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_with_opcode("Unhandled opcode", 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_with_decoration("Unhandled decoration", dec
->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_with_decoration("Unhandled decoration", dec
->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: %s (%u)",
1075 spirv_imageformat_to_string(format
), format
);
1079 static struct vtn_type
*
1080 vtn_type_layout_std430(struct vtn_builder
*b
, struct vtn_type
*type
,
1081 uint32_t *size_out
, uint32_t *align_out
)
1083 switch (type
->base_type
) {
1084 case vtn_base_type_scalar
: {
1085 uint32_t comp_size
= glsl_type_is_boolean(type
->type
)
1086 ? 4 : glsl_get_bit_size(type
->type
) / 8;
1087 *size_out
= comp_size
;
1088 *align_out
= comp_size
;
1092 case vtn_base_type_vector
: {
1093 uint32_t comp_size
= glsl_type_is_boolean(type
->type
)
1094 ? 4 : glsl_get_bit_size(type
->type
) / 8;
1095 unsigned align_comps
= type
->length
== 3 ? 4 : type
->length
;
1096 *size_out
= comp_size
* type
->length
,
1097 *align_out
= comp_size
* align_comps
;
1101 case vtn_base_type_matrix
:
1102 case vtn_base_type_array
: {
1103 /* We're going to add an array stride */
1104 type
= vtn_type_copy(b
, type
);
1105 uint32_t elem_size
, elem_align
;
1106 type
->array_element
= vtn_type_layout_std430(b
, type
->array_element
,
1107 &elem_size
, &elem_align
);
1108 type
->stride
= vtn_align_u32(elem_size
, elem_align
);
1109 *size_out
= type
->stride
* type
->length
;
1110 *align_out
= elem_align
;
1114 case vtn_base_type_struct
: {
1115 /* We're going to add member offsets */
1116 type
= vtn_type_copy(b
, type
);
1117 uint32_t offset
= 0;
1119 for (unsigned i
= 0; i
< type
->length
; i
++) {
1120 uint32_t mem_size
, mem_align
;
1121 type
->members
[i
] = vtn_type_layout_std430(b
, type
->members
[i
],
1122 &mem_size
, &mem_align
);
1123 offset
= vtn_align_u32(offset
, mem_align
);
1124 type
->offsets
[i
] = offset
;
1126 align
= MAX2(align
, mem_align
);
1134 unreachable("Invalid SPIR-V type for std430");
1139 vtn_handle_type(struct vtn_builder
*b
, SpvOp opcode
,
1140 const uint32_t *w
, unsigned count
)
1142 struct vtn_value
*val
= NULL
;
1144 /* In order to properly handle forward declarations, we have to defer
1145 * allocation for pointer types.
1147 if (opcode
!= SpvOpTypePointer
&& opcode
!= SpvOpTypeForwardPointer
) {
1148 val
= vtn_push_value(b
, w
[1], vtn_value_type_type
);
1149 vtn_fail_if(val
->type
!= NULL
,
1150 "Only pointers can have forward declarations");
1151 val
->type
= rzalloc(b
, struct vtn_type
);
1152 val
->type
->id
= w
[1];
1157 val
->type
->base_type
= vtn_base_type_void
;
1158 val
->type
->type
= glsl_void_type();
1161 val
->type
->base_type
= vtn_base_type_scalar
;
1162 val
->type
->type
= glsl_bool_type();
1163 val
->type
->length
= 1;
1165 case SpvOpTypeInt
: {
1166 int bit_size
= w
[2];
1167 const bool signedness
= w
[3];
1168 val
->type
->base_type
= vtn_base_type_scalar
;
1171 val
->type
->type
= (signedness
? glsl_int64_t_type() : glsl_uint64_t_type());
1174 val
->type
->type
= (signedness
? glsl_int_type() : glsl_uint_type());
1177 val
->type
->type
= (signedness
? glsl_int16_t_type() : glsl_uint16_t_type());
1180 val
->type
->type
= (signedness
? glsl_int8_t_type() : glsl_uint8_t_type());
1183 vtn_fail("Invalid int bit size: %u", bit_size
);
1185 val
->type
->length
= 1;
1189 case SpvOpTypeFloat
: {
1190 int bit_size
= w
[2];
1191 val
->type
->base_type
= vtn_base_type_scalar
;
1194 val
->type
->type
= glsl_float16_t_type();
1197 val
->type
->type
= glsl_float_type();
1200 val
->type
->type
= glsl_double_type();
1203 vtn_fail("Invalid float bit size: %u", bit_size
);
1205 val
->type
->length
= 1;
1209 case SpvOpTypeVector
: {
1210 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1211 unsigned elems
= w
[3];
1213 vtn_fail_if(base
->base_type
!= vtn_base_type_scalar
,
1214 "Base type for OpTypeVector must be a scalar");
1215 vtn_fail_if((elems
< 2 || elems
> 4) && (elems
!= 8) && (elems
!= 16),
1216 "Invalid component count for OpTypeVector");
1218 val
->type
->base_type
= vtn_base_type_vector
;
1219 val
->type
->type
= glsl_vector_type(glsl_get_base_type(base
->type
), elems
);
1220 val
->type
->length
= elems
;
1221 val
->type
->stride
= glsl_type_is_boolean(val
->type
->type
)
1222 ? 4 : glsl_get_bit_size(base
->type
) / 8;
1223 val
->type
->array_element
= base
;
1227 case SpvOpTypeMatrix
: {
1228 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1229 unsigned columns
= w
[3];
1231 vtn_fail_if(base
->base_type
!= vtn_base_type_vector
,
1232 "Base type for OpTypeMatrix must be a vector");
1233 vtn_fail_if(columns
< 2 || columns
> 4,
1234 "Invalid column count for OpTypeMatrix");
1236 val
->type
->base_type
= vtn_base_type_matrix
;
1237 val
->type
->type
= glsl_matrix_type(glsl_get_base_type(base
->type
),
1238 glsl_get_vector_elements(base
->type
),
1240 vtn_fail_if(glsl_type_is_error(val
->type
->type
),
1241 "Unsupported base type for OpTypeMatrix");
1242 assert(!glsl_type_is_error(val
->type
->type
));
1243 val
->type
->length
= columns
;
1244 val
->type
->array_element
= base
;
1245 val
->type
->row_major
= false;
1246 val
->type
->stride
= 0;
1250 case SpvOpTypeRuntimeArray
:
1251 case SpvOpTypeArray
: {
1252 struct vtn_type
*array_element
=
1253 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1255 if (opcode
== SpvOpTypeRuntimeArray
) {
1256 /* A length of 0 is used to denote unsized arrays */
1257 val
->type
->length
= 0;
1260 vtn_value(b
, w
[3], vtn_value_type_constant
)->constant
->values
[0][0].u32
;
1263 val
->type
->base_type
= vtn_base_type_array
;
1264 val
->type
->array_element
= array_element
;
1265 if (b
->shader
->info
.stage
== MESA_SHADER_KERNEL
)
1266 val
->type
->stride
= glsl_get_cl_size(array_element
->type
);
1268 vtn_foreach_decoration(b
, val
, array_stride_decoration_cb
, NULL
);
1269 val
->type
->type
= glsl_array_type(array_element
->type
, val
->type
->length
,
1274 case SpvOpTypeStruct
: {
1275 unsigned num_fields
= count
- 2;
1276 val
->type
->base_type
= vtn_base_type_struct
;
1277 val
->type
->length
= num_fields
;
1278 val
->type
->members
= ralloc_array(b
, struct vtn_type
*, num_fields
);
1279 val
->type
->offsets
= ralloc_array(b
, unsigned, num_fields
);
1280 val
->type
->packed
= false;
1282 NIR_VLA(struct glsl_struct_field
, fields
, count
);
1283 for (unsigned i
= 0; i
< num_fields
; i
++) {
1284 val
->type
->members
[i
] =
1285 vtn_value(b
, w
[i
+ 2], vtn_value_type_type
)->type
;
1286 fields
[i
] = (struct glsl_struct_field
) {
1287 .type
= val
->type
->members
[i
]->type
,
1288 .name
= ralloc_asprintf(b
, "field%d", i
),
1294 if (b
->shader
->info
.stage
== MESA_SHADER_KERNEL
) {
1295 unsigned offset
= 0;
1296 for (unsigned i
= 0; i
< num_fields
; i
++) {
1297 offset
= align(offset
, glsl_get_cl_alignment(fields
[i
].type
));
1298 fields
[i
].offset
= offset
;
1299 offset
+= glsl_get_cl_size(fields
[i
].type
);
1303 struct member_decoration_ctx ctx
= {
1304 .num_fields
= num_fields
,
1309 vtn_foreach_decoration(b
, val
, struct_member_decoration_cb
, &ctx
);
1310 vtn_foreach_decoration(b
, val
, struct_member_matrix_stride_cb
, &ctx
);
1312 vtn_foreach_decoration(b
, val
, struct_block_decoration_cb
, NULL
);
1314 const char *name
= val
->name
;
1316 if (val
->type
->block
|| val
->type
->buffer_block
) {
1317 /* Packing will be ignored since types coming from SPIR-V are
1318 * explicitly laid out.
1320 val
->type
->type
= glsl_interface_type(fields
, num_fields
,
1321 /* packing */ 0, false,
1322 name
? name
: "block");
1324 val
->type
->type
= glsl_struct_type(fields
, num_fields
,
1325 name
? name
: "struct", false);
1330 case SpvOpTypeFunction
: {
1331 val
->type
->base_type
= vtn_base_type_function
;
1332 val
->type
->type
= NULL
;
1334 val
->type
->return_type
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1336 const unsigned num_params
= count
- 3;
1337 val
->type
->length
= num_params
;
1338 val
->type
->params
= ralloc_array(b
, struct vtn_type
*, num_params
);
1339 for (unsigned i
= 0; i
< count
- 3; i
++) {
1340 val
->type
->params
[i
] =
1341 vtn_value(b
, w
[i
+ 3], vtn_value_type_type
)->type
;
1346 case SpvOpTypePointer
:
1347 case SpvOpTypeForwardPointer
: {
1348 /* We can't blindly push the value because it might be a forward
1351 val
= vtn_untyped_value(b
, w
[1]);
1353 SpvStorageClass storage_class
= w
[2];
1355 if (val
->value_type
== vtn_value_type_invalid
) {
1356 val
->value_type
= vtn_value_type_type
;
1357 val
->type
= rzalloc(b
, struct vtn_type
);
1358 val
->type
->id
= w
[1];
1359 val
->type
->base_type
= vtn_base_type_pointer
;
1360 val
->type
->storage_class
= storage_class
;
1362 /* These can actually be stored to nir_variables and used as SSA
1363 * values so they need a real glsl_type.
1365 switch (storage_class
) {
1366 case SpvStorageClassUniform
:
1367 val
->type
->type
= b
->options
->ubo_ptr_type
;
1369 case SpvStorageClassStorageBuffer
:
1370 val
->type
->type
= b
->options
->ssbo_ptr_type
;
1372 case SpvStorageClassPhysicalStorageBufferEXT
:
1373 val
->type
->type
= b
->options
->phys_ssbo_ptr_type
;
1375 case SpvStorageClassPushConstant
:
1376 val
->type
->type
= b
->options
->push_const_ptr_type
;
1378 case SpvStorageClassWorkgroup
:
1379 val
->type
->type
= b
->options
->shared_ptr_type
;
1381 case SpvStorageClassCrossWorkgroup
:
1382 val
->type
->type
= b
->options
->global_ptr_type
;
1384 case SpvStorageClassFunction
:
1385 if (b
->physical_ptrs
)
1386 val
->type
->type
= b
->options
->temp_ptr_type
;
1389 /* In this case, no variable pointers are allowed so all deref
1390 * chains are complete back to the variable and it doesn't matter
1391 * what type gets used so we leave it NULL.
1396 vtn_fail_if(val
->type
->storage_class
!= storage_class
,
1397 "The storage classes of an OpTypePointer and any "
1398 "OpTypeForwardPointers that provide forward "
1399 "declarations of it must match.");
1402 if (opcode
== SpvOpTypePointer
) {
1403 vtn_fail_if(val
->type
->deref
!= NULL
,
1404 "While OpTypeForwardPointer can be used to provide a "
1405 "forward declaration of a pointer, OpTypePointer can "
1406 "only be used once for a given id.");
1408 val
->type
->deref
= vtn_value(b
, w
[3], vtn_value_type_type
)->type
;
1410 vtn_foreach_decoration(b
, val
, array_stride_decoration_cb
, NULL
);
1412 if (b
->physical_ptrs
) {
1413 switch (storage_class
) {
1414 case SpvStorageClassFunction
:
1415 case SpvStorageClassWorkgroup
:
1416 case SpvStorageClassCrossWorkgroup
:
1417 val
->type
->stride
= align(glsl_get_cl_size(val
->type
->deref
->type
),
1418 glsl_get_cl_alignment(val
->type
->deref
->type
));
1425 if (storage_class
== SpvStorageClassWorkgroup
&&
1426 b
->options
->lower_workgroup_access_to_offsets
) {
1427 uint32_t size
, align
;
1428 val
->type
->deref
= vtn_type_layout_std430(b
, val
->type
->deref
,
1430 val
->type
->length
= size
;
1431 val
->type
->align
= align
;
1437 case SpvOpTypeImage
: {
1438 val
->type
->base_type
= vtn_base_type_image
;
1440 const struct vtn_type
*sampled_type
=
1441 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1443 vtn_fail_if(sampled_type
->base_type
!= vtn_base_type_scalar
||
1444 glsl_get_bit_size(sampled_type
->type
) != 32,
1445 "Sampled type of OpTypeImage must be a 32-bit scalar");
1447 enum glsl_sampler_dim dim
;
1448 switch ((SpvDim
)w
[3]) {
1449 case SpvDim1D
: dim
= GLSL_SAMPLER_DIM_1D
; break;
1450 case SpvDim2D
: dim
= GLSL_SAMPLER_DIM_2D
; break;
1451 case SpvDim3D
: dim
= GLSL_SAMPLER_DIM_3D
; break;
1452 case SpvDimCube
: dim
= GLSL_SAMPLER_DIM_CUBE
; break;
1453 case SpvDimRect
: dim
= GLSL_SAMPLER_DIM_RECT
; break;
1454 case SpvDimBuffer
: dim
= GLSL_SAMPLER_DIM_BUF
; break;
1455 case SpvDimSubpassData
: dim
= GLSL_SAMPLER_DIM_SUBPASS
; break;
1457 vtn_fail("Invalid SPIR-V image dimensionality: %s (%u)",
1458 spirv_dim_to_string((SpvDim
)w
[3]), w
[3]);
1461 /* w[4]: as per Vulkan spec "Validation Rules within a Module",
1462 * The “Depth” operand of OpTypeImage is ignored.
1464 bool is_array
= w
[5];
1465 bool multisampled
= w
[6];
1466 unsigned sampled
= w
[7];
1467 SpvImageFormat format
= w
[8];
1470 val
->type
->access_qualifier
= w
[9];
1472 val
->type
->access_qualifier
= SpvAccessQualifierReadWrite
;
1475 if (dim
== GLSL_SAMPLER_DIM_2D
)
1476 dim
= GLSL_SAMPLER_DIM_MS
;
1477 else if (dim
== GLSL_SAMPLER_DIM_SUBPASS
)
1478 dim
= GLSL_SAMPLER_DIM_SUBPASS_MS
;
1480 vtn_fail("Unsupported multisampled image type");
1483 val
->type
->image_format
= translate_image_format(b
, format
);
1485 enum glsl_base_type sampled_base_type
=
1486 glsl_get_base_type(sampled_type
->type
);
1488 val
->type
->sampled
= true;
1489 val
->type
->type
= glsl_sampler_type(dim
, false, is_array
,
1491 } else if (sampled
== 2) {
1492 val
->type
->sampled
= false;
1493 val
->type
->type
= glsl_image_type(dim
, is_array
, sampled_base_type
);
1495 vtn_fail("We need to know if the image will be sampled");
1500 case SpvOpTypeSampledImage
:
1501 val
->type
->base_type
= vtn_base_type_sampled_image
;
1502 val
->type
->image
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1503 val
->type
->type
= val
->type
->image
->type
;
1506 case SpvOpTypeSampler
:
1507 /* The actual sampler type here doesn't really matter. It gets
1508 * thrown away the moment you combine it with an image. What really
1509 * matters is that it's a sampler type as opposed to an integer type
1510 * so the backend knows what to do.
1512 val
->type
->base_type
= vtn_base_type_sampler
;
1513 val
->type
->type
= glsl_bare_sampler_type();
1516 case SpvOpTypeOpaque
:
1517 case SpvOpTypeEvent
:
1518 case SpvOpTypeDeviceEvent
:
1519 case SpvOpTypeReserveId
:
1520 case SpvOpTypeQueue
:
1523 vtn_fail_with_opcode("Unhandled opcode", opcode
);
1526 vtn_foreach_decoration(b
, val
, type_decoration_cb
, NULL
);
1528 if (val
->type
->base_type
== vtn_base_type_struct
&&
1529 (val
->type
->block
|| val
->type
->buffer_block
)) {
1530 for (unsigned i
= 0; i
< val
->type
->length
; i
++) {
1531 vtn_fail_if(vtn_type_contains_block(b
, val
->type
->members
[i
]),
1532 "Block and BufferBlock decorations cannot decorate a "
1533 "structure type that is nested at any level inside "
1534 "another structure type decorated with Block or "
1540 static nir_constant
*
1541 vtn_null_constant(struct vtn_builder
*b
, const struct glsl_type
*type
)
1543 nir_constant
*c
= rzalloc(b
, nir_constant
);
1545 /* For pointers and other typeless things, we have to return something but
1546 * it doesn't matter what.
1551 switch (glsl_get_base_type(type
)) {
1553 case GLSL_TYPE_UINT
:
1554 case GLSL_TYPE_INT16
:
1555 case GLSL_TYPE_UINT16
:
1556 case GLSL_TYPE_UINT8
:
1557 case GLSL_TYPE_INT8
:
1558 case GLSL_TYPE_INT64
:
1559 case GLSL_TYPE_UINT64
:
1560 case GLSL_TYPE_BOOL
:
1561 case GLSL_TYPE_FLOAT
:
1562 case GLSL_TYPE_FLOAT16
:
1563 case GLSL_TYPE_DOUBLE
:
1564 /* Nothing to do here. It's already initialized to zero */
1567 case GLSL_TYPE_ARRAY
:
1568 vtn_assert(glsl_get_length(type
) > 0);
1569 c
->num_elements
= glsl_get_length(type
);
1570 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1572 c
->elements
[0] = vtn_null_constant(b
, glsl_get_array_element(type
));
1573 for (unsigned i
= 1; i
< c
->num_elements
; i
++)
1574 c
->elements
[i
] = c
->elements
[0];
1577 case GLSL_TYPE_STRUCT
:
1578 c
->num_elements
= glsl_get_length(type
);
1579 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1581 for (unsigned i
= 0; i
< c
->num_elements
; i
++) {
1582 c
->elements
[i
] = vtn_null_constant(b
, glsl_get_struct_field(type
, i
));
1587 vtn_fail("Invalid type for null constant");
1594 spec_constant_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*v
,
1595 int member
, const struct vtn_decoration
*dec
,
1598 vtn_assert(member
== -1);
1599 if (dec
->decoration
!= SpvDecorationSpecId
)
1602 struct spec_constant_value
*const_value
= data
;
1604 for (unsigned i
= 0; i
< b
->num_specializations
; i
++) {
1605 if (b
->specializations
[i
].id
== dec
->literals
[0]) {
1606 if (const_value
->is_double
)
1607 const_value
->data64
= b
->specializations
[i
].data64
;
1609 const_value
->data32
= b
->specializations
[i
].data32
;
1616 get_specialization(struct vtn_builder
*b
, struct vtn_value
*val
,
1617 uint32_t const_value
)
1619 struct spec_constant_value data
;
1620 data
.is_double
= false;
1621 data
.data32
= const_value
;
1622 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1627 get_specialization64(struct vtn_builder
*b
, struct vtn_value
*val
,
1628 uint64_t const_value
)
1630 struct spec_constant_value data
;
1631 data
.is_double
= true;
1632 data
.data64
= const_value
;
1633 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1638 handle_workgroup_size_decoration_cb(struct vtn_builder
*b
,
1639 struct vtn_value
*val
,
1641 const struct vtn_decoration
*dec
,
1644 vtn_assert(member
== -1);
1645 if (dec
->decoration
!= SpvDecorationBuiltIn
||
1646 dec
->literals
[0] != SpvBuiltInWorkgroupSize
)
1649 vtn_assert(val
->type
->type
== glsl_vector_type(GLSL_TYPE_UINT
, 3));
1650 b
->workgroup_size_builtin
= val
;
1654 vtn_handle_constant(struct vtn_builder
*b
, SpvOp opcode
,
1655 const uint32_t *w
, unsigned count
)
1657 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_constant
);
1658 val
->constant
= rzalloc(b
, nir_constant
);
1660 case SpvOpConstantTrue
:
1661 case SpvOpConstantFalse
:
1662 case SpvOpSpecConstantTrue
:
1663 case SpvOpSpecConstantFalse
: {
1664 vtn_fail_if(val
->type
->type
!= glsl_bool_type(),
1665 "Result type of %s must be OpTypeBool",
1666 spirv_op_to_string(opcode
));
1668 uint32_t int_val
= (opcode
== SpvOpConstantTrue
||
1669 opcode
== SpvOpSpecConstantTrue
);
1671 if (opcode
== SpvOpSpecConstantTrue
||
1672 opcode
== SpvOpSpecConstantFalse
)
1673 int_val
= get_specialization(b
, val
, int_val
);
1675 val
->constant
->values
[0][0].b
= int_val
!= 0;
1679 case SpvOpConstant
: {
1680 vtn_fail_if(val
->type
->base_type
!= vtn_base_type_scalar
,
1681 "Result type of %s must be a scalar",
1682 spirv_op_to_string(opcode
));
1683 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1686 val
->constant
->values
[0][0].u64
= vtn_u64_literal(&w
[3]);
1689 val
->constant
->values
[0][0].u32
= w
[3];
1692 val
->constant
->values
[0][0].u16
= w
[3];
1695 val
->constant
->values
[0][0].u8
= w
[3];
1698 vtn_fail("Unsupported SpvOpConstant bit size: %u", bit_size
);
1703 case SpvOpSpecConstant
: {
1704 vtn_fail_if(val
->type
->base_type
!= vtn_base_type_scalar
,
1705 "Result type of %s must be a scalar",
1706 spirv_op_to_string(opcode
));
1707 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1710 val
->constant
->values
[0][0].u64
=
1711 get_specialization64(b
, val
, vtn_u64_literal(&w
[3]));
1714 val
->constant
->values
[0][0].u32
= get_specialization(b
, val
, w
[3]);
1717 val
->constant
->values
[0][0].u16
= get_specialization(b
, val
, w
[3]);
1720 val
->constant
->values
[0][0].u8
= get_specialization(b
, val
, w
[3]);
1723 vtn_fail("Unsupported SpvOpSpecConstant bit size");
1728 case SpvOpSpecConstantComposite
:
1729 case SpvOpConstantComposite
: {
1730 unsigned elem_count
= count
- 3;
1731 vtn_fail_if(elem_count
!= val
->type
->length
,
1732 "%s has %u constituents, expected %u",
1733 spirv_op_to_string(opcode
), elem_count
, val
->type
->length
);
1735 nir_constant
**elems
= ralloc_array(b
, nir_constant
*, elem_count
);
1736 for (unsigned i
= 0; i
< elem_count
; i
++) {
1737 struct vtn_value
*val
= vtn_untyped_value(b
, w
[i
+ 3]);
1739 if (val
->value_type
== vtn_value_type_constant
) {
1740 elems
[i
] = val
->constant
;
1742 vtn_fail_if(val
->value_type
!= vtn_value_type_undef
,
1743 "only constants or undefs allowed for "
1744 "SpvOpConstantComposite");
1745 /* to make it easier, just insert a NULL constant for now */
1746 elems
[i
] = vtn_null_constant(b
, val
->type
->type
);
1750 switch (val
->type
->base_type
) {
1751 case vtn_base_type_vector
: {
1752 assert(glsl_type_is_vector(val
->type
->type
));
1753 for (unsigned i
= 0; i
< elem_count
; i
++)
1754 val
->constant
->values
[0][i
] = elems
[i
]->values
[0][0];
1758 case vtn_base_type_matrix
:
1759 assert(glsl_type_is_matrix(val
->type
->type
));
1760 for (unsigned i
= 0; i
< elem_count
; i
++) {
1761 unsigned components
=
1762 glsl_get_components(glsl_get_column_type(val
->type
->type
));
1763 memcpy(val
->constant
->values
[i
], elems
[i
]->values
,
1764 sizeof(nir_const_value
) * components
);
1768 case vtn_base_type_struct
:
1769 case vtn_base_type_array
:
1770 ralloc_steal(val
->constant
, elems
);
1771 val
->constant
->num_elements
= elem_count
;
1772 val
->constant
->elements
= elems
;
1776 vtn_fail("Result type of %s must be a composite type",
1777 spirv_op_to_string(opcode
));
1782 case SpvOpSpecConstantOp
: {
1783 SpvOp opcode
= get_specialization(b
, val
, w
[3]);
1785 case SpvOpVectorShuffle
: {
1786 struct vtn_value
*v0
= &b
->values
[w
[4]];
1787 struct vtn_value
*v1
= &b
->values
[w
[5]];
1789 vtn_assert(v0
->value_type
== vtn_value_type_constant
||
1790 v0
->value_type
== vtn_value_type_undef
);
1791 vtn_assert(v1
->value_type
== vtn_value_type_constant
||
1792 v1
->value_type
== vtn_value_type_undef
);
1794 unsigned len0
= glsl_get_vector_elements(v0
->type
->type
);
1795 unsigned len1
= glsl_get_vector_elements(v1
->type
->type
);
1797 vtn_assert(len0
+ len1
< 16);
1799 unsigned bit_size
= glsl_get_bit_size(val
->type
->type
);
1800 unsigned bit_size0
= glsl_get_bit_size(v0
->type
->type
);
1801 unsigned bit_size1
= glsl_get_bit_size(v1
->type
->type
);
1803 vtn_assert(bit_size
== bit_size0
&& bit_size
== bit_size1
);
1804 (void)bit_size0
; (void)bit_size1
;
1806 if (bit_size
== 64) {
1808 if (v0
->value_type
== vtn_value_type_constant
) {
1809 for (unsigned i
= 0; i
< len0
; i
++)
1810 u64
[i
] = v0
->constant
->values
[0][i
].u64
;
1812 if (v1
->value_type
== vtn_value_type_constant
) {
1813 for (unsigned i
= 0; i
< len1
; i
++)
1814 u64
[len0
+ i
] = v1
->constant
->values
[0][i
].u64
;
1817 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1818 uint32_t comp
= w
[i
+ 6];
1819 /* If component is not used, set the value to a known constant
1820 * to detect if it is wrongly used.
1822 if (comp
== (uint32_t)-1)
1823 val
->constant
->values
[0][j
].u64
= 0xdeadbeefdeadbeef;
1825 val
->constant
->values
[0][j
].u64
= u64
[comp
];
1828 /* This is for both 32-bit and 16-bit values */
1830 if (v0
->value_type
== vtn_value_type_constant
) {
1831 for (unsigned i
= 0; i
< len0
; i
++)
1832 u32
[i
] = v0
->constant
->values
[0][i
].u32
;
1834 if (v1
->value_type
== vtn_value_type_constant
) {
1835 for (unsigned i
= 0; i
< len1
; i
++)
1836 u32
[len0
+ i
] = v1
->constant
->values
[0][i
].u32
;
1839 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1840 uint32_t comp
= w
[i
+ 6];
1841 /* If component is not used, set the value to a known constant
1842 * to detect if it is wrongly used.
1844 if (comp
== (uint32_t)-1)
1845 val
->constant
->values
[0][j
].u32
= 0xdeadbeef;
1847 val
->constant
->values
[0][j
].u32
= u32
[comp
];
1853 case SpvOpCompositeExtract
:
1854 case SpvOpCompositeInsert
: {
1855 struct vtn_value
*comp
;
1856 unsigned deref_start
;
1857 struct nir_constant
**c
;
1858 if (opcode
== SpvOpCompositeExtract
) {
1859 comp
= vtn_value(b
, w
[4], vtn_value_type_constant
);
1861 c
= &comp
->constant
;
1863 comp
= vtn_value(b
, w
[5], vtn_value_type_constant
);
1865 val
->constant
= nir_constant_clone(comp
->constant
,
1872 const struct vtn_type
*type
= comp
->type
;
1873 for (unsigned i
= deref_start
; i
< count
; i
++) {
1874 vtn_fail_if(w
[i
] > type
->length
,
1875 "%uth index of %s is %u but the type has only "
1876 "%u elements", i
- deref_start
,
1877 spirv_op_to_string(opcode
), w
[i
], type
->length
);
1879 switch (type
->base_type
) {
1880 case vtn_base_type_vector
:
1882 type
= type
->array_element
;
1885 case vtn_base_type_matrix
:
1886 assert(col
== 0 && elem
== -1);
1889 type
= type
->array_element
;
1892 case vtn_base_type_array
:
1893 c
= &(*c
)->elements
[w
[i
]];
1894 type
= type
->array_element
;
1897 case vtn_base_type_struct
:
1898 c
= &(*c
)->elements
[w
[i
]];
1899 type
= type
->members
[w
[i
]];
1903 vtn_fail("%s must only index into composite types",
1904 spirv_op_to_string(opcode
));
1908 if (opcode
== SpvOpCompositeExtract
) {
1912 unsigned num_components
= type
->length
;
1913 for (unsigned i
= 0; i
< num_components
; i
++)
1914 val
->constant
->values
[0][i
] = (*c
)->values
[col
][elem
+ i
];
1917 struct vtn_value
*insert
=
1918 vtn_value(b
, w
[4], vtn_value_type_constant
);
1919 vtn_assert(insert
->type
== type
);
1921 *c
= insert
->constant
;
1923 unsigned num_components
= type
->length
;
1924 for (unsigned i
= 0; i
< num_components
; i
++)
1925 (*c
)->values
[col
][elem
+ i
] = insert
->constant
->values
[0][i
];
1933 nir_alu_type dst_alu_type
= nir_get_nir_type_for_glsl_type(val
->type
->type
);
1934 nir_alu_type src_alu_type
= dst_alu_type
;
1935 unsigned num_components
= glsl_get_vector_elements(val
->type
->type
);
1938 vtn_assert(count
<= 7);
1943 /* We have a source in a conversion */
1945 nir_get_nir_type_for_glsl_type(
1946 vtn_value(b
, w
[4], vtn_value_type_constant
)->type
->type
);
1947 /* We use the bitsize of the conversion source to evaluate the opcode later */
1948 bit_size
= glsl_get_bit_size(
1949 vtn_value(b
, w
[4], vtn_value_type_constant
)->type
->type
);
1952 bit_size
= glsl_get_bit_size(val
->type
->type
);
1955 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(b
, opcode
, &swap
,
1956 nir_alu_type_get_type_size(src_alu_type
),
1957 nir_alu_type_get_type_size(dst_alu_type
));
1958 nir_const_value src
[3][NIR_MAX_VEC_COMPONENTS
];
1960 for (unsigned i
= 0; i
< count
- 4; i
++) {
1961 struct vtn_value
*src_val
=
1962 vtn_value(b
, w
[4 + i
], vtn_value_type_constant
);
1964 /* If this is an unsized source, pull the bit size from the
1965 * source; otherwise, we'll use the bit size from the destination.
1967 if (!nir_alu_type_get_type_size(nir_op_infos
[op
].input_types
[i
]))
1968 bit_size
= glsl_get_bit_size(src_val
->type
->type
);
1970 unsigned j
= swap
? 1 - i
: i
;
1971 memcpy(src
[j
], src_val
->constant
->values
[0], sizeof(src
[j
]));
1974 /* fix up fixed size sources */
1981 for (unsigned i
= 0; i
< num_components
; ++i
) {
1983 case 64: src
[1][i
].u32
= src
[1][i
].u64
; break;
1984 case 16: src
[1][i
].u32
= src
[1][i
].u16
; break;
1985 case 8: src
[1][i
].u32
= src
[1][i
].u8
; break;
1994 nir_const_value
*srcs
[3] = {
1995 src
[0], src
[1], src
[2],
1997 nir_eval_const_opcode(op
, val
->constant
->values
[0], num_components
, bit_size
, srcs
);
2004 case SpvOpConstantNull
:
2005 val
->constant
= vtn_null_constant(b
, val
->type
->type
);
2008 case SpvOpConstantSampler
:
2009 vtn_fail("OpConstantSampler requires Kernel Capability");
2013 vtn_fail_with_opcode("Unhandled opcode", opcode
);
2016 /* Now that we have the value, update the workgroup size if needed */
2017 vtn_foreach_decoration(b
, val
, handle_workgroup_size_decoration_cb
, NULL
);
2020 struct vtn_ssa_value
*
2021 vtn_create_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
2023 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
2026 if (!glsl_type_is_vector_or_scalar(type
)) {
2027 unsigned elems
= glsl_get_length(type
);
2028 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
2029 for (unsigned i
= 0; i
< elems
; i
++) {
2030 const struct glsl_type
*child_type
;
2032 switch (glsl_get_base_type(type
)) {
2034 case GLSL_TYPE_UINT
:
2035 case GLSL_TYPE_INT16
:
2036 case GLSL_TYPE_UINT16
:
2037 case GLSL_TYPE_UINT8
:
2038 case GLSL_TYPE_INT8
:
2039 case GLSL_TYPE_INT64
:
2040 case GLSL_TYPE_UINT64
:
2041 case GLSL_TYPE_BOOL
:
2042 case GLSL_TYPE_FLOAT
:
2043 case GLSL_TYPE_FLOAT16
:
2044 case GLSL_TYPE_DOUBLE
:
2045 child_type
= glsl_get_column_type(type
);
2047 case GLSL_TYPE_ARRAY
:
2048 child_type
= glsl_get_array_element(type
);
2050 case GLSL_TYPE_STRUCT
:
2051 case GLSL_TYPE_INTERFACE
:
2052 child_type
= glsl_get_struct_field(type
, i
);
2055 vtn_fail("unkown base type");
2058 val
->elems
[i
] = vtn_create_ssa_value(b
, child_type
);
2066 vtn_tex_src(struct vtn_builder
*b
, unsigned index
, nir_tex_src_type type
)
2069 src
.src
= nir_src_for_ssa(vtn_ssa_value(b
, index
)->def
);
2070 src
.src_type
= type
;
2075 vtn_handle_texture(struct vtn_builder
*b
, SpvOp opcode
,
2076 const uint32_t *w
, unsigned count
)
2078 if (opcode
== SpvOpSampledImage
) {
2079 struct vtn_value
*val
=
2080 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
2081 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
2082 val
->sampled_image
->type
=
2083 vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2084 val
->sampled_image
->image
=
2085 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2086 val
->sampled_image
->sampler
=
2087 vtn_value(b
, w
[4], vtn_value_type_pointer
)->pointer
;
2089 } else if (opcode
== SpvOpImage
) {
2090 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
2091 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[3]);
2092 if (src_val
->value_type
== vtn_value_type_sampled_image
) {
2093 val
->pointer
= src_val
->sampled_image
->image
;
2095 vtn_assert(src_val
->value_type
== vtn_value_type_pointer
);
2096 val
->pointer
= src_val
->pointer
;
2101 struct vtn_type
*ret_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2102 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2104 struct vtn_sampled_image sampled
;
2105 struct vtn_value
*sampled_val
= vtn_untyped_value(b
, w
[3]);
2106 if (sampled_val
->value_type
== vtn_value_type_sampled_image
) {
2107 sampled
= *sampled_val
->sampled_image
;
2109 vtn_assert(sampled_val
->value_type
== vtn_value_type_pointer
);
2110 sampled
.type
= sampled_val
->pointer
->type
;
2111 sampled
.image
= NULL
;
2112 sampled
.sampler
= sampled_val
->pointer
;
2115 const struct glsl_type
*image_type
= sampled
.type
->type
;
2116 const enum glsl_sampler_dim sampler_dim
= glsl_get_sampler_dim(image_type
);
2117 const bool is_array
= glsl_sampler_type_is_array(image_type
);
2119 /* Figure out the base texture operation */
2122 case SpvOpImageSampleImplicitLod
:
2123 case SpvOpImageSampleDrefImplicitLod
:
2124 case SpvOpImageSampleProjImplicitLod
:
2125 case SpvOpImageSampleProjDrefImplicitLod
:
2126 texop
= nir_texop_tex
;
2129 case SpvOpImageSampleExplicitLod
:
2130 case SpvOpImageSampleDrefExplicitLod
:
2131 case SpvOpImageSampleProjExplicitLod
:
2132 case SpvOpImageSampleProjDrefExplicitLod
:
2133 texop
= nir_texop_txl
;
2136 case SpvOpImageFetch
:
2137 if (glsl_get_sampler_dim(image_type
) == GLSL_SAMPLER_DIM_MS
) {
2138 texop
= nir_texop_txf_ms
;
2140 texop
= nir_texop_txf
;
2144 case SpvOpImageGather
:
2145 case SpvOpImageDrefGather
:
2146 texop
= nir_texop_tg4
;
2149 case SpvOpImageQuerySizeLod
:
2150 case SpvOpImageQuerySize
:
2151 texop
= nir_texop_txs
;
2154 case SpvOpImageQueryLod
:
2155 texop
= nir_texop_lod
;
2158 case SpvOpImageQueryLevels
:
2159 texop
= nir_texop_query_levels
;
2162 case SpvOpImageQuerySamples
:
2163 texop
= nir_texop_texture_samples
;
2167 vtn_fail_with_opcode("Unhandled opcode", opcode
);
2170 nir_tex_src srcs
[10]; /* 10 should be enough */
2171 nir_tex_src
*p
= srcs
;
2173 nir_deref_instr
*sampler
= vtn_pointer_to_deref(b
, sampled
.sampler
);
2174 nir_deref_instr
*texture
=
2175 sampled
.image
? vtn_pointer_to_deref(b
, sampled
.image
) : sampler
;
2177 p
->src
= nir_src_for_ssa(&texture
->dest
.ssa
);
2178 p
->src_type
= nir_tex_src_texture_deref
;
2188 /* These operations require a sampler */
2189 p
->src
= nir_src_for_ssa(&sampler
->dest
.ssa
);
2190 p
->src_type
= nir_tex_src_sampler_deref
;
2194 case nir_texop_txf_ms
:
2196 case nir_texop_query_levels
:
2197 case nir_texop_texture_samples
:
2198 case nir_texop_samples_identical
:
2201 case nir_texop_txf_ms_mcs
:
2202 vtn_fail("unexpected nir_texop_txf_ms_mcs");
2207 struct nir_ssa_def
*coord
;
2208 unsigned coord_components
;
2210 case SpvOpImageSampleImplicitLod
:
2211 case SpvOpImageSampleExplicitLod
:
2212 case SpvOpImageSampleDrefImplicitLod
:
2213 case SpvOpImageSampleDrefExplicitLod
:
2214 case SpvOpImageSampleProjImplicitLod
:
2215 case SpvOpImageSampleProjExplicitLod
:
2216 case SpvOpImageSampleProjDrefImplicitLod
:
2217 case SpvOpImageSampleProjDrefExplicitLod
:
2218 case SpvOpImageFetch
:
2219 case SpvOpImageGather
:
2220 case SpvOpImageDrefGather
:
2221 case SpvOpImageQueryLod
: {
2222 /* All these types have the coordinate as their first real argument */
2223 switch (sampler_dim
) {
2224 case GLSL_SAMPLER_DIM_1D
:
2225 case GLSL_SAMPLER_DIM_BUF
:
2226 coord_components
= 1;
2228 case GLSL_SAMPLER_DIM_2D
:
2229 case GLSL_SAMPLER_DIM_RECT
:
2230 case GLSL_SAMPLER_DIM_MS
:
2231 coord_components
= 2;
2233 case GLSL_SAMPLER_DIM_3D
:
2234 case GLSL_SAMPLER_DIM_CUBE
:
2235 coord_components
= 3;
2238 vtn_fail("Invalid sampler type");
2241 if (is_array
&& texop
!= nir_texop_lod
)
2244 coord
= vtn_ssa_value(b
, w
[idx
++])->def
;
2245 p
->src
= nir_src_for_ssa(nir_channels(&b
->nb
, coord
,
2246 (1 << coord_components
) - 1));
2247 p
->src_type
= nir_tex_src_coord
;
2254 coord_components
= 0;
2259 case SpvOpImageSampleProjImplicitLod
:
2260 case SpvOpImageSampleProjExplicitLod
:
2261 case SpvOpImageSampleProjDrefImplicitLod
:
2262 case SpvOpImageSampleProjDrefExplicitLod
:
2263 /* These have the projector as the last coordinate component */
2264 p
->src
= nir_src_for_ssa(nir_channel(&b
->nb
, coord
, coord_components
));
2265 p
->src_type
= nir_tex_src_projector
;
2273 bool is_shadow
= false;
2274 unsigned gather_component
= 0;
2276 case SpvOpImageSampleDrefImplicitLod
:
2277 case SpvOpImageSampleDrefExplicitLod
:
2278 case SpvOpImageSampleProjDrefImplicitLod
:
2279 case SpvOpImageSampleProjDrefExplicitLod
:
2280 case SpvOpImageDrefGather
:
2281 /* These all have an explicit depth value as their next source */
2283 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_comparator
);
2286 case SpvOpImageGather
:
2287 /* This has a component as its next source */
2289 vtn_value(b
, w
[idx
++], vtn_value_type_constant
)->constant
->values
[0][0].u32
;
2296 /* For OpImageQuerySizeLod, we always have an LOD */
2297 if (opcode
== SpvOpImageQuerySizeLod
)
2298 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
2300 /* Now we need to handle some number of optional arguments */
2301 struct vtn_value
*gather_offsets
= NULL
;
2303 uint32_t operands
= w
[idx
++];
2305 if (operands
& SpvImageOperandsBiasMask
) {
2306 vtn_assert(texop
== nir_texop_tex
);
2307 texop
= nir_texop_txb
;
2308 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_bias
);
2311 if (operands
& SpvImageOperandsLodMask
) {
2312 vtn_assert(texop
== nir_texop_txl
|| texop
== nir_texop_txf
||
2313 texop
== nir_texop_txs
);
2314 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
2317 if (operands
& SpvImageOperandsGradMask
) {
2318 vtn_assert(texop
== nir_texop_txl
);
2319 texop
= nir_texop_txd
;
2320 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddx
);
2321 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddy
);
2324 if (operands
& SpvImageOperandsOffsetMask
||
2325 operands
& SpvImageOperandsConstOffsetMask
)
2326 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_offset
);
2328 if (operands
& SpvImageOperandsConstOffsetsMask
) {
2329 vtn_assert(texop
== nir_texop_tg4
);
2330 gather_offsets
= vtn_value(b
, w
[idx
++], vtn_value_type_constant
);
2333 if (operands
& SpvImageOperandsSampleMask
) {
2334 vtn_assert(texop
== nir_texop_txf_ms
);
2335 texop
= nir_texop_txf_ms
;
2336 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ms_index
);
2339 if (operands
& SpvImageOperandsMinLodMask
) {
2340 vtn_assert(texop
== nir_texop_tex
||
2341 texop
== nir_texop_txb
||
2342 texop
== nir_texop_txd
);
2343 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_min_lod
);
2346 /* We should have now consumed exactly all of the arguments */
2347 vtn_assert(idx
== count
);
2349 nir_tex_instr
*instr
= nir_tex_instr_create(b
->shader
, p
- srcs
);
2352 memcpy(instr
->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
2354 instr
->coord_components
= coord_components
;
2355 instr
->sampler_dim
= sampler_dim
;
2356 instr
->is_array
= is_array
;
2357 instr
->is_shadow
= is_shadow
;
2358 instr
->is_new_style_shadow
=
2359 is_shadow
&& glsl_get_components(ret_type
->type
) == 1;
2360 instr
->component
= gather_component
;
2362 if (sampled
.image
&& (sampled
.image
->access
& ACCESS_NON_UNIFORM
))
2363 instr
->texture_non_uniform
= true;
2365 if (sampled
.sampler
&& (sampled
.sampler
->access
& ACCESS_NON_UNIFORM
))
2366 instr
->sampler_non_uniform
= true;
2368 switch (glsl_get_sampler_result_type(image_type
)) {
2369 case GLSL_TYPE_FLOAT
: instr
->dest_type
= nir_type_float
; break;
2370 case GLSL_TYPE_INT
: instr
->dest_type
= nir_type_int
; break;
2371 case GLSL_TYPE_UINT
: instr
->dest_type
= nir_type_uint
; break;
2372 case GLSL_TYPE_BOOL
: instr
->dest_type
= nir_type_bool
; break;
2374 vtn_fail("Invalid base type for sampler result");
2377 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
2378 nir_tex_instr_dest_size(instr
), 32, NULL
);
2380 vtn_assert(glsl_get_vector_elements(ret_type
->type
) ==
2381 nir_tex_instr_dest_size(instr
));
2383 if (gather_offsets
) {
2384 vtn_fail_if(gather_offsets
->type
->base_type
!= vtn_base_type_array
||
2385 gather_offsets
->type
->length
!= 4,
2386 "ConstOffsets must be an array of size four of vectors "
2387 "of two integer components");
2389 struct vtn_type
*vec_type
= gather_offsets
->type
->array_element
;
2390 vtn_fail_if(vec_type
->base_type
!= vtn_base_type_vector
||
2391 vec_type
->length
!= 2 ||
2392 !glsl_type_is_integer(vec_type
->type
),
2393 "ConstOffsets must be an array of size four of vectors "
2394 "of two integer components");
2396 unsigned bit_size
= glsl_get_bit_size(vec_type
->type
);
2397 for (uint32_t i
= 0; i
< 4; i
++) {
2398 const nir_const_value
*cvec
=
2399 gather_offsets
->constant
->elements
[i
]->values
[0];
2400 for (uint32_t j
= 0; j
< 2; j
++) {
2402 case 8: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i8
; break;
2403 case 16: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i16
; break;
2404 case 32: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i32
; break;
2405 case 64: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i64
; break;
2407 vtn_fail("Unsupported bit size: %u", bit_size
);
2413 val
->ssa
= vtn_create_ssa_value(b
, ret_type
->type
);
2414 val
->ssa
->def
= &instr
->dest
.ssa
;
2416 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
2420 fill_common_atomic_sources(struct vtn_builder
*b
, SpvOp opcode
,
2421 const uint32_t *w
, nir_src
*src
)
2424 case SpvOpAtomicIIncrement
:
2425 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
2428 case SpvOpAtomicIDecrement
:
2429 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
2432 case SpvOpAtomicISub
:
2434 nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
2437 case SpvOpAtomicCompareExchange
:
2438 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[8])->def
);
2439 src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
2442 case SpvOpAtomicExchange
:
2443 case SpvOpAtomicIAdd
:
2444 case SpvOpAtomicSMin
:
2445 case SpvOpAtomicUMin
:
2446 case SpvOpAtomicSMax
:
2447 case SpvOpAtomicUMax
:
2448 case SpvOpAtomicAnd
:
2450 case SpvOpAtomicXor
:
2451 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
2455 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
2459 static nir_ssa_def
*
2460 get_image_coord(struct vtn_builder
*b
, uint32_t value
)
2462 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, value
);
2464 /* The image_load_store intrinsics assume a 4-dim coordinate */
2465 unsigned dim
= glsl_get_vector_elements(coord
->type
);
2466 unsigned swizzle
[4];
2467 for (unsigned i
= 0; i
< 4; i
++)
2468 swizzle
[i
] = MIN2(i
, dim
- 1);
2470 return nir_swizzle(&b
->nb
, coord
->def
, swizzle
, 4, false);
2473 static nir_ssa_def
*
2474 expand_to_vec4(nir_builder
*b
, nir_ssa_def
*value
)
2476 if (value
->num_components
== 4)
2480 for (unsigned i
= 0; i
< 4; i
++)
2481 swiz
[i
] = i
< value
->num_components
? i
: 0;
2482 return nir_swizzle(b
, value
, swiz
, 4, false);
2486 vtn_handle_image(struct vtn_builder
*b
, SpvOp opcode
,
2487 const uint32_t *w
, unsigned count
)
2489 /* Just get this one out of the way */
2490 if (opcode
== SpvOpImageTexelPointer
) {
2491 struct vtn_value
*val
=
2492 vtn_push_value(b
, w
[2], vtn_value_type_image_pointer
);
2493 val
->image
= ralloc(b
, struct vtn_image_pointer
);
2495 val
->image
->image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2496 val
->image
->coord
= get_image_coord(b
, w
[4]);
2497 val
->image
->sample
= vtn_ssa_value(b
, w
[5])->def
;
2501 struct vtn_image_pointer image
;
2504 case SpvOpAtomicExchange
:
2505 case SpvOpAtomicCompareExchange
:
2506 case SpvOpAtomicCompareExchangeWeak
:
2507 case SpvOpAtomicIIncrement
:
2508 case SpvOpAtomicIDecrement
:
2509 case SpvOpAtomicIAdd
:
2510 case SpvOpAtomicISub
:
2511 case SpvOpAtomicLoad
:
2512 case SpvOpAtomicSMin
:
2513 case SpvOpAtomicUMin
:
2514 case SpvOpAtomicSMax
:
2515 case SpvOpAtomicUMax
:
2516 case SpvOpAtomicAnd
:
2518 case SpvOpAtomicXor
:
2519 image
= *vtn_value(b
, w
[3], vtn_value_type_image_pointer
)->image
;
2522 case SpvOpAtomicStore
:
2523 image
= *vtn_value(b
, w
[1], vtn_value_type_image_pointer
)->image
;
2526 case SpvOpImageQuerySize
:
2527 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2529 image
.sample
= NULL
;
2532 case SpvOpImageRead
:
2533 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2534 image
.coord
= get_image_coord(b
, w
[4]);
2536 if (count
> 5 && (w
[5] & SpvImageOperandsSampleMask
)) {
2537 vtn_assert(w
[5] == SpvImageOperandsSampleMask
);
2538 image
.sample
= vtn_ssa_value(b
, w
[6])->def
;
2540 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2544 case SpvOpImageWrite
:
2545 image
.image
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
2546 image
.coord
= get_image_coord(b
, w
[2]);
2550 if (count
> 4 && (w
[4] & SpvImageOperandsSampleMask
)) {
2551 vtn_assert(w
[4] == SpvImageOperandsSampleMask
);
2552 image
.sample
= vtn_ssa_value(b
, w
[5])->def
;
2554 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2559 vtn_fail_with_opcode("Invalid image opcode", opcode
);
2562 nir_intrinsic_op op
;
2564 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_deref_##N; break;
2565 OP(ImageQuerySize
, size
)
2567 OP(ImageWrite
, store
)
2568 OP(AtomicLoad
, load
)
2569 OP(AtomicStore
, store
)
2570 OP(AtomicExchange
, atomic_exchange
)
2571 OP(AtomicCompareExchange
, atomic_comp_swap
)
2572 OP(AtomicIIncrement
, atomic_add
)
2573 OP(AtomicIDecrement
, atomic_add
)
2574 OP(AtomicIAdd
, atomic_add
)
2575 OP(AtomicISub
, atomic_add
)
2576 OP(AtomicSMin
, atomic_min
)
2577 OP(AtomicUMin
, atomic_min
)
2578 OP(AtomicSMax
, atomic_max
)
2579 OP(AtomicUMax
, atomic_max
)
2580 OP(AtomicAnd
, atomic_and
)
2581 OP(AtomicOr
, atomic_or
)
2582 OP(AtomicXor
, atomic_xor
)
2585 vtn_fail_with_opcode("Invalid image opcode", opcode
);
2588 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
2590 nir_deref_instr
*image_deref
= vtn_pointer_to_deref(b
, image
.image
);
2591 intrin
->src
[0] = nir_src_for_ssa(&image_deref
->dest
.ssa
);
2593 /* ImageQuerySize doesn't take any extra parameters */
2594 if (opcode
!= SpvOpImageQuerySize
) {
2595 /* The image coordinate is always 4 components but we may not have that
2596 * many. Swizzle to compensate.
2598 intrin
->src
[1] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, image
.coord
));
2599 intrin
->src
[2] = nir_src_for_ssa(image
.sample
);
2603 case SpvOpAtomicLoad
:
2604 case SpvOpImageQuerySize
:
2605 case SpvOpImageRead
:
2607 case SpvOpAtomicStore
:
2608 case SpvOpImageWrite
: {
2609 const uint32_t value_id
= opcode
== SpvOpAtomicStore
? w
[4] : w
[3];
2610 nir_ssa_def
*value
= vtn_ssa_value(b
, value_id
)->def
;
2611 /* nir_intrinsic_image_deref_store always takes a vec4 value */
2612 assert(op
== nir_intrinsic_image_deref_store
);
2613 intrin
->num_components
= 4;
2614 intrin
->src
[3] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, value
));
2618 case SpvOpAtomicCompareExchange
:
2619 case SpvOpAtomicIIncrement
:
2620 case SpvOpAtomicIDecrement
:
2621 case SpvOpAtomicExchange
:
2622 case SpvOpAtomicIAdd
:
2623 case SpvOpAtomicISub
:
2624 case SpvOpAtomicSMin
:
2625 case SpvOpAtomicUMin
:
2626 case SpvOpAtomicSMax
:
2627 case SpvOpAtomicUMax
:
2628 case SpvOpAtomicAnd
:
2630 case SpvOpAtomicXor
:
2631 fill_common_atomic_sources(b
, opcode
, w
, &intrin
->src
[3]);
2635 vtn_fail_with_opcode("Invalid image opcode", opcode
);
2638 if (opcode
!= SpvOpImageWrite
&& opcode
!= SpvOpAtomicStore
) {
2639 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2640 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2642 unsigned dest_components
= glsl_get_vector_elements(type
->type
);
2643 intrin
->num_components
= nir_intrinsic_infos
[op
].dest_components
;
2644 if (intrin
->num_components
== 0)
2645 intrin
->num_components
= dest_components
;
2647 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
,
2648 intrin
->num_components
, 32, NULL
);
2650 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2652 nir_ssa_def
*result
= &intrin
->dest
.ssa
;
2653 if (intrin
->num_components
!= dest_components
)
2654 result
= nir_channels(&b
->nb
, result
, (1 << dest_components
) - 1);
2656 val
->ssa
= vtn_create_ssa_value(b
, type
->type
);
2657 val
->ssa
->def
= result
;
2659 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2663 static nir_intrinsic_op
2664 get_ssbo_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2667 case SpvOpAtomicLoad
: return nir_intrinsic_load_ssbo
;
2668 case SpvOpAtomicStore
: return nir_intrinsic_store_ssbo
;
2669 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
2670 OP(AtomicExchange
, atomic_exchange
)
2671 OP(AtomicCompareExchange
, atomic_comp_swap
)
2672 OP(AtomicIIncrement
, atomic_add
)
2673 OP(AtomicIDecrement
, atomic_add
)
2674 OP(AtomicIAdd
, atomic_add
)
2675 OP(AtomicISub
, atomic_add
)
2676 OP(AtomicSMin
, atomic_imin
)
2677 OP(AtomicUMin
, atomic_umin
)
2678 OP(AtomicSMax
, atomic_imax
)
2679 OP(AtomicUMax
, atomic_umax
)
2680 OP(AtomicAnd
, atomic_and
)
2681 OP(AtomicOr
, atomic_or
)
2682 OP(AtomicXor
, atomic_xor
)
2685 vtn_fail_with_opcode("Invalid SSBO atomic", opcode
);
2689 static nir_intrinsic_op
2690 get_uniform_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2693 #define OP(S, N) case SpvOp##S: return nir_intrinsic_atomic_counter_ ##N;
2694 OP(AtomicLoad
, read_deref
)
2695 OP(AtomicExchange
, exchange
)
2696 OP(AtomicCompareExchange
, comp_swap
)
2697 OP(AtomicIIncrement
, inc_deref
)
2698 OP(AtomicIDecrement
, post_dec_deref
)
2699 OP(AtomicIAdd
, add_deref
)
2700 OP(AtomicISub
, add_deref
)
2701 OP(AtomicUMin
, min_deref
)
2702 OP(AtomicUMax
, max_deref
)
2703 OP(AtomicAnd
, and_deref
)
2704 OP(AtomicOr
, or_deref
)
2705 OP(AtomicXor
, xor_deref
)
2708 /* We left the following out: AtomicStore, AtomicSMin and
2709 * AtomicSmax. Right now there are not nir intrinsics for them. At this
2710 * moment Atomic Counter support is needed for ARB_spirv support, so is
2711 * only need to support GLSL Atomic Counters that are uints and don't
2712 * allow direct storage.
2714 unreachable("Invalid uniform atomic");
2718 static nir_intrinsic_op
2719 get_shared_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2722 case SpvOpAtomicLoad
: return nir_intrinsic_load_shared
;
2723 case SpvOpAtomicStore
: return nir_intrinsic_store_shared
;
2724 #define OP(S, N) case SpvOp##S: return nir_intrinsic_shared_##N;
2725 OP(AtomicExchange
, atomic_exchange
)
2726 OP(AtomicCompareExchange
, atomic_comp_swap
)
2727 OP(AtomicIIncrement
, atomic_add
)
2728 OP(AtomicIDecrement
, atomic_add
)
2729 OP(AtomicIAdd
, atomic_add
)
2730 OP(AtomicISub
, atomic_add
)
2731 OP(AtomicSMin
, atomic_imin
)
2732 OP(AtomicUMin
, atomic_umin
)
2733 OP(AtomicSMax
, atomic_imax
)
2734 OP(AtomicUMax
, atomic_umax
)
2735 OP(AtomicAnd
, atomic_and
)
2736 OP(AtomicOr
, atomic_or
)
2737 OP(AtomicXor
, atomic_xor
)
2740 vtn_fail_with_opcode("Invalid shared atomic", opcode
);
2744 static nir_intrinsic_op
2745 get_deref_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2748 case SpvOpAtomicLoad
: return nir_intrinsic_load_deref
;
2749 case SpvOpAtomicStore
: return nir_intrinsic_store_deref
;
2750 #define OP(S, N) case SpvOp##S: return nir_intrinsic_deref_##N;
2751 OP(AtomicExchange
, atomic_exchange
)
2752 OP(AtomicCompareExchange
, atomic_comp_swap
)
2753 OP(AtomicIIncrement
, atomic_add
)
2754 OP(AtomicIDecrement
, atomic_add
)
2755 OP(AtomicIAdd
, atomic_add
)
2756 OP(AtomicISub
, atomic_add
)
2757 OP(AtomicSMin
, atomic_imin
)
2758 OP(AtomicUMin
, atomic_umin
)
2759 OP(AtomicSMax
, atomic_imax
)
2760 OP(AtomicUMax
, atomic_umax
)
2761 OP(AtomicAnd
, atomic_and
)
2762 OP(AtomicOr
, atomic_or
)
2763 OP(AtomicXor
, atomic_xor
)
2766 vtn_fail_with_opcode("Invalid shared atomic", opcode
);
2771 * Handles shared atomics, ssbo atomics and atomic counters.
2774 vtn_handle_atomics(struct vtn_builder
*b
, SpvOp opcode
,
2775 const uint32_t *w
, unsigned count
)
2777 struct vtn_pointer
*ptr
;
2778 nir_intrinsic_instr
*atomic
;
2781 case SpvOpAtomicLoad
:
2782 case SpvOpAtomicExchange
:
2783 case SpvOpAtomicCompareExchange
:
2784 case SpvOpAtomicCompareExchangeWeak
:
2785 case SpvOpAtomicIIncrement
:
2786 case SpvOpAtomicIDecrement
:
2787 case SpvOpAtomicIAdd
:
2788 case SpvOpAtomicISub
:
2789 case SpvOpAtomicSMin
:
2790 case SpvOpAtomicUMin
:
2791 case SpvOpAtomicSMax
:
2792 case SpvOpAtomicUMax
:
2793 case SpvOpAtomicAnd
:
2795 case SpvOpAtomicXor
:
2796 ptr
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2799 case SpvOpAtomicStore
:
2800 ptr
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
2804 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
2808 SpvScope scope = w[4];
2809 SpvMemorySemanticsMask semantics = w[5];
2812 /* uniform as "atomic counter uniform" */
2813 if (ptr
->mode
== vtn_variable_mode_uniform
) {
2814 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
2815 const struct glsl_type
*deref_type
= deref
->type
;
2816 nir_intrinsic_op op
= get_uniform_nir_atomic_op(b
, opcode
);
2817 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2818 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
2820 /* SSBO needs to initialize index/offset. In this case we don't need to,
2821 * as that info is already stored on the ptr->var->var nir_variable (see
2822 * vtn_create_variable)
2826 case SpvOpAtomicLoad
:
2827 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2830 case SpvOpAtomicStore
:
2831 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2832 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2835 case SpvOpAtomicExchange
:
2836 case SpvOpAtomicCompareExchange
:
2837 case SpvOpAtomicCompareExchangeWeak
:
2838 case SpvOpAtomicIIncrement
:
2839 case SpvOpAtomicIDecrement
:
2840 case SpvOpAtomicIAdd
:
2841 case SpvOpAtomicISub
:
2842 case SpvOpAtomicSMin
:
2843 case SpvOpAtomicUMin
:
2844 case SpvOpAtomicSMax
:
2845 case SpvOpAtomicUMax
:
2846 case SpvOpAtomicAnd
:
2848 case SpvOpAtomicXor
:
2849 /* Nothing: we don't need to call fill_common_atomic_sources here, as
2850 * atomic counter uniforms doesn't have sources
2855 unreachable("Invalid SPIR-V atomic");
2858 } else if (vtn_pointer_uses_ssa_offset(b
, ptr
)) {
2859 nir_ssa_def
*offset
, *index
;
2860 offset
= vtn_pointer_to_offset(b
, ptr
, &index
);
2862 nir_intrinsic_op op
;
2863 if (ptr
->mode
== vtn_variable_mode_ssbo
) {
2864 op
= get_ssbo_nir_atomic_op(b
, opcode
);
2866 vtn_assert(ptr
->mode
== vtn_variable_mode_workgroup
&&
2867 b
->options
->lower_workgroup_access_to_offsets
);
2868 op
= get_shared_nir_atomic_op(b
, opcode
);
2871 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2875 case SpvOpAtomicLoad
:
2876 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
2877 nir_intrinsic_set_align(atomic
, 4, 0);
2878 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2879 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2880 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2883 case SpvOpAtomicStore
:
2884 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
2885 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2886 nir_intrinsic_set_align(atomic
, 4, 0);
2887 atomic
->src
[src
++] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2888 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2889 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2890 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2893 case SpvOpAtomicExchange
:
2894 case SpvOpAtomicCompareExchange
:
2895 case SpvOpAtomicCompareExchangeWeak
:
2896 case SpvOpAtomicIIncrement
:
2897 case SpvOpAtomicIDecrement
:
2898 case SpvOpAtomicIAdd
:
2899 case SpvOpAtomicISub
:
2900 case SpvOpAtomicSMin
:
2901 case SpvOpAtomicUMin
:
2902 case SpvOpAtomicSMax
:
2903 case SpvOpAtomicUMax
:
2904 case SpvOpAtomicAnd
:
2906 case SpvOpAtomicXor
:
2907 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2908 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2909 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2910 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[src
]);
2914 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
2917 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
2918 const struct glsl_type
*deref_type
= deref
->type
;
2919 nir_intrinsic_op op
= get_deref_nir_atomic_op(b
, opcode
);
2920 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2921 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
2924 case SpvOpAtomicLoad
:
2925 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2928 case SpvOpAtomicStore
:
2929 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2930 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2931 atomic
->src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2934 case SpvOpAtomicExchange
:
2935 case SpvOpAtomicCompareExchange
:
2936 case SpvOpAtomicCompareExchangeWeak
:
2937 case SpvOpAtomicIIncrement
:
2938 case SpvOpAtomicIDecrement
:
2939 case SpvOpAtomicIAdd
:
2940 case SpvOpAtomicISub
:
2941 case SpvOpAtomicSMin
:
2942 case SpvOpAtomicUMin
:
2943 case SpvOpAtomicSMax
:
2944 case SpvOpAtomicUMax
:
2945 case SpvOpAtomicAnd
:
2947 case SpvOpAtomicXor
:
2948 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[1]);
2952 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
2956 if (opcode
!= SpvOpAtomicStore
) {
2957 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2959 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
,
2960 glsl_get_vector_elements(type
->type
),
2961 glsl_get_bit_size(type
->type
), NULL
);
2963 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2964 val
->ssa
= rzalloc(b
, struct vtn_ssa_value
);
2965 val
->ssa
->def
= &atomic
->dest
.ssa
;
2966 val
->ssa
->type
= type
->type
;
2969 nir_builder_instr_insert(&b
->nb
, &atomic
->instr
);
2972 static nir_alu_instr
*
2973 create_vec(struct vtn_builder
*b
, unsigned num_components
, unsigned bit_size
)
2976 switch (num_components
) {
2977 case 1: op
= nir_op_imov
; break;
2978 case 2: op
= nir_op_vec2
; break;
2979 case 3: op
= nir_op_vec3
; break;
2980 case 4: op
= nir_op_vec4
; break;
2981 default: vtn_fail("bad vector size: %u", num_components
);
2984 nir_alu_instr
*vec
= nir_alu_instr_create(b
->shader
, op
);
2985 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, num_components
,
2987 vec
->dest
.write_mask
= (1 << num_components
) - 1;
2992 struct vtn_ssa_value
*
2993 vtn_ssa_transpose(struct vtn_builder
*b
, struct vtn_ssa_value
*src
)
2995 if (src
->transposed
)
2996 return src
->transposed
;
2998 struct vtn_ssa_value
*dest
=
2999 vtn_create_ssa_value(b
, glsl_transposed_type(src
->type
));
3001 for (unsigned i
= 0; i
< glsl_get_matrix_columns(dest
->type
); i
++) {
3002 nir_alu_instr
*vec
= create_vec(b
, glsl_get_matrix_columns(src
->type
),
3003 glsl_get_bit_size(src
->type
));
3004 if (glsl_type_is_vector_or_scalar(src
->type
)) {
3005 vec
->src
[0].src
= nir_src_for_ssa(src
->def
);
3006 vec
->src
[0].swizzle
[0] = i
;
3008 for (unsigned j
= 0; j
< glsl_get_matrix_columns(src
->type
); j
++) {
3009 vec
->src
[j
].src
= nir_src_for_ssa(src
->elems
[j
]->def
);
3010 vec
->src
[j
].swizzle
[0] = i
;
3013 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3014 dest
->elems
[i
]->def
= &vec
->dest
.dest
.ssa
;
3017 dest
->transposed
= src
;
3023 vtn_vector_extract(struct vtn_builder
*b
, nir_ssa_def
*src
, unsigned index
)
3025 return nir_channel(&b
->nb
, src
, index
);
3029 vtn_vector_insert(struct vtn_builder
*b
, nir_ssa_def
*src
, nir_ssa_def
*insert
,
3032 nir_alu_instr
*vec
= create_vec(b
, src
->num_components
,
3035 for (unsigned i
= 0; i
< src
->num_components
; i
++) {
3037 vec
->src
[i
].src
= nir_src_for_ssa(insert
);
3039 vec
->src
[i
].src
= nir_src_for_ssa(src
);
3040 vec
->src
[i
].swizzle
[0] = i
;
3044 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3046 return &vec
->dest
.dest
.ssa
;
3049 static nir_ssa_def
*
3050 nir_ieq_imm(nir_builder
*b
, nir_ssa_def
*x
, uint64_t i
)
3052 return nir_ieq(b
, x
, nir_imm_intN_t(b
, i
, x
->bit_size
));
3056 vtn_vector_extract_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
3059 return nir_vector_extract(&b
->nb
, src
, nir_i2i(&b
->nb
, index
, 32));
3063 vtn_vector_insert_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
3064 nir_ssa_def
*insert
, nir_ssa_def
*index
)
3066 nir_ssa_def
*dest
= vtn_vector_insert(b
, src
, insert
, 0);
3067 for (unsigned i
= 1; i
< src
->num_components
; i
++)
3068 dest
= nir_bcsel(&b
->nb
, nir_ieq_imm(&b
->nb
, index
, i
),
3069 vtn_vector_insert(b
, src
, insert
, i
), dest
);
3074 static nir_ssa_def
*
3075 vtn_vector_shuffle(struct vtn_builder
*b
, unsigned num_components
,
3076 nir_ssa_def
*src0
, nir_ssa_def
*src1
,
3077 const uint32_t *indices
)
3079 nir_alu_instr
*vec
= create_vec(b
, num_components
, src0
->bit_size
);
3081 for (unsigned i
= 0; i
< num_components
; i
++) {
3082 uint32_t index
= indices
[i
];
3083 if (index
== 0xffffffff) {
3085 nir_src_for_ssa(nir_ssa_undef(&b
->nb
, 1, src0
->bit_size
));
3086 } else if (index
< src0
->num_components
) {
3087 vec
->src
[i
].src
= nir_src_for_ssa(src0
);
3088 vec
->src
[i
].swizzle
[0] = index
;
3090 vec
->src
[i
].src
= nir_src_for_ssa(src1
);
3091 vec
->src
[i
].swizzle
[0] = index
- src0
->num_components
;
3095 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3097 return &vec
->dest
.dest
.ssa
;
3101 * Concatentates a number of vectors/scalars together to produce a vector
3103 static nir_ssa_def
*
3104 vtn_vector_construct(struct vtn_builder
*b
, unsigned num_components
,
3105 unsigned num_srcs
, nir_ssa_def
**srcs
)
3107 nir_alu_instr
*vec
= create_vec(b
, num_components
, srcs
[0]->bit_size
);
3109 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
3111 * "When constructing a vector, there must be at least two Constituent
3114 vtn_assert(num_srcs
>= 2);
3116 unsigned dest_idx
= 0;
3117 for (unsigned i
= 0; i
< num_srcs
; i
++) {
3118 nir_ssa_def
*src
= srcs
[i
];
3119 vtn_assert(dest_idx
+ src
->num_components
<= num_components
);
3120 for (unsigned j
= 0; j
< src
->num_components
; j
++) {
3121 vec
->src
[dest_idx
].src
= nir_src_for_ssa(src
);
3122 vec
->src
[dest_idx
].swizzle
[0] = j
;
3127 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
3129 * "When constructing a vector, the total number of components in all
3130 * the operands must equal the number of components in Result Type."
3132 vtn_assert(dest_idx
== num_components
);
3134 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3136 return &vec
->dest
.dest
.ssa
;
3139 static struct vtn_ssa_value
*
3140 vtn_composite_copy(void *mem_ctx
, struct vtn_ssa_value
*src
)
3142 struct vtn_ssa_value
*dest
= rzalloc(mem_ctx
, struct vtn_ssa_value
);
3143 dest
->type
= src
->type
;
3145 if (glsl_type_is_vector_or_scalar(src
->type
)) {
3146 dest
->def
= src
->def
;
3148 unsigned elems
= glsl_get_length(src
->type
);
3150 dest
->elems
= ralloc_array(mem_ctx
, struct vtn_ssa_value
*, elems
);
3151 for (unsigned i
= 0; i
< elems
; i
++)
3152 dest
->elems
[i
] = vtn_composite_copy(mem_ctx
, src
->elems
[i
]);
3158 static struct vtn_ssa_value
*
3159 vtn_composite_insert(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
3160 struct vtn_ssa_value
*insert
, const uint32_t *indices
,
3161 unsigned num_indices
)
3163 struct vtn_ssa_value
*dest
= vtn_composite_copy(b
, src
);
3165 struct vtn_ssa_value
*cur
= dest
;
3167 for (i
= 0; i
< num_indices
- 1; i
++) {
3168 cur
= cur
->elems
[indices
[i
]];
3171 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
3172 /* According to the SPIR-V spec, OpCompositeInsert may work down to
3173 * the component granularity. In that case, the last index will be
3174 * the index to insert the scalar into the vector.
3177 cur
->def
= vtn_vector_insert(b
, cur
->def
, insert
->def
, indices
[i
]);
3179 cur
->elems
[indices
[i
]] = insert
;
3185 static struct vtn_ssa_value
*
3186 vtn_composite_extract(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
3187 const uint32_t *indices
, unsigned num_indices
)
3189 struct vtn_ssa_value
*cur
= src
;
3190 for (unsigned i
= 0; i
< num_indices
; i
++) {
3191 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
3192 vtn_assert(i
== num_indices
- 1);
3193 /* According to the SPIR-V spec, OpCompositeExtract may work down to
3194 * the component granularity. The last index will be the index of the
3195 * vector to extract.
3198 struct vtn_ssa_value
*ret
= rzalloc(b
, struct vtn_ssa_value
);
3199 ret
->type
= glsl_scalar_type(glsl_get_base_type(cur
->type
));
3200 ret
->def
= vtn_vector_extract(b
, cur
->def
, indices
[i
]);
3203 cur
= cur
->elems
[indices
[i
]];
3211 vtn_handle_composite(struct vtn_builder
*b
, SpvOp opcode
,
3212 const uint32_t *w
, unsigned count
)
3214 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
3215 const struct glsl_type
*type
=
3216 vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
3217 val
->ssa
= vtn_create_ssa_value(b
, type
);
3220 case SpvOpVectorExtractDynamic
:
3221 val
->ssa
->def
= vtn_vector_extract_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
3222 vtn_ssa_value(b
, w
[4])->def
);
3225 case SpvOpVectorInsertDynamic
:
3226 val
->ssa
->def
= vtn_vector_insert_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
3227 vtn_ssa_value(b
, w
[4])->def
,
3228 vtn_ssa_value(b
, w
[5])->def
);
3231 case SpvOpVectorShuffle
:
3232 val
->ssa
->def
= vtn_vector_shuffle(b
, glsl_get_vector_elements(type
),
3233 vtn_ssa_value(b
, w
[3])->def
,
3234 vtn_ssa_value(b
, w
[4])->def
,
3238 case SpvOpCompositeConstruct
: {
3239 unsigned elems
= count
- 3;
3241 if (glsl_type_is_vector_or_scalar(type
)) {
3242 nir_ssa_def
*srcs
[NIR_MAX_VEC_COMPONENTS
];
3243 for (unsigned i
= 0; i
< elems
; i
++)
3244 srcs
[i
] = vtn_ssa_value(b
, w
[3 + i
])->def
;
3246 vtn_vector_construct(b
, glsl_get_vector_elements(type
),
3249 val
->ssa
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
3250 for (unsigned i
= 0; i
< elems
; i
++)
3251 val
->ssa
->elems
[i
] = vtn_ssa_value(b
, w
[3 + i
]);
3255 case SpvOpCompositeExtract
:
3256 val
->ssa
= vtn_composite_extract(b
, vtn_ssa_value(b
, w
[3]),
3260 case SpvOpCompositeInsert
:
3261 val
->ssa
= vtn_composite_insert(b
, vtn_ssa_value(b
, w
[4]),
3262 vtn_ssa_value(b
, w
[3]),
3266 case SpvOpCopyObject
:
3267 val
->ssa
= vtn_composite_copy(b
, vtn_ssa_value(b
, w
[3]));
3271 vtn_fail_with_opcode("unknown composite operation", opcode
);
3276 vtn_emit_barrier(struct vtn_builder
*b
, nir_intrinsic_op op
)
3278 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
3279 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3283 vtn_emit_memory_barrier(struct vtn_builder
*b
, SpvScope scope
,
3284 SpvMemorySemanticsMask semantics
)
3286 static const SpvMemorySemanticsMask all_memory_semantics
=
3287 SpvMemorySemanticsUniformMemoryMask
|
3288 SpvMemorySemanticsWorkgroupMemoryMask
|
3289 SpvMemorySemanticsAtomicCounterMemoryMask
|
3290 SpvMemorySemanticsImageMemoryMask
;
3292 /* If we're not actually doing a memory barrier, bail */
3293 if (!(semantics
& all_memory_semantics
))
3296 /* GL and Vulkan don't have these */
3297 vtn_assert(scope
!= SpvScopeCrossDevice
);
3299 if (scope
== SpvScopeSubgroup
)
3300 return; /* Nothing to do here */
3302 if (scope
== SpvScopeWorkgroup
) {
3303 vtn_emit_barrier(b
, nir_intrinsic_group_memory_barrier
);
3307 /* There's only two scopes thing left */
3308 vtn_assert(scope
== SpvScopeInvocation
|| scope
== SpvScopeDevice
);
3310 if ((semantics
& all_memory_semantics
) == all_memory_semantics
) {
3311 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier
);
3315 /* Issue a bunch of more specific barriers */
3316 uint32_t bits
= semantics
;
3318 SpvMemorySemanticsMask semantic
= 1 << u_bit_scan(&bits
);
3320 case SpvMemorySemanticsUniformMemoryMask
:
3321 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_buffer
);
3323 case SpvMemorySemanticsWorkgroupMemoryMask
:
3324 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_shared
);
3326 case SpvMemorySemanticsAtomicCounterMemoryMask
:
3327 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_atomic_counter
);
3329 case SpvMemorySemanticsImageMemoryMask
:
3330 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_image
);
3339 vtn_handle_barrier(struct vtn_builder
*b
, SpvOp opcode
,
3340 const uint32_t *w
, unsigned count
)
3343 case SpvOpEmitVertex
:
3344 case SpvOpEmitStreamVertex
:
3345 case SpvOpEndPrimitive
:
3346 case SpvOpEndStreamPrimitive
: {
3347 nir_intrinsic_op intrinsic_op
;
3349 case SpvOpEmitVertex
:
3350 case SpvOpEmitStreamVertex
:
3351 intrinsic_op
= nir_intrinsic_emit_vertex
;
3353 case SpvOpEndPrimitive
:
3354 case SpvOpEndStreamPrimitive
:
3355 intrinsic_op
= nir_intrinsic_end_primitive
;
3358 unreachable("Invalid opcode");
3361 nir_intrinsic_instr
*intrin
=
3362 nir_intrinsic_instr_create(b
->shader
, intrinsic_op
);
3365 case SpvOpEmitStreamVertex
:
3366 case SpvOpEndStreamPrimitive
: {
3367 unsigned stream
= vtn_constant_uint(b
, w
[1]);
3368 nir_intrinsic_set_stream_id(intrin
, stream
);
3376 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3380 case SpvOpMemoryBarrier
: {
3381 SpvScope scope
= vtn_constant_uint(b
, w
[1]);
3382 SpvMemorySemanticsMask semantics
= vtn_constant_uint(b
, w
[2]);
3383 vtn_emit_memory_barrier(b
, scope
, semantics
);
3387 case SpvOpControlBarrier
: {
3388 SpvScope execution_scope
= vtn_constant_uint(b
, w
[1]);
3389 if (execution_scope
== SpvScopeWorkgroup
)
3390 vtn_emit_barrier(b
, nir_intrinsic_barrier
);
3392 SpvScope memory_scope
= vtn_constant_uint(b
, w
[2]);
3393 SpvMemorySemanticsMask memory_semantics
= vtn_constant_uint(b
, w
[3]);
3394 vtn_emit_memory_barrier(b
, memory_scope
, memory_semantics
);
3399 unreachable("unknown barrier instruction");
3404 gl_primitive_from_spv_execution_mode(struct vtn_builder
*b
,
3405 SpvExecutionMode mode
)
3408 case SpvExecutionModeInputPoints
:
3409 case SpvExecutionModeOutputPoints
:
3410 return 0; /* GL_POINTS */
3411 case SpvExecutionModeInputLines
:
3412 return 1; /* GL_LINES */
3413 case SpvExecutionModeInputLinesAdjacency
:
3414 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
3415 case SpvExecutionModeTriangles
:
3416 return 4; /* GL_TRIANGLES */
3417 case SpvExecutionModeInputTrianglesAdjacency
:
3418 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
3419 case SpvExecutionModeQuads
:
3420 return 7; /* GL_QUADS */
3421 case SpvExecutionModeIsolines
:
3422 return 0x8E7A; /* GL_ISOLINES */
3423 case SpvExecutionModeOutputLineStrip
:
3424 return 3; /* GL_LINE_STRIP */
3425 case SpvExecutionModeOutputTriangleStrip
:
3426 return 5; /* GL_TRIANGLE_STRIP */
3428 vtn_fail("Invalid primitive type: %s (%u)",
3429 spirv_executionmode_to_string(mode
), mode
);
3434 vertices_in_from_spv_execution_mode(struct vtn_builder
*b
,
3435 SpvExecutionMode mode
)
3438 case SpvExecutionModeInputPoints
:
3440 case SpvExecutionModeInputLines
:
3442 case SpvExecutionModeInputLinesAdjacency
:
3444 case SpvExecutionModeTriangles
:
3446 case SpvExecutionModeInputTrianglesAdjacency
:
3449 vtn_fail("Invalid GS input mode: %s (%u)",
3450 spirv_executionmode_to_string(mode
), mode
);
3454 static gl_shader_stage
3455 stage_for_execution_model(struct vtn_builder
*b
, SpvExecutionModel model
)
3458 case SpvExecutionModelVertex
:
3459 return MESA_SHADER_VERTEX
;
3460 case SpvExecutionModelTessellationControl
:
3461 return MESA_SHADER_TESS_CTRL
;
3462 case SpvExecutionModelTessellationEvaluation
:
3463 return MESA_SHADER_TESS_EVAL
;
3464 case SpvExecutionModelGeometry
:
3465 return MESA_SHADER_GEOMETRY
;
3466 case SpvExecutionModelFragment
:
3467 return MESA_SHADER_FRAGMENT
;
3468 case SpvExecutionModelGLCompute
:
3469 return MESA_SHADER_COMPUTE
;
3470 case SpvExecutionModelKernel
:
3471 return MESA_SHADER_KERNEL
;
3473 vtn_fail("Unsupported execution model: %s (%u)",
3474 spirv_executionmodel_to_string(model
), model
);
3478 #define spv_check_supported(name, cap) do { \
3479 if (!(b->options && b->options->caps.name)) \
3480 vtn_warn("Unsupported SPIR-V capability: %s (%u)", \
3481 spirv_capability_to_string(cap), cap); \
3486 vtn_handle_entry_point(struct vtn_builder
*b
, const uint32_t *w
,
3489 struct vtn_value
*entry_point
= &b
->values
[w
[2]];
3490 /* Let this be a name label regardless */
3491 unsigned name_words
;
3492 entry_point
->name
= vtn_string_literal(b
, &w
[3], count
- 3, &name_words
);
3494 if (strcmp(entry_point
->name
, b
->entry_point_name
) != 0 ||
3495 stage_for_execution_model(b
, w
[1]) != b
->entry_point_stage
)
3498 vtn_assert(b
->entry_point
== NULL
);
3499 b
->entry_point
= entry_point
;
3503 vtn_handle_preamble_instruction(struct vtn_builder
*b
, SpvOp opcode
,
3504 const uint32_t *w
, unsigned count
)
3511 case SpvSourceLanguageUnknown
: lang
= "unknown"; break;
3512 case SpvSourceLanguageESSL
: lang
= "ESSL"; break;
3513 case SpvSourceLanguageGLSL
: lang
= "GLSL"; break;
3514 case SpvSourceLanguageOpenCL_C
: lang
= "OpenCL C"; break;
3515 case SpvSourceLanguageOpenCL_CPP
: lang
= "OpenCL C++"; break;
3516 case SpvSourceLanguageHLSL
: lang
= "HLSL"; break;
3519 uint32_t version
= w
[2];
3522 (count
> 3) ? vtn_value(b
, w
[3], vtn_value_type_string
)->str
: "";
3524 vtn_info("Parsing SPIR-V from %s %u source file %s", lang
, version
, file
);
3528 case SpvOpSourceExtension
:
3529 case SpvOpSourceContinued
:
3530 case SpvOpExtension
:
3531 case SpvOpModuleProcessed
:
3532 /* Unhandled, but these are for debug so that's ok. */
3535 case SpvOpCapability
: {
3536 SpvCapability cap
= w
[1];
3538 case SpvCapabilityMatrix
:
3539 case SpvCapabilityShader
:
3540 case SpvCapabilityGeometry
:
3541 case SpvCapabilityGeometryPointSize
:
3542 case SpvCapabilityUniformBufferArrayDynamicIndexing
:
3543 case SpvCapabilitySampledImageArrayDynamicIndexing
:
3544 case SpvCapabilityStorageBufferArrayDynamicIndexing
:
3545 case SpvCapabilityStorageImageArrayDynamicIndexing
:
3546 case SpvCapabilityImageRect
:
3547 case SpvCapabilitySampledRect
:
3548 case SpvCapabilitySampled1D
:
3549 case SpvCapabilityImage1D
:
3550 case SpvCapabilitySampledCubeArray
:
3551 case SpvCapabilityImageCubeArray
:
3552 case SpvCapabilitySampledBuffer
:
3553 case SpvCapabilityImageBuffer
:
3554 case SpvCapabilityImageQuery
:
3555 case SpvCapabilityDerivativeControl
:
3556 case SpvCapabilityInterpolationFunction
:
3557 case SpvCapabilityMultiViewport
:
3558 case SpvCapabilitySampleRateShading
:
3559 case SpvCapabilityClipDistance
:
3560 case SpvCapabilityCullDistance
:
3561 case SpvCapabilityInputAttachment
:
3562 case SpvCapabilityImageGatherExtended
:
3563 case SpvCapabilityStorageImageExtendedFormats
:
3566 case SpvCapabilityLinkage
:
3567 case SpvCapabilityVector16
:
3568 case SpvCapabilityFloat16Buffer
:
3569 case SpvCapabilitySparseResidency
:
3570 vtn_warn("Unsupported SPIR-V capability: %s",
3571 spirv_capability_to_string(cap
));
3574 case SpvCapabilityMinLod
:
3575 spv_check_supported(min_lod
, cap
);
3578 case SpvCapabilityAtomicStorage
:
3579 spv_check_supported(atomic_storage
, cap
);
3582 case SpvCapabilityFloat64
:
3583 spv_check_supported(float64
, cap
);
3585 case SpvCapabilityInt64
:
3586 spv_check_supported(int64
, cap
);
3588 case SpvCapabilityInt16
:
3589 spv_check_supported(int16
, cap
);
3592 case SpvCapabilityTransformFeedback
:
3593 spv_check_supported(transform_feedback
, cap
);
3596 case SpvCapabilityGeometryStreams
:
3597 spv_check_supported(geometry_streams
, cap
);
3600 case SpvCapabilityInt64Atomics
:
3601 spv_check_supported(int64_atomics
, cap
);
3604 case SpvCapabilityInt8
:
3605 spv_check_supported(int8
, cap
);
3608 case SpvCapabilityStorageImageMultisample
:
3609 spv_check_supported(storage_image_ms
, cap
);
3612 case SpvCapabilityAddresses
:
3613 spv_check_supported(address
, cap
);
3616 case SpvCapabilityKernel
:
3617 spv_check_supported(kernel
, cap
);
3620 case SpvCapabilityImageBasic
:
3621 case SpvCapabilityImageReadWrite
:
3622 case SpvCapabilityImageMipmap
:
3623 case SpvCapabilityPipes
:
3624 case SpvCapabilityGroups
:
3625 case SpvCapabilityDeviceEnqueue
:
3626 case SpvCapabilityLiteralSampler
:
3627 case SpvCapabilityGenericPointer
:
3628 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
3629 spirv_capability_to_string(cap
));
3632 case SpvCapabilityImageMSArray
:
3633 spv_check_supported(image_ms_array
, cap
);
3636 case SpvCapabilityTessellation
:
3637 case SpvCapabilityTessellationPointSize
:
3638 spv_check_supported(tessellation
, cap
);
3641 case SpvCapabilityDrawParameters
:
3642 spv_check_supported(draw_parameters
, cap
);
3645 case SpvCapabilityStorageImageReadWithoutFormat
:
3646 spv_check_supported(image_read_without_format
, cap
);
3649 case SpvCapabilityStorageImageWriteWithoutFormat
:
3650 spv_check_supported(image_write_without_format
, cap
);
3653 case SpvCapabilityDeviceGroup
:
3654 spv_check_supported(device_group
, cap
);
3657 case SpvCapabilityMultiView
:
3658 spv_check_supported(multiview
, cap
);
3661 case SpvCapabilityGroupNonUniform
:
3662 spv_check_supported(subgroup_basic
, cap
);
3665 case SpvCapabilityGroupNonUniformVote
:
3666 spv_check_supported(subgroup_vote
, cap
);
3669 case SpvCapabilitySubgroupBallotKHR
:
3670 case SpvCapabilityGroupNonUniformBallot
:
3671 spv_check_supported(subgroup_ballot
, cap
);
3674 case SpvCapabilityGroupNonUniformShuffle
:
3675 case SpvCapabilityGroupNonUniformShuffleRelative
:
3676 spv_check_supported(subgroup_shuffle
, cap
);
3679 case SpvCapabilityGroupNonUniformQuad
:
3680 spv_check_supported(subgroup_quad
, cap
);
3683 case SpvCapabilityGroupNonUniformArithmetic
:
3684 case SpvCapabilityGroupNonUniformClustered
:
3685 spv_check_supported(subgroup_arithmetic
, cap
);
3688 case SpvCapabilityVariablePointersStorageBuffer
:
3689 case SpvCapabilityVariablePointers
:
3690 spv_check_supported(variable_pointers
, cap
);
3691 b
->variable_pointers
= true;
3694 case SpvCapabilityStorageUniformBufferBlock16
:
3695 case SpvCapabilityStorageUniform16
:
3696 case SpvCapabilityStoragePushConstant16
:
3697 case SpvCapabilityStorageInputOutput16
:
3698 spv_check_supported(storage_16bit
, cap
);
3701 case SpvCapabilityShaderViewportIndexLayerEXT
:
3702 spv_check_supported(shader_viewport_index_layer
, cap
);
3705 case SpvCapabilityStorageBuffer8BitAccess
:
3706 case SpvCapabilityUniformAndStorageBuffer8BitAccess
:
3707 case SpvCapabilityStoragePushConstant8
:
3708 spv_check_supported(storage_8bit
, cap
);
3711 case SpvCapabilityInputAttachmentArrayDynamicIndexingEXT
:
3712 case SpvCapabilityUniformTexelBufferArrayDynamicIndexingEXT
:
3713 case SpvCapabilityStorageTexelBufferArrayDynamicIndexingEXT
:
3714 spv_check_supported(descriptor_array_dynamic_indexing
, cap
);
3717 case SpvCapabilityRuntimeDescriptorArrayEXT
:
3718 spv_check_supported(runtime_descriptor_array
, cap
);
3721 case SpvCapabilityStencilExportEXT
:
3722 spv_check_supported(stencil_export
, cap
);
3725 case SpvCapabilitySampleMaskPostDepthCoverage
:
3726 spv_check_supported(post_depth_coverage
, cap
);
3729 case SpvCapabilityPhysicalStorageBufferAddressesEXT
:
3730 spv_check_supported(physical_storage_buffer_address
, cap
);
3733 case SpvCapabilityComputeDerivativeGroupQuadsNV
:
3734 case SpvCapabilityComputeDerivativeGroupLinearNV
:
3735 spv_check_supported(derivative_group
, cap
);
3738 case SpvCapabilityFloat16
:
3739 spv_check_supported(float16
, cap
);
3743 vtn_fail("Unhandled capability: %s (%u)",
3744 spirv_capability_to_string(cap
), cap
);
3749 case SpvOpExtInstImport
:
3750 vtn_handle_extension(b
, opcode
, w
, count
);
3753 case SpvOpMemoryModel
:
3755 case SpvAddressingModelPhysical32
:
3756 vtn_fail_if(b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
,
3757 "AddressingModelPhysical32 only supported for kernels");
3758 b
->shader
->info
.cs
.ptr_size
= 32;
3759 b
->physical_ptrs
= true;
3760 b
->options
->shared_ptr_type
= glsl_uint_type();
3761 b
->options
->global_ptr_type
= glsl_uint_type();
3762 b
->options
->temp_ptr_type
= glsl_uint_type();
3764 case SpvAddressingModelPhysical64
:
3765 vtn_fail_if(b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
,
3766 "AddressingModelPhysical64 only supported for kernels");
3767 b
->shader
->info
.cs
.ptr_size
= 64;
3768 b
->physical_ptrs
= true;
3769 b
->options
->shared_ptr_type
= glsl_uint64_t_type();
3770 b
->options
->global_ptr_type
= glsl_uint64_t_type();
3771 b
->options
->temp_ptr_type
= glsl_uint64_t_type();
3773 case SpvAddressingModelLogical
:
3774 vtn_fail_if(b
->shader
->info
.stage
>= MESA_SHADER_STAGES
,
3775 "AddressingModelLogical only supported for shaders");
3776 b
->shader
->info
.cs
.ptr_size
= 0;
3777 b
->physical_ptrs
= false;
3779 case SpvAddressingModelPhysicalStorageBuffer64EXT
:
3780 vtn_fail_if(!b
->options
||
3781 !b
->options
->caps
.physical_storage_buffer_address
,
3782 "AddressingModelPhysicalStorageBuffer64EXT not supported");
3785 vtn_fail("Unknown addressing model: %s (%u)",
3786 spirv_addressingmodel_to_string(w
[1]), w
[1]);
3790 vtn_assert(w
[2] == SpvMemoryModelSimple
||
3791 w
[2] == SpvMemoryModelGLSL450
||
3792 w
[2] == SpvMemoryModelOpenCL
);
3795 case SpvOpEntryPoint
:
3796 vtn_handle_entry_point(b
, w
, count
);
3800 vtn_push_value(b
, w
[1], vtn_value_type_string
)->str
=
3801 vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
3805 b
->values
[w
[1]].name
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
3808 case SpvOpMemberName
:
3812 case SpvOpExecutionMode
:
3813 case SpvOpExecutionModeId
:
3814 case SpvOpDecorationGroup
:
3816 case SpvOpMemberDecorate
:
3817 case SpvOpGroupDecorate
:
3818 case SpvOpGroupMemberDecorate
:
3819 case SpvOpDecorateStringGOOGLE
:
3820 case SpvOpMemberDecorateStringGOOGLE
:
3821 vtn_handle_decoration(b
, opcode
, w
, count
);
3825 return false; /* End of preamble */
3832 vtn_handle_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
3833 const struct vtn_decoration
*mode
, void *data
)
3835 vtn_assert(b
->entry_point
== entry_point
);
3837 switch(mode
->exec_mode
) {
3838 case SpvExecutionModeOriginUpperLeft
:
3839 case SpvExecutionModeOriginLowerLeft
:
3840 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3841 b
->shader
->info
.fs
.origin_upper_left
=
3842 (mode
->exec_mode
== SpvExecutionModeOriginUpperLeft
);
3845 case SpvExecutionModeEarlyFragmentTests
:
3846 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3847 b
->shader
->info
.fs
.early_fragment_tests
= true;
3850 case SpvExecutionModePostDepthCoverage
:
3851 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3852 b
->shader
->info
.fs
.post_depth_coverage
= true;
3855 case SpvExecutionModeInvocations
:
3856 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3857 b
->shader
->info
.gs
.invocations
= MAX2(1, mode
->literals
[0]);
3860 case SpvExecutionModeDepthReplacing
:
3861 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3862 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_ANY
;
3864 case SpvExecutionModeDepthGreater
:
3865 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3866 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_GREATER
;
3868 case SpvExecutionModeDepthLess
:
3869 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3870 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_LESS
;
3872 case SpvExecutionModeDepthUnchanged
:
3873 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3874 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
3877 case SpvExecutionModeLocalSize
:
3878 vtn_assert(gl_shader_stage_is_compute(b
->shader
->info
.stage
));
3879 b
->shader
->info
.cs
.local_size
[0] = mode
->literals
[0];
3880 b
->shader
->info
.cs
.local_size
[1] = mode
->literals
[1];
3881 b
->shader
->info
.cs
.local_size
[2] = mode
->literals
[2];
3884 case SpvExecutionModeLocalSizeId
:
3885 b
->shader
->info
.cs
.local_size
[0] = vtn_constant_uint(b
, mode
->literals
[0]);
3886 b
->shader
->info
.cs
.local_size
[1] = vtn_constant_uint(b
, mode
->literals
[1]);
3887 b
->shader
->info
.cs
.local_size
[2] = vtn_constant_uint(b
, mode
->literals
[2]);
3890 case SpvExecutionModeLocalSizeHint
:
3891 case SpvExecutionModeLocalSizeHintId
:
3892 break; /* Nothing to do with this */
3894 case SpvExecutionModeOutputVertices
:
3895 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3896 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
3897 b
->shader
->info
.tess
.tcs_vertices_out
= mode
->literals
[0];
3899 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3900 b
->shader
->info
.gs
.vertices_out
= mode
->literals
[0];
3904 case SpvExecutionModeInputPoints
:
3905 case SpvExecutionModeInputLines
:
3906 case SpvExecutionModeInputLinesAdjacency
:
3907 case SpvExecutionModeTriangles
:
3908 case SpvExecutionModeInputTrianglesAdjacency
:
3909 case SpvExecutionModeQuads
:
3910 case SpvExecutionModeIsolines
:
3911 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3912 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
3913 b
->shader
->info
.tess
.primitive_mode
=
3914 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
3916 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3917 b
->shader
->info
.gs
.vertices_in
=
3918 vertices_in_from_spv_execution_mode(b
, mode
->exec_mode
);
3919 b
->shader
->info
.gs
.input_primitive
=
3920 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
3924 case SpvExecutionModeOutputPoints
:
3925 case SpvExecutionModeOutputLineStrip
:
3926 case SpvExecutionModeOutputTriangleStrip
:
3927 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3928 b
->shader
->info
.gs
.output_primitive
=
3929 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
3932 case SpvExecutionModeSpacingEqual
:
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_EQUAL
;
3937 case SpvExecutionModeSpacingFractionalEven
:
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
.spacing
= TESS_SPACING_FRACTIONAL_EVEN
;
3942 case SpvExecutionModeSpacingFractionalOdd
:
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
.spacing
= TESS_SPACING_FRACTIONAL_ODD
;
3947 case SpvExecutionModeVertexOrderCw
:
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
.ccw
= false;
3952 case SpvExecutionModeVertexOrderCcw
:
3953 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3954 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3955 b
->shader
->info
.tess
.ccw
= true;
3957 case SpvExecutionModePointMode
:
3958 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3959 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3960 b
->shader
->info
.tess
.point_mode
= true;
3963 case SpvExecutionModePixelCenterInteger
:
3964 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3965 b
->shader
->info
.fs
.pixel_center_integer
= true;
3968 case SpvExecutionModeXfb
:
3969 b
->shader
->info
.has_transform_feedback_varyings
= true;
3972 case SpvExecutionModeVecTypeHint
:
3975 case SpvExecutionModeContractionOff
:
3976 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
)
3977 vtn_warn("ExectionMode only allowed for CL-style kernels: %s",
3978 spirv_executionmode_to_string(mode
->exec_mode
));
3983 case SpvExecutionModeStencilRefReplacingEXT
:
3984 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3987 case SpvExecutionModeDerivativeGroupQuadsNV
:
3988 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
3989 b
->shader
->info
.cs
.derivative_group
= DERIVATIVE_GROUP_QUADS
;
3992 case SpvExecutionModeDerivativeGroupLinearNV
:
3993 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
3994 b
->shader
->info
.cs
.derivative_group
= DERIVATIVE_GROUP_LINEAR
;
3998 vtn_fail("Unhandled execution mode: %s (%u)",
3999 spirv_executionmode_to_string(mode
->exec_mode
),
4005 vtn_handle_variable_or_type_instruction(struct vtn_builder
*b
, SpvOp opcode
,
4006 const uint32_t *w
, unsigned count
)
4008 vtn_set_instruction_result_type(b
, opcode
, w
, count
);
4012 case SpvOpSourceContinued
:
4013 case SpvOpSourceExtension
:
4014 case SpvOpExtension
:
4015 case SpvOpCapability
:
4016 case SpvOpExtInstImport
:
4017 case SpvOpMemoryModel
:
4018 case SpvOpEntryPoint
:
4019 case SpvOpExecutionMode
:
4022 case SpvOpMemberName
:
4023 case SpvOpDecorationGroup
:
4025 case SpvOpMemberDecorate
:
4026 case SpvOpGroupDecorate
:
4027 case SpvOpGroupMemberDecorate
:
4028 case SpvOpDecorateStringGOOGLE
:
4029 case SpvOpMemberDecorateStringGOOGLE
:
4030 vtn_fail("Invalid opcode types and variables section");
4036 case SpvOpTypeFloat
:
4037 case SpvOpTypeVector
:
4038 case SpvOpTypeMatrix
:
4039 case SpvOpTypeImage
:
4040 case SpvOpTypeSampler
:
4041 case SpvOpTypeSampledImage
:
4042 case SpvOpTypeArray
:
4043 case SpvOpTypeRuntimeArray
:
4044 case SpvOpTypeStruct
:
4045 case SpvOpTypeOpaque
:
4046 case SpvOpTypePointer
:
4047 case SpvOpTypeForwardPointer
:
4048 case SpvOpTypeFunction
:
4049 case SpvOpTypeEvent
:
4050 case SpvOpTypeDeviceEvent
:
4051 case SpvOpTypeReserveId
:
4052 case SpvOpTypeQueue
:
4054 vtn_handle_type(b
, opcode
, w
, count
);
4057 case SpvOpConstantTrue
:
4058 case SpvOpConstantFalse
:
4060 case SpvOpConstantComposite
:
4061 case SpvOpConstantSampler
:
4062 case SpvOpConstantNull
:
4063 case SpvOpSpecConstantTrue
:
4064 case SpvOpSpecConstantFalse
:
4065 case SpvOpSpecConstant
:
4066 case SpvOpSpecConstantComposite
:
4067 case SpvOpSpecConstantOp
:
4068 vtn_handle_constant(b
, opcode
, w
, count
);
4073 vtn_handle_variables(b
, opcode
, w
, count
);
4077 return false; /* End of preamble */
4084 vtn_handle_body_instruction(struct vtn_builder
*b
, SpvOp opcode
,
4085 const uint32_t *w
, unsigned count
)
4091 case SpvOpLoopMerge
:
4092 case SpvOpSelectionMerge
:
4093 /* This is handled by cfg pre-pass and walk_blocks */
4097 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
4098 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
4103 vtn_handle_extension(b
, opcode
, w
, count
);
4109 case SpvOpCopyMemory
:
4110 case SpvOpCopyMemorySized
:
4111 case SpvOpAccessChain
:
4112 case SpvOpPtrAccessChain
:
4113 case SpvOpInBoundsAccessChain
:
4114 case SpvOpInBoundsPtrAccessChain
:
4115 case SpvOpArrayLength
:
4116 case SpvOpConvertPtrToU
:
4117 case SpvOpConvertUToPtr
:
4118 vtn_handle_variables(b
, opcode
, w
, count
);
4121 case SpvOpFunctionCall
:
4122 vtn_handle_function_call(b
, opcode
, w
, count
);
4125 case SpvOpSampledImage
:
4127 case SpvOpImageSampleImplicitLod
:
4128 case SpvOpImageSampleExplicitLod
:
4129 case SpvOpImageSampleDrefImplicitLod
:
4130 case SpvOpImageSampleDrefExplicitLod
:
4131 case SpvOpImageSampleProjImplicitLod
:
4132 case SpvOpImageSampleProjExplicitLod
:
4133 case SpvOpImageSampleProjDrefImplicitLod
:
4134 case SpvOpImageSampleProjDrefExplicitLod
:
4135 case SpvOpImageFetch
:
4136 case SpvOpImageGather
:
4137 case SpvOpImageDrefGather
:
4138 case SpvOpImageQuerySizeLod
:
4139 case SpvOpImageQueryLod
:
4140 case SpvOpImageQueryLevels
:
4141 case SpvOpImageQuerySamples
:
4142 vtn_handle_texture(b
, opcode
, w
, count
);
4145 case SpvOpImageRead
:
4146 case SpvOpImageWrite
:
4147 case SpvOpImageTexelPointer
:
4148 vtn_handle_image(b
, opcode
, w
, count
);
4151 case SpvOpImageQuerySize
: {
4152 struct vtn_pointer
*image
=
4153 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
4154 if (glsl_type_is_image(image
->type
->type
)) {
4155 vtn_handle_image(b
, opcode
, w
, count
);
4157 vtn_assert(glsl_type_is_sampler(image
->type
->type
));
4158 vtn_handle_texture(b
, opcode
, w
, count
);
4163 case SpvOpAtomicLoad
:
4164 case SpvOpAtomicExchange
:
4165 case SpvOpAtomicCompareExchange
:
4166 case SpvOpAtomicCompareExchangeWeak
:
4167 case SpvOpAtomicIIncrement
:
4168 case SpvOpAtomicIDecrement
:
4169 case SpvOpAtomicIAdd
:
4170 case SpvOpAtomicISub
:
4171 case SpvOpAtomicSMin
:
4172 case SpvOpAtomicUMin
:
4173 case SpvOpAtomicSMax
:
4174 case SpvOpAtomicUMax
:
4175 case SpvOpAtomicAnd
:
4177 case SpvOpAtomicXor
: {
4178 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[3]);
4179 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
4180 vtn_handle_image(b
, opcode
, w
, count
);
4182 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
4183 vtn_handle_atomics(b
, opcode
, w
, count
);
4188 case SpvOpAtomicStore
: {
4189 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[1]);
4190 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
4191 vtn_handle_image(b
, opcode
, w
, count
);
4193 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
4194 vtn_handle_atomics(b
, opcode
, w
, count
);
4200 /* Handle OpSelect up-front here because it needs to be able to handle
4201 * pointers and not just regular vectors and scalars.
4203 struct vtn_value
*res_val
= vtn_untyped_value(b
, w
[2]);
4204 struct vtn_value
*sel_val
= vtn_untyped_value(b
, w
[3]);
4205 struct vtn_value
*obj1_val
= vtn_untyped_value(b
, w
[4]);
4206 struct vtn_value
*obj2_val
= vtn_untyped_value(b
, w
[5]);
4208 const struct glsl_type
*sel_type
;
4209 switch (res_val
->type
->base_type
) {
4210 case vtn_base_type_scalar
:
4211 sel_type
= glsl_bool_type();
4213 case vtn_base_type_vector
:
4214 sel_type
= glsl_vector_type(GLSL_TYPE_BOOL
, res_val
->type
->length
);
4216 case vtn_base_type_pointer
:
4217 /* We need to have actual storage for pointer types */
4218 vtn_fail_if(res_val
->type
->type
== NULL
,
4219 "Invalid pointer result type for OpSelect");
4220 sel_type
= glsl_bool_type();
4223 vtn_fail("Result type of OpSelect must be a scalar, vector, or pointer");
4226 if (unlikely(sel_val
->type
->type
!= sel_type
)) {
4227 if (sel_val
->type
->type
== glsl_bool_type()) {
4228 /* This case is illegal but some older versions of GLSLang produce
4229 * it. The GLSLang issue was fixed on March 30, 2017:
4231 * https://github.com/KhronosGroup/glslang/issues/809
4233 * Unfortunately, there are applications in the wild which are
4234 * shipping with this bug so it isn't nice to fail on them so we
4235 * throw a warning instead. It's not actually a problem for us as
4236 * nir_builder will just splat the condition out which is most
4237 * likely what the client wanted anyway.
4239 vtn_warn("Condition type of OpSelect must have the same number "
4240 "of components as Result Type");
4242 vtn_fail("Condition type of OpSelect must be a scalar or vector "
4243 "of Boolean type. It must have the same number of "
4244 "components as Result Type");
4248 vtn_fail_if(obj1_val
->type
!= res_val
->type
||
4249 obj2_val
->type
!= res_val
->type
,
4250 "Object types must match the result type in OpSelect");
4252 struct vtn_type
*res_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
4253 struct vtn_ssa_value
*ssa
= vtn_create_ssa_value(b
, res_type
->type
);
4254 ssa
->def
= nir_bcsel(&b
->nb
, vtn_ssa_value(b
, w
[3])->def
,
4255 vtn_ssa_value(b
, w
[4])->def
,
4256 vtn_ssa_value(b
, w
[5])->def
);
4257 vtn_push_ssa(b
, w
[2], res_type
, ssa
);
4266 case SpvOpConvertFToU
:
4267 case SpvOpConvertFToS
:
4268 case SpvOpConvertSToF
:
4269 case SpvOpConvertUToF
:
4273 case SpvOpQuantizeToF16
:
4274 case SpvOpPtrCastToGeneric
:
4275 case SpvOpGenericCastToPtr
:
4281 case SpvOpSignBitSet
:
4282 case SpvOpLessOrGreater
:
4284 case SpvOpUnordered
:
4299 case SpvOpVectorTimesScalar
:
4301 case SpvOpIAddCarry
:
4302 case SpvOpISubBorrow
:
4303 case SpvOpUMulExtended
:
4304 case SpvOpSMulExtended
:
4305 case SpvOpShiftRightLogical
:
4306 case SpvOpShiftRightArithmetic
:
4307 case SpvOpShiftLeftLogical
:
4308 case SpvOpLogicalEqual
:
4309 case SpvOpLogicalNotEqual
:
4310 case SpvOpLogicalOr
:
4311 case SpvOpLogicalAnd
:
4312 case SpvOpLogicalNot
:
4313 case SpvOpBitwiseOr
:
4314 case SpvOpBitwiseXor
:
4315 case SpvOpBitwiseAnd
:
4317 case SpvOpFOrdEqual
:
4318 case SpvOpFUnordEqual
:
4319 case SpvOpINotEqual
:
4320 case SpvOpFOrdNotEqual
:
4321 case SpvOpFUnordNotEqual
:
4322 case SpvOpULessThan
:
4323 case SpvOpSLessThan
:
4324 case SpvOpFOrdLessThan
:
4325 case SpvOpFUnordLessThan
:
4326 case SpvOpUGreaterThan
:
4327 case SpvOpSGreaterThan
:
4328 case SpvOpFOrdGreaterThan
:
4329 case SpvOpFUnordGreaterThan
:
4330 case SpvOpULessThanEqual
:
4331 case SpvOpSLessThanEqual
:
4332 case SpvOpFOrdLessThanEqual
:
4333 case SpvOpFUnordLessThanEqual
:
4334 case SpvOpUGreaterThanEqual
:
4335 case SpvOpSGreaterThanEqual
:
4336 case SpvOpFOrdGreaterThanEqual
:
4337 case SpvOpFUnordGreaterThanEqual
:
4343 case SpvOpFwidthFine
:
4344 case SpvOpDPdxCoarse
:
4345 case SpvOpDPdyCoarse
:
4346 case SpvOpFwidthCoarse
:
4347 case SpvOpBitFieldInsert
:
4348 case SpvOpBitFieldSExtract
:
4349 case SpvOpBitFieldUExtract
:
4350 case SpvOpBitReverse
:
4352 case SpvOpTranspose
:
4353 case SpvOpOuterProduct
:
4354 case SpvOpMatrixTimesScalar
:
4355 case SpvOpVectorTimesMatrix
:
4356 case SpvOpMatrixTimesVector
:
4357 case SpvOpMatrixTimesMatrix
:
4358 vtn_handle_alu(b
, opcode
, w
, count
);
4361 case SpvOpVectorExtractDynamic
:
4362 case SpvOpVectorInsertDynamic
:
4363 case SpvOpVectorShuffle
:
4364 case SpvOpCompositeConstruct
:
4365 case SpvOpCompositeExtract
:
4366 case SpvOpCompositeInsert
:
4367 case SpvOpCopyObject
:
4368 vtn_handle_composite(b
, opcode
, w
, count
);
4371 case SpvOpEmitVertex
:
4372 case SpvOpEndPrimitive
:
4373 case SpvOpEmitStreamVertex
:
4374 case SpvOpEndStreamPrimitive
:
4375 case SpvOpControlBarrier
:
4376 case SpvOpMemoryBarrier
:
4377 vtn_handle_barrier(b
, opcode
, w
, count
);
4380 case SpvOpGroupNonUniformElect
:
4381 case SpvOpGroupNonUniformAll
:
4382 case SpvOpGroupNonUniformAny
:
4383 case SpvOpGroupNonUniformAllEqual
:
4384 case SpvOpGroupNonUniformBroadcast
:
4385 case SpvOpGroupNonUniformBroadcastFirst
:
4386 case SpvOpGroupNonUniformBallot
:
4387 case SpvOpGroupNonUniformInverseBallot
:
4388 case SpvOpGroupNonUniformBallotBitExtract
:
4389 case SpvOpGroupNonUniformBallotBitCount
:
4390 case SpvOpGroupNonUniformBallotFindLSB
:
4391 case SpvOpGroupNonUniformBallotFindMSB
:
4392 case SpvOpGroupNonUniformShuffle
:
4393 case SpvOpGroupNonUniformShuffleXor
:
4394 case SpvOpGroupNonUniformShuffleUp
:
4395 case SpvOpGroupNonUniformShuffleDown
:
4396 case SpvOpGroupNonUniformIAdd
:
4397 case SpvOpGroupNonUniformFAdd
:
4398 case SpvOpGroupNonUniformIMul
:
4399 case SpvOpGroupNonUniformFMul
:
4400 case SpvOpGroupNonUniformSMin
:
4401 case SpvOpGroupNonUniformUMin
:
4402 case SpvOpGroupNonUniformFMin
:
4403 case SpvOpGroupNonUniformSMax
:
4404 case SpvOpGroupNonUniformUMax
:
4405 case SpvOpGroupNonUniformFMax
:
4406 case SpvOpGroupNonUniformBitwiseAnd
:
4407 case SpvOpGroupNonUniformBitwiseOr
:
4408 case SpvOpGroupNonUniformBitwiseXor
:
4409 case SpvOpGroupNonUniformLogicalAnd
:
4410 case SpvOpGroupNonUniformLogicalOr
:
4411 case SpvOpGroupNonUniformLogicalXor
:
4412 case SpvOpGroupNonUniformQuadBroadcast
:
4413 case SpvOpGroupNonUniformQuadSwap
:
4414 vtn_handle_subgroup(b
, opcode
, w
, count
);
4418 vtn_fail_with_opcode("Unhandled opcode", opcode
);
4425 vtn_create_builder(const uint32_t *words
, size_t word_count
,
4426 gl_shader_stage stage
, const char *entry_point_name
,
4427 const struct spirv_to_nir_options
*options
)
4429 /* Initialize the vtn_builder object */
4430 struct vtn_builder
*b
= rzalloc(NULL
, struct vtn_builder
);
4431 struct spirv_to_nir_options
*dup_options
=
4432 ralloc(b
, struct spirv_to_nir_options
);
4433 *dup_options
= *options
;
4436 b
->spirv_word_count
= word_count
;
4440 exec_list_make_empty(&b
->functions
);
4441 b
->entry_point_stage
= stage
;
4442 b
->entry_point_name
= entry_point_name
;
4443 b
->options
= dup_options
;
4446 * Handle the SPIR-V header (first 5 dwords).
4447 * Can't use vtx_assert() as the setjmp(3) target isn't initialized yet.
4449 if (word_count
<= 5)
4452 if (words
[0] != SpvMagicNumber
) {
4453 vtn_err("words[0] was 0x%x, want 0x%x", words
[0], SpvMagicNumber
);
4456 if (words
[1] < 0x10000) {
4457 vtn_err("words[1] was 0x%x, want >= 0x10000", words
[1]);
4461 uint16_t generator_id
= words
[2] >> 16;
4462 uint16_t generator_version
= words
[2];
4464 /* The first GLSLang version bump actually 1.5 years after #179 was fixed
4465 * but this should at least let us shut the workaround off for modern
4466 * versions of GLSLang.
4468 b
->wa_glslang_179
= (generator_id
== 8 && generator_version
== 1);
4470 /* words[2] == generator magic */
4471 unsigned value_id_bound
= words
[3];
4472 if (words
[4] != 0) {
4473 vtn_err("words[4] was %u, want 0", words
[4]);
4477 b
->value_id_bound
= value_id_bound
;
4478 b
->values
= rzalloc_array(b
, struct vtn_value
, value_id_bound
);
4486 static nir_function
*
4487 vtn_emit_kernel_entry_point_wrapper(struct vtn_builder
*b
,
4488 nir_function
*entry_point
)
4490 vtn_assert(entry_point
== b
->entry_point
->func
->impl
->function
);
4491 vtn_fail_if(!entry_point
->name
, "entry points are required to have a name");
4492 const char *func_name
=
4493 ralloc_asprintf(b
->shader
, "__wrapped_%s", entry_point
->name
);
4495 /* we shouldn't have any inputs yet */
4496 vtn_assert(!entry_point
->shader
->num_inputs
);
4497 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_KERNEL
);
4499 nir_function
*main_entry_point
= nir_function_create(b
->shader
, func_name
);
4500 main_entry_point
->impl
= nir_function_impl_create(main_entry_point
);
4501 nir_builder_init(&b
->nb
, main_entry_point
->impl
);
4502 b
->nb
.cursor
= nir_after_cf_list(&main_entry_point
->impl
->body
);
4503 b
->func_param_idx
= 0;
4505 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, entry_point
);
4507 for (unsigned i
= 0; i
< entry_point
->num_params
; ++i
) {
4508 struct vtn_type
*param_type
= b
->entry_point
->func
->type
->params
[i
];
4510 /* consider all pointers to function memory to be parameters passed
4513 bool is_by_val
= param_type
->base_type
== vtn_base_type_pointer
&&
4514 param_type
->storage_class
== SpvStorageClassFunction
;
4516 /* input variable */
4517 nir_variable
*in_var
= rzalloc(b
->nb
.shader
, nir_variable
);
4518 in_var
->data
.mode
= nir_var_shader_in
;
4519 in_var
->data
.read_only
= true;
4520 in_var
->data
.location
= i
;
4523 in_var
->type
= param_type
->deref
->type
;
4525 in_var
->type
= param_type
->type
;
4527 nir_shader_add_variable(b
->nb
.shader
, in_var
);
4528 b
->nb
.shader
->num_inputs
++;
4530 /* we have to copy the entire variable into function memory */
4532 nir_variable
*copy_var
=
4533 nir_local_variable_create(main_entry_point
->impl
, in_var
->type
,
4535 nir_copy_var(&b
->nb
, copy_var
, in_var
);
4537 nir_src_for_ssa(&nir_build_deref_var(&b
->nb
, copy_var
)->dest
.ssa
);
4539 call
->params
[i
] = nir_src_for_ssa(nir_load_var(&b
->nb
, in_var
));
4543 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
4545 return main_entry_point
;
4549 spirv_to_nir(const uint32_t *words
, size_t word_count
,
4550 struct nir_spirv_specialization
*spec
, unsigned num_spec
,
4551 gl_shader_stage stage
, const char *entry_point_name
,
4552 const struct spirv_to_nir_options
*options
,
4553 const nir_shader_compiler_options
*nir_options
)
4556 const uint32_t *word_end
= words
+ word_count
;
4558 struct vtn_builder
*b
= vtn_create_builder(words
, word_count
,
4559 stage
, entry_point_name
,
4565 /* See also _vtn_fail() */
4566 if (setjmp(b
->fail_jump
)) {
4571 /* Skip the SPIR-V header, handled at vtn_create_builder */
4574 b
->shader
= nir_shader_create(b
, stage
, nir_options
, NULL
);
4576 /* Handle all the preamble instructions */
4577 words
= vtn_foreach_instruction(b
, words
, word_end
,
4578 vtn_handle_preamble_instruction
);
4580 if (b
->entry_point
== NULL
) {
4581 vtn_fail("Entry point not found");
4586 /* Set shader info defaults */
4587 b
->shader
->info
.gs
.invocations
= 1;
4589 b
->specializations
= spec
;
4590 b
->num_specializations
= num_spec
;
4592 /* Handle all variable, type, and constant instructions */
4593 words
= vtn_foreach_instruction(b
, words
, word_end
,
4594 vtn_handle_variable_or_type_instruction
);
4596 /* Parse execution modes */
4597 vtn_foreach_execution_mode(b
, b
->entry_point
,
4598 vtn_handle_execution_mode
, NULL
);
4600 if (b
->workgroup_size_builtin
) {
4601 vtn_assert(b
->workgroup_size_builtin
->type
->type
==
4602 glsl_vector_type(GLSL_TYPE_UINT
, 3));
4604 nir_const_value
*const_size
=
4605 b
->workgroup_size_builtin
->constant
->values
[0];
4607 b
->shader
->info
.cs
.local_size
[0] = const_size
[0].u32
;
4608 b
->shader
->info
.cs
.local_size
[1] = const_size
[1].u32
;
4609 b
->shader
->info
.cs
.local_size
[2] = const_size
[2].u32
;
4612 /* Set types on all vtn_values */
4613 vtn_foreach_instruction(b
, words
, word_end
, vtn_set_instruction_result_type
);
4615 vtn_build_cfg(b
, words
, word_end
);
4617 assert(b
->entry_point
->value_type
== vtn_value_type_function
);
4618 b
->entry_point
->func
->referenced
= true;
4623 foreach_list_typed(struct vtn_function
, func
, node
, &b
->functions
) {
4624 if (func
->referenced
&& !func
->emitted
) {
4625 b
->const_table
= _mesa_pointer_hash_table_create(b
);
4627 vtn_function_emit(b
, func
, vtn_handle_body_instruction
);
4633 vtn_assert(b
->entry_point
->value_type
== vtn_value_type_function
);
4634 nir_function
*entry_point
= b
->entry_point
->func
->impl
->function
;
4635 vtn_assert(entry_point
);
4637 /* post process entry_points with input params */
4638 if (entry_point
->num_params
&& b
->shader
->info
.stage
== MESA_SHADER_KERNEL
)
4639 entry_point
= vtn_emit_kernel_entry_point_wrapper(b
, entry_point
);
4641 entry_point
->is_entrypoint
= true;
4643 /* When multiple shader stages exist in the same SPIR-V module, we
4644 * generate input and output variables for every stage, in the same
4645 * NIR program. These dead variables can be invalid NIR. For example,
4646 * TCS outputs must be per-vertex arrays (or decorated 'patch'), while
4647 * VS output variables wouldn't be.
4649 * To ensure we have valid NIR, we eliminate any dead inputs and outputs
4650 * right away. In order to do so, we must lower any constant initializers
4651 * on outputs so nir_remove_dead_variables sees that they're written to.
4653 nir_lower_constant_initializers(b
->shader
, nir_var_shader_out
);
4654 nir_remove_dead_variables(b
->shader
,
4655 nir_var_shader_in
| nir_var_shader_out
);
4657 /* We sometimes generate bogus derefs that, while never used, give the
4658 * validator a bit of heartburn. Run dead code to get rid of them.
4660 nir_opt_dce(b
->shader
);
4662 /* Unparent the shader from the vtn_builder before we delete the builder */
4663 ralloc_steal(NULL
, b
->shader
);