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
25 #include "vtn_private.h"
26 #include "spirv_info.h"
29 * Normally, column vectors in SPIR-V correspond to a single NIR SSA
30 * definition. But for matrix multiplies, we want to do one routine for
31 * multiplying a matrix by a matrix and then pretend that vectors are matrices
32 * with one column. So we "wrap" these things, and unwrap the result before we
36 static struct vtn_ssa_value
*
37 wrap_matrix(struct vtn_builder
*b
, struct vtn_ssa_value
*val
)
42 if (glsl_type_is_matrix(val
->type
))
45 struct vtn_ssa_value
*dest
= rzalloc(b
, struct vtn_ssa_value
);
46 dest
->type
= val
->type
;
47 dest
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, 1);
53 static struct vtn_ssa_value
*
54 unwrap_matrix(struct vtn_ssa_value
*val
)
56 if (glsl_type_is_matrix(val
->type
))
62 static struct vtn_ssa_value
*
63 matrix_multiply(struct vtn_builder
*b
,
64 struct vtn_ssa_value
*_src0
, struct vtn_ssa_value
*_src1
)
67 struct vtn_ssa_value
*src0
= wrap_matrix(b
, _src0
);
68 struct vtn_ssa_value
*src1
= wrap_matrix(b
, _src1
);
69 struct vtn_ssa_value
*src0_transpose
= wrap_matrix(b
, _src0
->transposed
);
70 struct vtn_ssa_value
*src1_transpose
= wrap_matrix(b
, _src1
->transposed
);
72 unsigned src0_rows
= glsl_get_vector_elements(src0
->type
);
73 unsigned src0_columns
= glsl_get_matrix_columns(src0
->type
);
74 unsigned src1_columns
= glsl_get_matrix_columns(src1
->type
);
76 const struct glsl_type
*dest_type
;
77 if (src1_columns
> 1) {
78 dest_type
= glsl_matrix_type(glsl_get_base_type(src0
->type
),
79 src0_rows
, src1_columns
);
81 dest_type
= glsl_vector_type(glsl_get_base_type(src0
->type
), src0_rows
);
83 struct vtn_ssa_value
*dest
= vtn_create_ssa_value(b
, dest_type
);
85 dest
= wrap_matrix(b
, dest
);
87 bool transpose_result
= false;
88 if (src0_transpose
&& src1_transpose
) {
89 /* transpose(A) * transpose(B) = transpose(B * A) */
90 src1
= src0_transpose
;
91 src0
= src1_transpose
;
92 src0_transpose
= NULL
;
93 src1_transpose
= NULL
;
94 transpose_result
= true;
97 if (src0_transpose
&& !src1_transpose
&&
98 glsl_get_base_type(src0
->type
) == GLSL_TYPE_FLOAT
) {
99 /* We already have the rows of src0 and the columns of src1 available,
100 * so we can just take the dot product of each row with each column to
104 for (unsigned i
= 0; i
< src1_columns
; i
++) {
105 nir_ssa_def
*vec_src
[4];
106 for (unsigned j
= 0; j
< src0_rows
; j
++) {
107 vec_src
[j
] = nir_fdot(&b
->nb
, src0_transpose
->elems
[j
]->def
,
108 src1
->elems
[i
]->def
);
110 dest
->elems
[i
]->def
= nir_vec(&b
->nb
, vec_src
, src0_rows
);
113 /* We don't handle the case where src1 is transposed but not src0, since
114 * the general case only uses individual components of src1 so the
115 * optimizer should chew through the transpose we emitted for src1.
118 for (unsigned i
= 0; i
< src1_columns
; i
++) {
119 /* dest[i] = sum(src0[j] * src1[i][j] for all j) */
120 dest
->elems
[i
]->def
=
121 nir_fmul(&b
->nb
, src0
->elems
[0]->def
,
122 nir_channel(&b
->nb
, src1
->elems
[i
]->def
, 0));
123 for (unsigned j
= 1; j
< src0_columns
; j
++) {
124 dest
->elems
[i
]->def
=
125 nir_fadd(&b
->nb
, dest
->elems
[i
]->def
,
126 nir_fmul(&b
->nb
, src0
->elems
[j
]->def
,
127 nir_channel(&b
->nb
, src1
->elems
[i
]->def
, j
)));
132 dest
= unwrap_matrix(dest
);
134 if (transpose_result
)
135 dest
= vtn_ssa_transpose(b
, dest
);
140 static struct vtn_ssa_value
*
141 mat_times_scalar(struct vtn_builder
*b
,
142 struct vtn_ssa_value
*mat
,
145 struct vtn_ssa_value
*dest
= vtn_create_ssa_value(b
, mat
->type
);
146 for (unsigned i
= 0; i
< glsl_get_matrix_columns(mat
->type
); i
++) {
147 if (glsl_base_type_is_integer(glsl_get_base_type(mat
->type
)))
148 dest
->elems
[i
]->def
= nir_imul(&b
->nb
, mat
->elems
[i
]->def
, scalar
);
150 dest
->elems
[i
]->def
= nir_fmul(&b
->nb
, mat
->elems
[i
]->def
, scalar
);
157 vtn_handle_matrix_alu(struct vtn_builder
*b
, SpvOp opcode
,
158 struct vtn_value
*dest
,
159 struct vtn_ssa_value
*src0
, struct vtn_ssa_value
*src1
)
163 dest
->ssa
= vtn_create_ssa_value(b
, src0
->type
);
164 unsigned cols
= glsl_get_matrix_columns(src0
->type
);
165 for (unsigned i
= 0; i
< cols
; i
++)
166 dest
->ssa
->elems
[i
]->def
= nir_fneg(&b
->nb
, src0
->elems
[i
]->def
);
171 dest
->ssa
= vtn_create_ssa_value(b
, src0
->type
);
172 unsigned cols
= glsl_get_matrix_columns(src0
->type
);
173 for (unsigned i
= 0; i
< cols
; i
++)
174 dest
->ssa
->elems
[i
]->def
=
175 nir_fadd(&b
->nb
, src0
->elems
[i
]->def
, src1
->elems
[i
]->def
);
180 dest
->ssa
= vtn_create_ssa_value(b
, src0
->type
);
181 unsigned cols
= glsl_get_matrix_columns(src0
->type
);
182 for (unsigned i
= 0; i
< cols
; i
++)
183 dest
->ssa
->elems
[i
]->def
=
184 nir_fsub(&b
->nb
, src0
->elems
[i
]->def
, src1
->elems
[i
]->def
);
189 dest
->ssa
= vtn_ssa_transpose(b
, src0
);
192 case SpvOpMatrixTimesScalar
:
193 if (src0
->transposed
) {
194 dest
->ssa
= vtn_ssa_transpose(b
, mat_times_scalar(b
, src0
->transposed
,
197 dest
->ssa
= mat_times_scalar(b
, src0
, src1
->def
);
201 case SpvOpVectorTimesMatrix
:
202 case SpvOpMatrixTimesVector
:
203 case SpvOpMatrixTimesMatrix
:
204 if (opcode
== SpvOpVectorTimesMatrix
) {
205 dest
->ssa
= matrix_multiply(b
, vtn_ssa_transpose(b
, src1
), src0
);
207 dest
->ssa
= matrix_multiply(b
, src0
, src1
);
211 default: vtn_fail_with_opcode("unknown matrix opcode", opcode
);
216 vtn_nir_alu_op_for_spirv_opcode(struct vtn_builder
*b
,
217 SpvOp opcode
, bool *swap
,
218 unsigned src_bit_size
, unsigned dst_bit_size
)
220 /* Indicates that the first two arguments should be swapped. This is
221 * used for implementing greater-than and less-than-or-equal.
226 case SpvOpSNegate
: return nir_op_ineg
;
227 case SpvOpFNegate
: return nir_op_fneg
;
228 case SpvOpNot
: return nir_op_inot
;
229 case SpvOpIAdd
: return nir_op_iadd
;
230 case SpvOpFAdd
: return nir_op_fadd
;
231 case SpvOpISub
: return nir_op_isub
;
232 case SpvOpFSub
: return nir_op_fsub
;
233 case SpvOpIMul
: return nir_op_imul
;
234 case SpvOpFMul
: return nir_op_fmul
;
235 case SpvOpUDiv
: return nir_op_udiv
;
236 case SpvOpSDiv
: return nir_op_idiv
;
237 case SpvOpFDiv
: return nir_op_fdiv
;
238 case SpvOpUMod
: return nir_op_umod
;
239 case SpvOpSMod
: return nir_op_imod
;
240 case SpvOpFMod
: return nir_op_fmod
;
241 case SpvOpSRem
: return nir_op_irem
;
242 case SpvOpFRem
: return nir_op_frem
;
244 case SpvOpShiftRightLogical
: return nir_op_ushr
;
245 case SpvOpShiftRightArithmetic
: return nir_op_ishr
;
246 case SpvOpShiftLeftLogical
: return nir_op_ishl
;
247 case SpvOpLogicalOr
: return nir_op_ior
;
248 case SpvOpLogicalEqual
: return nir_op_ieq
;
249 case SpvOpLogicalNotEqual
: return nir_op_ine
;
250 case SpvOpLogicalAnd
: return nir_op_iand
;
251 case SpvOpLogicalNot
: return nir_op_inot
;
252 case SpvOpBitwiseOr
: return nir_op_ior
;
253 case SpvOpBitwiseXor
: return nir_op_ixor
;
254 case SpvOpBitwiseAnd
: return nir_op_iand
;
255 case SpvOpSelect
: return nir_op_bcsel
;
256 case SpvOpIEqual
: return nir_op_ieq
;
258 case SpvOpBitFieldInsert
: return nir_op_bitfield_insert
;
259 case SpvOpBitFieldSExtract
: return nir_op_ibitfield_extract
;
260 case SpvOpBitFieldUExtract
: return nir_op_ubitfield_extract
;
261 case SpvOpBitReverse
: return nir_op_bitfield_reverse
;
262 case SpvOpBitCount
: return nir_op_bit_count
;
264 case SpvOpUCountLeadingZerosINTEL
: return nir_op_uclz
;
265 /* SpvOpUCountTrailingZerosINTEL is handled elsewhere. */
266 case SpvOpAbsISubINTEL
: return nir_op_uabs_isub
;
267 case SpvOpAbsUSubINTEL
: return nir_op_uabs_usub
;
268 case SpvOpIAddSatINTEL
: return nir_op_iadd_sat
;
269 case SpvOpUAddSatINTEL
: return nir_op_uadd_sat
;
270 case SpvOpIAverageINTEL
: return nir_op_ihadd
;
271 case SpvOpUAverageINTEL
: return nir_op_uhadd
;
272 case SpvOpIAverageRoundedINTEL
: return nir_op_irhadd
;
273 case SpvOpUAverageRoundedINTEL
: return nir_op_urhadd
;
274 case SpvOpISubSatINTEL
: return nir_op_isub_sat
;
275 case SpvOpUSubSatINTEL
: return nir_op_usub_sat
;
276 case SpvOpIMul32x16INTEL
: return nir_op_imul_32x16
;
277 case SpvOpUMul32x16INTEL
: return nir_op_umul_32x16
;
279 /* The ordered / unordered operators need special implementation besides
280 * the logical operator to use since they also need to check if operands are
283 case SpvOpFOrdEqual
: return nir_op_feq
;
284 case SpvOpFUnordEqual
: return nir_op_feq
;
285 case SpvOpINotEqual
: return nir_op_ine
;
286 case SpvOpFOrdNotEqual
: return nir_op_fne
;
287 case SpvOpFUnordNotEqual
: return nir_op_fne
;
288 case SpvOpULessThan
: return nir_op_ult
;
289 case SpvOpSLessThan
: return nir_op_ilt
;
290 case SpvOpFOrdLessThan
: return nir_op_flt
;
291 case SpvOpFUnordLessThan
: return nir_op_flt
;
292 case SpvOpUGreaterThan
: *swap
= true; return nir_op_ult
;
293 case SpvOpSGreaterThan
: *swap
= true; return nir_op_ilt
;
294 case SpvOpFOrdGreaterThan
: *swap
= true; return nir_op_flt
;
295 case SpvOpFUnordGreaterThan
: *swap
= true; return nir_op_flt
;
296 case SpvOpULessThanEqual
: *swap
= true; return nir_op_uge
;
297 case SpvOpSLessThanEqual
: *swap
= true; return nir_op_ige
;
298 case SpvOpFOrdLessThanEqual
: *swap
= true; return nir_op_fge
;
299 case SpvOpFUnordLessThanEqual
: *swap
= true; return nir_op_fge
;
300 case SpvOpUGreaterThanEqual
: return nir_op_uge
;
301 case SpvOpSGreaterThanEqual
: return nir_op_ige
;
302 case SpvOpFOrdGreaterThanEqual
: return nir_op_fge
;
303 case SpvOpFUnordGreaterThanEqual
: return nir_op_fge
;
306 case SpvOpQuantizeToF16
: return nir_op_fquantize2f16
;
308 case SpvOpConvertFToU
:
309 case SpvOpConvertFToS
:
310 case SpvOpConvertSToF
:
311 case SpvOpConvertUToF
:
313 case SpvOpFConvert
: {
314 nir_alu_type src_type
;
315 nir_alu_type dst_type
;
318 case SpvOpConvertFToS
:
319 src_type
= nir_type_float
;
320 dst_type
= nir_type_int
;
322 case SpvOpConvertFToU
:
323 src_type
= nir_type_float
;
324 dst_type
= nir_type_uint
;
327 src_type
= dst_type
= nir_type_float
;
329 case SpvOpConvertSToF
:
330 src_type
= nir_type_int
;
331 dst_type
= nir_type_float
;
334 src_type
= dst_type
= nir_type_int
;
336 case SpvOpConvertUToF
:
337 src_type
= nir_type_uint
;
338 dst_type
= nir_type_float
;
341 src_type
= dst_type
= nir_type_uint
;
344 unreachable("Invalid opcode");
346 src_type
|= src_bit_size
;
347 dst_type
|= dst_bit_size
;
348 return nir_type_conversion_op(src_type
, dst_type
, nir_rounding_mode_undef
);
351 case SpvOpDPdx
: return nir_op_fddx
;
352 case SpvOpDPdy
: return nir_op_fddy
;
353 case SpvOpDPdxFine
: return nir_op_fddx_fine
;
354 case SpvOpDPdyFine
: return nir_op_fddy_fine
;
355 case SpvOpDPdxCoarse
: return nir_op_fddx_coarse
;
356 case SpvOpDPdyCoarse
: return nir_op_fddy_coarse
;
359 vtn_fail("No NIR equivalent: %u", opcode
);
364 handle_no_contraction(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
365 const struct vtn_decoration
*dec
, void *_void
)
367 vtn_assert(dec
->scope
== VTN_DEC_DECORATION
);
368 if (dec
->decoration
!= SpvDecorationNoContraction
)
375 handle_rounding_mode(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
376 const struct vtn_decoration
*dec
, void *_out_rounding_mode
)
378 nir_rounding_mode
*out_rounding_mode
= _out_rounding_mode
;
379 assert(dec
->scope
== VTN_DEC_DECORATION
);
380 if (dec
->decoration
!= SpvDecorationFPRoundingMode
)
382 switch (dec
->operands
[0]) {
383 case SpvFPRoundingModeRTE
:
384 *out_rounding_mode
= nir_rounding_mode_rtne
;
386 case SpvFPRoundingModeRTZ
:
387 *out_rounding_mode
= nir_rounding_mode_rtz
;
390 unreachable("Not supported rounding mode");
396 handle_no_wrap(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
397 const struct vtn_decoration
*dec
, void *_alu
)
399 nir_alu_instr
*alu
= _alu
;
400 switch (dec
->decoration
) {
401 case SpvDecorationNoSignedWrap
:
402 alu
->no_signed_wrap
= true;
404 case SpvDecorationNoUnsignedWrap
:
405 alu
->no_unsigned_wrap
= true;
414 vtn_handle_alu(struct vtn_builder
*b
, SpvOp opcode
,
415 const uint32_t *w
, unsigned count
)
417 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
418 const struct glsl_type
*type
= vtn_get_type(b
, w
[1])->type
;
420 vtn_foreach_decoration(b
, val
, handle_no_contraction
, NULL
);
422 /* Collect the various SSA sources */
423 const unsigned num_inputs
= count
- 3;
424 struct vtn_ssa_value
*vtn_src
[4] = { NULL
, };
425 for (unsigned i
= 0; i
< num_inputs
; i
++)
426 vtn_src
[i
] = vtn_ssa_value(b
, w
[i
+ 3]);
428 if (glsl_type_is_matrix(vtn_src
[0]->type
) ||
429 (num_inputs
>= 2 && glsl_type_is_matrix(vtn_src
[1]->type
))) {
430 vtn_handle_matrix_alu(b
, opcode
, val
, vtn_src
[0], vtn_src
[1]);
431 b
->nb
.exact
= b
->exact
;
435 val
->ssa
= vtn_create_ssa_value(b
, type
);
436 nir_ssa_def
*src
[4] = { NULL
, };
437 for (unsigned i
= 0; i
< num_inputs
; i
++) {
438 vtn_assert(glsl_type_is_vector_or_scalar(vtn_src
[i
]->type
));
439 src
[i
] = vtn_src
[i
]->def
;
444 val
->ssa
->def
= nir_bany(&b
->nb
, src
[0]);
448 val
->ssa
->def
= nir_ball(&b
->nb
, src
[0]);
451 case SpvOpOuterProduct
: {
452 for (unsigned i
= 0; i
< src
[1]->num_components
; i
++) {
453 val
->ssa
->elems
[i
]->def
=
454 nir_fmul(&b
->nb
, src
[0], nir_channel(&b
->nb
, src
[1], i
));
460 val
->ssa
->def
= nir_fdot(&b
->nb
, src
[0], src
[1]);
464 vtn_assert(glsl_type_is_struct_or_ifc(val
->ssa
->type
));
465 val
->ssa
->elems
[0]->def
= nir_iadd(&b
->nb
, src
[0], src
[1]);
466 val
->ssa
->elems
[1]->def
= nir_uadd_carry(&b
->nb
, src
[0], src
[1]);
469 case SpvOpISubBorrow
:
470 vtn_assert(glsl_type_is_struct_or_ifc(val
->ssa
->type
));
471 val
->ssa
->elems
[0]->def
= nir_isub(&b
->nb
, src
[0], src
[1]);
472 val
->ssa
->elems
[1]->def
= nir_usub_borrow(&b
->nb
, src
[0], src
[1]);
475 case SpvOpUMulExtended
: {
476 vtn_assert(glsl_type_is_struct_or_ifc(val
->ssa
->type
));
477 nir_ssa_def
*umul
= nir_umul_2x32_64(&b
->nb
, src
[0], src
[1]);
478 val
->ssa
->elems
[0]->def
= nir_unpack_64_2x32_split_x(&b
->nb
, umul
);
479 val
->ssa
->elems
[1]->def
= nir_unpack_64_2x32_split_y(&b
->nb
, umul
);
483 case SpvOpSMulExtended
: {
484 vtn_assert(glsl_type_is_struct_or_ifc(val
->ssa
->type
));
485 nir_ssa_def
*smul
= nir_imul_2x32_64(&b
->nb
, src
[0], src
[1]);
486 val
->ssa
->elems
[0]->def
= nir_unpack_64_2x32_split_x(&b
->nb
, smul
);
487 val
->ssa
->elems
[1]->def
= nir_unpack_64_2x32_split_y(&b
->nb
, smul
);
492 val
->ssa
->def
= nir_fadd(&b
->nb
,
493 nir_fabs(&b
->nb
, nir_fddx(&b
->nb
, src
[0])),
494 nir_fabs(&b
->nb
, nir_fddy(&b
->nb
, src
[0])));
496 case SpvOpFwidthFine
:
497 val
->ssa
->def
= nir_fadd(&b
->nb
,
498 nir_fabs(&b
->nb
, nir_fddx_fine(&b
->nb
, src
[0])),
499 nir_fabs(&b
->nb
, nir_fddy_fine(&b
->nb
, src
[0])));
501 case SpvOpFwidthCoarse
:
502 val
->ssa
->def
= nir_fadd(&b
->nb
,
503 nir_fabs(&b
->nb
, nir_fddx_coarse(&b
->nb
, src
[0])),
504 nir_fabs(&b
->nb
, nir_fddy_coarse(&b
->nb
, src
[0])));
507 case SpvOpVectorTimesScalar
:
508 /* The builder will take care of splatting for us. */
509 val
->ssa
->def
= nir_fmul(&b
->nb
, src
[0], src
[1]);
513 val
->ssa
->def
= nir_fne(&b
->nb
, src
[0], src
[0]);
517 nir_ssa_def
*inf
= nir_imm_floatN_t(&b
->nb
, INFINITY
, src
[0]->bit_size
);
518 val
->ssa
->def
= nir_ieq(&b
->nb
, nir_fabs(&b
->nb
, src
[0]), inf
);
522 case SpvOpFUnordEqual
:
523 case SpvOpFUnordNotEqual
:
524 case SpvOpFUnordLessThan
:
525 case SpvOpFUnordGreaterThan
:
526 case SpvOpFUnordLessThanEqual
:
527 case SpvOpFUnordGreaterThanEqual
: {
529 unsigned src_bit_size
= glsl_get_bit_size(vtn_src
[0]->type
);
530 unsigned dst_bit_size
= glsl_get_bit_size(type
);
531 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(b
, opcode
, &swap
,
532 src_bit_size
, dst_bit_size
);
535 nir_ssa_def
*tmp
= src
[0];
542 nir_build_alu(&b
->nb
, op
, src
[0], src
[1], NULL
, NULL
),
544 nir_fne(&b
->nb
, src
[0], src
[0]),
545 nir_fne(&b
->nb
, src
[1], src
[1])));
549 case SpvOpFOrdNotEqual
: {
550 /* For all the SpvOpFOrd* comparisons apart from NotEqual, the value
551 * from the ALU will probably already be false if the operands are not
552 * ordered so we don’t need to handle it specially.
555 unsigned src_bit_size
= glsl_get_bit_size(vtn_src
[0]->type
);
556 unsigned dst_bit_size
= glsl_get_bit_size(type
);
557 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(b
, opcode
, &swap
,
558 src_bit_size
, dst_bit_size
);
564 nir_build_alu(&b
->nb
, op
, src
[0], src
[1], NULL
, NULL
),
566 nir_feq(&b
->nb
, src
[0], src
[0]),
567 nir_feq(&b
->nb
, src
[1], src
[1])));
571 case SpvOpFConvert
: {
572 nir_alu_type src_alu_type
= nir_get_nir_type_for_glsl_type(vtn_src
[0]->type
);
573 nir_alu_type dst_alu_type
= nir_get_nir_type_for_glsl_type(type
);
574 nir_rounding_mode rounding_mode
= nir_rounding_mode_undef
;
576 vtn_foreach_decoration(b
, val
, handle_rounding_mode
, &rounding_mode
);
577 nir_op op
= nir_type_conversion_op(src_alu_type
, dst_alu_type
, rounding_mode
);
579 val
->ssa
->def
= nir_build_alu(&b
->nb
, op
, src
[0], src
[1], NULL
, NULL
);
583 case SpvOpBitFieldInsert
:
584 case SpvOpBitFieldSExtract
:
585 case SpvOpBitFieldUExtract
:
586 case SpvOpShiftLeftLogical
:
587 case SpvOpShiftRightArithmetic
:
588 case SpvOpShiftRightLogical
: {
590 unsigned src0_bit_size
= glsl_get_bit_size(vtn_src
[0]->type
);
591 unsigned dst_bit_size
= glsl_get_bit_size(type
);
592 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(b
, opcode
, &swap
,
593 src0_bit_size
, dst_bit_size
);
595 assert (op
== nir_op_ushr
|| op
== nir_op_ishr
|| op
== nir_op_ishl
||
596 op
== nir_op_bitfield_insert
|| op
== nir_op_ubitfield_extract
||
597 op
== nir_op_ibitfield_extract
);
599 for (unsigned i
= 0; i
< nir_op_infos
[op
].num_inputs
; i
++) {
600 unsigned src_bit_size
=
601 nir_alu_type_get_type_size(nir_op_infos
[op
].input_types
[i
]);
602 if (src_bit_size
== 0)
604 if (src_bit_size
!= src
[i
]->bit_size
) {
605 assert(src_bit_size
== 32);
606 /* Convert the Shift, Offset and Count operands to 32 bits, which is the bitsize
607 * supported by the NIR instructions. See discussion here:
609 * https://lists.freedesktop.org/archives/mesa-dev/2018-April/193026.html
611 src
[i
] = nir_u2u32(&b
->nb
, src
[i
]);
614 val
->ssa
->def
= nir_build_alu(&b
->nb
, op
, src
[0], src
[1], src
[2], src
[3]);
618 case SpvOpSignBitSet
:
619 val
->ssa
->def
= nir_i2b(&b
->nb
,
620 nir_ushr(&b
->nb
, src
[0], nir_imm_int(&b
->nb
, src
[0]->bit_size
- 1)));
623 case SpvOpUCountTrailingZerosINTEL
:
624 val
->ssa
->def
= nir_umin(&b
->nb
,
625 nir_find_lsb(&b
->nb
, src
[0]),
626 nir_imm_int(&b
->nb
, 32u));
631 unsigned src_bit_size
= glsl_get_bit_size(vtn_src
[0]->type
);
632 unsigned dst_bit_size
= glsl_get_bit_size(type
);
633 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(b
, opcode
, &swap
,
634 src_bit_size
, dst_bit_size
);
637 nir_ssa_def
*tmp
= src
[0];
646 if (src
[1]->bit_size
!= 32)
647 src
[1] = nir_u2u32(&b
->nb
, src
[1]);
653 val
->ssa
->def
= nir_build_alu(&b
->nb
, op
, src
[0], src
[1], src
[2], src
[3]);
662 case SpvOpShiftLeftLogical
:
664 nir_alu_instr
*alu
= nir_instr_as_alu(val
->ssa
->def
->parent_instr
);
665 vtn_foreach_decoration(b
, val
, handle_no_wrap
, alu
);
673 b
->nb
.exact
= b
->exact
;
677 vtn_handle_bitcast(struct vtn_builder
*b
, const uint32_t *w
, unsigned count
)
679 vtn_assert(count
== 4);
680 /* From the definition of OpBitcast in the SPIR-V 1.2 spec:
682 * "If Result Type has the same number of components as Operand, they
683 * must also have the same component width, and results are computed per
686 * If Result Type has a different number of components than Operand, the
687 * total number of bits in Result Type must equal the total number of
688 * bits in Operand. Let L be the type, either Result Type or Operand’s
689 * type, that has the larger number of components. Let S be the other
690 * type, with the smaller number of components. The number of components
691 * in L must be an integer multiple of the number of components in S.
692 * The first component (that is, the only or lowest-numbered component)
693 * of S maps to the first components of L, and so on, up to the last
694 * component of S mapping to the last components of L. Within this
695 * mapping, any single component of S (mapping to multiple components of
696 * L) maps its lower-ordered bits to the lower-numbered components of L."
699 struct vtn_type
*type
= vtn_get_type(b
, w
[1]);
700 struct nir_ssa_def
*src
= vtn_get_nir_ssa(b
, w
[3]);
702 vtn_fail_if(src
->num_components
* src
->bit_size
!=
703 glsl_get_vector_elements(type
->type
) * glsl_get_bit_size(type
->type
),
704 "Source and destination of OpBitcast must have the same "
705 "total number of bits");
707 nir_bitcast_vector(&b
->nb
, src
, glsl_get_bit_size(type
->type
));
708 vtn_push_nir_ssa(b
, w
[2], val
);