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 "nir/nir_deref.h"
33 #include "spirv_info.h"
38 vtn_log(struct vtn_builder
*b
, enum nir_spirv_debug_level level
,
39 size_t spirv_offset
, const char *message
)
41 if (b
->options
->debug
.func
) {
42 b
->options
->debug
.func(b
->options
->debug
.private_data
,
43 level
, spirv_offset
, message
);
47 if (level
>= NIR_SPIRV_DEBUG_LEVEL_WARNING
)
48 fprintf(stderr
, "%s\n", message
);
53 vtn_logf(struct vtn_builder
*b
, enum nir_spirv_debug_level level
,
54 size_t spirv_offset
, const char *fmt
, ...)
60 msg
= ralloc_vasprintf(NULL
, fmt
, args
);
63 vtn_log(b
, level
, spirv_offset
, msg
);
69 vtn_log_err(struct vtn_builder
*b
,
70 enum nir_spirv_debug_level level
, const char *prefix
,
71 const char *file
, unsigned line
,
72 const char *fmt
, va_list args
)
76 msg
= ralloc_strdup(NULL
, prefix
);
79 ralloc_asprintf_append(&msg
, " In file %s:%u\n", file
, line
);
82 ralloc_asprintf_append(&msg
, " ");
84 ralloc_vasprintf_append(&msg
, fmt
, args
);
86 ralloc_asprintf_append(&msg
, "\n %zu bytes into the SPIR-V binary",
90 ralloc_asprintf_append(&msg
,
91 "\n in SPIR-V source file %s, line %d, col %d",
92 b
->file
, b
->line
, b
->col
);
95 vtn_log(b
, level
, b
->spirv_offset
, msg
);
101 vtn_dump_shader(struct vtn_builder
*b
, const char *path
, const char *prefix
)
106 int len
= snprintf(filename
, sizeof(filename
), "%s/%s-%d.spirv",
107 path
, prefix
, idx
++);
108 if (len
< 0 || len
>= sizeof(filename
))
111 FILE *f
= fopen(filename
, "w");
115 fwrite(b
->spirv
, sizeof(*b
->spirv
), b
->spirv_word_count
, f
);
118 vtn_info("SPIR-V shader dumped to %s", filename
);
122 _vtn_warn(struct vtn_builder
*b
, const char *file
, unsigned line
,
123 const char *fmt
, ...)
128 vtn_log_err(b
, NIR_SPIRV_DEBUG_LEVEL_WARNING
, "SPIR-V WARNING:\n",
129 file
, line
, fmt
, args
);
134 _vtn_err(struct vtn_builder
*b
, const char *file
, unsigned line
,
135 const char *fmt
, ...)
140 vtn_log_err(b
, NIR_SPIRV_DEBUG_LEVEL_ERROR
, "SPIR-V ERROR:\n",
141 file
, line
, fmt
, args
);
146 _vtn_fail(struct vtn_builder
*b
, const char *file
, unsigned line
,
147 const char *fmt
, ...)
152 vtn_log_err(b
, NIR_SPIRV_DEBUG_LEVEL_ERROR
, "SPIR-V parsing FAILED:\n",
153 file
, line
, fmt
, args
);
156 const char *dump_path
= getenv("MESA_SPIRV_FAIL_DUMP_PATH");
158 vtn_dump_shader(b
, dump_path
, "fail");
160 longjmp(b
->fail_jump
, 1);
163 struct spec_constant_value
{
171 static struct vtn_ssa_value
*
172 vtn_undef_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
174 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
177 if (glsl_type_is_vector_or_scalar(type
)) {
178 unsigned num_components
= glsl_get_vector_elements(val
->type
);
179 unsigned bit_size
= glsl_get_bit_size(val
->type
);
180 val
->def
= nir_ssa_undef(&b
->nb
, num_components
, bit_size
);
182 unsigned elems
= glsl_get_length(val
->type
);
183 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
184 if (glsl_type_is_matrix(type
)) {
185 const struct glsl_type
*elem_type
=
186 glsl_vector_type(glsl_get_base_type(type
),
187 glsl_get_vector_elements(type
));
189 for (unsigned i
= 0; i
< elems
; i
++)
190 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
191 } else if (glsl_type_is_array(type
)) {
192 const struct glsl_type
*elem_type
= glsl_get_array_element(type
);
193 for (unsigned i
= 0; i
< elems
; i
++)
194 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
196 for (unsigned i
= 0; i
< elems
; i
++) {
197 const struct glsl_type
*elem_type
= glsl_get_struct_field(type
, i
);
198 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
206 static struct vtn_ssa_value
*
207 vtn_const_ssa_value(struct vtn_builder
*b
, nir_constant
*constant
,
208 const struct glsl_type
*type
)
210 struct hash_entry
*entry
= _mesa_hash_table_search(b
->const_table
, constant
);
215 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
218 switch (glsl_get_base_type(type
)) {
221 case GLSL_TYPE_INT16
:
222 case GLSL_TYPE_UINT16
:
223 case GLSL_TYPE_UINT8
:
225 case GLSL_TYPE_INT64
:
226 case GLSL_TYPE_UINT64
:
228 case GLSL_TYPE_FLOAT
:
229 case GLSL_TYPE_FLOAT16
:
230 case GLSL_TYPE_DOUBLE
: {
231 int bit_size
= glsl_get_bit_size(type
);
232 if (glsl_type_is_vector_or_scalar(type
)) {
233 unsigned num_components
= glsl_get_vector_elements(val
->type
);
234 nir_load_const_instr
*load
=
235 nir_load_const_instr_create(b
->shader
, num_components
, bit_size
);
237 load
->value
= constant
->values
[0];
239 nir_instr_insert_before_cf_list(&b
->nb
.impl
->body
, &load
->instr
);
240 val
->def
= &load
->def
;
242 assert(glsl_type_is_matrix(type
));
243 unsigned rows
= glsl_get_vector_elements(val
->type
);
244 unsigned columns
= glsl_get_matrix_columns(val
->type
);
245 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, columns
);
247 for (unsigned i
= 0; i
< columns
; i
++) {
248 struct vtn_ssa_value
*col_val
= rzalloc(b
, struct vtn_ssa_value
);
249 col_val
->type
= glsl_get_column_type(val
->type
);
250 nir_load_const_instr
*load
=
251 nir_load_const_instr_create(b
->shader
, rows
, bit_size
);
253 load
->value
= constant
->values
[i
];
255 nir_instr_insert_before_cf_list(&b
->nb
.impl
->body
, &load
->instr
);
256 col_val
->def
= &load
->def
;
258 val
->elems
[i
] = col_val
;
264 case GLSL_TYPE_ARRAY
: {
265 unsigned elems
= glsl_get_length(val
->type
);
266 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
267 const struct glsl_type
*elem_type
= glsl_get_array_element(val
->type
);
268 for (unsigned i
= 0; i
< elems
; i
++)
269 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
274 case GLSL_TYPE_STRUCT
: {
275 unsigned elems
= glsl_get_length(val
->type
);
276 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
277 for (unsigned i
= 0; i
< elems
; i
++) {
278 const struct glsl_type
*elem_type
=
279 glsl_get_struct_field(val
->type
, i
);
280 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
287 vtn_fail("bad constant type");
293 struct vtn_ssa_value
*
294 vtn_ssa_value(struct vtn_builder
*b
, uint32_t value_id
)
296 struct vtn_value
*val
= vtn_untyped_value(b
, value_id
);
297 switch (val
->value_type
) {
298 case vtn_value_type_undef
:
299 return vtn_undef_ssa_value(b
, val
->type
->type
);
301 case vtn_value_type_constant
:
302 return vtn_const_ssa_value(b
, val
->constant
, val
->type
->type
);
304 case vtn_value_type_ssa
:
307 case vtn_value_type_pointer
:
308 vtn_assert(val
->pointer
->ptr_type
&& val
->pointer
->ptr_type
->type
);
309 struct vtn_ssa_value
*ssa
=
310 vtn_create_ssa_value(b
, val
->pointer
->ptr_type
->type
);
311 ssa
->def
= vtn_pointer_to_ssa(b
, val
->pointer
);
315 vtn_fail("Invalid type for an SSA value");
320 vtn_string_literal(struct vtn_builder
*b
, const uint32_t *words
,
321 unsigned word_count
, unsigned *words_used
)
323 char *dup
= ralloc_strndup(b
, (char *)words
, word_count
* sizeof(*words
));
325 /* Ammount of space taken by the string (including the null) */
326 unsigned len
= strlen(dup
) + 1;
327 *words_used
= DIV_ROUND_UP(len
, sizeof(*words
));
333 vtn_foreach_instruction(struct vtn_builder
*b
, const uint32_t *start
,
334 const uint32_t *end
, vtn_instruction_handler handler
)
340 const uint32_t *w
= start
;
342 SpvOp opcode
= w
[0] & SpvOpCodeMask
;
343 unsigned count
= w
[0] >> SpvWordCountShift
;
344 vtn_assert(count
>= 1 && w
+ count
<= end
);
346 b
->spirv_offset
= (uint8_t *)w
- (uint8_t *)b
->spirv
;
350 break; /* Do nothing */
353 b
->file
= vtn_value(b
, w
[1], vtn_value_type_string
)->str
;
365 if (!handler(b
, opcode
, w
, count
))
383 vtn_handle_extension(struct vtn_builder
*b
, SpvOp opcode
,
384 const uint32_t *w
, unsigned count
)
386 const char *ext
= (const char *)&w
[2];
388 case SpvOpExtInstImport
: {
389 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_extension
);
390 if (strcmp(ext
, "GLSL.std.450") == 0) {
391 val
->ext_handler
= vtn_handle_glsl450_instruction
;
392 } else if ((strcmp(ext
, "SPV_AMD_gcn_shader") == 0)
393 && (b
->options
&& b
->options
->caps
.gcn_shader
)) {
394 val
->ext_handler
= vtn_handle_amd_gcn_shader_instruction
;
395 } else if ((strcmp(ext
, "SPV_AMD_shader_trinary_minmax") == 0)
396 && (b
->options
&& b
->options
->caps
.trinary_minmax
)) {
397 val
->ext_handler
= vtn_handle_amd_shader_trinary_minmax_instruction
;
399 vtn_fail("Unsupported extension: %s", ext
);
405 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
406 bool handled
= val
->ext_handler(b
, w
[4], w
, count
);
412 vtn_fail("Unhandled opcode");
417 _foreach_decoration_helper(struct vtn_builder
*b
,
418 struct vtn_value
*base_value
,
420 struct vtn_value
*value
,
421 vtn_decoration_foreach_cb cb
, void *data
)
423 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
425 if (dec
->scope
== VTN_DEC_DECORATION
) {
426 member
= parent_member
;
427 } else if (dec
->scope
>= VTN_DEC_STRUCT_MEMBER0
) {
428 vtn_fail_if(value
->value_type
!= vtn_value_type_type
||
429 value
->type
->base_type
!= vtn_base_type_struct
,
430 "OpMemberDecorate and OpGroupMemberDecorate are only "
431 "allowed on OpTypeStruct");
432 /* This means we haven't recursed yet */
433 assert(value
== base_value
);
435 member
= dec
->scope
- VTN_DEC_STRUCT_MEMBER0
;
437 vtn_fail_if(member
>= base_value
->type
->length
,
438 "OpMemberDecorate specifies member %d but the "
439 "OpTypeStruct has only %u members",
440 member
, base_value
->type
->length
);
442 /* Not a decoration */
443 assert(dec
->scope
== VTN_DEC_EXECUTION_MODE
);
448 assert(dec
->group
->value_type
== vtn_value_type_decoration_group
);
449 _foreach_decoration_helper(b
, base_value
, member
, dec
->group
,
452 cb(b
, base_value
, member
, dec
, data
);
457 /** Iterates (recursively if needed) over all of the decorations on a value
459 * This function iterates over all of the decorations applied to a given
460 * value. If it encounters a decoration group, it recurses into the group
461 * and iterates over all of those decorations as well.
464 vtn_foreach_decoration(struct vtn_builder
*b
, struct vtn_value
*value
,
465 vtn_decoration_foreach_cb cb
, void *data
)
467 _foreach_decoration_helper(b
, value
, -1, value
, cb
, data
);
471 vtn_foreach_execution_mode(struct vtn_builder
*b
, struct vtn_value
*value
,
472 vtn_execution_mode_foreach_cb cb
, void *data
)
474 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
475 if (dec
->scope
!= VTN_DEC_EXECUTION_MODE
)
478 assert(dec
->group
== NULL
);
479 cb(b
, value
, dec
, data
);
484 vtn_handle_decoration(struct vtn_builder
*b
, SpvOp opcode
,
485 const uint32_t *w
, unsigned count
)
487 const uint32_t *w_end
= w
+ count
;
488 const uint32_t target
= w
[1];
492 case SpvOpDecorationGroup
:
493 vtn_push_value(b
, target
, vtn_value_type_decoration_group
);
497 case SpvOpMemberDecorate
:
498 case SpvOpDecorateStringGOOGLE
:
499 case SpvOpMemberDecorateStringGOOGLE
:
500 case SpvOpExecutionMode
: {
501 struct vtn_value
*val
= vtn_untyped_value(b
, target
);
503 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
506 case SpvOpDecorateStringGOOGLE
:
507 dec
->scope
= VTN_DEC_DECORATION
;
509 case SpvOpMemberDecorate
:
510 case SpvOpMemberDecorateStringGOOGLE
:
511 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(w
++);
512 vtn_fail_if(dec
->scope
< VTN_DEC_STRUCT_MEMBER0
, /* overflow */
513 "Member argument of OpMemberDecorate too large");
515 case SpvOpExecutionMode
:
516 dec
->scope
= VTN_DEC_EXECUTION_MODE
;
519 unreachable("Invalid decoration opcode");
521 dec
->decoration
= *(w
++);
524 /* Link into the list */
525 dec
->next
= val
->decoration
;
526 val
->decoration
= dec
;
530 case SpvOpGroupMemberDecorate
:
531 case SpvOpGroupDecorate
: {
532 struct vtn_value
*group
=
533 vtn_value(b
, target
, vtn_value_type_decoration_group
);
535 for (; w
< w_end
; w
++) {
536 struct vtn_value
*val
= vtn_untyped_value(b
, *w
);
537 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
540 if (opcode
== SpvOpGroupDecorate
) {
541 dec
->scope
= VTN_DEC_DECORATION
;
543 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(++w
);
544 vtn_fail_if(dec
->scope
< 0, /* Check for overflow */
545 "Member argument of OpGroupMemberDecorate too large");
548 /* Link into the list */
549 dec
->next
= val
->decoration
;
550 val
->decoration
= dec
;
556 unreachable("Unhandled opcode");
560 struct member_decoration_ctx
{
562 struct glsl_struct_field
*fields
;
563 struct vtn_type
*type
;
566 /** Returns true if two types are "compatible", i.e. you can do an OpLoad,
567 * OpStore, or OpCopyMemory between them without breaking anything.
568 * Technically, the SPIR-V rules require the exact same type ID but this lets
569 * us internally be a bit looser.
572 vtn_types_compatible(struct vtn_builder
*b
,
573 struct vtn_type
*t1
, struct vtn_type
*t2
)
575 if (t1
->id
== t2
->id
)
578 if (t1
->base_type
!= t2
->base_type
)
581 switch (t1
->base_type
) {
582 case vtn_base_type_void
:
583 case vtn_base_type_scalar
:
584 case vtn_base_type_vector
:
585 case vtn_base_type_matrix
:
586 case vtn_base_type_image
:
587 case vtn_base_type_sampler
:
588 case vtn_base_type_sampled_image
:
589 return t1
->type
== t2
->type
;
591 case vtn_base_type_array
:
592 return t1
->length
== t2
->length
&&
593 vtn_types_compatible(b
, t1
->array_element
, t2
->array_element
);
595 case vtn_base_type_pointer
:
596 return vtn_types_compatible(b
, t1
->deref
, t2
->deref
);
598 case vtn_base_type_struct
:
599 if (t1
->length
!= t2
->length
)
602 for (unsigned i
= 0; i
< t1
->length
; i
++) {
603 if (!vtn_types_compatible(b
, t1
->members
[i
], t2
->members
[i
]))
608 case vtn_base_type_function
:
609 /* This case shouldn't get hit since you can't copy around function
610 * types. Just require them to be identical.
615 vtn_fail("Invalid base type");
618 /* does a shallow copy of a vtn_type */
620 static struct vtn_type
*
621 vtn_type_copy(struct vtn_builder
*b
, struct vtn_type
*src
)
623 struct vtn_type
*dest
= ralloc(b
, struct vtn_type
);
626 switch (src
->base_type
) {
627 case vtn_base_type_void
:
628 case vtn_base_type_scalar
:
629 case vtn_base_type_vector
:
630 case vtn_base_type_matrix
:
631 case vtn_base_type_array
:
632 case vtn_base_type_pointer
:
633 case vtn_base_type_image
:
634 case vtn_base_type_sampler
:
635 case vtn_base_type_sampled_image
:
636 /* Nothing more to do */
639 case vtn_base_type_struct
:
640 dest
->members
= ralloc_array(b
, struct vtn_type
*, src
->length
);
641 memcpy(dest
->members
, src
->members
,
642 src
->length
* sizeof(src
->members
[0]));
644 dest
->offsets
= ralloc_array(b
, unsigned, src
->length
);
645 memcpy(dest
->offsets
, src
->offsets
,
646 src
->length
* sizeof(src
->offsets
[0]));
649 case vtn_base_type_function
:
650 dest
->params
= ralloc_array(b
, struct vtn_type
*, src
->length
);
651 memcpy(dest
->params
, src
->params
, src
->length
* sizeof(src
->params
[0]));
658 static struct vtn_type
*
659 mutable_matrix_member(struct vtn_builder
*b
, struct vtn_type
*type
, int member
)
661 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
662 type
= type
->members
[member
];
664 /* We may have an array of matrices.... Oh, joy! */
665 while (glsl_type_is_array(type
->type
)) {
666 type
->array_element
= vtn_type_copy(b
, type
->array_element
);
667 type
= type
->array_element
;
670 vtn_assert(glsl_type_is_matrix(type
->type
));
676 vtn_handle_access_qualifier(struct vtn_builder
*b
, struct vtn_type
*type
,
677 int member
, enum gl_access_qualifier access
)
679 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
680 type
= type
->members
[member
];
682 type
->access
|= access
;
686 struct_member_decoration_cb(struct vtn_builder
*b
,
687 struct vtn_value
*val
, int member
,
688 const struct vtn_decoration
*dec
, void *void_ctx
)
690 struct member_decoration_ctx
*ctx
= void_ctx
;
695 assert(member
< ctx
->num_fields
);
697 switch (dec
->decoration
) {
698 case SpvDecorationRelaxedPrecision
:
699 case SpvDecorationUniform
:
700 break; /* FIXME: Do nothing with this for now. */
701 case SpvDecorationNonWritable
:
702 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_NON_WRITEABLE
);
704 case SpvDecorationNonReadable
:
705 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_NON_READABLE
);
707 case SpvDecorationVolatile
:
708 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_VOLATILE
);
710 case SpvDecorationCoherent
:
711 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_COHERENT
);
713 case SpvDecorationNoPerspective
:
714 ctx
->fields
[member
].interpolation
= INTERP_MODE_NOPERSPECTIVE
;
716 case SpvDecorationFlat
:
717 ctx
->fields
[member
].interpolation
= INTERP_MODE_FLAT
;
719 case SpvDecorationCentroid
:
720 ctx
->fields
[member
].centroid
= true;
722 case SpvDecorationSample
:
723 ctx
->fields
[member
].sample
= true;
725 case SpvDecorationStream
:
726 /* Vulkan only allows one GS stream */
727 vtn_assert(dec
->literals
[0] == 0);
729 case SpvDecorationLocation
:
730 ctx
->fields
[member
].location
= dec
->literals
[0];
732 case SpvDecorationComponent
:
733 break; /* FIXME: What should we do with these? */
734 case SpvDecorationBuiltIn
:
735 ctx
->type
->members
[member
] = vtn_type_copy(b
, ctx
->type
->members
[member
]);
736 ctx
->type
->members
[member
]->is_builtin
= true;
737 ctx
->type
->members
[member
]->builtin
= dec
->literals
[0];
738 ctx
->type
->builtin_block
= true;
740 case SpvDecorationOffset
:
741 ctx
->type
->offsets
[member
] = dec
->literals
[0];
743 case SpvDecorationMatrixStride
:
744 /* Handled as a second pass */
746 case SpvDecorationColMajor
:
747 break; /* Nothing to do here. Column-major is the default. */
748 case SpvDecorationRowMajor
:
749 mutable_matrix_member(b
, ctx
->type
, member
)->row_major
= true;
752 case SpvDecorationPatch
:
755 case SpvDecorationSpecId
:
756 case SpvDecorationBlock
:
757 case SpvDecorationBufferBlock
:
758 case SpvDecorationArrayStride
:
759 case SpvDecorationGLSLShared
:
760 case SpvDecorationGLSLPacked
:
761 case SpvDecorationInvariant
:
762 case SpvDecorationRestrict
:
763 case SpvDecorationAliased
:
764 case SpvDecorationConstant
:
765 case SpvDecorationIndex
:
766 case SpvDecorationBinding
:
767 case SpvDecorationDescriptorSet
:
768 case SpvDecorationLinkageAttributes
:
769 case SpvDecorationNoContraction
:
770 case SpvDecorationInputAttachmentIndex
:
771 vtn_warn("Decoration not allowed on struct members: %s",
772 spirv_decoration_to_string(dec
->decoration
));
775 case SpvDecorationXfbBuffer
:
776 case SpvDecorationXfbStride
:
777 vtn_warn("Vulkan does not have transform feedback");
780 case SpvDecorationCPacked
:
781 case SpvDecorationSaturatedConversion
:
782 case SpvDecorationFuncParamAttr
:
783 case SpvDecorationFPRoundingMode
:
784 case SpvDecorationFPFastMathMode
:
785 case SpvDecorationAlignment
:
786 vtn_warn("Decoration only allowed for CL-style kernels: %s",
787 spirv_decoration_to_string(dec
->decoration
));
791 vtn_fail("Unhandled decoration");
795 /* Matrix strides are handled as a separate pass because we need to know
796 * whether the matrix is row-major or not first.
799 struct_member_matrix_stride_cb(struct vtn_builder
*b
,
800 struct vtn_value
*val
, int member
,
801 const struct vtn_decoration
*dec
,
804 if (dec
->decoration
!= SpvDecorationMatrixStride
)
807 vtn_fail_if(member
< 0,
808 "The MatrixStride decoration is only allowed on members "
811 struct member_decoration_ctx
*ctx
= void_ctx
;
813 struct vtn_type
*mat_type
= mutable_matrix_member(b
, ctx
->type
, member
);
814 if (mat_type
->row_major
) {
815 mat_type
->array_element
= vtn_type_copy(b
, mat_type
->array_element
);
816 mat_type
->stride
= mat_type
->array_element
->stride
;
817 mat_type
->array_element
->stride
= dec
->literals
[0];
819 vtn_assert(mat_type
->array_element
->stride
> 0);
820 mat_type
->stride
= dec
->literals
[0];
825 type_decoration_cb(struct vtn_builder
*b
,
826 struct vtn_value
*val
, int member
,
827 const struct vtn_decoration
*dec
, void *ctx
)
829 struct vtn_type
*type
= val
->type
;
832 /* This should have been handled by OpTypeStruct */
833 assert(val
->type
->base_type
== vtn_base_type_struct
);
834 assert(member
>= 0 && member
< val
->type
->length
);
838 switch (dec
->decoration
) {
839 case SpvDecorationArrayStride
:
840 vtn_assert(type
->base_type
== vtn_base_type_matrix
||
841 type
->base_type
== vtn_base_type_array
||
842 type
->base_type
== vtn_base_type_pointer
);
843 type
->stride
= dec
->literals
[0];
845 case SpvDecorationBlock
:
846 vtn_assert(type
->base_type
== vtn_base_type_struct
);
849 case SpvDecorationBufferBlock
:
850 vtn_assert(type
->base_type
== vtn_base_type_struct
);
851 type
->buffer_block
= true;
853 case SpvDecorationGLSLShared
:
854 case SpvDecorationGLSLPacked
:
855 /* Ignore these, since we get explicit offsets anyways */
858 case SpvDecorationRowMajor
:
859 case SpvDecorationColMajor
:
860 case SpvDecorationMatrixStride
:
861 case SpvDecorationBuiltIn
:
862 case SpvDecorationNoPerspective
:
863 case SpvDecorationFlat
:
864 case SpvDecorationPatch
:
865 case SpvDecorationCentroid
:
866 case SpvDecorationSample
:
867 case SpvDecorationVolatile
:
868 case SpvDecorationCoherent
:
869 case SpvDecorationNonWritable
:
870 case SpvDecorationNonReadable
:
871 case SpvDecorationUniform
:
872 case SpvDecorationLocation
:
873 case SpvDecorationComponent
:
874 case SpvDecorationOffset
:
875 case SpvDecorationXfbBuffer
:
876 case SpvDecorationXfbStride
:
877 vtn_warn("Decoration only allowed for struct members: %s",
878 spirv_decoration_to_string(dec
->decoration
));
881 case SpvDecorationStream
:
882 /* We don't need to do anything here, as stream is filled up when
883 * aplying the decoration to a variable, just check that if it is not a
884 * struct member, it should be a struct.
886 vtn_assert(type
->base_type
== vtn_base_type_struct
);
889 case SpvDecorationRelaxedPrecision
:
890 case SpvDecorationSpecId
:
891 case SpvDecorationInvariant
:
892 case SpvDecorationRestrict
:
893 case SpvDecorationAliased
:
894 case SpvDecorationConstant
:
895 case SpvDecorationIndex
:
896 case SpvDecorationBinding
:
897 case SpvDecorationDescriptorSet
:
898 case SpvDecorationLinkageAttributes
:
899 case SpvDecorationNoContraction
:
900 case SpvDecorationInputAttachmentIndex
:
901 vtn_warn("Decoration not allowed on types: %s",
902 spirv_decoration_to_string(dec
->decoration
));
905 case SpvDecorationCPacked
:
906 case SpvDecorationSaturatedConversion
:
907 case SpvDecorationFuncParamAttr
:
908 case SpvDecorationFPRoundingMode
:
909 case SpvDecorationFPFastMathMode
:
910 case SpvDecorationAlignment
:
911 vtn_warn("Decoration only allowed for CL-style kernels: %s",
912 spirv_decoration_to_string(dec
->decoration
));
916 vtn_fail("Unhandled decoration");
921 translate_image_format(struct vtn_builder
*b
, SpvImageFormat format
)
924 case SpvImageFormatUnknown
: return 0; /* GL_NONE */
925 case SpvImageFormatRgba32f
: return 0x8814; /* GL_RGBA32F */
926 case SpvImageFormatRgba16f
: return 0x881A; /* GL_RGBA16F */
927 case SpvImageFormatR32f
: return 0x822E; /* GL_R32F */
928 case SpvImageFormatRgba8
: return 0x8058; /* GL_RGBA8 */
929 case SpvImageFormatRgba8Snorm
: return 0x8F97; /* GL_RGBA8_SNORM */
930 case SpvImageFormatRg32f
: return 0x8230; /* GL_RG32F */
931 case SpvImageFormatRg16f
: return 0x822F; /* GL_RG16F */
932 case SpvImageFormatR11fG11fB10f
: return 0x8C3A; /* GL_R11F_G11F_B10F */
933 case SpvImageFormatR16f
: return 0x822D; /* GL_R16F */
934 case SpvImageFormatRgba16
: return 0x805B; /* GL_RGBA16 */
935 case SpvImageFormatRgb10A2
: return 0x8059; /* GL_RGB10_A2 */
936 case SpvImageFormatRg16
: return 0x822C; /* GL_RG16 */
937 case SpvImageFormatRg8
: return 0x822B; /* GL_RG8 */
938 case SpvImageFormatR16
: return 0x822A; /* GL_R16 */
939 case SpvImageFormatR8
: return 0x8229; /* GL_R8 */
940 case SpvImageFormatRgba16Snorm
: return 0x8F9B; /* GL_RGBA16_SNORM */
941 case SpvImageFormatRg16Snorm
: return 0x8F99; /* GL_RG16_SNORM */
942 case SpvImageFormatRg8Snorm
: return 0x8F95; /* GL_RG8_SNORM */
943 case SpvImageFormatR16Snorm
: return 0x8F98; /* GL_R16_SNORM */
944 case SpvImageFormatR8Snorm
: return 0x8F94; /* GL_R8_SNORM */
945 case SpvImageFormatRgba32i
: return 0x8D82; /* GL_RGBA32I */
946 case SpvImageFormatRgba16i
: return 0x8D88; /* GL_RGBA16I */
947 case SpvImageFormatRgba8i
: return 0x8D8E; /* GL_RGBA8I */
948 case SpvImageFormatR32i
: return 0x8235; /* GL_R32I */
949 case SpvImageFormatRg32i
: return 0x823B; /* GL_RG32I */
950 case SpvImageFormatRg16i
: return 0x8239; /* GL_RG16I */
951 case SpvImageFormatRg8i
: return 0x8237; /* GL_RG8I */
952 case SpvImageFormatR16i
: return 0x8233; /* GL_R16I */
953 case SpvImageFormatR8i
: return 0x8231; /* GL_R8I */
954 case SpvImageFormatRgba32ui
: return 0x8D70; /* GL_RGBA32UI */
955 case SpvImageFormatRgba16ui
: return 0x8D76; /* GL_RGBA16UI */
956 case SpvImageFormatRgba8ui
: return 0x8D7C; /* GL_RGBA8UI */
957 case SpvImageFormatR32ui
: return 0x8236; /* GL_R32UI */
958 case SpvImageFormatRgb10a2ui
: return 0x906F; /* GL_RGB10_A2UI */
959 case SpvImageFormatRg32ui
: return 0x823C; /* GL_RG32UI */
960 case SpvImageFormatRg16ui
: return 0x823A; /* GL_RG16UI */
961 case SpvImageFormatRg8ui
: return 0x8238; /* GL_RG8UI */
962 case SpvImageFormatR16ui
: return 0x8234; /* GL_R16UI */
963 case SpvImageFormatR8ui
: return 0x8232; /* GL_R8UI */
965 vtn_fail("Invalid image format");
969 static struct vtn_type
*
970 vtn_type_layout_std430(struct vtn_builder
*b
, struct vtn_type
*type
,
971 uint32_t *size_out
, uint32_t *align_out
)
973 switch (type
->base_type
) {
974 case vtn_base_type_scalar
: {
975 uint32_t comp_size
= glsl_get_bit_size(type
->type
) / 8;
976 *size_out
= comp_size
;
977 *align_out
= comp_size
;
981 case vtn_base_type_vector
: {
982 uint32_t comp_size
= glsl_get_bit_size(type
->type
) / 8;
983 unsigned align_comps
= type
->length
== 3 ? 4 : type
->length
;
984 *size_out
= comp_size
* type
->length
,
985 *align_out
= comp_size
* align_comps
;
989 case vtn_base_type_matrix
:
990 case vtn_base_type_array
: {
991 /* We're going to add an array stride */
992 type
= vtn_type_copy(b
, type
);
993 uint32_t elem_size
, elem_align
;
994 type
->array_element
= vtn_type_layout_std430(b
, type
->array_element
,
995 &elem_size
, &elem_align
);
996 type
->stride
= vtn_align_u32(elem_size
, elem_align
);
997 *size_out
= type
->stride
* type
->length
;
998 *align_out
= elem_align
;
1002 case vtn_base_type_struct
: {
1003 /* We're going to add member offsets */
1004 type
= vtn_type_copy(b
, type
);
1005 uint32_t offset
= 0;
1007 for (unsigned i
= 0; i
< type
->length
; i
++) {
1008 uint32_t mem_size
, mem_align
;
1009 type
->members
[i
] = vtn_type_layout_std430(b
, type
->members
[i
],
1010 &mem_size
, &mem_align
);
1011 offset
= vtn_align_u32(offset
, mem_align
);
1012 type
->offsets
[i
] = offset
;
1014 align
= MAX2(align
, mem_align
);
1022 unreachable("Invalid SPIR-V type for std430");
1027 vtn_handle_type(struct vtn_builder
*b
, SpvOp opcode
,
1028 const uint32_t *w
, unsigned count
)
1030 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_type
);
1032 val
->type
= rzalloc(b
, struct vtn_type
);
1033 val
->type
->id
= w
[1];
1037 val
->type
->base_type
= vtn_base_type_void
;
1038 val
->type
->type
= glsl_void_type();
1041 val
->type
->base_type
= vtn_base_type_scalar
;
1042 val
->type
->type
= glsl_bool_type();
1043 val
->type
->length
= 1;
1045 case SpvOpTypeInt
: {
1046 int bit_size
= w
[2];
1047 const bool signedness
= w
[3];
1048 val
->type
->base_type
= vtn_base_type_scalar
;
1051 val
->type
->type
= (signedness
? glsl_int64_t_type() : glsl_uint64_t_type());
1054 val
->type
->type
= (signedness
? glsl_int_type() : glsl_uint_type());
1057 val
->type
->type
= (signedness
? glsl_int16_t_type() : glsl_uint16_t_type());
1060 val
->type
->type
= (signedness
? glsl_int8_t_type() : glsl_uint8_t_type());
1063 vtn_fail("Invalid int bit size");
1065 val
->type
->length
= 1;
1069 case SpvOpTypeFloat
: {
1070 int bit_size
= w
[2];
1071 val
->type
->base_type
= vtn_base_type_scalar
;
1074 val
->type
->type
= glsl_float16_t_type();
1077 val
->type
->type
= glsl_float_type();
1080 val
->type
->type
= glsl_double_type();
1083 vtn_fail("Invalid float bit size");
1085 val
->type
->length
= 1;
1089 case SpvOpTypeVector
: {
1090 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1091 unsigned elems
= w
[3];
1093 vtn_fail_if(base
->base_type
!= vtn_base_type_scalar
,
1094 "Base type for OpTypeVector must be a scalar");
1095 vtn_fail_if((elems
< 2 || elems
> 4) && (elems
!= 8) && (elems
!= 16),
1096 "Invalid component count for OpTypeVector");
1098 val
->type
->base_type
= vtn_base_type_vector
;
1099 val
->type
->type
= glsl_vector_type(glsl_get_base_type(base
->type
), elems
);
1100 val
->type
->length
= elems
;
1101 val
->type
->stride
= glsl_get_bit_size(base
->type
) / 8;
1102 val
->type
->array_element
= base
;
1106 case SpvOpTypeMatrix
: {
1107 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1108 unsigned columns
= w
[3];
1110 vtn_fail_if(base
->base_type
!= vtn_base_type_vector
,
1111 "Base type for OpTypeMatrix must be a vector");
1112 vtn_fail_if(columns
< 2 || columns
> 4,
1113 "Invalid column count for OpTypeMatrix");
1115 val
->type
->base_type
= vtn_base_type_matrix
;
1116 val
->type
->type
= glsl_matrix_type(glsl_get_base_type(base
->type
),
1117 glsl_get_vector_elements(base
->type
),
1119 vtn_fail_if(glsl_type_is_error(val
->type
->type
),
1120 "Unsupported base type for OpTypeMatrix");
1121 assert(!glsl_type_is_error(val
->type
->type
));
1122 val
->type
->length
= columns
;
1123 val
->type
->array_element
= base
;
1124 val
->type
->row_major
= false;
1125 val
->type
->stride
= 0;
1129 case SpvOpTypeRuntimeArray
:
1130 case SpvOpTypeArray
: {
1131 struct vtn_type
*array_element
=
1132 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1134 if (opcode
== SpvOpTypeRuntimeArray
) {
1135 /* A length of 0 is used to denote unsized arrays */
1136 val
->type
->length
= 0;
1139 vtn_value(b
, w
[3], vtn_value_type_constant
)->constant
->values
[0].u32
[0];
1142 val
->type
->base_type
= vtn_base_type_array
;
1143 val
->type
->type
= glsl_array_type(array_element
->type
, val
->type
->length
);
1144 val
->type
->array_element
= array_element
;
1145 val
->type
->stride
= 0;
1149 case SpvOpTypeStruct
: {
1150 unsigned num_fields
= count
- 2;
1151 val
->type
->base_type
= vtn_base_type_struct
;
1152 val
->type
->length
= num_fields
;
1153 val
->type
->members
= ralloc_array(b
, struct vtn_type
*, num_fields
);
1154 val
->type
->offsets
= ralloc_array(b
, unsigned, num_fields
);
1156 NIR_VLA(struct glsl_struct_field
, fields
, count
);
1157 for (unsigned i
= 0; i
< num_fields
; i
++) {
1158 val
->type
->members
[i
] =
1159 vtn_value(b
, w
[i
+ 2], vtn_value_type_type
)->type
;
1160 fields
[i
] = (struct glsl_struct_field
) {
1161 .type
= val
->type
->members
[i
]->type
,
1162 .name
= ralloc_asprintf(b
, "field%d", i
),
1167 struct member_decoration_ctx ctx
= {
1168 .num_fields
= num_fields
,
1173 vtn_foreach_decoration(b
, val
, struct_member_decoration_cb
, &ctx
);
1174 vtn_foreach_decoration(b
, val
, struct_member_matrix_stride_cb
, &ctx
);
1176 const char *name
= val
->name
? val
->name
: "struct";
1178 val
->type
->type
= glsl_struct_type(fields
, num_fields
, name
);
1182 case SpvOpTypeFunction
: {
1183 val
->type
->base_type
= vtn_base_type_function
;
1184 val
->type
->type
= NULL
;
1186 val
->type
->return_type
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1188 const unsigned num_params
= count
- 3;
1189 val
->type
->length
= num_params
;
1190 val
->type
->params
= ralloc_array(b
, struct vtn_type
*, num_params
);
1191 for (unsigned i
= 0; i
< count
- 3; i
++) {
1192 val
->type
->params
[i
] =
1193 vtn_value(b
, w
[i
+ 3], vtn_value_type_type
)->type
;
1198 case SpvOpTypePointer
: {
1199 SpvStorageClass storage_class
= w
[2];
1200 struct vtn_type
*deref_type
=
1201 vtn_value(b
, w
[3], vtn_value_type_type
)->type
;
1203 val
->type
->base_type
= vtn_base_type_pointer
;
1204 val
->type
->storage_class
= storage_class
;
1205 val
->type
->deref
= deref_type
;
1207 if (storage_class
== SpvStorageClassUniform
||
1208 storage_class
== SpvStorageClassStorageBuffer
) {
1209 /* These can actually be stored to nir_variables and used as SSA
1210 * values so they need a real glsl_type.
1212 val
->type
->type
= glsl_vector_type(GLSL_TYPE_UINT
, 2);
1215 if (storage_class
== SpvStorageClassPushConstant
) {
1216 /* These can actually be stored to nir_variables and used as SSA
1217 * values so they need a real glsl_type.
1219 val
->type
->type
= glsl_uint_type();
1222 if (storage_class
== SpvStorageClassWorkgroup
&&
1223 b
->options
->lower_workgroup_access_to_offsets
) {
1224 uint32_t size
, align
;
1225 val
->type
->deref
= vtn_type_layout_std430(b
, val
->type
->deref
,
1227 val
->type
->length
= size
;
1228 val
->type
->align
= align
;
1229 /* These can actually be stored to nir_variables and used as SSA
1230 * values so they need a real glsl_type.
1232 val
->type
->type
= glsl_uint_type();
1237 case SpvOpTypeImage
: {
1238 val
->type
->base_type
= vtn_base_type_image
;
1240 const struct vtn_type
*sampled_type
=
1241 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1243 vtn_fail_if(sampled_type
->base_type
!= vtn_base_type_scalar
||
1244 glsl_get_bit_size(sampled_type
->type
) != 32,
1245 "Sampled type of OpTypeImage must be a 32-bit scalar");
1247 enum glsl_sampler_dim dim
;
1248 switch ((SpvDim
)w
[3]) {
1249 case SpvDim1D
: dim
= GLSL_SAMPLER_DIM_1D
; break;
1250 case SpvDim2D
: dim
= GLSL_SAMPLER_DIM_2D
; break;
1251 case SpvDim3D
: dim
= GLSL_SAMPLER_DIM_3D
; break;
1252 case SpvDimCube
: dim
= GLSL_SAMPLER_DIM_CUBE
; break;
1253 case SpvDimRect
: dim
= GLSL_SAMPLER_DIM_RECT
; break;
1254 case SpvDimBuffer
: dim
= GLSL_SAMPLER_DIM_BUF
; break;
1255 case SpvDimSubpassData
: dim
= GLSL_SAMPLER_DIM_SUBPASS
; break;
1257 vtn_fail("Invalid SPIR-V image dimensionality");
1260 bool is_shadow
= w
[4];
1261 bool is_array
= w
[5];
1262 bool multisampled
= w
[6];
1263 unsigned sampled
= w
[7];
1264 SpvImageFormat format
= w
[8];
1267 val
->type
->access_qualifier
= w
[9];
1269 val
->type
->access_qualifier
= SpvAccessQualifierReadWrite
;
1272 if (dim
== GLSL_SAMPLER_DIM_2D
)
1273 dim
= GLSL_SAMPLER_DIM_MS
;
1274 else if (dim
== GLSL_SAMPLER_DIM_SUBPASS
)
1275 dim
= GLSL_SAMPLER_DIM_SUBPASS_MS
;
1277 vtn_fail("Unsupported multisampled image type");
1280 val
->type
->image_format
= translate_image_format(b
, format
);
1282 enum glsl_base_type sampled_base_type
=
1283 glsl_get_base_type(sampled_type
->type
);
1285 val
->type
->sampled
= true;
1286 val
->type
->type
= glsl_sampler_type(dim
, is_shadow
, is_array
,
1288 } else if (sampled
== 2) {
1289 vtn_assert(!is_shadow
);
1290 val
->type
->sampled
= false;
1291 val
->type
->type
= glsl_image_type(dim
, is_array
, sampled_base_type
);
1293 vtn_fail("We need to know if the image will be sampled");
1298 case SpvOpTypeSampledImage
:
1299 val
->type
->base_type
= vtn_base_type_sampled_image
;
1300 val
->type
->image
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1301 val
->type
->type
= val
->type
->image
->type
;
1304 case SpvOpTypeSampler
:
1305 /* The actual sampler type here doesn't really matter. It gets
1306 * thrown away the moment you combine it with an image. What really
1307 * matters is that it's a sampler type as opposed to an integer type
1308 * so the backend knows what to do.
1310 val
->type
->base_type
= vtn_base_type_sampler
;
1311 val
->type
->type
= glsl_bare_sampler_type();
1314 case SpvOpTypeOpaque
:
1315 case SpvOpTypeEvent
:
1316 case SpvOpTypeDeviceEvent
:
1317 case SpvOpTypeReserveId
:
1318 case SpvOpTypeQueue
:
1321 vtn_fail("Unhandled opcode");
1324 vtn_foreach_decoration(b
, val
, type_decoration_cb
, NULL
);
1327 static nir_constant
*
1328 vtn_null_constant(struct vtn_builder
*b
, const struct glsl_type
*type
)
1330 nir_constant
*c
= rzalloc(b
, nir_constant
);
1332 /* For pointers and other typeless things, we have to return something but
1333 * it doesn't matter what.
1338 switch (glsl_get_base_type(type
)) {
1340 case GLSL_TYPE_UINT
:
1341 case GLSL_TYPE_INT16
:
1342 case GLSL_TYPE_UINT16
:
1343 case GLSL_TYPE_UINT8
:
1344 case GLSL_TYPE_INT8
:
1345 case GLSL_TYPE_INT64
:
1346 case GLSL_TYPE_UINT64
:
1347 case GLSL_TYPE_BOOL
:
1348 case GLSL_TYPE_FLOAT
:
1349 case GLSL_TYPE_FLOAT16
:
1350 case GLSL_TYPE_DOUBLE
:
1351 /* Nothing to do here. It's already initialized to zero */
1354 case GLSL_TYPE_ARRAY
:
1355 vtn_assert(glsl_get_length(type
) > 0);
1356 c
->num_elements
= glsl_get_length(type
);
1357 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1359 c
->elements
[0] = vtn_null_constant(b
, glsl_get_array_element(type
));
1360 for (unsigned i
= 1; i
< c
->num_elements
; i
++)
1361 c
->elements
[i
] = c
->elements
[0];
1364 case GLSL_TYPE_STRUCT
:
1365 c
->num_elements
= glsl_get_length(type
);
1366 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1368 for (unsigned i
= 0; i
< c
->num_elements
; i
++) {
1369 c
->elements
[i
] = vtn_null_constant(b
, glsl_get_struct_field(type
, i
));
1374 vtn_fail("Invalid type for null constant");
1381 spec_constant_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*v
,
1382 int member
, const struct vtn_decoration
*dec
,
1385 vtn_assert(member
== -1);
1386 if (dec
->decoration
!= SpvDecorationSpecId
)
1389 struct spec_constant_value
*const_value
= data
;
1391 for (unsigned i
= 0; i
< b
->num_specializations
; i
++) {
1392 if (b
->specializations
[i
].id
== dec
->literals
[0]) {
1393 if (const_value
->is_double
)
1394 const_value
->data64
= b
->specializations
[i
].data64
;
1396 const_value
->data32
= b
->specializations
[i
].data32
;
1403 get_specialization(struct vtn_builder
*b
, struct vtn_value
*val
,
1404 uint32_t const_value
)
1406 struct spec_constant_value data
;
1407 data
.is_double
= false;
1408 data
.data32
= const_value
;
1409 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1414 get_specialization64(struct vtn_builder
*b
, struct vtn_value
*val
,
1415 uint64_t const_value
)
1417 struct spec_constant_value data
;
1418 data
.is_double
= true;
1419 data
.data64
= const_value
;
1420 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1425 handle_workgroup_size_decoration_cb(struct vtn_builder
*b
,
1426 struct vtn_value
*val
,
1428 const struct vtn_decoration
*dec
,
1431 vtn_assert(member
== -1);
1432 if (dec
->decoration
!= SpvDecorationBuiltIn
||
1433 dec
->literals
[0] != SpvBuiltInWorkgroupSize
)
1436 vtn_assert(val
->type
->type
== glsl_vector_type(GLSL_TYPE_UINT
, 3));
1438 b
->shader
->info
.cs
.local_size
[0] = val
->constant
->values
[0].u32
[0];
1439 b
->shader
->info
.cs
.local_size
[1] = val
->constant
->values
[0].u32
[1];
1440 b
->shader
->info
.cs
.local_size
[2] = val
->constant
->values
[0].u32
[2];
1444 vtn_handle_constant(struct vtn_builder
*b
, SpvOp opcode
,
1445 const uint32_t *w
, unsigned count
)
1447 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_constant
);
1448 val
->constant
= rzalloc(b
, nir_constant
);
1450 case SpvOpConstantTrue
:
1451 case SpvOpConstantFalse
:
1452 case SpvOpSpecConstantTrue
:
1453 case SpvOpSpecConstantFalse
: {
1454 vtn_fail_if(val
->type
->type
!= glsl_bool_type(),
1455 "Result type of %s must be OpTypeBool",
1456 spirv_op_to_string(opcode
));
1458 uint32_t int_val
= (opcode
== SpvOpConstantTrue
||
1459 opcode
== SpvOpSpecConstantTrue
);
1461 if (opcode
== SpvOpSpecConstantTrue
||
1462 opcode
== SpvOpSpecConstantFalse
)
1463 int_val
= get_specialization(b
, val
, int_val
);
1465 val
->constant
->values
[0].u32
[0] = int_val
? NIR_TRUE
: NIR_FALSE
;
1469 case SpvOpConstant
: {
1470 vtn_fail_if(val
->type
->base_type
!= vtn_base_type_scalar
,
1471 "Result type of %s must be a scalar",
1472 spirv_op_to_string(opcode
));
1473 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1476 val
->constant
->values
->u64
[0] = vtn_u64_literal(&w
[3]);
1479 val
->constant
->values
->u32
[0] = w
[3];
1482 val
->constant
->values
->u16
[0] = w
[3];
1485 val
->constant
->values
->u8
[0] = w
[3];
1488 vtn_fail("Unsupported SpvOpConstant bit size");
1493 case SpvOpSpecConstant
: {
1494 vtn_fail_if(val
->type
->base_type
!= vtn_base_type_scalar
,
1495 "Result type of %s must be a scalar",
1496 spirv_op_to_string(opcode
));
1497 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1500 val
->constant
->values
[0].u64
[0] =
1501 get_specialization64(b
, val
, vtn_u64_literal(&w
[3]));
1504 val
->constant
->values
[0].u32
[0] = get_specialization(b
, val
, w
[3]);
1507 val
->constant
->values
[0].u16
[0] = get_specialization(b
, val
, w
[3]);
1510 val
->constant
->values
[0].u8
[0] = get_specialization(b
, val
, w
[3]);
1513 vtn_fail("Unsupported SpvOpSpecConstant bit size");
1518 case SpvOpSpecConstantComposite
:
1519 case SpvOpConstantComposite
: {
1520 unsigned elem_count
= count
- 3;
1521 vtn_fail_if(elem_count
!= val
->type
->length
,
1522 "%s has %u constituents, expected %u",
1523 spirv_op_to_string(opcode
), elem_count
, val
->type
->length
);
1525 nir_constant
**elems
= ralloc_array(b
, nir_constant
*, elem_count
);
1526 for (unsigned i
= 0; i
< elem_count
; i
++) {
1527 struct vtn_value
*val
= vtn_untyped_value(b
, w
[i
+ 3]);
1529 if (val
->value_type
== vtn_value_type_constant
) {
1530 elems
[i
] = val
->constant
;
1532 vtn_fail_if(val
->value_type
!= vtn_value_type_undef
,
1533 "only constants or undefs allowed for "
1534 "SpvOpConstantComposite");
1535 /* to make it easier, just insert a NULL constant for now */
1536 elems
[i
] = vtn_null_constant(b
, val
->type
->type
);
1540 switch (val
->type
->base_type
) {
1541 case vtn_base_type_vector
: {
1542 assert(glsl_type_is_vector(val
->type
->type
));
1543 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1544 for (unsigned i
= 0; i
< elem_count
; i
++) {
1547 val
->constant
->values
[0].u64
[i
] = elems
[i
]->values
[0].u64
[0];
1550 val
->constant
->values
[0].u32
[i
] = elems
[i
]->values
[0].u32
[0];
1553 val
->constant
->values
[0].u16
[i
] = elems
[i
]->values
[0].u16
[0];
1556 val
->constant
->values
[0].u8
[i
] = elems
[i
]->values
[0].u8
[0];
1559 vtn_fail("Invalid SpvOpConstantComposite bit size");
1565 case vtn_base_type_matrix
:
1566 assert(glsl_type_is_matrix(val
->type
->type
));
1567 for (unsigned i
= 0; i
< elem_count
; i
++)
1568 val
->constant
->values
[i
] = elems
[i
]->values
[0];
1571 case vtn_base_type_struct
:
1572 case vtn_base_type_array
:
1573 ralloc_steal(val
->constant
, elems
);
1574 val
->constant
->num_elements
= elem_count
;
1575 val
->constant
->elements
= elems
;
1579 vtn_fail("Result type of %s must be a composite type",
1580 spirv_op_to_string(opcode
));
1585 case SpvOpSpecConstantOp
: {
1586 SpvOp opcode
= get_specialization(b
, val
, w
[3]);
1588 case SpvOpVectorShuffle
: {
1589 struct vtn_value
*v0
= &b
->values
[w
[4]];
1590 struct vtn_value
*v1
= &b
->values
[w
[5]];
1592 vtn_assert(v0
->value_type
== vtn_value_type_constant
||
1593 v0
->value_type
== vtn_value_type_undef
);
1594 vtn_assert(v1
->value_type
== vtn_value_type_constant
||
1595 v1
->value_type
== vtn_value_type_undef
);
1597 unsigned len0
= glsl_get_vector_elements(v0
->type
->type
);
1598 unsigned len1
= glsl_get_vector_elements(v1
->type
->type
);
1600 vtn_assert(len0
+ len1
< 16);
1602 unsigned bit_size
= glsl_get_bit_size(val
->type
->type
);
1603 unsigned bit_size0
= glsl_get_bit_size(v0
->type
->type
);
1604 unsigned bit_size1
= glsl_get_bit_size(v1
->type
->type
);
1606 vtn_assert(bit_size
== bit_size0
&& bit_size
== bit_size1
);
1607 (void)bit_size0
; (void)bit_size1
;
1609 if (bit_size
== 64) {
1611 if (v0
->value_type
== vtn_value_type_constant
) {
1612 for (unsigned i
= 0; i
< len0
; i
++)
1613 u64
[i
] = v0
->constant
->values
[0].u64
[i
];
1615 if (v1
->value_type
== vtn_value_type_constant
) {
1616 for (unsigned i
= 0; i
< len1
; i
++)
1617 u64
[len0
+ i
] = v1
->constant
->values
[0].u64
[i
];
1620 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1621 uint32_t comp
= w
[i
+ 6];
1622 /* If component is not used, set the value to a known constant
1623 * to detect if it is wrongly used.
1625 if (comp
== (uint32_t)-1)
1626 val
->constant
->values
[0].u64
[j
] = 0xdeadbeefdeadbeef;
1628 val
->constant
->values
[0].u64
[j
] = u64
[comp
];
1631 /* This is for both 32-bit and 16-bit values */
1633 if (v0
->value_type
== vtn_value_type_constant
) {
1634 for (unsigned i
= 0; i
< len0
; i
++)
1635 u32
[i
] = v0
->constant
->values
[0].u32
[i
];
1637 if (v1
->value_type
== vtn_value_type_constant
) {
1638 for (unsigned i
= 0; i
< len1
; i
++)
1639 u32
[len0
+ i
] = v1
->constant
->values
[0].u32
[i
];
1642 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1643 uint32_t comp
= w
[i
+ 6];
1644 /* If component is not used, set the value to a known constant
1645 * to detect if it is wrongly used.
1647 if (comp
== (uint32_t)-1)
1648 val
->constant
->values
[0].u32
[j
] = 0xdeadbeef;
1650 val
->constant
->values
[0].u32
[j
] = u32
[comp
];
1656 case SpvOpCompositeExtract
:
1657 case SpvOpCompositeInsert
: {
1658 struct vtn_value
*comp
;
1659 unsigned deref_start
;
1660 struct nir_constant
**c
;
1661 if (opcode
== SpvOpCompositeExtract
) {
1662 comp
= vtn_value(b
, w
[4], vtn_value_type_constant
);
1664 c
= &comp
->constant
;
1666 comp
= vtn_value(b
, w
[5], vtn_value_type_constant
);
1668 val
->constant
= nir_constant_clone(comp
->constant
,
1675 const struct vtn_type
*type
= comp
->type
;
1676 for (unsigned i
= deref_start
; i
< count
; i
++) {
1677 vtn_fail_if(w
[i
] > type
->length
,
1678 "%uth index of %s is %u but the type has only "
1679 "%u elements", i
- deref_start
,
1680 spirv_op_to_string(opcode
), w
[i
], type
->length
);
1682 switch (type
->base_type
) {
1683 case vtn_base_type_vector
:
1685 type
= type
->array_element
;
1688 case vtn_base_type_matrix
:
1689 assert(col
== 0 && elem
== -1);
1692 type
= type
->array_element
;
1695 case vtn_base_type_array
:
1696 c
= &(*c
)->elements
[w
[i
]];
1697 type
= type
->array_element
;
1700 case vtn_base_type_struct
:
1701 c
= &(*c
)->elements
[w
[i
]];
1702 type
= type
->members
[w
[i
]];
1706 vtn_fail("%s must only index into composite types",
1707 spirv_op_to_string(opcode
));
1711 if (opcode
== SpvOpCompositeExtract
) {
1715 unsigned num_components
= type
->length
;
1716 unsigned bit_size
= glsl_get_bit_size(type
->type
);
1717 for (unsigned i
= 0; i
< num_components
; i
++)
1720 val
->constant
->values
[0].u64
[i
] = (*c
)->values
[col
].u64
[elem
+ i
];
1723 val
->constant
->values
[0].u32
[i
] = (*c
)->values
[col
].u32
[elem
+ i
];
1726 val
->constant
->values
[0].u16
[i
] = (*c
)->values
[col
].u16
[elem
+ i
];
1729 val
->constant
->values
[0].u8
[i
] = (*c
)->values
[col
].u8
[elem
+ i
];
1732 vtn_fail("Invalid SpvOpCompositeExtract bit size");
1736 struct vtn_value
*insert
=
1737 vtn_value(b
, w
[4], vtn_value_type_constant
);
1738 vtn_assert(insert
->type
== type
);
1740 *c
= insert
->constant
;
1742 unsigned num_components
= type
->length
;
1743 unsigned bit_size
= glsl_get_bit_size(type
->type
);
1744 for (unsigned i
= 0; i
< num_components
; i
++)
1747 (*c
)->values
[col
].u64
[elem
+ i
] = insert
->constant
->values
[0].u64
[i
];
1750 (*c
)->values
[col
].u32
[elem
+ i
] = insert
->constant
->values
[0].u32
[i
];
1753 (*c
)->values
[col
].u16
[elem
+ i
] = insert
->constant
->values
[0].u16
[i
];
1756 (*c
)->values
[col
].u8
[elem
+ i
] = insert
->constant
->values
[0].u8
[i
];
1759 vtn_fail("Invalid SpvOpCompositeInsert bit size");
1768 nir_alu_type dst_alu_type
= nir_get_nir_type_for_glsl_type(val
->type
->type
);
1769 nir_alu_type src_alu_type
= dst_alu_type
;
1770 unsigned num_components
= glsl_get_vector_elements(val
->type
->type
);
1773 vtn_assert(count
<= 7);
1778 /* We have a source in a conversion */
1780 nir_get_nir_type_for_glsl_type(
1781 vtn_value(b
, w
[4], vtn_value_type_constant
)->type
->type
);
1782 /* We use the bitsize of the conversion source to evaluate the opcode later */
1783 bit_size
= glsl_get_bit_size(
1784 vtn_value(b
, w
[4], vtn_value_type_constant
)->type
->type
);
1787 bit_size
= glsl_get_bit_size(val
->type
->type
);
1790 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(b
, opcode
, &swap
,
1791 nir_alu_type_get_type_size(src_alu_type
),
1792 nir_alu_type_get_type_size(dst_alu_type
));
1793 nir_const_value src
[4];
1795 for (unsigned i
= 0; i
< count
- 4; i
++) {
1797 vtn_value(b
, w
[4 + i
], vtn_value_type_constant
)->constant
;
1799 unsigned j
= swap
? 1 - i
: i
;
1800 src
[j
] = c
->values
[0];
1803 val
->constant
->values
[0] =
1804 nir_eval_const_opcode(op
, num_components
, bit_size
, src
);
1811 case SpvOpConstantNull
:
1812 val
->constant
= vtn_null_constant(b
, val
->type
->type
);
1815 case SpvOpConstantSampler
:
1816 vtn_fail("OpConstantSampler requires Kernel Capability");
1820 vtn_fail("Unhandled opcode");
1823 /* Now that we have the value, update the workgroup size if needed */
1824 vtn_foreach_decoration(b
, val
, handle_workgroup_size_decoration_cb
, NULL
);
1827 struct vtn_ssa_value
*
1828 vtn_create_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
1830 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
1833 if (!glsl_type_is_vector_or_scalar(type
)) {
1834 unsigned elems
= glsl_get_length(type
);
1835 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
1836 for (unsigned i
= 0; i
< elems
; i
++) {
1837 const struct glsl_type
*child_type
;
1839 switch (glsl_get_base_type(type
)) {
1841 case GLSL_TYPE_UINT
:
1842 case GLSL_TYPE_INT16
:
1843 case GLSL_TYPE_UINT16
:
1844 case GLSL_TYPE_UINT8
:
1845 case GLSL_TYPE_INT8
:
1846 case GLSL_TYPE_INT64
:
1847 case GLSL_TYPE_UINT64
:
1848 case GLSL_TYPE_BOOL
:
1849 case GLSL_TYPE_FLOAT
:
1850 case GLSL_TYPE_FLOAT16
:
1851 case GLSL_TYPE_DOUBLE
:
1852 child_type
= glsl_get_column_type(type
);
1854 case GLSL_TYPE_ARRAY
:
1855 child_type
= glsl_get_array_element(type
);
1857 case GLSL_TYPE_STRUCT
:
1858 child_type
= glsl_get_struct_field(type
, i
);
1861 vtn_fail("unkown base type");
1864 val
->elems
[i
] = vtn_create_ssa_value(b
, child_type
);
1872 vtn_tex_src(struct vtn_builder
*b
, unsigned index
, nir_tex_src_type type
)
1875 src
.src
= nir_src_for_ssa(vtn_ssa_value(b
, index
)->def
);
1876 src
.src_type
= type
;
1881 vtn_handle_texture(struct vtn_builder
*b
, SpvOp opcode
,
1882 const uint32_t *w
, unsigned count
)
1884 if (opcode
== SpvOpSampledImage
) {
1885 struct vtn_value
*val
=
1886 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
1887 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
1888 val
->sampled_image
->type
=
1889 vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1890 val
->sampled_image
->image
=
1891 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
1892 val
->sampled_image
->sampler
=
1893 vtn_value(b
, w
[4], vtn_value_type_pointer
)->pointer
;
1895 } else if (opcode
== SpvOpImage
) {
1896 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
1897 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[3]);
1898 if (src_val
->value_type
== vtn_value_type_sampled_image
) {
1899 val
->pointer
= src_val
->sampled_image
->image
;
1901 vtn_assert(src_val
->value_type
== vtn_value_type_pointer
);
1902 val
->pointer
= src_val
->pointer
;
1907 struct vtn_type
*ret_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1908 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1910 struct vtn_sampled_image sampled
;
1911 struct vtn_value
*sampled_val
= vtn_untyped_value(b
, w
[3]);
1912 if (sampled_val
->value_type
== vtn_value_type_sampled_image
) {
1913 sampled
= *sampled_val
->sampled_image
;
1915 vtn_assert(sampled_val
->value_type
== vtn_value_type_pointer
);
1916 sampled
.type
= sampled_val
->pointer
->type
;
1917 sampled
.image
= NULL
;
1918 sampled
.sampler
= sampled_val
->pointer
;
1921 const struct glsl_type
*image_type
= sampled
.type
->type
;
1922 const enum glsl_sampler_dim sampler_dim
= glsl_get_sampler_dim(image_type
);
1923 const bool is_array
= glsl_sampler_type_is_array(image_type
);
1925 /* Figure out the base texture operation */
1928 case SpvOpImageSampleImplicitLod
:
1929 case SpvOpImageSampleDrefImplicitLod
:
1930 case SpvOpImageSampleProjImplicitLod
:
1931 case SpvOpImageSampleProjDrefImplicitLod
:
1932 texop
= nir_texop_tex
;
1935 case SpvOpImageSampleExplicitLod
:
1936 case SpvOpImageSampleDrefExplicitLod
:
1937 case SpvOpImageSampleProjExplicitLod
:
1938 case SpvOpImageSampleProjDrefExplicitLod
:
1939 texop
= nir_texop_txl
;
1942 case SpvOpImageFetch
:
1943 if (glsl_get_sampler_dim(image_type
) == GLSL_SAMPLER_DIM_MS
) {
1944 texop
= nir_texop_txf_ms
;
1946 texop
= nir_texop_txf
;
1950 case SpvOpImageGather
:
1951 case SpvOpImageDrefGather
:
1952 texop
= nir_texop_tg4
;
1955 case SpvOpImageQuerySizeLod
:
1956 case SpvOpImageQuerySize
:
1957 texop
= nir_texop_txs
;
1960 case SpvOpImageQueryLod
:
1961 texop
= nir_texop_lod
;
1964 case SpvOpImageQueryLevels
:
1965 texop
= nir_texop_query_levels
;
1968 case SpvOpImageQuerySamples
:
1969 texop
= nir_texop_texture_samples
;
1973 vtn_fail("Unhandled opcode");
1976 nir_tex_src srcs
[10]; /* 10 should be enough */
1977 nir_tex_src
*p
= srcs
;
1979 nir_deref_instr
*sampler
= vtn_pointer_to_deref(b
, sampled
.sampler
);
1980 nir_deref_instr
*texture
=
1981 sampled
.image
? vtn_pointer_to_deref(b
, sampled
.image
) : sampler
;
1983 p
->src
= nir_src_for_ssa(&texture
->dest
.ssa
);
1984 p
->src_type
= nir_tex_src_texture_deref
;
1993 /* These operations require a sampler */
1994 p
->src
= nir_src_for_ssa(&sampler
->dest
.ssa
);
1995 p
->src_type
= nir_tex_src_sampler_deref
;
1999 case nir_texop_txf_ms
:
2002 case nir_texop_query_levels
:
2003 case nir_texop_texture_samples
:
2004 case nir_texop_samples_identical
:
2007 case nir_texop_txf_ms_mcs
:
2008 vtn_fail("unexpected nir_texop_txf_ms_mcs");
2013 struct nir_ssa_def
*coord
;
2014 unsigned coord_components
;
2016 case SpvOpImageSampleImplicitLod
:
2017 case SpvOpImageSampleExplicitLod
:
2018 case SpvOpImageSampleDrefImplicitLod
:
2019 case SpvOpImageSampleDrefExplicitLod
:
2020 case SpvOpImageSampleProjImplicitLod
:
2021 case SpvOpImageSampleProjExplicitLod
:
2022 case SpvOpImageSampleProjDrefImplicitLod
:
2023 case SpvOpImageSampleProjDrefExplicitLod
:
2024 case SpvOpImageFetch
:
2025 case SpvOpImageGather
:
2026 case SpvOpImageDrefGather
:
2027 case SpvOpImageQueryLod
: {
2028 /* All these types have the coordinate as their first real argument */
2029 switch (sampler_dim
) {
2030 case GLSL_SAMPLER_DIM_1D
:
2031 case GLSL_SAMPLER_DIM_BUF
:
2032 coord_components
= 1;
2034 case GLSL_SAMPLER_DIM_2D
:
2035 case GLSL_SAMPLER_DIM_RECT
:
2036 case GLSL_SAMPLER_DIM_MS
:
2037 coord_components
= 2;
2039 case GLSL_SAMPLER_DIM_3D
:
2040 case GLSL_SAMPLER_DIM_CUBE
:
2041 coord_components
= 3;
2044 vtn_fail("Invalid sampler type");
2047 if (is_array
&& texop
!= nir_texop_lod
)
2050 coord
= vtn_ssa_value(b
, w
[idx
++])->def
;
2051 p
->src
= nir_src_for_ssa(nir_channels(&b
->nb
, coord
,
2052 (1 << coord_components
) - 1));
2053 p
->src_type
= nir_tex_src_coord
;
2060 coord_components
= 0;
2065 case SpvOpImageSampleProjImplicitLod
:
2066 case SpvOpImageSampleProjExplicitLod
:
2067 case SpvOpImageSampleProjDrefImplicitLod
:
2068 case SpvOpImageSampleProjDrefExplicitLod
:
2069 /* These have the projector as the last coordinate component */
2070 p
->src
= nir_src_for_ssa(nir_channel(&b
->nb
, coord
, coord_components
));
2071 p
->src_type
= nir_tex_src_projector
;
2079 bool is_shadow
= false;
2080 unsigned gather_component
= 0;
2082 case SpvOpImageSampleDrefImplicitLod
:
2083 case SpvOpImageSampleDrefExplicitLod
:
2084 case SpvOpImageSampleProjDrefImplicitLod
:
2085 case SpvOpImageSampleProjDrefExplicitLod
:
2086 case SpvOpImageDrefGather
:
2087 /* These all have an explicit depth value as their next source */
2089 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_comparator
);
2092 case SpvOpImageGather
:
2093 /* This has a component as its next source */
2095 vtn_value(b
, w
[idx
++], vtn_value_type_constant
)->constant
->values
[0].u32
[0];
2102 /* For OpImageQuerySizeLod, we always have an LOD */
2103 if (opcode
== SpvOpImageQuerySizeLod
)
2104 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
2106 /* Now we need to handle some number of optional arguments */
2107 const struct vtn_ssa_value
*gather_offsets
= NULL
;
2109 uint32_t operands
= w
[idx
++];
2111 if (operands
& SpvImageOperandsBiasMask
) {
2112 vtn_assert(texop
== nir_texop_tex
);
2113 texop
= nir_texop_txb
;
2114 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_bias
);
2117 if (operands
& SpvImageOperandsLodMask
) {
2118 vtn_assert(texop
== nir_texop_txl
|| texop
== nir_texop_txf
||
2119 texop
== nir_texop_txs
);
2120 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
2123 if (operands
& SpvImageOperandsGradMask
) {
2124 vtn_assert(texop
== nir_texop_txl
);
2125 texop
= nir_texop_txd
;
2126 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddx
);
2127 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddy
);
2130 if (operands
& SpvImageOperandsOffsetMask
||
2131 operands
& SpvImageOperandsConstOffsetMask
)
2132 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_offset
);
2134 if (operands
& SpvImageOperandsConstOffsetsMask
) {
2135 nir_tex_src none
= {0};
2136 gather_offsets
= vtn_ssa_value(b
, w
[idx
++]);
2140 if (operands
& SpvImageOperandsSampleMask
) {
2141 vtn_assert(texop
== nir_texop_txf_ms
);
2142 texop
= nir_texop_txf_ms
;
2143 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ms_index
);
2146 /* We should have now consumed exactly all of the arguments */
2147 vtn_assert(idx
== count
);
2149 nir_tex_instr
*instr
= nir_tex_instr_create(b
->shader
, p
- srcs
);
2152 memcpy(instr
->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
2154 instr
->coord_components
= coord_components
;
2155 instr
->sampler_dim
= sampler_dim
;
2156 instr
->is_array
= is_array
;
2157 instr
->is_shadow
= is_shadow
;
2158 instr
->is_new_style_shadow
=
2159 is_shadow
&& glsl_get_components(ret_type
->type
) == 1;
2160 instr
->component
= gather_component
;
2162 switch (glsl_get_sampler_result_type(image_type
)) {
2163 case GLSL_TYPE_FLOAT
: instr
->dest_type
= nir_type_float
; break;
2164 case GLSL_TYPE_INT
: instr
->dest_type
= nir_type_int
; break;
2165 case GLSL_TYPE_UINT
: instr
->dest_type
= nir_type_uint
; break;
2166 case GLSL_TYPE_BOOL
: instr
->dest_type
= nir_type_bool
; break;
2168 vtn_fail("Invalid base type for sampler result");
2171 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
2172 nir_tex_instr_dest_size(instr
), 32, NULL
);
2174 vtn_assert(glsl_get_vector_elements(ret_type
->type
) ==
2175 nir_tex_instr_dest_size(instr
));
2178 nir_instr
*instruction
;
2179 if (gather_offsets
) {
2180 vtn_assert(glsl_get_base_type(gather_offsets
->type
) == GLSL_TYPE_ARRAY
);
2181 vtn_assert(glsl_get_length(gather_offsets
->type
) == 4);
2182 nir_tex_instr
*instrs
[4] = {instr
, NULL
, NULL
, NULL
};
2184 /* Copy the current instruction 4x */
2185 for (uint32_t i
= 1; i
< 4; i
++) {
2186 instrs
[i
] = nir_tex_instr_create(b
->shader
, instr
->num_srcs
);
2187 instrs
[i
]->op
= instr
->op
;
2188 instrs
[i
]->coord_components
= instr
->coord_components
;
2189 instrs
[i
]->sampler_dim
= instr
->sampler_dim
;
2190 instrs
[i
]->is_array
= instr
->is_array
;
2191 instrs
[i
]->is_shadow
= instr
->is_shadow
;
2192 instrs
[i
]->is_new_style_shadow
= instr
->is_new_style_shadow
;
2193 instrs
[i
]->component
= instr
->component
;
2194 instrs
[i
]->dest_type
= instr
->dest_type
;
2196 memcpy(instrs
[i
]->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
2198 nir_ssa_dest_init(&instrs
[i
]->instr
, &instrs
[i
]->dest
,
2199 nir_tex_instr_dest_size(instr
), 32, NULL
);
2202 /* Fill in the last argument with the offset from the passed in offsets
2203 * and insert the instruction into the stream.
2205 for (uint32_t i
= 0; i
< 4; i
++) {
2207 src
.src
= nir_src_for_ssa(gather_offsets
->elems
[i
]->def
);
2208 src
.src_type
= nir_tex_src_offset
;
2209 instrs
[i
]->src
[instrs
[i
]->num_srcs
- 1] = src
;
2210 nir_builder_instr_insert(&b
->nb
, &instrs
[i
]->instr
);
2213 /* Combine the results of the 4 instructions by taking their .w
2216 nir_alu_instr
*vec4
= nir_alu_instr_create(b
->shader
, nir_op_vec4
);
2217 nir_ssa_dest_init(&vec4
->instr
, &vec4
->dest
.dest
, 4, 32, NULL
);
2218 vec4
->dest
.write_mask
= 0xf;
2219 for (uint32_t i
= 0; i
< 4; i
++) {
2220 vec4
->src
[i
].src
= nir_src_for_ssa(&instrs
[i
]->dest
.ssa
);
2221 vec4
->src
[i
].swizzle
[0] = 3;
2223 def
= &vec4
->dest
.dest
.ssa
;
2224 instruction
= &vec4
->instr
;
2226 def
= &instr
->dest
.ssa
;
2227 instruction
= &instr
->instr
;
2230 val
->ssa
= vtn_create_ssa_value(b
, ret_type
->type
);
2231 val
->ssa
->def
= def
;
2233 nir_builder_instr_insert(&b
->nb
, instruction
);
2237 fill_common_atomic_sources(struct vtn_builder
*b
, SpvOp opcode
,
2238 const uint32_t *w
, nir_src
*src
)
2241 case SpvOpAtomicIIncrement
:
2242 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
2245 case SpvOpAtomicIDecrement
:
2246 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
2249 case SpvOpAtomicISub
:
2251 nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
2254 case SpvOpAtomicCompareExchange
:
2255 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[8])->def
);
2256 src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
2259 case SpvOpAtomicExchange
:
2260 case SpvOpAtomicIAdd
:
2261 case SpvOpAtomicSMin
:
2262 case SpvOpAtomicUMin
:
2263 case SpvOpAtomicSMax
:
2264 case SpvOpAtomicUMax
:
2265 case SpvOpAtomicAnd
:
2267 case SpvOpAtomicXor
:
2268 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
2272 vtn_fail("Invalid SPIR-V atomic");
2276 static nir_ssa_def
*
2277 get_image_coord(struct vtn_builder
*b
, uint32_t value
)
2279 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, value
);
2281 /* The image_load_store intrinsics assume a 4-dim coordinate */
2282 unsigned dim
= glsl_get_vector_elements(coord
->type
);
2283 unsigned swizzle
[4];
2284 for (unsigned i
= 0; i
< 4; i
++)
2285 swizzle
[i
] = MIN2(i
, dim
- 1);
2287 return nir_swizzle(&b
->nb
, coord
->def
, swizzle
, 4, false);
2290 static nir_ssa_def
*
2291 expand_to_vec4(nir_builder
*b
, nir_ssa_def
*value
)
2293 if (value
->num_components
== 4)
2297 for (unsigned i
= 0; i
< 4; i
++)
2298 swiz
[i
] = i
< value
->num_components
? i
: 0;
2299 return nir_swizzle(b
, value
, swiz
, 4, false);
2303 vtn_handle_image(struct vtn_builder
*b
, SpvOp opcode
,
2304 const uint32_t *w
, unsigned count
)
2306 /* Just get this one out of the way */
2307 if (opcode
== SpvOpImageTexelPointer
) {
2308 struct vtn_value
*val
=
2309 vtn_push_value(b
, w
[2], vtn_value_type_image_pointer
);
2310 val
->image
= ralloc(b
, struct vtn_image_pointer
);
2312 val
->image
->image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2313 val
->image
->coord
= get_image_coord(b
, w
[4]);
2314 val
->image
->sample
= vtn_ssa_value(b
, w
[5])->def
;
2318 struct vtn_image_pointer image
;
2321 case SpvOpAtomicExchange
:
2322 case SpvOpAtomicCompareExchange
:
2323 case SpvOpAtomicCompareExchangeWeak
:
2324 case SpvOpAtomicIIncrement
:
2325 case SpvOpAtomicIDecrement
:
2326 case SpvOpAtomicIAdd
:
2327 case SpvOpAtomicISub
:
2328 case SpvOpAtomicLoad
:
2329 case SpvOpAtomicSMin
:
2330 case SpvOpAtomicUMin
:
2331 case SpvOpAtomicSMax
:
2332 case SpvOpAtomicUMax
:
2333 case SpvOpAtomicAnd
:
2335 case SpvOpAtomicXor
:
2336 image
= *vtn_value(b
, w
[3], vtn_value_type_image_pointer
)->image
;
2339 case SpvOpAtomicStore
:
2340 image
= *vtn_value(b
, w
[1], vtn_value_type_image_pointer
)->image
;
2343 case SpvOpImageQuerySize
:
2344 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2346 image
.sample
= NULL
;
2349 case SpvOpImageRead
:
2350 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2351 image
.coord
= get_image_coord(b
, w
[4]);
2353 if (count
> 5 && (w
[5] & SpvImageOperandsSampleMask
)) {
2354 vtn_assert(w
[5] == SpvImageOperandsSampleMask
);
2355 image
.sample
= vtn_ssa_value(b
, w
[6])->def
;
2357 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2361 case SpvOpImageWrite
:
2362 image
.image
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
2363 image
.coord
= get_image_coord(b
, w
[2]);
2367 if (count
> 4 && (w
[4] & SpvImageOperandsSampleMask
)) {
2368 vtn_assert(w
[4] == SpvImageOperandsSampleMask
);
2369 image
.sample
= vtn_ssa_value(b
, w
[5])->def
;
2371 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2376 vtn_fail("Invalid image opcode");
2379 nir_intrinsic_op op
;
2381 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_deref_##N; break;
2382 OP(ImageQuerySize
, size
)
2384 OP(ImageWrite
, store
)
2385 OP(AtomicLoad
, load
)
2386 OP(AtomicStore
, store
)
2387 OP(AtomicExchange
, atomic_exchange
)
2388 OP(AtomicCompareExchange
, atomic_comp_swap
)
2389 OP(AtomicIIncrement
, atomic_add
)
2390 OP(AtomicIDecrement
, atomic_add
)
2391 OP(AtomicIAdd
, atomic_add
)
2392 OP(AtomicISub
, atomic_add
)
2393 OP(AtomicSMin
, atomic_min
)
2394 OP(AtomicUMin
, atomic_min
)
2395 OP(AtomicSMax
, atomic_max
)
2396 OP(AtomicUMax
, atomic_max
)
2397 OP(AtomicAnd
, atomic_and
)
2398 OP(AtomicOr
, atomic_or
)
2399 OP(AtomicXor
, atomic_xor
)
2402 vtn_fail("Invalid image opcode");
2405 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
2407 nir_deref_instr
*image_deref
= vtn_pointer_to_deref(b
, image
.image
);
2408 intrin
->src
[0] = nir_src_for_ssa(&image_deref
->dest
.ssa
);
2410 /* ImageQuerySize doesn't take any extra parameters */
2411 if (opcode
!= SpvOpImageQuerySize
) {
2412 /* The image coordinate is always 4 components but we may not have that
2413 * many. Swizzle to compensate.
2415 intrin
->src
[1] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, image
.coord
));
2416 intrin
->src
[2] = nir_src_for_ssa(image
.sample
);
2420 case SpvOpAtomicLoad
:
2421 case SpvOpImageQuerySize
:
2422 case SpvOpImageRead
:
2424 case SpvOpAtomicStore
:
2425 case SpvOpImageWrite
: {
2426 const uint32_t value_id
= opcode
== SpvOpAtomicStore
? w
[4] : w
[3];
2427 nir_ssa_def
*value
= vtn_ssa_value(b
, value_id
)->def
;
2428 /* nir_intrinsic_image_deref_store always takes a vec4 value */
2429 assert(op
== nir_intrinsic_image_deref_store
);
2430 intrin
->num_components
= 4;
2431 intrin
->src
[3] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, value
));
2435 case SpvOpAtomicCompareExchange
:
2436 case SpvOpAtomicIIncrement
:
2437 case SpvOpAtomicIDecrement
:
2438 case SpvOpAtomicExchange
:
2439 case SpvOpAtomicIAdd
:
2440 case SpvOpAtomicISub
:
2441 case SpvOpAtomicSMin
:
2442 case SpvOpAtomicUMin
:
2443 case SpvOpAtomicSMax
:
2444 case SpvOpAtomicUMax
:
2445 case SpvOpAtomicAnd
:
2447 case SpvOpAtomicXor
:
2448 fill_common_atomic_sources(b
, opcode
, w
, &intrin
->src
[3]);
2452 vtn_fail("Invalid image opcode");
2455 if (opcode
!= SpvOpImageWrite
&& opcode
!= SpvOpAtomicStore
) {
2456 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2457 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2459 unsigned dest_components
= glsl_get_vector_elements(type
->type
);
2460 intrin
->num_components
= nir_intrinsic_infos
[op
].dest_components
;
2461 if (intrin
->num_components
== 0)
2462 intrin
->num_components
= dest_components
;
2464 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
,
2465 intrin
->num_components
, 32, NULL
);
2467 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2469 nir_ssa_def
*result
= &intrin
->dest
.ssa
;
2470 if (intrin
->num_components
!= dest_components
)
2471 result
= nir_channels(&b
->nb
, result
, (1 << dest_components
) - 1);
2473 val
->ssa
= vtn_create_ssa_value(b
, type
->type
);
2474 val
->ssa
->def
= result
;
2476 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2480 static nir_intrinsic_op
2481 get_ssbo_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2484 case SpvOpAtomicLoad
: return nir_intrinsic_load_ssbo
;
2485 case SpvOpAtomicStore
: return nir_intrinsic_store_ssbo
;
2486 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
2487 OP(AtomicExchange
, atomic_exchange
)
2488 OP(AtomicCompareExchange
, atomic_comp_swap
)
2489 OP(AtomicIIncrement
, atomic_add
)
2490 OP(AtomicIDecrement
, atomic_add
)
2491 OP(AtomicIAdd
, atomic_add
)
2492 OP(AtomicISub
, atomic_add
)
2493 OP(AtomicSMin
, atomic_imin
)
2494 OP(AtomicUMin
, atomic_umin
)
2495 OP(AtomicSMax
, atomic_imax
)
2496 OP(AtomicUMax
, atomic_umax
)
2497 OP(AtomicAnd
, atomic_and
)
2498 OP(AtomicOr
, atomic_or
)
2499 OP(AtomicXor
, atomic_xor
)
2502 vtn_fail("Invalid SSBO atomic");
2506 static nir_intrinsic_op
2507 get_uniform_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2510 #define OP(S, N) case SpvOp##S: return nir_intrinsic_atomic_counter_ ##N;
2511 OP(AtomicLoad
, read_deref
)
2512 OP(AtomicExchange
, exchange
)
2513 OP(AtomicCompareExchange
, comp_swap
)
2514 OP(AtomicIIncrement
, inc_deref
)
2515 OP(AtomicIDecrement
, post_dec_deref
)
2516 OP(AtomicIAdd
, add_deref
)
2517 OP(AtomicISub
, add_deref
)
2518 OP(AtomicUMin
, min_deref
)
2519 OP(AtomicUMax
, max_deref
)
2520 OP(AtomicAnd
, and_deref
)
2521 OP(AtomicOr
, or_deref
)
2522 OP(AtomicXor
, xor_deref
)
2525 /* We left the following out: AtomicStore, AtomicSMin and
2526 * AtomicSmax. Right now there are not nir intrinsics for them. At this
2527 * moment Atomic Counter support is needed for ARB_spirv support, so is
2528 * only need to support GLSL Atomic Counters that are uints and don't
2529 * allow direct storage.
2531 unreachable("Invalid uniform atomic");
2535 static nir_intrinsic_op
2536 get_shared_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2539 case SpvOpAtomicLoad
: return nir_intrinsic_load_shared
;
2540 case SpvOpAtomicStore
: return nir_intrinsic_store_shared
;
2541 #define OP(S, N) case SpvOp##S: return nir_intrinsic_shared_##N;
2542 OP(AtomicExchange
, atomic_exchange
)
2543 OP(AtomicCompareExchange
, atomic_comp_swap
)
2544 OP(AtomicIIncrement
, atomic_add
)
2545 OP(AtomicIDecrement
, atomic_add
)
2546 OP(AtomicIAdd
, atomic_add
)
2547 OP(AtomicISub
, atomic_add
)
2548 OP(AtomicSMin
, atomic_imin
)
2549 OP(AtomicUMin
, atomic_umin
)
2550 OP(AtomicSMax
, atomic_imax
)
2551 OP(AtomicUMax
, atomic_umax
)
2552 OP(AtomicAnd
, atomic_and
)
2553 OP(AtomicOr
, atomic_or
)
2554 OP(AtomicXor
, atomic_xor
)
2557 vtn_fail("Invalid shared atomic");
2561 static nir_intrinsic_op
2562 get_deref_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2565 case SpvOpAtomicLoad
: return nir_intrinsic_load_deref
;
2566 case SpvOpAtomicStore
: return nir_intrinsic_store_deref
;
2567 #define OP(S, N) case SpvOp##S: return nir_intrinsic_deref_##N;
2568 OP(AtomicExchange
, atomic_exchange
)
2569 OP(AtomicCompareExchange
, atomic_comp_swap
)
2570 OP(AtomicIIncrement
, atomic_add
)
2571 OP(AtomicIDecrement
, atomic_add
)
2572 OP(AtomicIAdd
, atomic_add
)
2573 OP(AtomicISub
, atomic_add
)
2574 OP(AtomicSMin
, atomic_imin
)
2575 OP(AtomicUMin
, atomic_umin
)
2576 OP(AtomicSMax
, atomic_imax
)
2577 OP(AtomicUMax
, atomic_umax
)
2578 OP(AtomicAnd
, atomic_and
)
2579 OP(AtomicOr
, atomic_or
)
2580 OP(AtomicXor
, atomic_xor
)
2583 vtn_fail("Invalid shared atomic");
2588 * Handles shared atomics, ssbo atomics and atomic counters.
2591 vtn_handle_atomics(struct vtn_builder
*b
, SpvOp opcode
,
2592 const uint32_t *w
, unsigned count
)
2594 struct vtn_pointer
*ptr
;
2595 nir_intrinsic_instr
*atomic
;
2598 case SpvOpAtomicLoad
:
2599 case SpvOpAtomicExchange
:
2600 case SpvOpAtomicCompareExchange
:
2601 case SpvOpAtomicCompareExchangeWeak
:
2602 case SpvOpAtomicIIncrement
:
2603 case SpvOpAtomicIDecrement
:
2604 case SpvOpAtomicIAdd
:
2605 case SpvOpAtomicISub
:
2606 case SpvOpAtomicSMin
:
2607 case SpvOpAtomicUMin
:
2608 case SpvOpAtomicSMax
:
2609 case SpvOpAtomicUMax
:
2610 case SpvOpAtomicAnd
:
2612 case SpvOpAtomicXor
:
2613 ptr
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2616 case SpvOpAtomicStore
:
2617 ptr
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
2621 vtn_fail("Invalid SPIR-V atomic");
2625 SpvScope scope = w[4];
2626 SpvMemorySemanticsMask semantics = w[5];
2629 /* uniform as "atomic counter uniform" */
2630 if (ptr
->mode
== vtn_variable_mode_uniform
) {
2631 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
2632 const struct glsl_type
*deref_type
= deref
->type
;
2633 nir_intrinsic_op op
= get_uniform_nir_atomic_op(b
, opcode
);
2634 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2635 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
2637 /* SSBO needs to initialize index/offset. In this case we don't need to,
2638 * as that info is already stored on the ptr->var->var nir_variable (see
2639 * vtn_create_variable)
2643 case SpvOpAtomicLoad
:
2644 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2647 case SpvOpAtomicStore
:
2648 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2649 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2652 case SpvOpAtomicExchange
:
2653 case SpvOpAtomicCompareExchange
:
2654 case SpvOpAtomicCompareExchangeWeak
:
2655 case SpvOpAtomicIIncrement
:
2656 case SpvOpAtomicIDecrement
:
2657 case SpvOpAtomicIAdd
:
2658 case SpvOpAtomicISub
:
2659 case SpvOpAtomicSMin
:
2660 case SpvOpAtomicUMin
:
2661 case SpvOpAtomicSMax
:
2662 case SpvOpAtomicUMax
:
2663 case SpvOpAtomicAnd
:
2665 case SpvOpAtomicXor
:
2666 /* Nothing: we don't need to call fill_common_atomic_sources here, as
2667 * atomic counter uniforms doesn't have sources
2672 unreachable("Invalid SPIR-V atomic");
2675 } else if (ptr
->mode
== vtn_variable_mode_workgroup
&&
2676 !b
->options
->lower_workgroup_access_to_offsets
) {
2677 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
2678 const struct glsl_type
*deref_type
= deref
->type
;
2679 nir_intrinsic_op op
= get_deref_nir_atomic_op(b
, opcode
);
2680 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2681 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
2684 case SpvOpAtomicLoad
:
2685 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2688 case SpvOpAtomicStore
:
2689 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2690 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2691 atomic
->src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2694 case SpvOpAtomicExchange
:
2695 case SpvOpAtomicCompareExchange
:
2696 case SpvOpAtomicCompareExchangeWeak
:
2697 case SpvOpAtomicIIncrement
:
2698 case SpvOpAtomicIDecrement
:
2699 case SpvOpAtomicIAdd
:
2700 case SpvOpAtomicISub
:
2701 case SpvOpAtomicSMin
:
2702 case SpvOpAtomicUMin
:
2703 case SpvOpAtomicSMax
:
2704 case SpvOpAtomicUMax
:
2705 case SpvOpAtomicAnd
:
2707 case SpvOpAtomicXor
:
2708 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[1]);
2712 vtn_fail("Invalid SPIR-V atomic");
2716 nir_ssa_def
*offset
, *index
;
2717 offset
= vtn_pointer_to_offset(b
, ptr
, &index
);
2719 nir_intrinsic_op op
;
2720 if (ptr
->mode
== vtn_variable_mode_ssbo
) {
2721 op
= get_ssbo_nir_atomic_op(b
, opcode
);
2723 vtn_assert(ptr
->mode
== vtn_variable_mode_workgroup
&&
2724 b
->options
->lower_workgroup_access_to_offsets
);
2725 op
= get_shared_nir_atomic_op(b
, opcode
);
2728 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2732 case SpvOpAtomicLoad
:
2733 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
2734 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2735 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2736 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2739 case SpvOpAtomicStore
:
2740 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
2741 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2742 atomic
->src
[src
++] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2743 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2744 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2745 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2748 case SpvOpAtomicExchange
:
2749 case SpvOpAtomicCompareExchange
:
2750 case SpvOpAtomicCompareExchangeWeak
:
2751 case SpvOpAtomicIIncrement
:
2752 case SpvOpAtomicIDecrement
:
2753 case SpvOpAtomicIAdd
:
2754 case SpvOpAtomicISub
:
2755 case SpvOpAtomicSMin
:
2756 case SpvOpAtomicUMin
:
2757 case SpvOpAtomicSMax
:
2758 case SpvOpAtomicUMax
:
2759 case SpvOpAtomicAnd
:
2761 case SpvOpAtomicXor
:
2762 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2763 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2764 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2765 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[src
]);
2769 vtn_fail("Invalid SPIR-V atomic");
2773 if (opcode
!= SpvOpAtomicStore
) {
2774 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2776 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
,
2777 glsl_get_vector_elements(type
->type
),
2778 glsl_get_bit_size(type
->type
), NULL
);
2780 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2781 val
->ssa
= rzalloc(b
, struct vtn_ssa_value
);
2782 val
->ssa
->def
= &atomic
->dest
.ssa
;
2783 val
->ssa
->type
= type
->type
;
2786 nir_builder_instr_insert(&b
->nb
, &atomic
->instr
);
2789 static nir_alu_instr
*
2790 create_vec(struct vtn_builder
*b
, unsigned num_components
, unsigned bit_size
)
2793 switch (num_components
) {
2794 case 1: op
= nir_op_imov
; break;
2795 case 2: op
= nir_op_vec2
; break;
2796 case 3: op
= nir_op_vec3
; break;
2797 case 4: op
= nir_op_vec4
; break;
2798 default: vtn_fail("bad vector size");
2801 nir_alu_instr
*vec
= nir_alu_instr_create(b
->shader
, op
);
2802 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, num_components
,
2804 vec
->dest
.write_mask
= (1 << num_components
) - 1;
2809 struct vtn_ssa_value
*
2810 vtn_ssa_transpose(struct vtn_builder
*b
, struct vtn_ssa_value
*src
)
2812 if (src
->transposed
)
2813 return src
->transposed
;
2815 struct vtn_ssa_value
*dest
=
2816 vtn_create_ssa_value(b
, glsl_transposed_type(src
->type
));
2818 for (unsigned i
= 0; i
< glsl_get_matrix_columns(dest
->type
); i
++) {
2819 nir_alu_instr
*vec
= create_vec(b
, glsl_get_matrix_columns(src
->type
),
2820 glsl_get_bit_size(src
->type
));
2821 if (glsl_type_is_vector_or_scalar(src
->type
)) {
2822 vec
->src
[0].src
= nir_src_for_ssa(src
->def
);
2823 vec
->src
[0].swizzle
[0] = i
;
2825 for (unsigned j
= 0; j
< glsl_get_matrix_columns(src
->type
); j
++) {
2826 vec
->src
[j
].src
= nir_src_for_ssa(src
->elems
[j
]->def
);
2827 vec
->src
[j
].swizzle
[0] = i
;
2830 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2831 dest
->elems
[i
]->def
= &vec
->dest
.dest
.ssa
;
2834 dest
->transposed
= src
;
2840 vtn_vector_extract(struct vtn_builder
*b
, nir_ssa_def
*src
, unsigned index
)
2842 return nir_channel(&b
->nb
, src
, index
);
2846 vtn_vector_insert(struct vtn_builder
*b
, nir_ssa_def
*src
, nir_ssa_def
*insert
,
2849 nir_alu_instr
*vec
= create_vec(b
, src
->num_components
,
2852 for (unsigned i
= 0; i
< src
->num_components
; i
++) {
2854 vec
->src
[i
].src
= nir_src_for_ssa(insert
);
2856 vec
->src
[i
].src
= nir_src_for_ssa(src
);
2857 vec
->src
[i
].swizzle
[0] = i
;
2861 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2863 return &vec
->dest
.dest
.ssa
;
2867 vtn_vector_extract_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
2870 nir_ssa_def
*dest
= vtn_vector_extract(b
, src
, 0);
2871 for (unsigned i
= 1; i
< src
->num_components
; i
++)
2872 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
2873 vtn_vector_extract(b
, src
, i
), dest
);
2879 vtn_vector_insert_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
2880 nir_ssa_def
*insert
, nir_ssa_def
*index
)
2882 nir_ssa_def
*dest
= vtn_vector_insert(b
, src
, insert
, 0);
2883 for (unsigned i
= 1; i
< src
->num_components
; i
++)
2884 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
2885 vtn_vector_insert(b
, src
, insert
, i
), dest
);
2890 static nir_ssa_def
*
2891 vtn_vector_shuffle(struct vtn_builder
*b
, unsigned num_components
,
2892 nir_ssa_def
*src0
, nir_ssa_def
*src1
,
2893 const uint32_t *indices
)
2895 nir_alu_instr
*vec
= create_vec(b
, num_components
, src0
->bit_size
);
2897 for (unsigned i
= 0; i
< num_components
; i
++) {
2898 uint32_t index
= indices
[i
];
2899 if (index
== 0xffffffff) {
2901 nir_src_for_ssa(nir_ssa_undef(&b
->nb
, 1, src0
->bit_size
));
2902 } else if (index
< src0
->num_components
) {
2903 vec
->src
[i
].src
= nir_src_for_ssa(src0
);
2904 vec
->src
[i
].swizzle
[0] = index
;
2906 vec
->src
[i
].src
= nir_src_for_ssa(src1
);
2907 vec
->src
[i
].swizzle
[0] = index
- src0
->num_components
;
2911 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2913 return &vec
->dest
.dest
.ssa
;
2917 * Concatentates a number of vectors/scalars together to produce a vector
2919 static nir_ssa_def
*
2920 vtn_vector_construct(struct vtn_builder
*b
, unsigned num_components
,
2921 unsigned num_srcs
, nir_ssa_def
**srcs
)
2923 nir_alu_instr
*vec
= create_vec(b
, num_components
, srcs
[0]->bit_size
);
2925 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2927 * "When constructing a vector, there must be at least two Constituent
2930 vtn_assert(num_srcs
>= 2);
2932 unsigned dest_idx
= 0;
2933 for (unsigned i
= 0; i
< num_srcs
; i
++) {
2934 nir_ssa_def
*src
= srcs
[i
];
2935 vtn_assert(dest_idx
+ src
->num_components
<= num_components
);
2936 for (unsigned j
= 0; j
< src
->num_components
; j
++) {
2937 vec
->src
[dest_idx
].src
= nir_src_for_ssa(src
);
2938 vec
->src
[dest_idx
].swizzle
[0] = j
;
2943 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2945 * "When constructing a vector, the total number of components in all
2946 * the operands must equal the number of components in Result Type."
2948 vtn_assert(dest_idx
== num_components
);
2950 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2952 return &vec
->dest
.dest
.ssa
;
2955 static struct vtn_ssa_value
*
2956 vtn_composite_copy(void *mem_ctx
, struct vtn_ssa_value
*src
)
2958 struct vtn_ssa_value
*dest
= rzalloc(mem_ctx
, struct vtn_ssa_value
);
2959 dest
->type
= src
->type
;
2961 if (glsl_type_is_vector_or_scalar(src
->type
)) {
2962 dest
->def
= src
->def
;
2964 unsigned elems
= glsl_get_length(src
->type
);
2966 dest
->elems
= ralloc_array(mem_ctx
, struct vtn_ssa_value
*, elems
);
2967 for (unsigned i
= 0; i
< elems
; i
++)
2968 dest
->elems
[i
] = vtn_composite_copy(mem_ctx
, src
->elems
[i
]);
2974 static struct vtn_ssa_value
*
2975 vtn_composite_insert(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
2976 struct vtn_ssa_value
*insert
, const uint32_t *indices
,
2977 unsigned num_indices
)
2979 struct vtn_ssa_value
*dest
= vtn_composite_copy(b
, src
);
2981 struct vtn_ssa_value
*cur
= dest
;
2983 for (i
= 0; i
< num_indices
- 1; i
++) {
2984 cur
= cur
->elems
[indices
[i
]];
2987 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
2988 /* According to the SPIR-V spec, OpCompositeInsert may work down to
2989 * the component granularity. In that case, the last index will be
2990 * the index to insert the scalar into the vector.
2993 cur
->def
= vtn_vector_insert(b
, cur
->def
, insert
->def
, indices
[i
]);
2995 cur
->elems
[indices
[i
]] = insert
;
3001 static struct vtn_ssa_value
*
3002 vtn_composite_extract(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
3003 const uint32_t *indices
, unsigned num_indices
)
3005 struct vtn_ssa_value
*cur
= src
;
3006 for (unsigned i
= 0; i
< num_indices
; i
++) {
3007 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
3008 vtn_assert(i
== num_indices
- 1);
3009 /* According to the SPIR-V spec, OpCompositeExtract may work down to
3010 * the component granularity. The last index will be the index of the
3011 * vector to extract.
3014 struct vtn_ssa_value
*ret
= rzalloc(b
, struct vtn_ssa_value
);
3015 ret
->type
= glsl_scalar_type(glsl_get_base_type(cur
->type
));
3016 ret
->def
= vtn_vector_extract(b
, cur
->def
, indices
[i
]);
3019 cur
= cur
->elems
[indices
[i
]];
3027 vtn_handle_composite(struct vtn_builder
*b
, SpvOp opcode
,
3028 const uint32_t *w
, unsigned count
)
3030 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
3031 const struct glsl_type
*type
=
3032 vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
3033 val
->ssa
= vtn_create_ssa_value(b
, type
);
3036 case SpvOpVectorExtractDynamic
:
3037 val
->ssa
->def
= vtn_vector_extract_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
3038 vtn_ssa_value(b
, w
[4])->def
);
3041 case SpvOpVectorInsertDynamic
:
3042 val
->ssa
->def
= vtn_vector_insert_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
3043 vtn_ssa_value(b
, w
[4])->def
,
3044 vtn_ssa_value(b
, w
[5])->def
);
3047 case SpvOpVectorShuffle
:
3048 val
->ssa
->def
= vtn_vector_shuffle(b
, glsl_get_vector_elements(type
),
3049 vtn_ssa_value(b
, w
[3])->def
,
3050 vtn_ssa_value(b
, w
[4])->def
,
3054 case SpvOpCompositeConstruct
: {
3055 unsigned elems
= count
- 3;
3057 if (glsl_type_is_vector_or_scalar(type
)) {
3058 nir_ssa_def
*srcs
[NIR_MAX_VEC_COMPONENTS
];
3059 for (unsigned i
= 0; i
< elems
; i
++)
3060 srcs
[i
] = vtn_ssa_value(b
, w
[3 + i
])->def
;
3062 vtn_vector_construct(b
, glsl_get_vector_elements(type
),
3065 val
->ssa
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
3066 for (unsigned i
= 0; i
< elems
; i
++)
3067 val
->ssa
->elems
[i
] = vtn_ssa_value(b
, w
[3 + i
]);
3071 case SpvOpCompositeExtract
:
3072 val
->ssa
= vtn_composite_extract(b
, vtn_ssa_value(b
, w
[3]),
3076 case SpvOpCompositeInsert
:
3077 val
->ssa
= vtn_composite_insert(b
, vtn_ssa_value(b
, w
[4]),
3078 vtn_ssa_value(b
, w
[3]),
3082 case SpvOpCopyObject
:
3083 val
->ssa
= vtn_composite_copy(b
, vtn_ssa_value(b
, w
[3]));
3087 vtn_fail("unknown composite operation");
3092 vtn_emit_barrier(struct vtn_builder
*b
, nir_intrinsic_op op
)
3094 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
3095 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3099 vtn_emit_memory_barrier(struct vtn_builder
*b
, SpvScope scope
,
3100 SpvMemorySemanticsMask semantics
)
3102 static const SpvMemorySemanticsMask all_memory_semantics
=
3103 SpvMemorySemanticsUniformMemoryMask
|
3104 SpvMemorySemanticsWorkgroupMemoryMask
|
3105 SpvMemorySemanticsAtomicCounterMemoryMask
|
3106 SpvMemorySemanticsImageMemoryMask
;
3108 /* If we're not actually doing a memory barrier, bail */
3109 if (!(semantics
& all_memory_semantics
))
3112 /* GL and Vulkan don't have these */
3113 vtn_assert(scope
!= SpvScopeCrossDevice
);
3115 if (scope
== SpvScopeSubgroup
)
3116 return; /* Nothing to do here */
3118 if (scope
== SpvScopeWorkgroup
) {
3119 vtn_emit_barrier(b
, nir_intrinsic_group_memory_barrier
);
3123 /* There's only two scopes thing left */
3124 vtn_assert(scope
== SpvScopeInvocation
|| scope
== SpvScopeDevice
);
3126 if ((semantics
& all_memory_semantics
) == all_memory_semantics
) {
3127 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier
);
3131 /* Issue a bunch of more specific barriers */
3132 uint32_t bits
= semantics
;
3134 SpvMemorySemanticsMask semantic
= 1 << u_bit_scan(&bits
);
3136 case SpvMemorySemanticsUniformMemoryMask
:
3137 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_buffer
);
3139 case SpvMemorySemanticsWorkgroupMemoryMask
:
3140 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_shared
);
3142 case SpvMemorySemanticsAtomicCounterMemoryMask
:
3143 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_atomic_counter
);
3145 case SpvMemorySemanticsImageMemoryMask
:
3146 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_image
);
3155 vtn_handle_barrier(struct vtn_builder
*b
, SpvOp opcode
,
3156 const uint32_t *w
, unsigned count
)
3159 case SpvOpEmitVertex
:
3160 case SpvOpEmitStreamVertex
:
3161 case SpvOpEndPrimitive
:
3162 case SpvOpEndStreamPrimitive
: {
3163 nir_intrinsic_op intrinsic_op
;
3165 case SpvOpEmitVertex
:
3166 case SpvOpEmitStreamVertex
:
3167 intrinsic_op
= nir_intrinsic_emit_vertex
;
3169 case SpvOpEndPrimitive
:
3170 case SpvOpEndStreamPrimitive
:
3171 intrinsic_op
= nir_intrinsic_end_primitive
;
3174 unreachable("Invalid opcode");
3177 nir_intrinsic_instr
*intrin
=
3178 nir_intrinsic_instr_create(b
->shader
, intrinsic_op
);
3181 case SpvOpEmitStreamVertex
:
3182 case SpvOpEndStreamPrimitive
: {
3183 unsigned stream
= vtn_constant_value(b
, w
[1])->values
[0].u32
[0];
3184 nir_intrinsic_set_stream_id(intrin
, stream
);
3192 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3196 case SpvOpMemoryBarrier
: {
3197 SpvScope scope
= vtn_constant_value(b
, w
[1])->values
[0].u32
[0];
3198 SpvMemorySemanticsMask semantics
=
3199 vtn_constant_value(b
, w
[2])->values
[0].u32
[0];
3200 vtn_emit_memory_barrier(b
, scope
, semantics
);
3204 case SpvOpControlBarrier
: {
3205 SpvScope execution_scope
=
3206 vtn_constant_value(b
, w
[1])->values
[0].u32
[0];
3207 if (execution_scope
== SpvScopeWorkgroup
)
3208 vtn_emit_barrier(b
, nir_intrinsic_barrier
);
3210 SpvScope memory_scope
=
3211 vtn_constant_value(b
, w
[2])->values
[0].u32
[0];
3212 SpvMemorySemanticsMask memory_semantics
=
3213 vtn_constant_value(b
, w
[3])->values
[0].u32
[0];
3214 vtn_emit_memory_barrier(b
, memory_scope
, memory_semantics
);
3219 unreachable("unknown barrier instruction");
3224 gl_primitive_from_spv_execution_mode(struct vtn_builder
*b
,
3225 SpvExecutionMode mode
)
3228 case SpvExecutionModeInputPoints
:
3229 case SpvExecutionModeOutputPoints
:
3230 return 0; /* GL_POINTS */
3231 case SpvExecutionModeInputLines
:
3232 return 1; /* GL_LINES */
3233 case SpvExecutionModeInputLinesAdjacency
:
3234 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
3235 case SpvExecutionModeTriangles
:
3236 return 4; /* GL_TRIANGLES */
3237 case SpvExecutionModeInputTrianglesAdjacency
:
3238 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
3239 case SpvExecutionModeQuads
:
3240 return 7; /* GL_QUADS */
3241 case SpvExecutionModeIsolines
:
3242 return 0x8E7A; /* GL_ISOLINES */
3243 case SpvExecutionModeOutputLineStrip
:
3244 return 3; /* GL_LINE_STRIP */
3245 case SpvExecutionModeOutputTriangleStrip
:
3246 return 5; /* GL_TRIANGLE_STRIP */
3248 vtn_fail("Invalid primitive type");
3253 vertices_in_from_spv_execution_mode(struct vtn_builder
*b
,
3254 SpvExecutionMode mode
)
3257 case SpvExecutionModeInputPoints
:
3259 case SpvExecutionModeInputLines
:
3261 case SpvExecutionModeInputLinesAdjacency
:
3263 case SpvExecutionModeTriangles
:
3265 case SpvExecutionModeInputTrianglesAdjacency
:
3268 vtn_fail("Invalid GS input mode");
3272 static gl_shader_stage
3273 stage_for_execution_model(struct vtn_builder
*b
, SpvExecutionModel model
)
3276 case SpvExecutionModelVertex
:
3277 return MESA_SHADER_VERTEX
;
3278 case SpvExecutionModelTessellationControl
:
3279 return MESA_SHADER_TESS_CTRL
;
3280 case SpvExecutionModelTessellationEvaluation
:
3281 return MESA_SHADER_TESS_EVAL
;
3282 case SpvExecutionModelGeometry
:
3283 return MESA_SHADER_GEOMETRY
;
3284 case SpvExecutionModelFragment
:
3285 return MESA_SHADER_FRAGMENT
;
3286 case SpvExecutionModelGLCompute
:
3287 return MESA_SHADER_COMPUTE
;
3289 vtn_fail("Unsupported execution model");
3293 #define spv_check_supported(name, cap) do { \
3294 if (!(b->options && b->options->caps.name)) \
3295 vtn_warn("Unsupported SPIR-V capability: %s", \
3296 spirv_capability_to_string(cap)); \
3301 vtn_handle_entry_point(struct vtn_builder
*b
, const uint32_t *w
,
3304 struct vtn_value
*entry_point
= &b
->values
[w
[2]];
3305 /* Let this be a name label regardless */
3306 unsigned name_words
;
3307 entry_point
->name
= vtn_string_literal(b
, &w
[3], count
- 3, &name_words
);
3309 if (strcmp(entry_point
->name
, b
->entry_point_name
) != 0 ||
3310 stage_for_execution_model(b
, w
[1]) != b
->entry_point_stage
)
3313 vtn_assert(b
->entry_point
== NULL
);
3314 b
->entry_point
= entry_point
;
3318 vtn_handle_preamble_instruction(struct vtn_builder
*b
, SpvOp opcode
,
3319 const uint32_t *w
, unsigned count
)
3326 case SpvSourceLanguageUnknown
: lang
= "unknown"; break;
3327 case SpvSourceLanguageESSL
: lang
= "ESSL"; break;
3328 case SpvSourceLanguageGLSL
: lang
= "GLSL"; break;
3329 case SpvSourceLanguageOpenCL_C
: lang
= "OpenCL C"; break;
3330 case SpvSourceLanguageOpenCL_CPP
: lang
= "OpenCL C++"; break;
3331 case SpvSourceLanguageHLSL
: lang
= "HLSL"; break;
3334 uint32_t version
= w
[2];
3337 (count
> 3) ? vtn_value(b
, w
[3], vtn_value_type_string
)->str
: "";
3339 vtn_info("Parsing SPIR-V from %s %u source file %s", lang
, version
, file
);
3343 case SpvOpSourceExtension
:
3344 case SpvOpSourceContinued
:
3345 case SpvOpExtension
:
3346 case SpvOpModuleProcessed
:
3347 /* Unhandled, but these are for debug so that's ok. */
3350 case SpvOpCapability
: {
3351 SpvCapability cap
= w
[1];
3353 case SpvCapabilityMatrix
:
3354 case SpvCapabilityShader
:
3355 case SpvCapabilityGeometry
:
3356 case SpvCapabilityGeometryPointSize
:
3357 case SpvCapabilityUniformBufferArrayDynamicIndexing
:
3358 case SpvCapabilitySampledImageArrayDynamicIndexing
:
3359 case SpvCapabilityStorageBufferArrayDynamicIndexing
:
3360 case SpvCapabilityStorageImageArrayDynamicIndexing
:
3361 case SpvCapabilityImageRect
:
3362 case SpvCapabilitySampledRect
:
3363 case SpvCapabilitySampled1D
:
3364 case SpvCapabilityImage1D
:
3365 case SpvCapabilitySampledCubeArray
:
3366 case SpvCapabilityImageCubeArray
:
3367 case SpvCapabilitySampledBuffer
:
3368 case SpvCapabilityImageBuffer
:
3369 case SpvCapabilityImageQuery
:
3370 case SpvCapabilityDerivativeControl
:
3371 case SpvCapabilityInterpolationFunction
:
3372 case SpvCapabilityMultiViewport
:
3373 case SpvCapabilitySampleRateShading
:
3374 case SpvCapabilityClipDistance
:
3375 case SpvCapabilityCullDistance
:
3376 case SpvCapabilityInputAttachment
:
3377 case SpvCapabilityImageGatherExtended
:
3378 case SpvCapabilityStorageImageExtendedFormats
:
3381 case SpvCapabilityLinkage
:
3382 case SpvCapabilityVector16
:
3383 case SpvCapabilityFloat16Buffer
:
3384 case SpvCapabilityFloat16
:
3385 case SpvCapabilityInt64Atomics
:
3386 case SpvCapabilityStorageImageMultisample
:
3387 case SpvCapabilityInt8
:
3388 case SpvCapabilitySparseResidency
:
3389 case SpvCapabilityMinLod
:
3390 vtn_warn("Unsupported SPIR-V capability: %s",
3391 spirv_capability_to_string(cap
));
3394 case SpvCapabilityAtomicStorage
:
3395 spv_check_supported(atomic_storage
, cap
);
3398 case SpvCapabilityFloat64
:
3399 spv_check_supported(float64
, cap
);
3401 case SpvCapabilityInt64
:
3402 spv_check_supported(int64
, cap
);
3404 case SpvCapabilityInt16
:
3405 spv_check_supported(int16
, cap
);
3408 case SpvCapabilityTransformFeedback
:
3409 spv_check_supported(transform_feedback
, cap
);
3412 case SpvCapabilityGeometryStreams
:
3413 spv_check_supported(geometry_streams
, cap
);
3416 case SpvCapabilityAddresses
:
3417 case SpvCapabilityKernel
:
3418 case SpvCapabilityImageBasic
:
3419 case SpvCapabilityImageReadWrite
:
3420 case SpvCapabilityImageMipmap
:
3421 case SpvCapabilityPipes
:
3422 case SpvCapabilityGroups
:
3423 case SpvCapabilityDeviceEnqueue
:
3424 case SpvCapabilityLiteralSampler
:
3425 case SpvCapabilityGenericPointer
:
3426 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
3427 spirv_capability_to_string(cap
));
3430 case SpvCapabilityImageMSArray
:
3431 spv_check_supported(image_ms_array
, cap
);
3434 case SpvCapabilityTessellation
:
3435 case SpvCapabilityTessellationPointSize
:
3436 spv_check_supported(tessellation
, cap
);
3439 case SpvCapabilityDrawParameters
:
3440 spv_check_supported(draw_parameters
, cap
);
3443 case SpvCapabilityStorageImageReadWithoutFormat
:
3444 spv_check_supported(image_read_without_format
, cap
);
3447 case SpvCapabilityStorageImageWriteWithoutFormat
:
3448 spv_check_supported(image_write_without_format
, cap
);
3451 case SpvCapabilityDeviceGroup
:
3452 spv_check_supported(device_group
, cap
);
3455 case SpvCapabilityMultiView
:
3456 spv_check_supported(multiview
, cap
);
3459 case SpvCapabilityGroupNonUniform
:
3460 spv_check_supported(subgroup_basic
, cap
);
3463 case SpvCapabilityGroupNonUniformVote
:
3464 spv_check_supported(subgroup_vote
, cap
);
3467 case SpvCapabilitySubgroupBallotKHR
:
3468 case SpvCapabilityGroupNonUniformBallot
:
3469 spv_check_supported(subgroup_ballot
, cap
);
3472 case SpvCapabilityGroupNonUniformShuffle
:
3473 case SpvCapabilityGroupNonUniformShuffleRelative
:
3474 spv_check_supported(subgroup_shuffle
, cap
);
3477 case SpvCapabilityGroupNonUniformQuad
:
3478 spv_check_supported(subgroup_quad
, cap
);
3481 case SpvCapabilityGroupNonUniformArithmetic
:
3482 case SpvCapabilityGroupNonUniformClustered
:
3483 spv_check_supported(subgroup_arithmetic
, cap
);
3486 case SpvCapabilityVariablePointersStorageBuffer
:
3487 case SpvCapabilityVariablePointers
:
3488 spv_check_supported(variable_pointers
, cap
);
3491 case SpvCapabilityStorageUniformBufferBlock16
:
3492 case SpvCapabilityStorageUniform16
:
3493 case SpvCapabilityStoragePushConstant16
:
3494 case SpvCapabilityStorageInputOutput16
:
3495 spv_check_supported(storage_16bit
, cap
);
3498 case SpvCapabilityShaderViewportIndexLayerEXT
:
3499 spv_check_supported(shader_viewport_index_layer
, cap
);
3502 case SpvCapabilityStorageBuffer8BitAccess
:
3503 case SpvCapabilityUniformAndStorageBuffer8BitAccess
:
3504 case SpvCapabilityStoragePushConstant8
:
3505 spv_check_supported(storage_8bit
, cap
);
3508 case SpvCapabilityInputAttachmentArrayDynamicIndexingEXT
:
3509 case SpvCapabilityUniformTexelBufferArrayDynamicIndexingEXT
:
3510 case SpvCapabilityStorageTexelBufferArrayDynamicIndexingEXT
:
3511 spv_check_supported(descriptor_array_dynamic_indexing
, cap
);
3514 case SpvCapabilityRuntimeDescriptorArrayEXT
:
3515 spv_check_supported(runtime_descriptor_array
, cap
);
3518 case SpvCapabilityStencilExportEXT
:
3519 spv_check_supported(stencil_export
, cap
);
3522 case SpvCapabilitySampleMaskPostDepthCoverage
:
3523 spv_check_supported(post_depth_coverage
, cap
);
3527 vtn_fail("Unhandled capability");
3532 case SpvOpExtInstImport
:
3533 vtn_handle_extension(b
, opcode
, w
, count
);
3536 case SpvOpMemoryModel
:
3537 vtn_assert(w
[1] == SpvAddressingModelLogical
);
3538 vtn_assert(w
[2] == SpvMemoryModelSimple
||
3539 w
[2] == SpvMemoryModelGLSL450
);
3542 case SpvOpEntryPoint
:
3543 vtn_handle_entry_point(b
, w
, count
);
3547 vtn_push_value(b
, w
[1], vtn_value_type_string
)->str
=
3548 vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
3552 b
->values
[w
[1]].name
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
3555 case SpvOpMemberName
:
3559 case SpvOpExecutionMode
:
3560 case SpvOpDecorationGroup
:
3562 case SpvOpMemberDecorate
:
3563 case SpvOpGroupDecorate
:
3564 case SpvOpGroupMemberDecorate
:
3565 case SpvOpDecorateStringGOOGLE
:
3566 case SpvOpMemberDecorateStringGOOGLE
:
3567 vtn_handle_decoration(b
, opcode
, w
, count
);
3571 return false; /* End of preamble */
3578 vtn_handle_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
3579 const struct vtn_decoration
*mode
, void *data
)
3581 vtn_assert(b
->entry_point
== entry_point
);
3583 switch(mode
->exec_mode
) {
3584 case SpvExecutionModeOriginUpperLeft
:
3585 case SpvExecutionModeOriginLowerLeft
:
3586 b
->origin_upper_left
=
3587 (mode
->exec_mode
== SpvExecutionModeOriginUpperLeft
);
3590 case SpvExecutionModeEarlyFragmentTests
:
3591 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3592 b
->shader
->info
.fs
.early_fragment_tests
= true;
3595 case SpvExecutionModePostDepthCoverage
:
3596 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3597 b
->shader
->info
.fs
.post_depth_coverage
= true;
3600 case SpvExecutionModeInvocations
:
3601 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3602 b
->shader
->info
.gs
.invocations
= MAX2(1, mode
->literals
[0]);
3605 case SpvExecutionModeDepthReplacing
:
3606 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3607 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_ANY
;
3609 case SpvExecutionModeDepthGreater
:
3610 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3611 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_GREATER
;
3613 case SpvExecutionModeDepthLess
:
3614 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3615 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_LESS
;
3617 case SpvExecutionModeDepthUnchanged
:
3618 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3619 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
3622 case SpvExecutionModeLocalSize
:
3623 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
3624 b
->shader
->info
.cs
.local_size
[0] = mode
->literals
[0];
3625 b
->shader
->info
.cs
.local_size
[1] = mode
->literals
[1];
3626 b
->shader
->info
.cs
.local_size
[2] = mode
->literals
[2];
3628 case SpvExecutionModeLocalSizeHint
:
3629 break; /* Nothing to do with this */
3631 case SpvExecutionModeOutputVertices
:
3632 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3633 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
3634 b
->shader
->info
.tess
.tcs_vertices_out
= mode
->literals
[0];
3636 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3637 b
->shader
->info
.gs
.vertices_out
= mode
->literals
[0];
3641 case SpvExecutionModeInputPoints
:
3642 case SpvExecutionModeInputLines
:
3643 case SpvExecutionModeInputLinesAdjacency
:
3644 case SpvExecutionModeTriangles
:
3645 case SpvExecutionModeInputTrianglesAdjacency
:
3646 case SpvExecutionModeQuads
:
3647 case SpvExecutionModeIsolines
:
3648 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3649 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
3650 b
->shader
->info
.tess
.primitive_mode
=
3651 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
3653 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3654 b
->shader
->info
.gs
.vertices_in
=
3655 vertices_in_from_spv_execution_mode(b
, mode
->exec_mode
);
3656 b
->shader
->info
.gs
.input_primitive
=
3657 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
3661 case SpvExecutionModeOutputPoints
:
3662 case SpvExecutionModeOutputLineStrip
:
3663 case SpvExecutionModeOutputTriangleStrip
:
3664 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3665 b
->shader
->info
.gs
.output_primitive
=
3666 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
3669 case SpvExecutionModeSpacingEqual
:
3670 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3671 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3672 b
->shader
->info
.tess
.spacing
= TESS_SPACING_EQUAL
;
3674 case SpvExecutionModeSpacingFractionalEven
:
3675 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3676 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3677 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_EVEN
;
3679 case SpvExecutionModeSpacingFractionalOdd
:
3680 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3681 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3682 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_ODD
;
3684 case SpvExecutionModeVertexOrderCw
:
3685 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3686 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3687 b
->shader
->info
.tess
.ccw
= false;
3689 case SpvExecutionModeVertexOrderCcw
:
3690 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3691 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3692 b
->shader
->info
.tess
.ccw
= true;
3694 case SpvExecutionModePointMode
:
3695 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3696 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3697 b
->shader
->info
.tess
.point_mode
= true;
3700 case SpvExecutionModePixelCenterInteger
:
3701 b
->pixel_center_integer
= true;
3704 case SpvExecutionModeXfb
:
3705 b
->shader
->info
.has_transform_feedback_varyings
= true;
3708 case SpvExecutionModeVecTypeHint
:
3709 case SpvExecutionModeContractionOff
:
3712 case SpvExecutionModeStencilRefReplacingEXT
:
3713 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3717 vtn_fail("Unhandled execution mode");
3722 vtn_handle_variable_or_type_instruction(struct vtn_builder
*b
, SpvOp opcode
,
3723 const uint32_t *w
, unsigned count
)
3725 vtn_set_instruction_result_type(b
, opcode
, w
, count
);
3729 case SpvOpSourceContinued
:
3730 case SpvOpSourceExtension
:
3731 case SpvOpExtension
:
3732 case SpvOpCapability
:
3733 case SpvOpExtInstImport
:
3734 case SpvOpMemoryModel
:
3735 case SpvOpEntryPoint
:
3736 case SpvOpExecutionMode
:
3739 case SpvOpMemberName
:
3740 case SpvOpDecorationGroup
:
3742 case SpvOpMemberDecorate
:
3743 case SpvOpGroupDecorate
:
3744 case SpvOpGroupMemberDecorate
:
3745 case SpvOpDecorateStringGOOGLE
:
3746 case SpvOpMemberDecorateStringGOOGLE
:
3747 vtn_fail("Invalid opcode types and variables section");
3753 case SpvOpTypeFloat
:
3754 case SpvOpTypeVector
:
3755 case SpvOpTypeMatrix
:
3756 case SpvOpTypeImage
:
3757 case SpvOpTypeSampler
:
3758 case SpvOpTypeSampledImage
:
3759 case SpvOpTypeArray
:
3760 case SpvOpTypeRuntimeArray
:
3761 case SpvOpTypeStruct
:
3762 case SpvOpTypeOpaque
:
3763 case SpvOpTypePointer
:
3764 case SpvOpTypeFunction
:
3765 case SpvOpTypeEvent
:
3766 case SpvOpTypeDeviceEvent
:
3767 case SpvOpTypeReserveId
:
3768 case SpvOpTypeQueue
:
3770 vtn_handle_type(b
, opcode
, w
, count
);
3773 case SpvOpConstantTrue
:
3774 case SpvOpConstantFalse
:
3776 case SpvOpConstantComposite
:
3777 case SpvOpConstantSampler
:
3778 case SpvOpConstantNull
:
3779 case SpvOpSpecConstantTrue
:
3780 case SpvOpSpecConstantFalse
:
3781 case SpvOpSpecConstant
:
3782 case SpvOpSpecConstantComposite
:
3783 case SpvOpSpecConstantOp
:
3784 vtn_handle_constant(b
, opcode
, w
, count
);
3789 vtn_handle_variables(b
, opcode
, w
, count
);
3793 return false; /* End of preamble */
3800 vtn_handle_body_instruction(struct vtn_builder
*b
, SpvOp opcode
,
3801 const uint32_t *w
, unsigned count
)
3807 case SpvOpLoopMerge
:
3808 case SpvOpSelectionMerge
:
3809 /* This is handled by cfg pre-pass and walk_blocks */
3813 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
3814 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
3819 vtn_handle_extension(b
, opcode
, w
, count
);
3825 case SpvOpCopyMemory
:
3826 case SpvOpCopyMemorySized
:
3827 case SpvOpAccessChain
:
3828 case SpvOpPtrAccessChain
:
3829 case SpvOpInBoundsAccessChain
:
3830 case SpvOpArrayLength
:
3831 vtn_handle_variables(b
, opcode
, w
, count
);
3834 case SpvOpFunctionCall
:
3835 vtn_handle_function_call(b
, opcode
, w
, count
);
3838 case SpvOpSampledImage
:
3840 case SpvOpImageSampleImplicitLod
:
3841 case SpvOpImageSampleExplicitLod
:
3842 case SpvOpImageSampleDrefImplicitLod
:
3843 case SpvOpImageSampleDrefExplicitLod
:
3844 case SpvOpImageSampleProjImplicitLod
:
3845 case SpvOpImageSampleProjExplicitLod
:
3846 case SpvOpImageSampleProjDrefImplicitLod
:
3847 case SpvOpImageSampleProjDrefExplicitLod
:
3848 case SpvOpImageFetch
:
3849 case SpvOpImageGather
:
3850 case SpvOpImageDrefGather
:
3851 case SpvOpImageQuerySizeLod
:
3852 case SpvOpImageQueryLod
:
3853 case SpvOpImageQueryLevels
:
3854 case SpvOpImageQuerySamples
:
3855 vtn_handle_texture(b
, opcode
, w
, count
);
3858 case SpvOpImageRead
:
3859 case SpvOpImageWrite
:
3860 case SpvOpImageTexelPointer
:
3861 vtn_handle_image(b
, opcode
, w
, count
);
3864 case SpvOpImageQuerySize
: {
3865 struct vtn_pointer
*image
=
3866 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
3867 if (glsl_type_is_image(image
->type
->type
)) {
3868 vtn_handle_image(b
, opcode
, w
, count
);
3870 vtn_assert(glsl_type_is_sampler(image
->type
->type
));
3871 vtn_handle_texture(b
, opcode
, w
, count
);
3876 case SpvOpAtomicLoad
:
3877 case SpvOpAtomicExchange
:
3878 case SpvOpAtomicCompareExchange
:
3879 case SpvOpAtomicCompareExchangeWeak
:
3880 case SpvOpAtomicIIncrement
:
3881 case SpvOpAtomicIDecrement
:
3882 case SpvOpAtomicIAdd
:
3883 case SpvOpAtomicISub
:
3884 case SpvOpAtomicSMin
:
3885 case SpvOpAtomicUMin
:
3886 case SpvOpAtomicSMax
:
3887 case SpvOpAtomicUMax
:
3888 case SpvOpAtomicAnd
:
3890 case SpvOpAtomicXor
: {
3891 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[3]);
3892 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
3893 vtn_handle_image(b
, opcode
, w
, count
);
3895 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
3896 vtn_handle_atomics(b
, opcode
, w
, count
);
3901 case SpvOpAtomicStore
: {
3902 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[1]);
3903 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
3904 vtn_handle_image(b
, opcode
, w
, count
);
3906 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
3907 vtn_handle_atomics(b
, opcode
, w
, count
);
3913 /* Handle OpSelect up-front here because it needs to be able to handle
3914 * pointers and not just regular vectors and scalars.
3916 struct vtn_value
*res_val
= vtn_untyped_value(b
, w
[2]);
3917 struct vtn_value
*sel_val
= vtn_untyped_value(b
, w
[3]);
3918 struct vtn_value
*obj1_val
= vtn_untyped_value(b
, w
[4]);
3919 struct vtn_value
*obj2_val
= vtn_untyped_value(b
, w
[5]);
3921 const struct glsl_type
*sel_type
;
3922 switch (res_val
->type
->base_type
) {
3923 case vtn_base_type_scalar
:
3924 sel_type
= glsl_bool_type();
3926 case vtn_base_type_vector
:
3927 sel_type
= glsl_vector_type(GLSL_TYPE_BOOL
, res_val
->type
->length
);
3929 case vtn_base_type_pointer
:
3930 /* We need to have actual storage for pointer types */
3931 vtn_fail_if(res_val
->type
->type
== NULL
,
3932 "Invalid pointer result type for OpSelect");
3933 sel_type
= glsl_bool_type();
3936 vtn_fail("Result type of OpSelect must be a scalar, vector, or pointer");
3939 if (unlikely(sel_val
->type
->type
!= sel_type
)) {
3940 if (sel_val
->type
->type
== glsl_bool_type()) {
3941 /* This case is illegal but some older versions of GLSLang produce
3942 * it. The GLSLang issue was fixed on March 30, 2017:
3944 * https://github.com/KhronosGroup/glslang/issues/809
3946 * Unfortunately, there are applications in the wild which are
3947 * shipping with this bug so it isn't nice to fail on them so we
3948 * throw a warning instead. It's not actually a problem for us as
3949 * nir_builder will just splat the condition out which is most
3950 * likely what the client wanted anyway.
3952 vtn_warn("Condition type of OpSelect must have the same number "
3953 "of components as Result Type");
3955 vtn_fail("Condition type of OpSelect must be a scalar or vector "
3956 "of Boolean type. It must have the same number of "
3957 "components as Result Type");
3961 vtn_fail_if(obj1_val
->type
!= res_val
->type
||
3962 obj2_val
->type
!= res_val
->type
,
3963 "Object types must match the result type in OpSelect");
3965 struct vtn_type
*res_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
3966 struct vtn_ssa_value
*ssa
= vtn_create_ssa_value(b
, res_type
->type
);
3967 ssa
->def
= nir_bcsel(&b
->nb
, vtn_ssa_value(b
, w
[3])->def
,
3968 vtn_ssa_value(b
, w
[4])->def
,
3969 vtn_ssa_value(b
, w
[5])->def
);
3970 vtn_push_ssa(b
, w
[2], res_type
, ssa
);
3979 case SpvOpConvertFToU
:
3980 case SpvOpConvertFToS
:
3981 case SpvOpConvertSToF
:
3982 case SpvOpConvertUToF
:
3986 case SpvOpQuantizeToF16
:
3987 case SpvOpConvertPtrToU
:
3988 case SpvOpConvertUToPtr
:
3989 case SpvOpPtrCastToGeneric
:
3990 case SpvOpGenericCastToPtr
:
3996 case SpvOpSignBitSet
:
3997 case SpvOpLessOrGreater
:
3999 case SpvOpUnordered
:
4014 case SpvOpVectorTimesScalar
:
4016 case SpvOpIAddCarry
:
4017 case SpvOpISubBorrow
:
4018 case SpvOpUMulExtended
:
4019 case SpvOpSMulExtended
:
4020 case SpvOpShiftRightLogical
:
4021 case SpvOpShiftRightArithmetic
:
4022 case SpvOpShiftLeftLogical
:
4023 case SpvOpLogicalEqual
:
4024 case SpvOpLogicalNotEqual
:
4025 case SpvOpLogicalOr
:
4026 case SpvOpLogicalAnd
:
4027 case SpvOpLogicalNot
:
4028 case SpvOpBitwiseOr
:
4029 case SpvOpBitwiseXor
:
4030 case SpvOpBitwiseAnd
:
4032 case SpvOpFOrdEqual
:
4033 case SpvOpFUnordEqual
:
4034 case SpvOpINotEqual
:
4035 case SpvOpFOrdNotEqual
:
4036 case SpvOpFUnordNotEqual
:
4037 case SpvOpULessThan
:
4038 case SpvOpSLessThan
:
4039 case SpvOpFOrdLessThan
:
4040 case SpvOpFUnordLessThan
:
4041 case SpvOpUGreaterThan
:
4042 case SpvOpSGreaterThan
:
4043 case SpvOpFOrdGreaterThan
:
4044 case SpvOpFUnordGreaterThan
:
4045 case SpvOpULessThanEqual
:
4046 case SpvOpSLessThanEqual
:
4047 case SpvOpFOrdLessThanEqual
:
4048 case SpvOpFUnordLessThanEqual
:
4049 case SpvOpUGreaterThanEqual
:
4050 case SpvOpSGreaterThanEqual
:
4051 case SpvOpFOrdGreaterThanEqual
:
4052 case SpvOpFUnordGreaterThanEqual
:
4058 case SpvOpFwidthFine
:
4059 case SpvOpDPdxCoarse
:
4060 case SpvOpDPdyCoarse
:
4061 case SpvOpFwidthCoarse
:
4062 case SpvOpBitFieldInsert
:
4063 case SpvOpBitFieldSExtract
:
4064 case SpvOpBitFieldUExtract
:
4065 case SpvOpBitReverse
:
4067 case SpvOpTranspose
:
4068 case SpvOpOuterProduct
:
4069 case SpvOpMatrixTimesScalar
:
4070 case SpvOpVectorTimesMatrix
:
4071 case SpvOpMatrixTimesVector
:
4072 case SpvOpMatrixTimesMatrix
:
4073 vtn_handle_alu(b
, opcode
, w
, count
);
4076 case SpvOpVectorExtractDynamic
:
4077 case SpvOpVectorInsertDynamic
:
4078 case SpvOpVectorShuffle
:
4079 case SpvOpCompositeConstruct
:
4080 case SpvOpCompositeExtract
:
4081 case SpvOpCompositeInsert
:
4082 case SpvOpCopyObject
:
4083 vtn_handle_composite(b
, opcode
, w
, count
);
4086 case SpvOpEmitVertex
:
4087 case SpvOpEndPrimitive
:
4088 case SpvOpEmitStreamVertex
:
4089 case SpvOpEndStreamPrimitive
:
4090 case SpvOpControlBarrier
:
4091 case SpvOpMemoryBarrier
:
4092 vtn_handle_barrier(b
, opcode
, w
, count
);
4095 case SpvOpGroupNonUniformElect
:
4096 case SpvOpGroupNonUniformAll
:
4097 case SpvOpGroupNonUniformAny
:
4098 case SpvOpGroupNonUniformAllEqual
:
4099 case SpvOpGroupNonUniformBroadcast
:
4100 case SpvOpGroupNonUniformBroadcastFirst
:
4101 case SpvOpGroupNonUniformBallot
:
4102 case SpvOpGroupNonUniformInverseBallot
:
4103 case SpvOpGroupNonUniformBallotBitExtract
:
4104 case SpvOpGroupNonUniformBallotBitCount
:
4105 case SpvOpGroupNonUniformBallotFindLSB
:
4106 case SpvOpGroupNonUniformBallotFindMSB
:
4107 case SpvOpGroupNonUniformShuffle
:
4108 case SpvOpGroupNonUniformShuffleXor
:
4109 case SpvOpGroupNonUniformShuffleUp
:
4110 case SpvOpGroupNonUniformShuffleDown
:
4111 case SpvOpGroupNonUniformIAdd
:
4112 case SpvOpGroupNonUniformFAdd
:
4113 case SpvOpGroupNonUniformIMul
:
4114 case SpvOpGroupNonUniformFMul
:
4115 case SpvOpGroupNonUniformSMin
:
4116 case SpvOpGroupNonUniformUMin
:
4117 case SpvOpGroupNonUniformFMin
:
4118 case SpvOpGroupNonUniformSMax
:
4119 case SpvOpGroupNonUniformUMax
:
4120 case SpvOpGroupNonUniformFMax
:
4121 case SpvOpGroupNonUniformBitwiseAnd
:
4122 case SpvOpGroupNonUniformBitwiseOr
:
4123 case SpvOpGroupNonUniformBitwiseXor
:
4124 case SpvOpGroupNonUniformLogicalAnd
:
4125 case SpvOpGroupNonUniformLogicalOr
:
4126 case SpvOpGroupNonUniformLogicalXor
:
4127 case SpvOpGroupNonUniformQuadBroadcast
:
4128 case SpvOpGroupNonUniformQuadSwap
:
4129 vtn_handle_subgroup(b
, opcode
, w
, count
);
4133 vtn_fail("Unhandled opcode");
4140 vtn_create_builder(const uint32_t *words
, size_t word_count
,
4141 gl_shader_stage stage
, const char *entry_point_name
,
4142 const struct spirv_to_nir_options
*options
)
4144 /* Initialize the vtn_builder object */
4145 struct vtn_builder
*b
= rzalloc(NULL
, struct vtn_builder
);
4147 b
->spirv_word_count
= word_count
;
4151 exec_list_make_empty(&b
->functions
);
4152 b
->entry_point_stage
= stage
;
4153 b
->entry_point_name
= entry_point_name
;
4154 b
->options
= options
;
4157 * Handle the SPIR-V header (first 5 dwords).
4158 * Can't use vtx_assert() as the setjmp(3) target isn't initialized yet.
4160 if (word_count
<= 5)
4163 if (words
[0] != SpvMagicNumber
) {
4164 vtn_err("words[0] was 0x%x, want 0x%x", words
[0], SpvMagicNumber
);
4167 if (words
[1] < 0x10000) {
4168 vtn_err("words[1] was 0x%x, want >= 0x10000", words
[1]);
4172 /* words[2] == generator magic */
4173 unsigned value_id_bound
= words
[3];
4174 if (words
[4] != 0) {
4175 vtn_err("words[4] was %u, want 0", words
[4]);
4179 b
->value_id_bound
= value_id_bound
;
4180 b
->values
= rzalloc_array(b
, struct vtn_value
, value_id_bound
);
4189 spirv_to_nir(const uint32_t *words
, size_t word_count
,
4190 struct nir_spirv_specialization
*spec
, unsigned num_spec
,
4191 gl_shader_stage stage
, const char *entry_point_name
,
4192 const struct spirv_to_nir_options
*options
,
4193 const nir_shader_compiler_options
*nir_options
)
4196 const uint32_t *word_end
= words
+ word_count
;
4198 struct vtn_builder
*b
= vtn_create_builder(words
, word_count
,
4199 stage
, entry_point_name
,
4205 /* See also _vtn_fail() */
4206 if (setjmp(b
->fail_jump
)) {
4211 /* Skip the SPIR-V header, handled at vtn_create_builder */
4214 /* Handle all the preamble instructions */
4215 words
= vtn_foreach_instruction(b
, words
, word_end
,
4216 vtn_handle_preamble_instruction
);
4218 if (b
->entry_point
== NULL
) {
4219 vtn_fail("Entry point not found");
4224 b
->shader
= nir_shader_create(b
, stage
, nir_options
, NULL
);
4226 /* Set shader info defaults */
4227 b
->shader
->info
.gs
.invocations
= 1;
4229 /* Parse execution modes */
4230 vtn_foreach_execution_mode(b
, b
->entry_point
,
4231 vtn_handle_execution_mode
, NULL
);
4233 b
->specializations
= spec
;
4234 b
->num_specializations
= num_spec
;
4236 /* Handle all variable, type, and constant instructions */
4237 words
= vtn_foreach_instruction(b
, words
, word_end
,
4238 vtn_handle_variable_or_type_instruction
);
4240 /* Set types on all vtn_values */
4241 vtn_foreach_instruction(b
, words
, word_end
, vtn_set_instruction_result_type
);
4243 vtn_build_cfg(b
, words
, word_end
);
4245 assert(b
->entry_point
->value_type
== vtn_value_type_function
);
4246 b
->entry_point
->func
->referenced
= true;
4251 foreach_list_typed(struct vtn_function
, func
, node
, &b
->functions
) {
4252 if (func
->referenced
&& !func
->emitted
) {
4253 b
->const_table
= _mesa_hash_table_create(b
, _mesa_hash_pointer
,
4254 _mesa_key_pointer_equal
);
4256 vtn_function_emit(b
, func
, vtn_handle_body_instruction
);
4262 /* We sometimes generate bogus derefs that, while never used, give the
4263 * validator a bit of heartburn. Run dead code to get rid of them.
4265 nir_opt_dce(b
->shader
);
4267 vtn_assert(b
->entry_point
->value_type
== vtn_value_type_function
);
4268 nir_function
*entry_point
= b
->entry_point
->func
->impl
->function
;
4269 vtn_assert(entry_point
);
4271 /* Unparent the shader from the vtn_builder before we delete the builder */
4272 ralloc_steal(NULL
, b
->shader
);