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/format/u_format.h"
36 #include "util/u_math.h"
41 vtn_log(struct vtn_builder
*b
, enum nir_spirv_debug_level level
,
42 size_t spirv_offset
, const char *message
)
44 if (b
->options
->debug
.func
) {
45 b
->options
->debug
.func(b
->options
->debug
.private_data
,
46 level
, spirv_offset
, message
);
50 if (level
>= NIR_SPIRV_DEBUG_LEVEL_WARNING
)
51 fprintf(stderr
, "%s\n", message
);
56 vtn_logf(struct vtn_builder
*b
, enum nir_spirv_debug_level level
,
57 size_t spirv_offset
, const char *fmt
, ...)
63 msg
= ralloc_vasprintf(NULL
, fmt
, args
);
66 vtn_log(b
, level
, spirv_offset
, msg
);
72 vtn_log_err(struct vtn_builder
*b
,
73 enum nir_spirv_debug_level level
, const char *prefix
,
74 const char *file
, unsigned line
,
75 const char *fmt
, va_list args
)
79 msg
= ralloc_strdup(NULL
, prefix
);
82 ralloc_asprintf_append(&msg
, " In file %s:%u\n", file
, line
);
85 ralloc_asprintf_append(&msg
, " ");
87 ralloc_vasprintf_append(&msg
, fmt
, args
);
89 ralloc_asprintf_append(&msg
, "\n %zu bytes into the SPIR-V binary",
93 ralloc_asprintf_append(&msg
,
94 "\n in SPIR-V source file %s, line %d, col %d",
95 b
->file
, b
->line
, b
->col
);
98 vtn_log(b
, level
, b
->spirv_offset
, msg
);
104 vtn_dump_shader(struct vtn_builder
*b
, const char *path
, const char *prefix
)
109 int len
= snprintf(filename
, sizeof(filename
), "%s/%s-%d.spirv",
110 path
, prefix
, idx
++);
111 if (len
< 0 || len
>= sizeof(filename
))
114 FILE *f
= fopen(filename
, "w");
118 fwrite(b
->spirv
, sizeof(*b
->spirv
), b
->spirv_word_count
, f
);
121 vtn_info("SPIR-V shader dumped to %s", filename
);
125 _vtn_warn(struct vtn_builder
*b
, const char *file
, unsigned line
,
126 const char *fmt
, ...)
131 vtn_log_err(b
, NIR_SPIRV_DEBUG_LEVEL_WARNING
, "SPIR-V WARNING:\n",
132 file
, line
, fmt
, args
);
137 _vtn_err(struct vtn_builder
*b
, const char *file
, unsigned line
,
138 const char *fmt
, ...)
143 vtn_log_err(b
, NIR_SPIRV_DEBUG_LEVEL_ERROR
, "SPIR-V ERROR:\n",
144 file
, line
, fmt
, args
);
149 _vtn_fail(struct vtn_builder
*b
, const char *file
, unsigned line
,
150 const char *fmt
, ...)
155 vtn_log_err(b
, NIR_SPIRV_DEBUG_LEVEL_ERROR
, "SPIR-V parsing FAILED:\n",
156 file
, line
, fmt
, args
);
159 const char *dump_path
= getenv("MESA_SPIRV_FAIL_DUMP_PATH");
161 vtn_dump_shader(b
, dump_path
, "fail");
163 longjmp(b
->fail_jump
, 1);
166 struct spec_constant_value
{
174 static struct vtn_ssa_value
*
175 vtn_undef_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
177 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
180 if (glsl_type_is_vector_or_scalar(type
)) {
181 unsigned num_components
= glsl_get_vector_elements(val
->type
);
182 unsigned bit_size
= glsl_get_bit_size(val
->type
);
183 val
->def
= nir_ssa_undef(&b
->nb
, num_components
, bit_size
);
185 unsigned elems
= glsl_get_length(val
->type
);
186 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
187 if (glsl_type_is_matrix(type
)) {
188 const struct glsl_type
*elem_type
=
189 glsl_vector_type(glsl_get_base_type(type
),
190 glsl_get_vector_elements(type
));
192 for (unsigned i
= 0; i
< elems
; i
++)
193 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
194 } else if (glsl_type_is_array(type
)) {
195 const struct glsl_type
*elem_type
= glsl_get_array_element(type
);
196 for (unsigned i
= 0; i
< elems
; i
++)
197 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
199 for (unsigned i
= 0; i
< elems
; i
++) {
200 const struct glsl_type
*elem_type
= glsl_get_struct_field(type
, i
);
201 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
209 static struct vtn_ssa_value
*
210 vtn_const_ssa_value(struct vtn_builder
*b
, nir_constant
*constant
,
211 const struct glsl_type
*type
)
213 struct hash_entry
*entry
= _mesa_hash_table_search(b
->const_table
, constant
);
218 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
221 switch (glsl_get_base_type(type
)) {
224 case GLSL_TYPE_INT16
:
225 case GLSL_TYPE_UINT16
:
226 case GLSL_TYPE_UINT8
:
228 case GLSL_TYPE_INT64
:
229 case GLSL_TYPE_UINT64
:
231 case GLSL_TYPE_FLOAT
:
232 case GLSL_TYPE_FLOAT16
:
233 case GLSL_TYPE_DOUBLE
: {
234 int bit_size
= glsl_get_bit_size(type
);
235 if (glsl_type_is_vector_or_scalar(type
)) {
236 unsigned num_components
= glsl_get_vector_elements(val
->type
);
237 nir_load_const_instr
*load
=
238 nir_load_const_instr_create(b
->shader
, num_components
, bit_size
);
240 memcpy(load
->value
, constant
->values
,
241 sizeof(nir_const_value
) * load
->def
.num_components
);
243 nir_instr_insert_before_cf_list(&b
->nb
.impl
->body
, &load
->instr
);
244 val
->def
= &load
->def
;
246 assert(glsl_type_is_matrix(type
));
247 unsigned columns
= glsl_get_matrix_columns(val
->type
);
248 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, columns
);
249 const struct glsl_type
*column_type
= glsl_get_column_type(val
->type
);
250 for (unsigned i
= 0; i
< columns
; i
++)
251 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
257 case GLSL_TYPE_ARRAY
: {
258 unsigned elems
= glsl_get_length(val
->type
);
259 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
260 const struct glsl_type
*elem_type
= glsl_get_array_element(val
->type
);
261 for (unsigned i
= 0; i
< elems
; i
++)
262 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
267 case GLSL_TYPE_STRUCT
: {
268 unsigned elems
= glsl_get_length(val
->type
);
269 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
270 for (unsigned i
= 0; i
< elems
; i
++) {
271 const struct glsl_type
*elem_type
=
272 glsl_get_struct_field(val
->type
, i
);
273 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
280 vtn_fail("bad constant type");
286 struct vtn_ssa_value
*
287 vtn_ssa_value(struct vtn_builder
*b
, uint32_t value_id
)
289 struct vtn_value
*val
= vtn_untyped_value(b
, value_id
);
290 switch (val
->value_type
) {
291 case vtn_value_type_undef
:
292 return vtn_undef_ssa_value(b
, val
->type
->type
);
294 case vtn_value_type_constant
:
295 return vtn_const_ssa_value(b
, val
->constant
, val
->type
->type
);
297 case vtn_value_type_ssa
:
300 case vtn_value_type_pointer
:
301 vtn_assert(val
->pointer
->ptr_type
&& val
->pointer
->ptr_type
->type
);
302 struct vtn_ssa_value
*ssa
=
303 vtn_create_ssa_value(b
, val
->pointer
->ptr_type
->type
);
304 ssa
->def
= vtn_pointer_to_ssa(b
, val
->pointer
);
308 vtn_fail("Invalid type for an SSA value");
313 vtn_string_literal(struct vtn_builder
*b
, const uint32_t *words
,
314 unsigned word_count
, unsigned *words_used
)
316 char *dup
= ralloc_strndup(b
, (char *)words
, word_count
* sizeof(*words
));
318 /* Ammount of space taken by the string (including the null) */
319 unsigned len
= strlen(dup
) + 1;
320 *words_used
= DIV_ROUND_UP(len
, sizeof(*words
));
326 vtn_foreach_instruction(struct vtn_builder
*b
, const uint32_t *start
,
327 const uint32_t *end
, vtn_instruction_handler handler
)
333 const uint32_t *w
= start
;
335 SpvOp opcode
= w
[0] & SpvOpCodeMask
;
336 unsigned count
= w
[0] >> SpvWordCountShift
;
337 vtn_assert(count
>= 1 && w
+ count
<= end
);
339 b
->spirv_offset
= (uint8_t *)w
- (uint8_t *)b
->spirv
;
343 break; /* Do nothing */
346 b
->file
= vtn_value(b
, w
[1], vtn_value_type_string
)->str
;
358 if (!handler(b
, opcode
, w
, count
))
376 vtn_handle_non_semantic_instruction(struct vtn_builder
*b
, SpvOp ext_opcode
,
377 const uint32_t *w
, unsigned count
)
384 vtn_handle_extension(struct vtn_builder
*b
, SpvOp opcode
,
385 const uint32_t *w
, unsigned count
)
387 const char *ext
= (const char *)&w
[2];
389 case SpvOpExtInstImport
: {
390 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_extension
);
391 if (strcmp(ext
, "GLSL.std.450") == 0) {
392 val
->ext_handler
= vtn_handle_glsl450_instruction
;
393 } else if ((strcmp(ext
, "SPV_AMD_gcn_shader") == 0)
394 && (b
->options
&& b
->options
->caps
.amd_gcn_shader
)) {
395 val
->ext_handler
= vtn_handle_amd_gcn_shader_instruction
;
396 } else if ((strcmp(ext
, "SPV_AMD_shader_ballot") == 0)
397 && (b
->options
&& b
->options
->caps
.amd_shader_ballot
)) {
398 val
->ext_handler
= vtn_handle_amd_shader_ballot_instruction
;
399 } else if ((strcmp(ext
, "SPV_AMD_shader_trinary_minmax") == 0)
400 && (b
->options
&& b
->options
->caps
.amd_trinary_minmax
)) {
401 val
->ext_handler
= vtn_handle_amd_shader_trinary_minmax_instruction
;
402 } else if ((strcmp(ext
, "SPV_AMD_shader_explicit_vertex_parameter") == 0)
403 && (b
->options
&& b
->options
->caps
.amd_shader_explicit_vertex_parameter
)) {
404 val
->ext_handler
= vtn_handle_amd_shader_explicit_vertex_parameter_instruction
;
405 } else if (strcmp(ext
, "OpenCL.std") == 0) {
406 val
->ext_handler
= vtn_handle_opencl_instruction
;
407 } else if (strstr(ext
, "NonSemantic.") == ext
) {
408 val
->ext_handler
= vtn_handle_non_semantic_instruction
;
410 vtn_fail("Unsupported extension: %s", ext
);
416 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
417 bool handled
= val
->ext_handler(b
, w
[4], w
, count
);
423 vtn_fail_with_opcode("Unhandled opcode", opcode
);
428 _foreach_decoration_helper(struct vtn_builder
*b
,
429 struct vtn_value
*base_value
,
431 struct vtn_value
*value
,
432 vtn_decoration_foreach_cb cb
, void *data
)
434 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
436 if (dec
->scope
== VTN_DEC_DECORATION
) {
437 member
= parent_member
;
438 } else if (dec
->scope
>= VTN_DEC_STRUCT_MEMBER0
) {
439 vtn_fail_if(value
->value_type
!= vtn_value_type_type
||
440 value
->type
->base_type
!= vtn_base_type_struct
,
441 "OpMemberDecorate and OpGroupMemberDecorate are only "
442 "allowed on OpTypeStruct");
443 /* This means we haven't recursed yet */
444 assert(value
== base_value
);
446 member
= dec
->scope
- VTN_DEC_STRUCT_MEMBER0
;
448 vtn_fail_if(member
>= base_value
->type
->length
,
449 "OpMemberDecorate specifies member %d but the "
450 "OpTypeStruct has only %u members",
451 member
, base_value
->type
->length
);
453 /* Not a decoration */
454 assert(dec
->scope
== VTN_DEC_EXECUTION_MODE
);
459 assert(dec
->group
->value_type
== vtn_value_type_decoration_group
);
460 _foreach_decoration_helper(b
, base_value
, member
, dec
->group
,
463 cb(b
, base_value
, member
, dec
, data
);
468 /** Iterates (recursively if needed) over all of the decorations on a value
470 * This function iterates over all of the decorations applied to a given
471 * value. If it encounters a decoration group, it recurses into the group
472 * and iterates over all of those decorations as well.
475 vtn_foreach_decoration(struct vtn_builder
*b
, struct vtn_value
*value
,
476 vtn_decoration_foreach_cb cb
, void *data
)
478 _foreach_decoration_helper(b
, value
, -1, value
, cb
, data
);
482 vtn_foreach_execution_mode(struct vtn_builder
*b
, struct vtn_value
*value
,
483 vtn_execution_mode_foreach_cb cb
, void *data
)
485 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
486 if (dec
->scope
!= VTN_DEC_EXECUTION_MODE
)
489 assert(dec
->group
== NULL
);
490 cb(b
, value
, dec
, data
);
495 vtn_handle_decoration(struct vtn_builder
*b
, SpvOp opcode
,
496 const uint32_t *w
, unsigned count
)
498 const uint32_t *w_end
= w
+ count
;
499 const uint32_t target
= w
[1];
503 case SpvOpDecorationGroup
:
504 vtn_push_value(b
, target
, vtn_value_type_decoration_group
);
508 case SpvOpDecorateId
:
509 case SpvOpMemberDecorate
:
510 case SpvOpDecorateString
:
511 case SpvOpMemberDecorateString
:
512 case SpvOpExecutionMode
:
513 case SpvOpExecutionModeId
: {
514 struct vtn_value
*val
= vtn_untyped_value(b
, target
);
516 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
519 case SpvOpDecorateId
:
520 case SpvOpDecorateString
:
521 dec
->scope
= VTN_DEC_DECORATION
;
523 case SpvOpMemberDecorate
:
524 case SpvOpMemberDecorateString
:
525 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(w
++);
526 vtn_fail_if(dec
->scope
< VTN_DEC_STRUCT_MEMBER0
, /* overflow */
527 "Member argument of OpMemberDecorate too large");
529 case SpvOpExecutionMode
:
530 case SpvOpExecutionModeId
:
531 dec
->scope
= VTN_DEC_EXECUTION_MODE
;
534 unreachable("Invalid decoration opcode");
536 dec
->decoration
= *(w
++);
539 /* Link into the list */
540 dec
->next
= val
->decoration
;
541 val
->decoration
= dec
;
545 case SpvOpGroupMemberDecorate
:
546 case SpvOpGroupDecorate
: {
547 struct vtn_value
*group
=
548 vtn_value(b
, target
, vtn_value_type_decoration_group
);
550 for (; w
< w_end
; w
++) {
551 struct vtn_value
*val
= vtn_untyped_value(b
, *w
);
552 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
555 if (opcode
== SpvOpGroupDecorate
) {
556 dec
->scope
= VTN_DEC_DECORATION
;
558 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(++w
);
559 vtn_fail_if(dec
->scope
< 0, /* Check for overflow */
560 "Member argument of OpGroupMemberDecorate too large");
563 /* Link into the list */
564 dec
->next
= val
->decoration
;
565 val
->decoration
= dec
;
571 unreachable("Unhandled opcode");
575 struct member_decoration_ctx
{
577 struct glsl_struct_field
*fields
;
578 struct vtn_type
*type
;
582 * Returns true if the given type contains a struct decorated Block or
586 vtn_type_contains_block(struct vtn_builder
*b
, struct vtn_type
*type
)
588 switch (type
->base_type
) {
589 case vtn_base_type_array
:
590 return vtn_type_contains_block(b
, type
->array_element
);
591 case vtn_base_type_struct
:
592 if (type
->block
|| type
->buffer_block
)
594 for (unsigned i
= 0; i
< type
->length
; i
++) {
595 if (vtn_type_contains_block(b
, type
->members
[i
]))
604 /** Returns true if two types are "compatible", i.e. you can do an OpLoad,
605 * OpStore, or OpCopyMemory between them without breaking anything.
606 * Technically, the SPIR-V rules require the exact same type ID but this lets
607 * us internally be a bit looser.
610 vtn_types_compatible(struct vtn_builder
*b
,
611 struct vtn_type
*t1
, struct vtn_type
*t2
)
613 if (t1
->id
== t2
->id
)
616 if (t1
->base_type
!= t2
->base_type
)
619 switch (t1
->base_type
) {
620 case vtn_base_type_void
:
621 case vtn_base_type_scalar
:
622 case vtn_base_type_vector
:
623 case vtn_base_type_matrix
:
624 case vtn_base_type_image
:
625 case vtn_base_type_sampler
:
626 case vtn_base_type_sampled_image
:
627 return t1
->type
== t2
->type
;
629 case vtn_base_type_array
:
630 return t1
->length
== t2
->length
&&
631 vtn_types_compatible(b
, t1
->array_element
, t2
->array_element
);
633 case vtn_base_type_pointer
:
634 return vtn_types_compatible(b
, t1
->deref
, t2
->deref
);
636 case vtn_base_type_struct
:
637 if (t1
->length
!= t2
->length
)
640 for (unsigned i
= 0; i
< t1
->length
; i
++) {
641 if (!vtn_types_compatible(b
, t1
->members
[i
], t2
->members
[i
]))
646 case vtn_base_type_function
:
647 /* This case shouldn't get hit since you can't copy around function
648 * types. Just require them to be identical.
653 vtn_fail("Invalid base type");
657 vtn_type_without_array(struct vtn_type
*type
)
659 while (type
->base_type
== vtn_base_type_array
)
660 type
= type
->array_element
;
664 /* does a shallow copy of a vtn_type */
666 static struct vtn_type
*
667 vtn_type_copy(struct vtn_builder
*b
, struct vtn_type
*src
)
669 struct vtn_type
*dest
= ralloc(b
, struct vtn_type
);
672 switch (src
->base_type
) {
673 case vtn_base_type_void
:
674 case vtn_base_type_scalar
:
675 case vtn_base_type_vector
:
676 case vtn_base_type_matrix
:
677 case vtn_base_type_array
:
678 case vtn_base_type_pointer
:
679 case vtn_base_type_image
:
680 case vtn_base_type_sampler
:
681 case vtn_base_type_sampled_image
:
682 /* Nothing more to do */
685 case vtn_base_type_struct
:
686 dest
->members
= ralloc_array(b
, struct vtn_type
*, src
->length
);
687 memcpy(dest
->members
, src
->members
,
688 src
->length
* sizeof(src
->members
[0]));
690 dest
->offsets
= ralloc_array(b
, unsigned, src
->length
);
691 memcpy(dest
->offsets
, src
->offsets
,
692 src
->length
* sizeof(src
->offsets
[0]));
695 case vtn_base_type_function
:
696 dest
->params
= ralloc_array(b
, struct vtn_type
*, src
->length
);
697 memcpy(dest
->params
, src
->params
, src
->length
* sizeof(src
->params
[0]));
704 static struct vtn_type
*
705 mutable_matrix_member(struct vtn_builder
*b
, struct vtn_type
*type
, int member
)
707 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
708 type
= type
->members
[member
];
710 /* We may have an array of matrices.... Oh, joy! */
711 while (glsl_type_is_array(type
->type
)) {
712 type
->array_element
= vtn_type_copy(b
, type
->array_element
);
713 type
= type
->array_element
;
716 vtn_assert(glsl_type_is_matrix(type
->type
));
722 vtn_handle_access_qualifier(struct vtn_builder
*b
, struct vtn_type
*type
,
723 int member
, enum gl_access_qualifier access
)
725 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
726 type
= type
->members
[member
];
728 type
->access
|= access
;
732 array_stride_decoration_cb(struct vtn_builder
*b
,
733 struct vtn_value
*val
, int member
,
734 const struct vtn_decoration
*dec
, void *void_ctx
)
736 struct vtn_type
*type
= val
->type
;
738 if (dec
->decoration
== SpvDecorationArrayStride
) {
739 if (vtn_type_contains_block(b
, type
)) {
740 vtn_warn("The ArrayStride decoration cannot be applied to an array "
741 "type which contains a structure type decorated Block "
743 /* Ignore the decoration */
745 vtn_fail_if(dec
->operands
[0] == 0, "ArrayStride must be non-zero");
746 type
->stride
= dec
->operands
[0];
752 struct_member_decoration_cb(struct vtn_builder
*b
,
753 UNUSED
struct vtn_value
*val
, int member
,
754 const struct vtn_decoration
*dec
, void *void_ctx
)
756 struct member_decoration_ctx
*ctx
= void_ctx
;
761 assert(member
< ctx
->num_fields
);
763 switch (dec
->decoration
) {
764 case SpvDecorationRelaxedPrecision
:
765 case SpvDecorationUniform
:
766 case SpvDecorationUniformId
:
767 break; /* FIXME: Do nothing with this for now. */
768 case SpvDecorationNonWritable
:
769 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_NON_WRITEABLE
);
771 case SpvDecorationNonReadable
:
772 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_NON_READABLE
);
774 case SpvDecorationVolatile
:
775 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_VOLATILE
);
777 case SpvDecorationCoherent
:
778 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_COHERENT
);
780 case SpvDecorationNoPerspective
:
781 ctx
->fields
[member
].interpolation
= INTERP_MODE_NOPERSPECTIVE
;
783 case SpvDecorationFlat
:
784 ctx
->fields
[member
].interpolation
= INTERP_MODE_FLAT
;
786 case SpvDecorationExplicitInterpAMD
:
787 ctx
->fields
[member
].interpolation
= INTERP_MODE_EXPLICIT
;
789 case SpvDecorationCentroid
:
790 ctx
->fields
[member
].centroid
= true;
792 case SpvDecorationSample
:
793 ctx
->fields
[member
].sample
= true;
795 case SpvDecorationStream
:
796 /* This is handled later by var_decoration_cb in vtn_variables.c */
798 case SpvDecorationLocation
:
799 ctx
->fields
[member
].location
= dec
->operands
[0];
801 case SpvDecorationComponent
:
802 break; /* FIXME: What should we do with these? */
803 case SpvDecorationBuiltIn
:
804 ctx
->type
->members
[member
] = vtn_type_copy(b
, ctx
->type
->members
[member
]);
805 ctx
->type
->members
[member
]->is_builtin
= true;
806 ctx
->type
->members
[member
]->builtin
= dec
->operands
[0];
807 ctx
->type
->builtin_block
= true;
809 case SpvDecorationOffset
:
810 ctx
->type
->offsets
[member
] = dec
->operands
[0];
811 ctx
->fields
[member
].offset
= dec
->operands
[0];
813 case SpvDecorationMatrixStride
:
814 /* Handled as a second pass */
816 case SpvDecorationColMajor
:
817 break; /* Nothing to do here. Column-major is the default. */
818 case SpvDecorationRowMajor
:
819 mutable_matrix_member(b
, ctx
->type
, member
)->row_major
= true;
822 case SpvDecorationPatch
:
825 case SpvDecorationSpecId
:
826 case SpvDecorationBlock
:
827 case SpvDecorationBufferBlock
:
828 case SpvDecorationArrayStride
:
829 case SpvDecorationGLSLShared
:
830 case SpvDecorationGLSLPacked
:
831 case SpvDecorationInvariant
:
832 case SpvDecorationRestrict
:
833 case SpvDecorationAliased
:
834 case SpvDecorationConstant
:
835 case SpvDecorationIndex
:
836 case SpvDecorationBinding
:
837 case SpvDecorationDescriptorSet
:
838 case SpvDecorationLinkageAttributes
:
839 case SpvDecorationNoContraction
:
840 case SpvDecorationInputAttachmentIndex
:
841 vtn_warn("Decoration not allowed on struct members: %s",
842 spirv_decoration_to_string(dec
->decoration
));
845 case SpvDecorationXfbBuffer
:
846 case SpvDecorationXfbStride
:
847 /* This is handled later by var_decoration_cb in vtn_variables.c */
850 case SpvDecorationCPacked
:
851 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
)
852 vtn_warn("Decoration only allowed for CL-style kernels: %s",
853 spirv_decoration_to_string(dec
->decoration
));
855 ctx
->type
->packed
= true;
858 case SpvDecorationSaturatedConversion
:
859 case SpvDecorationFuncParamAttr
:
860 case SpvDecorationFPRoundingMode
:
861 case SpvDecorationFPFastMathMode
:
862 case SpvDecorationAlignment
:
863 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
) {
864 vtn_warn("Decoration only allowed for CL-style kernels: %s",
865 spirv_decoration_to_string(dec
->decoration
));
869 case SpvDecorationUserSemantic
:
870 /* User semantic decorations can safely be ignored by the driver. */
874 vtn_fail_with_decoration("Unhandled decoration", dec
->decoration
);
878 /** Chases the array type all the way down to the tail and rewrites the
879 * glsl_types to be based off the tail's glsl_type.
882 vtn_array_type_rewrite_glsl_type(struct vtn_type
*type
)
884 if (type
->base_type
!= vtn_base_type_array
)
887 vtn_array_type_rewrite_glsl_type(type
->array_element
);
889 type
->type
= glsl_array_type(type
->array_element
->type
,
890 type
->length
, type
->stride
);
893 /* Matrix strides are handled as a separate pass because we need to know
894 * whether the matrix is row-major or not first.
897 struct_member_matrix_stride_cb(struct vtn_builder
*b
,
898 UNUSED
struct vtn_value
*val
, int member
,
899 const struct vtn_decoration
*dec
,
902 if (dec
->decoration
!= SpvDecorationMatrixStride
)
905 vtn_fail_if(member
< 0,
906 "The MatrixStride decoration is only allowed on members "
908 vtn_fail_if(dec
->operands
[0] == 0, "MatrixStride must be non-zero");
910 struct member_decoration_ctx
*ctx
= void_ctx
;
912 struct vtn_type
*mat_type
= mutable_matrix_member(b
, ctx
->type
, member
);
913 if (mat_type
->row_major
) {
914 mat_type
->array_element
= vtn_type_copy(b
, mat_type
->array_element
);
915 mat_type
->stride
= mat_type
->array_element
->stride
;
916 mat_type
->array_element
->stride
= dec
->operands
[0];
918 mat_type
->type
= glsl_explicit_matrix_type(mat_type
->type
,
919 dec
->operands
[0], true);
920 mat_type
->array_element
->type
= glsl_get_column_type(mat_type
->type
);
922 vtn_assert(mat_type
->array_element
->stride
> 0);
923 mat_type
->stride
= dec
->operands
[0];
925 mat_type
->type
= glsl_explicit_matrix_type(mat_type
->type
,
926 dec
->operands
[0], false);
929 /* Now that we've replaced the glsl_type with a properly strided matrix
930 * type, rewrite the member type so that it's an array of the proper kind
933 vtn_array_type_rewrite_glsl_type(ctx
->type
->members
[member
]);
934 ctx
->fields
[member
].type
= ctx
->type
->members
[member
]->type
;
938 struct_block_decoration_cb(struct vtn_builder
*b
,
939 struct vtn_value
*val
, int member
,
940 const struct vtn_decoration
*dec
, void *ctx
)
945 struct vtn_type
*type
= val
->type
;
946 if (dec
->decoration
== SpvDecorationBlock
)
948 else if (dec
->decoration
== SpvDecorationBufferBlock
)
949 type
->buffer_block
= true;
953 type_decoration_cb(struct vtn_builder
*b
,
954 struct vtn_value
*val
, int member
,
955 const struct vtn_decoration
*dec
, UNUSED
void *ctx
)
957 struct vtn_type
*type
= val
->type
;
960 /* This should have been handled by OpTypeStruct */
961 assert(val
->type
->base_type
== vtn_base_type_struct
);
962 assert(member
>= 0 && member
< val
->type
->length
);
966 switch (dec
->decoration
) {
967 case SpvDecorationArrayStride
:
968 vtn_assert(type
->base_type
== vtn_base_type_array
||
969 type
->base_type
== vtn_base_type_pointer
);
971 case SpvDecorationBlock
:
972 vtn_assert(type
->base_type
== vtn_base_type_struct
);
973 vtn_assert(type
->block
);
975 case SpvDecorationBufferBlock
:
976 vtn_assert(type
->base_type
== vtn_base_type_struct
);
977 vtn_assert(type
->buffer_block
);
979 case SpvDecorationGLSLShared
:
980 case SpvDecorationGLSLPacked
:
981 /* Ignore these, since we get explicit offsets anyways */
984 case SpvDecorationRowMajor
:
985 case SpvDecorationColMajor
:
986 case SpvDecorationMatrixStride
:
987 case SpvDecorationBuiltIn
:
988 case SpvDecorationNoPerspective
:
989 case SpvDecorationFlat
:
990 case SpvDecorationPatch
:
991 case SpvDecorationCentroid
:
992 case SpvDecorationSample
:
993 case SpvDecorationExplicitInterpAMD
:
994 case SpvDecorationVolatile
:
995 case SpvDecorationCoherent
:
996 case SpvDecorationNonWritable
:
997 case SpvDecorationNonReadable
:
998 case SpvDecorationUniform
:
999 case SpvDecorationUniformId
:
1000 case SpvDecorationLocation
:
1001 case SpvDecorationComponent
:
1002 case SpvDecorationOffset
:
1003 case SpvDecorationXfbBuffer
:
1004 case SpvDecorationXfbStride
:
1005 case SpvDecorationUserSemantic
:
1006 vtn_warn("Decoration only allowed for struct members: %s",
1007 spirv_decoration_to_string(dec
->decoration
));
1010 case SpvDecorationStream
:
1011 /* We don't need to do anything here, as stream is filled up when
1012 * aplying the decoration to a variable, just check that if it is not a
1013 * struct member, it should be a struct.
1015 vtn_assert(type
->base_type
== vtn_base_type_struct
);
1018 case SpvDecorationRelaxedPrecision
:
1019 case SpvDecorationSpecId
:
1020 case SpvDecorationInvariant
:
1021 case SpvDecorationRestrict
:
1022 case SpvDecorationAliased
:
1023 case SpvDecorationConstant
:
1024 case SpvDecorationIndex
:
1025 case SpvDecorationBinding
:
1026 case SpvDecorationDescriptorSet
:
1027 case SpvDecorationLinkageAttributes
:
1028 case SpvDecorationNoContraction
:
1029 case SpvDecorationInputAttachmentIndex
:
1030 vtn_warn("Decoration not allowed on types: %s",
1031 spirv_decoration_to_string(dec
->decoration
));
1034 case SpvDecorationCPacked
:
1035 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
)
1036 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1037 spirv_decoration_to_string(dec
->decoration
));
1039 type
->packed
= true;
1042 case SpvDecorationSaturatedConversion
:
1043 case SpvDecorationFuncParamAttr
:
1044 case SpvDecorationFPRoundingMode
:
1045 case SpvDecorationFPFastMathMode
:
1046 case SpvDecorationAlignment
:
1047 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1048 spirv_decoration_to_string(dec
->decoration
));
1052 vtn_fail_with_decoration("Unhandled decoration", dec
->decoration
);
1057 translate_image_format(struct vtn_builder
*b
, SpvImageFormat format
)
1060 case SpvImageFormatUnknown
: return PIPE_FORMAT_NONE
;
1061 case SpvImageFormatRgba32f
: return PIPE_FORMAT_R32G32B32A32_FLOAT
;
1062 case SpvImageFormatRgba16f
: return PIPE_FORMAT_R16G16B16A16_FLOAT
;
1063 case SpvImageFormatR32f
: return PIPE_FORMAT_R32_FLOAT
;
1064 case SpvImageFormatRgba8
: return PIPE_FORMAT_R8G8B8A8_UNORM
;
1065 case SpvImageFormatRgba8Snorm
: return PIPE_FORMAT_R8G8B8A8_SNORM
;
1066 case SpvImageFormatRg32f
: return PIPE_FORMAT_R32G32_FLOAT
;
1067 case SpvImageFormatRg16f
: return PIPE_FORMAT_R16G16_FLOAT
;
1068 case SpvImageFormatR11fG11fB10f
: return PIPE_FORMAT_R11G11B10_FLOAT
;
1069 case SpvImageFormatR16f
: return PIPE_FORMAT_R16_FLOAT
;
1070 case SpvImageFormatRgba16
: return PIPE_FORMAT_R16G16B16A16_UNORM
;
1071 case SpvImageFormatRgb10A2
: return PIPE_FORMAT_R10G10B10A2_UNORM
;
1072 case SpvImageFormatRg16
: return PIPE_FORMAT_R16G16_UNORM
;
1073 case SpvImageFormatRg8
: return PIPE_FORMAT_R8G8_UNORM
;
1074 case SpvImageFormatR16
: return PIPE_FORMAT_R16_UNORM
;
1075 case SpvImageFormatR8
: return PIPE_FORMAT_R8_UNORM
;
1076 case SpvImageFormatRgba16Snorm
: return PIPE_FORMAT_R16G16B16A16_SNORM
;
1077 case SpvImageFormatRg16Snorm
: return PIPE_FORMAT_R16G16_SNORM
;
1078 case SpvImageFormatRg8Snorm
: return PIPE_FORMAT_R8G8_SNORM
;
1079 case SpvImageFormatR16Snorm
: return PIPE_FORMAT_R16_SNORM
;
1080 case SpvImageFormatR8Snorm
: return PIPE_FORMAT_R8_SNORM
;
1081 case SpvImageFormatRgba32i
: return PIPE_FORMAT_R32G32B32A32_SINT
;
1082 case SpvImageFormatRgba16i
: return PIPE_FORMAT_R16G16B16A16_SINT
;
1083 case SpvImageFormatRgba8i
: return PIPE_FORMAT_R8G8B8A8_SINT
;
1084 case SpvImageFormatR32i
: return PIPE_FORMAT_R32_SINT
;
1085 case SpvImageFormatRg32i
: return PIPE_FORMAT_R32G32_SINT
;
1086 case SpvImageFormatRg16i
: return PIPE_FORMAT_R16G16_SINT
;
1087 case SpvImageFormatRg8i
: return PIPE_FORMAT_R8G8_SINT
;
1088 case SpvImageFormatR16i
: return PIPE_FORMAT_R16_SINT
;
1089 case SpvImageFormatR8i
: return PIPE_FORMAT_R8_SINT
;
1090 case SpvImageFormatRgba32ui
: return PIPE_FORMAT_R32G32B32A32_UINT
;
1091 case SpvImageFormatRgba16ui
: return PIPE_FORMAT_R16G16B16A16_UINT
;
1092 case SpvImageFormatRgba8ui
: return PIPE_FORMAT_R8G8B8A8_UINT
;
1093 case SpvImageFormatR32ui
: return PIPE_FORMAT_R32_UINT
;
1094 case SpvImageFormatRgb10a2ui
: return PIPE_FORMAT_R10G10B10A2_UINT
;
1095 case SpvImageFormatRg32ui
: return PIPE_FORMAT_R32G32_UINT
;
1096 case SpvImageFormatRg16ui
: return PIPE_FORMAT_R16G16_UINT
;
1097 case SpvImageFormatRg8ui
: return PIPE_FORMAT_R8G8_UINT
;
1098 case SpvImageFormatR16ui
: return PIPE_FORMAT_R16_UINT
;
1099 case SpvImageFormatR8ui
: return PIPE_FORMAT_R8_UINT
;
1101 vtn_fail("Invalid image format: %s (%u)",
1102 spirv_imageformat_to_string(format
), format
);
1107 vtn_handle_type(struct vtn_builder
*b
, SpvOp opcode
,
1108 const uint32_t *w
, unsigned count
)
1110 struct vtn_value
*val
= NULL
;
1112 /* In order to properly handle forward declarations, we have to defer
1113 * allocation for pointer types.
1115 if (opcode
!= SpvOpTypePointer
&& opcode
!= SpvOpTypeForwardPointer
) {
1116 val
= vtn_push_value(b
, w
[1], vtn_value_type_type
);
1117 vtn_fail_if(val
->type
!= NULL
,
1118 "Only pointers can have forward declarations");
1119 val
->type
= rzalloc(b
, struct vtn_type
);
1120 val
->type
->id
= w
[1];
1125 val
->type
->base_type
= vtn_base_type_void
;
1126 val
->type
->type
= glsl_void_type();
1129 val
->type
->base_type
= vtn_base_type_scalar
;
1130 val
->type
->type
= glsl_bool_type();
1131 val
->type
->length
= 1;
1133 case SpvOpTypeInt
: {
1134 int bit_size
= w
[2];
1135 const bool signedness
= w
[3];
1136 val
->type
->base_type
= vtn_base_type_scalar
;
1139 val
->type
->type
= (signedness
? glsl_int64_t_type() : glsl_uint64_t_type());
1142 val
->type
->type
= (signedness
? glsl_int_type() : glsl_uint_type());
1145 val
->type
->type
= (signedness
? glsl_int16_t_type() : glsl_uint16_t_type());
1148 val
->type
->type
= (signedness
? glsl_int8_t_type() : glsl_uint8_t_type());
1151 vtn_fail("Invalid int bit size: %u", bit_size
);
1153 val
->type
->length
= 1;
1157 case SpvOpTypeFloat
: {
1158 int bit_size
= w
[2];
1159 val
->type
->base_type
= vtn_base_type_scalar
;
1162 val
->type
->type
= glsl_float16_t_type();
1165 val
->type
->type
= glsl_float_type();
1168 val
->type
->type
= glsl_double_type();
1171 vtn_fail("Invalid float bit size: %u", bit_size
);
1173 val
->type
->length
= 1;
1177 case SpvOpTypeVector
: {
1178 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1179 unsigned elems
= w
[3];
1181 vtn_fail_if(base
->base_type
!= vtn_base_type_scalar
,
1182 "Base type for OpTypeVector must be a scalar");
1183 vtn_fail_if((elems
< 2 || elems
> 4) && (elems
!= 8) && (elems
!= 16),
1184 "Invalid component count for OpTypeVector");
1186 val
->type
->base_type
= vtn_base_type_vector
;
1187 val
->type
->type
= glsl_vector_type(glsl_get_base_type(base
->type
), elems
);
1188 val
->type
->length
= elems
;
1189 val
->type
->stride
= glsl_type_is_boolean(val
->type
->type
)
1190 ? 4 : glsl_get_bit_size(base
->type
) / 8;
1191 val
->type
->array_element
= base
;
1195 case SpvOpTypeMatrix
: {
1196 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1197 unsigned columns
= w
[3];
1199 vtn_fail_if(base
->base_type
!= vtn_base_type_vector
,
1200 "Base type for OpTypeMatrix must be a vector");
1201 vtn_fail_if(columns
< 2 || columns
> 4,
1202 "Invalid column count for OpTypeMatrix");
1204 val
->type
->base_type
= vtn_base_type_matrix
;
1205 val
->type
->type
= glsl_matrix_type(glsl_get_base_type(base
->type
),
1206 glsl_get_vector_elements(base
->type
),
1208 vtn_fail_if(glsl_type_is_error(val
->type
->type
),
1209 "Unsupported base type for OpTypeMatrix");
1210 assert(!glsl_type_is_error(val
->type
->type
));
1211 val
->type
->length
= columns
;
1212 val
->type
->array_element
= base
;
1213 val
->type
->row_major
= false;
1214 val
->type
->stride
= 0;
1218 case SpvOpTypeRuntimeArray
:
1219 case SpvOpTypeArray
: {
1220 struct vtn_type
*array_element
=
1221 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1223 if (opcode
== SpvOpTypeRuntimeArray
) {
1224 /* A length of 0 is used to denote unsized arrays */
1225 val
->type
->length
= 0;
1227 val
->type
->length
= vtn_constant_uint(b
, w
[3]);
1230 val
->type
->base_type
= vtn_base_type_array
;
1231 val
->type
->array_element
= array_element
;
1232 if (b
->shader
->info
.stage
== MESA_SHADER_KERNEL
)
1233 val
->type
->stride
= glsl_get_cl_size(array_element
->type
);
1235 vtn_foreach_decoration(b
, val
, array_stride_decoration_cb
, NULL
);
1236 val
->type
->type
= glsl_array_type(array_element
->type
, val
->type
->length
,
1241 case SpvOpTypeStruct
: {
1242 unsigned num_fields
= count
- 2;
1243 val
->type
->base_type
= vtn_base_type_struct
;
1244 val
->type
->length
= num_fields
;
1245 val
->type
->members
= ralloc_array(b
, struct vtn_type
*, num_fields
);
1246 val
->type
->offsets
= ralloc_array(b
, unsigned, num_fields
);
1247 val
->type
->packed
= false;
1249 NIR_VLA(struct glsl_struct_field
, fields
, count
);
1250 for (unsigned i
= 0; i
< num_fields
; i
++) {
1251 val
->type
->members
[i
] =
1252 vtn_value(b
, w
[i
+ 2], vtn_value_type_type
)->type
;
1253 fields
[i
] = (struct glsl_struct_field
) {
1254 .type
= val
->type
->members
[i
]->type
,
1255 .name
= ralloc_asprintf(b
, "field%d", i
),
1261 if (b
->shader
->info
.stage
== MESA_SHADER_KERNEL
) {
1262 unsigned offset
= 0;
1263 for (unsigned i
= 0; i
< num_fields
; i
++) {
1264 offset
= align(offset
, glsl_get_cl_alignment(fields
[i
].type
));
1265 fields
[i
].offset
= offset
;
1266 offset
+= glsl_get_cl_size(fields
[i
].type
);
1270 struct member_decoration_ctx ctx
= {
1271 .num_fields
= num_fields
,
1276 vtn_foreach_decoration(b
, val
, struct_member_decoration_cb
, &ctx
);
1277 vtn_foreach_decoration(b
, val
, struct_member_matrix_stride_cb
, &ctx
);
1279 vtn_foreach_decoration(b
, val
, struct_block_decoration_cb
, NULL
);
1281 const char *name
= val
->name
;
1283 if (val
->type
->block
|| val
->type
->buffer_block
) {
1284 /* Packing will be ignored since types coming from SPIR-V are
1285 * explicitly laid out.
1287 val
->type
->type
= glsl_interface_type(fields
, num_fields
,
1288 /* packing */ 0, false,
1289 name
? name
: "block");
1291 val
->type
->type
= glsl_struct_type(fields
, num_fields
,
1292 name
? name
: "struct", false);
1297 case SpvOpTypeFunction
: {
1298 val
->type
->base_type
= vtn_base_type_function
;
1299 val
->type
->type
= NULL
;
1301 val
->type
->return_type
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1303 const unsigned num_params
= count
- 3;
1304 val
->type
->length
= num_params
;
1305 val
->type
->params
= ralloc_array(b
, struct vtn_type
*, num_params
);
1306 for (unsigned i
= 0; i
< count
- 3; i
++) {
1307 val
->type
->params
[i
] =
1308 vtn_value(b
, w
[i
+ 3], vtn_value_type_type
)->type
;
1313 case SpvOpTypePointer
:
1314 case SpvOpTypeForwardPointer
: {
1315 /* We can't blindly push the value because it might be a forward
1318 val
= vtn_untyped_value(b
, w
[1]);
1320 SpvStorageClass storage_class
= w
[2];
1322 if (val
->value_type
== vtn_value_type_invalid
) {
1323 val
->value_type
= vtn_value_type_type
;
1324 val
->type
= rzalloc(b
, struct vtn_type
);
1325 val
->type
->id
= w
[1];
1326 val
->type
->base_type
= vtn_base_type_pointer
;
1327 val
->type
->storage_class
= storage_class
;
1329 /* These can actually be stored to nir_variables and used as SSA
1330 * values so they need a real glsl_type.
1332 enum vtn_variable_mode mode
= vtn_storage_class_to_mode(
1333 b
, storage_class
, NULL
, NULL
);
1334 val
->type
->type
= nir_address_format_to_glsl_type(
1335 vtn_mode_to_address_format(b
, mode
));
1337 vtn_fail_if(val
->type
->storage_class
!= storage_class
,
1338 "The storage classes of an OpTypePointer and any "
1339 "OpTypeForwardPointers that provide forward "
1340 "declarations of it must match.");
1343 if (opcode
== SpvOpTypePointer
) {
1344 vtn_fail_if(val
->type
->deref
!= NULL
,
1345 "While OpTypeForwardPointer can be used to provide a "
1346 "forward declaration of a pointer, OpTypePointer can "
1347 "only be used once for a given id.");
1349 val
->type
->deref
= vtn_value(b
, w
[3], vtn_value_type_type
)->type
;
1351 /* Only certain storage classes use ArrayStride. The others (in
1352 * particular Workgroup) are expected to be laid out by the driver.
1354 switch (storage_class
) {
1355 case SpvStorageClassUniform
:
1356 case SpvStorageClassPushConstant
:
1357 case SpvStorageClassStorageBuffer
:
1358 case SpvStorageClassPhysicalStorageBuffer
:
1359 vtn_foreach_decoration(b
, val
, array_stride_decoration_cb
, NULL
);
1362 /* Nothing to do. */
1366 if (b
->physical_ptrs
) {
1367 switch (storage_class
) {
1368 case SpvStorageClassFunction
:
1369 case SpvStorageClassWorkgroup
:
1370 case SpvStorageClassCrossWorkgroup
:
1371 case SpvStorageClassUniformConstant
:
1372 val
->type
->stride
= align(glsl_get_cl_size(val
->type
->deref
->type
),
1373 glsl_get_cl_alignment(val
->type
->deref
->type
));
1383 case SpvOpTypeImage
: {
1384 val
->type
->base_type
= vtn_base_type_image
;
1386 const struct vtn_type
*sampled_type
=
1387 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1389 vtn_fail_if(sampled_type
->base_type
!= vtn_base_type_scalar
||
1390 glsl_get_bit_size(sampled_type
->type
) != 32,
1391 "Sampled type of OpTypeImage must be a 32-bit scalar");
1393 enum glsl_sampler_dim dim
;
1394 switch ((SpvDim
)w
[3]) {
1395 case SpvDim1D
: dim
= GLSL_SAMPLER_DIM_1D
; break;
1396 case SpvDim2D
: dim
= GLSL_SAMPLER_DIM_2D
; break;
1397 case SpvDim3D
: dim
= GLSL_SAMPLER_DIM_3D
; break;
1398 case SpvDimCube
: dim
= GLSL_SAMPLER_DIM_CUBE
; break;
1399 case SpvDimRect
: dim
= GLSL_SAMPLER_DIM_RECT
; break;
1400 case SpvDimBuffer
: dim
= GLSL_SAMPLER_DIM_BUF
; break;
1401 case SpvDimSubpassData
: dim
= GLSL_SAMPLER_DIM_SUBPASS
; break;
1403 vtn_fail("Invalid SPIR-V image dimensionality: %s (%u)",
1404 spirv_dim_to_string((SpvDim
)w
[3]), w
[3]);
1407 /* w[4]: as per Vulkan spec "Validation Rules within a Module",
1408 * The “Depth” operand of OpTypeImage is ignored.
1410 bool is_array
= w
[5];
1411 bool multisampled
= w
[6];
1412 unsigned sampled
= w
[7];
1413 SpvImageFormat format
= w
[8];
1416 val
->type
->access_qualifier
= w
[9];
1418 val
->type
->access_qualifier
= SpvAccessQualifierReadWrite
;
1421 if (dim
== GLSL_SAMPLER_DIM_2D
)
1422 dim
= GLSL_SAMPLER_DIM_MS
;
1423 else if (dim
== GLSL_SAMPLER_DIM_SUBPASS
)
1424 dim
= GLSL_SAMPLER_DIM_SUBPASS_MS
;
1426 vtn_fail("Unsupported multisampled image type");
1429 val
->type
->image_format
= translate_image_format(b
, format
);
1431 enum glsl_base_type sampled_base_type
=
1432 glsl_get_base_type(sampled_type
->type
);
1434 val
->type
->sampled
= true;
1435 val
->type
->type
= glsl_sampler_type(dim
, false, is_array
,
1437 } else if (sampled
== 2) {
1438 val
->type
->sampled
= false;
1439 val
->type
->type
= glsl_image_type(dim
, is_array
, sampled_base_type
);
1441 vtn_fail("We need to know if the image will be sampled");
1446 case SpvOpTypeSampledImage
:
1447 val
->type
->base_type
= vtn_base_type_sampled_image
;
1448 val
->type
->image
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1449 val
->type
->type
= val
->type
->image
->type
;
1452 case SpvOpTypeSampler
:
1453 /* The actual sampler type here doesn't really matter. It gets
1454 * thrown away the moment you combine it with an image. What really
1455 * matters is that it's a sampler type as opposed to an integer type
1456 * so the backend knows what to do.
1458 val
->type
->base_type
= vtn_base_type_sampler
;
1459 val
->type
->type
= glsl_bare_sampler_type();
1462 case SpvOpTypeOpaque
:
1463 case SpvOpTypeEvent
:
1464 case SpvOpTypeDeviceEvent
:
1465 case SpvOpTypeReserveId
:
1466 case SpvOpTypeQueue
:
1469 vtn_fail_with_opcode("Unhandled opcode", opcode
);
1472 vtn_foreach_decoration(b
, val
, type_decoration_cb
, NULL
);
1474 if (val
->type
->base_type
== vtn_base_type_struct
&&
1475 (val
->type
->block
|| val
->type
->buffer_block
)) {
1476 for (unsigned i
= 0; i
< val
->type
->length
; i
++) {
1477 vtn_fail_if(vtn_type_contains_block(b
, val
->type
->members
[i
]),
1478 "Block and BufferBlock decorations cannot decorate a "
1479 "structure type that is nested at any level inside "
1480 "another structure type decorated with Block or "
1486 static nir_constant
*
1487 vtn_null_constant(struct vtn_builder
*b
, struct vtn_type
*type
)
1489 nir_constant
*c
= rzalloc(b
, nir_constant
);
1491 switch (type
->base_type
) {
1492 case vtn_base_type_scalar
:
1493 case vtn_base_type_vector
:
1494 /* Nothing to do here. It's already initialized to zero */
1497 case vtn_base_type_pointer
: {
1498 enum vtn_variable_mode mode
= vtn_storage_class_to_mode(
1499 b
, type
->storage_class
, type
->deref
, NULL
);
1500 nir_address_format addr_format
= vtn_mode_to_address_format(b
, mode
);
1502 const nir_const_value
*null_value
= nir_address_format_null_value(addr_format
);
1503 memcpy(c
->values
, null_value
,
1504 sizeof(nir_const_value
) * nir_address_format_num_components(addr_format
));
1508 case vtn_base_type_void
:
1509 case vtn_base_type_image
:
1510 case vtn_base_type_sampler
:
1511 case vtn_base_type_sampled_image
:
1512 case vtn_base_type_function
:
1513 /* For those we have to return something but it doesn't matter what. */
1516 case vtn_base_type_matrix
:
1517 case vtn_base_type_array
:
1518 vtn_assert(type
->length
> 0);
1519 c
->num_elements
= type
->length
;
1520 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1522 c
->elements
[0] = vtn_null_constant(b
, type
->array_element
);
1523 for (unsigned i
= 1; i
< c
->num_elements
; i
++)
1524 c
->elements
[i
] = c
->elements
[0];
1527 case vtn_base_type_struct
:
1528 c
->num_elements
= type
->length
;
1529 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1530 for (unsigned i
= 0; i
< c
->num_elements
; i
++)
1531 c
->elements
[i
] = vtn_null_constant(b
, type
->members
[i
]);
1535 vtn_fail("Invalid type for null constant");
1542 spec_constant_decoration_cb(struct vtn_builder
*b
, UNUSED
struct vtn_value
*val
,
1543 ASSERTED
int member
,
1544 const struct vtn_decoration
*dec
, void *data
)
1546 vtn_assert(member
== -1);
1547 if (dec
->decoration
!= SpvDecorationSpecId
)
1550 struct spec_constant_value
*const_value
= data
;
1552 for (unsigned i
= 0; i
< b
->num_specializations
; i
++) {
1553 if (b
->specializations
[i
].id
== dec
->operands
[0]) {
1554 if (const_value
->is_double
)
1555 const_value
->data64
= b
->specializations
[i
].data64
;
1557 const_value
->data32
= b
->specializations
[i
].data32
;
1564 get_specialization(struct vtn_builder
*b
, struct vtn_value
*val
,
1565 uint32_t const_value
)
1567 struct spec_constant_value data
;
1568 data
.is_double
= false;
1569 data
.data32
= const_value
;
1570 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1575 get_specialization64(struct vtn_builder
*b
, struct vtn_value
*val
,
1576 uint64_t const_value
)
1578 struct spec_constant_value data
;
1579 data
.is_double
= true;
1580 data
.data64
= const_value
;
1581 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1586 handle_workgroup_size_decoration_cb(struct vtn_builder
*b
,
1587 struct vtn_value
*val
,
1588 ASSERTED
int member
,
1589 const struct vtn_decoration
*dec
,
1592 vtn_assert(member
== -1);
1593 if (dec
->decoration
!= SpvDecorationBuiltIn
||
1594 dec
->operands
[0] != SpvBuiltInWorkgroupSize
)
1597 vtn_assert(val
->type
->type
== glsl_vector_type(GLSL_TYPE_UINT
, 3));
1598 b
->workgroup_size_builtin
= val
;
1602 vtn_handle_constant(struct vtn_builder
*b
, SpvOp opcode
,
1603 const uint32_t *w
, unsigned count
)
1605 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_constant
);
1606 val
->constant
= rzalloc(b
, nir_constant
);
1608 case SpvOpConstantTrue
:
1609 case SpvOpConstantFalse
:
1610 case SpvOpSpecConstantTrue
:
1611 case SpvOpSpecConstantFalse
: {
1612 vtn_fail_if(val
->type
->type
!= glsl_bool_type(),
1613 "Result type of %s must be OpTypeBool",
1614 spirv_op_to_string(opcode
));
1616 uint32_t int_val
= (opcode
== SpvOpConstantTrue
||
1617 opcode
== SpvOpSpecConstantTrue
);
1619 if (opcode
== SpvOpSpecConstantTrue
||
1620 opcode
== SpvOpSpecConstantFalse
)
1621 int_val
= get_specialization(b
, val
, int_val
);
1623 val
->constant
->values
[0].b
= int_val
!= 0;
1627 case SpvOpConstant
: {
1628 vtn_fail_if(val
->type
->base_type
!= vtn_base_type_scalar
,
1629 "Result type of %s must be a scalar",
1630 spirv_op_to_string(opcode
));
1631 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1634 val
->constant
->values
[0].u64
= vtn_u64_literal(&w
[3]);
1637 val
->constant
->values
[0].u32
= w
[3];
1640 val
->constant
->values
[0].u16
= w
[3];
1643 val
->constant
->values
[0].u8
= w
[3];
1646 vtn_fail("Unsupported SpvOpConstant bit size: %u", bit_size
);
1651 case SpvOpSpecConstant
: {
1652 vtn_fail_if(val
->type
->base_type
!= vtn_base_type_scalar
,
1653 "Result type of %s must be a scalar",
1654 spirv_op_to_string(opcode
));
1655 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1658 val
->constant
->values
[0].u64
=
1659 get_specialization64(b
, val
, vtn_u64_literal(&w
[3]));
1662 val
->constant
->values
[0].u32
= get_specialization(b
, val
, w
[3]);
1665 val
->constant
->values
[0].u16
= get_specialization(b
, val
, w
[3]);
1668 val
->constant
->values
[0].u8
= get_specialization(b
, val
, w
[3]);
1671 vtn_fail("Unsupported SpvOpSpecConstant bit size");
1676 case SpvOpSpecConstantComposite
:
1677 case SpvOpConstantComposite
: {
1678 unsigned elem_count
= count
- 3;
1679 vtn_fail_if(elem_count
!= val
->type
->length
,
1680 "%s has %u constituents, expected %u",
1681 spirv_op_to_string(opcode
), elem_count
, val
->type
->length
);
1683 nir_constant
**elems
= ralloc_array(b
, nir_constant
*, elem_count
);
1684 for (unsigned i
= 0; i
< elem_count
; i
++) {
1685 struct vtn_value
*val
= vtn_untyped_value(b
, w
[i
+ 3]);
1687 if (val
->value_type
== vtn_value_type_constant
) {
1688 elems
[i
] = val
->constant
;
1690 vtn_fail_if(val
->value_type
!= vtn_value_type_undef
,
1691 "only constants or undefs allowed for "
1692 "SpvOpConstantComposite");
1693 /* to make it easier, just insert a NULL constant for now */
1694 elems
[i
] = vtn_null_constant(b
, val
->type
);
1698 switch (val
->type
->base_type
) {
1699 case vtn_base_type_vector
: {
1700 assert(glsl_type_is_vector(val
->type
->type
));
1701 for (unsigned i
= 0; i
< elem_count
; i
++)
1702 val
->constant
->values
[i
] = elems
[i
]->values
[0];
1706 case vtn_base_type_matrix
:
1707 case vtn_base_type_struct
:
1708 case vtn_base_type_array
:
1709 ralloc_steal(val
->constant
, elems
);
1710 val
->constant
->num_elements
= elem_count
;
1711 val
->constant
->elements
= elems
;
1715 vtn_fail("Result type of %s must be a composite type",
1716 spirv_op_to_string(opcode
));
1721 case SpvOpSpecConstantOp
: {
1722 SpvOp opcode
= get_specialization(b
, val
, w
[3]);
1724 case SpvOpVectorShuffle
: {
1725 struct vtn_value
*v0
= &b
->values
[w
[4]];
1726 struct vtn_value
*v1
= &b
->values
[w
[5]];
1728 vtn_assert(v0
->value_type
== vtn_value_type_constant
||
1729 v0
->value_type
== vtn_value_type_undef
);
1730 vtn_assert(v1
->value_type
== vtn_value_type_constant
||
1731 v1
->value_type
== vtn_value_type_undef
);
1733 unsigned len0
= glsl_get_vector_elements(v0
->type
->type
);
1734 unsigned len1
= glsl_get_vector_elements(v1
->type
->type
);
1736 vtn_assert(len0
+ len1
< 16);
1738 unsigned bit_size
= glsl_get_bit_size(val
->type
->type
);
1739 unsigned bit_size0
= glsl_get_bit_size(v0
->type
->type
);
1740 unsigned bit_size1
= glsl_get_bit_size(v1
->type
->type
);
1742 vtn_assert(bit_size
== bit_size0
&& bit_size
== bit_size1
);
1743 (void)bit_size0
; (void)bit_size1
;
1745 nir_const_value undef
= { .u64
= 0xdeadbeefdeadbeef };
1746 nir_const_value combined
[NIR_MAX_VEC_COMPONENTS
* 2];
1748 if (v0
->value_type
== vtn_value_type_constant
) {
1749 for (unsigned i
= 0; i
< len0
; i
++)
1750 combined
[i
] = v0
->constant
->values
[i
];
1752 if (v1
->value_type
== vtn_value_type_constant
) {
1753 for (unsigned i
= 0; i
< len1
; i
++)
1754 combined
[len0
+ i
] = v1
->constant
->values
[i
];
1757 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1758 uint32_t comp
= w
[i
+ 6];
1759 if (comp
== (uint32_t)-1) {
1760 /* If component is not used, set the value to a known constant
1761 * to detect if it is wrongly used.
1763 val
->constant
->values
[j
] = undef
;
1765 vtn_fail_if(comp
>= len0
+ len1
,
1766 "All Component literals must either be FFFFFFFF "
1767 "or in [0, N - 1] (inclusive).");
1768 val
->constant
->values
[j
] = combined
[comp
];
1774 case SpvOpCompositeExtract
:
1775 case SpvOpCompositeInsert
: {
1776 struct vtn_value
*comp
;
1777 unsigned deref_start
;
1778 struct nir_constant
**c
;
1779 if (opcode
== SpvOpCompositeExtract
) {
1780 comp
= vtn_value(b
, w
[4], vtn_value_type_constant
);
1782 c
= &comp
->constant
;
1784 comp
= vtn_value(b
, w
[5], vtn_value_type_constant
);
1786 val
->constant
= nir_constant_clone(comp
->constant
,
1792 const struct vtn_type
*type
= comp
->type
;
1793 for (unsigned i
= deref_start
; i
< count
; i
++) {
1794 vtn_fail_if(w
[i
] > type
->length
,
1795 "%uth index of %s is %u but the type has only "
1796 "%u elements", i
- deref_start
,
1797 spirv_op_to_string(opcode
), w
[i
], type
->length
);
1799 switch (type
->base_type
) {
1800 case vtn_base_type_vector
:
1802 type
= type
->array_element
;
1805 case vtn_base_type_matrix
:
1806 case vtn_base_type_array
:
1807 c
= &(*c
)->elements
[w
[i
]];
1808 type
= type
->array_element
;
1811 case vtn_base_type_struct
:
1812 c
= &(*c
)->elements
[w
[i
]];
1813 type
= type
->members
[w
[i
]];
1817 vtn_fail("%s must only index into composite types",
1818 spirv_op_to_string(opcode
));
1822 if (opcode
== SpvOpCompositeExtract
) {
1826 unsigned num_components
= type
->length
;
1827 for (unsigned i
= 0; i
< num_components
; i
++)
1828 val
->constant
->values
[i
] = (*c
)->values
[elem
+ i
];
1831 struct vtn_value
*insert
=
1832 vtn_value(b
, w
[4], vtn_value_type_constant
);
1833 vtn_assert(insert
->type
== type
);
1835 *c
= insert
->constant
;
1837 unsigned num_components
= type
->length
;
1838 for (unsigned i
= 0; i
< num_components
; i
++)
1839 (*c
)->values
[elem
+ i
] = insert
->constant
->values
[i
];
1847 nir_alu_type dst_alu_type
= nir_get_nir_type_for_glsl_type(val
->type
->type
);
1848 nir_alu_type src_alu_type
= dst_alu_type
;
1849 unsigned num_components
= glsl_get_vector_elements(val
->type
->type
);
1852 vtn_assert(count
<= 7);
1858 /* We have a source in a conversion */
1860 nir_get_nir_type_for_glsl_type(
1861 vtn_value(b
, w
[4], vtn_value_type_constant
)->type
->type
);
1862 /* We use the bitsize of the conversion source to evaluate the opcode later */
1863 bit_size
= glsl_get_bit_size(
1864 vtn_value(b
, w
[4], vtn_value_type_constant
)->type
->type
);
1867 bit_size
= glsl_get_bit_size(val
->type
->type
);
1870 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(b
, opcode
, &swap
,
1871 nir_alu_type_get_type_size(src_alu_type
),
1872 nir_alu_type_get_type_size(dst_alu_type
));
1873 nir_const_value src
[3][NIR_MAX_VEC_COMPONENTS
];
1875 for (unsigned i
= 0; i
< count
- 4; i
++) {
1876 struct vtn_value
*src_val
=
1877 vtn_value(b
, w
[4 + i
], vtn_value_type_constant
);
1879 /* If this is an unsized source, pull the bit size from the
1880 * source; otherwise, we'll use the bit size from the destination.
1882 if (!nir_alu_type_get_type_size(nir_op_infos
[op
].input_types
[i
]))
1883 bit_size
= glsl_get_bit_size(src_val
->type
->type
);
1885 unsigned src_comps
= nir_op_infos
[op
].input_sizes
[i
] ?
1886 nir_op_infos
[op
].input_sizes
[i
] :
1889 unsigned j
= swap
? 1 - i
: i
;
1890 for (unsigned c
= 0; c
< src_comps
; c
++)
1891 src
[j
][c
] = src_val
->constant
->values
[c
];
1894 /* fix up fixed size sources */
1901 for (unsigned i
= 0; i
< num_components
; ++i
) {
1903 case 64: src
[1][i
].u32
= src
[1][i
].u64
; break;
1904 case 16: src
[1][i
].u32
= src
[1][i
].u16
; break;
1905 case 8: src
[1][i
].u32
= src
[1][i
].u8
; break;
1914 nir_const_value
*srcs
[3] = {
1915 src
[0], src
[1], src
[2],
1917 nir_eval_const_opcode(op
, val
->constant
->values
,
1918 num_components
, bit_size
, srcs
,
1919 b
->shader
->info
.float_controls_execution_mode
);
1926 case SpvOpConstantNull
:
1927 val
->constant
= vtn_null_constant(b
, val
->type
);
1930 case SpvOpConstantSampler
:
1931 vtn_fail("OpConstantSampler requires Kernel Capability");
1935 vtn_fail_with_opcode("Unhandled opcode", opcode
);
1938 /* Now that we have the value, update the workgroup size if needed */
1939 vtn_foreach_decoration(b
, val
, handle_workgroup_size_decoration_cb
, NULL
);
1942 SpvMemorySemanticsMask
1943 vtn_storage_class_to_memory_semantics(SpvStorageClass sc
)
1946 case SpvStorageClassStorageBuffer
:
1947 case SpvStorageClassPhysicalStorageBuffer
:
1948 return SpvMemorySemanticsUniformMemoryMask
;
1949 case SpvStorageClassWorkgroup
:
1950 return SpvMemorySemanticsWorkgroupMemoryMask
;
1952 return SpvMemorySemanticsMaskNone
;
1957 vtn_split_barrier_semantics(struct vtn_builder
*b
,
1958 SpvMemorySemanticsMask semantics
,
1959 SpvMemorySemanticsMask
*before
,
1960 SpvMemorySemanticsMask
*after
)
1962 /* For memory semantics embedded in operations, we split them into up to
1963 * two barriers, to be added before and after the operation. This is less
1964 * strict than if we propagated until the final backend stage, but still
1965 * result in correct execution.
1967 * A further improvement could be pipe this information (and use!) into the
1968 * next compiler layers, at the expense of making the handling of barriers
1972 *before
= SpvMemorySemanticsMaskNone
;
1973 *after
= SpvMemorySemanticsMaskNone
;
1975 SpvMemorySemanticsMask order_semantics
=
1976 semantics
& (SpvMemorySemanticsAcquireMask
|
1977 SpvMemorySemanticsReleaseMask
|
1978 SpvMemorySemanticsAcquireReleaseMask
|
1979 SpvMemorySemanticsSequentiallyConsistentMask
);
1981 if (util_bitcount(order_semantics
) > 1) {
1982 /* Old GLSLang versions incorrectly set all the ordering bits. This was
1983 * fixed in c51287d744fb6e7e9ccc09f6f8451e6c64b1dad6 of glslang repo,
1984 * and it is in GLSLang since revision "SPIRV99.1321" (from Jul-2016).
1986 vtn_warn("Multiple memory ordering semantics specified, "
1987 "assuming AcquireRelease.");
1988 order_semantics
= SpvMemorySemanticsAcquireReleaseMask
;
1991 const SpvMemorySemanticsMask av_vis_semantics
=
1992 semantics
& (SpvMemorySemanticsMakeAvailableMask
|
1993 SpvMemorySemanticsMakeVisibleMask
);
1995 const SpvMemorySemanticsMask storage_semantics
=
1996 semantics
& (SpvMemorySemanticsUniformMemoryMask
|
1997 SpvMemorySemanticsSubgroupMemoryMask
|
1998 SpvMemorySemanticsWorkgroupMemoryMask
|
1999 SpvMemorySemanticsCrossWorkgroupMemoryMask
|
2000 SpvMemorySemanticsAtomicCounterMemoryMask
|
2001 SpvMemorySemanticsImageMemoryMask
|
2002 SpvMemorySemanticsOutputMemoryMask
);
2004 const SpvMemorySemanticsMask other_semantics
=
2005 semantics
& ~(order_semantics
| av_vis_semantics
| storage_semantics
);
2007 if (other_semantics
)
2008 vtn_warn("Ignoring unhandled memory semantics: %u\n", other_semantics
);
2010 /* SequentiallyConsistent is treated as AcquireRelease. */
2012 /* The RELEASE barrier happens BEFORE the operation, and it is usually
2013 * associated with a Store. All the write operations with a matching
2014 * semantics will not be reordered after the Store.
2016 if (order_semantics
& (SpvMemorySemanticsReleaseMask
|
2017 SpvMemorySemanticsAcquireReleaseMask
|
2018 SpvMemorySemanticsSequentiallyConsistentMask
)) {
2019 *before
|= SpvMemorySemanticsReleaseMask
| storage_semantics
;
2022 /* The ACQUIRE barrier happens AFTER the operation, and it is usually
2023 * associated with a Load. All the operations with a matching semantics
2024 * will not be reordered before the Load.
2026 if (order_semantics
& (SpvMemorySemanticsAcquireMask
|
2027 SpvMemorySemanticsAcquireReleaseMask
|
2028 SpvMemorySemanticsSequentiallyConsistentMask
)) {
2029 *after
|= SpvMemorySemanticsAcquireMask
| storage_semantics
;
2032 if (av_vis_semantics
& SpvMemorySemanticsMakeVisibleMask
)
2033 *before
|= SpvMemorySemanticsMakeVisibleMask
| storage_semantics
;
2035 if (av_vis_semantics
& SpvMemorySemanticsMakeAvailableMask
)
2036 *after
|= SpvMemorySemanticsMakeAvailableMask
| storage_semantics
;
2040 vtn_emit_scoped_memory_barrier(struct vtn_builder
*b
, SpvScope scope
,
2041 SpvMemorySemanticsMask semantics
)
2043 nir_memory_semantics nir_semantics
= 0;
2045 SpvMemorySemanticsMask order_semantics
=
2046 semantics
& (SpvMemorySemanticsAcquireMask
|
2047 SpvMemorySemanticsReleaseMask
|
2048 SpvMemorySemanticsAcquireReleaseMask
|
2049 SpvMemorySemanticsSequentiallyConsistentMask
);
2051 if (util_bitcount(order_semantics
) > 1) {
2052 /* Old GLSLang versions incorrectly set all the ordering bits. This was
2053 * fixed in c51287d744fb6e7e9ccc09f6f8451e6c64b1dad6 of glslang repo,
2054 * and it is in GLSLang since revision "SPIRV99.1321" (from Jul-2016).
2056 vtn_warn("Multiple memory ordering semantics bits specified, "
2057 "assuming AcquireRelease.");
2058 order_semantics
= SpvMemorySemanticsAcquireReleaseMask
;
2061 switch (order_semantics
) {
2063 /* Not an ordering barrier. */
2066 case SpvMemorySemanticsAcquireMask
:
2067 nir_semantics
= NIR_MEMORY_ACQUIRE
;
2070 case SpvMemorySemanticsReleaseMask
:
2071 nir_semantics
= NIR_MEMORY_RELEASE
;
2074 case SpvMemorySemanticsSequentiallyConsistentMask
:
2075 /* Fall through. Treated as AcquireRelease in Vulkan. */
2076 case SpvMemorySemanticsAcquireReleaseMask
:
2077 nir_semantics
= NIR_MEMORY_ACQUIRE
| NIR_MEMORY_RELEASE
;
2081 unreachable("Invalid memory order semantics");
2084 if (semantics
& SpvMemorySemanticsMakeAvailableMask
) {
2085 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
2086 "To use MakeAvailable memory semantics the VulkanMemoryModel "
2087 "capability must be declared.");
2088 nir_semantics
|= NIR_MEMORY_MAKE_AVAILABLE
;
2091 if (semantics
& SpvMemorySemanticsMakeVisibleMask
) {
2092 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
2093 "To use MakeVisible memory semantics the VulkanMemoryModel "
2094 "capability must be declared.");
2095 nir_semantics
|= NIR_MEMORY_MAKE_VISIBLE
;
2098 /* Vulkan Environment for SPIR-V says "SubgroupMemory, CrossWorkgroupMemory,
2099 * and AtomicCounterMemory are ignored".
2101 semantics
&= ~(SpvMemorySemanticsSubgroupMemoryMask
|
2102 SpvMemorySemanticsCrossWorkgroupMemoryMask
|
2103 SpvMemorySemanticsAtomicCounterMemoryMask
);
2105 /* TODO: Consider adding nir_var_mem_image mode to NIR so it can be used
2106 * for SpvMemorySemanticsImageMemoryMask.
2109 nir_variable_mode modes
= 0;
2110 if (semantics
& (SpvMemorySemanticsUniformMemoryMask
|
2111 SpvMemorySemanticsImageMemoryMask
)) {
2112 modes
|= nir_var_uniform
|
2117 if (semantics
& SpvMemorySemanticsWorkgroupMemoryMask
)
2118 modes
|= nir_var_mem_shared
;
2119 if (semantics
& SpvMemorySemanticsOutputMemoryMask
) {
2120 modes
|= nir_var_shader_out
;
2123 /* No barrier to add. */
2124 if (nir_semantics
== 0 || modes
== 0)
2127 nir_scope nir_scope
;
2129 case SpvScopeDevice
:
2130 vtn_fail_if(b
->options
->caps
.vk_memory_model
&&
2131 !b
->options
->caps
.vk_memory_model_device_scope
,
2132 "If the Vulkan memory model is declared and any instruction "
2133 "uses Device scope, the VulkanMemoryModelDeviceScope "
2134 "capability must be declared.");
2135 nir_scope
= NIR_SCOPE_DEVICE
;
2138 case SpvScopeQueueFamily
:
2139 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
2140 "To use Queue Family scope, the VulkanMemoryModel capability "
2141 "must be declared.");
2142 nir_scope
= NIR_SCOPE_QUEUE_FAMILY
;
2145 case SpvScopeWorkgroup
:
2146 nir_scope
= NIR_SCOPE_WORKGROUP
;
2149 case SpvScopeSubgroup
:
2150 nir_scope
= NIR_SCOPE_SUBGROUP
;
2153 case SpvScopeInvocation
:
2154 nir_scope
= NIR_SCOPE_INVOCATION
;
2158 vtn_fail("Invalid memory scope");
2161 nir_intrinsic_instr
*intrin
=
2162 nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_scoped_memory_barrier
);
2163 nir_intrinsic_set_memory_semantics(intrin
, nir_semantics
);
2165 nir_intrinsic_set_memory_modes(intrin
, modes
);
2166 nir_intrinsic_set_memory_scope(intrin
, nir_scope
);
2167 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2170 struct vtn_ssa_value
*
2171 vtn_create_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
2173 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
2176 if (!glsl_type_is_vector_or_scalar(type
)) {
2177 unsigned elems
= glsl_get_length(type
);
2178 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
2179 for (unsigned i
= 0; i
< elems
; i
++) {
2180 const struct glsl_type
*child_type
;
2182 switch (glsl_get_base_type(type
)) {
2184 case GLSL_TYPE_UINT
:
2185 case GLSL_TYPE_INT16
:
2186 case GLSL_TYPE_UINT16
:
2187 case GLSL_TYPE_UINT8
:
2188 case GLSL_TYPE_INT8
:
2189 case GLSL_TYPE_INT64
:
2190 case GLSL_TYPE_UINT64
:
2191 case GLSL_TYPE_BOOL
:
2192 case GLSL_TYPE_FLOAT
:
2193 case GLSL_TYPE_FLOAT16
:
2194 case GLSL_TYPE_DOUBLE
:
2195 child_type
= glsl_get_column_type(type
);
2197 case GLSL_TYPE_ARRAY
:
2198 child_type
= glsl_get_array_element(type
);
2200 case GLSL_TYPE_STRUCT
:
2201 case GLSL_TYPE_INTERFACE
:
2202 child_type
= glsl_get_struct_field(type
, i
);
2205 vtn_fail("unkown base type");
2208 val
->elems
[i
] = vtn_create_ssa_value(b
, child_type
);
2216 vtn_tex_src(struct vtn_builder
*b
, unsigned index
, nir_tex_src_type type
)
2219 src
.src
= nir_src_for_ssa(vtn_ssa_value(b
, index
)->def
);
2220 src
.src_type
= type
;
2225 image_operand_arg(struct vtn_builder
*b
, const uint32_t *w
, uint32_t count
,
2226 uint32_t mask_idx
, SpvImageOperandsMask op
)
2228 static const SpvImageOperandsMask ops_with_arg
=
2229 SpvImageOperandsBiasMask
|
2230 SpvImageOperandsLodMask
|
2231 SpvImageOperandsGradMask
|
2232 SpvImageOperandsConstOffsetMask
|
2233 SpvImageOperandsOffsetMask
|
2234 SpvImageOperandsConstOffsetsMask
|
2235 SpvImageOperandsSampleMask
|
2236 SpvImageOperandsMinLodMask
|
2237 SpvImageOperandsMakeTexelAvailableMask
|
2238 SpvImageOperandsMakeTexelVisibleMask
;
2240 assert(util_bitcount(op
) == 1);
2241 assert(w
[mask_idx
] & op
);
2242 assert(op
& ops_with_arg
);
2244 uint32_t idx
= util_bitcount(w
[mask_idx
] & (op
- 1) & ops_with_arg
) + 1;
2246 /* Adjust indices for operands with two arguments. */
2247 static const SpvImageOperandsMask ops_with_two_args
=
2248 SpvImageOperandsGradMask
;
2249 idx
+= util_bitcount(w
[mask_idx
] & (op
- 1) & ops_with_two_args
);
2253 vtn_fail_if(idx
+ (op
& ops_with_two_args
? 1 : 0) >= count
,
2254 "Image op claims to have %s but does not enough "
2255 "following operands", spirv_imageoperands_to_string(op
));
2261 vtn_handle_texture(struct vtn_builder
*b
, SpvOp opcode
,
2262 const uint32_t *w
, unsigned count
)
2264 if (opcode
== SpvOpSampledImage
) {
2265 struct vtn_value
*val
=
2266 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
2267 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
2268 val
->sampled_image
->image
=
2269 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2270 val
->sampled_image
->sampler
=
2271 vtn_value(b
, w
[4], vtn_value_type_pointer
)->pointer
;
2273 } else if (opcode
== SpvOpImage
) {
2274 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[3]);
2275 if (src_val
->value_type
== vtn_value_type_sampled_image
) {
2276 vtn_push_value_pointer(b
, w
[2], src_val
->sampled_image
->image
);
2278 vtn_assert(src_val
->value_type
== vtn_value_type_pointer
);
2279 vtn_push_value_pointer(b
, w
[2], src_val
->pointer
);
2284 struct vtn_type
*ret_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2286 struct vtn_pointer
*image
= NULL
, *sampler
= NULL
;
2287 struct vtn_value
*sampled_val
= vtn_untyped_value(b
, w
[3]);
2288 if (sampled_val
->value_type
== vtn_value_type_sampled_image
) {
2289 image
= sampled_val
->sampled_image
->image
;
2290 sampler
= sampled_val
->sampled_image
->sampler
;
2292 vtn_assert(sampled_val
->value_type
== vtn_value_type_pointer
);
2293 image
= sampled_val
->pointer
;
2296 nir_deref_instr
*image_deref
= vtn_pointer_to_deref(b
, image
);
2297 nir_deref_instr
*sampler_deref
=
2298 sampler
? vtn_pointer_to_deref(b
, sampler
) : NULL
;
2300 const struct glsl_type
*image_type
= sampled_val
->type
->type
;
2301 const enum glsl_sampler_dim sampler_dim
= glsl_get_sampler_dim(image_type
);
2302 const bool is_array
= glsl_sampler_type_is_array(image_type
);
2303 nir_alu_type dest_type
= nir_type_invalid
;
2305 /* Figure out the base texture operation */
2308 case SpvOpImageSampleImplicitLod
:
2309 case SpvOpImageSampleDrefImplicitLod
:
2310 case SpvOpImageSampleProjImplicitLod
:
2311 case SpvOpImageSampleProjDrefImplicitLod
:
2312 texop
= nir_texop_tex
;
2315 case SpvOpImageSampleExplicitLod
:
2316 case SpvOpImageSampleDrefExplicitLod
:
2317 case SpvOpImageSampleProjExplicitLod
:
2318 case SpvOpImageSampleProjDrefExplicitLod
:
2319 texop
= nir_texop_txl
;
2322 case SpvOpImageFetch
:
2323 if (sampler_dim
== GLSL_SAMPLER_DIM_MS
) {
2324 texop
= nir_texop_txf_ms
;
2326 texop
= nir_texop_txf
;
2330 case SpvOpImageGather
:
2331 case SpvOpImageDrefGather
:
2332 texop
= nir_texop_tg4
;
2335 case SpvOpImageQuerySizeLod
:
2336 case SpvOpImageQuerySize
:
2337 texop
= nir_texop_txs
;
2338 dest_type
= nir_type_int
;
2341 case SpvOpImageQueryLod
:
2342 texop
= nir_texop_lod
;
2343 dest_type
= nir_type_float
;
2346 case SpvOpImageQueryLevels
:
2347 texop
= nir_texop_query_levels
;
2348 dest_type
= nir_type_int
;
2351 case SpvOpImageQuerySamples
:
2352 texop
= nir_texop_texture_samples
;
2353 dest_type
= nir_type_int
;
2356 case SpvOpFragmentFetchAMD
:
2357 texop
= nir_texop_fragment_fetch
;
2360 case SpvOpFragmentMaskFetchAMD
:
2361 texop
= nir_texop_fragment_mask_fetch
;
2365 vtn_fail_with_opcode("Unhandled opcode", opcode
);
2368 nir_tex_src srcs
[10]; /* 10 should be enough */
2369 nir_tex_src
*p
= srcs
;
2371 p
->src
= nir_src_for_ssa(&image_deref
->dest
.ssa
);
2372 p
->src_type
= nir_tex_src_texture_deref
;
2382 vtn_fail_if(sampler
== NULL
,
2383 "%s requires an image of type OpTypeSampledImage",
2384 spirv_op_to_string(opcode
));
2385 p
->src
= nir_src_for_ssa(&sampler_deref
->dest
.ssa
);
2386 p
->src_type
= nir_tex_src_sampler_deref
;
2390 case nir_texop_txf_ms
:
2392 case nir_texop_query_levels
:
2393 case nir_texop_texture_samples
:
2394 case nir_texop_samples_identical
:
2395 case nir_texop_fragment_fetch
:
2396 case nir_texop_fragment_mask_fetch
:
2399 case nir_texop_txf_ms_fb
:
2400 vtn_fail("unexpected nir_texop_txf_ms_fb");
2402 case nir_texop_txf_ms_mcs
:
2403 vtn_fail("unexpected nir_texop_txf_ms_mcs");
2404 case nir_texop_tex_prefetch
:
2405 vtn_fail("unexpected nir_texop_tex_prefetch");
2410 struct nir_ssa_def
*coord
;
2411 unsigned coord_components
;
2413 case SpvOpImageSampleImplicitLod
:
2414 case SpvOpImageSampleExplicitLod
:
2415 case SpvOpImageSampleDrefImplicitLod
:
2416 case SpvOpImageSampleDrefExplicitLod
:
2417 case SpvOpImageSampleProjImplicitLod
:
2418 case SpvOpImageSampleProjExplicitLod
:
2419 case SpvOpImageSampleProjDrefImplicitLod
:
2420 case SpvOpImageSampleProjDrefExplicitLod
:
2421 case SpvOpImageFetch
:
2422 case SpvOpImageGather
:
2423 case SpvOpImageDrefGather
:
2424 case SpvOpImageQueryLod
:
2425 case SpvOpFragmentFetchAMD
:
2426 case SpvOpFragmentMaskFetchAMD
: {
2427 /* All these types have the coordinate as their first real argument */
2428 coord_components
= glsl_get_sampler_dim_coordinate_components(sampler_dim
);
2430 if (is_array
&& texop
!= nir_texop_lod
)
2433 coord
= vtn_ssa_value(b
, w
[idx
++])->def
;
2434 p
->src
= nir_src_for_ssa(nir_channels(&b
->nb
, coord
,
2435 (1 << coord_components
) - 1));
2436 p
->src_type
= nir_tex_src_coord
;
2443 coord_components
= 0;
2448 case SpvOpImageSampleProjImplicitLod
:
2449 case SpvOpImageSampleProjExplicitLod
:
2450 case SpvOpImageSampleProjDrefImplicitLod
:
2451 case SpvOpImageSampleProjDrefExplicitLod
:
2452 /* These have the projector as the last coordinate component */
2453 p
->src
= nir_src_for_ssa(nir_channel(&b
->nb
, coord
, coord_components
));
2454 p
->src_type
= nir_tex_src_projector
;
2462 bool is_shadow
= false;
2463 unsigned gather_component
= 0;
2465 case SpvOpImageSampleDrefImplicitLod
:
2466 case SpvOpImageSampleDrefExplicitLod
:
2467 case SpvOpImageSampleProjDrefImplicitLod
:
2468 case SpvOpImageSampleProjDrefExplicitLod
:
2469 case SpvOpImageDrefGather
:
2470 /* These all have an explicit depth value as their next source */
2472 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_comparator
);
2475 case SpvOpImageGather
:
2476 /* This has a component as its next source */
2477 gather_component
= vtn_constant_uint(b
, w
[idx
++]);
2484 /* For OpImageQuerySizeLod, we always have an LOD */
2485 if (opcode
== SpvOpImageQuerySizeLod
)
2486 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
2488 /* For OpFragmentFetchAMD, we always have a multisample index */
2489 if (opcode
== SpvOpFragmentFetchAMD
)
2490 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ms_index
);
2492 /* Now we need to handle some number of optional arguments */
2493 struct vtn_value
*gather_offsets
= NULL
;
2495 uint32_t operands
= w
[idx
];
2497 if (operands
& SpvImageOperandsBiasMask
) {
2498 vtn_assert(texop
== nir_texop_tex
);
2499 texop
= nir_texop_txb
;
2500 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2501 SpvImageOperandsBiasMask
);
2502 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_bias
);
2505 if (operands
& SpvImageOperandsLodMask
) {
2506 vtn_assert(texop
== nir_texop_txl
|| texop
== nir_texop_txf
||
2507 texop
== nir_texop_txs
);
2508 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2509 SpvImageOperandsLodMask
);
2510 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_lod
);
2513 if (operands
& SpvImageOperandsGradMask
) {
2514 vtn_assert(texop
== nir_texop_txl
);
2515 texop
= nir_texop_txd
;
2516 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2517 SpvImageOperandsGradMask
);
2518 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_ddx
);
2519 (*p
++) = vtn_tex_src(b
, w
[arg
+ 1], nir_tex_src_ddy
);
2522 vtn_fail_if(util_bitcount(operands
& (SpvImageOperandsConstOffsetsMask
|
2523 SpvImageOperandsOffsetMask
|
2524 SpvImageOperandsConstOffsetMask
)) > 1,
2525 "At most one of the ConstOffset, Offset, and ConstOffsets "
2526 "image operands can be used on a given instruction.");
2528 if (operands
& SpvImageOperandsOffsetMask
) {
2529 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2530 SpvImageOperandsOffsetMask
);
2531 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_offset
);
2534 if (operands
& SpvImageOperandsConstOffsetMask
) {
2535 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2536 SpvImageOperandsConstOffsetMask
);
2537 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_offset
);
2540 if (operands
& SpvImageOperandsConstOffsetsMask
) {
2541 vtn_assert(texop
== nir_texop_tg4
);
2542 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2543 SpvImageOperandsConstOffsetsMask
);
2544 gather_offsets
= vtn_value(b
, w
[arg
], vtn_value_type_constant
);
2547 if (operands
& SpvImageOperandsSampleMask
) {
2548 vtn_assert(texop
== nir_texop_txf_ms
);
2549 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2550 SpvImageOperandsSampleMask
);
2551 texop
= nir_texop_txf_ms
;
2552 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_ms_index
);
2555 if (operands
& SpvImageOperandsMinLodMask
) {
2556 vtn_assert(texop
== nir_texop_tex
||
2557 texop
== nir_texop_txb
||
2558 texop
== nir_texop_txd
);
2559 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2560 SpvImageOperandsMinLodMask
);
2561 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_min_lod
);
2565 nir_tex_instr
*instr
= nir_tex_instr_create(b
->shader
, p
- srcs
);
2568 memcpy(instr
->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
2570 instr
->coord_components
= coord_components
;
2571 instr
->sampler_dim
= sampler_dim
;
2572 instr
->is_array
= is_array
;
2573 instr
->is_shadow
= is_shadow
;
2574 instr
->is_new_style_shadow
=
2575 is_shadow
&& glsl_get_components(ret_type
->type
) == 1;
2576 instr
->component
= gather_component
;
2578 if (image
&& (image
->access
& ACCESS_NON_UNIFORM
))
2579 instr
->texture_non_uniform
= true;
2581 if (sampler
&& (sampler
->access
& ACCESS_NON_UNIFORM
))
2582 instr
->sampler_non_uniform
= true;
2584 /* for non-query ops, get dest_type from sampler type */
2585 if (dest_type
== nir_type_invalid
) {
2586 switch (glsl_get_sampler_result_type(image_type
)) {
2587 case GLSL_TYPE_FLOAT
: dest_type
= nir_type_float
; break;
2588 case GLSL_TYPE_INT
: dest_type
= nir_type_int
; break;
2589 case GLSL_TYPE_UINT
: dest_type
= nir_type_uint
; break;
2590 case GLSL_TYPE_BOOL
: dest_type
= nir_type_bool
; break;
2592 vtn_fail("Invalid base type for sampler result");
2596 instr
->dest_type
= dest_type
;
2598 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
2599 nir_tex_instr_dest_size(instr
), 32, NULL
);
2601 vtn_assert(glsl_get_vector_elements(ret_type
->type
) ==
2602 nir_tex_instr_dest_size(instr
));
2604 if (gather_offsets
) {
2605 vtn_fail_if(gather_offsets
->type
->base_type
!= vtn_base_type_array
||
2606 gather_offsets
->type
->length
!= 4,
2607 "ConstOffsets must be an array of size four of vectors "
2608 "of two integer components");
2610 struct vtn_type
*vec_type
= gather_offsets
->type
->array_element
;
2611 vtn_fail_if(vec_type
->base_type
!= vtn_base_type_vector
||
2612 vec_type
->length
!= 2 ||
2613 !glsl_type_is_integer(vec_type
->type
),
2614 "ConstOffsets must be an array of size four of vectors "
2615 "of two integer components");
2617 unsigned bit_size
= glsl_get_bit_size(vec_type
->type
);
2618 for (uint32_t i
= 0; i
< 4; i
++) {
2619 const nir_const_value
*cvec
=
2620 gather_offsets
->constant
->elements
[i
]->values
;
2621 for (uint32_t j
= 0; j
< 2; j
++) {
2623 case 8: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i8
; break;
2624 case 16: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i16
; break;
2625 case 32: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i32
; break;
2626 case 64: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i64
; break;
2628 vtn_fail("Unsupported bit size: %u", bit_size
);
2634 struct vtn_ssa_value
*ssa
= vtn_create_ssa_value(b
, ret_type
->type
);
2635 ssa
->def
= &instr
->dest
.ssa
;
2636 vtn_push_ssa(b
, w
[2], ret_type
, ssa
);
2638 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
2642 fill_common_atomic_sources(struct vtn_builder
*b
, SpvOp opcode
,
2643 const uint32_t *w
, nir_src
*src
)
2646 case SpvOpAtomicIIncrement
:
2647 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
2650 case SpvOpAtomicIDecrement
:
2651 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
2654 case SpvOpAtomicISub
:
2656 nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
2659 case SpvOpAtomicCompareExchange
:
2660 case SpvOpAtomicCompareExchangeWeak
:
2661 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[8])->def
);
2662 src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
2665 case SpvOpAtomicExchange
:
2666 case SpvOpAtomicIAdd
:
2667 case SpvOpAtomicSMin
:
2668 case SpvOpAtomicUMin
:
2669 case SpvOpAtomicSMax
:
2670 case SpvOpAtomicUMax
:
2671 case SpvOpAtomicAnd
:
2673 case SpvOpAtomicXor
:
2674 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
2678 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
2682 static nir_ssa_def
*
2683 get_image_coord(struct vtn_builder
*b
, uint32_t value
)
2685 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, value
);
2687 /* The image_load_store intrinsics assume a 4-dim coordinate */
2688 unsigned dim
= glsl_get_vector_elements(coord
->type
);
2689 unsigned swizzle
[4];
2690 for (unsigned i
= 0; i
< 4; i
++)
2691 swizzle
[i
] = MIN2(i
, dim
- 1);
2693 return nir_swizzle(&b
->nb
, coord
->def
, swizzle
, 4);
2696 static nir_ssa_def
*
2697 expand_to_vec4(nir_builder
*b
, nir_ssa_def
*value
)
2699 if (value
->num_components
== 4)
2703 for (unsigned i
= 0; i
< 4; i
++)
2704 swiz
[i
] = i
< value
->num_components
? i
: 0;
2705 return nir_swizzle(b
, value
, swiz
, 4);
2709 vtn_handle_image(struct vtn_builder
*b
, SpvOp opcode
,
2710 const uint32_t *w
, unsigned count
)
2712 /* Just get this one out of the way */
2713 if (opcode
== SpvOpImageTexelPointer
) {
2714 struct vtn_value
*val
=
2715 vtn_push_value(b
, w
[2], vtn_value_type_image_pointer
);
2716 val
->image
= ralloc(b
, struct vtn_image_pointer
);
2718 val
->image
->image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2719 val
->image
->coord
= get_image_coord(b
, w
[4]);
2720 val
->image
->sample
= vtn_ssa_value(b
, w
[5])->def
;
2721 val
->image
->lod
= nir_imm_int(&b
->nb
, 0);
2725 struct vtn_image_pointer image
;
2726 SpvScope scope
= SpvScopeInvocation
;
2727 SpvMemorySemanticsMask semantics
= 0;
2730 case SpvOpAtomicExchange
:
2731 case SpvOpAtomicCompareExchange
:
2732 case SpvOpAtomicCompareExchangeWeak
:
2733 case SpvOpAtomicIIncrement
:
2734 case SpvOpAtomicIDecrement
:
2735 case SpvOpAtomicIAdd
:
2736 case SpvOpAtomicISub
:
2737 case SpvOpAtomicLoad
:
2738 case SpvOpAtomicSMin
:
2739 case SpvOpAtomicUMin
:
2740 case SpvOpAtomicSMax
:
2741 case SpvOpAtomicUMax
:
2742 case SpvOpAtomicAnd
:
2744 case SpvOpAtomicXor
:
2745 image
= *vtn_value(b
, w
[3], vtn_value_type_image_pointer
)->image
;
2746 scope
= vtn_constant_uint(b
, w
[4]);
2747 semantics
= vtn_constant_uint(b
, w
[5]);
2750 case SpvOpAtomicStore
:
2751 image
= *vtn_value(b
, w
[1], vtn_value_type_image_pointer
)->image
;
2752 scope
= vtn_constant_uint(b
, w
[2]);
2753 semantics
= vtn_constant_uint(b
, w
[3]);
2756 case SpvOpImageQuerySize
:
2757 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2759 image
.sample
= NULL
;
2763 case SpvOpImageRead
: {
2764 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2765 image
.coord
= get_image_coord(b
, w
[4]);
2767 const SpvImageOperandsMask operands
=
2768 count
> 5 ? w
[5] : SpvImageOperandsMaskNone
;
2770 if (operands
& SpvImageOperandsSampleMask
) {
2771 uint32_t arg
= image_operand_arg(b
, w
, count
, 5,
2772 SpvImageOperandsSampleMask
);
2773 image
.sample
= vtn_ssa_value(b
, w
[arg
])->def
;
2775 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2778 if (operands
& SpvImageOperandsMakeTexelVisibleMask
) {
2779 vtn_fail_if((operands
& SpvImageOperandsNonPrivateTexelMask
) == 0,
2780 "MakeTexelVisible requires NonPrivateTexel to also be set.");
2781 uint32_t arg
= image_operand_arg(b
, w
, count
, 5,
2782 SpvImageOperandsMakeTexelVisibleMask
);
2783 semantics
= SpvMemorySemanticsMakeVisibleMask
;
2784 scope
= vtn_constant_uint(b
, w
[arg
]);
2787 if (operands
& SpvImageOperandsLodMask
) {
2788 uint32_t arg
= image_operand_arg(b
, w
, count
, 5,
2789 SpvImageOperandsLodMask
);
2790 image
.lod
= vtn_ssa_value(b
, w
[arg
])->def
;
2792 image
.lod
= nir_imm_int(&b
->nb
, 0);
2795 /* TODO: Volatile. */
2800 case SpvOpImageWrite
: {
2801 image
.image
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
2802 image
.coord
= get_image_coord(b
, w
[2]);
2806 const SpvImageOperandsMask operands
=
2807 count
> 4 ? w
[4] : SpvImageOperandsMaskNone
;
2809 if (operands
& SpvImageOperandsSampleMask
) {
2810 uint32_t arg
= image_operand_arg(b
, w
, count
, 4,
2811 SpvImageOperandsSampleMask
);
2812 image
.sample
= vtn_ssa_value(b
, w
[arg
])->def
;
2814 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2817 if (operands
& SpvImageOperandsMakeTexelAvailableMask
) {
2818 vtn_fail_if((operands
& SpvImageOperandsNonPrivateTexelMask
) == 0,
2819 "MakeTexelAvailable requires NonPrivateTexel to also be set.");
2820 uint32_t arg
= image_operand_arg(b
, w
, count
, 4,
2821 SpvImageOperandsMakeTexelAvailableMask
);
2822 semantics
= SpvMemorySemanticsMakeAvailableMask
;
2823 scope
= vtn_constant_uint(b
, w
[arg
]);
2826 if (operands
& SpvImageOperandsLodMask
) {
2827 uint32_t arg
= image_operand_arg(b
, w
, count
, 4,
2828 SpvImageOperandsLodMask
);
2829 image
.lod
= vtn_ssa_value(b
, w
[arg
])->def
;
2831 image
.lod
= nir_imm_int(&b
->nb
, 0);
2834 /* TODO: Volatile. */
2840 vtn_fail_with_opcode("Invalid image opcode", opcode
);
2843 nir_intrinsic_op op
;
2845 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_deref_##N; break;
2846 OP(ImageQuerySize
, size
)
2848 OP(ImageWrite
, store
)
2849 OP(AtomicLoad
, load
)
2850 OP(AtomicStore
, store
)
2851 OP(AtomicExchange
, atomic_exchange
)
2852 OP(AtomicCompareExchange
, atomic_comp_swap
)
2853 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
2854 OP(AtomicIIncrement
, atomic_add
)
2855 OP(AtomicIDecrement
, atomic_add
)
2856 OP(AtomicIAdd
, atomic_add
)
2857 OP(AtomicISub
, atomic_add
)
2858 OP(AtomicSMin
, atomic_imin
)
2859 OP(AtomicUMin
, atomic_umin
)
2860 OP(AtomicSMax
, atomic_imax
)
2861 OP(AtomicUMax
, atomic_umax
)
2862 OP(AtomicAnd
, atomic_and
)
2863 OP(AtomicOr
, atomic_or
)
2864 OP(AtomicXor
, atomic_xor
)
2867 vtn_fail_with_opcode("Invalid image opcode", opcode
);
2870 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
2872 nir_deref_instr
*image_deref
= vtn_pointer_to_deref(b
, image
.image
);
2873 intrin
->src
[0] = nir_src_for_ssa(&image_deref
->dest
.ssa
);
2875 /* ImageQuerySize doesn't take any extra parameters */
2876 if (opcode
!= SpvOpImageQuerySize
) {
2877 /* The image coordinate is always 4 components but we may not have that
2878 * many. Swizzle to compensate.
2880 intrin
->src
[1] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, image
.coord
));
2881 intrin
->src
[2] = nir_src_for_ssa(image
.sample
);
2884 nir_intrinsic_set_access(intrin
, image
.image
->access
);
2887 case SpvOpAtomicLoad
:
2888 case SpvOpImageQuerySize
:
2889 case SpvOpImageRead
:
2890 if (opcode
== SpvOpImageRead
|| opcode
== SpvOpAtomicLoad
) {
2891 /* Only OpImageRead can support a lod parameter if
2892 * SPV_AMD_shader_image_load_store_lod is used but the current NIR
2893 * intrinsics definition for atomics requires us to set it for
2896 intrin
->src
[3] = nir_src_for_ssa(image
.lod
);
2899 case SpvOpAtomicStore
:
2900 case SpvOpImageWrite
: {
2901 const uint32_t value_id
= opcode
== SpvOpAtomicStore
? w
[4] : w
[3];
2902 nir_ssa_def
*value
= vtn_ssa_value(b
, value_id
)->def
;
2903 /* nir_intrinsic_image_deref_store always takes a vec4 value */
2904 assert(op
== nir_intrinsic_image_deref_store
);
2905 intrin
->num_components
= 4;
2906 intrin
->src
[3] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, value
));
2907 /* Only OpImageWrite can support a lod parameter if
2908 * SPV_AMD_shader_image_load_store_lod is used but the current NIR
2909 * intrinsics definition for atomics requires us to set it for
2912 intrin
->src
[4] = nir_src_for_ssa(image
.lod
);
2916 case SpvOpAtomicCompareExchange
:
2917 case SpvOpAtomicCompareExchangeWeak
:
2918 case SpvOpAtomicIIncrement
:
2919 case SpvOpAtomicIDecrement
:
2920 case SpvOpAtomicExchange
:
2921 case SpvOpAtomicIAdd
:
2922 case SpvOpAtomicISub
:
2923 case SpvOpAtomicSMin
:
2924 case SpvOpAtomicUMin
:
2925 case SpvOpAtomicSMax
:
2926 case SpvOpAtomicUMax
:
2927 case SpvOpAtomicAnd
:
2929 case SpvOpAtomicXor
:
2930 fill_common_atomic_sources(b
, opcode
, w
, &intrin
->src
[3]);
2934 vtn_fail_with_opcode("Invalid image opcode", opcode
);
2937 /* Image operations implicitly have the Image storage memory semantics. */
2938 semantics
|= SpvMemorySemanticsImageMemoryMask
;
2940 SpvMemorySemanticsMask before_semantics
;
2941 SpvMemorySemanticsMask after_semantics
;
2942 vtn_split_barrier_semantics(b
, semantics
, &before_semantics
, &after_semantics
);
2944 if (before_semantics
)
2945 vtn_emit_memory_barrier(b
, scope
, before_semantics
);
2947 if (opcode
!= SpvOpImageWrite
&& opcode
!= SpvOpAtomicStore
) {
2948 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2950 unsigned dest_components
= glsl_get_vector_elements(type
->type
);
2951 intrin
->num_components
= nir_intrinsic_infos
[op
].dest_components
;
2952 if (intrin
->num_components
== 0)
2953 intrin
->num_components
= dest_components
;
2955 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
,
2956 intrin
->num_components
, 32, NULL
);
2958 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2960 nir_ssa_def
*result
= &intrin
->dest
.ssa
;
2961 if (intrin
->num_components
!= dest_components
)
2962 result
= nir_channels(&b
->nb
, result
, (1 << dest_components
) - 1);
2964 struct vtn_value
*val
=
2965 vtn_push_ssa(b
, w
[2], type
, vtn_create_ssa_value(b
, type
->type
));
2966 val
->ssa
->def
= result
;
2968 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2971 if (after_semantics
)
2972 vtn_emit_memory_barrier(b
, scope
, after_semantics
);
2975 static nir_intrinsic_op
2976 get_ssbo_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2979 case SpvOpAtomicLoad
: return nir_intrinsic_load_ssbo
;
2980 case SpvOpAtomicStore
: return nir_intrinsic_store_ssbo
;
2981 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
2982 OP(AtomicExchange
, atomic_exchange
)
2983 OP(AtomicCompareExchange
, atomic_comp_swap
)
2984 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
2985 OP(AtomicIIncrement
, atomic_add
)
2986 OP(AtomicIDecrement
, atomic_add
)
2987 OP(AtomicIAdd
, atomic_add
)
2988 OP(AtomicISub
, atomic_add
)
2989 OP(AtomicSMin
, atomic_imin
)
2990 OP(AtomicUMin
, atomic_umin
)
2991 OP(AtomicSMax
, atomic_imax
)
2992 OP(AtomicUMax
, atomic_umax
)
2993 OP(AtomicAnd
, atomic_and
)
2994 OP(AtomicOr
, atomic_or
)
2995 OP(AtomicXor
, atomic_xor
)
2998 vtn_fail_with_opcode("Invalid SSBO atomic", opcode
);
3002 static nir_intrinsic_op
3003 get_uniform_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
3006 #define OP(S, N) case SpvOp##S: return nir_intrinsic_atomic_counter_ ##N;
3007 OP(AtomicLoad
, read_deref
)
3008 OP(AtomicExchange
, exchange
)
3009 OP(AtomicCompareExchange
, comp_swap
)
3010 OP(AtomicCompareExchangeWeak
, comp_swap
)
3011 OP(AtomicIIncrement
, inc_deref
)
3012 OP(AtomicIDecrement
, post_dec_deref
)
3013 OP(AtomicIAdd
, add_deref
)
3014 OP(AtomicISub
, add_deref
)
3015 OP(AtomicUMin
, min_deref
)
3016 OP(AtomicUMax
, max_deref
)
3017 OP(AtomicAnd
, and_deref
)
3018 OP(AtomicOr
, or_deref
)
3019 OP(AtomicXor
, xor_deref
)
3022 /* We left the following out: AtomicStore, AtomicSMin and
3023 * AtomicSmax. Right now there are not nir intrinsics for them. At this
3024 * moment Atomic Counter support is needed for ARB_spirv support, so is
3025 * only need to support GLSL Atomic Counters that are uints and don't
3026 * allow direct storage.
3028 vtn_fail("Invalid uniform atomic");
3032 static nir_intrinsic_op
3033 get_deref_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
3036 case SpvOpAtomicLoad
: return nir_intrinsic_load_deref
;
3037 case SpvOpAtomicStore
: return nir_intrinsic_store_deref
;
3038 #define OP(S, N) case SpvOp##S: return nir_intrinsic_deref_##N;
3039 OP(AtomicExchange
, atomic_exchange
)
3040 OP(AtomicCompareExchange
, atomic_comp_swap
)
3041 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
3042 OP(AtomicIIncrement
, atomic_add
)
3043 OP(AtomicIDecrement
, atomic_add
)
3044 OP(AtomicIAdd
, atomic_add
)
3045 OP(AtomicISub
, atomic_add
)
3046 OP(AtomicSMin
, atomic_imin
)
3047 OP(AtomicUMin
, atomic_umin
)
3048 OP(AtomicSMax
, atomic_imax
)
3049 OP(AtomicUMax
, atomic_umax
)
3050 OP(AtomicAnd
, atomic_and
)
3051 OP(AtomicOr
, atomic_or
)
3052 OP(AtomicXor
, atomic_xor
)
3055 vtn_fail_with_opcode("Invalid shared atomic", opcode
);
3060 * Handles shared atomics, ssbo atomics and atomic counters.
3063 vtn_handle_atomics(struct vtn_builder
*b
, SpvOp opcode
,
3064 const uint32_t *w
, UNUSED
unsigned count
)
3066 struct vtn_pointer
*ptr
;
3067 nir_intrinsic_instr
*atomic
;
3069 SpvScope scope
= SpvScopeInvocation
;
3070 SpvMemorySemanticsMask semantics
= 0;
3073 case SpvOpAtomicLoad
:
3074 case SpvOpAtomicExchange
:
3075 case SpvOpAtomicCompareExchange
:
3076 case SpvOpAtomicCompareExchangeWeak
:
3077 case SpvOpAtomicIIncrement
:
3078 case SpvOpAtomicIDecrement
:
3079 case SpvOpAtomicIAdd
:
3080 case SpvOpAtomicISub
:
3081 case SpvOpAtomicSMin
:
3082 case SpvOpAtomicUMin
:
3083 case SpvOpAtomicSMax
:
3084 case SpvOpAtomicUMax
:
3085 case SpvOpAtomicAnd
:
3087 case SpvOpAtomicXor
:
3088 ptr
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
3089 scope
= vtn_constant_uint(b
, w
[4]);
3090 semantics
= vtn_constant_uint(b
, w
[5]);
3093 case SpvOpAtomicStore
:
3094 ptr
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
3095 scope
= vtn_constant_uint(b
, w
[2]);
3096 semantics
= vtn_constant_uint(b
, w
[3]);
3100 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
3103 /* uniform as "atomic counter uniform" */
3104 if (ptr
->mode
== vtn_variable_mode_uniform
) {
3105 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
3106 const struct glsl_type
*deref_type
= deref
->type
;
3107 nir_intrinsic_op op
= get_uniform_nir_atomic_op(b
, opcode
);
3108 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
3109 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
3111 /* SSBO needs to initialize index/offset. In this case we don't need to,
3112 * as that info is already stored on the ptr->var->var nir_variable (see
3113 * vtn_create_variable)
3117 case SpvOpAtomicLoad
:
3118 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
3121 case SpvOpAtomicStore
:
3122 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
3123 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
3126 case SpvOpAtomicExchange
:
3127 case SpvOpAtomicCompareExchange
:
3128 case SpvOpAtomicCompareExchangeWeak
:
3129 case SpvOpAtomicIIncrement
:
3130 case SpvOpAtomicIDecrement
:
3131 case SpvOpAtomicIAdd
:
3132 case SpvOpAtomicISub
:
3133 case SpvOpAtomicSMin
:
3134 case SpvOpAtomicUMin
:
3135 case SpvOpAtomicSMax
:
3136 case SpvOpAtomicUMax
:
3137 case SpvOpAtomicAnd
:
3139 case SpvOpAtomicXor
:
3140 /* Nothing: we don't need to call fill_common_atomic_sources here, as
3141 * atomic counter uniforms doesn't have sources
3146 unreachable("Invalid SPIR-V atomic");
3149 } else if (vtn_pointer_uses_ssa_offset(b
, ptr
)) {
3150 nir_ssa_def
*offset
, *index
;
3151 offset
= vtn_pointer_to_offset(b
, ptr
, &index
);
3153 assert(ptr
->mode
== vtn_variable_mode_ssbo
);
3155 nir_intrinsic_op op
= get_ssbo_nir_atomic_op(b
, opcode
);
3156 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
3160 case SpvOpAtomicLoad
:
3161 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
3162 nir_intrinsic_set_align(atomic
, 4, 0);
3163 if (ptr
->mode
== vtn_variable_mode_ssbo
)
3164 atomic
->src
[src
++] = nir_src_for_ssa(index
);
3165 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
3168 case SpvOpAtomicStore
:
3169 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
3170 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
3171 nir_intrinsic_set_align(atomic
, 4, 0);
3172 atomic
->src
[src
++] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
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 SpvOpAtomicExchange
:
3179 case SpvOpAtomicCompareExchange
:
3180 case SpvOpAtomicCompareExchangeWeak
:
3181 case SpvOpAtomicIIncrement
:
3182 case SpvOpAtomicIDecrement
:
3183 case SpvOpAtomicIAdd
:
3184 case SpvOpAtomicISub
:
3185 case SpvOpAtomicSMin
:
3186 case SpvOpAtomicUMin
:
3187 case SpvOpAtomicSMax
:
3188 case SpvOpAtomicUMax
:
3189 case SpvOpAtomicAnd
:
3191 case SpvOpAtomicXor
:
3192 if (ptr
->mode
== vtn_variable_mode_ssbo
)
3193 atomic
->src
[src
++] = nir_src_for_ssa(index
);
3194 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
3195 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[src
]);
3199 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
3202 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
3203 const struct glsl_type
*deref_type
= deref
->type
;
3204 nir_intrinsic_op op
= get_deref_nir_atomic_op(b
, opcode
);
3205 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
3206 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
3209 case SpvOpAtomicLoad
:
3210 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
3213 case SpvOpAtomicStore
:
3214 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
3215 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
3216 atomic
->src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
3219 case SpvOpAtomicExchange
:
3220 case SpvOpAtomicCompareExchange
:
3221 case SpvOpAtomicCompareExchangeWeak
:
3222 case SpvOpAtomicIIncrement
:
3223 case SpvOpAtomicIDecrement
:
3224 case SpvOpAtomicIAdd
:
3225 case SpvOpAtomicISub
:
3226 case SpvOpAtomicSMin
:
3227 case SpvOpAtomicUMin
:
3228 case SpvOpAtomicSMax
:
3229 case SpvOpAtomicUMax
:
3230 case SpvOpAtomicAnd
:
3232 case SpvOpAtomicXor
:
3233 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[1]);
3237 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
3241 /* Atomic ordering operations will implicitly apply to the atomic operation
3242 * storage class, so include that too.
3244 semantics
|= vtn_storage_class_to_memory_semantics(ptr
->ptr_type
->storage_class
);
3246 SpvMemorySemanticsMask before_semantics
;
3247 SpvMemorySemanticsMask after_semantics
;
3248 vtn_split_barrier_semantics(b
, semantics
, &before_semantics
, &after_semantics
);
3250 if (before_semantics
)
3251 vtn_emit_memory_barrier(b
, scope
, before_semantics
);
3253 if (opcode
!= SpvOpAtomicStore
) {
3254 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
3256 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
,
3257 glsl_get_vector_elements(type
->type
),
3258 glsl_get_bit_size(type
->type
), NULL
);
3260 struct vtn_ssa_value
*ssa
= rzalloc(b
, struct vtn_ssa_value
);
3261 ssa
->def
= &atomic
->dest
.ssa
;
3262 ssa
->type
= type
->type
;
3263 vtn_push_ssa(b
, w
[2], type
, ssa
);
3266 nir_builder_instr_insert(&b
->nb
, &atomic
->instr
);
3268 if (after_semantics
)
3269 vtn_emit_memory_barrier(b
, scope
, after_semantics
);
3272 static nir_alu_instr
*
3273 create_vec(struct vtn_builder
*b
, unsigned num_components
, unsigned bit_size
)
3275 nir_op op
= nir_op_vec(num_components
);
3276 nir_alu_instr
*vec
= nir_alu_instr_create(b
->shader
, op
);
3277 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, num_components
,
3279 vec
->dest
.write_mask
= (1 << num_components
) - 1;
3284 struct vtn_ssa_value
*
3285 vtn_ssa_transpose(struct vtn_builder
*b
, struct vtn_ssa_value
*src
)
3287 if (src
->transposed
)
3288 return src
->transposed
;
3290 struct vtn_ssa_value
*dest
=
3291 vtn_create_ssa_value(b
, glsl_transposed_type(src
->type
));
3293 for (unsigned i
= 0; i
< glsl_get_matrix_columns(dest
->type
); i
++) {
3294 nir_alu_instr
*vec
= create_vec(b
, glsl_get_matrix_columns(src
->type
),
3295 glsl_get_bit_size(src
->type
));
3296 if (glsl_type_is_vector_or_scalar(src
->type
)) {
3297 vec
->src
[0].src
= nir_src_for_ssa(src
->def
);
3298 vec
->src
[0].swizzle
[0] = i
;
3300 for (unsigned j
= 0; j
< glsl_get_matrix_columns(src
->type
); j
++) {
3301 vec
->src
[j
].src
= nir_src_for_ssa(src
->elems
[j
]->def
);
3302 vec
->src
[j
].swizzle
[0] = i
;
3305 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3306 dest
->elems
[i
]->def
= &vec
->dest
.dest
.ssa
;
3309 dest
->transposed
= src
;
3315 vtn_vector_extract(struct vtn_builder
*b
, nir_ssa_def
*src
, unsigned index
)
3317 return nir_channel(&b
->nb
, src
, index
);
3321 vtn_vector_insert(struct vtn_builder
*b
, nir_ssa_def
*src
, nir_ssa_def
*insert
,
3324 nir_alu_instr
*vec
= create_vec(b
, src
->num_components
,
3327 for (unsigned i
= 0; i
< src
->num_components
; i
++) {
3329 vec
->src
[i
].src
= nir_src_for_ssa(insert
);
3331 vec
->src
[i
].src
= nir_src_for_ssa(src
);
3332 vec
->src
[i
].swizzle
[0] = i
;
3336 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3338 return &vec
->dest
.dest
.ssa
;
3341 static nir_ssa_def
*
3342 nir_ieq_imm(nir_builder
*b
, nir_ssa_def
*x
, uint64_t i
)
3344 return nir_ieq(b
, x
, nir_imm_intN_t(b
, i
, x
->bit_size
));
3348 vtn_vector_extract_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
3351 return nir_vector_extract(&b
->nb
, src
, nir_i2i(&b
->nb
, index
, 32));
3355 vtn_vector_insert_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
3356 nir_ssa_def
*insert
, nir_ssa_def
*index
)
3358 nir_ssa_def
*dest
= vtn_vector_insert(b
, src
, insert
, 0);
3359 for (unsigned i
= 1; i
< src
->num_components
; i
++)
3360 dest
= nir_bcsel(&b
->nb
, nir_ieq_imm(&b
->nb
, index
, i
),
3361 vtn_vector_insert(b
, src
, insert
, i
), dest
);
3366 static nir_ssa_def
*
3367 vtn_vector_shuffle(struct vtn_builder
*b
, unsigned num_components
,
3368 nir_ssa_def
*src0
, nir_ssa_def
*src1
,
3369 const uint32_t *indices
)
3371 nir_alu_instr
*vec
= create_vec(b
, num_components
, src0
->bit_size
);
3373 for (unsigned i
= 0; i
< num_components
; i
++) {
3374 uint32_t index
= indices
[i
];
3375 if (index
== 0xffffffff) {
3377 nir_src_for_ssa(nir_ssa_undef(&b
->nb
, 1, src0
->bit_size
));
3378 } else if (index
< src0
->num_components
) {
3379 vec
->src
[i
].src
= nir_src_for_ssa(src0
);
3380 vec
->src
[i
].swizzle
[0] = index
;
3382 vec
->src
[i
].src
= nir_src_for_ssa(src1
);
3383 vec
->src
[i
].swizzle
[0] = index
- src0
->num_components
;
3387 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3389 return &vec
->dest
.dest
.ssa
;
3393 * Concatentates a number of vectors/scalars together to produce a vector
3395 static nir_ssa_def
*
3396 vtn_vector_construct(struct vtn_builder
*b
, unsigned num_components
,
3397 unsigned num_srcs
, nir_ssa_def
**srcs
)
3399 nir_alu_instr
*vec
= create_vec(b
, num_components
, srcs
[0]->bit_size
);
3401 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
3403 * "When constructing a vector, there must be at least two Constituent
3406 vtn_assert(num_srcs
>= 2);
3408 unsigned dest_idx
= 0;
3409 for (unsigned i
= 0; i
< num_srcs
; i
++) {
3410 nir_ssa_def
*src
= srcs
[i
];
3411 vtn_assert(dest_idx
+ src
->num_components
<= num_components
);
3412 for (unsigned j
= 0; j
< src
->num_components
; j
++) {
3413 vec
->src
[dest_idx
].src
= nir_src_for_ssa(src
);
3414 vec
->src
[dest_idx
].swizzle
[0] = j
;
3419 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
3421 * "When constructing a vector, the total number of components in all
3422 * the operands must equal the number of components in Result Type."
3424 vtn_assert(dest_idx
== num_components
);
3426 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3428 return &vec
->dest
.dest
.ssa
;
3431 static struct vtn_ssa_value
*
3432 vtn_composite_copy(void *mem_ctx
, struct vtn_ssa_value
*src
)
3434 struct vtn_ssa_value
*dest
= rzalloc(mem_ctx
, struct vtn_ssa_value
);
3435 dest
->type
= src
->type
;
3437 if (glsl_type_is_vector_or_scalar(src
->type
)) {
3438 dest
->def
= src
->def
;
3440 unsigned elems
= glsl_get_length(src
->type
);
3442 dest
->elems
= ralloc_array(mem_ctx
, struct vtn_ssa_value
*, elems
);
3443 for (unsigned i
= 0; i
< elems
; i
++)
3444 dest
->elems
[i
] = vtn_composite_copy(mem_ctx
, src
->elems
[i
]);
3450 static struct vtn_ssa_value
*
3451 vtn_composite_insert(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
3452 struct vtn_ssa_value
*insert
, const uint32_t *indices
,
3453 unsigned num_indices
)
3455 struct vtn_ssa_value
*dest
= vtn_composite_copy(b
, src
);
3457 struct vtn_ssa_value
*cur
= dest
;
3459 for (i
= 0; i
< num_indices
- 1; i
++) {
3460 cur
= cur
->elems
[indices
[i
]];
3463 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
3464 /* According to the SPIR-V spec, OpCompositeInsert may work down to
3465 * the component granularity. In that case, the last index will be
3466 * the index to insert the scalar into the vector.
3469 cur
->def
= vtn_vector_insert(b
, cur
->def
, insert
->def
, indices
[i
]);
3471 cur
->elems
[indices
[i
]] = insert
;
3477 static struct vtn_ssa_value
*
3478 vtn_composite_extract(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
3479 const uint32_t *indices
, unsigned num_indices
)
3481 struct vtn_ssa_value
*cur
= src
;
3482 for (unsigned i
= 0; i
< num_indices
; i
++) {
3483 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
3484 vtn_assert(i
== num_indices
- 1);
3485 /* According to the SPIR-V spec, OpCompositeExtract may work down to
3486 * the component granularity. The last index will be the index of the
3487 * vector to extract.
3490 struct vtn_ssa_value
*ret
= rzalloc(b
, struct vtn_ssa_value
);
3491 ret
->type
= glsl_scalar_type(glsl_get_base_type(cur
->type
));
3492 ret
->def
= vtn_vector_extract(b
, cur
->def
, indices
[i
]);
3495 cur
= cur
->elems
[indices
[i
]];
3503 vtn_handle_composite(struct vtn_builder
*b
, SpvOp opcode
,
3504 const uint32_t *w
, unsigned count
)
3506 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
3507 struct vtn_ssa_value
*ssa
= vtn_create_ssa_value(b
, type
->type
);
3510 case SpvOpVectorExtractDynamic
:
3511 ssa
->def
= vtn_vector_extract_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
3512 vtn_ssa_value(b
, w
[4])->def
);
3515 case SpvOpVectorInsertDynamic
:
3516 ssa
->def
= vtn_vector_insert_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
3517 vtn_ssa_value(b
, w
[4])->def
,
3518 vtn_ssa_value(b
, w
[5])->def
);
3521 case SpvOpVectorShuffle
:
3522 ssa
->def
= vtn_vector_shuffle(b
, glsl_get_vector_elements(type
->type
),
3523 vtn_ssa_value(b
, w
[3])->def
,
3524 vtn_ssa_value(b
, w
[4])->def
,
3528 case SpvOpCompositeConstruct
: {
3529 unsigned elems
= count
- 3;
3531 if (glsl_type_is_vector_or_scalar(type
->type
)) {
3532 nir_ssa_def
*srcs
[NIR_MAX_VEC_COMPONENTS
];
3533 for (unsigned i
= 0; i
< elems
; i
++)
3534 srcs
[i
] = vtn_ssa_value(b
, w
[3 + i
])->def
;
3536 vtn_vector_construct(b
, glsl_get_vector_elements(type
->type
),
3539 ssa
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
3540 for (unsigned i
= 0; i
< elems
; i
++)
3541 ssa
->elems
[i
] = vtn_ssa_value(b
, w
[3 + i
]);
3545 case SpvOpCompositeExtract
:
3546 ssa
= vtn_composite_extract(b
, vtn_ssa_value(b
, w
[3]),
3550 case SpvOpCompositeInsert
:
3551 ssa
= vtn_composite_insert(b
, vtn_ssa_value(b
, w
[4]),
3552 vtn_ssa_value(b
, w
[3]),
3556 case SpvOpCopyLogical
:
3557 case SpvOpCopyObject
:
3558 ssa
= vtn_composite_copy(b
, vtn_ssa_value(b
, w
[3]));
3562 vtn_fail_with_opcode("unknown composite operation", opcode
);
3565 vtn_push_ssa(b
, w
[2], type
, ssa
);
3569 vtn_emit_barrier(struct vtn_builder
*b
, nir_intrinsic_op op
)
3571 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
3572 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3576 vtn_emit_memory_barrier(struct vtn_builder
*b
, SpvScope scope
,
3577 SpvMemorySemanticsMask semantics
)
3579 if (b
->shader
->options
->use_scoped_memory_barrier
) {
3580 vtn_emit_scoped_memory_barrier(b
, scope
, semantics
);
3584 static const SpvMemorySemanticsMask all_memory_semantics
=
3585 SpvMemorySemanticsUniformMemoryMask
|
3586 SpvMemorySemanticsWorkgroupMemoryMask
|
3587 SpvMemorySemanticsAtomicCounterMemoryMask
|
3588 SpvMemorySemanticsImageMemoryMask
|
3589 SpvMemorySemanticsOutputMemoryMask
;
3591 /* If we're not actually doing a memory barrier, bail */
3592 if (!(semantics
& all_memory_semantics
))
3595 /* GL and Vulkan don't have these */
3596 vtn_assert(scope
!= SpvScopeCrossDevice
);
3598 if (scope
== SpvScopeSubgroup
)
3599 return; /* Nothing to do here */
3601 if (scope
== SpvScopeWorkgroup
) {
3602 vtn_emit_barrier(b
, nir_intrinsic_group_memory_barrier
);
3606 /* There's only two scopes thing left */
3607 vtn_assert(scope
== SpvScopeInvocation
|| scope
== SpvScopeDevice
);
3609 /* Map the GLSL memoryBarrier() construct to the corresponding NIR one. */
3610 static const SpvMemorySemanticsMask glsl_memory_barrier
=
3611 SpvMemorySemanticsUniformMemoryMask
|
3612 SpvMemorySemanticsWorkgroupMemoryMask
|
3613 SpvMemorySemanticsImageMemoryMask
;
3614 if ((semantics
& glsl_memory_barrier
) == glsl_memory_barrier
) {
3615 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier
);
3616 semantics
&= ~(glsl_memory_barrier
| SpvMemorySemanticsAtomicCounterMemoryMask
);
3619 /* Issue a bunch of more specific barriers */
3620 uint32_t bits
= semantics
;
3622 SpvMemorySemanticsMask semantic
= 1 << u_bit_scan(&bits
);
3624 case SpvMemorySemanticsUniformMemoryMask
:
3625 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_buffer
);
3627 case SpvMemorySemanticsWorkgroupMemoryMask
:
3628 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_shared
);
3630 case SpvMemorySemanticsAtomicCounterMemoryMask
:
3631 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_atomic_counter
);
3633 case SpvMemorySemanticsImageMemoryMask
:
3634 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_image
);
3636 case SpvMemorySemanticsOutputMemoryMask
:
3637 if (b
->nb
.shader
->info
.stage
== MESA_SHADER_TESS_CTRL
)
3638 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_tcs_patch
);
3647 vtn_handle_barrier(struct vtn_builder
*b
, SpvOp opcode
,
3648 const uint32_t *w
, UNUSED
unsigned count
)
3651 case SpvOpEmitVertex
:
3652 case SpvOpEmitStreamVertex
:
3653 case SpvOpEndPrimitive
:
3654 case SpvOpEndStreamPrimitive
: {
3655 nir_intrinsic_op intrinsic_op
;
3657 case SpvOpEmitVertex
:
3658 case SpvOpEmitStreamVertex
:
3659 intrinsic_op
= nir_intrinsic_emit_vertex
;
3661 case SpvOpEndPrimitive
:
3662 case SpvOpEndStreamPrimitive
:
3663 intrinsic_op
= nir_intrinsic_end_primitive
;
3666 unreachable("Invalid opcode");
3669 nir_intrinsic_instr
*intrin
=
3670 nir_intrinsic_instr_create(b
->shader
, intrinsic_op
);
3673 case SpvOpEmitStreamVertex
:
3674 case SpvOpEndStreamPrimitive
: {
3675 unsigned stream
= vtn_constant_uint(b
, w
[1]);
3676 nir_intrinsic_set_stream_id(intrin
, stream
);
3684 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3688 case SpvOpMemoryBarrier
: {
3689 SpvScope scope
= vtn_constant_uint(b
, w
[1]);
3690 SpvMemorySemanticsMask semantics
= vtn_constant_uint(b
, w
[2]);
3691 vtn_emit_memory_barrier(b
, scope
, semantics
);
3695 case SpvOpControlBarrier
: {
3696 SpvScope execution_scope
= vtn_constant_uint(b
, w
[1]);
3697 SpvScope memory_scope
= vtn_constant_uint(b
, w
[2]);
3698 SpvMemorySemanticsMask memory_semantics
= vtn_constant_uint(b
, w
[3]);
3700 /* GLSLang, prior to commit 8297936dd6eb3, emitted OpControlBarrier with
3701 * memory semantics of None for GLSL barrier().
3703 if (b
->wa_glslang_cs_barrier
&&
3704 b
->nb
.shader
->info
.stage
== MESA_SHADER_COMPUTE
&&
3705 execution_scope
== SpvScopeWorkgroup
&&
3706 memory_semantics
== SpvMemorySemanticsMaskNone
) {
3707 memory_scope
= SpvScopeWorkgroup
;
3708 memory_semantics
= SpvMemorySemanticsAcquireReleaseMask
|
3709 SpvMemorySemanticsWorkgroupMemoryMask
;
3712 /* From the SPIR-V spec:
3714 * "When used with the TessellationControl execution model, it also
3715 * implicitly synchronizes the Output Storage Class: Writes to Output
3716 * variables performed by any invocation executed prior to a
3717 * OpControlBarrier will be visible to any other invocation after
3718 * return from that OpControlBarrier."
3720 if (b
->nb
.shader
->info
.stage
== MESA_SHADER_TESS_CTRL
) {
3721 memory_semantics
&= ~(SpvMemorySemanticsAcquireMask
|
3722 SpvMemorySemanticsReleaseMask
|
3723 SpvMemorySemanticsAcquireReleaseMask
|
3724 SpvMemorySemanticsSequentiallyConsistentMask
);
3725 memory_semantics
|= SpvMemorySemanticsAcquireReleaseMask
|
3726 SpvMemorySemanticsOutputMemoryMask
;
3729 vtn_emit_memory_barrier(b
, memory_scope
, memory_semantics
);
3731 if (execution_scope
== SpvScopeWorkgroup
)
3732 vtn_emit_barrier(b
, nir_intrinsic_control_barrier
);
3737 unreachable("unknown barrier instruction");
3742 gl_primitive_from_spv_execution_mode(struct vtn_builder
*b
,
3743 SpvExecutionMode mode
)
3746 case SpvExecutionModeInputPoints
:
3747 case SpvExecutionModeOutputPoints
:
3748 return 0; /* GL_POINTS */
3749 case SpvExecutionModeInputLines
:
3750 return 1; /* GL_LINES */
3751 case SpvExecutionModeInputLinesAdjacency
:
3752 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
3753 case SpvExecutionModeTriangles
:
3754 return 4; /* GL_TRIANGLES */
3755 case SpvExecutionModeInputTrianglesAdjacency
:
3756 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
3757 case SpvExecutionModeQuads
:
3758 return 7; /* GL_QUADS */
3759 case SpvExecutionModeIsolines
:
3760 return 0x8E7A; /* GL_ISOLINES */
3761 case SpvExecutionModeOutputLineStrip
:
3762 return 3; /* GL_LINE_STRIP */
3763 case SpvExecutionModeOutputTriangleStrip
:
3764 return 5; /* GL_TRIANGLE_STRIP */
3766 vtn_fail("Invalid primitive type: %s (%u)",
3767 spirv_executionmode_to_string(mode
), mode
);
3772 vertices_in_from_spv_execution_mode(struct vtn_builder
*b
,
3773 SpvExecutionMode mode
)
3776 case SpvExecutionModeInputPoints
:
3778 case SpvExecutionModeInputLines
:
3780 case SpvExecutionModeInputLinesAdjacency
:
3782 case SpvExecutionModeTriangles
:
3784 case SpvExecutionModeInputTrianglesAdjacency
:
3787 vtn_fail("Invalid GS input mode: %s (%u)",
3788 spirv_executionmode_to_string(mode
), mode
);
3792 static gl_shader_stage
3793 stage_for_execution_model(struct vtn_builder
*b
, SpvExecutionModel model
)
3796 case SpvExecutionModelVertex
:
3797 return MESA_SHADER_VERTEX
;
3798 case SpvExecutionModelTessellationControl
:
3799 return MESA_SHADER_TESS_CTRL
;
3800 case SpvExecutionModelTessellationEvaluation
:
3801 return MESA_SHADER_TESS_EVAL
;
3802 case SpvExecutionModelGeometry
:
3803 return MESA_SHADER_GEOMETRY
;
3804 case SpvExecutionModelFragment
:
3805 return MESA_SHADER_FRAGMENT
;
3806 case SpvExecutionModelGLCompute
:
3807 return MESA_SHADER_COMPUTE
;
3808 case SpvExecutionModelKernel
:
3809 return MESA_SHADER_KERNEL
;
3811 vtn_fail("Unsupported execution model: %s (%u)",
3812 spirv_executionmodel_to_string(model
), model
);
3816 #define spv_check_supported(name, cap) do { \
3817 if (!(b->options && b->options->caps.name)) \
3818 vtn_warn("Unsupported SPIR-V capability: %s (%u)", \
3819 spirv_capability_to_string(cap), cap); \
3824 vtn_handle_entry_point(struct vtn_builder
*b
, const uint32_t *w
,
3827 struct vtn_value
*entry_point
= &b
->values
[w
[2]];
3828 /* Let this be a name label regardless */
3829 unsigned name_words
;
3830 entry_point
->name
= vtn_string_literal(b
, &w
[3], count
- 3, &name_words
);
3832 if (strcmp(entry_point
->name
, b
->entry_point_name
) != 0 ||
3833 stage_for_execution_model(b
, w
[1]) != b
->entry_point_stage
)
3836 vtn_assert(b
->entry_point
== NULL
);
3837 b
->entry_point
= entry_point
;
3841 vtn_handle_preamble_instruction(struct vtn_builder
*b
, SpvOp opcode
,
3842 const uint32_t *w
, unsigned count
)
3849 case SpvSourceLanguageUnknown
: lang
= "unknown"; break;
3850 case SpvSourceLanguageESSL
: lang
= "ESSL"; break;
3851 case SpvSourceLanguageGLSL
: lang
= "GLSL"; break;
3852 case SpvSourceLanguageOpenCL_C
: lang
= "OpenCL C"; break;
3853 case SpvSourceLanguageOpenCL_CPP
: lang
= "OpenCL C++"; break;
3854 case SpvSourceLanguageHLSL
: lang
= "HLSL"; break;
3857 uint32_t version
= w
[2];
3860 (count
> 3) ? vtn_value(b
, w
[3], vtn_value_type_string
)->str
: "";
3862 vtn_info("Parsing SPIR-V from %s %u source file %s", lang
, version
, file
);
3866 case SpvOpSourceExtension
:
3867 case SpvOpSourceContinued
:
3868 case SpvOpExtension
:
3869 case SpvOpModuleProcessed
:
3870 /* Unhandled, but these are for debug so that's ok. */
3873 case SpvOpCapability
: {
3874 SpvCapability cap
= w
[1];
3876 case SpvCapabilityMatrix
:
3877 case SpvCapabilityShader
:
3878 case SpvCapabilityGeometry
:
3879 case SpvCapabilityGeometryPointSize
:
3880 case SpvCapabilityUniformBufferArrayDynamicIndexing
:
3881 case SpvCapabilitySampledImageArrayDynamicIndexing
:
3882 case SpvCapabilityStorageBufferArrayDynamicIndexing
:
3883 case SpvCapabilityStorageImageArrayDynamicIndexing
:
3884 case SpvCapabilityImageRect
:
3885 case SpvCapabilitySampledRect
:
3886 case SpvCapabilitySampled1D
:
3887 case SpvCapabilityImage1D
:
3888 case SpvCapabilitySampledCubeArray
:
3889 case SpvCapabilityImageCubeArray
:
3890 case SpvCapabilitySampledBuffer
:
3891 case SpvCapabilityImageBuffer
:
3892 case SpvCapabilityImageQuery
:
3893 case SpvCapabilityDerivativeControl
:
3894 case SpvCapabilityInterpolationFunction
:
3895 case SpvCapabilityMultiViewport
:
3896 case SpvCapabilitySampleRateShading
:
3897 case SpvCapabilityClipDistance
:
3898 case SpvCapabilityCullDistance
:
3899 case SpvCapabilityInputAttachment
:
3900 case SpvCapabilityImageGatherExtended
:
3901 case SpvCapabilityStorageImageExtendedFormats
:
3902 case SpvCapabilityVector16
:
3905 case SpvCapabilityLinkage
:
3906 case SpvCapabilityFloat16Buffer
:
3907 case SpvCapabilitySparseResidency
:
3908 vtn_warn("Unsupported SPIR-V capability: %s",
3909 spirv_capability_to_string(cap
));
3912 case SpvCapabilityMinLod
:
3913 spv_check_supported(min_lod
, cap
);
3916 case SpvCapabilityAtomicStorage
:
3917 spv_check_supported(atomic_storage
, cap
);
3920 case SpvCapabilityFloat64
:
3921 spv_check_supported(float64
, cap
);
3923 case SpvCapabilityInt64
:
3924 spv_check_supported(int64
, cap
);
3926 case SpvCapabilityInt16
:
3927 spv_check_supported(int16
, cap
);
3929 case SpvCapabilityInt8
:
3930 spv_check_supported(int8
, cap
);
3933 case SpvCapabilityTransformFeedback
:
3934 spv_check_supported(transform_feedback
, cap
);
3937 case SpvCapabilityGeometryStreams
:
3938 spv_check_supported(geometry_streams
, cap
);
3941 case SpvCapabilityInt64Atomics
:
3942 spv_check_supported(int64_atomics
, cap
);
3945 case SpvCapabilityStorageImageMultisample
:
3946 spv_check_supported(storage_image_ms
, cap
);
3949 case SpvCapabilityAddresses
:
3950 spv_check_supported(address
, cap
);
3953 case SpvCapabilityKernel
:
3954 spv_check_supported(kernel
, cap
);
3957 case SpvCapabilityImageBasic
:
3958 case SpvCapabilityImageReadWrite
:
3959 case SpvCapabilityImageMipmap
:
3960 case SpvCapabilityPipes
:
3961 case SpvCapabilityDeviceEnqueue
:
3962 case SpvCapabilityLiteralSampler
:
3963 case SpvCapabilityGenericPointer
:
3964 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
3965 spirv_capability_to_string(cap
));
3968 case SpvCapabilityImageMSArray
:
3969 spv_check_supported(image_ms_array
, cap
);
3972 case SpvCapabilityTessellation
:
3973 case SpvCapabilityTessellationPointSize
:
3974 spv_check_supported(tessellation
, cap
);
3977 case SpvCapabilityDrawParameters
:
3978 spv_check_supported(draw_parameters
, cap
);
3981 case SpvCapabilityStorageImageReadWithoutFormat
:
3982 spv_check_supported(image_read_without_format
, cap
);
3985 case SpvCapabilityStorageImageWriteWithoutFormat
:
3986 spv_check_supported(image_write_without_format
, cap
);
3989 case SpvCapabilityDeviceGroup
:
3990 spv_check_supported(device_group
, cap
);
3993 case SpvCapabilityMultiView
:
3994 spv_check_supported(multiview
, cap
);
3997 case SpvCapabilityGroupNonUniform
:
3998 spv_check_supported(subgroup_basic
, cap
);
4001 case SpvCapabilitySubgroupVoteKHR
:
4002 case SpvCapabilityGroupNonUniformVote
:
4003 spv_check_supported(subgroup_vote
, cap
);
4006 case SpvCapabilitySubgroupBallotKHR
:
4007 case SpvCapabilityGroupNonUniformBallot
:
4008 spv_check_supported(subgroup_ballot
, cap
);
4011 case SpvCapabilityGroupNonUniformShuffle
:
4012 case SpvCapabilityGroupNonUniformShuffleRelative
:
4013 spv_check_supported(subgroup_shuffle
, cap
);
4016 case SpvCapabilityGroupNonUniformQuad
:
4017 spv_check_supported(subgroup_quad
, cap
);
4020 case SpvCapabilityGroupNonUniformArithmetic
:
4021 case SpvCapabilityGroupNonUniformClustered
:
4022 spv_check_supported(subgroup_arithmetic
, cap
);
4025 case SpvCapabilityGroups
:
4026 spv_check_supported(amd_shader_ballot
, cap
);
4029 case SpvCapabilityVariablePointersStorageBuffer
:
4030 case SpvCapabilityVariablePointers
:
4031 spv_check_supported(variable_pointers
, cap
);
4032 b
->variable_pointers
= true;
4035 case SpvCapabilityStorageUniformBufferBlock16
:
4036 case SpvCapabilityStorageUniform16
:
4037 case SpvCapabilityStoragePushConstant16
:
4038 case SpvCapabilityStorageInputOutput16
:
4039 spv_check_supported(storage_16bit
, cap
);
4042 case SpvCapabilityShaderLayer
:
4043 case SpvCapabilityShaderViewportIndex
:
4044 case SpvCapabilityShaderViewportIndexLayerEXT
:
4045 spv_check_supported(shader_viewport_index_layer
, cap
);
4048 case SpvCapabilityStorageBuffer8BitAccess
:
4049 case SpvCapabilityUniformAndStorageBuffer8BitAccess
:
4050 case SpvCapabilityStoragePushConstant8
:
4051 spv_check_supported(storage_8bit
, cap
);
4054 case SpvCapabilityShaderNonUniformEXT
:
4055 spv_check_supported(descriptor_indexing
, cap
);
4058 case SpvCapabilityInputAttachmentArrayDynamicIndexingEXT
:
4059 case SpvCapabilityUniformTexelBufferArrayDynamicIndexingEXT
:
4060 case SpvCapabilityStorageTexelBufferArrayDynamicIndexingEXT
:
4061 spv_check_supported(descriptor_array_dynamic_indexing
, cap
);
4064 case SpvCapabilityUniformBufferArrayNonUniformIndexingEXT
:
4065 case SpvCapabilitySampledImageArrayNonUniformIndexingEXT
:
4066 case SpvCapabilityStorageBufferArrayNonUniformIndexingEXT
:
4067 case SpvCapabilityStorageImageArrayNonUniformIndexingEXT
:
4068 case SpvCapabilityInputAttachmentArrayNonUniformIndexingEXT
:
4069 case SpvCapabilityUniformTexelBufferArrayNonUniformIndexingEXT
:
4070 case SpvCapabilityStorageTexelBufferArrayNonUniformIndexingEXT
:
4071 spv_check_supported(descriptor_array_non_uniform_indexing
, cap
);
4074 case SpvCapabilityRuntimeDescriptorArrayEXT
:
4075 spv_check_supported(runtime_descriptor_array
, cap
);
4078 case SpvCapabilityStencilExportEXT
:
4079 spv_check_supported(stencil_export
, cap
);
4082 case SpvCapabilitySampleMaskPostDepthCoverage
:
4083 spv_check_supported(post_depth_coverage
, cap
);
4086 case SpvCapabilityDenormFlushToZero
:
4087 case SpvCapabilityDenormPreserve
:
4088 case SpvCapabilitySignedZeroInfNanPreserve
:
4089 case SpvCapabilityRoundingModeRTE
:
4090 case SpvCapabilityRoundingModeRTZ
:
4091 spv_check_supported(float_controls
, cap
);
4094 case SpvCapabilityPhysicalStorageBufferAddresses
:
4095 spv_check_supported(physical_storage_buffer_address
, cap
);
4098 case SpvCapabilityComputeDerivativeGroupQuadsNV
:
4099 case SpvCapabilityComputeDerivativeGroupLinearNV
:
4100 spv_check_supported(derivative_group
, cap
);
4103 case SpvCapabilityFloat16
:
4104 spv_check_supported(float16
, cap
);
4107 case SpvCapabilityFragmentShaderSampleInterlockEXT
:
4108 spv_check_supported(fragment_shader_sample_interlock
, cap
);
4111 case SpvCapabilityFragmentShaderPixelInterlockEXT
:
4112 spv_check_supported(fragment_shader_pixel_interlock
, cap
);
4115 case SpvCapabilityDemoteToHelperInvocationEXT
:
4116 spv_check_supported(demote_to_helper_invocation
, cap
);
4119 case SpvCapabilityShaderClockKHR
:
4120 spv_check_supported(shader_clock
, cap
);
4123 case SpvCapabilityVulkanMemoryModel
:
4124 spv_check_supported(vk_memory_model
, cap
);
4127 case SpvCapabilityVulkanMemoryModelDeviceScope
:
4128 spv_check_supported(vk_memory_model_device_scope
, cap
);
4131 case SpvCapabilityImageReadWriteLodAMD
:
4132 spv_check_supported(amd_image_read_write_lod
, cap
);
4135 case SpvCapabilityIntegerFunctions2INTEL
:
4136 spv_check_supported(integer_functions2
, cap
);
4139 case SpvCapabilityFragmentMaskAMD
:
4140 spv_check_supported(amd_fragment_mask
, cap
);
4144 vtn_fail("Unhandled capability: %s (%u)",
4145 spirv_capability_to_string(cap
), cap
);
4150 case SpvOpExtInstImport
:
4151 vtn_handle_extension(b
, opcode
, w
, count
);
4154 case SpvOpMemoryModel
:
4156 case SpvAddressingModelPhysical32
:
4157 vtn_fail_if(b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
,
4158 "AddressingModelPhysical32 only supported for kernels");
4159 b
->shader
->info
.cs
.ptr_size
= 32;
4160 b
->physical_ptrs
= true;
4161 b
->options
->shared_addr_format
= nir_address_format_32bit_global
;
4162 b
->options
->global_addr_format
= nir_address_format_32bit_global
;
4163 b
->options
->temp_addr_format
= nir_address_format_32bit_global
;
4165 case SpvAddressingModelPhysical64
:
4166 vtn_fail_if(b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
,
4167 "AddressingModelPhysical64 only supported for kernels");
4168 b
->shader
->info
.cs
.ptr_size
= 64;
4169 b
->physical_ptrs
= true;
4170 b
->options
->shared_addr_format
= nir_address_format_64bit_global
;
4171 b
->options
->global_addr_format
= nir_address_format_64bit_global
;
4172 b
->options
->temp_addr_format
= nir_address_format_64bit_global
;
4174 case SpvAddressingModelLogical
:
4175 vtn_fail_if(b
->shader
->info
.stage
== MESA_SHADER_KERNEL
,
4176 "AddressingModelLogical only supported for shaders");
4177 b
->physical_ptrs
= false;
4179 case SpvAddressingModelPhysicalStorageBuffer64
:
4180 vtn_fail_if(!b
->options
||
4181 !b
->options
->caps
.physical_storage_buffer_address
,
4182 "AddressingModelPhysicalStorageBuffer64 not supported");
4185 vtn_fail("Unknown addressing model: %s (%u)",
4186 spirv_addressingmodel_to_string(w
[1]), w
[1]);
4191 case SpvMemoryModelSimple
:
4192 case SpvMemoryModelGLSL450
:
4193 case SpvMemoryModelOpenCL
:
4195 case SpvMemoryModelVulkan
:
4196 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
4197 "Vulkan memory model is unsupported by this driver");
4200 vtn_fail("Unsupported memory model: %s",
4201 spirv_memorymodel_to_string(w
[2]));
4206 case SpvOpEntryPoint
:
4207 vtn_handle_entry_point(b
, w
, count
);
4211 vtn_push_value(b
, w
[1], vtn_value_type_string
)->str
=
4212 vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
4216 b
->values
[w
[1]].name
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
4219 case SpvOpMemberName
:
4223 case SpvOpExecutionMode
:
4224 case SpvOpExecutionModeId
:
4225 case SpvOpDecorationGroup
:
4227 case SpvOpDecorateId
:
4228 case SpvOpMemberDecorate
:
4229 case SpvOpGroupDecorate
:
4230 case SpvOpGroupMemberDecorate
:
4231 case SpvOpDecorateString
:
4232 case SpvOpMemberDecorateString
:
4233 vtn_handle_decoration(b
, opcode
, w
, count
);
4236 case SpvOpExtInst
: {
4237 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
4238 if (val
->ext_handler
== vtn_handle_non_semantic_instruction
) {
4239 /* NonSemantic extended instructions are acceptable in preamble. */
4240 vtn_handle_non_semantic_instruction(b
, w
[4], w
, count
);
4243 return false; /* End of preamble. */
4248 return false; /* End of preamble */
4255 vtn_handle_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
4256 const struct vtn_decoration
*mode
, UNUSED
void *data
)
4258 vtn_assert(b
->entry_point
== entry_point
);
4260 switch(mode
->exec_mode
) {
4261 case SpvExecutionModeOriginUpperLeft
:
4262 case SpvExecutionModeOriginLowerLeft
:
4263 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4264 b
->shader
->info
.fs
.origin_upper_left
=
4265 (mode
->exec_mode
== SpvExecutionModeOriginUpperLeft
);
4268 case SpvExecutionModeEarlyFragmentTests
:
4269 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4270 b
->shader
->info
.fs
.early_fragment_tests
= true;
4273 case SpvExecutionModePostDepthCoverage
:
4274 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4275 b
->shader
->info
.fs
.post_depth_coverage
= true;
4278 case SpvExecutionModeInvocations
:
4279 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4280 b
->shader
->info
.gs
.invocations
= MAX2(1, mode
->operands
[0]);
4283 case SpvExecutionModeDepthReplacing
:
4284 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4285 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_ANY
;
4287 case SpvExecutionModeDepthGreater
:
4288 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4289 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_GREATER
;
4291 case SpvExecutionModeDepthLess
:
4292 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4293 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_LESS
;
4295 case SpvExecutionModeDepthUnchanged
:
4296 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4297 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
4300 case SpvExecutionModeLocalSize
:
4301 vtn_assert(gl_shader_stage_is_compute(b
->shader
->info
.stage
));
4302 b
->shader
->info
.cs
.local_size
[0] = mode
->operands
[0];
4303 b
->shader
->info
.cs
.local_size
[1] = mode
->operands
[1];
4304 b
->shader
->info
.cs
.local_size
[2] = mode
->operands
[2];
4307 case SpvExecutionModeLocalSizeId
:
4308 b
->shader
->info
.cs
.local_size
[0] = vtn_constant_uint(b
, mode
->operands
[0]);
4309 b
->shader
->info
.cs
.local_size
[1] = vtn_constant_uint(b
, mode
->operands
[1]);
4310 b
->shader
->info
.cs
.local_size
[2] = vtn_constant_uint(b
, mode
->operands
[2]);
4313 case SpvExecutionModeLocalSizeHint
:
4314 case SpvExecutionModeLocalSizeHintId
:
4315 break; /* Nothing to do with this */
4317 case SpvExecutionModeOutputVertices
:
4318 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4319 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
4320 b
->shader
->info
.tess
.tcs_vertices_out
= mode
->operands
[0];
4322 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4323 b
->shader
->info
.gs
.vertices_out
= mode
->operands
[0];
4327 case SpvExecutionModeInputPoints
:
4328 case SpvExecutionModeInputLines
:
4329 case SpvExecutionModeInputLinesAdjacency
:
4330 case SpvExecutionModeTriangles
:
4331 case SpvExecutionModeInputTrianglesAdjacency
:
4332 case SpvExecutionModeQuads
:
4333 case SpvExecutionModeIsolines
:
4334 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4335 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
4336 b
->shader
->info
.tess
.primitive_mode
=
4337 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
4339 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4340 b
->shader
->info
.gs
.vertices_in
=
4341 vertices_in_from_spv_execution_mode(b
, mode
->exec_mode
);
4342 b
->shader
->info
.gs
.input_primitive
=
4343 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
4347 case SpvExecutionModeOutputPoints
:
4348 case SpvExecutionModeOutputLineStrip
:
4349 case SpvExecutionModeOutputTriangleStrip
:
4350 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4351 b
->shader
->info
.gs
.output_primitive
=
4352 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
4355 case SpvExecutionModeSpacingEqual
:
4356 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4357 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4358 b
->shader
->info
.tess
.spacing
= TESS_SPACING_EQUAL
;
4360 case SpvExecutionModeSpacingFractionalEven
:
4361 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4362 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4363 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_EVEN
;
4365 case SpvExecutionModeSpacingFractionalOdd
:
4366 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4367 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4368 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_ODD
;
4370 case SpvExecutionModeVertexOrderCw
:
4371 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4372 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4373 b
->shader
->info
.tess
.ccw
= false;
4375 case SpvExecutionModeVertexOrderCcw
:
4376 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4377 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4378 b
->shader
->info
.tess
.ccw
= true;
4380 case SpvExecutionModePointMode
:
4381 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4382 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4383 b
->shader
->info
.tess
.point_mode
= true;
4386 case SpvExecutionModePixelCenterInteger
:
4387 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4388 b
->shader
->info
.fs
.pixel_center_integer
= true;
4391 case SpvExecutionModeXfb
:
4392 b
->shader
->info
.has_transform_feedback_varyings
= true;
4395 case SpvExecutionModeVecTypeHint
:
4398 case SpvExecutionModeContractionOff
:
4399 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
)
4400 vtn_warn("ExectionMode only allowed for CL-style kernels: %s",
4401 spirv_executionmode_to_string(mode
->exec_mode
));
4406 case SpvExecutionModeStencilRefReplacingEXT
:
4407 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4410 case SpvExecutionModeDerivativeGroupQuadsNV
:
4411 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
4412 b
->shader
->info
.cs
.derivative_group
= DERIVATIVE_GROUP_QUADS
;
4415 case SpvExecutionModeDerivativeGroupLinearNV
:
4416 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
4417 b
->shader
->info
.cs
.derivative_group
= DERIVATIVE_GROUP_LINEAR
;
4420 case SpvExecutionModePixelInterlockOrderedEXT
:
4421 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4422 b
->shader
->info
.fs
.pixel_interlock_ordered
= true;
4425 case SpvExecutionModePixelInterlockUnorderedEXT
:
4426 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4427 b
->shader
->info
.fs
.pixel_interlock_unordered
= true;
4430 case SpvExecutionModeSampleInterlockOrderedEXT
:
4431 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4432 b
->shader
->info
.fs
.sample_interlock_ordered
= true;
4435 case SpvExecutionModeSampleInterlockUnorderedEXT
:
4436 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4437 b
->shader
->info
.fs
.sample_interlock_unordered
= true;
4440 case SpvExecutionModeDenormPreserve
:
4441 case SpvExecutionModeDenormFlushToZero
:
4442 case SpvExecutionModeSignedZeroInfNanPreserve
:
4443 case SpvExecutionModeRoundingModeRTE
:
4444 case SpvExecutionModeRoundingModeRTZ
:
4445 /* Already handled in vtn_handle_rounding_mode_in_execution_mode() */
4449 vtn_fail("Unhandled execution mode: %s (%u)",
4450 spirv_executionmode_to_string(mode
->exec_mode
),
4456 vtn_handle_rounding_mode_in_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
4457 const struct vtn_decoration
*mode
, void *data
)
4459 vtn_assert(b
->entry_point
== entry_point
);
4461 unsigned execution_mode
= 0;
4463 switch(mode
->exec_mode
) {
4464 case SpvExecutionModeDenormPreserve
:
4465 switch (mode
->operands
[0]) {
4466 case 16: execution_mode
= FLOAT_CONTROLS_DENORM_PRESERVE_FP16
; break;
4467 case 32: execution_mode
= FLOAT_CONTROLS_DENORM_PRESERVE_FP32
; break;
4468 case 64: execution_mode
= FLOAT_CONTROLS_DENORM_PRESERVE_FP64
; break;
4469 default: vtn_fail("Floating point type not supported");
4472 case SpvExecutionModeDenormFlushToZero
:
4473 switch (mode
->operands
[0]) {
4474 case 16: execution_mode
= FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16
; break;
4475 case 32: execution_mode
= FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32
; break;
4476 case 64: execution_mode
= FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP64
; break;
4477 default: vtn_fail("Floating point type not supported");
4480 case SpvExecutionModeSignedZeroInfNanPreserve
:
4481 switch (mode
->operands
[0]) {
4482 case 16: execution_mode
= FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP16
; break;
4483 case 32: execution_mode
= FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP32
; break;
4484 case 64: execution_mode
= FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP64
; break;
4485 default: vtn_fail("Floating point type not supported");
4488 case SpvExecutionModeRoundingModeRTE
:
4489 switch (mode
->operands
[0]) {
4490 case 16: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
; break;
4491 case 32: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
; break;
4492 case 64: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
; break;
4493 default: vtn_fail("Floating point type not supported");
4496 case SpvExecutionModeRoundingModeRTZ
:
4497 switch (mode
->operands
[0]) {
4498 case 16: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
; break;
4499 case 32: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
; break;
4500 case 64: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
; break;
4501 default: vtn_fail("Floating point type not supported");
4509 b
->shader
->info
.float_controls_execution_mode
|= execution_mode
;
4513 vtn_handle_variable_or_type_instruction(struct vtn_builder
*b
, SpvOp opcode
,
4514 const uint32_t *w
, unsigned count
)
4516 vtn_set_instruction_result_type(b
, opcode
, w
, count
);
4520 case SpvOpSourceContinued
:
4521 case SpvOpSourceExtension
:
4522 case SpvOpExtension
:
4523 case SpvOpCapability
:
4524 case SpvOpExtInstImport
:
4525 case SpvOpMemoryModel
:
4526 case SpvOpEntryPoint
:
4527 case SpvOpExecutionMode
:
4530 case SpvOpMemberName
:
4531 case SpvOpDecorationGroup
:
4533 case SpvOpDecorateId
:
4534 case SpvOpMemberDecorate
:
4535 case SpvOpGroupDecorate
:
4536 case SpvOpGroupMemberDecorate
:
4537 case SpvOpDecorateString
:
4538 case SpvOpMemberDecorateString
:
4539 vtn_fail("Invalid opcode types and variables section");
4545 case SpvOpTypeFloat
:
4546 case SpvOpTypeVector
:
4547 case SpvOpTypeMatrix
:
4548 case SpvOpTypeImage
:
4549 case SpvOpTypeSampler
:
4550 case SpvOpTypeSampledImage
:
4551 case SpvOpTypeArray
:
4552 case SpvOpTypeRuntimeArray
:
4553 case SpvOpTypeStruct
:
4554 case SpvOpTypeOpaque
:
4555 case SpvOpTypePointer
:
4556 case SpvOpTypeForwardPointer
:
4557 case SpvOpTypeFunction
:
4558 case SpvOpTypeEvent
:
4559 case SpvOpTypeDeviceEvent
:
4560 case SpvOpTypeReserveId
:
4561 case SpvOpTypeQueue
:
4563 vtn_handle_type(b
, opcode
, w
, count
);
4566 case SpvOpConstantTrue
:
4567 case SpvOpConstantFalse
:
4569 case SpvOpConstantComposite
:
4570 case SpvOpConstantSampler
:
4571 case SpvOpConstantNull
:
4572 case SpvOpSpecConstantTrue
:
4573 case SpvOpSpecConstantFalse
:
4574 case SpvOpSpecConstant
:
4575 case SpvOpSpecConstantComposite
:
4576 case SpvOpSpecConstantOp
:
4577 vtn_handle_constant(b
, opcode
, w
, count
);
4582 vtn_handle_variables(b
, opcode
, w
, count
);
4585 case SpvOpExtInst
: {
4586 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
4587 /* NonSemantic extended instructions are acceptable in preamble, others
4588 * will indicate the end of preamble.
4590 return val
->ext_handler
== vtn_handle_non_semantic_instruction
;
4594 return false; /* End of preamble */
4600 static struct vtn_ssa_value
*
4601 vtn_nir_select(struct vtn_builder
*b
, struct vtn_ssa_value
*src0
,
4602 struct vtn_ssa_value
*src1
, struct vtn_ssa_value
*src2
)
4604 struct vtn_ssa_value
*dest
= rzalloc(b
, struct vtn_ssa_value
);
4605 dest
->type
= src1
->type
;
4607 if (glsl_type_is_vector_or_scalar(src1
->type
)) {
4608 dest
->def
= nir_bcsel(&b
->nb
, src0
->def
, src1
->def
, src2
->def
);
4610 unsigned elems
= glsl_get_length(src1
->type
);
4612 dest
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
4613 for (unsigned i
= 0; i
< elems
; i
++) {
4614 dest
->elems
[i
] = vtn_nir_select(b
, src0
,
4615 src1
->elems
[i
], src2
->elems
[i
]);
4623 vtn_handle_select(struct vtn_builder
*b
, SpvOp opcode
,
4624 const uint32_t *w
, unsigned count
)
4626 /* Handle OpSelect up-front here because it needs to be able to handle
4627 * pointers and not just regular vectors and scalars.
4629 struct vtn_value
*res_val
= vtn_untyped_value(b
, w
[2]);
4630 struct vtn_value
*cond_val
= vtn_untyped_value(b
, w
[3]);
4631 struct vtn_value
*obj1_val
= vtn_untyped_value(b
, w
[4]);
4632 struct vtn_value
*obj2_val
= vtn_untyped_value(b
, w
[5]);
4634 vtn_fail_if(obj1_val
->type
!= res_val
->type
||
4635 obj2_val
->type
!= res_val
->type
,
4636 "Object types must match the result type in OpSelect");
4638 vtn_fail_if((cond_val
->type
->base_type
!= vtn_base_type_scalar
&&
4639 cond_val
->type
->base_type
!= vtn_base_type_vector
) ||
4640 !glsl_type_is_boolean(cond_val
->type
->type
),
4641 "OpSelect must have either a vector of booleans or "
4642 "a boolean as Condition type");
4644 vtn_fail_if(cond_val
->type
->base_type
== vtn_base_type_vector
&&
4645 (res_val
->type
->base_type
!= vtn_base_type_vector
||
4646 res_val
->type
->length
!= cond_val
->type
->length
),
4647 "When Condition type in OpSelect is a vector, the Result "
4648 "type must be a vector of the same length");
4650 switch (res_val
->type
->base_type
) {
4651 case vtn_base_type_scalar
:
4652 case vtn_base_type_vector
:
4653 case vtn_base_type_matrix
:
4654 case vtn_base_type_array
:
4655 case vtn_base_type_struct
:
4658 case vtn_base_type_pointer
:
4659 /* We need to have actual storage for pointer types. */
4660 vtn_fail_if(res_val
->type
->type
== NULL
,
4661 "Invalid pointer result type for OpSelect");
4664 vtn_fail("Result type of OpSelect must be a scalar, composite, or pointer");
4667 struct vtn_type
*res_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
4668 struct vtn_ssa_value
*ssa
= vtn_nir_select(b
,
4669 vtn_ssa_value(b
, w
[3]), vtn_ssa_value(b
, w
[4]), vtn_ssa_value(b
, w
[5]));
4671 vtn_push_ssa(b
, w
[2], res_type
, ssa
);
4675 vtn_handle_ptr(struct vtn_builder
*b
, SpvOp opcode
,
4676 const uint32_t *w
, unsigned count
)
4678 struct vtn_type
*type1
= vtn_untyped_value(b
, w
[3])->type
;
4679 struct vtn_type
*type2
= vtn_untyped_value(b
, w
[4])->type
;
4680 vtn_fail_if(type1
->base_type
!= vtn_base_type_pointer
||
4681 type2
->base_type
!= vtn_base_type_pointer
,
4682 "%s operands must have pointer types",
4683 spirv_op_to_string(opcode
));
4684 vtn_fail_if(type1
->storage_class
!= type2
->storage_class
,
4685 "%s operands must have the same storage class",
4686 spirv_op_to_string(opcode
));
4688 struct vtn_type
*vtn_type
=
4689 vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
4690 const struct glsl_type
*type
= vtn_type
->type
;
4692 nir_address_format addr_format
= vtn_mode_to_address_format(
4693 b
, vtn_storage_class_to_mode(b
, type1
->storage_class
, NULL
, NULL
));
4698 case SpvOpPtrDiff
: {
4699 /* OpPtrDiff returns the difference in number of elements (not byte offset). */
4700 unsigned elem_size
, elem_align
;
4701 glsl_get_natural_size_align_bytes(type1
->deref
->type
,
4702 &elem_size
, &elem_align
);
4704 def
= nir_build_addr_isub(&b
->nb
,
4705 vtn_ssa_value(b
, w
[3])->def
,
4706 vtn_ssa_value(b
, w
[4])->def
,
4708 def
= nir_idiv(&b
->nb
, def
, nir_imm_intN_t(&b
->nb
, elem_size
, def
->bit_size
));
4709 def
= nir_i2i(&b
->nb
, def
, glsl_get_bit_size(type
));
4714 case SpvOpPtrNotEqual
: {
4715 def
= nir_build_addr_ieq(&b
->nb
,
4716 vtn_ssa_value(b
, w
[3])->def
,
4717 vtn_ssa_value(b
, w
[4])->def
,
4719 if (opcode
== SpvOpPtrNotEqual
)
4720 def
= nir_inot(&b
->nb
, def
);
4725 unreachable("Invalid ptr operation");
4728 struct vtn_ssa_value
*ssa_value
= vtn_create_ssa_value(b
, type
);
4729 ssa_value
->def
= def
;
4730 vtn_push_ssa(b
, w
[2], vtn_type
, ssa_value
);
4734 vtn_handle_body_instruction(struct vtn_builder
*b
, SpvOp opcode
,
4735 const uint32_t *w
, unsigned count
)
4741 case SpvOpLoopMerge
:
4742 case SpvOpSelectionMerge
:
4743 /* This is handled by cfg pre-pass and walk_blocks */
4747 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
4748 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
4753 vtn_handle_extension(b
, opcode
, w
, count
);
4759 case SpvOpCopyMemory
:
4760 case SpvOpCopyMemorySized
:
4761 case SpvOpAccessChain
:
4762 case SpvOpPtrAccessChain
:
4763 case SpvOpInBoundsAccessChain
:
4764 case SpvOpInBoundsPtrAccessChain
:
4765 case SpvOpArrayLength
:
4766 case SpvOpConvertPtrToU
:
4767 case SpvOpConvertUToPtr
:
4768 vtn_handle_variables(b
, opcode
, w
, count
);
4771 case SpvOpFunctionCall
:
4772 vtn_handle_function_call(b
, opcode
, w
, count
);
4775 case SpvOpSampledImage
:
4777 case SpvOpImageSampleImplicitLod
:
4778 case SpvOpImageSampleExplicitLod
:
4779 case SpvOpImageSampleDrefImplicitLod
:
4780 case SpvOpImageSampleDrefExplicitLod
:
4781 case SpvOpImageSampleProjImplicitLod
:
4782 case SpvOpImageSampleProjExplicitLod
:
4783 case SpvOpImageSampleProjDrefImplicitLod
:
4784 case SpvOpImageSampleProjDrefExplicitLod
:
4785 case SpvOpImageFetch
:
4786 case SpvOpImageGather
:
4787 case SpvOpImageDrefGather
:
4788 case SpvOpImageQuerySizeLod
:
4789 case SpvOpImageQueryLod
:
4790 case SpvOpImageQueryLevels
:
4791 case SpvOpImageQuerySamples
:
4792 vtn_handle_texture(b
, opcode
, w
, count
);
4795 case SpvOpImageRead
:
4796 case SpvOpImageWrite
:
4797 case SpvOpImageTexelPointer
:
4798 vtn_handle_image(b
, opcode
, w
, count
);
4801 case SpvOpImageQuerySize
: {
4802 struct vtn_pointer
*image
=
4803 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
4804 if (glsl_type_is_image(image
->type
->type
)) {
4805 vtn_handle_image(b
, opcode
, w
, count
);
4807 vtn_assert(glsl_type_is_sampler(image
->type
->type
));
4808 vtn_handle_texture(b
, opcode
, w
, count
);
4813 case SpvOpFragmentMaskFetchAMD
:
4814 case SpvOpFragmentFetchAMD
:
4815 vtn_handle_texture(b
, opcode
, w
, count
);
4818 case SpvOpAtomicLoad
:
4819 case SpvOpAtomicExchange
:
4820 case SpvOpAtomicCompareExchange
:
4821 case SpvOpAtomicCompareExchangeWeak
:
4822 case SpvOpAtomicIIncrement
:
4823 case SpvOpAtomicIDecrement
:
4824 case SpvOpAtomicIAdd
:
4825 case SpvOpAtomicISub
:
4826 case SpvOpAtomicSMin
:
4827 case SpvOpAtomicUMin
:
4828 case SpvOpAtomicSMax
:
4829 case SpvOpAtomicUMax
:
4830 case SpvOpAtomicAnd
:
4832 case SpvOpAtomicXor
: {
4833 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[3]);
4834 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
4835 vtn_handle_image(b
, opcode
, w
, count
);
4837 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
4838 vtn_handle_atomics(b
, opcode
, w
, count
);
4843 case SpvOpAtomicStore
: {
4844 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[1]);
4845 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
4846 vtn_handle_image(b
, opcode
, w
, count
);
4848 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
4849 vtn_handle_atomics(b
, opcode
, w
, count
);
4855 vtn_handle_select(b
, opcode
, w
, count
);
4863 case SpvOpConvertFToU
:
4864 case SpvOpConvertFToS
:
4865 case SpvOpConvertSToF
:
4866 case SpvOpConvertUToF
:
4870 case SpvOpQuantizeToF16
:
4871 case SpvOpPtrCastToGeneric
:
4872 case SpvOpGenericCastToPtr
:
4877 case SpvOpSignBitSet
:
4878 case SpvOpLessOrGreater
:
4880 case SpvOpUnordered
:
4895 case SpvOpVectorTimesScalar
:
4897 case SpvOpIAddCarry
:
4898 case SpvOpISubBorrow
:
4899 case SpvOpUMulExtended
:
4900 case SpvOpSMulExtended
:
4901 case SpvOpShiftRightLogical
:
4902 case SpvOpShiftRightArithmetic
:
4903 case SpvOpShiftLeftLogical
:
4904 case SpvOpLogicalEqual
:
4905 case SpvOpLogicalNotEqual
:
4906 case SpvOpLogicalOr
:
4907 case SpvOpLogicalAnd
:
4908 case SpvOpLogicalNot
:
4909 case SpvOpBitwiseOr
:
4910 case SpvOpBitwiseXor
:
4911 case SpvOpBitwiseAnd
:
4913 case SpvOpFOrdEqual
:
4914 case SpvOpFUnordEqual
:
4915 case SpvOpINotEqual
:
4916 case SpvOpFOrdNotEqual
:
4917 case SpvOpFUnordNotEqual
:
4918 case SpvOpULessThan
:
4919 case SpvOpSLessThan
:
4920 case SpvOpFOrdLessThan
:
4921 case SpvOpFUnordLessThan
:
4922 case SpvOpUGreaterThan
:
4923 case SpvOpSGreaterThan
:
4924 case SpvOpFOrdGreaterThan
:
4925 case SpvOpFUnordGreaterThan
:
4926 case SpvOpULessThanEqual
:
4927 case SpvOpSLessThanEqual
:
4928 case SpvOpFOrdLessThanEqual
:
4929 case SpvOpFUnordLessThanEqual
:
4930 case SpvOpUGreaterThanEqual
:
4931 case SpvOpSGreaterThanEqual
:
4932 case SpvOpFOrdGreaterThanEqual
:
4933 case SpvOpFUnordGreaterThanEqual
:
4939 case SpvOpFwidthFine
:
4940 case SpvOpDPdxCoarse
:
4941 case SpvOpDPdyCoarse
:
4942 case SpvOpFwidthCoarse
:
4943 case SpvOpBitFieldInsert
:
4944 case SpvOpBitFieldSExtract
:
4945 case SpvOpBitFieldUExtract
:
4946 case SpvOpBitReverse
:
4948 case SpvOpTranspose
:
4949 case SpvOpOuterProduct
:
4950 case SpvOpMatrixTimesScalar
:
4951 case SpvOpVectorTimesMatrix
:
4952 case SpvOpMatrixTimesVector
:
4953 case SpvOpMatrixTimesMatrix
:
4954 case SpvOpUCountLeadingZerosINTEL
:
4955 case SpvOpUCountTrailingZerosINTEL
:
4956 case SpvOpAbsISubINTEL
:
4957 case SpvOpAbsUSubINTEL
:
4958 case SpvOpIAddSatINTEL
:
4959 case SpvOpUAddSatINTEL
:
4960 case SpvOpIAverageINTEL
:
4961 case SpvOpUAverageINTEL
:
4962 case SpvOpIAverageRoundedINTEL
:
4963 case SpvOpUAverageRoundedINTEL
:
4964 case SpvOpISubSatINTEL
:
4965 case SpvOpUSubSatINTEL
:
4966 case SpvOpIMul32x16INTEL
:
4967 case SpvOpUMul32x16INTEL
:
4968 vtn_handle_alu(b
, opcode
, w
, count
);
4972 vtn_handle_bitcast(b
, w
, count
);
4975 case SpvOpVectorExtractDynamic
:
4976 case SpvOpVectorInsertDynamic
:
4977 case SpvOpVectorShuffle
:
4978 case SpvOpCompositeConstruct
:
4979 case SpvOpCompositeExtract
:
4980 case SpvOpCompositeInsert
:
4981 case SpvOpCopyLogical
:
4982 case SpvOpCopyObject
:
4983 vtn_handle_composite(b
, opcode
, w
, count
);
4986 case SpvOpEmitVertex
:
4987 case SpvOpEndPrimitive
:
4988 case SpvOpEmitStreamVertex
:
4989 case SpvOpEndStreamPrimitive
:
4990 case SpvOpControlBarrier
:
4991 case SpvOpMemoryBarrier
:
4992 vtn_handle_barrier(b
, opcode
, w
, count
);
4995 case SpvOpGroupNonUniformElect
:
4996 case SpvOpGroupNonUniformAll
:
4997 case SpvOpGroupNonUniformAny
:
4998 case SpvOpGroupNonUniformAllEqual
:
4999 case SpvOpGroupNonUniformBroadcast
:
5000 case SpvOpGroupNonUniformBroadcastFirst
:
5001 case SpvOpGroupNonUniformBallot
:
5002 case SpvOpGroupNonUniformInverseBallot
:
5003 case SpvOpGroupNonUniformBallotBitExtract
:
5004 case SpvOpGroupNonUniformBallotBitCount
:
5005 case SpvOpGroupNonUniformBallotFindLSB
:
5006 case SpvOpGroupNonUniformBallotFindMSB
:
5007 case SpvOpGroupNonUniformShuffle
:
5008 case SpvOpGroupNonUniformShuffleXor
:
5009 case SpvOpGroupNonUniformShuffleUp
:
5010 case SpvOpGroupNonUniformShuffleDown
:
5011 case SpvOpGroupNonUniformIAdd
:
5012 case SpvOpGroupNonUniformFAdd
:
5013 case SpvOpGroupNonUniformIMul
:
5014 case SpvOpGroupNonUniformFMul
:
5015 case SpvOpGroupNonUniformSMin
:
5016 case SpvOpGroupNonUniformUMin
:
5017 case SpvOpGroupNonUniformFMin
:
5018 case SpvOpGroupNonUniformSMax
:
5019 case SpvOpGroupNonUniformUMax
:
5020 case SpvOpGroupNonUniformFMax
:
5021 case SpvOpGroupNonUniformBitwiseAnd
:
5022 case SpvOpGroupNonUniformBitwiseOr
:
5023 case SpvOpGroupNonUniformBitwiseXor
:
5024 case SpvOpGroupNonUniformLogicalAnd
:
5025 case SpvOpGroupNonUniformLogicalOr
:
5026 case SpvOpGroupNonUniformLogicalXor
:
5027 case SpvOpGroupNonUniformQuadBroadcast
:
5028 case SpvOpGroupNonUniformQuadSwap
:
5031 case SpvOpGroupBroadcast
:
5032 case SpvOpGroupIAdd
:
5033 case SpvOpGroupFAdd
:
5034 case SpvOpGroupFMin
:
5035 case SpvOpGroupUMin
:
5036 case SpvOpGroupSMin
:
5037 case SpvOpGroupFMax
:
5038 case SpvOpGroupUMax
:
5039 case SpvOpGroupSMax
:
5040 case SpvOpSubgroupBallotKHR
:
5041 case SpvOpSubgroupFirstInvocationKHR
:
5042 case SpvOpSubgroupReadInvocationKHR
:
5043 case SpvOpSubgroupAllKHR
:
5044 case SpvOpSubgroupAnyKHR
:
5045 case SpvOpSubgroupAllEqualKHR
:
5046 case SpvOpGroupIAddNonUniformAMD
:
5047 case SpvOpGroupFAddNonUniformAMD
:
5048 case SpvOpGroupFMinNonUniformAMD
:
5049 case SpvOpGroupUMinNonUniformAMD
:
5050 case SpvOpGroupSMinNonUniformAMD
:
5051 case SpvOpGroupFMaxNonUniformAMD
:
5052 case SpvOpGroupUMaxNonUniformAMD
:
5053 case SpvOpGroupSMaxNonUniformAMD
:
5054 vtn_handle_subgroup(b
, opcode
, w
, count
);
5059 case SpvOpPtrNotEqual
:
5060 vtn_handle_ptr(b
, opcode
, w
, count
);
5063 case SpvOpBeginInvocationInterlockEXT
:
5064 vtn_emit_barrier(b
, nir_intrinsic_begin_invocation_interlock
);
5067 case SpvOpEndInvocationInterlockEXT
:
5068 vtn_emit_barrier(b
, nir_intrinsic_end_invocation_interlock
);
5071 case SpvOpDemoteToHelperInvocationEXT
: {
5072 nir_intrinsic_instr
*intrin
=
5073 nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_demote
);
5074 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
5078 case SpvOpIsHelperInvocationEXT
: {
5079 nir_intrinsic_instr
*intrin
=
5080 nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_is_helper_invocation
);
5081 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
, 1, 1, NULL
);
5082 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
5084 struct vtn_type
*res_type
=
5085 vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
5086 struct vtn_ssa_value
*val
= vtn_create_ssa_value(b
, res_type
->type
);
5087 val
->def
= &intrin
->dest
.ssa
;
5089 vtn_push_ssa(b
, w
[2], res_type
, val
);
5093 case SpvOpReadClockKHR
: {
5094 assert(vtn_constant_uint(b
, w
[3]) == SpvScopeSubgroup
);
5096 /* Operation supports two result types: uvec2 and uint64_t. The NIR
5097 * intrinsic gives uvec2, so pack the result for the other case.
5099 nir_intrinsic_instr
*intrin
=
5100 nir_intrinsic_instr_create(b
->nb
.shader
, nir_intrinsic_shader_clock
);
5101 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
, 2, 32, NULL
);
5102 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
5104 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
5105 const struct glsl_type
*dest_type
= type
->type
;
5106 nir_ssa_def
*result
;
5108 if (glsl_type_is_vector(dest_type
)) {
5109 assert(dest_type
== glsl_vector_type(GLSL_TYPE_UINT
, 2));
5110 result
= &intrin
->dest
.ssa
;
5112 assert(glsl_type_is_scalar(dest_type
));
5113 assert(glsl_get_base_type(dest_type
) == GLSL_TYPE_UINT64
);
5114 result
= nir_pack_64_2x32(&b
->nb
, &intrin
->dest
.ssa
);
5117 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
5119 val
->ssa
= vtn_create_ssa_value(b
, dest_type
);
5120 val
->ssa
->def
= result
;
5124 case SpvOpLifetimeStart
:
5125 case SpvOpLifetimeStop
:
5129 vtn_fail_with_opcode("Unhandled opcode", opcode
);
5136 vtn_create_builder(const uint32_t *words
, size_t word_count
,
5137 gl_shader_stage stage
, const char *entry_point_name
,
5138 const struct spirv_to_nir_options
*options
)
5140 /* Initialize the vtn_builder object */
5141 struct vtn_builder
*b
= rzalloc(NULL
, struct vtn_builder
);
5142 struct spirv_to_nir_options
*dup_options
=
5143 ralloc(b
, struct spirv_to_nir_options
);
5144 *dup_options
= *options
;
5147 b
->spirv_word_count
= word_count
;
5151 list_inithead(&b
->functions
);
5152 b
->entry_point_stage
= stage
;
5153 b
->entry_point_name
= entry_point_name
;
5154 b
->options
= dup_options
;
5157 * Handle the SPIR-V header (first 5 dwords).
5158 * Can't use vtx_assert() as the setjmp(3) target isn't initialized yet.
5160 if (word_count
<= 5)
5163 if (words
[0] != SpvMagicNumber
) {
5164 vtn_err("words[0] was 0x%x, want 0x%x", words
[0], SpvMagicNumber
);
5167 if (words
[1] < 0x10000) {
5168 vtn_err("words[1] was 0x%x, want >= 0x10000", words
[1]);
5172 uint16_t generator_id
= words
[2] >> 16;
5173 uint16_t generator_version
= words
[2];
5175 /* The first GLSLang version bump actually 1.5 years after #179 was fixed
5176 * but this should at least let us shut the workaround off for modern
5177 * versions of GLSLang.
5179 b
->wa_glslang_179
= (generator_id
== 8 && generator_version
== 1);
5181 /* In GLSLang commit 8297936dd6eb3, their handling of barrier() was fixed
5182 * to provide correct memory semantics on compute shader barrier()
5183 * commands. Prior to that, we need to fix them up ourselves. This
5184 * GLSLang fix caused them to bump to generator version 3.
5186 b
->wa_glslang_cs_barrier
= (generator_id
== 8 && generator_version
< 3);
5188 /* words[2] == generator magic */
5189 unsigned value_id_bound
= words
[3];
5190 if (words
[4] != 0) {
5191 vtn_err("words[4] was %u, want 0", words
[4]);
5195 b
->value_id_bound
= value_id_bound
;
5196 b
->values
= rzalloc_array(b
, struct vtn_value
, value_id_bound
);
5204 static nir_function
*
5205 vtn_emit_kernel_entry_point_wrapper(struct vtn_builder
*b
,
5206 nir_function
*entry_point
)
5208 vtn_assert(entry_point
== b
->entry_point
->func
->impl
->function
);
5209 vtn_fail_if(!entry_point
->name
, "entry points are required to have a name");
5210 const char *func_name
=
5211 ralloc_asprintf(b
->shader
, "__wrapped_%s", entry_point
->name
);
5213 /* we shouldn't have any inputs yet */
5214 vtn_assert(!entry_point
->shader
->num_inputs
);
5215 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_KERNEL
);
5217 nir_function
*main_entry_point
= nir_function_create(b
->shader
, func_name
);
5218 main_entry_point
->impl
= nir_function_impl_create(main_entry_point
);
5219 nir_builder_init(&b
->nb
, main_entry_point
->impl
);
5220 b
->nb
.cursor
= nir_after_cf_list(&main_entry_point
->impl
->body
);
5221 b
->func_param_idx
= 0;
5223 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, entry_point
);
5225 for (unsigned i
= 0; i
< entry_point
->num_params
; ++i
) {
5226 struct vtn_type
*param_type
= b
->entry_point
->func
->type
->params
[i
];
5228 /* consider all pointers to function memory to be parameters passed
5231 bool is_by_val
= param_type
->base_type
== vtn_base_type_pointer
&&
5232 param_type
->storage_class
== SpvStorageClassFunction
;
5234 /* input variable */
5235 nir_variable
*in_var
= rzalloc(b
->nb
.shader
, nir_variable
);
5236 in_var
->data
.mode
= nir_var_shader_in
;
5237 in_var
->data
.read_only
= true;
5238 in_var
->data
.location
= i
;
5241 in_var
->type
= param_type
->deref
->type
;
5243 in_var
->type
= param_type
->type
;
5245 nir_shader_add_variable(b
->nb
.shader
, in_var
);
5246 b
->nb
.shader
->num_inputs
++;
5248 /* we have to copy the entire variable into function memory */
5250 nir_variable
*copy_var
=
5251 nir_local_variable_create(main_entry_point
->impl
, in_var
->type
,
5253 nir_copy_var(&b
->nb
, copy_var
, in_var
);
5255 nir_src_for_ssa(&nir_build_deref_var(&b
->nb
, copy_var
)->dest
.ssa
);
5257 call
->params
[i
] = nir_src_for_ssa(nir_load_var(&b
->nb
, in_var
));
5261 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
5263 return main_entry_point
;
5267 spirv_to_nir(const uint32_t *words
, size_t word_count
,
5268 struct nir_spirv_specialization
*spec
, unsigned num_spec
,
5269 gl_shader_stage stage
, const char *entry_point_name
,
5270 const struct spirv_to_nir_options
*options
,
5271 const nir_shader_compiler_options
*nir_options
)
5274 const uint32_t *word_end
= words
+ word_count
;
5276 struct vtn_builder
*b
= vtn_create_builder(words
, word_count
,
5277 stage
, entry_point_name
,
5283 /* See also _vtn_fail() */
5284 if (setjmp(b
->fail_jump
)) {
5289 /* Skip the SPIR-V header, handled at vtn_create_builder */
5292 b
->shader
= nir_shader_create(b
, stage
, nir_options
, NULL
);
5294 /* Handle all the preamble instructions */
5295 words
= vtn_foreach_instruction(b
, words
, word_end
,
5296 vtn_handle_preamble_instruction
);
5298 if (b
->entry_point
== NULL
) {
5299 vtn_fail("Entry point not found");
5304 /* Set shader info defaults */
5305 if (stage
== MESA_SHADER_GEOMETRY
)
5306 b
->shader
->info
.gs
.invocations
= 1;
5308 /* Parse rounding mode execution modes. This has to happen earlier than
5309 * other changes in the execution modes since they can affect, for example,
5310 * the result of the floating point constants.
5312 vtn_foreach_execution_mode(b
, b
->entry_point
,
5313 vtn_handle_rounding_mode_in_execution_mode
, NULL
);
5315 b
->specializations
= spec
;
5316 b
->num_specializations
= num_spec
;
5318 /* Handle all variable, type, and constant instructions */
5319 words
= vtn_foreach_instruction(b
, words
, word_end
,
5320 vtn_handle_variable_or_type_instruction
);
5322 /* Parse execution modes */
5323 vtn_foreach_execution_mode(b
, b
->entry_point
,
5324 vtn_handle_execution_mode
, NULL
);
5326 if (b
->workgroup_size_builtin
) {
5327 vtn_assert(b
->workgroup_size_builtin
->type
->type
==
5328 glsl_vector_type(GLSL_TYPE_UINT
, 3));
5330 nir_const_value
*const_size
=
5331 b
->workgroup_size_builtin
->constant
->values
;
5333 b
->shader
->info
.cs
.local_size
[0] = const_size
[0].u32
;
5334 b
->shader
->info
.cs
.local_size
[1] = const_size
[1].u32
;
5335 b
->shader
->info
.cs
.local_size
[2] = const_size
[2].u32
;
5338 /* Set types on all vtn_values */
5339 vtn_foreach_instruction(b
, words
, word_end
, vtn_set_instruction_result_type
);
5341 vtn_build_cfg(b
, words
, word_end
);
5343 assert(b
->entry_point
->value_type
== vtn_value_type_function
);
5344 b
->entry_point
->func
->referenced
= true;
5349 vtn_foreach_cf_node(node
, &b
->functions
) {
5350 struct vtn_function
*func
= vtn_cf_node_as_function(node
);
5351 if (func
->referenced
&& !func
->emitted
) {
5352 b
->const_table
= _mesa_pointer_hash_table_create(b
);
5354 vtn_function_emit(b
, func
, vtn_handle_body_instruction
);
5360 vtn_assert(b
->entry_point
->value_type
== vtn_value_type_function
);
5361 nir_function
*entry_point
= b
->entry_point
->func
->impl
->function
;
5362 vtn_assert(entry_point
);
5364 /* post process entry_points with input params */
5365 if (entry_point
->num_params
&& b
->shader
->info
.stage
== MESA_SHADER_KERNEL
)
5366 entry_point
= vtn_emit_kernel_entry_point_wrapper(b
, entry_point
);
5368 entry_point
->is_entrypoint
= true;
5370 /* When multiple shader stages exist in the same SPIR-V module, we
5371 * generate input and output variables for every stage, in the same
5372 * NIR program. These dead variables can be invalid NIR. For example,
5373 * TCS outputs must be per-vertex arrays (or decorated 'patch'), while
5374 * VS output variables wouldn't be.
5376 * To ensure we have valid NIR, we eliminate any dead inputs and outputs
5377 * right away. In order to do so, we must lower any constant initializers
5378 * on outputs so nir_remove_dead_variables sees that they're written to.
5380 nir_lower_variable_initializers(b
->shader
, nir_var_shader_out
);
5381 nir_remove_dead_variables(b
->shader
,
5382 nir_var_shader_in
| nir_var_shader_out
);
5384 /* We sometimes generate bogus derefs that, while never used, give the
5385 * validator a bit of heartburn. Run dead code to get rid of them.
5387 nir_opt_dce(b
->shader
);
5389 /* Unparent the shader from the vtn_builder before we delete the builder */
5390 ralloc_steal(NULL
, b
->shader
);
5392 nir_shader
*shader
= b
->shader
;