2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Jason Ekstrand (jason@jlekstrand.net)
28 #include "vtn_private.h"
29 #include "nir/nir_vla.h"
30 #include "nir/nir_control_flow.h"
31 #include "nir/nir_constant_expressions.h"
32 #include "spirv_info.h"
35 vtn_log(struct vtn_builder
*b
, enum nir_spirv_debug_level level
,
36 size_t spirv_offset
, const char *message
)
38 if (b
->options
->debug
.func
) {
39 b
->options
->debug
.func(b
->options
->debug
.private_data
,
40 level
, spirv_offset
, message
);
44 if (level
>= NIR_SPIRV_DEBUG_LEVEL_WARNING
)
45 fprintf(stderr
, "%s\n", message
);
50 vtn_logf(struct vtn_builder
*b
, enum nir_spirv_debug_level level
,
51 size_t spirv_offset
, const char *fmt
, ...)
57 msg
= ralloc_vasprintf(NULL
, fmt
, args
);
60 vtn_log(b
, level
, spirv_offset
, msg
);
66 vtn_log_err(struct vtn_builder
*b
,
67 enum nir_spirv_debug_level level
, const char *prefix
,
68 const char *file
, unsigned line
,
69 const char *fmt
, va_list args
)
73 msg
= ralloc_strdup(NULL
, prefix
);
76 ralloc_asprintf_append(&msg
, " In file %s:%u\n", file
, line
);
79 ralloc_asprintf_append(&msg
, " ");
81 ralloc_vasprintf_append(&msg
, fmt
, args
);
83 ralloc_asprintf_append(&msg
, "\n %zu bytes into the SPIR-V binary",
87 ralloc_asprintf_append(&msg
,
88 "\n in SPIR-V source file %s, line %d, col %d",
89 b
->file
, b
->line
, b
->col
);
92 vtn_log(b
, level
, b
->spirv_offset
, msg
);
98 _vtn_warn(struct vtn_builder
*b
, const char *file
, unsigned line
,
104 vtn_log_err(b
, NIR_SPIRV_DEBUG_LEVEL_WARNING
, "SPIR-V WARNING:\n",
105 file
, line
, fmt
, args
);
110 _vtn_fail(struct vtn_builder
*b
, const char *file
, unsigned line
,
111 const char *fmt
, ...)
116 vtn_log_err(b
, NIR_SPIRV_DEBUG_LEVEL_ERROR
, "SPIR-V parsing FAILED:\n",
117 file
, line
, fmt
, args
);
120 longjmp(b
->fail_jump
, 1);
123 struct spec_constant_value
{
131 static struct vtn_ssa_value
*
132 vtn_undef_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
134 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
137 if (glsl_type_is_vector_or_scalar(type
)) {
138 unsigned num_components
= glsl_get_vector_elements(val
->type
);
139 unsigned bit_size
= glsl_get_bit_size(val
->type
);
140 val
->def
= nir_ssa_undef(&b
->nb
, num_components
, bit_size
);
142 unsigned elems
= glsl_get_length(val
->type
);
143 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
144 if (glsl_type_is_matrix(type
)) {
145 const struct glsl_type
*elem_type
=
146 glsl_vector_type(glsl_get_base_type(type
),
147 glsl_get_vector_elements(type
));
149 for (unsigned i
= 0; i
< elems
; i
++)
150 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
151 } else if (glsl_type_is_array(type
)) {
152 const struct glsl_type
*elem_type
= glsl_get_array_element(type
);
153 for (unsigned i
= 0; i
< elems
; i
++)
154 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
156 for (unsigned i
= 0; i
< elems
; i
++) {
157 const struct glsl_type
*elem_type
= glsl_get_struct_field(type
, i
);
158 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
166 static struct vtn_ssa_value
*
167 vtn_const_ssa_value(struct vtn_builder
*b
, nir_constant
*constant
,
168 const struct glsl_type
*type
)
170 struct hash_entry
*entry
= _mesa_hash_table_search(b
->const_table
, constant
);
175 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
178 switch (glsl_get_base_type(type
)) {
181 case GLSL_TYPE_INT16
:
182 case GLSL_TYPE_UINT16
:
183 case GLSL_TYPE_INT64
:
184 case GLSL_TYPE_UINT64
:
186 case GLSL_TYPE_FLOAT
:
187 case GLSL_TYPE_FLOAT16
:
188 case GLSL_TYPE_DOUBLE
: {
189 int bit_size
= glsl_get_bit_size(type
);
190 if (glsl_type_is_vector_or_scalar(type
)) {
191 unsigned num_components
= glsl_get_vector_elements(val
->type
);
192 nir_load_const_instr
*load
=
193 nir_load_const_instr_create(b
->shader
, num_components
, bit_size
);
195 load
->value
= constant
->values
[0];
197 nir_instr_insert_before_cf_list(&b
->nb
.impl
->body
, &load
->instr
);
198 val
->def
= &load
->def
;
200 assert(glsl_type_is_matrix(type
));
201 unsigned rows
= glsl_get_vector_elements(val
->type
);
202 unsigned columns
= glsl_get_matrix_columns(val
->type
);
203 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, columns
);
205 for (unsigned i
= 0; i
< columns
; i
++) {
206 struct vtn_ssa_value
*col_val
= rzalloc(b
, struct vtn_ssa_value
);
207 col_val
->type
= glsl_get_column_type(val
->type
);
208 nir_load_const_instr
*load
=
209 nir_load_const_instr_create(b
->shader
, rows
, bit_size
);
211 load
->value
= constant
->values
[i
];
213 nir_instr_insert_before_cf_list(&b
->nb
.impl
->body
, &load
->instr
);
214 col_val
->def
= &load
->def
;
216 val
->elems
[i
] = col_val
;
222 case GLSL_TYPE_ARRAY
: {
223 unsigned elems
= glsl_get_length(val
->type
);
224 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
225 const struct glsl_type
*elem_type
= glsl_get_array_element(val
->type
);
226 for (unsigned i
= 0; i
< elems
; i
++)
227 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
232 case GLSL_TYPE_STRUCT
: {
233 unsigned elems
= glsl_get_length(val
->type
);
234 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
235 for (unsigned i
= 0; i
< elems
; i
++) {
236 const struct glsl_type
*elem_type
=
237 glsl_get_struct_field(val
->type
, i
);
238 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
245 vtn_fail("bad constant type");
251 struct vtn_ssa_value
*
252 vtn_ssa_value(struct vtn_builder
*b
, uint32_t value_id
)
254 struct vtn_value
*val
= vtn_untyped_value(b
, value_id
);
255 switch (val
->value_type
) {
256 case vtn_value_type_undef
:
257 return vtn_undef_ssa_value(b
, val
->type
->type
);
259 case vtn_value_type_constant
:
260 return vtn_const_ssa_value(b
, val
->constant
, val
->type
->type
);
262 case vtn_value_type_ssa
:
265 case vtn_value_type_pointer
:
266 vtn_assert(val
->pointer
->ptr_type
&& val
->pointer
->ptr_type
->type
);
267 struct vtn_ssa_value
*ssa
=
268 vtn_create_ssa_value(b
, val
->pointer
->ptr_type
->type
);
269 ssa
->def
= vtn_pointer_to_ssa(b
, val
->pointer
);
273 vtn_fail("Invalid type for an SSA value");
278 vtn_string_literal(struct vtn_builder
*b
, const uint32_t *words
,
279 unsigned word_count
, unsigned *words_used
)
281 char *dup
= ralloc_strndup(b
, (char *)words
, word_count
* sizeof(*words
));
283 /* Ammount of space taken by the string (including the null) */
284 unsigned len
= strlen(dup
) + 1;
285 *words_used
= DIV_ROUND_UP(len
, sizeof(*words
));
291 vtn_foreach_instruction(struct vtn_builder
*b
, const uint32_t *start
,
292 const uint32_t *end
, vtn_instruction_handler handler
)
298 const uint32_t *w
= start
;
300 SpvOp opcode
= w
[0] & SpvOpCodeMask
;
301 unsigned count
= w
[0] >> SpvWordCountShift
;
302 vtn_assert(count
>= 1 && w
+ count
<= end
);
304 b
->spirv_offset
= (uint8_t *)w
- (uint8_t *)b
->spirv
;
308 break; /* Do nothing */
311 b
->file
= vtn_value(b
, w
[1], vtn_value_type_string
)->str
;
323 if (!handler(b
, opcode
, w
, count
))
341 vtn_handle_extension(struct vtn_builder
*b
, SpvOp opcode
,
342 const uint32_t *w
, unsigned count
)
345 case SpvOpExtInstImport
: {
346 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_extension
);
347 if (strcmp((const char *)&w
[2], "GLSL.std.450") == 0) {
348 val
->ext_handler
= vtn_handle_glsl450_instruction
;
350 vtn_fail("Unsupported extension");
356 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
357 bool handled
= val
->ext_handler(b
, w
[4], w
, count
);
363 vtn_fail("Unhandled opcode");
368 _foreach_decoration_helper(struct vtn_builder
*b
,
369 struct vtn_value
*base_value
,
371 struct vtn_value
*value
,
372 vtn_decoration_foreach_cb cb
, void *data
)
374 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
376 if (dec
->scope
== VTN_DEC_DECORATION
) {
377 member
= parent_member
;
378 } else if (dec
->scope
>= VTN_DEC_STRUCT_MEMBER0
) {
379 vtn_assert(parent_member
== -1);
380 member
= dec
->scope
- VTN_DEC_STRUCT_MEMBER0
;
382 /* Not a decoration */
387 vtn_assert(dec
->group
->value_type
== vtn_value_type_decoration_group
);
388 _foreach_decoration_helper(b
, base_value
, member
, dec
->group
,
391 cb(b
, base_value
, member
, dec
, data
);
396 /** Iterates (recursively if needed) over all of the decorations on a value
398 * This function iterates over all of the decorations applied to a given
399 * value. If it encounters a decoration group, it recurses into the group
400 * and iterates over all of those decorations as well.
403 vtn_foreach_decoration(struct vtn_builder
*b
, struct vtn_value
*value
,
404 vtn_decoration_foreach_cb cb
, void *data
)
406 _foreach_decoration_helper(b
, value
, -1, value
, cb
, data
);
410 vtn_foreach_execution_mode(struct vtn_builder
*b
, struct vtn_value
*value
,
411 vtn_execution_mode_foreach_cb cb
, void *data
)
413 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
414 if (dec
->scope
!= VTN_DEC_EXECUTION_MODE
)
417 vtn_assert(dec
->group
== NULL
);
418 cb(b
, value
, dec
, data
);
423 vtn_handle_decoration(struct vtn_builder
*b
, SpvOp opcode
,
424 const uint32_t *w
, unsigned count
)
426 const uint32_t *w_end
= w
+ count
;
427 const uint32_t target
= w
[1];
431 case SpvOpDecorationGroup
:
432 vtn_push_value(b
, target
, vtn_value_type_decoration_group
);
436 case SpvOpMemberDecorate
:
437 case SpvOpExecutionMode
: {
438 struct vtn_value
*val
= &b
->values
[target
];
440 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
443 dec
->scope
= VTN_DEC_DECORATION
;
445 case SpvOpMemberDecorate
:
446 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(w
++);
448 case SpvOpExecutionMode
:
449 dec
->scope
= VTN_DEC_EXECUTION_MODE
;
452 vtn_fail("Invalid decoration opcode");
454 dec
->decoration
= *(w
++);
457 /* Link into the list */
458 dec
->next
= val
->decoration
;
459 val
->decoration
= dec
;
463 case SpvOpGroupMemberDecorate
:
464 case SpvOpGroupDecorate
: {
465 struct vtn_value
*group
=
466 vtn_value(b
, target
, vtn_value_type_decoration_group
);
468 for (; w
< w_end
; w
++) {
469 struct vtn_value
*val
= vtn_untyped_value(b
, *w
);
470 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
473 if (opcode
== SpvOpGroupDecorate
) {
474 dec
->scope
= VTN_DEC_DECORATION
;
476 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(++w
);
479 /* Link into the list */
480 dec
->next
= val
->decoration
;
481 val
->decoration
= dec
;
487 vtn_fail("Unhandled opcode");
491 struct member_decoration_ctx
{
493 struct glsl_struct_field
*fields
;
494 struct vtn_type
*type
;
497 /* does a shallow copy of a vtn_type */
499 static struct vtn_type
*
500 vtn_type_copy(struct vtn_builder
*b
, struct vtn_type
*src
)
502 struct vtn_type
*dest
= ralloc(b
, struct vtn_type
);
505 switch (src
->base_type
) {
506 case vtn_base_type_void
:
507 case vtn_base_type_scalar
:
508 case vtn_base_type_vector
:
509 case vtn_base_type_matrix
:
510 case vtn_base_type_array
:
511 case vtn_base_type_pointer
:
512 case vtn_base_type_image
:
513 case vtn_base_type_sampler
:
514 case vtn_base_type_sampled_image
:
515 /* Nothing more to do */
518 case vtn_base_type_struct
:
519 dest
->members
= ralloc_array(b
, struct vtn_type
*, src
->length
);
520 memcpy(dest
->members
, src
->members
,
521 src
->length
* sizeof(src
->members
[0]));
523 dest
->offsets
= ralloc_array(b
, unsigned, src
->length
);
524 memcpy(dest
->offsets
, src
->offsets
,
525 src
->length
* sizeof(src
->offsets
[0]));
528 case vtn_base_type_function
:
529 dest
->params
= ralloc_array(b
, struct vtn_type
*, src
->length
);
530 memcpy(dest
->params
, src
->params
, src
->length
* sizeof(src
->params
[0]));
537 static struct vtn_type
*
538 mutable_matrix_member(struct vtn_builder
*b
, struct vtn_type
*type
, int member
)
540 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
541 type
= type
->members
[member
];
543 /* We may have an array of matrices.... Oh, joy! */
544 while (glsl_type_is_array(type
->type
)) {
545 type
->array_element
= vtn_type_copy(b
, type
->array_element
);
546 type
= type
->array_element
;
549 vtn_assert(glsl_type_is_matrix(type
->type
));
555 struct_member_decoration_cb(struct vtn_builder
*b
,
556 struct vtn_value
*val
, int member
,
557 const struct vtn_decoration
*dec
, void *void_ctx
)
559 struct member_decoration_ctx
*ctx
= void_ctx
;
564 vtn_assert(member
< ctx
->num_fields
);
566 switch (dec
->decoration
) {
567 case SpvDecorationNonWritable
:
568 case SpvDecorationNonReadable
:
569 case SpvDecorationRelaxedPrecision
:
570 case SpvDecorationVolatile
:
571 case SpvDecorationCoherent
:
572 case SpvDecorationUniform
:
573 break; /* FIXME: Do nothing with this for now. */
574 case SpvDecorationNoPerspective
:
575 ctx
->fields
[member
].interpolation
= INTERP_MODE_NOPERSPECTIVE
;
577 case SpvDecorationFlat
:
578 ctx
->fields
[member
].interpolation
= INTERP_MODE_FLAT
;
580 case SpvDecorationCentroid
:
581 ctx
->fields
[member
].centroid
= true;
583 case SpvDecorationSample
:
584 ctx
->fields
[member
].sample
= true;
586 case SpvDecorationStream
:
587 /* Vulkan only allows one GS stream */
588 vtn_assert(dec
->literals
[0] == 0);
590 case SpvDecorationLocation
:
591 ctx
->fields
[member
].location
= dec
->literals
[0];
593 case SpvDecorationComponent
:
594 break; /* FIXME: What should we do with these? */
595 case SpvDecorationBuiltIn
:
596 ctx
->type
->members
[member
] = vtn_type_copy(b
, ctx
->type
->members
[member
]);
597 ctx
->type
->members
[member
]->is_builtin
= true;
598 ctx
->type
->members
[member
]->builtin
= dec
->literals
[0];
599 ctx
->type
->builtin_block
= true;
601 case SpvDecorationOffset
:
602 ctx
->type
->offsets
[member
] = dec
->literals
[0];
604 case SpvDecorationMatrixStride
:
605 /* Handled as a second pass */
607 case SpvDecorationColMajor
:
608 break; /* Nothing to do here. Column-major is the default. */
609 case SpvDecorationRowMajor
:
610 mutable_matrix_member(b
, ctx
->type
, member
)->row_major
= true;
613 case SpvDecorationPatch
:
616 case SpvDecorationSpecId
:
617 case SpvDecorationBlock
:
618 case SpvDecorationBufferBlock
:
619 case SpvDecorationArrayStride
:
620 case SpvDecorationGLSLShared
:
621 case SpvDecorationGLSLPacked
:
622 case SpvDecorationInvariant
:
623 case SpvDecorationRestrict
:
624 case SpvDecorationAliased
:
625 case SpvDecorationConstant
:
626 case SpvDecorationIndex
:
627 case SpvDecorationBinding
:
628 case SpvDecorationDescriptorSet
:
629 case SpvDecorationLinkageAttributes
:
630 case SpvDecorationNoContraction
:
631 case SpvDecorationInputAttachmentIndex
:
632 vtn_warn("Decoration not allowed on struct members: %s",
633 spirv_decoration_to_string(dec
->decoration
));
636 case SpvDecorationXfbBuffer
:
637 case SpvDecorationXfbStride
:
638 vtn_warn("Vulkan does not have transform feedback");
641 case SpvDecorationCPacked
:
642 case SpvDecorationSaturatedConversion
:
643 case SpvDecorationFuncParamAttr
:
644 case SpvDecorationFPRoundingMode
:
645 case SpvDecorationFPFastMathMode
:
646 case SpvDecorationAlignment
:
647 vtn_warn("Decoration only allowed for CL-style kernels: %s",
648 spirv_decoration_to_string(dec
->decoration
));
652 vtn_fail("Unhandled decoration");
656 /* Matrix strides are handled as a separate pass because we need to know
657 * whether the matrix is row-major or not first.
660 struct_member_matrix_stride_cb(struct vtn_builder
*b
,
661 struct vtn_value
*val
, int member
,
662 const struct vtn_decoration
*dec
,
665 if (dec
->decoration
!= SpvDecorationMatrixStride
)
667 vtn_assert(member
>= 0);
669 struct member_decoration_ctx
*ctx
= void_ctx
;
671 struct vtn_type
*mat_type
= mutable_matrix_member(b
, ctx
->type
, member
);
672 if (mat_type
->row_major
) {
673 mat_type
->array_element
= vtn_type_copy(b
, mat_type
->array_element
);
674 mat_type
->stride
= mat_type
->array_element
->stride
;
675 mat_type
->array_element
->stride
= dec
->literals
[0];
677 vtn_assert(mat_type
->array_element
->stride
> 0);
678 mat_type
->stride
= dec
->literals
[0];
683 type_decoration_cb(struct vtn_builder
*b
,
684 struct vtn_value
*val
, int member
,
685 const struct vtn_decoration
*dec
, void *ctx
)
687 struct vtn_type
*type
= val
->type
;
692 switch (dec
->decoration
) {
693 case SpvDecorationArrayStride
:
694 vtn_assert(type
->base_type
== vtn_base_type_matrix
||
695 type
->base_type
== vtn_base_type_array
||
696 type
->base_type
== vtn_base_type_pointer
);
697 type
->stride
= dec
->literals
[0];
699 case SpvDecorationBlock
:
700 vtn_assert(type
->base_type
== vtn_base_type_struct
);
703 case SpvDecorationBufferBlock
:
704 vtn_assert(type
->base_type
== vtn_base_type_struct
);
705 type
->buffer_block
= true;
707 case SpvDecorationGLSLShared
:
708 case SpvDecorationGLSLPacked
:
709 /* Ignore these, since we get explicit offsets anyways */
712 case SpvDecorationRowMajor
:
713 case SpvDecorationColMajor
:
714 case SpvDecorationMatrixStride
:
715 case SpvDecorationBuiltIn
:
716 case SpvDecorationNoPerspective
:
717 case SpvDecorationFlat
:
718 case SpvDecorationPatch
:
719 case SpvDecorationCentroid
:
720 case SpvDecorationSample
:
721 case SpvDecorationVolatile
:
722 case SpvDecorationCoherent
:
723 case SpvDecorationNonWritable
:
724 case SpvDecorationNonReadable
:
725 case SpvDecorationUniform
:
726 case SpvDecorationStream
:
727 case SpvDecorationLocation
:
728 case SpvDecorationComponent
:
729 case SpvDecorationOffset
:
730 case SpvDecorationXfbBuffer
:
731 case SpvDecorationXfbStride
:
732 vtn_warn("Decoration only allowed for struct members: %s",
733 spirv_decoration_to_string(dec
->decoration
));
736 case SpvDecorationRelaxedPrecision
:
737 case SpvDecorationSpecId
:
738 case SpvDecorationInvariant
:
739 case SpvDecorationRestrict
:
740 case SpvDecorationAliased
:
741 case SpvDecorationConstant
:
742 case SpvDecorationIndex
:
743 case SpvDecorationBinding
:
744 case SpvDecorationDescriptorSet
:
745 case SpvDecorationLinkageAttributes
:
746 case SpvDecorationNoContraction
:
747 case SpvDecorationInputAttachmentIndex
:
748 vtn_warn("Decoration not allowed on types: %s",
749 spirv_decoration_to_string(dec
->decoration
));
752 case SpvDecorationCPacked
:
753 case SpvDecorationSaturatedConversion
:
754 case SpvDecorationFuncParamAttr
:
755 case SpvDecorationFPRoundingMode
:
756 case SpvDecorationFPFastMathMode
:
757 case SpvDecorationAlignment
:
758 vtn_warn("Decoration only allowed for CL-style kernels: %s",
759 spirv_decoration_to_string(dec
->decoration
));
763 vtn_fail("Unhandled decoration");
768 translate_image_format(struct vtn_builder
*b
, SpvImageFormat format
)
771 case SpvImageFormatUnknown
: return 0; /* GL_NONE */
772 case SpvImageFormatRgba32f
: return 0x8814; /* GL_RGBA32F */
773 case SpvImageFormatRgba16f
: return 0x881A; /* GL_RGBA16F */
774 case SpvImageFormatR32f
: return 0x822E; /* GL_R32F */
775 case SpvImageFormatRgba8
: return 0x8058; /* GL_RGBA8 */
776 case SpvImageFormatRgba8Snorm
: return 0x8F97; /* GL_RGBA8_SNORM */
777 case SpvImageFormatRg32f
: return 0x8230; /* GL_RG32F */
778 case SpvImageFormatRg16f
: return 0x822F; /* GL_RG16F */
779 case SpvImageFormatR11fG11fB10f
: return 0x8C3A; /* GL_R11F_G11F_B10F */
780 case SpvImageFormatR16f
: return 0x822D; /* GL_R16F */
781 case SpvImageFormatRgba16
: return 0x805B; /* GL_RGBA16 */
782 case SpvImageFormatRgb10A2
: return 0x8059; /* GL_RGB10_A2 */
783 case SpvImageFormatRg16
: return 0x822C; /* GL_RG16 */
784 case SpvImageFormatRg8
: return 0x822B; /* GL_RG8 */
785 case SpvImageFormatR16
: return 0x822A; /* GL_R16 */
786 case SpvImageFormatR8
: return 0x8229; /* GL_R8 */
787 case SpvImageFormatRgba16Snorm
: return 0x8F9B; /* GL_RGBA16_SNORM */
788 case SpvImageFormatRg16Snorm
: return 0x8F99; /* GL_RG16_SNORM */
789 case SpvImageFormatRg8Snorm
: return 0x8F95; /* GL_RG8_SNORM */
790 case SpvImageFormatR16Snorm
: return 0x8F98; /* GL_R16_SNORM */
791 case SpvImageFormatR8Snorm
: return 0x8F94; /* GL_R8_SNORM */
792 case SpvImageFormatRgba32i
: return 0x8D82; /* GL_RGBA32I */
793 case SpvImageFormatRgba16i
: return 0x8D88; /* GL_RGBA16I */
794 case SpvImageFormatRgba8i
: return 0x8D8E; /* GL_RGBA8I */
795 case SpvImageFormatR32i
: return 0x8235; /* GL_R32I */
796 case SpvImageFormatRg32i
: return 0x823B; /* GL_RG32I */
797 case SpvImageFormatRg16i
: return 0x8239; /* GL_RG16I */
798 case SpvImageFormatRg8i
: return 0x8237; /* GL_RG8I */
799 case SpvImageFormatR16i
: return 0x8233; /* GL_R16I */
800 case SpvImageFormatR8i
: return 0x8231; /* GL_R8I */
801 case SpvImageFormatRgba32ui
: return 0x8D70; /* GL_RGBA32UI */
802 case SpvImageFormatRgba16ui
: return 0x8D76; /* GL_RGBA16UI */
803 case SpvImageFormatRgba8ui
: return 0x8D7C; /* GL_RGBA8UI */
804 case SpvImageFormatR32ui
: return 0x8236; /* GL_R32UI */
805 case SpvImageFormatRgb10a2ui
: return 0x906F; /* GL_RGB10_A2UI */
806 case SpvImageFormatRg32ui
: return 0x823C; /* GL_RG32UI */
807 case SpvImageFormatRg16ui
: return 0x823A; /* GL_RG16UI */
808 case SpvImageFormatRg8ui
: return 0x8238; /* GL_RG8UI */
809 case SpvImageFormatR16ui
: return 0x8234; /* GL_R16UI */
810 case SpvImageFormatR8ui
: return 0x8232; /* GL_R8UI */
812 vtn_fail("Invalid image format");
816 static struct vtn_type
*
817 vtn_type_layout_std430(struct vtn_builder
*b
, struct vtn_type
*type
,
818 uint32_t *size_out
, uint32_t *align_out
)
820 switch (type
->base_type
) {
821 case vtn_base_type_scalar
: {
822 uint32_t comp_size
= glsl_get_bit_size(type
->type
) / 8;
823 *size_out
= comp_size
;
824 *align_out
= comp_size
;
828 case vtn_base_type_vector
: {
829 uint32_t comp_size
= glsl_get_bit_size(type
->type
) / 8;
830 assert(type
->length
> 0 && type
->length
<= 4);
831 unsigned align_comps
= type
->length
== 3 ? 4 : type
->length
;
832 *size_out
= comp_size
* type
->length
,
833 *align_out
= comp_size
* align_comps
;
837 case vtn_base_type_matrix
:
838 case vtn_base_type_array
: {
839 /* We're going to add an array stride */
840 type
= vtn_type_copy(b
, type
);
841 uint32_t elem_size
, elem_align
;
842 type
->array_element
= vtn_type_layout_std430(b
, type
->array_element
,
843 &elem_size
, &elem_align
);
844 type
->stride
= vtn_align_u32(elem_size
, elem_align
);
845 *size_out
= type
->stride
* type
->length
;
846 *align_out
= elem_align
;
850 case vtn_base_type_struct
: {
851 /* We're going to add member offsets */
852 type
= vtn_type_copy(b
, type
);
855 for (unsigned i
= 0; i
< type
->length
; i
++) {
856 uint32_t mem_size
, mem_align
;
857 type
->members
[i
] = vtn_type_layout_std430(b
, type
->members
[i
],
858 &mem_size
, &mem_align
);
859 offset
= vtn_align_u32(offset
, mem_align
);
860 type
->offsets
[i
] = offset
;
862 align
= MAX2(align
, mem_align
);
870 unreachable("Invalid SPIR-V type for std430");
875 vtn_handle_type(struct vtn_builder
*b
, SpvOp opcode
,
876 const uint32_t *w
, unsigned count
)
878 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_type
);
880 val
->type
= rzalloc(b
, struct vtn_type
);
881 val
->type
->val
= val
;
885 val
->type
->base_type
= vtn_base_type_void
;
886 val
->type
->type
= glsl_void_type();
889 val
->type
->base_type
= vtn_base_type_scalar
;
890 val
->type
->type
= glsl_bool_type();
891 val
->type
->length
= 1;
895 const bool signedness
= w
[3];
896 val
->type
->base_type
= vtn_base_type_scalar
;
899 val
->type
->type
= (signedness
? glsl_int64_t_type() : glsl_uint64_t_type());
902 val
->type
->type
= (signedness
? glsl_int_type() : glsl_uint_type());
905 val
->type
->type
= (signedness
? glsl_int16_t_type() : glsl_uint16_t_type());
908 vtn_fail("Invalid int bit size");
910 val
->type
->length
= 1;
914 case SpvOpTypeFloat
: {
916 val
->type
->base_type
= vtn_base_type_scalar
;
919 val
->type
->type
= glsl_float16_t_type();
922 val
->type
->type
= glsl_float_type();
925 val
->type
->type
= glsl_double_type();
928 vtn_fail("Invalid float bit size");
930 val
->type
->length
= 1;
934 case SpvOpTypeVector
: {
935 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
936 unsigned elems
= w
[3];
938 vtn_fail_if(base
->base_type
!= vtn_base_type_scalar
,
939 "Base type for OpTypeVector must be a scalar");
940 vtn_fail_if(elems
< 2 || elems
> 4,
941 "Invalid component count for OpTypeVector");
943 val
->type
->base_type
= vtn_base_type_vector
;
944 val
->type
->type
= glsl_vector_type(glsl_get_base_type(base
->type
), elems
);
945 val
->type
->length
= elems
;
946 val
->type
->stride
= glsl_get_bit_size(base
->type
) / 8;
947 val
->type
->array_element
= base
;
951 case SpvOpTypeMatrix
: {
952 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
953 unsigned columns
= w
[3];
955 vtn_fail_if(base
->base_type
!= vtn_base_type_vector
,
956 "Base type for OpTypeMatrix must be a vector");
957 vtn_fail_if(columns
< 2 || columns
> 4,
958 "Invalid column count for OpTypeMatrix");
960 val
->type
->base_type
= vtn_base_type_matrix
;
961 val
->type
->type
= glsl_matrix_type(glsl_get_base_type(base
->type
),
962 glsl_get_vector_elements(base
->type
),
964 vtn_fail_if(glsl_type_is_error(val
->type
->type
),
965 "Unsupported base type for OpTypeMatrix");
966 assert(!glsl_type_is_error(val
->type
->type
));
967 val
->type
->length
= columns
;
968 val
->type
->array_element
= base
;
969 val
->type
->row_major
= false;
970 val
->type
->stride
= 0;
974 case SpvOpTypeRuntimeArray
:
975 case SpvOpTypeArray
: {
976 struct vtn_type
*array_element
=
977 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
979 if (opcode
== SpvOpTypeRuntimeArray
) {
980 /* A length of 0 is used to denote unsized arrays */
981 val
->type
->length
= 0;
984 vtn_value(b
, w
[3], vtn_value_type_constant
)->constant
->values
[0].u32
[0];
987 val
->type
->base_type
= vtn_base_type_array
;
988 val
->type
->type
= glsl_array_type(array_element
->type
, val
->type
->length
);
989 val
->type
->array_element
= array_element
;
990 val
->type
->stride
= 0;
994 case SpvOpTypeStruct
: {
995 unsigned num_fields
= count
- 2;
996 val
->type
->base_type
= vtn_base_type_struct
;
997 val
->type
->length
= num_fields
;
998 val
->type
->members
= ralloc_array(b
, struct vtn_type
*, num_fields
);
999 val
->type
->offsets
= ralloc_array(b
, unsigned, num_fields
);
1001 NIR_VLA(struct glsl_struct_field
, fields
, count
);
1002 for (unsigned i
= 0; i
< num_fields
; i
++) {
1003 val
->type
->members
[i
] =
1004 vtn_value(b
, w
[i
+ 2], vtn_value_type_type
)->type
;
1005 fields
[i
] = (struct glsl_struct_field
) {
1006 .type
= val
->type
->members
[i
]->type
,
1007 .name
= ralloc_asprintf(b
, "field%d", i
),
1012 struct member_decoration_ctx ctx
= {
1013 .num_fields
= num_fields
,
1018 vtn_foreach_decoration(b
, val
, struct_member_decoration_cb
, &ctx
);
1019 vtn_foreach_decoration(b
, val
, struct_member_matrix_stride_cb
, &ctx
);
1021 const char *name
= val
->name
? val
->name
: "struct";
1023 val
->type
->type
= glsl_struct_type(fields
, num_fields
, name
);
1027 case SpvOpTypeFunction
: {
1028 val
->type
->base_type
= vtn_base_type_function
;
1029 val
->type
->type
= NULL
;
1031 val
->type
->return_type
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1033 const unsigned num_params
= count
- 3;
1034 val
->type
->length
= num_params
;
1035 val
->type
->params
= ralloc_array(b
, struct vtn_type
*, num_params
);
1036 for (unsigned i
= 0; i
< count
- 3; i
++) {
1037 val
->type
->params
[i
] =
1038 vtn_value(b
, w
[i
+ 3], vtn_value_type_type
)->type
;
1043 case SpvOpTypePointer
: {
1044 SpvStorageClass storage_class
= w
[2];
1045 struct vtn_type
*deref_type
=
1046 vtn_value(b
, w
[3], vtn_value_type_type
)->type
;
1048 val
->type
->base_type
= vtn_base_type_pointer
;
1049 val
->type
->storage_class
= storage_class
;
1050 val
->type
->deref
= deref_type
;
1052 if (storage_class
== SpvStorageClassUniform
||
1053 storage_class
== SpvStorageClassStorageBuffer
) {
1054 /* These can actually be stored to nir_variables and used as SSA
1055 * values so they need a real glsl_type.
1057 val
->type
->type
= glsl_vector_type(GLSL_TYPE_UINT
, 2);
1060 if (storage_class
== SpvStorageClassWorkgroup
&&
1061 b
->options
->lower_workgroup_access_to_offsets
) {
1062 uint32_t size
, align
;
1063 val
->type
->deref
= vtn_type_layout_std430(b
, val
->type
->deref
,
1065 val
->type
->length
= size
;
1066 val
->type
->align
= align
;
1067 /* These can actually be stored to nir_variables and used as SSA
1068 * values so they need a real glsl_type.
1070 val
->type
->type
= glsl_uint_type();
1075 case SpvOpTypeImage
: {
1076 val
->type
->base_type
= vtn_base_type_image
;
1078 const struct glsl_type
*sampled_type
=
1079 vtn_value(b
, w
[2], vtn_value_type_type
)->type
->type
;
1081 vtn_assert(glsl_type_is_vector_or_scalar(sampled_type
));
1083 enum glsl_sampler_dim dim
;
1084 switch ((SpvDim
)w
[3]) {
1085 case SpvDim1D
: dim
= GLSL_SAMPLER_DIM_1D
; break;
1086 case SpvDim2D
: dim
= GLSL_SAMPLER_DIM_2D
; break;
1087 case SpvDim3D
: dim
= GLSL_SAMPLER_DIM_3D
; break;
1088 case SpvDimCube
: dim
= GLSL_SAMPLER_DIM_CUBE
; break;
1089 case SpvDimRect
: dim
= GLSL_SAMPLER_DIM_RECT
; break;
1090 case SpvDimBuffer
: dim
= GLSL_SAMPLER_DIM_BUF
; break;
1091 case SpvDimSubpassData
: dim
= GLSL_SAMPLER_DIM_SUBPASS
; break;
1093 vtn_fail("Invalid SPIR-V Sampler dimension");
1096 bool is_shadow
= w
[4];
1097 bool is_array
= w
[5];
1098 bool multisampled
= w
[6];
1099 unsigned sampled
= w
[7];
1100 SpvImageFormat format
= w
[8];
1103 val
->type
->access_qualifier
= w
[9];
1105 val
->type
->access_qualifier
= SpvAccessQualifierReadWrite
;
1108 if (dim
== GLSL_SAMPLER_DIM_2D
)
1109 dim
= GLSL_SAMPLER_DIM_MS
;
1110 else if (dim
== GLSL_SAMPLER_DIM_SUBPASS
)
1111 dim
= GLSL_SAMPLER_DIM_SUBPASS_MS
;
1113 vtn_fail("Unsupported multisampled image type");
1116 val
->type
->image_format
= translate_image_format(b
, format
);
1119 val
->type
->sampled
= true;
1120 val
->type
->type
= glsl_sampler_type(dim
, is_shadow
, is_array
,
1121 glsl_get_base_type(sampled_type
));
1122 } else if (sampled
== 2) {
1123 vtn_assert(!is_shadow
);
1124 val
->type
->sampled
= false;
1125 val
->type
->type
= glsl_image_type(dim
, is_array
,
1126 glsl_get_base_type(sampled_type
));
1128 vtn_fail("We need to know if the image will be sampled");
1133 case SpvOpTypeSampledImage
:
1134 val
->type
->base_type
= vtn_base_type_sampled_image
;
1135 val
->type
->image
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1136 val
->type
->type
= val
->type
->image
->type
;
1139 case SpvOpTypeSampler
:
1140 /* The actual sampler type here doesn't really matter. It gets
1141 * thrown away the moment you combine it with an image. What really
1142 * matters is that it's a sampler type as opposed to an integer type
1143 * so the backend knows what to do.
1145 val
->type
->base_type
= vtn_base_type_sampler
;
1146 val
->type
->type
= glsl_bare_sampler_type();
1149 case SpvOpTypeOpaque
:
1150 case SpvOpTypeEvent
:
1151 case SpvOpTypeDeviceEvent
:
1152 case SpvOpTypeReserveId
:
1153 case SpvOpTypeQueue
:
1156 vtn_fail("Unhandled opcode");
1159 vtn_foreach_decoration(b
, val
, type_decoration_cb
, NULL
);
1162 static nir_constant
*
1163 vtn_null_constant(struct vtn_builder
*b
, const struct glsl_type
*type
)
1165 nir_constant
*c
= rzalloc(b
, nir_constant
);
1167 /* For pointers and other typeless things, we have to return something but
1168 * it doesn't matter what.
1173 switch (glsl_get_base_type(type
)) {
1175 case GLSL_TYPE_UINT
:
1176 case GLSL_TYPE_INT16
:
1177 case GLSL_TYPE_UINT16
:
1178 case GLSL_TYPE_INT64
:
1179 case GLSL_TYPE_UINT64
:
1180 case GLSL_TYPE_BOOL
:
1181 case GLSL_TYPE_FLOAT
:
1182 case GLSL_TYPE_FLOAT16
:
1183 case GLSL_TYPE_DOUBLE
:
1184 /* Nothing to do here. It's already initialized to zero */
1187 case GLSL_TYPE_ARRAY
:
1188 vtn_assert(glsl_get_length(type
) > 0);
1189 c
->num_elements
= glsl_get_length(type
);
1190 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1192 c
->elements
[0] = vtn_null_constant(b
, glsl_get_array_element(type
));
1193 for (unsigned i
= 1; i
< c
->num_elements
; i
++)
1194 c
->elements
[i
] = c
->elements
[0];
1197 case GLSL_TYPE_STRUCT
:
1198 c
->num_elements
= glsl_get_length(type
);
1199 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1201 for (unsigned i
= 0; i
< c
->num_elements
; i
++) {
1202 c
->elements
[i
] = vtn_null_constant(b
, glsl_get_struct_field(type
, i
));
1207 vtn_fail("Invalid type for null constant");
1214 spec_constant_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*v
,
1215 int member
, const struct vtn_decoration
*dec
,
1218 vtn_assert(member
== -1);
1219 if (dec
->decoration
!= SpvDecorationSpecId
)
1222 struct spec_constant_value
*const_value
= data
;
1224 for (unsigned i
= 0; i
< b
->num_specializations
; i
++) {
1225 if (b
->specializations
[i
].id
== dec
->literals
[0]) {
1226 if (const_value
->is_double
)
1227 const_value
->data64
= b
->specializations
[i
].data64
;
1229 const_value
->data32
= b
->specializations
[i
].data32
;
1236 get_specialization(struct vtn_builder
*b
, struct vtn_value
*val
,
1237 uint32_t const_value
)
1239 struct spec_constant_value data
;
1240 data
.is_double
= false;
1241 data
.data32
= const_value
;
1242 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1247 get_specialization64(struct vtn_builder
*b
, struct vtn_value
*val
,
1248 uint64_t const_value
)
1250 struct spec_constant_value data
;
1251 data
.is_double
= true;
1252 data
.data64
= const_value
;
1253 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1258 handle_workgroup_size_decoration_cb(struct vtn_builder
*b
,
1259 struct vtn_value
*val
,
1261 const struct vtn_decoration
*dec
,
1264 vtn_assert(member
== -1);
1265 if (dec
->decoration
!= SpvDecorationBuiltIn
||
1266 dec
->literals
[0] != SpvBuiltInWorkgroupSize
)
1269 vtn_assert(val
->type
->type
== glsl_vector_type(GLSL_TYPE_UINT
, 3));
1271 b
->shader
->info
.cs
.local_size
[0] = val
->constant
->values
[0].u32
[0];
1272 b
->shader
->info
.cs
.local_size
[1] = val
->constant
->values
[0].u32
[1];
1273 b
->shader
->info
.cs
.local_size
[2] = val
->constant
->values
[0].u32
[2];
1277 vtn_handle_constant(struct vtn_builder
*b
, SpvOp opcode
,
1278 const uint32_t *w
, unsigned count
)
1280 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_constant
);
1281 val
->constant
= rzalloc(b
, nir_constant
);
1283 case SpvOpConstantTrue
:
1284 case SpvOpConstantFalse
:
1285 case SpvOpSpecConstantTrue
:
1286 case SpvOpSpecConstantFalse
: {
1287 vtn_fail_if(val
->type
->type
!= glsl_bool_type(),
1288 "Result type of %s must be OpTypeBool",
1289 spirv_op_to_string(opcode
));
1291 uint32_t int_val
= (opcode
== SpvOpConstantTrue
||
1292 opcode
== SpvOpSpecConstantTrue
);
1294 if (opcode
== SpvOpSpecConstantTrue
||
1295 opcode
== SpvOpSpecConstantFalse
)
1296 int_val
= get_specialization(b
, val
, int_val
);
1298 val
->constant
->values
[0].u32
[0] = int_val
? NIR_TRUE
: NIR_FALSE
;
1302 case SpvOpConstant
: {
1303 vtn_assert(glsl_type_is_scalar(val
->type
->type
));
1304 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1307 val
->constant
->values
->u64
[0] = vtn_u64_literal(&w
[3]);
1310 val
->constant
->values
->u32
[0] = w
[3];
1313 val
->constant
->values
->u16
[0] = w
[3];
1316 vtn_fail("Unsupported SpvOpConstant bit size");
1320 case SpvOpSpecConstant
: {
1321 vtn_assert(glsl_type_is_scalar(val
->type
->type
));
1322 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1325 val
->constant
->values
[0].u64
[0] =
1326 get_specialization64(b
, val
, vtn_u64_literal(&w
[3]));
1329 val
->constant
->values
[0].u32
[0] = get_specialization(b
, val
, w
[3]);
1332 val
->constant
->values
[0].u16
[0] = get_specialization(b
, val
, w
[3]);
1335 vtn_fail("Unsupported SpvOpSpecConstant bit size");
1339 case SpvOpSpecConstantComposite
:
1340 case SpvOpConstantComposite
: {
1341 unsigned elem_count
= count
- 3;
1342 nir_constant
**elems
= ralloc_array(b
, nir_constant
*, elem_count
);
1343 for (unsigned i
= 0; i
< elem_count
; i
++)
1344 elems
[i
] = vtn_value(b
, w
[i
+ 3], vtn_value_type_constant
)->constant
;
1346 switch (glsl_get_base_type(val
->type
->type
)) {
1347 case GLSL_TYPE_UINT
:
1349 case GLSL_TYPE_UINT16
:
1350 case GLSL_TYPE_INT16
:
1351 case GLSL_TYPE_UINT64
:
1352 case GLSL_TYPE_INT64
:
1353 case GLSL_TYPE_FLOAT
:
1354 case GLSL_TYPE_FLOAT16
:
1355 case GLSL_TYPE_BOOL
:
1356 case GLSL_TYPE_DOUBLE
: {
1357 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1358 if (glsl_type_is_matrix(val
->type
->type
)) {
1359 vtn_assert(glsl_get_matrix_columns(val
->type
->type
) == elem_count
);
1360 for (unsigned i
= 0; i
< elem_count
; i
++)
1361 val
->constant
->values
[i
] = elems
[i
]->values
[0];
1363 vtn_assert(glsl_type_is_vector(val
->type
->type
));
1364 vtn_assert(glsl_get_vector_elements(val
->type
->type
) == elem_count
);
1365 for (unsigned i
= 0; i
< elem_count
; i
++) {
1368 val
->constant
->values
[0].u64
[i
] = elems
[i
]->values
[0].u64
[0];
1371 val
->constant
->values
[0].u32
[i
] = elems
[i
]->values
[0].u32
[0];
1374 val
->constant
->values
[0].u16
[i
] = elems
[i
]->values
[0].u16
[0];
1377 vtn_fail("Invalid SpvOpConstantComposite bit size");
1384 case GLSL_TYPE_STRUCT
:
1385 case GLSL_TYPE_ARRAY
:
1386 ralloc_steal(val
->constant
, elems
);
1387 val
->constant
->num_elements
= elem_count
;
1388 val
->constant
->elements
= elems
;
1392 vtn_fail("Unsupported type for constants");
1397 case SpvOpSpecConstantOp
: {
1398 SpvOp opcode
= get_specialization(b
, val
, w
[3]);
1400 case SpvOpVectorShuffle
: {
1401 struct vtn_value
*v0
= &b
->values
[w
[4]];
1402 struct vtn_value
*v1
= &b
->values
[w
[5]];
1404 vtn_assert(v0
->value_type
== vtn_value_type_constant
||
1405 v0
->value_type
== vtn_value_type_undef
);
1406 vtn_assert(v1
->value_type
== vtn_value_type_constant
||
1407 v1
->value_type
== vtn_value_type_undef
);
1409 unsigned len0
= glsl_get_vector_elements(v0
->type
->type
);
1410 unsigned len1
= glsl_get_vector_elements(v1
->type
->type
);
1412 vtn_assert(len0
+ len1
< 16);
1414 unsigned bit_size
= glsl_get_bit_size(val
->type
->type
);
1415 unsigned bit_size0
= glsl_get_bit_size(v0
->type
->type
);
1416 unsigned bit_size1
= glsl_get_bit_size(v1
->type
->type
);
1418 vtn_assert(bit_size
== bit_size0
&& bit_size
== bit_size1
);
1419 (void)bit_size0
; (void)bit_size1
;
1421 if (bit_size
== 64) {
1423 if (v0
->value_type
== vtn_value_type_constant
) {
1424 for (unsigned i
= 0; i
< len0
; i
++)
1425 u64
[i
] = v0
->constant
->values
[0].u64
[i
];
1427 if (v1
->value_type
== vtn_value_type_constant
) {
1428 for (unsigned i
= 0; i
< len1
; i
++)
1429 u64
[len0
+ i
] = v1
->constant
->values
[0].u64
[i
];
1432 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1433 uint32_t comp
= w
[i
+ 6];
1434 /* If component is not used, set the value to a known constant
1435 * to detect if it is wrongly used.
1437 if (comp
== (uint32_t)-1)
1438 val
->constant
->values
[0].u64
[j
] = 0xdeadbeefdeadbeef;
1440 val
->constant
->values
[0].u64
[j
] = u64
[comp
];
1443 /* This is for both 32-bit and 16-bit values */
1445 if (v0
->value_type
== vtn_value_type_constant
) {
1446 for (unsigned i
= 0; i
< len0
; i
++)
1447 u32
[i
] = v0
->constant
->values
[0].u32
[i
];
1449 if (v1
->value_type
== vtn_value_type_constant
) {
1450 for (unsigned i
= 0; i
< len1
; i
++)
1451 u32
[len0
+ i
] = v1
->constant
->values
[0].u32
[i
];
1454 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1455 uint32_t comp
= w
[i
+ 6];
1456 /* If component is not used, set the value to a known constant
1457 * to detect if it is wrongly used.
1459 if (comp
== (uint32_t)-1)
1460 val
->constant
->values
[0].u32
[j
] = 0xdeadbeef;
1462 val
->constant
->values
[0].u32
[j
] = u32
[comp
];
1468 case SpvOpCompositeExtract
:
1469 case SpvOpCompositeInsert
: {
1470 struct vtn_value
*comp
;
1471 unsigned deref_start
;
1472 struct nir_constant
**c
;
1473 if (opcode
== SpvOpCompositeExtract
) {
1474 comp
= vtn_value(b
, w
[4], vtn_value_type_constant
);
1476 c
= &comp
->constant
;
1478 comp
= vtn_value(b
, w
[5], vtn_value_type_constant
);
1480 val
->constant
= nir_constant_clone(comp
->constant
,
1487 const struct glsl_type
*type
= comp
->type
->type
;
1488 for (unsigned i
= deref_start
; i
< count
; i
++) {
1489 switch (glsl_get_base_type(type
)) {
1490 case GLSL_TYPE_UINT
:
1492 case GLSL_TYPE_UINT16
:
1493 case GLSL_TYPE_INT16
:
1494 case GLSL_TYPE_UINT64
:
1495 case GLSL_TYPE_INT64
:
1496 case GLSL_TYPE_FLOAT
:
1497 case GLSL_TYPE_FLOAT16
:
1498 case GLSL_TYPE_DOUBLE
:
1499 case GLSL_TYPE_BOOL
:
1500 /* If we hit this granularity, we're picking off an element */
1501 if (glsl_type_is_matrix(type
)) {
1502 vtn_assert(col
== 0 && elem
== -1);
1505 type
= glsl_get_column_type(type
);
1507 vtn_assert(elem
<= 0 && glsl_type_is_vector(type
));
1509 type
= glsl_scalar_type(glsl_get_base_type(type
));
1513 case GLSL_TYPE_ARRAY
:
1514 c
= &(*c
)->elements
[w
[i
]];
1515 type
= glsl_get_array_element(type
);
1518 case GLSL_TYPE_STRUCT
:
1519 c
= &(*c
)->elements
[w
[i
]];
1520 type
= glsl_get_struct_field(type
, w
[i
]);
1524 vtn_fail("Invalid constant type");
1528 if (opcode
== SpvOpCompositeExtract
) {
1532 unsigned num_components
= glsl_get_vector_elements(type
);
1533 unsigned bit_size
= glsl_get_bit_size(type
);
1534 for (unsigned i
= 0; i
< num_components
; i
++)
1537 val
->constant
->values
[0].u64
[i
] = (*c
)->values
[col
].u64
[elem
+ i
];
1540 val
->constant
->values
[0].u32
[i
] = (*c
)->values
[col
].u32
[elem
+ i
];
1543 val
->constant
->values
[0].u16
[i
] = (*c
)->values
[col
].u16
[elem
+ i
];
1546 vtn_fail("Invalid SpvOpCompositeExtract bit size");
1550 struct vtn_value
*insert
=
1551 vtn_value(b
, w
[4], vtn_value_type_constant
);
1552 vtn_assert(insert
->type
->type
== type
);
1554 *c
= insert
->constant
;
1556 unsigned num_components
= glsl_get_vector_elements(type
);
1557 unsigned bit_size
= glsl_get_bit_size(type
);
1558 for (unsigned i
= 0; i
< num_components
; i
++)
1561 (*c
)->values
[col
].u64
[elem
+ i
] = insert
->constant
->values
[0].u64
[i
];
1564 (*c
)->values
[col
].u32
[elem
+ i
] = insert
->constant
->values
[0].u32
[i
];
1567 (*c
)->values
[col
].u16
[elem
+ i
] = insert
->constant
->values
[0].u16
[i
];
1570 vtn_fail("Invalid SpvOpCompositeInsert bit size");
1579 nir_alu_type dst_alu_type
= nir_get_nir_type_for_glsl_type(val
->type
->type
);
1580 nir_alu_type src_alu_type
= dst_alu_type
;
1581 unsigned num_components
= glsl_get_vector_elements(val
->type
->type
);
1584 vtn_assert(count
<= 7);
1589 /* We have a source in a conversion */
1591 nir_get_nir_type_for_glsl_type(
1592 vtn_value(b
, w
[4], vtn_value_type_constant
)->type
->type
);
1593 /* We use the bitsize of the conversion source to evaluate the opcode later */
1594 bit_size
= glsl_get_bit_size(
1595 vtn_value(b
, w
[4], vtn_value_type_constant
)->type
->type
);
1598 bit_size
= glsl_get_bit_size(val
->type
->type
);
1601 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(b
, opcode
, &swap
,
1604 nir_const_value src
[4];
1606 for (unsigned i
= 0; i
< count
- 4; i
++) {
1608 vtn_value(b
, w
[4 + i
], vtn_value_type_constant
)->constant
;
1610 unsigned j
= swap
? 1 - i
: i
;
1611 src
[j
] = c
->values
[0];
1614 val
->constant
->values
[0] =
1615 nir_eval_const_opcode(op
, num_components
, bit_size
, src
);
1622 case SpvOpConstantNull
:
1623 val
->constant
= vtn_null_constant(b
, val
->type
->type
);
1626 case SpvOpConstantSampler
:
1627 vtn_fail("OpConstantSampler requires Kernel Capability");
1631 vtn_fail("Unhandled opcode");
1634 /* Now that we have the value, update the workgroup size if needed */
1635 vtn_foreach_decoration(b
, val
, handle_workgroup_size_decoration_cb
, NULL
);
1639 vtn_handle_function_call(struct vtn_builder
*b
, SpvOp opcode
,
1640 const uint32_t *w
, unsigned count
)
1642 struct vtn_type
*res_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1643 struct vtn_function
*vtn_callee
=
1644 vtn_value(b
, w
[3], vtn_value_type_function
)->func
;
1645 struct nir_function
*callee
= vtn_callee
->impl
->function
;
1647 vtn_callee
->referenced
= true;
1649 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, callee
);
1650 for (unsigned i
= 0; i
< call
->num_params
; i
++) {
1651 unsigned arg_id
= w
[4 + i
];
1652 struct vtn_value
*arg
= vtn_untyped_value(b
, arg_id
);
1653 if (arg
->value_type
== vtn_value_type_pointer
&&
1654 arg
->pointer
->ptr_type
->type
== NULL
) {
1655 nir_deref_var
*d
= vtn_pointer_to_deref(b
, arg
->pointer
);
1656 call
->params
[i
] = nir_deref_var_clone(d
, call
);
1658 struct vtn_ssa_value
*arg_ssa
= vtn_ssa_value(b
, arg_id
);
1660 /* Make a temporary to store the argument in */
1662 nir_local_variable_create(b
->nb
.impl
, arg_ssa
->type
, "arg_tmp");
1663 call
->params
[i
] = nir_deref_var_create(call
, tmp
);
1665 vtn_local_store(b
, arg_ssa
, call
->params
[i
]);
1669 nir_variable
*out_tmp
= NULL
;
1670 vtn_assert(res_type
->type
== callee
->return_type
);
1671 if (!glsl_type_is_void(callee
->return_type
)) {
1672 out_tmp
= nir_local_variable_create(b
->nb
.impl
, callee
->return_type
,
1674 call
->return_deref
= nir_deref_var_create(call
, out_tmp
);
1677 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
1679 if (glsl_type_is_void(callee
->return_type
)) {
1680 vtn_push_value(b
, w
[2], vtn_value_type_undef
);
1682 vtn_push_ssa(b
, w
[2], res_type
, vtn_local_load(b
, call
->return_deref
));
1686 struct vtn_ssa_value
*
1687 vtn_create_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
1689 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
1692 if (!glsl_type_is_vector_or_scalar(type
)) {
1693 unsigned elems
= glsl_get_length(type
);
1694 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
1695 for (unsigned i
= 0; i
< elems
; i
++) {
1696 const struct glsl_type
*child_type
;
1698 switch (glsl_get_base_type(type
)) {
1700 case GLSL_TYPE_UINT
:
1701 case GLSL_TYPE_INT16
:
1702 case GLSL_TYPE_UINT16
:
1703 case GLSL_TYPE_INT64
:
1704 case GLSL_TYPE_UINT64
:
1705 case GLSL_TYPE_BOOL
:
1706 case GLSL_TYPE_FLOAT
:
1707 case GLSL_TYPE_FLOAT16
:
1708 case GLSL_TYPE_DOUBLE
:
1709 child_type
= glsl_get_column_type(type
);
1711 case GLSL_TYPE_ARRAY
:
1712 child_type
= glsl_get_array_element(type
);
1714 case GLSL_TYPE_STRUCT
:
1715 child_type
= glsl_get_struct_field(type
, i
);
1718 vtn_fail("unkown base type");
1721 val
->elems
[i
] = vtn_create_ssa_value(b
, child_type
);
1729 vtn_tex_src(struct vtn_builder
*b
, unsigned index
, nir_tex_src_type type
)
1732 src
.src
= nir_src_for_ssa(vtn_ssa_value(b
, index
)->def
);
1733 src
.src_type
= type
;
1738 vtn_handle_texture(struct vtn_builder
*b
, SpvOp opcode
,
1739 const uint32_t *w
, unsigned count
)
1741 if (opcode
== SpvOpSampledImage
) {
1742 struct vtn_value
*val
=
1743 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
1744 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
1745 val
->sampled_image
->type
=
1746 vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1747 val
->sampled_image
->image
=
1748 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
1749 val
->sampled_image
->sampler
=
1750 vtn_value(b
, w
[4], vtn_value_type_pointer
)->pointer
;
1752 } else if (opcode
== SpvOpImage
) {
1753 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
1754 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[3]);
1755 if (src_val
->value_type
== vtn_value_type_sampled_image
) {
1756 val
->pointer
= src_val
->sampled_image
->image
;
1758 vtn_assert(src_val
->value_type
== vtn_value_type_pointer
);
1759 val
->pointer
= src_val
->pointer
;
1764 struct vtn_type
*ret_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1765 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1767 struct vtn_sampled_image sampled
;
1768 struct vtn_value
*sampled_val
= vtn_untyped_value(b
, w
[3]);
1769 if (sampled_val
->value_type
== vtn_value_type_sampled_image
) {
1770 sampled
= *sampled_val
->sampled_image
;
1772 vtn_assert(sampled_val
->value_type
== vtn_value_type_pointer
);
1773 sampled
.type
= sampled_val
->pointer
->type
;
1774 sampled
.image
= NULL
;
1775 sampled
.sampler
= sampled_val
->pointer
;
1778 const struct glsl_type
*image_type
= sampled
.type
->type
;
1779 const enum glsl_sampler_dim sampler_dim
= glsl_get_sampler_dim(image_type
);
1780 const bool is_array
= glsl_sampler_type_is_array(image_type
);
1781 const bool is_shadow
= glsl_sampler_type_is_shadow(image_type
);
1783 /* Figure out the base texture operation */
1786 case SpvOpImageSampleImplicitLod
:
1787 case SpvOpImageSampleDrefImplicitLod
:
1788 case SpvOpImageSampleProjImplicitLod
:
1789 case SpvOpImageSampleProjDrefImplicitLod
:
1790 texop
= nir_texop_tex
;
1793 case SpvOpImageSampleExplicitLod
:
1794 case SpvOpImageSampleDrefExplicitLod
:
1795 case SpvOpImageSampleProjExplicitLod
:
1796 case SpvOpImageSampleProjDrefExplicitLod
:
1797 texop
= nir_texop_txl
;
1800 case SpvOpImageFetch
:
1801 if (glsl_get_sampler_dim(image_type
) == GLSL_SAMPLER_DIM_MS
) {
1802 texop
= nir_texop_txf_ms
;
1804 texop
= nir_texop_txf
;
1808 case SpvOpImageGather
:
1809 case SpvOpImageDrefGather
:
1810 texop
= nir_texop_tg4
;
1813 case SpvOpImageQuerySizeLod
:
1814 case SpvOpImageQuerySize
:
1815 texop
= nir_texop_txs
;
1818 case SpvOpImageQueryLod
:
1819 texop
= nir_texop_lod
;
1822 case SpvOpImageQueryLevels
:
1823 texop
= nir_texop_query_levels
;
1826 case SpvOpImageQuerySamples
:
1827 texop
= nir_texop_texture_samples
;
1831 vtn_fail("Unhandled opcode");
1834 nir_tex_src srcs
[8]; /* 8 should be enough */
1835 nir_tex_src
*p
= srcs
;
1839 struct nir_ssa_def
*coord
;
1840 unsigned coord_components
;
1842 case SpvOpImageSampleImplicitLod
:
1843 case SpvOpImageSampleExplicitLod
:
1844 case SpvOpImageSampleDrefImplicitLod
:
1845 case SpvOpImageSampleDrefExplicitLod
:
1846 case SpvOpImageSampleProjImplicitLod
:
1847 case SpvOpImageSampleProjExplicitLod
:
1848 case SpvOpImageSampleProjDrefImplicitLod
:
1849 case SpvOpImageSampleProjDrefExplicitLod
:
1850 case SpvOpImageFetch
:
1851 case SpvOpImageGather
:
1852 case SpvOpImageDrefGather
:
1853 case SpvOpImageQueryLod
: {
1854 /* All these types have the coordinate as their first real argument */
1855 switch (sampler_dim
) {
1856 case GLSL_SAMPLER_DIM_1D
:
1857 case GLSL_SAMPLER_DIM_BUF
:
1858 coord_components
= 1;
1860 case GLSL_SAMPLER_DIM_2D
:
1861 case GLSL_SAMPLER_DIM_RECT
:
1862 case GLSL_SAMPLER_DIM_MS
:
1863 coord_components
= 2;
1865 case GLSL_SAMPLER_DIM_3D
:
1866 case GLSL_SAMPLER_DIM_CUBE
:
1867 coord_components
= 3;
1870 vtn_fail("Invalid sampler type");
1873 if (is_array
&& texop
!= nir_texop_lod
)
1876 coord
= vtn_ssa_value(b
, w
[idx
++])->def
;
1877 p
->src
= nir_src_for_ssa(nir_channels(&b
->nb
, coord
,
1878 (1 << coord_components
) - 1));
1879 p
->src_type
= nir_tex_src_coord
;
1886 coord_components
= 0;
1891 case SpvOpImageSampleProjImplicitLod
:
1892 case SpvOpImageSampleProjExplicitLod
:
1893 case SpvOpImageSampleProjDrefImplicitLod
:
1894 case SpvOpImageSampleProjDrefExplicitLod
:
1895 /* These have the projector as the last coordinate component */
1896 p
->src
= nir_src_for_ssa(nir_channel(&b
->nb
, coord
, coord_components
));
1897 p
->src_type
= nir_tex_src_projector
;
1905 unsigned gather_component
= 0;
1907 case SpvOpImageSampleDrefImplicitLod
:
1908 case SpvOpImageSampleDrefExplicitLod
:
1909 case SpvOpImageSampleProjDrefImplicitLod
:
1910 case SpvOpImageSampleProjDrefExplicitLod
:
1911 case SpvOpImageDrefGather
:
1912 /* These all have an explicit depth value as their next source */
1913 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_comparator
);
1916 case SpvOpImageGather
:
1917 /* This has a component as its next source */
1919 vtn_value(b
, w
[idx
++], vtn_value_type_constant
)->constant
->values
[0].u32
[0];
1926 /* For OpImageQuerySizeLod, we always have an LOD */
1927 if (opcode
== SpvOpImageQuerySizeLod
)
1928 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
1930 /* Now we need to handle some number of optional arguments */
1931 const struct vtn_ssa_value
*gather_offsets
= NULL
;
1933 uint32_t operands
= w
[idx
++];
1935 if (operands
& SpvImageOperandsBiasMask
) {
1936 vtn_assert(texop
== nir_texop_tex
);
1937 texop
= nir_texop_txb
;
1938 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_bias
);
1941 if (operands
& SpvImageOperandsLodMask
) {
1942 vtn_assert(texop
== nir_texop_txl
|| texop
== nir_texop_txf
||
1943 texop
== nir_texop_txs
);
1944 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
1947 if (operands
& SpvImageOperandsGradMask
) {
1948 vtn_assert(texop
== nir_texop_txl
);
1949 texop
= nir_texop_txd
;
1950 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddx
);
1951 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddy
);
1954 if (operands
& SpvImageOperandsOffsetMask
||
1955 operands
& SpvImageOperandsConstOffsetMask
)
1956 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_offset
);
1958 if (operands
& SpvImageOperandsConstOffsetsMask
) {
1959 gather_offsets
= vtn_ssa_value(b
, w
[idx
++]);
1960 (*p
++) = (nir_tex_src
){};
1963 if (operands
& SpvImageOperandsSampleMask
) {
1964 vtn_assert(texop
== nir_texop_txf_ms
);
1965 texop
= nir_texop_txf_ms
;
1966 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ms_index
);
1969 /* We should have now consumed exactly all of the arguments */
1970 vtn_assert(idx
== count
);
1972 nir_tex_instr
*instr
= nir_tex_instr_create(b
->shader
, p
- srcs
);
1975 memcpy(instr
->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
1977 instr
->coord_components
= coord_components
;
1978 instr
->sampler_dim
= sampler_dim
;
1979 instr
->is_array
= is_array
;
1980 instr
->is_shadow
= is_shadow
;
1981 instr
->is_new_style_shadow
=
1982 is_shadow
&& glsl_get_components(ret_type
->type
) == 1;
1983 instr
->component
= gather_component
;
1985 switch (glsl_get_sampler_result_type(image_type
)) {
1986 case GLSL_TYPE_FLOAT
: instr
->dest_type
= nir_type_float
; break;
1987 case GLSL_TYPE_INT
: instr
->dest_type
= nir_type_int
; break;
1988 case GLSL_TYPE_UINT
: instr
->dest_type
= nir_type_uint
; break;
1989 case GLSL_TYPE_BOOL
: instr
->dest_type
= nir_type_bool
; break;
1991 vtn_fail("Invalid base type for sampler result");
1994 nir_deref_var
*sampler
= vtn_pointer_to_deref(b
, sampled
.sampler
);
1995 nir_deref_var
*texture
;
1996 if (sampled
.image
) {
1997 nir_deref_var
*image
= vtn_pointer_to_deref(b
, sampled
.image
);
2003 instr
->texture
= nir_deref_var_clone(texture
, instr
);
2005 switch (instr
->op
) {
2011 /* These operations require a sampler */
2012 instr
->sampler
= nir_deref_var_clone(sampler
, instr
);
2015 case nir_texop_txf_ms
:
2018 case nir_texop_query_levels
:
2019 case nir_texop_texture_samples
:
2020 case nir_texop_samples_identical
:
2022 instr
->sampler
= NULL
;
2024 case nir_texop_txf_ms_mcs
:
2025 vtn_fail("unexpected nir_texop_txf_ms_mcs");
2028 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
2029 nir_tex_instr_dest_size(instr
), 32, NULL
);
2031 vtn_assert(glsl_get_vector_elements(ret_type
->type
) ==
2032 nir_tex_instr_dest_size(instr
));
2035 nir_instr
*instruction
;
2036 if (gather_offsets
) {
2037 vtn_assert(glsl_get_base_type(gather_offsets
->type
) == GLSL_TYPE_ARRAY
);
2038 vtn_assert(glsl_get_length(gather_offsets
->type
) == 4);
2039 nir_tex_instr
*instrs
[4] = {instr
, NULL
, NULL
, NULL
};
2041 /* Copy the current instruction 4x */
2042 for (uint32_t i
= 1; i
< 4; i
++) {
2043 instrs
[i
] = nir_tex_instr_create(b
->shader
, instr
->num_srcs
);
2044 instrs
[i
]->op
= instr
->op
;
2045 instrs
[i
]->coord_components
= instr
->coord_components
;
2046 instrs
[i
]->sampler_dim
= instr
->sampler_dim
;
2047 instrs
[i
]->is_array
= instr
->is_array
;
2048 instrs
[i
]->is_shadow
= instr
->is_shadow
;
2049 instrs
[i
]->is_new_style_shadow
= instr
->is_new_style_shadow
;
2050 instrs
[i
]->component
= instr
->component
;
2051 instrs
[i
]->dest_type
= instr
->dest_type
;
2052 instrs
[i
]->texture
= nir_deref_var_clone(texture
, instrs
[i
]);
2053 instrs
[i
]->sampler
= NULL
;
2055 memcpy(instrs
[i
]->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
2057 nir_ssa_dest_init(&instrs
[i
]->instr
, &instrs
[i
]->dest
,
2058 nir_tex_instr_dest_size(instr
), 32, NULL
);
2061 /* Fill in the last argument with the offset from the passed in offsets
2062 * and insert the instruction into the stream.
2064 for (uint32_t i
= 0; i
< 4; i
++) {
2066 src
.src
= nir_src_for_ssa(gather_offsets
->elems
[i
]->def
);
2067 src
.src_type
= nir_tex_src_offset
;
2068 instrs
[i
]->src
[instrs
[i
]->num_srcs
- 1] = src
;
2069 nir_builder_instr_insert(&b
->nb
, &instrs
[i
]->instr
);
2072 /* Combine the results of the 4 instructions by taking their .w
2075 nir_alu_instr
*vec4
= nir_alu_instr_create(b
->shader
, nir_op_vec4
);
2076 nir_ssa_dest_init(&vec4
->instr
, &vec4
->dest
.dest
, 4, 32, NULL
);
2077 vec4
->dest
.write_mask
= 0xf;
2078 for (uint32_t i
= 0; i
< 4; i
++) {
2079 vec4
->src
[i
].src
= nir_src_for_ssa(&instrs
[i
]->dest
.ssa
);
2080 vec4
->src
[i
].swizzle
[0] = 3;
2082 def
= &vec4
->dest
.dest
.ssa
;
2083 instruction
= &vec4
->instr
;
2085 def
= &instr
->dest
.ssa
;
2086 instruction
= &instr
->instr
;
2089 val
->ssa
= vtn_create_ssa_value(b
, ret_type
->type
);
2090 val
->ssa
->def
= def
;
2092 nir_builder_instr_insert(&b
->nb
, instruction
);
2096 fill_common_atomic_sources(struct vtn_builder
*b
, SpvOp opcode
,
2097 const uint32_t *w
, nir_src
*src
)
2100 case SpvOpAtomicIIncrement
:
2101 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
2104 case SpvOpAtomicIDecrement
:
2105 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
2108 case SpvOpAtomicISub
:
2110 nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
2113 case SpvOpAtomicCompareExchange
:
2114 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[8])->def
);
2115 src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
2118 case SpvOpAtomicExchange
:
2119 case SpvOpAtomicIAdd
:
2120 case SpvOpAtomicSMin
:
2121 case SpvOpAtomicUMin
:
2122 case SpvOpAtomicSMax
:
2123 case SpvOpAtomicUMax
:
2124 case SpvOpAtomicAnd
:
2126 case SpvOpAtomicXor
:
2127 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
2131 vtn_fail("Invalid SPIR-V atomic");
2135 static nir_ssa_def
*
2136 get_image_coord(struct vtn_builder
*b
, uint32_t value
)
2138 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, value
);
2140 /* The image_load_store intrinsics assume a 4-dim coordinate */
2141 unsigned dim
= glsl_get_vector_elements(coord
->type
);
2142 unsigned swizzle
[4];
2143 for (unsigned i
= 0; i
< 4; i
++)
2144 swizzle
[i
] = MIN2(i
, dim
- 1);
2146 return nir_swizzle(&b
->nb
, coord
->def
, swizzle
, 4, false);
2150 vtn_handle_image(struct vtn_builder
*b
, SpvOp opcode
,
2151 const uint32_t *w
, unsigned count
)
2153 /* Just get this one out of the way */
2154 if (opcode
== SpvOpImageTexelPointer
) {
2155 struct vtn_value
*val
=
2156 vtn_push_value(b
, w
[2], vtn_value_type_image_pointer
);
2157 val
->image
= ralloc(b
, struct vtn_image_pointer
);
2159 val
->image
->image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2160 val
->image
->coord
= get_image_coord(b
, w
[4]);
2161 val
->image
->sample
= vtn_ssa_value(b
, w
[5])->def
;
2165 struct vtn_image_pointer image
;
2168 case SpvOpAtomicExchange
:
2169 case SpvOpAtomicCompareExchange
:
2170 case SpvOpAtomicCompareExchangeWeak
:
2171 case SpvOpAtomicIIncrement
:
2172 case SpvOpAtomicIDecrement
:
2173 case SpvOpAtomicIAdd
:
2174 case SpvOpAtomicISub
:
2175 case SpvOpAtomicLoad
:
2176 case SpvOpAtomicSMin
:
2177 case SpvOpAtomicUMin
:
2178 case SpvOpAtomicSMax
:
2179 case SpvOpAtomicUMax
:
2180 case SpvOpAtomicAnd
:
2182 case SpvOpAtomicXor
:
2183 image
= *vtn_value(b
, w
[3], vtn_value_type_image_pointer
)->image
;
2186 case SpvOpAtomicStore
:
2187 image
= *vtn_value(b
, w
[1], vtn_value_type_image_pointer
)->image
;
2190 case SpvOpImageQuerySize
:
2191 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2193 image
.sample
= NULL
;
2196 case SpvOpImageRead
:
2197 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2198 image
.coord
= get_image_coord(b
, w
[4]);
2200 if (count
> 5 && (w
[5] & SpvImageOperandsSampleMask
)) {
2201 vtn_assert(w
[5] == SpvImageOperandsSampleMask
);
2202 image
.sample
= vtn_ssa_value(b
, w
[6])->def
;
2204 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2208 case SpvOpImageWrite
:
2209 image
.image
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
2210 image
.coord
= get_image_coord(b
, w
[2]);
2214 if (count
> 4 && (w
[4] & SpvImageOperandsSampleMask
)) {
2215 vtn_assert(w
[4] == SpvImageOperandsSampleMask
);
2216 image
.sample
= vtn_ssa_value(b
, w
[5])->def
;
2218 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2223 vtn_fail("Invalid image opcode");
2226 nir_intrinsic_op op
;
2228 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_##N; break;
2229 OP(ImageQuerySize
, size
)
2231 OP(ImageWrite
, store
)
2232 OP(AtomicLoad
, load
)
2233 OP(AtomicStore
, store
)
2234 OP(AtomicExchange
, atomic_exchange
)
2235 OP(AtomicCompareExchange
, atomic_comp_swap
)
2236 OP(AtomicIIncrement
, atomic_add
)
2237 OP(AtomicIDecrement
, atomic_add
)
2238 OP(AtomicIAdd
, atomic_add
)
2239 OP(AtomicISub
, atomic_add
)
2240 OP(AtomicSMin
, atomic_min
)
2241 OP(AtomicUMin
, atomic_min
)
2242 OP(AtomicSMax
, atomic_max
)
2243 OP(AtomicUMax
, atomic_max
)
2244 OP(AtomicAnd
, atomic_and
)
2245 OP(AtomicOr
, atomic_or
)
2246 OP(AtomicXor
, atomic_xor
)
2249 vtn_fail("Invalid image opcode");
2252 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
2254 nir_deref_var
*image_deref
= vtn_pointer_to_deref(b
, image
.image
);
2255 intrin
->variables
[0] = nir_deref_var_clone(image_deref
, intrin
);
2257 /* ImageQuerySize doesn't take any extra parameters */
2258 if (opcode
!= SpvOpImageQuerySize
) {
2259 /* The image coordinate is always 4 components but we may not have that
2260 * many. Swizzle to compensate.
2263 for (unsigned i
= 0; i
< 4; i
++)
2264 swiz
[i
] = i
< image
.coord
->num_components
? i
: 0;
2265 intrin
->src
[0] = nir_src_for_ssa(nir_swizzle(&b
->nb
, image
.coord
,
2267 intrin
->src
[1] = nir_src_for_ssa(image
.sample
);
2271 case SpvOpAtomicLoad
:
2272 case SpvOpImageQuerySize
:
2273 case SpvOpImageRead
:
2275 case SpvOpAtomicStore
:
2276 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2278 case SpvOpImageWrite
:
2279 intrin
->src
[2] = nir_src_for_ssa(vtn_ssa_value(b
, w
[3])->def
);
2282 case SpvOpAtomicCompareExchange
:
2283 case SpvOpAtomicIIncrement
:
2284 case SpvOpAtomicIDecrement
:
2285 case SpvOpAtomicExchange
:
2286 case SpvOpAtomicIAdd
:
2287 case SpvOpAtomicISub
:
2288 case SpvOpAtomicSMin
:
2289 case SpvOpAtomicUMin
:
2290 case SpvOpAtomicSMax
:
2291 case SpvOpAtomicUMax
:
2292 case SpvOpAtomicAnd
:
2294 case SpvOpAtomicXor
:
2295 fill_common_atomic_sources(b
, opcode
, w
, &intrin
->src
[2]);
2299 vtn_fail("Invalid image opcode");
2302 if (opcode
!= SpvOpImageWrite
) {
2303 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2304 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2306 unsigned dest_components
=
2307 nir_intrinsic_infos
[intrin
->intrinsic
].dest_components
;
2308 if (intrin
->intrinsic
== nir_intrinsic_image_size
) {
2309 dest_components
= intrin
->num_components
=
2310 glsl_get_vector_elements(type
->type
);
2313 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
,
2314 dest_components
, 32, NULL
);
2316 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2318 val
->ssa
= vtn_create_ssa_value(b
, type
->type
);
2319 val
->ssa
->def
= &intrin
->dest
.ssa
;
2321 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2325 static nir_intrinsic_op
2326 get_ssbo_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2329 case SpvOpAtomicLoad
: return nir_intrinsic_load_ssbo
;
2330 case SpvOpAtomicStore
: return nir_intrinsic_store_ssbo
;
2331 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
2332 OP(AtomicExchange
, atomic_exchange
)
2333 OP(AtomicCompareExchange
, atomic_comp_swap
)
2334 OP(AtomicIIncrement
, atomic_add
)
2335 OP(AtomicIDecrement
, atomic_add
)
2336 OP(AtomicIAdd
, atomic_add
)
2337 OP(AtomicISub
, atomic_add
)
2338 OP(AtomicSMin
, atomic_imin
)
2339 OP(AtomicUMin
, atomic_umin
)
2340 OP(AtomicSMax
, atomic_imax
)
2341 OP(AtomicUMax
, atomic_umax
)
2342 OP(AtomicAnd
, atomic_and
)
2343 OP(AtomicOr
, atomic_or
)
2344 OP(AtomicXor
, atomic_xor
)
2347 vtn_fail("Invalid SSBO atomic");
2351 static nir_intrinsic_op
2352 get_shared_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2355 case SpvOpAtomicLoad
: return nir_intrinsic_load_shared
;
2356 case SpvOpAtomicStore
: return nir_intrinsic_store_shared
;
2357 #define OP(S, N) case SpvOp##S: return nir_intrinsic_shared_##N;
2358 OP(AtomicExchange
, atomic_exchange
)
2359 OP(AtomicCompareExchange
, atomic_comp_swap
)
2360 OP(AtomicIIncrement
, atomic_add
)
2361 OP(AtomicIDecrement
, atomic_add
)
2362 OP(AtomicIAdd
, atomic_add
)
2363 OP(AtomicISub
, atomic_add
)
2364 OP(AtomicSMin
, atomic_imin
)
2365 OP(AtomicUMin
, atomic_umin
)
2366 OP(AtomicSMax
, atomic_imax
)
2367 OP(AtomicUMax
, atomic_umax
)
2368 OP(AtomicAnd
, atomic_and
)
2369 OP(AtomicOr
, atomic_or
)
2370 OP(AtomicXor
, atomic_xor
)
2373 vtn_fail("Invalid shared atomic");
2377 static nir_intrinsic_op
2378 get_var_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2381 case SpvOpAtomicLoad
: return nir_intrinsic_load_var
;
2382 case SpvOpAtomicStore
: return nir_intrinsic_store_var
;
2383 #define OP(S, N) case SpvOp##S: return nir_intrinsic_var_##N;
2384 OP(AtomicExchange
, atomic_exchange
)
2385 OP(AtomicCompareExchange
, atomic_comp_swap
)
2386 OP(AtomicIIncrement
, atomic_add
)
2387 OP(AtomicIDecrement
, atomic_add
)
2388 OP(AtomicIAdd
, atomic_add
)
2389 OP(AtomicISub
, atomic_add
)
2390 OP(AtomicSMin
, atomic_imin
)
2391 OP(AtomicUMin
, atomic_umin
)
2392 OP(AtomicSMax
, atomic_imax
)
2393 OP(AtomicUMax
, atomic_umax
)
2394 OP(AtomicAnd
, atomic_and
)
2395 OP(AtomicOr
, atomic_or
)
2396 OP(AtomicXor
, atomic_xor
)
2399 vtn_fail("Invalid shared atomic");
2404 vtn_handle_ssbo_or_shared_atomic(struct vtn_builder
*b
, SpvOp opcode
,
2405 const uint32_t *w
, unsigned count
)
2407 struct vtn_pointer
*ptr
;
2408 nir_intrinsic_instr
*atomic
;
2411 case SpvOpAtomicLoad
:
2412 case SpvOpAtomicExchange
:
2413 case SpvOpAtomicCompareExchange
:
2414 case SpvOpAtomicCompareExchangeWeak
:
2415 case SpvOpAtomicIIncrement
:
2416 case SpvOpAtomicIDecrement
:
2417 case SpvOpAtomicIAdd
:
2418 case SpvOpAtomicISub
:
2419 case SpvOpAtomicSMin
:
2420 case SpvOpAtomicUMin
:
2421 case SpvOpAtomicSMax
:
2422 case SpvOpAtomicUMax
:
2423 case SpvOpAtomicAnd
:
2425 case SpvOpAtomicXor
:
2426 ptr
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2429 case SpvOpAtomicStore
:
2430 ptr
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
2434 vtn_fail("Invalid SPIR-V atomic");
2438 SpvScope scope = w[4];
2439 SpvMemorySemanticsMask semantics = w[5];
2442 if (ptr
->mode
== vtn_variable_mode_workgroup
&&
2443 !b
->options
->lower_workgroup_access_to_offsets
) {
2444 nir_deref_var
*deref
= vtn_pointer_to_deref(b
, ptr
);
2445 const struct glsl_type
*deref_type
= nir_deref_tail(&deref
->deref
)->type
;
2446 nir_intrinsic_op op
= get_var_nir_atomic_op(b
, opcode
);
2447 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2448 atomic
->variables
[0] = nir_deref_var_clone(deref
, atomic
);
2451 case SpvOpAtomicLoad
:
2452 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2455 case SpvOpAtomicStore
:
2456 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2457 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2458 atomic
->src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2461 case SpvOpAtomicExchange
:
2462 case SpvOpAtomicCompareExchange
:
2463 case SpvOpAtomicCompareExchangeWeak
:
2464 case SpvOpAtomicIIncrement
:
2465 case SpvOpAtomicIDecrement
:
2466 case SpvOpAtomicIAdd
:
2467 case SpvOpAtomicISub
:
2468 case SpvOpAtomicSMin
:
2469 case SpvOpAtomicUMin
:
2470 case SpvOpAtomicSMax
:
2471 case SpvOpAtomicUMax
:
2472 case SpvOpAtomicAnd
:
2474 case SpvOpAtomicXor
:
2475 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[0]);
2479 vtn_fail("Invalid SPIR-V atomic");
2483 nir_ssa_def
*offset
, *index
;
2484 offset
= vtn_pointer_to_offset(b
, ptr
, &index
, NULL
);
2486 nir_intrinsic_op op
;
2487 if (ptr
->mode
== vtn_variable_mode_ssbo
) {
2488 op
= get_ssbo_nir_atomic_op(b
, opcode
);
2490 vtn_assert(ptr
->mode
== vtn_variable_mode_workgroup
&&
2491 b
->options
->lower_workgroup_access_to_offsets
);
2492 op
= get_shared_nir_atomic_op(b
, opcode
);
2495 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2499 case SpvOpAtomicLoad
:
2500 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
2501 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2502 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2503 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2506 case SpvOpAtomicStore
:
2507 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
2508 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2509 atomic
->src
[src
++] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2510 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2511 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2512 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2515 case SpvOpAtomicExchange
:
2516 case SpvOpAtomicCompareExchange
:
2517 case SpvOpAtomicCompareExchangeWeak
:
2518 case SpvOpAtomicIIncrement
:
2519 case SpvOpAtomicIDecrement
:
2520 case SpvOpAtomicIAdd
:
2521 case SpvOpAtomicISub
:
2522 case SpvOpAtomicSMin
:
2523 case SpvOpAtomicUMin
:
2524 case SpvOpAtomicSMax
:
2525 case SpvOpAtomicUMax
:
2526 case SpvOpAtomicAnd
:
2528 case SpvOpAtomicXor
:
2529 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2530 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2531 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2532 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[src
]);
2536 vtn_fail("Invalid SPIR-V atomic");
2540 if (opcode
!= SpvOpAtomicStore
) {
2541 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2543 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
,
2544 glsl_get_vector_elements(type
->type
),
2545 glsl_get_bit_size(type
->type
), NULL
);
2547 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2548 val
->ssa
= rzalloc(b
, struct vtn_ssa_value
);
2549 val
->ssa
->def
= &atomic
->dest
.ssa
;
2550 val
->ssa
->type
= type
->type
;
2553 nir_builder_instr_insert(&b
->nb
, &atomic
->instr
);
2556 static nir_alu_instr
*
2557 create_vec(struct vtn_builder
*b
, unsigned num_components
, unsigned bit_size
)
2560 switch (num_components
) {
2561 case 1: op
= nir_op_fmov
; break;
2562 case 2: op
= nir_op_vec2
; break;
2563 case 3: op
= nir_op_vec3
; break;
2564 case 4: op
= nir_op_vec4
; break;
2565 default: vtn_fail("bad vector size");
2568 nir_alu_instr
*vec
= nir_alu_instr_create(b
->shader
, op
);
2569 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, num_components
,
2571 vec
->dest
.write_mask
= (1 << num_components
) - 1;
2576 struct vtn_ssa_value
*
2577 vtn_ssa_transpose(struct vtn_builder
*b
, struct vtn_ssa_value
*src
)
2579 if (src
->transposed
)
2580 return src
->transposed
;
2582 struct vtn_ssa_value
*dest
=
2583 vtn_create_ssa_value(b
, glsl_transposed_type(src
->type
));
2585 for (unsigned i
= 0; i
< glsl_get_matrix_columns(dest
->type
); i
++) {
2586 nir_alu_instr
*vec
= create_vec(b
, glsl_get_matrix_columns(src
->type
),
2587 glsl_get_bit_size(src
->type
));
2588 if (glsl_type_is_vector_or_scalar(src
->type
)) {
2589 vec
->src
[0].src
= nir_src_for_ssa(src
->def
);
2590 vec
->src
[0].swizzle
[0] = i
;
2592 for (unsigned j
= 0; j
< glsl_get_matrix_columns(src
->type
); j
++) {
2593 vec
->src
[j
].src
= nir_src_for_ssa(src
->elems
[j
]->def
);
2594 vec
->src
[j
].swizzle
[0] = i
;
2597 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2598 dest
->elems
[i
]->def
= &vec
->dest
.dest
.ssa
;
2601 dest
->transposed
= src
;
2607 vtn_vector_extract(struct vtn_builder
*b
, nir_ssa_def
*src
, unsigned index
)
2609 unsigned swiz
[4] = { index
};
2610 return nir_swizzle(&b
->nb
, src
, swiz
, 1, true);
2614 vtn_vector_insert(struct vtn_builder
*b
, nir_ssa_def
*src
, nir_ssa_def
*insert
,
2617 nir_alu_instr
*vec
= create_vec(b
, src
->num_components
,
2620 for (unsigned i
= 0; i
< src
->num_components
; i
++) {
2622 vec
->src
[i
].src
= nir_src_for_ssa(insert
);
2624 vec
->src
[i
].src
= nir_src_for_ssa(src
);
2625 vec
->src
[i
].swizzle
[0] = i
;
2629 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2631 return &vec
->dest
.dest
.ssa
;
2635 vtn_vector_extract_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
2638 nir_ssa_def
*dest
= vtn_vector_extract(b
, src
, 0);
2639 for (unsigned i
= 1; i
< src
->num_components
; i
++)
2640 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
2641 vtn_vector_extract(b
, src
, i
), dest
);
2647 vtn_vector_insert_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
2648 nir_ssa_def
*insert
, nir_ssa_def
*index
)
2650 nir_ssa_def
*dest
= vtn_vector_insert(b
, src
, insert
, 0);
2651 for (unsigned i
= 1; i
< src
->num_components
; i
++)
2652 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
2653 vtn_vector_insert(b
, src
, insert
, i
), dest
);
2658 static nir_ssa_def
*
2659 vtn_vector_shuffle(struct vtn_builder
*b
, unsigned num_components
,
2660 nir_ssa_def
*src0
, nir_ssa_def
*src1
,
2661 const uint32_t *indices
)
2663 nir_alu_instr
*vec
= create_vec(b
, num_components
, src0
->bit_size
);
2665 for (unsigned i
= 0; i
< num_components
; i
++) {
2666 uint32_t index
= indices
[i
];
2667 if (index
== 0xffffffff) {
2669 nir_src_for_ssa(nir_ssa_undef(&b
->nb
, 1, src0
->bit_size
));
2670 } else if (index
< src0
->num_components
) {
2671 vec
->src
[i
].src
= nir_src_for_ssa(src0
);
2672 vec
->src
[i
].swizzle
[0] = index
;
2674 vec
->src
[i
].src
= nir_src_for_ssa(src1
);
2675 vec
->src
[i
].swizzle
[0] = index
- src0
->num_components
;
2679 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2681 return &vec
->dest
.dest
.ssa
;
2685 * Concatentates a number of vectors/scalars together to produce a vector
2687 static nir_ssa_def
*
2688 vtn_vector_construct(struct vtn_builder
*b
, unsigned num_components
,
2689 unsigned num_srcs
, nir_ssa_def
**srcs
)
2691 nir_alu_instr
*vec
= create_vec(b
, num_components
, srcs
[0]->bit_size
);
2693 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2695 * "When constructing a vector, there must be at least two Constituent
2698 vtn_assert(num_srcs
>= 2);
2700 unsigned dest_idx
= 0;
2701 for (unsigned i
= 0; i
< num_srcs
; i
++) {
2702 nir_ssa_def
*src
= srcs
[i
];
2703 vtn_assert(dest_idx
+ src
->num_components
<= num_components
);
2704 for (unsigned j
= 0; j
< src
->num_components
; j
++) {
2705 vec
->src
[dest_idx
].src
= nir_src_for_ssa(src
);
2706 vec
->src
[dest_idx
].swizzle
[0] = j
;
2711 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2713 * "When constructing a vector, the total number of components in all
2714 * the operands must equal the number of components in Result Type."
2716 vtn_assert(dest_idx
== num_components
);
2718 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2720 return &vec
->dest
.dest
.ssa
;
2723 static struct vtn_ssa_value
*
2724 vtn_composite_copy(void *mem_ctx
, struct vtn_ssa_value
*src
)
2726 struct vtn_ssa_value
*dest
= rzalloc(mem_ctx
, struct vtn_ssa_value
);
2727 dest
->type
= src
->type
;
2729 if (glsl_type_is_vector_or_scalar(src
->type
)) {
2730 dest
->def
= src
->def
;
2732 unsigned elems
= glsl_get_length(src
->type
);
2734 dest
->elems
= ralloc_array(mem_ctx
, struct vtn_ssa_value
*, elems
);
2735 for (unsigned i
= 0; i
< elems
; i
++)
2736 dest
->elems
[i
] = vtn_composite_copy(mem_ctx
, src
->elems
[i
]);
2742 static struct vtn_ssa_value
*
2743 vtn_composite_insert(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
2744 struct vtn_ssa_value
*insert
, const uint32_t *indices
,
2745 unsigned num_indices
)
2747 struct vtn_ssa_value
*dest
= vtn_composite_copy(b
, src
);
2749 struct vtn_ssa_value
*cur
= dest
;
2751 for (i
= 0; i
< num_indices
- 1; i
++) {
2752 cur
= cur
->elems
[indices
[i
]];
2755 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
2756 /* According to the SPIR-V spec, OpCompositeInsert may work down to
2757 * the component granularity. In that case, the last index will be
2758 * the index to insert the scalar into the vector.
2761 cur
->def
= vtn_vector_insert(b
, cur
->def
, insert
->def
, indices
[i
]);
2763 cur
->elems
[indices
[i
]] = insert
;
2769 static struct vtn_ssa_value
*
2770 vtn_composite_extract(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
2771 const uint32_t *indices
, unsigned num_indices
)
2773 struct vtn_ssa_value
*cur
= src
;
2774 for (unsigned i
= 0; i
< num_indices
; i
++) {
2775 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
2776 vtn_assert(i
== num_indices
- 1);
2777 /* According to the SPIR-V spec, OpCompositeExtract may work down to
2778 * the component granularity. The last index will be the index of the
2779 * vector to extract.
2782 struct vtn_ssa_value
*ret
= rzalloc(b
, struct vtn_ssa_value
);
2783 ret
->type
= glsl_scalar_type(glsl_get_base_type(cur
->type
));
2784 ret
->def
= vtn_vector_extract(b
, cur
->def
, indices
[i
]);
2787 cur
= cur
->elems
[indices
[i
]];
2795 vtn_handle_composite(struct vtn_builder
*b
, SpvOp opcode
,
2796 const uint32_t *w
, unsigned count
)
2798 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2799 const struct glsl_type
*type
=
2800 vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
2801 val
->ssa
= vtn_create_ssa_value(b
, type
);
2804 case SpvOpVectorExtractDynamic
:
2805 val
->ssa
->def
= vtn_vector_extract_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
2806 vtn_ssa_value(b
, w
[4])->def
);
2809 case SpvOpVectorInsertDynamic
:
2810 val
->ssa
->def
= vtn_vector_insert_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
2811 vtn_ssa_value(b
, w
[4])->def
,
2812 vtn_ssa_value(b
, w
[5])->def
);
2815 case SpvOpVectorShuffle
:
2816 val
->ssa
->def
= vtn_vector_shuffle(b
, glsl_get_vector_elements(type
),
2817 vtn_ssa_value(b
, w
[3])->def
,
2818 vtn_ssa_value(b
, w
[4])->def
,
2822 case SpvOpCompositeConstruct
: {
2823 unsigned elems
= count
- 3;
2824 if (glsl_type_is_vector_or_scalar(type
)) {
2825 nir_ssa_def
*srcs
[4];
2826 for (unsigned i
= 0; i
< elems
; i
++)
2827 srcs
[i
] = vtn_ssa_value(b
, w
[3 + i
])->def
;
2829 vtn_vector_construct(b
, glsl_get_vector_elements(type
),
2832 val
->ssa
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
2833 for (unsigned i
= 0; i
< elems
; i
++)
2834 val
->ssa
->elems
[i
] = vtn_ssa_value(b
, w
[3 + i
]);
2838 case SpvOpCompositeExtract
:
2839 val
->ssa
= vtn_composite_extract(b
, vtn_ssa_value(b
, w
[3]),
2843 case SpvOpCompositeInsert
:
2844 val
->ssa
= vtn_composite_insert(b
, vtn_ssa_value(b
, w
[4]),
2845 vtn_ssa_value(b
, w
[3]),
2849 case SpvOpCopyObject
:
2850 val
->ssa
= vtn_composite_copy(b
, vtn_ssa_value(b
, w
[3]));
2854 vtn_fail("unknown composite operation");
2859 vtn_handle_barrier(struct vtn_builder
*b
, SpvOp opcode
,
2860 const uint32_t *w
, unsigned count
)
2862 nir_intrinsic_op intrinsic_op
;
2864 case SpvOpEmitVertex
:
2865 case SpvOpEmitStreamVertex
:
2866 intrinsic_op
= nir_intrinsic_emit_vertex
;
2868 case SpvOpEndPrimitive
:
2869 case SpvOpEndStreamPrimitive
:
2870 intrinsic_op
= nir_intrinsic_end_primitive
;
2872 case SpvOpMemoryBarrier
:
2873 intrinsic_op
= nir_intrinsic_memory_barrier
;
2875 case SpvOpControlBarrier
:
2876 intrinsic_op
= nir_intrinsic_barrier
;
2879 vtn_fail("unknown barrier instruction");
2882 nir_intrinsic_instr
*intrin
=
2883 nir_intrinsic_instr_create(b
->shader
, intrinsic_op
);
2885 if (opcode
== SpvOpEmitStreamVertex
|| opcode
== SpvOpEndStreamPrimitive
)
2886 nir_intrinsic_set_stream_id(intrin
, w
[1]);
2888 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2892 gl_primitive_from_spv_execution_mode(struct vtn_builder
*b
,
2893 SpvExecutionMode mode
)
2896 case SpvExecutionModeInputPoints
:
2897 case SpvExecutionModeOutputPoints
:
2898 return 0; /* GL_POINTS */
2899 case SpvExecutionModeInputLines
:
2900 return 1; /* GL_LINES */
2901 case SpvExecutionModeInputLinesAdjacency
:
2902 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
2903 case SpvExecutionModeTriangles
:
2904 return 4; /* GL_TRIANGLES */
2905 case SpvExecutionModeInputTrianglesAdjacency
:
2906 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
2907 case SpvExecutionModeQuads
:
2908 return 7; /* GL_QUADS */
2909 case SpvExecutionModeIsolines
:
2910 return 0x8E7A; /* GL_ISOLINES */
2911 case SpvExecutionModeOutputLineStrip
:
2912 return 3; /* GL_LINE_STRIP */
2913 case SpvExecutionModeOutputTriangleStrip
:
2914 return 5; /* GL_TRIANGLE_STRIP */
2916 vtn_fail("Invalid primitive type");
2921 vertices_in_from_spv_execution_mode(struct vtn_builder
*b
,
2922 SpvExecutionMode mode
)
2925 case SpvExecutionModeInputPoints
:
2927 case SpvExecutionModeInputLines
:
2929 case SpvExecutionModeInputLinesAdjacency
:
2931 case SpvExecutionModeTriangles
:
2933 case SpvExecutionModeInputTrianglesAdjacency
:
2936 vtn_fail("Invalid GS input mode");
2940 static gl_shader_stage
2941 stage_for_execution_model(struct vtn_builder
*b
, SpvExecutionModel model
)
2944 case SpvExecutionModelVertex
:
2945 return MESA_SHADER_VERTEX
;
2946 case SpvExecutionModelTessellationControl
:
2947 return MESA_SHADER_TESS_CTRL
;
2948 case SpvExecutionModelTessellationEvaluation
:
2949 return MESA_SHADER_TESS_EVAL
;
2950 case SpvExecutionModelGeometry
:
2951 return MESA_SHADER_GEOMETRY
;
2952 case SpvExecutionModelFragment
:
2953 return MESA_SHADER_FRAGMENT
;
2954 case SpvExecutionModelGLCompute
:
2955 return MESA_SHADER_COMPUTE
;
2957 vtn_fail("Unsupported execution model");
2961 #define spv_check_supported(name, cap) do { \
2962 if (!(b->options && b->options->caps.name)) \
2963 vtn_warn("Unsupported SPIR-V capability: %s", \
2964 spirv_capability_to_string(cap)); \
2968 vtn_handle_preamble_instruction(struct vtn_builder
*b
, SpvOp opcode
,
2969 const uint32_t *w
, unsigned count
)
2976 case SpvSourceLanguageUnknown
: lang
= "unknown"; break;
2977 case SpvSourceLanguageESSL
: lang
= "ESSL"; break;
2978 case SpvSourceLanguageGLSL
: lang
= "GLSL"; break;
2979 case SpvSourceLanguageOpenCL_C
: lang
= "OpenCL C"; break;
2980 case SpvSourceLanguageOpenCL_CPP
: lang
= "OpenCL C++"; break;
2981 case SpvSourceLanguageHLSL
: lang
= "HLSL"; break;
2984 uint32_t version
= w
[2];
2987 (count
> 3) ? vtn_value(b
, w
[3], vtn_value_type_string
)->str
: "";
2989 vtn_info("Parsing SPIR-V from %s %u source file %s", lang
, version
, file
);
2993 case SpvOpSourceExtension
:
2994 case SpvOpSourceContinued
:
2995 case SpvOpExtension
:
2996 /* Unhandled, but these are for debug so that's ok. */
2999 case SpvOpCapability
: {
3000 SpvCapability cap
= w
[1];
3002 case SpvCapabilityMatrix
:
3003 case SpvCapabilityShader
:
3004 case SpvCapabilityGeometry
:
3005 case SpvCapabilityGeometryPointSize
:
3006 case SpvCapabilityUniformBufferArrayDynamicIndexing
:
3007 case SpvCapabilitySampledImageArrayDynamicIndexing
:
3008 case SpvCapabilityStorageBufferArrayDynamicIndexing
:
3009 case SpvCapabilityStorageImageArrayDynamicIndexing
:
3010 case SpvCapabilityImageRect
:
3011 case SpvCapabilitySampledRect
:
3012 case SpvCapabilitySampled1D
:
3013 case SpvCapabilityImage1D
:
3014 case SpvCapabilitySampledCubeArray
:
3015 case SpvCapabilityImageCubeArray
:
3016 case SpvCapabilitySampledBuffer
:
3017 case SpvCapabilityImageBuffer
:
3018 case SpvCapabilityImageQuery
:
3019 case SpvCapabilityDerivativeControl
:
3020 case SpvCapabilityInterpolationFunction
:
3021 case SpvCapabilityMultiViewport
:
3022 case SpvCapabilitySampleRateShading
:
3023 case SpvCapabilityClipDistance
:
3024 case SpvCapabilityCullDistance
:
3025 case SpvCapabilityInputAttachment
:
3026 case SpvCapabilityImageGatherExtended
:
3027 case SpvCapabilityStorageImageExtendedFormats
:
3030 case SpvCapabilityGeometryStreams
:
3031 case SpvCapabilityLinkage
:
3032 case SpvCapabilityVector16
:
3033 case SpvCapabilityFloat16Buffer
:
3034 case SpvCapabilityFloat16
:
3035 case SpvCapabilityInt64Atomics
:
3036 case SpvCapabilityAtomicStorage
:
3037 case SpvCapabilityInt16
:
3038 case SpvCapabilityStorageImageMultisample
:
3039 case SpvCapabilityInt8
:
3040 case SpvCapabilitySparseResidency
:
3041 case SpvCapabilityMinLod
:
3042 case SpvCapabilityTransformFeedback
:
3043 vtn_warn("Unsupported SPIR-V capability: %s",
3044 spirv_capability_to_string(cap
));
3047 case SpvCapabilityFloat64
:
3048 spv_check_supported(float64
, cap
);
3050 case SpvCapabilityInt64
:
3051 spv_check_supported(int64
, cap
);
3054 case SpvCapabilityAddresses
:
3055 case SpvCapabilityKernel
:
3056 case SpvCapabilityImageBasic
:
3057 case SpvCapabilityImageReadWrite
:
3058 case SpvCapabilityImageMipmap
:
3059 case SpvCapabilityPipes
:
3060 case SpvCapabilityGroups
:
3061 case SpvCapabilityDeviceEnqueue
:
3062 case SpvCapabilityLiteralSampler
:
3063 case SpvCapabilityGenericPointer
:
3064 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
3065 spirv_capability_to_string(cap
));
3068 case SpvCapabilityImageMSArray
:
3069 spv_check_supported(image_ms_array
, cap
);
3072 case SpvCapabilityTessellation
:
3073 case SpvCapabilityTessellationPointSize
:
3074 spv_check_supported(tessellation
, cap
);
3077 case SpvCapabilityDrawParameters
:
3078 spv_check_supported(draw_parameters
, cap
);
3081 case SpvCapabilityStorageImageReadWithoutFormat
:
3082 spv_check_supported(image_read_without_format
, cap
);
3085 case SpvCapabilityStorageImageWriteWithoutFormat
:
3086 spv_check_supported(image_write_without_format
, cap
);
3089 case SpvCapabilityMultiView
:
3090 spv_check_supported(multiview
, cap
);
3093 case SpvCapabilityVariablePointersStorageBuffer
:
3094 case SpvCapabilityVariablePointers
:
3095 spv_check_supported(variable_pointers
, cap
);
3098 case SpvCapabilityStorageUniformBufferBlock16
:
3099 case SpvCapabilityStorageUniform16
:
3100 case SpvCapabilityStoragePushConstant16
:
3101 case SpvCapabilityStorageInputOutput16
:
3102 spv_check_supported(storage_16bit
, cap
);
3106 vtn_fail("Unhandled capability");
3111 case SpvOpExtInstImport
:
3112 vtn_handle_extension(b
, opcode
, w
, count
);
3115 case SpvOpMemoryModel
:
3116 vtn_assert(w
[1] == SpvAddressingModelLogical
);
3117 vtn_assert(w
[2] == SpvMemoryModelSimple
||
3118 w
[2] == SpvMemoryModelGLSL450
);
3121 case SpvOpEntryPoint
: {
3122 struct vtn_value
*entry_point
= &b
->values
[w
[2]];
3123 /* Let this be a name label regardless */
3124 unsigned name_words
;
3125 entry_point
->name
= vtn_string_literal(b
, &w
[3], count
- 3, &name_words
);
3127 if (strcmp(entry_point
->name
, b
->entry_point_name
) != 0 ||
3128 stage_for_execution_model(b
, w
[1]) != b
->entry_point_stage
)
3131 vtn_assert(b
->entry_point
== NULL
);
3132 b
->entry_point
= entry_point
;
3137 vtn_push_value(b
, w
[1], vtn_value_type_string
)->str
=
3138 vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
3142 b
->values
[w
[1]].name
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
3145 case SpvOpMemberName
:
3149 case SpvOpExecutionMode
:
3150 case SpvOpDecorationGroup
:
3152 case SpvOpMemberDecorate
:
3153 case SpvOpGroupDecorate
:
3154 case SpvOpGroupMemberDecorate
:
3155 vtn_handle_decoration(b
, opcode
, w
, count
);
3159 return false; /* End of preamble */
3166 vtn_handle_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
3167 const struct vtn_decoration
*mode
, void *data
)
3169 vtn_assert(b
->entry_point
== entry_point
);
3171 switch(mode
->exec_mode
) {
3172 case SpvExecutionModeOriginUpperLeft
:
3173 case SpvExecutionModeOriginLowerLeft
:
3174 b
->origin_upper_left
=
3175 (mode
->exec_mode
== SpvExecutionModeOriginUpperLeft
);
3178 case SpvExecutionModeEarlyFragmentTests
:
3179 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3180 b
->shader
->info
.fs
.early_fragment_tests
= true;
3183 case SpvExecutionModeInvocations
:
3184 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3185 b
->shader
->info
.gs
.invocations
= MAX2(1, mode
->literals
[0]);
3188 case SpvExecutionModeDepthReplacing
:
3189 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3190 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_ANY
;
3192 case SpvExecutionModeDepthGreater
:
3193 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3194 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_GREATER
;
3196 case SpvExecutionModeDepthLess
:
3197 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3198 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_LESS
;
3200 case SpvExecutionModeDepthUnchanged
:
3201 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3202 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
3205 case SpvExecutionModeLocalSize
:
3206 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
3207 b
->shader
->info
.cs
.local_size
[0] = mode
->literals
[0];
3208 b
->shader
->info
.cs
.local_size
[1] = mode
->literals
[1];
3209 b
->shader
->info
.cs
.local_size
[2] = mode
->literals
[2];
3211 case SpvExecutionModeLocalSizeHint
:
3212 break; /* Nothing to do with this */
3214 case SpvExecutionModeOutputVertices
:
3215 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3216 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
3217 b
->shader
->info
.tess
.tcs_vertices_out
= mode
->literals
[0];
3219 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3220 b
->shader
->info
.gs
.vertices_out
= mode
->literals
[0];
3224 case SpvExecutionModeInputPoints
:
3225 case SpvExecutionModeInputLines
:
3226 case SpvExecutionModeInputLinesAdjacency
:
3227 case SpvExecutionModeTriangles
:
3228 case SpvExecutionModeInputTrianglesAdjacency
:
3229 case SpvExecutionModeQuads
:
3230 case SpvExecutionModeIsolines
:
3231 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3232 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
3233 b
->shader
->info
.tess
.primitive_mode
=
3234 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
3236 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3237 b
->shader
->info
.gs
.vertices_in
=
3238 vertices_in_from_spv_execution_mode(b
, mode
->exec_mode
);
3242 case SpvExecutionModeOutputPoints
:
3243 case SpvExecutionModeOutputLineStrip
:
3244 case SpvExecutionModeOutputTriangleStrip
:
3245 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3246 b
->shader
->info
.gs
.output_primitive
=
3247 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
3250 case SpvExecutionModeSpacingEqual
:
3251 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3252 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3253 b
->shader
->info
.tess
.spacing
= TESS_SPACING_EQUAL
;
3255 case SpvExecutionModeSpacingFractionalEven
:
3256 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3257 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3258 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_EVEN
;
3260 case SpvExecutionModeSpacingFractionalOdd
:
3261 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3262 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3263 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_ODD
;
3265 case SpvExecutionModeVertexOrderCw
:
3266 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3267 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3268 b
->shader
->info
.tess
.ccw
= false;
3270 case SpvExecutionModeVertexOrderCcw
:
3271 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3272 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3273 b
->shader
->info
.tess
.ccw
= true;
3275 case SpvExecutionModePointMode
:
3276 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3277 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3278 b
->shader
->info
.tess
.point_mode
= true;
3281 case SpvExecutionModePixelCenterInteger
:
3282 b
->pixel_center_integer
= true;
3285 case SpvExecutionModeXfb
:
3286 vtn_fail("Unhandled execution mode");
3289 case SpvExecutionModeVecTypeHint
:
3290 case SpvExecutionModeContractionOff
:
3294 vtn_fail("Unhandled execution mode");
3299 vtn_handle_variable_or_type_instruction(struct vtn_builder
*b
, SpvOp opcode
,
3300 const uint32_t *w
, unsigned count
)
3302 vtn_set_instruction_result_type(b
, opcode
, w
, count
);
3306 case SpvOpSourceContinued
:
3307 case SpvOpSourceExtension
:
3308 case SpvOpExtension
:
3309 case SpvOpCapability
:
3310 case SpvOpExtInstImport
:
3311 case SpvOpMemoryModel
:
3312 case SpvOpEntryPoint
:
3313 case SpvOpExecutionMode
:
3316 case SpvOpMemberName
:
3317 case SpvOpDecorationGroup
:
3319 case SpvOpMemberDecorate
:
3320 case SpvOpGroupDecorate
:
3321 case SpvOpGroupMemberDecorate
:
3322 vtn_fail("Invalid opcode types and variables section");
3328 case SpvOpTypeFloat
:
3329 case SpvOpTypeVector
:
3330 case SpvOpTypeMatrix
:
3331 case SpvOpTypeImage
:
3332 case SpvOpTypeSampler
:
3333 case SpvOpTypeSampledImage
:
3334 case SpvOpTypeArray
:
3335 case SpvOpTypeRuntimeArray
:
3336 case SpvOpTypeStruct
:
3337 case SpvOpTypeOpaque
:
3338 case SpvOpTypePointer
:
3339 case SpvOpTypeFunction
:
3340 case SpvOpTypeEvent
:
3341 case SpvOpTypeDeviceEvent
:
3342 case SpvOpTypeReserveId
:
3343 case SpvOpTypeQueue
:
3345 vtn_handle_type(b
, opcode
, w
, count
);
3348 case SpvOpConstantTrue
:
3349 case SpvOpConstantFalse
:
3351 case SpvOpConstantComposite
:
3352 case SpvOpConstantSampler
:
3353 case SpvOpConstantNull
:
3354 case SpvOpSpecConstantTrue
:
3355 case SpvOpSpecConstantFalse
:
3356 case SpvOpSpecConstant
:
3357 case SpvOpSpecConstantComposite
:
3358 case SpvOpSpecConstantOp
:
3359 vtn_handle_constant(b
, opcode
, w
, count
);
3364 vtn_handle_variables(b
, opcode
, w
, count
);
3368 return false; /* End of preamble */
3375 vtn_handle_body_instruction(struct vtn_builder
*b
, SpvOp opcode
,
3376 const uint32_t *w
, unsigned count
)
3382 case SpvOpLoopMerge
:
3383 case SpvOpSelectionMerge
:
3384 /* This is handled by cfg pre-pass and walk_blocks */
3388 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
3389 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
3394 vtn_handle_extension(b
, opcode
, w
, count
);
3400 case SpvOpCopyMemory
:
3401 case SpvOpCopyMemorySized
:
3402 case SpvOpAccessChain
:
3403 case SpvOpPtrAccessChain
:
3404 case SpvOpInBoundsAccessChain
:
3405 case SpvOpArrayLength
:
3406 vtn_handle_variables(b
, opcode
, w
, count
);
3409 case SpvOpFunctionCall
:
3410 vtn_handle_function_call(b
, opcode
, w
, count
);
3413 case SpvOpSampledImage
:
3415 case SpvOpImageSampleImplicitLod
:
3416 case SpvOpImageSampleExplicitLod
:
3417 case SpvOpImageSampleDrefImplicitLod
:
3418 case SpvOpImageSampleDrefExplicitLod
:
3419 case SpvOpImageSampleProjImplicitLod
:
3420 case SpvOpImageSampleProjExplicitLod
:
3421 case SpvOpImageSampleProjDrefImplicitLod
:
3422 case SpvOpImageSampleProjDrefExplicitLod
:
3423 case SpvOpImageFetch
:
3424 case SpvOpImageGather
:
3425 case SpvOpImageDrefGather
:
3426 case SpvOpImageQuerySizeLod
:
3427 case SpvOpImageQueryLod
:
3428 case SpvOpImageQueryLevels
:
3429 case SpvOpImageQuerySamples
:
3430 vtn_handle_texture(b
, opcode
, w
, count
);
3433 case SpvOpImageRead
:
3434 case SpvOpImageWrite
:
3435 case SpvOpImageTexelPointer
:
3436 vtn_handle_image(b
, opcode
, w
, count
);
3439 case SpvOpImageQuerySize
: {
3440 struct vtn_pointer
*image
=
3441 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
3442 if (image
->mode
== vtn_variable_mode_image
) {
3443 vtn_handle_image(b
, opcode
, w
, count
);
3445 vtn_assert(image
->mode
== vtn_variable_mode_sampler
);
3446 vtn_handle_texture(b
, opcode
, w
, count
);
3451 case SpvOpAtomicLoad
:
3452 case SpvOpAtomicExchange
:
3453 case SpvOpAtomicCompareExchange
:
3454 case SpvOpAtomicCompareExchangeWeak
:
3455 case SpvOpAtomicIIncrement
:
3456 case SpvOpAtomicIDecrement
:
3457 case SpvOpAtomicIAdd
:
3458 case SpvOpAtomicISub
:
3459 case SpvOpAtomicSMin
:
3460 case SpvOpAtomicUMin
:
3461 case SpvOpAtomicSMax
:
3462 case SpvOpAtomicUMax
:
3463 case SpvOpAtomicAnd
:
3465 case SpvOpAtomicXor
: {
3466 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[3]);
3467 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
3468 vtn_handle_image(b
, opcode
, w
, count
);
3470 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
3471 vtn_handle_ssbo_or_shared_atomic(b
, opcode
, w
, count
);
3476 case SpvOpAtomicStore
: {
3477 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[1]);
3478 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
3479 vtn_handle_image(b
, opcode
, w
, count
);
3481 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
3482 vtn_handle_ssbo_or_shared_atomic(b
, opcode
, w
, count
);
3488 /* Handle OpSelect up-front here because it needs to be able to handle
3489 * pointers and not just regular vectors and scalars.
3491 struct vtn_value
*res_val
= vtn_untyped_value(b
, w
[2]);
3492 struct vtn_value
*sel_val
= vtn_untyped_value(b
, w
[3]);
3493 struct vtn_value
*obj1_val
= vtn_untyped_value(b
, w
[4]);
3494 struct vtn_value
*obj2_val
= vtn_untyped_value(b
, w
[5]);
3496 const struct glsl_type
*sel_type
;
3497 switch (res_val
->type
->base_type
) {
3498 case vtn_base_type_scalar
:
3499 sel_type
= glsl_bool_type();
3501 case vtn_base_type_vector
:
3502 sel_type
= glsl_vector_type(GLSL_TYPE_BOOL
, res_val
->type
->length
);
3504 case vtn_base_type_pointer
:
3505 /* We need to have actual storage for pointer types */
3506 vtn_fail_if(res_val
->type
->type
== NULL
,
3507 "Invalid pointer result type for OpSelect");
3508 sel_type
= glsl_bool_type();
3511 vtn_fail("Result type of OpSelect must be a scalar, vector, or pointer");
3514 if (unlikely(sel_val
->type
->type
!= sel_type
)) {
3515 if (sel_val
->type
->type
== glsl_bool_type()) {
3516 /* This case is illegal but some older versions of GLSLang produce
3517 * it. The GLSLang issue was fixed on March 30, 2017:
3519 * https://github.com/KhronosGroup/glslang/issues/809
3521 * Unfortunately, there are applications in the wild which are
3522 * shipping with this bug so it isn't nice to fail on them so we
3523 * throw a warning instead. It's not actually a problem for us as
3524 * nir_builder will just splat the condition out which is most
3525 * likely what the client wanted anyway.
3527 vtn_warn("Condition type of OpSelect must have the same number "
3528 "of components as Result Type");
3530 vtn_fail("Condition type of OpSelect must be a scalar or vector "
3531 "of Boolean type. It must have the same number of "
3532 "components as Result Type");
3536 vtn_fail_if(obj1_val
->type
!= res_val
->type
||
3537 obj2_val
->type
!= res_val
->type
,
3538 "Object types must match the result type in OpSelect");
3540 struct vtn_type
*res_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
3541 struct vtn_ssa_value
*ssa
= vtn_create_ssa_value(b
, res_type
->type
);
3542 ssa
->def
= nir_bcsel(&b
->nb
, vtn_ssa_value(b
, w
[3])->def
,
3543 vtn_ssa_value(b
, w
[4])->def
,
3544 vtn_ssa_value(b
, w
[5])->def
);
3545 vtn_push_ssa(b
, w
[2], res_type
, ssa
);
3554 case SpvOpConvertFToU
:
3555 case SpvOpConvertFToS
:
3556 case SpvOpConvertSToF
:
3557 case SpvOpConvertUToF
:
3561 case SpvOpQuantizeToF16
:
3562 case SpvOpConvertPtrToU
:
3563 case SpvOpConvertUToPtr
:
3564 case SpvOpPtrCastToGeneric
:
3565 case SpvOpGenericCastToPtr
:
3571 case SpvOpSignBitSet
:
3572 case SpvOpLessOrGreater
:
3574 case SpvOpUnordered
:
3589 case SpvOpVectorTimesScalar
:
3591 case SpvOpIAddCarry
:
3592 case SpvOpISubBorrow
:
3593 case SpvOpUMulExtended
:
3594 case SpvOpSMulExtended
:
3595 case SpvOpShiftRightLogical
:
3596 case SpvOpShiftRightArithmetic
:
3597 case SpvOpShiftLeftLogical
:
3598 case SpvOpLogicalEqual
:
3599 case SpvOpLogicalNotEqual
:
3600 case SpvOpLogicalOr
:
3601 case SpvOpLogicalAnd
:
3602 case SpvOpLogicalNot
:
3603 case SpvOpBitwiseOr
:
3604 case SpvOpBitwiseXor
:
3605 case SpvOpBitwiseAnd
:
3607 case SpvOpFOrdEqual
:
3608 case SpvOpFUnordEqual
:
3609 case SpvOpINotEqual
:
3610 case SpvOpFOrdNotEqual
:
3611 case SpvOpFUnordNotEqual
:
3612 case SpvOpULessThan
:
3613 case SpvOpSLessThan
:
3614 case SpvOpFOrdLessThan
:
3615 case SpvOpFUnordLessThan
:
3616 case SpvOpUGreaterThan
:
3617 case SpvOpSGreaterThan
:
3618 case SpvOpFOrdGreaterThan
:
3619 case SpvOpFUnordGreaterThan
:
3620 case SpvOpULessThanEqual
:
3621 case SpvOpSLessThanEqual
:
3622 case SpvOpFOrdLessThanEqual
:
3623 case SpvOpFUnordLessThanEqual
:
3624 case SpvOpUGreaterThanEqual
:
3625 case SpvOpSGreaterThanEqual
:
3626 case SpvOpFOrdGreaterThanEqual
:
3627 case SpvOpFUnordGreaterThanEqual
:
3633 case SpvOpFwidthFine
:
3634 case SpvOpDPdxCoarse
:
3635 case SpvOpDPdyCoarse
:
3636 case SpvOpFwidthCoarse
:
3637 case SpvOpBitFieldInsert
:
3638 case SpvOpBitFieldSExtract
:
3639 case SpvOpBitFieldUExtract
:
3640 case SpvOpBitReverse
:
3642 case SpvOpTranspose
:
3643 case SpvOpOuterProduct
:
3644 case SpvOpMatrixTimesScalar
:
3645 case SpvOpVectorTimesMatrix
:
3646 case SpvOpMatrixTimesVector
:
3647 case SpvOpMatrixTimesMatrix
:
3648 vtn_handle_alu(b
, opcode
, w
, count
);
3651 case SpvOpVectorExtractDynamic
:
3652 case SpvOpVectorInsertDynamic
:
3653 case SpvOpVectorShuffle
:
3654 case SpvOpCompositeConstruct
:
3655 case SpvOpCompositeExtract
:
3656 case SpvOpCompositeInsert
:
3657 case SpvOpCopyObject
:
3658 vtn_handle_composite(b
, opcode
, w
, count
);
3661 case SpvOpEmitVertex
:
3662 case SpvOpEndPrimitive
:
3663 case SpvOpEmitStreamVertex
:
3664 case SpvOpEndStreamPrimitive
:
3665 case SpvOpControlBarrier
:
3666 case SpvOpMemoryBarrier
:
3667 vtn_handle_barrier(b
, opcode
, w
, count
);
3671 vtn_fail("Unhandled opcode");
3678 spirv_to_nir(const uint32_t *words
, size_t word_count
,
3679 struct nir_spirv_specialization
*spec
, unsigned num_spec
,
3680 gl_shader_stage stage
, const char *entry_point_name
,
3681 const struct spirv_to_nir_options
*options
,
3682 const nir_shader_compiler_options
*nir_options
)
3684 /* Initialize the stn_builder object */
3685 struct vtn_builder
*b
= rzalloc(NULL
, struct vtn_builder
);
3690 exec_list_make_empty(&b
->functions
);
3691 b
->entry_point_stage
= stage
;
3692 b
->entry_point_name
= entry_point_name
;
3693 b
->options
= options
;
3695 /* See also _vtn_fail() */
3696 if (setjmp(b
->fail_jump
)) {
3701 const uint32_t *word_end
= words
+ word_count
;
3703 /* Handle the SPIR-V header (first 4 dwords) */
3704 vtn_assert(word_count
> 5);
3706 vtn_assert(words
[0] == SpvMagicNumber
);
3707 vtn_assert(words
[1] >= 0x10000);
3708 /* words[2] == generator magic */
3709 unsigned value_id_bound
= words
[3];
3710 vtn_assert(words
[4] == 0);
3714 b
->value_id_bound
= value_id_bound
;
3715 b
->values
= rzalloc_array(b
, struct vtn_value
, value_id_bound
);
3717 /* Handle all the preamble instructions */
3718 words
= vtn_foreach_instruction(b
, words
, word_end
,
3719 vtn_handle_preamble_instruction
);
3721 if (b
->entry_point
== NULL
) {
3722 vtn_fail("Entry point not found");
3727 b
->shader
= nir_shader_create(b
, stage
, nir_options
, NULL
);
3729 /* Set shader info defaults */
3730 b
->shader
->info
.gs
.invocations
= 1;
3732 /* Parse execution modes */
3733 vtn_foreach_execution_mode(b
, b
->entry_point
,
3734 vtn_handle_execution_mode
, NULL
);
3736 b
->specializations
= spec
;
3737 b
->num_specializations
= num_spec
;
3739 /* Handle all variable, type, and constant instructions */
3740 words
= vtn_foreach_instruction(b
, words
, word_end
,
3741 vtn_handle_variable_or_type_instruction
);
3743 /* Set types on all vtn_values */
3744 vtn_foreach_instruction(b
, words
, word_end
, vtn_set_instruction_result_type
);
3746 vtn_build_cfg(b
, words
, word_end
);
3748 assert(b
->entry_point
->value_type
== vtn_value_type_function
);
3749 b
->entry_point
->func
->referenced
= true;
3754 foreach_list_typed(struct vtn_function
, func
, node
, &b
->functions
) {
3755 if (func
->referenced
&& !func
->emitted
) {
3756 b
->const_table
= _mesa_hash_table_create(b
, _mesa_hash_pointer
,
3757 _mesa_key_pointer_equal
);
3759 vtn_function_emit(b
, func
, vtn_handle_body_instruction
);
3765 vtn_assert(b
->entry_point
->value_type
== vtn_value_type_function
);
3766 nir_function
*entry_point
= b
->entry_point
->func
->impl
->function
;
3767 vtn_assert(entry_point
);
3769 /* Unparent the shader from the vtn_builder before we delete the builder */
3770 ralloc_steal(NULL
, b
->shader
);