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
->const_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();
893 const bool signedness
= w
[3];
894 val
->type
->base_type
= vtn_base_type_scalar
;
897 val
->type
->type
= (signedness
? glsl_int64_t_type() : glsl_uint64_t_type());
900 val
->type
->type
= (signedness
? glsl_int_type() : glsl_uint_type());
903 val
->type
->type
= (signedness
? glsl_int16_t_type() : glsl_uint16_t_type());
906 vtn_fail("Invalid int bit size");
911 case SpvOpTypeFloat
: {
913 val
->type
->base_type
= vtn_base_type_scalar
;
916 val
->type
->type
= glsl_float16_t_type();
919 val
->type
->type
= glsl_float_type();
922 val
->type
->type
= glsl_double_type();
925 vtn_fail("Invalid float bit size");
930 case SpvOpTypeVector
: {
931 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
932 unsigned elems
= w
[3];
934 vtn_assert(glsl_type_is_scalar(base
->type
));
935 val
->type
->base_type
= vtn_base_type_vector
;
936 val
->type
->type
= glsl_vector_type(glsl_get_base_type(base
->type
), elems
);
937 val
->type
->stride
= glsl_get_bit_size(base
->type
) / 8;
938 val
->type
->array_element
= base
;
942 case SpvOpTypeMatrix
: {
943 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
944 unsigned columns
= w
[3];
946 vtn_assert(glsl_type_is_vector(base
->type
));
947 val
->type
->base_type
= vtn_base_type_matrix
;
948 val
->type
->type
= glsl_matrix_type(glsl_get_base_type(base
->type
),
949 glsl_get_vector_elements(base
->type
),
951 vtn_assert(!glsl_type_is_error(val
->type
->type
));
952 val
->type
->length
= columns
;
953 val
->type
->array_element
= base
;
954 val
->type
->row_major
= false;
955 val
->type
->stride
= 0;
959 case SpvOpTypeRuntimeArray
:
960 case SpvOpTypeArray
: {
961 struct vtn_type
*array_element
=
962 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
964 if (opcode
== SpvOpTypeRuntimeArray
) {
965 /* A length of 0 is used to denote unsized arrays */
966 val
->type
->length
= 0;
969 vtn_value(b
, w
[3], vtn_value_type_constant
)->constant
->values
[0].u32
[0];
972 val
->type
->base_type
= vtn_base_type_array
;
973 val
->type
->type
= glsl_array_type(array_element
->type
, val
->type
->length
);
974 val
->type
->array_element
= array_element
;
975 val
->type
->stride
= 0;
979 case SpvOpTypeStruct
: {
980 unsigned num_fields
= count
- 2;
981 val
->type
->base_type
= vtn_base_type_struct
;
982 val
->type
->length
= num_fields
;
983 val
->type
->members
= ralloc_array(b
, struct vtn_type
*, num_fields
);
984 val
->type
->offsets
= ralloc_array(b
, unsigned, num_fields
);
986 NIR_VLA(struct glsl_struct_field
, fields
, count
);
987 for (unsigned i
= 0; i
< num_fields
; i
++) {
988 val
->type
->members
[i
] =
989 vtn_value(b
, w
[i
+ 2], vtn_value_type_type
)->type
;
990 fields
[i
] = (struct glsl_struct_field
) {
991 .type
= val
->type
->members
[i
]->type
,
992 .name
= ralloc_asprintf(b
, "field%d", i
),
997 struct member_decoration_ctx ctx
= {
998 .num_fields
= num_fields
,
1003 vtn_foreach_decoration(b
, val
, struct_member_decoration_cb
, &ctx
);
1004 vtn_foreach_decoration(b
, val
, struct_member_matrix_stride_cb
, &ctx
);
1006 const char *name
= val
->name
? val
->name
: "struct";
1008 val
->type
->type
= glsl_struct_type(fields
, num_fields
, name
);
1012 case SpvOpTypeFunction
: {
1013 val
->type
->base_type
= vtn_base_type_function
;
1014 val
->type
->type
= NULL
;
1016 val
->type
->return_type
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1018 const unsigned num_params
= count
- 3;
1019 val
->type
->length
= num_params
;
1020 val
->type
->params
= ralloc_array(b
, struct vtn_type
*, num_params
);
1021 for (unsigned i
= 0; i
< count
- 3; i
++) {
1022 val
->type
->params
[i
] =
1023 vtn_value(b
, w
[i
+ 3], vtn_value_type_type
)->type
;
1028 case SpvOpTypePointer
: {
1029 SpvStorageClass storage_class
= w
[2];
1030 struct vtn_type
*deref_type
=
1031 vtn_value(b
, w
[3], vtn_value_type_type
)->type
;
1033 val
->type
->base_type
= vtn_base_type_pointer
;
1034 val
->type
->storage_class
= storage_class
;
1035 val
->type
->deref
= deref_type
;
1037 if (storage_class
== SpvStorageClassUniform
||
1038 storage_class
== SpvStorageClassStorageBuffer
) {
1039 /* These can actually be stored to nir_variables and used as SSA
1040 * values so they need a real glsl_type.
1042 val
->type
->type
= glsl_vector_type(GLSL_TYPE_UINT
, 2);
1045 if (storage_class
== SpvStorageClassWorkgroup
&&
1046 b
->options
->lower_workgroup_access_to_offsets
) {
1047 uint32_t size
, align
;
1048 val
->type
->deref
= vtn_type_layout_std430(b
, val
->type
->deref
,
1050 val
->type
->length
= size
;
1051 val
->type
->align
= align
;
1052 /* These can actually be stored to nir_variables and used as SSA
1053 * values so they need a real glsl_type.
1055 val
->type
->type
= glsl_uint_type();
1060 case SpvOpTypeImage
: {
1061 val
->type
->base_type
= vtn_base_type_image
;
1063 const struct glsl_type
*sampled_type
=
1064 vtn_value(b
, w
[2], vtn_value_type_type
)->type
->type
;
1066 vtn_assert(glsl_type_is_vector_or_scalar(sampled_type
));
1068 enum glsl_sampler_dim dim
;
1069 switch ((SpvDim
)w
[3]) {
1070 case SpvDim1D
: dim
= GLSL_SAMPLER_DIM_1D
; break;
1071 case SpvDim2D
: dim
= GLSL_SAMPLER_DIM_2D
; break;
1072 case SpvDim3D
: dim
= GLSL_SAMPLER_DIM_3D
; break;
1073 case SpvDimCube
: dim
= GLSL_SAMPLER_DIM_CUBE
; break;
1074 case SpvDimRect
: dim
= GLSL_SAMPLER_DIM_RECT
; break;
1075 case SpvDimBuffer
: dim
= GLSL_SAMPLER_DIM_BUF
; break;
1076 case SpvDimSubpassData
: dim
= GLSL_SAMPLER_DIM_SUBPASS
; break;
1078 vtn_fail("Invalid SPIR-V Sampler dimension");
1081 bool is_shadow
= w
[4];
1082 bool is_array
= w
[5];
1083 bool multisampled
= w
[6];
1084 unsigned sampled
= w
[7];
1085 SpvImageFormat format
= w
[8];
1088 val
->type
->access_qualifier
= w
[9];
1090 val
->type
->access_qualifier
= SpvAccessQualifierReadWrite
;
1093 if (dim
== GLSL_SAMPLER_DIM_2D
)
1094 dim
= GLSL_SAMPLER_DIM_MS
;
1095 else if (dim
== GLSL_SAMPLER_DIM_SUBPASS
)
1096 dim
= GLSL_SAMPLER_DIM_SUBPASS_MS
;
1098 vtn_fail("Unsupported multisampled image type");
1101 val
->type
->image_format
= translate_image_format(b
, format
);
1104 val
->type
->sampled
= true;
1105 val
->type
->type
= glsl_sampler_type(dim
, is_shadow
, is_array
,
1106 glsl_get_base_type(sampled_type
));
1107 } else if (sampled
== 2) {
1108 vtn_assert(!is_shadow
);
1109 val
->type
->sampled
= false;
1110 val
->type
->type
= glsl_image_type(dim
, is_array
,
1111 glsl_get_base_type(sampled_type
));
1113 vtn_fail("We need to know if the image will be sampled");
1118 case SpvOpTypeSampledImage
:
1119 val
->type
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1122 case SpvOpTypeSampler
:
1123 /* The actual sampler type here doesn't really matter. It gets
1124 * thrown away the moment you combine it with an image. What really
1125 * matters is that it's a sampler type as opposed to an integer type
1126 * so the backend knows what to do.
1128 val
->type
->base_type
= vtn_base_type_sampler
;
1129 val
->type
->type
= glsl_bare_sampler_type();
1132 case SpvOpTypeOpaque
:
1133 case SpvOpTypeEvent
:
1134 case SpvOpTypeDeviceEvent
:
1135 case SpvOpTypeReserveId
:
1136 case SpvOpTypeQueue
:
1139 vtn_fail("Unhandled opcode");
1142 vtn_foreach_decoration(b
, val
, type_decoration_cb
, NULL
);
1145 static nir_constant
*
1146 vtn_null_constant(struct vtn_builder
*b
, const struct glsl_type
*type
)
1148 nir_constant
*c
= rzalloc(b
, nir_constant
);
1150 /* For pointers and other typeless things, we have to return something but
1151 * it doesn't matter what.
1156 switch (glsl_get_base_type(type
)) {
1158 case GLSL_TYPE_UINT
:
1159 case GLSL_TYPE_INT16
:
1160 case GLSL_TYPE_UINT16
:
1161 case GLSL_TYPE_INT64
:
1162 case GLSL_TYPE_UINT64
:
1163 case GLSL_TYPE_BOOL
:
1164 case GLSL_TYPE_FLOAT
:
1165 case GLSL_TYPE_FLOAT16
:
1166 case GLSL_TYPE_DOUBLE
:
1167 /* Nothing to do here. It's already initialized to zero */
1170 case GLSL_TYPE_ARRAY
:
1171 vtn_assert(glsl_get_length(type
) > 0);
1172 c
->num_elements
= glsl_get_length(type
);
1173 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1175 c
->elements
[0] = vtn_null_constant(b
, glsl_get_array_element(type
));
1176 for (unsigned i
= 1; i
< c
->num_elements
; i
++)
1177 c
->elements
[i
] = c
->elements
[0];
1180 case GLSL_TYPE_STRUCT
:
1181 c
->num_elements
= glsl_get_length(type
);
1182 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1184 for (unsigned i
= 0; i
< c
->num_elements
; i
++) {
1185 c
->elements
[i
] = vtn_null_constant(b
, glsl_get_struct_field(type
, i
));
1190 vtn_fail("Invalid type for null constant");
1197 spec_constant_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*v
,
1198 int member
, const struct vtn_decoration
*dec
,
1201 vtn_assert(member
== -1);
1202 if (dec
->decoration
!= SpvDecorationSpecId
)
1205 struct spec_constant_value
*const_value
= data
;
1207 for (unsigned i
= 0; i
< b
->num_specializations
; i
++) {
1208 if (b
->specializations
[i
].id
== dec
->literals
[0]) {
1209 if (const_value
->is_double
)
1210 const_value
->data64
= b
->specializations
[i
].data64
;
1212 const_value
->data32
= b
->specializations
[i
].data32
;
1219 get_specialization(struct vtn_builder
*b
, struct vtn_value
*val
,
1220 uint32_t const_value
)
1222 struct spec_constant_value data
;
1223 data
.is_double
= false;
1224 data
.data32
= const_value
;
1225 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1230 get_specialization64(struct vtn_builder
*b
, struct vtn_value
*val
,
1231 uint64_t const_value
)
1233 struct spec_constant_value data
;
1234 data
.is_double
= true;
1235 data
.data64
= const_value
;
1236 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1241 handle_workgroup_size_decoration_cb(struct vtn_builder
*b
,
1242 struct vtn_value
*val
,
1244 const struct vtn_decoration
*dec
,
1247 vtn_assert(member
== -1);
1248 if (dec
->decoration
!= SpvDecorationBuiltIn
||
1249 dec
->literals
[0] != SpvBuiltInWorkgroupSize
)
1252 vtn_assert(val
->const_type
== glsl_vector_type(GLSL_TYPE_UINT
, 3));
1254 b
->shader
->info
.cs
.local_size
[0] = val
->constant
->values
[0].u32
[0];
1255 b
->shader
->info
.cs
.local_size
[1] = val
->constant
->values
[0].u32
[1];
1256 b
->shader
->info
.cs
.local_size
[2] = val
->constant
->values
[0].u32
[2];
1260 vtn_handle_constant(struct vtn_builder
*b
, SpvOp opcode
,
1261 const uint32_t *w
, unsigned count
)
1263 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_constant
);
1264 val
->const_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
1265 val
->constant
= rzalloc(b
, nir_constant
);
1267 case SpvOpConstantTrue
:
1268 vtn_assert(val
->const_type
== glsl_bool_type());
1269 val
->constant
->values
[0].u32
[0] = NIR_TRUE
;
1271 case SpvOpConstantFalse
:
1272 vtn_assert(val
->const_type
== glsl_bool_type());
1273 val
->constant
->values
[0].u32
[0] = NIR_FALSE
;
1276 case SpvOpSpecConstantTrue
:
1277 case SpvOpSpecConstantFalse
: {
1278 vtn_assert(val
->const_type
== glsl_bool_type());
1280 get_specialization(b
, val
, (opcode
== SpvOpSpecConstantTrue
));
1281 val
->constant
->values
[0].u32
[0] = int_val
? NIR_TRUE
: NIR_FALSE
;
1285 case SpvOpConstant
: {
1286 vtn_assert(glsl_type_is_scalar(val
->const_type
));
1287 int bit_size
= glsl_get_bit_size(val
->const_type
);
1290 val
->constant
->values
->u64
[0] = vtn_u64_literal(&w
[3]);
1293 val
->constant
->values
->u32
[0] = w
[3];
1296 val
->constant
->values
->u16
[0] = w
[3];
1299 vtn_fail("Unsupported SpvOpConstant bit size");
1303 case SpvOpSpecConstant
: {
1304 vtn_assert(glsl_type_is_scalar(val
->const_type
));
1305 val
->constant
->values
[0].u32
[0] = get_specialization(b
, val
, w
[3]);
1306 int bit_size
= glsl_get_bit_size(val
->const_type
);
1309 val
->constant
->values
[0].u64
[0] =
1310 get_specialization64(b
, val
, vtn_u64_literal(&w
[3]));
1313 val
->constant
->values
[0].u32
[0] = get_specialization(b
, val
, w
[3]);
1316 val
->constant
->values
[0].u16
[0] = get_specialization(b
, val
, w
[3]);
1319 vtn_fail("Unsupported SpvOpSpecConstant bit size");
1323 case SpvOpSpecConstantComposite
:
1324 case SpvOpConstantComposite
: {
1325 unsigned elem_count
= count
- 3;
1326 nir_constant
**elems
= ralloc_array(b
, nir_constant
*, elem_count
);
1327 for (unsigned i
= 0; i
< elem_count
; i
++)
1328 elems
[i
] = vtn_value(b
, w
[i
+ 3], vtn_value_type_constant
)->constant
;
1330 switch (glsl_get_base_type(val
->const_type
)) {
1331 case GLSL_TYPE_UINT
:
1333 case GLSL_TYPE_UINT16
:
1334 case GLSL_TYPE_INT16
:
1335 case GLSL_TYPE_UINT64
:
1336 case GLSL_TYPE_INT64
:
1337 case GLSL_TYPE_FLOAT
:
1338 case GLSL_TYPE_FLOAT16
:
1339 case GLSL_TYPE_BOOL
:
1340 case GLSL_TYPE_DOUBLE
: {
1341 int bit_size
= glsl_get_bit_size(val
->const_type
);
1342 if (glsl_type_is_matrix(val
->const_type
)) {
1343 vtn_assert(glsl_get_matrix_columns(val
->const_type
) == elem_count
);
1344 for (unsigned i
= 0; i
< elem_count
; i
++)
1345 val
->constant
->values
[i
] = elems
[i
]->values
[0];
1347 vtn_assert(glsl_type_is_vector(val
->const_type
));
1348 vtn_assert(glsl_get_vector_elements(val
->const_type
) == elem_count
);
1349 for (unsigned i
= 0; i
< elem_count
; i
++) {
1352 val
->constant
->values
[0].u64
[i
] = elems
[i
]->values
[0].u64
[0];
1355 val
->constant
->values
[0].u32
[i
] = elems
[i
]->values
[0].u32
[0];
1358 val
->constant
->values
[0].u16
[i
] = elems
[i
]->values
[0].u16
[0];
1361 vtn_fail("Invalid SpvOpConstantComposite bit size");
1368 case GLSL_TYPE_STRUCT
:
1369 case GLSL_TYPE_ARRAY
:
1370 ralloc_steal(val
->constant
, elems
);
1371 val
->constant
->num_elements
= elem_count
;
1372 val
->constant
->elements
= elems
;
1376 vtn_fail("Unsupported type for constants");
1381 case SpvOpSpecConstantOp
: {
1382 SpvOp opcode
= get_specialization(b
, val
, w
[3]);
1384 case SpvOpVectorShuffle
: {
1385 struct vtn_value
*v0
= &b
->values
[w
[4]];
1386 struct vtn_value
*v1
= &b
->values
[w
[5]];
1388 vtn_assert(v0
->value_type
== vtn_value_type_constant
||
1389 v0
->value_type
== vtn_value_type_undef
);
1390 vtn_assert(v1
->value_type
== vtn_value_type_constant
||
1391 v1
->value_type
== vtn_value_type_undef
);
1393 unsigned len0
= v0
->value_type
== vtn_value_type_constant
?
1394 glsl_get_vector_elements(v0
->const_type
) :
1395 glsl_get_vector_elements(v0
->type
->type
);
1396 unsigned len1
= v1
->value_type
== vtn_value_type_constant
?
1397 glsl_get_vector_elements(v1
->const_type
) :
1398 glsl_get_vector_elements(v1
->type
->type
);
1400 vtn_assert(len0
+ len1
< 16);
1402 unsigned bit_size
= glsl_get_bit_size(val
->const_type
);
1403 unsigned bit_size0
= v0
->value_type
== vtn_value_type_constant
?
1404 glsl_get_bit_size(v0
->const_type
) :
1405 glsl_get_bit_size(v0
->type
->type
);
1406 unsigned bit_size1
= v1
->value_type
== vtn_value_type_constant
?
1407 glsl_get_bit_size(v1
->const_type
) :
1408 glsl_get_bit_size(v1
->type
->type
);
1410 vtn_assert(bit_size
== bit_size0
&& bit_size
== bit_size1
);
1411 (void)bit_size0
; (void)bit_size1
;
1413 if (bit_size
== 64) {
1415 if (v0
->value_type
== vtn_value_type_constant
) {
1416 for (unsigned i
= 0; i
< len0
; i
++)
1417 u64
[i
] = v0
->constant
->values
[0].u64
[i
];
1419 if (v1
->value_type
== vtn_value_type_constant
) {
1420 for (unsigned i
= 0; i
< len1
; i
++)
1421 u64
[len0
+ i
] = v1
->constant
->values
[0].u64
[i
];
1424 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1425 uint32_t comp
= w
[i
+ 6];
1426 /* If component is not used, set the value to a known constant
1427 * to detect if it is wrongly used.
1429 if (comp
== (uint32_t)-1)
1430 val
->constant
->values
[0].u64
[j
] = 0xdeadbeefdeadbeef;
1432 val
->constant
->values
[0].u64
[j
] = u64
[comp
];
1435 /* This is for both 32-bit and 16-bit values */
1437 if (v0
->value_type
== vtn_value_type_constant
) {
1438 for (unsigned i
= 0; i
< len0
; i
++)
1439 u32
[i
] = v0
->constant
->values
[0].u32
[i
];
1441 if (v1
->value_type
== vtn_value_type_constant
) {
1442 for (unsigned i
= 0; i
< len1
; i
++)
1443 u32
[len0
+ i
] = v1
->constant
->values
[0].u32
[i
];
1446 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1447 uint32_t comp
= w
[i
+ 6];
1448 /* If component is not used, set the value to a known constant
1449 * to detect if it is wrongly used.
1451 if (comp
== (uint32_t)-1)
1452 val
->constant
->values
[0].u32
[j
] = 0xdeadbeef;
1454 val
->constant
->values
[0].u32
[j
] = u32
[comp
];
1460 case SpvOpCompositeExtract
:
1461 case SpvOpCompositeInsert
: {
1462 struct vtn_value
*comp
;
1463 unsigned deref_start
;
1464 struct nir_constant
**c
;
1465 if (opcode
== SpvOpCompositeExtract
) {
1466 comp
= vtn_value(b
, w
[4], vtn_value_type_constant
);
1468 c
= &comp
->constant
;
1470 comp
= vtn_value(b
, w
[5], vtn_value_type_constant
);
1472 val
->constant
= nir_constant_clone(comp
->constant
,
1479 const struct glsl_type
*type
= comp
->const_type
;
1480 for (unsigned i
= deref_start
; i
< count
; i
++) {
1481 switch (glsl_get_base_type(type
)) {
1482 case GLSL_TYPE_UINT
:
1484 case GLSL_TYPE_UINT16
:
1485 case GLSL_TYPE_INT16
:
1486 case GLSL_TYPE_UINT64
:
1487 case GLSL_TYPE_INT64
:
1488 case GLSL_TYPE_FLOAT
:
1489 case GLSL_TYPE_FLOAT16
:
1490 case GLSL_TYPE_DOUBLE
:
1491 case GLSL_TYPE_BOOL
:
1492 /* If we hit this granularity, we're picking off an element */
1493 if (glsl_type_is_matrix(type
)) {
1494 vtn_assert(col
== 0 && elem
== -1);
1497 type
= glsl_get_column_type(type
);
1499 vtn_assert(elem
<= 0 && glsl_type_is_vector(type
));
1501 type
= glsl_scalar_type(glsl_get_base_type(type
));
1505 case GLSL_TYPE_ARRAY
:
1506 c
= &(*c
)->elements
[w
[i
]];
1507 type
= glsl_get_array_element(type
);
1510 case GLSL_TYPE_STRUCT
:
1511 c
= &(*c
)->elements
[w
[i
]];
1512 type
= glsl_get_struct_field(type
, w
[i
]);
1516 vtn_fail("Invalid constant type");
1520 if (opcode
== SpvOpCompositeExtract
) {
1524 unsigned num_components
= glsl_get_vector_elements(type
);
1525 unsigned bit_size
= glsl_get_bit_size(type
);
1526 for (unsigned i
= 0; i
< num_components
; i
++)
1529 val
->constant
->values
[0].u64
[i
] = (*c
)->values
[col
].u64
[elem
+ i
];
1532 val
->constant
->values
[0].u32
[i
] = (*c
)->values
[col
].u32
[elem
+ i
];
1535 val
->constant
->values
[0].u16
[i
] = (*c
)->values
[col
].u16
[elem
+ i
];
1538 vtn_fail("Invalid SpvOpCompositeExtract bit size");
1542 struct vtn_value
*insert
=
1543 vtn_value(b
, w
[4], vtn_value_type_constant
);
1544 vtn_assert(insert
->const_type
== type
);
1546 *c
= insert
->constant
;
1548 unsigned num_components
= glsl_get_vector_elements(type
);
1549 unsigned bit_size
= glsl_get_bit_size(type
);
1550 for (unsigned i
= 0; i
< num_components
; i
++)
1553 (*c
)->values
[col
].u64
[elem
+ i
] = insert
->constant
->values
[0].u64
[i
];
1556 (*c
)->values
[col
].u32
[elem
+ i
] = insert
->constant
->values
[0].u32
[i
];
1559 (*c
)->values
[col
].u16
[elem
+ i
] = insert
->constant
->values
[0].u16
[i
];
1562 vtn_fail("Invalid SpvOpCompositeInsert bit size");
1571 nir_alu_type dst_alu_type
= nir_get_nir_type_for_glsl_type(val
->const_type
);
1572 nir_alu_type src_alu_type
= dst_alu_type
;
1573 unsigned num_components
= glsl_get_vector_elements(val
->const_type
);
1576 vtn_assert(count
<= 7);
1581 /* We have a source in a conversion */
1583 nir_get_nir_type_for_glsl_type(
1584 vtn_value(b
, w
[4], vtn_value_type_constant
)->const_type
);
1585 /* We use the bitsize of the conversion source to evaluate the opcode later */
1586 bit_size
= glsl_get_bit_size(
1587 vtn_value(b
, w
[4], vtn_value_type_constant
)->const_type
);
1590 bit_size
= glsl_get_bit_size(val
->const_type
);
1593 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(b
, opcode
, &swap
,
1596 nir_const_value src
[4];
1598 for (unsigned i
= 0; i
< count
- 4; i
++) {
1600 vtn_value(b
, w
[4 + i
], vtn_value_type_constant
)->constant
;
1602 unsigned j
= swap
? 1 - i
: i
;
1603 src
[j
] = c
->values
[0];
1606 val
->constant
->values
[0] =
1607 nir_eval_const_opcode(op
, num_components
, bit_size
, src
);
1614 case SpvOpConstantNull
:
1615 val
->constant
= vtn_null_constant(b
, val
->const_type
);
1618 case SpvOpConstantSampler
:
1619 vtn_fail("OpConstantSampler requires Kernel Capability");
1623 vtn_fail("Unhandled opcode");
1626 /* Now that we have the value, update the workgroup size if needed */
1627 vtn_foreach_decoration(b
, val
, handle_workgroup_size_decoration_cb
, NULL
);
1631 vtn_handle_function_call(struct vtn_builder
*b
, SpvOp opcode
,
1632 const uint32_t *w
, unsigned count
)
1634 struct vtn_type
*res_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1635 struct vtn_function
*vtn_callee
=
1636 vtn_value(b
, w
[3], vtn_value_type_function
)->func
;
1637 struct nir_function
*callee
= vtn_callee
->impl
->function
;
1639 vtn_callee
->referenced
= true;
1641 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, callee
);
1642 for (unsigned i
= 0; i
< call
->num_params
; i
++) {
1643 unsigned arg_id
= w
[4 + i
];
1644 struct vtn_value
*arg
= vtn_untyped_value(b
, arg_id
);
1645 if (arg
->value_type
== vtn_value_type_pointer
&&
1646 arg
->pointer
->ptr_type
->type
== NULL
) {
1647 nir_deref_var
*d
= vtn_pointer_to_deref(b
, arg
->pointer
);
1648 call
->params
[i
] = nir_deref_var_clone(d
, call
);
1650 struct vtn_ssa_value
*arg_ssa
= vtn_ssa_value(b
, arg_id
);
1652 /* Make a temporary to store the argument in */
1654 nir_local_variable_create(b
->nb
.impl
, arg_ssa
->type
, "arg_tmp");
1655 call
->params
[i
] = nir_deref_var_create(call
, tmp
);
1657 vtn_local_store(b
, arg_ssa
, call
->params
[i
]);
1661 nir_variable
*out_tmp
= NULL
;
1662 vtn_assert(res_type
->type
== callee
->return_type
);
1663 if (!glsl_type_is_void(callee
->return_type
)) {
1664 out_tmp
= nir_local_variable_create(b
->nb
.impl
, callee
->return_type
,
1666 call
->return_deref
= nir_deref_var_create(call
, out_tmp
);
1669 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
1671 if (glsl_type_is_void(callee
->return_type
)) {
1672 vtn_push_value(b
, w
[2], vtn_value_type_undef
);
1674 vtn_push_ssa(b
, w
[2], res_type
, vtn_local_load(b
, call
->return_deref
));
1678 struct vtn_ssa_value
*
1679 vtn_create_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
1681 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
1684 if (!glsl_type_is_vector_or_scalar(type
)) {
1685 unsigned elems
= glsl_get_length(type
);
1686 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
1687 for (unsigned i
= 0; i
< elems
; i
++) {
1688 const struct glsl_type
*child_type
;
1690 switch (glsl_get_base_type(type
)) {
1692 case GLSL_TYPE_UINT
:
1693 case GLSL_TYPE_INT16
:
1694 case GLSL_TYPE_UINT16
:
1695 case GLSL_TYPE_INT64
:
1696 case GLSL_TYPE_UINT64
:
1697 case GLSL_TYPE_BOOL
:
1698 case GLSL_TYPE_FLOAT
:
1699 case GLSL_TYPE_FLOAT16
:
1700 case GLSL_TYPE_DOUBLE
:
1701 child_type
= glsl_get_column_type(type
);
1703 case GLSL_TYPE_ARRAY
:
1704 child_type
= glsl_get_array_element(type
);
1706 case GLSL_TYPE_STRUCT
:
1707 child_type
= glsl_get_struct_field(type
, i
);
1710 vtn_fail("unkown base type");
1713 val
->elems
[i
] = vtn_create_ssa_value(b
, child_type
);
1721 vtn_tex_src(struct vtn_builder
*b
, unsigned index
, nir_tex_src_type type
)
1724 src
.src
= nir_src_for_ssa(vtn_ssa_value(b
, index
)->def
);
1725 src
.src_type
= type
;
1730 vtn_handle_texture(struct vtn_builder
*b
, SpvOp opcode
,
1731 const uint32_t *w
, unsigned count
)
1733 if (opcode
== SpvOpSampledImage
) {
1734 struct vtn_value
*val
=
1735 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
1736 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
1737 val
->sampled_image
->type
=
1738 vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1739 val
->sampled_image
->image
=
1740 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
1741 val
->sampled_image
->sampler
=
1742 vtn_value(b
, w
[4], vtn_value_type_pointer
)->pointer
;
1744 } else if (opcode
== SpvOpImage
) {
1745 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
1746 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[3]);
1747 if (src_val
->value_type
== vtn_value_type_sampled_image
) {
1748 val
->pointer
= src_val
->sampled_image
->image
;
1750 vtn_assert(src_val
->value_type
== vtn_value_type_pointer
);
1751 val
->pointer
= src_val
->pointer
;
1756 struct vtn_type
*ret_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1757 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1759 struct vtn_sampled_image sampled
;
1760 struct vtn_value
*sampled_val
= vtn_untyped_value(b
, w
[3]);
1761 if (sampled_val
->value_type
== vtn_value_type_sampled_image
) {
1762 sampled
= *sampled_val
->sampled_image
;
1764 vtn_assert(sampled_val
->value_type
== vtn_value_type_pointer
);
1765 sampled
.type
= sampled_val
->pointer
->type
;
1766 sampled
.image
= NULL
;
1767 sampled
.sampler
= sampled_val
->pointer
;
1770 const struct glsl_type
*image_type
= sampled
.type
->type
;
1771 const enum glsl_sampler_dim sampler_dim
= glsl_get_sampler_dim(image_type
);
1772 const bool is_array
= glsl_sampler_type_is_array(image_type
);
1773 const bool is_shadow
= glsl_sampler_type_is_shadow(image_type
);
1775 /* Figure out the base texture operation */
1778 case SpvOpImageSampleImplicitLod
:
1779 case SpvOpImageSampleDrefImplicitLod
:
1780 case SpvOpImageSampleProjImplicitLod
:
1781 case SpvOpImageSampleProjDrefImplicitLod
:
1782 texop
= nir_texop_tex
;
1785 case SpvOpImageSampleExplicitLod
:
1786 case SpvOpImageSampleDrefExplicitLod
:
1787 case SpvOpImageSampleProjExplicitLod
:
1788 case SpvOpImageSampleProjDrefExplicitLod
:
1789 texop
= nir_texop_txl
;
1792 case SpvOpImageFetch
:
1793 if (glsl_get_sampler_dim(image_type
) == GLSL_SAMPLER_DIM_MS
) {
1794 texop
= nir_texop_txf_ms
;
1796 texop
= nir_texop_txf
;
1800 case SpvOpImageGather
:
1801 case SpvOpImageDrefGather
:
1802 texop
= nir_texop_tg4
;
1805 case SpvOpImageQuerySizeLod
:
1806 case SpvOpImageQuerySize
:
1807 texop
= nir_texop_txs
;
1810 case SpvOpImageQueryLod
:
1811 texop
= nir_texop_lod
;
1814 case SpvOpImageQueryLevels
:
1815 texop
= nir_texop_query_levels
;
1818 case SpvOpImageQuerySamples
:
1819 texop
= nir_texop_texture_samples
;
1823 vtn_fail("Unhandled opcode");
1826 nir_tex_src srcs
[8]; /* 8 should be enough */
1827 nir_tex_src
*p
= srcs
;
1831 struct nir_ssa_def
*coord
;
1832 unsigned coord_components
;
1834 case SpvOpImageSampleImplicitLod
:
1835 case SpvOpImageSampleExplicitLod
:
1836 case SpvOpImageSampleDrefImplicitLod
:
1837 case SpvOpImageSampleDrefExplicitLod
:
1838 case SpvOpImageSampleProjImplicitLod
:
1839 case SpvOpImageSampleProjExplicitLod
:
1840 case SpvOpImageSampleProjDrefImplicitLod
:
1841 case SpvOpImageSampleProjDrefExplicitLod
:
1842 case SpvOpImageFetch
:
1843 case SpvOpImageGather
:
1844 case SpvOpImageDrefGather
:
1845 case SpvOpImageQueryLod
: {
1846 /* All these types have the coordinate as their first real argument */
1847 switch (sampler_dim
) {
1848 case GLSL_SAMPLER_DIM_1D
:
1849 case GLSL_SAMPLER_DIM_BUF
:
1850 coord_components
= 1;
1852 case GLSL_SAMPLER_DIM_2D
:
1853 case GLSL_SAMPLER_DIM_RECT
:
1854 case GLSL_SAMPLER_DIM_MS
:
1855 coord_components
= 2;
1857 case GLSL_SAMPLER_DIM_3D
:
1858 case GLSL_SAMPLER_DIM_CUBE
:
1859 coord_components
= 3;
1862 vtn_fail("Invalid sampler type");
1865 if (is_array
&& texop
!= nir_texop_lod
)
1868 coord
= vtn_ssa_value(b
, w
[idx
++])->def
;
1869 p
->src
= nir_src_for_ssa(nir_channels(&b
->nb
, coord
,
1870 (1 << coord_components
) - 1));
1871 p
->src_type
= nir_tex_src_coord
;
1878 coord_components
= 0;
1883 case SpvOpImageSampleProjImplicitLod
:
1884 case SpvOpImageSampleProjExplicitLod
:
1885 case SpvOpImageSampleProjDrefImplicitLod
:
1886 case SpvOpImageSampleProjDrefExplicitLod
:
1887 /* These have the projector as the last coordinate component */
1888 p
->src
= nir_src_for_ssa(nir_channel(&b
->nb
, coord
, coord_components
));
1889 p
->src_type
= nir_tex_src_projector
;
1897 unsigned gather_component
= 0;
1899 case SpvOpImageSampleDrefImplicitLod
:
1900 case SpvOpImageSampleDrefExplicitLod
:
1901 case SpvOpImageSampleProjDrefImplicitLod
:
1902 case SpvOpImageSampleProjDrefExplicitLod
:
1903 case SpvOpImageDrefGather
:
1904 /* These all have an explicit depth value as their next source */
1905 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_comparator
);
1908 case SpvOpImageGather
:
1909 /* This has a component as its next source */
1911 vtn_value(b
, w
[idx
++], vtn_value_type_constant
)->constant
->values
[0].u32
[0];
1918 /* For OpImageQuerySizeLod, we always have an LOD */
1919 if (opcode
== SpvOpImageQuerySizeLod
)
1920 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
1922 /* Now we need to handle some number of optional arguments */
1923 const struct vtn_ssa_value
*gather_offsets
= NULL
;
1925 uint32_t operands
= w
[idx
++];
1927 if (operands
& SpvImageOperandsBiasMask
) {
1928 vtn_assert(texop
== nir_texop_tex
);
1929 texop
= nir_texop_txb
;
1930 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_bias
);
1933 if (operands
& SpvImageOperandsLodMask
) {
1934 vtn_assert(texop
== nir_texop_txl
|| texop
== nir_texop_txf
||
1935 texop
== nir_texop_txs
);
1936 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
1939 if (operands
& SpvImageOperandsGradMask
) {
1940 vtn_assert(texop
== nir_texop_txl
);
1941 texop
= nir_texop_txd
;
1942 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddx
);
1943 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddy
);
1946 if (operands
& SpvImageOperandsOffsetMask
||
1947 operands
& SpvImageOperandsConstOffsetMask
)
1948 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_offset
);
1950 if (operands
& SpvImageOperandsConstOffsetsMask
) {
1951 gather_offsets
= vtn_ssa_value(b
, w
[idx
++]);
1952 (*p
++) = (nir_tex_src
){};
1955 if (operands
& SpvImageOperandsSampleMask
) {
1956 vtn_assert(texop
== nir_texop_txf_ms
);
1957 texop
= nir_texop_txf_ms
;
1958 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ms_index
);
1961 /* We should have now consumed exactly all of the arguments */
1962 vtn_assert(idx
== count
);
1964 nir_tex_instr
*instr
= nir_tex_instr_create(b
->shader
, p
- srcs
);
1967 memcpy(instr
->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
1969 instr
->coord_components
= coord_components
;
1970 instr
->sampler_dim
= sampler_dim
;
1971 instr
->is_array
= is_array
;
1972 instr
->is_shadow
= is_shadow
;
1973 instr
->is_new_style_shadow
=
1974 is_shadow
&& glsl_get_components(ret_type
->type
) == 1;
1975 instr
->component
= gather_component
;
1977 switch (glsl_get_sampler_result_type(image_type
)) {
1978 case GLSL_TYPE_FLOAT
: instr
->dest_type
= nir_type_float
; break;
1979 case GLSL_TYPE_INT
: instr
->dest_type
= nir_type_int
; break;
1980 case GLSL_TYPE_UINT
: instr
->dest_type
= nir_type_uint
; break;
1981 case GLSL_TYPE_BOOL
: instr
->dest_type
= nir_type_bool
; break;
1983 vtn_fail("Invalid base type for sampler result");
1986 nir_deref_var
*sampler
= vtn_pointer_to_deref(b
, sampled
.sampler
);
1987 nir_deref_var
*texture
;
1988 if (sampled
.image
) {
1989 nir_deref_var
*image
= vtn_pointer_to_deref(b
, sampled
.image
);
1995 instr
->texture
= nir_deref_var_clone(texture
, instr
);
1997 switch (instr
->op
) {
2003 /* These operations require a sampler */
2004 instr
->sampler
= nir_deref_var_clone(sampler
, instr
);
2007 case nir_texop_txf_ms
:
2010 case nir_texop_query_levels
:
2011 case nir_texop_texture_samples
:
2012 case nir_texop_samples_identical
:
2014 instr
->sampler
= NULL
;
2016 case nir_texop_txf_ms_mcs
:
2017 vtn_fail("unexpected nir_texop_txf_ms_mcs");
2020 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
2021 nir_tex_instr_dest_size(instr
), 32, NULL
);
2023 vtn_assert(glsl_get_vector_elements(ret_type
->type
) ==
2024 nir_tex_instr_dest_size(instr
));
2027 nir_instr
*instruction
;
2028 if (gather_offsets
) {
2029 vtn_assert(glsl_get_base_type(gather_offsets
->type
) == GLSL_TYPE_ARRAY
);
2030 vtn_assert(glsl_get_length(gather_offsets
->type
) == 4);
2031 nir_tex_instr
*instrs
[4] = {instr
, NULL
, NULL
, NULL
};
2033 /* Copy the current instruction 4x */
2034 for (uint32_t i
= 1; i
< 4; i
++) {
2035 instrs
[i
] = nir_tex_instr_create(b
->shader
, instr
->num_srcs
);
2036 instrs
[i
]->op
= instr
->op
;
2037 instrs
[i
]->coord_components
= instr
->coord_components
;
2038 instrs
[i
]->sampler_dim
= instr
->sampler_dim
;
2039 instrs
[i
]->is_array
= instr
->is_array
;
2040 instrs
[i
]->is_shadow
= instr
->is_shadow
;
2041 instrs
[i
]->is_new_style_shadow
= instr
->is_new_style_shadow
;
2042 instrs
[i
]->component
= instr
->component
;
2043 instrs
[i
]->dest_type
= instr
->dest_type
;
2044 instrs
[i
]->texture
= nir_deref_var_clone(texture
, instrs
[i
]);
2045 instrs
[i
]->sampler
= NULL
;
2047 memcpy(instrs
[i
]->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
2049 nir_ssa_dest_init(&instrs
[i
]->instr
, &instrs
[i
]->dest
,
2050 nir_tex_instr_dest_size(instr
), 32, NULL
);
2053 /* Fill in the last argument with the offset from the passed in offsets
2054 * and insert the instruction into the stream.
2056 for (uint32_t i
= 0; i
< 4; i
++) {
2058 src
.src
= nir_src_for_ssa(gather_offsets
->elems
[i
]->def
);
2059 src
.src_type
= nir_tex_src_offset
;
2060 instrs
[i
]->src
[instrs
[i
]->num_srcs
- 1] = src
;
2061 nir_builder_instr_insert(&b
->nb
, &instrs
[i
]->instr
);
2064 /* Combine the results of the 4 instructions by taking their .w
2067 nir_alu_instr
*vec4
= nir_alu_instr_create(b
->shader
, nir_op_vec4
);
2068 nir_ssa_dest_init(&vec4
->instr
, &vec4
->dest
.dest
, 4, 32, NULL
);
2069 vec4
->dest
.write_mask
= 0xf;
2070 for (uint32_t i
= 0; i
< 4; i
++) {
2071 vec4
->src
[i
].src
= nir_src_for_ssa(&instrs
[i
]->dest
.ssa
);
2072 vec4
->src
[i
].swizzle
[0] = 3;
2074 def
= &vec4
->dest
.dest
.ssa
;
2075 instruction
= &vec4
->instr
;
2077 def
= &instr
->dest
.ssa
;
2078 instruction
= &instr
->instr
;
2081 val
->ssa
= vtn_create_ssa_value(b
, ret_type
->type
);
2082 val
->ssa
->def
= def
;
2084 nir_builder_instr_insert(&b
->nb
, instruction
);
2088 fill_common_atomic_sources(struct vtn_builder
*b
, SpvOp opcode
,
2089 const uint32_t *w
, nir_src
*src
)
2092 case SpvOpAtomicIIncrement
:
2093 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
2096 case SpvOpAtomicIDecrement
:
2097 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
2100 case SpvOpAtomicISub
:
2102 nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
2105 case SpvOpAtomicCompareExchange
:
2106 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[8])->def
);
2107 src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
2110 case SpvOpAtomicExchange
:
2111 case SpvOpAtomicIAdd
:
2112 case SpvOpAtomicSMin
:
2113 case SpvOpAtomicUMin
:
2114 case SpvOpAtomicSMax
:
2115 case SpvOpAtomicUMax
:
2116 case SpvOpAtomicAnd
:
2118 case SpvOpAtomicXor
:
2119 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
2123 vtn_fail("Invalid SPIR-V atomic");
2127 static nir_ssa_def
*
2128 get_image_coord(struct vtn_builder
*b
, uint32_t value
)
2130 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, value
);
2132 /* The image_load_store intrinsics assume a 4-dim coordinate */
2133 unsigned dim
= glsl_get_vector_elements(coord
->type
);
2134 unsigned swizzle
[4];
2135 for (unsigned i
= 0; i
< 4; i
++)
2136 swizzle
[i
] = MIN2(i
, dim
- 1);
2138 return nir_swizzle(&b
->nb
, coord
->def
, swizzle
, 4, false);
2142 vtn_handle_image(struct vtn_builder
*b
, SpvOp opcode
,
2143 const uint32_t *w
, unsigned count
)
2145 /* Just get this one out of the way */
2146 if (opcode
== SpvOpImageTexelPointer
) {
2147 struct vtn_value
*val
=
2148 vtn_push_value(b
, w
[2], vtn_value_type_image_pointer
);
2149 val
->image
= ralloc(b
, struct vtn_image_pointer
);
2151 val
->image
->image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2152 val
->image
->coord
= get_image_coord(b
, w
[4]);
2153 val
->image
->sample
= vtn_ssa_value(b
, w
[5])->def
;
2157 struct vtn_image_pointer image
;
2160 case SpvOpAtomicExchange
:
2161 case SpvOpAtomicCompareExchange
:
2162 case SpvOpAtomicCompareExchangeWeak
:
2163 case SpvOpAtomicIIncrement
:
2164 case SpvOpAtomicIDecrement
:
2165 case SpvOpAtomicIAdd
:
2166 case SpvOpAtomicISub
:
2167 case SpvOpAtomicLoad
:
2168 case SpvOpAtomicSMin
:
2169 case SpvOpAtomicUMin
:
2170 case SpvOpAtomicSMax
:
2171 case SpvOpAtomicUMax
:
2172 case SpvOpAtomicAnd
:
2174 case SpvOpAtomicXor
:
2175 image
= *vtn_value(b
, w
[3], vtn_value_type_image_pointer
)->image
;
2178 case SpvOpAtomicStore
:
2179 image
= *vtn_value(b
, w
[1], vtn_value_type_image_pointer
)->image
;
2182 case SpvOpImageQuerySize
:
2183 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2185 image
.sample
= NULL
;
2188 case SpvOpImageRead
:
2189 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2190 image
.coord
= get_image_coord(b
, w
[4]);
2192 if (count
> 5 && (w
[5] & SpvImageOperandsSampleMask
)) {
2193 vtn_assert(w
[5] == SpvImageOperandsSampleMask
);
2194 image
.sample
= vtn_ssa_value(b
, w
[6])->def
;
2196 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2200 case SpvOpImageWrite
:
2201 image
.image
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
2202 image
.coord
= get_image_coord(b
, w
[2]);
2206 if (count
> 4 && (w
[4] & SpvImageOperandsSampleMask
)) {
2207 vtn_assert(w
[4] == SpvImageOperandsSampleMask
);
2208 image
.sample
= vtn_ssa_value(b
, w
[5])->def
;
2210 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2215 vtn_fail("Invalid image opcode");
2218 nir_intrinsic_op op
;
2220 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_##N; break;
2221 OP(ImageQuerySize
, size
)
2223 OP(ImageWrite
, store
)
2224 OP(AtomicLoad
, load
)
2225 OP(AtomicStore
, store
)
2226 OP(AtomicExchange
, atomic_exchange
)
2227 OP(AtomicCompareExchange
, atomic_comp_swap
)
2228 OP(AtomicIIncrement
, atomic_add
)
2229 OP(AtomicIDecrement
, atomic_add
)
2230 OP(AtomicIAdd
, atomic_add
)
2231 OP(AtomicISub
, atomic_add
)
2232 OP(AtomicSMin
, atomic_min
)
2233 OP(AtomicUMin
, atomic_min
)
2234 OP(AtomicSMax
, atomic_max
)
2235 OP(AtomicUMax
, atomic_max
)
2236 OP(AtomicAnd
, atomic_and
)
2237 OP(AtomicOr
, atomic_or
)
2238 OP(AtomicXor
, atomic_xor
)
2241 vtn_fail("Invalid image opcode");
2244 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
2246 nir_deref_var
*image_deref
= vtn_pointer_to_deref(b
, image
.image
);
2247 intrin
->variables
[0] = nir_deref_var_clone(image_deref
, intrin
);
2249 /* ImageQuerySize doesn't take any extra parameters */
2250 if (opcode
!= SpvOpImageQuerySize
) {
2251 /* The image coordinate is always 4 components but we may not have that
2252 * many. Swizzle to compensate.
2255 for (unsigned i
= 0; i
< 4; i
++)
2256 swiz
[i
] = i
< image
.coord
->num_components
? i
: 0;
2257 intrin
->src
[0] = nir_src_for_ssa(nir_swizzle(&b
->nb
, image
.coord
,
2259 intrin
->src
[1] = nir_src_for_ssa(image
.sample
);
2263 case SpvOpAtomicLoad
:
2264 case SpvOpImageQuerySize
:
2265 case SpvOpImageRead
:
2267 case SpvOpAtomicStore
:
2268 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2270 case SpvOpImageWrite
:
2271 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[3])->def
);
2274 case SpvOpAtomicCompareExchange
:
2275 case SpvOpAtomicIIncrement
:
2276 case SpvOpAtomicIDecrement
:
2277 case SpvOpAtomicExchange
:
2278 case SpvOpAtomicIAdd
:
2279 case SpvOpAtomicISub
:
2280 case SpvOpAtomicSMin
:
2281 case SpvOpAtomicUMin
:
2282 case SpvOpAtomicSMax
:
2283 case SpvOpAtomicUMax
:
2284 case SpvOpAtomicAnd
:
2286 case SpvOpAtomicXor
:
2287 fill_common_atomic_sources(b
, opcode
, w
, &intrin
->src
[2]);
2291 vtn_fail("Invalid image opcode");
2294 if (opcode
!= SpvOpImageWrite
) {
2295 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2296 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2298 unsigned dest_components
=
2299 nir_intrinsic_infos
[intrin
->intrinsic
].dest_components
;
2300 if (intrin
->intrinsic
== nir_intrinsic_image_size
) {
2301 dest_components
= intrin
->num_components
=
2302 glsl_get_vector_elements(type
->type
);
2305 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
,
2306 dest_components
, 32, NULL
);
2308 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2310 val
->ssa
= vtn_create_ssa_value(b
, type
->type
);
2311 val
->ssa
->def
= &intrin
->dest
.ssa
;
2313 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2317 static nir_intrinsic_op
2318 get_ssbo_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2321 case SpvOpAtomicLoad
: return nir_intrinsic_load_ssbo
;
2322 case SpvOpAtomicStore
: return nir_intrinsic_store_ssbo
;
2323 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
2324 OP(AtomicExchange
, atomic_exchange
)
2325 OP(AtomicCompareExchange
, atomic_comp_swap
)
2326 OP(AtomicIIncrement
, atomic_add
)
2327 OP(AtomicIDecrement
, atomic_add
)
2328 OP(AtomicIAdd
, atomic_add
)
2329 OP(AtomicISub
, atomic_add
)
2330 OP(AtomicSMin
, atomic_imin
)
2331 OP(AtomicUMin
, atomic_umin
)
2332 OP(AtomicSMax
, atomic_imax
)
2333 OP(AtomicUMax
, atomic_umax
)
2334 OP(AtomicAnd
, atomic_and
)
2335 OP(AtomicOr
, atomic_or
)
2336 OP(AtomicXor
, atomic_xor
)
2339 vtn_fail("Invalid SSBO atomic");
2343 static nir_intrinsic_op
2344 get_shared_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2347 case SpvOpAtomicLoad
: return nir_intrinsic_load_shared
;
2348 case SpvOpAtomicStore
: return nir_intrinsic_store_shared
;
2349 #define OP(S, N) case SpvOp##S: return nir_intrinsic_shared_##N;
2350 OP(AtomicExchange
, atomic_exchange
)
2351 OP(AtomicCompareExchange
, atomic_comp_swap
)
2352 OP(AtomicIIncrement
, atomic_add
)
2353 OP(AtomicIDecrement
, atomic_add
)
2354 OP(AtomicIAdd
, atomic_add
)
2355 OP(AtomicISub
, atomic_add
)
2356 OP(AtomicSMin
, atomic_imin
)
2357 OP(AtomicUMin
, atomic_umin
)
2358 OP(AtomicSMax
, atomic_imax
)
2359 OP(AtomicUMax
, atomic_umax
)
2360 OP(AtomicAnd
, atomic_and
)
2361 OP(AtomicOr
, atomic_or
)
2362 OP(AtomicXor
, atomic_xor
)
2365 vtn_fail("Invalid shared atomic");
2369 static nir_intrinsic_op
2370 get_var_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2373 case SpvOpAtomicLoad
: return nir_intrinsic_load_var
;
2374 case SpvOpAtomicStore
: return nir_intrinsic_store_var
;
2375 #define OP(S, N) case SpvOp##S: return nir_intrinsic_var_##N;
2376 OP(AtomicExchange
, atomic_exchange
)
2377 OP(AtomicCompareExchange
, atomic_comp_swap
)
2378 OP(AtomicIIncrement
, atomic_add
)
2379 OP(AtomicIDecrement
, atomic_add
)
2380 OP(AtomicIAdd
, atomic_add
)
2381 OP(AtomicISub
, atomic_add
)
2382 OP(AtomicSMin
, atomic_imin
)
2383 OP(AtomicUMin
, atomic_umin
)
2384 OP(AtomicSMax
, atomic_imax
)
2385 OP(AtomicUMax
, atomic_umax
)
2386 OP(AtomicAnd
, atomic_and
)
2387 OP(AtomicOr
, atomic_or
)
2388 OP(AtomicXor
, atomic_xor
)
2391 vtn_fail("Invalid shared atomic");
2396 vtn_handle_ssbo_or_shared_atomic(struct vtn_builder
*b
, SpvOp opcode
,
2397 const uint32_t *w
, unsigned count
)
2399 struct vtn_pointer
*ptr
;
2400 nir_intrinsic_instr
*atomic
;
2403 case SpvOpAtomicLoad
:
2404 case SpvOpAtomicExchange
:
2405 case SpvOpAtomicCompareExchange
:
2406 case SpvOpAtomicCompareExchangeWeak
:
2407 case SpvOpAtomicIIncrement
:
2408 case SpvOpAtomicIDecrement
:
2409 case SpvOpAtomicIAdd
:
2410 case SpvOpAtomicISub
:
2411 case SpvOpAtomicSMin
:
2412 case SpvOpAtomicUMin
:
2413 case SpvOpAtomicSMax
:
2414 case SpvOpAtomicUMax
:
2415 case SpvOpAtomicAnd
:
2417 case SpvOpAtomicXor
:
2418 ptr
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2421 case SpvOpAtomicStore
:
2422 ptr
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
2426 vtn_fail("Invalid SPIR-V atomic");
2430 SpvScope scope = w[4];
2431 SpvMemorySemanticsMask semantics = w[5];
2434 if (ptr
->mode
== vtn_variable_mode_workgroup
&&
2435 !b
->options
->lower_workgroup_access_to_offsets
) {
2436 nir_deref_var
*deref
= vtn_pointer_to_deref(b
, ptr
);
2437 const struct glsl_type
*deref_type
= nir_deref_tail(&deref
->deref
)->type
;
2438 nir_intrinsic_op op
= get_var_nir_atomic_op(b
, opcode
);
2439 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2440 atomic
->variables
[0] = nir_deref_var_clone(deref
, atomic
);
2443 case SpvOpAtomicLoad
:
2444 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2447 case SpvOpAtomicStore
:
2448 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2449 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2450 atomic
->src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2453 case SpvOpAtomicExchange
:
2454 case SpvOpAtomicCompareExchange
:
2455 case SpvOpAtomicCompareExchangeWeak
:
2456 case SpvOpAtomicIIncrement
:
2457 case SpvOpAtomicIDecrement
:
2458 case SpvOpAtomicIAdd
:
2459 case SpvOpAtomicISub
:
2460 case SpvOpAtomicSMin
:
2461 case SpvOpAtomicUMin
:
2462 case SpvOpAtomicSMax
:
2463 case SpvOpAtomicUMax
:
2464 case SpvOpAtomicAnd
:
2466 case SpvOpAtomicXor
:
2467 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[0]);
2471 vtn_fail("Invalid SPIR-V atomic");
2475 nir_ssa_def
*offset
, *index
;
2476 offset
= vtn_pointer_to_offset(b
, ptr
, &index
, NULL
);
2478 nir_intrinsic_op op
;
2479 if (ptr
->mode
== vtn_variable_mode_ssbo
) {
2480 op
= get_ssbo_nir_atomic_op(b
, opcode
);
2482 vtn_assert(ptr
->mode
== vtn_variable_mode_workgroup
&&
2483 b
->options
->lower_workgroup_access_to_offsets
);
2484 op
= get_shared_nir_atomic_op(b
, opcode
);
2487 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2491 case SpvOpAtomicLoad
:
2492 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
2493 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2494 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2495 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2498 case SpvOpAtomicStore
:
2499 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
2500 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2501 atomic
->src
[src
++] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2502 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2503 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2504 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2507 case SpvOpAtomicExchange
:
2508 case SpvOpAtomicCompareExchange
:
2509 case SpvOpAtomicCompareExchangeWeak
:
2510 case SpvOpAtomicIIncrement
:
2511 case SpvOpAtomicIDecrement
:
2512 case SpvOpAtomicIAdd
:
2513 case SpvOpAtomicISub
:
2514 case SpvOpAtomicSMin
:
2515 case SpvOpAtomicUMin
:
2516 case SpvOpAtomicSMax
:
2517 case SpvOpAtomicUMax
:
2518 case SpvOpAtomicAnd
:
2520 case SpvOpAtomicXor
:
2521 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2522 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2523 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2524 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[src
]);
2528 vtn_fail("Invalid SPIR-V atomic");
2532 if (opcode
!= SpvOpAtomicStore
) {
2533 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2535 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
,
2536 glsl_get_vector_elements(type
->type
),
2537 glsl_get_bit_size(type
->type
), NULL
);
2539 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2540 val
->ssa
= rzalloc(b
, struct vtn_ssa_value
);
2541 val
->ssa
->def
= &atomic
->dest
.ssa
;
2542 val
->ssa
->type
= type
->type
;
2545 nir_builder_instr_insert(&b
->nb
, &atomic
->instr
);
2548 static nir_alu_instr
*
2549 create_vec(struct vtn_builder
*b
, unsigned num_components
, unsigned bit_size
)
2552 switch (num_components
) {
2553 case 1: op
= nir_op_fmov
; break;
2554 case 2: op
= nir_op_vec2
; break;
2555 case 3: op
= nir_op_vec3
; break;
2556 case 4: op
= nir_op_vec4
; break;
2557 default: vtn_fail("bad vector size");
2560 nir_alu_instr
*vec
= nir_alu_instr_create(b
->shader
, op
);
2561 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, num_components
,
2563 vec
->dest
.write_mask
= (1 << num_components
) - 1;
2568 struct vtn_ssa_value
*
2569 vtn_ssa_transpose(struct vtn_builder
*b
, struct vtn_ssa_value
*src
)
2571 if (src
->transposed
)
2572 return src
->transposed
;
2574 struct vtn_ssa_value
*dest
=
2575 vtn_create_ssa_value(b
, glsl_transposed_type(src
->type
));
2577 for (unsigned i
= 0; i
< glsl_get_matrix_columns(dest
->type
); i
++) {
2578 nir_alu_instr
*vec
= create_vec(b
, glsl_get_matrix_columns(src
->type
),
2579 glsl_get_bit_size(src
->type
));
2580 if (glsl_type_is_vector_or_scalar(src
->type
)) {
2581 vec
->src
[0].src
= nir_src_for_ssa(src
->def
);
2582 vec
->src
[0].swizzle
[0] = i
;
2584 for (unsigned j
= 0; j
< glsl_get_matrix_columns(src
->type
); j
++) {
2585 vec
->src
[j
].src
= nir_src_for_ssa(src
->elems
[j
]->def
);
2586 vec
->src
[j
].swizzle
[0] = i
;
2589 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2590 dest
->elems
[i
]->def
= &vec
->dest
.dest
.ssa
;
2593 dest
->transposed
= src
;
2599 vtn_vector_extract(struct vtn_builder
*b
, nir_ssa_def
*src
, unsigned index
)
2601 unsigned swiz
[4] = { index
};
2602 return nir_swizzle(&b
->nb
, src
, swiz
, 1, true);
2606 vtn_vector_insert(struct vtn_builder
*b
, nir_ssa_def
*src
, nir_ssa_def
*insert
,
2609 nir_alu_instr
*vec
= create_vec(b
, src
->num_components
,
2612 for (unsigned i
= 0; i
< src
->num_components
; i
++) {
2614 vec
->src
[i
].src
= nir_src_for_ssa(insert
);
2616 vec
->src
[i
].src
= nir_src_for_ssa(src
);
2617 vec
->src
[i
].swizzle
[0] = i
;
2621 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2623 return &vec
->dest
.dest
.ssa
;
2627 vtn_vector_extract_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
2630 nir_ssa_def
*dest
= vtn_vector_extract(b
, src
, 0);
2631 for (unsigned i
= 1; i
< src
->num_components
; i
++)
2632 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
2633 vtn_vector_extract(b
, src
, i
), dest
);
2639 vtn_vector_insert_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
2640 nir_ssa_def
*insert
, nir_ssa_def
*index
)
2642 nir_ssa_def
*dest
= vtn_vector_insert(b
, src
, insert
, 0);
2643 for (unsigned i
= 1; i
< src
->num_components
; i
++)
2644 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
2645 vtn_vector_insert(b
, src
, insert
, i
), dest
);
2650 static nir_ssa_def
*
2651 vtn_vector_shuffle(struct vtn_builder
*b
, unsigned num_components
,
2652 nir_ssa_def
*src0
, nir_ssa_def
*src1
,
2653 const uint32_t *indices
)
2655 nir_alu_instr
*vec
= create_vec(b
, num_components
, src0
->bit_size
);
2657 for (unsigned i
= 0; i
< num_components
; i
++) {
2658 uint32_t index
= indices
[i
];
2659 if (index
== 0xffffffff) {
2661 nir_src_for_ssa(nir_ssa_undef(&b
->nb
, 1, src0
->bit_size
));
2662 } else if (index
< src0
->num_components
) {
2663 vec
->src
[i
].src
= nir_src_for_ssa(src0
);
2664 vec
->src
[i
].swizzle
[0] = index
;
2666 vec
->src
[i
].src
= nir_src_for_ssa(src1
);
2667 vec
->src
[i
].swizzle
[0] = index
- src0
->num_components
;
2671 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2673 return &vec
->dest
.dest
.ssa
;
2677 * Concatentates a number of vectors/scalars together to produce a vector
2679 static nir_ssa_def
*
2680 vtn_vector_construct(struct vtn_builder
*b
, unsigned num_components
,
2681 unsigned num_srcs
, nir_ssa_def
**srcs
)
2683 nir_alu_instr
*vec
= create_vec(b
, num_components
, srcs
[0]->bit_size
);
2685 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2687 * "When constructing a vector, there must be at least two Constituent
2690 vtn_assert(num_srcs
>= 2);
2692 unsigned dest_idx
= 0;
2693 for (unsigned i
= 0; i
< num_srcs
; i
++) {
2694 nir_ssa_def
*src
= srcs
[i
];
2695 vtn_assert(dest_idx
+ src
->num_components
<= num_components
);
2696 for (unsigned j
= 0; j
< src
->num_components
; j
++) {
2697 vec
->src
[dest_idx
].src
= nir_src_for_ssa(src
);
2698 vec
->src
[dest_idx
].swizzle
[0] = j
;
2703 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2705 * "When constructing a vector, the total number of components in all
2706 * the operands must equal the number of components in Result Type."
2708 vtn_assert(dest_idx
== num_components
);
2710 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2712 return &vec
->dest
.dest
.ssa
;
2715 static struct vtn_ssa_value
*
2716 vtn_composite_copy(void *mem_ctx
, struct vtn_ssa_value
*src
)
2718 struct vtn_ssa_value
*dest
= rzalloc(mem_ctx
, struct vtn_ssa_value
);
2719 dest
->type
= src
->type
;
2721 if (glsl_type_is_vector_or_scalar(src
->type
)) {
2722 dest
->def
= src
->def
;
2724 unsigned elems
= glsl_get_length(src
->type
);
2726 dest
->elems
= ralloc_array(mem_ctx
, struct vtn_ssa_value
*, elems
);
2727 for (unsigned i
= 0; i
< elems
; i
++)
2728 dest
->elems
[i
] = vtn_composite_copy(mem_ctx
, src
->elems
[i
]);
2734 static struct vtn_ssa_value
*
2735 vtn_composite_insert(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
2736 struct vtn_ssa_value
*insert
, const uint32_t *indices
,
2737 unsigned num_indices
)
2739 struct vtn_ssa_value
*dest
= vtn_composite_copy(b
, src
);
2741 struct vtn_ssa_value
*cur
= dest
;
2743 for (i
= 0; i
< num_indices
- 1; i
++) {
2744 cur
= cur
->elems
[indices
[i
]];
2747 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
2748 /* According to the SPIR-V spec, OpCompositeInsert may work down to
2749 * the component granularity. In that case, the last index will be
2750 * the index to insert the scalar into the vector.
2753 cur
->def
= vtn_vector_insert(b
, cur
->def
, insert
->def
, indices
[i
]);
2755 cur
->elems
[indices
[i
]] = insert
;
2761 static struct vtn_ssa_value
*
2762 vtn_composite_extract(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
2763 const uint32_t *indices
, unsigned num_indices
)
2765 struct vtn_ssa_value
*cur
= src
;
2766 for (unsigned i
= 0; i
< num_indices
; i
++) {
2767 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
2768 vtn_assert(i
== num_indices
- 1);
2769 /* According to the SPIR-V spec, OpCompositeExtract may work down to
2770 * the component granularity. The last index will be the index of the
2771 * vector to extract.
2774 struct vtn_ssa_value
*ret
= rzalloc(b
, struct vtn_ssa_value
);
2775 ret
->type
= glsl_scalar_type(glsl_get_base_type(cur
->type
));
2776 ret
->def
= vtn_vector_extract(b
, cur
->def
, indices
[i
]);
2779 cur
= cur
->elems
[indices
[i
]];
2787 vtn_handle_composite(struct vtn_builder
*b
, SpvOp opcode
,
2788 const uint32_t *w
, unsigned count
)
2790 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2791 const struct glsl_type
*type
=
2792 vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
2793 val
->ssa
= vtn_create_ssa_value(b
, type
);
2796 case SpvOpVectorExtractDynamic
:
2797 val
->ssa
->def
= vtn_vector_extract_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
2798 vtn_ssa_value(b
, w
[4])->def
);
2801 case SpvOpVectorInsertDynamic
:
2802 val
->ssa
->def
= vtn_vector_insert_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
2803 vtn_ssa_value(b
, w
[4])->def
,
2804 vtn_ssa_value(b
, w
[5])->def
);
2807 case SpvOpVectorShuffle
:
2808 val
->ssa
->def
= vtn_vector_shuffle(b
, glsl_get_vector_elements(type
),
2809 vtn_ssa_value(b
, w
[3])->def
,
2810 vtn_ssa_value(b
, w
[4])->def
,
2814 case SpvOpCompositeConstruct
: {
2815 unsigned elems
= count
- 3;
2816 if (glsl_type_is_vector_or_scalar(type
)) {
2817 nir_ssa_def
*srcs
[4];
2818 for (unsigned i
= 0; i
< elems
; i
++)
2819 srcs
[i
] = vtn_ssa_value(b
, w
[3 + i
])->def
;
2821 vtn_vector_construct(b
, glsl_get_vector_elements(type
),
2824 val
->ssa
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
2825 for (unsigned i
= 0; i
< elems
; i
++)
2826 val
->ssa
->elems
[i
] = vtn_ssa_value(b
, w
[3 + i
]);
2830 case SpvOpCompositeExtract
:
2831 val
->ssa
= vtn_composite_extract(b
, vtn_ssa_value(b
, w
[3]),
2835 case SpvOpCompositeInsert
:
2836 val
->ssa
= vtn_composite_insert(b
, vtn_ssa_value(b
, w
[4]),
2837 vtn_ssa_value(b
, w
[3]),
2841 case SpvOpCopyObject
:
2842 val
->ssa
= vtn_composite_copy(b
, vtn_ssa_value(b
, w
[3]));
2846 vtn_fail("unknown composite operation");
2851 vtn_handle_barrier(struct vtn_builder
*b
, SpvOp opcode
,
2852 const uint32_t *w
, unsigned count
)
2854 nir_intrinsic_op intrinsic_op
;
2856 case SpvOpEmitVertex
:
2857 case SpvOpEmitStreamVertex
:
2858 intrinsic_op
= nir_intrinsic_emit_vertex
;
2860 case SpvOpEndPrimitive
:
2861 case SpvOpEndStreamPrimitive
:
2862 intrinsic_op
= nir_intrinsic_end_primitive
;
2864 case SpvOpMemoryBarrier
:
2865 intrinsic_op
= nir_intrinsic_memory_barrier
;
2867 case SpvOpControlBarrier
:
2868 intrinsic_op
= nir_intrinsic_barrier
;
2871 vtn_fail("unknown barrier instruction");
2874 nir_intrinsic_instr
*intrin
=
2875 nir_intrinsic_instr_create(b
->shader
, intrinsic_op
);
2877 if (opcode
== SpvOpEmitStreamVertex
|| opcode
== SpvOpEndStreamPrimitive
)
2878 nir_intrinsic_set_stream_id(intrin
, w
[1]);
2880 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2884 gl_primitive_from_spv_execution_mode(struct vtn_builder
*b
,
2885 SpvExecutionMode mode
)
2888 case SpvExecutionModeInputPoints
:
2889 case SpvExecutionModeOutputPoints
:
2890 return 0; /* GL_POINTS */
2891 case SpvExecutionModeInputLines
:
2892 return 1; /* GL_LINES */
2893 case SpvExecutionModeInputLinesAdjacency
:
2894 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
2895 case SpvExecutionModeTriangles
:
2896 return 4; /* GL_TRIANGLES */
2897 case SpvExecutionModeInputTrianglesAdjacency
:
2898 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
2899 case SpvExecutionModeQuads
:
2900 return 7; /* GL_QUADS */
2901 case SpvExecutionModeIsolines
:
2902 return 0x8E7A; /* GL_ISOLINES */
2903 case SpvExecutionModeOutputLineStrip
:
2904 return 3; /* GL_LINE_STRIP */
2905 case SpvExecutionModeOutputTriangleStrip
:
2906 return 5; /* GL_TRIANGLE_STRIP */
2908 vtn_fail("Invalid primitive type");
2913 vertices_in_from_spv_execution_mode(struct vtn_builder
*b
,
2914 SpvExecutionMode mode
)
2917 case SpvExecutionModeInputPoints
:
2919 case SpvExecutionModeInputLines
:
2921 case SpvExecutionModeInputLinesAdjacency
:
2923 case SpvExecutionModeTriangles
:
2925 case SpvExecutionModeInputTrianglesAdjacency
:
2928 vtn_fail("Invalid GS input mode");
2932 static gl_shader_stage
2933 stage_for_execution_model(struct vtn_builder
*b
, SpvExecutionModel model
)
2936 case SpvExecutionModelVertex
:
2937 return MESA_SHADER_VERTEX
;
2938 case SpvExecutionModelTessellationControl
:
2939 return MESA_SHADER_TESS_CTRL
;
2940 case SpvExecutionModelTessellationEvaluation
:
2941 return MESA_SHADER_TESS_EVAL
;
2942 case SpvExecutionModelGeometry
:
2943 return MESA_SHADER_GEOMETRY
;
2944 case SpvExecutionModelFragment
:
2945 return MESA_SHADER_FRAGMENT
;
2946 case SpvExecutionModelGLCompute
:
2947 return MESA_SHADER_COMPUTE
;
2949 vtn_fail("Unsupported execution model");
2953 #define spv_check_supported(name, cap) do { \
2954 if (!(b->options && b->options->caps.name)) \
2955 vtn_warn("Unsupported SPIR-V capability: %s", \
2956 spirv_capability_to_string(cap)); \
2960 vtn_handle_preamble_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2961 const uint32_t *w
, unsigned count
)
2968 case SpvSourceLanguageUnknown
: lang
= "unknown"; break;
2969 case SpvSourceLanguageESSL
: lang
= "ESSL"; break;
2970 case SpvSourceLanguageGLSL
: lang
= "GLSL"; break;
2971 case SpvSourceLanguageOpenCL_C
: lang
= "OpenCL C"; break;
2972 case SpvSourceLanguageOpenCL_CPP
: lang
= "OpenCL C++"; break;
2973 case SpvSourceLanguageHLSL
: lang
= "HLSL"; break;
2976 uint32_t version
= w
[2];
2979 (count
> 3) ? vtn_value(b
, w
[3], vtn_value_type_string
)->str
: "";
2981 vtn_info("Parsing SPIR-V from %s %u source file %s", lang
, version
, file
);
2985 case SpvOpSourceExtension
:
2986 case SpvOpSourceContinued
:
2987 case SpvOpExtension
:
2988 /* Unhandled, but these are for debug so that's ok. */
2991 case SpvOpCapability
: {
2992 SpvCapability cap
= w
[1];
2994 case SpvCapabilityMatrix
:
2995 case SpvCapabilityShader
:
2996 case SpvCapabilityGeometry
:
2997 case SpvCapabilityGeometryPointSize
:
2998 case SpvCapabilityUniformBufferArrayDynamicIndexing
:
2999 case SpvCapabilitySampledImageArrayDynamicIndexing
:
3000 case SpvCapabilityStorageBufferArrayDynamicIndexing
:
3001 case SpvCapabilityStorageImageArrayDynamicIndexing
:
3002 case SpvCapabilityImageRect
:
3003 case SpvCapabilitySampledRect
:
3004 case SpvCapabilitySampled1D
:
3005 case SpvCapabilityImage1D
:
3006 case SpvCapabilitySampledCubeArray
:
3007 case SpvCapabilityImageCubeArray
:
3008 case SpvCapabilitySampledBuffer
:
3009 case SpvCapabilityImageBuffer
:
3010 case SpvCapabilityImageQuery
:
3011 case SpvCapabilityDerivativeControl
:
3012 case SpvCapabilityInterpolationFunction
:
3013 case SpvCapabilityMultiViewport
:
3014 case SpvCapabilitySampleRateShading
:
3015 case SpvCapabilityClipDistance
:
3016 case SpvCapabilityCullDistance
:
3017 case SpvCapabilityInputAttachment
:
3018 case SpvCapabilityImageGatherExtended
:
3019 case SpvCapabilityStorageImageExtendedFormats
:
3022 case SpvCapabilityGeometryStreams
:
3023 case SpvCapabilityLinkage
:
3024 case SpvCapabilityVector16
:
3025 case SpvCapabilityFloat16Buffer
:
3026 case SpvCapabilityFloat16
:
3027 case SpvCapabilityInt64Atomics
:
3028 case SpvCapabilityAtomicStorage
:
3029 case SpvCapabilityInt16
:
3030 case SpvCapabilityStorageImageMultisample
:
3031 case SpvCapabilityInt8
:
3032 case SpvCapabilitySparseResidency
:
3033 case SpvCapabilityMinLod
:
3034 case SpvCapabilityTransformFeedback
:
3035 vtn_warn("Unsupported SPIR-V capability: %s",
3036 spirv_capability_to_string(cap
));
3039 case SpvCapabilityFloat64
:
3040 spv_check_supported(float64
, cap
);
3042 case SpvCapabilityInt64
:
3043 spv_check_supported(int64
, cap
);
3046 case SpvCapabilityAddresses
:
3047 case SpvCapabilityKernel
:
3048 case SpvCapabilityImageBasic
:
3049 case SpvCapabilityImageReadWrite
:
3050 case SpvCapabilityImageMipmap
:
3051 case SpvCapabilityPipes
:
3052 case SpvCapabilityGroups
:
3053 case SpvCapabilityDeviceEnqueue
:
3054 case SpvCapabilityLiteralSampler
:
3055 case SpvCapabilityGenericPointer
:
3056 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
3057 spirv_capability_to_string(cap
));
3060 case SpvCapabilityImageMSArray
:
3061 spv_check_supported(image_ms_array
, cap
);
3064 case SpvCapabilityTessellation
:
3065 case SpvCapabilityTessellationPointSize
:
3066 spv_check_supported(tessellation
, cap
);
3069 case SpvCapabilityDrawParameters
:
3070 spv_check_supported(draw_parameters
, cap
);
3073 case SpvCapabilityStorageImageReadWithoutFormat
:
3074 spv_check_supported(image_read_without_format
, cap
);
3077 case SpvCapabilityStorageImageWriteWithoutFormat
:
3078 spv_check_supported(image_write_without_format
, cap
);
3081 case SpvCapabilityMultiView
:
3082 spv_check_supported(multiview
, cap
);
3085 case SpvCapabilityVariablePointersStorageBuffer
:
3086 case SpvCapabilityVariablePointers
:
3087 spv_check_supported(variable_pointers
, cap
);
3090 case SpvCapabilityStorageUniformBufferBlock16
:
3091 case SpvCapabilityStorageUniform16
:
3092 case SpvCapabilityStoragePushConstant16
:
3093 case SpvCapabilityStorageInputOutput16
:
3094 spv_check_supported(storage_16bit
, cap
);
3098 vtn_fail("Unhandled capability");
3103 case SpvOpExtInstImport
:
3104 vtn_handle_extension(b
, opcode
, w
, count
);
3107 case SpvOpMemoryModel
:
3108 vtn_assert(w
[1] == SpvAddressingModelLogical
);
3109 vtn_assert(w
[2] == SpvMemoryModelSimple
||
3110 w
[2] == SpvMemoryModelGLSL450
);
3113 case SpvOpEntryPoint
: {
3114 struct vtn_value
*entry_point
= &b
->values
[w
[2]];
3115 /* Let this be a name label regardless */
3116 unsigned name_words
;
3117 entry_point
->name
= vtn_string_literal(b
, &w
[3], count
- 3, &name_words
);
3119 if (strcmp(entry_point
->name
, b
->entry_point_name
) != 0 ||
3120 stage_for_execution_model(b
, w
[1]) != b
->entry_point_stage
)
3123 vtn_assert(b
->entry_point
== NULL
);
3124 b
->entry_point
= entry_point
;
3129 vtn_push_value(b
, w
[1], vtn_value_type_string
)->str
=
3130 vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
3134 b
->values
[w
[1]].name
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
3137 case SpvOpMemberName
:
3141 case SpvOpExecutionMode
:
3142 case SpvOpDecorationGroup
:
3144 case SpvOpMemberDecorate
:
3145 case SpvOpGroupDecorate
:
3146 case SpvOpGroupMemberDecorate
:
3147 vtn_handle_decoration(b
, opcode
, w
, count
);
3151 return false; /* End of preamble */
3158 vtn_handle_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
3159 const struct vtn_decoration
*mode
, void *data
)
3161 vtn_assert(b
->entry_point
== entry_point
);
3163 switch(mode
->exec_mode
) {
3164 case SpvExecutionModeOriginUpperLeft
:
3165 case SpvExecutionModeOriginLowerLeft
:
3166 b
->origin_upper_left
=
3167 (mode
->exec_mode
== SpvExecutionModeOriginUpperLeft
);
3170 case SpvExecutionModeEarlyFragmentTests
:
3171 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3172 b
->shader
->info
.fs
.early_fragment_tests
= true;
3175 case SpvExecutionModeInvocations
:
3176 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3177 b
->shader
->info
.gs
.invocations
= MAX2(1, mode
->literals
[0]);
3180 case SpvExecutionModeDepthReplacing
:
3181 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3182 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_ANY
;
3184 case SpvExecutionModeDepthGreater
:
3185 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3186 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_GREATER
;
3188 case SpvExecutionModeDepthLess
:
3189 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3190 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_LESS
;
3192 case SpvExecutionModeDepthUnchanged
:
3193 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3194 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
3197 case SpvExecutionModeLocalSize
:
3198 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
3199 b
->shader
->info
.cs
.local_size
[0] = mode
->literals
[0];
3200 b
->shader
->info
.cs
.local_size
[1] = mode
->literals
[1];
3201 b
->shader
->info
.cs
.local_size
[2] = mode
->literals
[2];
3203 case SpvExecutionModeLocalSizeHint
:
3204 break; /* Nothing to do with this */
3206 case SpvExecutionModeOutputVertices
:
3207 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3208 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
3209 b
->shader
->info
.tess
.tcs_vertices_out
= mode
->literals
[0];
3211 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3212 b
->shader
->info
.gs
.vertices_out
= mode
->literals
[0];
3216 case SpvExecutionModeInputPoints
:
3217 case SpvExecutionModeInputLines
:
3218 case SpvExecutionModeInputLinesAdjacency
:
3219 case SpvExecutionModeTriangles
:
3220 case SpvExecutionModeInputTrianglesAdjacency
:
3221 case SpvExecutionModeQuads
:
3222 case SpvExecutionModeIsolines
:
3223 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3224 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
3225 b
->shader
->info
.tess
.primitive_mode
=
3226 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
3228 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3229 b
->shader
->info
.gs
.vertices_in
=
3230 vertices_in_from_spv_execution_mode(b
, mode
->exec_mode
);
3234 case SpvExecutionModeOutputPoints
:
3235 case SpvExecutionModeOutputLineStrip
:
3236 case SpvExecutionModeOutputTriangleStrip
:
3237 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3238 b
->shader
->info
.gs
.output_primitive
=
3239 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
3242 case SpvExecutionModeSpacingEqual
:
3243 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3244 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3245 b
->shader
->info
.tess
.spacing
= TESS_SPACING_EQUAL
;
3247 case SpvExecutionModeSpacingFractionalEven
:
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_FRACTIONAL_EVEN
;
3252 case SpvExecutionModeSpacingFractionalOdd
:
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_ODD
;
3257 case SpvExecutionModeVertexOrderCw
:
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
.ccw
= false;
3262 case SpvExecutionModeVertexOrderCcw
:
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
= true;
3267 case SpvExecutionModePointMode
:
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
.point_mode
= true;
3273 case SpvExecutionModePixelCenterInteger
:
3274 b
->pixel_center_integer
= true;
3277 case SpvExecutionModeXfb
:
3278 vtn_fail("Unhandled execution mode");
3281 case SpvExecutionModeVecTypeHint
:
3282 case SpvExecutionModeContractionOff
:
3286 vtn_fail("Unhandled execution mode");
3291 vtn_handle_variable_or_type_instruction(struct vtn_builder
*b
, SpvOp opcode
,
3292 const uint32_t *w
, unsigned count
)
3296 case SpvOpSourceContinued
:
3297 case SpvOpSourceExtension
:
3298 case SpvOpExtension
:
3299 case SpvOpCapability
:
3300 case SpvOpExtInstImport
:
3301 case SpvOpMemoryModel
:
3302 case SpvOpEntryPoint
:
3303 case SpvOpExecutionMode
:
3306 case SpvOpMemberName
:
3307 case SpvOpDecorationGroup
:
3309 case SpvOpMemberDecorate
:
3310 case SpvOpGroupDecorate
:
3311 case SpvOpGroupMemberDecorate
:
3312 vtn_fail("Invalid opcode types and variables section");
3318 case SpvOpTypeFloat
:
3319 case SpvOpTypeVector
:
3320 case SpvOpTypeMatrix
:
3321 case SpvOpTypeImage
:
3322 case SpvOpTypeSampler
:
3323 case SpvOpTypeSampledImage
:
3324 case SpvOpTypeArray
:
3325 case SpvOpTypeRuntimeArray
:
3326 case SpvOpTypeStruct
:
3327 case SpvOpTypeOpaque
:
3328 case SpvOpTypePointer
:
3329 case SpvOpTypeFunction
:
3330 case SpvOpTypeEvent
:
3331 case SpvOpTypeDeviceEvent
:
3332 case SpvOpTypeReserveId
:
3333 case SpvOpTypeQueue
:
3335 vtn_handle_type(b
, opcode
, w
, count
);
3338 case SpvOpConstantTrue
:
3339 case SpvOpConstantFalse
:
3341 case SpvOpConstantComposite
:
3342 case SpvOpConstantSampler
:
3343 case SpvOpConstantNull
:
3344 case SpvOpSpecConstantTrue
:
3345 case SpvOpSpecConstantFalse
:
3346 case SpvOpSpecConstant
:
3347 case SpvOpSpecConstantComposite
:
3348 case SpvOpSpecConstantOp
:
3349 vtn_handle_constant(b
, opcode
, w
, count
);
3354 vtn_handle_variables(b
, opcode
, w
, count
);
3358 return false; /* End of preamble */
3365 vtn_handle_body_instruction(struct vtn_builder
*b
, SpvOp opcode
,
3366 const uint32_t *w
, unsigned count
)
3372 case SpvOpLoopMerge
:
3373 case SpvOpSelectionMerge
:
3374 /* This is handled by cfg pre-pass and walk_blocks */
3378 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
3379 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
3384 vtn_handle_extension(b
, opcode
, w
, count
);
3390 case SpvOpCopyMemory
:
3391 case SpvOpCopyMemorySized
:
3392 case SpvOpAccessChain
:
3393 case SpvOpPtrAccessChain
:
3394 case SpvOpInBoundsAccessChain
:
3395 case SpvOpArrayLength
:
3396 vtn_handle_variables(b
, opcode
, w
, count
);
3399 case SpvOpFunctionCall
:
3400 vtn_handle_function_call(b
, opcode
, w
, count
);
3403 case SpvOpSampledImage
:
3405 case SpvOpImageSampleImplicitLod
:
3406 case SpvOpImageSampleExplicitLod
:
3407 case SpvOpImageSampleDrefImplicitLod
:
3408 case SpvOpImageSampleDrefExplicitLod
:
3409 case SpvOpImageSampleProjImplicitLod
:
3410 case SpvOpImageSampleProjExplicitLod
:
3411 case SpvOpImageSampleProjDrefImplicitLod
:
3412 case SpvOpImageSampleProjDrefExplicitLod
:
3413 case SpvOpImageFetch
:
3414 case SpvOpImageGather
:
3415 case SpvOpImageDrefGather
:
3416 case SpvOpImageQuerySizeLod
:
3417 case SpvOpImageQueryLod
:
3418 case SpvOpImageQueryLevels
:
3419 case SpvOpImageQuerySamples
:
3420 vtn_handle_texture(b
, opcode
, w
, count
);
3423 case SpvOpImageRead
:
3424 case SpvOpImageWrite
:
3425 case SpvOpImageTexelPointer
:
3426 vtn_handle_image(b
, opcode
, w
, count
);
3429 case SpvOpImageQuerySize
: {
3430 struct vtn_pointer
*image
=
3431 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
3432 if (image
->mode
== vtn_variable_mode_image
) {
3433 vtn_handle_image(b
, opcode
, w
, count
);
3435 vtn_assert(image
->mode
== vtn_variable_mode_sampler
);
3436 vtn_handle_texture(b
, opcode
, w
, count
);
3441 case SpvOpAtomicLoad
:
3442 case SpvOpAtomicExchange
:
3443 case SpvOpAtomicCompareExchange
:
3444 case SpvOpAtomicCompareExchangeWeak
:
3445 case SpvOpAtomicIIncrement
:
3446 case SpvOpAtomicIDecrement
:
3447 case SpvOpAtomicIAdd
:
3448 case SpvOpAtomicISub
:
3449 case SpvOpAtomicSMin
:
3450 case SpvOpAtomicUMin
:
3451 case SpvOpAtomicSMax
:
3452 case SpvOpAtomicUMax
:
3453 case SpvOpAtomicAnd
:
3455 case SpvOpAtomicXor
: {
3456 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[3]);
3457 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
3458 vtn_handle_image(b
, opcode
, w
, count
);
3460 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
3461 vtn_handle_ssbo_or_shared_atomic(b
, opcode
, w
, count
);
3466 case SpvOpAtomicStore
: {
3467 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[1]);
3468 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
3469 vtn_handle_image(b
, opcode
, w
, count
);
3471 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
3472 vtn_handle_ssbo_or_shared_atomic(b
, opcode
, w
, count
);
3478 /* Handle OpSelect up-front here because it needs to be able to handle
3479 * pointers and not just regular vectors and scalars.
3481 struct vtn_type
*res_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
3482 struct vtn_ssa_value
*ssa
= vtn_create_ssa_value(b
, res_type
->type
);
3483 ssa
->def
= nir_bcsel(&b
->nb
, vtn_ssa_value(b
, w
[3])->def
,
3484 vtn_ssa_value(b
, w
[4])->def
,
3485 vtn_ssa_value(b
, w
[5])->def
);
3486 vtn_push_ssa(b
, w
[2], res_type
, ssa
);
3495 case SpvOpConvertFToU
:
3496 case SpvOpConvertFToS
:
3497 case SpvOpConvertSToF
:
3498 case SpvOpConvertUToF
:
3502 case SpvOpQuantizeToF16
:
3503 case SpvOpConvertPtrToU
:
3504 case SpvOpConvertUToPtr
:
3505 case SpvOpPtrCastToGeneric
:
3506 case SpvOpGenericCastToPtr
:
3512 case SpvOpSignBitSet
:
3513 case SpvOpLessOrGreater
:
3515 case SpvOpUnordered
:
3530 case SpvOpVectorTimesScalar
:
3532 case SpvOpIAddCarry
:
3533 case SpvOpISubBorrow
:
3534 case SpvOpUMulExtended
:
3535 case SpvOpSMulExtended
:
3536 case SpvOpShiftRightLogical
:
3537 case SpvOpShiftRightArithmetic
:
3538 case SpvOpShiftLeftLogical
:
3539 case SpvOpLogicalEqual
:
3540 case SpvOpLogicalNotEqual
:
3541 case SpvOpLogicalOr
:
3542 case SpvOpLogicalAnd
:
3543 case SpvOpLogicalNot
:
3544 case SpvOpBitwiseOr
:
3545 case SpvOpBitwiseXor
:
3546 case SpvOpBitwiseAnd
:
3548 case SpvOpFOrdEqual
:
3549 case SpvOpFUnordEqual
:
3550 case SpvOpINotEqual
:
3551 case SpvOpFOrdNotEqual
:
3552 case SpvOpFUnordNotEqual
:
3553 case SpvOpULessThan
:
3554 case SpvOpSLessThan
:
3555 case SpvOpFOrdLessThan
:
3556 case SpvOpFUnordLessThan
:
3557 case SpvOpUGreaterThan
:
3558 case SpvOpSGreaterThan
:
3559 case SpvOpFOrdGreaterThan
:
3560 case SpvOpFUnordGreaterThan
:
3561 case SpvOpULessThanEqual
:
3562 case SpvOpSLessThanEqual
:
3563 case SpvOpFOrdLessThanEqual
:
3564 case SpvOpFUnordLessThanEqual
:
3565 case SpvOpUGreaterThanEqual
:
3566 case SpvOpSGreaterThanEqual
:
3567 case SpvOpFOrdGreaterThanEqual
:
3568 case SpvOpFUnordGreaterThanEqual
:
3574 case SpvOpFwidthFine
:
3575 case SpvOpDPdxCoarse
:
3576 case SpvOpDPdyCoarse
:
3577 case SpvOpFwidthCoarse
:
3578 case SpvOpBitFieldInsert
:
3579 case SpvOpBitFieldSExtract
:
3580 case SpvOpBitFieldUExtract
:
3581 case SpvOpBitReverse
:
3583 case SpvOpTranspose
:
3584 case SpvOpOuterProduct
:
3585 case SpvOpMatrixTimesScalar
:
3586 case SpvOpVectorTimesMatrix
:
3587 case SpvOpMatrixTimesVector
:
3588 case SpvOpMatrixTimesMatrix
:
3589 vtn_handle_alu(b
, opcode
, w
, count
);
3592 case SpvOpVectorExtractDynamic
:
3593 case SpvOpVectorInsertDynamic
:
3594 case SpvOpVectorShuffle
:
3595 case SpvOpCompositeConstruct
:
3596 case SpvOpCompositeExtract
:
3597 case SpvOpCompositeInsert
:
3598 case SpvOpCopyObject
:
3599 vtn_handle_composite(b
, opcode
, w
, count
);
3602 case SpvOpEmitVertex
:
3603 case SpvOpEndPrimitive
:
3604 case SpvOpEmitStreamVertex
:
3605 case SpvOpEndStreamPrimitive
:
3606 case SpvOpControlBarrier
:
3607 case SpvOpMemoryBarrier
:
3608 vtn_handle_barrier(b
, opcode
, w
, count
);
3612 vtn_fail("Unhandled opcode");
3619 spirv_to_nir(const uint32_t *words
, size_t word_count
,
3620 struct nir_spirv_specialization
*spec
, unsigned num_spec
,
3621 gl_shader_stage stage
, const char *entry_point_name
,
3622 const struct spirv_to_nir_options
*options
,
3623 const nir_shader_compiler_options
*nir_options
)
3625 /* Initialize the stn_builder object */
3626 struct vtn_builder
*b
= rzalloc(NULL
, struct vtn_builder
);
3631 exec_list_make_empty(&b
->functions
);
3632 b
->entry_point_stage
= stage
;
3633 b
->entry_point_name
= entry_point_name
;
3634 b
->options
= options
;
3636 /* See also _vtn_fail() */
3637 if (setjmp(b
->fail_jump
)) {
3642 const uint32_t *word_end
= words
+ word_count
;
3644 /* Handle the SPIR-V header (first 4 dwords) */
3645 vtn_assert(word_count
> 5);
3647 vtn_assert(words
[0] == SpvMagicNumber
);
3648 vtn_assert(words
[1] >= 0x10000);
3649 /* words[2] == generator magic */
3650 unsigned value_id_bound
= words
[3];
3651 vtn_assert(words
[4] == 0);
3655 b
->value_id_bound
= value_id_bound
;
3656 b
->values
= rzalloc_array(b
, struct vtn_value
, value_id_bound
);
3658 /* Handle all the preamble instructions */
3659 words
= vtn_foreach_instruction(b
, words
, word_end
,
3660 vtn_handle_preamble_instruction
);
3662 if (b
->entry_point
== NULL
) {
3663 vtn_fail("Entry point not found");
3668 b
->shader
= nir_shader_create(b
, stage
, nir_options
, NULL
);
3670 /* Set shader info defaults */
3671 b
->shader
->info
.gs
.invocations
= 1;
3673 /* Parse execution modes */
3674 vtn_foreach_execution_mode(b
, b
->entry_point
,
3675 vtn_handle_execution_mode
, NULL
);
3677 b
->specializations
= spec
;
3678 b
->num_specializations
= num_spec
;
3680 /* Handle all variable, type, and constant instructions */
3681 words
= vtn_foreach_instruction(b
, words
, word_end
,
3682 vtn_handle_variable_or_type_instruction
);
3684 vtn_build_cfg(b
, words
, word_end
);
3686 assert(b
->entry_point
->value_type
== vtn_value_type_function
);
3687 b
->entry_point
->func
->referenced
= true;
3692 foreach_list_typed(struct vtn_function
, func
, node
, &b
->functions
) {
3693 if (func
->referenced
&& !func
->emitted
) {
3694 b
->const_table
= _mesa_hash_table_create(b
, _mesa_hash_pointer
,
3695 _mesa_key_pointer_equal
);
3697 vtn_function_emit(b
, func
, vtn_handle_body_instruction
);
3703 vtn_assert(b
->entry_point
->value_type
== vtn_value_type_function
);
3704 nir_function
*entry_point
= b
->entry_point
->func
->impl
->function
;
3705 vtn_assert(entry_point
);
3707 /* Unparent the shader from the vtn_builder before we delete the builder */
3708 ralloc_steal(NULL
, b
->shader
);