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"
35 #include "util/u_math.h"
40 vtn_log(struct vtn_builder
*b
, enum nir_spirv_debug_level level
,
41 size_t spirv_offset
, const char *message
)
43 if (b
->options
->debug
.func
) {
44 b
->options
->debug
.func(b
->options
->debug
.private_data
,
45 level
, spirv_offset
, message
);
49 if (level
>= NIR_SPIRV_DEBUG_LEVEL_WARNING
)
50 fprintf(stderr
, "%s\n", message
);
55 vtn_logf(struct vtn_builder
*b
, enum nir_spirv_debug_level level
,
56 size_t spirv_offset
, const char *fmt
, ...)
62 msg
= ralloc_vasprintf(NULL
, fmt
, args
);
65 vtn_log(b
, level
, spirv_offset
, msg
);
71 vtn_log_err(struct vtn_builder
*b
,
72 enum nir_spirv_debug_level level
, const char *prefix
,
73 const char *file
, unsigned line
,
74 const char *fmt
, va_list args
)
78 msg
= ralloc_strdup(NULL
, prefix
);
81 ralloc_asprintf_append(&msg
, " In file %s:%u\n", file
, line
);
84 ralloc_asprintf_append(&msg
, " ");
86 ralloc_vasprintf_append(&msg
, fmt
, args
);
88 ralloc_asprintf_append(&msg
, "\n %zu bytes into the SPIR-V binary",
92 ralloc_asprintf_append(&msg
,
93 "\n in SPIR-V source file %s, line %d, col %d",
94 b
->file
, b
->line
, b
->col
);
97 vtn_log(b
, level
, b
->spirv_offset
, msg
);
103 vtn_dump_shader(struct vtn_builder
*b
, const char *path
, const char *prefix
)
108 int len
= snprintf(filename
, sizeof(filename
), "%s/%s-%d.spirv",
109 path
, prefix
, idx
++);
110 if (len
< 0 || len
>= sizeof(filename
))
113 FILE *f
= fopen(filename
, "w");
117 fwrite(b
->spirv
, sizeof(*b
->spirv
), b
->spirv_word_count
, f
);
120 vtn_info("SPIR-V shader dumped to %s", filename
);
124 _vtn_warn(struct vtn_builder
*b
, const char *file
, unsigned line
,
125 const char *fmt
, ...)
130 vtn_log_err(b
, NIR_SPIRV_DEBUG_LEVEL_WARNING
, "SPIR-V WARNING:\n",
131 file
, line
, fmt
, args
);
136 _vtn_err(struct vtn_builder
*b
, const char *file
, unsigned line
,
137 const char *fmt
, ...)
142 vtn_log_err(b
, NIR_SPIRV_DEBUG_LEVEL_ERROR
, "SPIR-V ERROR:\n",
143 file
, line
, fmt
, args
);
148 _vtn_fail(struct vtn_builder
*b
, const char *file
, unsigned line
,
149 const char *fmt
, ...)
154 vtn_log_err(b
, NIR_SPIRV_DEBUG_LEVEL_ERROR
, "SPIR-V parsing FAILED:\n",
155 file
, line
, fmt
, args
);
158 const char *dump_path
= getenv("MESA_SPIRV_FAIL_DUMP_PATH");
160 vtn_dump_shader(b
, dump_path
, "fail");
162 longjmp(b
->fail_jump
, 1);
165 struct spec_constant_value
{
173 static struct vtn_ssa_value
*
174 vtn_undef_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
176 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
179 if (glsl_type_is_vector_or_scalar(type
)) {
180 unsigned num_components
= glsl_get_vector_elements(val
->type
);
181 unsigned bit_size
= glsl_get_bit_size(val
->type
);
182 val
->def
= nir_ssa_undef(&b
->nb
, num_components
, bit_size
);
184 unsigned elems
= glsl_get_length(val
->type
);
185 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
186 if (glsl_type_is_matrix(type
)) {
187 const struct glsl_type
*elem_type
=
188 glsl_vector_type(glsl_get_base_type(type
),
189 glsl_get_vector_elements(type
));
191 for (unsigned i
= 0; i
< elems
; i
++)
192 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
193 } else if (glsl_type_is_array(type
)) {
194 const struct glsl_type
*elem_type
= glsl_get_array_element(type
);
195 for (unsigned i
= 0; i
< elems
; i
++)
196 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
198 for (unsigned i
= 0; i
< elems
; i
++) {
199 const struct glsl_type
*elem_type
= glsl_get_struct_field(type
, i
);
200 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
208 static struct vtn_ssa_value
*
209 vtn_const_ssa_value(struct vtn_builder
*b
, nir_constant
*constant
,
210 const struct glsl_type
*type
)
212 struct hash_entry
*entry
= _mesa_hash_table_search(b
->const_table
, constant
);
217 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
220 switch (glsl_get_base_type(type
)) {
223 case GLSL_TYPE_INT16
:
224 case GLSL_TYPE_UINT16
:
225 case GLSL_TYPE_UINT8
:
227 case GLSL_TYPE_INT64
:
228 case GLSL_TYPE_UINT64
:
230 case GLSL_TYPE_FLOAT
:
231 case GLSL_TYPE_FLOAT16
:
232 case GLSL_TYPE_DOUBLE
: {
233 int bit_size
= glsl_get_bit_size(type
);
234 if (glsl_type_is_vector_or_scalar(type
)) {
235 unsigned num_components
= glsl_get_vector_elements(val
->type
);
236 nir_load_const_instr
*load
=
237 nir_load_const_instr_create(b
->shader
, num_components
, bit_size
);
239 memcpy(load
->value
, constant
->values
,
240 sizeof(nir_const_value
) * load
->def
.num_components
);
242 nir_instr_insert_before_cf_list(&b
->nb
.impl
->body
, &load
->instr
);
243 val
->def
= &load
->def
;
245 assert(glsl_type_is_matrix(type
));
246 unsigned columns
= glsl_get_matrix_columns(val
->type
);
247 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, columns
);
248 const struct glsl_type
*column_type
= glsl_get_column_type(val
->type
);
249 for (unsigned i
= 0; i
< columns
; i
++)
250 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
256 case GLSL_TYPE_ARRAY
: {
257 unsigned elems
= glsl_get_length(val
->type
);
258 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
259 const struct glsl_type
*elem_type
= glsl_get_array_element(val
->type
);
260 for (unsigned i
= 0; i
< elems
; i
++)
261 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
266 case GLSL_TYPE_STRUCT
: {
267 unsigned elems
= glsl_get_length(val
->type
);
268 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
269 for (unsigned i
= 0; i
< elems
; i
++) {
270 const struct glsl_type
*elem_type
=
271 glsl_get_struct_field(val
->type
, i
);
272 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
279 vtn_fail("bad constant type");
285 struct vtn_ssa_value
*
286 vtn_ssa_value(struct vtn_builder
*b
, uint32_t value_id
)
288 struct vtn_value
*val
= vtn_untyped_value(b
, value_id
);
289 switch (val
->value_type
) {
290 case vtn_value_type_undef
:
291 return vtn_undef_ssa_value(b
, val
->type
->type
);
293 case vtn_value_type_constant
:
294 return vtn_const_ssa_value(b
, val
->constant
, val
->type
->type
);
296 case vtn_value_type_ssa
:
299 case vtn_value_type_pointer
:
300 vtn_assert(val
->pointer
->ptr_type
&& val
->pointer
->ptr_type
->type
);
301 struct vtn_ssa_value
*ssa
=
302 vtn_create_ssa_value(b
, val
->pointer
->ptr_type
->type
);
303 ssa
->def
= vtn_pointer_to_ssa(b
, val
->pointer
);
307 vtn_fail("Invalid type for an SSA value");
312 vtn_string_literal(struct vtn_builder
*b
, const uint32_t *words
,
313 unsigned word_count
, unsigned *words_used
)
315 char *dup
= ralloc_strndup(b
, (char *)words
, word_count
* sizeof(*words
));
317 /* Ammount of space taken by the string (including the null) */
318 unsigned len
= strlen(dup
) + 1;
319 *words_used
= DIV_ROUND_UP(len
, sizeof(*words
));
325 vtn_foreach_instruction(struct vtn_builder
*b
, const uint32_t *start
,
326 const uint32_t *end
, vtn_instruction_handler handler
)
332 const uint32_t *w
= start
;
334 SpvOp opcode
= w
[0] & SpvOpCodeMask
;
335 unsigned count
= w
[0] >> SpvWordCountShift
;
336 vtn_assert(count
>= 1 && w
+ count
<= end
);
338 b
->spirv_offset
= (uint8_t *)w
- (uint8_t *)b
->spirv
;
342 break; /* Do nothing */
345 b
->file
= vtn_value(b
, w
[1], vtn_value_type_string
)->str
;
357 if (!handler(b
, opcode
, w
, count
))
375 vtn_handle_non_semantic_instruction(struct vtn_builder
*b
, SpvOp ext_opcode
,
376 const uint32_t *w
, unsigned 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
.amd_gcn_shader
)) {
394 val
->ext_handler
= vtn_handle_amd_gcn_shader_instruction
;
395 } else if ((strcmp(ext
, "SPV_AMD_shader_ballot") == 0)
396 && (b
->options
&& b
->options
->caps
.amd_shader_ballot
)) {
397 val
->ext_handler
= vtn_handle_amd_shader_ballot_instruction
;
398 } else if ((strcmp(ext
, "SPV_AMD_shader_trinary_minmax") == 0)
399 && (b
->options
&& b
->options
->caps
.amd_trinary_minmax
)) {
400 val
->ext_handler
= vtn_handle_amd_shader_trinary_minmax_instruction
;
401 } else if (strcmp(ext
, "OpenCL.std") == 0) {
402 val
->ext_handler
= vtn_handle_opencl_instruction
;
403 } else if (strstr(ext
, "NonSemantic.") == ext
) {
404 val
->ext_handler
= vtn_handle_non_semantic_instruction
;
406 vtn_fail("Unsupported extension: %s", ext
);
412 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
413 bool handled
= val
->ext_handler(b
, w
[4], w
, count
);
419 vtn_fail_with_opcode("Unhandled opcode", opcode
);
424 _foreach_decoration_helper(struct vtn_builder
*b
,
425 struct vtn_value
*base_value
,
427 struct vtn_value
*value
,
428 vtn_decoration_foreach_cb cb
, void *data
)
430 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
432 if (dec
->scope
== VTN_DEC_DECORATION
) {
433 member
= parent_member
;
434 } else if (dec
->scope
>= VTN_DEC_STRUCT_MEMBER0
) {
435 vtn_fail_if(value
->value_type
!= vtn_value_type_type
||
436 value
->type
->base_type
!= vtn_base_type_struct
,
437 "OpMemberDecorate and OpGroupMemberDecorate are only "
438 "allowed on OpTypeStruct");
439 /* This means we haven't recursed yet */
440 assert(value
== base_value
);
442 member
= dec
->scope
- VTN_DEC_STRUCT_MEMBER0
;
444 vtn_fail_if(member
>= base_value
->type
->length
,
445 "OpMemberDecorate specifies member %d but the "
446 "OpTypeStruct has only %u members",
447 member
, base_value
->type
->length
);
449 /* Not a decoration */
450 assert(dec
->scope
== VTN_DEC_EXECUTION_MODE
);
455 assert(dec
->group
->value_type
== vtn_value_type_decoration_group
);
456 _foreach_decoration_helper(b
, base_value
, member
, dec
->group
,
459 cb(b
, base_value
, member
, dec
, data
);
464 /** Iterates (recursively if needed) over all of the decorations on a value
466 * This function iterates over all of the decorations applied to a given
467 * value. If it encounters a decoration group, it recurses into the group
468 * and iterates over all of those decorations as well.
471 vtn_foreach_decoration(struct vtn_builder
*b
, struct vtn_value
*value
,
472 vtn_decoration_foreach_cb cb
, void *data
)
474 _foreach_decoration_helper(b
, value
, -1, value
, cb
, data
);
478 vtn_foreach_execution_mode(struct vtn_builder
*b
, struct vtn_value
*value
,
479 vtn_execution_mode_foreach_cb cb
, void *data
)
481 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
482 if (dec
->scope
!= VTN_DEC_EXECUTION_MODE
)
485 assert(dec
->group
== NULL
);
486 cb(b
, value
, dec
, data
);
491 vtn_handle_decoration(struct vtn_builder
*b
, SpvOp opcode
,
492 const uint32_t *w
, unsigned count
)
494 const uint32_t *w_end
= w
+ count
;
495 const uint32_t target
= w
[1];
499 case SpvOpDecorationGroup
:
500 vtn_push_value(b
, target
, vtn_value_type_decoration_group
);
504 case SpvOpDecorateId
:
505 case SpvOpMemberDecorate
:
506 case SpvOpDecorateString
:
507 case SpvOpMemberDecorateString
:
508 case SpvOpExecutionMode
:
509 case SpvOpExecutionModeId
: {
510 struct vtn_value
*val
= vtn_untyped_value(b
, target
);
512 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
515 case SpvOpDecorateId
:
516 case SpvOpDecorateString
:
517 dec
->scope
= VTN_DEC_DECORATION
;
519 case SpvOpMemberDecorate
:
520 case SpvOpMemberDecorateString
:
521 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(w
++);
522 vtn_fail_if(dec
->scope
< VTN_DEC_STRUCT_MEMBER0
, /* overflow */
523 "Member argument of OpMemberDecorate too large");
525 case SpvOpExecutionMode
:
526 case SpvOpExecutionModeId
:
527 dec
->scope
= VTN_DEC_EXECUTION_MODE
;
530 unreachable("Invalid decoration opcode");
532 dec
->decoration
= *(w
++);
535 /* Link into the list */
536 dec
->next
= val
->decoration
;
537 val
->decoration
= dec
;
541 case SpvOpGroupMemberDecorate
:
542 case SpvOpGroupDecorate
: {
543 struct vtn_value
*group
=
544 vtn_value(b
, target
, vtn_value_type_decoration_group
);
546 for (; w
< w_end
; w
++) {
547 struct vtn_value
*val
= vtn_untyped_value(b
, *w
);
548 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
551 if (opcode
== SpvOpGroupDecorate
) {
552 dec
->scope
= VTN_DEC_DECORATION
;
554 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(++w
);
555 vtn_fail_if(dec
->scope
< 0, /* Check for overflow */
556 "Member argument of OpGroupMemberDecorate too large");
559 /* Link into the list */
560 dec
->next
= val
->decoration
;
561 val
->decoration
= dec
;
567 unreachable("Unhandled opcode");
571 struct member_decoration_ctx
{
573 struct glsl_struct_field
*fields
;
574 struct vtn_type
*type
;
578 * Returns true if the given type contains a struct decorated Block or
582 vtn_type_contains_block(struct vtn_builder
*b
, struct vtn_type
*type
)
584 switch (type
->base_type
) {
585 case vtn_base_type_array
:
586 return vtn_type_contains_block(b
, type
->array_element
);
587 case vtn_base_type_struct
:
588 if (type
->block
|| type
->buffer_block
)
590 for (unsigned i
= 0; i
< type
->length
; i
++) {
591 if (vtn_type_contains_block(b
, type
->members
[i
]))
600 /** Returns true if two types are "compatible", i.e. you can do an OpLoad,
601 * OpStore, or OpCopyMemory between them without breaking anything.
602 * Technically, the SPIR-V rules require the exact same type ID but this lets
603 * us internally be a bit looser.
606 vtn_types_compatible(struct vtn_builder
*b
,
607 struct vtn_type
*t1
, struct vtn_type
*t2
)
609 if (t1
->id
== t2
->id
)
612 if (t1
->base_type
!= t2
->base_type
)
615 switch (t1
->base_type
) {
616 case vtn_base_type_void
:
617 case vtn_base_type_scalar
:
618 case vtn_base_type_vector
:
619 case vtn_base_type_matrix
:
620 case vtn_base_type_image
:
621 case vtn_base_type_sampler
:
622 case vtn_base_type_sampled_image
:
623 return t1
->type
== t2
->type
;
625 case vtn_base_type_array
:
626 return t1
->length
== t2
->length
&&
627 vtn_types_compatible(b
, t1
->array_element
, t2
->array_element
);
629 case vtn_base_type_pointer
:
630 return vtn_types_compatible(b
, t1
->deref
, t2
->deref
);
632 case vtn_base_type_struct
:
633 if (t1
->length
!= t2
->length
)
636 for (unsigned i
= 0; i
< t1
->length
; i
++) {
637 if (!vtn_types_compatible(b
, t1
->members
[i
], t2
->members
[i
]))
642 case vtn_base_type_function
:
643 /* This case shouldn't get hit since you can't copy around function
644 * types. Just require them to be identical.
649 vtn_fail("Invalid base type");
653 vtn_type_without_array(struct vtn_type
*type
)
655 while (type
->base_type
== vtn_base_type_array
)
656 type
= type
->array_element
;
660 /* does a shallow copy of a vtn_type */
662 static struct vtn_type
*
663 vtn_type_copy(struct vtn_builder
*b
, struct vtn_type
*src
)
665 struct vtn_type
*dest
= ralloc(b
, struct vtn_type
);
668 switch (src
->base_type
) {
669 case vtn_base_type_void
:
670 case vtn_base_type_scalar
:
671 case vtn_base_type_vector
:
672 case vtn_base_type_matrix
:
673 case vtn_base_type_array
:
674 case vtn_base_type_pointer
:
675 case vtn_base_type_image
:
676 case vtn_base_type_sampler
:
677 case vtn_base_type_sampled_image
:
678 /* Nothing more to do */
681 case vtn_base_type_struct
:
682 dest
->members
= ralloc_array(b
, struct vtn_type
*, src
->length
);
683 memcpy(dest
->members
, src
->members
,
684 src
->length
* sizeof(src
->members
[0]));
686 dest
->offsets
= ralloc_array(b
, unsigned, src
->length
);
687 memcpy(dest
->offsets
, src
->offsets
,
688 src
->length
* sizeof(src
->offsets
[0]));
691 case vtn_base_type_function
:
692 dest
->params
= ralloc_array(b
, struct vtn_type
*, src
->length
);
693 memcpy(dest
->params
, src
->params
, src
->length
* sizeof(src
->params
[0]));
700 static struct vtn_type
*
701 mutable_matrix_member(struct vtn_builder
*b
, struct vtn_type
*type
, int member
)
703 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
704 type
= type
->members
[member
];
706 /* We may have an array of matrices.... Oh, joy! */
707 while (glsl_type_is_array(type
->type
)) {
708 type
->array_element
= vtn_type_copy(b
, type
->array_element
);
709 type
= type
->array_element
;
712 vtn_assert(glsl_type_is_matrix(type
->type
));
718 vtn_handle_access_qualifier(struct vtn_builder
*b
, struct vtn_type
*type
,
719 int member
, enum gl_access_qualifier access
)
721 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
722 type
= type
->members
[member
];
724 type
->access
|= access
;
728 array_stride_decoration_cb(struct vtn_builder
*b
,
729 struct vtn_value
*val
, int member
,
730 const struct vtn_decoration
*dec
, void *void_ctx
)
732 struct vtn_type
*type
= val
->type
;
734 if (dec
->decoration
== SpvDecorationArrayStride
) {
735 if (vtn_type_contains_block(b
, type
)) {
736 vtn_warn("The ArrayStride decoration cannot be applied to an array "
737 "type which contains a structure type decorated Block "
739 /* Ignore the decoration */
741 vtn_fail_if(dec
->operands
[0] == 0, "ArrayStride must be non-zero");
742 type
->stride
= dec
->operands
[0];
748 struct_member_decoration_cb(struct vtn_builder
*b
,
749 struct vtn_value
*val
, int member
,
750 const struct vtn_decoration
*dec
, void *void_ctx
)
752 struct member_decoration_ctx
*ctx
= void_ctx
;
757 assert(member
< ctx
->num_fields
);
759 switch (dec
->decoration
) {
760 case SpvDecorationRelaxedPrecision
:
761 case SpvDecorationUniform
:
762 case SpvDecorationUniformId
:
763 break; /* FIXME: Do nothing with this for now. */
764 case SpvDecorationNonWritable
:
765 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_NON_WRITEABLE
);
767 case SpvDecorationNonReadable
:
768 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_NON_READABLE
);
770 case SpvDecorationVolatile
:
771 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_VOLATILE
);
773 case SpvDecorationCoherent
:
774 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_COHERENT
);
776 case SpvDecorationNoPerspective
:
777 ctx
->fields
[member
].interpolation
= INTERP_MODE_NOPERSPECTIVE
;
779 case SpvDecorationFlat
:
780 ctx
->fields
[member
].interpolation
= INTERP_MODE_FLAT
;
782 case SpvDecorationCentroid
:
783 ctx
->fields
[member
].centroid
= true;
785 case SpvDecorationSample
:
786 ctx
->fields
[member
].sample
= true;
788 case SpvDecorationStream
:
789 /* Vulkan only allows one GS stream */
790 vtn_assert(dec
->operands
[0] == 0);
792 case SpvDecorationLocation
:
793 ctx
->fields
[member
].location
= dec
->operands
[0];
795 case SpvDecorationComponent
:
796 break; /* FIXME: What should we do with these? */
797 case SpvDecorationBuiltIn
:
798 ctx
->type
->members
[member
] = vtn_type_copy(b
, ctx
->type
->members
[member
]);
799 ctx
->type
->members
[member
]->is_builtin
= true;
800 ctx
->type
->members
[member
]->builtin
= dec
->operands
[0];
801 ctx
->type
->builtin_block
= true;
803 case SpvDecorationOffset
:
804 ctx
->type
->offsets
[member
] = dec
->operands
[0];
805 ctx
->fields
[member
].offset
= dec
->operands
[0];
807 case SpvDecorationMatrixStride
:
808 /* Handled as a second pass */
810 case SpvDecorationColMajor
:
811 break; /* Nothing to do here. Column-major is the default. */
812 case SpvDecorationRowMajor
:
813 mutable_matrix_member(b
, ctx
->type
, member
)->row_major
= true;
816 case SpvDecorationPatch
:
819 case SpvDecorationSpecId
:
820 case SpvDecorationBlock
:
821 case SpvDecorationBufferBlock
:
822 case SpvDecorationArrayStride
:
823 case SpvDecorationGLSLShared
:
824 case SpvDecorationGLSLPacked
:
825 case SpvDecorationInvariant
:
826 case SpvDecorationRestrict
:
827 case SpvDecorationAliased
:
828 case SpvDecorationConstant
:
829 case SpvDecorationIndex
:
830 case SpvDecorationBinding
:
831 case SpvDecorationDescriptorSet
:
832 case SpvDecorationLinkageAttributes
:
833 case SpvDecorationNoContraction
:
834 case SpvDecorationInputAttachmentIndex
:
835 vtn_warn("Decoration not allowed on struct members: %s",
836 spirv_decoration_to_string(dec
->decoration
));
839 case SpvDecorationXfbBuffer
:
840 case SpvDecorationXfbStride
:
841 vtn_warn("Vulkan does not have transform feedback");
844 case SpvDecorationCPacked
:
845 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
)
846 vtn_warn("Decoration only allowed for CL-style kernels: %s",
847 spirv_decoration_to_string(dec
->decoration
));
849 ctx
->type
->packed
= true;
852 case SpvDecorationSaturatedConversion
:
853 case SpvDecorationFuncParamAttr
:
854 case SpvDecorationFPRoundingMode
:
855 case SpvDecorationFPFastMathMode
:
856 case SpvDecorationAlignment
:
857 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
) {
858 vtn_warn("Decoration only allowed for CL-style kernels: %s",
859 spirv_decoration_to_string(dec
->decoration
));
863 case SpvDecorationUserSemantic
:
864 /* User semantic decorations can safely be ignored by the driver. */
868 vtn_fail_with_decoration("Unhandled decoration", dec
->decoration
);
872 /** Chases the array type all the way down to the tail and rewrites the
873 * glsl_types to be based off the tail's glsl_type.
876 vtn_array_type_rewrite_glsl_type(struct vtn_type
*type
)
878 if (type
->base_type
!= vtn_base_type_array
)
881 vtn_array_type_rewrite_glsl_type(type
->array_element
);
883 type
->type
= glsl_array_type(type
->array_element
->type
,
884 type
->length
, type
->stride
);
887 /* Matrix strides are handled as a separate pass because we need to know
888 * whether the matrix is row-major or not first.
891 struct_member_matrix_stride_cb(struct vtn_builder
*b
,
892 struct vtn_value
*val
, int member
,
893 const struct vtn_decoration
*dec
,
896 if (dec
->decoration
!= SpvDecorationMatrixStride
)
899 vtn_fail_if(member
< 0,
900 "The MatrixStride decoration is only allowed on members "
902 vtn_fail_if(dec
->operands
[0] == 0, "MatrixStride must be non-zero");
904 struct member_decoration_ctx
*ctx
= void_ctx
;
906 struct vtn_type
*mat_type
= mutable_matrix_member(b
, ctx
->type
, member
);
907 if (mat_type
->row_major
) {
908 mat_type
->array_element
= vtn_type_copy(b
, mat_type
->array_element
);
909 mat_type
->stride
= mat_type
->array_element
->stride
;
910 mat_type
->array_element
->stride
= dec
->operands
[0];
912 mat_type
->type
= glsl_explicit_matrix_type(mat_type
->type
,
913 dec
->operands
[0], true);
914 mat_type
->array_element
->type
= glsl_get_column_type(mat_type
->type
);
916 vtn_assert(mat_type
->array_element
->stride
> 0);
917 mat_type
->stride
= dec
->operands
[0];
919 mat_type
->type
= glsl_explicit_matrix_type(mat_type
->type
,
920 dec
->operands
[0], false);
923 /* Now that we've replaced the glsl_type with a properly strided matrix
924 * type, rewrite the member type so that it's an array of the proper kind
927 vtn_array_type_rewrite_glsl_type(ctx
->type
->members
[member
]);
928 ctx
->fields
[member
].type
= ctx
->type
->members
[member
]->type
;
932 struct_block_decoration_cb(struct vtn_builder
*b
,
933 struct vtn_value
*val
, int member
,
934 const struct vtn_decoration
*dec
, void *ctx
)
939 struct vtn_type
*type
= val
->type
;
940 if (dec
->decoration
== SpvDecorationBlock
)
942 else if (dec
->decoration
== SpvDecorationBufferBlock
)
943 type
->buffer_block
= true;
947 type_decoration_cb(struct vtn_builder
*b
,
948 struct vtn_value
*val
, int member
,
949 const struct vtn_decoration
*dec
, void *ctx
)
951 struct vtn_type
*type
= val
->type
;
954 /* This should have been handled by OpTypeStruct */
955 assert(val
->type
->base_type
== vtn_base_type_struct
);
956 assert(member
>= 0 && member
< val
->type
->length
);
960 switch (dec
->decoration
) {
961 case SpvDecorationArrayStride
:
962 vtn_assert(type
->base_type
== vtn_base_type_array
||
963 type
->base_type
== vtn_base_type_pointer
);
965 case SpvDecorationBlock
:
966 vtn_assert(type
->base_type
== vtn_base_type_struct
);
967 vtn_assert(type
->block
);
969 case SpvDecorationBufferBlock
:
970 vtn_assert(type
->base_type
== vtn_base_type_struct
);
971 vtn_assert(type
->buffer_block
);
973 case SpvDecorationGLSLShared
:
974 case SpvDecorationGLSLPacked
:
975 /* Ignore these, since we get explicit offsets anyways */
978 case SpvDecorationRowMajor
:
979 case SpvDecorationColMajor
:
980 case SpvDecorationMatrixStride
:
981 case SpvDecorationBuiltIn
:
982 case SpvDecorationNoPerspective
:
983 case SpvDecorationFlat
:
984 case SpvDecorationPatch
:
985 case SpvDecorationCentroid
:
986 case SpvDecorationSample
:
987 case SpvDecorationVolatile
:
988 case SpvDecorationCoherent
:
989 case SpvDecorationNonWritable
:
990 case SpvDecorationNonReadable
:
991 case SpvDecorationUniform
:
992 case SpvDecorationUniformId
:
993 case SpvDecorationLocation
:
994 case SpvDecorationComponent
:
995 case SpvDecorationOffset
:
996 case SpvDecorationXfbBuffer
:
997 case SpvDecorationXfbStride
:
998 case SpvDecorationUserSemantic
:
999 vtn_warn("Decoration only allowed for struct members: %s",
1000 spirv_decoration_to_string(dec
->decoration
));
1003 case SpvDecorationStream
:
1004 /* We don't need to do anything here, as stream is filled up when
1005 * aplying the decoration to a variable, just check that if it is not a
1006 * struct member, it should be a struct.
1008 vtn_assert(type
->base_type
== vtn_base_type_struct
);
1011 case SpvDecorationRelaxedPrecision
:
1012 case SpvDecorationSpecId
:
1013 case SpvDecorationInvariant
:
1014 case SpvDecorationRestrict
:
1015 case SpvDecorationAliased
:
1016 case SpvDecorationConstant
:
1017 case SpvDecorationIndex
:
1018 case SpvDecorationBinding
:
1019 case SpvDecorationDescriptorSet
:
1020 case SpvDecorationLinkageAttributes
:
1021 case SpvDecorationNoContraction
:
1022 case SpvDecorationInputAttachmentIndex
:
1023 vtn_warn("Decoration not allowed on types: %s",
1024 spirv_decoration_to_string(dec
->decoration
));
1027 case SpvDecorationCPacked
:
1028 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
)
1029 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1030 spirv_decoration_to_string(dec
->decoration
));
1032 type
->packed
= true;
1035 case SpvDecorationSaturatedConversion
:
1036 case SpvDecorationFuncParamAttr
:
1037 case SpvDecorationFPRoundingMode
:
1038 case SpvDecorationFPFastMathMode
:
1039 case SpvDecorationAlignment
:
1040 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1041 spirv_decoration_to_string(dec
->decoration
));
1045 vtn_fail_with_decoration("Unhandled decoration", dec
->decoration
);
1050 translate_image_format(struct vtn_builder
*b
, SpvImageFormat format
)
1053 case SpvImageFormatUnknown
: return 0; /* GL_NONE */
1054 case SpvImageFormatRgba32f
: return 0x8814; /* GL_RGBA32F */
1055 case SpvImageFormatRgba16f
: return 0x881A; /* GL_RGBA16F */
1056 case SpvImageFormatR32f
: return 0x822E; /* GL_R32F */
1057 case SpvImageFormatRgba8
: return 0x8058; /* GL_RGBA8 */
1058 case SpvImageFormatRgba8Snorm
: return 0x8F97; /* GL_RGBA8_SNORM */
1059 case SpvImageFormatRg32f
: return 0x8230; /* GL_RG32F */
1060 case SpvImageFormatRg16f
: return 0x822F; /* GL_RG16F */
1061 case SpvImageFormatR11fG11fB10f
: return 0x8C3A; /* GL_R11F_G11F_B10F */
1062 case SpvImageFormatR16f
: return 0x822D; /* GL_R16F */
1063 case SpvImageFormatRgba16
: return 0x805B; /* GL_RGBA16 */
1064 case SpvImageFormatRgb10A2
: return 0x8059; /* GL_RGB10_A2 */
1065 case SpvImageFormatRg16
: return 0x822C; /* GL_RG16 */
1066 case SpvImageFormatRg8
: return 0x822B; /* GL_RG8 */
1067 case SpvImageFormatR16
: return 0x822A; /* GL_R16 */
1068 case SpvImageFormatR8
: return 0x8229; /* GL_R8 */
1069 case SpvImageFormatRgba16Snorm
: return 0x8F9B; /* GL_RGBA16_SNORM */
1070 case SpvImageFormatRg16Snorm
: return 0x8F99; /* GL_RG16_SNORM */
1071 case SpvImageFormatRg8Snorm
: return 0x8F95; /* GL_RG8_SNORM */
1072 case SpvImageFormatR16Snorm
: return 0x8F98; /* GL_R16_SNORM */
1073 case SpvImageFormatR8Snorm
: return 0x8F94; /* GL_R8_SNORM */
1074 case SpvImageFormatRgba32i
: return 0x8D82; /* GL_RGBA32I */
1075 case SpvImageFormatRgba16i
: return 0x8D88; /* GL_RGBA16I */
1076 case SpvImageFormatRgba8i
: return 0x8D8E; /* GL_RGBA8I */
1077 case SpvImageFormatR32i
: return 0x8235; /* GL_R32I */
1078 case SpvImageFormatRg32i
: return 0x823B; /* GL_RG32I */
1079 case SpvImageFormatRg16i
: return 0x8239; /* GL_RG16I */
1080 case SpvImageFormatRg8i
: return 0x8237; /* GL_RG8I */
1081 case SpvImageFormatR16i
: return 0x8233; /* GL_R16I */
1082 case SpvImageFormatR8i
: return 0x8231; /* GL_R8I */
1083 case SpvImageFormatRgba32ui
: return 0x8D70; /* GL_RGBA32UI */
1084 case SpvImageFormatRgba16ui
: return 0x8D76; /* GL_RGBA16UI */
1085 case SpvImageFormatRgba8ui
: return 0x8D7C; /* GL_RGBA8UI */
1086 case SpvImageFormatR32ui
: return 0x8236; /* GL_R32UI */
1087 case SpvImageFormatRgb10a2ui
: return 0x906F; /* GL_RGB10_A2UI */
1088 case SpvImageFormatRg32ui
: return 0x823C; /* GL_RG32UI */
1089 case SpvImageFormatRg16ui
: return 0x823A; /* GL_RG16UI */
1090 case SpvImageFormatRg8ui
: return 0x8238; /* GL_RG8UI */
1091 case SpvImageFormatR16ui
: return 0x8234; /* GL_R16UI */
1092 case SpvImageFormatR8ui
: return 0x8232; /* GL_R8UI */
1094 vtn_fail("Invalid image format: %s (%u)",
1095 spirv_imageformat_to_string(format
), format
);
1100 vtn_handle_type(struct vtn_builder
*b
, SpvOp opcode
,
1101 const uint32_t *w
, unsigned count
)
1103 struct vtn_value
*val
= NULL
;
1105 /* In order to properly handle forward declarations, we have to defer
1106 * allocation for pointer types.
1108 if (opcode
!= SpvOpTypePointer
&& opcode
!= SpvOpTypeForwardPointer
) {
1109 val
= vtn_push_value(b
, w
[1], vtn_value_type_type
);
1110 vtn_fail_if(val
->type
!= NULL
,
1111 "Only pointers can have forward declarations");
1112 val
->type
= rzalloc(b
, struct vtn_type
);
1113 val
->type
->id
= w
[1];
1118 val
->type
->base_type
= vtn_base_type_void
;
1119 val
->type
->type
= glsl_void_type();
1122 val
->type
->base_type
= vtn_base_type_scalar
;
1123 val
->type
->type
= glsl_bool_type();
1124 val
->type
->length
= 1;
1126 case SpvOpTypeInt
: {
1127 int bit_size
= w
[2];
1128 const bool signedness
= w
[3];
1129 val
->type
->base_type
= vtn_base_type_scalar
;
1132 val
->type
->type
= (signedness
? glsl_int64_t_type() : glsl_uint64_t_type());
1135 val
->type
->type
= (signedness
? glsl_int_type() : glsl_uint_type());
1138 val
->type
->type
= (signedness
? glsl_int16_t_type() : glsl_uint16_t_type());
1141 val
->type
->type
= (signedness
? glsl_int8_t_type() : glsl_uint8_t_type());
1144 vtn_fail("Invalid int bit size: %u", bit_size
);
1146 val
->type
->length
= 1;
1150 case SpvOpTypeFloat
: {
1151 int bit_size
= w
[2];
1152 val
->type
->base_type
= vtn_base_type_scalar
;
1155 val
->type
->type
= glsl_float16_t_type();
1158 val
->type
->type
= glsl_float_type();
1161 val
->type
->type
= glsl_double_type();
1164 vtn_fail("Invalid float bit size: %u", bit_size
);
1166 val
->type
->length
= 1;
1170 case SpvOpTypeVector
: {
1171 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1172 unsigned elems
= w
[3];
1174 vtn_fail_if(base
->base_type
!= vtn_base_type_scalar
,
1175 "Base type for OpTypeVector must be a scalar");
1176 vtn_fail_if((elems
< 2 || elems
> 4) && (elems
!= 8) && (elems
!= 16),
1177 "Invalid component count for OpTypeVector");
1179 val
->type
->base_type
= vtn_base_type_vector
;
1180 val
->type
->type
= glsl_vector_type(glsl_get_base_type(base
->type
), elems
);
1181 val
->type
->length
= elems
;
1182 val
->type
->stride
= glsl_type_is_boolean(val
->type
->type
)
1183 ? 4 : glsl_get_bit_size(base
->type
) / 8;
1184 val
->type
->array_element
= base
;
1188 case SpvOpTypeMatrix
: {
1189 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1190 unsigned columns
= w
[3];
1192 vtn_fail_if(base
->base_type
!= vtn_base_type_vector
,
1193 "Base type for OpTypeMatrix must be a vector");
1194 vtn_fail_if(columns
< 2 || columns
> 4,
1195 "Invalid column count for OpTypeMatrix");
1197 val
->type
->base_type
= vtn_base_type_matrix
;
1198 val
->type
->type
= glsl_matrix_type(glsl_get_base_type(base
->type
),
1199 glsl_get_vector_elements(base
->type
),
1201 vtn_fail_if(glsl_type_is_error(val
->type
->type
),
1202 "Unsupported base type for OpTypeMatrix");
1203 assert(!glsl_type_is_error(val
->type
->type
));
1204 val
->type
->length
= columns
;
1205 val
->type
->array_element
= base
;
1206 val
->type
->row_major
= false;
1207 val
->type
->stride
= 0;
1211 case SpvOpTypeRuntimeArray
:
1212 case SpvOpTypeArray
: {
1213 struct vtn_type
*array_element
=
1214 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1216 if (opcode
== SpvOpTypeRuntimeArray
) {
1217 /* A length of 0 is used to denote unsized arrays */
1218 val
->type
->length
= 0;
1220 val
->type
->length
= vtn_constant_uint(b
, w
[3]);
1223 val
->type
->base_type
= vtn_base_type_array
;
1224 val
->type
->array_element
= array_element
;
1225 if (b
->shader
->info
.stage
== MESA_SHADER_KERNEL
)
1226 val
->type
->stride
= glsl_get_cl_size(array_element
->type
);
1228 vtn_foreach_decoration(b
, val
, array_stride_decoration_cb
, NULL
);
1229 val
->type
->type
= glsl_array_type(array_element
->type
, val
->type
->length
,
1234 case SpvOpTypeStruct
: {
1235 unsigned num_fields
= count
- 2;
1236 val
->type
->base_type
= vtn_base_type_struct
;
1237 val
->type
->length
= num_fields
;
1238 val
->type
->members
= ralloc_array(b
, struct vtn_type
*, num_fields
);
1239 val
->type
->offsets
= ralloc_array(b
, unsigned, num_fields
);
1240 val
->type
->packed
= false;
1242 NIR_VLA(struct glsl_struct_field
, fields
, count
);
1243 for (unsigned i
= 0; i
< num_fields
; i
++) {
1244 val
->type
->members
[i
] =
1245 vtn_value(b
, w
[i
+ 2], vtn_value_type_type
)->type
;
1246 fields
[i
] = (struct glsl_struct_field
) {
1247 .type
= val
->type
->members
[i
]->type
,
1248 .name
= ralloc_asprintf(b
, "field%d", i
),
1254 if (b
->shader
->info
.stage
== MESA_SHADER_KERNEL
) {
1255 unsigned offset
= 0;
1256 for (unsigned i
= 0; i
< num_fields
; i
++) {
1257 offset
= align(offset
, glsl_get_cl_alignment(fields
[i
].type
));
1258 fields
[i
].offset
= offset
;
1259 offset
+= glsl_get_cl_size(fields
[i
].type
);
1263 struct member_decoration_ctx ctx
= {
1264 .num_fields
= num_fields
,
1269 vtn_foreach_decoration(b
, val
, struct_member_decoration_cb
, &ctx
);
1270 vtn_foreach_decoration(b
, val
, struct_member_matrix_stride_cb
, &ctx
);
1272 vtn_foreach_decoration(b
, val
, struct_block_decoration_cb
, NULL
);
1274 const char *name
= val
->name
;
1276 if (val
->type
->block
|| val
->type
->buffer_block
) {
1277 /* Packing will be ignored since types coming from SPIR-V are
1278 * explicitly laid out.
1280 val
->type
->type
= glsl_interface_type(fields
, num_fields
,
1281 /* packing */ 0, false,
1282 name
? name
: "block");
1284 val
->type
->type
= glsl_struct_type(fields
, num_fields
,
1285 name
? name
: "struct", false);
1290 case SpvOpTypeFunction
: {
1291 val
->type
->base_type
= vtn_base_type_function
;
1292 val
->type
->type
= NULL
;
1294 val
->type
->return_type
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1296 const unsigned num_params
= count
- 3;
1297 val
->type
->length
= num_params
;
1298 val
->type
->params
= ralloc_array(b
, struct vtn_type
*, num_params
);
1299 for (unsigned i
= 0; i
< count
- 3; i
++) {
1300 val
->type
->params
[i
] =
1301 vtn_value(b
, w
[i
+ 3], vtn_value_type_type
)->type
;
1306 case SpvOpTypePointer
:
1307 case SpvOpTypeForwardPointer
: {
1308 /* We can't blindly push the value because it might be a forward
1311 val
= vtn_untyped_value(b
, w
[1]);
1313 SpvStorageClass storage_class
= w
[2];
1315 if (val
->value_type
== vtn_value_type_invalid
) {
1316 val
->value_type
= vtn_value_type_type
;
1317 val
->type
= rzalloc(b
, struct vtn_type
);
1318 val
->type
->id
= w
[1];
1319 val
->type
->base_type
= vtn_base_type_pointer
;
1320 val
->type
->storage_class
= storage_class
;
1322 /* These can actually be stored to nir_variables and used as SSA
1323 * values so they need a real glsl_type.
1325 enum vtn_variable_mode mode
= vtn_storage_class_to_mode(
1326 b
, storage_class
, NULL
, NULL
);
1327 val
->type
->type
= nir_address_format_to_glsl_type(
1328 vtn_mode_to_address_format(b
, mode
));
1330 vtn_fail_if(val
->type
->storage_class
!= storage_class
,
1331 "The storage classes of an OpTypePointer and any "
1332 "OpTypeForwardPointers that provide forward "
1333 "declarations of it must match.");
1336 if (opcode
== SpvOpTypePointer
) {
1337 vtn_fail_if(val
->type
->deref
!= NULL
,
1338 "While OpTypeForwardPointer can be used to provide a "
1339 "forward declaration of a pointer, OpTypePointer can "
1340 "only be used once for a given id.");
1342 val
->type
->deref
= vtn_value(b
, w
[3], vtn_value_type_type
)->type
;
1344 /* Only certain storage classes use ArrayStride. The others (in
1345 * particular Workgroup) are expected to be laid out by the driver.
1347 switch (storage_class
) {
1348 case SpvStorageClassUniform
:
1349 case SpvStorageClassPushConstant
:
1350 case SpvStorageClassStorageBuffer
:
1351 case SpvStorageClassPhysicalStorageBufferEXT
:
1352 vtn_foreach_decoration(b
, val
, array_stride_decoration_cb
, NULL
);
1355 /* Nothing to do. */
1359 if (b
->physical_ptrs
) {
1360 switch (storage_class
) {
1361 case SpvStorageClassFunction
:
1362 case SpvStorageClassWorkgroup
:
1363 case SpvStorageClassCrossWorkgroup
:
1364 case SpvStorageClassUniformConstant
:
1365 val
->type
->stride
= align(glsl_get_cl_size(val
->type
->deref
->type
),
1366 glsl_get_cl_alignment(val
->type
->deref
->type
));
1376 case SpvOpTypeImage
: {
1377 val
->type
->base_type
= vtn_base_type_image
;
1379 const struct vtn_type
*sampled_type
=
1380 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1382 vtn_fail_if(sampled_type
->base_type
!= vtn_base_type_scalar
||
1383 glsl_get_bit_size(sampled_type
->type
) != 32,
1384 "Sampled type of OpTypeImage must be a 32-bit scalar");
1386 enum glsl_sampler_dim dim
;
1387 switch ((SpvDim
)w
[3]) {
1388 case SpvDim1D
: dim
= GLSL_SAMPLER_DIM_1D
; break;
1389 case SpvDim2D
: dim
= GLSL_SAMPLER_DIM_2D
; break;
1390 case SpvDim3D
: dim
= GLSL_SAMPLER_DIM_3D
; break;
1391 case SpvDimCube
: dim
= GLSL_SAMPLER_DIM_CUBE
; break;
1392 case SpvDimRect
: dim
= GLSL_SAMPLER_DIM_RECT
; break;
1393 case SpvDimBuffer
: dim
= GLSL_SAMPLER_DIM_BUF
; break;
1394 case SpvDimSubpassData
: dim
= GLSL_SAMPLER_DIM_SUBPASS
; break;
1396 vtn_fail("Invalid SPIR-V image dimensionality: %s (%u)",
1397 spirv_dim_to_string((SpvDim
)w
[3]), w
[3]);
1400 /* w[4]: as per Vulkan spec "Validation Rules within a Module",
1401 * The “Depth” operand of OpTypeImage is ignored.
1403 bool is_array
= w
[5];
1404 bool multisampled
= w
[6];
1405 unsigned sampled
= w
[7];
1406 SpvImageFormat format
= w
[8];
1409 val
->type
->access_qualifier
= w
[9];
1411 val
->type
->access_qualifier
= SpvAccessQualifierReadWrite
;
1414 if (dim
== GLSL_SAMPLER_DIM_2D
)
1415 dim
= GLSL_SAMPLER_DIM_MS
;
1416 else if (dim
== GLSL_SAMPLER_DIM_SUBPASS
)
1417 dim
= GLSL_SAMPLER_DIM_SUBPASS_MS
;
1419 vtn_fail("Unsupported multisampled image type");
1422 val
->type
->image_format
= translate_image_format(b
, format
);
1424 enum glsl_base_type sampled_base_type
=
1425 glsl_get_base_type(sampled_type
->type
);
1427 val
->type
->sampled
= true;
1428 val
->type
->type
= glsl_sampler_type(dim
, false, is_array
,
1430 } else if (sampled
== 2) {
1431 val
->type
->sampled
= false;
1432 val
->type
->type
= glsl_image_type(dim
, is_array
, sampled_base_type
);
1434 vtn_fail("We need to know if the image will be sampled");
1439 case SpvOpTypeSampledImage
:
1440 val
->type
->base_type
= vtn_base_type_sampled_image
;
1441 val
->type
->image
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1442 val
->type
->type
= val
->type
->image
->type
;
1445 case SpvOpTypeSampler
:
1446 /* The actual sampler type here doesn't really matter. It gets
1447 * thrown away the moment you combine it with an image. What really
1448 * matters is that it's a sampler type as opposed to an integer type
1449 * so the backend knows what to do.
1451 val
->type
->base_type
= vtn_base_type_sampler
;
1452 val
->type
->type
= glsl_bare_sampler_type();
1455 case SpvOpTypeOpaque
:
1456 case SpvOpTypeEvent
:
1457 case SpvOpTypeDeviceEvent
:
1458 case SpvOpTypeReserveId
:
1459 case SpvOpTypeQueue
:
1462 vtn_fail_with_opcode("Unhandled opcode", opcode
);
1465 vtn_foreach_decoration(b
, val
, type_decoration_cb
, NULL
);
1467 if (val
->type
->base_type
== vtn_base_type_struct
&&
1468 (val
->type
->block
|| val
->type
->buffer_block
)) {
1469 for (unsigned i
= 0; i
< val
->type
->length
; i
++) {
1470 vtn_fail_if(vtn_type_contains_block(b
, val
->type
->members
[i
]),
1471 "Block and BufferBlock decorations cannot decorate a "
1472 "structure type that is nested at any level inside "
1473 "another structure type decorated with Block or "
1479 static nir_constant
*
1480 vtn_null_constant(struct vtn_builder
*b
, struct vtn_type
*type
)
1482 nir_constant
*c
= rzalloc(b
, nir_constant
);
1484 switch (type
->base_type
) {
1485 case vtn_base_type_scalar
:
1486 case vtn_base_type_vector
:
1487 /* Nothing to do here. It's already initialized to zero */
1490 case vtn_base_type_pointer
: {
1491 enum vtn_variable_mode mode
= vtn_storage_class_to_mode(
1492 b
, type
->storage_class
, type
->deref
, NULL
);
1493 nir_address_format addr_format
= vtn_mode_to_address_format(b
, mode
);
1495 const nir_const_value
*null_value
= nir_address_format_null_value(addr_format
);
1496 memcpy(c
->values
, null_value
,
1497 sizeof(nir_const_value
) * nir_address_format_num_components(addr_format
));
1501 case vtn_base_type_void
:
1502 case vtn_base_type_image
:
1503 case vtn_base_type_sampler
:
1504 case vtn_base_type_sampled_image
:
1505 case vtn_base_type_function
:
1506 /* For those we have to return something but it doesn't matter what. */
1509 case vtn_base_type_matrix
:
1510 case vtn_base_type_array
:
1511 vtn_assert(type
->length
> 0);
1512 c
->num_elements
= type
->length
;
1513 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1515 c
->elements
[0] = vtn_null_constant(b
, type
->array_element
);
1516 for (unsigned i
= 1; i
< c
->num_elements
; i
++)
1517 c
->elements
[i
] = c
->elements
[0];
1520 case vtn_base_type_struct
:
1521 c
->num_elements
= type
->length
;
1522 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1523 for (unsigned i
= 0; i
< c
->num_elements
; i
++)
1524 c
->elements
[i
] = vtn_null_constant(b
, type
->members
[i
]);
1528 vtn_fail("Invalid type for null constant");
1535 spec_constant_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*v
,
1536 int member
, const struct vtn_decoration
*dec
,
1539 vtn_assert(member
== -1);
1540 if (dec
->decoration
!= SpvDecorationSpecId
)
1543 struct spec_constant_value
*const_value
= data
;
1545 for (unsigned i
= 0; i
< b
->num_specializations
; i
++) {
1546 if (b
->specializations
[i
].id
== dec
->operands
[0]) {
1547 if (const_value
->is_double
)
1548 const_value
->data64
= b
->specializations
[i
].data64
;
1550 const_value
->data32
= b
->specializations
[i
].data32
;
1557 get_specialization(struct vtn_builder
*b
, struct vtn_value
*val
,
1558 uint32_t const_value
)
1560 struct spec_constant_value data
;
1561 data
.is_double
= false;
1562 data
.data32
= const_value
;
1563 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1568 get_specialization64(struct vtn_builder
*b
, struct vtn_value
*val
,
1569 uint64_t const_value
)
1571 struct spec_constant_value data
;
1572 data
.is_double
= true;
1573 data
.data64
= const_value
;
1574 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1579 handle_workgroup_size_decoration_cb(struct vtn_builder
*b
,
1580 struct vtn_value
*val
,
1582 const struct vtn_decoration
*dec
,
1585 vtn_assert(member
== -1);
1586 if (dec
->decoration
!= SpvDecorationBuiltIn
||
1587 dec
->operands
[0] != SpvBuiltInWorkgroupSize
)
1590 vtn_assert(val
->type
->type
== glsl_vector_type(GLSL_TYPE_UINT
, 3));
1591 b
->workgroup_size_builtin
= val
;
1595 vtn_handle_constant(struct vtn_builder
*b
, SpvOp opcode
,
1596 const uint32_t *w
, unsigned count
)
1598 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_constant
);
1599 val
->constant
= rzalloc(b
, nir_constant
);
1601 case SpvOpConstantTrue
:
1602 case SpvOpConstantFalse
:
1603 case SpvOpSpecConstantTrue
:
1604 case SpvOpSpecConstantFalse
: {
1605 vtn_fail_if(val
->type
->type
!= glsl_bool_type(),
1606 "Result type of %s must be OpTypeBool",
1607 spirv_op_to_string(opcode
));
1609 uint32_t int_val
= (opcode
== SpvOpConstantTrue
||
1610 opcode
== SpvOpSpecConstantTrue
);
1612 if (opcode
== SpvOpSpecConstantTrue
||
1613 opcode
== SpvOpSpecConstantFalse
)
1614 int_val
= get_specialization(b
, val
, int_val
);
1616 val
->constant
->values
[0].b
= int_val
!= 0;
1620 case SpvOpConstant
: {
1621 vtn_fail_if(val
->type
->base_type
!= vtn_base_type_scalar
,
1622 "Result type of %s must be a scalar",
1623 spirv_op_to_string(opcode
));
1624 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1627 val
->constant
->values
[0].u64
= vtn_u64_literal(&w
[3]);
1630 val
->constant
->values
[0].u32
= w
[3];
1633 val
->constant
->values
[0].u16
= w
[3];
1636 val
->constant
->values
[0].u8
= w
[3];
1639 vtn_fail("Unsupported SpvOpConstant bit size: %u", bit_size
);
1644 case SpvOpSpecConstant
: {
1645 vtn_fail_if(val
->type
->base_type
!= vtn_base_type_scalar
,
1646 "Result type of %s must be a scalar",
1647 spirv_op_to_string(opcode
));
1648 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1651 val
->constant
->values
[0].u64
=
1652 get_specialization64(b
, val
, vtn_u64_literal(&w
[3]));
1655 val
->constant
->values
[0].u32
= get_specialization(b
, val
, w
[3]);
1658 val
->constant
->values
[0].u16
= get_specialization(b
, val
, w
[3]);
1661 val
->constant
->values
[0].u8
= get_specialization(b
, val
, w
[3]);
1664 vtn_fail("Unsupported SpvOpSpecConstant bit size");
1669 case SpvOpSpecConstantComposite
:
1670 case SpvOpConstantComposite
: {
1671 unsigned elem_count
= count
- 3;
1672 vtn_fail_if(elem_count
!= val
->type
->length
,
1673 "%s has %u constituents, expected %u",
1674 spirv_op_to_string(opcode
), elem_count
, val
->type
->length
);
1676 nir_constant
**elems
= ralloc_array(b
, nir_constant
*, elem_count
);
1677 for (unsigned i
= 0; i
< elem_count
; i
++) {
1678 struct vtn_value
*val
= vtn_untyped_value(b
, w
[i
+ 3]);
1680 if (val
->value_type
== vtn_value_type_constant
) {
1681 elems
[i
] = val
->constant
;
1683 vtn_fail_if(val
->value_type
!= vtn_value_type_undef
,
1684 "only constants or undefs allowed for "
1685 "SpvOpConstantComposite");
1686 /* to make it easier, just insert a NULL constant for now */
1687 elems
[i
] = vtn_null_constant(b
, val
->type
);
1691 switch (val
->type
->base_type
) {
1692 case vtn_base_type_vector
: {
1693 assert(glsl_type_is_vector(val
->type
->type
));
1694 for (unsigned i
= 0; i
< elem_count
; i
++)
1695 val
->constant
->values
[i
] = elems
[i
]->values
[0];
1699 case vtn_base_type_matrix
:
1700 case vtn_base_type_struct
:
1701 case vtn_base_type_array
:
1702 ralloc_steal(val
->constant
, elems
);
1703 val
->constant
->num_elements
= elem_count
;
1704 val
->constant
->elements
= elems
;
1708 vtn_fail("Result type of %s must be a composite type",
1709 spirv_op_to_string(opcode
));
1714 case SpvOpSpecConstantOp
: {
1715 SpvOp opcode
= get_specialization(b
, val
, w
[3]);
1717 case SpvOpVectorShuffle
: {
1718 struct vtn_value
*v0
= &b
->values
[w
[4]];
1719 struct vtn_value
*v1
= &b
->values
[w
[5]];
1721 vtn_assert(v0
->value_type
== vtn_value_type_constant
||
1722 v0
->value_type
== vtn_value_type_undef
);
1723 vtn_assert(v1
->value_type
== vtn_value_type_constant
||
1724 v1
->value_type
== vtn_value_type_undef
);
1726 unsigned len0
= glsl_get_vector_elements(v0
->type
->type
);
1727 unsigned len1
= glsl_get_vector_elements(v1
->type
->type
);
1729 vtn_assert(len0
+ len1
< 16);
1731 unsigned bit_size
= glsl_get_bit_size(val
->type
->type
);
1732 unsigned bit_size0
= glsl_get_bit_size(v0
->type
->type
);
1733 unsigned bit_size1
= glsl_get_bit_size(v1
->type
->type
);
1735 vtn_assert(bit_size
== bit_size0
&& bit_size
== bit_size1
);
1736 (void)bit_size0
; (void)bit_size1
;
1738 nir_const_value undef
= { .u64
= 0xdeadbeefdeadbeef };
1739 nir_const_value combined
[NIR_MAX_VEC_COMPONENTS
* 2];
1741 if (v0
->value_type
== vtn_value_type_constant
) {
1742 for (unsigned i
= 0; i
< len0
; i
++)
1743 combined
[i
] = v0
->constant
->values
[i
];
1745 if (v1
->value_type
== vtn_value_type_constant
) {
1746 for (unsigned i
= 0; i
< len1
; i
++)
1747 combined
[len0
+ i
] = v1
->constant
->values
[i
];
1750 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1751 uint32_t comp
= w
[i
+ 6];
1752 if (comp
== (uint32_t)-1) {
1753 /* If component is not used, set the value to a known constant
1754 * to detect if it is wrongly used.
1756 val
->constant
->values
[j
] = undef
;
1758 vtn_fail_if(comp
>= len0
+ len1
,
1759 "All Component literals must either be FFFFFFFF "
1760 "or in [0, N - 1] (inclusive).");
1761 val
->constant
->values
[j
] = combined
[comp
];
1767 case SpvOpCompositeExtract
:
1768 case SpvOpCompositeInsert
: {
1769 struct vtn_value
*comp
;
1770 unsigned deref_start
;
1771 struct nir_constant
**c
;
1772 if (opcode
== SpvOpCompositeExtract
) {
1773 comp
= vtn_value(b
, w
[4], vtn_value_type_constant
);
1775 c
= &comp
->constant
;
1777 comp
= vtn_value(b
, w
[5], vtn_value_type_constant
);
1779 val
->constant
= nir_constant_clone(comp
->constant
,
1785 const struct vtn_type
*type
= comp
->type
;
1786 for (unsigned i
= deref_start
; i
< count
; i
++) {
1787 vtn_fail_if(w
[i
] > type
->length
,
1788 "%uth index of %s is %u but the type has only "
1789 "%u elements", i
- deref_start
,
1790 spirv_op_to_string(opcode
), w
[i
], type
->length
);
1792 switch (type
->base_type
) {
1793 case vtn_base_type_vector
:
1795 type
= type
->array_element
;
1798 case vtn_base_type_matrix
:
1799 case vtn_base_type_array
:
1800 c
= &(*c
)->elements
[w
[i
]];
1801 type
= type
->array_element
;
1804 case vtn_base_type_struct
:
1805 c
= &(*c
)->elements
[w
[i
]];
1806 type
= type
->members
[w
[i
]];
1810 vtn_fail("%s must only index into composite types",
1811 spirv_op_to_string(opcode
));
1815 if (opcode
== SpvOpCompositeExtract
) {
1819 unsigned num_components
= type
->length
;
1820 for (unsigned i
= 0; i
< num_components
; i
++)
1821 val
->constant
->values
[i
] = (*c
)->values
[elem
+ i
];
1824 struct vtn_value
*insert
=
1825 vtn_value(b
, w
[4], vtn_value_type_constant
);
1826 vtn_assert(insert
->type
== type
);
1828 *c
= insert
->constant
;
1830 unsigned num_components
= type
->length
;
1831 for (unsigned i
= 0; i
< num_components
; i
++)
1832 (*c
)->values
[elem
+ i
] = insert
->constant
->values
[i
];
1840 nir_alu_type dst_alu_type
= nir_get_nir_type_for_glsl_type(val
->type
->type
);
1841 nir_alu_type src_alu_type
= dst_alu_type
;
1842 unsigned num_components
= glsl_get_vector_elements(val
->type
->type
);
1845 vtn_assert(count
<= 7);
1851 /* We have a source in a conversion */
1853 nir_get_nir_type_for_glsl_type(
1854 vtn_value(b
, w
[4], vtn_value_type_constant
)->type
->type
);
1855 /* We use the bitsize of the conversion source to evaluate the opcode later */
1856 bit_size
= glsl_get_bit_size(
1857 vtn_value(b
, w
[4], vtn_value_type_constant
)->type
->type
);
1860 bit_size
= glsl_get_bit_size(val
->type
->type
);
1863 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(b
, opcode
, &swap
,
1864 nir_alu_type_get_type_size(src_alu_type
),
1865 nir_alu_type_get_type_size(dst_alu_type
));
1866 nir_const_value src
[3][NIR_MAX_VEC_COMPONENTS
];
1868 for (unsigned i
= 0; i
< count
- 4; i
++) {
1869 struct vtn_value
*src_val
=
1870 vtn_value(b
, w
[4 + i
], vtn_value_type_constant
);
1872 /* If this is an unsized source, pull the bit size from the
1873 * source; otherwise, we'll use the bit size from the destination.
1875 if (!nir_alu_type_get_type_size(nir_op_infos
[op
].input_types
[i
]))
1876 bit_size
= glsl_get_bit_size(src_val
->type
->type
);
1878 unsigned src_comps
= nir_op_infos
[op
].input_sizes
[i
] ?
1879 nir_op_infos
[op
].input_sizes
[i
] :
1882 unsigned j
= swap
? 1 - i
: i
;
1883 for (unsigned c
= 0; c
< src_comps
; c
++)
1884 src
[j
][c
] = src_val
->constant
->values
[c
];
1887 /* fix up fixed size sources */
1894 for (unsigned i
= 0; i
< num_components
; ++i
) {
1896 case 64: src
[1][i
].u32
= src
[1][i
].u64
; break;
1897 case 16: src
[1][i
].u32
= src
[1][i
].u16
; break;
1898 case 8: src
[1][i
].u32
= src
[1][i
].u8
; break;
1907 nir_const_value
*srcs
[3] = {
1908 src
[0], src
[1], src
[2],
1910 nir_eval_const_opcode(op
, val
->constant
->values
,
1911 num_components
, bit_size
, srcs
,
1912 b
->shader
->info
.float_controls_execution_mode
);
1919 case SpvOpConstantNull
:
1920 val
->constant
= vtn_null_constant(b
, val
->type
);
1923 case SpvOpConstantSampler
:
1924 vtn_fail("OpConstantSampler requires Kernel Capability");
1928 vtn_fail_with_opcode("Unhandled opcode", opcode
);
1931 /* Now that we have the value, update the workgroup size if needed */
1932 vtn_foreach_decoration(b
, val
, handle_workgroup_size_decoration_cb
, NULL
);
1935 SpvMemorySemanticsMask
1936 vtn_storage_class_to_memory_semantics(SpvStorageClass sc
)
1939 case SpvStorageClassStorageBuffer
:
1940 case SpvStorageClassPhysicalStorageBufferEXT
:
1941 return SpvMemorySemanticsUniformMemoryMask
;
1942 case SpvStorageClassWorkgroup
:
1943 return SpvMemorySemanticsWorkgroupMemoryMask
;
1945 return SpvMemorySemanticsMaskNone
;
1950 vtn_split_barrier_semantics(struct vtn_builder
*b
,
1951 SpvMemorySemanticsMask semantics
,
1952 SpvMemorySemanticsMask
*before
,
1953 SpvMemorySemanticsMask
*after
)
1955 /* For memory semantics embedded in operations, we split them into up to
1956 * two barriers, to be added before and after the operation. This is less
1957 * strict than if we propagated until the final backend stage, but still
1958 * result in correct execution.
1960 * A further improvement could be pipe this information (and use!) into the
1961 * next compiler layers, at the expense of making the handling of barriers
1965 *before
= SpvMemorySemanticsMaskNone
;
1966 *after
= SpvMemorySemanticsMaskNone
;
1968 SpvMemorySemanticsMask order_semantics
=
1969 semantics
& (SpvMemorySemanticsAcquireMask
|
1970 SpvMemorySemanticsReleaseMask
|
1971 SpvMemorySemanticsAcquireReleaseMask
|
1972 SpvMemorySemanticsSequentiallyConsistentMask
);
1974 if (util_bitcount(order_semantics
) > 1) {
1975 /* Old GLSLang versions incorrectly set all the ordering bits. This was
1976 * fixed in c51287d744fb6e7e9ccc09f6f8451e6c64b1dad6 of glslang repo,
1977 * and it is in GLSLang since revision "SPIRV99.1321" (from Jul-2016).
1979 vtn_warn("Multiple memory ordering semantics specified, "
1980 "assuming AcquireRelease.");
1981 order_semantics
= SpvMemorySemanticsAcquireReleaseMask
;
1984 const SpvMemorySemanticsMask av_vis_semantics
=
1985 semantics
& (SpvMemorySemanticsMakeAvailableMask
|
1986 SpvMemorySemanticsMakeVisibleMask
);
1988 const SpvMemorySemanticsMask storage_semantics
=
1989 semantics
& (SpvMemorySemanticsUniformMemoryMask
|
1990 SpvMemorySemanticsSubgroupMemoryMask
|
1991 SpvMemorySemanticsWorkgroupMemoryMask
|
1992 SpvMemorySemanticsCrossWorkgroupMemoryMask
|
1993 SpvMemorySemanticsAtomicCounterMemoryMask
|
1994 SpvMemorySemanticsImageMemoryMask
|
1995 SpvMemorySemanticsOutputMemoryMask
);
1997 const SpvMemorySemanticsMask other_semantics
=
1998 semantics
& ~(order_semantics
| av_vis_semantics
| storage_semantics
);
2000 if (other_semantics
)
2001 vtn_warn("Ignoring unhandled memory semantics: %u\n", other_semantics
);
2003 /* SequentiallyConsistent is treated as AcquireRelease. */
2005 /* The RELEASE barrier happens BEFORE the operation, and it is usually
2006 * associated with a Store. All the write operations with a matching
2007 * semantics will not be reordered after the Store.
2009 if (order_semantics
& (SpvMemorySemanticsReleaseMask
|
2010 SpvMemorySemanticsAcquireReleaseMask
|
2011 SpvMemorySemanticsSequentiallyConsistentMask
)) {
2012 *before
|= SpvMemorySemanticsReleaseMask
| storage_semantics
;
2015 /* The ACQUIRE barrier happens AFTER the operation, and it is usually
2016 * associated with a Load. All the operations with a matching semantics
2017 * will not be reordered before the Load.
2019 if (order_semantics
& (SpvMemorySemanticsAcquireMask
|
2020 SpvMemorySemanticsAcquireReleaseMask
|
2021 SpvMemorySemanticsSequentiallyConsistentMask
)) {
2022 *after
|= SpvMemorySemanticsAcquireMask
| storage_semantics
;
2025 if (av_vis_semantics
& SpvMemorySemanticsMakeVisibleMask
)
2026 *before
|= SpvMemorySemanticsMakeVisibleMask
| storage_semantics
;
2028 if (av_vis_semantics
& SpvMemorySemanticsMakeAvailableMask
)
2029 *after
|= SpvMemorySemanticsMakeAvailableMask
| storage_semantics
;
2033 vtn_emit_scoped_memory_barrier(struct vtn_builder
*b
, SpvScope scope
,
2034 SpvMemorySemanticsMask semantics
)
2036 nir_memory_semantics nir_semantics
= 0;
2038 SpvMemorySemanticsMask order_semantics
=
2039 semantics
& (SpvMemorySemanticsAcquireMask
|
2040 SpvMemorySemanticsReleaseMask
|
2041 SpvMemorySemanticsAcquireReleaseMask
|
2042 SpvMemorySemanticsSequentiallyConsistentMask
);
2044 if (util_bitcount(order_semantics
) > 1) {
2045 /* Old GLSLang versions incorrectly set all the ordering bits. This was
2046 * fixed in c51287d744fb6e7e9ccc09f6f8451e6c64b1dad6 of glslang repo,
2047 * and it is in GLSLang since revision "SPIRV99.1321" (from Jul-2016).
2049 vtn_warn("Multiple memory ordering semantics bits specified, "
2050 "assuming AcquireRelease.");
2051 order_semantics
= SpvMemorySemanticsAcquireReleaseMask
;
2054 switch (order_semantics
) {
2056 /* Not an ordering barrier. */
2059 case SpvMemorySemanticsAcquireMask
:
2060 nir_semantics
= NIR_MEMORY_ACQUIRE
;
2063 case SpvMemorySemanticsReleaseMask
:
2064 nir_semantics
= NIR_MEMORY_RELEASE
;
2067 case SpvMemorySemanticsSequentiallyConsistentMask
:
2068 /* Fall through. Treated as AcquireRelease in Vulkan. */
2069 case SpvMemorySemanticsAcquireReleaseMask
:
2070 nir_semantics
= NIR_MEMORY_ACQUIRE
| NIR_MEMORY_RELEASE
;
2074 unreachable("Invalid memory order semantics");
2077 if (semantics
& SpvMemorySemanticsMakeAvailableMask
) {
2078 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
2079 "To use MakeAvailable memory semantics the VulkanMemoryModel "
2080 "capability must be declared.");
2081 nir_semantics
|= NIR_MEMORY_MAKE_AVAILABLE
;
2084 if (semantics
& SpvMemorySemanticsMakeVisibleMask
) {
2085 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
2086 "To use MakeVisible memory semantics the VulkanMemoryModel "
2087 "capability must be declared.");
2088 nir_semantics
|= NIR_MEMORY_MAKE_VISIBLE
;
2091 /* Vulkan Environment for SPIR-V says "SubgroupMemory, CrossWorkgroupMemory,
2092 * and AtomicCounterMemory are ignored".
2094 semantics
&= ~(SpvMemorySemanticsSubgroupMemoryMask
|
2095 SpvMemorySemanticsCrossWorkgroupMemoryMask
|
2096 SpvMemorySemanticsAtomicCounterMemoryMask
);
2098 /* TODO: Consider adding nir_var_mem_image mode to NIR so it can be used
2099 * for SpvMemorySemanticsImageMemoryMask.
2102 nir_variable_mode modes
= 0;
2103 if (semantics
& (SpvMemorySemanticsUniformMemoryMask
|
2104 SpvMemorySemanticsImageMemoryMask
))
2105 modes
|= nir_var_mem_ubo
| nir_var_mem_ssbo
| nir_var_uniform
;
2106 if (semantics
& SpvMemorySemanticsWorkgroupMemoryMask
)
2107 modes
|= nir_var_mem_shared
;
2108 if (semantics
& SpvMemorySemanticsOutputMemoryMask
) {
2109 modes
|= nir_var_shader_out
;
2112 /* No barrier to add. */
2113 if (nir_semantics
== 0 || modes
== 0)
2116 nir_scope nir_scope
;
2118 case SpvScopeDevice
:
2119 vtn_fail_if(b
->options
->caps
.vk_memory_model
&&
2120 !b
->options
->caps
.vk_memory_model_device_scope
,
2121 "If the Vulkan memory model is declared and any instruction "
2122 "uses Device scope, the VulkanMemoryModelDeviceScope "
2123 "capability must be declared.");
2124 nir_scope
= NIR_SCOPE_DEVICE
;
2127 case SpvScopeQueueFamily
:
2128 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
2129 "To use Queue Family scope, the VulkanMemoryModel capability "
2130 "must be declared.");
2131 nir_scope
= NIR_SCOPE_QUEUE_FAMILY
;
2134 case SpvScopeWorkgroup
:
2135 nir_scope
= NIR_SCOPE_WORKGROUP
;
2138 case SpvScopeSubgroup
:
2139 nir_scope
= NIR_SCOPE_SUBGROUP
;
2142 case SpvScopeInvocation
:
2143 nir_scope
= NIR_SCOPE_INVOCATION
;
2147 vtn_fail("Invalid memory scope");
2150 nir_intrinsic_instr
*intrin
=
2151 nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_scoped_memory_barrier
);
2152 nir_intrinsic_set_memory_semantics(intrin
, nir_semantics
);
2154 nir_intrinsic_set_memory_modes(intrin
, modes
);
2155 nir_intrinsic_set_memory_scope(intrin
, nir_scope
);
2156 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2159 struct vtn_ssa_value
*
2160 vtn_create_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
2162 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
2165 if (!glsl_type_is_vector_or_scalar(type
)) {
2166 unsigned elems
= glsl_get_length(type
);
2167 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
2168 for (unsigned i
= 0; i
< elems
; i
++) {
2169 const struct glsl_type
*child_type
;
2171 switch (glsl_get_base_type(type
)) {
2173 case GLSL_TYPE_UINT
:
2174 case GLSL_TYPE_INT16
:
2175 case GLSL_TYPE_UINT16
:
2176 case GLSL_TYPE_UINT8
:
2177 case GLSL_TYPE_INT8
:
2178 case GLSL_TYPE_INT64
:
2179 case GLSL_TYPE_UINT64
:
2180 case GLSL_TYPE_BOOL
:
2181 case GLSL_TYPE_FLOAT
:
2182 case GLSL_TYPE_FLOAT16
:
2183 case GLSL_TYPE_DOUBLE
:
2184 child_type
= glsl_get_column_type(type
);
2186 case GLSL_TYPE_ARRAY
:
2187 child_type
= glsl_get_array_element(type
);
2189 case GLSL_TYPE_STRUCT
:
2190 case GLSL_TYPE_INTERFACE
:
2191 child_type
= glsl_get_struct_field(type
, i
);
2194 vtn_fail("unkown base type");
2197 val
->elems
[i
] = vtn_create_ssa_value(b
, child_type
);
2205 vtn_tex_src(struct vtn_builder
*b
, unsigned index
, nir_tex_src_type type
)
2208 src
.src
= nir_src_for_ssa(vtn_ssa_value(b
, index
)->def
);
2209 src
.src_type
= type
;
2214 image_operand_arg(struct vtn_builder
*b
, const uint32_t *w
, uint32_t count
,
2215 uint32_t mask_idx
, SpvImageOperandsMask op
)
2217 static const SpvImageOperandsMask ops_with_arg
=
2218 SpvImageOperandsBiasMask
|
2219 SpvImageOperandsLodMask
|
2220 SpvImageOperandsGradMask
|
2221 SpvImageOperandsConstOffsetMask
|
2222 SpvImageOperandsOffsetMask
|
2223 SpvImageOperandsConstOffsetsMask
|
2224 SpvImageOperandsSampleMask
|
2225 SpvImageOperandsMinLodMask
|
2226 SpvImageOperandsMakeTexelAvailableMask
|
2227 SpvImageOperandsMakeTexelVisibleMask
;
2229 assert(util_bitcount(op
) == 1);
2230 assert(w
[mask_idx
] & op
);
2231 assert(op
& ops_with_arg
);
2233 uint32_t idx
= util_bitcount(w
[mask_idx
] & (op
- 1) & ops_with_arg
) + 1;
2235 /* Adjust indices for operands with two arguments. */
2236 static const SpvImageOperandsMask ops_with_two_args
=
2237 SpvImageOperandsGradMask
;
2238 idx
+= util_bitcount(w
[mask_idx
] & (op
- 1) & ops_with_two_args
);
2242 vtn_fail_if(idx
+ (op
& ops_with_two_args
? 1 : 0) >= count
,
2243 "Image op claims to have %s but does not enough "
2244 "following operands", spirv_imageoperands_to_string(op
));
2250 vtn_handle_texture(struct vtn_builder
*b
, SpvOp opcode
,
2251 const uint32_t *w
, unsigned count
)
2253 if (opcode
== SpvOpSampledImage
) {
2254 struct vtn_value
*val
=
2255 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
2256 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
2257 val
->sampled_image
->image
=
2258 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2259 val
->sampled_image
->sampler
=
2260 vtn_value(b
, w
[4], vtn_value_type_pointer
)->pointer
;
2262 } else if (opcode
== SpvOpImage
) {
2263 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[3]);
2264 if (src_val
->value_type
== vtn_value_type_sampled_image
) {
2265 vtn_push_value_pointer(b
, w
[2], src_val
->sampled_image
->image
);
2267 vtn_assert(src_val
->value_type
== vtn_value_type_pointer
);
2268 vtn_push_value_pointer(b
, w
[2], src_val
->pointer
);
2273 struct vtn_type
*ret_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2275 struct vtn_pointer
*image
= NULL
, *sampler
= NULL
;
2276 struct vtn_value
*sampled_val
= vtn_untyped_value(b
, w
[3]);
2277 if (sampled_val
->value_type
== vtn_value_type_sampled_image
) {
2278 image
= sampled_val
->sampled_image
->image
;
2279 sampler
= sampled_val
->sampled_image
->sampler
;
2281 vtn_assert(sampled_val
->value_type
== vtn_value_type_pointer
);
2282 image
= sampled_val
->pointer
;
2285 nir_deref_instr
*image_deref
= vtn_pointer_to_deref(b
, image
);
2286 nir_deref_instr
*sampler_deref
=
2287 sampler
? vtn_pointer_to_deref(b
, sampler
) : NULL
;
2289 const struct glsl_type
*image_type
= sampled_val
->type
->type
;
2290 const enum glsl_sampler_dim sampler_dim
= glsl_get_sampler_dim(image_type
);
2291 const bool is_array
= glsl_sampler_type_is_array(image_type
);
2292 nir_alu_type dest_type
= nir_type_invalid
;
2294 /* Figure out the base texture operation */
2297 case SpvOpImageSampleImplicitLod
:
2298 case SpvOpImageSampleDrefImplicitLod
:
2299 case SpvOpImageSampleProjImplicitLod
:
2300 case SpvOpImageSampleProjDrefImplicitLod
:
2301 texop
= nir_texop_tex
;
2304 case SpvOpImageSampleExplicitLod
:
2305 case SpvOpImageSampleDrefExplicitLod
:
2306 case SpvOpImageSampleProjExplicitLod
:
2307 case SpvOpImageSampleProjDrefExplicitLod
:
2308 texop
= nir_texop_txl
;
2311 case SpvOpImageFetch
:
2312 if (sampler_dim
== GLSL_SAMPLER_DIM_MS
) {
2313 texop
= nir_texop_txf_ms
;
2315 texop
= nir_texop_txf
;
2319 case SpvOpImageGather
:
2320 case SpvOpImageDrefGather
:
2321 texop
= nir_texop_tg4
;
2324 case SpvOpImageQuerySizeLod
:
2325 case SpvOpImageQuerySize
:
2326 texop
= nir_texop_txs
;
2327 dest_type
= nir_type_int
;
2330 case SpvOpImageQueryLod
:
2331 texop
= nir_texop_lod
;
2332 dest_type
= nir_type_float
;
2335 case SpvOpImageQueryLevels
:
2336 texop
= nir_texop_query_levels
;
2337 dest_type
= nir_type_int
;
2340 case SpvOpImageQuerySamples
:
2341 texop
= nir_texop_texture_samples
;
2342 dest_type
= nir_type_int
;
2346 vtn_fail_with_opcode("Unhandled opcode", opcode
);
2349 nir_tex_src srcs
[10]; /* 10 should be enough */
2350 nir_tex_src
*p
= srcs
;
2352 p
->src
= nir_src_for_ssa(&image_deref
->dest
.ssa
);
2353 p
->src_type
= nir_tex_src_texture_deref
;
2363 vtn_fail_if(sampler
== NULL
,
2364 "%s requires an image of type OpTypeSampledImage",
2365 spirv_op_to_string(opcode
));
2366 p
->src
= nir_src_for_ssa(&sampler_deref
->dest
.ssa
);
2367 p
->src_type
= nir_tex_src_sampler_deref
;
2371 case nir_texop_txf_ms
:
2373 case nir_texop_query_levels
:
2374 case nir_texop_texture_samples
:
2375 case nir_texop_samples_identical
:
2378 case nir_texop_txf_ms_fb
:
2379 vtn_fail("unexpected nir_texop_txf_ms_fb");
2381 case nir_texop_txf_ms_mcs
:
2382 vtn_fail("unexpected nir_texop_txf_ms_mcs");
2383 case nir_texop_tex_prefetch
:
2384 vtn_fail("unexpected nir_texop_tex_prefetch");
2389 struct nir_ssa_def
*coord
;
2390 unsigned coord_components
;
2392 case SpvOpImageSampleImplicitLod
:
2393 case SpvOpImageSampleExplicitLod
:
2394 case SpvOpImageSampleDrefImplicitLod
:
2395 case SpvOpImageSampleDrefExplicitLod
:
2396 case SpvOpImageSampleProjImplicitLod
:
2397 case SpvOpImageSampleProjExplicitLod
:
2398 case SpvOpImageSampleProjDrefImplicitLod
:
2399 case SpvOpImageSampleProjDrefExplicitLod
:
2400 case SpvOpImageFetch
:
2401 case SpvOpImageGather
:
2402 case SpvOpImageDrefGather
:
2403 case SpvOpImageQueryLod
: {
2404 /* All these types have the coordinate as their first real argument */
2405 switch (sampler_dim
) {
2406 case GLSL_SAMPLER_DIM_1D
:
2407 case GLSL_SAMPLER_DIM_BUF
:
2408 coord_components
= 1;
2410 case GLSL_SAMPLER_DIM_2D
:
2411 case GLSL_SAMPLER_DIM_RECT
:
2412 case GLSL_SAMPLER_DIM_MS
:
2413 coord_components
= 2;
2415 case GLSL_SAMPLER_DIM_3D
:
2416 case GLSL_SAMPLER_DIM_CUBE
:
2417 coord_components
= 3;
2420 vtn_fail("Invalid sampler type");
2423 if (is_array
&& texop
!= nir_texop_lod
)
2426 coord
= vtn_ssa_value(b
, w
[idx
++])->def
;
2427 p
->src
= nir_src_for_ssa(nir_channels(&b
->nb
, coord
,
2428 (1 << coord_components
) - 1));
2429 p
->src_type
= nir_tex_src_coord
;
2436 coord_components
= 0;
2441 case SpvOpImageSampleProjImplicitLod
:
2442 case SpvOpImageSampleProjExplicitLod
:
2443 case SpvOpImageSampleProjDrefImplicitLod
:
2444 case SpvOpImageSampleProjDrefExplicitLod
:
2445 /* These have the projector as the last coordinate component */
2446 p
->src
= nir_src_for_ssa(nir_channel(&b
->nb
, coord
, coord_components
));
2447 p
->src_type
= nir_tex_src_projector
;
2455 bool is_shadow
= false;
2456 unsigned gather_component
= 0;
2458 case SpvOpImageSampleDrefImplicitLod
:
2459 case SpvOpImageSampleDrefExplicitLod
:
2460 case SpvOpImageSampleProjDrefImplicitLod
:
2461 case SpvOpImageSampleProjDrefExplicitLod
:
2462 case SpvOpImageDrefGather
:
2463 /* These all have an explicit depth value as their next source */
2465 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_comparator
);
2468 case SpvOpImageGather
:
2469 /* This has a component as its next source */
2470 gather_component
= vtn_constant_uint(b
, w
[idx
++]);
2477 /* For OpImageQuerySizeLod, we always have an LOD */
2478 if (opcode
== SpvOpImageQuerySizeLod
)
2479 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
2481 /* Now we need to handle some number of optional arguments */
2482 struct vtn_value
*gather_offsets
= NULL
;
2484 uint32_t operands
= w
[idx
];
2486 if (operands
& SpvImageOperandsBiasMask
) {
2487 vtn_assert(texop
== nir_texop_tex
);
2488 texop
= nir_texop_txb
;
2489 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2490 SpvImageOperandsBiasMask
);
2491 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_bias
);
2494 if (operands
& SpvImageOperandsLodMask
) {
2495 vtn_assert(texop
== nir_texop_txl
|| texop
== nir_texop_txf
||
2496 texop
== nir_texop_txs
);
2497 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2498 SpvImageOperandsLodMask
);
2499 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_lod
);
2502 if (operands
& SpvImageOperandsGradMask
) {
2503 vtn_assert(texop
== nir_texop_txl
);
2504 texop
= nir_texop_txd
;
2505 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2506 SpvImageOperandsGradMask
);
2507 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_ddx
);
2508 (*p
++) = vtn_tex_src(b
, w
[arg
+ 1], nir_tex_src_ddy
);
2511 vtn_fail_if(util_bitcount(operands
& (SpvImageOperandsConstOffsetsMask
|
2512 SpvImageOperandsOffsetMask
|
2513 SpvImageOperandsConstOffsetMask
)) > 1,
2514 "At most one of the ConstOffset, Offset, and ConstOffsets "
2515 "image operands can be used on a given instruction.");
2517 if (operands
& SpvImageOperandsOffsetMask
) {
2518 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2519 SpvImageOperandsOffsetMask
);
2520 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_offset
);
2523 if (operands
& SpvImageOperandsConstOffsetMask
) {
2524 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2525 SpvImageOperandsConstOffsetMask
);
2526 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_offset
);
2529 if (operands
& SpvImageOperandsConstOffsetsMask
) {
2530 vtn_assert(texop
== nir_texop_tg4
);
2531 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2532 SpvImageOperandsConstOffsetsMask
);
2533 gather_offsets
= vtn_value(b
, w
[arg
], vtn_value_type_constant
);
2536 if (operands
& SpvImageOperandsSampleMask
) {
2537 vtn_assert(texop
== nir_texop_txf_ms
);
2538 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2539 SpvImageOperandsSampleMask
);
2540 texop
= nir_texop_txf_ms
;
2541 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_ms_index
);
2544 if (operands
& SpvImageOperandsMinLodMask
) {
2545 vtn_assert(texop
== nir_texop_tex
||
2546 texop
== nir_texop_txb
||
2547 texop
== nir_texop_txd
);
2548 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2549 SpvImageOperandsMinLodMask
);
2550 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_min_lod
);
2554 nir_tex_instr
*instr
= nir_tex_instr_create(b
->shader
, p
- srcs
);
2557 memcpy(instr
->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
2559 instr
->coord_components
= coord_components
;
2560 instr
->sampler_dim
= sampler_dim
;
2561 instr
->is_array
= is_array
;
2562 instr
->is_shadow
= is_shadow
;
2563 instr
->is_new_style_shadow
=
2564 is_shadow
&& glsl_get_components(ret_type
->type
) == 1;
2565 instr
->component
= gather_component
;
2567 if (image
&& (image
->access
& ACCESS_NON_UNIFORM
))
2568 instr
->texture_non_uniform
= true;
2570 if (sampler
&& (sampler
->access
& ACCESS_NON_UNIFORM
))
2571 instr
->sampler_non_uniform
= true;
2573 /* for non-query ops, get dest_type from sampler type */
2574 if (dest_type
== nir_type_invalid
) {
2575 switch (glsl_get_sampler_result_type(image_type
)) {
2576 case GLSL_TYPE_FLOAT
: dest_type
= nir_type_float
; break;
2577 case GLSL_TYPE_INT
: dest_type
= nir_type_int
; break;
2578 case GLSL_TYPE_UINT
: dest_type
= nir_type_uint
; break;
2579 case GLSL_TYPE_BOOL
: dest_type
= nir_type_bool
; break;
2581 vtn_fail("Invalid base type for sampler result");
2585 instr
->dest_type
= dest_type
;
2587 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
2588 nir_tex_instr_dest_size(instr
), 32, NULL
);
2590 vtn_assert(glsl_get_vector_elements(ret_type
->type
) ==
2591 nir_tex_instr_dest_size(instr
));
2593 if (gather_offsets
) {
2594 vtn_fail_if(gather_offsets
->type
->base_type
!= vtn_base_type_array
||
2595 gather_offsets
->type
->length
!= 4,
2596 "ConstOffsets must be an array of size four of vectors "
2597 "of two integer components");
2599 struct vtn_type
*vec_type
= gather_offsets
->type
->array_element
;
2600 vtn_fail_if(vec_type
->base_type
!= vtn_base_type_vector
||
2601 vec_type
->length
!= 2 ||
2602 !glsl_type_is_integer(vec_type
->type
),
2603 "ConstOffsets must be an array of size four of vectors "
2604 "of two integer components");
2606 unsigned bit_size
= glsl_get_bit_size(vec_type
->type
);
2607 for (uint32_t i
= 0; i
< 4; i
++) {
2608 const nir_const_value
*cvec
=
2609 gather_offsets
->constant
->elements
[i
]->values
;
2610 for (uint32_t j
= 0; j
< 2; j
++) {
2612 case 8: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i8
; break;
2613 case 16: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i16
; break;
2614 case 32: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i32
; break;
2615 case 64: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i64
; break;
2617 vtn_fail("Unsupported bit size: %u", bit_size
);
2623 struct vtn_ssa_value
*ssa
= vtn_create_ssa_value(b
, ret_type
->type
);
2624 ssa
->def
= &instr
->dest
.ssa
;
2625 vtn_push_ssa(b
, w
[2], ret_type
, ssa
);
2627 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
2631 fill_common_atomic_sources(struct vtn_builder
*b
, SpvOp opcode
,
2632 const uint32_t *w
, nir_src
*src
)
2635 case SpvOpAtomicIIncrement
:
2636 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
2639 case SpvOpAtomicIDecrement
:
2640 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
2643 case SpvOpAtomicISub
:
2645 nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
2648 case SpvOpAtomicCompareExchange
:
2649 case SpvOpAtomicCompareExchangeWeak
:
2650 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[8])->def
);
2651 src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
2654 case SpvOpAtomicExchange
:
2655 case SpvOpAtomicIAdd
:
2656 case SpvOpAtomicSMin
:
2657 case SpvOpAtomicUMin
:
2658 case SpvOpAtomicSMax
:
2659 case SpvOpAtomicUMax
:
2660 case SpvOpAtomicAnd
:
2662 case SpvOpAtomicXor
:
2663 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
2667 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
2671 static nir_ssa_def
*
2672 get_image_coord(struct vtn_builder
*b
, uint32_t value
)
2674 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, value
);
2676 /* The image_load_store intrinsics assume a 4-dim coordinate */
2677 unsigned dim
= glsl_get_vector_elements(coord
->type
);
2678 unsigned swizzle
[4];
2679 for (unsigned i
= 0; i
< 4; i
++)
2680 swizzle
[i
] = MIN2(i
, dim
- 1);
2682 return nir_swizzle(&b
->nb
, coord
->def
, swizzle
, 4);
2685 static nir_ssa_def
*
2686 expand_to_vec4(nir_builder
*b
, nir_ssa_def
*value
)
2688 if (value
->num_components
== 4)
2692 for (unsigned i
= 0; i
< 4; i
++)
2693 swiz
[i
] = i
< value
->num_components
? i
: 0;
2694 return nir_swizzle(b
, value
, swiz
, 4);
2698 vtn_handle_image(struct vtn_builder
*b
, SpvOp opcode
,
2699 const uint32_t *w
, unsigned count
)
2701 /* Just get this one out of the way */
2702 if (opcode
== SpvOpImageTexelPointer
) {
2703 struct vtn_value
*val
=
2704 vtn_push_value(b
, w
[2], vtn_value_type_image_pointer
);
2705 val
->image
= ralloc(b
, struct vtn_image_pointer
);
2707 val
->image
->image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2708 val
->image
->coord
= get_image_coord(b
, w
[4]);
2709 val
->image
->sample
= vtn_ssa_value(b
, w
[5])->def
;
2710 val
->image
->lod
= nir_imm_int(&b
->nb
, 0);
2714 struct vtn_image_pointer image
;
2715 SpvScope scope
= SpvScopeInvocation
;
2716 SpvMemorySemanticsMask semantics
= 0;
2719 case SpvOpAtomicExchange
:
2720 case SpvOpAtomicCompareExchange
:
2721 case SpvOpAtomicCompareExchangeWeak
:
2722 case SpvOpAtomicIIncrement
:
2723 case SpvOpAtomicIDecrement
:
2724 case SpvOpAtomicIAdd
:
2725 case SpvOpAtomicISub
:
2726 case SpvOpAtomicLoad
:
2727 case SpvOpAtomicSMin
:
2728 case SpvOpAtomicUMin
:
2729 case SpvOpAtomicSMax
:
2730 case SpvOpAtomicUMax
:
2731 case SpvOpAtomicAnd
:
2733 case SpvOpAtomicXor
:
2734 image
= *vtn_value(b
, w
[3], vtn_value_type_image_pointer
)->image
;
2735 scope
= vtn_constant_uint(b
, w
[4]);
2736 semantics
= vtn_constant_uint(b
, w
[5]);
2739 case SpvOpAtomicStore
:
2740 image
= *vtn_value(b
, w
[1], vtn_value_type_image_pointer
)->image
;
2741 scope
= vtn_constant_uint(b
, w
[2]);
2742 semantics
= vtn_constant_uint(b
, w
[3]);
2745 case SpvOpImageQuerySize
:
2746 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2748 image
.sample
= NULL
;
2752 case SpvOpImageRead
: {
2753 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2754 image
.coord
= get_image_coord(b
, w
[4]);
2756 const SpvImageOperandsMask operands
=
2757 count
> 5 ? w
[5] : SpvImageOperandsMaskNone
;
2759 if (operands
& SpvImageOperandsSampleMask
) {
2760 uint32_t arg
= image_operand_arg(b
, w
, count
, 5,
2761 SpvImageOperandsSampleMask
);
2762 image
.sample
= vtn_ssa_value(b
, w
[arg
])->def
;
2764 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2767 if (operands
& SpvImageOperandsMakeTexelVisibleMask
) {
2768 vtn_fail_if((operands
& SpvImageOperandsNonPrivateTexelMask
) == 0,
2769 "MakeTexelVisible requires NonPrivateTexel to also be set.");
2770 uint32_t arg
= image_operand_arg(b
, w
, count
, 5,
2771 SpvImageOperandsMakeTexelVisibleMask
);
2772 semantics
= SpvMemorySemanticsMakeVisibleMask
;
2773 scope
= vtn_constant_uint(b
, w
[arg
]);
2776 if (operands
& SpvImageOperandsLodMask
) {
2777 uint32_t arg
= image_operand_arg(b
, w
, count
, 5,
2778 SpvImageOperandsLodMask
);
2779 image
.lod
= vtn_ssa_value(b
, w
[arg
])->def
;
2781 image
.lod
= nir_imm_int(&b
->nb
, 0);
2784 /* TODO: Volatile. */
2789 case SpvOpImageWrite
: {
2790 image
.image
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
2791 image
.coord
= get_image_coord(b
, w
[2]);
2795 const SpvImageOperandsMask operands
=
2796 count
> 4 ? w
[4] : SpvImageOperandsMaskNone
;
2798 if (operands
& SpvImageOperandsSampleMask
) {
2799 uint32_t arg
= image_operand_arg(b
, w
, count
, 4,
2800 SpvImageOperandsSampleMask
);
2801 image
.sample
= vtn_ssa_value(b
, w
[arg
])->def
;
2803 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2806 if (operands
& SpvImageOperandsMakeTexelAvailableMask
) {
2807 vtn_fail_if((operands
& SpvImageOperandsNonPrivateTexelMask
) == 0,
2808 "MakeTexelAvailable requires NonPrivateTexel to also be set.");
2809 uint32_t arg
= image_operand_arg(b
, w
, count
, 4,
2810 SpvImageOperandsMakeTexelAvailableMask
);
2811 semantics
= SpvMemorySemanticsMakeAvailableMask
;
2812 scope
= vtn_constant_uint(b
, w
[arg
]);
2815 if (operands
& SpvImageOperandsLodMask
) {
2816 uint32_t arg
= image_operand_arg(b
, w
, count
, 4,
2817 SpvImageOperandsLodMask
);
2818 image
.lod
= vtn_ssa_value(b
, w
[arg
])->def
;
2820 image
.lod
= nir_imm_int(&b
->nb
, 0);
2823 /* TODO: Volatile. */
2829 vtn_fail_with_opcode("Invalid image opcode", opcode
);
2832 nir_intrinsic_op op
;
2834 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_deref_##N; break;
2835 OP(ImageQuerySize
, size
)
2837 OP(ImageWrite
, store
)
2838 OP(AtomicLoad
, load
)
2839 OP(AtomicStore
, store
)
2840 OP(AtomicExchange
, atomic_exchange
)
2841 OP(AtomicCompareExchange
, atomic_comp_swap
)
2842 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
2843 OP(AtomicIIncrement
, atomic_add
)
2844 OP(AtomicIDecrement
, atomic_add
)
2845 OP(AtomicIAdd
, atomic_add
)
2846 OP(AtomicISub
, atomic_add
)
2847 OP(AtomicSMin
, atomic_imin
)
2848 OP(AtomicUMin
, atomic_umin
)
2849 OP(AtomicSMax
, atomic_imax
)
2850 OP(AtomicUMax
, atomic_umax
)
2851 OP(AtomicAnd
, atomic_and
)
2852 OP(AtomicOr
, atomic_or
)
2853 OP(AtomicXor
, atomic_xor
)
2856 vtn_fail_with_opcode("Invalid image opcode", opcode
);
2859 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
2861 nir_deref_instr
*image_deref
= vtn_pointer_to_deref(b
, image
.image
);
2862 intrin
->src
[0] = nir_src_for_ssa(&image_deref
->dest
.ssa
);
2864 /* ImageQuerySize doesn't take any extra parameters */
2865 if (opcode
!= SpvOpImageQuerySize
) {
2866 /* The image coordinate is always 4 components but we may not have that
2867 * many. Swizzle to compensate.
2869 intrin
->src
[1] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, image
.coord
));
2870 intrin
->src
[2] = nir_src_for_ssa(image
.sample
);
2873 nir_intrinsic_set_access(intrin
, image
.image
->access
);
2876 case SpvOpAtomicLoad
:
2877 case SpvOpImageQuerySize
:
2878 case SpvOpImageRead
:
2879 if (opcode
== SpvOpImageRead
|| opcode
== SpvOpAtomicLoad
) {
2880 /* Only OpImageRead can support a lod parameter if
2881 * SPV_AMD_shader_image_load_store_lod is used but the current NIR
2882 * intrinsics definition for atomics requires us to set it for
2885 intrin
->src
[3] = nir_src_for_ssa(image
.lod
);
2888 case SpvOpAtomicStore
:
2889 case SpvOpImageWrite
: {
2890 const uint32_t value_id
= opcode
== SpvOpAtomicStore
? w
[4] : w
[3];
2891 nir_ssa_def
*value
= vtn_ssa_value(b
, value_id
)->def
;
2892 /* nir_intrinsic_image_deref_store always takes a vec4 value */
2893 assert(op
== nir_intrinsic_image_deref_store
);
2894 intrin
->num_components
= 4;
2895 intrin
->src
[3] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, value
));
2896 /* Only OpImageWrite can support a lod parameter if
2897 * SPV_AMD_shader_image_load_store_lod is used but the current NIR
2898 * intrinsics definition for atomics requires us to set it for
2901 intrin
->src
[4] = nir_src_for_ssa(image
.lod
);
2905 case SpvOpAtomicCompareExchange
:
2906 case SpvOpAtomicCompareExchangeWeak
:
2907 case SpvOpAtomicIIncrement
:
2908 case SpvOpAtomicIDecrement
:
2909 case SpvOpAtomicExchange
:
2910 case SpvOpAtomicIAdd
:
2911 case SpvOpAtomicISub
:
2912 case SpvOpAtomicSMin
:
2913 case SpvOpAtomicUMin
:
2914 case SpvOpAtomicSMax
:
2915 case SpvOpAtomicUMax
:
2916 case SpvOpAtomicAnd
:
2918 case SpvOpAtomicXor
:
2919 fill_common_atomic_sources(b
, opcode
, w
, &intrin
->src
[3]);
2923 vtn_fail_with_opcode("Invalid image opcode", opcode
);
2926 /* Image operations implicitly have the Image storage memory semantics. */
2927 semantics
|= SpvMemorySemanticsImageMemoryMask
;
2929 SpvMemorySemanticsMask before_semantics
;
2930 SpvMemorySemanticsMask after_semantics
;
2931 vtn_split_barrier_semantics(b
, semantics
, &before_semantics
, &after_semantics
);
2933 if (before_semantics
)
2934 vtn_emit_memory_barrier(b
, scope
, before_semantics
);
2936 if (opcode
!= SpvOpImageWrite
&& opcode
!= SpvOpAtomicStore
) {
2937 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2939 unsigned dest_components
= glsl_get_vector_elements(type
->type
);
2940 intrin
->num_components
= nir_intrinsic_infos
[op
].dest_components
;
2941 if (intrin
->num_components
== 0)
2942 intrin
->num_components
= dest_components
;
2944 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
,
2945 intrin
->num_components
, 32, NULL
);
2947 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2949 nir_ssa_def
*result
= &intrin
->dest
.ssa
;
2950 if (intrin
->num_components
!= dest_components
)
2951 result
= nir_channels(&b
->nb
, result
, (1 << dest_components
) - 1);
2953 struct vtn_value
*val
=
2954 vtn_push_ssa(b
, w
[2], type
, vtn_create_ssa_value(b
, type
->type
));
2955 val
->ssa
->def
= result
;
2957 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2960 if (after_semantics
)
2961 vtn_emit_memory_barrier(b
, scope
, after_semantics
);
2964 static nir_intrinsic_op
2965 get_ssbo_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2968 case SpvOpAtomicLoad
: return nir_intrinsic_load_ssbo
;
2969 case SpvOpAtomicStore
: return nir_intrinsic_store_ssbo
;
2970 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
2971 OP(AtomicExchange
, atomic_exchange
)
2972 OP(AtomicCompareExchange
, atomic_comp_swap
)
2973 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
2974 OP(AtomicIIncrement
, atomic_add
)
2975 OP(AtomicIDecrement
, atomic_add
)
2976 OP(AtomicIAdd
, atomic_add
)
2977 OP(AtomicISub
, atomic_add
)
2978 OP(AtomicSMin
, atomic_imin
)
2979 OP(AtomicUMin
, atomic_umin
)
2980 OP(AtomicSMax
, atomic_imax
)
2981 OP(AtomicUMax
, atomic_umax
)
2982 OP(AtomicAnd
, atomic_and
)
2983 OP(AtomicOr
, atomic_or
)
2984 OP(AtomicXor
, atomic_xor
)
2987 vtn_fail_with_opcode("Invalid SSBO atomic", opcode
);
2991 static nir_intrinsic_op
2992 get_uniform_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2995 #define OP(S, N) case SpvOp##S: return nir_intrinsic_atomic_counter_ ##N;
2996 OP(AtomicLoad
, read_deref
)
2997 OP(AtomicExchange
, exchange
)
2998 OP(AtomicCompareExchange
, comp_swap
)
2999 OP(AtomicCompareExchangeWeak
, comp_swap
)
3000 OP(AtomicIIncrement
, inc_deref
)
3001 OP(AtomicIDecrement
, post_dec_deref
)
3002 OP(AtomicIAdd
, add_deref
)
3003 OP(AtomicISub
, add_deref
)
3004 OP(AtomicUMin
, min_deref
)
3005 OP(AtomicUMax
, max_deref
)
3006 OP(AtomicAnd
, and_deref
)
3007 OP(AtomicOr
, or_deref
)
3008 OP(AtomicXor
, xor_deref
)
3011 /* We left the following out: AtomicStore, AtomicSMin and
3012 * AtomicSmax. Right now there are not nir intrinsics for them. At this
3013 * moment Atomic Counter support is needed for ARB_spirv support, so is
3014 * only need to support GLSL Atomic Counters that are uints and don't
3015 * allow direct storage.
3017 unreachable("Invalid uniform atomic");
3021 static nir_intrinsic_op
3022 get_deref_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
3025 case SpvOpAtomicLoad
: return nir_intrinsic_load_deref
;
3026 case SpvOpAtomicStore
: return nir_intrinsic_store_deref
;
3027 #define OP(S, N) case SpvOp##S: return nir_intrinsic_deref_##N;
3028 OP(AtomicExchange
, atomic_exchange
)
3029 OP(AtomicCompareExchange
, atomic_comp_swap
)
3030 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
3031 OP(AtomicIIncrement
, atomic_add
)
3032 OP(AtomicIDecrement
, atomic_add
)
3033 OP(AtomicIAdd
, atomic_add
)
3034 OP(AtomicISub
, atomic_add
)
3035 OP(AtomicSMin
, atomic_imin
)
3036 OP(AtomicUMin
, atomic_umin
)
3037 OP(AtomicSMax
, atomic_imax
)
3038 OP(AtomicUMax
, atomic_umax
)
3039 OP(AtomicAnd
, atomic_and
)
3040 OP(AtomicOr
, atomic_or
)
3041 OP(AtomicXor
, atomic_xor
)
3044 vtn_fail_with_opcode("Invalid shared atomic", opcode
);
3049 * Handles shared atomics, ssbo atomics and atomic counters.
3052 vtn_handle_atomics(struct vtn_builder
*b
, SpvOp opcode
,
3053 const uint32_t *w
, unsigned count
)
3055 struct vtn_pointer
*ptr
;
3056 nir_intrinsic_instr
*atomic
;
3058 SpvScope scope
= SpvScopeInvocation
;
3059 SpvMemorySemanticsMask semantics
= 0;
3062 case SpvOpAtomicLoad
:
3063 case SpvOpAtomicExchange
:
3064 case SpvOpAtomicCompareExchange
:
3065 case SpvOpAtomicCompareExchangeWeak
:
3066 case SpvOpAtomicIIncrement
:
3067 case SpvOpAtomicIDecrement
:
3068 case SpvOpAtomicIAdd
:
3069 case SpvOpAtomicISub
:
3070 case SpvOpAtomicSMin
:
3071 case SpvOpAtomicUMin
:
3072 case SpvOpAtomicSMax
:
3073 case SpvOpAtomicUMax
:
3074 case SpvOpAtomicAnd
:
3076 case SpvOpAtomicXor
:
3077 ptr
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
3078 scope
= vtn_constant_uint(b
, w
[4]);
3079 semantics
= vtn_constant_uint(b
, w
[5]);
3082 case SpvOpAtomicStore
:
3083 ptr
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
3084 scope
= vtn_constant_uint(b
, w
[2]);
3085 semantics
= vtn_constant_uint(b
, w
[3]);
3089 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
3092 /* uniform as "atomic counter uniform" */
3093 if (ptr
->mode
== vtn_variable_mode_uniform
) {
3094 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
3095 const struct glsl_type
*deref_type
= deref
->type
;
3096 nir_intrinsic_op op
= get_uniform_nir_atomic_op(b
, opcode
);
3097 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
3098 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
3100 /* SSBO needs to initialize index/offset. In this case we don't need to,
3101 * as that info is already stored on the ptr->var->var nir_variable (see
3102 * vtn_create_variable)
3106 case SpvOpAtomicLoad
:
3107 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
3110 case SpvOpAtomicStore
:
3111 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
3112 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
3115 case SpvOpAtomicExchange
:
3116 case SpvOpAtomicCompareExchange
:
3117 case SpvOpAtomicCompareExchangeWeak
:
3118 case SpvOpAtomicIIncrement
:
3119 case SpvOpAtomicIDecrement
:
3120 case SpvOpAtomicIAdd
:
3121 case SpvOpAtomicISub
:
3122 case SpvOpAtomicSMin
:
3123 case SpvOpAtomicUMin
:
3124 case SpvOpAtomicSMax
:
3125 case SpvOpAtomicUMax
:
3126 case SpvOpAtomicAnd
:
3128 case SpvOpAtomicXor
:
3129 /* Nothing: we don't need to call fill_common_atomic_sources here, as
3130 * atomic counter uniforms doesn't have sources
3135 unreachable("Invalid SPIR-V atomic");
3138 } else if (vtn_pointer_uses_ssa_offset(b
, ptr
)) {
3139 nir_ssa_def
*offset
, *index
;
3140 offset
= vtn_pointer_to_offset(b
, ptr
, &index
);
3142 assert(ptr
->mode
== vtn_variable_mode_ssbo
);
3144 nir_intrinsic_op op
= get_ssbo_nir_atomic_op(b
, opcode
);
3145 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
3149 case SpvOpAtomicLoad
:
3150 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
3151 nir_intrinsic_set_align(atomic
, 4, 0);
3152 if (ptr
->mode
== vtn_variable_mode_ssbo
)
3153 atomic
->src
[src
++] = nir_src_for_ssa(index
);
3154 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
3157 case SpvOpAtomicStore
:
3158 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
3159 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
3160 nir_intrinsic_set_align(atomic
, 4, 0);
3161 atomic
->src
[src
++] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
3162 if (ptr
->mode
== vtn_variable_mode_ssbo
)
3163 atomic
->src
[src
++] = nir_src_for_ssa(index
);
3164 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
3167 case SpvOpAtomicExchange
:
3168 case SpvOpAtomicCompareExchange
:
3169 case SpvOpAtomicCompareExchangeWeak
:
3170 case SpvOpAtomicIIncrement
:
3171 case SpvOpAtomicIDecrement
:
3172 case SpvOpAtomicIAdd
:
3173 case SpvOpAtomicISub
:
3174 case SpvOpAtomicSMin
:
3175 case SpvOpAtomicUMin
:
3176 case SpvOpAtomicSMax
:
3177 case SpvOpAtomicUMax
:
3178 case SpvOpAtomicAnd
:
3180 case SpvOpAtomicXor
:
3181 if (ptr
->mode
== vtn_variable_mode_ssbo
)
3182 atomic
->src
[src
++] = nir_src_for_ssa(index
);
3183 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
3184 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[src
]);
3188 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
3191 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
3192 const struct glsl_type
*deref_type
= deref
->type
;
3193 nir_intrinsic_op op
= get_deref_nir_atomic_op(b
, opcode
);
3194 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
3195 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
3198 case SpvOpAtomicLoad
:
3199 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
3202 case SpvOpAtomicStore
:
3203 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
3204 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
3205 atomic
->src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
3208 case SpvOpAtomicExchange
:
3209 case SpvOpAtomicCompareExchange
:
3210 case SpvOpAtomicCompareExchangeWeak
:
3211 case SpvOpAtomicIIncrement
:
3212 case SpvOpAtomicIDecrement
:
3213 case SpvOpAtomicIAdd
:
3214 case SpvOpAtomicISub
:
3215 case SpvOpAtomicSMin
:
3216 case SpvOpAtomicUMin
:
3217 case SpvOpAtomicSMax
:
3218 case SpvOpAtomicUMax
:
3219 case SpvOpAtomicAnd
:
3221 case SpvOpAtomicXor
:
3222 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[1]);
3226 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
3230 /* Atomic ordering operations will implicitly apply to the atomic operation
3231 * storage class, so include that too.
3233 semantics
|= vtn_storage_class_to_memory_semantics(ptr
->ptr_type
->storage_class
);
3235 SpvMemorySemanticsMask before_semantics
;
3236 SpvMemorySemanticsMask after_semantics
;
3237 vtn_split_barrier_semantics(b
, semantics
, &before_semantics
, &after_semantics
);
3239 if (before_semantics
)
3240 vtn_emit_memory_barrier(b
, scope
, before_semantics
);
3242 if (opcode
!= SpvOpAtomicStore
) {
3243 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
3245 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
,
3246 glsl_get_vector_elements(type
->type
),
3247 glsl_get_bit_size(type
->type
), NULL
);
3249 struct vtn_ssa_value
*ssa
= rzalloc(b
, struct vtn_ssa_value
);
3250 ssa
->def
= &atomic
->dest
.ssa
;
3251 ssa
->type
= type
->type
;
3252 vtn_push_ssa(b
, w
[2], type
, ssa
);
3255 nir_builder_instr_insert(&b
->nb
, &atomic
->instr
);
3257 if (after_semantics
)
3258 vtn_emit_memory_barrier(b
, scope
, after_semantics
);
3261 static nir_alu_instr
*
3262 create_vec(struct vtn_builder
*b
, unsigned num_components
, unsigned bit_size
)
3264 nir_op op
= nir_op_vec(num_components
);
3265 nir_alu_instr
*vec
= nir_alu_instr_create(b
->shader
, op
);
3266 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, num_components
,
3268 vec
->dest
.write_mask
= (1 << num_components
) - 1;
3273 struct vtn_ssa_value
*
3274 vtn_ssa_transpose(struct vtn_builder
*b
, struct vtn_ssa_value
*src
)
3276 if (src
->transposed
)
3277 return src
->transposed
;
3279 struct vtn_ssa_value
*dest
=
3280 vtn_create_ssa_value(b
, glsl_transposed_type(src
->type
));
3282 for (unsigned i
= 0; i
< glsl_get_matrix_columns(dest
->type
); i
++) {
3283 nir_alu_instr
*vec
= create_vec(b
, glsl_get_matrix_columns(src
->type
),
3284 glsl_get_bit_size(src
->type
));
3285 if (glsl_type_is_vector_or_scalar(src
->type
)) {
3286 vec
->src
[0].src
= nir_src_for_ssa(src
->def
);
3287 vec
->src
[0].swizzle
[0] = i
;
3289 for (unsigned j
= 0; j
< glsl_get_matrix_columns(src
->type
); j
++) {
3290 vec
->src
[j
].src
= nir_src_for_ssa(src
->elems
[j
]->def
);
3291 vec
->src
[j
].swizzle
[0] = i
;
3294 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3295 dest
->elems
[i
]->def
= &vec
->dest
.dest
.ssa
;
3298 dest
->transposed
= src
;
3304 vtn_vector_extract(struct vtn_builder
*b
, nir_ssa_def
*src
, unsigned index
)
3306 return nir_channel(&b
->nb
, src
, index
);
3310 vtn_vector_insert(struct vtn_builder
*b
, nir_ssa_def
*src
, nir_ssa_def
*insert
,
3313 nir_alu_instr
*vec
= create_vec(b
, src
->num_components
,
3316 for (unsigned i
= 0; i
< src
->num_components
; i
++) {
3318 vec
->src
[i
].src
= nir_src_for_ssa(insert
);
3320 vec
->src
[i
].src
= nir_src_for_ssa(src
);
3321 vec
->src
[i
].swizzle
[0] = i
;
3325 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3327 return &vec
->dest
.dest
.ssa
;
3330 static nir_ssa_def
*
3331 nir_ieq_imm(nir_builder
*b
, nir_ssa_def
*x
, uint64_t i
)
3333 return nir_ieq(b
, x
, nir_imm_intN_t(b
, i
, x
->bit_size
));
3337 vtn_vector_extract_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
3340 return nir_vector_extract(&b
->nb
, src
, nir_i2i(&b
->nb
, index
, 32));
3344 vtn_vector_insert_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
3345 nir_ssa_def
*insert
, nir_ssa_def
*index
)
3347 nir_ssa_def
*dest
= vtn_vector_insert(b
, src
, insert
, 0);
3348 for (unsigned i
= 1; i
< src
->num_components
; i
++)
3349 dest
= nir_bcsel(&b
->nb
, nir_ieq_imm(&b
->nb
, index
, i
),
3350 vtn_vector_insert(b
, src
, insert
, i
), dest
);
3355 static nir_ssa_def
*
3356 vtn_vector_shuffle(struct vtn_builder
*b
, unsigned num_components
,
3357 nir_ssa_def
*src0
, nir_ssa_def
*src1
,
3358 const uint32_t *indices
)
3360 nir_alu_instr
*vec
= create_vec(b
, num_components
, src0
->bit_size
);
3362 for (unsigned i
= 0; i
< num_components
; i
++) {
3363 uint32_t index
= indices
[i
];
3364 if (index
== 0xffffffff) {
3366 nir_src_for_ssa(nir_ssa_undef(&b
->nb
, 1, src0
->bit_size
));
3367 } else if (index
< src0
->num_components
) {
3368 vec
->src
[i
].src
= nir_src_for_ssa(src0
);
3369 vec
->src
[i
].swizzle
[0] = index
;
3371 vec
->src
[i
].src
= nir_src_for_ssa(src1
);
3372 vec
->src
[i
].swizzle
[0] = index
- src0
->num_components
;
3376 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3378 return &vec
->dest
.dest
.ssa
;
3382 * Concatentates a number of vectors/scalars together to produce a vector
3384 static nir_ssa_def
*
3385 vtn_vector_construct(struct vtn_builder
*b
, unsigned num_components
,
3386 unsigned num_srcs
, nir_ssa_def
**srcs
)
3388 nir_alu_instr
*vec
= create_vec(b
, num_components
, srcs
[0]->bit_size
);
3390 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
3392 * "When constructing a vector, there must be at least two Constituent
3395 vtn_assert(num_srcs
>= 2);
3397 unsigned dest_idx
= 0;
3398 for (unsigned i
= 0; i
< num_srcs
; i
++) {
3399 nir_ssa_def
*src
= srcs
[i
];
3400 vtn_assert(dest_idx
+ src
->num_components
<= num_components
);
3401 for (unsigned j
= 0; j
< src
->num_components
; j
++) {
3402 vec
->src
[dest_idx
].src
= nir_src_for_ssa(src
);
3403 vec
->src
[dest_idx
].swizzle
[0] = j
;
3408 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
3410 * "When constructing a vector, the total number of components in all
3411 * the operands must equal the number of components in Result Type."
3413 vtn_assert(dest_idx
== num_components
);
3415 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3417 return &vec
->dest
.dest
.ssa
;
3420 static struct vtn_ssa_value
*
3421 vtn_composite_copy(void *mem_ctx
, struct vtn_ssa_value
*src
)
3423 struct vtn_ssa_value
*dest
= rzalloc(mem_ctx
, struct vtn_ssa_value
);
3424 dest
->type
= src
->type
;
3426 if (glsl_type_is_vector_or_scalar(src
->type
)) {
3427 dest
->def
= src
->def
;
3429 unsigned elems
= glsl_get_length(src
->type
);
3431 dest
->elems
= ralloc_array(mem_ctx
, struct vtn_ssa_value
*, elems
);
3432 for (unsigned i
= 0; i
< elems
; i
++)
3433 dest
->elems
[i
] = vtn_composite_copy(mem_ctx
, src
->elems
[i
]);
3439 static struct vtn_ssa_value
*
3440 vtn_composite_insert(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
3441 struct vtn_ssa_value
*insert
, const uint32_t *indices
,
3442 unsigned num_indices
)
3444 struct vtn_ssa_value
*dest
= vtn_composite_copy(b
, src
);
3446 struct vtn_ssa_value
*cur
= dest
;
3448 for (i
= 0; i
< num_indices
- 1; i
++) {
3449 cur
= cur
->elems
[indices
[i
]];
3452 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
3453 /* According to the SPIR-V spec, OpCompositeInsert may work down to
3454 * the component granularity. In that case, the last index will be
3455 * the index to insert the scalar into the vector.
3458 cur
->def
= vtn_vector_insert(b
, cur
->def
, insert
->def
, indices
[i
]);
3460 cur
->elems
[indices
[i
]] = insert
;
3466 static struct vtn_ssa_value
*
3467 vtn_composite_extract(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
3468 const uint32_t *indices
, unsigned num_indices
)
3470 struct vtn_ssa_value
*cur
= src
;
3471 for (unsigned i
= 0; i
< num_indices
; i
++) {
3472 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
3473 vtn_assert(i
== num_indices
- 1);
3474 /* According to the SPIR-V spec, OpCompositeExtract may work down to
3475 * the component granularity. The last index will be the index of the
3476 * vector to extract.
3479 struct vtn_ssa_value
*ret
= rzalloc(b
, struct vtn_ssa_value
);
3480 ret
->type
= glsl_scalar_type(glsl_get_base_type(cur
->type
));
3481 ret
->def
= vtn_vector_extract(b
, cur
->def
, indices
[i
]);
3484 cur
= cur
->elems
[indices
[i
]];
3492 vtn_handle_composite(struct vtn_builder
*b
, SpvOp opcode
,
3493 const uint32_t *w
, unsigned count
)
3495 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
3496 struct vtn_ssa_value
*ssa
= vtn_create_ssa_value(b
, type
->type
);
3499 case SpvOpVectorExtractDynamic
:
3500 ssa
->def
= vtn_vector_extract_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
3501 vtn_ssa_value(b
, w
[4])->def
);
3504 case SpvOpVectorInsertDynamic
:
3505 ssa
->def
= vtn_vector_insert_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
3506 vtn_ssa_value(b
, w
[4])->def
,
3507 vtn_ssa_value(b
, w
[5])->def
);
3510 case SpvOpVectorShuffle
:
3511 ssa
->def
= vtn_vector_shuffle(b
, glsl_get_vector_elements(type
->type
),
3512 vtn_ssa_value(b
, w
[3])->def
,
3513 vtn_ssa_value(b
, w
[4])->def
,
3517 case SpvOpCompositeConstruct
: {
3518 unsigned elems
= count
- 3;
3520 if (glsl_type_is_vector_or_scalar(type
->type
)) {
3521 nir_ssa_def
*srcs
[NIR_MAX_VEC_COMPONENTS
];
3522 for (unsigned i
= 0; i
< elems
; i
++)
3523 srcs
[i
] = vtn_ssa_value(b
, w
[3 + i
])->def
;
3525 vtn_vector_construct(b
, glsl_get_vector_elements(type
->type
),
3528 ssa
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
3529 for (unsigned i
= 0; i
< elems
; i
++)
3530 ssa
->elems
[i
] = vtn_ssa_value(b
, w
[3 + i
]);
3534 case SpvOpCompositeExtract
:
3535 ssa
= vtn_composite_extract(b
, vtn_ssa_value(b
, w
[3]),
3539 case SpvOpCompositeInsert
:
3540 ssa
= vtn_composite_insert(b
, vtn_ssa_value(b
, w
[4]),
3541 vtn_ssa_value(b
, w
[3]),
3545 case SpvOpCopyLogical
:
3546 case SpvOpCopyObject
:
3547 ssa
= vtn_composite_copy(b
, vtn_ssa_value(b
, w
[3]));
3551 vtn_fail_with_opcode("unknown composite operation", opcode
);
3554 vtn_push_ssa(b
, w
[2], type
, ssa
);
3558 vtn_emit_barrier(struct vtn_builder
*b
, nir_intrinsic_op op
)
3560 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
3561 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3565 vtn_emit_memory_barrier(struct vtn_builder
*b
, SpvScope scope
,
3566 SpvMemorySemanticsMask semantics
)
3568 if (b
->options
->use_scoped_memory_barrier
) {
3569 vtn_emit_scoped_memory_barrier(b
, scope
, semantics
);
3573 static const SpvMemorySemanticsMask all_memory_semantics
=
3574 SpvMemorySemanticsUniformMemoryMask
|
3575 SpvMemorySemanticsWorkgroupMemoryMask
|
3576 SpvMemorySemanticsAtomicCounterMemoryMask
|
3577 SpvMemorySemanticsImageMemoryMask
;
3579 /* If we're not actually doing a memory barrier, bail */
3580 if (!(semantics
& all_memory_semantics
))
3583 /* GL and Vulkan don't have these */
3584 vtn_assert(scope
!= SpvScopeCrossDevice
);
3586 if (scope
== SpvScopeSubgroup
)
3587 return; /* Nothing to do here */
3589 if (scope
== SpvScopeWorkgroup
) {
3590 vtn_emit_barrier(b
, nir_intrinsic_group_memory_barrier
);
3594 /* There's only two scopes thing left */
3595 vtn_assert(scope
== SpvScopeInvocation
|| scope
== SpvScopeDevice
);
3597 if ((semantics
& all_memory_semantics
) == all_memory_semantics
) {
3598 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier
);
3602 /* Issue a bunch of more specific barriers */
3603 uint32_t bits
= semantics
;
3605 SpvMemorySemanticsMask semantic
= 1 << u_bit_scan(&bits
);
3607 case SpvMemorySemanticsUniformMemoryMask
:
3608 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_buffer
);
3610 case SpvMemorySemanticsWorkgroupMemoryMask
:
3611 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_shared
);
3613 case SpvMemorySemanticsAtomicCounterMemoryMask
:
3614 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_atomic_counter
);
3616 case SpvMemorySemanticsImageMemoryMask
:
3617 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_image
);
3619 case SpvMemorySemanticsOutputMemoryMask
:
3620 if (b
->nb
.shader
->info
.stage
== MESA_SHADER_TESS_CTRL
)
3621 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_tcs_patch
);
3630 vtn_handle_barrier(struct vtn_builder
*b
, SpvOp opcode
,
3631 const uint32_t *w
, unsigned count
)
3634 case SpvOpEmitVertex
:
3635 case SpvOpEmitStreamVertex
:
3636 case SpvOpEndPrimitive
:
3637 case SpvOpEndStreamPrimitive
: {
3638 nir_intrinsic_op intrinsic_op
;
3640 case SpvOpEmitVertex
:
3641 case SpvOpEmitStreamVertex
:
3642 intrinsic_op
= nir_intrinsic_emit_vertex
;
3644 case SpvOpEndPrimitive
:
3645 case SpvOpEndStreamPrimitive
:
3646 intrinsic_op
= nir_intrinsic_end_primitive
;
3649 unreachable("Invalid opcode");
3652 nir_intrinsic_instr
*intrin
=
3653 nir_intrinsic_instr_create(b
->shader
, intrinsic_op
);
3656 case SpvOpEmitStreamVertex
:
3657 case SpvOpEndStreamPrimitive
: {
3658 unsigned stream
= vtn_constant_uint(b
, w
[1]);
3659 nir_intrinsic_set_stream_id(intrin
, stream
);
3667 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3671 case SpvOpMemoryBarrier
: {
3672 SpvScope scope
= vtn_constant_uint(b
, w
[1]);
3673 SpvMemorySemanticsMask semantics
= vtn_constant_uint(b
, w
[2]);
3674 vtn_emit_memory_barrier(b
, scope
, semantics
);
3678 case SpvOpControlBarrier
: {
3679 SpvScope execution_scope
= vtn_constant_uint(b
, w
[1]);
3680 SpvScope memory_scope
= vtn_constant_uint(b
, w
[2]);
3681 SpvMemorySemanticsMask memory_semantics
= vtn_constant_uint(b
, w
[3]);
3683 /* GLSLang, prior to commit 8297936dd6eb3, emitted OpControlBarrier with
3684 * memory semantics of None for GLSL barrier().
3686 if (b
->wa_glslang_cs_barrier
&&
3687 b
->nb
.shader
->info
.stage
== MESA_SHADER_COMPUTE
&&
3688 execution_scope
== SpvScopeWorkgroup
&&
3689 memory_semantics
== SpvMemorySemanticsMaskNone
) {
3690 memory_scope
= SpvScopeWorkgroup
;
3691 memory_semantics
= SpvMemorySemanticsAcquireReleaseMask
|
3692 SpvMemorySemanticsWorkgroupMemoryMask
;
3695 /* From the SPIR-V spec:
3697 * "When used with the TessellationControl execution model, it also
3698 * implicitly synchronizes the Output Storage Class: Writes to Output
3699 * variables performed by any invocation executed prior to a
3700 * OpControlBarrier will be visible to any other invocation after
3701 * return from that OpControlBarrier."
3703 if (b
->nb
.shader
->info
.stage
== MESA_SHADER_TESS_CTRL
) {
3704 memory_semantics
&= ~(SpvMemorySemanticsAcquireMask
|
3705 SpvMemorySemanticsReleaseMask
|
3706 SpvMemorySemanticsAcquireReleaseMask
|
3707 SpvMemorySemanticsSequentiallyConsistentMask
);
3708 memory_semantics
|= SpvMemorySemanticsAcquireReleaseMask
|
3709 SpvMemorySemanticsOutputMemoryMask
;
3712 vtn_emit_memory_barrier(b
, memory_scope
, memory_semantics
);
3714 if (execution_scope
== SpvScopeWorkgroup
)
3715 vtn_emit_barrier(b
, nir_intrinsic_control_barrier
);
3720 unreachable("unknown barrier instruction");
3725 gl_primitive_from_spv_execution_mode(struct vtn_builder
*b
,
3726 SpvExecutionMode mode
)
3729 case SpvExecutionModeInputPoints
:
3730 case SpvExecutionModeOutputPoints
:
3731 return 0; /* GL_POINTS */
3732 case SpvExecutionModeInputLines
:
3733 return 1; /* GL_LINES */
3734 case SpvExecutionModeInputLinesAdjacency
:
3735 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
3736 case SpvExecutionModeTriangles
:
3737 return 4; /* GL_TRIANGLES */
3738 case SpvExecutionModeInputTrianglesAdjacency
:
3739 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
3740 case SpvExecutionModeQuads
:
3741 return 7; /* GL_QUADS */
3742 case SpvExecutionModeIsolines
:
3743 return 0x8E7A; /* GL_ISOLINES */
3744 case SpvExecutionModeOutputLineStrip
:
3745 return 3; /* GL_LINE_STRIP */
3746 case SpvExecutionModeOutputTriangleStrip
:
3747 return 5; /* GL_TRIANGLE_STRIP */
3749 vtn_fail("Invalid primitive type: %s (%u)",
3750 spirv_executionmode_to_string(mode
), mode
);
3755 vertices_in_from_spv_execution_mode(struct vtn_builder
*b
,
3756 SpvExecutionMode mode
)
3759 case SpvExecutionModeInputPoints
:
3761 case SpvExecutionModeInputLines
:
3763 case SpvExecutionModeInputLinesAdjacency
:
3765 case SpvExecutionModeTriangles
:
3767 case SpvExecutionModeInputTrianglesAdjacency
:
3770 vtn_fail("Invalid GS input mode: %s (%u)",
3771 spirv_executionmode_to_string(mode
), mode
);
3775 static gl_shader_stage
3776 stage_for_execution_model(struct vtn_builder
*b
, SpvExecutionModel model
)
3779 case SpvExecutionModelVertex
:
3780 return MESA_SHADER_VERTEX
;
3781 case SpvExecutionModelTessellationControl
:
3782 return MESA_SHADER_TESS_CTRL
;
3783 case SpvExecutionModelTessellationEvaluation
:
3784 return MESA_SHADER_TESS_EVAL
;
3785 case SpvExecutionModelGeometry
:
3786 return MESA_SHADER_GEOMETRY
;
3787 case SpvExecutionModelFragment
:
3788 return MESA_SHADER_FRAGMENT
;
3789 case SpvExecutionModelGLCompute
:
3790 return MESA_SHADER_COMPUTE
;
3791 case SpvExecutionModelKernel
:
3792 return MESA_SHADER_KERNEL
;
3794 vtn_fail("Unsupported execution model: %s (%u)",
3795 spirv_executionmodel_to_string(model
), model
);
3799 #define spv_check_supported(name, cap) do { \
3800 if (!(b->options && b->options->caps.name)) \
3801 vtn_warn("Unsupported SPIR-V capability: %s (%u)", \
3802 spirv_capability_to_string(cap), cap); \
3807 vtn_handle_entry_point(struct vtn_builder
*b
, const uint32_t *w
,
3810 struct vtn_value
*entry_point
= &b
->values
[w
[2]];
3811 /* Let this be a name label regardless */
3812 unsigned name_words
;
3813 entry_point
->name
= vtn_string_literal(b
, &w
[3], count
- 3, &name_words
);
3815 if (strcmp(entry_point
->name
, b
->entry_point_name
) != 0 ||
3816 stage_for_execution_model(b
, w
[1]) != b
->entry_point_stage
)
3819 vtn_assert(b
->entry_point
== NULL
);
3820 b
->entry_point
= entry_point
;
3824 vtn_handle_preamble_instruction(struct vtn_builder
*b
, SpvOp opcode
,
3825 const uint32_t *w
, unsigned count
)
3832 case SpvSourceLanguageUnknown
: lang
= "unknown"; break;
3833 case SpvSourceLanguageESSL
: lang
= "ESSL"; break;
3834 case SpvSourceLanguageGLSL
: lang
= "GLSL"; break;
3835 case SpvSourceLanguageOpenCL_C
: lang
= "OpenCL C"; break;
3836 case SpvSourceLanguageOpenCL_CPP
: lang
= "OpenCL C++"; break;
3837 case SpvSourceLanguageHLSL
: lang
= "HLSL"; break;
3840 uint32_t version
= w
[2];
3843 (count
> 3) ? vtn_value(b
, w
[3], vtn_value_type_string
)->str
: "";
3845 vtn_info("Parsing SPIR-V from %s %u source file %s", lang
, version
, file
);
3849 case SpvOpSourceExtension
:
3850 case SpvOpSourceContinued
:
3851 case SpvOpExtension
:
3852 case SpvOpModuleProcessed
:
3853 /* Unhandled, but these are for debug so that's ok. */
3856 case SpvOpCapability
: {
3857 SpvCapability cap
= w
[1];
3859 case SpvCapabilityMatrix
:
3860 case SpvCapabilityShader
:
3861 case SpvCapabilityGeometry
:
3862 case SpvCapabilityGeometryPointSize
:
3863 case SpvCapabilityUniformBufferArrayDynamicIndexing
:
3864 case SpvCapabilitySampledImageArrayDynamicIndexing
:
3865 case SpvCapabilityStorageBufferArrayDynamicIndexing
:
3866 case SpvCapabilityStorageImageArrayDynamicIndexing
:
3867 case SpvCapabilityImageRect
:
3868 case SpvCapabilitySampledRect
:
3869 case SpvCapabilitySampled1D
:
3870 case SpvCapabilityImage1D
:
3871 case SpvCapabilitySampledCubeArray
:
3872 case SpvCapabilityImageCubeArray
:
3873 case SpvCapabilitySampledBuffer
:
3874 case SpvCapabilityImageBuffer
:
3875 case SpvCapabilityImageQuery
:
3876 case SpvCapabilityDerivativeControl
:
3877 case SpvCapabilityInterpolationFunction
:
3878 case SpvCapabilityMultiViewport
:
3879 case SpvCapabilitySampleRateShading
:
3880 case SpvCapabilityClipDistance
:
3881 case SpvCapabilityCullDistance
:
3882 case SpvCapabilityInputAttachment
:
3883 case SpvCapabilityImageGatherExtended
:
3884 case SpvCapabilityStorageImageExtendedFormats
:
3885 case SpvCapabilityVector16
:
3888 case SpvCapabilityLinkage
:
3889 case SpvCapabilityFloat16Buffer
:
3890 case SpvCapabilitySparseResidency
:
3891 vtn_warn("Unsupported SPIR-V capability: %s",
3892 spirv_capability_to_string(cap
));
3895 case SpvCapabilityMinLod
:
3896 spv_check_supported(min_lod
, cap
);
3899 case SpvCapabilityAtomicStorage
:
3900 spv_check_supported(atomic_storage
, cap
);
3903 case SpvCapabilityFloat64
:
3904 spv_check_supported(float64
, cap
);
3906 case SpvCapabilityInt64
:
3907 spv_check_supported(int64
, cap
);
3909 case SpvCapabilityInt16
:
3910 spv_check_supported(int16
, cap
);
3912 case SpvCapabilityInt8
:
3913 spv_check_supported(int8
, cap
);
3916 case SpvCapabilityTransformFeedback
:
3917 spv_check_supported(transform_feedback
, cap
);
3920 case SpvCapabilityGeometryStreams
:
3921 spv_check_supported(geometry_streams
, cap
);
3924 case SpvCapabilityInt64Atomics
:
3925 spv_check_supported(int64_atomics
, cap
);
3928 case SpvCapabilityStorageImageMultisample
:
3929 spv_check_supported(storage_image_ms
, cap
);
3932 case SpvCapabilityAddresses
:
3933 spv_check_supported(address
, cap
);
3936 case SpvCapabilityKernel
:
3937 spv_check_supported(kernel
, cap
);
3940 case SpvCapabilityImageBasic
:
3941 case SpvCapabilityImageReadWrite
:
3942 case SpvCapabilityImageMipmap
:
3943 case SpvCapabilityPipes
:
3944 case SpvCapabilityDeviceEnqueue
:
3945 case SpvCapabilityLiteralSampler
:
3946 case SpvCapabilityGenericPointer
:
3947 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
3948 spirv_capability_to_string(cap
));
3951 case SpvCapabilityImageMSArray
:
3952 spv_check_supported(image_ms_array
, cap
);
3955 case SpvCapabilityTessellation
:
3956 case SpvCapabilityTessellationPointSize
:
3957 spv_check_supported(tessellation
, cap
);
3960 case SpvCapabilityDrawParameters
:
3961 spv_check_supported(draw_parameters
, cap
);
3964 case SpvCapabilityStorageImageReadWithoutFormat
:
3965 spv_check_supported(image_read_without_format
, cap
);
3968 case SpvCapabilityStorageImageWriteWithoutFormat
:
3969 spv_check_supported(image_write_without_format
, cap
);
3972 case SpvCapabilityDeviceGroup
:
3973 spv_check_supported(device_group
, cap
);
3976 case SpvCapabilityMultiView
:
3977 spv_check_supported(multiview
, cap
);
3980 case SpvCapabilityGroupNonUniform
:
3981 spv_check_supported(subgroup_basic
, cap
);
3984 case SpvCapabilitySubgroupVoteKHR
:
3985 case SpvCapabilityGroupNonUniformVote
:
3986 spv_check_supported(subgroup_vote
, cap
);
3989 case SpvCapabilitySubgroupBallotKHR
:
3990 case SpvCapabilityGroupNonUniformBallot
:
3991 spv_check_supported(subgroup_ballot
, cap
);
3994 case SpvCapabilityGroupNonUniformShuffle
:
3995 case SpvCapabilityGroupNonUniformShuffleRelative
:
3996 spv_check_supported(subgroup_shuffle
, cap
);
3999 case SpvCapabilityGroupNonUniformQuad
:
4000 spv_check_supported(subgroup_quad
, cap
);
4003 case SpvCapabilityGroupNonUniformArithmetic
:
4004 case SpvCapabilityGroupNonUniformClustered
:
4005 spv_check_supported(subgroup_arithmetic
, cap
);
4008 case SpvCapabilityGroups
:
4009 spv_check_supported(amd_shader_ballot
, cap
);
4012 case SpvCapabilityVariablePointersStorageBuffer
:
4013 case SpvCapabilityVariablePointers
:
4014 spv_check_supported(variable_pointers
, cap
);
4015 b
->variable_pointers
= true;
4018 case SpvCapabilityStorageUniformBufferBlock16
:
4019 case SpvCapabilityStorageUniform16
:
4020 case SpvCapabilityStoragePushConstant16
:
4021 case SpvCapabilityStorageInputOutput16
:
4022 spv_check_supported(storage_16bit
, cap
);
4025 case SpvCapabilityShaderLayer
:
4026 case SpvCapabilityShaderViewportIndex
:
4027 case SpvCapabilityShaderViewportIndexLayerEXT
:
4028 spv_check_supported(shader_viewport_index_layer
, cap
);
4031 case SpvCapabilityStorageBuffer8BitAccess
:
4032 case SpvCapabilityUniformAndStorageBuffer8BitAccess
:
4033 case SpvCapabilityStoragePushConstant8
:
4034 spv_check_supported(storage_8bit
, cap
);
4037 case SpvCapabilityShaderNonUniformEXT
:
4038 spv_check_supported(descriptor_indexing
, cap
);
4041 case SpvCapabilityInputAttachmentArrayDynamicIndexingEXT
:
4042 case SpvCapabilityUniformTexelBufferArrayDynamicIndexingEXT
:
4043 case SpvCapabilityStorageTexelBufferArrayDynamicIndexingEXT
:
4044 spv_check_supported(descriptor_array_dynamic_indexing
, cap
);
4047 case SpvCapabilityUniformBufferArrayNonUniformIndexingEXT
:
4048 case SpvCapabilitySampledImageArrayNonUniformIndexingEXT
:
4049 case SpvCapabilityStorageBufferArrayNonUniformIndexingEXT
:
4050 case SpvCapabilityStorageImageArrayNonUniformIndexingEXT
:
4051 case SpvCapabilityInputAttachmentArrayNonUniformIndexingEXT
:
4052 case SpvCapabilityUniformTexelBufferArrayNonUniformIndexingEXT
:
4053 case SpvCapabilityStorageTexelBufferArrayNonUniformIndexingEXT
:
4054 spv_check_supported(descriptor_array_non_uniform_indexing
, cap
);
4057 case SpvCapabilityRuntimeDescriptorArrayEXT
:
4058 spv_check_supported(runtime_descriptor_array
, cap
);
4061 case SpvCapabilityStencilExportEXT
:
4062 spv_check_supported(stencil_export
, cap
);
4065 case SpvCapabilitySampleMaskPostDepthCoverage
:
4066 spv_check_supported(post_depth_coverage
, cap
);
4069 case SpvCapabilityDenormFlushToZero
:
4070 case SpvCapabilityDenormPreserve
:
4071 case SpvCapabilitySignedZeroInfNanPreserve
:
4072 case SpvCapabilityRoundingModeRTE
:
4073 case SpvCapabilityRoundingModeRTZ
:
4074 spv_check_supported(float_controls
, cap
);
4077 case SpvCapabilityPhysicalStorageBufferAddressesEXT
:
4078 spv_check_supported(physical_storage_buffer_address
, cap
);
4081 case SpvCapabilityComputeDerivativeGroupQuadsNV
:
4082 case SpvCapabilityComputeDerivativeGroupLinearNV
:
4083 spv_check_supported(derivative_group
, cap
);
4086 case SpvCapabilityFloat16
:
4087 spv_check_supported(float16
, cap
);
4090 case SpvCapabilityFragmentShaderSampleInterlockEXT
:
4091 spv_check_supported(fragment_shader_sample_interlock
, cap
);
4094 case SpvCapabilityFragmentShaderPixelInterlockEXT
:
4095 spv_check_supported(fragment_shader_pixel_interlock
, cap
);
4098 case SpvCapabilityDemoteToHelperInvocationEXT
:
4099 spv_check_supported(demote_to_helper_invocation
, cap
);
4102 case SpvCapabilityShaderClockKHR
:
4103 spv_check_supported(shader_clock
, cap
);
4106 case SpvCapabilityVulkanMemoryModel
:
4107 spv_check_supported(vk_memory_model
, cap
);
4110 case SpvCapabilityVulkanMemoryModelDeviceScope
:
4111 spv_check_supported(vk_memory_model_device_scope
, cap
);
4114 case SpvCapabilityImageReadWriteLodAMD
:
4115 spv_check_supported(amd_image_read_write_lod
, cap
);
4119 vtn_fail("Unhandled capability: %s (%u)",
4120 spirv_capability_to_string(cap
), cap
);
4125 case SpvOpExtInstImport
:
4126 vtn_handle_extension(b
, opcode
, w
, count
);
4129 case SpvOpMemoryModel
:
4131 case SpvAddressingModelPhysical32
:
4132 vtn_fail_if(b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
,
4133 "AddressingModelPhysical32 only supported for kernels");
4134 b
->shader
->info
.cs
.ptr_size
= 32;
4135 b
->physical_ptrs
= true;
4136 b
->options
->shared_addr_format
= nir_address_format_32bit_global
;
4137 b
->options
->global_addr_format
= nir_address_format_32bit_global
;
4138 b
->options
->temp_addr_format
= nir_address_format_32bit_global
;
4140 case SpvAddressingModelPhysical64
:
4141 vtn_fail_if(b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
,
4142 "AddressingModelPhysical64 only supported for kernels");
4143 b
->shader
->info
.cs
.ptr_size
= 64;
4144 b
->physical_ptrs
= true;
4145 b
->options
->shared_addr_format
= nir_address_format_64bit_global
;
4146 b
->options
->global_addr_format
= nir_address_format_64bit_global
;
4147 b
->options
->temp_addr_format
= nir_address_format_64bit_global
;
4149 case SpvAddressingModelLogical
:
4150 vtn_fail_if(b
->shader
->info
.stage
>= MESA_SHADER_STAGES
,
4151 "AddressingModelLogical only supported for shaders");
4152 b
->physical_ptrs
= false;
4154 case SpvAddressingModelPhysicalStorageBuffer64EXT
:
4155 vtn_fail_if(!b
->options
||
4156 !b
->options
->caps
.physical_storage_buffer_address
,
4157 "AddressingModelPhysicalStorageBuffer64EXT not supported");
4160 vtn_fail("Unknown addressing model: %s (%u)",
4161 spirv_addressingmodel_to_string(w
[1]), w
[1]);
4166 case SpvMemoryModelSimple
:
4167 case SpvMemoryModelGLSL450
:
4168 case SpvMemoryModelOpenCL
:
4170 case SpvMemoryModelVulkan
:
4171 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
4172 "Vulkan memory model is unsupported by this driver");
4175 vtn_fail("Unsupported memory model: %s",
4176 spirv_memorymodel_to_string(w
[2]));
4181 case SpvOpEntryPoint
:
4182 vtn_handle_entry_point(b
, w
, count
);
4186 vtn_push_value(b
, w
[1], vtn_value_type_string
)->str
=
4187 vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
4191 b
->values
[w
[1]].name
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
4194 case SpvOpMemberName
:
4198 case SpvOpExecutionMode
:
4199 case SpvOpExecutionModeId
:
4200 case SpvOpDecorationGroup
:
4202 case SpvOpDecorateId
:
4203 case SpvOpMemberDecorate
:
4204 case SpvOpGroupDecorate
:
4205 case SpvOpGroupMemberDecorate
:
4206 case SpvOpDecorateString
:
4207 case SpvOpMemberDecorateString
:
4208 vtn_handle_decoration(b
, opcode
, w
, count
);
4211 case SpvOpExtInst
: {
4212 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
4213 if (val
->ext_handler
== vtn_handle_non_semantic_instruction
) {
4214 /* NonSemantic extended instructions are acceptable in preamble. */
4215 vtn_handle_non_semantic_instruction(b
, w
[4], w
, count
);
4218 return false; /* End of preamble. */
4223 return false; /* End of preamble */
4230 vtn_handle_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
4231 const struct vtn_decoration
*mode
, void *data
)
4233 vtn_assert(b
->entry_point
== entry_point
);
4235 switch(mode
->exec_mode
) {
4236 case SpvExecutionModeOriginUpperLeft
:
4237 case SpvExecutionModeOriginLowerLeft
:
4238 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4239 b
->shader
->info
.fs
.origin_upper_left
=
4240 (mode
->exec_mode
== SpvExecutionModeOriginUpperLeft
);
4243 case SpvExecutionModeEarlyFragmentTests
:
4244 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4245 b
->shader
->info
.fs
.early_fragment_tests
= true;
4248 case SpvExecutionModePostDepthCoverage
:
4249 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4250 b
->shader
->info
.fs
.post_depth_coverage
= true;
4253 case SpvExecutionModeInvocations
:
4254 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4255 b
->shader
->info
.gs
.invocations
= MAX2(1, mode
->operands
[0]);
4258 case SpvExecutionModeDepthReplacing
:
4259 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4260 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_ANY
;
4262 case SpvExecutionModeDepthGreater
:
4263 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4264 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_GREATER
;
4266 case SpvExecutionModeDepthLess
:
4267 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4268 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_LESS
;
4270 case SpvExecutionModeDepthUnchanged
:
4271 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4272 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
4275 case SpvExecutionModeLocalSize
:
4276 vtn_assert(gl_shader_stage_is_compute(b
->shader
->info
.stage
));
4277 b
->shader
->info
.cs
.local_size
[0] = mode
->operands
[0];
4278 b
->shader
->info
.cs
.local_size
[1] = mode
->operands
[1];
4279 b
->shader
->info
.cs
.local_size
[2] = mode
->operands
[2];
4282 case SpvExecutionModeLocalSizeId
:
4283 b
->shader
->info
.cs
.local_size
[0] = vtn_constant_uint(b
, mode
->operands
[0]);
4284 b
->shader
->info
.cs
.local_size
[1] = vtn_constant_uint(b
, mode
->operands
[1]);
4285 b
->shader
->info
.cs
.local_size
[2] = vtn_constant_uint(b
, mode
->operands
[2]);
4288 case SpvExecutionModeLocalSizeHint
:
4289 case SpvExecutionModeLocalSizeHintId
:
4290 break; /* Nothing to do with this */
4292 case SpvExecutionModeOutputVertices
:
4293 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4294 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
4295 b
->shader
->info
.tess
.tcs_vertices_out
= mode
->operands
[0];
4297 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4298 b
->shader
->info
.gs
.vertices_out
= mode
->operands
[0];
4302 case SpvExecutionModeInputPoints
:
4303 case SpvExecutionModeInputLines
:
4304 case SpvExecutionModeInputLinesAdjacency
:
4305 case SpvExecutionModeTriangles
:
4306 case SpvExecutionModeInputTrianglesAdjacency
:
4307 case SpvExecutionModeQuads
:
4308 case SpvExecutionModeIsolines
:
4309 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4310 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
4311 b
->shader
->info
.tess
.primitive_mode
=
4312 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
4314 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4315 b
->shader
->info
.gs
.vertices_in
=
4316 vertices_in_from_spv_execution_mode(b
, mode
->exec_mode
);
4317 b
->shader
->info
.gs
.input_primitive
=
4318 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
4322 case SpvExecutionModeOutputPoints
:
4323 case SpvExecutionModeOutputLineStrip
:
4324 case SpvExecutionModeOutputTriangleStrip
:
4325 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4326 b
->shader
->info
.gs
.output_primitive
=
4327 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
4330 case SpvExecutionModeSpacingEqual
:
4331 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4332 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4333 b
->shader
->info
.tess
.spacing
= TESS_SPACING_EQUAL
;
4335 case SpvExecutionModeSpacingFractionalEven
:
4336 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4337 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4338 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_EVEN
;
4340 case SpvExecutionModeSpacingFractionalOdd
:
4341 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4342 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4343 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_ODD
;
4345 case SpvExecutionModeVertexOrderCw
:
4346 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4347 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4348 b
->shader
->info
.tess
.ccw
= false;
4350 case SpvExecutionModeVertexOrderCcw
:
4351 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4352 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4353 b
->shader
->info
.tess
.ccw
= true;
4355 case SpvExecutionModePointMode
:
4356 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4357 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4358 b
->shader
->info
.tess
.point_mode
= true;
4361 case SpvExecutionModePixelCenterInteger
:
4362 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4363 b
->shader
->info
.fs
.pixel_center_integer
= true;
4366 case SpvExecutionModeXfb
:
4367 b
->shader
->info
.has_transform_feedback_varyings
= true;
4370 case SpvExecutionModeVecTypeHint
:
4373 case SpvExecutionModeContractionOff
:
4374 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
)
4375 vtn_warn("ExectionMode only allowed for CL-style kernels: %s",
4376 spirv_executionmode_to_string(mode
->exec_mode
));
4381 case SpvExecutionModeStencilRefReplacingEXT
:
4382 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4385 case SpvExecutionModeDerivativeGroupQuadsNV
:
4386 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
4387 b
->shader
->info
.cs
.derivative_group
= DERIVATIVE_GROUP_QUADS
;
4390 case SpvExecutionModeDerivativeGroupLinearNV
:
4391 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
4392 b
->shader
->info
.cs
.derivative_group
= DERIVATIVE_GROUP_LINEAR
;
4395 case SpvExecutionModePixelInterlockOrderedEXT
:
4396 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4397 b
->shader
->info
.fs
.pixel_interlock_ordered
= true;
4400 case SpvExecutionModePixelInterlockUnorderedEXT
:
4401 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4402 b
->shader
->info
.fs
.pixel_interlock_unordered
= true;
4405 case SpvExecutionModeSampleInterlockOrderedEXT
:
4406 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4407 b
->shader
->info
.fs
.sample_interlock_ordered
= true;
4410 case SpvExecutionModeSampleInterlockUnorderedEXT
:
4411 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4412 b
->shader
->info
.fs
.sample_interlock_unordered
= true;
4415 case SpvExecutionModeDenormPreserve
:
4416 case SpvExecutionModeDenormFlushToZero
:
4417 case SpvExecutionModeSignedZeroInfNanPreserve
:
4418 case SpvExecutionModeRoundingModeRTE
:
4419 case SpvExecutionModeRoundingModeRTZ
:
4420 /* Already handled in vtn_handle_rounding_mode_in_execution_mode() */
4424 vtn_fail("Unhandled execution mode: %s (%u)",
4425 spirv_executionmode_to_string(mode
->exec_mode
),
4431 vtn_handle_rounding_mode_in_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
4432 const struct vtn_decoration
*mode
, void *data
)
4434 vtn_assert(b
->entry_point
== entry_point
);
4436 unsigned execution_mode
= 0;
4438 switch(mode
->exec_mode
) {
4439 case SpvExecutionModeDenormPreserve
:
4440 switch (mode
->operands
[0]) {
4441 case 16: execution_mode
= FLOAT_CONTROLS_DENORM_PRESERVE_FP16
; break;
4442 case 32: execution_mode
= FLOAT_CONTROLS_DENORM_PRESERVE_FP32
; break;
4443 case 64: execution_mode
= FLOAT_CONTROLS_DENORM_PRESERVE_FP64
; break;
4444 default: vtn_fail("Floating point type not supported");
4447 case SpvExecutionModeDenormFlushToZero
:
4448 switch (mode
->operands
[0]) {
4449 case 16: execution_mode
= FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16
; break;
4450 case 32: execution_mode
= FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32
; break;
4451 case 64: execution_mode
= FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP64
; break;
4452 default: vtn_fail("Floating point type not supported");
4455 case SpvExecutionModeSignedZeroInfNanPreserve
:
4456 switch (mode
->operands
[0]) {
4457 case 16: execution_mode
= FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP16
; break;
4458 case 32: execution_mode
= FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP32
; break;
4459 case 64: execution_mode
= FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP64
; break;
4460 default: vtn_fail("Floating point type not supported");
4463 case SpvExecutionModeRoundingModeRTE
:
4464 switch (mode
->operands
[0]) {
4465 case 16: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
; break;
4466 case 32: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
; break;
4467 case 64: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
; break;
4468 default: vtn_fail("Floating point type not supported");
4471 case SpvExecutionModeRoundingModeRTZ
:
4472 switch (mode
->operands
[0]) {
4473 case 16: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
; break;
4474 case 32: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
; break;
4475 case 64: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
; break;
4476 default: vtn_fail("Floating point type not supported");
4484 b
->shader
->info
.float_controls_execution_mode
|= execution_mode
;
4488 vtn_handle_variable_or_type_instruction(struct vtn_builder
*b
, SpvOp opcode
,
4489 const uint32_t *w
, unsigned count
)
4491 vtn_set_instruction_result_type(b
, opcode
, w
, count
);
4495 case SpvOpSourceContinued
:
4496 case SpvOpSourceExtension
:
4497 case SpvOpExtension
:
4498 case SpvOpCapability
:
4499 case SpvOpExtInstImport
:
4500 case SpvOpMemoryModel
:
4501 case SpvOpEntryPoint
:
4502 case SpvOpExecutionMode
:
4505 case SpvOpMemberName
:
4506 case SpvOpDecorationGroup
:
4508 case SpvOpDecorateId
:
4509 case SpvOpMemberDecorate
:
4510 case SpvOpGroupDecorate
:
4511 case SpvOpGroupMemberDecorate
:
4512 case SpvOpDecorateString
:
4513 case SpvOpMemberDecorateString
:
4514 vtn_fail("Invalid opcode types and variables section");
4520 case SpvOpTypeFloat
:
4521 case SpvOpTypeVector
:
4522 case SpvOpTypeMatrix
:
4523 case SpvOpTypeImage
:
4524 case SpvOpTypeSampler
:
4525 case SpvOpTypeSampledImage
:
4526 case SpvOpTypeArray
:
4527 case SpvOpTypeRuntimeArray
:
4528 case SpvOpTypeStruct
:
4529 case SpvOpTypeOpaque
:
4530 case SpvOpTypePointer
:
4531 case SpvOpTypeForwardPointer
:
4532 case SpvOpTypeFunction
:
4533 case SpvOpTypeEvent
:
4534 case SpvOpTypeDeviceEvent
:
4535 case SpvOpTypeReserveId
:
4536 case SpvOpTypeQueue
:
4538 vtn_handle_type(b
, opcode
, w
, count
);
4541 case SpvOpConstantTrue
:
4542 case SpvOpConstantFalse
:
4544 case SpvOpConstantComposite
:
4545 case SpvOpConstantSampler
:
4546 case SpvOpConstantNull
:
4547 case SpvOpSpecConstantTrue
:
4548 case SpvOpSpecConstantFalse
:
4549 case SpvOpSpecConstant
:
4550 case SpvOpSpecConstantComposite
:
4551 case SpvOpSpecConstantOp
:
4552 vtn_handle_constant(b
, opcode
, w
, count
);
4557 vtn_handle_variables(b
, opcode
, w
, count
);
4560 case SpvOpExtInst
: {
4561 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
4562 /* NonSemantic extended instructions are acceptable in preamble, others
4563 * will indicate the end of preamble.
4565 return val
->ext_handler
== vtn_handle_non_semantic_instruction
;
4569 return false; /* End of preamble */
4575 static struct vtn_ssa_value
*
4576 vtn_nir_select(struct vtn_builder
*b
, struct vtn_ssa_value
*src0
,
4577 struct vtn_ssa_value
*src1
, struct vtn_ssa_value
*src2
)
4579 struct vtn_ssa_value
*dest
= rzalloc(b
, struct vtn_ssa_value
);
4580 dest
->type
= src1
->type
;
4582 if (glsl_type_is_vector_or_scalar(src1
->type
)) {
4583 dest
->def
= nir_bcsel(&b
->nb
, src0
->def
, src1
->def
, src2
->def
);
4585 unsigned elems
= glsl_get_length(src1
->type
);
4587 dest
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
4588 for (unsigned i
= 0; i
< elems
; i
++) {
4589 dest
->elems
[i
] = vtn_nir_select(b
, src0
,
4590 src1
->elems
[i
], src2
->elems
[i
]);
4598 vtn_handle_select(struct vtn_builder
*b
, SpvOp opcode
,
4599 const uint32_t *w
, unsigned count
)
4601 /* Handle OpSelect up-front here because it needs to be able to handle
4602 * pointers and not just regular vectors and scalars.
4604 struct vtn_value
*res_val
= vtn_untyped_value(b
, w
[2]);
4605 struct vtn_value
*cond_val
= vtn_untyped_value(b
, w
[3]);
4606 struct vtn_value
*obj1_val
= vtn_untyped_value(b
, w
[4]);
4607 struct vtn_value
*obj2_val
= vtn_untyped_value(b
, w
[5]);
4609 vtn_fail_if(obj1_val
->type
!= res_val
->type
||
4610 obj2_val
->type
!= res_val
->type
,
4611 "Object types must match the result type in OpSelect");
4613 vtn_fail_if((cond_val
->type
->base_type
!= vtn_base_type_scalar
&&
4614 cond_val
->type
->base_type
!= vtn_base_type_vector
) ||
4615 !glsl_type_is_boolean(cond_val
->type
->type
),
4616 "OpSelect must have either a vector of booleans or "
4617 "a boolean as Condition type");
4619 vtn_fail_if(cond_val
->type
->base_type
== vtn_base_type_vector
&&
4620 (res_val
->type
->base_type
!= vtn_base_type_vector
||
4621 res_val
->type
->length
!= cond_val
->type
->length
),
4622 "When Condition type in OpSelect is a vector, the Result "
4623 "type must be a vector of the same length");
4625 switch (res_val
->type
->base_type
) {
4626 case vtn_base_type_scalar
:
4627 case vtn_base_type_vector
:
4628 case vtn_base_type_matrix
:
4629 case vtn_base_type_array
:
4630 case vtn_base_type_struct
:
4633 case vtn_base_type_pointer
:
4634 /* We need to have actual storage for pointer types. */
4635 vtn_fail_if(res_val
->type
->type
== NULL
,
4636 "Invalid pointer result type for OpSelect");
4639 vtn_fail("Result type of OpSelect must be a scalar, composite, or pointer");
4642 struct vtn_type
*res_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
4643 struct vtn_ssa_value
*ssa
= vtn_nir_select(b
,
4644 vtn_ssa_value(b
, w
[3]), vtn_ssa_value(b
, w
[4]), vtn_ssa_value(b
, w
[5]));
4646 vtn_push_ssa(b
, w
[2], res_type
, ssa
);
4650 vtn_handle_ptr(struct vtn_builder
*b
, SpvOp opcode
,
4651 const uint32_t *w
, unsigned count
)
4653 struct vtn_type
*type1
= vtn_untyped_value(b
, w
[3])->type
;
4654 struct vtn_type
*type2
= vtn_untyped_value(b
, w
[4])->type
;
4655 vtn_fail_if(type1
->base_type
!= vtn_base_type_pointer
||
4656 type2
->base_type
!= vtn_base_type_pointer
,
4657 "%s operands must have pointer types",
4658 spirv_op_to_string(opcode
));
4659 vtn_fail_if(type1
->storage_class
!= type2
->storage_class
,
4660 "%s operands must have the same storage class",
4661 spirv_op_to_string(opcode
));
4663 struct vtn_type
*vtn_type
=
4664 vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
4665 const struct glsl_type
*type
= vtn_type
->type
;
4667 nir_address_format addr_format
= vtn_mode_to_address_format(
4668 b
, vtn_storage_class_to_mode(b
, type1
->storage_class
, NULL
, NULL
));
4673 case SpvOpPtrDiff
: {
4674 /* OpPtrDiff returns the difference in number of elements (not byte offset). */
4675 unsigned elem_size
, elem_align
;
4676 glsl_get_natural_size_align_bytes(type1
->deref
->type
,
4677 &elem_size
, &elem_align
);
4679 def
= nir_build_addr_isub(&b
->nb
,
4680 vtn_ssa_value(b
, w
[3])->def
,
4681 vtn_ssa_value(b
, w
[4])->def
,
4683 def
= nir_idiv(&b
->nb
, def
, nir_imm_intN_t(&b
->nb
, elem_size
, def
->bit_size
));
4684 def
= nir_i2i(&b
->nb
, def
, glsl_get_bit_size(type
));
4689 case SpvOpPtrNotEqual
: {
4690 def
= nir_build_addr_ieq(&b
->nb
,
4691 vtn_ssa_value(b
, w
[3])->def
,
4692 vtn_ssa_value(b
, w
[4])->def
,
4694 if (opcode
== SpvOpPtrNotEqual
)
4695 def
= nir_inot(&b
->nb
, def
);
4700 unreachable("Invalid ptr operation");
4703 struct vtn_ssa_value
*ssa_value
= vtn_create_ssa_value(b
, type
);
4704 ssa_value
->def
= def
;
4705 vtn_push_ssa(b
, w
[2], vtn_type
, ssa_value
);
4709 vtn_handle_body_instruction(struct vtn_builder
*b
, SpvOp opcode
,
4710 const uint32_t *w
, unsigned count
)
4716 case SpvOpLoopMerge
:
4717 case SpvOpSelectionMerge
:
4718 /* This is handled by cfg pre-pass and walk_blocks */
4722 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
4723 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
4728 vtn_handle_extension(b
, opcode
, w
, count
);
4734 case SpvOpCopyMemory
:
4735 case SpvOpCopyMemorySized
:
4736 case SpvOpAccessChain
:
4737 case SpvOpPtrAccessChain
:
4738 case SpvOpInBoundsAccessChain
:
4739 case SpvOpInBoundsPtrAccessChain
:
4740 case SpvOpArrayLength
:
4741 case SpvOpConvertPtrToU
:
4742 case SpvOpConvertUToPtr
:
4743 vtn_handle_variables(b
, opcode
, w
, count
);
4746 case SpvOpFunctionCall
:
4747 vtn_handle_function_call(b
, opcode
, w
, count
);
4750 case SpvOpSampledImage
:
4752 case SpvOpImageSampleImplicitLod
:
4753 case SpvOpImageSampleExplicitLod
:
4754 case SpvOpImageSampleDrefImplicitLod
:
4755 case SpvOpImageSampleDrefExplicitLod
:
4756 case SpvOpImageSampleProjImplicitLod
:
4757 case SpvOpImageSampleProjExplicitLod
:
4758 case SpvOpImageSampleProjDrefImplicitLod
:
4759 case SpvOpImageSampleProjDrefExplicitLod
:
4760 case SpvOpImageFetch
:
4761 case SpvOpImageGather
:
4762 case SpvOpImageDrefGather
:
4763 case SpvOpImageQuerySizeLod
:
4764 case SpvOpImageQueryLod
:
4765 case SpvOpImageQueryLevels
:
4766 case SpvOpImageQuerySamples
:
4767 vtn_handle_texture(b
, opcode
, w
, count
);
4770 case SpvOpImageRead
:
4771 case SpvOpImageWrite
:
4772 case SpvOpImageTexelPointer
:
4773 vtn_handle_image(b
, opcode
, w
, count
);
4776 case SpvOpImageQuerySize
: {
4777 struct vtn_pointer
*image
=
4778 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
4779 if (glsl_type_is_image(image
->type
->type
)) {
4780 vtn_handle_image(b
, opcode
, w
, count
);
4782 vtn_assert(glsl_type_is_sampler(image
->type
->type
));
4783 vtn_handle_texture(b
, opcode
, w
, count
);
4788 case SpvOpAtomicLoad
:
4789 case SpvOpAtomicExchange
:
4790 case SpvOpAtomicCompareExchange
:
4791 case SpvOpAtomicCompareExchangeWeak
:
4792 case SpvOpAtomicIIncrement
:
4793 case SpvOpAtomicIDecrement
:
4794 case SpvOpAtomicIAdd
:
4795 case SpvOpAtomicISub
:
4796 case SpvOpAtomicSMin
:
4797 case SpvOpAtomicUMin
:
4798 case SpvOpAtomicSMax
:
4799 case SpvOpAtomicUMax
:
4800 case SpvOpAtomicAnd
:
4802 case SpvOpAtomicXor
: {
4803 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[3]);
4804 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
4805 vtn_handle_image(b
, opcode
, w
, count
);
4807 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
4808 vtn_handle_atomics(b
, opcode
, w
, count
);
4813 case SpvOpAtomicStore
: {
4814 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[1]);
4815 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
4816 vtn_handle_image(b
, opcode
, w
, count
);
4818 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
4819 vtn_handle_atomics(b
, opcode
, w
, count
);
4825 vtn_handle_select(b
, opcode
, w
, count
);
4833 case SpvOpConvertFToU
:
4834 case SpvOpConvertFToS
:
4835 case SpvOpConvertSToF
:
4836 case SpvOpConvertUToF
:
4840 case SpvOpQuantizeToF16
:
4841 case SpvOpPtrCastToGeneric
:
4842 case SpvOpGenericCastToPtr
:
4847 case SpvOpSignBitSet
:
4848 case SpvOpLessOrGreater
:
4850 case SpvOpUnordered
:
4865 case SpvOpVectorTimesScalar
:
4867 case SpvOpIAddCarry
:
4868 case SpvOpISubBorrow
:
4869 case SpvOpUMulExtended
:
4870 case SpvOpSMulExtended
:
4871 case SpvOpShiftRightLogical
:
4872 case SpvOpShiftRightArithmetic
:
4873 case SpvOpShiftLeftLogical
:
4874 case SpvOpLogicalEqual
:
4875 case SpvOpLogicalNotEqual
:
4876 case SpvOpLogicalOr
:
4877 case SpvOpLogicalAnd
:
4878 case SpvOpLogicalNot
:
4879 case SpvOpBitwiseOr
:
4880 case SpvOpBitwiseXor
:
4881 case SpvOpBitwiseAnd
:
4883 case SpvOpFOrdEqual
:
4884 case SpvOpFUnordEqual
:
4885 case SpvOpINotEqual
:
4886 case SpvOpFOrdNotEqual
:
4887 case SpvOpFUnordNotEqual
:
4888 case SpvOpULessThan
:
4889 case SpvOpSLessThan
:
4890 case SpvOpFOrdLessThan
:
4891 case SpvOpFUnordLessThan
:
4892 case SpvOpUGreaterThan
:
4893 case SpvOpSGreaterThan
:
4894 case SpvOpFOrdGreaterThan
:
4895 case SpvOpFUnordGreaterThan
:
4896 case SpvOpULessThanEqual
:
4897 case SpvOpSLessThanEqual
:
4898 case SpvOpFOrdLessThanEqual
:
4899 case SpvOpFUnordLessThanEqual
:
4900 case SpvOpUGreaterThanEqual
:
4901 case SpvOpSGreaterThanEqual
:
4902 case SpvOpFOrdGreaterThanEqual
:
4903 case SpvOpFUnordGreaterThanEqual
:
4909 case SpvOpFwidthFine
:
4910 case SpvOpDPdxCoarse
:
4911 case SpvOpDPdyCoarse
:
4912 case SpvOpFwidthCoarse
:
4913 case SpvOpBitFieldInsert
:
4914 case SpvOpBitFieldSExtract
:
4915 case SpvOpBitFieldUExtract
:
4916 case SpvOpBitReverse
:
4918 case SpvOpTranspose
:
4919 case SpvOpOuterProduct
:
4920 case SpvOpMatrixTimesScalar
:
4921 case SpvOpVectorTimesMatrix
:
4922 case SpvOpMatrixTimesVector
:
4923 case SpvOpMatrixTimesMatrix
:
4924 vtn_handle_alu(b
, opcode
, w
, count
);
4928 vtn_handle_bitcast(b
, w
, count
);
4931 case SpvOpVectorExtractDynamic
:
4932 case SpvOpVectorInsertDynamic
:
4933 case SpvOpVectorShuffle
:
4934 case SpvOpCompositeConstruct
:
4935 case SpvOpCompositeExtract
:
4936 case SpvOpCompositeInsert
:
4937 case SpvOpCopyLogical
:
4938 case SpvOpCopyObject
:
4939 vtn_handle_composite(b
, opcode
, w
, count
);
4942 case SpvOpEmitVertex
:
4943 case SpvOpEndPrimitive
:
4944 case SpvOpEmitStreamVertex
:
4945 case SpvOpEndStreamPrimitive
:
4946 case SpvOpControlBarrier
:
4947 case SpvOpMemoryBarrier
:
4948 vtn_handle_barrier(b
, opcode
, w
, count
);
4951 case SpvOpGroupNonUniformElect
:
4952 case SpvOpGroupNonUniformAll
:
4953 case SpvOpGroupNonUniformAny
:
4954 case SpvOpGroupNonUniformAllEqual
:
4955 case SpvOpGroupNonUniformBroadcast
:
4956 case SpvOpGroupNonUniformBroadcastFirst
:
4957 case SpvOpGroupNonUniformBallot
:
4958 case SpvOpGroupNonUniformInverseBallot
:
4959 case SpvOpGroupNonUniformBallotBitExtract
:
4960 case SpvOpGroupNonUniformBallotBitCount
:
4961 case SpvOpGroupNonUniformBallotFindLSB
:
4962 case SpvOpGroupNonUniformBallotFindMSB
:
4963 case SpvOpGroupNonUniformShuffle
:
4964 case SpvOpGroupNonUniformShuffleXor
:
4965 case SpvOpGroupNonUniformShuffleUp
:
4966 case SpvOpGroupNonUniformShuffleDown
:
4967 case SpvOpGroupNonUniformIAdd
:
4968 case SpvOpGroupNonUniformFAdd
:
4969 case SpvOpGroupNonUniformIMul
:
4970 case SpvOpGroupNonUniformFMul
:
4971 case SpvOpGroupNonUniformSMin
:
4972 case SpvOpGroupNonUniformUMin
:
4973 case SpvOpGroupNonUniformFMin
:
4974 case SpvOpGroupNonUniformSMax
:
4975 case SpvOpGroupNonUniformUMax
:
4976 case SpvOpGroupNonUniformFMax
:
4977 case SpvOpGroupNonUniformBitwiseAnd
:
4978 case SpvOpGroupNonUniformBitwiseOr
:
4979 case SpvOpGroupNonUniformBitwiseXor
:
4980 case SpvOpGroupNonUniformLogicalAnd
:
4981 case SpvOpGroupNonUniformLogicalOr
:
4982 case SpvOpGroupNonUniformLogicalXor
:
4983 case SpvOpGroupNonUniformQuadBroadcast
:
4984 case SpvOpGroupNonUniformQuadSwap
:
4987 case SpvOpGroupBroadcast
:
4988 case SpvOpGroupIAdd
:
4989 case SpvOpGroupFAdd
:
4990 case SpvOpGroupFMin
:
4991 case SpvOpGroupUMin
:
4992 case SpvOpGroupSMin
:
4993 case SpvOpGroupFMax
:
4994 case SpvOpGroupUMax
:
4995 case SpvOpGroupSMax
:
4996 case SpvOpSubgroupBallotKHR
:
4997 case SpvOpSubgroupFirstInvocationKHR
:
4998 case SpvOpSubgroupReadInvocationKHR
:
4999 case SpvOpSubgroupAllKHR
:
5000 case SpvOpSubgroupAnyKHR
:
5001 case SpvOpSubgroupAllEqualKHR
:
5002 case SpvOpGroupIAddNonUniformAMD
:
5003 case SpvOpGroupFAddNonUniformAMD
:
5004 case SpvOpGroupFMinNonUniformAMD
:
5005 case SpvOpGroupUMinNonUniformAMD
:
5006 case SpvOpGroupSMinNonUniformAMD
:
5007 case SpvOpGroupFMaxNonUniformAMD
:
5008 case SpvOpGroupUMaxNonUniformAMD
:
5009 case SpvOpGroupSMaxNonUniformAMD
:
5010 vtn_handle_subgroup(b
, opcode
, w
, count
);
5015 case SpvOpPtrNotEqual
:
5016 vtn_handle_ptr(b
, opcode
, w
, count
);
5019 case SpvOpBeginInvocationInterlockEXT
:
5020 vtn_emit_barrier(b
, nir_intrinsic_begin_invocation_interlock
);
5023 case SpvOpEndInvocationInterlockEXT
:
5024 vtn_emit_barrier(b
, nir_intrinsic_end_invocation_interlock
);
5027 case SpvOpDemoteToHelperInvocationEXT
: {
5028 nir_intrinsic_instr
*intrin
=
5029 nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_demote
);
5030 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
5034 case SpvOpIsHelperInvocationEXT
: {
5035 nir_intrinsic_instr
*intrin
=
5036 nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_is_helper_invocation
);
5037 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
, 1, 1, NULL
);
5038 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
5040 struct vtn_type
*res_type
=
5041 vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
5042 struct vtn_ssa_value
*val
= vtn_create_ssa_value(b
, res_type
->type
);
5043 val
->def
= &intrin
->dest
.ssa
;
5045 vtn_push_ssa(b
, w
[2], res_type
, val
);
5049 case SpvOpReadClockKHR
: {
5050 assert(vtn_constant_uint(b
, w
[3]) == SpvScopeSubgroup
);
5052 /* Operation supports two result types: uvec2 and uint64_t. The NIR
5053 * intrinsic gives uvec2, so pack the result for the other case.
5055 nir_intrinsic_instr
*intrin
=
5056 nir_intrinsic_instr_create(b
->nb
.shader
, nir_intrinsic_shader_clock
);
5057 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
, 2, 32, NULL
);
5058 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
5060 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
5061 const struct glsl_type
*dest_type
= type
->type
;
5062 nir_ssa_def
*result
;
5064 if (glsl_type_is_vector(dest_type
)) {
5065 assert(dest_type
== glsl_vector_type(GLSL_TYPE_UINT
, 2));
5066 result
= &intrin
->dest
.ssa
;
5068 assert(glsl_type_is_scalar(dest_type
));
5069 assert(glsl_get_base_type(dest_type
) == GLSL_TYPE_UINT64
);
5070 result
= nir_pack_64_2x32(&b
->nb
, &intrin
->dest
.ssa
);
5073 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
5075 val
->ssa
= vtn_create_ssa_value(b
, dest_type
);
5076 val
->ssa
->def
= result
;
5080 case SpvOpLifetimeStart
:
5081 case SpvOpLifetimeStop
:
5085 vtn_fail_with_opcode("Unhandled opcode", opcode
);
5092 vtn_create_builder(const uint32_t *words
, size_t word_count
,
5093 gl_shader_stage stage
, const char *entry_point_name
,
5094 const struct spirv_to_nir_options
*options
)
5096 /* Initialize the vtn_builder object */
5097 struct vtn_builder
*b
= rzalloc(NULL
, struct vtn_builder
);
5098 struct spirv_to_nir_options
*dup_options
=
5099 ralloc(b
, struct spirv_to_nir_options
);
5100 *dup_options
= *options
;
5103 b
->spirv_word_count
= word_count
;
5107 exec_list_make_empty(&b
->functions
);
5108 b
->entry_point_stage
= stage
;
5109 b
->entry_point_name
= entry_point_name
;
5110 b
->options
= dup_options
;
5113 * Handle the SPIR-V header (first 5 dwords).
5114 * Can't use vtx_assert() as the setjmp(3) target isn't initialized yet.
5116 if (word_count
<= 5)
5119 if (words
[0] != SpvMagicNumber
) {
5120 vtn_err("words[0] was 0x%x, want 0x%x", words
[0], SpvMagicNumber
);
5123 if (words
[1] < 0x10000) {
5124 vtn_err("words[1] was 0x%x, want >= 0x10000", words
[1]);
5128 uint16_t generator_id
= words
[2] >> 16;
5129 uint16_t generator_version
= words
[2];
5131 /* The first GLSLang version bump actually 1.5 years after #179 was fixed
5132 * but this should at least let us shut the workaround off for modern
5133 * versions of GLSLang.
5135 b
->wa_glslang_179
= (generator_id
== 8 && generator_version
== 1);
5137 /* In GLSLang commit 8297936dd6eb3, their handling of barrier() was fixed
5138 * to provide correct memory semantics on compute shader barrier()
5139 * commands. Prior to that, we need to fix them up ourselves. This
5140 * GLSLang fix caused them to bump to generator version 3.
5142 b
->wa_glslang_cs_barrier
= (generator_id
== 8 && generator_version
< 3);
5144 /* words[2] == generator magic */
5145 unsigned value_id_bound
= words
[3];
5146 if (words
[4] != 0) {
5147 vtn_err("words[4] was %u, want 0", words
[4]);
5151 b
->value_id_bound
= value_id_bound
;
5152 b
->values
= rzalloc_array(b
, struct vtn_value
, value_id_bound
);
5160 static nir_function
*
5161 vtn_emit_kernel_entry_point_wrapper(struct vtn_builder
*b
,
5162 nir_function
*entry_point
)
5164 vtn_assert(entry_point
== b
->entry_point
->func
->impl
->function
);
5165 vtn_fail_if(!entry_point
->name
, "entry points are required to have a name");
5166 const char *func_name
=
5167 ralloc_asprintf(b
->shader
, "__wrapped_%s", entry_point
->name
);
5169 /* we shouldn't have any inputs yet */
5170 vtn_assert(!entry_point
->shader
->num_inputs
);
5171 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_KERNEL
);
5173 nir_function
*main_entry_point
= nir_function_create(b
->shader
, func_name
);
5174 main_entry_point
->impl
= nir_function_impl_create(main_entry_point
);
5175 nir_builder_init(&b
->nb
, main_entry_point
->impl
);
5176 b
->nb
.cursor
= nir_after_cf_list(&main_entry_point
->impl
->body
);
5177 b
->func_param_idx
= 0;
5179 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, entry_point
);
5181 for (unsigned i
= 0; i
< entry_point
->num_params
; ++i
) {
5182 struct vtn_type
*param_type
= b
->entry_point
->func
->type
->params
[i
];
5184 /* consider all pointers to function memory to be parameters passed
5187 bool is_by_val
= param_type
->base_type
== vtn_base_type_pointer
&&
5188 param_type
->storage_class
== SpvStorageClassFunction
;
5190 /* input variable */
5191 nir_variable
*in_var
= rzalloc(b
->nb
.shader
, nir_variable
);
5192 in_var
->data
.mode
= nir_var_shader_in
;
5193 in_var
->data
.read_only
= true;
5194 in_var
->data
.location
= i
;
5197 in_var
->type
= param_type
->deref
->type
;
5199 in_var
->type
= param_type
->type
;
5201 nir_shader_add_variable(b
->nb
.shader
, in_var
);
5202 b
->nb
.shader
->num_inputs
++;
5204 /* we have to copy the entire variable into function memory */
5206 nir_variable
*copy_var
=
5207 nir_local_variable_create(main_entry_point
->impl
, in_var
->type
,
5209 nir_copy_var(&b
->nb
, copy_var
, in_var
);
5211 nir_src_for_ssa(&nir_build_deref_var(&b
->nb
, copy_var
)->dest
.ssa
);
5213 call
->params
[i
] = nir_src_for_ssa(nir_load_var(&b
->nb
, in_var
));
5217 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
5219 return main_entry_point
;
5223 spirv_to_nir(const uint32_t *words
, size_t word_count
,
5224 struct nir_spirv_specialization
*spec
, unsigned num_spec
,
5225 gl_shader_stage stage
, const char *entry_point_name
,
5226 const struct spirv_to_nir_options
*options
,
5227 const nir_shader_compiler_options
*nir_options
)
5230 const uint32_t *word_end
= words
+ word_count
;
5232 struct vtn_builder
*b
= vtn_create_builder(words
, word_count
,
5233 stage
, entry_point_name
,
5239 /* See also _vtn_fail() */
5240 if (setjmp(b
->fail_jump
)) {
5245 /* Skip the SPIR-V header, handled at vtn_create_builder */
5248 b
->shader
= nir_shader_create(b
, stage
, nir_options
, NULL
);
5250 /* Handle all the preamble instructions */
5251 words
= vtn_foreach_instruction(b
, words
, word_end
,
5252 vtn_handle_preamble_instruction
);
5254 if (b
->entry_point
== NULL
) {
5255 vtn_fail("Entry point not found");
5260 /* Set shader info defaults */
5261 if (stage
== MESA_SHADER_GEOMETRY
)
5262 b
->shader
->info
.gs
.invocations
= 1;
5264 /* Parse rounding mode execution modes. This has to happen earlier than
5265 * other changes in the execution modes since they can affect, for example,
5266 * the result of the floating point constants.
5268 vtn_foreach_execution_mode(b
, b
->entry_point
,
5269 vtn_handle_rounding_mode_in_execution_mode
, NULL
);
5271 b
->specializations
= spec
;
5272 b
->num_specializations
= num_spec
;
5274 /* Handle all variable, type, and constant instructions */
5275 words
= vtn_foreach_instruction(b
, words
, word_end
,
5276 vtn_handle_variable_or_type_instruction
);
5278 /* Parse execution modes */
5279 vtn_foreach_execution_mode(b
, b
->entry_point
,
5280 vtn_handle_execution_mode
, NULL
);
5282 if (b
->workgroup_size_builtin
) {
5283 vtn_assert(b
->workgroup_size_builtin
->type
->type
==
5284 glsl_vector_type(GLSL_TYPE_UINT
, 3));
5286 nir_const_value
*const_size
=
5287 b
->workgroup_size_builtin
->constant
->values
;
5289 b
->shader
->info
.cs
.local_size
[0] = const_size
[0].u32
;
5290 b
->shader
->info
.cs
.local_size
[1] = const_size
[1].u32
;
5291 b
->shader
->info
.cs
.local_size
[2] = const_size
[2].u32
;
5294 /* Set types on all vtn_values */
5295 vtn_foreach_instruction(b
, words
, word_end
, vtn_set_instruction_result_type
);
5297 vtn_build_cfg(b
, words
, word_end
);
5299 assert(b
->entry_point
->value_type
== vtn_value_type_function
);
5300 b
->entry_point
->func
->referenced
= true;
5305 foreach_list_typed(struct vtn_function
, func
, node
, &b
->functions
) {
5306 if (func
->referenced
&& !func
->emitted
) {
5307 b
->const_table
= _mesa_pointer_hash_table_create(b
);
5309 vtn_function_emit(b
, func
, vtn_handle_body_instruction
);
5315 vtn_assert(b
->entry_point
->value_type
== vtn_value_type_function
);
5316 nir_function
*entry_point
= b
->entry_point
->func
->impl
->function
;
5317 vtn_assert(entry_point
);
5319 /* post process entry_points with input params */
5320 if (entry_point
->num_params
&& b
->shader
->info
.stage
== MESA_SHADER_KERNEL
)
5321 entry_point
= vtn_emit_kernel_entry_point_wrapper(b
, entry_point
);
5323 entry_point
->is_entrypoint
= true;
5325 /* When multiple shader stages exist in the same SPIR-V module, we
5326 * generate input and output variables for every stage, in the same
5327 * NIR program. These dead variables can be invalid NIR. For example,
5328 * TCS outputs must be per-vertex arrays (or decorated 'patch'), while
5329 * VS output variables wouldn't be.
5331 * To ensure we have valid NIR, we eliminate any dead inputs and outputs
5332 * right away. In order to do so, we must lower any constant initializers
5333 * on outputs so nir_remove_dead_variables sees that they're written to.
5335 nir_lower_constant_initializers(b
->shader
, nir_var_shader_out
);
5336 nir_remove_dead_variables(b
->shader
,
5337 nir_var_shader_in
| nir_var_shader_out
);
5339 /* We sometimes generate bogus derefs that, while never used, give the
5340 * validator a bit of heartburn. Run dead code to get rid of them.
5342 nir_opt_dce(b
->shader
);
5344 /* Unparent the shader from the vtn_builder before we delete the builder */
5345 ralloc_steal(NULL
, b
->shader
);
5347 nir_shader
*shader
= b
->shader
;