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
[0],
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 rows
= glsl_get_vector_elements(val
->type
);
247 unsigned columns
= glsl_get_matrix_columns(val
->type
);
248 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, columns
);
250 for (unsigned i
= 0; i
< columns
; i
++) {
251 struct vtn_ssa_value
*col_val
= rzalloc(b
, struct vtn_ssa_value
);
252 col_val
->type
= glsl_get_column_type(val
->type
);
253 nir_load_const_instr
*load
=
254 nir_load_const_instr_create(b
->shader
, rows
, bit_size
);
256 memcpy(load
->value
, constant
->values
[i
],
257 sizeof(nir_const_value
) * load
->def
.num_components
);
259 nir_instr_insert_before_cf_list(&b
->nb
.impl
->body
, &load
->instr
);
260 col_val
->def
= &load
->def
;
262 val
->elems
[i
] = col_val
;
268 case GLSL_TYPE_ARRAY
: {
269 unsigned elems
= glsl_get_length(val
->type
);
270 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
271 const struct glsl_type
*elem_type
= glsl_get_array_element(val
->type
);
272 for (unsigned i
= 0; i
< elems
; i
++)
273 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
278 case GLSL_TYPE_STRUCT
: {
279 unsigned elems
= glsl_get_length(val
->type
);
280 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
281 for (unsigned i
= 0; i
< elems
; i
++) {
282 const struct glsl_type
*elem_type
=
283 glsl_get_struct_field(val
->type
, i
);
284 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
291 vtn_fail("bad constant type");
297 struct vtn_ssa_value
*
298 vtn_ssa_value(struct vtn_builder
*b
, uint32_t value_id
)
300 struct vtn_value
*val
= vtn_untyped_value(b
, value_id
);
301 switch (val
->value_type
) {
302 case vtn_value_type_undef
:
303 return vtn_undef_ssa_value(b
, val
->type
->type
);
305 case vtn_value_type_constant
:
306 return vtn_const_ssa_value(b
, val
->constant
, val
->type
->type
);
308 case vtn_value_type_ssa
:
311 case vtn_value_type_pointer
:
312 vtn_assert(val
->pointer
->ptr_type
&& val
->pointer
->ptr_type
->type
);
313 struct vtn_ssa_value
*ssa
=
314 vtn_create_ssa_value(b
, val
->pointer
->ptr_type
->type
);
315 ssa
->def
= vtn_pointer_to_ssa(b
, val
->pointer
);
319 vtn_fail("Invalid type for an SSA value");
324 vtn_string_literal(struct vtn_builder
*b
, const uint32_t *words
,
325 unsigned word_count
, unsigned *words_used
)
327 char *dup
= ralloc_strndup(b
, (char *)words
, word_count
* sizeof(*words
));
329 /* Ammount of space taken by the string (including the null) */
330 unsigned len
= strlen(dup
) + 1;
331 *words_used
= DIV_ROUND_UP(len
, sizeof(*words
));
337 vtn_foreach_instruction(struct vtn_builder
*b
, const uint32_t *start
,
338 const uint32_t *end
, vtn_instruction_handler handler
)
344 const uint32_t *w
= start
;
346 SpvOp opcode
= w
[0] & SpvOpCodeMask
;
347 unsigned count
= w
[0] >> SpvWordCountShift
;
348 vtn_assert(count
>= 1 && w
+ count
<= end
);
350 b
->spirv_offset
= (uint8_t *)w
- (uint8_t *)b
->spirv
;
354 break; /* Do nothing */
357 b
->file
= vtn_value(b
, w
[1], vtn_value_type_string
)->str
;
369 if (!handler(b
, opcode
, w
, count
))
387 vtn_handle_extension(struct vtn_builder
*b
, SpvOp opcode
,
388 const uint32_t *w
, unsigned count
)
390 const char *ext
= (const char *)&w
[2];
392 case SpvOpExtInstImport
: {
393 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_extension
);
394 if (strcmp(ext
, "GLSL.std.450") == 0) {
395 val
->ext_handler
= vtn_handle_glsl450_instruction
;
396 } else if ((strcmp(ext
, "SPV_AMD_gcn_shader") == 0)
397 && (b
->options
&& b
->options
->caps
.gcn_shader
)) {
398 val
->ext_handler
= vtn_handle_amd_gcn_shader_instruction
;
399 } else if ((strcmp(ext
, "SPV_AMD_shader_trinary_minmax") == 0)
400 && (b
->options
&& b
->options
->caps
.trinary_minmax
)) {
401 val
->ext_handler
= vtn_handle_amd_shader_trinary_minmax_instruction
;
402 } else if (strcmp(ext
, "OpenCL.std") == 0) {
403 val
->ext_handler
= vtn_handle_opencl_instruction
;
405 vtn_fail("Unsupported extension: %s", ext
);
411 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
412 bool handled
= val
->ext_handler(b
, w
[4], w
, count
);
418 vtn_fail("Unhandled opcode");
423 _foreach_decoration_helper(struct vtn_builder
*b
,
424 struct vtn_value
*base_value
,
426 struct vtn_value
*value
,
427 vtn_decoration_foreach_cb cb
, void *data
)
429 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
431 if (dec
->scope
== VTN_DEC_DECORATION
) {
432 member
= parent_member
;
433 } else if (dec
->scope
>= VTN_DEC_STRUCT_MEMBER0
) {
434 vtn_fail_if(value
->value_type
!= vtn_value_type_type
||
435 value
->type
->base_type
!= vtn_base_type_struct
,
436 "OpMemberDecorate and OpGroupMemberDecorate are only "
437 "allowed on OpTypeStruct");
438 /* This means we haven't recursed yet */
439 assert(value
== base_value
);
441 member
= dec
->scope
- VTN_DEC_STRUCT_MEMBER0
;
443 vtn_fail_if(member
>= base_value
->type
->length
,
444 "OpMemberDecorate specifies member %d but the "
445 "OpTypeStruct has only %u members",
446 member
, base_value
->type
->length
);
448 /* Not a decoration */
449 assert(dec
->scope
== VTN_DEC_EXECUTION_MODE
);
454 assert(dec
->group
->value_type
== vtn_value_type_decoration_group
);
455 _foreach_decoration_helper(b
, base_value
, member
, dec
->group
,
458 cb(b
, base_value
, member
, dec
, data
);
463 /** Iterates (recursively if needed) over all of the decorations on a value
465 * This function iterates over all of the decorations applied to a given
466 * value. If it encounters a decoration group, it recurses into the group
467 * and iterates over all of those decorations as well.
470 vtn_foreach_decoration(struct vtn_builder
*b
, struct vtn_value
*value
,
471 vtn_decoration_foreach_cb cb
, void *data
)
473 _foreach_decoration_helper(b
, value
, -1, value
, cb
, data
);
477 vtn_foreach_execution_mode(struct vtn_builder
*b
, struct vtn_value
*value
,
478 vtn_execution_mode_foreach_cb cb
, void *data
)
480 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
481 if (dec
->scope
!= VTN_DEC_EXECUTION_MODE
)
484 assert(dec
->group
== NULL
);
485 cb(b
, value
, dec
, data
);
490 vtn_handle_decoration(struct vtn_builder
*b
, SpvOp opcode
,
491 const uint32_t *w
, unsigned count
)
493 const uint32_t *w_end
= w
+ count
;
494 const uint32_t target
= w
[1];
498 case SpvOpDecorationGroup
:
499 vtn_push_value(b
, target
, vtn_value_type_decoration_group
);
503 case SpvOpMemberDecorate
:
504 case SpvOpDecorateStringGOOGLE
:
505 case SpvOpMemberDecorateStringGOOGLE
:
506 case SpvOpExecutionMode
:
507 case SpvOpExecutionModeId
: {
508 struct vtn_value
*val
= vtn_untyped_value(b
, target
);
510 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
513 case SpvOpDecorateStringGOOGLE
:
514 dec
->scope
= VTN_DEC_DECORATION
;
516 case SpvOpMemberDecorate
:
517 case SpvOpMemberDecorateStringGOOGLE
:
518 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(w
++);
519 vtn_fail_if(dec
->scope
< VTN_DEC_STRUCT_MEMBER0
, /* overflow */
520 "Member argument of OpMemberDecorate too large");
522 case SpvOpExecutionMode
:
523 case SpvOpExecutionModeId
:
524 dec
->scope
= VTN_DEC_EXECUTION_MODE
;
527 unreachable("Invalid decoration opcode");
529 dec
->decoration
= *(w
++);
532 /* Link into the list */
533 dec
->next
= val
->decoration
;
534 val
->decoration
= dec
;
538 case SpvOpGroupMemberDecorate
:
539 case SpvOpGroupDecorate
: {
540 struct vtn_value
*group
=
541 vtn_value(b
, target
, vtn_value_type_decoration_group
);
543 for (; w
< w_end
; w
++) {
544 struct vtn_value
*val
= vtn_untyped_value(b
, *w
);
545 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
548 if (opcode
== SpvOpGroupDecorate
) {
549 dec
->scope
= VTN_DEC_DECORATION
;
551 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(++w
);
552 vtn_fail_if(dec
->scope
< 0, /* Check for overflow */
553 "Member argument of OpGroupMemberDecorate too large");
556 /* Link into the list */
557 dec
->next
= val
->decoration
;
558 val
->decoration
= dec
;
564 unreachable("Unhandled opcode");
568 struct member_decoration_ctx
{
570 struct glsl_struct_field
*fields
;
571 struct vtn_type
*type
;
575 * Returns true if the given type contains a struct decorated Block or
579 vtn_type_contains_block(struct vtn_builder
*b
, struct vtn_type
*type
)
581 switch (type
->base_type
) {
582 case vtn_base_type_array
:
583 return vtn_type_contains_block(b
, type
->array_element
);
584 case vtn_base_type_struct
:
585 if (type
->block
|| type
->buffer_block
)
587 for (unsigned i
= 0; i
< type
->length
; i
++) {
588 if (vtn_type_contains_block(b
, type
->members
[i
]))
597 /** Returns true if two types are "compatible", i.e. you can do an OpLoad,
598 * OpStore, or OpCopyMemory between them without breaking anything.
599 * Technically, the SPIR-V rules require the exact same type ID but this lets
600 * us internally be a bit looser.
603 vtn_types_compatible(struct vtn_builder
*b
,
604 struct vtn_type
*t1
, struct vtn_type
*t2
)
606 if (t1
->id
== t2
->id
)
609 if (t1
->base_type
!= t2
->base_type
)
612 switch (t1
->base_type
) {
613 case vtn_base_type_void
:
614 case vtn_base_type_scalar
:
615 case vtn_base_type_vector
:
616 case vtn_base_type_matrix
:
617 case vtn_base_type_image
:
618 case vtn_base_type_sampler
:
619 case vtn_base_type_sampled_image
:
620 return t1
->type
== t2
->type
;
622 case vtn_base_type_array
:
623 return t1
->length
== t2
->length
&&
624 vtn_types_compatible(b
, t1
->array_element
, t2
->array_element
);
626 case vtn_base_type_pointer
:
627 return vtn_types_compatible(b
, t1
->deref
, t2
->deref
);
629 case vtn_base_type_struct
:
630 if (t1
->length
!= t2
->length
)
633 for (unsigned i
= 0; i
< t1
->length
; i
++) {
634 if (!vtn_types_compatible(b
, t1
->members
[i
], t2
->members
[i
]))
639 case vtn_base_type_function
:
640 /* This case shouldn't get hit since you can't copy around function
641 * types. Just require them to be identical.
646 vtn_fail("Invalid base type");
649 /* does a shallow copy of a vtn_type */
651 static struct vtn_type
*
652 vtn_type_copy(struct vtn_builder
*b
, struct vtn_type
*src
)
654 struct vtn_type
*dest
= ralloc(b
, struct vtn_type
);
657 switch (src
->base_type
) {
658 case vtn_base_type_void
:
659 case vtn_base_type_scalar
:
660 case vtn_base_type_vector
:
661 case vtn_base_type_matrix
:
662 case vtn_base_type_array
:
663 case vtn_base_type_pointer
:
664 case vtn_base_type_image
:
665 case vtn_base_type_sampler
:
666 case vtn_base_type_sampled_image
:
667 /* Nothing more to do */
670 case vtn_base_type_struct
:
671 dest
->members
= ralloc_array(b
, struct vtn_type
*, src
->length
);
672 memcpy(dest
->members
, src
->members
,
673 src
->length
* sizeof(src
->members
[0]));
675 dest
->offsets
= ralloc_array(b
, unsigned, src
->length
);
676 memcpy(dest
->offsets
, src
->offsets
,
677 src
->length
* sizeof(src
->offsets
[0]));
680 case vtn_base_type_function
:
681 dest
->params
= ralloc_array(b
, struct vtn_type
*, src
->length
);
682 memcpy(dest
->params
, src
->params
, src
->length
* sizeof(src
->params
[0]));
689 static struct vtn_type
*
690 mutable_matrix_member(struct vtn_builder
*b
, struct vtn_type
*type
, int member
)
692 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
693 type
= type
->members
[member
];
695 /* We may have an array of matrices.... Oh, joy! */
696 while (glsl_type_is_array(type
->type
)) {
697 type
->array_element
= vtn_type_copy(b
, type
->array_element
);
698 type
= type
->array_element
;
701 vtn_assert(glsl_type_is_matrix(type
->type
));
707 vtn_handle_access_qualifier(struct vtn_builder
*b
, struct vtn_type
*type
,
708 int member
, enum gl_access_qualifier access
)
710 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
711 type
= type
->members
[member
];
713 type
->access
|= access
;
717 array_stride_decoration_cb(struct vtn_builder
*b
,
718 struct vtn_value
*val
, int member
,
719 const struct vtn_decoration
*dec
, void *void_ctx
)
721 struct vtn_type
*type
= val
->type
;
723 if (dec
->decoration
== SpvDecorationArrayStride
) {
724 vtn_fail_if(dec
->literals
[0] == 0, "ArrayStride must be non-zero");
725 type
->stride
= dec
->literals
[0];
730 struct_member_decoration_cb(struct vtn_builder
*b
,
731 struct vtn_value
*val
, int member
,
732 const struct vtn_decoration
*dec
, void *void_ctx
)
734 struct member_decoration_ctx
*ctx
= void_ctx
;
739 assert(member
< ctx
->num_fields
);
741 switch (dec
->decoration
) {
742 case SpvDecorationRelaxedPrecision
:
743 case SpvDecorationUniform
:
744 break; /* FIXME: Do nothing with this for now. */
745 case SpvDecorationNonWritable
:
746 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_NON_WRITEABLE
);
748 case SpvDecorationNonReadable
:
749 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_NON_READABLE
);
751 case SpvDecorationVolatile
:
752 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_VOLATILE
);
754 case SpvDecorationCoherent
:
755 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_COHERENT
);
757 case SpvDecorationNoPerspective
:
758 ctx
->fields
[member
].interpolation
= INTERP_MODE_NOPERSPECTIVE
;
760 case SpvDecorationFlat
:
761 ctx
->fields
[member
].interpolation
= INTERP_MODE_FLAT
;
763 case SpvDecorationCentroid
:
764 ctx
->fields
[member
].centroid
= true;
766 case SpvDecorationSample
:
767 ctx
->fields
[member
].sample
= true;
769 case SpvDecorationStream
:
770 /* Vulkan only allows one GS stream */
771 vtn_assert(dec
->literals
[0] == 0);
773 case SpvDecorationLocation
:
774 ctx
->fields
[member
].location
= dec
->literals
[0];
776 case SpvDecorationComponent
:
777 break; /* FIXME: What should we do with these? */
778 case SpvDecorationBuiltIn
:
779 ctx
->type
->members
[member
] = vtn_type_copy(b
, ctx
->type
->members
[member
]);
780 ctx
->type
->members
[member
]->is_builtin
= true;
781 ctx
->type
->members
[member
]->builtin
= dec
->literals
[0];
782 ctx
->type
->builtin_block
= true;
784 case SpvDecorationOffset
:
785 ctx
->type
->offsets
[member
] = dec
->literals
[0];
786 ctx
->fields
[member
].offset
= dec
->literals
[0];
788 case SpvDecorationMatrixStride
:
789 /* Handled as a second pass */
791 case SpvDecorationColMajor
:
792 break; /* Nothing to do here. Column-major is the default. */
793 case SpvDecorationRowMajor
:
794 mutable_matrix_member(b
, ctx
->type
, member
)->row_major
= true;
797 case SpvDecorationPatch
:
800 case SpvDecorationSpecId
:
801 case SpvDecorationBlock
:
802 case SpvDecorationBufferBlock
:
803 case SpvDecorationArrayStride
:
804 case SpvDecorationGLSLShared
:
805 case SpvDecorationGLSLPacked
:
806 case SpvDecorationInvariant
:
807 case SpvDecorationRestrict
:
808 case SpvDecorationAliased
:
809 case SpvDecorationConstant
:
810 case SpvDecorationIndex
:
811 case SpvDecorationBinding
:
812 case SpvDecorationDescriptorSet
:
813 case SpvDecorationLinkageAttributes
:
814 case SpvDecorationNoContraction
:
815 case SpvDecorationInputAttachmentIndex
:
816 vtn_warn("Decoration not allowed on struct members: %s",
817 spirv_decoration_to_string(dec
->decoration
));
820 case SpvDecorationXfbBuffer
:
821 case SpvDecorationXfbStride
:
822 vtn_warn("Vulkan does not have transform feedback");
825 case SpvDecorationCPacked
:
826 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
)
827 vtn_warn("Decoration only allowed for CL-style kernels: %s",
828 spirv_decoration_to_string(dec
->decoration
));
830 ctx
->type
->packed
= true;
833 case SpvDecorationSaturatedConversion
:
834 case SpvDecorationFuncParamAttr
:
835 case SpvDecorationFPRoundingMode
:
836 case SpvDecorationFPFastMathMode
:
837 case SpvDecorationAlignment
:
838 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
) {
839 vtn_warn("Decoration only allowed for CL-style kernels: %s",
840 spirv_decoration_to_string(dec
->decoration
));
844 case SpvDecorationHlslSemanticGOOGLE
:
845 /* HLSL semantic decorations can safely be ignored by the driver. */
849 vtn_fail("Unhandled decoration");
853 /** Chases the array type all the way down to the tail and rewrites the
854 * glsl_types to be based off the tail's glsl_type.
857 vtn_array_type_rewrite_glsl_type(struct vtn_type
*type
)
859 if (type
->base_type
!= vtn_base_type_array
)
862 vtn_array_type_rewrite_glsl_type(type
->array_element
);
864 type
->type
= glsl_array_type(type
->array_element
->type
,
865 type
->length
, type
->stride
);
868 /* Matrix strides are handled as a separate pass because we need to know
869 * whether the matrix is row-major or not first.
872 struct_member_matrix_stride_cb(struct vtn_builder
*b
,
873 struct vtn_value
*val
, int member
,
874 const struct vtn_decoration
*dec
,
877 if (dec
->decoration
!= SpvDecorationMatrixStride
)
880 vtn_fail_if(member
< 0,
881 "The MatrixStride decoration is only allowed on members "
883 vtn_fail_if(dec
->literals
[0] == 0, "MatrixStride must be non-zero");
885 struct member_decoration_ctx
*ctx
= void_ctx
;
887 struct vtn_type
*mat_type
= mutable_matrix_member(b
, ctx
->type
, member
);
888 if (mat_type
->row_major
) {
889 mat_type
->array_element
= vtn_type_copy(b
, mat_type
->array_element
);
890 mat_type
->stride
= mat_type
->array_element
->stride
;
891 mat_type
->array_element
->stride
= dec
->literals
[0];
893 mat_type
->type
= glsl_explicit_matrix_type(mat_type
->type
,
894 dec
->literals
[0], true);
895 mat_type
->array_element
->type
= glsl_get_column_type(mat_type
->type
);
897 vtn_assert(mat_type
->array_element
->stride
> 0);
898 mat_type
->stride
= dec
->literals
[0];
900 mat_type
->type
= glsl_explicit_matrix_type(mat_type
->type
,
901 dec
->literals
[0], false);
904 /* Now that we've replaced the glsl_type with a properly strided matrix
905 * type, rewrite the member type so that it's an array of the proper kind
908 vtn_array_type_rewrite_glsl_type(ctx
->type
->members
[member
]);
909 ctx
->fields
[member
].type
= ctx
->type
->members
[member
]->type
;
913 struct_block_decoration_cb(struct vtn_builder
*b
,
914 struct vtn_value
*val
, int member
,
915 const struct vtn_decoration
*dec
, void *ctx
)
920 struct vtn_type
*type
= val
->type
;
921 if (dec
->decoration
== SpvDecorationBlock
)
923 else if (dec
->decoration
== SpvDecorationBufferBlock
)
924 type
->buffer_block
= true;
928 type_decoration_cb(struct vtn_builder
*b
,
929 struct vtn_value
*val
, int member
,
930 const struct vtn_decoration
*dec
, void *ctx
)
932 struct vtn_type
*type
= val
->type
;
935 /* This should have been handled by OpTypeStruct */
936 assert(val
->type
->base_type
== vtn_base_type_struct
);
937 assert(member
>= 0 && member
< val
->type
->length
);
941 switch (dec
->decoration
) {
942 case SpvDecorationArrayStride
:
943 vtn_assert(type
->base_type
== vtn_base_type_array
||
944 type
->base_type
== vtn_base_type_pointer
);
946 case SpvDecorationBlock
:
947 vtn_assert(type
->base_type
== vtn_base_type_struct
);
948 vtn_assert(type
->block
);
950 case SpvDecorationBufferBlock
:
951 vtn_assert(type
->base_type
== vtn_base_type_struct
);
952 vtn_assert(type
->buffer_block
);
954 case SpvDecorationGLSLShared
:
955 case SpvDecorationGLSLPacked
:
956 /* Ignore these, since we get explicit offsets anyways */
959 case SpvDecorationRowMajor
:
960 case SpvDecorationColMajor
:
961 case SpvDecorationMatrixStride
:
962 case SpvDecorationBuiltIn
:
963 case SpvDecorationNoPerspective
:
964 case SpvDecorationFlat
:
965 case SpvDecorationPatch
:
966 case SpvDecorationCentroid
:
967 case SpvDecorationSample
:
968 case SpvDecorationVolatile
:
969 case SpvDecorationCoherent
:
970 case SpvDecorationNonWritable
:
971 case SpvDecorationNonReadable
:
972 case SpvDecorationUniform
:
973 case SpvDecorationLocation
:
974 case SpvDecorationComponent
:
975 case SpvDecorationOffset
:
976 case SpvDecorationXfbBuffer
:
977 case SpvDecorationXfbStride
:
978 case SpvDecorationHlslSemanticGOOGLE
:
979 vtn_warn("Decoration only allowed for struct members: %s",
980 spirv_decoration_to_string(dec
->decoration
));
983 case SpvDecorationStream
:
984 /* We don't need to do anything here, as stream is filled up when
985 * aplying the decoration to a variable, just check that if it is not a
986 * struct member, it should be a struct.
988 vtn_assert(type
->base_type
== vtn_base_type_struct
);
991 case SpvDecorationRelaxedPrecision
:
992 case SpvDecorationSpecId
:
993 case SpvDecorationInvariant
:
994 case SpvDecorationRestrict
:
995 case SpvDecorationAliased
:
996 case SpvDecorationConstant
:
997 case SpvDecorationIndex
:
998 case SpvDecorationBinding
:
999 case SpvDecorationDescriptorSet
:
1000 case SpvDecorationLinkageAttributes
:
1001 case SpvDecorationNoContraction
:
1002 case SpvDecorationInputAttachmentIndex
:
1003 vtn_warn("Decoration not allowed on types: %s",
1004 spirv_decoration_to_string(dec
->decoration
));
1007 case SpvDecorationCPacked
:
1008 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
)
1009 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1010 spirv_decoration_to_string(dec
->decoration
));
1012 type
->packed
= true;
1015 case SpvDecorationSaturatedConversion
:
1016 case SpvDecorationFuncParamAttr
:
1017 case SpvDecorationFPRoundingMode
:
1018 case SpvDecorationFPFastMathMode
:
1019 case SpvDecorationAlignment
:
1020 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1021 spirv_decoration_to_string(dec
->decoration
));
1025 vtn_fail("Unhandled decoration");
1030 translate_image_format(struct vtn_builder
*b
, SpvImageFormat format
)
1033 case SpvImageFormatUnknown
: return 0; /* GL_NONE */
1034 case SpvImageFormatRgba32f
: return 0x8814; /* GL_RGBA32F */
1035 case SpvImageFormatRgba16f
: return 0x881A; /* GL_RGBA16F */
1036 case SpvImageFormatR32f
: return 0x822E; /* GL_R32F */
1037 case SpvImageFormatRgba8
: return 0x8058; /* GL_RGBA8 */
1038 case SpvImageFormatRgba8Snorm
: return 0x8F97; /* GL_RGBA8_SNORM */
1039 case SpvImageFormatRg32f
: return 0x8230; /* GL_RG32F */
1040 case SpvImageFormatRg16f
: return 0x822F; /* GL_RG16F */
1041 case SpvImageFormatR11fG11fB10f
: return 0x8C3A; /* GL_R11F_G11F_B10F */
1042 case SpvImageFormatR16f
: return 0x822D; /* GL_R16F */
1043 case SpvImageFormatRgba16
: return 0x805B; /* GL_RGBA16 */
1044 case SpvImageFormatRgb10A2
: return 0x8059; /* GL_RGB10_A2 */
1045 case SpvImageFormatRg16
: return 0x822C; /* GL_RG16 */
1046 case SpvImageFormatRg8
: return 0x822B; /* GL_RG8 */
1047 case SpvImageFormatR16
: return 0x822A; /* GL_R16 */
1048 case SpvImageFormatR8
: return 0x8229; /* GL_R8 */
1049 case SpvImageFormatRgba16Snorm
: return 0x8F9B; /* GL_RGBA16_SNORM */
1050 case SpvImageFormatRg16Snorm
: return 0x8F99; /* GL_RG16_SNORM */
1051 case SpvImageFormatRg8Snorm
: return 0x8F95; /* GL_RG8_SNORM */
1052 case SpvImageFormatR16Snorm
: return 0x8F98; /* GL_R16_SNORM */
1053 case SpvImageFormatR8Snorm
: return 0x8F94; /* GL_R8_SNORM */
1054 case SpvImageFormatRgba32i
: return 0x8D82; /* GL_RGBA32I */
1055 case SpvImageFormatRgba16i
: return 0x8D88; /* GL_RGBA16I */
1056 case SpvImageFormatRgba8i
: return 0x8D8E; /* GL_RGBA8I */
1057 case SpvImageFormatR32i
: return 0x8235; /* GL_R32I */
1058 case SpvImageFormatRg32i
: return 0x823B; /* GL_RG32I */
1059 case SpvImageFormatRg16i
: return 0x8239; /* GL_RG16I */
1060 case SpvImageFormatRg8i
: return 0x8237; /* GL_RG8I */
1061 case SpvImageFormatR16i
: return 0x8233; /* GL_R16I */
1062 case SpvImageFormatR8i
: return 0x8231; /* GL_R8I */
1063 case SpvImageFormatRgba32ui
: return 0x8D70; /* GL_RGBA32UI */
1064 case SpvImageFormatRgba16ui
: return 0x8D76; /* GL_RGBA16UI */
1065 case SpvImageFormatRgba8ui
: return 0x8D7C; /* GL_RGBA8UI */
1066 case SpvImageFormatR32ui
: return 0x8236; /* GL_R32UI */
1067 case SpvImageFormatRgb10a2ui
: return 0x906F; /* GL_RGB10_A2UI */
1068 case SpvImageFormatRg32ui
: return 0x823C; /* GL_RG32UI */
1069 case SpvImageFormatRg16ui
: return 0x823A; /* GL_RG16UI */
1070 case SpvImageFormatRg8ui
: return 0x8238; /* GL_RG8UI */
1071 case SpvImageFormatR16ui
: return 0x8234; /* GL_R16UI */
1072 case SpvImageFormatR8ui
: return 0x8232; /* GL_R8UI */
1074 vtn_fail("Invalid image format");
1078 static struct vtn_type
*
1079 vtn_type_layout_std430(struct vtn_builder
*b
, struct vtn_type
*type
,
1080 uint32_t *size_out
, uint32_t *align_out
)
1082 switch (type
->base_type
) {
1083 case vtn_base_type_scalar
: {
1084 uint32_t comp_size
= glsl_type_is_boolean(type
->type
)
1085 ? 4 : glsl_get_bit_size(type
->type
) / 8;
1086 *size_out
= comp_size
;
1087 *align_out
= comp_size
;
1091 case vtn_base_type_vector
: {
1092 uint32_t comp_size
= glsl_type_is_boolean(type
->type
)
1093 ? 4 : glsl_get_bit_size(type
->type
) / 8;
1094 unsigned align_comps
= type
->length
== 3 ? 4 : type
->length
;
1095 *size_out
= comp_size
* type
->length
,
1096 *align_out
= comp_size
* align_comps
;
1100 case vtn_base_type_matrix
:
1101 case vtn_base_type_array
: {
1102 /* We're going to add an array stride */
1103 type
= vtn_type_copy(b
, type
);
1104 uint32_t elem_size
, elem_align
;
1105 type
->array_element
= vtn_type_layout_std430(b
, type
->array_element
,
1106 &elem_size
, &elem_align
);
1107 type
->stride
= vtn_align_u32(elem_size
, elem_align
);
1108 *size_out
= type
->stride
* type
->length
;
1109 *align_out
= elem_align
;
1113 case vtn_base_type_struct
: {
1114 /* We're going to add member offsets */
1115 type
= vtn_type_copy(b
, type
);
1116 uint32_t offset
= 0;
1118 for (unsigned i
= 0; i
< type
->length
; i
++) {
1119 uint32_t mem_size
, mem_align
;
1120 type
->members
[i
] = vtn_type_layout_std430(b
, type
->members
[i
],
1121 &mem_size
, &mem_align
);
1122 offset
= vtn_align_u32(offset
, mem_align
);
1123 type
->offsets
[i
] = offset
;
1125 align
= MAX2(align
, mem_align
);
1133 unreachable("Invalid SPIR-V type for std430");
1138 vtn_handle_type(struct vtn_builder
*b
, SpvOp opcode
,
1139 const uint32_t *w
, unsigned count
)
1141 struct vtn_value
*val
= NULL
;
1143 /* In order to properly handle forward declarations, we have to defer
1144 * allocation for pointer types.
1146 if (opcode
!= SpvOpTypePointer
&& opcode
!= SpvOpTypeForwardPointer
) {
1147 val
= vtn_push_value(b
, w
[1], vtn_value_type_type
);
1148 vtn_fail_if(val
->type
!= NULL
,
1149 "Only pointers can have forward declarations");
1150 val
->type
= rzalloc(b
, struct vtn_type
);
1151 val
->type
->id
= w
[1];
1156 val
->type
->base_type
= vtn_base_type_void
;
1157 val
->type
->type
= glsl_void_type();
1160 val
->type
->base_type
= vtn_base_type_scalar
;
1161 val
->type
->type
= glsl_bool_type();
1162 val
->type
->length
= 1;
1164 case SpvOpTypeInt
: {
1165 int bit_size
= w
[2];
1166 const bool signedness
= w
[3];
1167 val
->type
->base_type
= vtn_base_type_scalar
;
1170 val
->type
->type
= (signedness
? glsl_int64_t_type() : glsl_uint64_t_type());
1173 val
->type
->type
= (signedness
? glsl_int_type() : glsl_uint_type());
1176 val
->type
->type
= (signedness
? glsl_int16_t_type() : glsl_uint16_t_type());
1179 val
->type
->type
= (signedness
? glsl_int8_t_type() : glsl_uint8_t_type());
1182 vtn_fail("Invalid int bit size");
1184 val
->type
->length
= 1;
1188 case SpvOpTypeFloat
: {
1189 int bit_size
= w
[2];
1190 val
->type
->base_type
= vtn_base_type_scalar
;
1193 val
->type
->type
= glsl_float16_t_type();
1196 val
->type
->type
= glsl_float_type();
1199 val
->type
->type
= glsl_double_type();
1202 vtn_fail("Invalid float bit size");
1204 val
->type
->length
= 1;
1208 case SpvOpTypeVector
: {
1209 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1210 unsigned elems
= w
[3];
1212 vtn_fail_if(base
->base_type
!= vtn_base_type_scalar
,
1213 "Base type for OpTypeVector must be a scalar");
1214 vtn_fail_if((elems
< 2 || elems
> 4) && (elems
!= 8) && (elems
!= 16),
1215 "Invalid component count for OpTypeVector");
1217 val
->type
->base_type
= vtn_base_type_vector
;
1218 val
->type
->type
= glsl_vector_type(glsl_get_base_type(base
->type
), elems
);
1219 val
->type
->length
= elems
;
1220 val
->type
->stride
= glsl_type_is_boolean(val
->type
->type
)
1221 ? 4 : glsl_get_bit_size(base
->type
) / 8;
1222 val
->type
->array_element
= base
;
1226 case SpvOpTypeMatrix
: {
1227 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1228 unsigned columns
= w
[3];
1230 vtn_fail_if(base
->base_type
!= vtn_base_type_vector
,
1231 "Base type for OpTypeMatrix must be a vector");
1232 vtn_fail_if(columns
< 2 || columns
> 4,
1233 "Invalid column count for OpTypeMatrix");
1235 val
->type
->base_type
= vtn_base_type_matrix
;
1236 val
->type
->type
= glsl_matrix_type(glsl_get_base_type(base
->type
),
1237 glsl_get_vector_elements(base
->type
),
1239 vtn_fail_if(glsl_type_is_error(val
->type
->type
),
1240 "Unsupported base type for OpTypeMatrix");
1241 assert(!glsl_type_is_error(val
->type
->type
));
1242 val
->type
->length
= columns
;
1243 val
->type
->array_element
= base
;
1244 val
->type
->row_major
= false;
1245 val
->type
->stride
= 0;
1249 case SpvOpTypeRuntimeArray
:
1250 case SpvOpTypeArray
: {
1251 struct vtn_type
*array_element
=
1252 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1254 if (opcode
== SpvOpTypeRuntimeArray
) {
1255 /* A length of 0 is used to denote unsized arrays */
1256 val
->type
->length
= 0;
1259 vtn_value(b
, w
[3], vtn_value_type_constant
)->constant
->values
[0][0].u32
;
1262 val
->type
->base_type
= vtn_base_type_array
;
1263 val
->type
->array_element
= array_element
;
1264 if (b
->shader
->info
.stage
== MESA_SHADER_KERNEL
)
1265 val
->type
->stride
= glsl_get_cl_size(array_element
->type
);
1267 vtn_foreach_decoration(b
, val
, array_stride_decoration_cb
, NULL
);
1268 val
->type
->type
= glsl_array_type(array_element
->type
, val
->type
->length
,
1273 case SpvOpTypeStruct
: {
1274 unsigned num_fields
= count
- 2;
1275 val
->type
->base_type
= vtn_base_type_struct
;
1276 val
->type
->length
= num_fields
;
1277 val
->type
->members
= ralloc_array(b
, struct vtn_type
*, num_fields
);
1278 val
->type
->offsets
= ralloc_array(b
, unsigned, num_fields
);
1279 val
->type
->packed
= false;
1281 NIR_VLA(struct glsl_struct_field
, fields
, count
);
1282 for (unsigned i
= 0; i
< num_fields
; i
++) {
1283 val
->type
->members
[i
] =
1284 vtn_value(b
, w
[i
+ 2], vtn_value_type_type
)->type
;
1285 fields
[i
] = (struct glsl_struct_field
) {
1286 .type
= val
->type
->members
[i
]->type
,
1287 .name
= ralloc_asprintf(b
, "field%d", i
),
1293 if (b
->shader
->info
.stage
== MESA_SHADER_KERNEL
) {
1294 unsigned offset
= 0;
1295 for (unsigned i
= 0; i
< num_fields
; i
++) {
1296 offset
= align(offset
, glsl_get_cl_alignment(fields
[i
].type
));
1297 fields
[i
].offset
= offset
;
1298 offset
+= glsl_get_cl_size(fields
[i
].type
);
1302 struct member_decoration_ctx ctx
= {
1303 .num_fields
= num_fields
,
1308 vtn_foreach_decoration(b
, val
, struct_member_decoration_cb
, &ctx
);
1309 vtn_foreach_decoration(b
, val
, struct_member_matrix_stride_cb
, &ctx
);
1311 vtn_foreach_decoration(b
, val
, struct_block_decoration_cb
, NULL
);
1313 const char *name
= val
->name
;
1315 if (val
->type
->block
|| val
->type
->buffer_block
) {
1316 /* Packing will be ignored since types coming from SPIR-V are
1317 * explicitly laid out.
1319 val
->type
->type
= glsl_interface_type(fields
, num_fields
,
1320 /* packing */ 0, false,
1321 name
? name
: "block");
1323 val
->type
->type
= glsl_struct_type(fields
, num_fields
,
1324 name
? name
: "struct", false);
1329 case SpvOpTypeFunction
: {
1330 val
->type
->base_type
= vtn_base_type_function
;
1331 val
->type
->type
= NULL
;
1333 val
->type
->return_type
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1335 const unsigned num_params
= count
- 3;
1336 val
->type
->length
= num_params
;
1337 val
->type
->params
= ralloc_array(b
, struct vtn_type
*, num_params
);
1338 for (unsigned i
= 0; i
< count
- 3; i
++) {
1339 val
->type
->params
[i
] =
1340 vtn_value(b
, w
[i
+ 3], vtn_value_type_type
)->type
;
1345 case SpvOpTypePointer
:
1346 case SpvOpTypeForwardPointer
: {
1347 /* We can't blindly push the value because it might be a forward
1350 val
= vtn_untyped_value(b
, w
[1]);
1352 SpvStorageClass storage_class
= w
[2];
1354 if (val
->value_type
== vtn_value_type_invalid
) {
1355 val
->value_type
= vtn_value_type_type
;
1356 val
->type
= rzalloc(b
, struct vtn_type
);
1357 val
->type
->id
= w
[1];
1358 val
->type
->base_type
= vtn_base_type_pointer
;
1359 val
->type
->storage_class
= storage_class
;
1361 /* These can actually be stored to nir_variables and used as SSA
1362 * values so they need a real glsl_type.
1364 switch (storage_class
) {
1365 case SpvStorageClassUniform
:
1366 val
->type
->type
= b
->options
->ubo_ptr_type
;
1368 case SpvStorageClassStorageBuffer
:
1369 val
->type
->type
= b
->options
->ssbo_ptr_type
;
1371 case SpvStorageClassPhysicalStorageBufferEXT
:
1372 val
->type
->type
= b
->options
->phys_ssbo_ptr_type
;
1374 case SpvStorageClassPushConstant
:
1375 val
->type
->type
= b
->options
->push_const_ptr_type
;
1377 case SpvStorageClassWorkgroup
:
1378 val
->type
->type
= b
->options
->shared_ptr_type
;
1380 case SpvStorageClassCrossWorkgroup
:
1381 val
->type
->type
= b
->options
->global_ptr_type
;
1383 case SpvStorageClassFunction
:
1384 if (b
->physical_ptrs
)
1385 val
->type
->type
= b
->options
->temp_ptr_type
;
1388 /* In this case, no variable pointers are allowed so all deref
1389 * chains are complete back to the variable and it doesn't matter
1390 * what type gets used so we leave it NULL.
1395 vtn_fail_if(val
->type
->storage_class
!= storage_class
,
1396 "The storage classes of an OpTypePointer and any "
1397 "OpTypeForwardPointers that provide forward "
1398 "declarations of it must match.");
1401 if (opcode
== SpvOpTypePointer
) {
1402 vtn_fail_if(val
->type
->deref
!= NULL
,
1403 "While OpTypeForwardPointer can be used to provide a "
1404 "forward declaration of a pointer, OpTypePointer can "
1405 "only be used once for a given id.");
1407 val
->type
->deref
= vtn_value(b
, w
[3], vtn_value_type_type
)->type
;
1409 vtn_foreach_decoration(b
, val
, array_stride_decoration_cb
, NULL
);
1411 if (b
->physical_ptrs
) {
1412 switch (storage_class
) {
1413 case SpvStorageClassFunction
:
1414 case SpvStorageClassWorkgroup
:
1415 case SpvStorageClassCrossWorkgroup
:
1416 val
->type
->stride
= align(glsl_get_cl_size(val
->type
->deref
->type
),
1417 glsl_get_cl_alignment(val
->type
->deref
->type
));
1424 if (storage_class
== SpvStorageClassWorkgroup
&&
1425 b
->options
->lower_workgroup_access_to_offsets
) {
1426 uint32_t size
, align
;
1427 val
->type
->deref
= vtn_type_layout_std430(b
, val
->type
->deref
,
1429 val
->type
->length
= size
;
1430 val
->type
->align
= align
;
1436 case SpvOpTypeImage
: {
1437 val
->type
->base_type
= vtn_base_type_image
;
1439 const struct vtn_type
*sampled_type
=
1440 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1442 vtn_fail_if(sampled_type
->base_type
!= vtn_base_type_scalar
||
1443 glsl_get_bit_size(sampled_type
->type
) != 32,
1444 "Sampled type of OpTypeImage must be a 32-bit scalar");
1446 enum glsl_sampler_dim dim
;
1447 switch ((SpvDim
)w
[3]) {
1448 case SpvDim1D
: dim
= GLSL_SAMPLER_DIM_1D
; break;
1449 case SpvDim2D
: dim
= GLSL_SAMPLER_DIM_2D
; break;
1450 case SpvDim3D
: dim
= GLSL_SAMPLER_DIM_3D
; break;
1451 case SpvDimCube
: dim
= GLSL_SAMPLER_DIM_CUBE
; break;
1452 case SpvDimRect
: dim
= GLSL_SAMPLER_DIM_RECT
; break;
1453 case SpvDimBuffer
: dim
= GLSL_SAMPLER_DIM_BUF
; break;
1454 case SpvDimSubpassData
: dim
= GLSL_SAMPLER_DIM_SUBPASS
; break;
1456 vtn_fail("Invalid SPIR-V image dimensionality");
1459 /* w[4]: as per Vulkan spec "Validation Rules within a Module",
1460 * The “Depth” operand of OpTypeImage is ignored.
1462 bool is_array
= w
[5];
1463 bool multisampled
= w
[6];
1464 unsigned sampled
= w
[7];
1465 SpvImageFormat format
= w
[8];
1468 val
->type
->access_qualifier
= w
[9];
1470 val
->type
->access_qualifier
= SpvAccessQualifierReadWrite
;
1473 if (dim
== GLSL_SAMPLER_DIM_2D
)
1474 dim
= GLSL_SAMPLER_DIM_MS
;
1475 else if (dim
== GLSL_SAMPLER_DIM_SUBPASS
)
1476 dim
= GLSL_SAMPLER_DIM_SUBPASS_MS
;
1478 vtn_fail("Unsupported multisampled image type");
1481 val
->type
->image_format
= translate_image_format(b
, format
);
1483 enum glsl_base_type sampled_base_type
=
1484 glsl_get_base_type(sampled_type
->type
);
1486 val
->type
->sampled
= true;
1487 val
->type
->type
= glsl_sampler_type(dim
, false, is_array
,
1489 } else if (sampled
== 2) {
1490 val
->type
->sampled
= false;
1491 val
->type
->type
= glsl_image_type(dim
, is_array
, sampled_base_type
);
1493 vtn_fail("We need to know if the image will be sampled");
1498 case SpvOpTypeSampledImage
:
1499 val
->type
->base_type
= vtn_base_type_sampled_image
;
1500 val
->type
->image
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1501 val
->type
->type
= val
->type
->image
->type
;
1504 case SpvOpTypeSampler
:
1505 /* The actual sampler type here doesn't really matter. It gets
1506 * thrown away the moment you combine it with an image. What really
1507 * matters is that it's a sampler type as opposed to an integer type
1508 * so the backend knows what to do.
1510 val
->type
->base_type
= vtn_base_type_sampler
;
1511 val
->type
->type
= glsl_bare_sampler_type();
1514 case SpvOpTypeOpaque
:
1515 case SpvOpTypeEvent
:
1516 case SpvOpTypeDeviceEvent
:
1517 case SpvOpTypeReserveId
:
1518 case SpvOpTypeQueue
:
1521 vtn_fail("Unhandled opcode");
1524 vtn_foreach_decoration(b
, val
, type_decoration_cb
, NULL
);
1526 if (val
->type
->base_type
== vtn_base_type_struct
&&
1527 (val
->type
->block
|| val
->type
->buffer_block
)) {
1528 for (unsigned i
= 0; i
< val
->type
->length
; i
++) {
1529 vtn_fail_if(vtn_type_contains_block(b
, val
->type
->members
[i
]),
1530 "Block and BufferBlock decorations cannot decorate a "
1531 "structure type that is nested at any level inside "
1532 "another structure type decorated with Block or "
1538 static nir_constant
*
1539 vtn_null_constant(struct vtn_builder
*b
, const struct glsl_type
*type
)
1541 nir_constant
*c
= rzalloc(b
, nir_constant
);
1543 /* For pointers and other typeless things, we have to return something but
1544 * it doesn't matter what.
1549 switch (glsl_get_base_type(type
)) {
1551 case GLSL_TYPE_UINT
:
1552 case GLSL_TYPE_INT16
:
1553 case GLSL_TYPE_UINT16
:
1554 case GLSL_TYPE_UINT8
:
1555 case GLSL_TYPE_INT8
:
1556 case GLSL_TYPE_INT64
:
1557 case GLSL_TYPE_UINT64
:
1558 case GLSL_TYPE_BOOL
:
1559 case GLSL_TYPE_FLOAT
:
1560 case GLSL_TYPE_FLOAT16
:
1561 case GLSL_TYPE_DOUBLE
:
1562 /* Nothing to do here. It's already initialized to zero */
1565 case GLSL_TYPE_ARRAY
:
1566 vtn_assert(glsl_get_length(type
) > 0);
1567 c
->num_elements
= glsl_get_length(type
);
1568 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1570 c
->elements
[0] = vtn_null_constant(b
, glsl_get_array_element(type
));
1571 for (unsigned i
= 1; i
< c
->num_elements
; i
++)
1572 c
->elements
[i
] = c
->elements
[0];
1575 case GLSL_TYPE_STRUCT
:
1576 c
->num_elements
= glsl_get_length(type
);
1577 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1579 for (unsigned i
= 0; i
< c
->num_elements
; i
++) {
1580 c
->elements
[i
] = vtn_null_constant(b
, glsl_get_struct_field(type
, i
));
1585 vtn_fail("Invalid type for null constant");
1592 spec_constant_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*v
,
1593 int member
, const struct vtn_decoration
*dec
,
1596 vtn_assert(member
== -1);
1597 if (dec
->decoration
!= SpvDecorationSpecId
)
1600 struct spec_constant_value
*const_value
= data
;
1602 for (unsigned i
= 0; i
< b
->num_specializations
; i
++) {
1603 if (b
->specializations
[i
].id
== dec
->literals
[0]) {
1604 if (const_value
->is_double
)
1605 const_value
->data64
= b
->specializations
[i
].data64
;
1607 const_value
->data32
= b
->specializations
[i
].data32
;
1614 get_specialization(struct vtn_builder
*b
, struct vtn_value
*val
,
1615 uint32_t const_value
)
1617 struct spec_constant_value data
;
1618 data
.is_double
= false;
1619 data
.data32
= const_value
;
1620 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1625 get_specialization64(struct vtn_builder
*b
, struct vtn_value
*val
,
1626 uint64_t const_value
)
1628 struct spec_constant_value data
;
1629 data
.is_double
= true;
1630 data
.data64
= const_value
;
1631 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1636 handle_workgroup_size_decoration_cb(struct vtn_builder
*b
,
1637 struct vtn_value
*val
,
1639 const struct vtn_decoration
*dec
,
1642 vtn_assert(member
== -1);
1643 if (dec
->decoration
!= SpvDecorationBuiltIn
||
1644 dec
->literals
[0] != SpvBuiltInWorkgroupSize
)
1647 vtn_assert(val
->type
->type
== glsl_vector_type(GLSL_TYPE_UINT
, 3));
1648 b
->workgroup_size_builtin
= val
;
1652 vtn_handle_constant(struct vtn_builder
*b
, SpvOp opcode
,
1653 const uint32_t *w
, unsigned count
)
1655 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_constant
);
1656 val
->constant
= rzalloc(b
, nir_constant
);
1658 case SpvOpConstantTrue
:
1659 case SpvOpConstantFalse
:
1660 case SpvOpSpecConstantTrue
:
1661 case SpvOpSpecConstantFalse
: {
1662 vtn_fail_if(val
->type
->type
!= glsl_bool_type(),
1663 "Result type of %s must be OpTypeBool",
1664 spirv_op_to_string(opcode
));
1666 uint32_t int_val
= (opcode
== SpvOpConstantTrue
||
1667 opcode
== SpvOpSpecConstantTrue
);
1669 if (opcode
== SpvOpSpecConstantTrue
||
1670 opcode
== SpvOpSpecConstantFalse
)
1671 int_val
= get_specialization(b
, val
, int_val
);
1673 val
->constant
->values
[0][0].b
= int_val
!= 0;
1677 case SpvOpConstant
: {
1678 vtn_fail_if(val
->type
->base_type
!= vtn_base_type_scalar
,
1679 "Result type of %s must be a scalar",
1680 spirv_op_to_string(opcode
));
1681 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1684 val
->constant
->values
[0][0].u64
= vtn_u64_literal(&w
[3]);
1687 val
->constant
->values
[0][0].u32
= w
[3];
1690 val
->constant
->values
[0][0].u16
= w
[3];
1693 val
->constant
->values
[0][0].u8
= w
[3];
1696 vtn_fail("Unsupported SpvOpConstant bit size");
1701 case SpvOpSpecConstant
: {
1702 vtn_fail_if(val
->type
->base_type
!= vtn_base_type_scalar
,
1703 "Result type of %s must be a scalar",
1704 spirv_op_to_string(opcode
));
1705 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1708 val
->constant
->values
[0][0].u64
=
1709 get_specialization64(b
, val
, vtn_u64_literal(&w
[3]));
1712 val
->constant
->values
[0][0].u32
= get_specialization(b
, val
, w
[3]);
1715 val
->constant
->values
[0][0].u16
= get_specialization(b
, val
, w
[3]);
1718 val
->constant
->values
[0][0].u8
= get_specialization(b
, val
, w
[3]);
1721 vtn_fail("Unsupported SpvOpSpecConstant bit size");
1726 case SpvOpSpecConstantComposite
:
1727 case SpvOpConstantComposite
: {
1728 unsigned elem_count
= count
- 3;
1729 vtn_fail_if(elem_count
!= val
->type
->length
,
1730 "%s has %u constituents, expected %u",
1731 spirv_op_to_string(opcode
), elem_count
, val
->type
->length
);
1733 nir_constant
**elems
= ralloc_array(b
, nir_constant
*, elem_count
);
1734 for (unsigned i
= 0; i
< elem_count
; i
++) {
1735 struct vtn_value
*val
= vtn_untyped_value(b
, w
[i
+ 3]);
1737 if (val
->value_type
== vtn_value_type_constant
) {
1738 elems
[i
] = val
->constant
;
1740 vtn_fail_if(val
->value_type
!= vtn_value_type_undef
,
1741 "only constants or undefs allowed for "
1742 "SpvOpConstantComposite");
1743 /* to make it easier, just insert a NULL constant for now */
1744 elems
[i
] = vtn_null_constant(b
, val
->type
->type
);
1748 switch (val
->type
->base_type
) {
1749 case vtn_base_type_vector
: {
1750 assert(glsl_type_is_vector(val
->type
->type
));
1751 for (unsigned i
= 0; i
< elem_count
; i
++)
1752 val
->constant
->values
[0][i
] = elems
[i
]->values
[0][0];
1756 case vtn_base_type_matrix
:
1757 assert(glsl_type_is_matrix(val
->type
->type
));
1758 for (unsigned i
= 0; i
< elem_count
; i
++) {
1759 unsigned components
=
1760 glsl_get_components(glsl_get_column_type(val
->type
->type
));
1761 memcpy(val
->constant
->values
[i
], elems
[i
]->values
,
1762 sizeof(nir_const_value
) * components
);
1766 case vtn_base_type_struct
:
1767 case vtn_base_type_array
:
1768 ralloc_steal(val
->constant
, elems
);
1769 val
->constant
->num_elements
= elem_count
;
1770 val
->constant
->elements
= elems
;
1774 vtn_fail("Result type of %s must be a composite type",
1775 spirv_op_to_string(opcode
));
1780 case SpvOpSpecConstantOp
: {
1781 SpvOp opcode
= get_specialization(b
, val
, w
[3]);
1783 case SpvOpVectorShuffle
: {
1784 struct vtn_value
*v0
= &b
->values
[w
[4]];
1785 struct vtn_value
*v1
= &b
->values
[w
[5]];
1787 vtn_assert(v0
->value_type
== vtn_value_type_constant
||
1788 v0
->value_type
== vtn_value_type_undef
);
1789 vtn_assert(v1
->value_type
== vtn_value_type_constant
||
1790 v1
->value_type
== vtn_value_type_undef
);
1792 unsigned len0
= glsl_get_vector_elements(v0
->type
->type
);
1793 unsigned len1
= glsl_get_vector_elements(v1
->type
->type
);
1795 vtn_assert(len0
+ len1
< 16);
1797 unsigned bit_size
= glsl_get_bit_size(val
->type
->type
);
1798 unsigned bit_size0
= glsl_get_bit_size(v0
->type
->type
);
1799 unsigned bit_size1
= glsl_get_bit_size(v1
->type
->type
);
1801 vtn_assert(bit_size
== bit_size0
&& bit_size
== bit_size1
);
1802 (void)bit_size0
; (void)bit_size1
;
1804 if (bit_size
== 64) {
1806 if (v0
->value_type
== vtn_value_type_constant
) {
1807 for (unsigned i
= 0; i
< len0
; i
++)
1808 u64
[i
] = v0
->constant
->values
[0][i
].u64
;
1810 if (v1
->value_type
== vtn_value_type_constant
) {
1811 for (unsigned i
= 0; i
< len1
; i
++)
1812 u64
[len0
+ i
] = v1
->constant
->values
[0][i
].u64
;
1815 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1816 uint32_t comp
= w
[i
+ 6];
1817 /* If component is not used, set the value to a known constant
1818 * to detect if it is wrongly used.
1820 if (comp
== (uint32_t)-1)
1821 val
->constant
->values
[0][j
].u64
= 0xdeadbeefdeadbeef;
1823 val
->constant
->values
[0][j
].u64
= u64
[comp
];
1826 /* This is for both 32-bit and 16-bit values */
1828 if (v0
->value_type
== vtn_value_type_constant
) {
1829 for (unsigned i
= 0; i
< len0
; i
++)
1830 u32
[i
] = v0
->constant
->values
[0][i
].u32
;
1832 if (v1
->value_type
== vtn_value_type_constant
) {
1833 for (unsigned i
= 0; i
< len1
; i
++)
1834 u32
[len0
+ i
] = v1
->constant
->values
[0][i
].u32
;
1837 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1838 uint32_t comp
= w
[i
+ 6];
1839 /* If component is not used, set the value to a known constant
1840 * to detect if it is wrongly used.
1842 if (comp
== (uint32_t)-1)
1843 val
->constant
->values
[0][j
].u32
= 0xdeadbeef;
1845 val
->constant
->values
[0][j
].u32
= u32
[comp
];
1851 case SpvOpCompositeExtract
:
1852 case SpvOpCompositeInsert
: {
1853 struct vtn_value
*comp
;
1854 unsigned deref_start
;
1855 struct nir_constant
**c
;
1856 if (opcode
== SpvOpCompositeExtract
) {
1857 comp
= vtn_value(b
, w
[4], vtn_value_type_constant
);
1859 c
= &comp
->constant
;
1861 comp
= vtn_value(b
, w
[5], vtn_value_type_constant
);
1863 val
->constant
= nir_constant_clone(comp
->constant
,
1870 const struct vtn_type
*type
= comp
->type
;
1871 for (unsigned i
= deref_start
; i
< count
; i
++) {
1872 vtn_fail_if(w
[i
] > type
->length
,
1873 "%uth index of %s is %u but the type has only "
1874 "%u elements", i
- deref_start
,
1875 spirv_op_to_string(opcode
), w
[i
], type
->length
);
1877 switch (type
->base_type
) {
1878 case vtn_base_type_vector
:
1880 type
= type
->array_element
;
1883 case vtn_base_type_matrix
:
1884 assert(col
== 0 && elem
== -1);
1887 type
= type
->array_element
;
1890 case vtn_base_type_array
:
1891 c
= &(*c
)->elements
[w
[i
]];
1892 type
= type
->array_element
;
1895 case vtn_base_type_struct
:
1896 c
= &(*c
)->elements
[w
[i
]];
1897 type
= type
->members
[w
[i
]];
1901 vtn_fail("%s must only index into composite types",
1902 spirv_op_to_string(opcode
));
1906 if (opcode
== SpvOpCompositeExtract
) {
1910 unsigned num_components
= type
->length
;
1911 for (unsigned i
= 0; i
< num_components
; i
++)
1912 val
->constant
->values
[0][i
] = (*c
)->values
[col
][elem
+ i
];
1915 struct vtn_value
*insert
=
1916 vtn_value(b
, w
[4], vtn_value_type_constant
);
1917 vtn_assert(insert
->type
== type
);
1919 *c
= insert
->constant
;
1921 unsigned num_components
= type
->length
;
1922 for (unsigned i
= 0; i
< num_components
; i
++)
1923 (*c
)->values
[col
][elem
+ i
] = insert
->constant
->values
[0][i
];
1931 nir_alu_type dst_alu_type
= nir_get_nir_type_for_glsl_type(val
->type
->type
);
1932 nir_alu_type src_alu_type
= dst_alu_type
;
1933 unsigned num_components
= glsl_get_vector_elements(val
->type
->type
);
1936 vtn_assert(count
<= 7);
1941 /* We have a source in a conversion */
1943 nir_get_nir_type_for_glsl_type(
1944 vtn_value(b
, w
[4], vtn_value_type_constant
)->type
->type
);
1945 /* We use the bitsize of the conversion source to evaluate the opcode later */
1946 bit_size
= glsl_get_bit_size(
1947 vtn_value(b
, w
[4], vtn_value_type_constant
)->type
->type
);
1950 bit_size
= glsl_get_bit_size(val
->type
->type
);
1953 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(b
, opcode
, &swap
,
1954 nir_alu_type_get_type_size(src_alu_type
),
1955 nir_alu_type_get_type_size(dst_alu_type
));
1956 nir_const_value src
[3][NIR_MAX_VEC_COMPONENTS
];
1958 for (unsigned i
= 0; i
< count
- 4; i
++) {
1959 struct vtn_value
*src_val
=
1960 vtn_value(b
, w
[4 + i
], vtn_value_type_constant
);
1962 /* If this is an unsized source, pull the bit size from the
1963 * source; otherwise, we'll use the bit size from the destination.
1965 if (!nir_alu_type_get_type_size(nir_op_infos
[op
].input_types
[i
]))
1966 bit_size
= glsl_get_bit_size(src_val
->type
->type
);
1968 unsigned j
= swap
? 1 - i
: i
;
1969 memcpy(src
[j
], src_val
->constant
->values
[0], sizeof(src
[j
]));
1972 /* fix up fixed size sources */
1979 for (unsigned i
= 0; i
< num_components
; ++i
) {
1981 case 64: src
[1][i
].u32
= src
[1][i
].u64
; break;
1982 case 16: src
[1][i
].u32
= src
[1][i
].u16
; break;
1983 case 8: src
[1][i
].u32
= src
[1][i
].u8
; break;
1992 nir_const_value
*srcs
[3] = {
1993 src
[0], src
[1], src
[2],
1995 nir_eval_const_opcode(op
, val
->constant
->values
[0], num_components
, bit_size
, srcs
);
2002 case SpvOpConstantNull
:
2003 val
->constant
= vtn_null_constant(b
, val
->type
->type
);
2006 case SpvOpConstantSampler
:
2007 vtn_fail("OpConstantSampler requires Kernel Capability");
2011 vtn_fail("Unhandled opcode");
2014 /* Now that we have the value, update the workgroup size if needed */
2015 vtn_foreach_decoration(b
, val
, handle_workgroup_size_decoration_cb
, NULL
);
2018 struct vtn_ssa_value
*
2019 vtn_create_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
2021 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
2024 if (!glsl_type_is_vector_or_scalar(type
)) {
2025 unsigned elems
= glsl_get_length(type
);
2026 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
2027 for (unsigned i
= 0; i
< elems
; i
++) {
2028 const struct glsl_type
*child_type
;
2030 switch (glsl_get_base_type(type
)) {
2032 case GLSL_TYPE_UINT
:
2033 case GLSL_TYPE_INT16
:
2034 case GLSL_TYPE_UINT16
:
2035 case GLSL_TYPE_UINT8
:
2036 case GLSL_TYPE_INT8
:
2037 case GLSL_TYPE_INT64
:
2038 case GLSL_TYPE_UINT64
:
2039 case GLSL_TYPE_BOOL
:
2040 case GLSL_TYPE_FLOAT
:
2041 case GLSL_TYPE_FLOAT16
:
2042 case GLSL_TYPE_DOUBLE
:
2043 child_type
= glsl_get_column_type(type
);
2045 case GLSL_TYPE_ARRAY
:
2046 child_type
= glsl_get_array_element(type
);
2048 case GLSL_TYPE_STRUCT
:
2049 case GLSL_TYPE_INTERFACE
:
2050 child_type
= glsl_get_struct_field(type
, i
);
2053 vtn_fail("unkown base type");
2056 val
->elems
[i
] = vtn_create_ssa_value(b
, child_type
);
2064 vtn_tex_src(struct vtn_builder
*b
, unsigned index
, nir_tex_src_type type
)
2067 src
.src
= nir_src_for_ssa(vtn_ssa_value(b
, index
)->def
);
2068 src
.src_type
= type
;
2073 vtn_handle_texture(struct vtn_builder
*b
, SpvOp opcode
,
2074 const uint32_t *w
, unsigned count
)
2076 if (opcode
== SpvOpSampledImage
) {
2077 struct vtn_value
*val
=
2078 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
2079 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
2080 val
->sampled_image
->type
=
2081 vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2082 val
->sampled_image
->image
=
2083 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2084 val
->sampled_image
->sampler
=
2085 vtn_value(b
, w
[4], vtn_value_type_pointer
)->pointer
;
2087 } else if (opcode
== SpvOpImage
) {
2088 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
2089 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[3]);
2090 if (src_val
->value_type
== vtn_value_type_sampled_image
) {
2091 val
->pointer
= src_val
->sampled_image
->image
;
2093 vtn_assert(src_val
->value_type
== vtn_value_type_pointer
);
2094 val
->pointer
= src_val
->pointer
;
2099 struct vtn_type
*ret_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2100 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2102 struct vtn_sampled_image sampled
;
2103 struct vtn_value
*sampled_val
= vtn_untyped_value(b
, w
[3]);
2104 if (sampled_val
->value_type
== vtn_value_type_sampled_image
) {
2105 sampled
= *sampled_val
->sampled_image
;
2107 vtn_assert(sampled_val
->value_type
== vtn_value_type_pointer
);
2108 sampled
.type
= sampled_val
->pointer
->type
;
2109 sampled
.image
= NULL
;
2110 sampled
.sampler
= sampled_val
->pointer
;
2113 const struct glsl_type
*image_type
= sampled
.type
->type
;
2114 const enum glsl_sampler_dim sampler_dim
= glsl_get_sampler_dim(image_type
);
2115 const bool is_array
= glsl_sampler_type_is_array(image_type
);
2117 /* Figure out the base texture operation */
2120 case SpvOpImageSampleImplicitLod
:
2121 case SpvOpImageSampleDrefImplicitLod
:
2122 case SpvOpImageSampleProjImplicitLod
:
2123 case SpvOpImageSampleProjDrefImplicitLod
:
2124 texop
= nir_texop_tex
;
2127 case SpvOpImageSampleExplicitLod
:
2128 case SpvOpImageSampleDrefExplicitLod
:
2129 case SpvOpImageSampleProjExplicitLod
:
2130 case SpvOpImageSampleProjDrefExplicitLod
:
2131 texop
= nir_texop_txl
;
2134 case SpvOpImageFetch
:
2135 if (glsl_get_sampler_dim(image_type
) == GLSL_SAMPLER_DIM_MS
) {
2136 texop
= nir_texop_txf_ms
;
2138 texop
= nir_texop_txf
;
2142 case SpvOpImageGather
:
2143 case SpvOpImageDrefGather
:
2144 texop
= nir_texop_tg4
;
2147 case SpvOpImageQuerySizeLod
:
2148 case SpvOpImageQuerySize
:
2149 texop
= nir_texop_txs
;
2152 case SpvOpImageQueryLod
:
2153 texop
= nir_texop_lod
;
2156 case SpvOpImageQueryLevels
:
2157 texop
= nir_texop_query_levels
;
2160 case SpvOpImageQuerySamples
:
2161 texop
= nir_texop_texture_samples
;
2165 vtn_fail("Unhandled opcode");
2168 nir_tex_src srcs
[10]; /* 10 should be enough */
2169 nir_tex_src
*p
= srcs
;
2171 nir_deref_instr
*sampler
= vtn_pointer_to_deref(b
, sampled
.sampler
);
2172 nir_deref_instr
*texture
=
2173 sampled
.image
? vtn_pointer_to_deref(b
, sampled
.image
) : sampler
;
2175 p
->src
= nir_src_for_ssa(&texture
->dest
.ssa
);
2176 p
->src_type
= nir_tex_src_texture_deref
;
2186 /* These operations require a sampler */
2187 p
->src
= nir_src_for_ssa(&sampler
->dest
.ssa
);
2188 p
->src_type
= nir_tex_src_sampler_deref
;
2192 case nir_texop_txf_ms
:
2194 case nir_texop_query_levels
:
2195 case nir_texop_texture_samples
:
2196 case nir_texop_samples_identical
:
2199 case nir_texop_txf_ms_mcs
:
2200 vtn_fail("unexpected nir_texop_txf_ms_mcs");
2205 struct nir_ssa_def
*coord
;
2206 unsigned coord_components
;
2208 case SpvOpImageSampleImplicitLod
:
2209 case SpvOpImageSampleExplicitLod
:
2210 case SpvOpImageSampleDrefImplicitLod
:
2211 case SpvOpImageSampleDrefExplicitLod
:
2212 case SpvOpImageSampleProjImplicitLod
:
2213 case SpvOpImageSampleProjExplicitLod
:
2214 case SpvOpImageSampleProjDrefImplicitLod
:
2215 case SpvOpImageSampleProjDrefExplicitLod
:
2216 case SpvOpImageFetch
:
2217 case SpvOpImageGather
:
2218 case SpvOpImageDrefGather
:
2219 case SpvOpImageQueryLod
: {
2220 /* All these types have the coordinate as their first real argument */
2221 switch (sampler_dim
) {
2222 case GLSL_SAMPLER_DIM_1D
:
2223 case GLSL_SAMPLER_DIM_BUF
:
2224 coord_components
= 1;
2226 case GLSL_SAMPLER_DIM_2D
:
2227 case GLSL_SAMPLER_DIM_RECT
:
2228 case GLSL_SAMPLER_DIM_MS
:
2229 coord_components
= 2;
2231 case GLSL_SAMPLER_DIM_3D
:
2232 case GLSL_SAMPLER_DIM_CUBE
:
2233 coord_components
= 3;
2236 vtn_fail("Invalid sampler type");
2239 if (is_array
&& texop
!= nir_texop_lod
)
2242 coord
= vtn_ssa_value(b
, w
[idx
++])->def
;
2243 p
->src
= nir_src_for_ssa(nir_channels(&b
->nb
, coord
,
2244 (1 << coord_components
) - 1));
2245 p
->src_type
= nir_tex_src_coord
;
2252 coord_components
= 0;
2257 case SpvOpImageSampleProjImplicitLod
:
2258 case SpvOpImageSampleProjExplicitLod
:
2259 case SpvOpImageSampleProjDrefImplicitLod
:
2260 case SpvOpImageSampleProjDrefExplicitLod
:
2261 /* These have the projector as the last coordinate component */
2262 p
->src
= nir_src_for_ssa(nir_channel(&b
->nb
, coord
, coord_components
));
2263 p
->src_type
= nir_tex_src_projector
;
2271 bool is_shadow
= false;
2272 unsigned gather_component
= 0;
2274 case SpvOpImageSampleDrefImplicitLod
:
2275 case SpvOpImageSampleDrefExplicitLod
:
2276 case SpvOpImageSampleProjDrefImplicitLod
:
2277 case SpvOpImageSampleProjDrefExplicitLod
:
2278 case SpvOpImageDrefGather
:
2279 /* These all have an explicit depth value as their next source */
2281 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_comparator
);
2284 case SpvOpImageGather
:
2285 /* This has a component as its next source */
2287 vtn_value(b
, w
[idx
++], vtn_value_type_constant
)->constant
->values
[0][0].u32
;
2294 /* For OpImageQuerySizeLod, we always have an LOD */
2295 if (opcode
== SpvOpImageQuerySizeLod
)
2296 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
2298 /* Now we need to handle some number of optional arguments */
2299 struct vtn_value
*gather_offsets
= NULL
;
2301 uint32_t operands
= w
[idx
++];
2303 if (operands
& SpvImageOperandsBiasMask
) {
2304 vtn_assert(texop
== nir_texop_tex
);
2305 texop
= nir_texop_txb
;
2306 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_bias
);
2309 if (operands
& SpvImageOperandsLodMask
) {
2310 vtn_assert(texop
== nir_texop_txl
|| texop
== nir_texop_txf
||
2311 texop
== nir_texop_txs
);
2312 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
2315 if (operands
& SpvImageOperandsGradMask
) {
2316 vtn_assert(texop
== nir_texop_txl
);
2317 texop
= nir_texop_txd
;
2318 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddx
);
2319 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddy
);
2322 if (operands
& SpvImageOperandsOffsetMask
||
2323 operands
& SpvImageOperandsConstOffsetMask
)
2324 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_offset
);
2326 if (operands
& SpvImageOperandsConstOffsetsMask
) {
2327 vtn_assert(texop
== nir_texop_tg4
);
2328 gather_offsets
= vtn_value(b
, w
[idx
++], vtn_value_type_constant
);
2331 if (operands
& SpvImageOperandsSampleMask
) {
2332 vtn_assert(texop
== nir_texop_txf_ms
);
2333 texop
= nir_texop_txf_ms
;
2334 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ms_index
);
2337 if (operands
& SpvImageOperandsMinLodMask
) {
2338 vtn_assert(texop
== nir_texop_tex
||
2339 texop
== nir_texop_txb
||
2340 texop
== nir_texop_txd
);
2341 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_min_lod
);
2344 /* We should have now consumed exactly all of the arguments */
2345 vtn_assert(idx
== count
);
2347 nir_tex_instr
*instr
= nir_tex_instr_create(b
->shader
, p
- srcs
);
2350 memcpy(instr
->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
2352 instr
->coord_components
= coord_components
;
2353 instr
->sampler_dim
= sampler_dim
;
2354 instr
->is_array
= is_array
;
2355 instr
->is_shadow
= is_shadow
;
2356 instr
->is_new_style_shadow
=
2357 is_shadow
&& glsl_get_components(ret_type
->type
) == 1;
2358 instr
->component
= gather_component
;
2360 if (sampled
.image
&& (sampled
.image
->access
& ACCESS_NON_UNIFORM
))
2361 instr
->texture_non_uniform
= true;
2363 if (sampled
.sampler
&& (sampled
.sampler
->access
& ACCESS_NON_UNIFORM
))
2364 instr
->sampler_non_uniform
= true;
2366 switch (glsl_get_sampler_result_type(image_type
)) {
2367 case GLSL_TYPE_FLOAT
: instr
->dest_type
= nir_type_float
; break;
2368 case GLSL_TYPE_INT
: instr
->dest_type
= nir_type_int
; break;
2369 case GLSL_TYPE_UINT
: instr
->dest_type
= nir_type_uint
; break;
2370 case GLSL_TYPE_BOOL
: instr
->dest_type
= nir_type_bool
; break;
2372 vtn_fail("Invalid base type for sampler result");
2375 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
2376 nir_tex_instr_dest_size(instr
), 32, NULL
);
2378 vtn_assert(glsl_get_vector_elements(ret_type
->type
) ==
2379 nir_tex_instr_dest_size(instr
));
2381 if (gather_offsets
) {
2382 vtn_fail_if(gather_offsets
->type
->base_type
!= vtn_base_type_array
||
2383 gather_offsets
->type
->length
!= 4,
2384 "ConstOffsets must be an array of size four of vectors "
2385 "of two integer components");
2387 struct vtn_type
*vec_type
= gather_offsets
->type
->array_element
;
2388 vtn_fail_if(vec_type
->base_type
!= vtn_base_type_vector
||
2389 vec_type
->length
!= 2 ||
2390 !glsl_type_is_integer(vec_type
->type
),
2391 "ConstOffsets must be an array of size four of vectors "
2392 "of two integer components");
2394 unsigned bit_size
= glsl_get_bit_size(vec_type
->type
);
2395 for (uint32_t i
= 0; i
< 4; i
++) {
2396 const nir_const_value
*cvec
=
2397 gather_offsets
->constant
->elements
[i
]->values
[0];
2398 for (uint32_t j
= 0; j
< 2; j
++) {
2400 case 8: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i8
; break;
2401 case 16: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i16
; break;
2402 case 32: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i32
; break;
2403 case 64: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i64
; break;
2405 vtn_fail("Unsupported bit size");
2411 val
->ssa
= vtn_create_ssa_value(b
, ret_type
->type
);
2412 val
->ssa
->def
= &instr
->dest
.ssa
;
2414 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
2418 fill_common_atomic_sources(struct vtn_builder
*b
, SpvOp opcode
,
2419 const uint32_t *w
, nir_src
*src
)
2422 case SpvOpAtomicIIncrement
:
2423 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
2426 case SpvOpAtomicIDecrement
:
2427 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
2430 case SpvOpAtomicISub
:
2432 nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
2435 case SpvOpAtomicCompareExchange
:
2436 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[8])->def
);
2437 src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
2440 case SpvOpAtomicExchange
:
2441 case SpvOpAtomicIAdd
:
2442 case SpvOpAtomicSMin
:
2443 case SpvOpAtomicUMin
:
2444 case SpvOpAtomicSMax
:
2445 case SpvOpAtomicUMax
:
2446 case SpvOpAtomicAnd
:
2448 case SpvOpAtomicXor
:
2449 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
2453 vtn_fail("Invalid SPIR-V atomic");
2457 static nir_ssa_def
*
2458 get_image_coord(struct vtn_builder
*b
, uint32_t value
)
2460 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, value
);
2462 /* The image_load_store intrinsics assume a 4-dim coordinate */
2463 unsigned dim
= glsl_get_vector_elements(coord
->type
);
2464 unsigned swizzle
[4];
2465 for (unsigned i
= 0; i
< 4; i
++)
2466 swizzle
[i
] = MIN2(i
, dim
- 1);
2468 return nir_swizzle(&b
->nb
, coord
->def
, swizzle
, 4, false);
2471 static nir_ssa_def
*
2472 expand_to_vec4(nir_builder
*b
, nir_ssa_def
*value
)
2474 if (value
->num_components
== 4)
2478 for (unsigned i
= 0; i
< 4; i
++)
2479 swiz
[i
] = i
< value
->num_components
? i
: 0;
2480 return nir_swizzle(b
, value
, swiz
, 4, false);
2484 vtn_handle_image(struct vtn_builder
*b
, SpvOp opcode
,
2485 const uint32_t *w
, unsigned count
)
2487 /* Just get this one out of the way */
2488 if (opcode
== SpvOpImageTexelPointer
) {
2489 struct vtn_value
*val
=
2490 vtn_push_value(b
, w
[2], vtn_value_type_image_pointer
);
2491 val
->image
= ralloc(b
, struct vtn_image_pointer
);
2493 val
->image
->image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2494 val
->image
->coord
= get_image_coord(b
, w
[4]);
2495 val
->image
->sample
= vtn_ssa_value(b
, w
[5])->def
;
2499 struct vtn_image_pointer image
;
2502 case SpvOpAtomicExchange
:
2503 case SpvOpAtomicCompareExchange
:
2504 case SpvOpAtomicCompareExchangeWeak
:
2505 case SpvOpAtomicIIncrement
:
2506 case SpvOpAtomicIDecrement
:
2507 case SpvOpAtomicIAdd
:
2508 case SpvOpAtomicISub
:
2509 case SpvOpAtomicLoad
:
2510 case SpvOpAtomicSMin
:
2511 case SpvOpAtomicUMin
:
2512 case SpvOpAtomicSMax
:
2513 case SpvOpAtomicUMax
:
2514 case SpvOpAtomicAnd
:
2516 case SpvOpAtomicXor
:
2517 image
= *vtn_value(b
, w
[3], vtn_value_type_image_pointer
)->image
;
2520 case SpvOpAtomicStore
:
2521 image
= *vtn_value(b
, w
[1], vtn_value_type_image_pointer
)->image
;
2524 case SpvOpImageQuerySize
:
2525 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2527 image
.sample
= NULL
;
2530 case SpvOpImageRead
:
2531 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2532 image
.coord
= get_image_coord(b
, w
[4]);
2534 if (count
> 5 && (w
[5] & SpvImageOperandsSampleMask
)) {
2535 vtn_assert(w
[5] == SpvImageOperandsSampleMask
);
2536 image
.sample
= vtn_ssa_value(b
, w
[6])->def
;
2538 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2542 case SpvOpImageWrite
:
2543 image
.image
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
2544 image
.coord
= get_image_coord(b
, w
[2]);
2548 if (count
> 4 && (w
[4] & SpvImageOperandsSampleMask
)) {
2549 vtn_assert(w
[4] == SpvImageOperandsSampleMask
);
2550 image
.sample
= vtn_ssa_value(b
, w
[5])->def
;
2552 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2557 vtn_fail("Invalid image opcode");
2560 nir_intrinsic_op op
;
2562 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_deref_##N; break;
2563 OP(ImageQuerySize
, size
)
2565 OP(ImageWrite
, store
)
2566 OP(AtomicLoad
, load
)
2567 OP(AtomicStore
, store
)
2568 OP(AtomicExchange
, atomic_exchange
)
2569 OP(AtomicCompareExchange
, atomic_comp_swap
)
2570 OP(AtomicIIncrement
, atomic_add
)
2571 OP(AtomicIDecrement
, atomic_add
)
2572 OP(AtomicIAdd
, atomic_add
)
2573 OP(AtomicISub
, atomic_add
)
2574 OP(AtomicSMin
, atomic_min
)
2575 OP(AtomicUMin
, atomic_min
)
2576 OP(AtomicSMax
, atomic_max
)
2577 OP(AtomicUMax
, atomic_max
)
2578 OP(AtomicAnd
, atomic_and
)
2579 OP(AtomicOr
, atomic_or
)
2580 OP(AtomicXor
, atomic_xor
)
2583 vtn_fail("Invalid image opcode");
2586 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
2588 nir_deref_instr
*image_deref
= vtn_pointer_to_deref(b
, image
.image
);
2589 intrin
->src
[0] = nir_src_for_ssa(&image_deref
->dest
.ssa
);
2591 /* ImageQuerySize doesn't take any extra parameters */
2592 if (opcode
!= SpvOpImageQuerySize
) {
2593 /* The image coordinate is always 4 components but we may not have that
2594 * many. Swizzle to compensate.
2596 intrin
->src
[1] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, image
.coord
));
2597 intrin
->src
[2] = nir_src_for_ssa(image
.sample
);
2601 case SpvOpAtomicLoad
:
2602 case SpvOpImageQuerySize
:
2603 case SpvOpImageRead
:
2605 case SpvOpAtomicStore
:
2606 case SpvOpImageWrite
: {
2607 const uint32_t value_id
= opcode
== SpvOpAtomicStore
? w
[4] : w
[3];
2608 nir_ssa_def
*value
= vtn_ssa_value(b
, value_id
)->def
;
2609 /* nir_intrinsic_image_deref_store always takes a vec4 value */
2610 assert(op
== nir_intrinsic_image_deref_store
);
2611 intrin
->num_components
= 4;
2612 intrin
->src
[3] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, value
));
2616 case SpvOpAtomicCompareExchange
:
2617 case SpvOpAtomicIIncrement
:
2618 case SpvOpAtomicIDecrement
:
2619 case SpvOpAtomicExchange
:
2620 case SpvOpAtomicIAdd
:
2621 case SpvOpAtomicISub
:
2622 case SpvOpAtomicSMin
:
2623 case SpvOpAtomicUMin
:
2624 case SpvOpAtomicSMax
:
2625 case SpvOpAtomicUMax
:
2626 case SpvOpAtomicAnd
:
2628 case SpvOpAtomicXor
:
2629 fill_common_atomic_sources(b
, opcode
, w
, &intrin
->src
[3]);
2633 vtn_fail("Invalid image opcode");
2636 if (opcode
!= SpvOpImageWrite
&& opcode
!= SpvOpAtomicStore
) {
2637 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2638 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2640 unsigned dest_components
= glsl_get_vector_elements(type
->type
);
2641 intrin
->num_components
= nir_intrinsic_infos
[op
].dest_components
;
2642 if (intrin
->num_components
== 0)
2643 intrin
->num_components
= dest_components
;
2645 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
,
2646 intrin
->num_components
, 32, NULL
);
2648 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2650 nir_ssa_def
*result
= &intrin
->dest
.ssa
;
2651 if (intrin
->num_components
!= dest_components
)
2652 result
= nir_channels(&b
->nb
, result
, (1 << dest_components
) - 1);
2654 val
->ssa
= vtn_create_ssa_value(b
, type
->type
);
2655 val
->ssa
->def
= result
;
2657 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2661 static nir_intrinsic_op
2662 get_ssbo_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2665 case SpvOpAtomicLoad
: return nir_intrinsic_load_ssbo
;
2666 case SpvOpAtomicStore
: return nir_intrinsic_store_ssbo
;
2667 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
2668 OP(AtomicExchange
, atomic_exchange
)
2669 OP(AtomicCompareExchange
, atomic_comp_swap
)
2670 OP(AtomicIIncrement
, atomic_add
)
2671 OP(AtomicIDecrement
, atomic_add
)
2672 OP(AtomicIAdd
, atomic_add
)
2673 OP(AtomicISub
, atomic_add
)
2674 OP(AtomicSMin
, atomic_imin
)
2675 OP(AtomicUMin
, atomic_umin
)
2676 OP(AtomicSMax
, atomic_imax
)
2677 OP(AtomicUMax
, atomic_umax
)
2678 OP(AtomicAnd
, atomic_and
)
2679 OP(AtomicOr
, atomic_or
)
2680 OP(AtomicXor
, atomic_xor
)
2683 vtn_fail("Invalid SSBO atomic");
2687 static nir_intrinsic_op
2688 get_uniform_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2691 #define OP(S, N) case SpvOp##S: return nir_intrinsic_atomic_counter_ ##N;
2692 OP(AtomicLoad
, read_deref
)
2693 OP(AtomicExchange
, exchange
)
2694 OP(AtomicCompareExchange
, comp_swap
)
2695 OP(AtomicIIncrement
, inc_deref
)
2696 OP(AtomicIDecrement
, post_dec_deref
)
2697 OP(AtomicIAdd
, add_deref
)
2698 OP(AtomicISub
, add_deref
)
2699 OP(AtomicUMin
, min_deref
)
2700 OP(AtomicUMax
, max_deref
)
2701 OP(AtomicAnd
, and_deref
)
2702 OP(AtomicOr
, or_deref
)
2703 OP(AtomicXor
, xor_deref
)
2706 /* We left the following out: AtomicStore, AtomicSMin and
2707 * AtomicSmax. Right now there are not nir intrinsics for them. At this
2708 * moment Atomic Counter support is needed for ARB_spirv support, so is
2709 * only need to support GLSL Atomic Counters that are uints and don't
2710 * allow direct storage.
2712 unreachable("Invalid uniform atomic");
2716 static nir_intrinsic_op
2717 get_shared_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2720 case SpvOpAtomicLoad
: return nir_intrinsic_load_shared
;
2721 case SpvOpAtomicStore
: return nir_intrinsic_store_shared
;
2722 #define OP(S, N) case SpvOp##S: return nir_intrinsic_shared_##N;
2723 OP(AtomicExchange
, atomic_exchange
)
2724 OP(AtomicCompareExchange
, atomic_comp_swap
)
2725 OP(AtomicIIncrement
, atomic_add
)
2726 OP(AtomicIDecrement
, atomic_add
)
2727 OP(AtomicIAdd
, atomic_add
)
2728 OP(AtomicISub
, atomic_add
)
2729 OP(AtomicSMin
, atomic_imin
)
2730 OP(AtomicUMin
, atomic_umin
)
2731 OP(AtomicSMax
, atomic_imax
)
2732 OP(AtomicUMax
, atomic_umax
)
2733 OP(AtomicAnd
, atomic_and
)
2734 OP(AtomicOr
, atomic_or
)
2735 OP(AtomicXor
, atomic_xor
)
2738 vtn_fail("Invalid shared atomic");
2742 static nir_intrinsic_op
2743 get_deref_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2746 case SpvOpAtomicLoad
: return nir_intrinsic_load_deref
;
2747 case SpvOpAtomicStore
: return nir_intrinsic_store_deref
;
2748 #define OP(S, N) case SpvOp##S: return nir_intrinsic_deref_##N;
2749 OP(AtomicExchange
, atomic_exchange
)
2750 OP(AtomicCompareExchange
, atomic_comp_swap
)
2751 OP(AtomicIIncrement
, atomic_add
)
2752 OP(AtomicIDecrement
, atomic_add
)
2753 OP(AtomicIAdd
, atomic_add
)
2754 OP(AtomicISub
, atomic_add
)
2755 OP(AtomicSMin
, atomic_imin
)
2756 OP(AtomicUMin
, atomic_umin
)
2757 OP(AtomicSMax
, atomic_imax
)
2758 OP(AtomicUMax
, atomic_umax
)
2759 OP(AtomicAnd
, atomic_and
)
2760 OP(AtomicOr
, atomic_or
)
2761 OP(AtomicXor
, atomic_xor
)
2764 vtn_fail("Invalid shared atomic");
2769 * Handles shared atomics, ssbo atomics and atomic counters.
2772 vtn_handle_atomics(struct vtn_builder
*b
, SpvOp opcode
,
2773 const uint32_t *w
, unsigned count
)
2775 struct vtn_pointer
*ptr
;
2776 nir_intrinsic_instr
*atomic
;
2779 case SpvOpAtomicLoad
:
2780 case SpvOpAtomicExchange
:
2781 case SpvOpAtomicCompareExchange
:
2782 case SpvOpAtomicCompareExchangeWeak
:
2783 case SpvOpAtomicIIncrement
:
2784 case SpvOpAtomicIDecrement
:
2785 case SpvOpAtomicIAdd
:
2786 case SpvOpAtomicISub
:
2787 case SpvOpAtomicSMin
:
2788 case SpvOpAtomicUMin
:
2789 case SpvOpAtomicSMax
:
2790 case SpvOpAtomicUMax
:
2791 case SpvOpAtomicAnd
:
2793 case SpvOpAtomicXor
:
2794 ptr
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2797 case SpvOpAtomicStore
:
2798 ptr
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
2802 vtn_fail("Invalid SPIR-V atomic");
2806 SpvScope scope = w[4];
2807 SpvMemorySemanticsMask semantics = w[5];
2810 /* uniform as "atomic counter uniform" */
2811 if (ptr
->mode
== vtn_variable_mode_uniform
) {
2812 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
2813 const struct glsl_type
*deref_type
= deref
->type
;
2814 nir_intrinsic_op op
= get_uniform_nir_atomic_op(b
, opcode
);
2815 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2816 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
2818 /* SSBO needs to initialize index/offset. In this case we don't need to,
2819 * as that info is already stored on the ptr->var->var nir_variable (see
2820 * vtn_create_variable)
2824 case SpvOpAtomicLoad
:
2825 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2828 case SpvOpAtomicStore
:
2829 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2830 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2833 case SpvOpAtomicExchange
:
2834 case SpvOpAtomicCompareExchange
:
2835 case SpvOpAtomicCompareExchangeWeak
:
2836 case SpvOpAtomicIIncrement
:
2837 case SpvOpAtomicIDecrement
:
2838 case SpvOpAtomicIAdd
:
2839 case SpvOpAtomicISub
:
2840 case SpvOpAtomicSMin
:
2841 case SpvOpAtomicUMin
:
2842 case SpvOpAtomicSMax
:
2843 case SpvOpAtomicUMax
:
2844 case SpvOpAtomicAnd
:
2846 case SpvOpAtomicXor
:
2847 /* Nothing: we don't need to call fill_common_atomic_sources here, as
2848 * atomic counter uniforms doesn't have sources
2853 unreachable("Invalid SPIR-V atomic");
2856 } else if (vtn_pointer_uses_ssa_offset(b
, ptr
)) {
2857 nir_ssa_def
*offset
, *index
;
2858 offset
= vtn_pointer_to_offset(b
, ptr
, &index
);
2860 nir_intrinsic_op op
;
2861 if (ptr
->mode
== vtn_variable_mode_ssbo
) {
2862 op
= get_ssbo_nir_atomic_op(b
, opcode
);
2864 vtn_assert(ptr
->mode
== vtn_variable_mode_workgroup
&&
2865 b
->options
->lower_workgroup_access_to_offsets
);
2866 op
= get_shared_nir_atomic_op(b
, opcode
);
2869 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2873 case SpvOpAtomicLoad
:
2874 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
2875 nir_intrinsic_set_align(atomic
, 4, 0);
2876 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2877 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2878 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2881 case SpvOpAtomicStore
:
2882 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
2883 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2884 nir_intrinsic_set_align(atomic
, 4, 0);
2885 atomic
->src
[src
++] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2886 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2887 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2888 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2891 case SpvOpAtomicExchange
:
2892 case SpvOpAtomicCompareExchange
:
2893 case SpvOpAtomicCompareExchangeWeak
:
2894 case SpvOpAtomicIIncrement
:
2895 case SpvOpAtomicIDecrement
:
2896 case SpvOpAtomicIAdd
:
2897 case SpvOpAtomicISub
:
2898 case SpvOpAtomicSMin
:
2899 case SpvOpAtomicUMin
:
2900 case SpvOpAtomicSMax
:
2901 case SpvOpAtomicUMax
:
2902 case SpvOpAtomicAnd
:
2904 case SpvOpAtomicXor
:
2905 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2906 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2907 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2908 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[src
]);
2912 vtn_fail("Invalid SPIR-V atomic");
2915 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
2916 const struct glsl_type
*deref_type
= deref
->type
;
2917 nir_intrinsic_op op
= get_deref_nir_atomic_op(b
, opcode
);
2918 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2919 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
2922 case SpvOpAtomicLoad
:
2923 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2926 case SpvOpAtomicStore
:
2927 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2928 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2929 atomic
->src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2932 case SpvOpAtomicExchange
:
2933 case SpvOpAtomicCompareExchange
:
2934 case SpvOpAtomicCompareExchangeWeak
:
2935 case SpvOpAtomicIIncrement
:
2936 case SpvOpAtomicIDecrement
:
2937 case SpvOpAtomicIAdd
:
2938 case SpvOpAtomicISub
:
2939 case SpvOpAtomicSMin
:
2940 case SpvOpAtomicUMin
:
2941 case SpvOpAtomicSMax
:
2942 case SpvOpAtomicUMax
:
2943 case SpvOpAtomicAnd
:
2945 case SpvOpAtomicXor
:
2946 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[1]);
2950 vtn_fail("Invalid SPIR-V atomic");
2954 if (opcode
!= SpvOpAtomicStore
) {
2955 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2957 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
,
2958 glsl_get_vector_elements(type
->type
),
2959 glsl_get_bit_size(type
->type
), NULL
);
2961 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2962 val
->ssa
= rzalloc(b
, struct vtn_ssa_value
);
2963 val
->ssa
->def
= &atomic
->dest
.ssa
;
2964 val
->ssa
->type
= type
->type
;
2967 nir_builder_instr_insert(&b
->nb
, &atomic
->instr
);
2970 static nir_alu_instr
*
2971 create_vec(struct vtn_builder
*b
, unsigned num_components
, unsigned bit_size
)
2974 switch (num_components
) {
2975 case 1: op
= nir_op_imov
; break;
2976 case 2: op
= nir_op_vec2
; break;
2977 case 3: op
= nir_op_vec3
; break;
2978 case 4: op
= nir_op_vec4
; break;
2979 default: vtn_fail("bad vector size");
2982 nir_alu_instr
*vec
= nir_alu_instr_create(b
->shader
, op
);
2983 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, num_components
,
2985 vec
->dest
.write_mask
= (1 << num_components
) - 1;
2990 struct vtn_ssa_value
*
2991 vtn_ssa_transpose(struct vtn_builder
*b
, struct vtn_ssa_value
*src
)
2993 if (src
->transposed
)
2994 return src
->transposed
;
2996 struct vtn_ssa_value
*dest
=
2997 vtn_create_ssa_value(b
, glsl_transposed_type(src
->type
));
2999 for (unsigned i
= 0; i
< glsl_get_matrix_columns(dest
->type
); i
++) {
3000 nir_alu_instr
*vec
= create_vec(b
, glsl_get_matrix_columns(src
->type
),
3001 glsl_get_bit_size(src
->type
));
3002 if (glsl_type_is_vector_or_scalar(src
->type
)) {
3003 vec
->src
[0].src
= nir_src_for_ssa(src
->def
);
3004 vec
->src
[0].swizzle
[0] = i
;
3006 for (unsigned j
= 0; j
< glsl_get_matrix_columns(src
->type
); j
++) {
3007 vec
->src
[j
].src
= nir_src_for_ssa(src
->elems
[j
]->def
);
3008 vec
->src
[j
].swizzle
[0] = i
;
3011 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3012 dest
->elems
[i
]->def
= &vec
->dest
.dest
.ssa
;
3015 dest
->transposed
= src
;
3021 vtn_vector_extract(struct vtn_builder
*b
, nir_ssa_def
*src
, unsigned index
)
3023 return nir_channel(&b
->nb
, src
, index
);
3027 vtn_vector_insert(struct vtn_builder
*b
, nir_ssa_def
*src
, nir_ssa_def
*insert
,
3030 nir_alu_instr
*vec
= create_vec(b
, src
->num_components
,
3033 for (unsigned i
= 0; i
< src
->num_components
; i
++) {
3035 vec
->src
[i
].src
= nir_src_for_ssa(insert
);
3037 vec
->src
[i
].src
= nir_src_for_ssa(src
);
3038 vec
->src
[i
].swizzle
[0] = i
;
3042 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3044 return &vec
->dest
.dest
.ssa
;
3047 static nir_ssa_def
*
3048 nir_ieq_imm(nir_builder
*b
, nir_ssa_def
*x
, uint64_t i
)
3050 return nir_ieq(b
, x
, nir_imm_intN_t(b
, i
, x
->bit_size
));
3054 vtn_vector_extract_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
3057 return nir_vector_extract(&b
->nb
, src
, nir_i2i(&b
->nb
, index
, 32));
3061 vtn_vector_insert_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
3062 nir_ssa_def
*insert
, nir_ssa_def
*index
)
3064 nir_ssa_def
*dest
= vtn_vector_insert(b
, src
, insert
, 0);
3065 for (unsigned i
= 1; i
< src
->num_components
; i
++)
3066 dest
= nir_bcsel(&b
->nb
, nir_ieq_imm(&b
->nb
, index
, i
),
3067 vtn_vector_insert(b
, src
, insert
, i
), dest
);
3072 static nir_ssa_def
*
3073 vtn_vector_shuffle(struct vtn_builder
*b
, unsigned num_components
,
3074 nir_ssa_def
*src0
, nir_ssa_def
*src1
,
3075 const uint32_t *indices
)
3077 nir_alu_instr
*vec
= create_vec(b
, num_components
, src0
->bit_size
);
3079 for (unsigned i
= 0; i
< num_components
; i
++) {
3080 uint32_t index
= indices
[i
];
3081 if (index
== 0xffffffff) {
3083 nir_src_for_ssa(nir_ssa_undef(&b
->nb
, 1, src0
->bit_size
));
3084 } else if (index
< src0
->num_components
) {
3085 vec
->src
[i
].src
= nir_src_for_ssa(src0
);
3086 vec
->src
[i
].swizzle
[0] = index
;
3088 vec
->src
[i
].src
= nir_src_for_ssa(src1
);
3089 vec
->src
[i
].swizzle
[0] = index
- src0
->num_components
;
3093 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3095 return &vec
->dest
.dest
.ssa
;
3099 * Concatentates a number of vectors/scalars together to produce a vector
3101 static nir_ssa_def
*
3102 vtn_vector_construct(struct vtn_builder
*b
, unsigned num_components
,
3103 unsigned num_srcs
, nir_ssa_def
**srcs
)
3105 nir_alu_instr
*vec
= create_vec(b
, num_components
, srcs
[0]->bit_size
);
3107 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
3109 * "When constructing a vector, there must be at least two Constituent
3112 vtn_assert(num_srcs
>= 2);
3114 unsigned dest_idx
= 0;
3115 for (unsigned i
= 0; i
< num_srcs
; i
++) {
3116 nir_ssa_def
*src
= srcs
[i
];
3117 vtn_assert(dest_idx
+ src
->num_components
<= num_components
);
3118 for (unsigned j
= 0; j
< src
->num_components
; j
++) {
3119 vec
->src
[dest_idx
].src
= nir_src_for_ssa(src
);
3120 vec
->src
[dest_idx
].swizzle
[0] = j
;
3125 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
3127 * "When constructing a vector, the total number of components in all
3128 * the operands must equal the number of components in Result Type."
3130 vtn_assert(dest_idx
== num_components
);
3132 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3134 return &vec
->dest
.dest
.ssa
;
3137 static struct vtn_ssa_value
*
3138 vtn_composite_copy(void *mem_ctx
, struct vtn_ssa_value
*src
)
3140 struct vtn_ssa_value
*dest
= rzalloc(mem_ctx
, struct vtn_ssa_value
);
3141 dest
->type
= src
->type
;
3143 if (glsl_type_is_vector_or_scalar(src
->type
)) {
3144 dest
->def
= src
->def
;
3146 unsigned elems
= glsl_get_length(src
->type
);
3148 dest
->elems
= ralloc_array(mem_ctx
, struct vtn_ssa_value
*, elems
);
3149 for (unsigned i
= 0; i
< elems
; i
++)
3150 dest
->elems
[i
] = vtn_composite_copy(mem_ctx
, src
->elems
[i
]);
3156 static struct vtn_ssa_value
*
3157 vtn_composite_insert(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
3158 struct vtn_ssa_value
*insert
, const uint32_t *indices
,
3159 unsigned num_indices
)
3161 struct vtn_ssa_value
*dest
= vtn_composite_copy(b
, src
);
3163 struct vtn_ssa_value
*cur
= dest
;
3165 for (i
= 0; i
< num_indices
- 1; i
++) {
3166 cur
= cur
->elems
[indices
[i
]];
3169 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
3170 /* According to the SPIR-V spec, OpCompositeInsert may work down to
3171 * the component granularity. In that case, the last index will be
3172 * the index to insert the scalar into the vector.
3175 cur
->def
= vtn_vector_insert(b
, cur
->def
, insert
->def
, indices
[i
]);
3177 cur
->elems
[indices
[i
]] = insert
;
3183 static struct vtn_ssa_value
*
3184 vtn_composite_extract(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
3185 const uint32_t *indices
, unsigned num_indices
)
3187 struct vtn_ssa_value
*cur
= src
;
3188 for (unsigned i
= 0; i
< num_indices
; i
++) {
3189 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
3190 vtn_assert(i
== num_indices
- 1);
3191 /* According to the SPIR-V spec, OpCompositeExtract may work down to
3192 * the component granularity. The last index will be the index of the
3193 * vector to extract.
3196 struct vtn_ssa_value
*ret
= rzalloc(b
, struct vtn_ssa_value
);
3197 ret
->type
= glsl_scalar_type(glsl_get_base_type(cur
->type
));
3198 ret
->def
= vtn_vector_extract(b
, cur
->def
, indices
[i
]);
3201 cur
= cur
->elems
[indices
[i
]];
3209 vtn_handle_composite(struct vtn_builder
*b
, SpvOp opcode
,
3210 const uint32_t *w
, unsigned count
)
3212 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
3213 const struct glsl_type
*type
=
3214 vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
3215 val
->ssa
= vtn_create_ssa_value(b
, type
);
3218 case SpvOpVectorExtractDynamic
:
3219 val
->ssa
->def
= vtn_vector_extract_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
3220 vtn_ssa_value(b
, w
[4])->def
);
3223 case SpvOpVectorInsertDynamic
:
3224 val
->ssa
->def
= vtn_vector_insert_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
3225 vtn_ssa_value(b
, w
[4])->def
,
3226 vtn_ssa_value(b
, w
[5])->def
);
3229 case SpvOpVectorShuffle
:
3230 val
->ssa
->def
= vtn_vector_shuffle(b
, glsl_get_vector_elements(type
),
3231 vtn_ssa_value(b
, w
[3])->def
,
3232 vtn_ssa_value(b
, w
[4])->def
,
3236 case SpvOpCompositeConstruct
: {
3237 unsigned elems
= count
- 3;
3239 if (glsl_type_is_vector_or_scalar(type
)) {
3240 nir_ssa_def
*srcs
[NIR_MAX_VEC_COMPONENTS
];
3241 for (unsigned i
= 0; i
< elems
; i
++)
3242 srcs
[i
] = vtn_ssa_value(b
, w
[3 + i
])->def
;
3244 vtn_vector_construct(b
, glsl_get_vector_elements(type
),
3247 val
->ssa
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
3248 for (unsigned i
= 0; i
< elems
; i
++)
3249 val
->ssa
->elems
[i
] = vtn_ssa_value(b
, w
[3 + i
]);
3253 case SpvOpCompositeExtract
:
3254 val
->ssa
= vtn_composite_extract(b
, vtn_ssa_value(b
, w
[3]),
3258 case SpvOpCompositeInsert
:
3259 val
->ssa
= vtn_composite_insert(b
, vtn_ssa_value(b
, w
[4]),
3260 vtn_ssa_value(b
, w
[3]),
3264 case SpvOpCopyObject
:
3265 val
->ssa
= vtn_composite_copy(b
, vtn_ssa_value(b
, w
[3]));
3269 vtn_fail("unknown composite operation");
3274 vtn_emit_barrier(struct vtn_builder
*b
, nir_intrinsic_op op
)
3276 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
3277 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3281 vtn_emit_memory_barrier(struct vtn_builder
*b
, SpvScope scope
,
3282 SpvMemorySemanticsMask semantics
)
3284 static const SpvMemorySemanticsMask all_memory_semantics
=
3285 SpvMemorySemanticsUniformMemoryMask
|
3286 SpvMemorySemanticsWorkgroupMemoryMask
|
3287 SpvMemorySemanticsAtomicCounterMemoryMask
|
3288 SpvMemorySemanticsImageMemoryMask
;
3290 /* If we're not actually doing a memory barrier, bail */
3291 if (!(semantics
& all_memory_semantics
))
3294 /* GL and Vulkan don't have these */
3295 vtn_assert(scope
!= SpvScopeCrossDevice
);
3297 if (scope
== SpvScopeSubgroup
)
3298 return; /* Nothing to do here */
3300 if (scope
== SpvScopeWorkgroup
) {
3301 vtn_emit_barrier(b
, nir_intrinsic_group_memory_barrier
);
3305 /* There's only two scopes thing left */
3306 vtn_assert(scope
== SpvScopeInvocation
|| scope
== SpvScopeDevice
);
3308 if ((semantics
& all_memory_semantics
) == all_memory_semantics
) {
3309 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier
);
3313 /* Issue a bunch of more specific barriers */
3314 uint32_t bits
= semantics
;
3316 SpvMemorySemanticsMask semantic
= 1 << u_bit_scan(&bits
);
3318 case SpvMemorySemanticsUniformMemoryMask
:
3319 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_buffer
);
3321 case SpvMemorySemanticsWorkgroupMemoryMask
:
3322 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_shared
);
3324 case SpvMemorySemanticsAtomicCounterMemoryMask
:
3325 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_atomic_counter
);
3327 case SpvMemorySemanticsImageMemoryMask
:
3328 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_image
);
3337 vtn_handle_barrier(struct vtn_builder
*b
, SpvOp opcode
,
3338 const uint32_t *w
, unsigned count
)
3341 case SpvOpEmitVertex
:
3342 case SpvOpEmitStreamVertex
:
3343 case SpvOpEndPrimitive
:
3344 case SpvOpEndStreamPrimitive
: {
3345 nir_intrinsic_op intrinsic_op
;
3347 case SpvOpEmitVertex
:
3348 case SpvOpEmitStreamVertex
:
3349 intrinsic_op
= nir_intrinsic_emit_vertex
;
3351 case SpvOpEndPrimitive
:
3352 case SpvOpEndStreamPrimitive
:
3353 intrinsic_op
= nir_intrinsic_end_primitive
;
3356 unreachable("Invalid opcode");
3359 nir_intrinsic_instr
*intrin
=
3360 nir_intrinsic_instr_create(b
->shader
, intrinsic_op
);
3363 case SpvOpEmitStreamVertex
:
3364 case SpvOpEndStreamPrimitive
: {
3365 unsigned stream
= vtn_constant_uint(b
, w
[1]);
3366 nir_intrinsic_set_stream_id(intrin
, stream
);
3374 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3378 case SpvOpMemoryBarrier
: {
3379 SpvScope scope
= vtn_constant_uint(b
, w
[1]);
3380 SpvMemorySemanticsMask semantics
= vtn_constant_uint(b
, w
[2]);
3381 vtn_emit_memory_barrier(b
, scope
, semantics
);
3385 case SpvOpControlBarrier
: {
3386 SpvScope execution_scope
= vtn_constant_uint(b
, w
[1]);
3387 if (execution_scope
== SpvScopeWorkgroup
)
3388 vtn_emit_barrier(b
, nir_intrinsic_barrier
);
3390 SpvScope memory_scope
= vtn_constant_uint(b
, w
[2]);
3391 SpvMemorySemanticsMask memory_semantics
= vtn_constant_uint(b
, w
[3]);
3392 vtn_emit_memory_barrier(b
, memory_scope
, memory_semantics
);
3397 unreachable("unknown barrier instruction");
3402 gl_primitive_from_spv_execution_mode(struct vtn_builder
*b
,
3403 SpvExecutionMode mode
)
3406 case SpvExecutionModeInputPoints
:
3407 case SpvExecutionModeOutputPoints
:
3408 return 0; /* GL_POINTS */
3409 case SpvExecutionModeInputLines
:
3410 return 1; /* GL_LINES */
3411 case SpvExecutionModeInputLinesAdjacency
:
3412 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
3413 case SpvExecutionModeTriangles
:
3414 return 4; /* GL_TRIANGLES */
3415 case SpvExecutionModeInputTrianglesAdjacency
:
3416 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
3417 case SpvExecutionModeQuads
:
3418 return 7; /* GL_QUADS */
3419 case SpvExecutionModeIsolines
:
3420 return 0x8E7A; /* GL_ISOLINES */
3421 case SpvExecutionModeOutputLineStrip
:
3422 return 3; /* GL_LINE_STRIP */
3423 case SpvExecutionModeOutputTriangleStrip
:
3424 return 5; /* GL_TRIANGLE_STRIP */
3426 vtn_fail("Invalid primitive type");
3431 vertices_in_from_spv_execution_mode(struct vtn_builder
*b
,
3432 SpvExecutionMode mode
)
3435 case SpvExecutionModeInputPoints
:
3437 case SpvExecutionModeInputLines
:
3439 case SpvExecutionModeInputLinesAdjacency
:
3441 case SpvExecutionModeTriangles
:
3443 case SpvExecutionModeInputTrianglesAdjacency
:
3446 vtn_fail("Invalid GS input mode");
3450 static gl_shader_stage
3451 stage_for_execution_model(struct vtn_builder
*b
, SpvExecutionModel model
)
3454 case SpvExecutionModelVertex
:
3455 return MESA_SHADER_VERTEX
;
3456 case SpvExecutionModelTessellationControl
:
3457 return MESA_SHADER_TESS_CTRL
;
3458 case SpvExecutionModelTessellationEvaluation
:
3459 return MESA_SHADER_TESS_EVAL
;
3460 case SpvExecutionModelGeometry
:
3461 return MESA_SHADER_GEOMETRY
;
3462 case SpvExecutionModelFragment
:
3463 return MESA_SHADER_FRAGMENT
;
3464 case SpvExecutionModelGLCompute
:
3465 return MESA_SHADER_COMPUTE
;
3466 case SpvExecutionModelKernel
:
3467 return MESA_SHADER_KERNEL
;
3469 vtn_fail("Unsupported execution model");
3473 #define spv_check_supported(name, cap) do { \
3474 if (!(b->options && b->options->caps.name)) \
3475 vtn_warn("Unsupported SPIR-V capability: %s", \
3476 spirv_capability_to_string(cap)); \
3481 vtn_handle_entry_point(struct vtn_builder
*b
, const uint32_t *w
,
3484 struct vtn_value
*entry_point
= &b
->values
[w
[2]];
3485 /* Let this be a name label regardless */
3486 unsigned name_words
;
3487 entry_point
->name
= vtn_string_literal(b
, &w
[3], count
- 3, &name_words
);
3489 if (strcmp(entry_point
->name
, b
->entry_point_name
) != 0 ||
3490 stage_for_execution_model(b
, w
[1]) != b
->entry_point_stage
)
3493 vtn_assert(b
->entry_point
== NULL
);
3494 b
->entry_point
= entry_point
;
3498 vtn_handle_preamble_instruction(struct vtn_builder
*b
, SpvOp opcode
,
3499 const uint32_t *w
, unsigned count
)
3506 case SpvSourceLanguageUnknown
: lang
= "unknown"; break;
3507 case SpvSourceLanguageESSL
: lang
= "ESSL"; break;
3508 case SpvSourceLanguageGLSL
: lang
= "GLSL"; break;
3509 case SpvSourceLanguageOpenCL_C
: lang
= "OpenCL C"; break;
3510 case SpvSourceLanguageOpenCL_CPP
: lang
= "OpenCL C++"; break;
3511 case SpvSourceLanguageHLSL
: lang
= "HLSL"; break;
3514 uint32_t version
= w
[2];
3517 (count
> 3) ? vtn_value(b
, w
[3], vtn_value_type_string
)->str
: "";
3519 vtn_info("Parsing SPIR-V from %s %u source file %s", lang
, version
, file
);
3523 case SpvOpSourceExtension
:
3524 case SpvOpSourceContinued
:
3525 case SpvOpExtension
:
3526 case SpvOpModuleProcessed
:
3527 /* Unhandled, but these are for debug so that's ok. */
3530 case SpvOpCapability
: {
3531 SpvCapability cap
= w
[1];
3533 case SpvCapabilityMatrix
:
3534 case SpvCapabilityShader
:
3535 case SpvCapabilityGeometry
:
3536 case SpvCapabilityGeometryPointSize
:
3537 case SpvCapabilityUniformBufferArrayDynamicIndexing
:
3538 case SpvCapabilitySampledImageArrayDynamicIndexing
:
3539 case SpvCapabilityStorageBufferArrayDynamicIndexing
:
3540 case SpvCapabilityStorageImageArrayDynamicIndexing
:
3541 case SpvCapabilityImageRect
:
3542 case SpvCapabilitySampledRect
:
3543 case SpvCapabilitySampled1D
:
3544 case SpvCapabilityImage1D
:
3545 case SpvCapabilitySampledCubeArray
:
3546 case SpvCapabilityImageCubeArray
:
3547 case SpvCapabilitySampledBuffer
:
3548 case SpvCapabilityImageBuffer
:
3549 case SpvCapabilityImageQuery
:
3550 case SpvCapabilityDerivativeControl
:
3551 case SpvCapabilityInterpolationFunction
:
3552 case SpvCapabilityMultiViewport
:
3553 case SpvCapabilitySampleRateShading
:
3554 case SpvCapabilityClipDistance
:
3555 case SpvCapabilityCullDistance
:
3556 case SpvCapabilityInputAttachment
:
3557 case SpvCapabilityImageGatherExtended
:
3558 case SpvCapabilityStorageImageExtendedFormats
:
3561 case SpvCapabilityLinkage
:
3562 case SpvCapabilityVector16
:
3563 case SpvCapabilityFloat16Buffer
:
3564 case SpvCapabilityFloat16
:
3565 case SpvCapabilitySparseResidency
:
3566 vtn_warn("Unsupported SPIR-V capability: %s",
3567 spirv_capability_to_string(cap
));
3570 case SpvCapabilityMinLod
:
3571 spv_check_supported(min_lod
, cap
);
3574 case SpvCapabilityAtomicStorage
:
3575 spv_check_supported(atomic_storage
, cap
);
3578 case SpvCapabilityFloat64
:
3579 spv_check_supported(float64
, cap
);
3581 case SpvCapabilityInt64
:
3582 spv_check_supported(int64
, cap
);
3584 case SpvCapabilityInt16
:
3585 spv_check_supported(int16
, cap
);
3588 case SpvCapabilityTransformFeedback
:
3589 spv_check_supported(transform_feedback
, cap
);
3592 case SpvCapabilityGeometryStreams
:
3593 spv_check_supported(geometry_streams
, cap
);
3596 case SpvCapabilityInt64Atomics
:
3597 spv_check_supported(int64_atomics
, cap
);
3600 case SpvCapabilityInt8
:
3601 spv_check_supported(int8
, cap
);
3604 case SpvCapabilityStorageImageMultisample
:
3605 spv_check_supported(storage_image_ms
, cap
);
3608 case SpvCapabilityAddresses
:
3609 spv_check_supported(address
, cap
);
3612 case SpvCapabilityKernel
:
3613 spv_check_supported(kernel
, cap
);
3616 case SpvCapabilityImageBasic
:
3617 case SpvCapabilityImageReadWrite
:
3618 case SpvCapabilityImageMipmap
:
3619 case SpvCapabilityPipes
:
3620 case SpvCapabilityGroups
:
3621 case SpvCapabilityDeviceEnqueue
:
3622 case SpvCapabilityLiteralSampler
:
3623 case SpvCapabilityGenericPointer
:
3624 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
3625 spirv_capability_to_string(cap
));
3628 case SpvCapabilityImageMSArray
:
3629 spv_check_supported(image_ms_array
, cap
);
3632 case SpvCapabilityTessellation
:
3633 case SpvCapabilityTessellationPointSize
:
3634 spv_check_supported(tessellation
, cap
);
3637 case SpvCapabilityDrawParameters
:
3638 spv_check_supported(draw_parameters
, cap
);
3641 case SpvCapabilityStorageImageReadWithoutFormat
:
3642 spv_check_supported(image_read_without_format
, cap
);
3645 case SpvCapabilityStorageImageWriteWithoutFormat
:
3646 spv_check_supported(image_write_without_format
, cap
);
3649 case SpvCapabilityDeviceGroup
:
3650 spv_check_supported(device_group
, cap
);
3653 case SpvCapabilityMultiView
:
3654 spv_check_supported(multiview
, cap
);
3657 case SpvCapabilityGroupNonUniform
:
3658 spv_check_supported(subgroup_basic
, cap
);
3661 case SpvCapabilityGroupNonUniformVote
:
3662 spv_check_supported(subgroup_vote
, cap
);
3665 case SpvCapabilitySubgroupBallotKHR
:
3666 case SpvCapabilityGroupNonUniformBallot
:
3667 spv_check_supported(subgroup_ballot
, cap
);
3670 case SpvCapabilityGroupNonUniformShuffle
:
3671 case SpvCapabilityGroupNonUniformShuffleRelative
:
3672 spv_check_supported(subgroup_shuffle
, cap
);
3675 case SpvCapabilityGroupNonUniformQuad
:
3676 spv_check_supported(subgroup_quad
, cap
);
3679 case SpvCapabilityGroupNonUniformArithmetic
:
3680 case SpvCapabilityGroupNonUniformClustered
:
3681 spv_check_supported(subgroup_arithmetic
, cap
);
3684 case SpvCapabilityVariablePointersStorageBuffer
:
3685 case SpvCapabilityVariablePointers
:
3686 spv_check_supported(variable_pointers
, cap
);
3687 b
->variable_pointers
= true;
3690 case SpvCapabilityStorageUniformBufferBlock16
:
3691 case SpvCapabilityStorageUniform16
:
3692 case SpvCapabilityStoragePushConstant16
:
3693 case SpvCapabilityStorageInputOutput16
:
3694 spv_check_supported(storage_16bit
, cap
);
3697 case SpvCapabilityShaderViewportIndexLayerEXT
:
3698 spv_check_supported(shader_viewport_index_layer
, cap
);
3701 case SpvCapabilityStorageBuffer8BitAccess
:
3702 case SpvCapabilityUniformAndStorageBuffer8BitAccess
:
3703 case SpvCapabilityStoragePushConstant8
:
3704 spv_check_supported(storage_8bit
, cap
);
3707 case SpvCapabilityInputAttachmentArrayDynamicIndexingEXT
:
3708 case SpvCapabilityUniformTexelBufferArrayDynamicIndexingEXT
:
3709 case SpvCapabilityStorageTexelBufferArrayDynamicIndexingEXT
:
3710 spv_check_supported(descriptor_array_dynamic_indexing
, cap
);
3713 case SpvCapabilityRuntimeDescriptorArrayEXT
:
3714 spv_check_supported(runtime_descriptor_array
, cap
);
3717 case SpvCapabilityStencilExportEXT
:
3718 spv_check_supported(stencil_export
, cap
);
3721 case SpvCapabilitySampleMaskPostDepthCoverage
:
3722 spv_check_supported(post_depth_coverage
, cap
);
3725 case SpvCapabilityPhysicalStorageBufferAddressesEXT
:
3726 spv_check_supported(physical_storage_buffer_address
, cap
);
3729 case SpvCapabilityComputeDerivativeGroupQuadsNV
:
3730 case SpvCapabilityComputeDerivativeGroupLinearNV
:
3731 spv_check_supported(derivative_group
, cap
);
3735 vtn_fail("Unhandled capability");
3740 case SpvOpExtInstImport
:
3741 vtn_handle_extension(b
, opcode
, w
, count
);
3744 case SpvOpMemoryModel
:
3746 case SpvAddressingModelPhysical32
:
3747 vtn_fail_if(b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
,
3748 "AddressingModelPhysical32 only supported for kernels");
3749 b
->shader
->info
.cs
.ptr_size
= 32;
3750 b
->physical_ptrs
= true;
3751 b
->options
->shared_ptr_type
= glsl_uint_type();
3752 b
->options
->global_ptr_type
= glsl_uint_type();
3753 b
->options
->temp_ptr_type
= glsl_uint_type();
3755 case SpvAddressingModelPhysical64
:
3756 vtn_fail_if(b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
,
3757 "AddressingModelPhysical64 only supported for kernels");
3758 b
->shader
->info
.cs
.ptr_size
= 64;
3759 b
->physical_ptrs
= true;
3760 b
->options
->shared_ptr_type
= glsl_uint64_t_type();
3761 b
->options
->global_ptr_type
= glsl_uint64_t_type();
3762 b
->options
->temp_ptr_type
= glsl_uint64_t_type();
3764 case SpvAddressingModelLogical
:
3765 vtn_fail_if(b
->shader
->info
.stage
>= MESA_SHADER_STAGES
,
3766 "AddressingModelLogical only supported for shaders");
3767 b
->shader
->info
.cs
.ptr_size
= 0;
3768 b
->physical_ptrs
= false;
3770 case SpvAddressingModelPhysicalStorageBuffer64EXT
:
3771 vtn_fail_if(!b
->options
||
3772 !b
->options
->caps
.physical_storage_buffer_address
,
3773 "AddressingModelPhysicalStorageBuffer64EXT not supported");
3776 vtn_fail("Unknown addressing model");
3780 vtn_assert(w
[2] == SpvMemoryModelSimple
||
3781 w
[2] == SpvMemoryModelGLSL450
||
3782 w
[2] == SpvMemoryModelOpenCL
);
3785 case SpvOpEntryPoint
:
3786 vtn_handle_entry_point(b
, w
, count
);
3790 vtn_push_value(b
, w
[1], vtn_value_type_string
)->str
=
3791 vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
3795 b
->values
[w
[1]].name
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
3798 case SpvOpMemberName
:
3802 case SpvOpExecutionMode
:
3803 case SpvOpExecutionModeId
:
3804 case SpvOpDecorationGroup
:
3806 case SpvOpMemberDecorate
:
3807 case SpvOpGroupDecorate
:
3808 case SpvOpGroupMemberDecorate
:
3809 case SpvOpDecorateStringGOOGLE
:
3810 case SpvOpMemberDecorateStringGOOGLE
:
3811 vtn_handle_decoration(b
, opcode
, w
, count
);
3815 return false; /* End of preamble */
3822 vtn_handle_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
3823 const struct vtn_decoration
*mode
, void *data
)
3825 vtn_assert(b
->entry_point
== entry_point
);
3827 switch(mode
->exec_mode
) {
3828 case SpvExecutionModeOriginUpperLeft
:
3829 case SpvExecutionModeOriginLowerLeft
:
3830 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3831 b
->shader
->info
.fs
.origin_upper_left
=
3832 (mode
->exec_mode
== SpvExecutionModeOriginUpperLeft
);
3835 case SpvExecutionModeEarlyFragmentTests
:
3836 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3837 b
->shader
->info
.fs
.early_fragment_tests
= true;
3840 case SpvExecutionModePostDepthCoverage
:
3841 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3842 b
->shader
->info
.fs
.post_depth_coverage
= true;
3845 case SpvExecutionModeInvocations
:
3846 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3847 b
->shader
->info
.gs
.invocations
= MAX2(1, mode
->literals
[0]);
3850 case SpvExecutionModeDepthReplacing
:
3851 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3852 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_ANY
;
3854 case SpvExecutionModeDepthGreater
:
3855 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3856 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_GREATER
;
3858 case SpvExecutionModeDepthLess
:
3859 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3860 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_LESS
;
3862 case SpvExecutionModeDepthUnchanged
:
3863 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3864 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
3867 case SpvExecutionModeLocalSize
:
3868 vtn_assert(gl_shader_stage_is_compute(b
->shader
->info
.stage
));
3869 b
->shader
->info
.cs
.local_size
[0] = mode
->literals
[0];
3870 b
->shader
->info
.cs
.local_size
[1] = mode
->literals
[1];
3871 b
->shader
->info
.cs
.local_size
[2] = mode
->literals
[2];
3874 case SpvExecutionModeLocalSizeId
:
3875 b
->shader
->info
.cs
.local_size
[0] = vtn_constant_uint(b
, mode
->literals
[0]);
3876 b
->shader
->info
.cs
.local_size
[1] = vtn_constant_uint(b
, mode
->literals
[1]);
3877 b
->shader
->info
.cs
.local_size
[2] = vtn_constant_uint(b
, mode
->literals
[2]);
3880 case SpvExecutionModeLocalSizeHint
:
3881 case SpvExecutionModeLocalSizeHintId
:
3882 break; /* Nothing to do with this */
3884 case SpvExecutionModeOutputVertices
:
3885 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3886 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
3887 b
->shader
->info
.tess
.tcs_vertices_out
= mode
->literals
[0];
3889 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3890 b
->shader
->info
.gs
.vertices_out
= mode
->literals
[0];
3894 case SpvExecutionModeInputPoints
:
3895 case SpvExecutionModeInputLines
:
3896 case SpvExecutionModeInputLinesAdjacency
:
3897 case SpvExecutionModeTriangles
:
3898 case SpvExecutionModeInputTrianglesAdjacency
:
3899 case SpvExecutionModeQuads
:
3900 case SpvExecutionModeIsolines
:
3901 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3902 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
3903 b
->shader
->info
.tess
.primitive_mode
=
3904 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
3906 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3907 b
->shader
->info
.gs
.vertices_in
=
3908 vertices_in_from_spv_execution_mode(b
, mode
->exec_mode
);
3909 b
->shader
->info
.gs
.input_primitive
=
3910 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
3914 case SpvExecutionModeOutputPoints
:
3915 case SpvExecutionModeOutputLineStrip
:
3916 case SpvExecutionModeOutputTriangleStrip
:
3917 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3918 b
->shader
->info
.gs
.output_primitive
=
3919 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
3922 case SpvExecutionModeSpacingEqual
:
3923 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3924 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3925 b
->shader
->info
.tess
.spacing
= TESS_SPACING_EQUAL
;
3927 case SpvExecutionModeSpacingFractionalEven
:
3928 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3929 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3930 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_EVEN
;
3932 case SpvExecutionModeSpacingFractionalOdd
:
3933 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3934 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3935 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_ODD
;
3937 case SpvExecutionModeVertexOrderCw
:
3938 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3939 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3940 b
->shader
->info
.tess
.ccw
= false;
3942 case SpvExecutionModeVertexOrderCcw
:
3943 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3944 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3945 b
->shader
->info
.tess
.ccw
= true;
3947 case SpvExecutionModePointMode
:
3948 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3949 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3950 b
->shader
->info
.tess
.point_mode
= true;
3953 case SpvExecutionModePixelCenterInteger
:
3954 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3955 b
->shader
->info
.fs
.pixel_center_integer
= true;
3958 case SpvExecutionModeXfb
:
3959 b
->shader
->info
.has_transform_feedback_varyings
= true;
3962 case SpvExecutionModeVecTypeHint
:
3965 case SpvExecutionModeContractionOff
:
3966 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
)
3967 vtn_warn("ExectionMode only allowed for CL-style kernels: %s",
3968 spirv_executionmode_to_string(mode
->exec_mode
));
3973 case SpvExecutionModeStencilRefReplacingEXT
:
3974 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3977 case SpvExecutionModeDerivativeGroupQuadsNV
:
3978 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
3979 b
->shader
->info
.cs
.derivative_group
= DERIVATIVE_GROUP_QUADS
;
3982 case SpvExecutionModeDerivativeGroupLinearNV
:
3983 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
3984 b
->shader
->info
.cs
.derivative_group
= DERIVATIVE_GROUP_LINEAR
;
3988 vtn_fail("Unhandled execution mode");
3993 vtn_handle_variable_or_type_instruction(struct vtn_builder
*b
, SpvOp opcode
,
3994 const uint32_t *w
, unsigned count
)
3996 vtn_set_instruction_result_type(b
, opcode
, w
, count
);
4000 case SpvOpSourceContinued
:
4001 case SpvOpSourceExtension
:
4002 case SpvOpExtension
:
4003 case SpvOpCapability
:
4004 case SpvOpExtInstImport
:
4005 case SpvOpMemoryModel
:
4006 case SpvOpEntryPoint
:
4007 case SpvOpExecutionMode
:
4010 case SpvOpMemberName
:
4011 case SpvOpDecorationGroup
:
4013 case SpvOpMemberDecorate
:
4014 case SpvOpGroupDecorate
:
4015 case SpvOpGroupMemberDecorate
:
4016 case SpvOpDecorateStringGOOGLE
:
4017 case SpvOpMemberDecorateStringGOOGLE
:
4018 vtn_fail("Invalid opcode types and variables section");
4024 case SpvOpTypeFloat
:
4025 case SpvOpTypeVector
:
4026 case SpvOpTypeMatrix
:
4027 case SpvOpTypeImage
:
4028 case SpvOpTypeSampler
:
4029 case SpvOpTypeSampledImage
:
4030 case SpvOpTypeArray
:
4031 case SpvOpTypeRuntimeArray
:
4032 case SpvOpTypeStruct
:
4033 case SpvOpTypeOpaque
:
4034 case SpvOpTypePointer
:
4035 case SpvOpTypeForwardPointer
:
4036 case SpvOpTypeFunction
:
4037 case SpvOpTypeEvent
:
4038 case SpvOpTypeDeviceEvent
:
4039 case SpvOpTypeReserveId
:
4040 case SpvOpTypeQueue
:
4042 vtn_handle_type(b
, opcode
, w
, count
);
4045 case SpvOpConstantTrue
:
4046 case SpvOpConstantFalse
:
4048 case SpvOpConstantComposite
:
4049 case SpvOpConstantSampler
:
4050 case SpvOpConstantNull
:
4051 case SpvOpSpecConstantTrue
:
4052 case SpvOpSpecConstantFalse
:
4053 case SpvOpSpecConstant
:
4054 case SpvOpSpecConstantComposite
:
4055 case SpvOpSpecConstantOp
:
4056 vtn_handle_constant(b
, opcode
, w
, count
);
4061 vtn_handle_variables(b
, opcode
, w
, count
);
4065 return false; /* End of preamble */
4072 vtn_handle_body_instruction(struct vtn_builder
*b
, SpvOp opcode
,
4073 const uint32_t *w
, unsigned count
)
4079 case SpvOpLoopMerge
:
4080 case SpvOpSelectionMerge
:
4081 /* This is handled by cfg pre-pass and walk_blocks */
4085 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
4086 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
4091 vtn_handle_extension(b
, opcode
, w
, count
);
4097 case SpvOpCopyMemory
:
4098 case SpvOpCopyMemorySized
:
4099 case SpvOpAccessChain
:
4100 case SpvOpPtrAccessChain
:
4101 case SpvOpInBoundsAccessChain
:
4102 case SpvOpInBoundsPtrAccessChain
:
4103 case SpvOpArrayLength
:
4104 case SpvOpConvertPtrToU
:
4105 case SpvOpConvertUToPtr
:
4106 vtn_handle_variables(b
, opcode
, w
, count
);
4109 case SpvOpFunctionCall
:
4110 vtn_handle_function_call(b
, opcode
, w
, count
);
4113 case SpvOpSampledImage
:
4115 case SpvOpImageSampleImplicitLod
:
4116 case SpvOpImageSampleExplicitLod
:
4117 case SpvOpImageSampleDrefImplicitLod
:
4118 case SpvOpImageSampleDrefExplicitLod
:
4119 case SpvOpImageSampleProjImplicitLod
:
4120 case SpvOpImageSampleProjExplicitLod
:
4121 case SpvOpImageSampleProjDrefImplicitLod
:
4122 case SpvOpImageSampleProjDrefExplicitLod
:
4123 case SpvOpImageFetch
:
4124 case SpvOpImageGather
:
4125 case SpvOpImageDrefGather
:
4126 case SpvOpImageQuerySizeLod
:
4127 case SpvOpImageQueryLod
:
4128 case SpvOpImageQueryLevels
:
4129 case SpvOpImageQuerySamples
:
4130 vtn_handle_texture(b
, opcode
, w
, count
);
4133 case SpvOpImageRead
:
4134 case SpvOpImageWrite
:
4135 case SpvOpImageTexelPointer
:
4136 vtn_handle_image(b
, opcode
, w
, count
);
4139 case SpvOpImageQuerySize
: {
4140 struct vtn_pointer
*image
=
4141 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
4142 if (glsl_type_is_image(image
->type
->type
)) {
4143 vtn_handle_image(b
, opcode
, w
, count
);
4145 vtn_assert(glsl_type_is_sampler(image
->type
->type
));
4146 vtn_handle_texture(b
, opcode
, w
, count
);
4151 case SpvOpAtomicLoad
:
4152 case SpvOpAtomicExchange
:
4153 case SpvOpAtomicCompareExchange
:
4154 case SpvOpAtomicCompareExchangeWeak
:
4155 case SpvOpAtomicIIncrement
:
4156 case SpvOpAtomicIDecrement
:
4157 case SpvOpAtomicIAdd
:
4158 case SpvOpAtomicISub
:
4159 case SpvOpAtomicSMin
:
4160 case SpvOpAtomicUMin
:
4161 case SpvOpAtomicSMax
:
4162 case SpvOpAtomicUMax
:
4163 case SpvOpAtomicAnd
:
4165 case SpvOpAtomicXor
: {
4166 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[3]);
4167 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
4168 vtn_handle_image(b
, opcode
, w
, count
);
4170 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
4171 vtn_handle_atomics(b
, opcode
, w
, count
);
4176 case SpvOpAtomicStore
: {
4177 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[1]);
4178 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
4179 vtn_handle_image(b
, opcode
, w
, count
);
4181 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
4182 vtn_handle_atomics(b
, opcode
, w
, count
);
4188 /* Handle OpSelect up-front here because it needs to be able to handle
4189 * pointers and not just regular vectors and scalars.
4191 struct vtn_value
*res_val
= vtn_untyped_value(b
, w
[2]);
4192 struct vtn_value
*sel_val
= vtn_untyped_value(b
, w
[3]);
4193 struct vtn_value
*obj1_val
= vtn_untyped_value(b
, w
[4]);
4194 struct vtn_value
*obj2_val
= vtn_untyped_value(b
, w
[5]);
4196 const struct glsl_type
*sel_type
;
4197 switch (res_val
->type
->base_type
) {
4198 case vtn_base_type_scalar
:
4199 sel_type
= glsl_bool_type();
4201 case vtn_base_type_vector
:
4202 sel_type
= glsl_vector_type(GLSL_TYPE_BOOL
, res_val
->type
->length
);
4204 case vtn_base_type_pointer
:
4205 /* We need to have actual storage for pointer types */
4206 vtn_fail_if(res_val
->type
->type
== NULL
,
4207 "Invalid pointer result type for OpSelect");
4208 sel_type
= glsl_bool_type();
4211 vtn_fail("Result type of OpSelect must be a scalar, vector, or pointer");
4214 if (unlikely(sel_val
->type
->type
!= sel_type
)) {
4215 if (sel_val
->type
->type
== glsl_bool_type()) {
4216 /* This case is illegal but some older versions of GLSLang produce
4217 * it. The GLSLang issue was fixed on March 30, 2017:
4219 * https://github.com/KhronosGroup/glslang/issues/809
4221 * Unfortunately, there are applications in the wild which are
4222 * shipping with this bug so it isn't nice to fail on them so we
4223 * throw a warning instead. It's not actually a problem for us as
4224 * nir_builder will just splat the condition out which is most
4225 * likely what the client wanted anyway.
4227 vtn_warn("Condition type of OpSelect must have the same number "
4228 "of components as Result Type");
4230 vtn_fail("Condition type of OpSelect must be a scalar or vector "
4231 "of Boolean type. It must have the same number of "
4232 "components as Result Type");
4236 vtn_fail_if(obj1_val
->type
!= res_val
->type
||
4237 obj2_val
->type
!= res_val
->type
,
4238 "Object types must match the result type in OpSelect");
4240 struct vtn_type
*res_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
4241 struct vtn_ssa_value
*ssa
= vtn_create_ssa_value(b
, res_type
->type
);
4242 ssa
->def
= nir_bcsel(&b
->nb
, vtn_ssa_value(b
, w
[3])->def
,
4243 vtn_ssa_value(b
, w
[4])->def
,
4244 vtn_ssa_value(b
, w
[5])->def
);
4245 vtn_push_ssa(b
, w
[2], res_type
, ssa
);
4254 case SpvOpConvertFToU
:
4255 case SpvOpConvertFToS
:
4256 case SpvOpConvertSToF
:
4257 case SpvOpConvertUToF
:
4261 case SpvOpQuantizeToF16
:
4262 case SpvOpPtrCastToGeneric
:
4263 case SpvOpGenericCastToPtr
:
4269 case SpvOpSignBitSet
:
4270 case SpvOpLessOrGreater
:
4272 case SpvOpUnordered
:
4287 case SpvOpVectorTimesScalar
:
4289 case SpvOpIAddCarry
:
4290 case SpvOpISubBorrow
:
4291 case SpvOpUMulExtended
:
4292 case SpvOpSMulExtended
:
4293 case SpvOpShiftRightLogical
:
4294 case SpvOpShiftRightArithmetic
:
4295 case SpvOpShiftLeftLogical
:
4296 case SpvOpLogicalEqual
:
4297 case SpvOpLogicalNotEqual
:
4298 case SpvOpLogicalOr
:
4299 case SpvOpLogicalAnd
:
4300 case SpvOpLogicalNot
:
4301 case SpvOpBitwiseOr
:
4302 case SpvOpBitwiseXor
:
4303 case SpvOpBitwiseAnd
:
4305 case SpvOpFOrdEqual
:
4306 case SpvOpFUnordEqual
:
4307 case SpvOpINotEqual
:
4308 case SpvOpFOrdNotEqual
:
4309 case SpvOpFUnordNotEqual
:
4310 case SpvOpULessThan
:
4311 case SpvOpSLessThan
:
4312 case SpvOpFOrdLessThan
:
4313 case SpvOpFUnordLessThan
:
4314 case SpvOpUGreaterThan
:
4315 case SpvOpSGreaterThan
:
4316 case SpvOpFOrdGreaterThan
:
4317 case SpvOpFUnordGreaterThan
:
4318 case SpvOpULessThanEqual
:
4319 case SpvOpSLessThanEqual
:
4320 case SpvOpFOrdLessThanEqual
:
4321 case SpvOpFUnordLessThanEqual
:
4322 case SpvOpUGreaterThanEqual
:
4323 case SpvOpSGreaterThanEqual
:
4324 case SpvOpFOrdGreaterThanEqual
:
4325 case SpvOpFUnordGreaterThanEqual
:
4331 case SpvOpFwidthFine
:
4332 case SpvOpDPdxCoarse
:
4333 case SpvOpDPdyCoarse
:
4334 case SpvOpFwidthCoarse
:
4335 case SpvOpBitFieldInsert
:
4336 case SpvOpBitFieldSExtract
:
4337 case SpvOpBitFieldUExtract
:
4338 case SpvOpBitReverse
:
4340 case SpvOpTranspose
:
4341 case SpvOpOuterProduct
:
4342 case SpvOpMatrixTimesScalar
:
4343 case SpvOpVectorTimesMatrix
:
4344 case SpvOpMatrixTimesVector
:
4345 case SpvOpMatrixTimesMatrix
:
4346 vtn_handle_alu(b
, opcode
, w
, count
);
4349 case SpvOpVectorExtractDynamic
:
4350 case SpvOpVectorInsertDynamic
:
4351 case SpvOpVectorShuffle
:
4352 case SpvOpCompositeConstruct
:
4353 case SpvOpCompositeExtract
:
4354 case SpvOpCompositeInsert
:
4355 case SpvOpCopyObject
:
4356 vtn_handle_composite(b
, opcode
, w
, count
);
4359 case SpvOpEmitVertex
:
4360 case SpvOpEndPrimitive
:
4361 case SpvOpEmitStreamVertex
:
4362 case SpvOpEndStreamPrimitive
:
4363 case SpvOpControlBarrier
:
4364 case SpvOpMemoryBarrier
:
4365 vtn_handle_barrier(b
, opcode
, w
, count
);
4368 case SpvOpGroupNonUniformElect
:
4369 case SpvOpGroupNonUniformAll
:
4370 case SpvOpGroupNonUniformAny
:
4371 case SpvOpGroupNonUniformAllEqual
:
4372 case SpvOpGroupNonUniformBroadcast
:
4373 case SpvOpGroupNonUniformBroadcastFirst
:
4374 case SpvOpGroupNonUniformBallot
:
4375 case SpvOpGroupNonUniformInverseBallot
:
4376 case SpvOpGroupNonUniformBallotBitExtract
:
4377 case SpvOpGroupNonUniformBallotBitCount
:
4378 case SpvOpGroupNonUniformBallotFindLSB
:
4379 case SpvOpGroupNonUniformBallotFindMSB
:
4380 case SpvOpGroupNonUniformShuffle
:
4381 case SpvOpGroupNonUniformShuffleXor
:
4382 case SpvOpGroupNonUniformShuffleUp
:
4383 case SpvOpGroupNonUniformShuffleDown
:
4384 case SpvOpGroupNonUniformIAdd
:
4385 case SpvOpGroupNonUniformFAdd
:
4386 case SpvOpGroupNonUniformIMul
:
4387 case SpvOpGroupNonUniformFMul
:
4388 case SpvOpGroupNonUniformSMin
:
4389 case SpvOpGroupNonUniformUMin
:
4390 case SpvOpGroupNonUniformFMin
:
4391 case SpvOpGroupNonUniformSMax
:
4392 case SpvOpGroupNonUniformUMax
:
4393 case SpvOpGroupNonUniformFMax
:
4394 case SpvOpGroupNonUniformBitwiseAnd
:
4395 case SpvOpGroupNonUniformBitwiseOr
:
4396 case SpvOpGroupNonUniformBitwiseXor
:
4397 case SpvOpGroupNonUniformLogicalAnd
:
4398 case SpvOpGroupNonUniformLogicalOr
:
4399 case SpvOpGroupNonUniformLogicalXor
:
4400 case SpvOpGroupNonUniformQuadBroadcast
:
4401 case SpvOpGroupNonUniformQuadSwap
:
4402 vtn_handle_subgroup(b
, opcode
, w
, count
);
4406 vtn_fail("Unhandled opcode");
4413 vtn_create_builder(const uint32_t *words
, size_t word_count
,
4414 gl_shader_stage stage
, const char *entry_point_name
,
4415 const struct spirv_to_nir_options
*options
)
4417 /* Initialize the vtn_builder object */
4418 struct vtn_builder
*b
= rzalloc(NULL
, struct vtn_builder
);
4419 struct spirv_to_nir_options
*dup_options
=
4420 ralloc(b
, struct spirv_to_nir_options
);
4421 *dup_options
= *options
;
4424 b
->spirv_word_count
= word_count
;
4428 exec_list_make_empty(&b
->functions
);
4429 b
->entry_point_stage
= stage
;
4430 b
->entry_point_name
= entry_point_name
;
4431 b
->options
= dup_options
;
4434 * Handle the SPIR-V header (first 5 dwords).
4435 * Can't use vtx_assert() as the setjmp(3) target isn't initialized yet.
4437 if (word_count
<= 5)
4440 if (words
[0] != SpvMagicNumber
) {
4441 vtn_err("words[0] was 0x%x, want 0x%x", words
[0], SpvMagicNumber
);
4444 if (words
[1] < 0x10000) {
4445 vtn_err("words[1] was 0x%x, want >= 0x10000", words
[1]);
4449 uint16_t generator_id
= words
[2] >> 16;
4450 uint16_t generator_version
= words
[2];
4452 /* The first GLSLang version bump actually 1.5 years after #179 was fixed
4453 * but this should at least let us shut the workaround off for modern
4454 * versions of GLSLang.
4456 b
->wa_glslang_179
= (generator_id
== 8 && generator_version
== 1);
4458 /* words[2] == generator magic */
4459 unsigned value_id_bound
= words
[3];
4460 if (words
[4] != 0) {
4461 vtn_err("words[4] was %u, want 0", words
[4]);
4465 b
->value_id_bound
= value_id_bound
;
4466 b
->values
= rzalloc_array(b
, struct vtn_value
, value_id_bound
);
4474 static nir_function
*
4475 vtn_emit_kernel_entry_point_wrapper(struct vtn_builder
*b
,
4476 nir_function
*entry_point
)
4478 vtn_assert(entry_point
== b
->entry_point
->func
->impl
->function
);
4479 vtn_fail_if(!entry_point
->name
, "entry points are required to have a name");
4480 const char *func_name
=
4481 ralloc_asprintf(b
->shader
, "__wrapped_%s", entry_point
->name
);
4483 /* we shouldn't have any inputs yet */
4484 vtn_assert(!entry_point
->shader
->num_inputs
);
4485 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_KERNEL
);
4487 nir_function
*main_entry_point
= nir_function_create(b
->shader
, func_name
);
4488 main_entry_point
->impl
= nir_function_impl_create(main_entry_point
);
4489 nir_builder_init(&b
->nb
, main_entry_point
->impl
);
4490 b
->nb
.cursor
= nir_after_cf_list(&main_entry_point
->impl
->body
);
4491 b
->func_param_idx
= 0;
4493 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, entry_point
);
4495 for (unsigned i
= 0; i
< entry_point
->num_params
; ++i
) {
4496 struct vtn_type
*param_type
= b
->entry_point
->func
->type
->params
[i
];
4498 /* consider all pointers to function memory to be parameters passed
4501 bool is_by_val
= param_type
->base_type
== vtn_base_type_pointer
&&
4502 param_type
->storage_class
== SpvStorageClassFunction
;
4504 /* input variable */
4505 nir_variable
*in_var
= rzalloc(b
->nb
.shader
, nir_variable
);
4506 in_var
->data
.mode
= nir_var_shader_in
;
4507 in_var
->data
.read_only
= true;
4508 in_var
->data
.location
= i
;
4511 in_var
->type
= param_type
->deref
->type
;
4513 in_var
->type
= param_type
->type
;
4515 nir_shader_add_variable(b
->nb
.shader
, in_var
);
4516 b
->nb
.shader
->num_inputs
++;
4518 /* we have to copy the entire variable into function memory */
4520 nir_variable
*copy_var
=
4521 nir_local_variable_create(main_entry_point
->impl
, in_var
->type
,
4523 nir_copy_var(&b
->nb
, copy_var
, in_var
);
4525 nir_src_for_ssa(&nir_build_deref_var(&b
->nb
, copy_var
)->dest
.ssa
);
4527 call
->params
[i
] = nir_src_for_ssa(nir_load_var(&b
->nb
, in_var
));
4531 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
4533 return main_entry_point
;
4537 spirv_to_nir(const uint32_t *words
, size_t word_count
,
4538 struct nir_spirv_specialization
*spec
, unsigned num_spec
,
4539 gl_shader_stage stage
, const char *entry_point_name
,
4540 const struct spirv_to_nir_options
*options
,
4541 const nir_shader_compiler_options
*nir_options
)
4544 const uint32_t *word_end
= words
+ word_count
;
4546 struct vtn_builder
*b
= vtn_create_builder(words
, word_count
,
4547 stage
, entry_point_name
,
4553 /* See also _vtn_fail() */
4554 if (setjmp(b
->fail_jump
)) {
4559 /* Skip the SPIR-V header, handled at vtn_create_builder */
4562 b
->shader
= nir_shader_create(b
, stage
, nir_options
, NULL
);
4564 /* Handle all the preamble instructions */
4565 words
= vtn_foreach_instruction(b
, words
, word_end
,
4566 vtn_handle_preamble_instruction
);
4568 if (b
->entry_point
== NULL
) {
4569 vtn_fail("Entry point not found");
4574 /* Set shader info defaults */
4575 b
->shader
->info
.gs
.invocations
= 1;
4577 b
->specializations
= spec
;
4578 b
->num_specializations
= num_spec
;
4580 /* Handle all variable, type, and constant instructions */
4581 words
= vtn_foreach_instruction(b
, words
, word_end
,
4582 vtn_handle_variable_or_type_instruction
);
4584 /* Parse execution modes */
4585 vtn_foreach_execution_mode(b
, b
->entry_point
,
4586 vtn_handle_execution_mode
, NULL
);
4588 if (b
->workgroup_size_builtin
) {
4589 vtn_assert(b
->workgroup_size_builtin
->type
->type
==
4590 glsl_vector_type(GLSL_TYPE_UINT
, 3));
4592 nir_const_value
*const_size
=
4593 b
->workgroup_size_builtin
->constant
->values
[0];
4595 b
->shader
->info
.cs
.local_size
[0] = const_size
[0].u32
;
4596 b
->shader
->info
.cs
.local_size
[1] = const_size
[1].u32
;
4597 b
->shader
->info
.cs
.local_size
[2] = const_size
[2].u32
;
4600 /* Set types on all vtn_values */
4601 vtn_foreach_instruction(b
, words
, word_end
, vtn_set_instruction_result_type
);
4603 vtn_build_cfg(b
, words
, word_end
);
4605 assert(b
->entry_point
->value_type
== vtn_value_type_function
);
4606 b
->entry_point
->func
->referenced
= true;
4611 foreach_list_typed(struct vtn_function
, func
, node
, &b
->functions
) {
4612 if (func
->referenced
&& !func
->emitted
) {
4613 b
->const_table
= _mesa_pointer_hash_table_create(b
);
4615 vtn_function_emit(b
, func
, vtn_handle_body_instruction
);
4621 vtn_assert(b
->entry_point
->value_type
== vtn_value_type_function
);
4622 nir_function
*entry_point
= b
->entry_point
->func
->impl
->function
;
4623 vtn_assert(entry_point
);
4625 /* post process entry_points with input params */
4626 if (entry_point
->num_params
&& b
->shader
->info
.stage
== MESA_SHADER_KERNEL
)
4627 entry_point
= vtn_emit_kernel_entry_point_wrapper(b
, entry_point
);
4629 entry_point
->is_entrypoint
= true;
4631 /* When multiple shader stages exist in the same SPIR-V module, we
4632 * generate input and output variables for every stage, in the same
4633 * NIR program. These dead variables can be invalid NIR. For example,
4634 * TCS outputs must be per-vertex arrays (or decorated 'patch'), while
4635 * VS output variables wouldn't be.
4637 * To ensure we have valid NIR, we eliminate any dead inputs and outputs
4638 * right away. In order to do so, we must lower any constant initializers
4639 * on outputs so nir_remove_dead_variables sees that they're written to.
4641 nir_lower_constant_initializers(b
->shader
, nir_var_shader_out
);
4642 nir_remove_dead_variables(b
->shader
,
4643 nir_var_shader_in
| nir_var_shader_out
);
4645 /* We sometimes generate bogus derefs that, while never used, give the
4646 * validator a bit of heartburn. Run dead code to get rid of them.
4648 nir_opt_dce(b
->shader
);
4650 /* Unparent the shader from the vtn_builder before we delete the builder */
4651 ralloc_steal(NULL
, b
->shader
);