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 "glsl/nir/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::int_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;
436 this->reladdr
= NULL
;
438 if (this->file
== IMM
&&
439 (this->type
!= BRW_REGISTER_TYPE_V
&&
440 this->type
!= BRW_REGISTER_TYPE_UV
&&
441 this->type
!= BRW_REGISTER_TYPE_VF
)) {
447 fs_reg::equals(const fs_reg
&r
) const
449 return (this->backend_reg::equals(r
) &&
450 subreg_offset
== r
.subreg_offset
&&
451 !reladdr
&& !r
.reladdr
&&
456 fs_reg::set_smear(unsigned subreg
)
458 assert(file
!= ARF
&& file
!= FIXED_GRF
&& file
!= IMM
);
459 subreg_offset
= subreg
* type_sz(type
);
465 fs_reg::is_contiguous() const
471 fs_reg::component_size(unsigned width
) const
473 const unsigned stride
= ((file
!= ARF
&& file
!= FIXED_GRF
) ? this->stride
:
476 return MAX2(width
* stride
, 1) * type_sz(type
);
480 type_size_scalar(const struct glsl_type
*type
)
482 unsigned int size
, i
;
484 switch (type
->base_type
) {
487 case GLSL_TYPE_FLOAT
:
489 return type
->components();
490 case GLSL_TYPE_ARRAY
:
491 return type_size_scalar(type
->fields
.array
) * type
->length
;
492 case GLSL_TYPE_STRUCT
:
494 for (i
= 0; i
< type
->length
; i
++) {
495 size
+= type_size_scalar(type
->fields
.structure
[i
].type
);
498 case GLSL_TYPE_SAMPLER
:
499 /* Samplers take up no register space, since they're baked in at
503 case GLSL_TYPE_ATOMIC_UINT
:
505 case GLSL_TYPE_SUBROUTINE
:
507 case GLSL_TYPE_IMAGE
:
508 return BRW_IMAGE_PARAM_SIZE
;
510 case GLSL_TYPE_ERROR
:
511 case GLSL_TYPE_INTERFACE
:
512 case GLSL_TYPE_DOUBLE
:
513 case GLSL_TYPE_FUNCTION
:
514 unreachable("not reached");
521 * Returns the number of scalar components needed to store type, assuming
522 * that vectors are padded out to vec4.
524 * This has the packing rules of type_size_vec4(), but counts components
525 * similar to type_size_scalar().
528 type_size_vec4_times_4(const struct glsl_type
*type
)
530 return 4 * type_size_vec4(type
);
534 * Create a MOV to read the timestamp register.
536 * The caller is responsible for emitting the MOV. The return value is
537 * the destination of the MOV, with extra parameters set.
540 fs_visitor::get_timestamp(const fs_builder
&bld
)
542 assert(devinfo
->gen
>= 7);
544 fs_reg ts
= fs_reg(retype(brw_vec4_reg(BRW_ARCHITECTURE_REGISTER_FILE
,
547 BRW_REGISTER_TYPE_UD
));
549 fs_reg dst
= fs_reg(VGRF
, alloc
.allocate(1), BRW_REGISTER_TYPE_UD
);
551 /* We want to read the 3 fields we care about even if it's not enabled in
554 bld
.group(4, 0).exec_all().MOV(dst
, ts
);
560 fs_visitor::emit_shader_time_begin()
562 shader_start_time
= get_timestamp(bld
.annotate("shader time start"));
564 /* We want only the low 32 bits of the timestamp. Since it's running
565 * at the GPU clock rate of ~1.2ghz, it will roll over every ~3 seconds,
566 * which is plenty of time for our purposes. It is identical across the
567 * EUs, but since it's tracking GPU core speed it will increment at a
568 * varying rate as render P-states change.
570 shader_start_time
.set_smear(0);
574 fs_visitor::emit_shader_time_end()
576 /* Insert our code just before the final SEND with EOT. */
577 exec_node
*end
= this->instructions
.get_tail();
578 assert(end
&& ((fs_inst
*) end
)->eot
);
579 const fs_builder ibld
= bld
.annotate("shader time end")
580 .exec_all().at(NULL
, end
);
582 fs_reg shader_end_time
= get_timestamp(ibld
);
584 /* We only use the low 32 bits of the timestamp - see
585 * emit_shader_time_begin()).
587 * We could also check if render P-states have changed (or anything
588 * else that might disrupt timing) by setting smear to 2 and checking if
589 * that field is != 0.
591 shader_end_time
.set_smear(0);
593 /* Check that there weren't any timestamp reset events (assuming these
594 * were the only two timestamp reads that happened).
596 fs_reg reset
= shader_end_time
;
598 set_condmod(BRW_CONDITIONAL_Z
,
599 ibld
.AND(ibld
.null_reg_ud(), reset
, brw_imm_ud(1u)));
600 ibld
.IF(BRW_PREDICATE_NORMAL
);
602 fs_reg start
= shader_start_time
;
604 fs_reg diff
= fs_reg(VGRF
, alloc
.allocate(1), BRW_REGISTER_TYPE_UD
);
607 const fs_builder cbld
= ibld
.group(1, 0);
608 cbld
.group(1, 0).ADD(diff
, start
, shader_end_time
);
610 /* If there were no instructions between the two timestamp gets, the diff
611 * is 2 cycles. Remove that overhead, so I can forget about that when
612 * trying to determine the time taken for single instructions.
614 cbld
.ADD(diff
, diff
, brw_imm_ud(-2u));
615 SHADER_TIME_ADD(cbld
, 0, diff
);
616 SHADER_TIME_ADD(cbld
, 1, brw_imm_ud(1u));
617 ibld
.emit(BRW_OPCODE_ELSE
);
618 SHADER_TIME_ADD(cbld
, 2, brw_imm_ud(1u));
619 ibld
.emit(BRW_OPCODE_ENDIF
);
623 fs_visitor::SHADER_TIME_ADD(const fs_builder
&bld
,
624 int shader_time_subindex
,
627 int index
= shader_time_index
* 3 + shader_time_subindex
;
628 struct brw_reg offset
= brw_imm_d(index
* SHADER_TIME_STRIDE
);
631 if (dispatch_width
== 8)
632 payload
= vgrf(glsl_type::uvec2_type
);
634 payload
= vgrf(glsl_type::uint_type
);
636 bld
.emit(SHADER_OPCODE_SHADER_TIME_ADD
, fs_reg(), payload
, offset
, value
);
640 fs_visitor::vfail(const char *format
, va_list va
)
649 msg
= ralloc_vasprintf(mem_ctx
, format
, va
);
650 msg
= ralloc_asprintf(mem_ctx
, "%s compile failed: %s\n", stage_abbrev
, msg
);
652 this->fail_msg
= msg
;
655 fprintf(stderr
, "%s", msg
);
660 fs_visitor::fail(const char *format
, ...)
664 va_start(va
, format
);
670 * Mark this program as impossible to compile in SIMD16 mode.
672 * During the SIMD8 compile (which happens first), we can detect and flag
673 * things that are unsupported in SIMD16 mode, so the compiler can skip
674 * the SIMD16 compile altogether.
676 * During a SIMD16 compile (if one happens anyway), this just calls fail().
679 fs_visitor::no16(const char *msg
)
681 if (dispatch_width
== 16) {
684 simd16_unsupported
= true;
686 compiler
->shader_perf_log(log_data
,
687 "SIMD16 shader failed to compile: %s", msg
);
692 * Returns true if the instruction has a flag that means it won't
693 * update an entire destination register.
695 * For example, dead code elimination and live variable analysis want to know
696 * when a write to a variable screens off any preceding values that were in
700 fs_inst::is_partial_write() const
702 return ((this->predicate
&& this->opcode
!= BRW_OPCODE_SEL
) ||
703 (this->exec_size
* type_sz(this->dst
.type
)) < 32 ||
704 !this->dst
.is_contiguous());
708 fs_inst::components_read(unsigned i
) const
711 case FS_OPCODE_LINTERP
:
717 case FS_OPCODE_PIXEL_X
:
718 case FS_OPCODE_PIXEL_Y
:
722 case FS_OPCODE_FB_WRITE_LOGICAL
:
723 assert(src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].file
== IMM
);
724 /* First/second FB write color. */
726 return src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].ud
;
730 case SHADER_OPCODE_TEX_LOGICAL
:
731 case SHADER_OPCODE_TXD_LOGICAL
:
732 case SHADER_OPCODE_TXF_LOGICAL
:
733 case SHADER_OPCODE_TXL_LOGICAL
:
734 case SHADER_OPCODE_TXS_LOGICAL
:
735 case FS_OPCODE_TXB_LOGICAL
:
736 case SHADER_OPCODE_TXF_CMS_LOGICAL
:
737 case SHADER_OPCODE_TXF_CMS_W_LOGICAL
:
738 case SHADER_OPCODE_TXF_UMS_LOGICAL
:
739 case SHADER_OPCODE_TXF_MCS_LOGICAL
:
740 case SHADER_OPCODE_LOD_LOGICAL
:
741 case SHADER_OPCODE_TG4_LOGICAL
:
742 case SHADER_OPCODE_TG4_OFFSET_LOGICAL
:
743 assert(src
[9].file
== IMM
&& src
[10].file
== IMM
);
744 /* Texture coordinates. */
747 /* Texture derivatives. */
748 else if ((i
== 2 || i
== 3) && opcode
== SHADER_OPCODE_TXD_LOGICAL
)
750 /* Texture offset. */
754 else if (i
== 5 && 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
== FIXED_GRF
) {
855 /* If the start of the region is not register aligned, then
856 * there's some portion of the register that's technically
857 * unread at the beginning.
859 * However, the register allocator works in terms of whole
860 * registers, and does not use subnr. It assumes that the
861 * read starts at the beginning of the register, and extends
862 * regs_read() whole registers beyond that.
864 * To compensate, we extend the region length to include this
865 * unread portion at the beginning.
868 region_length
+= src
[0].subnr
* type_sz(src
[0].type
);
870 return DIV_ROUND_UP(region_length
, REG_SIZE
);
872 assert(!"Invalid register file");
878 if (is_tex() && arg
== 0 && src
[0].file
== VGRF
)
883 switch (src
[arg
].file
) {
893 return DIV_ROUND_UP(components_read(arg
) *
894 src
[arg
].component_size(exec_size
),
897 unreachable("MRF registers are not allowed as sources");
903 fs_inst::reads_flag() const
909 fs_inst::writes_flag() const
911 return (conditional_mod
&& (opcode
!= BRW_OPCODE_SEL
&&
912 opcode
!= BRW_OPCODE_IF
&&
913 opcode
!= BRW_OPCODE_WHILE
)) ||
914 opcode
== FS_OPCODE_MOV_DISPATCH_TO_FLAGS
;
918 * Returns how many MRFs an FS opcode will write over.
920 * Note that this is not the 0 or 1 implied writes in an actual gen
921 * instruction -- the FS opcodes often generate MOVs in addition.
924 fs_visitor::implied_mrf_writes(fs_inst
*inst
)
929 if (inst
->base_mrf
== -1)
932 switch (inst
->opcode
) {
933 case SHADER_OPCODE_RCP
:
934 case SHADER_OPCODE_RSQ
:
935 case SHADER_OPCODE_SQRT
:
936 case SHADER_OPCODE_EXP2
:
937 case SHADER_OPCODE_LOG2
:
938 case SHADER_OPCODE_SIN
:
939 case SHADER_OPCODE_COS
:
940 return 1 * dispatch_width
/ 8;
941 case SHADER_OPCODE_POW
:
942 case SHADER_OPCODE_INT_QUOTIENT
:
943 case SHADER_OPCODE_INT_REMAINDER
:
944 return 2 * dispatch_width
/ 8;
945 case SHADER_OPCODE_TEX
:
947 case SHADER_OPCODE_TXD
:
948 case SHADER_OPCODE_TXF
:
949 case SHADER_OPCODE_TXF_CMS
:
950 case SHADER_OPCODE_TXF_CMS_W
:
951 case SHADER_OPCODE_TXF_MCS
:
952 case SHADER_OPCODE_TG4
:
953 case SHADER_OPCODE_TG4_OFFSET
:
954 case SHADER_OPCODE_TXL
:
955 case SHADER_OPCODE_TXS
:
956 case SHADER_OPCODE_LOD
:
957 case SHADER_OPCODE_SAMPLEINFO
:
959 case FS_OPCODE_FB_WRITE
:
961 case FS_OPCODE_GET_BUFFER_SIZE
:
962 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
963 case SHADER_OPCODE_GEN4_SCRATCH_READ
:
965 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD
:
967 case SHADER_OPCODE_GEN4_SCRATCH_WRITE
:
969 case SHADER_OPCODE_UNTYPED_ATOMIC
:
970 case SHADER_OPCODE_UNTYPED_SURFACE_READ
:
971 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE
:
972 case SHADER_OPCODE_TYPED_ATOMIC
:
973 case SHADER_OPCODE_TYPED_SURFACE_READ
:
974 case SHADER_OPCODE_TYPED_SURFACE_WRITE
:
975 case SHADER_OPCODE_URB_WRITE_SIMD8
:
976 case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
977 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
:
978 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
979 case FS_OPCODE_INTERPOLATE_AT_CENTROID
:
980 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
981 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
982 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
985 unreachable("not reached");
990 fs_visitor::vgrf(const glsl_type
*const type
)
992 int reg_width
= dispatch_width
/ 8;
993 return fs_reg(VGRF
, alloc
.allocate(type_size_scalar(type
) * reg_width
),
994 brw_type_for_base_type(type
));
997 fs_reg::fs_reg(enum brw_reg_file file
, int nr
)
1002 this->type
= BRW_REGISTER_TYPE_F
;
1003 this->stride
= (file
== UNIFORM
? 0 : 1);
1006 fs_reg::fs_reg(enum brw_reg_file file
, int nr
, enum brw_reg_type type
)
1012 this->stride
= (file
== UNIFORM
? 0 : 1);
1015 /* For SIMD16, we need to follow from the uniform setup of SIMD8 dispatch.
1016 * This brings in those uniform definitions
1019 fs_visitor::import_uniforms(fs_visitor
*v
)
1021 this->push_constant_loc
= v
->push_constant_loc
;
1022 this->pull_constant_loc
= v
->pull_constant_loc
;
1023 this->uniforms
= v
->uniforms
;
1024 this->param_size
= v
->param_size
;
1028 fs_visitor::emit_fragcoord_interpolation(bool pixel_center_integer
,
1029 bool origin_upper_left
)
1031 assert(stage
== MESA_SHADER_FRAGMENT
);
1032 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1033 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::vec4_type
));
1035 bool flip
= !origin_upper_left
^ key
->render_to_fbo
;
1037 /* gl_FragCoord.x */
1038 if (pixel_center_integer
) {
1039 bld
.MOV(wpos
, this->pixel_x
);
1041 bld
.ADD(wpos
, this->pixel_x
, brw_imm_f(0.5f
));
1043 wpos
= offset(wpos
, bld
, 1);
1045 /* gl_FragCoord.y */
1046 if (!flip
&& pixel_center_integer
) {
1047 bld
.MOV(wpos
, this->pixel_y
);
1049 fs_reg pixel_y
= this->pixel_y
;
1050 float offset
= (pixel_center_integer
? 0.0f
: 0.5f
);
1053 pixel_y
.negate
= true;
1054 offset
+= key
->drawable_height
- 1.0f
;
1057 bld
.ADD(wpos
, pixel_y
, brw_imm_f(offset
));
1059 wpos
= offset(wpos
, bld
, 1);
1061 /* gl_FragCoord.z */
1062 if (devinfo
->gen
>= 6) {
1063 bld
.MOV(wpos
, fs_reg(brw_vec8_grf(payload
.source_depth_reg
, 0)));
1065 bld
.emit(FS_OPCODE_LINTERP
, wpos
,
1066 this->delta_xy
[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
],
1067 interp_reg(VARYING_SLOT_POS
, 2));
1069 wpos
= offset(wpos
, bld
, 1);
1071 /* gl_FragCoord.w: Already set up in emit_interpolation */
1072 bld
.MOV(wpos
, this->wpos_w
);
1078 fs_visitor::emit_linterp(const fs_reg
&attr
, const fs_reg
&interp
,
1079 glsl_interp_qualifier interpolation_mode
,
1080 bool is_centroid
, bool is_sample
)
1082 brw_wm_barycentric_interp_mode barycoord_mode
;
1083 if (devinfo
->gen
>= 6) {
1085 if (interpolation_mode
== INTERP_QUALIFIER_SMOOTH
)
1086 barycoord_mode
= BRW_WM_PERSPECTIVE_CENTROID_BARYCENTRIC
;
1088 barycoord_mode
= BRW_WM_NONPERSPECTIVE_CENTROID_BARYCENTRIC
;
1089 } else if (is_sample
) {
1090 if (interpolation_mode
== INTERP_QUALIFIER_SMOOTH
)
1091 barycoord_mode
= BRW_WM_PERSPECTIVE_SAMPLE_BARYCENTRIC
;
1093 barycoord_mode
= BRW_WM_NONPERSPECTIVE_SAMPLE_BARYCENTRIC
;
1095 if (interpolation_mode
== INTERP_QUALIFIER_SMOOTH
)
1096 barycoord_mode
= BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
;
1098 barycoord_mode
= BRW_WM_NONPERSPECTIVE_PIXEL_BARYCENTRIC
;
1101 /* On Ironlake and below, there is only one interpolation mode.
1102 * Centroid interpolation doesn't mean anything on this hardware --
1103 * there is no multisampling.
1105 barycoord_mode
= BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
;
1107 return bld
.emit(FS_OPCODE_LINTERP
, attr
,
1108 this->delta_xy
[barycoord_mode
], interp
);
1112 fs_visitor::emit_general_interpolation(fs_reg
*attr
, const char *name
,
1113 const glsl_type
*type
,
1114 glsl_interp_qualifier interpolation_mode
,
1115 int *location
, bool mod_centroid
,
1118 assert(stage
== MESA_SHADER_FRAGMENT
);
1119 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
1120 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1122 if (interpolation_mode
== INTERP_QUALIFIER_NONE
) {
1124 *location
== VARYING_SLOT_COL0
|| *location
== VARYING_SLOT_COL1
;
1125 if (key
->flat_shade
&& is_gl_Color
) {
1126 interpolation_mode
= INTERP_QUALIFIER_FLAT
;
1128 interpolation_mode
= INTERP_QUALIFIER_SMOOTH
;
1132 if (type
->is_array() || type
->is_matrix()) {
1133 const glsl_type
*elem_type
= glsl_get_array_element(type
);
1134 const unsigned length
= glsl_get_length(type
);
1136 for (unsigned i
= 0; i
< length
; i
++) {
1137 emit_general_interpolation(attr
, name
, elem_type
, interpolation_mode
,
1138 location
, mod_centroid
, mod_sample
);
1140 } else if (type
->is_record()) {
1141 for (unsigned i
= 0; i
< type
->length
; i
++) {
1142 const glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
1143 emit_general_interpolation(attr
, name
, field_type
, interpolation_mode
,
1144 location
, mod_centroid
, mod_sample
);
1147 assert(type
->is_scalar() || type
->is_vector());
1149 if (prog_data
->urb_setup
[*location
] == -1) {
1150 /* If there's no incoming setup data for this slot, don't
1151 * emit interpolation for it.
1153 *attr
= offset(*attr
, bld
, type
->vector_elements
);
1158 attr
->type
= brw_type_for_base_type(type
->get_scalar_type());
1160 if (interpolation_mode
== INTERP_QUALIFIER_FLAT
) {
1161 /* Constant interpolation (flat shading) case. The SF has
1162 * handed us defined values in only the constant offset
1163 * field of the setup reg.
1165 for (unsigned int i
= 0; i
< type
->vector_elements
; i
++) {
1166 struct brw_reg interp
= interp_reg(*location
, i
);
1167 interp
= suboffset(interp
, 3);
1168 interp
.type
= attr
->type
;
1169 bld
.emit(FS_OPCODE_CINTERP
, *attr
, fs_reg(interp
));
1170 *attr
= offset(*attr
, bld
, 1);
1173 /* Smooth/noperspective interpolation case. */
1174 for (unsigned int i
= 0; i
< type
->vector_elements
; i
++) {
1175 struct brw_reg interp
= interp_reg(*location
, i
);
1176 if (devinfo
->needs_unlit_centroid_workaround
&& mod_centroid
) {
1177 /* Get the pixel/sample mask into f0 so that we know
1178 * which pixels are lit. Then, for each channel that is
1179 * unlit, replace the centroid data with non-centroid
1182 bld
.emit(FS_OPCODE_MOV_DISPATCH_TO_FLAGS
);
1185 inst
= emit_linterp(*attr
, fs_reg(interp
), interpolation_mode
,
1187 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1188 inst
->predicate_inverse
= true;
1189 if (devinfo
->has_pln
)
1190 inst
->no_dd_clear
= true;
1192 inst
= emit_linterp(*attr
, fs_reg(interp
), interpolation_mode
,
1193 mod_centroid
&& !key
->persample_shading
,
1194 mod_sample
|| key
->persample_shading
);
1195 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1196 inst
->predicate_inverse
= false;
1197 if (devinfo
->has_pln
)
1198 inst
->no_dd_check
= true;
1201 emit_linterp(*attr
, fs_reg(interp
), interpolation_mode
,
1202 mod_centroid
&& !key
->persample_shading
,
1203 mod_sample
|| key
->persample_shading
);
1205 if (devinfo
->gen
< 6 && interpolation_mode
== INTERP_QUALIFIER_SMOOTH
) {
1206 bld
.MUL(*attr
, *attr
, this->pixel_w
);
1208 *attr
= offset(*attr
, bld
, 1);
1216 fs_visitor::emit_frontfacing_interpolation()
1218 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::bool_type
));
1220 if (devinfo
->gen
>= 6) {
1221 /* Bit 15 of g0.0 is 0 if the polygon is front facing. We want to create
1222 * a boolean result from this (~0/true or 0/false).
1224 * We can use the fact that bit 15 is the MSB of g0.0:W to accomplish
1225 * this task in only one instruction:
1226 * - a negation source modifier will flip the bit; and
1227 * - a W -> D type conversion will sign extend the bit into the high
1228 * word of the destination.
1230 * An ASR 15 fills the low word of the destination.
1232 fs_reg g0
= fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_W
));
1235 bld
.ASR(*reg
, g0
, brw_imm_d(15));
1237 /* Bit 31 of g1.6 is 0 if the polygon is front facing. We want to create
1238 * a boolean result from this (1/true or 0/false).
1240 * Like in the above case, since the bit is the MSB of g1.6:UD we can use
1241 * the negation source modifier to flip it. Unfortunately the SHR
1242 * instruction only operates on UD (or D with an abs source modifier)
1243 * sources without negation.
1245 * Instead, use ASR (which will give ~0/true or 0/false).
1247 fs_reg g1_6
= fs_reg(retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_D
));
1250 bld
.ASR(*reg
, g1_6
, brw_imm_d(31));
1257 fs_visitor::compute_sample_position(fs_reg dst
, fs_reg int_sample_pos
)
1259 assert(stage
== MESA_SHADER_FRAGMENT
);
1260 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1261 assert(dst
.type
== BRW_REGISTER_TYPE_F
);
1263 if (key
->compute_pos_offset
) {
1264 /* Convert int_sample_pos to floating point */
1265 bld
.MOV(dst
, int_sample_pos
);
1266 /* Scale to the range [0, 1] */
1267 bld
.MUL(dst
, dst
, brw_imm_f(1 / 16.0f
));
1270 /* From ARB_sample_shading specification:
1271 * "When rendering to a non-multisample buffer, or if multisample
1272 * rasterization is disabled, gl_SamplePosition will always be
1275 bld
.MOV(dst
, brw_imm_f(0.5f
));
1280 fs_visitor::emit_samplepos_setup()
1282 assert(devinfo
->gen
>= 6);
1284 const fs_builder abld
= bld
.annotate("compute sample position");
1285 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::vec2_type
));
1287 fs_reg int_sample_x
= vgrf(glsl_type::int_type
);
1288 fs_reg int_sample_y
= vgrf(glsl_type::int_type
);
1290 /* WM will be run in MSDISPMODE_PERSAMPLE. So, only one of SIMD8 or SIMD16
1291 * mode will be enabled.
1293 * From the Ivy Bridge PRM, volume 2 part 1, page 344:
1294 * R31.1:0 Position Offset X/Y for Slot[3:0]
1295 * R31.3:2 Position Offset X/Y for Slot[7:4]
1298 * The X, Y sample positions come in as bytes in thread payload. So, read
1299 * the positions using vstride=16, width=8, hstride=2.
1301 struct brw_reg sample_pos_reg
=
1302 stride(retype(brw_vec1_grf(payload
.sample_pos_reg
, 0),
1303 BRW_REGISTER_TYPE_B
), 16, 8, 2);
1305 if (dispatch_width
== 8) {
1306 abld
.MOV(int_sample_x
, fs_reg(sample_pos_reg
));
1308 abld
.half(0).MOV(half(int_sample_x
, 0), fs_reg(sample_pos_reg
));
1309 abld
.half(1).MOV(half(int_sample_x
, 1),
1310 fs_reg(suboffset(sample_pos_reg
, 16)));
1312 /* Compute gl_SamplePosition.x */
1313 compute_sample_position(pos
, int_sample_x
);
1314 pos
= offset(pos
, abld
, 1);
1315 if (dispatch_width
== 8) {
1316 abld
.MOV(int_sample_y
, fs_reg(suboffset(sample_pos_reg
, 1)));
1318 abld
.half(0).MOV(half(int_sample_y
, 0),
1319 fs_reg(suboffset(sample_pos_reg
, 1)));
1320 abld
.half(1).MOV(half(int_sample_y
, 1),
1321 fs_reg(suboffset(sample_pos_reg
, 17)));
1323 /* Compute gl_SamplePosition.y */
1324 compute_sample_position(pos
, int_sample_y
);
1329 fs_visitor::emit_sampleid_setup()
1331 assert(stage
== MESA_SHADER_FRAGMENT
);
1332 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1333 assert(devinfo
->gen
>= 6);
1335 const fs_builder abld
= bld
.annotate("compute sample id");
1336 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::int_type
));
1338 if (key
->compute_sample_id
) {
1339 fs_reg
t1(VGRF
, alloc
.allocate(1), BRW_REGISTER_TYPE_D
);
1341 fs_reg
t2(VGRF
, alloc
.allocate(1), BRW_REGISTER_TYPE_W
);
1343 /* The PS will be run in MSDISPMODE_PERSAMPLE. For example with
1344 * 8x multisampling, subspan 0 will represent sample N (where N
1345 * is 0, 2, 4 or 6), subspan 1 will represent sample 1, 3, 5 or
1346 * 7. We can find the value of N by looking at R0.0 bits 7:6
1347 * ("Starting Sample Pair Index (SSPI)") and multiplying by two
1348 * (since samples are always delivered in pairs). That is, we
1349 * compute 2*((R0.0 & 0xc0) >> 6) == (R0.0 & 0xc0) >> 5. Then
1350 * we need to add N to the sequence (0, 0, 0, 0, 1, 1, 1, 1) in
1351 * case of SIMD8 and sequence (0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2,
1352 * 2, 3, 3, 3, 3) in case of SIMD16. We compute this sequence by
1353 * populating a temporary variable with the sequence (0, 1, 2, 3),
1354 * and then reading from it using vstride=1, width=4, hstride=0.
1355 * These computations hold good for 4x multisampling as well.
1357 * For 2x MSAA and SIMD16, we want to use the sequence (0, 1, 0, 1):
1358 * the first four slots are sample 0 of subspan 0; the next four
1359 * are sample 1 of subspan 0; the third group is sample 0 of
1360 * subspan 1, and finally sample 1 of subspan 1.
1363 /* SKL+ has an extra bit for the Starting Sample Pair Index to
1364 * accomodate 16x MSAA.
1366 unsigned sspi_mask
= devinfo
->gen
>= 9 ? 0x1c0 : 0xc0;
1368 abld
.exec_all().group(1, 0)
1369 .AND(t1
, fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_D
)),
1370 brw_imm_ud(sspi_mask
));
1371 abld
.exec_all().group(1, 0).SHR(t1
, t1
, brw_imm_d(5));
1373 /* This works for both SIMD8 and SIMD16 */
1374 abld
.exec_all().group(4, 0)
1375 .MOV(t2
, brw_imm_v(key
->persample_2x
? 0x1010 : 0x3210));
1377 /* This special instruction takes care of setting vstride=1,
1378 * width=4, hstride=0 of t2 during an ADD instruction.
1380 abld
.emit(FS_OPCODE_SET_SAMPLE_ID
, *reg
, t1
, t2
);
1382 /* As per GL_ARB_sample_shading specification:
1383 * "When rendering to a non-multisample buffer, or if multisample
1384 * rasterization is disabled, gl_SampleID will always be zero."
1386 abld
.MOV(*reg
, brw_imm_d(0));
1393 fs_visitor::resolve_source_modifiers(const fs_reg
&src
)
1395 if (!src
.abs
&& !src
.negate
)
1398 fs_reg temp
= bld
.vgrf(src
.type
);
1405 fs_visitor::emit_discard_jump()
1407 assert(((brw_wm_prog_data
*) this->prog_data
)->uses_kill
);
1409 /* For performance, after a discard, jump to the end of the
1410 * shader if all relevant channels have been discarded.
1412 fs_inst
*discard_jump
= bld
.emit(FS_OPCODE_DISCARD_JUMP
);
1413 discard_jump
->flag_subreg
= 1;
1415 discard_jump
->predicate
= (dispatch_width
== 8)
1416 ? BRW_PREDICATE_ALIGN1_ANY8H
1417 : BRW_PREDICATE_ALIGN1_ANY16H
;
1418 discard_jump
->predicate_inverse
= true;
1422 fs_visitor::emit_gs_thread_end()
1424 assert(stage
== MESA_SHADER_GEOMETRY
);
1426 struct brw_gs_prog_data
*gs_prog_data
=
1427 (struct brw_gs_prog_data
*) prog_data
;
1429 if (gs_compile
->control_data_header_size_bits
> 0) {
1430 emit_gs_control_data_bits(this->final_gs_vertex_count
);
1433 const fs_builder abld
= bld
.annotate("thread end");
1436 if (gs_prog_data
->static_vertex_count
!= -1) {
1437 foreach_in_list_reverse(fs_inst
, prev
, &this->instructions
) {
1438 if (prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8
||
1439 prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
||
1440 prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
||
1441 prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
) {
1444 /* Delete now dead instructions. */
1445 foreach_in_list_reverse_safe(exec_node
, dead
, &this->instructions
) {
1451 } else if (prev
->is_control_flow() || prev
->has_side_effects()) {
1455 fs_reg hdr
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1456 abld
.MOV(hdr
, fs_reg(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
)));
1457 inst
= abld
.emit(SHADER_OPCODE_URB_WRITE_SIMD8
, reg_undef
, hdr
);
1460 fs_reg payload
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
1461 fs_reg
*sources
= ralloc_array(mem_ctx
, fs_reg
, 2);
1462 sources
[0] = fs_reg(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
));
1463 sources
[1] = this->final_gs_vertex_count
;
1464 abld
.LOAD_PAYLOAD(payload
, sources
, 2, 2);
1465 inst
= abld
.emit(SHADER_OPCODE_URB_WRITE_SIMD8
, reg_undef
, payload
);
1473 fs_visitor::assign_curb_setup()
1475 if (dispatch_width
== 8) {
1476 prog_data
->dispatch_grf_start_reg
= payload
.num_regs
;
1478 if (stage
== MESA_SHADER_FRAGMENT
) {
1479 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
1480 prog_data
->dispatch_grf_start_reg_16
= payload
.num_regs
;
1481 } else if (stage
== MESA_SHADER_COMPUTE
) {
1482 brw_cs_prog_data
*prog_data
= (brw_cs_prog_data
*) this->prog_data
;
1483 prog_data
->dispatch_grf_start_reg_16
= payload
.num_regs
;
1485 unreachable("Unsupported shader type!");
1489 prog_data
->curb_read_length
= ALIGN(stage_prog_data
->nr_params
, 8) / 8;
1491 /* Map the offsets in the UNIFORM file to fixed HW regs. */
1492 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1493 for (unsigned int i
= 0; i
< inst
->sources
; i
++) {
1494 if (inst
->src
[i
].file
== UNIFORM
) {
1495 int uniform_nr
= inst
->src
[i
].nr
+ inst
->src
[i
].reg_offset
;
1497 if (uniform_nr
>= 0 && uniform_nr
< (int) uniforms
) {
1498 constant_nr
= push_constant_loc
[uniform_nr
];
1500 /* Section 5.11 of the OpenGL 4.1 spec says:
1501 * "Out-of-bounds reads return undefined values, which include
1502 * values from other variables of the active program or zero."
1503 * Just return the first push constant.
1508 struct brw_reg brw_reg
= brw_vec1_grf(payload
.num_regs
+
1511 brw_reg
.abs
= inst
->src
[i
].abs
;
1512 brw_reg
.negate
= inst
->src
[i
].negate
;
1514 assert(inst
->src
[i
].stride
== 0);
1515 inst
->src
[i
] = byte_offset(
1516 retype(brw_reg
, inst
->src
[i
].type
),
1517 inst
->src
[i
].subreg_offset
);
1522 /* This may be updated in assign_urb_setup or assign_vs_urb_setup. */
1523 this->first_non_payload_grf
= payload
.num_regs
+ prog_data
->curb_read_length
;
1527 fs_visitor::calculate_urb_setup()
1529 assert(stage
== MESA_SHADER_FRAGMENT
);
1530 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
1531 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1533 memset(prog_data
->urb_setup
, -1,
1534 sizeof(prog_data
->urb_setup
[0]) * VARYING_SLOT_MAX
);
1537 /* Figure out where each of the incoming setup attributes lands. */
1538 if (devinfo
->gen
>= 6) {
1539 if (_mesa_bitcount_64(nir
->info
.inputs_read
&
1540 BRW_FS_VARYING_INPUT_MASK
) <= 16) {
1541 /* The SF/SBE pipeline stage can do arbitrary rearrangement of the
1542 * first 16 varying inputs, so we can put them wherever we want.
1543 * Just put them in order.
1545 * This is useful because it means that (a) inputs not used by the
1546 * fragment shader won't take up valuable register space, and (b) we
1547 * won't have to recompile the fragment shader if it gets paired with
1548 * a different vertex (or geometry) shader.
1550 for (unsigned int i
= 0; i
< VARYING_SLOT_MAX
; i
++) {
1551 if (nir
->info
.inputs_read
& BRW_FS_VARYING_INPUT_MASK
&
1552 BITFIELD64_BIT(i
)) {
1553 prog_data
->urb_setup
[i
] = urb_next
++;
1557 bool include_vue_header
=
1558 nir
->info
.inputs_read
& (VARYING_BIT_LAYER
| VARYING_BIT_VIEWPORT
);
1560 /* We have enough input varyings that the SF/SBE pipeline stage can't
1561 * arbitrarily rearrange them to suit our whim; we have to put them
1562 * in an order that matches the output of the previous pipeline stage
1563 * (geometry or vertex shader).
1565 struct brw_vue_map prev_stage_vue_map
;
1566 brw_compute_vue_map(devinfo
, &prev_stage_vue_map
,
1567 key
->input_slots_valid
,
1568 nir
->info
.separate_shader
);
1570 include_vue_header
? 0 : 2 * BRW_SF_URB_ENTRY_READ_OFFSET
;
1572 assert(prev_stage_vue_map
.num_slots
<= first_slot
+ 32);
1573 for (int slot
= first_slot
; slot
< prev_stage_vue_map
.num_slots
;
1575 int varying
= prev_stage_vue_map
.slot_to_varying
[slot
];
1576 if (varying
!= BRW_VARYING_SLOT_PAD
&&
1577 (nir
->info
.inputs_read
& BRW_FS_VARYING_INPUT_MASK
&
1578 BITFIELD64_BIT(varying
))) {
1579 prog_data
->urb_setup
[varying
] = slot
- first_slot
;
1582 urb_next
= prev_stage_vue_map
.num_slots
- first_slot
;
1585 /* FINISHME: The sf doesn't map VS->FS inputs for us very well. */
1586 for (unsigned int i
= 0; i
< VARYING_SLOT_MAX
; i
++) {
1587 /* Point size is packed into the header, not as a general attribute */
1588 if (i
== VARYING_SLOT_PSIZ
)
1591 if (key
->input_slots_valid
& BITFIELD64_BIT(i
)) {
1592 /* The back color slot is skipped when the front color is
1593 * also written to. In addition, some slots can be
1594 * written in the vertex shader and not read in the
1595 * fragment shader. So the register number must always be
1596 * incremented, mapped or not.
1598 if (_mesa_varying_slot_in_fs((gl_varying_slot
) i
))
1599 prog_data
->urb_setup
[i
] = urb_next
;
1605 * It's a FS only attribute, and we did interpolation for this attribute
1606 * in SF thread. So, count it here, too.
1608 * See compile_sf_prog() for more info.
1610 if (nir
->info
.inputs_read
& BITFIELD64_BIT(VARYING_SLOT_PNTC
))
1611 prog_data
->urb_setup
[VARYING_SLOT_PNTC
] = urb_next
++;
1614 prog_data
->num_varying_inputs
= urb_next
;
1618 fs_visitor::assign_urb_setup()
1620 assert(stage
== MESA_SHADER_FRAGMENT
);
1621 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
1623 int urb_start
= payload
.num_regs
+ prog_data
->base
.curb_read_length
;
1625 /* Offset all the urb_setup[] index by the actual position of the
1626 * setup regs, now that the location of the constants has been chosen.
1628 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1629 if (inst
->opcode
== FS_OPCODE_LINTERP
) {
1630 assert(inst
->src
[1].file
== FIXED_GRF
);
1631 inst
->src
[1].nr
+= urb_start
;
1634 if (inst
->opcode
== FS_OPCODE_CINTERP
) {
1635 assert(inst
->src
[0].file
== FIXED_GRF
);
1636 inst
->src
[0].nr
+= urb_start
;
1640 /* Each attribute is 4 setup channels, each of which is half a reg. */
1641 this->first_non_payload_grf
+= prog_data
->num_varying_inputs
* 2;
1645 fs_visitor::convert_attr_sources_to_hw_regs(fs_inst
*inst
)
1647 for (int i
= 0; i
< inst
->sources
; i
++) {
1648 if (inst
->src
[i
].file
== ATTR
) {
1649 int grf
= payload
.num_regs
+
1650 prog_data
->curb_read_length
+
1652 inst
->src
[i
].reg_offset
;
1654 unsigned width
= inst
->src
[i
].stride
== 0 ? 1 : inst
->exec_size
;
1655 struct brw_reg reg
=
1656 stride(byte_offset(retype(brw_vec8_grf(grf
, 0), inst
->src
[i
].type
),
1657 inst
->src
[i
].subreg_offset
),
1658 inst
->exec_size
* inst
->src
[i
].stride
,
1659 width
, inst
->src
[i
].stride
);
1660 reg
.abs
= inst
->src
[i
].abs
;
1661 reg
.negate
= inst
->src
[i
].negate
;
1669 fs_visitor::assign_vs_urb_setup()
1671 brw_vs_prog_data
*vs_prog_data
= (brw_vs_prog_data
*) prog_data
;
1673 assert(stage
== MESA_SHADER_VERTEX
);
1674 int count
= _mesa_bitcount_64(vs_prog_data
->inputs_read
);
1675 if (vs_prog_data
->uses_vertexid
|| vs_prog_data
->uses_instanceid
)
1678 /* Each attribute is 4 regs. */
1679 this->first_non_payload_grf
+= 4 * vs_prog_data
->nr_attributes
;
1681 assert(vs_prog_data
->base
.urb_read_length
<= 15);
1683 /* Rewrite all ATTR file references to the hw grf that they land in. */
1684 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1685 convert_attr_sources_to_hw_regs(inst
);
1690 fs_visitor::assign_tes_urb_setup()
1692 assert(stage
== MESA_SHADER_TESS_EVAL
);
1694 brw_vue_prog_data
*vue_prog_data
= (brw_vue_prog_data
*) prog_data
;
1696 first_non_payload_grf
+= 8 * vue_prog_data
->urb_read_length
;
1698 /* Rewrite all ATTR file references to HW_REGs. */
1699 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1700 convert_attr_sources_to_hw_regs(inst
);
1705 fs_visitor::assign_gs_urb_setup()
1707 assert(stage
== MESA_SHADER_GEOMETRY
);
1709 brw_vue_prog_data
*vue_prog_data
= (brw_vue_prog_data
*) prog_data
;
1711 first_non_payload_grf
+=
1712 8 * vue_prog_data
->urb_read_length
* nir
->info
.gs
.vertices_in
;
1714 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1715 /* Rewrite all ATTR file references to GRFs. */
1716 convert_attr_sources_to_hw_regs(inst
);
1722 * Split large virtual GRFs into separate components if we can.
1724 * This is mostly duplicated with what brw_fs_vector_splitting does,
1725 * but that's really conservative because it's afraid of doing
1726 * splitting that doesn't result in real progress after the rest of
1727 * the optimization phases, which would cause infinite looping in
1728 * optimization. We can do it once here, safely. This also has the
1729 * opportunity to split interpolated values, or maybe even uniforms,
1730 * which we don't have at the IR level.
1732 * We want to split, because virtual GRFs are what we register
1733 * allocate and spill (due to contiguousness requirements for some
1734 * instructions), and they're what we naturally generate in the
1735 * codegen process, but most virtual GRFs don't actually need to be
1736 * contiguous sets of GRFs. If we split, we'll end up with reduced
1737 * live intervals and better dead code elimination and coalescing.
1740 fs_visitor::split_virtual_grfs()
1742 int num_vars
= this->alloc
.count
;
1744 /* Count the total number of registers */
1746 int vgrf_to_reg
[num_vars
];
1747 for (int i
= 0; i
< num_vars
; i
++) {
1748 vgrf_to_reg
[i
] = reg_count
;
1749 reg_count
+= alloc
.sizes
[i
];
1752 /* An array of "split points". For each register slot, this indicates
1753 * if this slot can be separated from the previous slot. Every time an
1754 * instruction uses multiple elements of a register (as a source or
1755 * destination), we mark the used slots as inseparable. Then we go
1756 * through and split the registers into the smallest pieces we can.
1758 bool split_points
[reg_count
];
1759 memset(split_points
, 0, sizeof(split_points
));
1761 /* Mark all used registers as fully splittable */
1762 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1763 if (inst
->dst
.file
== VGRF
) {
1764 int reg
= vgrf_to_reg
[inst
->dst
.nr
];
1765 for (unsigned j
= 1; j
< this->alloc
.sizes
[inst
->dst
.nr
]; j
++)
1766 split_points
[reg
+ j
] = true;
1769 for (int i
= 0; i
< inst
->sources
; i
++) {
1770 if (inst
->src
[i
].file
== VGRF
) {
1771 int reg
= vgrf_to_reg
[inst
->src
[i
].nr
];
1772 for (unsigned j
= 1; j
< this->alloc
.sizes
[inst
->src
[i
].nr
]; j
++)
1773 split_points
[reg
+ j
] = true;
1778 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1779 if (inst
->dst
.file
== VGRF
) {
1780 int reg
= vgrf_to_reg
[inst
->dst
.nr
] + inst
->dst
.reg_offset
;
1781 for (int j
= 1; j
< inst
->regs_written
; j
++)
1782 split_points
[reg
+ j
] = false;
1784 for (int i
= 0; i
< inst
->sources
; i
++) {
1785 if (inst
->src
[i
].file
== VGRF
) {
1786 int reg
= vgrf_to_reg
[inst
->src
[i
].nr
] + inst
->src
[i
].reg_offset
;
1787 for (int j
= 1; j
< inst
->regs_read(i
); j
++)
1788 split_points
[reg
+ j
] = false;
1793 int new_virtual_grf
[reg_count
];
1794 int new_reg_offset
[reg_count
];
1797 for (int i
= 0; i
< num_vars
; i
++) {
1798 /* The first one should always be 0 as a quick sanity check. */
1799 assert(split_points
[reg
] == false);
1802 new_reg_offset
[reg
] = 0;
1807 for (unsigned j
= 1; j
< alloc
.sizes
[i
]; j
++) {
1808 /* If this is a split point, reset the offset to 0 and allocate a
1809 * new virtual GRF for the previous offset many registers
1811 if (split_points
[reg
]) {
1812 assert(offset
<= MAX_VGRF_SIZE
);
1813 int grf
= alloc
.allocate(offset
);
1814 for (int k
= reg
- offset
; k
< reg
; k
++)
1815 new_virtual_grf
[k
] = grf
;
1818 new_reg_offset
[reg
] = offset
;
1823 /* The last one gets the original register number */
1824 assert(offset
<= MAX_VGRF_SIZE
);
1825 alloc
.sizes
[i
] = offset
;
1826 for (int k
= reg
- offset
; k
< reg
; k
++)
1827 new_virtual_grf
[k
] = i
;
1829 assert(reg
== reg_count
);
1831 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1832 if (inst
->dst
.file
== VGRF
) {
1833 reg
= vgrf_to_reg
[inst
->dst
.nr
] + inst
->dst
.reg_offset
;
1834 inst
->dst
.nr
= new_virtual_grf
[reg
];
1835 inst
->dst
.reg_offset
= new_reg_offset
[reg
];
1836 assert((unsigned)new_reg_offset
[reg
] < alloc
.sizes
[new_virtual_grf
[reg
]]);
1838 for (int i
= 0; i
< inst
->sources
; i
++) {
1839 if (inst
->src
[i
].file
== VGRF
) {
1840 reg
= vgrf_to_reg
[inst
->src
[i
].nr
] + inst
->src
[i
].reg_offset
;
1841 inst
->src
[i
].nr
= new_virtual_grf
[reg
];
1842 inst
->src
[i
].reg_offset
= new_reg_offset
[reg
];
1843 assert((unsigned)new_reg_offset
[reg
] < alloc
.sizes
[new_virtual_grf
[reg
]]);
1847 invalidate_live_intervals();
1851 * Remove unused virtual GRFs and compact the virtual_grf_* arrays.
1853 * During code generation, we create tons of temporary variables, many of
1854 * which get immediately killed and are never used again. Yet, in later
1855 * optimization and analysis passes, such as compute_live_intervals, we need
1856 * to loop over all the virtual GRFs. Compacting them can save a lot of
1860 fs_visitor::compact_virtual_grfs()
1862 bool progress
= false;
1863 int remap_table
[this->alloc
.count
];
1864 memset(remap_table
, -1, sizeof(remap_table
));
1866 /* Mark which virtual GRFs are used. */
1867 foreach_block_and_inst(block
, const fs_inst
, inst
, cfg
) {
1868 if (inst
->dst
.file
== VGRF
)
1869 remap_table
[inst
->dst
.nr
] = 0;
1871 for (int i
= 0; i
< inst
->sources
; i
++) {
1872 if (inst
->src
[i
].file
== VGRF
)
1873 remap_table
[inst
->src
[i
].nr
] = 0;
1877 /* Compact the GRF arrays. */
1879 for (unsigned i
= 0; i
< this->alloc
.count
; i
++) {
1880 if (remap_table
[i
] == -1) {
1881 /* We just found an unused register. This means that we are
1882 * actually going to compact something.
1886 remap_table
[i
] = new_index
;
1887 alloc
.sizes
[new_index
] = alloc
.sizes
[i
];
1888 invalidate_live_intervals();
1893 this->alloc
.count
= new_index
;
1895 /* Patch all the instructions to use the newly renumbered registers */
1896 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1897 if (inst
->dst
.file
== VGRF
)
1898 inst
->dst
.nr
= remap_table
[inst
->dst
.nr
];
1900 for (int i
= 0; i
< inst
->sources
; i
++) {
1901 if (inst
->src
[i
].file
== VGRF
)
1902 inst
->src
[i
].nr
= remap_table
[inst
->src
[i
].nr
];
1906 /* Patch all the references to delta_xy, since they're used in register
1907 * allocation. If they're unused, switch them to BAD_FILE so we don't
1908 * think some random VGRF is delta_xy.
1910 for (unsigned i
= 0; i
< ARRAY_SIZE(delta_xy
); i
++) {
1911 if (delta_xy
[i
].file
== VGRF
) {
1912 if (remap_table
[delta_xy
[i
].nr
] != -1) {
1913 delta_xy
[i
].nr
= remap_table
[delta_xy
[i
].nr
];
1915 delta_xy
[i
].file
= BAD_FILE
;
1924 * Assign UNIFORM file registers to either push constants or pull constants.
1926 * We allow a fragment shader to have more than the specified minimum
1927 * maximum number of fragment shader uniform components (64). If
1928 * there are too many of these, they'd fill up all of register space.
1929 * So, this will push some of them out to the pull constant buffer and
1930 * update the program to load them. We also use pull constants for all
1931 * indirect constant loads because we don't support indirect accesses in
1935 fs_visitor::assign_constant_locations()
1937 /* Only the first compile (SIMD8 mode) gets to decide on locations. */
1938 if (dispatch_width
!= 8)
1941 unsigned int num_pull_constants
= 0;
1943 pull_constant_loc
= ralloc_array(mem_ctx
, int, uniforms
);
1944 memset(pull_constant_loc
, -1, sizeof(pull_constant_loc
[0]) * uniforms
);
1946 bool is_live
[uniforms
];
1947 memset(is_live
, 0, sizeof(is_live
));
1949 /* First, we walk through the instructions and do two things:
1951 * 1) Figure out which uniforms are live.
1953 * 2) Find all indirect access of uniform arrays and flag them as needing
1954 * to go into the pull constant buffer.
1956 * Note that we don't move constant-indexed accesses to arrays. No
1957 * testing has been done of the performance impact of this choice.
1959 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
1960 for (int i
= 0 ; i
< inst
->sources
; i
++) {
1961 if (inst
->src
[i
].file
!= UNIFORM
)
1964 if (inst
->src
[i
].reladdr
) {
1965 int uniform
= inst
->src
[i
].nr
;
1967 /* If this array isn't already present in the pull constant buffer,
1970 if (pull_constant_loc
[uniform
] == -1) {
1971 assert(param_size
[uniform
]);
1972 for (int j
= 0; j
< param_size
[uniform
]; j
++)
1973 pull_constant_loc
[uniform
+ j
] = num_pull_constants
++;
1976 /* Mark the the one accessed uniform as live */
1977 int constant_nr
= inst
->src
[i
].nr
+ inst
->src
[i
].reg_offset
;
1978 if (constant_nr
>= 0 && constant_nr
< (int) uniforms
)
1979 is_live
[constant_nr
] = true;
1984 /* Only allow 16 registers (128 uniform components) as push constants.
1986 * Just demote the end of the list. We could probably do better
1987 * here, demoting things that are rarely used in the program first.
1989 * If changing this value, note the limitation about total_regs in
1992 unsigned int max_push_components
= 16 * 8;
1993 unsigned int num_push_constants
= 0;
1995 push_constant_loc
= ralloc_array(mem_ctx
, int, uniforms
);
1997 for (unsigned int i
= 0; i
< uniforms
; i
++) {
1998 if (!is_live
[i
] || pull_constant_loc
[i
] != -1) {
1999 /* This UNIFORM register is either dead, or has already been demoted
2000 * to a pull const. Mark it as no longer living in the param[] array.
2002 push_constant_loc
[i
] = -1;
2006 if (num_push_constants
< max_push_components
) {
2007 /* Retain as a push constant. Record the location in the params[]
2010 push_constant_loc
[i
] = num_push_constants
++;
2012 /* Demote to a pull constant. */
2013 push_constant_loc
[i
] = -1;
2014 pull_constant_loc
[i
] = num_pull_constants
++;
2018 stage_prog_data
->nr_params
= num_push_constants
;
2019 stage_prog_data
->nr_pull_params
= num_pull_constants
;
2021 /* Up until now, the param[] array has been indexed by reg + reg_offset
2022 * of UNIFORM registers. Move pull constants into pull_param[] and
2023 * condense param[] to only contain the uniforms we chose to push.
2025 * NOTE: Because we are condensing the params[] array, we know that
2026 * push_constant_loc[i] <= i and we can do it in one smooth loop without
2027 * having to make a copy.
2029 for (unsigned int i
= 0; i
< uniforms
; i
++) {
2030 const gl_constant_value
*value
= stage_prog_data
->param
[i
];
2032 if (pull_constant_loc
[i
] != -1) {
2033 stage_prog_data
->pull_param
[pull_constant_loc
[i
]] = value
;
2034 } else if (push_constant_loc
[i
] != -1) {
2035 stage_prog_data
->param
[push_constant_loc
[i
]] = value
;
2041 * Replace UNIFORM register file access with either UNIFORM_PULL_CONSTANT_LOAD
2042 * or VARYING_PULL_CONSTANT_LOAD instructions which load values into VGRFs.
2045 fs_visitor::demote_pull_constants()
2047 foreach_block_and_inst (block
, fs_inst
, inst
, cfg
) {
2048 for (int i
= 0; i
< inst
->sources
; i
++) {
2049 if (inst
->src
[i
].file
!= UNIFORM
)
2053 unsigned location
= inst
->src
[i
].nr
+ inst
->src
[i
].reg_offset
;
2054 if (location
>= uniforms
) /* Out of bounds access */
2057 pull_index
= pull_constant_loc
[location
];
2059 if (pull_index
== -1)
2062 /* Set up the annotation tracking for new generated instructions. */
2063 const fs_builder
ibld(this, block
, inst
);
2064 const unsigned index
= stage_prog_data
->binding_table
.pull_constants_start
;
2065 fs_reg dst
= vgrf(glsl_type::float_type
);
2067 assert(inst
->src
[i
].stride
== 0);
2069 /* Generate a pull load into dst. */
2070 if (inst
->src
[i
].reladdr
) {
2071 VARYING_PULL_CONSTANT_LOAD(ibld
, dst
,
2073 *inst
->src
[i
].reladdr
,
2075 inst
->src
[i
].reladdr
= NULL
;
2076 inst
->src
[i
].stride
= 1;
2078 const fs_builder ubld
= ibld
.exec_all().group(8, 0);
2079 struct brw_reg offset
= brw_imm_ud((unsigned)(pull_index
* 4) & ~15);
2080 ubld
.emit(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
,
2081 dst
, brw_imm_ud(index
), offset
);
2082 inst
->src
[i
].set_smear(pull_index
& 3);
2084 brw_mark_surface_used(prog_data
, index
);
2086 /* Rewrite the instruction to use the temporary VGRF. */
2087 inst
->src
[i
].file
= VGRF
;
2088 inst
->src
[i
].nr
= dst
.nr
;
2089 inst
->src
[i
].reg_offset
= 0;
2092 invalidate_live_intervals();
2096 fs_visitor::opt_algebraic()
2098 bool progress
= false;
2100 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2101 switch (inst
->opcode
) {
2102 case BRW_OPCODE_MOV
:
2103 if (inst
->src
[0].file
!= IMM
)
2106 if (inst
->saturate
) {
2107 if (inst
->dst
.type
!= inst
->src
[0].type
)
2108 assert(!"unimplemented: saturate mixed types");
2110 if (brw_saturate_immediate(inst
->dst
.type
,
2111 &inst
->src
[0].as_brw_reg())) {
2112 inst
->saturate
= false;
2118 case BRW_OPCODE_MUL
:
2119 if (inst
->src
[1].file
!= IMM
)
2123 if (inst
->src
[1].is_one()) {
2124 inst
->opcode
= BRW_OPCODE_MOV
;
2125 inst
->src
[1] = reg_undef
;
2131 if (inst
->src
[1].is_negative_one()) {
2132 inst
->opcode
= BRW_OPCODE_MOV
;
2133 inst
->src
[0].negate
= !inst
->src
[0].negate
;
2134 inst
->src
[1] = reg_undef
;
2140 if (inst
->src
[1].is_zero()) {
2141 inst
->opcode
= BRW_OPCODE_MOV
;
2142 inst
->src
[0] = inst
->src
[1];
2143 inst
->src
[1] = reg_undef
;
2148 if (inst
->src
[0].file
== IMM
) {
2149 assert(inst
->src
[0].type
== BRW_REGISTER_TYPE_F
);
2150 inst
->opcode
= BRW_OPCODE_MOV
;
2151 inst
->src
[0].f
*= inst
->src
[1].f
;
2152 inst
->src
[1] = reg_undef
;
2157 case BRW_OPCODE_ADD
:
2158 if (inst
->src
[1].file
!= IMM
)
2162 if (inst
->src
[1].is_zero()) {
2163 inst
->opcode
= BRW_OPCODE_MOV
;
2164 inst
->src
[1] = reg_undef
;
2169 if (inst
->src
[0].file
== IMM
) {
2170 assert(inst
->src
[0].type
== BRW_REGISTER_TYPE_F
);
2171 inst
->opcode
= BRW_OPCODE_MOV
;
2172 inst
->src
[0].f
+= inst
->src
[1].f
;
2173 inst
->src
[1] = reg_undef
;
2179 if (inst
->src
[0].equals(inst
->src
[1])) {
2180 inst
->opcode
= BRW_OPCODE_MOV
;
2181 inst
->src
[1] = reg_undef
;
2186 case BRW_OPCODE_LRP
:
2187 if (inst
->src
[1].equals(inst
->src
[2])) {
2188 inst
->opcode
= BRW_OPCODE_MOV
;
2189 inst
->src
[0] = inst
->src
[1];
2190 inst
->src
[1] = reg_undef
;
2191 inst
->src
[2] = reg_undef
;
2196 case BRW_OPCODE_CMP
:
2197 if (inst
->conditional_mod
== BRW_CONDITIONAL_GE
&&
2199 inst
->src
[0].negate
&&
2200 inst
->src
[1].is_zero()) {
2201 inst
->src
[0].abs
= false;
2202 inst
->src
[0].negate
= false;
2203 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
2208 case BRW_OPCODE_SEL
:
2209 if (inst
->src
[0].equals(inst
->src
[1])) {
2210 inst
->opcode
= BRW_OPCODE_MOV
;
2211 inst
->src
[1] = reg_undef
;
2212 inst
->predicate
= BRW_PREDICATE_NONE
;
2213 inst
->predicate_inverse
= false;
2215 } else if (inst
->saturate
&& inst
->src
[1].file
== IMM
) {
2216 switch (inst
->conditional_mod
) {
2217 case BRW_CONDITIONAL_LE
:
2218 case BRW_CONDITIONAL_L
:
2219 switch (inst
->src
[1].type
) {
2220 case BRW_REGISTER_TYPE_F
:
2221 if (inst
->src
[1].f
>= 1.0f
) {
2222 inst
->opcode
= BRW_OPCODE_MOV
;
2223 inst
->src
[1] = reg_undef
;
2224 inst
->conditional_mod
= BRW_CONDITIONAL_NONE
;
2232 case BRW_CONDITIONAL_GE
:
2233 case BRW_CONDITIONAL_G
:
2234 switch (inst
->src
[1].type
) {
2235 case BRW_REGISTER_TYPE_F
:
2236 if (inst
->src
[1].f
<= 0.0f
) {
2237 inst
->opcode
= BRW_OPCODE_MOV
;
2238 inst
->src
[1] = reg_undef
;
2239 inst
->conditional_mod
= BRW_CONDITIONAL_NONE
;
2251 case BRW_OPCODE_MAD
:
2252 if (inst
->src
[1].is_zero() || inst
->src
[2].is_zero()) {
2253 inst
->opcode
= BRW_OPCODE_MOV
;
2254 inst
->src
[1] = reg_undef
;
2255 inst
->src
[2] = reg_undef
;
2257 } else if (inst
->src
[0].is_zero()) {
2258 inst
->opcode
= BRW_OPCODE_MUL
;
2259 inst
->src
[0] = inst
->src
[2];
2260 inst
->src
[2] = reg_undef
;
2262 } else if (inst
->src
[1].is_one()) {
2263 inst
->opcode
= BRW_OPCODE_ADD
;
2264 inst
->src
[1] = inst
->src
[2];
2265 inst
->src
[2] = reg_undef
;
2267 } else if (inst
->src
[2].is_one()) {
2268 inst
->opcode
= BRW_OPCODE_ADD
;
2269 inst
->src
[2] = reg_undef
;
2271 } else if (inst
->src
[1].file
== IMM
&& inst
->src
[2].file
== IMM
) {
2272 inst
->opcode
= BRW_OPCODE_ADD
;
2273 inst
->src
[1].f
*= inst
->src
[2].f
;
2274 inst
->src
[2] = reg_undef
;
2278 case SHADER_OPCODE_RCP
: {
2279 fs_inst
*prev
= (fs_inst
*)inst
->prev
;
2280 if (prev
->opcode
== SHADER_OPCODE_SQRT
) {
2281 if (inst
->src
[0].equals(prev
->dst
)) {
2282 inst
->opcode
= SHADER_OPCODE_RSQ
;
2283 inst
->src
[0] = prev
->src
[0];
2289 case SHADER_OPCODE_BROADCAST
:
2290 if (is_uniform(inst
->src
[0])) {
2291 inst
->opcode
= BRW_OPCODE_MOV
;
2293 inst
->force_writemask_all
= true;
2295 } else if (inst
->src
[1].file
== IMM
) {
2296 inst
->opcode
= BRW_OPCODE_MOV
;
2297 inst
->src
[0] = component(inst
->src
[0],
2300 inst
->force_writemask_all
= true;
2309 /* Swap if src[0] is immediate. */
2310 if (progress
&& inst
->is_commutative()) {
2311 if (inst
->src
[0].file
== IMM
) {
2312 fs_reg tmp
= inst
->src
[1];
2313 inst
->src
[1] = inst
->src
[0];
2322 * Optimize sample messages that have constant zero values for the trailing
2323 * texture coordinates. We can just reduce the message length for these
2324 * instructions instead of reserving a register for it. Trailing parameters
2325 * that aren't sent default to zero anyway. This will cause the dead code
2326 * eliminator to remove the MOV instruction that would otherwise be emitted to
2327 * set up the zero value.
2330 fs_visitor::opt_zero_samples()
2332 /* Gen4 infers the texturing opcode based on the message length so we can't
2335 if (devinfo
->gen
< 5)
2338 bool progress
= false;
2340 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2341 if (!inst
->is_tex())
2344 fs_inst
*load_payload
= (fs_inst
*) inst
->prev
;
2346 if (load_payload
->is_head_sentinel() ||
2347 load_payload
->opcode
!= SHADER_OPCODE_LOAD_PAYLOAD
)
2350 /* We don't want to remove the message header or the first parameter.
2351 * Removing the first parameter is not allowed, see the Haswell PRM
2352 * volume 7, page 149:
2354 * "Parameter 0 is required except for the sampleinfo message, which
2355 * has no parameter 0"
2357 while (inst
->mlen
> inst
->header_size
+ inst
->exec_size
/ 8 &&
2358 load_payload
->src
[(inst
->mlen
- inst
->header_size
) /
2359 (inst
->exec_size
/ 8) +
2360 inst
->header_size
- 1].is_zero()) {
2361 inst
->mlen
-= inst
->exec_size
/ 8;
2367 invalidate_live_intervals();
2373 * Optimize sample messages which are followed by the final RT write.
2375 * CHV, and GEN9+ can mark a texturing SEND instruction with EOT to have its
2376 * results sent directly to the framebuffer, bypassing the EU. Recognize the
2377 * final texturing results copied to the framebuffer write payload and modify
2378 * them to write to the framebuffer directly.
2381 fs_visitor::opt_sampler_eot()
2383 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
2385 if (stage
!= MESA_SHADER_FRAGMENT
)
2388 if (devinfo
->gen
< 9 && !devinfo
->is_cherryview
)
2391 /* FINISHME: It should be possible to implement this optimization when there
2392 * are multiple drawbuffers.
2394 if (key
->nr_color_regions
!= 1)
2397 /* Look for a texturing instruction immediately before the final FB_WRITE. */
2398 bblock_t
*block
= cfg
->blocks
[cfg
->num_blocks
- 1];
2399 fs_inst
*fb_write
= (fs_inst
*)block
->end();
2400 assert(fb_write
->eot
);
2401 assert(fb_write
->opcode
== FS_OPCODE_FB_WRITE
);
2403 fs_inst
*tex_inst
= (fs_inst
*) fb_write
->prev
;
2405 /* There wasn't one; nothing to do. */
2406 if (unlikely(tex_inst
->is_head_sentinel()) || !tex_inst
->is_tex())
2409 /* 3D Sampler » Messages » Message Format
2411 * “Response Length of zero is allowed on all SIMD8* and SIMD16* sampler
2412 * messages except sample+killpix, resinfo, sampleinfo, LOD, and gather4*”
2414 if (tex_inst
->opcode
== SHADER_OPCODE_TXS
||
2415 tex_inst
->opcode
== SHADER_OPCODE_SAMPLEINFO
||
2416 tex_inst
->opcode
== SHADER_OPCODE_LOD
||
2417 tex_inst
->opcode
== SHADER_OPCODE_TG4
||
2418 tex_inst
->opcode
== SHADER_OPCODE_TG4_OFFSET
)
2421 /* If there's no header present, we need to munge the LOAD_PAYLOAD as well.
2422 * It's very likely to be the previous instruction.
2424 fs_inst
*load_payload
= (fs_inst
*) tex_inst
->prev
;
2425 if (load_payload
->is_head_sentinel() ||
2426 load_payload
->opcode
!= SHADER_OPCODE_LOAD_PAYLOAD
)
2429 assert(!tex_inst
->eot
); /* We can't get here twice */
2430 assert((tex_inst
->offset
& (0xff << 24)) == 0);
2432 const fs_builder
ibld(this, block
, tex_inst
);
2434 tex_inst
->offset
|= fb_write
->target
<< 24;
2435 tex_inst
->eot
= true;
2436 tex_inst
->dst
= ibld
.null_reg_ud();
2437 fb_write
->remove(cfg
->blocks
[cfg
->num_blocks
- 1]);
2439 /* If a header is present, marking the eot is sufficient. Otherwise, we need
2440 * to create a new LOAD_PAYLOAD command with the same sources and a space
2441 * saved for the header. Using a new destination register not only makes sure
2442 * we have enough space, but it will make sure the dead code eliminator kills
2443 * the instruction that this will replace.
2445 if (tex_inst
->header_size
!= 0)
2448 fs_reg send_header
= ibld
.vgrf(BRW_REGISTER_TYPE_F
,
2449 load_payload
->sources
+ 1);
2450 fs_reg
*new_sources
=
2451 ralloc_array(mem_ctx
, fs_reg
, load_payload
->sources
+ 1);
2453 new_sources
[0] = fs_reg();
2454 for (int i
= 0; i
< load_payload
->sources
; i
++)
2455 new_sources
[i
+1] = load_payload
->src
[i
];
2457 /* The LOAD_PAYLOAD helper seems like the obvious choice here. However, it
2458 * requires a lot of information about the sources to appropriately figure
2459 * out the number of registers needed to be used. Given this stage in our
2460 * optimization, we may not have the appropriate GRFs required by
2461 * LOAD_PAYLOAD at this point (copy propagation). Therefore, we need to
2462 * manually emit the instruction.
2464 fs_inst
*new_load_payload
= new(mem_ctx
) fs_inst(SHADER_OPCODE_LOAD_PAYLOAD
,
2465 load_payload
->exec_size
,
2468 load_payload
->sources
+ 1);
2470 new_load_payload
->regs_written
= load_payload
->regs_written
+ 1;
2471 new_load_payload
->header_size
= 1;
2473 tex_inst
->header_size
= 1;
2474 tex_inst
->insert_before(cfg
->blocks
[cfg
->num_blocks
- 1], new_load_payload
);
2475 tex_inst
->src
[0] = send_header
;
2481 fs_visitor::opt_register_renaming()
2483 bool progress
= false;
2486 int remap
[alloc
.count
];
2487 memset(remap
, -1, sizeof(int) * alloc
.count
);
2489 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2490 if (inst
->opcode
== BRW_OPCODE_IF
|| inst
->opcode
== BRW_OPCODE_DO
) {
2492 } else if (inst
->opcode
== BRW_OPCODE_ENDIF
||
2493 inst
->opcode
== BRW_OPCODE_WHILE
) {
2497 /* Rewrite instruction sources. */
2498 for (int i
= 0; i
< inst
->sources
; i
++) {
2499 if (inst
->src
[i
].file
== VGRF
&&
2500 remap
[inst
->src
[i
].nr
] != -1 &&
2501 remap
[inst
->src
[i
].nr
] != inst
->src
[i
].nr
) {
2502 inst
->src
[i
].nr
= remap
[inst
->src
[i
].nr
];
2507 const int dst
= inst
->dst
.nr
;
2510 inst
->dst
.file
== VGRF
&&
2511 alloc
.sizes
[inst
->dst
.nr
] == inst
->exec_size
/ 8 &&
2512 !inst
->is_partial_write()) {
2513 if (remap
[dst
] == -1) {
2516 remap
[dst
] = alloc
.allocate(inst
->exec_size
/ 8);
2517 inst
->dst
.nr
= remap
[dst
];
2520 } else if (inst
->dst
.file
== VGRF
&&
2522 remap
[dst
] != dst
) {
2523 inst
->dst
.nr
= remap
[dst
];
2529 invalidate_live_intervals();
2531 for (unsigned i
= 0; i
< ARRAY_SIZE(delta_xy
); i
++) {
2532 if (delta_xy
[i
].file
== VGRF
&& remap
[delta_xy
[i
].nr
] != -1) {
2533 delta_xy
[i
].nr
= remap
[delta_xy
[i
].nr
];
2542 * Remove redundant or useless discard jumps.
2544 * For example, we can eliminate jumps in the following sequence:
2546 * discard-jump (redundant with the next jump)
2547 * discard-jump (useless; jumps to the next instruction)
2551 fs_visitor::opt_redundant_discard_jumps()
2553 bool progress
= false;
2555 bblock_t
*last_bblock
= cfg
->blocks
[cfg
->num_blocks
- 1];
2557 fs_inst
*placeholder_halt
= NULL
;
2558 foreach_inst_in_block_reverse(fs_inst
, inst
, last_bblock
) {
2559 if (inst
->opcode
== FS_OPCODE_PLACEHOLDER_HALT
) {
2560 placeholder_halt
= inst
;
2565 if (!placeholder_halt
)
2568 /* Delete any HALTs immediately before the placeholder halt. */
2569 for (fs_inst
*prev
= (fs_inst
*) placeholder_halt
->prev
;
2570 !prev
->is_head_sentinel() && prev
->opcode
== FS_OPCODE_DISCARD_JUMP
;
2571 prev
= (fs_inst
*) placeholder_halt
->prev
) {
2572 prev
->remove(last_bblock
);
2577 invalidate_live_intervals();
2583 fs_visitor::compute_to_mrf()
2585 bool progress
= false;
2588 /* No MRFs on Gen >= 7. */
2589 if (devinfo
->gen
>= 7)
2592 calculate_live_intervals();
2594 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
2598 if (inst
->opcode
!= BRW_OPCODE_MOV
||
2599 inst
->is_partial_write() ||
2600 inst
->dst
.file
!= MRF
|| inst
->src
[0].file
!= VGRF
||
2601 inst
->dst
.type
!= inst
->src
[0].type
||
2602 inst
->src
[0].abs
|| inst
->src
[0].negate
||
2603 !inst
->src
[0].is_contiguous() ||
2604 inst
->src
[0].subreg_offset
)
2607 /* Work out which hardware MRF registers are written by this
2610 int mrf_low
= inst
->dst
.nr
& ~BRW_MRF_COMPR4
;
2612 if (inst
->dst
.nr
& BRW_MRF_COMPR4
) {
2613 mrf_high
= mrf_low
+ 4;
2614 } else if (inst
->exec_size
== 16) {
2615 mrf_high
= mrf_low
+ 1;
2620 /* Can't compute-to-MRF this GRF if someone else was going to
2623 if (this->virtual_grf_end
[inst
->src
[0].nr
] > ip
)
2626 /* Found a move of a GRF to a MRF. Let's see if we can go
2627 * rewrite the thing that made this GRF to write into the MRF.
2629 foreach_inst_in_block_reverse_starting_from(fs_inst
, scan_inst
, inst
) {
2630 if (scan_inst
->dst
.file
== VGRF
&&
2631 scan_inst
->dst
.nr
== inst
->src
[0].nr
) {
2632 /* Found the last thing to write our reg we want to turn
2633 * into a compute-to-MRF.
2636 /* If this one instruction didn't populate all the
2637 * channels, bail. We might be able to rewrite everything
2638 * that writes that reg, but it would require smarter
2639 * tracking to delay the rewriting until complete success.
2641 if (scan_inst
->is_partial_write())
2644 /* Things returning more than one register would need us to
2645 * understand coalescing out more than one MOV at a time.
2647 if (scan_inst
->regs_written
> scan_inst
->exec_size
/ 8)
2650 /* SEND instructions can't have MRF as a destination. */
2651 if (scan_inst
->mlen
)
2654 if (devinfo
->gen
== 6) {
2655 /* gen6 math instructions must have the destination be
2656 * GRF, so no compute-to-MRF for them.
2658 if (scan_inst
->is_math()) {
2663 if (scan_inst
->dst
.reg_offset
== inst
->src
[0].reg_offset
) {
2664 /* Found the creator of our MRF's source value. */
2665 scan_inst
->dst
.file
= MRF
;
2666 scan_inst
->dst
.nr
= inst
->dst
.nr
;
2667 scan_inst
->saturate
|= inst
->saturate
;
2668 inst
->remove(block
);
2674 /* We don't handle control flow here. Most computation of
2675 * values that end up in MRFs are shortly before the MRF
2678 if (block
->start() == scan_inst
)
2681 /* You can't read from an MRF, so if someone else reads our
2682 * MRF's source GRF that we wanted to rewrite, that stops us.
2684 bool interfered
= false;
2685 for (int i
= 0; i
< scan_inst
->sources
; i
++) {
2686 if (scan_inst
->src
[i
].file
== VGRF
&&
2687 scan_inst
->src
[i
].nr
== inst
->src
[0].nr
&&
2688 scan_inst
->src
[i
].reg_offset
== inst
->src
[0].reg_offset
) {
2695 if (scan_inst
->dst
.file
== MRF
) {
2696 /* If somebody else writes our MRF here, we can't
2697 * compute-to-MRF before that.
2699 int scan_mrf_low
= scan_inst
->dst
.nr
& ~BRW_MRF_COMPR4
;
2702 if (scan_inst
->dst
.nr
& BRW_MRF_COMPR4
) {
2703 scan_mrf_high
= scan_mrf_low
+ 4;
2704 } else if (scan_inst
->exec_size
== 16) {
2705 scan_mrf_high
= scan_mrf_low
+ 1;
2707 scan_mrf_high
= scan_mrf_low
;
2710 if (mrf_low
== scan_mrf_low
||
2711 mrf_low
== scan_mrf_high
||
2712 mrf_high
== scan_mrf_low
||
2713 mrf_high
== scan_mrf_high
) {
2718 if (scan_inst
->mlen
> 0 && scan_inst
->base_mrf
!= -1) {
2719 /* Found a SEND instruction, which means that there are
2720 * live values in MRFs from base_mrf to base_mrf +
2721 * scan_inst->mlen - 1. Don't go pushing our MRF write up
2724 if (mrf_low
>= scan_inst
->base_mrf
&&
2725 mrf_low
< scan_inst
->base_mrf
+ scan_inst
->mlen
) {
2728 if (mrf_high
>= scan_inst
->base_mrf
&&
2729 mrf_high
< scan_inst
->base_mrf
+ scan_inst
->mlen
) {
2737 invalidate_live_intervals();
2743 * Eliminate FIND_LIVE_CHANNEL instructions occurring outside any control
2744 * flow. We could probably do better here with some form of divergence
2748 fs_visitor::eliminate_find_live_channel()
2750 bool progress
= false;
2753 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
2754 switch (inst
->opcode
) {
2760 case BRW_OPCODE_ENDIF
:
2761 case BRW_OPCODE_WHILE
:
2765 case FS_OPCODE_DISCARD_JUMP
:
2766 /* This can potentially make control flow non-uniform until the end
2771 case SHADER_OPCODE_FIND_LIVE_CHANNEL
:
2773 inst
->opcode
= BRW_OPCODE_MOV
;
2774 inst
->src
[0] = brw_imm_ud(0u);
2776 inst
->force_writemask_all
= true;
2790 * Once we've generated code, try to convert normal FS_OPCODE_FB_WRITE
2791 * instructions to FS_OPCODE_REP_FB_WRITE.
2794 fs_visitor::emit_repclear_shader()
2796 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
2798 int color_mrf
= base_mrf
+ 2;
2801 if (uniforms
== 1) {
2802 mov
= bld
.exec_all().group(4, 0)
2803 .MOV(brw_message_reg(color_mrf
),
2804 fs_reg(UNIFORM
, 0, BRW_REGISTER_TYPE_F
));
2806 struct brw_reg reg
=
2807 brw_reg(BRW_GENERAL_REGISTER_FILE
,
2808 2, 3, 0, 0, BRW_REGISTER_TYPE_F
,
2809 BRW_VERTICAL_STRIDE_8
,
2811 BRW_HORIZONTAL_STRIDE_4
, BRW_SWIZZLE_XYZW
, WRITEMASK_XYZW
);
2813 mov
= bld
.exec_all().group(4, 0)
2814 .MOV(vec4(brw_message_reg(color_mrf
)), fs_reg(reg
));
2818 if (key
->nr_color_regions
== 1) {
2819 write
= bld
.emit(FS_OPCODE_REP_FB_WRITE
);
2820 write
->saturate
= key
->clamp_fragment_color
;
2821 write
->base_mrf
= color_mrf
;
2823 write
->header_size
= 0;
2826 assume(key
->nr_color_regions
> 0);
2827 for (int i
= 0; i
< key
->nr_color_regions
; ++i
) {
2828 write
= bld
.emit(FS_OPCODE_REP_FB_WRITE
);
2829 write
->saturate
= key
->clamp_fragment_color
;
2830 write
->base_mrf
= base_mrf
;
2832 write
->header_size
= 2;
2840 assign_constant_locations();
2841 assign_curb_setup();
2843 /* Now that we have the uniform assigned, go ahead and force it to a vec4. */
2844 if (uniforms
== 1) {
2845 assert(mov
->src
[0].file
== FIXED_GRF
);
2846 mov
->src
[0] = brw_vec4_grf(mov
->src
[0].nr
, 0);
2851 * Walks through basic blocks, looking for repeated MRF writes and
2852 * removing the later ones.
2855 fs_visitor::remove_duplicate_mrf_writes()
2857 fs_inst
*last_mrf_move
[BRW_MAX_MRF(devinfo
->gen
)];
2858 bool progress
= false;
2860 /* Need to update the MRF tracking for compressed instructions. */
2861 if (dispatch_width
== 16)
2864 memset(last_mrf_move
, 0, sizeof(last_mrf_move
));
2866 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
2867 if (inst
->is_control_flow()) {
2868 memset(last_mrf_move
, 0, sizeof(last_mrf_move
));
2871 if (inst
->opcode
== BRW_OPCODE_MOV
&&
2872 inst
->dst
.file
== MRF
) {
2873 fs_inst
*prev_inst
= last_mrf_move
[inst
->dst
.nr
];
2874 if (prev_inst
&& inst
->equals(prev_inst
)) {
2875 inst
->remove(block
);
2881 /* Clear out the last-write records for MRFs that were overwritten. */
2882 if (inst
->dst
.file
== MRF
) {
2883 last_mrf_move
[inst
->dst
.nr
] = NULL
;
2886 if (inst
->mlen
> 0 && inst
->base_mrf
!= -1) {
2887 /* Found a SEND instruction, which will include two or fewer
2888 * implied MRF writes. We could do better here.
2890 for (int i
= 0; i
< implied_mrf_writes(inst
); i
++) {
2891 last_mrf_move
[inst
->base_mrf
+ i
] = NULL
;
2895 /* Clear out any MRF move records whose sources got overwritten. */
2896 if (inst
->dst
.file
== VGRF
) {
2897 for (unsigned int i
= 0; i
< ARRAY_SIZE(last_mrf_move
); i
++) {
2898 if (last_mrf_move
[i
] &&
2899 last_mrf_move
[i
]->src
[0].nr
== inst
->dst
.nr
) {
2900 last_mrf_move
[i
] = NULL
;
2905 if (inst
->opcode
== BRW_OPCODE_MOV
&&
2906 inst
->dst
.file
== MRF
&&
2907 inst
->src
[0].file
== VGRF
&&
2908 !inst
->is_partial_write()) {
2909 last_mrf_move
[inst
->dst
.nr
] = inst
;
2914 invalidate_live_intervals();
2920 clear_deps_for_inst_src(fs_inst
*inst
, bool *deps
, int first_grf
, int grf_len
)
2922 /* Clear the flag for registers that actually got read (as expected). */
2923 for (int i
= 0; i
< inst
->sources
; i
++) {
2925 if (inst
->src
[i
].file
== VGRF
|| inst
->src
[i
].file
== FIXED_GRF
) {
2926 grf
= inst
->src
[i
].nr
;
2931 if (grf
>= first_grf
&&
2932 grf
< first_grf
+ grf_len
) {
2933 deps
[grf
- first_grf
] = false;
2934 if (inst
->exec_size
== 16)
2935 deps
[grf
- first_grf
+ 1] = false;
2941 * Implements this workaround for the original 965:
2943 * "[DevBW, DevCL] Implementation Restrictions: As the hardware does not
2944 * check for post destination dependencies on this instruction, software
2945 * must ensure that there is no destination hazard for the case of ‘write
2946 * followed by a posted write’ shown in the following example.
2949 * 2. send r3.xy <rest of send instruction>
2952 * Due to no post-destination dependency check on the ‘send’, the above
2953 * code sequence could have two instructions (1 and 2) in flight at the
2954 * same time that both consider ‘r3’ as the target of their final writes.
2957 fs_visitor::insert_gen4_pre_send_dependency_workarounds(bblock_t
*block
,
2960 int write_len
= inst
->regs_written
;
2961 int first_write_grf
= inst
->dst
.nr
;
2962 bool needs_dep
[BRW_MAX_MRF(devinfo
->gen
)];
2963 assert(write_len
< (int)sizeof(needs_dep
) - 1);
2965 memset(needs_dep
, false, sizeof(needs_dep
));
2966 memset(needs_dep
, true, write_len
);
2968 clear_deps_for_inst_src(inst
, needs_dep
, first_write_grf
, write_len
);
2970 /* Walk backwards looking for writes to registers we're writing which
2971 * aren't read since being written. If we hit the start of the program,
2972 * we assume that there are no outstanding dependencies on entry to the
2975 foreach_inst_in_block_reverse_starting_from(fs_inst
, scan_inst
, inst
) {
2976 /* If we hit control flow, assume that there *are* outstanding
2977 * dependencies, and force their cleanup before our instruction.
2979 if (block
->start() == scan_inst
) {
2980 for (int i
= 0; i
< write_len
; i
++) {
2982 DEP_RESOLVE_MOV(fs_builder(this, block
, inst
),
2983 first_write_grf
+ i
);
2988 /* We insert our reads as late as possible on the assumption that any
2989 * instruction but a MOV that might have left us an outstanding
2990 * dependency has more latency than a MOV.
2992 if (scan_inst
->dst
.file
== VGRF
) {
2993 for (int i
= 0; i
< scan_inst
->regs_written
; i
++) {
2994 int reg
= scan_inst
->dst
.nr
+ i
;
2996 if (reg
>= first_write_grf
&&
2997 reg
< first_write_grf
+ write_len
&&
2998 needs_dep
[reg
- first_write_grf
]) {
2999 DEP_RESOLVE_MOV(fs_builder(this, block
, inst
), reg
);
3000 needs_dep
[reg
- first_write_grf
] = false;
3001 if (scan_inst
->exec_size
== 16)
3002 needs_dep
[reg
- first_write_grf
+ 1] = false;
3007 /* Clear the flag for registers that actually got read (as expected). */
3008 clear_deps_for_inst_src(scan_inst
, needs_dep
, first_write_grf
, write_len
);
3010 /* Continue the loop only if we haven't resolved all the dependencies */
3012 for (i
= 0; i
< write_len
; i
++) {
3022 * Implements this workaround for the original 965:
3024 * "[DevBW, DevCL] Errata: A destination register from a send can not be
3025 * used as a destination register until after it has been sourced by an
3026 * instruction with a different destination register.
3029 fs_visitor::insert_gen4_post_send_dependency_workarounds(bblock_t
*block
, fs_inst
*inst
)
3031 int write_len
= inst
->regs_written
;
3032 int first_write_grf
= inst
->dst
.nr
;
3033 bool needs_dep
[BRW_MAX_MRF(devinfo
->gen
)];
3034 assert(write_len
< (int)sizeof(needs_dep
) - 1);
3036 memset(needs_dep
, false, sizeof(needs_dep
));
3037 memset(needs_dep
, true, write_len
);
3038 /* Walk forwards looking for writes to registers we're writing which aren't
3039 * read before being written.
3041 foreach_inst_in_block_starting_from(fs_inst
, scan_inst
, inst
) {
3042 /* If we hit control flow, force resolve all remaining dependencies. */
3043 if (block
->end() == scan_inst
) {
3044 for (int i
= 0; i
< write_len
; i
++) {
3046 DEP_RESOLVE_MOV(fs_builder(this, block
, scan_inst
),
3047 first_write_grf
+ i
);
3052 /* Clear the flag for registers that actually got read (as expected). */
3053 clear_deps_for_inst_src(scan_inst
, needs_dep
, first_write_grf
, write_len
);
3055 /* We insert our reads as late as possible since they're reading the
3056 * result of a SEND, which has massive latency.
3058 if (scan_inst
->dst
.file
== VGRF
&&
3059 scan_inst
->dst
.nr
>= first_write_grf
&&
3060 scan_inst
->dst
.nr
< first_write_grf
+ write_len
&&
3061 needs_dep
[scan_inst
->dst
.nr
- first_write_grf
]) {
3062 DEP_RESOLVE_MOV(fs_builder(this, block
, scan_inst
),
3064 needs_dep
[scan_inst
->dst
.nr
- first_write_grf
] = false;
3067 /* Continue the loop only if we haven't resolved all the dependencies */
3069 for (i
= 0; i
< write_len
; i
++) {
3079 fs_visitor::insert_gen4_send_dependency_workarounds()
3081 if (devinfo
->gen
!= 4 || devinfo
->is_g4x
)
3084 bool progress
= false;
3086 /* Note that we're done with register allocation, so GRF fs_regs always
3087 * have a .reg_offset of 0.
3090 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
3091 if (inst
->mlen
!= 0 && inst
->dst
.file
== VGRF
) {
3092 insert_gen4_pre_send_dependency_workarounds(block
, inst
);
3093 insert_gen4_post_send_dependency_workarounds(block
, inst
);
3099 invalidate_live_intervals();
3103 * Turns the generic expression-style uniform pull constant load instruction
3104 * into a hardware-specific series of instructions for loading a pull
3107 * The expression style allows the CSE pass before this to optimize out
3108 * repeated loads from the same offset, and gives the pre-register-allocation
3109 * scheduling full flexibility, while the conversion to native instructions
3110 * allows the post-register-allocation scheduler the best information
3113 * Note that execution masking for setting up pull constant loads is special:
3114 * the channels that need to be written are unrelated to the current execution
3115 * mask, since a later instruction will use one of the result channels as a
3116 * source operand for all 8 or 16 of its channels.
3119 fs_visitor::lower_uniform_pull_constant_loads()
3121 foreach_block_and_inst (block
, fs_inst
, inst
, cfg
) {
3122 if (inst
->opcode
!= FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
)
3125 if (devinfo
->gen
>= 7) {
3126 /* The offset arg is a vec4-aligned immediate byte offset. */
3127 fs_reg const_offset_reg
= inst
->src
[1];
3128 assert(const_offset_reg
.file
== IMM
&&
3129 const_offset_reg
.type
== BRW_REGISTER_TYPE_UD
);
3130 assert(const_offset_reg
.ud
% 16 == 0);
3132 fs_reg payload
, offset
;
3133 if (devinfo
->gen
>= 9) {
3134 /* We have to use a message header on Skylake to get SIMD4x2
3135 * mode. Reserve space for the register.
3137 offset
= payload
= fs_reg(VGRF
, alloc
.allocate(2));
3138 offset
.reg_offset
++;
3141 offset
= payload
= fs_reg(VGRF
, alloc
.allocate(1));
3145 /* This is actually going to be a MOV, but since only the first dword
3146 * is accessed, we have a special opcode to do just that one. Note
3147 * that this needs to be an operation that will be considered a def
3148 * by live variable analysis, or register allocation will explode.
3150 fs_inst
*setup
= new(mem_ctx
) fs_inst(FS_OPCODE_SET_SIMD4X2_OFFSET
,
3151 8, offset
, const_offset_reg
);
3152 setup
->force_writemask_all
= true;
3154 setup
->ir
= inst
->ir
;
3155 setup
->annotation
= inst
->annotation
;
3156 inst
->insert_before(block
, setup
);
3158 /* Similarly, this will only populate the first 4 channels of the
3159 * result register (since we only use smear values from 0-3), but we
3160 * don't tell the optimizer.
3162 inst
->opcode
= FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7
;
3163 inst
->src
[1] = payload
;
3164 inst
->base_mrf
= -1;
3166 invalidate_live_intervals();
3168 /* Before register allocation, we didn't tell the scheduler about the
3169 * MRF we use. We know it's safe to use this MRF because nothing
3170 * else does except for register spill/unspill, which generates and
3171 * uses its MRF within a single IR instruction.
3173 inst
->base_mrf
= FIRST_PULL_LOAD_MRF(devinfo
->gen
) + 1;
3180 fs_visitor::lower_load_payload()
3182 bool progress
= false;
3184 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
3185 if (inst
->opcode
!= SHADER_OPCODE_LOAD_PAYLOAD
)
3188 assert(inst
->dst
.file
== MRF
|| inst
->dst
.file
== VGRF
);
3189 assert(inst
->saturate
== false);
3190 fs_reg dst
= inst
->dst
;
3192 /* Get rid of COMPR4. We'll add it back in if we need it */
3193 if (dst
.file
== MRF
)
3194 dst
.nr
= dst
.nr
& ~BRW_MRF_COMPR4
;
3196 const fs_builder
ibld(this, block
, inst
);
3197 const fs_builder hbld
= ibld
.exec_all().group(8, 0);
3199 for (uint8_t i
= 0; i
< inst
->header_size
; i
++) {
3200 if (inst
->src
[i
].file
!= BAD_FILE
) {
3201 fs_reg mov_dst
= retype(dst
, BRW_REGISTER_TYPE_UD
);
3202 fs_reg mov_src
= retype(inst
->src
[i
], BRW_REGISTER_TYPE_UD
);
3203 hbld
.MOV(mov_dst
, mov_src
);
3205 dst
= offset(dst
, hbld
, 1);
3208 if (inst
->dst
.file
== MRF
&& (inst
->dst
.nr
& BRW_MRF_COMPR4
) &&
3209 inst
->exec_size
> 8) {
3210 /* In this case, the payload portion of the LOAD_PAYLOAD isn't
3211 * a straightforward copy. Instead, the result of the
3212 * LOAD_PAYLOAD is treated as interleaved and the first four
3213 * non-header sources are unpacked as:
3224 * This is used for gen <= 5 fb writes.
3226 assert(inst
->exec_size
== 16);
3227 assert(inst
->header_size
+ 4 <= inst
->sources
);
3228 for (uint8_t i
= inst
->header_size
; i
< inst
->header_size
+ 4; i
++) {
3229 if (inst
->src
[i
].file
!= BAD_FILE
) {
3230 if (devinfo
->has_compr4
) {
3231 fs_reg compr4_dst
= retype(dst
, inst
->src
[i
].type
);
3232 compr4_dst
.nr
|= BRW_MRF_COMPR4
;
3233 ibld
.MOV(compr4_dst
, inst
->src
[i
]);
3235 /* Platform doesn't have COMPR4. We have to fake it */
3236 fs_reg mov_dst
= retype(dst
, inst
->src
[i
].type
);
3237 ibld
.half(0).MOV(mov_dst
, half(inst
->src
[i
], 0));
3239 ibld
.half(1).MOV(mov_dst
, half(inst
->src
[i
], 1));
3246 /* The loop above only ever incremented us through the first set
3247 * of 4 registers. However, thanks to the magic of COMPR4, we
3248 * actually wrote to the first 8 registers, so we need to take
3249 * that into account now.
3253 /* The COMPR4 code took care of the first 4 sources. We'll let
3254 * the regular path handle any remaining sources. Yes, we are
3255 * modifying the instruction but we're about to delete it so
3256 * this really doesn't hurt anything.
3258 inst
->header_size
+= 4;
3261 for (uint8_t i
= inst
->header_size
; i
< inst
->sources
; i
++) {
3262 if (inst
->src
[i
].file
!= BAD_FILE
)
3263 ibld
.MOV(retype(dst
, inst
->src
[i
].type
), inst
->src
[i
]);
3264 dst
= offset(dst
, ibld
, 1);
3267 inst
->remove(block
);
3272 invalidate_live_intervals();
3278 fs_visitor::lower_integer_multiplication()
3280 bool progress
= false;
3282 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
3283 const fs_builder
ibld(this, block
, inst
);
3285 if (inst
->opcode
== BRW_OPCODE_MUL
) {
3286 if (inst
->dst
.is_accumulator() ||
3287 (inst
->dst
.type
!= BRW_REGISTER_TYPE_D
&&
3288 inst
->dst
.type
!= BRW_REGISTER_TYPE_UD
))
3291 /* Gen8's MUL instruction can do a 32-bit x 32-bit -> 32-bit
3292 * operation directly, but CHV/BXT cannot.
3294 if (devinfo
->gen
>= 8 &&
3295 !devinfo
->is_cherryview
&& !devinfo
->is_broxton
)
3298 if (inst
->src
[1].file
== IMM
&&
3299 inst
->src
[1].ud
< (1 << 16)) {
3300 /* The MUL instruction isn't commutative. On Gen <= 6, only the low
3301 * 16-bits of src0 are read, and on Gen >= 7 only the low 16-bits of
3304 * If multiplying by an immediate value that fits in 16-bits, do a
3305 * single MUL instruction with that value in the proper location.
3307 if (devinfo
->gen
< 7) {
3308 fs_reg
imm(VGRF
, alloc
.allocate(dispatch_width
/ 8),
3310 ibld
.MOV(imm
, inst
->src
[1]);
3311 ibld
.MUL(inst
->dst
, imm
, inst
->src
[0]);
3313 ibld
.MUL(inst
->dst
, inst
->src
[0], inst
->src
[1]);
3316 /* Gen < 8 (and some Gen8+ low-power parts like Cherryview) cannot
3317 * do 32-bit integer multiplication in one instruction, but instead
3318 * must do a sequence (which actually calculates a 64-bit result):
3320 * mul(8) acc0<1>D g3<8,8,1>D g4<8,8,1>D
3321 * mach(8) null g3<8,8,1>D g4<8,8,1>D
3322 * mov(8) g2<1>D acc0<8,8,1>D
3324 * But on Gen > 6, the ability to use second accumulator register
3325 * (acc1) for non-float data types was removed, preventing a simple
3326 * implementation in SIMD16. A 16-channel result can be calculated by
3327 * executing the three instructions twice in SIMD8, once with quarter
3328 * control of 1Q for the first eight channels and again with 2Q for
3329 * the second eight channels.
3331 * Which accumulator register is implicitly accessed (by AccWrEnable
3332 * for instance) is determined by the quarter control. Unfortunately
3333 * Ivybridge (and presumably Baytrail) has a hardware bug in which an
3334 * implicit accumulator access by an instruction with 2Q will access
3335 * acc1 regardless of whether the data type is usable in acc1.
3337 * Specifically, the 2Q mach(8) writes acc1 which does not exist for
3338 * integer data types.
3340 * Since we only want the low 32-bits of the result, we can do two
3341 * 32-bit x 16-bit multiplies (like the mul and mach are doing), and
3342 * adjust the high result and add them (like the mach is doing):
3344 * mul(8) g7<1>D g3<8,8,1>D g4.0<8,8,1>UW
3345 * mul(8) g8<1>D g3<8,8,1>D g4.1<8,8,1>UW
3346 * shl(8) g9<1>D g8<8,8,1>D 16D
3347 * add(8) g2<1>D g7<8,8,1>D g8<8,8,1>D
3349 * We avoid the shl instruction by realizing that we only want to add
3350 * the low 16-bits of the "high" result to the high 16-bits of the
3351 * "low" result and using proper regioning on the add:
3353 * mul(8) g7<1>D g3<8,8,1>D g4.0<16,8,2>UW
3354 * mul(8) g8<1>D g3<8,8,1>D g4.1<16,8,2>UW
3355 * add(8) g7.1<2>UW g7.1<16,8,2>UW g8<16,8,2>UW
3357 * Since it does not use the (single) accumulator register, we can
3358 * schedule multi-component multiplications much better.
3361 fs_reg orig_dst
= inst
->dst
;
3362 if (orig_dst
.is_null() || orig_dst
.file
== MRF
) {
3363 inst
->dst
= fs_reg(VGRF
, alloc
.allocate(dispatch_width
/ 8),
3366 fs_reg low
= inst
->dst
;
3367 fs_reg
high(VGRF
, alloc
.allocate(dispatch_width
/ 8),
3370 if (devinfo
->gen
>= 7) {
3371 fs_reg src1_0_w
= inst
->src
[1];
3372 fs_reg src1_1_w
= inst
->src
[1];
3374 if (inst
->src
[1].file
== IMM
) {
3375 src1_0_w
.ud
&= 0xffff;
3378 src1_0_w
.type
= BRW_REGISTER_TYPE_UW
;
3379 if (src1_0_w
.stride
!= 0) {
3380 assert(src1_0_w
.stride
== 1);
3381 src1_0_w
.stride
= 2;
3384 src1_1_w
.type
= BRW_REGISTER_TYPE_UW
;
3385 if (src1_1_w
.stride
!= 0) {
3386 assert(src1_1_w
.stride
== 1);
3387 src1_1_w
.stride
= 2;
3389 src1_1_w
.subreg_offset
+= type_sz(BRW_REGISTER_TYPE_UW
);
3391 ibld
.MUL(low
, inst
->src
[0], src1_0_w
);
3392 ibld
.MUL(high
, inst
->src
[0], src1_1_w
);
3394 fs_reg src0_0_w
= inst
->src
[0];
3395 fs_reg src0_1_w
= inst
->src
[0];
3397 src0_0_w
.type
= BRW_REGISTER_TYPE_UW
;
3398 if (src0_0_w
.stride
!= 0) {
3399 assert(src0_0_w
.stride
== 1);
3400 src0_0_w
.stride
= 2;
3403 src0_1_w
.type
= BRW_REGISTER_TYPE_UW
;
3404 if (src0_1_w
.stride
!= 0) {
3405 assert(src0_1_w
.stride
== 1);
3406 src0_1_w
.stride
= 2;
3408 src0_1_w
.subreg_offset
+= type_sz(BRW_REGISTER_TYPE_UW
);
3410 ibld
.MUL(low
, src0_0_w
, inst
->src
[1]);
3411 ibld
.MUL(high
, src0_1_w
, inst
->src
[1]);
3414 fs_reg dst
= inst
->dst
;
3415 dst
.type
= BRW_REGISTER_TYPE_UW
;
3416 dst
.subreg_offset
= 2;
3419 high
.type
= BRW_REGISTER_TYPE_UW
;
3422 low
.type
= BRW_REGISTER_TYPE_UW
;
3423 low
.subreg_offset
= 2;
3426 ibld
.ADD(dst
, low
, high
);
3428 if (inst
->conditional_mod
|| orig_dst
.file
== MRF
) {
3429 set_condmod(inst
->conditional_mod
,
3430 ibld
.MOV(orig_dst
, inst
->dst
));
3434 } else if (inst
->opcode
== SHADER_OPCODE_MULH
) {
3435 /* Should have been lowered to 8-wide. */
3436 assert(inst
->exec_size
<= 8);
3437 const fs_reg acc
= retype(brw_acc_reg(inst
->exec_size
),
3439 fs_inst
*mul
= ibld
.MUL(acc
, inst
->src
[0], inst
->src
[1]);
3440 fs_inst
*mach
= ibld
.MACH(inst
->dst
, inst
->src
[0], inst
->src
[1]);
3442 if (devinfo
->gen
>= 8) {
3443 /* Until Gen8, integer multiplies read 32-bits from one source,
3444 * and 16-bits from the other, and relying on the MACH instruction
3445 * to generate the high bits of the result.
3447 * On Gen8, the multiply instruction does a full 32x32-bit
3448 * multiply, but in order to do a 64-bit multiply we can simulate
3449 * the previous behavior and then use a MACH instruction.
3451 * FINISHME: Don't use source modifiers on src1.
3453 assert(mul
->src
[1].type
== BRW_REGISTER_TYPE_D
||
3454 mul
->src
[1].type
== BRW_REGISTER_TYPE_UD
);
3455 mul
->src
[1].type
= (type_is_signed(mul
->src
[1].type
) ?
3456 BRW_REGISTER_TYPE_W
: BRW_REGISTER_TYPE_UW
);
3457 mul
->src
[1].stride
*= 2;
3459 } else if (devinfo
->gen
== 7 && !devinfo
->is_haswell
&&
3460 inst
->force_sechalf
) {
3461 /* Among other things the quarter control bits influence which
3462 * accumulator register is used by the hardware for instructions
3463 * that access the accumulator implicitly (e.g. MACH). A
3464 * second-half instruction would normally map to acc1, which
3465 * doesn't exist on Gen7 and up (the hardware does emulate it for
3466 * floating-point instructions *only* by taking advantage of the
3467 * extra precision of acc0 not normally used for floating point
3470 * HSW and up are careful enough not to try to access an
3471 * accumulator register that doesn't exist, but on earlier Gen7
3472 * hardware we need to make sure that the quarter control bits are
3473 * zero to avoid non-deterministic behaviour and emit an extra MOV
3474 * to get the result masked correctly according to the current
3477 mach
->force_sechalf
= false;
3478 mach
->force_writemask_all
= true;
3479 mach
->dst
= ibld
.vgrf(inst
->dst
.type
);
3480 ibld
.MOV(inst
->dst
, mach
->dst
);
3486 inst
->remove(block
);
3491 invalidate_live_intervals();
3497 setup_color_payload(const fs_builder
&bld
, const brw_wm_prog_key
*key
,
3498 fs_reg
*dst
, fs_reg color
, unsigned components
)
3500 if (key
->clamp_fragment_color
) {
3501 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 4);
3502 assert(color
.type
== BRW_REGISTER_TYPE_F
);
3504 for (unsigned i
= 0; i
< components
; i
++)
3506 bld
.MOV(offset(tmp
, bld
, i
), offset(color
, bld
, i
)));
3511 for (unsigned i
= 0; i
< components
; i
++)
3512 dst
[i
] = offset(color
, bld
, i
);
3516 lower_fb_write_logical_send(const fs_builder
&bld
, fs_inst
*inst
,
3517 const brw_wm_prog_data
*prog_data
,
3518 const brw_wm_prog_key
*key
,
3519 const fs_visitor::thread_payload
&payload
)
3521 assert(inst
->src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].file
== IMM
);
3522 const brw_device_info
*devinfo
= bld
.shader
->devinfo
;
3523 const fs_reg
&color0
= inst
->src
[FB_WRITE_LOGICAL_SRC_COLOR0
];
3524 const fs_reg
&color1
= inst
->src
[FB_WRITE_LOGICAL_SRC_COLOR1
];
3525 const fs_reg
&src0_alpha
= inst
->src
[FB_WRITE_LOGICAL_SRC_SRC0_ALPHA
];
3526 const fs_reg
&src_depth
= inst
->src
[FB_WRITE_LOGICAL_SRC_SRC_DEPTH
];
3527 const fs_reg
&dst_depth
= inst
->src
[FB_WRITE_LOGICAL_SRC_DST_DEPTH
];
3528 const fs_reg
&src_stencil
= inst
->src
[FB_WRITE_LOGICAL_SRC_SRC_STENCIL
];
3529 fs_reg sample_mask
= inst
->src
[FB_WRITE_LOGICAL_SRC_OMASK
];
3530 const unsigned components
=
3531 inst
->src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].ud
;
3533 /* We can potentially have a message length of up to 15, so we have to set
3534 * base_mrf to either 0 or 1 in order to fit in m0..m15.
3537 int header_size
= 2, payload_header_size
;
3538 unsigned length
= 0;
3540 /* From the Sandy Bridge PRM, volume 4, page 198:
3542 * "Dispatched Pixel Enables. One bit per pixel indicating
3543 * which pixels were originally enabled when the thread was
3544 * dispatched. This field is only required for the end-of-
3545 * thread message and on all dual-source messages."
3547 if (devinfo
->gen
>= 6 &&
3548 (devinfo
->is_haswell
|| devinfo
->gen
>= 8 || !prog_data
->uses_kill
) &&
3549 color1
.file
== BAD_FILE
&&
3550 key
->nr_color_regions
== 1) {
3554 if (header_size
!= 0) {
3555 assert(header_size
== 2);
3556 /* Allocate 2 registers for a header */
3560 if (payload
.aa_dest_stencil_reg
) {
3561 sources
[length
] = fs_reg(VGRF
, bld
.shader
->alloc
.allocate(1));
3562 bld
.group(8, 0).exec_all().annotate("FB write stencil/AA alpha")
3563 .MOV(sources
[length
],
3564 fs_reg(brw_vec8_grf(payload
.aa_dest_stencil_reg
, 0)));
3568 if (prog_data
->uses_omask
) {
3569 sources
[length
] = fs_reg(VGRF
, bld
.shader
->alloc
.allocate(1),
3570 BRW_REGISTER_TYPE_UD
);
3572 /* Hand over gl_SampleMask. Only the lower 16 bits of each channel are
3573 * relevant. Since it's unsigned single words one vgrf is always
3574 * 16-wide, but only the lower or higher 8 channels will be used by the
3575 * hardware when doing a SIMD8 write depending on whether we have
3576 * selected the subspans for the first or second half respectively.
3578 assert(sample_mask
.file
!= BAD_FILE
&& type_sz(sample_mask
.type
) == 4);
3579 sample_mask
.type
= BRW_REGISTER_TYPE_UW
;
3580 sample_mask
.stride
*= 2;
3582 bld
.exec_all().annotate("FB write oMask")
3583 .MOV(half(retype(sources
[length
], BRW_REGISTER_TYPE_UW
),
3584 inst
->force_sechalf
),
3589 payload_header_size
= length
;
3591 if (src0_alpha
.file
!= BAD_FILE
) {
3592 /* FIXME: This is being passed at the wrong location in the payload and
3593 * doesn't work when gl_SampleMask and MRTs are used simultaneously.
3594 * It's supposed to be immediately before oMask but there seems to be no
3595 * reasonable way to pass them in the correct order because LOAD_PAYLOAD
3596 * requires header sources to form a contiguous segment at the beginning
3597 * of the message and src0_alpha has per-channel semantics.
3599 setup_color_payload(bld
, key
, &sources
[length
], src0_alpha
, 1);
3603 setup_color_payload(bld
, key
, &sources
[length
], color0
, components
);
3606 if (color1
.file
!= BAD_FILE
) {
3607 setup_color_payload(bld
, key
, &sources
[length
], color1
, components
);
3611 if (src_depth
.file
!= BAD_FILE
) {
3612 sources
[length
] = src_depth
;
3616 if (dst_depth
.file
!= BAD_FILE
) {
3617 sources
[length
] = dst_depth
;
3621 if (src_stencil
.file
!= BAD_FILE
) {
3622 assert(devinfo
->gen
>= 9);
3623 assert(bld
.dispatch_width() != 16);
3625 /* XXX: src_stencil is only available on gen9+. dst_depth is never
3626 * available on gen9+. As such it's impossible to have both enabled at the
3627 * same time and therefore length cannot overrun the array.
3629 assert(length
< 15);
3631 sources
[length
] = bld
.vgrf(BRW_REGISTER_TYPE_UD
);
3632 bld
.exec_all().annotate("FB write OS")
3633 .emit(FS_OPCODE_PACK_STENCIL_REF
, sources
[length
],
3634 retype(src_stencil
, BRW_REGISTER_TYPE_UB
));
3639 if (devinfo
->gen
>= 7) {
3640 /* Send from the GRF */
3641 fs_reg payload
= fs_reg(VGRF
, -1, BRW_REGISTER_TYPE_F
);
3642 load
= bld
.LOAD_PAYLOAD(payload
, sources
, length
, payload_header_size
);
3643 payload
.nr
= bld
.shader
->alloc
.allocate(load
->regs_written
);
3644 load
->dst
= payload
;
3646 inst
->src
[0] = payload
;
3647 inst
->resize_sources(1);
3648 inst
->base_mrf
= -1;
3650 /* Send from the MRF */
3651 load
= bld
.LOAD_PAYLOAD(fs_reg(MRF
, 1, BRW_REGISTER_TYPE_F
),
3652 sources
, length
, payload_header_size
);
3654 /* On pre-SNB, we have to interlace the color values. LOAD_PAYLOAD
3655 * will do this for us if we just give it a COMPR4 destination.
3657 if (devinfo
->gen
< 6 && bld
.dispatch_width() == 16)
3658 load
->dst
.nr
|= BRW_MRF_COMPR4
;
3660 inst
->resize_sources(0);
3664 inst
->opcode
= FS_OPCODE_FB_WRITE
;
3665 inst
->mlen
= load
->regs_written
;
3666 inst
->header_size
= header_size
;
3670 lower_sampler_logical_send_gen4(const fs_builder
&bld
, fs_inst
*inst
, opcode op
,
3671 const fs_reg
&coordinate
,
3672 const fs_reg
&shadow_c
,
3673 const fs_reg
&lod
, const fs_reg
&lod2
,
3674 const fs_reg
&surface
,
3675 const fs_reg
&sampler
,
3676 unsigned coord_components
,
3677 unsigned grad_components
)
3679 const bool has_lod
= (op
== SHADER_OPCODE_TXL
|| op
== FS_OPCODE_TXB
||
3680 op
== SHADER_OPCODE_TXF
|| op
== SHADER_OPCODE_TXS
);
3681 fs_reg
msg_begin(MRF
, 1, BRW_REGISTER_TYPE_F
);
3682 fs_reg msg_end
= msg_begin
;
3685 msg_end
= offset(msg_end
, bld
.group(8, 0), 1);
3687 for (unsigned i
= 0; i
< coord_components
; i
++)
3688 bld
.MOV(retype(offset(msg_end
, bld
, i
), coordinate
.type
),
3689 offset(coordinate
, bld
, i
));
3691 msg_end
= offset(msg_end
, bld
, coord_components
);
3693 /* Messages other than SAMPLE and RESINFO in SIMD16 and TXD in SIMD8
3694 * require all three components to be present and zero if they are unused.
3696 if (coord_components
> 0 &&
3697 (has_lod
|| shadow_c
.file
!= BAD_FILE
||
3698 (op
== SHADER_OPCODE_TEX
&& bld
.dispatch_width() == 8))) {
3699 for (unsigned i
= coord_components
; i
< 3; i
++)
3700 bld
.MOV(offset(msg_end
, bld
, i
), brw_imm_f(0.0f
));
3702 msg_end
= offset(msg_end
, bld
, 3 - coord_components
);
3705 if (op
== SHADER_OPCODE_TXD
) {
3706 /* TXD unsupported in SIMD16 mode. */
3707 assert(bld
.dispatch_width() == 8);
3709 /* the slots for u and v are always present, but r is optional */
3710 if (coord_components
< 2)
3711 msg_end
= offset(msg_end
, bld
, 2 - coord_components
);
3714 * dPdx = dudx, dvdx, drdx
3715 * dPdy = dudy, dvdy, drdy
3717 * 1-arg: Does not exist.
3719 * 2-arg: dudx dvdx dudy dvdy
3720 * dPdx.x dPdx.y dPdy.x dPdy.y
3723 * 3-arg: dudx dvdx drdx dudy dvdy drdy
3724 * dPdx.x dPdx.y dPdx.z dPdy.x dPdy.y dPdy.z
3725 * m5 m6 m7 m8 m9 m10
3727 for (unsigned i
= 0; i
< grad_components
; i
++)
3728 bld
.MOV(offset(msg_end
, bld
, i
), offset(lod
, bld
, i
));
3730 msg_end
= offset(msg_end
, bld
, MAX2(grad_components
, 2));
3732 for (unsigned i
= 0; i
< grad_components
; i
++)
3733 bld
.MOV(offset(msg_end
, bld
, i
), offset(lod2
, bld
, i
));
3735 msg_end
= offset(msg_end
, bld
, MAX2(grad_components
, 2));
3739 /* Bias/LOD with shadow comparitor is unsupported in SIMD16 -- *Without*
3740 * shadow comparitor (including RESINFO) it's unsupported in SIMD8 mode.
3742 assert(shadow_c
.file
!= BAD_FILE
? bld
.dispatch_width() == 8 :
3743 bld
.dispatch_width() == 16);
3745 const brw_reg_type type
=
3746 (op
== SHADER_OPCODE_TXF
|| op
== SHADER_OPCODE_TXS
?
3747 BRW_REGISTER_TYPE_UD
: BRW_REGISTER_TYPE_F
);
3748 bld
.MOV(retype(msg_end
, type
), lod
);
3749 msg_end
= offset(msg_end
, bld
, 1);
3752 if (shadow_c
.file
!= BAD_FILE
) {
3753 if (op
== SHADER_OPCODE_TEX
&& bld
.dispatch_width() == 8) {
3754 /* There's no plain shadow compare message, so we use shadow
3755 * compare with a bias of 0.0.
3757 bld
.MOV(msg_end
, brw_imm_f(0.0f
));
3758 msg_end
= offset(msg_end
, bld
, 1);
3761 bld
.MOV(msg_end
, shadow_c
);
3762 msg_end
= offset(msg_end
, bld
, 1);
3766 inst
->src
[0] = reg_undef
;
3767 inst
->src
[1] = surface
;
3768 inst
->src
[2] = sampler
;
3769 inst
->resize_sources(3);
3770 inst
->base_mrf
= msg_begin
.nr
;
3771 inst
->mlen
= msg_end
.nr
- msg_begin
.nr
;
3772 inst
->header_size
= 1;
3776 lower_sampler_logical_send_gen5(const fs_builder
&bld
, fs_inst
*inst
, opcode op
,
3778 const fs_reg
&shadow_c
,
3779 fs_reg lod
, fs_reg lod2
,
3780 const fs_reg
&sample_index
,
3781 const fs_reg
&surface
,
3782 const fs_reg
&sampler
,
3783 const fs_reg
&offset_value
,
3784 unsigned coord_components
,
3785 unsigned grad_components
)
3787 fs_reg
message(MRF
, 2, BRW_REGISTER_TYPE_F
);
3788 fs_reg msg_coords
= message
;
3789 unsigned header_size
= 0;
3791 if (offset_value
.file
!= BAD_FILE
) {
3792 /* The offsets set up by the visitor are in the m1 header, so we can't
3799 for (unsigned i
= 0; i
< coord_components
; i
++) {
3800 bld
.MOV(retype(offset(msg_coords
, bld
, i
), coordinate
.type
), coordinate
);
3801 coordinate
= offset(coordinate
, bld
, 1);
3803 fs_reg msg_end
= offset(msg_coords
, bld
, coord_components
);
3804 fs_reg msg_lod
= offset(msg_coords
, bld
, 4);
3806 if (shadow_c
.file
!= BAD_FILE
) {
3807 fs_reg msg_shadow
= msg_lod
;
3808 bld
.MOV(msg_shadow
, shadow_c
);
3809 msg_lod
= offset(msg_shadow
, bld
, 1);
3814 case SHADER_OPCODE_TXL
:
3816 bld
.MOV(msg_lod
, lod
);
3817 msg_end
= offset(msg_lod
, bld
, 1);
3819 case SHADER_OPCODE_TXD
:
3822 * dPdx = dudx, dvdx, drdx
3823 * dPdy = dudy, dvdy, drdy
3825 * Load up these values:
3826 * - dudx dudy dvdx dvdy drdx drdy
3827 * - dPdx.x dPdy.x dPdx.y dPdy.y dPdx.z dPdy.z
3830 for (unsigned i
= 0; i
< grad_components
; i
++) {
3831 bld
.MOV(msg_end
, lod
);
3832 lod
= offset(lod
, bld
, 1);
3833 msg_end
= offset(msg_end
, bld
, 1);
3835 bld
.MOV(msg_end
, lod2
);
3836 lod2
= offset(lod2
, bld
, 1);
3837 msg_end
= offset(msg_end
, bld
, 1);
3840 case SHADER_OPCODE_TXS
:
3841 msg_lod
= retype(msg_end
, BRW_REGISTER_TYPE_UD
);
3842 bld
.MOV(msg_lod
, lod
);
3843 msg_end
= offset(msg_lod
, bld
, 1);
3845 case SHADER_OPCODE_TXF
:
3846 msg_lod
= offset(msg_coords
, bld
, 3);
3847 bld
.MOV(retype(msg_lod
, BRW_REGISTER_TYPE_UD
), lod
);
3848 msg_end
= offset(msg_lod
, bld
, 1);
3850 case SHADER_OPCODE_TXF_CMS
:
3851 msg_lod
= offset(msg_coords
, bld
, 3);
3853 bld
.MOV(retype(msg_lod
, BRW_REGISTER_TYPE_UD
), brw_imm_ud(0u));
3855 bld
.MOV(retype(offset(msg_lod
, bld
, 1), BRW_REGISTER_TYPE_UD
), sample_index
);
3856 msg_end
= offset(msg_lod
, bld
, 2);
3863 inst
->src
[0] = reg_undef
;
3864 inst
->src
[1] = surface
;
3865 inst
->src
[2] = sampler
;
3866 inst
->resize_sources(3);
3867 inst
->base_mrf
= message
.nr
;
3868 inst
->mlen
= msg_end
.nr
- message
.nr
;
3869 inst
->header_size
= header_size
;
3871 /* Message length > MAX_SAMPLER_MESSAGE_SIZE disallowed by hardware. */
3872 assert(inst
->mlen
<= MAX_SAMPLER_MESSAGE_SIZE
);
3876 is_high_sampler(const struct brw_device_info
*devinfo
, const fs_reg
&sampler
)
3878 if (devinfo
->gen
< 8 && !devinfo
->is_haswell
)
3881 return sampler
.file
!= IMM
|| sampler
.ud
>= 16;
3885 lower_sampler_logical_send_gen7(const fs_builder
&bld
, fs_inst
*inst
, opcode op
,
3887 const fs_reg
&shadow_c
,
3888 fs_reg lod
, fs_reg lod2
,
3889 const fs_reg
&sample_index
,
3891 const fs_reg
&surface
,
3892 const fs_reg
&sampler
,
3893 fs_reg offset_value
,
3894 unsigned coord_components
,
3895 unsigned grad_components
)
3897 const brw_device_info
*devinfo
= bld
.shader
->devinfo
;
3898 int reg_width
= bld
.dispatch_width() / 8;
3899 unsigned header_size
= 0, length
= 0;
3900 fs_reg sources
[MAX_SAMPLER_MESSAGE_SIZE
];
3901 for (unsigned i
= 0; i
< ARRAY_SIZE(sources
); i
++)
3902 sources
[i
] = bld
.vgrf(BRW_REGISTER_TYPE_F
);
3904 if (op
== SHADER_OPCODE_TG4
|| op
== SHADER_OPCODE_TG4_OFFSET
||
3905 offset_value
.file
!= BAD_FILE
||
3906 is_high_sampler(devinfo
, sampler
)) {
3907 /* For general texture offsets (no txf workaround), we need a header to
3908 * put them in. Note that we're only reserving space for it in the
3909 * message payload as it will be initialized implicitly by the
3912 * TG4 needs to place its channel select in the header, for interaction
3913 * with ARB_texture_swizzle. The sampler index is only 4-bits, so for
3914 * larger sampler numbers we need to offset the Sampler State Pointer in
3918 sources
[0] = fs_reg();
3922 if (shadow_c
.file
!= BAD_FILE
) {
3923 bld
.MOV(sources
[length
], shadow_c
);
3927 bool coordinate_done
= false;
3929 /* The sampler can only meaningfully compute LOD for fragment shader
3930 * messages. For all other stages, we change the opcode to TXL and
3931 * hardcode the LOD to 0.
3933 if (bld
.shader
->stage
!= MESA_SHADER_FRAGMENT
&&
3934 op
== SHADER_OPCODE_TEX
) {
3935 op
= SHADER_OPCODE_TXL
;
3936 lod
= brw_imm_f(0.0f
);
3939 /* Set up the LOD info */
3942 case SHADER_OPCODE_TXL
:
3943 bld
.MOV(sources
[length
], lod
);
3946 case SHADER_OPCODE_TXD
:
3947 /* TXD should have been lowered in SIMD16 mode. */
3948 assert(bld
.dispatch_width() == 8);
3950 /* Load dPdx and the coordinate together:
3951 * [hdr], [ref], x, dPdx.x, dPdy.x, y, dPdx.y, dPdy.y, z, dPdx.z, dPdy.z
3953 for (unsigned i
= 0; i
< coord_components
; i
++) {
3954 bld
.MOV(sources
[length
], coordinate
);
3955 coordinate
= offset(coordinate
, bld
, 1);
3958 /* For cube map array, the coordinate is (u,v,r,ai) but there are
3959 * only derivatives for (u, v, r).
3961 if (i
< grad_components
) {
3962 bld
.MOV(sources
[length
], lod
);
3963 lod
= offset(lod
, bld
, 1);
3966 bld
.MOV(sources
[length
], lod2
);
3967 lod2
= offset(lod2
, bld
, 1);
3972 coordinate_done
= true;
3974 case SHADER_OPCODE_TXS
:
3975 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
), lod
);
3978 case SHADER_OPCODE_TXF
:
3979 /* Unfortunately, the parameters for LD are intermixed: u, lod, v, r.
3980 * On Gen9 they are u, v, lod, r
3982 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
), coordinate
);
3983 coordinate
= offset(coordinate
, bld
, 1);
3986 if (devinfo
->gen
>= 9) {
3987 if (coord_components
>= 2) {
3988 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
), coordinate
);
3989 coordinate
= offset(coordinate
, bld
, 1);
3994 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
), lod
);
3997 for (unsigned i
= devinfo
->gen
>= 9 ? 2 : 1; i
< coord_components
; i
++) {
3998 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
), coordinate
);
3999 coordinate
= offset(coordinate
, bld
, 1);
4003 coordinate_done
= true;
4005 case SHADER_OPCODE_TXF_CMS
:
4006 case SHADER_OPCODE_TXF_CMS_W
:
4007 case SHADER_OPCODE_TXF_UMS
:
4008 case SHADER_OPCODE_TXF_MCS
:
4009 if (op
== SHADER_OPCODE_TXF_UMS
||
4010 op
== SHADER_OPCODE_TXF_CMS
||
4011 op
== SHADER_OPCODE_TXF_CMS_W
) {
4012 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
), sample_index
);
4016 if (op
== SHADER_OPCODE_TXF_CMS
|| op
== SHADER_OPCODE_TXF_CMS_W
) {
4017 /* Data from the multisample control surface. */
4018 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
), mcs
);
4021 /* On Gen9+ we'll use ld2dms_w instead which has two registers for
4024 if (op
== SHADER_OPCODE_TXF_CMS_W
) {
4025 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
),
4028 offset(mcs
, bld
, 1));
4033 /* There is no offsetting for this message; just copy in the integer
4034 * texture coordinates.
4036 for (unsigned i
= 0; i
< coord_components
; i
++) {
4037 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
), coordinate
);
4038 coordinate
= offset(coordinate
, bld
, 1);
4042 coordinate_done
= true;
4044 case SHADER_OPCODE_TG4_OFFSET
:
4045 /* gather4_po_c should have been lowered in SIMD16 mode. */
4046 assert(bld
.dispatch_width() == 8 || shadow_c
.file
== BAD_FILE
);
4048 /* More crazy intermixing */
4049 for (unsigned i
= 0; i
< 2; i
++) { /* u, v */
4050 bld
.MOV(sources
[length
], coordinate
);
4051 coordinate
= offset(coordinate
, bld
, 1);
4055 for (unsigned i
= 0; i
< 2; i
++) { /* offu, offv */
4056 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
), offset_value
);
4057 offset_value
= offset(offset_value
, bld
, 1);
4061 if (coord_components
== 3) { /* r if present */
4062 bld
.MOV(sources
[length
], coordinate
);
4063 coordinate
= offset(coordinate
, bld
, 1);
4067 coordinate_done
= true;
4073 /* Set up the coordinate (except for cases where it was done above) */
4074 if (!coordinate_done
) {
4075 for (unsigned i
= 0; i
< coord_components
; i
++) {
4076 bld
.MOV(sources
[length
], coordinate
);
4077 coordinate
= offset(coordinate
, bld
, 1);
4084 mlen
= length
* reg_width
- header_size
;
4086 mlen
= length
* reg_width
;
4088 const fs_reg src_payload
= fs_reg(VGRF
, bld
.shader
->alloc
.allocate(mlen
),
4089 BRW_REGISTER_TYPE_F
);
4090 bld
.LOAD_PAYLOAD(src_payload
, sources
, length
, header_size
);
4092 /* Generate the SEND. */
4094 inst
->src
[0] = src_payload
;
4095 inst
->src
[1] = surface
;
4096 inst
->src
[2] = sampler
;
4097 inst
->resize_sources(3);
4098 inst
->base_mrf
= -1;
4100 inst
->header_size
= header_size
;
4102 /* Message length > MAX_SAMPLER_MESSAGE_SIZE disallowed by hardware. */
4103 assert(inst
->mlen
<= MAX_SAMPLER_MESSAGE_SIZE
);
4107 lower_sampler_logical_send(const fs_builder
&bld
, fs_inst
*inst
, opcode op
)
4109 const brw_device_info
*devinfo
= bld
.shader
->devinfo
;
4110 const fs_reg
&coordinate
= inst
->src
[0];
4111 const fs_reg
&shadow_c
= inst
->src
[1];
4112 const fs_reg
&lod
= inst
->src
[2];
4113 const fs_reg
&lod2
= inst
->src
[3];
4114 const fs_reg
&sample_index
= inst
->src
[4];
4115 const fs_reg
&mcs
= inst
->src
[5];
4116 const fs_reg
&surface
= inst
->src
[6];
4117 const fs_reg
&sampler
= inst
->src
[7];
4118 const fs_reg
&offset_value
= inst
->src
[8];
4119 assert(inst
->src
[9].file
== IMM
&& inst
->src
[10].file
== IMM
);
4120 const unsigned coord_components
= inst
->src
[9].ud
;
4121 const unsigned grad_components
= inst
->src
[10].ud
;
4123 if (devinfo
->gen
>= 7) {
4124 lower_sampler_logical_send_gen7(bld
, inst
, op
, coordinate
,
4125 shadow_c
, lod
, lod2
, sample_index
,
4126 mcs
, surface
, sampler
, offset_value
,
4127 coord_components
, grad_components
);
4128 } else if (devinfo
->gen
>= 5) {
4129 lower_sampler_logical_send_gen5(bld
, inst
, op
, coordinate
,
4130 shadow_c
, lod
, lod2
, sample_index
,
4131 surface
, sampler
, offset_value
,
4132 coord_components
, grad_components
);
4134 lower_sampler_logical_send_gen4(bld
, inst
, op
, coordinate
,
4135 shadow_c
, lod
, lod2
,
4137 coord_components
, grad_components
);
4142 * Initialize the header present in some typed and untyped surface
4146 emit_surface_header(const fs_builder
&bld
, const fs_reg
&sample_mask
)
4148 fs_builder ubld
= bld
.exec_all().group(8, 0);
4149 const fs_reg dst
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
4150 ubld
.MOV(dst
, brw_imm_d(0));
4151 ubld
.MOV(component(dst
, 7), sample_mask
);
4156 lower_surface_logical_send(const fs_builder
&bld
, fs_inst
*inst
, opcode op
,
4157 const fs_reg
&sample_mask
)
4159 /* Get the logical send arguments. */
4160 const fs_reg
&addr
= inst
->src
[0];
4161 const fs_reg
&src
= inst
->src
[1];
4162 const fs_reg
&surface
= inst
->src
[2];
4163 const UNUSED fs_reg
&dims
= inst
->src
[3];
4164 const fs_reg
&arg
= inst
->src
[4];
4166 /* Calculate the total number of components of the payload. */
4167 const unsigned addr_sz
= inst
->components_read(0);
4168 const unsigned src_sz
= inst
->components_read(1);
4169 const unsigned header_sz
= (sample_mask
.file
== BAD_FILE
? 0 : 1);
4170 const unsigned sz
= header_sz
+ addr_sz
+ src_sz
;
4172 /* Allocate space for the payload. */
4173 fs_reg
*const components
= new fs_reg
[sz
];
4174 const fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, sz
);
4177 /* Construct the payload. */
4179 components
[n
++] = emit_surface_header(bld
, sample_mask
);
4181 for (unsigned i
= 0; i
< addr_sz
; i
++)
4182 components
[n
++] = offset(addr
, bld
, i
);
4184 for (unsigned i
= 0; i
< src_sz
; i
++)
4185 components
[n
++] = offset(src
, bld
, i
);
4187 bld
.LOAD_PAYLOAD(payload
, components
, sz
, header_sz
);
4189 /* Update the original instruction. */
4191 inst
->mlen
= header_sz
+ (addr_sz
+ src_sz
) * inst
->exec_size
/ 8;
4192 inst
->header_size
= header_sz
;
4194 inst
->src
[0] = payload
;
4195 inst
->src
[1] = surface
;
4197 inst
->resize_sources(3);
4199 delete[] components
;
4203 fs_visitor::lower_logical_sends()
4205 bool progress
= false;
4207 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
4208 const fs_builder
ibld(this, block
, inst
);
4210 switch (inst
->opcode
) {
4211 case FS_OPCODE_FB_WRITE_LOGICAL
:
4212 assert(stage
== MESA_SHADER_FRAGMENT
);
4213 lower_fb_write_logical_send(ibld
, inst
,
4214 (const brw_wm_prog_data
*)prog_data
,
4215 (const brw_wm_prog_key
*)key
,
4219 case SHADER_OPCODE_TEX_LOGICAL
:
4220 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TEX
);
4223 case SHADER_OPCODE_TXD_LOGICAL
:
4224 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXD
);
4227 case SHADER_OPCODE_TXF_LOGICAL
:
4228 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF
);
4231 case SHADER_OPCODE_TXL_LOGICAL
:
4232 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXL
);
4235 case SHADER_OPCODE_TXS_LOGICAL
:
4236 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXS
);
4239 case FS_OPCODE_TXB_LOGICAL
:
4240 lower_sampler_logical_send(ibld
, inst
, FS_OPCODE_TXB
);
4243 case SHADER_OPCODE_TXF_CMS_LOGICAL
:
4244 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_CMS
);
4247 case SHADER_OPCODE_TXF_CMS_W_LOGICAL
:
4248 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_CMS_W
);
4251 case SHADER_OPCODE_TXF_UMS_LOGICAL
:
4252 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_UMS
);
4255 case SHADER_OPCODE_TXF_MCS_LOGICAL
:
4256 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_MCS
);
4259 case SHADER_OPCODE_LOD_LOGICAL
:
4260 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_LOD
);
4263 case SHADER_OPCODE_TG4_LOGICAL
:
4264 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TG4
);
4267 case SHADER_OPCODE_TG4_OFFSET_LOGICAL
:
4268 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TG4_OFFSET
);
4271 case SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
:
4272 lower_surface_logical_send(ibld
, inst
,
4273 SHADER_OPCODE_UNTYPED_SURFACE_READ
,
4277 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
:
4278 lower_surface_logical_send(ibld
, inst
,
4279 SHADER_OPCODE_UNTYPED_SURFACE_WRITE
,
4280 ibld
.sample_mask_reg());
4283 case SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL
:
4284 lower_surface_logical_send(ibld
, inst
,
4285 SHADER_OPCODE_UNTYPED_ATOMIC
,
4286 ibld
.sample_mask_reg());
4289 case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
:
4290 lower_surface_logical_send(ibld
, inst
,
4291 SHADER_OPCODE_TYPED_SURFACE_READ
,
4295 case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
:
4296 lower_surface_logical_send(ibld
, inst
,
4297 SHADER_OPCODE_TYPED_SURFACE_WRITE
,
4298 ibld
.sample_mask_reg());
4301 case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
:
4302 lower_surface_logical_send(ibld
, inst
,
4303 SHADER_OPCODE_TYPED_ATOMIC
,
4304 ibld
.sample_mask_reg());
4315 invalidate_live_intervals();
4321 * Get the closest native SIMD width supported by the hardware for instruction
4322 * \p inst. The instruction will be left untouched by
4323 * fs_visitor::lower_simd_width() if the returned value is equal to the
4324 * original execution size.
4327 get_lowered_simd_width(const struct brw_device_info
*devinfo
,
4328 const fs_inst
*inst
)
4330 switch (inst
->opcode
) {
4331 case BRW_OPCODE_MOV
:
4332 case BRW_OPCODE_SEL
:
4333 case BRW_OPCODE_NOT
:
4334 case BRW_OPCODE_AND
:
4336 case BRW_OPCODE_XOR
:
4337 case BRW_OPCODE_SHR
:
4338 case BRW_OPCODE_SHL
:
4339 case BRW_OPCODE_ASR
:
4340 case BRW_OPCODE_CMP
:
4341 case BRW_OPCODE_CMPN
:
4342 case BRW_OPCODE_CSEL
:
4343 case BRW_OPCODE_F32TO16
:
4344 case BRW_OPCODE_F16TO32
:
4345 case BRW_OPCODE_BFREV
:
4346 case BRW_OPCODE_BFE
:
4347 case BRW_OPCODE_BFI1
:
4348 case BRW_OPCODE_BFI2
:
4349 case BRW_OPCODE_ADD
:
4350 case BRW_OPCODE_MUL
:
4351 case BRW_OPCODE_AVG
:
4352 case BRW_OPCODE_FRC
:
4353 case BRW_OPCODE_RNDU
:
4354 case BRW_OPCODE_RNDD
:
4355 case BRW_OPCODE_RNDE
:
4356 case BRW_OPCODE_RNDZ
:
4357 case BRW_OPCODE_LZD
:
4358 case BRW_OPCODE_FBH
:
4359 case BRW_OPCODE_FBL
:
4360 case BRW_OPCODE_CBIT
:
4361 case BRW_OPCODE_SAD2
:
4362 case BRW_OPCODE_MAD
:
4363 case BRW_OPCODE_LRP
:
4364 case SHADER_OPCODE_RCP
:
4365 case SHADER_OPCODE_RSQ
:
4366 case SHADER_OPCODE_SQRT
:
4367 case SHADER_OPCODE_EXP2
:
4368 case SHADER_OPCODE_LOG2
:
4369 case SHADER_OPCODE_POW
:
4370 case SHADER_OPCODE_INT_QUOTIENT
:
4371 case SHADER_OPCODE_INT_REMAINDER
:
4372 case SHADER_OPCODE_SIN
:
4373 case SHADER_OPCODE_COS
: {
4374 /* According to the PRMs:
4375 * "A. In Direct Addressing mode, a source cannot span more than 2
4376 * adjacent GRF registers.
4377 * B. A destination cannot span more than 2 adjacent GRF registers."
4379 * Look for the source or destination with the largest register region
4380 * which is the one that is going to limit the overal execution size of
4381 * the instruction due to this rule.
4383 unsigned reg_count
= inst
->regs_written
;
4385 for (unsigned i
= 0; i
< inst
->sources
; i
++)
4386 reg_count
= MAX2(reg_count
, (unsigned)inst
->regs_read(i
));
4388 /* Calculate the maximum execution size of the instruction based on the
4389 * factor by which it goes over the hardware limit of 2 GRFs.
4391 return inst
->exec_size
/ DIV_ROUND_UP(reg_count
, 2);
4393 case SHADER_OPCODE_MULH
:
4394 /* MULH is lowered to the MUL/MACH sequence using the accumulator, which
4395 * is 8-wide on Gen7+.
4397 return (devinfo
->gen
>= 7 ? 8 : inst
->exec_size
);
4399 case FS_OPCODE_FB_WRITE_LOGICAL
:
4400 /* Gen6 doesn't support SIMD16 depth writes but we cannot handle them
4403 assert(devinfo
->gen
!= 6 ||
4404 inst
->src
[FB_WRITE_LOGICAL_SRC_SRC_DEPTH
].file
== BAD_FILE
||
4405 inst
->exec_size
== 8);
4406 /* Dual-source FB writes are unsupported in SIMD16 mode. */
4407 return (inst
->src
[FB_WRITE_LOGICAL_SRC_COLOR1
].file
!= BAD_FILE
?
4408 8 : inst
->exec_size
);
4410 case SHADER_OPCODE_TXD_LOGICAL
:
4411 /* TXD is unsupported in SIMD16 mode. */
4414 case SHADER_OPCODE_TG4_OFFSET_LOGICAL
: {
4415 /* gather4_po_c is unsupported in SIMD16 mode. */
4416 const fs_reg
&shadow_c
= inst
->src
[1];
4417 return (shadow_c
.file
!= BAD_FILE
? 8 : inst
->exec_size
);
4419 case SHADER_OPCODE_TXL_LOGICAL
:
4420 case FS_OPCODE_TXB_LOGICAL
: {
4421 /* Gen4 doesn't have SIMD8 non-shadow-compare bias/LOD instructions, and
4422 * Gen4-6 can't support TXL and TXB with shadow comparison in SIMD16
4423 * mode because the message exceeds the maximum length of 11.
4425 const fs_reg
&shadow_c
= inst
->src
[1];
4426 if (devinfo
->gen
== 4 && shadow_c
.file
== BAD_FILE
)
4428 else if (devinfo
->gen
< 7 && shadow_c
.file
!= BAD_FILE
)
4431 return inst
->exec_size
;
4433 case SHADER_OPCODE_TXF_LOGICAL
:
4434 case SHADER_OPCODE_TXS_LOGICAL
:
4435 /* Gen4 doesn't have SIMD8 variants for the RESINFO and LD-with-LOD
4436 * messages. Use SIMD16 instead.
4438 if (devinfo
->gen
== 4)
4441 return inst
->exec_size
;
4443 case SHADER_OPCODE_TXF_CMS_W_LOGICAL
: {
4444 /* This opcode can take up to 6 arguments which means that in some
4445 * circumstances it can end up with a message that is too long in SIMD16
4448 const unsigned coord_components
= inst
->src
[8].ud
;
4449 /* First three arguments are the sample index and the two arguments for
4452 if ((coord_components
+ 3) * 2 > MAX_SAMPLER_MESSAGE_SIZE
)
4455 return inst
->exec_size
;
4458 case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
:
4459 case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
:
4460 case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
:
4464 return inst
->exec_size
;
4469 * The \p rows array of registers represents a \p num_rows by \p num_columns
4470 * matrix in row-major order, write it in column-major order into the register
4471 * passed as destination. \p stride gives the separation between matrix
4472 * elements in the input in fs_builder::dispatch_width() units.
4475 emit_transpose(const fs_builder
&bld
,
4476 const fs_reg
&dst
, const fs_reg
*rows
,
4477 unsigned num_rows
, unsigned num_columns
, unsigned stride
)
4479 fs_reg
*const components
= new fs_reg
[num_rows
* num_columns
];
4481 for (unsigned i
= 0; i
< num_columns
; ++i
) {
4482 for (unsigned j
= 0; j
< num_rows
; ++j
)
4483 components
[num_rows
* i
+ j
] = offset(rows
[j
], bld
, stride
* i
);
4486 bld
.LOAD_PAYLOAD(dst
, components
, num_rows
* num_columns
, 0);
4488 delete[] components
;
4492 fs_visitor::lower_simd_width()
4494 bool progress
= false;
4496 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
4497 const unsigned lower_width
= get_lowered_simd_width(devinfo
, inst
);
4499 if (lower_width
!= inst
->exec_size
) {
4500 /* Builder matching the original instruction. We may also need to
4501 * emit an instruction of width larger than the original, set the
4502 * execution size of the builder to the highest of both for now so
4503 * we're sure that both cases can be handled.
4505 const fs_builder ibld
= bld
.at(block
, inst
)
4506 .exec_all(inst
->force_writemask_all
)
4507 .group(MAX2(inst
->exec_size
, lower_width
),
4508 inst
->force_sechalf
);
4510 /* Split the copies in chunks of the execution width of either the
4511 * original or the lowered instruction, whichever is lower.
4513 const unsigned copy_width
= MIN2(lower_width
, inst
->exec_size
);
4514 const unsigned n
= inst
->exec_size
/ copy_width
;
4515 const unsigned dst_size
= inst
->regs_written
* REG_SIZE
/
4516 inst
->dst
.component_size(inst
->exec_size
);
4519 assert(n
> 0 && n
<= ARRAY_SIZE(dsts
) &&
4520 !inst
->writes_accumulator
&& !inst
->mlen
);
4522 for (unsigned i
= 0; i
< n
; i
++) {
4523 /* Emit a copy of the original instruction with the lowered width.
4524 * If the EOT flag was set throw it away except for the last
4525 * instruction to avoid killing the thread prematurely.
4527 fs_inst split_inst
= *inst
;
4528 split_inst
.exec_size
= lower_width
;
4529 split_inst
.eot
= inst
->eot
&& i
== n
- 1;
4531 /* Select the correct channel enables for the i-th group, then
4532 * transform the sources and destination and emit the lowered
4535 const fs_builder lbld
= ibld
.group(lower_width
, i
);
4537 for (unsigned j
= 0; j
< inst
->sources
; j
++) {
4538 if (inst
->src
[j
].file
!= BAD_FILE
&&
4539 !is_uniform(inst
->src
[j
])) {
4540 /* Get the i-th copy_width-wide chunk of the source. */
4541 const fs_reg src
= horiz_offset(inst
->src
[j
], copy_width
* i
);
4542 const unsigned src_size
= inst
->components_read(j
);
4544 /* Use a trivial transposition to copy one every n
4545 * copy_width-wide components of the register into a
4546 * temporary passed as source to the lowered instruction.
4548 split_inst
.src
[j
] = lbld
.vgrf(inst
->src
[j
].type
, src_size
);
4549 emit_transpose(lbld
.group(copy_width
, 0),
4550 split_inst
.src
[j
], &src
, 1, src_size
, n
);
4554 if (inst
->regs_written
) {
4555 /* Allocate enough space to hold the result of the lowered
4556 * instruction and fix up the number of registers written.
4558 split_inst
.dst
= dsts
[i
] =
4559 lbld
.vgrf(inst
->dst
.type
, dst_size
);
4560 split_inst
.regs_written
=
4561 DIV_ROUND_UP(inst
->regs_written
* lower_width
,
4565 lbld
.emit(split_inst
);
4568 if (inst
->regs_written
) {
4569 /* Distance between useful channels in the temporaries, skipping
4570 * garbage if the lowered instruction is wider than the original.
4572 const unsigned m
= lower_width
/ copy_width
;
4574 /* Interleave the components of the result from the lowered
4575 * instructions. We need to set exec_all() when copying more than
4576 * one half per component, because LOAD_PAYLOAD (in terms of which
4577 * emit_transpose is implemented) can only use the same channel
4578 * enable signals for all of its non-header sources.
4580 emit_transpose(ibld
.exec_all(inst
->exec_size
> copy_width
)
4581 .group(copy_width
, 0),
4582 inst
->dst
, dsts
, n
, dst_size
, m
);
4585 inst
->remove(block
);
4591 invalidate_live_intervals();
4597 fs_visitor::dump_instructions()
4599 dump_instructions(NULL
);
4603 fs_visitor::dump_instructions(const char *name
)
4605 FILE *file
= stderr
;
4606 if (name
&& geteuid() != 0) {
4607 file
= fopen(name
, "w");
4613 calculate_register_pressure();
4614 int ip
= 0, max_pressure
= 0;
4615 foreach_block_and_inst(block
, backend_instruction
, inst
, cfg
) {
4616 max_pressure
= MAX2(max_pressure
, regs_live_at_ip
[ip
]);
4617 fprintf(file
, "{%3d} %4d: ", regs_live_at_ip
[ip
], ip
);
4618 dump_instruction(inst
, file
);
4621 fprintf(file
, "Maximum %3d registers live at once.\n", max_pressure
);
4624 foreach_in_list(backend_instruction
, inst
, &instructions
) {
4625 fprintf(file
, "%4d: ", ip
++);
4626 dump_instruction(inst
, file
);
4630 if (file
!= stderr
) {
4636 fs_visitor::dump_instruction(backend_instruction
*be_inst
)
4638 dump_instruction(be_inst
, stderr
);
4642 fs_visitor::dump_instruction(backend_instruction
*be_inst
, FILE *file
)
4644 fs_inst
*inst
= (fs_inst
*)be_inst
;
4646 if (inst
->predicate
) {
4647 fprintf(file
, "(%cf0.%d) ",
4648 inst
->predicate_inverse
? '-' : '+',
4652 fprintf(file
, "%s", brw_instruction_name(inst
->opcode
));
4654 fprintf(file
, ".sat");
4655 if (inst
->conditional_mod
) {
4656 fprintf(file
, "%s", conditional_modifier
[inst
->conditional_mod
]);
4657 if (!inst
->predicate
&&
4658 (devinfo
->gen
< 5 || (inst
->opcode
!= BRW_OPCODE_SEL
&&
4659 inst
->opcode
!= BRW_OPCODE_IF
&&
4660 inst
->opcode
!= BRW_OPCODE_WHILE
))) {
4661 fprintf(file
, ".f0.%d", inst
->flag_subreg
);
4664 fprintf(file
, "(%d) ", inst
->exec_size
);
4667 fprintf(file
, "(mlen: %d) ", inst
->mlen
);
4670 switch (inst
->dst
.file
) {
4672 fprintf(file
, "vgrf%d", inst
->dst
.nr
);
4673 if (alloc
.sizes
[inst
->dst
.nr
] != inst
->regs_written
||
4674 inst
->dst
.subreg_offset
)
4675 fprintf(file
, "+%d.%d",
4676 inst
->dst
.reg_offset
, inst
->dst
.subreg_offset
);
4679 fprintf(file
, "g%d", inst
->dst
.nr
);
4682 fprintf(file
, "m%d", inst
->dst
.nr
);
4685 fprintf(file
, "(null)");
4688 fprintf(file
, "***u%d***", inst
->dst
.nr
+ inst
->dst
.reg_offset
);
4691 fprintf(file
, "***attr%d***", inst
->dst
.nr
+ inst
->dst
.reg_offset
);
4694 switch (inst
->dst
.nr
) {
4696 fprintf(file
, "null");
4698 case BRW_ARF_ADDRESS
:
4699 fprintf(file
, "a0.%d", inst
->dst
.subnr
);
4701 case BRW_ARF_ACCUMULATOR
:
4702 fprintf(file
, "acc%d", inst
->dst
.subnr
);
4705 fprintf(file
, "f%d.%d", inst
->dst
.nr
& 0xf, inst
->dst
.subnr
);
4708 fprintf(file
, "arf%d.%d", inst
->dst
.nr
& 0xf, inst
->dst
.subnr
);
4711 if (inst
->dst
.subnr
)
4712 fprintf(file
, "+%d", inst
->dst
.subnr
);
4715 unreachable("not reached");
4717 if (inst
->dst
.stride
!= 1)
4718 fprintf(file
, "<%u>", inst
->dst
.stride
);
4719 fprintf(file
, ":%s, ", brw_reg_type_letters(inst
->dst
.type
));
4721 for (int i
= 0; i
< inst
->sources
; i
++) {
4722 if (inst
->src
[i
].negate
)
4724 if (inst
->src
[i
].abs
)
4726 switch (inst
->src
[i
].file
) {
4728 fprintf(file
, "vgrf%d", inst
->src
[i
].nr
);
4729 if (alloc
.sizes
[inst
->src
[i
].nr
] != (unsigned)inst
->regs_read(i
) ||
4730 inst
->src
[i
].subreg_offset
)
4731 fprintf(file
, "+%d.%d", inst
->src
[i
].reg_offset
,
4732 inst
->src
[i
].subreg_offset
);
4735 fprintf(file
, "g%d", inst
->src
[i
].nr
);
4738 fprintf(file
, "***m%d***", inst
->src
[i
].nr
);
4741 fprintf(file
, "attr%d+%d", inst
->src
[i
].nr
, inst
->src
[i
].reg_offset
);
4744 fprintf(file
, "u%d", inst
->src
[i
].nr
+ inst
->src
[i
].reg_offset
);
4745 if (inst
->src
[i
].reladdr
) {
4746 fprintf(file
, "+reladdr");
4747 } else if (inst
->src
[i
].subreg_offset
) {
4748 fprintf(file
, "+%d.%d", inst
->src
[i
].reg_offset
,
4749 inst
->src
[i
].subreg_offset
);
4753 fprintf(file
, "(null)");
4756 switch (inst
->src
[i
].type
) {
4757 case BRW_REGISTER_TYPE_F
:
4758 fprintf(file
, "%ff", inst
->src
[i
].f
);
4760 case BRW_REGISTER_TYPE_W
:
4761 case BRW_REGISTER_TYPE_D
:
4762 fprintf(file
, "%dd", inst
->src
[i
].d
);
4764 case BRW_REGISTER_TYPE_UW
:
4765 case BRW_REGISTER_TYPE_UD
:
4766 fprintf(file
, "%uu", inst
->src
[i
].ud
);
4768 case BRW_REGISTER_TYPE_VF
:
4769 fprintf(file
, "[%-gF, %-gF, %-gF, %-gF]",
4770 brw_vf_to_float((inst
->src
[i
].ud
>> 0) & 0xff),
4771 brw_vf_to_float((inst
->src
[i
].ud
>> 8) & 0xff),
4772 brw_vf_to_float((inst
->src
[i
].ud
>> 16) & 0xff),
4773 brw_vf_to_float((inst
->src
[i
].ud
>> 24) & 0xff));
4776 fprintf(file
, "???");
4781 switch (inst
->src
[i
].nr
) {
4783 fprintf(file
, "null");
4785 case BRW_ARF_ADDRESS
:
4786 fprintf(file
, "a0.%d", inst
->src
[i
].subnr
);
4788 case BRW_ARF_ACCUMULATOR
:
4789 fprintf(file
, "acc%d", inst
->src
[i
].subnr
);
4792 fprintf(file
, "f%d.%d", inst
->src
[i
].nr
& 0xf, inst
->src
[i
].subnr
);
4795 fprintf(file
, "arf%d.%d", inst
->src
[i
].nr
& 0xf, inst
->src
[i
].subnr
);
4798 if (inst
->src
[i
].subnr
)
4799 fprintf(file
, "+%d", inst
->src
[i
].subnr
);
4802 if (inst
->src
[i
].abs
)
4805 if (inst
->src
[i
].file
!= IMM
) {
4807 if (inst
->src
[i
].file
== ARF
|| inst
->src
[i
].file
== FIXED_GRF
) {
4808 unsigned hstride
= inst
->src
[i
].hstride
;
4809 stride
= (hstride
== 0 ? 0 : (1 << (hstride
- 1)));
4811 stride
= inst
->src
[i
].stride
;
4814 fprintf(file
, "<%u>", stride
);
4816 fprintf(file
, ":%s", brw_reg_type_letters(inst
->src
[i
].type
));
4819 if (i
< inst
->sources
- 1 && inst
->src
[i
+ 1].file
!= BAD_FILE
)
4820 fprintf(file
, ", ");
4825 if (inst
->force_writemask_all
)
4826 fprintf(file
, "NoMask ");
4828 if (dispatch_width
== 16 && inst
->exec_size
== 8) {
4829 if (inst
->force_sechalf
)
4830 fprintf(file
, "2ndhalf ");
4832 fprintf(file
, "1sthalf ");
4835 fprintf(file
, "\n");
4839 * Possibly returns an instruction that set up @param reg.
4841 * Sometimes we want to take the result of some expression/variable
4842 * dereference tree and rewrite the instruction generating the result
4843 * of the tree. When processing the tree, we know that the
4844 * instructions generated are all writing temporaries that are dead
4845 * outside of this tree. So, if we have some instructions that write
4846 * a temporary, we're free to point that temp write somewhere else.
4848 * Note that this doesn't guarantee that the instruction generated
4849 * only reg -- it might be the size=4 destination of a texture instruction.
4852 fs_visitor::get_instruction_generating_reg(fs_inst
*start
,
4857 end
->is_partial_write() ||
4859 !reg
.equals(end
->dst
)) {
4867 fs_visitor::setup_payload_gen6()
4870 (nir
->info
.inputs_read
& (1 << VARYING_SLOT_POS
)) != 0;
4871 unsigned barycentric_interp_modes
=
4872 (stage
== MESA_SHADER_FRAGMENT
) ?
4873 ((brw_wm_prog_data
*) this->prog_data
)->barycentric_interp_modes
: 0;
4875 assert(devinfo
->gen
>= 6);
4877 /* R0-1: masks, pixel X/Y coordinates. */
4878 payload
.num_regs
= 2;
4879 /* R2: only for 32-pixel dispatch.*/
4881 /* R3-26: barycentric interpolation coordinates. These appear in the
4882 * same order that they appear in the brw_wm_barycentric_interp_mode
4883 * enum. Each set of coordinates occupies 2 registers if dispatch width
4884 * == 8 and 4 registers if dispatch width == 16. Coordinates only
4885 * appear if they were enabled using the "Barycentric Interpolation
4886 * Mode" bits in WM_STATE.
4888 for (int i
= 0; i
< BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT
; ++i
) {
4889 if (barycentric_interp_modes
& (1 << i
)) {
4890 payload
.barycentric_coord_reg
[i
] = payload
.num_regs
;
4891 payload
.num_regs
+= 2;
4892 if (dispatch_width
== 16) {
4893 payload
.num_regs
+= 2;
4898 /* R27: interpolated depth if uses source depth */
4900 payload
.source_depth_reg
= payload
.num_regs
;
4902 if (dispatch_width
== 16) {
4903 /* R28: interpolated depth if not SIMD8. */
4907 /* R29: interpolated W set if GEN6_WM_USES_SOURCE_W. */
4909 payload
.source_w_reg
= payload
.num_regs
;
4911 if (dispatch_width
== 16) {
4912 /* R30: interpolated W if not SIMD8. */
4917 if (stage
== MESA_SHADER_FRAGMENT
) {
4918 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
4919 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
4920 prog_data
->uses_pos_offset
= key
->compute_pos_offset
;
4921 /* R31: MSAA position offsets. */
4922 if (prog_data
->uses_pos_offset
) {
4923 payload
.sample_pos_reg
= payload
.num_regs
;
4928 /* R32: MSAA input coverage mask */
4929 if (nir
->info
.system_values_read
& SYSTEM_BIT_SAMPLE_MASK_IN
) {
4930 assert(devinfo
->gen
>= 7);
4931 payload
.sample_mask_in_reg
= payload
.num_regs
;
4933 if (dispatch_width
== 16) {
4934 /* R33: input coverage mask if not SIMD8. */
4939 /* R34-: bary for 32-pixel. */
4940 /* R58-59: interp W for 32-pixel. */
4942 if (nir
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_DEPTH
)) {
4943 source_depth_to_render_target
= true;
4948 fs_visitor::setup_vs_payload()
4950 /* R0: thread header, R1: urb handles */
4951 payload
.num_regs
= 2;
4955 * We are building the local ID push constant data using the simplest possible
4956 * method. We simply push the local IDs directly as they should appear in the
4957 * registers for the uvec3 gl_LocalInvocationID variable.
4959 * Therefore, for SIMD8, we use 3 full registers, and for SIMD16 we use 6
4960 * registers worth of push constant space.
4962 * Note: Any updates to brw_cs_prog_local_id_payload_dwords,
4963 * fill_local_id_payload or fs_visitor::emit_cs_local_invocation_id_setup need
4966 * FINISHME: There are a few easy optimizations to consider.
4968 * 1. If gl_WorkGroupSize x, y or z is 1, we can just use zero, and there is
4969 * no need for using push constant space for that dimension.
4971 * 2. Since GL_MAX_COMPUTE_WORK_GROUP_SIZE is currently 1024 or less, we can
4972 * easily use 16-bit words rather than 32-bit dwords in the push constant
4975 * 3. If gl_WorkGroupSize x, y or z is small, then we can use bytes for
4976 * conveying the data, and thereby reduce push constant usage.
4980 fs_visitor::setup_gs_payload()
4982 assert(stage
== MESA_SHADER_GEOMETRY
);
4984 struct brw_gs_prog_data
*gs_prog_data
=
4985 (struct brw_gs_prog_data
*) prog_data
;
4986 struct brw_vue_prog_data
*vue_prog_data
=
4987 (struct brw_vue_prog_data
*) prog_data
;
4989 /* R0: thread header, R1: output URB handles */
4990 payload
.num_regs
= 2;
4992 if (gs_prog_data
->include_primitive_id
) {
4993 /* R2: Primitive ID 0..7 */
4997 /* Use a maximum of 32 registers for push-model inputs. */
4998 const unsigned max_push_components
= 32;
5000 /* If pushing our inputs would take too many registers, reduce the URB read
5001 * length (which is in HWords, or 8 registers), and resort to pulling.
5003 * Note that the GS reads <URB Read Length> HWords for every vertex - so we
5004 * have to multiply by VerticesIn to obtain the total storage requirement.
5006 if (8 * vue_prog_data
->urb_read_length
* nir
->info
.gs
.vertices_in
>
5007 max_push_components
) {
5008 gs_prog_data
->base
.include_vue_handles
= true;
5010 /* R3..RN: ICP Handles for each incoming vertex (when using pull model) */
5011 payload
.num_regs
+= nir
->info
.gs
.vertices_in
;
5013 vue_prog_data
->urb_read_length
=
5014 ROUND_DOWN_TO(max_push_components
/ nir
->info
.gs
.vertices_in
, 8) / 8;
5019 fs_visitor::setup_cs_payload()
5021 assert(devinfo
->gen
>= 7);
5022 brw_cs_prog_data
*prog_data
= (brw_cs_prog_data
*) this->prog_data
;
5024 payload
.num_regs
= 1;
5026 if (nir
->info
.system_values_read
& SYSTEM_BIT_LOCAL_INVOCATION_ID
) {
5027 prog_data
->local_invocation_id_regs
= dispatch_width
* 3 / 8;
5028 payload
.local_invocation_id_reg
= payload
.num_regs
;
5029 payload
.num_regs
+= prog_data
->local_invocation_id_regs
;
5034 fs_visitor::calculate_register_pressure()
5036 invalidate_live_intervals();
5037 calculate_live_intervals();
5039 unsigned num_instructions
= 0;
5040 foreach_block(block
, cfg
)
5041 num_instructions
+= block
->instructions
.length();
5043 regs_live_at_ip
= rzalloc_array(mem_ctx
, int, num_instructions
);
5045 for (unsigned reg
= 0; reg
< alloc
.count
; reg
++) {
5046 for (int ip
= virtual_grf_start
[reg
]; ip
<= virtual_grf_end
[reg
]; ip
++)
5047 regs_live_at_ip
[ip
] += alloc
.sizes
[reg
];
5052 fs_visitor::optimize()
5054 /* Start by validating the shader we currently have. */
5057 /* bld is the common builder object pointing at the end of the program we
5058 * used to translate it into i965 IR. For the optimization and lowering
5059 * passes coming next, any code added after the end of the program without
5060 * having explicitly called fs_builder::at() clearly points at a mistake.
5061 * Ideally optimization passes wouldn't be part of the visitor so they
5062 * wouldn't have access to bld at all, but they do, so just in case some
5063 * pass forgets to ask for a location explicitly set it to NULL here to
5064 * make it trip. The dispatch width is initialized to a bogus value to
5065 * make sure that optimizations set the execution controls explicitly to
5066 * match the code they are manipulating instead of relying on the defaults.
5068 bld
= fs_builder(this, 64);
5070 assign_constant_locations();
5071 demote_pull_constants();
5075 split_virtual_grfs();
5078 #define OPT(pass, args...) ({ \
5080 bool this_progress = pass(args); \
5082 if (unlikely(INTEL_DEBUG & DEBUG_OPTIMIZER) && this_progress) { \
5083 char filename[64]; \
5084 snprintf(filename, 64, "%s%d-%s-%02d-%02d-" #pass, \
5085 stage_abbrev, dispatch_width, nir->info.name, iteration, pass_num); \
5087 backend_shader::dump_instructions(filename); \
5092 progress = progress || this_progress; \
5096 if (unlikely(INTEL_DEBUG
& DEBUG_OPTIMIZER
)) {
5098 snprintf(filename
, 64, "%s%d-%s-00-start",
5099 stage_abbrev
, dispatch_width
, nir
->info
.name
);
5101 backend_shader::dump_instructions(filename
);
5104 bool progress
= false;
5108 OPT(lower_simd_width
);
5109 OPT(lower_logical_sends
);
5116 OPT(remove_duplicate_mrf_writes
);
5120 OPT(opt_copy_propagate
);
5121 OPT(opt_predicated_break
, this);
5122 OPT(opt_cmod_propagation
);
5123 OPT(dead_code_eliminate
);
5124 OPT(opt_peephole_sel
);
5125 OPT(dead_control_flow_eliminate
, this);
5126 OPT(opt_register_renaming
);
5127 OPT(opt_redundant_discard_jumps
);
5128 OPT(opt_saturate_propagation
);
5129 OPT(opt_zero_samples
);
5130 OPT(register_coalesce
);
5131 OPT(compute_to_mrf
);
5132 OPT(eliminate_find_live_channel
);
5134 OPT(compact_virtual_grfs
);
5139 OPT(opt_sampler_eot
);
5141 if (OPT(lower_load_payload
)) {
5142 split_virtual_grfs();
5143 OPT(register_coalesce
);
5144 OPT(compute_to_mrf
);
5145 OPT(dead_code_eliminate
);
5148 OPT(opt_combine_constants
);
5149 OPT(lower_integer_multiplication
);
5151 lower_uniform_pull_constant_loads();
5157 * Three source instruction must have a GRF/MRF destination register.
5158 * ARF NULL is not allowed. Fix that up by allocating a temporary GRF.
5161 fs_visitor::fixup_3src_null_dest()
5163 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
5164 if (inst
->is_3src() && inst
->dst
.is_null()) {
5165 inst
->dst
= fs_reg(VGRF
, alloc
.allocate(dispatch_width
/ 8),
5172 fs_visitor::allocate_registers()
5174 bool allocated_without_spills
;
5176 static const enum instruction_scheduler_mode pre_modes
[] = {
5178 SCHEDULE_PRE_NON_LIFO
,
5182 /* Try each scheduling heuristic to see if it can successfully register
5183 * allocate without spilling. They should be ordered by decreasing
5184 * performance but increasing likelihood of allocating.
5186 for (unsigned i
= 0; i
< ARRAY_SIZE(pre_modes
); i
++) {
5187 schedule_instructions(pre_modes
[i
]);
5190 assign_regs_trivial();
5191 allocated_without_spills
= true;
5193 allocated_without_spills
= assign_regs(false);
5195 if (allocated_without_spills
)
5199 if (!allocated_without_spills
) {
5200 /* We assume that any spilling is worse than just dropping back to
5201 * SIMD8. There's probably actually some intermediate point where
5202 * SIMD16 with a couple of spills is still better.
5204 if (dispatch_width
== 16) {
5205 fail("Failure to register allocate. Reduce number of "
5206 "live scalar values to avoid this.");
5208 compiler
->shader_perf_log(log_data
,
5209 "%s shader triggered register spilling. "
5210 "Try reducing the number of live scalar "
5211 "values to improve performance.\n",
5215 /* Since we're out of heuristics, just go spill registers until we
5216 * get an allocation.
5218 while (!assign_regs(true)) {
5224 /* This must come after all optimization and register allocation, since
5225 * it inserts dead code that happens to have side effects, and it does
5226 * so based on the actual physical registers in use.
5228 insert_gen4_send_dependency_workarounds();
5233 schedule_instructions(SCHEDULE_POST
);
5235 if (last_scratch
> 0)
5236 prog_data
->total_scratch
= brw_get_scratch_size(last_scratch
);
5240 fs_visitor::run_vs(gl_clip_plane
*clip_planes
)
5242 assert(stage
== MESA_SHADER_VERTEX
);
5246 if (shader_time_index
>= 0)
5247 emit_shader_time_begin();
5254 compute_clip_distance(clip_planes
);
5258 if (shader_time_index
>= 0)
5259 emit_shader_time_end();
5265 assign_curb_setup();
5266 assign_vs_urb_setup();
5268 fixup_3src_null_dest();
5269 allocate_registers();
5275 fs_visitor::run_tes()
5277 assert(stage
== MESA_SHADER_TESS_EVAL
);
5279 /* R0: thread header, R1-3: gl_TessCoord.xyz, R4: URB handles */
5280 payload
.num_regs
= 5;
5282 if (shader_time_index
>= 0)
5283 emit_shader_time_begin();
5292 if (shader_time_index
>= 0)
5293 emit_shader_time_end();
5299 assign_curb_setup();
5300 assign_tes_urb_setup();
5302 fixup_3src_null_dest();
5303 allocate_registers();
5309 fs_visitor::run_gs()
5311 assert(stage
== MESA_SHADER_GEOMETRY
);
5315 this->final_gs_vertex_count
= vgrf(glsl_type::uint_type
);
5317 if (gs_compile
->control_data_header_size_bits
> 0) {
5318 /* Create a VGRF to store accumulated control data bits. */
5319 this->control_data_bits
= vgrf(glsl_type::uint_type
);
5321 /* If we're outputting more than 32 control data bits, then EmitVertex()
5322 * will set control_data_bits to 0 after emitting the first vertex.
5323 * Otherwise, we need to initialize it to 0 here.
5325 if (gs_compile
->control_data_header_size_bits
<= 32) {
5326 const fs_builder abld
= bld
.annotate("initialize control data bits");
5327 abld
.MOV(this->control_data_bits
, brw_imm_ud(0u));
5331 if (shader_time_index
>= 0)
5332 emit_shader_time_begin();
5336 emit_gs_thread_end();
5338 if (shader_time_index
>= 0)
5339 emit_shader_time_end();
5348 assign_curb_setup();
5349 assign_gs_urb_setup();
5351 fixup_3src_null_dest();
5352 allocate_registers();
5358 fs_visitor::run_fs(bool do_rep_send
)
5360 brw_wm_prog_data
*wm_prog_data
= (brw_wm_prog_data
*) this->prog_data
;
5361 brw_wm_prog_key
*wm_key
= (brw_wm_prog_key
*) this->key
;
5363 assert(stage
== MESA_SHADER_FRAGMENT
);
5365 if (devinfo
->gen
>= 6)
5366 setup_payload_gen6();
5368 setup_payload_gen4();
5372 } else if (do_rep_send
) {
5373 assert(dispatch_width
== 16);
5374 emit_repclear_shader();
5376 if (shader_time_index
>= 0)
5377 emit_shader_time_begin();
5379 calculate_urb_setup();
5380 if (nir
->info
.inputs_read
> 0) {
5381 if (devinfo
->gen
< 6)
5382 emit_interpolation_setup_gen4();
5384 emit_interpolation_setup_gen6();
5387 /* We handle discards by keeping track of the still-live pixels in f0.1.
5388 * Initialize it with the dispatched pixels.
5390 if (wm_prog_data
->uses_kill
) {
5391 fs_inst
*discard_init
= bld
.emit(FS_OPCODE_MOV_DISPATCH_TO_FLAGS
);
5392 discard_init
->flag_subreg
= 1;
5395 /* Generate FS IR for main(). (the visitor only descends into
5396 * functions called "main").
5403 if (wm_prog_data
->uses_kill
)
5404 bld
.emit(FS_OPCODE_PLACEHOLDER_HALT
);
5406 if (wm_key
->alpha_test_func
)
5411 if (shader_time_index
>= 0)
5412 emit_shader_time_end();
5418 assign_curb_setup();
5421 fixup_3src_null_dest();
5422 allocate_registers();
5428 if (dispatch_width
== 8)
5429 wm_prog_data
->reg_blocks
= brw_register_blocks(grf_used
);
5431 wm_prog_data
->reg_blocks_16
= brw_register_blocks(grf_used
);
5437 fs_visitor::run_cs()
5439 assert(stage
== MESA_SHADER_COMPUTE
);
5443 if (shader_time_index
>= 0)
5444 emit_shader_time_begin();
5451 emit_cs_terminate();
5453 if (shader_time_index
>= 0)
5454 emit_shader_time_end();
5460 assign_curb_setup();
5462 fixup_3src_null_dest();
5463 allocate_registers();
5472 * Return a bitfield where bit n is set if barycentric interpolation mode n
5473 * (see enum brw_wm_barycentric_interp_mode) is needed by the fragment shader.
5476 brw_compute_barycentric_interp_modes(const struct brw_device_info
*devinfo
,
5477 bool shade_model_flat
,
5478 bool persample_shading
,
5479 const nir_shader
*shader
)
5481 unsigned barycentric_interp_modes
= 0;
5483 nir_foreach_variable(var
, &shader
->inputs
) {
5484 enum glsl_interp_qualifier interp_qualifier
=
5485 (enum glsl_interp_qualifier
)var
->data
.interpolation
;
5486 bool is_centroid
= var
->data
.centroid
&& !persample_shading
;
5487 bool is_sample
= var
->data
.sample
|| persample_shading
;
5488 bool is_gl_Color
= (var
->data
.location
== VARYING_SLOT_COL0
) ||
5489 (var
->data
.location
== VARYING_SLOT_COL1
);
5491 /* Ignore WPOS and FACE, because they don't require interpolation. */
5492 if (var
->data
.location
== VARYING_SLOT_POS
||
5493 var
->data
.location
== VARYING_SLOT_FACE
)
5496 /* Determine the set (or sets) of barycentric coordinates needed to
5497 * interpolate this variable. Note that when
5498 * brw->needs_unlit_centroid_workaround is set, centroid interpolation
5499 * uses PIXEL interpolation for unlit pixels and CENTROID interpolation
5500 * for lit pixels, so we need both sets of barycentric coordinates.
5502 if (interp_qualifier
== INTERP_QUALIFIER_NOPERSPECTIVE
) {
5504 barycentric_interp_modes
|=
5505 1 << BRW_WM_NONPERSPECTIVE_CENTROID_BARYCENTRIC
;
5506 } else if (is_sample
) {
5507 barycentric_interp_modes
|=
5508 1 << BRW_WM_NONPERSPECTIVE_SAMPLE_BARYCENTRIC
;
5510 if ((!is_centroid
&& !is_sample
) ||
5511 devinfo
->needs_unlit_centroid_workaround
) {
5512 barycentric_interp_modes
|=
5513 1 << BRW_WM_NONPERSPECTIVE_PIXEL_BARYCENTRIC
;
5515 } else if (interp_qualifier
== INTERP_QUALIFIER_SMOOTH
||
5516 (!(shade_model_flat
&& is_gl_Color
) &&
5517 interp_qualifier
== INTERP_QUALIFIER_NONE
)) {
5519 barycentric_interp_modes
|=
5520 1 << BRW_WM_PERSPECTIVE_CENTROID_BARYCENTRIC
;
5521 } else if (is_sample
) {
5522 barycentric_interp_modes
|=
5523 1 << BRW_WM_PERSPECTIVE_SAMPLE_BARYCENTRIC
;
5525 if ((!is_centroid
&& !is_sample
) ||
5526 devinfo
->needs_unlit_centroid_workaround
) {
5527 barycentric_interp_modes
|=
5528 1 << BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
;
5533 return barycentric_interp_modes
;
5537 computed_depth_mode(const nir_shader
*shader
)
5539 if (shader
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_DEPTH
)) {
5540 switch (shader
->info
.fs
.depth_layout
) {
5541 case FRAG_DEPTH_LAYOUT_NONE
:
5542 case FRAG_DEPTH_LAYOUT_ANY
:
5543 return BRW_PSCDEPTH_ON
;
5544 case FRAG_DEPTH_LAYOUT_GREATER
:
5545 return BRW_PSCDEPTH_ON_GE
;
5546 case FRAG_DEPTH_LAYOUT_LESS
:
5547 return BRW_PSCDEPTH_ON_LE
;
5548 case FRAG_DEPTH_LAYOUT_UNCHANGED
:
5549 return BRW_PSCDEPTH_OFF
;
5552 return BRW_PSCDEPTH_OFF
;
5556 brw_compile_fs(const struct brw_compiler
*compiler
, void *log_data
,
5558 const struct brw_wm_prog_key
*key
,
5559 struct brw_wm_prog_data
*prog_data
,
5560 const nir_shader
*src_shader
,
5561 struct gl_program
*prog
,
5562 int shader_time_index8
, int shader_time_index16
,
5564 unsigned *final_assembly_size
,
5567 nir_shader
*shader
= nir_shader_clone(mem_ctx
, src_shader
);
5568 shader
= brw_nir_apply_sampler_key(shader
, compiler
->devinfo
, &key
->tex
,
5570 shader
= brw_postprocess_nir(shader
, compiler
->devinfo
, true);
5572 /* key->alpha_test_func means simulating alpha testing via discards,
5573 * so the shader definitely kills pixels.
5575 prog_data
->uses_kill
= shader
->info
.fs
.uses_discard
|| key
->alpha_test_func
;
5576 prog_data
->uses_omask
=
5577 shader
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_SAMPLE_MASK
);
5578 prog_data
->computed_depth_mode
= computed_depth_mode(shader
);
5579 prog_data
->computed_stencil
=
5580 shader
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_STENCIL
);
5582 prog_data
->early_fragment_tests
= shader
->info
.fs
.early_fragment_tests
;
5584 prog_data
->barycentric_interp_modes
=
5585 brw_compute_barycentric_interp_modes(compiler
->devinfo
,
5587 key
->persample_shading
,
5590 fs_visitor
v(compiler
, log_data
, mem_ctx
, key
,
5591 &prog_data
->base
, prog
, shader
, 8,
5592 shader_time_index8
);
5593 if (!v
.run_fs(false /* do_rep_send */)) {
5595 *error_str
= ralloc_strdup(mem_ctx
, v
.fail_msg
);
5600 cfg_t
*simd16_cfg
= NULL
;
5601 fs_visitor
v2(compiler
, log_data
, mem_ctx
, key
,
5602 &prog_data
->base
, prog
, shader
, 16,
5603 shader_time_index16
);
5604 if (likely(!(INTEL_DEBUG
& DEBUG_NO16
) || use_rep_send
)) {
5605 if (!v
.simd16_unsupported
) {
5606 /* Try a SIMD16 compile */
5607 v2
.import_uniforms(&v
);
5608 if (!v2
.run_fs(use_rep_send
)) {
5609 compiler
->shader_perf_log(log_data
,
5610 "SIMD16 shader failed to compile: %s",
5613 simd16_cfg
= v2
.cfg
;
5619 int no_simd8
= (INTEL_DEBUG
& DEBUG_NO8
) || use_rep_send
;
5620 if ((no_simd8
|| compiler
->devinfo
->gen
< 5) && simd16_cfg
) {
5622 prog_data
->no_8
= true;
5625 prog_data
->no_8
= false;
5628 fs_generator
g(compiler
, log_data
, mem_ctx
, (void *) key
, &prog_data
->base
,
5629 v
.promoted_constants
, v
.runtime_check_aads_emit
, "FS");
5631 if (unlikely(INTEL_DEBUG
& DEBUG_WM
)) {
5632 g
.enable_debug(ralloc_asprintf(mem_ctx
, "%s fragment shader %s",
5633 shader
->info
.label
? shader
->info
.label
:
5635 shader
->info
.name
));
5639 g
.generate_code(simd8_cfg
, 8);
5641 prog_data
->prog_offset_16
= g
.generate_code(simd16_cfg
, 16);
5643 return g
.get_assembly(final_assembly_size
);
5647 fs_visitor::emit_cs_local_invocation_id_setup()
5649 assert(stage
== MESA_SHADER_COMPUTE
);
5651 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::uvec3_type
));
5653 struct brw_reg src
=
5654 brw_vec8_grf(payload
.local_invocation_id_reg
, 0);
5655 src
= retype(src
, BRW_REGISTER_TYPE_UD
);
5657 src
.nr
+= dispatch_width
/ 8;
5658 bld
.MOV(offset(*reg
, bld
, 1), src
);
5659 src
.nr
+= dispatch_width
/ 8;
5660 bld
.MOV(offset(*reg
, bld
, 2), src
);
5666 fs_visitor::emit_cs_work_group_id_setup()
5668 assert(stage
== MESA_SHADER_COMPUTE
);
5670 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::uvec3_type
));
5672 struct brw_reg
r0_1(retype(brw_vec1_grf(0, 1), BRW_REGISTER_TYPE_UD
));
5673 struct brw_reg
r0_6(retype(brw_vec1_grf(0, 6), BRW_REGISTER_TYPE_UD
));
5674 struct brw_reg
r0_7(retype(brw_vec1_grf(0, 7), BRW_REGISTER_TYPE_UD
));
5676 bld
.MOV(*reg
, r0_1
);
5677 bld
.MOV(offset(*reg
, bld
, 1), r0_6
);
5678 bld
.MOV(offset(*reg
, bld
, 2), r0_7
);
5684 brw_compile_cs(const struct brw_compiler
*compiler
, void *log_data
,
5686 const struct brw_cs_prog_key
*key
,
5687 struct brw_cs_prog_data
*prog_data
,
5688 const nir_shader
*src_shader
,
5689 int shader_time_index
,
5690 unsigned *final_assembly_size
,
5693 nir_shader
*shader
= nir_shader_clone(mem_ctx
, src_shader
);
5694 shader
= brw_nir_apply_sampler_key(shader
, compiler
->devinfo
, &key
->tex
,
5696 shader
= brw_postprocess_nir(shader
, compiler
->devinfo
, true);
5698 prog_data
->local_size
[0] = shader
->info
.cs
.local_size
[0];
5699 prog_data
->local_size
[1] = shader
->info
.cs
.local_size
[1];
5700 prog_data
->local_size
[2] = shader
->info
.cs
.local_size
[2];
5701 unsigned local_workgroup_size
=
5702 shader
->info
.cs
.local_size
[0] * shader
->info
.cs
.local_size
[1] *
5703 shader
->info
.cs
.local_size
[2];
5705 unsigned max_cs_threads
= compiler
->devinfo
->max_cs_threads
;
5708 const char *fail_msg
= NULL
;
5710 /* Now the main event: Visit the shader IR and generate our CS IR for it.
5712 fs_visitor
v8(compiler
, log_data
, mem_ctx
, key
, &prog_data
->base
,
5713 NULL
, /* Never used in core profile */
5714 shader
, 8, shader_time_index
);
5716 fail_msg
= v8
.fail_msg
;
5717 } else if (local_workgroup_size
<= 8 * max_cs_threads
) {
5719 prog_data
->simd_size
= 8;
5722 fs_visitor
v16(compiler
, log_data
, mem_ctx
, key
, &prog_data
->base
,
5723 NULL
, /* Never used in core profile */
5724 shader
, 16, shader_time_index
);
5725 if (likely(!(INTEL_DEBUG
& DEBUG_NO16
)) &&
5726 !fail_msg
&& !v8
.simd16_unsupported
&&
5727 local_workgroup_size
<= 16 * max_cs_threads
) {
5728 /* Try a SIMD16 compile */
5729 v16
.import_uniforms(&v8
);
5730 if (!v16
.run_cs()) {
5731 compiler
->shader_perf_log(log_data
,
5732 "SIMD16 shader failed to compile: %s",
5736 "Couldn't generate SIMD16 program and not "
5737 "enough threads for SIMD8";
5741 prog_data
->simd_size
= 16;
5745 if (unlikely(cfg
== NULL
)) {
5748 *error_str
= ralloc_strdup(mem_ctx
, fail_msg
);
5753 fs_generator
g(compiler
, log_data
, mem_ctx
, (void*) key
, &prog_data
->base
,
5754 v8
.promoted_constants
, v8
.runtime_check_aads_emit
, "CS");
5755 if (INTEL_DEBUG
& DEBUG_CS
) {
5756 char *name
= ralloc_asprintf(mem_ctx
, "%s compute shader %s",
5757 shader
->info
.label
? shader
->info
.label
:
5760 g
.enable_debug(name
);
5763 g
.generate_code(cfg
, prog_data
->simd_size
);
5765 return g
.get_assembly(final_assembly_size
);
5769 brw_cs_fill_local_id_payload(const struct brw_cs_prog_data
*prog_data
,
5770 void *buffer
, uint32_t threads
, uint32_t stride
)
5772 if (prog_data
->local_invocation_id_regs
== 0)
5775 /* 'stride' should be an integer number of registers, that is, a multiple
5778 assert(stride
% 32 == 0);
5780 unsigned x
= 0, y
= 0, z
= 0;
5781 for (unsigned t
= 0; t
< threads
; t
++) {
5782 uint32_t *param
= (uint32_t *) buffer
+ stride
* t
/ 4;
5784 for (unsigned i
= 0; i
< prog_data
->simd_size
; i
++) {
5785 param
[0 * prog_data
->simd_size
+ i
] = x
;
5786 param
[1 * prog_data
->simd_size
+ i
] = y
;
5787 param
[2 * prog_data
->simd_size
+ i
] = z
;
5790 if (x
== prog_data
->local_size
[0]) {
5793 if (y
== prog_data
->local_size
[1]) {
5796 if (z
== prog_data
->local_size
[2])