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"
34 #include "brw_fs_live_variables.h"
36 #include "brw_vec4_gs_visitor.h"
38 #include "brw_dead_control_flow.h"
39 #include "dev/gen_debug.h"
40 #include "compiler/glsl_types.h"
41 #include "compiler/nir/nir_builder.h"
42 #include "program/prog_parameter.h"
43 #include "util/u_math.h"
47 static unsigned get_lowered_simd_width(const struct gen_device_info
*devinfo
,
51 fs_inst::init(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
52 const fs_reg
*src
, unsigned sources
)
54 memset((void*)this, 0, sizeof(*this));
56 this->src
= new fs_reg
[MAX2(sources
, 3)];
57 for (unsigned i
= 0; i
< sources
; i
++)
58 this->src
[i
] = src
[i
];
60 this->opcode
= opcode
;
62 this->sources
= sources
;
63 this->exec_size
= exec_size
;
66 assert(dst
.file
!= IMM
&& dst
.file
!= UNIFORM
);
68 assert(this->exec_size
!= 0);
70 this->conditional_mod
= BRW_CONDITIONAL_NONE
;
72 /* This will be the case for almost all instructions. */
79 this->size_written
= dst
.component_size(exec_size
);
82 this->size_written
= 0;
86 unreachable("Invalid destination register file");
89 this->writes_accumulator
= false;
94 init(BRW_OPCODE_NOP
, 8, dst
, NULL
, 0);
97 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
)
99 init(opcode
, exec_size
, reg_undef
, NULL
, 0);
102 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
)
104 init(opcode
, exec_size
, dst
, NULL
, 0);
107 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
110 const fs_reg src
[1] = { src0
};
111 init(opcode
, exec_size
, dst
, src
, 1);
114 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
115 const fs_reg
&src0
, const fs_reg
&src1
)
117 const fs_reg src
[2] = { src0
, src1
};
118 init(opcode
, exec_size
, dst
, src
, 2);
121 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
122 const fs_reg
&src0
, const fs_reg
&src1
, const fs_reg
&src2
)
124 const fs_reg src
[3] = { src0
, src1
, src2
};
125 init(opcode
, exec_size
, dst
, src
, 3);
128 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_width
, const fs_reg
&dst
,
129 const fs_reg src
[], unsigned sources
)
131 init(opcode
, exec_width
, dst
, src
, sources
);
134 fs_inst::fs_inst(const fs_inst
&that
)
136 memcpy((void*)this, &that
, sizeof(that
));
138 this->src
= new fs_reg
[MAX2(that
.sources
, 3)];
140 for (unsigned i
= 0; i
< that
.sources
; i
++)
141 this->src
[i
] = that
.src
[i
];
150 fs_inst::resize_sources(uint8_t num_sources
)
152 if (this->sources
!= num_sources
) {
153 fs_reg
*src
= new fs_reg
[MAX2(num_sources
, 3)];
155 for (unsigned i
= 0; i
< MIN2(this->sources
, num_sources
); ++i
)
156 src
[i
] = this->src
[i
];
160 this->sources
= num_sources
;
165 fs_visitor::VARYING_PULL_CONSTANT_LOAD(const fs_builder
&bld
,
167 const fs_reg
&surf_index
,
168 const fs_reg
&varying_offset
,
169 uint32_t const_offset
)
171 /* We have our constant surface use a pitch of 4 bytes, so our index can
172 * be any component of a vector, and then we load 4 contiguous
173 * components starting from that.
175 * We break down the const_offset to a portion added to the variable offset
176 * and a portion done using fs_reg::offset, which means that if you have
177 * GLSL using something like "uniform vec4 a[20]; gl_FragColor = a[i]",
178 * we'll temporarily generate 4 vec4 loads from offset i * 4, and CSE can
179 * later notice that those loads are all the same and eliminate the
182 fs_reg vec4_offset
= vgrf(glsl_type::uint_type
);
183 bld
.ADD(vec4_offset
, varying_offset
, brw_imm_ud(const_offset
& ~0xf));
185 /* The pull load message will load a vec4 (16 bytes). If we are loading
186 * a double this means we are only loading 2 elements worth of data.
187 * We also want to use a 32-bit data type for the dst of the load operation
188 * so other parts of the driver don't get confused about the size of the
191 fs_reg vec4_result
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 4);
192 fs_inst
*inst
= bld
.emit(FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_LOGICAL
,
193 vec4_result
, surf_index
, vec4_offset
);
194 inst
->size_written
= 4 * vec4_result
.component_size(inst
->exec_size
);
196 shuffle_from_32bit_read(bld
, dst
, vec4_result
,
197 (const_offset
& 0xf) / type_sz(dst
.type
), 1);
201 * A helper for MOV generation for fixing up broken hardware SEND dependency
205 fs_visitor::DEP_RESOLVE_MOV(const fs_builder
&bld
, int grf
)
207 /* The caller always wants uncompressed to emit the minimal extra
208 * dependencies, and to avoid having to deal with aligning its regs to 2.
210 const fs_builder ubld
= bld
.annotate("send dependency resolve")
213 ubld
.MOV(ubld
.null_reg_f(), fs_reg(VGRF
, grf
, BRW_REGISTER_TYPE_F
));
217 fs_inst::is_send_from_grf() const
220 case SHADER_OPCODE_SEND
:
221 case SHADER_OPCODE_SHADER_TIME_ADD
:
222 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
223 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
224 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
225 case SHADER_OPCODE_URB_WRITE_SIMD8
:
226 case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
227 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
:
228 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
229 case SHADER_OPCODE_URB_READ_SIMD8
:
230 case SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
:
231 case SHADER_OPCODE_INTERLOCK
:
232 case SHADER_OPCODE_MEMORY_FENCE
:
233 case SHADER_OPCODE_BARRIER
:
235 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
236 return src
[1].file
== VGRF
;
237 case FS_OPCODE_FB_WRITE
:
238 case FS_OPCODE_FB_READ
:
239 return src
[0].file
== VGRF
;
242 return src
[0].file
== VGRF
;
249 fs_inst::is_control_source(unsigned arg
) const
252 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
253 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7
:
254 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN4
:
257 case SHADER_OPCODE_BROADCAST
:
258 case SHADER_OPCODE_SHUFFLE
:
259 case SHADER_OPCODE_QUAD_SWIZZLE
:
260 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
261 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
262 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
263 case SHADER_OPCODE_GET_BUFFER_SIZE
:
266 case SHADER_OPCODE_MOV_INDIRECT
:
267 case SHADER_OPCODE_CLUSTER_BROADCAST
:
268 case SHADER_OPCODE_TEX
:
270 case SHADER_OPCODE_TXD
:
271 case SHADER_OPCODE_TXF
:
272 case SHADER_OPCODE_TXF_LZ
:
273 case SHADER_OPCODE_TXF_CMS
:
274 case SHADER_OPCODE_TXF_CMS_W
:
275 case SHADER_OPCODE_TXF_UMS
:
276 case SHADER_OPCODE_TXF_MCS
:
277 case SHADER_OPCODE_TXL
:
278 case SHADER_OPCODE_TXL_LZ
:
279 case SHADER_OPCODE_TXS
:
280 case SHADER_OPCODE_LOD
:
281 case SHADER_OPCODE_TG4
:
282 case SHADER_OPCODE_TG4_OFFSET
:
283 case SHADER_OPCODE_SAMPLEINFO
:
284 return arg
== 1 || arg
== 2;
286 case SHADER_OPCODE_SEND
:
287 return arg
== 0 || arg
== 1;
295 fs_inst::is_payload(unsigned arg
) const
298 case FS_OPCODE_FB_WRITE
:
299 case FS_OPCODE_FB_READ
:
300 case SHADER_OPCODE_URB_WRITE_SIMD8
:
301 case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
302 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
:
303 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
304 case SHADER_OPCODE_URB_READ_SIMD8
:
305 case SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
:
306 case VEC4_OPCODE_UNTYPED_ATOMIC
:
307 case VEC4_OPCODE_UNTYPED_SURFACE_READ
:
308 case VEC4_OPCODE_UNTYPED_SURFACE_WRITE
:
309 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
310 case SHADER_OPCODE_SHADER_TIME_ADD
:
311 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
312 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
313 case SHADER_OPCODE_INTERLOCK
:
314 case SHADER_OPCODE_MEMORY_FENCE
:
315 case SHADER_OPCODE_BARRIER
:
318 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7
:
321 case SHADER_OPCODE_SEND
:
322 return arg
== 2 || arg
== 3;
333 * Returns true if this instruction's sources and destinations cannot
334 * safely be the same register.
336 * In most cases, a register can be written over safely by the same
337 * instruction that is its last use. For a single instruction, the
338 * sources are dereferenced before writing of the destination starts
341 * However, there are a few cases where this can be problematic:
343 * - Virtual opcodes that translate to multiple instructions in the
344 * code generator: if src == dst and one instruction writes the
345 * destination before a later instruction reads the source, then
346 * src will have been clobbered.
348 * - SIMD16 compressed instructions with certain regioning (see below).
350 * The register allocator uses this information to set up conflicts between
351 * GRF sources and the destination.
354 fs_inst::has_source_and_destination_hazard() const
357 case FS_OPCODE_PACK_HALF_2x16_SPLIT
:
358 /* Multiple partial writes to the destination */
360 case SHADER_OPCODE_SHUFFLE
:
361 /* This instruction returns an arbitrary channel from the source and
362 * gets split into smaller instructions in the generator. It's possible
363 * that one of the instructions will read from a channel corresponding
364 * to an earlier instruction.
366 case SHADER_OPCODE_SEL_EXEC
:
367 /* This is implemented as
369 * mov(16) g4<1>D 0D { align1 WE_all 1H };
370 * mov(16) g4<1>D g5<8,8,1>D { align1 1H }
372 * Because the source is only read in the second instruction, the first
373 * may stomp all over it.
376 case SHADER_OPCODE_QUAD_SWIZZLE
:
378 case BRW_SWIZZLE_XXXX
:
379 case BRW_SWIZZLE_YYYY
:
380 case BRW_SWIZZLE_ZZZZ
:
381 case BRW_SWIZZLE_WWWW
:
382 case BRW_SWIZZLE_XXZZ
:
383 case BRW_SWIZZLE_YYWW
:
384 case BRW_SWIZZLE_XYXY
:
385 case BRW_SWIZZLE_ZWZW
:
386 /* These can be implemented as a single Align1 region on all
387 * platforms, so there's never a hazard between source and
388 * destination. C.f. fs_generator::generate_quad_swizzle().
392 return !is_uniform(src
[0]);
395 /* The SIMD16 compressed instruction
397 * add(16) g4<1>F g4<8,8,1>F g6<8,8,1>F
399 * is actually decoded in hardware as:
401 * add(8) g4<1>F g4<8,8,1>F g6<8,8,1>F
402 * add(8) g5<1>F g5<8,8,1>F g7<8,8,1>F
404 * Which is safe. However, if we have uniform accesses
405 * happening, we get into trouble:
407 * add(8) g4<1>F g4<0,1,0>F g6<8,8,1>F
408 * add(8) g5<1>F g4<0,1,0>F g7<8,8,1>F
410 * Now our destination for the first instruction overwrote the
411 * second instruction's src0, and we get garbage for those 8
412 * pixels. There's a similar issue for the pre-gen6
413 * pixel_x/pixel_y, which are registers of 16-bit values and thus
414 * would get stomped by the first decode as well.
416 if (exec_size
== 16) {
417 for (int i
= 0; i
< sources
; i
++) {
418 if (src
[i
].file
== VGRF
&& (src
[i
].stride
== 0 ||
419 src
[i
].type
== BRW_REGISTER_TYPE_UW
||
420 src
[i
].type
== BRW_REGISTER_TYPE_W
||
421 src
[i
].type
== BRW_REGISTER_TYPE_UB
||
422 src
[i
].type
== BRW_REGISTER_TYPE_B
)) {
432 fs_inst::can_do_source_mods(const struct gen_device_info
*devinfo
) const
434 if (devinfo
->gen
== 6 && is_math())
437 if (is_send_from_grf())
440 /* From GEN:BUG:1604601757:
442 * "When multiplying a DW and any lower precision integer, source modifier
445 if (devinfo
->gen
>= 12 && (opcode
== BRW_OPCODE_MUL
||
446 opcode
== BRW_OPCODE_MAD
)) {
447 const brw_reg_type exec_type
= get_exec_type(this);
448 const unsigned min_type_sz
= opcode
== BRW_OPCODE_MAD
?
449 MIN2(type_sz(src
[1].type
), type_sz(src
[2].type
)) :
450 MIN2(type_sz(src
[0].type
), type_sz(src
[1].type
));
452 if (brw_reg_type_is_integer(exec_type
) &&
453 type_sz(exec_type
) >= 4 &&
454 type_sz(exec_type
) != min_type_sz
)
458 if (!backend_instruction::can_do_source_mods())
465 fs_inst::can_do_cmod()
467 if (!backend_instruction::can_do_cmod())
470 /* The accumulator result appears to get used for the conditional modifier
471 * generation. When negating a UD value, there is a 33rd bit generated for
472 * the sign in the accumulator value, so now you can't check, for example,
473 * equality with a 32-bit value. See piglit fs-op-neg-uvec4.
475 for (unsigned i
= 0; i
< sources
; i
++) {
476 if (type_is_unsigned_int(src
[i
].type
) && src
[i
].negate
)
484 fs_inst::can_change_types() const
486 return dst
.type
== src
[0].type
&&
487 !src
[0].abs
&& !src
[0].negate
&& !saturate
&&
488 (opcode
== BRW_OPCODE_MOV
||
489 (opcode
== BRW_OPCODE_SEL
&&
490 dst
.type
== src
[1].type
&&
491 predicate
!= BRW_PREDICATE_NONE
&&
492 !src
[1].abs
&& !src
[1].negate
));
498 memset((void*)this, 0, sizeof(*this));
499 type
= BRW_REGISTER_TYPE_UD
;
503 /** Generic unset register constructor. */
507 this->file
= BAD_FILE
;
510 fs_reg::fs_reg(struct ::brw_reg reg
) :
515 if (this->file
== IMM
&&
516 (this->type
!= BRW_REGISTER_TYPE_V
&&
517 this->type
!= BRW_REGISTER_TYPE_UV
&&
518 this->type
!= BRW_REGISTER_TYPE_VF
)) {
524 fs_reg::equals(const fs_reg
&r
) const
526 return (this->backend_reg::equals(r
) &&
531 fs_reg::negative_equals(const fs_reg
&r
) const
533 return (this->backend_reg::negative_equals(r
) &&
538 fs_reg::is_contiguous() const
543 return hstride
== BRW_HORIZONTAL_STRIDE_1
&&
544 vstride
== width
+ hstride
;
555 unreachable("Invalid register file");
559 fs_reg::component_size(unsigned width
) const
561 const unsigned stride
= ((file
!= ARF
&& file
!= FIXED_GRF
) ? this->stride
:
564 return MAX2(width
* stride
, 1) * type_sz(type
);
568 * Create a MOV to read the timestamp register.
571 fs_visitor::get_timestamp(const fs_builder
&bld
)
573 assert(devinfo
->gen
>= 7);
575 fs_reg ts
= fs_reg(retype(brw_vec4_reg(BRW_ARCHITECTURE_REGISTER_FILE
,
578 BRW_REGISTER_TYPE_UD
));
580 fs_reg dst
= fs_reg(VGRF
, alloc
.allocate(1), BRW_REGISTER_TYPE_UD
);
582 /* We want to read the 3 fields we care about even if it's not enabled in
585 bld
.group(4, 0).exec_all().MOV(dst
, ts
);
591 fs_visitor::emit_shader_time_begin()
593 /* We want only the low 32 bits of the timestamp. Since it's running
594 * at the GPU clock rate of ~1.2ghz, it will roll over every ~3 seconds,
595 * which is plenty of time for our purposes. It is identical across the
596 * EUs, but since it's tracking GPU core speed it will increment at a
597 * varying rate as render P-states change.
599 shader_start_time
= component(
600 get_timestamp(bld
.annotate("shader time start")), 0);
604 fs_visitor::emit_shader_time_end()
606 /* Insert our code just before the final SEND with EOT. */
607 exec_node
*end
= this->instructions
.get_tail();
608 assert(end
&& ((fs_inst
*) end
)->eot
);
609 const fs_builder ibld
= bld
.annotate("shader time end")
610 .exec_all().at(NULL
, end
);
611 const fs_reg timestamp
= get_timestamp(ibld
);
613 /* We only use the low 32 bits of the timestamp - see
614 * emit_shader_time_begin()).
616 * We could also check if render P-states have changed (or anything
617 * else that might disrupt timing) by setting smear to 2 and checking if
618 * that field is != 0.
620 const fs_reg shader_end_time
= component(timestamp
, 0);
622 /* Check that there weren't any timestamp reset events (assuming these
623 * were the only two timestamp reads that happened).
625 const fs_reg reset
= component(timestamp
, 2);
626 set_condmod(BRW_CONDITIONAL_Z
,
627 ibld
.AND(ibld
.null_reg_ud(), reset
, brw_imm_ud(1u)));
628 ibld
.IF(BRW_PREDICATE_NORMAL
);
630 fs_reg start
= shader_start_time
;
632 const fs_reg diff
= component(fs_reg(VGRF
, alloc
.allocate(1),
633 BRW_REGISTER_TYPE_UD
),
635 const fs_builder cbld
= ibld
.group(1, 0);
636 cbld
.group(1, 0).ADD(diff
, start
, shader_end_time
);
638 /* If there were no instructions between the two timestamp gets, the diff
639 * is 2 cycles. Remove that overhead, so I can forget about that when
640 * trying to determine the time taken for single instructions.
642 cbld
.ADD(diff
, diff
, brw_imm_ud(-2u));
643 SHADER_TIME_ADD(cbld
, 0, diff
);
644 SHADER_TIME_ADD(cbld
, 1, brw_imm_ud(1u));
645 ibld
.emit(BRW_OPCODE_ELSE
);
646 SHADER_TIME_ADD(cbld
, 2, brw_imm_ud(1u));
647 ibld
.emit(BRW_OPCODE_ENDIF
);
651 fs_visitor::SHADER_TIME_ADD(const fs_builder
&bld
,
652 int shader_time_subindex
,
655 int index
= shader_time_index
* 3 + shader_time_subindex
;
656 struct brw_reg offset
= brw_imm_d(index
* BRW_SHADER_TIME_STRIDE
);
659 if (dispatch_width
== 8)
660 payload
= vgrf(glsl_type::uvec2_type
);
662 payload
= vgrf(glsl_type::uint_type
);
664 bld
.emit(SHADER_OPCODE_SHADER_TIME_ADD
, fs_reg(), payload
, offset
, value
);
668 fs_visitor::vfail(const char *format
, va_list va
)
677 msg
= ralloc_vasprintf(mem_ctx
, format
, va
);
678 msg
= ralloc_asprintf(mem_ctx
, "%s compile failed: %s\n", stage_abbrev
, msg
);
680 this->fail_msg
= msg
;
683 fprintf(stderr
, "%s", msg
);
688 fs_visitor::fail(const char *format
, ...)
692 va_start(va
, format
);
698 * Mark this program as impossible to compile with dispatch width greater
701 * During the SIMD8 compile (which happens first), we can detect and flag
702 * things that are unsupported in SIMD16+ mode, so the compiler can skip the
703 * SIMD16+ compile altogether.
705 * During a compile of dispatch width greater than n (if one happens anyway),
706 * this just calls fail().
709 fs_visitor::limit_dispatch_width(unsigned n
, const char *msg
)
711 if (dispatch_width
> n
) {
714 max_dispatch_width
= n
;
715 compiler
->shader_perf_log(log_data
,
716 "Shader dispatch width limited to SIMD%d: %s",
722 * Returns true if the instruction has a flag that means it won't
723 * update an entire destination register.
725 * For example, dead code elimination and live variable analysis want to know
726 * when a write to a variable screens off any preceding values that were in
730 fs_inst::is_partial_write() const
732 return ((this->predicate
&& this->opcode
!= BRW_OPCODE_SEL
) ||
733 (this->exec_size
* type_sz(this->dst
.type
)) < 32 ||
734 !this->dst
.is_contiguous() ||
735 this->dst
.offset
% REG_SIZE
!= 0);
739 fs_inst::components_read(unsigned i
) const
741 /* Return zero if the source is not present. */
742 if (src
[i
].file
== BAD_FILE
)
746 case FS_OPCODE_LINTERP
:
752 case FS_OPCODE_PIXEL_X
:
753 case FS_OPCODE_PIXEL_Y
:
757 case FS_OPCODE_FB_WRITE_LOGICAL
:
758 assert(src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].file
== IMM
);
759 /* First/second FB write color. */
761 return src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].ud
;
765 case SHADER_OPCODE_TEX_LOGICAL
:
766 case SHADER_OPCODE_TXD_LOGICAL
:
767 case SHADER_OPCODE_TXF_LOGICAL
:
768 case SHADER_OPCODE_TXL_LOGICAL
:
769 case SHADER_OPCODE_TXS_LOGICAL
:
770 case SHADER_OPCODE_IMAGE_SIZE_LOGICAL
:
771 case FS_OPCODE_TXB_LOGICAL
:
772 case SHADER_OPCODE_TXF_CMS_LOGICAL
:
773 case SHADER_OPCODE_TXF_CMS_W_LOGICAL
:
774 case SHADER_OPCODE_TXF_UMS_LOGICAL
:
775 case SHADER_OPCODE_TXF_MCS_LOGICAL
:
776 case SHADER_OPCODE_LOD_LOGICAL
:
777 case SHADER_OPCODE_TG4_LOGICAL
:
778 case SHADER_OPCODE_TG4_OFFSET_LOGICAL
:
779 case SHADER_OPCODE_SAMPLEINFO_LOGICAL
:
780 assert(src
[TEX_LOGICAL_SRC_COORD_COMPONENTS
].file
== IMM
&&
781 src
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
].file
== IMM
);
782 /* Texture coordinates. */
783 if (i
== TEX_LOGICAL_SRC_COORDINATE
)
784 return src
[TEX_LOGICAL_SRC_COORD_COMPONENTS
].ud
;
785 /* Texture derivatives. */
786 else if ((i
== TEX_LOGICAL_SRC_LOD
|| i
== TEX_LOGICAL_SRC_LOD2
) &&
787 opcode
== SHADER_OPCODE_TXD_LOGICAL
)
788 return src
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
].ud
;
789 /* Texture offset. */
790 else if (i
== TEX_LOGICAL_SRC_TG4_OFFSET
)
793 else if (i
== TEX_LOGICAL_SRC_MCS
&& opcode
== SHADER_OPCODE_TXF_CMS_W_LOGICAL
)
798 case SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
:
799 case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
:
800 assert(src
[SURFACE_LOGICAL_SRC_IMM_DIMS
].file
== IMM
);
801 /* Surface coordinates. */
802 if (i
== SURFACE_LOGICAL_SRC_ADDRESS
)
803 return src
[SURFACE_LOGICAL_SRC_IMM_DIMS
].ud
;
804 /* Surface operation source (ignored for reads). */
805 else if (i
== SURFACE_LOGICAL_SRC_DATA
)
810 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
:
811 case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
:
812 assert(src
[SURFACE_LOGICAL_SRC_IMM_DIMS
].file
== IMM
&&
813 src
[SURFACE_LOGICAL_SRC_IMM_ARG
].file
== IMM
);
814 /* Surface coordinates. */
815 if (i
== SURFACE_LOGICAL_SRC_ADDRESS
)
816 return src
[SURFACE_LOGICAL_SRC_IMM_DIMS
].ud
;
817 /* Surface operation source. */
818 else if (i
== SURFACE_LOGICAL_SRC_DATA
)
819 return src
[SURFACE_LOGICAL_SRC_IMM_ARG
].ud
;
823 case SHADER_OPCODE_A64_UNTYPED_READ_LOGICAL
:
824 assert(src
[2].file
== IMM
);
827 case SHADER_OPCODE_A64_UNTYPED_WRITE_LOGICAL
:
828 assert(src
[2].file
== IMM
);
829 return i
== 1 ? src
[2].ud
: 1;
831 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_LOGICAL
:
832 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_INT64_LOGICAL
:
833 assert(src
[2].file
== IMM
);
836 const unsigned op
= src
[2].ud
;
851 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_FLOAT_LOGICAL
:
852 assert(src
[2].file
== IMM
);
855 const unsigned op
= src
[2].ud
;
856 return op
== BRW_AOP_FCMPWR
? 2 : 1;
861 case SHADER_OPCODE_BYTE_SCATTERED_READ_LOGICAL
:
862 case SHADER_OPCODE_DWORD_SCATTERED_READ_LOGICAL
:
863 /* Scattered logical opcodes use the following params:
864 * src[0] Surface coordinates
865 * src[1] Surface operation source (ignored for reads)
867 * src[3] IMM with always 1 dimension.
868 * src[4] IMM with arg bitsize for scattered read/write 8, 16, 32
870 assert(src
[SURFACE_LOGICAL_SRC_IMM_DIMS
].file
== IMM
&&
871 src
[SURFACE_LOGICAL_SRC_IMM_ARG
].file
== IMM
);
872 return i
== SURFACE_LOGICAL_SRC_DATA
? 0 : 1;
874 case SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL
:
875 case SHADER_OPCODE_DWORD_SCATTERED_WRITE_LOGICAL
:
876 assert(src
[SURFACE_LOGICAL_SRC_IMM_DIMS
].file
== IMM
&&
877 src
[SURFACE_LOGICAL_SRC_IMM_ARG
].file
== IMM
);
880 case SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL
:
881 case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
: {
882 assert(src
[SURFACE_LOGICAL_SRC_IMM_DIMS
].file
== IMM
&&
883 src
[SURFACE_LOGICAL_SRC_IMM_ARG
].file
== IMM
);
884 const unsigned op
= src
[SURFACE_LOGICAL_SRC_IMM_ARG
].ud
;
885 /* Surface coordinates. */
886 if (i
== SURFACE_LOGICAL_SRC_ADDRESS
)
887 return src
[SURFACE_LOGICAL_SRC_IMM_DIMS
].ud
;
888 /* Surface operation source. */
889 else if (i
== SURFACE_LOGICAL_SRC_DATA
&& op
== BRW_AOP_CMPWR
)
891 else if (i
== SURFACE_LOGICAL_SRC_DATA
&&
892 (op
== BRW_AOP_INC
|| op
== BRW_AOP_DEC
|| op
== BRW_AOP_PREDEC
))
897 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
898 return (i
== 0 ? 2 : 1);
900 case SHADER_OPCODE_UNTYPED_ATOMIC_FLOAT_LOGICAL
: {
901 assert(src
[SURFACE_LOGICAL_SRC_IMM_DIMS
].file
== IMM
&&
902 src
[SURFACE_LOGICAL_SRC_IMM_ARG
].file
== IMM
);
903 const unsigned op
= src
[SURFACE_LOGICAL_SRC_IMM_ARG
].ud
;
904 /* Surface coordinates. */
905 if (i
== SURFACE_LOGICAL_SRC_ADDRESS
)
906 return src
[SURFACE_LOGICAL_SRC_IMM_DIMS
].ud
;
907 /* Surface operation source. */
908 else if (i
== SURFACE_LOGICAL_SRC_DATA
&& op
== BRW_AOP_FCMPWR
)
920 fs_inst::size_read(int arg
) const
923 case SHADER_OPCODE_SEND
:
925 return mlen
* REG_SIZE
;
926 } else if (arg
== 3) {
927 return ex_mlen
* REG_SIZE
;
931 case FS_OPCODE_FB_WRITE
:
932 case FS_OPCODE_REP_FB_WRITE
:
935 return src
[0].file
== BAD_FILE
? 0 : 2 * REG_SIZE
;
937 return mlen
* REG_SIZE
;
941 case FS_OPCODE_FB_READ
:
942 case SHADER_OPCODE_URB_WRITE_SIMD8
:
943 case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
944 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
:
945 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
946 case SHADER_OPCODE_URB_READ_SIMD8
:
947 case SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
:
948 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
949 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
951 return mlen
* REG_SIZE
;
954 case FS_OPCODE_SET_SAMPLE_ID
:
959 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7
:
960 /* The payload is actually stored in src1 */
962 return mlen
* REG_SIZE
;
965 case FS_OPCODE_LINTERP
:
970 case SHADER_OPCODE_LOAD_PAYLOAD
:
971 if (arg
< this->header_size
)
975 case CS_OPCODE_CS_TERMINATE
:
976 case SHADER_OPCODE_BARRIER
:
979 case SHADER_OPCODE_MOV_INDIRECT
:
981 assert(src
[2].file
== IMM
);
987 if (is_tex() && arg
== 0 && src
[0].file
== VGRF
)
988 return mlen
* REG_SIZE
;
992 switch (src
[arg
].file
) {
995 return components_read(arg
) * type_sz(src
[arg
].type
);
1001 return components_read(arg
) * src
[arg
].component_size(exec_size
);
1003 unreachable("MRF registers are not allowed as sources");
1010 predicate_width(brw_predicate predicate
)
1012 switch (predicate
) {
1013 case BRW_PREDICATE_NONE
: return 1;
1014 case BRW_PREDICATE_NORMAL
: return 1;
1015 case BRW_PREDICATE_ALIGN1_ANY2H
: return 2;
1016 case BRW_PREDICATE_ALIGN1_ALL2H
: return 2;
1017 case BRW_PREDICATE_ALIGN1_ANY4H
: return 4;
1018 case BRW_PREDICATE_ALIGN1_ALL4H
: return 4;
1019 case BRW_PREDICATE_ALIGN1_ANY8H
: return 8;
1020 case BRW_PREDICATE_ALIGN1_ALL8H
: return 8;
1021 case BRW_PREDICATE_ALIGN1_ANY16H
: return 16;
1022 case BRW_PREDICATE_ALIGN1_ALL16H
: return 16;
1023 case BRW_PREDICATE_ALIGN1_ANY32H
: return 32;
1024 case BRW_PREDICATE_ALIGN1_ALL32H
: return 32;
1025 default: unreachable("Unsupported predicate");
1029 /* Return the subset of flag registers that an instruction could
1030 * potentially read or write based on the execution controls and flag
1031 * subregister number of the instruction.
1034 flag_mask(const fs_inst
*inst
, unsigned width
)
1036 assert(util_is_power_of_two_nonzero(width
));
1037 const unsigned start
= (inst
->flag_subreg
* 16 + inst
->group
) &
1039 const unsigned end
= start
+ ALIGN(inst
->exec_size
, width
);
1040 return ((1 << DIV_ROUND_UP(end
, 8)) - 1) & ~((1 << (start
/ 8)) - 1);
1044 bit_mask(unsigned n
)
1046 return (n
>= CHAR_BIT
* sizeof(bit_mask(n
)) ? ~0u : (1u << n
) - 1);
1050 flag_mask(const fs_reg
&r
, unsigned sz
)
1052 if (r
.file
== ARF
) {
1053 const unsigned start
= (r
.nr
- BRW_ARF_FLAG
) * 4 + r
.subnr
;
1054 const unsigned end
= start
+ sz
;
1055 return bit_mask(end
) & ~bit_mask(start
);
1063 fs_inst::flags_read(const gen_device_info
*devinfo
) const
1065 if (predicate
== BRW_PREDICATE_ALIGN1_ANYV
||
1066 predicate
== BRW_PREDICATE_ALIGN1_ALLV
) {
1067 /* The vertical predication modes combine corresponding bits from
1068 * f0.0 and f1.0 on Gen7+, and f0.0 and f0.1 on older hardware.
1070 const unsigned shift
= devinfo
->gen
>= 7 ? 4 : 2;
1071 return flag_mask(this, 1) << shift
| flag_mask(this, 1);
1072 } else if (predicate
) {
1073 return flag_mask(this, predicate_width(predicate
));
1076 for (int i
= 0; i
< sources
; i
++) {
1077 mask
|= flag_mask(src
[i
], size_read(i
));
1084 fs_inst::flags_written() const
1086 if ((conditional_mod
&& (opcode
!= BRW_OPCODE_SEL
&&
1087 opcode
!= BRW_OPCODE_CSEL
&&
1088 opcode
!= BRW_OPCODE_IF
&&
1089 opcode
!= BRW_OPCODE_WHILE
)) ||
1090 opcode
== FS_OPCODE_FB_WRITE
) {
1091 return flag_mask(this, 1);
1092 } else if (opcode
== SHADER_OPCODE_FIND_LIVE_CHANNEL
||
1093 opcode
== FS_OPCODE_LOAD_LIVE_CHANNELS
) {
1094 return flag_mask(this, 32);
1096 return flag_mask(dst
, size_written
);
1101 * Returns how many MRFs an FS opcode will write over.
1103 * Note that this is not the 0 or 1 implied writes in an actual gen
1104 * instruction -- the FS opcodes often generate MOVs in addition.
1107 fs_inst::implied_mrf_writes() const
1116 case SHADER_OPCODE_RCP
:
1117 case SHADER_OPCODE_RSQ
:
1118 case SHADER_OPCODE_SQRT
:
1119 case SHADER_OPCODE_EXP2
:
1120 case SHADER_OPCODE_LOG2
:
1121 case SHADER_OPCODE_SIN
:
1122 case SHADER_OPCODE_COS
:
1123 return 1 * exec_size
/ 8;
1124 case SHADER_OPCODE_POW
:
1125 case SHADER_OPCODE_INT_QUOTIENT
:
1126 case SHADER_OPCODE_INT_REMAINDER
:
1127 return 2 * exec_size
/ 8;
1128 case SHADER_OPCODE_TEX
:
1130 case SHADER_OPCODE_TXD
:
1131 case SHADER_OPCODE_TXF
:
1132 case SHADER_OPCODE_TXF_CMS
:
1133 case SHADER_OPCODE_TXF_MCS
:
1134 case SHADER_OPCODE_TG4
:
1135 case SHADER_OPCODE_TG4_OFFSET
:
1136 case SHADER_OPCODE_TXL
:
1137 case SHADER_OPCODE_TXS
:
1138 case SHADER_OPCODE_LOD
:
1139 case SHADER_OPCODE_SAMPLEINFO
:
1141 case FS_OPCODE_FB_WRITE
:
1142 case FS_OPCODE_REP_FB_WRITE
:
1143 return src
[0].file
== BAD_FILE
? 0 : 2;
1144 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
1145 case SHADER_OPCODE_GEN4_SCRATCH_READ
:
1147 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN4
:
1149 case SHADER_OPCODE_GEN4_SCRATCH_WRITE
:
1152 unreachable("not reached");
1157 fs_visitor::vgrf(const glsl_type
*const type
)
1159 int reg_width
= dispatch_width
/ 8;
1161 alloc
.allocate(glsl_count_dword_slots(type
, false) * reg_width
),
1162 brw_type_for_base_type(type
));
1165 fs_reg::fs_reg(enum brw_reg_file file
, int nr
)
1170 this->type
= BRW_REGISTER_TYPE_F
;
1171 this->stride
= (file
== UNIFORM
? 0 : 1);
1174 fs_reg::fs_reg(enum brw_reg_file file
, int nr
, enum brw_reg_type type
)
1180 this->stride
= (file
== UNIFORM
? 0 : 1);
1183 /* For SIMD16, we need to follow from the uniform setup of SIMD8 dispatch.
1184 * This brings in those uniform definitions
1187 fs_visitor::import_uniforms(fs_visitor
*v
)
1189 this->push_constant_loc
= v
->push_constant_loc
;
1190 this->pull_constant_loc
= v
->pull_constant_loc
;
1191 this->uniforms
= v
->uniforms
;
1192 this->subgroup_id
= v
->subgroup_id
;
1196 fs_visitor::emit_fragcoord_interpolation(fs_reg wpos
)
1198 assert(stage
== MESA_SHADER_FRAGMENT
);
1200 /* gl_FragCoord.x */
1201 bld
.MOV(wpos
, this->pixel_x
);
1202 wpos
= offset(wpos
, bld
, 1);
1204 /* gl_FragCoord.y */
1205 bld
.MOV(wpos
, this->pixel_y
);
1206 wpos
= offset(wpos
, bld
, 1);
1208 /* gl_FragCoord.z */
1209 if (devinfo
->gen
>= 6) {
1210 bld
.MOV(wpos
, fetch_payload_reg(bld
, payload
.source_depth_reg
));
1212 bld
.emit(FS_OPCODE_LINTERP
, wpos
,
1213 this->delta_xy
[BRW_BARYCENTRIC_PERSPECTIVE_PIXEL
],
1214 component(interp_reg(VARYING_SLOT_POS
, 2), 0));
1216 wpos
= offset(wpos
, bld
, 1);
1218 /* gl_FragCoord.w: Already set up in emit_interpolation */
1219 bld
.MOV(wpos
, this->wpos_w
);
1222 enum brw_barycentric_mode
1223 brw_barycentric_mode(enum glsl_interp_mode mode
, nir_intrinsic_op op
)
1225 /* Barycentric modes don't make sense for flat inputs. */
1226 assert(mode
!= INTERP_MODE_FLAT
);
1230 case nir_intrinsic_load_barycentric_pixel
:
1231 case nir_intrinsic_load_barycentric_at_offset
:
1232 bary
= BRW_BARYCENTRIC_PERSPECTIVE_PIXEL
;
1234 case nir_intrinsic_load_barycentric_centroid
:
1235 bary
= BRW_BARYCENTRIC_PERSPECTIVE_CENTROID
;
1237 case nir_intrinsic_load_barycentric_sample
:
1238 case nir_intrinsic_load_barycentric_at_sample
:
1239 bary
= BRW_BARYCENTRIC_PERSPECTIVE_SAMPLE
;
1242 unreachable("invalid intrinsic");
1245 if (mode
== INTERP_MODE_NOPERSPECTIVE
)
1248 return (enum brw_barycentric_mode
) bary
;
1252 * Turn one of the two CENTROID barycentric modes into PIXEL mode.
1254 static enum brw_barycentric_mode
1255 centroid_to_pixel(enum brw_barycentric_mode bary
)
1257 assert(bary
== BRW_BARYCENTRIC_PERSPECTIVE_CENTROID
||
1258 bary
== BRW_BARYCENTRIC_NONPERSPECTIVE_CENTROID
);
1259 return (enum brw_barycentric_mode
) ((unsigned) bary
- 1);
1263 fs_visitor::emit_frontfacing_interpolation()
1265 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::bool_type
));
1267 if (devinfo
->gen
>= 12) {
1268 fs_reg g1
= fs_reg(retype(brw_vec1_grf(1, 1), BRW_REGISTER_TYPE_W
));
1270 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_W
);
1271 bld
.ASR(tmp
, g1
, brw_imm_d(15));
1273 } else if (devinfo
->gen
>= 6) {
1274 /* Bit 15 of g0.0 is 0 if the polygon is front facing. We want to create
1275 * a boolean result from this (~0/true or 0/false).
1277 * We can use the fact that bit 15 is the MSB of g0.0:W to accomplish
1278 * this task in only one instruction:
1279 * - a negation source modifier will flip the bit; and
1280 * - a W -> D type conversion will sign extend the bit into the high
1281 * word of the destination.
1283 * An ASR 15 fills the low word of the destination.
1285 fs_reg g0
= fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_W
));
1288 bld
.ASR(*reg
, g0
, brw_imm_d(15));
1290 /* Bit 31 of g1.6 is 0 if the polygon is front facing. We want to create
1291 * a boolean result from this (1/true or 0/false).
1293 * Like in the above case, since the bit is the MSB of g1.6:UD we can use
1294 * the negation source modifier to flip it. Unfortunately the SHR
1295 * instruction only operates on UD (or D with an abs source modifier)
1296 * sources without negation.
1298 * Instead, use ASR (which will give ~0/true or 0/false).
1300 fs_reg g1_6
= fs_reg(retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_D
));
1303 bld
.ASR(*reg
, g1_6
, brw_imm_d(31));
1310 fs_visitor::compute_sample_position(fs_reg dst
, fs_reg int_sample_pos
)
1312 assert(stage
== MESA_SHADER_FRAGMENT
);
1313 struct brw_wm_prog_data
*wm_prog_data
= brw_wm_prog_data(this->prog_data
);
1314 assert(dst
.type
== BRW_REGISTER_TYPE_F
);
1316 if (wm_prog_data
->persample_dispatch
) {
1317 /* Convert int_sample_pos to floating point */
1318 bld
.MOV(dst
, int_sample_pos
);
1319 /* Scale to the range [0, 1] */
1320 bld
.MUL(dst
, dst
, brw_imm_f(1 / 16.0f
));
1323 /* From ARB_sample_shading specification:
1324 * "When rendering to a non-multisample buffer, or if multisample
1325 * rasterization is disabled, gl_SamplePosition will always be
1328 bld
.MOV(dst
, brw_imm_f(0.5f
));
1333 fs_visitor::emit_samplepos_setup()
1335 assert(devinfo
->gen
>= 6);
1337 const fs_builder abld
= bld
.annotate("compute sample position");
1338 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::vec2_type
));
1340 fs_reg int_sample_x
= vgrf(glsl_type::int_type
);
1341 fs_reg int_sample_y
= vgrf(glsl_type::int_type
);
1343 /* WM will be run in MSDISPMODE_PERSAMPLE. So, only one of SIMD8 or SIMD16
1344 * mode will be enabled.
1346 * From the Ivy Bridge PRM, volume 2 part 1, page 344:
1347 * R31.1:0 Position Offset X/Y for Slot[3:0]
1348 * R31.3:2 Position Offset X/Y for Slot[7:4]
1351 * The X, Y sample positions come in as bytes in thread payload. So, read
1352 * the positions using vstride=16, width=8, hstride=2.
1354 const fs_reg sample_pos_reg
=
1355 fetch_payload_reg(abld
, payload
.sample_pos_reg
, BRW_REGISTER_TYPE_W
);
1357 /* Compute gl_SamplePosition.x */
1358 abld
.MOV(int_sample_x
, subscript(sample_pos_reg
, BRW_REGISTER_TYPE_B
, 0));
1359 compute_sample_position(offset(pos
, abld
, 0), int_sample_x
);
1361 /* Compute gl_SamplePosition.y */
1362 abld
.MOV(int_sample_y
, subscript(sample_pos_reg
, BRW_REGISTER_TYPE_B
, 1));
1363 compute_sample_position(offset(pos
, abld
, 1), int_sample_y
);
1368 fs_visitor::emit_sampleid_setup()
1370 assert(stage
== MESA_SHADER_FRAGMENT
);
1371 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1372 assert(devinfo
->gen
>= 6);
1374 const fs_builder abld
= bld
.annotate("compute sample id");
1375 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::uint_type
));
1377 if (!key
->multisample_fbo
) {
1378 /* As per GL_ARB_sample_shading specification:
1379 * "When rendering to a non-multisample buffer, or if multisample
1380 * rasterization is disabled, gl_SampleID will always be zero."
1382 abld
.MOV(*reg
, brw_imm_d(0));
1383 } else if (devinfo
->gen
>= 8) {
1384 /* Sample ID comes in as 4-bit numbers in g1.0:
1386 * 15:12 Slot 3 SampleID (only used in SIMD16)
1387 * 11:8 Slot 2 SampleID (only used in SIMD16)
1388 * 7:4 Slot 1 SampleID
1389 * 3:0 Slot 0 SampleID
1391 * Each slot corresponds to four channels, so we want to replicate each
1392 * half-byte value to 4 channels in a row:
1394 * dst+0: .7 .6 .5 .4 .3 .2 .1 .0
1395 * 7:4 7:4 7:4 7:4 3:0 3:0 3:0 3:0
1397 * dst+1: .7 .6 .5 .4 .3 .2 .1 .0 (if SIMD16)
1398 * 15:12 15:12 15:12 15:12 11:8 11:8 11:8 11:8
1400 * First, we read g1.0 with a <1,8,0>UB region, causing the first 8
1401 * channels to read the first byte (7:0), and the second group of 8
1402 * channels to read the second byte (15:8). Then, we shift right by
1403 * a vector immediate of <4, 4, 4, 4, 0, 0, 0, 0>, moving the slot 1 / 3
1404 * values into place. Finally, we AND with 0xf to keep the low nibble.
1406 * shr(16) tmp<1>W g1.0<1,8,0>B 0x44440000:V
1407 * and(16) dst<1>D tmp<8,8,1>W 0xf:W
1409 * TODO: These payload bits exist on Gen7 too, but they appear to always
1410 * be zero, so this code fails to work. We should find out why.
1412 const fs_reg tmp
= abld
.vgrf(BRW_REGISTER_TYPE_UW
);
1414 for (unsigned i
= 0; i
< DIV_ROUND_UP(dispatch_width
, 16); i
++) {
1415 const fs_builder hbld
= abld
.group(MIN2(16, dispatch_width
), i
);
1416 hbld
.SHR(offset(tmp
, hbld
, i
),
1417 stride(retype(brw_vec1_grf(1 + i
, 0), BRW_REGISTER_TYPE_UB
),
1419 brw_imm_v(0x44440000));
1422 abld
.AND(*reg
, tmp
, brw_imm_w(0xf));
1424 const fs_reg t1
= component(abld
.vgrf(BRW_REGISTER_TYPE_UD
), 0);
1425 const fs_reg t2
= abld
.vgrf(BRW_REGISTER_TYPE_UW
);
1427 /* The PS will be run in MSDISPMODE_PERSAMPLE. For example with
1428 * 8x multisampling, subspan 0 will represent sample N (where N
1429 * is 0, 2, 4 or 6), subspan 1 will represent sample 1, 3, 5 or
1430 * 7. We can find the value of N by looking at R0.0 bits 7:6
1431 * ("Starting Sample Pair Index (SSPI)") and multiplying by two
1432 * (since samples are always delivered in pairs). That is, we
1433 * compute 2*((R0.0 & 0xc0) >> 6) == (R0.0 & 0xc0) >> 5. Then
1434 * we need to add N to the sequence (0, 0, 0, 0, 1, 1, 1, 1) in
1435 * case of SIMD8 and sequence (0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2,
1436 * 2, 3, 3, 3, 3) in case of SIMD16. We compute this sequence by
1437 * populating a temporary variable with the sequence (0, 1, 2, 3),
1438 * and then reading from it using vstride=1, width=4, hstride=0.
1439 * These computations hold good for 4x multisampling as well.
1441 * For 2x MSAA and SIMD16, we want to use the sequence (0, 1, 0, 1):
1442 * the first four slots are sample 0 of subspan 0; the next four
1443 * are sample 1 of subspan 0; the third group is sample 0 of
1444 * subspan 1, and finally sample 1 of subspan 1.
1447 /* SKL+ has an extra bit for the Starting Sample Pair Index to
1448 * accomodate 16x MSAA.
1450 abld
.exec_all().group(1, 0)
1451 .AND(t1
, fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
)),
1453 abld
.exec_all().group(1, 0).SHR(t1
, t1
, brw_imm_d(5));
1455 /* This works for SIMD8-SIMD16. It also works for SIMD32 but only if we
1456 * can assume 4x MSAA. Disallow it on IVB+
1458 * FINISHME: One day, we could come up with a way to do this that
1459 * actually works on gen7.
1461 if (devinfo
->gen
>= 7)
1462 limit_dispatch_width(16, "gl_SampleId is unsupported in SIMD32 on gen7");
1463 abld
.exec_all().group(8, 0).MOV(t2
, brw_imm_v(0x32103210));
1465 /* This special instruction takes care of setting vstride=1,
1466 * width=4, hstride=0 of t2 during an ADD instruction.
1468 abld
.emit(FS_OPCODE_SET_SAMPLE_ID
, *reg
, t1
, t2
);
1475 fs_visitor::emit_samplemaskin_setup()
1477 assert(stage
== MESA_SHADER_FRAGMENT
);
1478 struct brw_wm_prog_data
*wm_prog_data
= brw_wm_prog_data(this->prog_data
);
1479 assert(devinfo
->gen
>= 6);
1481 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::int_type
));
1483 fs_reg coverage_mask
=
1484 fetch_payload_reg(bld
, payload
.sample_mask_in_reg
, BRW_REGISTER_TYPE_D
);
1486 if (wm_prog_data
->persample_dispatch
) {
1487 /* gl_SampleMaskIn[] comes from two sources: the input coverage mask,
1488 * and a mask representing which sample is being processed by the
1489 * current shader invocation.
1491 * From the OES_sample_variables specification:
1492 * "When per-sample shading is active due to the use of a fragment input
1493 * qualified by "sample" or due to the use of the gl_SampleID or
1494 * gl_SamplePosition variables, only the bit for the current sample is
1495 * set in gl_SampleMaskIn."
1497 const fs_builder abld
= bld
.annotate("compute gl_SampleMaskIn");
1499 if (nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
].file
== BAD_FILE
)
1500 nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
] = *emit_sampleid_setup();
1502 fs_reg one
= vgrf(glsl_type::int_type
);
1503 fs_reg enabled_mask
= vgrf(glsl_type::int_type
);
1504 abld
.MOV(one
, brw_imm_d(1));
1505 abld
.SHL(enabled_mask
, one
, nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
]);
1506 abld
.AND(*reg
, enabled_mask
, coverage_mask
);
1508 /* In per-pixel mode, the coverage mask is sufficient. */
1509 *reg
= coverage_mask
;
1515 fs_visitor::resolve_source_modifiers(const fs_reg
&src
)
1517 if (!src
.abs
&& !src
.negate
)
1520 fs_reg temp
= bld
.vgrf(src
.type
);
1527 fs_visitor::emit_discard_jump()
1529 assert(brw_wm_prog_data(this->prog_data
)->uses_kill
);
1531 /* For performance, after a discard, jump to the end of the
1532 * shader if all relevant channels have been discarded.
1534 fs_inst
*discard_jump
= bld
.emit(FS_OPCODE_DISCARD_JUMP
);
1535 discard_jump
->flag_subreg
= sample_mask_flag_subreg(this);
1537 discard_jump
->predicate
= BRW_PREDICATE_ALIGN1_ANY4H
;
1538 discard_jump
->predicate_inverse
= true;
1542 fs_visitor::emit_gs_thread_end()
1544 assert(stage
== MESA_SHADER_GEOMETRY
);
1546 struct brw_gs_prog_data
*gs_prog_data
= brw_gs_prog_data(prog_data
);
1548 if (gs_compile
->control_data_header_size_bits
> 0) {
1549 emit_gs_control_data_bits(this->final_gs_vertex_count
);
1552 const fs_builder abld
= bld
.annotate("thread end");
1555 if (gs_prog_data
->static_vertex_count
!= -1) {
1556 foreach_in_list_reverse(fs_inst
, prev
, &this->instructions
) {
1557 if (prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8
||
1558 prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
||
1559 prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
||
1560 prev
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
) {
1563 /* Delete now dead instructions. */
1564 foreach_in_list_reverse_safe(exec_node
, dead
, &this->instructions
) {
1570 } else if (prev
->is_control_flow() || prev
->has_side_effects()) {
1574 fs_reg hdr
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1575 abld
.MOV(hdr
, fs_reg(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
)));
1576 inst
= abld
.emit(SHADER_OPCODE_URB_WRITE_SIMD8
, reg_undef
, hdr
);
1579 fs_reg payload
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
1580 fs_reg
*sources
= ralloc_array(mem_ctx
, fs_reg
, 2);
1581 sources
[0] = fs_reg(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
));
1582 sources
[1] = this->final_gs_vertex_count
;
1583 abld
.LOAD_PAYLOAD(payload
, sources
, 2, 2);
1584 inst
= abld
.emit(SHADER_OPCODE_URB_WRITE_SIMD8
, reg_undef
, payload
);
1592 fs_visitor::assign_curb_setup()
1594 unsigned uniform_push_length
= DIV_ROUND_UP(stage_prog_data
->nr_params
, 8);
1596 unsigned ubo_push_length
= 0;
1597 unsigned ubo_push_start
[4];
1598 for (int i
= 0; i
< 4; i
++) {
1599 ubo_push_start
[i
] = 8 * (ubo_push_length
+ uniform_push_length
);
1600 ubo_push_length
+= stage_prog_data
->ubo_ranges
[i
].length
;
1603 prog_data
->curb_read_length
= uniform_push_length
+ ubo_push_length
;
1605 /* Map the offsets in the UNIFORM file to fixed HW regs. */
1606 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1607 for (unsigned int i
= 0; i
< inst
->sources
; i
++) {
1608 if (inst
->src
[i
].file
== UNIFORM
) {
1609 int uniform_nr
= inst
->src
[i
].nr
+ inst
->src
[i
].offset
/ 4;
1611 if (inst
->src
[i
].nr
>= UBO_START
) {
1612 /* constant_nr is in 32-bit units, the rest are in bytes */
1613 constant_nr
= ubo_push_start
[inst
->src
[i
].nr
- UBO_START
] +
1614 inst
->src
[i
].offset
/ 4;
1615 } else if (uniform_nr
>= 0 && uniform_nr
< (int) uniforms
) {
1616 constant_nr
= push_constant_loc
[uniform_nr
];
1618 /* Section 5.11 of the OpenGL 4.1 spec says:
1619 * "Out-of-bounds reads return undefined values, which include
1620 * values from other variables of the active program or zero."
1621 * Just return the first push constant.
1626 struct brw_reg brw_reg
= brw_vec1_grf(payload
.num_regs
+
1629 brw_reg
.abs
= inst
->src
[i
].abs
;
1630 brw_reg
.negate
= inst
->src
[i
].negate
;
1632 assert(inst
->src
[i
].stride
== 0);
1633 inst
->src
[i
] = byte_offset(
1634 retype(brw_reg
, inst
->src
[i
].type
),
1635 inst
->src
[i
].offset
% 4);
1640 /* This may be updated in assign_urb_setup or assign_vs_urb_setup. */
1641 this->first_non_payload_grf
= payload
.num_regs
+ prog_data
->curb_read_length
;
1645 * Build up an array of indices into the urb_setup array that
1646 * references the active entries of the urb_setup array.
1647 * Used to accelerate walking the active entries of the urb_setup array
1651 brw_compute_urb_setup_index(struct brw_wm_prog_data
*wm_prog_data
)
1653 /* Make sure uint8_t is sufficient */
1654 STATIC_ASSERT(VARYING_SLOT_MAX
<= 0xff);
1656 for (uint8_t attr
= 0; attr
< VARYING_SLOT_MAX
; attr
++) {
1657 if (wm_prog_data
->urb_setup
[attr
] >= 0) {
1658 wm_prog_data
->urb_setup_attribs
[index
++] = attr
;
1661 wm_prog_data
->urb_setup_attribs_count
= index
;
1665 calculate_urb_setup(const struct gen_device_info
*devinfo
,
1666 const struct brw_wm_prog_key
*key
,
1667 struct brw_wm_prog_data
*prog_data
,
1668 const nir_shader
*nir
)
1670 memset(prog_data
->urb_setup
, -1,
1671 sizeof(prog_data
->urb_setup
[0]) * VARYING_SLOT_MAX
);
1674 /* Figure out where each of the incoming setup attributes lands. */
1675 if (devinfo
->gen
>= 6) {
1676 if (util_bitcount64(nir
->info
.inputs_read
&
1677 BRW_FS_VARYING_INPUT_MASK
) <= 16) {
1678 /* The SF/SBE pipeline stage can do arbitrary rearrangement of the
1679 * first 16 varying inputs, so we can put them wherever we want.
1680 * Just put them in order.
1682 * This is useful because it means that (a) inputs not used by the
1683 * fragment shader won't take up valuable register space, and (b) we
1684 * won't have to recompile the fragment shader if it gets paired with
1685 * a different vertex (or geometry) shader.
1687 for (unsigned int i
= 0; i
< VARYING_SLOT_MAX
; i
++) {
1688 if (nir
->info
.inputs_read
& BRW_FS_VARYING_INPUT_MASK
&
1689 BITFIELD64_BIT(i
)) {
1690 prog_data
->urb_setup
[i
] = urb_next
++;
1694 /* We have enough input varyings that the SF/SBE pipeline stage can't
1695 * arbitrarily rearrange them to suit our whim; we have to put them
1696 * in an order that matches the output of the previous pipeline stage
1697 * (geometry or vertex shader).
1699 struct brw_vue_map prev_stage_vue_map
;
1700 brw_compute_vue_map(devinfo
, &prev_stage_vue_map
,
1701 key
->input_slots_valid
,
1702 nir
->info
.separate_shader
, 1);
1705 brw_compute_first_urb_slot_required(nir
->info
.inputs_read
,
1706 &prev_stage_vue_map
);
1708 assert(prev_stage_vue_map
.num_slots
<= first_slot
+ 32);
1709 for (int slot
= first_slot
; slot
< prev_stage_vue_map
.num_slots
;
1711 int varying
= prev_stage_vue_map
.slot_to_varying
[slot
];
1712 if (varying
!= BRW_VARYING_SLOT_PAD
&&
1713 (nir
->info
.inputs_read
& BRW_FS_VARYING_INPUT_MASK
&
1714 BITFIELD64_BIT(varying
))) {
1715 prog_data
->urb_setup
[varying
] = slot
- first_slot
;
1718 urb_next
= prev_stage_vue_map
.num_slots
- first_slot
;
1721 /* FINISHME: The sf doesn't map VS->FS inputs for us very well. */
1722 for (unsigned int i
= 0; i
< VARYING_SLOT_MAX
; i
++) {
1723 /* Point size is packed into the header, not as a general attribute */
1724 if (i
== VARYING_SLOT_PSIZ
)
1727 if (key
->input_slots_valid
& BITFIELD64_BIT(i
)) {
1728 /* The back color slot is skipped when the front color is
1729 * also written to. In addition, some slots can be
1730 * written in the vertex shader and not read in the
1731 * fragment shader. So the register number must always be
1732 * incremented, mapped or not.
1734 if (_mesa_varying_slot_in_fs((gl_varying_slot
) i
))
1735 prog_data
->urb_setup
[i
] = urb_next
;
1741 * It's a FS only attribute, and we did interpolation for this attribute
1742 * in SF thread. So, count it here, too.
1744 * See compile_sf_prog() for more info.
1746 if (nir
->info
.inputs_read
& BITFIELD64_BIT(VARYING_SLOT_PNTC
))
1747 prog_data
->urb_setup
[VARYING_SLOT_PNTC
] = urb_next
++;
1750 prog_data
->num_varying_inputs
= urb_next
;
1751 prog_data
->inputs
= nir
->info
.inputs_read
;
1753 brw_compute_urb_setup_index(prog_data
);
1757 fs_visitor::assign_urb_setup()
1759 assert(stage
== MESA_SHADER_FRAGMENT
);
1760 struct brw_wm_prog_data
*prog_data
= brw_wm_prog_data(this->prog_data
);
1762 int urb_start
= payload
.num_regs
+ prog_data
->base
.curb_read_length
;
1764 /* Offset all the urb_setup[] index by the actual position of the
1765 * setup regs, now that the location of the constants has been chosen.
1767 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1768 for (int i
= 0; i
< inst
->sources
; i
++) {
1769 if (inst
->src
[i
].file
== ATTR
) {
1770 /* ATTR regs in the FS are in units of logical scalar inputs each
1771 * of which consumes half of a GRF register.
1773 assert(inst
->src
[i
].offset
< REG_SIZE
/ 2);
1774 const unsigned grf
= urb_start
+ inst
->src
[i
].nr
/ 2;
1775 const unsigned offset
= (inst
->src
[i
].nr
% 2) * (REG_SIZE
/ 2) +
1776 inst
->src
[i
].offset
;
1777 const unsigned width
= inst
->src
[i
].stride
== 0 ?
1778 1 : MIN2(inst
->exec_size
, 8);
1779 struct brw_reg reg
= stride(
1780 byte_offset(retype(brw_vec8_grf(grf
, 0), inst
->src
[i
].type
),
1782 width
* inst
->src
[i
].stride
,
1783 width
, inst
->src
[i
].stride
);
1784 reg
.abs
= inst
->src
[i
].abs
;
1785 reg
.negate
= inst
->src
[i
].negate
;
1791 /* Each attribute is 4 setup channels, each of which is half a reg. */
1792 this->first_non_payload_grf
+= prog_data
->num_varying_inputs
* 2;
1796 fs_visitor::convert_attr_sources_to_hw_regs(fs_inst
*inst
)
1798 for (int i
= 0; i
< inst
->sources
; i
++) {
1799 if (inst
->src
[i
].file
== ATTR
) {
1800 int grf
= payload
.num_regs
+
1801 prog_data
->curb_read_length
+
1803 inst
->src
[i
].offset
/ REG_SIZE
;
1805 /* As explained at brw_reg_from_fs_reg, From the Haswell PRM:
1807 * VertStride must be used to cross GRF register boundaries. This
1808 * rule implies that elements within a 'Width' cannot cross GRF
1811 * So, for registers that are large enough, we have to split the exec
1812 * size in two and trust the compression state to sort it out.
1814 unsigned total_size
= inst
->exec_size
*
1815 inst
->src
[i
].stride
*
1816 type_sz(inst
->src
[i
].type
);
1818 assert(total_size
<= 2 * REG_SIZE
);
1819 const unsigned exec_size
=
1820 (total_size
<= REG_SIZE
) ? inst
->exec_size
: inst
->exec_size
/ 2;
1822 unsigned width
= inst
->src
[i
].stride
== 0 ? 1 : exec_size
;
1823 struct brw_reg reg
=
1824 stride(byte_offset(retype(brw_vec8_grf(grf
, 0), inst
->src
[i
].type
),
1825 inst
->src
[i
].offset
% REG_SIZE
),
1826 exec_size
* inst
->src
[i
].stride
,
1827 width
, inst
->src
[i
].stride
);
1828 reg
.abs
= inst
->src
[i
].abs
;
1829 reg
.negate
= inst
->src
[i
].negate
;
1837 fs_visitor::assign_vs_urb_setup()
1839 struct brw_vs_prog_data
*vs_prog_data
= brw_vs_prog_data(prog_data
);
1841 assert(stage
== MESA_SHADER_VERTEX
);
1843 /* Each attribute is 4 regs. */
1844 this->first_non_payload_grf
+= 4 * vs_prog_data
->nr_attribute_slots
;
1846 assert(vs_prog_data
->base
.urb_read_length
<= 15);
1848 /* Rewrite all ATTR file references to the hw grf that they land in. */
1849 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1850 convert_attr_sources_to_hw_regs(inst
);
1855 fs_visitor::assign_tcs_urb_setup()
1857 assert(stage
== MESA_SHADER_TESS_CTRL
);
1859 /* Rewrite all ATTR file references to HW_REGs. */
1860 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1861 convert_attr_sources_to_hw_regs(inst
);
1866 fs_visitor::assign_tes_urb_setup()
1868 assert(stage
== MESA_SHADER_TESS_EVAL
);
1870 struct brw_vue_prog_data
*vue_prog_data
= brw_vue_prog_data(prog_data
);
1872 first_non_payload_grf
+= 8 * vue_prog_data
->urb_read_length
;
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_gs_urb_setup()
1883 assert(stage
== MESA_SHADER_GEOMETRY
);
1885 struct brw_vue_prog_data
*vue_prog_data
= brw_vue_prog_data(prog_data
);
1887 first_non_payload_grf
+=
1888 8 * vue_prog_data
->urb_read_length
* nir
->info
.gs
.vertices_in
;
1890 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1891 /* Rewrite all ATTR file references to GRFs. */
1892 convert_attr_sources_to_hw_regs(inst
);
1898 * Split large virtual GRFs into separate components if we can.
1900 * This is mostly duplicated with what brw_fs_vector_splitting does,
1901 * but that's really conservative because it's afraid of doing
1902 * splitting that doesn't result in real progress after the rest of
1903 * the optimization phases, which would cause infinite looping in
1904 * optimization. We can do it once here, safely. This also has the
1905 * opportunity to split interpolated values, or maybe even uniforms,
1906 * which we don't have at the IR level.
1908 * We want to split, because virtual GRFs are what we register
1909 * allocate and spill (due to contiguousness requirements for some
1910 * instructions), and they're what we naturally generate in the
1911 * codegen process, but most virtual GRFs don't actually need to be
1912 * contiguous sets of GRFs. If we split, we'll end up with reduced
1913 * live intervals and better dead code elimination and coalescing.
1916 fs_visitor::split_virtual_grfs()
1918 /* Compact the register file so we eliminate dead vgrfs. This
1919 * only defines split points for live registers, so if we have
1920 * too large dead registers they will hit assertions later.
1922 compact_virtual_grfs();
1924 int num_vars
= this->alloc
.count
;
1926 /* Count the total number of registers */
1928 int vgrf_to_reg
[num_vars
];
1929 for (int i
= 0; i
< num_vars
; i
++) {
1930 vgrf_to_reg
[i
] = reg_count
;
1931 reg_count
+= alloc
.sizes
[i
];
1934 /* An array of "split points". For each register slot, this indicates
1935 * if this slot can be separated from the previous slot. Every time an
1936 * instruction uses multiple elements of a register (as a source or
1937 * destination), we mark the used slots as inseparable. Then we go
1938 * through and split the registers into the smallest pieces we can.
1940 bool *split_points
= new bool[reg_count
];
1941 memset(split_points
, 0, reg_count
* sizeof(*split_points
));
1943 /* Mark all used registers as fully splittable */
1944 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1945 if (inst
->dst
.file
== VGRF
) {
1946 int reg
= vgrf_to_reg
[inst
->dst
.nr
];
1947 for (unsigned j
= 1; j
< this->alloc
.sizes
[inst
->dst
.nr
]; j
++)
1948 split_points
[reg
+ j
] = true;
1951 for (int i
= 0; i
< inst
->sources
; i
++) {
1952 if (inst
->src
[i
].file
== VGRF
) {
1953 int reg
= vgrf_to_reg
[inst
->src
[i
].nr
];
1954 for (unsigned j
= 1; j
< this->alloc
.sizes
[inst
->src
[i
].nr
]; j
++)
1955 split_points
[reg
+ j
] = true;
1960 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1961 /* We fix up undef instructions later */
1962 if (inst
->opcode
== SHADER_OPCODE_UNDEF
) {
1963 /* UNDEF instructions are currently only used to undef entire
1964 * registers. We need this invariant later when we split them.
1966 assert(inst
->dst
.file
== VGRF
);
1967 assert(inst
->dst
.offset
== 0);
1968 assert(inst
->size_written
== alloc
.sizes
[inst
->dst
.nr
] * REG_SIZE
);
1972 if (inst
->dst
.file
== VGRF
) {
1973 int reg
= vgrf_to_reg
[inst
->dst
.nr
] + inst
->dst
.offset
/ REG_SIZE
;
1974 for (unsigned j
= 1; j
< regs_written(inst
); j
++)
1975 split_points
[reg
+ j
] = false;
1977 for (int i
= 0; i
< inst
->sources
; i
++) {
1978 if (inst
->src
[i
].file
== VGRF
) {
1979 int reg
= vgrf_to_reg
[inst
->src
[i
].nr
] + inst
->src
[i
].offset
/ REG_SIZE
;
1980 for (unsigned j
= 1; j
< regs_read(inst
, i
); j
++)
1981 split_points
[reg
+ j
] = false;
1986 int *new_virtual_grf
= new int[reg_count
];
1987 int *new_reg_offset
= new int[reg_count
];
1990 for (int i
= 0; i
< num_vars
; i
++) {
1991 /* The first one should always be 0 as a quick sanity check. */
1992 assert(split_points
[reg
] == false);
1995 new_reg_offset
[reg
] = 0;
2000 for (unsigned j
= 1; j
< alloc
.sizes
[i
]; j
++) {
2001 /* If this is a split point, reset the offset to 0 and allocate a
2002 * new virtual GRF for the previous offset many registers
2004 if (split_points
[reg
]) {
2005 assert(offset
<= MAX_VGRF_SIZE
);
2006 int grf
= alloc
.allocate(offset
);
2007 for (int k
= reg
- offset
; k
< reg
; k
++)
2008 new_virtual_grf
[k
] = grf
;
2011 new_reg_offset
[reg
] = offset
;
2016 /* The last one gets the original register number */
2017 assert(offset
<= MAX_VGRF_SIZE
);
2018 alloc
.sizes
[i
] = offset
;
2019 for (int k
= reg
- offset
; k
< reg
; k
++)
2020 new_virtual_grf
[k
] = i
;
2022 assert(reg
== reg_count
);
2024 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
2025 if (inst
->opcode
== SHADER_OPCODE_UNDEF
) {
2026 const fs_builder
ibld(this, block
, inst
);
2027 assert(inst
->size_written
% REG_SIZE
== 0);
2028 unsigned reg_offset
= 0;
2029 while (reg_offset
< inst
->size_written
/ REG_SIZE
) {
2030 reg
= vgrf_to_reg
[inst
->dst
.nr
] + reg_offset
;
2031 ibld
.UNDEF(fs_reg(VGRF
, new_virtual_grf
[reg
], inst
->dst
.type
));
2032 reg_offset
+= alloc
.sizes
[new_virtual_grf
[reg
]];
2034 inst
->remove(block
);
2038 if (inst
->dst
.file
== VGRF
) {
2039 reg
= vgrf_to_reg
[inst
->dst
.nr
] + inst
->dst
.offset
/ REG_SIZE
;
2040 inst
->dst
.nr
= new_virtual_grf
[reg
];
2041 inst
->dst
.offset
= new_reg_offset
[reg
] * REG_SIZE
+
2042 inst
->dst
.offset
% REG_SIZE
;
2043 assert((unsigned)new_reg_offset
[reg
] < alloc
.sizes
[new_virtual_grf
[reg
]]);
2045 for (int i
= 0; i
< inst
->sources
; i
++) {
2046 if (inst
->src
[i
].file
== VGRF
) {
2047 reg
= vgrf_to_reg
[inst
->src
[i
].nr
] + inst
->src
[i
].offset
/ REG_SIZE
;
2048 inst
->src
[i
].nr
= new_virtual_grf
[reg
];
2049 inst
->src
[i
].offset
= new_reg_offset
[reg
] * REG_SIZE
+
2050 inst
->src
[i
].offset
% REG_SIZE
;
2051 assert((unsigned)new_reg_offset
[reg
] < alloc
.sizes
[new_virtual_grf
[reg
]]);
2055 invalidate_analysis(DEPENDENCY_INSTRUCTION_DETAIL
| DEPENDENCY_VARIABLES
);
2057 delete[] split_points
;
2058 delete[] new_virtual_grf
;
2059 delete[] new_reg_offset
;
2063 * Remove unused virtual GRFs and compact the vgrf_* arrays.
2065 * During code generation, we create tons of temporary variables, many of
2066 * which get immediately killed and are never used again. Yet, in later
2067 * optimization and analysis passes, such as compute_live_intervals, we need
2068 * to loop over all the virtual GRFs. Compacting them can save a lot of
2072 fs_visitor::compact_virtual_grfs()
2074 bool progress
= false;
2075 int *remap_table
= new int[this->alloc
.count
];
2076 memset(remap_table
, -1, this->alloc
.count
* sizeof(int));
2078 /* Mark which virtual GRFs are used. */
2079 foreach_block_and_inst(block
, const fs_inst
, inst
, cfg
) {
2080 if (inst
->dst
.file
== VGRF
)
2081 remap_table
[inst
->dst
.nr
] = 0;
2083 for (int i
= 0; i
< inst
->sources
; i
++) {
2084 if (inst
->src
[i
].file
== VGRF
)
2085 remap_table
[inst
->src
[i
].nr
] = 0;
2089 /* Compact the GRF arrays. */
2091 for (unsigned i
= 0; i
< this->alloc
.count
; i
++) {
2092 if (remap_table
[i
] == -1) {
2093 /* We just found an unused register. This means that we are
2094 * actually going to compact something.
2098 remap_table
[i
] = new_index
;
2099 alloc
.sizes
[new_index
] = alloc
.sizes
[i
];
2100 invalidate_analysis(DEPENDENCY_INSTRUCTION_DETAIL
| DEPENDENCY_VARIABLES
);
2105 this->alloc
.count
= new_index
;
2107 /* Patch all the instructions to use the newly renumbered registers */
2108 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2109 if (inst
->dst
.file
== VGRF
)
2110 inst
->dst
.nr
= remap_table
[inst
->dst
.nr
];
2112 for (int i
= 0; i
< inst
->sources
; i
++) {
2113 if (inst
->src
[i
].file
== VGRF
)
2114 inst
->src
[i
].nr
= remap_table
[inst
->src
[i
].nr
];
2118 /* Patch all the references to delta_xy, since they're used in register
2119 * allocation. If they're unused, switch them to BAD_FILE so we don't
2120 * think some random VGRF is delta_xy.
2122 for (unsigned i
= 0; i
< ARRAY_SIZE(delta_xy
); i
++) {
2123 if (delta_xy
[i
].file
== VGRF
) {
2124 if (remap_table
[delta_xy
[i
].nr
] != -1) {
2125 delta_xy
[i
].nr
= remap_table
[delta_xy
[i
].nr
];
2127 delta_xy
[i
].file
= BAD_FILE
;
2132 delete[] remap_table
;
2138 get_subgroup_id_param_index(const brw_stage_prog_data
*prog_data
)
2140 if (prog_data
->nr_params
== 0)
2143 /* The local thread id is always the last parameter in the list */
2144 uint32_t last_param
= prog_data
->param
[prog_data
->nr_params
- 1];
2145 if (last_param
== BRW_PARAM_BUILTIN_SUBGROUP_ID
)
2146 return prog_data
->nr_params
- 1;
2152 * Struct for handling complex alignments.
2154 * A complex alignment is stored as multiplier and an offset. A value is
2155 * considered to be aligned if it is {offset} larger than a multiple of {mul}.
2156 * For instance, with an alignment of {8, 2}, cplx_align_apply would do the
2159 * N | cplx_align_apply({8, 2}, N)
2160 * ----+-----------------------------
2174 #define CPLX_ALIGN_MAX_MUL 8
2177 cplx_align_assert_sane(struct cplx_align a
)
2179 assert(a
.mul
> 0 && util_is_power_of_two_nonzero(a
.mul
));
2180 assert(a
.offset
< a
.mul
);
2184 * Combines two alignments to produce a least multiple of sorts.
2186 * The returned alignment is the smallest (in terms of multiplier) such that
2187 * anything aligned to both a and b will be aligned to the new alignment.
2188 * This function will assert-fail if a and b are not compatible, i.e. if the
2189 * offset parameters are such that no common alignment is possible.
2191 static struct cplx_align
2192 cplx_align_combine(struct cplx_align a
, struct cplx_align b
)
2194 cplx_align_assert_sane(a
);
2195 cplx_align_assert_sane(b
);
2197 /* Assert that the alignments agree. */
2198 assert((a
.offset
& (b
.mul
- 1)) == (b
.offset
& (a
.mul
- 1)));
2200 return a
.mul
> b
.mul
? a
: b
;
2204 * Apply a complex alignment
2206 * This function will return the smallest number greater than or equal to
2207 * offset that is aligned to align.
2210 cplx_align_apply(struct cplx_align align
, unsigned offset
)
2212 return ALIGN(offset
- align
.offset
, align
.mul
) + align
.offset
;
2215 #define UNIFORM_SLOT_SIZE 4
2217 struct uniform_slot_info
{
2218 /** True if the given uniform slot is live */
2221 /** True if this slot and the next slot must remain contiguous */
2222 unsigned contiguous
:1;
2224 struct cplx_align align
;
2228 mark_uniform_slots_read(struct uniform_slot_info
*slots
,
2229 unsigned num_slots
, unsigned alignment
)
2231 assert(alignment
> 0 && util_is_power_of_two_nonzero(alignment
));
2232 assert(alignment
<= CPLX_ALIGN_MAX_MUL
);
2234 /* We can't align a slot to anything less than the slot size */
2235 alignment
= MAX2(alignment
, UNIFORM_SLOT_SIZE
);
2237 struct cplx_align align
= {alignment
, 0};
2238 cplx_align_assert_sane(align
);
2240 for (unsigned i
= 0; i
< num_slots
; i
++) {
2241 slots
[i
].is_live
= true;
2242 if (i
< num_slots
- 1)
2243 slots
[i
].contiguous
= true;
2245 align
.offset
= (i
* UNIFORM_SLOT_SIZE
) & (align
.mul
- 1);
2246 if (slots
[i
].align
.mul
== 0) {
2247 slots
[i
].align
= align
;
2249 slots
[i
].align
= cplx_align_combine(slots
[i
].align
, align
);
2255 * Assign UNIFORM file registers to either push constants or pull constants.
2257 * We allow a fragment shader to have more than the specified minimum
2258 * maximum number of fragment shader uniform components (64). If
2259 * there are too many of these, they'd fill up all of register space.
2260 * So, this will push some of them out to the pull constant buffer and
2261 * update the program to load them.
2264 fs_visitor::assign_constant_locations()
2266 /* Only the first compile gets to decide on locations. */
2267 if (push_constant_loc
) {
2268 assert(pull_constant_loc
);
2272 if (compiler
->compact_params
) {
2273 struct uniform_slot_info slots
[uniforms
+ 1];
2274 memset(slots
, 0, sizeof(slots
));
2276 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
2277 for (int i
= 0 ; i
< inst
->sources
; i
++) {
2278 if (inst
->src
[i
].file
!= UNIFORM
)
2281 /* NIR tightly packs things so the uniform number might not be
2282 * aligned (if we have a double right after a float, for
2283 * instance). This is fine because the process of re-arranging
2284 * them will ensure that things are properly aligned. The offset
2285 * into that uniform, however, must be aligned.
2287 * In Vulkan, we have explicit offsets but everything is crammed
2288 * into a single "variable" so inst->src[i].nr will always be 0.
2289 * Everything will be properly aligned relative to that one base.
2291 assert(inst
->src
[i
].offset
% type_sz(inst
->src
[i
].type
) == 0);
2293 unsigned u
= inst
->src
[i
].nr
+
2294 inst
->src
[i
].offset
/ UNIFORM_SLOT_SIZE
;
2299 unsigned slots_read
;
2300 if (inst
->opcode
== SHADER_OPCODE_MOV_INDIRECT
&& i
== 0) {
2301 slots_read
= DIV_ROUND_UP(inst
->src
[2].ud
, UNIFORM_SLOT_SIZE
);
2303 unsigned bytes_read
= inst
->components_read(i
) *
2304 type_sz(inst
->src
[i
].type
);
2305 slots_read
= DIV_ROUND_UP(bytes_read
, UNIFORM_SLOT_SIZE
);
2308 assert(u
+ slots_read
<= uniforms
);
2309 mark_uniform_slots_read(&slots
[u
], slots_read
,
2310 type_sz(inst
->src
[i
].type
));
2314 int subgroup_id_index
= get_subgroup_id_param_index(stage_prog_data
);
2316 /* Only allow 16 registers (128 uniform components) as push constants.
2318 * Just demote the end of the list. We could probably do better
2319 * here, demoting things that are rarely used in the program first.
2321 * If changing this value, note the limitation about total_regs in
2324 unsigned int max_push_components
= 16 * 8;
2325 if (subgroup_id_index
>= 0)
2326 max_push_components
--; /* Save a slot for the thread ID */
2328 /* We push small arrays, but no bigger than 16 floats. This is big
2329 * enough for a vec4 but hopefully not large enough to push out other
2330 * stuff. We should probably use a better heuristic at some point.
2332 const unsigned int max_chunk_size
= 16;
2334 unsigned int num_push_constants
= 0;
2335 unsigned int num_pull_constants
= 0;
2337 push_constant_loc
= ralloc_array(mem_ctx
, int, uniforms
);
2338 pull_constant_loc
= ralloc_array(mem_ctx
, int, uniforms
);
2340 /* Default to -1 meaning no location */
2341 memset(push_constant_loc
, -1, uniforms
* sizeof(*push_constant_loc
));
2342 memset(pull_constant_loc
, -1, uniforms
* sizeof(*pull_constant_loc
));
2344 int chunk_start
= -1;
2345 struct cplx_align align
;
2346 for (unsigned u
= 0; u
< uniforms
; u
++) {
2347 if (!slots
[u
].is_live
) {
2348 assert(chunk_start
== -1);
2352 /* Skip subgroup_id_index to put it in the last push register. */
2353 if (subgroup_id_index
== (int)u
)
2356 if (chunk_start
== -1) {
2358 align
= slots
[u
].align
;
2360 /* Offset into the chunk */
2361 unsigned chunk_offset
= (u
- chunk_start
) * UNIFORM_SLOT_SIZE
;
2363 /* Shift the slot alignment down by the chunk offset so it is
2364 * comparable with the base chunk alignment.
2366 struct cplx_align slot_align
= slots
[u
].align
;
2368 (slot_align
.offset
- chunk_offset
) & (align
.mul
- 1);
2370 align
= cplx_align_combine(align
, slot_align
);
2373 /* Sanity check the alignment */
2374 cplx_align_assert_sane(align
);
2376 if (slots
[u
].contiguous
)
2379 /* Adjust the alignment to be in terms of slots, not bytes */
2380 assert((align
.mul
& (UNIFORM_SLOT_SIZE
- 1)) == 0);
2381 assert((align
.offset
& (UNIFORM_SLOT_SIZE
- 1)) == 0);
2382 align
.mul
/= UNIFORM_SLOT_SIZE
;
2383 align
.offset
/= UNIFORM_SLOT_SIZE
;
2385 unsigned push_start_align
= cplx_align_apply(align
, num_push_constants
);
2386 unsigned chunk_size
= u
- chunk_start
+ 1;
2387 if ((!compiler
->supports_pull_constants
&& u
< UBO_START
) ||
2388 (chunk_size
< max_chunk_size
&&
2389 push_start_align
+ chunk_size
<= max_push_components
)) {
2390 /* Align up the number of push constants */
2391 num_push_constants
= push_start_align
;
2392 for (unsigned i
= 0; i
< chunk_size
; i
++)
2393 push_constant_loc
[chunk_start
+ i
] = num_push_constants
++;
2395 /* We need to pull this one */
2396 num_pull_constants
= cplx_align_apply(align
, num_pull_constants
);
2397 for (unsigned i
= 0; i
< chunk_size
; i
++)
2398 pull_constant_loc
[chunk_start
+ i
] = num_pull_constants
++;
2401 /* Reset the chunk and start again */
2405 /* Add the CS local thread ID uniform at the end of the push constants */
2406 if (subgroup_id_index
>= 0)
2407 push_constant_loc
[subgroup_id_index
] = num_push_constants
++;
2409 /* As the uniforms are going to be reordered, stash the old array and
2410 * create two new arrays for push/pull params.
2412 uint32_t *param
= stage_prog_data
->param
;
2413 stage_prog_data
->nr_params
= num_push_constants
;
2414 if (num_push_constants
) {
2415 stage_prog_data
->param
= rzalloc_array(mem_ctx
, uint32_t,
2416 num_push_constants
);
2418 stage_prog_data
->param
= NULL
;
2420 assert(stage_prog_data
->nr_pull_params
== 0);
2421 assert(stage_prog_data
->pull_param
== NULL
);
2422 if (num_pull_constants
> 0) {
2423 stage_prog_data
->nr_pull_params
= num_pull_constants
;
2424 stage_prog_data
->pull_param
= rzalloc_array(mem_ctx
, uint32_t,
2425 num_pull_constants
);
2428 /* Up until now, the param[] array has been indexed by reg + offset
2429 * of UNIFORM registers. Move pull constants into pull_param[] and
2430 * condense param[] to only contain the uniforms we chose to push.
2432 * NOTE: Because we are condensing the params[] array, we know that
2433 * push_constant_loc[i] <= i and we can do it in one smooth loop without
2434 * having to make a copy.
2436 for (unsigned int i
= 0; i
< uniforms
; i
++) {
2437 uint32_t value
= param
[i
];
2438 if (pull_constant_loc
[i
] != -1) {
2439 stage_prog_data
->pull_param
[pull_constant_loc
[i
]] = value
;
2440 } else if (push_constant_loc
[i
] != -1) {
2441 stage_prog_data
->param
[push_constant_loc
[i
]] = value
;
2446 /* If we don't want to compact anything, just set up dummy push/pull
2447 * arrays. All the rest of the compiler cares about are these arrays.
2449 push_constant_loc
= ralloc_array(mem_ctx
, int, uniforms
);
2450 pull_constant_loc
= ralloc_array(mem_ctx
, int, uniforms
);
2452 for (unsigned u
= 0; u
< uniforms
; u
++)
2453 push_constant_loc
[u
] = u
;
2455 memset(pull_constant_loc
, -1, uniforms
* sizeof(*pull_constant_loc
));
2458 /* Now that we know how many regular uniforms we'll push, reduce the
2459 * UBO push ranges so we don't exceed the 3DSTATE_CONSTANT limits.
2461 unsigned push_length
= DIV_ROUND_UP(stage_prog_data
->nr_params
, 8);
2462 for (int i
= 0; i
< 4; i
++) {
2463 struct brw_ubo_range
*range
= &prog_data
->ubo_ranges
[i
];
2465 if (push_length
+ range
->length
> 64)
2466 range
->length
= 64 - push_length
;
2468 push_length
+= range
->length
;
2470 assert(push_length
<= 64);
2474 fs_visitor::get_pull_locs(const fs_reg
&src
,
2475 unsigned *out_surf_index
,
2476 unsigned *out_pull_index
)
2478 assert(src
.file
== UNIFORM
);
2480 if (src
.nr
>= UBO_START
) {
2481 const struct brw_ubo_range
*range
=
2482 &prog_data
->ubo_ranges
[src
.nr
- UBO_START
];
2484 /* If this access is in our (reduced) range, use the push data. */
2485 if (src
.offset
/ 32 < range
->length
)
2488 *out_surf_index
= prog_data
->binding_table
.ubo_start
+ range
->block
;
2489 *out_pull_index
= (32 * range
->start
+ src
.offset
) / 4;
2491 prog_data
->has_ubo_pull
= true;
2495 const unsigned location
= src
.nr
+ src
.offset
/ 4;
2497 if (location
< uniforms
&& pull_constant_loc
[location
] != -1) {
2498 /* A regular uniform push constant */
2499 *out_surf_index
= stage_prog_data
->binding_table
.pull_constants_start
;
2500 *out_pull_index
= pull_constant_loc
[location
];
2502 prog_data
->has_ubo_pull
= true;
2510 * Replace UNIFORM register file access with either UNIFORM_PULL_CONSTANT_LOAD
2511 * or VARYING_PULL_CONSTANT_LOAD instructions which load values into VGRFs.
2514 fs_visitor::lower_constant_loads()
2516 unsigned index
, pull_index
;
2518 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
2519 /* Set up the annotation tracking for new generated instructions. */
2520 const fs_builder
ibld(this, block
, inst
);
2522 for (int i
= 0; i
< inst
->sources
; i
++) {
2523 if (inst
->src
[i
].file
!= UNIFORM
)
2526 /* We'll handle this case later */
2527 if (inst
->opcode
== SHADER_OPCODE_MOV_INDIRECT
&& i
== 0)
2530 if (!get_pull_locs(inst
->src
[i
], &index
, &pull_index
))
2533 assert(inst
->src
[i
].stride
== 0);
2535 const unsigned block_sz
= 64; /* Fetch one cacheline at a time. */
2536 const fs_builder ubld
= ibld
.exec_all().group(block_sz
/ 4, 0);
2537 const fs_reg dst
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
2538 const unsigned base
= pull_index
* 4;
2540 ubld
.emit(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
,
2541 dst
, brw_imm_ud(index
), brw_imm_ud(base
& ~(block_sz
- 1)));
2543 /* Rewrite the instruction to use the temporary VGRF. */
2544 inst
->src
[i
].file
= VGRF
;
2545 inst
->src
[i
].nr
= dst
.nr
;
2546 inst
->src
[i
].offset
= (base
& (block_sz
- 1)) +
2547 inst
->src
[i
].offset
% 4;
2550 if (inst
->opcode
== SHADER_OPCODE_MOV_INDIRECT
&&
2551 inst
->src
[0].file
== UNIFORM
) {
2553 if (!get_pull_locs(inst
->src
[0], &index
, &pull_index
))
2556 VARYING_PULL_CONSTANT_LOAD(ibld
, inst
->dst
,
2560 inst
->remove(block
);
2563 invalidate_analysis(DEPENDENCY_INSTRUCTIONS
);
2567 fs_visitor::opt_algebraic()
2569 bool progress
= false;
2571 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
2572 switch (inst
->opcode
) {
2573 case BRW_OPCODE_MOV
:
2574 if (!devinfo
->has_64bit_float
&&
2575 !devinfo
->has_64bit_int
&&
2576 (inst
->dst
.type
== BRW_REGISTER_TYPE_DF
||
2577 inst
->dst
.type
== BRW_REGISTER_TYPE_UQ
||
2578 inst
->dst
.type
== BRW_REGISTER_TYPE_Q
)) {
2579 assert(inst
->dst
.type
== inst
->src
[0].type
);
2580 assert(!inst
->saturate
);
2581 assert(!inst
->src
[0].abs
);
2582 assert(!inst
->src
[0].negate
);
2583 const brw::fs_builder
ibld(this, block
, inst
);
2585 if (inst
->src
[0].file
== IMM
) {
2586 ibld
.MOV(subscript(inst
->dst
, BRW_REGISTER_TYPE_UD
, 1),
2587 brw_imm_ud(inst
->src
[0].u64
>> 32));
2588 ibld
.MOV(subscript(inst
->dst
, BRW_REGISTER_TYPE_UD
, 0),
2589 brw_imm_ud(inst
->src
[0].u64
));
2591 ibld
.MOV(subscript(inst
->dst
, BRW_REGISTER_TYPE_UD
, 1),
2592 subscript(inst
->src
[0], BRW_REGISTER_TYPE_UD
, 1));
2593 ibld
.MOV(subscript(inst
->dst
, BRW_REGISTER_TYPE_UD
, 0),
2594 subscript(inst
->src
[0], BRW_REGISTER_TYPE_UD
, 0));
2597 inst
->remove(block
);
2601 if ((inst
->conditional_mod
== BRW_CONDITIONAL_Z
||
2602 inst
->conditional_mod
== BRW_CONDITIONAL_NZ
) &&
2603 inst
->dst
.is_null() &&
2604 (inst
->src
[0].abs
|| inst
->src
[0].negate
)) {
2605 inst
->src
[0].abs
= false;
2606 inst
->src
[0].negate
= false;
2611 if (inst
->src
[0].file
!= IMM
)
2614 if (inst
->saturate
) {
2615 /* Full mixed-type saturates don't happen. However, we can end up
2618 * mov.sat(8) g21<1>DF -1F
2620 * Other mixed-size-but-same-base-type cases may also be possible.
2622 if (inst
->dst
.type
!= inst
->src
[0].type
&&
2623 inst
->dst
.type
!= BRW_REGISTER_TYPE_DF
&&
2624 inst
->src
[0].type
!= BRW_REGISTER_TYPE_F
)
2625 assert(!"unimplemented: saturate mixed types");
2627 if (brw_saturate_immediate(inst
->src
[0].type
,
2628 &inst
->src
[0].as_brw_reg())) {
2629 inst
->saturate
= false;
2635 case BRW_OPCODE_MUL
:
2636 if (inst
->src
[1].file
!= IMM
)
2640 if (inst
->src
[1].is_one()) {
2641 inst
->opcode
= BRW_OPCODE_MOV
;
2642 inst
->src
[1] = reg_undef
;
2648 if (inst
->src
[1].is_negative_one()) {
2649 inst
->opcode
= BRW_OPCODE_MOV
;
2650 inst
->src
[0].negate
= !inst
->src
[0].negate
;
2651 inst
->src
[1] = reg_undef
;
2656 if (inst
->src
[0].file
== IMM
) {
2657 assert(inst
->src
[0].type
== BRW_REGISTER_TYPE_F
);
2658 inst
->opcode
= BRW_OPCODE_MOV
;
2659 inst
->src
[0].f
*= inst
->src
[1].f
;
2660 inst
->src
[1] = reg_undef
;
2665 case BRW_OPCODE_ADD
:
2666 if (inst
->src
[1].file
!= IMM
)
2669 if (inst
->src
[0].file
== IMM
) {
2670 assert(inst
->src
[0].type
== BRW_REGISTER_TYPE_F
);
2671 inst
->opcode
= BRW_OPCODE_MOV
;
2672 inst
->src
[0].f
+= inst
->src
[1].f
;
2673 inst
->src
[1] = reg_undef
;
2679 if (inst
->src
[0].equals(inst
->src
[1]) ||
2680 inst
->src
[1].is_zero()) {
2681 /* On Gen8+, the OR instruction can have a source modifier that
2682 * performs logical not on the operand. Cases of 'OR r0, ~r1, 0'
2683 * or 'OR r0, ~r1, ~r1' should become a NOT instead of a MOV.
2685 if (inst
->src
[0].negate
) {
2686 inst
->opcode
= BRW_OPCODE_NOT
;
2687 inst
->src
[0].negate
= false;
2689 inst
->opcode
= BRW_OPCODE_MOV
;
2691 inst
->src
[1] = reg_undef
;
2696 case BRW_OPCODE_CMP
:
2697 if ((inst
->conditional_mod
== BRW_CONDITIONAL_Z
||
2698 inst
->conditional_mod
== BRW_CONDITIONAL_NZ
) &&
2699 inst
->src
[1].is_zero() &&
2700 (inst
->src
[0].abs
|| inst
->src
[0].negate
)) {
2701 inst
->src
[0].abs
= false;
2702 inst
->src
[0].negate
= false;
2707 case BRW_OPCODE_SEL
:
2708 if (!devinfo
->has_64bit_float
&&
2709 !devinfo
->has_64bit_int
&&
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];
2843 invalidate_analysis(DEPENDENCY_INSTRUCTION_DATA_FLOW
|
2844 DEPENDENCY_INSTRUCTION_DETAIL
);
2850 * Optimize sample messages that have constant zero values for the trailing
2851 * texture coordinates. We can just reduce the message length for these
2852 * instructions instead of reserving a register for it. Trailing parameters
2853 * that aren't sent default to zero anyway. This will cause the dead code
2854 * eliminator to remove the MOV instruction that would otherwise be emitted to
2855 * set up the zero value.
2858 fs_visitor::opt_zero_samples()
2860 /* Gen4 infers the texturing opcode based on the message length so we can't
2863 if (devinfo
->gen
< 5)
2866 bool progress
= false;
2868 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2869 if (!inst
->is_tex())
2872 fs_inst
*load_payload
= (fs_inst
*) inst
->prev
;
2874 if (load_payload
->is_head_sentinel() ||
2875 load_payload
->opcode
!= SHADER_OPCODE_LOAD_PAYLOAD
)
2878 /* We don't want to remove the message header or the first parameter.
2879 * Removing the first parameter is not allowed, see the Haswell PRM
2880 * volume 7, page 149:
2882 * "Parameter 0 is required except for the sampleinfo message, which
2883 * has no parameter 0"
2885 while (inst
->mlen
> inst
->header_size
+ inst
->exec_size
/ 8 &&
2886 load_payload
->src
[(inst
->mlen
- inst
->header_size
) /
2887 (inst
->exec_size
/ 8) +
2888 inst
->header_size
- 1].is_zero()) {
2889 inst
->mlen
-= inst
->exec_size
/ 8;
2895 invalidate_analysis(DEPENDENCY_INSTRUCTION_DETAIL
);
2901 * Optimize sample messages which are followed by the final RT write.
2903 * CHV, and GEN9+ can mark a texturing SEND instruction with EOT to have its
2904 * results sent directly to the framebuffer, bypassing the EU. Recognize the
2905 * final texturing results copied to the framebuffer write payload and modify
2906 * them to write to the framebuffer directly.
2909 fs_visitor::opt_sampler_eot()
2911 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
2913 if (stage
!= MESA_SHADER_FRAGMENT
|| dispatch_width
> 16)
2916 if (devinfo
->gen
!= 9 && !devinfo
->is_cherryview
)
2919 /* FINISHME: It should be possible to implement this optimization when there
2920 * are multiple drawbuffers.
2922 if (key
->nr_color_regions
!= 1)
2925 /* Requires emitting a bunch of saturating MOV instructions during logical
2926 * send lowering to clamp the color payload, which the sampler unit isn't
2927 * going to do for us.
2929 if (key
->clamp_fragment_color
)
2932 /* Look for a texturing instruction immediately before the final FB_WRITE. */
2933 bblock_t
*block
= cfg
->blocks
[cfg
->num_blocks
- 1];
2934 fs_inst
*fb_write
= (fs_inst
*)block
->end();
2935 assert(fb_write
->eot
);
2936 assert(fb_write
->opcode
== FS_OPCODE_FB_WRITE_LOGICAL
);
2938 /* There wasn't one; nothing to do. */
2939 if (unlikely(fb_write
->prev
->is_head_sentinel()))
2942 fs_inst
*tex_inst
= (fs_inst
*) fb_write
->prev
;
2944 /* 3D Sampler » Messages » Message Format
2946 * “Response Length of zero is allowed on all SIMD8* and SIMD16* sampler
2947 * messages except sample+killpix, resinfo, sampleinfo, LOD, and gather4*”
2949 if (tex_inst
->opcode
!= SHADER_OPCODE_TEX_LOGICAL
&&
2950 tex_inst
->opcode
!= SHADER_OPCODE_TXD_LOGICAL
&&
2951 tex_inst
->opcode
!= SHADER_OPCODE_TXF_LOGICAL
&&
2952 tex_inst
->opcode
!= SHADER_OPCODE_TXL_LOGICAL
&&
2953 tex_inst
->opcode
!= FS_OPCODE_TXB_LOGICAL
&&
2954 tex_inst
->opcode
!= SHADER_OPCODE_TXF_CMS_LOGICAL
&&
2955 tex_inst
->opcode
!= SHADER_OPCODE_TXF_CMS_W_LOGICAL
&&
2956 tex_inst
->opcode
!= SHADER_OPCODE_TXF_UMS_LOGICAL
)
2959 /* XXX - This shouldn't be necessary. */
2960 if (tex_inst
->prev
->is_head_sentinel())
2963 /* Check that the FB write sources are fully initialized by the single
2964 * texturing instruction.
2966 for (unsigned i
= 0; i
< FB_WRITE_LOGICAL_NUM_SRCS
; i
++) {
2967 if (i
== FB_WRITE_LOGICAL_SRC_COLOR0
) {
2968 if (!fb_write
->src
[i
].equals(tex_inst
->dst
) ||
2969 fb_write
->size_read(i
) != tex_inst
->size_written
)
2971 } else if (i
!= FB_WRITE_LOGICAL_SRC_COMPONENTS
) {
2972 if (fb_write
->src
[i
].file
!= BAD_FILE
)
2977 assert(!tex_inst
->eot
); /* We can't get here twice */
2978 assert((tex_inst
->offset
& (0xff << 24)) == 0);
2980 const fs_builder
ibld(this, block
, tex_inst
);
2982 tex_inst
->offset
|= fb_write
->target
<< 24;
2983 tex_inst
->eot
= true;
2984 tex_inst
->dst
= ibld
.null_reg_ud();
2985 tex_inst
->size_written
= 0;
2986 fb_write
->remove(cfg
->blocks
[cfg
->num_blocks
- 1]);
2988 /* Marking EOT is sufficient, lower_logical_sends() will notice the EOT
2989 * flag and submit a header together with the sampler message as required
2992 invalidate_analysis(DEPENDENCY_INSTRUCTIONS
| DEPENDENCY_VARIABLES
);
2997 fs_visitor::opt_register_renaming()
2999 bool progress
= false;
3002 unsigned remap
[alloc
.count
];
3003 memset(remap
, ~0u, sizeof(unsigned) * alloc
.count
);
3005 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
3006 if (inst
->opcode
== BRW_OPCODE_IF
|| inst
->opcode
== BRW_OPCODE_DO
) {
3008 } else if (inst
->opcode
== BRW_OPCODE_ENDIF
||
3009 inst
->opcode
== BRW_OPCODE_WHILE
) {
3013 /* Rewrite instruction sources. */
3014 for (int i
= 0; i
< inst
->sources
; i
++) {
3015 if (inst
->src
[i
].file
== VGRF
&&
3016 remap
[inst
->src
[i
].nr
] != ~0u &&
3017 remap
[inst
->src
[i
].nr
] != inst
->src
[i
].nr
) {
3018 inst
->src
[i
].nr
= remap
[inst
->src
[i
].nr
];
3023 const unsigned dst
= inst
->dst
.nr
;
3026 inst
->dst
.file
== VGRF
&&
3027 alloc
.sizes
[inst
->dst
.nr
] * REG_SIZE
== inst
->size_written
&&
3028 !inst
->is_partial_write()) {
3029 if (remap
[dst
] == ~0u) {
3032 remap
[dst
] = alloc
.allocate(regs_written(inst
));
3033 inst
->dst
.nr
= remap
[dst
];
3036 } else if (inst
->dst
.file
== VGRF
&&
3037 remap
[dst
] != ~0u &&
3038 remap
[dst
] != dst
) {
3039 inst
->dst
.nr
= remap
[dst
];
3045 invalidate_analysis(DEPENDENCY_INSTRUCTION_DETAIL
|
3046 DEPENDENCY_VARIABLES
);
3048 for (unsigned i
= 0; i
< ARRAY_SIZE(delta_xy
); i
++) {
3049 if (delta_xy
[i
].file
== VGRF
&& remap
[delta_xy
[i
].nr
] != ~0u) {
3050 delta_xy
[i
].nr
= remap
[delta_xy
[i
].nr
];
3059 * Remove redundant or useless discard jumps.
3061 * For example, we can eliminate jumps in the following sequence:
3063 * discard-jump (redundant with the next jump)
3064 * discard-jump (useless; jumps to the next instruction)
3068 fs_visitor::opt_redundant_discard_jumps()
3070 bool progress
= false;
3072 bblock_t
*last_bblock
= cfg
->blocks
[cfg
->num_blocks
- 1];
3074 fs_inst
*placeholder_halt
= NULL
;
3075 foreach_inst_in_block_reverse(fs_inst
, inst
, last_bblock
) {
3076 if (inst
->opcode
== FS_OPCODE_PLACEHOLDER_HALT
) {
3077 placeholder_halt
= inst
;
3082 if (!placeholder_halt
)
3085 /* Delete any HALTs immediately before the placeholder halt. */
3086 for (fs_inst
*prev
= (fs_inst
*) placeholder_halt
->prev
;
3087 !prev
->is_head_sentinel() && prev
->opcode
== FS_OPCODE_DISCARD_JUMP
;
3088 prev
= (fs_inst
*) placeholder_halt
->prev
) {
3089 prev
->remove(last_bblock
);
3094 invalidate_analysis(DEPENDENCY_INSTRUCTIONS
);
3100 * Compute a bitmask with GRF granularity with a bit set for each GRF starting
3101 * from \p r.offset which overlaps the region starting at \p s.offset and
3102 * spanning \p ds bytes.
3104 static inline unsigned
3105 mask_relative_to(const fs_reg
&r
, const fs_reg
&s
, unsigned ds
)
3107 const int rel_offset
= reg_offset(s
) - reg_offset(r
);
3108 const int shift
= rel_offset
/ REG_SIZE
;
3109 const unsigned n
= DIV_ROUND_UP(rel_offset
% REG_SIZE
+ ds
, REG_SIZE
);
3110 assert(reg_space(r
) == reg_space(s
) &&
3111 shift
>= 0 && shift
< int(8 * sizeof(unsigned)));
3112 return ((1 << n
) - 1) << shift
;
3116 fs_visitor::compute_to_mrf()
3118 bool progress
= false;
3121 /* No MRFs on Gen >= 7. */
3122 if (devinfo
->gen
>= 7)
3125 const fs_live_variables
&live
= live_analysis
.require();
3127 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
3131 if (inst
->opcode
!= BRW_OPCODE_MOV
||
3132 inst
->is_partial_write() ||
3133 inst
->dst
.file
!= MRF
|| inst
->src
[0].file
!= VGRF
||
3134 inst
->dst
.type
!= inst
->src
[0].type
||
3135 inst
->src
[0].abs
|| inst
->src
[0].negate
||
3136 !inst
->src
[0].is_contiguous() ||
3137 inst
->src
[0].offset
% REG_SIZE
!= 0)
3140 /* Can't compute-to-MRF this GRF if someone else was going to
3143 if (live
.vgrf_end
[inst
->src
[0].nr
] > ip
)
3146 /* Found a move of a GRF to a MRF. Let's see if we can go rewrite the
3147 * things that computed the value of all GRFs of the source region. The
3148 * regs_left bitset keeps track of the registers we haven't yet found a
3149 * generating instruction for.
3151 unsigned regs_left
= (1 << regs_read(inst
, 0)) - 1;
3153 foreach_inst_in_block_reverse_starting_from(fs_inst
, scan_inst
, inst
) {
3154 if (regions_overlap(scan_inst
->dst
, scan_inst
->size_written
,
3155 inst
->src
[0], inst
->size_read(0))) {
3156 /* Found the last thing to write our reg we want to turn
3157 * into a compute-to-MRF.
3160 /* If this one instruction didn't populate all the
3161 * channels, bail. We might be able to rewrite everything
3162 * that writes that reg, but it would require smarter
3165 if (scan_inst
->is_partial_write())
3168 /* Handling things not fully contained in the source of the copy
3169 * would need us to understand coalescing out more than one MOV at
3172 if (!region_contained_in(scan_inst
->dst
, scan_inst
->size_written
,
3173 inst
->src
[0], inst
->size_read(0)))
3176 /* SEND instructions can't have MRF as a destination. */
3177 if (scan_inst
->mlen
)
3180 if (devinfo
->gen
== 6) {
3181 /* gen6 math instructions must have the destination be
3182 * GRF, so no compute-to-MRF for them.
3184 if (scan_inst
->is_math()) {
3189 /* Clear the bits for any registers this instruction overwrites. */
3190 regs_left
&= ~mask_relative_to(
3191 inst
->src
[0], scan_inst
->dst
, scan_inst
->size_written
);
3196 /* We don't handle control flow here. Most computation of
3197 * values that end up in MRFs are shortly before the MRF
3200 if (block
->start() == scan_inst
)
3203 /* You can't read from an MRF, so if someone else reads our
3204 * MRF's source GRF that we wanted to rewrite, that stops us.
3206 bool interfered
= false;
3207 for (int i
= 0; i
< scan_inst
->sources
; i
++) {
3208 if (regions_overlap(scan_inst
->src
[i
], scan_inst
->size_read(i
),
3209 inst
->src
[0], inst
->size_read(0))) {
3216 if (regions_overlap(scan_inst
->dst
, scan_inst
->size_written
,
3217 inst
->dst
, inst
->size_written
)) {
3218 /* If somebody else writes our MRF here, we can't
3219 * compute-to-MRF before that.
3224 if (scan_inst
->mlen
> 0 && scan_inst
->base_mrf
!= -1 &&
3225 regions_overlap(fs_reg(MRF
, scan_inst
->base_mrf
), scan_inst
->mlen
* REG_SIZE
,
3226 inst
->dst
, inst
->size_written
)) {
3227 /* Found a SEND instruction, which means that there are
3228 * live values in MRFs from base_mrf to base_mrf +
3229 * scan_inst->mlen - 1. Don't go pushing our MRF write up
3239 /* Found all generating instructions of our MRF's source value, so it
3240 * should be safe to rewrite them to point to the MRF directly.
3242 regs_left
= (1 << regs_read(inst
, 0)) - 1;
3244 foreach_inst_in_block_reverse_starting_from(fs_inst
, scan_inst
, inst
) {
3245 if (regions_overlap(scan_inst
->dst
, scan_inst
->size_written
,
3246 inst
->src
[0], inst
->size_read(0))) {
3247 /* Clear the bits for any registers this instruction overwrites. */
3248 regs_left
&= ~mask_relative_to(
3249 inst
->src
[0], scan_inst
->dst
, scan_inst
->size_written
);
3251 const unsigned rel_offset
= reg_offset(scan_inst
->dst
) -
3252 reg_offset(inst
->src
[0]);
3254 if (inst
->dst
.nr
& BRW_MRF_COMPR4
) {
3255 /* Apply the same address transformation done by the hardware
3256 * for COMPR4 MRF writes.
3258 assert(rel_offset
< 2 * REG_SIZE
);
3259 scan_inst
->dst
.nr
= inst
->dst
.nr
+ rel_offset
/ REG_SIZE
* 4;
3261 /* Clear the COMPR4 bit if the generating instruction is not
3264 if (scan_inst
->size_written
< 2 * REG_SIZE
)
3265 scan_inst
->dst
.nr
&= ~BRW_MRF_COMPR4
;
3268 /* Calculate the MRF number the result of this instruction is
3269 * ultimately written to.
3271 scan_inst
->dst
.nr
= inst
->dst
.nr
+ rel_offset
/ REG_SIZE
;
3274 scan_inst
->dst
.file
= MRF
;
3275 scan_inst
->dst
.offset
= inst
->dst
.offset
+ rel_offset
% REG_SIZE
;
3276 scan_inst
->saturate
|= inst
->saturate
;
3283 inst
->remove(block
);
3288 invalidate_analysis(DEPENDENCY_INSTRUCTIONS
);
3294 * Eliminate FIND_LIVE_CHANNEL instructions occurring outside any control
3295 * flow. We could probably do better here with some form of divergence
3299 fs_visitor::eliminate_find_live_channel()
3301 bool progress
= false;
3304 if (!brw_stage_has_packed_dispatch(devinfo
, stage
, stage_prog_data
)) {
3305 /* The optimization below assumes that channel zero is live on thread
3306 * dispatch, which may not be the case if the fixed function dispatches
3312 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
3313 switch (inst
->opcode
) {
3319 case BRW_OPCODE_ENDIF
:
3320 case BRW_OPCODE_WHILE
:
3324 case FS_OPCODE_DISCARD_JUMP
:
3325 /* This can potentially make control flow non-uniform until the end
3330 case SHADER_OPCODE_FIND_LIVE_CHANNEL
:
3332 inst
->opcode
= BRW_OPCODE_MOV
;
3333 inst
->src
[0] = brw_imm_ud(0u);
3335 inst
->force_writemask_all
= true;
3346 invalidate_analysis(DEPENDENCY_INSTRUCTION_DETAIL
);
3352 * Once we've generated code, try to convert normal FS_OPCODE_FB_WRITE
3353 * instructions to FS_OPCODE_REP_FB_WRITE.
3356 fs_visitor::emit_repclear_shader()
3358 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
3360 int color_mrf
= base_mrf
+ 2;
3364 mov
= bld
.exec_all().group(4, 0)
3365 .MOV(brw_message_reg(color_mrf
),
3366 fs_reg(UNIFORM
, 0, BRW_REGISTER_TYPE_F
));
3368 struct brw_reg reg
=
3369 brw_reg(BRW_GENERAL_REGISTER_FILE
, 2, 3, 0, 0, BRW_REGISTER_TYPE_F
,
3370 BRW_VERTICAL_STRIDE_8
, BRW_WIDTH_2
, BRW_HORIZONTAL_STRIDE_4
,
3371 BRW_SWIZZLE_XYZW
, WRITEMASK_XYZW
);
3373 mov
= bld
.exec_all().group(4, 0)
3374 .MOV(vec4(brw_message_reg(color_mrf
)), fs_reg(reg
));
3377 fs_inst
*write
= NULL
;
3378 if (key
->nr_color_regions
== 1) {
3379 write
= bld
.emit(FS_OPCODE_REP_FB_WRITE
);
3380 write
->saturate
= key
->clamp_fragment_color
;
3381 write
->base_mrf
= color_mrf
;
3383 write
->header_size
= 0;
3386 assume(key
->nr_color_regions
> 0);
3388 struct brw_reg header
=
3389 retype(brw_message_reg(base_mrf
), BRW_REGISTER_TYPE_UD
);
3390 bld
.exec_all().group(16, 0)
3391 .MOV(header
, retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
));
3393 for (int i
= 0; i
< key
->nr_color_regions
; ++i
) {
3395 bld
.exec_all().group(1, 0)
3396 .MOV(component(header
, 2), brw_imm_ud(i
));
3399 write
= bld
.emit(FS_OPCODE_REP_FB_WRITE
);
3400 write
->saturate
= key
->clamp_fragment_color
;
3401 write
->base_mrf
= base_mrf
;
3403 write
->header_size
= 2;
3408 write
->last_rt
= true;
3412 assign_constant_locations();
3413 assign_curb_setup();
3415 /* Now that we have the uniform assigned, go ahead and force it to a vec4. */
3417 assert(mov
->src
[0].file
== FIXED_GRF
);
3418 mov
->src
[0] = brw_vec4_grf(mov
->src
[0].nr
, 0);
3425 * Walks through basic blocks, looking for repeated MRF writes and
3426 * removing the later ones.
3429 fs_visitor::remove_duplicate_mrf_writes()
3431 fs_inst
*last_mrf_move
[BRW_MAX_MRF(devinfo
->gen
)];
3432 bool progress
= false;
3434 /* Need to update the MRF tracking for compressed instructions. */
3435 if (dispatch_width
>= 16)
3438 memset(last_mrf_move
, 0, sizeof(last_mrf_move
));
3440 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
3441 if (inst
->is_control_flow()) {
3442 memset(last_mrf_move
, 0, sizeof(last_mrf_move
));
3445 if (inst
->opcode
== BRW_OPCODE_MOV
&&
3446 inst
->dst
.file
== MRF
) {
3447 fs_inst
*prev_inst
= last_mrf_move
[inst
->dst
.nr
];
3448 if (prev_inst
&& prev_inst
->opcode
== BRW_OPCODE_MOV
&&
3449 inst
->dst
.equals(prev_inst
->dst
) &&
3450 inst
->src
[0].equals(prev_inst
->src
[0]) &&
3451 inst
->saturate
== prev_inst
->saturate
&&
3452 inst
->predicate
== prev_inst
->predicate
&&
3453 inst
->conditional_mod
== prev_inst
->conditional_mod
&&
3454 inst
->exec_size
== prev_inst
->exec_size
) {
3455 inst
->remove(block
);
3461 /* Clear out the last-write records for MRFs that were overwritten. */
3462 if (inst
->dst
.file
== MRF
) {
3463 last_mrf_move
[inst
->dst
.nr
] = NULL
;
3466 if (inst
->mlen
> 0 && inst
->base_mrf
!= -1) {
3467 /* Found a SEND instruction, which will include two or fewer
3468 * implied MRF writes. We could do better here.
3470 for (unsigned i
= 0; i
< inst
->implied_mrf_writes(); i
++) {
3471 last_mrf_move
[inst
->base_mrf
+ i
] = NULL
;
3475 /* Clear out any MRF move records whose sources got overwritten. */
3476 for (unsigned i
= 0; i
< ARRAY_SIZE(last_mrf_move
); i
++) {
3477 if (last_mrf_move
[i
] &&
3478 regions_overlap(inst
->dst
, inst
->size_written
,
3479 last_mrf_move
[i
]->src
[0],
3480 last_mrf_move
[i
]->size_read(0))) {
3481 last_mrf_move
[i
] = NULL
;
3485 if (inst
->opcode
== BRW_OPCODE_MOV
&&
3486 inst
->dst
.file
== MRF
&&
3487 inst
->src
[0].file
!= ARF
&&
3488 !inst
->is_partial_write()) {
3489 last_mrf_move
[inst
->dst
.nr
] = inst
;
3494 invalidate_analysis(DEPENDENCY_INSTRUCTIONS
);
3500 * Rounding modes for conversion instructions are included for each
3501 * conversion, but right now it is a state. So once it is set,
3502 * we don't need to call it again for subsequent calls.
3504 * This is useful for vector/matrices conversions, as setting the
3505 * mode once is enough for the full vector/matrix
3508 fs_visitor::remove_extra_rounding_modes()
3510 bool progress
= false;
3511 unsigned execution_mode
= this->nir
->info
.float_controls_execution_mode
;
3513 brw_rnd_mode base_mode
= BRW_RND_MODE_UNSPECIFIED
;
3514 if ((FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
|
3515 FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
|
3516 FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
) &
3518 base_mode
= BRW_RND_MODE_RTNE
;
3519 if ((FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
|
3520 FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
|
3521 FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
) &
3523 base_mode
= BRW_RND_MODE_RTZ
;
3525 foreach_block (block
, cfg
) {
3526 brw_rnd_mode prev_mode
= base_mode
;
3528 foreach_inst_in_block_safe (fs_inst
, inst
, block
) {
3529 if (inst
->opcode
== SHADER_OPCODE_RND_MODE
) {
3530 assert(inst
->src
[0].file
== BRW_IMMEDIATE_VALUE
);
3531 const brw_rnd_mode mode
= (brw_rnd_mode
) inst
->src
[0].d
;
3532 if (mode
== prev_mode
) {
3533 inst
->remove(block
);
3543 invalidate_analysis(DEPENDENCY_INSTRUCTIONS
);
3549 clear_deps_for_inst_src(fs_inst
*inst
, bool *deps
, int first_grf
, int grf_len
)
3551 /* Clear the flag for registers that actually got read (as expected). */
3552 for (int i
= 0; i
< inst
->sources
; i
++) {
3554 if (inst
->src
[i
].file
== VGRF
|| inst
->src
[i
].file
== FIXED_GRF
) {
3555 grf
= inst
->src
[i
].nr
;
3560 if (grf
>= first_grf
&&
3561 grf
< first_grf
+ grf_len
) {
3562 deps
[grf
- first_grf
] = false;
3563 if (inst
->exec_size
== 16)
3564 deps
[grf
- first_grf
+ 1] = false;
3570 * Implements this workaround for the original 965:
3572 * "[DevBW, DevCL] Implementation Restrictions: As the hardware does not
3573 * check for post destination dependencies on this instruction, software
3574 * must ensure that there is no destination hazard for the case of ‘write
3575 * followed by a posted write’ shown in the following example.
3578 * 2. send r3.xy <rest of send instruction>
3581 * Due to no post-destination dependency check on the ‘send’, the above
3582 * code sequence could have two instructions (1 and 2) in flight at the
3583 * same time that both consider ‘r3’ as the target of their final writes.
3586 fs_visitor::insert_gen4_pre_send_dependency_workarounds(bblock_t
*block
,
3589 int write_len
= regs_written(inst
);
3590 int first_write_grf
= inst
->dst
.nr
;
3591 bool needs_dep
[BRW_MAX_MRF(devinfo
->gen
)];
3592 assert(write_len
< (int)sizeof(needs_dep
) - 1);
3594 memset(needs_dep
, false, sizeof(needs_dep
));
3595 memset(needs_dep
, true, write_len
);
3597 clear_deps_for_inst_src(inst
, needs_dep
, first_write_grf
, write_len
);
3599 /* Walk backwards looking for writes to registers we're writing which
3600 * aren't read since being written. If we hit the start of the program,
3601 * we assume that there are no outstanding dependencies on entry to the
3604 foreach_inst_in_block_reverse_starting_from(fs_inst
, scan_inst
, inst
) {
3605 /* If we hit control flow, assume that there *are* outstanding
3606 * dependencies, and force their cleanup before our instruction.
3608 if (block
->start() == scan_inst
&& block
->num
!= 0) {
3609 for (int i
= 0; i
< write_len
; i
++) {
3611 DEP_RESOLVE_MOV(fs_builder(this, block
, inst
),
3612 first_write_grf
+ i
);
3617 /* We insert our reads as late as possible on the assumption that any
3618 * instruction but a MOV that might have left us an outstanding
3619 * dependency has more latency than a MOV.
3621 if (scan_inst
->dst
.file
== VGRF
) {
3622 for (unsigned i
= 0; i
< regs_written(scan_inst
); i
++) {
3623 int reg
= scan_inst
->dst
.nr
+ i
;
3625 if (reg
>= first_write_grf
&&
3626 reg
< first_write_grf
+ write_len
&&
3627 needs_dep
[reg
- first_write_grf
]) {
3628 DEP_RESOLVE_MOV(fs_builder(this, block
, inst
), reg
);
3629 needs_dep
[reg
- first_write_grf
] = false;
3630 if (scan_inst
->exec_size
== 16)
3631 needs_dep
[reg
- first_write_grf
+ 1] = false;
3636 /* Clear the flag for registers that actually got read (as expected). */
3637 clear_deps_for_inst_src(scan_inst
, needs_dep
, first_write_grf
, write_len
);
3639 /* Continue the loop only if we haven't resolved all the dependencies */
3641 for (i
= 0; i
< write_len
; i
++) {
3651 * Implements this workaround for the original 965:
3653 * "[DevBW, DevCL] Errata: A destination register from a send can not be
3654 * used as a destination register until after it has been sourced by an
3655 * instruction with a different destination register.
3658 fs_visitor::insert_gen4_post_send_dependency_workarounds(bblock_t
*block
, fs_inst
*inst
)
3660 int write_len
= regs_written(inst
);
3661 unsigned first_write_grf
= inst
->dst
.nr
;
3662 bool needs_dep
[BRW_MAX_MRF(devinfo
->gen
)];
3663 assert(write_len
< (int)sizeof(needs_dep
) - 1);
3665 memset(needs_dep
, false, sizeof(needs_dep
));
3666 memset(needs_dep
, true, write_len
);
3667 /* Walk forwards looking for writes to registers we're writing which aren't
3668 * read before being written.
3670 foreach_inst_in_block_starting_from(fs_inst
, scan_inst
, inst
) {
3671 /* If we hit control flow, force resolve all remaining dependencies. */
3672 if (block
->end() == scan_inst
&& block
->num
!= cfg
->num_blocks
- 1) {
3673 for (int i
= 0; i
< write_len
; i
++) {
3675 DEP_RESOLVE_MOV(fs_builder(this, block
, scan_inst
),
3676 first_write_grf
+ i
);
3681 /* Clear the flag for registers that actually got read (as expected). */
3682 clear_deps_for_inst_src(scan_inst
, needs_dep
, first_write_grf
, write_len
);
3684 /* We insert our reads as late as possible since they're reading the
3685 * result of a SEND, which has massive latency.
3687 if (scan_inst
->dst
.file
== VGRF
&&
3688 scan_inst
->dst
.nr
>= first_write_grf
&&
3689 scan_inst
->dst
.nr
< first_write_grf
+ write_len
&&
3690 needs_dep
[scan_inst
->dst
.nr
- first_write_grf
]) {
3691 DEP_RESOLVE_MOV(fs_builder(this, block
, scan_inst
),
3693 needs_dep
[scan_inst
->dst
.nr
- first_write_grf
] = false;
3696 /* Continue the loop only if we haven't resolved all the dependencies */
3698 for (i
= 0; i
< write_len
; i
++) {
3708 fs_visitor::insert_gen4_send_dependency_workarounds()
3710 if (devinfo
->gen
!= 4 || devinfo
->is_g4x
)
3713 bool progress
= false;
3715 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
3716 if (inst
->mlen
!= 0 && inst
->dst
.file
== VGRF
) {
3717 insert_gen4_pre_send_dependency_workarounds(block
, inst
);
3718 insert_gen4_post_send_dependency_workarounds(block
, inst
);
3724 invalidate_analysis(DEPENDENCY_INSTRUCTIONS
);
3728 * Turns the generic expression-style uniform pull constant load instruction
3729 * into a hardware-specific series of instructions for loading a pull
3732 * The expression style allows the CSE pass before this to optimize out
3733 * repeated loads from the same offset, and gives the pre-register-allocation
3734 * scheduling full flexibility, while the conversion to native instructions
3735 * allows the post-register-allocation scheduler the best information
3738 * Note that execution masking for setting up pull constant loads is special:
3739 * the channels that need to be written are unrelated to the current execution
3740 * mask, since a later instruction will use one of the result channels as a
3741 * source operand for all 8 or 16 of its channels.
3744 fs_visitor::lower_uniform_pull_constant_loads()
3746 foreach_block_and_inst (block
, fs_inst
, inst
, cfg
) {
3747 if (inst
->opcode
!= FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
)
3750 if (devinfo
->gen
>= 7) {
3751 const fs_builder ubld
= fs_builder(this, block
, inst
).exec_all();
3752 const fs_reg payload
= ubld
.group(8, 0).vgrf(BRW_REGISTER_TYPE_UD
);
3754 ubld
.group(8, 0).MOV(payload
,
3755 retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
));
3756 ubld
.group(1, 0).MOV(component(payload
, 2),
3757 brw_imm_ud(inst
->src
[1].ud
/ 16));
3759 inst
->opcode
= FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7
;
3760 inst
->src
[1] = payload
;
3761 inst
->header_size
= 1;
3764 invalidate_analysis(DEPENDENCY_INSTRUCTIONS
| DEPENDENCY_VARIABLES
);
3766 /* Before register allocation, we didn't tell the scheduler about the
3767 * MRF we use. We know it's safe to use this MRF because nothing
3768 * else does except for register spill/unspill, which generates and
3769 * uses its MRF within a single IR instruction.
3771 inst
->base_mrf
= FIRST_PULL_LOAD_MRF(devinfo
->gen
) + 1;
3778 fs_visitor::lower_load_payload()
3780 bool progress
= false;
3782 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
3783 if (inst
->opcode
!= SHADER_OPCODE_LOAD_PAYLOAD
)
3786 assert(inst
->dst
.file
== MRF
|| inst
->dst
.file
== VGRF
);
3787 assert(inst
->saturate
== false);
3788 fs_reg dst
= inst
->dst
;
3790 /* Get rid of COMPR4. We'll add it back in if we need it */
3791 if (dst
.file
== MRF
)
3792 dst
.nr
= dst
.nr
& ~BRW_MRF_COMPR4
;
3794 const fs_builder
ibld(this, block
, inst
);
3795 const fs_builder ubld
= ibld
.exec_all();
3797 for (uint8_t i
= 0; i
< inst
->header_size
;) {
3798 /* Number of header GRFs to initialize at once with a single MOV
3802 (i
+ 1 < inst
->header_size
&& inst
->src
[i
].stride
== 1 &&
3803 inst
->src
[i
+ 1].equals(byte_offset(inst
->src
[i
], REG_SIZE
))) ?
3806 if (inst
->src
[i
].file
!= BAD_FILE
)
3807 ubld
.group(8 * n
, 0).MOV(retype(dst
, BRW_REGISTER_TYPE_UD
),
3808 retype(inst
->src
[i
], BRW_REGISTER_TYPE_UD
));
3810 dst
= byte_offset(dst
, n
* REG_SIZE
);
3814 if (inst
->dst
.file
== MRF
&& (inst
->dst
.nr
& BRW_MRF_COMPR4
) &&
3815 inst
->exec_size
> 8) {
3816 /* In this case, the payload portion of the LOAD_PAYLOAD isn't
3817 * a straightforward copy. Instead, the result of the
3818 * LOAD_PAYLOAD is treated as interleaved and the first four
3819 * non-header sources are unpacked as:
3830 * This is used for gen <= 5 fb writes.
3832 assert(inst
->exec_size
== 16);
3833 assert(inst
->header_size
+ 4 <= inst
->sources
);
3834 for (uint8_t i
= inst
->header_size
; i
< inst
->header_size
+ 4; i
++) {
3835 if (inst
->src
[i
].file
!= BAD_FILE
) {
3836 if (devinfo
->has_compr4
) {
3837 fs_reg compr4_dst
= retype(dst
, inst
->src
[i
].type
);
3838 compr4_dst
.nr
|= BRW_MRF_COMPR4
;
3839 ibld
.MOV(compr4_dst
, inst
->src
[i
]);
3841 /* Platform doesn't have COMPR4. We have to fake it */
3842 fs_reg mov_dst
= retype(dst
, inst
->src
[i
].type
);
3843 ibld
.half(0).MOV(mov_dst
, half(inst
->src
[i
], 0));
3845 ibld
.half(1).MOV(mov_dst
, half(inst
->src
[i
], 1));
3852 /* The loop above only ever incremented us through the first set
3853 * of 4 registers. However, thanks to the magic of COMPR4, we
3854 * actually wrote to the first 8 registers, so we need to take
3855 * that into account now.
3859 /* The COMPR4 code took care of the first 4 sources. We'll let
3860 * the regular path handle any remaining sources. Yes, we are
3861 * modifying the instruction but we're about to delete it so
3862 * this really doesn't hurt anything.
3864 inst
->header_size
+= 4;
3867 for (uint8_t i
= inst
->header_size
; i
< inst
->sources
; i
++) {
3868 if (inst
->src
[i
].file
!= BAD_FILE
) {
3869 dst
.type
= inst
->src
[i
].type
;
3870 ibld
.MOV(dst
, inst
->src
[i
]);
3872 dst
.type
= BRW_REGISTER_TYPE_UD
;
3874 dst
= offset(dst
, ibld
, 1);
3877 inst
->remove(block
);
3882 invalidate_analysis(DEPENDENCY_INSTRUCTIONS
);
3888 fs_visitor::lower_mul_dword_inst(fs_inst
*inst
, bblock_t
*block
)
3890 const fs_builder
ibld(this, block
, inst
);
3892 const bool ud
= (inst
->src
[1].type
== BRW_REGISTER_TYPE_UD
);
3893 if (inst
->src
[1].file
== IMM
&&
3894 (( ud
&& inst
->src
[1].ud
<= UINT16_MAX
) ||
3895 (!ud
&& inst
->src
[1].d
<= INT16_MAX
&& inst
->src
[1].d
>= INT16_MIN
))) {
3896 /* The MUL instruction isn't commutative. On Gen <= 6, only the low
3897 * 16-bits of src0 are read, and on Gen >= 7 only the low 16-bits of
3900 * If multiplying by an immediate value that fits in 16-bits, do a
3901 * single MUL instruction with that value in the proper location.
3903 if (devinfo
->gen
< 7) {
3904 fs_reg
imm(VGRF
, alloc
.allocate(dispatch_width
/ 8), inst
->dst
.type
);
3905 ibld
.MOV(imm
, inst
->src
[1]);
3906 ibld
.MUL(inst
->dst
, imm
, inst
->src
[0]);
3908 ibld
.MUL(inst
->dst
, inst
->src
[0],
3909 ud
? brw_imm_uw(inst
->src
[1].ud
)
3910 : brw_imm_w(inst
->src
[1].d
));
3913 /* Gen < 8 (and some Gen8+ low-power parts like Cherryview) cannot
3914 * do 32-bit integer multiplication in one instruction, but instead
3915 * must do a sequence (which actually calculates a 64-bit result):
3917 * mul(8) acc0<1>D g3<8,8,1>D g4<8,8,1>D
3918 * mach(8) null g3<8,8,1>D g4<8,8,1>D
3919 * mov(8) g2<1>D acc0<8,8,1>D
3921 * But on Gen > 6, the ability to use second accumulator register
3922 * (acc1) for non-float data types was removed, preventing a simple
3923 * implementation in SIMD16. A 16-channel result can be calculated by
3924 * executing the three instructions twice in SIMD8, once with quarter
3925 * control of 1Q for the first eight channels and again with 2Q for
3926 * the second eight channels.
3928 * Which accumulator register is implicitly accessed (by AccWrEnable
3929 * for instance) is determined by the quarter control. Unfortunately
3930 * Ivybridge (and presumably Baytrail) has a hardware bug in which an
3931 * implicit accumulator access by an instruction with 2Q will access
3932 * acc1 regardless of whether the data type is usable in acc1.
3934 * Specifically, the 2Q mach(8) writes acc1 which does not exist for
3935 * integer data types.
3937 * Since we only want the low 32-bits of the result, we can do two
3938 * 32-bit x 16-bit multiplies (like the mul and mach are doing), and
3939 * adjust the high result and add them (like the mach is doing):
3941 * mul(8) g7<1>D g3<8,8,1>D g4.0<8,8,1>UW
3942 * mul(8) g8<1>D g3<8,8,1>D g4.1<8,8,1>UW
3943 * shl(8) g9<1>D g8<8,8,1>D 16D
3944 * add(8) g2<1>D g7<8,8,1>D g8<8,8,1>D
3946 * We avoid the shl instruction by realizing that we only want to add
3947 * the low 16-bits of the "high" result to the high 16-bits of the
3948 * "low" result and using proper regioning on the add:
3950 * mul(8) g7<1>D g3<8,8,1>D g4.0<16,8,2>UW
3951 * mul(8) g8<1>D g3<8,8,1>D g4.1<16,8,2>UW
3952 * add(8) g7.1<2>UW g7.1<16,8,2>UW g8<16,8,2>UW
3954 * Since it does not use the (single) accumulator register, we can
3955 * schedule multi-component multiplications much better.
3958 bool needs_mov
= false;
3959 fs_reg orig_dst
= inst
->dst
;
3961 /* Get a new VGRF for the "low" 32x16-bit multiplication result if
3962 * reusing the original destination is impossible due to hardware
3963 * restrictions, source/destination overlap, or it being the null
3966 fs_reg low
= inst
->dst
;
3967 if (orig_dst
.is_null() || orig_dst
.file
== MRF
||
3968 regions_overlap(inst
->dst
, inst
->size_written
,
3969 inst
->src
[0], inst
->size_read(0)) ||
3970 regions_overlap(inst
->dst
, inst
->size_written
,
3971 inst
->src
[1], inst
->size_read(1)) ||
3972 inst
->dst
.stride
>= 4) {
3974 low
= fs_reg(VGRF
, alloc
.allocate(regs_written(inst
)),
3978 /* Get a new VGRF but keep the same stride as inst->dst */
3979 fs_reg
high(VGRF
, alloc
.allocate(regs_written(inst
)), inst
->dst
.type
);
3980 high
.stride
= inst
->dst
.stride
;
3981 high
.offset
= inst
->dst
.offset
% REG_SIZE
;
3983 if (devinfo
->gen
>= 7) {
3984 if (inst
->src
[1].abs
)
3985 lower_src_modifiers(this, block
, inst
, 1);
3987 if (inst
->src
[1].file
== IMM
) {
3988 ibld
.MUL(low
, inst
->src
[0],
3989 brw_imm_uw(inst
->src
[1].ud
& 0xffff));
3990 ibld
.MUL(high
, inst
->src
[0],
3991 brw_imm_uw(inst
->src
[1].ud
>> 16));
3993 ibld
.MUL(low
, inst
->src
[0],
3994 subscript(inst
->src
[1], BRW_REGISTER_TYPE_UW
, 0));
3995 ibld
.MUL(high
, inst
->src
[0],
3996 subscript(inst
->src
[1], BRW_REGISTER_TYPE_UW
, 1));
3999 if (inst
->src
[0].abs
)
4000 lower_src_modifiers(this, block
, inst
, 0);
4002 ibld
.MUL(low
, subscript(inst
->src
[0], BRW_REGISTER_TYPE_UW
, 0),
4004 ibld
.MUL(high
, subscript(inst
->src
[0], BRW_REGISTER_TYPE_UW
, 1),
4008 ibld
.ADD(subscript(low
, BRW_REGISTER_TYPE_UW
, 1),
4009 subscript(low
, BRW_REGISTER_TYPE_UW
, 1),
4010 subscript(high
, BRW_REGISTER_TYPE_UW
, 0));
4012 if (needs_mov
|| inst
->conditional_mod
)
4013 set_condmod(inst
->conditional_mod
, ibld
.MOV(orig_dst
, low
));
4018 fs_visitor::lower_mul_qword_inst(fs_inst
*inst
, bblock_t
*block
)
4020 const fs_builder
ibld(this, block
, inst
);
4022 /* Considering two 64-bit integers ab and cd where each letter ab
4023 * corresponds to 32 bits, we get a 128-bit result WXYZ. We * cd
4024 * only need to provide the YZ part of the result. -------
4026 * Only BD needs to be 64 bits. For AD and BC we only care + AD
4027 * about the lower 32 bits (since they are part of the upper + BC
4028 * 32 bits of our result). AC is not needed since it starts + AC
4029 * on the 65th bit of the result. -------
4032 unsigned int q_regs
= regs_written(inst
);
4033 unsigned int d_regs
= (q_regs
+ 1) / 2;
4035 fs_reg
bd(VGRF
, alloc
.allocate(q_regs
), BRW_REGISTER_TYPE_UQ
);
4036 fs_reg
ad(VGRF
, alloc
.allocate(d_regs
), BRW_REGISTER_TYPE_UD
);
4037 fs_reg
bc(VGRF
, alloc
.allocate(d_regs
), BRW_REGISTER_TYPE_UD
);
4039 /* Here we need the full 64 bit result for 32b * 32b. */
4040 if (devinfo
->has_integer_dword_mul
) {
4041 ibld
.MUL(bd
, subscript(inst
->src
[0], BRW_REGISTER_TYPE_UD
, 0),
4042 subscript(inst
->src
[1], BRW_REGISTER_TYPE_UD
, 0));
4044 fs_reg
bd_high(VGRF
, alloc
.allocate(d_regs
), BRW_REGISTER_TYPE_UD
);
4045 fs_reg
bd_low(VGRF
, alloc
.allocate(d_regs
), BRW_REGISTER_TYPE_UD
);
4046 fs_reg acc
= retype(brw_acc_reg(inst
->exec_size
), BRW_REGISTER_TYPE_UD
);
4048 fs_inst
*mul
= ibld
.MUL(acc
,
4049 subscript(inst
->src
[0], BRW_REGISTER_TYPE_UD
, 0),
4050 subscript(inst
->src
[1], BRW_REGISTER_TYPE_UW
, 0));
4051 mul
->writes_accumulator
= true;
4053 ibld
.MACH(bd_high
, subscript(inst
->src
[0], BRW_REGISTER_TYPE_UD
, 0),
4054 subscript(inst
->src
[1], BRW_REGISTER_TYPE_UD
, 0));
4055 ibld
.MOV(bd_low
, acc
);
4057 ibld
.MOV(subscript(bd
, BRW_REGISTER_TYPE_UD
, 0), bd_low
);
4058 ibld
.MOV(subscript(bd
, BRW_REGISTER_TYPE_UD
, 1), bd_high
);
4061 ibld
.MUL(ad
, subscript(inst
->src
[0], BRW_REGISTER_TYPE_UD
, 1),
4062 subscript(inst
->src
[1], BRW_REGISTER_TYPE_UD
, 0));
4063 ibld
.MUL(bc
, subscript(inst
->src
[0], BRW_REGISTER_TYPE_UD
, 0),
4064 subscript(inst
->src
[1], BRW_REGISTER_TYPE_UD
, 1));
4066 ibld
.ADD(ad
, ad
, bc
);
4067 ibld
.ADD(subscript(bd
, BRW_REGISTER_TYPE_UD
, 1),
4068 subscript(bd
, BRW_REGISTER_TYPE_UD
, 1), ad
);
4070 ibld
.MOV(inst
->dst
, bd
);
4074 fs_visitor::lower_mulh_inst(fs_inst
*inst
, bblock_t
*block
)
4076 const fs_builder
ibld(this, block
, inst
);
4078 /* According to the BDW+ BSpec page for the "Multiply Accumulate
4079 * High" instruction:
4081 * "An added preliminary mov is required for source modification on
4083 * mov (8) r3.0<1>:d -r3<8;8,1>:d
4084 * mul (8) acc0:d r2.0<8;8,1>:d r3.0<16;8,2>:uw
4085 * mach (8) r5.0<1>:d r2.0<8;8,1>:d r3.0<8;8,1>:d"
4087 if (devinfo
->gen
>= 8 && (inst
->src
[1].negate
|| inst
->src
[1].abs
))
4088 lower_src_modifiers(this, block
, inst
, 1);
4090 /* Should have been lowered to 8-wide. */
4091 assert(inst
->exec_size
<= get_lowered_simd_width(devinfo
, inst
));
4092 const fs_reg acc
= retype(brw_acc_reg(inst
->exec_size
), inst
->dst
.type
);
4093 fs_inst
*mul
= ibld
.MUL(acc
, inst
->src
[0], inst
->src
[1]);
4094 fs_inst
*mach
= ibld
.MACH(inst
->dst
, inst
->src
[0], inst
->src
[1]);
4096 if (devinfo
->gen
>= 8) {
4097 /* Until Gen8, integer multiplies read 32-bits from one source,
4098 * and 16-bits from the other, and relying on the MACH instruction
4099 * to generate the high bits of the result.
4101 * On Gen8, the multiply instruction does a full 32x32-bit
4102 * multiply, but in order to do a 64-bit multiply we can simulate
4103 * the previous behavior and then use a MACH instruction.
4105 assert(mul
->src
[1].type
== BRW_REGISTER_TYPE_D
||
4106 mul
->src
[1].type
== BRW_REGISTER_TYPE_UD
);
4107 mul
->src
[1].type
= BRW_REGISTER_TYPE_UW
;
4108 mul
->src
[1].stride
*= 2;
4110 if (mul
->src
[1].file
== IMM
) {
4111 mul
->src
[1] = brw_imm_uw(mul
->src
[1].ud
);
4113 } else if (devinfo
->gen
== 7 && !devinfo
->is_haswell
&&
4115 /* Among other things the quarter control bits influence which
4116 * accumulator register is used by the hardware for instructions
4117 * that access the accumulator implicitly (e.g. MACH). A
4118 * second-half instruction would normally map to acc1, which
4119 * doesn't exist on Gen7 and up (the hardware does emulate it for
4120 * floating-point instructions *only* by taking advantage of the
4121 * extra precision of acc0 not normally used for floating point
4124 * HSW and up are careful enough not to try to access an
4125 * accumulator register that doesn't exist, but on earlier Gen7
4126 * hardware we need to make sure that the quarter control bits are
4127 * zero to avoid non-deterministic behaviour and emit an extra MOV
4128 * to get the result masked correctly according to the current
4132 mach
->force_writemask_all
= true;
4133 mach
->dst
= ibld
.vgrf(inst
->dst
.type
);
4134 ibld
.MOV(inst
->dst
, mach
->dst
);
4139 fs_visitor::lower_integer_multiplication()
4141 bool progress
= false;
4143 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
4144 if (inst
->opcode
== BRW_OPCODE_MUL
) {
4145 /* If the instruction is already in a form that does not need lowering,
4148 if (devinfo
->gen
>= 7) {
4149 if (type_sz(inst
->src
[1].type
) < 4 && type_sz(inst
->src
[0].type
) <= 4)
4152 if (type_sz(inst
->src
[0].type
) < 4 && type_sz(inst
->src
[1].type
) <= 4)
4156 if ((inst
->dst
.type
== BRW_REGISTER_TYPE_Q
||
4157 inst
->dst
.type
== BRW_REGISTER_TYPE_UQ
) &&
4158 (inst
->src
[0].type
== BRW_REGISTER_TYPE_Q
||
4159 inst
->src
[0].type
== BRW_REGISTER_TYPE_UQ
) &&
4160 (inst
->src
[1].type
== BRW_REGISTER_TYPE_Q
||
4161 inst
->src
[1].type
== BRW_REGISTER_TYPE_UQ
)) {
4162 lower_mul_qword_inst(inst
, block
);
4163 inst
->remove(block
);
4165 } else if (!inst
->dst
.is_accumulator() &&
4166 (inst
->dst
.type
== BRW_REGISTER_TYPE_D
||
4167 inst
->dst
.type
== BRW_REGISTER_TYPE_UD
) &&
4168 !devinfo
->has_integer_dword_mul
) {
4169 lower_mul_dword_inst(inst
, block
);
4170 inst
->remove(block
);
4173 } else if (inst
->opcode
== SHADER_OPCODE_MULH
) {
4174 lower_mulh_inst(inst
, block
);
4175 inst
->remove(block
);
4182 invalidate_analysis(DEPENDENCY_INSTRUCTIONS
| DEPENDENCY_VARIABLES
);
4188 fs_visitor::lower_minmax()
4190 assert(devinfo
->gen
< 6);
4192 bool progress
= false;
4194 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
4195 const fs_builder
ibld(this, block
, inst
);
4197 if (inst
->opcode
== BRW_OPCODE_SEL
&&
4198 inst
->predicate
== BRW_PREDICATE_NONE
) {
4199 /* FIXME: Using CMP doesn't preserve the NaN propagation semantics of
4200 * the original SEL.L/GE instruction
4202 ibld
.CMP(ibld
.null_reg_d(), inst
->src
[0], inst
->src
[1],
4203 inst
->conditional_mod
);
4204 inst
->predicate
= BRW_PREDICATE_NORMAL
;
4205 inst
->conditional_mod
= BRW_CONDITIONAL_NONE
;
4212 invalidate_analysis(DEPENDENCY_INSTRUCTIONS
);
4218 fs_visitor::lower_sub_sat()
4220 bool progress
= false;
4222 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
4223 const fs_builder
ibld(this, block
, inst
);
4225 if (inst
->opcode
== SHADER_OPCODE_USUB_SAT
||
4226 inst
->opcode
== SHADER_OPCODE_ISUB_SAT
) {
4227 /* The fundamental problem is the hardware performs source negation
4228 * at the bit width of the source. If the source is 0x80000000D, the
4229 * negation is 0x80000000D. As a result, subtractSaturate(0,
4230 * 0x80000000) will produce 0x80000000 instead of 0x7fffffff. There
4231 * are at least three ways to resolve this:
4233 * 1. Use the accumulator for the negated source. The accumulator is
4234 * 33 bits, so our source 0x80000000 is sign-extended to
4235 * 0x1800000000. The negation of which is 0x080000000. This
4236 * doesn't help for 64-bit integers (which are already bigger than
4237 * 33 bits). There are also only 8 accumulators, so SIMD16 or
4238 * SIMD32 instructions would have to be split into multiple SIMD8
4241 * 2. Use slightly different math. For any n-bit value x, we know (x
4242 * >> 1) != -(x >> 1). We can use this fact to only do
4243 * subtractions involving (x >> 1). subtractSaturate(a, b) ==
4244 * subtractSaturate(subtractSaturate(a, (b >> 1)), b - (b >> 1)).
4246 * 3. For unsigned sources, it is sufficient to replace the
4247 * subtractSaturate with (a > b) ? a - b : 0.
4249 * It may also be possible to use the SUBB instruction. This
4250 * implicitly writes the accumulator, so it could only be used in the
4251 * same situations as #1 above. It is further limited by only
4252 * allowing UD sources.
4254 if (inst
->exec_size
== 8 && inst
->src
[0].type
!= BRW_REGISTER_TYPE_Q
&&
4255 inst
->src
[0].type
!= BRW_REGISTER_TYPE_UQ
) {
4256 fs_reg
acc(ARF
, BRW_ARF_ACCUMULATOR
, inst
->src
[1].type
);
4258 ibld
.MOV(acc
, inst
->src
[1]);
4259 fs_inst
*add
= ibld
.ADD(inst
->dst
, acc
, inst
->src
[0]);
4260 add
->saturate
= true;
4261 add
->src
[0].negate
= true;
4262 } else if (inst
->opcode
== SHADER_OPCODE_ISUB_SAT
) {
4264 * dst = add.sat(add.sat(src0, -tmp), -(src1 - tmp));
4266 fs_reg tmp1
= ibld
.vgrf(inst
->src
[0].type
);
4267 fs_reg tmp2
= ibld
.vgrf(inst
->src
[0].type
);
4268 fs_reg tmp3
= ibld
.vgrf(inst
->src
[0].type
);
4271 ibld
.SHR(tmp1
, inst
->src
[1], brw_imm_d(1));
4273 add
= ibld
.ADD(tmp2
, inst
->src
[1], tmp1
);
4274 add
->src
[1].negate
= true;
4276 add
= ibld
.ADD(tmp3
, inst
->src
[0], tmp1
);
4277 add
->src
[1].negate
= true;
4278 add
->saturate
= true;
4280 add
= ibld
.ADD(inst
->dst
, tmp3
, tmp2
);
4281 add
->src
[1].negate
= true;
4282 add
->saturate
= true;
4284 /* a > b ? a - b : 0 */
4285 ibld
.CMP(ibld
.null_reg_d(), inst
->src
[0], inst
->src
[1],
4288 fs_inst
*add
= ibld
.ADD(inst
->dst
, inst
->src
[0], inst
->src
[1]);
4289 add
->src
[1].negate
= !add
->src
[1].negate
;
4291 ibld
.SEL(inst
->dst
, inst
->dst
, brw_imm_ud(0))
4292 ->predicate
= BRW_PREDICATE_NORMAL
;
4295 inst
->remove(block
);
4301 invalidate_analysis(DEPENDENCY_INSTRUCTIONS
| DEPENDENCY_VARIABLES
);
4307 * Get the mask of SIMD channels enabled during dispatch and not yet disabled
4308 * by discard. Due to the layout of the sample mask in the fragment shader
4309 * thread payload, \p bld is required to have a dispatch_width() not greater
4310 * than 16 for fragment shaders.
4313 sample_mask_reg(const fs_builder
&bld
)
4315 const fs_visitor
*v
= static_cast<const fs_visitor
*>(bld
.shader
);
4317 if (v
->stage
!= MESA_SHADER_FRAGMENT
) {
4318 return brw_imm_ud(0xffffffff);
4319 } else if (brw_wm_prog_data(v
->stage_prog_data
)->uses_kill
) {
4320 assert(bld
.dispatch_width() <= 16);
4321 return brw_flag_subreg(sample_mask_flag_subreg(v
) + bld
.group() / 16);
4323 assert(v
->devinfo
->gen
>= 6 && bld
.dispatch_width() <= 16);
4324 return retype(brw_vec1_grf((bld
.group() >= 16 ? 2 : 1), 7),
4325 BRW_REGISTER_TYPE_UW
);
4330 setup_color_payload(const fs_builder
&bld
, const brw_wm_prog_key
*key
,
4331 fs_reg
*dst
, fs_reg color
, unsigned components
)
4333 if (key
->clamp_fragment_color
) {
4334 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 4);
4335 assert(color
.type
== BRW_REGISTER_TYPE_F
);
4337 for (unsigned i
= 0; i
< components
; i
++)
4339 bld
.MOV(offset(tmp
, bld
, i
), offset(color
, bld
, i
)));
4344 for (unsigned i
= 0; i
< components
; i
++)
4345 dst
[i
] = offset(color
, bld
, i
);
4349 brw_fb_write_msg_control(const fs_inst
*inst
,
4350 const struct brw_wm_prog_data
*prog_data
)
4354 if (inst
->opcode
== FS_OPCODE_REP_FB_WRITE
) {
4355 assert(inst
->group
== 0 && inst
->exec_size
== 16);
4356 mctl
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE_REPLICATED
;
4357 } else if (prog_data
->dual_src_blend
) {
4358 assert(inst
->exec_size
== 8);
4360 if (inst
->group
% 16 == 0)
4361 mctl
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_DUAL_SOURCE_SUBSPAN01
;
4362 else if (inst
->group
% 16 == 8)
4363 mctl
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_DUAL_SOURCE_SUBSPAN23
;
4365 unreachable("Invalid dual-source FB write instruction group");
4367 assert(inst
->group
== 0 || (inst
->group
== 16 && inst
->exec_size
== 16));
4369 if (inst
->exec_size
== 16)
4370 mctl
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE
;
4371 else if (inst
->exec_size
== 8)
4372 mctl
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_SINGLE_SOURCE_SUBSPAN01
;
4374 unreachable("Invalid FB write execution size");
4381 lower_fb_write_logical_send(const fs_builder
&bld
, fs_inst
*inst
,
4382 const struct brw_wm_prog_data
*prog_data
,
4383 const brw_wm_prog_key
*key
,
4384 const fs_visitor::thread_payload
&payload
)
4386 assert(inst
->src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].file
== IMM
);
4387 const gen_device_info
*devinfo
= bld
.shader
->devinfo
;
4388 const fs_reg
&color0
= inst
->src
[FB_WRITE_LOGICAL_SRC_COLOR0
];
4389 const fs_reg
&color1
= inst
->src
[FB_WRITE_LOGICAL_SRC_COLOR1
];
4390 const fs_reg
&src0_alpha
= inst
->src
[FB_WRITE_LOGICAL_SRC_SRC0_ALPHA
];
4391 const fs_reg
&src_depth
= inst
->src
[FB_WRITE_LOGICAL_SRC_SRC_DEPTH
];
4392 const fs_reg
&dst_depth
= inst
->src
[FB_WRITE_LOGICAL_SRC_DST_DEPTH
];
4393 const fs_reg
&src_stencil
= inst
->src
[FB_WRITE_LOGICAL_SRC_SRC_STENCIL
];
4394 fs_reg sample_mask
= inst
->src
[FB_WRITE_LOGICAL_SRC_OMASK
];
4395 const unsigned components
=
4396 inst
->src
[FB_WRITE_LOGICAL_SRC_COMPONENTS
].ud
;
4398 assert(inst
->target
!= 0 || src0_alpha
.file
== BAD_FILE
);
4400 /* We can potentially have a message length of up to 15, so we have to set
4401 * base_mrf to either 0 or 1 in order to fit in m0..m15.
4404 int header_size
= 2, payload_header_size
;
4405 unsigned length
= 0;
4407 if (devinfo
->gen
< 6) {
4408 /* TODO: Support SIMD32 on gen4-5 */
4409 assert(bld
.group() < 16);
4411 /* For gen4-5, we always have a header consisting of g0 and g1. We have
4412 * an implied MOV from g0,g1 to the start of the message. The MOV from
4413 * g0 is handled by the hardware and the MOV from g1 is provided by the
4414 * generator. This is required because, on gen4-5, the generator may
4415 * generate two write messages with different message lengths in order
4416 * to handle AA data properly.
4418 * Also, since the pixel mask goes in the g0 portion of the message and
4419 * since render target writes are the last thing in the shader, we write
4420 * the pixel mask directly into g0 and it will get copied as part of the
4423 if (prog_data
->uses_kill
) {
4424 bld
.exec_all().group(1, 0)
4425 .MOV(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW
),
4426 sample_mask_reg(bld
));
4429 assert(length
== 0);
4431 } else if ((devinfo
->gen
<= 7 && !devinfo
->is_haswell
&&
4432 prog_data
->uses_kill
) ||
4433 (devinfo
->gen
< 11 &&
4434 (color1
.file
!= BAD_FILE
|| key
->nr_color_regions
> 1))) {
4435 /* From the Sandy Bridge PRM, volume 4, page 198:
4437 * "Dispatched Pixel Enables. One bit per pixel indicating
4438 * which pixels were originally enabled when the thread was
4439 * dispatched. This field is only required for the end-of-
4440 * thread message and on all dual-source messages."
4442 const fs_builder ubld
= bld
.exec_all().group(8, 0);
4444 fs_reg header
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
4445 if (bld
.group() < 16) {
4446 /* The header starts off as g0 and g1 for the first half */
4447 ubld
.group(16, 0).MOV(header
, retype(brw_vec8_grf(0, 0),
4448 BRW_REGISTER_TYPE_UD
));
4450 /* The header starts off as g0 and g2 for the second half */
4451 assert(bld
.group() < 32);
4452 const fs_reg header_sources
[2] = {
4453 retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
),
4454 retype(brw_vec8_grf(2, 0), BRW_REGISTER_TYPE_UD
),
4456 ubld
.LOAD_PAYLOAD(header
, header_sources
, 2, 0);
4459 uint32_t g00_bits
= 0;
4461 /* Set "Source0 Alpha Present to RenderTarget" bit in message
4464 if (src0_alpha
.file
!= BAD_FILE
)
4465 g00_bits
|= 1 << 11;
4467 /* Set computes stencil to render target */
4468 if (prog_data
->computed_stencil
)
4469 g00_bits
|= 1 << 14;
4472 /* OR extra bits into g0.0 */
4473 ubld
.group(1, 0).OR(component(header
, 0),
4474 retype(brw_vec1_grf(0, 0),
4475 BRW_REGISTER_TYPE_UD
),
4476 brw_imm_ud(g00_bits
));
4479 /* Set the render target index for choosing BLEND_STATE. */
4480 if (inst
->target
> 0) {
4481 ubld
.group(1, 0).MOV(component(header
, 2), brw_imm_ud(inst
->target
));
4484 if (prog_data
->uses_kill
) {
4485 ubld
.group(1, 0).MOV(retype(component(header
, 15),
4486 BRW_REGISTER_TYPE_UW
),
4487 sample_mask_reg(bld
));
4490 assert(length
== 0);
4491 sources
[0] = header
;
4492 sources
[1] = horiz_offset(header
, 8);
4495 assert(length
== 0 || length
== 2);
4496 header_size
= length
;
4498 if (payload
.aa_dest_stencil_reg
[0]) {
4499 assert(inst
->group
< 16);
4500 sources
[length
] = fs_reg(VGRF
, bld
.shader
->alloc
.allocate(1));
4501 bld
.group(8, 0).exec_all().annotate("FB write stencil/AA alpha")
4502 .MOV(sources
[length
],
4503 fs_reg(brw_vec8_grf(payload
.aa_dest_stencil_reg
[0], 0)));
4507 if (src0_alpha
.file
!= BAD_FILE
) {
4508 for (unsigned i
= 0; i
< bld
.dispatch_width() / 8; i
++) {
4509 const fs_builder
&ubld
= bld
.exec_all().group(8, i
)
4510 .annotate("FB write src0 alpha");
4511 const fs_reg tmp
= ubld
.vgrf(BRW_REGISTER_TYPE_F
);
4512 ubld
.MOV(tmp
, horiz_offset(src0_alpha
, i
* 8));
4513 setup_color_payload(ubld
, key
, &sources
[length
], tmp
, 1);
4518 if (sample_mask
.file
!= BAD_FILE
) {
4519 sources
[length
] = fs_reg(VGRF
, bld
.shader
->alloc
.allocate(1),
4520 BRW_REGISTER_TYPE_UD
);
4522 /* Hand over gl_SampleMask. Only the lower 16 bits of each channel are
4523 * relevant. Since it's unsigned single words one vgrf is always
4524 * 16-wide, but only the lower or higher 8 channels will be used by the
4525 * hardware when doing a SIMD8 write depending on whether we have
4526 * selected the subspans for the first or second half respectively.
4528 assert(sample_mask
.file
!= BAD_FILE
&& type_sz(sample_mask
.type
) == 4);
4529 sample_mask
.type
= BRW_REGISTER_TYPE_UW
;
4530 sample_mask
.stride
*= 2;
4532 bld
.exec_all().annotate("FB write oMask")
4533 .MOV(horiz_offset(retype(sources
[length
], BRW_REGISTER_TYPE_UW
),
4539 payload_header_size
= length
;
4541 setup_color_payload(bld
, key
, &sources
[length
], color0
, components
);
4544 if (color1
.file
!= BAD_FILE
) {
4545 setup_color_payload(bld
, key
, &sources
[length
], color1
, components
);
4549 if (src_depth
.file
!= BAD_FILE
) {
4550 sources
[length
] = src_depth
;
4554 if (dst_depth
.file
!= BAD_FILE
) {
4555 sources
[length
] = dst_depth
;
4559 if (src_stencil
.file
!= BAD_FILE
) {
4560 assert(devinfo
->gen
>= 9);
4561 assert(bld
.dispatch_width() == 8);
4563 /* XXX: src_stencil is only available on gen9+. dst_depth is never
4564 * available on gen9+. As such it's impossible to have both enabled at the
4565 * same time and therefore length cannot overrun the array.
4567 assert(length
< 15);
4569 sources
[length
] = bld
.vgrf(BRW_REGISTER_TYPE_UD
);
4570 bld
.exec_all().annotate("FB write OS")
4571 .MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UB
),
4572 subscript(src_stencil
, BRW_REGISTER_TYPE_UB
, 0));
4577 if (devinfo
->gen
>= 7) {
4578 /* Send from the GRF */
4579 fs_reg payload
= fs_reg(VGRF
, -1, BRW_REGISTER_TYPE_F
);
4580 load
= bld
.LOAD_PAYLOAD(payload
, sources
, length
, payload_header_size
);
4581 payload
.nr
= bld
.shader
->alloc
.allocate(regs_written(load
));
4582 load
->dst
= payload
;
4584 uint32_t msg_ctl
= brw_fb_write_msg_control(inst
, prog_data
);
4585 uint32_t ex_desc
= 0;
4588 (inst
->group
/ 16) << 11 | /* rt slot group */
4589 brw_dp_write_desc(devinfo
, inst
->target
, msg_ctl
,
4590 GEN6_DATAPORT_WRITE_MESSAGE_RENDER_TARGET_WRITE
,
4591 inst
->last_rt
, false);
4593 if (devinfo
->gen
>= 11) {
4594 /* Set the "Render Target Index" and "Src0 Alpha Present" fields
4595 * in the extended message descriptor, in lieu of using a header.
4597 ex_desc
= inst
->target
<< 12 | (src0_alpha
.file
!= BAD_FILE
) << 15;
4599 if (key
->nr_color_regions
== 0)
4600 ex_desc
|= 1 << 20; /* Null Render Target */
4603 inst
->opcode
= SHADER_OPCODE_SEND
;
4604 inst
->resize_sources(3);
4605 inst
->sfid
= GEN6_SFID_DATAPORT_RENDER_CACHE
;
4606 inst
->src
[0] = brw_imm_ud(inst
->desc
);
4607 inst
->src
[1] = brw_imm_ud(ex_desc
);
4608 inst
->src
[2] = payload
;
4609 inst
->mlen
= regs_written(load
);
4611 inst
->header_size
= header_size
;
4612 inst
->check_tdr
= true;
4613 inst
->send_has_side_effects
= true;
4615 /* Send from the MRF */
4616 load
= bld
.LOAD_PAYLOAD(fs_reg(MRF
, 1, BRW_REGISTER_TYPE_F
),
4617 sources
, length
, payload_header_size
);
4619 /* On pre-SNB, we have to interlace the color values. LOAD_PAYLOAD
4620 * will do this for us if we just give it a COMPR4 destination.
4622 if (devinfo
->gen
< 6 && bld
.dispatch_width() == 16)
4623 load
->dst
.nr
|= BRW_MRF_COMPR4
;
4625 if (devinfo
->gen
< 6) {
4626 /* Set up src[0] for the implied MOV from grf0-1 */
4627 inst
->resize_sources(1);
4628 inst
->src
[0] = brw_vec8_grf(0, 0);
4630 inst
->resize_sources(0);
4633 inst
->opcode
= FS_OPCODE_FB_WRITE
;
4634 inst
->mlen
= regs_written(load
);
4635 inst
->header_size
= header_size
;
4640 lower_fb_read_logical_send(const fs_builder
&bld
, fs_inst
*inst
)
4642 const fs_builder
&ubld
= bld
.exec_all().group(8, 0);
4643 const unsigned length
= 2;
4644 const fs_reg header
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
, length
);
4646 if (bld
.group() < 16) {
4647 ubld
.group(16, 0).MOV(header
, retype(brw_vec8_grf(0, 0),
4648 BRW_REGISTER_TYPE_UD
));
4650 assert(bld
.group() < 32);
4651 const fs_reg header_sources
[] = {
4652 retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
),
4653 retype(brw_vec8_grf(2, 0), BRW_REGISTER_TYPE_UD
)
4655 ubld
.LOAD_PAYLOAD(header
, header_sources
, ARRAY_SIZE(header_sources
), 0);
4658 inst
->resize_sources(1);
4659 inst
->src
[0] = header
;
4660 inst
->opcode
= FS_OPCODE_FB_READ
;
4661 inst
->mlen
= length
;
4662 inst
->header_size
= length
;
4666 lower_sampler_logical_send_gen4(const fs_builder
&bld
, fs_inst
*inst
, opcode op
,
4667 const fs_reg
&coordinate
,
4668 const fs_reg
&shadow_c
,
4669 const fs_reg
&lod
, const fs_reg
&lod2
,
4670 const fs_reg
&surface
,
4671 const fs_reg
&sampler
,
4672 unsigned coord_components
,
4673 unsigned grad_components
)
4675 const bool has_lod
= (op
== SHADER_OPCODE_TXL
|| op
== FS_OPCODE_TXB
||
4676 op
== SHADER_OPCODE_TXF
|| op
== SHADER_OPCODE_TXS
);
4677 fs_reg
msg_begin(MRF
, 1, BRW_REGISTER_TYPE_F
);
4678 fs_reg msg_end
= msg_begin
;
4681 msg_end
= offset(msg_end
, bld
.group(8, 0), 1);
4683 for (unsigned i
= 0; i
< coord_components
; i
++)
4684 bld
.MOV(retype(offset(msg_end
, bld
, i
), coordinate
.type
),
4685 offset(coordinate
, bld
, i
));
4687 msg_end
= offset(msg_end
, bld
, coord_components
);
4689 /* Messages other than SAMPLE and RESINFO in SIMD16 and TXD in SIMD8
4690 * require all three components to be present and zero if they are unused.
4692 if (coord_components
> 0 &&
4693 (has_lod
|| shadow_c
.file
!= BAD_FILE
||
4694 (op
== SHADER_OPCODE_TEX
&& bld
.dispatch_width() == 8))) {
4695 for (unsigned i
= coord_components
; i
< 3; i
++)
4696 bld
.MOV(offset(msg_end
, bld
, i
), brw_imm_f(0.0f
));
4698 msg_end
= offset(msg_end
, bld
, 3 - coord_components
);
4701 if (op
== SHADER_OPCODE_TXD
) {
4702 /* TXD unsupported in SIMD16 mode. */
4703 assert(bld
.dispatch_width() == 8);
4705 /* the slots for u and v are always present, but r is optional */
4706 if (coord_components
< 2)
4707 msg_end
= offset(msg_end
, bld
, 2 - coord_components
);
4710 * dPdx = dudx, dvdx, drdx
4711 * dPdy = dudy, dvdy, drdy
4713 * 1-arg: Does not exist.
4715 * 2-arg: dudx dvdx dudy dvdy
4716 * dPdx.x dPdx.y dPdy.x dPdy.y
4719 * 3-arg: dudx dvdx drdx dudy dvdy drdy
4720 * dPdx.x dPdx.y dPdx.z dPdy.x dPdy.y dPdy.z
4721 * m5 m6 m7 m8 m9 m10
4723 for (unsigned i
= 0; i
< grad_components
; i
++)
4724 bld
.MOV(offset(msg_end
, bld
, i
), offset(lod
, bld
, i
));
4726 msg_end
= offset(msg_end
, bld
, MAX2(grad_components
, 2));
4728 for (unsigned i
= 0; i
< grad_components
; i
++)
4729 bld
.MOV(offset(msg_end
, bld
, i
), offset(lod2
, bld
, i
));
4731 msg_end
= offset(msg_end
, bld
, MAX2(grad_components
, 2));
4735 /* Bias/LOD with shadow comparator is unsupported in SIMD16 -- *Without*
4736 * shadow comparator (including RESINFO) it's unsupported in SIMD8 mode.
4738 assert(shadow_c
.file
!= BAD_FILE
? bld
.dispatch_width() == 8 :
4739 bld
.dispatch_width() == 16);
4741 const brw_reg_type type
=
4742 (op
== SHADER_OPCODE_TXF
|| op
== SHADER_OPCODE_TXS
?
4743 BRW_REGISTER_TYPE_UD
: BRW_REGISTER_TYPE_F
);
4744 bld
.MOV(retype(msg_end
, type
), lod
);
4745 msg_end
= offset(msg_end
, bld
, 1);
4748 if (shadow_c
.file
!= BAD_FILE
) {
4749 if (op
== SHADER_OPCODE_TEX
&& bld
.dispatch_width() == 8) {
4750 /* There's no plain shadow compare message, so we use shadow
4751 * compare with a bias of 0.0.
4753 bld
.MOV(msg_end
, brw_imm_f(0.0f
));
4754 msg_end
= offset(msg_end
, bld
, 1);
4757 bld
.MOV(msg_end
, shadow_c
);
4758 msg_end
= offset(msg_end
, bld
, 1);
4762 inst
->src
[0] = reg_undef
;
4763 inst
->src
[1] = surface
;
4764 inst
->src
[2] = sampler
;
4765 inst
->resize_sources(3);
4766 inst
->base_mrf
= msg_begin
.nr
;
4767 inst
->mlen
= msg_end
.nr
- msg_begin
.nr
;
4768 inst
->header_size
= 1;
4772 lower_sampler_logical_send_gen5(const fs_builder
&bld
, fs_inst
*inst
, opcode op
,
4773 const fs_reg
&coordinate
,
4774 const fs_reg
&shadow_c
,
4775 const fs_reg
&lod
, const fs_reg
&lod2
,
4776 const fs_reg
&sample_index
,
4777 const fs_reg
&surface
,
4778 const fs_reg
&sampler
,
4779 unsigned coord_components
,
4780 unsigned grad_components
)
4782 fs_reg
message(MRF
, 2, BRW_REGISTER_TYPE_F
);
4783 fs_reg msg_coords
= message
;
4784 unsigned header_size
= 0;
4786 if (inst
->offset
!= 0) {
4787 /* The offsets set up by the visitor are in the m1 header, so we can't
4794 for (unsigned i
= 0; i
< coord_components
; i
++)
4795 bld
.MOV(retype(offset(msg_coords
, bld
, i
), coordinate
.type
),
4796 offset(coordinate
, bld
, i
));
4798 fs_reg msg_end
= offset(msg_coords
, bld
, coord_components
);
4799 fs_reg msg_lod
= offset(msg_coords
, bld
, 4);
4801 if (shadow_c
.file
!= BAD_FILE
) {
4802 fs_reg msg_shadow
= msg_lod
;
4803 bld
.MOV(msg_shadow
, shadow_c
);
4804 msg_lod
= offset(msg_shadow
, bld
, 1);
4809 case SHADER_OPCODE_TXL
:
4811 bld
.MOV(msg_lod
, lod
);
4812 msg_end
= offset(msg_lod
, bld
, 1);
4814 case SHADER_OPCODE_TXD
:
4817 * dPdx = dudx, dvdx, drdx
4818 * dPdy = dudy, dvdy, drdy
4820 * Load up these values:
4821 * - dudx dudy dvdx dvdy drdx drdy
4822 * - dPdx.x dPdy.x dPdx.y dPdy.y dPdx.z dPdy.z
4825 for (unsigned i
= 0; i
< grad_components
; i
++) {
4826 bld
.MOV(msg_end
, offset(lod
, bld
, i
));
4827 msg_end
= offset(msg_end
, bld
, 1);
4829 bld
.MOV(msg_end
, offset(lod2
, bld
, i
));
4830 msg_end
= offset(msg_end
, bld
, 1);
4833 case SHADER_OPCODE_TXS
:
4834 msg_lod
= retype(msg_end
, BRW_REGISTER_TYPE_UD
);
4835 bld
.MOV(msg_lod
, lod
);
4836 msg_end
= offset(msg_lod
, bld
, 1);
4838 case SHADER_OPCODE_TXF
:
4839 msg_lod
= offset(msg_coords
, bld
, 3);
4840 bld
.MOV(retype(msg_lod
, BRW_REGISTER_TYPE_UD
), lod
);
4841 msg_end
= offset(msg_lod
, bld
, 1);
4843 case SHADER_OPCODE_TXF_CMS
:
4844 msg_lod
= offset(msg_coords
, bld
, 3);
4846 bld
.MOV(retype(msg_lod
, BRW_REGISTER_TYPE_UD
), brw_imm_ud(0u));
4848 bld
.MOV(retype(offset(msg_lod
, bld
, 1), BRW_REGISTER_TYPE_UD
), sample_index
);
4849 msg_end
= offset(msg_lod
, bld
, 2);
4856 inst
->src
[0] = reg_undef
;
4857 inst
->src
[1] = surface
;
4858 inst
->src
[2] = sampler
;
4859 inst
->resize_sources(3);
4860 inst
->base_mrf
= message
.nr
;
4861 inst
->mlen
= msg_end
.nr
- message
.nr
;
4862 inst
->header_size
= header_size
;
4864 /* Message length > MAX_SAMPLER_MESSAGE_SIZE disallowed by hardware. */
4865 assert(inst
->mlen
<= MAX_SAMPLER_MESSAGE_SIZE
);
4869 is_high_sampler(const struct gen_device_info
*devinfo
, const fs_reg
&sampler
)
4871 if (devinfo
->gen
< 8 && !devinfo
->is_haswell
)
4874 return sampler
.file
!= IMM
|| sampler
.ud
>= 16;
4878 sampler_msg_type(const gen_device_info
*devinfo
,
4879 opcode opcode
, bool shadow_compare
)
4881 assert(devinfo
->gen
>= 5);
4883 case SHADER_OPCODE_TEX
:
4884 return shadow_compare
? GEN5_SAMPLER_MESSAGE_SAMPLE_COMPARE
:
4885 GEN5_SAMPLER_MESSAGE_SAMPLE
;
4887 return shadow_compare
? GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS_COMPARE
:
4888 GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS
;
4889 case SHADER_OPCODE_TXL
:
4890 return shadow_compare
? GEN5_SAMPLER_MESSAGE_SAMPLE_LOD_COMPARE
:
4891 GEN5_SAMPLER_MESSAGE_SAMPLE_LOD
;
4892 case SHADER_OPCODE_TXL_LZ
:
4893 return shadow_compare
? GEN9_SAMPLER_MESSAGE_SAMPLE_C_LZ
:
4894 GEN9_SAMPLER_MESSAGE_SAMPLE_LZ
;
4895 case SHADER_OPCODE_TXS
:
4896 case SHADER_OPCODE_IMAGE_SIZE_LOGICAL
:
4897 return GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO
;
4898 case SHADER_OPCODE_TXD
:
4899 assert(!shadow_compare
|| devinfo
->gen
>= 8 || devinfo
->is_haswell
);
4900 return shadow_compare
? HSW_SAMPLER_MESSAGE_SAMPLE_DERIV_COMPARE
:
4901 GEN5_SAMPLER_MESSAGE_SAMPLE_DERIVS
;
4902 case SHADER_OPCODE_TXF
:
4903 return GEN5_SAMPLER_MESSAGE_SAMPLE_LD
;
4904 case SHADER_OPCODE_TXF_LZ
:
4905 assert(devinfo
->gen
>= 9);
4906 return GEN9_SAMPLER_MESSAGE_SAMPLE_LD_LZ
;
4907 case SHADER_OPCODE_TXF_CMS_W
:
4908 assert(devinfo
->gen
>= 9);
4909 return GEN9_SAMPLER_MESSAGE_SAMPLE_LD2DMS_W
;
4910 case SHADER_OPCODE_TXF_CMS
:
4911 return devinfo
->gen
>= 7 ? GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DMS
:
4912 GEN5_SAMPLER_MESSAGE_SAMPLE_LD
;
4913 case SHADER_OPCODE_TXF_UMS
:
4914 assert(devinfo
->gen
>= 7);
4915 return GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DSS
;
4916 case SHADER_OPCODE_TXF_MCS
:
4917 assert(devinfo
->gen
>= 7);
4918 return GEN7_SAMPLER_MESSAGE_SAMPLE_LD_MCS
;
4919 case SHADER_OPCODE_LOD
:
4920 return GEN5_SAMPLER_MESSAGE_LOD
;
4921 case SHADER_OPCODE_TG4
:
4922 assert(devinfo
->gen
>= 7);
4923 return shadow_compare
? GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_C
:
4924 GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4
;
4926 case SHADER_OPCODE_TG4_OFFSET
:
4927 assert(devinfo
->gen
>= 7);
4928 return shadow_compare
? GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO_C
:
4929 GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO
;
4930 case SHADER_OPCODE_SAMPLEINFO
:
4931 return GEN6_SAMPLER_MESSAGE_SAMPLE_SAMPLEINFO
;
4933 unreachable("not reached");
4938 lower_sampler_logical_send_gen7(const fs_builder
&bld
, fs_inst
*inst
, opcode op
,
4939 const fs_reg
&coordinate
,
4940 const fs_reg
&shadow_c
,
4941 fs_reg lod
, const fs_reg
&lod2
,
4942 const fs_reg
&min_lod
,
4943 const fs_reg
&sample_index
,
4945 const fs_reg
&surface
,
4946 const fs_reg
&sampler
,
4947 const fs_reg
&surface_handle
,
4948 const fs_reg
&sampler_handle
,
4949 const fs_reg
&tg4_offset
,
4950 unsigned coord_components
,
4951 unsigned grad_components
)
4953 const gen_device_info
*devinfo
= bld
.shader
->devinfo
;
4954 const brw_stage_prog_data
*prog_data
= bld
.shader
->stage_prog_data
;
4955 unsigned reg_width
= bld
.dispatch_width() / 8;
4956 unsigned header_size
= 0, length
= 0;
4957 fs_reg sources
[MAX_SAMPLER_MESSAGE_SIZE
];
4958 for (unsigned i
= 0; i
< ARRAY_SIZE(sources
); i
++)
4959 sources
[i
] = bld
.vgrf(BRW_REGISTER_TYPE_F
);
4961 /* We must have exactly one of surface/sampler and surface/sampler_handle */
4962 assert((surface
.file
== BAD_FILE
) != (surface_handle
.file
== BAD_FILE
));
4963 assert((sampler
.file
== BAD_FILE
) != (sampler_handle
.file
== BAD_FILE
));
4965 if (op
== SHADER_OPCODE_TG4
|| op
== SHADER_OPCODE_TG4_OFFSET
||
4966 inst
->offset
!= 0 || inst
->eot
||
4967 op
== SHADER_OPCODE_SAMPLEINFO
||
4968 sampler_handle
.file
!= BAD_FILE
||
4969 is_high_sampler(devinfo
, sampler
)) {
4970 /* For general texture offsets (no txf workaround), we need a header to
4973 * TG4 needs to place its channel select in the header, for interaction
4974 * with ARB_texture_swizzle. The sampler index is only 4-bits, so for
4975 * larger sampler numbers we need to offset the Sampler State Pointer in
4978 fs_reg header
= retype(sources
[0], BRW_REGISTER_TYPE_UD
);
4982 /* If we're requesting fewer than four channels worth of response,
4983 * and we have an explicit header, we need to set up the sampler
4984 * writemask. It's reversed from normal: 1 means "don't write".
4986 if (!inst
->eot
&& regs_written(inst
) != 4 * reg_width
) {
4987 assert(regs_written(inst
) % reg_width
== 0);
4988 unsigned mask
= ~((1 << (regs_written(inst
) / reg_width
)) - 1) & 0xf;
4989 inst
->offset
|= mask
<< 12;
4992 /* Build the actual header */
4993 const fs_builder ubld
= bld
.exec_all().group(8, 0);
4994 const fs_builder ubld1
= ubld
.group(1, 0);
4995 ubld
.MOV(header
, retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
));
4997 ubld1
.MOV(component(header
, 2), brw_imm_ud(inst
->offset
));
4998 } else if (bld
.shader
->stage
!= MESA_SHADER_VERTEX
&&
4999 bld
.shader
->stage
!= MESA_SHADER_FRAGMENT
) {
5000 /* The vertex and fragment stages have g0.2 set to 0, so
5001 * header0.2 is 0 when g0 is copied. Other stages may not, so we
5002 * must set it to 0 to avoid setting undesirable bits in the
5005 ubld1
.MOV(component(header
, 2), brw_imm_ud(0));
5008 if (sampler_handle
.file
!= BAD_FILE
) {
5009 /* Bindless sampler handles aren't relative to the sampler state
5010 * pointer passed into the shader through SAMPLER_STATE_POINTERS_*.
5011 * Instead, it's an absolute pointer relative to dynamic state base
5014 * Sampler states are 16 bytes each and the pointer we give here has
5015 * to be 32-byte aligned. In order to avoid more indirect messages
5016 * than required, we assume that all bindless sampler states are
5017 * 32-byte aligned. This sacrifices a bit of general state base
5018 * address space but means we can do something more efficient in the
5021 ubld1
.MOV(component(header
, 3), sampler_handle
);
5022 } else if (is_high_sampler(devinfo
, sampler
)) {
5023 if (sampler
.file
== BRW_IMMEDIATE_VALUE
) {
5024 assert(sampler
.ud
>= 16);
5025 const int sampler_state_size
= 16; /* 16 bytes */
5027 ubld1
.ADD(component(header
, 3),
5028 retype(brw_vec1_grf(0, 3), BRW_REGISTER_TYPE_UD
),
5029 brw_imm_ud(16 * (sampler
.ud
/ 16) * sampler_state_size
));
5031 fs_reg tmp
= ubld1
.vgrf(BRW_REGISTER_TYPE_UD
);
5032 ubld1
.AND(tmp
, sampler
, brw_imm_ud(0x0f0));
5033 ubld1
.SHL(tmp
, tmp
, brw_imm_ud(4));
5034 ubld1
.ADD(component(header
, 3),
5035 retype(brw_vec1_grf(0, 3), BRW_REGISTER_TYPE_UD
),
5041 if (shadow_c
.file
!= BAD_FILE
) {
5042 bld
.MOV(sources
[length
], shadow_c
);
5046 bool coordinate_done
= false;
5048 /* Set up the LOD info */
5051 case SHADER_OPCODE_TXL
:
5052 if (devinfo
->gen
>= 9 && op
== SHADER_OPCODE_TXL
&& lod
.is_zero()) {
5053 op
= SHADER_OPCODE_TXL_LZ
;
5056 bld
.MOV(sources
[length
], lod
);
5059 case SHADER_OPCODE_TXD
:
5060 /* TXD should have been lowered in SIMD16 mode. */
5061 assert(bld
.dispatch_width() == 8);
5063 /* Load dPdx and the coordinate together:
5064 * [hdr], [ref], x, dPdx.x, dPdy.x, y, dPdx.y, dPdy.y, z, dPdx.z, dPdy.z
5066 for (unsigned i
= 0; i
< coord_components
; i
++) {
5067 bld
.MOV(sources
[length
++], offset(coordinate
, bld
, i
));
5069 /* For cube map array, the coordinate is (u,v,r,ai) but there are
5070 * only derivatives for (u, v, r).
5072 if (i
< grad_components
) {
5073 bld
.MOV(sources
[length
++], offset(lod
, bld
, i
));
5074 bld
.MOV(sources
[length
++], offset(lod2
, bld
, i
));
5078 coordinate_done
= true;
5080 case SHADER_OPCODE_TXS
:
5081 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
), lod
);
5084 case SHADER_OPCODE_IMAGE_SIZE_LOGICAL
:
5085 /* We need an LOD; just use 0 */
5086 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
), brw_imm_ud(0));
5089 case SHADER_OPCODE_TXF
:
5090 /* Unfortunately, the parameters for LD are intermixed: u, lod, v, r.
5091 * On Gen9 they are u, v, lod, r
5093 bld
.MOV(retype(sources
[length
++], BRW_REGISTER_TYPE_D
), coordinate
);
5095 if (devinfo
->gen
>= 9) {
5096 if (coord_components
>= 2) {
5097 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
),
5098 offset(coordinate
, bld
, 1));
5100 sources
[length
] = brw_imm_d(0);
5105 if (devinfo
->gen
>= 9 && lod
.is_zero()) {
5106 op
= SHADER_OPCODE_TXF_LZ
;
5108 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_D
), lod
);
5112 for (unsigned i
= devinfo
->gen
>= 9 ? 2 : 1; i
< coord_components
; i
++)
5113 bld
.MOV(retype(sources
[length
++], BRW_REGISTER_TYPE_D
),
5114 offset(coordinate
, bld
, i
));
5116 coordinate_done
= true;
5119 case SHADER_OPCODE_TXF_CMS
:
5120 case SHADER_OPCODE_TXF_CMS_W
:
5121 case SHADER_OPCODE_TXF_UMS
:
5122 case SHADER_OPCODE_TXF_MCS
:
5123 if (op
== SHADER_OPCODE_TXF_UMS
||
5124 op
== SHADER_OPCODE_TXF_CMS
||
5125 op
== SHADER_OPCODE_TXF_CMS_W
) {
5126 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
), sample_index
);
5130 if (op
== SHADER_OPCODE_TXF_CMS
|| op
== SHADER_OPCODE_TXF_CMS_W
) {
5131 /* Data from the multisample control surface. */
5132 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
), mcs
);
5135 /* On Gen9+ we'll use ld2dms_w instead which has two registers for
5138 if (op
== SHADER_OPCODE_TXF_CMS_W
) {
5139 bld
.MOV(retype(sources
[length
], BRW_REGISTER_TYPE_UD
),
5142 offset(mcs
, bld
, 1));
5147 /* There is no offsetting for this message; just copy in the integer
5148 * texture coordinates.
5150 for (unsigned i
= 0; i
< coord_components
; i
++)
5151 bld
.MOV(retype(sources
[length
++], BRW_REGISTER_TYPE_D
),
5152 offset(coordinate
, bld
, i
));
5154 coordinate_done
= true;
5156 case SHADER_OPCODE_TG4_OFFSET
:
5157 /* More crazy intermixing */
5158 for (unsigned i
= 0; i
< 2; i
++) /* u, v */
5159 bld
.MOV(sources
[length
++], offset(coordinate
, bld
, i
));
5161 for (unsigned i
= 0; i
< 2; i
++) /* offu, offv */
5162 bld
.MOV(retype(sources
[length
++], BRW_REGISTER_TYPE_D
),
5163 offset(tg4_offset
, bld
, i
));
5165 if (coord_components
== 3) /* r if present */
5166 bld
.MOV(sources
[length
++], offset(coordinate
, bld
, 2));
5168 coordinate_done
= true;
5174 /* Set up the coordinate (except for cases where it was done above) */
5175 if (!coordinate_done
) {
5176 for (unsigned i
= 0; i
< coord_components
; i
++)
5177 bld
.MOV(sources
[length
++], offset(coordinate
, bld
, i
));
5180 if (min_lod
.file
!= BAD_FILE
) {
5181 /* Account for all of the missing coordinate sources */
5182 length
+= 4 - coord_components
;
5183 if (op
== SHADER_OPCODE_TXD
)
5184 length
+= (3 - grad_components
) * 2;
5186 bld
.MOV(sources
[length
++], min_lod
);
5191 mlen
= length
* reg_width
- header_size
;
5193 mlen
= length
* reg_width
;
5195 const fs_reg src_payload
= fs_reg(VGRF
, bld
.shader
->alloc
.allocate(mlen
),
5196 BRW_REGISTER_TYPE_F
);
5197 bld
.LOAD_PAYLOAD(src_payload
, sources
, length
, header_size
);
5199 /* Generate the SEND. */
5200 inst
->opcode
= SHADER_OPCODE_SEND
;
5202 inst
->header_size
= header_size
;
5204 const unsigned msg_type
=
5205 sampler_msg_type(devinfo
, op
, inst
->shadow_compare
);
5206 const unsigned simd_mode
=
5207 inst
->exec_size
<= 8 ? BRW_SAMPLER_SIMD_MODE_SIMD8
:
5208 BRW_SAMPLER_SIMD_MODE_SIMD16
;
5210 uint32_t base_binding_table_index
;
5212 case SHADER_OPCODE_TG4
:
5213 case SHADER_OPCODE_TG4_OFFSET
:
5214 base_binding_table_index
= prog_data
->binding_table
.gather_texture_start
;
5216 case SHADER_OPCODE_IMAGE_SIZE_LOGICAL
:
5217 base_binding_table_index
= prog_data
->binding_table
.image_start
;
5220 base_binding_table_index
= prog_data
->binding_table
.texture_start
;
5224 inst
->sfid
= BRW_SFID_SAMPLER
;
5225 if (surface
.file
== IMM
&&
5226 (sampler
.file
== IMM
|| sampler_handle
.file
!= BAD_FILE
)) {
5227 inst
->desc
= brw_sampler_desc(devinfo
,
5228 surface
.ud
+ base_binding_table_index
,
5229 sampler
.file
== IMM
? sampler
.ud
% 16 : 0,
5232 0 /* return_format unused on gen7+ */);
5233 inst
->src
[0] = brw_imm_ud(0);
5234 inst
->src
[1] = brw_imm_ud(0); /* ex_desc */
5235 } else if (surface_handle
.file
!= BAD_FILE
) {
5236 /* Bindless surface */
5237 assert(devinfo
->gen
>= 9);
5238 inst
->desc
= brw_sampler_desc(devinfo
,
5240 sampler
.file
== IMM
? sampler
.ud
% 16 : 0,
5243 0 /* return_format unused on gen7+ */);
5245 /* For bindless samplers, the entire address is included in the message
5246 * header so we can leave the portion in the message descriptor 0.
5248 if (sampler_handle
.file
!= BAD_FILE
|| sampler
.file
== IMM
) {
5249 inst
->src
[0] = brw_imm_ud(0);
5251 const fs_builder ubld
= bld
.group(1, 0).exec_all();
5252 fs_reg desc
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
5253 ubld
.SHL(desc
, sampler
, brw_imm_ud(8));
5254 inst
->src
[0] = desc
;
5257 /* We assume that the driver provided the handle in the top 20 bits so
5258 * we can use the surface handle directly as the extended descriptor.
5260 inst
->src
[1] = retype(surface_handle
, BRW_REGISTER_TYPE_UD
);
5262 /* Immediate portion of the descriptor */
5263 inst
->desc
= brw_sampler_desc(devinfo
,
5268 0 /* return_format unused on gen7+ */);
5269 const fs_builder ubld
= bld
.group(1, 0).exec_all();
5270 fs_reg desc
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
5271 if (surface
.equals(sampler
)) {
5272 /* This case is common in GL */
5273 ubld
.MUL(desc
, surface
, brw_imm_ud(0x101));
5275 if (sampler_handle
.file
!= BAD_FILE
) {
5276 ubld
.MOV(desc
, surface
);
5277 } else if (sampler
.file
== IMM
) {
5278 ubld
.OR(desc
, surface
, brw_imm_ud(sampler
.ud
<< 8));
5280 ubld
.SHL(desc
, sampler
, brw_imm_ud(8));
5281 ubld
.OR(desc
, desc
, surface
);
5284 if (base_binding_table_index
)
5285 ubld
.ADD(desc
, desc
, brw_imm_ud(base_binding_table_index
));
5286 ubld
.AND(desc
, desc
, brw_imm_ud(0xfff));
5288 inst
->src
[0] = component(desc
, 0);
5289 inst
->src
[1] = brw_imm_ud(0); /* ex_desc */
5292 inst
->src
[2] = src_payload
;
5293 inst
->resize_sources(3);
5296 /* EOT sampler messages don't make sense to split because it would
5297 * involve ending half of the thread early.
5299 assert(inst
->group
== 0);
5300 /* We need to use SENDC for EOT sampler messages */
5301 inst
->check_tdr
= true;
5302 inst
->send_has_side_effects
= true;
5305 /* Message length > MAX_SAMPLER_MESSAGE_SIZE disallowed by hardware. */
5306 assert(inst
->mlen
<= MAX_SAMPLER_MESSAGE_SIZE
);
5310 lower_sampler_logical_send(const fs_builder
&bld
, fs_inst
*inst
, opcode op
)
5312 const gen_device_info
*devinfo
= bld
.shader
->devinfo
;
5313 const fs_reg
&coordinate
= inst
->src
[TEX_LOGICAL_SRC_COORDINATE
];
5314 const fs_reg
&shadow_c
= inst
->src
[TEX_LOGICAL_SRC_SHADOW_C
];
5315 const fs_reg
&lod
= inst
->src
[TEX_LOGICAL_SRC_LOD
];
5316 const fs_reg
&lod2
= inst
->src
[TEX_LOGICAL_SRC_LOD2
];
5317 const fs_reg
&min_lod
= inst
->src
[TEX_LOGICAL_SRC_MIN_LOD
];
5318 const fs_reg
&sample_index
= inst
->src
[TEX_LOGICAL_SRC_SAMPLE_INDEX
];
5319 const fs_reg
&mcs
= inst
->src
[TEX_LOGICAL_SRC_MCS
];
5320 const fs_reg
&surface
= inst
->src
[TEX_LOGICAL_SRC_SURFACE
];
5321 const fs_reg
&sampler
= inst
->src
[TEX_LOGICAL_SRC_SAMPLER
];
5322 const fs_reg
&surface_handle
= inst
->src
[TEX_LOGICAL_SRC_SURFACE_HANDLE
];
5323 const fs_reg
&sampler_handle
= inst
->src
[TEX_LOGICAL_SRC_SAMPLER_HANDLE
];
5324 const fs_reg
&tg4_offset
= inst
->src
[TEX_LOGICAL_SRC_TG4_OFFSET
];
5325 assert(inst
->src
[TEX_LOGICAL_SRC_COORD_COMPONENTS
].file
== IMM
);
5326 const unsigned coord_components
= inst
->src
[TEX_LOGICAL_SRC_COORD_COMPONENTS
].ud
;
5327 assert(inst
->src
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
].file
== IMM
);
5328 const unsigned grad_components
= inst
->src
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
].ud
;
5330 if (devinfo
->gen
>= 7) {
5331 lower_sampler_logical_send_gen7(bld
, inst
, op
, coordinate
,
5332 shadow_c
, lod
, lod2
, min_lod
,
5334 mcs
, surface
, sampler
,
5335 surface_handle
, sampler_handle
,
5337 coord_components
, grad_components
);
5338 } else if (devinfo
->gen
>= 5) {
5339 lower_sampler_logical_send_gen5(bld
, inst
, op
, coordinate
,
5340 shadow_c
, lod
, lod2
, sample_index
,
5342 coord_components
, grad_components
);
5344 lower_sampler_logical_send_gen4(bld
, inst
, op
, coordinate
,
5345 shadow_c
, lod
, lod2
,
5347 coord_components
, grad_components
);
5352 * Predicate the specified instruction on the sample mask.
5355 emit_predicate_on_sample_mask(const fs_builder
&bld
, fs_inst
*inst
)
5357 assert(bld
.shader
->stage
== MESA_SHADER_FRAGMENT
&&
5358 bld
.group() == inst
->group
&&
5359 bld
.dispatch_width() == inst
->exec_size
);
5361 const fs_visitor
*v
= static_cast<const fs_visitor
*>(bld
.shader
);
5362 const fs_reg sample_mask
= sample_mask_reg(bld
);
5363 const unsigned subreg
= sample_mask_flag_subreg(v
);
5365 if (brw_wm_prog_data(v
->stage_prog_data
)->uses_kill
) {
5366 assert(sample_mask
.file
== ARF
&&
5367 sample_mask
.nr
== brw_flag_subreg(subreg
).nr
&&
5368 sample_mask
.subnr
== brw_flag_subreg(
5369 subreg
+ inst
->group
/ 16).subnr
);
5371 bld
.group(1, 0).exec_all()
5372 .MOV(brw_flag_subreg(subreg
+ inst
->group
/ 16), sample_mask
);
5375 if (inst
->predicate
) {
5376 assert(inst
->predicate
== BRW_PREDICATE_NORMAL
);
5377 assert(!inst
->predicate_inverse
);
5378 assert(inst
->flag_subreg
== 0);
5379 /* Combine the sample mask with the existing predicate by using a
5380 * vertical predication mode.
5382 inst
->predicate
= BRW_PREDICATE_ALIGN1_ALLV
;
5384 inst
->flag_subreg
= subreg
;
5385 inst
->predicate
= BRW_PREDICATE_NORMAL
;
5386 inst
->predicate_inverse
= false;
5391 lower_surface_logical_send(const fs_builder
&bld
, fs_inst
*inst
)
5393 const gen_device_info
*devinfo
= bld
.shader
->devinfo
;
5395 /* Get the logical send arguments. */
5396 const fs_reg
&addr
= inst
->src
[SURFACE_LOGICAL_SRC_ADDRESS
];
5397 const fs_reg
&src
= inst
->src
[SURFACE_LOGICAL_SRC_DATA
];
5398 const fs_reg
&surface
= inst
->src
[SURFACE_LOGICAL_SRC_SURFACE
];
5399 const fs_reg
&surface_handle
= inst
->src
[SURFACE_LOGICAL_SRC_SURFACE_HANDLE
];
5400 const UNUSED fs_reg
&dims
= inst
->src
[SURFACE_LOGICAL_SRC_IMM_DIMS
];
5401 const fs_reg
&arg
= inst
->src
[SURFACE_LOGICAL_SRC_IMM_ARG
];
5402 assert(arg
.file
== IMM
);
5404 /* We must have exactly one of surface and surface_handle */
5405 assert((surface
.file
== BAD_FILE
) != (surface_handle
.file
== BAD_FILE
));
5407 /* Calculate the total number of components of the payload. */
5408 const unsigned addr_sz
= inst
->components_read(SURFACE_LOGICAL_SRC_ADDRESS
);
5409 const unsigned src_sz
= inst
->components_read(SURFACE_LOGICAL_SRC_DATA
);
5411 const bool is_typed_access
=
5412 inst
->opcode
== SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
||
5413 inst
->opcode
== SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
||
5414 inst
->opcode
== SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
;
5416 const bool is_surface_access
= is_typed_access
||
5417 inst
->opcode
== SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
||
5418 inst
->opcode
== SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
||
5419 inst
->opcode
== SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL
;
5421 const bool is_stateless
=
5422 surface
.file
== IMM
&& (surface
.ud
== BRW_BTI_STATELESS
||
5423 surface
.ud
== GEN8_BTI_STATELESS_NON_COHERENT
);
5425 const bool has_side_effects
= inst
->has_side_effects();
5426 fs_reg sample_mask
= has_side_effects
? sample_mask_reg(bld
) :
5427 fs_reg(brw_imm_d(0xffff));
5429 /* From the BDW PRM Volume 7, page 147:
5431 * "For the Data Cache Data Port*, the header must be present for the
5432 * following message types: [...] Typed read/write/atomics"
5434 * Earlier generations have a similar wording. Because of this restriction
5435 * we don't attempt to implement sample masks via predication for such
5436 * messages prior to Gen9, since we have to provide a header anyway. On
5437 * Gen11+ the header has been removed so we can only use predication.
5439 * For all stateless A32 messages, we also need a header
5442 if ((devinfo
->gen
< 9 && is_typed_access
) || is_stateless
) {
5443 fs_builder ubld
= bld
.exec_all().group(8, 0);
5444 header
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
5445 ubld
.MOV(header
, brw_imm_d(0));
5447 /* Both the typed and scattered byte/dword A32 messages take a buffer
5448 * base address in R0.5:[31:0] (See MH1_A32_PSM for typed messages or
5449 * MH_A32_GO for byte/dword scattered messages in the SKL PRM Vol. 2d
5450 * for more details.) This is conveniently where the HW places the
5451 * scratch surface base address.
5453 * From the SKL PRM Vol. 7 "Per-Thread Scratch Space":
5455 * "When a thread becomes 'active' it is allocated a portion of
5456 * scratch space, sized according to PerThreadScratchSpace. The
5457 * starting location of each thread’s scratch space allocation,
5458 * ScratchSpaceOffset, is passed in the thread payload in
5459 * R0.5[31:10] and is specified as a 1KB-granular offset from the
5460 * GeneralStateBaseAddress. The computation of ScratchSpaceOffset
5461 * includes the starting address of the stage’s scratch space
5462 * allocation, as programmed by ScratchSpaceBasePointer."
5464 * The base address is passed in bits R0.5[31:10] and the bottom 10
5465 * bits of R0.5 are used for other things. Therefore, we have to
5466 * mask off the bottom 10 bits so that we don't get a garbage base
5469 ubld
.group(1, 0).AND(component(header
, 5),
5470 retype(brw_vec1_grf(0, 5), BRW_REGISTER_TYPE_UD
),
5471 brw_imm_ud(0xfffffc00));
5473 if (is_surface_access
)
5474 ubld
.group(1, 0).MOV(component(header
, 7), sample_mask
);
5476 const unsigned header_sz
= header
.file
!= BAD_FILE
? 1 : 0;
5478 fs_reg payload
, payload2
;
5479 unsigned mlen
, ex_mlen
= 0;
5480 if (devinfo
->gen
>= 9 &&
5481 (src
.file
== BAD_FILE
|| header
.file
== BAD_FILE
)) {
5482 /* We have split sends on gen9 and above */
5483 if (header
.file
== BAD_FILE
) {
5484 payload
= bld
.move_to_vgrf(addr
, addr_sz
);
5485 payload2
= bld
.move_to_vgrf(src
, src_sz
);
5486 mlen
= addr_sz
* (inst
->exec_size
/ 8);
5487 ex_mlen
= src_sz
* (inst
->exec_size
/ 8);
5489 assert(src
.file
== BAD_FILE
);
5491 payload2
= bld
.move_to_vgrf(addr
, addr_sz
);
5493 ex_mlen
= addr_sz
* (inst
->exec_size
/ 8);
5496 /* Allocate space for the payload. */
5497 const unsigned sz
= header_sz
+ addr_sz
+ src_sz
;
5498 payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, sz
);
5499 fs_reg
*const components
= new fs_reg
[sz
];
5502 /* Construct the payload. */
5503 if (header
.file
!= BAD_FILE
)
5504 components
[n
++] = header
;
5506 for (unsigned i
= 0; i
< addr_sz
; i
++)
5507 components
[n
++] = offset(addr
, bld
, i
);
5509 for (unsigned i
= 0; i
< src_sz
; i
++)
5510 components
[n
++] = offset(src
, bld
, i
);
5512 bld
.LOAD_PAYLOAD(payload
, components
, sz
, header_sz
);
5513 mlen
= header_sz
+ (addr_sz
+ src_sz
) * inst
->exec_size
/ 8;
5515 delete[] components
;
5518 /* Predicate the instruction on the sample mask if no header is
5521 if ((header
.file
== BAD_FILE
|| !is_surface_access
) &&
5522 sample_mask
.file
!= BAD_FILE
&& sample_mask
.file
!= IMM
)
5523 emit_predicate_on_sample_mask(bld
, inst
);
5526 switch (inst
->opcode
) {
5527 case SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL
:
5528 case SHADER_OPCODE_BYTE_SCATTERED_READ_LOGICAL
:
5529 /* Byte scattered opcodes go through the normal data cache */
5530 sfid
= GEN7_SFID_DATAPORT_DATA_CACHE
;
5533 case SHADER_OPCODE_DWORD_SCATTERED_READ_LOGICAL
:
5534 case SHADER_OPCODE_DWORD_SCATTERED_WRITE_LOGICAL
:
5535 sfid
= devinfo
->gen
>= 7 ? GEN7_SFID_DATAPORT_DATA_CACHE
:
5536 devinfo
->gen
>= 6 ? GEN6_SFID_DATAPORT_RENDER_CACHE
:
5537 BRW_DATAPORT_READ_TARGET_RENDER_CACHE
;
5540 case SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
:
5541 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
:
5542 case SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL
:
5543 case SHADER_OPCODE_UNTYPED_ATOMIC_FLOAT_LOGICAL
:
5544 /* Untyped Surface messages go through the data cache but the SFID value
5545 * changed on Haswell.
5547 sfid
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
?
5548 HSW_SFID_DATAPORT_DATA_CACHE_1
:
5549 GEN7_SFID_DATAPORT_DATA_CACHE
);
5552 case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
:
5553 case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
:
5554 case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
:
5555 /* Typed surface messages go through the render cache on IVB and the
5556 * data cache on HSW+.
5558 sfid
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
?
5559 HSW_SFID_DATAPORT_DATA_CACHE_1
:
5560 GEN6_SFID_DATAPORT_RENDER_CACHE
);
5564 unreachable("Unsupported surface opcode");
5568 switch (inst
->opcode
) {
5569 case SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
:
5570 desc
= brw_dp_untyped_surface_rw_desc(devinfo
, inst
->exec_size
,
5571 arg
.ud
, /* num_channels */
5575 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
:
5576 desc
= brw_dp_untyped_surface_rw_desc(devinfo
, inst
->exec_size
,
5577 arg
.ud
, /* num_channels */
5581 case SHADER_OPCODE_BYTE_SCATTERED_READ_LOGICAL
:
5582 desc
= brw_dp_byte_scattered_rw_desc(devinfo
, inst
->exec_size
,
5583 arg
.ud
, /* bit_size */
5587 case SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL
:
5588 desc
= brw_dp_byte_scattered_rw_desc(devinfo
, inst
->exec_size
,
5589 arg
.ud
, /* bit_size */
5593 case SHADER_OPCODE_DWORD_SCATTERED_READ_LOGICAL
:
5594 assert(arg
.ud
== 32); /* bit_size */
5595 desc
= brw_dp_dword_scattered_rw_desc(devinfo
, inst
->exec_size
,
5599 case SHADER_OPCODE_DWORD_SCATTERED_WRITE_LOGICAL
:
5600 assert(arg
.ud
== 32); /* bit_size */
5601 desc
= brw_dp_dword_scattered_rw_desc(devinfo
, inst
->exec_size
,
5605 case SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL
:
5606 desc
= brw_dp_untyped_atomic_desc(devinfo
, inst
->exec_size
,
5607 arg
.ud
, /* atomic_op */
5608 !inst
->dst
.is_null());
5611 case SHADER_OPCODE_UNTYPED_ATOMIC_FLOAT_LOGICAL
:
5612 desc
= brw_dp_untyped_atomic_float_desc(devinfo
, inst
->exec_size
,
5613 arg
.ud
, /* atomic_op */
5614 !inst
->dst
.is_null());
5617 case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
:
5618 desc
= brw_dp_typed_surface_rw_desc(devinfo
, inst
->exec_size
, inst
->group
,
5619 arg
.ud
, /* num_channels */
5623 case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
:
5624 desc
= brw_dp_typed_surface_rw_desc(devinfo
, inst
->exec_size
, inst
->group
,
5625 arg
.ud
, /* num_channels */
5629 case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
:
5630 desc
= brw_dp_typed_atomic_desc(devinfo
, inst
->exec_size
, inst
->group
,
5631 arg
.ud
, /* atomic_op */
5632 !inst
->dst
.is_null());
5636 unreachable("Unknown surface logical instruction");
5639 /* Update the original instruction. */
5640 inst
->opcode
= SHADER_OPCODE_SEND
;
5642 inst
->ex_mlen
= ex_mlen
;
5643 inst
->header_size
= header_sz
;
5644 inst
->send_has_side_effects
= has_side_effects
;
5645 inst
->send_is_volatile
= !has_side_effects
;
5647 /* Set up SFID and descriptors */
5650 if (surface
.file
== IMM
) {
5651 inst
->desc
|= surface
.ud
& 0xff;
5652 inst
->src
[0] = brw_imm_ud(0);
5653 inst
->src
[1] = brw_imm_ud(0); /* ex_desc */
5654 } else if (surface_handle
.file
!= BAD_FILE
) {
5655 /* Bindless surface */
5656 assert(devinfo
->gen
>= 9);
5657 inst
->desc
|= GEN9_BTI_BINDLESS
;
5658 inst
->src
[0] = brw_imm_ud(0);
5660 /* We assume that the driver provided the handle in the top 20 bits so
5661 * we can use the surface handle directly as the extended descriptor.
5663 inst
->src
[1] = retype(surface_handle
, BRW_REGISTER_TYPE_UD
);
5665 const fs_builder ubld
= bld
.exec_all().group(1, 0);
5666 fs_reg tmp
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
5667 ubld
.AND(tmp
, surface
, brw_imm_ud(0xff));
5668 inst
->src
[0] = component(tmp
, 0);
5669 inst
->src
[1] = brw_imm_ud(0); /* ex_desc */
5672 /* Finally, the payload */
5673 inst
->src
[2] = payload
;
5674 inst
->src
[3] = payload2
;
5676 inst
->resize_sources(4);
5680 lower_a64_logical_send(const fs_builder
&bld
, fs_inst
*inst
)
5682 const gen_device_info
*devinfo
= bld
.shader
->devinfo
;
5684 const fs_reg
&addr
= inst
->src
[0];
5685 const fs_reg
&src
= inst
->src
[1];
5686 const unsigned src_comps
= inst
->components_read(1);
5687 assert(inst
->src
[2].file
== IMM
);
5688 const unsigned arg
= inst
->src
[2].ud
;
5689 const bool has_side_effects
= inst
->has_side_effects();
5691 /* If the surface message has side effects and we're a fragment shader, we
5692 * have to predicate with the sample mask to avoid helper invocations.
5694 if (has_side_effects
&& bld
.shader
->stage
== MESA_SHADER_FRAGMENT
)
5695 emit_predicate_on_sample_mask(bld
, inst
);
5697 fs_reg payload
, payload2
;
5698 unsigned mlen
, ex_mlen
= 0;
5699 if (devinfo
->gen
>= 9) {
5700 /* On Skylake and above, we have SENDS */
5701 mlen
= 2 * (inst
->exec_size
/ 8);
5702 ex_mlen
= src_comps
* type_sz(src
.type
) * inst
->exec_size
/ REG_SIZE
;
5703 payload
= retype(bld
.move_to_vgrf(addr
, 1), BRW_REGISTER_TYPE_UD
);
5704 payload2
= retype(bld
.move_to_vgrf(src
, src_comps
),
5705 BRW_REGISTER_TYPE_UD
);
5707 /* Add two because the address is 64-bit */
5708 const unsigned dwords
= 2 + src_comps
;
5709 mlen
= dwords
* (inst
->exec_size
/ 8);
5715 for (unsigned i
= 0; i
< src_comps
; i
++)
5716 sources
[1 + i
] = offset(src
, bld
, i
);
5718 payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, dwords
);
5719 bld
.LOAD_PAYLOAD(payload
, sources
, 1 + src_comps
, 0);
5723 switch (inst
->opcode
) {
5724 case SHADER_OPCODE_A64_UNTYPED_READ_LOGICAL
:
5725 desc
= brw_dp_a64_untyped_surface_rw_desc(devinfo
, inst
->exec_size
,
5726 arg
, /* num_channels */
5730 case SHADER_OPCODE_A64_UNTYPED_WRITE_LOGICAL
:
5731 desc
= brw_dp_a64_untyped_surface_rw_desc(devinfo
, inst
->exec_size
,
5732 arg
, /* num_channels */
5736 case SHADER_OPCODE_A64_BYTE_SCATTERED_READ_LOGICAL
:
5737 desc
= brw_dp_a64_byte_scattered_rw_desc(devinfo
, inst
->exec_size
,
5742 case SHADER_OPCODE_A64_BYTE_SCATTERED_WRITE_LOGICAL
:
5743 desc
= brw_dp_a64_byte_scattered_rw_desc(devinfo
, inst
->exec_size
,
5748 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_LOGICAL
:
5749 desc
= brw_dp_a64_untyped_atomic_desc(devinfo
, inst
->exec_size
, 32,
5750 arg
, /* atomic_op */
5751 !inst
->dst
.is_null());
5754 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_INT64_LOGICAL
:
5755 desc
= brw_dp_a64_untyped_atomic_desc(devinfo
, inst
->exec_size
, 64,
5756 arg
, /* atomic_op */
5757 !inst
->dst
.is_null());
5761 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_FLOAT_LOGICAL
:
5762 desc
= brw_dp_a64_untyped_atomic_float_desc(devinfo
, inst
->exec_size
,
5763 arg
, /* atomic_op */
5764 !inst
->dst
.is_null());
5768 unreachable("Unknown A64 logical instruction");
5771 /* Update the original instruction. */
5772 inst
->opcode
= SHADER_OPCODE_SEND
;
5774 inst
->ex_mlen
= ex_mlen
;
5775 inst
->header_size
= 0;
5776 inst
->send_has_side_effects
= has_side_effects
;
5777 inst
->send_is_volatile
= !has_side_effects
;
5779 /* Set up SFID and descriptors */
5780 inst
->sfid
= HSW_SFID_DATAPORT_DATA_CACHE_1
;
5782 inst
->resize_sources(4);
5783 inst
->src
[0] = brw_imm_ud(0); /* desc */
5784 inst
->src
[1] = brw_imm_ud(0); /* ex_desc */
5785 inst
->src
[2] = payload
;
5786 inst
->src
[3] = payload2
;
5790 lower_varying_pull_constant_logical_send(const fs_builder
&bld
, fs_inst
*inst
)
5792 const gen_device_info
*devinfo
= bld
.shader
->devinfo
;
5794 if (devinfo
->gen
>= 7) {
5795 fs_reg index
= inst
->src
[0];
5796 /* We are switching the instruction from an ALU-like instruction to a
5797 * send-from-grf instruction. Since sends can't handle strides or
5798 * source modifiers, we have to make a copy of the offset source.
5800 fs_reg offset
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
5801 bld
.MOV(offset
, inst
->src
[1]);
5803 const unsigned simd_mode
=
5804 inst
->exec_size
<= 8 ? BRW_SAMPLER_SIMD_MODE_SIMD8
:
5805 BRW_SAMPLER_SIMD_MODE_SIMD16
;
5807 inst
->opcode
= SHADER_OPCODE_SEND
;
5808 inst
->mlen
= inst
->exec_size
/ 8;
5809 inst
->resize_sources(3);
5811 inst
->sfid
= BRW_SFID_SAMPLER
;
5812 inst
->desc
= brw_sampler_desc(devinfo
, 0, 0,
5813 GEN5_SAMPLER_MESSAGE_SAMPLE_LD
,
5815 if (index
.file
== IMM
) {
5816 inst
->desc
|= index
.ud
& 0xff;
5817 inst
->src
[0] = brw_imm_ud(0);
5819 const fs_builder ubld
= bld
.exec_all().group(1, 0);
5820 fs_reg tmp
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
5821 ubld
.AND(tmp
, index
, brw_imm_ud(0xff));
5822 inst
->src
[0] = component(tmp
, 0);
5824 inst
->src
[1] = brw_imm_ud(0); /* ex_desc */
5825 inst
->src
[2] = offset
; /* payload */
5827 const fs_reg
payload(MRF
, FIRST_PULL_LOAD_MRF(devinfo
->gen
),
5828 BRW_REGISTER_TYPE_UD
);
5830 bld
.MOV(byte_offset(payload
, REG_SIZE
), inst
->src
[1]);
5832 inst
->opcode
= FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN4
;
5833 inst
->resize_sources(1);
5834 inst
->base_mrf
= payload
.nr
;
5835 inst
->header_size
= 1;
5836 inst
->mlen
= 1 + inst
->exec_size
/ 8;
5841 lower_math_logical_send(const fs_builder
&bld
, fs_inst
*inst
)
5843 assert(bld
.shader
->devinfo
->gen
< 6);
5846 inst
->mlen
= inst
->sources
* inst
->exec_size
/ 8;
5848 if (inst
->sources
> 1) {
5849 /* From the Ironlake PRM, Volume 4, Part 1, Section 6.1.13
5850 * "Message Payload":
5852 * "Operand0[7]. For the INT DIV functions, this operand is the
5855 * "Operand1[7]. For the INT DIV functions, this operand is the
5858 const bool is_int_div
= inst
->opcode
!= SHADER_OPCODE_POW
;
5859 const fs_reg src0
= is_int_div
? inst
->src
[1] : inst
->src
[0];
5860 const fs_reg src1
= is_int_div
? inst
->src
[0] : inst
->src
[1];
5862 inst
->resize_sources(1);
5863 inst
->src
[0] = src0
;
5865 assert(inst
->exec_size
== 8);
5866 bld
.MOV(fs_reg(MRF
, inst
->base_mrf
+ 1, src1
.type
), src1
);
5871 fs_visitor::lower_logical_sends()
5873 bool progress
= false;
5875 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
5876 const fs_builder
ibld(this, block
, inst
);
5878 switch (inst
->opcode
) {
5879 case FS_OPCODE_FB_WRITE_LOGICAL
:
5880 assert(stage
== MESA_SHADER_FRAGMENT
);
5881 lower_fb_write_logical_send(ibld
, inst
,
5882 brw_wm_prog_data(prog_data
),
5883 (const brw_wm_prog_key
*)key
,
5887 case FS_OPCODE_FB_READ_LOGICAL
:
5888 lower_fb_read_logical_send(ibld
, inst
);
5891 case SHADER_OPCODE_TEX_LOGICAL
:
5892 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TEX
);
5895 case SHADER_OPCODE_TXD_LOGICAL
:
5896 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXD
);
5899 case SHADER_OPCODE_TXF_LOGICAL
:
5900 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF
);
5903 case SHADER_OPCODE_TXL_LOGICAL
:
5904 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXL
);
5907 case SHADER_OPCODE_TXS_LOGICAL
:
5908 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXS
);
5911 case SHADER_OPCODE_IMAGE_SIZE_LOGICAL
:
5912 lower_sampler_logical_send(ibld
, inst
,
5913 SHADER_OPCODE_IMAGE_SIZE_LOGICAL
);
5916 case FS_OPCODE_TXB_LOGICAL
:
5917 lower_sampler_logical_send(ibld
, inst
, FS_OPCODE_TXB
);
5920 case SHADER_OPCODE_TXF_CMS_LOGICAL
:
5921 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_CMS
);
5924 case SHADER_OPCODE_TXF_CMS_W_LOGICAL
:
5925 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_CMS_W
);
5928 case SHADER_OPCODE_TXF_UMS_LOGICAL
:
5929 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_UMS
);
5932 case SHADER_OPCODE_TXF_MCS_LOGICAL
:
5933 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TXF_MCS
);
5936 case SHADER_OPCODE_LOD_LOGICAL
:
5937 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_LOD
);
5940 case SHADER_OPCODE_TG4_LOGICAL
:
5941 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TG4
);
5944 case SHADER_OPCODE_TG4_OFFSET_LOGICAL
:
5945 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_TG4_OFFSET
);
5948 case SHADER_OPCODE_SAMPLEINFO_LOGICAL
:
5949 lower_sampler_logical_send(ibld
, inst
, SHADER_OPCODE_SAMPLEINFO
);
5952 case SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
:
5953 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
:
5954 case SHADER_OPCODE_BYTE_SCATTERED_READ_LOGICAL
:
5955 case SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL
:
5956 case SHADER_OPCODE_DWORD_SCATTERED_READ_LOGICAL
:
5957 case SHADER_OPCODE_DWORD_SCATTERED_WRITE_LOGICAL
:
5958 case SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL
:
5959 case SHADER_OPCODE_UNTYPED_ATOMIC_FLOAT_LOGICAL
:
5960 case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
:
5961 case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
:
5962 case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
:
5963 lower_surface_logical_send(ibld
, inst
);
5966 case SHADER_OPCODE_A64_UNTYPED_WRITE_LOGICAL
:
5967 case SHADER_OPCODE_A64_UNTYPED_READ_LOGICAL
:
5968 case SHADER_OPCODE_A64_BYTE_SCATTERED_WRITE_LOGICAL
:
5969 case SHADER_OPCODE_A64_BYTE_SCATTERED_READ_LOGICAL
:
5970 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_LOGICAL
:
5971 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_INT64_LOGICAL
:
5972 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_FLOAT_LOGICAL
:
5973 lower_a64_logical_send(ibld
, inst
);
5976 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_LOGICAL
:
5977 lower_varying_pull_constant_logical_send(ibld
, inst
);
5980 case SHADER_OPCODE_RCP
:
5981 case SHADER_OPCODE_RSQ
:
5982 case SHADER_OPCODE_SQRT
:
5983 case SHADER_OPCODE_EXP2
:
5984 case SHADER_OPCODE_LOG2
:
5985 case SHADER_OPCODE_SIN
:
5986 case SHADER_OPCODE_COS
:
5987 case SHADER_OPCODE_POW
:
5988 case SHADER_OPCODE_INT_QUOTIENT
:
5989 case SHADER_OPCODE_INT_REMAINDER
:
5990 /* The math opcodes are overloaded for the send-like and
5991 * expression-like instructions which seems kind of icky. Gen6+ has
5992 * a native (but rather quirky) MATH instruction so we don't need to
5993 * do anything here. On Gen4-5 we'll have to lower the Gen6-like
5994 * logical instructions (which we can easily recognize because they
5995 * have mlen = 0) into send-like virtual instructions.
5997 if (devinfo
->gen
< 6 && inst
->mlen
== 0) {
5998 lower_math_logical_send(ibld
, inst
);
6013 invalidate_analysis(DEPENDENCY_INSTRUCTIONS
| DEPENDENCY_VARIABLES
);
6019 is_mixed_float_with_fp32_dst(const fs_inst
*inst
)
6021 /* This opcode sometimes uses :W type on the source even if the operand is
6022 * a :HF, because in gen7 there is no support for :HF, and thus it uses :W.
6024 if (inst
->opcode
== BRW_OPCODE_F16TO32
)
6027 if (inst
->dst
.type
!= BRW_REGISTER_TYPE_F
)
6030 for (int i
= 0; i
< inst
->sources
; i
++) {
6031 if (inst
->src
[i
].type
== BRW_REGISTER_TYPE_HF
)
6039 is_mixed_float_with_packed_fp16_dst(const fs_inst
*inst
)
6041 /* This opcode sometimes uses :W type on the destination even if the
6042 * destination is a :HF, because in gen7 there is no support for :HF, and
6045 if (inst
->opcode
== BRW_OPCODE_F32TO16
&&
6046 inst
->dst
.stride
== 1)
6049 if (inst
->dst
.type
!= BRW_REGISTER_TYPE_HF
||
6050 inst
->dst
.stride
!= 1)
6053 for (int i
= 0; i
< inst
->sources
; i
++) {
6054 if (inst
->src
[i
].type
== BRW_REGISTER_TYPE_F
)
6062 * Get the closest allowed SIMD width for instruction \p inst accounting for
6063 * some common regioning and execution control restrictions that apply to FPU
6064 * instructions. These restrictions don't necessarily have any relevance to
6065 * instructions not executed by the FPU pipeline like extended math, control
6066 * flow or send message instructions.
6068 * For virtual opcodes it's really up to the instruction -- In some cases
6069 * (e.g. where a virtual instruction unrolls into a simple sequence of FPU
6070 * instructions) it may simplify virtual instruction lowering if we can
6071 * enforce FPU-like regioning restrictions already on the virtual instruction,
6072 * in other cases (e.g. virtual send-like instructions) this may be
6073 * excessively restrictive.
6076 get_fpu_lowered_simd_width(const struct gen_device_info
*devinfo
,
6077 const fs_inst
*inst
)
6079 /* Maximum execution size representable in the instruction controls. */
6080 unsigned max_width
= MIN2(32, inst
->exec_size
);
6082 /* According to the PRMs:
6083 * "A. In Direct Addressing mode, a source cannot span more than 2
6084 * adjacent GRF registers.
6085 * B. A destination cannot span more than 2 adjacent GRF registers."
6087 * Look for the source or destination with the largest register region
6088 * which is the one that is going to limit the overall execution size of
6089 * the instruction due to this rule.
6091 unsigned reg_count
= DIV_ROUND_UP(inst
->size_written
, REG_SIZE
);
6093 for (unsigned i
= 0; i
< inst
->sources
; i
++)
6094 reg_count
= MAX2(reg_count
, DIV_ROUND_UP(inst
->size_read(i
), REG_SIZE
));
6096 /* Calculate the maximum execution size of the instruction based on the
6097 * factor by which it goes over the hardware limit of 2 GRFs.
6100 max_width
= MIN2(max_width
, inst
->exec_size
/ DIV_ROUND_UP(reg_count
, 2));
6102 /* According to the IVB PRMs:
6103 * "When destination spans two registers, the source MUST span two
6104 * registers. The exception to the above rule:
6106 * - When source is scalar, the source registers are not incremented.
6107 * - When source is packed integer Word and destination is packed
6108 * integer DWord, the source register is not incremented but the
6109 * source sub register is incremented."
6111 * The hardware specs from Gen4 to Gen7.5 mention similar regioning
6112 * restrictions. The code below intentionally doesn't check whether the
6113 * destination type is integer because empirically the hardware doesn't
6114 * seem to care what the actual type is as long as it's dword-aligned.
6116 if (devinfo
->gen
< 8) {
6117 for (unsigned i
= 0; i
< inst
->sources
; i
++) {
6118 /* IVB implements DF scalars as <0;2,1> regions. */
6119 const bool is_scalar_exception
= is_uniform(inst
->src
[i
]) &&
6120 (devinfo
->is_haswell
|| type_sz(inst
->src
[i
].type
) != 8);
6121 const bool is_packed_word_exception
=
6122 type_sz(inst
->dst
.type
) == 4 && inst
->dst
.stride
== 1 &&
6123 type_sz(inst
->src
[i
].type
) == 2 && inst
->src
[i
].stride
== 1;
6125 /* We check size_read(i) against size_written instead of REG_SIZE
6126 * because we want to properly handle SIMD32. In SIMD32, you can end
6127 * up with writes to 4 registers and a source that reads 2 registers
6128 * and we may still need to lower all the way to SIMD8 in that case.
6130 if (inst
->size_written
> REG_SIZE
&&
6131 inst
->size_read(i
) != 0 &&
6132 inst
->size_read(i
) < inst
->size_written
&&
6133 !is_scalar_exception
&& !is_packed_word_exception
) {
6134 const unsigned reg_count
= DIV_ROUND_UP(inst
->size_written
, REG_SIZE
);
6135 max_width
= MIN2(max_width
, inst
->exec_size
/ reg_count
);
6140 if (devinfo
->gen
< 6) {
6141 /* From the G45 PRM, Volume 4 Page 361:
6143 * "Operand Alignment Rule: With the exceptions listed below, a
6144 * source/destination operand in general should be aligned to even
6145 * 256-bit physical register with a region size equal to two 256-bit
6146 * physical registers."
6148 * Normally we enforce this by allocating virtual registers to the
6149 * even-aligned class. But we need to handle payload registers.
6151 for (unsigned i
= 0; i
< inst
->sources
; i
++) {
6152 if (inst
->src
[i
].file
== FIXED_GRF
&& (inst
->src
[i
].nr
& 1) &&
6153 inst
->size_read(i
) > REG_SIZE
) {
6154 max_width
= MIN2(max_width
, 8);
6159 /* From the IVB PRMs:
6160 * "When an instruction is SIMD32, the low 16 bits of the execution mask
6161 * are applied for both halves of the SIMD32 instruction. If different
6162 * execution mask channels are required, split the instruction into two
6163 * SIMD16 instructions."
6165 * There is similar text in the HSW PRMs. Gen4-6 don't even implement
6166 * 32-wide control flow support in hardware and will behave similarly.
6168 if (devinfo
->gen
< 8 && !inst
->force_writemask_all
)
6169 max_width
= MIN2(max_width
, 16);
6171 /* From the IVB PRMs (applies to HSW too):
6172 * "Instructions with condition modifiers must not use SIMD32."
6174 * From the BDW PRMs (applies to later hardware too):
6175 * "Ternary instruction with condition modifiers must not use SIMD32."
6177 if (inst
->conditional_mod
&& (devinfo
->gen
< 8 || inst
->is_3src(devinfo
)))
6178 max_width
= MIN2(max_width
, 16);
6180 /* From the IVB PRMs (applies to other devices that don't have the
6181 * gen_device_info::supports_simd16_3src flag set):
6182 * "In Align16 access mode, SIMD16 is not allowed for DW operations and
6183 * SIMD8 is not allowed for DF operations."
6185 if (inst
->is_3src(devinfo
) && !devinfo
->supports_simd16_3src
)
6186 max_width
= MIN2(max_width
, inst
->exec_size
/ reg_count
);
6188 /* Pre-Gen8 EUs are hardwired to use the QtrCtrl+1 (where QtrCtrl is
6189 * the 8-bit quarter of the execution mask signals specified in the
6190 * instruction control fields) for the second compressed half of any
6191 * single-precision instruction (for double-precision instructions
6192 * it's hardwired to use NibCtrl+1, at least on HSW), which means that
6193 * the EU will apply the wrong execution controls for the second
6194 * sequential GRF write if the number of channels per GRF is not exactly
6195 * eight in single-precision mode (or four in double-float mode).
6197 * In this situation we calculate the maximum size of the split
6198 * instructions so they only ever write to a single register.
6200 if (devinfo
->gen
< 8 && inst
->size_written
> REG_SIZE
&&
6201 !inst
->force_writemask_all
) {
6202 const unsigned channels_per_grf
= inst
->exec_size
/
6203 DIV_ROUND_UP(inst
->size_written
, REG_SIZE
);
6204 const unsigned exec_type_size
= get_exec_type_size(inst
);
6205 assert(exec_type_size
);
6207 /* The hardware shifts exactly 8 channels per compressed half of the
6208 * instruction in single-precision mode and exactly 4 in double-precision.
6210 if (channels_per_grf
!= (exec_type_size
== 8 ? 4 : 8))
6211 max_width
= MIN2(max_width
, channels_per_grf
);
6213 /* Lower all non-force_writemask_all DF instructions to SIMD4 on IVB/BYT
6214 * because HW applies the same channel enable signals to both halves of
6215 * the compressed instruction which will be just wrong under
6216 * non-uniform control flow.
6218 if (devinfo
->gen
== 7 && !devinfo
->is_haswell
&&
6219 (exec_type_size
== 8 || type_sz(inst
->dst
.type
) == 8))
6220 max_width
= MIN2(max_width
, 4);
6223 /* From the SKL PRM, Special Restrictions for Handling Mixed Mode
6226 * "No SIMD16 in mixed mode when destination is f32. Instruction
6227 * execution size must be no more than 8."
6229 * FIXME: the simulator doesn't seem to complain if we don't do this and
6230 * empirical testing with existing CTS tests show that they pass just fine
6231 * without implementing this, however, since our interpretation of the PRM
6232 * is that conversion MOVs between HF and F are still mixed-float
6233 * instructions (and therefore subject to this restriction) we decided to
6234 * split them to be safe. Might be useful to do additional investigation to
6235 * lift the restriction if we can ensure that it is safe though, since these
6236 * conversions are common when half-float types are involved since many
6237 * instructions do not support HF types and conversions from/to F are
6240 if (is_mixed_float_with_fp32_dst(inst
))
6241 max_width
= MIN2(max_width
, 8);
6243 /* From the SKL PRM, Special Restrictions for Handling Mixed Mode
6246 * "No SIMD16 in mixed mode when destination is packed f16 for both
6247 * Align1 and Align16."
6249 if (is_mixed_float_with_packed_fp16_dst(inst
))
6250 max_width
= MIN2(max_width
, 8);
6252 /* Only power-of-two execution sizes are representable in the instruction
6255 return 1 << _mesa_logbase2(max_width
);
6259 * Get the maximum allowed SIMD width for instruction \p inst accounting for
6260 * various payload size restrictions that apply to sampler message
6263 * This is only intended to provide a maximum theoretical bound for the
6264 * execution size of the message based on the number of argument components
6265 * alone, which in most cases will determine whether the SIMD8 or SIMD16
6266 * variant of the message can be used, though some messages may have
6267 * additional restrictions not accounted for here (e.g. pre-ILK hardware uses
6268 * the message length to determine the exact SIMD width and argument count,
6269 * which makes a number of sampler message combinations impossible to
6273 get_sampler_lowered_simd_width(const struct gen_device_info
*devinfo
,
6274 const fs_inst
*inst
)
6276 /* If we have a min_lod parameter on anything other than a simple sample
6277 * message, it will push it over 5 arguments and we have to fall back to
6280 if (inst
->opcode
!= SHADER_OPCODE_TEX
&&
6281 inst
->components_read(TEX_LOGICAL_SRC_MIN_LOD
))
6284 /* Calculate the number of coordinate components that have to be present
6285 * assuming that additional arguments follow the texel coordinates in the
6286 * message payload. On IVB+ there is no need for padding, on ILK-SNB we
6287 * need to pad to four or three components depending on the message,
6288 * pre-ILK we need to pad to at most three components.
6290 const unsigned req_coord_components
=
6291 (devinfo
->gen
>= 7 ||
6292 !inst
->components_read(TEX_LOGICAL_SRC_COORDINATE
)) ? 0 :
6293 (devinfo
->gen
>= 5 && inst
->opcode
!= SHADER_OPCODE_TXF_LOGICAL
&&
6294 inst
->opcode
!= SHADER_OPCODE_TXF_CMS_LOGICAL
) ? 4 :
6297 /* On Gen9+ the LOD argument is for free if we're able to use the LZ
6298 * variant of the TXL or TXF message.
6300 const bool implicit_lod
= devinfo
->gen
>= 9 &&
6301 (inst
->opcode
== SHADER_OPCODE_TXL
||
6302 inst
->opcode
== SHADER_OPCODE_TXF
) &&
6303 inst
->src
[TEX_LOGICAL_SRC_LOD
].is_zero();
6305 /* Calculate the total number of argument components that need to be passed
6306 * to the sampler unit.
6308 const unsigned num_payload_components
=
6309 MAX2(inst
->components_read(TEX_LOGICAL_SRC_COORDINATE
),
6310 req_coord_components
) +
6311 inst
->components_read(TEX_LOGICAL_SRC_SHADOW_C
) +
6312 (implicit_lod
? 0 : inst
->components_read(TEX_LOGICAL_SRC_LOD
)) +
6313 inst
->components_read(TEX_LOGICAL_SRC_LOD2
) +
6314 inst
->components_read(TEX_LOGICAL_SRC_SAMPLE_INDEX
) +
6315 (inst
->opcode
== SHADER_OPCODE_TG4_OFFSET_LOGICAL
?
6316 inst
->components_read(TEX_LOGICAL_SRC_TG4_OFFSET
) : 0) +
6317 inst
->components_read(TEX_LOGICAL_SRC_MCS
);
6319 /* SIMD16 messages with more than five arguments exceed the maximum message
6320 * size supported by the sampler, regardless of whether a header is
6323 return MIN2(inst
->exec_size
,
6324 num_payload_components
> MAX_SAMPLER_MESSAGE_SIZE
/ 2 ? 8 : 16);
6328 * Get the closest native SIMD width supported by the hardware for instruction
6329 * \p inst. The instruction will be left untouched by
6330 * fs_visitor::lower_simd_width() if the returned value is equal to the
6331 * original execution size.
6334 get_lowered_simd_width(const struct gen_device_info
*devinfo
,
6335 const fs_inst
*inst
)
6337 switch (inst
->opcode
) {
6338 case BRW_OPCODE_MOV
:
6339 case BRW_OPCODE_SEL
:
6340 case BRW_OPCODE_NOT
:
6341 case BRW_OPCODE_AND
:
6343 case BRW_OPCODE_XOR
:
6344 case BRW_OPCODE_SHR
:
6345 case BRW_OPCODE_SHL
:
6346 case BRW_OPCODE_ASR
:
6347 case BRW_OPCODE_ROR
:
6348 case BRW_OPCODE_ROL
:
6349 case BRW_OPCODE_CMPN
:
6350 case BRW_OPCODE_CSEL
:
6351 case BRW_OPCODE_F32TO16
:
6352 case BRW_OPCODE_F16TO32
:
6353 case BRW_OPCODE_BFREV
:
6354 case BRW_OPCODE_BFE
:
6355 case BRW_OPCODE_ADD
:
6356 case BRW_OPCODE_MUL
:
6357 case BRW_OPCODE_AVG
:
6358 case BRW_OPCODE_FRC
:
6359 case BRW_OPCODE_RNDU
:
6360 case BRW_OPCODE_RNDD
:
6361 case BRW_OPCODE_RNDE
:
6362 case BRW_OPCODE_RNDZ
:
6363 case BRW_OPCODE_LZD
:
6364 case BRW_OPCODE_FBH
:
6365 case BRW_OPCODE_FBL
:
6366 case BRW_OPCODE_CBIT
:
6367 case BRW_OPCODE_SAD2
:
6368 case BRW_OPCODE_MAD
:
6369 case BRW_OPCODE_LRP
:
6370 case FS_OPCODE_PACK
:
6371 case SHADER_OPCODE_SEL_EXEC
:
6372 case SHADER_OPCODE_CLUSTER_BROADCAST
:
6373 return get_fpu_lowered_simd_width(devinfo
, inst
);
6375 case BRW_OPCODE_CMP
: {
6376 /* The Ivybridge/BayTrail WaCMPInstFlagDepClearedEarly workaround says that
6377 * when the destination is a GRF the dependency-clear bit on the flag
6378 * register is cleared early.
6380 * Suggested workarounds are to disable coissuing CMP instructions
6381 * or to split CMP(16) instructions into two CMP(8) instructions.
6383 * We choose to split into CMP(8) instructions since disabling
6384 * coissuing would affect CMP instructions not otherwise affected by
6387 const unsigned max_width
= (devinfo
->gen
== 7 && !devinfo
->is_haswell
&&
6388 !inst
->dst
.is_null() ? 8 : ~0);
6389 return MIN2(max_width
, get_fpu_lowered_simd_width(devinfo
, inst
));
6391 case BRW_OPCODE_BFI1
:
6392 case BRW_OPCODE_BFI2
:
6393 /* The Haswell WaForceSIMD8ForBFIInstruction workaround says that we
6395 * "Force BFI instructions to be executed always in SIMD8."
6397 return MIN2(devinfo
->is_haswell
? 8 : ~0u,
6398 get_fpu_lowered_simd_width(devinfo
, inst
));
6401 assert(inst
->src
[0].file
== BAD_FILE
|| inst
->exec_size
<= 16);
6402 return inst
->exec_size
;
6404 case SHADER_OPCODE_RCP
:
6405 case SHADER_OPCODE_RSQ
:
6406 case SHADER_OPCODE_SQRT
:
6407 case SHADER_OPCODE_EXP2
:
6408 case SHADER_OPCODE_LOG2
:
6409 case SHADER_OPCODE_SIN
:
6410 case SHADER_OPCODE_COS
: {
6411 /* Unary extended math instructions are limited to SIMD8 on Gen4 and
6412 * Gen6. Extended Math Function is limited to SIMD8 with half-float.
6414 if (devinfo
->gen
== 6 || (devinfo
->gen
== 4 && !devinfo
->is_g4x
))
6415 return MIN2(8, inst
->exec_size
);
6416 if (inst
->dst
.type
== BRW_REGISTER_TYPE_HF
)
6417 return MIN2(8, inst
->exec_size
);
6418 return MIN2(16, inst
->exec_size
);
6421 case SHADER_OPCODE_POW
: {
6422 /* SIMD16 is only allowed on Gen7+. Extended Math Function is limited
6423 * to SIMD8 with half-float
6425 if (devinfo
->gen
< 7)
6426 return MIN2(8, inst
->exec_size
);
6427 if (inst
->dst
.type
== BRW_REGISTER_TYPE_HF
)
6428 return MIN2(8, inst
->exec_size
);
6429 return MIN2(16, inst
->exec_size
);
6432 case SHADER_OPCODE_USUB_SAT
:
6433 case SHADER_OPCODE_ISUB_SAT
:
6434 return get_fpu_lowered_simd_width(devinfo
, inst
);
6436 case SHADER_OPCODE_INT_QUOTIENT
:
6437 case SHADER_OPCODE_INT_REMAINDER
:
6438 /* Integer division is limited to SIMD8 on all generations. */
6439 return MIN2(8, inst
->exec_size
);
6441 case FS_OPCODE_LINTERP
:
6442 case SHADER_OPCODE_GET_BUFFER_SIZE
:
6443 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
6444 case FS_OPCODE_PACK_HALF_2x16_SPLIT
:
6445 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
6446 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
6447 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
6448 return MIN2(16, inst
->exec_size
);
6450 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_LOGICAL
:
6451 /* Pre-ILK hardware doesn't have a SIMD8 variant of the texel fetch
6452 * message used to implement varying pull constant loads, so expand it
6453 * to SIMD16. An alternative with longer message payload length but
6454 * shorter return payload would be to use the SIMD8 sampler message that
6455 * takes (header, u, v, r) as parameters instead of (header, u).
6457 return (devinfo
->gen
== 4 ? 16 : MIN2(16, inst
->exec_size
));
6459 case FS_OPCODE_DDX_COARSE
:
6460 case FS_OPCODE_DDX_FINE
:
6461 case FS_OPCODE_DDY_COARSE
:
6462 case FS_OPCODE_DDY_FINE
:
6463 /* The implementation of this virtual opcode may require emitting
6464 * compressed Align16 instructions, which are severely limited on some
6467 * From the Ivy Bridge PRM, volume 4 part 3, section 3.3.9 (Register
6468 * Region Restrictions):
6470 * "In Align16 access mode, SIMD16 is not allowed for DW operations
6471 * and SIMD8 is not allowed for DF operations."
6473 * In this context, "DW operations" means "operations acting on 32-bit
6474 * values", so it includes operations on floats.
6476 * Gen4 has a similar restriction. From the i965 PRM, section 11.5.3
6477 * (Instruction Compression -> Rules and Restrictions):
6479 * "A compressed instruction must be in Align1 access mode. Align16
6480 * mode instructions cannot be compressed."
6482 * Similar text exists in the g45 PRM.
6484 * Empirically, compressed align16 instructions using odd register
6485 * numbers don't appear to work on Sandybridge either.
6487 return (devinfo
->gen
== 4 || devinfo
->gen
== 6 ||
6488 (devinfo
->gen
== 7 && !devinfo
->is_haswell
) ?
6489 MIN2(8, inst
->exec_size
) : MIN2(16, inst
->exec_size
));
6491 case SHADER_OPCODE_MULH
:
6492 /* MULH is lowered to the MUL/MACH sequence using the accumulator, which
6493 * is 8-wide on Gen7+.
6495 return (devinfo
->gen
>= 7 ? 8 :
6496 get_fpu_lowered_simd_width(devinfo
, inst
));
6498 case FS_OPCODE_FB_WRITE_LOGICAL
:
6499 /* Gen6 doesn't support SIMD16 depth writes but we cannot handle them
6502 assert(devinfo
->gen
!= 6 ||
6503 inst
->src
[FB_WRITE_LOGICAL_SRC_SRC_DEPTH
].file
== BAD_FILE
||
6504 inst
->exec_size
== 8);
6505 /* Dual-source FB writes are unsupported in SIMD16 mode. */
6506 return (inst
->src
[FB_WRITE_LOGICAL_SRC_COLOR1
].file
!= BAD_FILE
?
6507 8 : MIN2(16, inst
->exec_size
));
6509 case FS_OPCODE_FB_READ_LOGICAL
:
6510 return MIN2(16, inst
->exec_size
);
6512 case SHADER_OPCODE_TEX_LOGICAL
:
6513 case SHADER_OPCODE_TXF_CMS_LOGICAL
:
6514 case SHADER_OPCODE_TXF_UMS_LOGICAL
:
6515 case SHADER_OPCODE_TXF_MCS_LOGICAL
:
6516 case SHADER_OPCODE_LOD_LOGICAL
:
6517 case SHADER_OPCODE_TG4_LOGICAL
:
6518 case SHADER_OPCODE_SAMPLEINFO_LOGICAL
:
6519 case SHADER_OPCODE_TXF_CMS_W_LOGICAL
:
6520 case SHADER_OPCODE_TG4_OFFSET_LOGICAL
:
6521 return get_sampler_lowered_simd_width(devinfo
, inst
);
6523 case SHADER_OPCODE_TXD_LOGICAL
:
6524 /* TXD is unsupported in SIMD16 mode. */
6527 case SHADER_OPCODE_TXL_LOGICAL
:
6528 case FS_OPCODE_TXB_LOGICAL
:
6529 /* Only one execution size is representable pre-ILK depending on whether
6530 * the shadow reference argument is present.
6532 if (devinfo
->gen
== 4)
6533 return inst
->src
[TEX_LOGICAL_SRC_SHADOW_C
].file
== BAD_FILE
? 16 : 8;
6535 return get_sampler_lowered_simd_width(devinfo
, inst
);
6537 case SHADER_OPCODE_TXF_LOGICAL
:
6538 case SHADER_OPCODE_TXS_LOGICAL
:
6539 /* Gen4 doesn't have SIMD8 variants for the RESINFO and LD-with-LOD
6540 * messages. Use SIMD16 instead.
6542 if (devinfo
->gen
== 4)
6545 return get_sampler_lowered_simd_width(devinfo
, inst
);
6547 case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
:
6548 case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
:
6549 case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
:
6552 case SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL
:
6553 case SHADER_OPCODE_UNTYPED_ATOMIC_FLOAT_LOGICAL
:
6554 case SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
:
6555 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
:
6556 case SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL
:
6557 case SHADER_OPCODE_BYTE_SCATTERED_READ_LOGICAL
:
6558 case SHADER_OPCODE_DWORD_SCATTERED_WRITE_LOGICAL
:
6559 case SHADER_OPCODE_DWORD_SCATTERED_READ_LOGICAL
:
6560 return MIN2(16, inst
->exec_size
);
6562 case SHADER_OPCODE_A64_UNTYPED_WRITE_LOGICAL
:
6563 case SHADER_OPCODE_A64_UNTYPED_READ_LOGICAL
:
6564 case SHADER_OPCODE_A64_BYTE_SCATTERED_WRITE_LOGICAL
:
6565 case SHADER_OPCODE_A64_BYTE_SCATTERED_READ_LOGICAL
:
6566 return devinfo
->gen
<= 8 ? 8 : MIN2(16, inst
->exec_size
);
6568 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_LOGICAL
:
6569 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_INT64_LOGICAL
:
6570 case SHADER_OPCODE_A64_UNTYPED_ATOMIC_FLOAT_LOGICAL
:
6573 case SHADER_OPCODE_URB_READ_SIMD8
:
6574 case SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
:
6575 case SHADER_OPCODE_URB_WRITE_SIMD8
:
6576 case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
6577 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
:
6578 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
6579 return MIN2(8, inst
->exec_size
);
6581 case SHADER_OPCODE_QUAD_SWIZZLE
: {
6582 const unsigned swiz
= inst
->src
[1].ud
;
6583 return (is_uniform(inst
->src
[0]) ?
6584 get_fpu_lowered_simd_width(devinfo
, inst
) :
6585 devinfo
->gen
< 11 && type_sz(inst
->src
[0].type
) == 4 ? 8 :
6586 swiz
== BRW_SWIZZLE_XYXY
|| swiz
== BRW_SWIZZLE_ZWZW
? 4 :
6587 get_fpu_lowered_simd_width(devinfo
, inst
));
6589 case SHADER_OPCODE_MOV_INDIRECT
: {
6590 /* From IVB and HSW PRMs:
6592 * "2.When the destination requires two registers and the sources are
6593 * indirect, the sources must use 1x1 regioning mode.
6595 * In case of DF instructions in HSW/IVB, the exec_size is limited by
6596 * the EU decompression logic not handling VxH indirect addressing
6599 const unsigned max_size
= (devinfo
->gen
>= 8 ? 2 : 1) * REG_SIZE
;
6600 /* Prior to Broadwell, we only have 8 address subregisters. */
6601 return MIN3(devinfo
->gen
>= 8 ? 16 : 8,
6602 max_size
/ (inst
->dst
.stride
* type_sz(inst
->dst
.type
)),
6606 case SHADER_OPCODE_LOAD_PAYLOAD
: {
6607 const unsigned reg_count
=
6608 DIV_ROUND_UP(inst
->dst
.component_size(inst
->exec_size
), REG_SIZE
);
6610 if (reg_count
> 2) {
6611 /* Only LOAD_PAYLOAD instructions with per-channel destination region
6612 * can be easily lowered (which excludes headers and heterogeneous
6615 assert(!inst
->header_size
);
6616 for (unsigned i
= 0; i
< inst
->sources
; i
++)
6617 assert(type_sz(inst
->dst
.type
) == type_sz(inst
->src
[i
].type
) ||
6618 inst
->src
[i
].file
== BAD_FILE
);
6620 return inst
->exec_size
/ DIV_ROUND_UP(reg_count
, 2);
6622 return inst
->exec_size
;
6626 return inst
->exec_size
;
6631 * Return true if splitting out the group of channels of instruction \p inst
6632 * given by lbld.group() requires allocating a temporary for the i-th source
6633 * of the lowered instruction.
6636 needs_src_copy(const fs_builder
&lbld
, const fs_inst
*inst
, unsigned i
)
6638 return !(is_periodic(inst
->src
[i
], lbld
.dispatch_width()) ||
6639 (inst
->components_read(i
) == 1 &&
6640 lbld
.dispatch_width() <= inst
->exec_size
)) ||
6641 (inst
->flags_written() &
6642 flag_mask(inst
->src
[i
], type_sz(inst
->src
[i
].type
)));
6646 * Extract the data that would be consumed by the channel group given by
6647 * lbld.group() from the i-th source region of instruction \p inst and return
6648 * it as result in packed form.
6651 emit_unzip(const fs_builder
&lbld
, fs_inst
*inst
, unsigned i
)
6653 assert(lbld
.group() >= inst
->group
);
6655 /* Specified channel group from the source region. */
6656 const fs_reg src
= horiz_offset(inst
->src
[i
], lbld
.group() - inst
->group
);
6658 if (needs_src_copy(lbld
, inst
, i
)) {
6659 /* Builder of the right width to perform the copy avoiding uninitialized
6660 * data if the lowered execution size is greater than the original
6661 * execution size of the instruction.
6663 const fs_builder cbld
= lbld
.group(MIN2(lbld
.dispatch_width(),
6664 inst
->exec_size
), 0);
6665 const fs_reg tmp
= lbld
.vgrf(inst
->src
[i
].type
, inst
->components_read(i
));
6667 for (unsigned k
= 0; k
< inst
->components_read(i
); ++k
)
6668 cbld
.MOV(offset(tmp
, lbld
, k
), offset(src
, inst
->exec_size
, k
));
6672 } else if (is_periodic(inst
->src
[i
], lbld
.dispatch_width())) {
6673 /* The source is invariant for all dispatch_width-wide groups of the
6676 return inst
->src
[i
];
6679 /* We can just point the lowered instruction at the right channel group
6680 * from the original region.
6687 * Return true if splitting out the group of channels of instruction \p inst
6688 * given by lbld.group() requires allocating a temporary for the destination
6689 * of the lowered instruction and copying the data back to the original
6690 * destination region.
6693 needs_dst_copy(const fs_builder
&lbld
, const fs_inst
*inst
)
6695 /* If the instruction writes more than one component we'll have to shuffle
6696 * the results of multiple lowered instructions in order to make sure that
6697 * they end up arranged correctly in the original destination region.
6699 if (inst
->size_written
> inst
->dst
.component_size(inst
->exec_size
))
6702 /* If the lowered execution size is larger than the original the result of
6703 * the instruction won't fit in the original destination, so we'll have to
6704 * allocate a temporary in any case.
6706 if (lbld
.dispatch_width() > inst
->exec_size
)
6709 for (unsigned i
= 0; i
< inst
->sources
; i
++) {
6710 /* If we already made a copy of the source for other reasons there won't
6711 * be any overlap with the destination.
6713 if (needs_src_copy(lbld
, inst
, i
))
6716 /* In order to keep the logic simple we emit a copy whenever the
6717 * destination region doesn't exactly match an overlapping source, which
6718 * may point at the source and destination not being aligned group by
6719 * group which could cause one of the lowered instructions to overwrite
6720 * the data read from the same source by other lowered instructions.
6722 if (regions_overlap(inst
->dst
, inst
->size_written
,
6723 inst
->src
[i
], inst
->size_read(i
)) &&
6724 !inst
->dst
.equals(inst
->src
[i
]))
6732 * Insert data from a packed temporary into the channel group given by
6733 * lbld.group() of the destination region of instruction \p inst and return
6734 * the temporary as result. Any copy instructions that are required for
6735 * unzipping the previous value (in the case of partial writes) will be
6736 * inserted using \p lbld_before and any copy instructions required for
6737 * zipping up the destination of \p inst will be inserted using \p lbld_after.
6740 emit_zip(const fs_builder
&lbld_before
, const fs_builder
&lbld_after
,
6743 assert(lbld_before
.dispatch_width() == lbld_after
.dispatch_width());
6744 assert(lbld_before
.group() == lbld_after
.group());
6745 assert(lbld_after
.group() >= inst
->group
);
6747 /* Specified channel group from the destination region. */
6748 const fs_reg dst
= horiz_offset(inst
->dst
, lbld_after
.group() - inst
->group
);
6749 const unsigned dst_size
= inst
->size_written
/
6750 inst
->dst
.component_size(inst
->exec_size
);
6752 if (needs_dst_copy(lbld_after
, inst
)) {
6753 const fs_reg tmp
= lbld_after
.vgrf(inst
->dst
.type
, dst_size
);
6755 if (inst
->predicate
) {
6756 /* Handle predication by copying the original contents of
6757 * the destination into the temporary before emitting the
6758 * lowered instruction.
6760 const fs_builder gbld_before
=
6761 lbld_before
.group(MIN2(lbld_before
.dispatch_width(),
6762 inst
->exec_size
), 0);
6763 for (unsigned k
= 0; k
< dst_size
; ++k
) {
6764 gbld_before
.MOV(offset(tmp
, lbld_before
, k
),
6765 offset(dst
, inst
->exec_size
, k
));
6769 const fs_builder gbld_after
=
6770 lbld_after
.group(MIN2(lbld_after
.dispatch_width(),
6771 inst
->exec_size
), 0);
6772 for (unsigned k
= 0; k
< dst_size
; ++k
) {
6773 /* Use a builder of the right width to perform the copy avoiding
6774 * uninitialized data if the lowered execution size is greater than
6775 * the original execution size of the instruction.
6777 gbld_after
.MOV(offset(dst
, inst
->exec_size
, k
),
6778 offset(tmp
, lbld_after
, k
));
6784 /* No need to allocate a temporary for the lowered instruction, just
6785 * take the right group of channels from the original region.
6792 fs_visitor::lower_simd_width()
6794 bool progress
= false;
6796 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
6797 const unsigned lower_width
= get_lowered_simd_width(devinfo
, inst
);
6799 if (lower_width
!= inst
->exec_size
) {
6800 /* Builder matching the original instruction. We may also need to
6801 * emit an instruction of width larger than the original, set the
6802 * execution size of the builder to the highest of both for now so
6803 * we're sure that both cases can be handled.
6805 const unsigned max_width
= MAX2(inst
->exec_size
, lower_width
);
6806 const fs_builder ibld
= bld
.at(block
, inst
)
6807 .exec_all(inst
->force_writemask_all
)
6808 .group(max_width
, inst
->group
/ max_width
);
6810 /* Split the copies in chunks of the execution width of either the
6811 * original or the lowered instruction, whichever is lower.
6813 const unsigned n
= DIV_ROUND_UP(inst
->exec_size
, lower_width
);
6814 const unsigned dst_size
= inst
->size_written
/
6815 inst
->dst
.component_size(inst
->exec_size
);
6817 assert(!inst
->writes_accumulator
&& !inst
->mlen
);
6819 /* Inserting the zip, unzip, and duplicated instructions in all of
6820 * the right spots is somewhat tricky. All of the unzip and any
6821 * instructions from the zip which unzip the destination prior to
6822 * writing need to happen before all of the per-group instructions
6823 * and the zip instructions need to happen after. In order to sort
6824 * this all out, we insert the unzip instructions before \p inst,
6825 * insert the per-group instructions after \p inst (i.e. before
6826 * inst->next), and insert the zip instructions before the
6827 * instruction after \p inst. Since we are inserting instructions
6828 * after \p inst, inst->next is a moving target and we need to save
6829 * it off here so that we insert the zip instructions in the right
6832 * Since we're inserting split instructions after after_inst, the
6833 * instructions will end up in the reverse order that we insert them.
6834 * However, certain render target writes require that the low group
6835 * instructions come before the high group. From the Ivy Bridge PRM
6836 * Vol. 4, Pt. 1, Section 3.9.11:
6838 * "If multiple SIMD8 Dual Source messages are delivered by the
6839 * pixel shader thread, each SIMD8_DUALSRC_LO message must be
6840 * issued before the SIMD8_DUALSRC_HI message with the same Slot
6841 * Group Select setting."
6843 * And, from Section 3.9.11.1 of the same PRM:
6845 * "When SIMD32 or SIMD16 PS threads send render target writes
6846 * with multiple SIMD8 and SIMD16 messages, the following must
6849 * All the slots (as described above) must have a corresponding
6850 * render target write irrespective of the slot's validity. A slot
6851 * is considered valid when at least one sample is enabled. For
6852 * example, a SIMD16 PS thread must send two SIMD8 render target
6853 * writes to cover all the slots.
6855 * PS thread must send SIMD render target write messages with
6856 * increasing slot numbers. For example, SIMD16 thread has
6857 * Slot[15:0] and if two SIMD8 render target writes are used, the
6858 * first SIMD8 render target write must send Slot[7:0] and the
6859 * next one must send Slot[15:8]."
6861 * In order to make low group instructions come before high group
6862 * instructions (this is required for some render target writes), we
6863 * split from the highest group to lowest.
6865 exec_node
*const after_inst
= inst
->next
;
6866 for (int i
= n
- 1; i
>= 0; i
--) {
6867 /* Emit a copy of the original instruction with the lowered width.
6868 * If the EOT flag was set throw it away except for the last
6869 * instruction to avoid killing the thread prematurely.
6871 fs_inst split_inst
= *inst
;
6872 split_inst
.exec_size
= lower_width
;
6873 split_inst
.eot
= inst
->eot
&& i
== int(n
- 1);
6875 /* Select the correct channel enables for the i-th group, then
6876 * transform the sources and destination and emit the lowered
6879 const fs_builder lbld
= ibld
.group(lower_width
, i
);
6881 for (unsigned j
= 0; j
< inst
->sources
; j
++)
6882 split_inst
.src
[j
] = emit_unzip(lbld
.at(block
, inst
), inst
, j
);
6884 split_inst
.dst
= emit_zip(lbld
.at(block
, inst
),
6885 lbld
.at(block
, after_inst
), inst
);
6886 split_inst
.size_written
=
6887 split_inst
.dst
.component_size(lower_width
) * dst_size
;
6889 lbld
.at(block
, inst
->next
).emit(split_inst
);
6892 inst
->remove(block
);
6898 invalidate_analysis(DEPENDENCY_INSTRUCTIONS
| DEPENDENCY_VARIABLES
);
6904 * Transform barycentric vectors into the interleaved form expected by the PLN
6905 * instruction and returned by the Gen7+ PI shared function.
6907 * For channels 0-15 in SIMD16 mode they are expected to be laid out as
6908 * follows in the register file:
6915 * There is no need to handle SIMD32 here -- This is expected to be run after
6916 * SIMD lowering, since SIMD lowering relies on vectors having the standard
6920 fs_visitor::lower_barycentrics()
6922 const bool has_interleaved_layout
= devinfo
->has_pln
|| devinfo
->gen
>= 7;
6923 bool progress
= false;
6925 if (stage
!= MESA_SHADER_FRAGMENT
|| !has_interleaved_layout
)
6928 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
6929 if (inst
->exec_size
< 16)
6932 const fs_builder
ibld(this, block
, inst
);
6933 const fs_builder ubld
= ibld
.exec_all().group(8, 0);
6935 switch (inst
->opcode
) {
6936 case FS_OPCODE_LINTERP
: {
6937 assert(inst
->exec_size
== 16);
6938 const fs_reg tmp
= ibld
.vgrf(inst
->src
[0].type
, 2);
6941 for (unsigned i
= 0; i
< ARRAY_SIZE(srcs
); i
++)
6942 srcs
[i
] = horiz_offset(offset(inst
->src
[0], ibld
, i
% 2),
6945 ubld
.LOAD_PAYLOAD(tmp
, srcs
, ARRAY_SIZE(srcs
), ARRAY_SIZE(srcs
));
6951 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
6952 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
6953 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
: {
6954 assert(inst
->exec_size
== 16);
6955 const fs_reg tmp
= ibld
.vgrf(inst
->dst
.type
, 2);
6957 for (unsigned i
= 0; i
< 2; i
++) {
6958 for (unsigned g
= 0; g
< inst
->exec_size
/ 8; g
++) {
6959 fs_inst
*mov
= ibld
.at(block
, inst
->next
).group(8, g
)
6960 .MOV(horiz_offset(offset(inst
->dst
, ibld
, i
),
6962 offset(tmp
, ubld
, 2 * g
+ i
));
6963 mov
->predicate
= inst
->predicate
;
6964 mov
->predicate_inverse
= inst
->predicate_inverse
;
6965 mov
->flag_subreg
= inst
->flag_subreg
;
6979 invalidate_analysis(DEPENDENCY_INSTRUCTIONS
| DEPENDENCY_VARIABLES
);
6985 fs_visitor::dump_instructions() const
6987 dump_instructions(NULL
);
6991 fs_visitor::dump_instructions(const char *name
) const
6993 FILE *file
= stderr
;
6994 if (name
&& geteuid() != 0) {
6995 file
= fopen(name
, "w");
7001 const register_pressure
&rp
= regpressure_analysis
.require();
7002 unsigned ip
= 0, max_pressure
= 0;
7003 foreach_block_and_inst(block
, backend_instruction
, inst
, cfg
) {
7004 max_pressure
= MAX2(max_pressure
, rp
.regs_live_at_ip
[ip
]);
7005 fprintf(file
, "{%3d} %4d: ", rp
.regs_live_at_ip
[ip
], ip
);
7006 dump_instruction(inst
, file
);
7009 fprintf(file
, "Maximum %3d registers live at once.\n", max_pressure
);
7012 foreach_in_list(backend_instruction
, inst
, &instructions
) {
7013 fprintf(file
, "%4d: ", ip
++);
7014 dump_instruction(inst
, file
);
7018 if (file
!= stderr
) {
7024 fs_visitor::dump_instruction(const backend_instruction
*be_inst
) const
7026 dump_instruction(be_inst
, stderr
);
7030 fs_visitor::dump_instruction(const backend_instruction
*be_inst
, FILE *file
) const
7032 const fs_inst
*inst
= (const fs_inst
*)be_inst
;
7034 if (inst
->predicate
) {
7035 fprintf(file
, "(%cf%d.%d) ",
7036 inst
->predicate_inverse
? '-' : '+',
7037 inst
->flag_subreg
/ 2,
7038 inst
->flag_subreg
% 2);
7041 fprintf(file
, "%s", brw_instruction_name(devinfo
, inst
->opcode
));
7043 fprintf(file
, ".sat");
7044 if (inst
->conditional_mod
) {
7045 fprintf(file
, "%s", conditional_modifier
[inst
->conditional_mod
]);
7046 if (!inst
->predicate
&&
7047 (devinfo
->gen
< 5 || (inst
->opcode
!= BRW_OPCODE_SEL
&&
7048 inst
->opcode
!= BRW_OPCODE_CSEL
&&
7049 inst
->opcode
!= BRW_OPCODE_IF
&&
7050 inst
->opcode
!= BRW_OPCODE_WHILE
))) {
7051 fprintf(file
, ".f%d.%d", inst
->flag_subreg
/ 2,
7052 inst
->flag_subreg
% 2);
7055 fprintf(file
, "(%d) ", inst
->exec_size
);
7058 fprintf(file
, "(mlen: %d) ", inst
->mlen
);
7061 if (inst
->ex_mlen
) {
7062 fprintf(file
, "(ex_mlen: %d) ", inst
->ex_mlen
);
7066 fprintf(file
, "(EOT) ");
7069 switch (inst
->dst
.file
) {
7071 fprintf(file
, "vgrf%d", inst
->dst
.nr
);
7074 fprintf(file
, "g%d", inst
->dst
.nr
);
7077 fprintf(file
, "m%d", inst
->dst
.nr
);
7080 fprintf(file
, "(null)");
7083 fprintf(file
, "***u%d***", inst
->dst
.nr
);
7086 fprintf(file
, "***attr%d***", inst
->dst
.nr
);
7089 switch (inst
->dst
.nr
) {
7091 fprintf(file
, "null");
7093 case BRW_ARF_ADDRESS
:
7094 fprintf(file
, "a0.%d", inst
->dst
.subnr
);
7096 case BRW_ARF_ACCUMULATOR
:
7097 fprintf(file
, "acc%d", inst
->dst
.subnr
);
7100 fprintf(file
, "f%d.%d", inst
->dst
.nr
& 0xf, inst
->dst
.subnr
);
7103 fprintf(file
, "arf%d.%d", inst
->dst
.nr
& 0xf, inst
->dst
.subnr
);
7108 unreachable("not reached");
7111 if (inst
->dst
.offset
||
7112 (inst
->dst
.file
== VGRF
&&
7113 alloc
.sizes
[inst
->dst
.nr
] * REG_SIZE
!= inst
->size_written
)) {
7114 const unsigned reg_size
= (inst
->dst
.file
== UNIFORM
? 4 : REG_SIZE
);
7115 fprintf(file
, "+%d.%d", inst
->dst
.offset
/ reg_size
,
7116 inst
->dst
.offset
% reg_size
);
7119 if (inst
->dst
.stride
!= 1)
7120 fprintf(file
, "<%u>", inst
->dst
.stride
);
7121 fprintf(file
, ":%s, ", brw_reg_type_to_letters(inst
->dst
.type
));
7123 for (int i
= 0; i
< inst
->sources
; i
++) {
7124 if (inst
->src
[i
].negate
)
7126 if (inst
->src
[i
].abs
)
7128 switch (inst
->src
[i
].file
) {
7130 fprintf(file
, "vgrf%d", inst
->src
[i
].nr
);
7133 fprintf(file
, "g%d", inst
->src
[i
].nr
);
7136 fprintf(file
, "***m%d***", inst
->src
[i
].nr
);
7139 fprintf(file
, "attr%d", inst
->src
[i
].nr
);
7142 fprintf(file
, "u%d", inst
->src
[i
].nr
);
7145 fprintf(file
, "(null)");
7148 switch (inst
->src
[i
].type
) {
7149 case BRW_REGISTER_TYPE_F
:
7150 fprintf(file
, "%-gf", inst
->src
[i
].f
);
7152 case BRW_REGISTER_TYPE_DF
:
7153 fprintf(file
, "%fdf", inst
->src
[i
].df
);
7155 case BRW_REGISTER_TYPE_W
:
7156 case BRW_REGISTER_TYPE_D
:
7157 fprintf(file
, "%dd", inst
->src
[i
].d
);
7159 case BRW_REGISTER_TYPE_UW
:
7160 case BRW_REGISTER_TYPE_UD
:
7161 fprintf(file
, "%uu", inst
->src
[i
].ud
);
7163 case BRW_REGISTER_TYPE_Q
:
7164 fprintf(file
, "%" PRId64
"q", inst
->src
[i
].d64
);
7166 case BRW_REGISTER_TYPE_UQ
:
7167 fprintf(file
, "%" PRIu64
"uq", inst
->src
[i
].u64
);
7169 case BRW_REGISTER_TYPE_VF
:
7170 fprintf(file
, "[%-gF, %-gF, %-gF, %-gF]",
7171 brw_vf_to_float((inst
->src
[i
].ud
>> 0) & 0xff),
7172 brw_vf_to_float((inst
->src
[i
].ud
>> 8) & 0xff),
7173 brw_vf_to_float((inst
->src
[i
].ud
>> 16) & 0xff),
7174 brw_vf_to_float((inst
->src
[i
].ud
>> 24) & 0xff));
7176 case BRW_REGISTER_TYPE_V
:
7177 case BRW_REGISTER_TYPE_UV
:
7178 fprintf(file
, "%08x%s", inst
->src
[i
].ud
,
7179 inst
->src
[i
].type
== BRW_REGISTER_TYPE_V
? "V" : "UV");
7182 fprintf(file
, "???");
7187 switch (inst
->src
[i
].nr
) {
7189 fprintf(file
, "null");
7191 case BRW_ARF_ADDRESS
:
7192 fprintf(file
, "a0.%d", inst
->src
[i
].subnr
);
7194 case BRW_ARF_ACCUMULATOR
:
7195 fprintf(file
, "acc%d", inst
->src
[i
].subnr
);
7198 fprintf(file
, "f%d.%d", inst
->src
[i
].nr
& 0xf, inst
->src
[i
].subnr
);
7201 fprintf(file
, "arf%d.%d", inst
->src
[i
].nr
& 0xf, inst
->src
[i
].subnr
);
7207 if (inst
->src
[i
].offset
||
7208 (inst
->src
[i
].file
== VGRF
&&
7209 alloc
.sizes
[inst
->src
[i
].nr
] * REG_SIZE
!= inst
->size_read(i
))) {
7210 const unsigned reg_size
= (inst
->src
[i
].file
== UNIFORM
? 4 : REG_SIZE
);
7211 fprintf(file
, "+%d.%d", inst
->src
[i
].offset
/ reg_size
,
7212 inst
->src
[i
].offset
% reg_size
);
7215 if (inst
->src
[i
].abs
)
7218 if (inst
->src
[i
].file
!= IMM
) {
7220 if (inst
->src
[i
].file
== ARF
|| inst
->src
[i
].file
== FIXED_GRF
) {
7221 unsigned hstride
= inst
->src
[i
].hstride
;
7222 stride
= (hstride
== 0 ? 0 : (1 << (hstride
- 1)));
7224 stride
= inst
->src
[i
].stride
;
7227 fprintf(file
, "<%u>", stride
);
7229 fprintf(file
, ":%s", brw_reg_type_to_letters(inst
->src
[i
].type
));
7232 if (i
< inst
->sources
- 1 && inst
->src
[i
+ 1].file
!= BAD_FILE
)
7233 fprintf(file
, ", ");
7238 if (inst
->force_writemask_all
)
7239 fprintf(file
, "NoMask ");
7241 if (inst
->exec_size
!= dispatch_width
)
7242 fprintf(file
, "group%d ", inst
->group
);
7244 fprintf(file
, "\n");
7248 fs_visitor::setup_fs_payload_gen6()
7250 assert(stage
== MESA_SHADER_FRAGMENT
);
7251 struct brw_wm_prog_data
*prog_data
= brw_wm_prog_data(this->prog_data
);
7252 const unsigned payload_width
= MIN2(16, dispatch_width
);
7253 assert(dispatch_width
% payload_width
== 0);
7254 assert(devinfo
->gen
>= 6);
7256 prog_data
->uses_src_depth
= prog_data
->uses_src_w
=
7257 (nir
->info
.system_values_read
& (1ull << SYSTEM_VALUE_FRAG_COORD
)) != 0;
7259 prog_data
->uses_sample_mask
=
7260 (nir
->info
.system_values_read
& SYSTEM_BIT_SAMPLE_MASK_IN
) != 0;
7262 /* From the Ivy Bridge PRM documentation for 3DSTATE_PS:
7264 * "MSDISPMODE_PERSAMPLE is required in order to select
7267 * So we can only really get sample positions if we are doing real
7268 * per-sample dispatch. If we need gl_SamplePosition and we don't have
7269 * persample dispatch, we hard-code it to 0.5.
7271 prog_data
->uses_pos_offset
= prog_data
->persample_dispatch
&&
7272 (nir
->info
.system_values_read
& SYSTEM_BIT_SAMPLE_POS
);
7274 /* R0: PS thread payload header. */
7277 for (unsigned j
= 0; j
< dispatch_width
/ payload_width
; j
++) {
7278 /* R1: masks, pixel X/Y coordinates. */
7279 payload
.subspan_coord_reg
[j
] = payload
.num_regs
++;
7282 for (unsigned j
= 0; j
< dispatch_width
/ payload_width
; j
++) {
7283 /* R3-26: barycentric interpolation coordinates. These appear in the
7284 * same order that they appear in the brw_barycentric_mode enum. Each
7285 * set of coordinates occupies 2 registers if dispatch width == 8 and 4
7286 * registers if dispatch width == 16. Coordinates only appear if they
7287 * were enabled using the "Barycentric Interpolation Mode" bits in
7290 for (int i
= 0; i
< BRW_BARYCENTRIC_MODE_COUNT
; ++i
) {
7291 if (prog_data
->barycentric_interp_modes
& (1 << i
)) {
7292 payload
.barycentric_coord_reg
[i
][j
] = payload
.num_regs
;
7293 payload
.num_regs
+= payload_width
/ 4;
7297 /* R27-28: interpolated depth if uses source depth */
7298 if (prog_data
->uses_src_depth
) {
7299 payload
.source_depth_reg
[j
] = payload
.num_regs
;
7300 payload
.num_regs
+= payload_width
/ 8;
7303 /* R29-30: interpolated W set if GEN6_WM_USES_SOURCE_W. */
7304 if (prog_data
->uses_src_w
) {
7305 payload
.source_w_reg
[j
] = payload
.num_regs
;
7306 payload
.num_regs
+= payload_width
/ 8;
7309 /* R31: MSAA position offsets. */
7310 if (prog_data
->uses_pos_offset
) {
7311 payload
.sample_pos_reg
[j
] = payload
.num_regs
;
7315 /* R32-33: MSAA input coverage mask */
7316 if (prog_data
->uses_sample_mask
) {
7317 assert(devinfo
->gen
>= 7);
7318 payload
.sample_mask_in_reg
[j
] = payload
.num_regs
;
7319 payload
.num_regs
+= payload_width
/ 8;
7323 if (nir
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_DEPTH
)) {
7324 source_depth_to_render_target
= true;
7329 fs_visitor::setup_vs_payload()
7331 /* R0: thread header, R1: urb handles */
7332 payload
.num_regs
= 2;
7336 fs_visitor::setup_gs_payload()
7338 assert(stage
== MESA_SHADER_GEOMETRY
);
7340 struct brw_gs_prog_data
*gs_prog_data
= brw_gs_prog_data(prog_data
);
7341 struct brw_vue_prog_data
*vue_prog_data
= brw_vue_prog_data(prog_data
);
7343 /* R0: thread header, R1: output URB handles */
7344 payload
.num_regs
= 2;
7346 if (gs_prog_data
->include_primitive_id
) {
7347 /* R2: Primitive ID 0..7 */
7351 /* Always enable VUE handles so we can safely use pull model if needed.
7353 * The push model for a GS uses a ton of register space even for trivial
7354 * scenarios with just a few inputs, so just make things easier and a bit
7355 * safer by always having pull model available.
7357 gs_prog_data
->base
.include_vue_handles
= true;
7359 /* R3..RN: ICP Handles for each incoming vertex (when using pull model) */
7360 payload
.num_regs
+= nir
->info
.gs
.vertices_in
;
7362 /* Use a maximum of 24 registers for push-model inputs. */
7363 const unsigned max_push_components
= 24;
7365 /* If pushing our inputs would take too many registers, reduce the URB read
7366 * length (which is in HWords, or 8 registers), and resort to pulling.
7368 * Note that the GS reads <URB Read Length> HWords for every vertex - so we
7369 * have to multiply by VerticesIn to obtain the total storage requirement.
7371 if (8 * vue_prog_data
->urb_read_length
* nir
->info
.gs
.vertices_in
>
7372 max_push_components
) {
7373 vue_prog_data
->urb_read_length
=
7374 ROUND_DOWN_TO(max_push_components
/ nir
->info
.gs
.vertices_in
, 8) / 8;
7379 fs_visitor::setup_cs_payload()
7381 assert(devinfo
->gen
>= 7);
7382 payload
.num_regs
= 1;
7385 brw::register_pressure::register_pressure(const fs_visitor
*v
)
7387 const fs_live_variables
&live
= v
->live_analysis
.require();
7388 const unsigned num_instructions
= v
->cfg
->num_blocks
?
7389 v
->cfg
->blocks
[v
->cfg
->num_blocks
- 1]->end_ip
+ 1 : 0;
7391 regs_live_at_ip
= new unsigned[num_instructions
]();
7393 for (unsigned reg
= 0; reg
< v
->alloc
.count
; reg
++) {
7394 for (int ip
= live
.vgrf_start
[reg
]; ip
<= live
.vgrf_end
[reg
]; ip
++)
7395 regs_live_at_ip
[ip
] += v
->alloc
.sizes
[reg
];
7399 brw::register_pressure::~register_pressure()
7401 delete[] regs_live_at_ip
;
7405 fs_visitor::invalidate_analysis(brw::analysis_dependency_class c
)
7407 backend_shader::invalidate_analysis(c
);
7408 live_analysis
.invalidate(c
);
7409 regpressure_analysis
.invalidate(c
);
7413 fs_visitor::optimize()
7415 /* Start by validating the shader we currently have. */
7418 /* bld is the common builder object pointing at the end of the program we
7419 * used to translate it into i965 IR. For the optimization and lowering
7420 * passes coming next, any code added after the end of the program without
7421 * having explicitly called fs_builder::at() clearly points at a mistake.
7422 * Ideally optimization passes wouldn't be part of the visitor so they
7423 * wouldn't have access to bld at all, but they do, so just in case some
7424 * pass forgets to ask for a location explicitly set it to NULL here to
7425 * make it trip. The dispatch width is initialized to a bogus value to
7426 * make sure that optimizations set the execution controls explicitly to
7427 * match the code they are manipulating instead of relying on the defaults.
7429 bld
= fs_builder(this, 64);
7431 assign_constant_locations();
7432 lower_constant_loads();
7436 split_virtual_grfs();
7439 #define OPT(pass, args...) ({ \
7441 bool this_progress = pass(args); \
7443 if (unlikely(INTEL_DEBUG & DEBUG_OPTIMIZER) && this_progress) { \
7444 char filename[64]; \
7445 snprintf(filename, 64, "%s%d-%s-%02d-%02d-" #pass, \
7446 stage_abbrev, dispatch_width, nir->info.name, iteration, pass_num); \
7448 backend_shader::dump_instructions(filename); \
7453 progress = progress || this_progress; \
7457 if (unlikely(INTEL_DEBUG
& DEBUG_OPTIMIZER
)) {
7459 snprintf(filename
, 64, "%s%d-%s-00-00-start",
7460 stage_abbrev
, dispatch_width
, nir
->info
.name
);
7462 backend_shader::dump_instructions(filename
);
7465 bool progress
= false;
7469 /* Before anything else, eliminate dead code. The results of some NIR
7470 * instructions may effectively be calculated twice. Once when the
7471 * instruction is encountered, and again when the user of that result is
7472 * encountered. Wipe those away before algebraic optimizations and
7473 * especially copy propagation can mix things up.
7475 OPT(dead_code_eliminate
);
7477 OPT(remove_extra_rounding_modes
);
7484 OPT(remove_duplicate_mrf_writes
);
7488 OPT(opt_copy_propagation
);
7489 OPT(opt_predicated_break
, this);
7490 OPT(opt_cmod_propagation
);
7491 OPT(dead_code_eliminate
);
7492 OPT(opt_peephole_sel
);
7493 OPT(dead_control_flow_eliminate
, this);
7494 OPT(opt_register_renaming
);
7495 OPT(opt_saturate_propagation
);
7496 OPT(register_coalesce
);
7497 OPT(compute_to_mrf
);
7498 OPT(eliminate_find_live_channel
);
7500 OPT(compact_virtual_grfs
);
7506 if (OPT(lower_pack
)) {
7507 OPT(register_coalesce
);
7508 OPT(dead_code_eliminate
);
7511 OPT(lower_simd_width
);
7512 OPT(lower_barycentrics
);
7514 /* After SIMD lowering just in case we had to unroll the EOT send. */
7515 OPT(opt_sampler_eot
);
7517 OPT(lower_logical_sends
);
7519 /* After logical SEND lowering. */
7520 OPT(fixup_nomask_control_flow
);
7523 OPT(opt_copy_propagation
);
7524 /* Only run after logical send lowering because it's easier to implement
7525 * in terms of physical sends.
7527 if (OPT(opt_zero_samples
))
7528 OPT(opt_copy_propagation
);
7529 /* Run after logical send lowering to give it a chance to CSE the
7530 * LOAD_PAYLOAD instructions created to construct the payloads of
7531 * e.g. texturing messages in cases where it wasn't possible to CSE the
7532 * whole logical instruction.
7535 OPT(register_coalesce
);
7536 OPT(compute_to_mrf
);
7537 OPT(dead_code_eliminate
);
7538 OPT(remove_duplicate_mrf_writes
);
7539 OPT(opt_peephole_sel
);
7542 OPT(opt_redundant_discard_jumps
);
7544 if (OPT(lower_load_payload
)) {
7545 split_virtual_grfs();
7547 /* Lower 64 bit MOVs generated by payload lowering. */
7548 if (!devinfo
->has_64bit_float
&& !devinfo
->has_64bit_int
)
7551 OPT(register_coalesce
);
7552 OPT(lower_simd_width
);
7553 OPT(compute_to_mrf
);
7554 OPT(dead_code_eliminate
);
7557 OPT(opt_combine_constants
);
7558 OPT(lower_integer_multiplication
);
7561 if (devinfo
->gen
<= 5 && OPT(lower_minmax
)) {
7562 OPT(opt_cmod_propagation
);
7564 OPT(opt_copy_propagation
);
7565 OPT(dead_code_eliminate
);
7568 if (OPT(lower_regioning
)) {
7569 OPT(opt_copy_propagation
);
7570 OPT(dead_code_eliminate
);
7571 OPT(lower_simd_width
);
7574 OPT(fixup_sends_duplicate_payload
);
7576 lower_uniform_pull_constant_loads();
7582 * From the Skylake PRM Vol. 2a docs for sends:
7584 * "It is required that the second block of GRFs does not overlap with the
7587 * There are plenty of cases where we may accidentally violate this due to
7588 * having, for instance, both sources be the constant 0. This little pass
7589 * just adds a new vgrf for the second payload and copies it over.
7592 fs_visitor::fixup_sends_duplicate_payload()
7594 bool progress
= false;
7596 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
7597 if (inst
->opcode
== SHADER_OPCODE_SEND
&& inst
->ex_mlen
> 0 &&
7598 regions_overlap(inst
->src
[2], inst
->mlen
* REG_SIZE
,
7599 inst
->src
[3], inst
->ex_mlen
* REG_SIZE
)) {
7600 fs_reg tmp
= fs_reg(VGRF
, alloc
.allocate(inst
->ex_mlen
),
7601 BRW_REGISTER_TYPE_UD
);
7602 /* Sadly, we've lost all notion of channels and bit sizes at this
7603 * point. Just WE_all it.
7605 const fs_builder ibld
= bld
.at(block
, inst
).exec_all().group(16, 0);
7606 fs_reg copy_src
= retype(inst
->src
[3], BRW_REGISTER_TYPE_UD
);
7607 fs_reg copy_dst
= tmp
;
7608 for (unsigned i
= 0; i
< inst
->ex_mlen
; i
+= 2) {
7609 if (inst
->ex_mlen
== i
+ 1) {
7610 /* Only one register left; do SIMD8 */
7611 ibld
.group(8, 0).MOV(copy_dst
, copy_src
);
7613 ibld
.MOV(copy_dst
, copy_src
);
7615 copy_src
= offset(copy_src
, ibld
, 1);
7616 copy_dst
= offset(copy_dst
, ibld
, 1);
7624 invalidate_analysis(DEPENDENCY_INSTRUCTIONS
| DEPENDENCY_VARIABLES
);
7630 * Three source instruction must have a GRF/MRF destination register.
7631 * ARF NULL is not allowed. Fix that up by allocating a temporary GRF.
7634 fs_visitor::fixup_3src_null_dest()
7636 bool progress
= false;
7638 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
7639 if (inst
->is_3src(devinfo
) && inst
->dst
.is_null()) {
7640 inst
->dst
= fs_reg(VGRF
, alloc
.allocate(dispatch_width
/ 8),
7647 invalidate_analysis(DEPENDENCY_INSTRUCTION_DETAIL
|
7648 DEPENDENCY_VARIABLES
);
7652 * Find the first instruction in the program that might start a region of
7653 * divergent control flow due to a HALT jump. There is no
7654 * find_halt_control_flow_region_end(), the region of divergence extends until
7655 * the only FS_OPCODE_PLACEHOLDER_HALT in the program.
7657 static const fs_inst
*
7658 find_halt_control_flow_region_start(const fs_visitor
*v
)
7660 if (brw_wm_prog_data(v
->prog_data
)->uses_kill
) {
7661 foreach_block_and_inst(block
, fs_inst
, inst
, v
->cfg
) {
7662 if (inst
->opcode
== FS_OPCODE_DISCARD_JUMP
||
7663 inst
->opcode
== FS_OPCODE_PLACEHOLDER_HALT
)
7672 * Work around the Gen12 hardware bug filed as GEN:BUG:1407528679. EU fusion
7673 * can cause a BB to be executed with all channels disabled, which will lead
7674 * to the execution of any NoMask instructions in it, even though any
7675 * execution-masked instructions will be correctly shot down. This may break
7676 * assumptions of some NoMask SEND messages whose descriptor depends on data
7677 * generated by live invocations of the shader.
7679 * This avoids the problem by predicating certain instructions on an ANY
7680 * horizontal predicate that makes sure that their execution is omitted when
7681 * all channels of the program are disabled.
7684 fs_visitor::fixup_nomask_control_flow()
7686 if (devinfo
->gen
!= 12)
7689 const brw_predicate pred
= dispatch_width
> 16 ? BRW_PREDICATE_ALIGN1_ANY32H
:
7690 dispatch_width
> 8 ? BRW_PREDICATE_ALIGN1_ANY16H
:
7691 BRW_PREDICATE_ALIGN1_ANY8H
;
7692 const fs_inst
*halt_start
= find_halt_control_flow_region_start(this);
7694 bool progress
= false;
7696 const fs_live_variables
&live_vars
= live_analysis
.require();
7698 /* Scan the program backwards in order to be able to easily determine
7699 * whether the flag register is live at any point.
7701 foreach_block_reverse_safe(block
, cfg
) {
7702 BITSET_WORD flag_liveout
= live_vars
.block_data
[block
->num
]
7704 STATIC_ASSERT(ARRAY_SIZE(live_vars
.block_data
[0].flag_liveout
) == 1);
7706 foreach_inst_in_block_reverse_safe(fs_inst
, inst
, block
) {
7707 if (!inst
->predicate
&& inst
->exec_size
>= 8)
7708 flag_liveout
&= ~inst
->flags_written();
7710 switch (inst
->opcode
) {
7713 /* Note that this doesn't handle FS_OPCODE_DISCARD_JUMP since only
7714 * the first one in the program closes the region of divergent
7715 * control flow due to any HALT instructions -- Instead this is
7716 * handled with the halt_start check below.
7721 case BRW_OPCODE_WHILE
:
7722 case BRW_OPCODE_ENDIF
:
7723 case FS_OPCODE_PLACEHOLDER_HALT
:
7728 /* Note that the vast majority of NoMask SEND instructions in the
7729 * program are harmless while executed in a block with all
7730 * channels disabled, since any instructions with side effects we
7731 * could hit here should be execution-masked.
7733 * The main concern is NoMask SEND instructions where the message
7734 * descriptor or header depends on data generated by live
7735 * invocations of the shader (RESINFO and
7736 * FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD with a dynamically
7737 * computed surface index seem to be the only examples right now
7738 * where this could easily lead to GPU hangs). Unfortunately we
7739 * have no straightforward way to detect that currently, so just
7740 * predicate any NoMask SEND instructions we find under control
7743 * If this proves to have a measurable performance impact it can
7744 * be easily extended with a whitelist of messages we know we can
7745 * safely omit the predication for.
7747 if (depth
&& inst
->force_writemask_all
&&
7748 is_send(inst
) && !inst
->predicate
) {
7749 /* We need to load the execution mask into the flag register by
7750 * using a builder with channel group matching the whole shader
7751 * (rather than the default which is derived from the original
7752 * instruction), in order to avoid getting a right-shifted
7755 const fs_builder ubld
= fs_builder(this, block
, inst
)
7756 .exec_all().group(dispatch_width
, 0);
7757 const fs_reg flag
= retype(brw_flag_reg(0, 0),
7758 BRW_REGISTER_TYPE_UD
);
7760 /* Due to the lack of flag register allocation we need to save
7761 * and restore the flag register if it's live.
7763 const bool save_flag
= flag_liveout
&
7764 flag_mask(flag
, dispatch_width
/ 8);
7765 const fs_reg tmp
= ubld
.group(1, 0).vgrf(flag
.type
);
7768 ubld
.group(1, 0).MOV(tmp
, flag
);
7770 ubld
.emit(FS_OPCODE_LOAD_LIVE_CHANNELS
);
7772 set_predicate(pred
, inst
);
7773 inst
->flag_subreg
= 0;
7776 ubld
.group(1, 0).at(block
, inst
->next
).MOV(flag
, tmp
);
7783 if (inst
== halt_start
)
7786 flag_liveout
|= inst
->flags_read(devinfo
);
7791 invalidate_analysis(DEPENDENCY_INSTRUCTIONS
| DEPENDENCY_VARIABLES
);
7797 fs_visitor::allocate_registers(unsigned min_dispatch_width
, bool allow_spilling
)
7801 static const enum instruction_scheduler_mode pre_modes
[] = {
7803 SCHEDULE_PRE_NON_LIFO
,
7807 static const char *scheduler_mode_name
[] = {
7813 bool spill_all
= allow_spilling
&& (INTEL_DEBUG
& DEBUG_SPILL_FS
);
7815 /* Try each scheduling heuristic to see if it can successfully register
7816 * allocate without spilling. They should be ordered by decreasing
7817 * performance but increasing likelihood of allocating.
7819 for (unsigned i
= 0; i
< ARRAY_SIZE(pre_modes
); i
++) {
7820 schedule_instructions(pre_modes
[i
]);
7821 this->shader_stats
.scheduler_mode
= scheduler_mode_name
[i
];
7824 assign_regs_trivial();
7829 /* Scheduling may create additional opportunities for CMOD propagation,
7830 * so let's do it again. If CMOD propagation made any progress,
7831 * elminate dead code one more time.
7833 bool progress
= false;
7834 const int iteration
= 99;
7837 if (OPT(opt_cmod_propagation
)) {
7838 /* dead_code_eliminate "undoes" the fixing done by
7839 * fixup_3src_null_dest, so we have to do it again if
7840 * dead_code_eliminiate makes any progress.
7842 if (OPT(dead_code_eliminate
))
7843 fixup_3src_null_dest();
7847 /* We only allow spilling for the last schedule mode and only if the
7848 * allow_spilling parameter and dispatch width work out ok.
7850 bool can_spill
= allow_spilling
&&
7851 (i
== ARRAY_SIZE(pre_modes
) - 1) &&
7852 dispatch_width
== min_dispatch_width
;
7854 /* We should only spill registers on the last scheduling. */
7855 assert(!spilled_any_registers
);
7857 allocated
= assign_regs(can_spill
, spill_all
);
7863 if (!allow_spilling
)
7864 fail("Failure to register allocate and spilling is not allowed.");
7866 /* We assume that any spilling is worse than just dropping back to
7867 * SIMD8. There's probably actually some intermediate point where
7868 * SIMD16 with a couple of spills is still better.
7870 if (dispatch_width
> min_dispatch_width
) {
7871 fail("Failure to register allocate. Reduce number of "
7872 "live scalar values to avoid this.");
7875 /* If we failed to allocate, we must have a reason */
7877 } else if (spilled_any_registers
) {
7878 compiler
->shader_perf_log(log_data
,
7879 "%s shader triggered register spilling. "
7880 "Try reducing the number of live scalar "
7881 "values to improve performance.\n",
7885 /* This must come after all optimization and register allocation, since
7886 * it inserts dead code that happens to have side effects, and it does
7887 * so based on the actual physical registers in use.
7889 insert_gen4_send_dependency_workarounds();
7894 opt_bank_conflicts();
7896 schedule_instructions(SCHEDULE_POST
);
7898 if (last_scratch
> 0) {
7899 ASSERTED
unsigned max_scratch_size
= 2 * 1024 * 1024;
7901 prog_data
->total_scratch
= brw_get_scratch_size(last_scratch
);
7903 if (stage
== MESA_SHADER_COMPUTE
) {
7904 if (devinfo
->is_haswell
) {
7905 /* According to the MEDIA_VFE_STATE's "Per Thread Scratch Space"
7906 * field documentation, Haswell supports a minimum of 2kB of
7907 * scratch space for compute shaders, unlike every other stage
7910 prog_data
->total_scratch
= MAX2(prog_data
->total_scratch
, 2048);
7911 } else if (devinfo
->gen
<= 7) {
7912 /* According to the MEDIA_VFE_STATE's "Per Thread Scratch Space"
7913 * field documentation, platforms prior to Haswell measure scratch
7914 * size linearly with a range of [1kB, 12kB] and 1kB granularity.
7916 prog_data
->total_scratch
= ALIGN(last_scratch
, 1024);
7917 max_scratch_size
= 12 * 1024;
7921 /* We currently only support up to 2MB of scratch space. If we
7922 * need to support more eventually, the documentation suggests
7923 * that we could allocate a larger buffer, and partition it out
7924 * ourselves. We'd just have to undo the hardware's address
7925 * calculation by subtracting (FFTID * Per Thread Scratch Space)
7926 * and then add FFTID * (Larger Per Thread Scratch Space).
7928 * See 3D-Media-GPGPU Engine > Media GPGPU Pipeline >
7929 * Thread Group Tracking > Local Memory/Scratch Space.
7931 assert(prog_data
->total_scratch
< max_scratch_size
);
7938 fs_visitor::run_vs()
7940 assert(stage
== MESA_SHADER_VERTEX
);
7944 if (shader_time_index
>= 0)
7945 emit_shader_time_begin();
7954 if (shader_time_index
>= 0)
7955 emit_shader_time_end();
7961 assign_curb_setup();
7962 assign_vs_urb_setup();
7964 fixup_3src_null_dest();
7965 allocate_registers(8, true);
7971 fs_visitor::set_tcs_invocation_id()
7973 struct brw_tcs_prog_data
*tcs_prog_data
= brw_tcs_prog_data(prog_data
);
7974 struct brw_vue_prog_data
*vue_prog_data
= &tcs_prog_data
->base
;
7976 const unsigned instance_id_mask
=
7977 devinfo
->gen
>= 11 ? INTEL_MASK(22, 16) : INTEL_MASK(23, 17);
7978 const unsigned instance_id_shift
=
7979 devinfo
->gen
>= 11 ? 16 : 17;
7981 /* Get instance number from g0.2 bits 22:16 or 23:17 */
7982 fs_reg t
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
7983 bld
.AND(t
, fs_reg(retype(brw_vec1_grf(0, 2), BRW_REGISTER_TYPE_UD
)),
7984 brw_imm_ud(instance_id_mask
));
7986 invocation_id
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
7988 if (vue_prog_data
->dispatch_mode
== DISPATCH_MODE_TCS_8_PATCH
) {
7989 /* gl_InvocationID is just the thread number */
7990 bld
.SHR(invocation_id
, t
, brw_imm_ud(instance_id_shift
));
7994 assert(vue_prog_data
->dispatch_mode
== DISPATCH_MODE_TCS_SINGLE_PATCH
);
7996 fs_reg channels_uw
= bld
.vgrf(BRW_REGISTER_TYPE_UW
);
7997 fs_reg channels_ud
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
7998 bld
.MOV(channels_uw
, fs_reg(brw_imm_uv(0x76543210)));
7999 bld
.MOV(channels_ud
, channels_uw
);
8001 if (tcs_prog_data
->instances
== 1) {
8002 invocation_id
= channels_ud
;
8004 fs_reg instance_times_8
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
8005 bld
.SHR(instance_times_8
, t
, brw_imm_ud(instance_id_shift
- 3));
8006 bld
.ADD(invocation_id
, instance_times_8
, channels_ud
);
8011 fs_visitor::run_tcs()
8013 assert(stage
== MESA_SHADER_TESS_CTRL
);
8015 struct brw_vue_prog_data
*vue_prog_data
= brw_vue_prog_data(prog_data
);
8016 struct brw_tcs_prog_data
*tcs_prog_data
= brw_tcs_prog_data(prog_data
);
8017 struct brw_tcs_prog_key
*tcs_key
= (struct brw_tcs_prog_key
*) key
;
8019 assert(vue_prog_data
->dispatch_mode
== DISPATCH_MODE_TCS_SINGLE_PATCH
||
8020 vue_prog_data
->dispatch_mode
== DISPATCH_MODE_TCS_8_PATCH
);
8022 if (vue_prog_data
->dispatch_mode
== DISPATCH_MODE_TCS_SINGLE_PATCH
) {
8023 /* r1-r4 contain the ICP handles. */
8024 payload
.num_regs
= 5;
8026 assert(vue_prog_data
->dispatch_mode
== DISPATCH_MODE_TCS_8_PATCH
);
8027 assert(tcs_key
->input_vertices
> 0);
8028 /* r1 contains output handles, r2 may contain primitive ID, then the
8029 * ICP handles occupy the next 1-32 registers.
8031 payload
.num_regs
= 2 + tcs_prog_data
->include_primitive_id
+
8032 tcs_key
->input_vertices
;
8035 if (shader_time_index
>= 0)
8036 emit_shader_time_begin();
8038 /* Initialize gl_InvocationID */
8039 set_tcs_invocation_id();
8041 const bool fix_dispatch_mask
=
8042 vue_prog_data
->dispatch_mode
== DISPATCH_MODE_TCS_SINGLE_PATCH
&&
8043 (nir
->info
.tess
.tcs_vertices_out
% 8) != 0;
8045 /* Fix the disptach mask */
8046 if (fix_dispatch_mask
) {
8047 bld
.CMP(bld
.null_reg_ud(), invocation_id
,
8048 brw_imm_ud(nir
->info
.tess
.tcs_vertices_out
), BRW_CONDITIONAL_L
);
8049 bld
.IF(BRW_PREDICATE_NORMAL
);
8054 if (fix_dispatch_mask
) {
8055 bld
.emit(BRW_OPCODE_ENDIF
);
8058 /* Emit EOT write; set TR DS Cache bit */
8060 fs_reg(get_tcs_output_urb_handle()),
8061 fs_reg(brw_imm_ud(WRITEMASK_X
<< 16)),
8062 fs_reg(brw_imm_ud(0)),
8064 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 3);
8065 bld
.LOAD_PAYLOAD(payload
, srcs
, 3, 2);
8067 fs_inst
*inst
= bld
.emit(SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
,
8068 bld
.null_reg_ud(), payload
);
8072 if (shader_time_index
>= 0)
8073 emit_shader_time_end();
8082 assign_curb_setup();
8083 assign_tcs_urb_setup();
8085 fixup_3src_null_dest();
8086 allocate_registers(8, true);
8092 fs_visitor::run_tes()
8094 assert(stage
== MESA_SHADER_TESS_EVAL
);
8096 /* R0: thread header, R1-3: gl_TessCoord.xyz, R4: URB handles */
8097 payload
.num_regs
= 5;
8099 if (shader_time_index
>= 0)
8100 emit_shader_time_begin();
8109 if (shader_time_index
>= 0)
8110 emit_shader_time_end();
8116 assign_curb_setup();
8117 assign_tes_urb_setup();
8119 fixup_3src_null_dest();
8120 allocate_registers(8, true);
8126 fs_visitor::run_gs()
8128 assert(stage
== MESA_SHADER_GEOMETRY
);
8132 this->final_gs_vertex_count
= vgrf(glsl_type::uint_type
);
8134 if (gs_compile
->control_data_header_size_bits
> 0) {
8135 /* Create a VGRF to store accumulated control data bits. */
8136 this->control_data_bits
= vgrf(glsl_type::uint_type
);
8138 /* If we're outputting more than 32 control data bits, then EmitVertex()
8139 * will set control_data_bits to 0 after emitting the first vertex.
8140 * Otherwise, we need to initialize it to 0 here.
8142 if (gs_compile
->control_data_header_size_bits
<= 32) {
8143 const fs_builder abld
= bld
.annotate("initialize control data bits");
8144 abld
.MOV(this->control_data_bits
, brw_imm_ud(0u));
8148 if (shader_time_index
>= 0)
8149 emit_shader_time_begin();
8153 emit_gs_thread_end();
8155 if (shader_time_index
>= 0)
8156 emit_shader_time_end();
8165 assign_curb_setup();
8166 assign_gs_urb_setup();
8168 fixup_3src_null_dest();
8169 allocate_registers(8, true);
8174 /* From the SKL PRM, Volume 16, Workarounds:
8176 * 0877 3D Pixel Shader Hang possible when pixel shader dispatched with
8177 * only header phases (R0-R2)
8179 * WA: Enable a non-header phase (e.g. push constant) when dispatch would
8180 * have been header only.
8182 * Instead of enabling push constants one can alternatively enable one of the
8183 * inputs. Here one simply chooses "layer" which shouldn't impose much
8187 gen9_ps_header_only_workaround(struct brw_wm_prog_data
*wm_prog_data
)
8189 if (wm_prog_data
->num_varying_inputs
)
8192 if (wm_prog_data
->base
.curb_read_length
)
8195 wm_prog_data
->urb_setup
[VARYING_SLOT_LAYER
] = 0;
8196 wm_prog_data
->num_varying_inputs
= 1;
8198 brw_compute_urb_setup_index(wm_prog_data
);
8202 fs_visitor::run_fs(bool allow_spilling
, bool do_rep_send
)
8204 struct brw_wm_prog_data
*wm_prog_data
= brw_wm_prog_data(this->prog_data
);
8205 brw_wm_prog_key
*wm_key
= (brw_wm_prog_key
*) this->key
;
8207 assert(stage
== MESA_SHADER_FRAGMENT
);
8209 if (devinfo
->gen
>= 6)
8210 setup_fs_payload_gen6();
8212 setup_fs_payload_gen4();
8216 } else if (do_rep_send
) {
8217 assert(dispatch_width
== 16);
8218 emit_repclear_shader();
8220 if (shader_time_index
>= 0)
8221 emit_shader_time_begin();
8223 if (nir
->info
.inputs_read
> 0 ||
8224 (nir
->info
.system_values_read
& (1ull << SYSTEM_VALUE_FRAG_COORD
)) ||
8225 (nir
->info
.outputs_read
> 0 && !wm_key
->coherent_fb_fetch
)) {
8226 if (devinfo
->gen
< 6)
8227 emit_interpolation_setup_gen4();
8229 emit_interpolation_setup_gen6();
8232 /* We handle discards by keeping track of the still-live pixels in f0.1.
8233 * Initialize it with the dispatched pixels.
8235 if (wm_prog_data
->uses_kill
) {
8236 const unsigned lower_width
= MIN2(dispatch_width
, 16);
8237 for (unsigned i
= 0; i
< dispatch_width
/ lower_width
; i
++) {
8238 const fs_reg dispatch_mask
=
8239 devinfo
->gen
>= 6 ? brw_vec1_grf((i
? 2 : 1), 7) :
8241 bld
.exec_all().group(1, 0)
8242 .MOV(sample_mask_reg(bld
.group(lower_width
, i
)),
8243 retype(dispatch_mask
, BRW_REGISTER_TYPE_UW
));
8252 if (wm_prog_data
->uses_kill
)
8253 bld
.emit(FS_OPCODE_PLACEHOLDER_HALT
);
8255 if (wm_key
->alpha_test_func
)
8260 if (shader_time_index
>= 0)
8261 emit_shader_time_end();
8267 assign_curb_setup();
8269 if (devinfo
->gen
>= 9)
8270 gen9_ps_header_only_workaround(wm_prog_data
);
8274 fixup_3src_null_dest();
8275 allocate_registers(8, allow_spilling
);
8285 fs_visitor::run_cs(unsigned min_dispatch_width
)
8287 assert(stage
== MESA_SHADER_COMPUTE
);
8288 assert(dispatch_width
>= min_dispatch_width
);
8292 if (shader_time_index
>= 0)
8293 emit_shader_time_begin();
8295 if (devinfo
->is_haswell
&& prog_data
->total_shared
> 0) {
8296 /* Move SLM index from g0.0[27:24] to sr0.1[11:8] */
8297 const fs_builder abld
= bld
.exec_all().group(1, 0);
8298 abld
.MOV(retype(brw_sr0_reg(1), BRW_REGISTER_TYPE_UW
),
8299 suboffset(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW
), 1));
8307 emit_cs_terminate();
8309 if (shader_time_index
>= 0)
8310 emit_shader_time_end();
8316 assign_curb_setup();
8318 fixup_3src_null_dest();
8319 allocate_registers(min_dispatch_width
, true);
8328 is_used_in_not_interp_frag_coord(nir_ssa_def
*def
)
8330 nir_foreach_use(src
, def
) {
8331 if (src
->parent_instr
->type
!= nir_instr_type_intrinsic
)
8334 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(src
->parent_instr
);
8335 if (intrin
->intrinsic
!= nir_intrinsic_load_frag_coord
)
8339 nir_foreach_if_use(src
, def
)
8346 * Return a bitfield where bit n is set if barycentric interpolation mode n
8347 * (see enum brw_barycentric_mode) is needed by the fragment shader.
8349 * We examine the load_barycentric intrinsics rather than looking at input
8350 * variables so that we catch interpolateAtCentroid() messages too, which
8351 * also need the BRW_BARYCENTRIC_[NON]PERSPECTIVE_CENTROID mode set up.
8354 brw_compute_barycentric_interp_modes(const struct gen_device_info
*devinfo
,
8355 const nir_shader
*shader
)
8357 unsigned barycentric_interp_modes
= 0;
8359 nir_foreach_function(f
, shader
) {
8363 nir_foreach_block(block
, f
->impl
) {
8364 nir_foreach_instr(instr
, block
) {
8365 if (instr
->type
!= nir_instr_type_intrinsic
)
8368 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
8369 switch (intrin
->intrinsic
) {
8370 case nir_intrinsic_load_barycentric_pixel
:
8371 case nir_intrinsic_load_barycentric_centroid
:
8372 case nir_intrinsic_load_barycentric_sample
:
8378 /* Ignore WPOS; it doesn't require interpolation. */
8379 assert(intrin
->dest
.is_ssa
);
8380 if (!is_used_in_not_interp_frag_coord(&intrin
->dest
.ssa
))
8383 enum glsl_interp_mode interp
= (enum glsl_interp_mode
)
8384 nir_intrinsic_interp_mode(intrin
);
8385 nir_intrinsic_op bary_op
= intrin
->intrinsic
;
8386 enum brw_barycentric_mode bary
=
8387 brw_barycentric_mode(interp
, bary_op
);
8389 barycentric_interp_modes
|= 1 << bary
;
8391 if (devinfo
->needs_unlit_centroid_workaround
&&
8392 bary_op
== nir_intrinsic_load_barycentric_centroid
)
8393 barycentric_interp_modes
|= 1 << centroid_to_pixel(bary
);
8398 return barycentric_interp_modes
;
8402 brw_compute_flat_inputs(struct brw_wm_prog_data
*prog_data
,
8403 const nir_shader
*shader
)
8405 prog_data
->flat_inputs
= 0;
8407 nir_foreach_variable(var
, &shader
->inputs
) {
8408 unsigned slots
= glsl_count_attribute_slots(var
->type
, false);
8409 for (unsigned s
= 0; s
< slots
; s
++) {
8410 int input_index
= prog_data
->urb_setup
[var
->data
.location
+ s
];
8412 if (input_index
< 0)
8416 if (var
->data
.interpolation
== INTERP_MODE_FLAT
)
8417 prog_data
->flat_inputs
|= 1 << input_index
;
8423 computed_depth_mode(const nir_shader
*shader
)
8425 if (shader
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_DEPTH
)) {
8426 switch (shader
->info
.fs
.depth_layout
) {
8427 case FRAG_DEPTH_LAYOUT_NONE
:
8428 case FRAG_DEPTH_LAYOUT_ANY
:
8429 return BRW_PSCDEPTH_ON
;
8430 case FRAG_DEPTH_LAYOUT_GREATER
:
8431 return BRW_PSCDEPTH_ON_GE
;
8432 case FRAG_DEPTH_LAYOUT_LESS
:
8433 return BRW_PSCDEPTH_ON_LE
;
8434 case FRAG_DEPTH_LAYOUT_UNCHANGED
:
8435 return BRW_PSCDEPTH_OFF
;
8438 return BRW_PSCDEPTH_OFF
;
8442 * Move load_interpolated_input with simple (payload-based) barycentric modes
8443 * to the top of the program so we don't emit multiple PLNs for the same input.
8445 * This works around CSE not being able to handle non-dominating cases
8451 * interpolate the same exact input
8454 * This should be replaced by global value numbering someday.
8457 move_interpolation_to_top(nir_shader
*nir
)
8459 bool progress
= false;
8461 nir_foreach_function(f
, nir
) {
8465 nir_block
*top
= nir_start_block(f
->impl
);
8466 exec_node
*cursor_node
= NULL
;
8468 nir_foreach_block(block
, f
->impl
) {
8472 nir_foreach_instr_safe(instr
, block
) {
8473 if (instr
->type
!= nir_instr_type_intrinsic
)
8476 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
8477 if (intrin
->intrinsic
!= nir_intrinsic_load_interpolated_input
)
8479 nir_intrinsic_instr
*bary_intrinsic
=
8480 nir_instr_as_intrinsic(intrin
->src
[0].ssa
->parent_instr
);
8481 nir_intrinsic_op op
= bary_intrinsic
->intrinsic
;
8483 /* Leave interpolateAtSample/Offset() where they are. */
8484 if (op
== nir_intrinsic_load_barycentric_at_sample
||
8485 op
== nir_intrinsic_load_barycentric_at_offset
)
8488 nir_instr
*move
[3] = {
8489 &bary_intrinsic
->instr
,
8490 intrin
->src
[1].ssa
->parent_instr
,
8494 for (unsigned i
= 0; i
< ARRAY_SIZE(move
); i
++) {
8495 if (move
[i
]->block
!= top
) {
8496 move
[i
]->block
= top
;
8497 exec_node_remove(&move
[i
]->node
);
8499 exec_node_insert_after(cursor_node
, &move
[i
]->node
);
8501 exec_list_push_head(&top
->instr_list
, &move
[i
]->node
);
8503 cursor_node
= &move
[i
]->node
;
8509 nir_metadata_preserve(f
->impl
, (nir_metadata
)
8510 ((unsigned) nir_metadata_block_index
|
8511 (unsigned) nir_metadata_dominance
));
8518 * Demote per-sample barycentric intrinsics to centroid.
8520 * Useful when rendering to a non-multisampled buffer.
8523 demote_sample_qualifiers(nir_shader
*nir
)
8525 bool progress
= true;
8527 nir_foreach_function(f
, nir
) {
8532 nir_builder_init(&b
, f
->impl
);
8534 nir_foreach_block(block
, f
->impl
) {
8535 nir_foreach_instr_safe(instr
, block
) {
8536 if (instr
->type
!= nir_instr_type_intrinsic
)
8539 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
8540 if (intrin
->intrinsic
!= nir_intrinsic_load_barycentric_sample
&&
8541 intrin
->intrinsic
!= nir_intrinsic_load_barycentric_at_sample
)
8544 b
.cursor
= nir_before_instr(instr
);
8545 nir_ssa_def
*centroid
=
8546 nir_load_barycentric(&b
, nir_intrinsic_load_barycentric_centroid
,
8547 nir_intrinsic_interp_mode(intrin
));
8548 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
8549 nir_src_for_ssa(centroid
));
8550 nir_instr_remove(instr
);
8555 nir_metadata_preserve(f
->impl
, (nir_metadata
)
8556 ((unsigned) nir_metadata_block_index
|
8557 (unsigned) nir_metadata_dominance
));
8564 * Pre-gen6, the register file of the EUs was shared between threads,
8565 * and each thread used some subset allocated on a 16-register block
8566 * granularity. The unit states wanted these block counts.
8569 brw_register_blocks(int reg_count
)
8571 return ALIGN(reg_count
, 16) / 16 - 1;
8575 brw_compile_fs(const struct brw_compiler
*compiler
, void *log_data
,
8577 const struct brw_wm_prog_key
*key
,
8578 struct brw_wm_prog_data
*prog_data
,
8580 int shader_time_index8
, int shader_time_index16
,
8581 int shader_time_index32
, bool allow_spilling
,
8582 bool use_rep_send
, struct brw_vue_map
*vue_map
,
8583 struct brw_compile_stats
*stats
,
8586 const struct gen_device_info
*devinfo
= compiler
->devinfo
;
8588 unsigned max_subgroup_size
= unlikely(INTEL_DEBUG
& DEBUG_DO32
) ? 32 : 16;
8590 brw_nir_apply_key(shader
, compiler
, &key
->base
, max_subgroup_size
, true);
8591 brw_nir_lower_fs_inputs(shader
, devinfo
, key
);
8592 brw_nir_lower_fs_outputs(shader
);
8594 if (devinfo
->gen
< 6)
8595 brw_setup_vue_interpolation(vue_map
, shader
, prog_data
);
8597 /* From the SKL PRM, Volume 7, "Alpha Coverage":
8598 * "If Pixel Shader outputs oMask, AlphaToCoverage is disabled in
8599 * hardware, regardless of the state setting for this feature."
8601 if (devinfo
->gen
> 6 && key
->alpha_to_coverage
) {
8602 /* Run constant fold optimization in order to get the correct source
8603 * offset to determine render target 0 store instruction in
8604 * emit_alpha_to_coverage pass.
8606 NIR_PASS_V(shader
, nir_opt_constant_folding
);
8607 NIR_PASS_V(shader
, brw_nir_lower_alpha_to_coverage
);
8610 if (!key
->multisample_fbo
)
8611 NIR_PASS_V(shader
, demote_sample_qualifiers
);
8612 NIR_PASS_V(shader
, move_interpolation_to_top
);
8613 brw_postprocess_nir(shader
, compiler
, true);
8615 /* key->alpha_test_func means simulating alpha testing via discards,
8616 * so the shader definitely kills pixels.
8618 prog_data
->uses_kill
= shader
->info
.fs
.uses_discard
||
8619 key
->alpha_test_func
;
8620 prog_data
->uses_omask
= key
->multisample_fbo
&&
8621 shader
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_SAMPLE_MASK
);
8622 prog_data
->computed_depth_mode
= computed_depth_mode(shader
);
8623 prog_data
->computed_stencil
=
8624 shader
->info
.outputs_written
& BITFIELD64_BIT(FRAG_RESULT_STENCIL
);
8626 prog_data
->persample_dispatch
=
8627 key
->multisample_fbo
&&
8628 (key
->persample_interp
||
8629 (shader
->info
.system_values_read
& (SYSTEM_BIT_SAMPLE_ID
|
8630 SYSTEM_BIT_SAMPLE_POS
)) ||
8631 shader
->info
.fs
.uses_sample_qualifier
||
8632 shader
->info
.outputs_read
);
8634 prog_data
->has_render_target_reads
= shader
->info
.outputs_read
!= 0ull;
8636 prog_data
->early_fragment_tests
= shader
->info
.fs
.early_fragment_tests
;
8637 prog_data
->post_depth_coverage
= shader
->info
.fs
.post_depth_coverage
;
8638 prog_data
->inner_coverage
= shader
->info
.fs
.inner_coverage
;
8640 prog_data
->barycentric_interp_modes
=
8641 brw_compute_barycentric_interp_modes(compiler
->devinfo
, shader
);
8643 calculate_urb_setup(devinfo
, key
, prog_data
, shader
);
8644 brw_compute_flat_inputs(prog_data
, shader
);
8646 cfg_t
*simd8_cfg
= NULL
, *simd16_cfg
= NULL
, *simd32_cfg
= NULL
;
8647 struct shader_stats v8_shader_stats
, v16_shader_stats
, v32_shader_stats
;
8649 fs_visitor
v8(compiler
, log_data
, mem_ctx
, &key
->base
,
8650 &prog_data
->base
, shader
, 8,
8651 shader_time_index8
);
8652 if (!v8
.run_fs(allow_spilling
, false /* do_rep_send */)) {
8654 *error_str
= ralloc_strdup(mem_ctx
, v8
.fail_msg
);
8657 } else if (likely(!(INTEL_DEBUG
& DEBUG_NO8
))) {
8659 v8_shader_stats
= v8
.shader_stats
;
8660 prog_data
->base
.dispatch_grf_start_reg
= v8
.payload
.num_regs
;
8661 prog_data
->reg_blocks_8
= brw_register_blocks(v8
.grf_used
);
8664 /* Limit dispatch width to simd8 with dual source blending on gen8.
8665 * See: https://gitlab.freedesktop.org/mesa/mesa/issues/1917
8667 if (devinfo
->gen
== 8 && prog_data
->dual_src_blend
&&
8668 !(INTEL_DEBUG
& DEBUG_NO8
)) {
8669 assert(!use_rep_send
);
8670 v8
.limit_dispatch_width(8, "gen8 workaround: "
8671 "using SIMD8 when dual src blending.\n");
8674 if (v8
.max_dispatch_width
>= 16 &&
8675 likely(!(INTEL_DEBUG
& DEBUG_NO16
) || use_rep_send
)) {
8676 /* Try a SIMD16 compile */
8677 fs_visitor
v16(compiler
, log_data
, mem_ctx
, &key
->base
,
8678 &prog_data
->base
, shader
, 16,
8679 shader_time_index16
);
8680 v16
.import_uniforms(&v8
);
8681 if (!v16
.run_fs(allow_spilling
, use_rep_send
)) {
8682 compiler
->shader_perf_log(log_data
,
8683 "SIMD16 shader failed to compile: %s",
8686 simd16_cfg
= v16
.cfg
;
8687 v16_shader_stats
= v16
.shader_stats
;
8688 prog_data
->dispatch_grf_start_reg_16
= v16
.payload
.num_regs
;
8689 prog_data
->reg_blocks_16
= brw_register_blocks(v16
.grf_used
);
8693 /* Currently, the compiler only supports SIMD32 on SNB+ */
8694 if (v8
.max_dispatch_width
>= 32 && !use_rep_send
&&
8695 compiler
->devinfo
->gen
>= 6 &&
8696 unlikely(INTEL_DEBUG
& DEBUG_DO32
)) {
8697 /* Try a SIMD32 compile */
8698 fs_visitor
v32(compiler
, log_data
, mem_ctx
, &key
->base
,
8699 &prog_data
->base
, shader
, 32,
8700 shader_time_index32
);
8701 v32
.import_uniforms(&v8
);
8702 if (!v32
.run_fs(allow_spilling
, false)) {
8703 compiler
->shader_perf_log(log_data
,
8704 "SIMD32 shader failed to compile: %s",
8707 simd32_cfg
= v32
.cfg
;
8708 v32_shader_stats
= v32
.shader_stats
;
8709 prog_data
->dispatch_grf_start_reg_32
= v32
.payload
.num_regs
;
8710 prog_data
->reg_blocks_32
= brw_register_blocks(v32
.grf_used
);
8714 /* When the caller requests a repclear shader, they want SIMD16-only */
8718 /* Prior to Iron Lake, the PS had a single shader offset with a jump table
8719 * at the top to select the shader. We've never implemented that.
8720 * Instead, we just give them exactly one shader and we pick the widest one
8723 if (compiler
->devinfo
->gen
< 5) {
8724 if (simd32_cfg
|| simd16_cfg
)
8730 /* If computed depth is enabled SNB only allows SIMD8. */
8731 if (compiler
->devinfo
->gen
== 6 &&
8732 prog_data
->computed_depth_mode
!= BRW_PSCDEPTH_OFF
)
8733 assert(simd16_cfg
== NULL
&& simd32_cfg
== NULL
);
8735 if (compiler
->devinfo
->gen
<= 5 && !simd8_cfg
) {
8736 /* Iron lake and earlier only have one Dispatch GRF start field. Make
8737 * the data available in the base prog data struct for convenience.
8740 prog_data
->base
.dispatch_grf_start_reg
=
8741 prog_data
->dispatch_grf_start_reg_16
;
8742 } else if (simd32_cfg
) {
8743 prog_data
->base
.dispatch_grf_start_reg
=
8744 prog_data
->dispatch_grf_start_reg_32
;
8748 if (prog_data
->persample_dispatch
) {
8749 /* Starting with SandyBridge (where we first get MSAA), the different
8750 * pixel dispatch combinations are grouped into classifications A
8751 * through F (SNB PRM Vol. 2 Part 1 Section 7.7.1). On all hardware
8752 * generations, the only configurations supporting persample dispatch
8753 * are are this in which only one dispatch width is enabled.
8755 if (simd32_cfg
|| simd16_cfg
)
8761 fs_generator
g(compiler
, log_data
, mem_ctx
, &prog_data
->base
,
8762 v8
.runtime_check_aads_emit
, MESA_SHADER_FRAGMENT
);
8764 if (unlikely(INTEL_DEBUG
& DEBUG_WM
)) {
8765 g
.enable_debug(ralloc_asprintf(mem_ctx
, "%s fragment shader %s",
8766 shader
->info
.label
?
8767 shader
->info
.label
: "unnamed",
8768 shader
->info
.name
));
8772 prog_data
->dispatch_8
= true;
8773 g
.generate_code(simd8_cfg
, 8, v8_shader_stats
, stats
);
8774 stats
= stats
? stats
+ 1 : NULL
;
8778 prog_data
->dispatch_16
= true;
8779 prog_data
->prog_offset_16
= g
.generate_code(simd16_cfg
, 16, v16_shader_stats
, stats
);
8780 stats
= stats
? stats
+ 1 : NULL
;
8784 prog_data
->dispatch_32
= true;
8785 prog_data
->prog_offset_32
= g
.generate_code(simd32_cfg
, 32, v32_shader_stats
, stats
);
8786 stats
= stats
? stats
+ 1 : NULL
;
8789 return g
.get_assembly();
8793 fs_visitor::emit_cs_work_group_id_setup()
8795 assert(stage
== MESA_SHADER_COMPUTE
);
8797 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::uvec3_type
));
8799 struct brw_reg
r0_1(retype(brw_vec1_grf(0, 1), BRW_REGISTER_TYPE_UD
));
8800 struct brw_reg
r0_6(retype(brw_vec1_grf(0, 6), BRW_REGISTER_TYPE_UD
));
8801 struct brw_reg
r0_7(retype(brw_vec1_grf(0, 7), BRW_REGISTER_TYPE_UD
));
8803 bld
.MOV(*reg
, r0_1
);
8804 bld
.MOV(offset(*reg
, bld
, 1), r0_6
);
8805 bld
.MOV(offset(*reg
, bld
, 2), r0_7
);
8811 fill_push_const_block_info(struct brw_push_const_block
*block
, unsigned dwords
)
8813 block
->dwords
= dwords
;
8814 block
->regs
= DIV_ROUND_UP(dwords
, 8);
8815 block
->size
= block
->regs
* 32;
8819 cs_fill_push_const_info(const struct gen_device_info
*devinfo
,
8820 struct brw_cs_prog_data
*cs_prog_data
)
8822 const struct brw_stage_prog_data
*prog_data
= &cs_prog_data
->base
;
8823 int subgroup_id_index
= get_subgroup_id_param_index(prog_data
);
8824 bool cross_thread_supported
= devinfo
->gen
> 7 || devinfo
->is_haswell
;
8826 /* The thread ID should be stored in the last param dword */
8827 assert(subgroup_id_index
== -1 ||
8828 subgroup_id_index
== (int)prog_data
->nr_params
- 1);
8830 unsigned cross_thread_dwords
, per_thread_dwords
;
8831 if (!cross_thread_supported
) {
8832 cross_thread_dwords
= 0u;
8833 per_thread_dwords
= prog_data
->nr_params
;
8834 } else if (subgroup_id_index
>= 0) {
8835 /* Fill all but the last register with cross-thread payload */
8836 cross_thread_dwords
= 8 * (subgroup_id_index
/ 8);
8837 per_thread_dwords
= prog_data
->nr_params
- cross_thread_dwords
;
8838 assert(per_thread_dwords
> 0 && per_thread_dwords
<= 8);
8840 /* Fill all data using cross-thread payload */
8841 cross_thread_dwords
= prog_data
->nr_params
;
8842 per_thread_dwords
= 0u;
8845 fill_push_const_block_info(&cs_prog_data
->push
.cross_thread
, cross_thread_dwords
);
8846 fill_push_const_block_info(&cs_prog_data
->push
.per_thread
, per_thread_dwords
);
8848 unsigned total_dwords
=
8849 (cs_prog_data
->push
.per_thread
.size
* cs_prog_data
->threads
+
8850 cs_prog_data
->push
.cross_thread
.size
) / 4;
8851 fill_push_const_block_info(&cs_prog_data
->push
.total
, total_dwords
);
8853 assert(cs_prog_data
->push
.cross_thread
.dwords
% 8 == 0 ||
8854 cs_prog_data
->push
.per_thread
.size
== 0);
8855 assert(cs_prog_data
->push
.cross_thread
.dwords
+
8856 cs_prog_data
->push
.per_thread
.dwords
==
8857 prog_data
->nr_params
);
8861 cs_set_simd_size(struct brw_cs_prog_data
*cs_prog_data
, unsigned size
)
8863 cs_prog_data
->simd_size
= size
;
8864 unsigned group_size
= cs_prog_data
->local_size
[0] *
8865 cs_prog_data
->local_size
[1] * cs_prog_data
->local_size
[2];
8866 cs_prog_data
->threads
= (group_size
+ size
- 1) / size
;
8870 compile_cs_to_nir(const struct brw_compiler
*compiler
,
8872 const struct brw_cs_prog_key
*key
,
8873 const nir_shader
*src_shader
,
8874 unsigned dispatch_width
)
8876 nir_shader
*shader
= nir_shader_clone(mem_ctx
, src_shader
);
8877 brw_nir_apply_key(shader
, compiler
, &key
->base
, dispatch_width
, true);
8879 NIR_PASS_V(shader
, brw_nir_lower_cs_intrinsics
, dispatch_width
);
8881 /* Clean up after the local index and ID calculations. */
8882 NIR_PASS_V(shader
, nir_opt_constant_folding
);
8883 NIR_PASS_V(shader
, nir_opt_dce
);
8885 brw_postprocess_nir(shader
, compiler
, true);
8891 brw_compile_cs(const struct brw_compiler
*compiler
, void *log_data
,
8893 const struct brw_cs_prog_key
*key
,
8894 struct brw_cs_prog_data
*prog_data
,
8895 const nir_shader
*src_shader
,
8896 int shader_time_index
,
8897 struct brw_compile_stats
*stats
,
8900 prog_data
->base
.total_shared
= src_shader
->info
.cs
.shared_size
;
8901 prog_data
->local_size
[0] = src_shader
->info
.cs
.local_size
[0];
8902 prog_data
->local_size
[1] = src_shader
->info
.cs
.local_size
[1];
8903 prog_data
->local_size
[2] = src_shader
->info
.cs
.local_size
[2];
8904 prog_data
->slm_size
= src_shader
->num_shared
;
8905 unsigned local_workgroup_size
=
8906 src_shader
->info
.cs
.local_size
[0] * src_shader
->info
.cs
.local_size
[1] *
8907 src_shader
->info
.cs
.local_size
[2];
8909 /* Limit max_threads to 64 for the GPGPU_WALKER command */
8910 const uint32_t max_threads
= MIN2(64, compiler
->devinfo
->max_cs_threads
);
8911 unsigned min_dispatch_width
=
8912 DIV_ROUND_UP(local_workgroup_size
, max_threads
);
8913 min_dispatch_width
= MAX2(8, min_dispatch_width
);
8914 min_dispatch_width
= util_next_power_of_two(min_dispatch_width
);
8915 assert(min_dispatch_width
<= 32);
8916 unsigned max_dispatch_width
= 32;
8918 fs_visitor
*v8
= NULL
, *v16
= NULL
, *v32
= NULL
;
8919 fs_visitor
*v
= NULL
;
8920 const char *fail_msg
= NULL
;
8922 if ((int)key
->base
.subgroup_size_type
>= (int)BRW_SUBGROUP_SIZE_REQUIRE_8
) {
8923 /* These enum values are expressly chosen to be equal to the subgroup
8924 * size that they require.
8926 const unsigned required_dispatch_width
=
8927 (unsigned)key
->base
.subgroup_size_type
;
8928 assert(required_dispatch_width
== 8 ||
8929 required_dispatch_width
== 16 ||
8930 required_dispatch_width
== 32);
8931 if (required_dispatch_width
< min_dispatch_width
||
8932 required_dispatch_width
> max_dispatch_width
) {
8933 fail_msg
= "Cannot satisfy explicit subgroup size";
8935 min_dispatch_width
= max_dispatch_width
= required_dispatch_width
;
8939 /* Now the main event: Visit the shader IR and generate our CS IR for it.
8941 if (!fail_msg
&& min_dispatch_width
<= 8 && max_dispatch_width
>= 8) {
8942 nir_shader
*nir8
= compile_cs_to_nir(compiler
, mem_ctx
, key
,
8944 v8
= new fs_visitor(compiler
, log_data
, mem_ctx
, &key
->base
,
8946 nir8
, 8, shader_time_index
);
8947 if (!v8
->run_cs(min_dispatch_width
)) {
8948 fail_msg
= v8
->fail_msg
;
8950 /* We should always be able to do SIMD32 for compute shaders */
8951 assert(v8
->max_dispatch_width
>= 32);
8954 cs_set_simd_size(prog_data
, 8);
8955 cs_fill_push_const_info(compiler
->devinfo
, prog_data
);
8959 if (likely(!(INTEL_DEBUG
& DEBUG_NO16
)) &&
8960 !fail_msg
&& min_dispatch_width
<= 16 && max_dispatch_width
>= 16) {
8961 /* Try a SIMD16 compile */
8962 nir_shader
*nir16
= compile_cs_to_nir(compiler
, mem_ctx
, key
,
8964 v16
= new fs_visitor(compiler
, log_data
, mem_ctx
, &key
->base
,
8966 nir16
, 16, shader_time_index
);
8968 v16
->import_uniforms(v8
);
8970 if (!v16
->run_cs(min_dispatch_width
)) {
8971 compiler
->shader_perf_log(log_data
,
8972 "SIMD16 shader failed to compile: %s",
8976 "Couldn't generate SIMD16 program and not "
8977 "enough threads for SIMD8";
8980 /* We should always be able to do SIMD32 for compute shaders */
8981 assert(v16
->max_dispatch_width
>= 32);
8984 cs_set_simd_size(prog_data
, 16);
8985 cs_fill_push_const_info(compiler
->devinfo
, prog_data
);
8989 /* We should always be able to do SIMD32 for compute shaders */
8990 assert(!v16
|| v16
->max_dispatch_width
>= 32);
8992 if (!fail_msg
&& (min_dispatch_width
> 16 || (INTEL_DEBUG
& DEBUG_DO32
)) &&
8993 max_dispatch_width
>= 32) {
8994 /* Try a SIMD32 compile */
8995 nir_shader
*nir32
= compile_cs_to_nir(compiler
, mem_ctx
, key
,
8997 v32
= new fs_visitor(compiler
, log_data
, mem_ctx
, &key
->base
,
8999 nir32
, 32, shader_time_index
);
9001 v32
->import_uniforms(v8
);
9003 v32
->import_uniforms(v16
);
9005 if (!v32
->run_cs(min_dispatch_width
)) {
9006 compiler
->shader_perf_log(log_data
,
9007 "SIMD32 shader failed to compile: %s",
9011 "Couldn't generate SIMD32 program and not "
9012 "enough threads for SIMD16";
9016 cs_set_simd_size(prog_data
, 32);
9017 cs_fill_push_const_info(compiler
->devinfo
, prog_data
);
9021 const unsigned *ret
= NULL
;
9022 if (unlikely(v
== NULL
)) {
9025 *error_str
= ralloc_strdup(mem_ctx
, fail_msg
);
9027 fs_generator
g(compiler
, log_data
, mem_ctx
, &prog_data
->base
,
9028 v
->runtime_check_aads_emit
, MESA_SHADER_COMPUTE
);
9029 if (INTEL_DEBUG
& DEBUG_CS
) {
9030 char *name
= ralloc_asprintf(mem_ctx
, "%s compute shader %s",
9031 src_shader
->info
.label
?
9032 src_shader
->info
.label
: "unnamed",
9033 src_shader
->info
.name
);
9034 g
.enable_debug(name
);
9037 g
.generate_code(v
->cfg
, prog_data
->simd_size
, v
->shader_stats
, stats
);
9039 ret
= g
.get_assembly();
9050 * Test the dispatch mask packing assumptions of
9051 * brw_stage_has_packed_dispatch(). Call this from e.g. the top of
9052 * fs_visitor::emit_nir_code() to cause a GPU hang if any shader invocation is
9053 * executed with an unexpected dispatch mask.
9056 brw_fs_test_dispatch_packing(const fs_builder
&bld
)
9058 const gl_shader_stage stage
= bld
.shader
->stage
;
9060 if (brw_stage_has_packed_dispatch(bld
.shader
->devinfo
, stage
,
9061 bld
.shader
->stage_prog_data
)) {
9062 const fs_builder ubld
= bld
.exec_all().group(1, 0);
9063 const fs_reg tmp
= component(bld
.vgrf(BRW_REGISTER_TYPE_UD
), 0);
9064 const fs_reg mask
= (stage
== MESA_SHADER_FRAGMENT
? brw_vmask_reg() :
9067 ubld
.ADD(tmp
, mask
, brw_imm_ud(1));
9068 ubld
.AND(tmp
, mask
, tmp
);
9070 /* This will loop forever if the dispatch mask doesn't have the expected
9071 * form '2^n-1', in which case tmp will be non-zero.
9073 bld
.emit(BRW_OPCODE_DO
);
9074 bld
.CMP(bld
.null_reg_ud(), tmp
, brw_imm_ud(0), BRW_CONDITIONAL_NZ
);
9075 set_predicate(BRW_PREDICATE_NORMAL
, bld
.emit(BRW_OPCODE_WHILE
));
9080 fs_visitor::workgroup_size() const
9082 assert(stage
== MESA_SHADER_COMPUTE
);
9083 const struct brw_cs_prog_data
*cs
= brw_cs_prog_data(prog_data
);
9084 return cs
->local_size
[0] * cs
->local_size
[1] * cs
->local_size
[2];