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
.amd_gcn_shader
)) {
398 val
->ext_handler
= vtn_handle_amd_gcn_shader_instruction
;
399 } else if ((strcmp(ext
, "SPV_AMD_shader_ballot") == 0)
400 && (b
->options
&& b
->options
->caps
.amd_shader_ballot
)) {
401 val
->ext_handler
= vtn_handle_amd_shader_ballot_instruction
;
402 } else if ((strcmp(ext
, "SPV_AMD_shader_trinary_minmax") == 0)
403 && (b
->options
&& b
->options
->caps
.amd_trinary_minmax
)) {
404 val
->ext_handler
= vtn_handle_amd_shader_trinary_minmax_instruction
;
405 } else if (strcmp(ext
, "OpenCL.std") == 0) {
406 val
->ext_handler
= vtn_handle_opencl_instruction
;
408 vtn_fail("Unsupported extension: %s", ext
);
414 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
415 bool handled
= val
->ext_handler(b
, w
[4], w
, count
);
421 vtn_fail_with_opcode("Unhandled opcode", opcode
);
426 _foreach_decoration_helper(struct vtn_builder
*b
,
427 struct vtn_value
*base_value
,
429 struct vtn_value
*value
,
430 vtn_decoration_foreach_cb cb
, void *data
)
432 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
434 if (dec
->scope
== VTN_DEC_DECORATION
) {
435 member
= parent_member
;
436 } else if (dec
->scope
>= VTN_DEC_STRUCT_MEMBER0
) {
437 vtn_fail_if(value
->value_type
!= vtn_value_type_type
||
438 value
->type
->base_type
!= vtn_base_type_struct
,
439 "OpMemberDecorate and OpGroupMemberDecorate are only "
440 "allowed on OpTypeStruct");
441 /* This means we haven't recursed yet */
442 assert(value
== base_value
);
444 member
= dec
->scope
- VTN_DEC_STRUCT_MEMBER0
;
446 vtn_fail_if(member
>= base_value
->type
->length
,
447 "OpMemberDecorate specifies member %d but the "
448 "OpTypeStruct has only %u members",
449 member
, base_value
->type
->length
);
451 /* Not a decoration */
452 assert(dec
->scope
== VTN_DEC_EXECUTION_MODE
);
457 assert(dec
->group
->value_type
== vtn_value_type_decoration_group
);
458 _foreach_decoration_helper(b
, base_value
, member
, dec
->group
,
461 cb(b
, base_value
, member
, dec
, data
);
466 /** Iterates (recursively if needed) over all of the decorations on a value
468 * This function iterates over all of the decorations applied to a given
469 * value. If it encounters a decoration group, it recurses into the group
470 * and iterates over all of those decorations as well.
473 vtn_foreach_decoration(struct vtn_builder
*b
, struct vtn_value
*value
,
474 vtn_decoration_foreach_cb cb
, void *data
)
476 _foreach_decoration_helper(b
, value
, -1, value
, cb
, data
);
480 vtn_foreach_execution_mode(struct vtn_builder
*b
, struct vtn_value
*value
,
481 vtn_execution_mode_foreach_cb cb
, void *data
)
483 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
484 if (dec
->scope
!= VTN_DEC_EXECUTION_MODE
)
487 assert(dec
->group
== NULL
);
488 cb(b
, value
, dec
, data
);
493 vtn_handle_decoration(struct vtn_builder
*b
, SpvOp opcode
,
494 const uint32_t *w
, unsigned count
)
496 const uint32_t *w_end
= w
+ count
;
497 const uint32_t target
= w
[1];
501 case SpvOpDecorationGroup
:
502 vtn_push_value(b
, target
, vtn_value_type_decoration_group
);
506 case SpvOpDecorateId
:
507 case SpvOpMemberDecorate
:
508 case SpvOpDecorateString
:
509 case SpvOpMemberDecorateString
:
510 case SpvOpExecutionMode
:
511 case SpvOpExecutionModeId
: {
512 struct vtn_value
*val
= vtn_untyped_value(b
, target
);
514 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
517 case SpvOpDecorateId
:
518 case SpvOpDecorateString
:
519 dec
->scope
= VTN_DEC_DECORATION
;
521 case SpvOpMemberDecorate
:
522 case SpvOpMemberDecorateString
:
523 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(w
++);
524 vtn_fail_if(dec
->scope
< VTN_DEC_STRUCT_MEMBER0
, /* overflow */
525 "Member argument of OpMemberDecorate too large");
527 case SpvOpExecutionMode
:
528 case SpvOpExecutionModeId
:
529 dec
->scope
= VTN_DEC_EXECUTION_MODE
;
532 unreachable("Invalid decoration opcode");
534 dec
->decoration
= *(w
++);
537 /* Link into the list */
538 dec
->next
= val
->decoration
;
539 val
->decoration
= dec
;
543 case SpvOpGroupMemberDecorate
:
544 case SpvOpGroupDecorate
: {
545 struct vtn_value
*group
=
546 vtn_value(b
, target
, vtn_value_type_decoration_group
);
548 for (; w
< w_end
; w
++) {
549 struct vtn_value
*val
= vtn_untyped_value(b
, *w
);
550 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
553 if (opcode
== SpvOpGroupDecorate
) {
554 dec
->scope
= VTN_DEC_DECORATION
;
556 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(++w
);
557 vtn_fail_if(dec
->scope
< 0, /* Check for overflow */
558 "Member argument of OpGroupMemberDecorate too large");
561 /* Link into the list */
562 dec
->next
= val
->decoration
;
563 val
->decoration
= dec
;
569 unreachable("Unhandled opcode");
573 struct member_decoration_ctx
{
575 struct glsl_struct_field
*fields
;
576 struct vtn_type
*type
;
580 * Returns true if the given type contains a struct decorated Block or
584 vtn_type_contains_block(struct vtn_builder
*b
, struct vtn_type
*type
)
586 switch (type
->base_type
) {
587 case vtn_base_type_array
:
588 return vtn_type_contains_block(b
, type
->array_element
);
589 case vtn_base_type_struct
:
590 if (type
->block
|| type
->buffer_block
)
592 for (unsigned i
= 0; i
< type
->length
; i
++) {
593 if (vtn_type_contains_block(b
, type
->members
[i
]))
602 /** Returns true if two types are "compatible", i.e. you can do an OpLoad,
603 * OpStore, or OpCopyMemory between them without breaking anything.
604 * Technically, the SPIR-V rules require the exact same type ID but this lets
605 * us internally be a bit looser.
608 vtn_types_compatible(struct vtn_builder
*b
,
609 struct vtn_type
*t1
, struct vtn_type
*t2
)
611 if (t1
->id
== t2
->id
)
614 if (t1
->base_type
!= t2
->base_type
)
617 switch (t1
->base_type
) {
618 case vtn_base_type_void
:
619 case vtn_base_type_scalar
:
620 case vtn_base_type_vector
:
621 case vtn_base_type_matrix
:
622 case vtn_base_type_image
:
623 case vtn_base_type_sampler
:
624 case vtn_base_type_sampled_image
:
625 return t1
->type
== t2
->type
;
627 case vtn_base_type_array
:
628 return t1
->length
== t2
->length
&&
629 vtn_types_compatible(b
, t1
->array_element
, t2
->array_element
);
631 case vtn_base_type_pointer
:
632 return vtn_types_compatible(b
, t1
->deref
, t2
->deref
);
634 case vtn_base_type_struct
:
635 if (t1
->length
!= t2
->length
)
638 for (unsigned i
= 0; i
< t1
->length
; i
++) {
639 if (!vtn_types_compatible(b
, t1
->members
[i
], t2
->members
[i
]))
644 case vtn_base_type_function
:
645 /* This case shouldn't get hit since you can't copy around function
646 * types. Just require them to be identical.
651 vtn_fail("Invalid base type");
655 vtn_type_without_array(struct vtn_type
*type
)
657 while (type
->base_type
== vtn_base_type_array
)
658 type
= type
->array_element
;
662 /* does a shallow copy of a vtn_type */
664 static struct vtn_type
*
665 vtn_type_copy(struct vtn_builder
*b
, struct vtn_type
*src
)
667 struct vtn_type
*dest
= ralloc(b
, struct vtn_type
);
670 switch (src
->base_type
) {
671 case vtn_base_type_void
:
672 case vtn_base_type_scalar
:
673 case vtn_base_type_vector
:
674 case vtn_base_type_matrix
:
675 case vtn_base_type_array
:
676 case vtn_base_type_pointer
:
677 case vtn_base_type_image
:
678 case vtn_base_type_sampler
:
679 case vtn_base_type_sampled_image
:
680 /* Nothing more to do */
683 case vtn_base_type_struct
:
684 dest
->members
= ralloc_array(b
, struct vtn_type
*, src
->length
);
685 memcpy(dest
->members
, src
->members
,
686 src
->length
* sizeof(src
->members
[0]));
688 dest
->offsets
= ralloc_array(b
, unsigned, src
->length
);
689 memcpy(dest
->offsets
, src
->offsets
,
690 src
->length
* sizeof(src
->offsets
[0]));
693 case vtn_base_type_function
:
694 dest
->params
= ralloc_array(b
, struct vtn_type
*, src
->length
);
695 memcpy(dest
->params
, src
->params
, src
->length
* sizeof(src
->params
[0]));
702 static struct vtn_type
*
703 mutable_matrix_member(struct vtn_builder
*b
, struct vtn_type
*type
, int member
)
705 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
706 type
= type
->members
[member
];
708 /* We may have an array of matrices.... Oh, joy! */
709 while (glsl_type_is_array(type
->type
)) {
710 type
->array_element
= vtn_type_copy(b
, type
->array_element
);
711 type
= type
->array_element
;
714 vtn_assert(glsl_type_is_matrix(type
->type
));
720 vtn_handle_access_qualifier(struct vtn_builder
*b
, struct vtn_type
*type
,
721 int member
, enum gl_access_qualifier access
)
723 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
724 type
= type
->members
[member
];
726 type
->access
|= access
;
730 array_stride_decoration_cb(struct vtn_builder
*b
,
731 struct vtn_value
*val
, int member
,
732 const struct vtn_decoration
*dec
, void *void_ctx
)
734 struct vtn_type
*type
= val
->type
;
736 if (dec
->decoration
== SpvDecorationArrayStride
) {
737 vtn_fail_if(dec
->operands
[0] == 0, "ArrayStride must be non-zero");
738 type
->stride
= dec
->operands
[0];
743 struct_member_decoration_cb(struct vtn_builder
*b
,
744 struct vtn_value
*val
, int member
,
745 const struct vtn_decoration
*dec
, void *void_ctx
)
747 struct member_decoration_ctx
*ctx
= void_ctx
;
752 assert(member
< ctx
->num_fields
);
754 switch (dec
->decoration
) {
755 case SpvDecorationRelaxedPrecision
:
756 case SpvDecorationUniform
:
757 case SpvDecorationUniformId
:
758 break; /* FIXME: Do nothing with this for now. */
759 case SpvDecorationNonWritable
:
760 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_NON_WRITEABLE
);
762 case SpvDecorationNonReadable
:
763 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_NON_READABLE
);
765 case SpvDecorationVolatile
:
766 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_VOLATILE
);
768 case SpvDecorationCoherent
:
769 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_COHERENT
);
771 case SpvDecorationNoPerspective
:
772 ctx
->fields
[member
].interpolation
= INTERP_MODE_NOPERSPECTIVE
;
774 case SpvDecorationFlat
:
775 ctx
->fields
[member
].interpolation
= INTERP_MODE_FLAT
;
777 case SpvDecorationCentroid
:
778 ctx
->fields
[member
].centroid
= true;
780 case SpvDecorationSample
:
781 ctx
->fields
[member
].sample
= true;
783 case SpvDecorationStream
:
784 /* Vulkan only allows one GS stream */
785 vtn_assert(dec
->operands
[0] == 0);
787 case SpvDecorationLocation
:
788 ctx
->fields
[member
].location
= dec
->operands
[0];
790 case SpvDecorationComponent
:
791 break; /* FIXME: What should we do with these? */
792 case SpvDecorationBuiltIn
:
793 ctx
->type
->members
[member
] = vtn_type_copy(b
, ctx
->type
->members
[member
]);
794 ctx
->type
->members
[member
]->is_builtin
= true;
795 ctx
->type
->members
[member
]->builtin
= dec
->operands
[0];
796 ctx
->type
->builtin_block
= true;
798 case SpvDecorationOffset
:
799 ctx
->type
->offsets
[member
] = dec
->operands
[0];
800 ctx
->fields
[member
].offset
= dec
->operands
[0];
802 case SpvDecorationMatrixStride
:
803 /* Handled as a second pass */
805 case SpvDecorationColMajor
:
806 break; /* Nothing to do here. Column-major is the default. */
807 case SpvDecorationRowMajor
:
808 mutable_matrix_member(b
, ctx
->type
, member
)->row_major
= true;
811 case SpvDecorationPatch
:
814 case SpvDecorationSpecId
:
815 case SpvDecorationBlock
:
816 case SpvDecorationBufferBlock
:
817 case SpvDecorationArrayStride
:
818 case SpvDecorationGLSLShared
:
819 case SpvDecorationGLSLPacked
:
820 case SpvDecorationInvariant
:
821 case SpvDecorationRestrict
:
822 case SpvDecorationAliased
:
823 case SpvDecorationConstant
:
824 case SpvDecorationIndex
:
825 case SpvDecorationBinding
:
826 case SpvDecorationDescriptorSet
:
827 case SpvDecorationLinkageAttributes
:
828 case SpvDecorationNoContraction
:
829 case SpvDecorationInputAttachmentIndex
:
830 vtn_warn("Decoration not allowed on struct members: %s",
831 spirv_decoration_to_string(dec
->decoration
));
834 case SpvDecorationXfbBuffer
:
835 case SpvDecorationXfbStride
:
836 vtn_warn("Vulkan does not have transform feedback");
839 case SpvDecorationCPacked
:
840 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
)
841 vtn_warn("Decoration only allowed for CL-style kernels: %s",
842 spirv_decoration_to_string(dec
->decoration
));
844 ctx
->type
->packed
= true;
847 case SpvDecorationSaturatedConversion
:
848 case SpvDecorationFuncParamAttr
:
849 case SpvDecorationFPRoundingMode
:
850 case SpvDecorationFPFastMathMode
:
851 case SpvDecorationAlignment
:
852 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
) {
853 vtn_warn("Decoration only allowed for CL-style kernels: %s",
854 spirv_decoration_to_string(dec
->decoration
));
858 case SpvDecorationUserSemantic
:
859 /* User semantic decorations can safely be ignored by the driver. */
863 vtn_fail_with_decoration("Unhandled decoration", dec
->decoration
);
867 /** Chases the array type all the way down to the tail and rewrites the
868 * glsl_types to be based off the tail's glsl_type.
871 vtn_array_type_rewrite_glsl_type(struct vtn_type
*type
)
873 if (type
->base_type
!= vtn_base_type_array
)
876 vtn_array_type_rewrite_glsl_type(type
->array_element
);
878 type
->type
= glsl_array_type(type
->array_element
->type
,
879 type
->length
, type
->stride
);
882 /* Matrix strides are handled as a separate pass because we need to know
883 * whether the matrix is row-major or not first.
886 struct_member_matrix_stride_cb(struct vtn_builder
*b
,
887 struct vtn_value
*val
, int member
,
888 const struct vtn_decoration
*dec
,
891 if (dec
->decoration
!= SpvDecorationMatrixStride
)
894 vtn_fail_if(member
< 0,
895 "The MatrixStride decoration is only allowed on members "
897 vtn_fail_if(dec
->operands
[0] == 0, "MatrixStride must be non-zero");
899 struct member_decoration_ctx
*ctx
= void_ctx
;
901 struct vtn_type
*mat_type
= mutable_matrix_member(b
, ctx
->type
, member
);
902 if (mat_type
->row_major
) {
903 mat_type
->array_element
= vtn_type_copy(b
, mat_type
->array_element
);
904 mat_type
->stride
= mat_type
->array_element
->stride
;
905 mat_type
->array_element
->stride
= dec
->operands
[0];
907 mat_type
->type
= glsl_explicit_matrix_type(mat_type
->type
,
908 dec
->operands
[0], true);
909 mat_type
->array_element
->type
= glsl_get_column_type(mat_type
->type
);
911 vtn_assert(mat_type
->array_element
->stride
> 0);
912 mat_type
->stride
= dec
->operands
[0];
914 mat_type
->type
= glsl_explicit_matrix_type(mat_type
->type
,
915 dec
->operands
[0], false);
918 /* Now that we've replaced the glsl_type with a properly strided matrix
919 * type, rewrite the member type so that it's an array of the proper kind
922 vtn_array_type_rewrite_glsl_type(ctx
->type
->members
[member
]);
923 ctx
->fields
[member
].type
= ctx
->type
->members
[member
]->type
;
927 struct_block_decoration_cb(struct vtn_builder
*b
,
928 struct vtn_value
*val
, int member
,
929 const struct vtn_decoration
*dec
, void *ctx
)
934 struct vtn_type
*type
= val
->type
;
935 if (dec
->decoration
== SpvDecorationBlock
)
937 else if (dec
->decoration
== SpvDecorationBufferBlock
)
938 type
->buffer_block
= true;
942 type_decoration_cb(struct vtn_builder
*b
,
943 struct vtn_value
*val
, int member
,
944 const struct vtn_decoration
*dec
, void *ctx
)
946 struct vtn_type
*type
= val
->type
;
949 /* This should have been handled by OpTypeStruct */
950 assert(val
->type
->base_type
== vtn_base_type_struct
);
951 assert(member
>= 0 && member
< val
->type
->length
);
955 switch (dec
->decoration
) {
956 case SpvDecorationArrayStride
:
957 vtn_assert(type
->base_type
== vtn_base_type_array
||
958 type
->base_type
== vtn_base_type_pointer
);
960 case SpvDecorationBlock
:
961 vtn_assert(type
->base_type
== vtn_base_type_struct
);
962 vtn_assert(type
->block
);
964 case SpvDecorationBufferBlock
:
965 vtn_assert(type
->base_type
== vtn_base_type_struct
);
966 vtn_assert(type
->buffer_block
);
968 case SpvDecorationGLSLShared
:
969 case SpvDecorationGLSLPacked
:
970 /* Ignore these, since we get explicit offsets anyways */
973 case SpvDecorationRowMajor
:
974 case SpvDecorationColMajor
:
975 case SpvDecorationMatrixStride
:
976 case SpvDecorationBuiltIn
:
977 case SpvDecorationNoPerspective
:
978 case SpvDecorationFlat
:
979 case SpvDecorationPatch
:
980 case SpvDecorationCentroid
:
981 case SpvDecorationSample
:
982 case SpvDecorationVolatile
:
983 case SpvDecorationCoherent
:
984 case SpvDecorationNonWritable
:
985 case SpvDecorationNonReadable
:
986 case SpvDecorationUniform
:
987 case SpvDecorationUniformId
:
988 case SpvDecorationLocation
:
989 case SpvDecorationComponent
:
990 case SpvDecorationOffset
:
991 case SpvDecorationXfbBuffer
:
992 case SpvDecorationXfbStride
:
993 case SpvDecorationUserSemantic
:
994 vtn_warn("Decoration only allowed for struct members: %s",
995 spirv_decoration_to_string(dec
->decoration
));
998 case SpvDecorationStream
:
999 /* We don't need to do anything here, as stream is filled up when
1000 * aplying the decoration to a variable, just check that if it is not a
1001 * struct member, it should be a struct.
1003 vtn_assert(type
->base_type
== vtn_base_type_struct
);
1006 case SpvDecorationRelaxedPrecision
:
1007 case SpvDecorationSpecId
:
1008 case SpvDecorationInvariant
:
1009 case SpvDecorationRestrict
:
1010 case SpvDecorationAliased
:
1011 case SpvDecorationConstant
:
1012 case SpvDecorationIndex
:
1013 case SpvDecorationBinding
:
1014 case SpvDecorationDescriptorSet
:
1015 case SpvDecorationLinkageAttributes
:
1016 case SpvDecorationNoContraction
:
1017 case SpvDecorationInputAttachmentIndex
:
1018 vtn_warn("Decoration not allowed on types: %s",
1019 spirv_decoration_to_string(dec
->decoration
));
1022 case SpvDecorationCPacked
:
1023 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
)
1024 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1025 spirv_decoration_to_string(dec
->decoration
));
1027 type
->packed
= true;
1030 case SpvDecorationSaturatedConversion
:
1031 case SpvDecorationFuncParamAttr
:
1032 case SpvDecorationFPRoundingMode
:
1033 case SpvDecorationFPFastMathMode
:
1034 case SpvDecorationAlignment
:
1035 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1036 spirv_decoration_to_string(dec
->decoration
));
1040 vtn_fail_with_decoration("Unhandled decoration", dec
->decoration
);
1045 translate_image_format(struct vtn_builder
*b
, SpvImageFormat format
)
1048 case SpvImageFormatUnknown
: return 0; /* GL_NONE */
1049 case SpvImageFormatRgba32f
: return 0x8814; /* GL_RGBA32F */
1050 case SpvImageFormatRgba16f
: return 0x881A; /* GL_RGBA16F */
1051 case SpvImageFormatR32f
: return 0x822E; /* GL_R32F */
1052 case SpvImageFormatRgba8
: return 0x8058; /* GL_RGBA8 */
1053 case SpvImageFormatRgba8Snorm
: return 0x8F97; /* GL_RGBA8_SNORM */
1054 case SpvImageFormatRg32f
: return 0x8230; /* GL_RG32F */
1055 case SpvImageFormatRg16f
: return 0x822F; /* GL_RG16F */
1056 case SpvImageFormatR11fG11fB10f
: return 0x8C3A; /* GL_R11F_G11F_B10F */
1057 case SpvImageFormatR16f
: return 0x822D; /* GL_R16F */
1058 case SpvImageFormatRgba16
: return 0x805B; /* GL_RGBA16 */
1059 case SpvImageFormatRgb10A2
: return 0x8059; /* GL_RGB10_A2 */
1060 case SpvImageFormatRg16
: return 0x822C; /* GL_RG16 */
1061 case SpvImageFormatRg8
: return 0x822B; /* GL_RG8 */
1062 case SpvImageFormatR16
: return 0x822A; /* GL_R16 */
1063 case SpvImageFormatR8
: return 0x8229; /* GL_R8 */
1064 case SpvImageFormatRgba16Snorm
: return 0x8F9B; /* GL_RGBA16_SNORM */
1065 case SpvImageFormatRg16Snorm
: return 0x8F99; /* GL_RG16_SNORM */
1066 case SpvImageFormatRg8Snorm
: return 0x8F95; /* GL_RG8_SNORM */
1067 case SpvImageFormatR16Snorm
: return 0x8F98; /* GL_R16_SNORM */
1068 case SpvImageFormatR8Snorm
: return 0x8F94; /* GL_R8_SNORM */
1069 case SpvImageFormatRgba32i
: return 0x8D82; /* GL_RGBA32I */
1070 case SpvImageFormatRgba16i
: return 0x8D88; /* GL_RGBA16I */
1071 case SpvImageFormatRgba8i
: return 0x8D8E; /* GL_RGBA8I */
1072 case SpvImageFormatR32i
: return 0x8235; /* GL_R32I */
1073 case SpvImageFormatRg32i
: return 0x823B; /* GL_RG32I */
1074 case SpvImageFormatRg16i
: return 0x8239; /* GL_RG16I */
1075 case SpvImageFormatRg8i
: return 0x8237; /* GL_RG8I */
1076 case SpvImageFormatR16i
: return 0x8233; /* GL_R16I */
1077 case SpvImageFormatR8i
: return 0x8231; /* GL_R8I */
1078 case SpvImageFormatRgba32ui
: return 0x8D70; /* GL_RGBA32UI */
1079 case SpvImageFormatRgba16ui
: return 0x8D76; /* GL_RGBA16UI */
1080 case SpvImageFormatRgba8ui
: return 0x8D7C; /* GL_RGBA8UI */
1081 case SpvImageFormatR32ui
: return 0x8236; /* GL_R32UI */
1082 case SpvImageFormatRgb10a2ui
: return 0x906F; /* GL_RGB10_A2UI */
1083 case SpvImageFormatRg32ui
: return 0x823C; /* GL_RG32UI */
1084 case SpvImageFormatRg16ui
: return 0x823A; /* GL_RG16UI */
1085 case SpvImageFormatRg8ui
: return 0x8238; /* GL_RG8UI */
1086 case SpvImageFormatR16ui
: return 0x8234; /* GL_R16UI */
1087 case SpvImageFormatR8ui
: return 0x8232; /* GL_R8UI */
1089 vtn_fail("Invalid image format: %s (%u)",
1090 spirv_imageformat_to_string(format
), format
);
1094 static struct vtn_type
*
1095 vtn_type_layout_std430(struct vtn_builder
*b
, struct vtn_type
*type
,
1096 uint32_t *size_out
, uint32_t *align_out
)
1098 switch (type
->base_type
) {
1099 case vtn_base_type_scalar
: {
1100 uint32_t comp_size
= glsl_type_is_boolean(type
->type
)
1101 ? 4 : glsl_get_bit_size(type
->type
) / 8;
1102 *size_out
= comp_size
;
1103 *align_out
= comp_size
;
1107 case vtn_base_type_vector
: {
1108 uint32_t comp_size
= glsl_type_is_boolean(type
->type
)
1109 ? 4 : glsl_get_bit_size(type
->type
) / 8;
1110 unsigned align_comps
= type
->length
== 3 ? 4 : type
->length
;
1111 *size_out
= comp_size
* type
->length
,
1112 *align_out
= comp_size
* align_comps
;
1116 case vtn_base_type_matrix
:
1117 case vtn_base_type_array
: {
1118 /* We're going to add an array stride */
1119 type
= vtn_type_copy(b
, type
);
1120 uint32_t elem_size
, elem_align
;
1121 type
->array_element
= vtn_type_layout_std430(b
, type
->array_element
,
1122 &elem_size
, &elem_align
);
1123 type
->stride
= vtn_align_u32(elem_size
, elem_align
);
1124 *size_out
= type
->stride
* type
->length
;
1125 *align_out
= elem_align
;
1129 case vtn_base_type_struct
: {
1130 /* We're going to add member offsets */
1131 type
= vtn_type_copy(b
, type
);
1132 uint32_t offset
= 0;
1134 for (unsigned i
= 0; i
< type
->length
; i
++) {
1135 uint32_t mem_size
, mem_align
;
1136 type
->members
[i
] = vtn_type_layout_std430(b
, type
->members
[i
],
1137 &mem_size
, &mem_align
);
1138 offset
= vtn_align_u32(offset
, mem_align
);
1139 type
->offsets
[i
] = offset
;
1141 align
= MAX2(align
, mem_align
);
1149 unreachable("Invalid SPIR-V type for std430");
1154 vtn_handle_type(struct vtn_builder
*b
, SpvOp opcode
,
1155 const uint32_t *w
, unsigned count
)
1157 struct vtn_value
*val
= NULL
;
1159 /* In order to properly handle forward declarations, we have to defer
1160 * allocation for pointer types.
1162 if (opcode
!= SpvOpTypePointer
&& opcode
!= SpvOpTypeForwardPointer
) {
1163 val
= vtn_push_value(b
, w
[1], vtn_value_type_type
);
1164 vtn_fail_if(val
->type
!= NULL
,
1165 "Only pointers can have forward declarations");
1166 val
->type
= rzalloc(b
, struct vtn_type
);
1167 val
->type
->id
= w
[1];
1172 val
->type
->base_type
= vtn_base_type_void
;
1173 val
->type
->type
= glsl_void_type();
1176 val
->type
->base_type
= vtn_base_type_scalar
;
1177 val
->type
->type
= glsl_bool_type();
1178 val
->type
->length
= 1;
1180 case SpvOpTypeInt
: {
1181 int bit_size
= w
[2];
1182 const bool signedness
= w
[3];
1183 val
->type
->base_type
= vtn_base_type_scalar
;
1186 val
->type
->type
= (signedness
? glsl_int64_t_type() : glsl_uint64_t_type());
1189 val
->type
->type
= (signedness
? glsl_int_type() : glsl_uint_type());
1192 val
->type
->type
= (signedness
? glsl_int16_t_type() : glsl_uint16_t_type());
1195 val
->type
->type
= (signedness
? glsl_int8_t_type() : glsl_uint8_t_type());
1198 vtn_fail("Invalid int bit size: %u", bit_size
);
1200 val
->type
->length
= 1;
1204 case SpvOpTypeFloat
: {
1205 int bit_size
= w
[2];
1206 val
->type
->base_type
= vtn_base_type_scalar
;
1209 val
->type
->type
= glsl_float16_t_type();
1212 val
->type
->type
= glsl_float_type();
1215 val
->type
->type
= glsl_double_type();
1218 vtn_fail("Invalid float bit size: %u", bit_size
);
1220 val
->type
->length
= 1;
1224 case SpvOpTypeVector
: {
1225 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1226 unsigned elems
= w
[3];
1228 vtn_fail_if(base
->base_type
!= vtn_base_type_scalar
,
1229 "Base type for OpTypeVector must be a scalar");
1230 vtn_fail_if((elems
< 2 || elems
> 4) && (elems
!= 8) && (elems
!= 16),
1231 "Invalid component count for OpTypeVector");
1233 val
->type
->base_type
= vtn_base_type_vector
;
1234 val
->type
->type
= glsl_vector_type(glsl_get_base_type(base
->type
), elems
);
1235 val
->type
->length
= elems
;
1236 val
->type
->stride
= glsl_type_is_boolean(val
->type
->type
)
1237 ? 4 : glsl_get_bit_size(base
->type
) / 8;
1238 val
->type
->array_element
= base
;
1242 case SpvOpTypeMatrix
: {
1243 struct vtn_type
*base
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1244 unsigned columns
= w
[3];
1246 vtn_fail_if(base
->base_type
!= vtn_base_type_vector
,
1247 "Base type for OpTypeMatrix must be a vector");
1248 vtn_fail_if(columns
< 2 || columns
> 4,
1249 "Invalid column count for OpTypeMatrix");
1251 val
->type
->base_type
= vtn_base_type_matrix
;
1252 val
->type
->type
= glsl_matrix_type(glsl_get_base_type(base
->type
),
1253 glsl_get_vector_elements(base
->type
),
1255 vtn_fail_if(glsl_type_is_error(val
->type
->type
),
1256 "Unsupported base type for OpTypeMatrix");
1257 assert(!glsl_type_is_error(val
->type
->type
));
1258 val
->type
->length
= columns
;
1259 val
->type
->array_element
= base
;
1260 val
->type
->row_major
= false;
1261 val
->type
->stride
= 0;
1265 case SpvOpTypeRuntimeArray
:
1266 case SpvOpTypeArray
: {
1267 struct vtn_type
*array_element
=
1268 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1270 if (opcode
== SpvOpTypeRuntimeArray
) {
1271 /* A length of 0 is used to denote unsized arrays */
1272 val
->type
->length
= 0;
1274 val
->type
->length
= vtn_constant_uint(b
, w
[3]);
1277 val
->type
->base_type
= vtn_base_type_array
;
1278 val
->type
->array_element
= array_element
;
1279 if (b
->shader
->info
.stage
== MESA_SHADER_KERNEL
)
1280 val
->type
->stride
= glsl_get_cl_size(array_element
->type
);
1282 vtn_foreach_decoration(b
, val
, array_stride_decoration_cb
, NULL
);
1283 val
->type
->type
= glsl_array_type(array_element
->type
, val
->type
->length
,
1288 case SpvOpTypeStruct
: {
1289 unsigned num_fields
= count
- 2;
1290 val
->type
->base_type
= vtn_base_type_struct
;
1291 val
->type
->length
= num_fields
;
1292 val
->type
->members
= ralloc_array(b
, struct vtn_type
*, num_fields
);
1293 val
->type
->offsets
= ralloc_array(b
, unsigned, num_fields
);
1294 val
->type
->packed
= false;
1296 NIR_VLA(struct glsl_struct_field
, fields
, count
);
1297 for (unsigned i
= 0; i
< num_fields
; i
++) {
1298 val
->type
->members
[i
] =
1299 vtn_value(b
, w
[i
+ 2], vtn_value_type_type
)->type
;
1300 fields
[i
] = (struct glsl_struct_field
) {
1301 .type
= val
->type
->members
[i
]->type
,
1302 .name
= ralloc_asprintf(b
, "field%d", i
),
1308 if (b
->shader
->info
.stage
== MESA_SHADER_KERNEL
) {
1309 unsigned offset
= 0;
1310 for (unsigned i
= 0; i
< num_fields
; i
++) {
1311 offset
= align(offset
, glsl_get_cl_alignment(fields
[i
].type
));
1312 fields
[i
].offset
= offset
;
1313 offset
+= glsl_get_cl_size(fields
[i
].type
);
1317 struct member_decoration_ctx ctx
= {
1318 .num_fields
= num_fields
,
1323 vtn_foreach_decoration(b
, val
, struct_member_decoration_cb
, &ctx
);
1324 vtn_foreach_decoration(b
, val
, struct_member_matrix_stride_cb
, &ctx
);
1326 vtn_foreach_decoration(b
, val
, struct_block_decoration_cb
, NULL
);
1328 const char *name
= val
->name
;
1330 if (val
->type
->block
|| val
->type
->buffer_block
) {
1331 /* Packing will be ignored since types coming from SPIR-V are
1332 * explicitly laid out.
1334 val
->type
->type
= glsl_interface_type(fields
, num_fields
,
1335 /* packing */ 0, false,
1336 name
? name
: "block");
1338 val
->type
->type
= glsl_struct_type(fields
, num_fields
,
1339 name
? name
: "struct", false);
1344 case SpvOpTypeFunction
: {
1345 val
->type
->base_type
= vtn_base_type_function
;
1346 val
->type
->type
= NULL
;
1348 val
->type
->return_type
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1350 const unsigned num_params
= count
- 3;
1351 val
->type
->length
= num_params
;
1352 val
->type
->params
= ralloc_array(b
, struct vtn_type
*, num_params
);
1353 for (unsigned i
= 0; i
< count
- 3; i
++) {
1354 val
->type
->params
[i
] =
1355 vtn_value(b
, w
[i
+ 3], vtn_value_type_type
)->type
;
1360 case SpvOpTypePointer
:
1361 case SpvOpTypeForwardPointer
: {
1362 /* We can't blindly push the value because it might be a forward
1365 val
= vtn_untyped_value(b
, w
[1]);
1367 SpvStorageClass storage_class
= w
[2];
1369 if (val
->value_type
== vtn_value_type_invalid
) {
1370 val
->value_type
= vtn_value_type_type
;
1371 val
->type
= rzalloc(b
, struct vtn_type
);
1372 val
->type
->id
= w
[1];
1373 val
->type
->base_type
= vtn_base_type_pointer
;
1374 val
->type
->storage_class
= storage_class
;
1376 /* These can actually be stored to nir_variables and used as SSA
1377 * values so they need a real glsl_type.
1379 enum vtn_variable_mode mode
= vtn_storage_class_to_mode(
1380 b
, storage_class
, NULL
, NULL
);
1381 val
->type
->type
= nir_address_format_to_glsl_type(
1382 vtn_mode_to_address_format(b
, mode
));
1384 vtn_fail_if(val
->type
->storage_class
!= storage_class
,
1385 "The storage classes of an OpTypePointer and any "
1386 "OpTypeForwardPointers that provide forward "
1387 "declarations of it must match.");
1390 if (opcode
== SpvOpTypePointer
) {
1391 vtn_fail_if(val
->type
->deref
!= NULL
,
1392 "While OpTypeForwardPointer can be used to provide a "
1393 "forward declaration of a pointer, OpTypePointer can "
1394 "only be used once for a given id.");
1396 val
->type
->deref
= vtn_value(b
, w
[3], vtn_value_type_type
)->type
;
1398 vtn_foreach_decoration(b
, val
, array_stride_decoration_cb
, NULL
);
1400 if (b
->physical_ptrs
) {
1401 switch (storage_class
) {
1402 case SpvStorageClassFunction
:
1403 case SpvStorageClassWorkgroup
:
1404 case SpvStorageClassCrossWorkgroup
:
1405 val
->type
->stride
= align(glsl_get_cl_size(val
->type
->deref
->type
),
1406 glsl_get_cl_alignment(val
->type
->deref
->type
));
1413 if (storage_class
== SpvStorageClassWorkgroup
&&
1414 b
->options
->lower_workgroup_access_to_offsets
) {
1415 uint32_t size
, align
;
1416 val
->type
->deref
= vtn_type_layout_std430(b
, val
->type
->deref
,
1418 val
->type
->length
= size
;
1419 val
->type
->align
= align
;
1425 case SpvOpTypeImage
: {
1426 val
->type
->base_type
= vtn_base_type_image
;
1428 const struct vtn_type
*sampled_type
=
1429 vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1431 vtn_fail_if(sampled_type
->base_type
!= vtn_base_type_scalar
||
1432 glsl_get_bit_size(sampled_type
->type
) != 32,
1433 "Sampled type of OpTypeImage must be a 32-bit scalar");
1435 enum glsl_sampler_dim dim
;
1436 switch ((SpvDim
)w
[3]) {
1437 case SpvDim1D
: dim
= GLSL_SAMPLER_DIM_1D
; break;
1438 case SpvDim2D
: dim
= GLSL_SAMPLER_DIM_2D
; break;
1439 case SpvDim3D
: dim
= GLSL_SAMPLER_DIM_3D
; break;
1440 case SpvDimCube
: dim
= GLSL_SAMPLER_DIM_CUBE
; break;
1441 case SpvDimRect
: dim
= GLSL_SAMPLER_DIM_RECT
; break;
1442 case SpvDimBuffer
: dim
= GLSL_SAMPLER_DIM_BUF
; break;
1443 case SpvDimSubpassData
: dim
= GLSL_SAMPLER_DIM_SUBPASS
; break;
1445 vtn_fail("Invalid SPIR-V image dimensionality: %s (%u)",
1446 spirv_dim_to_string((SpvDim
)w
[3]), w
[3]);
1449 /* w[4]: as per Vulkan spec "Validation Rules within a Module",
1450 * The “Depth” operand of OpTypeImage is ignored.
1452 bool is_array
= w
[5];
1453 bool multisampled
= w
[6];
1454 unsigned sampled
= w
[7];
1455 SpvImageFormat format
= w
[8];
1458 val
->type
->access_qualifier
= w
[9];
1460 val
->type
->access_qualifier
= SpvAccessQualifierReadWrite
;
1463 if (dim
== GLSL_SAMPLER_DIM_2D
)
1464 dim
= GLSL_SAMPLER_DIM_MS
;
1465 else if (dim
== GLSL_SAMPLER_DIM_SUBPASS
)
1466 dim
= GLSL_SAMPLER_DIM_SUBPASS_MS
;
1468 vtn_fail("Unsupported multisampled image type");
1471 val
->type
->image_format
= translate_image_format(b
, format
);
1473 enum glsl_base_type sampled_base_type
=
1474 glsl_get_base_type(sampled_type
->type
);
1476 val
->type
->sampled
= true;
1477 val
->type
->type
= glsl_sampler_type(dim
, false, is_array
,
1479 } else if (sampled
== 2) {
1480 val
->type
->sampled
= false;
1481 val
->type
->type
= glsl_image_type(dim
, is_array
, sampled_base_type
);
1483 vtn_fail("We need to know if the image will be sampled");
1488 case SpvOpTypeSampledImage
:
1489 val
->type
->base_type
= vtn_base_type_sampled_image
;
1490 val
->type
->image
= vtn_value(b
, w
[2], vtn_value_type_type
)->type
;
1491 val
->type
->type
= val
->type
->image
->type
;
1494 case SpvOpTypeSampler
:
1495 /* The actual sampler type here doesn't really matter. It gets
1496 * thrown away the moment you combine it with an image. What really
1497 * matters is that it's a sampler type as opposed to an integer type
1498 * so the backend knows what to do.
1500 val
->type
->base_type
= vtn_base_type_sampler
;
1501 val
->type
->type
= glsl_bare_sampler_type();
1504 case SpvOpTypeOpaque
:
1505 case SpvOpTypeEvent
:
1506 case SpvOpTypeDeviceEvent
:
1507 case SpvOpTypeReserveId
:
1508 case SpvOpTypeQueue
:
1511 vtn_fail_with_opcode("Unhandled opcode", opcode
);
1514 vtn_foreach_decoration(b
, val
, type_decoration_cb
, NULL
);
1516 if (val
->type
->base_type
== vtn_base_type_struct
&&
1517 (val
->type
->block
|| val
->type
->buffer_block
)) {
1518 for (unsigned i
= 0; i
< val
->type
->length
; i
++) {
1519 vtn_fail_if(vtn_type_contains_block(b
, val
->type
->members
[i
]),
1520 "Block and BufferBlock decorations cannot decorate a "
1521 "structure type that is nested at any level inside "
1522 "another structure type decorated with Block or "
1528 static nir_constant
*
1529 vtn_null_constant(struct vtn_builder
*b
, struct vtn_type
*type
)
1531 nir_constant
*c
= rzalloc(b
, nir_constant
);
1533 switch (type
->base_type
) {
1534 case vtn_base_type_scalar
:
1535 case vtn_base_type_vector
:
1536 /* Nothing to do here. It's already initialized to zero */
1539 case vtn_base_type_pointer
: {
1540 enum vtn_variable_mode mode
= vtn_storage_class_to_mode(
1541 b
, type
->storage_class
, type
->deref
, NULL
);
1542 nir_address_format addr_format
= vtn_mode_to_address_format(b
, mode
);
1544 const nir_const_value
*null_value
= nir_address_format_null_value(addr_format
);
1545 memcpy(c
->values
[0], null_value
,
1546 sizeof(nir_const_value
) * nir_address_format_num_components(addr_format
));
1550 case vtn_base_type_void
:
1551 case vtn_base_type_image
:
1552 case vtn_base_type_sampler
:
1553 case vtn_base_type_sampled_image
:
1554 case vtn_base_type_function
:
1555 /* For those we have to return something but it doesn't matter what. */
1558 case vtn_base_type_matrix
:
1559 case vtn_base_type_array
:
1560 vtn_assert(type
->length
> 0);
1561 c
->num_elements
= type
->length
;
1562 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1564 c
->elements
[0] = vtn_null_constant(b
, type
->array_element
);
1565 for (unsigned i
= 1; i
< c
->num_elements
; i
++)
1566 c
->elements
[i
] = c
->elements
[0];
1569 case vtn_base_type_struct
:
1570 c
->num_elements
= type
->length
;
1571 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1572 for (unsigned i
= 0; i
< c
->num_elements
; i
++)
1573 c
->elements
[i
] = vtn_null_constant(b
, type
->members
[i
]);
1577 vtn_fail("Invalid type for null constant");
1584 spec_constant_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*v
,
1585 int member
, const struct vtn_decoration
*dec
,
1588 vtn_assert(member
== -1);
1589 if (dec
->decoration
!= SpvDecorationSpecId
)
1592 struct spec_constant_value
*const_value
= data
;
1594 for (unsigned i
= 0; i
< b
->num_specializations
; i
++) {
1595 if (b
->specializations
[i
].id
== dec
->operands
[0]) {
1596 if (const_value
->is_double
)
1597 const_value
->data64
= b
->specializations
[i
].data64
;
1599 const_value
->data32
= b
->specializations
[i
].data32
;
1606 get_specialization(struct vtn_builder
*b
, struct vtn_value
*val
,
1607 uint32_t const_value
)
1609 struct spec_constant_value data
;
1610 data
.is_double
= false;
1611 data
.data32
= const_value
;
1612 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1617 get_specialization64(struct vtn_builder
*b
, struct vtn_value
*val
,
1618 uint64_t const_value
)
1620 struct spec_constant_value data
;
1621 data
.is_double
= true;
1622 data
.data64
= const_value
;
1623 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &data
);
1628 handle_workgroup_size_decoration_cb(struct vtn_builder
*b
,
1629 struct vtn_value
*val
,
1631 const struct vtn_decoration
*dec
,
1634 vtn_assert(member
== -1);
1635 if (dec
->decoration
!= SpvDecorationBuiltIn
||
1636 dec
->operands
[0] != SpvBuiltInWorkgroupSize
)
1639 vtn_assert(val
->type
->type
== glsl_vector_type(GLSL_TYPE_UINT
, 3));
1640 b
->workgroup_size_builtin
= val
;
1644 vtn_handle_constant(struct vtn_builder
*b
, SpvOp opcode
,
1645 const uint32_t *w
, unsigned count
)
1647 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_constant
);
1648 val
->constant
= rzalloc(b
, nir_constant
);
1650 case SpvOpConstantTrue
:
1651 case SpvOpConstantFalse
:
1652 case SpvOpSpecConstantTrue
:
1653 case SpvOpSpecConstantFalse
: {
1654 vtn_fail_if(val
->type
->type
!= glsl_bool_type(),
1655 "Result type of %s must be OpTypeBool",
1656 spirv_op_to_string(opcode
));
1658 uint32_t int_val
= (opcode
== SpvOpConstantTrue
||
1659 opcode
== SpvOpSpecConstantTrue
);
1661 if (opcode
== SpvOpSpecConstantTrue
||
1662 opcode
== SpvOpSpecConstantFalse
)
1663 int_val
= get_specialization(b
, val
, int_val
);
1665 val
->constant
->values
[0][0].b
= int_val
!= 0;
1669 case SpvOpConstant
: {
1670 vtn_fail_if(val
->type
->base_type
!= vtn_base_type_scalar
,
1671 "Result type of %s must be a scalar",
1672 spirv_op_to_string(opcode
));
1673 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1676 val
->constant
->values
[0][0].u64
= vtn_u64_literal(&w
[3]);
1679 val
->constant
->values
[0][0].u32
= w
[3];
1682 val
->constant
->values
[0][0].u16
= w
[3];
1685 val
->constant
->values
[0][0].u8
= w
[3];
1688 vtn_fail("Unsupported SpvOpConstant bit size: %u", bit_size
);
1693 case SpvOpSpecConstant
: {
1694 vtn_fail_if(val
->type
->base_type
!= vtn_base_type_scalar
,
1695 "Result type of %s must be a scalar",
1696 spirv_op_to_string(opcode
));
1697 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1700 val
->constant
->values
[0][0].u64
=
1701 get_specialization64(b
, val
, vtn_u64_literal(&w
[3]));
1704 val
->constant
->values
[0][0].u32
= get_specialization(b
, val
, w
[3]);
1707 val
->constant
->values
[0][0].u16
= get_specialization(b
, val
, w
[3]);
1710 val
->constant
->values
[0][0].u8
= get_specialization(b
, val
, w
[3]);
1713 vtn_fail("Unsupported SpvOpSpecConstant bit size");
1718 case SpvOpSpecConstantComposite
:
1719 case SpvOpConstantComposite
: {
1720 unsigned elem_count
= count
- 3;
1721 vtn_fail_if(elem_count
!= val
->type
->length
,
1722 "%s has %u constituents, expected %u",
1723 spirv_op_to_string(opcode
), elem_count
, val
->type
->length
);
1725 nir_constant
**elems
= ralloc_array(b
, nir_constant
*, elem_count
);
1726 for (unsigned i
= 0; i
< elem_count
; i
++) {
1727 struct vtn_value
*val
= vtn_untyped_value(b
, w
[i
+ 3]);
1729 if (val
->value_type
== vtn_value_type_constant
) {
1730 elems
[i
] = val
->constant
;
1732 vtn_fail_if(val
->value_type
!= vtn_value_type_undef
,
1733 "only constants or undefs allowed for "
1734 "SpvOpConstantComposite");
1735 /* to make it easier, just insert a NULL constant for now */
1736 elems
[i
] = vtn_null_constant(b
, val
->type
);
1740 switch (val
->type
->base_type
) {
1741 case vtn_base_type_vector
: {
1742 assert(glsl_type_is_vector(val
->type
->type
));
1743 for (unsigned i
= 0; i
< elem_count
; i
++)
1744 val
->constant
->values
[0][i
] = elems
[i
]->values
[0][0];
1748 case vtn_base_type_matrix
:
1749 assert(glsl_type_is_matrix(val
->type
->type
));
1750 for (unsigned i
= 0; i
< elem_count
; i
++) {
1751 unsigned components
=
1752 glsl_get_components(glsl_get_column_type(val
->type
->type
));
1753 memcpy(val
->constant
->values
[i
], elems
[i
]->values
,
1754 sizeof(nir_const_value
) * components
);
1758 case vtn_base_type_struct
:
1759 case vtn_base_type_array
:
1760 ralloc_steal(val
->constant
, elems
);
1761 val
->constant
->num_elements
= elem_count
;
1762 val
->constant
->elements
= elems
;
1766 vtn_fail("Result type of %s must be a composite type",
1767 spirv_op_to_string(opcode
));
1772 case SpvOpSpecConstantOp
: {
1773 SpvOp opcode
= get_specialization(b
, val
, w
[3]);
1775 case SpvOpVectorShuffle
: {
1776 struct vtn_value
*v0
= &b
->values
[w
[4]];
1777 struct vtn_value
*v1
= &b
->values
[w
[5]];
1779 vtn_assert(v0
->value_type
== vtn_value_type_constant
||
1780 v0
->value_type
== vtn_value_type_undef
);
1781 vtn_assert(v1
->value_type
== vtn_value_type_constant
||
1782 v1
->value_type
== vtn_value_type_undef
);
1784 unsigned len0
= glsl_get_vector_elements(v0
->type
->type
);
1785 unsigned len1
= glsl_get_vector_elements(v1
->type
->type
);
1787 vtn_assert(len0
+ len1
< 16);
1789 unsigned bit_size
= glsl_get_bit_size(val
->type
->type
);
1790 unsigned bit_size0
= glsl_get_bit_size(v0
->type
->type
);
1791 unsigned bit_size1
= glsl_get_bit_size(v1
->type
->type
);
1793 vtn_assert(bit_size
== bit_size0
&& bit_size
== bit_size1
);
1794 (void)bit_size0
; (void)bit_size1
;
1796 nir_const_value undef
= { .u64
= 0xdeadbeefdeadbeef };
1797 nir_const_value combined
[NIR_MAX_VEC_COMPONENTS
* 2];
1799 if (v0
->value_type
== vtn_value_type_constant
) {
1800 for (unsigned i
= 0; i
< len0
; i
++)
1801 combined
[i
] = v0
->constant
->values
[0][i
];
1803 if (v1
->value_type
== vtn_value_type_constant
) {
1804 for (unsigned i
= 0; i
< len1
; i
++)
1805 combined
[len0
+ i
] = v1
->constant
->values
[0][i
];
1808 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1809 uint32_t comp
= w
[i
+ 6];
1810 if (comp
== (uint32_t)-1) {
1811 /* If component is not used, set the value to a known constant
1812 * to detect if it is wrongly used.
1814 val
->constant
->values
[0][j
] = undef
;
1816 vtn_fail_if(comp
>= len0
+ len1
,
1817 "All Component literals must either be FFFFFFFF "
1818 "or in [0, N - 1] (inclusive).");
1819 val
->constant
->values
[0][j
] = combined
[comp
];
1825 case SpvOpCompositeExtract
:
1826 case SpvOpCompositeInsert
: {
1827 struct vtn_value
*comp
;
1828 unsigned deref_start
;
1829 struct nir_constant
**c
;
1830 if (opcode
== SpvOpCompositeExtract
) {
1831 comp
= vtn_value(b
, w
[4], vtn_value_type_constant
);
1833 c
= &comp
->constant
;
1835 comp
= vtn_value(b
, w
[5], vtn_value_type_constant
);
1837 val
->constant
= nir_constant_clone(comp
->constant
,
1844 const struct vtn_type
*type
= comp
->type
;
1845 for (unsigned i
= deref_start
; i
< count
; i
++) {
1846 vtn_fail_if(w
[i
] > type
->length
,
1847 "%uth index of %s is %u but the type has only "
1848 "%u elements", i
- deref_start
,
1849 spirv_op_to_string(opcode
), w
[i
], type
->length
);
1851 switch (type
->base_type
) {
1852 case vtn_base_type_vector
:
1854 type
= type
->array_element
;
1857 case vtn_base_type_matrix
:
1858 assert(col
== 0 && elem
== -1);
1861 type
= type
->array_element
;
1864 case vtn_base_type_array
:
1865 c
= &(*c
)->elements
[w
[i
]];
1866 type
= type
->array_element
;
1869 case vtn_base_type_struct
:
1870 c
= &(*c
)->elements
[w
[i
]];
1871 type
= type
->members
[w
[i
]];
1875 vtn_fail("%s must only index into composite types",
1876 spirv_op_to_string(opcode
));
1880 if (opcode
== SpvOpCompositeExtract
) {
1884 unsigned num_components
= type
->length
;
1885 for (unsigned i
= 0; i
< num_components
; i
++)
1886 val
->constant
->values
[0][i
] = (*c
)->values
[col
][elem
+ i
];
1889 struct vtn_value
*insert
=
1890 vtn_value(b
, w
[4], vtn_value_type_constant
);
1891 vtn_assert(insert
->type
== type
);
1893 *c
= insert
->constant
;
1895 unsigned num_components
= type
->length
;
1896 for (unsigned i
= 0; i
< num_components
; i
++)
1897 (*c
)->values
[col
][elem
+ i
] = insert
->constant
->values
[0][i
];
1905 nir_alu_type dst_alu_type
= nir_get_nir_type_for_glsl_type(val
->type
->type
);
1906 nir_alu_type src_alu_type
= dst_alu_type
;
1907 unsigned num_components
= glsl_get_vector_elements(val
->type
->type
);
1910 vtn_assert(count
<= 7);
1916 /* We have a source in a conversion */
1918 nir_get_nir_type_for_glsl_type(
1919 vtn_value(b
, w
[4], vtn_value_type_constant
)->type
->type
);
1920 /* We use the bitsize of the conversion source to evaluate the opcode later */
1921 bit_size
= glsl_get_bit_size(
1922 vtn_value(b
, w
[4], vtn_value_type_constant
)->type
->type
);
1925 bit_size
= glsl_get_bit_size(val
->type
->type
);
1928 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(b
, opcode
, &swap
,
1929 nir_alu_type_get_type_size(src_alu_type
),
1930 nir_alu_type_get_type_size(dst_alu_type
));
1931 nir_const_value src
[3][NIR_MAX_VEC_COMPONENTS
];
1933 for (unsigned i
= 0; i
< count
- 4; i
++) {
1934 struct vtn_value
*src_val
=
1935 vtn_value(b
, w
[4 + i
], vtn_value_type_constant
);
1937 /* If this is an unsized source, pull the bit size from the
1938 * source; otherwise, we'll use the bit size from the destination.
1940 if (!nir_alu_type_get_type_size(nir_op_infos
[op
].input_types
[i
]))
1941 bit_size
= glsl_get_bit_size(src_val
->type
->type
);
1943 unsigned src_comps
= nir_op_infos
[op
].input_sizes
[i
] ?
1944 nir_op_infos
[op
].input_sizes
[i
] :
1947 unsigned j
= swap
? 1 - i
: i
;
1948 for (unsigned c
= 0; c
< src_comps
; c
++)
1949 src
[j
][c
] = src_val
->constant
->values
[0][c
];
1952 /* fix up fixed size sources */
1959 for (unsigned i
= 0; i
< num_components
; ++i
) {
1961 case 64: src
[1][i
].u32
= src
[1][i
].u64
; break;
1962 case 16: src
[1][i
].u32
= src
[1][i
].u16
; break;
1963 case 8: src
[1][i
].u32
= src
[1][i
].u8
; break;
1972 nir_const_value
*srcs
[3] = {
1973 src
[0], src
[1], src
[2],
1975 nir_eval_const_opcode(op
, val
->constant
->values
[0], num_components
, bit_size
, srcs
);
1982 case SpvOpConstantNull
:
1983 val
->constant
= vtn_null_constant(b
, val
->type
);
1986 case SpvOpConstantSampler
:
1987 vtn_fail("OpConstantSampler requires Kernel Capability");
1991 vtn_fail_with_opcode("Unhandled opcode", opcode
);
1994 /* Now that we have the value, update the workgroup size if needed */
1995 vtn_foreach_decoration(b
, val
, handle_workgroup_size_decoration_cb
, NULL
);
1998 struct vtn_ssa_value
*
1999 vtn_create_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
2001 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
2004 if (!glsl_type_is_vector_or_scalar(type
)) {
2005 unsigned elems
= glsl_get_length(type
);
2006 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
2007 for (unsigned i
= 0; i
< elems
; i
++) {
2008 const struct glsl_type
*child_type
;
2010 switch (glsl_get_base_type(type
)) {
2012 case GLSL_TYPE_UINT
:
2013 case GLSL_TYPE_INT16
:
2014 case GLSL_TYPE_UINT16
:
2015 case GLSL_TYPE_UINT8
:
2016 case GLSL_TYPE_INT8
:
2017 case GLSL_TYPE_INT64
:
2018 case GLSL_TYPE_UINT64
:
2019 case GLSL_TYPE_BOOL
:
2020 case GLSL_TYPE_FLOAT
:
2021 case GLSL_TYPE_FLOAT16
:
2022 case GLSL_TYPE_DOUBLE
:
2023 child_type
= glsl_get_column_type(type
);
2025 case GLSL_TYPE_ARRAY
:
2026 child_type
= glsl_get_array_element(type
);
2028 case GLSL_TYPE_STRUCT
:
2029 case GLSL_TYPE_INTERFACE
:
2030 child_type
= glsl_get_struct_field(type
, i
);
2033 vtn_fail("unkown base type");
2036 val
->elems
[i
] = vtn_create_ssa_value(b
, child_type
);
2044 vtn_tex_src(struct vtn_builder
*b
, unsigned index
, nir_tex_src_type type
)
2047 src
.src
= nir_src_for_ssa(vtn_ssa_value(b
, index
)->def
);
2048 src
.src_type
= type
;
2053 vtn_handle_texture(struct vtn_builder
*b
, SpvOp opcode
,
2054 const uint32_t *w
, unsigned count
)
2056 if (opcode
== SpvOpSampledImage
) {
2057 struct vtn_value
*val
=
2058 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
2059 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
2060 val
->sampled_image
->type
=
2061 vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2062 val
->sampled_image
->image
=
2063 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2064 val
->sampled_image
->sampler
=
2065 vtn_value(b
, w
[4], vtn_value_type_pointer
)->pointer
;
2067 } else if (opcode
== SpvOpImage
) {
2068 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
2069 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[3]);
2070 if (src_val
->value_type
== vtn_value_type_sampled_image
) {
2071 val
->pointer
= src_val
->sampled_image
->image
;
2073 vtn_assert(src_val
->value_type
== vtn_value_type_pointer
);
2074 val
->pointer
= src_val
->pointer
;
2079 struct vtn_type
*ret_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2080 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2082 struct vtn_sampled_image sampled
;
2083 struct vtn_value
*sampled_val
= vtn_untyped_value(b
, w
[3]);
2084 if (sampled_val
->value_type
== vtn_value_type_sampled_image
) {
2085 sampled
= *sampled_val
->sampled_image
;
2087 vtn_assert(sampled_val
->value_type
== vtn_value_type_pointer
);
2088 sampled
.type
= sampled_val
->pointer
->type
;
2089 sampled
.image
= NULL
;
2090 sampled
.sampler
= sampled_val
->pointer
;
2093 const struct glsl_type
*image_type
= sampled
.type
->type
;
2094 const enum glsl_sampler_dim sampler_dim
= glsl_get_sampler_dim(image_type
);
2095 const bool is_array
= glsl_sampler_type_is_array(image_type
);
2097 /* Figure out the base texture operation */
2100 case SpvOpImageSampleImplicitLod
:
2101 case SpvOpImageSampleDrefImplicitLod
:
2102 case SpvOpImageSampleProjImplicitLod
:
2103 case SpvOpImageSampleProjDrefImplicitLod
:
2104 texop
= nir_texop_tex
;
2107 case SpvOpImageSampleExplicitLod
:
2108 case SpvOpImageSampleDrefExplicitLod
:
2109 case SpvOpImageSampleProjExplicitLod
:
2110 case SpvOpImageSampleProjDrefExplicitLod
:
2111 texop
= nir_texop_txl
;
2114 case SpvOpImageFetch
:
2115 if (glsl_get_sampler_dim(image_type
) == GLSL_SAMPLER_DIM_MS
) {
2116 texop
= nir_texop_txf_ms
;
2118 texop
= nir_texop_txf
;
2122 case SpvOpImageGather
:
2123 case SpvOpImageDrefGather
:
2124 texop
= nir_texop_tg4
;
2127 case SpvOpImageQuerySizeLod
:
2128 case SpvOpImageQuerySize
:
2129 texop
= nir_texop_txs
;
2132 case SpvOpImageQueryLod
:
2133 texop
= nir_texop_lod
;
2136 case SpvOpImageQueryLevels
:
2137 texop
= nir_texop_query_levels
;
2140 case SpvOpImageQuerySamples
:
2141 texop
= nir_texop_texture_samples
;
2145 vtn_fail_with_opcode("Unhandled opcode", opcode
);
2148 nir_tex_src srcs
[10]; /* 10 should be enough */
2149 nir_tex_src
*p
= srcs
;
2151 nir_deref_instr
*sampler
= vtn_pointer_to_deref(b
, sampled
.sampler
);
2152 nir_deref_instr
*texture
=
2153 sampled
.image
? vtn_pointer_to_deref(b
, sampled
.image
) : sampler
;
2155 p
->src
= nir_src_for_ssa(&texture
->dest
.ssa
);
2156 p
->src_type
= nir_tex_src_texture_deref
;
2166 /* These operations require a sampler */
2167 p
->src
= nir_src_for_ssa(&sampler
->dest
.ssa
);
2168 p
->src_type
= nir_tex_src_sampler_deref
;
2172 case nir_texop_txf_ms
:
2174 case nir_texop_query_levels
:
2175 case nir_texop_texture_samples
:
2176 case nir_texop_samples_identical
:
2179 case nir_texop_txf_ms_fb
:
2180 vtn_fail("unexpected nir_texop_txf_ms_fb");
2182 case nir_texop_txf_ms_mcs
:
2183 vtn_fail("unexpected nir_texop_txf_ms_mcs");
2188 struct nir_ssa_def
*coord
;
2189 unsigned coord_components
;
2191 case SpvOpImageSampleImplicitLod
:
2192 case SpvOpImageSampleExplicitLod
:
2193 case SpvOpImageSampleDrefImplicitLod
:
2194 case SpvOpImageSampleDrefExplicitLod
:
2195 case SpvOpImageSampleProjImplicitLod
:
2196 case SpvOpImageSampleProjExplicitLod
:
2197 case SpvOpImageSampleProjDrefImplicitLod
:
2198 case SpvOpImageSampleProjDrefExplicitLod
:
2199 case SpvOpImageFetch
:
2200 case SpvOpImageGather
:
2201 case SpvOpImageDrefGather
:
2202 case SpvOpImageQueryLod
: {
2203 /* All these types have the coordinate as their first real argument */
2204 switch (sampler_dim
) {
2205 case GLSL_SAMPLER_DIM_1D
:
2206 case GLSL_SAMPLER_DIM_BUF
:
2207 coord_components
= 1;
2209 case GLSL_SAMPLER_DIM_2D
:
2210 case GLSL_SAMPLER_DIM_RECT
:
2211 case GLSL_SAMPLER_DIM_MS
:
2212 coord_components
= 2;
2214 case GLSL_SAMPLER_DIM_3D
:
2215 case GLSL_SAMPLER_DIM_CUBE
:
2216 coord_components
= 3;
2219 vtn_fail("Invalid sampler type");
2222 if (is_array
&& texop
!= nir_texop_lod
)
2225 coord
= vtn_ssa_value(b
, w
[idx
++])->def
;
2226 p
->src
= nir_src_for_ssa(nir_channels(&b
->nb
, coord
,
2227 (1 << coord_components
) - 1));
2228 p
->src_type
= nir_tex_src_coord
;
2235 coord_components
= 0;
2240 case SpvOpImageSampleProjImplicitLod
:
2241 case SpvOpImageSampleProjExplicitLod
:
2242 case SpvOpImageSampleProjDrefImplicitLod
:
2243 case SpvOpImageSampleProjDrefExplicitLod
:
2244 /* These have the projector as the last coordinate component */
2245 p
->src
= nir_src_for_ssa(nir_channel(&b
->nb
, coord
, coord_components
));
2246 p
->src_type
= nir_tex_src_projector
;
2254 bool is_shadow
= false;
2255 unsigned gather_component
= 0;
2257 case SpvOpImageSampleDrefImplicitLod
:
2258 case SpvOpImageSampleDrefExplicitLod
:
2259 case SpvOpImageSampleProjDrefImplicitLod
:
2260 case SpvOpImageSampleProjDrefExplicitLod
:
2261 case SpvOpImageDrefGather
:
2262 /* These all have an explicit depth value as their next source */
2264 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_comparator
);
2267 case SpvOpImageGather
:
2268 /* This has a component as its next source */
2269 gather_component
= vtn_constant_uint(b
, w
[idx
++]);
2276 /* For OpImageQuerySizeLod, we always have an LOD */
2277 if (opcode
== SpvOpImageQuerySizeLod
)
2278 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
2280 /* Now we need to handle some number of optional arguments */
2281 struct vtn_value
*gather_offsets
= NULL
;
2283 uint32_t operands
= w
[idx
++];
2285 if (operands
& SpvImageOperandsBiasMask
) {
2286 vtn_assert(texop
== nir_texop_tex
);
2287 texop
= nir_texop_txb
;
2288 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_bias
);
2291 if (operands
& SpvImageOperandsLodMask
) {
2292 vtn_assert(texop
== nir_texop_txl
|| texop
== nir_texop_txf
||
2293 texop
== nir_texop_txs
);
2294 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
2297 if (operands
& SpvImageOperandsGradMask
) {
2298 vtn_assert(texop
== nir_texop_txl
);
2299 texop
= nir_texop_txd
;
2300 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddx
);
2301 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ddy
);
2304 if (operands
& SpvImageOperandsOffsetMask
||
2305 operands
& SpvImageOperandsConstOffsetMask
)
2306 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_offset
);
2308 if (operands
& SpvImageOperandsConstOffsetsMask
) {
2309 vtn_assert(texop
== nir_texop_tg4
);
2310 gather_offsets
= vtn_value(b
, w
[idx
++], vtn_value_type_constant
);
2313 if (operands
& SpvImageOperandsSampleMask
) {
2314 vtn_assert(texop
== nir_texop_txf_ms
);
2315 texop
= nir_texop_txf_ms
;
2316 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ms_index
);
2319 if (operands
& SpvImageOperandsMinLodMask
) {
2320 vtn_assert(texop
== nir_texop_tex
||
2321 texop
== nir_texop_txb
||
2322 texop
== nir_texop_txd
);
2323 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_min_lod
);
2326 /* We should have now consumed exactly all of the arguments */
2327 vtn_assert(idx
== count
);
2329 nir_tex_instr
*instr
= nir_tex_instr_create(b
->shader
, p
- srcs
);
2332 memcpy(instr
->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
2334 instr
->coord_components
= coord_components
;
2335 instr
->sampler_dim
= sampler_dim
;
2336 instr
->is_array
= is_array
;
2337 instr
->is_shadow
= is_shadow
;
2338 instr
->is_new_style_shadow
=
2339 is_shadow
&& glsl_get_components(ret_type
->type
) == 1;
2340 instr
->component
= gather_component
;
2342 if (sampled
.image
&& (sampled
.image
->access
& ACCESS_NON_UNIFORM
))
2343 instr
->texture_non_uniform
= true;
2345 if (sampled
.sampler
&& (sampled
.sampler
->access
& ACCESS_NON_UNIFORM
))
2346 instr
->sampler_non_uniform
= true;
2348 switch (glsl_get_sampler_result_type(image_type
)) {
2349 case GLSL_TYPE_FLOAT
: instr
->dest_type
= nir_type_float
; break;
2350 case GLSL_TYPE_INT
: instr
->dest_type
= nir_type_int
; break;
2351 case GLSL_TYPE_UINT
: instr
->dest_type
= nir_type_uint
; break;
2352 case GLSL_TYPE_BOOL
: instr
->dest_type
= nir_type_bool
; break;
2354 vtn_fail("Invalid base type for sampler result");
2357 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
2358 nir_tex_instr_dest_size(instr
), 32, NULL
);
2360 vtn_assert(glsl_get_vector_elements(ret_type
->type
) ==
2361 nir_tex_instr_dest_size(instr
));
2363 if (gather_offsets
) {
2364 vtn_fail_if(gather_offsets
->type
->base_type
!= vtn_base_type_array
||
2365 gather_offsets
->type
->length
!= 4,
2366 "ConstOffsets must be an array of size four of vectors "
2367 "of two integer components");
2369 struct vtn_type
*vec_type
= gather_offsets
->type
->array_element
;
2370 vtn_fail_if(vec_type
->base_type
!= vtn_base_type_vector
||
2371 vec_type
->length
!= 2 ||
2372 !glsl_type_is_integer(vec_type
->type
),
2373 "ConstOffsets must be an array of size four of vectors "
2374 "of two integer components");
2376 unsigned bit_size
= glsl_get_bit_size(vec_type
->type
);
2377 for (uint32_t i
= 0; i
< 4; i
++) {
2378 const nir_const_value
*cvec
=
2379 gather_offsets
->constant
->elements
[i
]->values
[0];
2380 for (uint32_t j
= 0; j
< 2; j
++) {
2382 case 8: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i8
; break;
2383 case 16: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i16
; break;
2384 case 32: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i32
; break;
2385 case 64: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i64
; break;
2387 vtn_fail("Unsupported bit size: %u", bit_size
);
2393 val
->ssa
= vtn_create_ssa_value(b
, ret_type
->type
);
2394 val
->ssa
->def
= &instr
->dest
.ssa
;
2396 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
2400 fill_common_atomic_sources(struct vtn_builder
*b
, SpvOp opcode
,
2401 const uint32_t *w
, nir_src
*src
)
2404 case SpvOpAtomicIIncrement
:
2405 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
2408 case SpvOpAtomicIDecrement
:
2409 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
2412 case SpvOpAtomicISub
:
2414 nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_ssa_value(b
, w
[6])->def
));
2417 case SpvOpAtomicCompareExchange
:
2418 case SpvOpAtomicCompareExchangeWeak
:
2419 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[8])->def
);
2420 src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[7])->def
);
2423 case SpvOpAtomicExchange
:
2424 case SpvOpAtomicIAdd
:
2425 case SpvOpAtomicSMin
:
2426 case SpvOpAtomicUMin
:
2427 case SpvOpAtomicSMax
:
2428 case SpvOpAtomicUMax
:
2429 case SpvOpAtomicAnd
:
2431 case SpvOpAtomicXor
:
2432 src
[0] = nir_src_for_ssa(vtn_ssa_value(b
, w
[6])->def
);
2436 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
2440 static nir_ssa_def
*
2441 get_image_coord(struct vtn_builder
*b
, uint32_t value
)
2443 struct vtn_ssa_value
*coord
= vtn_ssa_value(b
, value
);
2445 /* The image_load_store intrinsics assume a 4-dim coordinate */
2446 unsigned dim
= glsl_get_vector_elements(coord
->type
);
2447 unsigned swizzle
[4];
2448 for (unsigned i
= 0; i
< 4; i
++)
2449 swizzle
[i
] = MIN2(i
, dim
- 1);
2451 return nir_swizzle(&b
->nb
, coord
->def
, swizzle
, 4);
2454 static nir_ssa_def
*
2455 expand_to_vec4(nir_builder
*b
, nir_ssa_def
*value
)
2457 if (value
->num_components
== 4)
2461 for (unsigned i
= 0; i
< 4; i
++)
2462 swiz
[i
] = i
< value
->num_components
? i
: 0;
2463 return nir_swizzle(b
, value
, swiz
, 4);
2467 vtn_handle_image(struct vtn_builder
*b
, SpvOp opcode
,
2468 const uint32_t *w
, unsigned count
)
2470 /* Just get this one out of the way */
2471 if (opcode
== SpvOpImageTexelPointer
) {
2472 struct vtn_value
*val
=
2473 vtn_push_value(b
, w
[2], vtn_value_type_image_pointer
);
2474 val
->image
= ralloc(b
, struct vtn_image_pointer
);
2476 val
->image
->image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2477 val
->image
->coord
= get_image_coord(b
, w
[4]);
2478 val
->image
->sample
= vtn_ssa_value(b
, w
[5])->def
;
2482 struct vtn_image_pointer image
;
2485 case SpvOpAtomicExchange
:
2486 case SpvOpAtomicCompareExchange
:
2487 case SpvOpAtomicCompareExchangeWeak
:
2488 case SpvOpAtomicIIncrement
:
2489 case SpvOpAtomicIDecrement
:
2490 case SpvOpAtomicIAdd
:
2491 case SpvOpAtomicISub
:
2492 case SpvOpAtomicLoad
:
2493 case SpvOpAtomicSMin
:
2494 case SpvOpAtomicUMin
:
2495 case SpvOpAtomicSMax
:
2496 case SpvOpAtomicUMax
:
2497 case SpvOpAtomicAnd
:
2499 case SpvOpAtomicXor
:
2500 image
= *vtn_value(b
, w
[3], vtn_value_type_image_pointer
)->image
;
2503 case SpvOpAtomicStore
:
2504 image
= *vtn_value(b
, w
[1], vtn_value_type_image_pointer
)->image
;
2507 case SpvOpImageQuerySize
:
2508 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2510 image
.sample
= NULL
;
2513 case SpvOpImageRead
:
2514 image
.image
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2515 image
.coord
= get_image_coord(b
, w
[4]);
2517 if (count
> 5 && (w
[5] & SpvImageOperandsSampleMask
)) {
2518 vtn_assert(w
[5] == SpvImageOperandsSampleMask
);
2519 image
.sample
= vtn_ssa_value(b
, w
[6])->def
;
2521 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2525 case SpvOpImageWrite
:
2526 image
.image
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
2527 image
.coord
= get_image_coord(b
, w
[2]);
2531 if (count
> 4 && (w
[4] & SpvImageOperandsSampleMask
)) {
2532 vtn_assert(w
[4] == SpvImageOperandsSampleMask
);
2533 image
.sample
= vtn_ssa_value(b
, w
[5])->def
;
2535 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2540 vtn_fail_with_opcode("Invalid image opcode", opcode
);
2543 nir_intrinsic_op op
;
2545 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_deref_##N; break;
2546 OP(ImageQuerySize
, size
)
2548 OP(ImageWrite
, store
)
2549 OP(AtomicLoad
, load
)
2550 OP(AtomicStore
, store
)
2551 OP(AtomicExchange
, atomic_exchange
)
2552 OP(AtomicCompareExchange
, atomic_comp_swap
)
2553 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
2554 OP(AtomicIIncrement
, atomic_add
)
2555 OP(AtomicIDecrement
, atomic_add
)
2556 OP(AtomicIAdd
, atomic_add
)
2557 OP(AtomicISub
, atomic_add
)
2558 OP(AtomicSMin
, atomic_min
)
2559 OP(AtomicUMin
, atomic_min
)
2560 OP(AtomicSMax
, atomic_max
)
2561 OP(AtomicUMax
, atomic_max
)
2562 OP(AtomicAnd
, atomic_and
)
2563 OP(AtomicOr
, atomic_or
)
2564 OP(AtomicXor
, atomic_xor
)
2567 vtn_fail_with_opcode("Invalid image opcode", opcode
);
2570 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
2572 nir_deref_instr
*image_deref
= vtn_pointer_to_deref(b
, image
.image
);
2573 intrin
->src
[0] = nir_src_for_ssa(&image_deref
->dest
.ssa
);
2575 /* ImageQuerySize doesn't take any extra parameters */
2576 if (opcode
!= SpvOpImageQuerySize
) {
2577 /* The image coordinate is always 4 components but we may not have that
2578 * many. Swizzle to compensate.
2580 intrin
->src
[1] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, image
.coord
));
2581 intrin
->src
[2] = nir_src_for_ssa(image
.sample
);
2585 case SpvOpAtomicLoad
:
2586 case SpvOpImageQuerySize
:
2587 case SpvOpImageRead
:
2589 case SpvOpAtomicStore
:
2590 case SpvOpImageWrite
: {
2591 const uint32_t value_id
= opcode
== SpvOpAtomicStore
? w
[4] : w
[3];
2592 nir_ssa_def
*value
= vtn_ssa_value(b
, value_id
)->def
;
2593 /* nir_intrinsic_image_deref_store always takes a vec4 value */
2594 assert(op
== nir_intrinsic_image_deref_store
);
2595 intrin
->num_components
= 4;
2596 intrin
->src
[3] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, value
));
2600 case SpvOpAtomicCompareExchange
:
2601 case SpvOpAtomicCompareExchangeWeak
:
2602 case SpvOpAtomicIIncrement
:
2603 case SpvOpAtomicIDecrement
:
2604 case SpvOpAtomicExchange
:
2605 case SpvOpAtomicIAdd
:
2606 case SpvOpAtomicISub
:
2607 case SpvOpAtomicSMin
:
2608 case SpvOpAtomicUMin
:
2609 case SpvOpAtomicSMax
:
2610 case SpvOpAtomicUMax
:
2611 case SpvOpAtomicAnd
:
2613 case SpvOpAtomicXor
:
2614 fill_common_atomic_sources(b
, opcode
, w
, &intrin
->src
[3]);
2618 vtn_fail_with_opcode("Invalid image opcode", opcode
);
2621 if (opcode
!= SpvOpImageWrite
&& opcode
!= SpvOpAtomicStore
) {
2622 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2623 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2625 unsigned dest_components
= glsl_get_vector_elements(type
->type
);
2626 intrin
->num_components
= nir_intrinsic_infos
[op
].dest_components
;
2627 if (intrin
->num_components
== 0)
2628 intrin
->num_components
= dest_components
;
2630 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
,
2631 intrin
->num_components
, 32, NULL
);
2633 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2635 nir_ssa_def
*result
= &intrin
->dest
.ssa
;
2636 if (intrin
->num_components
!= dest_components
)
2637 result
= nir_channels(&b
->nb
, result
, (1 << dest_components
) - 1);
2639 val
->ssa
= vtn_create_ssa_value(b
, type
->type
);
2640 val
->ssa
->def
= result
;
2642 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
2646 static nir_intrinsic_op
2647 get_ssbo_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2650 case SpvOpAtomicLoad
: return nir_intrinsic_load_ssbo
;
2651 case SpvOpAtomicStore
: return nir_intrinsic_store_ssbo
;
2652 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
2653 OP(AtomicExchange
, atomic_exchange
)
2654 OP(AtomicCompareExchange
, atomic_comp_swap
)
2655 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
2656 OP(AtomicIIncrement
, atomic_add
)
2657 OP(AtomicIDecrement
, atomic_add
)
2658 OP(AtomicIAdd
, atomic_add
)
2659 OP(AtomicISub
, atomic_add
)
2660 OP(AtomicSMin
, atomic_imin
)
2661 OP(AtomicUMin
, atomic_umin
)
2662 OP(AtomicSMax
, atomic_imax
)
2663 OP(AtomicUMax
, atomic_umax
)
2664 OP(AtomicAnd
, atomic_and
)
2665 OP(AtomicOr
, atomic_or
)
2666 OP(AtomicXor
, atomic_xor
)
2669 vtn_fail_with_opcode("Invalid SSBO atomic", opcode
);
2673 static nir_intrinsic_op
2674 get_uniform_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2677 #define OP(S, N) case SpvOp##S: return nir_intrinsic_atomic_counter_ ##N;
2678 OP(AtomicLoad
, read_deref
)
2679 OP(AtomicExchange
, exchange
)
2680 OP(AtomicCompareExchange
, comp_swap
)
2681 OP(AtomicCompareExchangeWeak
, comp_swap
)
2682 OP(AtomicIIncrement
, inc_deref
)
2683 OP(AtomicIDecrement
, post_dec_deref
)
2684 OP(AtomicIAdd
, add_deref
)
2685 OP(AtomicISub
, add_deref
)
2686 OP(AtomicUMin
, min_deref
)
2687 OP(AtomicUMax
, max_deref
)
2688 OP(AtomicAnd
, and_deref
)
2689 OP(AtomicOr
, or_deref
)
2690 OP(AtomicXor
, xor_deref
)
2693 /* We left the following out: AtomicStore, AtomicSMin and
2694 * AtomicSmax. Right now there are not nir intrinsics for them. At this
2695 * moment Atomic Counter support is needed for ARB_spirv support, so is
2696 * only need to support GLSL Atomic Counters that are uints and don't
2697 * allow direct storage.
2699 unreachable("Invalid uniform atomic");
2703 static nir_intrinsic_op
2704 get_shared_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2707 case SpvOpAtomicLoad
: return nir_intrinsic_load_shared
;
2708 case SpvOpAtomicStore
: return nir_intrinsic_store_shared
;
2709 #define OP(S, N) case SpvOp##S: return nir_intrinsic_shared_##N;
2710 OP(AtomicExchange
, atomic_exchange
)
2711 OP(AtomicCompareExchange
, atomic_comp_swap
)
2712 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
2713 OP(AtomicIIncrement
, atomic_add
)
2714 OP(AtomicIDecrement
, atomic_add
)
2715 OP(AtomicIAdd
, atomic_add
)
2716 OP(AtomicISub
, atomic_add
)
2717 OP(AtomicSMin
, atomic_imin
)
2718 OP(AtomicUMin
, atomic_umin
)
2719 OP(AtomicSMax
, atomic_imax
)
2720 OP(AtomicUMax
, atomic_umax
)
2721 OP(AtomicAnd
, atomic_and
)
2722 OP(AtomicOr
, atomic_or
)
2723 OP(AtomicXor
, atomic_xor
)
2726 vtn_fail_with_opcode("Invalid shared atomic", opcode
);
2730 static nir_intrinsic_op
2731 get_deref_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
2734 case SpvOpAtomicLoad
: return nir_intrinsic_load_deref
;
2735 case SpvOpAtomicStore
: return nir_intrinsic_store_deref
;
2736 #define OP(S, N) case SpvOp##S: return nir_intrinsic_deref_##N;
2737 OP(AtomicExchange
, atomic_exchange
)
2738 OP(AtomicCompareExchange
, atomic_comp_swap
)
2739 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
2740 OP(AtomicIIncrement
, atomic_add
)
2741 OP(AtomicIDecrement
, atomic_add
)
2742 OP(AtomicIAdd
, atomic_add
)
2743 OP(AtomicISub
, atomic_add
)
2744 OP(AtomicSMin
, atomic_imin
)
2745 OP(AtomicUMin
, atomic_umin
)
2746 OP(AtomicSMax
, atomic_imax
)
2747 OP(AtomicUMax
, atomic_umax
)
2748 OP(AtomicAnd
, atomic_and
)
2749 OP(AtomicOr
, atomic_or
)
2750 OP(AtomicXor
, atomic_xor
)
2753 vtn_fail_with_opcode("Invalid shared atomic", opcode
);
2758 * Handles shared atomics, ssbo atomics and atomic counters.
2761 vtn_handle_atomics(struct vtn_builder
*b
, SpvOp opcode
,
2762 const uint32_t *w
, unsigned count
)
2764 struct vtn_pointer
*ptr
;
2765 nir_intrinsic_instr
*atomic
;
2768 case SpvOpAtomicLoad
:
2769 case SpvOpAtomicExchange
:
2770 case SpvOpAtomicCompareExchange
:
2771 case SpvOpAtomicCompareExchangeWeak
:
2772 case SpvOpAtomicIIncrement
:
2773 case SpvOpAtomicIDecrement
:
2774 case SpvOpAtomicIAdd
:
2775 case SpvOpAtomicISub
:
2776 case SpvOpAtomicSMin
:
2777 case SpvOpAtomicUMin
:
2778 case SpvOpAtomicSMax
:
2779 case SpvOpAtomicUMax
:
2780 case SpvOpAtomicAnd
:
2782 case SpvOpAtomicXor
:
2783 ptr
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2786 case SpvOpAtomicStore
:
2787 ptr
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
2791 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
2795 SpvScope scope = w[4];
2796 SpvMemorySemanticsMask semantics = w[5];
2799 /* uniform as "atomic counter uniform" */
2800 if (ptr
->mode
== vtn_variable_mode_uniform
) {
2801 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
2802 const struct glsl_type
*deref_type
= deref
->type
;
2803 nir_intrinsic_op op
= get_uniform_nir_atomic_op(b
, opcode
);
2804 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2805 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
2807 /* SSBO needs to initialize index/offset. In this case we don't need to,
2808 * as that info is already stored on the ptr->var->var nir_variable (see
2809 * vtn_create_variable)
2813 case SpvOpAtomicLoad
:
2814 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2817 case SpvOpAtomicStore
:
2818 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2819 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2822 case SpvOpAtomicExchange
:
2823 case SpvOpAtomicCompareExchange
:
2824 case SpvOpAtomicCompareExchangeWeak
:
2825 case SpvOpAtomicIIncrement
:
2826 case SpvOpAtomicIDecrement
:
2827 case SpvOpAtomicIAdd
:
2828 case SpvOpAtomicISub
:
2829 case SpvOpAtomicSMin
:
2830 case SpvOpAtomicUMin
:
2831 case SpvOpAtomicSMax
:
2832 case SpvOpAtomicUMax
:
2833 case SpvOpAtomicAnd
:
2835 case SpvOpAtomicXor
:
2836 /* Nothing: we don't need to call fill_common_atomic_sources here, as
2837 * atomic counter uniforms doesn't have sources
2842 unreachable("Invalid SPIR-V atomic");
2845 } else if (vtn_pointer_uses_ssa_offset(b
, ptr
)) {
2846 nir_ssa_def
*offset
, *index
;
2847 offset
= vtn_pointer_to_offset(b
, ptr
, &index
);
2849 nir_intrinsic_op op
;
2850 if (ptr
->mode
== vtn_variable_mode_ssbo
) {
2851 op
= get_ssbo_nir_atomic_op(b
, opcode
);
2853 vtn_assert(ptr
->mode
== vtn_variable_mode_workgroup
&&
2854 b
->options
->lower_workgroup_access_to_offsets
);
2855 op
= get_shared_nir_atomic_op(b
, opcode
);
2858 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2862 case SpvOpAtomicLoad
:
2863 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
2864 nir_intrinsic_set_align(atomic
, 4, 0);
2865 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2866 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2867 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2870 case SpvOpAtomicStore
:
2871 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
2872 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2873 nir_intrinsic_set_align(atomic
, 4, 0);
2874 atomic
->src
[src
++] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2875 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2876 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2877 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2880 case SpvOpAtomicExchange
:
2881 case SpvOpAtomicCompareExchange
:
2882 case SpvOpAtomicCompareExchangeWeak
:
2883 case SpvOpAtomicIIncrement
:
2884 case SpvOpAtomicIDecrement
:
2885 case SpvOpAtomicIAdd
:
2886 case SpvOpAtomicISub
:
2887 case SpvOpAtomicSMin
:
2888 case SpvOpAtomicUMin
:
2889 case SpvOpAtomicSMax
:
2890 case SpvOpAtomicUMax
:
2891 case SpvOpAtomicAnd
:
2893 case SpvOpAtomicXor
:
2894 if (ptr
->mode
== vtn_variable_mode_ssbo
)
2895 atomic
->src
[src
++] = nir_src_for_ssa(index
);
2896 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
2897 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[src
]);
2901 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
2904 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
2905 const struct glsl_type
*deref_type
= deref
->type
;
2906 nir_intrinsic_op op
= get_deref_nir_atomic_op(b
, opcode
);
2907 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
2908 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
2911 case SpvOpAtomicLoad
:
2912 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2915 case SpvOpAtomicStore
:
2916 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
2917 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
2918 atomic
->src
[1] = nir_src_for_ssa(vtn_ssa_value(b
, w
[4])->def
);
2921 case SpvOpAtomicExchange
:
2922 case SpvOpAtomicCompareExchange
:
2923 case SpvOpAtomicCompareExchangeWeak
:
2924 case SpvOpAtomicIIncrement
:
2925 case SpvOpAtomicIDecrement
:
2926 case SpvOpAtomicIAdd
:
2927 case SpvOpAtomicISub
:
2928 case SpvOpAtomicSMin
:
2929 case SpvOpAtomicUMin
:
2930 case SpvOpAtomicSMax
:
2931 case SpvOpAtomicUMax
:
2932 case SpvOpAtomicAnd
:
2934 case SpvOpAtomicXor
:
2935 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[1]);
2939 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
2943 if (opcode
!= SpvOpAtomicStore
) {
2944 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2946 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
,
2947 glsl_get_vector_elements(type
->type
),
2948 glsl_get_bit_size(type
->type
), NULL
);
2950 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2951 val
->ssa
= rzalloc(b
, struct vtn_ssa_value
);
2952 val
->ssa
->def
= &atomic
->dest
.ssa
;
2953 val
->ssa
->type
= type
->type
;
2956 nir_builder_instr_insert(&b
->nb
, &atomic
->instr
);
2959 static nir_alu_instr
*
2960 create_vec(struct vtn_builder
*b
, unsigned num_components
, unsigned bit_size
)
2962 nir_op op
= nir_op_vec(num_components
);
2963 nir_alu_instr
*vec
= nir_alu_instr_create(b
->shader
, op
);
2964 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, num_components
,
2966 vec
->dest
.write_mask
= (1 << num_components
) - 1;
2971 struct vtn_ssa_value
*
2972 vtn_ssa_transpose(struct vtn_builder
*b
, struct vtn_ssa_value
*src
)
2974 if (src
->transposed
)
2975 return src
->transposed
;
2977 struct vtn_ssa_value
*dest
=
2978 vtn_create_ssa_value(b
, glsl_transposed_type(src
->type
));
2980 for (unsigned i
= 0; i
< glsl_get_matrix_columns(dest
->type
); i
++) {
2981 nir_alu_instr
*vec
= create_vec(b
, glsl_get_matrix_columns(src
->type
),
2982 glsl_get_bit_size(src
->type
));
2983 if (glsl_type_is_vector_or_scalar(src
->type
)) {
2984 vec
->src
[0].src
= nir_src_for_ssa(src
->def
);
2985 vec
->src
[0].swizzle
[0] = i
;
2987 for (unsigned j
= 0; j
< glsl_get_matrix_columns(src
->type
); j
++) {
2988 vec
->src
[j
].src
= nir_src_for_ssa(src
->elems
[j
]->def
);
2989 vec
->src
[j
].swizzle
[0] = i
;
2992 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
2993 dest
->elems
[i
]->def
= &vec
->dest
.dest
.ssa
;
2996 dest
->transposed
= src
;
3002 vtn_vector_extract(struct vtn_builder
*b
, nir_ssa_def
*src
, unsigned index
)
3004 return nir_channel(&b
->nb
, src
, index
);
3008 vtn_vector_insert(struct vtn_builder
*b
, nir_ssa_def
*src
, nir_ssa_def
*insert
,
3011 nir_alu_instr
*vec
= create_vec(b
, src
->num_components
,
3014 for (unsigned i
= 0; i
< src
->num_components
; i
++) {
3016 vec
->src
[i
].src
= nir_src_for_ssa(insert
);
3018 vec
->src
[i
].src
= nir_src_for_ssa(src
);
3019 vec
->src
[i
].swizzle
[0] = i
;
3023 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3025 return &vec
->dest
.dest
.ssa
;
3028 static nir_ssa_def
*
3029 nir_ieq_imm(nir_builder
*b
, nir_ssa_def
*x
, uint64_t i
)
3031 return nir_ieq(b
, x
, nir_imm_intN_t(b
, i
, x
->bit_size
));
3035 vtn_vector_extract_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
3038 return nir_vector_extract(&b
->nb
, src
, nir_i2i(&b
->nb
, index
, 32));
3042 vtn_vector_insert_dynamic(struct vtn_builder
*b
, nir_ssa_def
*src
,
3043 nir_ssa_def
*insert
, nir_ssa_def
*index
)
3045 nir_ssa_def
*dest
= vtn_vector_insert(b
, src
, insert
, 0);
3046 for (unsigned i
= 1; i
< src
->num_components
; i
++)
3047 dest
= nir_bcsel(&b
->nb
, nir_ieq_imm(&b
->nb
, index
, i
),
3048 vtn_vector_insert(b
, src
, insert
, i
), dest
);
3053 static nir_ssa_def
*
3054 vtn_vector_shuffle(struct vtn_builder
*b
, unsigned num_components
,
3055 nir_ssa_def
*src0
, nir_ssa_def
*src1
,
3056 const uint32_t *indices
)
3058 nir_alu_instr
*vec
= create_vec(b
, num_components
, src0
->bit_size
);
3060 for (unsigned i
= 0; i
< num_components
; i
++) {
3061 uint32_t index
= indices
[i
];
3062 if (index
== 0xffffffff) {
3064 nir_src_for_ssa(nir_ssa_undef(&b
->nb
, 1, src0
->bit_size
));
3065 } else if (index
< src0
->num_components
) {
3066 vec
->src
[i
].src
= nir_src_for_ssa(src0
);
3067 vec
->src
[i
].swizzle
[0] = index
;
3069 vec
->src
[i
].src
= nir_src_for_ssa(src1
);
3070 vec
->src
[i
].swizzle
[0] = index
- src0
->num_components
;
3074 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3076 return &vec
->dest
.dest
.ssa
;
3080 * Concatentates a number of vectors/scalars together to produce a vector
3082 static nir_ssa_def
*
3083 vtn_vector_construct(struct vtn_builder
*b
, unsigned num_components
,
3084 unsigned num_srcs
, nir_ssa_def
**srcs
)
3086 nir_alu_instr
*vec
= create_vec(b
, num_components
, srcs
[0]->bit_size
);
3088 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
3090 * "When constructing a vector, there must be at least two Constituent
3093 vtn_assert(num_srcs
>= 2);
3095 unsigned dest_idx
= 0;
3096 for (unsigned i
= 0; i
< num_srcs
; i
++) {
3097 nir_ssa_def
*src
= srcs
[i
];
3098 vtn_assert(dest_idx
+ src
->num_components
<= num_components
);
3099 for (unsigned j
= 0; j
< src
->num_components
; j
++) {
3100 vec
->src
[dest_idx
].src
= nir_src_for_ssa(src
);
3101 vec
->src
[dest_idx
].swizzle
[0] = j
;
3106 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
3108 * "When constructing a vector, the total number of components in all
3109 * the operands must equal the number of components in Result Type."
3111 vtn_assert(dest_idx
== num_components
);
3113 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3115 return &vec
->dest
.dest
.ssa
;
3118 static struct vtn_ssa_value
*
3119 vtn_composite_copy(void *mem_ctx
, struct vtn_ssa_value
*src
)
3121 struct vtn_ssa_value
*dest
= rzalloc(mem_ctx
, struct vtn_ssa_value
);
3122 dest
->type
= src
->type
;
3124 if (glsl_type_is_vector_or_scalar(src
->type
)) {
3125 dest
->def
= src
->def
;
3127 unsigned elems
= glsl_get_length(src
->type
);
3129 dest
->elems
= ralloc_array(mem_ctx
, struct vtn_ssa_value
*, elems
);
3130 for (unsigned i
= 0; i
< elems
; i
++)
3131 dest
->elems
[i
] = vtn_composite_copy(mem_ctx
, src
->elems
[i
]);
3137 static struct vtn_ssa_value
*
3138 vtn_composite_insert(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
3139 struct vtn_ssa_value
*insert
, const uint32_t *indices
,
3140 unsigned num_indices
)
3142 struct vtn_ssa_value
*dest
= vtn_composite_copy(b
, src
);
3144 struct vtn_ssa_value
*cur
= dest
;
3146 for (i
= 0; i
< num_indices
- 1; i
++) {
3147 cur
= cur
->elems
[indices
[i
]];
3150 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
3151 /* According to the SPIR-V spec, OpCompositeInsert may work down to
3152 * the component granularity. In that case, the last index will be
3153 * the index to insert the scalar into the vector.
3156 cur
->def
= vtn_vector_insert(b
, cur
->def
, insert
->def
, indices
[i
]);
3158 cur
->elems
[indices
[i
]] = insert
;
3164 static struct vtn_ssa_value
*
3165 vtn_composite_extract(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
3166 const uint32_t *indices
, unsigned num_indices
)
3168 struct vtn_ssa_value
*cur
= src
;
3169 for (unsigned i
= 0; i
< num_indices
; i
++) {
3170 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
3171 vtn_assert(i
== num_indices
- 1);
3172 /* According to the SPIR-V spec, OpCompositeExtract may work down to
3173 * the component granularity. The last index will be the index of the
3174 * vector to extract.
3177 struct vtn_ssa_value
*ret
= rzalloc(b
, struct vtn_ssa_value
);
3178 ret
->type
= glsl_scalar_type(glsl_get_base_type(cur
->type
));
3179 ret
->def
= vtn_vector_extract(b
, cur
->def
, indices
[i
]);
3182 cur
= cur
->elems
[indices
[i
]];
3190 vtn_handle_composite(struct vtn_builder
*b
, SpvOp opcode
,
3191 const uint32_t *w
, unsigned count
)
3193 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
3194 const struct glsl_type
*type
=
3195 vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
3196 val
->ssa
= vtn_create_ssa_value(b
, type
);
3199 case SpvOpVectorExtractDynamic
:
3200 val
->ssa
->def
= vtn_vector_extract_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
3201 vtn_ssa_value(b
, w
[4])->def
);
3204 case SpvOpVectorInsertDynamic
:
3205 val
->ssa
->def
= vtn_vector_insert_dynamic(b
, vtn_ssa_value(b
, w
[3])->def
,
3206 vtn_ssa_value(b
, w
[4])->def
,
3207 vtn_ssa_value(b
, w
[5])->def
);
3210 case SpvOpVectorShuffle
:
3211 val
->ssa
->def
= vtn_vector_shuffle(b
, glsl_get_vector_elements(type
),
3212 vtn_ssa_value(b
, w
[3])->def
,
3213 vtn_ssa_value(b
, w
[4])->def
,
3217 case SpvOpCompositeConstruct
: {
3218 unsigned elems
= count
- 3;
3220 if (glsl_type_is_vector_or_scalar(type
)) {
3221 nir_ssa_def
*srcs
[NIR_MAX_VEC_COMPONENTS
];
3222 for (unsigned i
= 0; i
< elems
; i
++)
3223 srcs
[i
] = vtn_ssa_value(b
, w
[3 + i
])->def
;
3225 vtn_vector_construct(b
, glsl_get_vector_elements(type
),
3228 val
->ssa
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
3229 for (unsigned i
= 0; i
< elems
; i
++)
3230 val
->ssa
->elems
[i
] = vtn_ssa_value(b
, w
[3 + i
]);
3234 case SpvOpCompositeExtract
:
3235 val
->ssa
= vtn_composite_extract(b
, vtn_ssa_value(b
, w
[3]),
3239 case SpvOpCompositeInsert
:
3240 val
->ssa
= vtn_composite_insert(b
, vtn_ssa_value(b
, w
[4]),
3241 vtn_ssa_value(b
, w
[3]),
3245 case SpvOpCopyLogical
:
3246 case SpvOpCopyObject
:
3247 val
->ssa
= vtn_composite_copy(b
, vtn_ssa_value(b
, w
[3]));
3251 vtn_fail_with_opcode("unknown composite operation", opcode
);
3256 vtn_emit_barrier(struct vtn_builder
*b
, nir_intrinsic_op op
)
3258 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
3259 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3263 vtn_emit_memory_barrier(struct vtn_builder
*b
, SpvScope scope
,
3264 SpvMemorySemanticsMask semantics
)
3266 static const SpvMemorySemanticsMask all_memory_semantics
=
3267 SpvMemorySemanticsUniformMemoryMask
|
3268 SpvMemorySemanticsWorkgroupMemoryMask
|
3269 SpvMemorySemanticsAtomicCounterMemoryMask
|
3270 SpvMemorySemanticsImageMemoryMask
;
3272 /* If we're not actually doing a memory barrier, bail */
3273 if (!(semantics
& all_memory_semantics
))
3276 /* GL and Vulkan don't have these */
3277 vtn_assert(scope
!= SpvScopeCrossDevice
);
3279 if (scope
== SpvScopeSubgroup
)
3280 return; /* Nothing to do here */
3282 if (scope
== SpvScopeWorkgroup
) {
3283 vtn_emit_barrier(b
, nir_intrinsic_group_memory_barrier
);
3287 /* There's only two scopes thing left */
3288 vtn_assert(scope
== SpvScopeInvocation
|| scope
== SpvScopeDevice
);
3290 if ((semantics
& all_memory_semantics
) == all_memory_semantics
) {
3291 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier
);
3295 /* Issue a bunch of more specific barriers */
3296 uint32_t bits
= semantics
;
3298 SpvMemorySemanticsMask semantic
= 1 << u_bit_scan(&bits
);
3300 case SpvMemorySemanticsUniformMemoryMask
:
3301 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_buffer
);
3303 case SpvMemorySemanticsWorkgroupMemoryMask
:
3304 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_shared
);
3306 case SpvMemorySemanticsAtomicCounterMemoryMask
:
3307 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_atomic_counter
);
3309 case SpvMemorySemanticsImageMemoryMask
:
3310 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_image
);
3319 vtn_handle_barrier(struct vtn_builder
*b
, SpvOp opcode
,
3320 const uint32_t *w
, unsigned count
)
3323 case SpvOpEmitVertex
:
3324 case SpvOpEmitStreamVertex
:
3325 case SpvOpEndPrimitive
:
3326 case SpvOpEndStreamPrimitive
: {
3327 nir_intrinsic_op intrinsic_op
;
3329 case SpvOpEmitVertex
:
3330 case SpvOpEmitStreamVertex
:
3331 intrinsic_op
= nir_intrinsic_emit_vertex
;
3333 case SpvOpEndPrimitive
:
3334 case SpvOpEndStreamPrimitive
:
3335 intrinsic_op
= nir_intrinsic_end_primitive
;
3338 unreachable("Invalid opcode");
3341 nir_intrinsic_instr
*intrin
=
3342 nir_intrinsic_instr_create(b
->shader
, intrinsic_op
);
3345 case SpvOpEmitStreamVertex
:
3346 case SpvOpEndStreamPrimitive
: {
3347 unsigned stream
= vtn_constant_uint(b
, w
[1]);
3348 nir_intrinsic_set_stream_id(intrin
, stream
);
3356 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3360 case SpvOpMemoryBarrier
: {
3361 SpvScope scope
= vtn_constant_uint(b
, w
[1]);
3362 SpvMemorySemanticsMask semantics
= vtn_constant_uint(b
, w
[2]);
3363 vtn_emit_memory_barrier(b
, scope
, semantics
);
3367 case SpvOpControlBarrier
: {
3368 SpvScope execution_scope
= vtn_constant_uint(b
, w
[1]);
3369 if (execution_scope
== SpvScopeWorkgroup
)
3370 vtn_emit_barrier(b
, nir_intrinsic_barrier
);
3372 SpvScope memory_scope
= vtn_constant_uint(b
, w
[2]);
3373 SpvMemorySemanticsMask memory_semantics
= vtn_constant_uint(b
, w
[3]);
3374 vtn_emit_memory_barrier(b
, memory_scope
, memory_semantics
);
3379 unreachable("unknown barrier instruction");
3384 gl_primitive_from_spv_execution_mode(struct vtn_builder
*b
,
3385 SpvExecutionMode mode
)
3388 case SpvExecutionModeInputPoints
:
3389 case SpvExecutionModeOutputPoints
:
3390 return 0; /* GL_POINTS */
3391 case SpvExecutionModeInputLines
:
3392 return 1; /* GL_LINES */
3393 case SpvExecutionModeInputLinesAdjacency
:
3394 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
3395 case SpvExecutionModeTriangles
:
3396 return 4; /* GL_TRIANGLES */
3397 case SpvExecutionModeInputTrianglesAdjacency
:
3398 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
3399 case SpvExecutionModeQuads
:
3400 return 7; /* GL_QUADS */
3401 case SpvExecutionModeIsolines
:
3402 return 0x8E7A; /* GL_ISOLINES */
3403 case SpvExecutionModeOutputLineStrip
:
3404 return 3; /* GL_LINE_STRIP */
3405 case SpvExecutionModeOutputTriangleStrip
:
3406 return 5; /* GL_TRIANGLE_STRIP */
3408 vtn_fail("Invalid primitive type: %s (%u)",
3409 spirv_executionmode_to_string(mode
), mode
);
3414 vertices_in_from_spv_execution_mode(struct vtn_builder
*b
,
3415 SpvExecutionMode mode
)
3418 case SpvExecutionModeInputPoints
:
3420 case SpvExecutionModeInputLines
:
3422 case SpvExecutionModeInputLinesAdjacency
:
3424 case SpvExecutionModeTriangles
:
3426 case SpvExecutionModeInputTrianglesAdjacency
:
3429 vtn_fail("Invalid GS input mode: %s (%u)",
3430 spirv_executionmode_to_string(mode
), mode
);
3434 static gl_shader_stage
3435 stage_for_execution_model(struct vtn_builder
*b
, SpvExecutionModel model
)
3438 case SpvExecutionModelVertex
:
3439 return MESA_SHADER_VERTEX
;
3440 case SpvExecutionModelTessellationControl
:
3441 return MESA_SHADER_TESS_CTRL
;
3442 case SpvExecutionModelTessellationEvaluation
:
3443 return MESA_SHADER_TESS_EVAL
;
3444 case SpvExecutionModelGeometry
:
3445 return MESA_SHADER_GEOMETRY
;
3446 case SpvExecutionModelFragment
:
3447 return MESA_SHADER_FRAGMENT
;
3448 case SpvExecutionModelGLCompute
:
3449 return MESA_SHADER_COMPUTE
;
3450 case SpvExecutionModelKernel
:
3451 return MESA_SHADER_KERNEL
;
3453 vtn_fail("Unsupported execution model: %s (%u)",
3454 spirv_executionmodel_to_string(model
), model
);
3458 #define spv_check_supported(name, cap) do { \
3459 if (!(b->options && b->options->caps.name)) \
3460 vtn_warn("Unsupported SPIR-V capability: %s (%u)", \
3461 spirv_capability_to_string(cap), cap); \
3466 vtn_handle_entry_point(struct vtn_builder
*b
, const uint32_t *w
,
3469 struct vtn_value
*entry_point
= &b
->values
[w
[2]];
3470 /* Let this be a name label regardless */
3471 unsigned name_words
;
3472 entry_point
->name
= vtn_string_literal(b
, &w
[3], count
- 3, &name_words
);
3474 if (strcmp(entry_point
->name
, b
->entry_point_name
) != 0 ||
3475 stage_for_execution_model(b
, w
[1]) != b
->entry_point_stage
)
3478 vtn_assert(b
->entry_point
== NULL
);
3479 b
->entry_point
= entry_point
;
3483 vtn_handle_preamble_instruction(struct vtn_builder
*b
, SpvOp opcode
,
3484 const uint32_t *w
, unsigned count
)
3491 case SpvSourceLanguageUnknown
: lang
= "unknown"; break;
3492 case SpvSourceLanguageESSL
: lang
= "ESSL"; break;
3493 case SpvSourceLanguageGLSL
: lang
= "GLSL"; break;
3494 case SpvSourceLanguageOpenCL_C
: lang
= "OpenCL C"; break;
3495 case SpvSourceLanguageOpenCL_CPP
: lang
= "OpenCL C++"; break;
3496 case SpvSourceLanguageHLSL
: lang
= "HLSL"; break;
3499 uint32_t version
= w
[2];
3502 (count
> 3) ? vtn_value(b
, w
[3], vtn_value_type_string
)->str
: "";
3504 vtn_info("Parsing SPIR-V from %s %u source file %s", lang
, version
, file
);
3508 case SpvOpSourceExtension
:
3509 case SpvOpSourceContinued
:
3510 case SpvOpExtension
:
3511 case SpvOpModuleProcessed
:
3512 /* Unhandled, but these are for debug so that's ok. */
3515 case SpvOpCapability
: {
3516 SpvCapability cap
= w
[1];
3518 case SpvCapabilityMatrix
:
3519 case SpvCapabilityShader
:
3520 case SpvCapabilityGeometry
:
3521 case SpvCapabilityGeometryPointSize
:
3522 case SpvCapabilityUniformBufferArrayDynamicIndexing
:
3523 case SpvCapabilitySampledImageArrayDynamicIndexing
:
3524 case SpvCapabilityStorageBufferArrayDynamicIndexing
:
3525 case SpvCapabilityStorageImageArrayDynamicIndexing
:
3526 case SpvCapabilityImageRect
:
3527 case SpvCapabilitySampledRect
:
3528 case SpvCapabilitySampled1D
:
3529 case SpvCapabilityImage1D
:
3530 case SpvCapabilitySampledCubeArray
:
3531 case SpvCapabilityImageCubeArray
:
3532 case SpvCapabilitySampledBuffer
:
3533 case SpvCapabilityImageBuffer
:
3534 case SpvCapabilityImageQuery
:
3535 case SpvCapabilityDerivativeControl
:
3536 case SpvCapabilityInterpolationFunction
:
3537 case SpvCapabilityMultiViewport
:
3538 case SpvCapabilitySampleRateShading
:
3539 case SpvCapabilityClipDistance
:
3540 case SpvCapabilityCullDistance
:
3541 case SpvCapabilityInputAttachment
:
3542 case SpvCapabilityImageGatherExtended
:
3543 case SpvCapabilityStorageImageExtendedFormats
:
3546 case SpvCapabilityLinkage
:
3547 case SpvCapabilityVector16
:
3548 case SpvCapabilityFloat16Buffer
:
3549 case SpvCapabilitySparseResidency
:
3550 vtn_warn("Unsupported SPIR-V capability: %s",
3551 spirv_capability_to_string(cap
));
3554 case SpvCapabilityMinLod
:
3555 spv_check_supported(min_lod
, cap
);
3558 case SpvCapabilityAtomicStorage
:
3559 spv_check_supported(atomic_storage
, cap
);
3562 case SpvCapabilityFloat64
:
3563 spv_check_supported(float64
, cap
);
3565 case SpvCapabilityInt64
:
3566 spv_check_supported(int64
, cap
);
3568 case SpvCapabilityInt16
:
3569 spv_check_supported(int16
, cap
);
3571 case SpvCapabilityInt8
:
3572 spv_check_supported(int8
, cap
);
3575 case SpvCapabilityTransformFeedback
:
3576 spv_check_supported(transform_feedback
, cap
);
3579 case SpvCapabilityGeometryStreams
:
3580 spv_check_supported(geometry_streams
, cap
);
3583 case SpvCapabilityInt64Atomics
:
3584 spv_check_supported(int64_atomics
, cap
);
3587 case SpvCapabilityStorageImageMultisample
:
3588 spv_check_supported(storage_image_ms
, cap
);
3591 case SpvCapabilityAddresses
:
3592 spv_check_supported(address
, cap
);
3595 case SpvCapabilityKernel
:
3596 spv_check_supported(kernel
, cap
);
3599 case SpvCapabilityImageBasic
:
3600 case SpvCapabilityImageReadWrite
:
3601 case SpvCapabilityImageMipmap
:
3602 case SpvCapabilityPipes
:
3603 case SpvCapabilityDeviceEnqueue
:
3604 case SpvCapabilityLiteralSampler
:
3605 case SpvCapabilityGenericPointer
:
3606 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
3607 spirv_capability_to_string(cap
));
3610 case SpvCapabilityImageMSArray
:
3611 spv_check_supported(image_ms_array
, cap
);
3614 case SpvCapabilityTessellation
:
3615 case SpvCapabilityTessellationPointSize
:
3616 spv_check_supported(tessellation
, cap
);
3619 case SpvCapabilityDrawParameters
:
3620 spv_check_supported(draw_parameters
, cap
);
3623 case SpvCapabilityStorageImageReadWithoutFormat
:
3624 spv_check_supported(image_read_without_format
, cap
);
3627 case SpvCapabilityStorageImageWriteWithoutFormat
:
3628 spv_check_supported(image_write_without_format
, cap
);
3631 case SpvCapabilityDeviceGroup
:
3632 spv_check_supported(device_group
, cap
);
3635 case SpvCapabilityMultiView
:
3636 spv_check_supported(multiview
, cap
);
3639 case SpvCapabilityGroupNonUniform
:
3640 spv_check_supported(subgroup_basic
, cap
);
3643 case SpvCapabilitySubgroupVoteKHR
:
3644 case SpvCapabilityGroupNonUniformVote
:
3645 spv_check_supported(subgroup_vote
, cap
);
3648 case SpvCapabilitySubgroupBallotKHR
:
3649 case SpvCapabilityGroupNonUniformBallot
:
3650 spv_check_supported(subgroup_ballot
, cap
);
3653 case SpvCapabilityGroupNonUniformShuffle
:
3654 case SpvCapabilityGroupNonUniformShuffleRelative
:
3655 spv_check_supported(subgroup_shuffle
, cap
);
3658 case SpvCapabilityGroupNonUniformQuad
:
3659 spv_check_supported(subgroup_quad
, cap
);
3662 case SpvCapabilityGroupNonUniformArithmetic
:
3663 case SpvCapabilityGroupNonUniformClustered
:
3664 spv_check_supported(subgroup_arithmetic
, cap
);
3667 case SpvCapabilityGroups
:
3668 spv_check_supported(amd_shader_ballot
, cap
);
3671 case SpvCapabilityVariablePointersStorageBuffer
:
3672 case SpvCapabilityVariablePointers
:
3673 spv_check_supported(variable_pointers
, cap
);
3674 b
->variable_pointers
= true;
3677 case SpvCapabilityStorageUniformBufferBlock16
:
3678 case SpvCapabilityStorageUniform16
:
3679 case SpvCapabilityStoragePushConstant16
:
3680 case SpvCapabilityStorageInputOutput16
:
3681 spv_check_supported(storage_16bit
, cap
);
3684 case SpvCapabilityShaderViewportIndexLayerEXT
:
3685 spv_check_supported(shader_viewport_index_layer
, cap
);
3688 case SpvCapabilityStorageBuffer8BitAccess
:
3689 case SpvCapabilityUniformAndStorageBuffer8BitAccess
:
3690 case SpvCapabilityStoragePushConstant8
:
3691 spv_check_supported(storage_8bit
, cap
);
3694 case SpvCapabilityShaderNonUniformEXT
:
3695 spv_check_supported(descriptor_indexing
, cap
);
3698 case SpvCapabilityInputAttachmentArrayDynamicIndexingEXT
:
3699 case SpvCapabilityUniformTexelBufferArrayDynamicIndexingEXT
:
3700 case SpvCapabilityStorageTexelBufferArrayDynamicIndexingEXT
:
3701 spv_check_supported(descriptor_array_dynamic_indexing
, cap
);
3704 case SpvCapabilityUniformBufferArrayNonUniformIndexingEXT
:
3705 case SpvCapabilitySampledImageArrayNonUniformIndexingEXT
:
3706 case SpvCapabilityStorageBufferArrayNonUniformIndexingEXT
:
3707 case SpvCapabilityStorageImageArrayNonUniformIndexingEXT
:
3708 case SpvCapabilityInputAttachmentArrayNonUniformIndexingEXT
:
3709 case SpvCapabilityUniformTexelBufferArrayNonUniformIndexingEXT
:
3710 case SpvCapabilityStorageTexelBufferArrayNonUniformIndexingEXT
:
3711 spv_check_supported(descriptor_array_non_uniform_indexing
, cap
);
3714 case SpvCapabilityRuntimeDescriptorArrayEXT
:
3715 spv_check_supported(runtime_descriptor_array
, cap
);
3718 case SpvCapabilityStencilExportEXT
:
3719 spv_check_supported(stencil_export
, cap
);
3722 case SpvCapabilitySampleMaskPostDepthCoverage
:
3723 spv_check_supported(post_depth_coverage
, cap
);
3726 case SpvCapabilityPhysicalStorageBufferAddressesEXT
:
3727 spv_check_supported(physical_storage_buffer_address
, cap
);
3730 case SpvCapabilityComputeDerivativeGroupQuadsNV
:
3731 case SpvCapabilityComputeDerivativeGroupLinearNV
:
3732 spv_check_supported(derivative_group
, cap
);
3735 case SpvCapabilityFloat16
:
3736 spv_check_supported(float16
, cap
);
3739 case SpvCapabilityFragmentShaderSampleInterlockEXT
:
3740 spv_check_supported(fragment_shader_sample_interlock
, cap
);
3743 case SpvCapabilityFragmentShaderPixelInterlockEXT
:
3744 spv_check_supported(fragment_shader_pixel_interlock
, cap
);
3748 vtn_fail("Unhandled capability: %s (%u)",
3749 spirv_capability_to_string(cap
), cap
);
3754 case SpvOpExtInstImport
:
3755 vtn_handle_extension(b
, opcode
, w
, count
);
3758 case SpvOpMemoryModel
:
3760 case SpvAddressingModelPhysical32
:
3761 vtn_fail_if(b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
,
3762 "AddressingModelPhysical32 only supported for kernels");
3763 b
->shader
->info
.cs
.ptr_size
= 32;
3764 b
->physical_ptrs
= true;
3765 b
->options
->shared_addr_format
= nir_address_format_32bit_global
;
3766 b
->options
->global_addr_format
= nir_address_format_32bit_global
;
3767 b
->options
->temp_addr_format
= nir_address_format_32bit_global
;
3769 case SpvAddressingModelPhysical64
:
3770 vtn_fail_if(b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
,
3771 "AddressingModelPhysical64 only supported for kernels");
3772 b
->shader
->info
.cs
.ptr_size
= 64;
3773 b
->physical_ptrs
= true;
3774 b
->options
->shared_addr_format
= nir_address_format_64bit_global
;
3775 b
->options
->global_addr_format
= nir_address_format_64bit_global
;
3776 b
->options
->temp_addr_format
= nir_address_format_64bit_global
;
3778 case SpvAddressingModelLogical
:
3779 vtn_fail_if(b
->shader
->info
.stage
>= MESA_SHADER_STAGES
,
3780 "AddressingModelLogical only supported for shaders");
3781 b
->shader
->info
.cs
.ptr_size
= 0;
3782 b
->physical_ptrs
= false;
3784 case SpvAddressingModelPhysicalStorageBuffer64EXT
:
3785 vtn_fail_if(!b
->options
||
3786 !b
->options
->caps
.physical_storage_buffer_address
,
3787 "AddressingModelPhysicalStorageBuffer64EXT not supported");
3790 vtn_fail("Unknown addressing model: %s (%u)",
3791 spirv_addressingmodel_to_string(w
[1]), w
[1]);
3795 vtn_assert(w
[2] == SpvMemoryModelSimple
||
3796 w
[2] == SpvMemoryModelGLSL450
||
3797 w
[2] == SpvMemoryModelOpenCL
);
3800 case SpvOpEntryPoint
:
3801 vtn_handle_entry_point(b
, w
, count
);
3805 vtn_push_value(b
, w
[1], vtn_value_type_string
)->str
=
3806 vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
3810 b
->values
[w
[1]].name
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
3813 case SpvOpMemberName
:
3817 case SpvOpExecutionMode
:
3818 case SpvOpExecutionModeId
:
3819 case SpvOpDecorationGroup
:
3821 case SpvOpDecorateId
:
3822 case SpvOpMemberDecorate
:
3823 case SpvOpGroupDecorate
:
3824 case SpvOpGroupMemberDecorate
:
3825 case SpvOpDecorateString
:
3826 case SpvOpMemberDecorateString
:
3827 vtn_handle_decoration(b
, opcode
, w
, count
);
3831 return false; /* End of preamble */
3838 vtn_handle_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
3839 const struct vtn_decoration
*mode
, void *data
)
3841 vtn_assert(b
->entry_point
== entry_point
);
3843 switch(mode
->exec_mode
) {
3844 case SpvExecutionModeOriginUpperLeft
:
3845 case SpvExecutionModeOriginLowerLeft
:
3846 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3847 b
->shader
->info
.fs
.origin_upper_left
=
3848 (mode
->exec_mode
== SpvExecutionModeOriginUpperLeft
);
3851 case SpvExecutionModeEarlyFragmentTests
:
3852 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3853 b
->shader
->info
.fs
.early_fragment_tests
= true;
3856 case SpvExecutionModePostDepthCoverage
:
3857 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3858 b
->shader
->info
.fs
.post_depth_coverage
= true;
3861 case SpvExecutionModeInvocations
:
3862 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3863 b
->shader
->info
.gs
.invocations
= MAX2(1, mode
->operands
[0]);
3866 case SpvExecutionModeDepthReplacing
:
3867 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3868 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_ANY
;
3870 case SpvExecutionModeDepthGreater
:
3871 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3872 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_GREATER
;
3874 case SpvExecutionModeDepthLess
:
3875 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3876 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_LESS
;
3878 case SpvExecutionModeDepthUnchanged
:
3879 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3880 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
3883 case SpvExecutionModeLocalSize
:
3884 vtn_assert(gl_shader_stage_is_compute(b
->shader
->info
.stage
));
3885 b
->shader
->info
.cs
.local_size
[0] = mode
->operands
[0];
3886 b
->shader
->info
.cs
.local_size
[1] = mode
->operands
[1];
3887 b
->shader
->info
.cs
.local_size
[2] = mode
->operands
[2];
3890 case SpvExecutionModeLocalSizeId
:
3891 b
->shader
->info
.cs
.local_size
[0] = vtn_constant_uint(b
, mode
->operands
[0]);
3892 b
->shader
->info
.cs
.local_size
[1] = vtn_constant_uint(b
, mode
->operands
[1]);
3893 b
->shader
->info
.cs
.local_size
[2] = vtn_constant_uint(b
, mode
->operands
[2]);
3896 case SpvExecutionModeLocalSizeHint
:
3897 case SpvExecutionModeLocalSizeHintId
:
3898 break; /* Nothing to do with this */
3900 case SpvExecutionModeOutputVertices
:
3901 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3902 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
3903 b
->shader
->info
.tess
.tcs_vertices_out
= mode
->operands
[0];
3905 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3906 b
->shader
->info
.gs
.vertices_out
= mode
->operands
[0];
3910 case SpvExecutionModeInputPoints
:
3911 case SpvExecutionModeInputLines
:
3912 case SpvExecutionModeInputLinesAdjacency
:
3913 case SpvExecutionModeTriangles
:
3914 case SpvExecutionModeInputTrianglesAdjacency
:
3915 case SpvExecutionModeQuads
:
3916 case SpvExecutionModeIsolines
:
3917 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3918 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
3919 b
->shader
->info
.tess
.primitive_mode
=
3920 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
3922 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3923 b
->shader
->info
.gs
.vertices_in
=
3924 vertices_in_from_spv_execution_mode(b
, mode
->exec_mode
);
3925 b
->shader
->info
.gs
.input_primitive
=
3926 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
3930 case SpvExecutionModeOutputPoints
:
3931 case SpvExecutionModeOutputLineStrip
:
3932 case SpvExecutionModeOutputTriangleStrip
:
3933 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
3934 b
->shader
->info
.gs
.output_primitive
=
3935 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
3938 case SpvExecutionModeSpacingEqual
:
3939 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3940 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3941 b
->shader
->info
.tess
.spacing
= TESS_SPACING_EQUAL
;
3943 case SpvExecutionModeSpacingFractionalEven
:
3944 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3945 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3946 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_EVEN
;
3948 case SpvExecutionModeSpacingFractionalOdd
:
3949 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3950 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3951 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_ODD
;
3953 case SpvExecutionModeVertexOrderCw
:
3954 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3955 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3956 b
->shader
->info
.tess
.ccw
= false;
3958 case SpvExecutionModeVertexOrderCcw
:
3959 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3960 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3961 b
->shader
->info
.tess
.ccw
= true;
3963 case SpvExecutionModePointMode
:
3964 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
3965 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
3966 b
->shader
->info
.tess
.point_mode
= true;
3969 case SpvExecutionModePixelCenterInteger
:
3970 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3971 b
->shader
->info
.fs
.pixel_center_integer
= true;
3974 case SpvExecutionModeXfb
:
3975 b
->shader
->info
.has_transform_feedback_varyings
= true;
3978 case SpvExecutionModeVecTypeHint
:
3981 case SpvExecutionModeContractionOff
:
3982 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
)
3983 vtn_warn("ExectionMode only allowed for CL-style kernels: %s",
3984 spirv_executionmode_to_string(mode
->exec_mode
));
3989 case SpvExecutionModeStencilRefReplacingEXT
:
3990 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
3993 case SpvExecutionModeDerivativeGroupQuadsNV
:
3994 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
3995 b
->shader
->info
.cs
.derivative_group
= DERIVATIVE_GROUP_QUADS
;
3998 case SpvExecutionModeDerivativeGroupLinearNV
:
3999 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
4000 b
->shader
->info
.cs
.derivative_group
= DERIVATIVE_GROUP_LINEAR
;
4003 case SpvExecutionModePixelInterlockOrderedEXT
:
4004 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4005 b
->shader
->info
.fs
.pixel_interlock_ordered
= true;
4008 case SpvExecutionModePixelInterlockUnorderedEXT
:
4009 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4010 b
->shader
->info
.fs
.pixel_interlock_unordered
= true;
4013 case SpvExecutionModeSampleInterlockOrderedEXT
:
4014 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4015 b
->shader
->info
.fs
.sample_interlock_ordered
= true;
4018 case SpvExecutionModeSampleInterlockUnorderedEXT
:
4019 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4020 b
->shader
->info
.fs
.sample_interlock_unordered
= true;
4024 vtn_fail("Unhandled execution mode: %s (%u)",
4025 spirv_executionmode_to_string(mode
->exec_mode
),
4031 vtn_handle_variable_or_type_instruction(struct vtn_builder
*b
, SpvOp opcode
,
4032 const uint32_t *w
, unsigned count
)
4034 vtn_set_instruction_result_type(b
, opcode
, w
, count
);
4038 case SpvOpSourceContinued
:
4039 case SpvOpSourceExtension
:
4040 case SpvOpExtension
:
4041 case SpvOpCapability
:
4042 case SpvOpExtInstImport
:
4043 case SpvOpMemoryModel
:
4044 case SpvOpEntryPoint
:
4045 case SpvOpExecutionMode
:
4048 case SpvOpMemberName
:
4049 case SpvOpDecorationGroup
:
4051 case SpvOpDecorateId
:
4052 case SpvOpMemberDecorate
:
4053 case SpvOpGroupDecorate
:
4054 case SpvOpGroupMemberDecorate
:
4055 case SpvOpDecorateString
:
4056 case SpvOpMemberDecorateString
:
4057 vtn_fail("Invalid opcode types and variables section");
4063 case SpvOpTypeFloat
:
4064 case SpvOpTypeVector
:
4065 case SpvOpTypeMatrix
:
4066 case SpvOpTypeImage
:
4067 case SpvOpTypeSampler
:
4068 case SpvOpTypeSampledImage
:
4069 case SpvOpTypeArray
:
4070 case SpvOpTypeRuntimeArray
:
4071 case SpvOpTypeStruct
:
4072 case SpvOpTypeOpaque
:
4073 case SpvOpTypePointer
:
4074 case SpvOpTypeForwardPointer
:
4075 case SpvOpTypeFunction
:
4076 case SpvOpTypeEvent
:
4077 case SpvOpTypeDeviceEvent
:
4078 case SpvOpTypeReserveId
:
4079 case SpvOpTypeQueue
:
4081 vtn_handle_type(b
, opcode
, w
, count
);
4084 case SpvOpConstantTrue
:
4085 case SpvOpConstantFalse
:
4087 case SpvOpConstantComposite
:
4088 case SpvOpConstantSampler
:
4089 case SpvOpConstantNull
:
4090 case SpvOpSpecConstantTrue
:
4091 case SpvOpSpecConstantFalse
:
4092 case SpvOpSpecConstant
:
4093 case SpvOpSpecConstantComposite
:
4094 case SpvOpSpecConstantOp
:
4095 vtn_handle_constant(b
, opcode
, w
, count
);
4100 vtn_handle_variables(b
, opcode
, w
, count
);
4104 return false; /* End of preamble */
4110 static struct vtn_ssa_value
*
4111 vtn_nir_select(struct vtn_builder
*b
, struct vtn_ssa_value
*src0
,
4112 struct vtn_ssa_value
*src1
, struct vtn_ssa_value
*src2
)
4114 struct vtn_ssa_value
*dest
= rzalloc(b
, struct vtn_ssa_value
);
4115 dest
->type
= src1
->type
;
4117 if (glsl_type_is_vector_or_scalar(src1
->type
)) {
4118 dest
->def
= nir_bcsel(&b
->nb
, src0
->def
, src1
->def
, src2
->def
);
4120 unsigned elems
= glsl_get_length(src1
->type
);
4122 dest
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
4123 for (unsigned i
= 0; i
< elems
; i
++) {
4124 dest
->elems
[i
] = vtn_nir_select(b
, src0
,
4125 src1
->elems
[i
], src2
->elems
[i
]);
4133 vtn_handle_select(struct vtn_builder
*b
, SpvOp opcode
,
4134 const uint32_t *w
, unsigned count
)
4136 /* Handle OpSelect up-front here because it needs to be able to handle
4137 * pointers and not just regular vectors and scalars.
4139 struct vtn_value
*res_val
= vtn_untyped_value(b
, w
[2]);
4140 struct vtn_value
*cond_val
= vtn_untyped_value(b
, w
[3]);
4141 struct vtn_value
*obj1_val
= vtn_untyped_value(b
, w
[4]);
4142 struct vtn_value
*obj2_val
= vtn_untyped_value(b
, w
[5]);
4144 vtn_fail_if(obj1_val
->type
!= res_val
->type
||
4145 obj2_val
->type
!= res_val
->type
,
4146 "Object types must match the result type in OpSelect");
4148 vtn_fail_if((cond_val
->type
->base_type
!= vtn_base_type_scalar
&&
4149 cond_val
->type
->base_type
!= vtn_base_type_vector
) ||
4150 !glsl_type_is_boolean(cond_val
->type
->type
),
4151 "OpSelect must have either a vector of booleans or "
4152 "a boolean as Condition type");
4154 vtn_fail_if(cond_val
->type
->base_type
== vtn_base_type_vector
&&
4155 (res_val
->type
->base_type
!= vtn_base_type_vector
||
4156 res_val
->type
->length
!= cond_val
->type
->length
),
4157 "When Condition type in OpSelect is a vector, the Result "
4158 "type must be a vector of the same length");
4160 switch (res_val
->type
->base_type
) {
4161 case vtn_base_type_scalar
:
4162 case vtn_base_type_vector
:
4163 case vtn_base_type_matrix
:
4164 case vtn_base_type_array
:
4165 case vtn_base_type_struct
:
4168 case vtn_base_type_pointer
:
4169 /* We need to have actual storage for pointer types. */
4170 vtn_fail_if(res_val
->type
->type
== NULL
,
4171 "Invalid pointer result type for OpSelect");
4174 vtn_fail("Result type of OpSelect must be a scalar, composite, or pointer");
4177 struct vtn_type
*res_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
4178 struct vtn_ssa_value
*ssa
= vtn_nir_select(b
,
4179 vtn_ssa_value(b
, w
[3]), vtn_ssa_value(b
, w
[4]), vtn_ssa_value(b
, w
[5]));
4181 vtn_push_ssa(b
, w
[2], res_type
, ssa
);
4185 vtn_handle_ptr(struct vtn_builder
*b
, SpvOp opcode
,
4186 const uint32_t *w
, unsigned count
)
4188 struct vtn_type
*type1
= vtn_untyped_value(b
, w
[3])->type
;
4189 struct vtn_type
*type2
= vtn_untyped_value(b
, w
[4])->type
;
4190 vtn_fail_if(type1
->base_type
!= vtn_base_type_pointer
||
4191 type2
->base_type
!= vtn_base_type_pointer
,
4192 "%s operands must have pointer types",
4193 spirv_op_to_string(opcode
));
4194 vtn_fail_if(type1
->storage_class
!= type2
->storage_class
,
4195 "%s operands must have the same storage class",
4196 spirv_op_to_string(opcode
));
4198 const struct glsl_type
*type
=
4199 vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
4201 nir_address_format addr_format
= vtn_mode_to_address_format(
4202 b
, vtn_storage_class_to_mode(b
, type1
->storage_class
, NULL
, NULL
));
4207 case SpvOpPtrDiff
: {
4208 /* OpPtrDiff returns the difference in number of elements (not byte offset). */
4209 unsigned elem_size
, elem_align
;
4210 glsl_get_natural_size_align_bytes(type1
->deref
->type
,
4211 &elem_size
, &elem_align
);
4213 def
= nir_build_addr_isub(&b
->nb
,
4214 vtn_ssa_value(b
, w
[3])->def
,
4215 vtn_ssa_value(b
, w
[4])->def
,
4217 def
= nir_idiv(&b
->nb
, def
, nir_imm_intN_t(&b
->nb
, elem_size
, def
->bit_size
));
4218 def
= nir_i2i(&b
->nb
, def
, glsl_get_bit_size(type
));
4223 case SpvOpPtrNotEqual
: {
4224 def
= nir_build_addr_ieq(&b
->nb
,
4225 vtn_ssa_value(b
, w
[3])->def
,
4226 vtn_ssa_value(b
, w
[4])->def
,
4228 if (opcode
== SpvOpPtrNotEqual
)
4229 def
= nir_inot(&b
->nb
, def
);
4234 unreachable("Invalid ptr operation");
4237 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
4238 val
->ssa
= vtn_create_ssa_value(b
, type
);
4239 val
->ssa
->def
= def
;
4243 vtn_handle_body_instruction(struct vtn_builder
*b
, SpvOp opcode
,
4244 const uint32_t *w
, unsigned count
)
4250 case SpvOpLoopMerge
:
4251 case SpvOpSelectionMerge
:
4252 /* This is handled by cfg pre-pass and walk_blocks */
4256 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
4257 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
4262 vtn_handle_extension(b
, opcode
, w
, count
);
4268 case SpvOpCopyMemory
:
4269 case SpvOpCopyMemorySized
:
4270 case SpvOpAccessChain
:
4271 case SpvOpPtrAccessChain
:
4272 case SpvOpInBoundsAccessChain
:
4273 case SpvOpInBoundsPtrAccessChain
:
4274 case SpvOpArrayLength
:
4275 case SpvOpConvertPtrToU
:
4276 case SpvOpConvertUToPtr
:
4277 vtn_handle_variables(b
, opcode
, w
, count
);
4280 case SpvOpFunctionCall
:
4281 vtn_handle_function_call(b
, opcode
, w
, count
);
4284 case SpvOpSampledImage
:
4286 case SpvOpImageSampleImplicitLod
:
4287 case SpvOpImageSampleExplicitLod
:
4288 case SpvOpImageSampleDrefImplicitLod
:
4289 case SpvOpImageSampleDrefExplicitLod
:
4290 case SpvOpImageSampleProjImplicitLod
:
4291 case SpvOpImageSampleProjExplicitLod
:
4292 case SpvOpImageSampleProjDrefImplicitLod
:
4293 case SpvOpImageSampleProjDrefExplicitLod
:
4294 case SpvOpImageFetch
:
4295 case SpvOpImageGather
:
4296 case SpvOpImageDrefGather
:
4297 case SpvOpImageQuerySizeLod
:
4298 case SpvOpImageQueryLod
:
4299 case SpvOpImageQueryLevels
:
4300 case SpvOpImageQuerySamples
:
4301 vtn_handle_texture(b
, opcode
, w
, count
);
4304 case SpvOpImageRead
:
4305 case SpvOpImageWrite
:
4306 case SpvOpImageTexelPointer
:
4307 vtn_handle_image(b
, opcode
, w
, count
);
4310 case SpvOpImageQuerySize
: {
4311 struct vtn_pointer
*image
=
4312 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
4313 if (glsl_type_is_image(image
->type
->type
)) {
4314 vtn_handle_image(b
, opcode
, w
, count
);
4316 vtn_assert(glsl_type_is_sampler(image
->type
->type
));
4317 vtn_handle_texture(b
, opcode
, w
, count
);
4322 case SpvOpAtomicLoad
:
4323 case SpvOpAtomicExchange
:
4324 case SpvOpAtomicCompareExchange
:
4325 case SpvOpAtomicCompareExchangeWeak
:
4326 case SpvOpAtomicIIncrement
:
4327 case SpvOpAtomicIDecrement
:
4328 case SpvOpAtomicIAdd
:
4329 case SpvOpAtomicISub
:
4330 case SpvOpAtomicSMin
:
4331 case SpvOpAtomicUMin
:
4332 case SpvOpAtomicSMax
:
4333 case SpvOpAtomicUMax
:
4334 case SpvOpAtomicAnd
:
4336 case SpvOpAtomicXor
: {
4337 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[3]);
4338 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
4339 vtn_handle_image(b
, opcode
, w
, count
);
4341 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
4342 vtn_handle_atomics(b
, opcode
, w
, count
);
4347 case SpvOpAtomicStore
: {
4348 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[1]);
4349 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
4350 vtn_handle_image(b
, opcode
, w
, count
);
4352 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
4353 vtn_handle_atomics(b
, opcode
, w
, count
);
4359 vtn_handle_select(b
, opcode
, w
, count
);
4367 case SpvOpConvertFToU
:
4368 case SpvOpConvertFToS
:
4369 case SpvOpConvertSToF
:
4370 case SpvOpConvertUToF
:
4374 case SpvOpQuantizeToF16
:
4375 case SpvOpPtrCastToGeneric
:
4376 case SpvOpGenericCastToPtr
:
4381 case SpvOpSignBitSet
:
4382 case SpvOpLessOrGreater
:
4384 case SpvOpUnordered
:
4399 case SpvOpVectorTimesScalar
:
4401 case SpvOpIAddCarry
:
4402 case SpvOpISubBorrow
:
4403 case SpvOpUMulExtended
:
4404 case SpvOpSMulExtended
:
4405 case SpvOpShiftRightLogical
:
4406 case SpvOpShiftRightArithmetic
:
4407 case SpvOpShiftLeftLogical
:
4408 case SpvOpLogicalEqual
:
4409 case SpvOpLogicalNotEqual
:
4410 case SpvOpLogicalOr
:
4411 case SpvOpLogicalAnd
:
4412 case SpvOpLogicalNot
:
4413 case SpvOpBitwiseOr
:
4414 case SpvOpBitwiseXor
:
4415 case SpvOpBitwiseAnd
:
4417 case SpvOpFOrdEqual
:
4418 case SpvOpFUnordEqual
:
4419 case SpvOpINotEqual
:
4420 case SpvOpFOrdNotEqual
:
4421 case SpvOpFUnordNotEqual
:
4422 case SpvOpULessThan
:
4423 case SpvOpSLessThan
:
4424 case SpvOpFOrdLessThan
:
4425 case SpvOpFUnordLessThan
:
4426 case SpvOpUGreaterThan
:
4427 case SpvOpSGreaterThan
:
4428 case SpvOpFOrdGreaterThan
:
4429 case SpvOpFUnordGreaterThan
:
4430 case SpvOpULessThanEqual
:
4431 case SpvOpSLessThanEqual
:
4432 case SpvOpFOrdLessThanEqual
:
4433 case SpvOpFUnordLessThanEqual
:
4434 case SpvOpUGreaterThanEqual
:
4435 case SpvOpSGreaterThanEqual
:
4436 case SpvOpFOrdGreaterThanEqual
:
4437 case SpvOpFUnordGreaterThanEqual
:
4443 case SpvOpFwidthFine
:
4444 case SpvOpDPdxCoarse
:
4445 case SpvOpDPdyCoarse
:
4446 case SpvOpFwidthCoarse
:
4447 case SpvOpBitFieldInsert
:
4448 case SpvOpBitFieldSExtract
:
4449 case SpvOpBitFieldUExtract
:
4450 case SpvOpBitReverse
:
4452 case SpvOpTranspose
:
4453 case SpvOpOuterProduct
:
4454 case SpvOpMatrixTimesScalar
:
4455 case SpvOpVectorTimesMatrix
:
4456 case SpvOpMatrixTimesVector
:
4457 case SpvOpMatrixTimesMatrix
:
4458 vtn_handle_alu(b
, opcode
, w
, count
);
4462 vtn_handle_bitcast(b
, w
, count
);
4465 case SpvOpVectorExtractDynamic
:
4466 case SpvOpVectorInsertDynamic
:
4467 case SpvOpVectorShuffle
:
4468 case SpvOpCompositeConstruct
:
4469 case SpvOpCompositeExtract
:
4470 case SpvOpCompositeInsert
:
4471 case SpvOpCopyLogical
:
4472 case SpvOpCopyObject
:
4473 vtn_handle_composite(b
, opcode
, w
, count
);
4476 case SpvOpEmitVertex
:
4477 case SpvOpEndPrimitive
:
4478 case SpvOpEmitStreamVertex
:
4479 case SpvOpEndStreamPrimitive
:
4480 case SpvOpControlBarrier
:
4481 case SpvOpMemoryBarrier
:
4482 vtn_handle_barrier(b
, opcode
, w
, count
);
4485 case SpvOpGroupNonUniformElect
:
4486 case SpvOpGroupNonUniformAll
:
4487 case SpvOpGroupNonUniformAny
:
4488 case SpvOpGroupNonUniformAllEqual
:
4489 case SpvOpGroupNonUniformBroadcast
:
4490 case SpvOpGroupNonUniformBroadcastFirst
:
4491 case SpvOpGroupNonUniformBallot
:
4492 case SpvOpGroupNonUniformInverseBallot
:
4493 case SpvOpGroupNonUniformBallotBitExtract
:
4494 case SpvOpGroupNonUniformBallotBitCount
:
4495 case SpvOpGroupNonUniformBallotFindLSB
:
4496 case SpvOpGroupNonUniformBallotFindMSB
:
4497 case SpvOpGroupNonUniformShuffle
:
4498 case SpvOpGroupNonUniformShuffleXor
:
4499 case SpvOpGroupNonUniformShuffleUp
:
4500 case SpvOpGroupNonUniformShuffleDown
:
4501 case SpvOpGroupNonUniformIAdd
:
4502 case SpvOpGroupNonUniformFAdd
:
4503 case SpvOpGroupNonUniformIMul
:
4504 case SpvOpGroupNonUniformFMul
:
4505 case SpvOpGroupNonUniformSMin
:
4506 case SpvOpGroupNonUniformUMin
:
4507 case SpvOpGroupNonUniformFMin
:
4508 case SpvOpGroupNonUniformSMax
:
4509 case SpvOpGroupNonUniformUMax
:
4510 case SpvOpGroupNonUniformFMax
:
4511 case SpvOpGroupNonUniformBitwiseAnd
:
4512 case SpvOpGroupNonUniformBitwiseOr
:
4513 case SpvOpGroupNonUniformBitwiseXor
:
4514 case SpvOpGroupNonUniformLogicalAnd
:
4515 case SpvOpGroupNonUniformLogicalOr
:
4516 case SpvOpGroupNonUniformLogicalXor
:
4517 case SpvOpGroupNonUniformQuadBroadcast
:
4518 case SpvOpGroupNonUniformQuadSwap
:
4521 case SpvOpGroupBroadcast
:
4522 case SpvOpGroupIAdd
:
4523 case SpvOpGroupFAdd
:
4524 case SpvOpGroupFMin
:
4525 case SpvOpGroupUMin
:
4526 case SpvOpGroupSMin
:
4527 case SpvOpGroupFMax
:
4528 case SpvOpGroupUMax
:
4529 case SpvOpGroupSMax
:
4530 case SpvOpSubgroupBallotKHR
:
4531 case SpvOpSubgroupFirstInvocationKHR
:
4532 case SpvOpSubgroupReadInvocationKHR
:
4533 case SpvOpSubgroupAllKHR
:
4534 case SpvOpSubgroupAnyKHR
:
4535 case SpvOpSubgroupAllEqualKHR
:
4536 case SpvOpGroupIAddNonUniformAMD
:
4537 case SpvOpGroupFAddNonUniformAMD
:
4538 case SpvOpGroupFMinNonUniformAMD
:
4539 case SpvOpGroupUMinNonUniformAMD
:
4540 case SpvOpGroupSMinNonUniformAMD
:
4541 case SpvOpGroupFMaxNonUniformAMD
:
4542 case SpvOpGroupUMaxNonUniformAMD
:
4543 case SpvOpGroupSMaxNonUniformAMD
:
4544 vtn_handle_subgroup(b
, opcode
, w
, count
);
4549 case SpvOpPtrNotEqual
:
4550 vtn_handle_ptr(b
, opcode
, w
, count
);
4553 case SpvOpBeginInvocationInterlockEXT
:
4554 vtn_emit_barrier(b
, nir_intrinsic_begin_invocation_interlock
);
4557 case SpvOpEndInvocationInterlockEXT
:
4558 vtn_emit_barrier(b
, nir_intrinsic_end_invocation_interlock
);
4562 vtn_fail_with_opcode("Unhandled opcode", opcode
);
4569 vtn_create_builder(const uint32_t *words
, size_t word_count
,
4570 gl_shader_stage stage
, const char *entry_point_name
,
4571 const struct spirv_to_nir_options
*options
)
4573 /* Initialize the vtn_builder object */
4574 struct vtn_builder
*b
= rzalloc(NULL
, struct vtn_builder
);
4575 struct spirv_to_nir_options
*dup_options
=
4576 ralloc(b
, struct spirv_to_nir_options
);
4577 *dup_options
= *options
;
4580 b
->spirv_word_count
= word_count
;
4584 exec_list_make_empty(&b
->functions
);
4585 b
->entry_point_stage
= stage
;
4586 b
->entry_point_name
= entry_point_name
;
4587 b
->options
= dup_options
;
4590 * Handle the SPIR-V header (first 5 dwords).
4591 * Can't use vtx_assert() as the setjmp(3) target isn't initialized yet.
4593 if (word_count
<= 5)
4596 if (words
[0] != SpvMagicNumber
) {
4597 vtn_err("words[0] was 0x%x, want 0x%x", words
[0], SpvMagicNumber
);
4600 if (words
[1] < 0x10000) {
4601 vtn_err("words[1] was 0x%x, want >= 0x10000", words
[1]);
4605 uint16_t generator_id
= words
[2] >> 16;
4606 uint16_t generator_version
= words
[2];
4608 /* The first GLSLang version bump actually 1.5 years after #179 was fixed
4609 * but this should at least let us shut the workaround off for modern
4610 * versions of GLSLang.
4612 b
->wa_glslang_179
= (generator_id
== 8 && generator_version
== 1);
4614 /* words[2] == generator magic */
4615 unsigned value_id_bound
= words
[3];
4616 if (words
[4] != 0) {
4617 vtn_err("words[4] was %u, want 0", words
[4]);
4621 b
->value_id_bound
= value_id_bound
;
4622 b
->values
= rzalloc_array(b
, struct vtn_value
, value_id_bound
);
4630 static nir_function
*
4631 vtn_emit_kernel_entry_point_wrapper(struct vtn_builder
*b
,
4632 nir_function
*entry_point
)
4634 vtn_assert(entry_point
== b
->entry_point
->func
->impl
->function
);
4635 vtn_fail_if(!entry_point
->name
, "entry points are required to have a name");
4636 const char *func_name
=
4637 ralloc_asprintf(b
->shader
, "__wrapped_%s", entry_point
->name
);
4639 /* we shouldn't have any inputs yet */
4640 vtn_assert(!entry_point
->shader
->num_inputs
);
4641 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_KERNEL
);
4643 nir_function
*main_entry_point
= nir_function_create(b
->shader
, func_name
);
4644 main_entry_point
->impl
= nir_function_impl_create(main_entry_point
);
4645 nir_builder_init(&b
->nb
, main_entry_point
->impl
);
4646 b
->nb
.cursor
= nir_after_cf_list(&main_entry_point
->impl
->body
);
4647 b
->func_param_idx
= 0;
4649 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, entry_point
);
4651 for (unsigned i
= 0; i
< entry_point
->num_params
; ++i
) {
4652 struct vtn_type
*param_type
= b
->entry_point
->func
->type
->params
[i
];
4654 /* consider all pointers to function memory to be parameters passed
4657 bool is_by_val
= param_type
->base_type
== vtn_base_type_pointer
&&
4658 param_type
->storage_class
== SpvStorageClassFunction
;
4660 /* input variable */
4661 nir_variable
*in_var
= rzalloc(b
->nb
.shader
, nir_variable
);
4662 in_var
->data
.mode
= nir_var_shader_in
;
4663 in_var
->data
.read_only
= true;
4664 in_var
->data
.location
= i
;
4667 in_var
->type
= param_type
->deref
->type
;
4669 in_var
->type
= param_type
->type
;
4671 nir_shader_add_variable(b
->nb
.shader
, in_var
);
4672 b
->nb
.shader
->num_inputs
++;
4674 /* we have to copy the entire variable into function memory */
4676 nir_variable
*copy_var
=
4677 nir_local_variable_create(main_entry_point
->impl
, in_var
->type
,
4679 nir_copy_var(&b
->nb
, copy_var
, in_var
);
4681 nir_src_for_ssa(&nir_build_deref_var(&b
->nb
, copy_var
)->dest
.ssa
);
4683 call
->params
[i
] = nir_src_for_ssa(nir_load_var(&b
->nb
, in_var
));
4687 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
4689 return main_entry_point
;
4693 spirv_to_nir(const uint32_t *words
, size_t word_count
,
4694 struct nir_spirv_specialization
*spec
, unsigned num_spec
,
4695 gl_shader_stage stage
, const char *entry_point_name
,
4696 const struct spirv_to_nir_options
*options
,
4697 const nir_shader_compiler_options
*nir_options
)
4700 const uint32_t *word_end
= words
+ word_count
;
4702 struct vtn_builder
*b
= vtn_create_builder(words
, word_count
,
4703 stage
, entry_point_name
,
4709 /* See also _vtn_fail() */
4710 if (setjmp(b
->fail_jump
)) {
4715 /* Skip the SPIR-V header, handled at vtn_create_builder */
4718 b
->shader
= nir_shader_create(b
, stage
, nir_options
, NULL
);
4720 /* Handle all the preamble instructions */
4721 words
= vtn_foreach_instruction(b
, words
, word_end
,
4722 vtn_handle_preamble_instruction
);
4724 if (b
->entry_point
== NULL
) {
4725 vtn_fail("Entry point not found");
4730 /* Set shader info defaults */
4731 b
->shader
->info
.gs
.invocations
= 1;
4733 b
->specializations
= spec
;
4734 b
->num_specializations
= num_spec
;
4736 /* Handle all variable, type, and constant instructions */
4737 words
= vtn_foreach_instruction(b
, words
, word_end
,
4738 vtn_handle_variable_or_type_instruction
);
4740 /* Parse execution modes */
4741 vtn_foreach_execution_mode(b
, b
->entry_point
,
4742 vtn_handle_execution_mode
, NULL
);
4744 if (b
->workgroup_size_builtin
) {
4745 vtn_assert(b
->workgroup_size_builtin
->type
->type
==
4746 glsl_vector_type(GLSL_TYPE_UINT
, 3));
4748 nir_const_value
*const_size
=
4749 b
->workgroup_size_builtin
->constant
->values
[0];
4751 b
->shader
->info
.cs
.local_size
[0] = const_size
[0].u32
;
4752 b
->shader
->info
.cs
.local_size
[1] = const_size
[1].u32
;
4753 b
->shader
->info
.cs
.local_size
[2] = const_size
[2].u32
;
4756 /* Set types on all vtn_values */
4757 vtn_foreach_instruction(b
, words
, word_end
, vtn_set_instruction_result_type
);
4759 vtn_build_cfg(b
, words
, word_end
);
4761 assert(b
->entry_point
->value_type
== vtn_value_type_function
);
4762 b
->entry_point
->func
->referenced
= true;
4767 foreach_list_typed(struct vtn_function
, func
, node
, &b
->functions
) {
4768 if (func
->referenced
&& !func
->emitted
) {
4769 b
->const_table
= _mesa_pointer_hash_table_create(b
);
4771 vtn_function_emit(b
, func
, vtn_handle_body_instruction
);
4777 vtn_assert(b
->entry_point
->value_type
== vtn_value_type_function
);
4778 nir_function
*entry_point
= b
->entry_point
->func
->impl
->function
;
4779 vtn_assert(entry_point
);
4781 /* post process entry_points with input params */
4782 if (entry_point
->num_params
&& b
->shader
->info
.stage
== MESA_SHADER_KERNEL
)
4783 entry_point
= vtn_emit_kernel_entry_point_wrapper(b
, entry_point
);
4785 entry_point
->is_entrypoint
= true;
4787 /* When multiple shader stages exist in the same SPIR-V module, we
4788 * generate input and output variables for every stage, in the same
4789 * NIR program. These dead variables can be invalid NIR. For example,
4790 * TCS outputs must be per-vertex arrays (or decorated 'patch'), while
4791 * VS output variables wouldn't be.
4793 * To ensure we have valid NIR, we eliminate any dead inputs and outputs
4794 * right away. In order to do so, we must lower any constant initializers
4795 * on outputs so nir_remove_dead_variables sees that they're written to.
4797 nir_lower_constant_initializers(b
->shader
, nir_var_shader_out
);
4798 nir_remove_dead_variables(b
->shader
,
4799 nir_var_shader_in
| nir_var_shader_out
);
4801 /* We sometimes generate bogus derefs that, while never used, give the
4802 * validator a bit of heartburn. Run dead code to get rid of them.
4804 nir_opt_dce(b
->shader
);
4806 /* Unparent the shader from the vtn_builder before we delete the builder */
4807 ralloc_steal(NULL
, b
->shader
);
4809 nir_shader
*shader
= b
->shader
;