2 * Copyright © 2016 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
24 #include "vtn_private.h"
27 * Normally, column vectors in SPIR-V correspond to a single NIR SSA
28 * definition. But for matrix multiplies, we want to do one routine for
29 * multiplying a matrix by a matrix and then pretend that vectors are matrices
30 * with one column. So we "wrap" these things, and unwrap the result before we
34 static struct vtn_ssa_value
*
35 wrap_matrix(struct vtn_builder
*b
, struct vtn_ssa_value
*val
)
40 if (glsl_type_is_matrix(val
->type
))
43 struct vtn_ssa_value
*dest
= rzalloc(b
, struct vtn_ssa_value
);
44 dest
->type
= val
->type
;
45 dest
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, 1);
51 static struct vtn_ssa_value
*
52 unwrap_matrix(struct vtn_ssa_value
*val
)
54 if (glsl_type_is_matrix(val
->type
))
60 static struct vtn_ssa_value
*
61 matrix_multiply(struct vtn_builder
*b
,
62 struct vtn_ssa_value
*_src0
, struct vtn_ssa_value
*_src1
)
65 struct vtn_ssa_value
*src0
= wrap_matrix(b
, _src0
);
66 struct vtn_ssa_value
*src1
= wrap_matrix(b
, _src1
);
67 struct vtn_ssa_value
*src0_transpose
= wrap_matrix(b
, _src0
->transposed
);
68 struct vtn_ssa_value
*src1_transpose
= wrap_matrix(b
, _src1
->transposed
);
70 unsigned src0_rows
= glsl_get_vector_elements(src0
->type
);
71 unsigned src0_columns
= glsl_get_matrix_columns(src0
->type
);
72 unsigned src1_columns
= glsl_get_matrix_columns(src1
->type
);
74 const struct glsl_type
*dest_type
;
75 if (src1_columns
> 1) {
76 dest_type
= glsl_matrix_type(glsl_get_base_type(src0
->type
),
77 src0_rows
, src1_columns
);
79 dest_type
= glsl_vector_type(glsl_get_base_type(src0
->type
), src0_rows
);
81 struct vtn_ssa_value
*dest
= vtn_create_ssa_value(b
, dest_type
);
83 dest
= wrap_matrix(b
, dest
);
85 bool transpose_result
= false;
86 if (src0_transpose
&& src1_transpose
) {
87 /* transpose(A) * transpose(B) = transpose(B * A) */
88 src1
= src0_transpose
;
89 src0
= src1_transpose
;
90 src0_transpose
= NULL
;
91 src1_transpose
= NULL
;
92 transpose_result
= true;
95 if (src0_transpose
&& !src1_transpose
&&
96 glsl_get_base_type(src0
->type
) == GLSL_TYPE_FLOAT
) {
97 /* We already have the rows of src0 and the columns of src1 available,
98 * so we can just take the dot product of each row with each column to
102 for (unsigned i
= 0; i
< src1_columns
; i
++) {
103 nir_ssa_def
*vec_src
[4];
104 for (unsigned j
= 0; j
< src0_rows
; j
++) {
105 vec_src
[j
] = nir_fdot(&b
->nb
, src0_transpose
->elems
[j
]->def
,
106 src1
->elems
[i
]->def
);
108 dest
->elems
[i
]->def
= nir_vec(&b
->nb
, vec_src
, src0_rows
);
111 /* We don't handle the case where src1 is transposed but not src0, since
112 * the general case only uses individual components of src1 so the
113 * optimizer should chew through the transpose we emitted for src1.
116 for (unsigned i
= 0; i
< src1_columns
; i
++) {
117 /* dest[i] = sum(src0[j] * src1[i][j] for all j) */
118 dest
->elems
[i
]->def
=
119 nir_fmul(&b
->nb
, src0
->elems
[0]->def
,
120 nir_channel(&b
->nb
, src1
->elems
[i
]->def
, 0));
121 for (unsigned j
= 1; j
< src0_columns
; j
++) {
122 dest
->elems
[i
]->def
=
123 nir_fadd(&b
->nb
, dest
->elems
[i
]->def
,
124 nir_fmul(&b
->nb
, src0
->elems
[j
]->def
,
125 nir_channel(&b
->nb
, src1
->elems
[i
]->def
, j
)));
130 dest
= unwrap_matrix(dest
);
132 if (transpose_result
)
133 dest
= vtn_ssa_transpose(b
, dest
);
138 static struct vtn_ssa_value
*
139 mat_times_scalar(struct vtn_builder
*b
,
140 struct vtn_ssa_value
*mat
,
143 struct vtn_ssa_value
*dest
= vtn_create_ssa_value(b
, mat
->type
);
144 for (unsigned i
= 0; i
< glsl_get_matrix_columns(mat
->type
); i
++) {
145 if (glsl_get_base_type(mat
->type
) == GLSL_TYPE_FLOAT
)
146 dest
->elems
[i
]->def
= nir_fmul(&b
->nb
, mat
->elems
[i
]->def
, scalar
);
148 dest
->elems
[i
]->def
= nir_imul(&b
->nb
, mat
->elems
[i
]->def
, scalar
);
155 vtn_handle_matrix_alu(struct vtn_builder
*b
, SpvOp opcode
,
156 struct vtn_value
*dest
,
157 struct vtn_ssa_value
*src0
, struct vtn_ssa_value
*src1
)
161 dest
->ssa
= vtn_create_ssa_value(b
, src0
->type
);
162 unsigned cols
= glsl_get_matrix_columns(src0
->type
);
163 for (unsigned i
= 0; i
< cols
; i
++)
164 dest
->ssa
->elems
[i
]->def
= nir_fneg(&b
->nb
, src0
->elems
[i
]->def
);
169 dest
->ssa
= vtn_create_ssa_value(b
, src0
->type
);
170 unsigned cols
= glsl_get_matrix_columns(src0
->type
);
171 for (unsigned i
= 0; i
< cols
; i
++)
172 dest
->ssa
->elems
[i
]->def
=
173 nir_fadd(&b
->nb
, src0
->elems
[i
]->def
, src1
->elems
[i
]->def
);
178 dest
->ssa
= vtn_create_ssa_value(b
, src0
->type
);
179 unsigned cols
= glsl_get_matrix_columns(src0
->type
);
180 for (unsigned i
= 0; i
< cols
; i
++)
181 dest
->ssa
->elems
[i
]->def
=
182 nir_fsub(&b
->nb
, src0
->elems
[i
]->def
, src1
->elems
[i
]->def
);
187 dest
->ssa
= vtn_ssa_transpose(b
, src0
);
190 case SpvOpMatrixTimesScalar
:
191 if (src0
->transposed
) {
192 dest
->ssa
= vtn_ssa_transpose(b
, mat_times_scalar(b
, src0
->transposed
,
195 dest
->ssa
= mat_times_scalar(b
, src0
, src1
->def
);
199 case SpvOpVectorTimesMatrix
:
200 case SpvOpMatrixTimesVector
:
201 case SpvOpMatrixTimesMatrix
:
202 if (opcode
== SpvOpVectorTimesMatrix
) {
203 dest
->ssa
= matrix_multiply(b
, vtn_ssa_transpose(b
, src1
), src0
);
205 dest
->ssa
= matrix_multiply(b
, src0
, src1
);
209 default: unreachable("unknown matrix opcode");
214 vtn_handle_bitcast(struct vtn_builder
*b
, struct vtn_ssa_value
*dest
,
215 struct nir_ssa_def
*src
)
217 if (glsl_get_vector_elements(dest
->type
) == src
->num_components
) {
218 /* From the definition of OpBitcast in the SPIR-V 1.2 spec:
220 * "If Result Type has the same number of components as Operand, they
221 * must also have the same component width, and results are computed per
224 dest
->def
= nir_imov(&b
->nb
, src
);
228 /* From the definition of OpBitcast in the SPIR-V 1.2 spec:
230 * "If Result Type has a different number of components than Operand, the
231 * total number of bits in Result Type must equal the total number of bits
232 * in Operand. Let L be the type, either Result Type or Operand’s type, that
233 * has the larger number of components. Let S be the other type, with the
234 * smaller number of components. The number of components in L must be an
235 * integer multiple of the number of components in S. The first component
236 * (that is, the only or lowest-numbered component) of S maps to the first
237 * components of L, and so on, up to the last component of S mapping to the
238 * last components of L. Within this mapping, any single component of S
239 * (mapping to multiple components of L) maps its lower-ordered bits to the
240 * lower-numbered components of L."
242 unsigned src_bit_size
= src
->bit_size
;
243 unsigned dest_bit_size
= glsl_get_bit_size(dest
->type
);
244 unsigned src_components
= src
->num_components
;
245 unsigned dest_components
= glsl_get_vector_elements(dest
->type
);
246 assert(src_bit_size
* src_components
== dest_bit_size
* dest_components
);
248 nir_ssa_def
*dest_chan
[4];
249 if (src_bit_size
> dest_bit_size
) {
250 assert(src_bit_size
% dest_bit_size
== 0);
251 unsigned divisor
= src_bit_size
/ dest_bit_size
;
252 for (unsigned comp
= 0; comp
< src_components
; comp
++) {
253 assert(src_bit_size
== 64);
254 assert(dest_bit_size
== 32);
256 nir_unpack_64_2x32(&b
->nb
, nir_channel(&b
->nb
, src
, comp
));
257 for (unsigned i
= 0; i
< divisor
; i
++)
258 dest_chan
[divisor
* comp
+ i
] = nir_channel(&b
->nb
, split
, i
);
261 assert(dest_bit_size
% src_bit_size
== 0);
262 unsigned divisor
= dest_bit_size
/ src_bit_size
;
263 for (unsigned comp
= 0; comp
< dest_components
; comp
++) {
264 unsigned channels
= ((1 << divisor
) - 1) << (comp
* divisor
);
265 nir_ssa_def
*src_chan
=
266 nir_channels(&b
->nb
, src
, channels
);
267 assert(dest_bit_size
== 64);
268 assert(src_bit_size
== 32);
269 dest_chan
[comp
] = nir_pack_64_2x32(&b
->nb
, src_chan
);
272 dest
->def
= nir_vec(&b
->nb
, dest_chan
, dest_components
);
276 vtn_nir_alu_op_for_spirv_opcode(SpvOp opcode
, bool *swap
,
277 nir_alu_type src
, nir_alu_type dst
)
279 /* Indicates that the first two arguments should be swapped. This is
280 * used for implementing greater-than and less-than-or-equal.
285 case SpvOpSNegate
: return nir_op_ineg
;
286 case SpvOpFNegate
: return nir_op_fneg
;
287 case SpvOpNot
: return nir_op_inot
;
288 case SpvOpIAdd
: return nir_op_iadd
;
289 case SpvOpFAdd
: return nir_op_fadd
;
290 case SpvOpISub
: return nir_op_isub
;
291 case SpvOpFSub
: return nir_op_fsub
;
292 case SpvOpIMul
: return nir_op_imul
;
293 case SpvOpFMul
: return nir_op_fmul
;
294 case SpvOpUDiv
: return nir_op_udiv
;
295 case SpvOpSDiv
: return nir_op_idiv
;
296 case SpvOpFDiv
: return nir_op_fdiv
;
297 case SpvOpUMod
: return nir_op_umod
;
298 case SpvOpSMod
: return nir_op_imod
;
299 case SpvOpFMod
: return nir_op_fmod
;
300 case SpvOpSRem
: return nir_op_irem
;
301 case SpvOpFRem
: return nir_op_frem
;
303 case SpvOpShiftRightLogical
: return nir_op_ushr
;
304 case SpvOpShiftRightArithmetic
: return nir_op_ishr
;
305 case SpvOpShiftLeftLogical
: return nir_op_ishl
;
306 case SpvOpLogicalOr
: return nir_op_ior
;
307 case SpvOpLogicalEqual
: return nir_op_ieq
;
308 case SpvOpLogicalNotEqual
: return nir_op_ine
;
309 case SpvOpLogicalAnd
: return nir_op_iand
;
310 case SpvOpLogicalNot
: return nir_op_inot
;
311 case SpvOpBitwiseOr
: return nir_op_ior
;
312 case SpvOpBitwiseXor
: return nir_op_ixor
;
313 case SpvOpBitwiseAnd
: return nir_op_iand
;
314 case SpvOpSelect
: return nir_op_bcsel
;
315 case SpvOpIEqual
: return nir_op_ieq
;
317 case SpvOpBitFieldInsert
: return nir_op_bitfield_insert
;
318 case SpvOpBitFieldSExtract
: return nir_op_ibitfield_extract
;
319 case SpvOpBitFieldUExtract
: return nir_op_ubitfield_extract
;
320 case SpvOpBitReverse
: return nir_op_bitfield_reverse
;
321 case SpvOpBitCount
: return nir_op_bit_count
;
323 /* The ordered / unordered operators need special implementation besides
324 * the logical operator to use since they also need to check if operands are
327 case SpvOpFOrdEqual
: return nir_op_feq
;
328 case SpvOpFUnordEqual
: return nir_op_feq
;
329 case SpvOpINotEqual
: return nir_op_ine
;
330 case SpvOpFOrdNotEqual
: return nir_op_fne
;
331 case SpvOpFUnordNotEqual
: return nir_op_fne
;
332 case SpvOpULessThan
: return nir_op_ult
;
333 case SpvOpSLessThan
: return nir_op_ilt
;
334 case SpvOpFOrdLessThan
: return nir_op_flt
;
335 case SpvOpFUnordLessThan
: return nir_op_flt
;
336 case SpvOpUGreaterThan
: *swap
= true; return nir_op_ult
;
337 case SpvOpSGreaterThan
: *swap
= true; return nir_op_ilt
;
338 case SpvOpFOrdGreaterThan
: *swap
= true; return nir_op_flt
;
339 case SpvOpFUnordGreaterThan
: *swap
= true; return nir_op_flt
;
340 case SpvOpULessThanEqual
: *swap
= true; return nir_op_uge
;
341 case SpvOpSLessThanEqual
: *swap
= true; return nir_op_ige
;
342 case SpvOpFOrdLessThanEqual
: *swap
= true; return nir_op_fge
;
343 case SpvOpFUnordLessThanEqual
: *swap
= true; return nir_op_fge
;
344 case SpvOpUGreaterThanEqual
: return nir_op_uge
;
345 case SpvOpSGreaterThanEqual
: return nir_op_ige
;
346 case SpvOpFOrdGreaterThanEqual
: return nir_op_fge
;
347 case SpvOpFUnordGreaterThanEqual
: return nir_op_fge
;
350 case SpvOpQuantizeToF16
: return nir_op_fquantize2f16
;
352 case SpvOpConvertFToU
:
353 case SpvOpConvertFToS
:
354 case SpvOpConvertSToF
:
355 case SpvOpConvertUToF
:
358 return nir_type_conversion_op(src
, dst
);
361 case SpvOpDPdx
: return nir_op_fddx
;
362 case SpvOpDPdy
: return nir_op_fddy
;
363 case SpvOpDPdxFine
: return nir_op_fddx_fine
;
364 case SpvOpDPdyFine
: return nir_op_fddy_fine
;
365 case SpvOpDPdxCoarse
: return nir_op_fddx_coarse
;
366 case SpvOpDPdyCoarse
: return nir_op_fddy_coarse
;
369 unreachable("No NIR equivalent");
374 handle_no_contraction(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
375 const struct vtn_decoration
*dec
, void *_void
)
377 assert(dec
->scope
== VTN_DEC_DECORATION
);
378 if (dec
->decoration
!= SpvDecorationNoContraction
)
385 vtn_handle_alu(struct vtn_builder
*b
, SpvOp opcode
,
386 const uint32_t *w
, unsigned count
)
388 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
389 const struct glsl_type
*type
=
390 vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
392 vtn_foreach_decoration(b
, val
, handle_no_contraction
, NULL
);
394 /* Collect the various SSA sources */
395 const unsigned num_inputs
= count
- 3;
396 struct vtn_ssa_value
*vtn_src
[4] = { NULL
, };
397 for (unsigned i
= 0; i
< num_inputs
; i
++)
398 vtn_src
[i
] = vtn_ssa_value(b
, w
[i
+ 3]);
400 if (glsl_type_is_matrix(vtn_src
[0]->type
) ||
401 (num_inputs
>= 2 && glsl_type_is_matrix(vtn_src
[1]->type
))) {
402 vtn_handle_matrix_alu(b
, opcode
, val
, vtn_src
[0], vtn_src
[1]);
407 val
->ssa
= vtn_create_ssa_value(b
, type
);
408 nir_ssa_def
*src
[4] = { NULL
, };
409 for (unsigned i
= 0; i
< num_inputs
; i
++) {
410 assert(glsl_type_is_vector_or_scalar(vtn_src
[i
]->type
));
411 src
[i
] = vtn_src
[i
]->def
;
416 if (src
[0]->num_components
== 1) {
417 val
->ssa
->def
= nir_imov(&b
->nb
, src
[0]);
420 switch (src
[0]->num_components
) {
421 case 2: op
= nir_op_bany_inequal2
; break;
422 case 3: op
= nir_op_bany_inequal3
; break;
423 case 4: op
= nir_op_bany_inequal4
; break;
424 default: unreachable("invalid number of components");
426 val
->ssa
->def
= nir_build_alu(&b
->nb
, op
, src
[0],
427 nir_imm_int(&b
->nb
, NIR_FALSE
),
433 if (src
[0]->num_components
== 1) {
434 val
->ssa
->def
= nir_imov(&b
->nb
, src
[0]);
437 switch (src
[0]->num_components
) {
438 case 2: op
= nir_op_ball_iequal2
; break;
439 case 3: op
= nir_op_ball_iequal3
; break;
440 case 4: op
= nir_op_ball_iequal4
; break;
441 default: unreachable("invalid number of components");
443 val
->ssa
->def
= nir_build_alu(&b
->nb
, op
, src
[0],
444 nir_imm_int(&b
->nb
, NIR_TRUE
),
449 case SpvOpOuterProduct
: {
450 for (unsigned i
= 0; i
< src
[1]->num_components
; i
++) {
451 val
->ssa
->elems
[i
]->def
=
452 nir_fmul(&b
->nb
, src
[0], nir_channel(&b
->nb
, src
[1], i
));
458 val
->ssa
->def
= nir_fdot(&b
->nb
, src
[0], src
[1]);
462 assert(glsl_type_is_struct(val
->ssa
->type
));
463 val
->ssa
->elems
[0]->def
= nir_iadd(&b
->nb
, src
[0], src
[1]);
464 val
->ssa
->elems
[1]->def
= nir_uadd_carry(&b
->nb
, src
[0], src
[1]);
467 case SpvOpISubBorrow
:
468 assert(glsl_type_is_struct(val
->ssa
->type
));
469 val
->ssa
->elems
[0]->def
= nir_isub(&b
->nb
, src
[0], src
[1]);
470 val
->ssa
->elems
[1]->def
= nir_usub_borrow(&b
->nb
, src
[0], src
[1]);
473 case SpvOpUMulExtended
:
474 assert(glsl_type_is_struct(val
->ssa
->type
));
475 val
->ssa
->elems
[0]->def
= nir_imul(&b
->nb
, src
[0], src
[1]);
476 val
->ssa
->elems
[1]->def
= nir_umul_high(&b
->nb
, src
[0], src
[1]);
479 case SpvOpSMulExtended
:
480 assert(glsl_type_is_struct(val
->ssa
->type
));
481 val
->ssa
->elems
[0]->def
= nir_imul(&b
->nb
, src
[0], src
[1]);
482 val
->ssa
->elems
[1]->def
= nir_imul_high(&b
->nb
, src
[0], src
[1]);
486 val
->ssa
->def
= nir_fadd(&b
->nb
,
487 nir_fabs(&b
->nb
, nir_fddx(&b
->nb
, src
[0])),
488 nir_fabs(&b
->nb
, nir_fddy(&b
->nb
, src
[0])));
490 case SpvOpFwidthFine
:
491 val
->ssa
->def
= nir_fadd(&b
->nb
,
492 nir_fabs(&b
->nb
, nir_fddx_fine(&b
->nb
, src
[0])),
493 nir_fabs(&b
->nb
, nir_fddy_fine(&b
->nb
, src
[0])));
495 case SpvOpFwidthCoarse
:
496 val
->ssa
->def
= nir_fadd(&b
->nb
,
497 nir_fabs(&b
->nb
, nir_fddx_coarse(&b
->nb
, src
[0])),
498 nir_fabs(&b
->nb
, nir_fddy_coarse(&b
->nb
, src
[0])));
501 case SpvOpVectorTimesScalar
:
502 /* The builder will take care of splatting for us. */
503 val
->ssa
->def
= nir_fmul(&b
->nb
, src
[0], src
[1]);
507 val
->ssa
->def
= nir_fne(&b
->nb
, src
[0], src
[0]);
511 val
->ssa
->def
= nir_ieq(&b
->nb
, nir_fabs(&b
->nb
, src
[0]),
512 nir_imm_float(&b
->nb
, INFINITY
));
515 case SpvOpFUnordEqual
:
516 case SpvOpFUnordNotEqual
:
517 case SpvOpFUnordLessThan
:
518 case SpvOpFUnordGreaterThan
:
519 case SpvOpFUnordLessThanEqual
:
520 case SpvOpFUnordGreaterThanEqual
: {
522 nir_alu_type src_alu_type
= nir_get_nir_type_for_glsl_type(vtn_src
[0]->type
);
523 nir_alu_type dst_alu_type
= nir_get_nir_type_for_glsl_type(type
);
524 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(opcode
, &swap
, src_alu_type
, dst_alu_type
);
527 nir_ssa_def
*tmp
= src
[0];
534 nir_build_alu(&b
->nb
, op
, src
[0], src
[1], NULL
, NULL
),
536 nir_fne(&b
->nb
, src
[0], src
[0]),
537 nir_fne(&b
->nb
, src
[1], src
[1])));
542 case SpvOpFOrdNotEqual
:
543 case SpvOpFOrdLessThan
:
544 case SpvOpFOrdGreaterThan
:
545 case SpvOpFOrdLessThanEqual
:
546 case SpvOpFOrdGreaterThanEqual
: {
548 nir_alu_type src_alu_type
= nir_get_nir_type_for_glsl_type(vtn_src
[0]->type
);
549 nir_alu_type dst_alu_type
= nir_get_nir_type_for_glsl_type(type
);
550 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(opcode
, &swap
, src_alu_type
, dst_alu_type
);
553 nir_ssa_def
*tmp
= src
[0];
560 nir_build_alu(&b
->nb
, op
, src
[0], src
[1], NULL
, NULL
),
562 nir_feq(&b
->nb
, src
[0], src
[0]),
563 nir_feq(&b
->nb
, src
[1], src
[1])));
568 vtn_handle_bitcast(b
, val
->ssa
, src
[0]);
573 nir_alu_type src_alu_type
= nir_get_nir_type_for_glsl_type(vtn_src
[0]->type
);
574 nir_alu_type dst_alu_type
= nir_get_nir_type_for_glsl_type(type
);
575 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(opcode
, &swap
, src_alu_type
, dst_alu_type
);
578 nir_ssa_def
*tmp
= src
[0];
583 val
->ssa
->def
= nir_build_alu(&b
->nb
, op
, src
[0], src
[1], src
[2], src
[3]);