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 UNUSED
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 UNUSED
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
, UNUSED
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 SpvStorageClassPhysicalStorageBuffer
:
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
, UNUSED
struct vtn_value
*val
,
1536 ASSERTED
int member
,
1537 const struct vtn_decoration
*dec
, void *data
)
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
,
1581 ASSERTED
int member
,
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 SpvStorageClassPhysicalStorageBuffer
:
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_uniform
|
2110 if (semantics
& SpvMemorySemanticsWorkgroupMemoryMask
)
2111 modes
|= nir_var_mem_shared
;
2112 if (semantics
& SpvMemorySemanticsOutputMemoryMask
) {
2113 modes
|= nir_var_shader_out
;
2116 /* No barrier to add. */
2117 if (nir_semantics
== 0 || modes
== 0)
2120 nir_scope nir_scope
;
2122 case SpvScopeDevice
:
2123 vtn_fail_if(b
->options
->caps
.vk_memory_model
&&
2124 !b
->options
->caps
.vk_memory_model_device_scope
,
2125 "If the Vulkan memory model is declared and any instruction "
2126 "uses Device scope, the VulkanMemoryModelDeviceScope "
2127 "capability must be declared.");
2128 nir_scope
= NIR_SCOPE_DEVICE
;
2131 case SpvScopeQueueFamily
:
2132 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
2133 "To use Queue Family scope, the VulkanMemoryModel capability "
2134 "must be declared.");
2135 nir_scope
= NIR_SCOPE_QUEUE_FAMILY
;
2138 case SpvScopeWorkgroup
:
2139 nir_scope
= NIR_SCOPE_WORKGROUP
;
2142 case SpvScopeSubgroup
:
2143 nir_scope
= NIR_SCOPE_SUBGROUP
;
2146 case SpvScopeInvocation
:
2147 nir_scope
= NIR_SCOPE_INVOCATION
;
2151 vtn_fail("Invalid memory scope");
2154 nir_intrinsic_instr
*intrin
=
2155 nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_scoped_memory_barrier
);
2156 nir_intrinsic_set_memory_semantics(intrin
, nir_semantics
);
2158 nir_intrinsic_set_memory_modes(intrin
, modes
);
2159 nir_intrinsic_set_memory_scope(intrin
, nir_scope
);
2160 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2163 struct vtn_ssa_value
*
2164 vtn_create_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
2166 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
2169 if (!glsl_type_is_vector_or_scalar(type
)) {
2170 unsigned elems
= glsl_get_length(type
);
2171 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
2172 for (unsigned i
= 0; i
< elems
; i
++) {
2173 const struct glsl_type
*child_type
;
2175 switch (glsl_get_base_type(type
)) {
2177 case GLSL_TYPE_UINT
:
2178 case GLSL_TYPE_INT16
:
2179 case GLSL_TYPE_UINT16
:
2180 case GLSL_TYPE_UINT8
:
2181 case GLSL_TYPE_INT8
:
2182 case GLSL_TYPE_INT64
:
2183 case GLSL_TYPE_UINT64
:
2184 case GLSL_TYPE_BOOL
:
2185 case GLSL_TYPE_FLOAT
:
2186 case GLSL_TYPE_FLOAT16
:
2187 case GLSL_TYPE_DOUBLE
:
2188 child_type
= glsl_get_column_type(type
);
2190 case GLSL_TYPE_ARRAY
:
2191 child_type
= glsl_get_array_element(type
);
2193 case GLSL_TYPE_STRUCT
:
2194 case GLSL_TYPE_INTERFACE
:
2195 child_type
= glsl_get_struct_field(type
, i
);
2198 vtn_fail("unkown base type");
2201 val
->elems
[i
] = vtn_create_ssa_value(b
, child_type
);
2209 vtn_tex_src(struct vtn_builder
*b
, unsigned index
, nir_tex_src_type type
)
2212 src
.src
= nir_src_for_ssa(vtn_ssa_value(b
, index
)->def
);
2213 src
.src_type
= type
;
2218 image_operand_arg(struct vtn_builder
*b
, const uint32_t *w
, uint32_t count
,
2219 uint32_t mask_idx
, SpvImageOperandsMask op
)
2221 static const SpvImageOperandsMask ops_with_arg
=
2222 SpvImageOperandsBiasMask
|
2223 SpvImageOperandsLodMask
|
2224 SpvImageOperandsGradMask
|
2225 SpvImageOperandsConstOffsetMask
|
2226 SpvImageOperandsOffsetMask
|
2227 SpvImageOperandsConstOffsetsMask
|
2228 SpvImageOperandsSampleMask
|
2229 SpvImageOperandsMinLodMask
|
2230 SpvImageOperandsMakeTexelAvailableMask
|
2231 SpvImageOperandsMakeTexelVisibleMask
;
2233 assert(util_bitcount(op
) == 1);
2234 assert(w
[mask_idx
] & op
);
2235 assert(op
& ops_with_arg
);
2237 uint32_t idx
= util_bitcount(w
[mask_idx
] & (op
- 1) & ops_with_arg
) + 1;
2239 /* Adjust indices for operands with two arguments. */
2240 static const SpvImageOperandsMask ops_with_two_args
=
2241 SpvImageOperandsGradMask
;
2242 idx
+= util_bitcount(w
[mask_idx
] & (op
- 1) & ops_with_two_args
);
2246 vtn_fail_if(idx
+ (op
& ops_with_two_args
? 1 : 0) >= count
,
2247 "Image op claims to have %s but does not enough "
2248 "following operands", spirv_imageoperands_to_string(op
));
2254 vtn_handle_texture(struct vtn_builder
*b
, SpvOp opcode
,
2255 const uint32_t *w
, unsigned count
)
2257 if (opcode
== SpvOpSampledImage
) {
2258 struct vtn_value
*val
=
2259 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
2260 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
2261 val
->sampled_image
->image
=
2262 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2263 val
->sampled_image
->sampler
=
2264 vtn_value(b
, w
[4], vtn_value_type_pointer
)->pointer
;
2266 } else if (opcode
== SpvOpImage
) {
2267 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[3]);
2268 if (src_val
->value_type
== vtn_value_type_sampled_image
) {
2269 vtn_push_value_pointer(b
, w
[2], src_val
->sampled_image
->image
);
2271 vtn_assert(src_val
->value_type
== vtn_value_type_pointer
);
2272 vtn_push_value_pointer(b
, w
[2], src_val
->pointer
);
2277 struct vtn_type
*ret_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2279 struct vtn_pointer
*image
= NULL
, *sampler
= NULL
;
2280 struct vtn_value
*sampled_val
= vtn_untyped_value(b
, w
[3]);
2281 if (sampled_val
->value_type
== vtn_value_type_sampled_image
) {
2282 image
= sampled_val
->sampled_image
->image
;
2283 sampler
= sampled_val
->sampled_image
->sampler
;
2285 vtn_assert(sampled_val
->value_type
== vtn_value_type_pointer
);
2286 image
= sampled_val
->pointer
;
2289 nir_deref_instr
*image_deref
= vtn_pointer_to_deref(b
, image
);
2290 nir_deref_instr
*sampler_deref
=
2291 sampler
? vtn_pointer_to_deref(b
, sampler
) : NULL
;
2293 const struct glsl_type
*image_type
= sampled_val
->type
->type
;
2294 const enum glsl_sampler_dim sampler_dim
= glsl_get_sampler_dim(image_type
);
2295 const bool is_array
= glsl_sampler_type_is_array(image_type
);
2296 nir_alu_type dest_type
= nir_type_invalid
;
2298 /* Figure out the base texture operation */
2301 case SpvOpImageSampleImplicitLod
:
2302 case SpvOpImageSampleDrefImplicitLod
:
2303 case SpvOpImageSampleProjImplicitLod
:
2304 case SpvOpImageSampleProjDrefImplicitLod
:
2305 texop
= nir_texop_tex
;
2308 case SpvOpImageSampleExplicitLod
:
2309 case SpvOpImageSampleDrefExplicitLod
:
2310 case SpvOpImageSampleProjExplicitLod
:
2311 case SpvOpImageSampleProjDrefExplicitLod
:
2312 texop
= nir_texop_txl
;
2315 case SpvOpImageFetch
:
2316 if (sampler_dim
== GLSL_SAMPLER_DIM_MS
) {
2317 texop
= nir_texop_txf_ms
;
2319 texop
= nir_texop_txf
;
2323 case SpvOpImageGather
:
2324 case SpvOpImageDrefGather
:
2325 texop
= nir_texop_tg4
;
2328 case SpvOpImageQuerySizeLod
:
2329 case SpvOpImageQuerySize
:
2330 texop
= nir_texop_txs
;
2331 dest_type
= nir_type_int
;
2334 case SpvOpImageQueryLod
:
2335 texop
= nir_texop_lod
;
2336 dest_type
= nir_type_float
;
2339 case SpvOpImageQueryLevels
:
2340 texop
= nir_texop_query_levels
;
2341 dest_type
= nir_type_int
;
2344 case SpvOpImageQuerySamples
:
2345 texop
= nir_texop_texture_samples
;
2346 dest_type
= nir_type_int
;
2349 case SpvOpFragmentFetchAMD
:
2350 texop
= nir_texop_fragment_fetch
;
2353 case SpvOpFragmentMaskFetchAMD
:
2354 texop
= nir_texop_fragment_mask_fetch
;
2358 vtn_fail_with_opcode("Unhandled opcode", opcode
);
2361 nir_tex_src srcs
[10]; /* 10 should be enough */
2362 nir_tex_src
*p
= srcs
;
2364 p
->src
= nir_src_for_ssa(&image_deref
->dest
.ssa
);
2365 p
->src_type
= nir_tex_src_texture_deref
;
2375 vtn_fail_if(sampler
== NULL
,
2376 "%s requires an image of type OpTypeSampledImage",
2377 spirv_op_to_string(opcode
));
2378 p
->src
= nir_src_for_ssa(&sampler_deref
->dest
.ssa
);
2379 p
->src_type
= nir_tex_src_sampler_deref
;
2383 case nir_texop_txf_ms
:
2385 case nir_texop_query_levels
:
2386 case nir_texop_texture_samples
:
2387 case nir_texop_samples_identical
:
2388 case nir_texop_fragment_fetch
:
2389 case nir_texop_fragment_mask_fetch
:
2392 case nir_texop_txf_ms_fb
:
2393 vtn_fail("unexpected nir_texop_txf_ms_fb");
2395 case nir_texop_txf_ms_mcs
:
2396 vtn_fail("unexpected nir_texop_txf_ms_mcs");
2397 case nir_texop_tex_prefetch
:
2398 vtn_fail("unexpected nir_texop_tex_prefetch");
2403 struct nir_ssa_def
*coord
;
2404 unsigned coord_components
;
2406 case SpvOpImageSampleImplicitLod
:
2407 case SpvOpImageSampleExplicitLod
:
2408 case SpvOpImageSampleDrefImplicitLod
:
2409 case SpvOpImageSampleDrefExplicitLod
:
2410 case SpvOpImageSampleProjImplicitLod
:
2411 case SpvOpImageSampleProjExplicitLod
:
2412 case SpvOpImageSampleProjDrefImplicitLod
:
2413 case SpvOpImageSampleProjDrefExplicitLod
:
2414 case SpvOpImageFetch
:
2415 case SpvOpImageGather
:
2416 case SpvOpImageDrefGather
:
2417 case SpvOpImageQueryLod
:
2418 case SpvOpFragmentFetchAMD
:
2419 case SpvOpFragmentMaskFetchAMD
: {
2420 /* All these types have the coordinate as their first real argument */
2421 switch (sampler_dim
) {
2422 case GLSL_SAMPLER_DIM_1D
:
2423 case GLSL_SAMPLER_DIM_BUF
:
2424 coord_components
= 1;
2426 case GLSL_SAMPLER_DIM_2D
:
2427 case GLSL_SAMPLER_DIM_RECT
:
2428 case GLSL_SAMPLER_DIM_MS
:
2429 case GLSL_SAMPLER_DIM_SUBPASS_MS
:
2430 coord_components
= 2;
2432 case GLSL_SAMPLER_DIM_3D
:
2433 case GLSL_SAMPLER_DIM_CUBE
:
2434 coord_components
= 3;
2437 vtn_fail("Invalid sampler type");
2440 if (is_array
&& texop
!= nir_texop_lod
)
2443 coord
= vtn_ssa_value(b
, w
[idx
++])->def
;
2444 p
->src
= nir_src_for_ssa(nir_channels(&b
->nb
, coord
,
2445 (1 << coord_components
) - 1));
2446 p
->src_type
= nir_tex_src_coord
;
2453 coord_components
= 0;
2458 case SpvOpImageSampleProjImplicitLod
:
2459 case SpvOpImageSampleProjExplicitLod
:
2460 case SpvOpImageSampleProjDrefImplicitLod
:
2461 case SpvOpImageSampleProjDrefExplicitLod
:
2462 /* These have the projector as the last coordinate component */
2463 p
->src
= nir_src_for_ssa(nir_channel(&b
->nb
, coord
, coord_components
));
2464 p
->src_type
= nir_tex_src_projector
;
2472 bool is_shadow
= false;
2473 unsigned gather_component
= 0;
2475 case SpvOpImageSampleDrefImplicitLod
:
2476 case SpvOpImageSampleDrefExplicitLod
:
2477 case SpvOpImageSampleProjDrefImplicitLod
:
2478 case SpvOpImageSampleProjDrefExplicitLod
:
2479 case SpvOpImageDrefGather
:
2480 /* These all have an explicit depth value as their next source */
2482 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_comparator
);
2485 case SpvOpImageGather
:
2486 /* This has a component as its next source */
2487 gather_component
= vtn_constant_uint(b
, w
[idx
++]);
2494 /* For OpImageQuerySizeLod, we always have an LOD */
2495 if (opcode
== SpvOpImageQuerySizeLod
)
2496 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
2498 /* For OpFragmentFetchAMD, we always have a multisample index */
2499 if (opcode
== SpvOpFragmentFetchAMD
)
2500 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ms_index
);
2502 /* Now we need to handle some number of optional arguments */
2503 struct vtn_value
*gather_offsets
= NULL
;
2505 uint32_t operands
= w
[idx
];
2507 if (operands
& SpvImageOperandsBiasMask
) {
2508 vtn_assert(texop
== nir_texop_tex
);
2509 texop
= nir_texop_txb
;
2510 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2511 SpvImageOperandsBiasMask
);
2512 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_bias
);
2515 if (operands
& SpvImageOperandsLodMask
) {
2516 vtn_assert(texop
== nir_texop_txl
|| texop
== nir_texop_txf
||
2517 texop
== nir_texop_txs
);
2518 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2519 SpvImageOperandsLodMask
);
2520 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_lod
);
2523 if (operands
& SpvImageOperandsGradMask
) {
2524 vtn_assert(texop
== nir_texop_txl
);
2525 texop
= nir_texop_txd
;
2526 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2527 SpvImageOperandsGradMask
);
2528 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_ddx
);
2529 (*p
++) = vtn_tex_src(b
, w
[arg
+ 1], nir_tex_src_ddy
);
2532 vtn_fail_if(util_bitcount(operands
& (SpvImageOperandsConstOffsetsMask
|
2533 SpvImageOperandsOffsetMask
|
2534 SpvImageOperandsConstOffsetMask
)) > 1,
2535 "At most one of the ConstOffset, Offset, and ConstOffsets "
2536 "image operands can be used on a given instruction.");
2538 if (operands
& SpvImageOperandsOffsetMask
) {
2539 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2540 SpvImageOperandsOffsetMask
);
2541 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_offset
);
2544 if (operands
& SpvImageOperandsConstOffsetMask
) {
2545 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2546 SpvImageOperandsConstOffsetMask
);
2547 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_offset
);
2550 if (operands
& SpvImageOperandsConstOffsetsMask
) {
2551 vtn_assert(texop
== nir_texop_tg4
);
2552 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2553 SpvImageOperandsConstOffsetsMask
);
2554 gather_offsets
= vtn_value(b
, w
[arg
], vtn_value_type_constant
);
2557 if (operands
& SpvImageOperandsSampleMask
) {
2558 vtn_assert(texop
== nir_texop_txf_ms
);
2559 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2560 SpvImageOperandsSampleMask
);
2561 texop
= nir_texop_txf_ms
;
2562 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_ms_index
);
2565 if (operands
& SpvImageOperandsMinLodMask
) {
2566 vtn_assert(texop
== nir_texop_tex
||
2567 texop
== nir_texop_txb
||
2568 texop
== nir_texop_txd
);
2569 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2570 SpvImageOperandsMinLodMask
);
2571 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_min_lod
);
2575 nir_tex_instr
*instr
= nir_tex_instr_create(b
->shader
, p
- srcs
);
2578 memcpy(instr
->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
2580 instr
->coord_components
= coord_components
;
2581 instr
->sampler_dim
= sampler_dim
;
2582 instr
->is_array
= is_array
;
2583 instr
->is_shadow
= is_shadow
;
2584 instr
->is_new_style_shadow
=
2585 is_shadow
&& glsl_get_components(ret_type
->type
) == 1;
2586 instr
->component
= gather_component
;
2588 if (image
&& (image
->access
& ACCESS_NON_UNIFORM
))
2589 instr
->texture_non_uniform
= true;
2591 if (sampler
&& (sampler
->access
& ACCESS_NON_UNIFORM
))
2592 instr
->sampler_non_uniform
= true;
2594 /* for non-query ops, get dest_type from sampler type */
2595 if (dest_type
== nir_type_invalid
) {
2596 switch (glsl_get_sampler_result_type(image_type
)) {
2597 case GLSL_TYPE_FLOAT
: dest_type
= nir_type_float
; break;
2598 case GLSL_TYPE_INT
: dest_type
= nir_type_int
; break;
2599 case GLSL_TYPE_UINT
: dest_type
= nir_type_uint
; break;
2600 case GLSL_TYPE_BOOL
: dest_type
= nir_type_bool
; break;
2602 vtn_fail("Invalid base type for sampler result");
2606 instr
->dest_type
= dest_type
;
2608 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
2609 nir_tex_instr_dest_size(instr
), 32, NULL
);
2611 vtn_assert(glsl_get_vector_elements(ret_type
->type
) ==
2612 nir_tex_instr_dest_size(instr
));
2614 if (gather_offsets
) {
2615 vtn_fail_if(gather_offsets
->type
->base_type
!= vtn_base_type_array
||
2616 gather_offsets
->type
->length
!= 4,
2617 "ConstOffsets must be an array of size four of vectors "
2618 "of two integer components");
2620 struct vtn_type
*vec_type
= gather_offsets
->type
->array_element
;
2621 vtn_fail_if(vec_type
->base_type
!= vtn_base_type_vector
||
2622 vec_type
->length
!= 2 ||
2623 !glsl_type_is_integer(vec_type
->type
),
2624 "ConstOffsets must be an array of size four of vectors "
2625 "of two integer components");
2627 unsigned bit_size
= glsl_get_bit_size(vec_type
->type
);
2628 for (uint32_t i
= 0; i
< 4; i
++) {
2629 const nir_const_value
*cvec
=
2630 gather_offsets
->constant
->elements
[i
]->values
;
2631 for (uint32_t j
= 0; j
< 2; j
++) {
2633 case 8: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i8
; break;
2634 case 16: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i16
; break;
2635 case 32: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i32
; break;
2636 case 64: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i64
; break;
2638 vtn_fail("Unsupported bit size: %u", bit_size
);
2644 struct vtn_ssa_value
*ssa
= vtn_create_ssa_value(b
, ret_type
->type
);
2645 ssa
->def
= &instr
->dest
.ssa
;
2646 vtn_push_ssa(b
, w
[2], ret_type
, ssa
);
2648 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
2652 fill_common_atomic_sources(struct vtn_builder
*b
, SpvOp opcode
,
2653 const uint32_t *w
, nir_src
*src
)
2656 case SpvOpAtomicIIncrement
:
2657 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
2660 case SpvOpAtomicIDecrement
:
2661 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
2664 case SpvOpAtomicISub
:
2666 nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
2669 case SpvOpAtomicCompareExchange
:
2670 case SpvOpAtomicCompareExchangeWeak
:
2671 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[8])->def
);
2672 src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
2675 case SpvOpAtomicExchange
:
2676 case SpvOpAtomicIAdd
:
2677 case SpvOpAtomicSMin
:
2678 case SpvOpAtomicUMin
:
2679 case SpvOpAtomicSMax
:
2680 case SpvOpAtomicUMax
:
2681 case SpvOpAtomicAnd
:
2683 case SpvOpAtomicXor
:
2684 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
2688 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
2692 static nir_ssa_def
*
2693 get_image_coord(struct vtn_builder
*b
, uint32_t value
)
2695 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, value
);
2697 /* The image_load_store intrinsics assume a 4-dim coordinate */
2698 unsigned dim
= glsl_get_vector_elements(coord
->type
);
2699 unsigned swizzle
[4];
2700 for (unsigned i
= 0; i
< 4; i
++)
2701 swizzle
[i
] = MIN2(i
, dim
- 1);
2703 return nir_swizzle(&b
->nb
, coord
->def
, swizzle
, 4);
2706 static nir_ssa_def
*
2707 expand_to_vec4(nir_builder
*b
, nir_ssa_def
*value
)
2709 if (value
->num_components
== 4)
2713 for (unsigned i
= 0; i
< 4; i
++)
2714 swiz
[i
] = i
< value
->num_components
? i
: 0;
2715 return nir_swizzle(b
, value
, swiz
, 4);
2719 vtn_handle_image(struct vtn_builder
*b
, SpvOp opcode
,
2720 const uint32_t *w
, unsigned count
)
2722 /* Just get this one out of the way */
2723 if (opcode
== SpvOpImageTexelPointer
) {
2724 struct vtn_value
*val
=
2725 vtn_push_value(b
, w
[2], vtn_value_type_image_pointer
);
2726 val
->image
= ralloc(b
, struct vtn_image_pointer
);
2728 val
->image
->image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2729 val
->image
->coord
= get_image_coord(b
, w
[4]);
2730 val
->image
->sample
= vtn_ssa_value(b
, w
[5])->def
;
2731 val
->image
->lod
= nir_imm_int(&b
->nb
, 0);
2735 struct vtn_image_pointer image
;
2736 SpvScope scope
= SpvScopeInvocation
;
2737 SpvMemorySemanticsMask semantics
= 0;
2740 case SpvOpAtomicExchange
:
2741 case SpvOpAtomicCompareExchange
:
2742 case SpvOpAtomicCompareExchangeWeak
:
2743 case SpvOpAtomicIIncrement
:
2744 case SpvOpAtomicIDecrement
:
2745 case SpvOpAtomicIAdd
:
2746 case SpvOpAtomicISub
:
2747 case SpvOpAtomicLoad
:
2748 case SpvOpAtomicSMin
:
2749 case SpvOpAtomicUMin
:
2750 case SpvOpAtomicSMax
:
2751 case SpvOpAtomicUMax
:
2752 case SpvOpAtomicAnd
:
2754 case SpvOpAtomicXor
:
2755 image
= *vtn_value(b
, w
[3], vtn_value_type_image_pointer
)->image
;
2756 scope
= vtn_constant_uint(b
, w
[4]);
2757 semantics
= vtn_constant_uint(b
, w
[5]);
2760 case SpvOpAtomicStore
:
2761 image
= *vtn_value(b
, w
[1], vtn_value_type_image_pointer
)->image
;
2762 scope
= vtn_constant_uint(b
, w
[2]);
2763 semantics
= vtn_constant_uint(b
, w
[3]);
2766 case SpvOpImageQuerySize
:
2767 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2769 image
.sample
= NULL
;
2773 case SpvOpImageRead
: {
2774 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2775 image
.coord
= get_image_coord(b
, w
[4]);
2777 const SpvImageOperandsMask operands
=
2778 count
> 5 ? w
[5] : SpvImageOperandsMaskNone
;
2780 if (operands
& SpvImageOperandsSampleMask
) {
2781 uint32_t arg
= image_operand_arg(b
, w
, count
, 5,
2782 SpvImageOperandsSampleMask
);
2783 image
.sample
= vtn_ssa_value(b
, w
[arg
])->def
;
2785 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2788 if (operands
& SpvImageOperandsMakeTexelVisibleMask
) {
2789 vtn_fail_if((operands
& SpvImageOperandsNonPrivateTexelMask
) == 0,
2790 "MakeTexelVisible requires NonPrivateTexel to also be set.");
2791 uint32_t arg
= image_operand_arg(b
, w
, count
, 5,
2792 SpvImageOperandsMakeTexelVisibleMask
);
2793 semantics
= SpvMemorySemanticsMakeVisibleMask
;
2794 scope
= vtn_constant_uint(b
, w
[arg
]);
2797 if (operands
& SpvImageOperandsLodMask
) {
2798 uint32_t arg
= image_operand_arg(b
, w
, count
, 5,
2799 SpvImageOperandsLodMask
);
2800 image
.lod
= vtn_ssa_value(b
, w
[arg
])->def
;
2802 image
.lod
= nir_imm_int(&b
->nb
, 0);
2805 /* TODO: Volatile. */
2810 case SpvOpImageWrite
: {
2811 image
.image
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
2812 image
.coord
= get_image_coord(b
, w
[2]);
2816 const SpvImageOperandsMask operands
=
2817 count
> 4 ? w
[4] : SpvImageOperandsMaskNone
;
2819 if (operands
& SpvImageOperandsSampleMask
) {
2820 uint32_t arg
= image_operand_arg(b
, w
, count
, 4,
2821 SpvImageOperandsSampleMask
);
2822 image
.sample
= vtn_ssa_value(b
, w
[arg
])->def
;
2824 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2827 if (operands
& SpvImageOperandsMakeTexelAvailableMask
) {
2828 vtn_fail_if((operands
& SpvImageOperandsNonPrivateTexelMask
) == 0,
2829 "MakeTexelAvailable requires NonPrivateTexel to also be set.");
2830 uint32_t arg
= image_operand_arg(b
, w
, count
, 4,
2831 SpvImageOperandsMakeTexelAvailableMask
);
2832 semantics
= SpvMemorySemanticsMakeAvailableMask
;
2833 scope
= vtn_constant_uint(b
, w
[arg
]);
2836 if (operands
& SpvImageOperandsLodMask
) {
2837 uint32_t arg
= image_operand_arg(b
, w
, count
, 4,
2838 SpvImageOperandsLodMask
);
2839 image
.lod
= vtn_ssa_value(b
, w
[arg
])->def
;
2841 image
.lod
= nir_imm_int(&b
->nb
, 0);
2844 /* TODO: Volatile. */
2850 vtn_fail_with_opcode("Invalid image opcode", opcode
);
2853 nir_intrinsic_op op
;
2855 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_deref_##N; break;
2856 OP(ImageQuerySize
, size
)
2858 OP(ImageWrite
, store
)
2859 OP(AtomicLoad
, load
)
2860 OP(AtomicStore
, store
)
2861 OP(AtomicExchange
, atomic_exchange
)
2862 OP(AtomicCompareExchange
, atomic_comp_swap
)
2863 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
2864 OP(AtomicIIncrement
, atomic_add
)
2865 OP(AtomicIDecrement
, atomic_add
)
2866 OP(AtomicIAdd
, atomic_add
)
2867 OP(AtomicISub
, atomic_add
)
2868 OP(AtomicSMin
, atomic_imin
)
2869 OP(AtomicUMin
, atomic_umin
)
2870 OP(AtomicSMax
, atomic_imax
)
2871 OP(AtomicUMax
, atomic_umax
)
2872 OP(AtomicAnd
, atomic_and
)
2873 OP(AtomicOr
, atomic_or
)
2874 OP(AtomicXor
, atomic_xor
)
2877 vtn_fail_with_opcode("Invalid image opcode", opcode
);
2880 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
2882 nir_deref_instr
*image_deref
= vtn_pointer_to_deref(b
, image
.image
);
2883 intrin
->src
[0] = nir_src_for_ssa(&image_deref
->dest
.ssa
);
2885 /* ImageQuerySize doesn't take any extra parameters */
2886 if (opcode
!= SpvOpImageQuerySize
) {
2887 /* The image coordinate is always 4 components but we may not have that
2888 * many. Swizzle to compensate.
2890 intrin
->src
[1] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, image
.coord
));
2891 intrin
->src
[2] = nir_src_for_ssa(image
.sample
);
2894 nir_intrinsic_set_access(intrin
, image
.image
->access
);
2897 case SpvOpAtomicLoad
:
2898 case SpvOpImageQuerySize
:
2899 case SpvOpImageRead
:
2900 if (opcode
== SpvOpImageRead
|| opcode
== SpvOpAtomicLoad
) {
2901 /* Only OpImageRead can support a lod parameter if
2902 * SPV_AMD_shader_image_load_store_lod is used but the current NIR
2903 * intrinsics definition for atomics requires us to set it for
2906 intrin
->src
[3] = nir_src_for_ssa(image
.lod
);
2909 case SpvOpAtomicStore
:
2910 case SpvOpImageWrite
: {
2911 const uint32_t value_id
= opcode
== SpvOpAtomicStore
? w
[4] : w
[3];
2912 nir_ssa_def
*value
= vtn_ssa_value(b
, value_id
)->def
;
2913 /* nir_intrinsic_image_deref_store always takes a vec4 value */
2914 assert(op
== nir_intrinsic_image_deref_store
);
2915 intrin
->num_components
= 4;
2916 intrin
->src
[3] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, value
));
2917 /* Only OpImageWrite can support a lod parameter if
2918 * SPV_AMD_shader_image_load_store_lod is used but the current NIR
2919 * intrinsics definition for atomics requires us to set it for
2922 intrin
->src
[4] = nir_src_for_ssa(image
.lod
);
2926 case SpvOpAtomicCompareExchange
:
2927 case SpvOpAtomicCompareExchangeWeak
:
2928 case SpvOpAtomicIIncrement
:
2929 case SpvOpAtomicIDecrement
:
2930 case SpvOpAtomicExchange
:
2931 case SpvOpAtomicIAdd
:
2932 case SpvOpAtomicISub
:
2933 case SpvOpAtomicSMin
:
2934 case SpvOpAtomicUMin
:
2935 case SpvOpAtomicSMax
:
2936 case SpvOpAtomicUMax
:
2937 case SpvOpAtomicAnd
:
2939 case SpvOpAtomicXor
:
2940 fill_common_atomic_sources(b
, opcode
, w
, &intrin
->src
[3]);
2944 vtn_fail_with_opcode("Invalid image opcode", opcode
);
2947 /* Image operations implicitly have the Image storage memory semantics. */
2948 semantics
|= SpvMemorySemanticsImageMemoryMask
;
2950 SpvMemorySemanticsMask before_semantics
;
2951 SpvMemorySemanticsMask after_semantics
;
2952 vtn_split_barrier_semantics(b
, semantics
, &before_semantics
, &after_semantics
);
2954 if (before_semantics
)
2955 vtn_emit_memory_barrier(b
, scope
, before_semantics
);
2957 if (opcode
!= SpvOpImageWrite
&& opcode
!= SpvOpAtomicStore
) {
2958 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2960 unsigned dest_components
= glsl_get_vector_elements(type
->type
);
2961 intrin
->num_components
= nir_intrinsic_infos
[op
].dest_components
;
2962 if (intrin
->num_components
== 0)
2963 intrin
->num_components
= dest_components
;
2965 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
,
2966 intrin
->num_components
, 32, NULL
);
2968 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2970 nir_ssa_def
*result
= &intrin
->dest
.ssa
;
2971 if (intrin
->num_components
!= dest_components
)
2972 result
= nir_channels(&b
->nb
, result
, (1 << dest_components
) - 1);
2974 struct vtn_value
*val
=
2975 vtn_push_ssa(b
, w
[2], type
, vtn_create_ssa_value(b
, type
->type
));
2976 val
->ssa
->def
= result
;
2978 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2981 if (after_semantics
)
2982 vtn_emit_memory_barrier(b
, scope
, after_semantics
);
2985 static nir_intrinsic_op
2986 get_ssbo_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2989 case SpvOpAtomicLoad
: return nir_intrinsic_load_ssbo
;
2990 case SpvOpAtomicStore
: return nir_intrinsic_store_ssbo
;
2991 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
2992 OP(AtomicExchange
, atomic_exchange
)
2993 OP(AtomicCompareExchange
, atomic_comp_swap
)
2994 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
2995 OP(AtomicIIncrement
, atomic_add
)
2996 OP(AtomicIDecrement
, atomic_add
)
2997 OP(AtomicIAdd
, atomic_add
)
2998 OP(AtomicISub
, atomic_add
)
2999 OP(AtomicSMin
, atomic_imin
)
3000 OP(AtomicUMin
, atomic_umin
)
3001 OP(AtomicSMax
, atomic_imax
)
3002 OP(AtomicUMax
, atomic_umax
)
3003 OP(AtomicAnd
, atomic_and
)
3004 OP(AtomicOr
, atomic_or
)
3005 OP(AtomicXor
, atomic_xor
)
3008 vtn_fail_with_opcode("Invalid SSBO atomic", opcode
);
3012 static nir_intrinsic_op
3013 get_uniform_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
3016 #define OP(S, N) case SpvOp##S: return nir_intrinsic_atomic_counter_ ##N;
3017 OP(AtomicLoad
, read_deref
)
3018 OP(AtomicExchange
, exchange
)
3019 OP(AtomicCompareExchange
, comp_swap
)
3020 OP(AtomicCompareExchangeWeak
, comp_swap
)
3021 OP(AtomicIIncrement
, inc_deref
)
3022 OP(AtomicIDecrement
, post_dec_deref
)
3023 OP(AtomicIAdd
, add_deref
)
3024 OP(AtomicISub
, add_deref
)
3025 OP(AtomicUMin
, min_deref
)
3026 OP(AtomicUMax
, max_deref
)
3027 OP(AtomicAnd
, and_deref
)
3028 OP(AtomicOr
, or_deref
)
3029 OP(AtomicXor
, xor_deref
)
3032 /* We left the following out: AtomicStore, AtomicSMin and
3033 * AtomicSmax. Right now there are not nir intrinsics for them. At this
3034 * moment Atomic Counter support is needed for ARB_spirv support, so is
3035 * only need to support GLSL Atomic Counters that are uints and don't
3036 * allow direct storage.
3038 vtn_fail("Invalid uniform atomic");
3042 static nir_intrinsic_op
3043 get_deref_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
3046 case SpvOpAtomicLoad
: return nir_intrinsic_load_deref
;
3047 case SpvOpAtomicStore
: return nir_intrinsic_store_deref
;
3048 #define OP(S, N) case SpvOp##S: return nir_intrinsic_deref_##N;
3049 OP(AtomicExchange
, atomic_exchange
)
3050 OP(AtomicCompareExchange
, atomic_comp_swap
)
3051 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
3052 OP(AtomicIIncrement
, atomic_add
)
3053 OP(AtomicIDecrement
, atomic_add
)
3054 OP(AtomicIAdd
, atomic_add
)
3055 OP(AtomicISub
, atomic_add
)
3056 OP(AtomicSMin
, atomic_imin
)
3057 OP(AtomicUMin
, atomic_umin
)
3058 OP(AtomicSMax
, atomic_imax
)
3059 OP(AtomicUMax
, atomic_umax
)
3060 OP(AtomicAnd
, atomic_and
)
3061 OP(AtomicOr
, atomic_or
)
3062 OP(AtomicXor
, atomic_xor
)
3065 vtn_fail_with_opcode("Invalid shared atomic", opcode
);
3070 * Handles shared atomics, ssbo atomics and atomic counters.
3073 vtn_handle_atomics(struct vtn_builder
*b
, SpvOp opcode
,
3074 const uint32_t *w
, UNUSED
unsigned count
)
3076 struct vtn_pointer
*ptr
;
3077 nir_intrinsic_instr
*atomic
;
3079 SpvScope scope
= SpvScopeInvocation
;
3080 SpvMemorySemanticsMask semantics
= 0;
3083 case SpvOpAtomicLoad
:
3084 case SpvOpAtomicExchange
:
3085 case SpvOpAtomicCompareExchange
:
3086 case SpvOpAtomicCompareExchangeWeak
:
3087 case SpvOpAtomicIIncrement
:
3088 case SpvOpAtomicIDecrement
:
3089 case SpvOpAtomicIAdd
:
3090 case SpvOpAtomicISub
:
3091 case SpvOpAtomicSMin
:
3092 case SpvOpAtomicUMin
:
3093 case SpvOpAtomicSMax
:
3094 case SpvOpAtomicUMax
:
3095 case SpvOpAtomicAnd
:
3097 case SpvOpAtomicXor
:
3098 ptr
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
3099 scope
= vtn_constant_uint(b
, w
[4]);
3100 semantics
= vtn_constant_uint(b
, w
[5]);
3103 case SpvOpAtomicStore
:
3104 ptr
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
3105 scope
= vtn_constant_uint(b
, w
[2]);
3106 semantics
= vtn_constant_uint(b
, w
[3]);
3110 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
3113 /* uniform as "atomic counter uniform" */
3114 if (ptr
->mode
== vtn_variable_mode_uniform
) {
3115 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
3116 const struct glsl_type
*deref_type
= deref
->type
;
3117 nir_intrinsic_op op
= get_uniform_nir_atomic_op(b
, opcode
);
3118 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
3119 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
3121 /* SSBO needs to initialize index/offset. In this case we don't need to,
3122 * as that info is already stored on the ptr->var->var nir_variable (see
3123 * vtn_create_variable)
3127 case SpvOpAtomicLoad
:
3128 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
3131 case SpvOpAtomicStore
:
3132 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
3133 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
3136 case SpvOpAtomicExchange
:
3137 case SpvOpAtomicCompareExchange
:
3138 case SpvOpAtomicCompareExchangeWeak
:
3139 case SpvOpAtomicIIncrement
:
3140 case SpvOpAtomicIDecrement
:
3141 case SpvOpAtomicIAdd
:
3142 case SpvOpAtomicISub
:
3143 case SpvOpAtomicSMin
:
3144 case SpvOpAtomicUMin
:
3145 case SpvOpAtomicSMax
:
3146 case SpvOpAtomicUMax
:
3147 case SpvOpAtomicAnd
:
3149 case SpvOpAtomicXor
:
3150 /* Nothing: we don't need to call fill_common_atomic_sources here, as
3151 * atomic counter uniforms doesn't have sources
3156 unreachable("Invalid SPIR-V atomic");
3159 } else if (vtn_pointer_uses_ssa_offset(b
, ptr
)) {
3160 nir_ssa_def
*offset
, *index
;
3161 offset
= vtn_pointer_to_offset(b
, ptr
, &index
);
3163 assert(ptr
->mode
== vtn_variable_mode_ssbo
);
3165 nir_intrinsic_op op
= get_ssbo_nir_atomic_op(b
, opcode
);
3166 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
3170 case SpvOpAtomicLoad
:
3171 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
3172 nir_intrinsic_set_align(atomic
, 4, 0);
3173 if (ptr
->mode
== vtn_variable_mode_ssbo
)
3174 atomic
->src
[src
++] = nir_src_for_ssa(index
);
3175 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
3178 case SpvOpAtomicStore
:
3179 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
3180 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
3181 nir_intrinsic_set_align(atomic
, 4, 0);
3182 atomic
->src
[src
++] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
3183 if (ptr
->mode
== vtn_variable_mode_ssbo
)
3184 atomic
->src
[src
++] = nir_src_for_ssa(index
);
3185 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
3188 case SpvOpAtomicExchange
:
3189 case SpvOpAtomicCompareExchange
:
3190 case SpvOpAtomicCompareExchangeWeak
:
3191 case SpvOpAtomicIIncrement
:
3192 case SpvOpAtomicIDecrement
:
3193 case SpvOpAtomicIAdd
:
3194 case SpvOpAtomicISub
:
3195 case SpvOpAtomicSMin
:
3196 case SpvOpAtomicUMin
:
3197 case SpvOpAtomicSMax
:
3198 case SpvOpAtomicUMax
:
3199 case SpvOpAtomicAnd
:
3201 case SpvOpAtomicXor
:
3202 if (ptr
->mode
== vtn_variable_mode_ssbo
)
3203 atomic
->src
[src
++] = nir_src_for_ssa(index
);
3204 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
3205 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[src
]);
3209 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
3212 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
3213 const struct glsl_type
*deref_type
= deref
->type
;
3214 nir_intrinsic_op op
= get_deref_nir_atomic_op(b
, opcode
);
3215 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
3216 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
3219 case SpvOpAtomicLoad
:
3220 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
3223 case SpvOpAtomicStore
:
3224 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
3225 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
3226 atomic
->src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
3229 case SpvOpAtomicExchange
:
3230 case SpvOpAtomicCompareExchange
:
3231 case SpvOpAtomicCompareExchangeWeak
:
3232 case SpvOpAtomicIIncrement
:
3233 case SpvOpAtomicIDecrement
:
3234 case SpvOpAtomicIAdd
:
3235 case SpvOpAtomicISub
:
3236 case SpvOpAtomicSMin
:
3237 case SpvOpAtomicUMin
:
3238 case SpvOpAtomicSMax
:
3239 case SpvOpAtomicUMax
:
3240 case SpvOpAtomicAnd
:
3242 case SpvOpAtomicXor
:
3243 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[1]);
3247 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
3251 /* Atomic ordering operations will implicitly apply to the atomic operation
3252 * storage class, so include that too.
3254 semantics
|= vtn_storage_class_to_memory_semantics(ptr
->ptr_type
->storage_class
);
3256 SpvMemorySemanticsMask before_semantics
;
3257 SpvMemorySemanticsMask after_semantics
;
3258 vtn_split_barrier_semantics(b
, semantics
, &before_semantics
, &after_semantics
);
3260 if (before_semantics
)
3261 vtn_emit_memory_barrier(b
, scope
, before_semantics
);
3263 if (opcode
!= SpvOpAtomicStore
) {
3264 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
3266 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
,
3267 glsl_get_vector_elements(type
->type
),
3268 glsl_get_bit_size(type
->type
), NULL
);
3270 struct vtn_ssa_value
*ssa
= rzalloc(b
, struct vtn_ssa_value
);
3271 ssa
->def
= &atomic
->dest
.ssa
;
3272 ssa
->type
= type
->type
;
3273 vtn_push_ssa(b
, w
[2], type
, ssa
);
3276 nir_builder_instr_insert(&b
->nb
, &atomic
->instr
);
3278 if (after_semantics
)
3279 vtn_emit_memory_barrier(b
, scope
, after_semantics
);
3282 static nir_alu_instr
*
3283 create_vec(struct vtn_builder
*b
, unsigned num_components
, unsigned bit_size
)
3285 nir_op op
= nir_op_vec(num_components
);
3286 nir_alu_instr
*vec
= nir_alu_instr_create(b
->shader
, op
);
3287 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, num_components
,
3289 vec
->dest
.write_mask
= (1 << num_components
) - 1;
3294 struct vtn_ssa_value
*
3295 vtn_ssa_transpose(struct vtn_builder
*b
, struct vtn_ssa_value
*src
)
3297 if (src
->transposed
)
3298 return src
->transposed
;
3300 struct vtn_ssa_value
*dest
=
3301 vtn_create_ssa_value(b
, glsl_transposed_type(src
->type
));
3303 for (unsigned i
= 0; i
< glsl_get_matrix_columns(dest
->type
); i
++) {
3304 nir_alu_instr
*vec
= create_vec(b
, glsl_get_matrix_columns(src
->type
),
3305 glsl_get_bit_size(src
->type
));
3306 if (glsl_type_is_vector_or_scalar(src
->type
)) {
3307 vec
->src
[0].src
= nir_src_for_ssa(src
->def
);
3308 vec
->src
[0].swizzle
[0] = i
;
3310 for (unsigned j
= 0; j
< glsl_get_matrix_columns(src
->type
); j
++) {
3311 vec
->src
[j
].src
= nir_src_for_ssa(src
->elems
[j
]->def
);
3312 vec
->src
[j
].swizzle
[0] = i
;
3315 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3316 dest
->elems
[i
]->def
= &vec
->dest
.dest
.ssa
;
3319 dest
->transposed
= src
;
3325 vtn_vector_extract(struct vtn_builder
*b
, nir_ssa_def
*src
, unsigned index
)
3327 return nir_channel(&b
->nb
, src
, index
);
3331 vtn_vector_insert(struct vtn_builder
*b
, nir_ssa_def
*src
, nir_ssa_def
*insert
,
3334 nir_alu_instr
*vec
= create_vec(b
, src
->num_components
,
3337 for (unsigned i
= 0; i
< src
->num_components
; i
++) {
3339 vec
->src
[i
].src
= nir_src_for_ssa(insert
);
3341 vec
->src
[i
].src
= nir_src_for_ssa(src
);
3342 vec
->src
[i
].swizzle
[0] = i
;
3346 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3348 return &vec
->dest
.dest
.ssa
;
3351 static nir_ssa_def
*
3352 nir_ieq_imm(nir_builder
*b
, nir_ssa_def
*x
, uint64_t i
)
3354 return nir_ieq(b
, x
, nir_imm_intN_t(b
, i
, x
->bit_size
));
3358 vtn_vector_extract_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
3361 return nir_vector_extract(&b
->nb
, src
, nir_i2i(&b
->nb
, index
, 32));
3365 vtn_vector_insert_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
3366 nir_ssa_def
*insert
, nir_ssa_def
*index
)
3368 nir_ssa_def
*dest
= vtn_vector_insert(b
, src
, insert
, 0);
3369 for (unsigned i
= 1; i
< src
->num_components
; i
++)
3370 dest
= nir_bcsel(&b
->nb
, nir_ieq_imm(&b
->nb
, index
, i
),
3371 vtn_vector_insert(b
, src
, insert
, i
), dest
);
3376 static nir_ssa_def
*
3377 vtn_vector_shuffle(struct vtn_builder
*b
, unsigned num_components
,
3378 nir_ssa_def
*src0
, nir_ssa_def
*src1
,
3379 const uint32_t *indices
)
3381 nir_alu_instr
*vec
= create_vec(b
, num_components
, src0
->bit_size
);
3383 for (unsigned i
= 0; i
< num_components
; i
++) {
3384 uint32_t index
= indices
[i
];
3385 if (index
== 0xffffffff) {
3387 nir_src_for_ssa(nir_ssa_undef(&b
->nb
, 1, src0
->bit_size
));
3388 } else if (index
< src0
->num_components
) {
3389 vec
->src
[i
].src
= nir_src_for_ssa(src0
);
3390 vec
->src
[i
].swizzle
[0] = index
;
3392 vec
->src
[i
].src
= nir_src_for_ssa(src1
);
3393 vec
->src
[i
].swizzle
[0] = index
- src0
->num_components
;
3397 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3399 return &vec
->dest
.dest
.ssa
;
3403 * Concatentates a number of vectors/scalars together to produce a vector
3405 static nir_ssa_def
*
3406 vtn_vector_construct(struct vtn_builder
*b
, unsigned num_components
,
3407 unsigned num_srcs
, nir_ssa_def
**srcs
)
3409 nir_alu_instr
*vec
= create_vec(b
, num_components
, srcs
[0]->bit_size
);
3411 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
3413 * "When constructing a vector, there must be at least two Constituent
3416 vtn_assert(num_srcs
>= 2);
3418 unsigned dest_idx
= 0;
3419 for (unsigned i
= 0; i
< num_srcs
; i
++) {
3420 nir_ssa_def
*src
= srcs
[i
];
3421 vtn_assert(dest_idx
+ src
->num_components
<= num_components
);
3422 for (unsigned j
= 0; j
< src
->num_components
; j
++) {
3423 vec
->src
[dest_idx
].src
= nir_src_for_ssa(src
);
3424 vec
->src
[dest_idx
].swizzle
[0] = j
;
3429 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
3431 * "When constructing a vector, the total number of components in all
3432 * the operands must equal the number of components in Result Type."
3434 vtn_assert(dest_idx
== num_components
);
3436 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3438 return &vec
->dest
.dest
.ssa
;
3441 static struct vtn_ssa_value
*
3442 vtn_composite_copy(void *mem_ctx
, struct vtn_ssa_value
*src
)
3444 struct vtn_ssa_value
*dest
= rzalloc(mem_ctx
, struct vtn_ssa_value
);
3445 dest
->type
= src
->type
;
3447 if (glsl_type_is_vector_or_scalar(src
->type
)) {
3448 dest
->def
= src
->def
;
3450 unsigned elems
= glsl_get_length(src
->type
);
3452 dest
->elems
= ralloc_array(mem_ctx
, struct vtn_ssa_value
*, elems
);
3453 for (unsigned i
= 0; i
< elems
; i
++)
3454 dest
->elems
[i
] = vtn_composite_copy(mem_ctx
, src
->elems
[i
]);
3460 static struct vtn_ssa_value
*
3461 vtn_composite_insert(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
3462 struct vtn_ssa_value
*insert
, const uint32_t *indices
,
3463 unsigned num_indices
)
3465 struct vtn_ssa_value
*dest
= vtn_composite_copy(b
, src
);
3467 struct vtn_ssa_value
*cur
= dest
;
3469 for (i
= 0; i
< num_indices
- 1; i
++) {
3470 cur
= cur
->elems
[indices
[i
]];
3473 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
3474 /* According to the SPIR-V spec, OpCompositeInsert may work down to
3475 * the component granularity. In that case, the last index will be
3476 * the index to insert the scalar into the vector.
3479 cur
->def
= vtn_vector_insert(b
, cur
->def
, insert
->def
, indices
[i
]);
3481 cur
->elems
[indices
[i
]] = insert
;
3487 static struct vtn_ssa_value
*
3488 vtn_composite_extract(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
3489 const uint32_t *indices
, unsigned num_indices
)
3491 struct vtn_ssa_value
*cur
= src
;
3492 for (unsigned i
= 0; i
< num_indices
; i
++) {
3493 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
3494 vtn_assert(i
== num_indices
- 1);
3495 /* According to the SPIR-V spec, OpCompositeExtract may work down to
3496 * the component granularity. The last index will be the index of the
3497 * vector to extract.
3500 struct vtn_ssa_value
*ret
= rzalloc(b
, struct vtn_ssa_value
);
3501 ret
->type
= glsl_scalar_type(glsl_get_base_type(cur
->type
));
3502 ret
->def
= vtn_vector_extract(b
, cur
->def
, indices
[i
]);
3505 cur
= cur
->elems
[indices
[i
]];
3513 vtn_handle_composite(struct vtn_builder
*b
, SpvOp opcode
,
3514 const uint32_t *w
, unsigned count
)
3516 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
3517 struct vtn_ssa_value
*ssa
= vtn_create_ssa_value(b
, type
->type
);
3520 case SpvOpVectorExtractDynamic
:
3521 ssa
->def
= vtn_vector_extract_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
3522 vtn_ssa_value(b
, w
[4])->def
);
3525 case SpvOpVectorInsertDynamic
:
3526 ssa
->def
= vtn_vector_insert_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
3527 vtn_ssa_value(b
, w
[4])->def
,
3528 vtn_ssa_value(b
, w
[5])->def
);
3531 case SpvOpVectorShuffle
:
3532 ssa
->def
= vtn_vector_shuffle(b
, glsl_get_vector_elements(type
->type
),
3533 vtn_ssa_value(b
, w
[3])->def
,
3534 vtn_ssa_value(b
, w
[4])->def
,
3538 case SpvOpCompositeConstruct
: {
3539 unsigned elems
= count
- 3;
3541 if (glsl_type_is_vector_or_scalar(type
->type
)) {
3542 nir_ssa_def
*srcs
[NIR_MAX_VEC_COMPONENTS
];
3543 for (unsigned i
= 0; i
< elems
; i
++)
3544 srcs
[i
] = vtn_ssa_value(b
, w
[3 + i
])->def
;
3546 vtn_vector_construct(b
, glsl_get_vector_elements(type
->type
),
3549 ssa
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
3550 for (unsigned i
= 0; i
< elems
; i
++)
3551 ssa
->elems
[i
] = vtn_ssa_value(b
, w
[3 + i
]);
3555 case SpvOpCompositeExtract
:
3556 ssa
= vtn_composite_extract(b
, vtn_ssa_value(b
, w
[3]),
3560 case SpvOpCompositeInsert
:
3561 ssa
= vtn_composite_insert(b
, vtn_ssa_value(b
, w
[4]),
3562 vtn_ssa_value(b
, w
[3]),
3566 case SpvOpCopyLogical
:
3567 case SpvOpCopyObject
:
3568 ssa
= vtn_composite_copy(b
, vtn_ssa_value(b
, w
[3]));
3572 vtn_fail_with_opcode("unknown composite operation", opcode
);
3575 vtn_push_ssa(b
, w
[2], type
, ssa
);
3579 vtn_emit_barrier(struct vtn_builder
*b
, nir_intrinsic_op op
)
3581 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
3582 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3586 vtn_emit_memory_barrier(struct vtn_builder
*b
, SpvScope scope
,
3587 SpvMemorySemanticsMask semantics
)
3589 if (b
->options
->use_scoped_memory_barrier
) {
3590 vtn_emit_scoped_memory_barrier(b
, scope
, semantics
);
3594 static const SpvMemorySemanticsMask all_memory_semantics
=
3595 SpvMemorySemanticsUniformMemoryMask
|
3596 SpvMemorySemanticsWorkgroupMemoryMask
|
3597 SpvMemorySemanticsAtomicCounterMemoryMask
|
3598 SpvMemorySemanticsImageMemoryMask
;
3600 /* If we're not actually doing a memory barrier, bail */
3601 if (!(semantics
& all_memory_semantics
))
3604 /* GL and Vulkan don't have these */
3605 vtn_assert(scope
!= SpvScopeCrossDevice
);
3607 if (scope
== SpvScopeSubgroup
)
3608 return; /* Nothing to do here */
3610 if (scope
== SpvScopeWorkgroup
) {
3611 vtn_emit_barrier(b
, nir_intrinsic_group_memory_barrier
);
3615 /* There's only two scopes thing left */
3616 vtn_assert(scope
== SpvScopeInvocation
|| scope
== SpvScopeDevice
);
3618 if ((semantics
& all_memory_semantics
) == all_memory_semantics
) {
3619 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier
);
3623 /* Issue a bunch of more specific barriers */
3624 uint32_t bits
= semantics
;
3626 SpvMemorySemanticsMask semantic
= 1 << u_bit_scan(&bits
);
3628 case SpvMemorySemanticsUniformMemoryMask
:
3629 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_buffer
);
3631 case SpvMemorySemanticsWorkgroupMemoryMask
:
3632 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_shared
);
3634 case SpvMemorySemanticsAtomicCounterMemoryMask
:
3635 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_atomic_counter
);
3637 case SpvMemorySemanticsImageMemoryMask
:
3638 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_image
);
3640 case SpvMemorySemanticsOutputMemoryMask
:
3641 if (b
->nb
.shader
->info
.stage
== MESA_SHADER_TESS_CTRL
)
3642 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_tcs_patch
);
3651 vtn_handle_barrier(struct vtn_builder
*b
, SpvOp opcode
,
3652 const uint32_t *w
, UNUSED
unsigned count
)
3655 case SpvOpEmitVertex
:
3656 case SpvOpEmitStreamVertex
:
3657 case SpvOpEndPrimitive
:
3658 case SpvOpEndStreamPrimitive
: {
3659 nir_intrinsic_op intrinsic_op
;
3661 case SpvOpEmitVertex
:
3662 case SpvOpEmitStreamVertex
:
3663 intrinsic_op
= nir_intrinsic_emit_vertex
;
3665 case SpvOpEndPrimitive
:
3666 case SpvOpEndStreamPrimitive
:
3667 intrinsic_op
= nir_intrinsic_end_primitive
;
3670 unreachable("Invalid opcode");
3673 nir_intrinsic_instr
*intrin
=
3674 nir_intrinsic_instr_create(b
->shader
, intrinsic_op
);
3677 case SpvOpEmitStreamVertex
:
3678 case SpvOpEndStreamPrimitive
: {
3679 unsigned stream
= vtn_constant_uint(b
, w
[1]);
3680 nir_intrinsic_set_stream_id(intrin
, stream
);
3688 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3692 case SpvOpMemoryBarrier
: {
3693 SpvScope scope
= vtn_constant_uint(b
, w
[1]);
3694 SpvMemorySemanticsMask semantics
= vtn_constant_uint(b
, w
[2]);
3695 vtn_emit_memory_barrier(b
, scope
, semantics
);
3699 case SpvOpControlBarrier
: {
3700 SpvScope execution_scope
= vtn_constant_uint(b
, w
[1]);
3701 SpvScope memory_scope
= vtn_constant_uint(b
, w
[2]);
3702 SpvMemorySemanticsMask memory_semantics
= vtn_constant_uint(b
, w
[3]);
3704 /* GLSLang, prior to commit 8297936dd6eb3, emitted OpControlBarrier with
3705 * memory semantics of None for GLSL barrier().
3707 if (b
->wa_glslang_cs_barrier
&&
3708 b
->nb
.shader
->info
.stage
== MESA_SHADER_COMPUTE
&&
3709 execution_scope
== SpvScopeWorkgroup
&&
3710 memory_semantics
== SpvMemorySemanticsMaskNone
) {
3711 memory_scope
= SpvScopeWorkgroup
;
3712 memory_semantics
= SpvMemorySemanticsAcquireReleaseMask
|
3713 SpvMemorySemanticsWorkgroupMemoryMask
;
3716 /* From the SPIR-V spec:
3718 * "When used with the TessellationControl execution model, it also
3719 * implicitly synchronizes the Output Storage Class: Writes to Output
3720 * variables performed by any invocation executed prior to a
3721 * OpControlBarrier will be visible to any other invocation after
3722 * return from that OpControlBarrier."
3724 if (b
->nb
.shader
->info
.stage
== MESA_SHADER_TESS_CTRL
) {
3725 memory_semantics
&= ~(SpvMemorySemanticsAcquireMask
|
3726 SpvMemorySemanticsReleaseMask
|
3727 SpvMemorySemanticsAcquireReleaseMask
|
3728 SpvMemorySemanticsSequentiallyConsistentMask
);
3729 memory_semantics
|= SpvMemorySemanticsAcquireReleaseMask
|
3730 SpvMemorySemanticsOutputMemoryMask
;
3733 vtn_emit_memory_barrier(b
, memory_scope
, memory_semantics
);
3735 if (execution_scope
== SpvScopeWorkgroup
)
3736 vtn_emit_barrier(b
, nir_intrinsic_control_barrier
);
3741 unreachable("unknown barrier instruction");
3746 gl_primitive_from_spv_execution_mode(struct vtn_builder
*b
,
3747 SpvExecutionMode mode
)
3750 case SpvExecutionModeInputPoints
:
3751 case SpvExecutionModeOutputPoints
:
3752 return 0; /* GL_POINTS */
3753 case SpvExecutionModeInputLines
:
3754 return 1; /* GL_LINES */
3755 case SpvExecutionModeInputLinesAdjacency
:
3756 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
3757 case SpvExecutionModeTriangles
:
3758 return 4; /* GL_TRIANGLES */
3759 case SpvExecutionModeInputTrianglesAdjacency
:
3760 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
3761 case SpvExecutionModeQuads
:
3762 return 7; /* GL_QUADS */
3763 case SpvExecutionModeIsolines
:
3764 return 0x8E7A; /* GL_ISOLINES */
3765 case SpvExecutionModeOutputLineStrip
:
3766 return 3; /* GL_LINE_STRIP */
3767 case SpvExecutionModeOutputTriangleStrip
:
3768 return 5; /* GL_TRIANGLE_STRIP */
3770 vtn_fail("Invalid primitive type: %s (%u)",
3771 spirv_executionmode_to_string(mode
), mode
);
3776 vertices_in_from_spv_execution_mode(struct vtn_builder
*b
,
3777 SpvExecutionMode mode
)
3780 case SpvExecutionModeInputPoints
:
3782 case SpvExecutionModeInputLines
:
3784 case SpvExecutionModeInputLinesAdjacency
:
3786 case SpvExecutionModeTriangles
:
3788 case SpvExecutionModeInputTrianglesAdjacency
:
3791 vtn_fail("Invalid GS input mode: %s (%u)",
3792 spirv_executionmode_to_string(mode
), mode
);
3796 static gl_shader_stage
3797 stage_for_execution_model(struct vtn_builder
*b
, SpvExecutionModel model
)
3800 case SpvExecutionModelVertex
:
3801 return MESA_SHADER_VERTEX
;
3802 case SpvExecutionModelTessellationControl
:
3803 return MESA_SHADER_TESS_CTRL
;
3804 case SpvExecutionModelTessellationEvaluation
:
3805 return MESA_SHADER_TESS_EVAL
;
3806 case SpvExecutionModelGeometry
:
3807 return MESA_SHADER_GEOMETRY
;
3808 case SpvExecutionModelFragment
:
3809 return MESA_SHADER_FRAGMENT
;
3810 case SpvExecutionModelGLCompute
:
3811 return MESA_SHADER_COMPUTE
;
3812 case SpvExecutionModelKernel
:
3813 return MESA_SHADER_KERNEL
;
3815 vtn_fail("Unsupported execution model: %s (%u)",
3816 spirv_executionmodel_to_string(model
), model
);
3820 #define spv_check_supported(name, cap) do { \
3821 if (!(b->options && b->options->caps.name)) \
3822 vtn_warn("Unsupported SPIR-V capability: %s (%u)", \
3823 spirv_capability_to_string(cap), cap); \
3828 vtn_handle_entry_point(struct vtn_builder
*b
, const uint32_t *w
,
3831 struct vtn_value
*entry_point
= &b
->values
[w
[2]];
3832 /* Let this be a name label regardless */
3833 unsigned name_words
;
3834 entry_point
->name
= vtn_string_literal(b
, &w
[3], count
- 3, &name_words
);
3836 if (strcmp(entry_point
->name
, b
->entry_point_name
) != 0 ||
3837 stage_for_execution_model(b
, w
[1]) != b
->entry_point_stage
)
3840 vtn_assert(b
->entry_point
== NULL
);
3841 b
->entry_point
= entry_point
;
3845 vtn_handle_preamble_instruction(struct vtn_builder
*b
, SpvOp opcode
,
3846 const uint32_t *w
, unsigned count
)
3853 case SpvSourceLanguageUnknown
: lang
= "unknown"; break;
3854 case SpvSourceLanguageESSL
: lang
= "ESSL"; break;
3855 case SpvSourceLanguageGLSL
: lang
= "GLSL"; break;
3856 case SpvSourceLanguageOpenCL_C
: lang
= "OpenCL C"; break;
3857 case SpvSourceLanguageOpenCL_CPP
: lang
= "OpenCL C++"; break;
3858 case SpvSourceLanguageHLSL
: lang
= "HLSL"; break;
3861 uint32_t version
= w
[2];
3864 (count
> 3) ? vtn_value(b
, w
[3], vtn_value_type_string
)->str
: "";
3866 vtn_info("Parsing SPIR-V from %s %u source file %s", lang
, version
, file
);
3870 case SpvOpSourceExtension
:
3871 case SpvOpSourceContinued
:
3872 case SpvOpExtension
:
3873 case SpvOpModuleProcessed
:
3874 /* Unhandled, but these are for debug so that's ok. */
3877 case SpvOpCapability
: {
3878 SpvCapability cap
= w
[1];
3880 case SpvCapabilityMatrix
:
3881 case SpvCapabilityShader
:
3882 case SpvCapabilityGeometry
:
3883 case SpvCapabilityGeometryPointSize
:
3884 case SpvCapabilityUniformBufferArrayDynamicIndexing
:
3885 case SpvCapabilitySampledImageArrayDynamicIndexing
:
3886 case SpvCapabilityStorageBufferArrayDynamicIndexing
:
3887 case SpvCapabilityStorageImageArrayDynamicIndexing
:
3888 case SpvCapabilityImageRect
:
3889 case SpvCapabilitySampledRect
:
3890 case SpvCapabilitySampled1D
:
3891 case SpvCapabilityImage1D
:
3892 case SpvCapabilitySampledCubeArray
:
3893 case SpvCapabilityImageCubeArray
:
3894 case SpvCapabilitySampledBuffer
:
3895 case SpvCapabilityImageBuffer
:
3896 case SpvCapabilityImageQuery
:
3897 case SpvCapabilityDerivativeControl
:
3898 case SpvCapabilityInterpolationFunction
:
3899 case SpvCapabilityMultiViewport
:
3900 case SpvCapabilitySampleRateShading
:
3901 case SpvCapabilityClipDistance
:
3902 case SpvCapabilityCullDistance
:
3903 case SpvCapabilityInputAttachment
:
3904 case SpvCapabilityImageGatherExtended
:
3905 case SpvCapabilityStorageImageExtendedFormats
:
3906 case SpvCapabilityVector16
:
3909 case SpvCapabilityLinkage
:
3910 case SpvCapabilityFloat16Buffer
:
3911 case SpvCapabilitySparseResidency
:
3912 vtn_warn("Unsupported SPIR-V capability: %s",
3913 spirv_capability_to_string(cap
));
3916 case SpvCapabilityMinLod
:
3917 spv_check_supported(min_lod
, cap
);
3920 case SpvCapabilityAtomicStorage
:
3921 spv_check_supported(atomic_storage
, cap
);
3924 case SpvCapabilityFloat64
:
3925 spv_check_supported(float64
, cap
);
3927 case SpvCapabilityInt64
:
3928 spv_check_supported(int64
, cap
);
3930 case SpvCapabilityInt16
:
3931 spv_check_supported(int16
, cap
);
3933 case SpvCapabilityInt8
:
3934 spv_check_supported(int8
, cap
);
3937 case SpvCapabilityTransformFeedback
:
3938 spv_check_supported(transform_feedback
, cap
);
3941 case SpvCapabilityGeometryStreams
:
3942 spv_check_supported(geometry_streams
, cap
);
3945 case SpvCapabilityInt64Atomics
:
3946 spv_check_supported(int64_atomics
, cap
);
3949 case SpvCapabilityStorageImageMultisample
:
3950 spv_check_supported(storage_image_ms
, cap
);
3953 case SpvCapabilityAddresses
:
3954 spv_check_supported(address
, cap
);
3957 case SpvCapabilityKernel
:
3958 spv_check_supported(kernel
, cap
);
3961 case SpvCapabilityImageBasic
:
3962 case SpvCapabilityImageReadWrite
:
3963 case SpvCapabilityImageMipmap
:
3964 case SpvCapabilityPipes
:
3965 case SpvCapabilityDeviceEnqueue
:
3966 case SpvCapabilityLiteralSampler
:
3967 case SpvCapabilityGenericPointer
:
3968 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
3969 spirv_capability_to_string(cap
));
3972 case SpvCapabilityImageMSArray
:
3973 spv_check_supported(image_ms_array
, cap
);
3976 case SpvCapabilityTessellation
:
3977 case SpvCapabilityTessellationPointSize
:
3978 spv_check_supported(tessellation
, cap
);
3981 case SpvCapabilityDrawParameters
:
3982 spv_check_supported(draw_parameters
, cap
);
3985 case SpvCapabilityStorageImageReadWithoutFormat
:
3986 spv_check_supported(image_read_without_format
, cap
);
3989 case SpvCapabilityStorageImageWriteWithoutFormat
:
3990 spv_check_supported(image_write_without_format
, cap
);
3993 case SpvCapabilityDeviceGroup
:
3994 spv_check_supported(device_group
, cap
);
3997 case SpvCapabilityMultiView
:
3998 spv_check_supported(multiview
, cap
);
4001 case SpvCapabilityGroupNonUniform
:
4002 spv_check_supported(subgroup_basic
, cap
);
4005 case SpvCapabilitySubgroupVoteKHR
:
4006 case SpvCapabilityGroupNonUniformVote
:
4007 spv_check_supported(subgroup_vote
, cap
);
4010 case SpvCapabilitySubgroupBallotKHR
:
4011 case SpvCapabilityGroupNonUniformBallot
:
4012 spv_check_supported(subgroup_ballot
, cap
);
4015 case SpvCapabilityGroupNonUniformShuffle
:
4016 case SpvCapabilityGroupNonUniformShuffleRelative
:
4017 spv_check_supported(subgroup_shuffle
, cap
);
4020 case SpvCapabilityGroupNonUniformQuad
:
4021 spv_check_supported(subgroup_quad
, cap
);
4024 case SpvCapabilityGroupNonUniformArithmetic
:
4025 case SpvCapabilityGroupNonUniformClustered
:
4026 spv_check_supported(subgroup_arithmetic
, cap
);
4029 case SpvCapabilityGroups
:
4030 spv_check_supported(amd_shader_ballot
, cap
);
4033 case SpvCapabilityVariablePointersStorageBuffer
:
4034 case SpvCapabilityVariablePointers
:
4035 spv_check_supported(variable_pointers
, cap
);
4036 b
->variable_pointers
= true;
4039 case SpvCapabilityStorageUniformBufferBlock16
:
4040 case SpvCapabilityStorageUniform16
:
4041 case SpvCapabilityStoragePushConstant16
:
4042 case SpvCapabilityStorageInputOutput16
:
4043 spv_check_supported(storage_16bit
, cap
);
4046 case SpvCapabilityShaderLayer
:
4047 case SpvCapabilityShaderViewportIndex
:
4048 case SpvCapabilityShaderViewportIndexLayerEXT
:
4049 spv_check_supported(shader_viewport_index_layer
, cap
);
4052 case SpvCapabilityStorageBuffer8BitAccess
:
4053 case SpvCapabilityUniformAndStorageBuffer8BitAccess
:
4054 case SpvCapabilityStoragePushConstant8
:
4055 spv_check_supported(storage_8bit
, cap
);
4058 case SpvCapabilityShaderNonUniformEXT
:
4059 spv_check_supported(descriptor_indexing
, cap
);
4062 case SpvCapabilityInputAttachmentArrayDynamicIndexingEXT
:
4063 case SpvCapabilityUniformTexelBufferArrayDynamicIndexingEXT
:
4064 case SpvCapabilityStorageTexelBufferArrayDynamicIndexingEXT
:
4065 spv_check_supported(descriptor_array_dynamic_indexing
, cap
);
4068 case SpvCapabilityUniformBufferArrayNonUniformIndexingEXT
:
4069 case SpvCapabilitySampledImageArrayNonUniformIndexingEXT
:
4070 case SpvCapabilityStorageBufferArrayNonUniformIndexingEXT
:
4071 case SpvCapabilityStorageImageArrayNonUniformIndexingEXT
:
4072 case SpvCapabilityInputAttachmentArrayNonUniformIndexingEXT
:
4073 case SpvCapabilityUniformTexelBufferArrayNonUniformIndexingEXT
:
4074 case SpvCapabilityStorageTexelBufferArrayNonUniformIndexingEXT
:
4075 spv_check_supported(descriptor_array_non_uniform_indexing
, cap
);
4078 case SpvCapabilityRuntimeDescriptorArrayEXT
:
4079 spv_check_supported(runtime_descriptor_array
, cap
);
4082 case SpvCapabilityStencilExportEXT
:
4083 spv_check_supported(stencil_export
, cap
);
4086 case SpvCapabilitySampleMaskPostDepthCoverage
:
4087 spv_check_supported(post_depth_coverage
, cap
);
4090 case SpvCapabilityDenormFlushToZero
:
4091 case SpvCapabilityDenormPreserve
:
4092 case SpvCapabilitySignedZeroInfNanPreserve
:
4093 case SpvCapabilityRoundingModeRTE
:
4094 case SpvCapabilityRoundingModeRTZ
:
4095 spv_check_supported(float_controls
, cap
);
4098 case SpvCapabilityPhysicalStorageBufferAddresses
:
4099 spv_check_supported(physical_storage_buffer_address
, cap
);
4102 case SpvCapabilityComputeDerivativeGroupQuadsNV
:
4103 case SpvCapabilityComputeDerivativeGroupLinearNV
:
4104 spv_check_supported(derivative_group
, cap
);
4107 case SpvCapabilityFloat16
:
4108 spv_check_supported(float16
, cap
);
4111 case SpvCapabilityFragmentShaderSampleInterlockEXT
:
4112 spv_check_supported(fragment_shader_sample_interlock
, cap
);
4115 case SpvCapabilityFragmentShaderPixelInterlockEXT
:
4116 spv_check_supported(fragment_shader_pixel_interlock
, cap
);
4119 case SpvCapabilityDemoteToHelperInvocationEXT
:
4120 spv_check_supported(demote_to_helper_invocation
, cap
);
4123 case SpvCapabilityShaderClockKHR
:
4124 spv_check_supported(shader_clock
, cap
);
4127 case SpvCapabilityVulkanMemoryModel
:
4128 spv_check_supported(vk_memory_model
, cap
);
4131 case SpvCapabilityVulkanMemoryModelDeviceScope
:
4132 spv_check_supported(vk_memory_model_device_scope
, cap
);
4135 case SpvCapabilityImageReadWriteLodAMD
:
4136 spv_check_supported(amd_image_read_write_lod
, cap
);
4139 case SpvCapabilityIntegerFunctions2INTEL
:
4140 spv_check_supported(integer_functions2
, cap
);
4143 case SpvCapabilityFragmentMaskAMD
:
4144 spv_check_supported(amd_fragment_mask
, cap
);
4148 vtn_fail("Unhandled capability: %s (%u)",
4149 spirv_capability_to_string(cap
), cap
);
4154 case SpvOpExtInstImport
:
4155 vtn_handle_extension(b
, opcode
, w
, count
);
4158 case SpvOpMemoryModel
:
4160 case SpvAddressingModelPhysical32
:
4161 vtn_fail_if(b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
,
4162 "AddressingModelPhysical32 only supported for kernels");
4163 b
->shader
->info
.cs
.ptr_size
= 32;
4164 b
->physical_ptrs
= true;
4165 b
->options
->shared_addr_format
= nir_address_format_32bit_global
;
4166 b
->options
->global_addr_format
= nir_address_format_32bit_global
;
4167 b
->options
->temp_addr_format
= nir_address_format_32bit_global
;
4169 case SpvAddressingModelPhysical64
:
4170 vtn_fail_if(b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
,
4171 "AddressingModelPhysical64 only supported for kernels");
4172 b
->shader
->info
.cs
.ptr_size
= 64;
4173 b
->physical_ptrs
= true;
4174 b
->options
->shared_addr_format
= nir_address_format_64bit_global
;
4175 b
->options
->global_addr_format
= nir_address_format_64bit_global
;
4176 b
->options
->temp_addr_format
= nir_address_format_64bit_global
;
4178 case SpvAddressingModelLogical
:
4179 vtn_fail_if(b
->shader
->info
.stage
>= MESA_SHADER_STAGES
,
4180 "AddressingModelLogical only supported for shaders");
4181 b
->physical_ptrs
= false;
4183 case SpvAddressingModelPhysicalStorageBuffer64
:
4184 vtn_fail_if(!b
->options
||
4185 !b
->options
->caps
.physical_storage_buffer_address
,
4186 "AddressingModelPhysicalStorageBuffer64 not supported");
4189 vtn_fail("Unknown addressing model: %s (%u)",
4190 spirv_addressingmodel_to_string(w
[1]), w
[1]);
4195 case SpvMemoryModelSimple
:
4196 case SpvMemoryModelGLSL450
:
4197 case SpvMemoryModelOpenCL
:
4199 case SpvMemoryModelVulkan
:
4200 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
4201 "Vulkan memory model is unsupported by this driver");
4204 vtn_fail("Unsupported memory model: %s",
4205 spirv_memorymodel_to_string(w
[2]));
4210 case SpvOpEntryPoint
:
4211 vtn_handle_entry_point(b
, w
, count
);
4215 vtn_push_value(b
, w
[1], vtn_value_type_string
)->str
=
4216 vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
4220 b
->values
[w
[1]].name
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
4223 case SpvOpMemberName
:
4227 case SpvOpExecutionMode
:
4228 case SpvOpExecutionModeId
:
4229 case SpvOpDecorationGroup
:
4231 case SpvOpDecorateId
:
4232 case SpvOpMemberDecorate
:
4233 case SpvOpGroupDecorate
:
4234 case SpvOpGroupMemberDecorate
:
4235 case SpvOpDecorateString
:
4236 case SpvOpMemberDecorateString
:
4237 vtn_handle_decoration(b
, opcode
, w
, count
);
4240 case SpvOpExtInst
: {
4241 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
4242 if (val
->ext_handler
== vtn_handle_non_semantic_instruction
) {
4243 /* NonSemantic extended instructions are acceptable in preamble. */
4244 vtn_handle_non_semantic_instruction(b
, w
[4], w
, count
);
4247 return false; /* End of preamble. */
4252 return false; /* End of preamble */
4259 vtn_handle_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
4260 const struct vtn_decoration
*mode
, UNUSED
void *data
)
4262 vtn_assert(b
->entry_point
== entry_point
);
4264 switch(mode
->exec_mode
) {
4265 case SpvExecutionModeOriginUpperLeft
:
4266 case SpvExecutionModeOriginLowerLeft
:
4267 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4268 b
->shader
->info
.fs
.origin_upper_left
=
4269 (mode
->exec_mode
== SpvExecutionModeOriginUpperLeft
);
4272 case SpvExecutionModeEarlyFragmentTests
:
4273 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4274 b
->shader
->info
.fs
.early_fragment_tests
= true;
4277 case SpvExecutionModePostDepthCoverage
:
4278 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4279 b
->shader
->info
.fs
.post_depth_coverage
= true;
4282 case SpvExecutionModeInvocations
:
4283 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4284 b
->shader
->info
.gs
.invocations
= MAX2(1, mode
->operands
[0]);
4287 case SpvExecutionModeDepthReplacing
:
4288 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4289 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_ANY
;
4291 case SpvExecutionModeDepthGreater
:
4292 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4293 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_GREATER
;
4295 case SpvExecutionModeDepthLess
:
4296 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4297 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_LESS
;
4299 case SpvExecutionModeDepthUnchanged
:
4300 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4301 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
4304 case SpvExecutionModeLocalSize
:
4305 vtn_assert(gl_shader_stage_is_compute(b
->shader
->info
.stage
));
4306 b
->shader
->info
.cs
.local_size
[0] = mode
->operands
[0];
4307 b
->shader
->info
.cs
.local_size
[1] = mode
->operands
[1];
4308 b
->shader
->info
.cs
.local_size
[2] = mode
->operands
[2];
4311 case SpvExecutionModeLocalSizeId
:
4312 b
->shader
->info
.cs
.local_size
[0] = vtn_constant_uint(b
, mode
->operands
[0]);
4313 b
->shader
->info
.cs
.local_size
[1] = vtn_constant_uint(b
, mode
->operands
[1]);
4314 b
->shader
->info
.cs
.local_size
[2] = vtn_constant_uint(b
, mode
->operands
[2]);
4317 case SpvExecutionModeLocalSizeHint
:
4318 case SpvExecutionModeLocalSizeHintId
:
4319 break; /* Nothing to do with this */
4321 case SpvExecutionModeOutputVertices
:
4322 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4323 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
4324 b
->shader
->info
.tess
.tcs_vertices_out
= mode
->operands
[0];
4326 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4327 b
->shader
->info
.gs
.vertices_out
= mode
->operands
[0];
4331 case SpvExecutionModeInputPoints
:
4332 case SpvExecutionModeInputLines
:
4333 case SpvExecutionModeInputLinesAdjacency
:
4334 case SpvExecutionModeTriangles
:
4335 case SpvExecutionModeInputTrianglesAdjacency
:
4336 case SpvExecutionModeQuads
:
4337 case SpvExecutionModeIsolines
:
4338 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4339 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
4340 b
->shader
->info
.tess
.primitive_mode
=
4341 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
4343 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4344 b
->shader
->info
.gs
.vertices_in
=
4345 vertices_in_from_spv_execution_mode(b
, mode
->exec_mode
);
4346 b
->shader
->info
.gs
.input_primitive
=
4347 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
4351 case SpvExecutionModeOutputPoints
:
4352 case SpvExecutionModeOutputLineStrip
:
4353 case SpvExecutionModeOutputTriangleStrip
:
4354 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4355 b
->shader
->info
.gs
.output_primitive
=
4356 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
4359 case SpvExecutionModeSpacingEqual
:
4360 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4361 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4362 b
->shader
->info
.tess
.spacing
= TESS_SPACING_EQUAL
;
4364 case SpvExecutionModeSpacingFractionalEven
:
4365 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4366 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4367 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_EVEN
;
4369 case SpvExecutionModeSpacingFractionalOdd
:
4370 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4371 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4372 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_ODD
;
4374 case SpvExecutionModeVertexOrderCw
:
4375 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4376 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4377 b
->shader
->info
.tess
.ccw
= false;
4379 case SpvExecutionModeVertexOrderCcw
:
4380 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4381 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4382 b
->shader
->info
.tess
.ccw
= true;
4384 case SpvExecutionModePointMode
:
4385 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4386 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4387 b
->shader
->info
.tess
.point_mode
= true;
4390 case SpvExecutionModePixelCenterInteger
:
4391 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4392 b
->shader
->info
.fs
.pixel_center_integer
= true;
4395 case SpvExecutionModeXfb
:
4396 b
->shader
->info
.has_transform_feedback_varyings
= true;
4399 case SpvExecutionModeVecTypeHint
:
4402 case SpvExecutionModeContractionOff
:
4403 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
)
4404 vtn_warn("ExectionMode only allowed for CL-style kernels: %s",
4405 spirv_executionmode_to_string(mode
->exec_mode
));
4410 case SpvExecutionModeStencilRefReplacingEXT
:
4411 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4414 case SpvExecutionModeDerivativeGroupQuadsNV
:
4415 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
4416 b
->shader
->info
.cs
.derivative_group
= DERIVATIVE_GROUP_QUADS
;
4419 case SpvExecutionModeDerivativeGroupLinearNV
:
4420 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
4421 b
->shader
->info
.cs
.derivative_group
= DERIVATIVE_GROUP_LINEAR
;
4424 case SpvExecutionModePixelInterlockOrderedEXT
:
4425 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4426 b
->shader
->info
.fs
.pixel_interlock_ordered
= true;
4429 case SpvExecutionModePixelInterlockUnorderedEXT
:
4430 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4431 b
->shader
->info
.fs
.pixel_interlock_unordered
= true;
4434 case SpvExecutionModeSampleInterlockOrderedEXT
:
4435 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4436 b
->shader
->info
.fs
.sample_interlock_ordered
= true;
4439 case SpvExecutionModeSampleInterlockUnorderedEXT
:
4440 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4441 b
->shader
->info
.fs
.sample_interlock_unordered
= true;
4444 case SpvExecutionModeDenormPreserve
:
4445 case SpvExecutionModeDenormFlushToZero
:
4446 case SpvExecutionModeSignedZeroInfNanPreserve
:
4447 case SpvExecutionModeRoundingModeRTE
:
4448 case SpvExecutionModeRoundingModeRTZ
:
4449 /* Already handled in vtn_handle_rounding_mode_in_execution_mode() */
4453 vtn_fail("Unhandled execution mode: %s (%u)",
4454 spirv_executionmode_to_string(mode
->exec_mode
),
4460 vtn_handle_rounding_mode_in_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
4461 const struct vtn_decoration
*mode
, void *data
)
4463 vtn_assert(b
->entry_point
== entry_point
);
4465 unsigned execution_mode
= 0;
4467 switch(mode
->exec_mode
) {
4468 case SpvExecutionModeDenormPreserve
:
4469 switch (mode
->operands
[0]) {
4470 case 16: execution_mode
= FLOAT_CONTROLS_DENORM_PRESERVE_FP16
; break;
4471 case 32: execution_mode
= FLOAT_CONTROLS_DENORM_PRESERVE_FP32
; break;
4472 case 64: execution_mode
= FLOAT_CONTROLS_DENORM_PRESERVE_FP64
; break;
4473 default: vtn_fail("Floating point type not supported");
4476 case SpvExecutionModeDenormFlushToZero
:
4477 switch (mode
->operands
[0]) {
4478 case 16: execution_mode
= FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16
; break;
4479 case 32: execution_mode
= FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32
; break;
4480 case 64: execution_mode
= FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP64
; break;
4481 default: vtn_fail("Floating point type not supported");
4484 case SpvExecutionModeSignedZeroInfNanPreserve
:
4485 switch (mode
->operands
[0]) {
4486 case 16: execution_mode
= FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP16
; break;
4487 case 32: execution_mode
= FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP32
; break;
4488 case 64: execution_mode
= FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP64
; break;
4489 default: vtn_fail("Floating point type not supported");
4492 case SpvExecutionModeRoundingModeRTE
:
4493 switch (mode
->operands
[0]) {
4494 case 16: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
; break;
4495 case 32: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
; break;
4496 case 64: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
; break;
4497 default: vtn_fail("Floating point type not supported");
4500 case SpvExecutionModeRoundingModeRTZ
:
4501 switch (mode
->operands
[0]) {
4502 case 16: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
; break;
4503 case 32: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
; break;
4504 case 64: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
; break;
4505 default: vtn_fail("Floating point type not supported");
4513 b
->shader
->info
.float_controls_execution_mode
|= execution_mode
;
4517 vtn_handle_variable_or_type_instruction(struct vtn_builder
*b
, SpvOp opcode
,
4518 const uint32_t *w
, unsigned count
)
4520 vtn_set_instruction_result_type(b
, opcode
, w
, count
);
4524 case SpvOpSourceContinued
:
4525 case SpvOpSourceExtension
:
4526 case SpvOpExtension
:
4527 case SpvOpCapability
:
4528 case SpvOpExtInstImport
:
4529 case SpvOpMemoryModel
:
4530 case SpvOpEntryPoint
:
4531 case SpvOpExecutionMode
:
4534 case SpvOpMemberName
:
4535 case SpvOpDecorationGroup
:
4537 case SpvOpDecorateId
:
4538 case SpvOpMemberDecorate
:
4539 case SpvOpGroupDecorate
:
4540 case SpvOpGroupMemberDecorate
:
4541 case SpvOpDecorateString
:
4542 case SpvOpMemberDecorateString
:
4543 vtn_fail("Invalid opcode types and variables section");
4549 case SpvOpTypeFloat
:
4550 case SpvOpTypeVector
:
4551 case SpvOpTypeMatrix
:
4552 case SpvOpTypeImage
:
4553 case SpvOpTypeSampler
:
4554 case SpvOpTypeSampledImage
:
4555 case SpvOpTypeArray
:
4556 case SpvOpTypeRuntimeArray
:
4557 case SpvOpTypeStruct
:
4558 case SpvOpTypeOpaque
:
4559 case SpvOpTypePointer
:
4560 case SpvOpTypeForwardPointer
:
4561 case SpvOpTypeFunction
:
4562 case SpvOpTypeEvent
:
4563 case SpvOpTypeDeviceEvent
:
4564 case SpvOpTypeReserveId
:
4565 case SpvOpTypeQueue
:
4567 vtn_handle_type(b
, opcode
, w
, count
);
4570 case SpvOpConstantTrue
:
4571 case SpvOpConstantFalse
:
4573 case SpvOpConstantComposite
:
4574 case SpvOpConstantSampler
:
4575 case SpvOpConstantNull
:
4576 case SpvOpSpecConstantTrue
:
4577 case SpvOpSpecConstantFalse
:
4578 case SpvOpSpecConstant
:
4579 case SpvOpSpecConstantComposite
:
4580 case SpvOpSpecConstantOp
:
4581 vtn_handle_constant(b
, opcode
, w
, count
);
4586 vtn_handle_variables(b
, opcode
, w
, count
);
4589 case SpvOpExtInst
: {
4590 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
4591 /* NonSemantic extended instructions are acceptable in preamble, others
4592 * will indicate the end of preamble.
4594 return val
->ext_handler
== vtn_handle_non_semantic_instruction
;
4598 return false; /* End of preamble */
4604 static struct vtn_ssa_value
*
4605 vtn_nir_select(struct vtn_builder
*b
, struct vtn_ssa_value
*src0
,
4606 struct vtn_ssa_value
*src1
, struct vtn_ssa_value
*src2
)
4608 struct vtn_ssa_value
*dest
= rzalloc(b
, struct vtn_ssa_value
);
4609 dest
->type
= src1
->type
;
4611 if (glsl_type_is_vector_or_scalar(src1
->type
)) {
4612 dest
->def
= nir_bcsel(&b
->nb
, src0
->def
, src1
->def
, src2
->def
);
4614 unsigned elems
= glsl_get_length(src1
->type
);
4616 dest
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
4617 for (unsigned i
= 0; i
< elems
; i
++) {
4618 dest
->elems
[i
] = vtn_nir_select(b
, src0
,
4619 src1
->elems
[i
], src2
->elems
[i
]);
4627 vtn_handle_select(struct vtn_builder
*b
, SpvOp opcode
,
4628 const uint32_t *w
, unsigned count
)
4630 /* Handle OpSelect up-front here because it needs to be able to handle
4631 * pointers and not just regular vectors and scalars.
4633 struct vtn_value
*res_val
= vtn_untyped_value(b
, w
[2]);
4634 struct vtn_value
*cond_val
= vtn_untyped_value(b
, w
[3]);
4635 struct vtn_value
*obj1_val
= vtn_untyped_value(b
, w
[4]);
4636 struct vtn_value
*obj2_val
= vtn_untyped_value(b
, w
[5]);
4638 vtn_fail_if(obj1_val
->type
!= res_val
->type
||
4639 obj2_val
->type
!= res_val
->type
,
4640 "Object types must match the result type in OpSelect");
4642 vtn_fail_if((cond_val
->type
->base_type
!= vtn_base_type_scalar
&&
4643 cond_val
->type
->base_type
!= vtn_base_type_vector
) ||
4644 !glsl_type_is_boolean(cond_val
->type
->type
),
4645 "OpSelect must have either a vector of booleans or "
4646 "a boolean as Condition type");
4648 vtn_fail_if(cond_val
->type
->base_type
== vtn_base_type_vector
&&
4649 (res_val
->type
->base_type
!= vtn_base_type_vector
||
4650 res_val
->type
->length
!= cond_val
->type
->length
),
4651 "When Condition type in OpSelect is a vector, the Result "
4652 "type must be a vector of the same length");
4654 switch (res_val
->type
->base_type
) {
4655 case vtn_base_type_scalar
:
4656 case vtn_base_type_vector
:
4657 case vtn_base_type_matrix
:
4658 case vtn_base_type_array
:
4659 case vtn_base_type_struct
:
4662 case vtn_base_type_pointer
:
4663 /* We need to have actual storage for pointer types. */
4664 vtn_fail_if(res_val
->type
->type
== NULL
,
4665 "Invalid pointer result type for OpSelect");
4668 vtn_fail("Result type of OpSelect must be a scalar, composite, or pointer");
4671 struct vtn_type
*res_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
4672 struct vtn_ssa_value
*ssa
= vtn_nir_select(b
,
4673 vtn_ssa_value(b
, w
[3]), vtn_ssa_value(b
, w
[4]), vtn_ssa_value(b
, w
[5]));
4675 vtn_push_ssa(b
, w
[2], res_type
, ssa
);
4679 vtn_handle_ptr(struct vtn_builder
*b
, SpvOp opcode
,
4680 const uint32_t *w
, unsigned count
)
4682 struct vtn_type
*type1
= vtn_untyped_value(b
, w
[3])->type
;
4683 struct vtn_type
*type2
= vtn_untyped_value(b
, w
[4])->type
;
4684 vtn_fail_if(type1
->base_type
!= vtn_base_type_pointer
||
4685 type2
->base_type
!= vtn_base_type_pointer
,
4686 "%s operands must have pointer types",
4687 spirv_op_to_string(opcode
));
4688 vtn_fail_if(type1
->storage_class
!= type2
->storage_class
,
4689 "%s operands must have the same storage class",
4690 spirv_op_to_string(opcode
));
4692 struct vtn_type
*vtn_type
=
4693 vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
4694 const struct glsl_type
*type
= vtn_type
->type
;
4696 nir_address_format addr_format
= vtn_mode_to_address_format(
4697 b
, vtn_storage_class_to_mode(b
, type1
->storage_class
, NULL
, NULL
));
4702 case SpvOpPtrDiff
: {
4703 /* OpPtrDiff returns the difference in number of elements (not byte offset). */
4704 unsigned elem_size
, elem_align
;
4705 glsl_get_natural_size_align_bytes(type1
->deref
->type
,
4706 &elem_size
, &elem_align
);
4708 def
= nir_build_addr_isub(&b
->nb
,
4709 vtn_ssa_value(b
, w
[3])->def
,
4710 vtn_ssa_value(b
, w
[4])->def
,
4712 def
= nir_idiv(&b
->nb
, def
, nir_imm_intN_t(&b
->nb
, elem_size
, def
->bit_size
));
4713 def
= nir_i2i(&b
->nb
, def
, glsl_get_bit_size(type
));
4718 case SpvOpPtrNotEqual
: {
4719 def
= nir_build_addr_ieq(&b
->nb
,
4720 vtn_ssa_value(b
, w
[3])->def
,
4721 vtn_ssa_value(b
, w
[4])->def
,
4723 if (opcode
== SpvOpPtrNotEqual
)
4724 def
= nir_inot(&b
->nb
, def
);
4729 unreachable("Invalid ptr operation");
4732 struct vtn_ssa_value
*ssa_value
= vtn_create_ssa_value(b
, type
);
4733 ssa_value
->def
= def
;
4734 vtn_push_ssa(b
, w
[2], vtn_type
, ssa_value
);
4738 vtn_handle_body_instruction(struct vtn_builder
*b
, SpvOp opcode
,
4739 const uint32_t *w
, unsigned count
)
4745 case SpvOpLoopMerge
:
4746 case SpvOpSelectionMerge
:
4747 /* This is handled by cfg pre-pass and walk_blocks */
4751 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
4752 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
4757 vtn_handle_extension(b
, opcode
, w
, count
);
4763 case SpvOpCopyMemory
:
4764 case SpvOpCopyMemorySized
:
4765 case SpvOpAccessChain
:
4766 case SpvOpPtrAccessChain
:
4767 case SpvOpInBoundsAccessChain
:
4768 case SpvOpInBoundsPtrAccessChain
:
4769 case SpvOpArrayLength
:
4770 case SpvOpConvertPtrToU
:
4771 case SpvOpConvertUToPtr
:
4772 vtn_handle_variables(b
, opcode
, w
, count
);
4775 case SpvOpFunctionCall
:
4776 vtn_handle_function_call(b
, opcode
, w
, count
);
4779 case SpvOpSampledImage
:
4781 case SpvOpImageSampleImplicitLod
:
4782 case SpvOpImageSampleExplicitLod
:
4783 case SpvOpImageSampleDrefImplicitLod
:
4784 case SpvOpImageSampleDrefExplicitLod
:
4785 case SpvOpImageSampleProjImplicitLod
:
4786 case SpvOpImageSampleProjExplicitLod
:
4787 case SpvOpImageSampleProjDrefImplicitLod
:
4788 case SpvOpImageSampleProjDrefExplicitLod
:
4789 case SpvOpImageFetch
:
4790 case SpvOpImageGather
:
4791 case SpvOpImageDrefGather
:
4792 case SpvOpImageQuerySizeLod
:
4793 case SpvOpImageQueryLod
:
4794 case SpvOpImageQueryLevels
:
4795 case SpvOpImageQuerySamples
:
4796 vtn_handle_texture(b
, opcode
, w
, count
);
4799 case SpvOpImageRead
:
4800 case SpvOpImageWrite
:
4801 case SpvOpImageTexelPointer
:
4802 vtn_handle_image(b
, opcode
, w
, count
);
4805 case SpvOpImageQuerySize
: {
4806 struct vtn_pointer
*image
=
4807 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
4808 if (glsl_type_is_image(image
->type
->type
)) {
4809 vtn_handle_image(b
, opcode
, w
, count
);
4811 vtn_assert(glsl_type_is_sampler(image
->type
->type
));
4812 vtn_handle_texture(b
, opcode
, w
, count
);
4817 case SpvOpFragmentMaskFetchAMD
:
4818 case SpvOpFragmentFetchAMD
:
4819 vtn_handle_texture(b
, opcode
, w
, count
);
4822 case SpvOpAtomicLoad
:
4823 case SpvOpAtomicExchange
:
4824 case SpvOpAtomicCompareExchange
:
4825 case SpvOpAtomicCompareExchangeWeak
:
4826 case SpvOpAtomicIIncrement
:
4827 case SpvOpAtomicIDecrement
:
4828 case SpvOpAtomicIAdd
:
4829 case SpvOpAtomicISub
:
4830 case SpvOpAtomicSMin
:
4831 case SpvOpAtomicUMin
:
4832 case SpvOpAtomicSMax
:
4833 case SpvOpAtomicUMax
:
4834 case SpvOpAtomicAnd
:
4836 case SpvOpAtomicXor
: {
4837 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[3]);
4838 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
4839 vtn_handle_image(b
, opcode
, w
, count
);
4841 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
4842 vtn_handle_atomics(b
, opcode
, w
, count
);
4847 case SpvOpAtomicStore
: {
4848 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[1]);
4849 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
4850 vtn_handle_image(b
, opcode
, w
, count
);
4852 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
4853 vtn_handle_atomics(b
, opcode
, w
, count
);
4859 vtn_handle_select(b
, opcode
, w
, count
);
4867 case SpvOpConvertFToU
:
4868 case SpvOpConvertFToS
:
4869 case SpvOpConvertSToF
:
4870 case SpvOpConvertUToF
:
4874 case SpvOpQuantizeToF16
:
4875 case SpvOpPtrCastToGeneric
:
4876 case SpvOpGenericCastToPtr
:
4881 case SpvOpSignBitSet
:
4882 case SpvOpLessOrGreater
:
4884 case SpvOpUnordered
:
4899 case SpvOpVectorTimesScalar
:
4901 case SpvOpIAddCarry
:
4902 case SpvOpISubBorrow
:
4903 case SpvOpUMulExtended
:
4904 case SpvOpSMulExtended
:
4905 case SpvOpShiftRightLogical
:
4906 case SpvOpShiftRightArithmetic
:
4907 case SpvOpShiftLeftLogical
:
4908 case SpvOpLogicalEqual
:
4909 case SpvOpLogicalNotEqual
:
4910 case SpvOpLogicalOr
:
4911 case SpvOpLogicalAnd
:
4912 case SpvOpLogicalNot
:
4913 case SpvOpBitwiseOr
:
4914 case SpvOpBitwiseXor
:
4915 case SpvOpBitwiseAnd
:
4917 case SpvOpFOrdEqual
:
4918 case SpvOpFUnordEqual
:
4919 case SpvOpINotEqual
:
4920 case SpvOpFOrdNotEqual
:
4921 case SpvOpFUnordNotEqual
:
4922 case SpvOpULessThan
:
4923 case SpvOpSLessThan
:
4924 case SpvOpFOrdLessThan
:
4925 case SpvOpFUnordLessThan
:
4926 case SpvOpUGreaterThan
:
4927 case SpvOpSGreaterThan
:
4928 case SpvOpFOrdGreaterThan
:
4929 case SpvOpFUnordGreaterThan
:
4930 case SpvOpULessThanEqual
:
4931 case SpvOpSLessThanEqual
:
4932 case SpvOpFOrdLessThanEqual
:
4933 case SpvOpFUnordLessThanEqual
:
4934 case SpvOpUGreaterThanEqual
:
4935 case SpvOpSGreaterThanEqual
:
4936 case SpvOpFOrdGreaterThanEqual
:
4937 case SpvOpFUnordGreaterThanEqual
:
4943 case SpvOpFwidthFine
:
4944 case SpvOpDPdxCoarse
:
4945 case SpvOpDPdyCoarse
:
4946 case SpvOpFwidthCoarse
:
4947 case SpvOpBitFieldInsert
:
4948 case SpvOpBitFieldSExtract
:
4949 case SpvOpBitFieldUExtract
:
4950 case SpvOpBitReverse
:
4952 case SpvOpTranspose
:
4953 case SpvOpOuterProduct
:
4954 case SpvOpMatrixTimesScalar
:
4955 case SpvOpVectorTimesMatrix
:
4956 case SpvOpMatrixTimesVector
:
4957 case SpvOpMatrixTimesMatrix
:
4958 case SpvOpUCountLeadingZerosINTEL
:
4959 case SpvOpUCountTrailingZerosINTEL
:
4960 case SpvOpAbsISubINTEL
:
4961 case SpvOpAbsUSubINTEL
:
4962 case SpvOpIAddSatINTEL
:
4963 case SpvOpUAddSatINTEL
:
4964 case SpvOpIAverageINTEL
:
4965 case SpvOpUAverageINTEL
:
4966 case SpvOpIAverageRoundedINTEL
:
4967 case SpvOpUAverageRoundedINTEL
:
4968 case SpvOpISubSatINTEL
:
4969 case SpvOpUSubSatINTEL
:
4970 case SpvOpIMul32x16INTEL
:
4971 case SpvOpUMul32x16INTEL
:
4972 vtn_handle_alu(b
, opcode
, w
, count
);
4976 vtn_handle_bitcast(b
, w
, count
);
4979 case SpvOpVectorExtractDynamic
:
4980 case SpvOpVectorInsertDynamic
:
4981 case SpvOpVectorShuffle
:
4982 case SpvOpCompositeConstruct
:
4983 case SpvOpCompositeExtract
:
4984 case SpvOpCompositeInsert
:
4985 case SpvOpCopyLogical
:
4986 case SpvOpCopyObject
:
4987 vtn_handle_composite(b
, opcode
, w
, count
);
4990 case SpvOpEmitVertex
:
4991 case SpvOpEndPrimitive
:
4992 case SpvOpEmitStreamVertex
:
4993 case SpvOpEndStreamPrimitive
:
4994 case SpvOpControlBarrier
:
4995 case SpvOpMemoryBarrier
:
4996 vtn_handle_barrier(b
, opcode
, w
, count
);
4999 case SpvOpGroupNonUniformElect
:
5000 case SpvOpGroupNonUniformAll
:
5001 case SpvOpGroupNonUniformAny
:
5002 case SpvOpGroupNonUniformAllEqual
:
5003 case SpvOpGroupNonUniformBroadcast
:
5004 case SpvOpGroupNonUniformBroadcastFirst
:
5005 case SpvOpGroupNonUniformBallot
:
5006 case SpvOpGroupNonUniformInverseBallot
:
5007 case SpvOpGroupNonUniformBallotBitExtract
:
5008 case SpvOpGroupNonUniformBallotBitCount
:
5009 case SpvOpGroupNonUniformBallotFindLSB
:
5010 case SpvOpGroupNonUniformBallotFindMSB
:
5011 case SpvOpGroupNonUniformShuffle
:
5012 case SpvOpGroupNonUniformShuffleXor
:
5013 case SpvOpGroupNonUniformShuffleUp
:
5014 case SpvOpGroupNonUniformShuffleDown
:
5015 case SpvOpGroupNonUniformIAdd
:
5016 case SpvOpGroupNonUniformFAdd
:
5017 case SpvOpGroupNonUniformIMul
:
5018 case SpvOpGroupNonUniformFMul
:
5019 case SpvOpGroupNonUniformSMin
:
5020 case SpvOpGroupNonUniformUMin
:
5021 case SpvOpGroupNonUniformFMin
:
5022 case SpvOpGroupNonUniformSMax
:
5023 case SpvOpGroupNonUniformUMax
:
5024 case SpvOpGroupNonUniformFMax
:
5025 case SpvOpGroupNonUniformBitwiseAnd
:
5026 case SpvOpGroupNonUniformBitwiseOr
:
5027 case SpvOpGroupNonUniformBitwiseXor
:
5028 case SpvOpGroupNonUniformLogicalAnd
:
5029 case SpvOpGroupNonUniformLogicalOr
:
5030 case SpvOpGroupNonUniformLogicalXor
:
5031 case SpvOpGroupNonUniformQuadBroadcast
:
5032 case SpvOpGroupNonUniformQuadSwap
:
5035 case SpvOpGroupBroadcast
:
5036 case SpvOpGroupIAdd
:
5037 case SpvOpGroupFAdd
:
5038 case SpvOpGroupFMin
:
5039 case SpvOpGroupUMin
:
5040 case SpvOpGroupSMin
:
5041 case SpvOpGroupFMax
:
5042 case SpvOpGroupUMax
:
5043 case SpvOpGroupSMax
:
5044 case SpvOpSubgroupBallotKHR
:
5045 case SpvOpSubgroupFirstInvocationKHR
:
5046 case SpvOpSubgroupReadInvocationKHR
:
5047 case SpvOpSubgroupAllKHR
:
5048 case SpvOpSubgroupAnyKHR
:
5049 case SpvOpSubgroupAllEqualKHR
:
5050 case SpvOpGroupIAddNonUniformAMD
:
5051 case SpvOpGroupFAddNonUniformAMD
:
5052 case SpvOpGroupFMinNonUniformAMD
:
5053 case SpvOpGroupUMinNonUniformAMD
:
5054 case SpvOpGroupSMinNonUniformAMD
:
5055 case SpvOpGroupFMaxNonUniformAMD
:
5056 case SpvOpGroupUMaxNonUniformAMD
:
5057 case SpvOpGroupSMaxNonUniformAMD
:
5058 vtn_handle_subgroup(b
, opcode
, w
, count
);
5063 case SpvOpPtrNotEqual
:
5064 vtn_handle_ptr(b
, opcode
, w
, count
);
5067 case SpvOpBeginInvocationInterlockEXT
:
5068 vtn_emit_barrier(b
, nir_intrinsic_begin_invocation_interlock
);
5071 case SpvOpEndInvocationInterlockEXT
:
5072 vtn_emit_barrier(b
, nir_intrinsic_end_invocation_interlock
);
5075 case SpvOpDemoteToHelperInvocationEXT
: {
5076 nir_intrinsic_instr
*intrin
=
5077 nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_demote
);
5078 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
5082 case SpvOpIsHelperInvocationEXT
: {
5083 nir_intrinsic_instr
*intrin
=
5084 nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_is_helper_invocation
);
5085 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
, 1, 1, NULL
);
5086 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
5088 struct vtn_type
*res_type
=
5089 vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
5090 struct vtn_ssa_value
*val
= vtn_create_ssa_value(b
, res_type
->type
);
5091 val
->def
= &intrin
->dest
.ssa
;
5093 vtn_push_ssa(b
, w
[2], res_type
, val
);
5097 case SpvOpReadClockKHR
: {
5098 assert(vtn_constant_uint(b
, w
[3]) == SpvScopeSubgroup
);
5100 /* Operation supports two result types: uvec2 and uint64_t. The NIR
5101 * intrinsic gives uvec2, so pack the result for the other case.
5103 nir_intrinsic_instr
*intrin
=
5104 nir_intrinsic_instr_create(b
->nb
.shader
, nir_intrinsic_shader_clock
);
5105 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
, 2, 32, NULL
);
5106 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
5108 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
5109 const struct glsl_type
*dest_type
= type
->type
;
5110 nir_ssa_def
*result
;
5112 if (glsl_type_is_vector(dest_type
)) {
5113 assert(dest_type
== glsl_vector_type(GLSL_TYPE_UINT
, 2));
5114 result
= &intrin
->dest
.ssa
;
5116 assert(glsl_type_is_scalar(dest_type
));
5117 assert(glsl_get_base_type(dest_type
) == GLSL_TYPE_UINT64
);
5118 result
= nir_pack_64_2x32(&b
->nb
, &intrin
->dest
.ssa
);
5121 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
5123 val
->ssa
= vtn_create_ssa_value(b
, dest_type
);
5124 val
->ssa
->def
= result
;
5128 case SpvOpLifetimeStart
:
5129 case SpvOpLifetimeStop
:
5133 vtn_fail_with_opcode("Unhandled opcode", opcode
);
5140 vtn_create_builder(const uint32_t *words
, size_t word_count
,
5141 gl_shader_stage stage
, const char *entry_point_name
,
5142 const struct spirv_to_nir_options
*options
)
5144 /* Initialize the vtn_builder object */
5145 struct vtn_builder
*b
= rzalloc(NULL
, struct vtn_builder
);
5146 struct spirv_to_nir_options
*dup_options
=
5147 ralloc(b
, struct spirv_to_nir_options
);
5148 *dup_options
= *options
;
5151 b
->spirv_word_count
= word_count
;
5155 exec_list_make_empty(&b
->functions
);
5156 b
->entry_point_stage
= stage
;
5157 b
->entry_point_name
= entry_point_name
;
5158 b
->options
= dup_options
;
5161 * Handle the SPIR-V header (first 5 dwords).
5162 * Can't use vtx_assert() as the setjmp(3) target isn't initialized yet.
5164 if (word_count
<= 5)
5167 if (words
[0] != SpvMagicNumber
) {
5168 vtn_err("words[0] was 0x%x, want 0x%x", words
[0], SpvMagicNumber
);
5171 if (words
[1] < 0x10000) {
5172 vtn_err("words[1] was 0x%x, want >= 0x10000", words
[1]);
5176 uint16_t generator_id
= words
[2] >> 16;
5177 uint16_t generator_version
= words
[2];
5179 /* The first GLSLang version bump actually 1.5 years after #179 was fixed
5180 * but this should at least let us shut the workaround off for modern
5181 * versions of GLSLang.
5183 b
->wa_glslang_179
= (generator_id
== 8 && generator_version
== 1);
5185 /* In GLSLang commit 8297936dd6eb3, their handling of barrier() was fixed
5186 * to provide correct memory semantics on compute shader barrier()
5187 * commands. Prior to that, we need to fix them up ourselves. This
5188 * GLSLang fix caused them to bump to generator version 3.
5190 b
->wa_glslang_cs_barrier
= (generator_id
== 8 && generator_version
< 3);
5192 /* words[2] == generator magic */
5193 unsigned value_id_bound
= words
[3];
5194 if (words
[4] != 0) {
5195 vtn_err("words[4] was %u, want 0", words
[4]);
5199 b
->value_id_bound
= value_id_bound
;
5200 b
->values
= rzalloc_array(b
, struct vtn_value
, value_id_bound
);
5208 static nir_function
*
5209 vtn_emit_kernel_entry_point_wrapper(struct vtn_builder
*b
,
5210 nir_function
*entry_point
)
5212 vtn_assert(entry_point
== b
->entry_point
->func
->impl
->function
);
5213 vtn_fail_if(!entry_point
->name
, "entry points are required to have a name");
5214 const char *func_name
=
5215 ralloc_asprintf(b
->shader
, "__wrapped_%s", entry_point
->name
);
5217 /* we shouldn't have any inputs yet */
5218 vtn_assert(!entry_point
->shader
->num_inputs
);
5219 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_KERNEL
);
5221 nir_function
*main_entry_point
= nir_function_create(b
->shader
, func_name
);
5222 main_entry_point
->impl
= nir_function_impl_create(main_entry_point
);
5223 nir_builder_init(&b
->nb
, main_entry_point
->impl
);
5224 b
->nb
.cursor
= nir_after_cf_list(&main_entry_point
->impl
->body
);
5225 b
->func_param_idx
= 0;
5227 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, entry_point
);
5229 for (unsigned i
= 0; i
< entry_point
->num_params
; ++i
) {
5230 struct vtn_type
*param_type
= b
->entry_point
->func
->type
->params
[i
];
5232 /* consider all pointers to function memory to be parameters passed
5235 bool is_by_val
= param_type
->base_type
== vtn_base_type_pointer
&&
5236 param_type
->storage_class
== SpvStorageClassFunction
;
5238 /* input variable */
5239 nir_variable
*in_var
= rzalloc(b
->nb
.shader
, nir_variable
);
5240 in_var
->data
.mode
= nir_var_shader_in
;
5241 in_var
->data
.read_only
= true;
5242 in_var
->data
.location
= i
;
5245 in_var
->type
= param_type
->deref
->type
;
5247 in_var
->type
= param_type
->type
;
5249 nir_shader_add_variable(b
->nb
.shader
, in_var
);
5250 b
->nb
.shader
->num_inputs
++;
5252 /* we have to copy the entire variable into function memory */
5254 nir_variable
*copy_var
=
5255 nir_local_variable_create(main_entry_point
->impl
, in_var
->type
,
5257 nir_copy_var(&b
->nb
, copy_var
, in_var
);
5259 nir_src_for_ssa(&nir_build_deref_var(&b
->nb
, copy_var
)->dest
.ssa
);
5261 call
->params
[i
] = nir_src_for_ssa(nir_load_var(&b
->nb
, in_var
));
5265 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
5267 return main_entry_point
;
5271 spirv_to_nir(const uint32_t *words
, size_t word_count
,
5272 struct nir_spirv_specialization
*spec
, unsigned num_spec
,
5273 gl_shader_stage stage
, const char *entry_point_name
,
5274 const struct spirv_to_nir_options
*options
,
5275 const nir_shader_compiler_options
*nir_options
)
5278 const uint32_t *word_end
= words
+ word_count
;
5280 struct vtn_builder
*b
= vtn_create_builder(words
, word_count
,
5281 stage
, entry_point_name
,
5287 /* See also _vtn_fail() */
5288 if (setjmp(b
->fail_jump
)) {
5293 /* Skip the SPIR-V header, handled at vtn_create_builder */
5296 b
->shader
= nir_shader_create(b
, stage
, nir_options
, NULL
);
5298 /* Handle all the preamble instructions */
5299 words
= vtn_foreach_instruction(b
, words
, word_end
,
5300 vtn_handle_preamble_instruction
);
5302 if (b
->entry_point
== NULL
) {
5303 vtn_fail("Entry point not found");
5308 /* Set shader info defaults */
5309 if (stage
== MESA_SHADER_GEOMETRY
)
5310 b
->shader
->info
.gs
.invocations
= 1;
5312 /* Parse rounding mode execution modes. This has to happen earlier than
5313 * other changes in the execution modes since they can affect, for example,
5314 * the result of the floating point constants.
5316 vtn_foreach_execution_mode(b
, b
->entry_point
,
5317 vtn_handle_rounding_mode_in_execution_mode
, NULL
);
5319 b
->specializations
= spec
;
5320 b
->num_specializations
= num_spec
;
5322 /* Handle all variable, type, and constant instructions */
5323 words
= vtn_foreach_instruction(b
, words
, word_end
,
5324 vtn_handle_variable_or_type_instruction
);
5326 /* Parse execution modes */
5327 vtn_foreach_execution_mode(b
, b
->entry_point
,
5328 vtn_handle_execution_mode
, NULL
);
5330 if (b
->workgroup_size_builtin
) {
5331 vtn_assert(b
->workgroup_size_builtin
->type
->type
==
5332 glsl_vector_type(GLSL_TYPE_UINT
, 3));
5334 nir_const_value
*const_size
=
5335 b
->workgroup_size_builtin
->constant
->values
;
5337 b
->shader
->info
.cs
.local_size
[0] = const_size
[0].u32
;
5338 b
->shader
->info
.cs
.local_size
[1] = const_size
[1].u32
;
5339 b
->shader
->info
.cs
.local_size
[2] = const_size
[2].u32
;
5342 /* Set types on all vtn_values */
5343 vtn_foreach_instruction(b
, words
, word_end
, vtn_set_instruction_result_type
);
5345 vtn_build_cfg(b
, words
, word_end
);
5347 assert(b
->entry_point
->value_type
== vtn_value_type_function
);
5348 b
->entry_point
->func
->referenced
= true;
5353 foreach_list_typed(struct vtn_function
, func
, node
, &b
->functions
) {
5354 if (func
->referenced
&& !func
->emitted
) {
5355 b
->const_table
= _mesa_pointer_hash_table_create(b
);
5357 vtn_function_emit(b
, func
, vtn_handle_body_instruction
);
5363 vtn_assert(b
->entry_point
->value_type
== vtn_value_type_function
);
5364 nir_function
*entry_point
= b
->entry_point
->func
->impl
->function
;
5365 vtn_assert(entry_point
);
5367 /* post process entry_points with input params */
5368 if (entry_point
->num_params
&& b
->shader
->info
.stage
== MESA_SHADER_KERNEL
)
5369 entry_point
= vtn_emit_kernel_entry_point_wrapper(b
, entry_point
);
5371 entry_point
->is_entrypoint
= true;
5373 /* When multiple shader stages exist in the same SPIR-V module, we
5374 * generate input and output variables for every stage, in the same
5375 * NIR program. These dead variables can be invalid NIR. For example,
5376 * TCS outputs must be per-vertex arrays (or decorated 'patch'), while
5377 * VS output variables wouldn't be.
5379 * To ensure we have valid NIR, we eliminate any dead inputs and outputs
5380 * right away. In order to do so, we must lower any constant initializers
5381 * on outputs so nir_remove_dead_variables sees that they're written to.
5383 nir_lower_constant_initializers(b
->shader
, nir_var_shader_out
);
5384 nir_remove_dead_variables(b
->shader
,
5385 nir_var_shader_in
| nir_var_shader_out
);
5387 /* We sometimes generate bogus derefs that, while never used, give the
5388 * validator a bit of heartburn. Run dead code to get rid of them.
5390 nir_opt_dce(b
->shader
);
5392 /* Unparent the shader from the vtn_builder before we delete the builder */
5393 ralloc_steal(NULL
, b
->shader
);
5395 nir_shader
*shader
= b
->shader
;