2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Jason Ekstrand (jason@jlekstrand.net)
28 #include "vtn_private.h"
29 #include "nir/nir_vla.h"
30 #include "nir/nir_control_flow.h"
31 #include "nir/nir_constant_expressions.h"
32 #include "nir/nir_deref.h"
33 #include "spirv_info.h"
35 #include "util/u_math.h"
40 vtn_log(struct vtn_builder
*b
, enum nir_spirv_debug_level level
,
41 size_t spirv_offset
, const char *message
)
43 if (b
->options
->debug
.func
) {
44 b
->options
->debug
.func(b
->options
->debug
.private_data
,
45 level
, spirv_offset
, message
);
49 if (level
>= NIR_SPIRV_DEBUG_LEVEL_WARNING
)
50 fprintf(stderr
, "%s\n", message
);
55 vtn_logf(struct vtn_builder
*b
, enum nir_spirv_debug_level level
,
56 size_t spirv_offset
, const char *fmt
, ...)
62 msg
= ralloc_vasprintf(NULL
, fmt
, args
);
65 vtn_log(b
, level
, spirv_offset
, msg
);
71 vtn_log_err(struct vtn_builder
*b
,
72 enum nir_spirv_debug_level level
, const char *prefix
,
73 const char *file
, unsigned line
,
74 const char *fmt
, va_list args
)
78 msg
= ralloc_strdup(NULL
, prefix
);
81 ralloc_asprintf_append(&msg
, " In file %s:%u\n", file
, line
);
84 ralloc_asprintf_append(&msg
, " ");
86 ralloc_vasprintf_append(&msg
, fmt
, args
);
88 ralloc_asprintf_append(&msg
, "\n %zu bytes into the SPIR-V binary",
92 ralloc_asprintf_append(&msg
,
93 "\n in SPIR-V source file %s, line %d, col %d",
94 b
->file
, b
->line
, b
->col
);
97 vtn_log(b
, level
, b
->spirv_offset
, msg
);
103 vtn_dump_shader(struct vtn_builder
*b
, const char *path
, const char *prefix
)
108 int len
= snprintf(filename
, sizeof(filename
), "%s/%s-%d.spirv",
109 path
, prefix
, idx
++);
110 if (len
< 0 || len
>= sizeof(filename
))
113 FILE *f
= fopen(filename
, "w");
117 fwrite(b
->spirv
, sizeof(*b
->spirv
), b
->spirv_word_count
, f
);
120 vtn_info("SPIR-V shader dumped to %s", filename
);
124 _vtn_warn(struct vtn_builder
*b
, const char *file
, unsigned line
,
125 const char *fmt
, ...)
130 vtn_log_err(b
, NIR_SPIRV_DEBUG_LEVEL_WARNING
, "SPIR-V WARNING:\n",
131 file
, line
, fmt
, args
);
136 _vtn_err(struct vtn_builder
*b
, const char *file
, unsigned line
,
137 const char *fmt
, ...)
142 vtn_log_err(b
, NIR_SPIRV_DEBUG_LEVEL_ERROR
, "SPIR-V ERROR:\n",
143 file
, line
, fmt
, args
);
148 _vtn_fail(struct vtn_builder
*b
, const char *file
, unsigned line
,
149 const char *fmt
, ...)
154 vtn_log_err(b
, NIR_SPIRV_DEBUG_LEVEL_ERROR
, "SPIR-V parsing FAILED:\n",
155 file
, line
, fmt
, args
);
158 const char *dump_path
= getenv("MESA_SPIRV_FAIL_DUMP_PATH");
160 vtn_dump_shader(b
, dump_path
, "fail");
162 longjmp(b
->fail_jump
, 1);
165 struct spec_constant_value
{
173 static struct vtn_ssa_value
*
174 vtn_undef_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
176 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
179 if (glsl_type_is_vector_or_scalar(type
)) {
180 unsigned num_components
= glsl_get_vector_elements(val
->type
);
181 unsigned bit_size
= glsl_get_bit_size(val
->type
);
182 val
->def
= nir_ssa_undef(&b
->nb
, num_components
, bit_size
);
184 unsigned elems
= glsl_get_length(val
->type
);
185 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
186 if (glsl_type_is_matrix(type
)) {
187 const struct glsl_type
*elem_type
=
188 glsl_vector_type(glsl_get_base_type(type
),
189 glsl_get_vector_elements(type
));
191 for (unsigned i
= 0; i
< elems
; i
++)
192 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
193 } else if (glsl_type_is_array(type
)) {
194 const struct glsl_type
*elem_type
= glsl_get_array_element(type
);
195 for (unsigned i
= 0; i
< elems
; i
++)
196 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
198 for (unsigned i
= 0; i
< elems
; i
++) {
199 const struct glsl_type
*elem_type
= glsl_get_struct_field(type
, i
);
200 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
208 static struct vtn_ssa_value
*
209 vtn_const_ssa_value(struct vtn_builder
*b
, nir_constant
*constant
,
210 const struct glsl_type
*type
)
212 struct hash_entry
*entry
= _mesa_hash_table_search(b
->const_table
, constant
);
217 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
220 switch (glsl_get_base_type(type
)) {
223 case GLSL_TYPE_INT16
:
224 case GLSL_TYPE_UINT16
:
225 case GLSL_TYPE_UINT8
:
227 case GLSL_TYPE_INT64
:
228 case GLSL_TYPE_UINT64
:
230 case GLSL_TYPE_FLOAT
:
231 case GLSL_TYPE_FLOAT16
:
232 case GLSL_TYPE_DOUBLE
: {
233 int bit_size
= glsl_get_bit_size(type
);
234 if (glsl_type_is_vector_or_scalar(type
)) {
235 unsigned num_components
= glsl_get_vector_elements(val
->type
);
236 nir_load_const_instr
*load
=
237 nir_load_const_instr_create(b
->shader
, num_components
, bit_size
);
239 memcpy(load
->value
, constant
->values
,
240 sizeof(nir_const_value
) * load
->def
.num_components
);
242 nir_instr_insert_before_cf_list(&b
->nb
.impl
->body
, &load
->instr
);
243 val
->def
= &load
->def
;
245 assert(glsl_type_is_matrix(type
));
246 unsigned columns
= glsl_get_matrix_columns(val
->type
);
247 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, columns
);
248 const struct glsl_type
*column_type
= glsl_get_column_type(val
->type
);
249 for (unsigned i
= 0; i
< columns
; i
++)
250 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
256 case GLSL_TYPE_ARRAY
: {
257 unsigned elems
= glsl_get_length(val
->type
);
258 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
259 const struct glsl_type
*elem_type
= glsl_get_array_element(val
->type
);
260 for (unsigned i
= 0; i
< elems
; i
++)
261 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
266 case GLSL_TYPE_STRUCT
: {
267 unsigned elems
= glsl_get_length(val
->type
);
268 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
269 for (unsigned i
= 0; i
< elems
; i
++) {
270 const struct glsl_type
*elem_type
=
271 glsl_get_struct_field(val
->type
, i
);
272 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
279 vtn_fail("bad constant type");
285 struct vtn_ssa_value
*
286 vtn_ssa_value(struct vtn_builder
*b
, uint32_t value_id
)
288 struct vtn_value
*val
= vtn_untyped_value(b
, value_id
);
289 switch (val
->value_type
) {
290 case vtn_value_type_undef
:
291 return vtn_undef_ssa_value(b
, val
->type
->type
);
293 case vtn_value_type_constant
:
294 return vtn_const_ssa_value(b
, val
->constant
, val
->type
->type
);
296 case vtn_value_type_ssa
:
299 case vtn_value_type_pointer
:
300 vtn_assert(val
->pointer
->ptr_type
&& val
->pointer
->ptr_type
->type
);
301 struct vtn_ssa_value
*ssa
=
302 vtn_create_ssa_value(b
, val
->pointer
->ptr_type
->type
);
303 ssa
->def
= vtn_pointer_to_ssa(b
, val
->pointer
);
307 vtn_fail("Invalid type for an SSA value");
312 vtn_string_literal(struct vtn_builder
*b
, const uint32_t *words
,
313 unsigned word_count
, unsigned *words_used
)
315 char *dup
= ralloc_strndup(b
, (char *)words
, word_count
* sizeof(*words
));
317 /* Ammount of space taken by the string (including the null) */
318 unsigned len
= strlen(dup
) + 1;
319 *words_used
= DIV_ROUND_UP(len
, sizeof(*words
));
325 vtn_foreach_instruction(struct vtn_builder
*b
, const uint32_t *start
,
326 const uint32_t *end
, vtn_instruction_handler handler
)
332 const uint32_t *w
= start
;
334 SpvOp opcode
= w
[0] & SpvOpCodeMask
;
335 unsigned count
= w
[0] >> SpvWordCountShift
;
336 vtn_assert(count
>= 1 && w
+ count
<= end
);
338 b
->spirv_offset
= (uint8_t *)w
- (uint8_t *)b
->spirv
;
342 break; /* Do nothing */
345 b
->file
= vtn_value(b
, w
[1], vtn_value_type_string
)->str
;
357 if (!handler(b
, opcode
, w
, count
))
375 vtn_handle_extension(struct vtn_builder
*b
, SpvOp opcode
,
376 const uint32_t *w
, unsigned count
)
378 const char *ext
= (const char *)&w
[2];
380 case SpvOpExtInstImport
: {
381 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_extension
);
382 if (strcmp(ext
, "GLSL.std.450") == 0) {
383 val
->ext_handler
= vtn_handle_glsl450_instruction
;
384 } else if ((strcmp(ext
, "SPV_AMD_gcn_shader") == 0)
385 && (b
->options
&& b
->options
->caps
.amd_gcn_shader
)) {
386 val
->ext_handler
= vtn_handle_amd_gcn_shader_instruction
;
387 } else if ((strcmp(ext
, "SPV_AMD_shader_ballot") == 0)
388 && (b
->options
&& b
->options
->caps
.amd_shader_ballot
)) {
389 val
->ext_handler
= vtn_handle_amd_shader_ballot_instruction
;
390 } else if ((strcmp(ext
, "SPV_AMD_shader_trinary_minmax") == 0)
391 && (b
->options
&& b
->options
->caps
.amd_trinary_minmax
)) {
392 val
->ext_handler
= vtn_handle_amd_shader_trinary_minmax_instruction
;
393 } else if (strcmp(ext
, "OpenCL.std") == 0) {
394 val
->ext_handler
= vtn_handle_opencl_instruction
;
396 vtn_fail("Unsupported extension: %s", ext
);
402 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
403 bool handled
= val
->ext_handler(b
, w
[4], w
, count
);
409 vtn_fail_with_opcode("Unhandled opcode", opcode
);
414 _foreach_decoration_helper(struct vtn_builder
*b
,
415 struct vtn_value
*base_value
,
417 struct vtn_value
*value
,
418 vtn_decoration_foreach_cb cb
, void *data
)
420 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
422 if (dec
->scope
== VTN_DEC_DECORATION
) {
423 member
= parent_member
;
424 } else if (dec
->scope
>= VTN_DEC_STRUCT_MEMBER0
) {
425 vtn_fail_if(value
->value_type
!= vtn_value_type_type
||
426 value
->type
->base_type
!= vtn_base_type_struct
,
427 "OpMemberDecorate and OpGroupMemberDecorate are only "
428 "allowed on OpTypeStruct");
429 /* This means we haven't recursed yet */
430 assert(value
== base_value
);
432 member
= dec
->scope
- VTN_DEC_STRUCT_MEMBER0
;
434 vtn_fail_if(member
>= base_value
->type
->length
,
435 "OpMemberDecorate specifies member %d but the "
436 "OpTypeStruct has only %u members",
437 member
, base_value
->type
->length
);
439 /* Not a decoration */
440 assert(dec
->scope
== VTN_DEC_EXECUTION_MODE
);
445 assert(dec
->group
->value_type
== vtn_value_type_decoration_group
);
446 _foreach_decoration_helper(b
, base_value
, member
, dec
->group
,
449 cb(b
, base_value
, member
, dec
, data
);
454 /** Iterates (recursively if needed) over all of the decorations on a value
456 * This function iterates over all of the decorations applied to a given
457 * value. If it encounters a decoration group, it recurses into the group
458 * and iterates over all of those decorations as well.
461 vtn_foreach_decoration(struct vtn_builder
*b
, struct vtn_value
*value
,
462 vtn_decoration_foreach_cb cb
, void *data
)
464 _foreach_decoration_helper(b
, value
, -1, value
, cb
, data
);
468 vtn_foreach_execution_mode(struct vtn_builder
*b
, struct vtn_value
*value
,
469 vtn_execution_mode_foreach_cb cb
, void *data
)
471 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
472 if (dec
->scope
!= VTN_DEC_EXECUTION_MODE
)
475 assert(dec
->group
== NULL
);
476 cb(b
, value
, dec
, data
);
481 vtn_handle_decoration(struct vtn_builder
*b
, SpvOp opcode
,
482 const uint32_t *w
, unsigned count
)
484 const uint32_t *w_end
= w
+ count
;
485 const uint32_t target
= w
[1];
489 case SpvOpDecorationGroup
:
490 vtn_push_value(b
, target
, vtn_value_type_decoration_group
);
494 case SpvOpDecorateId
:
495 case SpvOpMemberDecorate
:
496 case SpvOpDecorateString
:
497 case SpvOpMemberDecorateString
:
498 case SpvOpExecutionMode
:
499 case SpvOpExecutionModeId
: {
500 struct vtn_value
*val
= vtn_untyped_value(b
, target
);
502 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
505 case SpvOpDecorateId
:
506 case SpvOpDecorateString
:
507 dec
->scope
= VTN_DEC_DECORATION
;
509 case SpvOpMemberDecorate
:
510 case SpvOpMemberDecorateString
:
511 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(w
++);
512 vtn_fail_if(dec
->scope
< VTN_DEC_STRUCT_MEMBER0
, /* overflow */
513 "Member argument of OpMemberDecorate too large");
515 case SpvOpExecutionMode
:
516 case SpvOpExecutionModeId
:
517 dec
->scope
= VTN_DEC_EXECUTION_MODE
;
520 unreachable("Invalid decoration opcode");
522 dec
->decoration
= *(w
++);
525 /* Link into the list */
526 dec
->next
= val
->decoration
;
527 val
->decoration
= dec
;
531 case SpvOpGroupMemberDecorate
:
532 case SpvOpGroupDecorate
: {
533 struct vtn_value
*group
=
534 vtn_value(b
, target
, vtn_value_type_decoration_group
);
536 for (; w
< w_end
; w
++) {
537 struct vtn_value
*val
= vtn_untyped_value(b
, *w
);
538 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
541 if (opcode
== SpvOpGroupDecorate
) {
542 dec
->scope
= VTN_DEC_DECORATION
;
544 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(++w
);
545 vtn_fail_if(dec
->scope
< 0, /* Check for overflow */
546 "Member argument of OpGroupMemberDecorate too large");
549 /* Link into the list */
550 dec
->next
= val
->decoration
;
551 val
->decoration
= dec
;
557 unreachable("Unhandled opcode");
561 struct member_decoration_ctx
{
563 struct glsl_struct_field
*fields
;
564 struct vtn_type
*type
;
568 * Returns true if the given type contains a struct decorated Block or
572 vtn_type_contains_block(struct vtn_builder
*b
, struct vtn_type
*type
)
574 switch (type
->base_type
) {
575 case vtn_base_type_array
:
576 return vtn_type_contains_block(b
, type
->array_element
);
577 case vtn_base_type_struct
:
578 if (type
->block
|| type
->buffer_block
)
580 for (unsigned i
= 0; i
< type
->length
; i
++) {
581 if (vtn_type_contains_block(b
, type
->members
[i
]))
590 /** Returns true if two types are "compatible", i.e. you can do an OpLoad,
591 * OpStore, or OpCopyMemory between them without breaking anything.
592 * Technically, the SPIR-V rules require the exact same type ID but this lets
593 * us internally be a bit looser.
596 vtn_types_compatible(struct vtn_builder
*b
,
597 struct vtn_type
*t1
, struct vtn_type
*t2
)
599 if (t1
->id
== t2
->id
)
602 if (t1
->base_type
!= t2
->base_type
)
605 switch (t1
->base_type
) {
606 case vtn_base_type_void
:
607 case vtn_base_type_scalar
:
608 case vtn_base_type_vector
:
609 case vtn_base_type_matrix
:
610 case vtn_base_type_image
:
611 case vtn_base_type_sampler
:
612 case vtn_base_type_sampled_image
:
613 return t1
->type
== t2
->type
;
615 case vtn_base_type_array
:
616 return t1
->length
== t2
->length
&&
617 vtn_types_compatible(b
, t1
->array_element
, t2
->array_element
);
619 case vtn_base_type_pointer
:
620 return vtn_types_compatible(b
, t1
->deref
, t2
->deref
);
622 case vtn_base_type_struct
:
623 if (t1
->length
!= t2
->length
)
626 for (unsigned i
= 0; i
< t1
->length
; i
++) {
627 if (!vtn_types_compatible(b
, t1
->members
[i
], t2
->members
[i
]))
632 case vtn_base_type_function
:
633 /* This case shouldn't get hit since you can't copy around function
634 * types. Just require them to be identical.
639 vtn_fail("Invalid base type");
643 vtn_type_without_array(struct vtn_type
*type
)
645 while (type
->base_type
== vtn_base_type_array
)
646 type
= type
->array_element
;
650 /* does a shallow copy of a vtn_type */
652 static struct vtn_type
*
653 vtn_type_copy(struct vtn_builder
*b
, struct vtn_type
*src
)
655 struct vtn_type
*dest
= ralloc(b
, struct vtn_type
);
658 switch (src
->base_type
) {
659 case vtn_base_type_void
:
660 case vtn_base_type_scalar
:
661 case vtn_base_type_vector
:
662 case vtn_base_type_matrix
:
663 case vtn_base_type_array
:
664 case vtn_base_type_pointer
:
665 case vtn_base_type_image
:
666 case vtn_base_type_sampler
:
667 case vtn_base_type_sampled_image
:
668 /* Nothing more to do */
671 case vtn_base_type_struct
:
672 dest
->members
= ralloc_array(b
, struct vtn_type
*, src
->length
);
673 memcpy(dest
->members
, src
->members
,
674 src
->length
* sizeof(src
->members
[0]));
676 dest
->offsets
= ralloc_array(b
, unsigned, src
->length
);
677 memcpy(dest
->offsets
, src
->offsets
,
678 src
->length
* sizeof(src
->offsets
[0]));
681 case vtn_base_type_function
:
682 dest
->params
= ralloc_array(b
, struct vtn_type
*, src
->length
);
683 memcpy(dest
->params
, src
->params
, src
->length
* sizeof(src
->params
[0]));
690 static struct vtn_type
*
691 mutable_matrix_member(struct vtn_builder
*b
, struct vtn_type
*type
, int member
)
693 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
694 type
= type
->members
[member
];
696 /* We may have an array of matrices.... Oh, joy! */
697 while (glsl_type_is_array(type
->type
)) {
698 type
->array_element
= vtn_type_copy(b
, type
->array_element
);
699 type
= type
->array_element
;
702 vtn_assert(glsl_type_is_matrix(type
->type
));
708 vtn_handle_access_qualifier(struct vtn_builder
*b
, struct vtn_type
*type
,
709 int member
, enum gl_access_qualifier access
)
711 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
712 type
= type
->members
[member
];
714 type
->access
|= access
;
718 array_stride_decoration_cb(struct vtn_builder
*b
,
719 struct vtn_value
*val
, int member
,
720 const struct vtn_decoration
*dec
, void *void_ctx
)
722 struct vtn_type
*type
= val
->type
;
724 if (dec
->decoration
== SpvDecorationArrayStride
) {
725 if (vtn_type_contains_block(b
, type
)) {
726 vtn_warn("The ArrayStride decoration cannot be applied to an array "
727 "type which contains a structure type decorated Block "
729 /* Ignore the decoration */
731 vtn_fail_if(dec
->operands
[0] == 0, "ArrayStride must be non-zero");
732 type
->stride
= dec
->operands
[0];
738 struct_member_decoration_cb(struct vtn_builder
*b
,
739 struct vtn_value
*val
, int member
,
740 const struct vtn_decoration
*dec
, void *void_ctx
)
742 struct member_decoration_ctx
*ctx
= void_ctx
;
747 assert(member
< ctx
->num_fields
);
749 switch (dec
->decoration
) {
750 case SpvDecorationRelaxedPrecision
:
751 case SpvDecorationUniform
:
752 case SpvDecorationUniformId
:
753 break; /* FIXME: Do nothing with this for now. */
754 case SpvDecorationNonWritable
:
755 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_NON_WRITEABLE
);
757 case SpvDecorationNonReadable
:
758 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_NON_READABLE
);
760 case SpvDecorationVolatile
:
761 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_VOLATILE
);
763 case SpvDecorationCoherent
:
764 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_COHERENT
);
766 case SpvDecorationNoPerspective
:
767 ctx
->fields
[member
].interpolation
= INTERP_MODE_NOPERSPECTIVE
;
769 case SpvDecorationFlat
:
770 ctx
->fields
[member
].interpolation
= INTERP_MODE_FLAT
;
772 case SpvDecorationCentroid
:
773 ctx
->fields
[member
].centroid
= true;
775 case SpvDecorationSample
:
776 ctx
->fields
[member
].sample
= true;
778 case SpvDecorationStream
:
779 /* Vulkan only allows one GS stream */
780 vtn_assert(dec
->operands
[0] == 0);
782 case SpvDecorationLocation
:
783 ctx
->fields
[member
].location
= dec
->operands
[0];
785 case SpvDecorationComponent
:
786 break; /* FIXME: What should we do with these? */
787 case SpvDecorationBuiltIn
:
788 ctx
->type
->members
[member
] = vtn_type_copy(b
, ctx
->type
->members
[member
]);
789 ctx
->type
->members
[member
]->is_builtin
= true;
790 ctx
->type
->members
[member
]->builtin
= dec
->operands
[0];
791 ctx
->type
->builtin_block
= true;
793 case SpvDecorationOffset
:
794 ctx
->type
->offsets
[member
] = dec
->operands
[0];
795 ctx
->fields
[member
].offset
= dec
->operands
[0];
797 case SpvDecorationMatrixStride
:
798 /* Handled as a second pass */
800 case SpvDecorationColMajor
:
801 break; /* Nothing to do here. Column-major is the default. */
802 case SpvDecorationRowMajor
:
803 mutable_matrix_member(b
, ctx
->type
, member
)->row_major
= true;
806 case SpvDecorationPatch
:
809 case SpvDecorationSpecId
:
810 case SpvDecorationBlock
:
811 case SpvDecorationBufferBlock
:
812 case SpvDecorationArrayStride
:
813 case SpvDecorationGLSLShared
:
814 case SpvDecorationGLSLPacked
:
815 case SpvDecorationInvariant
:
816 case SpvDecorationRestrict
:
817 case SpvDecorationAliased
:
818 case SpvDecorationConstant
:
819 case SpvDecorationIndex
:
820 case SpvDecorationBinding
:
821 case SpvDecorationDescriptorSet
:
822 case SpvDecorationLinkageAttributes
:
823 case SpvDecorationNoContraction
:
824 case SpvDecorationInputAttachmentIndex
:
825 vtn_warn("Decoration not allowed on struct members: %s",
826 spirv_decoration_to_string(dec
->decoration
));
829 case SpvDecorationXfbBuffer
:
830 case SpvDecorationXfbStride
:
831 vtn_warn("Vulkan does not have transform feedback");
834 case SpvDecorationCPacked
:
835 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
)
836 vtn_warn("Decoration only allowed for CL-style kernels: %s",
837 spirv_decoration_to_string(dec
->decoration
));
839 ctx
->type
->packed
= true;
842 case SpvDecorationSaturatedConversion
:
843 case SpvDecorationFuncParamAttr
:
844 case SpvDecorationFPRoundingMode
:
845 case SpvDecorationFPFastMathMode
:
846 case SpvDecorationAlignment
:
847 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
) {
848 vtn_warn("Decoration only allowed for CL-style kernels: %s",
849 spirv_decoration_to_string(dec
->decoration
));
853 case SpvDecorationUserSemantic
:
854 /* User semantic decorations can safely be ignored by the driver. */
858 vtn_fail_with_decoration("Unhandled decoration", dec
->decoration
);
862 /** Chases the array type all the way down to the tail and rewrites the
863 * glsl_types to be based off the tail's glsl_type.
866 vtn_array_type_rewrite_glsl_type(struct vtn_type
*type
)
868 if (type
->base_type
!= vtn_base_type_array
)
871 vtn_array_type_rewrite_glsl_type(type
->array_element
);
873 type
->type
= glsl_array_type(type
->array_element
->type
,
874 type
->length
, type
->stride
);
877 /* Matrix strides are handled as a separate pass because we need to know
878 * whether the matrix is row-major or not first.
881 struct_member_matrix_stride_cb(struct vtn_builder
*b
,
882 struct vtn_value
*val
, int member
,
883 const struct vtn_decoration
*dec
,
886 if (dec
->decoration
!= SpvDecorationMatrixStride
)
889 vtn_fail_if(member
< 0,
890 "The MatrixStride decoration is only allowed on members "
892 vtn_fail_if(dec
->operands
[0] == 0, "MatrixStride must be non-zero");
894 struct member_decoration_ctx
*ctx
= void_ctx
;
896 struct vtn_type
*mat_type
= mutable_matrix_member(b
, ctx
->type
, member
);
897 if (mat_type
->row_major
) {
898 mat_type
->array_element
= vtn_type_copy(b
, mat_type
->array_element
);
899 mat_type
->stride
= mat_type
->array_element
->stride
;
900 mat_type
->array_element
->stride
= dec
->operands
[0];
902 mat_type
->type
= glsl_explicit_matrix_type(mat_type
->type
,
903 dec
->operands
[0], true);
904 mat_type
->array_element
->type
= glsl_get_column_type(mat_type
->type
);
906 vtn_assert(mat_type
->array_element
->stride
> 0);
907 mat_type
->stride
= dec
->operands
[0];
909 mat_type
->type
= glsl_explicit_matrix_type(mat_type
->type
,
910 dec
->operands
[0], false);
913 /* Now that we've replaced the glsl_type with a properly strided matrix
914 * type, rewrite the member type so that it's an array of the proper kind
917 vtn_array_type_rewrite_glsl_type(ctx
->type
->members
[member
]);
918 ctx
->fields
[member
].type
= ctx
->type
->members
[member
]->type
;
922 struct_block_decoration_cb(struct vtn_builder
*b
,
923 struct vtn_value
*val
, int member
,
924 const struct vtn_decoration
*dec
, void *ctx
)
929 struct vtn_type
*type
= val
->type
;
930 if (dec
->decoration
== SpvDecorationBlock
)
932 else if (dec
->decoration
== SpvDecorationBufferBlock
)
933 type
->buffer_block
= true;
937 type_decoration_cb(struct vtn_builder
*b
,
938 struct vtn_value
*val
, int member
,
939 const struct vtn_decoration
*dec
, void *ctx
)
941 struct vtn_type
*type
= val
->type
;
944 /* This should have been handled by OpTypeStruct */
945 assert(val
->type
->base_type
== vtn_base_type_struct
);
946 assert(member
>= 0 && member
< val
->type
->length
);
950 switch (dec
->decoration
) {
951 case SpvDecorationArrayStride
:
952 vtn_assert(type
->base_type
== vtn_base_type_array
||
953 type
->base_type
== vtn_base_type_pointer
);
955 case SpvDecorationBlock
:
956 vtn_assert(type
->base_type
== vtn_base_type_struct
);
957 vtn_assert(type
->block
);
959 case SpvDecorationBufferBlock
:
960 vtn_assert(type
->base_type
== vtn_base_type_struct
);
961 vtn_assert(type
->buffer_block
);
963 case SpvDecorationGLSLShared
:
964 case SpvDecorationGLSLPacked
:
965 /* Ignore these, since we get explicit offsets anyways */
968 case SpvDecorationRowMajor
:
969 case SpvDecorationColMajor
:
970 case SpvDecorationMatrixStride
:
971 case SpvDecorationBuiltIn
:
972 case SpvDecorationNoPerspective
:
973 case SpvDecorationFlat
:
974 case SpvDecorationPatch
:
975 case SpvDecorationCentroid
:
976 case SpvDecorationSample
:
977 case SpvDecorationVolatile
:
978 case SpvDecorationCoherent
:
979 case SpvDecorationNonWritable
:
980 case SpvDecorationNonReadable
:
981 case SpvDecorationUniform
:
982 case SpvDecorationUniformId
:
983 case SpvDecorationLocation
:
984 case SpvDecorationComponent
:
985 case SpvDecorationOffset
:
986 case SpvDecorationXfbBuffer
:
987 case SpvDecorationXfbStride
:
988 case SpvDecorationUserSemantic
:
989 vtn_warn("Decoration only allowed for struct members: %s",
990 spirv_decoration_to_string(dec
->decoration
));
993 case SpvDecorationStream
:
994 /* We don't need to do anything here, as stream is filled up when
995 * aplying the decoration to a variable, just check that if it is not a
996 * struct member, it should be a struct.
998 vtn_assert(type
->base_type
== vtn_base_type_struct
);
1001 case SpvDecorationRelaxedPrecision
:
1002 case SpvDecorationSpecId
:
1003 case SpvDecorationInvariant
:
1004 case SpvDecorationRestrict
:
1005 case SpvDecorationAliased
:
1006 case SpvDecorationConstant
:
1007 case SpvDecorationIndex
:
1008 case SpvDecorationBinding
:
1009 case SpvDecorationDescriptorSet
:
1010 case SpvDecorationLinkageAttributes
:
1011 case SpvDecorationNoContraction
:
1012 case SpvDecorationInputAttachmentIndex
:
1013 vtn_warn("Decoration not allowed on types: %s",
1014 spirv_decoration_to_string(dec
->decoration
));
1017 case SpvDecorationCPacked
:
1018 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
)
1019 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1020 spirv_decoration_to_string(dec
->decoration
));
1022 type
->packed
= true;
1025 case SpvDecorationSaturatedConversion
:
1026 case SpvDecorationFuncParamAttr
:
1027 case SpvDecorationFPRoundingMode
:
1028 case SpvDecorationFPFastMathMode
:
1029 case SpvDecorationAlignment
:
1030 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1031 spirv_decoration_to_string(dec
->decoration
));
1035 vtn_fail_with_decoration("Unhandled decoration", dec
->decoration
);
1040 translate_image_format(struct vtn_builder
*b
, SpvImageFormat format
)
1043 case SpvImageFormatUnknown
: return 0; /* GL_NONE */
1044 case SpvImageFormatRgba32f
: return 0x8814; /* GL_RGBA32F */
1045 case SpvImageFormatRgba16f
: return 0x881A; /* GL_RGBA16F */
1046 case SpvImageFormatR32f
: return 0x822E; /* GL_R32F */
1047 case SpvImageFormatRgba8
: return 0x8058; /* GL_RGBA8 */
1048 case SpvImageFormatRgba8Snorm
: return 0x8F97; /* GL_RGBA8_SNORM */
1049 case SpvImageFormatRg32f
: return 0x8230; /* GL_RG32F */
1050 case SpvImageFormatRg16f
: return 0x822F; /* GL_RG16F */
1051 case SpvImageFormatR11fG11fB10f
: return 0x8C3A; /* GL_R11F_G11F_B10F */
1052 case SpvImageFormatR16f
: return 0x822D; /* GL_R16F */
1053 case SpvImageFormatRgba16
: return 0x805B; /* GL_RGBA16 */
1054 case SpvImageFormatRgb10A2
: return 0x8059; /* GL_RGB10_A2 */
1055 case SpvImageFormatRg16
: return 0x822C; /* GL_RG16 */
1056 case SpvImageFormatRg8
: return 0x822B; /* GL_RG8 */
1057 case SpvImageFormatR16
: return 0x822A; /* GL_R16 */
1058 case SpvImageFormatR8
: return 0x8229; /* GL_R8 */
1059 case SpvImageFormatRgba16Snorm
: return 0x8F9B; /* GL_RGBA16_SNORM */
1060 case SpvImageFormatRg16Snorm
: return 0x8F99; /* GL_RG16_SNORM */
1061 case SpvImageFormatRg8Snorm
: return 0x8F95; /* GL_RG8_SNORM */
1062 case SpvImageFormatR16Snorm
: return 0x8F98; /* GL_R16_SNORM */
1063 case SpvImageFormatR8Snorm
: return 0x8F94; /* GL_R8_SNORM */
1064 case SpvImageFormatRgba32i
: return 0x8D82; /* GL_RGBA32I */
1065 case SpvImageFormatRgba16i
: return 0x8D88; /* GL_RGBA16I */
1066 case SpvImageFormatRgba8i
: return 0x8D8E; /* GL_RGBA8I */
1067 case SpvImageFormatR32i
: return 0x8235; /* GL_R32I */
1068 case SpvImageFormatRg32i
: return 0x823B; /* GL_RG32I */
1069 case SpvImageFormatRg16i
: return 0x8239; /* GL_RG16I */
1070 case SpvImageFormatRg8i
: return 0x8237; /* GL_RG8I */
1071 case SpvImageFormatR16i
: return 0x8233; /* GL_R16I */
1072 case SpvImageFormatR8i
: return 0x8231; /* GL_R8I */
1073 case SpvImageFormatRgba32ui
: return 0x8D70; /* GL_RGBA32UI */
1074 case SpvImageFormatRgba16ui
: return 0x8D76; /* GL_RGBA16UI */
1075 case SpvImageFormatRgba8ui
: return 0x8D7C; /* GL_RGBA8UI */
1076 case SpvImageFormatR32ui
: return 0x8236; /* GL_R32UI */
1077 case SpvImageFormatRgb10a2ui
: return 0x906F; /* GL_RGB10_A2UI */
1078 case SpvImageFormatRg32ui
: return 0x823C; /* GL_RG32UI */
1079 case SpvImageFormatRg16ui
: return 0x823A; /* GL_RG16UI */
1080 case SpvImageFormatRg8ui
: return 0x8238; /* GL_RG8UI */
1081 case SpvImageFormatR16ui
: return 0x8234; /* GL_R16UI */
1082 case SpvImageFormatR8ui
: return 0x8232; /* GL_R8UI */
1084 vtn_fail("Invalid image format: %s (%u)",
1085 spirv_imageformat_to_string(format
), format
);
1090 vtn_handle_type(struct vtn_builder
*b
, SpvOp opcode
,
1091 const uint32_t *w
, unsigned count
)
1093 struct vtn_value
*val
= NULL
;
1095 /* In order to properly handle forward declarations, we have to defer
1096 * allocation for pointer types.
1098 if (opcode
!= SpvOpTypePointer
&& opcode
!= SpvOpTypeForwardPointer
) {
1099 val
= vtn_push_value(b
, w
[1], vtn_value_type_type
);
1100 vtn_fail_if(val
->type
!= NULL
,
1101 "Only pointers can have forward declarations");
1102 val
->type
= rzalloc(b
, struct vtn_type
);
1103 val
->type
->id
= w
[1];
1108 val
->type
->base_type
= vtn_base_type_void
;
1109 val
->type
->type
= glsl_void_type();
1112 val
->type
->base_type
= vtn_base_type_scalar
;
1113 val
->type
->type
= glsl_bool_type();
1114 val
->type
->length
= 1;
1116 case SpvOpTypeInt
: {
1117 int bit_size
= w
[2];
1118 const bool signedness
= w
[3];
1119 val
->type
->base_type
= vtn_base_type_scalar
;
1122 val
->type
->type
= (signedness
? glsl_int64_t_type() : glsl_uint64_t_type());
1125 val
->type
->type
= (signedness
? glsl_int_type() : glsl_uint_type());
1128 val
->type
->type
= (signedness
? glsl_int16_t_type() : glsl_uint16_t_type());
1131 val
->type
->type
= (signedness
? glsl_int8_t_type() : glsl_uint8_t_type());
1134 vtn_fail("Invalid int bit size: %u", bit_size
);
1136 val
->type
->length
= 1;
1140 case SpvOpTypeFloat
: {
1141 int bit_size
= w
[2];
1142 val
->type
->base_type
= vtn_base_type_scalar
;
1145 val
->type
->type
= glsl_float16_t_type();
1148 val
->type
->type
= glsl_float_type();
1151 val
->type
->type
= glsl_double_type();
1154 vtn_fail("Invalid float bit size: %u", bit_size
);
1156 val
->type
->length
= 1;
1160 case SpvOpTypeVector
: {
1161 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1162 unsigned elems
= w
[3];
1164 vtn_fail_if(base
->base_type
!= vtn_base_type_scalar
,
1165 "Base type for OpTypeVector must be a scalar");
1166 vtn_fail_if((elems
< 2 || elems
> 4) && (elems
!= 8) && (elems
!= 16),
1167 "Invalid component count for OpTypeVector");
1169 val
->type
->base_type
= vtn_base_type_vector
;
1170 val
->type
->type
= glsl_vector_type(glsl_get_base_type(base
->type
), elems
);
1171 val
->type
->length
= elems
;
1172 val
->type
->stride
= glsl_type_is_boolean(val
->type
->type
)
1173 ? 4 : glsl_get_bit_size(base
->type
) / 8;
1174 val
->type
->array_element
= base
;
1178 case SpvOpTypeMatrix
: {
1179 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1180 unsigned columns
= w
[3];
1182 vtn_fail_if(base
->base_type
!= vtn_base_type_vector
,
1183 "Base type for OpTypeMatrix must be a vector");
1184 vtn_fail_if(columns
< 2 || columns
> 4,
1185 "Invalid column count for OpTypeMatrix");
1187 val
->type
->base_type
= vtn_base_type_matrix
;
1188 val
->type
->type
= glsl_matrix_type(glsl_get_base_type(base
->type
),
1189 glsl_get_vector_elements(base
->type
),
1191 vtn_fail_if(glsl_type_is_error(val
->type
->type
),
1192 "Unsupported base type for OpTypeMatrix");
1193 assert(!glsl_type_is_error(val
->type
->type
));
1194 val
->type
->length
= columns
;
1195 val
->type
->array_element
= base
;
1196 val
->type
->row_major
= false;
1197 val
->type
->stride
= 0;
1201 case SpvOpTypeRuntimeArray
:
1202 case SpvOpTypeArray
: {
1203 struct vtn_type
*array_element
=
1204 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1206 if (opcode
== SpvOpTypeRuntimeArray
) {
1207 /* A length of 0 is used to denote unsized arrays */
1208 val
->type
->length
= 0;
1210 val
->type
->length
= vtn_constant_uint(b
, w
[3]);
1213 val
->type
->base_type
= vtn_base_type_array
;
1214 val
->type
->array_element
= array_element
;
1215 if (b
->shader
->info
.stage
== MESA_SHADER_KERNEL
)
1216 val
->type
->stride
= glsl_get_cl_size(array_element
->type
);
1218 vtn_foreach_decoration(b
, val
, array_stride_decoration_cb
, NULL
);
1219 val
->type
->type
= glsl_array_type(array_element
->type
, val
->type
->length
,
1224 case SpvOpTypeStruct
: {
1225 unsigned num_fields
= count
- 2;
1226 val
->type
->base_type
= vtn_base_type_struct
;
1227 val
->type
->length
= num_fields
;
1228 val
->type
->members
= ralloc_array(b
, struct vtn_type
*, num_fields
);
1229 val
->type
->offsets
= ralloc_array(b
, unsigned, num_fields
);
1230 val
->type
->packed
= false;
1232 NIR_VLA(struct glsl_struct_field
, fields
, count
);
1233 for (unsigned i
= 0; i
< num_fields
; i
++) {
1234 val
->type
->members
[i
] =
1235 vtn_value(b
, w
[i
+ 2], vtn_value_type_type
)->type
;
1236 fields
[i
] = (struct glsl_struct_field
) {
1237 .type
= val
->type
->members
[i
]->type
,
1238 .name
= ralloc_asprintf(b
, "field%d", i
),
1244 if (b
->shader
->info
.stage
== MESA_SHADER_KERNEL
) {
1245 unsigned offset
= 0;
1246 for (unsigned i
= 0; i
< num_fields
; i
++) {
1247 offset
= align(offset
, glsl_get_cl_alignment(fields
[i
].type
));
1248 fields
[i
].offset
= offset
;
1249 offset
+= glsl_get_cl_size(fields
[i
].type
);
1253 struct member_decoration_ctx ctx
= {
1254 .num_fields
= num_fields
,
1259 vtn_foreach_decoration(b
, val
, struct_member_decoration_cb
, &ctx
);
1260 vtn_foreach_decoration(b
, val
, struct_member_matrix_stride_cb
, &ctx
);
1262 vtn_foreach_decoration(b
, val
, struct_block_decoration_cb
, NULL
);
1264 const char *name
= val
->name
;
1266 if (val
->type
->block
|| val
->type
->buffer_block
) {
1267 /* Packing will be ignored since types coming from SPIR-V are
1268 * explicitly laid out.
1270 val
->type
->type
= glsl_interface_type(fields
, num_fields
,
1271 /* packing */ 0, false,
1272 name
? name
: "block");
1274 val
->type
->type
= glsl_struct_type(fields
, num_fields
,
1275 name
? name
: "struct", false);
1280 case SpvOpTypeFunction
: {
1281 val
->type
->base_type
= vtn_base_type_function
;
1282 val
->type
->type
= NULL
;
1284 val
->type
->return_type
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1286 const unsigned num_params
= count
- 3;
1287 val
->type
->length
= num_params
;
1288 val
->type
->params
= ralloc_array(b
, struct vtn_type
*, num_params
);
1289 for (unsigned i
= 0; i
< count
- 3; i
++) {
1290 val
->type
->params
[i
] =
1291 vtn_value(b
, w
[i
+ 3], vtn_value_type_type
)->type
;
1296 case SpvOpTypePointer
:
1297 case SpvOpTypeForwardPointer
: {
1298 /* We can't blindly push the value because it might be a forward
1301 val
= vtn_untyped_value(b
, w
[1]);
1303 SpvStorageClass storage_class
= w
[2];
1305 if (val
->value_type
== vtn_value_type_invalid
) {
1306 val
->value_type
= vtn_value_type_type
;
1307 val
->type
= rzalloc(b
, struct vtn_type
);
1308 val
->type
->id
= w
[1];
1309 val
->type
->base_type
= vtn_base_type_pointer
;
1310 val
->type
->storage_class
= storage_class
;
1312 /* These can actually be stored to nir_variables and used as SSA
1313 * values so they need a real glsl_type.
1315 enum vtn_variable_mode mode
= vtn_storage_class_to_mode(
1316 b
, storage_class
, NULL
, NULL
);
1317 val
->type
->type
= nir_address_format_to_glsl_type(
1318 vtn_mode_to_address_format(b
, mode
));
1320 vtn_fail_if(val
->type
->storage_class
!= storage_class
,
1321 "The storage classes of an OpTypePointer and any "
1322 "OpTypeForwardPointers that provide forward "
1323 "declarations of it must match.");
1326 if (opcode
== SpvOpTypePointer
) {
1327 vtn_fail_if(val
->type
->deref
!= NULL
,
1328 "While OpTypeForwardPointer can be used to provide a "
1329 "forward declaration of a pointer, OpTypePointer can "
1330 "only be used once for a given id.");
1332 val
->type
->deref
= vtn_value(b
, w
[3], vtn_value_type_type
)->type
;
1334 /* Only certain storage classes use ArrayStride. The others (in
1335 * particular Workgroup) are expected to be laid out by the driver.
1337 switch (storage_class
) {
1338 case SpvStorageClassUniform
:
1339 case SpvStorageClassPushConstant
:
1340 case SpvStorageClassStorageBuffer
:
1341 case SpvStorageClassPhysicalStorageBufferEXT
:
1342 vtn_foreach_decoration(b
, val
, array_stride_decoration_cb
, NULL
);
1345 /* Nothing to do. */
1349 if (b
->physical_ptrs
) {
1350 switch (storage_class
) {
1351 case SpvStorageClassFunction
:
1352 case SpvStorageClassWorkgroup
:
1353 case SpvStorageClassCrossWorkgroup
:
1354 val
->type
->stride
= align(glsl_get_cl_size(val
->type
->deref
->type
),
1355 glsl_get_cl_alignment(val
->type
->deref
->type
));
1365 case SpvOpTypeImage
: {
1366 val
->type
->base_type
= vtn_base_type_image
;
1368 const struct vtn_type
*sampled_type
=
1369 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1371 vtn_fail_if(sampled_type
->base_type
!= vtn_base_type_scalar
||
1372 glsl_get_bit_size(sampled_type
->type
) != 32,
1373 "Sampled type of OpTypeImage must be a 32-bit scalar");
1375 enum glsl_sampler_dim dim
;
1376 switch ((SpvDim
)w
[3]) {
1377 case SpvDim1D
: dim
= GLSL_SAMPLER_DIM_1D
; break;
1378 case SpvDim2D
: dim
= GLSL_SAMPLER_DIM_2D
; break;
1379 case SpvDim3D
: dim
= GLSL_SAMPLER_DIM_3D
; break;
1380 case SpvDimCube
: dim
= GLSL_SAMPLER_DIM_CUBE
; break;
1381 case SpvDimRect
: dim
= GLSL_SAMPLER_DIM_RECT
; break;
1382 case SpvDimBuffer
: dim
= GLSL_SAMPLER_DIM_BUF
; break;
1383 case SpvDimSubpassData
: dim
= GLSL_SAMPLER_DIM_SUBPASS
; break;
1385 vtn_fail("Invalid SPIR-V image dimensionality: %s (%u)",
1386 spirv_dim_to_string((SpvDim
)w
[3]), w
[3]);
1389 /* w[4]: as per Vulkan spec "Validation Rules within a Module",
1390 * The “Depth” operand of OpTypeImage is ignored.
1392 bool is_array
= w
[5];
1393 bool multisampled
= w
[6];
1394 unsigned sampled
= w
[7];
1395 SpvImageFormat format
= w
[8];
1398 val
->type
->access_qualifier
= w
[9];
1400 val
->type
->access_qualifier
= SpvAccessQualifierReadWrite
;
1403 if (dim
== GLSL_SAMPLER_DIM_2D
)
1404 dim
= GLSL_SAMPLER_DIM_MS
;
1405 else if (dim
== GLSL_SAMPLER_DIM_SUBPASS
)
1406 dim
= GLSL_SAMPLER_DIM_SUBPASS_MS
;
1408 vtn_fail("Unsupported multisampled image type");
1411 val
->type
->image_format
= translate_image_format(b
, format
);
1413 enum glsl_base_type sampled_base_type
=
1414 glsl_get_base_type(sampled_type
->type
);
1416 val
->type
->sampled
= true;
1417 val
->type
->type
= glsl_sampler_type(dim
, false, is_array
,
1419 } else if (sampled
== 2) {
1420 val
->type
->sampled
= false;
1421 val
->type
->type
= glsl_image_type(dim
, is_array
, sampled_base_type
);
1423 vtn_fail("We need to know if the image will be sampled");
1428 case SpvOpTypeSampledImage
:
1429 val
->type
->base_type
= vtn_base_type_sampled_image
;
1430 val
->type
->image
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1431 val
->type
->type
= val
->type
->image
->type
;
1434 case SpvOpTypeSampler
:
1435 /* The actual sampler type here doesn't really matter. It gets
1436 * thrown away the moment you combine it with an image. What really
1437 * matters is that it's a sampler type as opposed to an integer type
1438 * so the backend knows what to do.
1440 val
->type
->base_type
= vtn_base_type_sampler
;
1441 val
->type
->type
= glsl_bare_sampler_type();
1444 case SpvOpTypeOpaque
:
1445 case SpvOpTypeEvent
:
1446 case SpvOpTypeDeviceEvent
:
1447 case SpvOpTypeReserveId
:
1448 case SpvOpTypeQueue
:
1451 vtn_fail_with_opcode("Unhandled opcode", opcode
);
1454 vtn_foreach_decoration(b
, val
, type_decoration_cb
, NULL
);
1456 if (val
->type
->base_type
== vtn_base_type_struct
&&
1457 (val
->type
->block
|| val
->type
->buffer_block
)) {
1458 for (unsigned i
= 0; i
< val
->type
->length
; i
++) {
1459 vtn_fail_if(vtn_type_contains_block(b
, val
->type
->members
[i
]),
1460 "Block and BufferBlock decorations cannot decorate a "
1461 "structure type that is nested at any level inside "
1462 "another structure type decorated with Block or "
1468 static nir_constant
*
1469 vtn_null_constant(struct vtn_builder
*b
, struct vtn_type
*type
)
1471 nir_constant
*c
= rzalloc(b
, nir_constant
);
1473 switch (type
->base_type
) {
1474 case vtn_base_type_scalar
:
1475 case vtn_base_type_vector
:
1476 /* Nothing to do here. It's already initialized to zero */
1479 case vtn_base_type_pointer
: {
1480 enum vtn_variable_mode mode
= vtn_storage_class_to_mode(
1481 b
, type
->storage_class
, type
->deref
, NULL
);
1482 nir_address_format addr_format
= vtn_mode_to_address_format(b
, mode
);
1484 const nir_const_value
*null_value
= nir_address_format_null_value(addr_format
);
1485 memcpy(c
->values
, null_value
,
1486 sizeof(nir_const_value
) * nir_address_format_num_components(addr_format
));
1490 case vtn_base_type_void
:
1491 case vtn_base_type_image
:
1492 case vtn_base_type_sampler
:
1493 case vtn_base_type_sampled_image
:
1494 case vtn_base_type_function
:
1495 /* For those we have to return something but it doesn't matter what. */
1498 case vtn_base_type_matrix
:
1499 case vtn_base_type_array
:
1500 vtn_assert(type
->length
> 0);
1501 c
->num_elements
= type
->length
;
1502 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1504 c
->elements
[0] = vtn_null_constant(b
, type
->array_element
);
1505 for (unsigned i
= 1; i
< c
->num_elements
; i
++)
1506 c
->elements
[i
] = c
->elements
[0];
1509 case vtn_base_type_struct
:
1510 c
->num_elements
= type
->length
;
1511 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1512 for (unsigned i
= 0; i
< c
->num_elements
; i
++)
1513 c
->elements
[i
] = vtn_null_constant(b
, type
->members
[i
]);
1517 vtn_fail("Invalid type for null constant");
1524 spec_constant_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*v
,
1525 int member
, const struct vtn_decoration
*dec
,
1528 vtn_assert(member
== -1);
1529 if (dec
->decoration
!= SpvDecorationSpecId
)
1532 struct spec_constant_value
*const_value
= data
;
1534 for (unsigned i
= 0; i
< b
->num_specializations
; i
++) {
1535 if (b
->specializations
[i
].id
== dec
->operands
[0]) {
1536 if (const_value
->is_double
)
1537 const_value
->data64
= b
->specializations
[i
].data64
;
1539 const_value
->data32
= b
->specializations
[i
].data32
;
1546 get_specialization(struct vtn_builder
*b
, struct vtn_value
*val
,
1547 uint32_t const_value
)
1549 struct spec_constant_value data
;
1550 data
.is_double
= false;
1551 data
.data32
= const_value
;
1552 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1557 get_specialization64(struct vtn_builder
*b
, struct vtn_value
*val
,
1558 uint64_t const_value
)
1560 struct spec_constant_value data
;
1561 data
.is_double
= true;
1562 data
.data64
= const_value
;
1563 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1568 handle_workgroup_size_decoration_cb(struct vtn_builder
*b
,
1569 struct vtn_value
*val
,
1571 const struct vtn_decoration
*dec
,
1574 vtn_assert(member
== -1);
1575 if (dec
->decoration
!= SpvDecorationBuiltIn
||
1576 dec
->operands
[0] != SpvBuiltInWorkgroupSize
)
1579 vtn_assert(val
->type
->type
== glsl_vector_type(GLSL_TYPE_UINT
, 3));
1580 b
->workgroup_size_builtin
= val
;
1584 vtn_handle_constant(struct vtn_builder
*b
, SpvOp opcode
,
1585 const uint32_t *w
, unsigned count
)
1587 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_constant
);
1588 val
->constant
= rzalloc(b
, nir_constant
);
1590 case SpvOpConstantTrue
:
1591 case SpvOpConstantFalse
:
1592 case SpvOpSpecConstantTrue
:
1593 case SpvOpSpecConstantFalse
: {
1594 vtn_fail_if(val
->type
->type
!= glsl_bool_type(),
1595 "Result type of %s must be OpTypeBool",
1596 spirv_op_to_string(opcode
));
1598 uint32_t int_val
= (opcode
== SpvOpConstantTrue
||
1599 opcode
== SpvOpSpecConstantTrue
);
1601 if (opcode
== SpvOpSpecConstantTrue
||
1602 opcode
== SpvOpSpecConstantFalse
)
1603 int_val
= get_specialization(b
, val
, int_val
);
1605 val
->constant
->values
[0].b
= int_val
!= 0;
1609 case SpvOpConstant
: {
1610 vtn_fail_if(val
->type
->base_type
!= vtn_base_type_scalar
,
1611 "Result type of %s must be a scalar",
1612 spirv_op_to_string(opcode
));
1613 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1616 val
->constant
->values
[0].u64
= vtn_u64_literal(&w
[3]);
1619 val
->constant
->values
[0].u32
= w
[3];
1622 val
->constant
->values
[0].u16
= w
[3];
1625 val
->constant
->values
[0].u8
= w
[3];
1628 vtn_fail("Unsupported SpvOpConstant bit size: %u", bit_size
);
1633 case SpvOpSpecConstant
: {
1634 vtn_fail_if(val
->type
->base_type
!= vtn_base_type_scalar
,
1635 "Result type of %s must be a scalar",
1636 spirv_op_to_string(opcode
));
1637 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1640 val
->constant
->values
[0].u64
=
1641 get_specialization64(b
, val
, vtn_u64_literal(&w
[3]));
1644 val
->constant
->values
[0].u32
= get_specialization(b
, val
, w
[3]);
1647 val
->constant
->values
[0].u16
= get_specialization(b
, val
, w
[3]);
1650 val
->constant
->values
[0].u8
= get_specialization(b
, val
, w
[3]);
1653 vtn_fail("Unsupported SpvOpSpecConstant bit size");
1658 case SpvOpSpecConstantComposite
:
1659 case SpvOpConstantComposite
: {
1660 unsigned elem_count
= count
- 3;
1661 vtn_fail_if(elem_count
!= val
->type
->length
,
1662 "%s has %u constituents, expected %u",
1663 spirv_op_to_string(opcode
), elem_count
, val
->type
->length
);
1665 nir_constant
**elems
= ralloc_array(b
, nir_constant
*, elem_count
);
1666 for (unsigned i
= 0; i
< elem_count
; i
++) {
1667 struct vtn_value
*val
= vtn_untyped_value(b
, w
[i
+ 3]);
1669 if (val
->value_type
== vtn_value_type_constant
) {
1670 elems
[i
] = val
->constant
;
1672 vtn_fail_if(val
->value_type
!= vtn_value_type_undef
,
1673 "only constants or undefs allowed for "
1674 "SpvOpConstantComposite");
1675 /* to make it easier, just insert a NULL constant for now */
1676 elems
[i
] = vtn_null_constant(b
, val
->type
);
1680 switch (val
->type
->base_type
) {
1681 case vtn_base_type_vector
: {
1682 assert(glsl_type_is_vector(val
->type
->type
));
1683 for (unsigned i
= 0; i
< elem_count
; i
++)
1684 val
->constant
->values
[i
] = elems
[i
]->values
[0];
1688 case vtn_base_type_matrix
:
1689 case vtn_base_type_struct
:
1690 case vtn_base_type_array
:
1691 ralloc_steal(val
->constant
, elems
);
1692 val
->constant
->num_elements
= elem_count
;
1693 val
->constant
->elements
= elems
;
1697 vtn_fail("Result type of %s must be a composite type",
1698 spirv_op_to_string(opcode
));
1703 case SpvOpSpecConstantOp
: {
1704 SpvOp opcode
= get_specialization(b
, val
, w
[3]);
1706 case SpvOpVectorShuffle
: {
1707 struct vtn_value
*v0
= &b
->values
[w
[4]];
1708 struct vtn_value
*v1
= &b
->values
[w
[5]];
1710 vtn_assert(v0
->value_type
== vtn_value_type_constant
||
1711 v0
->value_type
== vtn_value_type_undef
);
1712 vtn_assert(v1
->value_type
== vtn_value_type_constant
||
1713 v1
->value_type
== vtn_value_type_undef
);
1715 unsigned len0
= glsl_get_vector_elements(v0
->type
->type
);
1716 unsigned len1
= glsl_get_vector_elements(v1
->type
->type
);
1718 vtn_assert(len0
+ len1
< 16);
1720 unsigned bit_size
= glsl_get_bit_size(val
->type
->type
);
1721 unsigned bit_size0
= glsl_get_bit_size(v0
->type
->type
);
1722 unsigned bit_size1
= glsl_get_bit_size(v1
->type
->type
);
1724 vtn_assert(bit_size
== bit_size0
&& bit_size
== bit_size1
);
1725 (void)bit_size0
; (void)bit_size1
;
1727 nir_const_value undef
= { .u64
= 0xdeadbeefdeadbeef };
1728 nir_const_value combined
[NIR_MAX_VEC_COMPONENTS
* 2];
1730 if (v0
->value_type
== vtn_value_type_constant
) {
1731 for (unsigned i
= 0; i
< len0
; i
++)
1732 combined
[i
] = v0
->constant
->values
[i
];
1734 if (v1
->value_type
== vtn_value_type_constant
) {
1735 for (unsigned i
= 0; i
< len1
; i
++)
1736 combined
[len0
+ i
] = v1
->constant
->values
[i
];
1739 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1740 uint32_t comp
= w
[i
+ 6];
1741 if (comp
== (uint32_t)-1) {
1742 /* If component is not used, set the value to a known constant
1743 * to detect if it is wrongly used.
1745 val
->constant
->values
[j
] = undef
;
1747 vtn_fail_if(comp
>= len0
+ len1
,
1748 "All Component literals must either be FFFFFFFF "
1749 "or in [0, N - 1] (inclusive).");
1750 val
->constant
->values
[j
] = combined
[comp
];
1756 case SpvOpCompositeExtract
:
1757 case SpvOpCompositeInsert
: {
1758 struct vtn_value
*comp
;
1759 unsigned deref_start
;
1760 struct nir_constant
**c
;
1761 if (opcode
== SpvOpCompositeExtract
) {
1762 comp
= vtn_value(b
, w
[4], vtn_value_type_constant
);
1764 c
= &comp
->constant
;
1766 comp
= vtn_value(b
, w
[5], vtn_value_type_constant
);
1768 val
->constant
= nir_constant_clone(comp
->constant
,
1774 const struct vtn_type
*type
= comp
->type
;
1775 for (unsigned i
= deref_start
; i
< count
; i
++) {
1776 vtn_fail_if(w
[i
] > type
->length
,
1777 "%uth index of %s is %u but the type has only "
1778 "%u elements", i
- deref_start
,
1779 spirv_op_to_string(opcode
), w
[i
], type
->length
);
1781 switch (type
->base_type
) {
1782 case vtn_base_type_vector
:
1784 type
= type
->array_element
;
1787 case vtn_base_type_matrix
:
1788 case vtn_base_type_array
:
1789 c
= &(*c
)->elements
[w
[i
]];
1790 type
= type
->array_element
;
1793 case vtn_base_type_struct
:
1794 c
= &(*c
)->elements
[w
[i
]];
1795 type
= type
->members
[w
[i
]];
1799 vtn_fail("%s must only index into composite types",
1800 spirv_op_to_string(opcode
));
1804 if (opcode
== SpvOpCompositeExtract
) {
1808 unsigned num_components
= type
->length
;
1809 for (unsigned i
= 0; i
< num_components
; i
++)
1810 val
->constant
->values
[i
] = (*c
)->values
[elem
+ i
];
1813 struct vtn_value
*insert
=
1814 vtn_value(b
, w
[4], vtn_value_type_constant
);
1815 vtn_assert(insert
->type
== type
);
1817 *c
= insert
->constant
;
1819 unsigned num_components
= type
->length
;
1820 for (unsigned i
= 0; i
< num_components
; i
++)
1821 (*c
)->values
[elem
+ i
] = insert
->constant
->values
[i
];
1829 nir_alu_type dst_alu_type
= nir_get_nir_type_for_glsl_type(val
->type
->type
);
1830 nir_alu_type src_alu_type
= dst_alu_type
;
1831 unsigned num_components
= glsl_get_vector_elements(val
->type
->type
);
1834 vtn_assert(count
<= 7);
1840 /* We have a source in a conversion */
1842 nir_get_nir_type_for_glsl_type(
1843 vtn_value(b
, w
[4], vtn_value_type_constant
)->type
->type
);
1844 /* We use the bitsize of the conversion source to evaluate the opcode later */
1845 bit_size
= glsl_get_bit_size(
1846 vtn_value(b
, w
[4], vtn_value_type_constant
)->type
->type
);
1849 bit_size
= glsl_get_bit_size(val
->type
->type
);
1852 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(b
, opcode
, &swap
,
1853 nir_alu_type_get_type_size(src_alu_type
),
1854 nir_alu_type_get_type_size(dst_alu_type
));
1855 nir_const_value src
[3][NIR_MAX_VEC_COMPONENTS
];
1857 for (unsigned i
= 0; i
< count
- 4; i
++) {
1858 struct vtn_value
*src_val
=
1859 vtn_value(b
, w
[4 + i
], vtn_value_type_constant
);
1861 /* If this is an unsized source, pull the bit size from the
1862 * source; otherwise, we'll use the bit size from the destination.
1864 if (!nir_alu_type_get_type_size(nir_op_infos
[op
].input_types
[i
]))
1865 bit_size
= glsl_get_bit_size(src_val
->type
->type
);
1867 unsigned src_comps
= nir_op_infos
[op
].input_sizes
[i
] ?
1868 nir_op_infos
[op
].input_sizes
[i
] :
1871 unsigned j
= swap
? 1 - i
: i
;
1872 for (unsigned c
= 0; c
< src_comps
; c
++)
1873 src
[j
][c
] = src_val
->constant
->values
[c
];
1876 /* fix up fixed size sources */
1883 for (unsigned i
= 0; i
< num_components
; ++i
) {
1885 case 64: src
[1][i
].u32
= src
[1][i
].u64
; break;
1886 case 16: src
[1][i
].u32
= src
[1][i
].u16
; break;
1887 case 8: src
[1][i
].u32
= src
[1][i
].u8
; break;
1896 nir_const_value
*srcs
[3] = {
1897 src
[0], src
[1], src
[2],
1899 nir_eval_const_opcode(op
, val
->constant
->values
,
1900 num_components
, bit_size
, srcs
,
1901 b
->shader
->info
.float_controls_execution_mode
);
1908 case SpvOpConstantNull
:
1909 val
->constant
= vtn_null_constant(b
, val
->type
);
1912 case SpvOpConstantSampler
:
1913 vtn_fail("OpConstantSampler requires Kernel Capability");
1917 vtn_fail_with_opcode("Unhandled opcode", opcode
);
1920 /* Now that we have the value, update the workgroup size if needed */
1921 vtn_foreach_decoration(b
, val
, handle_workgroup_size_decoration_cb
, NULL
);
1924 SpvMemorySemanticsMask
1925 vtn_storage_class_to_memory_semantics(SpvStorageClass sc
)
1928 case SpvStorageClassStorageBuffer
:
1929 case SpvStorageClassPhysicalStorageBufferEXT
:
1930 return SpvMemorySemanticsUniformMemoryMask
;
1931 case SpvStorageClassWorkgroup
:
1932 return SpvMemorySemanticsWorkgroupMemoryMask
;
1934 return SpvMemorySemanticsMaskNone
;
1939 vtn_split_barrier_semantics(struct vtn_builder
*b
,
1940 SpvMemorySemanticsMask semantics
,
1941 SpvMemorySemanticsMask
*before
,
1942 SpvMemorySemanticsMask
*after
)
1944 /* For memory semantics embedded in operations, we split them into up to
1945 * two barriers, to be added before and after the operation. This is less
1946 * strict than if we propagated until the final backend stage, but still
1947 * result in correct execution.
1949 * A further improvement could be pipe this information (and use!) into the
1950 * next compiler layers, at the expense of making the handling of barriers
1954 *before
= SpvMemorySemanticsMaskNone
;
1955 *after
= SpvMemorySemanticsMaskNone
;
1957 SpvMemorySemanticsMask order_semantics
=
1958 semantics
& (SpvMemorySemanticsAcquireMask
|
1959 SpvMemorySemanticsReleaseMask
|
1960 SpvMemorySemanticsAcquireReleaseMask
|
1961 SpvMemorySemanticsSequentiallyConsistentMask
);
1963 if (util_bitcount(order_semantics
) > 1) {
1964 /* Old GLSLang versions incorrectly set all the ordering bits. This was
1965 * fixed in c51287d744fb6e7e9ccc09f6f8451e6c64b1dad6 of glslang repo,
1966 * and it is in GLSLang since revision "SPIRV99.1321" (from Jul-2016).
1968 vtn_warn("Multiple memory ordering semantics specified, "
1969 "assuming AcquireRelease.");
1970 order_semantics
= SpvMemorySemanticsAcquireReleaseMask
;
1973 const SpvMemorySemanticsMask av_vis_semantics
=
1974 semantics
& (SpvMemorySemanticsMakeAvailableMask
|
1975 SpvMemorySemanticsMakeVisibleMask
);
1977 const SpvMemorySemanticsMask storage_semantics
=
1978 semantics
& (SpvMemorySemanticsUniformMemoryMask
|
1979 SpvMemorySemanticsSubgroupMemoryMask
|
1980 SpvMemorySemanticsWorkgroupMemoryMask
|
1981 SpvMemorySemanticsCrossWorkgroupMemoryMask
|
1982 SpvMemorySemanticsAtomicCounterMemoryMask
|
1983 SpvMemorySemanticsImageMemoryMask
|
1984 SpvMemorySemanticsOutputMemoryMask
);
1986 const SpvMemorySemanticsMask other_semantics
=
1987 semantics
& ~(order_semantics
| av_vis_semantics
| storage_semantics
);
1989 if (other_semantics
)
1990 vtn_warn("Ignoring unhandled memory semantics: %u\n", other_semantics
);
1992 /* SequentiallyConsistent is treated as AcquireRelease. */
1994 /* The RELEASE barrier happens BEFORE the operation, and it is usually
1995 * associated with a Store. All the write operations with a matching
1996 * semantics will not be reordered after the Store.
1998 if (order_semantics
& (SpvMemorySemanticsReleaseMask
|
1999 SpvMemorySemanticsAcquireReleaseMask
|
2000 SpvMemorySemanticsSequentiallyConsistentMask
)) {
2001 *before
|= SpvMemorySemanticsReleaseMask
| storage_semantics
;
2004 /* The ACQUIRE barrier happens AFTER the operation, and it is usually
2005 * associated with a Load. All the operations with a matching semantics
2006 * will not be reordered before the Load.
2008 if (order_semantics
& (SpvMemorySemanticsAcquireMask
|
2009 SpvMemorySemanticsAcquireReleaseMask
|
2010 SpvMemorySemanticsSequentiallyConsistentMask
)) {
2011 *after
|= SpvMemorySemanticsAcquireMask
| storage_semantics
;
2014 if (av_vis_semantics
& SpvMemorySemanticsMakeVisibleMask
)
2015 *before
|= SpvMemorySemanticsMakeVisibleMask
| storage_semantics
;
2017 if (av_vis_semantics
& SpvMemorySemanticsMakeAvailableMask
)
2018 *after
|= SpvMemorySemanticsMakeAvailableMask
| storage_semantics
;
2022 vtn_emit_scoped_memory_barrier(struct vtn_builder
*b
, SpvScope scope
,
2023 SpvMemorySemanticsMask semantics
)
2025 nir_memory_semantics nir_semantics
= 0;
2027 SpvMemorySemanticsMask order_semantics
=
2028 semantics
& (SpvMemorySemanticsAcquireMask
|
2029 SpvMemorySemanticsReleaseMask
|
2030 SpvMemorySemanticsAcquireReleaseMask
|
2031 SpvMemorySemanticsSequentiallyConsistentMask
);
2033 if (util_bitcount(order_semantics
) > 1) {
2034 /* Old GLSLang versions incorrectly set all the ordering bits. This was
2035 * fixed in c51287d744fb6e7e9ccc09f6f8451e6c64b1dad6 of glslang repo,
2036 * and it is in GLSLang since revision "SPIRV99.1321" (from Jul-2016).
2038 vtn_warn("Multiple memory ordering semantics bits specified, "
2039 "assuming AcquireRelease.");
2040 order_semantics
= SpvMemorySemanticsAcquireReleaseMask
;
2043 switch (order_semantics
) {
2045 /* Not an ordering barrier. */
2048 case SpvMemorySemanticsAcquireMask
:
2049 nir_semantics
= NIR_MEMORY_ACQUIRE
;
2052 case SpvMemorySemanticsReleaseMask
:
2053 nir_semantics
= NIR_MEMORY_RELEASE
;
2056 case SpvMemorySemanticsSequentiallyConsistentMask
:
2057 /* Fall through. Treated as AcquireRelease in Vulkan. */
2058 case SpvMemorySemanticsAcquireReleaseMask
:
2059 nir_semantics
= NIR_MEMORY_ACQUIRE
| NIR_MEMORY_RELEASE
;
2063 unreachable("Invalid memory order semantics");
2066 if (semantics
& SpvMemorySemanticsMakeAvailableMask
) {
2067 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
2068 "To use MakeAvailable memory semantics the VulkanMemoryModel "
2069 "capability must be declared.");
2070 nir_semantics
|= NIR_MEMORY_MAKE_AVAILABLE
;
2073 if (semantics
& SpvMemorySemanticsMakeVisibleMask
) {
2074 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
2075 "To use MakeVisible memory semantics the VulkanMemoryModel "
2076 "capability must be declared.");
2077 nir_semantics
|= NIR_MEMORY_MAKE_VISIBLE
;
2080 /* Vulkan Environment for SPIR-V says "SubgroupMemory, CrossWorkgroupMemory,
2081 * and AtomicCounterMemory are ignored".
2083 semantics
&= ~(SpvMemorySemanticsSubgroupMemoryMask
|
2084 SpvMemorySemanticsCrossWorkgroupMemoryMask
|
2085 SpvMemorySemanticsAtomicCounterMemoryMask
);
2087 /* TODO: Consider adding nir_var_mem_image mode to NIR so it can be used
2088 * for SpvMemorySemanticsImageMemoryMask.
2091 nir_variable_mode modes
= 0;
2092 if (semantics
& (SpvMemorySemanticsUniformMemoryMask
|
2093 SpvMemorySemanticsImageMemoryMask
))
2094 modes
|= nir_var_mem_ubo
| nir_var_mem_ssbo
| nir_var_uniform
;
2095 if (semantics
& SpvMemorySemanticsWorkgroupMemoryMask
)
2096 modes
|= nir_var_mem_shared
;
2097 if (semantics
& SpvMemorySemanticsOutputMemoryMask
) {
2098 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
2099 "To use Output memory semantics, the VulkanMemoryModel "
2100 "capability must be declared.");
2101 modes
|= nir_var_shader_out
;
2104 /* No barrier to add. */
2105 if (nir_semantics
== 0 || modes
== 0)
2108 nir_scope nir_scope
;
2110 case SpvScopeDevice
:
2111 vtn_fail_if(b
->options
->caps
.vk_memory_model
&&
2112 !b
->options
->caps
.vk_memory_model_device_scope
,
2113 "If the Vulkan memory model is declared and any instruction "
2114 "uses Device scope, the VulkanMemoryModelDeviceScope "
2115 "capability must be declared.");
2116 nir_scope
= NIR_SCOPE_DEVICE
;
2119 case SpvScopeQueueFamily
:
2120 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
2121 "To use Queue Family scope, the VulkanMemoryModel capability "
2122 "must be declared.");
2123 nir_scope
= NIR_SCOPE_QUEUE_FAMILY
;
2126 case SpvScopeWorkgroup
:
2127 nir_scope
= NIR_SCOPE_WORKGROUP
;
2130 case SpvScopeSubgroup
:
2131 nir_scope
= NIR_SCOPE_SUBGROUP
;
2134 case SpvScopeInvocation
:
2135 nir_scope
= NIR_SCOPE_INVOCATION
;
2139 vtn_fail("Invalid memory scope");
2142 nir_intrinsic_instr
*intrin
=
2143 nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_scoped_memory_barrier
);
2144 nir_intrinsic_set_memory_semantics(intrin
, nir_semantics
);
2146 nir_intrinsic_set_memory_modes(intrin
, modes
);
2147 nir_intrinsic_set_memory_scope(intrin
, nir_scope
);
2148 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2151 struct vtn_ssa_value
*
2152 vtn_create_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
2154 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
2157 if (!glsl_type_is_vector_or_scalar(type
)) {
2158 unsigned elems
= glsl_get_length(type
);
2159 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
2160 for (unsigned i
= 0; i
< elems
; i
++) {
2161 const struct glsl_type
*child_type
;
2163 switch (glsl_get_base_type(type
)) {
2165 case GLSL_TYPE_UINT
:
2166 case GLSL_TYPE_INT16
:
2167 case GLSL_TYPE_UINT16
:
2168 case GLSL_TYPE_UINT8
:
2169 case GLSL_TYPE_INT8
:
2170 case GLSL_TYPE_INT64
:
2171 case GLSL_TYPE_UINT64
:
2172 case GLSL_TYPE_BOOL
:
2173 case GLSL_TYPE_FLOAT
:
2174 case GLSL_TYPE_FLOAT16
:
2175 case GLSL_TYPE_DOUBLE
:
2176 child_type
= glsl_get_column_type(type
);
2178 case GLSL_TYPE_ARRAY
:
2179 child_type
= glsl_get_array_element(type
);
2181 case GLSL_TYPE_STRUCT
:
2182 case GLSL_TYPE_INTERFACE
:
2183 child_type
= glsl_get_struct_field(type
, i
);
2186 vtn_fail("unkown base type");
2189 val
->elems
[i
] = vtn_create_ssa_value(b
, child_type
);
2197 vtn_tex_src(struct vtn_builder
*b
, unsigned index
, nir_tex_src_type type
)
2200 src
.src
= nir_src_for_ssa(vtn_ssa_value(b
, index
)->def
);
2201 src
.src_type
= type
;
2206 image_operand_arg(struct vtn_builder
*b
, const uint32_t *w
, uint32_t count
,
2207 uint32_t mask_idx
, SpvImageOperandsMask op
)
2209 static const SpvImageOperandsMask ops_with_arg
=
2210 SpvImageOperandsBiasMask
|
2211 SpvImageOperandsLodMask
|
2212 SpvImageOperandsGradMask
|
2213 SpvImageOperandsConstOffsetMask
|
2214 SpvImageOperandsOffsetMask
|
2215 SpvImageOperandsConstOffsetsMask
|
2216 SpvImageOperandsSampleMask
|
2217 SpvImageOperandsMinLodMask
|
2218 SpvImageOperandsMakeTexelAvailableMask
|
2219 SpvImageOperandsMakeTexelVisibleMask
;
2221 assert(util_bitcount(op
) == 1);
2222 assert(w
[mask_idx
] & op
);
2223 assert(op
& ops_with_arg
);
2225 uint32_t idx
= util_bitcount(w
[mask_idx
] & (op
- 1) & ops_with_arg
) + 1;
2227 /* Adjust indices for operands with two arguments. */
2228 static const SpvImageOperandsMask ops_with_two_args
=
2229 SpvImageOperandsGradMask
;
2230 idx
+= util_bitcount(w
[mask_idx
] & (op
- 1) & ops_with_two_args
);
2234 vtn_fail_if(idx
+ (op
& ops_with_two_args
? 1 : 0) >= count
,
2235 "Image op claims to have %s but does not enough "
2236 "following operands", spirv_imageoperands_to_string(op
));
2242 vtn_handle_texture(struct vtn_builder
*b
, SpvOp opcode
,
2243 const uint32_t *w
, unsigned count
)
2245 if (opcode
== SpvOpSampledImage
) {
2246 struct vtn_value
*val
=
2247 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
2248 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
2249 val
->sampled_image
->image
=
2250 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2251 val
->sampled_image
->sampler
=
2252 vtn_value(b
, w
[4], vtn_value_type_pointer
)->pointer
;
2254 } else if (opcode
== SpvOpImage
) {
2255 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[3]);
2256 if (src_val
->value_type
== vtn_value_type_sampled_image
) {
2257 vtn_push_value_pointer(b
, w
[2], src_val
->sampled_image
->image
);
2259 vtn_assert(src_val
->value_type
== vtn_value_type_pointer
);
2260 vtn_push_value_pointer(b
, w
[2], src_val
->pointer
);
2265 struct vtn_type
*ret_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2267 struct vtn_sampled_image sampled
;
2268 struct vtn_value
*sampled_val
= vtn_untyped_value(b
, w
[3]);
2269 if (sampled_val
->value_type
== vtn_value_type_sampled_image
) {
2270 sampled
= *sampled_val
->sampled_image
;
2272 vtn_assert(sampled_val
->value_type
== vtn_value_type_pointer
);
2273 sampled
.image
= NULL
;
2274 sampled
.sampler
= sampled_val
->pointer
;
2277 const struct glsl_type
*image_type
= sampled_val
->type
->type
;
2278 const enum glsl_sampler_dim sampler_dim
= glsl_get_sampler_dim(image_type
);
2279 const bool is_array
= glsl_sampler_type_is_array(image_type
);
2280 nir_alu_type dest_type
= nir_type_invalid
;
2282 /* Figure out the base texture operation */
2285 case SpvOpImageSampleImplicitLod
:
2286 case SpvOpImageSampleDrefImplicitLod
:
2287 case SpvOpImageSampleProjImplicitLod
:
2288 case SpvOpImageSampleProjDrefImplicitLod
:
2289 texop
= nir_texop_tex
;
2292 case SpvOpImageSampleExplicitLod
:
2293 case SpvOpImageSampleDrefExplicitLod
:
2294 case SpvOpImageSampleProjExplicitLod
:
2295 case SpvOpImageSampleProjDrefExplicitLod
:
2296 texop
= nir_texop_txl
;
2299 case SpvOpImageFetch
:
2300 if (sampler_dim
== GLSL_SAMPLER_DIM_MS
) {
2301 texop
= nir_texop_txf_ms
;
2303 texop
= nir_texop_txf
;
2307 case SpvOpImageGather
:
2308 case SpvOpImageDrefGather
:
2309 texop
= nir_texop_tg4
;
2312 case SpvOpImageQuerySizeLod
:
2313 case SpvOpImageQuerySize
:
2314 texop
= nir_texop_txs
;
2315 dest_type
= nir_type_int
;
2318 case SpvOpImageQueryLod
:
2319 texop
= nir_texop_lod
;
2320 dest_type
= nir_type_float
;
2323 case SpvOpImageQueryLevels
:
2324 texop
= nir_texop_query_levels
;
2325 dest_type
= nir_type_int
;
2328 case SpvOpImageQuerySamples
:
2329 texop
= nir_texop_texture_samples
;
2330 dest_type
= nir_type_int
;
2334 vtn_fail_with_opcode("Unhandled opcode", opcode
);
2337 nir_tex_src srcs
[10]; /* 10 should be enough */
2338 nir_tex_src
*p
= srcs
;
2340 nir_deref_instr
*sampler
= vtn_pointer_to_deref(b
, sampled
.sampler
);
2341 nir_deref_instr
*texture
=
2342 sampled
.image
? vtn_pointer_to_deref(b
, sampled
.image
) : sampler
;
2344 p
->src
= nir_src_for_ssa(&texture
->dest
.ssa
);
2345 p
->src_type
= nir_tex_src_texture_deref
;
2355 /* These operations require a sampler */
2356 p
->src
= nir_src_for_ssa(&sampler
->dest
.ssa
);
2357 p
->src_type
= nir_tex_src_sampler_deref
;
2361 case nir_texop_txf_ms
:
2363 case nir_texop_query_levels
:
2364 case nir_texop_texture_samples
:
2365 case nir_texop_samples_identical
:
2368 case nir_texop_txf_ms_fb
:
2369 vtn_fail("unexpected nir_texop_txf_ms_fb");
2371 case nir_texop_txf_ms_mcs
:
2372 vtn_fail("unexpected nir_texop_txf_ms_mcs");
2373 case nir_texop_tex_prefetch
:
2374 vtn_fail("unexpected nir_texop_tex_prefetch");
2379 struct nir_ssa_def
*coord
;
2380 unsigned coord_components
;
2382 case SpvOpImageSampleImplicitLod
:
2383 case SpvOpImageSampleExplicitLod
:
2384 case SpvOpImageSampleDrefImplicitLod
:
2385 case SpvOpImageSampleDrefExplicitLod
:
2386 case SpvOpImageSampleProjImplicitLod
:
2387 case SpvOpImageSampleProjExplicitLod
:
2388 case SpvOpImageSampleProjDrefImplicitLod
:
2389 case SpvOpImageSampleProjDrefExplicitLod
:
2390 case SpvOpImageFetch
:
2391 case SpvOpImageGather
:
2392 case SpvOpImageDrefGather
:
2393 case SpvOpImageQueryLod
: {
2394 /* All these types have the coordinate as their first real argument */
2395 switch (sampler_dim
) {
2396 case GLSL_SAMPLER_DIM_1D
:
2397 case GLSL_SAMPLER_DIM_BUF
:
2398 coord_components
= 1;
2400 case GLSL_SAMPLER_DIM_2D
:
2401 case GLSL_SAMPLER_DIM_RECT
:
2402 case GLSL_SAMPLER_DIM_MS
:
2403 coord_components
= 2;
2405 case GLSL_SAMPLER_DIM_3D
:
2406 case GLSL_SAMPLER_DIM_CUBE
:
2407 coord_components
= 3;
2410 vtn_fail("Invalid sampler type");
2413 if (is_array
&& texop
!= nir_texop_lod
)
2416 coord
= vtn_ssa_value(b
, w
[idx
++])->def
;
2417 p
->src
= nir_src_for_ssa(nir_channels(&b
->nb
, coord
,
2418 (1 << coord_components
) - 1));
2419 p
->src_type
= nir_tex_src_coord
;
2426 coord_components
= 0;
2431 case SpvOpImageSampleProjImplicitLod
:
2432 case SpvOpImageSampleProjExplicitLod
:
2433 case SpvOpImageSampleProjDrefImplicitLod
:
2434 case SpvOpImageSampleProjDrefExplicitLod
:
2435 /* These have the projector as the last coordinate component */
2436 p
->src
= nir_src_for_ssa(nir_channel(&b
->nb
, coord
, coord_components
));
2437 p
->src_type
= nir_tex_src_projector
;
2445 bool is_shadow
= false;
2446 unsigned gather_component
= 0;
2448 case SpvOpImageSampleDrefImplicitLod
:
2449 case SpvOpImageSampleDrefExplicitLod
:
2450 case SpvOpImageSampleProjDrefImplicitLod
:
2451 case SpvOpImageSampleProjDrefExplicitLod
:
2452 case SpvOpImageDrefGather
:
2453 /* These all have an explicit depth value as their next source */
2455 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_comparator
);
2458 case SpvOpImageGather
:
2459 /* This has a component as its next source */
2460 gather_component
= vtn_constant_uint(b
, w
[idx
++]);
2467 /* For OpImageQuerySizeLod, we always have an LOD */
2468 if (opcode
== SpvOpImageQuerySizeLod
)
2469 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
2471 /* Now we need to handle some number of optional arguments */
2472 struct vtn_value
*gather_offsets
= NULL
;
2474 uint32_t operands
= w
[idx
];
2476 if (operands
& SpvImageOperandsBiasMask
) {
2477 vtn_assert(texop
== nir_texop_tex
);
2478 texop
= nir_texop_txb
;
2479 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2480 SpvImageOperandsBiasMask
);
2481 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_bias
);
2484 if (operands
& SpvImageOperandsLodMask
) {
2485 vtn_assert(texop
== nir_texop_txl
|| texop
== nir_texop_txf
||
2486 texop
== nir_texop_txs
);
2487 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2488 SpvImageOperandsLodMask
);
2489 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_lod
);
2492 if (operands
& SpvImageOperandsGradMask
) {
2493 vtn_assert(texop
== nir_texop_txl
);
2494 texop
= nir_texop_txd
;
2495 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2496 SpvImageOperandsGradMask
);
2497 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_ddx
);
2498 (*p
++) = vtn_tex_src(b
, w
[arg
+ 1], nir_tex_src_ddy
);
2501 vtn_fail_if(util_bitcount(operands
& (SpvImageOperandsConstOffsetsMask
|
2502 SpvImageOperandsOffsetMask
|
2503 SpvImageOperandsConstOffsetMask
)) > 1,
2504 "At most one of the ConstOffset, Offset, and ConstOffsets "
2505 "image operands can be used on a given instruction.");
2507 if (operands
& SpvImageOperandsOffsetMask
) {
2508 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2509 SpvImageOperandsOffsetMask
);
2510 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_offset
);
2513 if (operands
& SpvImageOperandsConstOffsetMask
) {
2514 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2515 SpvImageOperandsConstOffsetMask
);
2516 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_offset
);
2519 if (operands
& SpvImageOperandsConstOffsetsMask
) {
2520 vtn_assert(texop
== nir_texop_tg4
);
2521 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2522 SpvImageOperandsConstOffsetsMask
);
2523 gather_offsets
= vtn_value(b
, w
[arg
], vtn_value_type_constant
);
2526 if (operands
& SpvImageOperandsSampleMask
) {
2527 vtn_assert(texop
== nir_texop_txf_ms
);
2528 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2529 SpvImageOperandsSampleMask
);
2530 texop
= nir_texop_txf_ms
;
2531 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_ms_index
);
2534 if (operands
& SpvImageOperandsMinLodMask
) {
2535 vtn_assert(texop
== nir_texop_tex
||
2536 texop
== nir_texop_txb
||
2537 texop
== nir_texop_txd
);
2538 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2539 SpvImageOperandsMinLodMask
);
2540 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_min_lod
);
2544 nir_tex_instr
*instr
= nir_tex_instr_create(b
->shader
, p
- srcs
);
2547 memcpy(instr
->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
2549 instr
->coord_components
= coord_components
;
2550 instr
->sampler_dim
= sampler_dim
;
2551 instr
->is_array
= is_array
;
2552 instr
->is_shadow
= is_shadow
;
2553 instr
->is_new_style_shadow
=
2554 is_shadow
&& glsl_get_components(ret_type
->type
) == 1;
2555 instr
->component
= gather_component
;
2557 if (sampled
.image
&& (sampled
.image
->access
& ACCESS_NON_UNIFORM
))
2558 instr
->texture_non_uniform
= true;
2560 if (sampled
.sampler
&& (sampled
.sampler
->access
& ACCESS_NON_UNIFORM
))
2561 instr
->sampler_non_uniform
= true;
2563 /* for non-query ops, get dest_type from sampler type */
2564 if (dest_type
== nir_type_invalid
) {
2565 switch (glsl_get_sampler_result_type(image_type
)) {
2566 case GLSL_TYPE_FLOAT
: dest_type
= nir_type_float
; break;
2567 case GLSL_TYPE_INT
: dest_type
= nir_type_int
; break;
2568 case GLSL_TYPE_UINT
: dest_type
= nir_type_uint
; break;
2569 case GLSL_TYPE_BOOL
: dest_type
= nir_type_bool
; break;
2571 vtn_fail("Invalid base type for sampler result");
2575 instr
->dest_type
= dest_type
;
2577 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
2578 nir_tex_instr_dest_size(instr
), 32, NULL
);
2580 vtn_assert(glsl_get_vector_elements(ret_type
->type
) ==
2581 nir_tex_instr_dest_size(instr
));
2583 if (gather_offsets
) {
2584 vtn_fail_if(gather_offsets
->type
->base_type
!= vtn_base_type_array
||
2585 gather_offsets
->type
->length
!= 4,
2586 "ConstOffsets must be an array of size four of vectors "
2587 "of two integer components");
2589 struct vtn_type
*vec_type
= gather_offsets
->type
->array_element
;
2590 vtn_fail_if(vec_type
->base_type
!= vtn_base_type_vector
||
2591 vec_type
->length
!= 2 ||
2592 !glsl_type_is_integer(vec_type
->type
),
2593 "ConstOffsets must be an array of size four of vectors "
2594 "of two integer components");
2596 unsigned bit_size
= glsl_get_bit_size(vec_type
->type
);
2597 for (uint32_t i
= 0; i
< 4; i
++) {
2598 const nir_const_value
*cvec
=
2599 gather_offsets
->constant
->elements
[i
]->values
;
2600 for (uint32_t j
= 0; j
< 2; j
++) {
2602 case 8: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i8
; break;
2603 case 16: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i16
; break;
2604 case 32: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i32
; break;
2605 case 64: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i64
; break;
2607 vtn_fail("Unsupported bit size: %u", bit_size
);
2613 struct vtn_ssa_value
*ssa
= vtn_create_ssa_value(b
, ret_type
->type
);
2614 ssa
->def
= &instr
->dest
.ssa
;
2615 vtn_push_ssa(b
, w
[2], ret_type
, ssa
);
2617 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
2621 fill_common_atomic_sources(struct vtn_builder
*b
, SpvOp opcode
,
2622 const uint32_t *w
, nir_src
*src
)
2625 case SpvOpAtomicIIncrement
:
2626 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
2629 case SpvOpAtomicIDecrement
:
2630 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
2633 case SpvOpAtomicISub
:
2635 nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
2638 case SpvOpAtomicCompareExchange
:
2639 case SpvOpAtomicCompareExchangeWeak
:
2640 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[8])->def
);
2641 src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
2644 case SpvOpAtomicExchange
:
2645 case SpvOpAtomicIAdd
:
2646 case SpvOpAtomicSMin
:
2647 case SpvOpAtomicUMin
:
2648 case SpvOpAtomicSMax
:
2649 case SpvOpAtomicUMax
:
2650 case SpvOpAtomicAnd
:
2652 case SpvOpAtomicXor
:
2653 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
2657 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
2661 static nir_ssa_def
*
2662 get_image_coord(struct vtn_builder
*b
, uint32_t value
)
2664 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, value
);
2666 /* The image_load_store intrinsics assume a 4-dim coordinate */
2667 unsigned dim
= glsl_get_vector_elements(coord
->type
);
2668 unsigned swizzle
[4];
2669 for (unsigned i
= 0; i
< 4; i
++)
2670 swizzle
[i
] = MIN2(i
, dim
- 1);
2672 return nir_swizzle(&b
->nb
, coord
->def
, swizzle
, 4);
2675 static nir_ssa_def
*
2676 expand_to_vec4(nir_builder
*b
, nir_ssa_def
*value
)
2678 if (value
->num_components
== 4)
2682 for (unsigned i
= 0; i
< 4; i
++)
2683 swiz
[i
] = i
< value
->num_components
? i
: 0;
2684 return nir_swizzle(b
, value
, swiz
, 4);
2688 vtn_handle_image(struct vtn_builder
*b
, SpvOp opcode
,
2689 const uint32_t *w
, unsigned count
)
2691 /* Just get this one out of the way */
2692 if (opcode
== SpvOpImageTexelPointer
) {
2693 struct vtn_value
*val
=
2694 vtn_push_value(b
, w
[2], vtn_value_type_image_pointer
);
2695 val
->image
= ralloc(b
, struct vtn_image_pointer
);
2697 val
->image
->image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2698 val
->image
->coord
= get_image_coord(b
, w
[4]);
2699 val
->image
->sample
= vtn_ssa_value(b
, w
[5])->def
;
2703 struct vtn_image_pointer image
;
2704 SpvScope scope
= SpvScopeInvocation
;
2705 SpvMemorySemanticsMask semantics
= 0;
2708 case SpvOpAtomicExchange
:
2709 case SpvOpAtomicCompareExchange
:
2710 case SpvOpAtomicCompareExchangeWeak
:
2711 case SpvOpAtomicIIncrement
:
2712 case SpvOpAtomicIDecrement
:
2713 case SpvOpAtomicIAdd
:
2714 case SpvOpAtomicISub
:
2715 case SpvOpAtomicLoad
:
2716 case SpvOpAtomicSMin
:
2717 case SpvOpAtomicUMin
:
2718 case SpvOpAtomicSMax
:
2719 case SpvOpAtomicUMax
:
2720 case SpvOpAtomicAnd
:
2722 case SpvOpAtomicXor
:
2723 image
= *vtn_value(b
, w
[3], vtn_value_type_image_pointer
)->image
;
2724 scope
= vtn_constant_uint(b
, w
[4]);
2725 semantics
= vtn_constant_uint(b
, w
[5]);
2728 case SpvOpAtomicStore
:
2729 image
= *vtn_value(b
, w
[1], vtn_value_type_image_pointer
)->image
;
2730 scope
= vtn_constant_uint(b
, w
[2]);
2731 semantics
= vtn_constant_uint(b
, w
[3]);
2734 case SpvOpImageQuerySize
:
2735 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2737 image
.sample
= NULL
;
2740 case SpvOpImageRead
: {
2741 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2742 image
.coord
= get_image_coord(b
, w
[4]);
2744 const SpvImageOperandsMask operands
=
2745 count
> 5 ? w
[5] : SpvImageOperandsMaskNone
;
2747 if (operands
& SpvImageOperandsSampleMask
) {
2748 uint32_t arg
= image_operand_arg(b
, w
, count
, 5,
2749 SpvImageOperandsSampleMask
);
2750 image
.sample
= vtn_ssa_value(b
, w
[arg
])->def
;
2752 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2755 if (operands
& SpvImageOperandsMakeTexelVisibleMask
) {
2756 vtn_fail_if((operands
& SpvImageOperandsNonPrivateTexelMask
) == 0,
2757 "MakeTexelVisible requires NonPrivateTexel to also be set.");
2758 uint32_t arg
= image_operand_arg(b
, w
, count
, 5,
2759 SpvImageOperandsMakeTexelVisibleMask
);
2760 semantics
= SpvMemorySemanticsMakeVisibleMask
;
2761 scope
= vtn_constant_uint(b
, w
[arg
]);
2764 /* TODO: Volatile. */
2769 case SpvOpImageWrite
: {
2770 image
.image
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
2771 image
.coord
= get_image_coord(b
, w
[2]);
2775 const SpvImageOperandsMask operands
=
2776 count
> 4 ? w
[4] : SpvImageOperandsMaskNone
;
2778 if (operands
& SpvImageOperandsSampleMask
) {
2779 uint32_t arg
= image_operand_arg(b
, w
, count
, 4,
2780 SpvImageOperandsSampleMask
);
2781 image
.sample
= vtn_ssa_value(b
, w
[arg
])->def
;
2783 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2786 if (operands
& SpvImageOperandsMakeTexelAvailableMask
) {
2787 vtn_fail_if((operands
& SpvImageOperandsNonPrivateTexelMask
) == 0,
2788 "MakeTexelAvailable requires NonPrivateTexel to also be set.");
2789 uint32_t arg
= image_operand_arg(b
, w
, count
, 4,
2790 SpvImageOperandsMakeTexelAvailableMask
);
2791 semantics
= SpvMemorySemanticsMakeAvailableMask
;
2792 scope
= vtn_constant_uint(b
, w
[arg
]);
2795 /* TODO: Volatile. */
2801 vtn_fail_with_opcode("Invalid image opcode", opcode
);
2804 nir_intrinsic_op op
;
2806 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_deref_##N; break;
2807 OP(ImageQuerySize
, size
)
2809 OP(ImageWrite
, store
)
2810 OP(AtomicLoad
, load
)
2811 OP(AtomicStore
, store
)
2812 OP(AtomicExchange
, atomic_exchange
)
2813 OP(AtomicCompareExchange
, atomic_comp_swap
)
2814 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
2815 OP(AtomicIIncrement
, atomic_add
)
2816 OP(AtomicIDecrement
, atomic_add
)
2817 OP(AtomicIAdd
, atomic_add
)
2818 OP(AtomicISub
, atomic_add
)
2819 OP(AtomicSMin
, atomic_imin
)
2820 OP(AtomicUMin
, atomic_umin
)
2821 OP(AtomicSMax
, atomic_imax
)
2822 OP(AtomicUMax
, atomic_umax
)
2823 OP(AtomicAnd
, atomic_and
)
2824 OP(AtomicOr
, atomic_or
)
2825 OP(AtomicXor
, atomic_xor
)
2828 vtn_fail_with_opcode("Invalid image opcode", opcode
);
2831 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
2833 nir_deref_instr
*image_deref
= vtn_pointer_to_deref(b
, image
.image
);
2834 intrin
->src
[0] = nir_src_for_ssa(&image_deref
->dest
.ssa
);
2836 /* ImageQuerySize doesn't take any extra parameters */
2837 if (opcode
!= SpvOpImageQuerySize
) {
2838 /* The image coordinate is always 4 components but we may not have that
2839 * many. Swizzle to compensate.
2841 intrin
->src
[1] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, image
.coord
));
2842 intrin
->src
[2] = nir_src_for_ssa(image
.sample
);
2845 nir_intrinsic_set_access(intrin
, image
.image
->access
);
2848 case SpvOpAtomicLoad
:
2849 case SpvOpImageQuerySize
:
2850 case SpvOpImageRead
:
2852 case SpvOpAtomicStore
:
2853 case SpvOpImageWrite
: {
2854 const uint32_t value_id
= opcode
== SpvOpAtomicStore
? w
[4] : w
[3];
2855 nir_ssa_def
*value
= vtn_ssa_value(b
, value_id
)->def
;
2856 /* nir_intrinsic_image_deref_store always takes a vec4 value */
2857 assert(op
== nir_intrinsic_image_deref_store
);
2858 intrin
->num_components
= 4;
2859 intrin
->src
[3] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, value
));
2863 case SpvOpAtomicCompareExchange
:
2864 case SpvOpAtomicCompareExchangeWeak
:
2865 case SpvOpAtomicIIncrement
:
2866 case SpvOpAtomicIDecrement
:
2867 case SpvOpAtomicExchange
:
2868 case SpvOpAtomicIAdd
:
2869 case SpvOpAtomicISub
:
2870 case SpvOpAtomicSMin
:
2871 case SpvOpAtomicUMin
:
2872 case SpvOpAtomicSMax
:
2873 case SpvOpAtomicUMax
:
2874 case SpvOpAtomicAnd
:
2876 case SpvOpAtomicXor
:
2877 fill_common_atomic_sources(b
, opcode
, w
, &intrin
->src
[3]);
2881 vtn_fail_with_opcode("Invalid image opcode", opcode
);
2884 /* Image operations implicitly have the Image storage memory semantics. */
2885 semantics
|= SpvMemorySemanticsImageMemoryMask
;
2887 SpvMemorySemanticsMask before_semantics
;
2888 SpvMemorySemanticsMask after_semantics
;
2889 vtn_split_barrier_semantics(b
, semantics
, &before_semantics
, &after_semantics
);
2891 if (before_semantics
)
2892 vtn_emit_memory_barrier(b
, scope
, before_semantics
);
2894 if (opcode
!= SpvOpImageWrite
&& opcode
!= SpvOpAtomicStore
) {
2895 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2897 unsigned dest_components
= glsl_get_vector_elements(type
->type
);
2898 intrin
->num_components
= nir_intrinsic_infos
[op
].dest_components
;
2899 if (intrin
->num_components
== 0)
2900 intrin
->num_components
= dest_components
;
2902 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
,
2903 intrin
->num_components
, 32, NULL
);
2905 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2907 nir_ssa_def
*result
= &intrin
->dest
.ssa
;
2908 if (intrin
->num_components
!= dest_components
)
2909 result
= nir_channels(&b
->nb
, result
, (1 << dest_components
) - 1);
2911 struct vtn_value
*val
=
2912 vtn_push_ssa(b
, w
[2], type
, vtn_create_ssa_value(b
, type
->type
));
2913 val
->ssa
->def
= result
;
2915 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2918 if (after_semantics
)
2919 vtn_emit_memory_barrier(b
, scope
, after_semantics
);
2922 static nir_intrinsic_op
2923 get_ssbo_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2926 case SpvOpAtomicLoad
: return nir_intrinsic_load_ssbo
;
2927 case SpvOpAtomicStore
: return nir_intrinsic_store_ssbo
;
2928 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
2929 OP(AtomicExchange
, atomic_exchange
)
2930 OP(AtomicCompareExchange
, atomic_comp_swap
)
2931 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
2932 OP(AtomicIIncrement
, atomic_add
)
2933 OP(AtomicIDecrement
, atomic_add
)
2934 OP(AtomicIAdd
, atomic_add
)
2935 OP(AtomicISub
, atomic_add
)
2936 OP(AtomicSMin
, atomic_imin
)
2937 OP(AtomicUMin
, atomic_umin
)
2938 OP(AtomicSMax
, atomic_imax
)
2939 OP(AtomicUMax
, atomic_umax
)
2940 OP(AtomicAnd
, atomic_and
)
2941 OP(AtomicOr
, atomic_or
)
2942 OP(AtomicXor
, atomic_xor
)
2945 vtn_fail_with_opcode("Invalid SSBO atomic", opcode
);
2949 static nir_intrinsic_op
2950 get_uniform_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2953 #define OP(S, N) case SpvOp##S: return nir_intrinsic_atomic_counter_ ##N;
2954 OP(AtomicLoad
, read_deref
)
2955 OP(AtomicExchange
, exchange
)
2956 OP(AtomicCompareExchange
, comp_swap
)
2957 OP(AtomicCompareExchangeWeak
, comp_swap
)
2958 OP(AtomicIIncrement
, inc_deref
)
2959 OP(AtomicIDecrement
, post_dec_deref
)
2960 OP(AtomicIAdd
, add_deref
)
2961 OP(AtomicISub
, add_deref
)
2962 OP(AtomicUMin
, min_deref
)
2963 OP(AtomicUMax
, max_deref
)
2964 OP(AtomicAnd
, and_deref
)
2965 OP(AtomicOr
, or_deref
)
2966 OP(AtomicXor
, xor_deref
)
2969 /* We left the following out: AtomicStore, AtomicSMin and
2970 * AtomicSmax. Right now there are not nir intrinsics for them. At this
2971 * moment Atomic Counter support is needed for ARB_spirv support, so is
2972 * only need to support GLSL Atomic Counters that are uints and don't
2973 * allow direct storage.
2975 unreachable("Invalid uniform atomic");
2979 static nir_intrinsic_op
2980 get_deref_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2983 case SpvOpAtomicLoad
: return nir_intrinsic_load_deref
;
2984 case SpvOpAtomicStore
: return nir_intrinsic_store_deref
;
2985 #define OP(S, N) case SpvOp##S: return nir_intrinsic_deref_##N;
2986 OP(AtomicExchange
, atomic_exchange
)
2987 OP(AtomicCompareExchange
, atomic_comp_swap
)
2988 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
2989 OP(AtomicIIncrement
, atomic_add
)
2990 OP(AtomicIDecrement
, atomic_add
)
2991 OP(AtomicIAdd
, atomic_add
)
2992 OP(AtomicISub
, atomic_add
)
2993 OP(AtomicSMin
, atomic_imin
)
2994 OP(AtomicUMin
, atomic_umin
)
2995 OP(AtomicSMax
, atomic_imax
)
2996 OP(AtomicUMax
, atomic_umax
)
2997 OP(AtomicAnd
, atomic_and
)
2998 OP(AtomicOr
, atomic_or
)
2999 OP(AtomicXor
, atomic_xor
)
3002 vtn_fail_with_opcode("Invalid shared atomic", opcode
);
3007 * Handles shared atomics, ssbo atomics and atomic counters.
3010 vtn_handle_atomics(struct vtn_builder
*b
, SpvOp opcode
,
3011 const uint32_t *w
, unsigned count
)
3013 struct vtn_pointer
*ptr
;
3014 nir_intrinsic_instr
*atomic
;
3016 SpvScope scope
= SpvScopeInvocation
;
3017 SpvMemorySemanticsMask semantics
= 0;
3020 case SpvOpAtomicLoad
:
3021 case SpvOpAtomicExchange
:
3022 case SpvOpAtomicCompareExchange
:
3023 case SpvOpAtomicCompareExchangeWeak
:
3024 case SpvOpAtomicIIncrement
:
3025 case SpvOpAtomicIDecrement
:
3026 case SpvOpAtomicIAdd
:
3027 case SpvOpAtomicISub
:
3028 case SpvOpAtomicSMin
:
3029 case SpvOpAtomicUMin
:
3030 case SpvOpAtomicSMax
:
3031 case SpvOpAtomicUMax
:
3032 case SpvOpAtomicAnd
:
3034 case SpvOpAtomicXor
:
3035 ptr
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
3036 scope
= vtn_constant_uint(b
, w
[4]);
3037 semantics
= vtn_constant_uint(b
, w
[5]);
3040 case SpvOpAtomicStore
:
3041 ptr
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
3042 scope
= vtn_constant_uint(b
, w
[2]);
3043 semantics
= vtn_constant_uint(b
, w
[3]);
3047 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
3050 /* uniform as "atomic counter uniform" */
3051 if (ptr
->mode
== vtn_variable_mode_uniform
) {
3052 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
3053 const struct glsl_type
*deref_type
= deref
->type
;
3054 nir_intrinsic_op op
= get_uniform_nir_atomic_op(b
, opcode
);
3055 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
3056 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
3058 /* SSBO needs to initialize index/offset. In this case we don't need to,
3059 * as that info is already stored on the ptr->var->var nir_variable (see
3060 * vtn_create_variable)
3064 case SpvOpAtomicLoad
:
3065 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
3068 case SpvOpAtomicStore
:
3069 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
3070 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
3073 case SpvOpAtomicExchange
:
3074 case SpvOpAtomicCompareExchange
:
3075 case SpvOpAtomicCompareExchangeWeak
:
3076 case SpvOpAtomicIIncrement
:
3077 case SpvOpAtomicIDecrement
:
3078 case SpvOpAtomicIAdd
:
3079 case SpvOpAtomicISub
:
3080 case SpvOpAtomicSMin
:
3081 case SpvOpAtomicUMin
:
3082 case SpvOpAtomicSMax
:
3083 case SpvOpAtomicUMax
:
3084 case SpvOpAtomicAnd
:
3086 case SpvOpAtomicXor
:
3087 /* Nothing: we don't need to call fill_common_atomic_sources here, as
3088 * atomic counter uniforms doesn't have sources
3093 unreachable("Invalid SPIR-V atomic");
3096 } else if (vtn_pointer_uses_ssa_offset(b
, ptr
)) {
3097 nir_ssa_def
*offset
, *index
;
3098 offset
= vtn_pointer_to_offset(b
, ptr
, &index
);
3100 assert(ptr
->mode
== vtn_variable_mode_ssbo
);
3102 nir_intrinsic_op op
= get_ssbo_nir_atomic_op(b
, opcode
);
3103 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
3107 case SpvOpAtomicLoad
:
3108 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
3109 nir_intrinsic_set_align(atomic
, 4, 0);
3110 if (ptr
->mode
== vtn_variable_mode_ssbo
)
3111 atomic
->src
[src
++] = nir_src_for_ssa(index
);
3112 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
3115 case SpvOpAtomicStore
:
3116 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
3117 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
3118 nir_intrinsic_set_align(atomic
, 4, 0);
3119 atomic
->src
[src
++] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
3120 if (ptr
->mode
== vtn_variable_mode_ssbo
)
3121 atomic
->src
[src
++] = nir_src_for_ssa(index
);
3122 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
3125 case SpvOpAtomicExchange
:
3126 case SpvOpAtomicCompareExchange
:
3127 case SpvOpAtomicCompareExchangeWeak
:
3128 case SpvOpAtomicIIncrement
:
3129 case SpvOpAtomicIDecrement
:
3130 case SpvOpAtomicIAdd
:
3131 case SpvOpAtomicISub
:
3132 case SpvOpAtomicSMin
:
3133 case SpvOpAtomicUMin
:
3134 case SpvOpAtomicSMax
:
3135 case SpvOpAtomicUMax
:
3136 case SpvOpAtomicAnd
:
3138 case SpvOpAtomicXor
:
3139 if (ptr
->mode
== vtn_variable_mode_ssbo
)
3140 atomic
->src
[src
++] = nir_src_for_ssa(index
);
3141 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
3142 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[src
]);
3146 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
3149 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
3150 const struct glsl_type
*deref_type
= deref
->type
;
3151 nir_intrinsic_op op
= get_deref_nir_atomic_op(b
, opcode
);
3152 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
3153 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
3156 case SpvOpAtomicLoad
:
3157 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
3160 case SpvOpAtomicStore
:
3161 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
3162 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
3163 atomic
->src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
3166 case SpvOpAtomicExchange
:
3167 case SpvOpAtomicCompareExchange
:
3168 case SpvOpAtomicCompareExchangeWeak
:
3169 case SpvOpAtomicIIncrement
:
3170 case SpvOpAtomicIDecrement
:
3171 case SpvOpAtomicIAdd
:
3172 case SpvOpAtomicISub
:
3173 case SpvOpAtomicSMin
:
3174 case SpvOpAtomicUMin
:
3175 case SpvOpAtomicSMax
:
3176 case SpvOpAtomicUMax
:
3177 case SpvOpAtomicAnd
:
3179 case SpvOpAtomicXor
:
3180 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[1]);
3184 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
3188 /* Atomic ordering operations will implicitly apply to the atomic operation
3189 * storage class, so include that too.
3191 semantics
|= vtn_storage_class_to_memory_semantics(ptr
->ptr_type
->storage_class
);
3193 SpvMemorySemanticsMask before_semantics
;
3194 SpvMemorySemanticsMask after_semantics
;
3195 vtn_split_barrier_semantics(b
, semantics
, &before_semantics
, &after_semantics
);
3197 if (before_semantics
)
3198 vtn_emit_memory_barrier(b
, scope
, before_semantics
);
3200 if (opcode
!= SpvOpAtomicStore
) {
3201 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
3203 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
,
3204 glsl_get_vector_elements(type
->type
),
3205 glsl_get_bit_size(type
->type
), NULL
);
3207 struct vtn_ssa_value
*ssa
= rzalloc(b
, struct vtn_ssa_value
);
3208 ssa
->def
= &atomic
->dest
.ssa
;
3209 ssa
->type
= type
->type
;
3210 vtn_push_ssa(b
, w
[2], type
, ssa
);
3213 nir_builder_instr_insert(&b
->nb
, &atomic
->instr
);
3215 if (after_semantics
)
3216 vtn_emit_memory_barrier(b
, scope
, after_semantics
);
3219 static nir_alu_instr
*
3220 create_vec(struct vtn_builder
*b
, unsigned num_components
, unsigned bit_size
)
3222 nir_op op
= nir_op_vec(num_components
);
3223 nir_alu_instr
*vec
= nir_alu_instr_create(b
->shader
, op
);
3224 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, num_components
,
3226 vec
->dest
.write_mask
= (1 << num_components
) - 1;
3231 struct vtn_ssa_value
*
3232 vtn_ssa_transpose(struct vtn_builder
*b
, struct vtn_ssa_value
*src
)
3234 if (src
->transposed
)
3235 return src
->transposed
;
3237 struct vtn_ssa_value
*dest
=
3238 vtn_create_ssa_value(b
, glsl_transposed_type(src
->type
));
3240 for (unsigned i
= 0; i
< glsl_get_matrix_columns(dest
->type
); i
++) {
3241 nir_alu_instr
*vec
= create_vec(b
, glsl_get_matrix_columns(src
->type
),
3242 glsl_get_bit_size(src
->type
));
3243 if (glsl_type_is_vector_or_scalar(src
->type
)) {
3244 vec
->src
[0].src
= nir_src_for_ssa(src
->def
);
3245 vec
->src
[0].swizzle
[0] = i
;
3247 for (unsigned j
= 0; j
< glsl_get_matrix_columns(src
->type
); j
++) {
3248 vec
->src
[j
].src
= nir_src_for_ssa(src
->elems
[j
]->def
);
3249 vec
->src
[j
].swizzle
[0] = i
;
3252 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3253 dest
->elems
[i
]->def
= &vec
->dest
.dest
.ssa
;
3256 dest
->transposed
= src
;
3262 vtn_vector_extract(struct vtn_builder
*b
, nir_ssa_def
*src
, unsigned index
)
3264 return nir_channel(&b
->nb
, src
, index
);
3268 vtn_vector_insert(struct vtn_builder
*b
, nir_ssa_def
*src
, nir_ssa_def
*insert
,
3271 nir_alu_instr
*vec
= create_vec(b
, src
->num_components
,
3274 for (unsigned i
= 0; i
< src
->num_components
; i
++) {
3276 vec
->src
[i
].src
= nir_src_for_ssa(insert
);
3278 vec
->src
[i
].src
= nir_src_for_ssa(src
);
3279 vec
->src
[i
].swizzle
[0] = i
;
3283 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3285 return &vec
->dest
.dest
.ssa
;
3288 static nir_ssa_def
*
3289 nir_ieq_imm(nir_builder
*b
, nir_ssa_def
*x
, uint64_t i
)
3291 return nir_ieq(b
, x
, nir_imm_intN_t(b
, i
, x
->bit_size
));
3295 vtn_vector_extract_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
3298 return nir_vector_extract(&b
->nb
, src
, nir_i2i(&b
->nb
, index
, 32));
3302 vtn_vector_insert_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
3303 nir_ssa_def
*insert
, nir_ssa_def
*index
)
3305 nir_ssa_def
*dest
= vtn_vector_insert(b
, src
, insert
, 0);
3306 for (unsigned i
= 1; i
< src
->num_components
; i
++)
3307 dest
= nir_bcsel(&b
->nb
, nir_ieq_imm(&b
->nb
, index
, i
),
3308 vtn_vector_insert(b
, src
, insert
, i
), dest
);
3313 static nir_ssa_def
*
3314 vtn_vector_shuffle(struct vtn_builder
*b
, unsigned num_components
,
3315 nir_ssa_def
*src0
, nir_ssa_def
*src1
,
3316 const uint32_t *indices
)
3318 nir_alu_instr
*vec
= create_vec(b
, num_components
, src0
->bit_size
);
3320 for (unsigned i
= 0; i
< num_components
; i
++) {
3321 uint32_t index
= indices
[i
];
3322 if (index
== 0xffffffff) {
3324 nir_src_for_ssa(nir_ssa_undef(&b
->nb
, 1, src0
->bit_size
));
3325 } else if (index
< src0
->num_components
) {
3326 vec
->src
[i
].src
= nir_src_for_ssa(src0
);
3327 vec
->src
[i
].swizzle
[0] = index
;
3329 vec
->src
[i
].src
= nir_src_for_ssa(src1
);
3330 vec
->src
[i
].swizzle
[0] = index
- src0
->num_components
;
3334 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3336 return &vec
->dest
.dest
.ssa
;
3340 * Concatentates a number of vectors/scalars together to produce a vector
3342 static nir_ssa_def
*
3343 vtn_vector_construct(struct vtn_builder
*b
, unsigned num_components
,
3344 unsigned num_srcs
, nir_ssa_def
**srcs
)
3346 nir_alu_instr
*vec
= create_vec(b
, num_components
, srcs
[0]->bit_size
);
3348 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
3350 * "When constructing a vector, there must be at least two Constituent
3353 vtn_assert(num_srcs
>= 2);
3355 unsigned dest_idx
= 0;
3356 for (unsigned i
= 0; i
< num_srcs
; i
++) {
3357 nir_ssa_def
*src
= srcs
[i
];
3358 vtn_assert(dest_idx
+ src
->num_components
<= num_components
);
3359 for (unsigned j
= 0; j
< src
->num_components
; j
++) {
3360 vec
->src
[dest_idx
].src
= nir_src_for_ssa(src
);
3361 vec
->src
[dest_idx
].swizzle
[0] = j
;
3366 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
3368 * "When constructing a vector, the total number of components in all
3369 * the operands must equal the number of components in Result Type."
3371 vtn_assert(dest_idx
== num_components
);
3373 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3375 return &vec
->dest
.dest
.ssa
;
3378 static struct vtn_ssa_value
*
3379 vtn_composite_copy(void *mem_ctx
, struct vtn_ssa_value
*src
)
3381 struct vtn_ssa_value
*dest
= rzalloc(mem_ctx
, struct vtn_ssa_value
);
3382 dest
->type
= src
->type
;
3384 if (glsl_type_is_vector_or_scalar(src
->type
)) {
3385 dest
->def
= src
->def
;
3387 unsigned elems
= glsl_get_length(src
->type
);
3389 dest
->elems
= ralloc_array(mem_ctx
, struct vtn_ssa_value
*, elems
);
3390 for (unsigned i
= 0; i
< elems
; i
++)
3391 dest
->elems
[i
] = vtn_composite_copy(mem_ctx
, src
->elems
[i
]);
3397 static struct vtn_ssa_value
*
3398 vtn_composite_insert(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
3399 struct vtn_ssa_value
*insert
, const uint32_t *indices
,
3400 unsigned num_indices
)
3402 struct vtn_ssa_value
*dest
= vtn_composite_copy(b
, src
);
3404 struct vtn_ssa_value
*cur
= dest
;
3406 for (i
= 0; i
< num_indices
- 1; i
++) {
3407 cur
= cur
->elems
[indices
[i
]];
3410 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
3411 /* According to the SPIR-V spec, OpCompositeInsert may work down to
3412 * the component granularity. In that case, the last index will be
3413 * the index to insert the scalar into the vector.
3416 cur
->def
= vtn_vector_insert(b
, cur
->def
, insert
->def
, indices
[i
]);
3418 cur
->elems
[indices
[i
]] = insert
;
3424 static struct vtn_ssa_value
*
3425 vtn_composite_extract(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
3426 const uint32_t *indices
, unsigned num_indices
)
3428 struct vtn_ssa_value
*cur
= src
;
3429 for (unsigned i
= 0; i
< num_indices
; i
++) {
3430 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
3431 vtn_assert(i
== num_indices
- 1);
3432 /* According to the SPIR-V spec, OpCompositeExtract may work down to
3433 * the component granularity. The last index will be the index of the
3434 * vector to extract.
3437 struct vtn_ssa_value
*ret
= rzalloc(b
, struct vtn_ssa_value
);
3438 ret
->type
= glsl_scalar_type(glsl_get_base_type(cur
->type
));
3439 ret
->def
= vtn_vector_extract(b
, cur
->def
, indices
[i
]);
3442 cur
= cur
->elems
[indices
[i
]];
3450 vtn_handle_composite(struct vtn_builder
*b
, SpvOp opcode
,
3451 const uint32_t *w
, unsigned count
)
3453 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
3454 struct vtn_ssa_value
*ssa
= vtn_create_ssa_value(b
, type
->type
);
3457 case SpvOpVectorExtractDynamic
:
3458 ssa
->def
= vtn_vector_extract_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
3459 vtn_ssa_value(b
, w
[4])->def
);
3462 case SpvOpVectorInsertDynamic
:
3463 ssa
->def
= vtn_vector_insert_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
3464 vtn_ssa_value(b
, w
[4])->def
,
3465 vtn_ssa_value(b
, w
[5])->def
);
3468 case SpvOpVectorShuffle
:
3469 ssa
->def
= vtn_vector_shuffle(b
, glsl_get_vector_elements(type
->type
),
3470 vtn_ssa_value(b
, w
[3])->def
,
3471 vtn_ssa_value(b
, w
[4])->def
,
3475 case SpvOpCompositeConstruct
: {
3476 unsigned elems
= count
- 3;
3478 if (glsl_type_is_vector_or_scalar(type
->type
)) {
3479 nir_ssa_def
*srcs
[NIR_MAX_VEC_COMPONENTS
];
3480 for (unsigned i
= 0; i
< elems
; i
++)
3481 srcs
[i
] = vtn_ssa_value(b
, w
[3 + i
])->def
;
3483 vtn_vector_construct(b
, glsl_get_vector_elements(type
->type
),
3486 ssa
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
3487 for (unsigned i
= 0; i
< elems
; i
++)
3488 ssa
->elems
[i
] = vtn_ssa_value(b
, w
[3 + i
]);
3492 case SpvOpCompositeExtract
:
3493 ssa
= vtn_composite_extract(b
, vtn_ssa_value(b
, w
[3]),
3497 case SpvOpCompositeInsert
:
3498 ssa
= vtn_composite_insert(b
, vtn_ssa_value(b
, w
[4]),
3499 vtn_ssa_value(b
, w
[3]),
3503 case SpvOpCopyLogical
:
3504 case SpvOpCopyObject
:
3505 ssa
= vtn_composite_copy(b
, vtn_ssa_value(b
, w
[3]));
3509 vtn_fail_with_opcode("unknown composite operation", opcode
);
3512 vtn_push_ssa(b
, w
[2], type
, ssa
);
3516 vtn_emit_barrier(struct vtn_builder
*b
, nir_intrinsic_op op
)
3518 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
3519 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3523 vtn_emit_memory_barrier(struct vtn_builder
*b
, SpvScope scope
,
3524 SpvMemorySemanticsMask semantics
)
3526 if (b
->options
->use_scoped_memory_barrier
) {
3527 vtn_emit_scoped_memory_barrier(b
, scope
, semantics
);
3531 static const SpvMemorySemanticsMask all_memory_semantics
=
3532 SpvMemorySemanticsUniformMemoryMask
|
3533 SpvMemorySemanticsWorkgroupMemoryMask
|
3534 SpvMemorySemanticsAtomicCounterMemoryMask
|
3535 SpvMemorySemanticsImageMemoryMask
;
3537 /* If we're not actually doing a memory barrier, bail */
3538 if (!(semantics
& all_memory_semantics
))
3541 /* GL and Vulkan don't have these */
3542 vtn_assert(scope
!= SpvScopeCrossDevice
);
3544 if (scope
== SpvScopeSubgroup
)
3545 return; /* Nothing to do here */
3547 if (scope
== SpvScopeWorkgroup
) {
3548 vtn_emit_barrier(b
, nir_intrinsic_group_memory_barrier
);
3552 /* There's only two scopes thing left */
3553 vtn_assert(scope
== SpvScopeInvocation
|| scope
== SpvScopeDevice
);
3555 if ((semantics
& all_memory_semantics
) == all_memory_semantics
) {
3556 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier
);
3560 /* Issue a bunch of more specific barriers */
3561 uint32_t bits
= semantics
;
3563 SpvMemorySemanticsMask semantic
= 1 << u_bit_scan(&bits
);
3565 case SpvMemorySemanticsUniformMemoryMask
:
3566 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_buffer
);
3568 case SpvMemorySemanticsWorkgroupMemoryMask
:
3569 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_shared
);
3571 case SpvMemorySemanticsAtomicCounterMemoryMask
:
3572 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_atomic_counter
);
3574 case SpvMemorySemanticsImageMemoryMask
:
3575 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_image
);
3584 vtn_handle_barrier(struct vtn_builder
*b
, SpvOp opcode
,
3585 const uint32_t *w
, unsigned count
)
3588 case SpvOpEmitVertex
:
3589 case SpvOpEmitStreamVertex
:
3590 case SpvOpEndPrimitive
:
3591 case SpvOpEndStreamPrimitive
: {
3592 nir_intrinsic_op intrinsic_op
;
3594 case SpvOpEmitVertex
:
3595 case SpvOpEmitStreamVertex
:
3596 intrinsic_op
= nir_intrinsic_emit_vertex
;
3598 case SpvOpEndPrimitive
:
3599 case SpvOpEndStreamPrimitive
:
3600 intrinsic_op
= nir_intrinsic_end_primitive
;
3603 unreachable("Invalid opcode");
3606 nir_intrinsic_instr
*intrin
=
3607 nir_intrinsic_instr_create(b
->shader
, intrinsic_op
);
3610 case SpvOpEmitStreamVertex
:
3611 case SpvOpEndStreamPrimitive
: {
3612 unsigned stream
= vtn_constant_uint(b
, w
[1]);
3613 nir_intrinsic_set_stream_id(intrin
, stream
);
3621 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3625 case SpvOpMemoryBarrier
: {
3626 SpvScope scope
= vtn_constant_uint(b
, w
[1]);
3627 SpvMemorySemanticsMask semantics
= vtn_constant_uint(b
, w
[2]);
3628 vtn_emit_memory_barrier(b
, scope
, semantics
);
3632 case SpvOpControlBarrier
: {
3633 SpvScope memory_scope
= vtn_constant_uint(b
, w
[2]);
3634 SpvMemorySemanticsMask memory_semantics
= vtn_constant_uint(b
, w
[3]);
3635 vtn_emit_memory_barrier(b
, memory_scope
, memory_semantics
);
3637 SpvScope execution_scope
= vtn_constant_uint(b
, w
[1]);
3638 if (execution_scope
== SpvScopeWorkgroup
)
3639 vtn_emit_barrier(b
, nir_intrinsic_barrier
);
3644 unreachable("unknown barrier instruction");
3649 gl_primitive_from_spv_execution_mode(struct vtn_builder
*b
,
3650 SpvExecutionMode mode
)
3653 case SpvExecutionModeInputPoints
:
3654 case SpvExecutionModeOutputPoints
:
3655 return 0; /* GL_POINTS */
3656 case SpvExecutionModeInputLines
:
3657 return 1; /* GL_LINES */
3658 case SpvExecutionModeInputLinesAdjacency
:
3659 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
3660 case SpvExecutionModeTriangles
:
3661 return 4; /* GL_TRIANGLES */
3662 case SpvExecutionModeInputTrianglesAdjacency
:
3663 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
3664 case SpvExecutionModeQuads
:
3665 return 7; /* GL_QUADS */
3666 case SpvExecutionModeIsolines
:
3667 return 0x8E7A; /* GL_ISOLINES */
3668 case SpvExecutionModeOutputLineStrip
:
3669 return 3; /* GL_LINE_STRIP */
3670 case SpvExecutionModeOutputTriangleStrip
:
3671 return 5; /* GL_TRIANGLE_STRIP */
3673 vtn_fail("Invalid primitive type: %s (%u)",
3674 spirv_executionmode_to_string(mode
), mode
);
3679 vertices_in_from_spv_execution_mode(struct vtn_builder
*b
,
3680 SpvExecutionMode mode
)
3683 case SpvExecutionModeInputPoints
:
3685 case SpvExecutionModeInputLines
:
3687 case SpvExecutionModeInputLinesAdjacency
:
3689 case SpvExecutionModeTriangles
:
3691 case SpvExecutionModeInputTrianglesAdjacency
:
3694 vtn_fail("Invalid GS input mode: %s (%u)",
3695 spirv_executionmode_to_string(mode
), mode
);
3699 static gl_shader_stage
3700 stage_for_execution_model(struct vtn_builder
*b
, SpvExecutionModel model
)
3703 case SpvExecutionModelVertex
:
3704 return MESA_SHADER_VERTEX
;
3705 case SpvExecutionModelTessellationControl
:
3706 return MESA_SHADER_TESS_CTRL
;
3707 case SpvExecutionModelTessellationEvaluation
:
3708 return MESA_SHADER_TESS_EVAL
;
3709 case SpvExecutionModelGeometry
:
3710 return MESA_SHADER_GEOMETRY
;
3711 case SpvExecutionModelFragment
:
3712 return MESA_SHADER_FRAGMENT
;
3713 case SpvExecutionModelGLCompute
:
3714 return MESA_SHADER_COMPUTE
;
3715 case SpvExecutionModelKernel
:
3716 return MESA_SHADER_KERNEL
;
3718 vtn_fail("Unsupported execution model: %s (%u)",
3719 spirv_executionmodel_to_string(model
), model
);
3723 #define spv_check_supported(name, cap) do { \
3724 if (!(b->options && b->options->caps.name)) \
3725 vtn_warn("Unsupported SPIR-V capability: %s (%u)", \
3726 spirv_capability_to_string(cap), cap); \
3731 vtn_handle_entry_point(struct vtn_builder
*b
, const uint32_t *w
,
3734 struct vtn_value
*entry_point
= &b
->values
[w
[2]];
3735 /* Let this be a name label regardless */
3736 unsigned name_words
;
3737 entry_point
->name
= vtn_string_literal(b
, &w
[3], count
- 3, &name_words
);
3739 if (strcmp(entry_point
->name
, b
->entry_point_name
) != 0 ||
3740 stage_for_execution_model(b
, w
[1]) != b
->entry_point_stage
)
3743 vtn_assert(b
->entry_point
== NULL
);
3744 b
->entry_point
= entry_point
;
3748 vtn_handle_preamble_instruction(struct vtn_builder
*b
, SpvOp opcode
,
3749 const uint32_t *w
, unsigned count
)
3756 case SpvSourceLanguageUnknown
: lang
= "unknown"; break;
3757 case SpvSourceLanguageESSL
: lang
= "ESSL"; break;
3758 case SpvSourceLanguageGLSL
: lang
= "GLSL"; break;
3759 case SpvSourceLanguageOpenCL_C
: lang
= "OpenCL C"; break;
3760 case SpvSourceLanguageOpenCL_CPP
: lang
= "OpenCL C++"; break;
3761 case SpvSourceLanguageHLSL
: lang
= "HLSL"; break;
3764 uint32_t version
= w
[2];
3767 (count
> 3) ? vtn_value(b
, w
[3], vtn_value_type_string
)->str
: "";
3769 vtn_info("Parsing SPIR-V from %s %u source file %s", lang
, version
, file
);
3773 case SpvOpSourceExtension
:
3774 case SpvOpSourceContinued
:
3775 case SpvOpExtension
:
3776 case SpvOpModuleProcessed
:
3777 /* Unhandled, but these are for debug so that's ok. */
3780 case SpvOpCapability
: {
3781 SpvCapability cap
= w
[1];
3783 case SpvCapabilityMatrix
:
3784 case SpvCapabilityShader
:
3785 case SpvCapabilityGeometry
:
3786 case SpvCapabilityGeometryPointSize
:
3787 case SpvCapabilityUniformBufferArrayDynamicIndexing
:
3788 case SpvCapabilitySampledImageArrayDynamicIndexing
:
3789 case SpvCapabilityStorageBufferArrayDynamicIndexing
:
3790 case SpvCapabilityStorageImageArrayDynamicIndexing
:
3791 case SpvCapabilityImageRect
:
3792 case SpvCapabilitySampledRect
:
3793 case SpvCapabilitySampled1D
:
3794 case SpvCapabilityImage1D
:
3795 case SpvCapabilitySampledCubeArray
:
3796 case SpvCapabilityImageCubeArray
:
3797 case SpvCapabilitySampledBuffer
:
3798 case SpvCapabilityImageBuffer
:
3799 case SpvCapabilityImageQuery
:
3800 case SpvCapabilityDerivativeControl
:
3801 case SpvCapabilityInterpolationFunction
:
3802 case SpvCapabilityMultiViewport
:
3803 case SpvCapabilitySampleRateShading
:
3804 case SpvCapabilityClipDistance
:
3805 case SpvCapabilityCullDistance
:
3806 case SpvCapabilityInputAttachment
:
3807 case SpvCapabilityImageGatherExtended
:
3808 case SpvCapabilityStorageImageExtendedFormats
:
3811 case SpvCapabilityLinkage
:
3812 case SpvCapabilityVector16
:
3813 case SpvCapabilityFloat16Buffer
:
3814 case SpvCapabilitySparseResidency
:
3815 vtn_warn("Unsupported SPIR-V capability: %s",
3816 spirv_capability_to_string(cap
));
3819 case SpvCapabilityMinLod
:
3820 spv_check_supported(min_lod
, cap
);
3823 case SpvCapabilityAtomicStorage
:
3824 spv_check_supported(atomic_storage
, cap
);
3827 case SpvCapabilityFloat64
:
3828 spv_check_supported(float64
, cap
);
3830 case SpvCapabilityInt64
:
3831 spv_check_supported(int64
, cap
);
3833 case SpvCapabilityInt16
:
3834 spv_check_supported(int16
, cap
);
3836 case SpvCapabilityInt8
:
3837 spv_check_supported(int8
, cap
);
3840 case SpvCapabilityTransformFeedback
:
3841 spv_check_supported(transform_feedback
, cap
);
3844 case SpvCapabilityGeometryStreams
:
3845 spv_check_supported(geometry_streams
, cap
);
3848 case SpvCapabilityInt64Atomics
:
3849 spv_check_supported(int64_atomics
, cap
);
3852 case SpvCapabilityStorageImageMultisample
:
3853 spv_check_supported(storage_image_ms
, cap
);
3856 case SpvCapabilityAddresses
:
3857 spv_check_supported(address
, cap
);
3860 case SpvCapabilityKernel
:
3861 spv_check_supported(kernel
, cap
);
3864 case SpvCapabilityImageBasic
:
3865 case SpvCapabilityImageReadWrite
:
3866 case SpvCapabilityImageMipmap
:
3867 case SpvCapabilityPipes
:
3868 case SpvCapabilityDeviceEnqueue
:
3869 case SpvCapabilityLiteralSampler
:
3870 case SpvCapabilityGenericPointer
:
3871 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
3872 spirv_capability_to_string(cap
));
3875 case SpvCapabilityImageMSArray
:
3876 spv_check_supported(image_ms_array
, cap
);
3879 case SpvCapabilityTessellation
:
3880 case SpvCapabilityTessellationPointSize
:
3881 spv_check_supported(tessellation
, cap
);
3884 case SpvCapabilityDrawParameters
:
3885 spv_check_supported(draw_parameters
, cap
);
3888 case SpvCapabilityStorageImageReadWithoutFormat
:
3889 spv_check_supported(image_read_without_format
, cap
);
3892 case SpvCapabilityStorageImageWriteWithoutFormat
:
3893 spv_check_supported(image_write_without_format
, cap
);
3896 case SpvCapabilityDeviceGroup
:
3897 spv_check_supported(device_group
, cap
);
3900 case SpvCapabilityMultiView
:
3901 spv_check_supported(multiview
, cap
);
3904 case SpvCapabilityGroupNonUniform
:
3905 spv_check_supported(subgroup_basic
, cap
);
3908 case SpvCapabilitySubgroupVoteKHR
:
3909 case SpvCapabilityGroupNonUniformVote
:
3910 spv_check_supported(subgroup_vote
, cap
);
3913 case SpvCapabilitySubgroupBallotKHR
:
3914 case SpvCapabilityGroupNonUniformBallot
:
3915 spv_check_supported(subgroup_ballot
, cap
);
3918 case SpvCapabilityGroupNonUniformShuffle
:
3919 case SpvCapabilityGroupNonUniformShuffleRelative
:
3920 spv_check_supported(subgroup_shuffle
, cap
);
3923 case SpvCapabilityGroupNonUniformQuad
:
3924 spv_check_supported(subgroup_quad
, cap
);
3927 case SpvCapabilityGroupNonUniformArithmetic
:
3928 case SpvCapabilityGroupNonUniformClustered
:
3929 spv_check_supported(subgroup_arithmetic
, cap
);
3932 case SpvCapabilityGroups
:
3933 spv_check_supported(amd_shader_ballot
, cap
);
3936 case SpvCapabilityVariablePointersStorageBuffer
:
3937 case SpvCapabilityVariablePointers
:
3938 spv_check_supported(variable_pointers
, cap
);
3939 b
->variable_pointers
= true;
3942 case SpvCapabilityStorageUniformBufferBlock16
:
3943 case SpvCapabilityStorageUniform16
:
3944 case SpvCapabilityStoragePushConstant16
:
3945 case SpvCapabilityStorageInputOutput16
:
3946 spv_check_supported(storage_16bit
, cap
);
3949 case SpvCapabilityShaderLayer
:
3950 case SpvCapabilityShaderViewportIndex
:
3951 case SpvCapabilityShaderViewportIndexLayerEXT
:
3952 spv_check_supported(shader_viewport_index_layer
, cap
);
3955 case SpvCapabilityStorageBuffer8BitAccess
:
3956 case SpvCapabilityUniformAndStorageBuffer8BitAccess
:
3957 case SpvCapabilityStoragePushConstant8
:
3958 spv_check_supported(storage_8bit
, cap
);
3961 case SpvCapabilityShaderNonUniformEXT
:
3962 spv_check_supported(descriptor_indexing
, cap
);
3965 case SpvCapabilityInputAttachmentArrayDynamicIndexingEXT
:
3966 case SpvCapabilityUniformTexelBufferArrayDynamicIndexingEXT
:
3967 case SpvCapabilityStorageTexelBufferArrayDynamicIndexingEXT
:
3968 spv_check_supported(descriptor_array_dynamic_indexing
, cap
);
3971 case SpvCapabilityUniformBufferArrayNonUniformIndexingEXT
:
3972 case SpvCapabilitySampledImageArrayNonUniformIndexingEXT
:
3973 case SpvCapabilityStorageBufferArrayNonUniformIndexingEXT
:
3974 case SpvCapabilityStorageImageArrayNonUniformIndexingEXT
:
3975 case SpvCapabilityInputAttachmentArrayNonUniformIndexingEXT
:
3976 case SpvCapabilityUniformTexelBufferArrayNonUniformIndexingEXT
:
3977 case SpvCapabilityStorageTexelBufferArrayNonUniformIndexingEXT
:
3978 spv_check_supported(descriptor_array_non_uniform_indexing
, cap
);
3981 case SpvCapabilityRuntimeDescriptorArrayEXT
:
3982 spv_check_supported(runtime_descriptor_array
, cap
);
3985 case SpvCapabilityStencilExportEXT
:
3986 spv_check_supported(stencil_export
, cap
);
3989 case SpvCapabilitySampleMaskPostDepthCoverage
:
3990 spv_check_supported(post_depth_coverage
, cap
);
3993 case SpvCapabilityDenormFlushToZero
:
3994 case SpvCapabilityDenormPreserve
:
3995 case SpvCapabilitySignedZeroInfNanPreserve
:
3996 case SpvCapabilityRoundingModeRTE
:
3997 case SpvCapabilityRoundingModeRTZ
:
3998 spv_check_supported(float_controls
, cap
);
4001 case SpvCapabilityPhysicalStorageBufferAddressesEXT
:
4002 spv_check_supported(physical_storage_buffer_address
, cap
);
4005 case SpvCapabilityComputeDerivativeGroupQuadsNV
:
4006 case SpvCapabilityComputeDerivativeGroupLinearNV
:
4007 spv_check_supported(derivative_group
, cap
);
4010 case SpvCapabilityFloat16
:
4011 spv_check_supported(float16
, cap
);
4014 case SpvCapabilityFragmentShaderSampleInterlockEXT
:
4015 spv_check_supported(fragment_shader_sample_interlock
, cap
);
4018 case SpvCapabilityFragmentShaderPixelInterlockEXT
:
4019 spv_check_supported(fragment_shader_pixel_interlock
, cap
);
4022 case SpvCapabilityDemoteToHelperInvocationEXT
:
4023 spv_check_supported(demote_to_helper_invocation
, cap
);
4026 case SpvCapabilityShaderClockKHR
:
4027 spv_check_supported(shader_clock
, cap
);
4030 case SpvCapabilityVulkanMemoryModel
:
4031 spv_check_supported(vk_memory_model
, cap
);
4034 case SpvCapabilityVulkanMemoryModelDeviceScope
:
4035 spv_check_supported(vk_memory_model_device_scope
, cap
);
4039 vtn_fail("Unhandled capability: %s (%u)",
4040 spirv_capability_to_string(cap
), cap
);
4045 case SpvOpExtInstImport
:
4046 vtn_handle_extension(b
, opcode
, w
, count
);
4049 case SpvOpMemoryModel
:
4051 case SpvAddressingModelPhysical32
:
4052 vtn_fail_if(b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
,
4053 "AddressingModelPhysical32 only supported for kernels");
4054 b
->shader
->info
.cs
.ptr_size
= 32;
4055 b
->physical_ptrs
= true;
4056 b
->options
->shared_addr_format
= nir_address_format_32bit_global
;
4057 b
->options
->global_addr_format
= nir_address_format_32bit_global
;
4058 b
->options
->temp_addr_format
= nir_address_format_32bit_global
;
4060 case SpvAddressingModelPhysical64
:
4061 vtn_fail_if(b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
,
4062 "AddressingModelPhysical64 only supported for kernels");
4063 b
->shader
->info
.cs
.ptr_size
= 64;
4064 b
->physical_ptrs
= true;
4065 b
->options
->shared_addr_format
= nir_address_format_64bit_global
;
4066 b
->options
->global_addr_format
= nir_address_format_64bit_global
;
4067 b
->options
->temp_addr_format
= nir_address_format_64bit_global
;
4069 case SpvAddressingModelLogical
:
4070 vtn_fail_if(b
->shader
->info
.stage
>= MESA_SHADER_STAGES
,
4071 "AddressingModelLogical only supported for shaders");
4072 b
->shader
->info
.cs
.ptr_size
= 0;
4073 b
->physical_ptrs
= false;
4075 case SpvAddressingModelPhysicalStorageBuffer64EXT
:
4076 vtn_fail_if(!b
->options
||
4077 !b
->options
->caps
.physical_storage_buffer_address
,
4078 "AddressingModelPhysicalStorageBuffer64EXT not supported");
4081 vtn_fail("Unknown addressing model: %s (%u)",
4082 spirv_addressingmodel_to_string(w
[1]), w
[1]);
4087 case SpvMemoryModelSimple
:
4088 case SpvMemoryModelGLSL450
:
4089 case SpvMemoryModelOpenCL
:
4091 case SpvMemoryModelVulkan
:
4092 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
4093 "Vulkan memory model is unsupported by this driver");
4096 vtn_fail("Unsupported memory model: %s",
4097 spirv_memorymodel_to_string(w
[2]));
4102 case SpvOpEntryPoint
:
4103 vtn_handle_entry_point(b
, w
, count
);
4107 vtn_push_value(b
, w
[1], vtn_value_type_string
)->str
=
4108 vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
4112 b
->values
[w
[1]].name
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
4115 case SpvOpMemberName
:
4119 case SpvOpExecutionMode
:
4120 case SpvOpExecutionModeId
:
4121 case SpvOpDecorationGroup
:
4123 case SpvOpDecorateId
:
4124 case SpvOpMemberDecorate
:
4125 case SpvOpGroupDecorate
:
4126 case SpvOpGroupMemberDecorate
:
4127 case SpvOpDecorateString
:
4128 case SpvOpMemberDecorateString
:
4129 vtn_handle_decoration(b
, opcode
, w
, count
);
4133 return false; /* End of preamble */
4140 vtn_handle_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
4141 const struct vtn_decoration
*mode
, void *data
)
4143 vtn_assert(b
->entry_point
== entry_point
);
4145 switch(mode
->exec_mode
) {
4146 case SpvExecutionModeOriginUpperLeft
:
4147 case SpvExecutionModeOriginLowerLeft
:
4148 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4149 b
->shader
->info
.fs
.origin_upper_left
=
4150 (mode
->exec_mode
== SpvExecutionModeOriginUpperLeft
);
4153 case SpvExecutionModeEarlyFragmentTests
:
4154 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4155 b
->shader
->info
.fs
.early_fragment_tests
= true;
4158 case SpvExecutionModePostDepthCoverage
:
4159 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4160 b
->shader
->info
.fs
.post_depth_coverage
= true;
4163 case SpvExecutionModeInvocations
:
4164 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4165 b
->shader
->info
.gs
.invocations
= MAX2(1, mode
->operands
[0]);
4168 case SpvExecutionModeDepthReplacing
:
4169 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4170 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_ANY
;
4172 case SpvExecutionModeDepthGreater
:
4173 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4174 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_GREATER
;
4176 case SpvExecutionModeDepthLess
:
4177 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4178 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_LESS
;
4180 case SpvExecutionModeDepthUnchanged
:
4181 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4182 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
4185 case SpvExecutionModeLocalSize
:
4186 vtn_assert(gl_shader_stage_is_compute(b
->shader
->info
.stage
));
4187 b
->shader
->info
.cs
.local_size
[0] = mode
->operands
[0];
4188 b
->shader
->info
.cs
.local_size
[1] = mode
->operands
[1];
4189 b
->shader
->info
.cs
.local_size
[2] = mode
->operands
[2];
4192 case SpvExecutionModeLocalSizeId
:
4193 b
->shader
->info
.cs
.local_size
[0] = vtn_constant_uint(b
, mode
->operands
[0]);
4194 b
->shader
->info
.cs
.local_size
[1] = vtn_constant_uint(b
, mode
->operands
[1]);
4195 b
->shader
->info
.cs
.local_size
[2] = vtn_constant_uint(b
, mode
->operands
[2]);
4198 case SpvExecutionModeLocalSizeHint
:
4199 case SpvExecutionModeLocalSizeHintId
:
4200 break; /* Nothing to do with this */
4202 case SpvExecutionModeOutputVertices
:
4203 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4204 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
4205 b
->shader
->info
.tess
.tcs_vertices_out
= mode
->operands
[0];
4207 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4208 b
->shader
->info
.gs
.vertices_out
= mode
->operands
[0];
4212 case SpvExecutionModeInputPoints
:
4213 case SpvExecutionModeInputLines
:
4214 case SpvExecutionModeInputLinesAdjacency
:
4215 case SpvExecutionModeTriangles
:
4216 case SpvExecutionModeInputTrianglesAdjacency
:
4217 case SpvExecutionModeQuads
:
4218 case SpvExecutionModeIsolines
:
4219 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4220 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
4221 b
->shader
->info
.tess
.primitive_mode
=
4222 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
4224 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4225 b
->shader
->info
.gs
.vertices_in
=
4226 vertices_in_from_spv_execution_mode(b
, mode
->exec_mode
);
4227 b
->shader
->info
.gs
.input_primitive
=
4228 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
4232 case SpvExecutionModeOutputPoints
:
4233 case SpvExecutionModeOutputLineStrip
:
4234 case SpvExecutionModeOutputTriangleStrip
:
4235 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4236 b
->shader
->info
.gs
.output_primitive
=
4237 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
4240 case SpvExecutionModeSpacingEqual
:
4241 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4242 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4243 b
->shader
->info
.tess
.spacing
= TESS_SPACING_EQUAL
;
4245 case SpvExecutionModeSpacingFractionalEven
:
4246 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4247 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4248 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_EVEN
;
4250 case SpvExecutionModeSpacingFractionalOdd
:
4251 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4252 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4253 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_ODD
;
4255 case SpvExecutionModeVertexOrderCw
:
4256 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4257 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4258 b
->shader
->info
.tess
.ccw
= false;
4260 case SpvExecutionModeVertexOrderCcw
:
4261 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4262 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4263 b
->shader
->info
.tess
.ccw
= true;
4265 case SpvExecutionModePointMode
:
4266 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4267 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4268 b
->shader
->info
.tess
.point_mode
= true;
4271 case SpvExecutionModePixelCenterInteger
:
4272 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4273 b
->shader
->info
.fs
.pixel_center_integer
= true;
4276 case SpvExecutionModeXfb
:
4277 b
->shader
->info
.has_transform_feedback_varyings
= true;
4280 case SpvExecutionModeVecTypeHint
:
4283 case SpvExecutionModeContractionOff
:
4284 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
)
4285 vtn_warn("ExectionMode only allowed for CL-style kernels: %s",
4286 spirv_executionmode_to_string(mode
->exec_mode
));
4291 case SpvExecutionModeStencilRefReplacingEXT
:
4292 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4295 case SpvExecutionModeDerivativeGroupQuadsNV
:
4296 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
4297 b
->shader
->info
.cs
.derivative_group
= DERIVATIVE_GROUP_QUADS
;
4300 case SpvExecutionModeDerivativeGroupLinearNV
:
4301 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
4302 b
->shader
->info
.cs
.derivative_group
= DERIVATIVE_GROUP_LINEAR
;
4305 case SpvExecutionModePixelInterlockOrderedEXT
:
4306 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4307 b
->shader
->info
.fs
.pixel_interlock_ordered
= true;
4310 case SpvExecutionModePixelInterlockUnorderedEXT
:
4311 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4312 b
->shader
->info
.fs
.pixel_interlock_unordered
= true;
4315 case SpvExecutionModeSampleInterlockOrderedEXT
:
4316 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4317 b
->shader
->info
.fs
.sample_interlock_ordered
= true;
4320 case SpvExecutionModeSampleInterlockUnorderedEXT
:
4321 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4322 b
->shader
->info
.fs
.sample_interlock_unordered
= true;
4325 case SpvExecutionModeDenormPreserve
:
4326 case SpvExecutionModeDenormFlushToZero
:
4327 case SpvExecutionModeSignedZeroInfNanPreserve
:
4328 case SpvExecutionModeRoundingModeRTE
:
4329 case SpvExecutionModeRoundingModeRTZ
:
4330 /* Already handled in vtn_handle_rounding_mode_in_execution_mode() */
4334 vtn_fail("Unhandled execution mode: %s (%u)",
4335 spirv_executionmode_to_string(mode
->exec_mode
),
4341 vtn_handle_rounding_mode_in_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
4342 const struct vtn_decoration
*mode
, void *data
)
4344 vtn_assert(b
->entry_point
== entry_point
);
4346 unsigned execution_mode
= 0;
4348 switch(mode
->exec_mode
) {
4349 case SpvExecutionModeDenormPreserve
:
4350 switch (mode
->operands
[0]) {
4351 case 16: execution_mode
= FLOAT_CONTROLS_DENORM_PRESERVE_FP16
; break;
4352 case 32: execution_mode
= FLOAT_CONTROLS_DENORM_PRESERVE_FP32
; break;
4353 case 64: execution_mode
= FLOAT_CONTROLS_DENORM_PRESERVE_FP64
; break;
4354 default: vtn_fail("Floating point type not supported");
4357 case SpvExecutionModeDenormFlushToZero
:
4358 switch (mode
->operands
[0]) {
4359 case 16: execution_mode
= FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16
; break;
4360 case 32: execution_mode
= FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32
; break;
4361 case 64: execution_mode
= FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP64
; break;
4362 default: vtn_fail("Floating point type not supported");
4365 case SpvExecutionModeSignedZeroInfNanPreserve
:
4366 switch (mode
->operands
[0]) {
4367 case 16: execution_mode
= FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP16
; break;
4368 case 32: execution_mode
= FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP32
; break;
4369 case 64: execution_mode
= FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP64
; break;
4370 default: vtn_fail("Floating point type not supported");
4373 case SpvExecutionModeRoundingModeRTE
:
4374 switch (mode
->operands
[0]) {
4375 case 16: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
; break;
4376 case 32: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
; break;
4377 case 64: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
; break;
4378 default: vtn_fail("Floating point type not supported");
4381 case SpvExecutionModeRoundingModeRTZ
:
4382 switch (mode
->operands
[0]) {
4383 case 16: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
; break;
4384 case 32: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
; break;
4385 case 64: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
; break;
4386 default: vtn_fail("Floating point type not supported");
4394 b
->shader
->info
.float_controls_execution_mode
|= execution_mode
;
4398 vtn_handle_variable_or_type_instruction(struct vtn_builder
*b
, SpvOp opcode
,
4399 const uint32_t *w
, unsigned count
)
4401 vtn_set_instruction_result_type(b
, opcode
, w
, count
);
4405 case SpvOpSourceContinued
:
4406 case SpvOpSourceExtension
:
4407 case SpvOpExtension
:
4408 case SpvOpCapability
:
4409 case SpvOpExtInstImport
:
4410 case SpvOpMemoryModel
:
4411 case SpvOpEntryPoint
:
4412 case SpvOpExecutionMode
:
4415 case SpvOpMemberName
:
4416 case SpvOpDecorationGroup
:
4418 case SpvOpDecorateId
:
4419 case SpvOpMemberDecorate
:
4420 case SpvOpGroupDecorate
:
4421 case SpvOpGroupMemberDecorate
:
4422 case SpvOpDecorateString
:
4423 case SpvOpMemberDecorateString
:
4424 vtn_fail("Invalid opcode types and variables section");
4430 case SpvOpTypeFloat
:
4431 case SpvOpTypeVector
:
4432 case SpvOpTypeMatrix
:
4433 case SpvOpTypeImage
:
4434 case SpvOpTypeSampler
:
4435 case SpvOpTypeSampledImage
:
4436 case SpvOpTypeArray
:
4437 case SpvOpTypeRuntimeArray
:
4438 case SpvOpTypeStruct
:
4439 case SpvOpTypeOpaque
:
4440 case SpvOpTypePointer
:
4441 case SpvOpTypeForwardPointer
:
4442 case SpvOpTypeFunction
:
4443 case SpvOpTypeEvent
:
4444 case SpvOpTypeDeviceEvent
:
4445 case SpvOpTypeReserveId
:
4446 case SpvOpTypeQueue
:
4448 vtn_handle_type(b
, opcode
, w
, count
);
4451 case SpvOpConstantTrue
:
4452 case SpvOpConstantFalse
:
4454 case SpvOpConstantComposite
:
4455 case SpvOpConstantSampler
:
4456 case SpvOpConstantNull
:
4457 case SpvOpSpecConstantTrue
:
4458 case SpvOpSpecConstantFalse
:
4459 case SpvOpSpecConstant
:
4460 case SpvOpSpecConstantComposite
:
4461 case SpvOpSpecConstantOp
:
4462 vtn_handle_constant(b
, opcode
, w
, count
);
4467 vtn_handle_variables(b
, opcode
, w
, count
);
4471 return false; /* End of preamble */
4477 static struct vtn_ssa_value
*
4478 vtn_nir_select(struct vtn_builder
*b
, struct vtn_ssa_value
*src0
,
4479 struct vtn_ssa_value
*src1
, struct vtn_ssa_value
*src2
)
4481 struct vtn_ssa_value
*dest
= rzalloc(b
, struct vtn_ssa_value
);
4482 dest
->type
= src1
->type
;
4484 if (glsl_type_is_vector_or_scalar(src1
->type
)) {
4485 dest
->def
= nir_bcsel(&b
->nb
, src0
->def
, src1
->def
, src2
->def
);
4487 unsigned elems
= glsl_get_length(src1
->type
);
4489 dest
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
4490 for (unsigned i
= 0; i
< elems
; i
++) {
4491 dest
->elems
[i
] = vtn_nir_select(b
, src0
,
4492 src1
->elems
[i
], src2
->elems
[i
]);
4500 vtn_handle_select(struct vtn_builder
*b
, SpvOp opcode
,
4501 const uint32_t *w
, unsigned count
)
4503 /* Handle OpSelect up-front here because it needs to be able to handle
4504 * pointers and not just regular vectors and scalars.
4506 struct vtn_value
*res_val
= vtn_untyped_value(b
, w
[2]);
4507 struct vtn_value
*cond_val
= vtn_untyped_value(b
, w
[3]);
4508 struct vtn_value
*obj1_val
= vtn_untyped_value(b
, w
[4]);
4509 struct vtn_value
*obj2_val
= vtn_untyped_value(b
, w
[5]);
4511 vtn_fail_if(obj1_val
->type
!= res_val
->type
||
4512 obj2_val
->type
!= res_val
->type
,
4513 "Object types must match the result type in OpSelect");
4515 vtn_fail_if((cond_val
->type
->base_type
!= vtn_base_type_scalar
&&
4516 cond_val
->type
->base_type
!= vtn_base_type_vector
) ||
4517 !glsl_type_is_boolean(cond_val
->type
->type
),
4518 "OpSelect must have either a vector of booleans or "
4519 "a boolean as Condition type");
4521 vtn_fail_if(cond_val
->type
->base_type
== vtn_base_type_vector
&&
4522 (res_val
->type
->base_type
!= vtn_base_type_vector
||
4523 res_val
->type
->length
!= cond_val
->type
->length
),
4524 "When Condition type in OpSelect is a vector, the Result "
4525 "type must be a vector of the same length");
4527 switch (res_val
->type
->base_type
) {
4528 case vtn_base_type_scalar
:
4529 case vtn_base_type_vector
:
4530 case vtn_base_type_matrix
:
4531 case vtn_base_type_array
:
4532 case vtn_base_type_struct
:
4535 case vtn_base_type_pointer
:
4536 /* We need to have actual storage for pointer types. */
4537 vtn_fail_if(res_val
->type
->type
== NULL
,
4538 "Invalid pointer result type for OpSelect");
4541 vtn_fail("Result type of OpSelect must be a scalar, composite, or pointer");
4544 struct vtn_type
*res_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
4545 struct vtn_ssa_value
*ssa
= vtn_nir_select(b
,
4546 vtn_ssa_value(b
, w
[3]), vtn_ssa_value(b
, w
[4]), vtn_ssa_value(b
, w
[5]));
4548 vtn_push_ssa(b
, w
[2], res_type
, ssa
);
4552 vtn_handle_ptr(struct vtn_builder
*b
, SpvOp opcode
,
4553 const uint32_t *w
, unsigned count
)
4555 struct vtn_type
*type1
= vtn_untyped_value(b
, w
[3])->type
;
4556 struct vtn_type
*type2
= vtn_untyped_value(b
, w
[4])->type
;
4557 vtn_fail_if(type1
->base_type
!= vtn_base_type_pointer
||
4558 type2
->base_type
!= vtn_base_type_pointer
,
4559 "%s operands must have pointer types",
4560 spirv_op_to_string(opcode
));
4561 vtn_fail_if(type1
->storage_class
!= type2
->storage_class
,
4562 "%s operands must have the same storage class",
4563 spirv_op_to_string(opcode
));
4565 struct vtn_type
*vtn_type
=
4566 vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
4567 const struct glsl_type
*type
= vtn_type
->type
;
4569 nir_address_format addr_format
= vtn_mode_to_address_format(
4570 b
, vtn_storage_class_to_mode(b
, type1
->storage_class
, NULL
, NULL
));
4575 case SpvOpPtrDiff
: {
4576 /* OpPtrDiff returns the difference in number of elements (not byte offset). */
4577 unsigned elem_size
, elem_align
;
4578 glsl_get_natural_size_align_bytes(type1
->deref
->type
,
4579 &elem_size
, &elem_align
);
4581 def
= nir_build_addr_isub(&b
->nb
,
4582 vtn_ssa_value(b
, w
[3])->def
,
4583 vtn_ssa_value(b
, w
[4])->def
,
4585 def
= nir_idiv(&b
->nb
, def
, nir_imm_intN_t(&b
->nb
, elem_size
, def
->bit_size
));
4586 def
= nir_i2i(&b
->nb
, def
, glsl_get_bit_size(type
));
4591 case SpvOpPtrNotEqual
: {
4592 def
= nir_build_addr_ieq(&b
->nb
,
4593 vtn_ssa_value(b
, w
[3])->def
,
4594 vtn_ssa_value(b
, w
[4])->def
,
4596 if (opcode
== SpvOpPtrNotEqual
)
4597 def
= nir_inot(&b
->nb
, def
);
4602 unreachable("Invalid ptr operation");
4605 struct vtn_ssa_value
*ssa_value
= vtn_create_ssa_value(b
, type
);
4606 ssa_value
->def
= def
;
4607 vtn_push_ssa(b
, w
[2], vtn_type
, ssa_value
);
4611 vtn_handle_body_instruction(struct vtn_builder
*b
, SpvOp opcode
,
4612 const uint32_t *w
, unsigned count
)
4618 case SpvOpLoopMerge
:
4619 case SpvOpSelectionMerge
:
4620 /* This is handled by cfg pre-pass and walk_blocks */
4624 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
4625 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
4630 vtn_handle_extension(b
, opcode
, w
, count
);
4636 case SpvOpCopyMemory
:
4637 case SpvOpCopyMemorySized
:
4638 case SpvOpAccessChain
:
4639 case SpvOpPtrAccessChain
:
4640 case SpvOpInBoundsAccessChain
:
4641 case SpvOpInBoundsPtrAccessChain
:
4642 case SpvOpArrayLength
:
4643 case SpvOpConvertPtrToU
:
4644 case SpvOpConvertUToPtr
:
4645 vtn_handle_variables(b
, opcode
, w
, count
);
4648 case SpvOpFunctionCall
:
4649 vtn_handle_function_call(b
, opcode
, w
, count
);
4652 case SpvOpSampledImage
:
4654 case SpvOpImageSampleImplicitLod
:
4655 case SpvOpImageSampleExplicitLod
:
4656 case SpvOpImageSampleDrefImplicitLod
:
4657 case SpvOpImageSampleDrefExplicitLod
:
4658 case SpvOpImageSampleProjImplicitLod
:
4659 case SpvOpImageSampleProjExplicitLod
:
4660 case SpvOpImageSampleProjDrefImplicitLod
:
4661 case SpvOpImageSampleProjDrefExplicitLod
:
4662 case SpvOpImageFetch
:
4663 case SpvOpImageGather
:
4664 case SpvOpImageDrefGather
:
4665 case SpvOpImageQuerySizeLod
:
4666 case SpvOpImageQueryLod
:
4667 case SpvOpImageQueryLevels
:
4668 case SpvOpImageQuerySamples
:
4669 vtn_handle_texture(b
, opcode
, w
, count
);
4672 case SpvOpImageRead
:
4673 case SpvOpImageWrite
:
4674 case SpvOpImageTexelPointer
:
4675 vtn_handle_image(b
, opcode
, w
, count
);
4678 case SpvOpImageQuerySize
: {
4679 struct vtn_pointer
*image
=
4680 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
4681 if (glsl_type_is_image(image
->type
->type
)) {
4682 vtn_handle_image(b
, opcode
, w
, count
);
4684 vtn_assert(glsl_type_is_sampler(image
->type
->type
));
4685 vtn_handle_texture(b
, opcode
, w
, count
);
4690 case SpvOpAtomicLoad
:
4691 case SpvOpAtomicExchange
:
4692 case SpvOpAtomicCompareExchange
:
4693 case SpvOpAtomicCompareExchangeWeak
:
4694 case SpvOpAtomicIIncrement
:
4695 case SpvOpAtomicIDecrement
:
4696 case SpvOpAtomicIAdd
:
4697 case SpvOpAtomicISub
:
4698 case SpvOpAtomicSMin
:
4699 case SpvOpAtomicUMin
:
4700 case SpvOpAtomicSMax
:
4701 case SpvOpAtomicUMax
:
4702 case SpvOpAtomicAnd
:
4704 case SpvOpAtomicXor
: {
4705 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[3]);
4706 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
4707 vtn_handle_image(b
, opcode
, w
, count
);
4709 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
4710 vtn_handle_atomics(b
, opcode
, w
, count
);
4715 case SpvOpAtomicStore
: {
4716 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[1]);
4717 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
4718 vtn_handle_image(b
, opcode
, w
, count
);
4720 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
4721 vtn_handle_atomics(b
, opcode
, w
, count
);
4727 vtn_handle_select(b
, opcode
, w
, count
);
4735 case SpvOpConvertFToU
:
4736 case SpvOpConvertFToS
:
4737 case SpvOpConvertSToF
:
4738 case SpvOpConvertUToF
:
4742 case SpvOpQuantizeToF16
:
4743 case SpvOpPtrCastToGeneric
:
4744 case SpvOpGenericCastToPtr
:
4749 case SpvOpSignBitSet
:
4750 case SpvOpLessOrGreater
:
4752 case SpvOpUnordered
:
4767 case SpvOpVectorTimesScalar
:
4769 case SpvOpIAddCarry
:
4770 case SpvOpISubBorrow
:
4771 case SpvOpUMulExtended
:
4772 case SpvOpSMulExtended
:
4773 case SpvOpShiftRightLogical
:
4774 case SpvOpShiftRightArithmetic
:
4775 case SpvOpShiftLeftLogical
:
4776 case SpvOpLogicalEqual
:
4777 case SpvOpLogicalNotEqual
:
4778 case SpvOpLogicalOr
:
4779 case SpvOpLogicalAnd
:
4780 case SpvOpLogicalNot
:
4781 case SpvOpBitwiseOr
:
4782 case SpvOpBitwiseXor
:
4783 case SpvOpBitwiseAnd
:
4785 case SpvOpFOrdEqual
:
4786 case SpvOpFUnordEqual
:
4787 case SpvOpINotEqual
:
4788 case SpvOpFOrdNotEqual
:
4789 case SpvOpFUnordNotEqual
:
4790 case SpvOpULessThan
:
4791 case SpvOpSLessThan
:
4792 case SpvOpFOrdLessThan
:
4793 case SpvOpFUnordLessThan
:
4794 case SpvOpUGreaterThan
:
4795 case SpvOpSGreaterThan
:
4796 case SpvOpFOrdGreaterThan
:
4797 case SpvOpFUnordGreaterThan
:
4798 case SpvOpULessThanEqual
:
4799 case SpvOpSLessThanEqual
:
4800 case SpvOpFOrdLessThanEqual
:
4801 case SpvOpFUnordLessThanEqual
:
4802 case SpvOpUGreaterThanEqual
:
4803 case SpvOpSGreaterThanEqual
:
4804 case SpvOpFOrdGreaterThanEqual
:
4805 case SpvOpFUnordGreaterThanEqual
:
4811 case SpvOpFwidthFine
:
4812 case SpvOpDPdxCoarse
:
4813 case SpvOpDPdyCoarse
:
4814 case SpvOpFwidthCoarse
:
4815 case SpvOpBitFieldInsert
:
4816 case SpvOpBitFieldSExtract
:
4817 case SpvOpBitFieldUExtract
:
4818 case SpvOpBitReverse
:
4820 case SpvOpTranspose
:
4821 case SpvOpOuterProduct
:
4822 case SpvOpMatrixTimesScalar
:
4823 case SpvOpVectorTimesMatrix
:
4824 case SpvOpMatrixTimesVector
:
4825 case SpvOpMatrixTimesMatrix
:
4826 vtn_handle_alu(b
, opcode
, w
, count
);
4830 vtn_handle_bitcast(b
, w
, count
);
4833 case SpvOpVectorExtractDynamic
:
4834 case SpvOpVectorInsertDynamic
:
4835 case SpvOpVectorShuffle
:
4836 case SpvOpCompositeConstruct
:
4837 case SpvOpCompositeExtract
:
4838 case SpvOpCompositeInsert
:
4839 case SpvOpCopyLogical
:
4840 case SpvOpCopyObject
:
4841 vtn_handle_composite(b
, opcode
, w
, count
);
4844 case SpvOpEmitVertex
:
4845 case SpvOpEndPrimitive
:
4846 case SpvOpEmitStreamVertex
:
4847 case SpvOpEndStreamPrimitive
:
4848 case SpvOpControlBarrier
:
4849 case SpvOpMemoryBarrier
:
4850 vtn_handle_barrier(b
, opcode
, w
, count
);
4853 case SpvOpGroupNonUniformElect
:
4854 case SpvOpGroupNonUniformAll
:
4855 case SpvOpGroupNonUniformAny
:
4856 case SpvOpGroupNonUniformAllEqual
:
4857 case SpvOpGroupNonUniformBroadcast
:
4858 case SpvOpGroupNonUniformBroadcastFirst
:
4859 case SpvOpGroupNonUniformBallot
:
4860 case SpvOpGroupNonUniformInverseBallot
:
4861 case SpvOpGroupNonUniformBallotBitExtract
:
4862 case SpvOpGroupNonUniformBallotBitCount
:
4863 case SpvOpGroupNonUniformBallotFindLSB
:
4864 case SpvOpGroupNonUniformBallotFindMSB
:
4865 case SpvOpGroupNonUniformShuffle
:
4866 case SpvOpGroupNonUniformShuffleXor
:
4867 case SpvOpGroupNonUniformShuffleUp
:
4868 case SpvOpGroupNonUniformShuffleDown
:
4869 case SpvOpGroupNonUniformIAdd
:
4870 case SpvOpGroupNonUniformFAdd
:
4871 case SpvOpGroupNonUniformIMul
:
4872 case SpvOpGroupNonUniformFMul
:
4873 case SpvOpGroupNonUniformSMin
:
4874 case SpvOpGroupNonUniformUMin
:
4875 case SpvOpGroupNonUniformFMin
:
4876 case SpvOpGroupNonUniformSMax
:
4877 case SpvOpGroupNonUniformUMax
:
4878 case SpvOpGroupNonUniformFMax
:
4879 case SpvOpGroupNonUniformBitwiseAnd
:
4880 case SpvOpGroupNonUniformBitwiseOr
:
4881 case SpvOpGroupNonUniformBitwiseXor
:
4882 case SpvOpGroupNonUniformLogicalAnd
:
4883 case SpvOpGroupNonUniformLogicalOr
:
4884 case SpvOpGroupNonUniformLogicalXor
:
4885 case SpvOpGroupNonUniformQuadBroadcast
:
4886 case SpvOpGroupNonUniformQuadSwap
:
4889 case SpvOpGroupBroadcast
:
4890 case SpvOpGroupIAdd
:
4891 case SpvOpGroupFAdd
:
4892 case SpvOpGroupFMin
:
4893 case SpvOpGroupUMin
:
4894 case SpvOpGroupSMin
:
4895 case SpvOpGroupFMax
:
4896 case SpvOpGroupUMax
:
4897 case SpvOpGroupSMax
:
4898 case SpvOpSubgroupBallotKHR
:
4899 case SpvOpSubgroupFirstInvocationKHR
:
4900 case SpvOpSubgroupReadInvocationKHR
:
4901 case SpvOpSubgroupAllKHR
:
4902 case SpvOpSubgroupAnyKHR
:
4903 case SpvOpSubgroupAllEqualKHR
:
4904 case SpvOpGroupIAddNonUniformAMD
:
4905 case SpvOpGroupFAddNonUniformAMD
:
4906 case SpvOpGroupFMinNonUniformAMD
:
4907 case SpvOpGroupUMinNonUniformAMD
:
4908 case SpvOpGroupSMinNonUniformAMD
:
4909 case SpvOpGroupFMaxNonUniformAMD
:
4910 case SpvOpGroupUMaxNonUniformAMD
:
4911 case SpvOpGroupSMaxNonUniformAMD
:
4912 vtn_handle_subgroup(b
, opcode
, w
, count
);
4917 case SpvOpPtrNotEqual
:
4918 vtn_handle_ptr(b
, opcode
, w
, count
);
4921 case SpvOpBeginInvocationInterlockEXT
:
4922 vtn_emit_barrier(b
, nir_intrinsic_begin_invocation_interlock
);
4925 case SpvOpEndInvocationInterlockEXT
:
4926 vtn_emit_barrier(b
, nir_intrinsic_end_invocation_interlock
);
4929 case SpvOpDemoteToHelperInvocationEXT
: {
4930 nir_intrinsic_instr
*intrin
=
4931 nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_demote
);
4932 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
4936 case SpvOpIsHelperInvocationEXT
: {
4937 nir_intrinsic_instr
*intrin
=
4938 nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_is_helper_invocation
);
4939 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
, 1, 1, NULL
);
4940 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
4942 struct vtn_type
*res_type
=
4943 vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
4944 struct vtn_ssa_value
*val
= vtn_create_ssa_value(b
, res_type
->type
);
4945 val
->def
= &intrin
->dest
.ssa
;
4947 vtn_push_ssa(b
, w
[2], res_type
, val
);
4951 case SpvOpReadClockKHR
: {
4952 assert(vtn_constant_uint(b
, w
[3]) == SpvScopeSubgroup
);
4954 /* Operation supports two result types: uvec2 and uint64_t. The NIR
4955 * intrinsic gives uvec2, so pack the result for the other case.
4957 nir_intrinsic_instr
*intrin
=
4958 nir_intrinsic_instr_create(b
->nb
.shader
, nir_intrinsic_shader_clock
);
4959 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
, 2, 32, NULL
);
4960 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
4962 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
4963 const struct glsl_type
*dest_type
= type
->type
;
4964 nir_ssa_def
*result
;
4966 if (glsl_type_is_vector(dest_type
)) {
4967 assert(dest_type
== glsl_vector_type(GLSL_TYPE_UINT
, 2));
4968 result
= &intrin
->dest
.ssa
;
4970 assert(glsl_type_is_scalar(dest_type
));
4971 assert(glsl_get_base_type(dest_type
) == GLSL_TYPE_UINT64
);
4972 result
= nir_pack_64_2x32(&b
->nb
, &intrin
->dest
.ssa
);
4975 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
4977 val
->ssa
= vtn_create_ssa_value(b
, dest_type
);
4978 val
->ssa
->def
= result
;
4983 vtn_fail_with_opcode("Unhandled opcode", opcode
);
4990 vtn_create_builder(const uint32_t *words
, size_t word_count
,
4991 gl_shader_stage stage
, const char *entry_point_name
,
4992 const struct spirv_to_nir_options
*options
)
4994 /* Initialize the vtn_builder object */
4995 struct vtn_builder
*b
= rzalloc(NULL
, struct vtn_builder
);
4996 struct spirv_to_nir_options
*dup_options
=
4997 ralloc(b
, struct spirv_to_nir_options
);
4998 *dup_options
= *options
;
5001 b
->spirv_word_count
= word_count
;
5005 exec_list_make_empty(&b
->functions
);
5006 b
->entry_point_stage
= stage
;
5007 b
->entry_point_name
= entry_point_name
;
5008 b
->options
= dup_options
;
5011 * Handle the SPIR-V header (first 5 dwords).
5012 * Can't use vtx_assert() as the setjmp(3) target isn't initialized yet.
5014 if (word_count
<= 5)
5017 if (words
[0] != SpvMagicNumber
) {
5018 vtn_err("words[0] was 0x%x, want 0x%x", words
[0], SpvMagicNumber
);
5021 if (words
[1] < 0x10000) {
5022 vtn_err("words[1] was 0x%x, want >= 0x10000", words
[1]);
5026 uint16_t generator_id
= words
[2] >> 16;
5027 uint16_t generator_version
= words
[2];
5029 /* The first GLSLang version bump actually 1.5 years after #179 was fixed
5030 * but this should at least let us shut the workaround off for modern
5031 * versions of GLSLang.
5033 b
->wa_glslang_179
= (generator_id
== 8 && generator_version
== 1);
5035 /* words[2] == generator magic */
5036 unsigned value_id_bound
= words
[3];
5037 if (words
[4] != 0) {
5038 vtn_err("words[4] was %u, want 0", words
[4]);
5042 b
->value_id_bound
= value_id_bound
;
5043 b
->values
= rzalloc_array(b
, struct vtn_value
, value_id_bound
);
5051 static nir_function
*
5052 vtn_emit_kernel_entry_point_wrapper(struct vtn_builder
*b
,
5053 nir_function
*entry_point
)
5055 vtn_assert(entry_point
== b
->entry_point
->func
->impl
->function
);
5056 vtn_fail_if(!entry_point
->name
, "entry points are required to have a name");
5057 const char *func_name
=
5058 ralloc_asprintf(b
->shader
, "__wrapped_%s", entry_point
->name
);
5060 /* we shouldn't have any inputs yet */
5061 vtn_assert(!entry_point
->shader
->num_inputs
);
5062 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_KERNEL
);
5064 nir_function
*main_entry_point
= nir_function_create(b
->shader
, func_name
);
5065 main_entry_point
->impl
= nir_function_impl_create(main_entry_point
);
5066 nir_builder_init(&b
->nb
, main_entry_point
->impl
);
5067 b
->nb
.cursor
= nir_after_cf_list(&main_entry_point
->impl
->body
);
5068 b
->func_param_idx
= 0;
5070 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, entry_point
);
5072 for (unsigned i
= 0; i
< entry_point
->num_params
; ++i
) {
5073 struct vtn_type
*param_type
= b
->entry_point
->func
->type
->params
[i
];
5075 /* consider all pointers to function memory to be parameters passed
5078 bool is_by_val
= param_type
->base_type
== vtn_base_type_pointer
&&
5079 param_type
->storage_class
== SpvStorageClassFunction
;
5081 /* input variable */
5082 nir_variable
*in_var
= rzalloc(b
->nb
.shader
, nir_variable
);
5083 in_var
->data
.mode
= nir_var_shader_in
;
5084 in_var
->data
.read_only
= true;
5085 in_var
->data
.location
= i
;
5088 in_var
->type
= param_type
->deref
->type
;
5090 in_var
->type
= param_type
->type
;
5092 nir_shader_add_variable(b
->nb
.shader
, in_var
);
5093 b
->nb
.shader
->num_inputs
++;
5095 /* we have to copy the entire variable into function memory */
5097 nir_variable
*copy_var
=
5098 nir_local_variable_create(main_entry_point
->impl
, in_var
->type
,
5100 nir_copy_var(&b
->nb
, copy_var
, in_var
);
5102 nir_src_for_ssa(&nir_build_deref_var(&b
->nb
, copy_var
)->dest
.ssa
);
5104 call
->params
[i
] = nir_src_for_ssa(nir_load_var(&b
->nb
, in_var
));
5108 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
5110 return main_entry_point
;
5114 spirv_to_nir(const uint32_t *words
, size_t word_count
,
5115 struct nir_spirv_specialization
*spec
, unsigned num_spec
,
5116 gl_shader_stage stage
, const char *entry_point_name
,
5117 const struct spirv_to_nir_options
*options
,
5118 const nir_shader_compiler_options
*nir_options
)
5121 const uint32_t *word_end
= words
+ word_count
;
5123 struct vtn_builder
*b
= vtn_create_builder(words
, word_count
,
5124 stage
, entry_point_name
,
5130 /* See also _vtn_fail() */
5131 if (setjmp(b
->fail_jump
)) {
5136 /* Skip the SPIR-V header, handled at vtn_create_builder */
5139 b
->shader
= nir_shader_create(b
, stage
, nir_options
, NULL
);
5141 /* Handle all the preamble instructions */
5142 words
= vtn_foreach_instruction(b
, words
, word_end
,
5143 vtn_handle_preamble_instruction
);
5145 if (b
->entry_point
== NULL
) {
5146 vtn_fail("Entry point not found");
5151 /* Set shader info defaults */
5152 if (stage
== MESA_SHADER_GEOMETRY
)
5153 b
->shader
->info
.gs
.invocations
= 1;
5155 /* Parse rounding mode execution modes. This has to happen earlier than
5156 * other changes in the execution modes since they can affect, for example,
5157 * the result of the floating point constants.
5159 vtn_foreach_execution_mode(b
, b
->entry_point
,
5160 vtn_handle_rounding_mode_in_execution_mode
, NULL
);
5162 b
->specializations
= spec
;
5163 b
->num_specializations
= num_spec
;
5165 /* Handle all variable, type, and constant instructions */
5166 words
= vtn_foreach_instruction(b
, words
, word_end
,
5167 vtn_handle_variable_or_type_instruction
);
5169 /* Parse execution modes */
5170 vtn_foreach_execution_mode(b
, b
->entry_point
,
5171 vtn_handle_execution_mode
, NULL
);
5173 if (b
->workgroup_size_builtin
) {
5174 vtn_assert(b
->workgroup_size_builtin
->type
->type
==
5175 glsl_vector_type(GLSL_TYPE_UINT
, 3));
5177 nir_const_value
*const_size
=
5178 b
->workgroup_size_builtin
->constant
->values
;
5180 b
->shader
->info
.cs
.local_size
[0] = const_size
[0].u32
;
5181 b
->shader
->info
.cs
.local_size
[1] = const_size
[1].u32
;
5182 b
->shader
->info
.cs
.local_size
[2] = const_size
[2].u32
;
5185 /* Set types on all vtn_values */
5186 vtn_foreach_instruction(b
, words
, word_end
, vtn_set_instruction_result_type
);
5188 vtn_build_cfg(b
, words
, word_end
);
5190 assert(b
->entry_point
->value_type
== vtn_value_type_function
);
5191 b
->entry_point
->func
->referenced
= true;
5196 foreach_list_typed(struct vtn_function
, func
, node
, &b
->functions
) {
5197 if (func
->referenced
&& !func
->emitted
) {
5198 b
->const_table
= _mesa_pointer_hash_table_create(b
);
5200 vtn_function_emit(b
, func
, vtn_handle_body_instruction
);
5206 vtn_assert(b
->entry_point
->value_type
== vtn_value_type_function
);
5207 nir_function
*entry_point
= b
->entry_point
->func
->impl
->function
;
5208 vtn_assert(entry_point
);
5210 /* post process entry_points with input params */
5211 if (entry_point
->num_params
&& b
->shader
->info
.stage
== MESA_SHADER_KERNEL
)
5212 entry_point
= vtn_emit_kernel_entry_point_wrapper(b
, entry_point
);
5214 entry_point
->is_entrypoint
= true;
5216 /* When multiple shader stages exist in the same SPIR-V module, we
5217 * generate input and output variables for every stage, in the same
5218 * NIR program. These dead variables can be invalid NIR. For example,
5219 * TCS outputs must be per-vertex arrays (or decorated 'patch'), while
5220 * VS output variables wouldn't be.
5222 * To ensure we have valid NIR, we eliminate any dead inputs and outputs
5223 * right away. In order to do so, we must lower any constant initializers
5224 * on outputs so nir_remove_dead_variables sees that they're written to.
5226 nir_lower_constant_initializers(b
->shader
, nir_var_shader_out
);
5227 nir_remove_dead_variables(b
->shader
,
5228 nir_var_shader_in
| nir_var_shader_out
);
5230 /* We sometimes generate bogus derefs that, while never used, give the
5231 * validator a bit of heartburn. Run dead code to get rid of them.
5233 nir_opt_dce(b
->shader
);
5235 /* Unparent the shader from the vtn_builder before we delete the builder */
5236 ralloc_steal(NULL
, b
->shader
);
5238 nir_shader
*shader
= b
->shader
;