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"
53 fs_inst::init(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
54 const fs_reg
*src
, unsigned sources
)
56 memset(this, 0, sizeof(*this));
58 this->src
= new fs_reg
[MAX2(sources
, 3)];
59 for (unsigned i
= 0; i
< sources
; i
++)
60 this->src
[i
] = src
[i
];
62 this->opcode
= opcode
;
64 this->sources
= sources
;
65 this->exec_size
= exec_size
;
67 assert(dst
.file
!= IMM
&& dst
.file
!= UNIFORM
);
69 /* If exec_size == 0, try to guess it from the registers. Since all
70 * manner of things may use hardware registers, we first try to guess
71 * based on GRF registers. If this fails, we will go ahead and take the
72 * width from the destination register.
74 if (this->exec_size
== 0) {
75 if (dst
.file
== GRF
) {
76 this->exec_size
= dst
.width
;
78 for (unsigned i
= 0; i
< sources
; ++i
) {
79 if (src
[i
].file
!= GRF
&& src
[i
].file
!= ATTR
)
82 if (this->exec_size
<= 1)
83 this->exec_size
= src
[i
].width
;
84 assert(src
[i
].width
== 1 || src
[i
].width
== this->exec_size
);
88 if (this->exec_size
== 0 && dst
.file
!= BAD_FILE
)
89 this->exec_size
= dst
.width
;
91 assert(this->exec_size
!= 0);
93 for (unsigned i
= 0; i
< sources
; ++i
) {
94 switch (this->src
[i
].file
) {
96 this->src
[i
].effective_width
= 8;
101 assert(this->src
[i
].width
> 0);
102 if (this->src
[i
].width
== 1) {
103 this->src
[i
].effective_width
= this->exec_size
;
105 this->src
[i
].effective_width
= this->src
[i
].width
;
110 this->src
[i
].effective_width
= this->exec_size
;
113 unreachable("Invalid source register file");
116 this->dst
.effective_width
= this->exec_size
;
118 this->conditional_mod
= BRW_CONDITIONAL_NONE
;
120 /* This will be the case for almost all instructions. */
127 DIV_ROUND_UP(MAX2(dst
.width
* dst
.stride
, 1) * type_sz(dst
.type
), 32);
130 this->regs_written
= 0;
134 unreachable("Invalid destination register file");
136 unreachable("Invalid register file");
139 this->writes_accumulator
= false;
144 init(BRW_OPCODE_NOP
, 8, dst
, NULL
, 0);
147 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
)
149 init(opcode
, exec_size
, reg_undef
, NULL
, 0);
152 fs_inst::fs_inst(enum opcode opcode
, const fs_reg
&dst
)
154 init(opcode
, 0, dst
, NULL
, 0);
157 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
160 const fs_reg src
[1] = { src0
};
161 init(opcode
, exec_size
, dst
, src
, 1);
164 fs_inst::fs_inst(enum opcode opcode
, const fs_reg
&dst
, const fs_reg
&src0
)
166 const fs_reg src
[1] = { src0
};
167 init(opcode
, 0, dst
, src
, 1);
170 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
171 const fs_reg
&src0
, const fs_reg
&src1
)
173 const fs_reg src
[2] = { src0
, src1
};
174 init(opcode
, exec_size
, dst
, src
, 2);
177 fs_inst::fs_inst(enum opcode opcode
, const fs_reg
&dst
, const fs_reg
&src0
,
180 const fs_reg src
[2] = { src0
, src1
};
181 init(opcode
, 0, dst
, src
, 2);
184 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_size
, const fs_reg
&dst
,
185 const fs_reg
&src0
, const fs_reg
&src1
, const fs_reg
&src2
)
187 const fs_reg src
[3] = { src0
, src1
, src2
};
188 init(opcode
, exec_size
, dst
, src
, 3);
191 fs_inst::fs_inst(enum opcode opcode
, const fs_reg
&dst
, const fs_reg
&src0
,
192 const fs_reg
&src1
, const fs_reg
&src2
)
194 const fs_reg src
[3] = { src0
, src1
, src2
};
195 init(opcode
, 0, dst
, src
, 3);
198 fs_inst::fs_inst(enum opcode opcode
, const fs_reg
&dst
,
199 const fs_reg src
[], unsigned sources
)
201 init(opcode
, 0, dst
, src
, sources
);
204 fs_inst::fs_inst(enum opcode opcode
, uint8_t exec_width
, const fs_reg
&dst
,
205 const fs_reg src
[], unsigned sources
)
207 init(opcode
, exec_width
, dst
, src
, sources
);
210 fs_inst::fs_inst(const fs_inst
&that
)
212 memcpy(this, &that
, sizeof(that
));
214 this->src
= new fs_reg
[MAX2(that
.sources
, 3)];
216 for (unsigned i
= 0; i
< that
.sources
; i
++)
217 this->src
[i
] = that
.src
[i
];
226 fs_inst::resize_sources(uint8_t num_sources
)
228 if (this->sources
!= num_sources
) {
229 fs_reg
*src
= new fs_reg
[MAX2(num_sources
, 3)];
231 for (unsigned i
= 0; i
< MIN2(this->sources
, num_sources
); ++i
)
232 src
[i
] = this->src
[i
];
236 this->sources
= num_sources
;
242 fs_visitor::op(const fs_reg &dst, const fs_reg &src0) \
244 return new(mem_ctx) fs_inst(BRW_OPCODE_##op, dst, src0); \
249 fs_visitor::op(const fs_reg &dst, const fs_reg &src0, \
250 const fs_reg &src1) \
252 return new(mem_ctx) fs_inst(BRW_OPCODE_##op, dst, src0, src1); \
255 #define ALU2_ACC(op) \
257 fs_visitor::op(const fs_reg &dst, const fs_reg &src0, \
258 const fs_reg &src1) \
260 fs_inst *inst = new(mem_ctx) fs_inst(BRW_OPCODE_##op, dst, src0, src1);\
261 inst->writes_accumulator = true; \
267 fs_visitor::op(const fs_reg &dst, const fs_reg &src0, \
268 const fs_reg &src1, const fs_reg &src2) \
270 return new(mem_ctx) fs_inst(BRW_OPCODE_##op, dst, src0, src1, src2);\
302 /** Gen4 predicated IF. */
304 fs_visitor::IF(enum brw_predicate predicate
)
306 fs_inst
*inst
= new(mem_ctx
) fs_inst(BRW_OPCODE_IF
, dispatch_width
);
307 inst
->predicate
= predicate
;
311 /** Gen6 IF with embedded comparison. */
313 fs_visitor::IF(const fs_reg
&src0
, const fs_reg
&src1
,
314 enum brw_conditional_mod condition
)
316 assert(brw
->gen
== 6);
317 fs_inst
*inst
= new(mem_ctx
) fs_inst(BRW_OPCODE_IF
, dispatch_width
,
318 reg_null_d
, src0
, src1
);
319 inst
->conditional_mod
= condition
;
324 * CMP: Sets the low bit of the destination channels with the result
325 * of the comparison, while the upper bits are undefined, and updates
326 * the flag register with the packed 16 bits of the result.
329 fs_visitor::CMP(fs_reg dst
, fs_reg src0
, fs_reg src1
,
330 enum brw_conditional_mod condition
)
334 /* Take the instruction:
336 * CMP null<d> src0<f> src1<f>
338 * Original gen4 does type conversion to the destination type before
339 * comparison, producing garbage results for floating point comparisons.
341 * The destination type doesn't matter on newer generations, so we set the
342 * type to match src0 so we can compact the instruction.
344 dst
.type
= src0
.type
;
345 if (dst
.file
== HW_REG
)
346 dst
.fixed_hw_reg
.type
= dst
.type
;
348 resolve_ud_negate(&src0
);
349 resolve_ud_negate(&src1
);
351 inst
= new(mem_ctx
) fs_inst(BRW_OPCODE_CMP
, dst
, src0
, src1
);
352 inst
->conditional_mod
= condition
;
358 fs_visitor::LOAD_PAYLOAD(const fs_reg
&dst
, fs_reg
*src
, int sources
)
360 uint8_t exec_size
= dst
.width
;
361 for (int i
= 0; i
< sources
; ++i
) {
362 assert(src
[i
].width
% dst
.width
== 0);
363 if (src
[i
].width
> exec_size
)
364 exec_size
= src
[i
].width
;
367 fs_inst
*inst
= new(mem_ctx
) fs_inst(SHADER_OPCODE_LOAD_PAYLOAD
, exec_size
,
369 inst
->regs_written
= 0;
370 for (int i
= 0; i
< sources
; ++i
) {
371 /* The LOAD_PAYLOAD instruction only really makes sense if we are
372 * dealing with whole registers. If this ever changes, we can deal
375 int size
= inst
->src
[i
].effective_width
* type_sz(src
[i
].type
);
376 assert(size
% 32 == 0);
377 inst
->regs_written
+= (size
+ 31) / 32;
384 fs_visitor::VARYING_PULL_CONSTANT_LOAD(const fs_reg
&dst
,
385 const fs_reg
&surf_index
,
386 const fs_reg
&varying_offset
,
387 uint32_t const_offset
)
389 exec_list instructions
;
392 /* We have our constant surface use a pitch of 4 bytes, so our index can
393 * be any component of a vector, and then we load 4 contiguous
394 * components starting from that.
396 * We break down the const_offset to a portion added to the variable
397 * offset and a portion done using reg_offset, which means that if you
398 * have GLSL using something like "uniform vec4 a[20]; gl_FragColor =
399 * a[i]", we'll temporarily generate 4 vec4 loads from offset i * 4, and
400 * CSE can later notice that those loads are all the same and eliminate
401 * the redundant ones.
403 fs_reg vec4_offset
= vgrf(glsl_type::int_type
);
404 instructions
.push_tail(ADD(vec4_offset
,
405 varying_offset
, fs_reg(const_offset
& ~3)));
408 if (brw
->gen
== 4 && dst
.width
== 8) {
409 /* Pre-gen5, we can either use a SIMD8 message that requires (header,
410 * u, v, r) as parameters, or we can just use the SIMD16 message
411 * consisting of (header, u). We choose the second, at the cost of a
412 * longer return length.
419 op
= FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7
;
421 op
= FS_OPCODE_VARYING_PULL_CONSTANT_LOAD
;
423 assert(dst
.width
% 8 == 0);
424 int regs_written
= 4 * (dst
.width
/ 8) * scale
;
425 fs_reg vec4_result
= fs_reg(GRF
, alloc
.allocate(regs_written
),
426 dst
.type
, dst
.width
);
427 inst
= new(mem_ctx
) fs_inst(op
, vec4_result
, surf_index
, vec4_offset
);
428 inst
->regs_written
= regs_written
;
429 instructions
.push_tail(inst
);
433 inst
->header_present
= true;
437 inst
->mlen
= 1 + dispatch_width
/ 8;
440 fs_reg result
= offset(vec4_result
, (const_offset
& 3) * scale
);
441 instructions
.push_tail(MOV(dst
, result
));
447 * A helper for MOV generation for fixing up broken hardware SEND dependency
451 fs_visitor::DEP_RESOLVE_MOV(int grf
)
453 fs_inst
*inst
= MOV(brw_null_reg(), fs_reg(GRF
, grf
, BRW_REGISTER_TYPE_F
));
456 inst
->annotation
= "send dependency resolve";
458 /* The caller always wants uncompressed to emit the minimal extra
459 * dependencies, and to avoid having to deal with aligning its regs to 2.
467 fs_inst::equals(fs_inst
*inst
) const
469 return (opcode
== inst
->opcode
&&
470 dst
.equals(inst
->dst
) &&
471 src
[0].equals(inst
->src
[0]) &&
472 src
[1].equals(inst
->src
[1]) &&
473 src
[2].equals(inst
->src
[2]) &&
474 saturate
== inst
->saturate
&&
475 predicate
== inst
->predicate
&&
476 conditional_mod
== inst
->conditional_mod
&&
477 mlen
== inst
->mlen
&&
478 base_mrf
== inst
->base_mrf
&&
479 target
== inst
->target
&&
481 header_present
== inst
->header_present
&&
482 shadow_compare
== inst
->shadow_compare
&&
483 exec_size
== inst
->exec_size
&&
484 offset
== inst
->offset
);
488 fs_inst::overwrites_reg(const fs_reg
®
) const
490 return reg
.in_range(dst
, regs_written
);
494 fs_inst::is_send_from_grf() const
497 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7
:
498 case SHADER_OPCODE_SHADER_TIME_ADD
:
499 case FS_OPCODE_INTERPOLATE_AT_CENTROID
:
500 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
501 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
502 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
503 case SHADER_OPCODE_UNTYPED_ATOMIC
:
504 case SHADER_OPCODE_UNTYPED_SURFACE_READ
:
505 case SHADER_OPCODE_URB_WRITE_SIMD8
:
507 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
508 return src
[1].file
== GRF
;
509 case FS_OPCODE_FB_WRITE
:
510 return src
[0].file
== GRF
;
513 return src
[0].file
== GRF
;
520 fs_inst::can_do_source_mods(struct brw_context
*brw
)
522 if (brw
->gen
== 6 && is_math())
525 if (is_send_from_grf())
528 if (!backend_instruction::can_do_source_mods())
537 memset(this, 0, sizeof(*this));
541 /** Generic unset register constructor. */
545 this->file
= BAD_FILE
;
548 /** Immediate value constructor. */
549 fs_reg::fs_reg(float f
)
553 this->type
= BRW_REGISTER_TYPE_F
;
554 this->fixed_hw_reg
.dw1
.f
= f
;
558 /** Immediate value constructor. */
559 fs_reg::fs_reg(int32_t i
)
563 this->type
= BRW_REGISTER_TYPE_D
;
564 this->fixed_hw_reg
.dw1
.d
= i
;
568 /** Immediate value constructor. */
569 fs_reg::fs_reg(uint32_t u
)
573 this->type
= BRW_REGISTER_TYPE_UD
;
574 this->fixed_hw_reg
.dw1
.ud
= u
;
578 /** Vector float immediate value constructor. */
579 fs_reg::fs_reg(uint8_t vf
[4])
583 this->type
= BRW_REGISTER_TYPE_VF
;
584 memcpy(&this->fixed_hw_reg
.dw1
.ud
, vf
, sizeof(unsigned));
587 /** Vector float immediate value constructor. */
588 fs_reg::fs_reg(uint8_t vf0
, uint8_t vf1
, uint8_t vf2
, uint8_t vf3
)
592 this->type
= BRW_REGISTER_TYPE_VF
;
593 this->fixed_hw_reg
.dw1
.ud
= (vf0
<< 0) |
599 /** Fixed brw_reg. */
600 fs_reg::fs_reg(struct brw_reg fixed_hw_reg
)
604 this->fixed_hw_reg
= fixed_hw_reg
;
605 this->type
= fixed_hw_reg
.type
;
606 this->width
= 1 << fixed_hw_reg
.width
;
610 fs_reg::equals(const fs_reg
&r
) const
612 return (file
== r
.file
&&
614 reg_offset
== r
.reg_offset
&&
615 subreg_offset
== r
.subreg_offset
&&
617 negate
== r
.negate
&&
619 !reladdr
&& !r
.reladdr
&&
620 memcmp(&fixed_hw_reg
, &r
.fixed_hw_reg
, sizeof(fixed_hw_reg
)) == 0 &&
626 fs_reg::set_smear(unsigned subreg
)
628 assert(file
!= HW_REG
&& file
!= IMM
);
629 subreg_offset
= subreg
* type_sz(type
);
635 fs_reg::is_contiguous() const
641 fs_visitor::type_size(const struct glsl_type
*type
)
643 unsigned int size
, i
;
645 switch (type
->base_type
) {
648 case GLSL_TYPE_FLOAT
:
650 return type
->components();
651 case GLSL_TYPE_ARRAY
:
652 return type_size(type
->fields
.array
) * type
->length
;
653 case GLSL_TYPE_STRUCT
:
655 for (i
= 0; i
< type
->length
; i
++) {
656 size
+= type_size(type
->fields
.structure
[i
].type
);
659 case GLSL_TYPE_SAMPLER
:
660 /* Samplers take up no register space, since they're baked in at
664 case GLSL_TYPE_ATOMIC_UINT
:
666 case GLSL_TYPE_IMAGE
:
668 case GLSL_TYPE_ERROR
:
669 case GLSL_TYPE_INTERFACE
:
670 case GLSL_TYPE_DOUBLE
:
671 unreachable("not reached");
678 * Create a MOV to read the timestamp register.
680 * The caller is responsible for emitting the MOV. The return value is
681 * the destination of the MOV, with extra parameters set.
684 fs_visitor::get_timestamp(fs_inst
**out_mov
)
686 assert(brw
->gen
>= 7);
688 fs_reg ts
= fs_reg(retype(brw_vec4_reg(BRW_ARCHITECTURE_REGISTER_FILE
,
691 BRW_REGISTER_TYPE_UD
));
693 fs_reg dst
= fs_reg(GRF
, alloc
.allocate(1), BRW_REGISTER_TYPE_UD
, 4);
695 fs_inst
*mov
= MOV(dst
, ts
);
696 /* We want to read the 3 fields we care about even if it's not enabled in
699 mov
->force_writemask_all
= true;
701 /* The caller wants the low 32 bits of the timestamp. Since it's running
702 * at the GPU clock rate of ~1.2ghz, it will roll over every ~3 seconds,
703 * which is plenty of time for our purposes. It is identical across the
704 * EUs, but since it's tracking GPU core speed it will increment at a
705 * varying rate as render P-states change.
707 * The caller could also check if render P-states have changed (or anything
708 * else that might disrupt timing) by setting smear to 2 and checking if
709 * that field is != 0.
718 fs_visitor::emit_shader_time_begin()
720 current_annotation
= "shader time start";
722 shader_start_time
= get_timestamp(&mov
);
727 fs_visitor::emit_shader_time_end()
729 current_annotation
= "shader time end";
731 enum shader_time_shader_type type
, written_type
, reset_type
;
733 case MESA_SHADER_VERTEX
:
735 written_type
= ST_VS_WRITTEN
;
736 reset_type
= ST_VS_RESET
;
738 case MESA_SHADER_GEOMETRY
:
740 written_type
= ST_GS_WRITTEN
;
741 reset_type
= ST_GS_RESET
;
743 case MESA_SHADER_FRAGMENT
:
744 if (dispatch_width
== 8) {
746 written_type
= ST_FS8_WRITTEN
;
747 reset_type
= ST_FS8_RESET
;
749 assert(dispatch_width
== 16);
751 written_type
= ST_FS16_WRITTEN
;
752 reset_type
= ST_FS16_RESET
;
756 unreachable("fs_visitor::emit_shader_time_end missing code");
759 /* Insert our code just before the final SEND with EOT. */
760 exec_node
*end
= this->instructions
.get_tail();
761 assert(end
&& ((fs_inst
*) end
)->eot
);
764 fs_reg shader_end_time
= get_timestamp(&tm_read
);
765 end
->insert_before(tm_read
);
767 /* Check that there weren't any timestamp reset events (assuming these
768 * were the only two timestamp reads that happened).
770 fs_reg reset
= shader_end_time
;
772 fs_inst
*test
= AND(reg_null_d
, reset
, fs_reg(1u));
773 test
->conditional_mod
= BRW_CONDITIONAL_Z
;
774 test
->force_writemask_all
= true;
775 end
->insert_before(test
);
776 end
->insert_before(IF(BRW_PREDICATE_NORMAL
));
778 fs_reg start
= shader_start_time
;
780 fs_reg diff
= fs_reg(GRF
, alloc
.allocate(1), BRW_REGISTER_TYPE_UD
, 1);
782 fs_inst
*add
= ADD(diff
, start
, shader_end_time
);
783 add
->force_writemask_all
= true;
784 end
->insert_before(add
);
786 /* If there were no instructions between the two timestamp gets, the diff
787 * is 2 cycles. Remove that overhead, so I can forget about that when
788 * trying to determine the time taken for single instructions.
790 add
= ADD(diff
, diff
, fs_reg(-2u));
791 add
->force_writemask_all
= true;
792 end
->insert_before(add
);
794 end
->insert_before(SHADER_TIME_ADD(type
, diff
));
795 end
->insert_before(SHADER_TIME_ADD(written_type
, fs_reg(1u)));
796 end
->insert_before(new(mem_ctx
) fs_inst(BRW_OPCODE_ELSE
, dispatch_width
));
797 end
->insert_before(SHADER_TIME_ADD(reset_type
, fs_reg(1u)));
798 end
->insert_before(new(mem_ctx
) fs_inst(BRW_OPCODE_ENDIF
, dispatch_width
));
802 fs_visitor::SHADER_TIME_ADD(enum shader_time_shader_type type
, fs_reg value
)
804 int shader_time_index
=
805 brw_get_shader_time_index(brw
, shader_prog
, prog
, type
);
806 fs_reg offset
= fs_reg(shader_time_index
* SHADER_TIME_STRIDE
);
809 if (dispatch_width
== 8)
810 payload
= vgrf(glsl_type::uvec2_type
);
812 payload
= vgrf(glsl_type::uint_type
);
814 return new(mem_ctx
) fs_inst(SHADER_OPCODE_SHADER_TIME_ADD
,
815 fs_reg(), payload
, offset
, value
);
819 fs_visitor::vfail(const char *format
, va_list va
)
828 msg
= ralloc_vasprintf(mem_ctx
, format
, va
);
829 msg
= ralloc_asprintf(mem_ctx
, "%s compile failed: %s\n", stage_abbrev
, msg
);
831 this->fail_msg
= msg
;
834 fprintf(stderr
, "%s", msg
);
839 fs_visitor::fail(const char *format
, ...)
843 va_start(va
, format
);
849 * Mark this program as impossible to compile in SIMD16 mode.
851 * During the SIMD8 compile (which happens first), we can detect and flag
852 * things that are unsupported in SIMD16 mode, so the compiler can skip
853 * the SIMD16 compile altogether.
855 * During a SIMD16 compile (if one happens anyway), this just calls fail().
858 fs_visitor::no16(const char *format
, ...)
862 va_start(va
, format
);
864 if (dispatch_width
== 16) {
867 simd16_unsupported
= true;
869 if (brw
->perf_debug
) {
871 ralloc_vasprintf_append(&no16_msg
, format
, va
);
873 no16_msg
= ralloc_vasprintf(mem_ctx
, format
, va
);
881 fs_visitor::emit(enum opcode opcode
)
883 return emit(new(mem_ctx
) fs_inst(opcode
, dispatch_width
));
887 fs_visitor::emit(enum opcode opcode
, const fs_reg
&dst
)
889 return emit(new(mem_ctx
) fs_inst(opcode
, dst
));
893 fs_visitor::emit(enum opcode opcode
, const fs_reg
&dst
, const fs_reg
&src0
)
895 return emit(new(mem_ctx
) fs_inst(opcode
, dst
, src0
));
899 fs_visitor::emit(enum opcode opcode
, const fs_reg
&dst
, const fs_reg
&src0
,
902 return emit(new(mem_ctx
) fs_inst(opcode
, dst
, src0
, src1
));
906 fs_visitor::emit(enum opcode opcode
, const fs_reg
&dst
, const fs_reg
&src0
,
907 const fs_reg
&src1
, const fs_reg
&src2
)
909 return emit(new(mem_ctx
) fs_inst(opcode
, dst
, src0
, src1
, src2
));
913 fs_visitor::emit(enum opcode opcode
, const fs_reg
&dst
,
914 fs_reg src
[], int sources
)
916 return emit(new(mem_ctx
) fs_inst(opcode
, dst
, src
, sources
));
920 * Returns true if the instruction has a flag that means it won't
921 * update an entire destination register.
923 * For example, dead code elimination and live variable analysis want to know
924 * when a write to a variable screens off any preceding values that were in
928 fs_inst::is_partial_write() const
930 return ((this->predicate
&& this->opcode
!= BRW_OPCODE_SEL
) ||
931 (this->dst
.width
* type_sz(this->dst
.type
)) < 32 ||
932 !this->dst
.is_contiguous());
936 fs_inst::regs_read(int arg
) const
938 if (is_tex() && arg
== 0 && src
[0].file
== GRF
) {
940 } else if (opcode
== FS_OPCODE_FB_WRITE
&& arg
== 0) {
942 } else if (opcode
== SHADER_OPCODE_URB_WRITE_SIMD8
&& arg
== 0) {
944 } else if (opcode
== SHADER_OPCODE_UNTYPED_ATOMIC
&& arg
== 0) {
946 } else if (opcode
== SHADER_OPCODE_UNTYPED_SURFACE_READ
&& arg
== 0) {
948 } else if (opcode
== FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
&& arg
== 0) {
952 switch (src
[arg
].file
) {
959 if (src
[arg
].stride
== 0) {
962 int size
= src
[arg
].width
* src
[arg
].stride
* type_sz(src
[arg
].type
);
963 return (size
+ 31) / 32;
966 unreachable("MRF registers are not allowed as sources");
968 unreachable("Invalid register file");
973 fs_inst::reads_flag() const
979 fs_inst::writes_flag() const
981 return (conditional_mod
&& (opcode
!= BRW_OPCODE_SEL
&&
982 opcode
!= BRW_OPCODE_IF
&&
983 opcode
!= BRW_OPCODE_WHILE
)) ||
984 opcode
== FS_OPCODE_MOV_DISPATCH_TO_FLAGS
;
988 * Returns how many MRFs an FS opcode will write over.
990 * Note that this is not the 0 or 1 implied writes in an actual gen
991 * instruction -- the FS opcodes often generate MOVs in addition.
994 fs_visitor::implied_mrf_writes(fs_inst
*inst
)
999 if (inst
->base_mrf
== -1)
1002 switch (inst
->opcode
) {
1003 case SHADER_OPCODE_RCP
:
1004 case SHADER_OPCODE_RSQ
:
1005 case SHADER_OPCODE_SQRT
:
1006 case SHADER_OPCODE_EXP2
:
1007 case SHADER_OPCODE_LOG2
:
1008 case SHADER_OPCODE_SIN
:
1009 case SHADER_OPCODE_COS
:
1010 return 1 * dispatch_width
/ 8;
1011 case SHADER_OPCODE_POW
:
1012 case SHADER_OPCODE_INT_QUOTIENT
:
1013 case SHADER_OPCODE_INT_REMAINDER
:
1014 return 2 * dispatch_width
/ 8;
1015 case SHADER_OPCODE_TEX
:
1017 case SHADER_OPCODE_TXD
:
1018 case SHADER_OPCODE_TXF
:
1019 case SHADER_OPCODE_TXF_CMS
:
1020 case SHADER_OPCODE_TXF_MCS
:
1021 case SHADER_OPCODE_TG4
:
1022 case SHADER_OPCODE_TG4_OFFSET
:
1023 case SHADER_OPCODE_TXL
:
1024 case SHADER_OPCODE_TXS
:
1025 case SHADER_OPCODE_LOD
:
1027 case FS_OPCODE_FB_WRITE
:
1029 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
1030 case SHADER_OPCODE_GEN4_SCRATCH_READ
:
1032 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD
:
1034 case SHADER_OPCODE_GEN4_SCRATCH_WRITE
:
1036 case SHADER_OPCODE_UNTYPED_ATOMIC
:
1037 case SHADER_OPCODE_UNTYPED_SURFACE_READ
:
1038 case SHADER_OPCODE_URB_WRITE_SIMD8
:
1039 case FS_OPCODE_INTERPOLATE_AT_CENTROID
:
1040 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
1041 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
1042 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
1045 unreachable("not reached");
1050 fs_visitor::vgrf(const glsl_type
*const type
)
1052 int reg_width
= dispatch_width
/ 8;
1053 return fs_reg(GRF
, alloc
.allocate(type_size(type
) * reg_width
),
1054 brw_type_for_base_type(type
), dispatch_width
);
1058 fs_visitor::vgrf(int num_components
)
1060 int reg_width
= dispatch_width
/ 8;
1061 return fs_reg(GRF
, alloc
.allocate(num_components
* reg_width
),
1062 BRW_REGISTER_TYPE_F
, dispatch_width
);
1065 /** Fixed HW reg constructor. */
1066 fs_reg::fs_reg(enum register_file file
, int reg
)
1071 this->type
= BRW_REGISTER_TYPE_F
;
1082 /** Fixed HW reg constructor. */
1083 fs_reg::fs_reg(enum register_file file
, int reg
, enum brw_reg_type type
)
1099 /** Fixed HW reg constructor. */
1100 fs_reg::fs_reg(enum register_file file
, int reg
, enum brw_reg_type type
,
1107 this->width
= width
;
1111 fs_visitor::variable_storage(ir_variable
*var
)
1113 return (fs_reg
*)hash_table_find(this->variable_ht
, var
);
1117 import_uniforms_callback(const void *key
,
1121 struct hash_table
*dst_ht
= (struct hash_table
*)closure
;
1122 const fs_reg
*reg
= (const fs_reg
*)data
;
1124 if (reg
->file
!= UNIFORM
)
1127 hash_table_insert(dst_ht
, data
, key
);
1130 /* For SIMD16, we need to follow from the uniform setup of SIMD8 dispatch.
1131 * This brings in those uniform definitions
1134 fs_visitor::import_uniforms(fs_visitor
*v
)
1136 hash_table_call_foreach(v
->variable_ht
,
1137 import_uniforms_callback
,
1139 this->push_constant_loc
= v
->push_constant_loc
;
1140 this->pull_constant_loc
= v
->pull_constant_loc
;
1141 this->uniforms
= v
->uniforms
;
1142 this->param_size
= v
->param_size
;
1145 /* Our support for uniforms is piggy-backed on the struct
1146 * gl_fragment_program, because that's where the values actually
1147 * get stored, rather than in some global gl_shader_program uniform
1151 fs_visitor::setup_uniform_values(ir_variable
*ir
)
1153 int namelen
= strlen(ir
->name
);
1155 /* The data for our (non-builtin) uniforms is stored in a series of
1156 * gl_uniform_driver_storage structs for each subcomponent that
1157 * glGetUniformLocation() could name. We know it's been set up in the same
1158 * order we'd walk the type, so walk the list of storage and find anything
1159 * with our name, or the prefix of a component that starts with our name.
1161 unsigned params_before
= uniforms
;
1162 for (unsigned u
= 0; u
< shader_prog
->NumUserUniformStorage
; u
++) {
1163 struct gl_uniform_storage
*storage
= &shader_prog
->UniformStorage
[u
];
1165 if (strncmp(ir
->name
, storage
->name
, namelen
) != 0 ||
1166 (storage
->name
[namelen
] != 0 &&
1167 storage
->name
[namelen
] != '.' &&
1168 storage
->name
[namelen
] != '[')) {
1172 unsigned slots
= storage
->type
->component_slots();
1173 if (storage
->array_elements
)
1174 slots
*= storage
->array_elements
;
1176 for (unsigned i
= 0; i
< slots
; i
++) {
1177 stage_prog_data
->param
[uniforms
++] = &storage
->storage
[i
];
1181 /* Make sure we actually initialized the right amount of stuff here. */
1182 assert(params_before
+ ir
->type
->component_slots() == uniforms
);
1183 (void)params_before
;
1187 /* Our support for builtin uniforms is even scarier than non-builtin.
1188 * It sits on top of the PROG_STATE_VAR parameters that are
1189 * automatically updated from GL context state.
1192 fs_visitor::setup_builtin_uniform_values(ir_variable
*ir
)
1194 const ir_state_slot
*const slots
= ir
->get_state_slots();
1195 assert(slots
!= NULL
);
1197 for (unsigned int i
= 0; i
< ir
->get_num_state_slots(); i
++) {
1198 /* This state reference has already been setup by ir_to_mesa, but we'll
1199 * get the same index back here.
1201 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
1202 (gl_state_index
*)slots
[i
].tokens
);
1204 /* Add each of the unique swizzles of the element as a parameter.
1205 * This'll end up matching the expected layout of the
1206 * array/matrix/structure we're trying to fill in.
1209 for (unsigned int j
= 0; j
< 4; j
++) {
1210 int swiz
= GET_SWZ(slots
[i
].swizzle
, j
);
1211 if (swiz
== last_swiz
)
1215 stage_prog_data
->param
[uniforms
++] =
1216 &prog
->Parameters
->ParameterValues
[index
][swiz
];
1222 fs_visitor::emit_fragcoord_interpolation(bool pixel_center_integer
,
1223 bool origin_upper_left
)
1225 assert(stage
== MESA_SHADER_FRAGMENT
);
1226 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1227 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::vec4_type
));
1229 bool flip
= !origin_upper_left
^ key
->render_to_fbo
;
1231 /* gl_FragCoord.x */
1232 if (pixel_center_integer
) {
1233 emit(MOV(wpos
, this->pixel_x
));
1235 emit(ADD(wpos
, this->pixel_x
, fs_reg(0.5f
)));
1237 wpos
= offset(wpos
, 1);
1239 /* gl_FragCoord.y */
1240 if (!flip
&& pixel_center_integer
) {
1241 emit(MOV(wpos
, this->pixel_y
));
1243 fs_reg pixel_y
= this->pixel_y
;
1244 float offset
= (pixel_center_integer
? 0.0 : 0.5);
1247 pixel_y
.negate
= true;
1248 offset
+= key
->drawable_height
- 1.0;
1251 emit(ADD(wpos
, pixel_y
, fs_reg(offset
)));
1253 wpos
= offset(wpos
, 1);
1255 /* gl_FragCoord.z */
1256 if (brw
->gen
>= 6) {
1257 emit(MOV(wpos
, fs_reg(brw_vec8_grf(payload
.source_depth_reg
, 0))));
1259 emit(FS_OPCODE_LINTERP
, wpos
,
1260 this->delta_x
[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
],
1261 this->delta_y
[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
],
1262 interp_reg(VARYING_SLOT_POS
, 2));
1264 wpos
= offset(wpos
, 1);
1266 /* gl_FragCoord.w: Already set up in emit_interpolation */
1267 emit(BRW_OPCODE_MOV
, wpos
, this->wpos_w
);
1273 fs_visitor::emit_linterp(const fs_reg
&attr
, const fs_reg
&interp
,
1274 glsl_interp_qualifier interpolation_mode
,
1275 bool is_centroid
, bool is_sample
)
1277 brw_wm_barycentric_interp_mode barycoord_mode
;
1278 if (brw
->gen
>= 6) {
1280 if (interpolation_mode
== INTERP_QUALIFIER_SMOOTH
)
1281 barycoord_mode
= BRW_WM_PERSPECTIVE_CENTROID_BARYCENTRIC
;
1283 barycoord_mode
= BRW_WM_NONPERSPECTIVE_CENTROID_BARYCENTRIC
;
1284 } else if (is_sample
) {
1285 if (interpolation_mode
== INTERP_QUALIFIER_SMOOTH
)
1286 barycoord_mode
= BRW_WM_PERSPECTIVE_SAMPLE_BARYCENTRIC
;
1288 barycoord_mode
= BRW_WM_NONPERSPECTIVE_SAMPLE_BARYCENTRIC
;
1290 if (interpolation_mode
== INTERP_QUALIFIER_SMOOTH
)
1291 barycoord_mode
= BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
;
1293 barycoord_mode
= BRW_WM_NONPERSPECTIVE_PIXEL_BARYCENTRIC
;
1296 /* On Ironlake and below, there is only one interpolation mode.
1297 * Centroid interpolation doesn't mean anything on this hardware --
1298 * there is no multisampling.
1300 barycoord_mode
= BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
;
1302 return emit(FS_OPCODE_LINTERP
, attr
,
1303 this->delta_x
[barycoord_mode
],
1304 this->delta_y
[barycoord_mode
], interp
);
1308 fs_visitor::emit_general_interpolation(fs_reg attr
, const char *name
,
1309 const glsl_type
*type
,
1310 glsl_interp_qualifier interpolation_mode
,
1311 int location
, bool mod_centroid
,
1314 attr
.type
= brw_type_for_base_type(type
->get_scalar_type());
1316 assert(stage
== MESA_SHADER_FRAGMENT
);
1317 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
1318 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1320 unsigned int array_elements
;
1322 if (type
->is_array()) {
1323 array_elements
= type
->length
;
1324 if (array_elements
== 0) {
1325 fail("dereferenced array '%s' has length 0\n", name
);
1327 type
= type
->fields
.array
;
1332 if (interpolation_mode
== INTERP_QUALIFIER_NONE
) {
1334 location
== VARYING_SLOT_COL0
|| location
== VARYING_SLOT_COL1
;
1335 if (key
->flat_shade
&& is_gl_Color
) {
1336 interpolation_mode
= INTERP_QUALIFIER_FLAT
;
1338 interpolation_mode
= INTERP_QUALIFIER_SMOOTH
;
1342 for (unsigned int i
= 0; i
< array_elements
; i
++) {
1343 for (unsigned int j
= 0; j
< type
->matrix_columns
; j
++) {
1344 if (prog_data
->urb_setup
[location
] == -1) {
1345 /* If there's no incoming setup data for this slot, don't
1346 * emit interpolation for it.
1348 attr
= offset(attr
, type
->vector_elements
);
1353 if (interpolation_mode
== INTERP_QUALIFIER_FLAT
) {
1354 /* Constant interpolation (flat shading) case. The SF has
1355 * handed us defined values in only the constant offset
1356 * field of the setup reg.
1358 for (unsigned int k
= 0; k
< type
->vector_elements
; k
++) {
1359 struct brw_reg interp
= interp_reg(location
, k
);
1360 interp
= suboffset(interp
, 3);
1361 interp
.type
= attr
.type
;
1362 emit(FS_OPCODE_CINTERP
, attr
, fs_reg(interp
));
1363 attr
= offset(attr
, 1);
1366 /* Smooth/noperspective interpolation case. */
1367 for (unsigned int k
= 0; k
< type
->vector_elements
; k
++) {
1368 struct brw_reg interp
= interp_reg(location
, k
);
1369 if (brw
->needs_unlit_centroid_workaround
&& mod_centroid
) {
1370 /* Get the pixel/sample mask into f0 so that we know
1371 * which pixels are lit. Then, for each channel that is
1372 * unlit, replace the centroid data with non-centroid
1375 emit(FS_OPCODE_MOV_DISPATCH_TO_FLAGS
);
1378 inst
= emit_linterp(attr
, fs_reg(interp
), interpolation_mode
,
1380 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1381 inst
->predicate_inverse
= true;
1383 inst
->no_dd_clear
= true;
1385 inst
= emit_linterp(attr
, fs_reg(interp
), interpolation_mode
,
1386 mod_centroid
&& !key
->persample_shading
,
1387 mod_sample
|| key
->persample_shading
);
1388 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1389 inst
->predicate_inverse
= false;
1391 inst
->no_dd_check
= true;
1394 emit_linterp(attr
, fs_reg(interp
), interpolation_mode
,
1395 mod_centroid
&& !key
->persample_shading
,
1396 mod_sample
|| key
->persample_shading
);
1398 if (brw
->gen
< 6 && interpolation_mode
== INTERP_QUALIFIER_SMOOTH
) {
1399 emit(BRW_OPCODE_MUL
, attr
, attr
, this->pixel_w
);
1401 attr
= offset(attr
, 1);
1411 fs_visitor::emit_frontfacing_interpolation()
1413 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::bool_type
));
1415 if (brw
->gen
>= 6) {
1416 /* Bit 15 of g0.0 is 0 if the polygon is front facing. We want to create
1417 * a boolean result from this (~0/true or 0/false).
1419 * We can use the fact that bit 15 is the MSB of g0.0:W to accomplish
1420 * this task in only one instruction:
1421 * - a negation source modifier will flip the bit; and
1422 * - a W -> D type conversion will sign extend the bit into the high
1423 * word of the destination.
1425 * An ASR 15 fills the low word of the destination.
1427 fs_reg g0
= fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_W
));
1430 emit(ASR(*reg
, g0
, fs_reg(15)));
1432 /* Bit 31 of g1.6 is 0 if the polygon is front facing. We want to create
1433 * a boolean result from this (1/true or 0/false).
1435 * Like in the above case, since the bit is the MSB of g1.6:UD we can use
1436 * the negation source modifier to flip it. Unfortunately the SHR
1437 * instruction only operates on UD (or D with an abs source modifier)
1438 * sources without negation.
1440 * Instead, use ASR (which will give ~0/true or 0/false).
1442 fs_reg g1_6
= fs_reg(retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_D
));
1445 emit(ASR(*reg
, g1_6
, fs_reg(31)));
1452 fs_visitor::compute_sample_position(fs_reg dst
, fs_reg int_sample_pos
)
1454 assert(stage
== MESA_SHADER_FRAGMENT
);
1455 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1456 assert(dst
.type
== BRW_REGISTER_TYPE_F
);
1458 if (key
->compute_pos_offset
) {
1459 /* Convert int_sample_pos to floating point */
1460 emit(MOV(dst
, int_sample_pos
));
1461 /* Scale to the range [0, 1] */
1462 emit(MUL(dst
, dst
, fs_reg(1 / 16.0f
)));
1465 /* From ARB_sample_shading specification:
1466 * "When rendering to a non-multisample buffer, or if multisample
1467 * rasterization is disabled, gl_SamplePosition will always be
1470 emit(MOV(dst
, fs_reg(0.5f
)));
1475 fs_visitor::emit_samplepos_setup()
1477 assert(brw
->gen
>= 6);
1479 this->current_annotation
= "compute sample position";
1480 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::vec2_type
));
1482 fs_reg int_sample_x
= vgrf(glsl_type::int_type
);
1483 fs_reg int_sample_y
= vgrf(glsl_type::int_type
);
1485 /* WM will be run in MSDISPMODE_PERSAMPLE. So, only one of SIMD8 or SIMD16
1486 * mode will be enabled.
1488 * From the Ivy Bridge PRM, volume 2 part 1, page 344:
1489 * R31.1:0 Position Offset X/Y for Slot[3:0]
1490 * R31.3:2 Position Offset X/Y for Slot[7:4]
1493 * The X, Y sample positions come in as bytes in thread payload. So, read
1494 * the positions using vstride=16, width=8, hstride=2.
1496 struct brw_reg sample_pos_reg
=
1497 stride(retype(brw_vec1_grf(payload
.sample_pos_reg
, 0),
1498 BRW_REGISTER_TYPE_B
), 16, 8, 2);
1500 if (dispatch_width
== 8) {
1501 emit(MOV(int_sample_x
, fs_reg(sample_pos_reg
)));
1503 emit(MOV(half(int_sample_x
, 0), fs_reg(sample_pos_reg
)));
1504 emit(MOV(half(int_sample_x
, 1), fs_reg(suboffset(sample_pos_reg
, 16))))
1505 ->force_sechalf
= true;
1507 /* Compute gl_SamplePosition.x */
1508 compute_sample_position(pos
, int_sample_x
);
1509 pos
= offset(pos
, 1);
1510 if (dispatch_width
== 8) {
1511 emit(MOV(int_sample_y
, fs_reg(suboffset(sample_pos_reg
, 1))));
1513 emit(MOV(half(int_sample_y
, 0),
1514 fs_reg(suboffset(sample_pos_reg
, 1))));
1515 emit(MOV(half(int_sample_y
, 1), fs_reg(suboffset(sample_pos_reg
, 17))))
1516 ->force_sechalf
= true;
1518 /* Compute gl_SamplePosition.y */
1519 compute_sample_position(pos
, int_sample_y
);
1524 fs_visitor::emit_sampleid_setup()
1526 assert(stage
== MESA_SHADER_FRAGMENT
);
1527 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1528 assert(brw
->gen
>= 6);
1530 this->current_annotation
= "compute sample id";
1531 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(vgrf(glsl_type::int_type
));
1533 if (key
->compute_sample_id
) {
1534 fs_reg t1
= vgrf(glsl_type::int_type
);
1535 fs_reg t2
= vgrf(glsl_type::int_type
);
1536 t2
.type
= BRW_REGISTER_TYPE_UW
;
1538 /* The PS will be run in MSDISPMODE_PERSAMPLE. For example with
1539 * 8x multisampling, subspan 0 will represent sample N (where N
1540 * is 0, 2, 4 or 6), subspan 1 will represent sample 1, 3, 5 or
1541 * 7. We can find the value of N by looking at R0.0 bits 7:6
1542 * ("Starting Sample Pair Index (SSPI)") and multiplying by two
1543 * (since samples are always delivered in pairs). That is, we
1544 * compute 2*((R0.0 & 0xc0) >> 6) == (R0.0 & 0xc0) >> 5. Then
1545 * we need to add N to the sequence (0, 0, 0, 0, 1, 1, 1, 1) in
1546 * case of SIMD8 and sequence (0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2,
1547 * 2, 3, 3, 3, 3) in case of SIMD16. We compute this sequence by
1548 * populating a temporary variable with the sequence (0, 1, 2, 3),
1549 * and then reading from it using vstride=1, width=4, hstride=0.
1550 * These computations hold good for 4x multisampling as well.
1552 * For 2x MSAA and SIMD16, we want to use the sequence (0, 1, 0, 1):
1553 * the first four slots are sample 0 of subspan 0; the next four
1554 * are sample 1 of subspan 0; the third group is sample 0 of
1555 * subspan 1, and finally sample 1 of subspan 1.
1558 inst
= emit(BRW_OPCODE_AND
, t1
,
1559 fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
)),
1561 inst
->force_writemask_all
= true;
1562 inst
= emit(BRW_OPCODE_SHR
, t1
, t1
, fs_reg(5));
1563 inst
->force_writemask_all
= true;
1564 /* This works for both SIMD8 and SIMD16 */
1565 inst
= emit(MOV(t2
, brw_imm_v(key
->persample_2x
? 0x1010 : 0x3210)));
1566 inst
->force_writemask_all
= true;
1567 /* This special instruction takes care of setting vstride=1,
1568 * width=4, hstride=0 of t2 during an ADD instruction.
1570 emit(FS_OPCODE_SET_SAMPLE_ID
, *reg
, t1
, t2
);
1572 /* As per GL_ARB_sample_shading specification:
1573 * "When rendering to a non-multisample buffer, or if multisample
1574 * rasterization is disabled, gl_SampleID will always be zero."
1576 emit(BRW_OPCODE_MOV
, *reg
, fs_reg(0));
1583 fs_visitor::resolve_source_modifiers(fs_reg
*src
)
1585 if (!src
->abs
&& !src
->negate
)
1588 fs_reg temp
= retype(vgrf(1), src
->type
);
1589 emit(MOV(temp
, *src
));
1594 fs_visitor::fix_math_operand(fs_reg src
)
1596 /* Can't do hstride == 0 args on gen6 math, so expand it out. We
1597 * might be able to do better by doing execsize = 1 math and then
1598 * expanding that result out, but we would need to be careful with
1601 * The hardware ignores source modifiers (negate and abs) on math
1602 * instructions, so we also move to a temp to set those up.
1604 if (brw
->gen
== 6 && src
.file
!= UNIFORM
&& src
.file
!= IMM
&&
1605 !src
.abs
&& !src
.negate
)
1608 /* Gen7 relaxes most of the above restrictions, but still can't use IMM
1611 if (brw
->gen
>= 7 && src
.file
!= IMM
)
1614 fs_reg expanded
= vgrf(glsl_type::float_type
);
1615 expanded
.type
= src
.type
;
1616 emit(BRW_OPCODE_MOV
, expanded
, src
);
1621 fs_visitor::emit_math(enum opcode opcode
, fs_reg dst
, fs_reg src
)
1624 case SHADER_OPCODE_RCP
:
1625 case SHADER_OPCODE_RSQ
:
1626 case SHADER_OPCODE_SQRT
:
1627 case SHADER_OPCODE_EXP2
:
1628 case SHADER_OPCODE_LOG2
:
1629 case SHADER_OPCODE_SIN
:
1630 case SHADER_OPCODE_COS
:
1633 unreachable("not reached: bad math opcode");
1636 /* Can't do hstride == 0 args to gen6 math, so expand it out. We
1637 * might be able to do better by doing execsize = 1 math and then
1638 * expanding that result out, but we would need to be careful with
1641 * Gen 6 hardware ignores source modifiers (negate and abs) on math
1642 * instructions, so we also move to a temp to set those up.
1644 if (brw
->gen
== 6 || brw
->gen
== 7)
1645 src
= fix_math_operand(src
);
1647 fs_inst
*inst
= emit(opcode
, dst
, src
);
1651 inst
->mlen
= dispatch_width
/ 8;
1658 fs_visitor::emit_math(enum opcode opcode
, fs_reg dst
, fs_reg src0
, fs_reg src1
)
1663 if (brw
->gen
>= 8) {
1664 inst
= emit(opcode
, dst
, src0
, src1
);
1665 } else if (brw
->gen
>= 6) {
1666 src0
= fix_math_operand(src0
);
1667 src1
= fix_math_operand(src1
);
1669 inst
= emit(opcode
, dst
, src0
, src1
);
1671 /* From the Ironlake PRM, Volume 4, Part 1, Section 6.1.13
1672 * "Message Payload":
1674 * "Operand0[7]. For the INT DIV functions, this operand is the
1677 * "Operand1[7]. For the INT DIV functions, this operand is the
1680 bool is_int_div
= opcode
!= SHADER_OPCODE_POW
;
1681 fs_reg
&op0
= is_int_div
? src1
: src0
;
1682 fs_reg
&op1
= is_int_div
? src0
: src1
;
1684 emit(MOV(fs_reg(MRF
, base_mrf
+ 1, op1
.type
, dispatch_width
), op1
));
1685 inst
= emit(opcode
, dst
, op0
, reg_null_f
);
1687 inst
->base_mrf
= base_mrf
;
1688 inst
->mlen
= 2 * dispatch_width
/ 8;
1694 fs_visitor::emit_discard_jump()
1696 /* For performance, after a discard, jump to the end of the
1697 * shader if all relevant channels have been discarded.
1699 fs_inst
*discard_jump
= emit(FS_OPCODE_DISCARD_JUMP
);
1700 discard_jump
->flag_subreg
= 1;
1702 discard_jump
->predicate
= (dispatch_width
== 8)
1703 ? BRW_PREDICATE_ALIGN1_ANY8H
1704 : BRW_PREDICATE_ALIGN1_ANY16H
;
1705 discard_jump
->predicate_inverse
= true;
1709 fs_visitor::assign_curb_setup()
1711 if (dispatch_width
== 8) {
1712 prog_data
->dispatch_grf_start_reg
= payload
.num_regs
;
1714 assert(stage
== MESA_SHADER_FRAGMENT
);
1715 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
1716 prog_data
->dispatch_grf_start_reg_16
= payload
.num_regs
;
1719 prog_data
->curb_read_length
= ALIGN(stage_prog_data
->nr_params
, 8) / 8;
1721 /* Map the offsets in the UNIFORM file to fixed HW regs. */
1722 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1723 for (unsigned int i
= 0; i
< inst
->sources
; i
++) {
1724 if (inst
->src
[i
].file
== UNIFORM
) {
1725 int uniform_nr
= inst
->src
[i
].reg
+ inst
->src
[i
].reg_offset
;
1727 if (uniform_nr
>= 0 && uniform_nr
< (int) uniforms
) {
1728 constant_nr
= push_constant_loc
[uniform_nr
];
1730 /* Section 5.11 of the OpenGL 4.1 spec says:
1731 * "Out-of-bounds reads return undefined values, which include
1732 * values from other variables of the active program or zero."
1733 * Just return the first push constant.
1738 struct brw_reg brw_reg
= brw_vec1_grf(payload
.num_regs
+
1742 inst
->src
[i
].file
= HW_REG
;
1743 inst
->src
[i
].fixed_hw_reg
= byte_offset(
1744 retype(brw_reg
, inst
->src
[i
].type
),
1745 inst
->src
[i
].subreg_offset
);
1752 fs_visitor::calculate_urb_setup()
1754 assert(stage
== MESA_SHADER_FRAGMENT
);
1755 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
1756 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
1758 memset(prog_data
->urb_setup
, -1,
1759 sizeof(prog_data
->urb_setup
[0]) * VARYING_SLOT_MAX
);
1762 /* Figure out where each of the incoming setup attributes lands. */
1763 if (brw
->gen
>= 6) {
1764 if (_mesa_bitcount_64(prog
->InputsRead
&
1765 BRW_FS_VARYING_INPUT_MASK
) <= 16) {
1766 /* The SF/SBE pipeline stage can do arbitrary rearrangement of the
1767 * first 16 varying inputs, so we can put them wherever we want.
1768 * Just put them in order.
1770 * This is useful because it means that (a) inputs not used by the
1771 * fragment shader won't take up valuable register space, and (b) we
1772 * won't have to recompile the fragment shader if it gets paired with
1773 * a different vertex (or geometry) shader.
1775 for (unsigned int i
= 0; i
< VARYING_SLOT_MAX
; i
++) {
1776 if (prog
->InputsRead
& BRW_FS_VARYING_INPUT_MASK
&
1777 BITFIELD64_BIT(i
)) {
1778 prog_data
->urb_setup
[i
] = urb_next
++;
1782 /* We have enough input varyings that the SF/SBE pipeline stage can't
1783 * arbitrarily rearrange them to suit our whim; we have to put them
1784 * in an order that matches the output of the previous pipeline stage
1785 * (geometry or vertex shader).
1787 struct brw_vue_map prev_stage_vue_map
;
1788 brw_compute_vue_map(brw
, &prev_stage_vue_map
,
1789 key
->input_slots_valid
);
1790 int first_slot
= 2 * BRW_SF_URB_ENTRY_READ_OFFSET
;
1791 assert(prev_stage_vue_map
.num_slots
<= first_slot
+ 32);
1792 for (int slot
= first_slot
; slot
< prev_stage_vue_map
.num_slots
;
1794 int varying
= prev_stage_vue_map
.slot_to_varying
[slot
];
1795 /* Note that varying == BRW_VARYING_SLOT_COUNT when a slot is
1798 if (varying
!= BRW_VARYING_SLOT_COUNT
&&
1799 (prog
->InputsRead
& BRW_FS_VARYING_INPUT_MASK
&
1800 BITFIELD64_BIT(varying
))) {
1801 prog_data
->urb_setup
[varying
] = slot
- first_slot
;
1804 urb_next
= prev_stage_vue_map
.num_slots
- first_slot
;
1807 /* FINISHME: The sf doesn't map VS->FS inputs for us very well. */
1808 for (unsigned int i
= 0; i
< VARYING_SLOT_MAX
; i
++) {
1809 /* Point size is packed into the header, not as a general attribute */
1810 if (i
== VARYING_SLOT_PSIZ
)
1813 if (key
->input_slots_valid
& BITFIELD64_BIT(i
)) {
1814 /* The back color slot is skipped when the front color is
1815 * also written to. In addition, some slots can be
1816 * written in the vertex shader and not read in the
1817 * fragment shader. So the register number must always be
1818 * incremented, mapped or not.
1820 if (_mesa_varying_slot_in_fs((gl_varying_slot
) i
))
1821 prog_data
->urb_setup
[i
] = urb_next
;
1827 * It's a FS only attribute, and we did interpolation for this attribute
1828 * in SF thread. So, count it here, too.
1830 * See compile_sf_prog() for more info.
1832 if (prog
->InputsRead
& BITFIELD64_BIT(VARYING_SLOT_PNTC
))
1833 prog_data
->urb_setup
[VARYING_SLOT_PNTC
] = urb_next
++;
1836 prog_data
->num_varying_inputs
= urb_next
;
1840 fs_visitor::assign_urb_setup()
1842 assert(stage
== MESA_SHADER_FRAGMENT
);
1843 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
1845 int urb_start
= payload
.num_regs
+ prog_data
->base
.curb_read_length
;
1847 /* Offset all the urb_setup[] index by the actual position of the
1848 * setup regs, now that the location of the constants has been chosen.
1850 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1851 if (inst
->opcode
== FS_OPCODE_LINTERP
) {
1852 assert(inst
->src
[2].file
== HW_REG
);
1853 inst
->src
[2].fixed_hw_reg
.nr
+= urb_start
;
1856 if (inst
->opcode
== FS_OPCODE_CINTERP
) {
1857 assert(inst
->src
[0].file
== HW_REG
);
1858 inst
->src
[0].fixed_hw_reg
.nr
+= urb_start
;
1862 /* Each attribute is 4 setup channels, each of which is half a reg. */
1863 this->first_non_payload_grf
=
1864 urb_start
+ prog_data
->num_varying_inputs
* 2;
1868 fs_visitor::assign_vs_urb_setup()
1870 brw_vs_prog_data
*vs_prog_data
= (brw_vs_prog_data
*) prog_data
;
1871 int grf
, count
, slot
, channel
, attr
;
1873 assert(stage
== MESA_SHADER_VERTEX
);
1874 count
= _mesa_bitcount_64(vs_prog_data
->inputs_read
);
1875 if (vs_prog_data
->uses_vertexid
|| vs_prog_data
->uses_instanceid
)
1878 /* Each attribute is 4 regs. */
1879 this->first_non_payload_grf
=
1880 payload
.num_regs
+ prog_data
->curb_read_length
+ count
* 4;
1882 unsigned vue_entries
=
1883 MAX2(count
, vs_prog_data
->base
.vue_map
.num_slots
);
1885 vs_prog_data
->base
.urb_entry_size
= ALIGN(vue_entries
, 4) / 4;
1886 vs_prog_data
->base
.urb_read_length
= (count
+ 1) / 2;
1888 assert(vs_prog_data
->base
.urb_read_length
<= 15);
1890 /* Rewrite all ATTR file references to the hw grf that they land in. */
1891 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1892 for (int i
= 0; i
< inst
->sources
; i
++) {
1893 if (inst
->src
[i
].file
== ATTR
) {
1895 if (inst
->src
[i
].reg
== VERT_ATTRIB_MAX
) {
1898 /* Attributes come in in a contiguous block, ordered by their
1899 * gl_vert_attrib value. That means we can compute the slot
1900 * number for an attribute by masking out the enabled
1901 * attributes before it and counting the bits.
1903 attr
= inst
->src
[i
].reg
+ inst
->src
[i
].reg_offset
/ 4;
1904 slot
= _mesa_bitcount_64(vs_prog_data
->inputs_read
&
1905 BITFIELD64_MASK(attr
));
1908 channel
= inst
->src
[i
].reg_offset
& 3;
1910 grf
= payload
.num_regs
+
1911 prog_data
->curb_read_length
+
1914 inst
->src
[i
].file
= HW_REG
;
1915 inst
->src
[i
].fixed_hw_reg
=
1916 retype(brw_vec8_grf(grf
, 0), inst
->src
[i
].type
);
1923 * Split large virtual GRFs into separate components if we can.
1925 * This is mostly duplicated with what brw_fs_vector_splitting does,
1926 * but that's really conservative because it's afraid of doing
1927 * splitting that doesn't result in real progress after the rest of
1928 * the optimization phases, which would cause infinite looping in
1929 * optimization. We can do it once here, safely. This also has the
1930 * opportunity to split interpolated values, or maybe even uniforms,
1931 * which we don't have at the IR level.
1933 * We want to split, because virtual GRFs are what we register
1934 * allocate and spill (due to contiguousness requirements for some
1935 * instructions), and they're what we naturally generate in the
1936 * codegen process, but most virtual GRFs don't actually need to be
1937 * contiguous sets of GRFs. If we split, we'll end up with reduced
1938 * live intervals and better dead code elimination and coalescing.
1941 fs_visitor::split_virtual_grfs()
1943 int num_vars
= this->alloc
.count
;
1945 /* Count the total number of registers */
1947 int vgrf_to_reg
[num_vars
];
1948 for (int i
= 0; i
< num_vars
; i
++) {
1949 vgrf_to_reg
[i
] = reg_count
;
1950 reg_count
+= alloc
.sizes
[i
];
1953 /* An array of "split points". For each register slot, this indicates
1954 * if this slot can be separated from the previous slot. Every time an
1955 * instruction uses multiple elements of a register (as a source or
1956 * destination), we mark the used slots as inseparable. Then we go
1957 * through and split the registers into the smallest pieces we can.
1959 bool split_points
[reg_count
];
1960 memset(split_points
, 0, sizeof(split_points
));
1962 /* Mark all used registers as fully splittable */
1963 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1964 if (inst
->dst
.file
== GRF
) {
1965 int reg
= vgrf_to_reg
[inst
->dst
.reg
];
1966 for (unsigned j
= 1; j
< this->alloc
.sizes
[inst
->dst
.reg
]; j
++)
1967 split_points
[reg
+ j
] = true;
1970 for (int i
= 0; i
< inst
->sources
; i
++) {
1971 if (inst
->src
[i
].file
== GRF
) {
1972 int reg
= vgrf_to_reg
[inst
->src
[i
].reg
];
1973 for (unsigned j
= 1; j
< this->alloc
.sizes
[inst
->src
[i
].reg
]; j
++)
1974 split_points
[reg
+ j
] = true;
1980 this->delta_x
[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
].file
== GRF
) {
1981 /* PLN opcodes rely on the delta_xy being contiguous. We only have to
1982 * check this for BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC, because prior to
1983 * Gen6, that was the only supported interpolation mode, and since Gen6,
1984 * delta_x and delta_y are in fixed hardware registers.
1986 int vgrf
= this->delta_x
[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
].reg
;
1987 split_points
[vgrf_to_reg
[vgrf
] + 1] = false;
1990 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
1991 if (inst
->dst
.file
== GRF
) {
1992 int reg
= vgrf_to_reg
[inst
->dst
.reg
] + inst
->dst
.reg_offset
;
1993 for (int j
= 1; j
< inst
->regs_written
; j
++)
1994 split_points
[reg
+ j
] = false;
1996 for (int i
= 0; i
< inst
->sources
; i
++) {
1997 if (inst
->src
[i
].file
== GRF
) {
1998 int reg
= vgrf_to_reg
[inst
->src
[i
].reg
] + inst
->src
[i
].reg_offset
;
1999 for (int j
= 1; j
< inst
->regs_read(i
); j
++)
2000 split_points
[reg
+ j
] = false;
2005 int new_virtual_grf
[reg_count
];
2006 int new_reg_offset
[reg_count
];
2009 for (int i
= 0; i
< num_vars
; i
++) {
2010 /* The first one should always be 0 as a quick sanity check. */
2011 assert(split_points
[reg
] == false);
2014 new_reg_offset
[reg
] = 0;
2019 for (unsigned j
= 1; j
< alloc
.sizes
[i
]; j
++) {
2020 /* If this is a split point, reset the offset to 0 and allocate a
2021 * new virtual GRF for the previous offset many registers
2023 if (split_points
[reg
]) {
2024 assert(offset
<= MAX_VGRF_SIZE
);
2025 int grf
= alloc
.allocate(offset
);
2026 for (int k
= reg
- offset
; k
< reg
; k
++)
2027 new_virtual_grf
[k
] = grf
;
2030 new_reg_offset
[reg
] = offset
;
2035 /* The last one gets the original register number */
2036 assert(offset
<= MAX_VGRF_SIZE
);
2037 alloc
.sizes
[i
] = offset
;
2038 for (int k
= reg
- offset
; k
< reg
; k
++)
2039 new_virtual_grf
[k
] = i
;
2041 assert(reg
== reg_count
);
2043 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2044 if (inst
->dst
.file
== GRF
) {
2045 reg
= vgrf_to_reg
[inst
->dst
.reg
] + inst
->dst
.reg_offset
;
2046 inst
->dst
.reg
= new_virtual_grf
[reg
];
2047 inst
->dst
.reg_offset
= new_reg_offset
[reg
];
2048 assert((unsigned)new_reg_offset
[reg
] < alloc
.sizes
[new_virtual_grf
[reg
]]);
2050 for (int i
= 0; i
< inst
->sources
; i
++) {
2051 if (inst
->src
[i
].file
== GRF
) {
2052 reg
= vgrf_to_reg
[inst
->src
[i
].reg
] + inst
->src
[i
].reg_offset
;
2053 inst
->src
[i
].reg
= new_virtual_grf
[reg
];
2054 inst
->src
[i
].reg_offset
= new_reg_offset
[reg
];
2055 assert((unsigned)new_reg_offset
[reg
] < alloc
.sizes
[new_virtual_grf
[reg
]]);
2059 invalidate_live_intervals();
2063 * Remove unused virtual GRFs and compact the virtual_grf_* arrays.
2065 * During code generation, we create tons of temporary variables, many of
2066 * which get immediately killed and are never used again. Yet, in later
2067 * optimization and analysis passes, such as compute_live_intervals, we need
2068 * to loop over all the virtual GRFs. Compacting them can save a lot of
2072 fs_visitor::compact_virtual_grfs()
2074 bool progress
= false;
2075 int remap_table
[this->alloc
.count
];
2076 memset(remap_table
, -1, sizeof(remap_table
));
2078 /* Mark which virtual GRFs are used. */
2079 foreach_block_and_inst(block
, const fs_inst
, inst
, cfg
) {
2080 if (inst
->dst
.file
== GRF
)
2081 remap_table
[inst
->dst
.reg
] = 0;
2083 for (int i
= 0; i
< inst
->sources
; i
++) {
2084 if (inst
->src
[i
].file
== GRF
)
2085 remap_table
[inst
->src
[i
].reg
] = 0;
2089 /* Compact the GRF arrays. */
2091 for (unsigned i
= 0; i
< this->alloc
.count
; i
++) {
2092 if (remap_table
[i
] == -1) {
2093 /* We just found an unused register. This means that we are
2094 * actually going to compact something.
2098 remap_table
[i
] = new_index
;
2099 alloc
.sizes
[new_index
] = alloc
.sizes
[i
];
2100 invalidate_live_intervals();
2105 this->alloc
.count
= new_index
;
2107 /* Patch all the instructions to use the newly renumbered registers */
2108 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2109 if (inst
->dst
.file
== GRF
)
2110 inst
->dst
.reg
= remap_table
[inst
->dst
.reg
];
2112 for (int i
= 0; i
< inst
->sources
; i
++) {
2113 if (inst
->src
[i
].file
== GRF
)
2114 inst
->src
[i
].reg
= remap_table
[inst
->src
[i
].reg
];
2118 /* Patch all the references to delta_x/delta_y, since they're used in
2119 * register allocation. If they're unused, switch them to BAD_FILE so
2120 * we don't think some random VGRF is delta_x/delta_y.
2122 for (unsigned i
= 0; i
< ARRAY_SIZE(delta_x
); i
++) {
2123 if (delta_x
[i
].file
== GRF
) {
2124 if (remap_table
[delta_x
[i
].reg
] != -1) {
2125 delta_x
[i
].reg
= remap_table
[delta_x
[i
].reg
];
2127 delta_x
[i
].file
= BAD_FILE
;
2131 for (unsigned i
= 0; i
< ARRAY_SIZE(delta_y
); i
++) {
2132 if (delta_y
[i
].file
== GRF
) {
2133 if (remap_table
[delta_y
[i
].reg
] != -1) {
2134 delta_y
[i
].reg
= remap_table
[delta_y
[i
].reg
];
2136 delta_y
[i
].file
= BAD_FILE
;
2145 * Implements array access of uniforms by inserting a
2146 * PULL_CONSTANT_LOAD instruction.
2148 * Unlike temporary GRF array access (where we don't support it due to
2149 * the difficulty of doing relative addressing on instruction
2150 * destinations), we could potentially do array access of uniforms
2151 * that were loaded in GRF space as push constants. In real-world
2152 * usage we've seen, though, the arrays being used are always larger
2153 * than we could load as push constants, so just always move all
2154 * uniform array access out to a pull constant buffer.
2157 fs_visitor::move_uniform_array_access_to_pull_constants()
2159 if (dispatch_width
!= 8)
2162 pull_constant_loc
= ralloc_array(mem_ctx
, int, uniforms
);
2163 memset(pull_constant_loc
, -1, sizeof(pull_constant_loc
[0]) * uniforms
);
2165 /* Walk through and find array access of uniforms. Put a copy of that
2166 * uniform in the pull constant buffer.
2168 * Note that we don't move constant-indexed accesses to arrays. No
2169 * testing has been done of the performance impact of this choice.
2171 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
2172 for (int i
= 0 ; i
< inst
->sources
; i
++) {
2173 if (inst
->src
[i
].file
!= UNIFORM
|| !inst
->src
[i
].reladdr
)
2176 int uniform
= inst
->src
[i
].reg
;
2178 /* If this array isn't already present in the pull constant buffer,
2181 if (pull_constant_loc
[uniform
] == -1) {
2182 const gl_constant_value
**values
= &stage_prog_data
->param
[uniform
];
2184 assert(param_size
[uniform
]);
2186 for (int j
= 0; j
< param_size
[uniform
]; j
++) {
2187 pull_constant_loc
[uniform
+ j
] = stage_prog_data
->nr_pull_params
;
2189 stage_prog_data
->pull_param
[stage_prog_data
->nr_pull_params
++] =
2198 * Assign UNIFORM file registers to either push constants or pull constants.
2200 * We allow a fragment shader to have more than the specified minimum
2201 * maximum number of fragment shader uniform components (64). If
2202 * there are too many of these, they'd fill up all of register space.
2203 * So, this will push some of them out to the pull constant buffer and
2204 * update the program to load them.
2207 fs_visitor::assign_constant_locations()
2209 /* Only the first compile (SIMD8 mode) gets to decide on locations. */
2210 if (dispatch_width
!= 8)
2213 /* Find which UNIFORM registers are still in use. */
2214 bool is_live
[uniforms
];
2215 for (unsigned int i
= 0; i
< uniforms
; i
++) {
2219 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2220 for (int i
= 0; i
< inst
->sources
; i
++) {
2221 if (inst
->src
[i
].file
!= UNIFORM
)
2224 int constant_nr
= inst
->src
[i
].reg
+ inst
->src
[i
].reg_offset
;
2225 if (constant_nr
>= 0 && constant_nr
< (int) uniforms
)
2226 is_live
[constant_nr
] = true;
2230 /* Only allow 16 registers (128 uniform components) as push constants.
2232 * Just demote the end of the list. We could probably do better
2233 * here, demoting things that are rarely used in the program first.
2235 * If changing this value, note the limitation about total_regs in
2238 unsigned int max_push_components
= 16 * 8;
2239 unsigned int num_push_constants
= 0;
2241 push_constant_loc
= ralloc_array(mem_ctx
, int, uniforms
);
2243 for (unsigned int i
= 0; i
< uniforms
; i
++) {
2244 if (!is_live
[i
] || pull_constant_loc
[i
] != -1) {
2245 /* This UNIFORM register is either dead, or has already been demoted
2246 * to a pull const. Mark it as no longer living in the param[] array.
2248 push_constant_loc
[i
] = -1;
2252 if (num_push_constants
< max_push_components
) {
2253 /* Retain as a push constant. Record the location in the params[]
2256 push_constant_loc
[i
] = num_push_constants
++;
2258 /* Demote to a pull constant. */
2259 push_constant_loc
[i
] = -1;
2261 int pull_index
= stage_prog_data
->nr_pull_params
++;
2262 stage_prog_data
->pull_param
[pull_index
] = stage_prog_data
->param
[i
];
2263 pull_constant_loc
[i
] = pull_index
;
2267 stage_prog_data
->nr_params
= num_push_constants
;
2269 /* Up until now, the param[] array has been indexed by reg + reg_offset
2270 * of UNIFORM registers. Condense it to only contain the uniforms we
2271 * chose to upload as push constants.
2273 for (unsigned int i
= 0; i
< uniforms
; i
++) {
2274 int remapped
= push_constant_loc
[i
];
2279 assert(remapped
<= (int)i
);
2280 stage_prog_data
->param
[remapped
] = stage_prog_data
->param
[i
];
2285 * Replace UNIFORM register file access with either UNIFORM_PULL_CONSTANT_LOAD
2286 * or VARYING_PULL_CONSTANT_LOAD instructions which load values into VGRFs.
2289 fs_visitor::demote_pull_constants()
2291 foreach_block_and_inst (block
, fs_inst
, inst
, cfg
) {
2292 for (int i
= 0; i
< inst
->sources
; i
++) {
2293 if (inst
->src
[i
].file
!= UNIFORM
)
2297 unsigned location
= inst
->src
[i
].reg
+ inst
->src
[i
].reg_offset
;
2298 if (location
>= uniforms
) /* Out of bounds access */
2301 pull_index
= pull_constant_loc
[location
];
2303 if (pull_index
== -1)
2306 /* Set up the annotation tracking for new generated instructions. */
2308 current_annotation
= inst
->annotation
;
2310 fs_reg
surf_index(stage_prog_data
->binding_table
.pull_constants_start
);
2311 fs_reg dst
= vgrf(glsl_type::float_type
);
2313 /* Generate a pull load into dst. */
2314 if (inst
->src
[i
].reladdr
) {
2315 exec_list list
= VARYING_PULL_CONSTANT_LOAD(dst
,
2317 *inst
->src
[i
].reladdr
,
2319 inst
->insert_before(block
, &list
);
2320 inst
->src
[i
].reladdr
= NULL
;
2322 fs_reg offset
= fs_reg((unsigned)(pull_index
* 4) & ~15);
2324 new(mem_ctx
) fs_inst(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
, 8,
2325 dst
, surf_index
, offset
);
2326 inst
->insert_before(block
, pull
);
2327 inst
->src
[i
].set_smear(pull_index
& 3);
2330 /* Rewrite the instruction to use the temporary VGRF. */
2331 inst
->src
[i
].file
= GRF
;
2332 inst
->src
[i
].reg
= dst
.reg
;
2333 inst
->src
[i
].reg_offset
= 0;
2334 inst
->src
[i
].width
= dispatch_width
;
2337 invalidate_live_intervals();
2341 fs_visitor::opt_algebraic()
2343 bool progress
= false;
2345 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2346 switch (inst
->opcode
) {
2347 case BRW_OPCODE_MOV
:
2348 if (inst
->src
[0].file
!= IMM
)
2351 if (inst
->saturate
) {
2352 if (inst
->dst
.type
!= inst
->src
[0].type
)
2353 assert(!"unimplemented: saturate mixed types");
2355 if (brw_saturate_immediate(inst
->dst
.type
,
2356 &inst
->src
[0].fixed_hw_reg
)) {
2357 inst
->saturate
= false;
2363 case BRW_OPCODE_MUL
:
2364 if (inst
->src
[1].file
!= IMM
)
2368 if (inst
->src
[1].is_one()) {
2369 inst
->opcode
= BRW_OPCODE_MOV
;
2370 inst
->src
[1] = reg_undef
;
2376 if (inst
->src
[1].is_negative_one()) {
2377 inst
->opcode
= BRW_OPCODE_MOV
;
2378 inst
->src
[0].negate
= !inst
->src
[0].negate
;
2379 inst
->src
[1] = reg_undef
;
2385 if (inst
->src
[1].is_zero()) {
2386 inst
->opcode
= BRW_OPCODE_MOV
;
2387 inst
->src
[0] = inst
->src
[1];
2388 inst
->src
[1] = reg_undef
;
2393 if (inst
->src
[0].file
== IMM
) {
2394 assert(inst
->src
[0].type
== BRW_REGISTER_TYPE_F
);
2395 inst
->opcode
= BRW_OPCODE_MOV
;
2396 inst
->src
[0].fixed_hw_reg
.dw1
.f
*= inst
->src
[1].fixed_hw_reg
.dw1
.f
;
2397 inst
->src
[1] = reg_undef
;
2402 case BRW_OPCODE_ADD
:
2403 if (inst
->src
[1].file
!= IMM
)
2407 if (inst
->src
[1].is_zero()) {
2408 inst
->opcode
= BRW_OPCODE_MOV
;
2409 inst
->src
[1] = reg_undef
;
2414 if (inst
->src
[0].file
== IMM
) {
2415 assert(inst
->src
[0].type
== BRW_REGISTER_TYPE_F
);
2416 inst
->opcode
= BRW_OPCODE_MOV
;
2417 inst
->src
[0].fixed_hw_reg
.dw1
.f
+= inst
->src
[1].fixed_hw_reg
.dw1
.f
;
2418 inst
->src
[1] = reg_undef
;
2424 if (inst
->src
[0].equals(inst
->src
[1])) {
2425 inst
->opcode
= BRW_OPCODE_MOV
;
2426 inst
->src
[1] = reg_undef
;
2431 case BRW_OPCODE_LRP
:
2432 if (inst
->src
[1].equals(inst
->src
[2])) {
2433 inst
->opcode
= BRW_OPCODE_MOV
;
2434 inst
->src
[0] = inst
->src
[1];
2435 inst
->src
[1] = reg_undef
;
2436 inst
->src
[2] = reg_undef
;
2441 case BRW_OPCODE_CMP
:
2442 if (inst
->conditional_mod
== BRW_CONDITIONAL_GE
&&
2444 inst
->src
[0].negate
&&
2445 inst
->src
[1].is_zero()) {
2446 inst
->src
[0].abs
= false;
2447 inst
->src
[0].negate
= false;
2448 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
2453 case BRW_OPCODE_SEL
:
2454 if (inst
->src
[0].equals(inst
->src
[1])) {
2455 inst
->opcode
= BRW_OPCODE_MOV
;
2456 inst
->src
[1] = reg_undef
;
2457 inst
->predicate
= BRW_PREDICATE_NONE
;
2458 inst
->predicate_inverse
= false;
2460 } else if (inst
->saturate
&& inst
->src
[1].file
== IMM
) {
2461 switch (inst
->conditional_mod
) {
2462 case BRW_CONDITIONAL_LE
:
2463 case BRW_CONDITIONAL_L
:
2464 switch (inst
->src
[1].type
) {
2465 case BRW_REGISTER_TYPE_F
:
2466 if (inst
->src
[1].fixed_hw_reg
.dw1
.f
>= 1.0f
) {
2467 inst
->opcode
= BRW_OPCODE_MOV
;
2468 inst
->src
[1] = reg_undef
;
2469 inst
->conditional_mod
= BRW_CONDITIONAL_NONE
;
2477 case BRW_CONDITIONAL_GE
:
2478 case BRW_CONDITIONAL_G
:
2479 switch (inst
->src
[1].type
) {
2480 case BRW_REGISTER_TYPE_F
:
2481 if (inst
->src
[1].fixed_hw_reg
.dw1
.f
<= 0.0f
) {
2482 inst
->opcode
= BRW_OPCODE_MOV
;
2483 inst
->src
[1] = reg_undef
;
2484 inst
->conditional_mod
= BRW_CONDITIONAL_NONE
;
2496 case BRW_OPCODE_MAD
:
2497 if (inst
->src
[1].is_zero() || inst
->src
[2].is_zero()) {
2498 inst
->opcode
= BRW_OPCODE_MOV
;
2499 inst
->src
[1] = reg_undef
;
2500 inst
->src
[2] = reg_undef
;
2502 } else if (inst
->src
[0].is_zero()) {
2503 inst
->opcode
= BRW_OPCODE_MUL
;
2504 inst
->src
[0] = inst
->src
[2];
2505 inst
->src
[2] = reg_undef
;
2507 } else if (inst
->src
[1].is_one()) {
2508 inst
->opcode
= BRW_OPCODE_ADD
;
2509 inst
->src
[1] = inst
->src
[2];
2510 inst
->src
[2] = reg_undef
;
2512 } else if (inst
->src
[2].is_one()) {
2513 inst
->opcode
= BRW_OPCODE_ADD
;
2514 inst
->src
[2] = reg_undef
;
2516 } else if (inst
->src
[1].file
== IMM
&& inst
->src
[2].file
== IMM
) {
2517 inst
->opcode
= BRW_OPCODE_ADD
;
2518 inst
->src
[1].fixed_hw_reg
.dw1
.f
*= inst
->src
[2].fixed_hw_reg
.dw1
.f
;
2519 inst
->src
[2] = reg_undef
;
2523 case SHADER_OPCODE_RCP
: {
2524 fs_inst
*prev
= (fs_inst
*)inst
->prev
;
2525 if (prev
->opcode
== SHADER_OPCODE_SQRT
) {
2526 if (inst
->src
[0].equals(prev
->dst
)) {
2527 inst
->opcode
= SHADER_OPCODE_RSQ
;
2528 inst
->src
[0] = prev
->src
[0];
2538 /* Swap if src[0] is immediate. */
2539 if (progress
&& inst
->is_commutative()) {
2540 if (inst
->src
[0].file
== IMM
) {
2541 fs_reg tmp
= inst
->src
[1];
2542 inst
->src
[1] = inst
->src
[0];
2551 fs_visitor::opt_register_renaming()
2553 bool progress
= false;
2556 int remap
[alloc
.count
];
2557 memset(remap
, -1, sizeof(int) * alloc
.count
);
2559 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
2560 if (inst
->opcode
== BRW_OPCODE_IF
|| inst
->opcode
== BRW_OPCODE_DO
) {
2562 } else if (inst
->opcode
== BRW_OPCODE_ENDIF
||
2563 inst
->opcode
== BRW_OPCODE_WHILE
) {
2567 /* Rewrite instruction sources. */
2568 for (int i
= 0; i
< inst
->sources
; i
++) {
2569 if (inst
->src
[i
].file
== GRF
&&
2570 remap
[inst
->src
[i
].reg
] != -1 &&
2571 remap
[inst
->src
[i
].reg
] != inst
->src
[i
].reg
) {
2572 inst
->src
[i
].reg
= remap
[inst
->src
[i
].reg
];
2577 const int dst
= inst
->dst
.reg
;
2580 inst
->dst
.file
== GRF
&&
2581 alloc
.sizes
[inst
->dst
.reg
] == inst
->dst
.width
/ 8 &&
2582 !inst
->is_partial_write()) {
2583 if (remap
[dst
] == -1) {
2586 remap
[dst
] = alloc
.allocate(inst
->dst
.width
/ 8);
2587 inst
->dst
.reg
= remap
[dst
];
2590 } else if (inst
->dst
.file
== GRF
&&
2592 remap
[dst
] != dst
) {
2593 inst
->dst
.reg
= remap
[dst
];
2599 invalidate_live_intervals();
2601 for (unsigned i
= 0; i
< ARRAY_SIZE(delta_x
); i
++) {
2602 if (delta_x
[i
].file
== GRF
&& remap
[delta_x
[i
].reg
] != -1) {
2603 delta_x
[i
].reg
= remap
[delta_x
[i
].reg
];
2606 for (unsigned i
= 0; i
< ARRAY_SIZE(delta_y
); i
++) {
2607 if (delta_y
[i
].file
== GRF
&& remap
[delta_y
[i
].reg
] != -1) {
2608 delta_y
[i
].reg
= remap
[delta_y
[i
].reg
];
2617 * Remove redundant or useless discard jumps.
2619 * For example, we can eliminate jumps in the following sequence:
2621 * discard-jump (redundant with the next jump)
2622 * discard-jump (useless; jumps to the next instruction)
2626 fs_visitor::opt_redundant_discard_jumps()
2628 bool progress
= false;
2630 bblock_t
*last_bblock
= cfg
->blocks
[cfg
->num_blocks
- 1];
2632 fs_inst
*placeholder_halt
= NULL
;
2633 foreach_inst_in_block_reverse(fs_inst
, inst
, last_bblock
) {
2634 if (inst
->opcode
== FS_OPCODE_PLACEHOLDER_HALT
) {
2635 placeholder_halt
= inst
;
2640 if (!placeholder_halt
)
2643 /* Delete any HALTs immediately before the placeholder halt. */
2644 for (fs_inst
*prev
= (fs_inst
*) placeholder_halt
->prev
;
2645 !prev
->is_head_sentinel() && prev
->opcode
== FS_OPCODE_DISCARD_JUMP
;
2646 prev
= (fs_inst
*) placeholder_halt
->prev
) {
2647 prev
->remove(last_bblock
);
2652 invalidate_live_intervals();
2658 fs_visitor::compute_to_mrf()
2660 bool progress
= false;
2663 /* No MRFs on Gen >= 7. */
2667 calculate_live_intervals();
2669 foreach_block_and_inst_safe(block
, fs_inst
, inst
, cfg
) {
2673 if (inst
->opcode
!= BRW_OPCODE_MOV
||
2674 inst
->is_partial_write() ||
2675 inst
->dst
.file
!= MRF
|| inst
->src
[0].file
!= GRF
||
2676 inst
->dst
.type
!= inst
->src
[0].type
||
2677 inst
->src
[0].abs
|| inst
->src
[0].negate
||
2678 !inst
->src
[0].is_contiguous() ||
2679 inst
->src
[0].subreg_offset
)
2682 /* Work out which hardware MRF registers are written by this
2685 int mrf_low
= inst
->dst
.reg
& ~BRW_MRF_COMPR4
;
2687 if (inst
->dst
.reg
& BRW_MRF_COMPR4
) {
2688 mrf_high
= mrf_low
+ 4;
2689 } else if (inst
->exec_size
== 16) {
2690 mrf_high
= mrf_low
+ 1;
2695 /* Can't compute-to-MRF this GRF if someone else was going to
2698 if (this->virtual_grf_end
[inst
->src
[0].reg
] > ip
)
2701 /* Found a move of a GRF to a MRF. Let's see if we can go
2702 * rewrite the thing that made this GRF to write into the MRF.
2704 foreach_inst_in_block_reverse_starting_from(fs_inst
, scan_inst
, inst
, block
) {
2705 if (scan_inst
->dst
.file
== GRF
&&
2706 scan_inst
->dst
.reg
== inst
->src
[0].reg
) {
2707 /* Found the last thing to write our reg we want to turn
2708 * into a compute-to-MRF.
2711 /* If this one instruction didn't populate all the
2712 * channels, bail. We might be able to rewrite everything
2713 * that writes that reg, but it would require smarter
2714 * tracking to delay the rewriting until complete success.
2716 if (scan_inst
->is_partial_write())
2719 /* Things returning more than one register would need us to
2720 * understand coalescing out more than one MOV at a time.
2722 if (scan_inst
->regs_written
> scan_inst
->dst
.width
/ 8)
2725 /* SEND instructions can't have MRF as a destination. */
2726 if (scan_inst
->mlen
)
2729 if (brw
->gen
== 6) {
2730 /* gen6 math instructions must have the destination be
2731 * GRF, so no compute-to-MRF for them.
2733 if (scan_inst
->is_math()) {
2738 if (scan_inst
->dst
.reg_offset
== inst
->src
[0].reg_offset
) {
2739 /* Found the creator of our MRF's source value. */
2740 scan_inst
->dst
.file
= MRF
;
2741 scan_inst
->dst
.reg
= inst
->dst
.reg
;
2742 scan_inst
->saturate
|= inst
->saturate
;
2743 inst
->remove(block
);
2749 /* We don't handle control flow here. Most computation of
2750 * values that end up in MRFs are shortly before the MRF
2753 if (block
->start() == scan_inst
)
2756 /* You can't read from an MRF, so if someone else reads our
2757 * MRF's source GRF that we wanted to rewrite, that stops us.
2759 bool interfered
= false;
2760 for (int i
= 0; i
< scan_inst
->sources
; i
++) {
2761 if (scan_inst
->src
[i
].file
== GRF
&&
2762 scan_inst
->src
[i
].reg
== inst
->src
[0].reg
&&
2763 scan_inst
->src
[i
].reg_offset
== inst
->src
[0].reg_offset
) {
2770 if (scan_inst
->dst
.file
== MRF
) {
2771 /* If somebody else writes our MRF here, we can't
2772 * compute-to-MRF before that.
2774 int scan_mrf_low
= scan_inst
->dst
.reg
& ~BRW_MRF_COMPR4
;
2777 if (scan_inst
->dst
.reg
& BRW_MRF_COMPR4
) {
2778 scan_mrf_high
= scan_mrf_low
+ 4;
2779 } else if (scan_inst
->exec_size
== 16) {
2780 scan_mrf_high
= scan_mrf_low
+ 1;
2782 scan_mrf_high
= scan_mrf_low
;
2785 if (mrf_low
== scan_mrf_low
||
2786 mrf_low
== scan_mrf_high
||
2787 mrf_high
== scan_mrf_low
||
2788 mrf_high
== scan_mrf_high
) {
2793 if (scan_inst
->mlen
> 0 && scan_inst
->base_mrf
!= -1) {
2794 /* Found a SEND instruction, which means that there are
2795 * live values in MRFs from base_mrf to base_mrf +
2796 * scan_inst->mlen - 1. Don't go pushing our MRF write up
2799 if (mrf_low
>= scan_inst
->base_mrf
&&
2800 mrf_low
< scan_inst
->base_mrf
+ scan_inst
->mlen
) {
2803 if (mrf_high
>= scan_inst
->base_mrf
&&
2804 mrf_high
< scan_inst
->base_mrf
+ scan_inst
->mlen
) {
2812 invalidate_live_intervals();
2818 * Once we've generated code, try to convert normal FS_OPCODE_FB_WRITE
2819 * instructions to FS_OPCODE_REP_FB_WRITE.
2822 fs_visitor::emit_repclear_shader()
2824 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
2826 int color_mrf
= base_mrf
+ 2;
2828 fs_inst
*mov
= emit(MOV(vec4(brw_message_reg(color_mrf
)),
2829 fs_reg(UNIFORM
, 0, BRW_REGISTER_TYPE_F
)));
2830 mov
->force_writemask_all
= true;
2833 if (key
->nr_color_regions
== 1) {
2834 write
= emit(FS_OPCODE_REP_FB_WRITE
);
2835 write
->saturate
= key
->clamp_fragment_color
;
2836 write
->base_mrf
= color_mrf
;
2838 write
->header_present
= false;
2841 assume(key
->nr_color_regions
> 0);
2842 for (int i
= 0; i
< key
->nr_color_regions
; ++i
) {
2843 write
= emit(FS_OPCODE_REP_FB_WRITE
);
2844 write
->saturate
= key
->clamp_fragment_color
;
2845 write
->base_mrf
= base_mrf
;
2847 write
->header_present
= true;
2855 assign_constant_locations();
2856 assign_curb_setup();
2858 /* Now that we have the uniform assigned, go ahead and force it to a vec4. */
2859 assert(mov
->src
[0].file
== HW_REG
);
2860 mov
->src
[0] = brw_vec4_grf(mov
->src
[0].fixed_hw_reg
.nr
, 0);
2864 * Walks through basic blocks, looking for repeated MRF writes and
2865 * removing the later ones.
2868 fs_visitor::remove_duplicate_mrf_writes()
2870 fs_inst
*last_mrf_move
[16];
2871 bool progress
= false;
2873 /* Need to update the MRF tracking for compressed instructions. */
2874 if (dispatch_width
== 16)
2877 memset(last_mrf_move
, 0, sizeof(last_mrf_move
));
2879 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
2880 if (inst
->is_control_flow()) {
2881 memset(last_mrf_move
, 0, sizeof(last_mrf_move
));
2884 if (inst
->opcode
== BRW_OPCODE_MOV
&&
2885 inst
->dst
.file
== MRF
) {
2886 fs_inst
*prev_inst
= last_mrf_move
[inst
->dst
.reg
];
2887 if (prev_inst
&& inst
->equals(prev_inst
)) {
2888 inst
->remove(block
);
2894 /* Clear out the last-write records for MRFs that were overwritten. */
2895 if (inst
->dst
.file
== MRF
) {
2896 last_mrf_move
[inst
->dst
.reg
] = NULL
;
2899 if (inst
->mlen
> 0 && inst
->base_mrf
!= -1) {
2900 /* Found a SEND instruction, which will include two or fewer
2901 * implied MRF writes. We could do better here.
2903 for (int i
= 0; i
< implied_mrf_writes(inst
); i
++) {
2904 last_mrf_move
[inst
->base_mrf
+ i
] = NULL
;
2908 /* Clear out any MRF move records whose sources got overwritten. */
2909 if (inst
->dst
.file
== GRF
) {
2910 for (unsigned int i
= 0; i
< ARRAY_SIZE(last_mrf_move
); i
++) {
2911 if (last_mrf_move
[i
] &&
2912 last_mrf_move
[i
]->src
[0].reg
== inst
->dst
.reg
) {
2913 last_mrf_move
[i
] = NULL
;
2918 if (inst
->opcode
== BRW_OPCODE_MOV
&&
2919 inst
->dst
.file
== MRF
&&
2920 inst
->src
[0].file
== GRF
&&
2921 !inst
->is_partial_write()) {
2922 last_mrf_move
[inst
->dst
.reg
] = inst
;
2927 invalidate_live_intervals();
2933 clear_deps_for_inst_src(fs_inst
*inst
, bool *deps
, int first_grf
, int grf_len
)
2935 /* Clear the flag for registers that actually got read (as expected). */
2936 for (int i
= 0; i
< inst
->sources
; i
++) {
2938 if (inst
->src
[i
].file
== GRF
) {
2939 grf
= inst
->src
[i
].reg
;
2940 } else if (inst
->src
[i
].file
== HW_REG
&&
2941 inst
->src
[i
].fixed_hw_reg
.file
== BRW_GENERAL_REGISTER_FILE
) {
2942 grf
= inst
->src
[i
].fixed_hw_reg
.nr
;
2947 if (grf
>= first_grf
&&
2948 grf
< first_grf
+ grf_len
) {
2949 deps
[grf
- first_grf
] = false;
2950 if (inst
->exec_size
== 16)
2951 deps
[grf
- first_grf
+ 1] = false;
2957 * Implements this workaround for the original 965:
2959 * "[DevBW, DevCL] Implementation Restrictions: As the hardware does not
2960 * check for post destination dependencies on this instruction, software
2961 * must ensure that there is no destination hazard for the case of ‘write
2962 * followed by a posted write’ shown in the following example.
2965 * 2. send r3.xy <rest of send instruction>
2968 * Due to no post-destination dependency check on the ‘send’, the above
2969 * code sequence could have two instructions (1 and 2) in flight at the
2970 * same time that both consider ‘r3’ as the target of their final writes.
2973 fs_visitor::insert_gen4_pre_send_dependency_workarounds(bblock_t
*block
,
2976 int write_len
= inst
->regs_written
;
2977 int first_write_grf
= inst
->dst
.reg
;
2978 bool needs_dep
[BRW_MAX_MRF
];
2979 assert(write_len
< (int)sizeof(needs_dep
) - 1);
2981 memset(needs_dep
, false, sizeof(needs_dep
));
2982 memset(needs_dep
, true, write_len
);
2984 clear_deps_for_inst_src(inst
, needs_dep
, first_write_grf
, write_len
);
2986 /* Walk backwards looking for writes to registers we're writing which
2987 * aren't read since being written. If we hit the start of the program,
2988 * we assume that there are no outstanding dependencies on entry to the
2991 foreach_inst_in_block_reverse_starting_from(fs_inst
, scan_inst
, inst
, block
) {
2992 /* If we hit control flow, assume that there *are* outstanding
2993 * dependencies, and force their cleanup before our instruction.
2995 if (block
->start() == scan_inst
) {
2996 for (int i
= 0; i
< write_len
; i
++) {
2998 inst
->insert_before(block
, DEP_RESOLVE_MOV(first_write_grf
+ i
));
3004 /* We insert our reads as late as possible on the assumption that any
3005 * instruction but a MOV that might have left us an outstanding
3006 * dependency has more latency than a MOV.
3008 if (scan_inst
->dst
.file
== GRF
) {
3009 for (int i
= 0; i
< scan_inst
->regs_written
; i
++) {
3010 int reg
= scan_inst
->dst
.reg
+ i
;
3012 if (reg
>= first_write_grf
&&
3013 reg
< first_write_grf
+ write_len
&&
3014 needs_dep
[reg
- first_write_grf
]) {
3015 inst
->insert_before(block
, DEP_RESOLVE_MOV(reg
));
3016 needs_dep
[reg
- first_write_grf
] = false;
3017 if (scan_inst
->exec_size
== 16)
3018 needs_dep
[reg
- first_write_grf
+ 1] = false;
3023 /* Clear the flag for registers that actually got read (as expected). */
3024 clear_deps_for_inst_src(scan_inst
, needs_dep
, first_write_grf
, write_len
);
3026 /* Continue the loop only if we haven't resolved all the dependencies */
3028 for (i
= 0; i
< write_len
; i
++) {
3038 * Implements this workaround for the original 965:
3040 * "[DevBW, DevCL] Errata: A destination register from a send can not be
3041 * used as a destination register until after it has been sourced by an
3042 * instruction with a different destination register.
3045 fs_visitor::insert_gen4_post_send_dependency_workarounds(bblock_t
*block
, fs_inst
*inst
)
3047 int write_len
= inst
->regs_written
;
3048 int first_write_grf
= inst
->dst
.reg
;
3049 bool needs_dep
[BRW_MAX_MRF
];
3050 assert(write_len
< (int)sizeof(needs_dep
) - 1);
3052 memset(needs_dep
, false, sizeof(needs_dep
));
3053 memset(needs_dep
, true, write_len
);
3054 /* Walk forwards looking for writes to registers we're writing which aren't
3055 * read before being written.
3057 foreach_inst_in_block_starting_from(fs_inst
, scan_inst
, inst
, block
) {
3058 /* If we hit control flow, force resolve all remaining dependencies. */
3059 if (block
->end() == scan_inst
) {
3060 for (int i
= 0; i
< write_len
; i
++) {
3062 scan_inst
->insert_before(block
,
3063 DEP_RESOLVE_MOV(first_write_grf
+ i
));
3068 /* Clear the flag for registers that actually got read (as expected). */
3069 clear_deps_for_inst_src(scan_inst
, needs_dep
, first_write_grf
, write_len
);
3071 /* We insert our reads as late as possible since they're reading the
3072 * result of a SEND, which has massive latency.
3074 if (scan_inst
->dst
.file
== GRF
&&
3075 scan_inst
->dst
.reg
>= first_write_grf
&&
3076 scan_inst
->dst
.reg
< first_write_grf
+ write_len
&&
3077 needs_dep
[scan_inst
->dst
.reg
- first_write_grf
]) {
3078 scan_inst
->insert_before(block
, DEP_RESOLVE_MOV(scan_inst
->dst
.reg
));
3079 needs_dep
[scan_inst
->dst
.reg
- first_write_grf
] = false;
3082 /* Continue the loop only if we haven't resolved all the dependencies */
3084 for (i
= 0; i
< write_len
; i
++) {
3094 fs_visitor::insert_gen4_send_dependency_workarounds()
3096 if (brw
->gen
!= 4 || brw
->is_g4x
)
3099 bool progress
= false;
3101 /* Note that we're done with register allocation, so GRF fs_regs always
3102 * have a .reg_offset of 0.
3105 foreach_block_and_inst(block
, fs_inst
, inst
, cfg
) {
3106 if (inst
->mlen
!= 0 && inst
->dst
.file
== GRF
) {
3107 insert_gen4_pre_send_dependency_workarounds(block
, inst
);
3108 insert_gen4_post_send_dependency_workarounds(block
, inst
);
3114 invalidate_live_intervals();
3118 * Turns the generic expression-style uniform pull constant load instruction
3119 * into a hardware-specific series of instructions for loading a pull
3122 * The expression style allows the CSE pass before this to optimize out
3123 * repeated loads from the same offset, and gives the pre-register-allocation
3124 * scheduling full flexibility, while the conversion to native instructions
3125 * allows the post-register-allocation scheduler the best information
3128 * Note that execution masking for setting up pull constant loads is special:
3129 * the channels that need to be written are unrelated to the current execution
3130 * mask, since a later instruction will use one of the result channels as a
3131 * source operand for all 8 or 16 of its channels.
3134 fs_visitor::lower_uniform_pull_constant_loads()
3136 foreach_block_and_inst (block
, fs_inst
, inst
, cfg
) {
3137 if (inst
->opcode
!= FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
)
3140 if (brw
->gen
>= 7) {
3141 /* The offset arg before was a vec4-aligned byte offset. We need to
3142 * turn it into a dword offset.
3144 fs_reg const_offset_reg
= inst
->src
[1];
3145 assert(const_offset_reg
.file
== IMM
&&
3146 const_offset_reg
.type
== BRW_REGISTER_TYPE_UD
);
3147 const_offset_reg
.fixed_hw_reg
.dw1
.ud
/= 4;
3148 fs_reg payload
= fs_reg(GRF
, alloc
.allocate(1));
3150 /* We have to use a message header on Skylake to get SIMD4x2 mode.
3151 * Reserve space for the register.
3153 if (brw
->gen
>= 9) {
3154 payload
.reg_offset
++;
3155 alloc
.sizes
[payload
.reg
] = 2;
3158 /* This is actually going to be a MOV, but since only the first dword
3159 * is accessed, we have a special opcode to do just that one. Note
3160 * that this needs to be an operation that will be considered a def
3161 * by live variable analysis, or register allocation will explode.
3163 fs_inst
*setup
= new(mem_ctx
) fs_inst(FS_OPCODE_SET_SIMD4X2_OFFSET
,
3164 8, payload
, const_offset_reg
);
3165 setup
->force_writemask_all
= true;
3167 setup
->ir
= inst
->ir
;
3168 setup
->annotation
= inst
->annotation
;
3169 inst
->insert_before(block
, setup
);
3171 /* Similarly, this will only populate the first 4 channels of the
3172 * result register (since we only use smear values from 0-3), but we
3173 * don't tell the optimizer.
3175 inst
->opcode
= FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7
;
3176 inst
->src
[1] = payload
;
3178 invalidate_live_intervals();
3180 /* Before register allocation, we didn't tell the scheduler about the
3181 * MRF we use. We know it's safe to use this MRF because nothing
3182 * else does except for register spill/unspill, which generates and
3183 * uses its MRF within a single IR instruction.
3185 inst
->base_mrf
= 14;
3192 fs_visitor::lower_load_payload()
3194 bool progress
= false;
3196 int vgrf_to_reg
[alloc
.count
];
3198 for (unsigned i
= 0; i
< alloc
.count
; ++i
) {
3199 vgrf_to_reg
[i
] = reg_count
;
3200 reg_count
+= alloc
.sizes
[i
];
3204 bool written
:1; /* Whether this register has ever been written */
3205 bool force_writemask_all
:1;
3206 bool force_sechalf
:1;
3207 } metadata
[reg_count
];
3208 memset(metadata
, 0, sizeof(metadata
));
3210 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
3211 if (inst
->dst
.file
== GRF
) {
3212 const int dst_reg
= vgrf_to_reg
[inst
->dst
.reg
] + inst
->dst
.reg_offset
;
3213 bool force_sechalf
= inst
->force_sechalf
&&
3214 !inst
->force_writemask_all
;
3215 bool toggle_sechalf
= inst
->dst
.width
== 16 &&
3216 type_sz(inst
->dst
.type
) == 4 &&
3217 !inst
->force_writemask_all
;
3218 for (int i
= 0; i
< inst
->regs_written
; ++i
) {
3219 metadata
[dst_reg
+ i
].written
= true;
3220 metadata
[dst_reg
+ i
].force_sechalf
= force_sechalf
;
3221 metadata
[dst_reg
+ i
].force_writemask_all
= inst
->force_writemask_all
;
3222 force_sechalf
= (toggle_sechalf
!= force_sechalf
);
3226 if (inst
->opcode
== SHADER_OPCODE_LOAD_PAYLOAD
) {
3227 assert(inst
->dst
.file
== MRF
|| inst
->dst
.file
== GRF
);
3228 fs_reg dst
= inst
->dst
;
3230 for (int i
= 0; i
< inst
->sources
; i
++) {
3231 dst
.width
= inst
->src
[i
].effective_width
;
3232 dst
.type
= inst
->src
[i
].type
;
3234 if (inst
->src
[i
].file
== BAD_FILE
) {
3235 /* Do nothing but otherwise increment as normal */
3236 } else if (dst
.file
== MRF
&&
3239 i
+ 4 < inst
->sources
&&
3240 inst
->src
[i
+ 4].equals(horiz_offset(inst
->src
[i
], 8))) {
3241 fs_reg compr4_dst
= dst
;
3242 compr4_dst
.reg
+= BRW_MRF_COMPR4
;
3243 compr4_dst
.width
= 16;
3244 fs_reg compr4_src
= inst
->src
[i
];
3245 compr4_src
.width
= 16;
3246 fs_inst
*mov
= MOV(compr4_dst
, compr4_src
);
3247 mov
->force_writemask_all
= true;
3248 inst
->insert_before(block
, mov
);
3249 /* Mark i+4 as BAD_FILE so we don't emit a MOV for it */
3250 inst
->src
[i
+ 4].file
= BAD_FILE
;
3252 fs_inst
*mov
= MOV(dst
, inst
->src
[i
]);
3253 if (inst
->src
[i
].file
== GRF
) {
3254 int src_reg
= vgrf_to_reg
[inst
->src
[i
].reg
] +
3255 inst
->src
[i
].reg_offset
;
3256 mov
->force_sechalf
= metadata
[src_reg
].force_sechalf
;
3257 mov
->force_writemask_all
= metadata
[src_reg
].force_writemask_all
;
3259 /* We don't have any useful metadata for immediates or
3260 * uniforms. Assume that any of the channels of the
3261 * destination may be used.
3263 assert(inst
->src
[i
].file
== IMM
||
3264 inst
->src
[i
].file
== UNIFORM
);
3265 mov
->force_writemask_all
= true;
3268 if (dst
.file
== GRF
) {
3269 const int dst_reg
= vgrf_to_reg
[dst
.reg
] + dst
.reg_offset
;
3270 const bool force_writemask
= mov
->force_writemask_all
;
3271 metadata
[dst_reg
].force_writemask_all
= force_writemask
;
3272 metadata
[dst_reg
].force_sechalf
= mov
->force_sechalf
;
3273 if (dst
.width
* type_sz(dst
.type
) > 32) {
3274 assert(!mov
->force_sechalf
);
3275 metadata
[dst_reg
+ 1].force_writemask_all
= force_writemask
;
3276 metadata
[dst_reg
+ 1].force_sechalf
= !force_writemask
;
3280 inst
->insert_before(block
, mov
);
3283 dst
= offset(dst
, 1);
3286 inst
->remove(block
);
3292 invalidate_live_intervals();
3298 fs_visitor::dump_instructions()
3300 dump_instructions(NULL
);
3304 fs_visitor::dump_instructions(const char *name
)
3306 FILE *file
= stderr
;
3307 if (name
&& geteuid() != 0) {
3308 file
= fopen(name
, "w");
3314 calculate_register_pressure();
3315 int ip
= 0, max_pressure
= 0;
3316 foreach_block_and_inst(block
, backend_instruction
, inst
, cfg
) {
3317 max_pressure
= MAX2(max_pressure
, regs_live_at_ip
[ip
]);
3318 fprintf(file
, "{%3d} %4d: ", regs_live_at_ip
[ip
], ip
);
3319 dump_instruction(inst
, file
);
3322 fprintf(file
, "Maximum %3d registers live at once.\n", max_pressure
);
3325 foreach_in_list(backend_instruction
, inst
, &instructions
) {
3326 fprintf(file
, "%4d: ", ip
++);
3327 dump_instruction(inst
, file
);
3331 if (file
!= stderr
) {
3337 fs_visitor::dump_instruction(backend_instruction
*be_inst
)
3339 dump_instruction(be_inst
, stderr
);
3343 fs_visitor::dump_instruction(backend_instruction
*be_inst
, FILE *file
)
3345 fs_inst
*inst
= (fs_inst
*)be_inst
;
3347 if (inst
->predicate
) {
3348 fprintf(file
, "(%cf0.%d) ",
3349 inst
->predicate_inverse
? '-' : '+',
3353 fprintf(file
, "%s", brw_instruction_name(inst
->opcode
));
3355 fprintf(file
, ".sat");
3356 if (inst
->conditional_mod
) {
3357 fprintf(file
, "%s", conditional_modifier
[inst
->conditional_mod
]);
3358 if (!inst
->predicate
&&
3359 (brw
->gen
< 5 || (inst
->opcode
!= BRW_OPCODE_SEL
&&
3360 inst
->opcode
!= BRW_OPCODE_IF
&&
3361 inst
->opcode
!= BRW_OPCODE_WHILE
))) {
3362 fprintf(file
, ".f0.%d", inst
->flag_subreg
);
3365 fprintf(file
, "(%d) ", inst
->exec_size
);
3368 switch (inst
->dst
.file
) {
3370 fprintf(file
, "vgrf%d", inst
->dst
.reg
);
3371 if (inst
->dst
.width
!= dispatch_width
)
3372 fprintf(file
, "@%d", inst
->dst
.width
);
3373 if (alloc
.sizes
[inst
->dst
.reg
] != inst
->dst
.width
/ 8 ||
3374 inst
->dst
.subreg_offset
)
3375 fprintf(file
, "+%d.%d",
3376 inst
->dst
.reg_offset
, inst
->dst
.subreg_offset
);
3379 fprintf(file
, "m%d", inst
->dst
.reg
);
3382 fprintf(file
, "(null)");
3385 fprintf(file
, "***u%d***", inst
->dst
.reg
+ inst
->dst
.reg_offset
);
3388 fprintf(file
, "***attr%d***", inst
->dst
.reg
+ inst
->dst
.reg_offset
);
3391 if (inst
->dst
.fixed_hw_reg
.file
== BRW_ARCHITECTURE_REGISTER_FILE
) {
3392 switch (inst
->dst
.fixed_hw_reg
.nr
) {
3394 fprintf(file
, "null");
3396 case BRW_ARF_ADDRESS
:
3397 fprintf(file
, "a0.%d", inst
->dst
.fixed_hw_reg
.subnr
);
3399 case BRW_ARF_ACCUMULATOR
:
3400 fprintf(file
, "acc%d", inst
->dst
.fixed_hw_reg
.subnr
);
3403 fprintf(file
, "f%d.%d", inst
->dst
.fixed_hw_reg
.nr
& 0xf,
3404 inst
->dst
.fixed_hw_reg
.subnr
);
3407 fprintf(file
, "arf%d.%d", inst
->dst
.fixed_hw_reg
.nr
& 0xf,
3408 inst
->dst
.fixed_hw_reg
.subnr
);
3412 fprintf(file
, "hw_reg%d", inst
->dst
.fixed_hw_reg
.nr
);
3414 if (inst
->dst
.fixed_hw_reg
.subnr
)
3415 fprintf(file
, "+%d", inst
->dst
.fixed_hw_reg
.subnr
);
3418 fprintf(file
, "???");
3421 fprintf(file
, ":%s, ", brw_reg_type_letters(inst
->dst
.type
));
3423 for (int i
= 0; i
< inst
->sources
; i
++) {
3424 if (inst
->src
[i
].negate
)
3426 if (inst
->src
[i
].abs
)
3428 switch (inst
->src
[i
].file
) {
3430 fprintf(file
, "vgrf%d", inst
->src
[i
].reg
);
3431 if (inst
->src
[i
].width
!= dispatch_width
)
3432 fprintf(file
, "@%d", inst
->src
[i
].width
);
3433 if (alloc
.sizes
[inst
->src
[i
].reg
] != inst
->src
[i
].width
/ 8 ||
3434 inst
->src
[i
].subreg_offset
)
3435 fprintf(file
, "+%d.%d", inst
->src
[i
].reg_offset
,
3436 inst
->src
[i
].subreg_offset
);
3439 fprintf(file
, "***m%d***", inst
->src
[i
].reg
);
3442 fprintf(file
, "attr%d", inst
->src
[i
].reg
+ inst
->src
[i
].reg_offset
);
3445 fprintf(file
, "u%d", inst
->src
[i
].reg
+ inst
->src
[i
].reg_offset
);
3446 if (inst
->src
[i
].reladdr
) {
3447 fprintf(file
, "+reladdr");
3448 } else if (inst
->src
[i
].subreg_offset
) {
3449 fprintf(file
, "+%d.%d", inst
->src
[i
].reg_offset
,
3450 inst
->src
[i
].subreg_offset
);
3454 fprintf(file
, "(null)");
3457 switch (inst
->src
[i
].type
) {
3458 case BRW_REGISTER_TYPE_F
:
3459 fprintf(file
, "%ff", inst
->src
[i
].fixed_hw_reg
.dw1
.f
);
3461 case BRW_REGISTER_TYPE_W
:
3462 case BRW_REGISTER_TYPE_D
:
3463 fprintf(file
, "%dd", inst
->src
[i
].fixed_hw_reg
.dw1
.d
);
3465 case BRW_REGISTER_TYPE_UW
:
3466 case BRW_REGISTER_TYPE_UD
:
3467 fprintf(file
, "%uu", inst
->src
[i
].fixed_hw_reg
.dw1
.ud
);
3469 case BRW_REGISTER_TYPE_VF
:
3470 fprintf(file
, "[%-gF, %-gF, %-gF, %-gF]",
3471 brw_vf_to_float((inst
->src
[i
].fixed_hw_reg
.dw1
.ud
>> 0) & 0xff),
3472 brw_vf_to_float((inst
->src
[i
].fixed_hw_reg
.dw1
.ud
>> 8) & 0xff),
3473 brw_vf_to_float((inst
->src
[i
].fixed_hw_reg
.dw1
.ud
>> 16) & 0xff),
3474 brw_vf_to_float((inst
->src
[i
].fixed_hw_reg
.dw1
.ud
>> 24) & 0xff));
3477 fprintf(file
, "???");
3482 if (inst
->src
[i
].fixed_hw_reg
.negate
)
3484 if (inst
->src
[i
].fixed_hw_reg
.abs
)
3486 if (inst
->src
[i
].fixed_hw_reg
.file
== BRW_ARCHITECTURE_REGISTER_FILE
) {
3487 switch (inst
->src
[i
].fixed_hw_reg
.nr
) {
3489 fprintf(file
, "null");
3491 case BRW_ARF_ADDRESS
:
3492 fprintf(file
, "a0.%d", inst
->src
[i
].fixed_hw_reg
.subnr
);
3494 case BRW_ARF_ACCUMULATOR
:
3495 fprintf(file
, "acc%d", inst
->src
[i
].fixed_hw_reg
.subnr
);
3498 fprintf(file
, "f%d.%d", inst
->src
[i
].fixed_hw_reg
.nr
& 0xf,
3499 inst
->src
[i
].fixed_hw_reg
.subnr
);
3502 fprintf(file
, "arf%d.%d", inst
->src
[i
].fixed_hw_reg
.nr
& 0xf,
3503 inst
->src
[i
].fixed_hw_reg
.subnr
);
3507 fprintf(file
, "hw_reg%d", inst
->src
[i
].fixed_hw_reg
.nr
);
3509 if (inst
->src
[i
].fixed_hw_reg
.subnr
)
3510 fprintf(file
, "+%d", inst
->src
[i
].fixed_hw_reg
.subnr
);
3511 if (inst
->src
[i
].fixed_hw_reg
.abs
)
3515 fprintf(file
, "???");
3518 if (inst
->src
[i
].abs
)
3521 if (inst
->src
[i
].file
!= IMM
) {
3522 fprintf(file
, ":%s", brw_reg_type_letters(inst
->src
[i
].type
));
3525 if (i
< inst
->sources
- 1 && inst
->src
[i
+ 1].file
!= BAD_FILE
)
3526 fprintf(file
, ", ");
3531 if (dispatch_width
== 16 && inst
->exec_size
== 8) {
3532 if (inst
->force_sechalf
)
3533 fprintf(file
, "2ndhalf ");
3535 fprintf(file
, "1sthalf ");
3538 fprintf(file
, "\n");
3542 * Possibly returns an instruction that set up @param reg.
3544 * Sometimes we want to take the result of some expression/variable
3545 * dereference tree and rewrite the instruction generating the result
3546 * of the tree. When processing the tree, we know that the
3547 * instructions generated are all writing temporaries that are dead
3548 * outside of this tree. So, if we have some instructions that write
3549 * a temporary, we're free to point that temp write somewhere else.
3551 * Note that this doesn't guarantee that the instruction generated
3552 * only reg -- it might be the size=4 destination of a texture instruction.
3555 fs_visitor::get_instruction_generating_reg(fs_inst
*start
,
3560 end
->is_partial_write() ||
3562 !reg
.equals(end
->dst
)) {
3570 fs_visitor::setup_payload_gen6()
3573 (prog
->InputsRead
& (1 << VARYING_SLOT_POS
)) != 0;
3574 unsigned barycentric_interp_modes
=
3575 (stage
== MESA_SHADER_FRAGMENT
) ?
3576 ((brw_wm_prog_data
*) this->prog_data
)->barycentric_interp_modes
: 0;
3578 assert(brw
->gen
>= 6);
3580 /* R0-1: masks, pixel X/Y coordinates. */
3581 payload
.num_regs
= 2;
3582 /* R2: only for 32-pixel dispatch.*/
3584 /* R3-26: barycentric interpolation coordinates. These appear in the
3585 * same order that they appear in the brw_wm_barycentric_interp_mode
3586 * enum. Each set of coordinates occupies 2 registers if dispatch width
3587 * == 8 and 4 registers if dispatch width == 16. Coordinates only
3588 * appear if they were enabled using the "Barycentric Interpolation
3589 * Mode" bits in WM_STATE.
3591 for (int i
= 0; i
< BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT
; ++i
) {
3592 if (barycentric_interp_modes
& (1 << i
)) {
3593 payload
.barycentric_coord_reg
[i
] = payload
.num_regs
;
3594 payload
.num_regs
+= 2;
3595 if (dispatch_width
== 16) {
3596 payload
.num_regs
+= 2;
3601 /* R27: interpolated depth if uses source depth */
3603 payload
.source_depth_reg
= payload
.num_regs
;
3605 if (dispatch_width
== 16) {
3606 /* R28: interpolated depth if not SIMD8. */
3610 /* R29: interpolated W set if GEN6_WM_USES_SOURCE_W. */
3612 payload
.source_w_reg
= payload
.num_regs
;
3614 if (dispatch_width
== 16) {
3615 /* R30: interpolated W if not SIMD8. */
3620 if (stage
== MESA_SHADER_FRAGMENT
) {
3621 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
3622 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
3623 prog_data
->uses_pos_offset
= key
->compute_pos_offset
;
3624 /* R31: MSAA position offsets. */
3625 if (prog_data
->uses_pos_offset
) {
3626 payload
.sample_pos_reg
= payload
.num_regs
;
3631 /* R32: MSAA input coverage mask */
3632 if (prog
->SystemValuesRead
& SYSTEM_BIT_SAMPLE_MASK_IN
) {
3633 assert(brw
->gen
>= 7);
3634 payload
.sample_mask_in_reg
= payload
.num_regs
;
3636 if (dispatch_width
== 16) {
3637 /* R33: input coverage mask if not SIMD8. */
3642 /* R34-: bary for 32-pixel. */
3643 /* R58-59: interp W for 32-pixel. */
3645 if (prog
->OutputsWritten
& BITFIELD64_BIT(FRAG_RESULT_DEPTH
)) {
3646 source_depth_to_render_target
= true;
3651 fs_visitor::setup_vs_payload()
3653 /* R0: thread header, R1: urb handles */
3654 payload
.num_regs
= 2;
3658 fs_visitor::assign_binding_table_offsets()
3660 assert(stage
== MESA_SHADER_FRAGMENT
);
3661 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
3662 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
3663 uint32_t next_binding_table_offset
= 0;
3665 /* If there are no color regions, we still perform an FB write to a null
3666 * renderbuffer, which we place at surface index 0.
3668 prog_data
->binding_table
.render_target_start
= next_binding_table_offset
;
3669 next_binding_table_offset
+= MAX2(key
->nr_color_regions
, 1);
3671 assign_common_binding_table_offsets(next_binding_table_offset
);
3675 fs_visitor::calculate_register_pressure()
3677 invalidate_live_intervals();
3678 calculate_live_intervals();
3680 unsigned num_instructions
= 0;
3681 foreach_block(block
, cfg
)
3682 num_instructions
+= block
->instructions
.length();
3684 regs_live_at_ip
= rzalloc_array(mem_ctx
, int, num_instructions
);
3686 for (unsigned reg
= 0; reg
< alloc
.count
; reg
++) {
3687 for (int ip
= virtual_grf_start
[reg
]; ip
<= virtual_grf_end
[reg
]; ip
++)
3688 regs_live_at_ip
[ip
] += alloc
.sizes
[reg
];
3693 fs_visitor::optimize()
3695 const char *stage_name
= stage
== MESA_SHADER_VERTEX
? "vs" : "fs";
3697 split_virtual_grfs();
3699 move_uniform_array_access_to_pull_constants();
3700 assign_constant_locations();
3701 demote_pull_constants();
3703 #define OPT(pass, args...) ({ \
3705 bool this_progress = pass(args); \
3707 if (unlikely(INTEL_DEBUG & DEBUG_OPTIMIZER) && this_progress) { \
3708 char filename[64]; \
3709 snprintf(filename, 64, "%s%d-%04d-%02d-%02d-" #pass, \
3710 stage_name, dispatch_width, shader_prog ? shader_prog->Name : 0, iteration, pass_num); \
3712 backend_visitor::dump_instructions(filename); \
3715 progress = progress || this_progress; \
3719 if (unlikely(INTEL_DEBUG
& DEBUG_OPTIMIZER
)) {
3721 snprintf(filename
, 64, "%s%d-%04d-00-start",
3722 stage_name
, dispatch_width
, shader_prog
? shader_prog
->Name
: 0);
3724 backend_visitor::dump_instructions(filename
);
3735 OPT(remove_duplicate_mrf_writes
);
3739 OPT(opt_copy_propagate
);
3740 OPT(opt_peephole_predicated_break
);
3741 OPT(opt_cmod_propagation
);
3742 OPT(dead_code_eliminate
);
3743 OPT(opt_peephole_sel
);
3744 OPT(dead_control_flow_eliminate
, this);
3745 OPT(opt_register_renaming
);
3746 OPT(opt_redundant_discard_jumps
);
3747 OPT(opt_saturate_propagation
);
3748 OPT(register_coalesce
);
3749 OPT(compute_to_mrf
);
3751 OPT(compact_virtual_grfs
);
3756 if (OPT(lower_load_payload
)) {
3757 split_virtual_grfs();
3758 OPT(register_coalesce
);
3759 OPT(compute_to_mrf
);
3760 OPT(dead_code_eliminate
);
3763 OPT(opt_combine_constants
);
3765 lower_uniform_pull_constant_loads();
3769 * Three source instruction must have a GRF/MRF destination register.
3770 * ARF NULL is not allowed. Fix that up by allocating a temporary GRF.
3773 fs_visitor::fixup_3src_null_dest()
3775 foreach_block_and_inst_safe (block
, fs_inst
, inst
, cfg
) {
3776 if (inst
->is_3src() && inst
->dst
.is_null()) {
3777 inst
->dst
= fs_reg(GRF
, alloc
.allocate(dispatch_width
/ 8),
3784 fs_visitor::allocate_registers()
3786 bool allocated_without_spills
;
3788 static const enum instruction_scheduler_mode pre_modes
[] = {
3790 SCHEDULE_PRE_NON_LIFO
,
3794 /* Try each scheduling heuristic to see if it can successfully register
3795 * allocate without spilling. They should be ordered by decreasing
3796 * performance but increasing likelihood of allocating.
3798 for (unsigned i
= 0; i
< ARRAY_SIZE(pre_modes
); i
++) {
3799 schedule_instructions(pre_modes
[i
]);
3802 assign_regs_trivial();
3803 allocated_without_spills
= true;
3805 allocated_without_spills
= assign_regs(false);
3807 if (allocated_without_spills
)
3811 if (!allocated_without_spills
) {
3812 const char *stage_name
= stage
== MESA_SHADER_VERTEX
?
3813 "Vertex" : "Fragment";
3815 /* We assume that any spilling is worse than just dropping back to
3816 * SIMD8. There's probably actually some intermediate point where
3817 * SIMD16 with a couple of spills is still better.
3819 if (dispatch_width
== 16) {
3820 fail("Failure to register allocate. Reduce number of "
3821 "live scalar values to avoid this.");
3823 perf_debug("%s shader triggered register spilling. "
3824 "Try reducing the number of live scalar values to "
3825 "improve performance.\n", stage_name
);
3828 /* Since we're out of heuristics, just go spill registers until we
3829 * get an allocation.
3831 while (!assign_regs(true)) {
3837 /* This must come after all optimization and register allocation, since
3838 * it inserts dead code that happens to have side effects, and it does
3839 * so based on the actual physical registers in use.
3841 insert_gen4_send_dependency_workarounds();
3846 if (!allocated_without_spills
)
3847 schedule_instructions(SCHEDULE_POST
);
3849 if (last_scratch
> 0)
3850 prog_data
->total_scratch
= brw_get_scratch_size(last_scratch
);
3854 env_var_as_boolean(const char *var_name
, bool default_value
)
3856 const char *str
= getenv(var_name
);
3858 return default_value
;
3860 if (strcmp(str
, "1") == 0 ||
3861 strcasecmp(str
, "true") == 0 ||
3862 strcasecmp(str
, "yes") == 0) {
3864 } else if (strcmp(str
, "0") == 0 ||
3865 strcasecmp(str
, "false") == 0 ||
3866 strcasecmp(str
, "no") == 0) {
3869 return default_value
;
3874 fs_visitor::run_vs()
3876 assert(stage
== MESA_SHADER_VERTEX
);
3878 assign_common_binding_table_offsets(0);
3881 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
)
3882 emit_shader_time_begin();
3884 if (env_var_as_boolean("INTEL_USE_NIR", false)) {
3887 foreach_in_list(ir_instruction
, ir
, shader
->base
.ir
) {
3889 this->result
= reg_undef
;
3904 assign_curb_setup();
3905 assign_vs_urb_setup();
3907 fixup_3src_null_dest();
3908 allocate_registers();
3914 fs_visitor::run_fs()
3916 brw_wm_prog_data
*wm_prog_data
= (brw_wm_prog_data
*) this->prog_data
;
3917 brw_wm_prog_key
*wm_key
= (brw_wm_prog_key
*) this->key
;
3919 assert(stage
== MESA_SHADER_FRAGMENT
);
3921 sanity_param_count
= prog
->Parameters
->NumParameters
;
3923 assign_binding_table_offsets();
3926 setup_payload_gen6();
3928 setup_payload_gen4();
3932 } else if (brw
->use_rep_send
&& dispatch_width
== 16) {
3933 emit_repclear_shader();
3935 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
)
3936 emit_shader_time_begin();
3938 calculate_urb_setup();
3939 if (prog
->InputsRead
> 0) {
3941 emit_interpolation_setup_gen4();
3943 emit_interpolation_setup_gen6();
3946 /* We handle discards by keeping track of the still-live pixels in f0.1.
3947 * Initialize it with the dispatched pixels.
3949 if (wm_prog_data
->uses_kill
) {
3950 fs_inst
*discard_init
= emit(FS_OPCODE_MOV_DISPATCH_TO_FLAGS
);
3951 discard_init
->flag_subreg
= 1;
3954 /* Generate FS IR for main(). (the visitor only descends into
3955 * functions called "main").
3958 if (env_var_as_boolean("INTEL_USE_NIR", false)) {
3961 foreach_in_list(ir_instruction
, ir
, shader
->base
.ir
) {
3963 this->result
= reg_undef
;
3968 emit_fragment_program_code();
3974 emit(FS_OPCODE_PLACEHOLDER_HALT
);
3976 if (wm_key
->alpha_test_func
)
3981 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
)
3982 emit_shader_time_end();
3988 assign_curb_setup();
3991 fixup_3src_null_dest();
3992 allocate_registers();
3998 if (dispatch_width
== 8)
3999 wm_prog_data
->reg_blocks
= brw_register_blocks(grf_used
);
4001 wm_prog_data
->reg_blocks_16
= brw_register_blocks(grf_used
);
4003 /* If any state parameters were appended, then ParameterValues could have
4004 * been realloced, in which case the driver uniform storage set up by
4005 * _mesa_associate_uniform_storage() would point to freed memory. Make
4006 * sure that didn't happen.
4008 assert(sanity_param_count
== prog
->Parameters
->NumParameters
);
4014 brw_wm_fs_emit(struct brw_context
*brw
,
4016 const struct brw_wm_prog_key
*key
,
4017 struct brw_wm_prog_data
*prog_data
,
4018 struct gl_fragment_program
*fp
,
4019 struct gl_shader_program
*prog
,
4020 unsigned *final_assembly_size
)
4022 bool start_busy
= false;
4023 double start_time
= 0;
4025 if (unlikely(brw
->perf_debug
)) {
4026 start_busy
= (brw
->batch
.last_bo
&&
4027 drm_intel_bo_busy(brw
->batch
.last_bo
));
4028 start_time
= get_time();
4031 struct brw_shader
*shader
= NULL
;
4033 shader
= (brw_shader
*) prog
->_LinkedShaders
[MESA_SHADER_FRAGMENT
];
4035 if (unlikely(INTEL_DEBUG
& DEBUG_WM
))
4036 brw_dump_ir("fragment", prog
, &shader
->base
, &fp
->Base
);
4038 /* Now the main event: Visit the shader IR and generate our FS IR for it.
4040 fs_visitor
v(brw
, mem_ctx
, key
, prog_data
, prog
, fp
, 8);
4043 prog
->LinkStatus
= false;
4044 ralloc_strcat(&prog
->InfoLog
, v
.fail_msg
);
4047 _mesa_problem(NULL
, "Failed to compile fragment shader: %s\n",
4053 cfg_t
*simd16_cfg
= NULL
;
4054 fs_visitor
v2(brw
, mem_ctx
, key
, prog_data
, prog
, fp
, 16);
4055 if (brw
->gen
>= 5 && likely(!(INTEL_DEBUG
& DEBUG_NO16
) ||
4056 brw
->use_rep_send
)) {
4057 if (!v
.simd16_unsupported
) {
4058 /* Try a SIMD16 compile */
4059 v2
.import_uniforms(&v
);
4061 perf_debug("SIMD16 shader failed to compile, falling back to "
4062 "SIMD8 at a 10-20%% performance cost: %s", v2
.fail_msg
);
4064 simd16_cfg
= v2
.cfg
;
4067 perf_debug("SIMD16 shader unsupported, falling back to "
4068 "SIMD8 at a 10-20%% performance cost: %s", v
.no16_msg
);
4073 int no_simd8
= (INTEL_DEBUG
& DEBUG_NO8
) || brw
->no_simd8
;
4074 if (no_simd8
&& simd16_cfg
) {
4076 prog_data
->no_8
= true;
4079 prog_data
->no_8
= false;
4082 fs_generator
g(brw
, mem_ctx
, (void *) key
, &prog_data
->base
,
4083 &fp
->Base
, v
.promoted_constants
, v
.runtime_check_aads_emit
, "FS");
4085 if (unlikely(INTEL_DEBUG
& DEBUG_WM
)) {
4088 name
= ralloc_asprintf(mem_ctx
, "%s fragment shader %d",
4089 prog
->Label
? prog
->Label
: "unnamed",
4092 name
= ralloc_asprintf(mem_ctx
, "fragment program %d", fp
->Base
.Id
);
4094 g
.enable_debug(name
);
4098 g
.generate_code(simd8_cfg
, 8);
4100 prog_data
->prog_offset_16
= g
.generate_code(simd16_cfg
, 16);
4102 if (unlikely(brw
->perf_debug
) && shader
) {
4103 if (shader
->compiled_once
)
4104 brw_wm_debug_recompile(brw
, prog
, key
);
4105 shader
->compiled_once
= true;
4107 if (start_busy
&& !drm_intel_bo_busy(brw
->batch
.last_bo
)) {
4108 perf_debug("FS compile took %.03f ms and stalled the GPU\n",
4109 (get_time() - start_time
) * 1000);
4113 return g
.get_assembly(final_assembly_size
);
4117 brw_fs_precompile(struct gl_context
*ctx
,
4118 struct gl_shader_program
*shader_prog
,
4119 struct gl_program
*prog
)
4121 struct brw_context
*brw
= brw_context(ctx
);
4122 struct brw_wm_prog_key key
;
4124 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*) prog
;
4125 struct brw_fragment_program
*bfp
= brw_fragment_program(fp
);
4126 bool program_uses_dfdy
= fp
->UsesDFdy
;
4128 memset(&key
, 0, sizeof(key
));
4132 key
.iz_lookup
|= IZ_PS_KILL_ALPHATEST_BIT
;
4134 if (fp
->Base
.OutputsWritten
& BITFIELD64_BIT(FRAG_RESULT_DEPTH
))
4135 key
.iz_lookup
|= IZ_PS_COMPUTES_DEPTH_BIT
;
4137 /* Just assume depth testing. */
4138 key
.iz_lookup
|= IZ_DEPTH_TEST_ENABLE_BIT
;
4139 key
.iz_lookup
|= IZ_DEPTH_WRITE_ENABLE_BIT
;
4142 if (brw
->gen
< 6 || _mesa_bitcount_64(fp
->Base
.InputsRead
&
4143 BRW_FS_VARYING_INPUT_MASK
) > 16)
4144 key
.input_slots_valid
= fp
->Base
.InputsRead
| VARYING_BIT_POS
;
4146 const bool has_shader_channel_select
= brw
->is_haswell
|| brw
->gen
>= 8;
4147 unsigned sampler_count
= _mesa_fls(fp
->Base
.SamplersUsed
);
4148 for (unsigned i
= 0; i
< sampler_count
; i
++) {
4149 if (!has_shader_channel_select
&& (fp
->Base
.ShadowSamplers
& (1 << i
))) {
4150 /* Assume DEPTH_TEXTURE_MODE is the default: X, X, X, 1 */
4151 key
.tex
.swizzles
[i
] =
4152 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_ONE
);
4154 /* Color sampler: assume no swizzling. */
4155 key
.tex
.swizzles
[i
] = SWIZZLE_XYZW
;
4159 if (fp
->Base
.InputsRead
& VARYING_BIT_POS
) {
4160 key
.drawable_height
= ctx
->DrawBuffer
->Height
;
4163 key
.nr_color_regions
= _mesa_bitcount_64(fp
->Base
.OutputsWritten
&
4164 ~(BITFIELD64_BIT(FRAG_RESULT_DEPTH
) |
4165 BITFIELD64_BIT(FRAG_RESULT_SAMPLE_MASK
)));
4167 if ((fp
->Base
.InputsRead
& VARYING_BIT_POS
) || program_uses_dfdy
) {
4168 key
.render_to_fbo
= _mesa_is_user_fbo(ctx
->DrawBuffer
) ||
4169 key
.nr_color_regions
> 1;
4172 key
.program_string_id
= bfp
->id
;
4174 uint32_t old_prog_offset
= brw
->wm
.base
.prog_offset
;
4175 struct brw_wm_prog_data
*old_prog_data
= brw
->wm
.prog_data
;
4177 bool success
= do_wm_prog(brw
, shader_prog
, bfp
, &key
);
4179 brw
->wm
.base
.prog_offset
= old_prog_offset
;
4180 brw
->wm
.prog_data
= old_prog_data
;