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"
32 #include "brw_context.h"
37 #include "brw_vec4_gs_visitor.h"
39 #include "brw_program.h"
40 #include "brw_dead_control_flow.h"
41 #include "compiler/glsl_types.h"
46 fs_inst::init(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
47 const fs_reg
*src
, unsigned sources
)
49 memset(this, 0, sizeof(*this));
51 this->src
= new fs_reg
[MAX2(sources
, 3)];
52 for (unsigned i
= 0; i
< sources
; i
++)
53 this->src
[i
] = src
[i
];
55 this->opcode
= opcode
;
57 this->sources
= sources
;
58 this->exec_size
= exec_size
;
60 assert(dst
.file
!= IMM
&& dst
.file
!= UNIFORM
);
62 assert(this->exec_size
!= 0);
64 this->conditional_mod
= BRW_CONDITIONAL_NONE
;
66 /* This will be the case for almost all instructions. */
73 this->regs_written
= DIV_ROUND_UP(dst
.component_size(exec_size
),
77 this->regs_written
= 0;
81 unreachable("Invalid destination register file");
84 this->writes_accumulator
= false;
89 init(BRW_OPCODE_NOP
, 8, dst
, NULL
, 0);
92 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
)
94 init(opcode
, exec_size
, reg_undef
, NULL
, 0);
97 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
)
99 init(opcode
, exec_size
, dst
, NULL
, 0);
102 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
105 const fs_reg src
[1] = { src0
};
106 init(opcode
, exec_size
, dst
, src
, 1);
109 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
110 const fs_reg
&src0
, const fs_reg
&src1
)
112 const fs_reg src
[2] = { src0
, src1
};
113 init(opcode
, exec_size
, dst
, src
, 2);
116 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
117 const fs_reg
&src0
, const fs_reg
&src1
, const fs_reg
&src2
)
119 const fs_reg src
[3] = { src0
, src1
, src2
};
120 init(opcode
, exec_size
, dst
, src
, 3);
123 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_width
, const fs_reg
&dst
,
124 const fs_reg src
[], unsigned sources
)
126 init(opcode
, exec_width
, dst
, src
, sources
);
129 fs_inst::fs_inst(const fs_inst
&that
)
131 memcpy(this, &that
, sizeof(that
));
133 this->src
= new fs_reg
[MAX2(that
.sources
, 3)];
135 for (unsigned i
= 0; i
< that
.sources
; i
++)
136 this->src
[i
] = that
.src
[i
];
145 fs_inst::resize_sources(uint8_t num_sources
)
147 if (this->sources
!= num_sources
) {
148 fs_reg
*src
= new fs_reg
[MAX2(num_sources
, 3)];
150 for (unsigned i
= 0; i
< MIN2(this->sources
, num_sources
); ++i
)
151 src
[i
] = this->src
[i
];
155 this->sources
= num_sources
;
160 fs_visitor::VARYING_PULL_CONSTANT_LOAD(const fs_builder
&bld
,
162 const fs_reg
&surf_index
,
163 const fs_reg
&varying_offset
,
164 uint32_t const_offset
)
166 /* We have our constant surface use a pitch of 4 bytes, so our index can
167 * be any component of a vector, and then we load 4 contiguous
168 * components starting from that.
170 * We break down the const_offset to a portion added to the variable
171 * offset and a portion done using reg_offset, which means that if you
172 * have GLSL using something like "uniform vec4 a[20]; gl_FragColor =
173 * a[i]", we'll temporarily generate 4 vec4 loads from offset i * 4, and
174 * CSE can later notice that those loads are all the same and eliminate
175 * the redundant ones.
177 fs_reg vec4_offset
= vgrf(glsl_type::uint_type
);
178 bld
.ADD(vec4_offset
, varying_offset
, brw_imm_ud(const_offset
& ~0xf));
181 if (devinfo
->gen
== 4 && bld
.dispatch_width() == 8) {
182 /* Pre-gen5, we can either use a SIMD8 message that requires (header,
183 * u, v, r) as parameters, or we can just use the SIMD16 message
184 * consisting of (header, u). We choose the second, at the cost of a
185 * longer return length.
191 if (devinfo
->gen
>= 7)
192 op
= FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7
;
194 op
= FS_OPCODE_VARYING_PULL_CONSTANT_LOAD
;
196 int regs_written
= 4 * (bld
.dispatch_width() / 8) * scale
;
197 fs_reg vec4_result
= fs_reg(VGRF
, alloc
.allocate(regs_written
), dst
.type
);
198 fs_inst
*inst
= bld
.emit(op
, vec4_result
, surf_index
, vec4_offset
);
199 inst
->regs_written
= regs_written
;
201 if (devinfo
->gen
< 7) {
202 inst
->base_mrf
= FIRST_PULL_LOAD_MRF(devinfo
->gen
);
203 inst
->header_size
= 1;
204 if (devinfo
->gen
== 4)
207 inst
->mlen
= 1 + bld
.dispatch_width() / 8;
210 bld
.MOV(dst
, offset(vec4_result
, bld
, ((const_offset
& 0xf) / 4) * scale
));
214 * A helper for MOV generation for fixing up broken hardware SEND dependency
218 fs_visitor::DEP_RESOLVE_MOV(const fs_builder
&bld
, int grf
)
220 /* The caller always wants uncompressed to emit the minimal extra
221 * dependencies, and to avoid having to deal with aligning its regs to 2.
223 const fs_builder ubld
= bld
.annotate("send dependency resolve")
226 ubld
.MOV(ubld
.null_reg_f(), fs_reg(VGRF
, grf
, BRW_REGISTER_TYPE_F
));
230 fs_inst::equals(fs_inst
*inst
) const
232 return (opcode
== inst
->opcode
&&
233 dst
.equals(inst
->dst
) &&
234 src
[0].equals(inst
->src
[0]) &&
235 src
[1].equals(inst
->src
[1]) &&
236 src
[2].equals(inst
->src
[2]) &&
237 saturate
== inst
->saturate
&&
238 predicate
== inst
->predicate
&&
239 conditional_mod
== inst
->conditional_mod
&&
240 mlen
== inst
->mlen
&&
241 base_mrf
== inst
->base_mrf
&&
242 target
== inst
->target
&&
244 header_size
== inst
->header_size
&&
245 shadow_compare
== inst
->shadow_compare
&&
246 exec_size
== inst
->exec_size
&&
247 offset
== inst
->offset
);
251 fs_inst::overwrites_reg(const fs_reg
®
) const
253 return reg
.in_range(dst
, regs_written
);
257 fs_inst::is_send_from_grf() const
260 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7
:
261 case SHADER_OPCODE_SHADER_TIME_ADD
:
262 case FS_OPCODE_INTERPOLATE_AT_CENTROID
:
263 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
264 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
265 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
266 case SHADER_OPCODE_UNTYPED_ATOMIC
:
267 case SHADER_OPCODE_UNTYPED_SURFACE_READ
:
268 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE
:
269 case SHADER_OPCODE_TYPED_ATOMIC
:
270 case SHADER_OPCODE_TYPED_SURFACE_READ
:
271 case SHADER_OPCODE_TYPED_SURFACE_WRITE
:
272 case SHADER_OPCODE_URB_WRITE_SIMD8
:
273 case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
274 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
:
275 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
276 case SHADER_OPCODE_URB_READ_SIMD8
:
277 case SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
:
279 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
280 return src
[1].file
== VGRF
;
281 case FS_OPCODE_FB_WRITE
:
282 return src
[0].file
== VGRF
;
285 return src
[0].file
== VGRF
;
292 * Returns true if this instruction's sources and destinations cannot
293 * safely be the same register.
295 * In most cases, a register can be written over safely by the same
296 * instruction that is its last use. For a single instruction, the
297 * sources are dereferenced before writing of the destination starts
300 * However, there are a few cases where this can be problematic:
302 * - Virtual opcodes that translate to multiple instructions in the
303 * code generator: if src == dst and one instruction writes the
304 * destination before a later instruction reads the source, then
305 * src will have been clobbered.
307 * - SIMD16 compressed instructions with certain regioning (see below).
309 * The register allocator uses this information to set up conflicts between
310 * GRF sources and the destination.
313 fs_inst::has_source_and_destination_hazard() const
316 case FS_OPCODE_PACK_HALF_2x16_SPLIT
:
317 /* Multiple partial writes to the destination */
320 /* The SIMD16 compressed instruction
322 * add(16) g4<1>F g4<8,8,1>F g6<8,8,1>F
324 * is actually decoded in hardware as:
326 * add(8) g4<1>F g4<8,8,1>F g6<8,8,1>F
327 * add(8) g5<1>F g5<8,8,1>F g7<8,8,1>F
329 * Which is safe. However, if we have uniform accesses
330 * happening, we get into trouble:
332 * add(8) g4<1>F g4<0,1,0>F g6<8,8,1>F
333 * add(8) g5<1>F g4<0,1,0>F g7<8,8,1>F
335 * Now our destination for the first instruction overwrote the
336 * second instruction's src0, and we get garbage for those 8
337 * pixels. There's a similar issue for the pre-gen6
338 * pixel_x/pixel_y, which are registers of 16-bit values and thus
339 * would get stomped by the first decode as well.
341 if (exec_size
== 16) {
342 for (int i
= 0; i
< sources
; i
++) {
343 if (src
[i
].file
== VGRF
&& (src
[i
].stride
== 0 ||
344 src
[i
].type
== BRW_REGISTER_TYPE_UW
||
345 src
[i
].type
== BRW_REGISTER_TYPE_W
||
346 src
[i
].type
== BRW_REGISTER_TYPE_UB
||
347 src
[i
].type
== BRW_REGISTER_TYPE_B
)) {
357 fs_inst::is_copy_payload(const brw::simple_allocator
&grf_alloc
) const
359 if (this->opcode
!= SHADER_OPCODE_LOAD_PAYLOAD
)
362 fs_reg reg
= this->src
[0];
363 if (reg
.file
!= VGRF
|| reg
.reg_offset
!= 0 || reg
.stride
== 0)
366 if (grf_alloc
.sizes
[reg
.nr
] != this->regs_written
)
369 for (int i
= 0; i
< this->sources
; i
++) {
370 reg
.type
= this->src
[i
].type
;
371 if (!this->src
[i
].equals(reg
))
374 if (i
< this->header_size
) {
377 reg
.reg_offset
+= this->exec_size
/ 8;
385 fs_inst::can_do_source_mods(const struct brw_device_info
*devinfo
)
387 if (devinfo
->gen
== 6 && is_math())
390 if (is_send_from_grf())
393 if (!backend_instruction::can_do_source_mods())
400 fs_inst::can_change_types() const
402 return dst
.type
== src
[0].type
&&
403 !src
[0].abs
&& !src
[0].negate
&& !saturate
&&
404 (opcode
== BRW_OPCODE_MOV
||
405 (opcode
== BRW_OPCODE_SEL
&&
406 dst
.type
== src
[1].type
&&
407 predicate
!= BRW_PREDICATE_NONE
&&
408 !src
[1].abs
&& !src
[1].negate
));
412 fs_inst::has_side_effects() const
414 return this->eot
|| backend_instruction::has_side_effects();
420 memset(this, 0, sizeof(*this));
424 /** Generic unset register constructor. */
428 this->file
= BAD_FILE
;
431 fs_reg::fs_reg(struct ::brw_reg reg
) :
434 this->reg_offset
= 0;
435 this->subreg_offset
= 0;
437 if (this->file
== IMM
&&
438 (this->type
!= BRW_REGISTER_TYPE_V
&&
439 this->type
!= BRW_REGISTER_TYPE_UV
&&
440 this->type
!= BRW_REGISTER_TYPE_VF
)) {
446 fs_reg::equals(const fs_reg
&r
) const
448 return (this->backend_reg::equals(r
) &&
449 subreg_offset
== r
.subreg_offset
&&
454 fs_reg::set_smear(unsigned subreg
)
456 assert(file
!= ARF
&& file
!= FIXED_GRF
&& file
!= IMM
);
457 subreg_offset
= subreg
* type_sz(type
);
463 fs_reg::is_contiguous() const
469 fs_reg::component_size(unsigned width
) const
471 const unsigned stride
= ((file
!= ARF
&& file
!= FIXED_GRF
) ? this->stride
:
474 return MAX2(width
* stride
, 1) * type_sz(type
);
478 type_size_scalar(const struct glsl_type
*type
)
480 unsigned int size
, i
;
482 switch (type
->base_type
) {
485 case GLSL_TYPE_FLOAT
:
487 return type
->components();
488 case GLSL_TYPE_ARRAY
:
489 return type_size_scalar(type
->fields
.array
) * type
->length
;
490 case GLSL_TYPE_STRUCT
:
492 for (i
= 0; i
< type
->length
; i
++) {
493 size
+= type_size_scalar(type
->fields
.structure
[i
].type
);
496 case GLSL_TYPE_SAMPLER
:
497 /* Samplers take up no register space, since they're baked in at
501 case GLSL_TYPE_ATOMIC_UINT
:
503 case GLSL_TYPE_SUBROUTINE
:
505 case GLSL_TYPE_IMAGE
:
506 return BRW_IMAGE_PARAM_SIZE
;
508 case GLSL_TYPE_ERROR
:
509 case GLSL_TYPE_INTERFACE
:
510 case GLSL_TYPE_DOUBLE
:
511 case GLSL_TYPE_FUNCTION
:
512 unreachable("not reached");
519 * Returns the number of scalar components needed to store type, assuming
520 * that vectors are padded out to vec4.
522 * This has the packing rules of type_size_vec4(), but counts components
523 * similar to type_size_scalar().
526 type_size_vec4_times_4(const struct glsl_type
*type
)
528 return 4 * type_size_vec4(type
);
532 * Create a MOV to read the timestamp register.
534 * The caller is responsible for emitting the MOV. The return value is
535 * the destination of the MOV, with extra parameters set.
538 fs_visitor::get_timestamp(const fs_builder
&bld
)
540 assert(devinfo
->gen
>= 7);
542 fs_reg ts
= fs_reg(retype(brw_vec4_reg(BRW_ARCHITECTURE_REGISTER_FILE
,
545 BRW_REGISTER_TYPE_UD
));
547 fs_reg dst
= fs_reg(VGRF
, alloc
.allocate(1), BRW_REGISTER_TYPE_UD
);
549 /* We want to read the 3 fields we care about even if it's not enabled in
552 bld
.group(4, 0).exec_all().MOV(dst
, ts
);
558 fs_visitor::emit_shader_time_begin()
560 shader_start_time
= get_timestamp(bld
.annotate("shader time start"));
562 /* We want only the low 32 bits of the timestamp. Since it's running
563 * at the GPU clock rate of ~1.2ghz, it will roll over every ~3 seconds,
564 * which is plenty of time for our purposes. It is identical across the
565 * EUs, but since it's tracking GPU core speed it will increment at a
566 * varying rate as render P-states change.
568 shader_start_time
.set_smear(0);
572 fs_visitor::emit_shader_time_end()
574 /* Insert our code just before the final SEND with EOT. */
575 exec_node
*end
= this->instructions
.get_tail();
576 assert(end
&& ((fs_inst
*) end
)->eot
);
577 const fs_builder ibld
= bld
.annotate("shader time end")
578 .exec_all().at(NULL
, end
);
580 fs_reg shader_end_time
= get_timestamp(ibld
);
582 /* We only use the low 32 bits of the timestamp - see
583 * emit_shader_time_begin()).
585 * We could also check if render P-states have changed (or anything
586 * else that might disrupt timing) by setting smear to 2 and checking if
587 * that field is != 0.
589 shader_end_time
.set_smear(0);
591 /* Check that there weren't any timestamp reset events (assuming these
592 * were the only two timestamp reads that happened).
594 fs_reg reset
= shader_end_time
;
596 set_condmod(BRW_CONDITIONAL_Z
,
597 ibld
.AND(ibld
.null_reg_ud(), reset
, brw_imm_ud(1u)));
598 ibld
.IF(BRW_PREDICATE_NORMAL
);
600 fs_reg start
= shader_start_time
;
602 fs_reg diff
= fs_reg(VGRF
, alloc
.allocate(1), BRW_REGISTER_TYPE_UD
);
605 const fs_builder cbld
= ibld
.group(1, 0);
606 cbld
.group(1, 0).ADD(diff
, start
, shader_end_time
);
608 /* If there were no instructions between the two timestamp gets, the diff
609 * is 2 cycles. Remove that overhead, so I can forget about that when
610 * trying to determine the time taken for single instructions.
612 cbld
.ADD(diff
, diff
, brw_imm_ud(-2u));
613 SHADER_TIME_ADD(cbld
, 0, diff
);
614 SHADER_TIME_ADD(cbld
, 1, brw_imm_ud(1u));
615 ibld
.emit(BRW_OPCODE_ELSE
);
616 SHADER_TIME_ADD(cbld
, 2, brw_imm_ud(1u));
617 ibld
.emit(BRW_OPCODE_ENDIF
);
621 fs_visitor::SHADER_TIME_ADD(const fs_builder
&bld
,
622 int shader_time_subindex
,
625 int index
= shader_time_index
* 3 + shader_time_subindex
;
626 struct brw_reg offset
= brw_imm_d(index
* SHADER_TIME_STRIDE
);
629 if (dispatch_width
== 8)
630 payload
= vgrf(glsl_type::uvec2_type
);
632 payload
= vgrf(glsl_type::uint_type
);
634 bld
.emit(SHADER_OPCODE_SHADER_TIME_ADD
, fs_reg(), payload
, offset
, value
);
638 fs_visitor::vfail(const char *format
, va_list va
)
647 msg
= ralloc_vasprintf(mem_ctx
, format
, va
);
648 msg
= ralloc_asprintf(mem_ctx
, "%s compile failed: %s\n", stage_abbrev
, msg
);
650 this->fail_msg
= msg
;
653 fprintf(stderr
, "%s", msg
);
658 fs_visitor::fail(const char *format
, ...)
662 va_start(va
, format
);
668 * Mark this program as impossible to compile in SIMD16 mode.
670 * During the SIMD8 compile (which happens first), we can detect and flag
671 * things that are unsupported in SIMD16 mode, so the compiler can skip
672 * the SIMD16 compile altogether.
674 * During a SIMD16 compile (if one happens anyway), this just calls fail().
677 fs_visitor::no16(const char *msg
)
679 if (dispatch_width
== 16) {
682 simd16_unsupported
= true;
684 compiler
->shader_perf_log(log_data
,
685 "SIMD16 shader failed to compile: %s", msg
);
690 * Returns true if the instruction has a flag that means it won't
691 * update an entire destination register.
693 * For example, dead code elimination and live variable analysis want to know
694 * when a write to a variable screens off any preceding values that were in
698 fs_inst::is_partial_write() const
700 return ((this->predicate
&& this->opcode
!= BRW_OPCODE_SEL
) ||
701 (this->exec_size
* type_sz(this->dst
.type
)) < 32 ||
702 !this->dst
.is_contiguous());
706 fs_inst::components_read(unsigned i
) const
709 case FS_OPCODE_LINTERP
:
715 case FS_OPCODE_PIXEL_X
:
716 case FS_OPCODE_PIXEL_Y
:
720 case FS_OPCODE_FB_WRITE_LOGICAL
:
721 assert(src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].file
== IMM
);
722 /* First/second FB write color. */
724 return src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].ud
;
728 case SHADER_OPCODE_TEX_LOGICAL
:
729 case SHADER_OPCODE_TXD_LOGICAL
:
730 case SHADER_OPCODE_TXF_LOGICAL
:
731 case SHADER_OPCODE_TXL_LOGICAL
:
732 case SHADER_OPCODE_TXS_LOGICAL
:
733 case FS_OPCODE_TXB_LOGICAL
:
734 case SHADER_OPCODE_TXF_CMS_LOGICAL
:
735 case SHADER_OPCODE_TXF_CMS_W_LOGICAL
:
736 case SHADER_OPCODE_TXF_UMS_LOGICAL
:
737 case SHADER_OPCODE_TXF_MCS_LOGICAL
:
738 case SHADER_OPCODE_LOD_LOGICAL
:
739 case SHADER_OPCODE_TG4_LOGICAL
:
740 case SHADER_OPCODE_TG4_OFFSET_LOGICAL
:
741 assert(src
[TEX_LOGICAL_SRC_COORD_COMPONENTS
].file
== IMM
&&
742 src
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
].file
== IMM
);
743 /* Texture coordinates. */
744 if (i
== TEX_LOGICAL_SRC_COORDINATE
)
745 return src
[TEX_LOGICAL_SRC_COORD_COMPONENTS
].ud
;
746 /* Texture derivatives. */
747 else if ((i
== TEX_LOGICAL_SRC_LOD
|| i
== TEX_LOGICAL_SRC_LOD2
) &&
748 opcode
== SHADER_OPCODE_TXD_LOGICAL
)
749 return src
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
].ud
;
750 /* Texture offset. */
751 else if (i
== TEX_LOGICAL_SRC_OFFSET_VALUE
)
754 else if (i
== TEX_LOGICAL_SRC_MCS
&& opcode
== SHADER_OPCODE_TXF_CMS_W_LOGICAL
)
759 case SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
:
760 case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
:
761 assert(src
[3].file
== IMM
);
762 /* Surface coordinates. */
765 /* Surface operation source (ignored for reads). */
771 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
:
772 case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
:
773 assert(src
[3].file
== IMM
&&
775 /* Surface coordinates. */
778 /* Surface operation source. */
784 case SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL
:
785 case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
: {
786 assert(src
[3].file
== IMM
&&
788 const unsigned op
= src
[4].ud
;
789 /* Surface coordinates. */
792 /* Surface operation source. */
793 else if (i
== 1 && op
== BRW_AOP_CMPWR
)
795 else if (i
== 1 && (op
== BRW_AOP_INC
|| op
== BRW_AOP_DEC
||
796 op
== BRW_AOP_PREDEC
))
808 fs_inst::regs_read(int arg
) const
811 case FS_OPCODE_FB_WRITE
:
812 case SHADER_OPCODE_URB_WRITE_SIMD8
:
813 case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
814 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
:
815 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
816 case SHADER_OPCODE_URB_READ_SIMD8
:
817 case SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
:
818 case SHADER_OPCODE_UNTYPED_ATOMIC
:
819 case SHADER_OPCODE_UNTYPED_SURFACE_READ
:
820 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE
:
821 case SHADER_OPCODE_TYPED_ATOMIC
:
822 case SHADER_OPCODE_TYPED_SURFACE_READ
:
823 case SHADER_OPCODE_TYPED_SURFACE_WRITE
:
824 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
829 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7
:
830 /* The payload is actually stored in src1 */
835 case FS_OPCODE_LINTERP
:
840 case SHADER_OPCODE_LOAD_PAYLOAD
:
841 if (arg
< this->header_size
)
845 case CS_OPCODE_CS_TERMINATE
:
846 case SHADER_OPCODE_BARRIER
:
849 case SHADER_OPCODE_MOV_INDIRECT
:
851 assert(src
[2].file
== IMM
);
852 unsigned region_length
= src
[2].ud
;
854 if (src
[0].file
== UNIFORM
) {
855 assert(region_length
% 4 == 0);
856 return region_length
/ 4;
857 } else if (src
[0].file
== FIXED_GRF
) {
858 /* If the start of the region is not register aligned, then
859 * there's some portion of the register that's technically
860 * unread at the beginning.
862 * However, the register allocator works in terms of whole
863 * registers, and does not use subnr. It assumes that the
864 * read starts at the beginning of the register, and extends
865 * regs_read() whole registers beyond that.
867 * To compensate, we extend the region length to include this
868 * unread portion at the beginning.
871 region_length
+= src
[0].subnr
;
873 return DIV_ROUND_UP(region_length
, REG_SIZE
);
875 assert(!"Invalid register file");
881 if (is_tex() && arg
== 0 && src
[0].file
== VGRF
)
886 switch (src
[arg
].file
) {
896 return DIV_ROUND_UP(components_read(arg
) *
897 src
[arg
].component_size(exec_size
),
900 unreachable("MRF registers are not allowed as sources");
906 fs_inst::reads_flag() const
912 fs_inst::writes_flag() const
914 return (conditional_mod
&& (opcode
!= BRW_OPCODE_SEL
&&
915 opcode
!= BRW_OPCODE_IF
&&
916 opcode
!= BRW_OPCODE_WHILE
)) ||
917 opcode
== FS_OPCODE_MOV_DISPATCH_TO_FLAGS
;
921 * Returns how many MRFs an FS opcode will write over.
923 * Note that this is not the 0 or 1 implied writes in an actual gen
924 * instruction -- the FS opcodes often generate MOVs in addition.
927 fs_visitor::implied_mrf_writes(fs_inst
*inst
)
932 if (inst
->base_mrf
== -1)
935 switch (inst
->opcode
) {
936 case SHADER_OPCODE_RCP
:
937 case SHADER_OPCODE_RSQ
:
938 case SHADER_OPCODE_SQRT
:
939 case SHADER_OPCODE_EXP2
:
940 case SHADER_OPCODE_LOG2
:
941 case SHADER_OPCODE_SIN
:
942 case SHADER_OPCODE_COS
:
943 return 1 * dispatch_width
/ 8;
944 case SHADER_OPCODE_POW
:
945 case SHADER_OPCODE_INT_QUOTIENT
:
946 case SHADER_OPCODE_INT_REMAINDER
:
947 return 2 * dispatch_width
/ 8;
948 case SHADER_OPCODE_TEX
:
950 case SHADER_OPCODE_TXD
:
951 case SHADER_OPCODE_TXF
:
952 case SHADER_OPCODE_TXF_CMS
:
953 case SHADER_OPCODE_TXF_CMS_W
:
954 case SHADER_OPCODE_TXF_MCS
:
955 case SHADER_OPCODE_TG4
:
956 case SHADER_OPCODE_TG4_OFFSET
:
957 case SHADER_OPCODE_TXL
:
958 case SHADER_OPCODE_TXS
:
959 case SHADER_OPCODE_LOD
:
960 case SHADER_OPCODE_SAMPLEINFO
:
962 case FS_OPCODE_FB_WRITE
:
964 case FS_OPCODE_GET_BUFFER_SIZE
:
965 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
966 case SHADER_OPCODE_GEN4_SCRATCH_READ
:
968 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD
:
970 case SHADER_OPCODE_GEN4_SCRATCH_WRITE
:
972 case SHADER_OPCODE_UNTYPED_ATOMIC
:
973 case SHADER_OPCODE_UNTYPED_SURFACE_READ
:
974 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE
:
975 case SHADER_OPCODE_TYPED_ATOMIC
:
976 case SHADER_OPCODE_TYPED_SURFACE_READ
:
977 case SHADER_OPCODE_TYPED_SURFACE_WRITE
:
978 case SHADER_OPCODE_URB_WRITE_SIMD8
:
979 case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
980 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
:
981 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
982 case FS_OPCODE_INTERPOLATE_AT_CENTROID
:
983 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
984 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
985 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
988 unreachable("not reached");
993 fs_visitor::vgrf(const glsl_type
*const type
)
995 int reg_width
= dispatch_width
/ 8;
996 return fs_reg(VGRF
, alloc
.allocate(type_size_scalar(type
) * reg_width
),
997 brw_type_for_base_type(type
));
1000 fs_reg::fs_reg(enum brw_reg_file file
, int nr
)
1005 this->type
= BRW_REGISTER_TYPE_F
;
1006 this->stride
= (file
== UNIFORM
? 0 : 1);
1009 fs_reg::fs_reg(enum brw_reg_file file
, int nr
, enum brw_reg_type type
)
1015 this->stride
= (file
== UNIFORM
? 0 : 1);
1018 /* For SIMD16, we need to follow from the uniform setup of SIMD8 dispatch.
1019 * This brings in those uniform definitions
1022 fs_visitor::import_uniforms(fs_visitor
*v
)
1024 this->push_constant_loc
= v
->push_constant_loc
;
1025 this->pull_constant_loc
= v
->pull_constant_loc
;
1026 this->uniforms
= v
->uniforms
;
1030 fs_visitor::emit_fragcoord_interpolation(bool pixel_center_integer
,
1031 bool origin_upper_left
)
1033 assert(stage
== MESA_SHADER_FRAGMENT
);
1034 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1035 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::vec4_type
));
1037 bool flip
= !origin_upper_left
^ key
->render_to_fbo
;
1039 /* gl_FragCoord.x */
1040 if (pixel_center_integer
) {
1041 bld
.MOV(wpos
, this->pixel_x
);
1043 bld
.ADD(wpos
, this->pixel_x
, brw_imm_f(0.5f
));
1045 wpos
= offset(wpos
, bld
, 1);
1047 /* gl_FragCoord.y */
1048 if (!flip
&& pixel_center_integer
) {
1049 bld
.MOV(wpos
, this->pixel_y
);
1051 fs_reg pixel_y
= this->pixel_y
;
1052 float offset
= (pixel_center_integer
? 0.0f
: 0.5f
);
1055 pixel_y
.negate
= true;
1056 offset
+= key
->drawable_height
- 1.0f
;
1059 bld
.ADD(wpos
, pixel_y
, brw_imm_f(offset
));
1061 wpos
= offset(wpos
, bld
, 1);
1063 /* gl_FragCoord.z */
1064 if (devinfo
->gen
>= 6) {
1065 bld
.MOV(wpos
, fs_reg(brw_vec8_grf(payload
.source_depth_reg
, 0)));
1067 bld
.emit(FS_OPCODE_LINTERP
, wpos
,
1068 this->delta_xy
[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
],
1069 interp_reg(VARYING_SLOT_POS
, 2));
1071 wpos
= offset(wpos
, bld
, 1);
1073 /* gl_FragCoord.w: Already set up in emit_interpolation */
1074 bld
.MOV(wpos
, this->wpos_w
);
1080 fs_visitor::emit_linterp(const fs_reg
&attr
, const fs_reg
&interp
,
1081 glsl_interp_qualifier interpolation_mode
,
1082 bool is_centroid
, bool is_sample
)
1084 brw_wm_barycentric_interp_mode barycoord_mode
;
1085 if (devinfo
->gen
>= 6) {
1087 if (interpolation_mode
== INTERP_QUALIFIER_SMOOTH
)
1088 barycoord_mode
= BRW_WM_PERSPECTIVE_CENTROID_BARYCENTRIC
;
1090 barycoord_mode
= BRW_WM_NONPERSPECTIVE_CENTROID_BARYCENTRIC
;
1091 } else if (is_sample
) {
1092 if (interpolation_mode
== INTERP_QUALIFIER_SMOOTH
)
1093 barycoord_mode
= BRW_WM_PERSPECTIVE_SAMPLE_BARYCENTRIC
;
1095 barycoord_mode
= BRW_WM_NONPERSPECTIVE_SAMPLE_BARYCENTRIC
;
1097 if (interpolation_mode
== INTERP_QUALIFIER_SMOOTH
)
1098 barycoord_mode
= BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
;
1100 barycoord_mode
= BRW_WM_NONPERSPECTIVE_PIXEL_BARYCENTRIC
;
1103 /* On Ironlake and below, there is only one interpolation mode.
1104 * Centroid interpolation doesn't mean anything on this hardware --
1105 * there is no multisampling.
1107 barycoord_mode
= BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
;
1109 return bld
.emit(FS_OPCODE_LINTERP
, attr
,
1110 this->delta_xy
[barycoord_mode
], interp
);
1114 fs_visitor::emit_general_interpolation(fs_reg
*attr
, const char *name
,
1115 const glsl_type
*type
,
1116 glsl_interp_qualifier interpolation_mode
,
1117 int *location
, bool mod_centroid
,
1120 assert(stage
== MESA_SHADER_FRAGMENT
);
1121 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
1122 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1124 if (interpolation_mode
== INTERP_QUALIFIER_NONE
) {
1126 *location
== VARYING_SLOT_COL0
|| *location
== VARYING_SLOT_COL1
;
1127 if (key
->flat_shade
&& is_gl_Color
) {
1128 interpolation_mode
= INTERP_QUALIFIER_FLAT
;
1130 interpolation_mode
= INTERP_QUALIFIER_SMOOTH
;
1134 if (type
->is_array() || type
->is_matrix()) {
1135 const glsl_type
*elem_type
= glsl_get_array_element(type
);
1136 const unsigned length
= glsl_get_length(type
);
1138 for (unsigned i
= 0; i
< length
; i
++) {
1139 emit_general_interpolation(attr
, name
, elem_type
, interpolation_mode
,
1140 location
, mod_centroid
, mod_sample
);
1142 } else if (type
->is_record()) {
1143 for (unsigned i
= 0; i
< type
->length
; i
++) {
1144 const glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
1145 emit_general_interpolation(attr
, name
, field_type
, interpolation_mode
,
1146 location
, mod_centroid
, mod_sample
);
1149 assert(type
->is_scalar() || type
->is_vector());
1151 if (prog_data
->urb_setup
[*location
] == -1) {
1152 /* If there's no incoming setup data for this slot, don't
1153 * emit interpolation for it.
1155 *attr
= offset(*attr
, bld
, type
->vector_elements
);
1160 attr
->type
= brw_type_for_base_type(type
->get_scalar_type());
1162 if (interpolation_mode
== INTERP_QUALIFIER_FLAT
) {
1163 /* Constant interpolation (flat shading) case. The SF has
1164 * handed us defined values in only the constant offset
1165 * field of the setup reg.
1167 for (unsigned int i
= 0; i
< type
->vector_elements
; i
++) {
1168 struct brw_reg interp
= interp_reg(*location
, i
);
1169 interp
= suboffset(interp
, 3);
1170 interp
.type
= attr
->type
;
1171 bld
.emit(FS_OPCODE_CINTERP
, *attr
, fs_reg(interp
));
1172 *attr
= offset(*attr
, bld
, 1);
1175 /* Smooth/noperspective interpolation case. */
1176 for (unsigned int i
= 0; i
< type
->vector_elements
; i
++) {
1177 struct brw_reg interp
= interp_reg(*location
, i
);
1178 if (devinfo
->needs_unlit_centroid_workaround
&& mod_centroid
) {
1179 /* Get the pixel/sample mask into f0 so that we know
1180 * which pixels are lit. Then, for each channel that is
1181 * unlit, replace the centroid data with non-centroid
1184 bld
.emit(FS_OPCODE_MOV_DISPATCH_TO_FLAGS
);
1187 inst
= emit_linterp(*attr
, fs_reg(interp
), interpolation_mode
,
1189 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1190 inst
->predicate_inverse
= true;
1191 if (devinfo
->has_pln
)
1192 inst
->no_dd_clear
= true;
1194 inst
= emit_linterp(*attr
, fs_reg(interp
), interpolation_mode
,
1195 mod_centroid
&& !key
->persample_shading
,
1196 mod_sample
|| key
->persample_shading
);
1197 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1198 inst
->predicate_inverse
= false;
1199 if (devinfo
->has_pln
)
1200 inst
->no_dd_check
= true;
1203 emit_linterp(*attr
, fs_reg(interp
), interpolation_mode
,
1204 mod_centroid
&& !key
->persample_shading
,
1205 mod_sample
|| key
->persample_shading
);
1207 if (devinfo
->gen
< 6 && interpolation_mode
== INTERP_QUALIFIER_SMOOTH
) {
1208 bld
.MUL(*attr
, *attr
, this->pixel_w
);
1210 *attr
= offset(*attr
, bld
, 1);
1218 fs_visitor::emit_frontfacing_interpolation()
1220 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::bool_type
));
1222 if (devinfo
->gen
>= 6) {
1223 /* Bit 15 of g0.0 is 0 if the polygon is front facing. We want to create
1224 * a boolean result from this (~0/true or 0/false).
1226 * We can use the fact that bit 15 is the MSB of g0.0:W to accomplish
1227 * this task in only one instruction:
1228 * - a negation source modifier will flip the bit; and
1229 * - a W -> D type conversion will sign extend the bit into the high
1230 * word of the destination.
1232 * An ASR 15 fills the low word of the destination.
1234 fs_reg g0
= fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_W
));
1237 bld
.ASR(*reg
, g0
, brw_imm_d(15));
1239 /* Bit 31 of g1.6 is 0 if the polygon is front facing. We want to create
1240 * a boolean result from this (1/true or 0/false).
1242 * Like in the above case, since the bit is the MSB of g1.6:UD we can use
1243 * the negation source modifier to flip it. Unfortunately the SHR
1244 * instruction only operates on UD (or D with an abs source modifier)
1245 * sources without negation.
1247 * Instead, use ASR (which will give ~0/true or 0/false).
1249 fs_reg g1_6
= fs_reg(retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_D
));
1252 bld
.ASR(*reg
, g1_6
, brw_imm_d(31));
1259 fs_visitor::compute_sample_position(fs_reg dst
, fs_reg int_sample_pos
)
1261 assert(stage
== MESA_SHADER_FRAGMENT
);
1262 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1263 assert(dst
.type
== BRW_REGISTER_TYPE_F
);
1265 if (key
->compute_pos_offset
) {
1266 /* Convert int_sample_pos to floating point */
1267 bld
.MOV(dst
, int_sample_pos
);
1268 /* Scale to the range [0, 1] */
1269 bld
.MUL(dst
, dst
, brw_imm_f(1 / 16.0f
));
1272 /* From ARB_sample_shading specification:
1273 * "When rendering to a non-multisample buffer, or if multisample
1274 * rasterization is disabled, gl_SamplePosition will always be
1277 bld
.MOV(dst
, brw_imm_f(0.5f
));
1282 fs_visitor::emit_samplepos_setup()
1284 assert(devinfo
->gen
>= 6);
1286 const fs_builder abld
= bld
.annotate("compute sample position");
1287 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::vec2_type
));
1289 fs_reg int_sample_x
= vgrf(glsl_type::int_type
);
1290 fs_reg int_sample_y
= vgrf(glsl_type::int_type
);
1292 /* WM will be run in MSDISPMODE_PERSAMPLE. So, only one of SIMD8 or SIMD16
1293 * mode will be enabled.
1295 * From the Ivy Bridge PRM, volume 2 part 1, page 344:
1296 * R31.1:0 Position Offset X/Y for Slot[3:0]
1297 * R31.3:2 Position Offset X/Y for Slot[7:4]
1300 * The X, Y sample positions come in as bytes in thread payload. So, read
1301 * the positions using vstride=16, width=8, hstride=2.
1303 struct brw_reg sample_pos_reg
=
1304 stride(retype(brw_vec1_grf(payload
.sample_pos_reg
, 0),
1305 BRW_REGISTER_TYPE_B
), 16, 8, 2);
1307 if (dispatch_width
== 8) {
1308 abld
.MOV(int_sample_x
, fs_reg(sample_pos_reg
));
1310 abld
.half(0).MOV(half(int_sample_x
, 0), fs_reg(sample_pos_reg
));
1311 abld
.half(1).MOV(half(int_sample_x
, 1),
1312 fs_reg(suboffset(sample_pos_reg
, 16)));
1314 /* Compute gl_SamplePosition.x */
1315 compute_sample_position(pos
, int_sample_x
);
1316 pos
= offset(pos
, abld
, 1);
1317 if (dispatch_width
== 8) {
1318 abld
.MOV(int_sample_y
, fs_reg(suboffset(sample_pos_reg
, 1)));
1320 abld
.half(0).MOV(half(int_sample_y
, 0),
1321 fs_reg(suboffset(sample_pos_reg
, 1)));
1322 abld
.half(1).MOV(half(int_sample_y
, 1),
1323 fs_reg(suboffset(sample_pos_reg
, 17)));
1325 /* Compute gl_SamplePosition.y */
1326 compute_sample_position(pos
, int_sample_y
);
1331 fs_visitor::emit_sampleid_setup()
1333 assert(stage
== MESA_SHADER_FRAGMENT
);
1334 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1335 assert(devinfo
->gen
>= 6);
1337 const fs_builder abld
= bld
.annotate("compute sample id");
1338 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::int_type
));
1340 if (!key
->multisample_fbo
) {
1341 /* As per GL_ARB_sample_shading specification:
1342 * "When rendering to a non-multisample buffer, or if multisample
1343 * rasterization is disabled, gl_SampleID will always be zero."
1345 abld
.MOV(*reg
, brw_imm_d(0));
1346 } else if (devinfo
->gen
>= 8) {
1347 /* Sample ID comes in as 4-bit numbers in g1.0:
1349 * 15:12 Slot 3 SampleID (only used in SIMD16)
1350 * 11:8 Slot 2 SampleID (only used in SIMD16)
1351 * 7:4 Slot 1 SampleID
1352 * 3:0 Slot 0 SampleID
1354 * Each slot corresponds to four channels, so we want to replicate each
1355 * half-byte value to 4 channels in a row:
1357 * dst+0: .7 .6 .5 .4 .3 .2 .1 .0
1358 * 7:4 7:4 7:4 7:4 3:0 3:0 3:0 3:0
1360 * dst+1: .7 .6 .5 .4 .3 .2 .1 .0 (if SIMD16)
1361 * 15:12 15:12 15:12 15:12 11:8 11:8 11:8 11:8
1363 * First, we read g1.0 with a <1,8,0>UB region, causing the first 8
1364 * channels to read the first byte (7:0), and the second group of 8
1365 * channels to read the second byte (15:8). Then, we shift right by
1366 * a vector immediate of <4, 4, 4, 4, 0, 0, 0, 0>, moving the slot 1 / 3
1367 * values into place. Finally, we AND with 0xf to keep the low nibble.
1369 * shr(16) tmp<1>W g1.0<1,8,0>B 0x44440000:V
1370 * and(16) dst<1>D tmp<8,8,1>W 0xf:W
1372 * TODO: These payload bits exist on Gen7 too, but they appear to always
1373 * be zero, so this code fails to work. We should find out why.
1375 fs_reg
tmp(VGRF
, alloc
.allocate(1), BRW_REGISTER_TYPE_W
);
1377 abld
.SHR(tmp
, fs_reg(stride(retype(brw_vec1_grf(1, 0),
1378 BRW_REGISTER_TYPE_B
), 1, 8, 0)),
1379 brw_imm_v(0x44440000));
1380 abld
.AND(*reg
, tmp
, brw_imm_w(0xf));
1382 fs_reg
t1(VGRF
, alloc
.allocate(1), BRW_REGISTER_TYPE_D
);
1384 fs_reg
t2(VGRF
, alloc
.allocate(1), BRW_REGISTER_TYPE_W
);
1386 /* The PS will be run in MSDISPMODE_PERSAMPLE. For example with
1387 * 8x multisampling, subspan 0 will represent sample N (where N
1388 * is 0, 2, 4 or 6), subspan 1 will represent sample 1, 3, 5 or
1389 * 7. We can find the value of N by looking at R0.0 bits 7:6
1390 * ("Starting Sample Pair Index (SSPI)") and multiplying by two
1391 * (since samples are always delivered in pairs). That is, we
1392 * compute 2*((R0.0 & 0xc0) >> 6) == (R0.0 & 0xc0) >> 5. Then
1393 * we need to add N to the sequence (0, 0, 0, 0, 1, 1, 1, 1) in
1394 * case of SIMD8 and sequence (0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2,
1395 * 2, 3, 3, 3, 3) in case of SIMD16. We compute this sequence by
1396 * populating a temporary variable with the sequence (0, 1, 2, 3),
1397 * and then reading from it using vstride=1, width=4, hstride=0.
1398 * These computations hold good for 4x multisampling as well.
1400 * For 2x MSAA and SIMD16, we want to use the sequence (0, 1, 0, 1):
1401 * the first four slots are sample 0 of subspan 0; the next four
1402 * are sample 1 of subspan 0; the third group is sample 0 of
1403 * subspan 1, and finally sample 1 of subspan 1.
1406 /* SKL+ has an extra bit for the Starting Sample Pair Index to
1407 * accomodate 16x MSAA.
1409 abld
.exec_all().group(1, 0)
1410 .AND(t1
, fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_D
)),
1412 abld
.exec_all().group(1, 0).SHR(t1
, t1
, brw_imm_d(5));
1414 /* This works for both SIMD8 and SIMD16 */
1415 abld
.exec_all().group(4, 0).MOV(t2
, brw_imm_v(0x3210));
1417 /* This special instruction takes care of setting vstride=1,
1418 * width=4, hstride=0 of t2 during an ADD instruction.
1420 abld
.emit(FS_OPCODE_SET_SAMPLE_ID
, *reg
, t1
, t2
);
1427 fs_visitor::emit_samplemaskin_setup()
1429 assert(stage
== MESA_SHADER_FRAGMENT
);
1430 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1431 assert(devinfo
->gen
>= 6);
1433 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::int_type
));
1435 fs_reg
coverage_mask(retype(brw_vec8_grf(payload
.sample_mask_in_reg
, 0),
1436 BRW_REGISTER_TYPE_D
));
1438 if (key
->persample_shading
) {
1439 /* gl_SampleMaskIn[] comes from two sources: the input coverage mask,
1440 * and a mask representing which sample is being processed by the
1441 * current shader invocation.
1443 * From the OES_sample_variables specification:
1444 * "When per-sample shading is active due to the use of a fragment input
1445 * qualified by "sample" or due to the use of the gl_SampleID or
1446 * gl_SamplePosition variables, only the bit for the current sample is
1447 * set in gl_SampleMaskIn."
1449 const fs_builder abld
= bld
.annotate("compute gl_SampleMaskIn");
1451 if (nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
].file
== BAD_FILE
)
1452 nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
] = *emit_sampleid_setup();
1454 fs_reg one
= vgrf(glsl_type::int_type
);
1455 fs_reg enabled_mask
= vgrf(glsl_type::int_type
);
1456 abld
.MOV(one
, brw_imm_d(1));
1457 abld
.SHL(enabled_mask
, one
, nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
]);
1458 abld
.AND(*reg
, enabled_mask
, coverage_mask
);
1460 /* In per-pixel mode, the coverage mask is sufficient. */
1461 *reg
= coverage_mask
;
1467 fs_visitor::resolve_source_modifiers(const fs_reg
&src
)
1469 if (!src
.abs
&& !src
.negate
)
1472 fs_reg temp
= bld
.vgrf(src
.type
);
1479 fs_visitor::emit_discard_jump()
1481 assert(((brw_wm_prog_data
*) this->prog_data
)->uses_kill
);
1483 /* For performance, after a discard, jump to the end of the
1484 * shader if all relevant channels have been discarded.
1486 fs_inst
*discard_jump
= bld
.emit(FS_OPCODE_DISCARD_JUMP
);
1487 discard_jump
->flag_subreg
= 1;
1489 discard_jump
->predicate
= (dispatch_width
== 8)
1490 ? BRW_PREDICATE_ALIGN1_ANY8H
1491 : BRW_PREDICATE_ALIGN1_ANY16H
;
1492 discard_jump
->predicate_inverse
= true;
1496 fs_visitor::emit_gs_thread_end()
1498 assert(stage
== MESA_SHADER_GEOMETRY
);
1500 struct brw_gs_prog_data
*gs_prog_data
=
1501 (struct brw_gs_prog_data
*) prog_data
;
1503 if (gs_compile
->control_data_header_size_bits
> 0) {
1504 emit_gs_control_data_bits(this->final_gs_vertex_count
);
1507 const fs_builder abld
= bld
.annotate("thread end");
1510 if (gs_prog_data
->static_vertex_count
!= -1) {
1511 foreach_in_list_reverse(fs_inst
, prev
, &this->instructions
) {
1512 if (prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8
||
1513 prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
||
1514 prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
||
1515 prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
) {
1518 /* Delete now dead instructions. */
1519 foreach_in_list_reverse_safe(exec_node
, dead
, &this->instructions
) {
1525 } else if (prev
->is_control_flow() || prev
->has_side_effects()) {
1529 fs_reg hdr
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1530 abld
.MOV(hdr
, fs_reg(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
)));
1531 inst
= abld
.emit(SHADER_OPCODE_URB_WRITE_SIMD8
, reg_undef
, hdr
);
1534 fs_reg payload
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
1535 fs_reg
*sources
= ralloc_array(mem_ctx
, fs_reg
, 2);
1536 sources
[0] = fs_reg(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
));
1537 sources
[1] = this->final_gs_vertex_count
;
1538 abld
.LOAD_PAYLOAD(payload
, sources
, 2, 2);
1539 inst
= abld
.emit(SHADER_OPCODE_URB_WRITE_SIMD8
, reg_undef
, payload
);
1547 fs_visitor::assign_curb_setup()
1549 if (dispatch_width
== 8) {
1550 prog_data
->dispatch_grf_start_reg
= payload
.num_regs
;
1552 if (stage
== MESA_SHADER_FRAGMENT
) {
1553 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
1554 prog_data
->dispatch_grf_start_reg_16
= payload
.num_regs
;
1555 } else if (stage
== MESA_SHADER_COMPUTE
) {
1556 brw_cs_prog_data
*prog_data
= (brw_cs_prog_data
*) this->prog_data
;
1557 prog_data
->dispatch_grf_start_reg_16
= payload
.num_regs
;
1559 unreachable("Unsupported shader type!");
1563 prog_data
->curb_read_length
= ALIGN(stage_prog_data
->nr_params
, 8) / 8;
1565 /* Map the offsets in the UNIFORM file to fixed HW regs. */
1566 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1567 for (unsigned int i
= 0; i
< inst
->sources
; i
++) {
1568 if (inst
->src
[i
].file
== UNIFORM
) {
1569 int uniform_nr
= inst
->src
[i
].nr
+ inst
->src
[i
].reg_offset
;
1571 if (uniform_nr
>= 0 && uniform_nr
< (int) uniforms
) {
1572 constant_nr
= push_constant_loc
[uniform_nr
];
1574 /* Section 5.11 of the OpenGL 4.1 spec says:
1575 * "Out-of-bounds reads return undefined values, which include
1576 * values from other variables of the active program or zero."
1577 * Just return the first push constant.
1582 struct brw_reg brw_reg
= brw_vec1_grf(payload
.num_regs
+
1585 brw_reg
.abs
= inst
->src
[i
].abs
;
1586 brw_reg
.negate
= inst
->src
[i
].negate
;
1588 assert(inst
->src
[i
].stride
== 0);
1589 inst
->src
[i
] = byte_offset(
1590 retype(brw_reg
, inst
->src
[i
].type
),
1591 inst
->src
[i
].subreg_offset
);
1596 /* This may be updated in assign_urb_setup or assign_vs_urb_setup. */
1597 this->first_non_payload_grf
= payload
.num_regs
+ prog_data
->curb_read_length
;
1601 fs_visitor::calculate_urb_setup()
1603 assert(stage
== MESA_SHADER_FRAGMENT
);
1604 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
1605 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1607 memset(prog_data
->urb_setup
, -1,
1608 sizeof(prog_data
->urb_setup
[0]) * VARYING_SLOT_MAX
);
1611 /* Figure out where each of the incoming setup attributes lands. */
1612 if (devinfo
->gen
>= 6) {
1613 if (_mesa_bitcount_64(nir
->info
.inputs_read
&
1614 BRW_FS_VARYING_INPUT_MASK
) <= 16) {
1615 /* The SF/SBE pipeline stage can do arbitrary rearrangement of the
1616 * first 16 varying inputs, so we can put them wherever we want.
1617 * Just put them in order.
1619 * This is useful because it means that (a) inputs not used by the
1620 * fragment shader won't take up valuable register space, and (b) we
1621 * won't have to recompile the fragment shader if it gets paired with
1622 * a different vertex (or geometry) shader.
1624 for (unsigned int i
= 0; i
< VARYING_SLOT_MAX
; i
++) {
1625 if (nir
->info
.inputs_read
& BRW_FS_VARYING_INPUT_MASK
&
1626 BITFIELD64_BIT(i
)) {
1627 prog_data
->urb_setup
[i
] = urb_next
++;
1631 bool include_vue_header
=
1632 nir
->info
.inputs_read
& (VARYING_BIT_LAYER
| VARYING_BIT_VIEWPORT
);
1634 /* We have enough input varyings that the SF/SBE pipeline stage can't
1635 * arbitrarily rearrange them to suit our whim; we have to put them
1636 * in an order that matches the output of the previous pipeline stage
1637 * (geometry or vertex shader).
1639 struct brw_vue_map prev_stage_vue_map
;
1640 brw_compute_vue_map(devinfo
, &prev_stage_vue_map
,
1641 key
->input_slots_valid
,
1642 nir
->info
.separate_shader
);
1644 include_vue_header
? 0 : 2 * BRW_SF_URB_ENTRY_READ_OFFSET
;
1646 assert(prev_stage_vue_map
.num_slots
<= first_slot
+ 32);
1647 for (int slot
= first_slot
; slot
< prev_stage_vue_map
.num_slots
;
1649 int varying
= prev_stage_vue_map
.slot_to_varying
[slot
];
1650 if (varying
!= BRW_VARYING_SLOT_PAD
&&
1651 (nir
->info
.inputs_read
& BRW_FS_VARYING_INPUT_MASK
&
1652 BITFIELD64_BIT(varying
))) {
1653 prog_data
->urb_setup
[varying
] = slot
- first_slot
;
1656 urb_next
= prev_stage_vue_map
.num_slots
- first_slot
;
1659 /* FINISHME: The sf doesn't map VS->FS inputs for us very well. */
1660 for (unsigned int i
= 0; i
< VARYING_SLOT_MAX
; i
++) {
1661 /* Point size is packed into the header, not as a general attribute */
1662 if (i
== VARYING_SLOT_PSIZ
)
1665 if (key
->input_slots_valid
& BITFIELD64_BIT(i
)) {
1666 /* The back color slot is skipped when the front color is
1667 * also written to. In addition, some slots can be
1668 * written in the vertex shader and not read in the
1669 * fragment shader. So the register number must always be
1670 * incremented, mapped or not.
1672 if (_mesa_varying_slot_in_fs((gl_varying_slot
) i
))
1673 prog_data
->urb_setup
[i
] = urb_next
;
1679 * It's a FS only attribute, and we did interpolation for this attribute
1680 * in SF thread. So, count it here, too.
1682 * See compile_sf_prog() for more info.
1684 if (nir
->info
.inputs_read
& BITFIELD64_BIT(VARYING_SLOT_PNTC
))
1685 prog_data
->urb_setup
[VARYING_SLOT_PNTC
] = urb_next
++;
1688 prog_data
->num_varying_inputs
= urb_next
;
1692 fs_visitor::assign_urb_setup()
1694 assert(stage
== MESA_SHADER_FRAGMENT
);
1695 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
1697 int urb_start
= payload
.num_regs
+ prog_data
->base
.curb_read_length
;
1699 /* Offset all the urb_setup[] index by the actual position of the
1700 * setup regs, now that the location of the constants has been chosen.
1702 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1703 if (inst
->opcode
== FS_OPCODE_LINTERP
) {
1704 assert(inst
->src
[1].file
== FIXED_GRF
);
1705 inst
->src
[1].nr
+= urb_start
;
1708 if (inst
->opcode
== FS_OPCODE_CINTERP
) {
1709 assert(inst
->src
[0].file
== FIXED_GRF
);
1710 inst
->src
[0].nr
+= urb_start
;
1714 /* Each attribute is 4 setup channels, each of which is half a reg. */
1715 this->first_non_payload_grf
+= prog_data
->num_varying_inputs
* 2;
1719 fs_visitor::convert_attr_sources_to_hw_regs(fs_inst
*inst
)
1721 for (int i
= 0; i
< inst
->sources
; i
++) {
1722 if (inst
->src
[i
].file
== ATTR
) {
1723 int grf
= payload
.num_regs
+
1724 prog_data
->curb_read_length
+
1726 inst
->src
[i
].reg_offset
;
1728 unsigned width
= inst
->src
[i
].stride
== 0 ? 1 : inst
->exec_size
;
1729 struct brw_reg reg
=
1730 stride(byte_offset(retype(brw_vec8_grf(grf
, 0), inst
->src
[i
].type
),
1731 inst
->src
[i
].subreg_offset
),
1732 inst
->exec_size
* inst
->src
[i
].stride
,
1733 width
, inst
->src
[i
].stride
);
1734 reg
.abs
= inst
->src
[i
].abs
;
1735 reg
.negate
= inst
->src
[i
].negate
;
1743 fs_visitor::assign_vs_urb_setup()
1745 brw_vs_prog_data
*vs_prog_data
= (brw_vs_prog_data
*) prog_data
;
1747 assert(stage
== MESA_SHADER_VERTEX
);
1749 /* Each attribute is 4 regs. */
1750 this->first_non_payload_grf
+= 4 * vs_prog_data
->nr_attributes
;
1752 assert(vs_prog_data
->base
.urb_read_length
<= 15);
1754 /* Rewrite all ATTR file references to the hw grf that they land in. */
1755 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1756 convert_attr_sources_to_hw_regs(inst
);
1761 fs_visitor::assign_tcs_single_patch_urb_setup()
1763 assert(stage
== MESA_SHADER_TESS_CTRL
);
1765 brw_vue_prog_data
*vue_prog_data
= (brw_vue_prog_data
*) prog_data
;
1767 /* Rewrite all ATTR file references to HW_REGs. */
1768 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1769 convert_attr_sources_to_hw_regs(inst
);
1774 fs_visitor::assign_tes_urb_setup()
1776 assert(stage
== MESA_SHADER_TESS_EVAL
);
1778 brw_vue_prog_data
*vue_prog_data
= (brw_vue_prog_data
*) prog_data
;
1780 first_non_payload_grf
+= 8 * vue_prog_data
->urb_read_length
;
1782 /* Rewrite all ATTR file references to HW_REGs. */
1783 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1784 convert_attr_sources_to_hw_regs(inst
);
1789 fs_visitor::assign_gs_urb_setup()
1791 assert(stage
== MESA_SHADER_GEOMETRY
);
1793 brw_vue_prog_data
*vue_prog_data
= (brw_vue_prog_data
*) prog_data
;
1795 first_non_payload_grf
+=
1796 8 * vue_prog_data
->urb_read_length
* nir
->info
.gs
.vertices_in
;
1798 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1799 /* Rewrite all ATTR file references to GRFs. */
1800 convert_attr_sources_to_hw_regs(inst
);
1806 * Split large virtual GRFs into separate components if we can.
1808 * This is mostly duplicated with what brw_fs_vector_splitting does,
1809 * but that's really conservative because it's afraid of doing
1810 * splitting that doesn't result in real progress after the rest of
1811 * the optimization phases, which would cause infinite looping in
1812 * optimization. We can do it once here, safely. This also has the
1813 * opportunity to split interpolated values, or maybe even uniforms,
1814 * which we don't have at the IR level.
1816 * We want to split, because virtual GRFs are what we register
1817 * allocate and spill (due to contiguousness requirements for some
1818 * instructions), and they're what we naturally generate in the
1819 * codegen process, but most virtual GRFs don't actually need to be
1820 * contiguous sets of GRFs. If we split, we'll end up with reduced
1821 * live intervals and better dead code elimination and coalescing.
1824 fs_visitor::split_virtual_grfs()
1826 int num_vars
= this->alloc
.count
;
1828 /* Count the total number of registers */
1830 int vgrf_to_reg
[num_vars
];
1831 for (int i
= 0; i
< num_vars
; i
++) {
1832 vgrf_to_reg
[i
] = reg_count
;
1833 reg_count
+= alloc
.sizes
[i
];
1836 /* An array of "split points". For each register slot, this indicates
1837 * if this slot can be separated from the previous slot. Every time an
1838 * instruction uses multiple elements of a register (as a source or
1839 * destination), we mark the used slots as inseparable. Then we go
1840 * through and split the registers into the smallest pieces we can.
1842 bool split_points
[reg_count
];
1843 memset(split_points
, 0, sizeof(split_points
));
1845 /* Mark all used registers as fully splittable */
1846 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1847 if (inst
->dst
.file
== VGRF
) {
1848 int reg
= vgrf_to_reg
[inst
->dst
.nr
];
1849 for (unsigned j
= 1; j
< this->alloc
.sizes
[inst
->dst
.nr
]; j
++)
1850 split_points
[reg
+ j
] = true;
1853 for (int i
= 0; i
< inst
->sources
; i
++) {
1854 if (inst
->src
[i
].file
== VGRF
) {
1855 int reg
= vgrf_to_reg
[inst
->src
[i
].nr
];
1856 for (unsigned j
= 1; j
< this->alloc
.sizes
[inst
->src
[i
].nr
]; j
++)
1857 split_points
[reg
+ j
] = true;
1862 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1863 if (inst
->dst
.file
== VGRF
) {
1864 int reg
= vgrf_to_reg
[inst
->dst
.nr
] + inst
->dst
.reg_offset
;
1865 for (int j
= 1; j
< inst
->regs_written
; j
++)
1866 split_points
[reg
+ j
] = false;
1868 for (int i
= 0; i
< inst
->sources
; i
++) {
1869 if (inst
->src
[i
].file
== VGRF
) {
1870 int reg
= vgrf_to_reg
[inst
->src
[i
].nr
] + inst
->src
[i
].reg_offset
;
1871 for (int j
= 1; j
< inst
->regs_read(i
); j
++)
1872 split_points
[reg
+ j
] = false;
1877 int new_virtual_grf
[reg_count
];
1878 int new_reg_offset
[reg_count
];
1881 for (int i
= 0; i
< num_vars
; i
++) {
1882 /* The first one should always be 0 as a quick sanity check. */
1883 assert(split_points
[reg
] == false);
1886 new_reg_offset
[reg
] = 0;
1891 for (unsigned j
= 1; j
< alloc
.sizes
[i
]; j
++) {
1892 /* If this is a split point, reset the offset to 0 and allocate a
1893 * new virtual GRF for the previous offset many registers
1895 if (split_points
[reg
]) {
1896 assert(offset
<= MAX_VGRF_SIZE
);
1897 int grf
= alloc
.allocate(offset
);
1898 for (int k
= reg
- offset
; k
< reg
; k
++)
1899 new_virtual_grf
[k
] = grf
;
1902 new_reg_offset
[reg
] = offset
;
1907 /* The last one gets the original register number */
1908 assert(offset
<= MAX_VGRF_SIZE
);
1909 alloc
.sizes
[i
] = offset
;
1910 for (int k
= reg
- offset
; k
< reg
; k
++)
1911 new_virtual_grf
[k
] = i
;
1913 assert(reg
== reg_count
);
1915 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1916 if (inst
->dst
.file
== VGRF
) {
1917 reg
= vgrf_to_reg
[inst
->dst
.nr
] + inst
->dst
.reg_offset
;
1918 inst
->dst
.nr
= new_virtual_grf
[reg
];
1919 inst
->dst
.reg_offset
= new_reg_offset
[reg
];
1920 assert((unsigned)new_reg_offset
[reg
] < alloc
.sizes
[new_virtual_grf
[reg
]]);
1922 for (int i
= 0; i
< inst
->sources
; i
++) {
1923 if (inst
->src
[i
].file
== VGRF
) {
1924 reg
= vgrf_to_reg
[inst
->src
[i
].nr
] + inst
->src
[i
].reg_offset
;
1925 inst
->src
[i
].nr
= new_virtual_grf
[reg
];
1926 inst
->src
[i
].reg_offset
= new_reg_offset
[reg
];
1927 assert((unsigned)new_reg_offset
[reg
] < alloc
.sizes
[new_virtual_grf
[reg
]]);
1931 invalidate_live_intervals();
1935 * Remove unused virtual GRFs and compact the virtual_grf_* arrays.
1937 * During code generation, we create tons of temporary variables, many of
1938 * which get immediately killed and are never used again. Yet, in later
1939 * optimization and analysis passes, such as compute_live_intervals, we need
1940 * to loop over all the virtual GRFs. Compacting them can save a lot of
1944 fs_visitor::compact_virtual_grfs()
1946 bool progress
= false;
1947 int remap_table
[this->alloc
.count
];
1948 memset(remap_table
, -1, sizeof(remap_table
));
1950 /* Mark which virtual GRFs are used. */
1951 foreach_block_and_inst(block
, const fs_inst
, inst
, cfg
) {
1952 if (inst
->dst
.file
== VGRF
)
1953 remap_table
[inst
->dst
.nr
] = 0;
1955 for (int i
= 0; i
< inst
->sources
; i
++) {
1956 if (inst
->src
[i
].file
== VGRF
)
1957 remap_table
[inst
->src
[i
].nr
] = 0;
1961 /* Compact the GRF arrays. */
1963 for (unsigned i
= 0; i
< this->alloc
.count
; i
++) {
1964 if (remap_table
[i
] == -1) {
1965 /* We just found an unused register. This means that we are
1966 * actually going to compact something.
1970 remap_table
[i
] = new_index
;
1971 alloc
.sizes
[new_index
] = alloc
.sizes
[i
];
1972 invalidate_live_intervals();
1977 this->alloc
.count
= new_index
;
1979 /* Patch all the instructions to use the newly renumbered registers */
1980 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1981 if (inst
->dst
.file
== VGRF
)
1982 inst
->dst
.nr
= remap_table
[inst
->dst
.nr
];
1984 for (int i
= 0; i
< inst
->sources
; i
++) {
1985 if (inst
->src
[i
].file
== VGRF
)
1986 inst
->src
[i
].nr
= remap_table
[inst
->src
[i
].nr
];
1990 /* Patch all the references to delta_xy, since they're used in register
1991 * allocation. If they're unused, switch them to BAD_FILE so we don't
1992 * think some random VGRF is delta_xy.
1994 for (unsigned i
= 0; i
< ARRAY_SIZE(delta_xy
); i
++) {
1995 if (delta_xy
[i
].file
== VGRF
) {
1996 if (remap_table
[delta_xy
[i
].nr
] != -1) {
1997 delta_xy
[i
].nr
= remap_table
[delta_xy
[i
].nr
];
1999 delta_xy
[i
].file
= BAD_FILE
;
2008 * Assign UNIFORM file registers to either push constants or pull constants.
2010 * We allow a fragment shader to have more than the specified minimum
2011 * maximum number of fragment shader uniform components (64). If
2012 * there are too many of these, they'd fill up all of register space.
2013 * So, this will push some of them out to the pull constant buffer and
2014 * update the program to load them.
2017 fs_visitor::assign_constant_locations()
2019 /* Only the first compile gets to decide on locations. */
2020 if (dispatch_width
!= min_dispatch_width
)
2023 bool is_live
[uniforms
];
2024 memset(is_live
, 0, sizeof(is_live
));
2026 /* For each uniform slot, a value of true indicates that the given slot and
2027 * the next slot must remain contiguous. This is used to keep us from
2028 * splitting arrays apart.
2030 bool contiguous
[uniforms
];
2031 memset(contiguous
, 0, sizeof(contiguous
));
2033 /* First, we walk through the instructions and do two things:
2035 * 1) Figure out which uniforms are live.
2037 * 2) Mark any indirectly used ranges of registers as contiguous.
2039 * Note that we don't move constant-indexed accesses to arrays. No
2040 * testing has been done of the performance impact of this choice.
2042 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
2043 for (int i
= 0 ; i
< inst
->sources
; i
++) {
2044 if (inst
->src
[i
].file
!= UNIFORM
)
2047 int constant_nr
= inst
->src
[i
].nr
+ inst
->src
[i
].reg_offset
;
2049 if (inst
->opcode
== SHADER_OPCODE_MOV_INDIRECT
&& i
== 0) {
2050 assert(inst
->src
[2].ud
% 4 == 0);
2051 unsigned last
= constant_nr
+ (inst
->src
[2].ud
/ 4) - 1;
2052 assert(last
< uniforms
);
2054 for (unsigned j
= constant_nr
; j
< last
; j
++) {
2056 contiguous
[j
] = true;
2058 is_live
[last
] = true;
2060 if (constant_nr
>= 0 && constant_nr
< (int) uniforms
)
2061 is_live
[constant_nr
] = true;
2066 /* Only allow 16 registers (128 uniform components) as push constants.
2068 * Just demote the end of the list. We could probably do better
2069 * here, demoting things that are rarely used in the program first.
2071 * If changing this value, note the limitation about total_regs in
2074 const unsigned int max_push_components
= 16 * 8;
2076 /* We push small arrays, but no bigger than 16 floats. This is big enough
2077 * for a vec4 but hopefully not large enough to push out other stuff. We
2078 * should probably use a better heuristic at some point.
2080 const unsigned int max_chunk_size
= 16;
2082 unsigned int num_push_constants
= 0;
2083 unsigned int num_pull_constants
= 0;
2085 push_constant_loc
= ralloc_array(mem_ctx
, int, uniforms
);
2086 pull_constant_loc
= ralloc_array(mem_ctx
, int, uniforms
);
2088 int chunk_start
= -1;
2089 for (unsigned u
= 0; u
< uniforms
; u
++) {
2090 push_constant_loc
[u
] = -1;
2091 pull_constant_loc
[u
] = -1;
2096 /* This is the first live uniform in the chunk */
2097 if (chunk_start
< 0)
2100 /* If this element does not need to be contiguous with the next, we
2101 * split at this point and everthing between chunk_start and u forms a
2104 if (!contiguous
[u
]) {
2105 unsigned chunk_size
= u
- chunk_start
+ 1;
2107 /* Decide whether we should push or pull this parameter. In the
2108 * Vulkan driver, push constants are explicitly exposed via the API
2109 * so we push everything. In GL, we only push small arrays.
2111 if (stage_prog_data
->pull_param
== NULL
||
2112 (num_push_constants
+ chunk_size
<= max_push_components
&&
2113 chunk_size
<= max_chunk_size
)) {
2114 assert(num_push_constants
+ chunk_size
<= max_push_components
);
2115 for (unsigned j
= chunk_start
; j
<= u
; j
++)
2116 push_constant_loc
[j
] = num_push_constants
++;
2118 for (unsigned j
= chunk_start
; j
<= u
; j
++)
2119 pull_constant_loc
[j
] = num_pull_constants
++;
2126 stage_prog_data
->nr_params
= num_push_constants
;
2127 stage_prog_data
->nr_pull_params
= num_pull_constants
;
2129 /* Up until now, the param[] array has been indexed by reg + reg_offset
2130 * of UNIFORM registers. Move pull constants into pull_param[] and
2131 * condense param[] to only contain the uniforms we chose to push.
2133 * NOTE: Because we are condensing the params[] array, we know that
2134 * push_constant_loc[i] <= i and we can do it in one smooth loop without
2135 * having to make a copy.
2137 for (unsigned int i
= 0; i
< uniforms
; i
++) {
2138 const gl_constant_value
*value
= stage_prog_data
->param
[i
];
2140 if (pull_constant_loc
[i
] != -1) {
2141 stage_prog_data
->pull_param
[pull_constant_loc
[i
]] = value
;
2142 } else if (push_constant_loc
[i
] != -1) {
2143 stage_prog_data
->param
[push_constant_loc
[i
]] = value
;
2149 * Replace UNIFORM register file access with either UNIFORM_PULL_CONSTANT_LOAD
2150 * or VARYING_PULL_CONSTANT_LOAD instructions which load values into VGRFs.
2153 fs_visitor::lower_constant_loads()
2155 const unsigned index
= stage_prog_data
->binding_table
.pull_constants_start
;
2157 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
2158 /* Set up the annotation tracking for new generated instructions. */
2159 const fs_builder
ibld(this, block
, inst
);
2161 for (int i
= 0; i
< inst
->sources
; i
++) {
2162 if (inst
->src
[i
].file
!= UNIFORM
)
2165 /* We'll handle this case later */
2166 if (inst
->opcode
== SHADER_OPCODE_MOV_INDIRECT
&& i
== 0)
2169 unsigned location
= inst
->src
[i
].nr
+ inst
->src
[i
].reg_offset
;
2170 if (location
>= uniforms
)
2171 continue; /* Out of bounds access */
2173 int pull_index
= pull_constant_loc
[location
];
2175 if (pull_index
== -1)
2178 assert(inst
->src
[i
].stride
== 0);
2180 fs_reg dst
= vgrf(glsl_type::float_type
);
2181 const fs_builder ubld
= ibld
.exec_all().group(8, 0);
2182 struct brw_reg offset
= brw_imm_ud((unsigned)(pull_index
* 4) & ~15);
2183 ubld
.emit(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
,
2184 dst
, brw_imm_ud(index
), offset
);
2186 /* Rewrite the instruction to use the temporary VGRF. */
2187 inst
->src
[i
].file
= VGRF
;
2188 inst
->src
[i
].nr
= dst
.nr
;
2189 inst
->src
[i
].reg_offset
= 0;
2190 inst
->src
[i
].set_smear(pull_index
& 3);
2192 brw_mark_surface_used(prog_data
, index
);
2195 if (inst
->opcode
== SHADER_OPCODE_MOV_INDIRECT
&&
2196 inst
->src
[0].file
== UNIFORM
) {
2198 unsigned location
= inst
->src
[0].nr
+ inst
->src
[0].reg_offset
;
2199 if (location
>= uniforms
)
2200 continue; /* Out of bounds access */
2202 int pull_index
= pull_constant_loc
[location
];
2204 if (pull_index
== -1)
2207 VARYING_PULL_CONSTANT_LOAD(ibld
, inst
->dst
,
2211 inst
->remove(block
);
2213 brw_mark_surface_used(prog_data
, index
);
2216 invalidate_live_intervals();
2220 fs_visitor::opt_algebraic()
2222 bool progress
= false;
2224 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2225 switch (inst
->opcode
) {
2226 case BRW_OPCODE_MOV
:
2227 if (inst
->src
[0].file
!= IMM
)
2230 if (inst
->saturate
) {
2231 if (inst
->dst
.type
!= inst
->src
[0].type
)
2232 assert(!"unimplemented: saturate mixed types");
2234 if (brw_saturate_immediate(inst
->dst
.type
,
2235 &inst
->src
[0].as_brw_reg())) {
2236 inst
->saturate
= false;
2242 case BRW_OPCODE_MUL
:
2243 if (inst
->src
[1].file
!= IMM
)
2247 if (inst
->src
[1].is_one()) {
2248 inst
->opcode
= BRW_OPCODE_MOV
;
2249 inst
->src
[1] = reg_undef
;
2255 if (inst
->src
[1].is_negative_one()) {
2256 inst
->opcode
= BRW_OPCODE_MOV
;
2257 inst
->src
[0].negate
= !inst
->src
[0].negate
;
2258 inst
->src
[1] = reg_undef
;
2264 if (inst
->src
[1].is_zero()) {
2265 inst
->opcode
= BRW_OPCODE_MOV
;
2266 inst
->src
[0] = inst
->src
[1];
2267 inst
->src
[1] = reg_undef
;
2272 if (inst
->src
[0].file
== IMM
) {
2273 assert(inst
->src
[0].type
== BRW_REGISTER_TYPE_F
);
2274 inst
->opcode
= BRW_OPCODE_MOV
;
2275 inst
->src
[0].f
*= inst
->src
[1].f
;
2276 inst
->src
[1] = reg_undef
;
2281 case BRW_OPCODE_ADD
:
2282 if (inst
->src
[1].file
!= IMM
)
2286 if (inst
->src
[1].is_zero()) {
2287 inst
->opcode
= BRW_OPCODE_MOV
;
2288 inst
->src
[1] = reg_undef
;
2293 if (inst
->src
[0].file
== IMM
) {
2294 assert(inst
->src
[0].type
== BRW_REGISTER_TYPE_F
);
2295 inst
->opcode
= BRW_OPCODE_MOV
;
2296 inst
->src
[0].f
+= inst
->src
[1].f
;
2297 inst
->src
[1] = reg_undef
;
2303 if (inst
->src
[0].equals(inst
->src
[1])) {
2304 inst
->opcode
= BRW_OPCODE_MOV
;
2305 inst
->src
[1] = reg_undef
;
2310 case BRW_OPCODE_LRP
:
2311 if (inst
->src
[1].equals(inst
->src
[2])) {
2312 inst
->opcode
= BRW_OPCODE_MOV
;
2313 inst
->src
[0] = inst
->src
[1];
2314 inst
->src
[1] = reg_undef
;
2315 inst
->src
[2] = reg_undef
;
2320 case BRW_OPCODE_CMP
:
2321 if (inst
->conditional_mod
== BRW_CONDITIONAL_GE
&&
2323 inst
->src
[0].negate
&&
2324 inst
->src
[1].is_zero()) {
2325 inst
->src
[0].abs
= false;
2326 inst
->src
[0].negate
= false;
2327 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
2332 case BRW_OPCODE_SEL
:
2333 if (inst
->src
[0].equals(inst
->src
[1])) {
2334 inst
->opcode
= BRW_OPCODE_MOV
;
2335 inst
->src
[1] = reg_undef
;
2336 inst
->predicate
= BRW_PREDICATE_NONE
;
2337 inst
->predicate_inverse
= false;
2339 } else if (inst
->saturate
&& inst
->src
[1].file
== IMM
) {
2340 switch (inst
->conditional_mod
) {
2341 case BRW_CONDITIONAL_LE
:
2342 case BRW_CONDITIONAL_L
:
2343 switch (inst
->src
[1].type
) {
2344 case BRW_REGISTER_TYPE_F
:
2345 if (inst
->src
[1].f
>= 1.0f
) {
2346 inst
->opcode
= BRW_OPCODE_MOV
;
2347 inst
->src
[1] = reg_undef
;
2348 inst
->conditional_mod
= BRW_CONDITIONAL_NONE
;
2356 case BRW_CONDITIONAL_GE
:
2357 case BRW_CONDITIONAL_G
:
2358 switch (inst
->src
[1].type
) {
2359 case BRW_REGISTER_TYPE_F
:
2360 if (inst
->src
[1].f
<= 0.0f
) {
2361 inst
->opcode
= BRW_OPCODE_MOV
;
2362 inst
->src
[1] = reg_undef
;
2363 inst
->conditional_mod
= BRW_CONDITIONAL_NONE
;
2375 case BRW_OPCODE_MAD
:
2376 if (inst
->src
[1].is_zero() || inst
->src
[2].is_zero()) {
2377 inst
->opcode
= BRW_OPCODE_MOV
;
2378 inst
->src
[1] = reg_undef
;
2379 inst
->src
[2] = reg_undef
;
2381 } else if (inst
->src
[0].is_zero()) {
2382 inst
->opcode
= BRW_OPCODE_MUL
;
2383 inst
->src
[0] = inst
->src
[2];
2384 inst
->src
[2] = reg_undef
;
2386 } else if (inst
->src
[1].is_one()) {
2387 inst
->opcode
= BRW_OPCODE_ADD
;
2388 inst
->src
[1] = inst
->src
[2];
2389 inst
->src
[2] = reg_undef
;
2391 } else if (inst
->src
[2].is_one()) {
2392 inst
->opcode
= BRW_OPCODE_ADD
;
2393 inst
->src
[2] = reg_undef
;
2395 } else if (inst
->src
[1].file
== IMM
&& inst
->src
[2].file
== IMM
) {
2396 inst
->opcode
= BRW_OPCODE_ADD
;
2397 inst
->src
[1].f
*= inst
->src
[2].f
;
2398 inst
->src
[2] = reg_undef
;
2402 case SHADER_OPCODE_BROADCAST
:
2403 if (is_uniform(inst
->src
[0])) {
2404 inst
->opcode
= BRW_OPCODE_MOV
;
2406 inst
->force_writemask_all
= true;
2408 } else if (inst
->src
[1].file
== IMM
) {
2409 inst
->opcode
= BRW_OPCODE_MOV
;
2410 inst
->src
[0] = component(inst
->src
[0],
2413 inst
->force_writemask_all
= true;
2422 /* Swap if src[0] is immediate. */
2423 if (progress
&& inst
->is_commutative()) {
2424 if (inst
->src
[0].file
== IMM
) {
2425 fs_reg tmp
= inst
->src
[1];
2426 inst
->src
[1] = inst
->src
[0];
2435 * Optimize sample messages that have constant zero values for the trailing
2436 * texture coordinates. We can just reduce the message length for these
2437 * instructions instead of reserving a register for it. Trailing parameters
2438 * that aren't sent default to zero anyway. This will cause the dead code
2439 * eliminator to remove the MOV instruction that would otherwise be emitted to
2440 * set up the zero value.
2443 fs_visitor::opt_zero_samples()
2445 /* Gen4 infers the texturing opcode based on the message length so we can't
2448 if (devinfo
->gen
< 5)
2451 bool progress
= false;
2453 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2454 if (!inst
->is_tex())
2457 fs_inst
*load_payload
= (fs_inst
*) inst
->prev
;
2459 if (load_payload
->is_head_sentinel() ||
2460 load_payload
->opcode
!= SHADER_OPCODE_LOAD_PAYLOAD
)
2463 /* We don't want to remove the message header or the first parameter.
2464 * Removing the first parameter is not allowed, see the Haswell PRM
2465 * volume 7, page 149:
2467 * "Parameter 0 is required except for the sampleinfo message, which
2468 * has no parameter 0"
2470 while (inst
->mlen
> inst
->header_size
+ inst
->exec_size
/ 8 &&
2471 load_payload
->src
[(inst
->mlen
- inst
->header_size
) /
2472 (inst
->exec_size
/ 8) +
2473 inst
->header_size
- 1].is_zero()) {
2474 inst
->mlen
-= inst
->exec_size
/ 8;
2480 invalidate_live_intervals();
2486 * Optimize sample messages which are followed by the final RT write.
2488 * CHV, and GEN9+ can mark a texturing SEND instruction with EOT to have its
2489 * results sent directly to the framebuffer, bypassing the EU. Recognize the
2490 * final texturing results copied to the framebuffer write payload and modify
2491 * them to write to the framebuffer directly.
2494 fs_visitor::opt_sampler_eot()
2496 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
2498 if (stage
!= MESA_SHADER_FRAGMENT
)
2501 if (devinfo
->gen
< 9 && !devinfo
->is_cherryview
)
2504 /* FINISHME: It should be possible to implement this optimization when there
2505 * are multiple drawbuffers.
2507 if (key
->nr_color_regions
!= 1)
2510 /* Look for a texturing instruction immediately before the final FB_WRITE. */
2511 bblock_t
*block
= cfg
->blocks
[cfg
->num_blocks
- 1];
2512 fs_inst
*fb_write
= (fs_inst
*)block
->end();
2513 assert(fb_write
->eot
);
2514 assert(fb_write
->opcode
== FS_OPCODE_FB_WRITE
);
2516 fs_inst
*tex_inst
= (fs_inst
*) fb_write
->prev
;
2518 /* There wasn't one; nothing to do. */
2519 if (unlikely(tex_inst
->is_head_sentinel()) || !tex_inst
->is_tex())
2522 /* 3D Sampler » Messages » Message Format
2524 * “Response Length of zero is allowed on all SIMD8* and SIMD16* sampler
2525 * messages except sample+killpix, resinfo, sampleinfo, LOD, and gather4*”
2527 if (tex_inst
->opcode
== SHADER_OPCODE_TXS
||
2528 tex_inst
->opcode
== SHADER_OPCODE_SAMPLEINFO
||
2529 tex_inst
->opcode
== SHADER_OPCODE_LOD
||
2530 tex_inst
->opcode
== SHADER_OPCODE_TG4
||
2531 tex_inst
->opcode
== SHADER_OPCODE_TG4_OFFSET
)
2534 /* If there's no header present, we need to munge the LOAD_PAYLOAD as well.
2535 * It's very likely to be the previous instruction.
2537 fs_inst
*load_payload
= (fs_inst
*) tex_inst
->prev
;
2538 if (load_payload
->is_head_sentinel() ||
2539 load_payload
->opcode
!= SHADER_OPCODE_LOAD_PAYLOAD
)
2542 assert(!tex_inst
->eot
); /* We can't get here twice */
2543 assert((tex_inst
->offset
& (0xff << 24)) == 0);
2545 const fs_builder
ibld(this, block
, tex_inst
);
2547 tex_inst
->offset
|= fb_write
->target
<< 24;
2548 tex_inst
->eot
= true;
2549 tex_inst
->dst
= ibld
.null_reg_ud();
2550 tex_inst
->regs_written
= 0;
2551 fb_write
->remove(cfg
->blocks
[cfg
->num_blocks
- 1]);
2553 /* If a header is present, marking the eot is sufficient. Otherwise, we need
2554 * to create a new LOAD_PAYLOAD command with the same sources and a space
2555 * saved for the header. Using a new destination register not only makes sure
2556 * we have enough space, but it will make sure the dead code eliminator kills
2557 * the instruction that this will replace.
2559 if (tex_inst
->header_size
!= 0) {
2560 invalidate_live_intervals();
2564 fs_reg send_header
= ibld
.vgrf(BRW_REGISTER_TYPE_F
,
2565 load_payload
->sources
+ 1);
2566 fs_reg
*new_sources
=
2567 ralloc_array(mem_ctx
, fs_reg
, load_payload
->sources
+ 1);
2569 new_sources
[0] = fs_reg();
2570 for (int i
= 0; i
< load_payload
->sources
; i
++)
2571 new_sources
[i
+1] = load_payload
->src
[i
];
2573 /* The LOAD_PAYLOAD helper seems like the obvious choice here. However, it
2574 * requires a lot of information about the sources to appropriately figure
2575 * out the number of registers needed to be used. Given this stage in our
2576 * optimization, we may not have the appropriate GRFs required by
2577 * LOAD_PAYLOAD at this point (copy propagation). Therefore, we need to
2578 * manually emit the instruction.
2580 fs_inst
*new_load_payload
= new(mem_ctx
) fs_inst(SHADER_OPCODE_LOAD_PAYLOAD
,
2581 load_payload
->exec_size
,
2584 load_payload
->sources
+ 1);
2586 new_load_payload
->regs_written
= load_payload
->regs_written
+ 1;
2587 new_load_payload
->header_size
= 1;
2589 tex_inst
->header_size
= 1;
2590 tex_inst
->insert_before(cfg
->blocks
[cfg
->num_blocks
- 1], new_load_payload
);
2591 tex_inst
->src
[0] = send_header
;
2593 invalidate_live_intervals();
2598 fs_visitor::opt_register_renaming()
2600 bool progress
= false;
2603 int remap
[alloc
.count
];
2604 memset(remap
, -1, sizeof(int) * alloc
.count
);
2606 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2607 if (inst
->opcode
== BRW_OPCODE_IF
|| inst
->opcode
== BRW_OPCODE_DO
) {
2609 } else if (inst
->opcode
== BRW_OPCODE_ENDIF
||
2610 inst
->opcode
== BRW_OPCODE_WHILE
) {
2614 /* Rewrite instruction sources. */
2615 for (int i
= 0; i
< inst
->sources
; i
++) {
2616 if (inst
->src
[i
].file
== VGRF
&&
2617 remap
[inst
->src
[i
].nr
] != -1 &&
2618 remap
[inst
->src
[i
].nr
] != inst
->src
[i
].nr
) {
2619 inst
->src
[i
].nr
= remap
[inst
->src
[i
].nr
];
2624 const int dst
= inst
->dst
.nr
;
2627 inst
->dst
.file
== VGRF
&&
2628 alloc
.sizes
[inst
->dst
.nr
] == inst
->exec_size
/ 8 &&
2629 !inst
->is_partial_write()) {
2630 if (remap
[dst
] == -1) {
2633 remap
[dst
] = alloc
.allocate(inst
->exec_size
/ 8);
2634 inst
->dst
.nr
= remap
[dst
];
2637 } else if (inst
->dst
.file
== VGRF
&&
2639 remap
[dst
] != dst
) {
2640 inst
->dst
.nr
= remap
[dst
];
2646 invalidate_live_intervals();
2648 for (unsigned i
= 0; i
< ARRAY_SIZE(delta_xy
); i
++) {
2649 if (delta_xy
[i
].file
== VGRF
&& remap
[delta_xy
[i
].nr
] != -1) {
2650 delta_xy
[i
].nr
= remap
[delta_xy
[i
].nr
];
2659 * Remove redundant or useless discard jumps.
2661 * For example, we can eliminate jumps in the following sequence:
2663 * discard-jump (redundant with the next jump)
2664 * discard-jump (useless; jumps to the next instruction)
2668 fs_visitor::opt_redundant_discard_jumps()
2670 bool progress
= false;
2672 bblock_t
*last_bblock
= cfg
->blocks
[cfg
->num_blocks
- 1];
2674 fs_inst
*placeholder_halt
= NULL
;
2675 foreach_inst_in_block_reverse(fs_inst
, inst
, last_bblock
) {
2676 if (inst
->opcode
== FS_OPCODE_PLACEHOLDER_HALT
) {
2677 placeholder_halt
= inst
;
2682 if (!placeholder_halt
)
2685 /* Delete any HALTs immediately before the placeholder halt. */
2686 for (fs_inst
*prev
= (fs_inst
*) placeholder_halt
->prev
;
2687 !prev
->is_head_sentinel() && prev
->opcode
== FS_OPCODE_DISCARD_JUMP
;
2688 prev
= (fs_inst
*) placeholder_halt
->prev
) {
2689 prev
->remove(last_bblock
);
2694 invalidate_live_intervals();
2700 fs_visitor::compute_to_mrf()
2702 bool progress
= false;
2705 /* No MRFs on Gen >= 7. */
2706 if (devinfo
->gen
>= 7)
2709 calculate_live_intervals();
2711 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
2715 if (inst
->opcode
!= BRW_OPCODE_MOV
||
2716 inst
->is_partial_write() ||
2717 inst
->dst
.file
!= MRF
|| inst
->src
[0].file
!= VGRF
||
2718 inst
->dst
.type
!= inst
->src
[0].type
||
2719 inst
->src
[0].abs
|| inst
->src
[0].negate
||
2720 !inst
->src
[0].is_contiguous() ||
2721 inst
->src
[0].subreg_offset
)
2724 /* Work out which hardware MRF registers are written by this
2727 int mrf_low
= inst
->dst
.nr
& ~BRW_MRF_COMPR4
;
2729 if (inst
->dst
.nr
& BRW_MRF_COMPR4
) {
2730 mrf_high
= mrf_low
+ 4;
2731 } else if (inst
->exec_size
== 16) {
2732 mrf_high
= mrf_low
+ 1;
2737 /* Can't compute-to-MRF this GRF if someone else was going to
2740 if (this->virtual_grf_end
[inst
->src
[0].nr
] > ip
)
2743 /* Found a move of a GRF to a MRF. Let's see if we can go
2744 * rewrite the thing that made this GRF to write into the MRF.
2746 foreach_inst_in_block_reverse_starting_from(fs_inst
, scan_inst
, inst
) {
2747 if (scan_inst
->dst
.file
== VGRF
&&
2748 scan_inst
->dst
.nr
== inst
->src
[0].nr
) {
2749 /* Found the last thing to write our reg we want to turn
2750 * into a compute-to-MRF.
2753 /* If this one instruction didn't populate all the
2754 * channels, bail. We might be able to rewrite everything
2755 * that writes that reg, but it would require smarter
2756 * tracking to delay the rewriting until complete success.
2758 if (scan_inst
->is_partial_write())
2761 /* Things returning more than one register would need us to
2762 * understand coalescing out more than one MOV at a time.
2764 if (scan_inst
->regs_written
> scan_inst
->exec_size
/ 8)
2767 /* SEND instructions can't have MRF as a destination. */
2768 if (scan_inst
->mlen
)
2771 if (devinfo
->gen
== 6) {
2772 /* gen6 math instructions must have the destination be
2773 * GRF, so no compute-to-MRF for them.
2775 if (scan_inst
->is_math()) {
2780 if (scan_inst
->dst
.reg_offset
== inst
->src
[0].reg_offset
) {
2781 /* Found the creator of our MRF's source value. */
2782 scan_inst
->dst
.file
= MRF
;
2783 scan_inst
->dst
.nr
= inst
->dst
.nr
;
2784 scan_inst
->saturate
|= inst
->saturate
;
2785 inst
->remove(block
);
2791 /* We don't handle control flow here. Most computation of
2792 * values that end up in MRFs are shortly before the MRF
2795 if (block
->start() == scan_inst
)
2798 /* You can't read from an MRF, so if someone else reads our
2799 * MRF's source GRF that we wanted to rewrite, that stops us.
2801 bool interfered
= false;
2802 for (int i
= 0; i
< scan_inst
->sources
; i
++) {
2803 if (scan_inst
->src
[i
].file
== VGRF
&&
2804 scan_inst
->src
[i
].nr
== inst
->src
[0].nr
&&
2805 scan_inst
->src
[i
].reg_offset
== inst
->src
[0].reg_offset
) {
2812 if (scan_inst
->dst
.file
== MRF
) {
2813 /* If somebody else writes our MRF here, we can't
2814 * compute-to-MRF before that.
2816 int scan_mrf_low
= scan_inst
->dst
.nr
& ~BRW_MRF_COMPR4
;
2819 if (scan_inst
->dst
.nr
& BRW_MRF_COMPR4
) {
2820 scan_mrf_high
= scan_mrf_low
+ 4;
2821 } else if (scan_inst
->exec_size
== 16) {
2822 scan_mrf_high
= scan_mrf_low
+ 1;
2824 scan_mrf_high
= scan_mrf_low
;
2827 if (mrf_low
== scan_mrf_low
||
2828 mrf_low
== scan_mrf_high
||
2829 mrf_high
== scan_mrf_low
||
2830 mrf_high
== scan_mrf_high
) {
2835 if (scan_inst
->mlen
> 0 && scan_inst
->base_mrf
!= -1) {
2836 /* Found a SEND instruction, which means that there are
2837 * live values in MRFs from base_mrf to base_mrf +
2838 * scan_inst->mlen - 1. Don't go pushing our MRF write up
2841 if (mrf_low
>= scan_inst
->base_mrf
&&
2842 mrf_low
< scan_inst
->base_mrf
+ scan_inst
->mlen
) {
2845 if (mrf_high
>= scan_inst
->base_mrf
&&
2846 mrf_high
< scan_inst
->base_mrf
+ scan_inst
->mlen
) {
2854 invalidate_live_intervals();
2860 * Eliminate FIND_LIVE_CHANNEL instructions occurring outside any control
2861 * flow. We could probably do better here with some form of divergence
2865 fs_visitor::eliminate_find_live_channel()
2867 bool progress
= false;
2870 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
2871 switch (inst
->opcode
) {
2877 case BRW_OPCODE_ENDIF
:
2878 case BRW_OPCODE_WHILE
:
2882 case FS_OPCODE_DISCARD_JUMP
:
2883 /* This can potentially make control flow non-uniform until the end
2888 case SHADER_OPCODE_FIND_LIVE_CHANNEL
:
2890 inst
->opcode
= BRW_OPCODE_MOV
;
2891 inst
->src
[0] = brw_imm_ud(0u);
2893 inst
->force_writemask_all
= true;
2907 * Once we've generated code, try to convert normal FS_OPCODE_FB_WRITE
2908 * instructions to FS_OPCODE_REP_FB_WRITE.
2911 fs_visitor::emit_repclear_shader()
2913 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
2915 int color_mrf
= base_mrf
+ 2;
2919 mov
= bld
.exec_all().group(4, 0)
2920 .MOV(brw_message_reg(color_mrf
),
2921 fs_reg(UNIFORM
, 0, BRW_REGISTER_TYPE_F
));
2923 struct brw_reg reg
=
2924 brw_reg(BRW_GENERAL_REGISTER_FILE
, 2, 3, 0, 0, BRW_REGISTER_TYPE_F
,
2925 BRW_VERTICAL_STRIDE_8
, BRW_WIDTH_2
, BRW_HORIZONTAL_STRIDE_4
,
2926 BRW_SWIZZLE_XYZW
, WRITEMASK_XYZW
);
2928 mov
= bld
.exec_all().group(4, 0)
2929 .MOV(vec4(brw_message_reg(color_mrf
)), fs_reg(reg
));
2933 if (key
->nr_color_regions
== 1) {
2934 write
= bld
.emit(FS_OPCODE_REP_FB_WRITE
);
2935 write
->saturate
= key
->clamp_fragment_color
;
2936 write
->base_mrf
= color_mrf
;
2938 write
->header_size
= 0;
2941 assume(key
->nr_color_regions
> 0);
2942 for (int i
= 0; i
< key
->nr_color_regions
; ++i
) {
2943 write
= bld
.emit(FS_OPCODE_REP_FB_WRITE
);
2944 write
->saturate
= key
->clamp_fragment_color
;
2945 write
->base_mrf
= base_mrf
;
2947 write
->header_size
= 2;
2955 assign_constant_locations();
2956 assign_curb_setup();
2958 /* Now that we have the uniform assigned, go ahead and force it to a vec4. */
2960 assert(mov
->src
[0].file
== FIXED_GRF
);
2961 mov
->src
[0] = brw_vec4_grf(mov
->src
[0].nr
, 0);
2966 * Walks through basic blocks, looking for repeated MRF writes and
2967 * removing the later ones.
2970 fs_visitor::remove_duplicate_mrf_writes()
2972 fs_inst
*last_mrf_move
[BRW_MAX_MRF(devinfo
->gen
)];
2973 bool progress
= false;
2975 /* Need to update the MRF tracking for compressed instructions. */
2976 if (dispatch_width
== 16)
2979 memset(last_mrf_move
, 0, sizeof(last_mrf_move
));
2981 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
2982 if (inst
->is_control_flow()) {
2983 memset(last_mrf_move
, 0, sizeof(last_mrf_move
));
2986 if (inst
->opcode
== BRW_OPCODE_MOV
&&
2987 inst
->dst
.file
== MRF
) {
2988 fs_inst
*prev_inst
= last_mrf_move
[inst
->dst
.nr
];
2989 if (prev_inst
&& inst
->equals(prev_inst
)) {
2990 inst
->remove(block
);
2996 /* Clear out the last-write records for MRFs that were overwritten. */
2997 if (inst
->dst
.file
== MRF
) {
2998 last_mrf_move
[inst
->dst
.nr
] = NULL
;
3001 if (inst
->mlen
> 0 && inst
->base_mrf
!= -1) {
3002 /* Found a SEND instruction, which will include two or fewer
3003 * implied MRF writes. We could do better here.
3005 for (int i
= 0; i
< implied_mrf_writes(inst
); i
++) {
3006 last_mrf_move
[inst
->base_mrf
+ i
] = NULL
;
3010 /* Clear out any MRF move records whose sources got overwritten. */
3011 if (inst
->dst
.file
== VGRF
) {
3012 for (unsigned int i
= 0; i
< ARRAY_SIZE(last_mrf_move
); i
++) {
3013 if (last_mrf_move
[i
] &&
3014 last_mrf_move
[i
]->src
[0].nr
== inst
->dst
.nr
) {
3015 last_mrf_move
[i
] = NULL
;
3020 if (inst
->opcode
== BRW_OPCODE_MOV
&&
3021 inst
->dst
.file
== MRF
&&
3022 inst
->src
[0].file
== VGRF
&&
3023 !inst
->is_partial_write()) {
3024 last_mrf_move
[inst
->dst
.nr
] = inst
;
3029 invalidate_live_intervals();
3035 clear_deps_for_inst_src(fs_inst
*inst
, bool *deps
, int first_grf
, int grf_len
)
3037 /* Clear the flag for registers that actually got read (as expected). */
3038 for (int i
= 0; i
< inst
->sources
; i
++) {
3040 if (inst
->src
[i
].file
== VGRF
|| inst
->src
[i
].file
== FIXED_GRF
) {
3041 grf
= inst
->src
[i
].nr
;
3046 if (grf
>= first_grf
&&
3047 grf
< first_grf
+ grf_len
) {
3048 deps
[grf
- first_grf
] = false;
3049 if (inst
->exec_size
== 16)
3050 deps
[grf
- first_grf
+ 1] = false;
3056 * Implements this workaround for the original 965:
3058 * "[DevBW, DevCL] Implementation Restrictions: As the hardware does not
3059 * check for post destination dependencies on this instruction, software
3060 * must ensure that there is no destination hazard for the case of ‘write
3061 * followed by a posted write’ shown in the following example.
3064 * 2. send r3.xy <rest of send instruction>
3067 * Due to no post-destination dependency check on the ‘send’, the above
3068 * code sequence could have two instructions (1 and 2) in flight at the
3069 * same time that both consider ‘r3’ as the target of their final writes.
3072 fs_visitor::insert_gen4_pre_send_dependency_workarounds(bblock_t
*block
,
3075 int write_len
= inst
->regs_written
;
3076 int first_write_grf
= inst
->dst
.nr
;
3077 bool needs_dep
[BRW_MAX_MRF(devinfo
->gen
)];
3078 assert(write_len
< (int)sizeof(needs_dep
) - 1);
3080 memset(needs_dep
, false, sizeof(needs_dep
));
3081 memset(needs_dep
, true, write_len
);
3083 clear_deps_for_inst_src(inst
, needs_dep
, first_write_grf
, write_len
);
3085 /* Walk backwards looking for writes to registers we're writing which
3086 * aren't read since being written. If we hit the start of the program,
3087 * we assume that there are no outstanding dependencies on entry to the
3090 foreach_inst_in_block_reverse_starting_from(fs_inst
, scan_inst
, inst
) {
3091 /* If we hit control flow, assume that there *are* outstanding
3092 * dependencies, and force their cleanup before our instruction.
3094 if (block
->start() == scan_inst
) {
3095 for (int i
= 0; i
< write_len
; i
++) {
3097 DEP_RESOLVE_MOV(fs_builder(this, block
, inst
),
3098 first_write_grf
+ i
);
3103 /* We insert our reads as late as possible on the assumption that any
3104 * instruction but a MOV that might have left us an outstanding
3105 * dependency has more latency than a MOV.
3107 if (scan_inst
->dst
.file
== VGRF
) {
3108 for (int i
= 0; i
< scan_inst
->regs_written
; i
++) {
3109 int reg
= scan_inst
->dst
.nr
+ i
;
3111 if (reg
>= first_write_grf
&&
3112 reg
< first_write_grf
+ write_len
&&
3113 needs_dep
[reg
- first_write_grf
]) {
3114 DEP_RESOLVE_MOV(fs_builder(this, block
, inst
), reg
);
3115 needs_dep
[reg
- first_write_grf
] = false;
3116 if (scan_inst
->exec_size
== 16)
3117 needs_dep
[reg
- first_write_grf
+ 1] = false;
3122 /* Clear the flag for registers that actually got read (as expected). */
3123 clear_deps_for_inst_src(scan_inst
, needs_dep
, first_write_grf
, write_len
);
3125 /* Continue the loop only if we haven't resolved all the dependencies */
3127 for (i
= 0; i
< write_len
; i
++) {
3137 * Implements this workaround for the original 965:
3139 * "[DevBW, DevCL] Errata: A destination register from a send can not be
3140 * used as a destination register until after it has been sourced by an
3141 * instruction with a different destination register.
3144 fs_visitor::insert_gen4_post_send_dependency_workarounds(bblock_t
*block
, fs_inst
*inst
)
3146 int write_len
= inst
->regs_written
;
3147 int first_write_grf
= inst
->dst
.nr
;
3148 bool needs_dep
[BRW_MAX_MRF(devinfo
->gen
)];
3149 assert(write_len
< (int)sizeof(needs_dep
) - 1);
3151 memset(needs_dep
, false, sizeof(needs_dep
));
3152 memset(needs_dep
, true, write_len
);
3153 /* Walk forwards looking for writes to registers we're writing which aren't
3154 * read before being written.
3156 foreach_inst_in_block_starting_from(fs_inst
, scan_inst
, inst
) {
3157 /* If we hit control flow, force resolve all remaining dependencies. */
3158 if (block
->end() == scan_inst
) {
3159 for (int i
= 0; i
< write_len
; i
++) {
3161 DEP_RESOLVE_MOV(fs_builder(this, block
, scan_inst
),
3162 first_write_grf
+ i
);
3167 /* Clear the flag for registers that actually got read (as expected). */
3168 clear_deps_for_inst_src(scan_inst
, needs_dep
, first_write_grf
, write_len
);
3170 /* We insert our reads as late as possible since they're reading the
3171 * result of a SEND, which has massive latency.
3173 if (scan_inst
->dst
.file
== VGRF
&&
3174 scan_inst
->dst
.nr
>= first_write_grf
&&
3175 scan_inst
->dst
.nr
< first_write_grf
+ write_len
&&
3176 needs_dep
[scan_inst
->dst
.nr
- first_write_grf
]) {
3177 DEP_RESOLVE_MOV(fs_builder(this, block
, scan_inst
),
3179 needs_dep
[scan_inst
->dst
.nr
- first_write_grf
] = false;
3182 /* Continue the loop only if we haven't resolved all the dependencies */
3184 for (i
= 0; i
< write_len
; i
++) {
3194 fs_visitor::insert_gen4_send_dependency_workarounds()
3196 if (devinfo
->gen
!= 4 || devinfo
->is_g4x
)
3199 bool progress
= false;
3201 /* Note that we're done with register allocation, so GRF fs_regs always
3202 * have a .reg_offset of 0.
3205 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
3206 if (inst
->mlen
!= 0 && inst
->dst
.file
== VGRF
) {
3207 insert_gen4_pre_send_dependency_workarounds(block
, inst
);
3208 insert_gen4_post_send_dependency_workarounds(block
, inst
);
3214 invalidate_live_intervals();
3218 * Turns the generic expression-style uniform pull constant load instruction
3219 * into a hardware-specific series of instructions for loading a pull
3222 * The expression style allows the CSE pass before this to optimize out
3223 * repeated loads from the same offset, and gives the pre-register-allocation
3224 * scheduling full flexibility, while the conversion to native instructions
3225 * allows the post-register-allocation scheduler the best information
3228 * Note that execution masking for setting up pull constant loads is special:
3229 * the channels that need to be written are unrelated to the current execution
3230 * mask, since a later instruction will use one of the result channels as a
3231 * source operand for all 8 or 16 of its channels.
3234 fs_visitor::lower_uniform_pull_constant_loads()
3236 foreach_block_and_inst (block
, fs_inst
, inst
, cfg
) {
3237 if (inst
->opcode
!= FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
)
3240 if (devinfo
->gen
>= 7) {
3241 /* The offset arg is a vec4-aligned immediate byte offset. */
3242 fs_reg const_offset_reg
= inst
->src
[1];
3243 assert(const_offset_reg
.file
== IMM
&&
3244 const_offset_reg
.type
== BRW_REGISTER_TYPE_UD
);
3245 assert(const_offset_reg
.ud
% 16 == 0);
3247 fs_reg payload
, offset
;
3248 if (devinfo
->gen
>= 9) {
3249 /* We have to use a message header on Skylake to get SIMD4x2
3250 * mode. Reserve space for the register.
3252 offset
= payload
= fs_reg(VGRF
, alloc
.allocate(2));
3253 offset
.reg_offset
++;
3256 offset
= payload
= fs_reg(VGRF
, alloc
.allocate(1));
3260 /* This is actually going to be a MOV, but since only the first dword
3261 * is accessed, we have a special opcode to do just that one. Note
3262 * that this needs to be an operation that will be considered a def
3263 * by live variable analysis, or register allocation will explode.
3265 fs_inst
*setup
= new(mem_ctx
) fs_inst(FS_OPCODE_SET_SIMD4X2_OFFSET
,
3266 8, offset
, const_offset_reg
);
3267 setup
->force_writemask_all
= true;
3269 setup
->ir
= inst
->ir
;
3270 setup
->annotation
= inst
->annotation
;
3271 inst
->insert_before(block
, setup
);
3273 /* Similarly, this will only populate the first 4 channels of the
3274 * result register (since we only use smear values from 0-3), but we
3275 * don't tell the optimizer.
3277 inst
->opcode
= FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7
;
3278 inst
->src
[1] = payload
;
3279 inst
->base_mrf
= -1;
3281 invalidate_live_intervals();
3283 /* Before register allocation, we didn't tell the scheduler about the
3284 * MRF we use. We know it's safe to use this MRF because nothing
3285 * else does except for register spill/unspill, which generates and
3286 * uses its MRF within a single IR instruction.
3288 inst
->base_mrf
= FIRST_PULL_LOAD_MRF(devinfo
->gen
) + 1;
3295 fs_visitor::lower_load_payload()
3297 bool progress
= false;
3299 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
3300 if (inst
->opcode
!= SHADER_OPCODE_LOAD_PAYLOAD
)
3303 assert(inst
->dst
.file
== MRF
|| inst
->dst
.file
== VGRF
);
3304 assert(inst
->saturate
== false);
3305 fs_reg dst
= inst
->dst
;
3307 /* Get rid of COMPR4. We'll add it back in if we need it */
3308 if (dst
.file
== MRF
)
3309 dst
.nr
= dst
.nr
& ~BRW_MRF_COMPR4
;
3311 const fs_builder
ibld(this, block
, inst
);
3312 const fs_builder hbld
= ibld
.exec_all().group(8, 0);
3314 for (uint8_t i
= 0; i
< inst
->header_size
; i
++) {
3315 if (inst
->src
[i
].file
!= BAD_FILE
) {
3316 fs_reg mov_dst
= retype(dst
, BRW_REGISTER_TYPE_UD
);
3317 fs_reg mov_src
= retype(inst
->src
[i
], BRW_REGISTER_TYPE_UD
);
3318 hbld
.MOV(mov_dst
, mov_src
);
3320 dst
= offset(dst
, hbld
, 1);
3323 if (inst
->dst
.file
== MRF
&& (inst
->dst
.nr
& BRW_MRF_COMPR4
) &&
3324 inst
->exec_size
> 8) {
3325 /* In this case, the payload portion of the LOAD_PAYLOAD isn't
3326 * a straightforward copy. Instead, the result of the
3327 * LOAD_PAYLOAD is treated as interleaved and the first four
3328 * non-header sources are unpacked as:
3339 * This is used for gen <= 5 fb writes.
3341 assert(inst
->exec_size
== 16);
3342 assert(inst
->header_size
+ 4 <= inst
->sources
);
3343 for (uint8_t i
= inst
->header_size
; i
< inst
->header_size
+ 4; i
++) {
3344 if (inst
->src
[i
].file
!= BAD_FILE
) {
3345 if (devinfo
->has_compr4
) {
3346 fs_reg compr4_dst
= retype(dst
, inst
->src
[i
].type
);
3347 compr4_dst
.nr
|= BRW_MRF_COMPR4
;
3348 ibld
.MOV(compr4_dst
, inst
->src
[i
]);
3350 /* Platform doesn't have COMPR4. We have to fake it */
3351 fs_reg mov_dst
= retype(dst
, inst
->src
[i
].type
);
3352 ibld
.half(0).MOV(mov_dst
, half(inst
->src
[i
], 0));
3354 ibld
.half(1).MOV(mov_dst
, half(inst
->src
[i
], 1));
3361 /* The loop above only ever incremented us through the first set
3362 * of 4 registers. However, thanks to the magic of COMPR4, we
3363 * actually wrote to the first 8 registers, so we need to take
3364 * that into account now.
3368 /* The COMPR4 code took care of the first 4 sources. We'll let
3369 * the regular path handle any remaining sources. Yes, we are
3370 * modifying the instruction but we're about to delete it so
3371 * this really doesn't hurt anything.
3373 inst
->header_size
+= 4;
3376 for (uint8_t i
= inst
->header_size
; i
< inst
->sources
; i
++) {
3377 if (inst
->src
[i
].file
!= BAD_FILE
)
3378 ibld
.MOV(retype(dst
, inst
->src
[i
].type
), inst
->src
[i
]);
3379 dst
= offset(dst
, ibld
, 1);
3382 inst
->remove(block
);
3387 invalidate_live_intervals();
3393 fs_visitor::lower_integer_multiplication()
3395 bool progress
= false;
3397 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
3398 const fs_builder
ibld(this, block
, inst
);
3400 if (inst
->opcode
== BRW_OPCODE_MUL
) {
3401 if (inst
->dst
.is_accumulator() ||
3402 (inst
->dst
.type
!= BRW_REGISTER_TYPE_D
&&
3403 inst
->dst
.type
!= BRW_REGISTER_TYPE_UD
))
3406 /* Gen8's MUL instruction can do a 32-bit x 32-bit -> 32-bit
3407 * operation directly, but CHV/BXT cannot.
3409 if (devinfo
->gen
>= 8 &&
3410 !devinfo
->is_cherryview
&& !devinfo
->is_broxton
)
3413 if (inst
->src
[1].file
== IMM
&&
3414 inst
->src
[1].ud
< (1 << 16)) {
3415 /* The MUL instruction isn't commutative. On Gen <= 6, only the low
3416 * 16-bits of src0 are read, and on Gen >= 7 only the low 16-bits of
3419 * If multiplying by an immediate value that fits in 16-bits, do a
3420 * single MUL instruction with that value in the proper location.
3422 if (devinfo
->gen
< 7) {
3423 fs_reg
imm(VGRF
, alloc
.allocate(dispatch_width
/ 8),
3425 ibld
.MOV(imm
, inst
->src
[1]);
3426 ibld
.MUL(inst
->dst
, imm
, inst
->src
[0]);
3428 ibld
.MUL(inst
->dst
, inst
->src
[0], inst
->src
[1]);
3431 /* Gen < 8 (and some Gen8+ low-power parts like Cherryview) cannot
3432 * do 32-bit integer multiplication in one instruction, but instead
3433 * must do a sequence (which actually calculates a 64-bit result):
3435 * mul(8) acc0<1>D g3<8,8,1>D g4<8,8,1>D
3436 * mach(8) null g3<8,8,1>D g4<8,8,1>D
3437 * mov(8) g2<1>D acc0<8,8,1>D
3439 * But on Gen > 6, the ability to use second accumulator register
3440 * (acc1) for non-float data types was removed, preventing a simple
3441 * implementation in SIMD16. A 16-channel result can be calculated by
3442 * executing the three instructions twice in SIMD8, once with quarter
3443 * control of 1Q for the first eight channels and again with 2Q for
3444 * the second eight channels.
3446 * Which accumulator register is implicitly accessed (by AccWrEnable
3447 * for instance) is determined by the quarter control. Unfortunately
3448 * Ivybridge (and presumably Baytrail) has a hardware bug in which an
3449 * implicit accumulator access by an instruction with 2Q will access
3450 * acc1 regardless of whether the data type is usable in acc1.
3452 * Specifically, the 2Q mach(8) writes acc1 which does not exist for
3453 * integer data types.
3455 * Since we only want the low 32-bits of the result, we can do two
3456 * 32-bit x 16-bit multiplies (like the mul and mach are doing), and
3457 * adjust the high result and add them (like the mach is doing):
3459 * mul(8) g7<1>D g3<8,8,1>D g4.0<8,8,1>UW
3460 * mul(8) g8<1>D g3<8,8,1>D g4.1<8,8,1>UW
3461 * shl(8) g9<1>D g8<8,8,1>D 16D
3462 * add(8) g2<1>D g7<8,8,1>D g8<8,8,1>D
3464 * We avoid the shl instruction by realizing that we only want to add
3465 * the low 16-bits of the "high" result to the high 16-bits of the
3466 * "low" result and using proper regioning on the add:
3468 * mul(8) g7<1>D g3<8,8,1>D g4.0<16,8,2>UW
3469 * mul(8) g8<1>D g3<8,8,1>D g4.1<16,8,2>UW
3470 * add(8) g7.1<2>UW g7.1<16,8,2>UW g8<16,8,2>UW
3472 * Since it does not use the (single) accumulator register, we can
3473 * schedule multi-component multiplications much better.
3476 fs_reg orig_dst
= inst
->dst
;
3477 if (orig_dst
.is_null() || orig_dst
.file
== MRF
) {
3478 inst
->dst
= fs_reg(VGRF
, alloc
.allocate(dispatch_width
/ 8),
3481 fs_reg low
= inst
->dst
;
3482 fs_reg
high(VGRF
, alloc
.allocate(dispatch_width
/ 8),
3485 if (devinfo
->gen
>= 7) {
3486 fs_reg src1_0_w
= inst
->src
[1];
3487 fs_reg src1_1_w
= inst
->src
[1];
3489 if (inst
->src
[1].file
== IMM
) {
3490 src1_0_w
.ud
&= 0xffff;
3493 src1_0_w
.type
= BRW_REGISTER_TYPE_UW
;
3494 if (src1_0_w
.stride
!= 0) {
3495 assert(src1_0_w
.stride
== 1);
3496 src1_0_w
.stride
= 2;
3499 src1_1_w
.type
= BRW_REGISTER_TYPE_UW
;
3500 if (src1_1_w
.stride
!= 0) {
3501 assert(src1_1_w
.stride
== 1);
3502 src1_1_w
.stride
= 2;
3504 src1_1_w
.subreg_offset
+= type_sz(BRW_REGISTER_TYPE_UW
);
3506 ibld
.MUL(low
, inst
->src
[0], src1_0_w
);
3507 ibld
.MUL(high
, inst
->src
[0], src1_1_w
);
3509 fs_reg src0_0_w
= inst
->src
[0];
3510 fs_reg src0_1_w
= inst
->src
[0];
3512 src0_0_w
.type
= BRW_REGISTER_TYPE_UW
;
3513 if (src0_0_w
.stride
!= 0) {
3514 assert(src0_0_w
.stride
== 1);
3515 src0_0_w
.stride
= 2;
3518 src0_1_w
.type
= BRW_REGISTER_TYPE_UW
;
3519 if (src0_1_w
.stride
!= 0) {
3520 assert(src0_1_w
.stride
== 1);
3521 src0_1_w
.stride
= 2;
3523 src0_1_w
.subreg_offset
+= type_sz(BRW_REGISTER_TYPE_UW
);
3525 ibld
.MUL(low
, src0_0_w
, inst
->src
[1]);
3526 ibld
.MUL(high
, src0_1_w
, inst
->src
[1]);
3529 fs_reg dst
= inst
->dst
;
3530 dst
.type
= BRW_REGISTER_TYPE_UW
;
3531 dst
.subreg_offset
= 2;
3534 high
.type
= BRW_REGISTER_TYPE_UW
;
3537 low
.type
= BRW_REGISTER_TYPE_UW
;
3538 low
.subreg_offset
= 2;
3541 ibld
.ADD(dst
, low
, high
);
3543 if (inst
->conditional_mod
|| orig_dst
.file
== MRF
) {
3544 set_condmod(inst
->conditional_mod
,
3545 ibld
.MOV(orig_dst
, inst
->dst
));
3549 } else if (inst
->opcode
== SHADER_OPCODE_MULH
) {
3550 /* Should have been lowered to 8-wide. */
3551 assert(inst
->exec_size
<= 8);
3552 const fs_reg acc
= retype(brw_acc_reg(inst
->exec_size
),
3554 fs_inst
*mul
= ibld
.MUL(acc
, inst
->src
[0], inst
->src
[1]);
3555 fs_inst
*mach
= ibld
.MACH(inst
->dst
, inst
->src
[0], inst
->src
[1]);
3557 if (devinfo
->gen
>= 8) {
3558 /* Until Gen8, integer multiplies read 32-bits from one source,
3559 * and 16-bits from the other, and relying on the MACH instruction
3560 * to generate the high bits of the result.
3562 * On Gen8, the multiply instruction does a full 32x32-bit
3563 * multiply, but in order to do a 64-bit multiply we can simulate
3564 * the previous behavior and then use a MACH instruction.
3566 * FINISHME: Don't use source modifiers on src1.
3568 assert(mul
->src
[1].type
== BRW_REGISTER_TYPE_D
||
3569 mul
->src
[1].type
== BRW_REGISTER_TYPE_UD
);
3570 mul
->src
[1].type
= BRW_REGISTER_TYPE_UW
;
3571 mul
->src
[1].stride
*= 2;
3573 } else if (devinfo
->gen
== 7 && !devinfo
->is_haswell
&&
3574 inst
->force_sechalf
) {
3575 /* Among other things the quarter control bits influence which
3576 * accumulator register is used by the hardware for instructions
3577 * that access the accumulator implicitly (e.g. MACH). A
3578 * second-half instruction would normally map to acc1, which
3579 * doesn't exist on Gen7 and up (the hardware does emulate it for
3580 * floating-point instructions *only* by taking advantage of the
3581 * extra precision of acc0 not normally used for floating point
3584 * HSW and up are careful enough not to try to access an
3585 * accumulator register that doesn't exist, but on earlier Gen7
3586 * hardware we need to make sure that the quarter control bits are
3587 * zero to avoid non-deterministic behaviour and emit an extra MOV
3588 * to get the result masked correctly according to the current
3591 mach
->force_sechalf
= false;
3592 mach
->force_writemask_all
= true;
3593 mach
->dst
= ibld
.vgrf(inst
->dst
.type
);
3594 ibld
.MOV(inst
->dst
, mach
->dst
);
3600 inst
->remove(block
);
3605 invalidate_live_intervals();
3611 fs_visitor::lower_minmax()
3613 assert(devinfo
->gen
< 6);
3615 bool progress
= false;
3617 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
3618 const fs_builder
ibld(this, block
, inst
);
3620 if (inst
->opcode
== BRW_OPCODE_SEL
&&
3621 inst
->predicate
== BRW_PREDICATE_NONE
) {
3622 /* FIXME: Using CMP doesn't preserve the NaN propagation semantics of
3623 * the original SEL.L/GE instruction
3625 ibld
.CMP(ibld
.null_reg_d(), inst
->src
[0], inst
->src
[1],
3626 inst
->conditional_mod
);
3627 inst
->predicate
= BRW_PREDICATE_NORMAL
;
3628 inst
->conditional_mod
= BRW_CONDITIONAL_NONE
;
3635 invalidate_live_intervals();
3641 setup_color_payload(const fs_builder
&bld
, const brw_wm_prog_key
*key
,
3642 fs_reg
*dst
, fs_reg color
, unsigned components
)
3644 if (key
->clamp_fragment_color
) {
3645 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 4);
3646 assert(color
.type
== BRW_REGISTER_TYPE_F
);
3648 for (unsigned i
= 0; i
< components
; i
++)
3650 bld
.MOV(offset(tmp
, bld
, i
), offset(color
, bld
, i
)));
3655 for (unsigned i
= 0; i
< components
; i
++)
3656 dst
[i
] = offset(color
, bld
, i
);
3660 lower_fb_write_logical_send(const fs_builder
&bld
, fs_inst
*inst
,
3661 const brw_wm_prog_data
*prog_data
,
3662 const brw_wm_prog_key
*key
,
3663 const fs_visitor::thread_payload
&payload
)
3665 assert(inst
->src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].file
== IMM
);
3666 const brw_device_info
*devinfo
= bld
.shader
->devinfo
;
3667 const fs_reg
&color0
= inst
->src
[FB_WRITE_LOGICAL_SRC_COLOR0
];
3668 const fs_reg
&color1
= inst
->src
[FB_WRITE_LOGICAL_SRC_COLOR1
];
3669 const fs_reg
&src0_alpha
= inst
->src
[FB_WRITE_LOGICAL_SRC_SRC0_ALPHA
];
3670 const fs_reg
&src_depth
= inst
->src
[FB_WRITE_LOGICAL_SRC_SRC_DEPTH
];
3671 const fs_reg
&dst_depth
= inst
->src
[FB_WRITE_LOGICAL_SRC_DST_DEPTH
];
3672 const fs_reg
&src_stencil
= inst
->src
[FB_WRITE_LOGICAL_SRC_SRC_STENCIL
];
3673 fs_reg sample_mask
= inst
->src
[FB_WRITE_LOGICAL_SRC_OMASK
];
3674 const unsigned components
=
3675 inst
->src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].ud
;
3677 /* We can potentially have a message length of up to 15, so we have to set
3678 * base_mrf to either 0 or 1 in order to fit in m0..m15.
3681 int header_size
= 2, payload_header_size
;
3682 unsigned length
= 0;
3684 /* From the Sandy Bridge PRM, volume 4, page 198:
3686 * "Dispatched Pixel Enables. One bit per pixel indicating
3687 * which pixels were originally enabled when the thread was
3688 * dispatched. This field is only required for the end-of-
3689 * thread message and on all dual-source messages."
3691 if (devinfo
->gen
>= 6 &&
3692 (devinfo
->is_haswell
|| devinfo
->gen
>= 8 || !prog_data
->uses_kill
) &&
3693 color1
.file
== BAD_FILE
&&
3694 key
->nr_color_regions
== 1) {
3698 if (header_size
!= 0) {
3699 assert(header_size
== 2);
3700 /* Allocate 2 registers for a header */
3704 if (payload
.aa_dest_stencil_reg
) {
3705 sources
[length
] = fs_reg(VGRF
, bld
.shader
->alloc
.allocate(1));
3706 bld
.group(8, 0).exec_all().annotate("FB write stencil/AA alpha")
3707 .MOV(sources
[length
],
3708 fs_reg(brw_vec8_grf(payload
.aa_dest_stencil_reg
, 0)));
3712 if (prog_data
->uses_omask
) {
3713 sources
[length
] = fs_reg(VGRF
, bld
.shader
->alloc
.allocate(1),
3714 BRW_REGISTER_TYPE_UD
);
3716 /* Hand over gl_SampleMask. Only the lower 16 bits of each channel are
3717 * relevant. Since it's unsigned single words one vgrf is always
3718 * 16-wide, but only the lower or higher 8 channels will be used by the
3719 * hardware when doing a SIMD8 write depending on whether we have
3720 * selected the subspans for the first or second half respectively.
3722 assert(sample_mask
.file
!= BAD_FILE
&& type_sz(sample_mask
.type
) == 4);
3723 sample_mask
.type
= BRW_REGISTER_TYPE_UW
;
3724 sample_mask
.stride
*= 2;
3726 bld
.exec_all().annotate("FB write oMask")
3727 .MOV(half(retype(sources
[length
], BRW_REGISTER_TYPE_UW
),
3728 inst
->force_sechalf
),
3733 payload_header_size
= length
;
3735 if (src0_alpha
.file
!= BAD_FILE
) {
3736 /* FIXME: This is being passed at the wrong location in the payload and
3737 * doesn't work when gl_SampleMask and MRTs are used simultaneously.
3738 * It's supposed to be immediately before oMask but there seems to be no
3739 * reasonable way to pass them in the correct order because LOAD_PAYLOAD
3740 * requires header sources to form a contiguous segment at the beginning
3741 * of the message and src0_alpha has per-channel semantics.
3743 setup_color_payload(bld
, key
, &sources
[length
], src0_alpha
, 1);
3747 setup_color_payload(bld
, key
, &sources
[length
], color0
, components
);
3750 if (color1
.file
!= BAD_FILE
) {
3751 setup_color_payload(bld
, key
, &sources
[length
], color1
, components
);
3755 if (src_depth
.file
!= BAD_FILE
) {
3756 sources
[length
] = src_depth
;
3760 if (dst_depth
.file
!= BAD_FILE
) {
3761 sources
[length
] = dst_depth
;
3765 if (src_stencil
.file
!= BAD_FILE
) {
3766 assert(devinfo
->gen
>= 9);
3767 assert(bld
.dispatch_width() != 16);
3769 /* XXX: src_stencil is only available on gen9+. dst_depth is never
3770 * available on gen9+. As such it's impossible to have both enabled at the
3771 * same time and therefore length cannot overrun the array.
3773 assert(length
< 15);
3775 sources
[length
] = bld
.vgrf(BRW_REGISTER_TYPE_UD
);
3776 bld
.exec_all().annotate("FB write OS")
3777 .emit(FS_OPCODE_PACK_STENCIL_REF
, sources
[length
],
3778 retype(src_stencil
, BRW_REGISTER_TYPE_UB
));
3783 if (devinfo
->gen
>= 7) {
3784 /* Send from the GRF */
3785 fs_reg payload
= fs_reg(VGRF
, -1, BRW_REGISTER_TYPE_F
);
3786 load
= bld
.LOAD_PAYLOAD(payload
, sources
, length
, payload_header_size
);
3787 payload
.nr
= bld
.shader
->alloc
.allocate(load
->regs_written
);
3788 load
->dst
= payload
;
3790 inst
->src
[0] = payload
;
3791 inst
->resize_sources(1);
3792 inst
->base_mrf
= -1;
3794 /* Send from the MRF */
3795 load
= bld
.LOAD_PAYLOAD(fs_reg(MRF
, 1, BRW_REGISTER_TYPE_F
),
3796 sources
, length
, payload_header_size
);
3798 /* On pre-SNB, we have to interlace the color values. LOAD_PAYLOAD
3799 * will do this for us if we just give it a COMPR4 destination.
3801 if (devinfo
->gen
< 6 && bld
.dispatch_width() == 16)
3802 load
->dst
.nr
|= BRW_MRF_COMPR4
;
3804 inst
->resize_sources(0);
3808 inst
->opcode
= FS_OPCODE_FB_WRITE
;
3809 inst
->mlen
= load
->regs_written
;
3810 inst
->header_size
= header_size
;
3814 lower_sampler_logical_send_gen4(const fs_builder
&bld
, fs_inst
*inst
, opcode op
,
3815 const fs_reg
&coordinate
,
3816 const fs_reg
&shadow_c
,
3817 const fs_reg
&lod
, const fs_reg
&lod2
,
3818 const fs_reg
&surface
,
3819 const fs_reg
&sampler
,
3820 unsigned coord_components
,
3821 unsigned grad_components
)
3823 const bool has_lod
= (op
== SHADER_OPCODE_TXL
|| op
== FS_OPCODE_TXB
||
3824 op
== SHADER_OPCODE_TXF
|| op
== SHADER_OPCODE_TXS
);
3825 fs_reg
msg_begin(MRF
, 1, BRW_REGISTER_TYPE_F
);
3826 fs_reg msg_end
= msg_begin
;
3829 msg_end
= offset(msg_end
, bld
.group(8, 0), 1);
3831 for (unsigned i
= 0; i
< coord_components
; i
++)
3832 bld
.MOV(retype(offset(msg_end
, bld
, i
), coordinate
.type
),
3833 offset(coordinate
, bld
, i
));
3835 msg_end
= offset(msg_end
, bld
, coord_components
);
3837 /* Messages other than SAMPLE and RESINFO in SIMD16 and TXD in SIMD8
3838 * require all three components to be present and zero if they are unused.
3840 if (coord_components
> 0 &&
3841 (has_lod
|| shadow_c
.file
!= BAD_FILE
||
3842 (op
== SHADER_OPCODE_TEX
&& bld
.dispatch_width() == 8))) {
3843 for (unsigned i
= coord_components
; i
< 3; i
++)
3844 bld
.MOV(offset(msg_end
, bld
, i
), brw_imm_f(0.0f
));
3846 msg_end
= offset(msg_end
, bld
, 3 - coord_components
);
3849 if (op
== SHADER_OPCODE_TXD
) {
3850 /* TXD unsupported in SIMD16 mode. */
3851 assert(bld
.dispatch_width() == 8);
3853 /* the slots for u and v are always present, but r is optional */
3854 if (coord_components
< 2)
3855 msg_end
= offset(msg_end
, bld
, 2 - coord_components
);
3858 * dPdx = dudx, dvdx, drdx
3859 * dPdy = dudy, dvdy, drdy
3861 * 1-arg: Does not exist.
3863 * 2-arg: dudx dvdx dudy dvdy
3864 * dPdx.x dPdx.y dPdy.x dPdy.y
3867 * 3-arg: dudx dvdx drdx dudy dvdy drdy
3868 * dPdx.x dPdx.y dPdx.z dPdy.x dPdy.y dPdy.z
3869 * m5 m6 m7 m8 m9 m10
3871 for (unsigned i
= 0; i
< grad_components
; i
++)
3872 bld
.MOV(offset(msg_end
, bld
, i
), offset(lod
, bld
, i
));
3874 msg_end
= offset(msg_end
, bld
, MAX2(grad_components
, 2));
3876 for (unsigned i
= 0; i
< grad_components
; i
++)
3877 bld
.MOV(offset(msg_end
, bld
, i
), offset(lod2
, bld
, i
));
3879 msg_end
= offset(msg_end
, bld
, MAX2(grad_components
, 2));
3883 /* Bias/LOD with shadow comparitor is unsupported in SIMD16 -- *Without*
3884 * shadow comparitor (including RESINFO) it's unsupported in SIMD8 mode.
3886 assert(shadow_c
.file
!= BAD_FILE
? bld
.dispatch_width() == 8 :
3887 bld
.dispatch_width() == 16);
3889 const brw_reg_type type
=
3890 (op
== SHADER_OPCODE_TXF
|| op
== SHADER_OPCODE_TXS
?
3891 BRW_REGISTER_TYPE_UD
: BRW_REGISTER_TYPE_F
);
3892 bld
.MOV(retype(msg_end
, type
), lod
);
3893 msg_end
= offset(msg_end
, bld
, 1);
3896 if (shadow_c
.file
!= BAD_FILE
) {
3897 if (op
== SHADER_OPCODE_TEX
&& bld
.dispatch_width() == 8) {
3898 /* There's no plain shadow compare message, so we use shadow
3899 * compare with a bias of 0.0.
3901 bld
.MOV(msg_end
, brw_imm_f(0.0f
));
3902 msg_end
= offset(msg_end
, bld
, 1);
3905 bld
.MOV(msg_end
, shadow_c
);
3906 msg_end
= offset(msg_end
, bld
, 1);
3910 inst
->src
[0] = reg_undef
;
3911 inst
->src
[1] = surface
;
3912 inst
->src
[2] = sampler
;
3913 inst
->resize_sources(3);
3914 inst
->base_mrf
= msg_begin
.nr
;
3915 inst
->mlen
= msg_end
.nr
- msg_begin
.nr
;
3916 inst
->header_size
= 1;
3920 lower_sampler_logical_send_gen5(const fs_builder
&bld
, fs_inst
*inst
, opcode op
,
3922 const fs_reg
&shadow_c
,
3923 fs_reg lod
, fs_reg lod2
,
3924 const fs_reg
&sample_index
,
3925 const fs_reg
&surface
,
3926 const fs_reg
&sampler
,
3927 const fs_reg
&offset_value
,
3928 unsigned coord_components
,
3929 unsigned grad_components
)
3931 fs_reg
message(MRF
, 2, BRW_REGISTER_TYPE_F
);
3932 fs_reg msg_coords
= message
;
3933 unsigned header_size
= 0;
3935 if (offset_value
.file
!= BAD_FILE
) {
3936 /* The offsets set up by the visitor are in the m1 header, so we can't
3943 for (unsigned i
= 0; i
< coord_components
; i
++) {
3944 bld
.MOV(retype(offset(msg_coords
, bld
, i
), coordinate
.type
), coordinate
);
3945 coordinate
= offset(coordinate
, bld
, 1);
3947 fs_reg msg_end
= offset(msg_coords
, bld
, coord_components
);
3948 fs_reg msg_lod
= offset(msg_coords
, bld
, 4);
3950 if (shadow_c
.file
!= BAD_FILE
) {
3951 fs_reg msg_shadow
= msg_lod
;
3952 bld
.MOV(msg_shadow
, shadow_c
);
3953 msg_lod
= offset(msg_shadow
, bld
, 1);
3958 case SHADER_OPCODE_TXL
:
3960 bld
.MOV(msg_lod
, lod
);
3961 msg_end
= offset(msg_lod
, bld
, 1);
3963 case SHADER_OPCODE_TXD
:
3966 * dPdx = dudx, dvdx, drdx
3967 * dPdy = dudy, dvdy, drdy
3969 * Load up these values:
3970 * - dudx dudy dvdx dvdy drdx drdy
3971 * - dPdx.x dPdy.x dPdx.y dPdy.y dPdx.z dPdy.z
3974 for (unsigned i
= 0; i
< grad_components
; i
++) {
3975 bld
.MOV(msg_end
, lod
);
3976 lod
= offset(lod
, bld
, 1);
3977 msg_end
= offset(msg_end
, bld
, 1);
3979 bld
.MOV(msg_end
, lod2
);
3980 lod2
= offset(lod2
, bld
, 1);
3981 msg_end
= offset(msg_end
, bld
, 1);
3984 case SHADER_OPCODE_TXS
:
3985 msg_lod
= retype(msg_end
, BRW_REGISTER_TYPE_UD
);
3986 bld
.MOV(msg_lod
, lod
);
3987 msg_end
= offset(msg_lod
, bld
, 1);
3989 case SHADER_OPCODE_TXF
:
3990 msg_lod
= offset(msg_coords
, bld
, 3);
3991 bld
.MOV(retype(msg_lod
, BRW_REGISTER_TYPE_UD
), lod
);
3992 msg_end
= offset(msg_lod
, bld
, 1);
3994 case SHADER_OPCODE_TXF_CMS
:
3995 msg_lod
= offset(msg_coords
, bld
, 3);
3997 bld
.MOV(retype(msg_lod
, BRW_REGISTER_TYPE_UD
), brw_imm_ud(0u));
3999 bld
.MOV(retype(offset(msg_lod
, bld
, 1), BRW_REGISTER_TYPE_UD
), sample_index
);
4000 msg_end
= offset(msg_lod
, bld
, 2);
4007 inst
->src
[0] = reg_undef
;
4008 inst
->src
[1] = surface
;
4009 inst
->src
[2] = sampler
;
4010 inst
->resize_sources(3);
4011 inst
->base_mrf
= message
.nr
;
4012 inst
->mlen
= msg_end
.nr
- message
.nr
;
4013 inst
->header_size
= header_size
;
4015 /* Message length > MAX_SAMPLER_MESSAGE_SIZE disallowed by hardware. */
4016 assert(inst
->mlen
<= MAX_SAMPLER_MESSAGE_SIZE
);
4020 is_high_sampler(const struct brw_device_info
*devinfo
, const fs_reg
&sampler
)
4022 if (devinfo
->gen
< 8 && !devinfo
->is_haswell
)
4025 return sampler
.file
!= IMM
|| sampler
.ud
>= 16;
4029 lower_sampler_logical_send_gen7(const fs_builder
&bld
, fs_inst
*inst
, opcode op
,
4031 const fs_reg
&shadow_c
,
4032 fs_reg lod
, fs_reg lod2
,
4033 const fs_reg
&sample_index
,
4035 const fs_reg
&surface
,
4036 const fs_reg
&sampler
,
4037 fs_reg offset_value
,
4038 unsigned coord_components
,
4039 unsigned grad_components
)
4041 const brw_device_info
*devinfo
= bld
.shader
->devinfo
;
4042 int reg_width
= bld
.dispatch_width() / 8;
4043 unsigned header_size
= 0, length
= 0;
4044 fs_reg sources
[MAX_SAMPLER_MESSAGE_SIZE
];
4045 for (unsigned i
= 0; i
< ARRAY_SIZE(sources
); i
++)
4046 sources
[i
] = bld
.vgrf(BRW_REGISTER_TYPE_F
);
4048 if (op
== SHADER_OPCODE_TG4
|| op
== SHADER_OPCODE_TG4_OFFSET
||
4049 offset_value
.file
!= BAD_FILE
||
4050 is_high_sampler(devinfo
, sampler
)) {
4051 /* For general texture offsets (no txf workaround), we need a header to
4052 * put them in. Note that we're only reserving space for it in the
4053 * message payload as it will be initialized implicitly by the
4056 * TG4 needs to place its channel select in the header, for interaction
4057 * with ARB_texture_swizzle. The sampler index is only 4-bits, so for
4058 * larger sampler numbers we need to offset the Sampler State Pointer in
4062 sources
[0] = fs_reg();
4065 /* If we're requesting fewer than four channels worth of response,
4066 * and we have an explicit header, we need to set up the sampler
4067 * writemask. It's reversed from normal: 1 means "don't write".
4069 if (inst
->regs_written
!= 4 * reg_width
) {
4070 assert((inst
->regs_written
% reg_width
) == 0);
4071 unsigned mask
= ~((1 << (inst
->regs_written
/ reg_width
)) - 1) & 0xf;
4072 inst
->offset
|= mask
<< 12;
4076 if (shadow_c
.file
!= BAD_FILE
) {
4077 bld
.MOV(sources
[length
], shadow_c
);
4081 bool coordinate_done
= false;
4083 /* The sampler can only meaningfully compute LOD for fragment shader
4084 * messages. For all other stages, we change the opcode to TXL and
4085 * hardcode the LOD to 0.
4087 if (bld
.shader
->stage
!= MESA_SHADER_FRAGMENT
&&
4088 op
== SHADER_OPCODE_TEX
) {
4089 op
= SHADER_OPCODE_TXL
;
4090 lod
= brw_imm_f(0.0f
);
4093 /* Set up the LOD info */
4096 case SHADER_OPCODE_TXL
:
4097 bld
.MOV(sources
[length
], lod
);
4100 case SHADER_OPCODE_TXD
:
4101 /* TXD should have been lowered in SIMD16 mode. */
4102 assert(bld
.dispatch_width() == 8);
4104 /* Load dPdx and the coordinate together:
4105 * [hdr], [ref], x, dPdx.x, dPdy.x, y, dPdx.y, dPdy.y, z, dPdx.z, dPdy.z
4107 for (unsigned i
= 0; i
< coord_components
; i
++) {
4108 bld
.MOV(sources
[length
], coordinate
);
4109 coordinate
= offset(coordinate
, bld
, 1);
4112 /* For cube map array, the coordinate is (u,v,r,ai) but there are
4113 * only derivatives for (u, v, r).
4115 if (i
< grad_components
) {
4116 bld
.MOV(sources
[length
], lod
);
4117 lod
= offset(lod
, bld
, 1);
4120 bld
.MOV(sources
[length
], lod2
);
4121 lod2
= offset(lod2
, bld
, 1);
4126 coordinate_done
= true;
4128 case SHADER_OPCODE_TXS
:
4129 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
), lod
);
4132 case SHADER_OPCODE_TXF
:
4133 /* Unfortunately, the parameters for LD are intermixed: u, lod, v, r.
4134 * On Gen9 they are u, v, lod, r
4136 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
), coordinate
);
4137 coordinate
= offset(coordinate
, bld
, 1);
4140 if (devinfo
->gen
>= 9) {
4141 if (coord_components
>= 2) {
4142 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
), coordinate
);
4143 coordinate
= offset(coordinate
, bld
, 1);
4148 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
), lod
);
4151 for (unsigned i
= devinfo
->gen
>= 9 ? 2 : 1; i
< coord_components
; i
++) {
4152 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
), coordinate
);
4153 coordinate
= offset(coordinate
, bld
, 1);
4157 coordinate_done
= true;
4159 case SHADER_OPCODE_TXF_CMS
:
4160 case SHADER_OPCODE_TXF_CMS_W
:
4161 case SHADER_OPCODE_TXF_UMS
:
4162 case SHADER_OPCODE_TXF_MCS
:
4163 if (op
== SHADER_OPCODE_TXF_UMS
||
4164 op
== SHADER_OPCODE_TXF_CMS
||
4165 op
== SHADER_OPCODE_TXF_CMS_W
) {
4166 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
), sample_index
);
4170 if (op
== SHADER_OPCODE_TXF_CMS
|| op
== SHADER_OPCODE_TXF_CMS_W
) {
4171 /* Data from the multisample control surface. */
4172 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
), mcs
);
4175 /* On Gen9+ we'll use ld2dms_w instead which has two registers for
4178 if (op
== SHADER_OPCODE_TXF_CMS_W
) {
4179 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
),
4182 offset(mcs
, bld
, 1));
4187 /* There is no offsetting for this message; just copy in the integer
4188 * texture coordinates.
4190 for (unsigned i
= 0; i
< coord_components
; i
++) {
4191 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
), coordinate
);
4192 coordinate
= offset(coordinate
, bld
, 1);
4196 coordinate_done
= true;
4198 case SHADER_OPCODE_TG4_OFFSET
:
4199 /* gather4_po_c should have been lowered in SIMD16 mode. */
4200 assert(bld
.dispatch_width() == 8 || shadow_c
.file
== BAD_FILE
);
4202 /* More crazy intermixing */
4203 for (unsigned i
= 0; i
< 2; i
++) { /* u, v */
4204 bld
.MOV(sources
[length
], coordinate
);
4205 coordinate
= offset(coordinate
, bld
, 1);
4209 for (unsigned i
= 0; i
< 2; i
++) { /* offu, offv */
4210 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
), offset_value
);
4211 offset_value
= offset(offset_value
, bld
, 1);
4215 if (coord_components
== 3) { /* r if present */
4216 bld
.MOV(sources
[length
], coordinate
);
4217 coordinate
= offset(coordinate
, bld
, 1);
4221 coordinate_done
= true;
4227 /* Set up the coordinate (except for cases where it was done above) */
4228 if (!coordinate_done
) {
4229 for (unsigned i
= 0; i
< coord_components
; i
++) {
4230 bld
.MOV(sources
[length
], coordinate
);
4231 coordinate
= offset(coordinate
, bld
, 1);
4238 mlen
= length
* reg_width
- header_size
;
4240 mlen
= length
* reg_width
;
4242 const fs_reg src_payload
= fs_reg(VGRF
, bld
.shader
->alloc
.allocate(mlen
),
4243 BRW_REGISTER_TYPE_F
);
4244 bld
.LOAD_PAYLOAD(src_payload
, sources
, length
, header_size
);
4246 /* Generate the SEND. */
4248 inst
->src
[0] = src_payload
;
4249 inst
->src
[1] = surface
;
4250 inst
->src
[2] = sampler
;
4251 inst
->resize_sources(3);
4252 inst
->base_mrf
= -1;
4254 inst
->header_size
= header_size
;
4256 /* Message length > MAX_SAMPLER_MESSAGE_SIZE disallowed by hardware. */
4257 assert(inst
->mlen
<= MAX_SAMPLER_MESSAGE_SIZE
);
4261 lower_sampler_logical_send(const fs_builder
&bld
, fs_inst
*inst
, opcode op
)
4263 const brw_device_info
*devinfo
= bld
.shader
->devinfo
;
4264 const fs_reg
&coordinate
= inst
->src
[TEX_LOGICAL_SRC_COORDINATE
];
4265 const fs_reg
&shadow_c
= inst
->src
[TEX_LOGICAL_SRC_SHADOW_C
];
4266 const fs_reg
&lod
= inst
->src
[TEX_LOGICAL_SRC_LOD
];
4267 const fs_reg
&lod2
= inst
->src
[TEX_LOGICAL_SRC_LOD2
];
4268 const fs_reg
&sample_index
= inst
->src
[TEX_LOGICAL_SRC_SAMPLE_INDEX
];
4269 const fs_reg
&mcs
= inst
->src
[TEX_LOGICAL_SRC_MCS
];
4270 const fs_reg
&surface
= inst
->src
[TEX_LOGICAL_SRC_SURFACE
];
4271 const fs_reg
&sampler
= inst
->src
[TEX_LOGICAL_SRC_SAMPLER
];
4272 const fs_reg
&offset_value
= inst
->src
[TEX_LOGICAL_SRC_OFFSET_VALUE
];
4273 assert(inst
->src
[TEX_LOGICAL_SRC_COORD_COMPONENTS
].file
== IMM
);
4274 const unsigned coord_components
= inst
->src
[TEX_LOGICAL_SRC_COORD_COMPONENTS
].ud
;
4275 assert(inst
->src
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
].file
== IMM
);
4276 const unsigned grad_components
= inst
->src
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
].ud
;
4278 if (devinfo
->gen
>= 7) {
4279 lower_sampler_logical_send_gen7(bld
, inst
, op
, coordinate
,
4280 shadow_c
, lod
, lod2
, sample_index
,
4281 mcs
, surface
, sampler
, offset_value
,
4282 coord_components
, grad_components
);
4283 } else if (devinfo
->gen
>= 5) {
4284 lower_sampler_logical_send_gen5(bld
, inst
, op
, coordinate
,
4285 shadow_c
, lod
, lod2
, sample_index
,
4286 surface
, sampler
, offset_value
,
4287 coord_components
, grad_components
);
4289 lower_sampler_logical_send_gen4(bld
, inst
, op
, coordinate
,
4290 shadow_c
, lod
, lod2
,
4292 coord_components
, grad_components
);
4297 * Initialize the header present in some typed and untyped surface
4301 emit_surface_header(const fs_builder
&bld
, const fs_reg
&sample_mask
)
4303 fs_builder ubld
= bld
.exec_all().group(8, 0);
4304 const fs_reg dst
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
4305 ubld
.MOV(dst
, brw_imm_d(0));
4306 ubld
.MOV(component(dst
, 7), sample_mask
);
4311 lower_surface_logical_send(const fs_builder
&bld
, fs_inst
*inst
, opcode op
,
4312 const fs_reg
&sample_mask
)
4314 /* Get the logical send arguments. */
4315 const fs_reg
&addr
= inst
->src
[0];
4316 const fs_reg
&src
= inst
->src
[1];
4317 const fs_reg
&surface
= inst
->src
[2];
4318 const UNUSED fs_reg
&dims
= inst
->src
[3];
4319 const fs_reg
&arg
= inst
->src
[4];
4321 /* Calculate the total number of components of the payload. */
4322 const unsigned addr_sz
= inst
->components_read(0);
4323 const unsigned src_sz
= inst
->components_read(1);
4324 const unsigned header_sz
= (sample_mask
.file
== BAD_FILE
? 0 : 1);
4325 const unsigned sz
= header_sz
+ addr_sz
+ src_sz
;
4327 /* Allocate space for the payload. */
4328 fs_reg
*const components
= new fs_reg
[sz
];
4329 const fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, sz
);
4332 /* Construct the payload. */
4334 components
[n
++] = emit_surface_header(bld
, sample_mask
);
4336 for (unsigned i
= 0; i
< addr_sz
; i
++)
4337 components
[n
++] = offset(addr
, bld
, i
);
4339 for (unsigned i
= 0; i
< src_sz
; i
++)
4340 components
[n
++] = offset(src
, bld
, i
);
4342 bld
.LOAD_PAYLOAD(payload
, components
, sz
, header_sz
);
4344 /* Update the original instruction. */
4346 inst
->mlen
= header_sz
+ (addr_sz
+ src_sz
) * inst
->exec_size
/ 8;
4347 inst
->header_size
= header_sz
;
4349 inst
->src
[0] = payload
;
4350 inst
->src
[1] = surface
;
4352 inst
->resize_sources(3);
4354 delete[] components
;
4358 fs_visitor::lower_logical_sends()
4360 bool progress
= false;
4362 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
4363 const fs_builder
ibld(this, block
, inst
);
4365 switch (inst
->opcode
) {
4366 case FS_OPCODE_FB_WRITE_LOGICAL
:
4367 assert(stage
== MESA_SHADER_FRAGMENT
);
4368 lower_fb_write_logical_send(ibld
, inst
,
4369 (const brw_wm_prog_data
*)prog_data
,
4370 (const brw_wm_prog_key
*)key
,
4374 case SHADER_OPCODE_TEX_LOGICAL
:
4375 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TEX
);
4378 case SHADER_OPCODE_TXD_LOGICAL
:
4379 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXD
);
4382 case SHADER_OPCODE_TXF_LOGICAL
:
4383 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF
);
4386 case SHADER_OPCODE_TXL_LOGICAL
:
4387 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXL
);
4390 case SHADER_OPCODE_TXS_LOGICAL
:
4391 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXS
);
4394 case FS_OPCODE_TXB_LOGICAL
:
4395 lower_sampler_logical_send(ibld
, inst
, FS_OPCODE_TXB
);
4398 case SHADER_OPCODE_TXF_CMS_LOGICAL
:
4399 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_CMS
);
4402 case SHADER_OPCODE_TXF_CMS_W_LOGICAL
:
4403 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_CMS_W
);
4406 case SHADER_OPCODE_TXF_UMS_LOGICAL
:
4407 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_UMS
);
4410 case SHADER_OPCODE_TXF_MCS_LOGICAL
:
4411 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_MCS
);
4414 case SHADER_OPCODE_LOD_LOGICAL
:
4415 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_LOD
);
4418 case SHADER_OPCODE_TG4_LOGICAL
:
4419 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TG4
);
4422 case SHADER_OPCODE_TG4_OFFSET_LOGICAL
:
4423 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TG4_OFFSET
);
4426 case SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
:
4427 lower_surface_logical_send(ibld
, inst
,
4428 SHADER_OPCODE_UNTYPED_SURFACE_READ
,
4432 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
:
4433 lower_surface_logical_send(ibld
, inst
,
4434 SHADER_OPCODE_UNTYPED_SURFACE_WRITE
,
4435 ibld
.sample_mask_reg());
4438 case SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL
:
4439 lower_surface_logical_send(ibld
, inst
,
4440 SHADER_OPCODE_UNTYPED_ATOMIC
,
4441 ibld
.sample_mask_reg());
4444 case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
:
4445 lower_surface_logical_send(ibld
, inst
,
4446 SHADER_OPCODE_TYPED_SURFACE_READ
,
4450 case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
:
4451 lower_surface_logical_send(ibld
, inst
,
4452 SHADER_OPCODE_TYPED_SURFACE_WRITE
,
4453 ibld
.sample_mask_reg());
4456 case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
:
4457 lower_surface_logical_send(ibld
, inst
,
4458 SHADER_OPCODE_TYPED_ATOMIC
,
4459 ibld
.sample_mask_reg());
4470 invalidate_live_intervals();
4476 * Get the closest native SIMD width supported by the hardware for instruction
4477 * \p inst. The instruction will be left untouched by
4478 * fs_visitor::lower_simd_width() if the returned value is equal to the
4479 * original execution size.
4482 get_lowered_simd_width(const struct brw_device_info
*devinfo
,
4483 const fs_inst
*inst
)
4485 switch (inst
->opcode
) {
4486 case BRW_OPCODE_MOV
:
4487 case BRW_OPCODE_SEL
:
4488 case BRW_OPCODE_NOT
:
4489 case BRW_OPCODE_AND
:
4491 case BRW_OPCODE_XOR
:
4492 case BRW_OPCODE_SHR
:
4493 case BRW_OPCODE_SHL
:
4494 case BRW_OPCODE_ASR
:
4495 case BRW_OPCODE_CMP
:
4496 case BRW_OPCODE_CMPN
:
4497 case BRW_OPCODE_CSEL
:
4498 case BRW_OPCODE_F32TO16
:
4499 case BRW_OPCODE_F16TO32
:
4500 case BRW_OPCODE_BFREV
:
4501 case BRW_OPCODE_BFE
:
4502 case BRW_OPCODE_BFI1
:
4503 case BRW_OPCODE_BFI2
:
4504 case BRW_OPCODE_ADD
:
4505 case BRW_OPCODE_MUL
:
4506 case BRW_OPCODE_AVG
:
4507 case BRW_OPCODE_FRC
:
4508 case BRW_OPCODE_RNDU
:
4509 case BRW_OPCODE_RNDD
:
4510 case BRW_OPCODE_RNDE
:
4511 case BRW_OPCODE_RNDZ
:
4512 case BRW_OPCODE_LZD
:
4513 case BRW_OPCODE_FBH
:
4514 case BRW_OPCODE_FBL
:
4515 case BRW_OPCODE_CBIT
:
4516 case BRW_OPCODE_SAD2
:
4517 case BRW_OPCODE_MAD
:
4518 case BRW_OPCODE_LRP
:
4519 case SHADER_OPCODE_RCP
:
4520 case SHADER_OPCODE_RSQ
:
4521 case SHADER_OPCODE_SQRT
:
4522 case SHADER_OPCODE_EXP2
:
4523 case SHADER_OPCODE_LOG2
:
4524 case SHADER_OPCODE_POW
:
4525 case SHADER_OPCODE_INT_QUOTIENT
:
4526 case SHADER_OPCODE_INT_REMAINDER
:
4527 case SHADER_OPCODE_SIN
:
4528 case SHADER_OPCODE_COS
: {
4529 /* According to the PRMs:
4530 * "A. In Direct Addressing mode, a source cannot span more than 2
4531 * adjacent GRF registers.
4532 * B. A destination cannot span more than 2 adjacent GRF registers."
4534 * Look for the source or destination with the largest register region
4535 * which is the one that is going to limit the overal execution size of
4536 * the instruction due to this rule.
4538 unsigned reg_count
= inst
->regs_written
;
4540 for (unsigned i
= 0; i
< inst
->sources
; i
++)
4541 reg_count
= MAX2(reg_count
, (unsigned)inst
->regs_read(i
));
4543 /* Calculate the maximum execution size of the instruction based on the
4544 * factor by which it goes over the hardware limit of 2 GRFs.
4546 return inst
->exec_size
/ DIV_ROUND_UP(reg_count
, 2);
4548 case SHADER_OPCODE_MULH
:
4549 /* MULH is lowered to the MUL/MACH sequence using the accumulator, which
4550 * is 8-wide on Gen7+.
4552 return (devinfo
->gen
>= 7 ? 8 : inst
->exec_size
);
4554 case FS_OPCODE_FB_WRITE_LOGICAL
:
4555 /* Gen6 doesn't support SIMD16 depth writes but we cannot handle them
4558 assert(devinfo
->gen
!= 6 ||
4559 inst
->src
[FB_WRITE_LOGICAL_SRC_SRC_DEPTH
].file
== BAD_FILE
||
4560 inst
->exec_size
== 8);
4561 /* Dual-source FB writes are unsupported in SIMD16 mode. */
4562 return (inst
->src
[FB_WRITE_LOGICAL_SRC_COLOR1
].file
!= BAD_FILE
?
4563 8 : inst
->exec_size
);
4565 case SHADER_OPCODE_TXD_LOGICAL
:
4566 /* TXD is unsupported in SIMD16 mode. */
4569 case SHADER_OPCODE_TG4_OFFSET_LOGICAL
: {
4570 /* gather4_po_c is unsupported in SIMD16 mode. */
4571 const fs_reg
&shadow_c
= inst
->src
[TEX_LOGICAL_SRC_SHADOW_C
];
4572 return (shadow_c
.file
!= BAD_FILE
? 8 : inst
->exec_size
);
4574 case SHADER_OPCODE_TXL_LOGICAL
:
4575 case FS_OPCODE_TXB_LOGICAL
: {
4576 /* Gen4 doesn't have SIMD8 non-shadow-compare bias/LOD instructions, and
4577 * Gen4-6 can't support TXL and TXB with shadow comparison in SIMD16
4578 * mode because the message exceeds the maximum length of 11.
4580 const fs_reg
&shadow_c
= inst
->src
[TEX_LOGICAL_SRC_SHADOW_C
];
4581 if (devinfo
->gen
== 4 && shadow_c
.file
== BAD_FILE
)
4583 else if (devinfo
->gen
< 7 && shadow_c
.file
!= BAD_FILE
)
4586 return inst
->exec_size
;
4588 case SHADER_OPCODE_TXF_LOGICAL
:
4589 case SHADER_OPCODE_TXS_LOGICAL
:
4590 /* Gen4 doesn't have SIMD8 variants for the RESINFO and LD-with-LOD
4591 * messages. Use SIMD16 instead.
4593 if (devinfo
->gen
== 4)
4596 return inst
->exec_size
;
4598 case SHADER_OPCODE_TXF_CMS_W_LOGICAL
: {
4599 /* This opcode can take up to 6 arguments which means that in some
4600 * circumstances it can end up with a message that is too long in SIMD16
4603 const unsigned coord_components
=
4604 inst
->src
[TEX_LOGICAL_SRC_COORD_COMPONENTS
].ud
;
4605 /* First three arguments are the sample index and the two arguments for
4608 if ((coord_components
+ 3) * 2 > MAX_SAMPLER_MESSAGE_SIZE
)
4611 return inst
->exec_size
;
4614 case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
:
4615 case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
:
4616 case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
:
4619 case SHADER_OPCODE_MOV_INDIRECT
:
4620 /* Prior to Broadwell, we only have 8 address subregisters */
4621 return devinfo
->gen
< 8 ? 8 : MIN2(inst
->exec_size
, 16);
4624 return inst
->exec_size
;
4629 * The \p rows array of registers represents a \p num_rows by \p num_columns
4630 * matrix in row-major order, write it in column-major order into the register
4631 * passed as destination. \p stride gives the separation between matrix
4632 * elements in the input in fs_builder::dispatch_width() units.
4635 emit_transpose(const fs_builder
&bld
,
4636 const fs_reg
&dst
, const fs_reg
*rows
,
4637 unsigned num_rows
, unsigned num_columns
, unsigned stride
)
4639 fs_reg
*const components
= new fs_reg
[num_rows
* num_columns
];
4641 for (unsigned i
= 0; i
< num_columns
; ++i
) {
4642 for (unsigned j
= 0; j
< num_rows
; ++j
)
4643 components
[num_rows
* i
+ j
] = offset(rows
[j
], bld
, stride
* i
);
4646 bld
.LOAD_PAYLOAD(dst
, components
, num_rows
* num_columns
, 0);
4648 delete[] components
;
4652 fs_visitor::lower_simd_width()
4654 bool progress
= false;
4656 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
4657 const unsigned lower_width
= get_lowered_simd_width(devinfo
, inst
);
4659 if (lower_width
!= inst
->exec_size
) {
4660 /* Builder matching the original instruction. We may also need to
4661 * emit an instruction of width larger than the original, set the
4662 * execution size of the builder to the highest of both for now so
4663 * we're sure that both cases can be handled.
4665 const fs_builder ibld
= bld
.at(block
, inst
)
4666 .exec_all(inst
->force_writemask_all
)
4667 .group(MAX2(inst
->exec_size
, lower_width
),
4668 inst
->force_sechalf
);
4670 /* Split the copies in chunks of the execution width of either the
4671 * original or the lowered instruction, whichever is lower.
4673 const unsigned copy_width
= MIN2(lower_width
, inst
->exec_size
);
4674 const unsigned n
= inst
->exec_size
/ copy_width
;
4675 const unsigned dst_size
= inst
->regs_written
* REG_SIZE
/
4676 inst
->dst
.component_size(inst
->exec_size
);
4679 assert(n
> 0 && n
<= ARRAY_SIZE(dsts
) &&
4680 !inst
->writes_accumulator
&& !inst
->mlen
);
4682 for (unsigned i
= 0; i
< n
; i
++) {
4683 /* Emit a copy of the original instruction with the lowered width.
4684 * If the EOT flag was set throw it away except for the last
4685 * instruction to avoid killing the thread prematurely.
4687 fs_inst split_inst
= *inst
;
4688 split_inst
.exec_size
= lower_width
;
4689 split_inst
.eot
= inst
->eot
&& i
== n
- 1;
4691 /* Select the correct channel enables for the i-th group, then
4692 * transform the sources and destination and emit the lowered
4695 const fs_builder lbld
= ibld
.group(lower_width
, i
);
4697 for (unsigned j
= 0; j
< inst
->sources
; j
++) {
4698 if (inst
->src
[j
].file
!= BAD_FILE
&&
4699 !is_uniform(inst
->src
[j
])) {
4700 /* Get the i-th copy_width-wide chunk of the source. */
4701 const fs_reg src
= horiz_offset(inst
->src
[j
], copy_width
* i
);
4702 const unsigned src_size
= inst
->components_read(j
);
4704 /* Use a trivial transposition to copy one every n
4705 * copy_width-wide components of the register into a
4706 * temporary passed as source to the lowered instruction.
4708 split_inst
.src
[j
] = lbld
.vgrf(inst
->src
[j
].type
, src_size
);
4709 emit_transpose(lbld
.group(copy_width
, 0),
4710 split_inst
.src
[j
], &src
, 1, src_size
, n
);
4714 if (inst
->regs_written
) {
4715 /* Allocate enough space to hold the result of the lowered
4716 * instruction and fix up the number of registers written.
4718 split_inst
.dst
= dsts
[i
] =
4719 lbld
.vgrf(inst
->dst
.type
, dst_size
);
4720 split_inst
.regs_written
=
4721 DIV_ROUND_UP(inst
->regs_written
* lower_width
,
4725 lbld
.emit(split_inst
);
4728 if (inst
->regs_written
) {
4729 /* Distance between useful channels in the temporaries, skipping
4730 * garbage if the lowered instruction is wider than the original.
4732 const unsigned m
= lower_width
/ copy_width
;
4734 /* Interleave the components of the result from the lowered
4735 * instructions. We need to set exec_all() when copying more than
4736 * one half per component, because LOAD_PAYLOAD (in terms of which
4737 * emit_transpose is implemented) can only use the same channel
4738 * enable signals for all of its non-header sources.
4740 emit_transpose(ibld
.exec_all(inst
->exec_size
> copy_width
)
4741 .group(copy_width
, 0),
4742 inst
->dst
, dsts
, n
, dst_size
, m
);
4745 inst
->remove(block
);
4751 invalidate_live_intervals();
4757 fs_visitor::dump_instructions()
4759 dump_instructions(NULL
);
4763 fs_visitor::dump_instructions(const char *name
)
4765 FILE *file
= stderr
;
4766 if (name
&& geteuid() != 0) {
4767 file
= fopen(name
, "w");
4773 calculate_register_pressure();
4774 int ip
= 0, max_pressure
= 0;
4775 foreach_block_and_inst(block
, backend_instruction
, inst
, cfg
) {
4776 max_pressure
= MAX2(max_pressure
, regs_live_at_ip
[ip
]);
4777 fprintf(file
, "{%3d} %4d: ", regs_live_at_ip
[ip
], ip
);
4778 dump_instruction(inst
, file
);
4781 fprintf(file
, "Maximum %3d registers live at once.\n", max_pressure
);
4784 foreach_in_list(backend_instruction
, inst
, &instructions
) {
4785 fprintf(file
, "%4d: ", ip
++);
4786 dump_instruction(inst
, file
);
4790 if (file
!= stderr
) {
4796 fs_visitor::dump_instruction(backend_instruction
*be_inst
)
4798 dump_instruction(be_inst
, stderr
);
4802 fs_visitor::dump_instruction(backend_instruction
*be_inst
, FILE *file
)
4804 fs_inst
*inst
= (fs_inst
*)be_inst
;
4806 if (inst
->predicate
) {
4807 fprintf(file
, "(%cf0.%d) ",
4808 inst
->predicate_inverse
? '-' : '+',
4812 fprintf(file
, "%s", brw_instruction_name(inst
->opcode
));
4814 fprintf(file
, ".sat");
4815 if (inst
->conditional_mod
) {
4816 fprintf(file
, "%s", conditional_modifier
[inst
->conditional_mod
]);
4817 if (!inst
->predicate
&&
4818 (devinfo
->gen
< 5 || (inst
->opcode
!= BRW_OPCODE_SEL
&&
4819 inst
->opcode
!= BRW_OPCODE_IF
&&
4820 inst
->opcode
!= BRW_OPCODE_WHILE
))) {
4821 fprintf(file
, ".f0.%d", inst
->flag_subreg
);
4824 fprintf(file
, "(%d) ", inst
->exec_size
);
4827 fprintf(file
, "(mlen: %d) ", inst
->mlen
);
4830 switch (inst
->dst
.file
) {
4832 fprintf(file
, "vgrf%d", inst
->dst
.nr
);
4833 if (alloc
.sizes
[inst
->dst
.nr
] != inst
->regs_written
||
4834 inst
->dst
.subreg_offset
)
4835 fprintf(file
, "+%d.%d",
4836 inst
->dst
.reg_offset
, inst
->dst
.subreg_offset
);
4839 fprintf(file
, "g%d", inst
->dst
.nr
);
4842 fprintf(file
, "m%d", inst
->dst
.nr
);
4845 fprintf(file
, "(null)");
4848 fprintf(file
, "***u%d***", inst
->dst
.nr
+ inst
->dst
.reg_offset
);
4851 fprintf(file
, "***attr%d***", inst
->dst
.nr
+ inst
->dst
.reg_offset
);
4854 switch (inst
->dst
.nr
) {
4856 fprintf(file
, "null");
4858 case BRW_ARF_ADDRESS
:
4859 fprintf(file
, "a0.%d", inst
->dst
.subnr
);
4861 case BRW_ARF_ACCUMULATOR
:
4862 fprintf(file
, "acc%d", inst
->dst
.subnr
);
4865 fprintf(file
, "f%d.%d", inst
->dst
.nr
& 0xf, inst
->dst
.subnr
);
4868 fprintf(file
, "arf%d.%d", inst
->dst
.nr
& 0xf, inst
->dst
.subnr
);
4871 if (inst
->dst
.subnr
)
4872 fprintf(file
, "+%d", inst
->dst
.subnr
);
4875 unreachable("not reached");
4877 if (inst
->dst
.stride
!= 1)
4878 fprintf(file
, "<%u>", inst
->dst
.stride
);
4879 fprintf(file
, ":%s, ", brw_reg_type_letters(inst
->dst
.type
));
4881 for (int i
= 0; i
< inst
->sources
; i
++) {
4882 if (inst
->src
[i
].negate
)
4884 if (inst
->src
[i
].abs
)
4886 switch (inst
->src
[i
].file
) {
4888 fprintf(file
, "vgrf%d", inst
->src
[i
].nr
);
4889 if (alloc
.sizes
[inst
->src
[i
].nr
] != (unsigned)inst
->regs_read(i
) ||
4890 inst
->src
[i
].subreg_offset
)
4891 fprintf(file
, "+%d.%d", inst
->src
[i
].reg_offset
,
4892 inst
->src
[i
].subreg_offset
);
4895 fprintf(file
, "g%d", inst
->src
[i
].nr
);
4898 fprintf(file
, "***m%d***", inst
->src
[i
].nr
);
4901 fprintf(file
, "attr%d+%d", inst
->src
[i
].nr
, inst
->src
[i
].reg_offset
);
4904 fprintf(file
, "u%d", inst
->src
[i
].nr
+ inst
->src
[i
].reg_offset
);
4905 if (inst
->src
[i
].subreg_offset
) {
4906 fprintf(file
, "+%d.%d", inst
->src
[i
].reg_offset
,
4907 inst
->src
[i
].subreg_offset
);
4911 fprintf(file
, "(null)");
4914 switch (inst
->src
[i
].type
) {
4915 case BRW_REGISTER_TYPE_F
:
4916 fprintf(file
, "%-gf", inst
->src
[i
].f
);
4918 case BRW_REGISTER_TYPE_W
:
4919 case BRW_REGISTER_TYPE_D
:
4920 fprintf(file
, "%dd", inst
->src
[i
].d
);
4922 case BRW_REGISTER_TYPE_UW
:
4923 case BRW_REGISTER_TYPE_UD
:
4924 fprintf(file
, "%uu", inst
->src
[i
].ud
);
4926 case BRW_REGISTER_TYPE_VF
:
4927 fprintf(file
, "[%-gF, %-gF, %-gF, %-gF]",
4928 brw_vf_to_float((inst
->src
[i
].ud
>> 0) & 0xff),
4929 brw_vf_to_float((inst
->src
[i
].ud
>> 8) & 0xff),
4930 brw_vf_to_float((inst
->src
[i
].ud
>> 16) & 0xff),
4931 brw_vf_to_float((inst
->src
[i
].ud
>> 24) & 0xff));
4934 fprintf(file
, "???");
4939 switch (inst
->src
[i
].nr
) {
4941 fprintf(file
, "null");
4943 case BRW_ARF_ADDRESS
:
4944 fprintf(file
, "a0.%d", inst
->src
[i
].subnr
);
4946 case BRW_ARF_ACCUMULATOR
:
4947 fprintf(file
, "acc%d", inst
->src
[i
].subnr
);
4950 fprintf(file
, "f%d.%d", inst
->src
[i
].nr
& 0xf, inst
->src
[i
].subnr
);
4953 fprintf(file
, "arf%d.%d", inst
->src
[i
].nr
& 0xf, inst
->src
[i
].subnr
);
4956 if (inst
->src
[i
].subnr
)
4957 fprintf(file
, "+%d", inst
->src
[i
].subnr
);
4960 if (inst
->src
[i
].abs
)
4963 if (inst
->src
[i
].file
!= IMM
) {
4965 if (inst
->src
[i
].file
== ARF
|| inst
->src
[i
].file
== FIXED_GRF
) {
4966 unsigned hstride
= inst
->src
[i
].hstride
;
4967 stride
= (hstride
== 0 ? 0 : (1 << (hstride
- 1)));
4969 stride
= inst
->src
[i
].stride
;
4972 fprintf(file
, "<%u>", stride
);
4974 fprintf(file
, ":%s", brw_reg_type_letters(inst
->src
[i
].type
));
4977 if (i
< inst
->sources
- 1 && inst
->src
[i
+ 1].file
!= BAD_FILE
)
4978 fprintf(file
, ", ");
4983 if (inst
->force_writemask_all
)
4984 fprintf(file
, "NoMask ");
4986 if (dispatch_width
== 16 && inst
->exec_size
== 8) {
4987 if (inst
->force_sechalf
)
4988 fprintf(file
, "2ndhalf ");
4990 fprintf(file
, "1sthalf ");
4993 fprintf(file
, "\n");
4997 * Possibly returns an instruction that set up @param reg.
4999 * Sometimes we want to take the result of some expression/variable
5000 * dereference tree and rewrite the instruction generating the result
5001 * of the tree. When processing the tree, we know that the
5002 * instructions generated are all writing temporaries that are dead
5003 * outside of this tree. So, if we have some instructions that write
5004 * a temporary, we're free to point that temp write somewhere else.
5006 * Note that this doesn't guarantee that the instruction generated
5007 * only reg -- it might be the size=4 destination of a texture instruction.
5010 fs_visitor::get_instruction_generating_reg(fs_inst
*start
,
5015 end
->is_partial_write() ||
5016 !reg
.equals(end
->dst
)) {
5024 fs_visitor::setup_fs_payload_gen6()
5026 assert(stage
== MESA_SHADER_FRAGMENT
);
5027 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
5028 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
5030 unsigned barycentric_interp_modes
=
5031 (stage
== MESA_SHADER_FRAGMENT
) ?
5032 ((brw_wm_prog_data
*) this->prog_data
)->barycentric_interp_modes
: 0;
5034 assert(devinfo
->gen
>= 6);
5036 /* R0-1: masks, pixel X/Y coordinates. */
5037 payload
.num_regs
= 2;
5038 /* R2: only for 32-pixel dispatch.*/
5040 /* R3-26: barycentric interpolation coordinates. These appear in the
5041 * same order that they appear in the brw_wm_barycentric_interp_mode
5042 * enum. Each set of coordinates occupies 2 registers if dispatch width
5043 * == 8 and 4 registers if dispatch width == 16. Coordinates only
5044 * appear if they were enabled using the "Barycentric Interpolation
5045 * Mode" bits in WM_STATE.
5047 for (int i
= 0; i
< BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT
; ++i
) {
5048 if (barycentric_interp_modes
& (1 << i
)) {
5049 payload
.barycentric_coord_reg
[i
] = payload
.num_regs
;
5050 payload
.num_regs
+= 2;
5051 if (dispatch_width
== 16) {
5052 payload
.num_regs
+= 2;
5057 /* R27: interpolated depth if uses source depth */
5058 prog_data
->uses_src_depth
=
5059 (nir
->info
.inputs_read
& (1 << VARYING_SLOT_POS
)) != 0;
5060 if (prog_data
->uses_src_depth
) {
5061 payload
.source_depth_reg
= payload
.num_regs
;
5063 if (dispatch_width
== 16) {
5064 /* R28: interpolated depth if not SIMD8. */
5069 /* R29: interpolated W set if GEN6_WM_USES_SOURCE_W. */
5070 prog_data
->uses_src_w
=
5071 (nir
->info
.inputs_read
& (1 << VARYING_SLOT_POS
)) != 0;
5072 if (prog_data
->uses_src_w
) {
5073 payload
.source_w_reg
= payload
.num_regs
;
5075 if (dispatch_width
== 16) {
5076 /* R30: interpolated W if not SIMD8. */
5081 prog_data
->uses_pos_offset
= key
->compute_pos_offset
;
5082 /* R31: MSAA position offsets. */
5083 if (prog_data
->uses_pos_offset
) {
5084 payload
.sample_pos_reg
= payload
.num_regs
;
5088 /* R32: MSAA input coverage mask */
5089 prog_data
->uses_sample_mask
=
5090 (nir
->info
.system_values_read
& SYSTEM_BIT_SAMPLE_MASK_IN
) != 0;
5091 if (prog_data
->uses_sample_mask
) {
5092 assert(devinfo
->gen
>= 7);
5093 payload
.sample_mask_in_reg
= payload
.num_regs
;
5095 if (dispatch_width
== 16) {
5096 /* R33: input coverage mask if not SIMD8. */
5101 /* R34-: bary for 32-pixel. */
5102 /* R58-59: interp W for 32-pixel. */
5104 if (nir
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_DEPTH
)) {
5105 source_depth_to_render_target
= true;
5110 fs_visitor::setup_vs_payload()
5112 /* R0: thread header, R1: urb handles */
5113 payload
.num_regs
= 2;
5117 * We are building the local ID push constant data using the simplest possible
5118 * method. We simply push the local IDs directly as they should appear in the
5119 * registers for the uvec3 gl_LocalInvocationID variable.
5121 * Therefore, for SIMD8, we use 3 full registers, and for SIMD16 we use 6
5122 * registers worth of push constant space.
5124 * Note: Any updates to brw_cs_prog_local_id_payload_dwords,
5125 * fill_local_id_payload or fs_visitor::emit_cs_local_invocation_id_setup need
5128 * FINISHME: There are a few easy optimizations to consider.
5130 * 1. If gl_WorkGroupSize x, y or z is 1, we can just use zero, and there is
5131 * no need for using push constant space for that dimension.
5133 * 2. Since GL_MAX_COMPUTE_WORK_GROUP_SIZE is currently 1024 or less, we can
5134 * easily use 16-bit words rather than 32-bit dwords in the push constant
5137 * 3. If gl_WorkGroupSize x, y or z is small, then we can use bytes for
5138 * conveying the data, and thereby reduce push constant usage.
5142 fs_visitor::setup_gs_payload()
5144 assert(stage
== MESA_SHADER_GEOMETRY
);
5146 struct brw_gs_prog_data
*gs_prog_data
=
5147 (struct brw_gs_prog_data
*) prog_data
;
5148 struct brw_vue_prog_data
*vue_prog_data
=
5149 (struct brw_vue_prog_data
*) prog_data
;
5151 /* R0: thread header, R1: output URB handles */
5152 payload
.num_regs
= 2;
5154 if (gs_prog_data
->include_primitive_id
) {
5155 /* R2: Primitive ID 0..7 */
5159 /* Use a maximum of 32 registers for push-model inputs. */
5160 const unsigned max_push_components
= 32;
5162 /* If pushing our inputs would take too many registers, reduce the URB read
5163 * length (which is in HWords, or 8 registers), and resort to pulling.
5165 * Note that the GS reads <URB Read Length> HWords for every vertex - so we
5166 * have to multiply by VerticesIn to obtain the total storage requirement.
5168 if (8 * vue_prog_data
->urb_read_length
* nir
->info
.gs
.vertices_in
>
5169 max_push_components
) {
5170 gs_prog_data
->base
.include_vue_handles
= true;
5172 /* R3..RN: ICP Handles for each incoming vertex (when using pull model) */
5173 payload
.num_regs
+= nir
->info
.gs
.vertices_in
;
5175 vue_prog_data
->urb_read_length
=
5176 ROUND_DOWN_TO(max_push_components
/ nir
->info
.gs
.vertices_in
, 8) / 8;
5181 fs_visitor::setup_cs_payload()
5183 assert(devinfo
->gen
>= 7);
5184 brw_cs_prog_data
*prog_data
= (brw_cs_prog_data
*) this->prog_data
;
5186 payload
.num_regs
= 1;
5188 if (nir
->info
.system_values_read
& SYSTEM_BIT_LOCAL_INVOCATION_ID
) {
5189 prog_data
->local_invocation_id_regs
= dispatch_width
* 3 / 8;
5190 payload
.local_invocation_id_reg
= payload
.num_regs
;
5191 payload
.num_regs
+= prog_data
->local_invocation_id_regs
;
5196 fs_visitor::calculate_register_pressure()
5198 invalidate_live_intervals();
5199 calculate_live_intervals();
5201 unsigned num_instructions
= 0;
5202 foreach_block(block
, cfg
)
5203 num_instructions
+= block
->instructions
.length();
5205 regs_live_at_ip
= rzalloc_array(mem_ctx
, int, num_instructions
);
5207 for (unsigned reg
= 0; reg
< alloc
.count
; reg
++) {
5208 for (int ip
= virtual_grf_start
[reg
]; ip
<= virtual_grf_end
[reg
]; ip
++)
5209 regs_live_at_ip
[ip
] += alloc
.sizes
[reg
];
5214 * Look for repeated FS_OPCODE_MOV_DISPATCH_TO_FLAGS and drop the later ones.
5216 * The needs_unlit_centroid_workaround ends up producing one of these per
5217 * channel of centroid input, so it's good to clean them up.
5219 * An assumption here is that nothing ever modifies the dispatched pixels
5220 * value that FS_OPCODE_MOV_DISPATCH_TO_FLAGS reads from, but the hardware
5221 * dictates that anyway.
5224 fs_visitor::opt_drop_redundant_mov_to_flags()
5226 bool flag_mov_found
[2] = {false};
5227 bool progress
= false;
5229 /* Instructions removed by this pass can only be added if this were true */
5230 if (!devinfo
->needs_unlit_centroid_workaround
)
5233 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
5234 if (inst
->is_control_flow()) {
5235 memset(flag_mov_found
, 0, sizeof(flag_mov_found
));
5236 } else if (inst
->opcode
== FS_OPCODE_MOV_DISPATCH_TO_FLAGS
) {
5237 if (!flag_mov_found
[inst
->flag_subreg
]) {
5238 flag_mov_found
[inst
->flag_subreg
] = true;
5240 inst
->remove(block
);
5243 } else if (inst
->writes_flag()) {
5244 flag_mov_found
[inst
->flag_subreg
] = false;
5252 fs_visitor::optimize()
5254 /* Start by validating the shader we currently have. */
5257 /* bld is the common builder object pointing at the end of the program we
5258 * used to translate it into i965 IR. For the optimization and lowering
5259 * passes coming next, any code added after the end of the program without
5260 * having explicitly called fs_builder::at() clearly points at a mistake.
5261 * Ideally optimization passes wouldn't be part of the visitor so they
5262 * wouldn't have access to bld at all, but they do, so just in case some
5263 * pass forgets to ask for a location explicitly set it to NULL here to
5264 * make it trip. The dispatch width is initialized to a bogus value to
5265 * make sure that optimizations set the execution controls explicitly to
5266 * match the code they are manipulating instead of relying on the defaults.
5268 bld
= fs_builder(this, 64);
5270 assign_constant_locations();
5271 lower_constant_loads();
5275 split_virtual_grfs();
5278 #define OPT(pass, args...) ({ \
5280 bool this_progress = pass(args); \
5282 if (unlikely(INTEL_DEBUG & DEBUG_OPTIMIZER) && this_progress) { \
5283 char filename[64]; \
5284 snprintf(filename, 64, "%s%d-%s-%02d-%02d-" #pass, \
5285 stage_abbrev, dispatch_width, nir->info.name, iteration, pass_num); \
5287 backend_shader::dump_instructions(filename); \
5292 progress = progress || this_progress; \
5296 if (unlikely(INTEL_DEBUG
& DEBUG_OPTIMIZER
)) {
5298 snprintf(filename
, 64, "%s%d-%s-00-00-start",
5299 stage_abbrev
, dispatch_width
, nir
->info
.name
);
5301 backend_shader::dump_instructions(filename
);
5304 bool progress
= false;
5308 OPT(opt_drop_redundant_mov_to_flags
);
5310 OPT(lower_simd_width
);
5311 OPT(lower_logical_sends
);
5318 OPT(remove_duplicate_mrf_writes
);
5322 OPT(opt_copy_propagate
);
5323 OPT(opt_predicated_break
, this);
5324 OPT(opt_cmod_propagation
);
5325 OPT(dead_code_eliminate
);
5326 OPT(opt_peephole_sel
);
5327 OPT(dead_control_flow_eliminate
, this);
5328 OPT(opt_register_renaming
);
5329 OPT(opt_redundant_discard_jumps
);
5330 OPT(opt_saturate_propagation
);
5331 OPT(opt_zero_samples
);
5332 OPT(register_coalesce
);
5333 OPT(compute_to_mrf
);
5334 OPT(eliminate_find_live_channel
);
5336 OPT(compact_virtual_grfs
);
5341 OPT(opt_sampler_eot
);
5343 if (OPT(lower_load_payload
)) {
5344 split_virtual_grfs();
5345 OPT(register_coalesce
);
5346 OPT(compute_to_mrf
);
5347 OPT(dead_code_eliminate
);
5350 OPT(opt_combine_constants
);
5351 OPT(lower_integer_multiplication
);
5353 if (devinfo
->gen
<= 5 && OPT(lower_minmax
)) {
5354 OPT(opt_cmod_propagation
);
5356 OPT(opt_copy_propagate
);
5357 OPT(dead_code_eliminate
);
5360 lower_uniform_pull_constant_loads();
5366 * Three source instruction must have a GRF/MRF destination register.
5367 * ARF NULL is not allowed. Fix that up by allocating a temporary GRF.
5370 fs_visitor::fixup_3src_null_dest()
5372 bool progress
= false;
5374 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
5375 if (inst
->is_3src() && inst
->dst
.is_null()) {
5376 inst
->dst
= fs_reg(VGRF
, alloc
.allocate(dispatch_width
/ 8),
5383 invalidate_live_intervals();
5387 fs_visitor::allocate_registers()
5389 bool allocated_without_spills
;
5391 static const enum instruction_scheduler_mode pre_modes
[] = {
5393 SCHEDULE_PRE_NON_LIFO
,
5397 /* Try each scheduling heuristic to see if it can successfully register
5398 * allocate without spilling. They should be ordered by decreasing
5399 * performance but increasing likelihood of allocating.
5401 for (unsigned i
= 0; i
< ARRAY_SIZE(pre_modes
); i
++) {
5402 schedule_instructions(pre_modes
[i
]);
5405 assign_regs_trivial();
5406 allocated_without_spills
= true;
5408 allocated_without_spills
= assign_regs(false);
5410 if (allocated_without_spills
)
5414 if (!allocated_without_spills
) {
5415 /* We assume that any spilling is worse than just dropping back to
5416 * SIMD8. There's probably actually some intermediate point where
5417 * SIMD16 with a couple of spills is still better.
5419 if (dispatch_width
== 16 && min_dispatch_width
<= 8) {
5420 fail("Failure to register allocate. Reduce number of "
5421 "live scalar values to avoid this.");
5423 compiler
->shader_perf_log(log_data
,
5424 "%s shader triggered register spilling. "
5425 "Try reducing the number of live scalar "
5426 "values to improve performance.\n",
5430 /* Since we're out of heuristics, just go spill registers until we
5431 * get an allocation.
5433 while (!assign_regs(true)) {
5439 /* This must come after all optimization and register allocation, since
5440 * it inserts dead code that happens to have side effects, and it does
5441 * so based on the actual physical registers in use.
5443 insert_gen4_send_dependency_workarounds();
5448 schedule_instructions(SCHEDULE_POST
);
5450 if (last_scratch
> 0)
5451 prog_data
->total_scratch
= brw_get_scratch_size(last_scratch
);
5455 fs_visitor::run_vs(gl_clip_plane
*clip_planes
)
5457 assert(stage
== MESA_SHADER_VERTEX
);
5461 if (shader_time_index
>= 0)
5462 emit_shader_time_begin();
5469 compute_clip_distance(clip_planes
);
5473 if (shader_time_index
>= 0)
5474 emit_shader_time_end();
5480 assign_curb_setup();
5481 assign_vs_urb_setup();
5483 fixup_3src_null_dest();
5484 allocate_registers();
5490 fs_visitor::run_tcs_single_patch()
5492 assert(stage
== MESA_SHADER_TESS_CTRL
);
5494 struct brw_tcs_prog_data
*tcs_prog_data
=
5495 (struct brw_tcs_prog_data
*) prog_data
;
5497 /* r1-r4 contain the ICP handles. */
5498 payload
.num_regs
= 5;
5500 if (shader_time_index
>= 0)
5501 emit_shader_time_begin();
5503 /* Initialize gl_InvocationID */
5504 fs_reg channels_uw
= bld
.vgrf(BRW_REGISTER_TYPE_UW
);
5505 fs_reg channels_ud
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
5506 bld
.MOV(channels_uw
, fs_reg(brw_imm_uv(0x76543210)));
5507 bld
.MOV(channels_ud
, channels_uw
);
5509 if (tcs_prog_data
->instances
== 1) {
5510 invocation_id
= channels_ud
;
5512 invocation_id
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
5514 /* Get instance number from g0.2 bits 23:17, and multiply it by 8. */
5515 fs_reg t
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
5516 fs_reg instance_times_8
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
5517 bld
.AND(t
, fs_reg(retype(brw_vec1_grf(0, 2), BRW_REGISTER_TYPE_UD
)),
5518 brw_imm_ud(INTEL_MASK(23, 17)));
5519 bld
.SHR(instance_times_8
, t
, brw_imm_ud(17 - 3));
5521 bld
.ADD(invocation_id
, instance_times_8
, channels_ud
);
5524 /* Fix the disptach mask */
5525 if (nir
->info
.tcs
.vertices_out
% 8) {
5526 bld
.CMP(bld
.null_reg_ud(), invocation_id
,
5527 brw_imm_ud(nir
->info
.tcs
.vertices_out
), BRW_CONDITIONAL_L
);
5528 bld
.IF(BRW_PREDICATE_NORMAL
);
5533 if (nir
->info
.tcs
.vertices_out
% 8) {
5534 bld
.emit(BRW_OPCODE_ENDIF
);
5537 /* Emit EOT write; set TR DS Cache bit */
5539 fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
)),
5540 fs_reg(brw_imm_ud(WRITEMASK_X
<< 16)),
5541 fs_reg(brw_imm_ud(0)),
5543 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 3);
5544 bld
.LOAD_PAYLOAD(payload
, srcs
, 3, 2);
5546 fs_inst
*inst
= bld
.emit(SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
,
5547 bld
.null_reg_ud(), payload
);
5549 inst
->base_mrf
= -1;
5552 if (shader_time_index
>= 0)
5553 emit_shader_time_end();
5562 assign_curb_setup();
5563 assign_tcs_single_patch_urb_setup();
5565 fixup_3src_null_dest();
5566 allocate_registers();
5572 fs_visitor::run_tes()
5574 assert(stage
== MESA_SHADER_TESS_EVAL
);
5576 /* R0: thread header, R1-3: gl_TessCoord.xyz, R4: URB handles */
5577 payload
.num_regs
= 5;
5579 if (shader_time_index
>= 0)
5580 emit_shader_time_begin();
5589 if (shader_time_index
>= 0)
5590 emit_shader_time_end();
5596 assign_curb_setup();
5597 assign_tes_urb_setup();
5599 fixup_3src_null_dest();
5600 allocate_registers();
5606 fs_visitor::run_gs()
5608 assert(stage
== MESA_SHADER_GEOMETRY
);
5612 this->final_gs_vertex_count
= vgrf(glsl_type::uint_type
);
5614 if (gs_compile
->control_data_header_size_bits
> 0) {
5615 /* Create a VGRF to store accumulated control data bits. */
5616 this->control_data_bits
= vgrf(glsl_type::uint_type
);
5618 /* If we're outputting more than 32 control data bits, then EmitVertex()
5619 * will set control_data_bits to 0 after emitting the first vertex.
5620 * Otherwise, we need to initialize it to 0 here.
5622 if (gs_compile
->control_data_header_size_bits
<= 32) {
5623 const fs_builder abld
= bld
.annotate("initialize control data bits");
5624 abld
.MOV(this->control_data_bits
, brw_imm_ud(0u));
5628 if (shader_time_index
>= 0)
5629 emit_shader_time_begin();
5633 emit_gs_thread_end();
5635 if (shader_time_index
>= 0)
5636 emit_shader_time_end();
5645 assign_curb_setup();
5646 assign_gs_urb_setup();
5648 fixup_3src_null_dest();
5649 allocate_registers();
5655 fs_visitor::run_fs(bool do_rep_send
)
5657 brw_wm_prog_data
*wm_prog_data
= (brw_wm_prog_data
*) this->prog_data
;
5658 brw_wm_prog_key
*wm_key
= (brw_wm_prog_key
*) this->key
;
5660 assert(stage
== MESA_SHADER_FRAGMENT
);
5662 if (devinfo
->gen
>= 6)
5663 setup_fs_payload_gen6();
5665 setup_fs_payload_gen4();
5669 } else if (do_rep_send
) {
5670 assert(dispatch_width
== 16);
5671 emit_repclear_shader();
5673 if (shader_time_index
>= 0)
5674 emit_shader_time_begin();
5676 calculate_urb_setup();
5677 if (nir
->info
.inputs_read
> 0) {
5678 if (devinfo
->gen
< 6)
5679 emit_interpolation_setup_gen4();
5681 emit_interpolation_setup_gen6();
5684 /* We handle discards by keeping track of the still-live pixels in f0.1.
5685 * Initialize it with the dispatched pixels.
5687 if (wm_prog_data
->uses_kill
) {
5688 fs_inst
*discard_init
= bld
.emit(FS_OPCODE_MOV_DISPATCH_TO_FLAGS
);
5689 discard_init
->flag_subreg
= 1;
5692 /* Generate FS IR for main(). (the visitor only descends into
5693 * functions called "main").
5700 if (wm_prog_data
->uses_kill
)
5701 bld
.emit(FS_OPCODE_PLACEHOLDER_HALT
);
5703 if (wm_key
->alpha_test_func
)
5708 if (shader_time_index
>= 0)
5709 emit_shader_time_end();
5715 assign_curb_setup();
5718 fixup_3src_null_dest();
5719 allocate_registers();
5725 if (dispatch_width
== 8)
5726 wm_prog_data
->reg_blocks
= brw_register_blocks(grf_used
);
5728 wm_prog_data
->reg_blocks_16
= brw_register_blocks(grf_used
);
5734 fs_visitor::run_cs()
5736 assert(stage
== MESA_SHADER_COMPUTE
);
5740 if (shader_time_index
>= 0)
5741 emit_shader_time_begin();
5743 if (devinfo
->is_haswell
&& prog_data
->total_shared
> 0) {
5744 /* Move SLM index from g0.0[27:24] to sr0.1[11:8] */
5745 const fs_builder abld
= bld
.exec_all().group(1, 0);
5746 abld
.MOV(retype(suboffset(brw_sr0_reg(), 1), BRW_REGISTER_TYPE_UW
),
5747 suboffset(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW
), 1));
5755 emit_cs_terminate();
5757 if (shader_time_index
>= 0)
5758 emit_shader_time_end();
5764 assign_curb_setup();
5766 fixup_3src_null_dest();
5767 allocate_registers();
5776 * Return a bitfield where bit n is set if barycentric interpolation mode n
5777 * (see enum brw_wm_barycentric_interp_mode) is needed by the fragment shader.
5780 brw_compute_barycentric_interp_modes(const struct brw_device_info
*devinfo
,
5781 bool shade_model_flat
,
5782 bool persample_shading
,
5783 const nir_shader
*shader
)
5785 unsigned barycentric_interp_modes
= 0;
5787 nir_foreach_variable(var
, &shader
->inputs
) {
5788 enum glsl_interp_qualifier interp_qualifier
=
5789 (enum glsl_interp_qualifier
)var
->data
.interpolation
;
5790 bool is_centroid
= var
->data
.centroid
&& !persample_shading
;
5791 bool is_sample
= var
->data
.sample
|| persample_shading
;
5792 bool is_gl_Color
= (var
->data
.location
== VARYING_SLOT_COL0
) ||
5793 (var
->data
.location
== VARYING_SLOT_COL1
);
5795 /* Ignore WPOS and FACE, because they don't require interpolation. */
5796 if (var
->data
.location
== VARYING_SLOT_POS
||
5797 var
->data
.location
== VARYING_SLOT_FACE
)
5800 /* Determine the set (or sets) of barycentric coordinates needed to
5801 * interpolate this variable. Note that when
5802 * brw->needs_unlit_centroid_workaround is set, centroid interpolation
5803 * uses PIXEL interpolation for unlit pixels and CENTROID interpolation
5804 * for lit pixels, so we need both sets of barycentric coordinates.
5806 if (interp_qualifier
== INTERP_QUALIFIER_NOPERSPECTIVE
) {
5808 barycentric_interp_modes
|=
5809 1 << BRW_WM_NONPERSPECTIVE_CENTROID_BARYCENTRIC
;
5810 } else if (is_sample
) {
5811 barycentric_interp_modes
|=
5812 1 << BRW_WM_NONPERSPECTIVE_SAMPLE_BARYCENTRIC
;
5814 if ((!is_centroid
&& !is_sample
) ||
5815 devinfo
->needs_unlit_centroid_workaround
) {
5816 barycentric_interp_modes
|=
5817 1 << BRW_WM_NONPERSPECTIVE_PIXEL_BARYCENTRIC
;
5819 } else if (interp_qualifier
== INTERP_QUALIFIER_SMOOTH
||
5820 (!(shade_model_flat
&& is_gl_Color
) &&
5821 interp_qualifier
== INTERP_QUALIFIER_NONE
)) {
5823 barycentric_interp_modes
|=
5824 1 << BRW_WM_PERSPECTIVE_CENTROID_BARYCENTRIC
;
5825 } else if (is_sample
) {
5826 barycentric_interp_modes
|=
5827 1 << BRW_WM_PERSPECTIVE_SAMPLE_BARYCENTRIC
;
5829 if ((!is_centroid
&& !is_sample
) ||
5830 devinfo
->needs_unlit_centroid_workaround
) {
5831 barycentric_interp_modes
|=
5832 1 << BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
;
5837 return barycentric_interp_modes
;
5841 brw_compute_flat_inputs(struct brw_wm_prog_data
*prog_data
,
5842 bool shade_model_flat
, const nir_shader
*shader
)
5844 prog_data
->flat_inputs
= 0;
5846 nir_foreach_variable(var
, &shader
->inputs
) {
5847 enum glsl_interp_qualifier interp_qualifier
=
5848 (enum glsl_interp_qualifier
)var
->data
.interpolation
;
5849 bool is_gl_Color
= (var
->data
.location
== VARYING_SLOT_COL0
) ||
5850 (var
->data
.location
== VARYING_SLOT_COL1
);
5852 int input_index
= prog_data
->urb_setup
[var
->data
.location
];
5854 if (input_index
< 0)
5858 if (interp_qualifier
== INTERP_QUALIFIER_FLAT
||
5859 (shade_model_flat
&& is_gl_Color
&&
5860 interp_qualifier
== INTERP_QUALIFIER_NONE
))
5861 prog_data
->flat_inputs
|= (1 << input_index
);
5866 computed_depth_mode(const nir_shader
*shader
)
5868 if (shader
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_DEPTH
)) {
5869 switch (shader
->info
.fs
.depth_layout
) {
5870 case FRAG_DEPTH_LAYOUT_NONE
:
5871 case FRAG_DEPTH_LAYOUT_ANY
:
5872 return BRW_PSCDEPTH_ON
;
5873 case FRAG_DEPTH_LAYOUT_GREATER
:
5874 return BRW_PSCDEPTH_ON_GE
;
5875 case FRAG_DEPTH_LAYOUT_LESS
:
5876 return BRW_PSCDEPTH_ON_LE
;
5877 case FRAG_DEPTH_LAYOUT_UNCHANGED
:
5878 return BRW_PSCDEPTH_OFF
;
5881 return BRW_PSCDEPTH_OFF
;
5885 brw_compile_fs(const struct brw_compiler
*compiler
, void *log_data
,
5887 const struct brw_wm_prog_key
*key
,
5888 struct brw_wm_prog_data
*prog_data
,
5889 const nir_shader
*src_shader
,
5890 struct gl_program
*prog
,
5891 int shader_time_index8
, int shader_time_index16
,
5893 unsigned *final_assembly_size
,
5896 nir_shader
*shader
= nir_shader_clone(mem_ctx
, src_shader
);
5897 shader
= brw_nir_apply_sampler_key(shader
, compiler
->devinfo
, &key
->tex
,
5899 brw_nir_lower_fs_inputs(shader
);
5900 brw_nir_lower_fs_outputs(shader
);
5901 shader
= brw_postprocess_nir(shader
, compiler
->devinfo
, true);
5903 /* key->alpha_test_func means simulating alpha testing via discards,
5904 * so the shader definitely kills pixels.
5906 prog_data
->uses_kill
= shader
->info
.fs
.uses_discard
|| key
->alpha_test_func
;
5907 prog_data
->uses_omask
= key
->multisample_fbo
&&
5908 shader
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_SAMPLE_MASK
);
5909 prog_data
->computed_depth_mode
= computed_depth_mode(shader
);
5910 prog_data
->computed_stencil
=
5911 shader
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_STENCIL
);
5913 prog_data
->early_fragment_tests
= shader
->info
.fs
.early_fragment_tests
;
5915 prog_data
->barycentric_interp_modes
=
5916 brw_compute_barycentric_interp_modes(compiler
->devinfo
,
5918 key
->persample_shading
,
5921 fs_visitor
v(compiler
, log_data
, mem_ctx
, key
,
5922 &prog_data
->base
, prog
, shader
, 8,
5923 shader_time_index8
);
5924 if (!v
.run_fs(false /* do_rep_send */)) {
5926 *error_str
= ralloc_strdup(mem_ctx
, v
.fail_msg
);
5931 cfg_t
*simd16_cfg
= NULL
;
5932 fs_visitor
v2(compiler
, log_data
, mem_ctx
, key
,
5933 &prog_data
->base
, prog
, shader
, 16,
5934 shader_time_index16
);
5935 if (likely(!(INTEL_DEBUG
& DEBUG_NO16
) || use_rep_send
)) {
5936 if (!v
.simd16_unsupported
) {
5937 /* Try a SIMD16 compile */
5938 v2
.import_uniforms(&v
);
5939 if (!v2
.run_fs(use_rep_send
)) {
5940 compiler
->shader_perf_log(log_data
,
5941 "SIMD16 shader failed to compile: %s",
5944 simd16_cfg
= v2
.cfg
;
5949 /* We have to compute the flat inputs after the visitor is finished running
5950 * because it relies on prog_data->urb_setup which is computed in
5951 * fs_visitor::calculate_urb_setup().
5953 brw_compute_flat_inputs(prog_data
, key
->flat_shade
, shader
);
5956 int no_simd8
= (INTEL_DEBUG
& DEBUG_NO8
) || use_rep_send
;
5957 if ((no_simd8
|| compiler
->devinfo
->gen
< 5) && simd16_cfg
) {
5959 prog_data
->no_8
= true;
5962 prog_data
->no_8
= false;
5965 fs_generator
g(compiler
, log_data
, mem_ctx
, (void *) key
, &prog_data
->base
,
5966 v
.promoted_constants
, v
.runtime_check_aads_emit
,
5967 MESA_SHADER_FRAGMENT
);
5969 if (unlikely(INTEL_DEBUG
& DEBUG_WM
)) {
5970 g
.enable_debug(ralloc_asprintf(mem_ctx
, "%s fragment shader %s",
5971 shader
->info
.label
? shader
->info
.label
:
5973 shader
->info
.name
));
5977 g
.generate_code(simd8_cfg
, 8);
5979 prog_data
->prog_offset_16
= g
.generate_code(simd16_cfg
, 16);
5981 return g
.get_assembly(final_assembly_size
);
5985 fs_visitor::emit_cs_local_invocation_id_setup()
5987 assert(stage
== MESA_SHADER_COMPUTE
);
5989 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::uvec3_type
));
5991 struct brw_reg src
=
5992 brw_vec8_grf(payload
.local_invocation_id_reg
, 0);
5993 src
= retype(src
, BRW_REGISTER_TYPE_UD
);
5995 src
.nr
+= dispatch_width
/ 8;
5996 bld
.MOV(offset(*reg
, bld
, 1), src
);
5997 src
.nr
+= dispatch_width
/ 8;
5998 bld
.MOV(offset(*reg
, bld
, 2), src
);
6004 fs_visitor::emit_cs_work_group_id_setup()
6006 assert(stage
== MESA_SHADER_COMPUTE
);
6008 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::uvec3_type
));
6010 struct brw_reg
r0_1(retype(brw_vec1_grf(0, 1), BRW_REGISTER_TYPE_UD
));
6011 struct brw_reg
r0_6(retype(brw_vec1_grf(0, 6), BRW_REGISTER_TYPE_UD
));
6012 struct brw_reg
r0_7(retype(brw_vec1_grf(0, 7), BRW_REGISTER_TYPE_UD
));
6014 bld
.MOV(*reg
, r0_1
);
6015 bld
.MOV(offset(*reg
, bld
, 1), r0_6
);
6016 bld
.MOV(offset(*reg
, bld
, 2), r0_7
);
6022 brw_compile_cs(const struct brw_compiler
*compiler
, void *log_data
,
6024 const struct brw_cs_prog_key
*key
,
6025 struct brw_cs_prog_data
*prog_data
,
6026 const nir_shader
*src_shader
,
6027 int shader_time_index
,
6028 unsigned *final_assembly_size
,
6031 nir_shader
*shader
= nir_shader_clone(mem_ctx
, src_shader
);
6032 shader
= brw_nir_apply_sampler_key(shader
, compiler
->devinfo
, &key
->tex
,
6034 brw_nir_lower_cs_shared(shader
);
6035 prog_data
->base
.total_shared
+= shader
->num_shared
;
6036 shader
= brw_postprocess_nir(shader
, compiler
->devinfo
, true);
6038 prog_data
->local_size
[0] = shader
->info
.cs
.local_size
[0];
6039 prog_data
->local_size
[1] = shader
->info
.cs
.local_size
[1];
6040 prog_data
->local_size
[2] = shader
->info
.cs
.local_size
[2];
6041 unsigned local_workgroup_size
=
6042 shader
->info
.cs
.local_size
[0] * shader
->info
.cs
.local_size
[1] *
6043 shader
->info
.cs
.local_size
[2];
6045 unsigned max_cs_threads
= compiler
->devinfo
->max_cs_threads
;
6046 unsigned simd_required
= DIV_ROUND_UP(local_workgroup_size
, max_cs_threads
);
6049 const char *fail_msg
= NULL
;
6051 /* Now the main event: Visit the shader IR and generate our CS IR for it.
6053 fs_visitor
v8(compiler
, log_data
, mem_ctx
, key
, &prog_data
->base
,
6054 NULL
, /* Never used in core profile */
6055 shader
, 8, shader_time_index
);
6056 if (simd_required
<= 8) {
6058 fail_msg
= v8
.fail_msg
;
6061 prog_data
->simd_size
= 8;
6065 fs_visitor
v16(compiler
, log_data
, mem_ctx
, key
, &prog_data
->base
,
6066 NULL
, /* Never used in core profile */
6067 shader
, 16, shader_time_index
);
6068 if (likely(!(INTEL_DEBUG
& DEBUG_NO16
)) &&
6069 !fail_msg
&& !v8
.simd16_unsupported
&&
6070 local_workgroup_size
<= 16 * max_cs_threads
) {
6071 /* Try a SIMD16 compile */
6072 if (simd_required
<= 8)
6073 v16
.import_uniforms(&v8
);
6074 if (!v16
.run_cs()) {
6075 compiler
->shader_perf_log(log_data
,
6076 "SIMD16 shader failed to compile: %s",
6080 "Couldn't generate SIMD16 program and not "
6081 "enough threads for SIMD8";
6085 prog_data
->simd_size
= 16;
6089 if (unlikely(cfg
== NULL
)) {
6092 *error_str
= ralloc_strdup(mem_ctx
, fail_msg
);
6097 fs_generator
g(compiler
, log_data
, mem_ctx
, (void*) key
, &prog_data
->base
,
6098 v8
.promoted_constants
, v8
.runtime_check_aads_emit
,
6099 MESA_SHADER_COMPUTE
);
6100 if (INTEL_DEBUG
& DEBUG_CS
) {
6101 char *name
= ralloc_asprintf(mem_ctx
, "%s compute shader %s",
6102 shader
->info
.label
? shader
->info
.label
:
6105 g
.enable_debug(name
);
6108 g
.generate_code(cfg
, prog_data
->simd_size
);
6110 return g
.get_assembly(final_assembly_size
);
6114 brw_cs_fill_local_id_payload(const struct brw_cs_prog_data
*prog_data
,
6115 void *buffer
, uint32_t threads
, uint32_t stride
)
6117 if (prog_data
->local_invocation_id_regs
== 0)
6120 /* 'stride' should be an integer number of registers, that is, a multiple
6123 assert(stride
% 32 == 0);
6125 unsigned x
= 0, y
= 0, z
= 0;
6126 for (unsigned t
= 0; t
< threads
; t
++) {
6127 uint32_t *param
= (uint32_t *) buffer
+ stride
* t
/ 4;
6129 for (unsigned i
= 0; i
< prog_data
->simd_size
; i
++) {
6130 param
[0 * prog_data
->simd_size
+ i
] = x
;
6131 param
[1 * prog_data
->simd_size
+ i
] = y
;
6132 param
[2 * prog_data
->simd_size
+ i
] = z
;
6135 if (x
== prog_data
->local_size
[0]) {
6138 if (y
== prog_data
->local_size
[1]) {
6141 if (z
== prog_data
->local_size
[2])