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
)) {
107 case GLSL_TYPE_INT64
:
108 case GLSL_TYPE_UINT64
:
110 case GLSL_TYPE_FLOAT
:
111 case GLSL_TYPE_DOUBLE
: {
112 int bit_size
= glsl_get_bit_size(type
);
113 if (glsl_type_is_vector_or_scalar(type
)) {
114 unsigned num_components
= glsl_get_vector_elements(val
->type
);
115 nir_load_const_instr
*load
=
116 nir_load_const_instr_create(b
->shader
, num_components
, bit_size
);
118 load
->value
= constant
->values
[0];
120 nir_instr_insert_before_cf_list(&b
->impl
->body
, &load
->instr
);
121 val
->def
= &load
->def
;
123 assert(glsl_type_is_matrix(type
));
124 unsigned rows
= glsl_get_vector_elements(val
->type
);
125 unsigned columns
= glsl_get_matrix_columns(val
->type
);
126 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, columns
);
128 for (unsigned i
= 0; i
< columns
; i
++) {
129 struct vtn_ssa_value
*col_val
= rzalloc(b
, struct vtn_ssa_value
);
130 col_val
->type
= glsl_get_column_type(val
->type
);
131 nir_load_const_instr
*load
=
132 nir_load_const_instr_create(b
->shader
, rows
, bit_size
);
134 load
->value
= constant
->values
[i
];
136 nir_instr_insert_before_cf_list(&b
->impl
->body
, &load
->instr
);
137 col_val
->def
= &load
->def
;
139 val
->elems
[i
] = col_val
;
145 case GLSL_TYPE_ARRAY
: {
146 unsigned elems
= glsl_get_length(val
->type
);
147 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
148 const struct glsl_type
*elem_type
= glsl_get_array_element(val
->type
);
149 for (unsigned i
= 0; i
< elems
; i
++)
150 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
155 case GLSL_TYPE_STRUCT
: {
156 unsigned elems
= glsl_get_length(val
->type
);
157 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
158 for (unsigned i
= 0; i
< elems
; i
++) {
159 const struct glsl_type
*elem_type
=
160 glsl_get_struct_field(val
->type
, i
);
161 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
168 unreachable("bad constant type");
174 struct vtn_ssa_value
*
175 vtn_ssa_value(struct vtn_builder
*b
, uint32_t value_id
)
177 struct vtn_value
*val
= vtn_untyped_value(b
, value_id
);
178 switch (val
->value_type
) {
179 case vtn_value_type_undef
:
180 return vtn_undef_ssa_value(b
, val
->type
->type
);
182 case vtn_value_type_constant
:
183 return vtn_const_ssa_value(b
, val
->constant
, val
->const_type
);
185 case vtn_value_type_ssa
:
188 case vtn_value_type_pointer
:
189 /* This is needed for function parameters */
190 return vtn_variable_load(b
, val
->pointer
);
193 unreachable("Invalid type for an SSA value");
198 vtn_string_literal(struct vtn_builder
*b
, const uint32_t *words
,
199 unsigned word_count
, unsigned *words_used
)
201 char *dup
= ralloc_strndup(b
, (char *)words
, word_count
* sizeof(*words
));
203 /* Ammount of space taken by the string (including the null) */
204 unsigned len
= strlen(dup
) + 1;
205 *words_used
= DIV_ROUND_UP(len
, sizeof(*words
));
211 vtn_foreach_instruction(struct vtn_builder
*b
, const uint32_t *start
,
212 const uint32_t *end
, vtn_instruction_handler handler
)
218 const uint32_t *w
= start
;
220 SpvOp opcode
= w
[0] & SpvOpCodeMask
;
221 unsigned count
= w
[0] >> SpvWordCountShift
;
222 assert(count
>= 1 && w
+ count
<= end
);
226 break; /* Do nothing */
229 b
->file
= vtn_value(b
, w
[1], vtn_value_type_string
)->str
;
241 if (!handler(b
, opcode
, w
, count
))
253 vtn_handle_extension(struct vtn_builder
*b
, SpvOp opcode
,
254 const uint32_t *w
, unsigned count
)
257 case SpvOpExtInstImport
: {
258 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_extension
);
259 if (strcmp((const char *)&w
[2], "GLSL.std.450") == 0) {
260 val
->ext_handler
= vtn_handle_glsl450_instruction
;
262 assert(!"Unsupported extension");
268 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
269 bool handled
= val
->ext_handler(b
, w
[4], w
, count
);
276 unreachable("Unhandled opcode");
281 _foreach_decoration_helper(struct vtn_builder
*b
,
282 struct vtn_value
*base_value
,
284 struct vtn_value
*value
,
285 vtn_decoration_foreach_cb cb
, void *data
)
287 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
289 if (dec
->scope
== VTN_DEC_DECORATION
) {
290 member
= parent_member
;
291 } else if (dec
->scope
>= VTN_DEC_STRUCT_MEMBER0
) {
292 assert(parent_member
== -1);
293 member
= dec
->scope
- VTN_DEC_STRUCT_MEMBER0
;
295 /* Not a decoration */
300 assert(dec
->group
->value_type
== vtn_value_type_decoration_group
);
301 _foreach_decoration_helper(b
, base_value
, member
, dec
->group
,
304 cb(b
, base_value
, member
, dec
, data
);
309 /** Iterates (recursively if needed) over all of the decorations on a value
311 * This function iterates over all of the decorations applied to a given
312 * value. If it encounters a decoration group, it recurses into the group
313 * and iterates over all of those decorations as well.
316 vtn_foreach_decoration(struct vtn_builder
*b
, struct vtn_value
*value
,
317 vtn_decoration_foreach_cb cb
, void *data
)
319 _foreach_decoration_helper(b
, value
, -1, value
, cb
, data
);
323 vtn_foreach_execution_mode(struct vtn_builder
*b
, struct vtn_value
*value
,
324 vtn_execution_mode_foreach_cb cb
, void *data
)
326 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
327 if (dec
->scope
!= VTN_DEC_EXECUTION_MODE
)
330 assert(dec
->group
== NULL
);
331 cb(b
, value
, dec
, data
);
336 vtn_handle_decoration(struct vtn_builder
*b
, SpvOp opcode
,
337 const uint32_t *w
, unsigned count
)
339 const uint32_t *w_end
= w
+ count
;
340 const uint32_t target
= w
[1];
344 case SpvOpDecorationGroup
:
345 vtn_push_value(b
, target
, vtn_value_type_decoration_group
);
349 case SpvOpMemberDecorate
:
350 case SpvOpExecutionMode
: {
351 struct vtn_value
*val
= &b
->values
[target
];
353 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
356 dec
->scope
= VTN_DEC_DECORATION
;
358 case SpvOpMemberDecorate
:
359 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(w
++);
361 case SpvOpExecutionMode
:
362 dec
->scope
= VTN_DEC_EXECUTION_MODE
;
365 unreachable("Invalid decoration opcode");
367 dec
->decoration
= *(w
++);
370 /* Link into the list */
371 dec
->next
= val
->decoration
;
372 val
->decoration
= dec
;
376 case SpvOpGroupMemberDecorate
:
377 case SpvOpGroupDecorate
: {
378 struct vtn_value
*group
=
379 vtn_value(b
, target
, vtn_value_type_decoration_group
);
381 for (; w
< w_end
; w
++) {
382 struct vtn_value
*val
= vtn_untyped_value(b
, *w
);
383 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
386 if (opcode
== SpvOpGroupDecorate
) {
387 dec
->scope
= VTN_DEC_DECORATION
;
389 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(++w
);
392 /* Link into the list */
393 dec
->next
= val
->decoration
;
394 val
->decoration
= dec
;
400 unreachable("Unhandled opcode");
404 struct member_decoration_ctx
{
406 struct glsl_struct_field
*fields
;
407 struct vtn_type
*type
;
410 /* does a shallow copy of a vtn_type */
412 static struct vtn_type
*
413 vtn_type_copy(struct vtn_builder
*b
, struct vtn_type
*src
)
415 struct vtn_type
*dest
= ralloc(b
, struct vtn_type
);
418 switch (src
->base_type
) {
419 case vtn_base_type_void
:
420 case vtn_base_type_scalar
:
421 case vtn_base_type_vector
:
422 case vtn_base_type_matrix
:
423 case vtn_base_type_array
:
424 case vtn_base_type_image
:
425 case vtn_base_type_sampler
:
426 case vtn_base_type_function
:
427 /* Nothing more to do */
430 case vtn_base_type_struct
:
431 dest
->members
= ralloc_array(b
, struct vtn_type
*, src
->length
);
432 memcpy(dest
->members
, src
->members
,
433 src
->length
* sizeof(src
->members
[0]));
435 dest
->offsets
= ralloc_array(b
, unsigned, src
->length
);
436 memcpy(dest
->offsets
, src
->offsets
,
437 src
->length
* sizeof(src
->offsets
[0]));
444 static struct vtn_type
*
445 mutable_matrix_member(struct vtn_builder
*b
, struct vtn_type
*type
, int member
)
447 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
448 type
= type
->members
[member
];
450 /* We may have an array of matrices.... Oh, joy! */
451 while (glsl_type_is_array(type
->type
)) {
452 type
->array_element
= vtn_type_copy(b
, type
->array_element
);
453 type
= type
->array_element
;
456 assert(glsl_type_is_matrix(type
->type
));
462 struct_member_decoration_cb(struct vtn_builder
*b
,
463 struct vtn_value
*val
, int member
,
464 const struct vtn_decoration
*dec
, void *void_ctx
)
466 struct member_decoration_ctx
*ctx
= void_ctx
;
471 assert(member
< ctx
->num_fields
);
473 switch (dec
->decoration
) {
474 case SpvDecorationNonWritable
:
475 case SpvDecorationNonReadable
:
476 case SpvDecorationRelaxedPrecision
:
477 case SpvDecorationVolatile
:
478 case SpvDecorationCoherent
:
479 case SpvDecorationUniform
:
480 break; /* FIXME: Do nothing with this for now. */
481 case SpvDecorationNoPerspective
:
482 ctx
->fields
[member
].interpolation
= INTERP_MODE_NOPERSPECTIVE
;
484 case SpvDecorationFlat
:
485 ctx
->fields
[member
].interpolation
= INTERP_MODE_FLAT
;
487 case SpvDecorationCentroid
:
488 ctx
->fields
[member
].centroid
= true;
490 case SpvDecorationSample
:
491 ctx
->fields
[member
].sample
= true;
493 case SpvDecorationStream
:
494 /* Vulkan only allows one GS stream */
495 assert(dec
->literals
[0] == 0);
497 case SpvDecorationLocation
:
498 ctx
->fields
[member
].location
= dec
->literals
[0];
500 case SpvDecorationComponent
:
501 break; /* FIXME: What should we do with these? */
502 case SpvDecorationBuiltIn
:
503 ctx
->type
->members
[member
] = vtn_type_copy(b
, ctx
->type
->members
[member
]);
504 ctx
->type
->members
[member
]->is_builtin
= true;
505 ctx
->type
->members
[member
]->builtin
= dec
->literals
[0];
506 ctx
->type
->builtin_block
= true;
508 case SpvDecorationOffset
:
509 ctx
->type
->offsets
[member
] = dec
->literals
[0];
511 case SpvDecorationMatrixStride
:
512 /* Handled as a second pass */
514 case SpvDecorationColMajor
:
515 break; /* Nothing to do here. Column-major is the default. */
516 case SpvDecorationRowMajor
:
517 mutable_matrix_member(b
, ctx
->type
, member
)->row_major
= true;
520 case SpvDecorationPatch
:
523 case SpvDecorationSpecId
:
524 case SpvDecorationBlock
:
525 case SpvDecorationBufferBlock
:
526 case SpvDecorationArrayStride
:
527 case SpvDecorationGLSLShared
:
528 case SpvDecorationGLSLPacked
:
529 case SpvDecorationInvariant
:
530 case SpvDecorationRestrict
:
531 case SpvDecorationAliased
:
532 case SpvDecorationConstant
:
533 case SpvDecorationIndex
:
534 case SpvDecorationBinding
:
535 case SpvDecorationDescriptorSet
:
536 case SpvDecorationLinkageAttributes
:
537 case SpvDecorationNoContraction
:
538 case SpvDecorationInputAttachmentIndex
:
539 vtn_warn("Decoration not allowed on struct members: %s",
540 spirv_decoration_to_string(dec
->decoration
));
543 case SpvDecorationXfbBuffer
:
544 case SpvDecorationXfbStride
:
545 vtn_warn("Vulkan does not have transform feedback");
548 case SpvDecorationCPacked
:
549 case SpvDecorationSaturatedConversion
:
550 case SpvDecorationFuncParamAttr
:
551 case SpvDecorationFPRoundingMode
:
552 case SpvDecorationFPFastMathMode
:
553 case SpvDecorationAlignment
:
554 vtn_warn("Decoration only allowed for CL-style kernels: %s",
555 spirv_decoration_to_string(dec
->decoration
));
559 unreachable("Unhandled decoration");
563 /* Matrix strides are handled as a separate pass because we need to know
564 * whether the matrix is row-major or not first.
567 struct_member_matrix_stride_cb(struct vtn_builder
*b
,
568 struct vtn_value
*val
, int member
,
569 const struct vtn_decoration
*dec
,
572 if (dec
->decoration
!= SpvDecorationMatrixStride
)
576 struct member_decoration_ctx
*ctx
= void_ctx
;
578 struct vtn_type
*mat_type
= mutable_matrix_member(b
, ctx
->type
, member
);
579 if (mat_type
->row_major
) {
580 mat_type
->array_element
= vtn_type_copy(b
, mat_type
->array_element
);
581 mat_type
->stride
= mat_type
->array_element
->stride
;
582 mat_type
->array_element
->stride
= dec
->literals
[0];
584 assert(mat_type
->array_element
->stride
> 0);
585 mat_type
->stride
= dec
->literals
[0];
590 type_decoration_cb(struct vtn_builder
*b
,
591 struct vtn_value
*val
, int member
,
592 const struct vtn_decoration
*dec
, void *ctx
)
594 struct vtn_type
*type
= val
->type
;
599 switch (dec
->decoration
) {
600 case SpvDecorationArrayStride
:
601 type
->stride
= dec
->literals
[0];
603 case SpvDecorationBlock
:
606 case SpvDecorationBufferBlock
:
607 type
->buffer_block
= true;
609 case SpvDecorationGLSLShared
:
610 case SpvDecorationGLSLPacked
:
611 /* Ignore these, since we get explicit offsets anyways */
614 case SpvDecorationRowMajor
:
615 case SpvDecorationColMajor
:
616 case SpvDecorationMatrixStride
:
617 case SpvDecorationBuiltIn
:
618 case SpvDecorationNoPerspective
:
619 case SpvDecorationFlat
:
620 case SpvDecorationPatch
:
621 case SpvDecorationCentroid
:
622 case SpvDecorationSample
:
623 case SpvDecorationVolatile
:
624 case SpvDecorationCoherent
:
625 case SpvDecorationNonWritable
:
626 case SpvDecorationNonReadable
:
627 case SpvDecorationUniform
:
628 case SpvDecorationStream
:
629 case SpvDecorationLocation
:
630 case SpvDecorationComponent
:
631 case SpvDecorationOffset
:
632 case SpvDecorationXfbBuffer
:
633 case SpvDecorationXfbStride
:
634 vtn_warn("Decoration only allowed for struct members: %s",
635 spirv_decoration_to_string(dec
->decoration
));
638 case SpvDecorationRelaxedPrecision
:
639 case SpvDecorationSpecId
:
640 case SpvDecorationInvariant
:
641 case SpvDecorationRestrict
:
642 case SpvDecorationAliased
:
643 case SpvDecorationConstant
:
644 case SpvDecorationIndex
:
645 case SpvDecorationBinding
:
646 case SpvDecorationDescriptorSet
:
647 case SpvDecorationLinkageAttributes
:
648 case SpvDecorationNoContraction
:
649 case SpvDecorationInputAttachmentIndex
:
650 vtn_warn("Decoration not allowed on types: %s",
651 spirv_decoration_to_string(dec
->decoration
));
654 case SpvDecorationCPacked
:
655 case SpvDecorationSaturatedConversion
:
656 case SpvDecorationFuncParamAttr
:
657 case SpvDecorationFPRoundingMode
:
658 case SpvDecorationFPFastMathMode
:
659 case SpvDecorationAlignment
:
660 vtn_warn("Decoration only allowed for CL-style kernels: %s",
661 spirv_decoration_to_string(dec
->decoration
));
665 unreachable("Unhandled decoration");
670 translate_image_format(SpvImageFormat format
)
673 case SpvImageFormatUnknown
: return 0; /* GL_NONE */
674 case SpvImageFormatRgba32f
: return 0x8814; /* GL_RGBA32F */
675 case SpvImageFormatRgba16f
: return 0x881A; /* GL_RGBA16F */
676 case SpvImageFormatR32f
: return 0x822E; /* GL_R32F */
677 case SpvImageFormatRgba8
: return 0x8058; /* GL_RGBA8 */
678 case SpvImageFormatRgba8Snorm
: return 0x8F97; /* GL_RGBA8_SNORM */
679 case SpvImageFormatRg32f
: return 0x8230; /* GL_RG32F */
680 case SpvImageFormatRg16f
: return 0x822F; /* GL_RG16F */
681 case SpvImageFormatR11fG11fB10f
: return 0x8C3A; /* GL_R11F_G11F_B10F */
682 case SpvImageFormatR16f
: return 0x822D; /* GL_R16F */
683 case SpvImageFormatRgba16
: return 0x805B; /* GL_RGBA16 */
684 case SpvImageFormatRgb10A2
: return 0x8059; /* GL_RGB10_A2 */
685 case SpvImageFormatRg16
: return 0x822C; /* GL_RG16 */
686 case SpvImageFormatRg8
: return 0x822B; /* GL_RG8 */
687 case SpvImageFormatR16
: return 0x822A; /* GL_R16 */
688 case SpvImageFormatR8
: return 0x8229; /* GL_R8 */
689 case SpvImageFormatRgba16Snorm
: return 0x8F9B; /* GL_RGBA16_SNORM */
690 case SpvImageFormatRg16Snorm
: return 0x8F99; /* GL_RG16_SNORM */
691 case SpvImageFormatRg8Snorm
: return 0x8F95; /* GL_RG8_SNORM */
692 case SpvImageFormatR16Snorm
: return 0x8F98; /* GL_R16_SNORM */
693 case SpvImageFormatR8Snorm
: return 0x8F94; /* GL_R8_SNORM */
694 case SpvImageFormatRgba32i
: return 0x8D82; /* GL_RGBA32I */
695 case SpvImageFormatRgba16i
: return 0x8D88; /* GL_RGBA16I */
696 case SpvImageFormatRgba8i
: return 0x8D8E; /* GL_RGBA8I */
697 case SpvImageFormatR32i
: return 0x8235; /* GL_R32I */
698 case SpvImageFormatRg32i
: return 0x823B; /* GL_RG32I */
699 case SpvImageFormatRg16i
: return 0x8239; /* GL_RG16I */
700 case SpvImageFormatRg8i
: return 0x8237; /* GL_RG8I */
701 case SpvImageFormatR16i
: return 0x8233; /* GL_R16I */
702 case SpvImageFormatR8i
: return 0x8231; /* GL_R8I */
703 case SpvImageFormatRgba32ui
: return 0x8D70; /* GL_RGBA32UI */
704 case SpvImageFormatRgba16ui
: return 0x8D76; /* GL_RGBA16UI */
705 case SpvImageFormatRgba8ui
: return 0x8D7C; /* GL_RGBA8UI */
706 case SpvImageFormatR32ui
: return 0x8236; /* GL_R32UI */
707 case SpvImageFormatRgb10a2ui
: return 0x906F; /* GL_RGB10_A2UI */
708 case SpvImageFormatRg32ui
: return 0x823C; /* GL_RG32UI */
709 case SpvImageFormatRg16ui
: return 0x823A; /* GL_RG16UI */
710 case SpvImageFormatRg8ui
: return 0x8238; /* GL_RG8UI */
711 case SpvImageFormatR16ui
: return 0x823A; /* GL_RG16UI */
712 case SpvImageFormatR8ui
: return 0x8232; /* GL_R8UI */
714 assert(!"Invalid image format");
720 vtn_handle_type(struct vtn_builder
*b
, SpvOp opcode
,
721 const uint32_t *w
, unsigned count
)
723 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_type
);
725 val
->type
= rzalloc(b
, struct vtn_type
);
726 val
->type
->val
= val
;
730 val
->type
->base_type
= vtn_base_type_void
;
731 val
->type
->type
= glsl_void_type();
734 val
->type
->base_type
= vtn_base_type_scalar
;
735 val
->type
->type
= glsl_bool_type();
739 const bool signedness
= w
[3];
740 val
->type
->base_type
= vtn_base_type_scalar
;
742 val
->type
->type
= (signedness
? glsl_int64_t_type() : glsl_uint64_t_type());
744 val
->type
->type
= (signedness
? glsl_int_type() : glsl_uint_type());
747 case SpvOpTypeFloat
: {
749 val
->type
->base_type
= vtn_base_type_scalar
;
750 val
->type
->type
= bit_size
== 64 ? glsl_double_type() : glsl_float_type();
754 case SpvOpTypeVector
: {
755 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
756 unsigned elems
= w
[3];
758 assert(glsl_type_is_scalar(base
->type
));
759 val
->type
->base_type
= vtn_base_type_vector
;
760 val
->type
->type
= glsl_vector_type(glsl_get_base_type(base
->type
), elems
);
762 /* Vectors implicitly have sizeof(base_type) stride. For now, this
763 * is always 4 bytes. This will have to change if we want to start
764 * supporting doubles or half-floats.
766 val
->type
->stride
= glsl_get_bit_size(base
->type
) / 8;
767 val
->type
->array_element
= base
;
771 case SpvOpTypeMatrix
: {
772 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
773 unsigned columns
= w
[3];
775 assert(glsl_type_is_vector(base
->type
));
776 val
->type
->base_type
= vtn_base_type_matrix
;
777 val
->type
->type
= glsl_matrix_type(glsl_get_base_type(base
->type
),
778 glsl_get_vector_elements(base
->type
),
780 assert(!glsl_type_is_error(val
->type
->type
));
781 val
->type
->length
= columns
;
782 val
->type
->array_element
= base
;
783 val
->type
->row_major
= false;
784 val
->type
->stride
= 0;
788 case SpvOpTypeRuntimeArray
:
789 case SpvOpTypeArray
: {
790 struct vtn_type
*array_element
=
791 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
793 if (opcode
== SpvOpTypeRuntimeArray
) {
794 /* A length of 0 is used to denote unsized arrays */
795 val
->type
->length
= 0;
798 vtn_value(b
, w
[3], vtn_value_type_constant
)->constant
->values
[0].u32
[0];
801 val
->type
->base_type
= vtn_base_type_array
;
802 val
->type
->type
= glsl_array_type(array_element
->type
, val
->type
->length
);
803 val
->type
->array_element
= array_element
;
804 val
->type
->stride
= 0;
808 case SpvOpTypeStruct
: {
809 unsigned num_fields
= count
- 2;
810 val
->type
->base_type
= vtn_base_type_struct
;
811 val
->type
->length
= num_fields
;
812 val
->type
->members
= ralloc_array(b
, struct vtn_type
*, num_fields
);
813 val
->type
->offsets
= ralloc_array(b
, unsigned, num_fields
);
815 NIR_VLA(struct glsl_struct_field
, fields
, count
);
816 for (unsigned i
= 0; i
< num_fields
; i
++) {
817 val
->type
->members
[i
] =
818 vtn_value(b
, w
[i
+ 2], vtn_value_type_type
)->type
;
819 fields
[i
] = (struct glsl_struct_field
) {
820 .type
= val
->type
->members
[i
]->type
,
821 .name
= ralloc_asprintf(b
, "field%d", i
),
826 struct member_decoration_ctx ctx
= {
827 .num_fields
= num_fields
,
832 vtn_foreach_decoration(b
, val
, struct_member_decoration_cb
, &ctx
);
833 vtn_foreach_decoration(b
, val
, struct_member_matrix_stride_cb
, &ctx
);
835 const char *name
= val
->name
? val
->name
: "struct";
837 val
->type
->type
= glsl_struct_type(fields
, num_fields
, name
);
841 case SpvOpTypeFunction
: {
842 val
->type
->base_type
= vtn_base_type_function
;
844 const struct glsl_type
*return_type
=
845 vtn_value(b
, w
[2], vtn_value_type_type
)->type
->type
;
846 NIR_VLA(struct glsl_function_param
, params
, count
- 3);
847 for (unsigned i
= 0; i
< count
- 3; i
++) {
848 params
[i
].type
= vtn_value(b
, w
[i
+ 3], vtn_value_type_type
)->type
->type
;
852 params
[i
].out
= true;
854 val
->type
->type
= glsl_function_type(return_type
, params
, count
- 3);
858 case SpvOpTypePointer
:
859 /* FIXME: For now, we'll just do the really lame thing and return
860 * the same type. The validator should ensure that the proper number
861 * of dereferences happen
863 val
->type
= vtn_value(b
, w
[3], vtn_value_type_type
)->type
;
866 case SpvOpTypeImage
: {
867 val
->type
->base_type
= vtn_base_type_image
;
869 const struct glsl_type
*sampled_type
=
870 vtn_value(b
, w
[2], vtn_value_type_type
)->type
->type
;
872 assert(glsl_type_is_vector_or_scalar(sampled_type
));
874 enum glsl_sampler_dim dim
;
875 switch ((SpvDim
)w
[3]) {
876 case SpvDim1D
: dim
= GLSL_SAMPLER_DIM_1D
; break;
877 case SpvDim2D
: dim
= GLSL_SAMPLER_DIM_2D
; break;
878 case SpvDim3D
: dim
= GLSL_SAMPLER_DIM_3D
; break;
879 case SpvDimCube
: dim
= GLSL_SAMPLER_DIM_CUBE
; break;
880 case SpvDimRect
: dim
= GLSL_SAMPLER_DIM_RECT
; break;
881 case SpvDimBuffer
: dim
= GLSL_SAMPLER_DIM_BUF
; break;
882 case SpvDimSubpassData
: dim
= GLSL_SAMPLER_DIM_SUBPASS
; break;
884 unreachable("Invalid SPIR-V Sampler dimension");
887 bool is_shadow
= w
[4];
888 bool is_array
= w
[5];
889 bool multisampled
= w
[6];
890 unsigned sampled
= w
[7];
891 SpvImageFormat format
= w
[8];
894 val
->type
->access_qualifier
= w
[9];
896 val
->type
->access_qualifier
= SpvAccessQualifierReadWrite
;
899 if (dim
== GLSL_SAMPLER_DIM_2D
)
900 dim
= GLSL_SAMPLER_DIM_MS
;
901 else if (dim
== GLSL_SAMPLER_DIM_SUBPASS
)
902 dim
= GLSL_SAMPLER_DIM_SUBPASS_MS
;
904 assert(!"Unsupported multisampled image type");
907 val
->type
->image_format
= translate_image_format(format
);
910 val
->type
->sampled
= true;
911 val
->type
->type
= glsl_sampler_type(dim
, is_shadow
, is_array
,
912 glsl_get_base_type(sampled_type
));
913 } else if (sampled
== 2) {
915 val
->type
->sampled
= false;
916 val
->type
->type
= glsl_image_type(dim
, is_array
,
917 glsl_get_base_type(sampled_type
));
919 assert(!"We need to know if the image will be sampled");
924 case SpvOpTypeSampledImage
:
925 val
->type
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
928 case SpvOpTypeSampler
:
929 /* The actual sampler type here doesn't really matter. It gets
930 * thrown away the moment you combine it with an image. What really
931 * matters is that it's a sampler type as opposed to an integer type
932 * so the backend knows what to do.
934 val
->type
->base_type
= vtn_base_type_sampler
;
935 val
->type
->type
= glsl_bare_sampler_type();
938 case SpvOpTypeOpaque
:
940 case SpvOpTypeDeviceEvent
:
941 case SpvOpTypeReserveId
:
945 unreachable("Unhandled opcode");
948 vtn_foreach_decoration(b
, val
, type_decoration_cb
, NULL
);
951 static nir_constant
*
952 vtn_null_constant(struct vtn_builder
*b
, const struct glsl_type
*type
)
954 nir_constant
*c
= rzalloc(b
, nir_constant
);
956 switch (glsl_get_base_type(type
)) {
959 case GLSL_TYPE_INT64
:
960 case GLSL_TYPE_UINT64
:
962 case GLSL_TYPE_FLOAT
:
963 case GLSL_TYPE_DOUBLE
:
964 /* Nothing to do here. It's already initialized to zero */
967 case GLSL_TYPE_ARRAY
:
968 assert(glsl_get_length(type
) > 0);
969 c
->num_elements
= glsl_get_length(type
);
970 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
972 c
->elements
[0] = vtn_null_constant(b
, glsl_get_array_element(type
));
973 for (unsigned i
= 1; i
< c
->num_elements
; i
++)
974 c
->elements
[i
] = c
->elements
[0];
977 case GLSL_TYPE_STRUCT
:
978 c
->num_elements
= glsl_get_length(type
);
979 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
981 for (unsigned i
= 0; i
< c
->num_elements
; i
++) {
982 c
->elements
[i
] = vtn_null_constant(b
, glsl_get_struct_field(type
, i
));
987 unreachable("Invalid type for null constant");
994 spec_constant_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*v
,
995 int member
, const struct vtn_decoration
*dec
,
998 assert(member
== -1);
999 if (dec
->decoration
!= SpvDecorationSpecId
)
1002 struct spec_constant_value
*const_value
= data
;
1004 for (unsigned i
= 0; i
< b
->num_specializations
; i
++) {
1005 if (b
->specializations
[i
].id
== dec
->literals
[0]) {
1006 if (const_value
->is_double
)
1007 const_value
->data64
= b
->specializations
[i
].data64
;
1009 const_value
->data32
= b
->specializations
[i
].data32
;
1016 get_specialization(struct vtn_builder
*b
, struct vtn_value
*val
,
1017 uint32_t const_value
)
1019 struct spec_constant_value data
;
1020 data
.is_double
= false;
1021 data
.data32
= const_value
;
1022 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1027 get_specialization64(struct vtn_builder
*b
, struct vtn_value
*val
,
1028 uint64_t const_value
)
1030 struct spec_constant_value data
;
1031 data
.is_double
= true;
1032 data
.data64
= const_value
;
1033 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1038 handle_workgroup_size_decoration_cb(struct vtn_builder
*b
,
1039 struct vtn_value
*val
,
1041 const struct vtn_decoration
*dec
,
1044 assert(member
== -1);
1045 if (dec
->decoration
!= SpvDecorationBuiltIn
||
1046 dec
->literals
[0] != SpvBuiltInWorkgroupSize
)
1049 assert(val
->const_type
== glsl_vector_type(GLSL_TYPE_UINT
, 3));
1051 b
->shader
->info
.cs
.local_size
[0] = val
->constant
->values
[0].u32
[0];
1052 b
->shader
->info
.cs
.local_size
[1] = val
->constant
->values
[0].u32
[1];
1053 b
->shader
->info
.cs
.local_size
[2] = val
->constant
->values
[0].u32
[2];
1057 vtn_handle_constant(struct vtn_builder
*b
, SpvOp opcode
,
1058 const uint32_t *w
, unsigned count
)
1060 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_constant
);
1061 val
->const_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
1062 val
->constant
= rzalloc(b
, nir_constant
);
1064 case SpvOpConstantTrue
:
1065 assert(val
->const_type
== glsl_bool_type());
1066 val
->constant
->values
[0].u32
[0] = NIR_TRUE
;
1068 case SpvOpConstantFalse
:
1069 assert(val
->const_type
== glsl_bool_type());
1070 val
->constant
->values
[0].u32
[0] = NIR_FALSE
;
1073 case SpvOpSpecConstantTrue
:
1074 case SpvOpSpecConstantFalse
: {
1075 assert(val
->const_type
== glsl_bool_type());
1077 get_specialization(b
, val
, (opcode
== SpvOpSpecConstantTrue
));
1078 val
->constant
->values
[0].u32
[0] = int_val
? NIR_TRUE
: NIR_FALSE
;
1082 case SpvOpConstant
: {
1083 assert(glsl_type_is_scalar(val
->const_type
));
1084 int bit_size
= glsl_get_bit_size(val
->const_type
);
1085 if (bit_size
== 64) {
1086 val
->constant
->values
->u32
[0] = w
[3];
1087 val
->constant
->values
->u32
[1] = w
[4];
1089 assert(bit_size
== 32);
1090 val
->constant
->values
->u32
[0] = w
[3];
1094 case SpvOpSpecConstant
: {
1095 assert(glsl_type_is_scalar(val
->const_type
));
1096 val
->constant
->values
[0].u32
[0] = get_specialization(b
, val
, w
[3]);
1097 int bit_size
= glsl_get_bit_size(val
->const_type
);
1099 val
->constant
->values
[0].u64
[0] =
1100 get_specialization64(b
, val
, vtn_u64_literal(&w
[3]));
1102 val
->constant
->values
[0].u32
[0] = get_specialization(b
, val
, w
[3]);
1105 case SpvOpSpecConstantComposite
:
1106 case SpvOpConstantComposite
: {
1107 unsigned elem_count
= count
- 3;
1108 nir_constant
**elems
= ralloc_array(b
, nir_constant
*, elem_count
);
1109 for (unsigned i
= 0; i
< elem_count
; i
++)
1110 elems
[i
] = vtn_value(b
, w
[i
+ 3], vtn_value_type_constant
)->constant
;
1112 switch (glsl_get_base_type(val
->const_type
)) {
1113 case GLSL_TYPE_UINT
:
1115 case GLSL_TYPE_UINT64
:
1116 case GLSL_TYPE_INT64
:
1117 case GLSL_TYPE_FLOAT
:
1118 case GLSL_TYPE_BOOL
:
1119 case GLSL_TYPE_DOUBLE
: {
1120 int bit_size
= glsl_get_bit_size(val
->const_type
);
1121 if (glsl_type_is_matrix(val
->const_type
)) {
1122 assert(glsl_get_matrix_columns(val
->const_type
) == elem_count
);
1123 for (unsigned i
= 0; i
< elem_count
; i
++)
1124 val
->constant
->values
[i
] = elems
[i
]->values
[0];
1126 assert(glsl_type_is_vector(val
->const_type
));
1127 assert(glsl_get_vector_elements(val
->const_type
) == elem_count
);
1128 for (unsigned i
= 0; i
< elem_count
; i
++) {
1129 if (bit_size
== 64) {
1130 val
->constant
->values
[0].u64
[i
] = elems
[i
]->values
[0].u64
[0];
1132 assert(bit_size
== 32);
1133 val
->constant
->values
[0].u32
[i
] = elems
[i
]->values
[0].u32
[0];
1140 case GLSL_TYPE_STRUCT
:
1141 case GLSL_TYPE_ARRAY
:
1142 ralloc_steal(val
->constant
, elems
);
1143 val
->constant
->num_elements
= elem_count
;
1144 val
->constant
->elements
= elems
;
1148 unreachable("Unsupported type for constants");
1153 case SpvOpSpecConstantOp
: {
1154 SpvOp opcode
= get_specialization(b
, val
, w
[3]);
1156 case SpvOpVectorShuffle
: {
1157 struct vtn_value
*v0
= &b
->values
[w
[4]];
1158 struct vtn_value
*v1
= &b
->values
[w
[5]];
1160 assert(v0
->value_type
== vtn_value_type_constant
||
1161 v0
->value_type
== vtn_value_type_undef
);
1162 assert(v1
->value_type
== vtn_value_type_constant
||
1163 v1
->value_type
== vtn_value_type_undef
);
1165 unsigned len0
= v0
->value_type
== vtn_value_type_constant
?
1166 glsl_get_vector_elements(v0
->const_type
) :
1167 glsl_get_vector_elements(v0
->type
->type
);
1168 unsigned len1
= v1
->value_type
== vtn_value_type_constant
?
1169 glsl_get_vector_elements(v1
->const_type
) :
1170 glsl_get_vector_elements(v1
->type
->type
);
1172 assert(len0
+ len1
< 16);
1174 unsigned bit_size
= glsl_get_bit_size(val
->const_type
);
1175 unsigned bit_size0
= v0
->value_type
== vtn_value_type_constant
?
1176 glsl_get_bit_size(v0
->const_type
) :
1177 glsl_get_bit_size(v0
->type
->type
);
1178 unsigned bit_size1
= v1
->value_type
== vtn_value_type_constant
?
1179 glsl_get_bit_size(v1
->const_type
) :
1180 glsl_get_bit_size(v1
->type
->type
);
1182 assert(bit_size
== bit_size0
&& bit_size
== bit_size1
);
1183 (void)bit_size0
; (void)bit_size1
;
1185 if (bit_size
== 64) {
1187 if (v0
->value_type
== vtn_value_type_constant
) {
1188 for (unsigned i
= 0; i
< len0
; i
++)
1189 u64
[i
] = v0
->constant
->values
[0].u64
[i
];
1191 if (v1
->value_type
== vtn_value_type_constant
) {
1192 for (unsigned i
= 0; i
< len1
; i
++)
1193 u64
[len0
+ i
] = v1
->constant
->values
[0].u64
[i
];
1196 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1197 uint32_t comp
= w
[i
+ 6];
1198 /* If component is not used, set the value to a known constant
1199 * to detect if it is wrongly used.
1201 if (comp
== (uint32_t)-1)
1202 val
->constant
->values
[0].u64
[j
] = 0xdeadbeefdeadbeef;
1204 val
->constant
->values
[0].u64
[j
] = u64
[comp
];
1208 if (v0
->value_type
== vtn_value_type_constant
) {
1209 for (unsigned i
= 0; i
< len0
; i
++)
1210 u32
[i
] = v0
->constant
->values
[0].u32
[i
];
1212 if (v1
->value_type
== vtn_value_type_constant
) {
1213 for (unsigned i
= 0; i
< len1
; i
++)
1214 u32
[len0
+ i
] = v1
->constant
->values
[0].u32
[i
];
1217 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1218 uint32_t comp
= w
[i
+ 6];
1219 /* If component is not used, set the value to a known constant
1220 * to detect if it is wrongly used.
1222 if (comp
== (uint32_t)-1)
1223 val
->constant
->values
[0].u32
[j
] = 0xdeadbeef;
1225 val
->constant
->values
[0].u32
[j
] = u32
[comp
];
1231 case SpvOpCompositeExtract
:
1232 case SpvOpCompositeInsert
: {
1233 struct vtn_value
*comp
;
1234 unsigned deref_start
;
1235 struct nir_constant
**c
;
1236 if (opcode
== SpvOpCompositeExtract
) {
1237 comp
= vtn_value(b
, w
[4], vtn_value_type_constant
);
1239 c
= &comp
->constant
;
1241 comp
= vtn_value(b
, w
[5], vtn_value_type_constant
);
1243 val
->constant
= nir_constant_clone(comp
->constant
,
1250 const struct glsl_type
*type
= comp
->const_type
;
1251 for (unsigned i
= deref_start
; i
< count
; i
++) {
1252 switch (glsl_get_base_type(type
)) {
1253 case GLSL_TYPE_UINT
:
1255 case GLSL_TYPE_UINT64
:
1256 case GLSL_TYPE_INT64
:
1257 case GLSL_TYPE_FLOAT
:
1258 case GLSL_TYPE_DOUBLE
:
1259 case GLSL_TYPE_BOOL
:
1260 /* If we hit this granularity, we're picking off an element */
1261 if (glsl_type_is_matrix(type
)) {
1262 assert(col
== 0 && elem
== -1);
1265 type
= glsl_get_column_type(type
);
1267 assert(elem
<= 0 && glsl_type_is_vector(type
));
1269 type
= glsl_scalar_type(glsl_get_base_type(type
));
1273 case GLSL_TYPE_ARRAY
:
1274 c
= &(*c
)->elements
[w
[i
]];
1275 type
= glsl_get_array_element(type
);
1278 case GLSL_TYPE_STRUCT
:
1279 c
= &(*c
)->elements
[w
[i
]];
1280 type
= glsl_get_struct_field(type
, w
[i
]);
1284 unreachable("Invalid constant type");
1288 if (opcode
== SpvOpCompositeExtract
) {
1292 unsigned num_components
= glsl_get_vector_elements(type
);
1293 unsigned bit_size
= glsl_get_bit_size(type
);
1294 for (unsigned i
= 0; i
< num_components
; i
++)
1295 if (bit_size
== 64) {
1296 val
->constant
->values
[0].u64
[i
] = (*c
)->values
[col
].u64
[elem
+ i
];
1298 assert(bit_size
== 32);
1299 val
->constant
->values
[0].u32
[i
] = (*c
)->values
[col
].u32
[elem
+ i
];
1303 struct vtn_value
*insert
=
1304 vtn_value(b
, w
[4], vtn_value_type_constant
);
1305 assert(insert
->const_type
== type
);
1307 *c
= insert
->constant
;
1309 unsigned num_components
= glsl_get_vector_elements(type
);
1310 unsigned bit_size
= glsl_get_bit_size(type
);
1311 for (unsigned i
= 0; i
< num_components
; i
++)
1312 if (bit_size
== 64) {
1313 (*c
)->values
[col
].u64
[elem
+ i
] = insert
->constant
->values
[0].u64
[i
];
1315 assert(bit_size
== 32);
1316 (*c
)->values
[col
].u32
[elem
+ i
] = insert
->constant
->values
[0].u32
[i
];
1325 nir_alu_type dst_alu_type
= nir_get_nir_type_for_glsl_type(val
->const_type
);
1326 nir_alu_type src_alu_type
= dst_alu_type
;
1327 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(opcode
, &swap
, src_alu_type
, dst_alu_type
);
1329 unsigned num_components
= glsl_get_vector_elements(val
->const_type
);
1331 glsl_get_bit_size(val
->const_type
);
1333 nir_const_value src
[4];
1335 for (unsigned i
= 0; i
< count
- 4; i
++) {
1337 vtn_value(b
, w
[4 + i
], vtn_value_type_constant
)->constant
;
1339 unsigned j
= swap
? 1 - i
: i
;
1340 assert(bit_size
== 32);
1341 src
[j
] = c
->values
[0];
1344 val
->constant
->values
[0] =
1345 nir_eval_const_opcode(op
, num_components
, bit_size
, src
);
1352 case SpvOpConstantNull
:
1353 val
->constant
= vtn_null_constant(b
, val
->const_type
);
1356 case SpvOpConstantSampler
:
1357 assert(!"OpConstantSampler requires Kernel Capability");
1361 unreachable("Unhandled opcode");
1364 /* Now that we have the value, update the workgroup size if needed */
1365 vtn_foreach_decoration(b
, val
, handle_workgroup_size_decoration_cb
, NULL
);
1369 vtn_handle_function_call(struct vtn_builder
*b
, SpvOp opcode
,
1370 const uint32_t *w
, unsigned count
)
1372 struct nir_function
*callee
=
1373 vtn_value(b
, w
[3], vtn_value_type_function
)->func
->impl
->function
;
1375 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, callee
);
1376 for (unsigned i
= 0; i
< call
->num_params
; i
++) {
1377 unsigned arg_id
= w
[4 + i
];
1378 struct vtn_value
*arg
= vtn_untyped_value(b
, arg_id
);
1379 if (arg
->value_type
== vtn_value_type_pointer
) {
1380 nir_deref_var
*d
= vtn_pointer_to_deref(b
, arg
->pointer
);
1381 call
->params
[i
] = nir_deref_var_clone(d
, call
);
1383 struct vtn_ssa_value
*arg_ssa
= vtn_ssa_value(b
, arg_id
);
1385 /* Make a temporary to store the argument in */
1387 nir_local_variable_create(b
->impl
, arg_ssa
->type
, "arg_tmp");
1388 call
->params
[i
] = nir_deref_var_create(call
, tmp
);
1390 vtn_local_store(b
, arg_ssa
, call
->params
[i
]);
1394 nir_variable
*out_tmp
= NULL
;
1395 if (!glsl_type_is_void(callee
->return_type
)) {
1396 out_tmp
= nir_local_variable_create(b
->impl
, callee
->return_type
,
1398 call
->return_deref
= nir_deref_var_create(call
, out_tmp
);
1401 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
1403 if (glsl_type_is_void(callee
->return_type
)) {
1404 vtn_push_value(b
, w
[2], vtn_value_type_undef
);
1406 struct vtn_value
*retval
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1407 retval
->ssa
= vtn_local_load(b
, call
->return_deref
);
1411 struct vtn_ssa_value
*
1412 vtn_create_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
1414 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
1417 if (!glsl_type_is_vector_or_scalar(type
)) {
1418 unsigned elems
= glsl_get_length(type
);
1419 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
1420 for (unsigned i
= 0; i
< elems
; i
++) {
1421 const struct glsl_type
*child_type
;
1423 switch (glsl_get_base_type(type
)) {
1425 case GLSL_TYPE_UINT
:
1426 case GLSL_TYPE_INT64
:
1427 case GLSL_TYPE_UINT64
:
1428 case GLSL_TYPE_BOOL
:
1429 case GLSL_TYPE_FLOAT
:
1430 case GLSL_TYPE_DOUBLE
:
1431 child_type
= glsl_get_column_type(type
);
1433 case GLSL_TYPE_ARRAY
:
1434 child_type
= glsl_get_array_element(type
);
1436 case GLSL_TYPE_STRUCT
:
1437 child_type
= glsl_get_struct_field(type
, i
);
1440 unreachable("unkown base type");
1443 val
->elems
[i
] = vtn_create_ssa_value(b
, child_type
);
1451 vtn_tex_src(struct vtn_builder
*b
, unsigned index
, nir_tex_src_type type
)
1454 src
.src
= nir_src_for_ssa(vtn_ssa_value(b
, index
)->def
);
1455 src
.src_type
= type
;
1460 vtn_handle_texture(struct vtn_builder
*b
, SpvOp opcode
,
1461 const uint32_t *w
, unsigned count
)
1463 if (opcode
== SpvOpSampledImage
) {
1464 struct vtn_value
*val
=
1465 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
1466 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
1467 val
->sampled_image
->image
=
1468 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
1469 val
->sampled_image
->sampler
=
1470 vtn_value(b
, w
[4], vtn_value_type_pointer
)->pointer
;
1472 } else if (opcode
== SpvOpImage
) {
1473 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
1474 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[3]);
1475 if (src_val
->value_type
== vtn_value_type_sampled_image
) {
1476 val
->pointer
= src_val
->sampled_image
->image
;
1478 assert(src_val
->value_type
== vtn_value_type_pointer
);
1479 val
->pointer
= src_val
->pointer
;
1484 struct vtn_type
*ret_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1485 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1487 struct vtn_sampled_image sampled
;
1488 struct vtn_value
*sampled_val
= vtn_untyped_value(b
, w
[3]);
1489 if (sampled_val
->value_type
== vtn_value_type_sampled_image
) {
1490 sampled
= *sampled_val
->sampled_image
;
1492 assert(sampled_val
->value_type
== vtn_value_type_pointer
);
1493 sampled
.image
= NULL
;
1494 sampled
.sampler
= sampled_val
->pointer
;
1497 const struct glsl_type
*image_type
;
1498 if (sampled
.image
) {
1499 image_type
= sampled
.image
->var
->var
->interface_type
;
1501 image_type
= sampled
.sampler
->var
->var
->interface_type
;
1503 const enum glsl_sampler_dim sampler_dim
= glsl_get_sampler_dim(image_type
);
1504 const bool is_array
= glsl_sampler_type_is_array(image_type
);
1505 const bool is_shadow
= glsl_sampler_type_is_shadow(image_type
);
1507 /* Figure out the base texture operation */
1510 case SpvOpImageSampleImplicitLod
:
1511 case SpvOpImageSampleDrefImplicitLod
:
1512 case SpvOpImageSampleProjImplicitLod
:
1513 case SpvOpImageSampleProjDrefImplicitLod
:
1514 texop
= nir_texop_tex
;
1517 case SpvOpImageSampleExplicitLod
:
1518 case SpvOpImageSampleDrefExplicitLod
:
1519 case SpvOpImageSampleProjExplicitLod
:
1520 case SpvOpImageSampleProjDrefExplicitLod
:
1521 texop
= nir_texop_txl
;
1524 case SpvOpImageFetch
:
1525 if (glsl_get_sampler_dim(image_type
) == GLSL_SAMPLER_DIM_MS
) {
1526 texop
= nir_texop_txf_ms
;
1528 texop
= nir_texop_txf
;
1532 case SpvOpImageGather
:
1533 case SpvOpImageDrefGather
:
1534 texop
= nir_texop_tg4
;
1537 case SpvOpImageQuerySizeLod
:
1538 case SpvOpImageQuerySize
:
1539 texop
= nir_texop_txs
;
1542 case SpvOpImageQueryLod
:
1543 texop
= nir_texop_lod
;
1546 case SpvOpImageQueryLevels
:
1547 texop
= nir_texop_query_levels
;
1550 case SpvOpImageQuerySamples
:
1551 texop
= nir_texop_texture_samples
;
1555 unreachable("Unhandled opcode");
1558 nir_tex_src srcs
[8]; /* 8 should be enough */
1559 nir_tex_src
*p
= srcs
;
1563 struct nir_ssa_def
*coord
;
1564 unsigned coord_components
;
1566 case SpvOpImageSampleImplicitLod
:
1567 case SpvOpImageSampleExplicitLod
:
1568 case SpvOpImageSampleDrefImplicitLod
:
1569 case SpvOpImageSampleDrefExplicitLod
:
1570 case SpvOpImageSampleProjImplicitLod
:
1571 case SpvOpImageSampleProjExplicitLod
:
1572 case SpvOpImageSampleProjDrefImplicitLod
:
1573 case SpvOpImageSampleProjDrefExplicitLod
:
1574 case SpvOpImageFetch
:
1575 case SpvOpImageGather
:
1576 case SpvOpImageDrefGather
:
1577 case SpvOpImageQueryLod
: {
1578 /* All these types have the coordinate as their first real argument */
1579 switch (sampler_dim
) {
1580 case GLSL_SAMPLER_DIM_1D
:
1581 case GLSL_SAMPLER_DIM_BUF
:
1582 coord_components
= 1;
1584 case GLSL_SAMPLER_DIM_2D
:
1585 case GLSL_SAMPLER_DIM_RECT
:
1586 case GLSL_SAMPLER_DIM_MS
:
1587 coord_components
= 2;
1589 case GLSL_SAMPLER_DIM_3D
:
1590 case GLSL_SAMPLER_DIM_CUBE
:
1591 coord_components
= 3;
1594 unreachable("Invalid sampler type");
1597 if (is_array
&& texop
!= nir_texop_lod
)
1600 coord
= vtn_ssa_value(b
, w
[idx
++])->def
;
1601 p
->src
= nir_src_for_ssa(nir_channels(&b
->nb
, coord
,
1602 (1 << coord_components
) - 1));
1603 p
->src_type
= nir_tex_src_coord
;
1610 coord_components
= 0;
1615 case SpvOpImageSampleProjImplicitLod
:
1616 case SpvOpImageSampleProjExplicitLod
:
1617 case SpvOpImageSampleProjDrefImplicitLod
:
1618 case SpvOpImageSampleProjDrefExplicitLod
:
1619 /* These have the projector as the last coordinate component */
1620 p
->src
= nir_src_for_ssa(nir_channel(&b
->nb
, coord
, coord_components
));
1621 p
->src_type
= nir_tex_src_projector
;
1629 unsigned gather_component
= 0;
1631 case SpvOpImageSampleDrefImplicitLod
:
1632 case SpvOpImageSampleDrefExplicitLod
:
1633 case SpvOpImageSampleProjDrefImplicitLod
:
1634 case SpvOpImageSampleProjDrefExplicitLod
:
1635 case SpvOpImageDrefGather
:
1636 /* These all have an explicit depth value as their next source */
1637 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_comparator
);
1640 case SpvOpImageGather
:
1641 /* This has a component as its next source */
1643 vtn_value(b
, w
[idx
++], vtn_value_type_constant
)->constant
->values
[0].u32
[0];
1650 /* For OpImageQuerySizeLod, we always have an LOD */
1651 if (opcode
== SpvOpImageQuerySizeLod
)
1652 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
1654 /* Now we need to handle some number of optional arguments */
1655 const struct vtn_ssa_value
*gather_offsets
= NULL
;
1657 uint32_t operands
= w
[idx
++];
1659 if (operands
& SpvImageOperandsBiasMask
) {
1660 assert(texop
== nir_texop_tex
);
1661 texop
= nir_texop_txb
;
1662 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_bias
);
1665 if (operands
& SpvImageOperandsLodMask
) {
1666 assert(texop
== nir_texop_txl
|| texop
== nir_texop_txf
||
1667 texop
== nir_texop_txs
);
1668 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
1671 if (operands
& SpvImageOperandsGradMask
) {
1672 assert(texop
== nir_texop_txl
);
1673 texop
= nir_texop_txd
;
1674 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddx
);
1675 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddy
);
1678 if (operands
& SpvImageOperandsOffsetMask
||
1679 operands
& SpvImageOperandsConstOffsetMask
)
1680 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_offset
);
1682 if (operands
& SpvImageOperandsConstOffsetsMask
) {
1683 gather_offsets
= vtn_ssa_value(b
, w
[idx
++]);
1684 (*p
++) = (nir_tex_src
){};
1687 if (operands
& SpvImageOperandsSampleMask
) {
1688 assert(texop
== nir_texop_txf_ms
);
1689 texop
= nir_texop_txf_ms
;
1690 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ms_index
);
1693 /* We should have now consumed exactly all of the arguments */
1694 assert(idx
== count
);
1696 nir_tex_instr
*instr
= nir_tex_instr_create(b
->shader
, p
- srcs
);
1699 memcpy(instr
->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
1701 instr
->coord_components
= coord_components
;
1702 instr
->sampler_dim
= sampler_dim
;
1703 instr
->is_array
= is_array
;
1704 instr
->is_shadow
= is_shadow
;
1705 instr
->is_new_style_shadow
=
1706 is_shadow
&& glsl_get_components(ret_type
->type
) == 1;
1707 instr
->component
= gather_component
;
1709 switch (glsl_get_sampler_result_type(image_type
)) {
1710 case GLSL_TYPE_FLOAT
: instr
->dest_type
= nir_type_float
; break;
1711 case GLSL_TYPE_INT
: instr
->dest_type
= nir_type_int
; break;
1712 case GLSL_TYPE_UINT
: instr
->dest_type
= nir_type_uint
; break;
1713 case GLSL_TYPE_BOOL
: instr
->dest_type
= nir_type_bool
; break;
1715 unreachable("Invalid base type for sampler result");
1718 nir_deref_var
*sampler
= vtn_pointer_to_deref(b
, sampled
.sampler
);
1719 nir_deref_var
*texture
;
1720 if (sampled
.image
) {
1721 nir_deref_var
*image
= vtn_pointer_to_deref(b
, sampled
.image
);
1727 instr
->texture
= nir_deref_var_clone(texture
, instr
);
1729 switch (instr
->op
) {
1734 /* These operations require a sampler */
1735 instr
->sampler
= nir_deref_var_clone(sampler
, instr
);
1738 case nir_texop_txf_ms
:
1742 case nir_texop_query_levels
:
1743 case nir_texop_texture_samples
:
1744 case nir_texop_samples_identical
:
1746 instr
->sampler
= NULL
;
1748 case nir_texop_txf_ms_mcs
:
1749 unreachable("unexpected nir_texop_txf_ms_mcs");
1752 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1753 nir_tex_instr_dest_size(instr
), 32, NULL
);
1755 assert(glsl_get_vector_elements(ret_type
->type
) ==
1756 nir_tex_instr_dest_size(instr
));
1759 nir_instr
*instruction
;
1760 if (gather_offsets
) {
1761 assert(glsl_get_base_type(gather_offsets
->type
) == GLSL_TYPE_ARRAY
);
1762 assert(glsl_get_length(gather_offsets
->type
) == 4);
1763 nir_tex_instr
*instrs
[4] = {instr
, NULL
, NULL
, NULL
};
1765 /* Copy the current instruction 4x */
1766 for (uint32_t i
= 1; i
< 4; i
++) {
1767 instrs
[i
] = nir_tex_instr_create(b
->shader
, instr
->num_srcs
);
1768 instrs
[i
]->op
= instr
->op
;
1769 instrs
[i
]->coord_components
= instr
->coord_components
;
1770 instrs
[i
]->sampler_dim
= instr
->sampler_dim
;
1771 instrs
[i
]->is_array
= instr
->is_array
;
1772 instrs
[i
]->is_shadow
= instr
->is_shadow
;
1773 instrs
[i
]->is_new_style_shadow
= instr
->is_new_style_shadow
;
1774 instrs
[i
]->component
= instr
->component
;
1775 instrs
[i
]->dest_type
= instr
->dest_type
;
1776 instrs
[i
]->texture
= nir_deref_var_clone(texture
, instrs
[i
]);
1777 instrs
[i
]->sampler
= NULL
;
1779 memcpy(instrs
[i
]->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
1781 nir_ssa_dest_init(&instrs
[i
]->instr
, &instrs
[i
]->dest
,
1782 nir_tex_instr_dest_size(instr
), 32, NULL
);
1785 /* Fill in the last argument with the offset from the passed in offsets
1786 * and insert the instruction into the stream.
1788 for (uint32_t i
= 0; i
< 4; i
++) {
1790 src
.src
= nir_src_for_ssa(gather_offsets
->elems
[i
]->def
);
1791 src
.src_type
= nir_tex_src_offset
;
1792 instrs
[i
]->src
[instrs
[i
]->num_srcs
- 1] = src
;
1793 nir_builder_instr_insert(&b
->nb
, &instrs
[i
]->instr
);
1796 /* Combine the results of the 4 instructions by taking their .w
1799 nir_alu_instr
*vec4
= nir_alu_instr_create(b
->shader
, nir_op_vec4
);
1800 nir_ssa_dest_init(&vec4
->instr
, &vec4
->dest
.dest
, 4, 32, NULL
);
1801 vec4
->dest
.write_mask
= 0xf;
1802 for (uint32_t i
= 0; i
< 4; i
++) {
1803 vec4
->src
[i
].src
= nir_src_for_ssa(&instrs
[i
]->dest
.ssa
);
1804 vec4
->src
[i
].swizzle
[0] = 3;
1806 def
= &vec4
->dest
.dest
.ssa
;
1807 instruction
= &vec4
->instr
;
1809 def
= &instr
->dest
.ssa
;
1810 instruction
= &instr
->instr
;
1813 val
->ssa
= vtn_create_ssa_value(b
, ret_type
->type
);
1814 val
->ssa
->def
= def
;
1816 nir_builder_instr_insert(&b
->nb
, instruction
);
1820 fill_common_atomic_sources(struct vtn_builder
*b
, SpvOp opcode
,
1821 const uint32_t *w
, nir_src
*src
)
1824 case SpvOpAtomicIIncrement
:
1825 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
1828 case SpvOpAtomicIDecrement
:
1829 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
1832 case SpvOpAtomicISub
:
1834 nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
1837 case SpvOpAtomicCompareExchange
:
1838 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[8])->def
);
1839 src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
1842 case SpvOpAtomicExchange
:
1843 case SpvOpAtomicIAdd
:
1844 case SpvOpAtomicSMin
:
1845 case SpvOpAtomicUMin
:
1846 case SpvOpAtomicSMax
:
1847 case SpvOpAtomicUMax
:
1848 case SpvOpAtomicAnd
:
1850 case SpvOpAtomicXor
:
1851 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
1855 unreachable("Invalid SPIR-V atomic");
1859 static nir_ssa_def
*
1860 get_image_coord(struct vtn_builder
*b
, uint32_t value
)
1862 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, value
);
1864 /* The image_load_store intrinsics assume a 4-dim coordinate */
1865 unsigned dim
= glsl_get_vector_elements(coord
->type
);
1866 unsigned swizzle
[4];
1867 for (unsigned i
= 0; i
< 4; i
++)
1868 swizzle
[i
] = MIN2(i
, dim
- 1);
1870 return nir_swizzle(&b
->nb
, coord
->def
, swizzle
, 4, false);
1874 vtn_handle_image(struct vtn_builder
*b
, SpvOp opcode
,
1875 const uint32_t *w
, unsigned count
)
1877 /* Just get this one out of the way */
1878 if (opcode
== SpvOpImageTexelPointer
) {
1879 struct vtn_value
*val
=
1880 vtn_push_value(b
, w
[2], vtn_value_type_image_pointer
);
1881 val
->image
= ralloc(b
, struct vtn_image_pointer
);
1883 val
->image
->image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
1884 val
->image
->coord
= get_image_coord(b
, w
[4]);
1885 val
->image
->sample
= vtn_ssa_value(b
, w
[5])->def
;
1889 struct vtn_image_pointer image
;
1892 case SpvOpAtomicExchange
:
1893 case SpvOpAtomicCompareExchange
:
1894 case SpvOpAtomicCompareExchangeWeak
:
1895 case SpvOpAtomicIIncrement
:
1896 case SpvOpAtomicIDecrement
:
1897 case SpvOpAtomicIAdd
:
1898 case SpvOpAtomicISub
:
1899 case SpvOpAtomicLoad
:
1900 case SpvOpAtomicSMin
:
1901 case SpvOpAtomicUMin
:
1902 case SpvOpAtomicSMax
:
1903 case SpvOpAtomicUMax
:
1904 case SpvOpAtomicAnd
:
1906 case SpvOpAtomicXor
:
1907 image
= *vtn_value(b
, w
[3], vtn_value_type_image_pointer
)->image
;
1910 case SpvOpAtomicStore
:
1911 image
= *vtn_value(b
, w
[1], vtn_value_type_image_pointer
)->image
;
1914 case SpvOpImageQuerySize
:
1915 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
1917 image
.sample
= NULL
;
1920 case SpvOpImageRead
:
1921 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
1922 image
.coord
= get_image_coord(b
, w
[4]);
1924 if (count
> 5 && (w
[5] & SpvImageOperandsSampleMask
)) {
1925 assert(w
[5] == SpvImageOperandsSampleMask
);
1926 image
.sample
= vtn_ssa_value(b
, w
[6])->def
;
1928 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
1932 case SpvOpImageWrite
:
1933 image
.image
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
1934 image
.coord
= get_image_coord(b
, w
[2]);
1938 if (count
> 4 && (w
[4] & SpvImageOperandsSampleMask
)) {
1939 assert(w
[4] == SpvImageOperandsSampleMask
);
1940 image
.sample
= vtn_ssa_value(b
, w
[5])->def
;
1942 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
1947 unreachable("Invalid image opcode");
1950 nir_intrinsic_op op
;
1952 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_##N; break;
1953 OP(ImageQuerySize
, size
)
1955 OP(ImageWrite
, store
)
1956 OP(AtomicLoad
, load
)
1957 OP(AtomicStore
, store
)
1958 OP(AtomicExchange
, atomic_exchange
)
1959 OP(AtomicCompareExchange
, atomic_comp_swap
)
1960 OP(AtomicIIncrement
, atomic_add
)
1961 OP(AtomicIDecrement
, atomic_add
)
1962 OP(AtomicIAdd
, atomic_add
)
1963 OP(AtomicISub
, atomic_add
)
1964 OP(AtomicSMin
, atomic_min
)
1965 OP(AtomicUMin
, atomic_min
)
1966 OP(AtomicSMax
, atomic_max
)
1967 OP(AtomicUMax
, atomic_max
)
1968 OP(AtomicAnd
, atomic_and
)
1969 OP(AtomicOr
, atomic_or
)
1970 OP(AtomicXor
, atomic_xor
)
1973 unreachable("Invalid image opcode");
1976 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
1978 nir_deref_var
*image_deref
= vtn_pointer_to_deref(b
, image
.image
);
1979 intrin
->variables
[0] = nir_deref_var_clone(image_deref
, intrin
);
1981 /* ImageQuerySize doesn't take any extra parameters */
1982 if (opcode
!= SpvOpImageQuerySize
) {
1983 /* The image coordinate is always 4 components but we may not have that
1984 * many. Swizzle to compensate.
1987 for (unsigned i
= 0; i
< 4; i
++)
1988 swiz
[i
] = i
< image
.coord
->num_components
? i
: 0;
1989 intrin
->src
[0] = nir_src_for_ssa(nir_swizzle(&b
->nb
, image
.coord
,
1991 intrin
->src
[1] = nir_src_for_ssa(image
.sample
);
1995 case SpvOpAtomicLoad
:
1996 case SpvOpImageQuerySize
:
1997 case SpvOpImageRead
:
1999 case SpvOpAtomicStore
:
2000 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2002 case SpvOpImageWrite
:
2003 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[3])->def
);
2006 case SpvOpAtomicIIncrement
:
2007 case SpvOpAtomicIDecrement
:
2008 case SpvOpAtomicExchange
:
2009 case SpvOpAtomicIAdd
:
2010 case SpvOpAtomicSMin
:
2011 case SpvOpAtomicUMin
:
2012 case SpvOpAtomicSMax
:
2013 case SpvOpAtomicUMax
:
2014 case SpvOpAtomicAnd
:
2016 case SpvOpAtomicXor
:
2017 fill_common_atomic_sources(b
, opcode
, w
, &intrin
->src
[2]);
2021 unreachable("Invalid image opcode");
2024 if (opcode
!= SpvOpImageWrite
) {
2025 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2026 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2028 unsigned dest_components
=
2029 nir_intrinsic_infos
[intrin
->intrinsic
].dest_components
;
2030 if (intrin
->intrinsic
== nir_intrinsic_image_size
) {
2031 dest_components
= intrin
->num_components
=
2032 glsl_get_vector_elements(type
->type
);
2035 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
,
2036 dest_components
, 32, NULL
);
2038 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2040 val
->ssa
= vtn_create_ssa_value(b
, type
->type
);
2041 val
->ssa
->def
= &intrin
->dest
.ssa
;
2043 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2047 static nir_intrinsic_op
2048 get_ssbo_nir_atomic_op(SpvOp opcode
)
2051 case SpvOpAtomicLoad
: return nir_intrinsic_load_ssbo
;
2052 case SpvOpAtomicStore
: return nir_intrinsic_store_ssbo
;
2053 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
2054 OP(AtomicExchange
, atomic_exchange
)
2055 OP(AtomicCompareExchange
, atomic_comp_swap
)
2056 OP(AtomicIIncrement
, atomic_add
)
2057 OP(AtomicIDecrement
, atomic_add
)
2058 OP(AtomicIAdd
, atomic_add
)
2059 OP(AtomicISub
, atomic_add
)
2060 OP(AtomicSMin
, atomic_imin
)
2061 OP(AtomicUMin
, atomic_umin
)
2062 OP(AtomicSMax
, atomic_imax
)
2063 OP(AtomicUMax
, atomic_umax
)
2064 OP(AtomicAnd
, atomic_and
)
2065 OP(AtomicOr
, atomic_or
)
2066 OP(AtomicXor
, atomic_xor
)
2069 unreachable("Invalid SSBO atomic");
2073 static nir_intrinsic_op
2074 get_shared_nir_atomic_op(SpvOp opcode
)
2077 case SpvOpAtomicLoad
: return nir_intrinsic_load_var
;
2078 case SpvOpAtomicStore
: return nir_intrinsic_store_var
;
2079 #define OP(S, N) case SpvOp##S: return nir_intrinsic_var_##N;
2080 OP(AtomicExchange
, atomic_exchange
)
2081 OP(AtomicCompareExchange
, atomic_comp_swap
)
2082 OP(AtomicIIncrement
, atomic_add
)
2083 OP(AtomicIDecrement
, atomic_add
)
2084 OP(AtomicIAdd
, atomic_add
)
2085 OP(AtomicISub
, atomic_add
)
2086 OP(AtomicSMin
, atomic_imin
)
2087 OP(AtomicUMin
, atomic_umin
)
2088 OP(AtomicSMax
, atomic_imax
)
2089 OP(AtomicUMax
, atomic_umax
)
2090 OP(AtomicAnd
, atomic_and
)
2091 OP(AtomicOr
, atomic_or
)
2092 OP(AtomicXor
, atomic_xor
)
2095 unreachable("Invalid shared atomic");
2100 vtn_handle_ssbo_or_shared_atomic(struct vtn_builder
*b
, SpvOp opcode
,
2101 const uint32_t *w
, unsigned count
)
2103 struct vtn_pointer
*ptr
;
2104 nir_intrinsic_instr
*atomic
;
2107 case SpvOpAtomicLoad
:
2108 case SpvOpAtomicExchange
:
2109 case SpvOpAtomicCompareExchange
:
2110 case SpvOpAtomicCompareExchangeWeak
:
2111 case SpvOpAtomicIIncrement
:
2112 case SpvOpAtomicIDecrement
:
2113 case SpvOpAtomicIAdd
:
2114 case SpvOpAtomicISub
:
2115 case SpvOpAtomicSMin
:
2116 case SpvOpAtomicUMin
:
2117 case SpvOpAtomicSMax
:
2118 case SpvOpAtomicUMax
:
2119 case SpvOpAtomicAnd
:
2121 case SpvOpAtomicXor
:
2122 ptr
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2125 case SpvOpAtomicStore
:
2126 ptr
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
2130 unreachable("Invalid SPIR-V atomic");
2134 SpvScope scope = w[4];
2135 SpvMemorySemanticsMask semantics = w[5];
2138 if (ptr
->mode
== vtn_variable_mode_workgroup
) {
2139 nir_deref_var
*deref
= vtn_pointer_to_deref(b
, ptr
);
2140 const struct glsl_type
*deref_type
= nir_deref_tail(&deref
->deref
)->type
;
2141 nir_intrinsic_op op
= get_shared_nir_atomic_op(opcode
);
2142 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2143 atomic
->variables
[0] = nir_deref_var_clone(deref
, atomic
);
2146 case SpvOpAtomicLoad
:
2147 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2150 case SpvOpAtomicStore
:
2151 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2152 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2153 atomic
->src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
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 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[0]);
2174 unreachable("Invalid SPIR-V atomic");
2178 assert(ptr
->mode
== vtn_variable_mode_ssbo
);
2179 nir_ssa_def
*offset
, *index
;
2180 offset
= vtn_pointer_to_offset(b
, ptr
, &index
, NULL
);
2182 nir_intrinsic_op op
= get_ssbo_nir_atomic_op(opcode
);
2184 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2187 case SpvOpAtomicLoad
:
2188 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
2189 atomic
->src
[0] = nir_src_for_ssa(index
);
2190 atomic
->src
[1] = nir_src_for_ssa(offset
);
2193 case SpvOpAtomicStore
:
2194 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
2195 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2196 atomic
->src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2197 atomic
->src
[1] = nir_src_for_ssa(index
);
2198 atomic
->src
[2] = nir_src_for_ssa(offset
);
2201 case SpvOpAtomicExchange
:
2202 case SpvOpAtomicCompareExchange
:
2203 case SpvOpAtomicCompareExchangeWeak
:
2204 case SpvOpAtomicIIncrement
:
2205 case SpvOpAtomicIDecrement
:
2206 case SpvOpAtomicIAdd
:
2207 case SpvOpAtomicISub
:
2208 case SpvOpAtomicSMin
:
2209 case SpvOpAtomicUMin
:
2210 case SpvOpAtomicSMax
:
2211 case SpvOpAtomicUMax
:
2212 case SpvOpAtomicAnd
:
2214 case SpvOpAtomicXor
:
2215 atomic
->src
[0] = nir_src_for_ssa(index
);
2216 atomic
->src
[1] = nir_src_for_ssa(offset
);
2217 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[2]);
2221 unreachable("Invalid SPIR-V atomic");
2225 if (opcode
!= SpvOpAtomicStore
) {
2226 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2228 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
,
2229 glsl_get_vector_elements(type
->type
),
2230 glsl_get_bit_size(type
->type
), NULL
);
2232 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2233 val
->ssa
= rzalloc(b
, struct vtn_ssa_value
);
2234 val
->ssa
->def
= &atomic
->dest
.ssa
;
2235 val
->ssa
->type
= type
->type
;
2238 nir_builder_instr_insert(&b
->nb
, &atomic
->instr
);
2241 static nir_alu_instr
*
2242 create_vec(nir_shader
*shader
, unsigned num_components
, unsigned bit_size
)
2245 switch (num_components
) {
2246 case 1: op
= nir_op_fmov
; break;
2247 case 2: op
= nir_op_vec2
; break;
2248 case 3: op
= nir_op_vec3
; break;
2249 case 4: op
= nir_op_vec4
; break;
2250 default: unreachable("bad vector size");
2253 nir_alu_instr
*vec
= nir_alu_instr_create(shader
, op
);
2254 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, num_components
,
2256 vec
->dest
.write_mask
= (1 << num_components
) - 1;
2261 struct vtn_ssa_value
*
2262 vtn_ssa_transpose(struct vtn_builder
*b
, struct vtn_ssa_value
*src
)
2264 if (src
->transposed
)
2265 return src
->transposed
;
2267 struct vtn_ssa_value
*dest
=
2268 vtn_create_ssa_value(b
, glsl_transposed_type(src
->type
));
2270 for (unsigned i
= 0; i
< glsl_get_matrix_columns(dest
->type
); i
++) {
2271 nir_alu_instr
*vec
= create_vec(b
->shader
,
2272 glsl_get_matrix_columns(src
->type
),
2273 glsl_get_bit_size(src
->type
));
2274 if (glsl_type_is_vector_or_scalar(src
->type
)) {
2275 vec
->src
[0].src
= nir_src_for_ssa(src
->def
);
2276 vec
->src
[0].swizzle
[0] = i
;
2278 for (unsigned j
= 0; j
< glsl_get_matrix_columns(src
->type
); j
++) {
2279 vec
->src
[j
].src
= nir_src_for_ssa(src
->elems
[j
]->def
);
2280 vec
->src
[j
].swizzle
[0] = i
;
2283 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2284 dest
->elems
[i
]->def
= &vec
->dest
.dest
.ssa
;
2287 dest
->transposed
= src
;
2293 vtn_vector_extract(struct vtn_builder
*b
, nir_ssa_def
*src
, unsigned index
)
2295 unsigned swiz
[4] = { index
};
2296 return nir_swizzle(&b
->nb
, src
, swiz
, 1, true);
2300 vtn_vector_insert(struct vtn_builder
*b
, nir_ssa_def
*src
, nir_ssa_def
*insert
,
2303 nir_alu_instr
*vec
= create_vec(b
->shader
, src
->num_components
,
2306 for (unsigned i
= 0; i
< src
->num_components
; i
++) {
2308 vec
->src
[i
].src
= nir_src_for_ssa(insert
);
2310 vec
->src
[i
].src
= nir_src_for_ssa(src
);
2311 vec
->src
[i
].swizzle
[0] = i
;
2315 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2317 return &vec
->dest
.dest
.ssa
;
2321 vtn_vector_extract_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
2324 nir_ssa_def
*dest
= vtn_vector_extract(b
, src
, 0);
2325 for (unsigned i
= 1; i
< src
->num_components
; i
++)
2326 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
2327 vtn_vector_extract(b
, src
, i
), dest
);
2333 vtn_vector_insert_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
2334 nir_ssa_def
*insert
, nir_ssa_def
*index
)
2336 nir_ssa_def
*dest
= vtn_vector_insert(b
, src
, insert
, 0);
2337 for (unsigned i
= 1; i
< src
->num_components
; i
++)
2338 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
2339 vtn_vector_insert(b
, src
, insert
, i
), dest
);
2344 static nir_ssa_def
*
2345 vtn_vector_shuffle(struct vtn_builder
*b
, unsigned num_components
,
2346 nir_ssa_def
*src0
, nir_ssa_def
*src1
,
2347 const uint32_t *indices
)
2349 nir_alu_instr
*vec
= create_vec(b
->shader
, num_components
, src0
->bit_size
);
2351 for (unsigned i
= 0; i
< num_components
; i
++) {
2352 uint32_t index
= indices
[i
];
2353 if (index
== 0xffffffff) {
2355 nir_src_for_ssa(nir_ssa_undef(&b
->nb
, 1, src0
->bit_size
));
2356 } else if (index
< src0
->num_components
) {
2357 vec
->src
[i
].src
= nir_src_for_ssa(src0
);
2358 vec
->src
[i
].swizzle
[0] = index
;
2360 vec
->src
[i
].src
= nir_src_for_ssa(src1
);
2361 vec
->src
[i
].swizzle
[0] = index
- src0
->num_components
;
2365 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2367 return &vec
->dest
.dest
.ssa
;
2371 * Concatentates a number of vectors/scalars together to produce a vector
2373 static nir_ssa_def
*
2374 vtn_vector_construct(struct vtn_builder
*b
, unsigned num_components
,
2375 unsigned num_srcs
, nir_ssa_def
**srcs
)
2377 nir_alu_instr
*vec
= create_vec(b
->shader
, num_components
,
2380 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2382 * "When constructing a vector, there must be at least two Constituent
2385 assert(num_srcs
>= 2);
2387 unsigned dest_idx
= 0;
2388 for (unsigned i
= 0; i
< num_srcs
; i
++) {
2389 nir_ssa_def
*src
= srcs
[i
];
2390 assert(dest_idx
+ src
->num_components
<= num_components
);
2391 for (unsigned j
= 0; j
< src
->num_components
; j
++) {
2392 vec
->src
[dest_idx
].src
= nir_src_for_ssa(src
);
2393 vec
->src
[dest_idx
].swizzle
[0] = j
;
2398 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2400 * "When constructing a vector, the total number of components in all
2401 * the operands must equal the number of components in Result Type."
2403 assert(dest_idx
== num_components
);
2405 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2407 return &vec
->dest
.dest
.ssa
;
2410 static struct vtn_ssa_value
*
2411 vtn_composite_copy(void *mem_ctx
, struct vtn_ssa_value
*src
)
2413 struct vtn_ssa_value
*dest
= rzalloc(mem_ctx
, struct vtn_ssa_value
);
2414 dest
->type
= src
->type
;
2416 if (glsl_type_is_vector_or_scalar(src
->type
)) {
2417 dest
->def
= src
->def
;
2419 unsigned elems
= glsl_get_length(src
->type
);
2421 dest
->elems
= ralloc_array(mem_ctx
, struct vtn_ssa_value
*, elems
);
2422 for (unsigned i
= 0; i
< elems
; i
++)
2423 dest
->elems
[i
] = vtn_composite_copy(mem_ctx
, src
->elems
[i
]);
2429 static struct vtn_ssa_value
*
2430 vtn_composite_insert(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
2431 struct vtn_ssa_value
*insert
, const uint32_t *indices
,
2432 unsigned num_indices
)
2434 struct vtn_ssa_value
*dest
= vtn_composite_copy(b
, src
);
2436 struct vtn_ssa_value
*cur
= dest
;
2438 for (i
= 0; i
< num_indices
- 1; i
++) {
2439 cur
= cur
->elems
[indices
[i
]];
2442 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
2443 /* According to the SPIR-V spec, OpCompositeInsert may work down to
2444 * the component granularity. In that case, the last index will be
2445 * the index to insert the scalar into the vector.
2448 cur
->def
= vtn_vector_insert(b
, cur
->def
, insert
->def
, indices
[i
]);
2450 cur
->elems
[indices
[i
]] = insert
;
2456 static struct vtn_ssa_value
*
2457 vtn_composite_extract(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
2458 const uint32_t *indices
, unsigned num_indices
)
2460 struct vtn_ssa_value
*cur
= src
;
2461 for (unsigned i
= 0; i
< num_indices
; i
++) {
2462 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
2463 assert(i
== num_indices
- 1);
2464 /* According to the SPIR-V spec, OpCompositeExtract may work down to
2465 * the component granularity. The last index will be the index of the
2466 * vector to extract.
2469 struct vtn_ssa_value
*ret
= rzalloc(b
, struct vtn_ssa_value
);
2470 ret
->type
= glsl_scalar_type(glsl_get_base_type(cur
->type
));
2471 ret
->def
= vtn_vector_extract(b
, cur
->def
, indices
[i
]);
2474 cur
= cur
->elems
[indices
[i
]];
2482 vtn_handle_composite(struct vtn_builder
*b
, SpvOp opcode
,
2483 const uint32_t *w
, unsigned count
)
2485 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2486 const struct glsl_type
*type
=
2487 vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
2488 val
->ssa
= vtn_create_ssa_value(b
, type
);
2491 case SpvOpVectorExtractDynamic
:
2492 val
->ssa
->def
= vtn_vector_extract_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
2493 vtn_ssa_value(b
, w
[4])->def
);
2496 case SpvOpVectorInsertDynamic
:
2497 val
->ssa
->def
= vtn_vector_insert_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
2498 vtn_ssa_value(b
, w
[4])->def
,
2499 vtn_ssa_value(b
, w
[5])->def
);
2502 case SpvOpVectorShuffle
:
2503 val
->ssa
->def
= vtn_vector_shuffle(b
, glsl_get_vector_elements(type
),
2504 vtn_ssa_value(b
, w
[3])->def
,
2505 vtn_ssa_value(b
, w
[4])->def
,
2509 case SpvOpCompositeConstruct
: {
2510 unsigned elems
= count
- 3;
2511 if (glsl_type_is_vector_or_scalar(type
)) {
2512 nir_ssa_def
*srcs
[4];
2513 for (unsigned i
= 0; i
< elems
; i
++)
2514 srcs
[i
] = vtn_ssa_value(b
, w
[3 + i
])->def
;
2516 vtn_vector_construct(b
, glsl_get_vector_elements(type
),
2519 val
->ssa
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
2520 for (unsigned i
= 0; i
< elems
; i
++)
2521 val
->ssa
->elems
[i
] = vtn_ssa_value(b
, w
[3 + i
]);
2525 case SpvOpCompositeExtract
:
2526 val
->ssa
= vtn_composite_extract(b
, vtn_ssa_value(b
, w
[3]),
2530 case SpvOpCompositeInsert
:
2531 val
->ssa
= vtn_composite_insert(b
, vtn_ssa_value(b
, w
[4]),
2532 vtn_ssa_value(b
, w
[3]),
2536 case SpvOpCopyObject
:
2537 val
->ssa
= vtn_composite_copy(b
, vtn_ssa_value(b
, w
[3]));
2541 unreachable("unknown composite operation");
2546 vtn_handle_barrier(struct vtn_builder
*b
, SpvOp opcode
,
2547 const uint32_t *w
, unsigned count
)
2549 nir_intrinsic_op intrinsic_op
;
2551 case SpvOpEmitVertex
:
2552 case SpvOpEmitStreamVertex
:
2553 intrinsic_op
= nir_intrinsic_emit_vertex
;
2555 case SpvOpEndPrimitive
:
2556 case SpvOpEndStreamPrimitive
:
2557 intrinsic_op
= nir_intrinsic_end_primitive
;
2559 case SpvOpMemoryBarrier
:
2560 intrinsic_op
= nir_intrinsic_memory_barrier
;
2562 case SpvOpControlBarrier
:
2563 intrinsic_op
= nir_intrinsic_barrier
;
2566 unreachable("unknown barrier instruction");
2569 nir_intrinsic_instr
*intrin
=
2570 nir_intrinsic_instr_create(b
->shader
, intrinsic_op
);
2572 if (opcode
== SpvOpEmitStreamVertex
|| opcode
== SpvOpEndStreamPrimitive
)
2573 nir_intrinsic_set_stream_id(intrin
, w
[1]);
2575 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2579 gl_primitive_from_spv_execution_mode(SpvExecutionMode mode
)
2582 case SpvExecutionModeInputPoints
:
2583 case SpvExecutionModeOutputPoints
:
2584 return 0; /* GL_POINTS */
2585 case SpvExecutionModeInputLines
:
2586 return 1; /* GL_LINES */
2587 case SpvExecutionModeInputLinesAdjacency
:
2588 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
2589 case SpvExecutionModeTriangles
:
2590 return 4; /* GL_TRIANGLES */
2591 case SpvExecutionModeInputTrianglesAdjacency
:
2592 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
2593 case SpvExecutionModeQuads
:
2594 return 7; /* GL_QUADS */
2595 case SpvExecutionModeIsolines
:
2596 return 0x8E7A; /* GL_ISOLINES */
2597 case SpvExecutionModeOutputLineStrip
:
2598 return 3; /* GL_LINE_STRIP */
2599 case SpvExecutionModeOutputTriangleStrip
:
2600 return 5; /* GL_TRIANGLE_STRIP */
2602 assert(!"Invalid primitive type");
2608 vertices_in_from_spv_execution_mode(SpvExecutionMode mode
)
2611 case SpvExecutionModeInputPoints
:
2613 case SpvExecutionModeInputLines
:
2615 case SpvExecutionModeInputLinesAdjacency
:
2617 case SpvExecutionModeTriangles
:
2619 case SpvExecutionModeInputTrianglesAdjacency
:
2622 assert(!"Invalid GS input mode");
2627 static gl_shader_stage
2628 stage_for_execution_model(SpvExecutionModel model
)
2631 case SpvExecutionModelVertex
:
2632 return MESA_SHADER_VERTEX
;
2633 case SpvExecutionModelTessellationControl
:
2634 return MESA_SHADER_TESS_CTRL
;
2635 case SpvExecutionModelTessellationEvaluation
:
2636 return MESA_SHADER_TESS_EVAL
;
2637 case SpvExecutionModelGeometry
:
2638 return MESA_SHADER_GEOMETRY
;
2639 case SpvExecutionModelFragment
:
2640 return MESA_SHADER_FRAGMENT
;
2641 case SpvExecutionModelGLCompute
:
2642 return MESA_SHADER_COMPUTE
;
2644 unreachable("Unsupported execution model");
2648 #define spv_check_supported(name, cap) do { \
2649 if (!(b->ext && b->ext->name)) \
2650 vtn_warn("Unsupported SPIR-V capability: %s", \
2651 spirv_capability_to_string(cap)); \
2655 vtn_handle_preamble_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2656 const uint32_t *w
, unsigned count
)
2660 case SpvOpSourceExtension
:
2661 case SpvOpSourceContinued
:
2662 case SpvOpExtension
:
2663 /* Unhandled, but these are for debug so that's ok. */
2666 case SpvOpCapability
: {
2667 SpvCapability cap
= w
[1];
2669 case SpvCapabilityMatrix
:
2670 case SpvCapabilityShader
:
2671 case SpvCapabilityGeometry
:
2672 case SpvCapabilityGeometryPointSize
:
2673 case SpvCapabilityUniformBufferArrayDynamicIndexing
:
2674 case SpvCapabilitySampledImageArrayDynamicIndexing
:
2675 case SpvCapabilityStorageBufferArrayDynamicIndexing
:
2676 case SpvCapabilityStorageImageArrayDynamicIndexing
:
2677 case SpvCapabilityImageRect
:
2678 case SpvCapabilitySampledRect
:
2679 case SpvCapabilitySampled1D
:
2680 case SpvCapabilityImage1D
:
2681 case SpvCapabilitySampledCubeArray
:
2682 case SpvCapabilitySampledBuffer
:
2683 case SpvCapabilityImageBuffer
:
2684 case SpvCapabilityImageQuery
:
2685 case SpvCapabilityDerivativeControl
:
2686 case SpvCapabilityInterpolationFunction
:
2687 case SpvCapabilityMultiViewport
:
2688 case SpvCapabilitySampleRateShading
:
2689 case SpvCapabilityClipDistance
:
2690 case SpvCapabilityCullDistance
:
2691 case SpvCapabilityInputAttachment
:
2692 case SpvCapabilityImageGatherExtended
:
2693 case SpvCapabilityStorageImageExtendedFormats
:
2696 case SpvCapabilityGeometryStreams
:
2697 case SpvCapabilityLinkage
:
2698 case SpvCapabilityVector16
:
2699 case SpvCapabilityFloat16Buffer
:
2700 case SpvCapabilityFloat16
:
2701 case SpvCapabilityInt64Atomics
:
2702 case SpvCapabilityAtomicStorage
:
2703 case SpvCapabilityInt16
:
2704 case SpvCapabilityStorageImageMultisample
:
2705 case SpvCapabilityImageCubeArray
:
2706 case SpvCapabilityInt8
:
2707 case SpvCapabilitySparseResidency
:
2708 case SpvCapabilityMinLod
:
2709 case SpvCapabilityTransformFeedback
:
2710 vtn_warn("Unsupported SPIR-V capability: %s",
2711 spirv_capability_to_string(cap
));
2714 case SpvCapabilityFloat64
:
2715 spv_check_supported(float64
, cap
);
2717 case SpvCapabilityInt64
:
2718 spv_check_supported(int64
, cap
);
2721 case SpvCapabilityAddresses
:
2722 case SpvCapabilityKernel
:
2723 case SpvCapabilityImageBasic
:
2724 case SpvCapabilityImageReadWrite
:
2725 case SpvCapabilityImageMipmap
:
2726 case SpvCapabilityPipes
:
2727 case SpvCapabilityGroups
:
2728 case SpvCapabilityDeviceEnqueue
:
2729 case SpvCapabilityLiteralSampler
:
2730 case SpvCapabilityGenericPointer
:
2731 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
2732 spirv_capability_to_string(cap
));
2735 case SpvCapabilityImageMSArray
:
2736 spv_check_supported(image_ms_array
, cap
);
2739 case SpvCapabilityTessellation
:
2740 case SpvCapabilityTessellationPointSize
:
2741 spv_check_supported(tessellation
, cap
);
2744 case SpvCapabilityDrawParameters
:
2745 spv_check_supported(draw_parameters
, cap
);
2748 case SpvCapabilityStorageImageReadWithoutFormat
:
2749 spv_check_supported(image_read_without_format
, cap
);
2752 case SpvCapabilityStorageImageWriteWithoutFormat
:
2753 spv_check_supported(image_write_without_format
, cap
);
2756 case SpvCapabilityMultiView
:
2757 spv_check_supported(multiview
, cap
);
2761 unreachable("Unhandled capability");
2766 case SpvOpExtInstImport
:
2767 vtn_handle_extension(b
, opcode
, w
, count
);
2770 case SpvOpMemoryModel
:
2771 assert(w
[1] == SpvAddressingModelLogical
);
2772 assert(w
[2] == SpvMemoryModelGLSL450
);
2775 case SpvOpEntryPoint
: {
2776 struct vtn_value
*entry_point
= &b
->values
[w
[2]];
2777 /* Let this be a name label regardless */
2778 unsigned name_words
;
2779 entry_point
->name
= vtn_string_literal(b
, &w
[3], count
- 3, &name_words
);
2781 if (strcmp(entry_point
->name
, b
->entry_point_name
) != 0 ||
2782 stage_for_execution_model(w
[1]) != b
->entry_point_stage
)
2785 assert(b
->entry_point
== NULL
);
2786 b
->entry_point
= entry_point
;
2791 vtn_push_value(b
, w
[1], vtn_value_type_string
)->str
=
2792 vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
2796 b
->values
[w
[1]].name
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
2799 case SpvOpMemberName
:
2803 case SpvOpExecutionMode
:
2804 case SpvOpDecorationGroup
:
2806 case SpvOpMemberDecorate
:
2807 case SpvOpGroupDecorate
:
2808 case SpvOpGroupMemberDecorate
:
2809 vtn_handle_decoration(b
, opcode
, w
, count
);
2813 return false; /* End of preamble */
2820 vtn_handle_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
2821 const struct vtn_decoration
*mode
, void *data
)
2823 assert(b
->entry_point
== entry_point
);
2825 switch(mode
->exec_mode
) {
2826 case SpvExecutionModeOriginUpperLeft
:
2827 case SpvExecutionModeOriginLowerLeft
:
2828 b
->origin_upper_left
=
2829 (mode
->exec_mode
== SpvExecutionModeOriginUpperLeft
);
2832 case SpvExecutionModeEarlyFragmentTests
:
2833 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2834 b
->shader
->info
.fs
.early_fragment_tests
= true;
2837 case SpvExecutionModeInvocations
:
2838 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2839 b
->shader
->info
.gs
.invocations
= MAX2(1, mode
->literals
[0]);
2842 case SpvExecutionModeDepthReplacing
:
2843 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2844 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_ANY
;
2846 case SpvExecutionModeDepthGreater
:
2847 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2848 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_GREATER
;
2850 case SpvExecutionModeDepthLess
:
2851 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2852 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_LESS
;
2854 case SpvExecutionModeDepthUnchanged
:
2855 assert(b
->shader
->stage
== MESA_SHADER_FRAGMENT
);
2856 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
2859 case SpvExecutionModeLocalSize
:
2860 assert(b
->shader
->stage
== MESA_SHADER_COMPUTE
);
2861 b
->shader
->info
.cs
.local_size
[0] = mode
->literals
[0];
2862 b
->shader
->info
.cs
.local_size
[1] = mode
->literals
[1];
2863 b
->shader
->info
.cs
.local_size
[2] = mode
->literals
[2];
2865 case SpvExecutionModeLocalSizeHint
:
2866 break; /* Nothing to do with this */
2868 case SpvExecutionModeOutputVertices
:
2869 if (b
->shader
->stage
== MESA_SHADER_TESS_CTRL
||
2870 b
->shader
->stage
== MESA_SHADER_TESS_EVAL
) {
2871 b
->shader
->info
.tess
.tcs_vertices_out
= mode
->literals
[0];
2873 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2874 b
->shader
->info
.gs
.vertices_out
= mode
->literals
[0];
2878 case SpvExecutionModeInputPoints
:
2879 case SpvExecutionModeInputLines
:
2880 case SpvExecutionModeInputLinesAdjacency
:
2881 case SpvExecutionModeTriangles
:
2882 case SpvExecutionModeInputTrianglesAdjacency
:
2883 case SpvExecutionModeQuads
:
2884 case SpvExecutionModeIsolines
:
2885 if (b
->shader
->stage
== MESA_SHADER_TESS_CTRL
||
2886 b
->shader
->stage
== MESA_SHADER_TESS_EVAL
) {
2887 b
->shader
->info
.tess
.primitive_mode
=
2888 gl_primitive_from_spv_execution_mode(mode
->exec_mode
);
2890 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2891 b
->shader
->info
.gs
.vertices_in
=
2892 vertices_in_from_spv_execution_mode(mode
->exec_mode
);
2896 case SpvExecutionModeOutputPoints
:
2897 case SpvExecutionModeOutputLineStrip
:
2898 case SpvExecutionModeOutputTriangleStrip
:
2899 assert(b
->shader
->stage
== MESA_SHADER_GEOMETRY
);
2900 b
->shader
->info
.gs
.output_primitive
=
2901 gl_primitive_from_spv_execution_mode(mode
->exec_mode
);
2904 case SpvExecutionModeSpacingEqual
:
2905 assert(b
->shader
->stage
== MESA_SHADER_TESS_CTRL
||
2906 b
->shader
->stage
== MESA_SHADER_TESS_EVAL
);
2907 b
->shader
->info
.tess
.spacing
= TESS_SPACING_EQUAL
;
2909 case SpvExecutionModeSpacingFractionalEven
:
2910 assert(b
->shader
->stage
== MESA_SHADER_TESS_CTRL
||
2911 b
->shader
->stage
== MESA_SHADER_TESS_EVAL
);
2912 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_EVEN
;
2914 case SpvExecutionModeSpacingFractionalOdd
:
2915 assert(b
->shader
->stage
== MESA_SHADER_TESS_CTRL
||
2916 b
->shader
->stage
== MESA_SHADER_TESS_EVAL
);
2917 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_ODD
;
2919 case SpvExecutionModeVertexOrderCw
:
2920 assert(b
->shader
->stage
== MESA_SHADER_TESS_CTRL
||
2921 b
->shader
->stage
== MESA_SHADER_TESS_EVAL
);
2922 /* Vulkan's notion of CCW seems to match the hardware backends,
2923 * but be the opposite of OpenGL. Currently NIR follows GL semantics,
2924 * so we set it backwards here.
2926 b
->shader
->info
.tess
.ccw
= true;
2928 case SpvExecutionModeVertexOrderCcw
:
2929 assert(b
->shader
->stage
== MESA_SHADER_TESS_CTRL
||
2930 b
->shader
->stage
== MESA_SHADER_TESS_EVAL
);
2931 /* Backwards; see above */
2932 b
->shader
->info
.tess
.ccw
= false;
2934 case SpvExecutionModePointMode
:
2935 assert(b
->shader
->stage
== MESA_SHADER_TESS_CTRL
||
2936 b
->shader
->stage
== MESA_SHADER_TESS_EVAL
);
2937 b
->shader
->info
.tess
.point_mode
= true;
2940 case SpvExecutionModePixelCenterInteger
:
2941 b
->pixel_center_integer
= true;
2944 case SpvExecutionModeXfb
:
2945 assert(!"Unhandled execution mode");
2948 case SpvExecutionModeVecTypeHint
:
2949 case SpvExecutionModeContractionOff
:
2953 unreachable("Unhandled execution mode");
2958 vtn_handle_variable_or_type_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2959 const uint32_t *w
, unsigned count
)
2963 case SpvOpSourceContinued
:
2964 case SpvOpSourceExtension
:
2965 case SpvOpExtension
:
2966 case SpvOpCapability
:
2967 case SpvOpExtInstImport
:
2968 case SpvOpMemoryModel
:
2969 case SpvOpEntryPoint
:
2970 case SpvOpExecutionMode
:
2973 case SpvOpMemberName
:
2974 case SpvOpDecorationGroup
:
2976 case SpvOpMemberDecorate
:
2977 case SpvOpGroupDecorate
:
2978 case SpvOpGroupMemberDecorate
:
2979 assert(!"Invalid opcode types and variables section");
2985 case SpvOpTypeFloat
:
2986 case SpvOpTypeVector
:
2987 case SpvOpTypeMatrix
:
2988 case SpvOpTypeImage
:
2989 case SpvOpTypeSampler
:
2990 case SpvOpTypeSampledImage
:
2991 case SpvOpTypeArray
:
2992 case SpvOpTypeRuntimeArray
:
2993 case SpvOpTypeStruct
:
2994 case SpvOpTypeOpaque
:
2995 case SpvOpTypePointer
:
2996 case SpvOpTypeFunction
:
2997 case SpvOpTypeEvent
:
2998 case SpvOpTypeDeviceEvent
:
2999 case SpvOpTypeReserveId
:
3000 case SpvOpTypeQueue
:
3002 vtn_handle_type(b
, opcode
, w
, count
);
3005 case SpvOpConstantTrue
:
3006 case SpvOpConstantFalse
:
3008 case SpvOpConstantComposite
:
3009 case SpvOpConstantSampler
:
3010 case SpvOpConstantNull
:
3011 case SpvOpSpecConstantTrue
:
3012 case SpvOpSpecConstantFalse
:
3013 case SpvOpSpecConstant
:
3014 case SpvOpSpecConstantComposite
:
3015 case SpvOpSpecConstantOp
:
3016 vtn_handle_constant(b
, opcode
, w
, count
);
3021 vtn_handle_variables(b
, opcode
, w
, count
);
3025 return false; /* End of preamble */
3032 vtn_handle_body_instruction(struct vtn_builder
*b
, SpvOp opcode
,
3033 const uint32_t *w
, unsigned count
)
3039 case SpvOpLoopMerge
:
3040 case SpvOpSelectionMerge
:
3041 /* This is handled by cfg pre-pass and walk_blocks */
3045 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
3046 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
3051 vtn_handle_extension(b
, opcode
, w
, count
);
3057 case SpvOpCopyMemory
:
3058 case SpvOpCopyMemorySized
:
3059 case SpvOpAccessChain
:
3060 case SpvOpInBoundsAccessChain
:
3061 case SpvOpArrayLength
:
3062 vtn_handle_variables(b
, opcode
, w
, count
);
3065 case SpvOpFunctionCall
:
3066 vtn_handle_function_call(b
, opcode
, w
, count
);
3069 case SpvOpSampledImage
:
3071 case SpvOpImageSampleImplicitLod
:
3072 case SpvOpImageSampleExplicitLod
:
3073 case SpvOpImageSampleDrefImplicitLod
:
3074 case SpvOpImageSampleDrefExplicitLod
:
3075 case SpvOpImageSampleProjImplicitLod
:
3076 case SpvOpImageSampleProjExplicitLod
:
3077 case SpvOpImageSampleProjDrefImplicitLod
:
3078 case SpvOpImageSampleProjDrefExplicitLod
:
3079 case SpvOpImageFetch
:
3080 case SpvOpImageGather
:
3081 case SpvOpImageDrefGather
:
3082 case SpvOpImageQuerySizeLod
:
3083 case SpvOpImageQueryLod
:
3084 case SpvOpImageQueryLevels
:
3085 case SpvOpImageQuerySamples
:
3086 vtn_handle_texture(b
, opcode
, w
, count
);
3089 case SpvOpImageRead
:
3090 case SpvOpImageWrite
:
3091 case SpvOpImageTexelPointer
:
3092 vtn_handle_image(b
, opcode
, w
, count
);
3095 case SpvOpImageQuerySize
: {
3096 struct vtn_pointer
*image
=
3097 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
3098 if (image
->mode
== vtn_variable_mode_image
) {
3099 vtn_handle_image(b
, opcode
, w
, count
);
3101 assert(image
->mode
== vtn_variable_mode_sampler
);
3102 vtn_handle_texture(b
, opcode
, w
, count
);
3107 case SpvOpAtomicLoad
:
3108 case SpvOpAtomicExchange
:
3109 case SpvOpAtomicCompareExchange
:
3110 case SpvOpAtomicCompareExchangeWeak
:
3111 case SpvOpAtomicIIncrement
:
3112 case SpvOpAtomicIDecrement
:
3113 case SpvOpAtomicIAdd
:
3114 case SpvOpAtomicISub
:
3115 case SpvOpAtomicSMin
:
3116 case SpvOpAtomicUMin
:
3117 case SpvOpAtomicSMax
:
3118 case SpvOpAtomicUMax
:
3119 case SpvOpAtomicAnd
:
3121 case SpvOpAtomicXor
: {
3122 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[3]);
3123 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
3124 vtn_handle_image(b
, opcode
, w
, count
);
3126 assert(pointer
->value_type
== vtn_value_type_pointer
);
3127 vtn_handle_ssbo_or_shared_atomic(b
, opcode
, w
, count
);
3132 case SpvOpAtomicStore
: {
3133 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[1]);
3134 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
3135 vtn_handle_image(b
, opcode
, w
, count
);
3137 assert(pointer
->value_type
== vtn_value_type_pointer
);
3138 vtn_handle_ssbo_or_shared_atomic(b
, opcode
, w
, count
);
3148 case SpvOpConvertFToU
:
3149 case SpvOpConvertFToS
:
3150 case SpvOpConvertSToF
:
3151 case SpvOpConvertUToF
:
3155 case SpvOpQuantizeToF16
:
3156 case SpvOpConvertPtrToU
:
3157 case SpvOpConvertUToPtr
:
3158 case SpvOpPtrCastToGeneric
:
3159 case SpvOpGenericCastToPtr
:
3165 case SpvOpSignBitSet
:
3166 case SpvOpLessOrGreater
:
3168 case SpvOpUnordered
:
3183 case SpvOpVectorTimesScalar
:
3185 case SpvOpIAddCarry
:
3186 case SpvOpISubBorrow
:
3187 case SpvOpUMulExtended
:
3188 case SpvOpSMulExtended
:
3189 case SpvOpShiftRightLogical
:
3190 case SpvOpShiftRightArithmetic
:
3191 case SpvOpShiftLeftLogical
:
3192 case SpvOpLogicalEqual
:
3193 case SpvOpLogicalNotEqual
:
3194 case SpvOpLogicalOr
:
3195 case SpvOpLogicalAnd
:
3196 case SpvOpLogicalNot
:
3197 case SpvOpBitwiseOr
:
3198 case SpvOpBitwiseXor
:
3199 case SpvOpBitwiseAnd
:
3202 case SpvOpFOrdEqual
:
3203 case SpvOpFUnordEqual
:
3204 case SpvOpINotEqual
:
3205 case SpvOpFOrdNotEqual
:
3206 case SpvOpFUnordNotEqual
:
3207 case SpvOpULessThan
:
3208 case SpvOpSLessThan
:
3209 case SpvOpFOrdLessThan
:
3210 case SpvOpFUnordLessThan
:
3211 case SpvOpUGreaterThan
:
3212 case SpvOpSGreaterThan
:
3213 case SpvOpFOrdGreaterThan
:
3214 case SpvOpFUnordGreaterThan
:
3215 case SpvOpULessThanEqual
:
3216 case SpvOpSLessThanEqual
:
3217 case SpvOpFOrdLessThanEqual
:
3218 case SpvOpFUnordLessThanEqual
:
3219 case SpvOpUGreaterThanEqual
:
3220 case SpvOpSGreaterThanEqual
:
3221 case SpvOpFOrdGreaterThanEqual
:
3222 case SpvOpFUnordGreaterThanEqual
:
3228 case SpvOpFwidthFine
:
3229 case SpvOpDPdxCoarse
:
3230 case SpvOpDPdyCoarse
:
3231 case SpvOpFwidthCoarse
:
3232 case SpvOpBitFieldInsert
:
3233 case SpvOpBitFieldSExtract
:
3234 case SpvOpBitFieldUExtract
:
3235 case SpvOpBitReverse
:
3237 case SpvOpTranspose
:
3238 case SpvOpOuterProduct
:
3239 case SpvOpMatrixTimesScalar
:
3240 case SpvOpVectorTimesMatrix
:
3241 case SpvOpMatrixTimesVector
:
3242 case SpvOpMatrixTimesMatrix
:
3243 vtn_handle_alu(b
, opcode
, w
, count
);
3246 case SpvOpVectorExtractDynamic
:
3247 case SpvOpVectorInsertDynamic
:
3248 case SpvOpVectorShuffle
:
3249 case SpvOpCompositeConstruct
:
3250 case SpvOpCompositeExtract
:
3251 case SpvOpCompositeInsert
:
3252 case SpvOpCopyObject
:
3253 vtn_handle_composite(b
, opcode
, w
, count
);
3256 case SpvOpEmitVertex
:
3257 case SpvOpEndPrimitive
:
3258 case SpvOpEmitStreamVertex
:
3259 case SpvOpEndStreamPrimitive
:
3260 case SpvOpControlBarrier
:
3261 case SpvOpMemoryBarrier
:
3262 vtn_handle_barrier(b
, opcode
, w
, count
);
3266 unreachable("Unhandled opcode");
3273 spirv_to_nir(const uint32_t *words
, size_t word_count
,
3274 struct nir_spirv_specialization
*spec
, unsigned num_spec
,
3275 gl_shader_stage stage
, const char *entry_point_name
,
3276 const struct nir_spirv_supported_extensions
*ext
,
3277 const nir_shader_compiler_options
*options
)
3279 const uint32_t *word_end
= words
+ word_count
;
3281 /* Handle the SPIR-V header (first 4 dwords) */
3282 assert(word_count
> 5);
3284 assert(words
[0] == SpvMagicNumber
);
3285 assert(words
[1] >= 0x10000);
3286 /* words[2] == generator magic */
3287 unsigned value_id_bound
= words
[3];
3288 assert(words
[4] == 0);
3292 /* Initialize the stn_builder object */
3293 struct vtn_builder
*b
= rzalloc(NULL
, struct vtn_builder
);
3294 b
->value_id_bound
= value_id_bound
;
3295 b
->values
= rzalloc_array(b
, struct vtn_value
, value_id_bound
);
3296 exec_list_make_empty(&b
->functions
);
3297 b
->entry_point_stage
= stage
;
3298 b
->entry_point_name
= entry_point_name
;
3301 /* Handle all the preamble instructions */
3302 words
= vtn_foreach_instruction(b
, words
, word_end
,
3303 vtn_handle_preamble_instruction
);
3305 if (b
->entry_point
== NULL
) {
3306 assert(!"Entry point not found");
3311 b
->shader
= nir_shader_create(NULL
, stage
, options
, NULL
);
3313 /* Set shader info defaults */
3314 b
->shader
->info
.gs
.invocations
= 1;
3316 /* Parse execution modes */
3317 vtn_foreach_execution_mode(b
, b
->entry_point
,
3318 vtn_handle_execution_mode
, NULL
);
3320 b
->specializations
= spec
;
3321 b
->num_specializations
= num_spec
;
3323 /* Handle all variable, type, and constant instructions */
3324 words
= vtn_foreach_instruction(b
, words
, word_end
,
3325 vtn_handle_variable_or_type_instruction
);
3327 vtn_build_cfg(b
, words
, word_end
);
3329 foreach_list_typed(struct vtn_function
, func
, node
, &b
->functions
) {
3330 b
->impl
= func
->impl
;
3331 b
->const_table
= _mesa_hash_table_create(b
, _mesa_hash_pointer
,
3332 _mesa_key_pointer_equal
);
3334 vtn_function_emit(b
, func
, vtn_handle_body_instruction
);
3337 assert(b
->entry_point
->value_type
== vtn_value_type_function
);
3338 nir_function
*entry_point
= b
->entry_point
->func
->impl
->function
;
3339 assert(entry_point
);