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 <sys/types.h>
33 #include "util/hash_table.h"
34 #include "main/macros.h"
35 #include "main/shaderobj.h"
36 #include "main/fbobject.h"
37 #include "program/prog_parameter.h"
38 #include "program/prog_print.h"
39 #include "util/register_allocate.h"
40 #include "program/hash_table.h"
41 #include "brw_context.h"
46 #include "brw_dead_control_flow.h"
47 #include "main/uniforms.h"
48 #include "brw_fs_live_variables.h"
49 #include "glsl/glsl_types.h"
50 #include "program/sampler.h"
55 fs_inst::init(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
56 const fs_reg
*src
, unsigned sources
)
58 memset(this, 0, sizeof(*this));
60 this->src
= new fs_reg
[MAX2(sources
, 3)];
61 for (unsigned i
= 0; i
< sources
; i
++)
62 this->src
[i
] = src
[i
];
64 this->opcode
= opcode
;
66 this->sources
= sources
;
67 this->exec_size
= exec_size
;
69 assert(dst
.file
!= IMM
&& dst
.file
!= UNIFORM
);
71 /* If exec_size == 0, try to guess it from the registers. Since all
72 * manner of things may use hardware registers, we first try to guess
73 * based on GRF registers. If this fails, we will go ahead and take the
74 * width from the destination register.
76 if (this->exec_size
== 0) {
77 if (dst
.file
== GRF
) {
78 this->exec_size
= dst
.width
;
80 for (unsigned i
= 0; i
< sources
; ++i
) {
81 if (src
[i
].file
!= GRF
&& src
[i
].file
!= ATTR
)
84 if (this->exec_size
<= 1)
85 this->exec_size
= src
[i
].width
;
86 assert(src
[i
].width
== 1 || src
[i
].width
== this->exec_size
);
90 if (this->exec_size
== 0 && dst
.file
!= BAD_FILE
)
91 this->exec_size
= dst
.width
;
93 assert(this->exec_size
!= 0);
95 this->conditional_mod
= BRW_CONDITIONAL_NONE
;
97 /* This will be the case for almost all instructions. */
104 DIV_ROUND_UP(MAX2(dst
.width
* dst
.stride
, 1) * type_sz(dst
.type
), 32);
107 this->regs_written
= 0;
111 unreachable("Invalid destination register file");
113 unreachable("Invalid register file");
116 this->writes_accumulator
= false;
121 init(BRW_OPCODE_NOP
, 8, dst
, NULL
, 0);
124 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
)
126 init(opcode
, exec_size
, reg_undef
, NULL
, 0);
129 fs_inst::fs_inst(enum opcode opcode
, const fs_reg
&dst
)
131 init(opcode
, 0, dst
, NULL
, 0);
134 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
137 const fs_reg src
[1] = { src0
};
138 init(opcode
, exec_size
, dst
, src
, 1);
141 fs_inst::fs_inst(enum opcode opcode
, const fs_reg
&dst
, const fs_reg
&src0
)
143 const fs_reg src
[1] = { src0
};
144 init(opcode
, 0, dst
, src
, 1);
147 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
148 const fs_reg
&src0
, const fs_reg
&src1
)
150 const fs_reg src
[2] = { src0
, src1
};
151 init(opcode
, exec_size
, dst
, src
, 2);
154 fs_inst::fs_inst(enum opcode opcode
, const fs_reg
&dst
, const fs_reg
&src0
,
157 const fs_reg src
[2] = { src0
, src1
};
158 init(opcode
, 0, dst
, src
, 2);
161 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
162 const fs_reg
&src0
, const fs_reg
&src1
, const fs_reg
&src2
)
164 const fs_reg src
[3] = { src0
, src1
, src2
};
165 init(opcode
, exec_size
, dst
, src
, 3);
168 fs_inst::fs_inst(enum opcode opcode
, const fs_reg
&dst
, const fs_reg
&src0
,
169 const fs_reg
&src1
, const fs_reg
&src2
)
171 const fs_reg src
[3] = { src0
, src1
, src2
};
172 init(opcode
, 0, dst
, src
, 3);
175 fs_inst::fs_inst(enum opcode opcode
, const fs_reg
&dst
,
176 const fs_reg src
[], unsigned sources
)
178 init(opcode
, 0, dst
, src
, sources
);
181 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_width
, const fs_reg
&dst
,
182 const fs_reg src
[], unsigned sources
)
184 init(opcode
, exec_width
, dst
, src
, sources
);
187 fs_inst::fs_inst(const fs_inst
&that
)
189 memcpy(this, &that
, sizeof(that
));
191 this->src
= new fs_reg
[MAX2(that
.sources
, 3)];
193 for (unsigned i
= 0; i
< that
.sources
; i
++)
194 this->src
[i
] = that
.src
[i
];
203 fs_inst::resize_sources(uint8_t num_sources
)
205 if (this->sources
!= num_sources
) {
206 fs_reg
*src
= new fs_reg
[MAX2(num_sources
, 3)];
208 for (unsigned i
= 0; i
< MIN2(this->sources
, num_sources
); ++i
)
209 src
[i
] = this->src
[i
];
213 this->sources
= num_sources
;
219 fs_visitor::op(const fs_reg &dst, const fs_reg &src0) \
221 return new(mem_ctx) fs_inst(BRW_OPCODE_##op, dst, src0); \
226 fs_visitor::op(const fs_reg &dst, const fs_reg &src0, \
227 const fs_reg &src1) \
229 return new(mem_ctx) fs_inst(BRW_OPCODE_##op, dst, src0, src1); \
232 #define ALU2_ACC(op) \
234 fs_visitor::op(const fs_reg &dst, const fs_reg &src0, \
235 const fs_reg &src1) \
237 fs_inst *inst = new(mem_ctx) fs_inst(BRW_OPCODE_##op, dst, src0, src1);\
238 inst->writes_accumulator = true; \
244 fs_visitor::op(const fs_reg &dst, const fs_reg &src0, \
245 const fs_reg &src1, const fs_reg &src2) \
247 return new(mem_ctx) fs_inst(BRW_OPCODE_##op, dst, src0, src1, src2);\
279 /** Gen4 predicated IF. */
281 fs_visitor::IF(enum brw_predicate predicate
)
283 fs_inst
*inst
= new(mem_ctx
) fs_inst(BRW_OPCODE_IF
, dispatch_width
);
284 inst
->predicate
= predicate
;
288 /** Gen6 IF with embedded comparison. */
290 fs_visitor::IF(const fs_reg
&src0
, const fs_reg
&src1
,
291 enum brw_conditional_mod condition
)
293 assert(devinfo
->gen
== 6);
294 fs_inst
*inst
= new(mem_ctx
) fs_inst(BRW_OPCODE_IF
, dispatch_width
,
295 reg_null_d
, src0
, src1
);
296 inst
->conditional_mod
= condition
;
301 * CMP: Sets the low bit of the destination channels with the result
302 * of the comparison, while the upper bits are undefined, and updates
303 * the flag register with the packed 16 bits of the result.
306 fs_visitor::CMP(fs_reg dst
, fs_reg src0
, fs_reg src1
,
307 enum brw_conditional_mod condition
)
311 /* Take the instruction:
313 * CMP null<d> src0<f> src1<f>
315 * Original gen4 does type conversion to the destination type before
316 * comparison, producing garbage results for floating point comparisons.
318 * The destination type doesn't matter on newer generations, so we set the
319 * type to match src0 so we can compact the instruction.
321 dst
.type
= src0
.type
;
322 if (dst
.file
== HW_REG
)
323 dst
.fixed_hw_reg
.type
= dst
.type
;
325 resolve_ud_negate(&src0
);
326 resolve_ud_negate(&src1
);
328 inst
= new(mem_ctx
) fs_inst(BRW_OPCODE_CMP
, dst
, src0
, src1
);
329 inst
->conditional_mod
= condition
;
335 fs_visitor::LOAD_PAYLOAD(const fs_reg
&dst
, fs_reg
*src
, int sources
,
338 assert(dst
.width
% 8 == 0);
339 fs_inst
*inst
= new(mem_ctx
) fs_inst(SHADER_OPCODE_LOAD_PAYLOAD
, dst
.width
,
341 inst
->header_size
= header_size
;
343 for (int i
= 0; i
< header_size
; i
++)
344 assert(src
[i
].file
!= GRF
|| src
[i
].width
* type_sz(src
[i
].type
) == 32);
345 inst
->regs_written
= header_size
;
347 for (int i
= header_size
; i
< sources
; ++i
)
348 assert(src
[i
].file
!= GRF
|| src
[i
].width
== dst
.width
);
349 inst
->regs_written
+= (sources
- header_size
) * (dst
.width
/ 8);
355 fs_visitor::VARYING_PULL_CONSTANT_LOAD(const fs_builder
&bld
,
357 const fs_reg
&surf_index
,
358 const fs_reg
&varying_offset
,
359 uint32_t const_offset
)
361 /* We have our constant surface use a pitch of 4 bytes, so our index can
362 * be any component of a vector, and then we load 4 contiguous
363 * components starting from that.
365 * We break down the const_offset to a portion added to the variable
366 * offset and a portion done using reg_offset, which means that if you
367 * have GLSL using something like "uniform vec4 a[20]; gl_FragColor =
368 * a[i]", we'll temporarily generate 4 vec4 loads from offset i * 4, and
369 * CSE can later notice that those loads are all the same and eliminate
370 * the redundant ones.
372 fs_reg vec4_offset
= vgrf(glsl_type::int_type
);
373 bld
.ADD(vec4_offset
, varying_offset
, fs_reg(const_offset
& ~3));
376 if (devinfo
->gen
== 4 && dst
.width
== 8) {
377 /* Pre-gen5, we can either use a SIMD8 message that requires (header,
378 * u, v, r) as parameters, or we can just use the SIMD16 message
379 * consisting of (header, u). We choose the second, at the cost of a
380 * longer return length.
386 if (devinfo
->gen
>= 7)
387 op
= FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7
;
389 op
= FS_OPCODE_VARYING_PULL_CONSTANT_LOAD
;
391 assert(dst
.width
% 8 == 0);
392 int regs_written
= 4 * (dst
.width
/ 8) * scale
;
393 fs_reg vec4_result
= fs_reg(GRF
, alloc
.allocate(regs_written
),
394 dst
.type
, dst
.width
);
395 fs_inst
*inst
= bld
.emit(op
, vec4_result
, surf_index
, vec4_offset
);
396 inst
->regs_written
= regs_written
;
398 if (devinfo
->gen
< 7) {
400 inst
->header_size
= 1;
401 if (devinfo
->gen
== 4)
404 inst
->mlen
= 1 + dispatch_width
/ 8;
407 bld
.MOV(dst
, offset(vec4_result
, (const_offset
& 3) * scale
));
411 * A helper for MOV generation for fixing up broken hardware SEND dependency
415 fs_visitor::DEP_RESOLVE_MOV(const fs_builder
&bld
, int grf
)
417 /* The caller always wants uncompressed to emit the minimal extra
418 * dependencies, and to avoid having to deal with aligning its regs to 2.
420 const fs_builder ubld
= bld
.annotate("send dependency resolve")
423 ubld
.MOV(ubld
.null_reg_f(), fs_reg(GRF
, grf
, BRW_REGISTER_TYPE_F
));
427 fs_inst::equals(fs_inst
*inst
) const
429 return (opcode
== inst
->opcode
&&
430 dst
.equals(inst
->dst
) &&
431 src
[0].equals(inst
->src
[0]) &&
432 src
[1].equals(inst
->src
[1]) &&
433 src
[2].equals(inst
->src
[2]) &&
434 saturate
== inst
->saturate
&&
435 predicate
== inst
->predicate
&&
436 conditional_mod
== inst
->conditional_mod
&&
437 mlen
== inst
->mlen
&&
438 base_mrf
== inst
->base_mrf
&&
439 target
== inst
->target
&&
441 header_size
== inst
->header_size
&&
442 shadow_compare
== inst
->shadow_compare
&&
443 exec_size
== inst
->exec_size
&&
444 offset
== inst
->offset
);
448 fs_inst::overwrites_reg(const fs_reg
®
) const
450 return reg
.in_range(dst
, regs_written
);
454 fs_inst::is_send_from_grf() const
457 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7
:
458 case SHADER_OPCODE_SHADER_TIME_ADD
:
459 case FS_OPCODE_INTERPOLATE_AT_CENTROID
:
460 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
461 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
462 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
463 case SHADER_OPCODE_UNTYPED_ATOMIC
:
464 case SHADER_OPCODE_UNTYPED_SURFACE_READ
:
465 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE
:
466 case SHADER_OPCODE_TYPED_ATOMIC
:
467 case SHADER_OPCODE_TYPED_SURFACE_READ
:
468 case SHADER_OPCODE_TYPED_SURFACE_WRITE
:
469 case SHADER_OPCODE_URB_WRITE_SIMD8
:
471 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
472 return src
[1].file
== GRF
;
473 case FS_OPCODE_FB_WRITE
:
474 return src
[0].file
== GRF
;
477 return src
[0].file
== GRF
;
484 fs_inst::is_copy_payload(const brw::simple_allocator
&grf_alloc
) const
486 if (this->opcode
!= SHADER_OPCODE_LOAD_PAYLOAD
)
489 fs_reg reg
= this->src
[0];
490 if (reg
.file
!= GRF
|| reg
.reg_offset
!= 0 || reg
.stride
== 0)
493 if (grf_alloc
.sizes
[reg
.reg
] != this->regs_written
)
496 for (int i
= 0; i
< this->sources
; i
++) {
497 reg
.type
= this->src
[i
].type
;
498 reg
.width
= this->src
[i
].width
;
499 if (!this->src
[i
].equals(reg
))
501 reg
= ::offset(reg
, 1);
508 fs_inst::can_do_source_mods(const struct brw_device_info
*devinfo
)
510 if (devinfo
->gen
== 6 && is_math())
513 if (is_send_from_grf())
516 if (!backend_instruction::can_do_source_mods())
523 fs_inst::has_side_effects() const
525 return this->eot
|| backend_instruction::has_side_effects();
531 memset(this, 0, sizeof(*this));
535 /** Generic unset register constructor. */
539 this->file
= BAD_FILE
;
542 /** Immediate value constructor. */
543 fs_reg::fs_reg(float f
)
547 this->type
= BRW_REGISTER_TYPE_F
;
548 this->fixed_hw_reg
.dw1
.f
= f
;
552 /** Immediate value constructor. */
553 fs_reg::fs_reg(int32_t i
)
557 this->type
= BRW_REGISTER_TYPE_D
;
558 this->fixed_hw_reg
.dw1
.d
= i
;
562 /** Immediate value constructor. */
563 fs_reg::fs_reg(uint32_t u
)
567 this->type
= BRW_REGISTER_TYPE_UD
;
568 this->fixed_hw_reg
.dw1
.ud
= u
;
572 /** Vector float immediate value constructor. */
573 fs_reg::fs_reg(uint8_t vf
[4])
577 this->type
= BRW_REGISTER_TYPE_VF
;
578 memcpy(&this->fixed_hw_reg
.dw1
.ud
, vf
, sizeof(unsigned));
581 /** Vector float immediate value constructor. */
582 fs_reg::fs_reg(uint8_t vf0
, uint8_t vf1
, uint8_t vf2
, uint8_t vf3
)
586 this->type
= BRW_REGISTER_TYPE_VF
;
587 this->fixed_hw_reg
.dw1
.ud
= (vf0
<< 0) |
593 /** Fixed brw_reg. */
594 fs_reg::fs_reg(struct brw_reg fixed_hw_reg
)
598 this->fixed_hw_reg
= fixed_hw_reg
;
599 this->type
= fixed_hw_reg
.type
;
600 this->width
= 1 << fixed_hw_reg
.width
;
604 fs_reg::equals(const fs_reg
&r
) const
606 return (file
== r
.file
&&
608 reg_offset
== r
.reg_offset
&&
609 subreg_offset
== r
.subreg_offset
&&
611 negate
== r
.negate
&&
613 !reladdr
&& !r
.reladdr
&&
614 memcmp(&fixed_hw_reg
, &r
.fixed_hw_reg
, sizeof(fixed_hw_reg
)) == 0 &&
620 fs_reg::set_smear(unsigned subreg
)
622 assert(file
!= HW_REG
&& file
!= IMM
);
623 subreg_offset
= subreg
* type_sz(type
);
629 fs_reg::is_contiguous() const
635 fs_visitor::type_size(const struct glsl_type
*type
)
637 unsigned int size
, i
;
639 switch (type
->base_type
) {
642 case GLSL_TYPE_FLOAT
:
644 return type
->components();
645 case GLSL_TYPE_ARRAY
:
646 return type_size(type
->fields
.array
) * type
->length
;
647 case GLSL_TYPE_STRUCT
:
649 for (i
= 0; i
< type
->length
; i
++) {
650 size
+= type_size(type
->fields
.structure
[i
].type
);
653 case GLSL_TYPE_SAMPLER
:
654 /* Samplers take up no register space, since they're baked in at
658 case GLSL_TYPE_ATOMIC_UINT
:
660 case GLSL_TYPE_IMAGE
:
662 case GLSL_TYPE_ERROR
:
663 case GLSL_TYPE_INTERFACE
:
664 case GLSL_TYPE_DOUBLE
:
665 unreachable("not reached");
672 * Create a MOV to read the timestamp register.
674 * The caller is responsible for emitting the MOV. The return value is
675 * the destination of the MOV, with extra parameters set.
678 fs_visitor::get_timestamp(const fs_builder
&bld
)
680 assert(devinfo
->gen
>= 7);
682 fs_reg ts
= fs_reg(retype(brw_vec4_reg(BRW_ARCHITECTURE_REGISTER_FILE
,
685 BRW_REGISTER_TYPE_UD
));
687 fs_reg dst
= fs_reg(GRF
, alloc
.allocate(1), BRW_REGISTER_TYPE_UD
, 4);
689 /* We want to read the 3 fields we care about even if it's not enabled in
692 bld
.exec_all().MOV(dst
, ts
);
694 /* The caller wants the low 32 bits of the timestamp. Since it's running
695 * at the GPU clock rate of ~1.2ghz, it will roll over every ~3 seconds,
696 * which is plenty of time for our purposes. It is identical across the
697 * EUs, but since it's tracking GPU core speed it will increment at a
698 * varying rate as render P-states change.
700 * The caller could also check if render P-states have changed (or anything
701 * else that might disrupt timing) by setting smear to 2 and checking if
702 * that field is != 0.
710 fs_visitor::emit_shader_time_begin()
712 shader_start_time
= get_timestamp(bld
.annotate("shader time start"));
716 fs_visitor::emit_shader_time_end()
718 enum shader_time_shader_type type
, written_type
, reset_type
;
720 case MESA_SHADER_VERTEX
:
722 written_type
= ST_VS_WRITTEN
;
723 reset_type
= ST_VS_RESET
;
725 case MESA_SHADER_GEOMETRY
:
727 written_type
= ST_GS_WRITTEN
;
728 reset_type
= ST_GS_RESET
;
730 case MESA_SHADER_FRAGMENT
:
731 if (dispatch_width
== 8) {
733 written_type
= ST_FS8_WRITTEN
;
734 reset_type
= ST_FS8_RESET
;
736 assert(dispatch_width
== 16);
738 written_type
= ST_FS16_WRITTEN
;
739 reset_type
= ST_FS16_RESET
;
742 case MESA_SHADER_COMPUTE
:
744 written_type
= ST_CS_WRITTEN
;
745 reset_type
= ST_CS_RESET
;
748 unreachable("fs_visitor::emit_shader_time_end missing code");
751 /* Insert our code just before the final SEND with EOT. */
752 exec_node
*end
= this->instructions
.get_tail();
753 assert(end
&& ((fs_inst
*) end
)->eot
);
754 const fs_builder ibld
= bld
.annotate("shader time end")
755 .exec_all().at(NULL
, end
);
757 fs_reg shader_end_time
= get_timestamp(ibld
);
759 /* Check that there weren't any timestamp reset events (assuming these
760 * were the only two timestamp reads that happened).
762 fs_reg reset
= shader_end_time
;
764 set_condmod(BRW_CONDITIONAL_Z
,
765 ibld
.AND(ibld
.null_reg_ud(), reset
, fs_reg(1u)));
766 ibld
.IF(BRW_PREDICATE_NORMAL
);
768 fs_reg start
= shader_start_time
;
770 fs_reg diff
= fs_reg(GRF
, alloc
.allocate(1), BRW_REGISTER_TYPE_UD
, 1);
772 ibld
.ADD(diff
, start
, shader_end_time
);
774 /* If there were no instructions between the two timestamp gets, the diff
775 * is 2 cycles. Remove that overhead, so I can forget about that when
776 * trying to determine the time taken for single instructions.
778 ibld
.ADD(diff
, diff
, fs_reg(-2u));
779 SHADER_TIME_ADD(ibld
, type
, diff
);
780 SHADER_TIME_ADD(ibld
, written_type
, fs_reg(1u));
781 ibld
.emit(BRW_OPCODE_ELSE
);
782 SHADER_TIME_ADD(ibld
, reset_type
, fs_reg(1u));
783 ibld
.emit(BRW_OPCODE_ENDIF
);
787 fs_visitor::SHADER_TIME_ADD(const fs_builder
&bld
,
788 enum shader_time_shader_type type
, fs_reg value
)
790 int shader_time_index
=
791 brw_get_shader_time_index(brw
, shader_prog
, prog
, type
);
792 fs_reg offset
= fs_reg(shader_time_index
* SHADER_TIME_STRIDE
);
795 if (dispatch_width
== 8)
796 payload
= vgrf(glsl_type::uvec2_type
);
798 payload
= vgrf(glsl_type::uint_type
);
800 bld
.emit(SHADER_OPCODE_SHADER_TIME_ADD
, fs_reg(), payload
, offset
, value
);
804 fs_visitor::vfail(const char *format
, va_list va
)
813 msg
= ralloc_vasprintf(mem_ctx
, format
, va
);
814 msg
= ralloc_asprintf(mem_ctx
, "%s compile failed: %s\n", stage_abbrev
, msg
);
816 this->fail_msg
= msg
;
819 fprintf(stderr
, "%s", msg
);
824 fs_visitor::fail(const char *format
, ...)
828 va_start(va
, format
);
834 * Mark this program as impossible to compile in SIMD16 mode.
836 * During the SIMD8 compile (which happens first), we can detect and flag
837 * things that are unsupported in SIMD16 mode, so the compiler can skip
838 * the SIMD16 compile altogether.
840 * During a SIMD16 compile (if one happens anyway), this just calls fail().
843 fs_visitor::no16(const char *format
, ...)
847 va_start(va
, format
);
849 if (dispatch_width
== 16) {
852 simd16_unsupported
= true;
854 if (brw
->perf_debug
) {
856 ralloc_vasprintf_append(&no16_msg
, format
, va
);
858 no16_msg
= ralloc_vasprintf(mem_ctx
, format
, va
);
866 fs_visitor::emit(enum opcode opcode
)
868 return emit(new(mem_ctx
) fs_inst(opcode
, dispatch_width
));
872 fs_visitor::emit(enum opcode opcode
, const fs_reg
&dst
)
874 return emit(new(mem_ctx
) fs_inst(opcode
, dst
));
878 fs_visitor::emit(enum opcode opcode
, const fs_reg
&dst
, const fs_reg
&src0
)
880 return emit(new(mem_ctx
) fs_inst(opcode
, dst
, src0
));
884 fs_visitor::emit(enum opcode opcode
, const fs_reg
&dst
, const fs_reg
&src0
,
887 return emit(new(mem_ctx
) fs_inst(opcode
, dst
, src0
, src1
));
891 fs_visitor::emit(enum opcode opcode
, const fs_reg
&dst
, const fs_reg
&src0
,
892 const fs_reg
&src1
, const fs_reg
&src2
)
894 return emit(new(mem_ctx
) fs_inst(opcode
, dst
, src0
, src1
, src2
));
898 fs_visitor::emit(enum opcode opcode
, const fs_reg
&dst
,
899 fs_reg src
[], int sources
)
901 return emit(new(mem_ctx
) fs_inst(opcode
, dst
, src
, sources
));
905 * Returns true if the instruction has a flag that means it won't
906 * update an entire destination register.
908 * For example, dead code elimination and live variable analysis want to know
909 * when a write to a variable screens off any preceding values that were in
913 fs_inst::is_partial_write() const
915 return ((this->predicate
&& this->opcode
!= BRW_OPCODE_SEL
) ||
916 (this->dst
.width
* type_sz(this->dst
.type
)) < 32 ||
917 !this->dst
.is_contiguous());
921 fs_inst::regs_read(int arg
) const
923 if (is_tex() && arg
== 0 && src
[0].file
== GRF
) {
925 } else if (opcode
== FS_OPCODE_FB_WRITE
&& arg
== 0) {
927 } else if (opcode
== SHADER_OPCODE_URB_WRITE_SIMD8
&& arg
== 0) {
929 } else if (opcode
== SHADER_OPCODE_UNTYPED_ATOMIC
&& arg
== 0) {
931 } else if (opcode
== SHADER_OPCODE_UNTYPED_SURFACE_READ
&& arg
== 0) {
933 } else if (opcode
== SHADER_OPCODE_UNTYPED_SURFACE_WRITE
&& arg
== 0) {
935 } else if (opcode
== SHADER_OPCODE_TYPED_ATOMIC
&& arg
== 0) {
937 } else if (opcode
== SHADER_OPCODE_TYPED_SURFACE_READ
&& arg
== 0) {
939 } else if (opcode
== SHADER_OPCODE_TYPED_SURFACE_WRITE
&& arg
== 0) {
941 } else if (opcode
== FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
&& arg
== 0) {
943 } else if (opcode
== FS_OPCODE_LINTERP
&& arg
== 0) {
944 return exec_size
/ 4;
947 switch (src
[arg
].file
) {
954 if (src
[arg
].stride
== 0) {
957 int size
= src
[arg
].width
* src
[arg
].stride
* type_sz(src
[arg
].type
);
958 return (size
+ 31) / 32;
961 unreachable("MRF registers are not allowed as sources");
963 unreachable("Invalid register file");
968 fs_inst::reads_flag() const
974 fs_inst::writes_flag() const
976 return (conditional_mod
&& (opcode
!= BRW_OPCODE_SEL
&&
977 opcode
!= BRW_OPCODE_IF
&&
978 opcode
!= BRW_OPCODE_WHILE
)) ||
979 opcode
== FS_OPCODE_MOV_DISPATCH_TO_FLAGS
;
983 * Returns how many MRFs an FS opcode will write over.
985 * Note that this is not the 0 or 1 implied writes in an actual gen
986 * instruction -- the FS opcodes often generate MOVs in addition.
989 fs_visitor::implied_mrf_writes(fs_inst
*inst
)
994 if (inst
->base_mrf
== -1)
997 switch (inst
->opcode
) {
998 case SHADER_OPCODE_RCP
:
999 case SHADER_OPCODE_RSQ
:
1000 case SHADER_OPCODE_SQRT
:
1001 case SHADER_OPCODE_EXP2
:
1002 case SHADER_OPCODE_LOG2
:
1003 case SHADER_OPCODE_SIN
:
1004 case SHADER_OPCODE_COS
:
1005 return 1 * dispatch_width
/ 8;
1006 case SHADER_OPCODE_POW
:
1007 case SHADER_OPCODE_INT_QUOTIENT
:
1008 case SHADER_OPCODE_INT_REMAINDER
:
1009 return 2 * dispatch_width
/ 8;
1010 case SHADER_OPCODE_TEX
:
1012 case SHADER_OPCODE_TXD
:
1013 case SHADER_OPCODE_TXF
:
1014 case SHADER_OPCODE_TXF_CMS
:
1015 case SHADER_OPCODE_TXF_MCS
:
1016 case SHADER_OPCODE_TG4
:
1017 case SHADER_OPCODE_TG4_OFFSET
:
1018 case SHADER_OPCODE_TXL
:
1019 case SHADER_OPCODE_TXS
:
1020 case SHADER_OPCODE_LOD
:
1022 case FS_OPCODE_FB_WRITE
:
1024 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
1025 case SHADER_OPCODE_GEN4_SCRATCH_READ
:
1027 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD
:
1029 case SHADER_OPCODE_GEN4_SCRATCH_WRITE
:
1031 case SHADER_OPCODE_UNTYPED_ATOMIC
:
1032 case SHADER_OPCODE_UNTYPED_SURFACE_READ
:
1033 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE
:
1034 case SHADER_OPCODE_TYPED_ATOMIC
:
1035 case SHADER_OPCODE_TYPED_SURFACE_READ
:
1036 case SHADER_OPCODE_TYPED_SURFACE_WRITE
:
1037 case SHADER_OPCODE_URB_WRITE_SIMD8
:
1038 case FS_OPCODE_INTERPOLATE_AT_CENTROID
:
1039 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
1040 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
1041 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
1044 unreachable("not reached");
1049 fs_visitor::vgrf(const glsl_type
*const type
)
1051 int reg_width
= dispatch_width
/ 8;
1052 return fs_reg(GRF
, alloc
.allocate(type_size(type
) * reg_width
),
1053 brw_type_for_base_type(type
), dispatch_width
);
1057 fs_visitor::vgrf(int num_components
)
1059 int reg_width
= dispatch_width
/ 8;
1060 return fs_reg(GRF
, alloc
.allocate(num_components
* reg_width
),
1061 BRW_REGISTER_TYPE_F
, dispatch_width
);
1064 /** Fixed HW reg constructor. */
1065 fs_reg::fs_reg(enum register_file file
, int reg
)
1070 this->type
= BRW_REGISTER_TYPE_F
;
1081 /** Fixed HW reg constructor. */
1082 fs_reg::fs_reg(enum register_file file
, int reg
, enum brw_reg_type type
)
1098 /** Fixed HW reg constructor. */
1099 fs_reg::fs_reg(enum register_file file
, int reg
, enum brw_reg_type type
,
1106 this->width
= width
;
1109 /* For SIMD16, we need to follow from the uniform setup of SIMD8 dispatch.
1110 * This brings in those uniform definitions
1113 fs_visitor::import_uniforms(fs_visitor
*v
)
1115 this->push_constant_loc
= v
->push_constant_loc
;
1116 this->pull_constant_loc
= v
->pull_constant_loc
;
1117 this->uniforms
= v
->uniforms
;
1118 this->param_size
= v
->param_size
;
1122 fs_visitor::emit_fragcoord_interpolation(bool pixel_center_integer
,
1123 bool origin_upper_left
)
1125 assert(stage
== MESA_SHADER_FRAGMENT
);
1126 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1127 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::vec4_type
));
1129 bool flip
= !origin_upper_left
^ key
->render_to_fbo
;
1131 /* gl_FragCoord.x */
1132 if (pixel_center_integer
) {
1133 bld
.MOV(wpos
, this->pixel_x
);
1135 bld
.ADD(wpos
, this->pixel_x
, fs_reg(0.5f
));
1137 wpos
= offset(wpos
, 1);
1139 /* gl_FragCoord.y */
1140 if (!flip
&& pixel_center_integer
) {
1141 bld
.MOV(wpos
, this->pixel_y
);
1143 fs_reg pixel_y
= this->pixel_y
;
1144 float offset
= (pixel_center_integer
? 0.0 : 0.5);
1147 pixel_y
.negate
= true;
1148 offset
+= key
->drawable_height
- 1.0;
1151 bld
.ADD(wpos
, pixel_y
, fs_reg(offset
));
1153 wpos
= offset(wpos
, 1);
1155 /* gl_FragCoord.z */
1156 if (devinfo
->gen
>= 6) {
1157 bld
.MOV(wpos
, fs_reg(brw_vec8_grf(payload
.source_depth_reg
, 0)));
1159 bld
.emit(FS_OPCODE_LINTERP
, wpos
,
1160 this->delta_xy
[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
],
1161 interp_reg(VARYING_SLOT_POS
, 2));
1163 wpos
= offset(wpos
, 1);
1165 /* gl_FragCoord.w: Already set up in emit_interpolation */
1166 bld
.MOV(wpos
, this->wpos_w
);
1172 fs_visitor::emit_linterp(const fs_reg
&attr
, const fs_reg
&interp
,
1173 glsl_interp_qualifier interpolation_mode
,
1174 bool is_centroid
, bool is_sample
)
1176 brw_wm_barycentric_interp_mode barycoord_mode
;
1177 if (devinfo
->gen
>= 6) {
1179 if (interpolation_mode
== INTERP_QUALIFIER_SMOOTH
)
1180 barycoord_mode
= BRW_WM_PERSPECTIVE_CENTROID_BARYCENTRIC
;
1182 barycoord_mode
= BRW_WM_NONPERSPECTIVE_CENTROID_BARYCENTRIC
;
1183 } else if (is_sample
) {
1184 if (interpolation_mode
== INTERP_QUALIFIER_SMOOTH
)
1185 barycoord_mode
= BRW_WM_PERSPECTIVE_SAMPLE_BARYCENTRIC
;
1187 barycoord_mode
= BRW_WM_NONPERSPECTIVE_SAMPLE_BARYCENTRIC
;
1189 if (interpolation_mode
== INTERP_QUALIFIER_SMOOTH
)
1190 barycoord_mode
= BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
;
1192 barycoord_mode
= BRW_WM_NONPERSPECTIVE_PIXEL_BARYCENTRIC
;
1195 /* On Ironlake and below, there is only one interpolation mode.
1196 * Centroid interpolation doesn't mean anything on this hardware --
1197 * there is no multisampling.
1199 barycoord_mode
= BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
;
1201 return bld
.emit(FS_OPCODE_LINTERP
, attr
,
1202 this->delta_xy
[barycoord_mode
], interp
);
1206 fs_visitor::emit_general_interpolation(fs_reg attr
, const char *name
,
1207 const glsl_type
*type
,
1208 glsl_interp_qualifier interpolation_mode
,
1209 int location
, bool mod_centroid
,
1212 attr
.type
= brw_type_for_base_type(type
->get_scalar_type());
1214 assert(stage
== MESA_SHADER_FRAGMENT
);
1215 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
1216 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1218 unsigned int array_elements
;
1220 if (type
->is_array()) {
1221 array_elements
= type
->length
;
1222 if (array_elements
== 0) {
1223 fail("dereferenced array '%s' has length 0\n", name
);
1225 type
= type
->fields
.array
;
1230 if (interpolation_mode
== INTERP_QUALIFIER_NONE
) {
1232 location
== VARYING_SLOT_COL0
|| location
== VARYING_SLOT_COL1
;
1233 if (key
->flat_shade
&& is_gl_Color
) {
1234 interpolation_mode
= INTERP_QUALIFIER_FLAT
;
1236 interpolation_mode
= INTERP_QUALIFIER_SMOOTH
;
1240 for (unsigned int i
= 0; i
< array_elements
; i
++) {
1241 for (unsigned int j
= 0; j
< type
->matrix_columns
; j
++) {
1242 if (prog_data
->urb_setup
[location
] == -1) {
1243 /* If there's no incoming setup data for this slot, don't
1244 * emit interpolation for it.
1246 attr
= offset(attr
, type
->vector_elements
);
1251 if (interpolation_mode
== INTERP_QUALIFIER_FLAT
) {
1252 /* Constant interpolation (flat shading) case. The SF has
1253 * handed us defined values in only the constant offset
1254 * field of the setup reg.
1256 for (unsigned int k
= 0; k
< type
->vector_elements
; k
++) {
1257 struct brw_reg interp
= interp_reg(location
, k
);
1258 interp
= suboffset(interp
, 3);
1259 interp
.type
= attr
.type
;
1260 bld
.emit(FS_OPCODE_CINTERP
, attr
, fs_reg(interp
));
1261 attr
= offset(attr
, 1);
1264 /* Smooth/noperspective interpolation case. */
1265 for (unsigned int k
= 0; k
< type
->vector_elements
; k
++) {
1266 struct brw_reg interp
= interp_reg(location
, k
);
1267 if (devinfo
->needs_unlit_centroid_workaround
&& mod_centroid
) {
1268 /* Get the pixel/sample mask into f0 so that we know
1269 * which pixels are lit. Then, for each channel that is
1270 * unlit, replace the centroid data with non-centroid
1273 bld
.emit(FS_OPCODE_MOV_DISPATCH_TO_FLAGS
);
1276 inst
= emit_linterp(attr
, fs_reg(interp
), interpolation_mode
,
1278 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1279 inst
->predicate_inverse
= true;
1280 if (devinfo
->has_pln
)
1281 inst
->no_dd_clear
= true;
1283 inst
= emit_linterp(attr
, fs_reg(interp
), interpolation_mode
,
1284 mod_centroid
&& !key
->persample_shading
,
1285 mod_sample
|| key
->persample_shading
);
1286 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1287 inst
->predicate_inverse
= false;
1288 if (devinfo
->has_pln
)
1289 inst
->no_dd_check
= true;
1292 emit_linterp(attr
, fs_reg(interp
), interpolation_mode
,
1293 mod_centroid
&& !key
->persample_shading
,
1294 mod_sample
|| key
->persample_shading
);
1296 if (devinfo
->gen
< 6 && interpolation_mode
== INTERP_QUALIFIER_SMOOTH
) {
1297 bld
.MUL(attr
, attr
, this->pixel_w
);
1299 attr
= offset(attr
, 1);
1309 fs_visitor::emit_frontfacing_interpolation()
1311 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::bool_type
));
1313 if (devinfo
->gen
>= 6) {
1314 /* Bit 15 of g0.0 is 0 if the polygon is front facing. We want to create
1315 * a boolean result from this (~0/true or 0/false).
1317 * We can use the fact that bit 15 is the MSB of g0.0:W to accomplish
1318 * this task in only one instruction:
1319 * - a negation source modifier will flip the bit; and
1320 * - a W -> D type conversion will sign extend the bit into the high
1321 * word of the destination.
1323 * An ASR 15 fills the low word of the destination.
1325 fs_reg g0
= fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_W
));
1328 bld
.ASR(*reg
, g0
, fs_reg(15));
1330 /* Bit 31 of g1.6 is 0 if the polygon is front facing. We want to create
1331 * a boolean result from this (1/true or 0/false).
1333 * Like in the above case, since the bit is the MSB of g1.6:UD we can use
1334 * the negation source modifier to flip it. Unfortunately the SHR
1335 * instruction only operates on UD (or D with an abs source modifier)
1336 * sources without negation.
1338 * Instead, use ASR (which will give ~0/true or 0/false).
1340 fs_reg g1_6
= fs_reg(retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_D
));
1343 bld
.ASR(*reg
, g1_6
, fs_reg(31));
1350 fs_visitor::compute_sample_position(fs_reg dst
, fs_reg int_sample_pos
)
1352 assert(stage
== MESA_SHADER_FRAGMENT
);
1353 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1354 assert(dst
.type
== BRW_REGISTER_TYPE_F
);
1356 if (key
->compute_pos_offset
) {
1357 /* Convert int_sample_pos to floating point */
1358 emit(MOV(dst
, int_sample_pos
));
1359 /* Scale to the range [0, 1] */
1360 emit(MUL(dst
, dst
, fs_reg(1 / 16.0f
)));
1363 /* From ARB_sample_shading specification:
1364 * "When rendering to a non-multisample buffer, or if multisample
1365 * rasterization is disabled, gl_SamplePosition will always be
1368 emit(MOV(dst
, fs_reg(0.5f
)));
1373 fs_visitor::emit_samplepos_setup()
1375 assert(devinfo
->gen
>= 6);
1377 this->current_annotation
= "compute sample position";
1378 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::vec2_type
));
1380 fs_reg int_sample_x
= vgrf(glsl_type::int_type
);
1381 fs_reg int_sample_y
= vgrf(glsl_type::int_type
);
1383 /* WM will be run in MSDISPMODE_PERSAMPLE. So, only one of SIMD8 or SIMD16
1384 * mode will be enabled.
1386 * From the Ivy Bridge PRM, volume 2 part 1, page 344:
1387 * R31.1:0 Position Offset X/Y for Slot[3:0]
1388 * R31.3:2 Position Offset X/Y for Slot[7:4]
1391 * The X, Y sample positions come in as bytes in thread payload. So, read
1392 * the positions using vstride=16, width=8, hstride=2.
1394 struct brw_reg sample_pos_reg
=
1395 stride(retype(brw_vec1_grf(payload
.sample_pos_reg
, 0),
1396 BRW_REGISTER_TYPE_B
), 16, 8, 2);
1398 if (dispatch_width
== 8) {
1399 emit(MOV(int_sample_x
, fs_reg(sample_pos_reg
)));
1401 emit(MOV(half(int_sample_x
, 0), fs_reg(sample_pos_reg
)));
1402 emit(MOV(half(int_sample_x
, 1), fs_reg(suboffset(sample_pos_reg
, 16))))
1403 ->force_sechalf
= true;
1405 /* Compute gl_SamplePosition.x */
1406 compute_sample_position(pos
, int_sample_x
);
1407 pos
= offset(pos
, 1);
1408 if (dispatch_width
== 8) {
1409 emit(MOV(int_sample_y
, fs_reg(suboffset(sample_pos_reg
, 1))));
1411 emit(MOV(half(int_sample_y
, 0),
1412 fs_reg(suboffset(sample_pos_reg
, 1))));
1413 emit(MOV(half(int_sample_y
, 1), fs_reg(suboffset(sample_pos_reg
, 17))))
1414 ->force_sechalf
= true;
1416 /* Compute gl_SamplePosition.y */
1417 compute_sample_position(pos
, int_sample_y
);
1422 fs_visitor::emit_sampleid_setup()
1424 assert(stage
== MESA_SHADER_FRAGMENT
);
1425 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1426 assert(devinfo
->gen
>= 6);
1428 this->current_annotation
= "compute sample id";
1429 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::int_type
));
1431 if (key
->compute_sample_id
) {
1432 fs_reg t1
= vgrf(glsl_type::int_type
);
1433 fs_reg t2
= vgrf(glsl_type::int_type
);
1434 t2
.type
= BRW_REGISTER_TYPE_UW
;
1436 /* The PS will be run in MSDISPMODE_PERSAMPLE. For example with
1437 * 8x multisampling, subspan 0 will represent sample N (where N
1438 * is 0, 2, 4 or 6), subspan 1 will represent sample 1, 3, 5 or
1439 * 7. We can find the value of N by looking at R0.0 bits 7:6
1440 * ("Starting Sample Pair Index (SSPI)") and multiplying by two
1441 * (since samples are always delivered in pairs). That is, we
1442 * compute 2*((R0.0 & 0xc0) >> 6) == (R0.0 & 0xc0) >> 5. Then
1443 * we need to add N to the sequence (0, 0, 0, 0, 1, 1, 1, 1) in
1444 * case of SIMD8 and sequence (0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2,
1445 * 2, 3, 3, 3, 3) in case of SIMD16. We compute this sequence by
1446 * populating a temporary variable with the sequence (0, 1, 2, 3),
1447 * and then reading from it using vstride=1, width=4, hstride=0.
1448 * These computations hold good for 4x multisampling as well.
1450 * For 2x MSAA and SIMD16, we want to use the sequence (0, 1, 0, 1):
1451 * the first four slots are sample 0 of subspan 0; the next four
1452 * are sample 1 of subspan 0; the third group is sample 0 of
1453 * subspan 1, and finally sample 1 of subspan 1.
1456 inst
= emit(BRW_OPCODE_AND
, t1
,
1457 fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
)),
1459 inst
->force_writemask_all
= true;
1460 inst
= emit(BRW_OPCODE_SHR
, t1
, t1
, fs_reg(5));
1461 inst
->force_writemask_all
= true;
1462 /* This works for both SIMD8 and SIMD16 */
1463 inst
= emit(MOV(t2
, brw_imm_v(key
->persample_2x
? 0x1010 : 0x3210)));
1464 inst
->force_writemask_all
= true;
1465 /* This special instruction takes care of setting vstride=1,
1466 * width=4, hstride=0 of t2 during an ADD instruction.
1468 emit(FS_OPCODE_SET_SAMPLE_ID
, *reg
, t1
, t2
);
1470 /* As per GL_ARB_sample_shading specification:
1471 * "When rendering to a non-multisample buffer, or if multisample
1472 * rasterization is disabled, gl_SampleID will always be zero."
1474 emit(BRW_OPCODE_MOV
, *reg
, fs_reg(0));
1481 fs_visitor::resolve_source_modifiers(fs_reg
*src
)
1483 if (!src
->abs
&& !src
->negate
)
1486 fs_reg temp
= retype(vgrf(1), src
->type
);
1487 emit(MOV(temp
, *src
));
1492 fs_visitor::fix_math_operand(fs_reg src
)
1494 /* Can't do hstride == 0 args on gen6 math, so expand it out. We
1495 * might be able to do better by doing execsize = 1 math and then
1496 * expanding that result out, but we would need to be careful with
1499 * The hardware ignores source modifiers (negate and abs) on math
1500 * instructions, so we also move to a temp to set those up.
1502 if (devinfo
->gen
== 6 && src
.file
!= UNIFORM
&& src
.file
!= IMM
&&
1503 !src
.abs
&& !src
.negate
)
1506 /* Gen7 relaxes most of the above restrictions, but still can't use IMM
1509 if (devinfo
->gen
>= 7 && src
.file
!= IMM
)
1512 fs_reg expanded
= vgrf(glsl_type::float_type
);
1513 expanded
.type
= src
.type
;
1514 emit(BRW_OPCODE_MOV
, expanded
, src
);
1519 fs_visitor::emit_math(enum opcode opcode
, fs_reg dst
, fs_reg src
)
1522 case SHADER_OPCODE_RCP
:
1523 case SHADER_OPCODE_RSQ
:
1524 case SHADER_OPCODE_SQRT
:
1525 case SHADER_OPCODE_EXP2
:
1526 case SHADER_OPCODE_LOG2
:
1527 case SHADER_OPCODE_SIN
:
1528 case SHADER_OPCODE_COS
:
1531 unreachable("not reached: bad math opcode");
1534 /* Can't do hstride == 0 args to gen6 math, so expand it out. We
1535 * might be able to do better by doing execsize = 1 math and then
1536 * expanding that result out, but we would need to be careful with
1539 * Gen 6 hardware ignores source modifiers (negate and abs) on math
1540 * instructions, so we also move to a temp to set those up.
1542 if (devinfo
->gen
== 6 || devinfo
->gen
== 7)
1543 src
= fix_math_operand(src
);
1545 fs_inst
*inst
= emit(opcode
, dst
, src
);
1547 if (devinfo
->gen
< 6) {
1549 inst
->mlen
= dispatch_width
/ 8;
1556 fs_visitor::emit_math(enum opcode opcode
, fs_reg dst
, fs_reg src0
, fs_reg src1
)
1561 if (devinfo
->gen
>= 8) {
1562 inst
= emit(opcode
, dst
, src0
, src1
);
1563 } else if (devinfo
->gen
>= 6) {
1564 src0
= fix_math_operand(src0
);
1565 src1
= fix_math_operand(src1
);
1567 inst
= emit(opcode
, dst
, src0
, src1
);
1569 /* From the Ironlake PRM, Volume 4, Part 1, Section 6.1.13
1570 * "Message Payload":
1572 * "Operand0[7]. For the INT DIV functions, this operand is the
1575 * "Operand1[7]. For the INT DIV functions, this operand is the
1578 bool is_int_div
= opcode
!= SHADER_OPCODE_POW
;
1579 fs_reg
&op0
= is_int_div
? src1
: src0
;
1580 fs_reg
&op1
= is_int_div
? src0
: src1
;
1582 emit(MOV(fs_reg(MRF
, base_mrf
+ 1, op1
.type
, dispatch_width
), op1
));
1583 inst
= emit(opcode
, dst
, op0
, reg_null_f
);
1585 inst
->base_mrf
= base_mrf
;
1586 inst
->mlen
= 2 * dispatch_width
/ 8;
1592 fs_visitor::emit_discard_jump()
1594 assert(((brw_wm_prog_data
*) this->prog_data
)->uses_kill
);
1596 /* For performance, after a discard, jump to the end of the
1597 * shader if all relevant channels have been discarded.
1599 fs_inst
*discard_jump
= emit(FS_OPCODE_DISCARD_JUMP
);
1600 discard_jump
->flag_subreg
= 1;
1602 discard_jump
->predicate
= (dispatch_width
== 8)
1603 ? BRW_PREDICATE_ALIGN1_ANY8H
1604 : BRW_PREDICATE_ALIGN1_ANY16H
;
1605 discard_jump
->predicate_inverse
= true;
1609 fs_visitor::assign_curb_setup()
1611 if (dispatch_width
== 8) {
1612 prog_data
->dispatch_grf_start_reg
= payload
.num_regs
;
1614 if (stage
== MESA_SHADER_FRAGMENT
) {
1615 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
1616 prog_data
->dispatch_grf_start_reg_16
= payload
.num_regs
;
1617 } else if (stage
== MESA_SHADER_COMPUTE
) {
1618 brw_cs_prog_data
*prog_data
= (brw_cs_prog_data
*) this->prog_data
;
1619 prog_data
->dispatch_grf_start_reg_16
= payload
.num_regs
;
1621 unreachable("Unsupported shader type!");
1625 prog_data
->curb_read_length
= ALIGN(stage_prog_data
->nr_params
, 8) / 8;
1627 /* Map the offsets in the UNIFORM file to fixed HW regs. */
1628 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1629 for (unsigned int i
= 0; i
< inst
->sources
; i
++) {
1630 if (inst
->src
[i
].file
== UNIFORM
) {
1631 int uniform_nr
= inst
->src
[i
].reg
+ inst
->src
[i
].reg_offset
;
1633 if (uniform_nr
>= 0 && uniform_nr
< (int) uniforms
) {
1634 constant_nr
= push_constant_loc
[uniform_nr
];
1636 /* Section 5.11 of the OpenGL 4.1 spec says:
1637 * "Out-of-bounds reads return undefined values, which include
1638 * values from other variables of the active program or zero."
1639 * Just return the first push constant.
1644 struct brw_reg brw_reg
= brw_vec1_grf(payload
.num_regs
+
1648 inst
->src
[i
].file
= HW_REG
;
1649 inst
->src
[i
].fixed_hw_reg
= byte_offset(
1650 retype(brw_reg
, inst
->src
[i
].type
),
1651 inst
->src
[i
].subreg_offset
);
1658 fs_visitor::calculate_urb_setup()
1660 assert(stage
== MESA_SHADER_FRAGMENT
);
1661 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
1662 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1664 memset(prog_data
->urb_setup
, -1,
1665 sizeof(prog_data
->urb_setup
[0]) * VARYING_SLOT_MAX
);
1668 /* Figure out where each of the incoming setup attributes lands. */
1669 if (devinfo
->gen
>= 6) {
1670 if (_mesa_bitcount_64(prog
->InputsRead
&
1671 BRW_FS_VARYING_INPUT_MASK
) <= 16) {
1672 /* The SF/SBE pipeline stage can do arbitrary rearrangement of the
1673 * first 16 varying inputs, so we can put them wherever we want.
1674 * Just put them in order.
1676 * This is useful because it means that (a) inputs not used by the
1677 * fragment shader won't take up valuable register space, and (b) we
1678 * won't have to recompile the fragment shader if it gets paired with
1679 * a different vertex (or geometry) shader.
1681 for (unsigned int i
= 0; i
< VARYING_SLOT_MAX
; i
++) {
1682 if (prog
->InputsRead
& BRW_FS_VARYING_INPUT_MASK
&
1683 BITFIELD64_BIT(i
)) {
1684 prog_data
->urb_setup
[i
] = urb_next
++;
1688 /* We have enough input varyings that the SF/SBE pipeline stage can't
1689 * arbitrarily rearrange them to suit our whim; we have to put them
1690 * in an order that matches the output of the previous pipeline stage
1691 * (geometry or vertex shader).
1693 struct brw_vue_map prev_stage_vue_map
;
1694 brw_compute_vue_map(devinfo
, &prev_stage_vue_map
,
1695 key
->input_slots_valid
);
1696 int first_slot
= 2 * BRW_SF_URB_ENTRY_READ_OFFSET
;
1697 assert(prev_stage_vue_map
.num_slots
<= first_slot
+ 32);
1698 for (int slot
= first_slot
; slot
< prev_stage_vue_map
.num_slots
;
1700 int varying
= prev_stage_vue_map
.slot_to_varying
[slot
];
1701 /* Note that varying == BRW_VARYING_SLOT_COUNT when a slot is
1704 if (varying
!= BRW_VARYING_SLOT_COUNT
&&
1705 (prog
->InputsRead
& BRW_FS_VARYING_INPUT_MASK
&
1706 BITFIELD64_BIT(varying
))) {
1707 prog_data
->urb_setup
[varying
] = slot
- first_slot
;
1710 urb_next
= prev_stage_vue_map
.num_slots
- first_slot
;
1713 /* FINISHME: The sf doesn't map VS->FS inputs for us very well. */
1714 for (unsigned int i
= 0; i
< VARYING_SLOT_MAX
; i
++) {
1715 /* Point size is packed into the header, not as a general attribute */
1716 if (i
== VARYING_SLOT_PSIZ
)
1719 if (key
->input_slots_valid
& BITFIELD64_BIT(i
)) {
1720 /* The back color slot is skipped when the front color is
1721 * also written to. In addition, some slots can be
1722 * written in the vertex shader and not read in the
1723 * fragment shader. So the register number must always be
1724 * incremented, mapped or not.
1726 if (_mesa_varying_slot_in_fs((gl_varying_slot
) i
))
1727 prog_data
->urb_setup
[i
] = urb_next
;
1733 * It's a FS only attribute, and we did interpolation for this attribute
1734 * in SF thread. So, count it here, too.
1736 * See compile_sf_prog() for more info.
1738 if (prog
->InputsRead
& BITFIELD64_BIT(VARYING_SLOT_PNTC
))
1739 prog_data
->urb_setup
[VARYING_SLOT_PNTC
] = urb_next
++;
1742 prog_data
->num_varying_inputs
= urb_next
;
1746 fs_visitor::assign_urb_setup()
1748 assert(stage
== MESA_SHADER_FRAGMENT
);
1749 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
1751 int urb_start
= payload
.num_regs
+ prog_data
->base
.curb_read_length
;
1753 /* Offset all the urb_setup[] index by the actual position of the
1754 * setup regs, now that the location of the constants has been chosen.
1756 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1757 if (inst
->opcode
== FS_OPCODE_LINTERP
) {
1758 assert(inst
->src
[1].file
== HW_REG
);
1759 inst
->src
[1].fixed_hw_reg
.nr
+= urb_start
;
1762 if (inst
->opcode
== FS_OPCODE_CINTERP
) {
1763 assert(inst
->src
[0].file
== HW_REG
);
1764 inst
->src
[0].fixed_hw_reg
.nr
+= urb_start
;
1768 /* Each attribute is 4 setup channels, each of which is half a reg. */
1769 this->first_non_payload_grf
=
1770 urb_start
+ prog_data
->num_varying_inputs
* 2;
1774 fs_visitor::assign_vs_urb_setup()
1776 brw_vs_prog_data
*vs_prog_data
= (brw_vs_prog_data
*) prog_data
;
1777 int grf
, count
, slot
, channel
, attr
;
1779 assert(stage
== MESA_SHADER_VERTEX
);
1780 count
= _mesa_bitcount_64(vs_prog_data
->inputs_read
);
1781 if (vs_prog_data
->uses_vertexid
|| vs_prog_data
->uses_instanceid
)
1784 /* Each attribute is 4 regs. */
1785 this->first_non_payload_grf
=
1786 payload
.num_regs
+ prog_data
->curb_read_length
+ count
* 4;
1788 unsigned vue_entries
=
1789 MAX2(count
, vs_prog_data
->base
.vue_map
.num_slots
);
1791 vs_prog_data
->base
.urb_entry_size
= ALIGN(vue_entries
, 4) / 4;
1792 vs_prog_data
->base
.urb_read_length
= (count
+ 1) / 2;
1794 assert(vs_prog_data
->base
.urb_read_length
<= 15);
1796 /* Rewrite all ATTR file references to the hw grf that they land in. */
1797 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1798 for (int i
= 0; i
< inst
->sources
; i
++) {
1799 if (inst
->src
[i
].file
== ATTR
) {
1801 if (inst
->src
[i
].reg
== VERT_ATTRIB_MAX
) {
1804 /* Attributes come in in a contiguous block, ordered by their
1805 * gl_vert_attrib value. That means we can compute the slot
1806 * number for an attribute by masking out the enabled
1807 * attributes before it and counting the bits.
1809 attr
= inst
->src
[i
].reg
+ inst
->src
[i
].reg_offset
/ 4;
1810 slot
= _mesa_bitcount_64(vs_prog_data
->inputs_read
&
1811 BITFIELD64_MASK(attr
));
1814 channel
= inst
->src
[i
].reg_offset
& 3;
1816 grf
= payload
.num_regs
+
1817 prog_data
->curb_read_length
+
1820 inst
->src
[i
].file
= HW_REG
;
1821 inst
->src
[i
].fixed_hw_reg
=
1822 retype(brw_vec8_grf(grf
, 0), inst
->src
[i
].type
);
1829 * Split large virtual GRFs into separate components if we can.
1831 * This is mostly duplicated with what brw_fs_vector_splitting does,
1832 * but that's really conservative because it's afraid of doing
1833 * splitting that doesn't result in real progress after the rest of
1834 * the optimization phases, which would cause infinite looping in
1835 * optimization. We can do it once here, safely. This also has the
1836 * opportunity to split interpolated values, or maybe even uniforms,
1837 * which we don't have at the IR level.
1839 * We want to split, because virtual GRFs are what we register
1840 * allocate and spill (due to contiguousness requirements for some
1841 * instructions), and they're what we naturally generate in the
1842 * codegen process, but most virtual GRFs don't actually need to be
1843 * contiguous sets of GRFs. If we split, we'll end up with reduced
1844 * live intervals and better dead code elimination and coalescing.
1847 fs_visitor::split_virtual_grfs()
1849 int num_vars
= this->alloc
.count
;
1851 /* Count the total number of registers */
1853 int vgrf_to_reg
[num_vars
];
1854 for (int i
= 0; i
< num_vars
; i
++) {
1855 vgrf_to_reg
[i
] = reg_count
;
1856 reg_count
+= alloc
.sizes
[i
];
1859 /* An array of "split points". For each register slot, this indicates
1860 * if this slot can be separated from the previous slot. Every time an
1861 * instruction uses multiple elements of a register (as a source or
1862 * destination), we mark the used slots as inseparable. Then we go
1863 * through and split the registers into the smallest pieces we can.
1865 bool split_points
[reg_count
];
1866 memset(split_points
, 0, sizeof(split_points
));
1868 /* Mark all used registers as fully splittable */
1869 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1870 if (inst
->dst
.file
== GRF
) {
1871 int reg
= vgrf_to_reg
[inst
->dst
.reg
];
1872 for (unsigned j
= 1; j
< this->alloc
.sizes
[inst
->dst
.reg
]; j
++)
1873 split_points
[reg
+ j
] = true;
1876 for (int i
= 0; i
< inst
->sources
; i
++) {
1877 if (inst
->src
[i
].file
== GRF
) {
1878 int reg
= vgrf_to_reg
[inst
->src
[i
].reg
];
1879 for (unsigned j
= 1; j
< this->alloc
.sizes
[inst
->src
[i
].reg
]; j
++)
1880 split_points
[reg
+ j
] = true;
1885 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1886 if (inst
->dst
.file
== GRF
) {
1887 int reg
= vgrf_to_reg
[inst
->dst
.reg
] + inst
->dst
.reg_offset
;
1888 for (int j
= 1; j
< inst
->regs_written
; j
++)
1889 split_points
[reg
+ j
] = false;
1891 for (int i
= 0; i
< inst
->sources
; i
++) {
1892 if (inst
->src
[i
].file
== GRF
) {
1893 int reg
= vgrf_to_reg
[inst
->src
[i
].reg
] + inst
->src
[i
].reg_offset
;
1894 for (int j
= 1; j
< inst
->regs_read(i
); j
++)
1895 split_points
[reg
+ j
] = false;
1900 int new_virtual_grf
[reg_count
];
1901 int new_reg_offset
[reg_count
];
1904 for (int i
= 0; i
< num_vars
; i
++) {
1905 /* The first one should always be 0 as a quick sanity check. */
1906 assert(split_points
[reg
] == false);
1909 new_reg_offset
[reg
] = 0;
1914 for (unsigned j
= 1; j
< alloc
.sizes
[i
]; j
++) {
1915 /* If this is a split point, reset the offset to 0 and allocate a
1916 * new virtual GRF for the previous offset many registers
1918 if (split_points
[reg
]) {
1919 assert(offset
<= MAX_VGRF_SIZE
);
1920 int grf
= alloc
.allocate(offset
);
1921 for (int k
= reg
- offset
; k
< reg
; k
++)
1922 new_virtual_grf
[k
] = grf
;
1925 new_reg_offset
[reg
] = offset
;
1930 /* The last one gets the original register number */
1931 assert(offset
<= MAX_VGRF_SIZE
);
1932 alloc
.sizes
[i
] = offset
;
1933 for (int k
= reg
- offset
; k
< reg
; k
++)
1934 new_virtual_grf
[k
] = i
;
1936 assert(reg
== reg_count
);
1938 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1939 if (inst
->dst
.file
== GRF
) {
1940 reg
= vgrf_to_reg
[inst
->dst
.reg
] + inst
->dst
.reg_offset
;
1941 inst
->dst
.reg
= new_virtual_grf
[reg
];
1942 inst
->dst
.reg_offset
= new_reg_offset
[reg
];
1943 assert((unsigned)new_reg_offset
[reg
] < alloc
.sizes
[new_virtual_grf
[reg
]]);
1945 for (int i
= 0; i
< inst
->sources
; i
++) {
1946 if (inst
->src
[i
].file
== GRF
) {
1947 reg
= vgrf_to_reg
[inst
->src
[i
].reg
] + inst
->src
[i
].reg_offset
;
1948 inst
->src
[i
].reg
= new_virtual_grf
[reg
];
1949 inst
->src
[i
].reg_offset
= new_reg_offset
[reg
];
1950 assert((unsigned)new_reg_offset
[reg
] < alloc
.sizes
[new_virtual_grf
[reg
]]);
1954 invalidate_live_intervals();
1958 * Remove unused virtual GRFs and compact the virtual_grf_* arrays.
1960 * During code generation, we create tons of temporary variables, many of
1961 * which get immediately killed and are never used again. Yet, in later
1962 * optimization and analysis passes, such as compute_live_intervals, we need
1963 * to loop over all the virtual GRFs. Compacting them can save a lot of
1967 fs_visitor::compact_virtual_grfs()
1969 bool progress
= false;
1970 int remap_table
[this->alloc
.count
];
1971 memset(remap_table
, -1, sizeof(remap_table
));
1973 /* Mark which virtual GRFs are used. */
1974 foreach_block_and_inst(block
, const fs_inst
, inst
, cfg
) {
1975 if (inst
->dst
.file
== GRF
)
1976 remap_table
[inst
->dst
.reg
] = 0;
1978 for (int i
= 0; i
< inst
->sources
; i
++) {
1979 if (inst
->src
[i
].file
== GRF
)
1980 remap_table
[inst
->src
[i
].reg
] = 0;
1984 /* Compact the GRF arrays. */
1986 for (unsigned i
= 0; i
< this->alloc
.count
; i
++) {
1987 if (remap_table
[i
] == -1) {
1988 /* We just found an unused register. This means that we are
1989 * actually going to compact something.
1993 remap_table
[i
] = new_index
;
1994 alloc
.sizes
[new_index
] = alloc
.sizes
[i
];
1995 invalidate_live_intervals();
2000 this->alloc
.count
= new_index
;
2002 /* Patch all the instructions to use the newly renumbered registers */
2003 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2004 if (inst
->dst
.file
== GRF
)
2005 inst
->dst
.reg
= remap_table
[inst
->dst
.reg
];
2007 for (int i
= 0; i
< inst
->sources
; i
++) {
2008 if (inst
->src
[i
].file
== GRF
)
2009 inst
->src
[i
].reg
= remap_table
[inst
->src
[i
].reg
];
2013 /* Patch all the references to delta_xy, since they're used in register
2014 * allocation. If they're unused, switch them to BAD_FILE so we don't
2015 * think some random VGRF is delta_xy.
2017 for (unsigned i
= 0; i
< ARRAY_SIZE(delta_xy
); i
++) {
2018 if (delta_xy
[i
].file
== GRF
) {
2019 if (remap_table
[delta_xy
[i
].reg
] != -1) {
2020 delta_xy
[i
].reg
= remap_table
[delta_xy
[i
].reg
];
2022 delta_xy
[i
].file
= BAD_FILE
;
2031 * Implements array access of uniforms by inserting a
2032 * PULL_CONSTANT_LOAD instruction.
2034 * Unlike temporary GRF array access (where we don't support it due to
2035 * the difficulty of doing relative addressing on instruction
2036 * destinations), we could potentially do array access of uniforms
2037 * that were loaded in GRF space as push constants. In real-world
2038 * usage we've seen, though, the arrays being used are always larger
2039 * than we could load as push constants, so just always move all
2040 * uniform array access out to a pull constant buffer.
2043 fs_visitor::move_uniform_array_access_to_pull_constants()
2045 if (dispatch_width
!= 8)
2048 pull_constant_loc
= ralloc_array(mem_ctx
, int, uniforms
);
2049 memset(pull_constant_loc
, -1, sizeof(pull_constant_loc
[0]) * uniforms
);
2051 /* Walk through and find array access of uniforms. Put a copy of that
2052 * uniform in the pull constant buffer.
2054 * Note that we don't move constant-indexed accesses to arrays. No
2055 * testing has been done of the performance impact of this choice.
2057 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
2058 for (int i
= 0 ; i
< inst
->sources
; i
++) {
2059 if (inst
->src
[i
].file
!= UNIFORM
|| !inst
->src
[i
].reladdr
)
2062 int uniform
= inst
->src
[i
].reg
;
2064 /* If this array isn't already present in the pull constant buffer,
2067 if (pull_constant_loc
[uniform
] == -1) {
2068 const gl_constant_value
**values
= &stage_prog_data
->param
[uniform
];
2070 assert(param_size
[uniform
]);
2072 for (int j
= 0; j
< param_size
[uniform
]; j
++) {
2073 pull_constant_loc
[uniform
+ j
] = stage_prog_data
->nr_pull_params
;
2075 stage_prog_data
->pull_param
[stage_prog_data
->nr_pull_params
++] =
2084 * Assign UNIFORM file registers to either push constants or pull constants.
2086 * We allow a fragment shader to have more than the specified minimum
2087 * maximum number of fragment shader uniform components (64). If
2088 * there are too many of these, they'd fill up all of register space.
2089 * So, this will push some of them out to the pull constant buffer and
2090 * update the program to load them.
2093 fs_visitor::assign_constant_locations()
2095 /* Only the first compile (SIMD8 mode) gets to decide on locations. */
2096 if (dispatch_width
!= 8)
2099 /* Find which UNIFORM registers are still in use. */
2100 bool is_live
[uniforms
];
2101 for (unsigned int i
= 0; i
< uniforms
; i
++) {
2105 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2106 for (int i
= 0; i
< inst
->sources
; i
++) {
2107 if (inst
->src
[i
].file
!= UNIFORM
)
2110 int constant_nr
= inst
->src
[i
].reg
+ inst
->src
[i
].reg_offset
;
2111 if (constant_nr
>= 0 && constant_nr
< (int) uniforms
)
2112 is_live
[constant_nr
] = true;
2116 /* Only allow 16 registers (128 uniform components) as push constants.
2118 * Just demote the end of the list. We could probably do better
2119 * here, demoting things that are rarely used in the program first.
2121 * If changing this value, note the limitation about total_regs in
2124 unsigned int max_push_components
= 16 * 8;
2125 unsigned int num_push_constants
= 0;
2127 push_constant_loc
= ralloc_array(mem_ctx
, int, uniforms
);
2129 for (unsigned int i
= 0; i
< uniforms
; i
++) {
2130 if (!is_live
[i
] || pull_constant_loc
[i
] != -1) {
2131 /* This UNIFORM register is either dead, or has already been demoted
2132 * to a pull const. Mark it as no longer living in the param[] array.
2134 push_constant_loc
[i
] = -1;
2138 if (num_push_constants
< max_push_components
) {
2139 /* Retain as a push constant. Record the location in the params[]
2142 push_constant_loc
[i
] = num_push_constants
++;
2144 /* Demote to a pull constant. */
2145 push_constant_loc
[i
] = -1;
2147 int pull_index
= stage_prog_data
->nr_pull_params
++;
2148 stage_prog_data
->pull_param
[pull_index
] = stage_prog_data
->param
[i
];
2149 pull_constant_loc
[i
] = pull_index
;
2153 stage_prog_data
->nr_params
= num_push_constants
;
2155 /* Up until now, the param[] array has been indexed by reg + reg_offset
2156 * of UNIFORM registers. Condense it to only contain the uniforms we
2157 * chose to upload as push constants.
2159 for (unsigned int i
= 0; i
< uniforms
; i
++) {
2160 int remapped
= push_constant_loc
[i
];
2165 assert(remapped
<= (int)i
);
2166 stage_prog_data
->param
[remapped
] = stage_prog_data
->param
[i
];
2171 * Replace UNIFORM register file access with either UNIFORM_PULL_CONSTANT_LOAD
2172 * or VARYING_PULL_CONSTANT_LOAD instructions which load values into VGRFs.
2175 fs_visitor::demote_pull_constants()
2177 foreach_block_and_inst (block
, fs_inst
, inst
, cfg
) {
2178 for (int i
= 0; i
< inst
->sources
; i
++) {
2179 if (inst
->src
[i
].file
!= UNIFORM
)
2183 unsigned location
= inst
->src
[i
].reg
+ inst
->src
[i
].reg_offset
;
2184 if (location
>= uniforms
) /* Out of bounds access */
2187 pull_index
= pull_constant_loc
[location
];
2189 if (pull_index
== -1)
2192 /* Set up the annotation tracking for new generated instructions. */
2193 const fs_builder ibld
= bld
.annotate(inst
->annotation
, inst
->ir
)
2195 fs_reg
surf_index(stage_prog_data
->binding_table
.pull_constants_start
);
2196 fs_reg dst
= vgrf(glsl_type::float_type
);
2198 /* Generate a pull load into dst. */
2199 if (inst
->src
[i
].reladdr
) {
2200 VARYING_PULL_CONSTANT_LOAD(ibld
, dst
,
2202 *inst
->src
[i
].reladdr
,
2204 inst
->src
[i
].reladdr
= NULL
;
2206 fs_reg offset
= fs_reg((unsigned)(pull_index
* 4) & ~15);
2207 ibld
.emit(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
,
2208 dst
, surf_index
, offset
);
2209 inst
->src
[i
].set_smear(pull_index
& 3);
2212 /* Rewrite the instruction to use the temporary VGRF. */
2213 inst
->src
[i
].file
= GRF
;
2214 inst
->src
[i
].reg
= dst
.reg
;
2215 inst
->src
[i
].reg_offset
= 0;
2216 inst
->src
[i
].width
= dispatch_width
;
2219 invalidate_live_intervals();
2223 fs_visitor::opt_algebraic()
2225 bool progress
= false;
2227 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2228 switch (inst
->opcode
) {
2229 case BRW_OPCODE_MOV
:
2230 if (inst
->src
[0].file
!= IMM
)
2233 if (inst
->saturate
) {
2234 if (inst
->dst
.type
!= inst
->src
[0].type
)
2235 assert(!"unimplemented: saturate mixed types");
2237 if (brw_saturate_immediate(inst
->dst
.type
,
2238 &inst
->src
[0].fixed_hw_reg
)) {
2239 inst
->saturate
= false;
2245 case BRW_OPCODE_MUL
:
2246 if (inst
->src
[1].file
!= IMM
)
2250 if (inst
->src
[1].is_one()) {
2251 inst
->opcode
= BRW_OPCODE_MOV
;
2252 inst
->src
[1] = reg_undef
;
2258 if (inst
->src
[1].is_negative_one()) {
2259 inst
->opcode
= BRW_OPCODE_MOV
;
2260 inst
->src
[0].negate
= !inst
->src
[0].negate
;
2261 inst
->src
[1] = reg_undef
;
2267 if (inst
->src
[1].is_zero()) {
2268 inst
->opcode
= BRW_OPCODE_MOV
;
2269 inst
->src
[0] = inst
->src
[1];
2270 inst
->src
[1] = reg_undef
;
2275 if (inst
->src
[0].file
== IMM
) {
2276 assert(inst
->src
[0].type
== BRW_REGISTER_TYPE_F
);
2277 inst
->opcode
= BRW_OPCODE_MOV
;
2278 inst
->src
[0].fixed_hw_reg
.dw1
.f
*= inst
->src
[1].fixed_hw_reg
.dw1
.f
;
2279 inst
->src
[1] = reg_undef
;
2284 case BRW_OPCODE_ADD
:
2285 if (inst
->src
[1].file
!= IMM
)
2289 if (inst
->src
[1].is_zero()) {
2290 inst
->opcode
= BRW_OPCODE_MOV
;
2291 inst
->src
[1] = reg_undef
;
2296 if (inst
->src
[0].file
== IMM
) {
2297 assert(inst
->src
[0].type
== BRW_REGISTER_TYPE_F
);
2298 inst
->opcode
= BRW_OPCODE_MOV
;
2299 inst
->src
[0].fixed_hw_reg
.dw1
.f
+= inst
->src
[1].fixed_hw_reg
.dw1
.f
;
2300 inst
->src
[1] = reg_undef
;
2306 if (inst
->src
[0].equals(inst
->src
[1])) {
2307 inst
->opcode
= BRW_OPCODE_MOV
;
2308 inst
->src
[1] = reg_undef
;
2313 case BRW_OPCODE_LRP
:
2314 if (inst
->src
[1].equals(inst
->src
[2])) {
2315 inst
->opcode
= BRW_OPCODE_MOV
;
2316 inst
->src
[0] = inst
->src
[1];
2317 inst
->src
[1] = reg_undef
;
2318 inst
->src
[2] = reg_undef
;
2323 case BRW_OPCODE_CMP
:
2324 if (inst
->conditional_mod
== BRW_CONDITIONAL_GE
&&
2326 inst
->src
[0].negate
&&
2327 inst
->src
[1].is_zero()) {
2328 inst
->src
[0].abs
= false;
2329 inst
->src
[0].negate
= false;
2330 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
2335 case BRW_OPCODE_SEL
:
2336 if (inst
->src
[0].equals(inst
->src
[1])) {
2337 inst
->opcode
= BRW_OPCODE_MOV
;
2338 inst
->src
[1] = reg_undef
;
2339 inst
->predicate
= BRW_PREDICATE_NONE
;
2340 inst
->predicate_inverse
= false;
2342 } else if (inst
->saturate
&& inst
->src
[1].file
== IMM
) {
2343 switch (inst
->conditional_mod
) {
2344 case BRW_CONDITIONAL_LE
:
2345 case BRW_CONDITIONAL_L
:
2346 switch (inst
->src
[1].type
) {
2347 case BRW_REGISTER_TYPE_F
:
2348 if (inst
->src
[1].fixed_hw_reg
.dw1
.f
>= 1.0f
) {
2349 inst
->opcode
= BRW_OPCODE_MOV
;
2350 inst
->src
[1] = reg_undef
;
2351 inst
->conditional_mod
= BRW_CONDITIONAL_NONE
;
2359 case BRW_CONDITIONAL_GE
:
2360 case BRW_CONDITIONAL_G
:
2361 switch (inst
->src
[1].type
) {
2362 case BRW_REGISTER_TYPE_F
:
2363 if (inst
->src
[1].fixed_hw_reg
.dw1
.f
<= 0.0f
) {
2364 inst
->opcode
= BRW_OPCODE_MOV
;
2365 inst
->src
[1] = reg_undef
;
2366 inst
->conditional_mod
= BRW_CONDITIONAL_NONE
;
2378 case BRW_OPCODE_MAD
:
2379 if (inst
->src
[1].is_zero() || inst
->src
[2].is_zero()) {
2380 inst
->opcode
= BRW_OPCODE_MOV
;
2381 inst
->src
[1] = reg_undef
;
2382 inst
->src
[2] = reg_undef
;
2384 } else if (inst
->src
[0].is_zero()) {
2385 inst
->opcode
= BRW_OPCODE_MUL
;
2386 inst
->src
[0] = inst
->src
[2];
2387 inst
->src
[2] = reg_undef
;
2389 } else if (inst
->src
[1].is_one()) {
2390 inst
->opcode
= BRW_OPCODE_ADD
;
2391 inst
->src
[1] = inst
->src
[2];
2392 inst
->src
[2] = reg_undef
;
2394 } else if (inst
->src
[2].is_one()) {
2395 inst
->opcode
= BRW_OPCODE_ADD
;
2396 inst
->src
[2] = reg_undef
;
2398 } else if (inst
->src
[1].file
== IMM
&& inst
->src
[2].file
== IMM
) {
2399 inst
->opcode
= BRW_OPCODE_ADD
;
2400 inst
->src
[1].fixed_hw_reg
.dw1
.f
*= inst
->src
[2].fixed_hw_reg
.dw1
.f
;
2401 inst
->src
[2] = reg_undef
;
2405 case SHADER_OPCODE_RCP
: {
2406 fs_inst
*prev
= (fs_inst
*)inst
->prev
;
2407 if (prev
->opcode
== SHADER_OPCODE_SQRT
) {
2408 if (inst
->src
[0].equals(prev
->dst
)) {
2409 inst
->opcode
= SHADER_OPCODE_RSQ
;
2410 inst
->src
[0] = prev
->src
[0];
2416 case SHADER_OPCODE_BROADCAST
:
2417 if (is_uniform(inst
->src
[0])) {
2418 inst
->opcode
= BRW_OPCODE_MOV
;
2420 inst
->force_writemask_all
= true;
2422 } else if (inst
->src
[1].file
== IMM
) {
2423 inst
->opcode
= BRW_OPCODE_MOV
;
2424 inst
->src
[0] = component(inst
->src
[0],
2425 inst
->src
[1].fixed_hw_reg
.dw1
.ud
);
2427 inst
->force_writemask_all
= true;
2436 /* Swap if src[0] is immediate. */
2437 if (progress
&& inst
->is_commutative()) {
2438 if (inst
->src
[0].file
== IMM
) {
2439 fs_reg tmp
= inst
->src
[1];
2440 inst
->src
[1] = inst
->src
[0];
2449 * Optimize sample messages that have constant zero values for the trailing
2450 * texture coordinates. We can just reduce the message length for these
2451 * instructions instead of reserving a register for it. Trailing parameters
2452 * that aren't sent default to zero anyway. This will cause the dead code
2453 * eliminator to remove the MOV instruction that would otherwise be emitted to
2454 * set up the zero value.
2457 fs_visitor::opt_zero_samples()
2459 /* Gen4 infers the texturing opcode based on the message length so we can't
2462 if (devinfo
->gen
< 5)
2465 bool progress
= false;
2467 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2468 if (!inst
->is_tex())
2471 fs_inst
*load_payload
= (fs_inst
*) inst
->prev
;
2473 if (load_payload
->is_head_sentinel() ||
2474 load_payload
->opcode
!= SHADER_OPCODE_LOAD_PAYLOAD
)
2477 /* We don't want to remove the message header or the first parameter.
2478 * Removing the first parameter is not allowed, see the Haswell PRM
2479 * volume 7, page 149:
2481 * "Parameter 0 is required except for the sampleinfo message, which
2482 * has no parameter 0"
2484 while (inst
->mlen
> inst
->header_size
+ dispatch_width
/ 8 &&
2485 load_payload
->src
[(inst
->mlen
- inst
->header_size
) /
2486 (dispatch_width
/ 8) +
2487 inst
->header_size
- 1].is_zero()) {
2488 inst
->mlen
-= dispatch_width
/ 8;
2494 invalidate_live_intervals();
2500 * Optimize sample messages which are followed by the final RT write.
2502 * CHV, and GEN9+ can mark a texturing SEND instruction with EOT to have its
2503 * results sent directly to the framebuffer, bypassing the EU. Recognize the
2504 * final texturing results copied to the framebuffer write payload and modify
2505 * them to write to the framebuffer directly.
2508 fs_visitor::opt_sampler_eot()
2510 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
2512 if (stage
!= MESA_SHADER_FRAGMENT
)
2515 if (devinfo
->gen
< 9 && !devinfo
->is_cherryview
)
2518 /* FINISHME: It should be possible to implement this optimization when there
2519 * are multiple drawbuffers.
2521 if (key
->nr_color_regions
!= 1)
2524 /* Look for a texturing instruction immediately before the final FB_WRITE. */
2525 fs_inst
*fb_write
= (fs_inst
*) cfg
->blocks
[cfg
->num_blocks
- 1]->end();
2526 assert(fb_write
->eot
);
2527 assert(fb_write
->opcode
== FS_OPCODE_FB_WRITE
);
2529 fs_inst
*tex_inst
= (fs_inst
*) fb_write
->prev
;
2531 /* There wasn't one; nothing to do. */
2532 if (unlikely(tex_inst
->is_head_sentinel()) || !tex_inst
->is_tex())
2535 /* This optimisation doesn't seem to work for textureGather for some
2536 * reason. I can't find any documentation or known workarounds to indicate
2537 * that this is expected, but considering that it is probably pretty
2538 * unlikely that a shader would directly write out the results from
2539 * textureGather we might as well just disable it.
2541 if (tex_inst
->opcode
== SHADER_OPCODE_TG4
||
2542 tex_inst
->opcode
== SHADER_OPCODE_TG4_OFFSET
)
2545 /* If there's no header present, we need to munge the LOAD_PAYLOAD as well.
2546 * It's very likely to be the previous instruction.
2548 fs_inst
*load_payload
= (fs_inst
*) tex_inst
->prev
;
2549 if (load_payload
->is_head_sentinel() ||
2550 load_payload
->opcode
!= SHADER_OPCODE_LOAD_PAYLOAD
)
2553 assert(!tex_inst
->eot
); /* We can't get here twice */
2554 assert((tex_inst
->offset
& (0xff << 24)) == 0);
2556 tex_inst
->offset
|= fb_write
->target
<< 24;
2557 tex_inst
->eot
= true;
2558 tex_inst
->dst
= bld
.null_reg_ud();
2559 fb_write
->remove(cfg
->blocks
[cfg
->num_blocks
- 1]);
2561 /* If a header is present, marking the eot is sufficient. Otherwise, we need
2562 * to create a new LOAD_PAYLOAD command with the same sources and a space
2563 * saved for the header. Using a new destination register not only makes sure
2564 * we have enough space, but it will make sure the dead code eliminator kills
2565 * the instruction that this will replace.
2567 if (tex_inst
->header_size
!= 0)
2570 fs_reg send_header
= bld
.vgrf(BRW_REGISTER_TYPE_F
,
2571 load_payload
->sources
+ 1);
2572 fs_reg
*new_sources
=
2573 ralloc_array(mem_ctx
, fs_reg
, load_payload
->sources
+ 1);
2575 new_sources
[0] = fs_reg();
2576 for (int i
= 0; i
< load_payload
->sources
; i
++)
2577 new_sources
[i
+1] = load_payload
->src
[i
];
2579 /* The LOAD_PAYLOAD helper seems like the obvious choice here. However, it
2580 * requires a lot of information about the sources to appropriately figure
2581 * out the number of registers needed to be used. Given this stage in our
2582 * optimization, we may not have the appropriate GRFs required by
2583 * LOAD_PAYLOAD at this point (copy propagation). Therefore, we need to
2584 * manually emit the instruction.
2586 fs_inst
*new_load_payload
= new(mem_ctx
) fs_inst(SHADER_OPCODE_LOAD_PAYLOAD
,
2587 load_payload
->exec_size
,
2590 load_payload
->sources
+ 1);
2592 new_load_payload
->regs_written
= load_payload
->regs_written
+ 1;
2593 new_load_payload
->header_size
= 1;
2595 tex_inst
->header_size
= 1;
2596 tex_inst
->insert_before(cfg
->blocks
[cfg
->num_blocks
- 1], new_load_payload
);
2597 tex_inst
->src
[0] = send_header
;
2603 fs_visitor::opt_register_renaming()
2605 bool progress
= false;
2608 int remap
[alloc
.count
];
2609 memset(remap
, -1, sizeof(int) * alloc
.count
);
2611 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2612 if (inst
->opcode
== BRW_OPCODE_IF
|| inst
->opcode
== BRW_OPCODE_DO
) {
2614 } else if (inst
->opcode
== BRW_OPCODE_ENDIF
||
2615 inst
->opcode
== BRW_OPCODE_WHILE
) {
2619 /* Rewrite instruction sources. */
2620 for (int i
= 0; i
< inst
->sources
; i
++) {
2621 if (inst
->src
[i
].file
== GRF
&&
2622 remap
[inst
->src
[i
].reg
] != -1 &&
2623 remap
[inst
->src
[i
].reg
] != inst
->src
[i
].reg
) {
2624 inst
->src
[i
].reg
= remap
[inst
->src
[i
].reg
];
2629 const int dst
= inst
->dst
.reg
;
2632 inst
->dst
.file
== GRF
&&
2633 alloc
.sizes
[inst
->dst
.reg
] == inst
->dst
.width
/ 8 &&
2634 !inst
->is_partial_write()) {
2635 if (remap
[dst
] == -1) {
2638 remap
[dst
] = alloc
.allocate(inst
->dst
.width
/ 8);
2639 inst
->dst
.reg
= remap
[dst
];
2642 } else if (inst
->dst
.file
== GRF
&&
2644 remap
[dst
] != dst
) {
2645 inst
->dst
.reg
= remap
[dst
];
2651 invalidate_live_intervals();
2653 for (unsigned i
= 0; i
< ARRAY_SIZE(delta_xy
); i
++) {
2654 if (delta_xy
[i
].file
== GRF
&& remap
[delta_xy
[i
].reg
] != -1) {
2655 delta_xy
[i
].reg
= remap
[delta_xy
[i
].reg
];
2664 * Remove redundant or useless discard jumps.
2666 * For example, we can eliminate jumps in the following sequence:
2668 * discard-jump (redundant with the next jump)
2669 * discard-jump (useless; jumps to the next instruction)
2673 fs_visitor::opt_redundant_discard_jumps()
2675 bool progress
= false;
2677 bblock_t
*last_bblock
= cfg
->blocks
[cfg
->num_blocks
- 1];
2679 fs_inst
*placeholder_halt
= NULL
;
2680 foreach_inst_in_block_reverse(fs_inst
, inst
, last_bblock
) {
2681 if (inst
->opcode
== FS_OPCODE_PLACEHOLDER_HALT
) {
2682 placeholder_halt
= inst
;
2687 if (!placeholder_halt
)
2690 /* Delete any HALTs immediately before the placeholder halt. */
2691 for (fs_inst
*prev
= (fs_inst
*) placeholder_halt
->prev
;
2692 !prev
->is_head_sentinel() && prev
->opcode
== FS_OPCODE_DISCARD_JUMP
;
2693 prev
= (fs_inst
*) placeholder_halt
->prev
) {
2694 prev
->remove(last_bblock
);
2699 invalidate_live_intervals();
2705 fs_visitor::compute_to_mrf()
2707 bool progress
= false;
2710 /* No MRFs on Gen >= 7. */
2711 if (devinfo
->gen
>= 7)
2714 calculate_live_intervals();
2716 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
2720 if (inst
->opcode
!= BRW_OPCODE_MOV
||
2721 inst
->is_partial_write() ||
2722 inst
->dst
.file
!= MRF
|| inst
->src
[0].file
!= GRF
||
2723 inst
->dst
.type
!= inst
->src
[0].type
||
2724 inst
->src
[0].abs
|| inst
->src
[0].negate
||
2725 !inst
->src
[0].is_contiguous() ||
2726 inst
->src
[0].subreg_offset
)
2729 /* Work out which hardware MRF registers are written by this
2732 int mrf_low
= inst
->dst
.reg
& ~BRW_MRF_COMPR4
;
2734 if (inst
->dst
.reg
& BRW_MRF_COMPR4
) {
2735 mrf_high
= mrf_low
+ 4;
2736 } else if (inst
->exec_size
== 16) {
2737 mrf_high
= mrf_low
+ 1;
2742 /* Can't compute-to-MRF this GRF if someone else was going to
2745 if (this->virtual_grf_end
[inst
->src
[0].reg
] > ip
)
2748 /* Found a move of a GRF to a MRF. Let's see if we can go
2749 * rewrite the thing that made this GRF to write into the MRF.
2751 foreach_inst_in_block_reverse_starting_from(fs_inst
, scan_inst
, inst
, block
) {
2752 if (scan_inst
->dst
.file
== GRF
&&
2753 scan_inst
->dst
.reg
== inst
->src
[0].reg
) {
2754 /* Found the last thing to write our reg we want to turn
2755 * into a compute-to-MRF.
2758 /* If this one instruction didn't populate all the
2759 * channels, bail. We might be able to rewrite everything
2760 * that writes that reg, but it would require smarter
2761 * tracking to delay the rewriting until complete success.
2763 if (scan_inst
->is_partial_write())
2766 /* Things returning more than one register would need us to
2767 * understand coalescing out more than one MOV at a time.
2769 if (scan_inst
->regs_written
> scan_inst
->dst
.width
/ 8)
2772 /* SEND instructions can't have MRF as a destination. */
2773 if (scan_inst
->mlen
)
2776 if (devinfo
->gen
== 6) {
2777 /* gen6 math instructions must have the destination be
2778 * GRF, so no compute-to-MRF for them.
2780 if (scan_inst
->is_math()) {
2785 if (scan_inst
->dst
.reg_offset
== inst
->src
[0].reg_offset
) {
2786 /* Found the creator of our MRF's source value. */
2787 scan_inst
->dst
.file
= MRF
;
2788 scan_inst
->dst
.reg
= inst
->dst
.reg
;
2789 scan_inst
->saturate
|= inst
->saturate
;
2790 inst
->remove(block
);
2796 /* We don't handle control flow here. Most computation of
2797 * values that end up in MRFs are shortly before the MRF
2800 if (block
->start() == scan_inst
)
2803 /* You can't read from an MRF, so if someone else reads our
2804 * MRF's source GRF that we wanted to rewrite, that stops us.
2806 bool interfered
= false;
2807 for (int i
= 0; i
< scan_inst
->sources
; i
++) {
2808 if (scan_inst
->src
[i
].file
== GRF
&&
2809 scan_inst
->src
[i
].reg
== inst
->src
[0].reg
&&
2810 scan_inst
->src
[i
].reg_offset
== inst
->src
[0].reg_offset
) {
2817 if (scan_inst
->dst
.file
== MRF
) {
2818 /* If somebody else writes our MRF here, we can't
2819 * compute-to-MRF before that.
2821 int scan_mrf_low
= scan_inst
->dst
.reg
& ~BRW_MRF_COMPR4
;
2824 if (scan_inst
->dst
.reg
& BRW_MRF_COMPR4
) {
2825 scan_mrf_high
= scan_mrf_low
+ 4;
2826 } else if (scan_inst
->exec_size
== 16) {
2827 scan_mrf_high
= scan_mrf_low
+ 1;
2829 scan_mrf_high
= scan_mrf_low
;
2832 if (mrf_low
== scan_mrf_low
||
2833 mrf_low
== scan_mrf_high
||
2834 mrf_high
== scan_mrf_low
||
2835 mrf_high
== scan_mrf_high
) {
2840 if (scan_inst
->mlen
> 0 && scan_inst
->base_mrf
!= -1) {
2841 /* Found a SEND instruction, which means that there are
2842 * live values in MRFs from base_mrf to base_mrf +
2843 * scan_inst->mlen - 1. Don't go pushing our MRF write up
2846 if (mrf_low
>= scan_inst
->base_mrf
&&
2847 mrf_low
< scan_inst
->base_mrf
+ scan_inst
->mlen
) {
2850 if (mrf_high
>= scan_inst
->base_mrf
&&
2851 mrf_high
< scan_inst
->base_mrf
+ scan_inst
->mlen
) {
2859 invalidate_live_intervals();
2865 * Eliminate FIND_LIVE_CHANNEL instructions occurring outside any control
2866 * flow. We could probably do better here with some form of divergence
2870 fs_visitor::eliminate_find_live_channel()
2872 bool progress
= false;
2875 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
2876 switch (inst
->opcode
) {
2882 case BRW_OPCODE_ENDIF
:
2883 case BRW_OPCODE_WHILE
:
2887 case FS_OPCODE_DISCARD_JUMP
:
2888 /* This can potentially make control flow non-uniform until the end
2893 case SHADER_OPCODE_FIND_LIVE_CHANNEL
:
2895 inst
->opcode
= BRW_OPCODE_MOV
;
2896 inst
->src
[0] = fs_reg(0);
2898 inst
->force_writemask_all
= true;
2912 * Once we've generated code, try to convert normal FS_OPCODE_FB_WRITE
2913 * instructions to FS_OPCODE_REP_FB_WRITE.
2916 fs_visitor::emit_repclear_shader()
2918 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
2920 int color_mrf
= base_mrf
+ 2;
2922 fs_inst
*mov
= emit(MOV(vec4(brw_message_reg(color_mrf
)),
2923 fs_reg(UNIFORM
, 0, BRW_REGISTER_TYPE_F
)));
2924 mov
->force_writemask_all
= true;
2927 if (key
->nr_color_regions
== 1) {
2928 write
= emit(FS_OPCODE_REP_FB_WRITE
);
2929 write
->saturate
= key
->clamp_fragment_color
;
2930 write
->base_mrf
= color_mrf
;
2932 write
->header_size
= 0;
2935 assume(key
->nr_color_regions
> 0);
2936 for (int i
= 0; i
< key
->nr_color_regions
; ++i
) {
2937 write
= emit(FS_OPCODE_REP_FB_WRITE
);
2938 write
->saturate
= key
->clamp_fragment_color
;
2939 write
->base_mrf
= base_mrf
;
2941 write
->header_size
= 2;
2949 assign_constant_locations();
2950 assign_curb_setup();
2952 /* Now that we have the uniform assigned, go ahead and force it to a vec4. */
2953 assert(mov
->src
[0].file
== HW_REG
);
2954 mov
->src
[0] = brw_vec4_grf(mov
->src
[0].fixed_hw_reg
.nr
, 0);
2958 * Walks through basic blocks, looking for repeated MRF writes and
2959 * removing the later ones.
2962 fs_visitor::remove_duplicate_mrf_writes()
2964 fs_inst
*last_mrf_move
[16];
2965 bool progress
= false;
2967 /* Need to update the MRF tracking for compressed instructions. */
2968 if (dispatch_width
== 16)
2971 memset(last_mrf_move
, 0, sizeof(last_mrf_move
));
2973 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
2974 if (inst
->is_control_flow()) {
2975 memset(last_mrf_move
, 0, sizeof(last_mrf_move
));
2978 if (inst
->opcode
== BRW_OPCODE_MOV
&&
2979 inst
->dst
.file
== MRF
) {
2980 fs_inst
*prev_inst
= last_mrf_move
[inst
->dst
.reg
];
2981 if (prev_inst
&& inst
->equals(prev_inst
)) {
2982 inst
->remove(block
);
2988 /* Clear out the last-write records for MRFs that were overwritten. */
2989 if (inst
->dst
.file
== MRF
) {
2990 last_mrf_move
[inst
->dst
.reg
] = NULL
;
2993 if (inst
->mlen
> 0 && inst
->base_mrf
!= -1) {
2994 /* Found a SEND instruction, which will include two or fewer
2995 * implied MRF writes. We could do better here.
2997 for (int i
= 0; i
< implied_mrf_writes(inst
); i
++) {
2998 last_mrf_move
[inst
->base_mrf
+ i
] = NULL
;
3002 /* Clear out any MRF move records whose sources got overwritten. */
3003 if (inst
->dst
.file
== GRF
) {
3004 for (unsigned int i
= 0; i
< ARRAY_SIZE(last_mrf_move
); i
++) {
3005 if (last_mrf_move
[i
] &&
3006 last_mrf_move
[i
]->src
[0].reg
== inst
->dst
.reg
) {
3007 last_mrf_move
[i
] = NULL
;
3012 if (inst
->opcode
== BRW_OPCODE_MOV
&&
3013 inst
->dst
.file
== MRF
&&
3014 inst
->src
[0].file
== GRF
&&
3015 !inst
->is_partial_write()) {
3016 last_mrf_move
[inst
->dst
.reg
] = inst
;
3021 invalidate_live_intervals();
3027 clear_deps_for_inst_src(fs_inst
*inst
, bool *deps
, int first_grf
, int grf_len
)
3029 /* Clear the flag for registers that actually got read (as expected). */
3030 for (int i
= 0; i
< inst
->sources
; i
++) {
3032 if (inst
->src
[i
].file
== GRF
) {
3033 grf
= inst
->src
[i
].reg
;
3034 } else if (inst
->src
[i
].file
== HW_REG
&&
3035 inst
->src
[i
].fixed_hw_reg
.file
== BRW_GENERAL_REGISTER_FILE
) {
3036 grf
= inst
->src
[i
].fixed_hw_reg
.nr
;
3041 if (grf
>= first_grf
&&
3042 grf
< first_grf
+ grf_len
) {
3043 deps
[grf
- first_grf
] = false;
3044 if (inst
->exec_size
== 16)
3045 deps
[grf
- first_grf
+ 1] = false;
3051 * Implements this workaround for the original 965:
3053 * "[DevBW, DevCL] Implementation Restrictions: As the hardware does not
3054 * check for post destination dependencies on this instruction, software
3055 * must ensure that there is no destination hazard for the case of ‘write
3056 * followed by a posted write’ shown in the following example.
3059 * 2. send r3.xy <rest of send instruction>
3062 * Due to no post-destination dependency check on the ‘send’, the above
3063 * code sequence could have two instructions (1 and 2) in flight at the
3064 * same time that both consider ‘r3’ as the target of their final writes.
3067 fs_visitor::insert_gen4_pre_send_dependency_workarounds(bblock_t
*block
,
3070 int write_len
= inst
->regs_written
;
3071 int first_write_grf
= inst
->dst
.reg
;
3072 bool needs_dep
[BRW_MAX_MRF
];
3073 assert(write_len
< (int)sizeof(needs_dep
) - 1);
3075 memset(needs_dep
, false, sizeof(needs_dep
));
3076 memset(needs_dep
, true, write_len
);
3078 clear_deps_for_inst_src(inst
, needs_dep
, first_write_grf
, write_len
);
3080 /* Walk backwards looking for writes to registers we're writing which
3081 * aren't read since being written. If we hit the start of the program,
3082 * we assume that there are no outstanding dependencies on entry to the
3085 foreach_inst_in_block_reverse_starting_from(fs_inst
, scan_inst
, inst
, block
) {
3086 /* If we hit control flow, assume that there *are* outstanding
3087 * dependencies, and force their cleanup before our instruction.
3089 if (block
->start() == scan_inst
) {
3090 for (int i
= 0; i
< write_len
; i
++) {
3092 DEP_RESOLVE_MOV(bld
.at(block
, inst
), first_write_grf
+ i
);
3097 /* We insert our reads as late as possible on the assumption that any
3098 * instruction but a MOV that might have left us an outstanding
3099 * dependency has more latency than a MOV.
3101 if (scan_inst
->dst
.file
== GRF
) {
3102 for (int i
= 0; i
< scan_inst
->regs_written
; i
++) {
3103 int reg
= scan_inst
->dst
.reg
+ i
;
3105 if (reg
>= first_write_grf
&&
3106 reg
< first_write_grf
+ write_len
&&
3107 needs_dep
[reg
- first_write_grf
]) {
3108 DEP_RESOLVE_MOV(bld
.at(block
, inst
), reg
);
3109 needs_dep
[reg
- first_write_grf
] = false;
3110 if (scan_inst
->exec_size
== 16)
3111 needs_dep
[reg
- first_write_grf
+ 1] = false;
3116 /* Clear the flag for registers that actually got read (as expected). */
3117 clear_deps_for_inst_src(scan_inst
, needs_dep
, first_write_grf
, write_len
);
3119 /* Continue the loop only if we haven't resolved all the dependencies */
3121 for (i
= 0; i
< write_len
; i
++) {
3131 * Implements this workaround for the original 965:
3133 * "[DevBW, DevCL] Errata: A destination register from a send can not be
3134 * used as a destination register until after it has been sourced by an
3135 * instruction with a different destination register.
3138 fs_visitor::insert_gen4_post_send_dependency_workarounds(bblock_t
*block
, fs_inst
*inst
)
3140 int write_len
= inst
->regs_written
;
3141 int first_write_grf
= inst
->dst
.reg
;
3142 bool needs_dep
[BRW_MAX_MRF
];
3143 assert(write_len
< (int)sizeof(needs_dep
) - 1);
3145 memset(needs_dep
, false, sizeof(needs_dep
));
3146 memset(needs_dep
, true, write_len
);
3147 /* Walk forwards looking for writes to registers we're writing which aren't
3148 * read before being written.
3150 foreach_inst_in_block_starting_from(fs_inst
, scan_inst
, inst
, block
) {
3151 /* If we hit control flow, force resolve all remaining dependencies. */
3152 if (block
->end() == scan_inst
) {
3153 for (int i
= 0; i
< write_len
; i
++) {
3155 DEP_RESOLVE_MOV(bld
.at(block
, scan_inst
), first_write_grf
+ i
);
3160 /* Clear the flag for registers that actually got read (as expected). */
3161 clear_deps_for_inst_src(scan_inst
, needs_dep
, first_write_grf
, write_len
);
3163 /* We insert our reads as late as possible since they're reading the
3164 * result of a SEND, which has massive latency.
3166 if (scan_inst
->dst
.file
== GRF
&&
3167 scan_inst
->dst
.reg
>= first_write_grf
&&
3168 scan_inst
->dst
.reg
< first_write_grf
+ write_len
&&
3169 needs_dep
[scan_inst
->dst
.reg
- first_write_grf
]) {
3170 DEP_RESOLVE_MOV(bld
.at(block
, scan_inst
), scan_inst
->dst
.reg
);
3171 needs_dep
[scan_inst
->dst
.reg
- first_write_grf
] = false;
3174 /* Continue the loop only if we haven't resolved all the dependencies */
3176 for (i
= 0; i
< write_len
; i
++) {
3186 fs_visitor::insert_gen4_send_dependency_workarounds()
3188 if (devinfo
->gen
!= 4 || devinfo
->is_g4x
)
3191 bool progress
= false;
3193 /* Note that we're done with register allocation, so GRF fs_regs always
3194 * have a .reg_offset of 0.
3197 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
3198 if (inst
->mlen
!= 0 && inst
->dst
.file
== GRF
) {
3199 insert_gen4_pre_send_dependency_workarounds(block
, inst
);
3200 insert_gen4_post_send_dependency_workarounds(block
, inst
);
3206 invalidate_live_intervals();
3210 * Turns the generic expression-style uniform pull constant load instruction
3211 * into a hardware-specific series of instructions for loading a pull
3214 * The expression style allows the CSE pass before this to optimize out
3215 * repeated loads from the same offset, and gives the pre-register-allocation
3216 * scheduling full flexibility, while the conversion to native instructions
3217 * allows the post-register-allocation scheduler the best information
3220 * Note that execution masking for setting up pull constant loads is special:
3221 * the channels that need to be written are unrelated to the current execution
3222 * mask, since a later instruction will use one of the result channels as a
3223 * source operand for all 8 or 16 of its channels.
3226 fs_visitor::lower_uniform_pull_constant_loads()
3228 foreach_block_and_inst (block
, fs_inst
, inst
, cfg
) {
3229 if (inst
->opcode
!= FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
)
3232 if (devinfo
->gen
>= 7) {
3233 /* The offset arg before was a vec4-aligned byte offset. We need to
3234 * turn it into a dword offset.
3236 fs_reg const_offset_reg
= inst
->src
[1];
3237 assert(const_offset_reg
.file
== IMM
&&
3238 const_offset_reg
.type
== BRW_REGISTER_TYPE_UD
);
3239 const_offset_reg
.fixed_hw_reg
.dw1
.ud
/= 4;
3240 fs_reg payload
= fs_reg(GRF
, alloc
.allocate(1));
3242 /* We have to use a message header on Skylake to get SIMD4x2 mode.
3243 * Reserve space for the register.
3245 if (devinfo
->gen
>= 9) {
3246 payload
.reg_offset
++;
3247 alloc
.sizes
[payload
.reg
] = 2;
3250 /* This is actually going to be a MOV, but since only the first dword
3251 * is accessed, we have a special opcode to do just that one. Note
3252 * that this needs to be an operation that will be considered a def
3253 * by live variable analysis, or register allocation will explode.
3255 fs_inst
*setup
= new(mem_ctx
) fs_inst(FS_OPCODE_SET_SIMD4X2_OFFSET
,
3256 8, payload
, const_offset_reg
);
3257 setup
->force_writemask_all
= true;
3259 setup
->ir
= inst
->ir
;
3260 setup
->annotation
= inst
->annotation
;
3261 inst
->insert_before(block
, setup
);
3263 /* Similarly, this will only populate the first 4 channels of the
3264 * result register (since we only use smear values from 0-3), but we
3265 * don't tell the optimizer.
3267 inst
->opcode
= FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7
;
3268 inst
->src
[1] = payload
;
3270 invalidate_live_intervals();
3272 /* Before register allocation, we didn't tell the scheduler about the
3273 * MRF we use. We know it's safe to use this MRF because nothing
3274 * else does except for register spill/unspill, which generates and
3275 * uses its MRF within a single IR instruction.
3277 inst
->base_mrf
= 14;
3284 fs_visitor::lower_load_payload()
3286 bool progress
= false;
3288 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
3289 if (inst
->opcode
!= SHADER_OPCODE_LOAD_PAYLOAD
)
3292 assert(inst
->dst
.file
== MRF
|| inst
->dst
.file
== GRF
);
3293 assert(inst
->saturate
== false);
3295 const fs_builder ibld
= bld
.group(inst
->exec_size
, inst
->force_sechalf
)
3296 .exec_all(inst
->force_writemask_all
)
3298 fs_reg dst
= inst
->dst
;
3300 /* Get rid of COMPR4. We'll add it back in if we need it */
3301 if (dst
.file
== MRF
)
3302 dst
.reg
= dst
.reg
& ~BRW_MRF_COMPR4
;
3305 for (uint8_t i
= 0; i
< inst
->header_size
; i
++) {
3306 if (inst
->src
[i
].file
!= BAD_FILE
) {
3307 fs_reg mov_dst
= retype(dst
, BRW_REGISTER_TYPE_UD
);
3308 fs_reg mov_src
= retype(inst
->src
[i
], BRW_REGISTER_TYPE_UD
);
3310 ibld
.exec_all().MOV(mov_dst
, mov_src
);
3312 dst
= offset(dst
, 1);
3315 dst
.width
= inst
->exec_size
;
3316 if (inst
->dst
.file
== MRF
&& (inst
->dst
.reg
& BRW_MRF_COMPR4
) &&
3317 inst
->exec_size
> 8) {
3318 /* In this case, the payload portion of the LOAD_PAYLOAD isn't
3319 * a straightforward copy. Instead, the result of the
3320 * LOAD_PAYLOAD is treated as interleaved and the first four
3321 * non-header sources are unpacked as:
3332 * This is used for gen <= 5 fb writes.
3334 assert(inst
->exec_size
== 16);
3335 assert(inst
->header_size
+ 4 <= inst
->sources
);
3336 for (uint8_t i
= inst
->header_size
; i
< inst
->header_size
+ 4; i
++) {
3337 if (inst
->src
[i
].file
!= BAD_FILE
) {
3338 if (devinfo
->has_compr4
) {
3339 fs_reg compr4_dst
= retype(dst
, inst
->src
[i
].type
);
3340 compr4_dst
.reg
|= BRW_MRF_COMPR4
;
3341 ibld
.MOV(compr4_dst
, inst
->src
[i
]);
3343 /* Platform doesn't have COMPR4. We have to fake it */
3344 fs_reg mov_dst
= retype(dst
, inst
->src
[i
].type
);
3346 ibld
.half(0).MOV(mov_dst
, half(inst
->src
[i
], 0));
3347 ibld
.half(1).MOV(offset(mov_dst
, 4), half(inst
->src
[i
], 1));
3354 /* The loop above only ever incremented us through the first set
3355 * of 4 registers. However, thanks to the magic of COMPR4, we
3356 * actually wrote to the first 8 registers, so we need to take
3357 * that into account now.
3361 /* The COMPR4 code took care of the first 4 sources. We'll let
3362 * the regular path handle any remaining sources. Yes, we are
3363 * modifying the instruction but we're about to delete it so
3364 * this really doesn't hurt anything.
3366 inst
->header_size
+= 4;
3369 for (uint8_t i
= inst
->header_size
; i
< inst
->sources
; i
++) {
3370 if (inst
->src
[i
].file
!= BAD_FILE
)
3371 ibld
.MOV(retype(dst
, inst
->src
[i
].type
), inst
->src
[i
]);
3372 dst
= offset(dst
, 1);
3375 inst
->remove(block
);
3380 invalidate_live_intervals();
3386 fs_visitor::lower_integer_multiplication()
3388 bool progress
= false;
3390 /* Gen8's MUL instruction can do a 32-bit x 32-bit -> 32-bit operation
3391 * directly, but Cherryview cannot.
3393 if (devinfo
->gen
>= 8 && !devinfo
->is_cherryview
)
3396 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
3397 if (inst
->opcode
!= BRW_OPCODE_MUL
||
3398 inst
->dst
.is_accumulator() ||
3399 (inst
->dst
.type
!= BRW_REGISTER_TYPE_D
&&
3400 inst
->dst
.type
!= BRW_REGISTER_TYPE_UD
))
3403 const fs_builder ibld
= bld
.at(block
, inst
);
3405 /* The MUL instruction isn't commutative. On Gen <= 6, only the low
3406 * 16-bits of src0 are read, and on Gen >= 7 only the low 16-bits of
3409 * If multiplying by an immediate value that fits in 16-bits, do a
3410 * single MUL instruction with that value in the proper location.
3412 if (inst
->src
[1].file
== IMM
&&
3413 inst
->src
[1].fixed_hw_reg
.dw1
.ud
< (1 << 16)) {
3414 if (devinfo
->gen
< 7) {
3415 fs_reg
imm(GRF
, alloc
.allocate(dispatch_width
/ 8),
3416 inst
->dst
.type
, dispatch_width
);
3417 ibld
.MOV(imm
, inst
->src
[1]);
3418 ibld
.MUL(inst
->dst
, imm
, inst
->src
[0]);
3420 ibld
.MUL(inst
->dst
, inst
->src
[0], inst
->src
[1]);
3423 /* Gen < 8 (and some Gen8+ low-power parts like Cherryview) cannot
3424 * do 32-bit integer multiplication in one instruction, but instead
3425 * must do a sequence (which actually calculates a 64-bit result):
3427 * mul(8) acc0<1>D g3<8,8,1>D g4<8,8,1>D
3428 * mach(8) null g3<8,8,1>D g4<8,8,1>D
3429 * mov(8) g2<1>D acc0<8,8,1>D
3431 * But on Gen > 6, the ability to use second accumulator register
3432 * (acc1) for non-float data types was removed, preventing a simple
3433 * implementation in SIMD16. A 16-channel result can be calculated by
3434 * executing the three instructions twice in SIMD8, once with quarter
3435 * control of 1Q for the first eight channels and again with 2Q for
3436 * the second eight channels.
3438 * Which accumulator register is implicitly accessed (by AccWrEnable
3439 * for instance) is determined by the quarter control. Unfortunately
3440 * Ivybridge (and presumably Baytrail) has a hardware bug in which an
3441 * implicit accumulator access by an instruction with 2Q will access
3442 * acc1 regardless of whether the data type is usable in acc1.
3444 * Specifically, the 2Q mach(8) writes acc1 which does not exist for
3445 * integer data types.
3447 * Since we only want the low 32-bits of the result, we can do two
3448 * 32-bit x 16-bit multiplies (like the mul and mach are doing), and
3449 * adjust the high result and add them (like the mach is doing):
3451 * mul(8) g7<1>D g3<8,8,1>D g4.0<8,8,1>UW
3452 * mul(8) g8<1>D g3<8,8,1>D g4.1<8,8,1>UW
3453 * shl(8) g9<1>D g8<8,8,1>D 16D
3454 * add(8) g2<1>D g7<8,8,1>D g8<8,8,1>D
3456 * We avoid the shl instruction by realizing that we only want to add
3457 * the low 16-bits of the "high" result to the high 16-bits of the
3458 * "low" result and using proper regioning on the add:
3460 * mul(8) g7<1>D g3<8,8,1>D g4.0<16,8,2>UW
3461 * mul(8) g8<1>D g3<8,8,1>D g4.1<16,8,2>UW
3462 * add(8) g7.1<2>UW g7.1<16,8,2>UW g8<16,8,2>UW
3464 * Since it does not use the (single) accumulator register, we can
3465 * schedule multi-component multiplications much better.
3468 if (inst
->conditional_mod
&& inst
->dst
.is_null()) {
3469 inst
->dst
= fs_reg(GRF
, alloc
.allocate(dispatch_width
/ 8),
3470 inst
->dst
.type
, dispatch_width
);
3472 fs_reg low
= inst
->dst
;
3473 fs_reg
high(GRF
, alloc
.allocate(dispatch_width
/ 8),
3474 inst
->dst
.type
, dispatch_width
);
3476 if (brw
->gen
>= 7) {
3477 fs_reg src1_0_w
= inst
->src
[1];
3478 fs_reg src1_1_w
= inst
->src
[1];
3480 if (inst
->src
[1].file
== IMM
) {
3481 src1_0_w
.fixed_hw_reg
.dw1
.ud
&= 0xffff;
3482 src1_1_w
.fixed_hw_reg
.dw1
.ud
>>= 16;
3484 src1_0_w
.type
= BRW_REGISTER_TYPE_UW
;
3485 src1_0_w
.stride
= 2;
3487 src1_1_w
.type
= BRW_REGISTER_TYPE_UW
;
3488 src1_1_w
.stride
= 2;
3489 src1_1_w
.subreg_offset
+= type_sz(BRW_REGISTER_TYPE_UW
);
3491 ibld
.MUL(low
, inst
->src
[0], src1_0_w
);
3492 ibld
.MUL(high
, inst
->src
[0], src1_1_w
);
3494 fs_reg src0_0_w
= inst
->src
[0];
3495 fs_reg src0_1_w
= inst
->src
[0];
3497 src0_0_w
.type
= BRW_REGISTER_TYPE_UW
;
3498 src0_0_w
.stride
= 2;
3500 src0_1_w
.type
= BRW_REGISTER_TYPE_UW
;
3501 src0_1_w
.stride
= 2;
3502 src0_1_w
.subreg_offset
+= type_sz(BRW_REGISTER_TYPE_UW
);
3504 ibld
.MUL(low
, src0_0_w
, inst
->src
[1]);
3505 ibld
.MUL(high
, src0_1_w
, inst
->src
[1]);
3508 fs_reg dst
= inst
->dst
;
3509 dst
.type
= BRW_REGISTER_TYPE_UW
;
3510 dst
.subreg_offset
= 2;
3513 high
.type
= BRW_REGISTER_TYPE_UW
;
3516 low
.type
= BRW_REGISTER_TYPE_UW
;
3517 low
.subreg_offset
= 2;
3520 ibld
.ADD(dst
, low
, high
);
3522 if (inst
->conditional_mod
) {
3523 fs_reg
null(retype(brw_null_reg(), inst
->dst
.type
));
3524 set_condmod(inst
->conditional_mod
,
3525 ibld
.MOV(null
, inst
->dst
));
3529 inst
->remove(block
);
3534 invalidate_live_intervals();
3540 fs_visitor::dump_instructions()
3542 dump_instructions(NULL
);
3546 fs_visitor::dump_instructions(const char *name
)
3548 FILE *file
= stderr
;
3549 if (name
&& geteuid() != 0) {
3550 file
= fopen(name
, "w");
3556 calculate_register_pressure();
3557 int ip
= 0, max_pressure
= 0;
3558 foreach_block_and_inst(block
, backend_instruction
, inst
, cfg
) {
3559 max_pressure
= MAX2(max_pressure
, regs_live_at_ip
[ip
]);
3560 fprintf(file
, "{%3d} %4d: ", regs_live_at_ip
[ip
], ip
);
3561 dump_instruction(inst
, file
);
3564 fprintf(file
, "Maximum %3d registers live at once.\n", max_pressure
);
3567 foreach_in_list(backend_instruction
, inst
, &instructions
) {
3568 fprintf(file
, "%4d: ", ip
++);
3569 dump_instruction(inst
, file
);
3573 if (file
!= stderr
) {
3579 fs_visitor::dump_instruction(backend_instruction
*be_inst
)
3581 dump_instruction(be_inst
, stderr
);
3585 fs_visitor::dump_instruction(backend_instruction
*be_inst
, FILE *file
)
3587 fs_inst
*inst
= (fs_inst
*)be_inst
;
3589 if (inst
->predicate
) {
3590 fprintf(file
, "(%cf0.%d) ",
3591 inst
->predicate_inverse
? '-' : '+',
3595 fprintf(file
, "%s", brw_instruction_name(inst
->opcode
));
3597 fprintf(file
, ".sat");
3598 if (inst
->conditional_mod
) {
3599 fprintf(file
, "%s", conditional_modifier
[inst
->conditional_mod
]);
3600 if (!inst
->predicate
&&
3601 (devinfo
->gen
< 5 || (inst
->opcode
!= BRW_OPCODE_SEL
&&
3602 inst
->opcode
!= BRW_OPCODE_IF
&&
3603 inst
->opcode
!= BRW_OPCODE_WHILE
))) {
3604 fprintf(file
, ".f0.%d", inst
->flag_subreg
);
3607 fprintf(file
, "(%d) ", inst
->exec_size
);
3610 fprintf(file
, "(mlen: %d) ", inst
->mlen
);
3613 switch (inst
->dst
.file
) {
3615 fprintf(file
, "vgrf%d", inst
->dst
.reg
);
3616 if (inst
->dst
.width
!= dispatch_width
)
3617 fprintf(file
, "@%d", inst
->dst
.width
);
3618 if (alloc
.sizes
[inst
->dst
.reg
] != inst
->dst
.width
/ 8 ||
3619 inst
->dst
.subreg_offset
)
3620 fprintf(file
, "+%d.%d",
3621 inst
->dst
.reg_offset
, inst
->dst
.subreg_offset
);
3624 fprintf(file
, "m%d", inst
->dst
.reg
);
3627 fprintf(file
, "(null)");
3630 fprintf(file
, "***u%d***", inst
->dst
.reg
+ inst
->dst
.reg_offset
);
3633 fprintf(file
, "***attr%d***", inst
->dst
.reg
+ inst
->dst
.reg_offset
);
3636 if (inst
->dst
.fixed_hw_reg
.file
== BRW_ARCHITECTURE_REGISTER_FILE
) {
3637 switch (inst
->dst
.fixed_hw_reg
.nr
) {
3639 fprintf(file
, "null");
3641 case BRW_ARF_ADDRESS
:
3642 fprintf(file
, "a0.%d", inst
->dst
.fixed_hw_reg
.subnr
);
3644 case BRW_ARF_ACCUMULATOR
:
3645 fprintf(file
, "acc%d", inst
->dst
.fixed_hw_reg
.subnr
);
3648 fprintf(file
, "f%d.%d", inst
->dst
.fixed_hw_reg
.nr
& 0xf,
3649 inst
->dst
.fixed_hw_reg
.subnr
);
3652 fprintf(file
, "arf%d.%d", inst
->dst
.fixed_hw_reg
.nr
& 0xf,
3653 inst
->dst
.fixed_hw_reg
.subnr
);
3657 fprintf(file
, "hw_reg%d", inst
->dst
.fixed_hw_reg
.nr
);
3659 if (inst
->dst
.fixed_hw_reg
.subnr
)
3660 fprintf(file
, "+%d", inst
->dst
.fixed_hw_reg
.subnr
);
3663 fprintf(file
, "???");
3666 fprintf(file
, ":%s, ", brw_reg_type_letters(inst
->dst
.type
));
3668 for (int i
= 0; i
< inst
->sources
; i
++) {
3669 if (inst
->src
[i
].negate
)
3671 if (inst
->src
[i
].abs
)
3673 switch (inst
->src
[i
].file
) {
3675 fprintf(file
, "vgrf%d", inst
->src
[i
].reg
);
3676 if (inst
->src
[i
].width
!= dispatch_width
)
3677 fprintf(file
, "@%d", inst
->src
[i
].width
);
3678 if (alloc
.sizes
[inst
->src
[i
].reg
] != inst
->src
[i
].width
/ 8 ||
3679 inst
->src
[i
].subreg_offset
)
3680 fprintf(file
, "+%d.%d", inst
->src
[i
].reg_offset
,
3681 inst
->src
[i
].subreg_offset
);
3684 fprintf(file
, "***m%d***", inst
->src
[i
].reg
);
3687 fprintf(file
, "attr%d", inst
->src
[i
].reg
+ inst
->src
[i
].reg_offset
);
3690 fprintf(file
, "u%d", inst
->src
[i
].reg
+ inst
->src
[i
].reg_offset
);
3691 if (inst
->src
[i
].reladdr
) {
3692 fprintf(file
, "+reladdr");
3693 } else if (inst
->src
[i
].subreg_offset
) {
3694 fprintf(file
, "+%d.%d", inst
->src
[i
].reg_offset
,
3695 inst
->src
[i
].subreg_offset
);
3699 fprintf(file
, "(null)");
3702 switch (inst
->src
[i
].type
) {
3703 case BRW_REGISTER_TYPE_F
:
3704 fprintf(file
, "%ff", inst
->src
[i
].fixed_hw_reg
.dw1
.f
);
3706 case BRW_REGISTER_TYPE_W
:
3707 case BRW_REGISTER_TYPE_D
:
3708 fprintf(file
, "%dd", inst
->src
[i
].fixed_hw_reg
.dw1
.d
);
3710 case BRW_REGISTER_TYPE_UW
:
3711 case BRW_REGISTER_TYPE_UD
:
3712 fprintf(file
, "%uu", inst
->src
[i
].fixed_hw_reg
.dw1
.ud
);
3714 case BRW_REGISTER_TYPE_VF
:
3715 fprintf(file
, "[%-gF, %-gF, %-gF, %-gF]",
3716 brw_vf_to_float((inst
->src
[i
].fixed_hw_reg
.dw1
.ud
>> 0) & 0xff),
3717 brw_vf_to_float((inst
->src
[i
].fixed_hw_reg
.dw1
.ud
>> 8) & 0xff),
3718 brw_vf_to_float((inst
->src
[i
].fixed_hw_reg
.dw1
.ud
>> 16) & 0xff),
3719 brw_vf_to_float((inst
->src
[i
].fixed_hw_reg
.dw1
.ud
>> 24) & 0xff));
3722 fprintf(file
, "???");
3727 if (inst
->src
[i
].fixed_hw_reg
.negate
)
3729 if (inst
->src
[i
].fixed_hw_reg
.abs
)
3731 if (inst
->src
[i
].fixed_hw_reg
.file
== BRW_ARCHITECTURE_REGISTER_FILE
) {
3732 switch (inst
->src
[i
].fixed_hw_reg
.nr
) {
3734 fprintf(file
, "null");
3736 case BRW_ARF_ADDRESS
:
3737 fprintf(file
, "a0.%d", inst
->src
[i
].fixed_hw_reg
.subnr
);
3739 case BRW_ARF_ACCUMULATOR
:
3740 fprintf(file
, "acc%d", inst
->src
[i
].fixed_hw_reg
.subnr
);
3743 fprintf(file
, "f%d.%d", inst
->src
[i
].fixed_hw_reg
.nr
& 0xf,
3744 inst
->src
[i
].fixed_hw_reg
.subnr
);
3747 fprintf(file
, "arf%d.%d", inst
->src
[i
].fixed_hw_reg
.nr
& 0xf,
3748 inst
->src
[i
].fixed_hw_reg
.subnr
);
3752 fprintf(file
, "hw_reg%d", inst
->src
[i
].fixed_hw_reg
.nr
);
3754 if (inst
->src
[i
].fixed_hw_reg
.subnr
)
3755 fprintf(file
, "+%d", inst
->src
[i
].fixed_hw_reg
.subnr
);
3756 if (inst
->src
[i
].fixed_hw_reg
.abs
)
3760 fprintf(file
, "???");
3763 if (inst
->src
[i
].abs
)
3766 if (inst
->src
[i
].file
!= IMM
) {
3767 fprintf(file
, ":%s", brw_reg_type_letters(inst
->src
[i
].type
));
3770 if (i
< inst
->sources
- 1 && inst
->src
[i
+ 1].file
!= BAD_FILE
)
3771 fprintf(file
, ", ");
3776 if (dispatch_width
== 16 && inst
->exec_size
== 8) {
3777 if (inst
->force_sechalf
)
3778 fprintf(file
, "2ndhalf ");
3780 fprintf(file
, "1sthalf ");
3783 fprintf(file
, "\n");
3787 * Possibly returns an instruction that set up @param reg.
3789 * Sometimes we want to take the result of some expression/variable
3790 * dereference tree and rewrite the instruction generating the result
3791 * of the tree. When processing the tree, we know that the
3792 * instructions generated are all writing temporaries that are dead
3793 * outside of this tree. So, if we have some instructions that write
3794 * a temporary, we're free to point that temp write somewhere else.
3796 * Note that this doesn't guarantee that the instruction generated
3797 * only reg -- it might be the size=4 destination of a texture instruction.
3800 fs_visitor::get_instruction_generating_reg(fs_inst
*start
,
3805 end
->is_partial_write() ||
3807 !reg
.equals(end
->dst
)) {
3815 fs_visitor::setup_payload_gen6()
3818 (prog
->InputsRead
& (1 << VARYING_SLOT_POS
)) != 0;
3819 unsigned barycentric_interp_modes
=
3820 (stage
== MESA_SHADER_FRAGMENT
) ?
3821 ((brw_wm_prog_data
*) this->prog_data
)->barycentric_interp_modes
: 0;
3823 assert(devinfo
->gen
>= 6);
3825 /* R0-1: masks, pixel X/Y coordinates. */
3826 payload
.num_regs
= 2;
3827 /* R2: only for 32-pixel dispatch.*/
3829 /* R3-26: barycentric interpolation coordinates. These appear in the
3830 * same order that they appear in the brw_wm_barycentric_interp_mode
3831 * enum. Each set of coordinates occupies 2 registers if dispatch width
3832 * == 8 and 4 registers if dispatch width == 16. Coordinates only
3833 * appear if they were enabled using the "Barycentric Interpolation
3834 * Mode" bits in WM_STATE.
3836 for (int i
= 0; i
< BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT
; ++i
) {
3837 if (barycentric_interp_modes
& (1 << i
)) {
3838 payload
.barycentric_coord_reg
[i
] = payload
.num_regs
;
3839 payload
.num_regs
+= 2;
3840 if (dispatch_width
== 16) {
3841 payload
.num_regs
+= 2;
3846 /* R27: interpolated depth if uses source depth */
3848 payload
.source_depth_reg
= payload
.num_regs
;
3850 if (dispatch_width
== 16) {
3851 /* R28: interpolated depth if not SIMD8. */
3855 /* R29: interpolated W set if GEN6_WM_USES_SOURCE_W. */
3857 payload
.source_w_reg
= payload
.num_regs
;
3859 if (dispatch_width
== 16) {
3860 /* R30: interpolated W if not SIMD8. */
3865 if (stage
== MESA_SHADER_FRAGMENT
) {
3866 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
3867 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
3868 prog_data
->uses_pos_offset
= key
->compute_pos_offset
;
3869 /* R31: MSAA position offsets. */
3870 if (prog_data
->uses_pos_offset
) {
3871 payload
.sample_pos_reg
= payload
.num_regs
;
3876 /* R32: MSAA input coverage mask */
3877 if (prog
->SystemValuesRead
& SYSTEM_BIT_SAMPLE_MASK_IN
) {
3878 assert(devinfo
->gen
>= 7);
3879 payload
.sample_mask_in_reg
= payload
.num_regs
;
3881 if (dispatch_width
== 16) {
3882 /* R33: input coverage mask if not SIMD8. */
3887 /* R34-: bary for 32-pixel. */
3888 /* R58-59: interp W for 32-pixel. */
3890 if (prog
->OutputsWritten
& BITFIELD64_BIT(FRAG_RESULT_DEPTH
)) {
3891 source_depth_to_render_target
= true;
3896 fs_visitor::setup_vs_payload()
3898 /* R0: thread header, R1: urb handles */
3899 payload
.num_regs
= 2;
3903 fs_visitor::setup_cs_payload()
3905 assert(brw
->gen
>= 7);
3907 payload
.num_regs
= 1;
3911 fs_visitor::assign_binding_table_offsets()
3913 assert(stage
== MESA_SHADER_FRAGMENT
);
3914 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
3915 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
3916 uint32_t next_binding_table_offset
= 0;
3918 /* If there are no color regions, we still perform an FB write to a null
3919 * renderbuffer, which we place at surface index 0.
3921 prog_data
->binding_table
.render_target_start
= next_binding_table_offset
;
3922 next_binding_table_offset
+= MAX2(key
->nr_color_regions
, 1);
3924 assign_common_binding_table_offsets(next_binding_table_offset
);
3928 fs_visitor::calculate_register_pressure()
3930 invalidate_live_intervals();
3931 calculate_live_intervals();
3933 unsigned num_instructions
= 0;
3934 foreach_block(block
, cfg
)
3935 num_instructions
+= block
->instructions
.length();
3937 regs_live_at_ip
= rzalloc_array(mem_ctx
, int, num_instructions
);
3939 for (unsigned reg
= 0; reg
< alloc
.count
; reg
++) {
3940 for (int ip
= virtual_grf_start
[reg
]; ip
<= virtual_grf_end
[reg
]; ip
++)
3941 regs_live_at_ip
[ip
] += alloc
.sizes
[reg
];
3946 fs_visitor::optimize()
3948 /* bld is the common builder object pointing at the end of the program we
3949 * used to translate it into i965 IR. For the optimization and lowering
3950 * passes coming next, any code added after the end of the program without
3951 * having explicitly called fs_builder::at() clearly points at a mistake.
3952 * Ideally optimization passes wouldn't be part of the visitor so they
3953 * wouldn't have access to bld at all, but they do, so just in case some
3954 * pass forgets to ask for a location explicitly set it to NULL here to
3957 bld
= bld
.at(NULL
, NULL
);
3959 split_virtual_grfs();
3961 move_uniform_array_access_to_pull_constants();
3962 assign_constant_locations();
3963 demote_pull_constants();
3965 #define OPT(pass, args...) ({ \
3967 bool this_progress = pass(args); \
3969 if (unlikely(INTEL_DEBUG & DEBUG_OPTIMIZER) && this_progress) { \
3970 char filename[64]; \
3971 snprintf(filename, 64, "%s%d-%04d-%02d-%02d-" #pass, \
3972 stage_abbrev, dispatch_width, shader_prog ? shader_prog->Name : 0, iteration, pass_num); \
3974 backend_shader::dump_instructions(filename); \
3977 progress = progress || this_progress; \
3981 if (unlikely(INTEL_DEBUG
& DEBUG_OPTIMIZER
)) {
3983 snprintf(filename
, 64, "%s%d-%04d-00-start",
3984 stage_abbrev
, dispatch_width
,
3985 shader_prog
? shader_prog
->Name
: 0);
3987 backend_shader::dump_instructions(filename
);
3998 OPT(remove_duplicate_mrf_writes
);
4002 OPT(opt_copy_propagate
);
4003 OPT(opt_peephole_predicated_break
);
4004 OPT(opt_cmod_propagation
);
4005 OPT(dead_code_eliminate
);
4006 OPT(opt_peephole_sel
);
4007 OPT(dead_control_flow_eliminate
, this);
4008 OPT(opt_register_renaming
);
4009 OPT(opt_redundant_discard_jumps
);
4010 OPT(opt_saturate_propagation
);
4011 OPT(opt_zero_samples
);
4012 OPT(register_coalesce
);
4013 OPT(compute_to_mrf
);
4014 OPT(eliminate_find_live_channel
);
4016 OPT(compact_virtual_grfs
);
4021 OPT(opt_sampler_eot
);
4023 if (OPT(lower_load_payload
)) {
4024 split_virtual_grfs();
4025 OPT(register_coalesce
);
4026 OPT(compute_to_mrf
);
4027 OPT(dead_code_eliminate
);
4030 OPT(opt_combine_constants
);
4031 OPT(lower_integer_multiplication
);
4033 lower_uniform_pull_constant_loads();
4037 * Three source instruction must have a GRF/MRF destination register.
4038 * ARF NULL is not allowed. Fix that up by allocating a temporary GRF.
4041 fs_visitor::fixup_3src_null_dest()
4043 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
4044 if (inst
->is_3src() && inst
->dst
.is_null()) {
4045 inst
->dst
= fs_reg(GRF
, alloc
.allocate(dispatch_width
/ 8),
4052 fs_visitor::allocate_registers()
4054 bool allocated_without_spills
;
4056 static const enum instruction_scheduler_mode pre_modes
[] = {
4058 SCHEDULE_PRE_NON_LIFO
,
4062 /* Try each scheduling heuristic to see if it can successfully register
4063 * allocate without spilling. They should be ordered by decreasing
4064 * performance but increasing likelihood of allocating.
4066 for (unsigned i
= 0; i
< ARRAY_SIZE(pre_modes
); i
++) {
4067 schedule_instructions(pre_modes
[i
]);
4070 assign_regs_trivial();
4071 allocated_without_spills
= true;
4073 allocated_without_spills
= assign_regs(false);
4075 if (allocated_without_spills
)
4079 if (!allocated_without_spills
) {
4080 /* We assume that any spilling is worse than just dropping back to
4081 * SIMD8. There's probably actually some intermediate point where
4082 * SIMD16 with a couple of spills is still better.
4084 if (dispatch_width
== 16) {
4085 fail("Failure to register allocate. Reduce number of "
4086 "live scalar values to avoid this.");
4088 perf_debug("%s shader triggered register spilling. "
4089 "Try reducing the number of live scalar values to "
4090 "improve performance.\n", stage_name
);
4093 /* Since we're out of heuristics, just go spill registers until we
4094 * get an allocation.
4096 while (!assign_regs(true)) {
4102 /* This must come after all optimization and register allocation, since
4103 * it inserts dead code that happens to have side effects, and it does
4104 * so based on the actual physical registers in use.
4106 insert_gen4_send_dependency_workarounds();
4111 if (!allocated_without_spills
)
4112 schedule_instructions(SCHEDULE_POST
);
4114 if (last_scratch
> 0)
4115 prog_data
->total_scratch
= brw_get_scratch_size(last_scratch
);
4119 fs_visitor::run_vs()
4121 assert(stage
== MESA_SHADER_VERTEX
);
4123 assign_common_binding_table_offsets(0);
4126 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
)
4127 emit_shader_time_begin();
4136 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
)
4137 emit_shader_time_end();
4143 assign_curb_setup();
4144 assign_vs_urb_setup();
4146 fixup_3src_null_dest();
4147 allocate_registers();
4153 fs_visitor::run_fs()
4155 brw_wm_prog_data
*wm_prog_data
= (brw_wm_prog_data
*) this->prog_data
;
4156 brw_wm_prog_key
*wm_key
= (brw_wm_prog_key
*) this->key
;
4158 assert(stage
== MESA_SHADER_FRAGMENT
);
4160 sanity_param_count
= prog
->Parameters
->NumParameters
;
4162 assign_binding_table_offsets();
4164 if (devinfo
->gen
>= 6)
4165 setup_payload_gen6();
4167 setup_payload_gen4();
4171 } else if (brw
->use_rep_send
&& dispatch_width
== 16) {
4172 emit_repclear_shader();
4174 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
)
4175 emit_shader_time_begin();
4177 calculate_urb_setup();
4178 if (prog
->InputsRead
> 0) {
4179 if (devinfo
->gen
< 6)
4180 emit_interpolation_setup_gen4();
4182 emit_interpolation_setup_gen6();
4185 /* We handle discards by keeping track of the still-live pixels in f0.1.
4186 * Initialize it with the dispatched pixels.
4188 if (wm_prog_data
->uses_kill
) {
4189 fs_inst
*discard_init
= emit(FS_OPCODE_MOV_DISPATCH_TO_FLAGS
);
4190 discard_init
->flag_subreg
= 1;
4193 /* Generate FS IR for main(). (the visitor only descends into
4194 * functions called "main").
4201 if (wm_prog_data
->uses_kill
)
4202 emit(FS_OPCODE_PLACEHOLDER_HALT
);
4204 if (wm_key
->alpha_test_func
)
4209 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
)
4210 emit_shader_time_end();
4216 assign_curb_setup();
4219 fixup_3src_null_dest();
4220 allocate_registers();
4226 if (dispatch_width
== 8)
4227 wm_prog_data
->reg_blocks
= brw_register_blocks(grf_used
);
4229 wm_prog_data
->reg_blocks_16
= brw_register_blocks(grf_used
);
4231 /* If any state parameters were appended, then ParameterValues could have
4232 * been realloced, in which case the driver uniform storage set up by
4233 * _mesa_associate_uniform_storage() would point to freed memory. Make
4234 * sure that didn't happen.
4236 assert(sanity_param_count
== prog
->Parameters
->NumParameters
);
4242 fs_visitor::run_cs()
4244 assert(stage
== MESA_SHADER_COMPUTE
);
4247 sanity_param_count
= prog
->Parameters
->NumParameters
;
4249 assign_common_binding_table_offsets(0);
4253 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
)
4254 emit_shader_time_begin();
4261 emit_cs_terminate();
4263 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
)
4264 emit_shader_time_end();
4270 assign_curb_setup();
4272 fixup_3src_null_dest();
4273 allocate_registers();
4278 /* If any state parameters were appended, then ParameterValues could have
4279 * been realloced, in which case the driver uniform storage set up by
4280 * _mesa_associate_uniform_storage() would point to freed memory. Make
4281 * sure that didn't happen.
4283 assert(sanity_param_count
== prog
->Parameters
->NumParameters
);
4289 brw_wm_fs_emit(struct brw_context
*brw
,
4291 const struct brw_wm_prog_key
*key
,
4292 struct brw_wm_prog_data
*prog_data
,
4293 struct gl_fragment_program
*fp
,
4294 struct gl_shader_program
*prog
,
4295 unsigned *final_assembly_size
)
4297 bool start_busy
= false;
4298 double start_time
= 0;
4300 if (unlikely(brw
->perf_debug
)) {
4301 start_busy
= (brw
->batch
.last_bo
&&
4302 drm_intel_bo_busy(brw
->batch
.last_bo
));
4303 start_time
= get_time();
4306 struct brw_shader
*shader
= NULL
;
4308 shader
= (brw_shader
*) prog
->_LinkedShaders
[MESA_SHADER_FRAGMENT
];
4310 if (unlikely(INTEL_DEBUG
& DEBUG_WM
))
4311 brw_dump_ir("fragment", prog
, &shader
->base
, &fp
->Base
);
4313 /* Now the main event: Visit the shader IR and generate our FS IR for it.
4315 fs_visitor
v(brw
, mem_ctx
, MESA_SHADER_FRAGMENT
, key
, &prog_data
->base
,
4316 prog
, &fp
->Base
, 8);
4319 prog
->LinkStatus
= false;
4320 ralloc_strcat(&prog
->InfoLog
, v
.fail_msg
);
4323 _mesa_problem(NULL
, "Failed to compile fragment shader: %s\n",
4329 cfg_t
*simd16_cfg
= NULL
;
4330 fs_visitor
v2(brw
, mem_ctx
, MESA_SHADER_FRAGMENT
, key
, &prog_data
->base
,
4331 prog
, &fp
->Base
, 16);
4332 if (likely(!(INTEL_DEBUG
& DEBUG_NO16
) || brw
->use_rep_send
)) {
4333 if (!v
.simd16_unsupported
) {
4334 /* Try a SIMD16 compile */
4335 v2
.import_uniforms(&v
);
4337 perf_debug("SIMD16 shader failed to compile, falling back to "
4338 "SIMD8 at a 10-20%% performance cost: %s", v2
.fail_msg
);
4340 simd16_cfg
= v2
.cfg
;
4343 perf_debug("SIMD16 shader unsupported, falling back to "
4344 "SIMD8 at a 10-20%% performance cost: %s", v
.no16_msg
);
4349 int no_simd8
= (INTEL_DEBUG
& DEBUG_NO8
) || brw
->no_simd8
;
4350 if ((no_simd8
|| brw
->gen
< 5) && simd16_cfg
) {
4352 prog_data
->no_8
= true;
4355 prog_data
->no_8
= false;
4358 fs_generator
g(brw
, mem_ctx
, (void *) key
, &prog_data
->base
,
4359 &fp
->Base
, v
.promoted_constants
, v
.runtime_check_aads_emit
, "FS");
4361 if (unlikely(INTEL_DEBUG
& DEBUG_WM
)) {
4364 name
= ralloc_asprintf(mem_ctx
, "%s fragment shader %d",
4365 prog
->Label
? prog
->Label
: "unnamed",
4368 name
= ralloc_asprintf(mem_ctx
, "fragment program %d", fp
->Base
.Id
);
4370 g
.enable_debug(name
);
4374 g
.generate_code(simd8_cfg
, 8);
4376 prog_data
->prog_offset_16
= g
.generate_code(simd16_cfg
, 16);
4378 if (unlikely(brw
->perf_debug
) && shader
) {
4379 if (shader
->compiled_once
)
4380 brw_wm_debug_recompile(brw
, prog
, key
);
4381 shader
->compiled_once
= true;
4383 if (start_busy
&& !drm_intel_bo_busy(brw
->batch
.last_bo
)) {
4384 perf_debug("FS compile took %.03f ms and stalled the GPU\n",
4385 (get_time() - start_time
) * 1000);
4389 return g
.get_assembly(final_assembly_size
);
4393 brw_fs_precompile(struct gl_context
*ctx
,
4394 struct gl_shader_program
*shader_prog
,
4395 struct gl_program
*prog
)
4397 struct brw_context
*brw
= brw_context(ctx
);
4398 struct brw_wm_prog_key key
;
4400 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*) prog
;
4401 struct brw_fragment_program
*bfp
= brw_fragment_program(fp
);
4402 bool program_uses_dfdy
= fp
->UsesDFdy
;
4404 memset(&key
, 0, sizeof(key
));
4408 key
.iz_lookup
|= IZ_PS_KILL_ALPHATEST_BIT
;
4410 if (fp
->Base
.OutputsWritten
& BITFIELD64_BIT(FRAG_RESULT_DEPTH
))
4411 key
.iz_lookup
|= IZ_PS_COMPUTES_DEPTH_BIT
;
4413 /* Just assume depth testing. */
4414 key
.iz_lookup
|= IZ_DEPTH_TEST_ENABLE_BIT
;
4415 key
.iz_lookup
|= IZ_DEPTH_WRITE_ENABLE_BIT
;
4418 if (brw
->gen
< 6 || _mesa_bitcount_64(fp
->Base
.InputsRead
&
4419 BRW_FS_VARYING_INPUT_MASK
) > 16)
4420 key
.input_slots_valid
= fp
->Base
.InputsRead
| VARYING_BIT_POS
;
4422 brw_setup_tex_for_precompile(brw
, &key
.tex
, &fp
->Base
);
4424 if (fp
->Base
.InputsRead
& VARYING_BIT_POS
) {
4425 key
.drawable_height
= ctx
->DrawBuffer
->Height
;
4428 key
.nr_color_regions
= _mesa_bitcount_64(fp
->Base
.OutputsWritten
&
4429 ~(BITFIELD64_BIT(FRAG_RESULT_DEPTH
) |
4430 BITFIELD64_BIT(FRAG_RESULT_SAMPLE_MASK
)));
4432 if ((fp
->Base
.InputsRead
& VARYING_BIT_POS
) || program_uses_dfdy
) {
4433 key
.render_to_fbo
= _mesa_is_user_fbo(ctx
->DrawBuffer
) ||
4434 key
.nr_color_regions
> 1;
4437 key
.program_string_id
= bfp
->id
;
4439 uint32_t old_prog_offset
= brw
->wm
.base
.prog_offset
;
4440 struct brw_wm_prog_data
*old_prog_data
= brw
->wm
.prog_data
;
4442 bool success
= brw_codegen_wm_prog(brw
, shader_prog
, bfp
, &key
);
4444 brw
->wm
.base
.prog_offset
= old_prog_offset
;
4445 brw
->wm
.prog_data
= old_prog_data
;
4451 brw_setup_tex_for_precompile(struct brw_context
*brw
,
4452 struct brw_sampler_prog_key_data
*tex
,
4453 struct gl_program
*prog
)
4455 const bool has_shader_channel_select
= brw
->is_haswell
|| brw
->gen
>= 8;
4456 unsigned sampler_count
= _mesa_fls(prog
->SamplersUsed
);
4457 for (unsigned i
= 0; i
< sampler_count
; i
++) {
4458 if (!has_shader_channel_select
&& (prog
->ShadowSamplers
& (1 << i
))) {
4459 /* Assume DEPTH_TEXTURE_MODE is the default: X, X, X, 1 */
4461 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_ONE
);
4463 /* Color sampler: assume no swizzling. */
4464 tex
->swizzles
[i
] = SWIZZLE_XYZW
;