants
;
2065 /* Up until now, the param[] array has been indexed by reg + offset
2066 * of UNIFORM registers. Move pull constants into pull_param[] and
2067 * condense param[] to only contain the uniforms we chose to push.
2069 * NOTE: Because we are condensing the params[] array, we know that
2070 * push_constant_loc[i] <= i and we can do it in one smooth loop without
2071 * having to make a copy.
2073 int new_thread_local_id_index
= -1;
2074 for (unsigned int i
= 0; i
< uniforms
; i
++) {
2075 const gl_constant_value
*value
= param
[i
];
2077 if (pull_constant_loc
[i
] != -1) {
2078 stage_prog_data
->pull_param
[pull_constant_loc
[i
]] = value
;
2079 } else if (push_constant_loc
[i
] != -1) {
2080 stage_prog_data
->param
[push_constant_loc
[i
]] = value
;
2081 if (thread_local_id_index
== (int)i
)
2082 new_thread_local_id_index
= push_constant_loc
[i
];
2087 if (stage
== MESA_SHADER_COMPUTE
)
2088 brw_cs_prog_data(stage_prog_data
)->thread_local_id_index
=
2089 new_thread_local_id_index
;
2093 * Replace UNIFORM register file access with either UNIFORM_PULL_CONSTANT_LOAD
2094 * or VARYING_PULL_CONSTANT_LOAD instructions which load values into VGRFs.
2097 fs_visitor::lower_constant_loads()
2099 const unsigned index
= stage_prog_data
->binding_table
.pull_constants_start
;
2101 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
2102 /* Set up the annotation tracking for new generated instructions. */
2103 const fs_builder
ibld(this, block
, inst
);
2105 for (int i
= 0; i
< inst
->sources
; i
++) {
2106 if (inst
->src
[i
].file
!= UNIFORM
)
2109 /* We'll handle this case later */
2110 if (inst
->opcode
== SHADER_OPCODE_MOV_INDIRECT
&& i
== 0)
2113 unsigned location
= inst
->src
[i
].nr
+ inst
->src
[i
].offset
/ 4;
2114 if (location
>= uniforms
)
2115 continue; /* Out of bounds access */
2117 int pull_index
= pull_constant_loc
[location
];
2119 if (pull_index
== -1)
2122 assert(inst
->src
[i
].stride
== 0);
2124 const unsigned block_sz
= 64; /* Fetch one cacheline at a time. */
2125 const fs_builder ubld
= ibld
.exec_all().group(block_sz
/ 4, 0);
2126 const fs_reg dst
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
2127 const unsigned base
= pull_index
* 4;
2129 ubld
.emit(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
,
2130 dst
, brw_imm_ud(index
), brw_imm_ud(base
& ~(block_sz
- 1)));
2132 /* Rewrite the instruction to use the temporary VGRF. */
2133 inst
->src
[i
].file
= VGRF
;
2134 inst
->src
[i
].nr
= dst
.nr
;
2135 inst
->src
[i
].offset
= (base
& (block_sz
- 1)) +
2136 inst
->src
[i
].offset
% 4;
2138 brw_mark_surface_used(prog_data
, index
);
2141 if (inst
->opcode
== SHADER_OPCODE_MOV_INDIRECT
&&
2142 inst
->src
[0].file
== UNIFORM
) {
2144 unsigned location
= inst
->src
[0].nr
+ inst
->src
[0].offset
/ 4;
2145 if (location
>= uniforms
)
2146 continue; /* Out of bounds access */
2148 int pull_index
= pull_constant_loc
[location
];
2150 if (pull_index
== -1)
2153 VARYING_PULL_CONSTANT_LOAD(ibld
, inst
->dst
,
2157 inst
->remove(block
);
2159 brw_mark_surface_used(prog_data
, index
);
2162 invalidate_live_intervals();
2166 fs_visitor::opt_algebraic()
2168 bool progress
= false;
2170 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2171 switch (inst
->opcode
) {
2172 case BRW_OPCODE_MOV
:
2173 if (inst
->src
[0].file
!= IMM
)
2176 if (inst
->saturate
) {
2177 if (inst
->dst
.type
!= inst
->src
[0].type
)
2178 assert(!"unimplemented: saturate mixed types");
2180 if (brw_saturate_immediate(inst
->dst
.type
,
2181 &inst
->src
[0].as_brw_reg())) {
2182 inst
->saturate
= false;
2188 case BRW_OPCODE_MUL
:
2189 if (inst
->src
[1].file
!= IMM
)
2193 if (inst
->src
[1].is_one()) {
2194 inst
->opcode
= BRW_OPCODE_MOV
;
2195 inst
->src
[1] = reg_undef
;
2201 if (inst
->src
[1].is_negative_one()) {
2202 inst
->opcode
= BRW_OPCODE_MOV
;
2203 inst
->src
[0].negate
= !inst
->src
[0].negate
;
2204 inst
->src
[1] = reg_undef
;
2210 if (inst
->src
[1].is_zero()) {
2211 inst
->opcode
= BRW_OPCODE_MOV
;
2212 inst
->src
[0] = inst
->src
[1];
2213 inst
->src
[1] = reg_undef
;
2218 if (inst
->src
[0].file
== IMM
) {
2219 assert(inst
->src
[0].type
== BRW_REGISTER_TYPE_F
);
2220 inst
->opcode
= BRW_OPCODE_MOV
;
2221 inst
->src
[0].f
*= inst
->src
[1].f
;
2222 inst
->src
[1] = reg_undef
;
2227 case BRW_OPCODE_ADD
:
2228 if (inst
->src
[1].file
!= IMM
)
2232 if (inst
->src
[1].is_zero()) {
2233 inst
->opcode
= BRW_OPCODE_MOV
;
2234 inst
->src
[1] = reg_undef
;
2239 if (inst
->src
[0].file
== IMM
) {
2240 assert(inst
->src
[0].type
== BRW_REGISTER_TYPE_F
);
2241 inst
->opcode
= BRW_OPCODE_MOV
;
2242 inst
->src
[0].f
+= inst
->src
[1].f
;
2243 inst
->src
[1] = reg_undef
;
2249 if (inst
->src
[0].equals(inst
->src
[1])) {
2250 inst
->opcode
= BRW_OPCODE_MOV
;
2251 inst
->src
[1] = reg_undef
;
2256 case BRW_OPCODE_LRP
:
2257 if (inst
->src
[1].equals(inst
->src
[2])) {
2258 inst
->opcode
= BRW_OPCODE_MOV
;
2259 inst
->src
[0] = inst
->src
[1];
2260 inst
->src
[1] = reg_undef
;
2261 inst
->src
[2] = reg_undef
;
2266 case BRW_OPCODE_CMP
:
2267 if (inst
->conditional_mod
== BRW_CONDITIONAL_GE
&&
2269 inst
->src
[0].negate
&&
2270 inst
->src
[1].is_zero()) {
2271 inst
->src
[0].abs
= false;
2272 inst
->src
[0].negate
= false;
2273 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
2278 case BRW_OPCODE_SEL
:
2279 if (inst
->src
[0].equals(inst
->src
[1])) {
2280 inst
->opcode
= BRW_OPCODE_MOV
;
2281 inst
->src
[1] = reg_undef
;
2282 inst
->predicate
= BRW_PREDICATE_NONE
;
2283 inst
->predicate_inverse
= false;
2285 } else if (inst
->saturate
&& inst
->src
[1].file
== IMM
) {
2286 switch (inst
->conditional_mod
) {
2287 case BRW_CONDITIONAL_LE
:
2288 case BRW_CONDITIONAL_L
:
2289 switch (inst
->src
[1].type
) {
2290 case BRW_REGISTER_TYPE_F
:
2291 if (inst
->src
[1].f
>= 1.0f
) {
2292 inst
->opcode
= BRW_OPCODE_MOV
;
2293 inst
->src
[1] = reg_undef
;
2294 inst
->conditional_mod
= BRW_CONDITIONAL_NONE
;
2302 case BRW_CONDITIONAL_GE
:
2303 case BRW_CONDITIONAL_G
:
2304 switch (inst
->src
[1].type
) {
2305 case BRW_REGISTER_TYPE_F
:
2306 if (inst
->src
[1].f
<= 0.0f
) {
2307 inst
->opcode
= BRW_OPCODE_MOV
;
2308 inst
->src
[1] = reg_undef
;
2309 inst
->conditional_mod
= BRW_CONDITIONAL_NONE
;
2321 case BRW_OPCODE_MAD
:
2322 if (inst
->src
[1].is_zero() || inst
->src
[2].is_zero()) {
2323 inst
->opcode
= BRW_OPCODE_MOV
;
2324 inst
->src
[1] = reg_undef
;
2325 inst
->src
[2] = reg_undef
;
2327 } else if (inst
->src
[0].is_zero()) {
2328 inst
->opcode
= BRW_OPCODE_MUL
;
2329 inst
->src
[0] = inst
->src
[2];
2330 inst
->src
[2] = reg_undef
;
2332 } else if (inst
->src
[1].is_one()) {
2333 inst
->opcode
= BRW_OPCODE_ADD
;
2334 inst
->src
[1] = inst
->src
[2];
2335 inst
->src
[2] = reg_undef
;
2337 } else if (inst
->src
[2].is_one()) {
2338 inst
->opcode
= BRW_OPCODE_ADD
;
2339 inst
->src
[2] = reg_undef
;
2341 } else if (inst
->src
[1].file
== IMM
&& inst
->src
[2].file
== IMM
) {
2342 inst
->opcode
= BRW_OPCODE_ADD
;
2343 inst
->src
[1].f
*= inst
->src
[2].f
;
2344 inst
->src
[2] = reg_undef
;
2348 case SHADER_OPCODE_BROADCAST
:
2349 if (is_uniform(inst
->src
[0])) {
2350 inst
->opcode
= BRW_OPCODE_MOV
;
2352 inst
->force_writemask_all
= true;
2354 } else if (inst
->src
[1].file
== IMM
) {
2355 inst
->opcode
= BRW_OPCODE_MOV
;
2356 inst
->src
[0] = component(inst
->src
[0],
2359 inst
->force_writemask_all
= true;
2368 /* Swap if src[0] is immediate. */
2369 if (progress
&& inst
->is_commutative()) {
2370 if (inst
->src
[0].file
== IMM
) {
2371 fs_reg tmp
= inst
->src
[1];
2372 inst
->src
[1] = inst
->src
[0];
2381 * Optimize sample messages that have constant zero values for the trailing
2382 * texture coordinates. We can just reduce the message length for these
2383 * instructions instead of reserving a register for it. Trailing parameters
2384 * that aren't sent default to zero anyway. This will cause the dead code
2385 * eliminator to remove the MOV instruction that would otherwise be emitted to
2386 * set up the zero value.
2389 fs_visitor::opt_zero_samples()
2391 /* Gen4 infers the texturing opcode based on the message length so we can't
2394 if (devinfo
->gen
< 5)
2397 bool progress
= false;
2399 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2400 if (!inst
->is_tex())
2403 fs_inst
*load_payload
= (fs_inst
*) inst
->prev
;
2405 if (load_payload
->is_head_sentinel() ||
2406 load_payload
->opcode
!= SHADER_OPCODE_LOAD_PAYLOAD
)
2409 /* We don't want to remove the message header or the first parameter.
2410 * Removing the first parameter is not allowed, see the Haswell PRM
2411 * volume 7, page 149:
2413 * "Parameter 0 is required except for the sampleinfo message, which
2414 * has no parameter 0"
2416 while (inst
->mlen
> inst
->header_size
+ inst
->exec_size
/ 8 &&
2417 load_payload
->src
[(inst
->mlen
- inst
->header_size
) /
2418 (inst
->exec_size
/ 8) +
2419 inst
->header_size
- 1].is_zero()) {
2420 inst
->mlen
-= inst
->exec_size
/ 8;
2426 invalidate_live_intervals();
2432 * Optimize sample messages which are followed by the final RT write.
2434 * CHV, and GEN9+ can mark a texturing SEND instruction with EOT to have its
2435 * results sent directly to the framebuffer, bypassing the EU. Recognize the
2436 * final texturing results copied to the framebuffer write payload and modify
2437 * them to write to the framebuffer directly.
2440 fs_visitor::opt_sampler_eot()
2442 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
2444 if (stage
!= MESA_SHADER_FRAGMENT
)
2447 if (devinfo
->gen
!= 9 && !devinfo
->is_cherryview
)
2450 /* FINISHME: It should be possible to implement this optimization when there
2451 * are multiple drawbuffers.
2453 if (key
->nr_color_regions
!= 1)
2456 /* Requires emitting a bunch of saturating MOV instructions during logical
2457 * send lowering to clamp the color payload, which the sampler unit isn't
2458 * going to do for us.
2460 if (key
->clamp_fragment_color
)
2463 /* Look for a texturing instruction immediately before the final FB_WRITE. */
2464 bblock_t
*block
= cfg
->blocks
[cfg
->num_blocks
- 1];
2465 fs_inst
*fb_write
= (fs_inst
*)block
->end();
2466 assert(fb_write
->eot
);
2467 assert(fb_write
->opcode
== FS_OPCODE_FB_WRITE_LOGICAL
);
2469 /* There wasn't one; nothing to do. */
2470 if (unlikely(fb_write
->prev
->is_head_sentinel()))
2473 fs_inst
*tex_inst
= (fs_inst
*) fb_write
->prev
;
2475 /* 3D Sampler » Messages » Message Format
2477 * “Response Length of zero is allowed on all SIMD8* and SIMD16* sampler
2478 * messages except sample+killpix, resinfo, sampleinfo, LOD, and gather4*”
2480 if (tex_inst
->opcode
!= SHADER_OPCODE_TEX_LOGICAL
&&
2481 tex_inst
->opcode
!= SHADER_OPCODE_TXD_LOGICAL
&&
2482 tex_inst
->opcode
!= SHADER_OPCODE_TXF_LOGICAL
&&
2483 tex_inst
->opcode
!= SHADER_OPCODE_TXL_LOGICAL
&&
2484 tex_inst
->opcode
!= FS_OPCODE_TXB_LOGICAL
&&
2485 tex_inst
->opcode
!= SHADER_OPCODE_TXF_CMS_LOGICAL
&&
2486 tex_inst
->opcode
!= SHADER_OPCODE_TXF_CMS_W_LOGICAL
&&
2487 tex_inst
->opcode
!= SHADER_OPCODE_TXF_UMS_LOGICAL
)
2490 /* XXX - This shouldn't be necessary. */
2491 if (tex_inst
->prev
->is_head_sentinel())
2494 /* Check that the FB write sources are fully initialized by the single
2495 * texturing instruction.
2497 for (unsigned i
= 0; i
< FB_WRITE_LOGICAL_NUM_SRCS
; i
++) {
2498 if (i
== FB_WRITE_LOGICAL_SRC_COLOR0
) {
2499 if (!fb_write
->src
[i
].equals(tex_inst
->dst
) ||
2500 fb_write
->size_read(i
) != tex_inst
->size_written
)
2502 } else if (i
!= FB_WRITE_LOGICAL_SRC_COMPONENTS
) {
2503 if (fb_write
->src
[i
].file
!= BAD_FILE
)
2508 assert(!tex_inst
->eot
); /* We can't get here twice */
2509 assert((tex_inst
->offset
& (0xff << 24)) == 0);
2511 const fs_builder
ibld(this, block
, tex_inst
);
2513 tex_inst
->offset
|= fb_write
->target
<< 24;
2514 tex_inst
->eot
= true;
2515 tex_inst
->dst
= ibld
.null_reg_ud();
2516 tex_inst
->size_written
= 0;
2517 fb_write
->remove(cfg
->blocks
[cfg
->num_blocks
- 1]);
2519 /* Marking EOT is sufficient, lower_logical_sends() will notice the EOT
2520 * flag and submit a header together with the sampler message as required
2523 invalidate_live_intervals();
2528 fs_visitor::opt_register_renaming()
2530 bool progress
= false;
2533 int remap
[alloc
.count
];
2534 memset(remap
, -1, sizeof(int) * alloc
.count
);
2536 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2537 if (inst
->opcode
== BRW_OPCODE_IF
|| inst
->opcode
== BRW_OPCODE_DO
) {
2539 } else if (inst
->opcode
== BRW_OPCODE_ENDIF
||
2540 inst
->opcode
== BRW_OPCODE_WHILE
) {
2544 /* Rewrite instruction sources. */
2545 for (int i
= 0; i
< inst
->sources
; i
++) {
2546 if (inst
->src
[i
].file
== VGRF
&&
2547 remap
[inst
->src
[i
].nr
] != -1 &&
2548 remap
[inst
->src
[i
].nr
] != inst
->src
[i
].nr
) {
2549 inst
->src
[i
].nr
= remap
[inst
->src
[i
].nr
];
2554 const int dst
= inst
->dst
.nr
;
2557 inst
->dst
.file
== VGRF
&&
2558 alloc
.sizes
[inst
->dst
.nr
] * REG_SIZE
== inst
->size_written
&&
2559 !inst
->is_partial_write()) {
2560 if (remap
[dst
] == -1) {
2563 remap
[dst
] = alloc
.allocate(regs_written(inst
));
2564 inst
->dst
.nr
= remap
[dst
];
2567 } else if (inst
->dst
.file
== VGRF
&&
2569 remap
[dst
] != dst
) {
2570 inst
->dst
.nr
= remap
[dst
];
2576 invalidate_live_intervals();
2578 for (unsigned i
= 0; i
< ARRAY_SIZE(delta_xy
); i
++) {
2579 if (delta_xy
[i
].file
== VGRF
&& remap
[delta_xy
[i
].nr
] != -1) {
2580 delta_xy
[i
].nr
= remap
[delta_xy
[i
].nr
];
2589 * Remove redundant or useless discard jumps.
2591 * For example, we can eliminate jumps in the following sequence:
2593 * discard-jump (redundant with the next jump)
2594 * discard-jump (useless; jumps to the next instruction)
2598 fs_visitor::opt_redundant_discard_jumps()
2600 bool progress
= false;
2602 bblock_t
*last_bblock
= cfg
->blocks
[cfg
->num_blocks
- 1];
2604 fs_inst
*placeholder_halt
= NULL
;
2605 foreach_inst_in_block_reverse(fs_inst
, inst
, last_bblock
) {
2606 if (inst
->opcode
== FS_OPCODE_PLACEHOLDER_HALT
) {
2607 placeholder_halt
= inst
;
2612 if (!placeholder_halt
)
2615 /* Delete any HALTs immediately before the placeholder halt. */
2616 for (fs_inst
*prev
= (fs_inst
*) placeholder_halt
->prev
;
2617 !prev
->is_head_sentinel() && prev
->opcode
== FS_OPCODE_DISCARD_JUMP
;
2618 prev
= (fs_inst
*) placeholder_halt
->prev
) {
2619 prev
->remove(last_bblock
);
2624 invalidate_live_intervals();
2630 * Compute a bitmask with GRF granularity with a bit set for each GRF starting
2631 * from \p r.offset which overlaps the region starting at \p s.offset and
2632 * spanning \p ds bytes.
2634 static inline unsigned
2635 mask_relative_to(const fs_reg
&r
, const fs_reg
&s
, unsigned ds
)
2637 const int rel_offset
= reg_offset(s
) - reg_offset(r
);
2638 const int shift
= rel_offset
/ REG_SIZE
;
2639 const unsigned n
= DIV_ROUND_UP(rel_offset
% REG_SIZE
+ ds
, REG_SIZE
);
2640 assert(reg_space(r
) == reg_space(s
) &&
2641 shift
>= 0 && shift
< int(8 * sizeof(unsigned)));
2642 return ((1 << n
) - 1) << shift
;
2646 fs_visitor::compute_to_mrf()
2648 bool progress
= false;
2651 /* No MRFs on Gen >= 7. */
2652 if (devinfo
->gen
>= 7)
2655 calculate_live_intervals();
2657 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
2661 if (inst
->opcode
!= BRW_OPCODE_MOV
||
2662 inst
->is_partial_write() ||
2663 inst
->dst
.file
!= MRF
|| inst
->src
[0].file
!= VGRF
||
2664 inst
->dst
.type
!= inst
->src
[0].type
||
2665 inst
->src
[0].abs
|| inst
->src
[0].negate
||
2666 !inst
->src
[0].is_contiguous() ||
2667 inst
->src
[0].offset
% REG_SIZE
!= 0)
2670 /* Can't compute-to-MRF this GRF if someone else was going to
2673 if (this->virtual_grf_end
[inst
->src
[0].nr
] > ip
)
2676 /* Found a move of a GRF to a MRF. Let's see if we can go rewrite the
2677 * things that computed the value of all GRFs of the source region. The
2678 * regs_left bitset keeps track of the registers we haven't yet found a
2679 * generating instruction for.
2681 unsigned regs_left
= (1 << regs_read(inst
, 0)) - 1;
2683 foreach_inst_in_block_reverse_starting_from(fs_inst
, scan_inst
, inst
) {
2684 if (regions_overlap(scan_inst
->dst
, scan_inst
->size_written
,
2685 inst
->src
[0], inst
->size_read(0))) {
2686 /* Found the last thing to write our reg we want to turn
2687 * into a compute-to-MRF.
2690 /* If this one instruction didn't populate all the
2691 * channels, bail. We might be able to rewrite everything
2692 * that writes that reg, but it would require smarter
2695 if (scan_inst
->is_partial_write())
2698 /* Handling things not fully contained in the source of the copy
2699 * would need us to understand coalescing out more than one MOV at
2702 if (!region_contained_in(scan_inst
->dst
, scan_inst
->size_written
,
2703 inst
->src
[0], inst
->size_read(0)))
2706 /* SEND instructions can't have MRF as a destination. */
2707 if (scan_inst
->mlen
)
2710 if (devinfo
->gen
== 6) {
2711 /* gen6 math instructions must have the destination be
2712 * GRF, so no compute-to-MRF for them.
2714 if (scan_inst
->is_math()) {
2719 /* Clear the bits for any registers this instruction overwrites. */
2720 regs_left
&= ~mask_relative_to(
2721 inst
->src
[0], scan_inst
->dst
, scan_inst
->size_written
);
2726 /* We don't handle control flow here. Most computation of
2727 * values that end up in MRFs are shortly before the MRF
2730 if (block
->start() == scan_inst
)
2733 /* You can't read from an MRF, so if someone else reads our
2734 * MRF's source GRF that we wanted to rewrite, that stops us.
2736 bool interfered
= false;
2737 for (int i
= 0; i
< scan_inst
->sources
; i
++) {
2738 if (regions_overlap(scan_inst
->src
[i
], scan_inst
->size_read(i
),
2739 inst
->src
[0], inst
->size_read(0))) {
2746 if (regions_overlap(scan_inst
->dst
, scan_inst
->size_written
,
2747 inst
->dst
, inst
->size_written
)) {
2748 /* If somebody else writes our MRF here, we can't
2749 * compute-to-MRF before that.
2754 if (scan_inst
->mlen
> 0 && scan_inst
->base_mrf
!= -1 &&
2755 regions_overlap(fs_reg(MRF
, scan_inst
->base_mrf
), scan_inst
->mlen
* REG_SIZE
,
2756 inst
->dst
, inst
->size_written
)) {
2757 /* Found a SEND instruction, which means that there are
2758 * live values in MRFs from base_mrf to base_mrf +
2759 * scan_inst->mlen - 1. Don't go pushing our MRF write up
2769 /* Found all generating instructions of our MRF's source value, so it
2770 * should be safe to rewrite them to point to the MRF directly.
2772 regs_left
= (1 << regs_read(inst
, 0)) - 1;
2774 foreach_inst_in_block_reverse_starting_from(fs_inst
, scan_inst
, inst
) {
2775 if (regions_overlap(scan_inst
->dst
, scan_inst
->size_written
,
2776 inst
->src
[0], inst
->size_read(0))) {
2777 /* Clear the bits for any registers this instruction overwrites. */
2778 regs_left
&= ~mask_relative_to(
2779 inst
->src
[0], scan_inst
->dst
, scan_inst
->size_written
);
2781 const unsigned rel_offset
= reg_offset(scan_inst
->dst
) -
2782 reg_offset(inst
->src
[0]);
2784 if (inst
->dst
.nr
& BRW_MRF_COMPR4
) {
2785 /* Apply the same address transformation done by the hardware
2786 * for COMPR4 MRF writes.
2788 assert(rel_offset
< 2 * REG_SIZE
);
2789 scan_inst
->dst
.nr
= inst
->dst
.nr
+ rel_offset
/ REG_SIZE
* 4;
2791 /* Clear the COMPR4 bit if the generating instruction is not
2794 if (scan_inst
->size_written
< 2 * REG_SIZE
)
2795 scan_inst
->dst
.nr
&= ~BRW_MRF_COMPR4
;
2798 /* Calculate the MRF number the result of this instruction is
2799 * ultimately written to.
2801 scan_inst
->dst
.nr
= inst
->dst
.nr
+ rel_offset
/ REG_SIZE
;
2804 scan_inst
->dst
.file
= MRF
;
2805 scan_inst
->dst
.offset
= inst
->dst
.offset
+ rel_offset
% REG_SIZE
;
2806 scan_inst
->saturate
|= inst
->saturate
;
2813 inst
->remove(block
);
2818 invalidate_live_intervals();
2824 * Eliminate FIND_LIVE_CHANNEL instructions occurring outside any control
2825 * flow. We could probably do better here with some form of divergence
2829 fs_visitor::eliminate_find_live_channel()
2831 bool progress
= false;
2834 if (!brw_stage_has_packed_dispatch(devinfo
, stage
, stage_prog_data
)) {
2835 /* The optimization below assumes that channel zero is live on thread
2836 * dispatch, which may not be the case if the fixed function dispatches
2842 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
2843 switch (inst
->opcode
) {
2849 case BRW_OPCODE_ENDIF
:
2850 case BRW_OPCODE_WHILE
:
2854 case FS_OPCODE_DISCARD_JUMP
:
2855 /* This can potentially make control flow non-uniform until the end
2860 case SHADER_OPCODE_FIND_LIVE_CHANNEL
:
2862 inst
->opcode
= BRW_OPCODE_MOV
;
2863 inst
->src
[0] = brw_imm_ud(0u);
2865 inst
->force_writemask_all
= true;
2879 * Once we've generated code, try to convert normal FS_OPCODE_FB_WRITE
2880 * instructions to FS_OPCODE_REP_FB_WRITE.
2883 fs_visitor::emit_repclear_shader()
2885 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
2887 int color_mrf
= base_mrf
+ 2;
2891 mov
= bld
.exec_all().group(4, 0)
2892 .MOV(brw_message_reg(color_mrf
),
2893 fs_reg(UNIFORM
, 0, BRW_REGISTER_TYPE_F
));
2895 struct brw_reg reg
=
2896 brw_reg(BRW_GENERAL_REGISTER_FILE
, 2, 3, 0, 0, BRW_REGISTER_TYPE_F
,
2897 BRW_VERTICAL_STRIDE_8
, BRW_WIDTH_2
, BRW_HORIZONTAL_STRIDE_4
,
2898 BRW_SWIZZLE_XYZW
, WRITEMASK_XYZW
);
2900 mov
= bld
.exec_all().group(4, 0)
2901 .MOV(vec4(brw_message_reg(color_mrf
)), fs_reg(reg
));
2905 if (key
->nr_color_regions
== 1) {
2906 write
= bld
.emit(FS_OPCODE_REP_FB_WRITE
);
2907 write
->saturate
= key
->clamp_fragment_color
;
2908 write
->base_mrf
= color_mrf
;
2910 write
->header_size
= 0;
2913 assume(key
->nr_color_regions
> 0);
2914 for (int i
= 0; i
< key
->nr_color_regions
; ++i
) {
2915 write
= bld
.emit(FS_OPCODE_REP_FB_WRITE
);
2916 write
->saturate
= key
->clamp_fragment_color
;
2917 write
->base_mrf
= base_mrf
;
2919 write
->header_size
= 2;
2927 assign_constant_locations();
2928 assign_curb_setup();
2930 /* Now that we have the uniform assigned, go ahead and force it to a vec4. */
2932 assert(mov
->src
[0].file
== FIXED_GRF
);
2933 mov
->src
[0] = brw_vec4_grf(mov
->src
[0].nr
, 0);
2938 * Walks through basic blocks, looking for repeated MRF writes and
2939 * removing the later ones.
2942 fs_visitor::remove_duplicate_mrf_writes()
2944 fs_inst
*last_mrf_move
[BRW_MAX_MRF(devinfo
->gen
)];
2945 bool progress
= false;
2947 /* Need to update the MRF tracking for compressed instructions. */
2948 if (dispatch_width
>= 16)
2951 memset(last_mrf_move
, 0, sizeof(last_mrf_move
));
2953 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
2954 if (inst
->is_control_flow()) {
2955 memset(last_mrf_move
, 0, sizeof(last_mrf_move
));
2958 if (inst
->opcode
== BRW_OPCODE_MOV
&&
2959 inst
->dst
.file
== MRF
) {
2960 fs_inst
*prev_inst
= last_mrf_move
[inst
->dst
.nr
];
2961 if (prev_inst
&& inst
->equals(prev_inst
)) {
2962 inst
->remove(block
);
2968 /* Clear out the last-write records for MRFs that were overwritten. */
2969 if (inst
->dst
.file
== MRF
) {
2970 last_mrf_move
[inst
->dst
.nr
] = NULL
;
2973 if (inst
->mlen
> 0 && inst
->base_mrf
!= -1) {
2974 /* Found a SEND instruction, which will include two or fewer
2975 * implied MRF writes. We could do better here.
2977 for (int i
= 0; i
< implied_mrf_writes(inst
); i
++) {
2978 last_mrf_move
[inst
->base_mrf
+ i
] = NULL
;
2982 /* Clear out any MRF move records whose sources got overwritten. */
2983 for (unsigned i
= 0; i
< ARRAY_SIZE(last_mrf_move
); i
++) {
2984 if (last_mrf_move
[i
] &&
2985 regions_overlap(inst
->dst
, inst
->size_written
,
2986 last_mrf_move
[i
]->src
[0],
2987 last_mrf_move
[i
]->size_read(0))) {
2988 last_mrf_move
[i
] = NULL
;
2992 if (inst
->opcode
== BRW_OPCODE_MOV
&&
2993 inst
->dst
.file
== MRF
&&
2994 inst
->src
[0].file
!= ARF
&&
2995 !inst
->is_partial_write()) {
2996 last_mrf_move
[inst
->dst
.nr
] = inst
;
3001 invalidate_live_intervals();
3007 clear_deps_for_inst_src(fs_inst
*inst
, bool *deps
, int first_grf
, int grf_len
)
3009 /* Clear the flag for registers that actually got read (as expected). */
3010 for (int i
= 0; i
< inst
->sources
; i
++) {
3012 if (inst
->src
[i
].file
== VGRF
|| inst
->src
[i
].file
== FIXED_GRF
) {
3013 grf
= inst
->src
[i
].nr
;
3018 if (grf
>= first_grf
&&
3019 grf
< first_grf
+ grf_len
) {
3020 deps
[grf
- first_grf
] = false;
3021 if (inst
->exec_size
== 16)
3022 deps
[grf
- first_grf
+ 1] = false;
3028 * Implements this workaround for the original 965:
3030 * "[DevBW, DevCL] Implementation Restrictions: As the hardware does not
3031 * check for post destination dependencies on this instruction, software
3032 * must ensure that there is no destination hazard for the case of ‘write
3033 * followed by a posted write’ shown in the following example.
3036 * 2. send r3.xy <rest of send instruction>
3039 * Due to no post-destination dependency check on the ‘send’, the above
3040 * code sequence could have two instructions (1 and 2) in flight at the
3041 * same time that both consider ‘r3’ as the target of their final writes.
3044 fs_visitor::insert_gen4_pre_send_dependency_workarounds(bblock_t
*block
,
3047 int write_len
= regs_written(inst
);
3048 int first_write_grf
= inst
->dst
.nr
;
3049 bool needs_dep
[BRW_MAX_MRF(devinfo
->gen
)];
3050 assert(write_len
< (int)sizeof(needs_dep
) - 1);
3052 memset(needs_dep
, false, sizeof(needs_dep
));
3053 memset(needs_dep
, true, write_len
);
3055 clear_deps_for_inst_src(inst
, needs_dep
, first_write_grf
, write_len
);
3057 /* Walk backwards looking for writes to registers we're writing which
3058 * aren't read since being written. If we hit the start of the program,
3059 * we assume that there are no outstanding dependencies on entry to the
3062 foreach_inst_in_block_reverse_starting_from(fs_inst
, scan_inst
, inst
) {
3063 /* If we hit control flow, assume that there *are* outstanding
3064 * dependencies, and force their cleanup before our instruction.
3066 if (block
->start() == scan_inst
&& block
->num
!= 0) {
3067 for (int i
= 0; i
< write_len
; i
++) {
3069 DEP_RESOLVE_MOV(fs_builder(this, block
, inst
),
3070 first_write_grf
+ i
);
3075 /* We insert our reads as late as possible on the assumption that any
3076 * instruction but a MOV that might have left us an outstanding
3077 * dependency has more latency than a MOV.
3079 if (scan_inst
->dst
.file
== VGRF
) {
3080 for (unsigned i
= 0; i
< regs_written(scan_inst
); i
++) {
3081 int reg
= scan_inst
->dst
.nr
+ i
;
3083 if (reg
>= first_write_grf
&&
3084 reg
< first_write_grf
+ write_len
&&
3085 needs_dep
[reg
- first_write_grf
]) {
3086 DEP_RESOLVE_MOV(fs_builder(this, block
, inst
), reg
);
3087 needs_dep
[reg
- first_write_grf
] = false;
3088 if (scan_inst
->exec_size
== 16)
3089 needs_dep
[reg
- first_write_grf
+ 1] = false;
3094 /* Clear the flag for registers that actually got read (as expected). */
3095 clear_deps_for_inst_src(scan_inst
, needs_dep
, first_write_grf
, write_len
);
3097 /* Continue the loop only if we haven't resolved all the dependencies */
3099 for (i
= 0; i
< write_len
; i
++) {
3109 * Implements this workaround for the original 965:
3111 * "[DevBW, DevCL] Errata: A destination register from a send can not be
3112 * used as a destination register until after it has been sourced by an
3113 * instruction with a different destination register.
3116 fs_visitor::insert_gen4_post_send_dependency_workarounds(bblock_t
*block
, fs_inst
*inst
)
3118 int write_len
= regs_written(inst
);
3119 int first_write_grf
= inst
->dst
.nr
;
3120 bool needs_dep
[BRW_MAX_MRF(devinfo
->gen
)];
3121 assert(write_len
< (int)sizeof(needs_dep
) - 1);
3123 memset(needs_dep
, false, sizeof(needs_dep
));
3124 memset(needs_dep
, true, write_len
);
3125 /* Walk forwards looking for writes to registers we're writing which aren't
3126 * read before being written.
3128 foreach_inst_in_block_starting_from(fs_inst
, scan_inst
, inst
) {
3129 /* If we hit control flow, force resolve all remaining dependencies. */
3130 if (block
->end() == scan_inst
&& block
->num
!= cfg
->num_blocks
- 1) {
3131 for (int i
= 0; i
< write_len
; i
++) {
3133 DEP_RESOLVE_MOV(fs_builder(this, block
, scan_inst
),
3134 first_write_grf
+ i
);
3139 /* Clear the flag for registers that actually got read (as expected). */
3140 clear_deps_for_inst_src(scan_inst
, needs_dep
, first_write_grf
, write_len
);
3142 /* We insert our reads as late as possible since they're reading the
3143 * result of a SEND, which has massive latency.
3145 if (scan_inst
->dst
.file
== VGRF
&&
3146 scan_inst
->dst
.nr
>= first_write_grf
&&
3147 scan_inst
->dst
.nr
< first_write_grf
+ write_len
&&
3148 needs_dep
[scan_inst
->dst
.nr
- first_write_grf
]) {
3149 DEP_RESOLVE_MOV(fs_builder(this, block
, scan_inst
),
3151 needs_dep
[scan_inst
->dst
.nr
- first_write_grf
] = false;
3154 /* Continue the loop only if we haven't resolved all the dependencies */
3156 for (i
= 0; i
< write_len
; i
++) {
3166 fs_visitor::insert_gen4_send_dependency_workarounds()
3168 if (devinfo
->gen
!= 4 || devinfo
->is_g4x
)
3171 bool progress
= false;
3173 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
3174 if (inst
->mlen
!= 0 && inst
->dst
.file
== VGRF
) {
3175 insert_gen4_pre_send_dependency_workarounds(block
, inst
);
3176 insert_gen4_post_send_dependency_workarounds(block
, inst
);
3182 invalidate_live_intervals();
3186 * Turns the generic expression-style uniform pull constant load instruction
3187 * into a hardware-specific series of instructions for loading a pull
3190 * The expression style allows the CSE pass before this to optimize out
3191 * repeated loads from the same offset, and gives the pre-register-allocation
3192 * scheduling full flexibility, while the conversion to native instructions
3193 * allows the post-register-allocation scheduler the best information
3196 * Note that execution masking for setting up pull constant loads is special:
3197 * the channels that need to be written are unrelated to the current execution
3198 * mask, since a later instruction will use one of the result channels as a
3199 * source operand for all 8 or 16 of its channels.
3202 fs_visitor::lower_uniform_pull_constant_loads()
3204 foreach_block_and_inst (block
, fs_inst
, inst
, cfg
) {
3205 if (inst
->opcode
!= FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
)
3208 if (devinfo
->gen
>= 7) {
3209 const fs_builder ubld
= fs_builder(this, block
, inst
).exec_all();
3210 const fs_reg payload
= ubld
.group(8, 0).vgrf(BRW_REGISTER_TYPE_UD
);
3212 ubld
.group(8, 0).MOV(payload
,
3213 retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
));
3214 ubld
.group(1, 0).MOV(component(payload
, 2),
3215 brw_imm_ud(inst
->src
[1].ud
/ 16));
3217 inst
->opcode
= FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7
;
3218 inst
->src
[1] = payload
;
3219 inst
->header_size
= 1;
3222 invalidate_live_intervals();
3224 /* Before register allocation, we didn't tell the scheduler about the
3225 * MRF we use. We know it's safe to use this MRF because nothing
3226 * else does except for register spill/unspill, which generates and
3227 * uses its MRF within a single IR instruction.
3229 inst
->base_mrf
= FIRST_PULL_LOAD_MRF(devinfo
->gen
) + 1;
3236 fs_visitor::lower_load_payload()
3238 bool progress
= false;
3240 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
3241 if (inst
->opcode
!= SHADER_OPCODE_LOAD_PAYLOAD
)
3244 assert(inst
->dst
.file
== MRF
|| inst
->dst
.file
== VGRF
);
3245 assert(inst
->saturate
== false);
3246 fs_reg dst
= inst
->dst
;
3248 /* Get rid of COMPR4. We'll add it back in if we need it */
3249 if (dst
.file
== MRF
)
3250 dst
.nr
= dst
.nr
& ~BRW_MRF_COMPR4
;
3252 const fs_builder
ibld(this, block
, inst
);
3253 const fs_builder hbld
= ibld
.exec_all().group(8, 0);
3255 for (uint8_t i
= 0; i
< inst
->header_size
; i
++) {
3256 if (inst
->src
[i
].file
!= BAD_FILE
) {
3257 fs_reg mov_dst
= retype(dst
, BRW_REGISTER_TYPE_UD
);
3258 fs_reg mov_src
= retype(inst
->src
[i
], BRW_REGISTER_TYPE_UD
);
3259 hbld
.MOV(mov_dst
, mov_src
);
3261 dst
= offset(dst
, hbld
, 1);
3264 if (inst
->dst
.file
== MRF
&& (inst
->dst
.nr
& BRW_MRF_COMPR4
) &&
3265 inst
->exec_size
> 8) {
3266 /* In this case, the payload portion of the LOAD_PAYLOAD isn't
3267 * a straightforward copy. Instead, the result of the
3268 * LOAD_PAYLOAD is treated as interleaved and the first four
3269 * non-header sources are unpacked as:
3280 * This is used for gen <= 5 fb writes.
3282 assert(inst
->exec_size
== 16);
3283 assert(inst
->header_size
+ 4 <= inst
->sources
);
3284 for (uint8_t i
= inst
->header_size
; i
< inst
->header_size
+ 4; i
++) {
3285 if (inst
->src
[i
].file
!= BAD_FILE
) {
3286 if (devinfo
->has_compr4
) {
3287 fs_reg compr4_dst
= retype(dst
, inst
->src
[i
].type
);
3288 compr4_dst
.nr
|= BRW_MRF_COMPR4
;
3289 ibld
.MOV(compr4_dst
, inst
->src
[i
]);
3291 /* Platform doesn't have COMPR4. We have to fake it */
3292 fs_reg mov_dst
= retype(dst
, inst
->src
[i
].type
);
3293 ibld
.half(0).MOV(mov_dst
, half(inst
->src
[i
], 0));
3295 ibld
.half(1).MOV(mov_dst
, half(inst
->src
[i
], 1));
3302 /* The loop above only ever incremented us through the first set
3303 * of 4 registers. However, thanks to the magic of COMPR4, we
3304 * actually wrote to the first 8 registers, so we need to take
3305 * that into account now.
3309 /* The COMPR4 code took care of the first 4 sources. We'll let
3310 * the regular path handle any remaining sources. Yes, we are
3311 * modifying the instruction but we're about to delete it so
3312 * this really doesn't hurt anything.
3314 inst
->header_size
+= 4;
3317 for (uint8_t i
= inst
->header_size
; i
< inst
->sources
; i
++) {
3318 if (inst
->src
[i
].file
!= BAD_FILE
)
3319 ibld
.MOV(retype(dst
, inst
->src
[i
].type
), inst
->src
[i
]);
3320 dst
= offset(dst
, ibld
, 1);
3323 inst
->remove(block
);
3328 invalidate_live_intervals();
3334 fs_visitor::lower_integer_multiplication()
3336 bool progress
= false;
3338 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
3339 const fs_builder
ibld(this, block
, inst
);
3341 if (inst
->opcode
== BRW_OPCODE_MUL
) {
3342 if (inst
->dst
.is_accumulator() ||
3343 (inst
->dst
.type
!= BRW_REGISTER_TYPE_D
&&
3344 inst
->dst
.type
!= BRW_REGISTER_TYPE_UD
))
3347 /* Gen8's MUL instruction can do a 32-bit x 32-bit -> 32-bit
3348 * operation directly, but CHV/BXT cannot.
3350 if (devinfo
->gen
>= 8 &&
3351 !devinfo
->is_cherryview
&& !gen_device_info_is_9lp(devinfo
))
3354 if (inst
->src
[1].file
== IMM
&&
3355 inst
->src
[1].ud
< (1 << 16)) {
3356 /* The MUL instruction isn't commutative. On Gen <= 6, only the low
3357 * 16-bits of src0 are read, and on Gen >= 7 only the low 16-bits of
3360 * If multiplying by an immediate value that fits in 16-bits, do a
3361 * single MUL instruction with that value in the proper location.
3363 if (devinfo
->gen
< 7) {
3364 fs_reg
imm(VGRF
, alloc
.allocate(dispatch_width
/ 8),
3366 ibld
.MOV(imm
, inst
->src
[1]);
3367 ibld
.MUL(inst
->dst
, imm
, inst
->src
[0]);
3369 const bool ud
= (inst
->src
[1].type
== BRW_REGISTER_TYPE_UD
);
3370 ibld
.MUL(inst
->dst
, inst
->src
[0],
3371 ud
? brw_imm_uw(inst
->src
[1].ud
)
3372 : brw_imm_w(inst
->src
[1].d
));
3375 /* Gen < 8 (and some Gen8+ low-power parts like Cherryview) cannot
3376 * do 32-bit integer multiplication in one instruction, but instead
3377 * must do a sequence (which actually calculates a 64-bit result):
3379 * mul(8) acc0<1>D g3<8,8,1>D g4<8,8,1>D
3380 * mach(8) null g3<8,8,1>D g4<8,8,1>D
3381 * mov(8) g2<1>D acc0<8,8,1>D
3383 * But on Gen > 6, the ability to use second accumulator register
3384 * (acc1) for non-float data types was removed, preventing a simple
3385 * implementation in SIMD16. A 16-channel result can be calculated by
3386 * executing the three instructions twice in SIMD8, once with quarter
3387 * control of 1Q for the first eight channels and again with 2Q for
3388 * the second eight channels.
3390 * Which accumulator register is implicitly accessed (by AccWrEnable
3391 * for instance) is determined by the quarter control. Unfortunately
3392 * Ivybridge (and presumably Baytrail) has a hardware bug in which an
3393 * implicit accumulator access by an instruction with 2Q will access
3394 * acc1 regardless of whether the data type is usable in acc1.
3396 * Specifically, the 2Q mach(8) writes acc1 which does not exist for
3397 * integer data types.
3399 * Since we only want the low 32-bits of the result, we can do two
3400 * 32-bit x 16-bit multiplies (like the mul and mach are doing), and
3401 * adjust the high result and add them (like the mach is doing):
3403 * mul(8) g7<1>D g3<8,8,1>D g4.0<8,8,1>UW
3404 * mul(8) g8<1>D g3<8,8,1>D g4.1<8,8,1>UW
3405 * shl(8) g9<1>D g8<8,8,1>D 16D
3406 * add(8) g2<1>D g7<8,8,1>D g8<8,8,1>D
3408 * We avoid the shl instruction by realizing that we only want to add
3409 * the low 16-bits of the "high" result to the high 16-bits of the
3410 * "low" result and using proper regioning on the add:
3412 * mul(8) g7<1>D g3<8,8,1>D g4.0<16,8,2>UW
3413 * mul(8) g8<1>D g3<8,8,1>D g4.1<16,8,2>UW
3414 * add(8) g7.1<2>UW g7.1<16,8,2>UW g8<16,8,2>UW
3416 * Since it does not use the (single) accumulator register, we can
3417 * schedule multi-component multiplications much better.
3420 fs_reg orig_dst
= inst
->dst
;
3421 if (orig_dst
.is_null() || orig_dst
.file
== MRF
) {
3422 inst
->dst
= fs_reg(VGRF
, alloc
.allocate(dispatch_width
/ 8),
3425 fs_reg low
= inst
->dst
;
3426 fs_reg
high(VGRF
, alloc
.allocate(dispatch_width
/ 8),
3429 if (devinfo
->gen
>= 7) {
3430 if (inst
->src
[1].file
== IMM
) {
3431 ibld
.MUL(low
, inst
->src
[0],
3432 brw_imm_uw(inst
->src
[1].ud
& 0xffff));
3433 ibld
.MUL(high
, inst
->src
[0],
3434 brw_imm_uw(inst
->src
[1].ud
>> 16));
3436 ibld
.MUL(low
, inst
->src
[0],
3437 subscript(inst
->src
[1], BRW_REGISTER_TYPE_UW
, 0));
3438 ibld
.MUL(high
, inst
->src
[0],
3439 subscript(inst
->src
[1], BRW_REGISTER_TYPE_UW
, 1));
3442 ibld
.MUL(low
, subscript(inst
->src
[0], BRW_REGISTER_TYPE_UW
, 0),
3444 ibld
.MUL(high
, subscript(inst
->src
[0], BRW_REGISTER_TYPE_UW
, 1),
3448 ibld
.ADD(subscript(inst
->dst
, BRW_REGISTER_TYPE_UW
, 1),
3449 subscript(low
, BRW_REGISTER_TYPE_UW
, 1),
3450 subscript(high
, BRW_REGISTER_TYPE_UW
, 0));
3452 if (inst
->conditional_mod
|| orig_dst
.file
== MRF
) {
3453 set_condmod(inst
->conditional_mod
,
3454 ibld
.MOV(orig_dst
, inst
->dst
));
3458 } else if (inst
->opcode
== SHADER_OPCODE_MULH
) {
3459 /* Should have been lowered to 8-wide. */
3460 assert(inst
->exec_size
<= get_lowered_simd_width(devinfo
, inst
));
3461 const fs_reg acc
= retype(brw_acc_reg(inst
->exec_size
),
3463 fs_inst
*mul
= ibld
.MUL(acc
, inst
->src
[0], inst
->src
[1]);
3464 fs_inst
*mach
= ibld
.MACH(inst
->dst
, inst
->src
[0], inst
->src
[1]);
3466 if (devinfo
->gen
>= 8) {
3467 /* Until Gen8, integer multiplies read 32-bits from one source,
3468 * and 16-bits from the other, and relying on the MACH instruction
3469 * to generate the high bits of the result.
3471 * On Gen8, the multiply instruction does a full 32x32-bit
3472 * multiply, but in order to do a 64-bit multiply we can simulate
3473 * the previous behavior and then use a MACH instruction.
3475 * FINISHME: Don't use source modifiers on src1.
3477 assert(mul
->src
[1].type
== BRW_REGISTER_TYPE_D
||
3478 mul
->src
[1].type
== BRW_REGISTER_TYPE_UD
);
3479 mul
->src
[1].type
= BRW_REGISTER_TYPE_UW
;
3480 mul
->src
[1].stride
*= 2;
3482 } else if (devinfo
->gen
== 7 && !devinfo
->is_haswell
&&
3484 /* Among other things the quarter control bits influence which
3485 * accumulator register is used by the hardware for instructions
3486 * that access the accumulator implicitly (e.g. MACH). A
3487 * second-half instruction would normally map to acc1, which
3488 * doesn't exist on Gen7 and up (the hardware does emulate it for
3489 * floating-point instructions *only* by taking advantage of the
3490 * extra precision of acc0 not normally used for floating point
3493 * HSW and up are careful enough not to try to access an
3494 * accumulator register that doesn't exist, but on earlier Gen7
3495 * hardware we need to make sure that the quarter control bits are
3496 * zero to avoid non-deterministic behaviour and emit an extra MOV
3497 * to get the result masked correctly according to the current
3501 mach
->force_writemask_all
= true;
3502 mach
->dst
= ibld
.vgrf(inst
->dst
.type
);
3503 ibld
.MOV(inst
->dst
, mach
->dst
);
3509 inst
->remove(block
);
3514 invalidate_live_intervals();
3520 fs_visitor::lower_minmax()
3522 assert(devinfo
->gen
< 6);
3524 bool progress
= false;
3526 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
3527 const fs_builder
ibld(this, block
, inst
);
3529 if (inst
->opcode
== BRW_OPCODE_SEL
&&
3530 inst
->predicate
== BRW_PREDICATE_NONE
) {
3531 /* FIXME: Using CMP doesn't preserve the NaN propagation semantics of
3532 * the original SEL.L/GE instruction
3534 ibld
.CMP(ibld
.null_reg_d(), inst
->src
[0], inst
->src
[1],
3535 inst
->conditional_mod
);
3536 inst
->predicate
= BRW_PREDICATE_NORMAL
;
3537 inst
->conditional_mod
= BRW_CONDITIONAL_NONE
;
3544 invalidate_live_intervals();
3550 setup_color_payload(const fs_builder
&bld
, const brw_wm_prog_key
*key
,
3551 fs_reg
*dst
, fs_reg color
, unsigned components
)
3553 if (key
->clamp_fragment_color
) {
3554 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 4);
3555 assert(color
.type
== BRW_REGISTER_TYPE_F
);
3557 for (unsigned i
= 0; i
< components
; i
++)
3559 bld
.MOV(offset(tmp
, bld
, i
), offset(color
, bld
, i
)));
3564 for (unsigned i
= 0; i
< components
; i
++)
3565 dst
[i
] = offset(color
, bld
, i
);
3569 lower_fb_write_logical_send(const fs_builder
&bld
, fs_inst
*inst
,
3570 const struct brw_wm_prog_data
*prog_data
,
3571 const brw_wm_prog_key
*key
,
3572 const fs_visitor::thread_payload
&payload
)
3574 assert(inst
->src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].file
== IMM
);
3575 const gen_device_info
*devinfo
= bld
.shader
->devinfo
;
3576 const fs_reg
&color0
= inst
->src
[FB_WRITE_LOGICAL_SRC_COLOR0
];
3577 const fs_reg
&color1
= inst
->src
[FB_WRITE_LOGICAL_SRC_COLOR1
];
3578 const fs_reg
&src0_alpha
= inst
->src
[FB_WRITE_LOGICAL_SRC_SRC0_ALPHA
];
3579 const fs_reg
&src_depth
= inst
->src
[FB_WRITE_LOGICAL_SRC_SRC_DEPTH
];
3580 const fs_reg
&dst_depth
= inst
->src
[FB_WRITE_LOGICAL_SRC_DST_DEPTH
];
3581 const fs_reg
&src_stencil
= inst
->src
[FB_WRITE_LOGICAL_SRC_SRC_STENCIL
];
3582 fs_reg sample_mask
= inst
->src
[FB_WRITE_LOGICAL_SRC_OMASK
];
3583 const unsigned components
=
3584 inst
->src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].ud
;
3586 /* We can potentially have a message length of up to 15, so we have to set
3587 * base_mrf to either 0 or 1 in order to fit in m0..m15.
3590 int header_size
= 2, payload_header_size
;
3591 unsigned length
= 0;
3593 /* From the Sandy Bridge PRM, volume 4, page 198:
3595 * "Dispatched Pixel Enables. One bit per pixel indicating
3596 * which pixels were originally enabled when the thread was
3597 * dispatched. This field is only required for the end-of-
3598 * thread message and on all dual-source messages."
3600 if (devinfo
->gen
>= 6 &&
3601 (devinfo
->is_haswell
|| devinfo
->gen
>= 8 || !prog_data
->uses_kill
) &&
3602 color1
.file
== BAD_FILE
&&
3603 key
->nr_color_regions
== 1) {
3607 if (header_size
!= 0) {
3608 assert(header_size
== 2);
3609 /* Allocate 2 registers for a header */
3613 if (payload
.aa_dest_stencil_reg
) {
3614 sources
[length
] = fs_reg(VGRF
, bld
.shader
->alloc
.allocate(1));
3615 bld
.group(8, 0).exec_all().annotate("FB write stencil/AA alpha")
3616 .MOV(sources
[length
],
3617 fs_reg(brw_vec8_grf(payload
.aa_dest_stencil_reg
, 0)));
3621 if (sample_mask
.file
!= BAD_FILE
) {
3622 sources
[length
] = fs_reg(VGRF
, bld
.shader
->alloc
.allocate(1),
3623 BRW_REGISTER_TYPE_UD
);
3625 /* Hand over gl_SampleMask. Only the lower 16 bits of each channel are
3626 * relevant. Since it's unsigned single words one vgrf is always
3627 * 16-wide, but only the lower or higher 8 channels will be used by the
3628 * hardware when doing a SIMD8 write depending on whether we have
3629 * selected the subspans for the first or second half respectively.
3631 assert(sample_mask
.file
!= BAD_FILE
&& type_sz(sample_mask
.type
) == 4);
3632 sample_mask
.type
= BRW_REGISTER_TYPE_UW
;
3633 sample_mask
.stride
*= 2;
3635 bld
.exec_all().annotate("FB write oMask")
3636 .MOV(horiz_offset(retype(sources
[length
], BRW_REGISTER_TYPE_UW
),
3642 payload_header_size
= length
;
3644 if (src0_alpha
.file
!= BAD_FILE
) {
3645 /* FIXME: This is being passed at the wrong location in the payload and
3646 * doesn't work when gl_SampleMask and MRTs are used simultaneously.
3647 * It's supposed to be immediately before oMask but there seems to be no
3648 * reasonable way to pass them in the correct order because LOAD_PAYLOAD
3649 * requires header sources to form a contiguous segment at the beginning
3650 * of the message and src0_alpha has per-channel semantics.
3652 setup_color_payload(bld
, key
, &sources
[length
], src0_alpha
, 1);
3654 } else if (key
->replicate_alpha
&& inst
->target
!= 0) {
3655 /* Handle the case when fragment shader doesn't write to draw buffer
3656 * zero. No need to call setup_color_payload() for src0_alpha because
3657 * alpha value will be undefined.
3662 setup_color_payload(bld
, key
, &sources
[length
], color0
, components
);
3665 if (color1
.file
!= BAD_FILE
) {
3666 setup_color_payload(bld
, key
, &sources
[length
], color1
, components
);
3670 if (src_depth
.file
!= BAD_FILE
) {
3671 sources
[length
] = src_depth
;
3675 if (dst_depth
.file
!= BAD_FILE
) {
3676 sources
[length
] = dst_depth
;
3680 if (src_stencil
.file
!= BAD_FILE
) {
3681 assert(devinfo
->gen
>= 9);
3682 assert(bld
.dispatch_width() != 16);
3684 /* XXX: src_stencil is only available on gen9+. dst_depth is never
3685 * available on gen9+. As such it's impossible to have both enabled at the
3686 * same time and therefore length cannot overrun the array.
3688 assert(length
< 15);
3690 sources
[length
] = bld
.vgrf(BRW_REGISTER_TYPE_UD
);
3691 bld
.exec_all().annotate("FB write OS")
3692 .MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UB
),
3693 subscript(src_stencil
, BRW_REGISTER_TYPE_UB
, 0));
3698 if (devinfo
->gen
>= 7) {
3699 /* Send from the GRF */
3700 fs_reg payload
= fs_reg(VGRF
, -1, BRW_REGISTER_TYPE_F
);
3701 load
= bld
.LOAD_PAYLOAD(payload
, sources
, length
, payload_header_size
);
3702 payload
.nr
= bld
.shader
->alloc
.allocate(regs_written(load
));
3703 load
->dst
= payload
;
3705 inst
->src
[0] = payload
;
3706 inst
->resize_sources(1);
3708 /* Send from the MRF */
3709 load
= bld
.LOAD_PAYLOAD(fs_reg(MRF
, 1, BRW_REGISTER_TYPE_F
),
3710 sources
, length
, payload_header_size
);
3712 /* On pre-SNB, we have to interlace the color values. LOAD_PAYLOAD
3713 * will do this for us if we just give it a COMPR4 destination.
3715 if (devinfo
->gen
< 6 && bld
.dispatch_width() == 16)
3716 load
->dst
.nr
|= BRW_MRF_COMPR4
;
3718 inst
->resize_sources(0);
3722 inst
->opcode
= FS_OPCODE_FB_WRITE
;
3723 inst
->mlen
= regs_written(load
);
3724 inst
->header_size
= header_size
;
3728 lower_fb_read_logical_send(const fs_builder
&bld
, fs_inst
*inst
)
3730 const fs_builder
&ubld
= bld
.exec_all();
3731 const unsigned length
= 2;
3732 const fs_reg header
= ubld
.group(8, 0).vgrf(BRW_REGISTER_TYPE_UD
, length
);
3735 .MOV(header
, retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
));
3737 inst
->resize_sources(1);
3738 inst
->src
[0] = header
;
3739 inst
->opcode
= FS_OPCODE_FB_READ
;
3740 inst
->mlen
= length
;
3741 inst
->header_size
= length
;
3745 lower_sampler_logical_send_gen4(const fs_builder
&bld
, fs_inst
*inst
, opcode op
,
3746 const fs_reg
&coordinate
,
3747 const fs_reg
&shadow_c
,
3748 const fs_reg
&lod
, const fs_reg
&lod2
,
3749 const fs_reg
&surface
,
3750 const fs_reg
&sampler
,
3751 unsigned coord_components
,
3752 unsigned grad_components
)
3754 const bool has_lod
= (op
== SHADER_OPCODE_TXL
|| op
== FS_OPCODE_TXB
||
3755 op
== SHADER_OPCODE_TXF
|| op
== SHADER_OPCODE_TXS
);
3756 fs_reg
msg_begin(MRF
, 1, BRW_REGISTER_TYPE_F
);
3757 fs_reg msg_end
= msg_begin
;
3760 msg_end
= offset(msg_end
, bld
.group(8, 0), 1);
3762 for (unsigned i
= 0; i
< coord_components
; i
++)
3763 bld
.MOV(retype(offset(msg_end
, bld
, i
), coordinate
.type
),
3764 offset(coordinate
, bld
, i
));
3766 msg_end
= offset(msg_end
, bld
, coord_components
);
3768 /* Messages other than SAMPLE and RESINFO in SIMD16 and TXD in SIMD8
3769 * require all three components to be present and zero if they are unused.
3771 if (coord_components
> 0 &&
3772 (has_lod
|| shadow_c
.file
!= BAD_FILE
||
3773 (op
== SHADER_OPCODE_TEX
&& bld
.dispatch_width() == 8))) {
3774 for (unsigned i
= coord_components
; i
< 3; i
++)
3775 bld
.MOV(offset(msg_end
, bld
, i
), brw_imm_f(0.0f
));
3777 msg_end
= offset(msg_end
, bld
, 3 - coord_components
);
3780 if (op
== SHADER_OPCODE_TXD
) {
3781 /* TXD unsupported in SIMD16 mode. */
3782 assert(bld
.dispatch_width() == 8);
3784 /* the slots for u and v are always present, but r is optional */
3785 if (coord_components
< 2)
3786 msg_end
= offset(msg_end
, bld
, 2 - coord_components
);
3789 * dPdx = dudx, dvdx, drdx
3790 * dPdy = dudy, dvdy, drdy
3792 * 1-arg: Does not exist.
3794 * 2-arg: dudx dvdx dudy dvdy
3795 * dPdx.x dPdx.y dPdy.x dPdy.y
3798 * 3-arg: dudx dvdx drdx dudy dvdy drdy
3799 * dPdx.x dPdx.y dPdx.z dPdy.x dPdy.y dPdy.z
3800 * m5 m6 m7 m8 m9 m10
3802 for (unsigned i
= 0; i
< grad_components
; i
++)
3803 bld
.MOV(offset(msg_end
, bld
, i
), offset(lod
, bld
, i
));
3805 msg_end
= offset(msg_end
, bld
, MAX2(grad_components
, 2));
3807 for (unsigned i
= 0; i
< grad_components
; i
++)
3808 bld
.MOV(offset(msg_end
, bld
, i
), offset(lod2
, bld
, i
));
3810 msg_end
= offset(msg_end
, bld
, MAX2(grad_components
, 2));
3814 /* Bias/LOD with shadow comparator is unsupported in SIMD16 -- *Without*
3815 * shadow comparator (including RESINFO) it's unsupported in SIMD8 mode.
3817 assert(shadow_c
.file
!= BAD_FILE
? bld
.dispatch_width() == 8 :
3818 bld
.dispatch_width() == 16);
3820 const brw_reg_type type
=
3821 (op
== SHADER_OPCODE_TXF
|| op
== SHADER_OPCODE_TXS
?
3822 BRW_REGISTER_TYPE_UD
: BRW_REGISTER_TYPE_F
);
3823 bld
.MOV(retype(msg_end
, type
), lod
);
3824 msg_end
= offset(msg_end
, bld
, 1);
3827 if (shadow_c
.file
!= BAD_FILE
) {
3828 if (op
== SHADER_OPCODE_TEX
&& bld
.dispatch_width() == 8) {
3829 /* There's no plain shadow compare message, so we use shadow
3830 * compare with a bias of 0.0.
3832 bld
.MOV(msg_end
, brw_imm_f(0.0f
));
3833 msg_end
= offset(msg_end
, bld
, 1);
3836 bld
.MOV(msg_end
, shadow_c
);
3837 msg_end
= offset(msg_end
, bld
, 1);
3841 inst
->src
[0] = reg_undef
;
3842 inst
->src
[1] = surface
;
3843 inst
->src
[2] = sampler
;
3844 inst
->resize_sources(3);
3845 inst
->base_mrf
= msg_begin
.nr
;
3846 inst
->mlen
= msg_end
.nr
- msg_begin
.nr
;
3847 inst
->header_size
= 1;
3851 lower_sampler_logical_send_gen5(const fs_builder
&bld
, fs_inst
*inst
, opcode op
,
3852 const fs_reg
&coordinate
,
3853 const fs_reg
&shadow_c
,
3854 const fs_reg
&lod
, const fs_reg
&lod2
,
3855 const fs_reg
&sample_index
,
3856 const fs_reg
&surface
,
3857 const fs_reg
&sampler
,
3858 unsigned coord_components
,
3859 unsigned grad_components
)
3861 fs_reg
message(MRF
, 2, BRW_REGISTER_TYPE_F
);
3862 fs_reg msg_coords
= message
;
3863 unsigned header_size
= 0;
3865 if (inst
->offset
!= 0) {
3866 /* The offsets set up by the visitor are in the m1 header, so we can't
3873 for (unsigned i
= 0; i
< coord_components
; i
++)
3874 bld
.MOV(retype(offset(msg_coords
, bld
, i
), coordinate
.type
),
3875 offset(coordinate
, bld
, i
));
3877 fs_reg msg_end
= offset(msg_coords
, bld
, coord_components
);
3878 fs_reg msg_lod
= offset(msg_coords
, bld
, 4);
3880 if (shadow_c
.file
!= BAD_FILE
) {
3881 fs_reg msg_shadow
= msg_lod
;
3882 bld
.MOV(msg_shadow
, shadow_c
);
3883 msg_lod
= offset(msg_shadow
, bld
, 1);
3888 case SHADER_OPCODE_TXL
:
3890 bld
.MOV(msg_lod
, lod
);
3891 msg_end
= offset(msg_lod
, bld
, 1);
3893 case SHADER_OPCODE_TXD
:
3896 * dPdx = dudx, dvdx, drdx
3897 * dPdy = dudy, dvdy, drdy
3899 * Load up these values:
3900 * - dudx dudy dvdx dvdy drdx drdy
3901 * - dPdx.x dPdy.x dPdx.y dPdy.y dPdx.z dPdy.z
3904 for (unsigned i
= 0; i
< grad_components
; i
++) {
3905 bld
.MOV(msg_end
, offset(lod
, bld
, i
));
3906 msg_end
= offset(msg_end
, bld
, 1);
3908 bld
.MOV(msg_end
, offset(lod2
, bld
, i
));
3909 msg_end
= offset(msg_end
, bld
, 1);
3912 case SHADER_OPCODE_TXS
:
3913 msg_lod
= retype(msg_end
, BRW_REGISTER_TYPE_UD
);
3914 bld
.MOV(msg_lod
, lod
);
3915 msg_end
= offset(msg_lod
, bld
, 1);
3917 case SHADER_OPCODE_TXF
:
3918 msg_lod
= offset(msg_coords
, bld
, 3);
3919 bld
.MOV(retype(msg_lod
, BRW_REGISTER_TYPE_UD
), lod
);
3920 msg_end
= offset(msg_lod
, bld
, 1);
3922 case SHADER_OPCODE_TXF_CMS
:
3923 msg_lod
= offset(msg_coords
, bld
, 3);
3925 bld
.MOV(retype(msg_lod
, BRW_REGISTER_TYPE_UD
), brw_imm_ud(0u));
3927 bld
.MOV(retype(offset(msg_lod
, bld
, 1), BRW_REGISTER_TYPE_UD
), sample_index
);
3928 msg_end
= offset(msg_lod
, bld
, 2);
3935 inst
->src
[0] = reg_undef
;
3936 inst
->src
[1] = surface
;
3937 inst
->src
[2] = sampler
;
3938 inst
->resize_sources(3);
3939 inst
->base_mrf
= message
.nr
;
3940 inst
->mlen
= msg_end
.nr
- message
.nr
;
3941 inst
->header_size
= header_size
;
3943 /* Message length > MAX_SAMPLER_MESSAGE_SIZE disallowed by hardware. */
3944 assert(inst
->mlen
<= MAX_SAMPLER_MESSAGE_SIZE
);
3948 is_high_sampler(const struct gen_device_info
*devinfo
, const fs_reg
&sampler
)
3950 if (devinfo
->gen
< 8 && !devinfo
->is_haswell
)
3953 return sampler
.file
!= IMM
|| sampler
.ud
>= 16;
3957 lower_sampler_logical_send_gen7(const fs_builder
&bld
, fs_inst
*inst
, opcode op
,
3958 const fs_reg
&coordinate
,
3959 const fs_reg
&shadow_c
,
3960 fs_reg lod
, const fs_reg
&lod2
,
3961 const fs_reg
&sample_index
,
3963 const fs_reg
&surface
,
3964 const fs_reg
&sampler
,
3965 const fs_reg
&tg4_offset
,
3966 unsigned coord_components
,
3967 unsigned grad_components
)
3969 const gen_device_info
*devinfo
= bld
.shader
->devinfo
;
3970 unsigned reg_width
= bld
.dispatch_width() / 8;
3971 unsigned header_size
= 0, length
= 0;
3972 fs_reg sources
[MAX_SAMPLER_MESSAGE_SIZE
];
3973 for (unsigned i
= 0; i
< ARRAY_SIZE(sources
); i
++)
3974 sources
[i
] = bld
.vgrf(BRW_REGISTER_TYPE_F
);
3976 if (op
== SHADER_OPCODE_TG4
|| op
== SHADER_OPCODE_TG4_OFFSET
||
3977 inst
->offset
!= 0 || inst
->eot
||
3978 op
== SHADER_OPCODE_SAMPLEINFO
||
3979 is_high_sampler(devinfo
, sampler
)) {
3980 /* For general texture offsets (no txf workaround), we need a header to
3981 * put them in. Note that we're only reserving space for it in the
3982 * message payload as it will be initialized implicitly by the
3985 * TG4 needs to place its channel select in the header, for interaction
3986 * with ARB_texture_swizzle. The sampler index is only 4-bits, so for
3987 * larger sampler numbers we need to offset the Sampler State Pointer in
3991 sources
[0] = fs_reg();
3994 /* If we're requesting fewer than four channels worth of response,
3995 * and we have an explicit header, we need to set up the sampler
3996 * writemask. It's reversed from normal: 1 means "don't write".
3998 if (!inst
->eot
&& regs_written(inst
) != 4 * reg_width
) {
3999 assert(regs_written(inst
) % reg_width
== 0);
4000 unsigned mask
= ~((1 << (regs_written(inst
) / reg_width
)) - 1) & 0xf;
4001 inst
->offset
|= mask
<< 12;
4005 if (shadow_c
.file
!= BAD_FILE
) {
4006 bld
.MOV(sources
[length
], shadow_c
);
4010 bool coordinate_done
= false;
4012 /* Set up the LOD info */
4015 case SHADER_OPCODE_TXL
:
4016 if (devinfo
->gen
>= 9 && op
== SHADER_OPCODE_TXL
&& lod
.is_zero()) {
4017 op
= SHADER_OPCODE_TXL_LZ
;
4020 bld
.MOV(sources
[length
], lod
);
4023 case SHADER_OPCODE_TXD
:
4024 /* TXD should have been lowered in SIMD16 mode. */
4025 assert(bld
.dispatch_width() == 8);
4027 /* Load dPdx and the coordinate together:
4028 * [hdr], [ref], x, dPdx.x, dPdy.x, y, dPdx.y, dPdy.y, z, dPdx.z, dPdy.z
4030 for (unsigned i
= 0; i
< coord_components
; i
++) {
4031 bld
.MOV(sources
[length
++], offset(coordinate
, bld
, i
));
4033 /* For cube map array, the coordinate is (u,v,r,ai) but there are
4034 * only derivatives for (u, v, r).
4036 if (i
< grad_components
) {
4037 bld
.MOV(sources
[length
++], offset(lod
, bld
, i
));
4038 bld
.MOV(sources
[length
++], offset(lod2
, bld
, i
));
4042 coordinate_done
= true;
4044 case SHADER_OPCODE_TXS
:
4045 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
), lod
);
4048 case SHADER_OPCODE_TXF
:
4049 /* Unfortunately, the parameters for LD are intermixed: u, lod, v, r.
4050 * On Gen9 they are u, v, lod, r
4052 bld
.MOV(retype(sources
[length
++], BRW_REGISTER_TYPE_D
), coordinate
);
4054 if (devinfo
->gen
>= 9) {
4055 if (coord_components
>= 2) {
4056 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
),
4057 offset(coordinate
, bld
, 1));
4059 sources
[length
] = brw_imm_d(0);
4064 if (devinfo
->gen
>= 9 && lod
.is_zero()) {
4065 op
= SHADER_OPCODE_TXF_LZ
;
4067 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
), lod
);
4071 for (unsigned i
= devinfo
->gen
>= 9 ? 2 : 1; i
< coord_components
; i
++)
4072 bld
.MOV(retype(sources
[length
++], BRW_REGISTER_TYPE_D
),
4073 offset(coordinate
, bld
, i
));
4075 coordinate_done
= true;
4078 case SHADER_OPCODE_TXF_CMS
:
4079 case SHADER_OPCODE_TXF_CMS_W
:
4080 case SHADER_OPCODE_TXF_UMS
:
4081 case SHADER_OPCODE_TXF_MCS
:
4082 if (op
== SHADER_OPCODE_TXF_UMS
||
4083 op
== SHADER_OPCODE_TXF_CMS
||
4084 op
== SHADER_OPCODE_TXF_CMS_W
) {
4085 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
), sample_index
);
4089 if (op
== SHADER_OPCODE_TXF_CMS
|| op
== SHADER_OPCODE_TXF_CMS_W
) {
4090 /* Data from the multisample control surface. */
4091 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
), mcs
);
4094 /* On Gen9+ we'll use ld2dms_w instead which has two registers for
4097 if (op
== SHADER_OPCODE_TXF_CMS_W
) {
4098 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
),
4101 offset(mcs
, bld
, 1));
4106 /* There is no offsetting for this message; just copy in the integer
4107 * texture coordinates.
4109 for (unsigned i
= 0; i
< coord_components
; i
++)
4110 bld
.MOV(retype(sources
[length
++], BRW_REGISTER_TYPE_D
),
4111 offset(coordinate
, bld
, i
));
4113 coordinate_done
= true;
4115 case SHADER_OPCODE_TG4_OFFSET
:
4116 /* More crazy intermixing */
4117 for (unsigned i
= 0; i
< 2; i
++) /* u, v */
4118 bld
.MOV(sources
[length
++], offset(coordinate
, bld
, i
));
4120 for (unsigned i
= 0; i
< 2; i
++) /* offu, offv */
4121 bld
.MOV(retype(sources
[length
++], BRW_REGISTER_TYPE_D
),
4122 offset(tg4_offset
, bld
, i
));
4124 if (coord_components
== 3) /* r if present */
4125 bld
.MOV(sources
[length
++], offset(coordinate
, bld
, 2));
4127 coordinate_done
= true;
4133 /* Set up the coordinate (except for cases where it was done above) */
4134 if (!coordinate_done
) {
4135 for (unsigned i
= 0; i
< coord_components
; i
++)
4136 bld
.MOV(sources
[length
++], offset(coordinate
, bld
, i
));
4141 mlen
= length
* reg_width
- header_size
;
4143 mlen
= length
* reg_width
;
4145 const fs_reg src_payload
= fs_reg(VGRF
, bld
.shader
->alloc
.allocate(mlen
),
4146 BRW_REGISTER_TYPE_F
);
4147 bld
.LOAD_PAYLOAD(src_payload
, sources
, length
, header_size
);
4149 /* Generate the SEND. */
4151 inst
->src
[0] = src_payload
;
4152 inst
->src
[1] = surface
;
4153 inst
->src
[2] = sampler
;
4154 inst
->resize_sources(3);
4156 inst
->header_size
= header_size
;
4158 /* Message length > MAX_SAMPLER_MESSAGE_SIZE disallowed by hardware. */
4159 assert(inst
->mlen
<= MAX_SAMPLER_MESSAGE_SIZE
);
4163 lower_sampler_logical_send(const fs_builder
&bld
, fs_inst
*inst
, opcode op
)
4165 const gen_device_info
*devinfo
= bld
.shader
->devinfo
;
4166 const fs_reg
&coordinate
= inst
->src
[TEX_LOGICAL_SRC_COORDINATE
];
4167 const fs_reg
&shadow_c
= inst
->src
[TEX_LOGICAL_SRC_SHADOW_C
];
4168 const fs_reg
&lod
= inst
->src
[TEX_LOGICAL_SRC_LOD
];
4169 const fs_reg
&lod2
= inst
->src
[TEX_LOGICAL_SRC_LOD2
];
4170 const fs_reg
&sample_index
= inst
->src
[TEX_LOGICAL_SRC_SAMPLE_INDEX
];
4171 const fs_reg
&mcs
= inst
->src
[TEX_LOGICAL_SRC_MCS
];
4172 const fs_reg
&surface
= inst
->src
[TEX_LOGICAL_SRC_SURFACE
];
4173 const fs_reg
&sampler
= inst
->src
[TEX_LOGICAL_SRC_SAMPLER
];
4174 const fs_reg
&tg4_offset
= inst
->src
[TEX_LOGICAL_SRC_TG4_OFFSET
];
4175 assert(inst
->src
[TEX_LOGICAL_SRC_COORD_COMPONENTS
].file
== IMM
);
4176 const unsigned coord_components
= inst
->src
[TEX_LOGICAL_SRC_COORD_COMPONENTS
].ud
;
4177 assert(inst
->src
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
].file
== IMM
);
4178 const unsigned grad_components
= inst
->src
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
].ud
;
4180 if (devinfo
->gen
>= 7) {
4181 lower_sampler_logical_send_gen7(bld
, inst
, op
, coordinate
,
4182 shadow_c
, lod
, lod2
, sample_index
,
4183 mcs
, surface
, sampler
, tg4_offset
,
4184 coord_components
, grad_components
);
4185 } else if (devinfo
->gen
>= 5) {
4186 lower_sampler_logical_send_gen5(bld
, inst
, op
, coordinate
,
4187 shadow_c
, lod
, lod2
, sample_index
,
4189 coord_components
, grad_components
);
4191 lower_sampler_logical_send_gen4(bld
, inst
, op
, coordinate
,
4192 shadow_c
, lod
, lod2
,
4194 coord_components
, grad_components
);
4199 * Initialize the header present in some typed and untyped surface
4203 emit_surface_header(const fs_builder
&bld
, const fs_reg
&sample_mask
)
4205 fs_builder ubld
= bld
.exec_all().group(8, 0);
4206 const fs_reg dst
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
4207 ubld
.MOV(dst
, brw_imm_d(0));
4208 ubld
.MOV(component(dst
, 7), sample_mask
);
4213 lower_surface_logical_send(const fs_builder
&bld
, fs_inst
*inst
, opcode op
,
4214 const fs_reg
&sample_mask
)
4216 /* Get the logical send arguments. */
4217 const fs_reg
&addr
= inst
->src
[0];
4218 const fs_reg
&src
= inst
->src
[1];
4219 const fs_reg
&surface
= inst
->src
[2];
4220 const UNUSED fs_reg
&dims
= inst
->src
[3];
4221 const fs_reg
&arg
= inst
->src
[4];
4223 /* Calculate the total number of components of the payload. */
4224 const unsigned addr_sz
= inst
->components_read(0);
4225 const unsigned src_sz
= inst
->components_read(1);
4226 const unsigned header_sz
= (sample_mask
.file
== BAD_FILE
? 0 : 1);
4227 const unsigned sz
= header_sz
+ addr_sz
+ src_sz
;
4229 /* Allocate space for the payload. */
4230 fs_reg
*const components
= new fs_reg
[sz
];
4231 const fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, sz
);
4234 /* Construct the payload. */
4236 components
[n
++] = emit_surface_header(bld
, sample_mask
);
4238 for (unsigned i
= 0; i
< addr_sz
; i
++)
4239 components
[n
++] = offset(addr
, bld
, i
);
4241 for (unsigned i
= 0; i
< src_sz
; i
++)
4242 components
[n
++] = offset(src
, bld
, i
);
4244 bld
.LOAD_PAYLOAD(payload
, components
, sz
, header_sz
);
4246 /* Update the original instruction. */
4248 inst
->mlen
= header_sz
+ (addr_sz
+ src_sz
) * inst
->exec_size
/ 8;
4249 inst
->header_size
= header_sz
;
4251 inst
->src
[0] = payload
;
4252 inst
->src
[1] = surface
;
4254 inst
->resize_sources(3);
4256 delete[] components
;
4260 lower_varying_pull_constant_logical_send(const fs_builder
&bld
, fs_inst
*inst
)
4262 const gen_device_info
*devinfo
= bld
.shader
->devinfo
;
4264 if (devinfo
->gen
>= 7) {
4265 /* We are switching the instruction from an ALU-like instruction to a
4266 * send-from-grf instruction. Since sends can't handle strides or
4267 * source modifiers, we have to make a copy of the offset source.
4269 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
4270 bld
.MOV(tmp
, inst
->src
[1]);
4273 inst
->opcode
= FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7
;
4276 const fs_reg
payload(MRF
, FIRST_PULL_LOAD_MRF(devinfo
->gen
),
4277 BRW_REGISTER_TYPE_UD
);
4279 bld
.MOV(byte_offset(payload
, REG_SIZE
), inst
->src
[1]);
4281 inst
->opcode
= FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN4
;
4282 inst
->resize_sources(1);
4283 inst
->base_mrf
= payload
.nr
;
4284 inst
->header_size
= 1;
4285 inst
->mlen
= 1 + inst
->exec_size
/ 8;
4290 lower_math_logical_send(const fs_builder
&bld
, fs_inst
*inst
)
4292 assert(bld
.shader
->devinfo
->gen
< 6);
4295 inst
->mlen
= inst
->sources
* inst
->exec_size
/ 8;
4297 if (inst
->sources
> 1) {
4298 /* From the Ironlake PRM, Volume 4, Part 1, Section 6.1.13
4299 * "Message Payload":
4301 * "Operand0[7]. For the INT DIV functions, this operand is the
4304 * "Operand1[7]. For the INT DIV functions, this operand is the
4307 const bool is_int_div
= inst
->opcode
!= SHADER_OPCODE_POW
;
4308 const fs_reg src0
= is_int_div
? inst
->src
[1] : inst
->src
[0];
4309 const fs_reg src1
= is_int_div
? inst
->src
[0] : inst
->src
[1];
4311 inst
->resize_sources(1);
4312 inst
->src
[0] = src0
;
4314 assert(inst
->exec_size
== 8);
4315 bld
.MOV(fs_reg(MRF
, inst
->base_mrf
+ 1, src1
.type
), src1
);
4320 fs_visitor::lower_logical_sends()
4322 bool progress
= false;
4324 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
4325 const fs_builder
ibld(this, block
, inst
);
4327 switch (inst
->opcode
) {
4328 case FS_OPCODE_FB_WRITE_LOGICAL
:
4329 assert(stage
== MESA_SHADER_FRAGMENT
);
4330 lower_fb_write_logical_send(ibld
, inst
,
4331 brw_wm_prog_data(prog_data
),
4332 (const brw_wm_prog_key
*)key
,
4336 case FS_OPCODE_FB_READ_LOGICAL
:
4337 lower_fb_read_logical_send(ibld
, inst
);
4340 case SHADER_OPCODE_TEX_LOGICAL
:
4341 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TEX
);
4344 case SHADER_OPCODE_TXD_LOGICAL
:
4345 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXD
);
4348 case SHADER_OPCODE_TXF_LOGICAL
:
4349 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF
);
4352 case SHADER_OPCODE_TXL_LOGICAL
:
4353 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXL
);
4356 case SHADER_OPCODE_TXS_LOGICAL
:
4357 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXS
);
4360 case FS_OPCODE_TXB_LOGICAL
:
4361 lower_sampler_logical_send(ibld
, inst
, FS_OPCODE_TXB
);
4364 case SHADER_OPCODE_TXF_CMS_LOGICAL
:
4365 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_CMS
);
4368 case SHADER_OPCODE_TXF_CMS_W_LOGICAL
:
4369 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_CMS_W
);
4372 case SHADER_OPCODE_TXF_UMS_LOGICAL
:
4373 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_UMS
);
4376 case SHADER_OPCODE_TXF_MCS_LOGICAL
:
4377 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_MCS
);
4380 case SHADER_OPCODE_LOD_LOGICAL
:
4381 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_LOD
);
4384 case SHADER_OPCODE_TG4_LOGICAL
:
4385 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TG4
);
4388 case SHADER_OPCODE_TG4_OFFSET_LOGICAL
:
4389 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TG4_OFFSET
);
4392 case SHADER_OPCODE_SAMPLEINFO_LOGICAL
:
4393 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_SAMPLEINFO
);
4396 case SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
:
4397 lower_surface_logical_send(ibld
, inst
,
4398 SHADER_OPCODE_UNTYPED_SURFACE_READ
,
4402 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
:
4403 lower_surface_logical_send(ibld
, inst
,
4404 SHADER_OPCODE_UNTYPED_SURFACE_WRITE
,
4405 ibld
.sample_mask_reg());
4408 case SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL
:
4409 lower_surface_logical_send(ibld
, inst
,
4410 SHADER_OPCODE_UNTYPED_ATOMIC
,
4411 ibld
.sample_mask_reg());
4414 case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
:
4415 lower_surface_logical_send(ibld
, inst
,
4416 SHADER_OPCODE_TYPED_SURFACE_READ
,
4420 case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
:
4421 lower_surface_logical_send(ibld
, inst
,
4422 SHADER_OPCODE_TYPED_SURFACE_WRITE
,
4423 ibld
.sample_mask_reg());
4426 case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
:
4427 lower_surface_logical_send(ibld
, inst
,
4428 SHADER_OPCODE_TYPED_ATOMIC
,
4429 ibld
.sample_mask_reg());
4432 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_LOGICAL
:
4433 lower_varying_pull_constant_logical_send(ibld
, inst
);
4436 case SHADER_OPCODE_RCP
:
4437 case SHADER_OPCODE_RSQ
:
4438 case SHADER_OPCODE_SQRT
:
4439 case SHADER_OPCODE_EXP2
:
4440 case SHADER_OPCODE_LOG2
:
4441 case SHADER_OPCODE_SIN
:
4442 case SHADER_OPCODE_COS
:
4443 case SHADER_OPCODE_POW
:
4444 case SHADER_OPCODE_INT_QUOTIENT
:
4445 case SHADER_OPCODE_INT_REMAINDER
:
4446 /* The math opcodes are overloaded for the send-like and
4447 * expression-like instructions which seems kind of icky. Gen6+ has
4448 * a native (but rather quirky) MATH instruction so we don't need to
4449 * do anything here. On Gen4-5 we'll have to lower the Gen6-like
4450 * logical instructions (which we can easily recognize because they
4451 * have mlen = 0) into send-like virtual instructions.
4453 if (devinfo
->gen
< 6 && inst
->mlen
== 0) {
4454 lower_math_logical_send(ibld
, inst
);
4469 invalidate_live_intervals();
4475 * Get the closest allowed SIMD width for instruction \p inst accounting for
4476 * some common regioning and execution control restrictions that apply to FPU
4477 * instructions. These restrictions don't necessarily have any relevance to
4478 * instructions not executed by the FPU pipeline like extended math, control
4479 * flow or send message instructions.
4481 * For virtual opcodes it's really up to the instruction -- In some cases
4482 * (e.g. where a virtual instruction unrolls into a simple sequence of FPU
4483 * instructions) it may simplify virtual instruction lowering if we can
4484 * enforce FPU-like regioning restrictions already on the virtual instruction,
4485 * in other cases (e.g. virtual send-like instructions) this may be
4486 * excessively restrictive.
4489 get_fpu_lowered_simd_width(const struct gen_device_info
*devinfo
,
4490 const fs_inst
*inst
)
4492 /* Maximum execution size representable in the instruction controls. */
4493 unsigned max_width
= MIN2(32, inst
->exec_size
);
4495 /* According to the PRMs:
4496 * "A. In Direct Addressing mode, a source cannot span more than 2
4497 * adjacent GRF registers.
4498 * B. A destination cannot span more than 2 adjacent GRF registers."
4500 * Look for the source or destination with the largest register region
4501 * which is the one that is going to limit the overall execution size of
4502 * the instruction due to this rule.
4504 unsigned reg_count
= DIV_ROUND_UP(inst
->size_written
, REG_SIZE
);
4506 for (unsigned i
= 0; i
< inst
->sources
; i
++)
4507 reg_count
= MAX2(reg_count
, DIV_ROUND_UP(inst
->size_read(i
), REG_SIZE
));
4509 /* Calculate the maximum execution size of the instruction based on the
4510 * factor by which it goes over the hardware limit of 2 GRFs.
4513 max_width
= MIN2(max_width
, inst
->exec_size
/ DIV_ROUND_UP(reg_count
, 2));
4515 /* According to the IVB PRMs:
4516 * "When destination spans two registers, the source MUST span two
4517 * registers. The exception to the above rule:
4519 * - When source is scalar, the source registers are not incremented.
4520 * - When source is packed integer Word and destination is packed
4521 * integer DWord, the source register is not incremented but the
4522 * source sub register is incremented."
4524 * The hardware specs from Gen4 to Gen7.5 mention similar regioning
4525 * restrictions. The code below intentionally doesn't check whether the
4526 * destination type is integer because empirically the hardware doesn't
4527 * seem to care what the actual type is as long as it's dword-aligned.
4529 if (devinfo
->gen
< 8) {
4530 for (unsigned i
= 0; i
< inst
->sources
; i
++) {
4531 /* IVB implements DF scalars as <0;2,1> regions. */
4532 const bool is_scalar_exception
= is_uniform(inst
->src
[i
]) &&
4533 (devinfo
->is_haswell
|| type_sz(inst
->src
[i
].type
) != 8);
4534 const bool is_packed_word_exception
=
4535 type_sz(inst
->dst
.type
) == 4 && inst
->dst
.stride
== 1 &&
4536 type_sz(inst
->src
[i
].type
) == 2 && inst
->src
[i
].stride
== 1;
4538 if (inst
->size_written
> REG_SIZE
&&
4539 inst
->size_read(i
) != 0 && inst
->size_read(i
) <= REG_SIZE
&&
4540 !is_scalar_exception
&& !is_packed_word_exception
) {
4541 const unsigned reg_count
= DIV_ROUND_UP(inst
->size_written
, REG_SIZE
);
4542 max_width
= MIN2(max_width
, inst
->exec_size
/ reg_count
);
4547 /* From the IVB PRMs:
4548 * "When an instruction is SIMD32, the low 16 bits of the execution mask
4549 * are applied for both halves of the SIMD32 instruction. If different
4550 * execution mask channels are required, split the instruction into two
4551 * SIMD16 instructions."
4553 * There is similar text in the HSW PRMs. Gen4-6 don't even implement
4554 * 32-wide control flow support in hardware and will behave similarly.
4556 if (devinfo
->gen
< 8 && !inst
->force_writemask_all
)
4557 max_width
= MIN2(max_width
, 16);
4559 /* From the IVB PRMs (applies to HSW too):
4560 * "Instructions with condition modifiers must not use SIMD32."
4562 * From the BDW PRMs (applies to later hardware too):
4563 * "Ternary instruction with condition modifiers must not use SIMD32."
4565 if (inst
->conditional_mod
&& (devinfo
->gen
< 8 || inst
->is_3src(devinfo
)))
4566 max_width
= MIN2(max_width
, 16);
4568 /* From the IVB PRMs (applies to other devices that don't have the
4569 * gen_device_info::supports_simd16_3src flag set):
4570 * "In Align16 access mode, SIMD16 is not allowed for DW operations and
4571 * SIMD8 is not allowed for DF operations."
4573 if (inst
->is_3src(devinfo
) && !devinfo
->supports_simd16_3src
)
4574 max_width
= MIN2(max_width
, inst
->exec_size
/ reg_count
);
4576 /* Pre-Gen8 EUs are hardwired to use the QtrCtrl+1 (where QtrCtrl is
4577 * the 8-bit quarter of the execution mask signals specified in the
4578 * instruction control fields) for the second compressed half of any
4579 * single-precision instruction (for double-precision instructions
4580 * it's hardwired to use NibCtrl+1, at least on HSW), which means that
4581 * the EU will apply the wrong execution controls for the second
4582 * sequential GRF write if the number of channels per GRF is not exactly
4583 * eight in single-precision mode (or four in double-float mode).
4585 * In this situation we calculate the maximum size of the split
4586 * instructions so they only ever write to a single register.
4588 if (devinfo
->gen
< 8 && inst
->size_written
> REG_SIZE
&&
4589 !inst
->force_writemask_all
) {
4590 const unsigned channels_per_grf
= inst
->exec_size
/
4591 DIV_ROUND_UP(inst
->size_written
, REG_SIZE
);
4592 const unsigned exec_type_size
= get_exec_type_size(inst
);
4593 assert(exec_type_size
);
4595 /* The hardware shifts exactly 8 channels per compressed half of the
4596 * instruction in single-precision mode and exactly 4 in double-precision.
4598 if (channels_per_grf
!= (exec_type_size
== 8 ? 4 : 8))
4599 max_width
= MIN2(max_width
, channels_per_grf
);
4601 /* Lower all non-force_writemask_all DF instructions to SIMD4 on IVB/BYT
4602 * because HW applies the same channel enable signals to both halves of
4603 * the compressed instruction which will be just wrong under
4604 * non-uniform control flow.
4606 if (devinfo
->gen
== 7 && !devinfo
->is_haswell
&&
4607 (exec_type_size
== 8 || type_sz(inst
->dst
.type
) == 8))
4608 max_width
= MIN2(max_width
, 4);
4611 /* Only power-of-two execution sizes are representable in the instruction
4614 return 1 << _mesa_logbase2(max_width
);
4618 * Get the maximum allowed SIMD width for instruction \p inst accounting for
4619 * various payload size restrictions that apply to sampler message
4622 * This is only intended to provide a maximum theoretical bound for the
4623 * execution size of the message based on the number of argument components
4624 * alone, which in most cases will determine whether the SIMD8 or SIMD16
4625 * variant of the message can be used, though some messages may have
4626 * additional restrictions not accounted for here (e.g. pre-ILK hardware uses
4627 * the message length to determine the exact SIMD width and argument count,
4628 * which makes a number of sampler message combinations impossible to
4632 get_sampler_lowered_simd_width(const struct gen_device_info
*devinfo
,
4633 const fs_inst
*inst
)
4635 /* Calculate the number of coordinate components that have to be present
4636 * assuming that additional arguments follow the texel coordinates in the
4637 * message payload. On IVB+ there is no need for padding, on ILK-SNB we
4638 * need to pad to four or three components depending on the message,
4639 * pre-ILK we need to pad to at most three components.
4641 const unsigned req_coord_components
=
4642 (devinfo
->gen
>= 7 ||
4643 !inst
->components_read(TEX_LOGICAL_SRC_COORDINATE
)) ? 0 :
4644 (devinfo
->gen
>= 5 && inst
->opcode
!= SHADER_OPCODE_TXF_LOGICAL
&&
4645 inst
->opcode
!= SHADER_OPCODE_TXF_CMS_LOGICAL
) ? 4 :
4648 /* On Gen9+ the LOD argument is for free if we're able to use the LZ
4649 * variant of the TXL or TXF message.
4651 const bool implicit_lod
= devinfo
->gen
>= 9 &&
4652 (inst
->opcode
== SHADER_OPCODE_TXL
||
4653 inst
->opcode
== SHADER_OPCODE_TXF
) &&
4654 inst
->src
[TEX_LOGICAL_SRC_LOD
].is_zero();
4656 /* Calculate the total number of argument components that need to be passed
4657 * to the sampler unit.
4659 const unsigned num_payload_components
=
4660 MAX2(inst
->components_read(TEX_LOGICAL_SRC_COORDINATE
),
4661 req_coord_components
) +
4662 inst
->components_read(TEX_LOGICAL_SRC_SHADOW_C
) +
4663 (implicit_lod
? 0 : inst
->components_read(TEX_LOGICAL_SRC_LOD
)) +
4664 inst
->components_read(TEX_LOGICAL_SRC_LOD2
) +
4665 inst
->components_read(TEX_LOGICAL_SRC_SAMPLE_INDEX
) +
4666 (inst
->opcode
== SHADER_OPCODE_TG4_OFFSET_LOGICAL
?
4667 inst
->components_read(TEX_LOGICAL_SRC_TG4_OFFSET
) : 0) +
4668 inst
->components_read(TEX_LOGICAL_SRC_MCS
);
4670 /* SIMD16 messages with more than five arguments exceed the maximum message
4671 * size supported by the sampler, regardless of whether a header is
4674 return MIN2(inst
->exec_size
,
4675 num_payload_components
> MAX_SAMPLER_MESSAGE_SIZE
/ 2 ? 8 : 16);
4679 * Get the closest native SIMD width supported by the hardware for instruction
4680 * \p inst. The instruction will be left untouched by
4681 * fs_visitor::lower_simd_width() if the returned value is equal to the
4682 * original execution size.
4685 get_lowered_simd_width(const struct gen_device_info
*devinfo
,
4686 const fs_inst
*inst
)
4688 switch (inst
->opcode
) {
4689 case BRW_OPCODE_MOV
:
4690 case BRW_OPCODE_SEL
:
4691 case BRW_OPCODE_NOT
:
4692 case BRW_OPCODE_AND
:
4694 case BRW_OPCODE_XOR
:
4695 case BRW_OPCODE_SHR
:
4696 case BRW_OPCODE_SHL
:
4697 case BRW_OPCODE_ASR
:
4698 case BRW_OPCODE_CMPN
:
4699 case BRW_OPCODE_CSEL
:
4700 case BRW_OPCODE_F32TO16
:
4701 case BRW_OPCODE_F16TO32
:
4702 case BRW_OPCODE_BFREV
:
4703 case BRW_OPCODE_BFE
:
4704 case BRW_OPCODE_ADD
:
4705 case BRW_OPCODE_MUL
:
4706 case BRW_OPCODE_AVG
:
4707 case BRW_OPCODE_FRC
:
4708 case BRW_OPCODE_RNDU
:
4709 case BRW_OPCODE_RNDD
:
4710 case BRW_OPCODE_RNDE
:
4711 case BRW_OPCODE_RNDZ
:
4712 case BRW_OPCODE_LZD
:
4713 case BRW_OPCODE_FBH
:
4714 case BRW_OPCODE_FBL
:
4715 case BRW_OPCODE_CBIT
:
4716 case BRW_OPCODE_SAD2
:
4717 case BRW_OPCODE_MAD
:
4718 case BRW_OPCODE_LRP
:
4719 case FS_OPCODE_PACK
:
4720 return get_fpu_lowered_simd_width(devinfo
, inst
);
4722 case BRW_OPCODE_CMP
: {
4723 /* The Ivybridge/BayTrail WaCMPInstFlagDepClearedEarly workaround says that
4724 * when the destination is a GRF the dependency-clear bit on the flag
4725 * register is cleared early.
4727 * Suggested workarounds are to disable coissuing CMP instructions
4728 * or to split CMP(16) instructions into two CMP(8) instructions.
4730 * We choose to split into CMP(8) instructions since disabling
4731 * coissuing would affect CMP instructions not otherwise affected by
4734 const unsigned max_width
= (devinfo
->gen
== 7 && !devinfo
->is_haswell
&&
4735 !inst
->dst
.is_null() ? 8 : ~0);
4736 return MIN2(max_width
, get_fpu_lowered_simd_width(devinfo
, inst
));
4738 case BRW_OPCODE_BFI1
:
4739 case BRW_OPCODE_BFI2
:
4740 /* The Haswell WaForceSIMD8ForBFIInstruction workaround says that we
4742 * "Force BFI instructions to be executed always in SIMD8."
4744 return MIN2(devinfo
->is_haswell
? 8 : ~0u,
4745 get_fpu_lowered_simd_width(devinfo
, inst
));
4748 assert(inst
->src
[0].file
== BAD_FILE
|| inst
->exec_size
<= 16);
4749 return inst
->exec_size
;
4751 case SHADER_OPCODE_RCP
:
4752 case SHADER_OPCODE_RSQ
:
4753 case SHADER_OPCODE_SQRT
:
4754 case SHADER_OPCODE_EXP2
:
4755 case SHADER_OPCODE_LOG2
:
4756 case SHADER_OPCODE_SIN
:
4757 case SHADER_OPCODE_COS
:
4758 /* Unary extended math instructions are limited to SIMD8 on Gen4 and
4761 return (devinfo
->gen
>= 7 ? MIN2(16, inst
->exec_size
) :
4762 devinfo
->gen
== 5 || devinfo
->is_g4x
? MIN2(16, inst
->exec_size
) :
4763 MIN2(8, inst
->exec_size
));
4765 case SHADER_OPCODE_POW
:
4766 /* SIMD16 is only allowed on Gen7+. */
4767 return (devinfo
->gen
>= 7 ? MIN2(16, inst
->exec_size
) :
4768 MIN2(8, inst
->exec_size
));
4770 case SHADER_OPCODE_INT_QUOTIENT
:
4771 case SHADER_OPCODE_INT_REMAINDER
:
4772 /* Integer division is limited to SIMD8 on all generations. */
4773 return MIN2(8, inst
->exec_size
);
4775 case FS_OPCODE_LINTERP
:
4776 case FS_OPCODE_GET_BUFFER_SIZE
:
4777 case FS_OPCODE_DDX_COARSE
:
4778 case FS_OPCODE_DDX_FINE
:
4779 case FS_OPCODE_DDY_COARSE
:
4780 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
4781 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7
:
4782 case FS_OPCODE_PACK_HALF_2x16_SPLIT
:
4783 case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X
:
4784 case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y
:
4785 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
4786 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
4787 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
4788 return MIN2(16, inst
->exec_size
);
4790 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_LOGICAL
:
4791 /* Pre-ILK hardware doesn't have a SIMD8 variant of the texel fetch
4792 * message used to implement varying pull constant loads, so expand it
4793 * to SIMD16. An alternative with longer message payload length but
4794 * shorter return payload would be to use the SIMD8 sampler message that
4795 * takes (header, u, v, r) as parameters instead of (header, u).
4797 return (devinfo
->gen
== 4 ? 16 : MIN2(16, inst
->exec_size
));
4799 case FS_OPCODE_DDY_FINE
:
4800 /* The implementation of this virtual opcode may require emitting
4801 * compressed Align16 instructions, which are severely limited on some
4804 * From the Ivy Bridge PRM, volume 4 part 3, section 3.3.9 (Register
4805 * Region Restrictions):
4807 * "In Align16 access mode, SIMD16 is not allowed for DW operations
4808 * and SIMD8 is not allowed for DF operations."
4810 * In this context, "DW operations" means "operations acting on 32-bit
4811 * values", so it includes operations on floats.
4813 * Gen4 has a similar restriction. From the i965 PRM, section 11.5.3
4814 * (Instruction Compression -> Rules and Restrictions):
4816 * "A compressed instruction must be in Align1 access mode. Align16
4817 * mode instructions cannot be compressed."
4819 * Similar text exists in the g45 PRM.
4821 * Empirically, compressed align16 instructions using odd register
4822 * numbers don't appear to work on Sandybridge either.
4824 return (devinfo
->gen
== 4 || devinfo
->gen
== 6 ||
4825 (devinfo
->gen
== 7 && !devinfo
->is_haswell
) ?
4826 MIN2(8, inst
->exec_size
) : MIN2(16, inst
->exec_size
));
4828 case SHADER_OPCODE_MULH
:
4829 /* MULH is lowered to the MUL/MACH sequence using the accumulator, which
4830 * is 8-wide on Gen7+.
4832 return (devinfo
->gen
>= 7 ? 8 :
4833 get_fpu_lowered_simd_width(devinfo
, inst
));
4835 case FS_OPCODE_FB_WRITE_LOGICAL
:
4836 /* Gen6 doesn't support SIMD16 depth writes but we cannot handle them
4839 assert(devinfo
->gen
!= 6 ||
4840 inst
->src
[FB_WRITE_LOGICAL_SRC_SRC_DEPTH
].file
== BAD_FILE
||
4841 inst
->exec_size
== 8);
4842 /* Dual-source FB writes are unsupported in SIMD16 mode. */
4843 return (inst
->src
[FB_WRITE_LOGICAL_SRC_COLOR1
].file
!= BAD_FILE
?
4844 8 : MIN2(16, inst
->exec_size
));
4846 case FS_OPCODE_FB_READ_LOGICAL
:
4847 return MIN2(16, inst
->exec_size
);
4849 case SHADER_OPCODE_TEX_LOGICAL
:
4850 case SHADER_OPCODE_TXF_CMS_LOGICAL
:
4851 case SHADER_OPCODE_TXF_UMS_LOGICAL
:
4852 case SHADER_OPCODE_TXF_MCS_LOGICAL
:
4853 case SHADER_OPCODE_LOD_LOGICAL
:
4854 case SHADER_OPCODE_TG4_LOGICAL
:
4855 case SHADER_OPCODE_SAMPLEINFO_LOGICAL
:
4856 case SHADER_OPCODE_TXF_CMS_W_LOGICAL
:
4857 case SHADER_OPCODE_TG4_OFFSET_LOGICAL
:
4858 return get_sampler_lowered_simd_width(devinfo
, inst
);
4860 case SHADER_OPCODE_TXD_LOGICAL
:
4861 /* TXD is unsupported in SIMD16 mode. */
4864 case SHADER_OPCODE_TXL_LOGICAL
:
4865 case FS_OPCODE_TXB_LOGICAL
:
4866 /* Only one execution size is representable pre-ILK depending on whether
4867 * the shadow reference argument is present.
4869 if (devinfo
->gen
== 4)
4870 return inst
->src
[TEX_LOGICAL_SRC_SHADOW_C
].file
== BAD_FILE
? 16 : 8;
4872 return get_sampler_lowered_simd_width(devinfo
, inst
);
4874 case SHADER_OPCODE_TXF_LOGICAL
:
4875 case SHADER_OPCODE_TXS_LOGICAL
:
4876 /* Gen4 doesn't have SIMD8 variants for the RESINFO and LD-with-LOD
4877 * messages. Use SIMD16 instead.
4879 if (devinfo
->gen
== 4)
4882 return get_sampler_lowered_simd_width(devinfo
, inst
);
4884 case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
:
4885 case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
:
4886 case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
:
4889 case SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL
:
4890 case SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
:
4891 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
:
4892 return MIN2(16, inst
->exec_size
);
4894 case SHADER_OPCODE_URB_READ_SIMD8
:
4895 case SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
:
4896 case SHADER_OPCODE_URB_WRITE_SIMD8
:
4897 case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
4898 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
:
4899 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
4900 return MIN2(8, inst
->exec_size
);
4902 case SHADER_OPCODE_MOV_INDIRECT
: {
4903 /* From IVB and HSW PRMs:
4905 * "2.When the destination requires two registers and the sources are
4906 * indirect, the sources must use 1x1 regioning mode.
4908 * In case of DF instructions in HSW/IVB, the exec_size is limited by
4909 * the EU decompression logic not handling VxH indirect addressing
4912 const unsigned max_size
= (devinfo
->gen
>= 8 ? 2 : 1) * REG_SIZE
;
4913 /* Prior to Broadwell, we only have 8 address subregisters. */
4914 return MIN3(devinfo
->gen
>= 8 ? 16 : 8,
4915 max_size
/ (inst
->dst
.stride
* type_sz(inst
->dst
.type
)),
4919 case SHADER_OPCODE_LOAD_PAYLOAD
: {
4920 const unsigned reg_count
=
4921 DIV_ROUND_UP(inst
->dst
.component_size(inst
->exec_size
), REG_SIZE
);
4923 if (reg_count
> 2) {
4924 /* Only LOAD_PAYLOAD instructions with per-channel destination region
4925 * can be easily lowered (which excludes headers and heterogeneous
4928 assert(!inst
->header_size
);
4929 for (unsigned i
= 0; i
< inst
->sources
; i
++)
4930 assert(type_sz(inst
->dst
.type
) == type_sz(inst
->src
[i
].type
) ||
4931 inst
->src
[i
].file
== BAD_FILE
);
4933 return inst
->exec_size
/ DIV_ROUND_UP(reg_count
, 2);
4935 return inst
->exec_size
;
4939 return inst
->exec_size
;
4944 * Return true if splitting out the group of channels of instruction \p inst
4945 * given by lbld.group() requires allocating a temporary for the i-th source
4946 * of the lowered instruction.
4949 needs_src_copy(const fs_builder
&lbld
, const fs_inst
*inst
, unsigned i
)
4951 return !(is_periodic(inst
->src
[i
], lbld
.dispatch_width()) ||
4952 (inst
->components_read(i
) == 1 &&
4953 lbld
.dispatch_width() <= inst
->exec_size
));
4957 * Extract the data that would be consumed by the channel group given by
4958 * lbld.group() from the i-th source region of instruction \p inst and return
4959 * it as result in packed form. If any copy instructions are required they
4960 * will be emitted before the given \p inst in \p block.
4963 emit_unzip(const fs_builder
&lbld
, bblock_t
*block
, fs_inst
*inst
,
4966 /* Specified channel group from the source region. */
4967 const fs_reg src
= horiz_offset(inst
->src
[i
], lbld
.group());
4969 if (needs_src_copy(lbld
, inst
, i
)) {
4970 /* Builder of the right width to perform the copy avoiding uninitialized
4971 * data if the lowered execution size is greater than the original
4972 * execution size of the instruction.
4974 const fs_builder cbld
= lbld
.group(MIN2(lbld
.dispatch_width(),
4975 inst
->exec_size
), 0);
4976 const fs_reg tmp
= lbld
.vgrf(inst
->src
[i
].type
, inst
->components_read(i
));
4978 for (unsigned k
= 0; k
< inst
->components_read(i
); ++k
)
4979 cbld
.at(block
, inst
)
4980 .MOV(offset(tmp
, lbld
, k
), offset(src
, inst
->exec_size
, k
));
4984 } else if (is_periodic(inst
->src
[i
], lbld
.dispatch_width())) {
4985 /* The source is invariant for all dispatch_width-wide groups of the
4988 return inst
->src
[i
];
4991 /* We can just point the lowered instruction at the right channel group
4992 * from the original region.
4999 * Return true if splitting out the group of channels of instruction \p inst
5000 * given by lbld.group() requires allocating a temporary for the destination
5001 * of the lowered instruction and copying the data back to the original
5002 * destination region.
5005 needs_dst_copy(const fs_builder
&lbld
, const fs_inst
*inst
)
5007 /* If the instruction writes more than one component we'll have to shuffle
5008 * the results of multiple lowered instructions in order to make sure that
5009 * they end up arranged correctly in the original destination region.
5011 if (inst
->size_written
> inst
->dst
.component_size(inst
->exec_size
))
5014 /* If the lowered execution size is larger than the original the result of
5015 * the instruction won't fit in the original destination, so we'll have to
5016 * allocate a temporary in any case.
5018 if (lbld
.dispatch_width() > inst
->exec_size
)
5021 for (unsigned i
= 0; i
< inst
->sources
; i
++) {
5022 /* If we already made a copy of the source for other reasons there won't
5023 * be any overlap with the destination.
5025 if (needs_src_copy(lbld
, inst
, i
))
5028 /* In order to keep the logic simple we emit a copy whenever the
5029 * destination region doesn't exactly match an overlapping source, which
5030 * may point at the source and destination not being aligned group by
5031 * group which could cause one of the lowered instructions to overwrite
5032 * the data read from the same source by other lowered instructions.
5034 if (regions_overlap(inst
->dst
, inst
->size_written
,
5035 inst
->src
[i
], inst
->size_read(i
)) &&
5036 !inst
->dst
.equals(inst
->src
[i
]))
5044 * Insert data from a packed temporary into the channel group given by
5045 * lbld.group() of the destination region of instruction \p inst and return
5046 * the temporary as result. If any copy instructions are required they will
5047 * be emitted around the given \p inst in \p block.
5050 emit_zip(const fs_builder
&lbld
, bblock_t
*block
, fs_inst
*inst
)
5052 /* Builder of the right width to perform the copy avoiding uninitialized
5053 * data if the lowered execution size is greater than the original
5054 * execution size of the instruction.
5056 const fs_builder cbld
= lbld
.group(MIN2(lbld
.dispatch_width(),
5057 inst
->exec_size
), 0);
5059 /* Specified channel group from the destination region. */
5060 const fs_reg dst
= horiz_offset(inst
->dst
, lbld
.group());
5061 const unsigned dst_size
= inst
->size_written
/
5062 inst
->dst
.component_size(inst
->exec_size
);
5064 if (needs_dst_copy(lbld
, inst
)) {
5065 const fs_reg tmp
= lbld
.vgrf(inst
->dst
.type
, dst_size
);
5067 if (inst
->predicate
) {
5068 /* Handle predication by copying the original contents of
5069 * the destination into the temporary before emitting the
5070 * lowered instruction.
5072 for (unsigned k
= 0; k
< dst_size
; ++k
)
5073 cbld
.at(block
, inst
)
5074 .MOV(offset(tmp
, lbld
, k
), offset(dst
, inst
->exec_size
, k
));
5077 for (unsigned k
= 0; k
< dst_size
; ++k
)
5078 cbld
.at(block
, inst
->next
)
5079 .MOV(offset(dst
, inst
->exec_size
, k
), offset(tmp
, lbld
, k
));
5084 /* No need to allocate a temporary for the lowered instruction, just
5085 * take the right group of channels from the original region.
5092 fs_visitor::lower_simd_width()
5094 bool progress
= false;
5096 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
5097 const unsigned lower_width
= get_lowered_simd_width(devinfo
, inst
);
5099 if (lower_width
!= inst
->exec_size
) {
5100 /* Builder matching the original instruction. We may also need to
5101 * emit an instruction of width larger than the original, set the
5102 * execution size of the builder to the highest of both for now so
5103 * we're sure that both cases can be handled.
5105 const unsigned max_width
= MAX2(inst
->exec_size
, lower_width
);
5106 const fs_builder ibld
= bld
.at(block
, inst
)
5107 .exec_all(inst
->force_writemask_all
)
5108 .group(max_width
, inst
->group
/ max_width
);
5110 /* Split the copies in chunks of the execution width of either the
5111 * original or the lowered instruction, whichever is lower.
5113 const unsigned n
= DIV_ROUND_UP(inst
->exec_size
, lower_width
);
5114 const unsigned dst_size
= inst
->size_written
/
5115 inst
->dst
.component_size(inst
->exec_size
);
5117 assert(!inst
->writes_accumulator
&& !inst
->mlen
);
5119 for (unsigned i
= 0; i
< n
; i
++) {
5120 /* Emit a copy of the original instruction with the lowered width.
5121 * If the EOT flag was set throw it away except for the last
5122 * instruction to avoid killing the thread prematurely.
5124 fs_inst split_inst
= *inst
;
5125 split_inst
.exec_size
= lower_width
;
5126 split_inst
.eot
= inst
->eot
&& i
== n
- 1;
5128 /* Select the correct channel enables for the i-th group, then
5129 * transform the sources and destination and emit the lowered
5132 const fs_builder lbld
= ibld
.group(lower_width
, i
);
5134 for (unsigned j
= 0; j
< inst
->sources
; j
++)
5135 split_inst
.src
[j
] = emit_unzip(lbld
, block
, inst
, j
);
5137 split_inst
.dst
= emit_zip(lbld
, block
, inst
);
5138 split_inst
.size_written
=
5139 split_inst
.dst
.component_size(lower_width
) * dst_size
;
5141 lbld
.emit(split_inst
);
5144 inst
->remove(block
);
5150 invalidate_live_intervals();
5156 fs_visitor::dump_instructions()
5158 dump_instructions(NULL
);
5162 fs_visitor::dump_instructions(const char *name
)
5164 FILE *file
= stderr
;
5165 if (name
&& geteuid() != 0) {
5166 file
= fopen(name
, "w");
5172 calculate_register_pressure();
5173 int ip
= 0, max_pressure
= 0;
5174 foreach_block_and_inst(block
, backend_instruction
, inst
, cfg
) {
5175 max_pressure
= MAX2(max_pressure
, regs_live_at_ip
[ip
]);
5176 fprintf(file
, "{%3d} %4d: ", regs_live_at_ip
[ip
], ip
);
5177 dump_instruction(inst
, file
);
5180 fprintf(file
, "Maximum %3d registers live at once.\n", max_pressure
);
5183 foreach_in_list(backend_instruction
, inst
, &instructions
) {
5184 fprintf(file
, "%4d: ", ip
++);
5185 dump_instruction(inst
, file
);
5189 if (file
!= stderr
) {
5195 fs_visitor::dump_instruction(backend_instruction
*be_inst
)
5197 dump_instruction(be_inst
, stderr
);
5201 fs_visitor::dump_instruction(backend_instruction
*be_inst
, FILE *file
)
5203 fs_inst
*inst
= (fs_inst
*)be_inst
;
5205 if (inst
->predicate
) {
5206 fprintf(file
, "(%cf0.%d) ",
5207 inst
->predicate_inverse
? '-' : '+',
5211 fprintf(file
, "%s", brw_instruction_name(devinfo
, inst
->opcode
));
5213 fprintf(file
, ".sat");
5214 if (inst
->conditional_mod
) {
5215 fprintf(file
, "%s", conditional_modifier
[inst
->conditional_mod
]);
5216 if (!inst
->predicate
&&
5217 (devinfo
->gen
< 5 || (inst
->opcode
!= BRW_OPCODE_SEL
&&
5218 inst
->opcode
!= BRW_OPCODE_IF
&&
5219 inst
->opcode
!= BRW_OPCODE_WHILE
))) {
5220 fprintf(file
, ".f0.%d", inst
->flag_subreg
);
5223 fprintf(file
, "(%d) ", inst
->exec_size
);
5226 fprintf(file
, "(mlen: %d) ", inst
->mlen
);
5230 fprintf(file
, "(EOT) ");
5233 switch (inst
->dst
.file
) {
5235 fprintf(file
, "vgrf%d", inst
->dst
.nr
);
5238 fprintf(file
, "g%d", inst
->dst
.nr
);
5241 fprintf(file
, "m%d", inst
->dst
.nr
);
5244 fprintf(file
, "(null)");
5247 fprintf(file
, "***u%d***", inst
->dst
.nr
);
5250 fprintf(file
, "***attr%d***", inst
->dst
.nr
);
5253 switch (inst
->dst
.nr
) {
5255 fprintf(file
, "null");
5257 case BRW_ARF_ADDRESS
:
5258 fprintf(file
, "a0.%d", inst
->dst
.subnr
);
5260 case BRW_ARF_ACCUMULATOR
:
5261 fprintf(file
, "acc%d", inst
->dst
.subnr
);
5264 fprintf(file
, "f%d.%d", inst
->dst
.nr
& 0xf, inst
->dst
.subnr
);
5267 fprintf(file
, "arf%d.%d", inst
->dst
.nr
& 0xf, inst
->dst
.subnr
);
5272 unreachable("not reached");
5275 if (inst
->dst
.offset
||
5276 (inst
->dst
.file
== VGRF
&&
5277 alloc
.sizes
[inst
->dst
.nr
] * REG_SIZE
!= inst
->size_written
)) {
5278 const unsigned reg_size
= (inst
->dst
.file
== UNIFORM
? 4 : REG_SIZE
);
5279 fprintf(file
, "+%d.%d", inst
->dst
.offset
/ reg_size
,
5280 inst
->dst
.offset
% reg_size
);
5283 if (inst
->dst
.stride
!= 1)
5284 fprintf(file
, "<%u>", inst
->dst
.stride
);
5285 fprintf(file
, ":%s, ", brw_reg_type_letters(inst
->dst
.type
));
5287 for (int i
= 0; i
< inst
->sources
; i
++) {
5288 if (inst
->src
[i
].negate
)
5290 if (inst
->src
[i
].abs
)
5292 switch (inst
->src
[i
].file
) {
5294 fprintf(file
, "vgrf%d", inst
->src
[i
].nr
);
5297 fprintf(file
, "g%d", inst
->src
[i
].nr
);
5300 fprintf(file
, "***m%d***", inst
->src
[i
].nr
);
5303 fprintf(file
, "attr%d", inst
->src
[i
].nr
);
5306 fprintf(file
, "u%d", inst
->src
[i
].nr
);
5309 fprintf(file
, "(null)");
5312 switch (inst
->src
[i
].type
) {
5313 case BRW_REGISTER_TYPE_F
:
5314 fprintf(file
, "%-gf", inst
->src
[i
].f
);
5316 case BRW_REGISTER_TYPE_DF
:
5317 fprintf(file
, "%fdf", inst
->src
[i
].df
);
5319 case BRW_REGISTER_TYPE_W
:
5320 case BRW_REGISTER_TYPE_D
:
5321 fprintf(file
, "%dd", inst
->src
[i
].d
);
5323 case BRW_REGISTER_TYPE_UW
:
5324 case BRW_REGISTER_TYPE_UD
:
5325 fprintf(file
, "%uu", inst
->src
[i
].ud
);
5327 case BRW_REGISTER_TYPE_VF
:
5328 fprintf(file
, "[%-gF, %-gF, %-gF, %-gF]",
5329 brw_vf_to_float((inst
->src
[i
].ud
>> 0) & 0xff),
5330 brw_vf_to_float((inst
->src
[i
].ud
>> 8) & 0xff),
5331 brw_vf_to_float((inst
->src
[i
].ud
>> 16) & 0xff),
5332 brw_vf_to_float((inst
->src
[i
].ud
>> 24) & 0xff));
5335 fprintf(file
, "???");
5340 switch (inst
->src
[i
].nr
) {
5342 fprintf(file
, "null");
5344 case BRW_ARF_ADDRESS
:
5345 fprintf(file
, "a0.%d", inst
->src
[i
].subnr
);
5347 case BRW_ARF_ACCUMULATOR
:
5348 fprintf(file
, "acc%d", inst
->src
[i
].subnr
);
5351 fprintf(file
, "f%d.%d", inst
->src
[i
].nr
& 0xf, inst
->src
[i
].subnr
);
5354 fprintf(file
, "arf%d.%d", inst
->src
[i
].nr
& 0xf, inst
->src
[i
].subnr
);
5360 if (inst
->src
[i
].offset
||
5361 (inst
->src
[i
].file
== VGRF
&&
5362 alloc
.sizes
[inst
->src
[i
].nr
] * REG_SIZE
!= inst
->size_read(i
))) {
5363 const unsigned reg_size
= (inst
->src
[i
].file
== UNIFORM
? 4 : REG_SIZE
);
5364 fprintf(file
, "+%d.%d", inst
->src
[i
].offset
/ reg_size
,
5365 inst
->src
[i
].offset
% reg_size
);
5368 if (inst
->src
[i
].abs
)
5371 if (inst
->src
[i
].file
!= IMM
) {
5373 if (inst
->src
[i
].file
== ARF
|| inst
->src
[i
].file
== FIXED_GRF
) {
5374 unsigned hstride
= inst
->src
[i
].hstride
;
5375 stride
= (hstride
== 0 ? 0 : (1 << (hstride
- 1)));
5377 stride
= inst
->src
[i
].stride
;
5380 fprintf(file
, "<%u>", stride
);
5382 fprintf(file
, ":%s", brw_reg_type_letters(inst
->src
[i
].type
));
5385 if (i
< inst
->sources
- 1 && inst
->src
[i
+ 1].file
!= BAD_FILE
)
5386 fprintf(file
, ", ");
5391 if (inst
->force_writemask_all
)
5392 fprintf(file
, "NoMask ");
5394 if (inst
->exec_size
!= dispatch_width
)
5395 fprintf(file
, "group%d ", inst
->group
);
5397 fprintf(file
, "\n");
5401 * Possibly returns an instruction that set up @param reg.
5403 * Sometimes we want to take the result of some expression/variable
5404 * dereference tree and rewrite the instruction generating the result
5405 * of the tree. When processing the tree, we know that the
5406 * instructions generated are all writing temporaries that are dead
5407 * outside of this tree. So, if we have some instructions that write
5408 * a temporary, we're free to point that temp write somewhere else.
5410 * Note that this doesn't guarantee that the instruction generated
5411 * only reg -- it might be the size=4 destination of a texture instruction.
5414 fs_visitor::get_instruction_generating_reg(fs_inst
*start
,
5419 end
->is_partial_write() ||
5420 !reg
.equals(end
->dst
)) {
5428 fs_visitor::setup_fs_payload_gen6()
5430 assert(stage
== MESA_SHADER_FRAGMENT
);
5431 struct brw_wm_prog_data
*prog_data
= brw_wm_prog_data(this->prog_data
);
5433 assert(devinfo
->gen
>= 6);
5435 /* R0-1: masks, pixel X/Y coordinates. */
5436 payload
.num_regs
= 2;
5437 /* R2: only for 32-pixel dispatch.*/
5439 /* R3-26: barycentric interpolation coordinates. These appear in the
5440 * same order that they appear in the brw_barycentric_mode
5441 * enum. Each set of coordinates occupies 2 registers if dispatch width
5442 * == 8 and 4 registers if dispatch width == 16. Coordinates only
5443 * appear if they were enabled using the "Barycentric Interpolation
5444 * Mode" bits in WM_STATE.
5446 for (int i
= 0; i
< BRW_BARYCENTRIC_MODE_COUNT
; ++i
) {
5447 if (prog_data
->barycentric_interp_modes
& (1 << i
)) {
5448 payload
.barycentric_coord_reg
[i
] = payload
.num_regs
;
5449 payload
.num_regs
+= 2;
5450 if (dispatch_width
== 16) {
5451 payload
.num_regs
+= 2;
5456 /* R27: interpolated depth if uses source depth */
5457 prog_data
->uses_src_depth
=
5458 (nir
->info
.inputs_read
& (1 << VARYING_SLOT_POS
)) != 0;
5459 if (prog_data
->uses_src_depth
) {
5460 payload
.source_depth_reg
= payload
.num_regs
;
5462 if (dispatch_width
== 16) {
5463 /* R28: interpolated depth if not SIMD8. */
5468 /* R29: interpolated W set if GEN6_WM_USES_SOURCE_W. */
5469 prog_data
->uses_src_w
=
5470 (nir
->info
.inputs_read
& (1 << VARYING_SLOT_POS
)) != 0;
5471 if (prog_data
->uses_src_w
) {
5472 payload
.source_w_reg
= payload
.num_regs
;
5474 if (dispatch_width
== 16) {
5475 /* R30: interpolated W if not SIMD8. */
5480 /* R31: MSAA position offsets. */
5481 if (prog_data
->persample_dispatch
&&
5482 (nir
->info
.system_values_read
& SYSTEM_BIT_SAMPLE_POS
)) {
5483 /* From the Ivy Bridge PRM documentation for 3DSTATE_PS:
5485 * "MSDISPMODE_PERSAMPLE is required in order to select
5488 * So we can only really get sample positions if we are doing real
5489 * per-sample dispatch. If we need gl_SamplePosition and we don't have
5490 * persample dispatch, we hard-code it to 0.5.
5492 prog_data
->uses_pos_offset
= true;
5493 payload
.sample_pos_reg
= payload
.num_regs
;
5497 /* R32: MSAA input coverage mask */
5498 prog_data
->uses_sample_mask
=
5499 (nir
->info
.system_values_read
& SYSTEM_BIT_SAMPLE_MASK_IN
) != 0;
5500 if (prog_data
->uses_sample_mask
) {
5501 assert(devinfo
->gen
>= 7);
5502 payload
.sample_mask_in_reg
= payload
.num_regs
;
5504 if (dispatch_width
== 16) {
5505 /* R33: input coverage mask if not SIMD8. */
5510 /* R34-: bary for 32-pixel. */
5511 /* R58-59: interp W for 32-pixel. */
5513 if (nir
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_DEPTH
)) {
5514 source_depth_to_render_target
= true;
5519 fs_visitor::setup_vs_payload()
5521 /* R0: thread header, R1: urb handles */
5522 payload
.num_regs
= 2;
5526 fs_visitor::setup_gs_payload()
5528 assert(stage
== MESA_SHADER_GEOMETRY
);
5530 struct brw_gs_prog_data
*gs_prog_data
= brw_gs_prog_data(prog_data
);
5531 struct brw_vue_prog_data
*vue_prog_data
= brw_vue_prog_data(prog_data
);
5533 /* R0: thread header, R1: output URB handles */
5534 payload
.num_regs
= 2;
5536 if (gs_prog_data
->include_primitive_id
) {
5537 /* R2: Primitive ID 0..7 */
5541 /* Use a maximum of 24 registers for push-model inputs. */
5542 const unsigned max_push_components
= 24;
5544 /* If pushing our inputs would take too many registers, reduce the URB read
5545 * length (which is in HWords, or 8 registers), and resort to pulling.
5547 * Note that the GS reads <URB Read Length> HWords for every vertex - so we
5548 * have to multiply by VerticesIn to obtain the total storage requirement.
5550 if (8 * vue_prog_data
->urb_read_length
* nir
->info
.gs
.vertices_in
>
5551 max_push_components
|| gs_prog_data
->invocations
> 1) {
5552 gs_prog_data
->base
.include_vue_handles
= true;
5554 /* R3..RN: ICP Handles for each incoming vertex (when using pull model) */
5555 payload
.num_regs
+= nir
->info
.gs
.vertices_in
;
5557 vue_prog_data
->urb_read_length
=
5558 ROUND_DOWN_TO(max_push_components
/ nir
->info
.gs
.vertices_in
, 8) / 8;
5563 fs_visitor::setup_cs_payload()
5565 assert(devinfo
->gen
>= 7);
5566 payload
.num_regs
= 1;
5570 fs_visitor::calculate_register_pressure()
5572 invalidate_live_intervals();
5573 calculate_live_intervals();
5575 unsigned num_instructions
= 0;
5576 foreach_block(block
, cfg
)
5577 num_instructions
+= block
->instructions
.length();
5579 regs_live_at_ip
= rzalloc_array(mem_ctx
, int, num_instructions
);
5581 for (unsigned reg
= 0; reg
< alloc
.count
; reg
++) {
5582 for (int ip
= virtual_grf_start
[reg
]; ip
<= virtual_grf_end
[reg
]; ip
++)
5583 regs_live_at_ip
[ip
] += alloc
.sizes
[reg
];
5588 * Look for repeated FS_OPCODE_MOV_DISPATCH_TO_FLAGS and drop the later ones.
5590 * The needs_unlit_centroid_workaround ends up producing one of these per
5591 * channel of centroid input, so it's good to clean them up.
5593 * An assumption here is that nothing ever modifies the dispatched pixels
5594 * value that FS_OPCODE_MOV_DISPATCH_TO_FLAGS reads from, but the hardware
5595 * dictates that anyway.
5598 fs_visitor::opt_drop_redundant_mov_to_flags()
5600 bool flag_mov_found
[2] = {false};
5601 bool progress
= false;
5603 /* Instructions removed by this pass can only be added if this were true */
5604 if (!devinfo
->needs_unlit_centroid_workaround
)
5607 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
5608 if (inst
->is_control_flow()) {
5609 memset(flag_mov_found
, 0, sizeof(flag_mov_found
));
5610 } else if (inst
->opcode
== FS_OPCODE_MOV_DISPATCH_TO_FLAGS
) {
5611 if (!flag_mov_found
[inst
->flag_subreg
]) {
5612 flag_mov_found
[inst
->flag_subreg
] = true;
5614 inst
->remove(block
);
5617 } else if (inst
->flags_written()) {
5618 flag_mov_found
[inst
->flag_subreg
] = false;
5626 fs_visitor::optimize()
5628 /* Start by validating the shader we currently have. */
5631 /* bld is the common builder object pointing at the end of the program we
5632 * used to translate it into i965 IR. For the optimization and lowering
5633 * passes coming next, any code added after the end of the program without
5634 * having explicitly called fs_builder::at() clearly points at a mistake.
5635 * Ideally optimization passes wouldn't be part of the visitor so they
5636 * wouldn't have access to bld at all, but they do, so just in case some
5637 * pass forgets to ask for a location explicitly set it to NULL here to
5638 * make it trip. The dispatch width is initialized to a bogus value to
5639 * make sure that optimizations set the execution controls explicitly to
5640 * match the code they are manipulating instead of relying on the defaults.
5642 bld
= fs_builder(this, 64);
5644 assign_constant_locations();
5645 lower_constant_loads();
5649 split_virtual_grfs();
5652 #define OPT(pass, args...) ({ \
5654 bool this_progress = pass(args); \
5656 if (unlikely(INTEL_DEBUG & DEBUG_OPTIMIZER) && this_progress) { \
5657 char filename[64]; \
5658 snprintf(filename, 64, "%s%d-%s-%02d-%02d-" #pass, \
5659 stage_abbrev, dispatch_width, nir->info.name, iteration, pass_num); \
5661 backend_shader::dump_instructions(filename); \
5666 progress = progress || this_progress; \
5670 if (unlikely(INTEL_DEBUG
& DEBUG_OPTIMIZER
)) {
5672 snprintf(filename
, 64, "%s%d-%s-00-00-start",
5673 stage_abbrev
, dispatch_width
, nir
->info
.name
);
5675 backend_shader::dump_instructions(filename
);
5678 bool progress
= false;
5682 OPT(opt_drop_redundant_mov_to_flags
);
5689 OPT(remove_duplicate_mrf_writes
);
5693 OPT(opt_copy_propagation
);
5694 OPT(opt_predicated_break
, this);
5695 OPT(opt_cmod_propagation
);
5696 OPT(dead_code_eliminate
);
5697 OPT(opt_peephole_sel
);
5698 OPT(dead_control_flow_eliminate
, this);
5699 OPT(opt_register_renaming
);
5700 OPT(opt_saturate_propagation
);
5701 OPT(register_coalesce
);
5702 OPT(compute_to_mrf
);
5703 OPT(eliminate_find_live_channel
);
5705 OPT(compact_virtual_grfs
);
5711 if (OPT(lower_pack
)) {
5712 OPT(register_coalesce
);
5713 OPT(dead_code_eliminate
);
5716 OPT(lower_simd_width
);
5718 /* After SIMD lowering just in case we had to unroll the EOT send. */
5719 OPT(opt_sampler_eot
);
5721 OPT(lower_logical_sends
);
5724 OPT(opt_copy_propagation
);
5725 /* Only run after logical send lowering because it's easier to implement
5726 * in terms of physical sends.
5728 if (OPT(opt_zero_samples
))
5729 OPT(opt_copy_propagation
);
5730 /* Run after logical send lowering to give it a chance to CSE the
5731 * LOAD_PAYLOAD instructions created to construct the payloads of
5732 * e.g. texturing messages in cases where it wasn't possible to CSE the
5733 * whole logical instruction.
5736 OPT(register_coalesce
);
5737 OPT(compute_to_mrf
);
5738 OPT(dead_code_eliminate
);
5739 OPT(remove_duplicate_mrf_writes
);
5740 OPT(opt_peephole_sel
);
5743 OPT(opt_redundant_discard_jumps
);
5745 if (OPT(lower_load_payload
)) {
5746 split_virtual_grfs();
5747 OPT(register_coalesce
);
5748 OPT(compute_to_mrf
);
5749 OPT(dead_code_eliminate
);
5752 OPT(opt_combine_constants
);
5753 OPT(lower_integer_multiplication
);
5755 if (devinfo
->gen
<= 5 && OPT(lower_minmax
)) {
5756 OPT(opt_cmod_propagation
);
5758 OPT(opt_copy_propagation
);
5759 OPT(dead_code_eliminate
);
5762 if (OPT(lower_conversions
)) {
5763 OPT(opt_copy_propagation
);
5764 OPT(dead_code_eliminate
);
5765 OPT(lower_simd_width
);
5768 lower_uniform_pull_constant_loads();
5774 * Three source instruction must have a GRF/MRF destination register.
5775 * ARF NULL is not allowed. Fix that up by allocating a temporary GRF.
5778 fs_visitor::fixup_3src_null_dest()
5780 bool progress
= false;
5782 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
5783 if (inst
->is_3src(devinfo
) && inst
->dst
.is_null()) {
5784 inst
->dst
= fs_reg(VGRF
, alloc
.allocate(dispatch_width
/ 8),
5791 invalidate_live_intervals();
5795 fs_visitor::allocate_registers(bool allow_spilling
)
5797 bool allocated_without_spills
;
5799 static const enum instruction_scheduler_mode pre_modes
[] = {
5801 SCHEDULE_PRE_NON_LIFO
,
5805 bool spill_all
= allow_spilling
&& (INTEL_DEBUG
& DEBUG_SPILL_FS
);
5807 /* Try each scheduling heuristic to see if it can successfully register
5808 * allocate without spilling. They should be ordered by decreasing
5809 * performance but increasing likelihood of allocating.
5811 for (unsigned i
= 0; i
< ARRAY_SIZE(pre_modes
); i
++) {
5812 schedule_instructions(pre_modes
[i
]);
5815 assign_regs_trivial();
5816 allocated_without_spills
= true;
5818 allocated_without_spills
= assign_regs(false, spill_all
);
5820 if (allocated_without_spills
)
5824 if (!allocated_without_spills
) {
5825 if (!allow_spilling
)
5826 fail("Failure to register allocate and spilling is not allowed.");
5828 /* We assume that any spilling is worse than just dropping back to
5829 * SIMD8. There's probably actually some intermediate point where
5830 * SIMD16 with a couple of spills is still better.
5832 if (dispatch_width
> min_dispatch_width
) {
5833 fail("Failure to register allocate. Reduce number of "
5834 "live scalar values to avoid this.");
5836 compiler
->shader_perf_log(log_data
,
5837 "%s shader triggered register spilling. "
5838 "Try reducing the number of live scalar "
5839 "values to improve performance.\n",
5843 /* Since we're out of heuristics, just go spill registers until we
5844 * get an allocation.
5846 while (!assign_regs(true, spill_all
)) {
5852 /* This must come after all optimization and register allocation, since
5853 * it inserts dead code that happens to have side effects, and it does
5854 * so based on the actual physical registers in use.
5856 insert_gen4_send_dependency_workarounds();
5861 schedule_instructions(SCHEDULE_POST
);
5863 if (last_scratch
> 0) {
5864 MAYBE_UNUSED
unsigned max_scratch_size
= 2 * 1024 * 1024;
5866 prog_data
->total_scratch
= brw_get_scratch_size(last_scratch
);
5868 if (stage
== MESA_SHADER_COMPUTE
) {
5869 if (devinfo
->is_haswell
) {
5870 /* According to the MEDIA_VFE_STATE's "Per Thread Scratch Space"
5871 * field documentation, Haswell supports a minimum of 2kB of
5872 * scratch space for compute shaders, unlike every other stage
5875 prog_data
->total_scratch
= MAX2(prog_data
->total_scratch
, 2048);
5876 } else if (devinfo
->gen
<= 7) {
5877 /* According to the MEDIA_VFE_STATE's "Per Thread Scratch Space"
5878 * field documentation, platforms prior to Haswell measure scratch
5879 * size linearly with a range of [1kB, 12kB] and 1kB granularity.
5881 prog_data
->total_scratch
= ALIGN(last_scratch
, 1024);
5882 max_scratch_size
= 12 * 1024;
5886 /* We currently only support up to 2MB of scratch space. If we
5887 * need to support more eventually, the documentation suggests
5888 * that we could allocate a larger buffer, and partition it out
5889 * ourselves. We'd just have to undo the hardware's address
5890 * calculation by subtracting (FFTID * Per Thread Scratch Space)
5891 * and then add FFTID * (Larger Per Thread Scratch Space).
5893 * See 3D-Media-GPGPU Engine > Media GPGPU Pipeline >
5894 * Thread Group Tracking > Local Memory/Scratch Space.
5896 assert(prog_data
->total_scratch
< max_scratch_size
);
5901 fs_visitor::run_vs(gl_clip_plane
*clip_planes
)
5903 assert(stage
== MESA_SHADER_VERTEX
);
5907 if (shader_time_index
>= 0)
5908 emit_shader_time_begin();
5915 compute_clip_distance(clip_planes
);
5919 if (shader_time_index
>= 0)
5920 emit_shader_time_end();
5926 assign_curb_setup();
5927 assign_vs_urb_setup();
5929 fixup_3src_null_dest();
5930 allocate_registers(true);
5936 fs_visitor::run_tcs_single_patch()
5938 assert(stage
== MESA_SHADER_TESS_CTRL
);
5940 struct brw_tcs_prog_data
*tcs_prog_data
= brw_tcs_prog_data(prog_data
);
5942 /* r1-r4 contain the ICP handles. */
5943 payload
.num_regs
= 5;
5945 if (shader_time_index
>= 0)
5946 emit_shader_time_begin();
5948 /* Initialize gl_InvocationID */
5949 fs_reg channels_uw
= bld
.vgrf(BRW_REGISTER_TYPE_UW
);
5950 fs_reg channels_ud
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
5951 bld
.MOV(channels_uw
, fs_reg(brw_imm_uv(0x76543210)));
5952 bld
.MOV(channels_ud
, channels_uw
);
5954 if (tcs_prog_data
->instances
== 1) {
5955 invocation_id
= channels_ud
;
5957 invocation_id
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
5959 /* Get instance number from g0.2 bits 23:17, and multiply it by 8. */
5960 fs_reg t
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
5961 fs_reg instance_times_8
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
5962 bld
.AND(t
, fs_reg(retype(brw_vec1_grf(0, 2), BRW_REGISTER_TYPE_UD
)),
5963 brw_imm_ud(INTEL_MASK(23, 17)));
5964 bld
.SHR(instance_times_8
, t
, brw_imm_ud(17 - 3));
5966 bld
.ADD(invocation_id
, instance_times_8
, channels_ud
);
5969 /* Fix the disptach mask */
5970 if (nir
->info
.tess
.tcs_vertices_out
% 8) {
5971 bld
.CMP(bld
.null_reg_ud(), invocation_id
,
5972 brw_imm_ud(nir
->info
.tess
.tcs_vertices_out
), BRW_CONDITIONAL_L
);
5973 bld
.IF(BRW_PREDICATE_NORMAL
);
5978 if (nir
->info
.tess
.tcs_vertices_out
% 8) {
5979 bld
.emit(BRW_OPCODE_ENDIF
);
5982 /* Emit EOT write; set TR DS Cache bit */
5984 fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
)),
5985 fs_reg(brw_imm_ud(WRITEMASK_X
<< 16)),
5986 fs_reg(brw_imm_ud(0)),
5988 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 3);
5989 bld
.LOAD_PAYLOAD(payload
, srcs
, 3, 2);
5991 fs_inst
*inst
= bld
.emit(SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
,
5992 bld
.null_reg_ud(), payload
);
5996 if (shader_time_index
>= 0)
5997 emit_shader_time_end();
6006 assign_curb_setup();
6007 assign_tcs_single_patch_urb_setup();
6009 fixup_3src_null_dest();
6010 allocate_registers(true);
6016 fs_visitor::run_tes()
6018 assert(stage
== MESA_SHADER_TESS_EVAL
);
6020 /* R0: thread header, R1-3: gl_TessCoord.xyz, R4: URB handles */
6021 payload
.num_regs
= 5;
6023 if (shader_time_index
>= 0)
6024 emit_shader_time_begin();
6033 if (shader_time_index
>= 0)
6034 emit_shader_time_end();
6040 assign_curb_setup();
6041 assign_tes_urb_setup();
6043 fixup_3src_null_dest();
6044 allocate_registers(true);
6050 fs_visitor::run_gs()
6052 assert(stage
== MESA_SHADER_GEOMETRY
);
6056 this->final_gs_vertex_count
= vgrf(glsl_type::uint_type
);
6058 if (gs_compile
->control_data_header_size_bits
> 0) {
6059 /* Create a VGRF to store accumulated control data bits. */
6060 this->control_data_bits
= vgrf(glsl_type::uint_type
);
6062 /* If we're outputting more than 32 control data bits, then EmitVertex()
6063 * will set control_data_bits to 0 after emitting the first vertex.
6064 * Otherwise, we need to initialize it to 0 here.
6066 if (gs_compile
->control_data_header_size_bits
<= 32) {
6067 const fs_builder abld
= bld
.annotate("initialize control data bits");
6068 abld
.MOV(this->control_data_bits
, brw_imm_ud(0u));
6072 if (shader_time_index
>= 0)
6073 emit_shader_time_begin();
6077 emit_gs_thread_end();
6079 if (shader_time_index
>= 0)
6080 emit_shader_time_end();
6089 assign_curb_setup();
6090 assign_gs_urb_setup();
6092 fixup_3src_null_dest();
6093 allocate_registers(true);
6099 fs_visitor::run_fs(bool allow_spilling
, bool do_rep_send
)
6101 struct brw_wm_prog_data
*wm_prog_data
= brw_wm_prog_data(this->prog_data
);
6102 brw_wm_prog_key
*wm_key
= (brw_wm_prog_key
*) this->key
;
6104 assert(stage
== MESA_SHADER_FRAGMENT
);
6106 if (devinfo
->gen
>= 6)
6107 setup_fs_payload_gen6();
6109 setup_fs_payload_gen4();
6113 } else if (do_rep_send
) {
6114 assert(dispatch_width
== 16);
6115 emit_repclear_shader();
6117 if (shader_time_index
>= 0)
6118 emit_shader_time_begin();
6120 calculate_urb_setup();
6121 if (nir
->info
.inputs_read
> 0 ||
6122 (nir
->info
.outputs_read
> 0 && !wm_key
->coherent_fb_fetch
)) {
6123 if (devinfo
->gen
< 6)
6124 emit_interpolation_setup_gen4();
6126 emit_interpolation_setup_gen6();
6129 /* We handle discards by keeping track of the still-live pixels in f0.1.
6130 * Initialize it with the dispatched pixels.
6132 if (wm_prog_data
->uses_kill
) {
6133 fs_inst
*discard_init
= bld
.emit(FS_OPCODE_MOV_DISPATCH_TO_FLAGS
);
6134 discard_init
->flag_subreg
= 1;
6137 /* Generate FS IR for main(). (the visitor only descends into
6138 * functions called "main").
6145 if (wm_prog_data
->uses_kill
)
6146 bld
.emit(FS_OPCODE_PLACEHOLDER_HALT
);
6148 if (wm_key
->alpha_test_func
)
6153 if (shader_time_index
>= 0)
6154 emit_shader_time_end();
6160 assign_curb_setup();
6163 fixup_3src_null_dest();
6164 allocate_registers(allow_spilling
);
6174 fs_visitor::run_cs()
6176 assert(stage
== MESA_SHADER_COMPUTE
);
6180 if (shader_time_index
>= 0)
6181 emit_shader_time_begin();
6183 if (devinfo
->is_haswell
&& prog_data
->total_shared
> 0) {
6184 /* Move SLM index from g0.0[27:24] to sr0.1[11:8] */
6185 const fs_builder abld
= bld
.exec_all().group(1, 0);
6186 abld
.MOV(retype(brw_sr0_reg(1), BRW_REGISTER_TYPE_UW
),
6187 suboffset(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW
), 1));
6195 emit_cs_terminate();
6197 if (shader_time_index
>= 0)
6198 emit_shader_time_end();
6204 assign_curb_setup();
6206 fixup_3src_null_dest();
6207 allocate_registers(true);
6216 * Return a bitfield where bit n is set if barycentric interpolation mode n
6217 * (see enum brw_barycentric_mode) is needed by the fragment shader.
6219 * We examine the load_barycentric intrinsics rather than looking at input
6220 * variables so that we catch interpolateAtCentroid() messages too, which
6221 * also need the BRW_BARYCENTRIC_[NON]PERSPECTIVE_CENTROID mode set up.
6224 brw_compute_barycentric_interp_modes(const struct gen_device_info
*devinfo
,
6225 const nir_shader
*shader
)
6227 unsigned barycentric_interp_modes
= 0;
6229 nir_foreach_function(f
, shader
) {
6233 nir_foreach_block(block
, f
->impl
) {
6234 nir_foreach_instr(instr
, block
) {
6235 if (instr
->type
!= nir_instr_type_intrinsic
)
6238 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
6239 if (intrin
->intrinsic
!= nir_intrinsic_load_interpolated_input
)
6242 /* Ignore WPOS; it doesn't require interpolation. */
6243 if (nir_intrinsic_base(intrin
) == VARYING_SLOT_POS
)
6246 intrin
= nir_instr_as_intrinsic(intrin
->src
[0].ssa
->parent_instr
);
6247 enum glsl_interp_mode interp
= (enum glsl_interp_mode
)
6248 nir_intrinsic_interp_mode(intrin
);
6249 nir_intrinsic_op bary_op
= intrin
->intrinsic
;
6250 enum brw_barycentric_mode bary
=
6251 brw_barycentric_mode(interp
, bary_op
);
6253 barycentric_interp_modes
|= 1 << bary
;
6255 if (devinfo
->needs_unlit_centroid_workaround
&&
6256 bary_op
== nir_intrinsic_load_barycentric_centroid
)
6257 barycentric_interp_modes
|= 1 << centroid_to_pixel(bary
);
6262 return barycentric_interp_modes
;
6266 brw_compute_flat_inputs(struct brw_wm_prog_data
*prog_data
,
6267 const nir_shader
*shader
)
6269 prog_data
->flat_inputs
= 0;
6271 nir_foreach_variable(var
, &shader
->inputs
) {
6272 int input_index
= prog_data
->urb_setup
[var
->data
.location
];
6274 if (input_index
< 0)
6278 if (var
->data
.interpolation
== INTERP_MODE_FLAT
)
6279 prog_data
->flat_inputs
|= (1 << input_index
);
6284 computed_depth_mode(const nir_shader
*shader
)
6286 if (shader
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_DEPTH
)) {
6287 switch (shader
->info
.fs
.depth_layout
) {
6288 case FRAG_DEPTH_LAYOUT_NONE
:
6289 case FRAG_DEPTH_LAYOUT_ANY
:
6290 return BRW_PSCDEPTH_ON
;
6291 case FRAG_DEPTH_LAYOUT_GREATER
:
6292 return BRW_PSCDEPTH_ON_GE
;
6293 case FRAG_DEPTH_LAYOUT_LESS
:
6294 return BRW_PSCDEPTH_ON_LE
;
6295 case FRAG_DEPTH_LAYOUT_UNCHANGED
:
6296 return BRW_PSCDEPTH_OFF
;
6299 return BRW_PSCDEPTH_OFF
;
6303 * Move load_interpolated_input with simple (payload-based) barycentric modes
6304 * to the top of the program so we don't emit multiple PLNs for the same input.
6306 * This works around CSE not being able to handle non-dominating cases
6312 * interpolate the same exact input
6315 * This should be replaced by global value numbering someday.
6318 move_interpolation_to_top(nir_shader
*nir
)
6320 bool progress
= false;
6322 nir_foreach_function(f
, nir
) {
6326 nir_block
*top
= nir_start_block(f
->impl
);
6327 exec_node
*cursor_node
= NULL
;
6329 nir_foreach_block(block
, f
->impl
) {
6333 nir_foreach_instr_safe(instr
, block
) {
6334 if (instr
->type
!= nir_instr_type_intrinsic
)
6337 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
6338 if (intrin
->intrinsic
!= nir_intrinsic_load_interpolated_input
)
6340 nir_intrinsic_instr
*bary_intrinsic
=
6341 nir_instr_as_intrinsic(intrin
->src
[0].ssa
->parent_instr
);
6342 nir_intrinsic_op op
= bary_intrinsic
->intrinsic
;
6344 /* Leave interpolateAtSample/Offset() where they are. */
6345 if (op
== nir_intrinsic_load_barycentric_at_sample
||
6346 op
== nir_intrinsic_load_barycentric_at_offset
)
6349 nir_instr
*move
[3] = {
6350 &bary_intrinsic
->instr
,
6351 intrin
->src
[1].ssa
->parent_instr
,
6355 for (unsigned i
= 0; i
< ARRAY_SIZE(move
); i
++) {
6356 if (move
[i
]->block
!= top
) {
6357 move
[i
]->block
= top
;
6358 exec_node_remove(&move
[i
]->node
);
6360 exec_node_insert_after(cursor_node
, &move
[i
]->node
);
6362 exec_list_push_head(&top
->instr_list
, &move
[i
]->node
);
6364 cursor_node
= &move
[i
]->node
;
6370 nir_metadata_preserve(f
->impl
, (nir_metadata
)
6371 ((unsigned) nir_metadata_block_index
|
6372 (unsigned) nir_metadata_dominance
));
6379 * Demote per-sample barycentric intrinsics to centroid.
6381 * Useful when rendering to a non-multisampled buffer.
6384 demote_sample_qualifiers(nir_shader
*nir
)
6386 bool progress
= true;
6388 nir_foreach_function(f
, nir
) {
6393 nir_builder_init(&b
, f
->impl
);
6395 nir_foreach_block(block
, f
->impl
) {
6396 nir_foreach_instr_safe(instr
, block
) {
6397 if (instr
->type
!= nir_instr_type_intrinsic
)
6400 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
6401 if (intrin
->intrinsic
!= nir_intrinsic_load_barycentric_sample
&&
6402 intrin
->intrinsic
!= nir_intrinsic_load_barycentric_at_sample
)
6405 b
.cursor
= nir_before_instr(instr
);
6406 nir_ssa_def
*centroid
=
6407 nir_load_barycentric(&b
, nir_intrinsic_load_barycentric_centroid
,
6408 nir_intrinsic_interp_mode(intrin
));
6409 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
6410 nir_src_for_ssa(centroid
));
6411 nir_instr_remove(instr
);
6416 nir_metadata_preserve(f
->impl
, (nir_metadata
)
6417 ((unsigned) nir_metadata_block_index
|
6418 (unsigned) nir_metadata_dominance
));
6425 * Pre-gen6, the register file of the EUs was shared between threads,
6426 * and each thread used some subset allocated on a 16-register block
6427 * granularity. The unit states wanted these block counts.
6430 brw_register_blocks(int reg_count
)
6432 return ALIGN(reg_count
, 16) / 16 - 1;
6436 brw_compile_fs(const struct brw_compiler
*compiler
, void *log_data
,
6438 const struct brw_wm_prog_key
*key
,
6439 struct brw_wm_prog_data
*prog_data
,
6440 const nir_shader
*src_shader
,
6441 struct gl_program
*prog
,
6442 int shader_time_index8
, int shader_time_index16
,
6443 bool allow_spilling
,
6444 bool use_rep_send
, struct brw_vue_map
*vue_map
,
6445 unsigned *final_assembly_size
,
6448 const struct gen_device_info
*devinfo
= compiler
->devinfo
;
6450 nir_shader
*shader
= nir_shader_clone(mem_ctx
, src_shader
);
6451 shader
= brw_nir_apply_sampler_key(shader
, compiler
, &key
->tex
, true);
6452 brw_nir_lower_fs_inputs(shader
, devinfo
, key
);
6453 brw_nir_lower_fs_outputs(shader
);
6455 if (devinfo
->gen
< 6) {
6456 brw_setup_vue_interpolation(vue_map
, shader
, prog_data
, devinfo
);
6459 if (!key
->multisample_fbo
)
6460 NIR_PASS_V(shader
, demote_sample_qualifiers
);
6461 NIR_PASS_V(shader
, move_interpolation_to_top
);
6462 shader
= brw_postprocess_nir(shader
, compiler
, true);
6464 /* key->alpha_test_func means simulating alpha testing via discards,
6465 * so the shader definitely kills pixels.
6467 prog_data
->uses_kill
= shader
->info
.fs
.uses_discard
||
6468 key
->alpha_test_func
;
6469 prog_data
->uses_omask
= key
->multisample_fbo
&&
6470 shader
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_SAMPLE_MASK
);
6471 prog_data
->computed_depth_mode
= computed_depth_mode(shader
);
6472 prog_data
->computed_stencil
=
6473 shader
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_STENCIL
);
6475 prog_data
->persample_dispatch
=
6476 key
->multisample_fbo
&&
6477 (key
->persample_interp
||
6478 (shader
->info
.system_values_read
& (SYSTEM_BIT_SAMPLE_ID
|
6479 SYSTEM_BIT_SAMPLE_POS
)) ||
6480 shader
->info
.fs
.uses_sample_qualifier
||
6481 shader
->info
.outputs_read
);
6483 prog_data
->early_fragment_tests
= shader
->info
.fs
.early_fragment_tests
;
6484 prog_data
->post_depth_coverage
= shader
->info
.fs
.post_depth_coverage
;
6485 prog_data
->inner_coverage
= shader
->info
.fs
.inner_coverage
;
6487 prog_data
->barycentric_interp_modes
=
6488 brw_compute_barycentric_interp_modes(compiler
->devinfo
, shader
);
6490 cfg_t
*simd8_cfg
= NULL
, *simd16_cfg
= NULL
;
6491 uint8_t simd8_grf_start
= 0, simd16_grf_start
= 0;
6492 unsigned simd8_grf_used
= 0, simd16_grf_used
= 0;
6494 fs_visitor
v8(compiler
, log_data
, mem_ctx
, key
,
6495 &prog_data
->base
, prog
, shader
, 8,
6496 shader_time_index8
);
6497 if (!v8
.run_fs(allow_spilling
, false /* do_rep_send */)) {
6499 *error_str
= ralloc_strdup(mem_ctx
, v8
.fail_msg
);
6502 } else if (likely(!(INTEL_DEBUG
& DEBUG_NO8
))) {
6504 simd8_grf_start
= v8
.payload
.num_regs
;
6505 simd8_grf_used
= v8
.grf_used
;
6508 if (v8
.max_dispatch_width
>= 16 &&
6509 likely(!(INTEL_DEBUG
& DEBUG_NO16
) || use_rep_send
)) {
6510 /* Try a SIMD16 compile */
6511 fs_visitor
v16(compiler
, log_data
, mem_ctx
, key
,
6512 &prog_data
->base
, prog
, shader
, 16,
6513 shader_time_index16
);
6514 v16
.import_uniforms(&v8
);
6515 if (!v16
.run_fs(allow_spilling
, use_rep_send
)) {
6516 compiler
->shader_perf_log(log_data
,
6517 "SIMD16 shader failed to compile: %s",
6520 simd16_cfg
= v16
.cfg
;
6521 simd16_grf_start
= v16
.payload
.num_regs
;
6522 simd16_grf_used
= v16
.grf_used
;
6526 /* When the caller requests a repclear shader, they want SIMD16-only */
6530 /* Prior to Iron Lake, the PS had a single shader offset with a jump table
6531 * at the top to select the shader. We've never implemented that.
6532 * Instead, we just give them exactly one shader and we pick the widest one
6535 if (compiler
->devinfo
->gen
< 5 && simd16_cfg
)
6538 if (prog_data
->persample_dispatch
) {
6539 /* Starting with SandyBridge (where we first get MSAA), the different
6540 * pixel dispatch combinations are grouped into classifications A
6541 * through F (SNB PRM Vol. 2 Part 1 Section 7.7.1). On all hardware
6542 * generations, the only configurations supporting persample dispatch
6543 * are are this in which only one dispatch width is enabled.
6545 * If computed depth is enabled, SNB only allows SIMD8 while IVB+
6546 * allow SIMD8 or SIMD16 so we choose SIMD16 if available.
6548 if (compiler
->devinfo
->gen
== 6 &&
6549 prog_data
->computed_depth_mode
!= BRW_PSCDEPTH_OFF
) {
6551 } else if (simd16_cfg
) {
6556 /* We have to compute the flat inputs after the visitor is finished running
6557 * because it relies on prog_data->urb_setup which is computed in
6558 * fs_visitor::calculate_urb_setup().
6560 brw_compute_flat_inputs(prog_data
, shader
);
6562 fs_generator
g(compiler
, log_data
, mem_ctx
, (void *) key
, &prog_data
->base
,
6563 v8
.promoted_constants
, v8
.runtime_check_aads_emit
,
6564 MESA_SHADER_FRAGMENT
);
6566 if (unlikely(INTEL_DEBUG
& DEBUG_WM
)) {
6567 g
.enable_debug(ralloc_asprintf(mem_ctx
, "%s fragment shader %s",
6568 shader
->info
.label
?
6569 shader
->info
.label
: "unnamed",
6570 shader
->info
.name
));
6574 prog_data
->dispatch_8
= true;
6575 g
.generate_code(simd8_cfg
, 8);
6576 prog_data
->base
.dispatch_grf_start_reg
= simd8_grf_start
;
6577 prog_data
->reg_blocks_0
= brw_register_blocks(simd8_grf_used
);
6580 prog_data
->dispatch_16
= true;
6581 prog_data
->prog_offset_2
= g
.generate_code(simd16_cfg
, 16);
6582 prog_data
->dispatch_grf_start_reg_2
= simd16_grf_start
;
6583 prog_data
->reg_blocks_2
= brw_register_blocks(simd16_grf_used
);
6585 } else if (simd16_cfg
) {
6586 prog_data
->dispatch_16
= true;
6587 g
.generate_code(simd16_cfg
, 16);
6588 prog_data
->base
.dispatch_grf_start_reg
= simd16_grf_start
;
6589 prog_data
->reg_blocks_0
= brw_register_blocks(simd16_grf_used
);
6592 return g
.get_assembly(final_assembly_size
);
6596 fs_visitor::emit_cs_work_group_id_setup()
6598 assert(stage
== MESA_SHADER_COMPUTE
);
6600 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::uvec3_type
));
6602 struct brw_reg
r0_1(retype(brw_vec1_grf(0, 1), BRW_REGISTER_TYPE_UD
));
6603 struct brw_reg
r0_6(retype(brw_vec1_grf(0, 6), BRW_REGISTER_TYPE_UD
));
6604 struct brw_reg
r0_7(retype(brw_vec1_grf(0, 7), BRW_REGISTER_TYPE_UD
));
6606 bld
.MOV(*reg
, r0_1
);
6607 bld
.MOV(offset(*reg
, bld
, 1), r0_6
);
6608 bld
.MOV(offset(*reg
, bld
, 2), r0_7
);
6614 fill_push_const_block_info(struct brw_push_const_block
*block
, unsigned dwords
)
6616 block
->dwords
= dwords
;
6617 block
->regs
= DIV_ROUND_UP(dwords
, 8);
6618 block
->size
= block
->regs
* 32;
6622 cs_fill_push_const_info(const struct gen_device_info
*devinfo
,
6623 struct brw_cs_prog_data
*cs_prog_data
)
6625 const struct brw_stage_prog_data
*prog_data
= &cs_prog_data
->base
;
6626 bool fill_thread_id
=
6627 cs_prog_data
->thread_local_id_index
>= 0 &&
6628 cs_prog_data
->thread_local_id_index
< (int)prog_data
->nr_params
;
6629 bool cross_thread_supported
= devinfo
->gen
> 7 || devinfo
->is_haswell
;
6631 /* The thread ID should be stored in the last param dword */
6632 assert(prog_data
->nr_params
> 0 || !fill_thread_id
);
6633 assert(!fill_thread_id
||
6634 cs_prog_data
->thread_local_id_index
==
6635 (int)prog_data
->nr_params
- 1);
6637 unsigned cross_thread_dwords
, per_thread_dwords
;
6638 if (!cross_thread_supported
) {
6639 cross_thread_dwords
= 0u;
6640 per_thread_dwords
= prog_data
->nr_params
;
6641 } else if (fill_thread_id
) {
6642 /* Fill all but the last register with cross-thread payload */
6643 cross_thread_dwords
= 8 * (cs_prog_data
->thread_local_id_index
/ 8);
6644 per_thread_dwords
= prog_data
->nr_params
- cross_thread_dwords
;
6645 assert(per_thread_dwords
> 0 && per_thread_dwords
<= 8);
6647 /* Fill all data using cross-thread payload */
6648 cross_thread_dwords
= prog_data
->nr_params
;
6649 per_thread_dwords
= 0u;
6652 fill_push_const_block_info(&cs_prog_data
->push
.cross_thread
, cross_thread_dwords
);
6653 fill_push_const_block_info(&cs_prog_data
->push
.per_thread
, per_thread_dwords
);
6655 unsigned total_dwords
=
6656 (cs_prog_data
->push
.per_thread
.size
* cs_prog_data
->threads
+
6657 cs_prog_data
->push
.cross_thread
.size
) / 4;
6658 fill_push_const_block_info(&cs_prog_data
->push
.total
, total_dwords
);
6660 assert(cs_prog_data
->push
.cross_thread
.dwords
% 8 == 0 ||
6661 cs_prog_data
->push
.per_thread
.size
== 0);
6662 assert(cs_prog_data
->push
.cross_thread
.dwords
+
6663 cs_prog_data
->push
.per_thread
.dwords
==
6664 prog_data
->nr_params
);
6668 cs_set_simd_size(struct brw_cs_prog_data
*cs_prog_data
, unsigned size
)
6670 cs_prog_data
->simd_size
= size
;
6671 unsigned group_size
= cs_prog_data
->local_size
[0] *
6672 cs_prog_data
->local_size
[1] * cs_prog_data
->local_size
[2];
6673 cs_prog_data
->threads
= (group_size
+ size
- 1) / size
;
6677 brw_compile_cs(const struct brw_compiler
*compiler
, void *log_data
,
6679 const struct brw_cs_prog_key
*key
,
6680 struct brw_cs_prog_data
*prog_data
,
6681 const nir_shader
*src_shader
,
6682 int shader_time_index
,
6683 unsigned *final_assembly_size
,
6686 nir_shader
*shader
= nir_shader_clone(mem_ctx
, src_shader
);
6687 shader
= brw_nir_apply_sampler_key(shader
, compiler
, &key
->tex
, true);
6688 brw_nir_lower_cs_shared(shader
);
6689 prog_data
->base
.total_shared
+= shader
->num_shared
;
6691 /* Now that we cloned the nir_shader, we can update num_uniforms based on
6692 * the thread_local_id_index.
6694 assert(prog_data
->thread_local_id_index
>= 0);
6695 shader
->num_uniforms
=
6696 MAX2(shader
->num_uniforms
,
6697 (unsigned)4 * (prog_data
->thread_local_id_index
+ 1));
6699 brw_nir_lower_intrinsics(shader
, &prog_data
->base
);
6700 shader
= brw_postprocess_nir(shader
, compiler
, true);
6702 prog_data
->local_size
[0] = shader
->info
.cs
.local_size
[0];
6703 prog_data
->local_size
[1] = shader
->info
.cs
.local_size
[1];
6704 prog_data
->local_size
[2] = shader
->info
.cs
.local_size
[2];
6705 unsigned local_workgroup_size
=
6706 shader
->info
.cs
.local_size
[0] * shader
->info
.cs
.local_size
[1] *
6707 shader
->info
.cs
.local_size
[2];
6709 unsigned max_cs_threads
= compiler
->devinfo
->max_cs_threads
;
6710 unsigned simd_required
= DIV_ROUND_UP(local_workgroup_size
, max_cs_threads
);
6713 const char *fail_msg
= NULL
;
6715 /* Now the main event: Visit the shader IR and generate our CS IR for it.
6717 fs_visitor
v8(compiler
, log_data
, mem_ctx
, key
, &prog_data
->base
,
6718 NULL
, /* Never used in core profile */
6719 shader
, 8, shader_time_index
);
6720 if (simd_required
<= 8) {
6722 fail_msg
= v8
.fail_msg
;
6725 cs_set_simd_size(prog_data
, 8);
6726 cs_fill_push_const_info(compiler
->devinfo
, prog_data
);
6727 prog_data
->base
.dispatch_grf_start_reg
= v8
.payload
.num_regs
;
6731 fs_visitor
v16(compiler
, log_data
, mem_ctx
, key
, &prog_data
->base
,
6732 NULL
, /* Never used in core profile */
6733 shader
, 16, shader_time_index
);
6734 if (likely(!(INTEL_DEBUG
& DEBUG_NO16
)) &&
6735 !fail_msg
&& v8
.max_dispatch_width
>= 16 &&
6736 simd_required
<= 16) {
6737 /* Try a SIMD16 compile */
6738 if (simd_required
<= 8)
6739 v16
.import_uniforms(&v8
);
6740 if (!v16
.run_cs()) {
6741 compiler
->shader_perf_log(log_data
,
6742 "SIMD16 shader failed to compile: %s",
6746 "Couldn't generate SIMD16 program and not "
6747 "enough threads for SIMD8";
6751 cs_set_simd_size(prog_data
, 16);
6752 cs_fill_push_const_info(compiler
->devinfo
, prog_data
);
6753 prog_data
->dispatch_grf_start_reg_16
= v16
.payload
.num_regs
;
6757 fs_visitor
v32(compiler
, log_data
, mem_ctx
, key
, &prog_data
->base
,
6758 NULL
, /* Never used in core profile */
6759 shader
, 32, shader_time_index
);
6760 if (!fail_msg
&& v8
.max_dispatch_width
>= 32 &&
6761 (simd_required
> 16 || (INTEL_DEBUG
& DEBUG_DO32
))) {
6762 /* Try a SIMD32 compile */
6763 if (simd_required
<= 8)
6764 v32
.import_uniforms(&v8
);
6765 else if (simd_required
<= 16)
6766 v32
.import_uniforms(&v16
);
6768 if (!v32
.run_cs()) {
6769 compiler
->shader_perf_log(log_data
,
6770 "SIMD32 shader failed to compile: %s",
6774 "Couldn't generate SIMD32 program and not "
6775 "enough threads for SIMD16";
6779 cs_set_simd_size(prog_data
, 32);
6780 cs_fill_push_const_info(compiler
->devinfo
, prog_data
);
6784 if (unlikely(cfg
== NULL
)) {
6787 *error_str
= ralloc_strdup(mem_ctx
, fail_msg
);
6792 fs_generator
g(compiler
, log_data
, mem_ctx
, (void*) key
, &prog_data
->base
,
6793 v8
.promoted_constants
, v8
.runtime_check_aads_emit
,
6794 MESA_SHADER_COMPUTE
);
6795 if (INTEL_DEBUG
& DEBUG_CS
) {
6796 char *name
= ralloc_asprintf(mem_ctx
, "%s compute shader %s",
6797 shader
->info
.label
? shader
->info
.label
:
6800 g
.enable_debug(name
);
6803 g
.generate_code(cfg
, prog_data
->simd_size
);
6805 return g
.get_assembly(final_assembly_size
);
6809 * Test the dispatch mask packing assumptions of
6810 * brw_stage_has_packed_dispatch(). Call this from e.g. the top of
6811 * fs_visitor::emit_nir_code() to cause a GPU hang if any shader invocation is
6812 * executed with an unexpected dispatch mask.
6815 brw_fs_test_dispatch_packing(const fs_builder
&bld
)
6817 const gl_shader_stage stage
= bld
.shader
->stage
;
6819 if (brw_stage_has_packed_dispatch(bld
.shader
->devinfo
, stage
,
6820 bld
.shader
->stage_prog_data
)) {
6821 const fs_builder ubld
= bld
.exec_all().group(1, 0);
6822 const fs_reg tmp
= component(bld
.vgrf(BRW_REGISTER_TYPE_UD
), 0);
6823 const fs_reg mask
= (stage
== MESA_SHADER_FRAGMENT
? brw_vmask_reg() :
6826 ubld
.ADD(tmp
, mask
, brw_imm_ud(1));
6827 ubld
.AND(tmp
, mask
, tmp
);
6829 /* This will loop forever if the dispatch mask doesn't have the expected
6830 * form '2^n-1', in which case tmp will be non-zero.
6832 bld
.emit(BRW_OPCODE_DO
);
6833 bld
.CMP(bld
.null_reg_ud(), tmp
, brw_imm_ud(0), BRW_CONDITIONAL_NZ
);
6834 set_predicate(BRW_PREDICATE_NORMAL
, bld
.emit(BRW_OPCODE_WHILE
));