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"
35 vtn_log(struct vtn_builder
*b
, enum nir_spirv_debug_level level
,
36 size_t spirv_offset
, const char *message
)
38 if (b
->options
->debug
.func
) {
39 b
->options
->debug
.func(b
->options
->debug
.private_data
,
40 level
, spirv_offset
, message
);
44 if (level
>= NIR_SPIRV_DEBUG_LEVEL_WARNING
)
45 fprintf(stderr
, "%s\n", message
);
50 vtn_logf(struct vtn_builder
*b
, enum nir_spirv_debug_level level
,
51 size_t spirv_offset
, const char *fmt
, ...)
57 msg
= ralloc_vasprintf(NULL
, fmt
, args
);
60 vtn_log(b
, level
, spirv_offset
, msg
);
66 vtn_log_err(struct vtn_builder
*b
,
67 enum nir_spirv_debug_level level
, const char *prefix
,
68 const char *file
, unsigned line
,
69 const char *fmt
, va_list args
)
73 msg
= ralloc_strdup(NULL
, prefix
);
76 ralloc_asprintf_append(&msg
, " In file %s:%u\n", file
, line
);
79 ralloc_asprintf_append(&msg
, " ");
81 ralloc_vasprintf_append(&msg
, fmt
, args
);
83 ralloc_asprintf_append(&msg
, "\n %zu bytes into the SPIR-V binary",
87 ralloc_asprintf_append(&msg
,
88 "\n in SPIR-V source file %s, line %d, col %d",
89 b
->file
, b
->line
, b
->col
);
92 vtn_log(b
, level
, b
->spirv_offset
, msg
);
98 _vtn_warn(struct vtn_builder
*b
, const char *file
, unsigned line
,
104 vtn_log_err(b
, NIR_SPIRV_DEBUG_LEVEL_WARNING
, "SPIR-V WARNING:\n",
105 file
, line
, fmt
, args
);
110 _vtn_fail(struct vtn_builder
*b
, const char *file
, unsigned line
,
111 const char *fmt
, ...)
116 vtn_log_err(b
, NIR_SPIRV_DEBUG_LEVEL_ERROR
, "SPIR-V parsing FAILED:\n",
117 file
, line
, fmt
, args
);
120 longjmp(b
->fail_jump
, 1);
123 struct spec_constant_value
{
131 static struct vtn_ssa_value
*
132 vtn_undef_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
134 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
137 if (glsl_type_is_vector_or_scalar(type
)) {
138 unsigned num_components
= glsl_get_vector_elements(val
->type
);
139 unsigned bit_size
= glsl_get_bit_size(val
->type
);
140 val
->def
= nir_ssa_undef(&b
->nb
, num_components
, bit_size
);
142 unsigned elems
= glsl_get_length(val
->type
);
143 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
144 if (glsl_type_is_matrix(type
)) {
145 const struct glsl_type
*elem_type
=
146 glsl_vector_type(glsl_get_base_type(type
),
147 glsl_get_vector_elements(type
));
149 for (unsigned i
= 0; i
< elems
; i
++)
150 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
151 } else if (glsl_type_is_array(type
)) {
152 const struct glsl_type
*elem_type
= glsl_get_array_element(type
);
153 for (unsigned i
= 0; i
< elems
; i
++)
154 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
156 for (unsigned i
= 0; i
< elems
; i
++) {
157 const struct glsl_type
*elem_type
= glsl_get_struct_field(type
, i
);
158 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
166 static struct vtn_ssa_value
*
167 vtn_const_ssa_value(struct vtn_builder
*b
, nir_constant
*constant
,
168 const struct glsl_type
*type
)
170 struct hash_entry
*entry
= _mesa_hash_table_search(b
->const_table
, constant
);
175 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
178 switch (glsl_get_base_type(type
)) {
181 case GLSL_TYPE_INT16
:
182 case GLSL_TYPE_UINT16
:
183 case GLSL_TYPE_INT64
:
184 case GLSL_TYPE_UINT64
:
186 case GLSL_TYPE_FLOAT
:
187 case GLSL_TYPE_FLOAT16
:
188 case GLSL_TYPE_DOUBLE
: {
189 int bit_size
= glsl_get_bit_size(type
);
190 if (glsl_type_is_vector_or_scalar(type
)) {
191 unsigned num_components
= glsl_get_vector_elements(val
->type
);
192 nir_load_const_instr
*load
=
193 nir_load_const_instr_create(b
->shader
, num_components
, bit_size
);
195 load
->value
= constant
->values
[0];
197 nir_instr_insert_before_cf_list(&b
->nb
.impl
->body
, &load
->instr
);
198 val
->def
= &load
->def
;
200 assert(glsl_type_is_matrix(type
));
201 unsigned rows
= glsl_get_vector_elements(val
->type
);
202 unsigned columns
= glsl_get_matrix_columns(val
->type
);
203 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, columns
);
205 for (unsigned i
= 0; i
< columns
; i
++) {
206 struct vtn_ssa_value
*col_val
= rzalloc(b
, struct vtn_ssa_value
);
207 col_val
->type
= glsl_get_column_type(val
->type
);
208 nir_load_const_instr
*load
=
209 nir_load_const_instr_create(b
->shader
, rows
, bit_size
);
211 load
->value
= constant
->values
[i
];
213 nir_instr_insert_before_cf_list(&b
->nb
.impl
->body
, &load
->instr
);
214 col_val
->def
= &load
->def
;
216 val
->elems
[i
] = col_val
;
222 case GLSL_TYPE_ARRAY
: {
223 unsigned elems
= glsl_get_length(val
->type
);
224 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
225 const struct glsl_type
*elem_type
= glsl_get_array_element(val
->type
);
226 for (unsigned i
= 0; i
< elems
; i
++)
227 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
232 case GLSL_TYPE_STRUCT
: {
233 unsigned elems
= glsl_get_length(val
->type
);
234 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
235 for (unsigned i
= 0; i
< elems
; i
++) {
236 const struct glsl_type
*elem_type
=
237 glsl_get_struct_field(val
->type
, i
);
238 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
245 vtn_fail("bad constant type");
251 struct vtn_ssa_value
*
252 vtn_ssa_value(struct vtn_builder
*b
, uint32_t value_id
)
254 struct vtn_value
*val
= vtn_untyped_value(b
, value_id
);
255 switch (val
->value_type
) {
256 case vtn_value_type_undef
:
257 return vtn_undef_ssa_value(b
, val
->type
->type
);
259 case vtn_value_type_constant
:
260 return vtn_const_ssa_value(b
, val
->constant
, val
->type
->type
);
262 case vtn_value_type_ssa
:
265 case vtn_value_type_pointer
:
266 vtn_assert(val
->pointer
->ptr_type
&& val
->pointer
->ptr_type
->type
);
267 struct vtn_ssa_value
*ssa
=
268 vtn_create_ssa_value(b
, val
->pointer
->ptr_type
->type
);
269 ssa
->def
= vtn_pointer_to_ssa(b
, val
->pointer
);
273 vtn_fail("Invalid type for an SSA value");
278 vtn_string_literal(struct vtn_builder
*b
, const uint32_t *words
,
279 unsigned word_count
, unsigned *words_used
)
281 char *dup
= ralloc_strndup(b
, (char *)words
, word_count
* sizeof(*words
));
283 /* Ammount of space taken by the string (including the null) */
284 unsigned len
= strlen(dup
) + 1;
285 *words_used
= DIV_ROUND_UP(len
, sizeof(*words
));
291 vtn_foreach_instruction(struct vtn_builder
*b
, const uint32_t *start
,
292 const uint32_t *end
, vtn_instruction_handler handler
)
298 const uint32_t *w
= start
;
300 SpvOp opcode
= w
[0] & SpvOpCodeMask
;
301 unsigned count
= w
[0] >> SpvWordCountShift
;
302 vtn_assert(count
>= 1 && w
+ count
<= end
);
304 b
->spirv_offset
= (uint8_t *)w
- (uint8_t *)b
->spirv
;
308 break; /* Do nothing */
311 b
->file
= vtn_value(b
, w
[1], vtn_value_type_string
)->str
;
323 if (!handler(b
, opcode
, w
, count
))
341 vtn_handle_extension(struct vtn_builder
*b
, SpvOp opcode
,
342 const uint32_t *w
, unsigned count
)
345 case SpvOpExtInstImport
: {
346 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_extension
);
347 if (strcmp((const char *)&w
[2], "GLSL.std.450") == 0) {
348 val
->ext_handler
= vtn_handle_glsl450_instruction
;
350 vtn_fail("Unsupported extension");
356 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
357 bool handled
= val
->ext_handler(b
, w
[4], w
, count
);
363 vtn_fail("Unhandled opcode");
368 _foreach_decoration_helper(struct vtn_builder
*b
,
369 struct vtn_value
*base_value
,
371 struct vtn_value
*value
,
372 vtn_decoration_foreach_cb cb
, void *data
)
374 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
376 if (dec
->scope
== VTN_DEC_DECORATION
) {
377 member
= parent_member
;
378 } else if (dec
->scope
>= VTN_DEC_STRUCT_MEMBER0
) {
379 vtn_assert(parent_member
== -1);
380 member
= dec
->scope
- VTN_DEC_STRUCT_MEMBER0
;
382 /* Not a decoration */
387 vtn_assert(dec
->group
->value_type
== vtn_value_type_decoration_group
);
388 _foreach_decoration_helper(b
, base_value
, member
, dec
->group
,
391 cb(b
, base_value
, member
, dec
, data
);
396 /** Iterates (recursively if needed) over all of the decorations on a value
398 * This function iterates over all of the decorations applied to a given
399 * value. If it encounters a decoration group, it recurses into the group
400 * and iterates over all of those decorations as well.
403 vtn_foreach_decoration(struct vtn_builder
*b
, struct vtn_value
*value
,
404 vtn_decoration_foreach_cb cb
, void *data
)
406 _foreach_decoration_helper(b
, value
, -1, value
, cb
, data
);
410 vtn_foreach_execution_mode(struct vtn_builder
*b
, struct vtn_value
*value
,
411 vtn_execution_mode_foreach_cb cb
, void *data
)
413 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
414 if (dec
->scope
!= VTN_DEC_EXECUTION_MODE
)
417 vtn_assert(dec
->group
== NULL
);
418 cb(b
, value
, dec
, data
);
423 vtn_handle_decoration(struct vtn_builder
*b
, SpvOp opcode
,
424 const uint32_t *w
, unsigned count
)
426 const uint32_t *w_end
= w
+ count
;
427 const uint32_t target
= w
[1];
431 case SpvOpDecorationGroup
:
432 vtn_push_value(b
, target
, vtn_value_type_decoration_group
);
436 case SpvOpMemberDecorate
:
437 case SpvOpExecutionMode
: {
438 struct vtn_value
*val
= &b
->values
[target
];
440 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
443 dec
->scope
= VTN_DEC_DECORATION
;
445 case SpvOpMemberDecorate
:
446 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(w
++);
448 case SpvOpExecutionMode
:
449 dec
->scope
= VTN_DEC_EXECUTION_MODE
;
452 vtn_fail("Invalid decoration opcode");
454 dec
->decoration
= *(w
++);
457 /* Link into the list */
458 dec
->next
= val
->decoration
;
459 val
->decoration
= dec
;
463 case SpvOpGroupMemberDecorate
:
464 case SpvOpGroupDecorate
: {
465 struct vtn_value
*group
=
466 vtn_value(b
, target
, vtn_value_type_decoration_group
);
468 for (; w
< w_end
; w
++) {
469 struct vtn_value
*val
= vtn_untyped_value(b
, *w
);
470 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
473 if (opcode
== SpvOpGroupDecorate
) {
474 dec
->scope
= VTN_DEC_DECORATION
;
476 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(++w
);
479 /* Link into the list */
480 dec
->next
= val
->decoration
;
481 val
->decoration
= dec
;
487 vtn_fail("Unhandled opcode");
491 struct member_decoration_ctx
{
493 struct glsl_struct_field
*fields
;
494 struct vtn_type
*type
;
497 /* does a shallow copy of a vtn_type */
499 static struct vtn_type
*
500 vtn_type_copy(struct vtn_builder
*b
, struct vtn_type
*src
)
502 struct vtn_type
*dest
= ralloc(b
, struct vtn_type
);
505 switch (src
->base_type
) {
506 case vtn_base_type_void
:
507 case vtn_base_type_scalar
:
508 case vtn_base_type_vector
:
509 case vtn_base_type_matrix
:
510 case vtn_base_type_array
:
511 case vtn_base_type_pointer
:
512 case vtn_base_type_image
:
513 case vtn_base_type_sampler
:
514 /* Nothing more to do */
517 case vtn_base_type_struct
:
518 dest
->members
= ralloc_array(b
, struct vtn_type
*, src
->length
);
519 memcpy(dest
->members
, src
->members
,
520 src
->length
* sizeof(src
->members
[0]));
522 dest
->offsets
= ralloc_array(b
, unsigned, src
->length
);
523 memcpy(dest
->offsets
, src
->offsets
,
524 src
->length
* sizeof(src
->offsets
[0]));
527 case vtn_base_type_function
:
528 dest
->params
= ralloc_array(b
, struct vtn_type
*, src
->length
);
529 memcpy(dest
->params
, src
->params
, src
->length
* sizeof(src
->params
[0]));
536 static struct vtn_type
*
537 mutable_matrix_member(struct vtn_builder
*b
, struct vtn_type
*type
, int member
)
539 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
540 type
= type
->members
[member
];
542 /* We may have an array of matrices.... Oh, joy! */
543 while (glsl_type_is_array(type
->type
)) {
544 type
->array_element
= vtn_type_copy(b
, type
->array_element
);
545 type
= type
->array_element
;
548 vtn_assert(glsl_type_is_matrix(type
->type
));
554 struct_member_decoration_cb(struct vtn_builder
*b
,
555 struct vtn_value
*val
, int member
,
556 const struct vtn_decoration
*dec
, void *void_ctx
)
558 struct member_decoration_ctx
*ctx
= void_ctx
;
563 vtn_assert(member
< ctx
->num_fields
);
565 switch (dec
->decoration
) {
566 case SpvDecorationNonWritable
:
567 case SpvDecorationNonReadable
:
568 case SpvDecorationRelaxedPrecision
:
569 case SpvDecorationVolatile
:
570 case SpvDecorationCoherent
:
571 case SpvDecorationUniform
:
572 break; /* FIXME: Do nothing with this for now. */
573 case SpvDecorationNoPerspective
:
574 ctx
->fields
[member
].interpolation
= INTERP_MODE_NOPERSPECTIVE
;
576 case SpvDecorationFlat
:
577 ctx
->fields
[member
].interpolation
= INTERP_MODE_FLAT
;
579 case SpvDecorationCentroid
:
580 ctx
->fields
[member
].centroid
= true;
582 case SpvDecorationSample
:
583 ctx
->fields
[member
].sample
= true;
585 case SpvDecorationStream
:
586 /* Vulkan only allows one GS stream */
587 vtn_assert(dec
->literals
[0] == 0);
589 case SpvDecorationLocation
:
590 ctx
->fields
[member
].location
= dec
->literals
[0];
592 case SpvDecorationComponent
:
593 break; /* FIXME: What should we do with these? */
594 case SpvDecorationBuiltIn
:
595 ctx
->type
->members
[member
] = vtn_type_copy(b
, ctx
->type
->members
[member
]);
596 ctx
->type
->members
[member
]->is_builtin
= true;
597 ctx
->type
->members
[member
]->builtin
= dec
->literals
[0];
598 ctx
->type
->builtin_block
= true;
600 case SpvDecorationOffset
:
601 ctx
->type
->offsets
[member
] = dec
->literals
[0];
603 case SpvDecorationMatrixStride
:
604 /* Handled as a second pass */
606 case SpvDecorationColMajor
:
607 break; /* Nothing to do here. Column-major is the default. */
608 case SpvDecorationRowMajor
:
609 mutable_matrix_member(b
, ctx
->type
, member
)->row_major
= true;
612 case SpvDecorationPatch
:
615 case SpvDecorationSpecId
:
616 case SpvDecorationBlock
:
617 case SpvDecorationBufferBlock
:
618 case SpvDecorationArrayStride
:
619 case SpvDecorationGLSLShared
:
620 case SpvDecorationGLSLPacked
:
621 case SpvDecorationInvariant
:
622 case SpvDecorationRestrict
:
623 case SpvDecorationAliased
:
624 case SpvDecorationConstant
:
625 case SpvDecorationIndex
:
626 case SpvDecorationBinding
:
627 case SpvDecorationDescriptorSet
:
628 case SpvDecorationLinkageAttributes
:
629 case SpvDecorationNoContraction
:
630 case SpvDecorationInputAttachmentIndex
:
631 vtn_warn("Decoration not allowed on struct members: %s",
632 spirv_decoration_to_string(dec
->decoration
));
635 case SpvDecorationXfbBuffer
:
636 case SpvDecorationXfbStride
:
637 vtn_warn("Vulkan does not have transform feedback");
640 case SpvDecorationCPacked
:
641 case SpvDecorationSaturatedConversion
:
642 case SpvDecorationFuncParamAttr
:
643 case SpvDecorationFPRoundingMode
:
644 case SpvDecorationFPFastMathMode
:
645 case SpvDecorationAlignment
:
646 vtn_warn("Decoration only allowed for CL-style kernels: %s",
647 spirv_decoration_to_string(dec
->decoration
));
651 vtn_fail("Unhandled decoration");
655 /* Matrix strides are handled as a separate pass because we need to know
656 * whether the matrix is row-major or not first.
659 struct_member_matrix_stride_cb(struct vtn_builder
*b
,
660 struct vtn_value
*val
, int member
,
661 const struct vtn_decoration
*dec
,
664 if (dec
->decoration
!= SpvDecorationMatrixStride
)
666 vtn_assert(member
>= 0);
668 struct member_decoration_ctx
*ctx
= void_ctx
;
670 struct vtn_type
*mat_type
= mutable_matrix_member(b
, ctx
->type
, member
);
671 if (mat_type
->row_major
) {
672 mat_type
->array_element
= vtn_type_copy(b
, mat_type
->array_element
);
673 mat_type
->stride
= mat_type
->array_element
->stride
;
674 mat_type
->array_element
->stride
= dec
->literals
[0];
676 vtn_assert(mat_type
->array_element
->stride
> 0);
677 mat_type
->stride
= dec
->literals
[0];
682 type_decoration_cb(struct vtn_builder
*b
,
683 struct vtn_value
*val
, int member
,
684 const struct vtn_decoration
*dec
, void *ctx
)
686 struct vtn_type
*type
= val
->type
;
691 switch (dec
->decoration
) {
692 case SpvDecorationArrayStride
:
693 vtn_assert(type
->base_type
== vtn_base_type_matrix
||
694 type
->base_type
== vtn_base_type_array
||
695 type
->base_type
== vtn_base_type_pointer
);
696 type
->stride
= dec
->literals
[0];
698 case SpvDecorationBlock
:
699 vtn_assert(type
->base_type
== vtn_base_type_struct
);
702 case SpvDecorationBufferBlock
:
703 vtn_assert(type
->base_type
== vtn_base_type_struct
);
704 type
->buffer_block
= true;
706 case SpvDecorationGLSLShared
:
707 case SpvDecorationGLSLPacked
:
708 /* Ignore these, since we get explicit offsets anyways */
711 case SpvDecorationRowMajor
:
712 case SpvDecorationColMajor
:
713 case SpvDecorationMatrixStride
:
714 case SpvDecorationBuiltIn
:
715 case SpvDecorationNoPerspective
:
716 case SpvDecorationFlat
:
717 case SpvDecorationPatch
:
718 case SpvDecorationCentroid
:
719 case SpvDecorationSample
:
720 case SpvDecorationVolatile
:
721 case SpvDecorationCoherent
:
722 case SpvDecorationNonWritable
:
723 case SpvDecorationNonReadable
:
724 case SpvDecorationUniform
:
725 case SpvDecorationStream
:
726 case SpvDecorationLocation
:
727 case SpvDecorationComponent
:
728 case SpvDecorationOffset
:
729 case SpvDecorationXfbBuffer
:
730 case SpvDecorationXfbStride
:
731 vtn_warn("Decoration only allowed for struct members: %s",
732 spirv_decoration_to_string(dec
->decoration
));
735 case SpvDecorationRelaxedPrecision
:
736 case SpvDecorationSpecId
:
737 case SpvDecorationInvariant
:
738 case SpvDecorationRestrict
:
739 case SpvDecorationAliased
:
740 case SpvDecorationConstant
:
741 case SpvDecorationIndex
:
742 case SpvDecorationBinding
:
743 case SpvDecorationDescriptorSet
:
744 case SpvDecorationLinkageAttributes
:
745 case SpvDecorationNoContraction
:
746 case SpvDecorationInputAttachmentIndex
:
747 vtn_warn("Decoration not allowed on types: %s",
748 spirv_decoration_to_string(dec
->decoration
));
751 case SpvDecorationCPacked
:
752 case SpvDecorationSaturatedConversion
:
753 case SpvDecorationFuncParamAttr
:
754 case SpvDecorationFPRoundingMode
:
755 case SpvDecorationFPFastMathMode
:
756 case SpvDecorationAlignment
:
757 vtn_warn("Decoration only allowed for CL-style kernels: %s",
758 spirv_decoration_to_string(dec
->decoration
));
762 vtn_fail("Unhandled decoration");
767 translate_image_format(struct vtn_builder
*b
, SpvImageFormat format
)
770 case SpvImageFormatUnknown
: return 0; /* GL_NONE */
771 case SpvImageFormatRgba32f
: return 0x8814; /* GL_RGBA32F */
772 case SpvImageFormatRgba16f
: return 0x881A; /* GL_RGBA16F */
773 case SpvImageFormatR32f
: return 0x822E; /* GL_R32F */
774 case SpvImageFormatRgba8
: return 0x8058; /* GL_RGBA8 */
775 case SpvImageFormatRgba8Snorm
: return 0x8F97; /* GL_RGBA8_SNORM */
776 case SpvImageFormatRg32f
: return 0x8230; /* GL_RG32F */
777 case SpvImageFormatRg16f
: return 0x822F; /* GL_RG16F */
778 case SpvImageFormatR11fG11fB10f
: return 0x8C3A; /* GL_R11F_G11F_B10F */
779 case SpvImageFormatR16f
: return 0x822D; /* GL_R16F */
780 case SpvImageFormatRgba16
: return 0x805B; /* GL_RGBA16 */
781 case SpvImageFormatRgb10A2
: return 0x8059; /* GL_RGB10_A2 */
782 case SpvImageFormatRg16
: return 0x822C; /* GL_RG16 */
783 case SpvImageFormatRg8
: return 0x822B; /* GL_RG8 */
784 case SpvImageFormatR16
: return 0x822A; /* GL_R16 */
785 case SpvImageFormatR8
: return 0x8229; /* GL_R8 */
786 case SpvImageFormatRgba16Snorm
: return 0x8F9B; /* GL_RGBA16_SNORM */
787 case SpvImageFormatRg16Snorm
: return 0x8F99; /* GL_RG16_SNORM */
788 case SpvImageFormatRg8Snorm
: return 0x8F95; /* GL_RG8_SNORM */
789 case SpvImageFormatR16Snorm
: return 0x8F98; /* GL_R16_SNORM */
790 case SpvImageFormatR8Snorm
: return 0x8F94; /* GL_R8_SNORM */
791 case SpvImageFormatRgba32i
: return 0x8D82; /* GL_RGBA32I */
792 case SpvImageFormatRgba16i
: return 0x8D88; /* GL_RGBA16I */
793 case SpvImageFormatRgba8i
: return 0x8D8E; /* GL_RGBA8I */
794 case SpvImageFormatR32i
: return 0x8235; /* GL_R32I */
795 case SpvImageFormatRg32i
: return 0x823B; /* GL_RG32I */
796 case SpvImageFormatRg16i
: return 0x8239; /* GL_RG16I */
797 case SpvImageFormatRg8i
: return 0x8237; /* GL_RG8I */
798 case SpvImageFormatR16i
: return 0x8233; /* GL_R16I */
799 case SpvImageFormatR8i
: return 0x8231; /* GL_R8I */
800 case SpvImageFormatRgba32ui
: return 0x8D70; /* GL_RGBA32UI */
801 case SpvImageFormatRgba16ui
: return 0x8D76; /* GL_RGBA16UI */
802 case SpvImageFormatRgba8ui
: return 0x8D7C; /* GL_RGBA8UI */
803 case SpvImageFormatR32ui
: return 0x8236; /* GL_R32UI */
804 case SpvImageFormatRgb10a2ui
: return 0x906F; /* GL_RGB10_A2UI */
805 case SpvImageFormatRg32ui
: return 0x823C; /* GL_RG32UI */
806 case SpvImageFormatRg16ui
: return 0x823A; /* GL_RG16UI */
807 case SpvImageFormatRg8ui
: return 0x8238; /* GL_RG8UI */
808 case SpvImageFormatR16ui
: return 0x8234; /* GL_R16UI */
809 case SpvImageFormatR8ui
: return 0x8232; /* GL_R8UI */
811 vtn_fail("Invalid image format");
815 static struct vtn_type
*
816 vtn_type_layout_std430(struct vtn_builder
*b
, struct vtn_type
*type
,
817 uint32_t *size_out
, uint32_t *align_out
)
819 switch (type
->base_type
) {
820 case vtn_base_type_scalar
: {
821 uint32_t comp_size
= glsl_get_bit_size(type
->type
) / 8;
822 *size_out
= comp_size
;
823 *align_out
= comp_size
;
827 case vtn_base_type_vector
: {
828 uint32_t comp_size
= glsl_get_bit_size(type
->type
) / 8;
829 assert(type
->length
> 0 && type
->length
<= 4);
830 unsigned align_comps
= type
->length
== 3 ? 4 : type
->length
;
831 *size_out
= comp_size
* type
->length
,
832 *align_out
= comp_size
* align_comps
;
836 case vtn_base_type_matrix
:
837 case vtn_base_type_array
: {
838 /* We're going to add an array stride */
839 type
= vtn_type_copy(b
, type
);
840 uint32_t elem_size
, elem_align
;
841 type
->array_element
= vtn_type_layout_std430(b
, type
->array_element
,
842 &elem_size
, &elem_align
);
843 type
->stride
= vtn_align_u32(elem_size
, elem_align
);
844 *size_out
= type
->stride
* type
->length
;
845 *align_out
= elem_align
;
849 case vtn_base_type_struct
: {
850 /* We're going to add member offsets */
851 type
= vtn_type_copy(b
, type
);
854 for (unsigned i
= 0; i
< type
->length
; i
++) {
855 uint32_t mem_size
, mem_align
;
856 type
->members
[i
] = vtn_type_layout_std430(b
, type
->members
[i
],
857 &mem_size
, &mem_align
);
858 offset
= vtn_align_u32(offset
, mem_align
);
859 type
->offsets
[i
] = offset
;
861 align
= MAX2(align
, mem_align
);
869 unreachable("Invalid SPIR-V type for std430");
874 vtn_handle_type(struct vtn_builder
*b
, SpvOp opcode
,
875 const uint32_t *w
, unsigned count
)
877 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_type
);
879 val
->type
= rzalloc(b
, struct vtn_type
);
880 val
->type
->val
= val
;
884 val
->type
->base_type
= vtn_base_type_void
;
885 val
->type
->type
= glsl_void_type();
888 val
->type
->base_type
= vtn_base_type_scalar
;
889 val
->type
->type
= glsl_bool_type();
890 val
->type
->length
= 1;
894 const bool signedness
= w
[3];
895 val
->type
->base_type
= vtn_base_type_scalar
;
898 val
->type
->type
= (signedness
? glsl_int64_t_type() : glsl_uint64_t_type());
901 val
->type
->type
= (signedness
? glsl_int_type() : glsl_uint_type());
904 val
->type
->type
= (signedness
? glsl_int16_t_type() : glsl_uint16_t_type());
907 vtn_fail("Invalid int bit size");
909 val
->type
->length
= 1;
913 case SpvOpTypeFloat
: {
915 val
->type
->base_type
= vtn_base_type_scalar
;
918 val
->type
->type
= glsl_float16_t_type();
921 val
->type
->type
= glsl_float_type();
924 val
->type
->type
= glsl_double_type();
927 vtn_fail("Invalid float bit size");
929 val
->type
->length
= 1;
933 case SpvOpTypeVector
: {
934 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
935 unsigned elems
= w
[3];
937 vtn_fail_if(base
->base_type
!= vtn_base_type_scalar
,
938 "Base type for OpTypeVector must be a scalar");
939 vtn_fail_if(elems
< 2 || elems
> 4,
940 "Invalid component count for OpTypeVector");
942 val
->type
->base_type
= vtn_base_type_vector
;
943 val
->type
->type
= glsl_vector_type(glsl_get_base_type(base
->type
), elems
);
944 val
->type
->length
= elems
;
945 val
->type
->stride
= glsl_get_bit_size(base
->type
) / 8;
946 val
->type
->array_element
= base
;
950 case SpvOpTypeMatrix
: {
951 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
952 unsigned columns
= w
[3];
954 vtn_fail_if(base
->base_type
!= vtn_base_type_vector
,
955 "Base type for OpTypeMatrix must be a vector");
956 vtn_fail_if(columns
< 2 || columns
> 4,
957 "Invalid column count for OpTypeMatrix");
959 val
->type
->base_type
= vtn_base_type_matrix
;
960 val
->type
->type
= glsl_matrix_type(glsl_get_base_type(base
->type
),
961 glsl_get_vector_elements(base
->type
),
963 vtn_fail_if(glsl_type_is_error(val
->type
->type
),
964 "Unsupported base type for OpTypeMatrix");
965 assert(!glsl_type_is_error(val
->type
->type
));
966 val
->type
->length
= columns
;
967 val
->type
->array_element
= base
;
968 val
->type
->row_major
= false;
969 val
->type
->stride
= 0;
973 case SpvOpTypeRuntimeArray
:
974 case SpvOpTypeArray
: {
975 struct vtn_type
*array_element
=
976 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
978 if (opcode
== SpvOpTypeRuntimeArray
) {
979 /* A length of 0 is used to denote unsized arrays */
980 val
->type
->length
= 0;
983 vtn_value(b
, w
[3], vtn_value_type_constant
)->constant
->values
[0].u32
[0];
986 val
->type
->base_type
= vtn_base_type_array
;
987 val
->type
->type
= glsl_array_type(array_element
->type
, val
->type
->length
);
988 val
->type
->array_element
= array_element
;
989 val
->type
->stride
= 0;
993 case SpvOpTypeStruct
: {
994 unsigned num_fields
= count
- 2;
995 val
->type
->base_type
= vtn_base_type_struct
;
996 val
->type
->length
= num_fields
;
997 val
->type
->members
= ralloc_array(b
, struct vtn_type
*, num_fields
);
998 val
->type
->offsets
= ralloc_array(b
, unsigned, num_fields
);
1000 NIR_VLA(struct glsl_struct_field
, fields
, count
);
1001 for (unsigned i
= 0; i
< num_fields
; i
++) {
1002 val
->type
->members
[i
] =
1003 vtn_value(b
, w
[i
+ 2], vtn_value_type_type
)->type
;
1004 fields
[i
] = (struct glsl_struct_field
) {
1005 .type
= val
->type
->members
[i
]->type
,
1006 .name
= ralloc_asprintf(b
, "field%d", i
),
1011 struct member_decoration_ctx ctx
= {
1012 .num_fields
= num_fields
,
1017 vtn_foreach_decoration(b
, val
, struct_member_decoration_cb
, &ctx
);
1018 vtn_foreach_decoration(b
, val
, struct_member_matrix_stride_cb
, &ctx
);
1020 const char *name
= val
->name
? val
->name
: "struct";
1022 val
->type
->type
= glsl_struct_type(fields
, num_fields
, name
);
1026 case SpvOpTypeFunction
: {
1027 val
->type
->base_type
= vtn_base_type_function
;
1028 val
->type
->type
= NULL
;
1030 val
->type
->return_type
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1032 const unsigned num_params
= count
- 3;
1033 val
->type
->length
= num_params
;
1034 val
->type
->params
= ralloc_array(b
, struct vtn_type
*, num_params
);
1035 for (unsigned i
= 0; i
< count
- 3; i
++) {
1036 val
->type
->params
[i
] =
1037 vtn_value(b
, w
[i
+ 3], vtn_value_type_type
)->type
;
1042 case SpvOpTypePointer
: {
1043 SpvStorageClass storage_class
= w
[2];
1044 struct vtn_type
*deref_type
=
1045 vtn_value(b
, w
[3], vtn_value_type_type
)->type
;
1047 val
->type
->base_type
= vtn_base_type_pointer
;
1048 val
->type
->storage_class
= storage_class
;
1049 val
->type
->deref
= deref_type
;
1051 if (storage_class
== SpvStorageClassUniform
||
1052 storage_class
== SpvStorageClassStorageBuffer
) {
1053 /* These can actually be stored to nir_variables and used as SSA
1054 * values so they need a real glsl_type.
1056 val
->type
->type
= glsl_vector_type(GLSL_TYPE_UINT
, 2);
1059 if (storage_class
== SpvStorageClassWorkgroup
&&
1060 b
->options
->lower_workgroup_access_to_offsets
) {
1061 uint32_t size
, align
;
1062 val
->type
->deref
= vtn_type_layout_std430(b
, val
->type
->deref
,
1064 val
->type
->length
= size
;
1065 val
->type
->align
= align
;
1066 /* These can actually be stored to nir_variables and used as SSA
1067 * values so they need a real glsl_type.
1069 val
->type
->type
= glsl_uint_type();
1074 case SpvOpTypeImage
: {
1075 val
->type
->base_type
= vtn_base_type_image
;
1077 const struct glsl_type
*sampled_type
=
1078 vtn_value(b
, w
[2], vtn_value_type_type
)->type
->type
;
1080 vtn_assert(glsl_type_is_vector_or_scalar(sampled_type
));
1082 enum glsl_sampler_dim dim
;
1083 switch ((SpvDim
)w
[3]) {
1084 case SpvDim1D
: dim
= GLSL_SAMPLER_DIM_1D
; break;
1085 case SpvDim2D
: dim
= GLSL_SAMPLER_DIM_2D
; break;
1086 case SpvDim3D
: dim
= GLSL_SAMPLER_DIM_3D
; break;
1087 case SpvDimCube
: dim
= GLSL_SAMPLER_DIM_CUBE
; break;
1088 case SpvDimRect
: dim
= GLSL_SAMPLER_DIM_RECT
; break;
1089 case SpvDimBuffer
: dim
= GLSL_SAMPLER_DIM_BUF
; break;
1090 case SpvDimSubpassData
: dim
= GLSL_SAMPLER_DIM_SUBPASS
; break;
1092 vtn_fail("Invalid SPIR-V Sampler dimension");
1095 bool is_shadow
= w
[4];
1096 bool is_array
= w
[5];
1097 bool multisampled
= w
[6];
1098 unsigned sampled
= w
[7];
1099 SpvImageFormat format
= w
[8];
1102 val
->type
->access_qualifier
= w
[9];
1104 val
->type
->access_qualifier
= SpvAccessQualifierReadWrite
;
1107 if (dim
== GLSL_SAMPLER_DIM_2D
)
1108 dim
= GLSL_SAMPLER_DIM_MS
;
1109 else if (dim
== GLSL_SAMPLER_DIM_SUBPASS
)
1110 dim
= GLSL_SAMPLER_DIM_SUBPASS_MS
;
1112 vtn_fail("Unsupported multisampled image type");
1115 val
->type
->image_format
= translate_image_format(b
, format
);
1118 val
->type
->sampled
= true;
1119 val
->type
->type
= glsl_sampler_type(dim
, is_shadow
, is_array
,
1120 glsl_get_base_type(sampled_type
));
1121 } else if (sampled
== 2) {
1122 vtn_assert(!is_shadow
);
1123 val
->type
->sampled
= false;
1124 val
->type
->type
= glsl_image_type(dim
, is_array
,
1125 glsl_get_base_type(sampled_type
));
1127 vtn_fail("We need to know if the image will be sampled");
1132 case SpvOpTypeSampledImage
:
1133 val
->type
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1136 case SpvOpTypeSampler
:
1137 /* The actual sampler type here doesn't really matter. It gets
1138 * thrown away the moment you combine it with an image. What really
1139 * matters is that it's a sampler type as opposed to an integer type
1140 * so the backend knows what to do.
1142 val
->type
->base_type
= vtn_base_type_sampler
;
1143 val
->type
->type
= glsl_bare_sampler_type();
1146 case SpvOpTypeOpaque
:
1147 case SpvOpTypeEvent
:
1148 case SpvOpTypeDeviceEvent
:
1149 case SpvOpTypeReserveId
:
1150 case SpvOpTypeQueue
:
1153 vtn_fail("Unhandled opcode");
1156 vtn_foreach_decoration(b
, val
, type_decoration_cb
, NULL
);
1159 static nir_constant
*
1160 vtn_null_constant(struct vtn_builder
*b
, const struct glsl_type
*type
)
1162 nir_constant
*c
= rzalloc(b
, nir_constant
);
1164 /* For pointers and other typeless things, we have to return something but
1165 * it doesn't matter what.
1170 switch (glsl_get_base_type(type
)) {
1172 case GLSL_TYPE_UINT
:
1173 case GLSL_TYPE_INT16
:
1174 case GLSL_TYPE_UINT16
:
1175 case GLSL_TYPE_INT64
:
1176 case GLSL_TYPE_UINT64
:
1177 case GLSL_TYPE_BOOL
:
1178 case GLSL_TYPE_FLOAT
:
1179 case GLSL_TYPE_FLOAT16
:
1180 case GLSL_TYPE_DOUBLE
:
1181 /* Nothing to do here. It's already initialized to zero */
1184 case GLSL_TYPE_ARRAY
:
1185 vtn_assert(glsl_get_length(type
) > 0);
1186 c
->num_elements
= glsl_get_length(type
);
1187 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1189 c
->elements
[0] = vtn_null_constant(b
, glsl_get_array_element(type
));
1190 for (unsigned i
= 1; i
< c
->num_elements
; i
++)
1191 c
->elements
[i
] = c
->elements
[0];
1194 case GLSL_TYPE_STRUCT
:
1195 c
->num_elements
= glsl_get_length(type
);
1196 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1198 for (unsigned i
= 0; i
< c
->num_elements
; i
++) {
1199 c
->elements
[i
] = vtn_null_constant(b
, glsl_get_struct_field(type
, i
));
1204 vtn_fail("Invalid type for null constant");
1211 spec_constant_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*v
,
1212 int member
, const struct vtn_decoration
*dec
,
1215 vtn_assert(member
== -1);
1216 if (dec
->decoration
!= SpvDecorationSpecId
)
1219 struct spec_constant_value
*const_value
= data
;
1221 for (unsigned i
= 0; i
< b
->num_specializations
; i
++) {
1222 if (b
->specializations
[i
].id
== dec
->literals
[0]) {
1223 if (const_value
->is_double
)
1224 const_value
->data64
= b
->specializations
[i
].data64
;
1226 const_value
->data32
= b
->specializations
[i
].data32
;
1233 get_specialization(struct vtn_builder
*b
, struct vtn_value
*val
,
1234 uint32_t const_value
)
1236 struct spec_constant_value data
;
1237 data
.is_double
= false;
1238 data
.data32
= const_value
;
1239 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1244 get_specialization64(struct vtn_builder
*b
, struct vtn_value
*val
,
1245 uint64_t const_value
)
1247 struct spec_constant_value data
;
1248 data
.is_double
= true;
1249 data
.data64
= const_value
;
1250 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1255 handle_workgroup_size_decoration_cb(struct vtn_builder
*b
,
1256 struct vtn_value
*val
,
1258 const struct vtn_decoration
*dec
,
1261 vtn_assert(member
== -1);
1262 if (dec
->decoration
!= SpvDecorationBuiltIn
||
1263 dec
->literals
[0] != SpvBuiltInWorkgroupSize
)
1266 vtn_assert(val
->type
->type
== glsl_vector_type(GLSL_TYPE_UINT
, 3));
1268 b
->shader
->info
.cs
.local_size
[0] = val
->constant
->values
[0].u32
[0];
1269 b
->shader
->info
.cs
.local_size
[1] = val
->constant
->values
[0].u32
[1];
1270 b
->shader
->info
.cs
.local_size
[2] = val
->constant
->values
[0].u32
[2];
1274 vtn_handle_constant(struct vtn_builder
*b
, SpvOp opcode
,
1275 const uint32_t *w
, unsigned count
)
1277 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_constant
);
1278 val
->constant
= rzalloc(b
, nir_constant
);
1280 case SpvOpConstantTrue
:
1281 vtn_assert(val
->type
->type
== glsl_bool_type());
1282 val
->constant
->values
[0].u32
[0] = NIR_TRUE
;
1284 case SpvOpConstantFalse
:
1285 vtn_assert(val
->type
->type
== glsl_bool_type());
1286 val
->constant
->values
[0].u32
[0] = NIR_FALSE
;
1289 case SpvOpSpecConstantTrue
:
1290 case SpvOpSpecConstantFalse
: {
1291 vtn_assert(val
->type
->type
== glsl_bool_type());
1293 get_specialization(b
, val
, (opcode
== SpvOpSpecConstantTrue
));
1294 val
->constant
->values
[0].u32
[0] = int_val
? NIR_TRUE
: NIR_FALSE
;
1298 case SpvOpConstant
: {
1299 vtn_assert(glsl_type_is_scalar(val
->type
->type
));
1300 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1303 val
->constant
->values
->u64
[0] = vtn_u64_literal(&w
[3]);
1306 val
->constant
->values
->u32
[0] = w
[3];
1309 val
->constant
->values
->u16
[0] = w
[3];
1312 vtn_fail("Unsupported SpvOpConstant bit size");
1316 case SpvOpSpecConstant
: {
1317 vtn_assert(glsl_type_is_scalar(val
->type
->type
));
1318 val
->constant
->values
[0].u32
[0] = get_specialization(b
, val
, w
[3]);
1319 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1322 val
->constant
->values
[0].u64
[0] =
1323 get_specialization64(b
, val
, vtn_u64_literal(&w
[3]));
1326 val
->constant
->values
[0].u32
[0] = get_specialization(b
, val
, w
[3]);
1329 val
->constant
->values
[0].u16
[0] = get_specialization(b
, val
, w
[3]);
1332 vtn_fail("Unsupported SpvOpSpecConstant bit size");
1336 case SpvOpSpecConstantComposite
:
1337 case SpvOpConstantComposite
: {
1338 unsigned elem_count
= count
- 3;
1339 nir_constant
**elems
= ralloc_array(b
, nir_constant
*, elem_count
);
1340 for (unsigned i
= 0; i
< elem_count
; i
++)
1341 elems
[i
] = vtn_value(b
, w
[i
+ 3], vtn_value_type_constant
)->constant
;
1343 switch (glsl_get_base_type(val
->type
->type
)) {
1344 case GLSL_TYPE_UINT
:
1346 case GLSL_TYPE_UINT16
:
1347 case GLSL_TYPE_INT16
:
1348 case GLSL_TYPE_UINT64
:
1349 case GLSL_TYPE_INT64
:
1350 case GLSL_TYPE_FLOAT
:
1351 case GLSL_TYPE_FLOAT16
:
1352 case GLSL_TYPE_BOOL
:
1353 case GLSL_TYPE_DOUBLE
: {
1354 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1355 if (glsl_type_is_matrix(val
->type
->type
)) {
1356 vtn_assert(glsl_get_matrix_columns(val
->type
->type
) == elem_count
);
1357 for (unsigned i
= 0; i
< elem_count
; i
++)
1358 val
->constant
->values
[i
] = elems
[i
]->values
[0];
1360 vtn_assert(glsl_type_is_vector(val
->type
->type
));
1361 vtn_assert(glsl_get_vector_elements(val
->type
->type
) == elem_count
);
1362 for (unsigned i
= 0; i
< elem_count
; i
++) {
1365 val
->constant
->values
[0].u64
[i
] = elems
[i
]->values
[0].u64
[0];
1368 val
->constant
->values
[0].u32
[i
] = elems
[i
]->values
[0].u32
[0];
1371 val
->constant
->values
[0].u16
[i
] = elems
[i
]->values
[0].u16
[0];
1374 vtn_fail("Invalid SpvOpConstantComposite bit size");
1381 case GLSL_TYPE_STRUCT
:
1382 case GLSL_TYPE_ARRAY
:
1383 ralloc_steal(val
->constant
, elems
);
1384 val
->constant
->num_elements
= elem_count
;
1385 val
->constant
->elements
= elems
;
1389 vtn_fail("Unsupported type for constants");
1394 case SpvOpSpecConstantOp
: {
1395 SpvOp opcode
= get_specialization(b
, val
, w
[3]);
1397 case SpvOpVectorShuffle
: {
1398 struct vtn_value
*v0
= &b
->values
[w
[4]];
1399 struct vtn_value
*v1
= &b
->values
[w
[5]];
1401 vtn_assert(v0
->value_type
== vtn_value_type_constant
||
1402 v0
->value_type
== vtn_value_type_undef
);
1403 vtn_assert(v1
->value_type
== vtn_value_type_constant
||
1404 v1
->value_type
== vtn_value_type_undef
);
1406 unsigned len0
= glsl_get_vector_elements(v0
->type
->type
);
1407 unsigned len1
= glsl_get_vector_elements(v1
->type
->type
);
1409 vtn_assert(len0
+ len1
< 16);
1411 unsigned bit_size
= glsl_get_bit_size(val
->type
->type
);
1412 unsigned bit_size0
= glsl_get_bit_size(v0
->type
->type
);
1413 unsigned bit_size1
= glsl_get_bit_size(v1
->type
->type
);
1415 vtn_assert(bit_size
== bit_size0
&& bit_size
== bit_size1
);
1416 (void)bit_size0
; (void)bit_size1
;
1418 if (bit_size
== 64) {
1420 if (v0
->value_type
== vtn_value_type_constant
) {
1421 for (unsigned i
= 0; i
< len0
; i
++)
1422 u64
[i
] = v0
->constant
->values
[0].u64
[i
];
1424 if (v1
->value_type
== vtn_value_type_constant
) {
1425 for (unsigned i
= 0; i
< len1
; i
++)
1426 u64
[len0
+ i
] = v1
->constant
->values
[0].u64
[i
];
1429 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1430 uint32_t comp
= w
[i
+ 6];
1431 /* If component is not used, set the value to a known constant
1432 * to detect if it is wrongly used.
1434 if (comp
== (uint32_t)-1)
1435 val
->constant
->values
[0].u64
[j
] = 0xdeadbeefdeadbeef;
1437 val
->constant
->values
[0].u64
[j
] = u64
[comp
];
1440 /* This is for both 32-bit and 16-bit values */
1442 if (v0
->value_type
== vtn_value_type_constant
) {
1443 for (unsigned i
= 0; i
< len0
; i
++)
1444 u32
[i
] = v0
->constant
->values
[0].u32
[i
];
1446 if (v1
->value_type
== vtn_value_type_constant
) {
1447 for (unsigned i
= 0; i
< len1
; i
++)
1448 u32
[len0
+ i
] = v1
->constant
->values
[0].u32
[i
];
1451 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1452 uint32_t comp
= w
[i
+ 6];
1453 /* If component is not used, set the value to a known constant
1454 * to detect if it is wrongly used.
1456 if (comp
== (uint32_t)-1)
1457 val
->constant
->values
[0].u32
[j
] = 0xdeadbeef;
1459 val
->constant
->values
[0].u32
[j
] = u32
[comp
];
1465 case SpvOpCompositeExtract
:
1466 case SpvOpCompositeInsert
: {
1467 struct vtn_value
*comp
;
1468 unsigned deref_start
;
1469 struct nir_constant
**c
;
1470 if (opcode
== SpvOpCompositeExtract
) {
1471 comp
= vtn_value(b
, w
[4], vtn_value_type_constant
);
1473 c
= &comp
->constant
;
1475 comp
= vtn_value(b
, w
[5], vtn_value_type_constant
);
1477 val
->constant
= nir_constant_clone(comp
->constant
,
1484 const struct glsl_type
*type
= comp
->type
->type
;
1485 for (unsigned i
= deref_start
; i
< count
; i
++) {
1486 switch (glsl_get_base_type(type
)) {
1487 case GLSL_TYPE_UINT
:
1489 case GLSL_TYPE_UINT16
:
1490 case GLSL_TYPE_INT16
:
1491 case GLSL_TYPE_UINT64
:
1492 case GLSL_TYPE_INT64
:
1493 case GLSL_TYPE_FLOAT
:
1494 case GLSL_TYPE_FLOAT16
:
1495 case GLSL_TYPE_DOUBLE
:
1496 case GLSL_TYPE_BOOL
:
1497 /* If we hit this granularity, we're picking off an element */
1498 if (glsl_type_is_matrix(type
)) {
1499 vtn_assert(col
== 0 && elem
== -1);
1502 type
= glsl_get_column_type(type
);
1504 vtn_assert(elem
<= 0 && glsl_type_is_vector(type
));
1506 type
= glsl_scalar_type(glsl_get_base_type(type
));
1510 case GLSL_TYPE_ARRAY
:
1511 c
= &(*c
)->elements
[w
[i
]];
1512 type
= glsl_get_array_element(type
);
1515 case GLSL_TYPE_STRUCT
:
1516 c
= &(*c
)->elements
[w
[i
]];
1517 type
= glsl_get_struct_field(type
, w
[i
]);
1521 vtn_fail("Invalid constant type");
1525 if (opcode
== SpvOpCompositeExtract
) {
1529 unsigned num_components
= glsl_get_vector_elements(type
);
1530 unsigned bit_size
= glsl_get_bit_size(type
);
1531 for (unsigned i
= 0; i
< num_components
; i
++)
1534 val
->constant
->values
[0].u64
[i
] = (*c
)->values
[col
].u64
[elem
+ i
];
1537 val
->constant
->values
[0].u32
[i
] = (*c
)->values
[col
].u32
[elem
+ i
];
1540 val
->constant
->values
[0].u16
[i
] = (*c
)->values
[col
].u16
[elem
+ i
];
1543 vtn_fail("Invalid SpvOpCompositeExtract bit size");
1547 struct vtn_value
*insert
=
1548 vtn_value(b
, w
[4], vtn_value_type_constant
);
1549 vtn_assert(insert
->type
->type
== type
);
1551 *c
= insert
->constant
;
1553 unsigned num_components
= glsl_get_vector_elements(type
);
1554 unsigned bit_size
= glsl_get_bit_size(type
);
1555 for (unsigned i
= 0; i
< num_components
; i
++)
1558 (*c
)->values
[col
].u64
[elem
+ i
] = insert
->constant
->values
[0].u64
[i
];
1561 (*c
)->values
[col
].u32
[elem
+ i
] = insert
->constant
->values
[0].u32
[i
];
1564 (*c
)->values
[col
].u16
[elem
+ i
] = insert
->constant
->values
[0].u16
[i
];
1567 vtn_fail("Invalid SpvOpCompositeInsert bit size");
1576 nir_alu_type dst_alu_type
= nir_get_nir_type_for_glsl_type(val
->type
->type
);
1577 nir_alu_type src_alu_type
= dst_alu_type
;
1578 unsigned num_components
= glsl_get_vector_elements(val
->type
->type
);
1581 vtn_assert(count
<= 7);
1586 /* We have a source in a conversion */
1588 nir_get_nir_type_for_glsl_type(
1589 vtn_value(b
, w
[4], vtn_value_type_constant
)->type
->type
);
1590 /* We use the bitsize of the conversion source to evaluate the opcode later */
1591 bit_size
= glsl_get_bit_size(
1592 vtn_value(b
, w
[4], vtn_value_type_constant
)->type
->type
);
1595 bit_size
= glsl_get_bit_size(val
->type
->type
);
1598 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(b
, opcode
, &swap
,
1601 nir_const_value src
[4];
1603 for (unsigned i
= 0; i
< count
- 4; i
++) {
1605 vtn_value(b
, w
[4 + i
], vtn_value_type_constant
)->constant
;
1607 unsigned j
= swap
? 1 - i
: i
;
1608 src
[j
] = c
->values
[0];
1611 val
->constant
->values
[0] =
1612 nir_eval_const_opcode(op
, num_components
, bit_size
, src
);
1619 case SpvOpConstantNull
:
1620 val
->constant
= vtn_null_constant(b
, val
->type
->type
);
1623 case SpvOpConstantSampler
:
1624 vtn_fail("OpConstantSampler requires Kernel Capability");
1628 vtn_fail("Unhandled opcode");
1631 /* Now that we have the value, update the workgroup size if needed */
1632 vtn_foreach_decoration(b
, val
, handle_workgroup_size_decoration_cb
, NULL
);
1636 vtn_handle_function_call(struct vtn_builder
*b
, SpvOp opcode
,
1637 const uint32_t *w
, unsigned count
)
1639 struct vtn_type
*res_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1640 struct vtn_function
*vtn_callee
=
1641 vtn_value(b
, w
[3], vtn_value_type_function
)->func
;
1642 struct nir_function
*callee
= vtn_callee
->impl
->function
;
1644 vtn_callee
->referenced
= true;
1646 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, callee
);
1647 for (unsigned i
= 0; i
< call
->num_params
; i
++) {
1648 unsigned arg_id
= w
[4 + i
];
1649 struct vtn_value
*arg
= vtn_untyped_value(b
, arg_id
);
1650 if (arg
->value_type
== vtn_value_type_pointer
&&
1651 arg
->pointer
->ptr_type
->type
== NULL
) {
1652 nir_deref_var
*d
= vtn_pointer_to_deref(b
, arg
->pointer
);
1653 call
->params
[i
] = nir_deref_var_clone(d
, call
);
1655 struct vtn_ssa_value
*arg_ssa
= vtn_ssa_value(b
, arg_id
);
1657 /* Make a temporary to store the argument in */
1659 nir_local_variable_create(b
->nb
.impl
, arg_ssa
->type
, "arg_tmp");
1660 call
->params
[i
] = nir_deref_var_create(call
, tmp
);
1662 vtn_local_store(b
, arg_ssa
, call
->params
[i
]);
1666 nir_variable
*out_tmp
= NULL
;
1667 vtn_assert(res_type
->type
== callee
->return_type
);
1668 if (!glsl_type_is_void(callee
->return_type
)) {
1669 out_tmp
= nir_local_variable_create(b
->nb
.impl
, callee
->return_type
,
1671 call
->return_deref
= nir_deref_var_create(call
, out_tmp
);
1674 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
1676 if (glsl_type_is_void(callee
->return_type
)) {
1677 vtn_push_value(b
, w
[2], vtn_value_type_undef
);
1679 vtn_push_ssa(b
, w
[2], res_type
, vtn_local_load(b
, call
->return_deref
));
1683 struct vtn_ssa_value
*
1684 vtn_create_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
1686 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
1689 if (!glsl_type_is_vector_or_scalar(type
)) {
1690 unsigned elems
= glsl_get_length(type
);
1691 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
1692 for (unsigned i
= 0; i
< elems
; i
++) {
1693 const struct glsl_type
*child_type
;
1695 switch (glsl_get_base_type(type
)) {
1697 case GLSL_TYPE_UINT
:
1698 case GLSL_TYPE_INT16
:
1699 case GLSL_TYPE_UINT16
:
1700 case GLSL_TYPE_INT64
:
1701 case GLSL_TYPE_UINT64
:
1702 case GLSL_TYPE_BOOL
:
1703 case GLSL_TYPE_FLOAT
:
1704 case GLSL_TYPE_FLOAT16
:
1705 case GLSL_TYPE_DOUBLE
:
1706 child_type
= glsl_get_column_type(type
);
1708 case GLSL_TYPE_ARRAY
:
1709 child_type
= glsl_get_array_element(type
);
1711 case GLSL_TYPE_STRUCT
:
1712 child_type
= glsl_get_struct_field(type
, i
);
1715 vtn_fail("unkown base type");
1718 val
->elems
[i
] = vtn_create_ssa_value(b
, child_type
);
1726 vtn_tex_src(struct vtn_builder
*b
, unsigned index
, nir_tex_src_type type
)
1729 src
.src
= nir_src_for_ssa(vtn_ssa_value(b
, index
)->def
);
1730 src
.src_type
= type
;
1735 vtn_handle_texture(struct vtn_builder
*b
, SpvOp opcode
,
1736 const uint32_t *w
, unsigned count
)
1738 if (opcode
== SpvOpSampledImage
) {
1739 struct vtn_value
*val
=
1740 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
1741 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
1742 val
->sampled_image
->type
=
1743 vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1744 val
->sampled_image
->image
=
1745 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
1746 val
->sampled_image
->sampler
=
1747 vtn_value(b
, w
[4], vtn_value_type_pointer
)->pointer
;
1749 } else if (opcode
== SpvOpImage
) {
1750 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
1751 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[3]);
1752 if (src_val
->value_type
== vtn_value_type_sampled_image
) {
1753 val
->pointer
= src_val
->sampled_image
->image
;
1755 vtn_assert(src_val
->value_type
== vtn_value_type_pointer
);
1756 val
->pointer
= src_val
->pointer
;
1761 struct vtn_type
*ret_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1762 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1764 struct vtn_sampled_image sampled
;
1765 struct vtn_value
*sampled_val
= vtn_untyped_value(b
, w
[3]);
1766 if (sampled_val
->value_type
== vtn_value_type_sampled_image
) {
1767 sampled
= *sampled_val
->sampled_image
;
1769 vtn_assert(sampled_val
->value_type
== vtn_value_type_pointer
);
1770 sampled
.type
= sampled_val
->pointer
->type
;
1771 sampled
.image
= NULL
;
1772 sampled
.sampler
= sampled_val
->pointer
;
1775 const struct glsl_type
*image_type
= sampled
.type
->type
;
1776 const enum glsl_sampler_dim sampler_dim
= glsl_get_sampler_dim(image_type
);
1777 const bool is_array
= glsl_sampler_type_is_array(image_type
);
1778 const bool is_shadow
= glsl_sampler_type_is_shadow(image_type
);
1780 /* Figure out the base texture operation */
1783 case SpvOpImageSampleImplicitLod
:
1784 case SpvOpImageSampleDrefImplicitLod
:
1785 case SpvOpImageSampleProjImplicitLod
:
1786 case SpvOpImageSampleProjDrefImplicitLod
:
1787 texop
= nir_texop_tex
;
1790 case SpvOpImageSampleExplicitLod
:
1791 case SpvOpImageSampleDrefExplicitLod
:
1792 case SpvOpImageSampleProjExplicitLod
:
1793 case SpvOpImageSampleProjDrefExplicitLod
:
1794 texop
= nir_texop_txl
;
1797 case SpvOpImageFetch
:
1798 if (glsl_get_sampler_dim(image_type
) == GLSL_SAMPLER_DIM_MS
) {
1799 texop
= nir_texop_txf_ms
;
1801 texop
= nir_texop_txf
;
1805 case SpvOpImageGather
:
1806 case SpvOpImageDrefGather
:
1807 texop
= nir_texop_tg4
;
1810 case SpvOpImageQuerySizeLod
:
1811 case SpvOpImageQuerySize
:
1812 texop
= nir_texop_txs
;
1815 case SpvOpImageQueryLod
:
1816 texop
= nir_texop_lod
;
1819 case SpvOpImageQueryLevels
:
1820 texop
= nir_texop_query_levels
;
1823 case SpvOpImageQuerySamples
:
1824 texop
= nir_texop_texture_samples
;
1828 vtn_fail("Unhandled opcode");
1831 nir_tex_src srcs
[8]; /* 8 should be enough */
1832 nir_tex_src
*p
= srcs
;
1836 struct nir_ssa_def
*coord
;
1837 unsigned coord_components
;
1839 case SpvOpImageSampleImplicitLod
:
1840 case SpvOpImageSampleExplicitLod
:
1841 case SpvOpImageSampleDrefImplicitLod
:
1842 case SpvOpImageSampleDrefExplicitLod
:
1843 case SpvOpImageSampleProjImplicitLod
:
1844 case SpvOpImageSampleProjExplicitLod
:
1845 case SpvOpImageSampleProjDrefImplicitLod
:
1846 case SpvOpImageSampleProjDrefExplicitLod
:
1847 case SpvOpImageFetch
:
1848 case SpvOpImageGather
:
1849 case SpvOpImageDrefGather
:
1850 case SpvOpImageQueryLod
: {
1851 /* All these types have the coordinate as their first real argument */
1852 switch (sampler_dim
) {
1853 case GLSL_SAMPLER_DIM_1D
:
1854 case GLSL_SAMPLER_DIM_BUF
:
1855 coord_components
= 1;
1857 case GLSL_SAMPLER_DIM_2D
:
1858 case GLSL_SAMPLER_DIM_RECT
:
1859 case GLSL_SAMPLER_DIM_MS
:
1860 coord_components
= 2;
1862 case GLSL_SAMPLER_DIM_3D
:
1863 case GLSL_SAMPLER_DIM_CUBE
:
1864 coord_components
= 3;
1867 vtn_fail("Invalid sampler type");
1870 if (is_array
&& texop
!= nir_texop_lod
)
1873 coord
= vtn_ssa_value(b
, w
[idx
++])->def
;
1874 p
->src
= nir_src_for_ssa(nir_channels(&b
->nb
, coord
,
1875 (1 << coord_components
) - 1));
1876 p
->src_type
= nir_tex_src_coord
;
1883 coord_components
= 0;
1888 case SpvOpImageSampleProjImplicitLod
:
1889 case SpvOpImageSampleProjExplicitLod
:
1890 case SpvOpImageSampleProjDrefImplicitLod
:
1891 case SpvOpImageSampleProjDrefExplicitLod
:
1892 /* These have the projector as the last coordinate component */
1893 p
->src
= nir_src_for_ssa(nir_channel(&b
->nb
, coord
, coord_components
));
1894 p
->src_type
= nir_tex_src_projector
;
1902 unsigned gather_component
= 0;
1904 case SpvOpImageSampleDrefImplicitLod
:
1905 case SpvOpImageSampleDrefExplicitLod
:
1906 case SpvOpImageSampleProjDrefImplicitLod
:
1907 case SpvOpImageSampleProjDrefExplicitLod
:
1908 case SpvOpImageDrefGather
:
1909 /* These all have an explicit depth value as their next source */
1910 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_comparator
);
1913 case SpvOpImageGather
:
1914 /* This has a component as its next source */
1916 vtn_value(b
, w
[idx
++], vtn_value_type_constant
)->constant
->values
[0].u32
[0];
1923 /* For OpImageQuerySizeLod, we always have an LOD */
1924 if (opcode
== SpvOpImageQuerySizeLod
)
1925 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
1927 /* Now we need to handle some number of optional arguments */
1928 const struct vtn_ssa_value
*gather_offsets
= NULL
;
1930 uint32_t operands
= w
[idx
++];
1932 if (operands
& SpvImageOperandsBiasMask
) {
1933 vtn_assert(texop
== nir_texop_tex
);
1934 texop
= nir_texop_txb
;
1935 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_bias
);
1938 if (operands
& SpvImageOperandsLodMask
) {
1939 vtn_assert(texop
== nir_texop_txl
|| texop
== nir_texop_txf
||
1940 texop
== nir_texop_txs
);
1941 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
1944 if (operands
& SpvImageOperandsGradMask
) {
1945 vtn_assert(texop
== nir_texop_txl
);
1946 texop
= nir_texop_txd
;
1947 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddx
);
1948 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddy
);
1951 if (operands
& SpvImageOperandsOffsetMask
||
1952 operands
& SpvImageOperandsConstOffsetMask
)
1953 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_offset
);
1955 if (operands
& SpvImageOperandsConstOffsetsMask
) {
1956 gather_offsets
= vtn_ssa_value(b
, w
[idx
++]);
1957 (*p
++) = (nir_tex_src
){};
1960 if (operands
& SpvImageOperandsSampleMask
) {
1961 vtn_assert(texop
== nir_texop_txf_ms
);
1962 texop
= nir_texop_txf_ms
;
1963 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ms_index
);
1966 /* We should have now consumed exactly all of the arguments */
1967 vtn_assert(idx
== count
);
1969 nir_tex_instr
*instr
= nir_tex_instr_create(b
->shader
, p
- srcs
);
1972 memcpy(instr
->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
1974 instr
->coord_components
= coord_components
;
1975 instr
->sampler_dim
= sampler_dim
;
1976 instr
->is_array
= is_array
;
1977 instr
->is_shadow
= is_shadow
;
1978 instr
->is_new_style_shadow
=
1979 is_shadow
&& glsl_get_components(ret_type
->type
) == 1;
1980 instr
->component
= gather_component
;
1982 switch (glsl_get_sampler_result_type(image_type
)) {
1983 case GLSL_TYPE_FLOAT
: instr
->dest_type
= nir_type_float
; break;
1984 case GLSL_TYPE_INT
: instr
->dest_type
= nir_type_int
; break;
1985 case GLSL_TYPE_UINT
: instr
->dest_type
= nir_type_uint
; break;
1986 case GLSL_TYPE_BOOL
: instr
->dest_type
= nir_type_bool
; break;
1988 vtn_fail("Invalid base type for sampler result");
1991 nir_deref_var
*sampler
= vtn_pointer_to_deref(b
, sampled
.sampler
);
1992 nir_deref_var
*texture
;
1993 if (sampled
.image
) {
1994 nir_deref_var
*image
= vtn_pointer_to_deref(b
, sampled
.image
);
2000 instr
->texture
= nir_deref_var_clone(texture
, instr
);
2002 switch (instr
->op
) {
2008 /* These operations require a sampler */
2009 instr
->sampler
= nir_deref_var_clone(sampler
, instr
);
2012 case nir_texop_txf_ms
:
2015 case nir_texop_query_levels
:
2016 case nir_texop_texture_samples
:
2017 case nir_texop_samples_identical
:
2019 instr
->sampler
= NULL
;
2021 case nir_texop_txf_ms_mcs
:
2022 vtn_fail("unexpected nir_texop_txf_ms_mcs");
2025 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
2026 nir_tex_instr_dest_size(instr
), 32, NULL
);
2028 vtn_assert(glsl_get_vector_elements(ret_type
->type
) ==
2029 nir_tex_instr_dest_size(instr
));
2032 nir_instr
*instruction
;
2033 if (gather_offsets
) {
2034 vtn_assert(glsl_get_base_type(gather_offsets
->type
) == GLSL_TYPE_ARRAY
);
2035 vtn_assert(glsl_get_length(gather_offsets
->type
) == 4);
2036 nir_tex_instr
*instrs
[4] = {instr
, NULL
, NULL
, NULL
};
2038 /* Copy the current instruction 4x */
2039 for (uint32_t i
= 1; i
< 4; i
++) {
2040 instrs
[i
] = nir_tex_instr_create(b
->shader
, instr
->num_srcs
);
2041 instrs
[i
]->op
= instr
->op
;
2042 instrs
[i
]->coord_components
= instr
->coord_components
;
2043 instrs
[i
]->sampler_dim
= instr
->sampler_dim
;
2044 instrs
[i
]->is_array
= instr
->is_array
;
2045 instrs
[i
]->is_shadow
= instr
->is_shadow
;
2046 instrs
[i
]->is_new_style_shadow
= instr
->is_new_style_shadow
;
2047 instrs
[i
]->component
= instr
->component
;
2048 instrs
[i
]->dest_type
= instr
->dest_type
;
2049 instrs
[i
]->texture
= nir_deref_var_clone(texture
, instrs
[i
]);
2050 instrs
[i
]->sampler
= NULL
;
2052 memcpy(instrs
[i
]->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
2054 nir_ssa_dest_init(&instrs
[i
]->instr
, &instrs
[i
]->dest
,
2055 nir_tex_instr_dest_size(instr
), 32, NULL
);
2058 /* Fill in the last argument with the offset from the passed in offsets
2059 * and insert the instruction into the stream.
2061 for (uint32_t i
= 0; i
< 4; i
++) {
2063 src
.src
= nir_src_for_ssa(gather_offsets
->elems
[i
]->def
);
2064 src
.src_type
= nir_tex_src_offset
;
2065 instrs
[i
]->src
[instrs
[i
]->num_srcs
- 1] = src
;
2066 nir_builder_instr_insert(&b
->nb
, &instrs
[i
]->instr
);
2069 /* Combine the results of the 4 instructions by taking their .w
2072 nir_alu_instr
*vec4
= nir_alu_instr_create(b
->shader
, nir_op_vec4
);
2073 nir_ssa_dest_init(&vec4
->instr
, &vec4
->dest
.dest
, 4, 32, NULL
);
2074 vec4
->dest
.write_mask
= 0xf;
2075 for (uint32_t i
= 0; i
< 4; i
++) {
2076 vec4
->src
[i
].src
= nir_src_for_ssa(&instrs
[i
]->dest
.ssa
);
2077 vec4
->src
[i
].swizzle
[0] = 3;
2079 def
= &vec4
->dest
.dest
.ssa
;
2080 instruction
= &vec4
->instr
;
2082 def
= &instr
->dest
.ssa
;
2083 instruction
= &instr
->instr
;
2086 val
->ssa
= vtn_create_ssa_value(b
, ret_type
->type
);
2087 val
->ssa
->def
= def
;
2089 nir_builder_instr_insert(&b
->nb
, instruction
);
2093 fill_common_atomic_sources(struct vtn_builder
*b
, SpvOp opcode
,
2094 const uint32_t *w
, nir_src
*src
)
2097 case SpvOpAtomicIIncrement
:
2098 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
2101 case SpvOpAtomicIDecrement
:
2102 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
2105 case SpvOpAtomicISub
:
2107 nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
2110 case SpvOpAtomicCompareExchange
:
2111 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[8])->def
);
2112 src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
2115 case SpvOpAtomicExchange
:
2116 case SpvOpAtomicIAdd
:
2117 case SpvOpAtomicSMin
:
2118 case SpvOpAtomicUMin
:
2119 case SpvOpAtomicSMax
:
2120 case SpvOpAtomicUMax
:
2121 case SpvOpAtomicAnd
:
2123 case SpvOpAtomicXor
:
2124 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
2128 vtn_fail("Invalid SPIR-V atomic");
2132 static nir_ssa_def
*
2133 get_image_coord(struct vtn_builder
*b
, uint32_t value
)
2135 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, value
);
2137 /* The image_load_store intrinsics assume a 4-dim coordinate */
2138 unsigned dim
= glsl_get_vector_elements(coord
->type
);
2139 unsigned swizzle
[4];
2140 for (unsigned i
= 0; i
< 4; i
++)
2141 swizzle
[i
] = MIN2(i
, dim
- 1);
2143 return nir_swizzle(&b
->nb
, coord
->def
, swizzle
, 4, false);
2147 vtn_handle_image(struct vtn_builder
*b
, SpvOp opcode
,
2148 const uint32_t *w
, unsigned count
)
2150 /* Just get this one out of the way */
2151 if (opcode
== SpvOpImageTexelPointer
) {
2152 struct vtn_value
*val
=
2153 vtn_push_value(b
, w
[2], vtn_value_type_image_pointer
);
2154 val
->image
= ralloc(b
, struct vtn_image_pointer
);
2156 val
->image
->image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2157 val
->image
->coord
= get_image_coord(b
, w
[4]);
2158 val
->image
->sample
= vtn_ssa_value(b
, w
[5])->def
;
2162 struct vtn_image_pointer image
;
2165 case SpvOpAtomicExchange
:
2166 case SpvOpAtomicCompareExchange
:
2167 case SpvOpAtomicCompareExchangeWeak
:
2168 case SpvOpAtomicIIncrement
:
2169 case SpvOpAtomicIDecrement
:
2170 case SpvOpAtomicIAdd
:
2171 case SpvOpAtomicISub
:
2172 case SpvOpAtomicLoad
:
2173 case SpvOpAtomicSMin
:
2174 case SpvOpAtomicUMin
:
2175 case SpvOpAtomicSMax
:
2176 case SpvOpAtomicUMax
:
2177 case SpvOpAtomicAnd
:
2179 case SpvOpAtomicXor
:
2180 image
= *vtn_value(b
, w
[3], vtn_value_type_image_pointer
)->image
;
2183 case SpvOpAtomicStore
:
2184 image
= *vtn_value(b
, w
[1], vtn_value_type_image_pointer
)->image
;
2187 case SpvOpImageQuerySize
:
2188 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2190 image
.sample
= NULL
;
2193 case SpvOpImageRead
:
2194 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2195 image
.coord
= get_image_coord(b
, w
[4]);
2197 if (count
> 5 && (w
[5] & SpvImageOperandsSampleMask
)) {
2198 vtn_assert(w
[5] == SpvImageOperandsSampleMask
);
2199 image
.sample
= vtn_ssa_value(b
, w
[6])->def
;
2201 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2205 case SpvOpImageWrite
:
2206 image
.image
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
2207 image
.coord
= get_image_coord(b
, w
[2]);
2211 if (count
> 4 && (w
[4] & SpvImageOperandsSampleMask
)) {
2212 vtn_assert(w
[4] == SpvImageOperandsSampleMask
);
2213 image
.sample
= vtn_ssa_value(b
, w
[5])->def
;
2215 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2220 vtn_fail("Invalid image opcode");
2223 nir_intrinsic_op op
;
2225 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_##N; break;
2226 OP(ImageQuerySize
, size
)
2228 OP(ImageWrite
, store
)
2229 OP(AtomicLoad
, load
)
2230 OP(AtomicStore
, store
)
2231 OP(AtomicExchange
, atomic_exchange
)
2232 OP(AtomicCompareExchange
, atomic_comp_swap
)
2233 OP(AtomicIIncrement
, atomic_add
)
2234 OP(AtomicIDecrement
, atomic_add
)
2235 OP(AtomicIAdd
, atomic_add
)
2236 OP(AtomicISub
, atomic_add
)
2237 OP(AtomicSMin
, atomic_min
)
2238 OP(AtomicUMin
, atomic_min
)
2239 OP(AtomicSMax
, atomic_max
)
2240 OP(AtomicUMax
, atomic_max
)
2241 OP(AtomicAnd
, atomic_and
)
2242 OP(AtomicOr
, atomic_or
)
2243 OP(AtomicXor
, atomic_xor
)
2246 vtn_fail("Invalid image opcode");
2249 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
2251 nir_deref_var
*image_deref
= vtn_pointer_to_deref(b
, image
.image
);
2252 intrin
->variables
[0] = nir_deref_var_clone(image_deref
, intrin
);
2254 /* ImageQuerySize doesn't take any extra parameters */
2255 if (opcode
!= SpvOpImageQuerySize
) {
2256 /* The image coordinate is always 4 components but we may not have that
2257 * many. Swizzle to compensate.
2260 for (unsigned i
= 0; i
< 4; i
++)
2261 swiz
[i
] = i
< image
.coord
->num_components
? i
: 0;
2262 intrin
->src
[0] = nir_src_for_ssa(nir_swizzle(&b
->nb
, image
.coord
,
2264 intrin
->src
[1] = nir_src_for_ssa(image
.sample
);
2268 case SpvOpAtomicLoad
:
2269 case SpvOpImageQuerySize
:
2270 case SpvOpImageRead
:
2272 case SpvOpAtomicStore
:
2273 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2275 case SpvOpImageWrite
:
2276 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[3])->def
);
2279 case SpvOpAtomicCompareExchange
:
2280 case SpvOpAtomicIIncrement
:
2281 case SpvOpAtomicIDecrement
:
2282 case SpvOpAtomicExchange
:
2283 case SpvOpAtomicIAdd
:
2284 case SpvOpAtomicISub
:
2285 case SpvOpAtomicSMin
:
2286 case SpvOpAtomicUMin
:
2287 case SpvOpAtomicSMax
:
2288 case SpvOpAtomicUMax
:
2289 case SpvOpAtomicAnd
:
2291 case SpvOpAtomicXor
:
2292 fill_common_atomic_sources(b
, opcode
, w
, &intrin
->src
[2]);
2296 vtn_fail("Invalid image opcode");
2299 if (opcode
!= SpvOpImageWrite
) {
2300 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2301 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2303 unsigned dest_components
=
2304 nir_intrinsic_infos
[intrin
->intrinsic
].dest_components
;
2305 if (intrin
->intrinsic
== nir_intrinsic_image_size
) {
2306 dest_components
= intrin
->num_components
=
2307 glsl_get_vector_elements(type
->type
);
2310 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
,
2311 dest_components
, 32, NULL
);
2313 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2315 val
->ssa
= vtn_create_ssa_value(b
, type
->type
);
2316 val
->ssa
->def
= &intrin
->dest
.ssa
;
2318 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2322 static nir_intrinsic_op
2323 get_ssbo_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2326 case SpvOpAtomicLoad
: return nir_intrinsic_load_ssbo
;
2327 case SpvOpAtomicStore
: return nir_intrinsic_store_ssbo
;
2328 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
2329 OP(AtomicExchange
, atomic_exchange
)
2330 OP(AtomicCompareExchange
, atomic_comp_swap
)
2331 OP(AtomicIIncrement
, atomic_add
)
2332 OP(AtomicIDecrement
, atomic_add
)
2333 OP(AtomicIAdd
, atomic_add
)
2334 OP(AtomicISub
, atomic_add
)
2335 OP(AtomicSMin
, atomic_imin
)
2336 OP(AtomicUMin
, atomic_umin
)
2337 OP(AtomicSMax
, atomic_imax
)
2338 OP(AtomicUMax
, atomic_umax
)
2339 OP(AtomicAnd
, atomic_and
)
2340 OP(AtomicOr
, atomic_or
)
2341 OP(AtomicXor
, atomic_xor
)
2344 vtn_fail("Invalid SSBO atomic");
2348 static nir_intrinsic_op
2349 get_shared_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2352 case SpvOpAtomicLoad
: return nir_intrinsic_load_shared
;
2353 case SpvOpAtomicStore
: return nir_intrinsic_store_shared
;
2354 #define OP(S, N) case SpvOp##S: return nir_intrinsic_shared_##N;
2355 OP(AtomicExchange
, atomic_exchange
)
2356 OP(AtomicCompareExchange
, atomic_comp_swap
)
2357 OP(AtomicIIncrement
, atomic_add
)
2358 OP(AtomicIDecrement
, atomic_add
)
2359 OP(AtomicIAdd
, atomic_add
)
2360 OP(AtomicISub
, atomic_add
)
2361 OP(AtomicSMin
, atomic_imin
)
2362 OP(AtomicUMin
, atomic_umin
)
2363 OP(AtomicSMax
, atomic_imax
)
2364 OP(AtomicUMax
, atomic_umax
)
2365 OP(AtomicAnd
, atomic_and
)
2366 OP(AtomicOr
, atomic_or
)
2367 OP(AtomicXor
, atomic_xor
)
2370 vtn_fail("Invalid shared atomic");
2374 static nir_intrinsic_op
2375 get_var_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2378 case SpvOpAtomicLoad
: return nir_intrinsic_load_var
;
2379 case SpvOpAtomicStore
: return nir_intrinsic_store_var
;
2380 #define OP(S, N) case SpvOp##S: return nir_intrinsic_var_##N;
2381 OP(AtomicExchange
, atomic_exchange
)
2382 OP(AtomicCompareExchange
, atomic_comp_swap
)
2383 OP(AtomicIIncrement
, atomic_add
)
2384 OP(AtomicIDecrement
, atomic_add
)
2385 OP(AtomicIAdd
, atomic_add
)
2386 OP(AtomicISub
, atomic_add
)
2387 OP(AtomicSMin
, atomic_imin
)
2388 OP(AtomicUMin
, atomic_umin
)
2389 OP(AtomicSMax
, atomic_imax
)
2390 OP(AtomicUMax
, atomic_umax
)
2391 OP(AtomicAnd
, atomic_and
)
2392 OP(AtomicOr
, atomic_or
)
2393 OP(AtomicXor
, atomic_xor
)
2396 vtn_fail("Invalid shared atomic");
2401 vtn_handle_ssbo_or_shared_atomic(struct vtn_builder
*b
, SpvOp opcode
,
2402 const uint32_t *w
, unsigned count
)
2404 struct vtn_pointer
*ptr
;
2405 nir_intrinsic_instr
*atomic
;
2408 case SpvOpAtomicLoad
:
2409 case SpvOpAtomicExchange
:
2410 case SpvOpAtomicCompareExchange
:
2411 case SpvOpAtomicCompareExchangeWeak
:
2412 case SpvOpAtomicIIncrement
:
2413 case SpvOpAtomicIDecrement
:
2414 case SpvOpAtomicIAdd
:
2415 case SpvOpAtomicISub
:
2416 case SpvOpAtomicSMin
:
2417 case SpvOpAtomicUMin
:
2418 case SpvOpAtomicSMax
:
2419 case SpvOpAtomicUMax
:
2420 case SpvOpAtomicAnd
:
2422 case SpvOpAtomicXor
:
2423 ptr
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2426 case SpvOpAtomicStore
:
2427 ptr
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
2431 vtn_fail("Invalid SPIR-V atomic");
2435 SpvScope scope = w[4];
2436 SpvMemorySemanticsMask semantics = w[5];
2439 if (ptr
->mode
== vtn_variable_mode_workgroup
&&
2440 !b
->options
->lower_workgroup_access_to_offsets
) {
2441 nir_deref_var
*deref
= vtn_pointer_to_deref(b
, ptr
);
2442 const struct glsl_type
*deref_type
= nir_deref_tail(&deref
->deref
)->type
;
2443 nir_intrinsic_op op
= get_var_nir_atomic_op(b
, opcode
);
2444 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2445 atomic
->variables
[0] = nir_deref_var_clone(deref
, atomic
);
2448 case SpvOpAtomicLoad
:
2449 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2452 case SpvOpAtomicStore
:
2453 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2454 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2455 atomic
->src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2458 case SpvOpAtomicExchange
:
2459 case SpvOpAtomicCompareExchange
:
2460 case SpvOpAtomicCompareExchangeWeak
:
2461 case SpvOpAtomicIIncrement
:
2462 case SpvOpAtomicIDecrement
:
2463 case SpvOpAtomicIAdd
:
2464 case SpvOpAtomicISub
:
2465 case SpvOpAtomicSMin
:
2466 case SpvOpAtomicUMin
:
2467 case SpvOpAtomicSMax
:
2468 case SpvOpAtomicUMax
:
2469 case SpvOpAtomicAnd
:
2471 case SpvOpAtomicXor
:
2472 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[0]);
2476 vtn_fail("Invalid SPIR-V atomic");
2480 nir_ssa_def
*offset
, *index
;
2481 offset
= vtn_pointer_to_offset(b
, ptr
, &index
, NULL
);
2483 nir_intrinsic_op op
;
2484 if (ptr
->mode
== vtn_variable_mode_ssbo
) {
2485 op
= get_ssbo_nir_atomic_op(b
, opcode
);
2487 vtn_assert(ptr
->mode
== vtn_variable_mode_workgroup
&&
2488 b
->options
->lower_workgroup_access_to_offsets
);
2489 op
= get_shared_nir_atomic_op(b
, opcode
);
2492 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2496 case SpvOpAtomicLoad
:
2497 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
2498 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2499 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2500 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2503 case SpvOpAtomicStore
:
2504 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
2505 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2506 atomic
->src
[src
++] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2507 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2508 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2509 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2512 case SpvOpAtomicExchange
:
2513 case SpvOpAtomicCompareExchange
:
2514 case SpvOpAtomicCompareExchangeWeak
:
2515 case SpvOpAtomicIIncrement
:
2516 case SpvOpAtomicIDecrement
:
2517 case SpvOpAtomicIAdd
:
2518 case SpvOpAtomicISub
:
2519 case SpvOpAtomicSMin
:
2520 case SpvOpAtomicUMin
:
2521 case SpvOpAtomicSMax
:
2522 case SpvOpAtomicUMax
:
2523 case SpvOpAtomicAnd
:
2525 case SpvOpAtomicXor
:
2526 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2527 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2528 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2529 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[src
]);
2533 vtn_fail("Invalid SPIR-V atomic");
2537 if (opcode
!= SpvOpAtomicStore
) {
2538 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2540 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
,
2541 glsl_get_vector_elements(type
->type
),
2542 glsl_get_bit_size(type
->type
), NULL
);
2544 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2545 val
->ssa
= rzalloc(b
, struct vtn_ssa_value
);
2546 val
->ssa
->def
= &atomic
->dest
.ssa
;
2547 val
->ssa
->type
= type
->type
;
2550 nir_builder_instr_insert(&b
->nb
, &atomic
->instr
);
2553 static nir_alu_instr
*
2554 create_vec(struct vtn_builder
*b
, unsigned num_components
, unsigned bit_size
)
2557 switch (num_components
) {
2558 case 1: op
= nir_op_fmov
; break;
2559 case 2: op
= nir_op_vec2
; break;
2560 case 3: op
= nir_op_vec3
; break;
2561 case 4: op
= nir_op_vec4
; break;
2562 default: vtn_fail("bad vector size");
2565 nir_alu_instr
*vec
= nir_alu_instr_create(b
->shader
, op
);
2566 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, num_components
,
2568 vec
->dest
.write_mask
= (1 << num_components
) - 1;
2573 struct vtn_ssa_value
*
2574 vtn_ssa_transpose(struct vtn_builder
*b
, struct vtn_ssa_value
*src
)
2576 if (src
->transposed
)
2577 return src
->transposed
;
2579 struct vtn_ssa_value
*dest
=
2580 vtn_create_ssa_value(b
, glsl_transposed_type(src
->type
));
2582 for (unsigned i
= 0; i
< glsl_get_matrix_columns(dest
->type
); i
++) {
2583 nir_alu_instr
*vec
= create_vec(b
, glsl_get_matrix_columns(src
->type
),
2584 glsl_get_bit_size(src
->type
));
2585 if (glsl_type_is_vector_or_scalar(src
->type
)) {
2586 vec
->src
[0].src
= nir_src_for_ssa(src
->def
);
2587 vec
->src
[0].swizzle
[0] = i
;
2589 for (unsigned j
= 0; j
< glsl_get_matrix_columns(src
->type
); j
++) {
2590 vec
->src
[j
].src
= nir_src_for_ssa(src
->elems
[j
]->def
);
2591 vec
->src
[j
].swizzle
[0] = i
;
2594 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2595 dest
->elems
[i
]->def
= &vec
->dest
.dest
.ssa
;
2598 dest
->transposed
= src
;
2604 vtn_vector_extract(struct vtn_builder
*b
, nir_ssa_def
*src
, unsigned index
)
2606 unsigned swiz
[4] = { index
};
2607 return nir_swizzle(&b
->nb
, src
, swiz
, 1, true);
2611 vtn_vector_insert(struct vtn_builder
*b
, nir_ssa_def
*src
, nir_ssa_def
*insert
,
2614 nir_alu_instr
*vec
= create_vec(b
, src
->num_components
,
2617 for (unsigned i
= 0; i
< src
->num_components
; i
++) {
2619 vec
->src
[i
].src
= nir_src_for_ssa(insert
);
2621 vec
->src
[i
].src
= nir_src_for_ssa(src
);
2622 vec
->src
[i
].swizzle
[0] = i
;
2626 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2628 return &vec
->dest
.dest
.ssa
;
2632 vtn_vector_extract_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
2635 nir_ssa_def
*dest
= vtn_vector_extract(b
, src
, 0);
2636 for (unsigned i
= 1; i
< src
->num_components
; i
++)
2637 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
2638 vtn_vector_extract(b
, src
, i
), dest
);
2644 vtn_vector_insert_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
2645 nir_ssa_def
*insert
, nir_ssa_def
*index
)
2647 nir_ssa_def
*dest
= vtn_vector_insert(b
, src
, insert
, 0);
2648 for (unsigned i
= 1; i
< src
->num_components
; i
++)
2649 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
2650 vtn_vector_insert(b
, src
, insert
, i
), dest
);
2655 static nir_ssa_def
*
2656 vtn_vector_shuffle(struct vtn_builder
*b
, unsigned num_components
,
2657 nir_ssa_def
*src0
, nir_ssa_def
*src1
,
2658 const uint32_t *indices
)
2660 nir_alu_instr
*vec
= create_vec(b
, num_components
, src0
->bit_size
);
2662 for (unsigned i
= 0; i
< num_components
; i
++) {
2663 uint32_t index
= indices
[i
];
2664 if (index
== 0xffffffff) {
2666 nir_src_for_ssa(nir_ssa_undef(&b
->nb
, 1, src0
->bit_size
));
2667 } else if (index
< src0
->num_components
) {
2668 vec
->src
[i
].src
= nir_src_for_ssa(src0
);
2669 vec
->src
[i
].swizzle
[0] = index
;
2671 vec
->src
[i
].src
= nir_src_for_ssa(src1
);
2672 vec
->src
[i
].swizzle
[0] = index
- src0
->num_components
;
2676 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2678 return &vec
->dest
.dest
.ssa
;
2682 * Concatentates a number of vectors/scalars together to produce a vector
2684 static nir_ssa_def
*
2685 vtn_vector_construct(struct vtn_builder
*b
, unsigned num_components
,
2686 unsigned num_srcs
, nir_ssa_def
**srcs
)
2688 nir_alu_instr
*vec
= create_vec(b
, num_components
, srcs
[0]->bit_size
);
2690 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2692 * "When constructing a vector, there must be at least two Constituent
2695 vtn_assert(num_srcs
>= 2);
2697 unsigned dest_idx
= 0;
2698 for (unsigned i
= 0; i
< num_srcs
; i
++) {
2699 nir_ssa_def
*src
= srcs
[i
];
2700 vtn_assert(dest_idx
+ src
->num_components
<= num_components
);
2701 for (unsigned j
= 0; j
< src
->num_components
; j
++) {
2702 vec
->src
[dest_idx
].src
= nir_src_for_ssa(src
);
2703 vec
->src
[dest_idx
].swizzle
[0] = j
;
2708 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2710 * "When constructing a vector, the total number of components in all
2711 * the operands must equal the number of components in Result Type."
2713 vtn_assert(dest_idx
== num_components
);
2715 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2717 return &vec
->dest
.dest
.ssa
;
2720 static struct vtn_ssa_value
*
2721 vtn_composite_copy(void *mem_ctx
, struct vtn_ssa_value
*src
)
2723 struct vtn_ssa_value
*dest
= rzalloc(mem_ctx
, struct vtn_ssa_value
);
2724 dest
->type
= src
->type
;
2726 if (glsl_type_is_vector_or_scalar(src
->type
)) {
2727 dest
->def
= src
->def
;
2729 unsigned elems
= glsl_get_length(src
->type
);
2731 dest
->elems
= ralloc_array(mem_ctx
, struct vtn_ssa_value
*, elems
);
2732 for (unsigned i
= 0; i
< elems
; i
++)
2733 dest
->elems
[i
] = vtn_composite_copy(mem_ctx
, src
->elems
[i
]);
2739 static struct vtn_ssa_value
*
2740 vtn_composite_insert(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
2741 struct vtn_ssa_value
*insert
, const uint32_t *indices
,
2742 unsigned num_indices
)
2744 struct vtn_ssa_value
*dest
= vtn_composite_copy(b
, src
);
2746 struct vtn_ssa_value
*cur
= dest
;
2748 for (i
= 0; i
< num_indices
- 1; i
++) {
2749 cur
= cur
->elems
[indices
[i
]];
2752 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
2753 /* According to the SPIR-V spec, OpCompositeInsert may work down to
2754 * the component granularity. In that case, the last index will be
2755 * the index to insert the scalar into the vector.
2758 cur
->def
= vtn_vector_insert(b
, cur
->def
, insert
->def
, indices
[i
]);
2760 cur
->elems
[indices
[i
]] = insert
;
2766 static struct vtn_ssa_value
*
2767 vtn_composite_extract(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
2768 const uint32_t *indices
, unsigned num_indices
)
2770 struct vtn_ssa_value
*cur
= src
;
2771 for (unsigned i
= 0; i
< num_indices
; i
++) {
2772 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
2773 vtn_assert(i
== num_indices
- 1);
2774 /* According to the SPIR-V spec, OpCompositeExtract may work down to
2775 * the component granularity. The last index will be the index of the
2776 * vector to extract.
2779 struct vtn_ssa_value
*ret
= rzalloc(b
, struct vtn_ssa_value
);
2780 ret
->type
= glsl_scalar_type(glsl_get_base_type(cur
->type
));
2781 ret
->def
= vtn_vector_extract(b
, cur
->def
, indices
[i
]);
2784 cur
= cur
->elems
[indices
[i
]];
2792 vtn_handle_composite(struct vtn_builder
*b
, SpvOp opcode
,
2793 const uint32_t *w
, unsigned count
)
2795 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2796 const struct glsl_type
*type
=
2797 vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
2798 val
->ssa
= vtn_create_ssa_value(b
, type
);
2801 case SpvOpVectorExtractDynamic
:
2802 val
->ssa
->def
= vtn_vector_extract_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
2803 vtn_ssa_value(b
, w
[4])->def
);
2806 case SpvOpVectorInsertDynamic
:
2807 val
->ssa
->def
= vtn_vector_insert_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
2808 vtn_ssa_value(b
, w
[4])->def
,
2809 vtn_ssa_value(b
, w
[5])->def
);
2812 case SpvOpVectorShuffle
:
2813 val
->ssa
->def
= vtn_vector_shuffle(b
, glsl_get_vector_elements(type
),
2814 vtn_ssa_value(b
, w
[3])->def
,
2815 vtn_ssa_value(b
, w
[4])->def
,
2819 case SpvOpCompositeConstruct
: {
2820 unsigned elems
= count
- 3;
2821 if (glsl_type_is_vector_or_scalar(type
)) {
2822 nir_ssa_def
*srcs
[4];
2823 for (unsigned i
= 0; i
< elems
; i
++)
2824 srcs
[i
] = vtn_ssa_value(b
, w
[3 + i
])->def
;
2826 vtn_vector_construct(b
, glsl_get_vector_elements(type
),
2829 val
->ssa
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
2830 for (unsigned i
= 0; i
< elems
; i
++)
2831 val
->ssa
->elems
[i
] = vtn_ssa_value(b
, w
[3 + i
]);
2835 case SpvOpCompositeExtract
:
2836 val
->ssa
= vtn_composite_extract(b
, vtn_ssa_value(b
, w
[3]),
2840 case SpvOpCompositeInsert
:
2841 val
->ssa
= vtn_composite_insert(b
, vtn_ssa_value(b
, w
[4]),
2842 vtn_ssa_value(b
, w
[3]),
2846 case SpvOpCopyObject
:
2847 val
->ssa
= vtn_composite_copy(b
, vtn_ssa_value(b
, w
[3]));
2851 vtn_fail("unknown composite operation");
2856 vtn_handle_barrier(struct vtn_builder
*b
, SpvOp opcode
,
2857 const uint32_t *w
, unsigned count
)
2859 nir_intrinsic_op intrinsic_op
;
2861 case SpvOpEmitVertex
:
2862 case SpvOpEmitStreamVertex
:
2863 intrinsic_op
= nir_intrinsic_emit_vertex
;
2865 case SpvOpEndPrimitive
:
2866 case SpvOpEndStreamPrimitive
:
2867 intrinsic_op
= nir_intrinsic_end_primitive
;
2869 case SpvOpMemoryBarrier
:
2870 intrinsic_op
= nir_intrinsic_memory_barrier
;
2872 case SpvOpControlBarrier
:
2873 intrinsic_op
= nir_intrinsic_barrier
;
2876 vtn_fail("unknown barrier instruction");
2879 nir_intrinsic_instr
*intrin
=
2880 nir_intrinsic_instr_create(b
->shader
, intrinsic_op
);
2882 if (opcode
== SpvOpEmitStreamVertex
|| opcode
== SpvOpEndStreamPrimitive
)
2883 nir_intrinsic_set_stream_id(intrin
, w
[1]);
2885 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2889 gl_primitive_from_spv_execution_mode(struct vtn_builder
*b
,
2890 SpvExecutionMode mode
)
2893 case SpvExecutionModeInputPoints
:
2894 case SpvExecutionModeOutputPoints
:
2895 return 0; /* GL_POINTS */
2896 case SpvExecutionModeInputLines
:
2897 return 1; /* GL_LINES */
2898 case SpvExecutionModeInputLinesAdjacency
:
2899 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
2900 case SpvExecutionModeTriangles
:
2901 return 4; /* GL_TRIANGLES */
2902 case SpvExecutionModeInputTrianglesAdjacency
:
2903 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
2904 case SpvExecutionModeQuads
:
2905 return 7; /* GL_QUADS */
2906 case SpvExecutionModeIsolines
:
2907 return 0x8E7A; /* GL_ISOLINES */
2908 case SpvExecutionModeOutputLineStrip
:
2909 return 3; /* GL_LINE_STRIP */
2910 case SpvExecutionModeOutputTriangleStrip
:
2911 return 5; /* GL_TRIANGLE_STRIP */
2913 vtn_fail("Invalid primitive type");
2918 vertices_in_from_spv_execution_mode(struct vtn_builder
*b
,
2919 SpvExecutionMode mode
)
2922 case SpvExecutionModeInputPoints
:
2924 case SpvExecutionModeInputLines
:
2926 case SpvExecutionModeInputLinesAdjacency
:
2928 case SpvExecutionModeTriangles
:
2930 case SpvExecutionModeInputTrianglesAdjacency
:
2933 vtn_fail("Invalid GS input mode");
2937 static gl_shader_stage
2938 stage_for_execution_model(struct vtn_builder
*b
, SpvExecutionModel model
)
2941 case SpvExecutionModelVertex
:
2942 return MESA_SHADER_VERTEX
;
2943 case SpvExecutionModelTessellationControl
:
2944 return MESA_SHADER_TESS_CTRL
;
2945 case SpvExecutionModelTessellationEvaluation
:
2946 return MESA_SHADER_TESS_EVAL
;
2947 case SpvExecutionModelGeometry
:
2948 return MESA_SHADER_GEOMETRY
;
2949 case SpvExecutionModelFragment
:
2950 return MESA_SHADER_FRAGMENT
;
2951 case SpvExecutionModelGLCompute
:
2952 return MESA_SHADER_COMPUTE
;
2954 vtn_fail("Unsupported execution model");
2958 #define spv_check_supported(name, cap) do { \
2959 if (!(b->options && b->options->caps.name)) \
2960 vtn_warn("Unsupported SPIR-V capability: %s", \
2961 spirv_capability_to_string(cap)); \
2965 vtn_handle_preamble_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2966 const uint32_t *w
, unsigned count
)
2973 case SpvSourceLanguageUnknown
: lang
= "unknown"; break;
2974 case SpvSourceLanguageESSL
: lang
= "ESSL"; break;
2975 case SpvSourceLanguageGLSL
: lang
= "GLSL"; break;
2976 case SpvSourceLanguageOpenCL_C
: lang
= "OpenCL C"; break;
2977 case SpvSourceLanguageOpenCL_CPP
: lang
= "OpenCL C++"; break;
2978 case SpvSourceLanguageHLSL
: lang
= "HLSL"; break;
2981 uint32_t version
= w
[2];
2984 (count
> 3) ? vtn_value(b
, w
[3], vtn_value_type_string
)->str
: "";
2986 vtn_info("Parsing SPIR-V from %s %u source file %s", lang
, version
, file
);
2990 case SpvOpSourceExtension
:
2991 case SpvOpSourceContinued
:
2992 case SpvOpExtension
:
2993 /* Unhandled, but these are for debug so that's ok. */
2996 case SpvOpCapability
: {
2997 SpvCapability cap
= w
[1];
2999 case SpvCapabilityMatrix
:
3000 case SpvCapabilityShader
:
3001 case SpvCapabilityGeometry
:
3002 case SpvCapabilityGeometryPointSize
:
3003 case SpvCapabilityUniformBufferArrayDynamicIndexing
:
3004 case SpvCapabilitySampledImageArrayDynamicIndexing
:
3005 case SpvCapabilityStorageBufferArrayDynamicIndexing
:
3006 case SpvCapabilityStorageImageArrayDynamicIndexing
:
3007 case SpvCapabilityImageRect
:
3008 case SpvCapabilitySampledRect
:
3009 case SpvCapabilitySampled1D
:
3010 case SpvCapabilityImage1D
:
3011 case SpvCapabilitySampledCubeArray
:
3012 case SpvCapabilityImageCubeArray
:
3013 case SpvCapabilitySampledBuffer
:
3014 case SpvCapabilityImageBuffer
:
3015 case SpvCapabilityImageQuery
:
3016 case SpvCapabilityDerivativeControl
:
3017 case SpvCapabilityInterpolationFunction
:
3018 case SpvCapabilityMultiViewport
:
3019 case SpvCapabilitySampleRateShading
:
3020 case SpvCapabilityClipDistance
:
3021 case SpvCapabilityCullDistance
:
3022 case SpvCapabilityInputAttachment
:
3023 case SpvCapabilityImageGatherExtended
:
3024 case SpvCapabilityStorageImageExtendedFormats
:
3027 case SpvCapabilityGeometryStreams
:
3028 case SpvCapabilityLinkage
:
3029 case SpvCapabilityVector16
:
3030 case SpvCapabilityFloat16Buffer
:
3031 case SpvCapabilityFloat16
:
3032 case SpvCapabilityInt64Atomics
:
3033 case SpvCapabilityAtomicStorage
:
3034 case SpvCapabilityInt16
:
3035 case SpvCapabilityStorageImageMultisample
:
3036 case SpvCapabilityInt8
:
3037 case SpvCapabilitySparseResidency
:
3038 case SpvCapabilityMinLod
:
3039 case SpvCapabilityTransformFeedback
:
3040 vtn_warn("Unsupported SPIR-V capability: %s",
3041 spirv_capability_to_string(cap
));
3044 case SpvCapabilityFloat64
:
3045 spv_check_supported(float64
, cap
);
3047 case SpvCapabilityInt64
:
3048 spv_check_supported(int64
, cap
);
3051 case SpvCapabilityAddresses
:
3052 case SpvCapabilityKernel
:
3053 case SpvCapabilityImageBasic
:
3054 case SpvCapabilityImageReadWrite
:
3055 case SpvCapabilityImageMipmap
:
3056 case SpvCapabilityPipes
:
3057 case SpvCapabilityGroups
:
3058 case SpvCapabilityDeviceEnqueue
:
3059 case SpvCapabilityLiteralSampler
:
3060 case SpvCapabilityGenericPointer
:
3061 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
3062 spirv_capability_to_string(cap
));
3065 case SpvCapabilityImageMSArray
:
3066 spv_check_supported(image_ms_array
, cap
);
3069 case SpvCapabilityTessellation
:
3070 case SpvCapabilityTessellationPointSize
:
3071 spv_check_supported(tessellation
, cap
);
3074 case SpvCapabilityDrawParameters
:
3075 spv_check_supported(draw_parameters
, cap
);
3078 case SpvCapabilityStorageImageReadWithoutFormat
:
3079 spv_check_supported(image_read_without_format
, cap
);
3082 case SpvCapabilityStorageImageWriteWithoutFormat
:
3083 spv_check_supported(image_write_without_format
, cap
);
3086 case SpvCapabilityMultiView
:
3087 spv_check_supported(multiview
, cap
);
3090 case SpvCapabilityVariablePointersStorageBuffer
:
3091 case SpvCapabilityVariablePointers
:
3092 spv_check_supported(variable_pointers
, cap
);
3095 case SpvCapabilityStorageUniformBufferBlock16
:
3096 case SpvCapabilityStorageUniform16
:
3097 case SpvCapabilityStoragePushConstant16
:
3098 case SpvCapabilityStorageInputOutput16
:
3099 spv_check_supported(storage_16bit
, cap
);
3103 vtn_fail("Unhandled capability");
3108 case SpvOpExtInstImport
:
3109 vtn_handle_extension(b
, opcode
, w
, count
);
3112 case SpvOpMemoryModel
:
3113 vtn_assert(w
[1] == SpvAddressingModelLogical
);
3114 vtn_assert(w
[2] == SpvMemoryModelSimple
||
3115 w
[2] == SpvMemoryModelGLSL450
);
3118 case SpvOpEntryPoint
: {
3119 struct vtn_value
*entry_point
= &b
->values
[w
[2]];
3120 /* Let this be a name label regardless */
3121 unsigned name_words
;
3122 entry_point
->name
= vtn_string_literal(b
, &w
[3], count
- 3, &name_words
);
3124 if (strcmp(entry_point
->name
, b
->entry_point_name
) != 0 ||
3125 stage_for_execution_model(b
, w
[1]) != b
->entry_point_stage
)
3128 vtn_assert(b
->entry_point
== NULL
);
3129 b
->entry_point
= entry_point
;
3134 vtn_push_value(b
, w
[1], vtn_value_type_string
)->str
=
3135 vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
3139 b
->values
[w
[1]].name
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
3142 case SpvOpMemberName
:
3146 case SpvOpExecutionMode
:
3147 case SpvOpDecorationGroup
:
3149 case SpvOpMemberDecorate
:
3150 case SpvOpGroupDecorate
:
3151 case SpvOpGroupMemberDecorate
:
3152 vtn_handle_decoration(b
, opcode
, w
, count
);
3156 return false; /* End of preamble */
3163 vtn_handle_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
3164 const struct vtn_decoration
*mode
, void *data
)
3166 vtn_assert(b
->entry_point
== entry_point
);
3168 switch(mode
->exec_mode
) {
3169 case SpvExecutionModeOriginUpperLeft
:
3170 case SpvExecutionModeOriginLowerLeft
:
3171 b
->origin_upper_left
=
3172 (mode
->exec_mode
== SpvExecutionModeOriginUpperLeft
);
3175 case SpvExecutionModeEarlyFragmentTests
:
3176 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3177 b
->shader
->info
.fs
.early_fragment_tests
= true;
3180 case SpvExecutionModeInvocations
:
3181 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3182 b
->shader
->info
.gs
.invocations
= MAX2(1, mode
->literals
[0]);
3185 case SpvExecutionModeDepthReplacing
:
3186 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3187 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_ANY
;
3189 case SpvExecutionModeDepthGreater
:
3190 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3191 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_GREATER
;
3193 case SpvExecutionModeDepthLess
:
3194 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3195 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_LESS
;
3197 case SpvExecutionModeDepthUnchanged
:
3198 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3199 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
3202 case SpvExecutionModeLocalSize
:
3203 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
3204 b
->shader
->info
.cs
.local_size
[0] = mode
->literals
[0];
3205 b
->shader
->info
.cs
.local_size
[1] = mode
->literals
[1];
3206 b
->shader
->info
.cs
.local_size
[2] = mode
->literals
[2];
3208 case SpvExecutionModeLocalSizeHint
:
3209 break; /* Nothing to do with this */
3211 case SpvExecutionModeOutputVertices
:
3212 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3213 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
3214 b
->shader
->info
.tess
.tcs_vertices_out
= mode
->literals
[0];
3216 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3217 b
->shader
->info
.gs
.vertices_out
= mode
->literals
[0];
3221 case SpvExecutionModeInputPoints
:
3222 case SpvExecutionModeInputLines
:
3223 case SpvExecutionModeInputLinesAdjacency
:
3224 case SpvExecutionModeTriangles
:
3225 case SpvExecutionModeInputTrianglesAdjacency
:
3226 case SpvExecutionModeQuads
:
3227 case SpvExecutionModeIsolines
:
3228 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3229 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
3230 b
->shader
->info
.tess
.primitive_mode
=
3231 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
3233 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3234 b
->shader
->info
.gs
.vertices_in
=
3235 vertices_in_from_spv_execution_mode(b
, mode
->exec_mode
);
3239 case SpvExecutionModeOutputPoints
:
3240 case SpvExecutionModeOutputLineStrip
:
3241 case SpvExecutionModeOutputTriangleStrip
:
3242 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3243 b
->shader
->info
.gs
.output_primitive
=
3244 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
3247 case SpvExecutionModeSpacingEqual
:
3248 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3249 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3250 b
->shader
->info
.tess
.spacing
= TESS_SPACING_EQUAL
;
3252 case SpvExecutionModeSpacingFractionalEven
:
3253 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3254 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3255 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_EVEN
;
3257 case SpvExecutionModeSpacingFractionalOdd
:
3258 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3259 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3260 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_ODD
;
3262 case SpvExecutionModeVertexOrderCw
:
3263 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3264 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3265 b
->shader
->info
.tess
.ccw
= false;
3267 case SpvExecutionModeVertexOrderCcw
:
3268 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3269 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3270 b
->shader
->info
.tess
.ccw
= true;
3272 case SpvExecutionModePointMode
:
3273 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3274 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3275 b
->shader
->info
.tess
.point_mode
= true;
3278 case SpvExecutionModePixelCenterInteger
:
3279 b
->pixel_center_integer
= true;
3282 case SpvExecutionModeXfb
:
3283 vtn_fail("Unhandled execution mode");
3286 case SpvExecutionModeVecTypeHint
:
3287 case SpvExecutionModeContractionOff
:
3291 vtn_fail("Unhandled execution mode");
3296 vtn_handle_variable_or_type_instruction(struct vtn_builder
*b
, SpvOp opcode
,
3297 const uint32_t *w
, unsigned count
)
3299 vtn_set_instruction_result_type(b
, opcode
, w
, count
);
3303 case SpvOpSourceContinued
:
3304 case SpvOpSourceExtension
:
3305 case SpvOpExtension
:
3306 case SpvOpCapability
:
3307 case SpvOpExtInstImport
:
3308 case SpvOpMemoryModel
:
3309 case SpvOpEntryPoint
:
3310 case SpvOpExecutionMode
:
3313 case SpvOpMemberName
:
3314 case SpvOpDecorationGroup
:
3316 case SpvOpMemberDecorate
:
3317 case SpvOpGroupDecorate
:
3318 case SpvOpGroupMemberDecorate
:
3319 vtn_fail("Invalid opcode types and variables section");
3325 case SpvOpTypeFloat
:
3326 case SpvOpTypeVector
:
3327 case SpvOpTypeMatrix
:
3328 case SpvOpTypeImage
:
3329 case SpvOpTypeSampler
:
3330 case SpvOpTypeSampledImage
:
3331 case SpvOpTypeArray
:
3332 case SpvOpTypeRuntimeArray
:
3333 case SpvOpTypeStruct
:
3334 case SpvOpTypeOpaque
:
3335 case SpvOpTypePointer
:
3336 case SpvOpTypeFunction
:
3337 case SpvOpTypeEvent
:
3338 case SpvOpTypeDeviceEvent
:
3339 case SpvOpTypeReserveId
:
3340 case SpvOpTypeQueue
:
3342 vtn_handle_type(b
, opcode
, w
, count
);
3345 case SpvOpConstantTrue
:
3346 case SpvOpConstantFalse
:
3348 case SpvOpConstantComposite
:
3349 case SpvOpConstantSampler
:
3350 case SpvOpConstantNull
:
3351 case SpvOpSpecConstantTrue
:
3352 case SpvOpSpecConstantFalse
:
3353 case SpvOpSpecConstant
:
3354 case SpvOpSpecConstantComposite
:
3355 case SpvOpSpecConstantOp
:
3356 vtn_handle_constant(b
, opcode
, w
, count
);
3361 vtn_handle_variables(b
, opcode
, w
, count
);
3365 return false; /* End of preamble */
3372 vtn_handle_body_instruction(struct vtn_builder
*b
, SpvOp opcode
,
3373 const uint32_t *w
, unsigned count
)
3379 case SpvOpLoopMerge
:
3380 case SpvOpSelectionMerge
:
3381 /* This is handled by cfg pre-pass and walk_blocks */
3385 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
3386 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
3391 vtn_handle_extension(b
, opcode
, w
, count
);
3397 case SpvOpCopyMemory
:
3398 case SpvOpCopyMemorySized
:
3399 case SpvOpAccessChain
:
3400 case SpvOpPtrAccessChain
:
3401 case SpvOpInBoundsAccessChain
:
3402 case SpvOpArrayLength
:
3403 vtn_handle_variables(b
, opcode
, w
, count
);
3406 case SpvOpFunctionCall
:
3407 vtn_handle_function_call(b
, opcode
, w
, count
);
3410 case SpvOpSampledImage
:
3412 case SpvOpImageSampleImplicitLod
:
3413 case SpvOpImageSampleExplicitLod
:
3414 case SpvOpImageSampleDrefImplicitLod
:
3415 case SpvOpImageSampleDrefExplicitLod
:
3416 case SpvOpImageSampleProjImplicitLod
:
3417 case SpvOpImageSampleProjExplicitLod
:
3418 case SpvOpImageSampleProjDrefImplicitLod
:
3419 case SpvOpImageSampleProjDrefExplicitLod
:
3420 case SpvOpImageFetch
:
3421 case SpvOpImageGather
:
3422 case SpvOpImageDrefGather
:
3423 case SpvOpImageQuerySizeLod
:
3424 case SpvOpImageQueryLod
:
3425 case SpvOpImageQueryLevels
:
3426 case SpvOpImageQuerySamples
:
3427 vtn_handle_texture(b
, opcode
, w
, count
);
3430 case SpvOpImageRead
:
3431 case SpvOpImageWrite
:
3432 case SpvOpImageTexelPointer
:
3433 vtn_handle_image(b
, opcode
, w
, count
);
3436 case SpvOpImageQuerySize
: {
3437 struct vtn_pointer
*image
=
3438 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
3439 if (image
->mode
== vtn_variable_mode_image
) {
3440 vtn_handle_image(b
, opcode
, w
, count
);
3442 vtn_assert(image
->mode
== vtn_variable_mode_sampler
);
3443 vtn_handle_texture(b
, opcode
, w
, count
);
3448 case SpvOpAtomicLoad
:
3449 case SpvOpAtomicExchange
:
3450 case SpvOpAtomicCompareExchange
:
3451 case SpvOpAtomicCompareExchangeWeak
:
3452 case SpvOpAtomicIIncrement
:
3453 case SpvOpAtomicIDecrement
:
3454 case SpvOpAtomicIAdd
:
3455 case SpvOpAtomicISub
:
3456 case SpvOpAtomicSMin
:
3457 case SpvOpAtomicUMin
:
3458 case SpvOpAtomicSMax
:
3459 case SpvOpAtomicUMax
:
3460 case SpvOpAtomicAnd
:
3462 case SpvOpAtomicXor
: {
3463 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[3]);
3464 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
3465 vtn_handle_image(b
, opcode
, w
, count
);
3467 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
3468 vtn_handle_ssbo_or_shared_atomic(b
, opcode
, w
, count
);
3473 case SpvOpAtomicStore
: {
3474 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[1]);
3475 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
3476 vtn_handle_image(b
, opcode
, w
, count
);
3478 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
3479 vtn_handle_ssbo_or_shared_atomic(b
, opcode
, w
, count
);
3485 /* Handle OpSelect up-front here because it needs to be able to handle
3486 * pointers and not just regular vectors and scalars.
3488 struct vtn_value
*res_val
= vtn_untyped_value(b
, w
[2]);
3489 struct vtn_value
*sel_val
= vtn_untyped_value(b
, w
[3]);
3490 struct vtn_value
*obj1_val
= vtn_untyped_value(b
, w
[4]);
3491 struct vtn_value
*obj2_val
= vtn_untyped_value(b
, w
[5]);
3493 const struct glsl_type
*sel_type
;
3494 switch (res_val
->type
->base_type
) {
3495 case vtn_base_type_scalar
:
3496 sel_type
= glsl_bool_type();
3498 case vtn_base_type_vector
:
3499 sel_type
= glsl_vector_type(GLSL_TYPE_BOOL
, res_val
->type
->length
);
3501 case vtn_base_type_pointer
:
3502 /* We need to have actual storage for pointer types */
3503 vtn_fail_if(res_val
->type
->type
== NULL
,
3504 "Invalid pointer result type for OpSelect");
3505 sel_type
= glsl_bool_type();
3508 vtn_fail("Result type of OpSelect must be a scalar, vector, or pointer");
3511 vtn_fail_if(sel_val
->type
->type
!= sel_type
,
3512 "Condition type of OpSelect must be a scalar or vector of "
3513 "Boolean type. It must have the same number of components "
3516 vtn_fail_if(obj1_val
->type
!= res_val
->type
||
3517 obj2_val
->type
!= res_val
->type
,
3518 "Object types must match the result type in OpSelect");
3520 struct vtn_type
*res_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
3521 struct vtn_ssa_value
*ssa
= vtn_create_ssa_value(b
, res_type
->type
);
3522 ssa
->def
= nir_bcsel(&b
->nb
, vtn_ssa_value(b
, w
[3])->def
,
3523 vtn_ssa_value(b
, w
[4])->def
,
3524 vtn_ssa_value(b
, w
[5])->def
);
3525 vtn_push_ssa(b
, w
[2], res_type
, ssa
);
3534 case SpvOpConvertFToU
:
3535 case SpvOpConvertFToS
:
3536 case SpvOpConvertSToF
:
3537 case SpvOpConvertUToF
:
3541 case SpvOpQuantizeToF16
:
3542 case SpvOpConvertPtrToU
:
3543 case SpvOpConvertUToPtr
:
3544 case SpvOpPtrCastToGeneric
:
3545 case SpvOpGenericCastToPtr
:
3551 case SpvOpSignBitSet
:
3552 case SpvOpLessOrGreater
:
3554 case SpvOpUnordered
:
3569 case SpvOpVectorTimesScalar
:
3571 case SpvOpIAddCarry
:
3572 case SpvOpISubBorrow
:
3573 case SpvOpUMulExtended
:
3574 case SpvOpSMulExtended
:
3575 case SpvOpShiftRightLogical
:
3576 case SpvOpShiftRightArithmetic
:
3577 case SpvOpShiftLeftLogical
:
3578 case SpvOpLogicalEqual
:
3579 case SpvOpLogicalNotEqual
:
3580 case SpvOpLogicalOr
:
3581 case SpvOpLogicalAnd
:
3582 case SpvOpLogicalNot
:
3583 case SpvOpBitwiseOr
:
3584 case SpvOpBitwiseXor
:
3585 case SpvOpBitwiseAnd
:
3587 case SpvOpFOrdEqual
:
3588 case SpvOpFUnordEqual
:
3589 case SpvOpINotEqual
:
3590 case SpvOpFOrdNotEqual
:
3591 case SpvOpFUnordNotEqual
:
3592 case SpvOpULessThan
:
3593 case SpvOpSLessThan
:
3594 case SpvOpFOrdLessThan
:
3595 case SpvOpFUnordLessThan
:
3596 case SpvOpUGreaterThan
:
3597 case SpvOpSGreaterThan
:
3598 case SpvOpFOrdGreaterThan
:
3599 case SpvOpFUnordGreaterThan
:
3600 case SpvOpULessThanEqual
:
3601 case SpvOpSLessThanEqual
:
3602 case SpvOpFOrdLessThanEqual
:
3603 case SpvOpFUnordLessThanEqual
:
3604 case SpvOpUGreaterThanEqual
:
3605 case SpvOpSGreaterThanEqual
:
3606 case SpvOpFOrdGreaterThanEqual
:
3607 case SpvOpFUnordGreaterThanEqual
:
3613 case SpvOpFwidthFine
:
3614 case SpvOpDPdxCoarse
:
3615 case SpvOpDPdyCoarse
:
3616 case SpvOpFwidthCoarse
:
3617 case SpvOpBitFieldInsert
:
3618 case SpvOpBitFieldSExtract
:
3619 case SpvOpBitFieldUExtract
:
3620 case SpvOpBitReverse
:
3622 case SpvOpTranspose
:
3623 case SpvOpOuterProduct
:
3624 case SpvOpMatrixTimesScalar
:
3625 case SpvOpVectorTimesMatrix
:
3626 case SpvOpMatrixTimesVector
:
3627 case SpvOpMatrixTimesMatrix
:
3628 vtn_handle_alu(b
, opcode
, w
, count
);
3631 case SpvOpVectorExtractDynamic
:
3632 case SpvOpVectorInsertDynamic
:
3633 case SpvOpVectorShuffle
:
3634 case SpvOpCompositeConstruct
:
3635 case SpvOpCompositeExtract
:
3636 case SpvOpCompositeInsert
:
3637 case SpvOpCopyObject
:
3638 vtn_handle_composite(b
, opcode
, w
, count
);
3641 case SpvOpEmitVertex
:
3642 case SpvOpEndPrimitive
:
3643 case SpvOpEmitStreamVertex
:
3644 case SpvOpEndStreamPrimitive
:
3645 case SpvOpControlBarrier
:
3646 case SpvOpMemoryBarrier
:
3647 vtn_handle_barrier(b
, opcode
, w
, count
);
3651 vtn_fail("Unhandled opcode");
3658 spirv_to_nir(const uint32_t *words
, size_t word_count
,
3659 struct nir_spirv_specialization
*spec
, unsigned num_spec
,
3660 gl_shader_stage stage
, const char *entry_point_name
,
3661 const struct spirv_to_nir_options
*options
,
3662 const nir_shader_compiler_options
*nir_options
)
3664 /* Initialize the stn_builder object */
3665 struct vtn_builder
*b
= rzalloc(NULL
, struct vtn_builder
);
3670 exec_list_make_empty(&b
->functions
);
3671 b
->entry_point_stage
= stage
;
3672 b
->entry_point_name
= entry_point_name
;
3673 b
->options
= options
;
3675 /* See also _vtn_fail() */
3676 if (setjmp(b
->fail_jump
)) {
3681 const uint32_t *word_end
= words
+ word_count
;
3683 /* Handle the SPIR-V header (first 4 dwords) */
3684 vtn_assert(word_count
> 5);
3686 vtn_assert(words
[0] == SpvMagicNumber
);
3687 vtn_assert(words
[1] >= 0x10000);
3688 /* words[2] == generator magic */
3689 unsigned value_id_bound
= words
[3];
3690 vtn_assert(words
[4] == 0);
3694 b
->value_id_bound
= value_id_bound
;
3695 b
->values
= rzalloc_array(b
, struct vtn_value
, value_id_bound
);
3697 /* Handle all the preamble instructions */
3698 words
= vtn_foreach_instruction(b
, words
, word_end
,
3699 vtn_handle_preamble_instruction
);
3701 if (b
->entry_point
== NULL
) {
3702 vtn_fail("Entry point not found");
3707 b
->shader
= nir_shader_create(b
, stage
, nir_options
, NULL
);
3709 /* Set shader info defaults */
3710 b
->shader
->info
.gs
.invocations
= 1;
3712 /* Parse execution modes */
3713 vtn_foreach_execution_mode(b
, b
->entry_point
,
3714 vtn_handle_execution_mode
, NULL
);
3716 b
->specializations
= spec
;
3717 b
->num_specializations
= num_spec
;
3719 /* Handle all variable, type, and constant instructions */
3720 words
= vtn_foreach_instruction(b
, words
, word_end
,
3721 vtn_handle_variable_or_type_instruction
);
3723 /* Set types on all vtn_values */
3724 vtn_foreach_instruction(b
, words
, word_end
, vtn_set_instruction_result_type
);
3726 vtn_build_cfg(b
, words
, word_end
);
3728 assert(b
->entry_point
->value_type
== vtn_value_type_function
);
3729 b
->entry_point
->func
->referenced
= true;
3734 foreach_list_typed(struct vtn_function
, func
, node
, &b
->functions
) {
3735 if (func
->referenced
&& !func
->emitted
) {
3736 b
->const_table
= _mesa_hash_table_create(b
, _mesa_hash_pointer
,
3737 _mesa_key_pointer_equal
);
3739 vtn_function_emit(b
, func
, vtn_handle_body_instruction
);
3745 vtn_assert(b
->entry_point
->value_type
== vtn_value_type_function
);
3746 nir_function
*entry_point
= b
->entry_point
->func
->impl
->function
;
3747 vtn_assert(entry_point
);
3749 /* Unparent the shader from the vtn_builder before we delete the builder */
3750 ralloc_steal(NULL
, b
->shader
);