2 * Copyright © 2015 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 * Jason Ekstrand (jason@jlekstrand.net)
28 #include "vtn_private.h"
29 #include "nir/nir_vla.h"
30 #include "nir/nir_control_flow.h"
31 #include "nir/nir_constant_expressions.h"
32 #include "spirv_info.h"
34 struct spec_constant_value
{
43 _vtn_warn(const char *file
, int line
, const char *msg
, ...)
49 formatted
= ralloc_vasprintf(NULL
, msg
, args
);
52 fprintf(stderr
, "%s:%d WARNING: %s\n", file
, line
, formatted
);
54 ralloc_free(formatted
);
57 static struct vtn_ssa_value
*
58 vtn_undef_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
60 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
63 if (glsl_type_is_vector_or_scalar(type
)) {
64 unsigned num_components
= glsl_get_vector_elements(val
->type
);
65 unsigned bit_size
= glsl_get_bit_size(val
->type
);
66 val
->def
= nir_ssa_undef(&b
->nb
, num_components
, bit_size
);
68 unsigned elems
= glsl_get_length(val
->type
);
69 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
70 if (glsl_type_is_matrix(type
)) {
71 const struct glsl_type
*elem_type
=
72 glsl_vector_type(glsl_get_base_type(type
),
73 glsl_get_vector_elements(type
));
75 for (unsigned i
= 0; i
< elems
; i
++)
76 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
77 } else if (glsl_type_is_array(type
)) {
78 const struct glsl_type
*elem_type
= glsl_get_array_element(type
);
79 for (unsigned i
= 0; i
< elems
; i
++)
80 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
82 for (unsigned i
= 0; i
< elems
; i
++) {
83 const struct glsl_type
*elem_type
= glsl_get_struct_field(type
, i
);
84 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
92 static struct vtn_ssa_value
*
93 vtn_const_ssa_value(struct vtn_builder
*b
, nir_constant
*constant
,
94 const struct glsl_type
*type
)
96 struct hash_entry
*entry
= _mesa_hash_table_search(b
->const_table
, constant
);
101 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
104 switch (glsl_get_base_type(type
)) {
108 case GLSL_TYPE_FLOAT
:
109 case GLSL_TYPE_DOUBLE
: {
110 int bit_size
= glsl_get_bit_size(type
);
111 if (glsl_type_is_vector_or_scalar(type
)) {
112 unsigned num_components
= glsl_get_vector_elements(val
->type
);
113 nir_load_const_instr
*load
=
114 nir_load_const_instr_create(b
->shader
, num_components
, bit_size
);
116 load
->value
= constant
->values
[0];
118 nir_instr_insert_before_cf_list(&b
->impl
->body
, &load
->instr
);
119 val
->def
= &load
->def
;
121 assert(glsl_type_is_matrix(type
));
122 unsigned rows
= glsl_get_vector_elements(val
->type
);
123 unsigned columns
= glsl_get_matrix_columns(val
->type
);
124 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, columns
);
126 for (unsigned i
= 0; i
< columns
; i
++) {
127 struct vtn_ssa_value
*col_val
= rzalloc(b
, struct vtn_ssa_value
);
128 col_val
->type
= glsl_get_column_type(val
->type
);
129 nir_load_const_instr
*load
=
130 nir_load_const_instr_create(b
->shader
, rows
, bit_size
);
132 load
->value
= constant
->values
[i
];
134 nir_instr_insert_before_cf_list(&b
->impl
->body
, &load
->instr
);
135 col_val
->def
= &load
->def
;
137 val
->elems
[i
] = col_val
;
143 case GLSL_TYPE_ARRAY
: {
144 unsigned elems
= glsl_get_length(val
->type
);
145 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
146 const struct glsl_type
*elem_type
= glsl_get_array_element(val
->type
);
147 for (unsigned i
= 0; i
< elems
; i
++)
148 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
153 case GLSL_TYPE_STRUCT
: {
154 unsigned elems
= glsl_get_length(val
->type
);
155 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
156 for (unsigned i
= 0; i
< elems
; i
++) {
157 const struct glsl_type
*elem_type
=
158 glsl_get_struct_field(val
->type
, i
);
159 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
166 unreachable("bad constant type");
172 struct vtn_ssa_value
*
173 vtn_ssa_value(struct vtn_builder
*b
, uint32_t value_id
)
175 struct vtn_value
*val
= vtn_untyped_value(b
, value_id
);
176 switch (val
->value_type
) {
177 case vtn_value_type_undef
:
178 return vtn_undef_ssa_value(b
, val
->type
->type
);
180 case vtn_value_type_constant
:
181 return vtn_const_ssa_value(b
, val
->constant
, val
->const_type
);
183 case vtn_value_type_ssa
:
186 case vtn_value_type_access_chain
:
187 /* This is needed for function parameters */
188 return vtn_variable_load(b
, val
->access_chain
);
191 unreachable("Invalid type for an SSA value");
196 vtn_string_literal(struct vtn_builder
*b
, const uint32_t *words
,
197 unsigned word_count
, unsigned *words_used
)
199 char *dup
= ralloc_strndup(b
, (char *)words
, word_count
* sizeof(*words
));
201 /* Ammount of space taken by the string (including the null) */
202 unsigned len
= strlen(dup
) + 1;
203 *words_used
= DIV_ROUND_UP(len
, sizeof(*words
));
209 vtn_foreach_instruction(struct vtn_builder
*b
, const uint32_t *start
,
210 const uint32_t *end
, vtn_instruction_handler handler
)
216 const uint32_t *w
= start
;
218 SpvOp opcode
= w
[0] & SpvOpCodeMask
;
219 unsigned count
= w
[0] >> SpvWordCountShift
;
220 assert(count
>= 1 && w
+ count
<= end
);
224 break; /* Do nothing */
227 b
->file
= vtn_value(b
, w
[1], vtn_value_type_string
)->str
;
239 if (!handler(b
, opcode
, w
, count
))
251 vtn_handle_extension(struct vtn_builder
*b
, SpvOp opcode
,
252 const uint32_t *w
, unsigned count
)
255 case SpvOpExtInstImport
: {
256 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_extension
);
257 if (strcmp((const char *)&w
[2], "GLSL.std.450") == 0) {
258 val
->ext_handler
= vtn_handle_glsl450_instruction
;
260 assert(!"Unsupported extension");
266 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
267 bool handled
= val
->ext_handler(b
, w
[4], w
, count
);
274 unreachable("Unhandled opcode");
279 _foreach_decoration_helper(struct vtn_builder
*b
,
280 struct vtn_value
*base_value
,
282 struct vtn_value
*value
,
283 vtn_decoration_foreach_cb cb
, void *data
)
285 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
287 if (dec
->scope
== VTN_DEC_DECORATION
) {
288 member
= parent_member
;
289 } else if (dec
->scope
>= VTN_DEC_STRUCT_MEMBER0
) {
290 assert(parent_member
== -1);
291 member
= dec
->scope
- VTN_DEC_STRUCT_MEMBER0
;
293 /* Not a decoration */
298 assert(dec
->group
->value_type
== vtn_value_type_decoration_group
);
299 _foreach_decoration_helper(b
, base_value
, member
, dec
->group
,
302 cb(b
, base_value
, member
, dec
, data
);
307 /** Iterates (recursively if needed) over all of the decorations on a value
309 * This function iterates over all of the decorations applied to a given
310 * value. If it encounters a decoration group, it recurses into the group
311 * and iterates over all of those decorations as well.
314 vtn_foreach_decoration(struct vtn_builder
*b
, struct vtn_value
*value
,
315 vtn_decoration_foreach_cb cb
, void *data
)
317 _foreach_decoration_helper(b
, value
, -1, value
, cb
, data
);
321 vtn_foreach_execution_mode(struct vtn_builder
*b
, struct vtn_value
*value
,
322 vtn_execution_mode_foreach_cb cb
, void *data
)
324 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
325 if (dec
->scope
!= VTN_DEC_EXECUTION_MODE
)
328 assert(dec
->group
== NULL
);
329 cb(b
, value
, dec
, data
);
334 vtn_handle_decoration(struct vtn_builder
*b
, SpvOp opcode
,
335 const uint32_t *w
, unsigned count
)
337 const uint32_t *w_end
= w
+ count
;
338 const uint32_t target
= w
[1];
342 case SpvOpDecorationGroup
:
343 vtn_push_value(b
, target
, vtn_value_type_decoration_group
);
347 case SpvOpMemberDecorate
:
348 case SpvOpExecutionMode
: {
349 struct vtn_value
*val
= &b
->values
[target
];
351 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
354 dec
->scope
= VTN_DEC_DECORATION
;
356 case SpvOpMemberDecorate
:
357 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(w
++);
359 case SpvOpExecutionMode
:
360 dec
->scope
= VTN_DEC_EXECUTION_MODE
;
363 unreachable("Invalid decoration opcode");
365 dec
->decoration
= *(w
++);
368 /* Link into the list */
369 dec
->next
= val
->decoration
;
370 val
->decoration
= dec
;
374 case SpvOpGroupMemberDecorate
:
375 case SpvOpGroupDecorate
: {
376 struct vtn_value
*group
=
377 vtn_value(b
, target
, vtn_value_type_decoration_group
);
379 for (; w
< w_end
; w
++) {
380 struct vtn_value
*val
= vtn_untyped_value(b
, *w
);
381 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
384 if (opcode
== SpvOpGroupDecorate
) {
385 dec
->scope
= VTN_DEC_DECORATION
;
387 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(++w
);
390 /* Link into the list */
391 dec
->next
= val
->decoration
;
392 val
->decoration
= dec
;
398 unreachable("Unhandled opcode");
402 struct member_decoration_ctx
{
404 struct glsl_struct_field
*fields
;
405 struct vtn_type
*type
;
408 /* does a shallow copy of a vtn_type */
410 static struct vtn_type
*
411 vtn_type_copy(struct vtn_builder
*b
, struct vtn_type
*src
)
413 struct vtn_type
*dest
= ralloc(b
, struct vtn_type
);
414 dest
->type
= src
->type
;
415 dest
->is_builtin
= src
->is_builtin
;
417 dest
->builtin
= src
->builtin
;
419 if (!glsl_type_is_scalar(src
->type
)) {
420 switch (glsl_get_base_type(src
->type
)) {
424 case GLSL_TYPE_FLOAT
:
425 case GLSL_TYPE_DOUBLE
:
426 case GLSL_TYPE_ARRAY
:
427 dest
->row_major
= src
->row_major
;
428 dest
->stride
= src
->stride
;
429 dest
->array_element
= src
->array_element
;
432 case GLSL_TYPE_STRUCT
: {
433 unsigned elems
= glsl_get_length(src
->type
);
435 dest
->members
= ralloc_array(b
, struct vtn_type
*, elems
);
436 memcpy(dest
->members
, src
->members
, elems
* sizeof(struct vtn_type
*));
438 dest
->offsets
= ralloc_array(b
, unsigned, elems
);
439 memcpy(dest
->offsets
, src
->offsets
, elems
* sizeof(unsigned));
444 unreachable("unhandled type");
451 static struct vtn_type
*
452 mutable_matrix_member(struct vtn_builder
*b
, struct vtn_type
*type
, int member
)
454 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
455 type
= type
->members
[member
];
457 /* We may have an array of matrices.... Oh, joy! */
458 while (glsl_type_is_array(type
->type
)) {
459 type
->array_element
= vtn_type_copy(b
, type
->array_element
);
460 type
= type
->array_element
;
463 assert(glsl_type_is_matrix(type
->type
));
469 struct_member_decoration_cb(struct vtn_builder
*b
,
470 struct vtn_value
*val
, int member
,
471 const struct vtn_decoration
*dec
, void *void_ctx
)
473 struct member_decoration_ctx
*ctx
= void_ctx
;
478 assert(member
< ctx
->num_fields
);
480 switch (dec
->decoration
) {
481 case SpvDecorationNonWritable
:
482 case SpvDecorationNonReadable
:
483 case SpvDecorationRelaxedPrecision
:
484 case SpvDecorationVolatile
:
485 case SpvDecorationCoherent
:
486 case SpvDecorationUniform
:
487 break; /* FIXME: Do nothing with this for now. */
488 case SpvDecorationNoPerspective
:
489 ctx
->fields
[member
].interpolation
= INTERP_MODE_NOPERSPECTIVE
;
491 case SpvDecorationFlat
:
492 ctx
->fields
[member
].interpolation
= INTERP_MODE_FLAT
;
494 case SpvDecorationCentroid
:
495 ctx
->fields
[member
].centroid
= true;
497 case SpvDecorationSample
:
498 ctx
->fields
[member
].sample
= true;
500 case SpvDecorationStream
:
501 /* Vulkan only allows one GS stream */
502 assert(dec
->literals
[0] == 0);
504 case SpvDecorationLocation
:
505 ctx
->fields
[member
].location
= dec
->literals
[0];
507 case SpvDecorationComponent
:
508 break; /* FIXME: What should we do with these? */
509 case SpvDecorationBuiltIn
:
510 ctx
->type
->members
[member
] = vtn_type_copy(b
, ctx
->type
->members
[member
]);
511 ctx
->type
->members
[member
]->is_builtin
= true;
512 ctx
->type
->members
[member
]->builtin
= dec
->literals
[0];
513 ctx
->type
->builtin_block
= true;
515 case SpvDecorationOffset
:
516 ctx
->type
->offsets
[member
] = dec
->literals
[0];
518 case SpvDecorationMatrixStride
:
519 mutable_matrix_member(b
, ctx
->type
, member
)->stride
= dec
->literals
[0];
521 case SpvDecorationColMajor
:
522 break; /* Nothing to do here. Column-major is the default. */
523 case SpvDecorationRowMajor
:
524 mutable_matrix_member(b
, ctx
->type
, member
)->row_major
= true;
527 case SpvDecorationPatch
:
530 case SpvDecorationSpecId
:
531 case SpvDecorationBlock
:
532 case SpvDecorationBufferBlock
:
533 case SpvDecorationArrayStride
:
534 case SpvDecorationGLSLShared
:
535 case SpvDecorationGLSLPacked
:
536 case SpvDecorationInvariant
:
537 case SpvDecorationRestrict
:
538 case SpvDecorationAliased
:
539 case SpvDecorationConstant
:
540 case SpvDecorationIndex
:
541 case SpvDecorationBinding
:
542 case SpvDecorationDescriptorSet
:
543 case SpvDecorationLinkageAttributes
:
544 case SpvDecorationNoContraction
:
545 case SpvDecorationInputAttachmentIndex
:
546 vtn_warn("Decoration not allowed on struct members: %s",
547 spirv_decoration_to_string(dec
->decoration
));
550 case SpvDecorationXfbBuffer
:
551 case SpvDecorationXfbStride
:
552 vtn_warn("Vulkan does not have transform feedback");
555 case SpvDecorationCPacked
:
556 case SpvDecorationSaturatedConversion
:
557 case SpvDecorationFuncParamAttr
:
558 case SpvDecorationFPRoundingMode
:
559 case SpvDecorationFPFastMathMode
:
560 case SpvDecorationAlignment
:
561 vtn_warn("Decoration only allowed for CL-style kernels: %s",
562 spirv_decoration_to_string(dec
->decoration
));
566 unreachable("Unhandled decoration");
571 type_decoration_cb(struct vtn_builder
*b
,
572 struct vtn_value
*val
, int member
,
573 const struct vtn_decoration
*dec
, void *ctx
)
575 struct vtn_type
*type
= val
->type
;
580 switch (dec
->decoration
) {
581 case SpvDecorationArrayStride
:
582 type
->stride
= dec
->literals
[0];
584 case SpvDecorationBlock
:
587 case SpvDecorationBufferBlock
:
588 type
->buffer_block
= true;
590 case SpvDecorationGLSLShared
:
591 case SpvDecorationGLSLPacked
:
592 /* Ignore these, since we get explicit offsets anyways */
595 case SpvDecorationRowMajor
:
596 case SpvDecorationColMajor
:
597 case SpvDecorationMatrixStride
:
598 case SpvDecorationBuiltIn
:
599 case SpvDecorationNoPerspective
:
600 case SpvDecorationFlat
:
601 case SpvDecorationPatch
:
602 case SpvDecorationCentroid
:
603 case SpvDecorationSample
:
604 case SpvDecorationVolatile
:
605 case SpvDecorationCoherent
:
606 case SpvDecorationNonWritable
:
607 case SpvDecorationNonReadable
:
608 case SpvDecorationUniform
:
609 case SpvDecorationStream
:
610 case SpvDecorationLocation
:
611 case SpvDecorationComponent
:
612 case SpvDecorationOffset
:
613 case SpvDecorationXfbBuffer
:
614 case SpvDecorationXfbStride
:
615 vtn_warn("Decoration only allowed for struct members: %s",
616 spirv_decoration_to_string(dec
->decoration
));
619 case SpvDecorationRelaxedPrecision
:
620 case SpvDecorationSpecId
:
621 case SpvDecorationInvariant
:
622 case SpvDecorationRestrict
:
623 case SpvDecorationAliased
:
624 case SpvDecorationConstant
:
625 case SpvDecorationIndex
:
626 case SpvDecorationBinding
:
627 case SpvDecorationDescriptorSet
:
628 case SpvDecorationLinkageAttributes
:
629 case SpvDecorationNoContraction
:
630 case SpvDecorationInputAttachmentIndex
:
631 vtn_warn("Decoration not allowed on types: %s",
632 spirv_decoration_to_string(dec
->decoration
));
635 case SpvDecorationCPacked
:
636 case SpvDecorationSaturatedConversion
:
637 case SpvDecorationFuncParamAttr
:
638 case SpvDecorationFPRoundingMode
:
639 case SpvDecorationFPFastMathMode
:
640 case SpvDecorationAlignment
:
641 vtn_warn("Decoration only allowed for CL-style kernels: %s",
642 spirv_decoration_to_string(dec
->decoration
));
646 unreachable("Unhandled decoration");
651 translate_image_format(SpvImageFormat format
)
654 case SpvImageFormatUnknown
: return 0; /* GL_NONE */
655 case SpvImageFormatRgba32f
: return 0x8814; /* GL_RGBA32F */
656 case SpvImageFormatRgba16f
: return 0x881A; /* GL_RGBA16F */
657 case SpvImageFormatR32f
: return 0x822E; /* GL_R32F */
658 case SpvImageFormatRgba8
: return 0x8058; /* GL_RGBA8 */
659 case SpvImageFormatRgba8Snorm
: return 0x8F97; /* GL_RGBA8_SNORM */
660 case SpvImageFormatRg32f
: return 0x8230; /* GL_RG32F */
661 case SpvImageFormatRg16f
: return 0x822F; /* GL_RG16F */
662 case SpvImageFormatR11fG11fB10f
: return 0x8C3A; /* GL_R11F_G11F_B10F */
663 case SpvImageFormatR16f
: return 0x822D; /* GL_R16F */
664 case SpvImageFormatRgba16
: return 0x805B; /* GL_RGBA16 */
665 case SpvImageFormatRgb10A2
: return 0x8059; /* GL_RGB10_A2 */
666 case SpvImageFormatRg16
: return 0x822C; /* GL_RG16 */
667 case SpvImageFormatRg8
: return 0x822B; /* GL_RG8 */
668 case SpvImageFormatR16
: return 0x822A; /* GL_R16 */
669 case SpvImageFormatR8
: return 0x8229; /* GL_R8 */
670 case SpvImageFormatRgba16Snorm
: return 0x8F9B; /* GL_RGBA16_SNORM */
671 case SpvImageFormatRg16Snorm
: return 0x8F99; /* GL_RG16_SNORM */
672 case SpvImageFormatRg8Snorm
: return 0x8F95; /* GL_RG8_SNORM */
673 case SpvImageFormatR16Snorm
: return 0x8F98; /* GL_R16_SNORM */
674 case SpvImageFormatR8Snorm
: return 0x8F94; /* GL_R8_SNORM */
675 case SpvImageFormatRgba32i
: return 0x8D82; /* GL_RGBA32I */
676 case SpvImageFormatRgba16i
: return 0x8D88; /* GL_RGBA16I */
677 case SpvImageFormatRgba8i
: return 0x8D8E; /* GL_RGBA8I */
678 case SpvImageFormatR32i
: return 0x8235; /* GL_R32I */
679 case SpvImageFormatRg32i
: return 0x823B; /* GL_RG32I */
680 case SpvImageFormatRg16i
: return 0x8239; /* GL_RG16I */
681 case SpvImageFormatRg8i
: return 0x8237; /* GL_RG8I */
682 case SpvImageFormatR16i
: return 0x8233; /* GL_R16I */
683 case SpvImageFormatR8i
: return 0x8231; /* GL_R8I */
684 case SpvImageFormatRgba32ui
: return 0x8D70; /* GL_RGBA32UI */
685 case SpvImageFormatRgba16ui
: return 0x8D76; /* GL_RGBA16UI */
686 case SpvImageFormatRgba8ui
: return 0x8D7C; /* GL_RGBA8UI */
687 case SpvImageFormatR32ui
: return 0x8236; /* GL_R32UI */
688 case SpvImageFormatRgb10a2ui
: return 0x906F; /* GL_RGB10_A2UI */
689 case SpvImageFormatRg32ui
: return 0x823C; /* GL_RG32UI */
690 case SpvImageFormatRg16ui
: return 0x823A; /* GL_RG16UI */
691 case SpvImageFormatRg8ui
: return 0x8238; /* GL_RG8UI */
692 case SpvImageFormatR16ui
: return 0x823A; /* GL_RG16UI */
693 case SpvImageFormatR8ui
: return 0x8232; /* GL_R8UI */
695 assert(!"Invalid image format");
701 vtn_handle_type(struct vtn_builder
*b
, SpvOp opcode
,
702 const uint32_t *w
, unsigned count
)
704 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_type
);
706 val
->type
= rzalloc(b
, struct vtn_type
);
707 val
->type
->is_builtin
= false;
708 val
->type
->val
= val
;
712 val
->type
->type
= glsl_void_type();
715 val
->type
->type
= glsl_bool_type();
718 const bool signedness
= w
[3];
719 val
->type
->type
= (signedness
? glsl_int_type() : glsl_uint_type());
722 case SpvOpTypeFloat
: {
724 val
->type
->type
= bit_size
== 64 ? glsl_double_type() : glsl_float_type();
728 case SpvOpTypeVector
: {
729 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
730 unsigned elems
= w
[3];
732 assert(glsl_type_is_scalar(base
->type
));
733 val
->type
->type
= glsl_vector_type(glsl_get_base_type(base
->type
), elems
);
735 /* Vectors implicitly have sizeof(base_type) stride. For now, this
736 * is always 4 bytes. This will have to change if we want to start
737 * supporting doubles or half-floats.
739 val
->type
->stride
= 4;
740 val
->type
->array_element
= base
;
744 case SpvOpTypeMatrix
: {
745 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
746 unsigned columns
= w
[3];
748 assert(glsl_type_is_vector(base
->type
));
749 val
->type
->type
= glsl_matrix_type(glsl_get_base_type(base
->type
),
750 glsl_get_vector_elements(base
->type
),
752 assert(!glsl_type_is_error(val
->type
->type
));
753 val
->type
->array_element
= base
;
754 val
->type
->row_major
= false;
755 val
->type
->stride
= 0;
759 case SpvOpTypeRuntimeArray
:
760 case SpvOpTypeArray
: {
761 struct vtn_type
*array_element
=
762 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
765 if (opcode
== SpvOpTypeRuntimeArray
) {
766 /* A length of 0 is used to denote unsized arrays */
770 vtn_value(b
, w
[3], vtn_value_type_constant
)->constant
->values
[0].u32
[0];
773 val
->type
->type
= glsl_array_type(array_element
->type
, length
);
774 val
->type
->array_element
= array_element
;
775 val
->type
->stride
= 0;
779 case SpvOpTypeStruct
: {
780 unsigned num_fields
= count
- 2;
781 val
->type
->members
= ralloc_array(b
, struct vtn_type
*, num_fields
);
782 val
->type
->offsets
= ralloc_array(b
, unsigned, num_fields
);
784 NIR_VLA(struct glsl_struct_field
, fields
, count
);
785 for (unsigned i
= 0; i
< num_fields
; i
++) {
786 val
->type
->members
[i
] =
787 vtn_value(b
, w
[i
+ 2], vtn_value_type_type
)->type
;
788 fields
[i
] = (struct glsl_struct_field
) {
789 .type
= val
->type
->members
[i
]->type
,
790 .name
= ralloc_asprintf(b
, "field%d", i
),
795 struct member_decoration_ctx ctx
= {
796 .num_fields
= num_fields
,
801 vtn_foreach_decoration(b
, val
, struct_member_decoration_cb
, &ctx
);
803 const char *name
= val
->name
? val
->name
: "struct";
805 val
->type
->type
= glsl_struct_type(fields
, num_fields
, name
);
809 case SpvOpTypeFunction
: {
810 const struct glsl_type
*return_type
=
811 vtn_value(b
, w
[2], vtn_value_type_type
)->type
->type
;
812 NIR_VLA(struct glsl_function_param
, params
, count
- 3);
813 for (unsigned i
= 0; i
< count
- 3; i
++) {
814 params
[i
].type
= vtn_value(b
, w
[i
+ 3], vtn_value_type_type
)->type
->type
;
818 params
[i
].out
= true;
820 val
->type
->type
= glsl_function_type(return_type
, params
, count
- 3);
824 case SpvOpTypePointer
:
825 /* FIXME: For now, we'll just do the really lame thing and return
826 * the same type. The validator should ensure that the proper number
827 * of dereferences happen
829 val
->type
= vtn_value(b
, w
[3], vtn_value_type_type
)->type
;
832 case SpvOpTypeImage
: {
833 const struct glsl_type
*sampled_type
=
834 vtn_value(b
, w
[2], vtn_value_type_type
)->type
->type
;
836 assert(glsl_type_is_vector_or_scalar(sampled_type
));
838 enum glsl_sampler_dim dim
;
839 switch ((SpvDim
)w
[3]) {
840 case SpvDim1D
: dim
= GLSL_SAMPLER_DIM_1D
; break;
841 case SpvDim2D
: dim
= GLSL_SAMPLER_DIM_2D
; break;
842 case SpvDim3D
: dim
= GLSL_SAMPLER_DIM_3D
; break;
843 case SpvDimCube
: dim
= GLSL_SAMPLER_DIM_CUBE
; break;
844 case SpvDimRect
: dim
= GLSL_SAMPLER_DIM_RECT
; break;
845 case SpvDimBuffer
: dim
= GLSL_SAMPLER_DIM_BUF
; break;
846 case SpvDimSubpassData
: dim
= GLSL_SAMPLER_DIM_SUBPASS
; break;
848 unreachable("Invalid SPIR-V Sampler dimension");
851 bool is_shadow
= w
[4];
852 bool is_array
= w
[5];
853 bool multisampled
= w
[6];
854 unsigned sampled
= w
[7];
855 SpvImageFormat format
= w
[8];
858 val
->type
->access_qualifier
= w
[9];
860 val
->type
->access_qualifier
= SpvAccessQualifierReadWrite
;
863 if (dim
== GLSL_SAMPLER_DIM_2D
)
864 dim
= GLSL_SAMPLER_DIM_MS
;
865 else if (dim
== GLSL_SAMPLER_DIM_SUBPASS
)
866 dim
= GLSL_SAMPLER_DIM_SUBPASS_MS
;
868 assert(!"Unsupported multisampled image type");
871 val
->type
->image_format
= translate_image_format(format
);
874 val
->type
->type
= glsl_sampler_type(dim
, is_shadow
, is_array
,
875 glsl_get_base_type(sampled_type
));
876 } else if (sampled
== 2) {
878 val
->type
->type
= glsl_image_type(dim
, is_array
,
879 glsl_get_base_type(sampled_type
));
881 assert(!"We need to know if the image will be sampled");
886 case SpvOpTypeSampledImage
:
887 val
->type
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
890 case SpvOpTypeSampler
:
891 /* The actual sampler type here doesn't really matter. It gets
892 * thrown away the moment you combine it with an image. What really
893 * matters is that it's a sampler type as opposed to an integer type
894 * so the backend knows what to do.
896 val
->type
->type
= glsl_bare_sampler_type();
899 case SpvOpTypeOpaque
:
901 case SpvOpTypeDeviceEvent
:
902 case SpvOpTypeReserveId
:
906 unreachable("Unhandled opcode");
909 vtn_foreach_decoration(b
, val
, type_decoration_cb
, NULL
);
912 static nir_constant
*
913 vtn_null_constant(struct vtn_builder
*b
, const struct glsl_type
*type
)
915 nir_constant
*c
= rzalloc(b
, nir_constant
);
917 switch (glsl_get_base_type(type
)) {
921 case GLSL_TYPE_FLOAT
:
922 case GLSL_TYPE_DOUBLE
:
923 /* Nothing to do here. It's already initialized to zero */
926 case GLSL_TYPE_ARRAY
:
927 assert(glsl_get_length(type
) > 0);
928 c
->num_elements
= glsl_get_length(type
);
929 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
931 c
->elements
[0] = vtn_null_constant(b
, glsl_get_array_element(type
));
932 for (unsigned i
= 1; i
< c
->num_elements
; i
++)
933 c
->elements
[i
] = c
->elements
[0];
936 case GLSL_TYPE_STRUCT
:
937 c
->num_elements
= glsl_get_length(type
);
938 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
940 for (unsigned i
= 0; i
< c
->num_elements
; i
++) {
941 c
->elements
[i
] = vtn_null_constant(b
, glsl_get_struct_field(type
, i
));
946 unreachable("Invalid type for null constant");
953 spec_constant_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*v
,
954 int member
, const struct vtn_decoration
*dec
,
957 assert(member
== -1);
958 if (dec
->decoration
!= SpvDecorationSpecId
)
961 struct spec_constant_value
*const_value
= data
;
963 for (unsigned i
= 0; i
< b
->num_specializations
; i
++) {
964 if (b
->specializations
[i
].id
== dec
->literals
[0]) {
965 if (const_value
->is_double
)
966 const_value
->data64
= b
->specializations
[i
].data64
;
968 const_value
->data32
= b
->specializations
[i
].data32
;
975 get_specialization(struct vtn_builder
*b
, struct vtn_value
*val
,
976 uint32_t const_value
)
978 struct spec_constant_value data
;
979 data
.is_double
= false;
980 data
.data32
= const_value
;
981 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
986 get_specialization64(struct vtn_builder
*b
, struct vtn_value
*val
,
987 uint64_t const_value
)
989 struct spec_constant_value data
;
990 data
.is_double
= true;
991 data
.data64
= const_value
;
992 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
997 handle_workgroup_size_decoration_cb(struct vtn_builder
*b
,
998 struct vtn_value
*val
,
1000 const struct vtn_decoration
*dec
,
1003 assert(member
== -1);
1004 if (dec
->decoration
!= SpvDecorationBuiltIn
||
1005 dec
->literals
[0] != SpvBuiltInWorkgroupSize
)
1008 assert(val
->const_type
== glsl_vector_type(GLSL_TYPE_UINT
, 3));
1010 b
->shader
->info
->cs
.local_size
[0] = val
->constant
->values
[0].u32
[0];
1011 b
->shader
->info
->cs
.local_size
[1] = val
->constant
->values
[0].u32
[1];
1012 b
->shader
->info
->cs
.local_size
[2] = val
->constant
->values
[0].u32
[2];
1016 vtn_handle_constant(struct vtn_builder
*b
, SpvOp opcode
,
1017 const uint32_t *w
, unsigned count
)
1019 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_constant
);
1020 val
->const_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
1021 val
->constant
= rzalloc(b
, nir_constant
);
1023 case SpvOpConstantTrue
:
1024 assert(val
->const_type
== glsl_bool_type());
1025 val
->constant
->values
[0].u32
[0] = NIR_TRUE
;
1027 case SpvOpConstantFalse
:
1028 assert(val
->const_type
== glsl_bool_type());
1029 val
->constant
->values
[0].u32
[0] = NIR_FALSE
;
1032 case SpvOpSpecConstantTrue
:
1033 case SpvOpSpecConstantFalse
: {
1034 assert(val
->const_type
== glsl_bool_type());
1036 get_specialization(b
, val
, (opcode
== SpvOpSpecConstantTrue
));
1037 val
->constant
->values
[0].u32
[0] = int_val
? NIR_TRUE
: NIR_FALSE
;
1041 case SpvOpConstant
: {
1042 assert(glsl_type_is_scalar(val
->const_type
));
1043 int bit_size
= glsl_get_bit_size(val
->const_type
);
1044 if (bit_size
== 64) {
1045 val
->constant
->values
->u32
[0] = w
[3];
1046 val
->constant
->values
->u32
[1] = w
[4];
1048 assert(bit_size
== 32);
1049 val
->constant
->values
->u32
[0] = w
[3];
1053 case SpvOpSpecConstant
: {
1054 assert(glsl_type_is_scalar(val
->const_type
));
1055 val
->constant
->values
[0].u32
[0] = get_specialization(b
, val
, w
[3]);
1056 int bit_size
= glsl_get_bit_size(val
->const_type
);
1058 val
->constant
->values
[0].u64
[0] =
1059 get_specialization64(b
, val
, vtn_u64_literal(&w
[3]));
1061 val
->constant
->values
[0].u32
[0] = get_specialization(b
, val
, w
[3]);
1064 case SpvOpSpecConstantComposite
:
1065 case SpvOpConstantComposite
: {
1066 unsigned elem_count
= count
- 3;
1067 nir_constant
**elems
= ralloc_array(b
, nir_constant
*, elem_count
);
1068 for (unsigned i
= 0; i
< elem_count
; i
++)
1069 elems
[i
] = vtn_value(b
, w
[i
+ 3], vtn_value_type_constant
)->constant
;
1071 switch (glsl_get_base_type(val
->const_type
)) {
1072 case GLSL_TYPE_UINT
:
1074 case GLSL_TYPE_FLOAT
:
1075 case GLSL_TYPE_BOOL
:
1076 case GLSL_TYPE_DOUBLE
: {
1077 int bit_size
= glsl_get_bit_size(val
->const_type
);
1078 if (glsl_type_is_matrix(val
->const_type
)) {
1079 assert(glsl_get_matrix_columns(val
->const_type
) == elem_count
);
1080 for (unsigned i
= 0; i
< elem_count
; i
++)
1081 val
->constant
->values
[i
] = elems
[i
]->values
[0];
1083 assert(glsl_type_is_vector(val
->const_type
));
1084 assert(glsl_get_vector_elements(val
->const_type
) == elem_count
);
1085 for (unsigned i
= 0; i
< elem_count
; i
++) {
1086 if (bit_size
== 64) {
1087 val
->constant
->values
[0].u64
[i
] = elems
[i
]->values
[0].u64
[0];
1089 assert(bit_size
== 32);
1090 val
->constant
->values
[0].u32
[i
] = elems
[i
]->values
[0].u32
[0];
1097 case GLSL_TYPE_STRUCT
:
1098 case GLSL_TYPE_ARRAY
:
1099 ralloc_steal(val
->constant
, elems
);
1100 val
->constant
->num_elements
= elem_count
;
1101 val
->constant
->elements
= elems
;
1105 unreachable("Unsupported type for constants");
1110 case SpvOpSpecConstantOp
: {
1111 SpvOp opcode
= get_specialization(b
, val
, w
[3]);
1113 case SpvOpVectorShuffle
: {
1114 struct vtn_value
*v0
= &b
->values
[w
[4]];
1115 struct vtn_value
*v1
= &b
->values
[w
[5]];
1117 assert(v0
->value_type
== vtn_value_type_constant
||
1118 v0
->value_type
== vtn_value_type_undef
);
1119 assert(v1
->value_type
== vtn_value_type_constant
||
1120 v1
->value_type
== vtn_value_type_undef
);
1122 unsigned len0
= v0
->value_type
== vtn_value_type_constant
?
1123 glsl_get_vector_elements(v0
->const_type
) :
1124 glsl_get_vector_elements(v0
->type
->type
);
1125 unsigned len1
= v1
->value_type
== vtn_value_type_constant
?
1126 glsl_get_vector_elements(v1
->const_type
) :
1127 glsl_get_vector_elements(v1
->type
->type
);
1129 assert(len0
+ len1
< 16);
1131 unsigned bit_size
= glsl_get_bit_size(val
->const_type
);
1132 unsigned bit_size0
= v0
->value_type
== vtn_value_type_constant
?
1133 glsl_get_bit_size(v0
->const_type
) :
1134 glsl_get_bit_size(v0
->type
->type
);
1135 unsigned bit_size1
= v1
->value_type
== vtn_value_type_constant
?
1136 glsl_get_bit_size(v1
->const_type
) :
1137 glsl_get_bit_size(v1
->type
->type
);
1139 assert(bit_size
== bit_size0
&& bit_size
== bit_size1
);
1141 if (bit_size
== 64) {
1143 if (v0
->value_type
== vtn_value_type_constant
) {
1144 for (unsigned i
= 0; i
< len0
; i
++)
1145 u64
[i
] = v0
->constant
->values
[0].u64
[i
];
1147 if (v1
->value_type
== vtn_value_type_constant
) {
1148 for (unsigned i
= 0; i
< len1
; i
++)
1149 u64
[len0
+ i
] = v1
->constant
->values
[0].u64
[i
];
1152 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1153 uint32_t comp
= w
[i
+ 6];
1154 /* If component is not used, set the value to a known constant
1155 * to detect if it is wrongly used.
1157 if (comp
== (uint32_t)-1)
1158 val
->constant
->values
[0].u64
[j
] = 0xdeadbeefdeadbeef;
1160 val
->constant
->values
[0].u64
[j
] = u64
[comp
];
1164 if (v0
->value_type
== vtn_value_type_constant
) {
1165 for (unsigned i
= 0; i
< len0
; i
++)
1166 u32
[i
] = v0
->constant
->values
[0].u32
[i
];
1168 if (v1
->value_type
== vtn_value_type_constant
) {
1169 for (unsigned i
= 0; i
< len1
; i
++)
1170 u32
[len0
+ i
] = v1
->constant
->values
[0].u32
[i
];
1173 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1174 uint32_t comp
= w
[i
+ 6];
1175 /* If component is not used, set the value to a known constant
1176 * to detect if it is wrongly used.
1178 if (comp
== (uint32_t)-1)
1179 val
->constant
->values
[0].u32
[j
] = 0xdeadbeef;
1181 val
->constant
->values
[0].u32
[j
] = u32
[comp
];
1187 case SpvOpCompositeExtract
:
1188 case SpvOpCompositeInsert
: {
1189 struct vtn_value
*comp
;
1190 unsigned deref_start
;
1191 struct nir_constant
**c
;
1192 if (opcode
== SpvOpCompositeExtract
) {
1193 comp
= vtn_value(b
, w
[4], vtn_value_type_constant
);
1195 c
= &comp
->constant
;
1197 comp
= vtn_value(b
, w
[5], vtn_value_type_constant
);
1199 val
->constant
= nir_constant_clone(comp
->constant
,
1206 const struct glsl_type
*type
= comp
->const_type
;
1207 for (unsigned i
= deref_start
; i
< count
; i
++) {
1208 switch (glsl_get_base_type(type
)) {
1209 case GLSL_TYPE_UINT
:
1211 case GLSL_TYPE_FLOAT
:
1212 case GLSL_TYPE_DOUBLE
:
1213 case GLSL_TYPE_BOOL
:
1214 /* If we hit this granularity, we're picking off an element */
1215 if (glsl_type_is_matrix(type
)) {
1216 assert(col
== 0 && elem
== -1);
1219 type
= glsl_get_column_type(type
);
1221 assert(elem
<= 0 && glsl_type_is_vector(type
));
1223 type
= glsl_scalar_type(glsl_get_base_type(type
));
1227 case GLSL_TYPE_ARRAY
:
1228 c
= &(*c
)->elements
[w
[i
]];
1229 type
= glsl_get_array_element(type
);
1232 case GLSL_TYPE_STRUCT
:
1233 c
= &(*c
)->elements
[w
[i
]];
1234 type
= glsl_get_struct_field(type
, w
[i
]);
1238 unreachable("Invalid constant type");
1242 if (opcode
== SpvOpCompositeExtract
) {
1246 unsigned num_components
= glsl_get_vector_elements(type
);
1247 unsigned bit_size
= glsl_get_bit_size(type
);
1248 for (unsigned i
= 0; i
< num_components
; i
++)
1249 if (bit_size
== 64) {
1250 val
->constant
->values
[0].u64
[i
] = (*c
)->values
[col
].u64
[elem
+ i
];
1252 assert(bit_size
== 32);
1253 val
->constant
->values
[0].u32
[i
] = (*c
)->values
[col
].u32
[elem
+ i
];
1257 struct vtn_value
*insert
=
1258 vtn_value(b
, w
[4], vtn_value_type_constant
);
1259 assert(insert
->const_type
== type
);
1261 *c
= insert
->constant
;
1263 unsigned num_components
= glsl_get_vector_elements(type
);
1264 unsigned bit_size
= glsl_get_bit_size(type
);
1265 for (unsigned i
= 0; i
< num_components
; i
++)
1266 if (bit_size
== 64) {
1267 (*c
)->values
[col
].u64
[elem
+ i
] = insert
->constant
->values
[0].u64
[i
];
1269 assert(bit_size
== 32);
1270 (*c
)->values
[col
].u32
[elem
+ i
] = insert
->constant
->values
[0].u32
[i
];
1279 nir_alu_type dst_alu_type
= nir_get_nir_type_for_glsl_type(val
->const_type
);
1280 nir_alu_type src_alu_type
= dst_alu_type
;
1281 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(opcode
, &swap
, src_alu_type
, dst_alu_type
);
1283 unsigned num_components
= glsl_get_vector_elements(val
->const_type
);
1285 glsl_get_bit_size(val
->const_type
);
1287 nir_const_value src
[4];
1289 for (unsigned i
= 0; i
< count
- 4; i
++) {
1291 vtn_value(b
, w
[4 + i
], vtn_value_type_constant
)->constant
;
1293 unsigned j
= swap
? 1 - i
: i
;
1294 assert(bit_size
== 32);
1295 src
[j
] = c
->values
[0];
1298 val
->constant
->values
[0] =
1299 nir_eval_const_opcode(op
, num_components
, bit_size
, src
);
1306 case SpvOpConstantNull
:
1307 val
->constant
= vtn_null_constant(b
, val
->const_type
);
1310 case SpvOpConstantSampler
:
1311 assert(!"OpConstantSampler requires Kernel Capability");
1315 unreachable("Unhandled opcode");
1318 /* Now that we have the value, update the workgroup size if needed */
1319 vtn_foreach_decoration(b
, val
, handle_workgroup_size_decoration_cb
, NULL
);
1323 vtn_handle_function_call(struct vtn_builder
*b
, SpvOp opcode
,
1324 const uint32_t *w
, unsigned count
)
1326 struct nir_function
*callee
=
1327 vtn_value(b
, w
[3], vtn_value_type_function
)->func
->impl
->function
;
1329 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, callee
);
1330 for (unsigned i
= 0; i
< call
->num_params
; i
++) {
1331 unsigned arg_id
= w
[4 + i
];
1332 struct vtn_value
*arg
= vtn_untyped_value(b
, arg_id
);
1333 if (arg
->value_type
== vtn_value_type_access_chain
) {
1334 nir_deref_var
*d
= vtn_access_chain_to_deref(b
, arg
->access_chain
);
1335 call
->params
[i
] = nir_deref_var_clone(d
, call
);
1337 struct vtn_ssa_value
*arg_ssa
= vtn_ssa_value(b
, arg_id
);
1339 /* Make a temporary to store the argument in */
1341 nir_local_variable_create(b
->impl
, arg_ssa
->type
, "arg_tmp");
1342 call
->params
[i
] = nir_deref_var_create(call
, tmp
);
1344 vtn_local_store(b
, arg_ssa
, call
->params
[i
]);
1348 nir_variable
*out_tmp
= NULL
;
1349 if (!glsl_type_is_void(callee
->return_type
)) {
1350 out_tmp
= nir_local_variable_create(b
->impl
, callee
->return_type
,
1352 call
->return_deref
= nir_deref_var_create(call
, out_tmp
);
1355 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
1357 if (glsl_type_is_void(callee
->return_type
)) {
1358 vtn_push_value(b
, w
[2], vtn_value_type_undef
);
1360 struct vtn_value
*retval
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1361 retval
->ssa
= vtn_local_load(b
, call
->return_deref
);
1365 struct vtn_ssa_value
*
1366 vtn_create_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
1368 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
1371 if (!glsl_type_is_vector_or_scalar(type
)) {
1372 unsigned elems
= glsl_get_length(type
);
1373 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
1374 for (unsigned i
= 0; i
< elems
; i
++) {
1375 const struct glsl_type
*child_type
;
1377 switch (glsl_get_base_type(type
)) {
1379 case GLSL_TYPE_UINT
:
1380 case GLSL_TYPE_BOOL
:
1381 case GLSL_TYPE_FLOAT
:
1382 case GLSL_TYPE_DOUBLE
:
1383 child_type
= glsl_get_column_type(type
);
1385 case GLSL_TYPE_ARRAY
:
1386 child_type
= glsl_get_array_element(type
);
1388 case GLSL_TYPE_STRUCT
:
1389 child_type
= glsl_get_struct_field(type
, i
);
1392 unreachable("unkown base type");
1395 val
->elems
[i
] = vtn_create_ssa_value(b
, child_type
);
1403 vtn_tex_src(struct vtn_builder
*b
, unsigned index
, nir_tex_src_type type
)
1406 src
.src
= nir_src_for_ssa(vtn_ssa_value(b
, index
)->def
);
1407 src
.src_type
= type
;
1412 vtn_handle_texture(struct vtn_builder
*b
, SpvOp opcode
,
1413 const uint32_t *w
, unsigned count
)
1415 if (opcode
== SpvOpSampledImage
) {
1416 struct vtn_value
*val
=
1417 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
1418 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
1419 val
->sampled_image
->image
=
1420 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1421 val
->sampled_image
->sampler
=
1422 vtn_value(b
, w
[4], vtn_value_type_access_chain
)->access_chain
;
1424 } else if (opcode
== SpvOpImage
) {
1425 struct vtn_value
*val
=
1426 vtn_push_value(b
, w
[2], vtn_value_type_access_chain
);
1427 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[3]);
1428 if (src_val
->value_type
== vtn_value_type_sampled_image
) {
1429 val
->access_chain
= src_val
->sampled_image
->image
;
1431 assert(src_val
->value_type
== vtn_value_type_access_chain
);
1432 val
->access_chain
= src_val
->access_chain
;
1437 struct vtn_type
*ret_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1438 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1440 struct vtn_sampled_image sampled
;
1441 struct vtn_value
*sampled_val
= vtn_untyped_value(b
, w
[3]);
1442 if (sampled_val
->value_type
== vtn_value_type_sampled_image
) {
1443 sampled
= *sampled_val
->sampled_image
;
1445 assert(sampled_val
->value_type
== vtn_value_type_access_chain
);
1446 sampled
.image
= NULL
;
1447 sampled
.sampler
= sampled_val
->access_chain
;
1450 const struct glsl_type
*image_type
;
1451 if (sampled
.image
) {
1452 image_type
= sampled
.image
->var
->var
->interface_type
;
1454 image_type
= sampled
.sampler
->var
->var
->interface_type
;
1456 const enum glsl_sampler_dim sampler_dim
= glsl_get_sampler_dim(image_type
);
1457 const bool is_array
= glsl_sampler_type_is_array(image_type
);
1458 const bool is_shadow
= glsl_sampler_type_is_shadow(image_type
);
1460 /* Figure out the base texture operation */
1463 case SpvOpImageSampleImplicitLod
:
1464 case SpvOpImageSampleDrefImplicitLod
:
1465 case SpvOpImageSampleProjImplicitLod
:
1466 case SpvOpImageSampleProjDrefImplicitLod
:
1467 texop
= nir_texop_tex
;
1470 case SpvOpImageSampleExplicitLod
:
1471 case SpvOpImageSampleDrefExplicitLod
:
1472 case SpvOpImageSampleProjExplicitLod
:
1473 case SpvOpImageSampleProjDrefExplicitLod
:
1474 texop
= nir_texop_txl
;
1477 case SpvOpImageFetch
:
1478 if (glsl_get_sampler_dim(image_type
) == GLSL_SAMPLER_DIM_MS
) {
1479 texop
= nir_texop_txf_ms
;
1481 texop
= nir_texop_txf
;
1485 case SpvOpImageGather
:
1486 case SpvOpImageDrefGather
:
1487 texop
= nir_texop_tg4
;
1490 case SpvOpImageQuerySizeLod
:
1491 case SpvOpImageQuerySize
:
1492 texop
= nir_texop_txs
;
1495 case SpvOpImageQueryLod
:
1496 texop
= nir_texop_lod
;
1499 case SpvOpImageQueryLevels
:
1500 texop
= nir_texop_query_levels
;
1503 case SpvOpImageQuerySamples
:
1504 texop
= nir_texop_texture_samples
;
1508 unreachable("Unhandled opcode");
1511 nir_tex_src srcs
[8]; /* 8 should be enough */
1512 nir_tex_src
*p
= srcs
;
1516 struct nir_ssa_def
*coord
;
1517 unsigned coord_components
;
1519 case SpvOpImageSampleImplicitLod
:
1520 case SpvOpImageSampleExplicitLod
:
1521 case SpvOpImageSampleDrefImplicitLod
:
1522 case SpvOpImageSampleDrefExplicitLod
:
1523 case SpvOpImageSampleProjImplicitLod
:
1524 case SpvOpImageSampleProjExplicitLod
:
1525 case SpvOpImageSampleProjDrefImplicitLod
:
1526 case SpvOpImageSampleProjDrefExplicitLod
:
1527 case SpvOpImageFetch
:
1528 case SpvOpImageGather
:
1529 case SpvOpImageDrefGather
:
1530 case SpvOpImageQueryLod
: {
1531 /* All these types have the coordinate as their first real argument */
1532 switch (sampler_dim
) {
1533 case GLSL_SAMPLER_DIM_1D
:
1534 case GLSL_SAMPLER_DIM_BUF
:
1535 coord_components
= 1;
1537 case GLSL_SAMPLER_DIM_2D
:
1538 case GLSL_SAMPLER_DIM_RECT
:
1539 case GLSL_SAMPLER_DIM_MS
:
1540 coord_components
= 2;
1542 case GLSL_SAMPLER_DIM_3D
:
1543 case GLSL_SAMPLER_DIM_CUBE
:
1544 coord_components
= 3;
1547 unreachable("Invalid sampler type");
1550 if (is_array
&& texop
!= nir_texop_lod
)
1553 coord
= vtn_ssa_value(b
, w
[idx
++])->def
;
1554 p
->src
= nir_src_for_ssa(coord
);
1555 p
->src_type
= nir_tex_src_coord
;
1562 coord_components
= 0;
1567 case SpvOpImageSampleProjImplicitLod
:
1568 case SpvOpImageSampleProjExplicitLod
:
1569 case SpvOpImageSampleProjDrefImplicitLod
:
1570 case SpvOpImageSampleProjDrefExplicitLod
:
1571 /* These have the projector as the last coordinate component */
1572 p
->src
= nir_src_for_ssa(nir_channel(&b
->nb
, coord
, coord_components
));
1573 p
->src_type
= nir_tex_src_projector
;
1581 unsigned gather_component
= 0;
1583 case SpvOpImageSampleDrefImplicitLod
:
1584 case SpvOpImageSampleDrefExplicitLod
:
1585 case SpvOpImageSampleProjDrefImplicitLod
:
1586 case SpvOpImageSampleProjDrefExplicitLod
:
1587 case SpvOpImageDrefGather
:
1588 /* These all have an explicit depth value as their next source */
1589 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_comparator
);
1592 case SpvOpImageGather
:
1593 /* This has a component as its next source */
1595 vtn_value(b
, w
[idx
++], vtn_value_type_constant
)->constant
->values
[0].u32
[0];
1602 /* For OpImageQuerySizeLod, we always have an LOD */
1603 if (opcode
== SpvOpImageQuerySizeLod
)
1604 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
1606 /* Now we need to handle some number of optional arguments */
1607 const struct vtn_ssa_value
*gather_offsets
= NULL
;
1609 uint32_t operands
= w
[idx
++];
1611 if (operands
& SpvImageOperandsBiasMask
) {
1612 assert(texop
== nir_texop_tex
);
1613 texop
= nir_texop_txb
;
1614 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_bias
);
1617 if (operands
& SpvImageOperandsLodMask
) {
1618 assert(texop
== nir_texop_txl
|| texop
== nir_texop_txf
||
1619 texop
== nir_texop_txs
);
1620 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
1623 if (operands
& SpvImageOperandsGradMask
) {
1624 assert(texop
== nir_texop_txl
);
1625 texop
= nir_texop_txd
;
1626 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddx
);
1627 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddy
);
1630 if (operands
& SpvImageOperandsOffsetMask
||
1631 operands
& SpvImageOperandsConstOffsetMask
)
1632 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_offset
);
1634 if (operands
& SpvImageOperandsConstOffsetsMask
) {
1635 gather_offsets
= vtn_ssa_value(b
, w
[idx
++]);
1636 (*p
++) = (nir_tex_src
){};
1639 if (operands
& SpvImageOperandsSampleMask
) {
1640 assert(texop
== nir_texop_txf_ms
);
1641 texop
= nir_texop_txf_ms
;
1642 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ms_index
);
1645 /* We should have now consumed exactly all of the arguments */
1646 assert(idx
== count
);
1648 nir_tex_instr
*instr
= nir_tex_instr_create(b
->shader
, p
- srcs
);
1651 memcpy(instr
->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
1653 instr
->coord_components
= coord_components
;
1654 instr
->sampler_dim
= sampler_dim
;
1655 instr
->is_array
= is_array
;
1656 instr
->is_shadow
= is_shadow
;
1657 instr
->is_new_style_shadow
=
1658 is_shadow
&& glsl_get_components(ret_type
->type
) == 1;
1659 instr
->component
= gather_component
;
1661 switch (glsl_get_sampler_result_type(image_type
)) {
1662 case GLSL_TYPE_FLOAT
: instr
->dest_type
= nir_type_float
; break;
1663 case GLSL_TYPE_INT
: instr
->dest_type
= nir_type_int
; break;
1664 case GLSL_TYPE_UINT
: instr
->dest_type
= nir_type_uint
; break;
1665 case GLSL_TYPE_BOOL
: instr
->dest_type
= nir_type_bool
; break;
1667 unreachable("Invalid base type for sampler result");
1670 nir_deref_var
*sampler
= vtn_access_chain_to_deref(b
, sampled
.sampler
);
1671 nir_deref_var
*texture
;
1672 if (sampled
.image
) {
1673 nir_deref_var
*image
= vtn_access_chain_to_deref(b
, sampled
.image
);
1679 instr
->texture
= nir_deref_var_clone(texture
, instr
);
1681 switch (instr
->op
) {
1686 /* These operations require a sampler */
1687 instr
->sampler
= nir_deref_var_clone(sampler
, instr
);
1690 case nir_texop_txf_ms
:
1694 case nir_texop_query_levels
:
1695 case nir_texop_texture_samples
:
1696 case nir_texop_samples_identical
:
1698 instr
->sampler
= NULL
;
1700 case nir_texop_txf_ms_mcs
:
1701 unreachable("unexpected nir_texop_txf_ms_mcs");
1704 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1705 nir_tex_instr_dest_size(instr
), 32, NULL
);
1707 assert(glsl_get_vector_elements(ret_type
->type
) ==
1708 nir_tex_instr_dest_size(instr
));
1711 nir_instr
*instruction
;
1712 if (gather_offsets
) {
1713 assert(glsl_get_base_type(gather_offsets
->type
) == GLSL_TYPE_ARRAY
);
1714 assert(glsl_get_length(gather_offsets
->type
) == 4);
1715 nir_tex_instr
*instrs
[4] = {instr
, NULL
, NULL
, NULL
};
1717 /* Copy the current instruction 4x */
1718 for (uint32_t i
= 1; i
< 4; i
++) {
1719 instrs
[i
] = nir_tex_instr_create(b
->shader
, instr
->num_srcs
);
1720 instrs
[i
]->op
= instr
->op
;
1721 instrs
[i
]->coord_components
= instr
->coord_components
;
1722 instrs
[i
]->sampler_dim
= instr
->sampler_dim
;
1723 instrs
[i
]->is_array
= instr
->is_array
;
1724 instrs
[i
]->is_shadow
= instr
->is_shadow
;
1725 instrs
[i
]->is_new_style_shadow
= instr
->is_new_style_shadow
;
1726 instrs
[i
]->component
= instr
->component
;
1727 instrs
[i
]->dest_type
= instr
->dest_type
;
1728 instrs
[i
]->texture
= nir_deref_var_clone(texture
, instrs
[i
]);
1729 instrs
[i
]->sampler
= NULL
;
1731 memcpy(instrs
[i
]->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
1733 nir_ssa_dest_init(&instrs
[i
]->instr
, &instrs
[i
]->dest
,
1734 nir_tex_instr_dest_size(instr
), 32, NULL
);
1737 /* Fill in the last argument with the offset from the passed in offsets
1738 * and insert the instruction into the stream.
1740 for (uint32_t i
= 0; i
< 4; i
++) {
1742 src
.src
= nir_src_for_ssa(gather_offsets
->elems
[i
]->def
);
1743 src
.src_type
= nir_tex_src_offset
;
1744 instrs
[i
]->src
[instrs
[i
]->num_srcs
- 1] = src
;
1745 nir_builder_instr_insert(&b
->nb
, &instrs
[i
]->instr
);
1748 /* Combine the results of the 4 instructions by taking their .w
1751 nir_alu_instr
*vec4
= nir_alu_instr_create(b
->shader
, nir_op_vec4
);
1752 nir_ssa_dest_init(&vec4
->instr
, &vec4
->dest
.dest
, 4, 32, NULL
);
1753 vec4
->dest
.write_mask
= 0xf;
1754 for (uint32_t i
= 0; i
< 4; i
++) {
1755 vec4
->src
[i
].src
= nir_src_for_ssa(&instrs
[i
]->dest
.ssa
);
1756 vec4
->src
[i
].swizzle
[0] = 3;
1758 def
= &vec4
->dest
.dest
.ssa
;
1759 instruction
= &vec4
->instr
;
1761 def
= &instr
->dest
.ssa
;
1762 instruction
= &instr
->instr
;
1765 val
->ssa
= vtn_create_ssa_value(b
, ret_type
->type
);
1766 val
->ssa
->def
= def
;
1768 nir_builder_instr_insert(&b
->nb
, instruction
);
1772 fill_common_atomic_sources(struct vtn_builder
*b
, SpvOp opcode
,
1773 const uint32_t *w
, nir_src
*src
)
1776 case SpvOpAtomicIIncrement
:
1777 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
1780 case SpvOpAtomicIDecrement
:
1781 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
1784 case SpvOpAtomicISub
:
1786 nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
1789 case SpvOpAtomicCompareExchange
:
1790 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[8])->def
);
1791 src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
1794 case SpvOpAtomicExchange
:
1795 case SpvOpAtomicIAdd
:
1796 case SpvOpAtomicSMin
:
1797 case SpvOpAtomicUMin
:
1798 case SpvOpAtomicSMax
:
1799 case SpvOpAtomicUMax
:
1800 case SpvOpAtomicAnd
:
1802 case SpvOpAtomicXor
:
1803 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
1807 unreachable("Invalid SPIR-V atomic");
1811 static nir_ssa_def
*
1812 get_image_coord(struct vtn_builder
*b
, uint32_t value
)
1814 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, value
);
1816 /* The image_load_store intrinsics assume a 4-dim coordinate */
1817 unsigned dim
= glsl_get_vector_elements(coord
->type
);
1818 unsigned swizzle
[4];
1819 for (unsigned i
= 0; i
< 4; i
++)
1820 swizzle
[i
] = MIN2(i
, dim
- 1);
1822 return nir_swizzle(&b
->nb
, coord
->def
, swizzle
, 4, false);
1826 vtn_handle_image(struct vtn_builder
*b
, SpvOp opcode
,
1827 const uint32_t *w
, unsigned count
)
1829 /* Just get this one out of the way */
1830 if (opcode
== SpvOpImageTexelPointer
) {
1831 struct vtn_value
*val
=
1832 vtn_push_value(b
, w
[2], vtn_value_type_image_pointer
);
1833 val
->image
= ralloc(b
, struct vtn_image_pointer
);
1836 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1837 val
->image
->coord
= get_image_coord(b
, w
[4]);
1838 val
->image
->sample
= vtn_ssa_value(b
, w
[5])->def
;
1842 struct vtn_image_pointer image
;
1845 case SpvOpAtomicExchange
:
1846 case SpvOpAtomicCompareExchange
:
1847 case SpvOpAtomicCompareExchangeWeak
:
1848 case SpvOpAtomicIIncrement
:
1849 case SpvOpAtomicIDecrement
:
1850 case SpvOpAtomicIAdd
:
1851 case SpvOpAtomicISub
:
1852 case SpvOpAtomicLoad
:
1853 case SpvOpAtomicSMin
:
1854 case SpvOpAtomicUMin
:
1855 case SpvOpAtomicSMax
:
1856 case SpvOpAtomicUMax
:
1857 case SpvOpAtomicAnd
:
1859 case SpvOpAtomicXor
:
1860 image
= *vtn_value(b
, w
[3], vtn_value_type_image_pointer
)->image
;
1863 case SpvOpAtomicStore
:
1864 image
= *vtn_value(b
, w
[1], vtn_value_type_image_pointer
)->image
;
1867 case SpvOpImageQuerySize
:
1869 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1871 image
.sample
= NULL
;
1874 case SpvOpImageRead
:
1876 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1877 image
.coord
= get_image_coord(b
, w
[4]);
1879 if (count
> 5 && (w
[5] & SpvImageOperandsSampleMask
)) {
1880 assert(w
[5] == SpvImageOperandsSampleMask
);
1881 image
.sample
= vtn_ssa_value(b
, w
[6])->def
;
1883 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
1887 case SpvOpImageWrite
:
1889 vtn_value(b
, w
[1], vtn_value_type_access_chain
)->access_chain
;
1890 image
.coord
= get_image_coord(b
, w
[2]);
1894 if (count
> 4 && (w
[4] & SpvImageOperandsSampleMask
)) {
1895 assert(w
[4] == SpvImageOperandsSampleMask
);
1896 image
.sample
= vtn_ssa_value(b
, w
[5])->def
;
1898 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
1903 unreachable("Invalid image opcode");
1906 nir_intrinsic_op op
;
1908 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_##N; break;
1909 OP(ImageQuerySize
, size
)
1911 OP(ImageWrite
, store
)
1912 OP(AtomicLoad
, load
)
1913 OP(AtomicStore
, store
)
1914 OP(AtomicExchange
, atomic_exchange
)
1915 OP(AtomicCompareExchange
, atomic_comp_swap
)
1916 OP(AtomicIIncrement
, atomic_add
)
1917 OP(AtomicIDecrement
, atomic_add
)
1918 OP(AtomicIAdd
, atomic_add
)
1919 OP(AtomicISub
, atomic_add
)
1920 OP(AtomicSMin
, atomic_min
)
1921 OP(AtomicUMin
, atomic_min
)
1922 OP(AtomicSMax
, atomic_max
)
1923 OP(AtomicUMax
, atomic_max
)
1924 OP(AtomicAnd
, atomic_and
)
1925 OP(AtomicOr
, atomic_or
)
1926 OP(AtomicXor
, atomic_xor
)
1929 unreachable("Invalid image opcode");
1932 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
1934 nir_deref_var
*image_deref
= vtn_access_chain_to_deref(b
, image
.image
);
1935 intrin
->variables
[0] = nir_deref_var_clone(image_deref
, intrin
);
1937 /* ImageQuerySize doesn't take any extra parameters */
1938 if (opcode
!= SpvOpImageQuerySize
) {
1939 /* The image coordinate is always 4 components but we may not have that
1940 * many. Swizzle to compensate.
1943 for (unsigned i
= 0; i
< 4; i
++)
1944 swiz
[i
] = i
< image
.coord
->num_components
? i
: 0;
1945 intrin
->src
[0] = nir_src_for_ssa(nir_swizzle(&b
->nb
, image
.coord
,
1947 intrin
->src
[1] = nir_src_for_ssa(image
.sample
);
1951 case SpvOpAtomicLoad
:
1952 case SpvOpImageQuerySize
:
1953 case SpvOpImageRead
:
1955 case SpvOpAtomicStore
:
1956 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
1958 case SpvOpImageWrite
:
1959 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[3])->def
);
1962 case SpvOpAtomicIIncrement
:
1963 case SpvOpAtomicIDecrement
:
1964 case SpvOpAtomicExchange
:
1965 case SpvOpAtomicIAdd
:
1966 case SpvOpAtomicSMin
:
1967 case SpvOpAtomicUMin
:
1968 case SpvOpAtomicSMax
:
1969 case SpvOpAtomicUMax
:
1970 case SpvOpAtomicAnd
:
1972 case SpvOpAtomicXor
:
1973 fill_common_atomic_sources(b
, opcode
, w
, &intrin
->src
[2]);
1977 unreachable("Invalid image opcode");
1980 if (opcode
!= SpvOpImageWrite
) {
1981 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1982 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1983 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
, 4, 32, NULL
);
1985 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
1987 /* The image intrinsics always return 4 channels but we may not want
1988 * that many. Emit a mov to trim it down.
1990 unsigned swiz
[4] = {0, 1, 2, 3};
1991 val
->ssa
= vtn_create_ssa_value(b
, type
->type
);
1992 val
->ssa
->def
= nir_swizzle(&b
->nb
, &intrin
->dest
.ssa
, swiz
,
1993 glsl_get_vector_elements(type
->type
), false);
1995 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
1999 static nir_intrinsic_op
2000 get_ssbo_nir_atomic_op(SpvOp opcode
)
2003 case SpvOpAtomicLoad
: return nir_intrinsic_load_ssbo
;
2004 case SpvOpAtomicStore
: return nir_intrinsic_store_ssbo
;
2005 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
2006 OP(AtomicExchange
, atomic_exchange
)
2007 OP(AtomicCompareExchange
, atomic_comp_swap
)
2008 OP(AtomicIIncrement
, atomic_add
)
2009 OP(AtomicIDecrement
, atomic_add
)
2010 OP(AtomicIAdd
, atomic_add
)
2011 OP(AtomicISub
, atomic_add
)
2012 OP(AtomicSMin
, atomic_imin
)
2013 OP(AtomicUMin
, atomic_umin
)
2014 OP(AtomicSMax
, atomic_imax
)
2015 OP(AtomicUMax
, atomic_umax
)
2016 OP(AtomicAnd
, atomic_and
)
2017 OP(AtomicOr
, atomic_or
)
2018 OP(AtomicXor
, atomic_xor
)
2021 unreachable("Invalid SSBO atomic");
2025 static nir_intrinsic_op
2026 get_shared_nir_atomic_op(SpvOp opcode
)
2029 case SpvOpAtomicLoad
: return nir_intrinsic_load_var
;
2030 case SpvOpAtomicStore
: return nir_intrinsic_store_var
;
2031 #define OP(S, N) case SpvOp##S: return nir_intrinsic_var_##N;
2032 OP(AtomicExchange
, atomic_exchange
)
2033 OP(AtomicCompareExchange
, atomic_comp_swap
)
2034 OP(AtomicIIncrement
, atomic_add
)
2035 OP(AtomicIDecrement
, atomic_add
)
2036 OP(AtomicIAdd
, atomic_add
)
2037 OP(AtomicISub
, atomic_add
)
2038 OP(AtomicSMin
, atomic_imin
)
2039 OP(AtomicUMin
, atomic_umin
)
2040 OP(AtomicSMax
, atomic_imax
)
2041 OP(AtomicUMax
, atomic_umax
)
2042 OP(AtomicAnd
, atomic_and
)
2043 OP(AtomicOr
, atomic_or
)
2044 OP(AtomicXor
, atomic_xor
)
2047 unreachable("Invalid shared atomic");
2052 vtn_handle_ssbo_or_shared_atomic(struct vtn_builder
*b
, SpvOp opcode
,
2053 const uint32_t *w
, unsigned count
)
2055 struct vtn_access_chain
*chain
;
2056 nir_intrinsic_instr
*atomic
;
2059 case SpvOpAtomicLoad
:
2060 case SpvOpAtomicExchange
:
2061 case SpvOpAtomicCompareExchange
:
2062 case SpvOpAtomicCompareExchangeWeak
:
2063 case SpvOpAtomicIIncrement
:
2064 case SpvOpAtomicIDecrement
:
2065 case SpvOpAtomicIAdd
:
2066 case SpvOpAtomicISub
:
2067 case SpvOpAtomicSMin
:
2068 case SpvOpAtomicUMin
:
2069 case SpvOpAtomicSMax
:
2070 case SpvOpAtomicUMax
:
2071 case SpvOpAtomicAnd
:
2073 case SpvOpAtomicXor
:
2075 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
2078 case SpvOpAtomicStore
:
2080 vtn_value(b
, w
[1], vtn_value_type_access_chain
)->access_chain
;
2084 unreachable("Invalid SPIR-V atomic");
2088 SpvScope scope = w[4];
2089 SpvMemorySemanticsMask semantics = w[5];
2092 if (chain
->var
->mode
== vtn_variable_mode_workgroup
) {
2093 struct vtn_type
*type
= chain
->var
->type
;
2094 nir_deref_var
*deref
= vtn_access_chain_to_deref(b
, chain
);
2095 nir_intrinsic_op op
= get_shared_nir_atomic_op(opcode
);
2096 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2097 atomic
->variables
[0] = nir_deref_var_clone(deref
, atomic
);
2100 case SpvOpAtomicLoad
:
2101 atomic
->num_components
= glsl_get_vector_elements(type
->type
);
2104 case SpvOpAtomicStore
:
2105 atomic
->num_components
= glsl_get_vector_elements(type
->type
);
2106 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2107 atomic
->src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2110 case SpvOpAtomicExchange
:
2111 case SpvOpAtomicCompareExchange
:
2112 case SpvOpAtomicCompareExchangeWeak
:
2113 case SpvOpAtomicIIncrement
:
2114 case SpvOpAtomicIDecrement
:
2115 case SpvOpAtomicIAdd
:
2116 case SpvOpAtomicISub
:
2117 case SpvOpAtomicSMin
:
2118 case SpvOpAtomicUMin
:
2119 case SpvOpAtomicSMax
:
2120 case SpvOpAtomicUMax
:
2121 case SpvOpAtomicAnd
:
2123 case SpvOpAtomicXor
:
2124 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[0]);
2128 unreachable("Invalid SPIR-V atomic");
2132 assert(chain
->var
->mode
== vtn_variable_mode_ssbo
);
2133 struct vtn_type
*type
;
2134 nir_ssa_def
*offset
, *index
;
2135 offset
= vtn_access_chain_to_offset(b
, chain
, &index
, &type
, NULL
, false);
2137 nir_intrinsic_op op
= get_ssbo_nir_atomic_op(opcode
);
2139 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2142 case SpvOpAtomicLoad
:
2143 atomic
->num_components
= glsl_get_vector_elements(type
->type
);
2144 atomic
->src
[0] = nir_src_for_ssa(index
);
2145 atomic
->src
[1] = nir_src_for_ssa(offset
);
2148 case SpvOpAtomicStore
:
2149 atomic
->num_components
= glsl_get_vector_elements(type
->type
);
2150 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2151 atomic
->src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2152 atomic
->src
[1] = nir_src_for_ssa(index
);
2153 atomic
->src
[2] = nir_src_for_ssa(offset
);
2156 case SpvOpAtomicExchange
:
2157 case SpvOpAtomicCompareExchange
:
2158 case SpvOpAtomicCompareExchangeWeak
:
2159 case SpvOpAtomicIIncrement
:
2160 case SpvOpAtomicIDecrement
:
2161 case SpvOpAtomicIAdd
:
2162 case SpvOpAtomicISub
:
2163 case SpvOpAtomicSMin
:
2164 case SpvOpAtomicUMin
:
2165 case SpvOpAtomicSMax
:
2166 case SpvOpAtomicUMax
:
2167 case SpvOpAtomicAnd
:
2169 case SpvOpAtomicXor
:
2170 atomic
->src
[0] = nir_src_for_ssa(index
);
2171 atomic
->src
[1] = nir_src_for_ssa(offset
);
2172 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[2]);
2176 unreachable("Invalid SPIR-V atomic");
2180 if (opcode
!= SpvOpAtomicStore
) {
2181 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2183 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
,
2184 glsl_get_vector_elements(type
->type
),
2185 glsl_get_bit_size(type
->type
), NULL
);
2187 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2188 val
->ssa
= rzalloc(b
, struct vtn_ssa_value
);
2189 val
->ssa
->def
= &atomic
->dest
.ssa
;
2190 val
->ssa
->type
= type
->type
;
2193 nir_builder_instr_insert(&b
->nb
, &atomic
->instr
);
2196 static nir_alu_instr
*
2197 create_vec(nir_shader
*shader
, unsigned num_components
, unsigned bit_size
)
2200 switch (num_components
) {
2201 case 1: op
= nir_op_fmov
; break;
2202 case 2: op
= nir_op_vec2
; break;
2203 case 3: op
= nir_op_vec3
; break;
2204 case 4: op
= nir_op_vec4
; break;
2205 default: unreachable("bad vector size");
2208 nir_alu_instr
*vec
= nir_alu_instr_create(shader
, op
);
2209 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, num_components
,
2211 vec
->dest
.write_mask
= (1 << num_components
) - 1;
2216 struct vtn_ssa_value
*
2217 vtn_ssa_transpose(struct vtn_builder
*b
, struct vtn_ssa_value
*src
)
2219 if (src
->transposed
)
2220 return src
->transposed
;
2222 struct vtn_ssa_value
*dest
=
2223 vtn_create_ssa_value(b
, glsl_transposed_type(src
->type
));
2225 for (unsigned i
= 0; i
< glsl_get_matrix_columns(dest
->type
); i
++) {
2226 nir_alu_instr
*vec
= create_vec(b
->shader
,
2227 glsl_get_matrix_columns(src
->type
),
2228 glsl_get_bit_size(src
->type
));
2229 if (glsl_type_is_vector_or_scalar(src
->type
)) {
2230 vec
->src
[0].src
= nir_src_for_ssa(src
->def
);
2231 vec
->src
[0].swizzle
[0] = i
;
2233 for (unsigned j
= 0; j
< glsl_get_matrix_columns(src
->type
); j
++) {
2234 vec
->src
[j
].src
= nir_src_for_ssa(src
->elems
[j
]->def
);
2235 vec
->src
[j
].swizzle
[0] = i
;
2238 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2239 dest
->elems
[i
]->def
= &vec
->dest
.dest
.ssa
;
2242 dest
->transposed
= src
;
2248 vtn_vector_extract(struct vtn_builder
*b
, nir_ssa_def
*src
, unsigned index
)
2250 unsigned swiz
[4] = { index
};
2251 return nir_swizzle(&b
->nb
, src
, swiz
, 1, true);
2255 vtn_vector_insert(struct vtn_builder
*b
, nir_ssa_def
*src
, nir_ssa_def
*insert
,
2258 nir_alu_instr
*vec
= create_vec(b
->shader
, src
->num_components
,
2261 for (unsigned i
= 0; i
< src
->num_components
; i
++) {
2263 vec
->src
[i
].src
= nir_src_for_ssa(insert
);
2265 vec
->src
[i
].src
= nir_src_for_ssa(src
);
2266 vec
->src
[i
].swizzle
[0] = i
;
2270 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2272 return &vec
->dest
.dest
.ssa
;
2276 vtn_vector_extract_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
2279 nir_ssa_def
*dest
= vtn_vector_extract(b
, src
, 0);
2280 for (unsigned i
= 1; i
< src
->num_components
; i
++)
2281 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
2282 vtn_vector_extract(b
, src
, i
), dest
);
2288 vtn_vector_insert_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
2289 nir_ssa_def
*insert
, nir_ssa_def
*index
)
2291 nir_ssa_def
*dest
= vtn_vector_insert(b
, src
, insert
, 0);
2292 for (unsigned i
= 1; i
< src
->num_components
; i
++)
2293 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
2294 vtn_vector_insert(b
, src
, insert
, i
), dest
);
2299 static nir_ssa_def
*
2300 vtn_vector_shuffle(struct vtn_builder
*b
, unsigned num_components
,
2301 nir_ssa_def
*src0
, nir_ssa_def
*src1
,
2302 const uint32_t *indices
)
2304 nir_alu_instr
*vec
= create_vec(b
->shader
, num_components
, src0
->bit_size
);
2306 for (unsigned i
= 0; i
< num_components
; i
++) {
2307 uint32_t index
= indices
[i
];
2308 if (index
== 0xffffffff) {
2310 nir_src_for_ssa(nir_ssa_undef(&b
->nb
, 1, src0
->bit_size
));
2311 } else if (index
< src0
->num_components
) {
2312 vec
->src
[i
].src
= nir_src_for_ssa(src0
);
2313 vec
->src
[i
].swizzle
[0] = index
;
2315 vec
->src
[i
].src
= nir_src_for_ssa(src1
);
2316 vec
->src
[i
].swizzle
[0] = index
- src0
->num_components
;
2320 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2322 return &vec
->dest
.dest
.ssa
;
2326 * Concatentates a number of vectors/scalars together to produce a vector
2328 static nir_ssa_def
*
2329 vtn_vector_construct(struct vtn_builder
*b
, unsigned num_components
,
2330 unsigned num_srcs
, nir_ssa_def
**srcs
)
2332 nir_alu_instr
*vec
= create_vec(b
->shader
, num_components
,
2335 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2337 * "When constructing a vector, there must be at least two Constituent
2340 assert(num_srcs
>= 2);
2342 unsigned dest_idx
= 0;
2343 for (unsigned i
= 0; i
< num_srcs
; i
++) {
2344 nir_ssa_def
*src
= srcs
[i
];
2345 assert(dest_idx
+ src
->num_components
<= num_components
);
2346 for (unsigned j
= 0; j
< src
->num_components
; j
++) {
2347 vec
->src
[dest_idx
].src
= nir_src_for_ssa(src
);
2348 vec
->src
[dest_idx
].swizzle
[0] = j
;
2353 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2355 * "When constructing a vector, the total number of components in all
2356 * the operands must equal the number of components in Result Type."
2358 assert(dest_idx
== num_components
);
2360 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2362 return &vec
->dest
.dest
.ssa
;
2365 static struct vtn_ssa_value
*
2366 vtn_composite_copy(void *mem_ctx
, struct vtn_ssa_value
*src
)
2368 struct vtn_ssa_value
*dest
= rzalloc(mem_ctx
, struct vtn_ssa_value
);
2369 dest
->type
= src
->type
;
2371 if (glsl_type_is_vector_or_scalar(src
->type
)) {
2372 dest
->def
= src
->def
;
2374 unsigned elems
= glsl_get_length(src
->type
);
2376 dest
->elems
= ralloc_array(mem_ctx
, struct vtn_ssa_value
*, elems
);
2377 for (unsigned i
= 0; i
< elems
; i
++)
2378 dest
->elems
[i
] = vtn_composite_copy(mem_ctx
, src
->elems
[i
]);
2384 static struct vtn_ssa_value
*
2385 vtn_composite_insert(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
2386 struct vtn_ssa_value
*insert
, const uint32_t *indices
,
2387 unsigned num_indices
)
2389 struct vtn_ssa_value
*dest
= vtn_composite_copy(b
, src
);
2391 struct vtn_ssa_value
*cur
= dest
;
2393 for (i
= 0; i
< num_indices
- 1; i
++) {
2394 cur
= cur
->elems
[indices
[i
]];
2397 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
2398 /* According to the SPIR-V spec, OpCompositeInsert may work down to
2399 * the component granularity. In that case, the last index will be
2400 * the index to insert the scalar into the vector.
2403 cur
->def
= vtn_vector_insert(b
, cur
->def
, insert
->def
, indices
[i
]);
2405 cur
->elems
[indices
[i
]] = insert
;
2411 static struct vtn_ssa_value
*
2412 vtn_composite_extract(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
2413 const uint32_t *indices
, unsigned num_indices
)
2415 struct vtn_ssa_value
*cur
= src
;
2416 for (unsigned i
= 0; i
< num_indices
; i
++) {
2417 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
2418 assert(i
== num_indices
- 1);
2419 /* According to the SPIR-V spec, OpCompositeExtract may work down to
2420 * the component granularity. The last index will be the index of the
2421 * vector to extract.
2424 struct vtn_ssa_value
*ret
= rzalloc(b
, struct vtn_ssa_value
);
2425 ret
->type
= glsl_scalar_type(glsl_get_base_type(cur
->type
));
2426 ret
->def
= vtn_vector_extract(b
, cur
->def
, indices
[i
]);
2429 cur
= cur
->elems
[indices
[i
]];
2437 vtn_handle_composite(struct vtn_builder
*b
, SpvOp opcode
,
2438 const uint32_t *w
, unsigned count
)
2440 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2441 const struct glsl_type
*type
=
2442 vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
2443 val
->ssa
= vtn_create_ssa_value(b
, type
);
2446 case SpvOpVectorExtractDynamic
:
2447 val
->ssa
->def
= vtn_vector_extract_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
2448 vtn_ssa_value(b
, w
[4])->def
);
2451 case SpvOpVectorInsertDynamic
:
2452 val
->ssa
->def
= vtn_vector_insert_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
2453 vtn_ssa_value(b
, w
[4])->def
,
2454 vtn_ssa_value(b
, w
[5])->def
);
2457 case SpvOpVectorShuffle
:
2458 val
->ssa
->def
= vtn_vector_shuffle(b
, glsl_get_vector_elements(type
),
2459 vtn_ssa_value(b
, w
[3])->def
,
2460 vtn_ssa_value(b
, w
[4])->def
,
2464 case SpvOpCompositeConstruct
: {
2465 unsigned elems
= count
- 3;
2466 if (glsl_type_is_vector_or_scalar(type
)) {
2467 nir_ssa_def
*srcs
[4];
2468 for (unsigned i
= 0; i
< elems
; i
++)
2469 srcs
[i
] = vtn_ssa_value(b
, w
[3 + i
])->def
;
2471 vtn_vector_construct(b
, glsl_get_vector_elements(type
),
2474 val
->ssa
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
2475 for (unsigned i
= 0; i
< elems
; i
++)
2476 val
->ssa
->elems
[i
] = vtn_ssa_value(b
, w
[3 + i
]);
2480 case SpvOpCompositeExtract
:
2481 val
->ssa
= vtn_composite_extract(b
, vtn_ssa_value(b
, w
[3]),
2485 case SpvOpCompositeInsert
:
2486 val
->ssa
= vtn_composite_insert(b
, vtn_ssa_value(b
, w
[4]),
2487 vtn_ssa_value(b
, w
[3]),
2491 case SpvOpCopyObject
:
2492 val
->ssa
= vtn_composite_copy(b
, vtn_ssa_value(b
, w
[3]));
2496 unreachable("unknown composite operation");
2501 vtn_handle_barrier(struct vtn_builder
*b
, SpvOp opcode
,
2502 const uint32_t *w
, unsigned count
)
2504 nir_intrinsic_op intrinsic_op
;
2506 case SpvOpEmitVertex
:
2507 case SpvOpEmitStreamVertex
:
2508 intrinsic_op
= nir_intrinsic_emit_vertex
;
2510 case SpvOpEndPrimitive
:
2511 case SpvOpEndStreamPrimitive
:
2512 intrinsic_op
= nir_intrinsic_end_primitive
;
2514 case SpvOpMemoryBarrier
:
2515 intrinsic_op
= nir_intrinsic_memory_barrier
;
2517 case SpvOpControlBarrier
:
2518 intrinsic_op
= nir_intrinsic_barrier
;
2521 unreachable("unknown barrier instruction");
2524 nir_intrinsic_instr
*intrin
=
2525 nir_intrinsic_instr_create(b
->shader
, intrinsic_op
);
2527 if (opcode
== SpvOpEmitStreamVertex
|| opcode
== SpvOpEndStreamPrimitive
)
2528 nir_intrinsic_set_stream_id(intrin
, w
[1]);
2530 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2534 gl_primitive_from_spv_execution_mode(SpvExecutionMode mode
)
2537 case SpvExecutionModeInputPoints
:
2538 case SpvExecutionModeOutputPoints
:
2539 return 0; /* GL_POINTS */
2540 case SpvExecutionModeInputLines
:
2541 return 1; /* GL_LINES */
2542 case SpvExecutionModeInputLinesAdjacency
:
2543 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
2544 case SpvExecutionModeTriangles
:
2545 return 4; /* GL_TRIANGLES */
2546 case SpvExecutionModeInputTrianglesAdjacency
:
2547 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
2548 case SpvExecutionModeQuads
:
2549 return 7; /* GL_QUADS */
2550 case SpvExecutionModeIsolines
:
2551 return 0x8E7A; /* GL_ISOLINES */
2552 case SpvExecutionModeOutputLineStrip
:
2553 return 3; /* GL_LINE_STRIP */
2554 case SpvExecutionModeOutputTriangleStrip
:
2555 return 5; /* GL_TRIANGLE_STRIP */
2557 assert(!"Invalid primitive type");
2563 vertices_in_from_spv_execution_mode(SpvExecutionMode mode
)
2566 case SpvExecutionModeInputPoints
:
2568 case SpvExecutionModeInputLines
:
2570 case SpvExecutionModeInputLinesAdjacency
:
2572 case SpvExecutionModeTriangles
:
2574 case SpvExecutionModeInputTrianglesAdjacency
:
2577 assert(!"Invalid GS input mode");
2582 static gl_shader_stage
2583 stage_for_execution_model(SpvExecutionModel model
)
2586 case SpvExecutionModelVertex
:
2587 return MESA_SHADER_VERTEX
;
2588 case SpvExecutionModelTessellationControl
:
2589 return MESA_SHADER_TESS_CTRL
;
2590 case SpvExecutionModelTessellationEvaluation
:
2591 return MESA_SHADER_TESS_EVAL
;
2592 case SpvExecutionModelGeometry
:
2593 return MESA_SHADER_GEOMETRY
;
2594 case SpvExecutionModelFragment
:
2595 return MESA_SHADER_FRAGMENT
;
2596 case SpvExecutionModelGLCompute
:
2597 return MESA_SHADER_COMPUTE
;
2599 unreachable("Unsupported execution model");
2603 #define spv_check_supported(name, cap) do { \
2604 if (!(b->ext && b->ext->name)) \
2605 vtn_warn("Unsupported SPIR-V capability: %s", \
2606 spirv_capability_to_string(cap)); \
2610 vtn_handle_preamble_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2611 const uint32_t *w
, unsigned count
)
2615 case SpvOpSourceExtension
:
2616 case SpvOpSourceContinued
:
2617 case SpvOpExtension
:
2618 /* Unhandled, but these are for debug so that's ok. */
2621 case SpvOpCapability
: {
2622 SpvCapability cap
= w
[1];
2624 case SpvCapabilityMatrix
:
2625 case SpvCapabilityShader
:
2626 case SpvCapabilityGeometry
:
2627 case SpvCapabilityGeometryPointSize
:
2628 case SpvCapabilityUniformBufferArrayDynamicIndexing
:
2629 case SpvCapabilitySampledImageArrayDynamicIndexing
:
2630 case SpvCapabilityStorageBufferArrayDynamicIndexing
:
2631 case SpvCapabilityStorageImageArrayDynamicIndexing
:
2632 case SpvCapabilityImageRect
:
2633 case SpvCapabilitySampledRect
:
2634 case SpvCapabilitySampled1D
:
2635 case SpvCapabilityImage1D
:
2636 case SpvCapabilitySampledCubeArray
:
2637 case SpvCapabilitySampledBuffer
:
2638 case SpvCapabilityImageBuffer
:
2639 case SpvCapabilityImageQuery
:
2640 case SpvCapabilityDerivativeControl
:
2641 case SpvCapabilityInterpolationFunction
:
2642 case SpvCapabilityMultiViewport
:
2643 case SpvCapabilitySampleRateShading
:
2644 case SpvCapabilityClipDistance
:
2645 case SpvCapabilityCullDistance
:
2646 case SpvCapabilityInputAttachment
:
2647 case SpvCapabilityImageGatherExtended
:
2648 case SpvCapabilityStorageImageExtendedFormats
:
2651 case SpvCapabilityGeometryStreams
:
2652 case SpvCapabilityLinkage
:
2653 case SpvCapabilityVector16
:
2654 case SpvCapabilityFloat16Buffer
:
2655 case SpvCapabilityFloat16
:
2656 case SpvCapabilityInt64
:
2657 case SpvCapabilityInt64Atomics
:
2658 case SpvCapabilityAtomicStorage
:
2659 case SpvCapabilityInt16
:
2660 case SpvCapabilityStorageImageMultisample
:
2661 case SpvCapabilityImageCubeArray
:
2662 case SpvCapabilityInt8
:
2663 case SpvCapabilitySparseResidency
:
2664 case SpvCapabilityMinLod
:
2665 case SpvCapabilityTransformFeedback
:
2666 vtn_warn("Unsupported SPIR-V capability: %s",
2667 spirv_capability_to_string(cap
));
2670 case SpvCapabilityFloat64
:
2671 spv_check_supported(float64
, cap
);
2674 case SpvCapabilityAddresses
:
2675 case SpvCapabilityKernel
:
2676 case SpvCapabilityImageBasic
:
2677 case SpvCapabilityImageReadWrite
:
2678 case SpvCapabilityImageMipmap
:
2679 case SpvCapabilityPipes
:
2680 case SpvCapabilityGroups
:
2681 case SpvCapabilityDeviceEnqueue
:
2682 case SpvCapabilityLiteralSampler
:
2683 case SpvCapabilityGenericPointer
:
2684 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
2685 spirv_capability_to_string(cap
));
2688 case SpvCapabilityImageMSArray
:
2689 spv_check_supported(image_ms_array
, cap
);
2692 case SpvCapabilityTessellation
:
2693 case SpvCapabilityTessellationPointSize
:
2694 spv_check_supported(tessellation
, cap
);
2697 case SpvCapabilityDrawParameters
:
2698 spv_check_supported(draw_parameters
, cap
);
2701 case SpvCapabilityStorageImageReadWithoutFormat
:
2702 spv_check_supported(image_read_without_format
, cap
);
2705 case SpvCapabilityStorageImageWriteWithoutFormat
:
2706 spv_check_supported(image_write_without_format
, cap
);
2710 unreachable("Unhandled capability");
2715 case SpvOpExtInstImport
:
2716 vtn_handle_extension(b
, opcode
, w
, count
);
2719 case SpvOpMemoryModel
:
2720 assert(w
[1] == SpvAddressingModelLogical
);
2721 assert(w
[2] == SpvMemoryModelGLSL450
);
2724 case SpvOpEntryPoint
: {
2725 struct vtn_value
*entry_point
= &b
->values
[w
[2]];
2726 /* Let this be a name label regardless */
2727 unsigned name_words
;
2728 entry_point
->name
= vtn_string_literal(b
, &w
[3], count
- 3, &name_words
);
2730 if (strcmp(entry_point
->name
, b
->entry_point_name
) != 0 ||
2731 stage_for_execution_model(w
[1]) != b
->entry_point_stage
)
2734 assert(b
->entry_point
== NULL
);
2735 b
->entry_point
= entry_point
;
2740 vtn_push_value(b
, w
[1], vtn_value_type_string
)->str
=
2741 vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
2745 b
->values
[w
[1]].name
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
2748 case SpvOpMemberName
:
2752 case SpvOpExecutionMode
:
2753 case SpvOpDecorationGroup
:
2755 case SpvOpMemberDecorate
:
2756 case SpvOpGroupDecorate
:
2757 case SpvOpGroupMemberDecorate
:
2758 vtn_handle_decoration(b
, opcode
, w
, count
);
2762 return false; /* End of preamble */
2769 vtn_handle_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
2770 const struct vtn_decoration
*mode
, void *data
)
2772 assert(b
->entry_point
== entry_point
);
2774 switch(mode
->exec_mode
) {
2775 case SpvExecutionModeOriginUpperLeft
:
2776 case SpvExecutionModeOriginLowerLeft
:
2777 b
->origin_upper_left
=
2778 (mode
->exec_mode
== SpvExecutionModeOriginUpperLeft
);
2781 case SpvExecutionModeEarlyFragmentTests
:
2782 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2783 b
->shader
->info
->fs
.early_fragment_tests
= true;
2786 case SpvExecutionModeInvocations
:
2787 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2788 b
->shader
->info
->gs
.invocations
= MAX2(1, mode
->literals
[0]);
2791 case SpvExecutionModeDepthReplacing
:
2792 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2793 b
->shader
->info
->fs
.depth_layout
= FRAG_DEPTH_LAYOUT_ANY
;
2795 case SpvExecutionModeDepthGreater
:
2796 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2797 b
->shader
->info
->fs
.depth_layout
= FRAG_DEPTH_LAYOUT_GREATER
;
2799 case SpvExecutionModeDepthLess
:
2800 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2801 b
->shader
->info
->fs
.depth_layout
= FRAG_DEPTH_LAYOUT_LESS
;
2803 case SpvExecutionModeDepthUnchanged
:
2804 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2805 b
->shader
->info
->fs
.depth_layout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
2808 case SpvExecutionModeLocalSize
:
2809 assert(b
->shader
->stage
== MESA_SHADER_COMPUTE
);
2810 b
->shader
->info
->cs
.local_size
[0] = mode
->literals
[0];
2811 b
->shader
->info
->cs
.local_size
[1] = mode
->literals
[1];
2812 b
->shader
->info
->cs
.local_size
[2] = mode
->literals
[2];
2814 case SpvExecutionModeLocalSizeHint
:
2815 break; /* Nothing to do with this */
2817 case SpvExecutionModeOutputVertices
:
2818 if (b
->shader
->stage
== MESA_SHADER_TESS_CTRL
||
2819 b
->shader
->stage
== MESA_SHADER_TESS_EVAL
) {
2820 b
->shader
->info
->tess
.tcs_vertices_out
= mode
->literals
[0];
2822 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2823 b
->shader
->info
->gs
.vertices_out
= mode
->literals
[0];
2827 case SpvExecutionModeInputPoints
:
2828 case SpvExecutionModeInputLines
:
2829 case SpvExecutionModeInputLinesAdjacency
:
2830 case SpvExecutionModeTriangles
:
2831 case SpvExecutionModeInputTrianglesAdjacency
:
2832 case SpvExecutionModeQuads
:
2833 case SpvExecutionModeIsolines
:
2834 if (b
->shader
->stage
== MESA_SHADER_TESS_CTRL
||
2835 b
->shader
->stage
== MESA_SHADER_TESS_EVAL
) {
2836 b
->shader
->info
->tess
.primitive_mode
=
2837 gl_primitive_from_spv_execution_mode(mode
->exec_mode
);
2839 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2840 b
->shader
->info
->gs
.vertices_in
=
2841 vertices_in_from_spv_execution_mode(mode
->exec_mode
);
2845 case SpvExecutionModeOutputPoints
:
2846 case SpvExecutionModeOutputLineStrip
:
2847 case SpvExecutionModeOutputTriangleStrip
:
2848 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2849 b
->shader
->info
->gs
.output_primitive
=
2850 gl_primitive_from_spv_execution_mode(mode
->exec_mode
);
2853 case SpvExecutionModeSpacingEqual
:
2854 assert(b
->shader
->stage
== MESA_SHADER_TESS_CTRL
||
2855 b
->shader
->stage
== MESA_SHADER_TESS_EVAL
);
2856 b
->shader
->info
->tess
.spacing
= TESS_SPACING_EQUAL
;
2858 case SpvExecutionModeSpacingFractionalEven
:
2859 assert(b
->shader
->stage
== MESA_SHADER_TESS_CTRL
||
2860 b
->shader
->stage
== MESA_SHADER_TESS_EVAL
);
2861 b
->shader
->info
->tess
.spacing
= TESS_SPACING_FRACTIONAL_EVEN
;
2863 case SpvExecutionModeSpacingFractionalOdd
:
2864 assert(b
->shader
->stage
== MESA_SHADER_TESS_CTRL
||
2865 b
->shader
->stage
== MESA_SHADER_TESS_EVAL
);
2866 b
->shader
->info
->tess
.spacing
= TESS_SPACING_FRACTIONAL_ODD
;
2868 case SpvExecutionModeVertexOrderCw
:
2869 assert(b
->shader
->stage
== MESA_SHADER_TESS_CTRL
||
2870 b
->shader
->stage
== MESA_SHADER_TESS_EVAL
);
2871 /* Vulkan's notion of CCW seems to match the hardware backends,
2872 * but be the opposite of OpenGL. Currently NIR follows GL semantics,
2873 * so we set it backwards here.
2875 b
->shader
->info
->tess
.ccw
= true;
2877 case SpvExecutionModeVertexOrderCcw
:
2878 assert(b
->shader
->stage
== MESA_SHADER_TESS_CTRL
||
2879 b
->shader
->stage
== MESA_SHADER_TESS_EVAL
);
2880 /* Backwards; see above */
2881 b
->shader
->info
->tess
.ccw
= false;
2883 case SpvExecutionModePointMode
:
2884 assert(b
->shader
->stage
== MESA_SHADER_TESS_CTRL
||
2885 b
->shader
->stage
== MESA_SHADER_TESS_EVAL
);
2886 b
->shader
->info
->tess
.point_mode
= true;
2889 case SpvExecutionModePixelCenterInteger
:
2890 b
->pixel_center_integer
= true;
2893 case SpvExecutionModeXfb
:
2894 assert(!"Unhandled execution mode");
2897 case SpvExecutionModeVecTypeHint
:
2898 case SpvExecutionModeContractionOff
:
2902 unreachable("Unhandled execution mode");
2907 vtn_handle_variable_or_type_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2908 const uint32_t *w
, unsigned count
)
2912 case SpvOpSourceContinued
:
2913 case SpvOpSourceExtension
:
2914 case SpvOpExtension
:
2915 case SpvOpCapability
:
2916 case SpvOpExtInstImport
:
2917 case SpvOpMemoryModel
:
2918 case SpvOpEntryPoint
:
2919 case SpvOpExecutionMode
:
2922 case SpvOpMemberName
:
2923 case SpvOpDecorationGroup
:
2925 case SpvOpMemberDecorate
:
2926 case SpvOpGroupDecorate
:
2927 case SpvOpGroupMemberDecorate
:
2928 assert(!"Invalid opcode types and variables section");
2934 case SpvOpTypeFloat
:
2935 case SpvOpTypeVector
:
2936 case SpvOpTypeMatrix
:
2937 case SpvOpTypeImage
:
2938 case SpvOpTypeSampler
:
2939 case SpvOpTypeSampledImage
:
2940 case SpvOpTypeArray
:
2941 case SpvOpTypeRuntimeArray
:
2942 case SpvOpTypeStruct
:
2943 case SpvOpTypeOpaque
:
2944 case SpvOpTypePointer
:
2945 case SpvOpTypeFunction
:
2946 case SpvOpTypeEvent
:
2947 case SpvOpTypeDeviceEvent
:
2948 case SpvOpTypeReserveId
:
2949 case SpvOpTypeQueue
:
2951 vtn_handle_type(b
, opcode
, w
, count
);
2954 case SpvOpConstantTrue
:
2955 case SpvOpConstantFalse
:
2957 case SpvOpConstantComposite
:
2958 case SpvOpConstantSampler
:
2959 case SpvOpConstantNull
:
2960 case SpvOpSpecConstantTrue
:
2961 case SpvOpSpecConstantFalse
:
2962 case SpvOpSpecConstant
:
2963 case SpvOpSpecConstantComposite
:
2964 case SpvOpSpecConstantOp
:
2965 vtn_handle_constant(b
, opcode
, w
, count
);
2970 vtn_handle_variables(b
, opcode
, w
, count
);
2974 return false; /* End of preamble */
2981 vtn_handle_body_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2982 const uint32_t *w
, unsigned count
)
2988 case SpvOpLoopMerge
:
2989 case SpvOpSelectionMerge
:
2990 /* This is handled by cfg pre-pass and walk_blocks */
2994 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
2995 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
3000 vtn_handle_extension(b
, opcode
, w
, count
);
3006 case SpvOpCopyMemory
:
3007 case SpvOpCopyMemorySized
:
3008 case SpvOpAccessChain
:
3009 case SpvOpInBoundsAccessChain
:
3010 case SpvOpArrayLength
:
3011 vtn_handle_variables(b
, opcode
, w
, count
);
3014 case SpvOpFunctionCall
:
3015 vtn_handle_function_call(b
, opcode
, w
, count
);
3018 case SpvOpSampledImage
:
3020 case SpvOpImageSampleImplicitLod
:
3021 case SpvOpImageSampleExplicitLod
:
3022 case SpvOpImageSampleDrefImplicitLod
:
3023 case SpvOpImageSampleDrefExplicitLod
:
3024 case SpvOpImageSampleProjImplicitLod
:
3025 case SpvOpImageSampleProjExplicitLod
:
3026 case SpvOpImageSampleProjDrefImplicitLod
:
3027 case SpvOpImageSampleProjDrefExplicitLod
:
3028 case SpvOpImageFetch
:
3029 case SpvOpImageGather
:
3030 case SpvOpImageDrefGather
:
3031 case SpvOpImageQuerySizeLod
:
3032 case SpvOpImageQueryLod
:
3033 case SpvOpImageQueryLevels
:
3034 case SpvOpImageQuerySamples
:
3035 vtn_handle_texture(b
, opcode
, w
, count
);
3038 case SpvOpImageRead
:
3039 case SpvOpImageWrite
:
3040 case SpvOpImageTexelPointer
:
3041 vtn_handle_image(b
, opcode
, w
, count
);
3044 case SpvOpImageQuerySize
: {
3045 struct vtn_access_chain
*image
=
3046 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
3047 if (glsl_type_is_image(image
->var
->var
->interface_type
)) {
3048 vtn_handle_image(b
, opcode
, w
, count
);
3050 vtn_handle_texture(b
, opcode
, w
, count
);
3055 case SpvOpAtomicLoad
:
3056 case SpvOpAtomicExchange
:
3057 case SpvOpAtomicCompareExchange
:
3058 case SpvOpAtomicCompareExchangeWeak
:
3059 case SpvOpAtomicIIncrement
:
3060 case SpvOpAtomicIDecrement
:
3061 case SpvOpAtomicIAdd
:
3062 case SpvOpAtomicISub
:
3063 case SpvOpAtomicSMin
:
3064 case SpvOpAtomicUMin
:
3065 case SpvOpAtomicSMax
:
3066 case SpvOpAtomicUMax
:
3067 case SpvOpAtomicAnd
:
3069 case SpvOpAtomicXor
: {
3070 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[3]);
3071 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
3072 vtn_handle_image(b
, opcode
, w
, count
);
3074 assert(pointer
->value_type
== vtn_value_type_access_chain
);
3075 vtn_handle_ssbo_or_shared_atomic(b
, opcode
, w
, count
);
3080 case SpvOpAtomicStore
: {
3081 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[1]);
3082 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
3083 vtn_handle_image(b
, opcode
, w
, count
);
3085 assert(pointer
->value_type
== vtn_value_type_access_chain
);
3086 vtn_handle_ssbo_or_shared_atomic(b
, opcode
, w
, count
);
3096 case SpvOpConvertFToU
:
3097 case SpvOpConvertFToS
:
3098 case SpvOpConvertSToF
:
3099 case SpvOpConvertUToF
:
3103 case SpvOpQuantizeToF16
:
3104 case SpvOpConvertPtrToU
:
3105 case SpvOpConvertUToPtr
:
3106 case SpvOpPtrCastToGeneric
:
3107 case SpvOpGenericCastToPtr
:
3113 case SpvOpSignBitSet
:
3114 case SpvOpLessOrGreater
:
3116 case SpvOpUnordered
:
3131 case SpvOpVectorTimesScalar
:
3133 case SpvOpIAddCarry
:
3134 case SpvOpISubBorrow
:
3135 case SpvOpUMulExtended
:
3136 case SpvOpSMulExtended
:
3137 case SpvOpShiftRightLogical
:
3138 case SpvOpShiftRightArithmetic
:
3139 case SpvOpShiftLeftLogical
:
3140 case SpvOpLogicalEqual
:
3141 case SpvOpLogicalNotEqual
:
3142 case SpvOpLogicalOr
:
3143 case SpvOpLogicalAnd
:
3144 case SpvOpLogicalNot
:
3145 case SpvOpBitwiseOr
:
3146 case SpvOpBitwiseXor
:
3147 case SpvOpBitwiseAnd
:
3150 case SpvOpFOrdEqual
:
3151 case SpvOpFUnordEqual
:
3152 case SpvOpINotEqual
:
3153 case SpvOpFOrdNotEqual
:
3154 case SpvOpFUnordNotEqual
:
3155 case SpvOpULessThan
:
3156 case SpvOpSLessThan
:
3157 case SpvOpFOrdLessThan
:
3158 case SpvOpFUnordLessThan
:
3159 case SpvOpUGreaterThan
:
3160 case SpvOpSGreaterThan
:
3161 case SpvOpFOrdGreaterThan
:
3162 case SpvOpFUnordGreaterThan
:
3163 case SpvOpULessThanEqual
:
3164 case SpvOpSLessThanEqual
:
3165 case SpvOpFOrdLessThanEqual
:
3166 case SpvOpFUnordLessThanEqual
:
3167 case SpvOpUGreaterThanEqual
:
3168 case SpvOpSGreaterThanEqual
:
3169 case SpvOpFOrdGreaterThanEqual
:
3170 case SpvOpFUnordGreaterThanEqual
:
3176 case SpvOpFwidthFine
:
3177 case SpvOpDPdxCoarse
:
3178 case SpvOpDPdyCoarse
:
3179 case SpvOpFwidthCoarse
:
3180 case SpvOpBitFieldInsert
:
3181 case SpvOpBitFieldSExtract
:
3182 case SpvOpBitFieldUExtract
:
3183 case SpvOpBitReverse
:
3185 case SpvOpTranspose
:
3186 case SpvOpOuterProduct
:
3187 case SpvOpMatrixTimesScalar
:
3188 case SpvOpVectorTimesMatrix
:
3189 case SpvOpMatrixTimesVector
:
3190 case SpvOpMatrixTimesMatrix
:
3191 vtn_handle_alu(b
, opcode
, w
, count
);
3194 case SpvOpVectorExtractDynamic
:
3195 case SpvOpVectorInsertDynamic
:
3196 case SpvOpVectorShuffle
:
3197 case SpvOpCompositeConstruct
:
3198 case SpvOpCompositeExtract
:
3199 case SpvOpCompositeInsert
:
3200 case SpvOpCopyObject
:
3201 vtn_handle_composite(b
, opcode
, w
, count
);
3204 case SpvOpEmitVertex
:
3205 case SpvOpEndPrimitive
:
3206 case SpvOpEmitStreamVertex
:
3207 case SpvOpEndStreamPrimitive
:
3208 case SpvOpControlBarrier
:
3209 case SpvOpMemoryBarrier
:
3210 vtn_handle_barrier(b
, opcode
, w
, count
);
3214 unreachable("Unhandled opcode");
3221 spirv_to_nir(const uint32_t *words
, size_t word_count
,
3222 struct nir_spirv_specialization
*spec
, unsigned num_spec
,
3223 gl_shader_stage stage
, const char *entry_point_name
,
3224 const struct nir_spirv_supported_extensions
*ext
,
3225 const nir_shader_compiler_options
*options
)
3227 const uint32_t *word_end
= words
+ word_count
;
3229 /* Handle the SPIR-V header (first 4 dwords) */
3230 assert(word_count
> 5);
3232 assert(words
[0] == SpvMagicNumber
);
3233 assert(words
[1] >= 0x10000);
3234 /* words[2] == generator magic */
3235 unsigned value_id_bound
= words
[3];
3236 assert(words
[4] == 0);
3240 /* Initialize the stn_builder object */
3241 struct vtn_builder
*b
= rzalloc(NULL
, struct vtn_builder
);
3242 b
->value_id_bound
= value_id_bound
;
3243 b
->values
= rzalloc_array(b
, struct vtn_value
, value_id_bound
);
3244 exec_list_make_empty(&b
->functions
);
3245 b
->entry_point_stage
= stage
;
3246 b
->entry_point_name
= entry_point_name
;
3249 /* Handle all the preamble instructions */
3250 words
= vtn_foreach_instruction(b
, words
, word_end
,
3251 vtn_handle_preamble_instruction
);
3253 if (b
->entry_point
== NULL
) {
3254 assert(!"Entry point not found");
3259 b
->shader
= nir_shader_create(NULL
, stage
, options
, NULL
);
3261 /* Set shader info defaults */
3262 b
->shader
->info
->gs
.invocations
= 1;
3264 /* Parse execution modes */
3265 vtn_foreach_execution_mode(b
, b
->entry_point
,
3266 vtn_handle_execution_mode
, NULL
);
3268 b
->specializations
= spec
;
3269 b
->num_specializations
= num_spec
;
3271 /* Handle all variable, type, and constant instructions */
3272 words
= vtn_foreach_instruction(b
, words
, word_end
,
3273 vtn_handle_variable_or_type_instruction
);
3275 vtn_build_cfg(b
, words
, word_end
);
3277 foreach_list_typed(struct vtn_function
, func
, node
, &b
->functions
) {
3278 b
->impl
= func
->impl
;
3279 b
->const_table
= _mesa_hash_table_create(b
, _mesa_hash_pointer
,
3280 _mesa_key_pointer_equal
);
3282 vtn_function_emit(b
, func
, vtn_handle_body_instruction
);
3285 assert(b
->entry_point
->value_type
== vtn_value_type_function
);
3286 nir_function
*entry_point
= b
->entry_point
->func
->impl
->function
;
3287 assert(entry_point
);