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
33 #include <sys/types.h>
35 #include "main/macros.h"
36 #include "main/shaderobj.h"
37 #include "main/uniforms.h"
38 #include "main/fbobject.h"
39 #include "program/prog_parameter.h"
40 #include "program/prog_print.h"
41 #include "program/register_allocate.h"
42 #include "program/sampler.h"
43 #include "program/hash_table.h"
44 #include "brw_context.h"
49 #include "glsl/glsl_types.h"
50 #include "glsl/ir_print_visitor.h"
55 memset(this, 0, sizeof(*this));
56 this->opcode
= BRW_OPCODE_NOP
;
57 this->conditional_mod
= BRW_CONDITIONAL_NONE
;
59 this->dst
= reg_undef
;
60 this->src
[0] = reg_undef
;
61 this->src
[1] = reg_undef
;
62 this->src
[2] = reg_undef
;
70 fs_inst::fs_inst(enum opcode opcode
)
73 this->opcode
= opcode
;
76 fs_inst::fs_inst(enum opcode opcode
, fs_reg dst
)
79 this->opcode
= opcode
;
83 assert(dst
.reg_offset
>= 0);
86 fs_inst::fs_inst(enum opcode opcode
, fs_reg dst
, fs_reg src0
)
89 this->opcode
= opcode
;
94 assert(dst
.reg_offset
>= 0);
95 if (src
[0].file
== GRF
)
96 assert(src
[0].reg_offset
>= 0);
99 fs_inst::fs_inst(enum opcode opcode
, fs_reg dst
, fs_reg src0
, fs_reg src1
)
102 this->opcode
= opcode
;
108 assert(dst
.reg_offset
>= 0);
109 if (src
[0].file
== GRF
)
110 assert(src
[0].reg_offset
>= 0);
111 if (src
[1].file
== GRF
)
112 assert(src
[1].reg_offset
>= 0);
115 fs_inst::fs_inst(enum opcode opcode
, fs_reg dst
,
116 fs_reg src0
, fs_reg src1
, fs_reg src2
)
119 this->opcode
= opcode
;
126 assert(dst
.reg_offset
>= 0);
127 if (src
[0].file
== GRF
)
128 assert(src
[0].reg_offset
>= 0);
129 if (src
[1].file
== GRF
)
130 assert(src
[1].reg_offset
>= 0);
131 if (src
[2].file
== GRF
)
132 assert(src
[2].reg_offset
>= 0);
137 fs_visitor::op(fs_reg dst, fs_reg src0) \
139 return new(mem_ctx) fs_inst(BRW_OPCODE_##op, dst, src0); \
144 fs_visitor::op(fs_reg dst, fs_reg src0, fs_reg src1) \
146 return new(mem_ctx) fs_inst(BRW_OPCODE_##op, dst, src0, src1); \
165 /** Gen4 predicated IF. */
167 fs_visitor::IF(uint32_t predicate
)
169 fs_inst
*inst
= new(mem_ctx
) fs_inst(BRW_OPCODE_IF
);
170 inst
->predicate
= predicate
;
174 /** Gen6+ IF with embedded comparison. */
176 fs_visitor::IF(fs_reg src0
, fs_reg src1
, uint32_t condition
)
178 assert(intel
->gen
>= 6);
179 fs_inst
*inst
= new(mem_ctx
) fs_inst(BRW_OPCODE_IF
,
180 reg_null_d
, src0
, src1
);
181 inst
->conditional_mod
= condition
;
186 * CMP: Sets the low bit of the destination channels with the result
187 * of the comparison, while the upper bits are undefined, and updates
188 * the flag register with the packed 16 bits of the result.
191 fs_visitor::CMP(fs_reg dst
, fs_reg src0
, fs_reg src1
, uint32_t condition
)
195 /* Take the instruction:
197 * CMP null<d> src0<f> src1<f>
199 * Original gen4 does type conversion to the destination type before
200 * comparison, producing garbage results for floating point comparisons.
201 * gen5 does the comparison on the execution type (resolved source types),
202 * so dst type doesn't matter. gen6 does comparison and then uses the
203 * result as if it was the dst type with no conversion, which happens to
204 * mostly work out for float-interpreted-as-int since our comparisons are
207 if (intel
->gen
== 4) {
208 dst
.type
= src0
.type
;
209 if (dst
.file
== FIXED_HW_REG
)
210 dst
.fixed_hw_reg
.type
= dst
.type
;
213 resolve_ud_negate(&src0
);
214 resolve_ud_negate(&src1
);
216 inst
= new(mem_ctx
) fs_inst(BRW_OPCODE_CMP
, dst
, src0
, src1
);
217 inst
->conditional_mod
= condition
;
223 fs_inst::equals(fs_inst
*inst
)
225 return (opcode
== inst
->opcode
&&
226 dst
.equals(inst
->dst
) &&
227 src
[0].equals(inst
->src
[0]) &&
228 src
[1].equals(inst
->src
[1]) &&
229 src
[2].equals(inst
->src
[2]) &&
230 saturate
== inst
->saturate
&&
231 predicate
== inst
->predicate
&&
232 conditional_mod
== inst
->conditional_mod
&&
233 mlen
== inst
->mlen
&&
234 base_mrf
== inst
->base_mrf
&&
235 sampler
== inst
->sampler
&&
236 target
== inst
->target
&&
238 header_present
== inst
->header_present
&&
239 shadow_compare
== inst
->shadow_compare
&&
240 offset
== inst
->offset
);
244 fs_inst::regs_written()
249 /* The SINCOS and INT_DIV_QUOTIENT_AND_REMAINDER math functions return 2,
250 * but we don't currently use them...nor do we have an opcode for them.
257 fs_inst::overwrites_reg(const fs_reg
®
)
259 return (reg
.file
== dst
.file
&&
260 reg
.reg
== dst
.reg
&&
261 reg
.reg_offset
>= dst
.reg_offset
&&
262 reg
.reg_offset
< dst
.reg_offset
+ regs_written());
268 return (opcode
== SHADER_OPCODE_TEX
||
269 opcode
== FS_OPCODE_TXB
||
270 opcode
== SHADER_OPCODE_TXD
||
271 opcode
== SHADER_OPCODE_TXF
||
272 opcode
== SHADER_OPCODE_TXL
||
273 opcode
== SHADER_OPCODE_TXS
);
279 return (opcode
== SHADER_OPCODE_RCP
||
280 opcode
== SHADER_OPCODE_RSQ
||
281 opcode
== SHADER_OPCODE_SQRT
||
282 opcode
== SHADER_OPCODE_EXP2
||
283 opcode
== SHADER_OPCODE_LOG2
||
284 opcode
== SHADER_OPCODE_SIN
||
285 opcode
== SHADER_OPCODE_COS
||
286 opcode
== SHADER_OPCODE_INT_QUOTIENT
||
287 opcode
== SHADER_OPCODE_INT_REMAINDER
||
288 opcode
== SHADER_OPCODE_POW
);
294 memset(this, 0, sizeof(*this));
298 /** Generic unset register constructor. */
302 this->file
= BAD_FILE
;
305 /** Immediate value constructor. */
306 fs_reg::fs_reg(float f
)
310 this->type
= BRW_REGISTER_TYPE_F
;
314 /** Immediate value constructor. */
315 fs_reg::fs_reg(int32_t i
)
319 this->type
= BRW_REGISTER_TYPE_D
;
323 /** Immediate value constructor. */
324 fs_reg::fs_reg(uint32_t u
)
328 this->type
= BRW_REGISTER_TYPE_UD
;
332 /** Fixed brw_reg Immediate value constructor. */
333 fs_reg::fs_reg(struct brw_reg fixed_hw_reg
)
336 this->file
= FIXED_HW_REG
;
337 this->fixed_hw_reg
= fixed_hw_reg
;
338 this->type
= fixed_hw_reg
.type
;
342 fs_reg::equals(const fs_reg
&r
) const
344 return (file
== r
.file
&&
346 reg_offset
== r
.reg_offset
&&
348 negate
== r
.negate
&&
350 memcmp(&fixed_hw_reg
, &r
.fixed_hw_reg
,
351 sizeof(fixed_hw_reg
)) == 0 &&
357 fs_reg::is_zero() const
362 return type
== BRW_REGISTER_TYPE_F
? imm
.f
== 0.0 : imm
.i
== 0;
366 fs_reg::is_one() const
371 return type
== BRW_REGISTER_TYPE_F
? imm
.f
== 1.0 : imm
.i
== 1;
375 fs_visitor::type_size(const struct glsl_type
*type
)
377 unsigned int size
, i
;
379 switch (type
->base_type
) {
382 case GLSL_TYPE_FLOAT
:
384 return type
->components();
385 case GLSL_TYPE_ARRAY
:
386 return type_size(type
->fields
.array
) * type
->length
;
387 case GLSL_TYPE_STRUCT
:
389 for (i
= 0; i
< type
->length
; i
++) {
390 size
+= type_size(type
->fields
.structure
[i
].type
);
393 case GLSL_TYPE_SAMPLER
:
394 /* Samplers take up no register space, since they're baked in at
399 assert(!"not reached");
405 fs_visitor::fail(const char *format
, ...)
415 va_start(va
, format
);
416 msg
= ralloc_vasprintf(mem_ctx
, format
, va
);
418 msg
= ralloc_asprintf(mem_ctx
, "FS compile failed: %s\n", msg
);
420 this->fail_msg
= msg
;
422 if (INTEL_DEBUG
& DEBUG_WM
) {
423 fprintf(stderr
, "%s", msg
);
428 fs_visitor::emit(enum opcode opcode
)
430 return emit(fs_inst(opcode
));
434 fs_visitor::emit(enum opcode opcode
, fs_reg dst
)
436 return emit(fs_inst(opcode
, dst
));
440 fs_visitor::emit(enum opcode opcode
, fs_reg dst
, fs_reg src0
)
442 return emit(fs_inst(opcode
, dst
, src0
));
446 fs_visitor::emit(enum opcode opcode
, fs_reg dst
, fs_reg src0
, fs_reg src1
)
448 return emit(fs_inst(opcode
, dst
, src0
, src1
));
452 fs_visitor::emit(enum opcode opcode
, fs_reg dst
,
453 fs_reg src0
, fs_reg src1
, fs_reg src2
)
455 return emit(fs_inst(opcode
, dst
, src0
, src1
, src2
));
459 fs_visitor::push_force_uncompressed()
461 force_uncompressed_stack
++;
465 fs_visitor::pop_force_uncompressed()
467 force_uncompressed_stack
--;
468 assert(force_uncompressed_stack
>= 0);
472 fs_visitor::push_force_sechalf()
474 force_sechalf_stack
++;
478 fs_visitor::pop_force_sechalf()
480 force_sechalf_stack
--;
481 assert(force_sechalf_stack
>= 0);
485 * Returns how many MRFs an FS opcode will write over.
487 * Note that this is not the 0 or 1 implied writes in an actual gen
488 * instruction -- the FS opcodes often generate MOVs in addition.
491 fs_visitor::implied_mrf_writes(fs_inst
*inst
)
496 switch (inst
->opcode
) {
497 case SHADER_OPCODE_RCP
:
498 case SHADER_OPCODE_RSQ
:
499 case SHADER_OPCODE_SQRT
:
500 case SHADER_OPCODE_EXP2
:
501 case SHADER_OPCODE_LOG2
:
502 case SHADER_OPCODE_SIN
:
503 case SHADER_OPCODE_COS
:
504 return 1 * dispatch_width
/ 8;
505 case SHADER_OPCODE_POW
:
506 case SHADER_OPCODE_INT_QUOTIENT
:
507 case SHADER_OPCODE_INT_REMAINDER
:
508 return 2 * dispatch_width
/ 8;
509 case SHADER_OPCODE_TEX
:
511 case SHADER_OPCODE_TXD
:
512 case SHADER_OPCODE_TXF
:
513 case SHADER_OPCODE_TXL
:
514 case SHADER_OPCODE_TXS
:
516 case FS_OPCODE_FB_WRITE
:
518 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
519 case FS_OPCODE_UNSPILL
:
521 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD
:
522 return inst
->header_present
;
523 case FS_OPCODE_SPILL
:
526 assert(!"not reached");
532 fs_visitor::virtual_grf_alloc(int size
)
534 if (virtual_grf_array_size
<= virtual_grf_count
) {
535 if (virtual_grf_array_size
== 0)
536 virtual_grf_array_size
= 16;
538 virtual_grf_array_size
*= 2;
539 virtual_grf_sizes
= reralloc(mem_ctx
, virtual_grf_sizes
, int,
540 virtual_grf_array_size
);
542 virtual_grf_sizes
[virtual_grf_count
] = size
;
543 return virtual_grf_count
++;
546 /** Fixed HW reg constructor. */
547 fs_reg::fs_reg(enum register_file file
, int reg
)
552 this->type
= BRW_REGISTER_TYPE_F
;
555 /** Fixed HW reg constructor. */
556 fs_reg::fs_reg(enum register_file file
, int reg
, uint32_t type
)
564 /** Automatic reg constructor. */
565 fs_reg::fs_reg(class fs_visitor
*v
, const struct glsl_type
*type
)
570 this->reg
= v
->virtual_grf_alloc(v
->type_size(type
));
571 this->reg_offset
= 0;
572 this->type
= brw_type_for_base_type(type
);
576 fs_visitor::variable_storage(ir_variable
*var
)
578 return (fs_reg
*)hash_table_find(this->variable_ht
, var
);
582 import_uniforms_callback(const void *key
,
586 struct hash_table
*dst_ht
= (struct hash_table
*)closure
;
587 const fs_reg
*reg
= (const fs_reg
*)data
;
589 if (reg
->file
!= UNIFORM
)
592 hash_table_insert(dst_ht
, data
, key
);
595 /* For 16-wide, we need to follow from the uniform setup of 8-wide dispatch.
596 * This brings in those uniform definitions
599 fs_visitor::import_uniforms(fs_visitor
*v
)
601 hash_table_call_foreach(v
->variable_ht
,
602 import_uniforms_callback
,
604 this->params_remap
= v
->params_remap
;
607 /* Our support for uniforms is piggy-backed on the struct
608 * gl_fragment_program, because that's where the values actually
609 * get stored, rather than in some global gl_shader_program uniform
613 fs_visitor::setup_uniform_values(int loc
, const glsl_type
*type
)
615 unsigned int offset
= 0;
617 if (type
->is_matrix()) {
618 const glsl_type
*column
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
619 type
->vector_elements
,
622 for (unsigned int i
= 0; i
< type
->matrix_columns
; i
++) {
623 offset
+= setup_uniform_values(loc
+ offset
, column
);
629 switch (type
->base_type
) {
630 case GLSL_TYPE_FLOAT
:
634 for (unsigned int i
= 0; i
< type
->vector_elements
; i
++) {
635 unsigned int param
= c
->prog_data
.nr_params
++;
637 this->param_index
[param
] = loc
;
638 this->param_offset
[param
] = i
;
642 case GLSL_TYPE_STRUCT
:
643 for (unsigned int i
= 0; i
< type
->length
; i
++) {
644 offset
+= setup_uniform_values(loc
+ offset
,
645 type
->fields
.structure
[i
].type
);
649 case GLSL_TYPE_ARRAY
:
650 for (unsigned int i
= 0; i
< type
->length
; i
++) {
651 offset
+= setup_uniform_values(loc
+ offset
, type
->fields
.array
);
655 case GLSL_TYPE_SAMPLER
:
656 /* The sampler takes up a slot, but we don't use any values from it. */
660 assert(!"not reached");
666 /* Our support for builtin uniforms is even scarier than non-builtin.
667 * It sits on top of the PROG_STATE_VAR parameters that are
668 * automatically updated from GL context state.
671 fs_visitor::setup_builtin_uniform_values(ir_variable
*ir
)
673 const ir_state_slot
*const slots
= ir
->state_slots
;
674 assert(ir
->state_slots
!= NULL
);
676 for (unsigned int i
= 0; i
< ir
->num_state_slots
; i
++) {
677 /* This state reference has already been setup by ir_to_mesa, but we'll
678 * get the same index back here.
680 int index
= _mesa_add_state_reference(this->fp
->Base
.Parameters
,
681 (gl_state_index
*)slots
[i
].tokens
);
683 /* Add each of the unique swizzles of the element as a parameter.
684 * This'll end up matching the expected layout of the
685 * array/matrix/structure we're trying to fill in.
688 for (unsigned int j
= 0; j
< 4; j
++) {
689 int swiz
= GET_SWZ(slots
[i
].swizzle
, j
);
690 if (swiz
== last_swiz
)
694 this->param_index
[c
->prog_data
.nr_params
] = index
;
695 this->param_offset
[c
->prog_data
.nr_params
] = swiz
;
696 c
->prog_data
.nr_params
++;
702 fs_visitor::emit_fragcoord_interpolation(ir_variable
*ir
)
704 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(this, ir
->type
);
706 bool flip
= !ir
->origin_upper_left
^ c
->key
.render_to_fbo
;
709 if (ir
->pixel_center_integer
) {
710 emit(MOV(wpos
, this->pixel_x
));
712 emit(ADD(wpos
, this->pixel_x
, fs_reg(0.5f
)));
717 if (!flip
&& ir
->pixel_center_integer
) {
718 emit(MOV(wpos
, this->pixel_y
));
720 fs_reg pixel_y
= this->pixel_y
;
721 float offset
= (ir
->pixel_center_integer
? 0.0 : 0.5);
724 pixel_y
.negate
= true;
725 offset
+= c
->key
.drawable_height
- 1.0;
728 emit(ADD(wpos
, pixel_y
, fs_reg(offset
)));
733 if (intel
->gen
>= 6) {
734 emit(MOV(wpos
, fs_reg(brw_vec8_grf(c
->source_depth_reg
, 0))));
736 emit(FS_OPCODE_LINTERP
, wpos
,
737 this->delta_x
[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
],
738 this->delta_y
[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
],
739 interp_reg(FRAG_ATTRIB_WPOS
, 2));
743 /* gl_FragCoord.w: Already set up in emit_interpolation */
744 emit(BRW_OPCODE_MOV
, wpos
, this->wpos_w
);
750 fs_visitor::emit_linterp(const fs_reg
&attr
, const fs_reg
&interp
,
751 glsl_interp_qualifier interpolation_mode
,
754 brw_wm_barycentric_interp_mode barycoord_mode
;
756 if (interpolation_mode
== INTERP_QUALIFIER_SMOOTH
)
757 barycoord_mode
= BRW_WM_PERSPECTIVE_CENTROID_BARYCENTRIC
;
759 barycoord_mode
= BRW_WM_NONPERSPECTIVE_CENTROID_BARYCENTRIC
;
761 if (interpolation_mode
== INTERP_QUALIFIER_SMOOTH
)
762 barycoord_mode
= BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
;
764 barycoord_mode
= BRW_WM_NONPERSPECTIVE_PIXEL_BARYCENTRIC
;
766 return emit(FS_OPCODE_LINTERP
, attr
,
767 this->delta_x
[barycoord_mode
],
768 this->delta_y
[barycoord_mode
], interp
);
772 fs_visitor::emit_general_interpolation(ir_variable
*ir
)
774 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(this, ir
->type
);
775 reg
->type
= brw_type_for_base_type(ir
->type
->get_scalar_type());
778 unsigned int array_elements
;
779 const glsl_type
*type
;
781 if (ir
->type
->is_array()) {
782 array_elements
= ir
->type
->length
;
783 if (array_elements
== 0) {
784 fail("dereferenced array '%s' has length 0\n", ir
->name
);
786 type
= ir
->type
->fields
.array
;
792 glsl_interp_qualifier interpolation_mode
=
793 ir
->determine_interpolation_mode(c
->key
.flat_shade
);
795 int location
= ir
->location
;
796 for (unsigned int i
= 0; i
< array_elements
; i
++) {
797 for (unsigned int j
= 0; j
< type
->matrix_columns
; j
++) {
798 if (urb_setup
[location
] == -1) {
799 /* If there's no incoming setup data for this slot, don't
800 * emit interpolation for it.
802 attr
.reg_offset
+= type
->vector_elements
;
807 if (interpolation_mode
== INTERP_QUALIFIER_FLAT
) {
808 /* Constant interpolation (flat shading) case. The SF has
809 * handed us defined values in only the constant offset
810 * field of the setup reg.
812 for (unsigned int k
= 0; k
< type
->vector_elements
; k
++) {
813 struct brw_reg interp
= interp_reg(location
, k
);
814 interp
= suboffset(interp
, 3);
815 interp
.type
= reg
->type
;
816 emit(FS_OPCODE_CINTERP
, attr
, fs_reg(interp
));
820 /* Smooth/noperspective interpolation case. */
821 for (unsigned int k
= 0; k
< type
->vector_elements
; k
++) {
822 /* FINISHME: At some point we probably want to push
823 * this farther by giving similar treatment to the
824 * other potentially constant components of the
825 * attribute, as well as making brw_vs_constval.c
826 * handle varyings other than gl_TexCoord.
828 if (location
>= FRAG_ATTRIB_TEX0
&&
829 location
<= FRAG_ATTRIB_TEX7
&&
830 k
== 3 && !(c
->key
.proj_attrib_mask
& (1 << location
))) {
831 emit(BRW_OPCODE_MOV
, attr
, fs_reg(1.0f
));
833 struct brw_reg interp
= interp_reg(location
, k
);
834 emit_linterp(attr
, fs_reg(interp
), interpolation_mode
,
836 if (brw
->needs_unlit_centroid_workaround
&& ir
->centroid
) {
837 /* Get the pixel/sample mask into f0 so that we know
838 * which pixels are lit. Then, for each channel that is
839 * unlit, replace the centroid data with non-centroid
842 emit(FS_OPCODE_MOV_DISPATCH_TO_FLAGS
, attr
);
843 fs_inst
*inst
= emit_linterp(attr
, fs_reg(interp
),
844 interpolation_mode
, false);
845 inst
->predicate
= BRW_PREDICATE_NORMAL
;
846 inst
->predicate_inverse
= true;
848 if (intel
->gen
< 6) {
849 emit(BRW_OPCODE_MUL
, attr
, attr
, this->pixel_w
);
864 fs_visitor::emit_frontfacing_interpolation(ir_variable
*ir
)
866 fs_reg
*reg
= new(this->mem_ctx
) fs_reg(this, ir
->type
);
868 /* The frontfacing comes in as a bit in the thread payload. */
869 if (intel
->gen
>= 6) {
870 emit(BRW_OPCODE_ASR
, *reg
,
871 fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_D
)),
873 emit(BRW_OPCODE_NOT
, *reg
, *reg
);
874 emit(BRW_OPCODE_AND
, *reg
, *reg
, fs_reg(1));
876 struct brw_reg r1_6ud
= retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_UD
);
877 /* bit 31 is "primitive is back face", so checking < (1 << 31) gives
880 emit(CMP(*reg
, fs_reg(r1_6ud
), fs_reg(1u << 31), BRW_CONDITIONAL_L
));
881 emit(BRW_OPCODE_AND
, *reg
, *reg
, fs_reg(1u));
888 fs_visitor::emit_math(enum opcode opcode
, fs_reg dst
, fs_reg src
)
891 case SHADER_OPCODE_RCP
:
892 case SHADER_OPCODE_RSQ
:
893 case SHADER_OPCODE_SQRT
:
894 case SHADER_OPCODE_EXP2
:
895 case SHADER_OPCODE_LOG2
:
896 case SHADER_OPCODE_SIN
:
897 case SHADER_OPCODE_COS
:
900 assert(!"not reached: bad math opcode");
904 /* Can't do hstride == 0 args to gen6 math, so expand it out. We
905 * might be able to do better by doing execsize = 1 math and then
906 * expanding that result out, but we would need to be careful with
909 * Gen 6 hardware ignores source modifiers (negate and abs) on math
910 * instructions, so we also move to a temp to set those up.
912 if (intel
->gen
== 6 && (src
.file
== UNIFORM
||
915 fs_reg expanded
= fs_reg(this, glsl_type::float_type
);
916 emit(BRW_OPCODE_MOV
, expanded
, src
);
920 fs_inst
*inst
= emit(opcode
, dst
, src
);
922 if (intel
->gen
< 6) {
924 inst
->mlen
= dispatch_width
/ 8;
931 fs_visitor::emit_math(enum opcode opcode
, fs_reg dst
, fs_reg src0
, fs_reg src1
)
937 case SHADER_OPCODE_POW
:
938 case SHADER_OPCODE_INT_QUOTIENT
:
939 case SHADER_OPCODE_INT_REMAINDER
:
942 assert(!"not reached: unsupported binary math opcode.");
946 if (intel
->gen
>= 7) {
947 inst
= emit(opcode
, dst
, src0
, src1
);
948 } else if (intel
->gen
== 6) {
949 /* Can't do hstride == 0 args to gen6 math, so expand it out.
951 * The hardware ignores source modifiers (negate and abs) on math
952 * instructions, so we also move to a temp to set those up.
954 if (src0
.file
== UNIFORM
|| src0
.abs
|| src0
.negate
) {
955 fs_reg expanded
= fs_reg(this, glsl_type::float_type
);
956 expanded
.type
= src0
.type
;
957 emit(BRW_OPCODE_MOV
, expanded
, src0
);
961 if (src1
.file
== UNIFORM
|| src1
.abs
|| src1
.negate
) {
962 fs_reg expanded
= fs_reg(this, glsl_type::float_type
);
963 expanded
.type
= src1
.type
;
964 emit(BRW_OPCODE_MOV
, expanded
, src1
);
968 inst
= emit(opcode
, dst
, src0
, src1
);
970 /* From the Ironlake PRM, Volume 4, Part 1, Section 6.1.13
973 * "Operand0[7]. For the INT DIV functions, this operand is the
976 * "Operand1[7]. For the INT DIV functions, this operand is the
979 bool is_int_div
= opcode
!= SHADER_OPCODE_POW
;
980 fs_reg
&op0
= is_int_div
? src1
: src0
;
981 fs_reg
&op1
= is_int_div
? src0
: src1
;
983 emit(BRW_OPCODE_MOV
, fs_reg(MRF
, base_mrf
+ 1, op1
.type
), op1
);
984 inst
= emit(opcode
, dst
, op0
, reg_null_f
);
986 inst
->base_mrf
= base_mrf
;
987 inst
->mlen
= 2 * dispatch_width
/ 8;
993 * To be called after the last _mesa_add_state_reference() call, to
994 * set up prog_data.param[] for assign_curb_setup() and
995 * setup_pull_constants().
998 fs_visitor::setup_paramvalues_refs()
1000 if (dispatch_width
!= 8)
1003 /* Set up the pointers to ParamValues now that that array is finalized. */
1004 for (unsigned int i
= 0; i
< c
->prog_data
.nr_params
; i
++) {
1005 c
->prog_data
.param
[i
] =
1006 (const float *)fp
->Base
.Parameters
->ParameterValues
[this->param_index
[i
]] +
1007 this->param_offset
[i
];
1012 fs_visitor::assign_curb_setup()
1014 c
->prog_data
.curb_read_length
= ALIGN(c
->prog_data
.nr_params
, 8) / 8;
1015 if (dispatch_width
== 8) {
1016 c
->prog_data
.first_curbe_grf
= c
->nr_payload_regs
;
1018 c
->prog_data
.first_curbe_grf_16
= c
->nr_payload_regs
;
1021 /* Map the offsets in the UNIFORM file to fixed HW regs. */
1022 foreach_list(node
, &this->instructions
) {
1023 fs_inst
*inst
= (fs_inst
*)node
;
1025 for (unsigned int i
= 0; i
< 3; i
++) {
1026 if (inst
->src
[i
].file
== UNIFORM
) {
1027 int constant_nr
= inst
->src
[i
].reg
+ inst
->src
[i
].reg_offset
;
1028 struct brw_reg brw_reg
= brw_vec1_grf(c
->nr_payload_regs
+
1032 inst
->src
[i
].file
= FIXED_HW_REG
;
1033 inst
->src
[i
].fixed_hw_reg
= retype(brw_reg
, inst
->src
[i
].type
);
1040 fs_visitor::calculate_urb_setup()
1042 for (unsigned int i
= 0; i
< FRAG_ATTRIB_MAX
; i
++) {
1047 /* Figure out where each of the incoming setup attributes lands. */
1048 if (intel
->gen
>= 6) {
1049 for (unsigned int i
= 0; i
< FRAG_ATTRIB_MAX
; i
++) {
1050 if (fp
->Base
.InputsRead
& BITFIELD64_BIT(i
)) {
1051 urb_setup
[i
] = urb_next
++;
1055 /* FINISHME: The sf doesn't map VS->FS inputs for us very well. */
1056 for (unsigned int i
= 0; i
< VERT_RESULT_MAX
; i
++) {
1057 /* Point size is packed into the header, not as a general attribute */
1058 if (i
== VERT_RESULT_PSIZ
)
1061 if (c
->key
.vp_outputs_written
& BITFIELD64_BIT(i
)) {
1062 int fp_index
= _mesa_vert_result_to_frag_attrib((gl_vert_result
) i
);
1064 /* The back color slot is skipped when the front color is
1065 * also written to. In addition, some slots can be
1066 * written in the vertex shader and not read in the
1067 * fragment shader. So the register number must always be
1068 * incremented, mapped or not.
1071 urb_setup
[fp_index
] = urb_next
;
1077 * It's a FS only attribute, and we did interpolation for this attribute
1078 * in SF thread. So, count it here, too.
1080 * See compile_sf_prog() for more info.
1082 if (fp
->Base
.InputsRead
& BITFIELD64_BIT(FRAG_ATTRIB_PNTC
))
1083 urb_setup
[FRAG_ATTRIB_PNTC
] = urb_next
++;
1086 /* Each attribute is 4 setup channels, each of which is half a reg. */
1087 c
->prog_data
.urb_read_length
= urb_next
* 2;
1091 fs_visitor::assign_urb_setup()
1093 int urb_start
= c
->nr_payload_regs
+ c
->prog_data
.curb_read_length
;
1095 /* Offset all the urb_setup[] index by the actual position of the
1096 * setup regs, now that the location of the constants has been chosen.
1098 foreach_list(node
, &this->instructions
) {
1099 fs_inst
*inst
= (fs_inst
*)node
;
1101 if (inst
->opcode
== FS_OPCODE_LINTERP
) {
1102 assert(inst
->src
[2].file
== FIXED_HW_REG
);
1103 inst
->src
[2].fixed_hw_reg
.nr
+= urb_start
;
1106 if (inst
->opcode
== FS_OPCODE_CINTERP
) {
1107 assert(inst
->src
[0].file
== FIXED_HW_REG
);
1108 inst
->src
[0].fixed_hw_reg
.nr
+= urb_start
;
1112 this->first_non_payload_grf
= urb_start
+ c
->prog_data
.urb_read_length
;
1116 * Split large virtual GRFs into separate components if we can.
1118 * This is mostly duplicated with what brw_fs_vector_splitting does,
1119 * but that's really conservative because it's afraid of doing
1120 * splitting that doesn't result in real progress after the rest of
1121 * the optimization phases, which would cause infinite looping in
1122 * optimization. We can do it once here, safely. This also has the
1123 * opportunity to split interpolated values, or maybe even uniforms,
1124 * which we don't have at the IR level.
1126 * We want to split, because virtual GRFs are what we register
1127 * allocate and spill (due to contiguousness requirements for some
1128 * instructions), and they're what we naturally generate in the
1129 * codegen process, but most virtual GRFs don't actually need to be
1130 * contiguous sets of GRFs. If we split, we'll end up with reduced
1131 * live intervals and better dead code elimination and coalescing.
1134 fs_visitor::split_virtual_grfs()
1136 int num_vars
= this->virtual_grf_count
;
1137 bool split_grf
[num_vars
];
1138 int new_virtual_grf
[num_vars
];
1140 /* Try to split anything > 0 sized. */
1141 for (int i
= 0; i
< num_vars
; i
++) {
1142 if (this->virtual_grf_sizes
[i
] != 1)
1143 split_grf
[i
] = true;
1145 split_grf
[i
] = false;
1149 this->delta_x
[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
].file
== GRF
) {
1150 /* PLN opcodes rely on the delta_xy being contiguous. We only have to
1151 * check this for BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC, because prior to
1152 * Gen6, that was the only supported interpolation mode, and since Gen6,
1153 * delta_x and delta_y are in fixed hardware registers.
1155 split_grf
[this->delta_x
[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
].reg
] =
1159 foreach_list(node
, &this->instructions
) {
1160 fs_inst
*inst
= (fs_inst
*)node
;
1162 /* If there's a SEND message that requires contiguous destination
1163 * registers, no splitting is allowed.
1165 if (inst
->regs_written() > 1) {
1166 split_grf
[inst
->dst
.reg
] = false;
1170 /* Allocate new space for split regs. Note that the virtual
1171 * numbers will be contiguous.
1173 for (int i
= 0; i
< num_vars
; i
++) {
1175 new_virtual_grf
[i
] = virtual_grf_alloc(1);
1176 for (int j
= 2; j
< this->virtual_grf_sizes
[i
]; j
++) {
1177 int reg
= virtual_grf_alloc(1);
1178 assert(reg
== new_virtual_grf
[i
] + j
- 1);
1181 this->virtual_grf_sizes
[i
] = 1;
1185 foreach_list(node
, &this->instructions
) {
1186 fs_inst
*inst
= (fs_inst
*)node
;
1188 if (inst
->dst
.file
== GRF
&&
1189 split_grf
[inst
->dst
.reg
] &&
1190 inst
->dst
.reg_offset
!= 0) {
1191 inst
->dst
.reg
= (new_virtual_grf
[inst
->dst
.reg
] +
1192 inst
->dst
.reg_offset
- 1);
1193 inst
->dst
.reg_offset
= 0;
1195 for (int i
= 0; i
< 3; i
++) {
1196 if (inst
->src
[i
].file
== GRF
&&
1197 split_grf
[inst
->src
[i
].reg
] &&
1198 inst
->src
[i
].reg_offset
!= 0) {
1199 inst
->src
[i
].reg
= (new_virtual_grf
[inst
->src
[i
].reg
] +
1200 inst
->src
[i
].reg_offset
- 1);
1201 inst
->src
[i
].reg_offset
= 0;
1205 this->live_intervals_valid
= false;
1209 * Remove unused virtual GRFs and compact the virtual_grf_* arrays.
1211 * During code generation, we create tons of temporary variables, many of
1212 * which get immediately killed and are never used again. Yet, in later
1213 * optimization and analysis passes, such as compute_live_intervals, we need
1214 * to loop over all the virtual GRFs. Compacting them can save a lot of
1218 fs_visitor::compact_virtual_grfs()
1220 /* Mark which virtual GRFs are used, and count how many. */
1221 int remap_table
[this->virtual_grf_count
];
1222 memset(remap_table
, -1, sizeof(remap_table
));
1224 foreach_list(node
, &this->instructions
) {
1225 const fs_inst
*inst
= (const fs_inst
*) node
;
1227 if (inst
->dst
.file
== GRF
)
1228 remap_table
[inst
->dst
.reg
] = 0;
1230 for (int i
= 0; i
< 3; i
++) {
1231 if (inst
->src
[i
].file
== GRF
)
1232 remap_table
[inst
->src
[i
].reg
] = 0;
1236 /* In addition to registers used in instructions, fs_visitor keeps
1237 * direct references to certain special values which must be patched:
1239 fs_reg
*special
[] = {
1240 &frag_depth
, &pixel_x
, &pixel_y
, &pixel_w
, &wpos_w
, &dual_src_output
,
1241 &outputs
[0], &outputs
[1], &outputs
[2], &outputs
[3],
1242 &outputs
[4], &outputs
[5], &outputs
[6], &outputs
[7],
1243 &delta_x
[0], &delta_x
[1], &delta_x
[2],
1244 &delta_x
[3], &delta_x
[4], &delta_x
[5],
1245 &delta_y
[0], &delta_y
[1], &delta_y
[2],
1246 &delta_y
[3], &delta_y
[4], &delta_y
[5],
1248 STATIC_ASSERT(BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT
== 6);
1249 STATIC_ASSERT(BRW_MAX_DRAW_BUFFERS
== 8);
1251 /* Treat all special values as used, to be conservative */
1252 for (unsigned i
= 0; i
< ARRAY_SIZE(special
); i
++) {
1253 if (special
[i
]->file
== GRF
)
1254 remap_table
[special
[i
]->reg
] = 0;
1257 /* Compact the GRF arrays. */
1259 for (int i
= 0; i
< this->virtual_grf_count
; i
++) {
1260 if (remap_table
[i
] != -1) {
1261 remap_table
[i
] = new_index
;
1262 virtual_grf_sizes
[new_index
] = virtual_grf_sizes
[i
];
1263 if (live_intervals_valid
) {
1264 virtual_grf_use
[new_index
] = virtual_grf_use
[i
];
1265 virtual_grf_def
[new_index
] = virtual_grf_def
[i
];
1271 this->virtual_grf_count
= new_index
;
1273 /* Patch all the instructions to use the newly renumbered registers */
1274 foreach_list(node
, &this->instructions
) {
1275 fs_inst
*inst
= (fs_inst
*) node
;
1277 if (inst
->dst
.file
== GRF
)
1278 inst
->dst
.reg
= remap_table
[inst
->dst
.reg
];
1280 for (int i
= 0; i
< 3; i
++) {
1281 if (inst
->src
[i
].file
== GRF
)
1282 inst
->src
[i
].reg
= remap_table
[inst
->src
[i
].reg
];
1286 /* Patch all the references to special values */
1287 for (unsigned i
= 0; i
< ARRAY_SIZE(special
); i
++) {
1288 if (special
[i
]->file
== GRF
&& remap_table
[special
[i
]->reg
] != -1)
1289 special
[i
]->reg
= remap_table
[special
[i
]->reg
];
1294 fs_visitor::remove_dead_constants()
1296 if (dispatch_width
== 8) {
1297 this->params_remap
= ralloc_array(mem_ctx
, int, c
->prog_data
.nr_params
);
1299 for (unsigned int i
= 0; i
< c
->prog_data
.nr_params
; i
++)
1300 this->params_remap
[i
] = -1;
1302 /* Find which params are still in use. */
1303 foreach_list(node
, &this->instructions
) {
1304 fs_inst
*inst
= (fs_inst
*)node
;
1306 for (int i
= 0; i
< 3; i
++) {
1307 int constant_nr
= inst
->src
[i
].reg
+ inst
->src
[i
].reg_offset
;
1309 if (inst
->src
[i
].file
!= UNIFORM
)
1312 assert(constant_nr
< (int)c
->prog_data
.nr_params
);
1314 /* For now, set this to non-negative. We'll give it the
1315 * actual new number in a moment, in order to keep the
1316 * register numbers nicely ordered.
1318 this->params_remap
[constant_nr
] = 0;
1322 /* Figure out what the new numbers for the params will be. At some
1323 * point when we're doing uniform array access, we're going to want
1324 * to keep the distinction between .reg and .reg_offset, but for
1325 * now we don't care.
1327 unsigned int new_nr_params
= 0;
1328 for (unsigned int i
= 0; i
< c
->prog_data
.nr_params
; i
++) {
1329 if (this->params_remap
[i
] != -1) {
1330 this->params_remap
[i
] = new_nr_params
++;
1334 /* Update the list of params to be uploaded to match our new numbering. */
1335 for (unsigned int i
= 0; i
< c
->prog_data
.nr_params
; i
++) {
1336 int remapped
= this->params_remap
[i
];
1341 /* We've already done setup_paramvalues_refs() so no need to worry
1342 * about param_index and param_offset.
1344 c
->prog_data
.param
[remapped
] = c
->prog_data
.param
[i
];
1347 c
->prog_data
.nr_params
= new_nr_params
;
1349 /* This should have been generated in the 8-wide pass already. */
1350 assert(this->params_remap
);
1353 /* Now do the renumbering of the shader to remove unused params. */
1354 foreach_list(node
, &this->instructions
) {
1355 fs_inst
*inst
= (fs_inst
*)node
;
1357 for (int i
= 0; i
< 3; i
++) {
1358 int constant_nr
= inst
->src
[i
].reg
+ inst
->src
[i
].reg_offset
;
1360 if (inst
->src
[i
].file
!= UNIFORM
)
1363 assert(this->params_remap
[constant_nr
] != -1);
1364 inst
->src
[i
].reg
= this->params_remap
[constant_nr
];
1365 inst
->src
[i
].reg_offset
= 0;
1373 * Choose accesses from the UNIFORM file to demote to using the pull
1376 * We allow a fragment shader to have more than the specified minimum
1377 * maximum number of fragment shader uniform components (64). If
1378 * there are too many of these, they'd fill up all of register space.
1379 * So, this will push some of them out to the pull constant buffer and
1380 * update the program to load them.
1383 fs_visitor::setup_pull_constants()
1385 /* Only allow 16 registers (128 uniform components) as push constants. */
1386 unsigned int max_uniform_components
= 16 * 8;
1387 if (c
->prog_data
.nr_params
<= max_uniform_components
)
1390 if (dispatch_width
== 16) {
1391 fail("Pull constants not supported in 16-wide\n");
1395 /* Just demote the end of the list. We could probably do better
1396 * here, demoting things that are rarely used in the program first.
1398 int pull_uniform_base
= max_uniform_components
;
1399 int pull_uniform_count
= c
->prog_data
.nr_params
- pull_uniform_base
;
1401 foreach_list(node
, &this->instructions
) {
1402 fs_inst
*inst
= (fs_inst
*)node
;
1404 for (int i
= 0; i
< 3; i
++) {
1405 if (inst
->src
[i
].file
!= UNIFORM
)
1408 int uniform_nr
= inst
->src
[i
].reg
+ inst
->src
[i
].reg_offset
;
1409 if (uniform_nr
< pull_uniform_base
)
1412 fs_reg dst
= fs_reg(this, glsl_type::float_type
);
1413 fs_reg index
= fs_reg((unsigned)SURF_INDEX_FRAG_CONST_BUFFER
);
1414 fs_reg offset
= fs_reg((unsigned)(((uniform_nr
-
1415 pull_uniform_base
) * 4) & ~15));
1417 new(mem_ctx
) fs_inst(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
,
1418 dst
, index
, offset
);
1419 pull
->ir
= inst
->ir
;
1420 pull
->annotation
= inst
->annotation
;
1421 pull
->base_mrf
= 14;
1424 inst
->insert_before(pull
);
1426 inst
->src
[i
].file
= GRF
;
1427 inst
->src
[i
].reg
= dst
.reg
;
1428 inst
->src
[i
].reg_offset
= 0;
1429 inst
->src
[i
].smear
= (uniform_nr
- pull_uniform_base
) & 3;
1433 for (int i
= 0; i
< pull_uniform_count
; i
++) {
1434 c
->prog_data
.pull_param
[i
] = c
->prog_data
.param
[pull_uniform_base
+ i
];
1436 c
->prog_data
.nr_params
-= pull_uniform_count
;
1437 c
->prog_data
.nr_pull_params
= pull_uniform_count
;
1441 fs_visitor::opt_algebraic()
1443 bool progress
= false;
1445 foreach_list(node
, &this->instructions
) {
1446 fs_inst
*inst
= (fs_inst
*)node
;
1448 switch (inst
->opcode
) {
1449 case BRW_OPCODE_MUL
:
1450 if (inst
->src
[1].file
!= IMM
)
1454 if (inst
->src
[1].is_one()) {
1455 inst
->opcode
= BRW_OPCODE_MOV
;
1456 inst
->src
[1] = reg_undef
;
1462 if (inst
->src
[1].is_zero()) {
1463 inst
->opcode
= BRW_OPCODE_MOV
;
1464 inst
->src
[0] = inst
->src
[1];
1465 inst
->src
[1] = reg_undef
;
1471 case BRW_OPCODE_ADD
:
1472 if (inst
->src
[1].file
!= IMM
)
1476 if (inst
->src
[1].is_zero()) {
1477 inst
->opcode
= BRW_OPCODE_MOV
;
1478 inst
->src
[1] = reg_undef
;
1492 * Must be called after calculate_live_intervales() to remove unused
1493 * writes to registers -- register allocation will fail otherwise
1494 * because something deffed but not used won't be considered to
1495 * interfere with other regs.
1498 fs_visitor::dead_code_eliminate()
1500 bool progress
= false;
1503 calculate_live_intervals();
1505 foreach_list_safe(node
, &this->instructions
) {
1506 fs_inst
*inst
= (fs_inst
*)node
;
1508 if (inst
->dst
.file
== GRF
&& this->virtual_grf_use
[inst
->dst
.reg
] <= pc
) {
1517 live_intervals_valid
= false;
1523 * Implements a second type of register coalescing: This one checks if
1524 * the two regs involved in a raw move don't interfere, in which case
1525 * they can both by stored in the same place and the MOV removed.
1528 fs_visitor::register_coalesce_2()
1530 bool progress
= false;
1532 calculate_live_intervals();
1534 foreach_list_safe(node
, &this->instructions
) {
1535 fs_inst
*inst
= (fs_inst
*)node
;
1537 if (inst
->opcode
!= BRW_OPCODE_MOV
||
1540 inst
->src
[0].file
!= GRF
||
1541 inst
->src
[0].negate
||
1543 inst
->src
[0].smear
!= -1 ||
1544 inst
->dst
.file
!= GRF
||
1545 inst
->dst
.type
!= inst
->src
[0].type
||
1546 virtual_grf_sizes
[inst
->src
[0].reg
] != 1 ||
1547 virtual_grf_interferes(inst
->dst
.reg
, inst
->src
[0].reg
)) {
1551 int reg_from
= inst
->src
[0].reg
;
1552 assert(inst
->src
[0].reg_offset
== 0);
1553 int reg_to
= inst
->dst
.reg
;
1554 int reg_to_offset
= inst
->dst
.reg_offset
;
1556 foreach_list_safe(node
, &this->instructions
) {
1557 fs_inst
*scan_inst
= (fs_inst
*)node
;
1559 if (scan_inst
->dst
.file
== GRF
&&
1560 scan_inst
->dst
.reg
== reg_from
) {
1561 scan_inst
->dst
.reg
= reg_to
;
1562 scan_inst
->dst
.reg_offset
= reg_to_offset
;
1564 for (int i
= 0; i
< 3; i
++) {
1565 if (scan_inst
->src
[i
].file
== GRF
&&
1566 scan_inst
->src
[i
].reg
== reg_from
) {
1567 scan_inst
->src
[i
].reg
= reg_to
;
1568 scan_inst
->src
[i
].reg_offset
= reg_to_offset
;
1574 live_intervals_valid
= false;
1583 fs_visitor::register_coalesce()
1585 bool progress
= false;
1589 foreach_list_safe(node
, &this->instructions
) {
1590 fs_inst
*inst
= (fs_inst
*)node
;
1592 /* Make sure that we dominate the instructions we're going to
1593 * scan for interfering with our coalescing, or we won't have
1594 * scanned enough to see if anything interferes with our
1595 * coalescing. We don't dominate the following instructions if
1596 * we're in a loop or an if block.
1598 switch (inst
->opcode
) {
1602 case BRW_OPCODE_WHILE
:
1608 case BRW_OPCODE_ENDIF
:
1614 if (loop_depth
|| if_depth
)
1617 if (inst
->opcode
!= BRW_OPCODE_MOV
||
1620 inst
->dst
.file
!= GRF
|| (inst
->src
[0].file
!= GRF
&&
1621 inst
->src
[0].file
!= UNIFORM
)||
1622 inst
->dst
.type
!= inst
->src
[0].type
)
1625 bool has_source_modifiers
= inst
->src
[0].abs
|| inst
->src
[0].negate
;
1627 /* Found a move of a GRF to a GRF. Let's see if we can coalesce
1628 * them: check for no writes to either one until the exit of the
1631 bool interfered
= false;
1633 for (fs_inst
*scan_inst
= (fs_inst
*)inst
->next
;
1634 !scan_inst
->is_tail_sentinel();
1635 scan_inst
= (fs_inst
*)scan_inst
->next
) {
1636 if (scan_inst
->dst
.file
== GRF
) {
1637 if (scan_inst
->overwrites_reg(inst
->dst
) ||
1638 scan_inst
->overwrites_reg(inst
->src
[0])) {
1644 /* The gen6 MATH instruction can't handle source modifiers or
1645 * unusual register regions, so avoid coalescing those for
1646 * now. We should do something more specific.
1648 if (intel
->gen
== 6 &&
1649 scan_inst
->is_math() &&
1650 (has_source_modifiers
|| inst
->src
[0].file
== UNIFORM
)) {
1655 /* The accumulator result appears to get used for the
1656 * conditional modifier generation. When negating a UD
1657 * value, there is a 33rd bit generated for the sign in the
1658 * accumulator value, so now you can't check, for example,
1659 * equality with a 32-bit value. See piglit fs-op-neg-uint.
1661 if (scan_inst
->conditional_mod
&&
1662 inst
->src
[0].negate
&&
1663 inst
->src
[0].type
== BRW_REGISTER_TYPE_UD
) {
1672 /* Rewrite the later usage to point at the source of the move to
1675 for (fs_inst
*scan_inst
= inst
;
1676 !scan_inst
->is_tail_sentinel();
1677 scan_inst
= (fs_inst
*)scan_inst
->next
) {
1678 for (int i
= 0; i
< 3; i
++) {
1679 if (scan_inst
->src
[i
].file
== GRF
&&
1680 scan_inst
->src
[i
].reg
== inst
->dst
.reg
&&
1681 scan_inst
->src
[i
].reg_offset
== inst
->dst
.reg_offset
) {
1682 fs_reg new_src
= inst
->src
[0];
1683 if (scan_inst
->src
[i
].abs
) {
1687 new_src
.negate
^= scan_inst
->src
[i
].negate
;
1688 scan_inst
->src
[i
] = new_src
;
1698 live_intervals_valid
= false;
1705 fs_visitor::compute_to_mrf()
1707 bool progress
= false;
1710 calculate_live_intervals();
1712 foreach_list_safe(node
, &this->instructions
) {
1713 fs_inst
*inst
= (fs_inst
*)node
;
1718 if (inst
->opcode
!= BRW_OPCODE_MOV
||
1720 inst
->dst
.file
!= MRF
|| inst
->src
[0].file
!= GRF
||
1721 inst
->dst
.type
!= inst
->src
[0].type
||
1722 inst
->src
[0].abs
|| inst
->src
[0].negate
|| inst
->src
[0].smear
!= -1)
1725 /* Work out which hardware MRF registers are written by this
1728 int mrf_low
= inst
->dst
.reg
& ~BRW_MRF_COMPR4
;
1730 if (inst
->dst
.reg
& BRW_MRF_COMPR4
) {
1731 mrf_high
= mrf_low
+ 4;
1732 } else if (dispatch_width
== 16 &&
1733 (!inst
->force_uncompressed
&& !inst
->force_sechalf
)) {
1734 mrf_high
= mrf_low
+ 1;
1739 /* Can't compute-to-MRF this GRF if someone else was going to
1742 if (this->virtual_grf_use
[inst
->src
[0].reg
] > ip
)
1745 /* Found a move of a GRF to a MRF. Let's see if we can go
1746 * rewrite the thing that made this GRF to write into the MRF.
1749 for (scan_inst
= (fs_inst
*)inst
->prev
;
1750 scan_inst
->prev
!= NULL
;
1751 scan_inst
= (fs_inst
*)scan_inst
->prev
) {
1752 if (scan_inst
->dst
.file
== GRF
&&
1753 scan_inst
->dst
.reg
== inst
->src
[0].reg
) {
1754 /* Found the last thing to write our reg we want to turn
1755 * into a compute-to-MRF.
1758 /* SENDs can only write to GRFs, so no compute-to-MRF. */
1759 if (scan_inst
->mlen
) {
1763 /* If it's predicated, it (probably) didn't populate all
1764 * the channels. We might be able to rewrite everything
1765 * that writes that reg, but it would require smarter
1766 * tracking to delay the rewriting until complete success.
1768 if (scan_inst
->predicate
)
1771 /* If it's half of register setup and not the same half as
1772 * our MOV we're trying to remove, bail for now.
1774 if (scan_inst
->force_uncompressed
!= inst
->force_uncompressed
||
1775 scan_inst
->force_sechalf
!= inst
->force_sechalf
) {
1779 /* SEND instructions can't have MRF as a destination. */
1780 if (scan_inst
->mlen
)
1783 if (intel
->gen
>= 6) {
1784 /* gen6 math instructions must have the destination be
1785 * GRF, so no compute-to-MRF for them.
1787 if (scan_inst
->is_math()) {
1792 if (scan_inst
->dst
.reg_offset
== inst
->src
[0].reg_offset
) {
1793 /* Found the creator of our MRF's source value. */
1794 scan_inst
->dst
.file
= MRF
;
1795 scan_inst
->dst
.reg
= inst
->dst
.reg
;
1796 scan_inst
->saturate
|= inst
->saturate
;
1803 /* We don't handle flow control here. Most computation of
1804 * values that end up in MRFs are shortly before the MRF
1807 if (scan_inst
->opcode
== BRW_OPCODE_DO
||
1808 scan_inst
->opcode
== BRW_OPCODE_WHILE
||
1809 scan_inst
->opcode
== BRW_OPCODE_ELSE
||
1810 scan_inst
->opcode
== BRW_OPCODE_ENDIF
) {
1814 /* You can't read from an MRF, so if someone else reads our
1815 * MRF's source GRF that we wanted to rewrite, that stops us.
1817 bool interfered
= false;
1818 for (int i
= 0; i
< 3; i
++) {
1819 if (scan_inst
->src
[i
].file
== GRF
&&
1820 scan_inst
->src
[i
].reg
== inst
->src
[0].reg
&&
1821 scan_inst
->src
[i
].reg_offset
== inst
->src
[0].reg_offset
) {
1828 if (scan_inst
->dst
.file
== MRF
) {
1829 /* If somebody else writes our MRF here, we can't
1830 * compute-to-MRF before that.
1832 int scan_mrf_low
= scan_inst
->dst
.reg
& ~BRW_MRF_COMPR4
;
1835 if (scan_inst
->dst
.reg
& BRW_MRF_COMPR4
) {
1836 scan_mrf_high
= scan_mrf_low
+ 4;
1837 } else if (dispatch_width
== 16 &&
1838 (!scan_inst
->force_uncompressed
&&
1839 !scan_inst
->force_sechalf
)) {
1840 scan_mrf_high
= scan_mrf_low
+ 1;
1842 scan_mrf_high
= scan_mrf_low
;
1845 if (mrf_low
== scan_mrf_low
||
1846 mrf_low
== scan_mrf_high
||
1847 mrf_high
== scan_mrf_low
||
1848 mrf_high
== scan_mrf_high
) {
1853 if (scan_inst
->mlen
> 0) {
1854 /* Found a SEND instruction, which means that there are
1855 * live values in MRFs from base_mrf to base_mrf +
1856 * scan_inst->mlen - 1. Don't go pushing our MRF write up
1859 if (mrf_low
>= scan_inst
->base_mrf
&&
1860 mrf_low
< scan_inst
->base_mrf
+ scan_inst
->mlen
) {
1863 if (mrf_high
>= scan_inst
->base_mrf
&&
1864 mrf_high
< scan_inst
->base_mrf
+ scan_inst
->mlen
) {
1872 live_intervals_valid
= false;
1878 * Walks through basic blocks, looking for repeated MRF writes and
1879 * removing the later ones.
1882 fs_visitor::remove_duplicate_mrf_writes()
1884 fs_inst
*last_mrf_move
[16];
1885 bool progress
= false;
1887 /* Need to update the MRF tracking for compressed instructions. */
1888 if (dispatch_width
== 16)
1891 memset(last_mrf_move
, 0, sizeof(last_mrf_move
));
1893 foreach_list_safe(node
, &this->instructions
) {
1894 fs_inst
*inst
= (fs_inst
*)node
;
1896 switch (inst
->opcode
) {
1898 case BRW_OPCODE_WHILE
:
1900 case BRW_OPCODE_ELSE
:
1901 case BRW_OPCODE_ENDIF
:
1902 memset(last_mrf_move
, 0, sizeof(last_mrf_move
));
1908 if (inst
->opcode
== BRW_OPCODE_MOV
&&
1909 inst
->dst
.file
== MRF
) {
1910 fs_inst
*prev_inst
= last_mrf_move
[inst
->dst
.reg
];
1911 if (prev_inst
&& inst
->equals(prev_inst
)) {
1918 /* Clear out the last-write records for MRFs that were overwritten. */
1919 if (inst
->dst
.file
== MRF
) {
1920 last_mrf_move
[inst
->dst
.reg
] = NULL
;
1923 if (inst
->mlen
> 0) {
1924 /* Found a SEND instruction, which will include two or fewer
1925 * implied MRF writes. We could do better here.
1927 for (int i
= 0; i
< implied_mrf_writes(inst
); i
++) {
1928 last_mrf_move
[inst
->base_mrf
+ i
] = NULL
;
1932 /* Clear out any MRF move records whose sources got overwritten. */
1933 if (inst
->dst
.file
== GRF
) {
1934 for (unsigned int i
= 0; i
< Elements(last_mrf_move
); i
++) {
1935 if (last_mrf_move
[i
] &&
1936 last_mrf_move
[i
]->src
[0].reg
== inst
->dst
.reg
) {
1937 last_mrf_move
[i
] = NULL
;
1942 if (inst
->opcode
== BRW_OPCODE_MOV
&&
1943 inst
->dst
.file
== MRF
&&
1944 inst
->src
[0].file
== GRF
&&
1946 last_mrf_move
[inst
->dst
.reg
] = inst
;
1951 live_intervals_valid
= false;
1957 fs_visitor::dump_instruction(fs_inst
*inst
)
1959 if (inst
->opcode
< ARRAY_SIZE(opcode_descs
) &&
1960 opcode_descs
[inst
->opcode
].name
) {
1961 printf("%s", opcode_descs
[inst
->opcode
].name
);
1963 printf("op%d", inst
->opcode
);
1969 switch (inst
->dst
.file
) {
1971 printf("vgrf%d", inst
->dst
.reg
);
1972 if (inst
->dst
.reg_offset
)
1973 printf("+%d", inst
->dst
.reg_offset
);
1976 printf("m%d", inst
->dst
.reg
);
1982 printf("***u%d***", inst
->dst
.reg
);
1990 for (int i
= 0; i
< 3; i
++) {
1991 if (inst
->src
[i
].negate
)
1993 if (inst
->src
[i
].abs
)
1995 switch (inst
->src
[i
].file
) {
1997 printf("vgrf%d", inst
->src
[i
].reg
);
1998 if (inst
->src
[i
].reg_offset
)
1999 printf("+%d", inst
->src
[i
].reg_offset
);
2002 printf("***m%d***", inst
->src
[i
].reg
);
2005 printf("u%d", inst
->src
[i
].reg
);
2006 if (inst
->src
[i
].reg_offset
)
2007 printf(".%d", inst
->src
[i
].reg_offset
);
2016 if (inst
->src
[i
].abs
)
2025 if (inst
->force_uncompressed
)
2028 if (inst
->force_sechalf
)
2035 fs_visitor::dump_instructions()
2038 foreach_list(node
, &this->instructions
) {
2039 fs_inst
*inst
= (fs_inst
*)node
;
2040 printf("%d: ", ip
++);
2041 dump_instruction(inst
);
2046 * Possibly returns an instruction that set up @param reg.
2048 * Sometimes we want to take the result of some expression/variable
2049 * dereference tree and rewrite the instruction generating the result
2050 * of the tree. When processing the tree, we know that the
2051 * instructions generated are all writing temporaries that are dead
2052 * outside of this tree. So, if we have some instructions that write
2053 * a temporary, we're free to point that temp write somewhere else.
2055 * Note that this doesn't guarantee that the instruction generated
2056 * only reg -- it might be the size=4 destination of a texture instruction.
2059 fs_visitor::get_instruction_generating_reg(fs_inst
*start
,
2065 end
->force_uncompressed
||
2066 end
->force_sechalf
||
2067 !reg
.equals(end
->dst
)) {
2075 fs_visitor::setup_payload_gen6()
2077 struct intel_context
*intel
= &brw
->intel
;
2079 (fp
->Base
.InputsRead
& (1 << FRAG_ATTRIB_WPOS
)) != 0;
2080 unsigned barycentric_interp_modes
= c
->prog_data
.barycentric_interp_modes
;
2082 assert(intel
->gen
>= 6);
2084 /* R0-1: masks, pixel X/Y coordinates. */
2085 c
->nr_payload_regs
= 2;
2086 /* R2: only for 32-pixel dispatch.*/
2088 /* R3-26: barycentric interpolation coordinates. These appear in the
2089 * same order that they appear in the brw_wm_barycentric_interp_mode
2090 * enum. Each set of coordinates occupies 2 registers if dispatch width
2091 * == 8 and 4 registers if dispatch width == 16. Coordinates only
2092 * appear if they were enabled using the "Barycentric Interpolation
2093 * Mode" bits in WM_STATE.
2095 for (int i
= 0; i
< BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT
; ++i
) {
2096 if (barycentric_interp_modes
& (1 << i
)) {
2097 c
->barycentric_coord_reg
[i
] = c
->nr_payload_regs
;
2098 c
->nr_payload_regs
+= 2;
2099 if (dispatch_width
== 16) {
2100 c
->nr_payload_regs
+= 2;
2105 /* R27: interpolated depth if uses source depth */
2107 c
->source_depth_reg
= c
->nr_payload_regs
;
2108 c
->nr_payload_regs
++;
2109 if (dispatch_width
== 16) {
2110 /* R28: interpolated depth if not 8-wide. */
2111 c
->nr_payload_regs
++;
2114 /* R29: interpolated W set if GEN6_WM_USES_SOURCE_W. */
2116 c
->source_w_reg
= c
->nr_payload_regs
;
2117 c
->nr_payload_regs
++;
2118 if (dispatch_width
== 16) {
2119 /* R30: interpolated W if not 8-wide. */
2120 c
->nr_payload_regs
++;
2123 /* R31: MSAA position offsets. */
2124 /* R32-: bary for 32-pixel. */
2125 /* R58-59: interp W for 32-pixel. */
2127 if (fp
->Base
.OutputsWritten
& BITFIELD64_BIT(FRAG_RESULT_DEPTH
)) {
2128 c
->source_depth_to_render_target
= true;
2135 uint32_t orig_nr_params
= c
->prog_data
.nr_params
;
2137 if (intel
->gen
>= 6)
2138 setup_payload_gen6();
2140 setup_payload_gen4();
2145 calculate_urb_setup();
2147 emit_interpolation_setup_gen4();
2149 emit_interpolation_setup_gen6();
2151 /* Generate FS IR for main(). (the visitor only descends into
2152 * functions called "main").
2155 foreach_list(node
, &*shader
->ir
) {
2156 ir_instruction
*ir
= (ir_instruction
*)node
;
2158 this->result
= reg_undef
;
2162 emit_fragment_program_code();
2170 split_virtual_grfs();
2172 setup_paramvalues_refs();
2173 setup_pull_constants();
2179 compact_virtual_grfs();
2181 progress
= remove_duplicate_mrf_writes() || progress
;
2183 progress
= opt_algebraic() || progress
;
2184 progress
= opt_cse() || progress
;
2185 progress
= opt_copy_propagate() || progress
;
2186 progress
= dead_code_eliminate() || progress
;
2187 progress
= register_coalesce() || progress
;
2188 progress
= register_coalesce_2() || progress
;
2189 progress
= compute_to_mrf() || progress
;
2192 remove_dead_constants();
2194 schedule_instructions();
2196 assign_curb_setup();
2200 /* Debug of register spilling: Go spill everything. */
2201 for (int i
= 0; i
< virtual_grf_count
; i
++) {
2207 assign_regs_trivial();
2209 while (!assign_regs()) {
2215 assert(force_uncompressed_stack
== 0);
2216 assert(force_sechalf_stack
== 0);
2221 if (dispatch_width
== 8) {
2222 c
->prog_data
.reg_blocks
= brw_register_blocks(grf_used
);
2224 c
->prog_data
.reg_blocks_16
= brw_register_blocks(grf_used
);
2226 /* Make sure we didn't try to sneak in an extra uniform */
2227 assert(orig_nr_params
== c
->prog_data
.nr_params
);
2228 (void) orig_nr_params
;
2235 brw_wm_fs_emit(struct brw_context
*brw
, struct brw_wm_compile
*c
,
2236 struct gl_fragment_program
*fp
,
2237 struct gl_shader_program
*prog
,
2238 unsigned *final_assembly_size
)
2240 struct intel_context
*intel
= &brw
->intel
;
2241 bool start_busy
= false;
2242 float start_time
= 0;
2244 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
)) {
2245 start_busy
= (intel
->batch
.last_bo
&&
2246 drm_intel_bo_busy(intel
->batch
.last_bo
));
2247 start_time
= get_time();
2250 struct brw_shader
*shader
= NULL
;
2252 shader
= (brw_shader
*) prog
->_LinkedShaders
[MESA_SHADER_FRAGMENT
];
2254 if (unlikely(INTEL_DEBUG
& DEBUG_WM
)) {
2256 printf("GLSL IR for native fragment shader %d:\n", prog
->Name
);
2257 _mesa_print_ir(shader
->ir
, NULL
);
2260 printf("ARB_fragment_program %d ir for native fragment shader\n",
2262 _mesa_print_program(&fp
->Base
);
2266 /* Now the main event: Visit the shader IR and generate our FS IR for it.
2268 fs_visitor
v(brw
, c
, prog
, fp
, 8);
2270 prog
->LinkStatus
= false;
2271 ralloc_strcat(&prog
->InfoLog
, v
.fail_msg
);
2273 _mesa_problem(NULL
, "Failed to compile fragment shader: %s\n",
2279 exec_list
*simd16_instructions
= NULL
;
2280 fs_visitor
v2(brw
, c
, prog
, fp
, 16);
2281 if (intel
->gen
>= 5 && c
->prog_data
.nr_pull_params
== 0) {
2282 v2
.import_uniforms(&v
);
2284 perf_debug("16-wide shader failed to compile, falling back to "
2285 "8-wide at a 10-20%% performance cost: %s", v2
.fail_msg
);
2287 simd16_instructions
= &v2
.instructions
;
2291 c
->prog_data
.dispatch_width
= 8;
2293 fs_generator
g(brw
, c
, prog
, fp
, v
.dual_src_output
.file
!= BAD_FILE
);
2294 const unsigned *generated
= g
.generate_assembly(&v
.instructions
,
2295 simd16_instructions
,
2296 final_assembly_size
);
2298 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
) && shader
) {
2299 if (shader
->compiled_once
)
2300 brw_wm_debug_recompile(brw
, prog
, &c
->key
);
2301 shader
->compiled_once
= true;
2303 if (start_busy
&& !drm_intel_bo_busy(intel
->batch
.last_bo
)) {
2304 perf_debug("FS compile took %.03f ms and stalled the GPU\n",
2305 (get_time() - start_time
) * 1000);
2313 brw_fs_precompile(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
2315 struct brw_context
*brw
= brw_context(ctx
);
2316 struct intel_context
*intel
= &brw
->intel
;
2317 struct brw_wm_prog_key key
;
2319 if (!prog
->_LinkedShaders
[MESA_SHADER_FRAGMENT
])
2322 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)
2323 prog
->_LinkedShaders
[MESA_SHADER_FRAGMENT
]->Program
;
2324 struct brw_fragment_program
*bfp
= brw_fragment_program(fp
);
2325 bool program_uses_dfdy
= fp
->UsesDFdy
;
2327 memset(&key
, 0, sizeof(key
));
2329 if (intel
->gen
< 6) {
2331 key
.iz_lookup
|= IZ_PS_KILL_ALPHATEST_BIT
;
2333 if (fp
->Base
.OutputsWritten
& BITFIELD64_BIT(FRAG_RESULT_DEPTH
))
2334 key
.iz_lookup
|= IZ_PS_COMPUTES_DEPTH_BIT
;
2336 /* Just assume depth testing. */
2337 key
.iz_lookup
|= IZ_DEPTH_TEST_ENABLE_BIT
;
2338 key
.iz_lookup
|= IZ_DEPTH_WRITE_ENABLE_BIT
;
2341 if (prog
->Name
!= 0)
2342 key
.proj_attrib_mask
= 0xffffffff;
2345 key
.vp_outputs_written
|= BITFIELD64_BIT(FRAG_ATTRIB_WPOS
);
2347 for (int i
= 0; i
< FRAG_ATTRIB_MAX
; i
++) {
2348 if (!(fp
->Base
.InputsRead
& BITFIELD64_BIT(i
)))
2351 if (prog
->Name
== 0)
2352 key
.proj_attrib_mask
|= 1 << i
;
2354 if (intel
->gen
< 6) {
2355 int vp_index
= _mesa_vert_result_to_frag_attrib((gl_vert_result
) i
);
2358 key
.vp_outputs_written
|= BITFIELD64_BIT(vp_index
);
2362 key
.clamp_fragment_color
= true;
2364 for (int i
= 0; i
< MAX_SAMPLERS
; i
++) {
2365 if (fp
->Base
.ShadowSamplers
& (1 << i
)) {
2366 /* Assume DEPTH_TEXTURE_MODE is the default: X, X, X, 1 */
2367 key
.tex
.swizzles
[i
] =
2368 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_ONE
);
2370 /* Color sampler: assume no swizzling. */
2371 key
.tex
.swizzles
[i
] = SWIZZLE_XYZW
;
2375 if (fp
->Base
.InputsRead
& FRAG_BIT_WPOS
) {
2376 key
.drawable_height
= ctx
->DrawBuffer
->Height
;
2379 if ((fp
->Base
.InputsRead
& FRAG_BIT_WPOS
) || program_uses_dfdy
) {
2380 key
.render_to_fbo
= _mesa_is_user_fbo(ctx
->DrawBuffer
);
2383 key
.nr_color_regions
= 1;
2385 key
.program_string_id
= bfp
->id
;
2387 uint32_t old_prog_offset
= brw
->wm
.prog_offset
;
2388 struct brw_wm_prog_data
*old_prog_data
= brw
->wm
.prog_data
;
2390 bool success
= do_wm_prog(brw
, prog
, bfp
, &key
);
2392 brw
->wm
.prog_offset
= old_prog_offset
;
2393 brw
->wm
.prog_data
= old_prog_data
;