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
));
790 case SpvDecorationHlslSemanticGOOGLE
:
791 /* HLSL semantic decorations can safely be ignored by the driver. */
795 vtn_fail("Unhandled decoration");
799 /* Matrix strides are handled as a separate pass because we need to know
800 * whether the matrix is row-major or not first.
803 struct_member_matrix_stride_cb(struct vtn_builder
*b
,
804 struct vtn_value
*val
, int member
,
805 const struct vtn_decoration
*dec
,
808 if (dec
->decoration
!= SpvDecorationMatrixStride
)
811 vtn_fail_if(member
< 0,
812 "The MatrixStride decoration is only allowed on members "
815 struct member_decoration_ctx
*ctx
= void_ctx
;
817 struct vtn_type
*mat_type
= mutable_matrix_member(b
, ctx
->type
, member
);
818 if (mat_type
->row_major
) {
819 mat_type
->array_element
= vtn_type_copy(b
, mat_type
->array_element
);
820 mat_type
->stride
= mat_type
->array_element
->stride
;
821 mat_type
->array_element
->stride
= dec
->literals
[0];
823 vtn_assert(mat_type
->array_element
->stride
> 0);
824 mat_type
->stride
= dec
->literals
[0];
829 type_decoration_cb(struct vtn_builder
*b
,
830 struct vtn_value
*val
, int member
,
831 const struct vtn_decoration
*dec
, void *ctx
)
833 struct vtn_type
*type
= val
->type
;
836 /* This should have been handled by OpTypeStruct */
837 assert(val
->type
->base_type
== vtn_base_type_struct
);
838 assert(member
>= 0 && member
< val
->type
->length
);
842 switch (dec
->decoration
) {
843 case SpvDecorationArrayStride
:
844 vtn_assert(type
->base_type
== vtn_base_type_matrix
||
845 type
->base_type
== vtn_base_type_array
||
846 type
->base_type
== vtn_base_type_pointer
);
847 type
->stride
= dec
->literals
[0];
849 case SpvDecorationBlock
:
850 vtn_assert(type
->base_type
== vtn_base_type_struct
);
853 case SpvDecorationBufferBlock
:
854 vtn_assert(type
->base_type
== vtn_base_type_struct
);
855 type
->buffer_block
= true;
857 case SpvDecorationGLSLShared
:
858 case SpvDecorationGLSLPacked
:
859 /* Ignore these, since we get explicit offsets anyways */
862 case SpvDecorationRowMajor
:
863 case SpvDecorationColMajor
:
864 case SpvDecorationMatrixStride
:
865 case SpvDecorationBuiltIn
:
866 case SpvDecorationNoPerspective
:
867 case SpvDecorationFlat
:
868 case SpvDecorationPatch
:
869 case SpvDecorationCentroid
:
870 case SpvDecorationSample
:
871 case SpvDecorationVolatile
:
872 case SpvDecorationCoherent
:
873 case SpvDecorationNonWritable
:
874 case SpvDecorationNonReadable
:
875 case SpvDecorationUniform
:
876 case SpvDecorationLocation
:
877 case SpvDecorationComponent
:
878 case SpvDecorationOffset
:
879 case SpvDecorationXfbBuffer
:
880 case SpvDecorationXfbStride
:
881 case SpvDecorationHlslSemanticGOOGLE
:
882 vtn_warn("Decoration only allowed for struct members: %s",
883 spirv_decoration_to_string(dec
->decoration
));
886 case SpvDecorationStream
:
887 /* We don't need to do anything here, as stream is filled up when
888 * aplying the decoration to a variable, just check that if it is not a
889 * struct member, it should be a struct.
891 vtn_assert(type
->base_type
== vtn_base_type_struct
);
894 case SpvDecorationRelaxedPrecision
:
895 case SpvDecorationSpecId
:
896 case SpvDecorationInvariant
:
897 case SpvDecorationRestrict
:
898 case SpvDecorationAliased
:
899 case SpvDecorationConstant
:
900 case SpvDecorationIndex
:
901 case SpvDecorationBinding
:
902 case SpvDecorationDescriptorSet
:
903 case SpvDecorationLinkageAttributes
:
904 case SpvDecorationNoContraction
:
905 case SpvDecorationInputAttachmentIndex
:
906 vtn_warn("Decoration not allowed on types: %s",
907 spirv_decoration_to_string(dec
->decoration
));
910 case SpvDecorationCPacked
:
911 case SpvDecorationSaturatedConversion
:
912 case SpvDecorationFuncParamAttr
:
913 case SpvDecorationFPRoundingMode
:
914 case SpvDecorationFPFastMathMode
:
915 case SpvDecorationAlignment
:
916 vtn_warn("Decoration only allowed for CL-style kernels: %s",
917 spirv_decoration_to_string(dec
->decoration
));
921 vtn_fail("Unhandled decoration");
926 translate_image_format(struct vtn_builder
*b
, SpvImageFormat format
)
929 case SpvImageFormatUnknown
: return 0; /* GL_NONE */
930 case SpvImageFormatRgba32f
: return 0x8814; /* GL_RGBA32F */
931 case SpvImageFormatRgba16f
: return 0x881A; /* GL_RGBA16F */
932 case SpvImageFormatR32f
: return 0x822E; /* GL_R32F */
933 case SpvImageFormatRgba8
: return 0x8058; /* GL_RGBA8 */
934 case SpvImageFormatRgba8Snorm
: return 0x8F97; /* GL_RGBA8_SNORM */
935 case SpvImageFormatRg32f
: return 0x8230; /* GL_RG32F */
936 case SpvImageFormatRg16f
: return 0x822F; /* GL_RG16F */
937 case SpvImageFormatR11fG11fB10f
: return 0x8C3A; /* GL_R11F_G11F_B10F */
938 case SpvImageFormatR16f
: return 0x822D; /* GL_R16F */
939 case SpvImageFormatRgba16
: return 0x805B; /* GL_RGBA16 */
940 case SpvImageFormatRgb10A2
: return 0x8059; /* GL_RGB10_A2 */
941 case SpvImageFormatRg16
: return 0x822C; /* GL_RG16 */
942 case SpvImageFormatRg8
: return 0x822B; /* GL_RG8 */
943 case SpvImageFormatR16
: return 0x822A; /* GL_R16 */
944 case SpvImageFormatR8
: return 0x8229; /* GL_R8 */
945 case SpvImageFormatRgba16Snorm
: return 0x8F9B; /* GL_RGBA16_SNORM */
946 case SpvImageFormatRg16Snorm
: return 0x8F99; /* GL_RG16_SNORM */
947 case SpvImageFormatRg8Snorm
: return 0x8F95; /* GL_RG8_SNORM */
948 case SpvImageFormatR16Snorm
: return 0x8F98; /* GL_R16_SNORM */
949 case SpvImageFormatR8Snorm
: return 0x8F94; /* GL_R8_SNORM */
950 case SpvImageFormatRgba32i
: return 0x8D82; /* GL_RGBA32I */
951 case SpvImageFormatRgba16i
: return 0x8D88; /* GL_RGBA16I */
952 case SpvImageFormatRgba8i
: return 0x8D8E; /* GL_RGBA8I */
953 case SpvImageFormatR32i
: return 0x8235; /* GL_R32I */
954 case SpvImageFormatRg32i
: return 0x823B; /* GL_RG32I */
955 case SpvImageFormatRg16i
: return 0x8239; /* GL_RG16I */
956 case SpvImageFormatRg8i
: return 0x8237; /* GL_RG8I */
957 case SpvImageFormatR16i
: return 0x8233; /* GL_R16I */
958 case SpvImageFormatR8i
: return 0x8231; /* GL_R8I */
959 case SpvImageFormatRgba32ui
: return 0x8D70; /* GL_RGBA32UI */
960 case SpvImageFormatRgba16ui
: return 0x8D76; /* GL_RGBA16UI */
961 case SpvImageFormatRgba8ui
: return 0x8D7C; /* GL_RGBA8UI */
962 case SpvImageFormatR32ui
: return 0x8236; /* GL_R32UI */
963 case SpvImageFormatRgb10a2ui
: return 0x906F; /* GL_RGB10_A2UI */
964 case SpvImageFormatRg32ui
: return 0x823C; /* GL_RG32UI */
965 case SpvImageFormatRg16ui
: return 0x823A; /* GL_RG16UI */
966 case SpvImageFormatRg8ui
: return 0x8238; /* GL_RG8UI */
967 case SpvImageFormatR16ui
: return 0x8234; /* GL_R16UI */
968 case SpvImageFormatR8ui
: return 0x8232; /* GL_R8UI */
970 vtn_fail("Invalid image format");
974 static struct vtn_type
*
975 vtn_type_layout_std430(struct vtn_builder
*b
, struct vtn_type
*type
,
976 uint32_t *size_out
, uint32_t *align_out
)
978 switch (type
->base_type
) {
979 case vtn_base_type_scalar
: {
980 uint32_t comp_size
= glsl_get_bit_size(type
->type
) / 8;
981 *size_out
= comp_size
;
982 *align_out
= comp_size
;
986 case vtn_base_type_vector
: {
987 uint32_t comp_size
= glsl_get_bit_size(type
->type
) / 8;
988 unsigned align_comps
= type
->length
== 3 ? 4 : type
->length
;
989 *size_out
= comp_size
* type
->length
,
990 *align_out
= comp_size
* align_comps
;
994 case vtn_base_type_matrix
:
995 case vtn_base_type_array
: {
996 /* We're going to add an array stride */
997 type
= vtn_type_copy(b
, type
);
998 uint32_t elem_size
, elem_align
;
999 type
->array_element
= vtn_type_layout_std430(b
, type
->array_element
,
1000 &elem_size
, &elem_align
);
1001 type
->stride
= vtn_align_u32(elem_size
, elem_align
);
1002 *size_out
= type
->stride
* type
->length
;
1003 *align_out
= elem_align
;
1007 case vtn_base_type_struct
: {
1008 /* We're going to add member offsets */
1009 type
= vtn_type_copy(b
, type
);
1010 uint32_t offset
= 0;
1012 for (unsigned i
= 0; i
< type
->length
; i
++) {
1013 uint32_t mem_size
, mem_align
;
1014 type
->members
[i
] = vtn_type_layout_std430(b
, type
->members
[i
],
1015 &mem_size
, &mem_align
);
1016 offset
= vtn_align_u32(offset
, mem_align
);
1017 type
->offsets
[i
] = offset
;
1019 align
= MAX2(align
, mem_align
);
1027 unreachable("Invalid SPIR-V type for std430");
1032 vtn_handle_type(struct vtn_builder
*b
, SpvOp opcode
,
1033 const uint32_t *w
, unsigned count
)
1035 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_type
);
1037 val
->type
= rzalloc(b
, struct vtn_type
);
1038 val
->type
->id
= w
[1];
1042 val
->type
->base_type
= vtn_base_type_void
;
1043 val
->type
->type
= glsl_void_type();
1046 val
->type
->base_type
= vtn_base_type_scalar
;
1047 val
->type
->type
= glsl_bool_type();
1048 val
->type
->length
= 1;
1050 case SpvOpTypeInt
: {
1051 int bit_size
= w
[2];
1052 const bool signedness
= w
[3];
1053 val
->type
->base_type
= vtn_base_type_scalar
;
1056 val
->type
->type
= (signedness
? glsl_int64_t_type() : glsl_uint64_t_type());
1059 val
->type
->type
= (signedness
? glsl_int_type() : glsl_uint_type());
1062 val
->type
->type
= (signedness
? glsl_int16_t_type() : glsl_uint16_t_type());
1065 val
->type
->type
= (signedness
? glsl_int8_t_type() : glsl_uint8_t_type());
1068 vtn_fail("Invalid int bit size");
1070 val
->type
->length
= 1;
1074 case SpvOpTypeFloat
: {
1075 int bit_size
= w
[2];
1076 val
->type
->base_type
= vtn_base_type_scalar
;
1079 val
->type
->type
= glsl_float16_t_type();
1082 val
->type
->type
= glsl_float_type();
1085 val
->type
->type
= glsl_double_type();
1088 vtn_fail("Invalid float bit size");
1090 val
->type
->length
= 1;
1094 case SpvOpTypeVector
: {
1095 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1096 unsigned elems
= w
[3];
1098 vtn_fail_if(base
->base_type
!= vtn_base_type_scalar
,
1099 "Base type for OpTypeVector must be a scalar");
1100 vtn_fail_if((elems
< 2 || elems
> 4) && (elems
!= 8) && (elems
!= 16),
1101 "Invalid component count for OpTypeVector");
1103 val
->type
->base_type
= vtn_base_type_vector
;
1104 val
->type
->type
= glsl_vector_type(glsl_get_base_type(base
->type
), elems
);
1105 val
->type
->length
= elems
;
1106 val
->type
->stride
= glsl_get_bit_size(base
->type
) / 8;
1107 val
->type
->array_element
= base
;
1111 case SpvOpTypeMatrix
: {
1112 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1113 unsigned columns
= w
[3];
1115 vtn_fail_if(base
->base_type
!= vtn_base_type_vector
,
1116 "Base type for OpTypeMatrix must be a vector");
1117 vtn_fail_if(columns
< 2 || columns
> 4,
1118 "Invalid column count for OpTypeMatrix");
1120 val
->type
->base_type
= vtn_base_type_matrix
;
1121 val
->type
->type
= glsl_matrix_type(glsl_get_base_type(base
->type
),
1122 glsl_get_vector_elements(base
->type
),
1124 vtn_fail_if(glsl_type_is_error(val
->type
->type
),
1125 "Unsupported base type for OpTypeMatrix");
1126 assert(!glsl_type_is_error(val
->type
->type
));
1127 val
->type
->length
= columns
;
1128 val
->type
->array_element
= base
;
1129 val
->type
->row_major
= false;
1130 val
->type
->stride
= 0;
1134 case SpvOpTypeRuntimeArray
:
1135 case SpvOpTypeArray
: {
1136 struct vtn_type
*array_element
=
1137 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1139 if (opcode
== SpvOpTypeRuntimeArray
) {
1140 /* A length of 0 is used to denote unsized arrays */
1141 val
->type
->length
= 0;
1144 vtn_value(b
, w
[3], vtn_value_type_constant
)->constant
->values
[0].u32
[0];
1147 val
->type
->base_type
= vtn_base_type_array
;
1148 val
->type
->type
= glsl_array_type(array_element
->type
, val
->type
->length
);
1149 val
->type
->array_element
= array_element
;
1150 val
->type
->stride
= 0;
1154 case SpvOpTypeStruct
: {
1155 unsigned num_fields
= count
- 2;
1156 val
->type
->base_type
= vtn_base_type_struct
;
1157 val
->type
->length
= num_fields
;
1158 val
->type
->members
= ralloc_array(b
, struct vtn_type
*, num_fields
);
1159 val
->type
->offsets
= ralloc_array(b
, unsigned, num_fields
);
1161 NIR_VLA(struct glsl_struct_field
, fields
, count
);
1162 for (unsigned i
= 0; i
< num_fields
; i
++) {
1163 val
->type
->members
[i
] =
1164 vtn_value(b
, w
[i
+ 2], vtn_value_type_type
)->type
;
1165 fields
[i
] = (struct glsl_struct_field
) {
1166 .type
= val
->type
->members
[i
]->type
,
1167 .name
= ralloc_asprintf(b
, "field%d", i
),
1172 struct member_decoration_ctx ctx
= {
1173 .num_fields
= num_fields
,
1178 vtn_foreach_decoration(b
, val
, struct_member_decoration_cb
, &ctx
);
1179 vtn_foreach_decoration(b
, val
, struct_member_matrix_stride_cb
, &ctx
);
1181 const char *name
= val
->name
? val
->name
: "struct";
1183 val
->type
->type
= glsl_struct_type(fields
, num_fields
, name
);
1187 case SpvOpTypeFunction
: {
1188 val
->type
->base_type
= vtn_base_type_function
;
1189 val
->type
->type
= NULL
;
1191 val
->type
->return_type
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1193 const unsigned num_params
= count
- 3;
1194 val
->type
->length
= num_params
;
1195 val
->type
->params
= ralloc_array(b
, struct vtn_type
*, num_params
);
1196 for (unsigned i
= 0; i
< count
- 3; i
++) {
1197 val
->type
->params
[i
] =
1198 vtn_value(b
, w
[i
+ 3], vtn_value_type_type
)->type
;
1203 case SpvOpTypePointer
: {
1204 SpvStorageClass storage_class
= w
[2];
1205 struct vtn_type
*deref_type
=
1206 vtn_value(b
, w
[3], vtn_value_type_type
)->type
;
1208 val
->type
->base_type
= vtn_base_type_pointer
;
1209 val
->type
->storage_class
= storage_class
;
1210 val
->type
->deref
= deref_type
;
1212 if (storage_class
== SpvStorageClassUniform
||
1213 storage_class
== SpvStorageClassStorageBuffer
) {
1214 /* These can actually be stored to nir_variables and used as SSA
1215 * values so they need a real glsl_type.
1217 val
->type
->type
= glsl_vector_type(GLSL_TYPE_UINT
, 2);
1220 if (storage_class
== SpvStorageClassPushConstant
) {
1221 /* These can actually be stored to nir_variables and used as SSA
1222 * values so they need a real glsl_type.
1224 val
->type
->type
= glsl_uint_type();
1227 if (storage_class
== SpvStorageClassWorkgroup
) {
1228 /* These can actually be stored to nir_variables and used as SSA
1229 * values so they need a real glsl_type.
1231 val
->type
->type
= glsl_uint_type();
1232 if (b
->options
->lower_workgroup_access_to_offsets
) {
1233 uint32_t size
, align
;
1234 val
->type
->deref
= vtn_type_layout_std430(b
, val
->type
->deref
,
1236 val
->type
->length
= size
;
1237 val
->type
->align
= align
;
1243 case SpvOpTypeImage
: {
1244 val
->type
->base_type
= vtn_base_type_image
;
1246 const struct vtn_type
*sampled_type
=
1247 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1249 vtn_fail_if(sampled_type
->base_type
!= vtn_base_type_scalar
||
1250 glsl_get_bit_size(sampled_type
->type
) != 32,
1251 "Sampled type of OpTypeImage must be a 32-bit scalar");
1253 enum glsl_sampler_dim dim
;
1254 switch ((SpvDim
)w
[3]) {
1255 case SpvDim1D
: dim
= GLSL_SAMPLER_DIM_1D
; break;
1256 case SpvDim2D
: dim
= GLSL_SAMPLER_DIM_2D
; break;
1257 case SpvDim3D
: dim
= GLSL_SAMPLER_DIM_3D
; break;
1258 case SpvDimCube
: dim
= GLSL_SAMPLER_DIM_CUBE
; break;
1259 case SpvDimRect
: dim
= GLSL_SAMPLER_DIM_RECT
; break;
1260 case SpvDimBuffer
: dim
= GLSL_SAMPLER_DIM_BUF
; break;
1261 case SpvDimSubpassData
: dim
= GLSL_SAMPLER_DIM_SUBPASS
; break;
1263 vtn_fail("Invalid SPIR-V image dimensionality");
1266 bool is_shadow
= w
[4];
1267 bool is_array
= w
[5];
1268 bool multisampled
= w
[6];
1269 unsigned sampled
= w
[7];
1270 SpvImageFormat format
= w
[8];
1273 val
->type
->access_qualifier
= w
[9];
1275 val
->type
->access_qualifier
= SpvAccessQualifierReadWrite
;
1278 if (dim
== GLSL_SAMPLER_DIM_2D
)
1279 dim
= GLSL_SAMPLER_DIM_MS
;
1280 else if (dim
== GLSL_SAMPLER_DIM_SUBPASS
)
1281 dim
= GLSL_SAMPLER_DIM_SUBPASS_MS
;
1283 vtn_fail("Unsupported multisampled image type");
1286 val
->type
->image_format
= translate_image_format(b
, format
);
1288 enum glsl_base_type sampled_base_type
=
1289 glsl_get_base_type(sampled_type
->type
);
1291 val
->type
->sampled
= true;
1292 val
->type
->type
= glsl_sampler_type(dim
, is_shadow
, is_array
,
1294 } else if (sampled
== 2) {
1295 vtn_assert(!is_shadow
);
1296 val
->type
->sampled
= false;
1297 val
->type
->type
= glsl_image_type(dim
, is_array
, sampled_base_type
);
1299 vtn_fail("We need to know if the image will be sampled");
1304 case SpvOpTypeSampledImage
:
1305 val
->type
->base_type
= vtn_base_type_sampled_image
;
1306 val
->type
->image
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1307 val
->type
->type
= val
->type
->image
->type
;
1310 case SpvOpTypeSampler
:
1311 /* The actual sampler type here doesn't really matter. It gets
1312 * thrown away the moment you combine it with an image. What really
1313 * matters is that it's a sampler type as opposed to an integer type
1314 * so the backend knows what to do.
1316 val
->type
->base_type
= vtn_base_type_sampler
;
1317 val
->type
->type
= glsl_bare_sampler_type();
1320 case SpvOpTypeOpaque
:
1321 case SpvOpTypeEvent
:
1322 case SpvOpTypeDeviceEvent
:
1323 case SpvOpTypeReserveId
:
1324 case SpvOpTypeQueue
:
1327 vtn_fail("Unhandled opcode");
1330 vtn_foreach_decoration(b
, val
, type_decoration_cb
, NULL
);
1333 static nir_constant
*
1334 vtn_null_constant(struct vtn_builder
*b
, const struct glsl_type
*type
)
1336 nir_constant
*c
= rzalloc(b
, nir_constant
);
1338 /* For pointers and other typeless things, we have to return something but
1339 * it doesn't matter what.
1344 switch (glsl_get_base_type(type
)) {
1346 case GLSL_TYPE_UINT
:
1347 case GLSL_TYPE_INT16
:
1348 case GLSL_TYPE_UINT16
:
1349 case GLSL_TYPE_UINT8
:
1350 case GLSL_TYPE_INT8
:
1351 case GLSL_TYPE_INT64
:
1352 case GLSL_TYPE_UINT64
:
1353 case GLSL_TYPE_BOOL
:
1354 case GLSL_TYPE_FLOAT
:
1355 case GLSL_TYPE_FLOAT16
:
1356 case GLSL_TYPE_DOUBLE
:
1357 /* Nothing to do here. It's already initialized to zero */
1360 case GLSL_TYPE_ARRAY
:
1361 vtn_assert(glsl_get_length(type
) > 0);
1362 c
->num_elements
= glsl_get_length(type
);
1363 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1365 c
->elements
[0] = vtn_null_constant(b
, glsl_get_array_element(type
));
1366 for (unsigned i
= 1; i
< c
->num_elements
; i
++)
1367 c
->elements
[i
] = c
->elements
[0];
1370 case GLSL_TYPE_STRUCT
:
1371 c
->num_elements
= glsl_get_length(type
);
1372 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1374 for (unsigned i
= 0; i
< c
->num_elements
; i
++) {
1375 c
->elements
[i
] = vtn_null_constant(b
, glsl_get_struct_field(type
, i
));
1380 vtn_fail("Invalid type for null constant");
1387 spec_constant_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*v
,
1388 int member
, const struct vtn_decoration
*dec
,
1391 vtn_assert(member
== -1);
1392 if (dec
->decoration
!= SpvDecorationSpecId
)
1395 struct spec_constant_value
*const_value
= data
;
1397 for (unsigned i
= 0; i
< b
->num_specializations
; i
++) {
1398 if (b
->specializations
[i
].id
== dec
->literals
[0]) {
1399 if (const_value
->is_double
)
1400 const_value
->data64
= b
->specializations
[i
].data64
;
1402 const_value
->data32
= b
->specializations
[i
].data32
;
1409 get_specialization(struct vtn_builder
*b
, struct vtn_value
*val
,
1410 uint32_t const_value
)
1412 struct spec_constant_value data
;
1413 data
.is_double
= false;
1414 data
.data32
= const_value
;
1415 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1420 get_specialization64(struct vtn_builder
*b
, struct vtn_value
*val
,
1421 uint64_t const_value
)
1423 struct spec_constant_value data
;
1424 data
.is_double
= true;
1425 data
.data64
= const_value
;
1426 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1431 handle_workgroup_size_decoration_cb(struct vtn_builder
*b
,
1432 struct vtn_value
*val
,
1434 const struct vtn_decoration
*dec
,
1437 vtn_assert(member
== -1);
1438 if (dec
->decoration
!= SpvDecorationBuiltIn
||
1439 dec
->literals
[0] != SpvBuiltInWorkgroupSize
)
1442 vtn_assert(val
->type
->type
== glsl_vector_type(GLSL_TYPE_UINT
, 3));
1444 b
->shader
->info
.cs
.local_size
[0] = val
->constant
->values
[0].u32
[0];
1445 b
->shader
->info
.cs
.local_size
[1] = val
->constant
->values
[0].u32
[1];
1446 b
->shader
->info
.cs
.local_size
[2] = val
->constant
->values
[0].u32
[2];
1450 vtn_handle_constant(struct vtn_builder
*b
, SpvOp opcode
,
1451 const uint32_t *w
, unsigned count
)
1453 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_constant
);
1454 val
->constant
= rzalloc(b
, nir_constant
);
1456 case SpvOpConstantTrue
:
1457 case SpvOpConstantFalse
:
1458 case SpvOpSpecConstantTrue
:
1459 case SpvOpSpecConstantFalse
: {
1460 vtn_fail_if(val
->type
->type
!= glsl_bool_type(),
1461 "Result type of %s must be OpTypeBool",
1462 spirv_op_to_string(opcode
));
1464 uint32_t int_val
= (opcode
== SpvOpConstantTrue
||
1465 opcode
== SpvOpSpecConstantTrue
);
1467 if (opcode
== SpvOpSpecConstantTrue
||
1468 opcode
== SpvOpSpecConstantFalse
)
1469 int_val
= get_specialization(b
, val
, int_val
);
1471 val
->constant
->values
[0].u32
[0] = int_val
? NIR_TRUE
: NIR_FALSE
;
1475 case SpvOpConstant
: {
1476 vtn_fail_if(val
->type
->base_type
!= vtn_base_type_scalar
,
1477 "Result type of %s must be a scalar",
1478 spirv_op_to_string(opcode
));
1479 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1482 val
->constant
->values
->u64
[0] = vtn_u64_literal(&w
[3]);
1485 val
->constant
->values
->u32
[0] = w
[3];
1488 val
->constant
->values
->u16
[0] = w
[3];
1491 val
->constant
->values
->u8
[0] = w
[3];
1494 vtn_fail("Unsupported SpvOpConstant bit size");
1499 case SpvOpSpecConstant
: {
1500 vtn_fail_if(val
->type
->base_type
!= vtn_base_type_scalar
,
1501 "Result type of %s must be a scalar",
1502 spirv_op_to_string(opcode
));
1503 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1506 val
->constant
->values
[0].u64
[0] =
1507 get_specialization64(b
, val
, vtn_u64_literal(&w
[3]));
1510 val
->constant
->values
[0].u32
[0] = get_specialization(b
, val
, w
[3]);
1513 val
->constant
->values
[0].u16
[0] = get_specialization(b
, val
, w
[3]);
1516 val
->constant
->values
[0].u8
[0] = get_specialization(b
, val
, w
[3]);
1519 vtn_fail("Unsupported SpvOpSpecConstant bit size");
1524 case SpvOpSpecConstantComposite
:
1525 case SpvOpConstantComposite
: {
1526 unsigned elem_count
= count
- 3;
1527 vtn_fail_if(elem_count
!= val
->type
->length
,
1528 "%s has %u constituents, expected %u",
1529 spirv_op_to_string(opcode
), elem_count
, val
->type
->length
);
1531 nir_constant
**elems
= ralloc_array(b
, nir_constant
*, elem_count
);
1532 for (unsigned i
= 0; i
< elem_count
; i
++) {
1533 struct vtn_value
*val
= vtn_untyped_value(b
, w
[i
+ 3]);
1535 if (val
->value_type
== vtn_value_type_constant
) {
1536 elems
[i
] = val
->constant
;
1538 vtn_fail_if(val
->value_type
!= vtn_value_type_undef
,
1539 "only constants or undefs allowed for "
1540 "SpvOpConstantComposite");
1541 /* to make it easier, just insert a NULL constant for now */
1542 elems
[i
] = vtn_null_constant(b
, val
->type
->type
);
1546 switch (val
->type
->base_type
) {
1547 case vtn_base_type_vector
: {
1548 assert(glsl_type_is_vector(val
->type
->type
));
1549 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1550 for (unsigned i
= 0; i
< elem_count
; i
++) {
1553 val
->constant
->values
[0].u64
[i
] = elems
[i
]->values
[0].u64
[0];
1556 val
->constant
->values
[0].u32
[i
] = elems
[i
]->values
[0].u32
[0];
1559 val
->constant
->values
[0].u16
[i
] = elems
[i
]->values
[0].u16
[0];
1562 val
->constant
->values
[0].u8
[i
] = elems
[i
]->values
[0].u8
[0];
1565 vtn_fail("Invalid SpvOpConstantComposite bit size");
1571 case vtn_base_type_matrix
:
1572 assert(glsl_type_is_matrix(val
->type
->type
));
1573 for (unsigned i
= 0; i
< elem_count
; i
++)
1574 val
->constant
->values
[i
] = elems
[i
]->values
[0];
1577 case vtn_base_type_struct
:
1578 case vtn_base_type_array
:
1579 ralloc_steal(val
->constant
, elems
);
1580 val
->constant
->num_elements
= elem_count
;
1581 val
->constant
->elements
= elems
;
1585 vtn_fail("Result type of %s must be a composite type",
1586 spirv_op_to_string(opcode
));
1591 case SpvOpSpecConstantOp
: {
1592 SpvOp opcode
= get_specialization(b
, val
, w
[3]);
1594 case SpvOpVectorShuffle
: {
1595 struct vtn_value
*v0
= &b
->values
[w
[4]];
1596 struct vtn_value
*v1
= &b
->values
[w
[5]];
1598 vtn_assert(v0
->value_type
== vtn_value_type_constant
||
1599 v0
->value_type
== vtn_value_type_undef
);
1600 vtn_assert(v1
->value_type
== vtn_value_type_constant
||
1601 v1
->value_type
== vtn_value_type_undef
);
1603 unsigned len0
= glsl_get_vector_elements(v0
->type
->type
);
1604 unsigned len1
= glsl_get_vector_elements(v1
->type
->type
);
1606 vtn_assert(len0
+ len1
< 16);
1608 unsigned bit_size
= glsl_get_bit_size(val
->type
->type
);
1609 unsigned bit_size0
= glsl_get_bit_size(v0
->type
->type
);
1610 unsigned bit_size1
= glsl_get_bit_size(v1
->type
->type
);
1612 vtn_assert(bit_size
== bit_size0
&& bit_size
== bit_size1
);
1613 (void)bit_size0
; (void)bit_size1
;
1615 if (bit_size
== 64) {
1617 if (v0
->value_type
== vtn_value_type_constant
) {
1618 for (unsigned i
= 0; i
< len0
; i
++)
1619 u64
[i
] = v0
->constant
->values
[0].u64
[i
];
1621 if (v1
->value_type
== vtn_value_type_constant
) {
1622 for (unsigned i
= 0; i
< len1
; i
++)
1623 u64
[len0
+ i
] = v1
->constant
->values
[0].u64
[i
];
1626 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1627 uint32_t comp
= w
[i
+ 6];
1628 /* If component is not used, set the value to a known constant
1629 * to detect if it is wrongly used.
1631 if (comp
== (uint32_t)-1)
1632 val
->constant
->values
[0].u64
[j
] = 0xdeadbeefdeadbeef;
1634 val
->constant
->values
[0].u64
[j
] = u64
[comp
];
1637 /* This is for both 32-bit and 16-bit values */
1639 if (v0
->value_type
== vtn_value_type_constant
) {
1640 for (unsigned i
= 0; i
< len0
; i
++)
1641 u32
[i
] = v0
->constant
->values
[0].u32
[i
];
1643 if (v1
->value_type
== vtn_value_type_constant
) {
1644 for (unsigned i
= 0; i
< len1
; i
++)
1645 u32
[len0
+ i
] = v1
->constant
->values
[0].u32
[i
];
1648 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1649 uint32_t comp
= w
[i
+ 6];
1650 /* If component is not used, set the value to a known constant
1651 * to detect if it is wrongly used.
1653 if (comp
== (uint32_t)-1)
1654 val
->constant
->values
[0].u32
[j
] = 0xdeadbeef;
1656 val
->constant
->values
[0].u32
[j
] = u32
[comp
];
1662 case SpvOpCompositeExtract
:
1663 case SpvOpCompositeInsert
: {
1664 struct vtn_value
*comp
;
1665 unsigned deref_start
;
1666 struct nir_constant
**c
;
1667 if (opcode
== SpvOpCompositeExtract
) {
1668 comp
= vtn_value(b
, w
[4], vtn_value_type_constant
);
1670 c
= &comp
->constant
;
1672 comp
= vtn_value(b
, w
[5], vtn_value_type_constant
);
1674 val
->constant
= nir_constant_clone(comp
->constant
,
1681 const struct vtn_type
*type
= comp
->type
;
1682 for (unsigned i
= deref_start
; i
< count
; i
++) {
1683 vtn_fail_if(w
[i
] > type
->length
,
1684 "%uth index of %s is %u but the type has only "
1685 "%u elements", i
- deref_start
,
1686 spirv_op_to_string(opcode
), w
[i
], type
->length
);
1688 switch (type
->base_type
) {
1689 case vtn_base_type_vector
:
1691 type
= type
->array_element
;
1694 case vtn_base_type_matrix
:
1695 assert(col
== 0 && elem
== -1);
1698 type
= type
->array_element
;
1701 case vtn_base_type_array
:
1702 c
= &(*c
)->elements
[w
[i
]];
1703 type
= type
->array_element
;
1706 case vtn_base_type_struct
:
1707 c
= &(*c
)->elements
[w
[i
]];
1708 type
= type
->members
[w
[i
]];
1712 vtn_fail("%s must only index into composite types",
1713 spirv_op_to_string(opcode
));
1717 if (opcode
== SpvOpCompositeExtract
) {
1721 unsigned num_components
= type
->length
;
1722 unsigned bit_size
= glsl_get_bit_size(type
->type
);
1723 for (unsigned i
= 0; i
< num_components
; i
++)
1726 val
->constant
->values
[0].u64
[i
] = (*c
)->values
[col
].u64
[elem
+ i
];
1729 val
->constant
->values
[0].u32
[i
] = (*c
)->values
[col
].u32
[elem
+ i
];
1732 val
->constant
->values
[0].u16
[i
] = (*c
)->values
[col
].u16
[elem
+ i
];
1735 val
->constant
->values
[0].u8
[i
] = (*c
)->values
[col
].u8
[elem
+ i
];
1738 vtn_fail("Invalid SpvOpCompositeExtract bit size");
1742 struct vtn_value
*insert
=
1743 vtn_value(b
, w
[4], vtn_value_type_constant
);
1744 vtn_assert(insert
->type
== type
);
1746 *c
= insert
->constant
;
1748 unsigned num_components
= type
->length
;
1749 unsigned bit_size
= glsl_get_bit_size(type
->type
);
1750 for (unsigned i
= 0; i
< num_components
; i
++)
1753 (*c
)->values
[col
].u64
[elem
+ i
] = insert
->constant
->values
[0].u64
[i
];
1756 (*c
)->values
[col
].u32
[elem
+ i
] = insert
->constant
->values
[0].u32
[i
];
1759 (*c
)->values
[col
].u16
[elem
+ i
] = insert
->constant
->values
[0].u16
[i
];
1762 (*c
)->values
[col
].u8
[elem
+ i
] = insert
->constant
->values
[0].u8
[i
];
1765 vtn_fail("Invalid SpvOpCompositeInsert bit size");
1774 nir_alu_type dst_alu_type
= nir_get_nir_type_for_glsl_type(val
->type
->type
);
1775 nir_alu_type src_alu_type
= dst_alu_type
;
1776 unsigned num_components
= glsl_get_vector_elements(val
->type
->type
);
1779 vtn_assert(count
<= 7);
1784 /* We have a source in a conversion */
1786 nir_get_nir_type_for_glsl_type(
1787 vtn_value(b
, w
[4], vtn_value_type_constant
)->type
->type
);
1788 /* We use the bitsize of the conversion source to evaluate the opcode later */
1789 bit_size
= glsl_get_bit_size(
1790 vtn_value(b
, w
[4], vtn_value_type_constant
)->type
->type
);
1793 bit_size
= glsl_get_bit_size(val
->type
->type
);
1796 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(b
, opcode
, &swap
,
1797 nir_alu_type_get_type_size(src_alu_type
),
1798 nir_alu_type_get_type_size(dst_alu_type
));
1799 nir_const_value src
[4];
1801 for (unsigned i
= 0; i
< count
- 4; i
++) {
1802 struct vtn_value
*src_val
=
1803 vtn_value(b
, w
[4 + i
], vtn_value_type_constant
);
1805 /* If this is an unsized source, pull the bit size from the
1806 * source; otherwise, we'll use the bit size from the destination.
1808 if (!nir_alu_type_get_type_size(nir_op_infos
[op
].input_types
[i
]))
1809 bit_size
= glsl_get_bit_size(src_val
->type
->type
);
1811 unsigned j
= swap
? 1 - i
: i
;
1812 src
[j
] = src_val
->constant
->values
[0];
1815 /* fix up fixed size sources */
1822 for (unsigned i
= 0; i
< num_components
; ++i
) {
1824 case 64: src
[1].u32
[i
] = src
[1].u64
[i
]; break;
1825 case 16: src
[1].u32
[i
] = src
[1].u16
[i
]; break;
1826 case 8: src
[1].u32
[i
] = src
[1].u8
[i
]; break;
1835 val
->constant
->values
[0] =
1836 nir_eval_const_opcode(op
, num_components
, bit_size
, src
);
1843 case SpvOpConstantNull
:
1844 val
->constant
= vtn_null_constant(b
, val
->type
->type
);
1847 case SpvOpConstantSampler
:
1848 vtn_fail("OpConstantSampler requires Kernel Capability");
1852 vtn_fail("Unhandled opcode");
1855 /* Now that we have the value, update the workgroup size if needed */
1856 vtn_foreach_decoration(b
, val
, handle_workgroup_size_decoration_cb
, NULL
);
1859 struct vtn_ssa_value
*
1860 vtn_create_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
1862 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
1865 if (!glsl_type_is_vector_or_scalar(type
)) {
1866 unsigned elems
= glsl_get_length(type
);
1867 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
1868 for (unsigned i
= 0; i
< elems
; i
++) {
1869 const struct glsl_type
*child_type
;
1871 switch (glsl_get_base_type(type
)) {
1873 case GLSL_TYPE_UINT
:
1874 case GLSL_TYPE_INT16
:
1875 case GLSL_TYPE_UINT16
:
1876 case GLSL_TYPE_UINT8
:
1877 case GLSL_TYPE_INT8
:
1878 case GLSL_TYPE_INT64
:
1879 case GLSL_TYPE_UINT64
:
1880 case GLSL_TYPE_BOOL
:
1881 case GLSL_TYPE_FLOAT
:
1882 case GLSL_TYPE_FLOAT16
:
1883 case GLSL_TYPE_DOUBLE
:
1884 child_type
= glsl_get_column_type(type
);
1886 case GLSL_TYPE_ARRAY
:
1887 child_type
= glsl_get_array_element(type
);
1889 case GLSL_TYPE_STRUCT
:
1890 child_type
= glsl_get_struct_field(type
, i
);
1893 vtn_fail("unkown base type");
1896 val
->elems
[i
] = vtn_create_ssa_value(b
, child_type
);
1904 vtn_tex_src(struct vtn_builder
*b
, unsigned index
, nir_tex_src_type type
)
1907 src
.src
= nir_src_for_ssa(vtn_ssa_value(b
, index
)->def
);
1908 src
.src_type
= type
;
1913 vtn_handle_texture(struct vtn_builder
*b
, SpvOp opcode
,
1914 const uint32_t *w
, unsigned count
)
1916 if (opcode
== SpvOpSampledImage
) {
1917 struct vtn_value
*val
=
1918 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
1919 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
1920 val
->sampled_image
->type
=
1921 vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1922 val
->sampled_image
->image
=
1923 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
1924 val
->sampled_image
->sampler
=
1925 vtn_value(b
, w
[4], vtn_value_type_pointer
)->pointer
;
1927 } else if (opcode
== SpvOpImage
) {
1928 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
1929 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[3]);
1930 if (src_val
->value_type
== vtn_value_type_sampled_image
) {
1931 val
->pointer
= src_val
->sampled_image
->image
;
1933 vtn_assert(src_val
->value_type
== vtn_value_type_pointer
);
1934 val
->pointer
= src_val
->pointer
;
1939 struct vtn_type
*ret_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1940 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1942 struct vtn_sampled_image sampled
;
1943 struct vtn_value
*sampled_val
= vtn_untyped_value(b
, w
[3]);
1944 if (sampled_val
->value_type
== vtn_value_type_sampled_image
) {
1945 sampled
= *sampled_val
->sampled_image
;
1947 vtn_assert(sampled_val
->value_type
== vtn_value_type_pointer
);
1948 sampled
.type
= sampled_val
->pointer
->type
;
1949 sampled
.image
= NULL
;
1950 sampled
.sampler
= sampled_val
->pointer
;
1953 const struct glsl_type
*image_type
= sampled
.type
->type
;
1954 const enum glsl_sampler_dim sampler_dim
= glsl_get_sampler_dim(image_type
);
1955 const bool is_array
= glsl_sampler_type_is_array(image_type
);
1957 /* Figure out the base texture operation */
1960 case SpvOpImageSampleImplicitLod
:
1961 case SpvOpImageSampleDrefImplicitLod
:
1962 case SpvOpImageSampleProjImplicitLod
:
1963 case SpvOpImageSampleProjDrefImplicitLod
:
1964 texop
= nir_texop_tex
;
1967 case SpvOpImageSampleExplicitLod
:
1968 case SpvOpImageSampleDrefExplicitLod
:
1969 case SpvOpImageSampleProjExplicitLod
:
1970 case SpvOpImageSampleProjDrefExplicitLod
:
1971 texop
= nir_texop_txl
;
1974 case SpvOpImageFetch
:
1975 if (glsl_get_sampler_dim(image_type
) == GLSL_SAMPLER_DIM_MS
) {
1976 texop
= nir_texop_txf_ms
;
1978 texop
= nir_texop_txf
;
1982 case SpvOpImageGather
:
1983 case SpvOpImageDrefGather
:
1984 texop
= nir_texop_tg4
;
1987 case SpvOpImageQuerySizeLod
:
1988 case SpvOpImageQuerySize
:
1989 texop
= nir_texop_txs
;
1992 case SpvOpImageQueryLod
:
1993 texop
= nir_texop_lod
;
1996 case SpvOpImageQueryLevels
:
1997 texop
= nir_texop_query_levels
;
2000 case SpvOpImageQuerySamples
:
2001 texop
= nir_texop_texture_samples
;
2005 vtn_fail("Unhandled opcode");
2008 nir_tex_src srcs
[10]; /* 10 should be enough */
2009 nir_tex_src
*p
= srcs
;
2011 nir_deref_instr
*sampler
= vtn_pointer_to_deref(b
, sampled
.sampler
);
2012 nir_deref_instr
*texture
=
2013 sampled
.image
? vtn_pointer_to_deref(b
, sampled
.image
) : sampler
;
2015 p
->src
= nir_src_for_ssa(&texture
->dest
.ssa
);
2016 p
->src_type
= nir_tex_src_texture_deref
;
2025 /* These operations require a sampler */
2026 p
->src
= nir_src_for_ssa(&sampler
->dest
.ssa
);
2027 p
->src_type
= nir_tex_src_sampler_deref
;
2031 case nir_texop_txf_ms
:
2034 case nir_texop_query_levels
:
2035 case nir_texop_texture_samples
:
2036 case nir_texop_samples_identical
:
2039 case nir_texop_txf_ms_mcs
:
2040 vtn_fail("unexpected nir_texop_txf_ms_mcs");
2045 struct nir_ssa_def
*coord
;
2046 unsigned coord_components
;
2048 case SpvOpImageSampleImplicitLod
:
2049 case SpvOpImageSampleExplicitLod
:
2050 case SpvOpImageSampleDrefImplicitLod
:
2051 case SpvOpImageSampleDrefExplicitLod
:
2052 case SpvOpImageSampleProjImplicitLod
:
2053 case SpvOpImageSampleProjExplicitLod
:
2054 case SpvOpImageSampleProjDrefImplicitLod
:
2055 case SpvOpImageSampleProjDrefExplicitLod
:
2056 case SpvOpImageFetch
:
2057 case SpvOpImageGather
:
2058 case SpvOpImageDrefGather
:
2059 case SpvOpImageQueryLod
: {
2060 /* All these types have the coordinate as their first real argument */
2061 switch (sampler_dim
) {
2062 case GLSL_SAMPLER_DIM_1D
:
2063 case GLSL_SAMPLER_DIM_BUF
:
2064 coord_components
= 1;
2066 case GLSL_SAMPLER_DIM_2D
:
2067 case GLSL_SAMPLER_DIM_RECT
:
2068 case GLSL_SAMPLER_DIM_MS
:
2069 coord_components
= 2;
2071 case GLSL_SAMPLER_DIM_3D
:
2072 case GLSL_SAMPLER_DIM_CUBE
:
2073 coord_components
= 3;
2076 vtn_fail("Invalid sampler type");
2079 if (is_array
&& texop
!= nir_texop_lod
)
2082 coord
= vtn_ssa_value(b
, w
[idx
++])->def
;
2083 p
->src
= nir_src_for_ssa(nir_channels(&b
->nb
, coord
,
2084 (1 << coord_components
) - 1));
2085 p
->src_type
= nir_tex_src_coord
;
2092 coord_components
= 0;
2097 case SpvOpImageSampleProjImplicitLod
:
2098 case SpvOpImageSampleProjExplicitLod
:
2099 case SpvOpImageSampleProjDrefImplicitLod
:
2100 case SpvOpImageSampleProjDrefExplicitLod
:
2101 /* These have the projector as the last coordinate component */
2102 p
->src
= nir_src_for_ssa(nir_channel(&b
->nb
, coord
, coord_components
));
2103 p
->src_type
= nir_tex_src_projector
;
2111 bool is_shadow
= false;
2112 unsigned gather_component
= 0;
2114 case SpvOpImageSampleDrefImplicitLod
:
2115 case SpvOpImageSampleDrefExplicitLod
:
2116 case SpvOpImageSampleProjDrefImplicitLod
:
2117 case SpvOpImageSampleProjDrefExplicitLod
:
2118 case SpvOpImageDrefGather
:
2119 /* These all have an explicit depth value as their next source */
2121 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_comparator
);
2124 case SpvOpImageGather
:
2125 /* This has a component as its next source */
2127 vtn_value(b
, w
[idx
++], vtn_value_type_constant
)->constant
->values
[0].u32
[0];
2134 /* For OpImageQuerySizeLod, we always have an LOD */
2135 if (opcode
== SpvOpImageQuerySizeLod
)
2136 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
2138 /* Now we need to handle some number of optional arguments */
2139 const struct vtn_ssa_value
*gather_offsets
= NULL
;
2141 uint32_t operands
= w
[idx
++];
2143 if (operands
& SpvImageOperandsBiasMask
) {
2144 vtn_assert(texop
== nir_texop_tex
);
2145 texop
= nir_texop_txb
;
2146 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_bias
);
2149 if (operands
& SpvImageOperandsLodMask
) {
2150 vtn_assert(texop
== nir_texop_txl
|| texop
== nir_texop_txf
||
2151 texop
== nir_texop_txs
);
2152 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
2155 if (operands
& SpvImageOperandsGradMask
) {
2156 vtn_assert(texop
== nir_texop_txl
);
2157 texop
= nir_texop_txd
;
2158 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddx
);
2159 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddy
);
2162 if (operands
& SpvImageOperandsOffsetMask
||
2163 operands
& SpvImageOperandsConstOffsetMask
)
2164 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_offset
);
2166 if (operands
& SpvImageOperandsConstOffsetsMask
) {
2167 nir_tex_src none
= {0};
2168 gather_offsets
= vtn_ssa_value(b
, w
[idx
++]);
2172 if (operands
& SpvImageOperandsSampleMask
) {
2173 vtn_assert(texop
== nir_texop_txf_ms
);
2174 texop
= nir_texop_txf_ms
;
2175 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ms_index
);
2178 /* We should have now consumed exactly all of the arguments */
2179 vtn_assert(idx
== count
);
2181 nir_tex_instr
*instr
= nir_tex_instr_create(b
->shader
, p
- srcs
);
2184 memcpy(instr
->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
2186 instr
->coord_components
= coord_components
;
2187 instr
->sampler_dim
= sampler_dim
;
2188 instr
->is_array
= is_array
;
2189 instr
->is_shadow
= is_shadow
;
2190 instr
->is_new_style_shadow
=
2191 is_shadow
&& glsl_get_components(ret_type
->type
) == 1;
2192 instr
->component
= gather_component
;
2194 switch (glsl_get_sampler_result_type(image_type
)) {
2195 case GLSL_TYPE_FLOAT
: instr
->dest_type
= nir_type_float
; break;
2196 case GLSL_TYPE_INT
: instr
->dest_type
= nir_type_int
; break;
2197 case GLSL_TYPE_UINT
: instr
->dest_type
= nir_type_uint
; break;
2198 case GLSL_TYPE_BOOL
: instr
->dest_type
= nir_type_bool
; break;
2200 vtn_fail("Invalid base type for sampler result");
2203 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
2204 nir_tex_instr_dest_size(instr
), 32, NULL
);
2206 vtn_assert(glsl_get_vector_elements(ret_type
->type
) ==
2207 nir_tex_instr_dest_size(instr
));
2210 nir_instr
*instruction
;
2211 if (gather_offsets
) {
2212 vtn_assert(glsl_get_base_type(gather_offsets
->type
) == GLSL_TYPE_ARRAY
);
2213 vtn_assert(glsl_get_length(gather_offsets
->type
) == 4);
2214 nir_tex_instr
*instrs
[4] = {instr
, NULL
, NULL
, NULL
};
2216 /* Copy the current instruction 4x */
2217 for (uint32_t i
= 1; i
< 4; i
++) {
2218 instrs
[i
] = nir_tex_instr_create(b
->shader
, instr
->num_srcs
);
2219 instrs
[i
]->op
= instr
->op
;
2220 instrs
[i
]->coord_components
= instr
->coord_components
;
2221 instrs
[i
]->sampler_dim
= instr
->sampler_dim
;
2222 instrs
[i
]->is_array
= instr
->is_array
;
2223 instrs
[i
]->is_shadow
= instr
->is_shadow
;
2224 instrs
[i
]->is_new_style_shadow
= instr
->is_new_style_shadow
;
2225 instrs
[i
]->component
= instr
->component
;
2226 instrs
[i
]->dest_type
= instr
->dest_type
;
2228 memcpy(instrs
[i
]->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
2230 nir_ssa_dest_init(&instrs
[i
]->instr
, &instrs
[i
]->dest
,
2231 nir_tex_instr_dest_size(instr
), 32, NULL
);
2234 /* Fill in the last argument with the offset from the passed in offsets
2235 * and insert the instruction into the stream.
2237 for (uint32_t i
= 0; i
< 4; i
++) {
2239 src
.src
= nir_src_for_ssa(gather_offsets
->elems
[i
]->def
);
2240 src
.src_type
= nir_tex_src_offset
;
2241 instrs
[i
]->src
[instrs
[i
]->num_srcs
- 1] = src
;
2242 nir_builder_instr_insert(&b
->nb
, &instrs
[i
]->instr
);
2245 /* Combine the results of the 4 instructions by taking their .w
2248 nir_alu_instr
*vec4
= nir_alu_instr_create(b
->shader
, nir_op_vec4
);
2249 nir_ssa_dest_init(&vec4
->instr
, &vec4
->dest
.dest
, 4, 32, NULL
);
2250 vec4
->dest
.write_mask
= 0xf;
2251 for (uint32_t i
= 0; i
< 4; i
++) {
2252 vec4
->src
[i
].src
= nir_src_for_ssa(&instrs
[i
]->dest
.ssa
);
2253 vec4
->src
[i
].swizzle
[0] = 3;
2255 def
= &vec4
->dest
.dest
.ssa
;
2256 instruction
= &vec4
->instr
;
2258 def
= &instr
->dest
.ssa
;
2259 instruction
= &instr
->instr
;
2262 val
->ssa
= vtn_create_ssa_value(b
, ret_type
->type
);
2263 val
->ssa
->def
= def
;
2265 nir_builder_instr_insert(&b
->nb
, instruction
);
2269 fill_common_atomic_sources(struct vtn_builder
*b
, SpvOp opcode
,
2270 const uint32_t *w
, nir_src
*src
)
2273 case SpvOpAtomicIIncrement
:
2274 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
2277 case SpvOpAtomicIDecrement
:
2278 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
2281 case SpvOpAtomicISub
:
2283 nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
2286 case SpvOpAtomicCompareExchange
:
2287 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[8])->def
);
2288 src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
2291 case SpvOpAtomicExchange
:
2292 case SpvOpAtomicIAdd
:
2293 case SpvOpAtomicSMin
:
2294 case SpvOpAtomicUMin
:
2295 case SpvOpAtomicSMax
:
2296 case SpvOpAtomicUMax
:
2297 case SpvOpAtomicAnd
:
2299 case SpvOpAtomicXor
:
2300 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
2304 vtn_fail("Invalid SPIR-V atomic");
2308 static nir_ssa_def
*
2309 get_image_coord(struct vtn_builder
*b
, uint32_t value
)
2311 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, value
);
2313 /* The image_load_store intrinsics assume a 4-dim coordinate */
2314 unsigned dim
= glsl_get_vector_elements(coord
->type
);
2315 unsigned swizzle
[4];
2316 for (unsigned i
= 0; i
< 4; i
++)
2317 swizzle
[i
] = MIN2(i
, dim
- 1);
2319 return nir_swizzle(&b
->nb
, coord
->def
, swizzle
, 4, false);
2322 static nir_ssa_def
*
2323 expand_to_vec4(nir_builder
*b
, nir_ssa_def
*value
)
2325 if (value
->num_components
== 4)
2329 for (unsigned i
= 0; i
< 4; i
++)
2330 swiz
[i
] = i
< value
->num_components
? i
: 0;
2331 return nir_swizzle(b
, value
, swiz
, 4, false);
2335 vtn_handle_image(struct vtn_builder
*b
, SpvOp opcode
,
2336 const uint32_t *w
, unsigned count
)
2338 /* Just get this one out of the way */
2339 if (opcode
== SpvOpImageTexelPointer
) {
2340 struct vtn_value
*val
=
2341 vtn_push_value(b
, w
[2], vtn_value_type_image_pointer
);
2342 val
->image
= ralloc(b
, struct vtn_image_pointer
);
2344 val
->image
->image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2345 val
->image
->coord
= get_image_coord(b
, w
[4]);
2346 val
->image
->sample
= vtn_ssa_value(b
, w
[5])->def
;
2350 struct vtn_image_pointer image
;
2353 case SpvOpAtomicExchange
:
2354 case SpvOpAtomicCompareExchange
:
2355 case SpvOpAtomicCompareExchangeWeak
:
2356 case SpvOpAtomicIIncrement
:
2357 case SpvOpAtomicIDecrement
:
2358 case SpvOpAtomicIAdd
:
2359 case SpvOpAtomicISub
:
2360 case SpvOpAtomicLoad
:
2361 case SpvOpAtomicSMin
:
2362 case SpvOpAtomicUMin
:
2363 case SpvOpAtomicSMax
:
2364 case SpvOpAtomicUMax
:
2365 case SpvOpAtomicAnd
:
2367 case SpvOpAtomicXor
:
2368 image
= *vtn_value(b
, w
[3], vtn_value_type_image_pointer
)->image
;
2371 case SpvOpAtomicStore
:
2372 image
= *vtn_value(b
, w
[1], vtn_value_type_image_pointer
)->image
;
2375 case SpvOpImageQuerySize
:
2376 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2378 image
.sample
= NULL
;
2381 case SpvOpImageRead
:
2382 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2383 image
.coord
= get_image_coord(b
, w
[4]);
2385 if (count
> 5 && (w
[5] & SpvImageOperandsSampleMask
)) {
2386 vtn_assert(w
[5] == SpvImageOperandsSampleMask
);
2387 image
.sample
= vtn_ssa_value(b
, w
[6])->def
;
2389 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2393 case SpvOpImageWrite
:
2394 image
.image
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
2395 image
.coord
= get_image_coord(b
, w
[2]);
2399 if (count
> 4 && (w
[4] & SpvImageOperandsSampleMask
)) {
2400 vtn_assert(w
[4] == SpvImageOperandsSampleMask
);
2401 image
.sample
= vtn_ssa_value(b
, w
[5])->def
;
2403 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2408 vtn_fail("Invalid image opcode");
2411 nir_intrinsic_op op
;
2413 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_deref_##N; break;
2414 OP(ImageQuerySize
, size
)
2416 OP(ImageWrite
, store
)
2417 OP(AtomicLoad
, load
)
2418 OP(AtomicStore
, store
)
2419 OP(AtomicExchange
, atomic_exchange
)
2420 OP(AtomicCompareExchange
, atomic_comp_swap
)
2421 OP(AtomicIIncrement
, atomic_add
)
2422 OP(AtomicIDecrement
, atomic_add
)
2423 OP(AtomicIAdd
, atomic_add
)
2424 OP(AtomicISub
, atomic_add
)
2425 OP(AtomicSMin
, atomic_min
)
2426 OP(AtomicUMin
, atomic_min
)
2427 OP(AtomicSMax
, atomic_max
)
2428 OP(AtomicUMax
, atomic_max
)
2429 OP(AtomicAnd
, atomic_and
)
2430 OP(AtomicOr
, atomic_or
)
2431 OP(AtomicXor
, atomic_xor
)
2434 vtn_fail("Invalid image opcode");
2437 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
2439 nir_deref_instr
*image_deref
= vtn_pointer_to_deref(b
, image
.image
);
2440 intrin
->src
[0] = nir_src_for_ssa(&image_deref
->dest
.ssa
);
2442 /* ImageQuerySize doesn't take any extra parameters */
2443 if (opcode
!= SpvOpImageQuerySize
) {
2444 /* The image coordinate is always 4 components but we may not have that
2445 * many. Swizzle to compensate.
2447 intrin
->src
[1] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, image
.coord
));
2448 intrin
->src
[2] = nir_src_for_ssa(image
.sample
);
2452 case SpvOpAtomicLoad
:
2453 case SpvOpImageQuerySize
:
2454 case SpvOpImageRead
:
2456 case SpvOpAtomicStore
:
2457 case SpvOpImageWrite
: {
2458 const uint32_t value_id
= opcode
== SpvOpAtomicStore
? w
[4] : w
[3];
2459 nir_ssa_def
*value
= vtn_ssa_value(b
, value_id
)->def
;
2460 /* nir_intrinsic_image_deref_store always takes a vec4 value */
2461 assert(op
== nir_intrinsic_image_deref_store
);
2462 intrin
->num_components
= 4;
2463 intrin
->src
[3] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, value
));
2467 case SpvOpAtomicCompareExchange
:
2468 case SpvOpAtomicIIncrement
:
2469 case SpvOpAtomicIDecrement
:
2470 case SpvOpAtomicExchange
:
2471 case SpvOpAtomicIAdd
:
2472 case SpvOpAtomicISub
:
2473 case SpvOpAtomicSMin
:
2474 case SpvOpAtomicUMin
:
2475 case SpvOpAtomicSMax
:
2476 case SpvOpAtomicUMax
:
2477 case SpvOpAtomicAnd
:
2479 case SpvOpAtomicXor
:
2480 fill_common_atomic_sources(b
, opcode
, w
, &intrin
->src
[3]);
2484 vtn_fail("Invalid image opcode");
2487 if (opcode
!= SpvOpImageWrite
&& opcode
!= SpvOpAtomicStore
) {
2488 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2489 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2491 unsigned dest_components
= glsl_get_vector_elements(type
->type
);
2492 intrin
->num_components
= nir_intrinsic_infos
[op
].dest_components
;
2493 if (intrin
->num_components
== 0)
2494 intrin
->num_components
= dest_components
;
2496 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
,
2497 intrin
->num_components
, 32, NULL
);
2499 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2501 nir_ssa_def
*result
= &intrin
->dest
.ssa
;
2502 if (intrin
->num_components
!= dest_components
)
2503 result
= nir_channels(&b
->nb
, result
, (1 << dest_components
) - 1);
2505 val
->ssa
= vtn_create_ssa_value(b
, type
->type
);
2506 val
->ssa
->def
= result
;
2508 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2512 static nir_intrinsic_op
2513 get_ssbo_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2516 case SpvOpAtomicLoad
: return nir_intrinsic_load_ssbo
;
2517 case SpvOpAtomicStore
: return nir_intrinsic_store_ssbo
;
2518 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
2519 OP(AtomicExchange
, atomic_exchange
)
2520 OP(AtomicCompareExchange
, atomic_comp_swap
)
2521 OP(AtomicIIncrement
, atomic_add
)
2522 OP(AtomicIDecrement
, atomic_add
)
2523 OP(AtomicIAdd
, atomic_add
)
2524 OP(AtomicISub
, atomic_add
)
2525 OP(AtomicSMin
, atomic_imin
)
2526 OP(AtomicUMin
, atomic_umin
)
2527 OP(AtomicSMax
, atomic_imax
)
2528 OP(AtomicUMax
, atomic_umax
)
2529 OP(AtomicAnd
, atomic_and
)
2530 OP(AtomicOr
, atomic_or
)
2531 OP(AtomicXor
, atomic_xor
)
2534 vtn_fail("Invalid SSBO atomic");
2538 static nir_intrinsic_op
2539 get_uniform_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2542 #define OP(S, N) case SpvOp##S: return nir_intrinsic_atomic_counter_ ##N;
2543 OP(AtomicLoad
, read_deref
)
2544 OP(AtomicExchange
, exchange
)
2545 OP(AtomicCompareExchange
, comp_swap
)
2546 OP(AtomicIIncrement
, inc_deref
)
2547 OP(AtomicIDecrement
, post_dec_deref
)
2548 OP(AtomicIAdd
, add_deref
)
2549 OP(AtomicISub
, add_deref
)
2550 OP(AtomicUMin
, min_deref
)
2551 OP(AtomicUMax
, max_deref
)
2552 OP(AtomicAnd
, and_deref
)
2553 OP(AtomicOr
, or_deref
)
2554 OP(AtomicXor
, xor_deref
)
2557 /* We left the following out: AtomicStore, AtomicSMin and
2558 * AtomicSmax. Right now there are not nir intrinsics for them. At this
2559 * moment Atomic Counter support is needed for ARB_spirv support, so is
2560 * only need to support GLSL Atomic Counters that are uints and don't
2561 * allow direct storage.
2563 unreachable("Invalid uniform atomic");
2567 static nir_intrinsic_op
2568 get_shared_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2571 case SpvOpAtomicLoad
: return nir_intrinsic_load_shared
;
2572 case SpvOpAtomicStore
: return nir_intrinsic_store_shared
;
2573 #define OP(S, N) case SpvOp##S: return nir_intrinsic_shared_##N;
2574 OP(AtomicExchange
, atomic_exchange
)
2575 OP(AtomicCompareExchange
, atomic_comp_swap
)
2576 OP(AtomicIIncrement
, atomic_add
)
2577 OP(AtomicIDecrement
, atomic_add
)
2578 OP(AtomicIAdd
, atomic_add
)
2579 OP(AtomicISub
, atomic_add
)
2580 OP(AtomicSMin
, atomic_imin
)
2581 OP(AtomicUMin
, atomic_umin
)
2582 OP(AtomicSMax
, atomic_imax
)
2583 OP(AtomicUMax
, atomic_umax
)
2584 OP(AtomicAnd
, atomic_and
)
2585 OP(AtomicOr
, atomic_or
)
2586 OP(AtomicXor
, atomic_xor
)
2589 vtn_fail("Invalid shared atomic");
2593 static nir_intrinsic_op
2594 get_deref_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2597 case SpvOpAtomicLoad
: return nir_intrinsic_load_deref
;
2598 case SpvOpAtomicStore
: return nir_intrinsic_store_deref
;
2599 #define OP(S, N) case SpvOp##S: return nir_intrinsic_deref_##N;
2600 OP(AtomicExchange
, atomic_exchange
)
2601 OP(AtomicCompareExchange
, atomic_comp_swap
)
2602 OP(AtomicIIncrement
, atomic_add
)
2603 OP(AtomicIDecrement
, atomic_add
)
2604 OP(AtomicIAdd
, atomic_add
)
2605 OP(AtomicISub
, atomic_add
)
2606 OP(AtomicSMin
, atomic_imin
)
2607 OP(AtomicUMin
, atomic_umin
)
2608 OP(AtomicSMax
, atomic_imax
)
2609 OP(AtomicUMax
, atomic_umax
)
2610 OP(AtomicAnd
, atomic_and
)
2611 OP(AtomicOr
, atomic_or
)
2612 OP(AtomicXor
, atomic_xor
)
2615 vtn_fail("Invalid shared atomic");
2620 * Handles shared atomics, ssbo atomics and atomic counters.
2623 vtn_handle_atomics(struct vtn_builder
*b
, SpvOp opcode
,
2624 const uint32_t *w
, unsigned count
)
2626 struct vtn_pointer
*ptr
;
2627 nir_intrinsic_instr
*atomic
;
2630 case SpvOpAtomicLoad
:
2631 case SpvOpAtomicExchange
:
2632 case SpvOpAtomicCompareExchange
:
2633 case SpvOpAtomicCompareExchangeWeak
:
2634 case SpvOpAtomicIIncrement
:
2635 case SpvOpAtomicIDecrement
:
2636 case SpvOpAtomicIAdd
:
2637 case SpvOpAtomicISub
:
2638 case SpvOpAtomicSMin
:
2639 case SpvOpAtomicUMin
:
2640 case SpvOpAtomicSMax
:
2641 case SpvOpAtomicUMax
:
2642 case SpvOpAtomicAnd
:
2644 case SpvOpAtomicXor
:
2645 ptr
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2648 case SpvOpAtomicStore
:
2649 ptr
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
2653 vtn_fail("Invalid SPIR-V atomic");
2657 SpvScope scope = w[4];
2658 SpvMemorySemanticsMask semantics = w[5];
2661 /* uniform as "atomic counter uniform" */
2662 if (ptr
->mode
== vtn_variable_mode_uniform
) {
2663 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
2664 const struct glsl_type
*deref_type
= deref
->type
;
2665 nir_intrinsic_op op
= get_uniform_nir_atomic_op(b
, opcode
);
2666 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2667 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
2669 /* SSBO needs to initialize index/offset. In this case we don't need to,
2670 * as that info is already stored on the ptr->var->var nir_variable (see
2671 * vtn_create_variable)
2675 case SpvOpAtomicLoad
:
2676 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2679 case SpvOpAtomicStore
:
2680 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2681 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2684 case SpvOpAtomicExchange
:
2685 case SpvOpAtomicCompareExchange
:
2686 case SpvOpAtomicCompareExchangeWeak
:
2687 case SpvOpAtomicIIncrement
:
2688 case SpvOpAtomicIDecrement
:
2689 case SpvOpAtomicIAdd
:
2690 case SpvOpAtomicISub
:
2691 case SpvOpAtomicSMin
:
2692 case SpvOpAtomicUMin
:
2693 case SpvOpAtomicSMax
:
2694 case SpvOpAtomicUMax
:
2695 case SpvOpAtomicAnd
:
2697 case SpvOpAtomicXor
:
2698 /* Nothing: we don't need to call fill_common_atomic_sources here, as
2699 * atomic counter uniforms doesn't have sources
2704 unreachable("Invalid SPIR-V atomic");
2707 } else if (ptr
->mode
== vtn_variable_mode_workgroup
&&
2708 !b
->options
->lower_workgroup_access_to_offsets
) {
2709 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
2710 const struct glsl_type
*deref_type
= deref
->type
;
2711 nir_intrinsic_op op
= get_deref_nir_atomic_op(b
, opcode
);
2712 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2713 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
2716 case SpvOpAtomicLoad
:
2717 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2720 case SpvOpAtomicStore
:
2721 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2722 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2723 atomic
->src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2726 case SpvOpAtomicExchange
:
2727 case SpvOpAtomicCompareExchange
:
2728 case SpvOpAtomicCompareExchangeWeak
:
2729 case SpvOpAtomicIIncrement
:
2730 case SpvOpAtomicIDecrement
:
2731 case SpvOpAtomicIAdd
:
2732 case SpvOpAtomicISub
:
2733 case SpvOpAtomicSMin
:
2734 case SpvOpAtomicUMin
:
2735 case SpvOpAtomicSMax
:
2736 case SpvOpAtomicUMax
:
2737 case SpvOpAtomicAnd
:
2739 case SpvOpAtomicXor
:
2740 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[1]);
2744 vtn_fail("Invalid SPIR-V atomic");
2748 nir_ssa_def
*offset
, *index
;
2749 offset
= vtn_pointer_to_offset(b
, ptr
, &index
);
2751 nir_intrinsic_op op
;
2752 if (ptr
->mode
== vtn_variable_mode_ssbo
) {
2753 op
= get_ssbo_nir_atomic_op(b
, opcode
);
2755 vtn_assert(ptr
->mode
== vtn_variable_mode_workgroup
&&
2756 b
->options
->lower_workgroup_access_to_offsets
);
2757 op
= get_shared_nir_atomic_op(b
, opcode
);
2760 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2764 case SpvOpAtomicLoad
:
2765 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
2766 nir_intrinsic_set_align(atomic
, 4, 0);
2767 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2768 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2769 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2772 case SpvOpAtomicStore
:
2773 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
2774 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2775 nir_intrinsic_set_align(atomic
, 4, 0);
2776 atomic
->src
[src
++] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2777 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2778 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2779 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2782 case SpvOpAtomicExchange
:
2783 case SpvOpAtomicCompareExchange
:
2784 case SpvOpAtomicCompareExchangeWeak
:
2785 case SpvOpAtomicIIncrement
:
2786 case SpvOpAtomicIDecrement
:
2787 case SpvOpAtomicIAdd
:
2788 case SpvOpAtomicISub
:
2789 case SpvOpAtomicSMin
:
2790 case SpvOpAtomicUMin
:
2791 case SpvOpAtomicSMax
:
2792 case SpvOpAtomicUMax
:
2793 case SpvOpAtomicAnd
:
2795 case SpvOpAtomicXor
:
2796 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2797 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2798 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2799 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[src
]);
2803 vtn_fail("Invalid SPIR-V atomic");
2807 if (opcode
!= SpvOpAtomicStore
) {
2808 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2810 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
,
2811 glsl_get_vector_elements(type
->type
),
2812 glsl_get_bit_size(type
->type
), NULL
);
2814 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2815 val
->ssa
= rzalloc(b
, struct vtn_ssa_value
);
2816 val
->ssa
->def
= &atomic
->dest
.ssa
;
2817 val
->ssa
->type
= type
->type
;
2820 nir_builder_instr_insert(&b
->nb
, &atomic
->instr
);
2823 static nir_alu_instr
*
2824 create_vec(struct vtn_builder
*b
, unsigned num_components
, unsigned bit_size
)
2827 switch (num_components
) {
2828 case 1: op
= nir_op_imov
; break;
2829 case 2: op
= nir_op_vec2
; break;
2830 case 3: op
= nir_op_vec3
; break;
2831 case 4: op
= nir_op_vec4
; break;
2832 default: vtn_fail("bad vector size");
2835 nir_alu_instr
*vec
= nir_alu_instr_create(b
->shader
, op
);
2836 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, num_components
,
2838 vec
->dest
.write_mask
= (1 << num_components
) - 1;
2843 struct vtn_ssa_value
*
2844 vtn_ssa_transpose(struct vtn_builder
*b
, struct vtn_ssa_value
*src
)
2846 if (src
->transposed
)
2847 return src
->transposed
;
2849 struct vtn_ssa_value
*dest
=
2850 vtn_create_ssa_value(b
, glsl_transposed_type(src
->type
));
2852 for (unsigned i
= 0; i
< glsl_get_matrix_columns(dest
->type
); i
++) {
2853 nir_alu_instr
*vec
= create_vec(b
, glsl_get_matrix_columns(src
->type
),
2854 glsl_get_bit_size(src
->type
));
2855 if (glsl_type_is_vector_or_scalar(src
->type
)) {
2856 vec
->src
[0].src
= nir_src_for_ssa(src
->def
);
2857 vec
->src
[0].swizzle
[0] = i
;
2859 for (unsigned j
= 0; j
< glsl_get_matrix_columns(src
->type
); j
++) {
2860 vec
->src
[j
].src
= nir_src_for_ssa(src
->elems
[j
]->def
);
2861 vec
->src
[j
].swizzle
[0] = i
;
2864 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2865 dest
->elems
[i
]->def
= &vec
->dest
.dest
.ssa
;
2868 dest
->transposed
= src
;
2874 vtn_vector_extract(struct vtn_builder
*b
, nir_ssa_def
*src
, unsigned index
)
2876 return nir_channel(&b
->nb
, src
, index
);
2880 vtn_vector_insert(struct vtn_builder
*b
, nir_ssa_def
*src
, nir_ssa_def
*insert
,
2883 nir_alu_instr
*vec
= create_vec(b
, src
->num_components
,
2886 for (unsigned i
= 0; i
< src
->num_components
; i
++) {
2888 vec
->src
[i
].src
= nir_src_for_ssa(insert
);
2890 vec
->src
[i
].src
= nir_src_for_ssa(src
);
2891 vec
->src
[i
].swizzle
[0] = i
;
2895 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2897 return &vec
->dest
.dest
.ssa
;
2901 vtn_vector_extract_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
2904 nir_ssa_def
*dest
= vtn_vector_extract(b
, src
, 0);
2905 for (unsigned i
= 1; i
< src
->num_components
; i
++)
2906 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
2907 vtn_vector_extract(b
, src
, i
), dest
);
2913 vtn_vector_insert_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
2914 nir_ssa_def
*insert
, nir_ssa_def
*index
)
2916 nir_ssa_def
*dest
= vtn_vector_insert(b
, src
, insert
, 0);
2917 for (unsigned i
= 1; i
< src
->num_components
; i
++)
2918 dest
= nir_bcsel(&b
->nb
, nir_ieq(&b
->nb
, index
, nir_imm_int(&b
->nb
, i
)),
2919 vtn_vector_insert(b
, src
, insert
, i
), dest
);
2924 static nir_ssa_def
*
2925 vtn_vector_shuffle(struct vtn_builder
*b
, unsigned num_components
,
2926 nir_ssa_def
*src0
, nir_ssa_def
*src1
,
2927 const uint32_t *indices
)
2929 nir_alu_instr
*vec
= create_vec(b
, num_components
, src0
->bit_size
);
2931 for (unsigned i
= 0; i
< num_components
; i
++) {
2932 uint32_t index
= indices
[i
];
2933 if (index
== 0xffffffff) {
2935 nir_src_for_ssa(nir_ssa_undef(&b
->nb
, 1, src0
->bit_size
));
2936 } else if (index
< src0
->num_components
) {
2937 vec
->src
[i
].src
= nir_src_for_ssa(src0
);
2938 vec
->src
[i
].swizzle
[0] = index
;
2940 vec
->src
[i
].src
= nir_src_for_ssa(src1
);
2941 vec
->src
[i
].swizzle
[0] = index
- src0
->num_components
;
2945 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2947 return &vec
->dest
.dest
.ssa
;
2951 * Concatentates a number of vectors/scalars together to produce a vector
2953 static nir_ssa_def
*
2954 vtn_vector_construct(struct vtn_builder
*b
, unsigned num_components
,
2955 unsigned num_srcs
, nir_ssa_def
**srcs
)
2957 nir_alu_instr
*vec
= create_vec(b
, num_components
, srcs
[0]->bit_size
);
2959 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2961 * "When constructing a vector, there must be at least two Constituent
2964 vtn_assert(num_srcs
>= 2);
2966 unsigned dest_idx
= 0;
2967 for (unsigned i
= 0; i
< num_srcs
; i
++) {
2968 nir_ssa_def
*src
= srcs
[i
];
2969 vtn_assert(dest_idx
+ src
->num_components
<= num_components
);
2970 for (unsigned j
= 0; j
< src
->num_components
; j
++) {
2971 vec
->src
[dest_idx
].src
= nir_src_for_ssa(src
);
2972 vec
->src
[dest_idx
].swizzle
[0] = j
;
2977 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2979 * "When constructing a vector, the total number of components in all
2980 * the operands must equal the number of components in Result Type."
2982 vtn_assert(dest_idx
== num_components
);
2984 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2986 return &vec
->dest
.dest
.ssa
;
2989 static struct vtn_ssa_value
*
2990 vtn_composite_copy(void *mem_ctx
, struct vtn_ssa_value
*src
)
2992 struct vtn_ssa_value
*dest
= rzalloc(mem_ctx
, struct vtn_ssa_value
);
2993 dest
->type
= src
->type
;
2995 if (glsl_type_is_vector_or_scalar(src
->type
)) {
2996 dest
->def
= src
->def
;
2998 unsigned elems
= glsl_get_length(src
->type
);
3000 dest
->elems
= ralloc_array(mem_ctx
, struct vtn_ssa_value
*, elems
);
3001 for (unsigned i
= 0; i
< elems
; i
++)
3002 dest
->elems
[i
] = vtn_composite_copy(mem_ctx
, src
->elems
[i
]);
3008 static struct vtn_ssa_value
*
3009 vtn_composite_insert(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
3010 struct vtn_ssa_value
*insert
, const uint32_t *indices
,
3011 unsigned num_indices
)
3013 struct vtn_ssa_value
*dest
= vtn_composite_copy(b
, src
);
3015 struct vtn_ssa_value
*cur
= dest
;
3017 for (i
= 0; i
< num_indices
- 1; i
++) {
3018 cur
= cur
->elems
[indices
[i
]];
3021 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
3022 /* According to the SPIR-V spec, OpCompositeInsert may work down to
3023 * the component granularity. In that case, the last index will be
3024 * the index to insert the scalar into the vector.
3027 cur
->def
= vtn_vector_insert(b
, cur
->def
, insert
->def
, indices
[i
]);
3029 cur
->elems
[indices
[i
]] = insert
;
3035 static struct vtn_ssa_value
*
3036 vtn_composite_extract(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
3037 const uint32_t *indices
, unsigned num_indices
)
3039 struct vtn_ssa_value
*cur
= src
;
3040 for (unsigned i
= 0; i
< num_indices
; i
++) {
3041 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
3042 vtn_assert(i
== num_indices
- 1);
3043 /* According to the SPIR-V spec, OpCompositeExtract may work down to
3044 * the component granularity. The last index will be the index of the
3045 * vector to extract.
3048 struct vtn_ssa_value
*ret
= rzalloc(b
, struct vtn_ssa_value
);
3049 ret
->type
= glsl_scalar_type(glsl_get_base_type(cur
->type
));
3050 ret
->def
= vtn_vector_extract(b
, cur
->def
, indices
[i
]);
3053 cur
= cur
->elems
[indices
[i
]];
3061 vtn_handle_composite(struct vtn_builder
*b
, SpvOp opcode
,
3062 const uint32_t *w
, unsigned count
)
3064 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
3065 const struct glsl_type
*type
=
3066 vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
3067 val
->ssa
= vtn_create_ssa_value(b
, type
);
3070 case SpvOpVectorExtractDynamic
:
3071 val
->ssa
->def
= vtn_vector_extract_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
3072 vtn_ssa_value(b
, w
[4])->def
);
3075 case SpvOpVectorInsertDynamic
:
3076 val
->ssa
->def
= vtn_vector_insert_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
3077 vtn_ssa_value(b
, w
[4])->def
,
3078 vtn_ssa_value(b
, w
[5])->def
);
3081 case SpvOpVectorShuffle
:
3082 val
->ssa
->def
= vtn_vector_shuffle(b
, glsl_get_vector_elements(type
),
3083 vtn_ssa_value(b
, w
[3])->def
,
3084 vtn_ssa_value(b
, w
[4])->def
,
3088 case SpvOpCompositeConstruct
: {
3089 unsigned elems
= count
- 3;
3091 if (glsl_type_is_vector_or_scalar(type
)) {
3092 nir_ssa_def
*srcs
[NIR_MAX_VEC_COMPONENTS
];
3093 for (unsigned i
= 0; i
< elems
; i
++)
3094 srcs
[i
] = vtn_ssa_value(b
, w
[3 + i
])->def
;
3096 vtn_vector_construct(b
, glsl_get_vector_elements(type
),
3099 val
->ssa
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
3100 for (unsigned i
= 0; i
< elems
; i
++)
3101 val
->ssa
->elems
[i
] = vtn_ssa_value(b
, w
[3 + i
]);
3105 case SpvOpCompositeExtract
:
3106 val
->ssa
= vtn_composite_extract(b
, vtn_ssa_value(b
, w
[3]),
3110 case SpvOpCompositeInsert
:
3111 val
->ssa
= vtn_composite_insert(b
, vtn_ssa_value(b
, w
[4]),
3112 vtn_ssa_value(b
, w
[3]),
3116 case SpvOpCopyObject
:
3117 val
->ssa
= vtn_composite_copy(b
, vtn_ssa_value(b
, w
[3]));
3121 vtn_fail("unknown composite operation");
3126 vtn_emit_barrier(struct vtn_builder
*b
, nir_intrinsic_op op
)
3128 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
3129 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3133 vtn_emit_memory_barrier(struct vtn_builder
*b
, SpvScope scope
,
3134 SpvMemorySemanticsMask semantics
)
3136 static const SpvMemorySemanticsMask all_memory_semantics
=
3137 SpvMemorySemanticsUniformMemoryMask
|
3138 SpvMemorySemanticsWorkgroupMemoryMask
|
3139 SpvMemorySemanticsAtomicCounterMemoryMask
|
3140 SpvMemorySemanticsImageMemoryMask
;
3142 /* If we're not actually doing a memory barrier, bail */
3143 if (!(semantics
& all_memory_semantics
))
3146 /* GL and Vulkan don't have these */
3147 vtn_assert(scope
!= SpvScopeCrossDevice
);
3149 if (scope
== SpvScopeSubgroup
)
3150 return; /* Nothing to do here */
3152 if (scope
== SpvScopeWorkgroup
) {
3153 vtn_emit_barrier(b
, nir_intrinsic_group_memory_barrier
);
3157 /* There's only two scopes thing left */
3158 vtn_assert(scope
== SpvScopeInvocation
|| scope
== SpvScopeDevice
);
3160 if ((semantics
& all_memory_semantics
) == all_memory_semantics
) {
3161 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier
);
3165 /* Issue a bunch of more specific barriers */
3166 uint32_t bits
= semantics
;
3168 SpvMemorySemanticsMask semantic
= 1 << u_bit_scan(&bits
);
3170 case SpvMemorySemanticsUniformMemoryMask
:
3171 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_buffer
);
3173 case SpvMemorySemanticsWorkgroupMemoryMask
:
3174 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_shared
);
3176 case SpvMemorySemanticsAtomicCounterMemoryMask
:
3177 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_atomic_counter
);
3179 case SpvMemorySemanticsImageMemoryMask
:
3180 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_image
);
3189 vtn_handle_barrier(struct vtn_builder
*b
, SpvOp opcode
,
3190 const uint32_t *w
, unsigned count
)
3193 case SpvOpEmitVertex
:
3194 case SpvOpEmitStreamVertex
:
3195 case SpvOpEndPrimitive
:
3196 case SpvOpEndStreamPrimitive
: {
3197 nir_intrinsic_op intrinsic_op
;
3199 case SpvOpEmitVertex
:
3200 case SpvOpEmitStreamVertex
:
3201 intrinsic_op
= nir_intrinsic_emit_vertex
;
3203 case SpvOpEndPrimitive
:
3204 case SpvOpEndStreamPrimitive
:
3205 intrinsic_op
= nir_intrinsic_end_primitive
;
3208 unreachable("Invalid opcode");
3211 nir_intrinsic_instr
*intrin
=
3212 nir_intrinsic_instr_create(b
->shader
, intrinsic_op
);
3215 case SpvOpEmitStreamVertex
:
3216 case SpvOpEndStreamPrimitive
: {
3217 unsigned stream
= vtn_constant_value(b
, w
[1])->values
[0].u32
[0];
3218 nir_intrinsic_set_stream_id(intrin
, stream
);
3226 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3230 case SpvOpMemoryBarrier
: {
3231 SpvScope scope
= vtn_constant_value(b
, w
[1])->values
[0].u32
[0];
3232 SpvMemorySemanticsMask semantics
=
3233 vtn_constant_value(b
, w
[2])->values
[0].u32
[0];
3234 vtn_emit_memory_barrier(b
, scope
, semantics
);
3238 case SpvOpControlBarrier
: {
3239 SpvScope execution_scope
=
3240 vtn_constant_value(b
, w
[1])->values
[0].u32
[0];
3241 if (execution_scope
== SpvScopeWorkgroup
)
3242 vtn_emit_barrier(b
, nir_intrinsic_barrier
);
3244 SpvScope memory_scope
=
3245 vtn_constant_value(b
, w
[2])->values
[0].u32
[0];
3246 SpvMemorySemanticsMask memory_semantics
=
3247 vtn_constant_value(b
, w
[3])->values
[0].u32
[0];
3248 vtn_emit_memory_barrier(b
, memory_scope
, memory_semantics
);
3253 unreachable("unknown barrier instruction");
3258 gl_primitive_from_spv_execution_mode(struct vtn_builder
*b
,
3259 SpvExecutionMode mode
)
3262 case SpvExecutionModeInputPoints
:
3263 case SpvExecutionModeOutputPoints
:
3264 return 0; /* GL_POINTS */
3265 case SpvExecutionModeInputLines
:
3266 return 1; /* GL_LINES */
3267 case SpvExecutionModeInputLinesAdjacency
:
3268 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
3269 case SpvExecutionModeTriangles
:
3270 return 4; /* GL_TRIANGLES */
3271 case SpvExecutionModeInputTrianglesAdjacency
:
3272 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
3273 case SpvExecutionModeQuads
:
3274 return 7; /* GL_QUADS */
3275 case SpvExecutionModeIsolines
:
3276 return 0x8E7A; /* GL_ISOLINES */
3277 case SpvExecutionModeOutputLineStrip
:
3278 return 3; /* GL_LINE_STRIP */
3279 case SpvExecutionModeOutputTriangleStrip
:
3280 return 5; /* GL_TRIANGLE_STRIP */
3282 vtn_fail("Invalid primitive type");
3287 vertices_in_from_spv_execution_mode(struct vtn_builder
*b
,
3288 SpvExecutionMode mode
)
3291 case SpvExecutionModeInputPoints
:
3293 case SpvExecutionModeInputLines
:
3295 case SpvExecutionModeInputLinesAdjacency
:
3297 case SpvExecutionModeTriangles
:
3299 case SpvExecutionModeInputTrianglesAdjacency
:
3302 vtn_fail("Invalid GS input mode");
3306 static gl_shader_stage
3307 stage_for_execution_model(struct vtn_builder
*b
, SpvExecutionModel model
)
3310 case SpvExecutionModelVertex
:
3311 return MESA_SHADER_VERTEX
;
3312 case SpvExecutionModelTessellationControl
:
3313 return MESA_SHADER_TESS_CTRL
;
3314 case SpvExecutionModelTessellationEvaluation
:
3315 return MESA_SHADER_TESS_EVAL
;
3316 case SpvExecutionModelGeometry
:
3317 return MESA_SHADER_GEOMETRY
;
3318 case SpvExecutionModelFragment
:
3319 return MESA_SHADER_FRAGMENT
;
3320 case SpvExecutionModelGLCompute
:
3321 return MESA_SHADER_COMPUTE
;
3323 vtn_fail("Unsupported execution model");
3327 #define spv_check_supported(name, cap) do { \
3328 if (!(b->options && b->options->caps.name)) \
3329 vtn_warn("Unsupported SPIR-V capability: %s", \
3330 spirv_capability_to_string(cap)); \
3335 vtn_handle_entry_point(struct vtn_builder
*b
, const uint32_t *w
,
3338 struct vtn_value
*entry_point
= &b
->values
[w
[2]];
3339 /* Let this be a name label regardless */
3340 unsigned name_words
;
3341 entry_point
->name
= vtn_string_literal(b
, &w
[3], count
- 3, &name_words
);
3343 if (strcmp(entry_point
->name
, b
->entry_point_name
) != 0 ||
3344 stage_for_execution_model(b
, w
[1]) != b
->entry_point_stage
)
3347 vtn_assert(b
->entry_point
== NULL
);
3348 b
->entry_point
= entry_point
;
3352 vtn_handle_preamble_instruction(struct vtn_builder
*b
, SpvOp opcode
,
3353 const uint32_t *w
, unsigned count
)
3360 case SpvSourceLanguageUnknown
: lang
= "unknown"; break;
3361 case SpvSourceLanguageESSL
: lang
= "ESSL"; break;
3362 case SpvSourceLanguageGLSL
: lang
= "GLSL"; break;
3363 case SpvSourceLanguageOpenCL_C
: lang
= "OpenCL C"; break;
3364 case SpvSourceLanguageOpenCL_CPP
: lang
= "OpenCL C++"; break;
3365 case SpvSourceLanguageHLSL
: lang
= "HLSL"; break;
3368 uint32_t version
= w
[2];
3371 (count
> 3) ? vtn_value(b
, w
[3], vtn_value_type_string
)->str
: "";
3373 vtn_info("Parsing SPIR-V from %s %u source file %s", lang
, version
, file
);
3377 case SpvOpSourceExtension
:
3378 case SpvOpSourceContinued
:
3379 case SpvOpExtension
:
3380 case SpvOpModuleProcessed
:
3381 /* Unhandled, but these are for debug so that's ok. */
3384 case SpvOpCapability
: {
3385 SpvCapability cap
= w
[1];
3387 case SpvCapabilityMatrix
:
3388 case SpvCapabilityShader
:
3389 case SpvCapabilityGeometry
:
3390 case SpvCapabilityGeometryPointSize
:
3391 case SpvCapabilityUniformBufferArrayDynamicIndexing
:
3392 case SpvCapabilitySampledImageArrayDynamicIndexing
:
3393 case SpvCapabilityStorageBufferArrayDynamicIndexing
:
3394 case SpvCapabilityStorageImageArrayDynamicIndexing
:
3395 case SpvCapabilityImageRect
:
3396 case SpvCapabilitySampledRect
:
3397 case SpvCapabilitySampled1D
:
3398 case SpvCapabilityImage1D
:
3399 case SpvCapabilitySampledCubeArray
:
3400 case SpvCapabilityImageCubeArray
:
3401 case SpvCapabilitySampledBuffer
:
3402 case SpvCapabilityImageBuffer
:
3403 case SpvCapabilityImageQuery
:
3404 case SpvCapabilityDerivativeControl
:
3405 case SpvCapabilityInterpolationFunction
:
3406 case SpvCapabilityMultiViewport
:
3407 case SpvCapabilitySampleRateShading
:
3408 case SpvCapabilityClipDistance
:
3409 case SpvCapabilityCullDistance
:
3410 case SpvCapabilityInputAttachment
:
3411 case SpvCapabilityImageGatherExtended
:
3412 case SpvCapabilityStorageImageExtendedFormats
:
3415 case SpvCapabilityLinkage
:
3416 case SpvCapabilityVector16
:
3417 case SpvCapabilityFloat16Buffer
:
3418 case SpvCapabilityFloat16
:
3419 case SpvCapabilityInt64Atomics
:
3420 case SpvCapabilityStorageImageMultisample
:
3421 case SpvCapabilityInt8
:
3422 case SpvCapabilitySparseResidency
:
3423 case SpvCapabilityMinLod
:
3424 vtn_warn("Unsupported SPIR-V capability: %s",
3425 spirv_capability_to_string(cap
));
3428 case SpvCapabilityAtomicStorage
:
3429 spv_check_supported(atomic_storage
, cap
);
3432 case SpvCapabilityFloat64
:
3433 spv_check_supported(float64
, cap
);
3435 case SpvCapabilityInt64
:
3436 spv_check_supported(int64
, cap
);
3438 case SpvCapabilityInt16
:
3439 spv_check_supported(int16
, cap
);
3442 case SpvCapabilityTransformFeedback
:
3443 spv_check_supported(transform_feedback
, cap
);
3446 case SpvCapabilityGeometryStreams
:
3447 spv_check_supported(geometry_streams
, cap
);
3450 case SpvCapabilityAddresses
:
3451 case SpvCapabilityKernel
:
3452 case SpvCapabilityImageBasic
:
3453 case SpvCapabilityImageReadWrite
:
3454 case SpvCapabilityImageMipmap
:
3455 case SpvCapabilityPipes
:
3456 case SpvCapabilityGroups
:
3457 case SpvCapabilityDeviceEnqueue
:
3458 case SpvCapabilityLiteralSampler
:
3459 case SpvCapabilityGenericPointer
:
3460 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
3461 spirv_capability_to_string(cap
));
3464 case SpvCapabilityImageMSArray
:
3465 spv_check_supported(image_ms_array
, cap
);
3468 case SpvCapabilityTessellation
:
3469 case SpvCapabilityTessellationPointSize
:
3470 spv_check_supported(tessellation
, cap
);
3473 case SpvCapabilityDrawParameters
:
3474 spv_check_supported(draw_parameters
, cap
);
3477 case SpvCapabilityStorageImageReadWithoutFormat
:
3478 spv_check_supported(image_read_without_format
, cap
);
3481 case SpvCapabilityStorageImageWriteWithoutFormat
:
3482 spv_check_supported(image_write_without_format
, cap
);
3485 case SpvCapabilityDeviceGroup
:
3486 spv_check_supported(device_group
, cap
);
3489 case SpvCapabilityMultiView
:
3490 spv_check_supported(multiview
, cap
);
3493 case SpvCapabilityGroupNonUniform
:
3494 spv_check_supported(subgroup_basic
, cap
);
3497 case SpvCapabilityGroupNonUniformVote
:
3498 spv_check_supported(subgroup_vote
, cap
);
3501 case SpvCapabilitySubgroupBallotKHR
:
3502 case SpvCapabilityGroupNonUniformBallot
:
3503 spv_check_supported(subgroup_ballot
, cap
);
3506 case SpvCapabilityGroupNonUniformShuffle
:
3507 case SpvCapabilityGroupNonUniformShuffleRelative
:
3508 spv_check_supported(subgroup_shuffle
, cap
);
3511 case SpvCapabilityGroupNonUniformQuad
:
3512 spv_check_supported(subgroup_quad
, cap
);
3515 case SpvCapabilityGroupNonUniformArithmetic
:
3516 case SpvCapabilityGroupNonUniformClustered
:
3517 spv_check_supported(subgroup_arithmetic
, cap
);
3520 case SpvCapabilityVariablePointersStorageBuffer
:
3521 case SpvCapabilityVariablePointers
:
3522 spv_check_supported(variable_pointers
, cap
);
3525 case SpvCapabilityStorageUniformBufferBlock16
:
3526 case SpvCapabilityStorageUniform16
:
3527 case SpvCapabilityStoragePushConstant16
:
3528 case SpvCapabilityStorageInputOutput16
:
3529 spv_check_supported(storage_16bit
, cap
);
3532 case SpvCapabilityShaderViewportIndexLayerEXT
:
3533 spv_check_supported(shader_viewport_index_layer
, cap
);
3536 case SpvCapabilityStorageBuffer8BitAccess
:
3537 case SpvCapabilityUniformAndStorageBuffer8BitAccess
:
3538 case SpvCapabilityStoragePushConstant8
:
3539 spv_check_supported(storage_8bit
, cap
);
3542 case SpvCapabilityInputAttachmentArrayDynamicIndexingEXT
:
3543 case SpvCapabilityUniformTexelBufferArrayDynamicIndexingEXT
:
3544 case SpvCapabilityStorageTexelBufferArrayDynamicIndexingEXT
:
3545 spv_check_supported(descriptor_array_dynamic_indexing
, cap
);
3548 case SpvCapabilityRuntimeDescriptorArrayEXT
:
3549 spv_check_supported(runtime_descriptor_array
, cap
);
3552 case SpvCapabilityStencilExportEXT
:
3553 spv_check_supported(stencil_export
, cap
);
3556 case SpvCapabilitySampleMaskPostDepthCoverage
:
3557 spv_check_supported(post_depth_coverage
, cap
);
3561 vtn_fail("Unhandled capability");
3566 case SpvOpExtInstImport
:
3567 vtn_handle_extension(b
, opcode
, w
, count
);
3570 case SpvOpMemoryModel
:
3571 vtn_assert(w
[1] == SpvAddressingModelLogical
);
3572 vtn_assert(w
[2] == SpvMemoryModelSimple
||
3573 w
[2] == SpvMemoryModelGLSL450
);
3576 case SpvOpEntryPoint
:
3577 vtn_handle_entry_point(b
, w
, count
);
3581 vtn_push_value(b
, w
[1], vtn_value_type_string
)->str
=
3582 vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
3586 b
->values
[w
[1]].name
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
3589 case SpvOpMemberName
:
3593 case SpvOpExecutionMode
:
3594 case SpvOpDecorationGroup
:
3596 case SpvOpMemberDecorate
:
3597 case SpvOpGroupDecorate
:
3598 case SpvOpGroupMemberDecorate
:
3599 case SpvOpDecorateStringGOOGLE
:
3600 case SpvOpMemberDecorateStringGOOGLE
:
3601 vtn_handle_decoration(b
, opcode
, w
, count
);
3605 return false; /* End of preamble */
3612 vtn_handle_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
3613 const struct vtn_decoration
*mode
, void *data
)
3615 vtn_assert(b
->entry_point
== entry_point
);
3617 switch(mode
->exec_mode
) {
3618 case SpvExecutionModeOriginUpperLeft
:
3619 case SpvExecutionModeOriginLowerLeft
:
3620 b
->origin_upper_left
=
3621 (mode
->exec_mode
== SpvExecutionModeOriginUpperLeft
);
3624 case SpvExecutionModeEarlyFragmentTests
:
3625 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3626 b
->shader
->info
.fs
.early_fragment_tests
= true;
3629 case SpvExecutionModePostDepthCoverage
:
3630 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3631 b
->shader
->info
.fs
.post_depth_coverage
= true;
3634 case SpvExecutionModeInvocations
:
3635 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3636 b
->shader
->info
.gs
.invocations
= MAX2(1, mode
->literals
[0]);
3639 case SpvExecutionModeDepthReplacing
:
3640 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3641 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_ANY
;
3643 case SpvExecutionModeDepthGreater
:
3644 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3645 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_GREATER
;
3647 case SpvExecutionModeDepthLess
:
3648 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3649 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_LESS
;
3651 case SpvExecutionModeDepthUnchanged
:
3652 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3653 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
3656 case SpvExecutionModeLocalSize
:
3657 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
3658 b
->shader
->info
.cs
.local_size
[0] = mode
->literals
[0];
3659 b
->shader
->info
.cs
.local_size
[1] = mode
->literals
[1];
3660 b
->shader
->info
.cs
.local_size
[2] = mode
->literals
[2];
3662 case SpvExecutionModeLocalSizeHint
:
3663 break; /* Nothing to do with this */
3665 case SpvExecutionModeOutputVertices
:
3666 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3667 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
3668 b
->shader
->info
.tess
.tcs_vertices_out
= mode
->literals
[0];
3670 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3671 b
->shader
->info
.gs
.vertices_out
= mode
->literals
[0];
3675 case SpvExecutionModeInputPoints
:
3676 case SpvExecutionModeInputLines
:
3677 case SpvExecutionModeInputLinesAdjacency
:
3678 case SpvExecutionModeTriangles
:
3679 case SpvExecutionModeInputTrianglesAdjacency
:
3680 case SpvExecutionModeQuads
:
3681 case SpvExecutionModeIsolines
:
3682 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3683 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
3684 b
->shader
->info
.tess
.primitive_mode
=
3685 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
3687 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3688 b
->shader
->info
.gs
.vertices_in
=
3689 vertices_in_from_spv_execution_mode(b
, mode
->exec_mode
);
3690 b
->shader
->info
.gs
.input_primitive
=
3691 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
3695 case SpvExecutionModeOutputPoints
:
3696 case SpvExecutionModeOutputLineStrip
:
3697 case SpvExecutionModeOutputTriangleStrip
:
3698 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3699 b
->shader
->info
.gs
.output_primitive
=
3700 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
3703 case SpvExecutionModeSpacingEqual
:
3704 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3705 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3706 b
->shader
->info
.tess
.spacing
= TESS_SPACING_EQUAL
;
3708 case SpvExecutionModeSpacingFractionalEven
:
3709 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3710 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3711 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_EVEN
;
3713 case SpvExecutionModeSpacingFractionalOdd
:
3714 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3715 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3716 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_ODD
;
3718 case SpvExecutionModeVertexOrderCw
:
3719 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3720 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3721 b
->shader
->info
.tess
.ccw
= false;
3723 case SpvExecutionModeVertexOrderCcw
:
3724 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3725 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3726 b
->shader
->info
.tess
.ccw
= true;
3728 case SpvExecutionModePointMode
:
3729 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3730 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3731 b
->shader
->info
.tess
.point_mode
= true;
3734 case SpvExecutionModePixelCenterInteger
:
3735 b
->pixel_center_integer
= true;
3738 case SpvExecutionModeXfb
:
3739 b
->shader
->info
.has_transform_feedback_varyings
= true;
3742 case SpvExecutionModeVecTypeHint
:
3743 case SpvExecutionModeContractionOff
:
3746 case SpvExecutionModeStencilRefReplacingEXT
:
3747 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3751 vtn_fail("Unhandled execution mode");
3756 vtn_handle_variable_or_type_instruction(struct vtn_builder
*b
, SpvOp opcode
,
3757 const uint32_t *w
, unsigned count
)
3759 vtn_set_instruction_result_type(b
, opcode
, w
, count
);
3763 case SpvOpSourceContinued
:
3764 case SpvOpSourceExtension
:
3765 case SpvOpExtension
:
3766 case SpvOpCapability
:
3767 case SpvOpExtInstImport
:
3768 case SpvOpMemoryModel
:
3769 case SpvOpEntryPoint
:
3770 case SpvOpExecutionMode
:
3773 case SpvOpMemberName
:
3774 case SpvOpDecorationGroup
:
3776 case SpvOpMemberDecorate
:
3777 case SpvOpGroupDecorate
:
3778 case SpvOpGroupMemberDecorate
:
3779 case SpvOpDecorateStringGOOGLE
:
3780 case SpvOpMemberDecorateStringGOOGLE
:
3781 vtn_fail("Invalid opcode types and variables section");
3787 case SpvOpTypeFloat
:
3788 case SpvOpTypeVector
:
3789 case SpvOpTypeMatrix
:
3790 case SpvOpTypeImage
:
3791 case SpvOpTypeSampler
:
3792 case SpvOpTypeSampledImage
:
3793 case SpvOpTypeArray
:
3794 case SpvOpTypeRuntimeArray
:
3795 case SpvOpTypeStruct
:
3796 case SpvOpTypeOpaque
:
3797 case SpvOpTypePointer
:
3798 case SpvOpTypeFunction
:
3799 case SpvOpTypeEvent
:
3800 case SpvOpTypeDeviceEvent
:
3801 case SpvOpTypeReserveId
:
3802 case SpvOpTypeQueue
:
3804 vtn_handle_type(b
, opcode
, w
, count
);
3807 case SpvOpConstantTrue
:
3808 case SpvOpConstantFalse
:
3810 case SpvOpConstantComposite
:
3811 case SpvOpConstantSampler
:
3812 case SpvOpConstantNull
:
3813 case SpvOpSpecConstantTrue
:
3814 case SpvOpSpecConstantFalse
:
3815 case SpvOpSpecConstant
:
3816 case SpvOpSpecConstantComposite
:
3817 case SpvOpSpecConstantOp
:
3818 vtn_handle_constant(b
, opcode
, w
, count
);
3823 vtn_handle_variables(b
, opcode
, w
, count
);
3827 return false; /* End of preamble */
3834 vtn_handle_body_instruction(struct vtn_builder
*b
, SpvOp opcode
,
3835 const uint32_t *w
, unsigned count
)
3841 case SpvOpLoopMerge
:
3842 case SpvOpSelectionMerge
:
3843 /* This is handled by cfg pre-pass and walk_blocks */
3847 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
3848 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
3853 vtn_handle_extension(b
, opcode
, w
, count
);
3859 case SpvOpCopyMemory
:
3860 case SpvOpCopyMemorySized
:
3861 case SpvOpAccessChain
:
3862 case SpvOpPtrAccessChain
:
3863 case SpvOpInBoundsAccessChain
:
3864 case SpvOpArrayLength
:
3865 vtn_handle_variables(b
, opcode
, w
, count
);
3868 case SpvOpFunctionCall
:
3869 vtn_handle_function_call(b
, opcode
, w
, count
);
3872 case SpvOpSampledImage
:
3874 case SpvOpImageSampleImplicitLod
:
3875 case SpvOpImageSampleExplicitLod
:
3876 case SpvOpImageSampleDrefImplicitLod
:
3877 case SpvOpImageSampleDrefExplicitLod
:
3878 case SpvOpImageSampleProjImplicitLod
:
3879 case SpvOpImageSampleProjExplicitLod
:
3880 case SpvOpImageSampleProjDrefImplicitLod
:
3881 case SpvOpImageSampleProjDrefExplicitLod
:
3882 case SpvOpImageFetch
:
3883 case SpvOpImageGather
:
3884 case SpvOpImageDrefGather
:
3885 case SpvOpImageQuerySizeLod
:
3886 case SpvOpImageQueryLod
:
3887 case SpvOpImageQueryLevels
:
3888 case SpvOpImageQuerySamples
:
3889 vtn_handle_texture(b
, opcode
, w
, count
);
3892 case SpvOpImageRead
:
3893 case SpvOpImageWrite
:
3894 case SpvOpImageTexelPointer
:
3895 vtn_handle_image(b
, opcode
, w
, count
);
3898 case SpvOpImageQuerySize
: {
3899 struct vtn_pointer
*image
=
3900 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
3901 if (glsl_type_is_image(image
->type
->type
)) {
3902 vtn_handle_image(b
, opcode
, w
, count
);
3904 vtn_assert(glsl_type_is_sampler(image
->type
->type
));
3905 vtn_handle_texture(b
, opcode
, w
, count
);
3910 case SpvOpAtomicLoad
:
3911 case SpvOpAtomicExchange
:
3912 case SpvOpAtomicCompareExchange
:
3913 case SpvOpAtomicCompareExchangeWeak
:
3914 case SpvOpAtomicIIncrement
:
3915 case SpvOpAtomicIDecrement
:
3916 case SpvOpAtomicIAdd
:
3917 case SpvOpAtomicISub
:
3918 case SpvOpAtomicSMin
:
3919 case SpvOpAtomicUMin
:
3920 case SpvOpAtomicSMax
:
3921 case SpvOpAtomicUMax
:
3922 case SpvOpAtomicAnd
:
3924 case SpvOpAtomicXor
: {
3925 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[3]);
3926 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
3927 vtn_handle_image(b
, opcode
, w
, count
);
3929 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
3930 vtn_handle_atomics(b
, opcode
, w
, count
);
3935 case SpvOpAtomicStore
: {
3936 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[1]);
3937 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
3938 vtn_handle_image(b
, opcode
, w
, count
);
3940 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
3941 vtn_handle_atomics(b
, opcode
, w
, count
);
3947 /* Handle OpSelect up-front here because it needs to be able to handle
3948 * pointers and not just regular vectors and scalars.
3950 struct vtn_value
*res_val
= vtn_untyped_value(b
, w
[2]);
3951 struct vtn_value
*sel_val
= vtn_untyped_value(b
, w
[3]);
3952 struct vtn_value
*obj1_val
= vtn_untyped_value(b
, w
[4]);
3953 struct vtn_value
*obj2_val
= vtn_untyped_value(b
, w
[5]);
3955 const struct glsl_type
*sel_type
;
3956 switch (res_val
->type
->base_type
) {
3957 case vtn_base_type_scalar
:
3958 sel_type
= glsl_bool_type();
3960 case vtn_base_type_vector
:
3961 sel_type
= glsl_vector_type(GLSL_TYPE_BOOL
, res_val
->type
->length
);
3963 case vtn_base_type_pointer
:
3964 /* We need to have actual storage for pointer types */
3965 vtn_fail_if(res_val
->type
->type
== NULL
,
3966 "Invalid pointer result type for OpSelect");
3967 sel_type
= glsl_bool_type();
3970 vtn_fail("Result type of OpSelect must be a scalar, vector, or pointer");
3973 if (unlikely(sel_val
->type
->type
!= sel_type
)) {
3974 if (sel_val
->type
->type
== glsl_bool_type()) {
3975 /* This case is illegal but some older versions of GLSLang produce
3976 * it. The GLSLang issue was fixed on March 30, 2017:
3978 * https://github.com/KhronosGroup/glslang/issues/809
3980 * Unfortunately, there are applications in the wild which are
3981 * shipping with this bug so it isn't nice to fail on them so we
3982 * throw a warning instead. It's not actually a problem for us as
3983 * nir_builder will just splat the condition out which is most
3984 * likely what the client wanted anyway.
3986 vtn_warn("Condition type of OpSelect must have the same number "
3987 "of components as Result Type");
3989 vtn_fail("Condition type of OpSelect must be a scalar or vector "
3990 "of Boolean type. It must have the same number of "
3991 "components as Result Type");
3995 vtn_fail_if(obj1_val
->type
!= res_val
->type
||
3996 obj2_val
->type
!= res_val
->type
,
3997 "Object types must match the result type in OpSelect");
3999 struct vtn_type
*res_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
4000 struct vtn_ssa_value
*ssa
= vtn_create_ssa_value(b
, res_type
->type
);
4001 ssa
->def
= nir_bcsel(&b
->nb
, vtn_ssa_value(b
, w
[3])->def
,
4002 vtn_ssa_value(b
, w
[4])->def
,
4003 vtn_ssa_value(b
, w
[5])->def
);
4004 vtn_push_ssa(b
, w
[2], res_type
, ssa
);
4013 case SpvOpConvertFToU
:
4014 case SpvOpConvertFToS
:
4015 case SpvOpConvertSToF
:
4016 case SpvOpConvertUToF
:
4020 case SpvOpQuantizeToF16
:
4021 case SpvOpConvertPtrToU
:
4022 case SpvOpConvertUToPtr
:
4023 case SpvOpPtrCastToGeneric
:
4024 case SpvOpGenericCastToPtr
:
4030 case SpvOpSignBitSet
:
4031 case SpvOpLessOrGreater
:
4033 case SpvOpUnordered
:
4048 case SpvOpVectorTimesScalar
:
4050 case SpvOpIAddCarry
:
4051 case SpvOpISubBorrow
:
4052 case SpvOpUMulExtended
:
4053 case SpvOpSMulExtended
:
4054 case SpvOpShiftRightLogical
:
4055 case SpvOpShiftRightArithmetic
:
4056 case SpvOpShiftLeftLogical
:
4057 case SpvOpLogicalEqual
:
4058 case SpvOpLogicalNotEqual
:
4059 case SpvOpLogicalOr
:
4060 case SpvOpLogicalAnd
:
4061 case SpvOpLogicalNot
:
4062 case SpvOpBitwiseOr
:
4063 case SpvOpBitwiseXor
:
4064 case SpvOpBitwiseAnd
:
4066 case SpvOpFOrdEqual
:
4067 case SpvOpFUnordEqual
:
4068 case SpvOpINotEqual
:
4069 case SpvOpFOrdNotEqual
:
4070 case SpvOpFUnordNotEqual
:
4071 case SpvOpULessThan
:
4072 case SpvOpSLessThan
:
4073 case SpvOpFOrdLessThan
:
4074 case SpvOpFUnordLessThan
:
4075 case SpvOpUGreaterThan
:
4076 case SpvOpSGreaterThan
:
4077 case SpvOpFOrdGreaterThan
:
4078 case SpvOpFUnordGreaterThan
:
4079 case SpvOpULessThanEqual
:
4080 case SpvOpSLessThanEqual
:
4081 case SpvOpFOrdLessThanEqual
:
4082 case SpvOpFUnordLessThanEqual
:
4083 case SpvOpUGreaterThanEqual
:
4084 case SpvOpSGreaterThanEqual
:
4085 case SpvOpFOrdGreaterThanEqual
:
4086 case SpvOpFUnordGreaterThanEqual
:
4092 case SpvOpFwidthFine
:
4093 case SpvOpDPdxCoarse
:
4094 case SpvOpDPdyCoarse
:
4095 case SpvOpFwidthCoarse
:
4096 case SpvOpBitFieldInsert
:
4097 case SpvOpBitFieldSExtract
:
4098 case SpvOpBitFieldUExtract
:
4099 case SpvOpBitReverse
:
4101 case SpvOpTranspose
:
4102 case SpvOpOuterProduct
:
4103 case SpvOpMatrixTimesScalar
:
4104 case SpvOpVectorTimesMatrix
:
4105 case SpvOpMatrixTimesVector
:
4106 case SpvOpMatrixTimesMatrix
:
4107 vtn_handle_alu(b
, opcode
, w
, count
);
4110 case SpvOpVectorExtractDynamic
:
4111 case SpvOpVectorInsertDynamic
:
4112 case SpvOpVectorShuffle
:
4113 case SpvOpCompositeConstruct
:
4114 case SpvOpCompositeExtract
:
4115 case SpvOpCompositeInsert
:
4116 case SpvOpCopyObject
:
4117 vtn_handle_composite(b
, opcode
, w
, count
);
4120 case SpvOpEmitVertex
:
4121 case SpvOpEndPrimitive
:
4122 case SpvOpEmitStreamVertex
:
4123 case SpvOpEndStreamPrimitive
:
4124 case SpvOpControlBarrier
:
4125 case SpvOpMemoryBarrier
:
4126 vtn_handle_barrier(b
, opcode
, w
, count
);
4129 case SpvOpGroupNonUniformElect
:
4130 case SpvOpGroupNonUniformAll
:
4131 case SpvOpGroupNonUniformAny
:
4132 case SpvOpGroupNonUniformAllEqual
:
4133 case SpvOpGroupNonUniformBroadcast
:
4134 case SpvOpGroupNonUniformBroadcastFirst
:
4135 case SpvOpGroupNonUniformBallot
:
4136 case SpvOpGroupNonUniformInverseBallot
:
4137 case SpvOpGroupNonUniformBallotBitExtract
:
4138 case SpvOpGroupNonUniformBallotBitCount
:
4139 case SpvOpGroupNonUniformBallotFindLSB
:
4140 case SpvOpGroupNonUniformBallotFindMSB
:
4141 case SpvOpGroupNonUniformShuffle
:
4142 case SpvOpGroupNonUniformShuffleXor
:
4143 case SpvOpGroupNonUniformShuffleUp
:
4144 case SpvOpGroupNonUniformShuffleDown
:
4145 case SpvOpGroupNonUniformIAdd
:
4146 case SpvOpGroupNonUniformFAdd
:
4147 case SpvOpGroupNonUniformIMul
:
4148 case SpvOpGroupNonUniformFMul
:
4149 case SpvOpGroupNonUniformSMin
:
4150 case SpvOpGroupNonUniformUMin
:
4151 case SpvOpGroupNonUniformFMin
:
4152 case SpvOpGroupNonUniformSMax
:
4153 case SpvOpGroupNonUniformUMax
:
4154 case SpvOpGroupNonUniformFMax
:
4155 case SpvOpGroupNonUniformBitwiseAnd
:
4156 case SpvOpGroupNonUniformBitwiseOr
:
4157 case SpvOpGroupNonUniformBitwiseXor
:
4158 case SpvOpGroupNonUniformLogicalAnd
:
4159 case SpvOpGroupNonUniformLogicalOr
:
4160 case SpvOpGroupNonUniformLogicalXor
:
4161 case SpvOpGroupNonUniformQuadBroadcast
:
4162 case SpvOpGroupNonUniformQuadSwap
:
4163 vtn_handle_subgroup(b
, opcode
, w
, count
);
4167 vtn_fail("Unhandled opcode");
4174 vtn_create_builder(const uint32_t *words
, size_t word_count
,
4175 gl_shader_stage stage
, const char *entry_point_name
,
4176 const struct spirv_to_nir_options
*options
)
4178 /* Initialize the vtn_builder object */
4179 struct vtn_builder
*b
= rzalloc(NULL
, struct vtn_builder
);
4181 b
->spirv_word_count
= word_count
;
4185 exec_list_make_empty(&b
->functions
);
4186 b
->entry_point_stage
= stage
;
4187 b
->entry_point_name
= entry_point_name
;
4188 b
->options
= options
;
4191 * Handle the SPIR-V header (first 5 dwords).
4192 * Can't use vtx_assert() as the setjmp(3) target isn't initialized yet.
4194 if (word_count
<= 5)
4197 if (words
[0] != SpvMagicNumber
) {
4198 vtn_err("words[0] was 0x%x, want 0x%x", words
[0], SpvMagicNumber
);
4201 if (words
[1] < 0x10000) {
4202 vtn_err("words[1] was 0x%x, want >= 0x10000", words
[1]);
4206 /* words[2] == generator magic */
4207 unsigned value_id_bound
= words
[3];
4208 if (words
[4] != 0) {
4209 vtn_err("words[4] was %u, want 0", words
[4]);
4213 b
->value_id_bound
= value_id_bound
;
4214 b
->values
= rzalloc_array(b
, struct vtn_value
, value_id_bound
);
4223 spirv_to_nir(const uint32_t *words
, size_t word_count
,
4224 struct nir_spirv_specialization
*spec
, unsigned num_spec
,
4225 gl_shader_stage stage
, const char *entry_point_name
,
4226 const struct spirv_to_nir_options
*options
,
4227 const nir_shader_compiler_options
*nir_options
)
4230 const uint32_t *word_end
= words
+ word_count
;
4232 struct vtn_builder
*b
= vtn_create_builder(words
, word_count
,
4233 stage
, entry_point_name
,
4239 /* See also _vtn_fail() */
4240 if (setjmp(b
->fail_jump
)) {
4245 /* Skip the SPIR-V header, handled at vtn_create_builder */
4248 /* Handle all the preamble instructions */
4249 words
= vtn_foreach_instruction(b
, words
, word_end
,
4250 vtn_handle_preamble_instruction
);
4252 if (b
->entry_point
== NULL
) {
4253 vtn_fail("Entry point not found");
4258 b
->shader
= nir_shader_create(b
, stage
, nir_options
, NULL
);
4260 /* Set shader info defaults */
4261 b
->shader
->info
.gs
.invocations
= 1;
4263 /* Parse execution modes */
4264 vtn_foreach_execution_mode(b
, b
->entry_point
,
4265 vtn_handle_execution_mode
, NULL
);
4267 b
->specializations
= spec
;
4268 b
->num_specializations
= num_spec
;
4270 /* Handle all variable, type, and constant instructions */
4271 words
= vtn_foreach_instruction(b
, words
, word_end
,
4272 vtn_handle_variable_or_type_instruction
);
4274 /* Set types on all vtn_values */
4275 vtn_foreach_instruction(b
, words
, word_end
, vtn_set_instruction_result_type
);
4277 vtn_build_cfg(b
, words
, word_end
);
4279 assert(b
->entry_point
->value_type
== vtn_value_type_function
);
4280 b
->entry_point
->func
->referenced
= true;
4285 foreach_list_typed(struct vtn_function
, func
, node
, &b
->functions
) {
4286 if (func
->referenced
&& !func
->emitted
) {
4287 b
->const_table
= _mesa_hash_table_create(b
, _mesa_hash_pointer
,
4288 _mesa_key_pointer_equal
);
4290 vtn_function_emit(b
, func
, vtn_handle_body_instruction
);
4296 /* We sometimes generate bogus derefs that, while never used, give the
4297 * validator a bit of heartburn. Run dead code to get rid of them.
4299 nir_opt_dce(b
->shader
);
4301 vtn_assert(b
->entry_point
->value_type
== vtn_value_type_function
);
4302 nir_function
*entry_point
= b
->entry_point
->func
->impl
->function
;
4303 vtn_assert(entry_point
);
4305 /* Unparent the shader from the vtn_builder before we delete the builder */
4306 ralloc_steal(NULL
, b
->shader
);