3 * Copyright © 2010-2015 Intel Corporation
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25 #ifndef BRW_FS_BUILDER_H
26 #define BRW_FS_BUILDER_H
28 #include "brw_ir_fs.h"
29 #include "brw_shader.h"
33 * Toolbox to assemble an FS IR program out of individual instructions.
35 * This object is meant to have an interface consistent with
36 * brw::vec4_builder. They cannot be fully interchangeable because
37 * brw::fs_builder generates scalar code while brw::vec4_builder generates
42 /** Type used in this IR to represent a source of an instruction. */
43 typedef fs_reg src_reg
;
45 /** Type used in this IR to represent the destination of an instruction. */
46 typedef fs_reg dst_reg
;
48 /** Type used in this IR to represent an instruction. */
49 typedef fs_inst instruction
;
52 * Construct an fs_builder that inserts instructions into \p shader.
53 * \p dispatch_width gives the native execution width of the program.
55 fs_builder(backend_shader
*shader
,
56 unsigned dispatch_width
) :
57 shader(shader
), block(NULL
), cursor(NULL
),
58 _dispatch_width(dispatch_width
),
60 force_writemask_all(false),
66 * Construct an fs_builder that inserts instructions into \p shader
67 * before instruction \p inst in basic block \p block. The default
68 * execution controls and debug annotation are initialized from the
69 * instruction passed as argument.
71 fs_builder(backend_shader
*shader
, bblock_t
*block
, fs_inst
*inst
) :
72 shader(shader
), block(block
), cursor(inst
),
73 _dispatch_width(inst
->exec_size
),
75 force_writemask_all(inst
->force_writemask_all
)
77 annotation
.str
= inst
->annotation
;
78 annotation
.ir
= inst
->ir
;
82 * Construct an fs_builder that inserts instructions before \p cursor in
83 * basic block \p block, inheriting other code generation parameters
87 at(bblock_t
*block
, exec_node
*cursor
) const
89 fs_builder bld
= *this;
96 * Construct an fs_builder appending instructions at the end of the
97 * instruction list of the shader, inheriting other code generation
98 * parameters from this.
103 return at(NULL
, (exec_node
*)&shader
->instructions
.tail_sentinel
);
107 * Construct a builder specifying the default SIMD width and group of
108 * channel enable signals, inheriting other code generation parameters
111 * \p n gives the default SIMD width, \p i gives the slot group used for
112 * predication and control flow masking in multiples of \p n channels.
115 group(unsigned n
, unsigned i
) const
117 fs_builder bld
= *this;
119 if (n
<= dispatch_width() && i
< dispatch_width() / n
) {
122 /* The requested channel group isn't a subset of the channel group
123 * of this builder, which means that the resulting instructions
124 * would use (potentially undefined) channel enable signals not
125 * specified by the parent builder. That's only valid if the
126 * instruction doesn't have per-channel semantics, in which case
127 * we should clear off the default group index in order to prevent
128 * emitting instructions with channel group not aligned to their
129 * own execution size.
131 assert(force_writemask_all
);
135 bld
._dispatch_width
= n
;
140 * Alias for group() with width equal to eight.
143 half(unsigned i
) const
149 * Construct a builder with per-channel control flow execution masking
150 * disabled if \p b is true. If control flow execution masking is
151 * already disabled this has no effect.
154 exec_all(bool b
= true) const
156 fs_builder bld
= *this;
158 bld
.force_writemask_all
= true;
163 * Construct a builder with the given debug annotation info.
166 annotate(const char *str
, const void *ir
= NULL
) const
168 fs_builder bld
= *this;
169 bld
.annotation
.str
= str
;
170 bld
.annotation
.ir
= ir
;
175 * Get the SIMD width in use.
178 dispatch_width() const
180 return _dispatch_width
;
184 * Get the channel group in use.
193 * Allocate a virtual register of natural vector size (one for this IR)
194 * and SIMD width. \p n gives the amount of space to allocate in
195 * dispatch_width units (which is just enough space for one logical
196 * component in this IR).
199 vgrf(enum brw_reg_type type
, unsigned n
= 1) const
201 assert(dispatch_width() <= 32);
204 return dst_reg(VGRF
, shader
->alloc
.allocate(
205 DIV_ROUND_UP(n
* type_sz(type
) * dispatch_width(),
209 return retype(null_reg_ud(), type
);
213 * Create a null register of floating type.
218 return dst_reg(retype(brw_null_reg(), BRW_REGISTER_TYPE_F
));
224 return dst_reg(retype(brw_null_reg(), BRW_REGISTER_TYPE_DF
));
228 * Create a null register of signed integer type.
233 return dst_reg(retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
237 * Create a null register of unsigned integer type.
242 return dst_reg(retype(brw_null_reg(), BRW_REGISTER_TYPE_UD
));
246 * Get the mask of SIMD channels enabled by dispatch and not yet
247 * disabled by discard.
250 sample_mask_reg() const
252 if (shader
->stage
!= MESA_SHADER_FRAGMENT
) {
253 return brw_imm_d(0xffffffff);
254 } else if (brw_wm_prog_data(shader
->stage_prog_data
)->uses_kill
) {
255 return brw_flag_reg(0, 1);
257 assert(shader
->devinfo
->gen
>= 6 && dispatch_width() <= 16);
258 return retype(brw_vec1_grf((_group
>= 16 ? 2 : 1), 7),
259 BRW_REGISTER_TYPE_UD
);
264 * Insert an instruction into the program.
267 emit(const instruction
&inst
) const
269 return emit(new(shader
->mem_ctx
) instruction(inst
));
273 * Create and insert a nullary control instruction into the program.
276 emit(enum opcode opcode
) const
278 return emit(instruction(opcode
, dispatch_width()));
282 * Create and insert a nullary instruction into the program.
285 emit(enum opcode opcode
, const dst_reg
&dst
) const
287 return emit(instruction(opcode
, dispatch_width(), dst
));
291 * Create and insert a unary instruction into the program.
294 emit(enum opcode opcode
, const dst_reg
&dst
, const src_reg
&src0
) const
297 case SHADER_OPCODE_RCP
:
298 case SHADER_OPCODE_RSQ
:
299 case SHADER_OPCODE_SQRT
:
300 case SHADER_OPCODE_EXP2
:
301 case SHADER_OPCODE_LOG2
:
302 case SHADER_OPCODE_SIN
:
303 case SHADER_OPCODE_COS
:
304 return emit(instruction(opcode
, dispatch_width(), dst
,
305 fix_math_operand(src0
)));
308 return emit(instruction(opcode
, dispatch_width(), dst
, src0
));
313 * Create and insert a binary instruction into the program.
316 emit(enum opcode opcode
, const dst_reg
&dst
, const src_reg
&src0
,
317 const src_reg
&src1
) const
320 case SHADER_OPCODE_POW
:
321 case SHADER_OPCODE_INT_QUOTIENT
:
322 case SHADER_OPCODE_INT_REMAINDER
:
323 return emit(instruction(opcode
, dispatch_width(), dst
,
324 fix_math_operand(src0
),
325 fix_math_operand(src1
)));
328 return emit(instruction(opcode
, dispatch_width(), dst
, src0
, src1
));
334 * Create and insert a ternary instruction into the program.
337 emit(enum opcode opcode
, const dst_reg
&dst
, const src_reg
&src0
,
338 const src_reg
&src1
, const src_reg
&src2
) const
342 case BRW_OPCODE_BFI2
:
345 return emit(instruction(opcode
, dispatch_width(), dst
,
346 fix_3src_operand(src0
),
347 fix_3src_operand(src1
),
348 fix_3src_operand(src2
)));
351 return emit(instruction(opcode
, dispatch_width(), dst
,
357 * Create and insert an instruction with a variable number of sources
361 emit(enum opcode opcode
, const dst_reg
&dst
, const src_reg srcs
[],
364 return emit(instruction(opcode
, dispatch_width(), dst
, srcs
, n
));
368 * Insert a preallocated instruction into the program.
371 emit(instruction
*inst
) const
373 assert(inst
->exec_size
<= 32);
374 assert(inst
->exec_size
== dispatch_width() ||
375 force_writemask_all
);
377 inst
->group
= _group
;
378 inst
->force_writemask_all
= force_writemask_all
;
379 inst
->annotation
= annotation
.str
;
380 inst
->ir
= annotation
.ir
;
383 static_cast<instruction
*>(cursor
)->insert_before(block
, inst
);
385 cursor
->insert_before(inst
);
391 * Select \p src0 if the comparison of both sources with the given
392 * conditional mod evaluates to true, otherwise select \p src1.
394 * Generally useful to get the minimum or maximum of two values.
397 emit_minmax(const dst_reg
&dst
, const src_reg
&src0
,
398 const src_reg
&src1
, brw_conditional_mod mod
) const
400 assert(mod
== BRW_CONDITIONAL_GE
|| mod
== BRW_CONDITIONAL_L
);
402 return set_condmod(mod
, SEL(dst
, fix_unsigned_negate(src0
),
403 fix_unsigned_negate(src1
)));
407 * Copy any live channel from \p src to the first channel of the result.
410 emit_uniformize(const src_reg
&src
) const
412 /* FIXME: We use a vector chan_index and dst to allow constant and
413 * copy propagration to move result all the way into the consuming
414 * instruction (typically a surface index or sampler index for a
415 * send). This uses 1 or 3 extra hw registers in 16 or 32 wide
416 * dispatch. Once we teach const/copy propagation about scalars we
417 * should go back to scalar destinations here.
419 const fs_builder ubld
= exec_all();
420 const dst_reg chan_index
= vgrf(BRW_REGISTER_TYPE_UD
);
421 const dst_reg dst
= vgrf(src
.type
);
423 ubld
.emit(SHADER_OPCODE_FIND_LIVE_CHANNEL
, chan_index
)->flag_subreg
= 2;
424 ubld
.emit(SHADER_OPCODE_BROADCAST
, dst
, src
, component(chan_index
, 0));
426 return src_reg(component(dst
, 0));
430 move_to_vgrf(const src_reg
&src
, unsigned num_components
) const
432 src_reg
*const src_comps
= new src_reg
[num_components
];
433 for (unsigned i
= 0; i
< num_components
; i
++)
434 src_comps
[i
] = offset(src
, dispatch_width(), i
);
436 const dst_reg dst
= vgrf(src
.type
, num_components
);
437 LOAD_PAYLOAD(dst
, src_comps
, num_components
, 0);
445 emit_scan(enum opcode opcode
, const dst_reg
&tmp
,
446 unsigned cluster_size
, brw_conditional_mod mod
) const
448 assert(dispatch_width() >= 8);
450 /* The instruction splitting code isn't advanced enough to split
451 * these so we need to handle that ourselves.
453 if (dispatch_width() * type_sz(tmp
.type
) > 2 * REG_SIZE
) {
454 const unsigned half_width
= dispatch_width() / 2;
455 const fs_builder ubld
= exec_all().group(half_width
, 0);
457 dst_reg right
= horiz_offset(tmp
, half_width
);
458 ubld
.emit_scan(opcode
, left
, cluster_size
, mod
);
459 ubld
.emit_scan(opcode
, right
, cluster_size
, mod
);
460 if (cluster_size
> half_width
) {
461 src_reg left_comp
= component(left
, half_width
- 1);
462 set_condmod(mod
, ubld
.emit(opcode
, right
, left_comp
, right
));
467 if (cluster_size
> 1) {
468 const fs_builder ubld
= exec_all().group(dispatch_width() / 2, 0);
469 const dst_reg left
= horiz_stride(tmp
, 2);
470 const dst_reg right
= horiz_stride(horiz_offset(tmp
, 1), 2);
471 set_condmod(mod
, ubld
.emit(opcode
, right
, left
, right
));
474 if (cluster_size
> 2) {
475 if (type_sz(tmp
.type
) <= 4) {
476 const fs_builder ubld
=
477 exec_all().group(dispatch_width() / 4, 0);
478 src_reg left
= horiz_stride(horiz_offset(tmp
, 1), 4);
480 dst_reg right
= horiz_stride(horiz_offset(tmp
, 2), 4);
481 set_condmod(mod
, ubld
.emit(opcode
, right
, left
, right
));
483 right
= horiz_stride(horiz_offset(tmp
, 3), 4);
484 set_condmod(mod
, ubld
.emit(opcode
, right
, left
, right
));
486 /* For 64-bit types, we have to do things differently because
487 * the code above would land us with destination strides that
488 * the hardware can't handle. Fortunately, we'll only be
489 * 8-wide in that case and it's the same number of
492 const fs_builder ubld
= exec_all().group(2, 0);
494 for (unsigned i
= 0; i
< dispatch_width(); i
+= 4) {
495 src_reg left
= component(tmp
, i
+ 1);
496 dst_reg right
= horiz_offset(tmp
, i
+ 2);
497 set_condmod(mod
, ubld
.emit(opcode
, right
, left
, right
));
502 if (cluster_size
> 4) {
503 const fs_builder ubld
= exec_all().group(4, 0);
504 src_reg left
= component(tmp
, 3);
505 dst_reg right
= horiz_offset(tmp
, 4);
506 set_condmod(mod
, ubld
.emit(opcode
, right
, left
, right
));
508 if (dispatch_width() > 8) {
509 left
= component(tmp
, 8 + 3);
510 right
= horiz_offset(tmp
, 8 + 4);
511 set_condmod(mod
, ubld
.emit(opcode
, right
, left
, right
));
515 if (cluster_size
> 8 && dispatch_width() > 8) {
516 const fs_builder ubld
= exec_all().group(8, 0);
517 src_reg left
= component(tmp
, 7);
518 dst_reg right
= horiz_offset(tmp
, 8);
519 set_condmod(mod
, ubld
.emit(opcode
, right
, left
, right
));
524 * Assorted arithmetic ops.
529 op(const dst_reg &dst, const src_reg &src0) const \
531 return emit(BRW_OPCODE_##op, dst, src0); \
536 op(const dst_reg &dst, const src_reg &src0, const src_reg &src1) const \
538 return emit(BRW_OPCODE_##op, dst, src0, src1); \
541 #define ALU2_ACC(op) \
543 op(const dst_reg &dst, const src_reg &src0, const src_reg &src1) const \
545 instruction *inst = emit(BRW_OPCODE_##op, dst, src0, src1); \
546 inst->writes_accumulator = true; \
552 op(const dst_reg &dst, const src_reg &src0, const src_reg &src1, \
553 const src_reg &src2) const \
555 return emit(BRW_OPCODE_##op, dst, src0, src1, src2); \
608 * CMP: Sets the low bit of the destination channels with the result
609 * of the comparison, while the upper bits are undefined, and updates
610 * the flag register with the packed 16 bits of the result.
613 CMP(const dst_reg
&dst
, const src_reg
&src0
, const src_reg
&src1
,
614 brw_conditional_mod condition
) const
616 /* Take the instruction:
618 * CMP null<d> src0<f> src1<f>
620 * Original gen4 does type conversion to the destination type
621 * before comparison, producing garbage results for floating
624 * The destination type doesn't matter on newer generations,
625 * so we set the type to match src0 so we can compact the
628 return set_condmod(condition
,
629 emit(BRW_OPCODE_CMP
, retype(dst
, src0
.type
),
630 fix_unsigned_negate(src0
),
631 fix_unsigned_negate(src1
)));
635 * Gen4 predicated IF.
638 IF(brw_predicate predicate
) const
640 return set_predicate(predicate
, emit(BRW_OPCODE_IF
));
644 * CSEL: dst = src2 <op> 0.0f ? src0 : src1
647 CSEL(const dst_reg
&dst
, const src_reg
&src0
, const src_reg
&src1
,
648 const src_reg
&src2
, brw_conditional_mod condition
) const
650 /* CSEL only operates on floats, so we can't do integer </<=/>=/>
651 * comparisons. Zero/non-zero (== and !=) comparisons almost work.
652 * 0x80000000 fails because it is -0.0, and -0.0 == 0.0.
654 assert(src2
.type
== BRW_REGISTER_TYPE_F
);
656 return set_condmod(condition
,
657 emit(BRW_OPCODE_CSEL
,
658 retype(dst
, BRW_REGISTER_TYPE_F
),
659 retype(src0
, BRW_REGISTER_TYPE_F
),
660 retype(src1
, BRW_REGISTER_TYPE_F
),
665 * Emit a linear interpolation instruction.
668 LRP(const dst_reg
&dst
, const src_reg
&x
, const src_reg
&y
,
669 const src_reg
&a
) const
671 if (shader
->devinfo
->gen
>= 6 && shader
->devinfo
->gen
<= 10) {
672 /* The LRP instruction actually does op1 * op0 + op2 * (1 - op0), so
673 * we need to reorder the operands.
675 return emit(BRW_OPCODE_LRP
, dst
, a
, y
, x
);
678 /* We can't use the LRP instruction. Emit x*(1-a) + y*a. */
679 const dst_reg y_times_a
= vgrf(dst
.type
);
680 const dst_reg one_minus_a
= vgrf(dst
.type
);
681 const dst_reg x_times_one_minus_a
= vgrf(dst
.type
);
683 MUL(y_times_a
, y
, a
);
684 ADD(one_minus_a
, negate(a
), brw_imm_f(1.0f
));
685 MUL(x_times_one_minus_a
, x
, src_reg(one_minus_a
));
686 return ADD(dst
, src_reg(x_times_one_minus_a
), src_reg(y_times_a
));
691 * Collect a number of registers in a contiguous range of registers.
694 LOAD_PAYLOAD(const dst_reg
&dst
, const src_reg
*src
,
695 unsigned sources
, unsigned header_size
) const
697 instruction
*inst
= emit(SHADER_OPCODE_LOAD_PAYLOAD
, dst
, src
, sources
);
698 inst
->header_size
= header_size
;
699 inst
->size_written
= header_size
* REG_SIZE
;
700 for (unsigned i
= header_size
; i
< sources
; i
++) {
701 inst
->size_written
+=
702 ALIGN(dispatch_width() * type_sz(src
[i
].type
) * dst
.stride
,
710 UNDEF(const dst_reg
&dst
) const
712 assert(dst
.file
== VGRF
);
713 instruction
*inst
= emit(SHADER_OPCODE_UNDEF
,
714 retype(dst
, BRW_REGISTER_TYPE_UD
));
715 inst
->size_written
= shader
->alloc
.sizes
[dst
.nr
] * REG_SIZE
;
720 backend_shader
*shader
;
724 * Workaround for negation of UD registers. See comment in
725 * fs_generator::generate_code() for more details.
728 fix_unsigned_negate(const src_reg
&src
) const
730 if (src
.type
== BRW_REGISTER_TYPE_UD
&&
732 dst_reg temp
= vgrf(BRW_REGISTER_TYPE_UD
);
734 return src_reg(temp
);
741 * Workaround for source register modes not supported by the ternary
742 * instruction encoding.
745 fix_3src_operand(const src_reg
&src
) const
749 /* FINISHME: Could handle scalar region, other stride=1 regions */
750 if (src
.vstride
!= BRW_VERTICAL_STRIDE_8
||
751 src
.width
!= BRW_WIDTH_8
||
752 src
.hstride
!= BRW_HORIZONTAL_STRIDE_1
)
764 dst_reg expanded
= vgrf(src
.type
);
770 * Workaround for source register modes not supported by the math
774 fix_math_operand(const src_reg
&src
) const
776 /* Can't do hstride == 0 args on gen6 math, so expand it out. We
777 * might be able to do better by doing execsize = 1 math and then
778 * expanding that result out, but we would need to be careful with
781 * Gen6 hardware ignores source modifiers (negate and abs) on math
782 * instructions, so we also move to a temp to set those up.
784 * Gen7 relaxes most of the above restrictions, but still can't use IMM
787 if ((shader
->devinfo
->gen
== 6 &&
788 (src
.file
== IMM
|| src
.file
== UNIFORM
||
789 src
.abs
|| src
.negate
)) ||
790 (shader
->devinfo
->gen
== 7 && src
.file
== IMM
)) {
791 const dst_reg tmp
= vgrf(src
.type
);
802 unsigned _dispatch_width
;
804 bool force_writemask_all
;
806 /** Debug annotation info. */