2 * Copyright © 2010 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
26 * This file drives the GLSL IR -> LIR translation, contains the
27 * optimizations on the LIR, and drives the generation of native code
31 #include "main/macros.h"
35 #include "brw_vec4_gs_visitor.h"
37 #include "brw_dead_control_flow.h"
38 #include "dev/gen_debug.h"
39 #include "compiler/glsl_types.h"
40 #include "compiler/nir/nir_builder.h"
41 #include "program/prog_parameter.h"
42 #include "util/u_math.h"
46 static unsigned get_lowered_simd_width(const struct gen_device_info
*devinfo
,
50 fs_inst::init(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
51 const fs_reg
*src
, unsigned sources
)
53 memset((void*)this, 0, sizeof(*this));
55 this->src
= new fs_reg
[MAX2(sources
, 3)];
56 for (unsigned i
= 0; i
< sources
; i
++)
57 this->src
[i
] = src
[i
];
59 this->opcode
= opcode
;
61 this->sources
= sources
;
62 this->exec_size
= exec_size
;
65 assert(dst
.file
!= IMM
&& dst
.file
!= UNIFORM
);
67 assert(this->exec_size
!= 0);
69 this->conditional_mod
= BRW_CONDITIONAL_NONE
;
71 /* This will be the case for almost all instructions. */
78 this->size_written
= dst
.component_size(exec_size
);
81 this->size_written
= 0;
85 unreachable("Invalid destination register file");
88 this->writes_accumulator
= false;
93 init(BRW_OPCODE_NOP
, 8, dst
, NULL
, 0);
96 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
)
98 init(opcode
, exec_size
, reg_undef
, NULL
, 0);
101 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
)
103 init(opcode
, exec_size
, dst
, NULL
, 0);
106 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
109 const fs_reg src
[1] = { src0
};
110 init(opcode
, exec_size
, dst
, src
, 1);
113 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
114 const fs_reg
&src0
, const fs_reg
&src1
)
116 const fs_reg src
[2] = { src0
, src1
};
117 init(opcode
, exec_size
, dst
, src
, 2);
120 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
121 const fs_reg
&src0
, const fs_reg
&src1
, const fs_reg
&src2
)
123 const fs_reg src
[3] = { src0
, src1
, src2
};
124 init(opcode
, exec_size
, dst
, src
, 3);
127 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_width
, const fs_reg
&dst
,
128 const fs_reg src
[], unsigned sources
)
130 init(opcode
, exec_width
, dst
, src
, sources
);
133 fs_inst::fs_inst(const fs_inst
&that
)
135 memcpy((void*)this, &that
, sizeof(that
));
137 this->src
= new fs_reg
[MAX2(that
.sources
, 3)];
139 for (unsigned i
= 0; i
< that
.sources
; i
++)
140 this->src
[i
] = that
.src
[i
];
149 fs_inst::resize_sources(uint8_t num_sources
)
151 if (this->sources
!= num_sources
) {
152 fs_reg
*src
= new fs_reg
[MAX2(num_sources
, 3)];
154 for (unsigned i
= 0; i
< MIN2(this->sources
, num_sources
); ++i
)
155 src
[i
] = this->src
[i
];
159 this->sources
= num_sources
;
164 fs_visitor::VARYING_PULL_CONSTANT_LOAD(const fs_builder
&bld
,
166 const fs_reg
&surf_index
,
167 const fs_reg
&varying_offset
,
168 uint32_t const_offset
)
170 /* We have our constant surface use a pitch of 4 bytes, so our index can
171 * be any component of a vector, and then we load 4 contiguous
172 * components starting from that.
174 * We break down the const_offset to a portion added to the variable offset
175 * and a portion done using fs_reg::offset, which means that if you have
176 * GLSL using something like "uniform vec4 a[20]; gl_FragColor = a[i]",
177 * we'll temporarily generate 4 vec4 loads from offset i * 4, and CSE can
178 * later notice that those loads are all the same and eliminate the
181 fs_reg vec4_offset
= vgrf(glsl_type::uint_type
);
182 bld
.ADD(vec4_offset
, varying_offset
, brw_imm_ud(const_offset
& ~0xf));
184 /* The pull load message will load a vec4 (16 bytes). If we are loading
185 * a double this means we are only loading 2 elements worth of data.
186 * We also want to use a 32-bit data type for the dst of the load operation
187 * so other parts of the driver don't get confused about the size of the
190 fs_reg vec4_result
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 4);
191 fs_inst
*inst
= bld
.emit(FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_LOGICAL
,
192 vec4_result
, surf_index
, vec4_offset
);
193 inst
->size_written
= 4 * vec4_result
.component_size(inst
->exec_size
);
195 shuffle_from_32bit_read(bld
, dst
, vec4_result
,
196 (const_offset
& 0xf) / type_sz(dst
.type
), 1);
200 * A helper for MOV generation for fixing up broken hardware SEND dependency
204 fs_visitor::DEP_RESOLVE_MOV(const fs_builder
&bld
, int grf
)
206 /* The caller always wants uncompressed to emit the minimal extra
207 * dependencies, and to avoid having to deal with aligning its regs to 2.
209 const fs_builder ubld
= bld
.annotate("send dependency resolve")
212 ubld
.MOV(ubld
.null_reg_f(), fs_reg(VGRF
, grf
, BRW_REGISTER_TYPE_F
));
216 fs_inst::is_send_from_grf() const
219 case SHADER_OPCODE_SEND
:
220 case SHADER_OPCODE_SHADER_TIME_ADD
:
221 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
222 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
223 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
224 case SHADER_OPCODE_URB_WRITE_SIMD8
:
225 case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
226 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
:
227 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
228 case SHADER_OPCODE_URB_READ_SIMD8
:
229 case SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
:
230 case SHADER_OPCODE_INTERLOCK
:
231 case SHADER_OPCODE_MEMORY_FENCE
:
232 case SHADER_OPCODE_BARRIER
:
234 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
235 return src
[1].file
== VGRF
;
236 case FS_OPCODE_FB_WRITE
:
237 case FS_OPCODE_FB_READ
:
238 return src
[0].file
== VGRF
;
241 return src
[0].file
== VGRF
;
248 fs_inst::is_control_source(unsigned arg
) const
251 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
252 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7
:
253 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN4
:
256 case SHADER_OPCODE_BROADCAST
:
257 case SHADER_OPCODE_SHUFFLE
:
258 case SHADER_OPCODE_QUAD_SWIZZLE
:
259 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
260 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
261 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
262 case SHADER_OPCODE_GET_BUFFER_SIZE
:
265 case SHADER_OPCODE_MOV_INDIRECT
:
266 case SHADER_OPCODE_CLUSTER_BROADCAST
:
267 case SHADER_OPCODE_TEX
:
269 case SHADER_OPCODE_TXD
:
270 case SHADER_OPCODE_TXF
:
271 case SHADER_OPCODE_TXF_LZ
:
272 case SHADER_OPCODE_TXF_CMS
:
273 case SHADER_OPCODE_TXF_CMS_W
:
274 case SHADER_OPCODE_TXF_UMS
:
275 case SHADER_OPCODE_TXF_MCS
:
276 case SHADER_OPCODE_TXL
:
277 case SHADER_OPCODE_TXL_LZ
:
278 case SHADER_OPCODE_TXS
:
279 case SHADER_OPCODE_LOD
:
280 case SHADER_OPCODE_TG4
:
281 case SHADER_OPCODE_TG4_OFFSET
:
282 case SHADER_OPCODE_SAMPLEINFO
:
283 return arg
== 1 || arg
== 2;
285 case SHADER_OPCODE_SEND
:
286 return arg
== 0 || arg
== 1;
294 fs_inst::is_payload(unsigned arg
) const
297 case FS_OPCODE_FB_WRITE
:
298 case FS_OPCODE_FB_READ
:
299 case SHADER_OPCODE_URB_WRITE_SIMD8
:
300 case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
301 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
:
302 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
303 case SHADER_OPCODE_URB_READ_SIMD8
:
304 case SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
:
305 case VEC4_OPCODE_UNTYPED_ATOMIC
:
306 case VEC4_OPCODE_UNTYPED_SURFACE_READ
:
307 case VEC4_OPCODE_UNTYPED_SURFACE_WRITE
:
308 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
309 case SHADER_OPCODE_SHADER_TIME_ADD
:
310 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
311 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
312 case SHADER_OPCODE_INTERLOCK
:
313 case SHADER_OPCODE_MEMORY_FENCE
:
314 case SHADER_OPCODE_BARRIER
:
317 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7
:
320 case SHADER_OPCODE_SEND
:
321 return arg
== 2 || arg
== 3;
332 * Returns true if this instruction's sources and destinations cannot
333 * safely be the same register.
335 * In most cases, a register can be written over safely by the same
336 * instruction that is its last use. For a single instruction, the
337 * sources are dereferenced before writing of the destination starts
340 * However, there are a few cases where this can be problematic:
342 * - Virtual opcodes that translate to multiple instructions in the
343 * code generator: if src == dst and one instruction writes the
344 * destination before a later instruction reads the source, then
345 * src will have been clobbered.
347 * - SIMD16 compressed instructions with certain regioning (see below).
349 * The register allocator uses this information to set up conflicts between
350 * GRF sources and the destination.
353 fs_inst::has_source_and_destination_hazard() const
356 case FS_OPCODE_PACK_HALF_2x16_SPLIT
:
357 /* Multiple partial writes to the destination */
359 case SHADER_OPCODE_SHUFFLE
:
360 /* This instruction returns an arbitrary channel from the source and
361 * gets split into smaller instructions in the generator. It's possible
362 * that one of the instructions will read from a channel corresponding
363 * to an earlier instruction.
365 case SHADER_OPCODE_SEL_EXEC
:
366 /* This is implemented as
368 * mov(16) g4<1>D 0D { align1 WE_all 1H };
369 * mov(16) g4<1>D g5<8,8,1>D { align1 1H }
371 * Because the source is only read in the second instruction, the first
372 * may stomp all over it.
375 case SHADER_OPCODE_QUAD_SWIZZLE
:
377 case BRW_SWIZZLE_XXXX
:
378 case BRW_SWIZZLE_YYYY
:
379 case BRW_SWIZZLE_ZZZZ
:
380 case BRW_SWIZZLE_WWWW
:
381 case BRW_SWIZZLE_XXZZ
:
382 case BRW_SWIZZLE_YYWW
:
383 case BRW_SWIZZLE_XYXY
:
384 case BRW_SWIZZLE_ZWZW
:
385 /* These can be implemented as a single Align1 region on all
386 * platforms, so there's never a hazard between source and
387 * destination. C.f. fs_generator::generate_quad_swizzle().
391 return !is_uniform(src
[0]);
394 /* The SIMD16 compressed instruction
396 * add(16) g4<1>F g4<8,8,1>F g6<8,8,1>F
398 * is actually decoded in hardware as:
400 * add(8) g4<1>F g4<8,8,1>F g6<8,8,1>F
401 * add(8) g5<1>F g5<8,8,1>F g7<8,8,1>F
403 * Which is safe. However, if we have uniform accesses
404 * happening, we get into trouble:
406 * add(8) g4<1>F g4<0,1,0>F g6<8,8,1>F
407 * add(8) g5<1>F g4<0,1,0>F g7<8,8,1>F
409 * Now our destination for the first instruction overwrote the
410 * second instruction's src0, and we get garbage for those 8
411 * pixels. There's a similar issue for the pre-gen6
412 * pixel_x/pixel_y, which are registers of 16-bit values and thus
413 * would get stomped by the first decode as well.
415 if (exec_size
== 16) {
416 for (int i
= 0; i
< sources
; i
++) {
417 if (src
[i
].file
== VGRF
&& (src
[i
].stride
== 0 ||
418 src
[i
].type
== BRW_REGISTER_TYPE_UW
||
419 src
[i
].type
== BRW_REGISTER_TYPE_W
||
420 src
[i
].type
== BRW_REGISTER_TYPE_UB
||
421 src
[i
].type
== BRW_REGISTER_TYPE_B
)) {
431 fs_inst::is_copy_payload(const brw::simple_allocator
&grf_alloc
) const
433 if (this->opcode
!= SHADER_OPCODE_LOAD_PAYLOAD
)
436 fs_reg reg
= this->src
[0];
437 if (reg
.file
!= VGRF
|| reg
.offset
!= 0 || reg
.stride
!= 1)
440 if (grf_alloc
.sizes
[reg
.nr
] * REG_SIZE
!= this->size_written
)
443 for (int i
= 0; i
< this->sources
; i
++) {
444 reg
.type
= this->src
[i
].type
;
445 if (!this->src
[i
].equals(reg
))
448 if (i
< this->header_size
) {
449 reg
.offset
+= REG_SIZE
;
451 reg
= horiz_offset(reg
, this->exec_size
);
459 fs_inst::can_do_source_mods(const struct gen_device_info
*devinfo
) const
461 if (devinfo
->gen
== 6 && is_math())
464 if (is_send_from_grf())
467 if (!backend_instruction::can_do_source_mods())
474 fs_inst::can_do_cmod()
476 if (!backend_instruction::can_do_cmod())
479 /* The accumulator result appears to get used for the conditional modifier
480 * generation. When negating a UD value, there is a 33rd bit generated for
481 * the sign in the accumulator value, so now you can't check, for example,
482 * equality with a 32-bit value. See piglit fs-op-neg-uvec4.
484 for (unsigned i
= 0; i
< sources
; i
++) {
485 if (type_is_unsigned_int(src
[i
].type
) && src
[i
].negate
)
493 fs_inst::can_change_types() const
495 return dst
.type
== src
[0].type
&&
496 !src
[0].abs
&& !src
[0].negate
&& !saturate
&&
497 (opcode
== BRW_OPCODE_MOV
||
498 (opcode
== BRW_OPCODE_SEL
&&
499 dst
.type
== src
[1].type
&&
500 predicate
!= BRW_PREDICATE_NONE
&&
501 !src
[1].abs
&& !src
[1].negate
));
507 memset((void*)this, 0, sizeof(*this));
508 type
= BRW_REGISTER_TYPE_UD
;
512 /** Generic unset register constructor. */
516 this->file
= BAD_FILE
;
519 fs_reg::fs_reg(struct ::brw_reg reg
) :
524 if (this->file
== IMM
&&
525 (this->type
!= BRW_REGISTER_TYPE_V
&&
526 this->type
!= BRW_REGISTER_TYPE_UV
&&
527 this->type
!= BRW_REGISTER_TYPE_VF
)) {
533 fs_reg::equals(const fs_reg
&r
) const
535 return (this->backend_reg::equals(r
) &&
540 fs_reg::negative_equals(const fs_reg
&r
) const
542 return (this->backend_reg::negative_equals(r
) &&
547 fs_reg::is_contiguous() const
553 fs_reg::component_size(unsigned width
) const
555 const unsigned stride
= ((file
!= ARF
&& file
!= FIXED_GRF
) ? this->stride
:
558 return MAX2(width
* stride
, 1) * type_sz(type
);
562 type_size_scalar(const struct glsl_type
*type
, bool bindless
)
564 unsigned int size
, i
;
566 switch (type
->base_type
) {
569 case GLSL_TYPE_FLOAT
:
571 return type
->components();
572 case GLSL_TYPE_UINT16
:
573 case GLSL_TYPE_INT16
:
574 case GLSL_TYPE_FLOAT16
:
575 return DIV_ROUND_UP(type
->components(), 2);
576 case GLSL_TYPE_UINT8
:
578 return DIV_ROUND_UP(type
->components(), 4);
579 case GLSL_TYPE_DOUBLE
:
580 case GLSL_TYPE_UINT64
:
581 case GLSL_TYPE_INT64
:
582 return type
->components() * 2;
583 case GLSL_TYPE_ARRAY
:
584 return type_size_scalar(type
->fields
.array
, bindless
) * type
->length
;
585 case GLSL_TYPE_STRUCT
:
586 case GLSL_TYPE_INTERFACE
:
588 for (i
= 0; i
< type
->length
; i
++) {
589 size
+= type_size_scalar(type
->fields
.structure
[i
].type
, bindless
);
592 case GLSL_TYPE_SAMPLER
:
593 case GLSL_TYPE_IMAGE
:
595 return type
->components() * 2;
596 case GLSL_TYPE_ATOMIC_UINT
:
597 /* Samplers, atomics, and images take up no register space, since
598 * they're baked in at link time.
601 case GLSL_TYPE_SUBROUTINE
:
604 case GLSL_TYPE_ERROR
:
605 case GLSL_TYPE_FUNCTION
:
606 unreachable("not reached");
613 * Create a MOV to read the timestamp register.
615 * The caller is responsible for emitting the MOV. The return value is
616 * the destination of the MOV, with extra parameters set.
619 fs_visitor::get_timestamp(const fs_builder
&bld
)
621 assert(devinfo
->gen
>= 7);
623 fs_reg ts
= fs_reg(retype(brw_vec4_reg(BRW_ARCHITECTURE_REGISTER_FILE
,
626 BRW_REGISTER_TYPE_UD
));
628 fs_reg dst
= fs_reg(VGRF
, alloc
.allocate(1), BRW_REGISTER_TYPE_UD
);
630 /* We want to read the 3 fields we care about even if it's not enabled in
633 bld
.group(4, 0).exec_all().MOV(dst
, ts
);
639 fs_visitor::emit_shader_time_begin()
641 /* We want only the low 32 bits of the timestamp. Since it's running
642 * at the GPU clock rate of ~1.2ghz, it will roll over every ~3 seconds,
643 * which is plenty of time for our purposes. It is identical across the
644 * EUs, but since it's tracking GPU core speed it will increment at a
645 * varying rate as render P-states change.
647 shader_start_time
= component(
648 get_timestamp(bld
.annotate("shader time start")), 0);
652 fs_visitor::emit_shader_time_end()
654 /* Insert our code just before the final SEND with EOT. */
655 exec_node
*end
= this->instructions
.get_tail();
656 assert(end
&& ((fs_inst
*) end
)->eot
);
657 const fs_builder ibld
= bld
.annotate("shader time end")
658 .exec_all().at(NULL
, end
);
659 const fs_reg timestamp
= get_timestamp(ibld
);
661 /* We only use the low 32 bits of the timestamp - see
662 * emit_shader_time_begin()).
664 * We could also check if render P-states have changed (or anything
665 * else that might disrupt timing) by setting smear to 2 and checking if
666 * that field is != 0.
668 const fs_reg shader_end_time
= component(timestamp
, 0);
670 /* Check that there weren't any timestamp reset events (assuming these
671 * were the only two timestamp reads that happened).
673 const fs_reg reset
= component(timestamp
, 2);
674 set_condmod(BRW_CONDITIONAL_Z
,
675 ibld
.AND(ibld
.null_reg_ud(), reset
, brw_imm_ud(1u)));
676 ibld
.IF(BRW_PREDICATE_NORMAL
);
678 fs_reg start
= shader_start_time
;
680 const fs_reg diff
= component(fs_reg(VGRF
, alloc
.allocate(1),
681 BRW_REGISTER_TYPE_UD
),
683 const fs_builder cbld
= ibld
.group(1, 0);
684 cbld
.group(1, 0).ADD(diff
, start
, shader_end_time
);
686 /* If there were no instructions between the two timestamp gets, the diff
687 * is 2 cycles. Remove that overhead, so I can forget about that when
688 * trying to determine the time taken for single instructions.
690 cbld
.ADD(diff
, diff
, brw_imm_ud(-2u));
691 SHADER_TIME_ADD(cbld
, 0, diff
);
692 SHADER_TIME_ADD(cbld
, 1, brw_imm_ud(1u));
693 ibld
.emit(BRW_OPCODE_ELSE
);
694 SHADER_TIME_ADD(cbld
, 2, brw_imm_ud(1u));
695 ibld
.emit(BRW_OPCODE_ENDIF
);
699 fs_visitor::SHADER_TIME_ADD(const fs_builder
&bld
,
700 int shader_time_subindex
,
703 int index
= shader_time_index
* 3 + shader_time_subindex
;
704 struct brw_reg offset
= brw_imm_d(index
* BRW_SHADER_TIME_STRIDE
);
707 if (dispatch_width
== 8)
708 payload
= vgrf(glsl_type::uvec2_type
);
710 payload
= vgrf(glsl_type::uint_type
);
712 bld
.emit(SHADER_OPCODE_SHADER_TIME_ADD
, fs_reg(), payload
, offset
, value
);
716 fs_visitor::vfail(const char *format
, va_list va
)
725 msg
= ralloc_vasprintf(mem_ctx
, format
, va
);
726 msg
= ralloc_asprintf(mem_ctx
, "%s compile failed: %s\n", stage_abbrev
, msg
);
728 this->fail_msg
= msg
;
731 fprintf(stderr
, "%s", msg
);
736 fs_visitor::fail(const char *format
, ...)
740 va_start(va
, format
);
746 * Mark this program as impossible to compile with dispatch width greater
749 * During the SIMD8 compile (which happens first), we can detect and flag
750 * things that are unsupported in SIMD16+ mode, so the compiler can skip the
751 * SIMD16+ compile altogether.
753 * During a compile of dispatch width greater than n (if one happens anyway),
754 * this just calls fail().
757 fs_visitor::limit_dispatch_width(unsigned n
, const char *msg
)
759 if (dispatch_width
> n
) {
762 max_dispatch_width
= n
;
763 compiler
->shader_perf_log(log_data
,
764 "Shader dispatch width limited to SIMD%d: %s",
770 * Returns true if the instruction has a flag that means it won't
771 * update an entire destination register.
773 * For example, dead code elimination and live variable analysis want to know
774 * when a write to a variable screens off any preceding values that were in
778 fs_inst::is_partial_write() const
780 return ((this->predicate
&& this->opcode
!= BRW_OPCODE_SEL
) ||
781 (this->exec_size
* type_sz(this->dst
.type
)) < 32 ||
782 !this->dst
.is_contiguous() ||
783 this->dst
.offset
% REG_SIZE
!= 0);
787 fs_inst::components_read(unsigned i
) const
789 /* Return zero if the source is not present. */
790 if (src
[i
].file
== BAD_FILE
)
794 case FS_OPCODE_LINTERP
:
800 case FS_OPCODE_PIXEL_X
:
801 case FS_OPCODE_PIXEL_Y
:
805 case FS_OPCODE_FB_WRITE_LOGICAL
:
806 assert(src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].file
== IMM
);
807 /* First/second FB write color. */
809 return src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].ud
;
813 case SHADER_OPCODE_TEX_LOGICAL
:
814 case SHADER_OPCODE_TXD_LOGICAL
:
815 case SHADER_OPCODE_TXF_LOGICAL
:
816 case SHADER_OPCODE_TXL_LOGICAL
:
817 case SHADER_OPCODE_TXS_LOGICAL
:
818 case SHADER_OPCODE_IMAGE_SIZE_LOGICAL
:
819 case FS_OPCODE_TXB_LOGICAL
:
820 case SHADER_OPCODE_TXF_CMS_LOGICAL
:
821 case SHADER_OPCODE_TXF_CMS_W_LOGICAL
:
822 case SHADER_OPCODE_TXF_UMS_LOGICAL
:
823 case SHADER_OPCODE_TXF_MCS_LOGICAL
:
824 case SHADER_OPCODE_LOD_LOGICAL
:
825 case SHADER_OPCODE_TG4_LOGICAL
:
826 case SHADER_OPCODE_TG4_OFFSET_LOGICAL
:
827 case SHADER_OPCODE_SAMPLEINFO_LOGICAL
:
828 assert(src
[TEX_LOGICAL_SRC_COORD_COMPONENTS
].file
== IMM
&&
829 src
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
].file
== IMM
);
830 /* Texture coordinates. */
831 if (i
== TEX_LOGICAL_SRC_COORDINATE
)
832 return src
[TEX_LOGICAL_SRC_COORD_COMPONENTS
].ud
;
833 /* Texture derivatives. */
834 else if ((i
== TEX_LOGICAL_SRC_LOD
|| i
== TEX_LOGICAL_SRC_LOD2
) &&
835 opcode
== SHADER_OPCODE_TXD_LOGICAL
)
836 return src
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
].ud
;
837 /* Texture offset. */
838 else if (i
== TEX_LOGICAL_SRC_TG4_OFFSET
)
841 else if (i
== TEX_LOGICAL_SRC_MCS
&& opcode
== SHADER_OPCODE_TXF_CMS_W_LOGICAL
)
846 case SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
:
847 case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
:
848 assert(src
[SURFACE_LOGICAL_SRC_IMM_DIMS
].file
== IMM
);
849 /* Surface coordinates. */
850 if (i
== SURFACE_LOGICAL_SRC_ADDRESS
)
851 return src
[SURFACE_LOGICAL_SRC_IMM_DIMS
].ud
;
852 /* Surface operation source (ignored for reads). */
853 else if (i
== SURFACE_LOGICAL_SRC_DATA
)
858 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
:
859 case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
:
860 assert(src
[SURFACE_LOGICAL_SRC_IMM_DIMS
].file
== IMM
&&
861 src
[SURFACE_LOGICAL_SRC_IMM_ARG
].file
== IMM
);
862 /* Surface coordinates. */
863 if (i
== SURFACE_LOGICAL_SRC_ADDRESS
)
864 return src
[SURFACE_LOGICAL_SRC_IMM_DIMS
].ud
;
865 /* Surface operation source. */
866 else if (i
== SURFACE_LOGICAL_SRC_DATA
)
867 return src
[SURFACE_LOGICAL_SRC_IMM_ARG
].ud
;
871 case SHADER_OPCODE_A64_UNTYPED_READ_LOGICAL
:
872 assert(src
[2].file
== IMM
);
875 case SHADER_OPCODE_A64_UNTYPED_WRITE_LOGICAL
:
876 assert(src
[2].file
== IMM
);
877 return i
== 1 ? src
[2].ud
: 1;
879 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_LOGICAL
:
880 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_INT64_LOGICAL
:
881 assert(src
[2].file
== IMM
);
884 const unsigned op
= src
[2].ud
;
899 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_FLOAT_LOGICAL
:
900 assert(src
[2].file
== IMM
);
903 const unsigned op
= src
[2].ud
;
904 return op
== BRW_AOP_FCMPWR
? 2 : 1;
909 case SHADER_OPCODE_BYTE_SCATTERED_READ_LOGICAL
:
910 /* Scattered logical opcodes use the following params:
911 * src[0] Surface coordinates
912 * src[1] Surface operation source (ignored for reads)
914 * src[3] IMM with always 1 dimension.
915 * src[4] IMM with arg bitsize for scattered read/write 8, 16, 32
917 assert(src
[SURFACE_LOGICAL_SRC_IMM_DIMS
].file
== IMM
&&
918 src
[SURFACE_LOGICAL_SRC_IMM_ARG
].file
== IMM
);
919 return i
== SURFACE_LOGICAL_SRC_DATA
? 0 : 1;
921 case SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL
:
922 assert(src
[SURFACE_LOGICAL_SRC_IMM_DIMS
].file
== IMM
&&
923 src
[SURFACE_LOGICAL_SRC_IMM_ARG
].file
== IMM
);
926 case SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL
:
927 case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
: {
928 assert(src
[SURFACE_LOGICAL_SRC_IMM_DIMS
].file
== IMM
&&
929 src
[SURFACE_LOGICAL_SRC_IMM_ARG
].file
== IMM
);
930 const unsigned op
= src
[SURFACE_LOGICAL_SRC_IMM_ARG
].ud
;
931 /* Surface coordinates. */
932 if (i
== SURFACE_LOGICAL_SRC_ADDRESS
)
933 return src
[SURFACE_LOGICAL_SRC_IMM_DIMS
].ud
;
934 /* Surface operation source. */
935 else if (i
== SURFACE_LOGICAL_SRC_DATA
&& op
== BRW_AOP_CMPWR
)
937 else if (i
== SURFACE_LOGICAL_SRC_DATA
&&
938 (op
== BRW_AOP_INC
|| op
== BRW_AOP_DEC
|| op
== BRW_AOP_PREDEC
))
943 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
944 return (i
== 0 ? 2 : 1);
946 case SHADER_OPCODE_UNTYPED_ATOMIC_FLOAT_LOGICAL
: {
947 assert(src
[SURFACE_LOGICAL_SRC_IMM_DIMS
].file
== IMM
&&
948 src
[SURFACE_LOGICAL_SRC_IMM_ARG
].file
== IMM
);
949 const unsigned op
= src
[SURFACE_LOGICAL_SRC_IMM_ARG
].ud
;
950 /* Surface coordinates. */
951 if (i
== SURFACE_LOGICAL_SRC_ADDRESS
)
952 return src
[SURFACE_LOGICAL_SRC_IMM_DIMS
].ud
;
953 /* Surface operation source. */
954 else if (i
== SURFACE_LOGICAL_SRC_DATA
&& op
== BRW_AOP_FCMPWR
)
966 fs_inst::size_read(int arg
) const
969 case SHADER_OPCODE_SEND
:
971 return mlen
* REG_SIZE
;
972 } else if (arg
== 3) {
973 return ex_mlen
* REG_SIZE
;
977 case FS_OPCODE_FB_WRITE
:
978 case FS_OPCODE_REP_FB_WRITE
:
981 return src
[0].file
== BAD_FILE
? 0 : 2 * REG_SIZE
;
983 return mlen
* REG_SIZE
;
987 case FS_OPCODE_FB_READ
:
988 case SHADER_OPCODE_URB_WRITE_SIMD8
:
989 case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
990 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
:
991 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
992 case SHADER_OPCODE_URB_READ_SIMD8
:
993 case SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
:
994 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
995 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
997 return mlen
* REG_SIZE
;
1000 case FS_OPCODE_SET_SAMPLE_ID
:
1005 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7
:
1006 /* The payload is actually stored in src1 */
1008 return mlen
* REG_SIZE
;
1011 case FS_OPCODE_LINTERP
:
1016 case SHADER_OPCODE_LOAD_PAYLOAD
:
1017 if (arg
< this->header_size
)
1021 case CS_OPCODE_CS_TERMINATE
:
1022 case SHADER_OPCODE_BARRIER
:
1025 case SHADER_OPCODE_MOV_INDIRECT
:
1027 assert(src
[2].file
== IMM
);
1033 if (is_tex() && arg
== 0 && src
[0].file
== VGRF
)
1034 return mlen
* REG_SIZE
;
1038 switch (src
[arg
].file
) {
1041 return components_read(arg
) * type_sz(src
[arg
].type
);
1047 return components_read(arg
) * src
[arg
].component_size(exec_size
);
1049 unreachable("MRF registers are not allowed as sources");
1056 predicate_width(brw_predicate predicate
)
1058 switch (predicate
) {
1059 case BRW_PREDICATE_NONE
: return 1;
1060 case BRW_PREDICATE_NORMAL
: return 1;
1061 case BRW_PREDICATE_ALIGN1_ANY2H
: return 2;
1062 case BRW_PREDICATE_ALIGN1_ALL2H
: return 2;
1063 case BRW_PREDICATE_ALIGN1_ANY4H
: return 4;
1064 case BRW_PREDICATE_ALIGN1_ALL4H
: return 4;
1065 case BRW_PREDICATE_ALIGN1_ANY8H
: return 8;
1066 case BRW_PREDICATE_ALIGN1_ALL8H
: return 8;
1067 case BRW_PREDICATE_ALIGN1_ANY16H
: return 16;
1068 case BRW_PREDICATE_ALIGN1_ALL16H
: return 16;
1069 case BRW_PREDICATE_ALIGN1_ANY32H
: return 32;
1070 case BRW_PREDICATE_ALIGN1_ALL32H
: return 32;
1071 default: unreachable("Unsupported predicate");
1075 /* Return the subset of flag registers that an instruction could
1076 * potentially read or write based on the execution controls and flag
1077 * subregister number of the instruction.
1080 flag_mask(const fs_inst
*inst
, unsigned width
)
1082 assert(util_is_power_of_two_nonzero(width
));
1083 const unsigned start
= (inst
->flag_subreg
* 16 + inst
->group
) &
1085 const unsigned end
= start
+ ALIGN(inst
->exec_size
, width
);
1086 return ((1 << DIV_ROUND_UP(end
, 8)) - 1) & ~((1 << (start
/ 8)) - 1);
1090 bit_mask(unsigned n
)
1092 return (n
>= CHAR_BIT
* sizeof(bit_mask(n
)) ? ~0u : (1u << n
) - 1);
1096 flag_mask(const fs_reg
&r
, unsigned sz
)
1098 if (r
.file
== ARF
) {
1099 const unsigned start
= (r
.nr
- BRW_ARF_FLAG
) * 4 + r
.subnr
;
1100 const unsigned end
= start
+ sz
;
1101 return bit_mask(end
) & ~bit_mask(start
);
1109 fs_inst::flags_read(const gen_device_info
*devinfo
) const
1111 if (predicate
== BRW_PREDICATE_ALIGN1_ANYV
||
1112 predicate
== BRW_PREDICATE_ALIGN1_ALLV
) {
1113 /* The vertical predication modes combine corresponding bits from
1114 * f0.0 and f1.0 on Gen7+, and f0.0 and f0.1 on older hardware.
1116 const unsigned shift
= devinfo
->gen
>= 7 ? 4 : 2;
1117 return flag_mask(this, 1) << shift
| flag_mask(this, 1);
1118 } else if (predicate
) {
1119 return flag_mask(this, predicate_width(predicate
));
1122 for (int i
= 0; i
< sources
; i
++) {
1123 mask
|= flag_mask(src
[i
], size_read(i
));
1130 fs_inst::flags_written() const
1132 if ((conditional_mod
&& (opcode
!= BRW_OPCODE_SEL
&&
1133 opcode
!= BRW_OPCODE_CSEL
&&
1134 opcode
!= BRW_OPCODE_IF
&&
1135 opcode
!= BRW_OPCODE_WHILE
)) ||
1136 opcode
== SHADER_OPCODE_FIND_LIVE_CHANNEL
||
1137 opcode
== FS_OPCODE_FB_WRITE
) {
1138 return flag_mask(this, 1);
1140 return flag_mask(dst
, size_written
);
1145 * Returns how many MRFs an FS opcode will write over.
1147 * Note that this is not the 0 or 1 implied writes in an actual gen
1148 * instruction -- the FS opcodes often generate MOVs in addition.
1151 fs_visitor::implied_mrf_writes(const fs_inst
*inst
) const
1153 if (inst
->mlen
== 0)
1156 if (inst
->base_mrf
== -1)
1159 switch (inst
->opcode
) {
1160 case SHADER_OPCODE_RCP
:
1161 case SHADER_OPCODE_RSQ
:
1162 case SHADER_OPCODE_SQRT
:
1163 case SHADER_OPCODE_EXP2
:
1164 case SHADER_OPCODE_LOG2
:
1165 case SHADER_OPCODE_SIN
:
1166 case SHADER_OPCODE_COS
:
1167 return 1 * dispatch_width
/ 8;
1168 case SHADER_OPCODE_POW
:
1169 case SHADER_OPCODE_INT_QUOTIENT
:
1170 case SHADER_OPCODE_INT_REMAINDER
:
1171 return 2 * dispatch_width
/ 8;
1172 case SHADER_OPCODE_TEX
:
1174 case SHADER_OPCODE_TXD
:
1175 case SHADER_OPCODE_TXF
:
1176 case SHADER_OPCODE_TXF_CMS
:
1177 case SHADER_OPCODE_TXF_MCS
:
1178 case SHADER_OPCODE_TG4
:
1179 case SHADER_OPCODE_TG4_OFFSET
:
1180 case SHADER_OPCODE_TXL
:
1181 case SHADER_OPCODE_TXS
:
1182 case SHADER_OPCODE_LOD
:
1183 case SHADER_OPCODE_SAMPLEINFO
:
1185 case FS_OPCODE_FB_WRITE
:
1186 case FS_OPCODE_REP_FB_WRITE
:
1187 return inst
->src
[0].file
== BAD_FILE
? 0 : 2;
1188 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
1189 case SHADER_OPCODE_GEN4_SCRATCH_READ
:
1191 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN4
:
1193 case SHADER_OPCODE_GEN4_SCRATCH_WRITE
:
1196 unreachable("not reached");
1201 fs_visitor::vgrf(const glsl_type
*const type
)
1203 int reg_width
= dispatch_width
/ 8;
1205 alloc
.allocate(type_size_scalar(type
, false) * reg_width
),
1206 brw_type_for_base_type(type
));
1209 fs_reg::fs_reg(enum brw_reg_file file
, int nr
)
1214 this->type
= BRW_REGISTER_TYPE_F
;
1215 this->stride
= (file
== UNIFORM
? 0 : 1);
1218 fs_reg::fs_reg(enum brw_reg_file file
, int nr
, enum brw_reg_type type
)
1224 this->stride
= (file
== UNIFORM
? 0 : 1);
1227 /* For SIMD16, we need to follow from the uniform setup of SIMD8 dispatch.
1228 * This brings in those uniform definitions
1231 fs_visitor::import_uniforms(fs_visitor
*v
)
1233 this->push_constant_loc
= v
->push_constant_loc
;
1234 this->pull_constant_loc
= v
->pull_constant_loc
;
1235 this->uniforms
= v
->uniforms
;
1236 this->subgroup_id
= v
->subgroup_id
;
1240 fs_visitor::emit_fragcoord_interpolation(fs_reg wpos
)
1242 assert(stage
== MESA_SHADER_FRAGMENT
);
1244 /* gl_FragCoord.x */
1245 bld
.MOV(wpos
, this->pixel_x
);
1246 wpos
= offset(wpos
, bld
, 1);
1248 /* gl_FragCoord.y */
1249 bld
.MOV(wpos
, this->pixel_y
);
1250 wpos
= offset(wpos
, bld
, 1);
1252 /* gl_FragCoord.z */
1253 if (devinfo
->gen
>= 6) {
1254 bld
.MOV(wpos
, fetch_payload_reg(bld
, payload
.source_depth_reg
));
1256 bld
.emit(FS_OPCODE_LINTERP
, wpos
,
1257 this->delta_xy
[BRW_BARYCENTRIC_PERSPECTIVE_PIXEL
],
1258 component(interp_reg(VARYING_SLOT_POS
, 2), 0));
1260 wpos
= offset(wpos
, bld
, 1);
1262 /* gl_FragCoord.w: Already set up in emit_interpolation */
1263 bld
.MOV(wpos
, this->wpos_w
);
1266 enum brw_barycentric_mode
1267 brw_barycentric_mode(enum glsl_interp_mode mode
, nir_intrinsic_op op
)
1269 /* Barycentric modes don't make sense for flat inputs. */
1270 assert(mode
!= INTERP_MODE_FLAT
);
1274 case nir_intrinsic_load_barycentric_pixel
:
1275 case nir_intrinsic_load_barycentric_at_offset
:
1276 bary
= BRW_BARYCENTRIC_PERSPECTIVE_PIXEL
;
1278 case nir_intrinsic_load_barycentric_centroid
:
1279 bary
= BRW_BARYCENTRIC_PERSPECTIVE_CENTROID
;
1281 case nir_intrinsic_load_barycentric_sample
:
1282 case nir_intrinsic_load_barycentric_at_sample
:
1283 bary
= BRW_BARYCENTRIC_PERSPECTIVE_SAMPLE
;
1286 unreachable("invalid intrinsic");
1289 if (mode
== INTERP_MODE_NOPERSPECTIVE
)
1292 return (enum brw_barycentric_mode
) bary
;
1296 * Turn one of the two CENTROID barycentric modes into PIXEL mode.
1298 static enum brw_barycentric_mode
1299 centroid_to_pixel(enum brw_barycentric_mode bary
)
1301 assert(bary
== BRW_BARYCENTRIC_PERSPECTIVE_CENTROID
||
1302 bary
== BRW_BARYCENTRIC_NONPERSPECTIVE_CENTROID
);
1303 return (enum brw_barycentric_mode
) ((unsigned) bary
- 1);
1307 fs_visitor::emit_frontfacing_interpolation()
1309 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::bool_type
));
1311 if (devinfo
->gen
>= 6) {
1312 /* Bit 15 of g0.0 is 0 if the polygon is front facing. We want to create
1313 * a boolean result from this (~0/true or 0/false).
1315 * We can use the fact that bit 15 is the MSB of g0.0:W to accomplish
1316 * this task in only one instruction:
1317 * - a negation source modifier will flip the bit; and
1318 * - a W -> D type conversion will sign extend the bit into the high
1319 * word of the destination.
1321 * An ASR 15 fills the low word of the destination.
1323 fs_reg g0
= fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_W
));
1326 bld
.ASR(*reg
, g0
, brw_imm_d(15));
1328 /* Bit 31 of g1.6 is 0 if the polygon is front facing. We want to create
1329 * a boolean result from this (1/true or 0/false).
1331 * Like in the above case, since the bit is the MSB of g1.6:UD we can use
1332 * the negation source modifier to flip it. Unfortunately the SHR
1333 * instruction only operates on UD (or D with an abs source modifier)
1334 * sources without negation.
1336 * Instead, use ASR (which will give ~0/true or 0/false).
1338 fs_reg g1_6
= fs_reg(retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_D
));
1341 bld
.ASR(*reg
, g1_6
, brw_imm_d(31));
1348 fs_visitor::compute_sample_position(fs_reg dst
, fs_reg int_sample_pos
)
1350 assert(stage
== MESA_SHADER_FRAGMENT
);
1351 struct brw_wm_prog_data
*wm_prog_data
= brw_wm_prog_data(this->prog_data
);
1352 assert(dst
.type
== BRW_REGISTER_TYPE_F
);
1354 if (wm_prog_data
->persample_dispatch
) {
1355 /* Convert int_sample_pos to floating point */
1356 bld
.MOV(dst
, int_sample_pos
);
1357 /* Scale to the range [0, 1] */
1358 bld
.MUL(dst
, dst
, brw_imm_f(1 / 16.0f
));
1361 /* From ARB_sample_shading specification:
1362 * "When rendering to a non-multisample buffer, or if multisample
1363 * rasterization is disabled, gl_SamplePosition will always be
1366 bld
.MOV(dst
, brw_imm_f(0.5f
));
1371 fs_visitor::emit_samplepos_setup()
1373 assert(devinfo
->gen
>= 6);
1375 const fs_builder abld
= bld
.annotate("compute sample position");
1376 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::vec2_type
));
1378 fs_reg int_sample_x
= vgrf(glsl_type::int_type
);
1379 fs_reg int_sample_y
= vgrf(glsl_type::int_type
);
1381 /* WM will be run in MSDISPMODE_PERSAMPLE. So, only one of SIMD8 or SIMD16
1382 * mode will be enabled.
1384 * From the Ivy Bridge PRM, volume 2 part 1, page 344:
1385 * R31.1:0 Position Offset X/Y for Slot[3:0]
1386 * R31.3:2 Position Offset X/Y for Slot[7:4]
1389 * The X, Y sample positions come in as bytes in thread payload. So, read
1390 * the positions using vstride=16, width=8, hstride=2.
1392 const fs_reg sample_pos_reg
=
1393 fetch_payload_reg(abld
, payload
.sample_pos_reg
, BRW_REGISTER_TYPE_W
);
1395 /* Compute gl_SamplePosition.x */
1396 abld
.MOV(int_sample_x
, subscript(sample_pos_reg
, BRW_REGISTER_TYPE_B
, 0));
1397 compute_sample_position(offset(pos
, abld
, 0), int_sample_x
);
1399 /* Compute gl_SamplePosition.y */
1400 abld
.MOV(int_sample_y
, subscript(sample_pos_reg
, BRW_REGISTER_TYPE_B
, 1));
1401 compute_sample_position(offset(pos
, abld
, 1), int_sample_y
);
1406 fs_visitor::emit_sampleid_setup()
1408 assert(stage
== MESA_SHADER_FRAGMENT
);
1409 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1410 assert(devinfo
->gen
>= 6);
1412 const fs_builder abld
= bld
.annotate("compute sample id");
1413 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::uint_type
));
1415 if (!key
->multisample_fbo
) {
1416 /* As per GL_ARB_sample_shading specification:
1417 * "When rendering to a non-multisample buffer, or if multisample
1418 * rasterization is disabled, gl_SampleID will always be zero."
1420 abld
.MOV(*reg
, brw_imm_d(0));
1421 } else if (devinfo
->gen
>= 8) {
1422 /* Sample ID comes in as 4-bit numbers in g1.0:
1424 * 15:12 Slot 3 SampleID (only used in SIMD16)
1425 * 11:8 Slot 2 SampleID (only used in SIMD16)
1426 * 7:4 Slot 1 SampleID
1427 * 3:0 Slot 0 SampleID
1429 * Each slot corresponds to four channels, so we want to replicate each
1430 * half-byte value to 4 channels in a row:
1432 * dst+0: .7 .6 .5 .4 .3 .2 .1 .0
1433 * 7:4 7:4 7:4 7:4 3:0 3:0 3:0 3:0
1435 * dst+1: .7 .6 .5 .4 .3 .2 .1 .0 (if SIMD16)
1436 * 15:12 15:12 15:12 15:12 11:8 11:8 11:8 11:8
1438 * First, we read g1.0 with a <1,8,0>UB region, causing the first 8
1439 * channels to read the first byte (7:0), and the second group of 8
1440 * channels to read the second byte (15:8). Then, we shift right by
1441 * a vector immediate of <4, 4, 4, 4, 0, 0, 0, 0>, moving the slot 1 / 3
1442 * values into place. Finally, we AND with 0xf to keep the low nibble.
1444 * shr(16) tmp<1>W g1.0<1,8,0>B 0x44440000:V
1445 * and(16) dst<1>D tmp<8,8,1>W 0xf:W
1447 * TODO: These payload bits exist on Gen7 too, but they appear to always
1448 * be zero, so this code fails to work. We should find out why.
1450 const fs_reg tmp
= abld
.vgrf(BRW_REGISTER_TYPE_UW
);
1452 for (unsigned i
= 0; i
< DIV_ROUND_UP(dispatch_width
, 16); i
++) {
1453 const fs_builder hbld
= abld
.group(MIN2(16, dispatch_width
), i
);
1454 hbld
.SHR(offset(tmp
, hbld
, i
),
1455 stride(retype(brw_vec1_grf(1 + i
, 0), BRW_REGISTER_TYPE_UB
),
1457 brw_imm_v(0x44440000));
1460 abld
.AND(*reg
, tmp
, brw_imm_w(0xf));
1462 const fs_reg t1
= component(abld
.vgrf(BRW_REGISTER_TYPE_UD
), 0);
1463 const fs_reg t2
= abld
.vgrf(BRW_REGISTER_TYPE_UW
);
1465 /* The PS will be run in MSDISPMODE_PERSAMPLE. For example with
1466 * 8x multisampling, subspan 0 will represent sample N (where N
1467 * is 0, 2, 4 or 6), subspan 1 will represent sample 1, 3, 5 or
1468 * 7. We can find the value of N by looking at R0.0 bits 7:6
1469 * ("Starting Sample Pair Index (SSPI)") and multiplying by two
1470 * (since samples are always delivered in pairs). That is, we
1471 * compute 2*((R0.0 & 0xc0) >> 6) == (R0.0 & 0xc0) >> 5. Then
1472 * we need to add N to the sequence (0, 0, 0, 0, 1, 1, 1, 1) in
1473 * case of SIMD8 and sequence (0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2,
1474 * 2, 3, 3, 3, 3) in case of SIMD16. We compute this sequence by
1475 * populating a temporary variable with the sequence (0, 1, 2, 3),
1476 * and then reading from it using vstride=1, width=4, hstride=0.
1477 * These computations hold good for 4x multisampling as well.
1479 * For 2x MSAA and SIMD16, we want to use the sequence (0, 1, 0, 1):
1480 * the first four slots are sample 0 of subspan 0; the next four
1481 * are sample 1 of subspan 0; the third group is sample 0 of
1482 * subspan 1, and finally sample 1 of subspan 1.
1485 /* SKL+ has an extra bit for the Starting Sample Pair Index to
1486 * accomodate 16x MSAA.
1488 abld
.exec_all().group(1, 0)
1489 .AND(t1
, fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
)),
1491 abld
.exec_all().group(1, 0).SHR(t1
, t1
, brw_imm_d(5));
1493 /* This works for SIMD8-SIMD16. It also works for SIMD32 but only if we
1494 * can assume 4x MSAA. Disallow it on IVB+
1496 * FINISHME: One day, we could come up with a way to do this that
1497 * actually works on gen7.
1499 if (devinfo
->gen
>= 7)
1500 limit_dispatch_width(16, "gl_SampleId is unsupported in SIMD32 on gen7");
1501 abld
.exec_all().group(8, 0).MOV(t2
, brw_imm_v(0x32103210));
1503 /* This special instruction takes care of setting vstride=1,
1504 * width=4, hstride=0 of t2 during an ADD instruction.
1506 abld
.emit(FS_OPCODE_SET_SAMPLE_ID
, *reg
, t1
, t2
);
1513 fs_visitor::emit_samplemaskin_setup()
1515 assert(stage
== MESA_SHADER_FRAGMENT
);
1516 struct brw_wm_prog_data
*wm_prog_data
= brw_wm_prog_data(this->prog_data
);
1517 assert(devinfo
->gen
>= 6);
1519 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::int_type
));
1521 fs_reg coverage_mask
=
1522 fetch_payload_reg(bld
, payload
.sample_mask_in_reg
, BRW_REGISTER_TYPE_D
);
1524 if (wm_prog_data
->persample_dispatch
) {
1525 /* gl_SampleMaskIn[] comes from two sources: the input coverage mask,
1526 * and a mask representing which sample is being processed by the
1527 * current shader invocation.
1529 * From the OES_sample_variables specification:
1530 * "When per-sample shading is active due to the use of a fragment input
1531 * qualified by "sample" or due to the use of the gl_SampleID or
1532 * gl_SamplePosition variables, only the bit for the current sample is
1533 * set in gl_SampleMaskIn."
1535 const fs_builder abld
= bld
.annotate("compute gl_SampleMaskIn");
1537 if (nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
].file
== BAD_FILE
)
1538 nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
] = *emit_sampleid_setup();
1540 fs_reg one
= vgrf(glsl_type::int_type
);
1541 fs_reg enabled_mask
= vgrf(glsl_type::int_type
);
1542 abld
.MOV(one
, brw_imm_d(1));
1543 abld
.SHL(enabled_mask
, one
, nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
]);
1544 abld
.AND(*reg
, enabled_mask
, coverage_mask
);
1546 /* In per-pixel mode, the coverage mask is sufficient. */
1547 *reg
= coverage_mask
;
1553 fs_visitor::resolve_source_modifiers(const fs_reg
&src
)
1555 if (!src
.abs
&& !src
.negate
)
1558 fs_reg temp
= bld
.vgrf(src
.type
);
1565 fs_visitor::emit_discard_jump()
1567 assert(brw_wm_prog_data(this->prog_data
)->uses_kill
);
1569 /* For performance, after a discard, jump to the end of the
1570 * shader if all relevant channels have been discarded.
1572 fs_inst
*discard_jump
= bld
.emit(FS_OPCODE_DISCARD_JUMP
);
1573 discard_jump
->flag_subreg
= 1;
1575 discard_jump
->predicate
= BRW_PREDICATE_ALIGN1_ANY4H
;
1576 discard_jump
->predicate_inverse
= true;
1580 fs_visitor::emit_gs_thread_end()
1582 assert(stage
== MESA_SHADER_GEOMETRY
);
1584 struct brw_gs_prog_data
*gs_prog_data
= brw_gs_prog_data(prog_data
);
1586 if (gs_compile
->control_data_header_size_bits
> 0) {
1587 emit_gs_control_data_bits(this->final_gs_vertex_count
);
1590 const fs_builder abld
= bld
.annotate("thread end");
1593 if (gs_prog_data
->static_vertex_count
!= -1) {
1594 foreach_in_list_reverse(fs_inst
, prev
, &this->instructions
) {
1595 if (prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8
||
1596 prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
||
1597 prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
||
1598 prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
) {
1601 /* Delete now dead instructions. */
1602 foreach_in_list_reverse_safe(exec_node
, dead
, &this->instructions
) {
1608 } else if (prev
->is_control_flow() || prev
->has_side_effects()) {
1612 fs_reg hdr
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1613 abld
.MOV(hdr
, fs_reg(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
)));
1614 inst
= abld
.emit(SHADER_OPCODE_URB_WRITE_SIMD8
, reg_undef
, hdr
);
1617 fs_reg payload
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
1618 fs_reg
*sources
= ralloc_array(mem_ctx
, fs_reg
, 2);
1619 sources
[0] = fs_reg(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
));
1620 sources
[1] = this->final_gs_vertex_count
;
1621 abld
.LOAD_PAYLOAD(payload
, sources
, 2, 2);
1622 inst
= abld
.emit(SHADER_OPCODE_URB_WRITE_SIMD8
, reg_undef
, payload
);
1630 fs_visitor::assign_curb_setup()
1632 unsigned uniform_push_length
= DIV_ROUND_UP(stage_prog_data
->nr_params
, 8);
1634 unsigned ubo_push_length
= 0;
1635 unsigned ubo_push_start
[4];
1636 for (int i
= 0; i
< 4; i
++) {
1637 ubo_push_start
[i
] = 8 * (ubo_push_length
+ uniform_push_length
);
1638 ubo_push_length
+= stage_prog_data
->ubo_ranges
[i
].length
;
1641 prog_data
->curb_read_length
= uniform_push_length
+ ubo_push_length
;
1643 /* Map the offsets in the UNIFORM file to fixed HW regs. */
1644 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1645 for (unsigned int i
= 0; i
< inst
->sources
; i
++) {
1646 if (inst
->src
[i
].file
== UNIFORM
) {
1647 int uniform_nr
= inst
->src
[i
].nr
+ inst
->src
[i
].offset
/ 4;
1649 if (inst
->src
[i
].nr
>= UBO_START
) {
1650 /* constant_nr is in 32-bit units, the rest are in bytes */
1651 constant_nr
= ubo_push_start
[inst
->src
[i
].nr
- UBO_START
] +
1652 inst
->src
[i
].offset
/ 4;
1653 } else if (uniform_nr
>= 0 && uniform_nr
< (int) uniforms
) {
1654 constant_nr
= push_constant_loc
[uniform_nr
];
1656 /* Section 5.11 of the OpenGL 4.1 spec says:
1657 * "Out-of-bounds reads return undefined values, which include
1658 * values from other variables of the active program or zero."
1659 * Just return the first push constant.
1664 struct brw_reg brw_reg
= brw_vec1_grf(payload
.num_regs
+
1667 brw_reg
.abs
= inst
->src
[i
].abs
;
1668 brw_reg
.negate
= inst
->src
[i
].negate
;
1670 assert(inst
->src
[i
].stride
== 0);
1671 inst
->src
[i
] = byte_offset(
1672 retype(brw_reg
, inst
->src
[i
].type
),
1673 inst
->src
[i
].offset
% 4);
1678 /* This may be updated in assign_urb_setup or assign_vs_urb_setup. */
1679 this->first_non_payload_grf
= payload
.num_regs
+ prog_data
->curb_read_length
;
1683 calculate_urb_setup(const struct gen_device_info
*devinfo
,
1684 const struct brw_wm_prog_key
*key
,
1685 struct brw_wm_prog_data
*prog_data
,
1686 const nir_shader
*nir
)
1688 memset(prog_data
->urb_setup
, -1,
1689 sizeof(prog_data
->urb_setup
[0]) * VARYING_SLOT_MAX
);
1692 /* Figure out where each of the incoming setup attributes lands. */
1693 if (devinfo
->gen
>= 6) {
1694 if (util_bitcount64(nir
->info
.inputs_read
&
1695 BRW_FS_VARYING_INPUT_MASK
) <= 16) {
1696 /* The SF/SBE pipeline stage can do arbitrary rearrangement of the
1697 * first 16 varying inputs, so we can put them wherever we want.
1698 * Just put them in order.
1700 * This is useful because it means that (a) inputs not used by the
1701 * fragment shader won't take up valuable register space, and (b) we
1702 * won't have to recompile the fragment shader if it gets paired with
1703 * a different vertex (or geometry) shader.
1705 for (unsigned int i
= 0; i
< VARYING_SLOT_MAX
; i
++) {
1706 if (nir
->info
.inputs_read
& BRW_FS_VARYING_INPUT_MASK
&
1707 BITFIELD64_BIT(i
)) {
1708 prog_data
->urb_setup
[i
] = urb_next
++;
1712 /* We have enough input varyings that the SF/SBE pipeline stage can't
1713 * arbitrarily rearrange them to suit our whim; we have to put them
1714 * in an order that matches the output of the previous pipeline stage
1715 * (geometry or vertex shader).
1717 struct brw_vue_map prev_stage_vue_map
;
1718 brw_compute_vue_map(devinfo
, &prev_stage_vue_map
,
1719 key
->input_slots_valid
,
1720 nir
->info
.separate_shader
);
1723 brw_compute_first_urb_slot_required(nir
->info
.inputs_read
,
1724 &prev_stage_vue_map
);
1726 assert(prev_stage_vue_map
.num_slots
<= first_slot
+ 32);
1727 for (int slot
= first_slot
; slot
< prev_stage_vue_map
.num_slots
;
1729 int varying
= prev_stage_vue_map
.slot_to_varying
[slot
];
1730 if (varying
!= BRW_VARYING_SLOT_PAD
&&
1731 (nir
->info
.inputs_read
& BRW_FS_VARYING_INPUT_MASK
&
1732 BITFIELD64_BIT(varying
))) {
1733 prog_data
->urb_setup
[varying
] = slot
- first_slot
;
1736 urb_next
= prev_stage_vue_map
.num_slots
- first_slot
;
1739 /* FINISHME: The sf doesn't map VS->FS inputs for us very well. */
1740 for (unsigned int i
= 0; i
< VARYING_SLOT_MAX
; i
++) {
1741 /* Point size is packed into the header, not as a general attribute */
1742 if (i
== VARYING_SLOT_PSIZ
)
1745 if (key
->input_slots_valid
& BITFIELD64_BIT(i
)) {
1746 /* The back color slot is skipped when the front color is
1747 * also written to. In addition, some slots can be
1748 * written in the vertex shader and not read in the
1749 * fragment shader. So the register number must always be
1750 * incremented, mapped or not.
1752 if (_mesa_varying_slot_in_fs((gl_varying_slot
) i
))
1753 prog_data
->urb_setup
[i
] = urb_next
;
1759 * It's a FS only attribute, and we did interpolation for this attribute
1760 * in SF thread. So, count it here, too.
1762 * See compile_sf_prog() for more info.
1764 if (nir
->info
.inputs_read
& BITFIELD64_BIT(VARYING_SLOT_PNTC
))
1765 prog_data
->urb_setup
[VARYING_SLOT_PNTC
] = urb_next
++;
1768 prog_data
->num_varying_inputs
= urb_next
;
1772 fs_visitor::assign_urb_setup()
1774 assert(stage
== MESA_SHADER_FRAGMENT
);
1775 struct brw_wm_prog_data
*prog_data
= brw_wm_prog_data(this->prog_data
);
1777 int urb_start
= payload
.num_regs
+ prog_data
->base
.curb_read_length
;
1779 /* Offset all the urb_setup[] index by the actual position of the
1780 * setup regs, now that the location of the constants has been chosen.
1782 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1783 for (int i
= 0; i
< inst
->sources
; i
++) {
1784 if (inst
->src
[i
].file
== ATTR
) {
1785 /* ATTR regs in the FS are in units of logical scalar inputs each
1786 * of which consumes half of a GRF register.
1788 assert(inst
->src
[i
].offset
< REG_SIZE
/ 2);
1789 const unsigned grf
= urb_start
+ inst
->src
[i
].nr
/ 2;
1790 const unsigned offset
= (inst
->src
[i
].nr
% 2) * (REG_SIZE
/ 2) +
1791 inst
->src
[i
].offset
;
1792 const unsigned width
= inst
->src
[i
].stride
== 0 ?
1793 1 : MIN2(inst
->exec_size
, 8);
1794 struct brw_reg reg
= stride(
1795 byte_offset(retype(brw_vec8_grf(grf
, 0), inst
->src
[i
].type
),
1797 width
* inst
->src
[i
].stride
,
1798 width
, inst
->src
[i
].stride
);
1799 reg
.abs
= inst
->src
[i
].abs
;
1800 reg
.negate
= inst
->src
[i
].negate
;
1806 /* Each attribute is 4 setup channels, each of which is half a reg. */
1807 this->first_non_payload_grf
+= prog_data
->num_varying_inputs
* 2;
1811 fs_visitor::convert_attr_sources_to_hw_regs(fs_inst
*inst
)
1813 for (int i
= 0; i
< inst
->sources
; i
++) {
1814 if (inst
->src
[i
].file
== ATTR
) {
1815 int grf
= payload
.num_regs
+
1816 prog_data
->curb_read_length
+
1818 inst
->src
[i
].offset
/ REG_SIZE
;
1820 /* As explained at brw_reg_from_fs_reg, From the Haswell PRM:
1822 * VertStride must be used to cross GRF register boundaries. This
1823 * rule implies that elements within a 'Width' cannot cross GRF
1826 * So, for registers that are large enough, we have to split the exec
1827 * size in two and trust the compression state to sort it out.
1829 unsigned total_size
= inst
->exec_size
*
1830 inst
->src
[i
].stride
*
1831 type_sz(inst
->src
[i
].type
);
1833 assert(total_size
<= 2 * REG_SIZE
);
1834 const unsigned exec_size
=
1835 (total_size
<= REG_SIZE
) ? inst
->exec_size
: inst
->exec_size
/ 2;
1837 unsigned width
= inst
->src
[i
].stride
== 0 ? 1 : exec_size
;
1838 struct brw_reg reg
=
1839 stride(byte_offset(retype(brw_vec8_grf(grf
, 0), inst
->src
[i
].type
),
1840 inst
->src
[i
].offset
% REG_SIZE
),
1841 exec_size
* inst
->src
[i
].stride
,
1842 width
, inst
->src
[i
].stride
);
1843 reg
.abs
= inst
->src
[i
].abs
;
1844 reg
.negate
= inst
->src
[i
].negate
;
1852 fs_visitor::assign_vs_urb_setup()
1854 struct brw_vs_prog_data
*vs_prog_data
= brw_vs_prog_data(prog_data
);
1856 assert(stage
== MESA_SHADER_VERTEX
);
1858 /* Each attribute is 4 regs. */
1859 this->first_non_payload_grf
+= 4 * vs_prog_data
->nr_attribute_slots
;
1861 assert(vs_prog_data
->base
.urb_read_length
<= 15);
1863 /* Rewrite all ATTR file references to the hw grf that they land in. */
1864 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1865 convert_attr_sources_to_hw_regs(inst
);
1870 fs_visitor::assign_tcs_urb_setup()
1872 assert(stage
== MESA_SHADER_TESS_CTRL
);
1874 /* Rewrite all ATTR file references to HW_REGs. */
1875 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1876 convert_attr_sources_to_hw_regs(inst
);
1881 fs_visitor::assign_tes_urb_setup()
1883 assert(stage
== MESA_SHADER_TESS_EVAL
);
1885 struct brw_vue_prog_data
*vue_prog_data
= brw_vue_prog_data(prog_data
);
1887 first_non_payload_grf
+= 8 * vue_prog_data
->urb_read_length
;
1889 /* Rewrite all ATTR file references to HW_REGs. */
1890 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1891 convert_attr_sources_to_hw_regs(inst
);
1896 fs_visitor::assign_gs_urb_setup()
1898 assert(stage
== MESA_SHADER_GEOMETRY
);
1900 struct brw_vue_prog_data
*vue_prog_data
= brw_vue_prog_data(prog_data
);
1902 first_non_payload_grf
+=
1903 8 * vue_prog_data
->urb_read_length
* nir
->info
.gs
.vertices_in
;
1905 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1906 /* Rewrite all ATTR file references to GRFs. */
1907 convert_attr_sources_to_hw_regs(inst
);
1913 * Split large virtual GRFs into separate components if we can.
1915 * This is mostly duplicated with what brw_fs_vector_splitting does,
1916 * but that's really conservative because it's afraid of doing
1917 * splitting that doesn't result in real progress after the rest of
1918 * the optimization phases, which would cause infinite looping in
1919 * optimization. We can do it once here, safely. This also has the
1920 * opportunity to split interpolated values, or maybe even uniforms,
1921 * which we don't have at the IR level.
1923 * We want to split, because virtual GRFs are what we register
1924 * allocate and spill (due to contiguousness requirements for some
1925 * instructions), and they're what we naturally generate in the
1926 * codegen process, but most virtual GRFs don't actually need to be
1927 * contiguous sets of GRFs. If we split, we'll end up with reduced
1928 * live intervals and better dead code elimination and coalescing.
1931 fs_visitor::split_virtual_grfs()
1933 /* Compact the register file so we eliminate dead vgrfs. This
1934 * only defines split points for live registers, so if we have
1935 * too large dead registers they will hit assertions later.
1937 compact_virtual_grfs();
1939 int num_vars
= this->alloc
.count
;
1941 /* Count the total number of registers */
1943 int vgrf_to_reg
[num_vars
];
1944 for (int i
= 0; i
< num_vars
; i
++) {
1945 vgrf_to_reg
[i
] = reg_count
;
1946 reg_count
+= alloc
.sizes
[i
];
1949 /* An array of "split points". For each register slot, this indicates
1950 * if this slot can be separated from the previous slot. Every time an
1951 * instruction uses multiple elements of a register (as a source or
1952 * destination), we mark the used slots as inseparable. Then we go
1953 * through and split the registers into the smallest pieces we can.
1955 bool *split_points
= new bool[reg_count
];
1956 memset(split_points
, 0, reg_count
* sizeof(*split_points
));
1958 /* Mark all used registers as fully splittable */
1959 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1960 if (inst
->dst
.file
== VGRF
) {
1961 int reg
= vgrf_to_reg
[inst
->dst
.nr
];
1962 for (unsigned j
= 1; j
< this->alloc
.sizes
[inst
->dst
.nr
]; j
++)
1963 split_points
[reg
+ j
] = true;
1966 for (int i
= 0; i
< inst
->sources
; i
++) {
1967 if (inst
->src
[i
].file
== VGRF
) {
1968 int reg
= vgrf_to_reg
[inst
->src
[i
].nr
];
1969 for (unsigned j
= 1; j
< this->alloc
.sizes
[inst
->src
[i
].nr
]; j
++)
1970 split_points
[reg
+ j
] = true;
1975 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1976 /* We fix up undef instructions later */
1977 if (inst
->opcode
== SHADER_OPCODE_UNDEF
) {
1978 /* UNDEF instructions are currently only used to undef entire
1979 * registers. We need this invariant later when we split them.
1981 assert(inst
->dst
.file
== VGRF
);
1982 assert(inst
->dst
.offset
== 0);
1983 assert(inst
->size_written
== alloc
.sizes
[inst
->dst
.nr
] * REG_SIZE
);
1987 if (inst
->dst
.file
== VGRF
) {
1988 int reg
= vgrf_to_reg
[inst
->dst
.nr
] + inst
->dst
.offset
/ REG_SIZE
;
1989 for (unsigned j
= 1; j
< regs_written(inst
); j
++)
1990 split_points
[reg
+ j
] = false;
1992 for (int i
= 0; i
< inst
->sources
; i
++) {
1993 if (inst
->src
[i
].file
== VGRF
) {
1994 int reg
= vgrf_to_reg
[inst
->src
[i
].nr
] + inst
->src
[i
].offset
/ REG_SIZE
;
1995 for (unsigned j
= 1; j
< regs_read(inst
, i
); j
++)
1996 split_points
[reg
+ j
] = false;
2001 int *new_virtual_grf
= new int[reg_count
];
2002 int *new_reg_offset
= new int[reg_count
];
2005 for (int i
= 0; i
< num_vars
; i
++) {
2006 /* The first one should always be 0 as a quick sanity check. */
2007 assert(split_points
[reg
] == false);
2010 new_reg_offset
[reg
] = 0;
2015 for (unsigned j
= 1; j
< alloc
.sizes
[i
]; j
++) {
2016 /* If this is a split point, reset the offset to 0 and allocate a
2017 * new virtual GRF for the previous offset many registers
2019 if (split_points
[reg
]) {
2020 assert(offset
<= MAX_VGRF_SIZE
);
2021 int grf
= alloc
.allocate(offset
);
2022 for (int k
= reg
- offset
; k
< reg
; k
++)
2023 new_virtual_grf
[k
] = grf
;
2026 new_reg_offset
[reg
] = offset
;
2031 /* The last one gets the original register number */
2032 assert(offset
<= MAX_VGRF_SIZE
);
2033 alloc
.sizes
[i
] = offset
;
2034 for (int k
= reg
- offset
; k
< reg
; k
++)
2035 new_virtual_grf
[k
] = i
;
2037 assert(reg
== reg_count
);
2039 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
2040 if (inst
->opcode
== SHADER_OPCODE_UNDEF
) {
2041 const fs_builder
ibld(this, block
, inst
);
2042 assert(inst
->size_written
% REG_SIZE
== 0);
2043 unsigned reg_offset
= 0;
2044 while (reg_offset
< inst
->size_written
/ REG_SIZE
) {
2045 reg
= vgrf_to_reg
[inst
->dst
.nr
] + reg_offset
;
2046 ibld
.UNDEF(fs_reg(VGRF
, new_virtual_grf
[reg
], inst
->dst
.type
));
2047 reg_offset
+= alloc
.sizes
[new_virtual_grf
[reg
]];
2049 inst
->remove(block
);
2053 if (inst
->dst
.file
== VGRF
) {
2054 reg
= vgrf_to_reg
[inst
->dst
.nr
] + inst
->dst
.offset
/ REG_SIZE
;
2055 inst
->dst
.nr
= new_virtual_grf
[reg
];
2056 inst
->dst
.offset
= new_reg_offset
[reg
] * REG_SIZE
+
2057 inst
->dst
.offset
% REG_SIZE
;
2058 assert((unsigned)new_reg_offset
[reg
] < alloc
.sizes
[new_virtual_grf
[reg
]]);
2060 for (int i
= 0; i
< inst
->sources
; i
++) {
2061 if (inst
->src
[i
].file
== VGRF
) {
2062 reg
= vgrf_to_reg
[inst
->src
[i
].nr
] + inst
->src
[i
].offset
/ REG_SIZE
;
2063 inst
->src
[i
].nr
= new_virtual_grf
[reg
];
2064 inst
->src
[i
].offset
= new_reg_offset
[reg
] * REG_SIZE
+
2065 inst
->src
[i
].offset
% REG_SIZE
;
2066 assert((unsigned)new_reg_offset
[reg
] < alloc
.sizes
[new_virtual_grf
[reg
]]);
2070 invalidate_live_intervals();
2072 delete[] split_points
;
2073 delete[] new_virtual_grf
;
2074 delete[] new_reg_offset
;
2078 * Remove unused virtual GRFs and compact the virtual_grf_* arrays.
2080 * During code generation, we create tons of temporary variables, many of
2081 * which get immediately killed and are never used again. Yet, in later
2082 * optimization and analysis passes, such as compute_live_intervals, we need
2083 * to loop over all the virtual GRFs. Compacting them can save a lot of
2087 fs_visitor::compact_virtual_grfs()
2089 bool progress
= false;
2090 int *remap_table
= new int[this->alloc
.count
];
2091 memset(remap_table
, -1, this->alloc
.count
* sizeof(int));
2093 /* Mark which virtual GRFs are used. */
2094 foreach_block_and_inst(block
, const fs_inst
, inst
, cfg
) {
2095 if (inst
->dst
.file
== VGRF
)
2096 remap_table
[inst
->dst
.nr
] = 0;
2098 for (int i
= 0; i
< inst
->sources
; i
++) {
2099 if (inst
->src
[i
].file
== VGRF
)
2100 remap_table
[inst
->src
[i
].nr
] = 0;
2104 /* Compact the GRF arrays. */
2106 for (unsigned i
= 0; i
< this->alloc
.count
; i
++) {
2107 if (remap_table
[i
] == -1) {
2108 /* We just found an unused register. This means that we are
2109 * actually going to compact something.
2113 remap_table
[i
] = new_index
;
2114 alloc
.sizes
[new_index
] = alloc
.sizes
[i
];
2115 invalidate_live_intervals();
2120 this->alloc
.count
= new_index
;
2122 /* Patch all the instructions to use the newly renumbered registers */
2123 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2124 if (inst
->dst
.file
== VGRF
)
2125 inst
->dst
.nr
= remap_table
[inst
->dst
.nr
];
2127 for (int i
= 0; i
< inst
->sources
; i
++) {
2128 if (inst
->src
[i
].file
== VGRF
)
2129 inst
->src
[i
].nr
= remap_table
[inst
->src
[i
].nr
];
2133 /* Patch all the references to delta_xy, since they're used in register
2134 * allocation. If they're unused, switch them to BAD_FILE so we don't
2135 * think some random VGRF is delta_xy.
2137 for (unsigned i
= 0; i
< ARRAY_SIZE(delta_xy
); i
++) {
2138 if (delta_xy
[i
].file
== VGRF
) {
2139 if (remap_table
[delta_xy
[i
].nr
] != -1) {
2140 delta_xy
[i
].nr
= remap_table
[delta_xy
[i
].nr
];
2142 delta_xy
[i
].file
= BAD_FILE
;
2147 delete[] remap_table
;
2153 get_subgroup_id_param_index(const brw_stage_prog_data
*prog_data
)
2155 if (prog_data
->nr_params
== 0)
2158 /* The local thread id is always the last parameter in the list */
2159 uint32_t last_param
= prog_data
->param
[prog_data
->nr_params
- 1];
2160 if (last_param
== BRW_PARAM_BUILTIN_SUBGROUP_ID
)
2161 return prog_data
->nr_params
- 1;
2167 * Struct for handling complex alignments.
2169 * A complex alignment is stored as multiplier and an offset. A value is
2170 * considered to be aligned if it is {offset} larger than a multiple of {mul}.
2171 * For instance, with an alignment of {8, 2}, cplx_align_apply would do the
2174 * N | cplx_align_apply({8, 2}, N)
2175 * ----+-----------------------------
2189 #define CPLX_ALIGN_MAX_MUL 8
2192 cplx_align_assert_sane(struct cplx_align a
)
2194 assert(a
.mul
> 0 && util_is_power_of_two_nonzero(a
.mul
));
2195 assert(a
.offset
< a
.mul
);
2199 * Combines two alignments to produce a least multiple of sorts.
2201 * The returned alignment is the smallest (in terms of multiplier) such that
2202 * anything aligned to both a and b will be aligned to the new alignment.
2203 * This function will assert-fail if a and b are not compatible, i.e. if the
2204 * offset parameters are such that no common alignment is possible.
2206 static struct cplx_align
2207 cplx_align_combine(struct cplx_align a
, struct cplx_align b
)
2209 cplx_align_assert_sane(a
);
2210 cplx_align_assert_sane(b
);
2212 /* Assert that the alignments agree. */
2213 assert((a
.offset
& (b
.mul
- 1)) == (b
.offset
& (a
.mul
- 1)));
2215 return a
.mul
> b
.mul
? a
: b
;
2219 * Apply a complex alignment
2221 * This function will return the smallest number greater than or equal to
2222 * offset that is aligned to align.
2225 cplx_align_apply(struct cplx_align align
, unsigned offset
)
2227 return ALIGN(offset
- align
.offset
, align
.mul
) + align
.offset
;
2230 #define UNIFORM_SLOT_SIZE 4
2232 struct uniform_slot_info
{
2233 /** True if the given uniform slot is live */
2236 /** True if this slot and the next slot must remain contiguous */
2237 unsigned contiguous
:1;
2239 struct cplx_align align
;
2243 mark_uniform_slots_read(struct uniform_slot_info
*slots
,
2244 unsigned num_slots
, unsigned alignment
)
2246 assert(alignment
> 0 && util_is_power_of_two_nonzero(alignment
));
2247 assert(alignment
<= CPLX_ALIGN_MAX_MUL
);
2249 /* We can't align a slot to anything less than the slot size */
2250 alignment
= MAX2(alignment
, UNIFORM_SLOT_SIZE
);
2252 struct cplx_align align
= {alignment
, 0};
2253 cplx_align_assert_sane(align
);
2255 for (unsigned i
= 0; i
< num_slots
; i
++) {
2256 slots
[i
].is_live
= true;
2257 if (i
< num_slots
- 1)
2258 slots
[i
].contiguous
= true;
2260 align
.offset
= (i
* UNIFORM_SLOT_SIZE
) & (align
.mul
- 1);
2261 if (slots
[i
].align
.mul
== 0) {
2262 slots
[i
].align
= align
;
2264 slots
[i
].align
= cplx_align_combine(slots
[i
].align
, align
);
2270 * Assign UNIFORM file registers to either push constants or pull constants.
2272 * We allow a fragment shader to have more than the specified minimum
2273 * maximum number of fragment shader uniform components (64). If
2274 * there are too many of these, they'd fill up all of register space.
2275 * So, this will push some of them out to the pull constant buffer and
2276 * update the program to load them.
2279 fs_visitor::assign_constant_locations()
2281 /* Only the first compile gets to decide on locations. */
2282 if (push_constant_loc
) {
2283 assert(pull_constant_loc
);
2287 struct uniform_slot_info slots
[uniforms
];
2288 memset(slots
, 0, sizeof(slots
));
2290 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
2291 for (int i
= 0 ; i
< inst
->sources
; i
++) {
2292 if (inst
->src
[i
].file
!= UNIFORM
)
2295 /* NIR tightly packs things so the uniform number might not be
2296 * aligned (if we have a double right after a float, for instance).
2297 * This is fine because the process of re-arranging them will ensure
2298 * that things are properly aligned. The offset into that uniform,
2299 * however, must be aligned.
2301 * In Vulkan, we have explicit offsets but everything is crammed
2302 * into a single "variable" so inst->src[i].nr will always be 0.
2303 * Everything will be properly aligned relative to that one base.
2305 assert(inst
->src
[i
].offset
% type_sz(inst
->src
[i
].type
) == 0);
2307 unsigned u
= inst
->src
[i
].nr
+
2308 inst
->src
[i
].offset
/ UNIFORM_SLOT_SIZE
;
2313 unsigned slots_read
;
2314 if (inst
->opcode
== SHADER_OPCODE_MOV_INDIRECT
&& i
== 0) {
2315 slots_read
= DIV_ROUND_UP(inst
->src
[2].ud
, UNIFORM_SLOT_SIZE
);
2317 unsigned bytes_read
= inst
->components_read(i
) *
2318 type_sz(inst
->src
[i
].type
);
2319 slots_read
= DIV_ROUND_UP(bytes_read
, UNIFORM_SLOT_SIZE
);
2322 assert(u
+ slots_read
<= uniforms
);
2323 mark_uniform_slots_read(&slots
[u
], slots_read
,
2324 type_sz(inst
->src
[i
].type
));
2328 int subgroup_id_index
= get_subgroup_id_param_index(stage_prog_data
);
2330 /* Only allow 16 registers (128 uniform components) as push constants.
2332 * Just demote the end of the list. We could probably do better
2333 * here, demoting things that are rarely used in the program first.
2335 * If changing this value, note the limitation about total_regs in
2338 unsigned int max_push_components
= 16 * 8;
2339 if (subgroup_id_index
>= 0)
2340 max_push_components
--; /* Save a slot for the thread ID */
2342 /* We push small arrays, but no bigger than 16 floats. This is big enough
2343 * for a vec4 but hopefully not large enough to push out other stuff. We
2344 * should probably use a better heuristic at some point.
2346 const unsigned int max_chunk_size
= 16;
2348 unsigned int num_push_constants
= 0;
2349 unsigned int num_pull_constants
= 0;
2351 push_constant_loc
= ralloc_array(mem_ctx
, int, uniforms
);
2352 pull_constant_loc
= ralloc_array(mem_ctx
, int, uniforms
);
2354 /* Default to -1 meaning no location */
2355 memset(push_constant_loc
, -1, uniforms
* sizeof(*push_constant_loc
));
2356 memset(pull_constant_loc
, -1, uniforms
* sizeof(*pull_constant_loc
));
2358 int chunk_start
= -1;
2359 struct cplx_align align
;
2360 for (unsigned u
= 0; u
< uniforms
; u
++) {
2361 if (!slots
[u
].is_live
) {
2362 assert(chunk_start
== -1);
2366 /* Skip subgroup_id_index to put it in the last push register. */
2367 if (subgroup_id_index
== (int)u
)
2370 if (chunk_start
== -1) {
2372 align
= slots
[u
].align
;
2374 /* Offset into the chunk */
2375 unsigned chunk_offset
= (u
- chunk_start
) * UNIFORM_SLOT_SIZE
;
2377 /* Shift the slot alignment down by the chunk offset so it is
2378 * comparable with the base chunk alignment.
2380 struct cplx_align slot_align
= slots
[u
].align
;
2382 (slot_align
.offset
- chunk_offset
) & (align
.mul
- 1);
2384 align
= cplx_align_combine(align
, slot_align
);
2387 /* Sanity check the alignment */
2388 cplx_align_assert_sane(align
);
2390 if (slots
[u
].contiguous
)
2393 /* Adjust the alignment to be in terms of slots, not bytes */
2394 assert((align
.mul
& (UNIFORM_SLOT_SIZE
- 1)) == 0);
2395 assert((align
.offset
& (UNIFORM_SLOT_SIZE
- 1)) == 0);
2396 align
.mul
/= UNIFORM_SLOT_SIZE
;
2397 align
.offset
/= UNIFORM_SLOT_SIZE
;
2399 unsigned push_start_align
= cplx_align_apply(align
, num_push_constants
);
2400 unsigned chunk_size
= u
- chunk_start
+ 1;
2401 if ((!compiler
->supports_pull_constants
&& u
< UBO_START
) ||
2402 (chunk_size
< max_chunk_size
&&
2403 push_start_align
+ chunk_size
<= max_push_components
)) {
2404 /* Align up the number of push constants */
2405 num_push_constants
= push_start_align
;
2406 for (unsigned i
= 0; i
< chunk_size
; i
++)
2407 push_constant_loc
[chunk_start
+ i
] = num_push_constants
++;
2409 /* We need to pull this one */
2410 num_pull_constants
= cplx_align_apply(align
, num_pull_constants
);
2411 for (unsigned i
= 0; i
< chunk_size
; i
++)
2412 pull_constant_loc
[chunk_start
+ i
] = num_pull_constants
++;
2415 /* Reset the chunk and start again */
2419 /* Add the CS local thread ID uniform at the end of the push constants */
2420 if (subgroup_id_index
>= 0)
2421 push_constant_loc
[subgroup_id_index
] = num_push_constants
++;
2423 /* As the uniforms are going to be reordered, stash the old array and
2424 * create two new arrays for push/pull params.
2426 uint32_t *param
= stage_prog_data
->param
;
2427 stage_prog_data
->nr_params
= num_push_constants
;
2428 if (num_push_constants
) {
2429 stage_prog_data
->param
= rzalloc_array(mem_ctx
, uint32_t,
2430 num_push_constants
);
2432 stage_prog_data
->param
= NULL
;
2434 assert(stage_prog_data
->nr_pull_params
== 0);
2435 assert(stage_prog_data
->pull_param
== NULL
);
2436 if (num_pull_constants
> 0) {
2437 stage_prog_data
->nr_pull_params
= num_pull_constants
;
2438 stage_prog_data
->pull_param
= rzalloc_array(mem_ctx
, uint32_t,
2439 num_pull_constants
);
2442 /* Now that we know how many regular uniforms we'll push, reduce the
2443 * UBO push ranges so we don't exceed the 3DSTATE_CONSTANT limits.
2445 unsigned push_length
= DIV_ROUND_UP(stage_prog_data
->nr_params
, 8);
2446 for (int i
= 0; i
< 4; i
++) {
2447 struct brw_ubo_range
*range
= &prog_data
->ubo_ranges
[i
];
2449 if (push_length
+ range
->length
> 64)
2450 range
->length
= 64 - push_length
;
2452 push_length
+= range
->length
;
2454 assert(push_length
<= 64);
2456 /* Up until now, the param[] array has been indexed by reg + offset
2457 * of UNIFORM registers. Move pull constants into pull_param[] and
2458 * condense param[] to only contain the uniforms we chose to push.
2460 * NOTE: Because we are condensing the params[] array, we know that
2461 * push_constant_loc[i] <= i and we can do it in one smooth loop without
2462 * having to make a copy.
2464 for (unsigned int i
= 0; i
< uniforms
; i
++) {
2465 uint32_t value
= param
[i
];
2466 if (pull_constant_loc
[i
] != -1) {
2467 stage_prog_data
->pull_param
[pull_constant_loc
[i
]] = value
;
2468 } else if (push_constant_loc
[i
] != -1) {
2469 stage_prog_data
->param
[push_constant_loc
[i
]] = value
;
2476 fs_visitor::get_pull_locs(const fs_reg
&src
,
2477 unsigned *out_surf_index
,
2478 unsigned *out_pull_index
)
2480 assert(src
.file
== UNIFORM
);
2482 if (src
.nr
>= UBO_START
) {
2483 const struct brw_ubo_range
*range
=
2484 &prog_data
->ubo_ranges
[src
.nr
- UBO_START
];
2486 /* If this access is in our (reduced) range, use the push data. */
2487 if (src
.offset
/ 32 < range
->length
)
2490 *out_surf_index
= prog_data
->binding_table
.ubo_start
+ range
->block
;
2491 *out_pull_index
= (32 * range
->start
+ src
.offset
) / 4;
2493 prog_data
->has_ubo_pull
= true;
2497 const unsigned location
= src
.nr
+ src
.offset
/ 4;
2499 if (location
< uniforms
&& pull_constant_loc
[location
] != -1) {
2500 /* A regular uniform push constant */
2501 *out_surf_index
= stage_prog_data
->binding_table
.pull_constants_start
;
2502 *out_pull_index
= pull_constant_loc
[location
];
2504 prog_data
->has_ubo_pull
= true;
2512 * Replace UNIFORM register file access with either UNIFORM_PULL_CONSTANT_LOAD
2513 * or VARYING_PULL_CONSTANT_LOAD instructions which load values into VGRFs.
2516 fs_visitor::lower_constant_loads()
2518 unsigned index
, pull_index
;
2520 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
2521 /* Set up the annotation tracking for new generated instructions. */
2522 const fs_builder
ibld(this, block
, inst
);
2524 for (int i
= 0; i
< inst
->sources
; i
++) {
2525 if (inst
->src
[i
].file
!= UNIFORM
)
2528 /* We'll handle this case later */
2529 if (inst
->opcode
== SHADER_OPCODE_MOV_INDIRECT
&& i
== 0)
2532 if (!get_pull_locs(inst
->src
[i
], &index
, &pull_index
))
2535 assert(inst
->src
[i
].stride
== 0);
2537 const unsigned block_sz
= 64; /* Fetch one cacheline at a time. */
2538 const fs_builder ubld
= ibld
.exec_all().group(block_sz
/ 4, 0);
2539 const fs_reg dst
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
2540 const unsigned base
= pull_index
* 4;
2542 ubld
.emit(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
,
2543 dst
, brw_imm_ud(index
), brw_imm_ud(base
& ~(block_sz
- 1)));
2545 /* Rewrite the instruction to use the temporary VGRF. */
2546 inst
->src
[i
].file
= VGRF
;
2547 inst
->src
[i
].nr
= dst
.nr
;
2548 inst
->src
[i
].offset
= (base
& (block_sz
- 1)) +
2549 inst
->src
[i
].offset
% 4;
2552 if (inst
->opcode
== SHADER_OPCODE_MOV_INDIRECT
&&
2553 inst
->src
[0].file
== UNIFORM
) {
2555 if (!get_pull_locs(inst
->src
[0], &index
, &pull_index
))
2558 VARYING_PULL_CONSTANT_LOAD(ibld
, inst
->dst
,
2562 inst
->remove(block
);
2565 invalidate_live_intervals();
2569 fs_visitor::opt_algebraic()
2571 bool progress
= false;
2573 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
2574 switch (inst
->opcode
) {
2575 case BRW_OPCODE_MOV
:
2576 if (!devinfo
->has_64bit_types
&&
2577 (inst
->dst
.type
== BRW_REGISTER_TYPE_DF
||
2578 inst
->dst
.type
== BRW_REGISTER_TYPE_UQ
||
2579 inst
->dst
.type
== BRW_REGISTER_TYPE_Q
)) {
2580 assert(inst
->dst
.type
== inst
->src
[0].type
);
2581 assert(!inst
->saturate
);
2582 assert(!inst
->src
[0].abs
);
2583 assert(!inst
->src
[0].negate
);
2584 const brw::fs_builder
ibld(this, block
, inst
);
2586 if (inst
->src
[0].file
== IMM
) {
2587 ibld
.MOV(subscript(inst
->dst
, BRW_REGISTER_TYPE_UD
, 1),
2588 brw_imm_ud(inst
->src
[0].u64
>> 32));
2589 ibld
.MOV(subscript(inst
->dst
, BRW_REGISTER_TYPE_UD
, 0),
2590 brw_imm_ud(inst
->src
[0].u64
));
2592 ibld
.MOV(subscript(inst
->dst
, BRW_REGISTER_TYPE_UD
, 1),
2593 subscript(inst
->src
[0], BRW_REGISTER_TYPE_UD
, 1));
2594 ibld
.MOV(subscript(inst
->dst
, BRW_REGISTER_TYPE_UD
, 0),
2595 subscript(inst
->src
[0], BRW_REGISTER_TYPE_UD
, 0));
2598 inst
->remove(block
);
2602 if ((inst
->conditional_mod
== BRW_CONDITIONAL_Z
||
2603 inst
->conditional_mod
== BRW_CONDITIONAL_NZ
) &&
2604 inst
->dst
.is_null() &&
2605 (inst
->src
[0].abs
|| inst
->src
[0].negate
)) {
2606 inst
->src
[0].abs
= false;
2607 inst
->src
[0].negate
= false;
2612 if (inst
->src
[0].file
!= IMM
)
2615 if (inst
->saturate
) {
2616 /* Full mixed-type saturates don't happen. However, we can end up
2619 * mov.sat(8) g21<1>DF -1F
2621 * Other mixed-size-but-same-base-type cases may also be possible.
2623 if (inst
->dst
.type
!= inst
->src
[0].type
&&
2624 inst
->dst
.type
!= BRW_REGISTER_TYPE_DF
&&
2625 inst
->src
[0].type
!= BRW_REGISTER_TYPE_F
)
2626 assert(!"unimplemented: saturate mixed types");
2628 if (brw_saturate_immediate(inst
->src
[0].type
,
2629 &inst
->src
[0].as_brw_reg())) {
2630 inst
->saturate
= false;
2636 case BRW_OPCODE_MUL
:
2637 if (inst
->src
[1].file
!= IMM
)
2641 if (inst
->src
[1].is_one()) {
2642 inst
->opcode
= BRW_OPCODE_MOV
;
2643 inst
->src
[1] = reg_undef
;
2649 if (inst
->src
[1].is_negative_one()) {
2650 inst
->opcode
= BRW_OPCODE_MOV
;
2651 inst
->src
[0].negate
= !inst
->src
[0].negate
;
2652 inst
->src
[1] = reg_undef
;
2657 if (inst
->src
[0].file
== IMM
) {
2658 assert(inst
->src
[0].type
== BRW_REGISTER_TYPE_F
);
2659 inst
->opcode
= BRW_OPCODE_MOV
;
2660 inst
->src
[0].f
*= inst
->src
[1].f
;
2661 inst
->src
[1] = reg_undef
;
2666 case BRW_OPCODE_ADD
:
2667 if (inst
->src
[1].file
!= IMM
)
2670 if (inst
->src
[0].file
== IMM
) {
2671 assert(inst
->src
[0].type
== BRW_REGISTER_TYPE_F
);
2672 inst
->opcode
= BRW_OPCODE_MOV
;
2673 inst
->src
[0].f
+= inst
->src
[1].f
;
2674 inst
->src
[1] = reg_undef
;
2680 if (inst
->src
[0].equals(inst
->src
[1]) ||
2681 inst
->src
[1].is_zero()) {
2682 /* On Gen8+, the OR instruction can have a source modifier that
2683 * performs logical not on the operand. Cases of 'OR r0, ~r1, 0'
2684 * or 'OR r0, ~r1, ~r1' should become a NOT instead of a MOV.
2686 if (inst
->src
[0].negate
) {
2687 inst
->opcode
= BRW_OPCODE_NOT
;
2688 inst
->src
[0].negate
= false;
2690 inst
->opcode
= BRW_OPCODE_MOV
;
2692 inst
->src
[1] = reg_undef
;
2697 case BRW_OPCODE_CMP
:
2698 if ((inst
->conditional_mod
== BRW_CONDITIONAL_Z
||
2699 inst
->conditional_mod
== BRW_CONDITIONAL_NZ
) &&
2700 inst
->src
[1].is_zero() &&
2701 (inst
->src
[0].abs
|| inst
->src
[0].negate
)) {
2702 inst
->src
[0].abs
= false;
2703 inst
->src
[0].negate
= false;
2708 case BRW_OPCODE_SEL
:
2709 if (!devinfo
->has_64bit_types
&&
2710 (inst
->dst
.type
== BRW_REGISTER_TYPE_DF
||
2711 inst
->dst
.type
== BRW_REGISTER_TYPE_UQ
||
2712 inst
->dst
.type
== BRW_REGISTER_TYPE_Q
)) {
2713 assert(inst
->dst
.type
== inst
->src
[0].type
);
2714 assert(!inst
->saturate
);
2715 assert(!inst
->src
[0].abs
&& !inst
->src
[0].negate
);
2716 assert(!inst
->src
[1].abs
&& !inst
->src
[1].negate
);
2717 const brw::fs_builder
ibld(this, block
, inst
);
2719 set_predicate(inst
->predicate
,
2720 ibld
.SEL(subscript(inst
->dst
, BRW_REGISTER_TYPE_UD
, 0),
2721 subscript(inst
->src
[0], BRW_REGISTER_TYPE_UD
, 0),
2722 subscript(inst
->src
[1], BRW_REGISTER_TYPE_UD
, 0)));
2723 set_predicate(inst
->predicate
,
2724 ibld
.SEL(subscript(inst
->dst
, BRW_REGISTER_TYPE_UD
, 1),
2725 subscript(inst
->src
[0], BRW_REGISTER_TYPE_UD
, 1),
2726 subscript(inst
->src
[1], BRW_REGISTER_TYPE_UD
, 1)));
2728 inst
->remove(block
);
2731 if (inst
->src
[0].equals(inst
->src
[1])) {
2732 inst
->opcode
= BRW_OPCODE_MOV
;
2733 inst
->src
[1] = reg_undef
;
2734 inst
->predicate
= BRW_PREDICATE_NONE
;
2735 inst
->predicate_inverse
= false;
2737 } else if (inst
->saturate
&& inst
->src
[1].file
== IMM
) {
2738 switch (inst
->conditional_mod
) {
2739 case BRW_CONDITIONAL_LE
:
2740 case BRW_CONDITIONAL_L
:
2741 switch (inst
->src
[1].type
) {
2742 case BRW_REGISTER_TYPE_F
:
2743 if (inst
->src
[1].f
>= 1.0f
) {
2744 inst
->opcode
= BRW_OPCODE_MOV
;
2745 inst
->src
[1] = reg_undef
;
2746 inst
->conditional_mod
= BRW_CONDITIONAL_NONE
;
2754 case BRW_CONDITIONAL_GE
:
2755 case BRW_CONDITIONAL_G
:
2756 switch (inst
->src
[1].type
) {
2757 case BRW_REGISTER_TYPE_F
:
2758 if (inst
->src
[1].f
<= 0.0f
) {
2759 inst
->opcode
= BRW_OPCODE_MOV
;
2760 inst
->src
[1] = reg_undef
;
2761 inst
->conditional_mod
= BRW_CONDITIONAL_NONE
;
2773 case BRW_OPCODE_MAD
:
2774 if (inst
->src
[0].type
!= BRW_REGISTER_TYPE_F
||
2775 inst
->src
[1].type
!= BRW_REGISTER_TYPE_F
||
2776 inst
->src
[2].type
!= BRW_REGISTER_TYPE_F
)
2778 if (inst
->src
[1].is_one()) {
2779 inst
->opcode
= BRW_OPCODE_ADD
;
2780 inst
->src
[1] = inst
->src
[2];
2781 inst
->src
[2] = reg_undef
;
2783 } else if (inst
->src
[2].is_one()) {
2784 inst
->opcode
= BRW_OPCODE_ADD
;
2785 inst
->src
[2] = reg_undef
;
2789 case SHADER_OPCODE_BROADCAST
:
2790 if (is_uniform(inst
->src
[0])) {
2791 inst
->opcode
= BRW_OPCODE_MOV
;
2793 inst
->force_writemask_all
= true;
2795 } else if (inst
->src
[1].file
== IMM
) {
2796 inst
->opcode
= BRW_OPCODE_MOV
;
2797 /* It's possible that the selected component will be too large and
2798 * overflow the register. This can happen if someone does a
2799 * readInvocation() from GLSL or SPIR-V and provides an OOB
2800 * invocationIndex. If this happens and we some how manage
2801 * to constant fold it in and get here, then component() may cause
2802 * us to start reading outside of the VGRF which will lead to an
2803 * assert later. Instead, just let it wrap around if it goes over
2806 const unsigned comp
= inst
->src
[1].ud
& (inst
->exec_size
- 1);
2807 inst
->src
[0] = component(inst
->src
[0], comp
);
2809 inst
->force_writemask_all
= true;
2814 case SHADER_OPCODE_SHUFFLE
:
2815 if (is_uniform(inst
->src
[0])) {
2816 inst
->opcode
= BRW_OPCODE_MOV
;
2819 } else if (inst
->src
[1].file
== IMM
) {
2820 inst
->opcode
= BRW_OPCODE_MOV
;
2821 inst
->src
[0] = component(inst
->src
[0],
2832 /* Swap if src[0] is immediate. */
2833 if (progress
&& inst
->is_commutative()) {
2834 if (inst
->src
[0].file
== IMM
) {
2835 fs_reg tmp
= inst
->src
[1];
2836 inst
->src
[1] = inst
->src
[0];
2845 * Optimize sample messages that have constant zero values for the trailing
2846 * texture coordinates. We can just reduce the message length for these
2847 * instructions instead of reserving a register for it. Trailing parameters
2848 * that aren't sent default to zero anyway. This will cause the dead code
2849 * eliminator to remove the MOV instruction that would otherwise be emitted to
2850 * set up the zero value.
2853 fs_visitor::opt_zero_samples()
2855 /* Gen4 infers the texturing opcode based on the message length so we can't
2858 if (devinfo
->gen
< 5)
2861 bool progress
= false;
2863 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2864 if (!inst
->is_tex())
2867 fs_inst
*load_payload
= (fs_inst
*) inst
->prev
;
2869 if (load_payload
->is_head_sentinel() ||
2870 load_payload
->opcode
!= SHADER_OPCODE_LOAD_PAYLOAD
)
2873 /* We don't want to remove the message header or the first parameter.
2874 * Removing the first parameter is not allowed, see the Haswell PRM
2875 * volume 7, page 149:
2877 * "Parameter 0 is required except for the sampleinfo message, which
2878 * has no parameter 0"
2880 while (inst
->mlen
> inst
->header_size
+ inst
->exec_size
/ 8 &&
2881 load_payload
->src
[(inst
->mlen
- inst
->header_size
) /
2882 (inst
->exec_size
/ 8) +
2883 inst
->header_size
- 1].is_zero()) {
2884 inst
->mlen
-= inst
->exec_size
/ 8;
2890 invalidate_live_intervals();
2896 * Optimize sample messages which are followed by the final RT write.
2898 * CHV, and GEN9+ can mark a texturing SEND instruction with EOT to have its
2899 * results sent directly to the framebuffer, bypassing the EU. Recognize the
2900 * final texturing results copied to the framebuffer write payload and modify
2901 * them to write to the framebuffer directly.
2904 fs_visitor::opt_sampler_eot()
2906 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
2908 if (stage
!= MESA_SHADER_FRAGMENT
|| dispatch_width
> 16)
2911 if (devinfo
->gen
!= 9 && !devinfo
->is_cherryview
)
2914 /* FINISHME: It should be possible to implement this optimization when there
2915 * are multiple drawbuffers.
2917 if (key
->nr_color_regions
!= 1)
2920 /* Requires emitting a bunch of saturating MOV instructions during logical
2921 * send lowering to clamp the color payload, which the sampler unit isn't
2922 * going to do for us.
2924 if (key
->clamp_fragment_color
)
2927 /* Look for a texturing instruction immediately before the final FB_WRITE. */
2928 bblock_t
*block
= cfg
->blocks
[cfg
->num_blocks
- 1];
2929 fs_inst
*fb_write
= (fs_inst
*)block
->end();
2930 assert(fb_write
->eot
);
2931 assert(fb_write
->opcode
== FS_OPCODE_FB_WRITE_LOGICAL
);
2933 /* There wasn't one; nothing to do. */
2934 if (unlikely(fb_write
->prev
->is_head_sentinel()))
2937 fs_inst
*tex_inst
= (fs_inst
*) fb_write
->prev
;
2939 /* 3D Sampler » Messages » Message Format
2941 * “Response Length of zero is allowed on all SIMD8* and SIMD16* sampler
2942 * messages except sample+killpix, resinfo, sampleinfo, LOD, and gather4*”
2944 if (tex_inst
->opcode
!= SHADER_OPCODE_TEX_LOGICAL
&&
2945 tex_inst
->opcode
!= SHADER_OPCODE_TXD_LOGICAL
&&
2946 tex_inst
->opcode
!= SHADER_OPCODE_TXF_LOGICAL
&&
2947 tex_inst
->opcode
!= SHADER_OPCODE_TXL_LOGICAL
&&
2948 tex_inst
->opcode
!= FS_OPCODE_TXB_LOGICAL
&&
2949 tex_inst
->opcode
!= SHADER_OPCODE_TXF_CMS_LOGICAL
&&
2950 tex_inst
->opcode
!= SHADER_OPCODE_TXF_CMS_W_LOGICAL
&&
2951 tex_inst
->opcode
!= SHADER_OPCODE_TXF_UMS_LOGICAL
)
2954 /* XXX - This shouldn't be necessary. */
2955 if (tex_inst
->prev
->is_head_sentinel())
2958 /* Check that the FB write sources are fully initialized by the single
2959 * texturing instruction.
2961 for (unsigned i
= 0; i
< FB_WRITE_LOGICAL_NUM_SRCS
; i
++) {
2962 if (i
== FB_WRITE_LOGICAL_SRC_COLOR0
) {
2963 if (!fb_write
->src
[i
].equals(tex_inst
->dst
) ||
2964 fb_write
->size_read(i
) != tex_inst
->size_written
)
2966 } else if (i
!= FB_WRITE_LOGICAL_SRC_COMPONENTS
) {
2967 if (fb_write
->src
[i
].file
!= BAD_FILE
)
2972 assert(!tex_inst
->eot
); /* We can't get here twice */
2973 assert((tex_inst
->offset
& (0xff << 24)) == 0);
2975 const fs_builder
ibld(this, block
, tex_inst
);
2977 tex_inst
->offset
|= fb_write
->target
<< 24;
2978 tex_inst
->eot
= true;
2979 tex_inst
->dst
= ibld
.null_reg_ud();
2980 tex_inst
->size_written
= 0;
2981 fb_write
->remove(cfg
->blocks
[cfg
->num_blocks
- 1]);
2983 /* Marking EOT is sufficient, lower_logical_sends() will notice the EOT
2984 * flag and submit a header together with the sampler message as required
2987 invalidate_live_intervals();
2992 fs_visitor::opt_register_renaming()
2994 bool progress
= false;
2997 unsigned remap
[alloc
.count
];
2998 memset(remap
, ~0u, sizeof(unsigned) * alloc
.count
);
3000 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
3001 if (inst
->opcode
== BRW_OPCODE_IF
|| inst
->opcode
== BRW_OPCODE_DO
) {
3003 } else if (inst
->opcode
== BRW_OPCODE_ENDIF
||
3004 inst
->opcode
== BRW_OPCODE_WHILE
) {
3008 /* Rewrite instruction sources. */
3009 for (int i
= 0; i
< inst
->sources
; i
++) {
3010 if (inst
->src
[i
].file
== VGRF
&&
3011 remap
[inst
->src
[i
].nr
] != ~0u &&
3012 remap
[inst
->src
[i
].nr
] != inst
->src
[i
].nr
) {
3013 inst
->src
[i
].nr
= remap
[inst
->src
[i
].nr
];
3018 const unsigned dst
= inst
->dst
.nr
;
3021 inst
->dst
.file
== VGRF
&&
3022 alloc
.sizes
[inst
->dst
.nr
] * REG_SIZE
== inst
->size_written
&&
3023 !inst
->is_partial_write()) {
3024 if (remap
[dst
] == ~0u) {
3027 remap
[dst
] = alloc
.allocate(regs_written(inst
));
3028 inst
->dst
.nr
= remap
[dst
];
3031 } else if (inst
->dst
.file
== VGRF
&&
3032 remap
[dst
] != ~0u &&
3033 remap
[dst
] != dst
) {
3034 inst
->dst
.nr
= remap
[dst
];
3040 invalidate_live_intervals();
3042 for (unsigned i
= 0; i
< ARRAY_SIZE(delta_xy
); i
++) {
3043 if (delta_xy
[i
].file
== VGRF
&& remap
[delta_xy
[i
].nr
] != ~0u) {
3044 delta_xy
[i
].nr
= remap
[delta_xy
[i
].nr
];
3053 * Remove redundant or useless discard jumps.
3055 * For example, we can eliminate jumps in the following sequence:
3057 * discard-jump (redundant with the next jump)
3058 * discard-jump (useless; jumps to the next instruction)
3062 fs_visitor::opt_redundant_discard_jumps()
3064 bool progress
= false;
3066 bblock_t
*last_bblock
= cfg
->blocks
[cfg
->num_blocks
- 1];
3068 fs_inst
*placeholder_halt
= NULL
;
3069 foreach_inst_in_block_reverse(fs_inst
, inst
, last_bblock
) {
3070 if (inst
->opcode
== FS_OPCODE_PLACEHOLDER_HALT
) {
3071 placeholder_halt
= inst
;
3076 if (!placeholder_halt
)
3079 /* Delete any HALTs immediately before the placeholder halt. */
3080 for (fs_inst
*prev
= (fs_inst
*) placeholder_halt
->prev
;
3081 !prev
->is_head_sentinel() && prev
->opcode
== FS_OPCODE_DISCARD_JUMP
;
3082 prev
= (fs_inst
*) placeholder_halt
->prev
) {
3083 prev
->remove(last_bblock
);
3088 invalidate_live_intervals();
3094 * Compute a bitmask with GRF granularity with a bit set for each GRF starting
3095 * from \p r.offset which overlaps the region starting at \p s.offset and
3096 * spanning \p ds bytes.
3098 static inline unsigned
3099 mask_relative_to(const fs_reg
&r
, const fs_reg
&s
, unsigned ds
)
3101 const int rel_offset
= reg_offset(s
) - reg_offset(r
);
3102 const int shift
= rel_offset
/ REG_SIZE
;
3103 const unsigned n
= DIV_ROUND_UP(rel_offset
% REG_SIZE
+ ds
, REG_SIZE
);
3104 assert(reg_space(r
) == reg_space(s
) &&
3105 shift
>= 0 && shift
< int(8 * sizeof(unsigned)));
3106 return ((1 << n
) - 1) << shift
;
3110 fs_visitor::opt_peephole_csel()
3112 if (devinfo
->gen
< 8)
3115 bool progress
= false;
3117 foreach_block_reverse(block
, cfg
) {
3118 int ip
= block
->end_ip
+ 1;
3120 foreach_inst_in_block_reverse_safe(fs_inst
, inst
, block
) {
3123 if (inst
->opcode
!= BRW_OPCODE_SEL
||
3124 inst
->predicate
!= BRW_PREDICATE_NORMAL
||
3125 (inst
->dst
.type
!= BRW_REGISTER_TYPE_F
&&
3126 inst
->dst
.type
!= BRW_REGISTER_TYPE_D
&&
3127 inst
->dst
.type
!= BRW_REGISTER_TYPE_UD
))
3130 /* Because it is a 3-src instruction, CSEL cannot have an immediate
3131 * value as a source, but we can sometimes handle zero.
3133 if ((inst
->src
[0].file
!= VGRF
&& inst
->src
[0].file
!= ATTR
&&
3134 inst
->src
[0].file
!= UNIFORM
) ||
3135 (inst
->src
[1].file
!= VGRF
&& inst
->src
[1].file
!= ATTR
&&
3136 inst
->src
[1].file
!= UNIFORM
&& !inst
->src
[1].is_zero()))
3139 foreach_inst_in_block_reverse_starting_from(fs_inst
, scan_inst
, inst
) {
3140 if (!scan_inst
->flags_written())
3143 if ((scan_inst
->opcode
!= BRW_OPCODE_CMP
&&
3144 scan_inst
->opcode
!= BRW_OPCODE_MOV
) ||
3145 scan_inst
->predicate
!= BRW_PREDICATE_NONE
||
3146 (scan_inst
->src
[0].file
!= VGRF
&&
3147 scan_inst
->src
[0].file
!= ATTR
&&
3148 scan_inst
->src
[0].file
!= UNIFORM
) ||
3149 scan_inst
->src
[0].type
!= BRW_REGISTER_TYPE_F
)
3152 if (scan_inst
->opcode
== BRW_OPCODE_CMP
&& !scan_inst
->src
[1].is_zero())
3155 const brw::fs_builder
ibld(this, block
, inst
);
3157 const enum brw_conditional_mod cond
=
3158 inst
->predicate_inverse
3159 ? brw_negate_cmod(scan_inst
->conditional_mod
)
3160 : scan_inst
->conditional_mod
;
3162 fs_inst
*csel_inst
= NULL
;
3164 if (inst
->src
[1].file
!= IMM
) {
3165 csel_inst
= ibld
.CSEL(inst
->dst
,
3170 } else if (cond
== BRW_CONDITIONAL_NZ
) {
3171 /* Consider the sequence
3173 * cmp.nz.f0 null<1>F g3<8,8,1>F 0F
3174 * (+f0) sel g124<1>UD g2<8,8,1>UD 0x00000000UD
3176 * The sel will pick the immediate value 0 if r0 is ±0.0.
3177 * Therefore, this sequence is equivalent:
3179 * cmp.nz.f0 null<1>F g3<8,8,1>F 0F
3180 * (+f0) sel g124<1>F g2<8,8,1>F (abs)g3<8,8,1>F
3182 * The abs is ensures that the result is 0UD when g3 is -0.0F.
3183 * By normal cmp-sel merging, this is also equivalent:
3185 * csel.nz g124<1>F g2<4,4,1>F (abs)g3<4,4,1>F g3<4,4,1>F
3187 csel_inst
= ibld
.CSEL(inst
->dst
,
3193 csel_inst
->src
[1].abs
= true;
3196 if (csel_inst
!= NULL
) {
3198 csel_inst
->saturate
= inst
->saturate
;
3199 inst
->remove(block
);
3211 fs_visitor::compute_to_mrf()
3213 bool progress
= false;
3216 /* No MRFs on Gen >= 7. */
3217 if (devinfo
->gen
>= 7)
3220 calculate_live_intervals();
3222 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
3226 if (inst
->opcode
!= BRW_OPCODE_MOV
||
3227 inst
->is_partial_write() ||
3228 inst
->dst
.file
!= MRF
|| inst
->src
[0].file
!= VGRF
||
3229 inst
->dst
.type
!= inst
->src
[0].type
||
3230 inst
->src
[0].abs
|| inst
->src
[0].negate
||
3231 !inst
->src
[0].is_contiguous() ||
3232 inst
->src
[0].offset
% REG_SIZE
!= 0)
3235 /* Can't compute-to-MRF this GRF if someone else was going to
3238 if (this->virtual_grf_end
[inst
->src
[0].nr
] > ip
)
3241 /* Found a move of a GRF to a MRF. Let's see if we can go rewrite the
3242 * things that computed the value of all GRFs of the source region. The
3243 * regs_left bitset keeps track of the registers we haven't yet found a
3244 * generating instruction for.
3246 unsigned regs_left
= (1 << regs_read(inst
, 0)) - 1;
3248 foreach_inst_in_block_reverse_starting_from(fs_inst
, scan_inst
, inst
) {
3249 if (regions_overlap(scan_inst
->dst
, scan_inst
->size_written
,
3250 inst
->src
[0], inst
->size_read(0))) {
3251 /* Found the last thing to write our reg we want to turn
3252 * into a compute-to-MRF.
3255 /* If this one instruction didn't populate all the
3256 * channels, bail. We might be able to rewrite everything
3257 * that writes that reg, but it would require smarter
3260 if (scan_inst
->is_partial_write())
3263 /* Handling things not fully contained in the source of the copy
3264 * would need us to understand coalescing out more than one MOV at
3267 if (!region_contained_in(scan_inst
->dst
, scan_inst
->size_written
,
3268 inst
->src
[0], inst
->size_read(0)))
3271 /* SEND instructions can't have MRF as a destination. */
3272 if (scan_inst
->mlen
)
3275 if (devinfo
->gen
== 6) {
3276 /* gen6 math instructions must have the destination be
3277 * GRF, so no compute-to-MRF for them.
3279 if (scan_inst
->is_math()) {
3284 /* Clear the bits for any registers this instruction overwrites. */
3285 regs_left
&= ~mask_relative_to(
3286 inst
->src
[0], scan_inst
->dst
, scan_inst
->size_written
);
3291 /* We don't handle control flow here. Most computation of
3292 * values that end up in MRFs are shortly before the MRF
3295 if (block
->start() == scan_inst
)
3298 /* You can't read from an MRF, so if someone else reads our
3299 * MRF's source GRF that we wanted to rewrite, that stops us.
3301 bool interfered
= false;
3302 for (int i
= 0; i
< scan_inst
->sources
; i
++) {
3303 if (regions_overlap(scan_inst
->src
[i
], scan_inst
->size_read(i
),
3304 inst
->src
[0], inst
->size_read(0))) {
3311 if (regions_overlap(scan_inst
->dst
, scan_inst
->size_written
,
3312 inst
->dst
, inst
->size_written
)) {
3313 /* If somebody else writes our MRF here, we can't
3314 * compute-to-MRF before that.
3319 if (scan_inst
->mlen
> 0 && scan_inst
->base_mrf
!= -1 &&
3320 regions_overlap(fs_reg(MRF
, scan_inst
->base_mrf
), scan_inst
->mlen
* REG_SIZE
,
3321 inst
->dst
, inst
->size_written
)) {
3322 /* Found a SEND instruction, which means that there are
3323 * live values in MRFs from base_mrf to base_mrf +
3324 * scan_inst->mlen - 1. Don't go pushing our MRF write up
3334 /* Found all generating instructions of our MRF's source value, so it
3335 * should be safe to rewrite them to point to the MRF directly.
3337 regs_left
= (1 << regs_read(inst
, 0)) - 1;
3339 foreach_inst_in_block_reverse_starting_from(fs_inst
, scan_inst
, inst
) {
3340 if (regions_overlap(scan_inst
->dst
, scan_inst
->size_written
,
3341 inst
->src
[0], inst
->size_read(0))) {
3342 /* Clear the bits for any registers this instruction overwrites. */
3343 regs_left
&= ~mask_relative_to(
3344 inst
->src
[0], scan_inst
->dst
, scan_inst
->size_written
);
3346 const unsigned rel_offset
= reg_offset(scan_inst
->dst
) -
3347 reg_offset(inst
->src
[0]);
3349 if (inst
->dst
.nr
& BRW_MRF_COMPR4
) {
3350 /* Apply the same address transformation done by the hardware
3351 * for COMPR4 MRF writes.
3353 assert(rel_offset
< 2 * REG_SIZE
);
3354 scan_inst
->dst
.nr
= inst
->dst
.nr
+ rel_offset
/ REG_SIZE
* 4;
3356 /* Clear the COMPR4 bit if the generating instruction is not
3359 if (scan_inst
->size_written
< 2 * REG_SIZE
)
3360 scan_inst
->dst
.nr
&= ~BRW_MRF_COMPR4
;
3363 /* Calculate the MRF number the result of this instruction is
3364 * ultimately written to.
3366 scan_inst
->dst
.nr
= inst
->dst
.nr
+ rel_offset
/ REG_SIZE
;
3369 scan_inst
->dst
.file
= MRF
;
3370 scan_inst
->dst
.offset
= inst
->dst
.offset
+ rel_offset
% REG_SIZE
;
3371 scan_inst
->saturate
|= inst
->saturate
;
3378 inst
->remove(block
);
3383 invalidate_live_intervals();
3389 * Eliminate FIND_LIVE_CHANNEL instructions occurring outside any control
3390 * flow. We could probably do better here with some form of divergence
3394 fs_visitor::eliminate_find_live_channel()
3396 bool progress
= false;
3399 if (!brw_stage_has_packed_dispatch(devinfo
, stage
, stage_prog_data
)) {
3400 /* The optimization below assumes that channel zero is live on thread
3401 * dispatch, which may not be the case if the fixed function dispatches
3407 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
3408 switch (inst
->opcode
) {
3414 case BRW_OPCODE_ENDIF
:
3415 case BRW_OPCODE_WHILE
:
3419 case FS_OPCODE_DISCARD_JUMP
:
3420 /* This can potentially make control flow non-uniform until the end
3425 case SHADER_OPCODE_FIND_LIVE_CHANNEL
:
3427 inst
->opcode
= BRW_OPCODE_MOV
;
3428 inst
->src
[0] = brw_imm_ud(0u);
3430 inst
->force_writemask_all
= true;
3444 * Once we've generated code, try to convert normal FS_OPCODE_FB_WRITE
3445 * instructions to FS_OPCODE_REP_FB_WRITE.
3448 fs_visitor::emit_repclear_shader()
3450 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
3452 int color_mrf
= base_mrf
+ 2;
3456 mov
= bld
.exec_all().group(4, 0)
3457 .MOV(brw_message_reg(color_mrf
),
3458 fs_reg(UNIFORM
, 0, BRW_REGISTER_TYPE_F
));
3460 struct brw_reg reg
=
3461 brw_reg(BRW_GENERAL_REGISTER_FILE
, 2, 3, 0, 0, BRW_REGISTER_TYPE_F
,
3462 BRW_VERTICAL_STRIDE_8
, BRW_WIDTH_2
, BRW_HORIZONTAL_STRIDE_4
,
3463 BRW_SWIZZLE_XYZW
, WRITEMASK_XYZW
);
3465 mov
= bld
.exec_all().group(4, 0)
3466 .MOV(vec4(brw_message_reg(color_mrf
)), fs_reg(reg
));
3469 fs_inst
*write
= NULL
;
3470 if (key
->nr_color_regions
== 1) {
3471 write
= bld
.emit(FS_OPCODE_REP_FB_WRITE
);
3472 write
->saturate
= key
->clamp_fragment_color
;
3473 write
->base_mrf
= color_mrf
;
3475 write
->header_size
= 0;
3478 assume(key
->nr_color_regions
> 0);
3480 struct brw_reg header
=
3481 retype(brw_message_reg(base_mrf
), BRW_REGISTER_TYPE_UD
);
3482 bld
.exec_all().group(16, 0)
3483 .MOV(header
, retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
));
3485 for (int i
= 0; i
< key
->nr_color_regions
; ++i
) {
3487 bld
.exec_all().group(1, 0)
3488 .MOV(component(header
, 2), brw_imm_ud(i
));
3491 write
= bld
.emit(FS_OPCODE_REP_FB_WRITE
);
3492 write
->saturate
= key
->clamp_fragment_color
;
3493 write
->base_mrf
= base_mrf
;
3495 write
->header_size
= 2;
3500 write
->last_rt
= true;
3504 assign_constant_locations();
3505 assign_curb_setup();
3507 /* Now that we have the uniform assigned, go ahead and force it to a vec4. */
3509 assert(mov
->src
[0].file
== FIXED_GRF
);
3510 mov
->src
[0] = brw_vec4_grf(mov
->src
[0].nr
, 0);
3517 * Walks through basic blocks, looking for repeated MRF writes and
3518 * removing the later ones.
3521 fs_visitor::remove_duplicate_mrf_writes()
3523 fs_inst
*last_mrf_move
[BRW_MAX_MRF(devinfo
->gen
)];
3524 bool progress
= false;
3526 /* Need to update the MRF tracking for compressed instructions. */
3527 if (dispatch_width
>= 16)
3530 memset(last_mrf_move
, 0, sizeof(last_mrf_move
));
3532 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
3533 if (inst
->is_control_flow()) {
3534 memset(last_mrf_move
, 0, sizeof(last_mrf_move
));
3537 if (inst
->opcode
== BRW_OPCODE_MOV
&&
3538 inst
->dst
.file
== MRF
) {
3539 fs_inst
*prev_inst
= last_mrf_move
[inst
->dst
.nr
];
3540 if (prev_inst
&& prev_inst
->opcode
== BRW_OPCODE_MOV
&&
3541 inst
->dst
.equals(prev_inst
->dst
) &&
3542 inst
->src
[0].equals(prev_inst
->src
[0]) &&
3543 inst
->saturate
== prev_inst
->saturate
&&
3544 inst
->predicate
== prev_inst
->predicate
&&
3545 inst
->conditional_mod
== prev_inst
->conditional_mod
&&
3546 inst
->exec_size
== prev_inst
->exec_size
) {
3547 inst
->remove(block
);
3553 /* Clear out the last-write records for MRFs that were overwritten. */
3554 if (inst
->dst
.file
== MRF
) {
3555 last_mrf_move
[inst
->dst
.nr
] = NULL
;
3558 if (inst
->mlen
> 0 && inst
->base_mrf
!= -1) {
3559 /* Found a SEND instruction, which will include two or fewer
3560 * implied MRF writes. We could do better here.
3562 for (int i
= 0; i
< implied_mrf_writes(inst
); i
++) {
3563 last_mrf_move
[inst
->base_mrf
+ i
] = NULL
;
3567 /* Clear out any MRF move records whose sources got overwritten. */
3568 for (unsigned i
= 0; i
< ARRAY_SIZE(last_mrf_move
); i
++) {
3569 if (last_mrf_move
[i
] &&
3570 regions_overlap(inst
->dst
, inst
->size_written
,
3571 last_mrf_move
[i
]->src
[0],
3572 last_mrf_move
[i
]->size_read(0))) {
3573 last_mrf_move
[i
] = NULL
;
3577 if (inst
->opcode
== BRW_OPCODE_MOV
&&
3578 inst
->dst
.file
== MRF
&&
3579 inst
->src
[0].file
!= ARF
&&
3580 !inst
->is_partial_write()) {
3581 last_mrf_move
[inst
->dst
.nr
] = inst
;
3586 invalidate_live_intervals();
3592 * Rounding modes for conversion instructions are included for each
3593 * conversion, but right now it is a state. So once it is set,
3594 * we don't need to call it again for subsequent calls.
3596 * This is useful for vector/matrices conversions, as setting the
3597 * mode once is enough for the full vector/matrix
3600 fs_visitor::remove_extra_rounding_modes()
3602 bool progress
= false;
3603 unsigned execution_mode
= this->nir
->info
.float_controls_execution_mode
;
3605 brw_rnd_mode base_mode
= BRW_RND_MODE_UNSPECIFIED
;
3606 if ((FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
|
3607 FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
|
3608 FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
) &
3610 base_mode
= BRW_RND_MODE_RTNE
;
3611 if ((FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
|
3612 FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
|
3613 FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
) &
3615 base_mode
= BRW_RND_MODE_RTZ
;
3617 foreach_block (block
, cfg
) {
3618 brw_rnd_mode prev_mode
= base_mode
;
3620 foreach_inst_in_block_safe (fs_inst
, inst
, block
) {
3621 if (inst
->opcode
== SHADER_OPCODE_RND_MODE
) {
3622 assert(inst
->src
[0].file
== BRW_IMMEDIATE_VALUE
);
3623 const brw_rnd_mode mode
= (brw_rnd_mode
) inst
->src
[0].d
;
3624 if (mode
== prev_mode
) {
3625 inst
->remove(block
);
3635 invalidate_live_intervals();
3641 clear_deps_for_inst_src(fs_inst
*inst
, bool *deps
, int first_grf
, int grf_len
)
3643 /* Clear the flag for registers that actually got read (as expected). */
3644 for (int i
= 0; i
< inst
->sources
; i
++) {
3646 if (inst
->src
[i
].file
== VGRF
|| inst
->src
[i
].file
== FIXED_GRF
) {
3647 grf
= inst
->src
[i
].nr
;
3652 if (grf
>= first_grf
&&
3653 grf
< first_grf
+ grf_len
) {
3654 deps
[grf
- first_grf
] = false;
3655 if (inst
->exec_size
== 16)
3656 deps
[grf
- first_grf
+ 1] = false;
3662 * Implements this workaround for the original 965:
3664 * "[DevBW, DevCL] Implementation Restrictions: As the hardware does not
3665 * check for post destination dependencies on this instruction, software
3666 * must ensure that there is no destination hazard for the case of ‘write
3667 * followed by a posted write’ shown in the following example.
3670 * 2. send r3.xy <rest of send instruction>
3673 * Due to no post-destination dependency check on the ‘send’, the above
3674 * code sequence could have two instructions (1 and 2) in flight at the
3675 * same time that both consider ‘r3’ as the target of their final writes.
3678 fs_visitor::insert_gen4_pre_send_dependency_workarounds(bblock_t
*block
,
3681 int write_len
= regs_written(inst
);
3682 int first_write_grf
= inst
->dst
.nr
;
3683 bool needs_dep
[BRW_MAX_MRF(devinfo
->gen
)];
3684 assert(write_len
< (int)sizeof(needs_dep
) - 1);
3686 memset(needs_dep
, false, sizeof(needs_dep
));
3687 memset(needs_dep
, true, write_len
);
3689 clear_deps_for_inst_src(inst
, needs_dep
, first_write_grf
, write_len
);
3691 /* Walk backwards looking for writes to registers we're writing which
3692 * aren't read since being written. If we hit the start of the program,
3693 * we assume that there are no outstanding dependencies on entry to the
3696 foreach_inst_in_block_reverse_starting_from(fs_inst
, scan_inst
, inst
) {
3697 /* If we hit control flow, assume that there *are* outstanding
3698 * dependencies, and force their cleanup before our instruction.
3700 if (block
->start() == scan_inst
&& block
->num
!= 0) {
3701 for (int i
= 0; i
< write_len
; i
++) {
3703 DEP_RESOLVE_MOV(fs_builder(this, block
, inst
),
3704 first_write_grf
+ i
);
3709 /* We insert our reads as late as possible on the assumption that any
3710 * instruction but a MOV that might have left us an outstanding
3711 * dependency has more latency than a MOV.
3713 if (scan_inst
->dst
.file
== VGRF
) {
3714 for (unsigned i
= 0; i
< regs_written(scan_inst
); i
++) {
3715 int reg
= scan_inst
->dst
.nr
+ i
;
3717 if (reg
>= first_write_grf
&&
3718 reg
< first_write_grf
+ write_len
&&
3719 needs_dep
[reg
- first_write_grf
]) {
3720 DEP_RESOLVE_MOV(fs_builder(this, block
, inst
), reg
);
3721 needs_dep
[reg
- first_write_grf
] = false;
3722 if (scan_inst
->exec_size
== 16)
3723 needs_dep
[reg
- first_write_grf
+ 1] = false;
3728 /* Clear the flag for registers that actually got read (as expected). */
3729 clear_deps_for_inst_src(scan_inst
, needs_dep
, first_write_grf
, write_len
);
3731 /* Continue the loop only if we haven't resolved all the dependencies */
3733 for (i
= 0; i
< write_len
; i
++) {
3743 * Implements this workaround for the original 965:
3745 * "[DevBW, DevCL] Errata: A destination register from a send can not be
3746 * used as a destination register until after it has been sourced by an
3747 * instruction with a different destination register.
3750 fs_visitor::insert_gen4_post_send_dependency_workarounds(bblock_t
*block
, fs_inst
*inst
)
3752 int write_len
= regs_written(inst
);
3753 unsigned first_write_grf
= inst
->dst
.nr
;
3754 bool needs_dep
[BRW_MAX_MRF(devinfo
->gen
)];
3755 assert(write_len
< (int)sizeof(needs_dep
) - 1);
3757 memset(needs_dep
, false, sizeof(needs_dep
));
3758 memset(needs_dep
, true, write_len
);
3759 /* Walk forwards looking for writes to registers we're writing which aren't
3760 * read before being written.
3762 foreach_inst_in_block_starting_from(fs_inst
, scan_inst
, inst
) {
3763 /* If we hit control flow, force resolve all remaining dependencies. */
3764 if (block
->end() == scan_inst
&& block
->num
!= cfg
->num_blocks
- 1) {
3765 for (int i
= 0; i
< write_len
; i
++) {
3767 DEP_RESOLVE_MOV(fs_builder(this, block
, scan_inst
),
3768 first_write_grf
+ i
);
3773 /* Clear the flag for registers that actually got read (as expected). */
3774 clear_deps_for_inst_src(scan_inst
, needs_dep
, first_write_grf
, write_len
);
3776 /* We insert our reads as late as possible since they're reading the
3777 * result of a SEND, which has massive latency.
3779 if (scan_inst
->dst
.file
== VGRF
&&
3780 scan_inst
->dst
.nr
>= first_write_grf
&&
3781 scan_inst
->dst
.nr
< first_write_grf
+ write_len
&&
3782 needs_dep
[scan_inst
->dst
.nr
- first_write_grf
]) {
3783 DEP_RESOLVE_MOV(fs_builder(this, block
, scan_inst
),
3785 needs_dep
[scan_inst
->dst
.nr
- first_write_grf
] = false;
3788 /* Continue the loop only if we haven't resolved all the dependencies */
3790 for (i
= 0; i
< write_len
; i
++) {
3800 fs_visitor::insert_gen4_send_dependency_workarounds()
3802 if (devinfo
->gen
!= 4 || devinfo
->is_g4x
)
3805 bool progress
= false;
3807 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
3808 if (inst
->mlen
!= 0 && inst
->dst
.file
== VGRF
) {
3809 insert_gen4_pre_send_dependency_workarounds(block
, inst
);
3810 insert_gen4_post_send_dependency_workarounds(block
, inst
);
3816 invalidate_live_intervals();
3820 * Turns the generic expression-style uniform pull constant load instruction
3821 * into a hardware-specific series of instructions for loading a pull
3824 * The expression style allows the CSE pass before this to optimize out
3825 * repeated loads from the same offset, and gives the pre-register-allocation
3826 * scheduling full flexibility, while the conversion to native instructions
3827 * allows the post-register-allocation scheduler the best information
3830 * Note that execution masking for setting up pull constant loads is special:
3831 * the channels that need to be written are unrelated to the current execution
3832 * mask, since a later instruction will use one of the result channels as a
3833 * source operand for all 8 or 16 of its channels.
3836 fs_visitor::lower_uniform_pull_constant_loads()
3838 foreach_block_and_inst (block
, fs_inst
, inst
, cfg
) {
3839 if (inst
->opcode
!= FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
)
3842 if (devinfo
->gen
>= 7) {
3843 const fs_builder ubld
= fs_builder(this, block
, inst
).exec_all();
3844 const fs_reg payload
= ubld
.group(8, 0).vgrf(BRW_REGISTER_TYPE_UD
);
3846 ubld
.group(8, 0).MOV(payload
,
3847 retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
));
3848 ubld
.group(1, 0).MOV(component(payload
, 2),
3849 brw_imm_ud(inst
->src
[1].ud
/ 16));
3851 inst
->opcode
= FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7
;
3852 inst
->src
[1] = payload
;
3853 inst
->header_size
= 1;
3856 invalidate_live_intervals();
3858 /* Before register allocation, we didn't tell the scheduler about the
3859 * MRF we use. We know it's safe to use this MRF because nothing
3860 * else does except for register spill/unspill, which generates and
3861 * uses its MRF within a single IR instruction.
3863 inst
->base_mrf
= FIRST_PULL_LOAD_MRF(devinfo
->gen
) + 1;
3870 fs_visitor::lower_load_payload()
3872 bool progress
= false;
3874 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
3875 if (inst
->opcode
!= SHADER_OPCODE_LOAD_PAYLOAD
)
3878 assert(inst
->dst
.file
== MRF
|| inst
->dst
.file
== VGRF
);
3879 assert(inst
->saturate
== false);
3880 fs_reg dst
= inst
->dst
;
3882 /* Get rid of COMPR4. We'll add it back in if we need it */
3883 if (dst
.file
== MRF
)
3884 dst
.nr
= dst
.nr
& ~BRW_MRF_COMPR4
;
3886 const fs_builder
ibld(this, block
, inst
);
3887 const fs_builder hbld
= ibld
.exec_all().group(8, 0);
3889 for (uint8_t i
= 0; i
< inst
->header_size
; i
++) {
3890 if (inst
->src
[i
].file
!= BAD_FILE
) {
3891 fs_reg mov_dst
= retype(dst
, BRW_REGISTER_TYPE_UD
);
3892 fs_reg mov_src
= retype(inst
->src
[i
], BRW_REGISTER_TYPE_UD
);
3893 hbld
.MOV(mov_dst
, mov_src
);
3895 dst
= offset(dst
, hbld
, 1);
3898 if (inst
->dst
.file
== MRF
&& (inst
->dst
.nr
& BRW_MRF_COMPR4
) &&
3899 inst
->exec_size
> 8) {
3900 /* In this case, the payload portion of the LOAD_PAYLOAD isn't
3901 * a straightforward copy. Instead, the result of the
3902 * LOAD_PAYLOAD is treated as interleaved and the first four
3903 * non-header sources are unpacked as:
3914 * This is used for gen <= 5 fb writes.
3916 assert(inst
->exec_size
== 16);
3917 assert(inst
->header_size
+ 4 <= inst
->sources
);
3918 for (uint8_t i
= inst
->header_size
; i
< inst
->header_size
+ 4; i
++) {
3919 if (inst
->src
[i
].file
!= BAD_FILE
) {
3920 if (devinfo
->has_compr4
) {
3921 fs_reg compr4_dst
= retype(dst
, inst
->src
[i
].type
);
3922 compr4_dst
.nr
|= BRW_MRF_COMPR4
;
3923 ibld
.MOV(compr4_dst
, inst
->src
[i
]);
3925 /* Platform doesn't have COMPR4. We have to fake it */
3926 fs_reg mov_dst
= retype(dst
, inst
->src
[i
].type
);
3927 ibld
.half(0).MOV(mov_dst
, half(inst
->src
[i
], 0));
3929 ibld
.half(1).MOV(mov_dst
, half(inst
->src
[i
], 1));
3936 /* The loop above only ever incremented us through the first set
3937 * of 4 registers. However, thanks to the magic of COMPR4, we
3938 * actually wrote to the first 8 registers, so we need to take
3939 * that into account now.
3943 /* The COMPR4 code took care of the first 4 sources. We'll let
3944 * the regular path handle any remaining sources. Yes, we are
3945 * modifying the instruction but we're about to delete it so
3946 * this really doesn't hurt anything.
3948 inst
->header_size
+= 4;
3951 for (uint8_t i
= inst
->header_size
; i
< inst
->sources
; i
++) {
3952 if (inst
->src
[i
].file
!= BAD_FILE
) {
3953 dst
.type
= inst
->src
[i
].type
;
3954 ibld
.MOV(dst
, inst
->src
[i
]);
3956 dst
.type
= BRW_REGISTER_TYPE_UD
;
3958 dst
= offset(dst
, ibld
, 1);
3961 inst
->remove(block
);
3966 invalidate_live_intervals();
3972 fs_visitor::lower_mul_dword_inst(fs_inst
*inst
, bblock_t
*block
)
3974 const fs_builder
ibld(this, block
, inst
);
3976 if (inst
->src
[1].file
== IMM
&& inst
->src
[1].ud
< (1 << 16)) {
3977 /* The MUL instruction isn't commutative. On Gen <= 6, only the low
3978 * 16-bits of src0 are read, and on Gen >= 7 only the low 16-bits of
3981 * If multiplying by an immediate value that fits in 16-bits, do a
3982 * single MUL instruction with that value in the proper location.
3984 if (devinfo
->gen
< 7) {
3985 fs_reg
imm(VGRF
, alloc
.allocate(dispatch_width
/ 8), inst
->dst
.type
);
3986 ibld
.MOV(imm
, inst
->src
[1]);
3987 ibld
.MUL(inst
->dst
, imm
, inst
->src
[0]);
3989 const bool ud
= (inst
->src
[1].type
== BRW_REGISTER_TYPE_UD
);
3990 ibld
.MUL(inst
->dst
, inst
->src
[0],
3991 ud
? brw_imm_uw(inst
->src
[1].ud
)
3992 : brw_imm_w(inst
->src
[1].d
));
3995 /* Gen < 8 (and some Gen8+ low-power parts like Cherryview) cannot
3996 * do 32-bit integer multiplication in one instruction, but instead
3997 * must do a sequence (which actually calculates a 64-bit result):
3999 * mul(8) acc0<1>D g3<8,8,1>D g4<8,8,1>D
4000 * mach(8) null g3<8,8,1>D g4<8,8,1>D
4001 * mov(8) g2<1>D acc0<8,8,1>D
4003 * But on Gen > 6, the ability to use second accumulator register
4004 * (acc1) for non-float data types was removed, preventing a simple
4005 * implementation in SIMD16. A 16-channel result can be calculated by
4006 * executing the three instructions twice in SIMD8, once with quarter
4007 * control of 1Q for the first eight channels and again with 2Q for
4008 * the second eight channels.
4010 * Which accumulator register is implicitly accessed (by AccWrEnable
4011 * for instance) is determined by the quarter control. Unfortunately
4012 * Ivybridge (and presumably Baytrail) has a hardware bug in which an
4013 * implicit accumulator access by an instruction with 2Q will access
4014 * acc1 regardless of whether the data type is usable in acc1.
4016 * Specifically, the 2Q mach(8) writes acc1 which does not exist for
4017 * integer data types.
4019 * Since we only want the low 32-bits of the result, we can do two
4020 * 32-bit x 16-bit multiplies (like the mul and mach are doing), and
4021 * adjust the high result and add them (like the mach is doing):
4023 * mul(8) g7<1>D g3<8,8,1>D g4.0<8,8,1>UW
4024 * mul(8) g8<1>D g3<8,8,1>D g4.1<8,8,1>UW
4025 * shl(8) g9<1>D g8<8,8,1>D 16D
4026 * add(8) g2<1>D g7<8,8,1>D g8<8,8,1>D
4028 * We avoid the shl instruction by realizing that we only want to add
4029 * the low 16-bits of the "high" result to the high 16-bits of the
4030 * "low" result and using proper regioning on the add:
4032 * mul(8) g7<1>D g3<8,8,1>D g4.0<16,8,2>UW
4033 * mul(8) g8<1>D g3<8,8,1>D g4.1<16,8,2>UW
4034 * add(8) g7.1<2>UW g7.1<16,8,2>UW g8<16,8,2>UW
4036 * Since it does not use the (single) accumulator register, we can
4037 * schedule multi-component multiplications much better.
4040 bool needs_mov
= false;
4041 fs_reg orig_dst
= inst
->dst
;
4043 /* Get a new VGRF for the "low" 32x16-bit multiplication result if
4044 * reusing the original destination is impossible due to hardware
4045 * restrictions, source/destination overlap, or it being the null
4048 fs_reg low
= inst
->dst
;
4049 if (orig_dst
.is_null() || orig_dst
.file
== MRF
||
4050 regions_overlap(inst
->dst
, inst
->size_written
,
4051 inst
->src
[0], inst
->size_read(0)) ||
4052 regions_overlap(inst
->dst
, inst
->size_written
,
4053 inst
->src
[1], inst
->size_read(1)) ||
4054 inst
->dst
.stride
>= 4) {
4056 low
= fs_reg(VGRF
, alloc
.allocate(regs_written(inst
)),
4060 /* Get a new VGRF but keep the same stride as inst->dst */
4061 fs_reg
high(VGRF
, alloc
.allocate(regs_written(inst
)), inst
->dst
.type
);
4062 high
.stride
= inst
->dst
.stride
;
4063 high
.offset
= inst
->dst
.offset
% REG_SIZE
;
4065 if (devinfo
->gen
>= 7) {
4066 if (inst
->src
[1].abs
)
4067 lower_src_modifiers(this, block
, inst
, 1);
4069 if (inst
->src
[1].file
== IMM
) {
4070 ibld
.MUL(low
, inst
->src
[0],
4071 brw_imm_uw(inst
->src
[1].ud
& 0xffff));
4072 ibld
.MUL(high
, inst
->src
[0],
4073 brw_imm_uw(inst
->src
[1].ud
>> 16));
4075 ibld
.MUL(low
, inst
->src
[0],
4076 subscript(inst
->src
[1], BRW_REGISTER_TYPE_UW
, 0));
4077 ibld
.MUL(high
, inst
->src
[0],
4078 subscript(inst
->src
[1], BRW_REGISTER_TYPE_UW
, 1));
4081 if (inst
->src
[0].abs
)
4082 lower_src_modifiers(this, block
, inst
, 0);
4084 ibld
.MUL(low
, subscript(inst
->src
[0], BRW_REGISTER_TYPE_UW
, 0),
4086 ibld
.MUL(high
, subscript(inst
->src
[0], BRW_REGISTER_TYPE_UW
, 1),
4090 ibld
.ADD(subscript(low
, BRW_REGISTER_TYPE_UW
, 1),
4091 subscript(low
, BRW_REGISTER_TYPE_UW
, 1),
4092 subscript(high
, BRW_REGISTER_TYPE_UW
, 0));
4094 if (needs_mov
|| inst
->conditional_mod
)
4095 set_condmod(inst
->conditional_mod
, ibld
.MOV(orig_dst
, low
));
4100 fs_visitor::lower_mul_qword_inst(fs_inst
*inst
, bblock_t
*block
)
4102 const fs_builder
ibld(this, block
, inst
);
4104 /* Considering two 64-bit integers ab and cd where each letter ab
4105 * corresponds to 32 bits, we get a 128-bit result WXYZ. We * cd
4106 * only need to provide the YZ part of the result. -------
4108 * Only BD needs to be 64 bits. For AD and BC we only care + AD
4109 * about the lower 32 bits (since they are part of the upper + BC
4110 * 32 bits of our result). AC is not needed since it starts + AC
4111 * on the 65th bit of the result. -------
4114 unsigned int q_regs
= regs_written(inst
);
4115 unsigned int d_regs
= (q_regs
+ 1) / 2;
4117 fs_reg
bd(VGRF
, alloc
.allocate(q_regs
), BRW_REGISTER_TYPE_UQ
);
4118 fs_reg
ad(VGRF
, alloc
.allocate(d_regs
), BRW_REGISTER_TYPE_UD
);
4119 fs_reg
bc(VGRF
, alloc
.allocate(d_regs
), BRW_REGISTER_TYPE_UD
);
4121 /* Here we need the full 64 bit result for 32b * 32b. */
4122 if (devinfo
->has_integer_dword_mul
) {
4123 ibld
.MUL(bd
, subscript(inst
->src
[0], BRW_REGISTER_TYPE_UD
, 0),
4124 subscript(inst
->src
[1], BRW_REGISTER_TYPE_UD
, 0));
4126 fs_reg
bd_high(VGRF
, alloc
.allocate(d_regs
), BRW_REGISTER_TYPE_UD
);
4127 fs_reg
bd_low(VGRF
, alloc
.allocate(d_regs
), BRW_REGISTER_TYPE_UD
);
4128 fs_reg acc
= retype(brw_acc_reg(inst
->exec_size
), BRW_REGISTER_TYPE_UD
);
4130 fs_inst
*mul
= ibld
.MUL(acc
,
4131 subscript(inst
->src
[0], BRW_REGISTER_TYPE_UD
, 0),
4132 subscript(inst
->src
[1], BRW_REGISTER_TYPE_UW
, 0));
4133 mul
->writes_accumulator
= true;
4135 ibld
.MACH(bd_high
, subscript(inst
->src
[0], BRW_REGISTER_TYPE_UD
, 0),
4136 subscript(inst
->src
[1], BRW_REGISTER_TYPE_UD
, 0));
4137 ibld
.MOV(bd_low
, acc
);
4139 ibld
.MOV(subscript(bd
, BRW_REGISTER_TYPE_UD
, 0), bd_low
);
4140 ibld
.MOV(subscript(bd
, BRW_REGISTER_TYPE_UD
, 1), bd_high
);
4143 ibld
.MUL(ad
, subscript(inst
->src
[0], BRW_REGISTER_TYPE_UD
, 1),
4144 subscript(inst
->src
[1], BRW_REGISTER_TYPE_UD
, 0));
4145 ibld
.MUL(bc
, subscript(inst
->src
[0], BRW_REGISTER_TYPE_UD
, 0),
4146 subscript(inst
->src
[1], BRW_REGISTER_TYPE_UD
, 1));
4148 ibld
.ADD(ad
, ad
, bc
);
4149 ibld
.ADD(subscript(bd
, BRW_REGISTER_TYPE_UD
, 1),
4150 subscript(bd
, BRW_REGISTER_TYPE_UD
, 1), ad
);
4152 ibld
.MOV(inst
->dst
, bd
);
4156 fs_visitor::lower_mulh_inst(fs_inst
*inst
, bblock_t
*block
)
4158 const fs_builder
ibld(this, block
, inst
);
4160 /* According to the BDW+ BSpec page for the "Multiply Accumulate
4161 * High" instruction:
4163 * "An added preliminary mov is required for source modification on
4165 * mov (8) r3.0<1>:d -r3<8;8,1>:d
4166 * mul (8) acc0:d r2.0<8;8,1>:d r3.0<16;8,2>:uw
4167 * mach (8) r5.0<1>:d r2.0<8;8,1>:d r3.0<8;8,1>:d"
4169 if (devinfo
->gen
>= 8 && (inst
->src
[1].negate
|| inst
->src
[1].abs
))
4170 lower_src_modifiers(this, block
, inst
, 1);
4172 /* Should have been lowered to 8-wide. */
4173 assert(inst
->exec_size
<= get_lowered_simd_width(devinfo
, inst
));
4174 const fs_reg acc
= retype(brw_acc_reg(inst
->exec_size
), inst
->dst
.type
);
4175 fs_inst
*mul
= ibld
.MUL(acc
, inst
->src
[0], inst
->src
[1]);
4176 fs_inst
*mach
= ibld
.MACH(inst
->dst
, inst
->src
[0], inst
->src
[1]);
4178 if (devinfo
->gen
>= 8) {
4179 /* Until Gen8, integer multiplies read 32-bits from one source,
4180 * and 16-bits from the other, and relying on the MACH instruction
4181 * to generate the high bits of the result.
4183 * On Gen8, the multiply instruction does a full 32x32-bit
4184 * multiply, but in order to do a 64-bit multiply we can simulate
4185 * the previous behavior and then use a MACH instruction.
4187 assert(mul
->src
[1].type
== BRW_REGISTER_TYPE_D
||
4188 mul
->src
[1].type
== BRW_REGISTER_TYPE_UD
);
4189 mul
->src
[1].type
= BRW_REGISTER_TYPE_UW
;
4190 mul
->src
[1].stride
*= 2;
4192 if (mul
->src
[1].file
== IMM
) {
4193 mul
->src
[1] = brw_imm_uw(mul
->src
[1].ud
);
4195 } else if (devinfo
->gen
== 7 && !devinfo
->is_haswell
&&
4197 /* Among other things the quarter control bits influence which
4198 * accumulator register is used by the hardware for instructions
4199 * that access the accumulator implicitly (e.g. MACH). A
4200 * second-half instruction would normally map to acc1, which
4201 * doesn't exist on Gen7 and up (the hardware does emulate it for
4202 * floating-point instructions *only* by taking advantage of the
4203 * extra precision of acc0 not normally used for floating point
4206 * HSW and up are careful enough not to try to access an
4207 * accumulator register that doesn't exist, but on earlier Gen7
4208 * hardware we need to make sure that the quarter control bits are
4209 * zero to avoid non-deterministic behaviour and emit an extra MOV
4210 * to get the result masked correctly according to the current
4214 mach
->force_writemask_all
= true;
4215 mach
->dst
= ibld
.vgrf(inst
->dst
.type
);
4216 ibld
.MOV(inst
->dst
, mach
->dst
);
4221 fs_visitor::lower_integer_multiplication()
4223 bool progress
= false;
4225 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
4226 if (inst
->opcode
== BRW_OPCODE_MUL
) {
4227 if ((inst
->dst
.type
== BRW_REGISTER_TYPE_Q
||
4228 inst
->dst
.type
== BRW_REGISTER_TYPE_UQ
) &&
4229 (inst
->src
[0].type
== BRW_REGISTER_TYPE_Q
||
4230 inst
->src
[0].type
== BRW_REGISTER_TYPE_UQ
) &&
4231 (inst
->src
[1].type
== BRW_REGISTER_TYPE_Q
||
4232 inst
->src
[1].type
== BRW_REGISTER_TYPE_UQ
)) {
4233 lower_mul_qword_inst(inst
, block
);
4234 inst
->remove(block
);
4236 } else if (!inst
->dst
.is_accumulator() &&
4237 (inst
->dst
.type
== BRW_REGISTER_TYPE_D
||
4238 inst
->dst
.type
== BRW_REGISTER_TYPE_UD
) &&
4239 !devinfo
->has_integer_dword_mul
) {
4240 lower_mul_dword_inst(inst
, block
);
4241 inst
->remove(block
);
4244 } else if (inst
->opcode
== SHADER_OPCODE_MULH
) {
4245 lower_mulh_inst(inst
, block
);
4246 inst
->remove(block
);
4253 invalidate_live_intervals();
4259 fs_visitor::lower_minmax()
4261 assert(devinfo
->gen
< 6);
4263 bool progress
= false;
4265 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
4266 const fs_builder
ibld(this, block
, inst
);
4268 if (inst
->opcode
== BRW_OPCODE_SEL
&&
4269 inst
->predicate
== BRW_PREDICATE_NONE
) {
4270 /* FIXME: Using CMP doesn't preserve the NaN propagation semantics of
4271 * the original SEL.L/GE instruction
4273 ibld
.CMP(ibld
.null_reg_d(), inst
->src
[0], inst
->src
[1],
4274 inst
->conditional_mod
);
4275 inst
->predicate
= BRW_PREDICATE_NORMAL
;
4276 inst
->conditional_mod
= BRW_CONDITIONAL_NONE
;
4283 invalidate_live_intervals();
4289 setup_color_payload(const fs_builder
&bld
, const brw_wm_prog_key
*key
,
4290 fs_reg
*dst
, fs_reg color
, unsigned components
)
4292 if (key
->clamp_fragment_color
) {
4293 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 4);
4294 assert(color
.type
== BRW_REGISTER_TYPE_F
);
4296 for (unsigned i
= 0; i
< components
; i
++)
4298 bld
.MOV(offset(tmp
, bld
, i
), offset(color
, bld
, i
)));
4303 for (unsigned i
= 0; i
< components
; i
++)
4304 dst
[i
] = offset(color
, bld
, i
);
4308 brw_fb_write_msg_control(const fs_inst
*inst
,
4309 const struct brw_wm_prog_data
*prog_data
)
4313 if (inst
->opcode
== FS_OPCODE_REP_FB_WRITE
) {
4314 assert(inst
->group
== 0 && inst
->exec_size
== 16);
4315 mctl
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE_REPLICATED
;
4316 } else if (prog_data
->dual_src_blend
) {
4317 assert(inst
->exec_size
== 8);
4319 if (inst
->group
% 16 == 0)
4320 mctl
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_DUAL_SOURCE_SUBSPAN01
;
4321 else if (inst
->group
% 16 == 8)
4322 mctl
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_DUAL_SOURCE_SUBSPAN23
;
4324 unreachable("Invalid dual-source FB write instruction group");
4326 assert(inst
->group
== 0 || (inst
->group
== 16 && inst
->exec_size
== 16));
4328 if (inst
->exec_size
== 16)
4329 mctl
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE
;
4330 else if (inst
->exec_size
== 8)
4331 mctl
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_SINGLE_SOURCE_SUBSPAN01
;
4333 unreachable("Invalid FB write execution size");
4340 lower_fb_write_logical_send(const fs_builder
&bld
, fs_inst
*inst
,
4341 const struct brw_wm_prog_data
*prog_data
,
4342 const brw_wm_prog_key
*key
,
4343 const fs_visitor::thread_payload
&payload
)
4345 assert(inst
->src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].file
== IMM
);
4346 const gen_device_info
*devinfo
= bld
.shader
->devinfo
;
4347 const fs_reg
&color0
= inst
->src
[FB_WRITE_LOGICAL_SRC_COLOR0
];
4348 const fs_reg
&color1
= inst
->src
[FB_WRITE_LOGICAL_SRC_COLOR1
];
4349 const fs_reg
&src0_alpha
= inst
->src
[FB_WRITE_LOGICAL_SRC_SRC0_ALPHA
];
4350 const fs_reg
&src_depth
= inst
->src
[FB_WRITE_LOGICAL_SRC_SRC_DEPTH
];
4351 const fs_reg
&dst_depth
= inst
->src
[FB_WRITE_LOGICAL_SRC_DST_DEPTH
];
4352 const fs_reg
&src_stencil
= inst
->src
[FB_WRITE_LOGICAL_SRC_SRC_STENCIL
];
4353 fs_reg sample_mask
= inst
->src
[FB_WRITE_LOGICAL_SRC_OMASK
];
4354 const unsigned components
=
4355 inst
->src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].ud
;
4357 /* We can potentially have a message length of up to 15, so we have to set
4358 * base_mrf to either 0 or 1 in order to fit in m0..m15.
4361 int header_size
= 2, payload_header_size
;
4362 unsigned length
= 0;
4364 if (devinfo
->gen
< 6) {
4365 /* TODO: Support SIMD32 on gen4-5 */
4366 assert(bld
.group() < 16);
4368 /* For gen4-5, we always have a header consisting of g0 and g1. We have
4369 * an implied MOV from g0,g1 to the start of the message. The MOV from
4370 * g0 is handled by the hardware and the MOV from g1 is provided by the
4371 * generator. This is required because, on gen4-5, the generator may
4372 * generate two write messages with different message lengths in order
4373 * to handle AA data properly.
4375 * Also, since the pixel mask goes in the g0 portion of the message and
4376 * since render target writes are the last thing in the shader, we write
4377 * the pixel mask directly into g0 and it will get copied as part of the
4380 if (prog_data
->uses_kill
) {
4381 bld
.exec_all().group(1, 0)
4382 .MOV(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW
),
4383 brw_flag_reg(0, 1));
4386 assert(length
== 0);
4388 } else if ((devinfo
->gen
<= 7 && !devinfo
->is_haswell
&&
4389 prog_data
->uses_kill
) ||
4390 (devinfo
->gen
< 11 &&
4391 (color1
.file
!= BAD_FILE
|| key
->nr_color_regions
> 1))) {
4392 /* From the Sandy Bridge PRM, volume 4, page 198:
4394 * "Dispatched Pixel Enables. One bit per pixel indicating
4395 * which pixels were originally enabled when the thread was
4396 * dispatched. This field is only required for the end-of-
4397 * thread message and on all dual-source messages."
4399 const fs_builder ubld
= bld
.exec_all().group(8, 0);
4401 fs_reg header
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
4402 if (bld
.group() < 16) {
4403 /* The header starts off as g0 and g1 for the first half */
4404 ubld
.group(16, 0).MOV(header
, retype(brw_vec8_grf(0, 0),
4405 BRW_REGISTER_TYPE_UD
));
4407 /* The header starts off as g0 and g2 for the second half */
4408 assert(bld
.group() < 32);
4409 const fs_reg header_sources
[2] = {
4410 retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
),
4411 retype(brw_vec8_grf(2, 0), BRW_REGISTER_TYPE_UD
),
4413 ubld
.LOAD_PAYLOAD(header
, header_sources
, 2, 0);
4416 uint32_t g00_bits
= 0;
4418 /* Set "Source0 Alpha Present to RenderTarget" bit in message
4421 if (inst
->target
> 0 && prog_data
->replicate_alpha
)
4422 g00_bits
|= 1 << 11;
4424 /* Set computes stencil to render target */
4425 if (prog_data
->computed_stencil
)
4426 g00_bits
|= 1 << 14;
4429 /* OR extra bits into g0.0 */
4430 ubld
.group(1, 0).OR(component(header
, 0),
4431 retype(brw_vec1_grf(0, 0),
4432 BRW_REGISTER_TYPE_UD
),
4433 brw_imm_ud(g00_bits
));
4436 /* Set the render target index for choosing BLEND_STATE. */
4437 if (inst
->target
> 0) {
4438 ubld
.group(1, 0).MOV(component(header
, 2), brw_imm_ud(inst
->target
));
4441 if (prog_data
->uses_kill
) {
4442 assert(bld
.group() < 16);
4443 ubld
.group(1, 0).MOV(retype(component(header
, 15),
4444 BRW_REGISTER_TYPE_UW
),
4445 brw_flag_reg(0, 1));
4448 assert(length
== 0);
4449 sources
[0] = header
;
4450 sources
[1] = horiz_offset(header
, 8);
4453 assert(length
== 0 || length
== 2);
4454 header_size
= length
;
4456 if (payload
.aa_dest_stencil_reg
[0]) {
4457 assert(inst
->group
< 16);
4458 sources
[length
] = fs_reg(VGRF
, bld
.shader
->alloc
.allocate(1));
4459 bld
.group(8, 0).exec_all().annotate("FB write stencil/AA alpha")
4460 .MOV(sources
[length
],
4461 fs_reg(brw_vec8_grf(payload
.aa_dest_stencil_reg
[0], 0)));
4465 bool src0_alpha_present
= false;
4467 if (src0_alpha
.file
!= BAD_FILE
) {
4468 for (unsigned i
= 0; i
< bld
.dispatch_width() / 8; i
++) {
4469 const fs_builder
&ubld
= bld
.exec_all().group(8, i
)
4470 .annotate("FB write src0 alpha");
4471 const fs_reg tmp
= ubld
.vgrf(BRW_REGISTER_TYPE_F
);
4472 ubld
.MOV(tmp
, horiz_offset(src0_alpha
, i
* 8));
4473 setup_color_payload(ubld
, key
, &sources
[length
], tmp
, 1);
4476 src0_alpha_present
= true;
4477 } else if (prog_data
->replicate_alpha
&& inst
->target
!= 0) {
4478 /* Handle the case when fragment shader doesn't write to draw buffer
4479 * zero. No need to call setup_color_payload() for src0_alpha because
4480 * alpha value will be undefined.
4482 length
+= bld
.dispatch_width() / 8;
4483 src0_alpha_present
= true;
4486 if (sample_mask
.file
!= BAD_FILE
) {
4487 sources
[length
] = fs_reg(VGRF
, bld
.shader
->alloc
.allocate(1),
4488 BRW_REGISTER_TYPE_UD
);
4490 /* Hand over gl_SampleMask. Only the lower 16 bits of each channel are
4491 * relevant. Since it's unsigned single words one vgrf is always
4492 * 16-wide, but only the lower or higher 8 channels will be used by the
4493 * hardware when doing a SIMD8 write depending on whether we have
4494 * selected the subspans for the first or second half respectively.
4496 assert(sample_mask
.file
!= BAD_FILE
&& type_sz(sample_mask
.type
) == 4);
4497 sample_mask
.type
= BRW_REGISTER_TYPE_UW
;
4498 sample_mask
.stride
*= 2;
4500 bld
.exec_all().annotate("FB write oMask")
4501 .MOV(horiz_offset(retype(sources
[length
], BRW_REGISTER_TYPE_UW
),
4507 payload_header_size
= length
;
4509 setup_color_payload(bld
, key
, &sources
[length
], color0
, components
);
4512 if (color1
.file
!= BAD_FILE
) {
4513 setup_color_payload(bld
, key
, &sources
[length
], color1
, components
);
4517 if (src_depth
.file
!= BAD_FILE
) {
4518 sources
[length
] = src_depth
;
4522 if (dst_depth
.file
!= BAD_FILE
) {
4523 sources
[length
] = dst_depth
;
4527 if (src_stencil
.file
!= BAD_FILE
) {
4528 assert(devinfo
->gen
>= 9);
4529 assert(bld
.dispatch_width() == 8);
4531 /* XXX: src_stencil is only available on gen9+. dst_depth is never
4532 * available on gen9+. As such it's impossible to have both enabled at the
4533 * same time and therefore length cannot overrun the array.
4535 assert(length
< 15);
4537 sources
[length
] = bld
.vgrf(BRW_REGISTER_TYPE_UD
);
4538 bld
.exec_all().annotate("FB write OS")
4539 .MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UB
),
4540 subscript(src_stencil
, BRW_REGISTER_TYPE_UB
, 0));
4545 if (devinfo
->gen
>= 7) {
4546 /* Send from the GRF */
4547 fs_reg payload
= fs_reg(VGRF
, -1, BRW_REGISTER_TYPE_F
);
4548 load
= bld
.LOAD_PAYLOAD(payload
, sources
, length
, payload_header_size
);
4549 payload
.nr
= bld
.shader
->alloc
.allocate(regs_written(load
));
4550 load
->dst
= payload
;
4552 uint32_t msg_ctl
= brw_fb_write_msg_control(inst
, prog_data
);
4553 uint32_t ex_desc
= 0;
4556 (inst
->group
/ 16) << 11 | /* rt slot group */
4557 brw_dp_write_desc(devinfo
, inst
->target
, msg_ctl
,
4558 GEN6_DATAPORT_WRITE_MESSAGE_RENDER_TARGET_WRITE
,
4559 inst
->last_rt
, false);
4561 if (devinfo
->gen
>= 11) {
4562 /* Set the "Render Target Index" and "Src0 Alpha Present" fields
4563 * in the extended message descriptor, in lieu of using a header.
4565 ex_desc
= inst
->target
<< 12 | src0_alpha_present
<< 15;
4567 if (key
->nr_color_regions
== 0)
4568 ex_desc
|= 1 << 20; /* Null Render Target */
4571 inst
->opcode
= SHADER_OPCODE_SEND
;
4572 inst
->resize_sources(3);
4573 inst
->sfid
= GEN6_SFID_DATAPORT_RENDER_CACHE
;
4574 inst
->src
[0] = brw_imm_ud(inst
->desc
);
4575 inst
->src
[1] = brw_imm_ud(ex_desc
);
4576 inst
->src
[2] = payload
;
4577 inst
->mlen
= regs_written(load
);
4579 inst
->header_size
= header_size
;
4580 inst
->check_tdr
= true;
4581 inst
->send_has_side_effects
= true;
4583 /* Send from the MRF */
4584 load
= bld
.LOAD_PAYLOAD(fs_reg(MRF
, 1, BRW_REGISTER_TYPE_F
),
4585 sources
, length
, payload_header_size
);
4587 /* On pre-SNB, we have to interlace the color values. LOAD_PAYLOAD
4588 * will do this for us if we just give it a COMPR4 destination.
4590 if (devinfo
->gen
< 6 && bld
.dispatch_width() == 16)
4591 load
->dst
.nr
|= BRW_MRF_COMPR4
;
4593 if (devinfo
->gen
< 6) {
4594 /* Set up src[0] for the implied MOV from grf0-1 */
4595 inst
->resize_sources(1);
4596 inst
->src
[0] = brw_vec8_grf(0, 0);
4598 inst
->resize_sources(0);
4601 inst
->opcode
= FS_OPCODE_FB_WRITE
;
4602 inst
->mlen
= regs_written(load
);
4603 inst
->header_size
= header_size
;
4608 lower_fb_read_logical_send(const fs_builder
&bld
, fs_inst
*inst
)
4610 const fs_builder
&ubld
= bld
.exec_all().group(8, 0);
4611 const unsigned length
= 2;
4612 const fs_reg header
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
, length
);
4614 if (bld
.group() < 16) {
4615 ubld
.group(16, 0).MOV(header
, retype(brw_vec8_grf(0, 0),
4616 BRW_REGISTER_TYPE_UD
));
4618 assert(bld
.group() < 32);
4619 const fs_reg header_sources
[] = {
4620 retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
),
4621 retype(brw_vec8_grf(2, 0), BRW_REGISTER_TYPE_UD
)
4623 ubld
.LOAD_PAYLOAD(header
, header_sources
, ARRAY_SIZE(header_sources
), 0);
4626 inst
->resize_sources(1);
4627 inst
->src
[0] = header
;
4628 inst
->opcode
= FS_OPCODE_FB_READ
;
4629 inst
->mlen
= length
;
4630 inst
->header_size
= length
;
4634 lower_sampler_logical_send_gen4(const fs_builder
&bld
, fs_inst
*inst
, opcode op
,
4635 const fs_reg
&coordinate
,
4636 const fs_reg
&shadow_c
,
4637 const fs_reg
&lod
, const fs_reg
&lod2
,
4638 const fs_reg
&surface
,
4639 const fs_reg
&sampler
,
4640 unsigned coord_components
,
4641 unsigned grad_components
)
4643 const bool has_lod
= (op
== SHADER_OPCODE_TXL
|| op
== FS_OPCODE_TXB
||
4644 op
== SHADER_OPCODE_TXF
|| op
== SHADER_OPCODE_TXS
);
4645 fs_reg
msg_begin(MRF
, 1, BRW_REGISTER_TYPE_F
);
4646 fs_reg msg_end
= msg_begin
;
4649 msg_end
= offset(msg_end
, bld
.group(8, 0), 1);
4651 for (unsigned i
= 0; i
< coord_components
; i
++)
4652 bld
.MOV(retype(offset(msg_end
, bld
, i
), coordinate
.type
),
4653 offset(coordinate
, bld
, i
));
4655 msg_end
= offset(msg_end
, bld
, coord_components
);
4657 /* Messages other than SAMPLE and RESINFO in SIMD16 and TXD in SIMD8
4658 * require all three components to be present and zero if they are unused.
4660 if (coord_components
> 0 &&
4661 (has_lod
|| shadow_c
.file
!= BAD_FILE
||
4662 (op
== SHADER_OPCODE_TEX
&& bld
.dispatch_width() == 8))) {
4663 for (unsigned i
= coord_components
; i
< 3; i
++)
4664 bld
.MOV(offset(msg_end
, bld
, i
), brw_imm_f(0.0f
));
4666 msg_end
= offset(msg_end
, bld
, 3 - coord_components
);
4669 if (op
== SHADER_OPCODE_TXD
) {
4670 /* TXD unsupported in SIMD16 mode. */
4671 assert(bld
.dispatch_width() == 8);
4673 /* the slots for u and v are always present, but r is optional */
4674 if (coord_components
< 2)
4675 msg_end
= offset(msg_end
, bld
, 2 - coord_components
);
4678 * dPdx = dudx, dvdx, drdx
4679 * dPdy = dudy, dvdy, drdy
4681 * 1-arg: Does not exist.
4683 * 2-arg: dudx dvdx dudy dvdy
4684 * dPdx.x dPdx.y dPdy.x dPdy.y
4687 * 3-arg: dudx dvdx drdx dudy dvdy drdy
4688 * dPdx.x dPdx.y dPdx.z dPdy.x dPdy.y dPdy.z
4689 * m5 m6 m7 m8 m9 m10
4691 for (unsigned i
= 0; i
< grad_components
; i
++)
4692 bld
.MOV(offset(msg_end
, bld
, i
), offset(lod
, bld
, i
));
4694 msg_end
= offset(msg_end
, bld
, MAX2(grad_components
, 2));
4696 for (unsigned i
= 0; i
< grad_components
; i
++)
4697 bld
.MOV(offset(msg_end
, bld
, i
), offset(lod2
, bld
, i
));
4699 msg_end
= offset(msg_end
, bld
, MAX2(grad_components
, 2));
4703 /* Bias/LOD with shadow comparator is unsupported in SIMD16 -- *Without*
4704 * shadow comparator (including RESINFO) it's unsupported in SIMD8 mode.
4706 assert(shadow_c
.file
!= BAD_FILE
? bld
.dispatch_width() == 8 :
4707 bld
.dispatch_width() == 16);
4709 const brw_reg_type type
=
4710 (op
== SHADER_OPCODE_TXF
|| op
== SHADER_OPCODE_TXS
?
4711 BRW_REGISTER_TYPE_UD
: BRW_REGISTER_TYPE_F
);
4712 bld
.MOV(retype(msg_end
, type
), lod
);
4713 msg_end
= offset(msg_end
, bld
, 1);
4716 if (shadow_c
.file
!= BAD_FILE
) {
4717 if (op
== SHADER_OPCODE_TEX
&& bld
.dispatch_width() == 8) {
4718 /* There's no plain shadow compare message, so we use shadow
4719 * compare with a bias of 0.0.
4721 bld
.MOV(msg_end
, brw_imm_f(0.0f
));
4722 msg_end
= offset(msg_end
, bld
, 1);
4725 bld
.MOV(msg_end
, shadow_c
);
4726 msg_end
= offset(msg_end
, bld
, 1);
4730 inst
->src
[0] = reg_undef
;
4731 inst
->src
[1] = surface
;
4732 inst
->src
[2] = sampler
;
4733 inst
->resize_sources(3);
4734 inst
->base_mrf
= msg_begin
.nr
;
4735 inst
->mlen
= msg_end
.nr
- msg_begin
.nr
;
4736 inst
->header_size
= 1;
4740 lower_sampler_logical_send_gen5(const fs_builder
&bld
, fs_inst
*inst
, opcode op
,
4741 const fs_reg
&coordinate
,
4742 const fs_reg
&shadow_c
,
4743 const fs_reg
&lod
, const fs_reg
&lod2
,
4744 const fs_reg
&sample_index
,
4745 const fs_reg
&surface
,
4746 const fs_reg
&sampler
,
4747 unsigned coord_components
,
4748 unsigned grad_components
)
4750 fs_reg
message(MRF
, 2, BRW_REGISTER_TYPE_F
);
4751 fs_reg msg_coords
= message
;
4752 unsigned header_size
= 0;
4754 if (inst
->offset
!= 0) {
4755 /* The offsets set up by the visitor are in the m1 header, so we can't
4762 for (unsigned i
= 0; i
< coord_components
; i
++)
4763 bld
.MOV(retype(offset(msg_coords
, bld
, i
), coordinate
.type
),
4764 offset(coordinate
, bld
, i
));
4766 fs_reg msg_end
= offset(msg_coords
, bld
, coord_components
);
4767 fs_reg msg_lod
= offset(msg_coords
, bld
, 4);
4769 if (shadow_c
.file
!= BAD_FILE
) {
4770 fs_reg msg_shadow
= msg_lod
;
4771 bld
.MOV(msg_shadow
, shadow_c
);
4772 msg_lod
= offset(msg_shadow
, bld
, 1);
4777 case SHADER_OPCODE_TXL
:
4779 bld
.MOV(msg_lod
, lod
);
4780 msg_end
= offset(msg_lod
, bld
, 1);
4782 case SHADER_OPCODE_TXD
:
4785 * dPdx = dudx, dvdx, drdx
4786 * dPdy = dudy, dvdy, drdy
4788 * Load up these values:
4789 * - dudx dudy dvdx dvdy drdx drdy
4790 * - dPdx.x dPdy.x dPdx.y dPdy.y dPdx.z dPdy.z
4793 for (unsigned i
= 0; i
< grad_components
; i
++) {
4794 bld
.MOV(msg_end
, offset(lod
, bld
, i
));
4795 msg_end
= offset(msg_end
, bld
, 1);
4797 bld
.MOV(msg_end
, offset(lod2
, bld
, i
));
4798 msg_end
= offset(msg_end
, bld
, 1);
4801 case SHADER_OPCODE_TXS
:
4802 msg_lod
= retype(msg_end
, BRW_REGISTER_TYPE_UD
);
4803 bld
.MOV(msg_lod
, lod
);
4804 msg_end
= offset(msg_lod
, bld
, 1);
4806 case SHADER_OPCODE_TXF
:
4807 msg_lod
= offset(msg_coords
, bld
, 3);
4808 bld
.MOV(retype(msg_lod
, BRW_REGISTER_TYPE_UD
), lod
);
4809 msg_end
= offset(msg_lod
, bld
, 1);
4811 case SHADER_OPCODE_TXF_CMS
:
4812 msg_lod
= offset(msg_coords
, bld
, 3);
4814 bld
.MOV(retype(msg_lod
, BRW_REGISTER_TYPE_UD
), brw_imm_ud(0u));
4816 bld
.MOV(retype(offset(msg_lod
, bld
, 1), BRW_REGISTER_TYPE_UD
), sample_index
);
4817 msg_end
= offset(msg_lod
, bld
, 2);
4824 inst
->src
[0] = reg_undef
;
4825 inst
->src
[1] = surface
;
4826 inst
->src
[2] = sampler
;
4827 inst
->resize_sources(3);
4828 inst
->base_mrf
= message
.nr
;
4829 inst
->mlen
= msg_end
.nr
- message
.nr
;
4830 inst
->header_size
= header_size
;
4832 /* Message length > MAX_SAMPLER_MESSAGE_SIZE disallowed by hardware. */
4833 assert(inst
->mlen
<= MAX_SAMPLER_MESSAGE_SIZE
);
4837 is_high_sampler(const struct gen_device_info
*devinfo
, const fs_reg
&sampler
)
4839 if (devinfo
->gen
< 8 && !devinfo
->is_haswell
)
4842 return sampler
.file
!= IMM
|| sampler
.ud
>= 16;
4846 sampler_msg_type(const gen_device_info
*devinfo
,
4847 opcode opcode
, bool shadow_compare
)
4849 assert(devinfo
->gen
>= 5);
4851 case SHADER_OPCODE_TEX
:
4852 return shadow_compare
? GEN5_SAMPLER_MESSAGE_SAMPLE_COMPARE
:
4853 GEN5_SAMPLER_MESSAGE_SAMPLE
;
4855 return shadow_compare
? GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS_COMPARE
:
4856 GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS
;
4857 case SHADER_OPCODE_TXL
:
4858 return shadow_compare
? GEN5_SAMPLER_MESSAGE_SAMPLE_LOD_COMPARE
:
4859 GEN5_SAMPLER_MESSAGE_SAMPLE_LOD
;
4860 case SHADER_OPCODE_TXL_LZ
:
4861 return shadow_compare
? GEN9_SAMPLER_MESSAGE_SAMPLE_C_LZ
:
4862 GEN9_SAMPLER_MESSAGE_SAMPLE_LZ
;
4863 case SHADER_OPCODE_TXS
:
4864 case SHADER_OPCODE_IMAGE_SIZE_LOGICAL
:
4865 return GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO
;
4866 case SHADER_OPCODE_TXD
:
4867 assert(!shadow_compare
|| devinfo
->gen
>= 8 || devinfo
->is_haswell
);
4868 return shadow_compare
? HSW_SAMPLER_MESSAGE_SAMPLE_DERIV_COMPARE
:
4869 GEN5_SAMPLER_MESSAGE_SAMPLE_DERIVS
;
4870 case SHADER_OPCODE_TXF
:
4871 return GEN5_SAMPLER_MESSAGE_SAMPLE_LD
;
4872 case SHADER_OPCODE_TXF_LZ
:
4873 assert(devinfo
->gen
>= 9);
4874 return GEN9_SAMPLER_MESSAGE_SAMPLE_LD_LZ
;
4875 case SHADER_OPCODE_TXF_CMS_W
:
4876 assert(devinfo
->gen
>= 9);
4877 return GEN9_SAMPLER_MESSAGE_SAMPLE_LD2DMS_W
;
4878 case SHADER_OPCODE_TXF_CMS
:
4879 return devinfo
->gen
>= 7 ? GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DMS
:
4880 GEN5_SAMPLER_MESSAGE_SAMPLE_LD
;
4881 case SHADER_OPCODE_TXF_UMS
:
4882 assert(devinfo
->gen
>= 7);
4883 return GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DSS
;
4884 case SHADER_OPCODE_TXF_MCS
:
4885 assert(devinfo
->gen
>= 7);
4886 return GEN7_SAMPLER_MESSAGE_SAMPLE_LD_MCS
;
4887 case SHADER_OPCODE_LOD
:
4888 return GEN5_SAMPLER_MESSAGE_LOD
;
4889 case SHADER_OPCODE_TG4
:
4890 assert(devinfo
->gen
>= 7);
4891 return shadow_compare
? GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_C
:
4892 GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4
;
4894 case SHADER_OPCODE_TG4_OFFSET
:
4895 assert(devinfo
->gen
>= 7);
4896 return shadow_compare
? GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO_C
:
4897 GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO
;
4898 case SHADER_OPCODE_SAMPLEINFO
:
4899 return GEN6_SAMPLER_MESSAGE_SAMPLE_SAMPLEINFO
;
4901 unreachable("not reached");
4906 lower_sampler_logical_send_gen7(const fs_builder
&bld
, fs_inst
*inst
, opcode op
,
4907 const fs_reg
&coordinate
,
4908 const fs_reg
&shadow_c
,
4909 fs_reg lod
, const fs_reg
&lod2
,
4910 const fs_reg
&min_lod
,
4911 const fs_reg
&sample_index
,
4913 const fs_reg
&surface
,
4914 const fs_reg
&sampler
,
4915 const fs_reg
&surface_handle
,
4916 const fs_reg
&sampler_handle
,
4917 const fs_reg
&tg4_offset
,
4918 unsigned coord_components
,
4919 unsigned grad_components
)
4921 const gen_device_info
*devinfo
= bld
.shader
->devinfo
;
4922 const brw_stage_prog_data
*prog_data
= bld
.shader
->stage_prog_data
;
4923 unsigned reg_width
= bld
.dispatch_width() / 8;
4924 unsigned header_size
= 0, length
= 0;
4925 fs_reg sources
[MAX_SAMPLER_MESSAGE_SIZE
];
4926 for (unsigned i
= 0; i
< ARRAY_SIZE(sources
); i
++)
4927 sources
[i
] = bld
.vgrf(BRW_REGISTER_TYPE_F
);
4929 /* We must have exactly one of surface/sampler and surface/sampler_handle */
4930 assert((surface
.file
== BAD_FILE
) != (surface_handle
.file
== BAD_FILE
));
4931 assert((sampler
.file
== BAD_FILE
) != (sampler_handle
.file
== BAD_FILE
));
4933 if (op
== SHADER_OPCODE_TG4
|| op
== SHADER_OPCODE_TG4_OFFSET
||
4934 inst
->offset
!= 0 || inst
->eot
||
4935 op
== SHADER_OPCODE_SAMPLEINFO
||
4936 sampler_handle
.file
!= BAD_FILE
||
4937 is_high_sampler(devinfo
, sampler
)) {
4938 /* For general texture offsets (no txf workaround), we need a header to
4941 * TG4 needs to place its channel select in the header, for interaction
4942 * with ARB_texture_swizzle. The sampler index is only 4-bits, so for
4943 * larger sampler numbers we need to offset the Sampler State Pointer in
4946 fs_reg header
= retype(sources
[0], BRW_REGISTER_TYPE_UD
);
4950 /* If we're requesting fewer than four channels worth of response,
4951 * and we have an explicit header, we need to set up the sampler
4952 * writemask. It's reversed from normal: 1 means "don't write".
4954 if (!inst
->eot
&& regs_written(inst
) != 4 * reg_width
) {
4955 assert(regs_written(inst
) % reg_width
== 0);
4956 unsigned mask
= ~((1 << (regs_written(inst
) / reg_width
)) - 1) & 0xf;
4957 inst
->offset
|= mask
<< 12;
4960 /* Build the actual header */
4961 const fs_builder ubld
= bld
.exec_all().group(8, 0);
4962 const fs_builder ubld1
= ubld
.group(1, 0);
4963 ubld
.MOV(header
, retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
));
4965 ubld1
.MOV(component(header
, 2), brw_imm_ud(inst
->offset
));
4966 } else if (bld
.shader
->stage
!= MESA_SHADER_VERTEX
&&
4967 bld
.shader
->stage
!= MESA_SHADER_FRAGMENT
) {
4968 /* The vertex and fragment stages have g0.2 set to 0, so
4969 * header0.2 is 0 when g0 is copied. Other stages may not, so we
4970 * must set it to 0 to avoid setting undesirable bits in the
4973 ubld1
.MOV(component(header
, 2), brw_imm_ud(0));
4976 if (sampler_handle
.file
!= BAD_FILE
) {
4977 /* Bindless sampler handles aren't relative to the sampler state
4978 * pointer passed into the shader through SAMPLER_STATE_POINTERS_*.
4979 * Instead, it's an absolute pointer relative to dynamic state base
4982 * Sampler states are 16 bytes each and the pointer we give here has
4983 * to be 32-byte aligned. In order to avoid more indirect messages
4984 * than required, we assume that all bindless sampler states are
4985 * 32-byte aligned. This sacrifices a bit of general state base
4986 * address space but means we can do something more efficient in the
4989 ubld1
.MOV(component(header
, 3), sampler_handle
);
4990 } else if (is_high_sampler(devinfo
, sampler
)) {
4991 if (sampler
.file
== BRW_IMMEDIATE_VALUE
) {
4992 assert(sampler
.ud
>= 16);
4993 const int sampler_state_size
= 16; /* 16 bytes */
4995 ubld1
.ADD(component(header
, 3),
4996 retype(brw_vec1_grf(0, 3), BRW_REGISTER_TYPE_UD
),
4997 brw_imm_ud(16 * (sampler
.ud
/ 16) * sampler_state_size
));
4999 fs_reg tmp
= ubld1
.vgrf(BRW_REGISTER_TYPE_UD
);
5000 ubld1
.AND(tmp
, sampler
, brw_imm_ud(0x0f0));
5001 ubld1
.SHL(tmp
, tmp
, brw_imm_ud(4));
5002 ubld1
.ADD(component(header
, 3),
5003 retype(brw_vec1_grf(0, 3), BRW_REGISTER_TYPE_UD
),
5009 if (shadow_c
.file
!= BAD_FILE
) {
5010 bld
.MOV(sources
[length
], shadow_c
);
5014 bool coordinate_done
= false;
5016 /* Set up the LOD info */
5019 case SHADER_OPCODE_TXL
:
5020 if (devinfo
->gen
>= 9 && op
== SHADER_OPCODE_TXL
&& lod
.is_zero()) {
5021 op
= SHADER_OPCODE_TXL_LZ
;
5024 bld
.MOV(sources
[length
], lod
);
5027 case SHADER_OPCODE_TXD
:
5028 /* TXD should have been lowered in SIMD16 mode. */
5029 assert(bld
.dispatch_width() == 8);
5031 /* Load dPdx and the coordinate together:
5032 * [hdr], [ref], x, dPdx.x, dPdy.x, y, dPdx.y, dPdy.y, z, dPdx.z, dPdy.z
5034 for (unsigned i
= 0; i
< coord_components
; i
++) {
5035 bld
.MOV(sources
[length
++], offset(coordinate
, bld
, i
));
5037 /* For cube map array, the coordinate is (u,v,r,ai) but there are
5038 * only derivatives for (u, v, r).
5040 if (i
< grad_components
) {
5041 bld
.MOV(sources
[length
++], offset(lod
, bld
, i
));
5042 bld
.MOV(sources
[length
++], offset(lod2
, bld
, i
));
5046 coordinate_done
= true;
5048 case SHADER_OPCODE_TXS
:
5049 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
), lod
);
5052 case SHADER_OPCODE_IMAGE_SIZE_LOGICAL
:
5053 /* We need an LOD; just use 0 */
5054 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
), brw_imm_ud(0));
5057 case SHADER_OPCODE_TXF
:
5058 /* Unfortunately, the parameters for LD are intermixed: u, lod, v, r.
5059 * On Gen9 they are u, v, lod, r
5061 bld
.MOV(retype(sources
[length
++], BRW_REGISTER_TYPE_D
), coordinate
);
5063 if (devinfo
->gen
>= 9) {
5064 if (coord_components
>= 2) {
5065 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
),
5066 offset(coordinate
, bld
, 1));
5068 sources
[length
] = brw_imm_d(0);
5073 if (devinfo
->gen
>= 9 && lod
.is_zero()) {
5074 op
= SHADER_OPCODE_TXF_LZ
;
5076 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
), lod
);
5080 for (unsigned i
= devinfo
->gen
>= 9 ? 2 : 1; i
< coord_components
; i
++)
5081 bld
.MOV(retype(sources
[length
++], BRW_REGISTER_TYPE_D
),
5082 offset(coordinate
, bld
, i
));
5084 coordinate_done
= true;
5087 case SHADER_OPCODE_TXF_CMS
:
5088 case SHADER_OPCODE_TXF_CMS_W
:
5089 case SHADER_OPCODE_TXF_UMS
:
5090 case SHADER_OPCODE_TXF_MCS
:
5091 if (op
== SHADER_OPCODE_TXF_UMS
||
5092 op
== SHADER_OPCODE_TXF_CMS
||
5093 op
== SHADER_OPCODE_TXF_CMS_W
) {
5094 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
), sample_index
);
5098 if (op
== SHADER_OPCODE_TXF_CMS
|| op
== SHADER_OPCODE_TXF_CMS_W
) {
5099 /* Data from the multisample control surface. */
5100 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
), mcs
);
5103 /* On Gen9+ we'll use ld2dms_w instead which has two registers for
5106 if (op
== SHADER_OPCODE_TXF_CMS_W
) {
5107 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
),
5110 offset(mcs
, bld
, 1));
5115 /* There is no offsetting for this message; just copy in the integer
5116 * texture coordinates.
5118 for (unsigned i
= 0; i
< coord_components
; i
++)
5119 bld
.MOV(retype(sources
[length
++], BRW_REGISTER_TYPE_D
),
5120 offset(coordinate
, bld
, i
));
5122 coordinate_done
= true;
5124 case SHADER_OPCODE_TG4_OFFSET
:
5125 /* More crazy intermixing */
5126 for (unsigned i
= 0; i
< 2; i
++) /* u, v */
5127 bld
.MOV(sources
[length
++], offset(coordinate
, bld
, i
));
5129 for (unsigned i
= 0; i
< 2; i
++) /* offu, offv */
5130 bld
.MOV(retype(sources
[length
++], BRW_REGISTER_TYPE_D
),
5131 offset(tg4_offset
, bld
, i
));
5133 if (coord_components
== 3) /* r if present */
5134 bld
.MOV(sources
[length
++], offset(coordinate
, bld
, 2));
5136 coordinate_done
= true;
5142 /* Set up the coordinate (except for cases where it was done above) */
5143 if (!coordinate_done
) {
5144 for (unsigned i
= 0; i
< coord_components
; i
++)
5145 bld
.MOV(sources
[length
++], offset(coordinate
, bld
, i
));
5148 if (min_lod
.file
!= BAD_FILE
) {
5149 /* Account for all of the missing coordinate sources */
5150 length
+= 4 - coord_components
;
5151 if (op
== SHADER_OPCODE_TXD
)
5152 length
+= (3 - grad_components
) * 2;
5154 bld
.MOV(sources
[length
++], min_lod
);
5159 mlen
= length
* reg_width
- header_size
;
5161 mlen
= length
* reg_width
;
5163 const fs_reg src_payload
= fs_reg(VGRF
, bld
.shader
->alloc
.allocate(mlen
),
5164 BRW_REGISTER_TYPE_F
);
5165 bld
.LOAD_PAYLOAD(src_payload
, sources
, length
, header_size
);
5167 /* Generate the SEND. */
5168 inst
->opcode
= SHADER_OPCODE_SEND
;
5170 inst
->header_size
= header_size
;
5172 const unsigned msg_type
=
5173 sampler_msg_type(devinfo
, op
, inst
->shadow_compare
);
5174 const unsigned simd_mode
=
5175 inst
->exec_size
<= 8 ? BRW_SAMPLER_SIMD_MODE_SIMD8
:
5176 BRW_SAMPLER_SIMD_MODE_SIMD16
;
5178 uint32_t base_binding_table_index
;
5180 case SHADER_OPCODE_TG4
:
5181 case SHADER_OPCODE_TG4_OFFSET
:
5182 base_binding_table_index
= prog_data
->binding_table
.gather_texture_start
;
5184 case SHADER_OPCODE_IMAGE_SIZE_LOGICAL
:
5185 base_binding_table_index
= prog_data
->binding_table
.image_start
;
5188 base_binding_table_index
= prog_data
->binding_table
.texture_start
;
5192 inst
->sfid
= BRW_SFID_SAMPLER
;
5193 if (surface
.file
== IMM
&&
5194 (sampler
.file
== IMM
|| sampler_handle
.file
!= BAD_FILE
)) {
5195 inst
->desc
= brw_sampler_desc(devinfo
,
5196 surface
.ud
+ base_binding_table_index
,
5197 sampler
.file
== IMM
? sampler
.ud
% 16 : 0,
5200 0 /* return_format unused on gen7+ */);
5201 inst
->src
[0] = brw_imm_ud(0);
5202 inst
->src
[1] = brw_imm_ud(0); /* ex_desc */
5203 } else if (surface_handle
.file
!= BAD_FILE
) {
5204 /* Bindless surface */
5205 assert(devinfo
->gen
>= 9);
5206 inst
->desc
= brw_sampler_desc(devinfo
,
5208 sampler
.file
== IMM
? sampler
.ud
% 16 : 0,
5211 0 /* return_format unused on gen7+ */);
5213 /* For bindless samplers, the entire address is included in the message
5214 * header so we can leave the portion in the message descriptor 0.
5216 if (sampler_handle
.file
!= BAD_FILE
|| sampler
.file
== IMM
) {
5217 inst
->src
[0] = brw_imm_ud(0);
5219 const fs_builder ubld
= bld
.group(1, 0).exec_all();
5220 fs_reg desc
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
5221 ubld
.SHL(desc
, sampler
, brw_imm_ud(8));
5222 inst
->src
[0] = desc
;
5225 /* We assume that the driver provided the handle in the top 20 bits so
5226 * we can use the surface handle directly as the extended descriptor.
5228 inst
->src
[1] = retype(surface_handle
, BRW_REGISTER_TYPE_UD
);
5230 /* Immediate portion of the descriptor */
5231 inst
->desc
= brw_sampler_desc(devinfo
,
5236 0 /* return_format unused on gen7+ */);
5237 const fs_builder ubld
= bld
.group(1, 0).exec_all();
5238 fs_reg desc
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
5239 if (surface
.equals(sampler
)) {
5240 /* This case is common in GL */
5241 ubld
.MUL(desc
, surface
, brw_imm_ud(0x101));
5243 if (sampler_handle
.file
!= BAD_FILE
) {
5244 ubld
.MOV(desc
, surface
);
5245 } else if (sampler
.file
== IMM
) {
5246 ubld
.OR(desc
, surface
, brw_imm_ud(sampler
.ud
<< 8));
5248 ubld
.SHL(desc
, sampler
, brw_imm_ud(8));
5249 ubld
.OR(desc
, desc
, surface
);
5252 if (base_binding_table_index
)
5253 ubld
.ADD(desc
, desc
, brw_imm_ud(base_binding_table_index
));
5254 ubld
.AND(desc
, desc
, brw_imm_ud(0xfff));
5256 inst
->src
[0] = component(desc
, 0);
5257 inst
->src
[1] = brw_imm_ud(0); /* ex_desc */
5260 inst
->src
[2] = src_payload
;
5261 inst
->resize_sources(3);
5264 /* EOT sampler messages don't make sense to split because it would
5265 * involve ending half of the thread early.
5267 assert(inst
->group
== 0);
5268 /* We need to use SENDC for EOT sampler messages */
5269 inst
->check_tdr
= true;
5270 inst
->send_has_side_effects
= true;
5273 /* Message length > MAX_SAMPLER_MESSAGE_SIZE disallowed by hardware. */
5274 assert(inst
->mlen
<= MAX_SAMPLER_MESSAGE_SIZE
);
5278 lower_sampler_logical_send(const fs_builder
&bld
, fs_inst
*inst
, opcode op
)
5280 const gen_device_info
*devinfo
= bld
.shader
->devinfo
;
5281 const fs_reg
&coordinate
= inst
->src
[TEX_LOGICAL_SRC_COORDINATE
];
5282 const fs_reg
&shadow_c
= inst
->src
[TEX_LOGICAL_SRC_SHADOW_C
];
5283 const fs_reg
&lod
= inst
->src
[TEX_LOGICAL_SRC_LOD
];
5284 const fs_reg
&lod2
= inst
->src
[TEX_LOGICAL_SRC_LOD2
];
5285 const fs_reg
&min_lod
= inst
->src
[TEX_LOGICAL_SRC_MIN_LOD
];
5286 const fs_reg
&sample_index
= inst
->src
[TEX_LOGICAL_SRC_SAMPLE_INDEX
];
5287 const fs_reg
&mcs
= inst
->src
[TEX_LOGICAL_SRC_MCS
];
5288 const fs_reg
&surface
= inst
->src
[TEX_LOGICAL_SRC_SURFACE
];
5289 const fs_reg
&sampler
= inst
->src
[TEX_LOGICAL_SRC_SAMPLER
];
5290 const fs_reg
&surface_handle
= inst
->src
[TEX_LOGICAL_SRC_SURFACE_HANDLE
];
5291 const fs_reg
&sampler_handle
= inst
->src
[TEX_LOGICAL_SRC_SAMPLER_HANDLE
];
5292 const fs_reg
&tg4_offset
= inst
->src
[TEX_LOGICAL_SRC_TG4_OFFSET
];
5293 assert(inst
->src
[TEX_LOGICAL_SRC_COORD_COMPONENTS
].file
== IMM
);
5294 const unsigned coord_components
= inst
->src
[TEX_LOGICAL_SRC_COORD_COMPONENTS
].ud
;
5295 assert(inst
->src
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
].file
== IMM
);
5296 const unsigned grad_components
= inst
->src
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
].ud
;
5298 if (devinfo
->gen
>= 7) {
5299 lower_sampler_logical_send_gen7(bld
, inst
, op
, coordinate
,
5300 shadow_c
, lod
, lod2
, min_lod
,
5302 mcs
, surface
, sampler
,
5303 surface_handle
, sampler_handle
,
5305 coord_components
, grad_components
);
5306 } else if (devinfo
->gen
>= 5) {
5307 lower_sampler_logical_send_gen5(bld
, inst
, op
, coordinate
,
5308 shadow_c
, lod
, lod2
, sample_index
,
5310 coord_components
, grad_components
);
5312 lower_sampler_logical_send_gen4(bld
, inst
, op
, coordinate
,
5313 shadow_c
, lod
, lod2
,
5315 coord_components
, grad_components
);
5320 * Initialize the header present in some typed and untyped surface
5324 emit_surface_header(const fs_builder
&bld
, const fs_reg
&sample_mask
)
5326 fs_builder ubld
= bld
.exec_all().group(8, 0);
5327 const fs_reg dst
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
5328 ubld
.MOV(dst
, brw_imm_d(0));
5329 ubld
.group(1, 0).MOV(component(dst
, 7), sample_mask
);
5334 lower_surface_logical_send(const fs_builder
&bld
, fs_inst
*inst
)
5336 const gen_device_info
*devinfo
= bld
.shader
->devinfo
;
5338 /* Get the logical send arguments. */
5339 const fs_reg
&addr
= inst
->src
[SURFACE_LOGICAL_SRC_ADDRESS
];
5340 const fs_reg
&src
= inst
->src
[SURFACE_LOGICAL_SRC_DATA
];
5341 const fs_reg
&surface
= inst
->src
[SURFACE_LOGICAL_SRC_SURFACE
];
5342 const fs_reg
&surface_handle
= inst
->src
[SURFACE_LOGICAL_SRC_SURFACE_HANDLE
];
5343 const UNUSED fs_reg
&dims
= inst
->src
[SURFACE_LOGICAL_SRC_IMM_DIMS
];
5344 const fs_reg
&arg
= inst
->src
[SURFACE_LOGICAL_SRC_IMM_ARG
];
5345 assert(arg
.file
== IMM
);
5347 /* We must have exactly one of surface and surface_handle */
5348 assert((surface
.file
== BAD_FILE
) != (surface_handle
.file
== BAD_FILE
));
5350 /* Calculate the total number of components of the payload. */
5351 const unsigned addr_sz
= inst
->components_read(SURFACE_LOGICAL_SRC_ADDRESS
);
5352 const unsigned src_sz
= inst
->components_read(SURFACE_LOGICAL_SRC_DATA
);
5354 const bool is_typed_access
=
5355 inst
->opcode
== SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
||
5356 inst
->opcode
== SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
||
5357 inst
->opcode
== SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
;
5359 /* From the BDW PRM Volume 7, page 147:
5361 * "For the Data Cache Data Port*, the header must be present for the
5362 * following message types: [...] Typed read/write/atomics"
5364 * Earlier generations have a similar wording. Because of this restriction
5365 * we don't attempt to implement sample masks via predication for such
5366 * messages prior to Gen9, since we have to provide a header anyway. On
5367 * Gen11+ the header has been removed so we can only use predication.
5369 const unsigned header_sz
= devinfo
->gen
< 9 && is_typed_access
? 1 : 0;
5371 const bool has_side_effects
= inst
->has_side_effects();
5372 fs_reg sample_mask
= has_side_effects
? bld
.sample_mask_reg() :
5373 fs_reg(brw_imm_d(0xffff));
5375 fs_reg payload
, payload2
;
5376 unsigned mlen
, ex_mlen
= 0;
5377 if (devinfo
->gen
>= 9) {
5378 /* We have split sends on gen9 and above */
5379 assert(header_sz
== 0);
5380 payload
= bld
.move_to_vgrf(addr
, addr_sz
);
5381 payload2
= bld
.move_to_vgrf(src
, src_sz
);
5382 mlen
= addr_sz
* (inst
->exec_size
/ 8);
5383 ex_mlen
= src_sz
* (inst
->exec_size
/ 8);
5385 /* Allocate space for the payload. */
5386 const unsigned sz
= header_sz
+ addr_sz
+ src_sz
;
5387 payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, sz
);
5388 fs_reg
*const components
= new fs_reg
[sz
];
5391 /* Construct the payload. */
5393 components
[n
++] = emit_surface_header(bld
, sample_mask
);
5395 for (unsigned i
= 0; i
< addr_sz
; i
++)
5396 components
[n
++] = offset(addr
, bld
, i
);
5398 for (unsigned i
= 0; i
< src_sz
; i
++)
5399 components
[n
++] = offset(src
, bld
, i
);
5401 bld
.LOAD_PAYLOAD(payload
, components
, sz
, header_sz
);
5402 mlen
= header_sz
+ (addr_sz
+ src_sz
) * inst
->exec_size
/ 8;
5404 delete[] components
;
5407 /* Predicate the instruction on the sample mask if no header is
5410 if (!header_sz
&& sample_mask
.file
!= BAD_FILE
&&
5411 sample_mask
.file
!= IMM
) {
5412 const fs_builder ubld
= bld
.group(1, 0).exec_all();
5413 if (inst
->predicate
) {
5414 assert(inst
->predicate
== BRW_PREDICATE_NORMAL
);
5415 assert(!inst
->predicate_inverse
);
5416 assert(inst
->flag_subreg
< 2);
5417 /* Combine the sample mask with the existing predicate by using a
5418 * vertical predication mode.
5420 inst
->predicate
= BRW_PREDICATE_ALIGN1_ALLV
;
5421 ubld
.MOV(retype(brw_flag_subreg(inst
->flag_subreg
+ 2),
5425 inst
->flag_subreg
= 2;
5426 inst
->predicate
= BRW_PREDICATE_NORMAL
;
5427 inst
->predicate_inverse
= false;
5428 ubld
.MOV(retype(brw_flag_subreg(inst
->flag_subreg
), sample_mask
.type
),
5434 switch (inst
->opcode
) {
5435 case SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL
:
5436 case SHADER_OPCODE_BYTE_SCATTERED_READ_LOGICAL
:
5437 /* Byte scattered opcodes go through the normal data cache */
5438 sfid
= GEN7_SFID_DATAPORT_DATA_CACHE
;
5441 case SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
:
5442 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
:
5443 case SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL
:
5444 case SHADER_OPCODE_UNTYPED_ATOMIC_FLOAT_LOGICAL
:
5445 /* Untyped Surface messages go through the data cache but the SFID value
5446 * changed on Haswell.
5448 sfid
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
?
5449 HSW_SFID_DATAPORT_DATA_CACHE_1
:
5450 GEN7_SFID_DATAPORT_DATA_CACHE
);
5453 case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
:
5454 case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
:
5455 case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
:
5456 /* Typed surface messages go through the render cache on IVB and the
5457 * data cache on HSW+.
5459 sfid
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
?
5460 HSW_SFID_DATAPORT_DATA_CACHE_1
:
5461 GEN6_SFID_DATAPORT_RENDER_CACHE
);
5465 unreachable("Unsupported surface opcode");
5469 switch (inst
->opcode
) {
5470 case SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
:
5471 desc
= brw_dp_untyped_surface_rw_desc(devinfo
, inst
->exec_size
,
5472 arg
.ud
, /* num_channels */
5476 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
:
5477 desc
= brw_dp_untyped_surface_rw_desc(devinfo
, inst
->exec_size
,
5478 arg
.ud
, /* num_channels */
5482 case SHADER_OPCODE_BYTE_SCATTERED_READ_LOGICAL
:
5483 desc
= brw_dp_byte_scattered_rw_desc(devinfo
, inst
->exec_size
,
5484 arg
.ud
, /* bit_size */
5488 case SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL
:
5489 desc
= brw_dp_byte_scattered_rw_desc(devinfo
, inst
->exec_size
,
5490 arg
.ud
, /* bit_size */
5494 case SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL
:
5495 desc
= brw_dp_untyped_atomic_desc(devinfo
, inst
->exec_size
,
5496 arg
.ud
, /* atomic_op */
5497 !inst
->dst
.is_null());
5500 case SHADER_OPCODE_UNTYPED_ATOMIC_FLOAT_LOGICAL
:
5501 desc
= brw_dp_untyped_atomic_float_desc(devinfo
, inst
->exec_size
,
5502 arg
.ud
, /* atomic_op */
5503 !inst
->dst
.is_null());
5506 case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
:
5507 desc
= brw_dp_typed_surface_rw_desc(devinfo
, inst
->exec_size
, inst
->group
,
5508 arg
.ud
, /* num_channels */
5512 case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
:
5513 desc
= brw_dp_typed_surface_rw_desc(devinfo
, inst
->exec_size
, inst
->group
,
5514 arg
.ud
, /* num_channels */
5518 case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
:
5519 desc
= brw_dp_typed_atomic_desc(devinfo
, inst
->exec_size
, inst
->group
,
5520 arg
.ud
, /* atomic_op */
5521 !inst
->dst
.is_null());
5525 unreachable("Unknown surface logical instruction");
5528 /* Update the original instruction. */
5529 inst
->opcode
= SHADER_OPCODE_SEND
;
5531 inst
->ex_mlen
= ex_mlen
;
5532 inst
->header_size
= header_sz
;
5533 inst
->send_has_side_effects
= has_side_effects
;
5534 inst
->send_is_volatile
= !has_side_effects
;
5536 /* Set up SFID and descriptors */
5539 if (surface
.file
== IMM
) {
5540 inst
->desc
|= surface
.ud
& 0xff;
5541 inst
->src
[0] = brw_imm_ud(0);
5542 inst
->src
[1] = brw_imm_ud(0); /* ex_desc */
5543 } else if (surface_handle
.file
!= BAD_FILE
) {
5544 /* Bindless surface */
5545 assert(devinfo
->gen
>= 9);
5546 inst
->desc
|= GEN9_BTI_BINDLESS
;
5547 inst
->src
[0] = brw_imm_ud(0);
5549 /* We assume that the driver provided the handle in the top 20 bits so
5550 * we can use the surface handle directly as the extended descriptor.
5552 inst
->src
[1] = retype(surface_handle
, BRW_REGISTER_TYPE_UD
);
5554 const fs_builder ubld
= bld
.exec_all().group(1, 0);
5555 fs_reg tmp
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
5556 ubld
.AND(tmp
, surface
, brw_imm_ud(0xff));
5557 inst
->src
[0] = component(tmp
, 0);
5558 inst
->src
[1] = brw_imm_ud(0); /* ex_desc */
5561 /* Finally, the payload */
5562 inst
->src
[2] = payload
;
5563 inst
->src
[3] = payload2
;
5565 inst
->resize_sources(4);
5569 lower_a64_logical_send(const fs_builder
&bld
, fs_inst
*inst
)
5571 const gen_device_info
*devinfo
= bld
.shader
->devinfo
;
5573 const fs_reg
&addr
= inst
->src
[0];
5574 const fs_reg
&src
= inst
->src
[1];
5575 const unsigned src_comps
= inst
->components_read(1);
5576 assert(inst
->src
[2].file
== IMM
);
5577 const unsigned arg
= inst
->src
[2].ud
;
5578 const bool has_side_effects
= inst
->has_side_effects();
5580 /* If the surface message has side effects and we're a fragment shader, we
5581 * have to predicate with the sample mask to avoid helper invocations.
5583 if (has_side_effects
&& bld
.shader
->stage
== MESA_SHADER_FRAGMENT
) {
5584 inst
->flag_subreg
= 2;
5585 inst
->predicate
= BRW_PREDICATE_NORMAL
;
5586 inst
->predicate_inverse
= false;
5588 fs_reg sample_mask
= bld
.sample_mask_reg();
5589 const fs_builder ubld
= bld
.group(1, 0).exec_all();
5590 ubld
.MOV(retype(brw_flag_subreg(inst
->flag_subreg
), sample_mask
.type
),
5594 fs_reg payload
, payload2
;
5595 unsigned mlen
, ex_mlen
= 0;
5596 if (devinfo
->gen
>= 9) {
5597 /* On Skylake and above, we have SENDS */
5598 mlen
= 2 * (inst
->exec_size
/ 8);
5599 ex_mlen
= src_comps
* type_sz(src
.type
) * inst
->exec_size
/ REG_SIZE
;
5600 payload
= retype(bld
.move_to_vgrf(addr
, 1), BRW_REGISTER_TYPE_UD
);
5601 payload2
= retype(bld
.move_to_vgrf(src
, src_comps
),
5602 BRW_REGISTER_TYPE_UD
);
5604 /* Add two because the address is 64-bit */
5605 const unsigned dwords
= 2 + src_comps
;
5606 mlen
= dwords
* (inst
->exec_size
/ 8);
5612 for (unsigned i
= 0; i
< src_comps
; i
++)
5613 sources
[1 + i
] = offset(src
, bld
, i
);
5615 payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, dwords
);
5616 bld
.LOAD_PAYLOAD(payload
, sources
, 1 + src_comps
, 0);
5620 switch (inst
->opcode
) {
5621 case SHADER_OPCODE_A64_UNTYPED_READ_LOGICAL
:
5622 desc
= brw_dp_a64_untyped_surface_rw_desc(devinfo
, inst
->exec_size
,
5623 arg
, /* num_channels */
5627 case SHADER_OPCODE_A64_UNTYPED_WRITE_LOGICAL
:
5628 desc
= brw_dp_a64_untyped_surface_rw_desc(devinfo
, inst
->exec_size
,
5629 arg
, /* num_channels */
5633 case SHADER_OPCODE_A64_BYTE_SCATTERED_READ_LOGICAL
:
5634 desc
= brw_dp_a64_byte_scattered_rw_desc(devinfo
, inst
->exec_size
,
5639 case SHADER_OPCODE_A64_BYTE_SCATTERED_WRITE_LOGICAL
:
5640 desc
= brw_dp_a64_byte_scattered_rw_desc(devinfo
, inst
->exec_size
,
5645 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_LOGICAL
:
5646 desc
= brw_dp_a64_untyped_atomic_desc(devinfo
, inst
->exec_size
, 32,
5647 arg
, /* atomic_op */
5648 !inst
->dst
.is_null());
5651 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_INT64_LOGICAL
:
5652 desc
= brw_dp_a64_untyped_atomic_desc(devinfo
, inst
->exec_size
, 64,
5653 arg
, /* atomic_op */
5654 !inst
->dst
.is_null());
5658 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_FLOAT_LOGICAL
:
5659 desc
= brw_dp_a64_untyped_atomic_float_desc(devinfo
, inst
->exec_size
,
5660 arg
, /* atomic_op */
5661 !inst
->dst
.is_null());
5665 unreachable("Unknown A64 logical instruction");
5668 /* Update the original instruction. */
5669 inst
->opcode
= SHADER_OPCODE_SEND
;
5671 inst
->ex_mlen
= ex_mlen
;
5672 inst
->header_size
= 0;
5673 inst
->send_has_side_effects
= has_side_effects
;
5674 inst
->send_is_volatile
= !has_side_effects
;
5676 /* Set up SFID and descriptors */
5677 inst
->sfid
= HSW_SFID_DATAPORT_DATA_CACHE_1
;
5679 inst
->resize_sources(4);
5680 inst
->src
[0] = brw_imm_ud(0); /* desc */
5681 inst
->src
[1] = brw_imm_ud(0); /* ex_desc */
5682 inst
->src
[2] = payload
;
5683 inst
->src
[3] = payload2
;
5687 lower_varying_pull_constant_logical_send(const fs_builder
&bld
, fs_inst
*inst
)
5689 const gen_device_info
*devinfo
= bld
.shader
->devinfo
;
5691 if (devinfo
->gen
>= 7) {
5692 fs_reg index
= inst
->src
[0];
5693 /* We are switching the instruction from an ALU-like instruction to a
5694 * send-from-grf instruction. Since sends can't handle strides or
5695 * source modifiers, we have to make a copy of the offset source.
5697 fs_reg offset
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
5698 bld
.MOV(offset
, inst
->src
[1]);
5700 const unsigned simd_mode
=
5701 inst
->exec_size
<= 8 ? BRW_SAMPLER_SIMD_MODE_SIMD8
:
5702 BRW_SAMPLER_SIMD_MODE_SIMD16
;
5704 inst
->opcode
= SHADER_OPCODE_SEND
;
5705 inst
->mlen
= inst
->exec_size
/ 8;
5706 inst
->resize_sources(3);
5708 inst
->sfid
= BRW_SFID_SAMPLER
;
5709 inst
->desc
= brw_sampler_desc(devinfo
, 0, 0,
5710 GEN5_SAMPLER_MESSAGE_SAMPLE_LD
,
5712 if (index
.file
== IMM
) {
5713 inst
->desc
|= index
.ud
& 0xff;
5714 inst
->src
[0] = brw_imm_ud(0);
5716 const fs_builder ubld
= bld
.exec_all().group(1, 0);
5717 fs_reg tmp
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
5718 ubld
.AND(tmp
, index
, brw_imm_ud(0xff));
5719 inst
->src
[0] = component(tmp
, 0);
5721 inst
->src
[1] = brw_imm_ud(0); /* ex_desc */
5722 inst
->src
[2] = offset
; /* payload */
5724 const fs_reg
payload(MRF
, FIRST_PULL_LOAD_MRF(devinfo
->gen
),
5725 BRW_REGISTER_TYPE_UD
);
5727 bld
.MOV(byte_offset(payload
, REG_SIZE
), inst
->src
[1]);
5729 inst
->opcode
= FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN4
;
5730 inst
->resize_sources(1);
5731 inst
->base_mrf
= payload
.nr
;
5732 inst
->header_size
= 1;
5733 inst
->mlen
= 1 + inst
->exec_size
/ 8;
5738 lower_math_logical_send(const fs_builder
&bld
, fs_inst
*inst
)
5740 assert(bld
.shader
->devinfo
->gen
< 6);
5743 inst
->mlen
= inst
->sources
* inst
->exec_size
/ 8;
5745 if (inst
->sources
> 1) {
5746 /* From the Ironlake PRM, Volume 4, Part 1, Section 6.1.13
5747 * "Message Payload":
5749 * "Operand0[7]. For the INT DIV functions, this operand is the
5752 * "Operand1[7]. For the INT DIV functions, this operand is the
5755 const bool is_int_div
= inst
->opcode
!= SHADER_OPCODE_POW
;
5756 const fs_reg src0
= is_int_div
? inst
->src
[1] : inst
->src
[0];
5757 const fs_reg src1
= is_int_div
? inst
->src
[0] : inst
->src
[1];
5759 inst
->resize_sources(1);
5760 inst
->src
[0] = src0
;
5762 assert(inst
->exec_size
== 8);
5763 bld
.MOV(fs_reg(MRF
, inst
->base_mrf
+ 1, src1
.type
), src1
);
5768 fs_visitor::lower_logical_sends()
5770 bool progress
= false;
5772 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
5773 const fs_builder
ibld(this, block
, inst
);
5775 switch (inst
->opcode
) {
5776 case FS_OPCODE_FB_WRITE_LOGICAL
:
5777 assert(stage
== MESA_SHADER_FRAGMENT
);
5778 lower_fb_write_logical_send(ibld
, inst
,
5779 brw_wm_prog_data(prog_data
),
5780 (const brw_wm_prog_key
*)key
,
5784 case FS_OPCODE_FB_READ_LOGICAL
:
5785 lower_fb_read_logical_send(ibld
, inst
);
5788 case SHADER_OPCODE_TEX_LOGICAL
:
5789 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TEX
);
5792 case SHADER_OPCODE_TXD_LOGICAL
:
5793 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXD
);
5796 case SHADER_OPCODE_TXF_LOGICAL
:
5797 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF
);
5800 case SHADER_OPCODE_TXL_LOGICAL
:
5801 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXL
);
5804 case SHADER_OPCODE_TXS_LOGICAL
:
5805 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXS
);
5808 case SHADER_OPCODE_IMAGE_SIZE_LOGICAL
:
5809 lower_sampler_logical_send(ibld
, inst
,
5810 SHADER_OPCODE_IMAGE_SIZE_LOGICAL
);
5813 case FS_OPCODE_TXB_LOGICAL
:
5814 lower_sampler_logical_send(ibld
, inst
, FS_OPCODE_TXB
);
5817 case SHADER_OPCODE_TXF_CMS_LOGICAL
:
5818 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_CMS
);
5821 case SHADER_OPCODE_TXF_CMS_W_LOGICAL
:
5822 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_CMS_W
);
5825 case SHADER_OPCODE_TXF_UMS_LOGICAL
:
5826 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_UMS
);
5829 case SHADER_OPCODE_TXF_MCS_LOGICAL
:
5830 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_MCS
);
5833 case SHADER_OPCODE_LOD_LOGICAL
:
5834 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_LOD
);
5837 case SHADER_OPCODE_TG4_LOGICAL
:
5838 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TG4
);
5841 case SHADER_OPCODE_TG4_OFFSET_LOGICAL
:
5842 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TG4_OFFSET
);
5845 case SHADER_OPCODE_SAMPLEINFO_LOGICAL
:
5846 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_SAMPLEINFO
);
5849 case SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
:
5850 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
:
5851 case SHADER_OPCODE_BYTE_SCATTERED_READ_LOGICAL
:
5852 case SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL
:
5853 case SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL
:
5854 case SHADER_OPCODE_UNTYPED_ATOMIC_FLOAT_LOGICAL
:
5855 case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
:
5856 case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
:
5857 case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
:
5858 lower_surface_logical_send(ibld
, inst
);
5861 case SHADER_OPCODE_A64_UNTYPED_WRITE_LOGICAL
:
5862 case SHADER_OPCODE_A64_UNTYPED_READ_LOGICAL
:
5863 case SHADER_OPCODE_A64_BYTE_SCATTERED_WRITE_LOGICAL
:
5864 case SHADER_OPCODE_A64_BYTE_SCATTERED_READ_LOGICAL
:
5865 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_LOGICAL
:
5866 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_INT64_LOGICAL
:
5867 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_FLOAT_LOGICAL
:
5868 lower_a64_logical_send(ibld
, inst
);
5871 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_LOGICAL
:
5872 lower_varying_pull_constant_logical_send(ibld
, inst
);
5875 case SHADER_OPCODE_RCP
:
5876 case SHADER_OPCODE_RSQ
:
5877 case SHADER_OPCODE_SQRT
:
5878 case SHADER_OPCODE_EXP2
:
5879 case SHADER_OPCODE_LOG2
:
5880 case SHADER_OPCODE_SIN
:
5881 case SHADER_OPCODE_COS
:
5882 case SHADER_OPCODE_POW
:
5883 case SHADER_OPCODE_INT_QUOTIENT
:
5884 case SHADER_OPCODE_INT_REMAINDER
:
5885 /* The math opcodes are overloaded for the send-like and
5886 * expression-like instructions which seems kind of icky. Gen6+ has
5887 * a native (but rather quirky) MATH instruction so we don't need to
5888 * do anything here. On Gen4-5 we'll have to lower the Gen6-like
5889 * logical instructions (which we can easily recognize because they
5890 * have mlen = 0) into send-like virtual instructions.
5892 if (devinfo
->gen
< 6 && inst
->mlen
== 0) {
5893 lower_math_logical_send(ibld
, inst
);
5908 invalidate_live_intervals();
5914 is_mixed_float_with_fp32_dst(const fs_inst
*inst
)
5916 /* This opcode sometimes uses :W type on the source even if the operand is
5917 * a :HF, because in gen7 there is no support for :HF, and thus it uses :W.
5919 if (inst
->opcode
== BRW_OPCODE_F16TO32
)
5922 if (inst
->dst
.type
!= BRW_REGISTER_TYPE_F
)
5925 for (int i
= 0; i
< inst
->sources
; i
++) {
5926 if (inst
->src
[i
].type
== BRW_REGISTER_TYPE_HF
)
5934 is_mixed_float_with_packed_fp16_dst(const fs_inst
*inst
)
5936 /* This opcode sometimes uses :W type on the destination even if the
5937 * destination is a :HF, because in gen7 there is no support for :HF, and
5940 if (inst
->opcode
== BRW_OPCODE_F32TO16
&&
5941 inst
->dst
.stride
== 1)
5944 if (inst
->dst
.type
!= BRW_REGISTER_TYPE_HF
||
5945 inst
->dst
.stride
!= 1)
5948 for (int i
= 0; i
< inst
->sources
; i
++) {
5949 if (inst
->src
[i
].type
== BRW_REGISTER_TYPE_F
)
5957 * Get the closest allowed SIMD width for instruction \p inst accounting for
5958 * some common regioning and execution control restrictions that apply to FPU
5959 * instructions. These restrictions don't necessarily have any relevance to
5960 * instructions not executed by the FPU pipeline like extended math, control
5961 * flow or send message instructions.
5963 * For virtual opcodes it's really up to the instruction -- In some cases
5964 * (e.g. where a virtual instruction unrolls into a simple sequence of FPU
5965 * instructions) it may simplify virtual instruction lowering if we can
5966 * enforce FPU-like regioning restrictions already on the virtual instruction,
5967 * in other cases (e.g. virtual send-like instructions) this may be
5968 * excessively restrictive.
5971 get_fpu_lowered_simd_width(const struct gen_device_info
*devinfo
,
5972 const fs_inst
*inst
)
5974 /* Maximum execution size representable in the instruction controls. */
5975 unsigned max_width
= MIN2(32, inst
->exec_size
);
5977 /* According to the PRMs:
5978 * "A. In Direct Addressing mode, a source cannot span more than 2
5979 * adjacent GRF registers.
5980 * B. A destination cannot span more than 2 adjacent GRF registers."
5982 * Look for the source or destination with the largest register region
5983 * which is the one that is going to limit the overall execution size of
5984 * the instruction due to this rule.
5986 unsigned reg_count
= DIV_ROUND_UP(inst
->size_written
, REG_SIZE
);
5988 for (unsigned i
= 0; i
< inst
->sources
; i
++)
5989 reg_count
= MAX2(reg_count
, DIV_ROUND_UP(inst
->size_read(i
), REG_SIZE
));
5991 /* Calculate the maximum execution size of the instruction based on the
5992 * factor by which it goes over the hardware limit of 2 GRFs.
5995 max_width
= MIN2(max_width
, inst
->exec_size
/ DIV_ROUND_UP(reg_count
, 2));
5997 /* According to the IVB PRMs:
5998 * "When destination spans two registers, the source MUST span two
5999 * registers. The exception to the above rule:
6001 * - When source is scalar, the source registers are not incremented.
6002 * - When source is packed integer Word and destination is packed
6003 * integer DWord, the source register is not incremented but the
6004 * source sub register is incremented."
6006 * The hardware specs from Gen4 to Gen7.5 mention similar regioning
6007 * restrictions. The code below intentionally doesn't check whether the
6008 * destination type is integer because empirically the hardware doesn't
6009 * seem to care what the actual type is as long as it's dword-aligned.
6011 if (devinfo
->gen
< 8) {
6012 for (unsigned i
= 0; i
< inst
->sources
; i
++) {
6013 /* IVB implements DF scalars as <0;2,1> regions. */
6014 const bool is_scalar_exception
= is_uniform(inst
->src
[i
]) &&
6015 (devinfo
->is_haswell
|| type_sz(inst
->src
[i
].type
) != 8);
6016 const bool is_packed_word_exception
=
6017 type_sz(inst
->dst
.type
) == 4 && inst
->dst
.stride
== 1 &&
6018 type_sz(inst
->src
[i
].type
) == 2 && inst
->src
[i
].stride
== 1;
6020 /* We check size_read(i) against size_written instead of REG_SIZE
6021 * because we want to properly handle SIMD32. In SIMD32, you can end
6022 * up with writes to 4 registers and a source that reads 2 registers
6023 * and we may still need to lower all the way to SIMD8 in that case.
6025 if (inst
->size_written
> REG_SIZE
&&
6026 inst
->size_read(i
) != 0 &&
6027 inst
->size_read(i
) < inst
->size_written
&&
6028 !is_scalar_exception
&& !is_packed_word_exception
) {
6029 const unsigned reg_count
= DIV_ROUND_UP(inst
->size_written
, REG_SIZE
);
6030 max_width
= MIN2(max_width
, inst
->exec_size
/ reg_count
);
6035 if (devinfo
->gen
< 6) {
6036 /* From the G45 PRM, Volume 4 Page 361:
6038 * "Operand Alignment Rule: With the exceptions listed below, a
6039 * source/destination operand in general should be aligned to even
6040 * 256-bit physical register with a region size equal to two 256-bit
6041 * physical registers."
6043 * Normally we enforce this by allocating virtual registers to the
6044 * even-aligned class. But we need to handle payload registers.
6046 for (unsigned i
= 0; i
< inst
->sources
; i
++) {
6047 if (inst
->src
[i
].file
== FIXED_GRF
&& (inst
->src
[i
].nr
& 1) &&
6048 inst
->size_read(i
) > REG_SIZE
) {
6049 max_width
= MIN2(max_width
, 8);
6054 /* From the IVB PRMs:
6055 * "When an instruction is SIMD32, the low 16 bits of the execution mask
6056 * are applied for both halves of the SIMD32 instruction. If different
6057 * execution mask channels are required, split the instruction into two
6058 * SIMD16 instructions."
6060 * There is similar text in the HSW PRMs. Gen4-6 don't even implement
6061 * 32-wide control flow support in hardware and will behave similarly.
6063 if (devinfo
->gen
< 8 && !inst
->force_writemask_all
)
6064 max_width
= MIN2(max_width
, 16);
6066 /* From the IVB PRMs (applies to HSW too):
6067 * "Instructions with condition modifiers must not use SIMD32."
6069 * From the BDW PRMs (applies to later hardware too):
6070 * "Ternary instruction with condition modifiers must not use SIMD32."
6072 if (inst
->conditional_mod
&& (devinfo
->gen
< 8 || inst
->is_3src(devinfo
)))
6073 max_width
= MIN2(max_width
, 16);
6075 /* From the IVB PRMs (applies to other devices that don't have the
6076 * gen_device_info::supports_simd16_3src flag set):
6077 * "In Align16 access mode, SIMD16 is not allowed for DW operations and
6078 * SIMD8 is not allowed for DF operations."
6080 if (inst
->is_3src(devinfo
) && !devinfo
->supports_simd16_3src
)
6081 max_width
= MIN2(max_width
, inst
->exec_size
/ reg_count
);
6083 /* Pre-Gen8 EUs are hardwired to use the QtrCtrl+1 (where QtrCtrl is
6084 * the 8-bit quarter of the execution mask signals specified in the
6085 * instruction control fields) for the second compressed half of any
6086 * single-precision instruction (for double-precision instructions
6087 * it's hardwired to use NibCtrl+1, at least on HSW), which means that
6088 * the EU will apply the wrong execution controls for the second
6089 * sequential GRF write if the number of channels per GRF is not exactly
6090 * eight in single-precision mode (or four in double-float mode).
6092 * In this situation we calculate the maximum size of the split
6093 * instructions so they only ever write to a single register.
6095 if (devinfo
->gen
< 8 && inst
->size_written
> REG_SIZE
&&
6096 !inst
->force_writemask_all
) {
6097 const unsigned channels_per_grf
= inst
->exec_size
/
6098 DIV_ROUND_UP(inst
->size_written
, REG_SIZE
);
6099 const unsigned exec_type_size
= get_exec_type_size(inst
);
6100 assert(exec_type_size
);
6102 /* The hardware shifts exactly 8 channels per compressed half of the
6103 * instruction in single-precision mode and exactly 4 in double-precision.
6105 if (channels_per_grf
!= (exec_type_size
== 8 ? 4 : 8))
6106 max_width
= MIN2(max_width
, channels_per_grf
);
6108 /* Lower all non-force_writemask_all DF instructions to SIMD4 on IVB/BYT
6109 * because HW applies the same channel enable signals to both halves of
6110 * the compressed instruction which will be just wrong under
6111 * non-uniform control flow.
6113 if (devinfo
->gen
== 7 && !devinfo
->is_haswell
&&
6114 (exec_type_size
== 8 || type_sz(inst
->dst
.type
) == 8))
6115 max_width
= MIN2(max_width
, 4);
6118 /* From the SKL PRM, Special Restrictions for Handling Mixed Mode
6121 * "No SIMD16 in mixed mode when destination is f32. Instruction
6122 * execution size must be no more than 8."
6124 * FIXME: the simulator doesn't seem to complain if we don't do this and
6125 * empirical testing with existing CTS tests show that they pass just fine
6126 * without implementing this, however, since our interpretation of the PRM
6127 * is that conversion MOVs between HF and F are still mixed-float
6128 * instructions (and therefore subject to this restriction) we decided to
6129 * split them to be safe. Might be useful to do additional investigation to
6130 * lift the restriction if we can ensure that it is safe though, since these
6131 * conversions are common when half-float types are involved since many
6132 * instructions do not support HF types and conversions from/to F are
6135 if (is_mixed_float_with_fp32_dst(inst
))
6136 max_width
= MIN2(max_width
, 8);
6138 /* From the SKL PRM, Special Restrictions for Handling Mixed Mode
6141 * "No SIMD16 in mixed mode when destination is packed f16 for both
6142 * Align1 and Align16."
6144 if (is_mixed_float_with_packed_fp16_dst(inst
))
6145 max_width
= MIN2(max_width
, 8);
6147 /* Only power-of-two execution sizes are representable in the instruction
6150 return 1 << _mesa_logbase2(max_width
);
6154 * Get the maximum allowed SIMD width for instruction \p inst accounting for
6155 * various payload size restrictions that apply to sampler message
6158 * This is only intended to provide a maximum theoretical bound for the
6159 * execution size of the message based on the number of argument components
6160 * alone, which in most cases will determine whether the SIMD8 or SIMD16
6161 * variant of the message can be used, though some messages may have
6162 * additional restrictions not accounted for here (e.g. pre-ILK hardware uses
6163 * the message length to determine the exact SIMD width and argument count,
6164 * which makes a number of sampler message combinations impossible to
6168 get_sampler_lowered_simd_width(const struct gen_device_info
*devinfo
,
6169 const fs_inst
*inst
)
6171 /* If we have a min_lod parameter on anything other than a simple sample
6172 * message, it will push it over 5 arguments and we have to fall back to
6175 if (inst
->opcode
!= SHADER_OPCODE_TEX
&&
6176 inst
->components_read(TEX_LOGICAL_SRC_MIN_LOD
))
6179 /* Calculate the number of coordinate components that have to be present
6180 * assuming that additional arguments follow the texel coordinates in the
6181 * message payload. On IVB+ there is no need for padding, on ILK-SNB we
6182 * need to pad to four or three components depending on the message,
6183 * pre-ILK we need to pad to at most three components.
6185 const unsigned req_coord_components
=
6186 (devinfo
->gen
>= 7 ||
6187 !inst
->components_read(TEX_LOGICAL_SRC_COORDINATE
)) ? 0 :
6188 (devinfo
->gen
>= 5 && inst
->opcode
!= SHADER_OPCODE_TXF_LOGICAL
&&
6189 inst
->opcode
!= SHADER_OPCODE_TXF_CMS_LOGICAL
) ? 4 :
6192 /* On Gen9+ the LOD argument is for free if we're able to use the LZ
6193 * variant of the TXL or TXF message.
6195 const bool implicit_lod
= devinfo
->gen
>= 9 &&
6196 (inst
->opcode
== SHADER_OPCODE_TXL
||
6197 inst
->opcode
== SHADER_OPCODE_TXF
) &&
6198 inst
->src
[TEX_LOGICAL_SRC_LOD
].is_zero();
6200 /* Calculate the total number of argument components that need to be passed
6201 * to the sampler unit.
6203 const unsigned num_payload_components
=
6204 MAX2(inst
->components_read(TEX_LOGICAL_SRC_COORDINATE
),
6205 req_coord_components
) +
6206 inst
->components_read(TEX_LOGICAL_SRC_SHADOW_C
) +
6207 (implicit_lod
? 0 : inst
->components_read(TEX_LOGICAL_SRC_LOD
)) +
6208 inst
->components_read(TEX_LOGICAL_SRC_LOD2
) +
6209 inst
->components_read(TEX_LOGICAL_SRC_SAMPLE_INDEX
) +
6210 (inst
->opcode
== SHADER_OPCODE_TG4_OFFSET_LOGICAL
?
6211 inst
->components_read(TEX_LOGICAL_SRC_TG4_OFFSET
) : 0) +
6212 inst
->components_read(TEX_LOGICAL_SRC_MCS
);
6214 /* SIMD16 messages with more than five arguments exceed the maximum message
6215 * size supported by the sampler, regardless of whether a header is
6218 return MIN2(inst
->exec_size
,
6219 num_payload_components
> MAX_SAMPLER_MESSAGE_SIZE
/ 2 ? 8 : 16);
6223 * Get the closest native SIMD width supported by the hardware for instruction
6224 * \p inst. The instruction will be left untouched by
6225 * fs_visitor::lower_simd_width() if the returned value is equal to the
6226 * original execution size.
6229 get_lowered_simd_width(const struct gen_device_info
*devinfo
,
6230 const fs_inst
*inst
)
6232 switch (inst
->opcode
) {
6233 case BRW_OPCODE_MOV
:
6234 case BRW_OPCODE_SEL
:
6235 case BRW_OPCODE_NOT
:
6236 case BRW_OPCODE_AND
:
6238 case BRW_OPCODE_XOR
:
6239 case BRW_OPCODE_SHR
:
6240 case BRW_OPCODE_SHL
:
6241 case BRW_OPCODE_ASR
:
6242 case BRW_OPCODE_CMPN
:
6243 case BRW_OPCODE_CSEL
:
6244 case BRW_OPCODE_F32TO16
:
6245 case BRW_OPCODE_F16TO32
:
6246 case BRW_OPCODE_BFREV
:
6247 case BRW_OPCODE_BFE
:
6248 case BRW_OPCODE_ADD
:
6249 case BRW_OPCODE_MUL
:
6250 case BRW_OPCODE_AVG
:
6251 case BRW_OPCODE_FRC
:
6252 case BRW_OPCODE_RNDU
:
6253 case BRW_OPCODE_RNDD
:
6254 case BRW_OPCODE_RNDE
:
6255 case BRW_OPCODE_RNDZ
:
6256 case BRW_OPCODE_LZD
:
6257 case BRW_OPCODE_FBH
:
6258 case BRW_OPCODE_FBL
:
6259 case BRW_OPCODE_CBIT
:
6260 case BRW_OPCODE_SAD2
:
6261 case BRW_OPCODE_MAD
:
6262 case BRW_OPCODE_LRP
:
6263 case FS_OPCODE_PACK
:
6264 case SHADER_OPCODE_SEL_EXEC
:
6265 case SHADER_OPCODE_CLUSTER_BROADCAST
:
6266 return get_fpu_lowered_simd_width(devinfo
, inst
);
6268 case BRW_OPCODE_CMP
: {
6269 /* The Ivybridge/BayTrail WaCMPInstFlagDepClearedEarly workaround says that
6270 * when the destination is a GRF the dependency-clear bit on the flag
6271 * register is cleared early.
6273 * Suggested workarounds are to disable coissuing CMP instructions
6274 * or to split CMP(16) instructions into two CMP(8) instructions.
6276 * We choose to split into CMP(8) instructions since disabling
6277 * coissuing would affect CMP instructions not otherwise affected by
6280 const unsigned max_width
= (devinfo
->gen
== 7 && !devinfo
->is_haswell
&&
6281 !inst
->dst
.is_null() ? 8 : ~0);
6282 return MIN2(max_width
, get_fpu_lowered_simd_width(devinfo
, inst
));
6284 case BRW_OPCODE_BFI1
:
6285 case BRW_OPCODE_BFI2
:
6286 /* The Haswell WaForceSIMD8ForBFIInstruction workaround says that we
6288 * "Force BFI instructions to be executed always in SIMD8."
6290 return MIN2(devinfo
->is_haswell
? 8 : ~0u,
6291 get_fpu_lowered_simd_width(devinfo
, inst
));
6294 assert(inst
->src
[0].file
== BAD_FILE
|| inst
->exec_size
<= 16);
6295 return inst
->exec_size
;
6297 case SHADER_OPCODE_RCP
:
6298 case SHADER_OPCODE_RSQ
:
6299 case SHADER_OPCODE_SQRT
:
6300 case SHADER_OPCODE_EXP2
:
6301 case SHADER_OPCODE_LOG2
:
6302 case SHADER_OPCODE_SIN
:
6303 case SHADER_OPCODE_COS
: {
6304 /* Unary extended math instructions are limited to SIMD8 on Gen4 and
6305 * Gen6. Extended Math Function is limited to SIMD8 with half-float.
6307 if (devinfo
->gen
== 6 || (devinfo
->gen
== 4 && !devinfo
->is_g4x
))
6308 return MIN2(8, inst
->exec_size
);
6309 if (inst
->dst
.type
== BRW_REGISTER_TYPE_HF
)
6310 return MIN2(8, inst
->exec_size
);
6311 return MIN2(16, inst
->exec_size
);
6314 case SHADER_OPCODE_POW
: {
6315 /* SIMD16 is only allowed on Gen7+. Extended Math Function is limited
6316 * to SIMD8 with half-float
6318 if (devinfo
->gen
< 7)
6319 return MIN2(8, inst
->exec_size
);
6320 if (inst
->dst
.type
== BRW_REGISTER_TYPE_HF
)
6321 return MIN2(8, inst
->exec_size
);
6322 return MIN2(16, inst
->exec_size
);
6325 case SHADER_OPCODE_INT_QUOTIENT
:
6326 case SHADER_OPCODE_INT_REMAINDER
:
6327 /* Integer division is limited to SIMD8 on all generations. */
6328 return MIN2(8, inst
->exec_size
);
6330 case FS_OPCODE_LINTERP
:
6331 case SHADER_OPCODE_GET_BUFFER_SIZE
:
6332 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
6333 case FS_OPCODE_PACK_HALF_2x16_SPLIT
:
6334 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
6335 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
6336 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
6337 return MIN2(16, inst
->exec_size
);
6339 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_LOGICAL
:
6340 /* Pre-ILK hardware doesn't have a SIMD8 variant of the texel fetch
6341 * message used to implement varying pull constant loads, so expand it
6342 * to SIMD16. An alternative with longer message payload length but
6343 * shorter return payload would be to use the SIMD8 sampler message that
6344 * takes (header, u, v, r) as parameters instead of (header, u).
6346 return (devinfo
->gen
== 4 ? 16 : MIN2(16, inst
->exec_size
));
6348 case FS_OPCODE_DDX_COARSE
:
6349 case FS_OPCODE_DDX_FINE
:
6350 case FS_OPCODE_DDY_COARSE
:
6351 case FS_OPCODE_DDY_FINE
:
6352 /* The implementation of this virtual opcode may require emitting
6353 * compressed Align16 instructions, which are severely limited on some
6356 * From the Ivy Bridge PRM, volume 4 part 3, section 3.3.9 (Register
6357 * Region Restrictions):
6359 * "In Align16 access mode, SIMD16 is not allowed for DW operations
6360 * and SIMD8 is not allowed for DF operations."
6362 * In this context, "DW operations" means "operations acting on 32-bit
6363 * values", so it includes operations on floats.
6365 * Gen4 has a similar restriction. From the i965 PRM, section 11.5.3
6366 * (Instruction Compression -> Rules and Restrictions):
6368 * "A compressed instruction must be in Align1 access mode. Align16
6369 * mode instructions cannot be compressed."
6371 * Similar text exists in the g45 PRM.
6373 * Empirically, compressed align16 instructions using odd register
6374 * numbers don't appear to work on Sandybridge either.
6376 return (devinfo
->gen
== 4 || devinfo
->gen
== 6 ||
6377 (devinfo
->gen
== 7 && !devinfo
->is_haswell
) ?
6378 MIN2(8, inst
->exec_size
) : MIN2(16, inst
->exec_size
));
6380 case SHADER_OPCODE_MULH
:
6381 /* MULH is lowered to the MUL/MACH sequence using the accumulator, which
6382 * is 8-wide on Gen7+.
6384 return (devinfo
->gen
>= 7 ? 8 :
6385 get_fpu_lowered_simd_width(devinfo
, inst
));
6387 case FS_OPCODE_FB_WRITE_LOGICAL
:
6388 /* Gen6 doesn't support SIMD16 depth writes but we cannot handle them
6391 assert(devinfo
->gen
!= 6 ||
6392 inst
->src
[FB_WRITE_LOGICAL_SRC_SRC_DEPTH
].file
== BAD_FILE
||
6393 inst
->exec_size
== 8);
6394 /* Dual-source FB writes are unsupported in SIMD16 mode. */
6395 return (inst
->src
[FB_WRITE_LOGICAL_SRC_COLOR1
].file
!= BAD_FILE
?
6396 8 : MIN2(16, inst
->exec_size
));
6398 case FS_OPCODE_FB_READ_LOGICAL
:
6399 return MIN2(16, inst
->exec_size
);
6401 case SHADER_OPCODE_TEX_LOGICAL
:
6402 case SHADER_OPCODE_TXF_CMS_LOGICAL
:
6403 case SHADER_OPCODE_TXF_UMS_LOGICAL
:
6404 case SHADER_OPCODE_TXF_MCS_LOGICAL
:
6405 case SHADER_OPCODE_LOD_LOGICAL
:
6406 case SHADER_OPCODE_TG4_LOGICAL
:
6407 case SHADER_OPCODE_SAMPLEINFO_LOGICAL
:
6408 case SHADER_OPCODE_TXF_CMS_W_LOGICAL
:
6409 case SHADER_OPCODE_TG4_OFFSET_LOGICAL
:
6410 return get_sampler_lowered_simd_width(devinfo
, inst
);
6412 case SHADER_OPCODE_TXD_LOGICAL
:
6413 /* TXD is unsupported in SIMD16 mode. */
6416 case SHADER_OPCODE_TXL_LOGICAL
:
6417 case FS_OPCODE_TXB_LOGICAL
:
6418 /* Only one execution size is representable pre-ILK depending on whether
6419 * the shadow reference argument is present.
6421 if (devinfo
->gen
== 4)
6422 return inst
->src
[TEX_LOGICAL_SRC_SHADOW_C
].file
== BAD_FILE
? 16 : 8;
6424 return get_sampler_lowered_simd_width(devinfo
, inst
);
6426 case SHADER_OPCODE_TXF_LOGICAL
:
6427 case SHADER_OPCODE_TXS_LOGICAL
:
6428 /* Gen4 doesn't have SIMD8 variants for the RESINFO and LD-with-LOD
6429 * messages. Use SIMD16 instead.
6431 if (devinfo
->gen
== 4)
6434 return get_sampler_lowered_simd_width(devinfo
, inst
);
6436 case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
:
6437 case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
:
6438 case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
:
6441 case SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL
:
6442 case SHADER_OPCODE_UNTYPED_ATOMIC_FLOAT_LOGICAL
:
6443 case SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
:
6444 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
:
6445 case SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL
:
6446 case SHADER_OPCODE_BYTE_SCATTERED_READ_LOGICAL
:
6447 return MIN2(16, inst
->exec_size
);
6449 case SHADER_OPCODE_A64_UNTYPED_WRITE_LOGICAL
:
6450 case SHADER_OPCODE_A64_UNTYPED_READ_LOGICAL
:
6451 case SHADER_OPCODE_A64_BYTE_SCATTERED_WRITE_LOGICAL
:
6452 case SHADER_OPCODE_A64_BYTE_SCATTERED_READ_LOGICAL
:
6453 return devinfo
->gen
<= 8 ? 8 : MIN2(16, inst
->exec_size
);
6455 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_LOGICAL
:
6456 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_INT64_LOGICAL
:
6457 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_FLOAT_LOGICAL
:
6460 case SHADER_OPCODE_URB_READ_SIMD8
:
6461 case SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
:
6462 case SHADER_OPCODE_URB_WRITE_SIMD8
:
6463 case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
6464 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
:
6465 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
6466 return MIN2(8, inst
->exec_size
);
6468 case SHADER_OPCODE_QUAD_SWIZZLE
: {
6469 const unsigned swiz
= inst
->src
[1].ud
;
6470 return (is_uniform(inst
->src
[0]) ?
6471 get_fpu_lowered_simd_width(devinfo
, inst
) :
6472 devinfo
->gen
< 11 && type_sz(inst
->src
[0].type
) == 4 ? 8 :
6473 swiz
== BRW_SWIZZLE_XYXY
|| swiz
== BRW_SWIZZLE_ZWZW
? 4 :
6474 get_fpu_lowered_simd_width(devinfo
, inst
));
6476 case SHADER_OPCODE_MOV_INDIRECT
: {
6477 /* From IVB and HSW PRMs:
6479 * "2.When the destination requires two registers and the sources are
6480 * indirect, the sources must use 1x1 regioning mode.
6482 * In case of DF instructions in HSW/IVB, the exec_size is limited by
6483 * the EU decompression logic not handling VxH indirect addressing
6486 const unsigned max_size
= (devinfo
->gen
>= 8 ? 2 : 1) * REG_SIZE
;
6487 /* Prior to Broadwell, we only have 8 address subregisters. */
6488 return MIN3(devinfo
->gen
>= 8 ? 16 : 8,
6489 max_size
/ (inst
->dst
.stride
* type_sz(inst
->dst
.type
)),
6493 case SHADER_OPCODE_LOAD_PAYLOAD
: {
6494 const unsigned reg_count
=
6495 DIV_ROUND_UP(inst
->dst
.component_size(inst
->exec_size
), REG_SIZE
);
6497 if (reg_count
> 2) {
6498 /* Only LOAD_PAYLOAD instructions with per-channel destination region
6499 * can be easily lowered (which excludes headers and heterogeneous
6502 assert(!inst
->header_size
);
6503 for (unsigned i
= 0; i
< inst
->sources
; i
++)
6504 assert(type_sz(inst
->dst
.type
) == type_sz(inst
->src
[i
].type
) ||
6505 inst
->src
[i
].file
== BAD_FILE
);
6507 return inst
->exec_size
/ DIV_ROUND_UP(reg_count
, 2);
6509 return inst
->exec_size
;
6513 return inst
->exec_size
;
6518 * Return true if splitting out the group of channels of instruction \p inst
6519 * given by lbld.group() requires allocating a temporary for the i-th source
6520 * of the lowered instruction.
6523 needs_src_copy(const fs_builder
&lbld
, const fs_inst
*inst
, unsigned i
)
6525 return !(is_periodic(inst
->src
[i
], lbld
.dispatch_width()) ||
6526 (inst
->components_read(i
) == 1 &&
6527 lbld
.dispatch_width() <= inst
->exec_size
)) ||
6528 (inst
->flags_written() &
6529 flag_mask(inst
->src
[i
], type_sz(inst
->src
[i
].type
)));
6533 * Extract the data that would be consumed by the channel group given by
6534 * lbld.group() from the i-th source region of instruction \p inst and return
6535 * it as result in packed form.
6538 emit_unzip(const fs_builder
&lbld
, fs_inst
*inst
, unsigned i
)
6540 assert(lbld
.group() >= inst
->group
);
6542 /* Specified channel group from the source region. */
6543 const fs_reg src
= horiz_offset(inst
->src
[i
], lbld
.group() - inst
->group
);
6545 if (needs_src_copy(lbld
, inst
, i
)) {
6546 /* Builder of the right width to perform the copy avoiding uninitialized
6547 * data if the lowered execution size is greater than the original
6548 * execution size of the instruction.
6550 const fs_builder cbld
= lbld
.group(MIN2(lbld
.dispatch_width(),
6551 inst
->exec_size
), 0);
6552 const fs_reg tmp
= lbld
.vgrf(inst
->src
[i
].type
, inst
->components_read(i
));
6554 for (unsigned k
= 0; k
< inst
->components_read(i
); ++k
)
6555 cbld
.MOV(offset(tmp
, lbld
, k
), offset(src
, inst
->exec_size
, k
));
6559 } else if (is_periodic(inst
->src
[i
], lbld
.dispatch_width())) {
6560 /* The source is invariant for all dispatch_width-wide groups of the
6563 return inst
->src
[i
];
6566 /* We can just point the lowered instruction at the right channel group
6567 * from the original region.
6574 * Return true if splitting out the group of channels of instruction \p inst
6575 * given by lbld.group() requires allocating a temporary for the destination
6576 * of the lowered instruction and copying the data back to the original
6577 * destination region.
6580 needs_dst_copy(const fs_builder
&lbld
, const fs_inst
*inst
)
6582 /* If the instruction writes more than one component we'll have to shuffle
6583 * the results of multiple lowered instructions in order to make sure that
6584 * they end up arranged correctly in the original destination region.
6586 if (inst
->size_written
> inst
->dst
.component_size(inst
->exec_size
))
6589 /* If the lowered execution size is larger than the original the result of
6590 * the instruction won't fit in the original destination, so we'll have to
6591 * allocate a temporary in any case.
6593 if (lbld
.dispatch_width() > inst
->exec_size
)
6596 for (unsigned i
= 0; i
< inst
->sources
; i
++) {
6597 /* If we already made a copy of the source for other reasons there won't
6598 * be any overlap with the destination.
6600 if (needs_src_copy(lbld
, inst
, i
))
6603 /* In order to keep the logic simple we emit a copy whenever the
6604 * destination region doesn't exactly match an overlapping source, which
6605 * may point at the source and destination not being aligned group by
6606 * group which could cause one of the lowered instructions to overwrite
6607 * the data read from the same source by other lowered instructions.
6609 if (regions_overlap(inst
->dst
, inst
->size_written
,
6610 inst
->src
[i
], inst
->size_read(i
)) &&
6611 !inst
->dst
.equals(inst
->src
[i
]))
6619 * Insert data from a packed temporary into the channel group given by
6620 * lbld.group() of the destination region of instruction \p inst and return
6621 * the temporary as result. Any copy instructions that are required for
6622 * unzipping the previous value (in the case of partial writes) will be
6623 * inserted using \p lbld_before and any copy instructions required for
6624 * zipping up the destination of \p inst will be inserted using \p lbld_after.
6627 emit_zip(const fs_builder
&lbld_before
, const fs_builder
&lbld_after
,
6630 assert(lbld_before
.dispatch_width() == lbld_after
.dispatch_width());
6631 assert(lbld_before
.group() == lbld_after
.group());
6632 assert(lbld_after
.group() >= inst
->group
);
6634 /* Specified channel group from the destination region. */
6635 const fs_reg dst
= horiz_offset(inst
->dst
, lbld_after
.group() - inst
->group
);
6636 const unsigned dst_size
= inst
->size_written
/
6637 inst
->dst
.component_size(inst
->exec_size
);
6639 if (needs_dst_copy(lbld_after
, inst
)) {
6640 const fs_reg tmp
= lbld_after
.vgrf(inst
->dst
.type
, dst_size
);
6642 if (inst
->predicate
) {
6643 /* Handle predication by copying the original contents of
6644 * the destination into the temporary before emitting the
6645 * lowered instruction.
6647 const fs_builder gbld_before
=
6648 lbld_before
.group(MIN2(lbld_before
.dispatch_width(),
6649 inst
->exec_size
), 0);
6650 for (unsigned k
= 0; k
< dst_size
; ++k
) {
6651 gbld_before
.MOV(offset(tmp
, lbld_before
, k
),
6652 offset(dst
, inst
->exec_size
, k
));
6656 const fs_builder gbld_after
=
6657 lbld_after
.group(MIN2(lbld_after
.dispatch_width(),
6658 inst
->exec_size
), 0);
6659 for (unsigned k
= 0; k
< dst_size
; ++k
) {
6660 /* Use a builder of the right width to perform the copy avoiding
6661 * uninitialized data if the lowered execution size is greater than
6662 * the original execution size of the instruction.
6664 gbld_after
.MOV(offset(dst
, inst
->exec_size
, k
),
6665 offset(tmp
, lbld_after
, k
));
6671 /* No need to allocate a temporary for the lowered instruction, just
6672 * take the right group of channels from the original region.
6679 fs_visitor::lower_simd_width()
6681 bool progress
= false;
6683 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
6684 const unsigned lower_width
= get_lowered_simd_width(devinfo
, inst
);
6686 if (lower_width
!= inst
->exec_size
) {
6687 /* Builder matching the original instruction. We may also need to
6688 * emit an instruction of width larger than the original, set the
6689 * execution size of the builder to the highest of both for now so
6690 * we're sure that both cases can be handled.
6692 const unsigned max_width
= MAX2(inst
->exec_size
, lower_width
);
6693 const fs_builder ibld
= bld
.at(block
, inst
)
6694 .exec_all(inst
->force_writemask_all
)
6695 .group(max_width
, inst
->group
/ max_width
);
6697 /* Split the copies in chunks of the execution width of either the
6698 * original or the lowered instruction, whichever is lower.
6700 const unsigned n
= DIV_ROUND_UP(inst
->exec_size
, lower_width
);
6701 const unsigned dst_size
= inst
->size_written
/
6702 inst
->dst
.component_size(inst
->exec_size
);
6704 assert(!inst
->writes_accumulator
&& !inst
->mlen
);
6706 /* Inserting the zip, unzip, and duplicated instructions in all of
6707 * the right spots is somewhat tricky. All of the unzip and any
6708 * instructions from the zip which unzip the destination prior to
6709 * writing need to happen before all of the per-group instructions
6710 * and the zip instructions need to happen after. In order to sort
6711 * this all out, we insert the unzip instructions before \p inst,
6712 * insert the per-group instructions after \p inst (i.e. before
6713 * inst->next), and insert the zip instructions before the
6714 * instruction after \p inst. Since we are inserting instructions
6715 * after \p inst, inst->next is a moving target and we need to save
6716 * it off here so that we insert the zip instructions in the right
6719 * Since we're inserting split instructions after after_inst, the
6720 * instructions will end up in the reverse order that we insert them.
6721 * However, certain render target writes require that the low group
6722 * instructions come before the high group. From the Ivy Bridge PRM
6723 * Vol. 4, Pt. 1, Section 3.9.11:
6725 * "If multiple SIMD8 Dual Source messages are delivered by the
6726 * pixel shader thread, each SIMD8_DUALSRC_LO message must be
6727 * issued before the SIMD8_DUALSRC_HI message with the same Slot
6728 * Group Select setting."
6730 * And, from Section 3.9.11.1 of the same PRM:
6732 * "When SIMD32 or SIMD16 PS threads send render target writes
6733 * with multiple SIMD8 and SIMD16 messages, the following must
6736 * All the slots (as described above) must have a corresponding
6737 * render target write irrespective of the slot's validity. A slot
6738 * is considered valid when at least one sample is enabled. For
6739 * example, a SIMD16 PS thread must send two SIMD8 render target
6740 * writes to cover all the slots.
6742 * PS thread must send SIMD render target write messages with
6743 * increasing slot numbers. For example, SIMD16 thread has
6744 * Slot[15:0] and if two SIMD8 render target writes are used, the
6745 * first SIMD8 render target write must send Slot[7:0] and the
6746 * next one must send Slot[15:8]."
6748 * In order to make low group instructions come before high group
6749 * instructions (this is required for some render target writes), we
6750 * split from the highest group to lowest.
6752 exec_node
*const after_inst
= inst
->next
;
6753 for (int i
= n
- 1; i
>= 0; i
--) {
6754 /* Emit a copy of the original instruction with the lowered width.
6755 * If the EOT flag was set throw it away except for the last
6756 * instruction to avoid killing the thread prematurely.
6758 fs_inst split_inst
= *inst
;
6759 split_inst
.exec_size
= lower_width
;
6760 split_inst
.eot
= inst
->eot
&& i
== int(n
- 1);
6762 /* Select the correct channel enables for the i-th group, then
6763 * transform the sources and destination and emit the lowered
6766 const fs_builder lbld
= ibld
.group(lower_width
, i
);
6768 for (unsigned j
= 0; j
< inst
->sources
; j
++)
6769 split_inst
.src
[j
] = emit_unzip(lbld
.at(block
, inst
), inst
, j
);
6771 split_inst
.dst
= emit_zip(lbld
.at(block
, inst
),
6772 lbld
.at(block
, after_inst
), inst
);
6773 split_inst
.size_written
=
6774 split_inst
.dst
.component_size(lower_width
) * dst_size
;
6776 lbld
.at(block
, inst
->next
).emit(split_inst
);
6779 inst
->remove(block
);
6785 invalidate_live_intervals();
6791 fs_visitor::dump_instructions()
6793 dump_instructions(NULL
);
6797 fs_visitor::dump_instructions(const char *name
)
6799 FILE *file
= stderr
;
6800 if (name
&& geteuid() != 0) {
6801 file
= fopen(name
, "w");
6807 calculate_register_pressure();
6808 int ip
= 0, max_pressure
= 0;
6809 foreach_block_and_inst(block
, backend_instruction
, inst
, cfg
) {
6810 max_pressure
= MAX2(max_pressure
, regs_live_at_ip
[ip
]);
6811 fprintf(file
, "{%3d} %4d: ", regs_live_at_ip
[ip
], ip
);
6812 dump_instruction(inst
, file
);
6815 fprintf(file
, "Maximum %3d registers live at once.\n", max_pressure
);
6818 foreach_in_list(backend_instruction
, inst
, &instructions
) {
6819 fprintf(file
, "%4d: ", ip
++);
6820 dump_instruction(inst
, file
);
6824 if (file
!= stderr
) {
6830 fs_visitor::dump_instruction(backend_instruction
*be_inst
)
6832 dump_instruction(be_inst
, stderr
);
6836 fs_visitor::dump_instruction(backend_instruction
*be_inst
, FILE *file
)
6838 fs_inst
*inst
= (fs_inst
*)be_inst
;
6840 if (inst
->predicate
) {
6841 fprintf(file
, "(%cf%d.%d) ",
6842 inst
->predicate_inverse
? '-' : '+',
6843 inst
->flag_subreg
/ 2,
6844 inst
->flag_subreg
% 2);
6847 fprintf(file
, "%s", brw_instruction_name(devinfo
, inst
->opcode
));
6849 fprintf(file
, ".sat");
6850 if (inst
->conditional_mod
) {
6851 fprintf(file
, "%s", conditional_modifier
[inst
->conditional_mod
]);
6852 if (!inst
->predicate
&&
6853 (devinfo
->gen
< 5 || (inst
->opcode
!= BRW_OPCODE_SEL
&&
6854 inst
->opcode
!= BRW_OPCODE_CSEL
&&
6855 inst
->opcode
!= BRW_OPCODE_IF
&&
6856 inst
->opcode
!= BRW_OPCODE_WHILE
))) {
6857 fprintf(file
, ".f%d.%d", inst
->flag_subreg
/ 2,
6858 inst
->flag_subreg
% 2);
6861 fprintf(file
, "(%d) ", inst
->exec_size
);
6864 fprintf(file
, "(mlen: %d) ", inst
->mlen
);
6867 if (inst
->ex_mlen
) {
6868 fprintf(file
, "(ex_mlen: %d) ", inst
->ex_mlen
);
6872 fprintf(file
, "(EOT) ");
6875 switch (inst
->dst
.file
) {
6877 fprintf(file
, "vgrf%d", inst
->dst
.nr
);
6880 fprintf(file
, "g%d", inst
->dst
.nr
);
6883 fprintf(file
, "m%d", inst
->dst
.nr
);
6886 fprintf(file
, "(null)");
6889 fprintf(file
, "***u%d***", inst
->dst
.nr
);
6892 fprintf(file
, "***attr%d***", inst
->dst
.nr
);
6895 switch (inst
->dst
.nr
) {
6897 fprintf(file
, "null");
6899 case BRW_ARF_ADDRESS
:
6900 fprintf(file
, "a0.%d", inst
->dst
.subnr
);
6902 case BRW_ARF_ACCUMULATOR
:
6903 fprintf(file
, "acc%d", inst
->dst
.subnr
);
6906 fprintf(file
, "f%d.%d", inst
->dst
.nr
& 0xf, inst
->dst
.subnr
);
6909 fprintf(file
, "arf%d.%d", inst
->dst
.nr
& 0xf, inst
->dst
.subnr
);
6914 unreachable("not reached");
6917 if (inst
->dst
.offset
||
6918 (inst
->dst
.file
== VGRF
&&
6919 alloc
.sizes
[inst
->dst
.nr
] * REG_SIZE
!= inst
->size_written
)) {
6920 const unsigned reg_size
= (inst
->dst
.file
== UNIFORM
? 4 : REG_SIZE
);
6921 fprintf(file
, "+%d.%d", inst
->dst
.offset
/ reg_size
,
6922 inst
->dst
.offset
% reg_size
);
6925 if (inst
->dst
.stride
!= 1)
6926 fprintf(file
, "<%u>", inst
->dst
.stride
);
6927 fprintf(file
, ":%s, ", brw_reg_type_to_letters(inst
->dst
.type
));
6929 for (int i
= 0; i
< inst
->sources
; i
++) {
6930 if (inst
->src
[i
].negate
)
6932 if (inst
->src
[i
].abs
)
6934 switch (inst
->src
[i
].file
) {
6936 fprintf(file
, "vgrf%d", inst
->src
[i
].nr
);
6939 fprintf(file
, "g%d", inst
->src
[i
].nr
);
6942 fprintf(file
, "***m%d***", inst
->src
[i
].nr
);
6945 fprintf(file
, "attr%d", inst
->src
[i
].nr
);
6948 fprintf(file
, "u%d", inst
->src
[i
].nr
);
6951 fprintf(file
, "(null)");
6954 switch (inst
->src
[i
].type
) {
6955 case BRW_REGISTER_TYPE_F
:
6956 fprintf(file
, "%-gf", inst
->src
[i
].f
);
6958 case BRW_REGISTER_TYPE_DF
:
6959 fprintf(file
, "%fdf", inst
->src
[i
].df
);
6961 case BRW_REGISTER_TYPE_W
:
6962 case BRW_REGISTER_TYPE_D
:
6963 fprintf(file
, "%dd", inst
->src
[i
].d
);
6965 case BRW_REGISTER_TYPE_UW
:
6966 case BRW_REGISTER_TYPE_UD
:
6967 fprintf(file
, "%uu", inst
->src
[i
].ud
);
6969 case BRW_REGISTER_TYPE_Q
:
6970 fprintf(file
, "%" PRId64
"q", inst
->src
[i
].d64
);
6972 case BRW_REGISTER_TYPE_UQ
:
6973 fprintf(file
, "%" PRIu64
"uq", inst
->src
[i
].u64
);
6975 case BRW_REGISTER_TYPE_VF
:
6976 fprintf(file
, "[%-gF, %-gF, %-gF, %-gF]",
6977 brw_vf_to_float((inst
->src
[i
].ud
>> 0) & 0xff),
6978 brw_vf_to_float((inst
->src
[i
].ud
>> 8) & 0xff),
6979 brw_vf_to_float((inst
->src
[i
].ud
>> 16) & 0xff),
6980 brw_vf_to_float((inst
->src
[i
].ud
>> 24) & 0xff));
6982 case BRW_REGISTER_TYPE_V
:
6983 case BRW_REGISTER_TYPE_UV
:
6984 fprintf(file
, "%08x%s", inst
->src
[i
].ud
,
6985 inst
->src
[i
].type
== BRW_REGISTER_TYPE_V
? "V" : "UV");
6988 fprintf(file
, "???");
6993 switch (inst
->src
[i
].nr
) {
6995 fprintf(file
, "null");
6997 case BRW_ARF_ADDRESS
:
6998 fprintf(file
, "a0.%d", inst
->src
[i
].subnr
);
7000 case BRW_ARF_ACCUMULATOR
:
7001 fprintf(file
, "acc%d", inst
->src
[i
].subnr
);
7004 fprintf(file
, "f%d.%d", inst
->src
[i
].nr
& 0xf, inst
->src
[i
].subnr
);
7007 fprintf(file
, "arf%d.%d", inst
->src
[i
].nr
& 0xf, inst
->src
[i
].subnr
);
7013 if (inst
->src
[i
].offset
||
7014 (inst
->src
[i
].file
== VGRF
&&
7015 alloc
.sizes
[inst
->src
[i
].nr
] * REG_SIZE
!= inst
->size_read(i
))) {
7016 const unsigned reg_size
= (inst
->src
[i
].file
== UNIFORM
? 4 : REG_SIZE
);
7017 fprintf(file
, "+%d.%d", inst
->src
[i
].offset
/ reg_size
,
7018 inst
->src
[i
].offset
% reg_size
);
7021 if (inst
->src
[i
].abs
)
7024 if (inst
->src
[i
].file
!= IMM
) {
7026 if (inst
->src
[i
].file
== ARF
|| inst
->src
[i
].file
== FIXED_GRF
) {
7027 unsigned hstride
= inst
->src
[i
].hstride
;
7028 stride
= (hstride
== 0 ? 0 : (1 << (hstride
- 1)));
7030 stride
= inst
->src
[i
].stride
;
7033 fprintf(file
, "<%u>", stride
);
7035 fprintf(file
, ":%s", brw_reg_type_to_letters(inst
->src
[i
].type
));
7038 if (i
< inst
->sources
- 1 && inst
->src
[i
+ 1].file
!= BAD_FILE
)
7039 fprintf(file
, ", ");
7044 if (inst
->force_writemask_all
)
7045 fprintf(file
, "NoMask ");
7047 if (inst
->exec_size
!= dispatch_width
)
7048 fprintf(file
, "group%d ", inst
->group
);
7050 fprintf(file
, "\n");
7054 fs_visitor::setup_fs_payload_gen6()
7056 assert(stage
== MESA_SHADER_FRAGMENT
);
7057 struct brw_wm_prog_data
*prog_data
= brw_wm_prog_data(this->prog_data
);
7058 const unsigned payload_width
= MIN2(16, dispatch_width
);
7059 assert(dispatch_width
% payload_width
== 0);
7060 assert(devinfo
->gen
>= 6);
7062 prog_data
->uses_src_depth
= prog_data
->uses_src_w
=
7063 (nir
->info
.system_values_read
& (1ull << SYSTEM_VALUE_FRAG_COORD
)) != 0;
7065 prog_data
->uses_sample_mask
=
7066 (nir
->info
.system_values_read
& SYSTEM_BIT_SAMPLE_MASK_IN
) != 0;
7068 /* From the Ivy Bridge PRM documentation for 3DSTATE_PS:
7070 * "MSDISPMODE_PERSAMPLE is required in order to select
7073 * So we can only really get sample positions if we are doing real
7074 * per-sample dispatch. If we need gl_SamplePosition and we don't have
7075 * persample dispatch, we hard-code it to 0.5.
7077 prog_data
->uses_pos_offset
= prog_data
->persample_dispatch
&&
7078 (nir
->info
.system_values_read
& SYSTEM_BIT_SAMPLE_POS
);
7080 /* R0: PS thread payload header. */
7083 for (unsigned j
= 0; j
< dispatch_width
/ payload_width
; j
++) {
7084 /* R1: masks, pixel X/Y coordinates. */
7085 payload
.subspan_coord_reg
[j
] = payload
.num_regs
++;
7088 for (unsigned j
= 0; j
< dispatch_width
/ payload_width
; j
++) {
7089 /* R3-26: barycentric interpolation coordinates. These appear in the
7090 * same order that they appear in the brw_barycentric_mode enum. Each
7091 * set of coordinates occupies 2 registers if dispatch width == 8 and 4
7092 * registers if dispatch width == 16. Coordinates only appear if they
7093 * were enabled using the "Barycentric Interpolation Mode" bits in
7096 for (int i
= 0; i
< BRW_BARYCENTRIC_MODE_COUNT
; ++i
) {
7097 if (prog_data
->barycentric_interp_modes
& (1 << i
)) {
7098 payload
.barycentric_coord_reg
[i
][j
] = payload
.num_regs
;
7099 payload
.num_regs
+= payload_width
/ 4;
7103 /* R27-28: interpolated depth if uses source depth */
7104 if (prog_data
->uses_src_depth
) {
7105 payload
.source_depth_reg
[j
] = payload
.num_regs
;
7106 payload
.num_regs
+= payload_width
/ 8;
7109 /* R29-30: interpolated W set if GEN6_WM_USES_SOURCE_W. */
7110 if (prog_data
->uses_src_w
) {
7111 payload
.source_w_reg
[j
] = payload
.num_regs
;
7112 payload
.num_regs
+= payload_width
/ 8;
7115 /* R31: MSAA position offsets. */
7116 if (prog_data
->uses_pos_offset
) {
7117 payload
.sample_pos_reg
[j
] = payload
.num_regs
;
7121 /* R32-33: MSAA input coverage mask */
7122 if (prog_data
->uses_sample_mask
) {
7123 assert(devinfo
->gen
>= 7);
7124 payload
.sample_mask_in_reg
[j
] = payload
.num_regs
;
7125 payload
.num_regs
+= payload_width
/ 8;
7129 if (nir
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_DEPTH
)) {
7130 source_depth_to_render_target
= true;
7135 fs_visitor::setup_vs_payload()
7137 /* R0: thread header, R1: urb handles */
7138 payload
.num_regs
= 2;
7142 fs_visitor::setup_gs_payload()
7144 assert(stage
== MESA_SHADER_GEOMETRY
);
7146 struct brw_gs_prog_data
*gs_prog_data
= brw_gs_prog_data(prog_data
);
7147 struct brw_vue_prog_data
*vue_prog_data
= brw_vue_prog_data(prog_data
);
7149 /* R0: thread header, R1: output URB handles */
7150 payload
.num_regs
= 2;
7152 if (gs_prog_data
->include_primitive_id
) {
7153 /* R2: Primitive ID 0..7 */
7157 /* Always enable VUE handles so we can safely use pull model if needed.
7159 * The push model for a GS uses a ton of register space even for trivial
7160 * scenarios with just a few inputs, so just make things easier and a bit
7161 * safer by always having pull model available.
7163 gs_prog_data
->base
.include_vue_handles
= true;
7165 /* R3..RN: ICP Handles for each incoming vertex (when using pull model) */
7166 payload
.num_regs
+= nir
->info
.gs
.vertices_in
;
7168 /* Use a maximum of 24 registers for push-model inputs. */
7169 const unsigned max_push_components
= 24;
7171 /* If pushing our inputs would take too many registers, reduce the URB read
7172 * length (which is in HWords, or 8 registers), and resort to pulling.
7174 * Note that the GS reads <URB Read Length> HWords for every vertex - so we
7175 * have to multiply by VerticesIn to obtain the total storage requirement.
7177 if (8 * vue_prog_data
->urb_read_length
* nir
->info
.gs
.vertices_in
>
7178 max_push_components
) {
7179 vue_prog_data
->urb_read_length
=
7180 ROUND_DOWN_TO(max_push_components
/ nir
->info
.gs
.vertices_in
, 8) / 8;
7185 fs_visitor::setup_cs_payload()
7187 assert(devinfo
->gen
>= 7);
7188 payload
.num_regs
= 1;
7192 fs_visitor::calculate_register_pressure()
7194 invalidate_live_intervals();
7195 calculate_live_intervals();
7197 unsigned num_instructions
= 0;
7198 foreach_block(block
, cfg
)
7199 num_instructions
+= block
->instructions
.length();
7201 regs_live_at_ip
= rzalloc_array(mem_ctx
, int, num_instructions
);
7203 for (unsigned reg
= 0; reg
< alloc
.count
; reg
++) {
7204 for (int ip
= virtual_grf_start
[reg
]; ip
<= virtual_grf_end
[reg
]; ip
++)
7205 regs_live_at_ip
[ip
] += alloc
.sizes
[reg
];
7210 fs_visitor::optimize()
7212 /* Start by validating the shader we currently have. */
7215 /* bld is the common builder object pointing at the end of the program we
7216 * used to translate it into i965 IR. For the optimization and lowering
7217 * passes coming next, any code added after the end of the program without
7218 * having explicitly called fs_builder::at() clearly points at a mistake.
7219 * Ideally optimization passes wouldn't be part of the visitor so they
7220 * wouldn't have access to bld at all, but they do, so just in case some
7221 * pass forgets to ask for a location explicitly set it to NULL here to
7222 * make it trip. The dispatch width is initialized to a bogus value to
7223 * make sure that optimizations set the execution controls explicitly to
7224 * match the code they are manipulating instead of relying on the defaults.
7226 bld
= fs_builder(this, 64);
7228 assign_constant_locations();
7229 lower_constant_loads();
7233 split_virtual_grfs();
7236 #define OPT(pass, args...) ({ \
7238 bool this_progress = pass(args); \
7240 if (unlikely(INTEL_DEBUG & DEBUG_OPTIMIZER) && this_progress) { \
7241 char filename[64]; \
7242 snprintf(filename, 64, "%s%d-%s-%02d-%02d-" #pass, \
7243 stage_abbrev, dispatch_width, nir->info.name, iteration, pass_num); \
7245 backend_shader::dump_instructions(filename); \
7250 progress = progress || this_progress; \
7254 if (unlikely(INTEL_DEBUG
& DEBUG_OPTIMIZER
)) {
7256 snprintf(filename
, 64, "%s%d-%s-00-00-start",
7257 stage_abbrev
, dispatch_width
, nir
->info
.name
);
7259 backend_shader::dump_instructions(filename
);
7262 bool progress
= false;
7266 /* Before anything else, eliminate dead code. The results of some NIR
7267 * instructions may effectively be calculated twice. Once when the
7268 * instruction is encountered, and again when the user of that result is
7269 * encountered. Wipe those away before algebraic optimizations and
7270 * especially copy propagation can mix things up.
7272 OPT(dead_code_eliminate
);
7274 OPT(remove_extra_rounding_modes
);
7281 OPT(remove_duplicate_mrf_writes
);
7285 OPT(opt_copy_propagation
);
7286 OPT(opt_predicated_break
, this);
7287 OPT(opt_cmod_propagation
);
7288 OPT(dead_code_eliminate
);
7289 OPT(opt_peephole_sel
);
7290 OPT(dead_control_flow_eliminate
, this);
7291 OPT(opt_register_renaming
);
7292 OPT(opt_saturate_propagation
);
7293 OPT(register_coalesce
);
7294 OPT(compute_to_mrf
);
7295 OPT(eliminate_find_live_channel
);
7297 OPT(compact_virtual_grfs
);
7300 /* Do this after cmod propagation has had every possible opportunity to
7301 * propagate results into SEL instructions.
7303 if (OPT(opt_peephole_csel
))
7304 OPT(dead_code_eliminate
);
7309 if (OPT(lower_pack
)) {
7310 OPT(register_coalesce
);
7311 OPT(dead_code_eliminate
);
7314 OPT(lower_simd_width
);
7316 /* After SIMD lowering just in case we had to unroll the EOT send. */
7317 OPT(opt_sampler_eot
);
7319 OPT(lower_logical_sends
);
7322 OPT(opt_copy_propagation
);
7323 /* Only run after logical send lowering because it's easier to implement
7324 * in terms of physical sends.
7326 if (OPT(opt_zero_samples
))
7327 OPT(opt_copy_propagation
);
7328 /* Run after logical send lowering to give it a chance to CSE the
7329 * LOAD_PAYLOAD instructions created to construct the payloads of
7330 * e.g. texturing messages in cases where it wasn't possible to CSE the
7331 * whole logical instruction.
7334 OPT(register_coalesce
);
7335 OPT(compute_to_mrf
);
7336 OPT(dead_code_eliminate
);
7337 OPT(remove_duplicate_mrf_writes
);
7338 OPT(opt_peephole_sel
);
7341 OPT(opt_redundant_discard_jumps
);
7343 if (OPT(lower_load_payload
)) {
7344 split_virtual_grfs();
7345 OPT(register_coalesce
);
7346 OPT(lower_simd_width
);
7347 OPT(compute_to_mrf
);
7348 OPT(dead_code_eliminate
);
7351 OPT(opt_combine_constants
);
7352 OPT(lower_integer_multiplication
);
7354 if (devinfo
->gen
<= 5 && OPT(lower_minmax
)) {
7355 OPT(opt_cmod_propagation
);
7357 OPT(opt_copy_propagation
);
7358 OPT(dead_code_eliminate
);
7361 if (OPT(lower_regioning
)) {
7362 OPT(opt_copy_propagation
);
7363 OPT(dead_code_eliminate
);
7364 OPT(lower_simd_width
);
7367 OPT(fixup_sends_duplicate_payload
);
7369 lower_uniform_pull_constant_loads();
7375 * From the Skylake PRM Vol. 2a docs for sends:
7377 * "It is required that the second block of GRFs does not overlap with the
7380 * There are plenty of cases where we may accidentally violate this due to
7381 * having, for instance, both sources be the constant 0. This little pass
7382 * just adds a new vgrf for the second payload and copies it over.
7385 fs_visitor::fixup_sends_duplicate_payload()
7387 bool progress
= false;
7389 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
7390 if (inst
->opcode
== SHADER_OPCODE_SEND
&& inst
->ex_mlen
> 0 &&
7391 regions_overlap(inst
->src
[2], inst
->mlen
* REG_SIZE
,
7392 inst
->src
[3], inst
->ex_mlen
* REG_SIZE
)) {
7393 fs_reg tmp
= fs_reg(VGRF
, alloc
.allocate(inst
->ex_mlen
),
7394 BRW_REGISTER_TYPE_UD
);
7395 /* Sadly, we've lost all notion of channels and bit sizes at this
7396 * point. Just WE_all it.
7398 const fs_builder ibld
= bld
.at(block
, inst
).exec_all().group(16, 0);
7399 fs_reg copy_src
= retype(inst
->src
[3], BRW_REGISTER_TYPE_UD
);
7400 fs_reg copy_dst
= tmp
;
7401 for (unsigned i
= 0; i
< inst
->ex_mlen
; i
+= 2) {
7402 if (inst
->ex_mlen
== i
+ 1) {
7403 /* Only one register left; do SIMD8 */
7404 ibld
.group(8, 0).MOV(copy_dst
, copy_src
);
7406 ibld
.MOV(copy_dst
, copy_src
);
7408 copy_src
= offset(copy_src
, ibld
, 1);
7409 copy_dst
= offset(copy_dst
, ibld
, 1);
7417 invalidate_live_intervals();
7423 * Three source instruction must have a GRF/MRF destination register.
7424 * ARF NULL is not allowed. Fix that up by allocating a temporary GRF.
7427 fs_visitor::fixup_3src_null_dest()
7429 bool progress
= false;
7431 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
7432 if (inst
->is_3src(devinfo
) && inst
->dst
.is_null()) {
7433 inst
->dst
= fs_reg(VGRF
, alloc
.allocate(dispatch_width
/ 8),
7440 invalidate_live_intervals();
7444 fs_visitor::allocate_registers(unsigned min_dispatch_width
, bool allow_spilling
)
7448 static const enum instruction_scheduler_mode pre_modes
[] = {
7450 SCHEDULE_PRE_NON_LIFO
,
7454 static const char *scheduler_mode_name
[] = {
7460 bool spill_all
= allow_spilling
&& (INTEL_DEBUG
& DEBUG_SPILL_FS
);
7462 /* Try each scheduling heuristic to see if it can successfully register
7463 * allocate without spilling. They should be ordered by decreasing
7464 * performance but increasing likelihood of allocating.
7466 for (unsigned i
= 0; i
< ARRAY_SIZE(pre_modes
); i
++) {
7467 schedule_instructions(pre_modes
[i
]);
7468 this->shader_stats
.scheduler_mode
= scheduler_mode_name
[i
];
7471 assign_regs_trivial();
7476 /* We only allow spilling for the last schedule mode and only if the
7477 * allow_spilling parameter and dispatch width work out ok.
7479 bool can_spill
= allow_spilling
&&
7480 (i
== ARRAY_SIZE(pre_modes
) - 1) &&
7481 dispatch_width
== min_dispatch_width
;
7483 /* We should only spill registers on the last scheduling. */
7484 assert(!spilled_any_registers
);
7486 allocated
= assign_regs(can_spill
, spill_all
);
7492 if (!allow_spilling
)
7493 fail("Failure to register allocate and spilling is not allowed.");
7495 /* We assume that any spilling is worse than just dropping back to
7496 * SIMD8. There's probably actually some intermediate point where
7497 * SIMD16 with a couple of spills is still better.
7499 if (dispatch_width
> min_dispatch_width
) {
7500 fail("Failure to register allocate. Reduce number of "
7501 "live scalar values to avoid this.");
7504 /* If we failed to allocate, we must have a reason */
7506 } else if (spilled_any_registers
) {
7507 compiler
->shader_perf_log(log_data
,
7508 "%s shader triggered register spilling. "
7509 "Try reducing the number of live scalar "
7510 "values to improve performance.\n",
7514 /* This must come after all optimization and register allocation, since
7515 * it inserts dead code that happens to have side effects, and it does
7516 * so based on the actual physical registers in use.
7518 insert_gen4_send_dependency_workarounds();
7523 opt_bank_conflicts();
7525 schedule_instructions(SCHEDULE_POST
);
7527 if (last_scratch
> 0) {
7528 ASSERTED
unsigned max_scratch_size
= 2 * 1024 * 1024;
7530 prog_data
->total_scratch
= brw_get_scratch_size(last_scratch
);
7532 if (stage
== MESA_SHADER_COMPUTE
) {
7533 if (devinfo
->is_haswell
) {
7534 /* According to the MEDIA_VFE_STATE's "Per Thread Scratch Space"
7535 * field documentation, Haswell supports a minimum of 2kB of
7536 * scratch space for compute shaders, unlike every other stage
7539 prog_data
->total_scratch
= MAX2(prog_data
->total_scratch
, 2048);
7540 } else if (devinfo
->gen
<= 7) {
7541 /* According to the MEDIA_VFE_STATE's "Per Thread Scratch Space"
7542 * field documentation, platforms prior to Haswell measure scratch
7543 * size linearly with a range of [1kB, 12kB] and 1kB granularity.
7545 prog_data
->total_scratch
= ALIGN(last_scratch
, 1024);
7546 max_scratch_size
= 12 * 1024;
7550 /* We currently only support up to 2MB of scratch space. If we
7551 * need to support more eventually, the documentation suggests
7552 * that we could allocate a larger buffer, and partition it out
7553 * ourselves. We'd just have to undo the hardware's address
7554 * calculation by subtracting (FFTID * Per Thread Scratch Space)
7555 * and then add FFTID * (Larger Per Thread Scratch Space).
7557 * See 3D-Media-GPGPU Engine > Media GPGPU Pipeline >
7558 * Thread Group Tracking > Local Memory/Scratch Space.
7560 assert(prog_data
->total_scratch
< max_scratch_size
);
7567 fs_visitor::run_vs()
7569 assert(stage
== MESA_SHADER_VERTEX
);
7573 if (shader_time_index
>= 0)
7574 emit_shader_time_begin();
7583 if (shader_time_index
>= 0)
7584 emit_shader_time_end();
7590 assign_curb_setup();
7591 assign_vs_urb_setup();
7593 fixup_3src_null_dest();
7594 allocate_registers(8, true);
7600 fs_visitor::set_tcs_invocation_id()
7602 struct brw_tcs_prog_data
*tcs_prog_data
= brw_tcs_prog_data(prog_data
);
7603 struct brw_vue_prog_data
*vue_prog_data
= &tcs_prog_data
->base
;
7605 const unsigned instance_id_mask
=
7606 devinfo
->gen
>= 11 ? INTEL_MASK(22, 16) : INTEL_MASK(23, 17);
7607 const unsigned instance_id_shift
=
7608 devinfo
->gen
>= 11 ? 16 : 17;
7610 /* Get instance number from g0.2 bits 22:16 or 23:17 */
7611 fs_reg t
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
7612 bld
.AND(t
, fs_reg(retype(brw_vec1_grf(0, 2), BRW_REGISTER_TYPE_UD
)),
7613 brw_imm_ud(instance_id_mask
));
7615 invocation_id
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
7617 if (vue_prog_data
->dispatch_mode
== DISPATCH_MODE_TCS_8_PATCH
) {
7618 /* gl_InvocationID is just the thread number */
7619 bld
.SHR(invocation_id
, t
, brw_imm_ud(instance_id_shift
));
7623 assert(vue_prog_data
->dispatch_mode
== DISPATCH_MODE_TCS_SINGLE_PATCH
);
7625 fs_reg channels_uw
= bld
.vgrf(BRW_REGISTER_TYPE_UW
);
7626 fs_reg channels_ud
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
7627 bld
.MOV(channels_uw
, fs_reg(brw_imm_uv(0x76543210)));
7628 bld
.MOV(channels_ud
, channels_uw
);
7630 if (tcs_prog_data
->instances
== 1) {
7631 invocation_id
= channels_ud
;
7633 fs_reg instance_times_8
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
7634 bld
.SHR(instance_times_8
, t
, brw_imm_ud(instance_id_shift
- 3));
7635 bld
.ADD(invocation_id
, instance_times_8
, channels_ud
);
7640 fs_visitor::run_tcs()
7642 assert(stage
== MESA_SHADER_TESS_CTRL
);
7644 struct brw_vue_prog_data
*vue_prog_data
= brw_vue_prog_data(prog_data
);
7645 struct brw_tcs_prog_data
*tcs_prog_data
= brw_tcs_prog_data(prog_data
);
7646 struct brw_tcs_prog_key
*tcs_key
= (struct brw_tcs_prog_key
*) key
;
7648 assert(vue_prog_data
->dispatch_mode
== DISPATCH_MODE_TCS_SINGLE_PATCH
||
7649 vue_prog_data
->dispatch_mode
== DISPATCH_MODE_TCS_8_PATCH
);
7651 if (vue_prog_data
->dispatch_mode
== DISPATCH_MODE_TCS_SINGLE_PATCH
) {
7652 /* r1-r4 contain the ICP handles. */
7653 payload
.num_regs
= 5;
7655 assert(vue_prog_data
->dispatch_mode
== DISPATCH_MODE_TCS_8_PATCH
);
7656 assert(tcs_key
->input_vertices
> 0);
7657 /* r1 contains output handles, r2 may contain primitive ID, then the
7658 * ICP handles occupy the next 1-32 registers.
7660 payload
.num_regs
= 2 + tcs_prog_data
->include_primitive_id
+
7661 tcs_key
->input_vertices
;
7664 if (shader_time_index
>= 0)
7665 emit_shader_time_begin();
7667 /* Initialize gl_InvocationID */
7668 set_tcs_invocation_id();
7670 const bool fix_dispatch_mask
=
7671 vue_prog_data
->dispatch_mode
== DISPATCH_MODE_TCS_SINGLE_PATCH
&&
7672 (nir
->info
.tess
.tcs_vertices_out
% 8) != 0;
7674 /* Fix the disptach mask */
7675 if (fix_dispatch_mask
) {
7676 bld
.CMP(bld
.null_reg_ud(), invocation_id
,
7677 brw_imm_ud(nir
->info
.tess
.tcs_vertices_out
), BRW_CONDITIONAL_L
);
7678 bld
.IF(BRW_PREDICATE_NORMAL
);
7683 if (fix_dispatch_mask
) {
7684 bld
.emit(BRW_OPCODE_ENDIF
);
7687 /* Emit EOT write; set TR DS Cache bit */
7689 fs_reg(get_tcs_output_urb_handle()),
7690 fs_reg(brw_imm_ud(WRITEMASK_X
<< 16)),
7691 fs_reg(brw_imm_ud(0)),
7693 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 3);
7694 bld
.LOAD_PAYLOAD(payload
, srcs
, 3, 2);
7696 fs_inst
*inst
= bld
.emit(SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
,
7697 bld
.null_reg_ud(), payload
);
7701 if (shader_time_index
>= 0)
7702 emit_shader_time_end();
7711 assign_curb_setup();
7712 assign_tcs_urb_setup();
7714 fixup_3src_null_dest();
7715 allocate_registers(8, true);
7721 fs_visitor::run_tes()
7723 assert(stage
== MESA_SHADER_TESS_EVAL
);
7725 /* R0: thread header, R1-3: gl_TessCoord.xyz, R4: URB handles */
7726 payload
.num_regs
= 5;
7728 if (shader_time_index
>= 0)
7729 emit_shader_time_begin();
7738 if (shader_time_index
>= 0)
7739 emit_shader_time_end();
7745 assign_curb_setup();
7746 assign_tes_urb_setup();
7748 fixup_3src_null_dest();
7749 allocate_registers(8, true);
7755 fs_visitor::run_gs()
7757 assert(stage
== MESA_SHADER_GEOMETRY
);
7761 this->final_gs_vertex_count
= vgrf(glsl_type::uint_type
);
7763 if (gs_compile
->control_data_header_size_bits
> 0) {
7764 /* Create a VGRF to store accumulated control data bits. */
7765 this->control_data_bits
= vgrf(glsl_type::uint_type
);
7767 /* If we're outputting more than 32 control data bits, then EmitVertex()
7768 * will set control_data_bits to 0 after emitting the first vertex.
7769 * Otherwise, we need to initialize it to 0 here.
7771 if (gs_compile
->control_data_header_size_bits
<= 32) {
7772 const fs_builder abld
= bld
.annotate("initialize control data bits");
7773 abld
.MOV(this->control_data_bits
, brw_imm_ud(0u));
7777 if (shader_time_index
>= 0)
7778 emit_shader_time_begin();
7782 emit_gs_thread_end();
7784 if (shader_time_index
>= 0)
7785 emit_shader_time_end();
7794 assign_curb_setup();
7795 assign_gs_urb_setup();
7797 fixup_3src_null_dest();
7798 allocate_registers(8, true);
7803 /* From the SKL PRM, Volume 16, Workarounds:
7805 * 0877 3D Pixel Shader Hang possible when pixel shader dispatched with
7806 * only header phases (R0-R2)
7808 * WA: Enable a non-header phase (e.g. push constant) when dispatch would
7809 * have been header only.
7811 * Instead of enabling push constants one can alternatively enable one of the
7812 * inputs. Here one simply chooses "layer" which shouldn't impose much
7816 gen9_ps_header_only_workaround(struct brw_wm_prog_data
*wm_prog_data
)
7818 if (wm_prog_data
->num_varying_inputs
)
7821 if (wm_prog_data
->base
.curb_read_length
)
7824 wm_prog_data
->urb_setup
[VARYING_SLOT_LAYER
] = 0;
7825 wm_prog_data
->num_varying_inputs
= 1;
7829 fs_visitor::run_fs(bool allow_spilling
, bool do_rep_send
)
7831 struct brw_wm_prog_data
*wm_prog_data
= brw_wm_prog_data(this->prog_data
);
7832 brw_wm_prog_key
*wm_key
= (brw_wm_prog_key
*) this->key
;
7834 assert(stage
== MESA_SHADER_FRAGMENT
);
7836 if (devinfo
->gen
>= 6)
7837 setup_fs_payload_gen6();
7839 setup_fs_payload_gen4();
7843 } else if (do_rep_send
) {
7844 assert(dispatch_width
== 16);
7845 emit_repclear_shader();
7847 if (shader_time_index
>= 0)
7848 emit_shader_time_begin();
7850 if (nir
->info
.inputs_read
> 0 ||
7851 (nir
->info
.system_values_read
& (1ull << SYSTEM_VALUE_FRAG_COORD
)) ||
7852 (nir
->info
.outputs_read
> 0 && !wm_key
->coherent_fb_fetch
)) {
7853 if (devinfo
->gen
< 6)
7854 emit_interpolation_setup_gen4();
7856 emit_interpolation_setup_gen6();
7859 /* We handle discards by keeping track of the still-live pixels in f0.1.
7860 * Initialize it with the dispatched pixels.
7862 if (wm_prog_data
->uses_kill
) {
7863 const fs_reg dispatch_mask
=
7864 devinfo
->gen
>= 6 ? brw_vec1_grf(1, 7) : brw_vec1_grf(0, 0);
7865 bld
.exec_all().group(1, 0)
7866 .MOV(retype(brw_flag_reg(0, 1), BRW_REGISTER_TYPE_UW
),
7867 retype(dispatch_mask
, BRW_REGISTER_TYPE_UW
));
7875 if (wm_prog_data
->uses_kill
)
7876 bld
.emit(FS_OPCODE_PLACEHOLDER_HALT
);
7878 if (wm_key
->alpha_test_func
)
7883 if (shader_time_index
>= 0)
7884 emit_shader_time_end();
7890 assign_curb_setup();
7892 if (devinfo
->gen
>= 9)
7893 gen9_ps_header_only_workaround(wm_prog_data
);
7897 fixup_3src_null_dest();
7898 allocate_registers(8, allow_spilling
);
7908 fs_visitor::run_cs(unsigned min_dispatch_width
)
7910 assert(stage
== MESA_SHADER_COMPUTE
);
7911 assert(dispatch_width
>= min_dispatch_width
);
7915 if (shader_time_index
>= 0)
7916 emit_shader_time_begin();
7918 if (devinfo
->is_haswell
&& prog_data
->total_shared
> 0) {
7919 /* Move SLM index from g0.0[27:24] to sr0.1[11:8] */
7920 const fs_builder abld
= bld
.exec_all().group(1, 0);
7921 abld
.MOV(retype(brw_sr0_reg(1), BRW_REGISTER_TYPE_UW
),
7922 suboffset(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW
), 1));
7930 emit_cs_terminate();
7932 if (shader_time_index
>= 0)
7933 emit_shader_time_end();
7939 assign_curb_setup();
7941 fixup_3src_null_dest();
7942 allocate_registers(min_dispatch_width
, true);
7951 is_used_in_not_interp_frag_coord(nir_ssa_def
*def
)
7953 nir_foreach_use(src
, def
) {
7954 if (src
->parent_instr
->type
!= nir_instr_type_intrinsic
)
7957 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(src
->parent_instr
);
7958 if (intrin
->intrinsic
!= nir_intrinsic_load_frag_coord
)
7962 nir_foreach_if_use(src
, def
)
7969 * Return a bitfield where bit n is set if barycentric interpolation mode n
7970 * (see enum brw_barycentric_mode) is needed by the fragment shader.
7972 * We examine the load_barycentric intrinsics rather than looking at input
7973 * variables so that we catch interpolateAtCentroid() messages too, which
7974 * also need the BRW_BARYCENTRIC_[NON]PERSPECTIVE_CENTROID mode set up.
7977 brw_compute_barycentric_interp_modes(const struct gen_device_info
*devinfo
,
7978 const nir_shader
*shader
)
7980 unsigned barycentric_interp_modes
= 0;
7982 nir_foreach_function(f
, shader
) {
7986 nir_foreach_block(block
, f
->impl
) {
7987 nir_foreach_instr(instr
, block
) {
7988 if (instr
->type
!= nir_instr_type_intrinsic
)
7991 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
7992 switch (intrin
->intrinsic
) {
7993 case nir_intrinsic_load_barycentric_pixel
:
7994 case nir_intrinsic_load_barycentric_centroid
:
7995 case nir_intrinsic_load_barycentric_sample
:
8001 /* Ignore WPOS; it doesn't require interpolation. */
8002 assert(intrin
->dest
.is_ssa
);
8003 if (!is_used_in_not_interp_frag_coord(&intrin
->dest
.ssa
))
8006 enum glsl_interp_mode interp
= (enum glsl_interp_mode
)
8007 nir_intrinsic_interp_mode(intrin
);
8008 nir_intrinsic_op bary_op
= intrin
->intrinsic
;
8009 enum brw_barycentric_mode bary
=
8010 brw_barycentric_mode(interp
, bary_op
);
8012 barycentric_interp_modes
|= 1 << bary
;
8014 if (devinfo
->needs_unlit_centroid_workaround
&&
8015 bary_op
== nir_intrinsic_load_barycentric_centroid
)
8016 barycentric_interp_modes
|= 1 << centroid_to_pixel(bary
);
8021 return barycentric_interp_modes
;
8025 brw_compute_flat_inputs(struct brw_wm_prog_data
*prog_data
,
8026 const nir_shader
*shader
)
8028 prog_data
->flat_inputs
= 0;
8030 nir_foreach_variable(var
, &shader
->inputs
) {
8031 unsigned slots
= glsl_count_attribute_slots(var
->type
, false);
8032 for (unsigned s
= 0; s
< slots
; s
++) {
8033 int input_index
= prog_data
->urb_setup
[var
->data
.location
+ s
];
8035 if (input_index
< 0)
8039 if (var
->data
.interpolation
== INTERP_MODE_FLAT
)
8040 prog_data
->flat_inputs
|= 1 << input_index
;
8046 computed_depth_mode(const nir_shader
*shader
)
8048 if (shader
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_DEPTH
)) {
8049 switch (shader
->info
.fs
.depth_layout
) {
8050 case FRAG_DEPTH_LAYOUT_NONE
:
8051 case FRAG_DEPTH_LAYOUT_ANY
:
8052 return BRW_PSCDEPTH_ON
;
8053 case FRAG_DEPTH_LAYOUT_GREATER
:
8054 return BRW_PSCDEPTH_ON_GE
;
8055 case FRAG_DEPTH_LAYOUT_LESS
:
8056 return BRW_PSCDEPTH_ON_LE
;
8057 case FRAG_DEPTH_LAYOUT_UNCHANGED
:
8058 return BRW_PSCDEPTH_OFF
;
8061 return BRW_PSCDEPTH_OFF
;
8065 * Move load_interpolated_input with simple (payload-based) barycentric modes
8066 * to the top of the program so we don't emit multiple PLNs for the same input.
8068 * This works around CSE not being able to handle non-dominating cases
8074 * interpolate the same exact input
8077 * This should be replaced by global value numbering someday.
8080 move_interpolation_to_top(nir_shader
*nir
)
8082 bool progress
= false;
8084 nir_foreach_function(f
, nir
) {
8088 nir_block
*top
= nir_start_block(f
->impl
);
8089 exec_node
*cursor_node
= NULL
;
8091 nir_foreach_block(block
, f
->impl
) {
8095 nir_foreach_instr_safe(instr
, block
) {
8096 if (instr
->type
!= nir_instr_type_intrinsic
)
8099 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
8100 if (intrin
->intrinsic
!= nir_intrinsic_load_interpolated_input
)
8102 nir_intrinsic_instr
*bary_intrinsic
=
8103 nir_instr_as_intrinsic(intrin
->src
[0].ssa
->parent_instr
);
8104 nir_intrinsic_op op
= bary_intrinsic
->intrinsic
;
8106 /* Leave interpolateAtSample/Offset() where they are. */
8107 if (op
== nir_intrinsic_load_barycentric_at_sample
||
8108 op
== nir_intrinsic_load_barycentric_at_offset
)
8111 nir_instr
*move
[3] = {
8112 &bary_intrinsic
->instr
,
8113 intrin
->src
[1].ssa
->parent_instr
,
8117 for (unsigned i
= 0; i
< ARRAY_SIZE(move
); i
++) {
8118 if (move
[i
]->block
!= top
) {
8119 move
[i
]->block
= top
;
8120 exec_node_remove(&move
[i
]->node
);
8122 exec_node_insert_after(cursor_node
, &move
[i
]->node
);
8124 exec_list_push_head(&top
->instr_list
, &move
[i
]->node
);
8126 cursor_node
= &move
[i
]->node
;
8132 nir_metadata_preserve(f
->impl
, (nir_metadata
)
8133 ((unsigned) nir_metadata_block_index
|
8134 (unsigned) nir_metadata_dominance
));
8141 * Demote per-sample barycentric intrinsics to centroid.
8143 * Useful when rendering to a non-multisampled buffer.
8146 demote_sample_qualifiers(nir_shader
*nir
)
8148 bool progress
= true;
8150 nir_foreach_function(f
, nir
) {
8155 nir_builder_init(&b
, f
->impl
);
8157 nir_foreach_block(block
, f
->impl
) {
8158 nir_foreach_instr_safe(instr
, block
) {
8159 if (instr
->type
!= nir_instr_type_intrinsic
)
8162 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
8163 if (intrin
->intrinsic
!= nir_intrinsic_load_barycentric_sample
&&
8164 intrin
->intrinsic
!= nir_intrinsic_load_barycentric_at_sample
)
8167 b
.cursor
= nir_before_instr(instr
);
8168 nir_ssa_def
*centroid
=
8169 nir_load_barycentric(&b
, nir_intrinsic_load_barycentric_centroid
,
8170 nir_intrinsic_interp_mode(intrin
));
8171 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
8172 nir_src_for_ssa(centroid
));
8173 nir_instr_remove(instr
);
8178 nir_metadata_preserve(f
->impl
, (nir_metadata
)
8179 ((unsigned) nir_metadata_block_index
|
8180 (unsigned) nir_metadata_dominance
));
8187 * Pre-gen6, the register file of the EUs was shared between threads,
8188 * and each thread used some subset allocated on a 16-register block
8189 * granularity. The unit states wanted these block counts.
8192 brw_register_blocks(int reg_count
)
8194 return ALIGN(reg_count
, 16) / 16 - 1;
8198 brw_compile_fs(const struct brw_compiler
*compiler
, void *log_data
,
8200 const struct brw_wm_prog_key
*key
,
8201 struct brw_wm_prog_data
*prog_data
,
8203 int shader_time_index8
, int shader_time_index16
,
8204 int shader_time_index32
, bool allow_spilling
,
8205 bool use_rep_send
, struct brw_vue_map
*vue_map
,
8206 struct brw_compile_stats
*stats
,
8209 const struct gen_device_info
*devinfo
= compiler
->devinfo
;
8211 unsigned max_subgroup_size
= unlikely(INTEL_DEBUG
& DEBUG_DO32
) ? 32 : 16;
8213 brw_nir_apply_key(shader
, compiler
, &key
->base
, max_subgroup_size
, true);
8214 brw_nir_lower_fs_inputs(shader
, devinfo
, key
);
8215 brw_nir_lower_fs_outputs(shader
);
8217 if (devinfo
->gen
< 6)
8218 brw_setup_vue_interpolation(vue_map
, shader
, prog_data
);
8220 if (!key
->multisample_fbo
)
8221 NIR_PASS_V(shader
, demote_sample_qualifiers
);
8222 NIR_PASS_V(shader
, move_interpolation_to_top
);
8223 brw_postprocess_nir(shader
, compiler
, true);
8225 /* key->alpha_test_func means simulating alpha testing via discards,
8226 * so the shader definitely kills pixels.
8228 prog_data
->uses_kill
= shader
->info
.fs
.uses_discard
||
8229 key
->alpha_test_func
;
8230 prog_data
->uses_omask
= key
->multisample_fbo
&&
8231 shader
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_SAMPLE_MASK
);
8232 prog_data
->computed_depth_mode
= computed_depth_mode(shader
);
8233 prog_data
->computed_stencil
=
8234 shader
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_STENCIL
);
8236 prog_data
->persample_dispatch
=
8237 key
->multisample_fbo
&&
8238 (key
->persample_interp
||
8239 (shader
->info
.system_values_read
& (SYSTEM_BIT_SAMPLE_ID
|
8240 SYSTEM_BIT_SAMPLE_POS
)) ||
8241 shader
->info
.fs
.uses_sample_qualifier
||
8242 shader
->info
.outputs_read
);
8244 prog_data
->has_render_target_reads
= shader
->info
.outputs_read
!= 0ull;
8246 prog_data
->early_fragment_tests
= shader
->info
.fs
.early_fragment_tests
;
8247 prog_data
->post_depth_coverage
= shader
->info
.fs
.post_depth_coverage
;
8248 prog_data
->inner_coverage
= shader
->info
.fs
.inner_coverage
;
8250 prog_data
->barycentric_interp_modes
=
8251 brw_compute_barycentric_interp_modes(compiler
->devinfo
, shader
);
8253 calculate_urb_setup(devinfo
, key
, prog_data
, shader
);
8254 brw_compute_flat_inputs(prog_data
, shader
);
8256 cfg_t
*simd8_cfg
= NULL
, *simd16_cfg
= NULL
, *simd32_cfg
= NULL
;
8258 fs_visitor
v8(compiler
, log_data
, mem_ctx
, &key
->base
,
8259 &prog_data
->base
, shader
, 8,
8260 shader_time_index8
);
8261 if (!v8
.run_fs(allow_spilling
, false /* do_rep_send */)) {
8263 *error_str
= ralloc_strdup(mem_ctx
, v8
.fail_msg
);
8266 } else if (likely(!(INTEL_DEBUG
& DEBUG_NO8
))) {
8268 prog_data
->base
.dispatch_grf_start_reg
= v8
.payload
.num_regs
;
8269 prog_data
->reg_blocks_8
= brw_register_blocks(v8
.grf_used
);
8272 if (v8
.max_dispatch_width
>= 16 &&
8273 likely(!(INTEL_DEBUG
& DEBUG_NO16
) || use_rep_send
)) {
8274 /* Try a SIMD16 compile */
8275 fs_visitor
v16(compiler
, log_data
, mem_ctx
, &key
->base
,
8276 &prog_data
->base
, shader
, 16,
8277 shader_time_index16
);
8278 v16
.import_uniforms(&v8
);
8279 if (!v16
.run_fs(allow_spilling
, use_rep_send
)) {
8280 compiler
->shader_perf_log(log_data
,
8281 "SIMD16 shader failed to compile: %s",
8284 simd16_cfg
= v16
.cfg
;
8285 prog_data
->dispatch_grf_start_reg_16
= v16
.payload
.num_regs
;
8286 prog_data
->reg_blocks_16
= brw_register_blocks(v16
.grf_used
);
8290 /* Currently, the compiler only supports SIMD32 on SNB+ */
8291 if (v8
.max_dispatch_width
>= 32 && !use_rep_send
&&
8292 compiler
->devinfo
->gen
>= 6 &&
8293 unlikely(INTEL_DEBUG
& DEBUG_DO32
)) {
8294 /* Try a SIMD32 compile */
8295 fs_visitor
v32(compiler
, log_data
, mem_ctx
, &key
->base
,
8296 &prog_data
->base
, shader
, 32,
8297 shader_time_index32
);
8298 v32
.import_uniforms(&v8
);
8299 if (!v32
.run_fs(allow_spilling
, false)) {
8300 compiler
->shader_perf_log(log_data
,
8301 "SIMD32 shader failed to compile: %s",
8304 simd32_cfg
= v32
.cfg
;
8305 prog_data
->dispatch_grf_start_reg_32
= v32
.payload
.num_regs
;
8306 prog_data
->reg_blocks_32
= brw_register_blocks(v32
.grf_used
);
8310 /* When the caller requests a repclear shader, they want SIMD16-only */
8314 /* Prior to Iron Lake, the PS had a single shader offset with a jump table
8315 * at the top to select the shader. We've never implemented that.
8316 * Instead, we just give them exactly one shader and we pick the widest one
8319 if (compiler
->devinfo
->gen
< 5) {
8320 if (simd32_cfg
|| simd16_cfg
)
8326 /* If computed depth is enabled SNB only allows SIMD8. */
8327 if (compiler
->devinfo
->gen
== 6 &&
8328 prog_data
->computed_depth_mode
!= BRW_PSCDEPTH_OFF
)
8329 assert(simd16_cfg
== NULL
&& simd32_cfg
== NULL
);
8331 if (compiler
->devinfo
->gen
<= 5 && !simd8_cfg
) {
8332 /* Iron lake and earlier only have one Dispatch GRF start field. Make
8333 * the data available in the base prog data struct for convenience.
8336 prog_data
->base
.dispatch_grf_start_reg
=
8337 prog_data
->dispatch_grf_start_reg_16
;
8338 } else if (simd32_cfg
) {
8339 prog_data
->base
.dispatch_grf_start_reg
=
8340 prog_data
->dispatch_grf_start_reg_32
;
8344 if (prog_data
->persample_dispatch
) {
8345 /* Starting with SandyBridge (where we first get MSAA), the different
8346 * pixel dispatch combinations are grouped into classifications A
8347 * through F (SNB PRM Vol. 2 Part 1 Section 7.7.1). On all hardware
8348 * generations, the only configurations supporting persample dispatch
8349 * are are this in which only one dispatch width is enabled.
8351 if (simd32_cfg
|| simd16_cfg
)
8357 fs_generator
g(compiler
, log_data
, mem_ctx
, &prog_data
->base
,
8358 v8
.shader_stats
, v8
.runtime_check_aads_emit
,
8359 MESA_SHADER_FRAGMENT
);
8361 if (unlikely(INTEL_DEBUG
& DEBUG_WM
)) {
8362 g
.enable_debug(ralloc_asprintf(mem_ctx
, "%s fragment shader %s",
8363 shader
->info
.label
?
8364 shader
->info
.label
: "unnamed",
8365 shader
->info
.name
));
8369 prog_data
->dispatch_8
= true;
8370 g
.generate_code(simd8_cfg
, 8, stats
);
8371 stats
= stats
? stats
+ 1 : NULL
;
8375 prog_data
->dispatch_16
= true;
8376 prog_data
->prog_offset_16
= g
.generate_code(simd16_cfg
, 16, stats
);
8377 stats
= stats
? stats
+ 1 : NULL
;
8381 prog_data
->dispatch_32
= true;
8382 prog_data
->prog_offset_32
= g
.generate_code(simd32_cfg
, 32, stats
);
8383 stats
= stats
? stats
+ 1 : NULL
;
8386 return g
.get_assembly();
8390 fs_visitor::emit_cs_work_group_id_setup()
8392 assert(stage
== MESA_SHADER_COMPUTE
);
8394 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::uvec3_type
));
8396 struct brw_reg
r0_1(retype(brw_vec1_grf(0, 1), BRW_REGISTER_TYPE_UD
));
8397 struct brw_reg
r0_6(retype(brw_vec1_grf(0, 6), BRW_REGISTER_TYPE_UD
));
8398 struct brw_reg
r0_7(retype(brw_vec1_grf(0, 7), BRW_REGISTER_TYPE_UD
));
8400 bld
.MOV(*reg
, r0_1
);
8401 bld
.MOV(offset(*reg
, bld
, 1), r0_6
);
8402 bld
.MOV(offset(*reg
, bld
, 2), r0_7
);
8408 fill_push_const_block_info(struct brw_push_const_block
*block
, unsigned dwords
)
8410 block
->dwords
= dwords
;
8411 block
->regs
= DIV_ROUND_UP(dwords
, 8);
8412 block
->size
= block
->regs
* 32;
8416 cs_fill_push_const_info(const struct gen_device_info
*devinfo
,
8417 struct brw_cs_prog_data
*cs_prog_data
)
8419 const struct brw_stage_prog_data
*prog_data
= &cs_prog_data
->base
;
8420 int subgroup_id_index
= get_subgroup_id_param_index(prog_data
);
8421 bool cross_thread_supported
= devinfo
->gen
> 7 || devinfo
->is_haswell
;
8423 /* The thread ID should be stored in the last param dword */
8424 assert(subgroup_id_index
== -1 ||
8425 subgroup_id_index
== (int)prog_data
->nr_params
- 1);
8427 unsigned cross_thread_dwords
, per_thread_dwords
;
8428 if (!cross_thread_supported
) {
8429 cross_thread_dwords
= 0u;
8430 per_thread_dwords
= prog_data
->nr_params
;
8431 } else if (subgroup_id_index
>= 0) {
8432 /* Fill all but the last register with cross-thread payload */
8433 cross_thread_dwords
= 8 * (subgroup_id_index
/ 8);
8434 per_thread_dwords
= prog_data
->nr_params
- cross_thread_dwords
;
8435 assert(per_thread_dwords
> 0 && per_thread_dwords
<= 8);
8437 /* Fill all data using cross-thread payload */
8438 cross_thread_dwords
= prog_data
->nr_params
;
8439 per_thread_dwords
= 0u;
8442 fill_push_const_block_info(&cs_prog_data
->push
.cross_thread
, cross_thread_dwords
);
8443 fill_push_const_block_info(&cs_prog_data
->push
.per_thread
, per_thread_dwords
);
8445 unsigned total_dwords
=
8446 (cs_prog_data
->push
.per_thread
.size
* cs_prog_data
->threads
+
8447 cs_prog_data
->push
.cross_thread
.size
) / 4;
8448 fill_push_const_block_info(&cs_prog_data
->push
.total
, total_dwords
);
8450 assert(cs_prog_data
->push
.cross_thread
.dwords
% 8 == 0 ||
8451 cs_prog_data
->push
.per_thread
.size
== 0);
8452 assert(cs_prog_data
->push
.cross_thread
.dwords
+
8453 cs_prog_data
->push
.per_thread
.dwords
==
8454 prog_data
->nr_params
);
8458 cs_set_simd_size(struct brw_cs_prog_data
*cs_prog_data
, unsigned size
)
8460 cs_prog_data
->simd_size
= size
;
8461 unsigned group_size
= cs_prog_data
->local_size
[0] *
8462 cs_prog_data
->local_size
[1] * cs_prog_data
->local_size
[2];
8463 cs_prog_data
->threads
= (group_size
+ size
- 1) / size
;
8467 compile_cs_to_nir(const struct brw_compiler
*compiler
,
8469 const struct brw_cs_prog_key
*key
,
8470 const nir_shader
*src_shader
,
8471 unsigned dispatch_width
)
8473 nir_shader
*shader
= nir_shader_clone(mem_ctx
, src_shader
);
8474 brw_nir_apply_key(shader
, compiler
, &key
->base
, dispatch_width
, true);
8476 NIR_PASS_V(shader
, brw_nir_lower_cs_intrinsics
, dispatch_width
);
8478 /* Clean up after the local index and ID calculations. */
8479 NIR_PASS_V(shader
, nir_opt_constant_folding
);
8480 NIR_PASS_V(shader
, nir_opt_dce
);
8482 brw_postprocess_nir(shader
, compiler
, true);
8488 brw_compile_cs(const struct brw_compiler
*compiler
, void *log_data
,
8490 const struct brw_cs_prog_key
*key
,
8491 struct brw_cs_prog_data
*prog_data
,
8492 const nir_shader
*src_shader
,
8493 int shader_time_index
,
8494 struct brw_compile_stats
*stats
,
8497 prog_data
->base
.total_shared
= src_shader
->info
.cs
.shared_size
;
8498 prog_data
->local_size
[0] = src_shader
->info
.cs
.local_size
[0];
8499 prog_data
->local_size
[1] = src_shader
->info
.cs
.local_size
[1];
8500 prog_data
->local_size
[2] = src_shader
->info
.cs
.local_size
[2];
8501 prog_data
->slm_size
= src_shader
->num_shared
;
8502 unsigned local_workgroup_size
=
8503 src_shader
->info
.cs
.local_size
[0] * src_shader
->info
.cs
.local_size
[1] *
8504 src_shader
->info
.cs
.local_size
[2];
8506 unsigned min_dispatch_width
=
8507 DIV_ROUND_UP(local_workgroup_size
, compiler
->devinfo
->max_cs_threads
);
8508 min_dispatch_width
= MAX2(8, min_dispatch_width
);
8509 min_dispatch_width
= util_next_power_of_two(min_dispatch_width
);
8510 assert(min_dispatch_width
<= 32);
8511 unsigned max_dispatch_width
= 32;
8513 fs_visitor
*v8
= NULL
, *v16
= NULL
, *v32
= NULL
;
8514 fs_visitor
*v
= NULL
;
8515 const char *fail_msg
= NULL
;
8517 if ((int)key
->base
.subgroup_size_type
>= (int)BRW_SUBGROUP_SIZE_REQUIRE_8
) {
8518 /* These enum values are expressly chosen to be equal to the subgroup
8519 * size that they require.
8521 const unsigned required_dispatch_width
=
8522 (unsigned)key
->base
.subgroup_size_type
;
8523 assert(required_dispatch_width
== 8 ||
8524 required_dispatch_width
== 16 ||
8525 required_dispatch_width
== 32);
8526 if (required_dispatch_width
< min_dispatch_width
||
8527 required_dispatch_width
> max_dispatch_width
) {
8528 fail_msg
= "Cannot satisfy explicit subgroup size";
8530 min_dispatch_width
= max_dispatch_width
= required_dispatch_width
;
8534 /* Now the main event: Visit the shader IR and generate our CS IR for it.
8536 if (!fail_msg
&& min_dispatch_width
<= 8 && max_dispatch_width
>= 8) {
8537 nir_shader
*nir8
= compile_cs_to_nir(compiler
, mem_ctx
, key
,
8539 v8
= new fs_visitor(compiler
, log_data
, mem_ctx
, &key
->base
,
8541 nir8
, 8, shader_time_index
);
8542 if (!v8
->run_cs(min_dispatch_width
)) {
8543 fail_msg
= v8
->fail_msg
;
8545 /* We should always be able to do SIMD32 for compute shaders */
8546 assert(v8
->max_dispatch_width
>= 32);
8549 cs_set_simd_size(prog_data
, 8);
8550 cs_fill_push_const_info(compiler
->devinfo
, prog_data
);
8554 if (likely(!(INTEL_DEBUG
& DEBUG_NO16
)) &&
8555 !fail_msg
&& min_dispatch_width
<= 16 && max_dispatch_width
>= 16) {
8556 /* Try a SIMD16 compile */
8557 nir_shader
*nir16
= compile_cs_to_nir(compiler
, mem_ctx
, key
,
8559 v16
= new fs_visitor(compiler
, log_data
, mem_ctx
, &key
->base
,
8561 nir16
, 16, shader_time_index
);
8563 v16
->import_uniforms(v8
);
8565 if (!v16
->run_cs(min_dispatch_width
)) {
8566 compiler
->shader_perf_log(log_data
,
8567 "SIMD16 shader failed to compile: %s",
8571 "Couldn't generate SIMD16 program and not "
8572 "enough threads for SIMD8";
8575 /* We should always be able to do SIMD32 for compute shaders */
8576 assert(v16
->max_dispatch_width
>= 32);
8579 cs_set_simd_size(prog_data
, 16);
8580 cs_fill_push_const_info(compiler
->devinfo
, prog_data
);
8584 /* We should always be able to do SIMD32 for compute shaders */
8585 assert(!v16
|| v16
->max_dispatch_width
>= 32);
8587 if (!fail_msg
&& (min_dispatch_width
> 16 || (INTEL_DEBUG
& DEBUG_DO32
)) &&
8588 max_dispatch_width
>= 32) {
8589 /* Try a SIMD32 compile */
8590 nir_shader
*nir32
= compile_cs_to_nir(compiler
, mem_ctx
, key
,
8592 v32
= new fs_visitor(compiler
, log_data
, mem_ctx
, &key
->base
,
8594 nir32
, 32, shader_time_index
);
8596 v32
->import_uniforms(v8
);
8598 v32
->import_uniforms(v16
);
8600 if (!v32
->run_cs(min_dispatch_width
)) {
8601 compiler
->shader_perf_log(log_data
,
8602 "SIMD32 shader failed to compile: %s",
8606 "Couldn't generate SIMD32 program and not "
8607 "enough threads for SIMD16";
8611 cs_set_simd_size(prog_data
, 32);
8612 cs_fill_push_const_info(compiler
->devinfo
, prog_data
);
8616 const unsigned *ret
= NULL
;
8617 if (unlikely(v
== NULL
)) {
8620 *error_str
= ralloc_strdup(mem_ctx
, fail_msg
);
8622 fs_generator
g(compiler
, log_data
, mem_ctx
, &prog_data
->base
,
8623 v
->shader_stats
, v
->runtime_check_aads_emit
,
8624 MESA_SHADER_COMPUTE
);
8625 if (INTEL_DEBUG
& DEBUG_CS
) {
8626 char *name
= ralloc_asprintf(mem_ctx
, "%s compute shader %s",
8627 src_shader
->info
.label
?
8628 src_shader
->info
.label
: "unnamed",
8629 src_shader
->info
.name
);
8630 g
.enable_debug(name
);
8633 g
.generate_code(v
->cfg
, prog_data
->simd_size
, stats
);
8635 ret
= g
.get_assembly();
8646 * Test the dispatch mask packing assumptions of
8647 * brw_stage_has_packed_dispatch(). Call this from e.g. the top of
8648 * fs_visitor::emit_nir_code() to cause a GPU hang if any shader invocation is
8649 * executed with an unexpected dispatch mask.
8652 brw_fs_test_dispatch_packing(const fs_builder
&bld
)
8654 const gl_shader_stage stage
= bld
.shader
->stage
;
8656 if (brw_stage_has_packed_dispatch(bld
.shader
->devinfo
, stage
,
8657 bld
.shader
->stage_prog_data
)) {
8658 const fs_builder ubld
= bld
.exec_all().group(1, 0);
8659 const fs_reg tmp
= component(bld
.vgrf(BRW_REGISTER_TYPE_UD
), 0);
8660 const fs_reg mask
= (stage
== MESA_SHADER_FRAGMENT
? brw_vmask_reg() :
8663 ubld
.ADD(tmp
, mask
, brw_imm_ud(1));
8664 ubld
.AND(tmp
, mask
, tmp
);
8666 /* This will loop forever if the dispatch mask doesn't have the expected
8667 * form '2^n-1', in which case tmp will be non-zero.
8669 bld
.emit(BRW_OPCODE_DO
);
8670 bld
.CMP(bld
.null_reg_ud(), tmp
, brw_imm_ud(0), BRW_CONDITIONAL_NZ
);
8671 set_predicate(BRW_PREDICATE_NORMAL
, bld
.emit(BRW_OPCODE_WHILE
));