2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Jason Ekstrand (jason@jlekstrand.net)
28 #include "vtn_private.h"
29 #include "nir/nir_vla.h"
30 #include "nir/nir_control_flow.h"
31 #include "nir/nir_constant_expressions.h"
32 #include "nir/nir_deref.h"
33 #include "spirv_info.h"
35 #include "util/format/u_format.h"
36 #include "util/u_math.h"
41 vtn_log(struct vtn_builder
*b
, enum nir_spirv_debug_level level
,
42 size_t spirv_offset
, const char *message
)
44 if (b
->options
->debug
.func
) {
45 b
->options
->debug
.func(b
->options
->debug
.private_data
,
46 level
, spirv_offset
, message
);
50 if (level
>= NIR_SPIRV_DEBUG_LEVEL_WARNING
)
51 fprintf(stderr
, "%s\n", message
);
56 vtn_logf(struct vtn_builder
*b
, enum nir_spirv_debug_level level
,
57 size_t spirv_offset
, const char *fmt
, ...)
63 msg
= ralloc_vasprintf(NULL
, fmt
, args
);
66 vtn_log(b
, level
, spirv_offset
, msg
);
72 vtn_log_err(struct vtn_builder
*b
,
73 enum nir_spirv_debug_level level
, const char *prefix
,
74 const char *file
, unsigned line
,
75 const char *fmt
, va_list args
)
79 msg
= ralloc_strdup(NULL
, prefix
);
82 ralloc_asprintf_append(&msg
, " In file %s:%u\n", file
, line
);
85 ralloc_asprintf_append(&msg
, " ");
87 ralloc_vasprintf_append(&msg
, fmt
, args
);
89 ralloc_asprintf_append(&msg
, "\n %zu bytes into the SPIR-V binary",
93 ralloc_asprintf_append(&msg
,
94 "\n in SPIR-V source file %s, line %d, col %d",
95 b
->file
, b
->line
, b
->col
);
98 vtn_log(b
, level
, b
->spirv_offset
, msg
);
104 vtn_dump_shader(struct vtn_builder
*b
, const char *path
, const char *prefix
)
109 int len
= snprintf(filename
, sizeof(filename
), "%s/%s-%d.spirv",
110 path
, prefix
, idx
++);
111 if (len
< 0 || len
>= sizeof(filename
))
114 FILE *f
= fopen(filename
, "w");
118 fwrite(b
->spirv
, sizeof(*b
->spirv
), b
->spirv_word_count
, f
);
121 vtn_info("SPIR-V shader dumped to %s", filename
);
125 _vtn_warn(struct vtn_builder
*b
, const char *file
, unsigned line
,
126 const char *fmt
, ...)
131 vtn_log_err(b
, NIR_SPIRV_DEBUG_LEVEL_WARNING
, "SPIR-V WARNING:\n",
132 file
, line
, fmt
, args
);
137 _vtn_err(struct vtn_builder
*b
, const char *file
, unsigned line
,
138 const char *fmt
, ...)
143 vtn_log_err(b
, NIR_SPIRV_DEBUG_LEVEL_ERROR
, "SPIR-V ERROR:\n",
144 file
, line
, fmt
, args
);
149 _vtn_fail(struct vtn_builder
*b
, const char *file
, unsigned line
,
150 const char *fmt
, ...)
155 vtn_log_err(b
, NIR_SPIRV_DEBUG_LEVEL_ERROR
, "SPIR-V parsing FAILED:\n",
156 file
, line
, fmt
, args
);
159 const char *dump_path
= getenv("MESA_SPIRV_FAIL_DUMP_PATH");
161 vtn_dump_shader(b
, dump_path
, "fail");
163 longjmp(b
->fail_jump
, 1);
166 static struct vtn_ssa_value
*
167 vtn_undef_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
169 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
170 val
->type
= glsl_get_bare_type(type
);
172 if (glsl_type_is_vector_or_scalar(type
)) {
173 unsigned num_components
= glsl_get_vector_elements(val
->type
);
174 unsigned bit_size
= glsl_get_bit_size(val
->type
);
175 val
->def
= nir_ssa_undef(&b
->nb
, num_components
, bit_size
);
177 unsigned elems
= glsl_get_length(val
->type
);
178 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
179 if (glsl_type_is_array_or_matrix(type
)) {
180 const struct glsl_type
*elem_type
= glsl_get_array_element(type
);
181 for (unsigned i
= 0; i
< elems
; i
++)
182 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
184 vtn_assert(glsl_type_is_struct_or_ifc(type
));
185 for (unsigned i
= 0; i
< elems
; i
++) {
186 const struct glsl_type
*elem_type
= glsl_get_struct_field(type
, i
);
187 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
195 static struct vtn_ssa_value
*
196 vtn_const_ssa_value(struct vtn_builder
*b
, nir_constant
*constant
,
197 const struct glsl_type
*type
)
199 struct hash_entry
*entry
= _mesa_hash_table_search(b
->const_table
, constant
);
204 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
205 val
->type
= glsl_get_bare_type(type
);
207 if (glsl_type_is_vector_or_scalar(type
)) {
208 unsigned num_components
= glsl_get_vector_elements(val
->type
);
209 unsigned bit_size
= glsl_get_bit_size(type
);
210 nir_load_const_instr
*load
=
211 nir_load_const_instr_create(b
->shader
, num_components
, bit_size
);
213 memcpy(load
->value
, constant
->values
,
214 sizeof(nir_const_value
) * num_components
);
216 nir_instr_insert_before_cf_list(&b
->nb
.impl
->body
, &load
->instr
);
217 val
->def
= &load
->def
;
219 unsigned elems
= glsl_get_length(val
->type
);
220 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
221 if (glsl_type_is_array_or_matrix(type
)) {
222 const struct glsl_type
*elem_type
= glsl_get_array_element(type
);
223 for (unsigned i
= 0; i
< elems
; i
++) {
224 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
228 vtn_assert(glsl_type_is_struct_or_ifc(type
));
229 for (unsigned i
= 0; i
< elems
; i
++) {
230 const struct glsl_type
*elem_type
= glsl_get_struct_field(type
, i
);
231 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
240 struct vtn_ssa_value
*
241 vtn_ssa_value(struct vtn_builder
*b
, uint32_t value_id
)
243 struct vtn_value
*val
= vtn_untyped_value(b
, value_id
);
244 switch (val
->value_type
) {
245 case vtn_value_type_undef
:
246 return vtn_undef_ssa_value(b
, val
->type
->type
);
248 case vtn_value_type_constant
:
249 return vtn_const_ssa_value(b
, val
->constant
, val
->type
->type
);
251 case vtn_value_type_ssa
:
254 case vtn_value_type_pointer
:
255 vtn_assert(val
->pointer
->ptr_type
&& val
->pointer
->ptr_type
->type
);
256 struct vtn_ssa_value
*ssa
=
257 vtn_create_ssa_value(b
, val
->pointer
->ptr_type
->type
);
258 ssa
->def
= vtn_pointer_to_ssa(b
, val
->pointer
);
262 vtn_fail("Invalid type for an SSA value");
267 vtn_push_ssa_value(struct vtn_builder
*b
, uint32_t value_id
,
268 struct vtn_ssa_value
*ssa
)
270 struct vtn_type
*type
= vtn_get_value_type(b
, value_id
);
272 /* See vtn_create_ssa_value */
273 vtn_fail_if(ssa
->type
!= glsl_get_bare_type(type
->type
),
274 "Type mismatch for SPIR-V SSA value");
276 struct vtn_value
*val
;
277 if (type
->base_type
== vtn_base_type_pointer
) {
278 val
= vtn_push_pointer(b
, value_id
, vtn_pointer_from_ssa(b
, ssa
->def
, type
));
280 /* Don't trip the value_type_ssa check in vtn_push_value */
281 val
= vtn_push_value(b
, value_id
, vtn_value_type_invalid
);
282 val
->value_type
= vtn_value_type_ssa
;
290 vtn_get_nir_ssa(struct vtn_builder
*b
, uint32_t value_id
)
292 struct vtn_ssa_value
*ssa
= vtn_ssa_value(b
, value_id
);
293 vtn_fail_if(!glsl_type_is_vector_or_scalar(ssa
->type
),
294 "Expected a vector or scalar type");
299 vtn_push_nir_ssa(struct vtn_builder
*b
, uint32_t value_id
, nir_ssa_def
*def
)
301 /* Types for all SPIR-V SSA values are set as part of a pre-pass so the
302 * type will be valid by the time we get here.
304 struct vtn_type
*type
= vtn_get_value_type(b
, value_id
);
305 vtn_fail_if(def
->num_components
!= glsl_get_vector_elements(type
->type
) ||
306 def
->bit_size
!= glsl_get_bit_size(type
->type
),
307 "Mismatch between NIR and SPIR-V type.");
308 struct vtn_ssa_value
*ssa
= vtn_create_ssa_value(b
, type
->type
);
310 return vtn_push_ssa_value(b
, value_id
, ssa
);
313 static nir_deref_instr
*
314 vtn_get_image(struct vtn_builder
*b
, uint32_t value_id
)
316 struct vtn_type
*type
= vtn_get_value_type(b
, value_id
);
317 vtn_assert(type
->base_type
== vtn_base_type_image
);
318 return nir_build_deref_cast(&b
->nb
, vtn_get_nir_ssa(b
, value_id
),
319 nir_var_uniform
, type
->glsl_image
, 0);
323 vtn_push_image(struct vtn_builder
*b
, uint32_t value_id
,
324 nir_deref_instr
*deref
)
326 struct vtn_type
*type
= vtn_get_value_type(b
, value_id
);
327 vtn_assert(type
->base_type
== vtn_base_type_image
);
328 vtn_push_nir_ssa(b
, value_id
, &deref
->dest
.ssa
);
331 static nir_deref_instr
*
332 vtn_get_sampler(struct vtn_builder
*b
, uint32_t value_id
)
334 struct vtn_type
*type
= vtn_get_value_type(b
, value_id
);
335 vtn_assert(type
->base_type
== vtn_base_type_sampler
);
336 return nir_build_deref_cast(&b
->nb
, vtn_get_nir_ssa(b
, value_id
),
337 nir_var_uniform
, glsl_bare_sampler_type(), 0);
341 vtn_sampled_image_to_nir_ssa(struct vtn_builder
*b
,
342 struct vtn_sampled_image si
)
344 return nir_vec2(&b
->nb
, &si
.image
->dest
.ssa
, &si
.sampler
->dest
.ssa
);
348 vtn_push_sampled_image(struct vtn_builder
*b
, uint32_t value_id
,
349 struct vtn_sampled_image si
)
351 struct vtn_type
*type
= vtn_get_value_type(b
, value_id
);
352 vtn_assert(type
->base_type
== vtn_base_type_sampled_image
);
353 vtn_push_nir_ssa(b
, value_id
, vtn_sampled_image_to_nir_ssa(b
, si
));
356 static struct vtn_sampled_image
357 vtn_get_sampled_image(struct vtn_builder
*b
, uint32_t value_id
)
359 struct vtn_type
*type
= vtn_get_value_type(b
, value_id
);
360 vtn_assert(type
->base_type
== vtn_base_type_sampled_image
);
361 nir_ssa_def
*si_vec2
= vtn_get_nir_ssa(b
, value_id
);
363 struct vtn_sampled_image si
= { NULL
, };
364 si
.image
= nir_build_deref_cast(&b
->nb
, nir_channel(&b
->nb
, si_vec2
, 0),
366 type
->image
->glsl_image
, 0);
367 si
.sampler
= nir_build_deref_cast(&b
->nb
, nir_channel(&b
->nb
, si_vec2
, 1),
369 glsl_bare_sampler_type(), 0);
374 vtn_string_literal(struct vtn_builder
*b
, const uint32_t *words
,
375 unsigned word_count
, unsigned *words_used
)
377 char *dup
= ralloc_strndup(b
, (char *)words
, word_count
* sizeof(*words
));
379 /* Ammount of space taken by the string (including the null) */
380 unsigned len
= strlen(dup
) + 1;
381 *words_used
= DIV_ROUND_UP(len
, sizeof(*words
));
387 vtn_foreach_instruction(struct vtn_builder
*b
, const uint32_t *start
,
388 const uint32_t *end
, vtn_instruction_handler handler
)
394 const uint32_t *w
= start
;
396 SpvOp opcode
= w
[0] & SpvOpCodeMask
;
397 unsigned count
= w
[0] >> SpvWordCountShift
;
398 vtn_assert(count
>= 1 && w
+ count
<= end
);
400 b
->spirv_offset
= (uint8_t *)w
- (uint8_t *)b
->spirv
;
404 break; /* Do nothing */
407 b
->file
= vtn_value(b
, w
[1], vtn_value_type_string
)->str
;
419 if (!handler(b
, opcode
, w
, count
))
437 vtn_handle_non_semantic_instruction(struct vtn_builder
*b
, SpvOp ext_opcode
,
438 const uint32_t *w
, unsigned count
)
445 vtn_handle_extension(struct vtn_builder
*b
, SpvOp opcode
,
446 const uint32_t *w
, unsigned count
)
448 const char *ext
= (const char *)&w
[2];
450 case SpvOpExtInstImport
: {
451 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_extension
);
452 if (strcmp(ext
, "GLSL.std.450") == 0) {
453 val
->ext_handler
= vtn_handle_glsl450_instruction
;
454 } else if ((strcmp(ext
, "SPV_AMD_gcn_shader") == 0)
455 && (b
->options
&& b
->options
->caps
.amd_gcn_shader
)) {
456 val
->ext_handler
= vtn_handle_amd_gcn_shader_instruction
;
457 } else if ((strcmp(ext
, "SPV_AMD_shader_ballot") == 0)
458 && (b
->options
&& b
->options
->caps
.amd_shader_ballot
)) {
459 val
->ext_handler
= vtn_handle_amd_shader_ballot_instruction
;
460 } else if ((strcmp(ext
, "SPV_AMD_shader_trinary_minmax") == 0)
461 && (b
->options
&& b
->options
->caps
.amd_trinary_minmax
)) {
462 val
->ext_handler
= vtn_handle_amd_shader_trinary_minmax_instruction
;
463 } else if ((strcmp(ext
, "SPV_AMD_shader_explicit_vertex_parameter") == 0)
464 && (b
->options
&& b
->options
->caps
.amd_shader_explicit_vertex_parameter
)) {
465 val
->ext_handler
= vtn_handle_amd_shader_explicit_vertex_parameter_instruction
;
466 } else if (strcmp(ext
, "OpenCL.std") == 0) {
467 val
->ext_handler
= vtn_handle_opencl_instruction
;
468 } else if (strstr(ext
, "NonSemantic.") == ext
) {
469 val
->ext_handler
= vtn_handle_non_semantic_instruction
;
471 vtn_fail("Unsupported extension: %s", ext
);
477 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
478 bool handled
= val
->ext_handler(b
, w
[4], w
, count
);
484 vtn_fail_with_opcode("Unhandled opcode", opcode
);
489 _foreach_decoration_helper(struct vtn_builder
*b
,
490 struct vtn_value
*base_value
,
492 struct vtn_value
*value
,
493 vtn_decoration_foreach_cb cb
, void *data
)
495 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
497 if (dec
->scope
== VTN_DEC_DECORATION
) {
498 member
= parent_member
;
499 } else if (dec
->scope
>= VTN_DEC_STRUCT_MEMBER0
) {
500 vtn_fail_if(value
->value_type
!= vtn_value_type_type
||
501 value
->type
->base_type
!= vtn_base_type_struct
,
502 "OpMemberDecorate and OpGroupMemberDecorate are only "
503 "allowed on OpTypeStruct");
504 /* This means we haven't recursed yet */
505 assert(value
== base_value
);
507 member
= dec
->scope
- VTN_DEC_STRUCT_MEMBER0
;
509 vtn_fail_if(member
>= base_value
->type
->length
,
510 "OpMemberDecorate specifies member %d but the "
511 "OpTypeStruct has only %u members",
512 member
, base_value
->type
->length
);
514 /* Not a decoration */
515 assert(dec
->scope
== VTN_DEC_EXECUTION_MODE
);
520 assert(dec
->group
->value_type
== vtn_value_type_decoration_group
);
521 _foreach_decoration_helper(b
, base_value
, member
, dec
->group
,
524 cb(b
, base_value
, member
, dec
, data
);
529 /** Iterates (recursively if needed) over all of the decorations on a value
531 * This function iterates over all of the decorations applied to a given
532 * value. If it encounters a decoration group, it recurses into the group
533 * and iterates over all of those decorations as well.
536 vtn_foreach_decoration(struct vtn_builder
*b
, struct vtn_value
*value
,
537 vtn_decoration_foreach_cb cb
, void *data
)
539 _foreach_decoration_helper(b
, value
, -1, value
, cb
, data
);
543 vtn_foreach_execution_mode(struct vtn_builder
*b
, struct vtn_value
*value
,
544 vtn_execution_mode_foreach_cb cb
, void *data
)
546 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
547 if (dec
->scope
!= VTN_DEC_EXECUTION_MODE
)
550 assert(dec
->group
== NULL
);
551 cb(b
, value
, dec
, data
);
556 vtn_handle_decoration(struct vtn_builder
*b
, SpvOp opcode
,
557 const uint32_t *w
, unsigned count
)
559 const uint32_t *w_end
= w
+ count
;
560 const uint32_t target
= w
[1];
564 case SpvOpDecorationGroup
:
565 vtn_push_value(b
, target
, vtn_value_type_decoration_group
);
569 case SpvOpDecorateId
:
570 case SpvOpMemberDecorate
:
571 case SpvOpDecorateString
:
572 case SpvOpMemberDecorateString
:
573 case SpvOpExecutionMode
:
574 case SpvOpExecutionModeId
: {
575 struct vtn_value
*val
= vtn_untyped_value(b
, target
);
577 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
580 case SpvOpDecorateId
:
581 case SpvOpDecorateString
:
582 dec
->scope
= VTN_DEC_DECORATION
;
584 case SpvOpMemberDecorate
:
585 case SpvOpMemberDecorateString
:
586 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(w
++);
587 vtn_fail_if(dec
->scope
< VTN_DEC_STRUCT_MEMBER0
, /* overflow */
588 "Member argument of OpMemberDecorate too large");
590 case SpvOpExecutionMode
:
591 case SpvOpExecutionModeId
:
592 dec
->scope
= VTN_DEC_EXECUTION_MODE
;
595 unreachable("Invalid decoration opcode");
597 dec
->decoration
= *(w
++);
600 /* Link into the list */
601 dec
->next
= val
->decoration
;
602 val
->decoration
= dec
;
606 case SpvOpGroupMemberDecorate
:
607 case SpvOpGroupDecorate
: {
608 struct vtn_value
*group
=
609 vtn_value(b
, target
, vtn_value_type_decoration_group
);
611 for (; w
< w_end
; w
++) {
612 struct vtn_value
*val
= vtn_untyped_value(b
, *w
);
613 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
616 if (opcode
== SpvOpGroupDecorate
) {
617 dec
->scope
= VTN_DEC_DECORATION
;
619 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(++w
);
620 vtn_fail_if(dec
->scope
< 0, /* Check for overflow */
621 "Member argument of OpGroupMemberDecorate too large");
624 /* Link into the list */
625 dec
->next
= val
->decoration
;
626 val
->decoration
= dec
;
632 unreachable("Unhandled opcode");
636 struct member_decoration_ctx
{
638 struct glsl_struct_field
*fields
;
639 struct vtn_type
*type
;
643 * Returns true if the given type contains a struct decorated Block or
647 vtn_type_contains_block(struct vtn_builder
*b
, struct vtn_type
*type
)
649 switch (type
->base_type
) {
650 case vtn_base_type_array
:
651 return vtn_type_contains_block(b
, type
->array_element
);
652 case vtn_base_type_struct
:
653 if (type
->block
|| type
->buffer_block
)
655 for (unsigned i
= 0; i
< type
->length
; i
++) {
656 if (vtn_type_contains_block(b
, type
->members
[i
]))
665 /** Returns true if two types are "compatible", i.e. you can do an OpLoad,
666 * OpStore, or OpCopyMemory between them without breaking anything.
667 * Technically, the SPIR-V rules require the exact same type ID but this lets
668 * us internally be a bit looser.
671 vtn_types_compatible(struct vtn_builder
*b
,
672 struct vtn_type
*t1
, struct vtn_type
*t2
)
674 if (t1
->id
== t2
->id
)
677 if (t1
->base_type
!= t2
->base_type
)
680 switch (t1
->base_type
) {
681 case vtn_base_type_void
:
682 case vtn_base_type_scalar
:
683 case vtn_base_type_vector
:
684 case vtn_base_type_matrix
:
685 case vtn_base_type_image
:
686 case vtn_base_type_sampler
:
687 case vtn_base_type_sampled_image
:
688 return t1
->type
== t2
->type
;
690 case vtn_base_type_array
:
691 return t1
->length
== t2
->length
&&
692 vtn_types_compatible(b
, t1
->array_element
, t2
->array_element
);
694 case vtn_base_type_pointer
:
695 return vtn_types_compatible(b
, t1
->deref
, t2
->deref
);
697 case vtn_base_type_struct
:
698 if (t1
->length
!= t2
->length
)
701 for (unsigned i
= 0; i
< t1
->length
; i
++) {
702 if (!vtn_types_compatible(b
, t1
->members
[i
], t2
->members
[i
]))
707 case vtn_base_type_function
:
708 /* This case shouldn't get hit since you can't copy around function
709 * types. Just require them to be identical.
714 vtn_fail("Invalid base type");
718 vtn_type_without_array(struct vtn_type
*type
)
720 while (type
->base_type
== vtn_base_type_array
)
721 type
= type
->array_element
;
725 /* does a shallow copy of a vtn_type */
727 static struct vtn_type
*
728 vtn_type_copy(struct vtn_builder
*b
, struct vtn_type
*src
)
730 struct vtn_type
*dest
= ralloc(b
, struct vtn_type
);
733 switch (src
->base_type
) {
734 case vtn_base_type_void
:
735 case vtn_base_type_scalar
:
736 case vtn_base_type_vector
:
737 case vtn_base_type_matrix
:
738 case vtn_base_type_array
:
739 case vtn_base_type_pointer
:
740 case vtn_base_type_image
:
741 case vtn_base_type_sampler
:
742 case vtn_base_type_sampled_image
:
743 /* Nothing more to do */
746 case vtn_base_type_struct
:
747 dest
->members
= ralloc_array(b
, struct vtn_type
*, src
->length
);
748 memcpy(dest
->members
, src
->members
,
749 src
->length
* sizeof(src
->members
[0]));
751 dest
->offsets
= ralloc_array(b
, unsigned, src
->length
);
752 memcpy(dest
->offsets
, src
->offsets
,
753 src
->length
* sizeof(src
->offsets
[0]));
756 case vtn_base_type_function
:
757 dest
->params
= ralloc_array(b
, struct vtn_type
*, src
->length
);
758 memcpy(dest
->params
, src
->params
, src
->length
* sizeof(src
->params
[0]));
765 static const struct glsl_type
*
766 wrap_type_in_array(const struct glsl_type
*type
,
767 const struct glsl_type
*array_type
)
769 if (!glsl_type_is_array(array_type
))
772 const struct glsl_type
*elem_type
=
773 wrap_type_in_array(type
, glsl_get_array_element(array_type
));
774 return glsl_array_type(elem_type
, glsl_get_length(array_type
),
775 glsl_get_explicit_stride(array_type
));
779 vtn_type_needs_explicit_layout(struct vtn_builder
*b
, enum vtn_variable_mode mode
)
781 /* For OpenCL we never want to strip the info from the types, and it makes
782 * type comparisons easier in later stages.
784 if (b
->options
->environment
== NIR_SPIRV_OPENCL
)
788 case vtn_variable_mode_input
:
789 case vtn_variable_mode_output
:
790 /* Layout decorations kept because we need offsets for XFB arrays of
793 return b
->shader
->info
.has_transform_feedback_varyings
;
795 case vtn_variable_mode_ssbo
:
796 case vtn_variable_mode_phys_ssbo
:
797 case vtn_variable_mode_ubo
:
805 const struct glsl_type
*
806 vtn_type_get_nir_type(struct vtn_builder
*b
, struct vtn_type
*type
,
807 enum vtn_variable_mode mode
)
809 if (mode
== vtn_variable_mode_atomic_counter
) {
810 vtn_fail_if(glsl_without_array(type
->type
) != glsl_uint_type(),
811 "Variables in the AtomicCounter storage class should be "
812 "(possibly arrays of arrays of) uint.");
813 return wrap_type_in_array(glsl_atomic_uint_type(), type
->type
);
816 if (mode
== vtn_variable_mode_uniform
) {
817 switch (type
->base_type
) {
818 case vtn_base_type_array
: {
819 const struct glsl_type
*elem_type
=
820 vtn_type_get_nir_type(b
, type
->array_element
, mode
);
822 return glsl_array_type(elem_type
, type
->length
,
823 glsl_get_explicit_stride(type
->type
));
826 case vtn_base_type_struct
: {
827 bool need_new_struct
= false;
828 const uint32_t num_fields
= type
->length
;
829 NIR_VLA(struct glsl_struct_field
, fields
, num_fields
);
830 for (unsigned i
= 0; i
< num_fields
; i
++) {
831 fields
[i
] = *glsl_get_struct_field_data(type
->type
, i
);
832 const struct glsl_type
*field_nir_type
=
833 vtn_type_get_nir_type(b
, type
->members
[i
], mode
);
834 if (fields
[i
].type
!= field_nir_type
) {
835 fields
[i
].type
= field_nir_type
;
836 need_new_struct
= true;
839 if (need_new_struct
) {
840 if (glsl_type_is_interface(type
->type
)) {
841 return glsl_interface_type(fields
, num_fields
,
842 /* packing */ 0, false,
843 glsl_get_type_name(type
->type
));
845 return glsl_struct_type(fields
, num_fields
,
846 glsl_get_type_name(type
->type
),
847 glsl_struct_type_is_packed(type
->type
));
850 /* No changes, just pass it on */
855 case vtn_base_type_image
:
856 return type
->glsl_image
;
858 case vtn_base_type_sampler
:
859 return glsl_bare_sampler_type();
861 case vtn_base_type_sampled_image
:
862 return type
->image
->glsl_image
;
869 /* Layout decorations are allowed but ignored in certain conditions,
870 * to allow SPIR-V generators perform type deduplication. Discard
871 * unnecessary ones when passing to NIR.
873 if (!vtn_type_needs_explicit_layout(b
, mode
))
874 return glsl_get_bare_type(type
->type
);
879 static struct vtn_type
*
880 mutable_matrix_member(struct vtn_builder
*b
, struct vtn_type
*type
, int member
)
882 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
883 type
= type
->members
[member
];
885 /* We may have an array of matrices.... Oh, joy! */
886 while (glsl_type_is_array(type
->type
)) {
887 type
->array_element
= vtn_type_copy(b
, type
->array_element
);
888 type
= type
->array_element
;
891 vtn_assert(glsl_type_is_matrix(type
->type
));
897 vtn_handle_access_qualifier(struct vtn_builder
*b
, struct vtn_type
*type
,
898 int member
, enum gl_access_qualifier access
)
900 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
901 type
= type
->members
[member
];
903 type
->access
|= access
;
907 array_stride_decoration_cb(struct vtn_builder
*b
,
908 struct vtn_value
*val
, int member
,
909 const struct vtn_decoration
*dec
, void *void_ctx
)
911 struct vtn_type
*type
= val
->type
;
913 if (dec
->decoration
== SpvDecorationArrayStride
) {
914 if (vtn_type_contains_block(b
, type
)) {
915 vtn_warn("The ArrayStride decoration cannot be applied to an array "
916 "type which contains a structure type decorated Block "
918 /* Ignore the decoration */
920 vtn_fail_if(dec
->operands
[0] == 0, "ArrayStride must be non-zero");
921 type
->stride
= dec
->operands
[0];
927 struct_member_decoration_cb(struct vtn_builder
*b
,
928 UNUSED
struct vtn_value
*val
, int member
,
929 const struct vtn_decoration
*dec
, void *void_ctx
)
931 struct member_decoration_ctx
*ctx
= void_ctx
;
936 assert(member
< ctx
->num_fields
);
938 switch (dec
->decoration
) {
939 case SpvDecorationRelaxedPrecision
:
940 case SpvDecorationUniform
:
941 case SpvDecorationUniformId
:
942 break; /* FIXME: Do nothing with this for now. */
943 case SpvDecorationNonWritable
:
944 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_NON_WRITEABLE
);
946 case SpvDecorationNonReadable
:
947 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_NON_READABLE
);
949 case SpvDecorationVolatile
:
950 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_VOLATILE
);
952 case SpvDecorationCoherent
:
953 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_COHERENT
);
955 case SpvDecorationNoPerspective
:
956 ctx
->fields
[member
].interpolation
= INTERP_MODE_NOPERSPECTIVE
;
958 case SpvDecorationFlat
:
959 ctx
->fields
[member
].interpolation
= INTERP_MODE_FLAT
;
961 case SpvDecorationExplicitInterpAMD
:
962 ctx
->fields
[member
].interpolation
= INTERP_MODE_EXPLICIT
;
964 case SpvDecorationCentroid
:
965 ctx
->fields
[member
].centroid
= true;
967 case SpvDecorationSample
:
968 ctx
->fields
[member
].sample
= true;
970 case SpvDecorationStream
:
971 /* This is handled later by var_decoration_cb in vtn_variables.c */
973 case SpvDecorationLocation
:
974 ctx
->fields
[member
].location
= dec
->operands
[0];
976 case SpvDecorationComponent
:
977 break; /* FIXME: What should we do with these? */
978 case SpvDecorationBuiltIn
:
979 ctx
->type
->members
[member
] = vtn_type_copy(b
, ctx
->type
->members
[member
]);
980 ctx
->type
->members
[member
]->is_builtin
= true;
981 ctx
->type
->members
[member
]->builtin
= dec
->operands
[0];
982 ctx
->type
->builtin_block
= true;
984 case SpvDecorationOffset
:
985 ctx
->type
->offsets
[member
] = dec
->operands
[0];
986 ctx
->fields
[member
].offset
= dec
->operands
[0];
988 case SpvDecorationMatrixStride
:
989 /* Handled as a second pass */
991 case SpvDecorationColMajor
:
992 break; /* Nothing to do here. Column-major is the default. */
993 case SpvDecorationRowMajor
:
994 mutable_matrix_member(b
, ctx
->type
, member
)->row_major
= true;
997 case SpvDecorationPatch
:
1000 case SpvDecorationSpecId
:
1001 case SpvDecorationBlock
:
1002 case SpvDecorationBufferBlock
:
1003 case SpvDecorationArrayStride
:
1004 case SpvDecorationGLSLShared
:
1005 case SpvDecorationGLSLPacked
:
1006 case SpvDecorationInvariant
:
1007 case SpvDecorationRestrict
:
1008 case SpvDecorationAliased
:
1009 case SpvDecorationConstant
:
1010 case SpvDecorationIndex
:
1011 case SpvDecorationBinding
:
1012 case SpvDecorationDescriptorSet
:
1013 case SpvDecorationLinkageAttributes
:
1014 case SpvDecorationNoContraction
:
1015 case SpvDecorationInputAttachmentIndex
:
1016 vtn_warn("Decoration not allowed on struct members: %s",
1017 spirv_decoration_to_string(dec
->decoration
));
1020 case SpvDecorationXfbBuffer
:
1021 case SpvDecorationXfbStride
:
1022 /* This is handled later by var_decoration_cb in vtn_variables.c */
1025 case SpvDecorationCPacked
:
1026 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
)
1027 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1028 spirv_decoration_to_string(dec
->decoration
));
1030 ctx
->type
->packed
= true;
1033 case SpvDecorationSaturatedConversion
:
1034 case SpvDecorationFuncParamAttr
:
1035 case SpvDecorationFPRoundingMode
:
1036 case SpvDecorationFPFastMathMode
:
1037 case SpvDecorationAlignment
:
1038 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
) {
1039 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1040 spirv_decoration_to_string(dec
->decoration
));
1044 case SpvDecorationUserSemantic
:
1045 case SpvDecorationUserTypeGOOGLE
:
1046 /* User semantic decorations can safely be ignored by the driver. */
1050 vtn_fail_with_decoration("Unhandled decoration", dec
->decoration
);
1054 /** Chases the array type all the way down to the tail and rewrites the
1055 * glsl_types to be based off the tail's glsl_type.
1058 vtn_array_type_rewrite_glsl_type(struct vtn_type
*type
)
1060 if (type
->base_type
!= vtn_base_type_array
)
1063 vtn_array_type_rewrite_glsl_type(type
->array_element
);
1065 type
->type
= glsl_array_type(type
->array_element
->type
,
1066 type
->length
, type
->stride
);
1069 /* Matrix strides are handled as a separate pass because we need to know
1070 * whether the matrix is row-major or not first.
1073 struct_member_matrix_stride_cb(struct vtn_builder
*b
,
1074 UNUSED
struct vtn_value
*val
, int member
,
1075 const struct vtn_decoration
*dec
,
1078 if (dec
->decoration
!= SpvDecorationMatrixStride
)
1081 vtn_fail_if(member
< 0,
1082 "The MatrixStride decoration is only allowed on members "
1084 vtn_fail_if(dec
->operands
[0] == 0, "MatrixStride must be non-zero");
1086 struct member_decoration_ctx
*ctx
= void_ctx
;
1088 struct vtn_type
*mat_type
= mutable_matrix_member(b
, ctx
->type
, member
);
1089 if (mat_type
->row_major
) {
1090 mat_type
->array_element
= vtn_type_copy(b
, mat_type
->array_element
);
1091 mat_type
->stride
= mat_type
->array_element
->stride
;
1092 mat_type
->array_element
->stride
= dec
->operands
[0];
1094 mat_type
->type
= glsl_explicit_matrix_type(mat_type
->type
,
1095 dec
->operands
[0], true);
1096 mat_type
->array_element
->type
= glsl_get_column_type(mat_type
->type
);
1098 vtn_assert(mat_type
->array_element
->stride
> 0);
1099 mat_type
->stride
= dec
->operands
[0];
1101 mat_type
->type
= glsl_explicit_matrix_type(mat_type
->type
,
1102 dec
->operands
[0], false);
1105 /* Now that we've replaced the glsl_type with a properly strided matrix
1106 * type, rewrite the member type so that it's an array of the proper kind
1109 vtn_array_type_rewrite_glsl_type(ctx
->type
->members
[member
]);
1110 ctx
->fields
[member
].type
= ctx
->type
->members
[member
]->type
;
1114 struct_block_decoration_cb(struct vtn_builder
*b
,
1115 struct vtn_value
*val
, int member
,
1116 const struct vtn_decoration
*dec
, void *ctx
)
1121 struct vtn_type
*type
= val
->type
;
1122 if (dec
->decoration
== SpvDecorationBlock
)
1124 else if (dec
->decoration
== SpvDecorationBufferBlock
)
1125 type
->buffer_block
= true;
1129 type_decoration_cb(struct vtn_builder
*b
,
1130 struct vtn_value
*val
, int member
,
1131 const struct vtn_decoration
*dec
, UNUSED
void *ctx
)
1133 struct vtn_type
*type
= val
->type
;
1136 /* This should have been handled by OpTypeStruct */
1137 assert(val
->type
->base_type
== vtn_base_type_struct
);
1138 assert(member
>= 0 && member
< val
->type
->length
);
1142 switch (dec
->decoration
) {
1143 case SpvDecorationArrayStride
:
1144 vtn_assert(type
->base_type
== vtn_base_type_array
||
1145 type
->base_type
== vtn_base_type_pointer
);
1147 case SpvDecorationBlock
:
1148 vtn_assert(type
->base_type
== vtn_base_type_struct
);
1149 vtn_assert(type
->block
);
1151 case SpvDecorationBufferBlock
:
1152 vtn_assert(type
->base_type
== vtn_base_type_struct
);
1153 vtn_assert(type
->buffer_block
);
1155 case SpvDecorationGLSLShared
:
1156 case SpvDecorationGLSLPacked
:
1157 /* Ignore these, since we get explicit offsets anyways */
1160 case SpvDecorationRowMajor
:
1161 case SpvDecorationColMajor
:
1162 case SpvDecorationMatrixStride
:
1163 case SpvDecorationBuiltIn
:
1164 case SpvDecorationNoPerspective
:
1165 case SpvDecorationFlat
:
1166 case SpvDecorationPatch
:
1167 case SpvDecorationCentroid
:
1168 case SpvDecorationSample
:
1169 case SpvDecorationExplicitInterpAMD
:
1170 case SpvDecorationVolatile
:
1171 case SpvDecorationCoherent
:
1172 case SpvDecorationNonWritable
:
1173 case SpvDecorationNonReadable
:
1174 case SpvDecorationUniform
:
1175 case SpvDecorationUniformId
:
1176 case SpvDecorationLocation
:
1177 case SpvDecorationComponent
:
1178 case SpvDecorationOffset
:
1179 case SpvDecorationXfbBuffer
:
1180 case SpvDecorationXfbStride
:
1181 case SpvDecorationUserSemantic
:
1182 vtn_warn("Decoration only allowed for struct members: %s",
1183 spirv_decoration_to_string(dec
->decoration
));
1186 case SpvDecorationStream
:
1187 /* We don't need to do anything here, as stream is filled up when
1188 * aplying the decoration to a variable, just check that if it is not a
1189 * struct member, it should be a struct.
1191 vtn_assert(type
->base_type
== vtn_base_type_struct
);
1194 case SpvDecorationRelaxedPrecision
:
1195 case SpvDecorationSpecId
:
1196 case SpvDecorationInvariant
:
1197 case SpvDecorationRestrict
:
1198 case SpvDecorationAliased
:
1199 case SpvDecorationConstant
:
1200 case SpvDecorationIndex
:
1201 case SpvDecorationBinding
:
1202 case SpvDecorationDescriptorSet
:
1203 case SpvDecorationLinkageAttributes
:
1204 case SpvDecorationNoContraction
:
1205 case SpvDecorationInputAttachmentIndex
:
1206 vtn_warn("Decoration not allowed on types: %s",
1207 spirv_decoration_to_string(dec
->decoration
));
1210 case SpvDecorationCPacked
:
1211 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
)
1212 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1213 spirv_decoration_to_string(dec
->decoration
));
1215 type
->packed
= true;
1218 case SpvDecorationSaturatedConversion
:
1219 case SpvDecorationFuncParamAttr
:
1220 case SpvDecorationFPRoundingMode
:
1221 case SpvDecorationFPFastMathMode
:
1222 case SpvDecorationAlignment
:
1223 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1224 spirv_decoration_to_string(dec
->decoration
));
1227 case SpvDecorationUserTypeGOOGLE
:
1228 /* User semantic decorations can safely be ignored by the driver. */
1232 vtn_fail_with_decoration("Unhandled decoration", dec
->decoration
);
1237 translate_image_format(struct vtn_builder
*b
, SpvImageFormat format
)
1240 case SpvImageFormatUnknown
: return PIPE_FORMAT_NONE
;
1241 case SpvImageFormatRgba32f
: return PIPE_FORMAT_R32G32B32A32_FLOAT
;
1242 case SpvImageFormatRgba16f
: return PIPE_FORMAT_R16G16B16A16_FLOAT
;
1243 case SpvImageFormatR32f
: return PIPE_FORMAT_R32_FLOAT
;
1244 case SpvImageFormatRgba8
: return PIPE_FORMAT_R8G8B8A8_UNORM
;
1245 case SpvImageFormatRgba8Snorm
: return PIPE_FORMAT_R8G8B8A8_SNORM
;
1246 case SpvImageFormatRg32f
: return PIPE_FORMAT_R32G32_FLOAT
;
1247 case SpvImageFormatRg16f
: return PIPE_FORMAT_R16G16_FLOAT
;
1248 case SpvImageFormatR11fG11fB10f
: return PIPE_FORMAT_R11G11B10_FLOAT
;
1249 case SpvImageFormatR16f
: return PIPE_FORMAT_R16_FLOAT
;
1250 case SpvImageFormatRgba16
: return PIPE_FORMAT_R16G16B16A16_UNORM
;
1251 case SpvImageFormatRgb10A2
: return PIPE_FORMAT_R10G10B10A2_UNORM
;
1252 case SpvImageFormatRg16
: return PIPE_FORMAT_R16G16_UNORM
;
1253 case SpvImageFormatRg8
: return PIPE_FORMAT_R8G8_UNORM
;
1254 case SpvImageFormatR16
: return PIPE_FORMAT_R16_UNORM
;
1255 case SpvImageFormatR8
: return PIPE_FORMAT_R8_UNORM
;
1256 case SpvImageFormatRgba16Snorm
: return PIPE_FORMAT_R16G16B16A16_SNORM
;
1257 case SpvImageFormatRg16Snorm
: return PIPE_FORMAT_R16G16_SNORM
;
1258 case SpvImageFormatRg8Snorm
: return PIPE_FORMAT_R8G8_SNORM
;
1259 case SpvImageFormatR16Snorm
: return PIPE_FORMAT_R16_SNORM
;
1260 case SpvImageFormatR8Snorm
: return PIPE_FORMAT_R8_SNORM
;
1261 case SpvImageFormatRgba32i
: return PIPE_FORMAT_R32G32B32A32_SINT
;
1262 case SpvImageFormatRgba16i
: return PIPE_FORMAT_R16G16B16A16_SINT
;
1263 case SpvImageFormatRgba8i
: return PIPE_FORMAT_R8G8B8A8_SINT
;
1264 case SpvImageFormatR32i
: return PIPE_FORMAT_R32_SINT
;
1265 case SpvImageFormatRg32i
: return PIPE_FORMAT_R32G32_SINT
;
1266 case SpvImageFormatRg16i
: return PIPE_FORMAT_R16G16_SINT
;
1267 case SpvImageFormatRg8i
: return PIPE_FORMAT_R8G8_SINT
;
1268 case SpvImageFormatR16i
: return PIPE_FORMAT_R16_SINT
;
1269 case SpvImageFormatR8i
: return PIPE_FORMAT_R8_SINT
;
1270 case SpvImageFormatRgba32ui
: return PIPE_FORMAT_R32G32B32A32_UINT
;
1271 case SpvImageFormatRgba16ui
: return PIPE_FORMAT_R16G16B16A16_UINT
;
1272 case SpvImageFormatRgba8ui
: return PIPE_FORMAT_R8G8B8A8_UINT
;
1273 case SpvImageFormatR32ui
: return PIPE_FORMAT_R32_UINT
;
1274 case SpvImageFormatRgb10a2ui
: return PIPE_FORMAT_R10G10B10A2_UINT
;
1275 case SpvImageFormatRg32ui
: return PIPE_FORMAT_R32G32_UINT
;
1276 case SpvImageFormatRg16ui
: return PIPE_FORMAT_R16G16_UINT
;
1277 case SpvImageFormatRg8ui
: return PIPE_FORMAT_R8G8_UINT
;
1278 case SpvImageFormatR16ui
: return PIPE_FORMAT_R16_UINT
;
1279 case SpvImageFormatR8ui
: return PIPE_FORMAT_R8_UINT
;
1281 vtn_fail("Invalid image format: %s (%u)",
1282 spirv_imageformat_to_string(format
), format
);
1287 vtn_handle_type(struct vtn_builder
*b
, SpvOp opcode
,
1288 const uint32_t *w
, unsigned count
)
1290 struct vtn_value
*val
= NULL
;
1292 /* In order to properly handle forward declarations, we have to defer
1293 * allocation for pointer types.
1295 if (opcode
!= SpvOpTypePointer
&& opcode
!= SpvOpTypeForwardPointer
) {
1296 val
= vtn_push_value(b
, w
[1], vtn_value_type_type
);
1297 vtn_fail_if(val
->type
!= NULL
,
1298 "Only pointers can have forward declarations");
1299 val
->type
= rzalloc(b
, struct vtn_type
);
1300 val
->type
->id
= w
[1];
1305 val
->type
->base_type
= vtn_base_type_void
;
1306 val
->type
->type
= glsl_void_type();
1309 val
->type
->base_type
= vtn_base_type_scalar
;
1310 val
->type
->type
= glsl_bool_type();
1311 val
->type
->length
= 1;
1313 case SpvOpTypeInt
: {
1314 int bit_size
= w
[2];
1315 const bool signedness
= w
[3];
1316 val
->type
->base_type
= vtn_base_type_scalar
;
1319 val
->type
->type
= (signedness
? glsl_int64_t_type() : glsl_uint64_t_type());
1322 val
->type
->type
= (signedness
? glsl_int_type() : glsl_uint_type());
1325 val
->type
->type
= (signedness
? glsl_int16_t_type() : glsl_uint16_t_type());
1328 val
->type
->type
= (signedness
? glsl_int8_t_type() : glsl_uint8_t_type());
1331 vtn_fail("Invalid int bit size: %u", bit_size
);
1333 val
->type
->length
= 1;
1337 case SpvOpTypeFloat
: {
1338 int bit_size
= w
[2];
1339 val
->type
->base_type
= vtn_base_type_scalar
;
1342 val
->type
->type
= glsl_float16_t_type();
1345 val
->type
->type
= glsl_float_type();
1348 val
->type
->type
= glsl_double_type();
1351 vtn_fail("Invalid float bit size: %u", bit_size
);
1353 val
->type
->length
= 1;
1357 case SpvOpTypeVector
: {
1358 struct vtn_type
*base
= vtn_get_type(b
, w
[2]);
1359 unsigned elems
= w
[3];
1361 vtn_fail_if(base
->base_type
!= vtn_base_type_scalar
,
1362 "Base type for OpTypeVector must be a scalar");
1363 vtn_fail_if((elems
< 2 || elems
> 4) && (elems
!= 8) && (elems
!= 16),
1364 "Invalid component count for OpTypeVector");
1366 val
->type
->base_type
= vtn_base_type_vector
;
1367 val
->type
->type
= glsl_vector_type(glsl_get_base_type(base
->type
), elems
);
1368 val
->type
->length
= elems
;
1369 val
->type
->stride
= glsl_type_is_boolean(val
->type
->type
)
1370 ? 4 : glsl_get_bit_size(base
->type
) / 8;
1371 val
->type
->array_element
= base
;
1375 case SpvOpTypeMatrix
: {
1376 struct vtn_type
*base
= vtn_get_type(b
, w
[2]);
1377 unsigned columns
= w
[3];
1379 vtn_fail_if(base
->base_type
!= vtn_base_type_vector
,
1380 "Base type for OpTypeMatrix must be a vector");
1381 vtn_fail_if(columns
< 2 || columns
> 4,
1382 "Invalid column count for OpTypeMatrix");
1384 val
->type
->base_type
= vtn_base_type_matrix
;
1385 val
->type
->type
= glsl_matrix_type(glsl_get_base_type(base
->type
),
1386 glsl_get_vector_elements(base
->type
),
1388 vtn_fail_if(glsl_type_is_error(val
->type
->type
),
1389 "Unsupported base type for OpTypeMatrix");
1390 assert(!glsl_type_is_error(val
->type
->type
));
1391 val
->type
->length
= columns
;
1392 val
->type
->array_element
= base
;
1393 val
->type
->row_major
= false;
1394 val
->type
->stride
= 0;
1398 case SpvOpTypeRuntimeArray
:
1399 case SpvOpTypeArray
: {
1400 struct vtn_type
*array_element
= vtn_get_type(b
, w
[2]);
1402 if (opcode
== SpvOpTypeRuntimeArray
) {
1403 /* A length of 0 is used to denote unsized arrays */
1404 val
->type
->length
= 0;
1406 val
->type
->length
= vtn_constant_uint(b
, w
[3]);
1409 val
->type
->base_type
= vtn_base_type_array
;
1410 val
->type
->array_element
= array_element
;
1411 if (b
->shader
->info
.stage
== MESA_SHADER_KERNEL
)
1412 val
->type
->stride
= glsl_get_cl_size(array_element
->type
);
1414 vtn_foreach_decoration(b
, val
, array_stride_decoration_cb
, NULL
);
1415 val
->type
->type
= glsl_array_type(array_element
->type
, val
->type
->length
,
1420 case SpvOpTypeStruct
: {
1421 unsigned num_fields
= count
- 2;
1422 val
->type
->base_type
= vtn_base_type_struct
;
1423 val
->type
->length
= num_fields
;
1424 val
->type
->members
= ralloc_array(b
, struct vtn_type
*, num_fields
);
1425 val
->type
->offsets
= ralloc_array(b
, unsigned, num_fields
);
1426 val
->type
->packed
= false;
1428 NIR_VLA(struct glsl_struct_field
, fields
, count
);
1429 for (unsigned i
= 0; i
< num_fields
; i
++) {
1430 val
->type
->members
[i
] = vtn_get_type(b
, w
[i
+ 2]);
1431 fields
[i
] = (struct glsl_struct_field
) {
1432 .type
= val
->type
->members
[i
]->type
,
1433 .name
= ralloc_asprintf(b
, "field%d", i
),
1439 if (b
->shader
->info
.stage
== MESA_SHADER_KERNEL
) {
1440 unsigned offset
= 0;
1441 for (unsigned i
= 0; i
< num_fields
; i
++) {
1442 offset
= align(offset
, glsl_get_cl_alignment(fields
[i
].type
));
1443 fields
[i
].offset
= offset
;
1444 offset
+= glsl_get_cl_size(fields
[i
].type
);
1448 struct member_decoration_ctx ctx
= {
1449 .num_fields
= num_fields
,
1454 vtn_foreach_decoration(b
, val
, struct_member_decoration_cb
, &ctx
);
1455 vtn_foreach_decoration(b
, val
, struct_member_matrix_stride_cb
, &ctx
);
1457 vtn_foreach_decoration(b
, val
, struct_block_decoration_cb
, NULL
);
1459 const char *name
= val
->name
;
1461 if (val
->type
->block
|| val
->type
->buffer_block
) {
1462 /* Packing will be ignored since types coming from SPIR-V are
1463 * explicitly laid out.
1465 val
->type
->type
= glsl_interface_type(fields
, num_fields
,
1466 /* packing */ 0, false,
1467 name
? name
: "block");
1469 val
->type
->type
= glsl_struct_type(fields
, num_fields
,
1470 name
? name
: "struct", false);
1475 case SpvOpTypeFunction
: {
1476 val
->type
->base_type
= vtn_base_type_function
;
1477 val
->type
->type
= NULL
;
1479 val
->type
->return_type
= vtn_get_type(b
, w
[2]);
1481 const unsigned num_params
= count
- 3;
1482 val
->type
->length
= num_params
;
1483 val
->type
->params
= ralloc_array(b
, struct vtn_type
*, num_params
);
1484 for (unsigned i
= 0; i
< count
- 3; i
++) {
1485 val
->type
->params
[i
] = vtn_get_type(b
, w
[i
+ 3]);
1490 case SpvOpTypePointer
:
1491 case SpvOpTypeForwardPointer
: {
1492 /* We can't blindly push the value because it might be a forward
1495 val
= vtn_untyped_value(b
, w
[1]);
1497 SpvStorageClass storage_class
= w
[2];
1499 if (val
->value_type
== vtn_value_type_invalid
) {
1500 val
->value_type
= vtn_value_type_type
;
1501 val
->type
= rzalloc(b
, struct vtn_type
);
1502 val
->type
->id
= w
[1];
1503 val
->type
->base_type
= vtn_base_type_pointer
;
1504 val
->type
->storage_class
= storage_class
;
1506 /* These can actually be stored to nir_variables and used as SSA
1507 * values so they need a real glsl_type.
1509 enum vtn_variable_mode mode
= vtn_storage_class_to_mode(
1510 b
, storage_class
, NULL
, NULL
);
1511 val
->type
->type
= nir_address_format_to_glsl_type(
1512 vtn_mode_to_address_format(b
, mode
));
1514 vtn_fail_if(val
->type
->storage_class
!= storage_class
,
1515 "The storage classes of an OpTypePointer and any "
1516 "OpTypeForwardPointers that provide forward "
1517 "declarations of it must match.");
1520 if (opcode
== SpvOpTypePointer
) {
1521 vtn_fail_if(val
->type
->deref
!= NULL
,
1522 "While OpTypeForwardPointer can be used to provide a "
1523 "forward declaration of a pointer, OpTypePointer can "
1524 "only be used once for a given id.");
1526 val
->type
->deref
= vtn_get_type(b
, w
[3]);
1528 /* Only certain storage classes use ArrayStride. The others (in
1529 * particular Workgroup) are expected to be laid out by the driver.
1531 switch (storage_class
) {
1532 case SpvStorageClassUniform
:
1533 case SpvStorageClassPushConstant
:
1534 case SpvStorageClassStorageBuffer
:
1535 case SpvStorageClassPhysicalStorageBuffer
:
1536 vtn_foreach_decoration(b
, val
, array_stride_decoration_cb
, NULL
);
1539 /* Nothing to do. */
1543 if (b
->physical_ptrs
) {
1544 switch (storage_class
) {
1545 case SpvStorageClassFunction
:
1546 case SpvStorageClassWorkgroup
:
1547 case SpvStorageClassCrossWorkgroup
:
1548 case SpvStorageClassUniformConstant
:
1549 val
->type
->stride
= align(glsl_get_cl_size(val
->type
->deref
->type
),
1550 glsl_get_cl_alignment(val
->type
->deref
->type
));
1560 case SpvOpTypeImage
: {
1561 val
->type
->base_type
= vtn_base_type_image
;
1563 /* Images are represented in NIR as a scalar SSA value that is the
1564 * result of a deref instruction. An OpLoad on an OpTypeImage pointer
1565 * from UniformConstant memory just takes the NIR deref from the pointer
1566 * and turns it into an SSA value.
1568 val
->type
->type
= nir_address_format_to_glsl_type(
1569 vtn_mode_to_address_format(b
, vtn_variable_mode_function
));
1571 const struct vtn_type
*sampled_type
= vtn_get_type(b
, w
[2]);
1572 vtn_fail_if(sampled_type
->base_type
!= vtn_base_type_scalar
||
1573 glsl_get_bit_size(sampled_type
->type
) != 32,
1574 "Sampled type of OpTypeImage must be a 32-bit scalar");
1576 enum glsl_sampler_dim dim
;
1577 switch ((SpvDim
)w
[3]) {
1578 case SpvDim1D
: dim
= GLSL_SAMPLER_DIM_1D
; break;
1579 case SpvDim2D
: dim
= GLSL_SAMPLER_DIM_2D
; break;
1580 case SpvDim3D
: dim
= GLSL_SAMPLER_DIM_3D
; break;
1581 case SpvDimCube
: dim
= GLSL_SAMPLER_DIM_CUBE
; break;
1582 case SpvDimRect
: dim
= GLSL_SAMPLER_DIM_RECT
; break;
1583 case SpvDimBuffer
: dim
= GLSL_SAMPLER_DIM_BUF
; break;
1584 case SpvDimSubpassData
: dim
= GLSL_SAMPLER_DIM_SUBPASS
; break;
1586 vtn_fail("Invalid SPIR-V image dimensionality: %s (%u)",
1587 spirv_dim_to_string((SpvDim
)w
[3]), w
[3]);
1590 /* w[4]: as per Vulkan spec "Validation Rules within a Module",
1591 * The “Depth” operand of OpTypeImage is ignored.
1593 bool is_array
= w
[5];
1594 bool multisampled
= w
[6];
1595 unsigned sampled
= w
[7];
1596 SpvImageFormat format
= w
[8];
1599 val
->type
->access_qualifier
= w
[9];
1601 val
->type
->access_qualifier
= SpvAccessQualifierReadWrite
;
1604 if (dim
== GLSL_SAMPLER_DIM_2D
)
1605 dim
= GLSL_SAMPLER_DIM_MS
;
1606 else if (dim
== GLSL_SAMPLER_DIM_SUBPASS
)
1607 dim
= GLSL_SAMPLER_DIM_SUBPASS_MS
;
1609 vtn_fail("Unsupported multisampled image type");
1612 val
->type
->image_format
= translate_image_format(b
, format
);
1614 enum glsl_base_type sampled_base_type
=
1615 glsl_get_base_type(sampled_type
->type
);
1617 val
->type
->glsl_image
= glsl_sampler_type(dim
, false, is_array
,
1619 } else if (sampled
== 2) {
1620 val
->type
->glsl_image
= glsl_image_type(dim
, is_array
,
1623 vtn_fail("We need to know if the image will be sampled");
1628 case SpvOpTypeSampledImage
: {
1629 val
->type
->base_type
= vtn_base_type_sampled_image
;
1630 val
->type
->image
= vtn_get_type(b
, w
[2]);
1632 /* Sampled images are represented NIR as a vec2 SSA value where each
1633 * component is the result of a deref instruction. The first component
1634 * is the image and the second is the sampler. An OpLoad on an
1635 * OpTypeSampledImage pointer from UniformConstant memory just takes
1636 * the NIR deref from the pointer and duplicates it to both vector
1639 nir_address_format addr_format
=
1640 vtn_mode_to_address_format(b
, vtn_variable_mode_function
);
1641 assert(nir_address_format_num_components(addr_format
) == 1);
1642 unsigned bit_size
= nir_address_format_bit_size(addr_format
);
1643 assert(bit_size
== 32 || bit_size
== 64);
1645 enum glsl_base_type base_type
=
1646 bit_size
== 32 ? GLSL_TYPE_UINT
: GLSL_TYPE_UINT64
;
1647 val
->type
->type
= glsl_vector_type(base_type
, 2);
1651 case SpvOpTypeSampler
:
1652 val
->type
->base_type
= vtn_base_type_sampler
;
1654 /* Samplers are represented in NIR as a scalar SSA value that is the
1655 * result of a deref instruction. An OpLoad on an OpTypeSampler pointer
1656 * from UniformConstant memory just takes the NIR deref from the pointer
1657 * and turns it into an SSA value.
1659 val
->type
->type
= nir_address_format_to_glsl_type(
1660 vtn_mode_to_address_format(b
, vtn_variable_mode_function
));
1663 case SpvOpTypeOpaque
:
1664 case SpvOpTypeEvent
:
1665 case SpvOpTypeDeviceEvent
:
1666 case SpvOpTypeReserveId
:
1667 case SpvOpTypeQueue
:
1670 vtn_fail_with_opcode("Unhandled opcode", opcode
);
1673 vtn_foreach_decoration(b
, val
, type_decoration_cb
, NULL
);
1675 if (val
->type
->base_type
== vtn_base_type_struct
&&
1676 (val
->type
->block
|| val
->type
->buffer_block
)) {
1677 for (unsigned i
= 0; i
< val
->type
->length
; i
++) {
1678 vtn_fail_if(vtn_type_contains_block(b
, val
->type
->members
[i
]),
1679 "Block and BufferBlock decorations cannot decorate a "
1680 "structure type that is nested at any level inside "
1681 "another structure type decorated with Block or "
1687 static nir_constant
*
1688 vtn_null_constant(struct vtn_builder
*b
, struct vtn_type
*type
)
1690 nir_constant
*c
= rzalloc(b
, nir_constant
);
1692 switch (type
->base_type
) {
1693 case vtn_base_type_scalar
:
1694 case vtn_base_type_vector
:
1695 /* Nothing to do here. It's already initialized to zero */
1698 case vtn_base_type_pointer
: {
1699 enum vtn_variable_mode mode
= vtn_storage_class_to_mode(
1700 b
, type
->storage_class
, type
->deref
, NULL
);
1701 nir_address_format addr_format
= vtn_mode_to_address_format(b
, mode
);
1703 const nir_const_value
*null_value
= nir_address_format_null_value(addr_format
);
1704 memcpy(c
->values
, null_value
,
1705 sizeof(nir_const_value
) * nir_address_format_num_components(addr_format
));
1709 case vtn_base_type_void
:
1710 case vtn_base_type_image
:
1711 case vtn_base_type_sampler
:
1712 case vtn_base_type_sampled_image
:
1713 case vtn_base_type_function
:
1714 /* For those we have to return something but it doesn't matter what. */
1717 case vtn_base_type_matrix
:
1718 case vtn_base_type_array
:
1719 vtn_assert(type
->length
> 0);
1720 c
->num_elements
= type
->length
;
1721 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1723 c
->elements
[0] = vtn_null_constant(b
, type
->array_element
);
1724 for (unsigned i
= 1; i
< c
->num_elements
; i
++)
1725 c
->elements
[i
] = c
->elements
[0];
1728 case vtn_base_type_struct
:
1729 c
->num_elements
= type
->length
;
1730 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1731 for (unsigned i
= 0; i
< c
->num_elements
; i
++)
1732 c
->elements
[i
] = vtn_null_constant(b
, type
->members
[i
]);
1736 vtn_fail("Invalid type for null constant");
1743 spec_constant_decoration_cb(struct vtn_builder
*b
, UNUSED
struct vtn_value
*val
,
1744 ASSERTED
int member
,
1745 const struct vtn_decoration
*dec
, void *data
)
1747 vtn_assert(member
== -1);
1748 if (dec
->decoration
!= SpvDecorationSpecId
)
1751 nir_const_value
*value
= data
;
1752 for (unsigned i
= 0; i
< b
->num_specializations
; i
++) {
1753 if (b
->specializations
[i
].id
== dec
->operands
[0]) {
1754 *value
= b
->specializations
[i
].value
;
1761 handle_workgroup_size_decoration_cb(struct vtn_builder
*b
,
1762 struct vtn_value
*val
,
1763 ASSERTED
int member
,
1764 const struct vtn_decoration
*dec
,
1767 vtn_assert(member
== -1);
1768 if (dec
->decoration
!= SpvDecorationBuiltIn
||
1769 dec
->operands
[0] != SpvBuiltInWorkgroupSize
)
1772 vtn_assert(val
->type
->type
== glsl_vector_type(GLSL_TYPE_UINT
, 3));
1773 b
->workgroup_size_builtin
= val
;
1777 vtn_handle_constant(struct vtn_builder
*b
, SpvOp opcode
,
1778 const uint32_t *w
, unsigned count
)
1780 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_constant
);
1781 val
->constant
= rzalloc(b
, nir_constant
);
1783 case SpvOpConstantTrue
:
1784 case SpvOpConstantFalse
:
1785 case SpvOpSpecConstantTrue
:
1786 case SpvOpSpecConstantFalse
: {
1787 vtn_fail_if(val
->type
->type
!= glsl_bool_type(),
1788 "Result type of %s must be OpTypeBool",
1789 spirv_op_to_string(opcode
));
1791 bool bval
= (opcode
== SpvOpConstantTrue
||
1792 opcode
== SpvOpSpecConstantTrue
);
1794 nir_const_value u32val
= nir_const_value_for_uint(bval
, 32);
1796 if (opcode
== SpvOpSpecConstantTrue
||
1797 opcode
== SpvOpSpecConstantFalse
)
1798 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &u32val
);
1800 val
->constant
->values
[0].b
= u32val
.u32
!= 0;
1805 case SpvOpSpecConstant
: {
1806 vtn_fail_if(val
->type
->base_type
!= vtn_base_type_scalar
,
1807 "Result type of %s must be a scalar",
1808 spirv_op_to_string(opcode
));
1809 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1812 val
->constant
->values
[0].u64
= vtn_u64_literal(&w
[3]);
1815 val
->constant
->values
[0].u32
= w
[3];
1818 val
->constant
->values
[0].u16
= w
[3];
1821 val
->constant
->values
[0].u8
= w
[3];
1824 vtn_fail("Unsupported SpvOpConstant bit size: %u", bit_size
);
1827 if (opcode
== SpvOpSpecConstant
)
1828 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
,
1829 &val
->constant
->values
[0]);
1833 case SpvOpSpecConstantComposite
:
1834 case SpvOpConstantComposite
: {
1835 unsigned elem_count
= count
- 3;
1836 vtn_fail_if(elem_count
!= val
->type
->length
,
1837 "%s has %u constituents, expected %u",
1838 spirv_op_to_string(opcode
), elem_count
, val
->type
->length
);
1840 nir_constant
**elems
= ralloc_array(b
, nir_constant
*, elem_count
);
1841 for (unsigned i
= 0; i
< elem_count
; i
++) {
1842 struct vtn_value
*val
= vtn_untyped_value(b
, w
[i
+ 3]);
1844 if (val
->value_type
== vtn_value_type_constant
) {
1845 elems
[i
] = val
->constant
;
1847 vtn_fail_if(val
->value_type
!= vtn_value_type_undef
,
1848 "only constants or undefs allowed for "
1849 "SpvOpConstantComposite");
1850 /* to make it easier, just insert a NULL constant for now */
1851 elems
[i
] = vtn_null_constant(b
, val
->type
);
1855 switch (val
->type
->base_type
) {
1856 case vtn_base_type_vector
: {
1857 assert(glsl_type_is_vector(val
->type
->type
));
1858 for (unsigned i
= 0; i
< elem_count
; i
++)
1859 val
->constant
->values
[i
] = elems
[i
]->values
[0];
1863 case vtn_base_type_matrix
:
1864 case vtn_base_type_struct
:
1865 case vtn_base_type_array
:
1866 ralloc_steal(val
->constant
, elems
);
1867 val
->constant
->num_elements
= elem_count
;
1868 val
->constant
->elements
= elems
;
1872 vtn_fail("Result type of %s must be a composite type",
1873 spirv_op_to_string(opcode
));
1878 case SpvOpSpecConstantOp
: {
1879 nir_const_value u32op
= nir_const_value_for_uint(w
[3], 32);
1880 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &u32op
);
1881 SpvOp opcode
= u32op
.u32
;
1883 case SpvOpVectorShuffle
: {
1884 struct vtn_value
*v0
= &b
->values
[w
[4]];
1885 struct vtn_value
*v1
= &b
->values
[w
[5]];
1887 vtn_assert(v0
->value_type
== vtn_value_type_constant
||
1888 v0
->value_type
== vtn_value_type_undef
);
1889 vtn_assert(v1
->value_type
== vtn_value_type_constant
||
1890 v1
->value_type
== vtn_value_type_undef
);
1892 unsigned len0
= glsl_get_vector_elements(v0
->type
->type
);
1893 unsigned len1
= glsl_get_vector_elements(v1
->type
->type
);
1895 vtn_assert(len0
+ len1
< 16);
1897 unsigned bit_size
= glsl_get_bit_size(val
->type
->type
);
1898 unsigned bit_size0
= glsl_get_bit_size(v0
->type
->type
);
1899 unsigned bit_size1
= glsl_get_bit_size(v1
->type
->type
);
1901 vtn_assert(bit_size
== bit_size0
&& bit_size
== bit_size1
);
1902 (void)bit_size0
; (void)bit_size1
;
1904 nir_const_value undef
= { .u64
= 0xdeadbeefdeadbeef };
1905 nir_const_value combined
[NIR_MAX_VEC_COMPONENTS
* 2];
1907 if (v0
->value_type
== vtn_value_type_constant
) {
1908 for (unsigned i
= 0; i
< len0
; i
++)
1909 combined
[i
] = v0
->constant
->values
[i
];
1911 if (v1
->value_type
== vtn_value_type_constant
) {
1912 for (unsigned i
= 0; i
< len1
; i
++)
1913 combined
[len0
+ i
] = v1
->constant
->values
[i
];
1916 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1917 uint32_t comp
= w
[i
+ 6];
1918 if (comp
== (uint32_t)-1) {
1919 /* If component is not used, set the value to a known constant
1920 * to detect if it is wrongly used.
1922 val
->constant
->values
[j
] = undef
;
1924 vtn_fail_if(comp
>= len0
+ len1
,
1925 "All Component literals must either be FFFFFFFF "
1926 "or in [0, N - 1] (inclusive).");
1927 val
->constant
->values
[j
] = combined
[comp
];
1933 case SpvOpCompositeExtract
:
1934 case SpvOpCompositeInsert
: {
1935 struct vtn_value
*comp
;
1936 unsigned deref_start
;
1937 struct nir_constant
**c
;
1938 if (opcode
== SpvOpCompositeExtract
) {
1939 comp
= vtn_value(b
, w
[4], vtn_value_type_constant
);
1941 c
= &comp
->constant
;
1943 comp
= vtn_value(b
, w
[5], vtn_value_type_constant
);
1945 val
->constant
= nir_constant_clone(comp
->constant
,
1951 const struct vtn_type
*type
= comp
->type
;
1952 for (unsigned i
= deref_start
; i
< count
; i
++) {
1953 vtn_fail_if(w
[i
] > type
->length
,
1954 "%uth index of %s is %u but the type has only "
1955 "%u elements", i
- deref_start
,
1956 spirv_op_to_string(opcode
), w
[i
], type
->length
);
1958 switch (type
->base_type
) {
1959 case vtn_base_type_vector
:
1961 type
= type
->array_element
;
1964 case vtn_base_type_matrix
:
1965 case vtn_base_type_array
:
1966 c
= &(*c
)->elements
[w
[i
]];
1967 type
= type
->array_element
;
1970 case vtn_base_type_struct
:
1971 c
= &(*c
)->elements
[w
[i
]];
1972 type
= type
->members
[w
[i
]];
1976 vtn_fail("%s must only index into composite types",
1977 spirv_op_to_string(opcode
));
1981 if (opcode
== SpvOpCompositeExtract
) {
1985 unsigned num_components
= type
->length
;
1986 for (unsigned i
= 0; i
< num_components
; i
++)
1987 val
->constant
->values
[i
] = (*c
)->values
[elem
+ i
];
1990 struct vtn_value
*insert
=
1991 vtn_value(b
, w
[4], vtn_value_type_constant
);
1992 vtn_assert(insert
->type
== type
);
1994 *c
= insert
->constant
;
1996 unsigned num_components
= type
->length
;
1997 for (unsigned i
= 0; i
< num_components
; i
++)
1998 (*c
)->values
[elem
+ i
] = insert
->constant
->values
[i
];
2006 nir_alu_type dst_alu_type
= nir_get_nir_type_for_glsl_type(val
->type
->type
);
2007 nir_alu_type src_alu_type
= dst_alu_type
;
2008 unsigned num_components
= glsl_get_vector_elements(val
->type
->type
);
2011 vtn_assert(count
<= 7);
2017 /* We have a source in a conversion */
2019 nir_get_nir_type_for_glsl_type(vtn_get_value_type(b
, w
[4])->type
);
2020 /* We use the bitsize of the conversion source to evaluate the opcode later */
2021 bit_size
= glsl_get_bit_size(vtn_get_value_type(b
, w
[4])->type
);
2024 bit_size
= glsl_get_bit_size(val
->type
->type
);
2027 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(b
, opcode
, &swap
,
2028 nir_alu_type_get_type_size(src_alu_type
),
2029 nir_alu_type_get_type_size(dst_alu_type
));
2030 nir_const_value src
[3][NIR_MAX_VEC_COMPONENTS
];
2032 for (unsigned i
= 0; i
< count
- 4; i
++) {
2033 struct vtn_value
*src_val
=
2034 vtn_value(b
, w
[4 + i
], vtn_value_type_constant
);
2036 /* If this is an unsized source, pull the bit size from the
2037 * source; otherwise, we'll use the bit size from the destination.
2039 if (!nir_alu_type_get_type_size(nir_op_infos
[op
].input_types
[i
]))
2040 bit_size
= glsl_get_bit_size(src_val
->type
->type
);
2042 unsigned src_comps
= nir_op_infos
[op
].input_sizes
[i
] ?
2043 nir_op_infos
[op
].input_sizes
[i
] :
2046 unsigned j
= swap
? 1 - i
: i
;
2047 for (unsigned c
= 0; c
< src_comps
; c
++)
2048 src
[j
][c
] = src_val
->constant
->values
[c
];
2051 /* fix up fixed size sources */
2058 for (unsigned i
= 0; i
< num_components
; ++i
) {
2060 case 64: src
[1][i
].u32
= src
[1][i
].u64
; break;
2061 case 16: src
[1][i
].u32
= src
[1][i
].u16
; break;
2062 case 8: src
[1][i
].u32
= src
[1][i
].u8
; break;
2071 nir_const_value
*srcs
[3] = {
2072 src
[0], src
[1], src
[2],
2074 nir_eval_const_opcode(op
, val
->constant
->values
,
2075 num_components
, bit_size
, srcs
,
2076 b
->shader
->info
.float_controls_execution_mode
);
2083 case SpvOpConstantNull
:
2084 val
->constant
= vtn_null_constant(b
, val
->type
);
2087 case SpvOpConstantSampler
:
2088 vtn_fail("OpConstantSampler requires Kernel Capability");
2092 vtn_fail_with_opcode("Unhandled opcode", opcode
);
2095 /* Now that we have the value, update the workgroup size if needed */
2096 vtn_foreach_decoration(b
, val
, handle_workgroup_size_decoration_cb
, NULL
);
2099 SpvMemorySemanticsMask
2100 vtn_storage_class_to_memory_semantics(SpvStorageClass sc
)
2103 case SpvStorageClassStorageBuffer
:
2104 case SpvStorageClassPhysicalStorageBuffer
:
2105 return SpvMemorySemanticsUniformMemoryMask
;
2106 case SpvStorageClassWorkgroup
:
2107 return SpvMemorySemanticsWorkgroupMemoryMask
;
2109 return SpvMemorySemanticsMaskNone
;
2114 vtn_split_barrier_semantics(struct vtn_builder
*b
,
2115 SpvMemorySemanticsMask semantics
,
2116 SpvMemorySemanticsMask
*before
,
2117 SpvMemorySemanticsMask
*after
)
2119 /* For memory semantics embedded in operations, we split them into up to
2120 * two barriers, to be added before and after the operation. This is less
2121 * strict than if we propagated until the final backend stage, but still
2122 * result in correct execution.
2124 * A further improvement could be pipe this information (and use!) into the
2125 * next compiler layers, at the expense of making the handling of barriers
2129 *before
= SpvMemorySemanticsMaskNone
;
2130 *after
= SpvMemorySemanticsMaskNone
;
2132 SpvMemorySemanticsMask order_semantics
=
2133 semantics
& (SpvMemorySemanticsAcquireMask
|
2134 SpvMemorySemanticsReleaseMask
|
2135 SpvMemorySemanticsAcquireReleaseMask
|
2136 SpvMemorySemanticsSequentiallyConsistentMask
);
2138 if (util_bitcount(order_semantics
) > 1) {
2139 /* Old GLSLang versions incorrectly set all the ordering bits. This was
2140 * fixed in c51287d744fb6e7e9ccc09f6f8451e6c64b1dad6 of glslang repo,
2141 * and it is in GLSLang since revision "SPIRV99.1321" (from Jul-2016).
2143 vtn_warn("Multiple memory ordering semantics specified, "
2144 "assuming AcquireRelease.");
2145 order_semantics
= SpvMemorySemanticsAcquireReleaseMask
;
2148 const SpvMemorySemanticsMask av_vis_semantics
=
2149 semantics
& (SpvMemorySemanticsMakeAvailableMask
|
2150 SpvMemorySemanticsMakeVisibleMask
);
2152 const SpvMemorySemanticsMask storage_semantics
=
2153 semantics
& (SpvMemorySemanticsUniformMemoryMask
|
2154 SpvMemorySemanticsSubgroupMemoryMask
|
2155 SpvMemorySemanticsWorkgroupMemoryMask
|
2156 SpvMemorySemanticsCrossWorkgroupMemoryMask
|
2157 SpvMemorySemanticsAtomicCounterMemoryMask
|
2158 SpvMemorySemanticsImageMemoryMask
|
2159 SpvMemorySemanticsOutputMemoryMask
);
2161 const SpvMemorySemanticsMask other_semantics
=
2162 semantics
& ~(order_semantics
| av_vis_semantics
| storage_semantics
);
2164 if (other_semantics
)
2165 vtn_warn("Ignoring unhandled memory semantics: %u\n", other_semantics
);
2167 /* SequentiallyConsistent is treated as AcquireRelease. */
2169 /* The RELEASE barrier happens BEFORE the operation, and it is usually
2170 * associated with a Store. All the write operations with a matching
2171 * semantics will not be reordered after the Store.
2173 if (order_semantics
& (SpvMemorySemanticsReleaseMask
|
2174 SpvMemorySemanticsAcquireReleaseMask
|
2175 SpvMemorySemanticsSequentiallyConsistentMask
)) {
2176 *before
|= SpvMemorySemanticsReleaseMask
| storage_semantics
;
2179 /* The ACQUIRE barrier happens AFTER the operation, and it is usually
2180 * associated with a Load. All the operations with a matching semantics
2181 * will not be reordered before the Load.
2183 if (order_semantics
& (SpvMemorySemanticsAcquireMask
|
2184 SpvMemorySemanticsAcquireReleaseMask
|
2185 SpvMemorySemanticsSequentiallyConsistentMask
)) {
2186 *after
|= SpvMemorySemanticsAcquireMask
| storage_semantics
;
2189 if (av_vis_semantics
& SpvMemorySemanticsMakeVisibleMask
)
2190 *before
|= SpvMemorySemanticsMakeVisibleMask
| storage_semantics
;
2192 if (av_vis_semantics
& SpvMemorySemanticsMakeAvailableMask
)
2193 *after
|= SpvMemorySemanticsMakeAvailableMask
| storage_semantics
;
2196 static nir_memory_semantics
2197 vtn_mem_semantics_to_nir_mem_semantics(struct vtn_builder
*b
,
2198 SpvMemorySemanticsMask semantics
)
2200 nir_memory_semantics nir_semantics
= 0;
2202 SpvMemorySemanticsMask order_semantics
=
2203 semantics
& (SpvMemorySemanticsAcquireMask
|
2204 SpvMemorySemanticsReleaseMask
|
2205 SpvMemorySemanticsAcquireReleaseMask
|
2206 SpvMemorySemanticsSequentiallyConsistentMask
);
2208 if (util_bitcount(order_semantics
) > 1) {
2209 /* Old GLSLang versions incorrectly set all the ordering bits. This was
2210 * fixed in c51287d744fb6e7e9ccc09f6f8451e6c64b1dad6 of glslang repo,
2211 * and it is in GLSLang since revision "SPIRV99.1321" (from Jul-2016).
2213 vtn_warn("Multiple memory ordering semantics bits specified, "
2214 "assuming AcquireRelease.");
2215 order_semantics
= SpvMemorySemanticsAcquireReleaseMask
;
2218 switch (order_semantics
) {
2220 /* Not an ordering barrier. */
2223 case SpvMemorySemanticsAcquireMask
:
2224 nir_semantics
= NIR_MEMORY_ACQUIRE
;
2227 case SpvMemorySemanticsReleaseMask
:
2228 nir_semantics
= NIR_MEMORY_RELEASE
;
2231 case SpvMemorySemanticsSequentiallyConsistentMask
:
2232 /* Fall through. Treated as AcquireRelease in Vulkan. */
2233 case SpvMemorySemanticsAcquireReleaseMask
:
2234 nir_semantics
= NIR_MEMORY_ACQUIRE
| NIR_MEMORY_RELEASE
;
2238 unreachable("Invalid memory order semantics");
2241 if (semantics
& SpvMemorySemanticsMakeAvailableMask
) {
2242 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
2243 "To use MakeAvailable memory semantics the VulkanMemoryModel "
2244 "capability must be declared.");
2245 nir_semantics
|= NIR_MEMORY_MAKE_AVAILABLE
;
2248 if (semantics
& SpvMemorySemanticsMakeVisibleMask
) {
2249 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
2250 "To use MakeVisible memory semantics the VulkanMemoryModel "
2251 "capability must be declared.");
2252 nir_semantics
|= NIR_MEMORY_MAKE_VISIBLE
;
2255 return nir_semantics
;
2258 static nir_variable_mode
2259 vtn_mem_sematics_to_nir_var_modes(struct vtn_builder
*b
,
2260 SpvMemorySemanticsMask semantics
)
2262 /* Vulkan Environment for SPIR-V says "SubgroupMemory, CrossWorkgroupMemory,
2263 * and AtomicCounterMemory are ignored".
2265 semantics
&= ~(SpvMemorySemanticsSubgroupMemoryMask
|
2266 SpvMemorySemanticsCrossWorkgroupMemoryMask
|
2267 SpvMemorySemanticsAtomicCounterMemoryMask
);
2269 /* TODO: Consider adding nir_var_mem_image mode to NIR so it can be used
2270 * for SpvMemorySemanticsImageMemoryMask.
2273 nir_variable_mode modes
= 0;
2274 if (semantics
& (SpvMemorySemanticsUniformMemoryMask
|
2275 SpvMemorySemanticsImageMemoryMask
)) {
2276 modes
|= nir_var_uniform
|
2281 if (semantics
& SpvMemorySemanticsWorkgroupMemoryMask
)
2282 modes
|= nir_var_mem_shared
;
2283 if (semantics
& SpvMemorySemanticsOutputMemoryMask
) {
2284 modes
|= nir_var_shader_out
;
2291 vtn_scope_to_nir_scope(struct vtn_builder
*b
, SpvScope scope
)
2293 nir_scope nir_scope
;
2295 case SpvScopeDevice
:
2296 vtn_fail_if(b
->options
->caps
.vk_memory_model
&&
2297 !b
->options
->caps
.vk_memory_model_device_scope
,
2298 "If the Vulkan memory model is declared and any instruction "
2299 "uses Device scope, the VulkanMemoryModelDeviceScope "
2300 "capability must be declared.");
2301 nir_scope
= NIR_SCOPE_DEVICE
;
2304 case SpvScopeQueueFamily
:
2305 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
2306 "To use Queue Family scope, the VulkanMemoryModel capability "
2307 "must be declared.");
2308 nir_scope
= NIR_SCOPE_QUEUE_FAMILY
;
2311 case SpvScopeWorkgroup
:
2312 nir_scope
= NIR_SCOPE_WORKGROUP
;
2315 case SpvScopeSubgroup
:
2316 nir_scope
= NIR_SCOPE_SUBGROUP
;
2319 case SpvScopeInvocation
:
2320 nir_scope
= NIR_SCOPE_INVOCATION
;
2324 vtn_fail("Invalid memory scope");
2331 vtn_emit_scoped_control_barrier(struct vtn_builder
*b
, SpvScope exec_scope
,
2333 SpvMemorySemanticsMask semantics
)
2335 nir_memory_semantics nir_semantics
=
2336 vtn_mem_semantics_to_nir_mem_semantics(b
, semantics
);
2337 nir_variable_mode modes
= vtn_mem_sematics_to_nir_var_modes(b
, semantics
);
2338 nir_scope nir_exec_scope
= vtn_scope_to_nir_scope(b
, exec_scope
);
2340 /* Memory semantics is optional for OpControlBarrier. */
2341 nir_scope nir_mem_scope
;
2342 if (nir_semantics
== 0 || modes
== 0)
2343 nir_mem_scope
= NIR_SCOPE_NONE
;
2345 nir_mem_scope
= vtn_scope_to_nir_scope(b
, mem_scope
);
2347 nir_scoped_barrier(&b
->nb
, nir_exec_scope
, nir_mem_scope
, nir_semantics
, modes
);
2351 vtn_emit_scoped_memory_barrier(struct vtn_builder
*b
, SpvScope scope
,
2352 SpvMemorySemanticsMask semantics
)
2354 nir_variable_mode modes
= vtn_mem_sematics_to_nir_var_modes(b
, semantics
);
2355 nir_memory_semantics nir_semantics
=
2356 vtn_mem_semantics_to_nir_mem_semantics(b
, semantics
);
2358 /* No barrier to add. */
2359 if (nir_semantics
== 0 || modes
== 0)
2362 nir_scope nir_mem_scope
= vtn_scope_to_nir_scope(b
, scope
);
2363 nir_scoped_barrier(&b
->nb
, NIR_SCOPE_NONE
, nir_mem_scope
, nir_semantics
, modes
);
2366 struct vtn_ssa_value
*
2367 vtn_create_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
2369 /* Always use bare types for SSA values for a couple of reasons:
2371 * 1. Code which emits deref chains should never listen to the explicit
2372 * layout information on the SSA value if any exists. If we've
2373 * accidentally been relying on this, we want to find those bugs.
2375 * 2. We want to be able to quickly check that an SSA value being assigned
2376 * to a SPIR-V value has the right type. Using bare types everywhere
2377 * ensures that we can pointer-compare.
2379 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
2380 val
->type
= glsl_get_bare_type(type
);
2383 if (!glsl_type_is_vector_or_scalar(type
)) {
2384 unsigned elems
= glsl_get_length(val
->type
);
2385 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
2386 if (glsl_type_is_array_or_matrix(type
)) {
2387 const struct glsl_type
*elem_type
= glsl_get_array_element(type
);
2388 for (unsigned i
= 0; i
< elems
; i
++)
2389 val
->elems
[i
] = vtn_create_ssa_value(b
, elem_type
);
2391 vtn_assert(glsl_type_is_struct_or_ifc(type
));
2392 for (unsigned i
= 0; i
< elems
; i
++) {
2393 const struct glsl_type
*elem_type
= glsl_get_struct_field(type
, i
);
2394 val
->elems
[i
] = vtn_create_ssa_value(b
, elem_type
);
2403 vtn_tex_src(struct vtn_builder
*b
, unsigned index
, nir_tex_src_type type
)
2406 src
.src
= nir_src_for_ssa(vtn_get_nir_ssa(b
, index
));
2407 src
.src_type
= type
;
2412 image_operand_arg(struct vtn_builder
*b
, const uint32_t *w
, uint32_t count
,
2413 uint32_t mask_idx
, SpvImageOperandsMask op
)
2415 static const SpvImageOperandsMask ops_with_arg
=
2416 SpvImageOperandsBiasMask
|
2417 SpvImageOperandsLodMask
|
2418 SpvImageOperandsGradMask
|
2419 SpvImageOperandsConstOffsetMask
|
2420 SpvImageOperandsOffsetMask
|
2421 SpvImageOperandsConstOffsetsMask
|
2422 SpvImageOperandsSampleMask
|
2423 SpvImageOperandsMinLodMask
|
2424 SpvImageOperandsMakeTexelAvailableMask
|
2425 SpvImageOperandsMakeTexelVisibleMask
;
2427 assert(util_bitcount(op
) == 1);
2428 assert(w
[mask_idx
] & op
);
2429 assert(op
& ops_with_arg
);
2431 uint32_t idx
= util_bitcount(w
[mask_idx
] & (op
- 1) & ops_with_arg
) + 1;
2433 /* Adjust indices for operands with two arguments. */
2434 static const SpvImageOperandsMask ops_with_two_args
=
2435 SpvImageOperandsGradMask
;
2436 idx
+= util_bitcount(w
[mask_idx
] & (op
- 1) & ops_with_two_args
);
2440 vtn_fail_if(idx
+ (op
& ops_with_two_args
? 1 : 0) >= count
,
2441 "Image op claims to have %s but does not enough "
2442 "following operands", spirv_imageoperands_to_string(op
));
2448 non_uniform_decoration_cb(struct vtn_builder
*b
,
2449 struct vtn_value
*val
, int member
,
2450 const struct vtn_decoration
*dec
, void *void_ctx
)
2452 enum gl_access_qualifier
*access
= void_ctx
;
2453 switch (dec
->decoration
) {
2454 case SpvDecorationNonUniformEXT
:
2455 *access
|= ACCESS_NON_UNIFORM
;
2464 vtn_handle_texture(struct vtn_builder
*b
, SpvOp opcode
,
2465 const uint32_t *w
, unsigned count
)
2467 struct vtn_type
*ret_type
= vtn_get_type(b
, w
[1]);
2469 if (opcode
== SpvOpSampledImage
) {
2470 struct vtn_sampled_image si
= {
2471 .image
= vtn_get_image(b
, w
[3]),
2472 .sampler
= vtn_get_sampler(b
, w
[4]),
2474 vtn_push_sampled_image(b
, w
[2], si
);
2476 } else if (opcode
== SpvOpImage
) {
2477 struct vtn_sampled_image si
= vtn_get_sampled_image(b
, w
[3]);
2478 vtn_push_image(b
, w
[2], si
.image
);
2482 nir_deref_instr
*image
= NULL
, *sampler
= NULL
;
2483 struct vtn_value
*sampled_val
= vtn_untyped_value(b
, w
[3]);
2484 if (sampled_val
->type
->base_type
== vtn_base_type_sampled_image
) {
2485 struct vtn_sampled_image si
= vtn_get_sampled_image(b
, w
[3]);
2487 sampler
= si
.sampler
;
2489 image
= vtn_get_image(b
, w
[3]);
2492 const enum glsl_sampler_dim sampler_dim
= glsl_get_sampler_dim(image
->type
);
2493 const bool is_array
= glsl_sampler_type_is_array(image
->type
);
2494 nir_alu_type dest_type
= nir_type_invalid
;
2496 /* Figure out the base texture operation */
2499 case SpvOpImageSampleImplicitLod
:
2500 case SpvOpImageSampleDrefImplicitLod
:
2501 case SpvOpImageSampleProjImplicitLod
:
2502 case SpvOpImageSampleProjDrefImplicitLod
:
2503 texop
= nir_texop_tex
;
2506 case SpvOpImageSampleExplicitLod
:
2507 case SpvOpImageSampleDrefExplicitLod
:
2508 case SpvOpImageSampleProjExplicitLod
:
2509 case SpvOpImageSampleProjDrefExplicitLod
:
2510 texop
= nir_texop_txl
;
2513 case SpvOpImageFetch
:
2514 if (sampler_dim
== GLSL_SAMPLER_DIM_MS
) {
2515 texop
= nir_texop_txf_ms
;
2517 texop
= nir_texop_txf
;
2521 case SpvOpImageGather
:
2522 case SpvOpImageDrefGather
:
2523 texop
= nir_texop_tg4
;
2526 case SpvOpImageQuerySizeLod
:
2527 case SpvOpImageQuerySize
:
2528 texop
= nir_texop_txs
;
2529 dest_type
= nir_type_int
;
2532 case SpvOpImageQueryLod
:
2533 texop
= nir_texop_lod
;
2534 dest_type
= nir_type_float
;
2537 case SpvOpImageQueryLevels
:
2538 texop
= nir_texop_query_levels
;
2539 dest_type
= nir_type_int
;
2542 case SpvOpImageQuerySamples
:
2543 texop
= nir_texop_texture_samples
;
2544 dest_type
= nir_type_int
;
2547 case SpvOpFragmentFetchAMD
:
2548 texop
= nir_texop_fragment_fetch
;
2551 case SpvOpFragmentMaskFetchAMD
:
2552 texop
= nir_texop_fragment_mask_fetch
;
2556 vtn_fail_with_opcode("Unhandled opcode", opcode
);
2559 nir_tex_src srcs
[10]; /* 10 should be enough */
2560 nir_tex_src
*p
= srcs
;
2562 p
->src
= nir_src_for_ssa(&image
->dest
.ssa
);
2563 p
->src_type
= nir_tex_src_texture_deref
;
2573 vtn_fail_if(sampler
== NULL
,
2574 "%s requires an image of type OpTypeSampledImage",
2575 spirv_op_to_string(opcode
));
2576 p
->src
= nir_src_for_ssa(&sampler
->dest
.ssa
);
2577 p
->src_type
= nir_tex_src_sampler_deref
;
2581 case nir_texop_txf_ms
:
2583 case nir_texop_query_levels
:
2584 case nir_texop_texture_samples
:
2585 case nir_texop_samples_identical
:
2586 case nir_texop_fragment_fetch
:
2587 case nir_texop_fragment_mask_fetch
:
2590 case nir_texop_txf_ms_fb
:
2591 vtn_fail("unexpected nir_texop_txf_ms_fb");
2593 case nir_texop_txf_ms_mcs
:
2594 vtn_fail("unexpected nir_texop_txf_ms_mcs");
2595 case nir_texop_tex_prefetch
:
2596 vtn_fail("unexpected nir_texop_tex_prefetch");
2601 struct nir_ssa_def
*coord
;
2602 unsigned coord_components
;
2604 case SpvOpImageSampleImplicitLod
:
2605 case SpvOpImageSampleExplicitLod
:
2606 case SpvOpImageSampleDrefImplicitLod
:
2607 case SpvOpImageSampleDrefExplicitLod
:
2608 case SpvOpImageSampleProjImplicitLod
:
2609 case SpvOpImageSampleProjExplicitLod
:
2610 case SpvOpImageSampleProjDrefImplicitLod
:
2611 case SpvOpImageSampleProjDrefExplicitLod
:
2612 case SpvOpImageFetch
:
2613 case SpvOpImageGather
:
2614 case SpvOpImageDrefGather
:
2615 case SpvOpImageQueryLod
:
2616 case SpvOpFragmentFetchAMD
:
2617 case SpvOpFragmentMaskFetchAMD
: {
2618 /* All these types have the coordinate as their first real argument */
2619 coord_components
= glsl_get_sampler_dim_coordinate_components(sampler_dim
);
2621 if (is_array
&& texop
!= nir_texop_lod
)
2624 coord
= vtn_get_nir_ssa(b
, w
[idx
++]);
2625 p
->src
= nir_src_for_ssa(nir_channels(&b
->nb
, coord
,
2626 (1 << coord_components
) - 1));
2627 p
->src_type
= nir_tex_src_coord
;
2634 coord_components
= 0;
2639 case SpvOpImageSampleProjImplicitLod
:
2640 case SpvOpImageSampleProjExplicitLod
:
2641 case SpvOpImageSampleProjDrefImplicitLod
:
2642 case SpvOpImageSampleProjDrefExplicitLod
:
2643 /* These have the projector as the last coordinate component */
2644 p
->src
= nir_src_for_ssa(nir_channel(&b
->nb
, coord
, coord_components
));
2645 p
->src_type
= nir_tex_src_projector
;
2653 bool is_shadow
= false;
2654 unsigned gather_component
= 0;
2656 case SpvOpImageSampleDrefImplicitLod
:
2657 case SpvOpImageSampleDrefExplicitLod
:
2658 case SpvOpImageSampleProjDrefImplicitLod
:
2659 case SpvOpImageSampleProjDrefExplicitLod
:
2660 case SpvOpImageDrefGather
:
2661 /* These all have an explicit depth value as their next source */
2663 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_comparator
);
2666 case SpvOpImageGather
:
2667 /* This has a component as its next source */
2668 gather_component
= vtn_constant_uint(b
, w
[idx
++]);
2675 /* For OpImageQuerySizeLod, we always have an LOD */
2676 if (opcode
== SpvOpImageQuerySizeLod
)
2677 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
2679 /* For OpFragmentFetchAMD, we always have a multisample index */
2680 if (opcode
== SpvOpFragmentFetchAMD
)
2681 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ms_index
);
2683 /* Now we need to handle some number of optional arguments */
2684 struct vtn_value
*gather_offsets
= NULL
;
2686 uint32_t operands
= w
[idx
];
2688 if (operands
& SpvImageOperandsBiasMask
) {
2689 vtn_assert(texop
== nir_texop_tex
||
2690 texop
== nir_texop_tg4
);
2691 if (texop
== nir_texop_tex
)
2692 texop
= nir_texop_txb
;
2693 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2694 SpvImageOperandsBiasMask
);
2695 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_bias
);
2698 if (operands
& SpvImageOperandsLodMask
) {
2699 vtn_assert(texop
== nir_texop_txl
|| texop
== nir_texop_txf
||
2700 texop
== nir_texop_txs
|| texop
== nir_texop_tg4
);
2701 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2702 SpvImageOperandsLodMask
);
2703 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_lod
);
2706 if (operands
& SpvImageOperandsGradMask
) {
2707 vtn_assert(texop
== nir_texop_txl
);
2708 texop
= nir_texop_txd
;
2709 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2710 SpvImageOperandsGradMask
);
2711 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_ddx
);
2712 (*p
++) = vtn_tex_src(b
, w
[arg
+ 1], nir_tex_src_ddy
);
2715 vtn_fail_if(util_bitcount(operands
& (SpvImageOperandsConstOffsetsMask
|
2716 SpvImageOperandsOffsetMask
|
2717 SpvImageOperandsConstOffsetMask
)) > 1,
2718 "At most one of the ConstOffset, Offset, and ConstOffsets "
2719 "image operands can be used on a given instruction.");
2721 if (operands
& SpvImageOperandsOffsetMask
) {
2722 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2723 SpvImageOperandsOffsetMask
);
2724 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_offset
);
2727 if (operands
& SpvImageOperandsConstOffsetMask
) {
2728 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2729 SpvImageOperandsConstOffsetMask
);
2730 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_offset
);
2733 if (operands
& SpvImageOperandsConstOffsetsMask
) {
2734 vtn_assert(texop
== nir_texop_tg4
);
2735 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2736 SpvImageOperandsConstOffsetsMask
);
2737 gather_offsets
= vtn_value(b
, w
[arg
], vtn_value_type_constant
);
2740 if (operands
& SpvImageOperandsSampleMask
) {
2741 vtn_assert(texop
== nir_texop_txf_ms
);
2742 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2743 SpvImageOperandsSampleMask
);
2744 texop
= nir_texop_txf_ms
;
2745 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_ms_index
);
2748 if (operands
& SpvImageOperandsMinLodMask
) {
2749 vtn_assert(texop
== nir_texop_tex
||
2750 texop
== nir_texop_txb
||
2751 texop
== nir_texop_txd
);
2752 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2753 SpvImageOperandsMinLodMask
);
2754 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_min_lod
);
2758 nir_tex_instr
*instr
= nir_tex_instr_create(b
->shader
, p
- srcs
);
2761 memcpy(instr
->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
2763 instr
->coord_components
= coord_components
;
2764 instr
->sampler_dim
= sampler_dim
;
2765 instr
->is_array
= is_array
;
2766 instr
->is_shadow
= is_shadow
;
2767 instr
->is_new_style_shadow
=
2768 is_shadow
&& glsl_get_components(ret_type
->type
) == 1;
2769 instr
->component
= gather_component
;
2771 /* The Vulkan spec says:
2773 * "If an instruction loads from or stores to a resource (including
2774 * atomics and image instructions) and the resource descriptor being
2775 * accessed is not dynamically uniform, then the operand corresponding
2776 * to that resource (e.g. the pointer or sampled image operand) must be
2777 * decorated with NonUniform."
2779 * It's very careful to specify that the exact operand must be decorated
2780 * NonUniform. The SPIR-V parser is not expected to chase through long
2781 * chains to find the NonUniform decoration. It's either right there or we
2782 * can assume it doesn't exist.
2784 enum gl_access_qualifier access
= 0;
2785 vtn_foreach_decoration(b
, sampled_val
, non_uniform_decoration_cb
, &access
);
2787 if (image
&& (access
& ACCESS_NON_UNIFORM
))
2788 instr
->texture_non_uniform
= true;
2790 if (sampler
&& (access
& ACCESS_NON_UNIFORM
))
2791 instr
->sampler_non_uniform
= true;
2793 /* for non-query ops, get dest_type from sampler type */
2794 if (dest_type
== nir_type_invalid
) {
2795 switch (glsl_get_sampler_result_type(image
->type
)) {
2796 case GLSL_TYPE_FLOAT
: dest_type
= nir_type_float
; break;
2797 case GLSL_TYPE_INT
: dest_type
= nir_type_int
; break;
2798 case GLSL_TYPE_UINT
: dest_type
= nir_type_uint
; break;
2799 case GLSL_TYPE_BOOL
: dest_type
= nir_type_bool
; break;
2801 vtn_fail("Invalid base type for sampler result");
2805 instr
->dest_type
= dest_type
;
2807 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
2808 nir_tex_instr_dest_size(instr
), 32, NULL
);
2810 vtn_assert(glsl_get_vector_elements(ret_type
->type
) ==
2811 nir_tex_instr_dest_size(instr
));
2813 if (gather_offsets
) {
2814 vtn_fail_if(gather_offsets
->type
->base_type
!= vtn_base_type_array
||
2815 gather_offsets
->type
->length
!= 4,
2816 "ConstOffsets must be an array of size four of vectors "
2817 "of two integer components");
2819 struct vtn_type
*vec_type
= gather_offsets
->type
->array_element
;
2820 vtn_fail_if(vec_type
->base_type
!= vtn_base_type_vector
||
2821 vec_type
->length
!= 2 ||
2822 !glsl_type_is_integer(vec_type
->type
),
2823 "ConstOffsets must be an array of size four of vectors "
2824 "of two integer components");
2826 unsigned bit_size
= glsl_get_bit_size(vec_type
->type
);
2827 for (uint32_t i
= 0; i
< 4; i
++) {
2828 const nir_const_value
*cvec
=
2829 gather_offsets
->constant
->elements
[i
]->values
;
2830 for (uint32_t j
= 0; j
< 2; j
++) {
2832 case 8: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i8
; break;
2833 case 16: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i16
; break;
2834 case 32: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i32
; break;
2835 case 64: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i64
; break;
2837 vtn_fail("Unsupported bit size: %u", bit_size
);
2843 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
2845 vtn_push_nir_ssa(b
, w
[2], &instr
->dest
.ssa
);
2849 fill_common_atomic_sources(struct vtn_builder
*b
, SpvOp opcode
,
2850 const uint32_t *w
, nir_src
*src
)
2853 case SpvOpAtomicIIncrement
:
2854 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 1));
2857 case SpvOpAtomicIDecrement
:
2858 src
[0] = nir_src_for_ssa(nir_imm_int(&b
->nb
, -1));
2861 case SpvOpAtomicISub
:
2863 nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_get_nir_ssa(b
, w
[6])));
2866 case SpvOpAtomicCompareExchange
:
2867 case SpvOpAtomicCompareExchangeWeak
:
2868 src
[0] = nir_src_for_ssa(vtn_get_nir_ssa(b
, w
[8]));
2869 src
[1] = nir_src_for_ssa(vtn_get_nir_ssa(b
, w
[7]));
2872 case SpvOpAtomicExchange
:
2873 case SpvOpAtomicIAdd
:
2874 case SpvOpAtomicSMin
:
2875 case SpvOpAtomicUMin
:
2876 case SpvOpAtomicSMax
:
2877 case SpvOpAtomicUMax
:
2878 case SpvOpAtomicAnd
:
2880 case SpvOpAtomicXor
:
2881 case SpvOpAtomicFAddEXT
:
2882 src
[0] = nir_src_for_ssa(vtn_get_nir_ssa(b
, w
[6]));
2886 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
2890 static nir_ssa_def
*
2891 get_image_coord(struct vtn_builder
*b
, uint32_t value
)
2893 nir_ssa_def
*coord
= vtn_get_nir_ssa(b
, value
);
2895 /* The image_load_store intrinsics assume a 4-dim coordinate */
2896 unsigned swizzle
[4];
2897 for (unsigned i
= 0; i
< 4; i
++)
2898 swizzle
[i
] = MIN2(i
, coord
->num_components
- 1);
2900 return nir_swizzle(&b
->nb
, coord
, swizzle
, 4);
2903 static nir_ssa_def
*
2904 expand_to_vec4(nir_builder
*b
, nir_ssa_def
*value
)
2906 if (value
->num_components
== 4)
2910 for (unsigned i
= 0; i
< 4; i
++)
2911 swiz
[i
] = i
< value
->num_components
? i
: 0;
2912 return nir_swizzle(b
, value
, swiz
, 4);
2916 vtn_handle_image(struct vtn_builder
*b
, SpvOp opcode
,
2917 const uint32_t *w
, unsigned count
)
2919 /* Just get this one out of the way */
2920 if (opcode
== SpvOpImageTexelPointer
) {
2921 struct vtn_value
*val
=
2922 vtn_push_value(b
, w
[2], vtn_value_type_image_pointer
);
2923 val
->image
= ralloc(b
, struct vtn_image_pointer
);
2925 val
->image
->image
= vtn_nir_deref(b
, w
[3]);
2926 val
->image
->coord
= get_image_coord(b
, w
[4]);
2927 val
->image
->sample
= vtn_get_nir_ssa(b
, w
[5]);
2928 val
->image
->lod
= nir_imm_int(&b
->nb
, 0);
2932 struct vtn_image_pointer image
;
2933 SpvScope scope
= SpvScopeInvocation
;
2934 SpvMemorySemanticsMask semantics
= 0;
2936 enum gl_access_qualifier access
= 0;
2938 struct vtn_value
*res_val
;
2940 case SpvOpAtomicExchange
:
2941 case SpvOpAtomicCompareExchange
:
2942 case SpvOpAtomicCompareExchangeWeak
:
2943 case SpvOpAtomicIIncrement
:
2944 case SpvOpAtomicIDecrement
:
2945 case SpvOpAtomicIAdd
:
2946 case SpvOpAtomicISub
:
2947 case SpvOpAtomicLoad
:
2948 case SpvOpAtomicSMin
:
2949 case SpvOpAtomicUMin
:
2950 case SpvOpAtomicSMax
:
2951 case SpvOpAtomicUMax
:
2952 case SpvOpAtomicAnd
:
2954 case SpvOpAtomicXor
:
2955 case SpvOpAtomicFAddEXT
:
2956 res_val
= vtn_value(b
, w
[3], vtn_value_type_image_pointer
);
2957 image
= *res_val
->image
;
2958 scope
= vtn_constant_uint(b
, w
[4]);
2959 semantics
= vtn_constant_uint(b
, w
[5]);
2960 access
|= ACCESS_COHERENT
;
2963 case SpvOpAtomicStore
:
2964 res_val
= vtn_value(b
, w
[1], vtn_value_type_image_pointer
);
2965 image
= *res_val
->image
;
2966 scope
= vtn_constant_uint(b
, w
[2]);
2967 semantics
= vtn_constant_uint(b
, w
[3]);
2968 access
|= ACCESS_COHERENT
;
2971 case SpvOpImageQuerySize
:
2972 res_val
= vtn_untyped_value(b
, w
[3]);
2973 image
.image
= vtn_get_image(b
, w
[3]);
2975 image
.sample
= NULL
;
2979 case SpvOpImageRead
: {
2980 res_val
= vtn_untyped_value(b
, w
[3]);
2981 image
.image
= vtn_get_image(b
, w
[3]);
2982 image
.coord
= get_image_coord(b
, w
[4]);
2984 const SpvImageOperandsMask operands
=
2985 count
> 5 ? w
[5] : SpvImageOperandsMaskNone
;
2987 if (operands
& SpvImageOperandsSampleMask
) {
2988 uint32_t arg
= image_operand_arg(b
, w
, count
, 5,
2989 SpvImageOperandsSampleMask
);
2990 image
.sample
= vtn_get_nir_ssa(b
, w
[arg
]);
2992 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
2995 if (operands
& SpvImageOperandsMakeTexelVisibleMask
) {
2996 vtn_fail_if((operands
& SpvImageOperandsNonPrivateTexelMask
) == 0,
2997 "MakeTexelVisible requires NonPrivateTexel to also be set.");
2998 uint32_t arg
= image_operand_arg(b
, w
, count
, 5,
2999 SpvImageOperandsMakeTexelVisibleMask
);
3000 semantics
= SpvMemorySemanticsMakeVisibleMask
;
3001 scope
= vtn_constant_uint(b
, w
[arg
]);
3004 if (operands
& SpvImageOperandsLodMask
) {
3005 uint32_t arg
= image_operand_arg(b
, w
, count
, 5,
3006 SpvImageOperandsLodMask
);
3007 image
.lod
= vtn_get_nir_ssa(b
, w
[arg
]);
3009 image
.lod
= nir_imm_int(&b
->nb
, 0);
3012 /* TODO: Volatile. */
3017 case SpvOpImageWrite
: {
3018 res_val
= vtn_untyped_value(b
, w
[1]);
3019 image
.image
= vtn_get_image(b
, w
[1]);
3020 image
.coord
= get_image_coord(b
, w
[2]);
3024 const SpvImageOperandsMask operands
=
3025 count
> 4 ? w
[4] : SpvImageOperandsMaskNone
;
3027 if (operands
& SpvImageOperandsSampleMask
) {
3028 uint32_t arg
= image_operand_arg(b
, w
, count
, 4,
3029 SpvImageOperandsSampleMask
);
3030 image
.sample
= vtn_get_nir_ssa(b
, w
[arg
]);
3032 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
3035 if (operands
& SpvImageOperandsMakeTexelAvailableMask
) {
3036 vtn_fail_if((operands
& SpvImageOperandsNonPrivateTexelMask
) == 0,
3037 "MakeTexelAvailable requires NonPrivateTexel to also be set.");
3038 uint32_t arg
= image_operand_arg(b
, w
, count
, 4,
3039 SpvImageOperandsMakeTexelAvailableMask
);
3040 semantics
= SpvMemorySemanticsMakeAvailableMask
;
3041 scope
= vtn_constant_uint(b
, w
[arg
]);
3044 if (operands
& SpvImageOperandsLodMask
) {
3045 uint32_t arg
= image_operand_arg(b
, w
, count
, 4,
3046 SpvImageOperandsLodMask
);
3047 image
.lod
= vtn_get_nir_ssa(b
, w
[arg
]);
3049 image
.lod
= nir_imm_int(&b
->nb
, 0);
3052 /* TODO: Volatile. */
3058 vtn_fail_with_opcode("Invalid image opcode", opcode
);
3061 nir_intrinsic_op op
;
3063 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_deref_##N; break;
3064 OP(ImageQuerySize
, size
)
3066 OP(ImageWrite
, store
)
3067 OP(AtomicLoad
, load
)
3068 OP(AtomicStore
, store
)
3069 OP(AtomicExchange
, atomic_exchange
)
3070 OP(AtomicCompareExchange
, atomic_comp_swap
)
3071 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
3072 OP(AtomicIIncrement
, atomic_add
)
3073 OP(AtomicIDecrement
, atomic_add
)
3074 OP(AtomicIAdd
, atomic_add
)
3075 OP(AtomicISub
, atomic_add
)
3076 OP(AtomicSMin
, atomic_imin
)
3077 OP(AtomicUMin
, atomic_umin
)
3078 OP(AtomicSMax
, atomic_imax
)
3079 OP(AtomicUMax
, atomic_umax
)
3080 OP(AtomicAnd
, atomic_and
)
3081 OP(AtomicOr
, atomic_or
)
3082 OP(AtomicXor
, atomic_xor
)
3083 OP(AtomicFAddEXT
, atomic_fadd
)
3086 vtn_fail_with_opcode("Invalid image opcode", opcode
);
3089 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
3091 intrin
->src
[0] = nir_src_for_ssa(&image
.image
->dest
.ssa
);
3093 /* ImageQuerySize doesn't take any extra parameters */
3094 if (opcode
!= SpvOpImageQuerySize
) {
3095 /* The image coordinate is always 4 components but we may not have that
3096 * many. Swizzle to compensate.
3098 intrin
->src
[1] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, image
.coord
));
3099 intrin
->src
[2] = nir_src_for_ssa(image
.sample
);
3102 /* The Vulkan spec says:
3104 * "If an instruction loads from or stores to a resource (including
3105 * atomics and image instructions) and the resource descriptor being
3106 * accessed is not dynamically uniform, then the operand corresponding
3107 * to that resource (e.g. the pointer or sampled image operand) must be
3108 * decorated with NonUniform."
3110 * It's very careful to specify that the exact operand must be decorated
3111 * NonUniform. The SPIR-V parser is not expected to chase through long
3112 * chains to find the NonUniform decoration. It's either right there or we
3113 * can assume it doesn't exist.
3115 vtn_foreach_decoration(b
, res_val
, non_uniform_decoration_cb
, &access
);
3116 nir_intrinsic_set_access(intrin
, access
);
3119 case SpvOpAtomicLoad
:
3120 case SpvOpImageQuerySize
:
3121 case SpvOpImageRead
:
3122 if (opcode
== SpvOpImageRead
|| opcode
== SpvOpAtomicLoad
) {
3123 /* Only OpImageRead can support a lod parameter if
3124 * SPV_AMD_shader_image_load_store_lod is used but the current NIR
3125 * intrinsics definition for atomics requires us to set it for
3128 intrin
->src
[3] = nir_src_for_ssa(image
.lod
);
3131 case SpvOpAtomicStore
:
3132 case SpvOpImageWrite
: {
3133 const uint32_t value_id
= opcode
== SpvOpAtomicStore
? w
[4] : w
[3];
3134 nir_ssa_def
*value
= vtn_get_nir_ssa(b
, value_id
);
3135 /* nir_intrinsic_image_deref_store always takes a vec4 value */
3136 assert(op
== nir_intrinsic_image_deref_store
);
3137 intrin
->num_components
= 4;
3138 intrin
->src
[3] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, value
));
3139 /* Only OpImageWrite can support a lod parameter if
3140 * SPV_AMD_shader_image_load_store_lod is used but the current NIR
3141 * intrinsics definition for atomics requires us to set it for
3144 intrin
->src
[4] = nir_src_for_ssa(image
.lod
);
3148 case SpvOpAtomicCompareExchange
:
3149 case SpvOpAtomicCompareExchangeWeak
:
3150 case SpvOpAtomicIIncrement
:
3151 case SpvOpAtomicIDecrement
:
3152 case SpvOpAtomicExchange
:
3153 case SpvOpAtomicIAdd
:
3154 case SpvOpAtomicISub
:
3155 case SpvOpAtomicSMin
:
3156 case SpvOpAtomicUMin
:
3157 case SpvOpAtomicSMax
:
3158 case SpvOpAtomicUMax
:
3159 case SpvOpAtomicAnd
:
3161 case SpvOpAtomicXor
:
3162 case SpvOpAtomicFAddEXT
:
3163 fill_common_atomic_sources(b
, opcode
, w
, &intrin
->src
[3]);
3167 vtn_fail_with_opcode("Invalid image opcode", opcode
);
3170 /* Image operations implicitly have the Image storage memory semantics. */
3171 semantics
|= SpvMemorySemanticsImageMemoryMask
;
3173 SpvMemorySemanticsMask before_semantics
;
3174 SpvMemorySemanticsMask after_semantics
;
3175 vtn_split_barrier_semantics(b
, semantics
, &before_semantics
, &after_semantics
);
3177 if (before_semantics
)
3178 vtn_emit_memory_barrier(b
, scope
, before_semantics
);
3180 if (opcode
!= SpvOpImageWrite
&& opcode
!= SpvOpAtomicStore
) {
3181 struct vtn_type
*type
= vtn_get_type(b
, w
[1]);
3183 unsigned dest_components
= glsl_get_vector_elements(type
->type
);
3184 if (nir_intrinsic_infos
[op
].dest_components
== 0)
3185 intrin
->num_components
= dest_components
;
3187 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
,
3188 nir_intrinsic_dest_components(intrin
), 32, NULL
);
3190 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3192 nir_ssa_def
*result
= &intrin
->dest
.ssa
;
3193 if (nir_intrinsic_dest_components(intrin
) != dest_components
)
3194 result
= nir_channels(&b
->nb
, result
, (1 << dest_components
) - 1);
3196 vtn_push_nir_ssa(b
, w
[2], result
);
3198 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3201 if (after_semantics
)
3202 vtn_emit_memory_barrier(b
, scope
, after_semantics
);
3205 static nir_intrinsic_op
3206 get_ssbo_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
3209 case SpvOpAtomicLoad
: return nir_intrinsic_load_ssbo
;
3210 case SpvOpAtomicStore
: return nir_intrinsic_store_ssbo
;
3211 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
3212 OP(AtomicExchange
, atomic_exchange
)
3213 OP(AtomicCompareExchange
, atomic_comp_swap
)
3214 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
3215 OP(AtomicIIncrement
, atomic_add
)
3216 OP(AtomicIDecrement
, atomic_add
)
3217 OP(AtomicIAdd
, atomic_add
)
3218 OP(AtomicISub
, atomic_add
)
3219 OP(AtomicSMin
, atomic_imin
)
3220 OP(AtomicUMin
, atomic_umin
)
3221 OP(AtomicSMax
, atomic_imax
)
3222 OP(AtomicUMax
, atomic_umax
)
3223 OP(AtomicAnd
, atomic_and
)
3224 OP(AtomicOr
, atomic_or
)
3225 OP(AtomicXor
, atomic_xor
)
3226 OP(AtomicFAddEXT
, atomic_fadd
)
3229 vtn_fail_with_opcode("Invalid SSBO atomic", opcode
);
3233 static nir_intrinsic_op
3234 get_uniform_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
3237 #define OP(S, N) case SpvOp##S: return nir_intrinsic_atomic_counter_ ##N;
3238 OP(AtomicLoad
, read_deref
)
3239 OP(AtomicExchange
, exchange
)
3240 OP(AtomicCompareExchange
, comp_swap
)
3241 OP(AtomicCompareExchangeWeak
, comp_swap
)
3242 OP(AtomicIIncrement
, inc_deref
)
3243 OP(AtomicIDecrement
, post_dec_deref
)
3244 OP(AtomicIAdd
, add_deref
)
3245 OP(AtomicISub
, add_deref
)
3246 OP(AtomicUMin
, min_deref
)
3247 OP(AtomicUMax
, max_deref
)
3248 OP(AtomicAnd
, and_deref
)
3249 OP(AtomicOr
, or_deref
)
3250 OP(AtomicXor
, xor_deref
)
3253 /* We left the following out: AtomicStore, AtomicSMin and
3254 * AtomicSmax. Right now there are not nir intrinsics for them. At this
3255 * moment Atomic Counter support is needed for ARB_spirv support, so is
3256 * only need to support GLSL Atomic Counters that are uints and don't
3257 * allow direct storage.
3259 vtn_fail("Invalid uniform atomic");
3263 static nir_intrinsic_op
3264 get_deref_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
3267 case SpvOpAtomicLoad
: return nir_intrinsic_load_deref
;
3268 case SpvOpAtomicStore
: return nir_intrinsic_store_deref
;
3269 #define OP(S, N) case SpvOp##S: return nir_intrinsic_deref_##N;
3270 OP(AtomicExchange
, atomic_exchange
)
3271 OP(AtomicCompareExchange
, atomic_comp_swap
)
3272 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
3273 OP(AtomicIIncrement
, atomic_add
)
3274 OP(AtomicIDecrement
, atomic_add
)
3275 OP(AtomicIAdd
, atomic_add
)
3276 OP(AtomicISub
, atomic_add
)
3277 OP(AtomicSMin
, atomic_imin
)
3278 OP(AtomicUMin
, atomic_umin
)
3279 OP(AtomicSMax
, atomic_imax
)
3280 OP(AtomicUMax
, atomic_umax
)
3281 OP(AtomicAnd
, atomic_and
)
3282 OP(AtomicOr
, atomic_or
)
3283 OP(AtomicXor
, atomic_xor
)
3284 OP(AtomicFAddEXT
, atomic_fadd
)
3287 vtn_fail_with_opcode("Invalid shared atomic", opcode
);
3292 * Handles shared atomics, ssbo atomics and atomic counters.
3295 vtn_handle_atomics(struct vtn_builder
*b
, SpvOp opcode
,
3296 const uint32_t *w
, UNUSED
unsigned count
)
3298 struct vtn_pointer
*ptr
;
3299 nir_intrinsic_instr
*atomic
;
3301 SpvScope scope
= SpvScopeInvocation
;
3302 SpvMemorySemanticsMask semantics
= 0;
3305 case SpvOpAtomicLoad
:
3306 case SpvOpAtomicExchange
:
3307 case SpvOpAtomicCompareExchange
:
3308 case SpvOpAtomicCompareExchangeWeak
:
3309 case SpvOpAtomicIIncrement
:
3310 case SpvOpAtomicIDecrement
:
3311 case SpvOpAtomicIAdd
:
3312 case SpvOpAtomicISub
:
3313 case SpvOpAtomicSMin
:
3314 case SpvOpAtomicUMin
:
3315 case SpvOpAtomicSMax
:
3316 case SpvOpAtomicUMax
:
3317 case SpvOpAtomicAnd
:
3319 case SpvOpAtomicXor
:
3320 case SpvOpAtomicFAddEXT
:
3321 ptr
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
3322 scope
= vtn_constant_uint(b
, w
[4]);
3323 semantics
= vtn_constant_uint(b
, w
[5]);
3326 case SpvOpAtomicStore
:
3327 ptr
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
3328 scope
= vtn_constant_uint(b
, w
[2]);
3329 semantics
= vtn_constant_uint(b
, w
[3]);
3333 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
3336 /* uniform as "atomic counter uniform" */
3337 if (ptr
->mode
== vtn_variable_mode_atomic_counter
) {
3338 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
3339 nir_intrinsic_op op
= get_uniform_nir_atomic_op(b
, opcode
);
3340 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
3341 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
3343 /* SSBO needs to initialize index/offset. In this case we don't need to,
3344 * as that info is already stored on the ptr->var->var nir_variable (see
3345 * vtn_create_variable)
3349 case SpvOpAtomicLoad
:
3350 case SpvOpAtomicExchange
:
3351 case SpvOpAtomicCompareExchange
:
3352 case SpvOpAtomicCompareExchangeWeak
:
3353 case SpvOpAtomicIIncrement
:
3354 case SpvOpAtomicIDecrement
:
3355 case SpvOpAtomicIAdd
:
3356 case SpvOpAtomicISub
:
3357 case SpvOpAtomicSMin
:
3358 case SpvOpAtomicUMin
:
3359 case SpvOpAtomicSMax
:
3360 case SpvOpAtomicUMax
:
3361 case SpvOpAtomicAnd
:
3363 case SpvOpAtomicXor
:
3364 /* Nothing: we don't need to call fill_common_atomic_sources here, as
3365 * atomic counter uniforms doesn't have sources
3370 unreachable("Invalid SPIR-V atomic");
3373 } else if (vtn_pointer_uses_ssa_offset(b
, ptr
)) {
3374 nir_ssa_def
*offset
, *index
;
3375 offset
= vtn_pointer_to_offset(b
, ptr
, &index
);
3377 assert(ptr
->mode
== vtn_variable_mode_ssbo
);
3379 nir_intrinsic_op op
= get_ssbo_nir_atomic_op(b
, opcode
);
3380 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
3382 nir_intrinsic_set_access(atomic
, ACCESS_COHERENT
);
3386 case SpvOpAtomicLoad
:
3387 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
3388 nir_intrinsic_set_align(atomic
, 4, 0);
3389 if (ptr
->mode
== vtn_variable_mode_ssbo
)
3390 atomic
->src
[src
++] = nir_src_for_ssa(index
);
3391 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
3394 case SpvOpAtomicStore
:
3395 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
3396 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
3397 nir_intrinsic_set_align(atomic
, 4, 0);
3398 atomic
->src
[src
++] = nir_src_for_ssa(vtn_get_nir_ssa(b
, w
[4]));
3399 if (ptr
->mode
== vtn_variable_mode_ssbo
)
3400 atomic
->src
[src
++] = nir_src_for_ssa(index
);
3401 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
3404 case SpvOpAtomicExchange
:
3405 case SpvOpAtomicCompareExchange
:
3406 case SpvOpAtomicCompareExchangeWeak
:
3407 case SpvOpAtomicIIncrement
:
3408 case SpvOpAtomicIDecrement
:
3409 case SpvOpAtomicIAdd
:
3410 case SpvOpAtomicISub
:
3411 case SpvOpAtomicSMin
:
3412 case SpvOpAtomicUMin
:
3413 case SpvOpAtomicSMax
:
3414 case SpvOpAtomicUMax
:
3415 case SpvOpAtomicAnd
:
3417 case SpvOpAtomicXor
:
3418 case SpvOpAtomicFAddEXT
:
3419 if (ptr
->mode
== vtn_variable_mode_ssbo
)
3420 atomic
->src
[src
++] = nir_src_for_ssa(index
);
3421 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
3422 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[src
]);
3426 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
3429 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
3430 const struct glsl_type
*deref_type
= deref
->type
;
3431 nir_intrinsic_op op
= get_deref_nir_atomic_op(b
, opcode
);
3432 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
3433 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
3435 if (ptr
->mode
!= vtn_variable_mode_workgroup
)
3436 nir_intrinsic_set_access(atomic
, ACCESS_COHERENT
);
3439 case SpvOpAtomicLoad
:
3440 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
3443 case SpvOpAtomicStore
:
3444 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
3445 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
3446 atomic
->src
[1] = nir_src_for_ssa(vtn_get_nir_ssa(b
, w
[4]));
3449 case SpvOpAtomicExchange
:
3450 case SpvOpAtomicCompareExchange
:
3451 case SpvOpAtomicCompareExchangeWeak
:
3452 case SpvOpAtomicIIncrement
:
3453 case SpvOpAtomicIDecrement
:
3454 case SpvOpAtomicIAdd
:
3455 case SpvOpAtomicISub
:
3456 case SpvOpAtomicSMin
:
3457 case SpvOpAtomicUMin
:
3458 case SpvOpAtomicSMax
:
3459 case SpvOpAtomicUMax
:
3460 case SpvOpAtomicAnd
:
3462 case SpvOpAtomicXor
:
3463 case SpvOpAtomicFAddEXT
:
3464 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[1]);
3468 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
3472 /* Atomic ordering operations will implicitly apply to the atomic operation
3473 * storage class, so include that too.
3475 semantics
|= vtn_storage_class_to_memory_semantics(ptr
->ptr_type
->storage_class
);
3477 SpvMemorySemanticsMask before_semantics
;
3478 SpvMemorySemanticsMask after_semantics
;
3479 vtn_split_barrier_semantics(b
, semantics
, &before_semantics
, &after_semantics
);
3481 if (before_semantics
)
3482 vtn_emit_memory_barrier(b
, scope
, before_semantics
);
3484 if (opcode
!= SpvOpAtomicStore
) {
3485 struct vtn_type
*type
= vtn_get_type(b
, w
[1]);
3487 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
,
3488 glsl_get_vector_elements(type
->type
),
3489 glsl_get_bit_size(type
->type
), NULL
);
3491 vtn_push_nir_ssa(b
, w
[2], &atomic
->dest
.ssa
);
3494 nir_builder_instr_insert(&b
->nb
, &atomic
->instr
);
3496 if (after_semantics
)
3497 vtn_emit_memory_barrier(b
, scope
, after_semantics
);
3500 static nir_alu_instr
*
3501 create_vec(struct vtn_builder
*b
, unsigned num_components
, unsigned bit_size
)
3503 nir_op op
= nir_op_vec(num_components
);
3504 nir_alu_instr
*vec
= nir_alu_instr_create(b
->shader
, op
);
3505 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, num_components
,
3507 vec
->dest
.write_mask
= (1 << num_components
) - 1;
3512 struct vtn_ssa_value
*
3513 vtn_ssa_transpose(struct vtn_builder
*b
, struct vtn_ssa_value
*src
)
3515 if (src
->transposed
)
3516 return src
->transposed
;
3518 struct vtn_ssa_value
*dest
=
3519 vtn_create_ssa_value(b
, glsl_transposed_type(src
->type
));
3521 for (unsigned i
= 0; i
< glsl_get_matrix_columns(dest
->type
); i
++) {
3522 nir_alu_instr
*vec
= create_vec(b
, glsl_get_matrix_columns(src
->type
),
3523 glsl_get_bit_size(src
->type
));
3524 if (glsl_type_is_vector_or_scalar(src
->type
)) {
3525 vec
->src
[0].src
= nir_src_for_ssa(src
->def
);
3526 vec
->src
[0].swizzle
[0] = i
;
3528 for (unsigned j
= 0; j
< glsl_get_matrix_columns(src
->type
); j
++) {
3529 vec
->src
[j
].src
= nir_src_for_ssa(src
->elems
[j
]->def
);
3530 vec
->src
[j
].swizzle
[0] = i
;
3533 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3534 dest
->elems
[i
]->def
= &vec
->dest
.dest
.ssa
;
3537 dest
->transposed
= src
;
3542 static nir_ssa_def
*
3543 vtn_vector_shuffle(struct vtn_builder
*b
, unsigned num_components
,
3544 nir_ssa_def
*src0
, nir_ssa_def
*src1
,
3545 const uint32_t *indices
)
3547 nir_alu_instr
*vec
= create_vec(b
, num_components
, src0
->bit_size
);
3549 for (unsigned i
= 0; i
< num_components
; i
++) {
3550 uint32_t index
= indices
[i
];
3551 if (index
== 0xffffffff) {
3553 nir_src_for_ssa(nir_ssa_undef(&b
->nb
, 1, src0
->bit_size
));
3554 } else if (index
< src0
->num_components
) {
3555 vec
->src
[i
].src
= nir_src_for_ssa(src0
);
3556 vec
->src
[i
].swizzle
[0] = index
;
3558 vec
->src
[i
].src
= nir_src_for_ssa(src1
);
3559 vec
->src
[i
].swizzle
[0] = index
- src0
->num_components
;
3563 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3565 return &vec
->dest
.dest
.ssa
;
3569 * Concatentates a number of vectors/scalars together to produce a vector
3571 static nir_ssa_def
*
3572 vtn_vector_construct(struct vtn_builder
*b
, unsigned num_components
,
3573 unsigned num_srcs
, nir_ssa_def
**srcs
)
3575 nir_alu_instr
*vec
= create_vec(b
, num_components
, srcs
[0]->bit_size
);
3577 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
3579 * "When constructing a vector, there must be at least two Constituent
3582 vtn_assert(num_srcs
>= 2);
3584 unsigned dest_idx
= 0;
3585 for (unsigned i
= 0; i
< num_srcs
; i
++) {
3586 nir_ssa_def
*src
= srcs
[i
];
3587 vtn_assert(dest_idx
+ src
->num_components
<= num_components
);
3588 for (unsigned j
= 0; j
< src
->num_components
; j
++) {
3589 vec
->src
[dest_idx
].src
= nir_src_for_ssa(src
);
3590 vec
->src
[dest_idx
].swizzle
[0] = j
;
3595 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
3597 * "When constructing a vector, the total number of components in all
3598 * the operands must equal the number of components in Result Type."
3600 vtn_assert(dest_idx
== num_components
);
3602 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3604 return &vec
->dest
.dest
.ssa
;
3607 static struct vtn_ssa_value
*
3608 vtn_composite_copy(void *mem_ctx
, struct vtn_ssa_value
*src
)
3610 struct vtn_ssa_value
*dest
= rzalloc(mem_ctx
, struct vtn_ssa_value
);
3611 dest
->type
= src
->type
;
3613 if (glsl_type_is_vector_or_scalar(src
->type
)) {
3614 dest
->def
= src
->def
;
3616 unsigned elems
= glsl_get_length(src
->type
);
3618 dest
->elems
= ralloc_array(mem_ctx
, struct vtn_ssa_value
*, elems
);
3619 for (unsigned i
= 0; i
< elems
; i
++)
3620 dest
->elems
[i
] = vtn_composite_copy(mem_ctx
, src
->elems
[i
]);
3626 static struct vtn_ssa_value
*
3627 vtn_composite_insert(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
3628 struct vtn_ssa_value
*insert
, const uint32_t *indices
,
3629 unsigned num_indices
)
3631 struct vtn_ssa_value
*dest
= vtn_composite_copy(b
, src
);
3633 struct vtn_ssa_value
*cur
= dest
;
3635 for (i
= 0; i
< num_indices
- 1; i
++) {
3636 /* If we got a vector here, that means the next index will be trying to
3637 * dereference a scalar.
3639 vtn_fail_if(glsl_type_is_vector_or_scalar(cur
->type
),
3640 "OpCompositeInsert has too many indices.");
3641 vtn_fail_if(indices
[i
] >= glsl_get_length(cur
->type
),
3642 "All indices in an OpCompositeInsert must be in-bounds");
3643 cur
= cur
->elems
[indices
[i
]];
3646 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
3647 vtn_fail_if(indices
[i
] >= glsl_get_vector_elements(cur
->type
),
3648 "All indices in an OpCompositeInsert must be in-bounds");
3650 /* According to the SPIR-V spec, OpCompositeInsert may work down to
3651 * the component granularity. In that case, the last index will be
3652 * the index to insert the scalar into the vector.
3655 cur
->def
= nir_vector_insert_imm(&b
->nb
, cur
->def
, insert
->def
, indices
[i
]);
3657 vtn_fail_if(indices
[i
] >= glsl_get_length(cur
->type
),
3658 "All indices in an OpCompositeInsert must be in-bounds");
3659 cur
->elems
[indices
[i
]] = insert
;
3665 static struct vtn_ssa_value
*
3666 vtn_composite_extract(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
3667 const uint32_t *indices
, unsigned num_indices
)
3669 struct vtn_ssa_value
*cur
= src
;
3670 for (unsigned i
= 0; i
< num_indices
; i
++) {
3671 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
3672 vtn_assert(i
== num_indices
- 1);
3673 vtn_fail_if(indices
[i
] >= glsl_get_vector_elements(cur
->type
),
3674 "All indices in an OpCompositeExtract must be in-bounds");
3676 /* According to the SPIR-V spec, OpCompositeExtract may work down to
3677 * the component granularity. The last index will be the index of the
3678 * vector to extract.
3681 const struct glsl_type
*scalar_type
=
3682 glsl_scalar_type(glsl_get_base_type(cur
->type
));
3683 struct vtn_ssa_value
*ret
= vtn_create_ssa_value(b
, scalar_type
);
3684 ret
->def
= nir_channel(&b
->nb
, cur
->def
, indices
[i
]);
3687 vtn_fail_if(indices
[i
] >= glsl_get_length(cur
->type
),
3688 "All indices in an OpCompositeExtract must be in-bounds");
3689 cur
= cur
->elems
[indices
[i
]];
3697 vtn_handle_composite(struct vtn_builder
*b
, SpvOp opcode
,
3698 const uint32_t *w
, unsigned count
)
3700 struct vtn_type
*type
= vtn_get_type(b
, w
[1]);
3701 struct vtn_ssa_value
*ssa
= vtn_create_ssa_value(b
, type
->type
);
3704 case SpvOpVectorExtractDynamic
:
3705 ssa
->def
= nir_vector_extract(&b
->nb
, vtn_get_nir_ssa(b
, w
[3]),
3706 vtn_get_nir_ssa(b
, w
[4]));
3709 case SpvOpVectorInsertDynamic
:
3710 ssa
->def
= nir_vector_insert(&b
->nb
, vtn_get_nir_ssa(b
, w
[3]),
3711 vtn_get_nir_ssa(b
, w
[4]),
3712 vtn_get_nir_ssa(b
, w
[5]));
3715 case SpvOpVectorShuffle
:
3716 ssa
->def
= vtn_vector_shuffle(b
, glsl_get_vector_elements(type
->type
),
3717 vtn_get_nir_ssa(b
, w
[3]),
3718 vtn_get_nir_ssa(b
, w
[4]),
3722 case SpvOpCompositeConstruct
: {
3723 unsigned elems
= count
- 3;
3725 if (glsl_type_is_vector_or_scalar(type
->type
)) {
3726 nir_ssa_def
*srcs
[NIR_MAX_VEC_COMPONENTS
];
3727 for (unsigned i
= 0; i
< elems
; i
++)
3728 srcs
[i
] = vtn_get_nir_ssa(b
, w
[3 + i
]);
3730 vtn_vector_construct(b
, glsl_get_vector_elements(type
->type
),
3733 ssa
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
3734 for (unsigned i
= 0; i
< elems
; i
++)
3735 ssa
->elems
[i
] = vtn_ssa_value(b
, w
[3 + i
]);
3739 case SpvOpCompositeExtract
:
3740 ssa
= vtn_composite_extract(b
, vtn_ssa_value(b
, w
[3]),
3744 case SpvOpCompositeInsert
:
3745 ssa
= vtn_composite_insert(b
, vtn_ssa_value(b
, w
[4]),
3746 vtn_ssa_value(b
, w
[3]),
3750 case SpvOpCopyLogical
:
3751 ssa
= vtn_composite_copy(b
, vtn_ssa_value(b
, w
[3]));
3753 case SpvOpCopyObject
:
3754 vtn_copy_value(b
, w
[3], w
[2]);
3758 vtn_fail_with_opcode("unknown composite operation", opcode
);
3761 vtn_push_ssa_value(b
, w
[2], ssa
);
3765 vtn_emit_barrier(struct vtn_builder
*b
, nir_intrinsic_op op
)
3767 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
3768 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3772 vtn_emit_memory_barrier(struct vtn_builder
*b
, SpvScope scope
,
3773 SpvMemorySemanticsMask semantics
)
3775 if (b
->shader
->options
->use_scoped_barrier
) {
3776 vtn_emit_scoped_memory_barrier(b
, scope
, semantics
);
3780 static const SpvMemorySemanticsMask all_memory_semantics
=
3781 SpvMemorySemanticsUniformMemoryMask
|
3782 SpvMemorySemanticsWorkgroupMemoryMask
|
3783 SpvMemorySemanticsAtomicCounterMemoryMask
|
3784 SpvMemorySemanticsImageMemoryMask
|
3785 SpvMemorySemanticsOutputMemoryMask
;
3787 /* If we're not actually doing a memory barrier, bail */
3788 if (!(semantics
& all_memory_semantics
))
3791 /* GL and Vulkan don't have these */
3792 vtn_assert(scope
!= SpvScopeCrossDevice
);
3794 if (scope
== SpvScopeSubgroup
)
3795 return; /* Nothing to do here */
3797 if (scope
== SpvScopeWorkgroup
) {
3798 vtn_emit_barrier(b
, nir_intrinsic_group_memory_barrier
);
3802 /* There's only two scopes thing left */
3803 vtn_assert(scope
== SpvScopeInvocation
|| scope
== SpvScopeDevice
);
3805 /* Map the GLSL memoryBarrier() construct and any barriers with more than one
3806 * semantic to the corresponding NIR one.
3808 if (util_bitcount(semantics
& all_memory_semantics
) > 1) {
3809 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier
);
3810 if (semantics
& SpvMemorySemanticsOutputMemoryMask
) {
3811 /* GLSL memoryBarrier() (and the corresponding NIR one) doesn't include
3812 * TCS outputs, so we have to emit it's own intrinsic for that. We
3813 * then need to emit another memory_barrier to prevent moving
3814 * non-output operations to before the tcs_patch barrier.
3816 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_tcs_patch
);
3817 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier
);
3822 /* Issue a more specific barrier */
3823 switch (semantics
& all_memory_semantics
) {
3824 case SpvMemorySemanticsUniformMemoryMask
:
3825 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_buffer
);
3827 case SpvMemorySemanticsWorkgroupMemoryMask
:
3828 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_shared
);
3830 case SpvMemorySemanticsAtomicCounterMemoryMask
:
3831 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_atomic_counter
);
3833 case SpvMemorySemanticsImageMemoryMask
:
3834 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_image
);
3836 case SpvMemorySemanticsOutputMemoryMask
:
3837 if (b
->nb
.shader
->info
.stage
== MESA_SHADER_TESS_CTRL
)
3838 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_tcs_patch
);
3846 vtn_handle_barrier(struct vtn_builder
*b
, SpvOp opcode
,
3847 const uint32_t *w
, UNUSED
unsigned count
)
3850 case SpvOpEmitVertex
:
3851 case SpvOpEmitStreamVertex
:
3852 case SpvOpEndPrimitive
:
3853 case SpvOpEndStreamPrimitive
: {
3854 nir_intrinsic_op intrinsic_op
;
3856 case SpvOpEmitVertex
:
3857 case SpvOpEmitStreamVertex
:
3858 intrinsic_op
= nir_intrinsic_emit_vertex
;
3860 case SpvOpEndPrimitive
:
3861 case SpvOpEndStreamPrimitive
:
3862 intrinsic_op
= nir_intrinsic_end_primitive
;
3865 unreachable("Invalid opcode");
3868 nir_intrinsic_instr
*intrin
=
3869 nir_intrinsic_instr_create(b
->shader
, intrinsic_op
);
3872 case SpvOpEmitStreamVertex
:
3873 case SpvOpEndStreamPrimitive
: {
3874 unsigned stream
= vtn_constant_uint(b
, w
[1]);
3875 nir_intrinsic_set_stream_id(intrin
, stream
);
3883 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3887 case SpvOpMemoryBarrier
: {
3888 SpvScope scope
= vtn_constant_uint(b
, w
[1]);
3889 SpvMemorySemanticsMask semantics
= vtn_constant_uint(b
, w
[2]);
3890 vtn_emit_memory_barrier(b
, scope
, semantics
);
3894 case SpvOpControlBarrier
: {
3895 SpvScope execution_scope
= vtn_constant_uint(b
, w
[1]);
3896 SpvScope memory_scope
= vtn_constant_uint(b
, w
[2]);
3897 SpvMemorySemanticsMask memory_semantics
= vtn_constant_uint(b
, w
[3]);
3899 /* GLSLang, prior to commit 8297936dd6eb3, emitted OpControlBarrier with
3900 * memory semantics of None for GLSL barrier().
3901 * And before that, prior to c3f1cdfa, emitted the OpControlBarrier with
3902 * Device instead of Workgroup for execution scope.
3904 if (b
->wa_glslang_cs_barrier
&&
3905 b
->nb
.shader
->info
.stage
== MESA_SHADER_COMPUTE
&&
3906 (execution_scope
== SpvScopeWorkgroup
||
3907 execution_scope
== SpvScopeDevice
) &&
3908 memory_semantics
== SpvMemorySemanticsMaskNone
) {
3909 execution_scope
= SpvScopeWorkgroup
;
3910 memory_scope
= SpvScopeWorkgroup
;
3911 memory_semantics
= SpvMemorySemanticsAcquireReleaseMask
|
3912 SpvMemorySemanticsWorkgroupMemoryMask
;
3915 /* From the SPIR-V spec:
3917 * "When used with the TessellationControl execution model, it also
3918 * implicitly synchronizes the Output Storage Class: Writes to Output
3919 * variables performed by any invocation executed prior to a
3920 * OpControlBarrier will be visible to any other invocation after
3921 * return from that OpControlBarrier."
3923 if (b
->nb
.shader
->info
.stage
== MESA_SHADER_TESS_CTRL
) {
3924 memory_semantics
&= ~(SpvMemorySemanticsAcquireMask
|
3925 SpvMemorySemanticsReleaseMask
|
3926 SpvMemorySemanticsAcquireReleaseMask
|
3927 SpvMemorySemanticsSequentiallyConsistentMask
);
3928 memory_semantics
|= SpvMemorySemanticsAcquireReleaseMask
|
3929 SpvMemorySemanticsOutputMemoryMask
;
3932 if (b
->shader
->options
->use_scoped_barrier
) {
3933 vtn_emit_scoped_control_barrier(b
, execution_scope
, memory_scope
,
3936 vtn_emit_memory_barrier(b
, memory_scope
, memory_semantics
);
3938 if (execution_scope
== SpvScopeWorkgroup
)
3939 vtn_emit_barrier(b
, nir_intrinsic_control_barrier
);
3945 unreachable("unknown barrier instruction");
3950 gl_primitive_from_spv_execution_mode(struct vtn_builder
*b
,
3951 SpvExecutionMode mode
)
3954 case SpvExecutionModeInputPoints
:
3955 case SpvExecutionModeOutputPoints
:
3956 return 0; /* GL_POINTS */
3957 case SpvExecutionModeInputLines
:
3958 return 1; /* GL_LINES */
3959 case SpvExecutionModeInputLinesAdjacency
:
3960 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
3961 case SpvExecutionModeTriangles
:
3962 return 4; /* GL_TRIANGLES */
3963 case SpvExecutionModeInputTrianglesAdjacency
:
3964 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
3965 case SpvExecutionModeQuads
:
3966 return 7; /* GL_QUADS */
3967 case SpvExecutionModeIsolines
:
3968 return 0x8E7A; /* GL_ISOLINES */
3969 case SpvExecutionModeOutputLineStrip
:
3970 return 3; /* GL_LINE_STRIP */
3971 case SpvExecutionModeOutputTriangleStrip
:
3972 return 5; /* GL_TRIANGLE_STRIP */
3974 vtn_fail("Invalid primitive type: %s (%u)",
3975 spirv_executionmode_to_string(mode
), mode
);
3980 vertices_in_from_spv_execution_mode(struct vtn_builder
*b
,
3981 SpvExecutionMode mode
)
3984 case SpvExecutionModeInputPoints
:
3986 case SpvExecutionModeInputLines
:
3988 case SpvExecutionModeInputLinesAdjacency
:
3990 case SpvExecutionModeTriangles
:
3992 case SpvExecutionModeInputTrianglesAdjacency
:
3995 vtn_fail("Invalid GS input mode: %s (%u)",
3996 spirv_executionmode_to_string(mode
), mode
);
4000 static gl_shader_stage
4001 stage_for_execution_model(struct vtn_builder
*b
, SpvExecutionModel model
)
4004 case SpvExecutionModelVertex
:
4005 return MESA_SHADER_VERTEX
;
4006 case SpvExecutionModelTessellationControl
:
4007 return MESA_SHADER_TESS_CTRL
;
4008 case SpvExecutionModelTessellationEvaluation
:
4009 return MESA_SHADER_TESS_EVAL
;
4010 case SpvExecutionModelGeometry
:
4011 return MESA_SHADER_GEOMETRY
;
4012 case SpvExecutionModelFragment
:
4013 return MESA_SHADER_FRAGMENT
;
4014 case SpvExecutionModelGLCompute
:
4015 return MESA_SHADER_COMPUTE
;
4016 case SpvExecutionModelKernel
:
4017 return MESA_SHADER_KERNEL
;
4019 vtn_fail("Unsupported execution model: %s (%u)",
4020 spirv_executionmodel_to_string(model
), model
);
4024 #define spv_check_supported(name, cap) do { \
4025 if (!(b->options && b->options->caps.name)) \
4026 vtn_warn("Unsupported SPIR-V capability: %s (%u)", \
4027 spirv_capability_to_string(cap), cap); \
4032 vtn_handle_entry_point(struct vtn_builder
*b
, const uint32_t *w
,
4035 struct vtn_value
*entry_point
= &b
->values
[w
[2]];
4036 /* Let this be a name label regardless */
4037 unsigned name_words
;
4038 entry_point
->name
= vtn_string_literal(b
, &w
[3], count
- 3, &name_words
);
4040 if (strcmp(entry_point
->name
, b
->entry_point_name
) != 0 ||
4041 stage_for_execution_model(b
, w
[1]) != b
->entry_point_stage
)
4044 vtn_assert(b
->entry_point
== NULL
);
4045 b
->entry_point
= entry_point
;
4049 vtn_handle_preamble_instruction(struct vtn_builder
*b
, SpvOp opcode
,
4050 const uint32_t *w
, unsigned count
)
4057 case SpvSourceLanguageUnknown
: lang
= "unknown"; break;
4058 case SpvSourceLanguageESSL
: lang
= "ESSL"; break;
4059 case SpvSourceLanguageGLSL
: lang
= "GLSL"; break;
4060 case SpvSourceLanguageOpenCL_C
: lang
= "OpenCL C"; break;
4061 case SpvSourceLanguageOpenCL_CPP
: lang
= "OpenCL C++"; break;
4062 case SpvSourceLanguageHLSL
: lang
= "HLSL"; break;
4065 uint32_t version
= w
[2];
4068 (count
> 3) ? vtn_value(b
, w
[3], vtn_value_type_string
)->str
: "";
4070 vtn_info("Parsing SPIR-V from %s %u source file %s", lang
, version
, file
);
4074 case SpvOpSourceExtension
:
4075 case SpvOpSourceContinued
:
4076 case SpvOpExtension
:
4077 case SpvOpModuleProcessed
:
4078 /* Unhandled, but these are for debug so that's ok. */
4081 case SpvOpCapability
: {
4082 SpvCapability cap
= w
[1];
4084 case SpvCapabilityMatrix
:
4085 case SpvCapabilityShader
:
4086 case SpvCapabilityGeometry
:
4087 case SpvCapabilityGeometryPointSize
:
4088 case SpvCapabilityUniformBufferArrayDynamicIndexing
:
4089 case SpvCapabilitySampledImageArrayDynamicIndexing
:
4090 case SpvCapabilityStorageBufferArrayDynamicIndexing
:
4091 case SpvCapabilityStorageImageArrayDynamicIndexing
:
4092 case SpvCapabilityImageRect
:
4093 case SpvCapabilitySampledRect
:
4094 case SpvCapabilitySampled1D
:
4095 case SpvCapabilityImage1D
:
4096 case SpvCapabilitySampledCubeArray
:
4097 case SpvCapabilityImageCubeArray
:
4098 case SpvCapabilitySampledBuffer
:
4099 case SpvCapabilityImageBuffer
:
4100 case SpvCapabilityImageQuery
:
4101 case SpvCapabilityDerivativeControl
:
4102 case SpvCapabilityInterpolationFunction
:
4103 case SpvCapabilityMultiViewport
:
4104 case SpvCapabilitySampleRateShading
:
4105 case SpvCapabilityClipDistance
:
4106 case SpvCapabilityCullDistance
:
4107 case SpvCapabilityInputAttachment
:
4108 case SpvCapabilityImageGatherExtended
:
4109 case SpvCapabilityStorageImageExtendedFormats
:
4110 case SpvCapabilityVector16
:
4113 case SpvCapabilityLinkage
:
4114 case SpvCapabilityFloat16Buffer
:
4115 case SpvCapabilitySparseResidency
:
4116 vtn_warn("Unsupported SPIR-V capability: %s",
4117 spirv_capability_to_string(cap
));
4120 case SpvCapabilityMinLod
:
4121 spv_check_supported(min_lod
, cap
);
4124 case SpvCapabilityAtomicStorage
:
4125 spv_check_supported(atomic_storage
, cap
);
4128 case SpvCapabilityFloat64
:
4129 spv_check_supported(float64
, cap
);
4131 case SpvCapabilityInt64
:
4132 spv_check_supported(int64
, cap
);
4134 case SpvCapabilityInt16
:
4135 spv_check_supported(int16
, cap
);
4137 case SpvCapabilityInt8
:
4138 spv_check_supported(int8
, cap
);
4141 case SpvCapabilityTransformFeedback
:
4142 spv_check_supported(transform_feedback
, cap
);
4145 case SpvCapabilityGeometryStreams
:
4146 spv_check_supported(geometry_streams
, cap
);
4149 case SpvCapabilityInt64Atomics
:
4150 spv_check_supported(int64_atomics
, cap
);
4153 case SpvCapabilityStorageImageMultisample
:
4154 spv_check_supported(storage_image_ms
, cap
);
4157 case SpvCapabilityAddresses
:
4158 spv_check_supported(address
, cap
);
4161 case SpvCapabilityKernel
:
4162 spv_check_supported(kernel
, cap
);
4165 case SpvCapabilityImageBasic
:
4166 case SpvCapabilityImageReadWrite
:
4167 case SpvCapabilityImageMipmap
:
4168 case SpvCapabilityPipes
:
4169 case SpvCapabilityDeviceEnqueue
:
4170 case SpvCapabilityLiteralSampler
:
4171 case SpvCapabilityGenericPointer
:
4172 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
4173 spirv_capability_to_string(cap
));
4176 case SpvCapabilityImageMSArray
:
4177 spv_check_supported(image_ms_array
, cap
);
4180 case SpvCapabilityTessellation
:
4181 case SpvCapabilityTessellationPointSize
:
4182 spv_check_supported(tessellation
, cap
);
4185 case SpvCapabilityDrawParameters
:
4186 spv_check_supported(draw_parameters
, cap
);
4189 case SpvCapabilityStorageImageReadWithoutFormat
:
4190 spv_check_supported(image_read_without_format
, cap
);
4193 case SpvCapabilityStorageImageWriteWithoutFormat
:
4194 spv_check_supported(image_write_without_format
, cap
);
4197 case SpvCapabilityDeviceGroup
:
4198 spv_check_supported(device_group
, cap
);
4201 case SpvCapabilityMultiView
:
4202 spv_check_supported(multiview
, cap
);
4205 case SpvCapabilityGroupNonUniform
:
4206 spv_check_supported(subgroup_basic
, cap
);
4209 case SpvCapabilitySubgroupVoteKHR
:
4210 case SpvCapabilityGroupNonUniformVote
:
4211 spv_check_supported(subgroup_vote
, cap
);
4214 case SpvCapabilitySubgroupBallotKHR
:
4215 case SpvCapabilityGroupNonUniformBallot
:
4216 spv_check_supported(subgroup_ballot
, cap
);
4219 case SpvCapabilityGroupNonUniformShuffle
:
4220 case SpvCapabilityGroupNonUniformShuffleRelative
:
4221 spv_check_supported(subgroup_shuffle
, cap
);
4224 case SpvCapabilityGroupNonUniformQuad
:
4225 spv_check_supported(subgroup_quad
, cap
);
4228 case SpvCapabilityGroupNonUniformArithmetic
:
4229 case SpvCapabilityGroupNonUniformClustered
:
4230 spv_check_supported(subgroup_arithmetic
, cap
);
4233 case SpvCapabilityGroups
:
4234 spv_check_supported(amd_shader_ballot
, cap
);
4237 case SpvCapabilityVariablePointersStorageBuffer
:
4238 case SpvCapabilityVariablePointers
:
4239 spv_check_supported(variable_pointers
, cap
);
4240 b
->variable_pointers
= true;
4243 case SpvCapabilityStorageUniformBufferBlock16
:
4244 case SpvCapabilityStorageUniform16
:
4245 case SpvCapabilityStoragePushConstant16
:
4246 case SpvCapabilityStorageInputOutput16
:
4247 spv_check_supported(storage_16bit
, cap
);
4250 case SpvCapabilityShaderLayer
:
4251 case SpvCapabilityShaderViewportIndex
:
4252 case SpvCapabilityShaderViewportIndexLayerEXT
:
4253 spv_check_supported(shader_viewport_index_layer
, cap
);
4256 case SpvCapabilityStorageBuffer8BitAccess
:
4257 case SpvCapabilityUniformAndStorageBuffer8BitAccess
:
4258 case SpvCapabilityStoragePushConstant8
:
4259 spv_check_supported(storage_8bit
, cap
);
4262 case SpvCapabilityShaderNonUniformEXT
:
4263 spv_check_supported(descriptor_indexing
, cap
);
4266 case SpvCapabilityInputAttachmentArrayDynamicIndexingEXT
:
4267 case SpvCapabilityUniformTexelBufferArrayDynamicIndexingEXT
:
4268 case SpvCapabilityStorageTexelBufferArrayDynamicIndexingEXT
:
4269 spv_check_supported(descriptor_array_dynamic_indexing
, cap
);
4272 case SpvCapabilityUniformBufferArrayNonUniformIndexingEXT
:
4273 case SpvCapabilitySampledImageArrayNonUniformIndexingEXT
:
4274 case SpvCapabilityStorageBufferArrayNonUniformIndexingEXT
:
4275 case SpvCapabilityStorageImageArrayNonUniformIndexingEXT
:
4276 case SpvCapabilityInputAttachmentArrayNonUniformIndexingEXT
:
4277 case SpvCapabilityUniformTexelBufferArrayNonUniformIndexingEXT
:
4278 case SpvCapabilityStorageTexelBufferArrayNonUniformIndexingEXT
:
4279 spv_check_supported(descriptor_array_non_uniform_indexing
, cap
);
4282 case SpvCapabilityRuntimeDescriptorArrayEXT
:
4283 spv_check_supported(runtime_descriptor_array
, cap
);
4286 case SpvCapabilityStencilExportEXT
:
4287 spv_check_supported(stencil_export
, cap
);
4290 case SpvCapabilitySampleMaskPostDepthCoverage
:
4291 spv_check_supported(post_depth_coverage
, cap
);
4294 case SpvCapabilityDenormFlushToZero
:
4295 case SpvCapabilityDenormPreserve
:
4296 case SpvCapabilitySignedZeroInfNanPreserve
:
4297 case SpvCapabilityRoundingModeRTE
:
4298 case SpvCapabilityRoundingModeRTZ
:
4299 spv_check_supported(float_controls
, cap
);
4302 case SpvCapabilityPhysicalStorageBufferAddresses
:
4303 spv_check_supported(physical_storage_buffer_address
, cap
);
4306 case SpvCapabilityComputeDerivativeGroupQuadsNV
:
4307 case SpvCapabilityComputeDerivativeGroupLinearNV
:
4308 spv_check_supported(derivative_group
, cap
);
4311 case SpvCapabilityFloat16
:
4312 spv_check_supported(float16
, cap
);
4315 case SpvCapabilityFragmentShaderSampleInterlockEXT
:
4316 spv_check_supported(fragment_shader_sample_interlock
, cap
);
4319 case SpvCapabilityFragmentShaderPixelInterlockEXT
:
4320 spv_check_supported(fragment_shader_pixel_interlock
, cap
);
4323 case SpvCapabilityDemoteToHelperInvocationEXT
:
4324 spv_check_supported(demote_to_helper_invocation
, cap
);
4327 case SpvCapabilityShaderClockKHR
:
4328 spv_check_supported(shader_clock
, cap
);
4331 case SpvCapabilityVulkanMemoryModel
:
4332 spv_check_supported(vk_memory_model
, cap
);
4335 case SpvCapabilityVulkanMemoryModelDeviceScope
:
4336 spv_check_supported(vk_memory_model_device_scope
, cap
);
4339 case SpvCapabilityImageReadWriteLodAMD
:
4340 spv_check_supported(amd_image_read_write_lod
, cap
);
4343 case SpvCapabilityIntegerFunctions2INTEL
:
4344 spv_check_supported(integer_functions2
, cap
);
4347 case SpvCapabilityFragmentMaskAMD
:
4348 spv_check_supported(amd_fragment_mask
, cap
);
4351 case SpvCapabilityImageGatherBiasLodAMD
:
4352 spv_check_supported(amd_image_gather_bias_lod
, cap
);
4355 case SpvCapabilityAtomicFloat32AddEXT
:
4356 spv_check_supported(float32_atomic_add
, cap
);
4359 case SpvCapabilityAtomicFloat64AddEXT
:
4360 spv_check_supported(float64_atomic_add
, cap
);
4364 vtn_fail("Unhandled capability: %s (%u)",
4365 spirv_capability_to_string(cap
), cap
);
4370 case SpvOpExtInstImport
:
4371 vtn_handle_extension(b
, opcode
, w
, count
);
4374 case SpvOpMemoryModel
:
4376 case SpvAddressingModelPhysical32
:
4377 vtn_fail_if(b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
,
4378 "AddressingModelPhysical32 only supported for kernels");
4379 b
->shader
->info
.cs
.ptr_size
= 32;
4380 b
->physical_ptrs
= true;
4381 b
->options
->shared_addr_format
= nir_address_format_32bit_global
;
4382 b
->options
->global_addr_format
= nir_address_format_32bit_global
;
4383 b
->options
->temp_addr_format
= nir_address_format_32bit_global
;
4385 case SpvAddressingModelPhysical64
:
4386 vtn_fail_if(b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
,
4387 "AddressingModelPhysical64 only supported for kernels");
4388 b
->shader
->info
.cs
.ptr_size
= 64;
4389 b
->physical_ptrs
= true;
4390 b
->options
->shared_addr_format
= nir_address_format_64bit_global
;
4391 b
->options
->global_addr_format
= nir_address_format_64bit_global
;
4392 b
->options
->temp_addr_format
= nir_address_format_64bit_global
;
4394 case SpvAddressingModelLogical
:
4395 vtn_fail_if(b
->shader
->info
.stage
== MESA_SHADER_KERNEL
,
4396 "AddressingModelLogical only supported for shaders");
4397 b
->physical_ptrs
= false;
4399 case SpvAddressingModelPhysicalStorageBuffer64
:
4400 vtn_fail_if(!b
->options
||
4401 !b
->options
->caps
.physical_storage_buffer_address
,
4402 "AddressingModelPhysicalStorageBuffer64 not supported");
4405 vtn_fail("Unknown addressing model: %s (%u)",
4406 spirv_addressingmodel_to_string(w
[1]), w
[1]);
4410 b
->mem_model
= w
[2];
4412 case SpvMemoryModelSimple
:
4413 case SpvMemoryModelGLSL450
:
4414 case SpvMemoryModelOpenCL
:
4416 case SpvMemoryModelVulkan
:
4417 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
4418 "Vulkan memory model is unsupported by this driver");
4421 vtn_fail("Unsupported memory model: %s",
4422 spirv_memorymodel_to_string(w
[2]));
4427 case SpvOpEntryPoint
:
4428 vtn_handle_entry_point(b
, w
, count
);
4432 vtn_push_value(b
, w
[1], vtn_value_type_string
)->str
=
4433 vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
4437 b
->values
[w
[1]].name
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
4440 case SpvOpMemberName
:
4444 case SpvOpExecutionMode
:
4445 case SpvOpExecutionModeId
:
4446 case SpvOpDecorationGroup
:
4448 case SpvOpDecorateId
:
4449 case SpvOpMemberDecorate
:
4450 case SpvOpGroupDecorate
:
4451 case SpvOpGroupMemberDecorate
:
4452 case SpvOpDecorateString
:
4453 case SpvOpMemberDecorateString
:
4454 vtn_handle_decoration(b
, opcode
, w
, count
);
4457 case SpvOpExtInst
: {
4458 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
4459 if (val
->ext_handler
== vtn_handle_non_semantic_instruction
) {
4460 /* NonSemantic extended instructions are acceptable in preamble. */
4461 vtn_handle_non_semantic_instruction(b
, w
[4], w
, count
);
4464 return false; /* End of preamble. */
4469 return false; /* End of preamble */
4476 vtn_handle_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
4477 const struct vtn_decoration
*mode
, UNUSED
void *data
)
4479 vtn_assert(b
->entry_point
== entry_point
);
4481 switch(mode
->exec_mode
) {
4482 case SpvExecutionModeOriginUpperLeft
:
4483 case SpvExecutionModeOriginLowerLeft
:
4484 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4485 b
->shader
->info
.fs
.origin_upper_left
=
4486 (mode
->exec_mode
== SpvExecutionModeOriginUpperLeft
);
4489 case SpvExecutionModeEarlyFragmentTests
:
4490 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4491 b
->shader
->info
.fs
.early_fragment_tests
= true;
4494 case SpvExecutionModePostDepthCoverage
:
4495 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4496 b
->shader
->info
.fs
.post_depth_coverage
= true;
4499 case SpvExecutionModeInvocations
:
4500 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4501 b
->shader
->info
.gs
.invocations
= MAX2(1, mode
->operands
[0]);
4504 case SpvExecutionModeDepthReplacing
:
4505 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4506 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_ANY
;
4508 case SpvExecutionModeDepthGreater
:
4509 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4510 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_GREATER
;
4512 case SpvExecutionModeDepthLess
:
4513 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4514 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_LESS
;
4516 case SpvExecutionModeDepthUnchanged
:
4517 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4518 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
4521 case SpvExecutionModeLocalSize
:
4522 vtn_assert(gl_shader_stage_is_compute(b
->shader
->info
.stage
));
4523 b
->shader
->info
.cs
.local_size
[0] = mode
->operands
[0];
4524 b
->shader
->info
.cs
.local_size
[1] = mode
->operands
[1];
4525 b
->shader
->info
.cs
.local_size
[2] = mode
->operands
[2];
4528 case SpvExecutionModeLocalSizeHint
:
4529 break; /* Nothing to do with this */
4531 case SpvExecutionModeOutputVertices
:
4532 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4533 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
4534 b
->shader
->info
.tess
.tcs_vertices_out
= mode
->operands
[0];
4536 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4537 b
->shader
->info
.gs
.vertices_out
= mode
->operands
[0];
4541 case SpvExecutionModeInputPoints
:
4542 case SpvExecutionModeInputLines
:
4543 case SpvExecutionModeInputLinesAdjacency
:
4544 case SpvExecutionModeTriangles
:
4545 case SpvExecutionModeInputTrianglesAdjacency
:
4546 case SpvExecutionModeQuads
:
4547 case SpvExecutionModeIsolines
:
4548 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4549 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
4550 b
->shader
->info
.tess
.primitive_mode
=
4551 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
4553 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4554 b
->shader
->info
.gs
.vertices_in
=
4555 vertices_in_from_spv_execution_mode(b
, mode
->exec_mode
);
4556 b
->shader
->info
.gs
.input_primitive
=
4557 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
4561 case SpvExecutionModeOutputPoints
:
4562 case SpvExecutionModeOutputLineStrip
:
4563 case SpvExecutionModeOutputTriangleStrip
:
4564 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4565 b
->shader
->info
.gs
.output_primitive
=
4566 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
4569 case SpvExecutionModeSpacingEqual
:
4570 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4571 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4572 b
->shader
->info
.tess
.spacing
= TESS_SPACING_EQUAL
;
4574 case SpvExecutionModeSpacingFractionalEven
:
4575 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4576 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4577 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_EVEN
;
4579 case SpvExecutionModeSpacingFractionalOdd
:
4580 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4581 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4582 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_ODD
;
4584 case SpvExecutionModeVertexOrderCw
:
4585 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4586 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4587 b
->shader
->info
.tess
.ccw
= false;
4589 case SpvExecutionModeVertexOrderCcw
:
4590 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4591 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4592 b
->shader
->info
.tess
.ccw
= true;
4594 case SpvExecutionModePointMode
:
4595 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4596 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4597 b
->shader
->info
.tess
.point_mode
= true;
4600 case SpvExecutionModePixelCenterInteger
:
4601 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4602 b
->shader
->info
.fs
.pixel_center_integer
= true;
4605 case SpvExecutionModeXfb
:
4606 b
->shader
->info
.has_transform_feedback_varyings
= true;
4609 case SpvExecutionModeVecTypeHint
:
4612 case SpvExecutionModeContractionOff
:
4613 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
)
4614 vtn_warn("ExectionMode only allowed for CL-style kernels: %s",
4615 spirv_executionmode_to_string(mode
->exec_mode
));
4620 case SpvExecutionModeStencilRefReplacingEXT
:
4621 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4624 case SpvExecutionModeDerivativeGroupQuadsNV
:
4625 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
4626 b
->shader
->info
.cs
.derivative_group
= DERIVATIVE_GROUP_QUADS
;
4629 case SpvExecutionModeDerivativeGroupLinearNV
:
4630 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
4631 b
->shader
->info
.cs
.derivative_group
= DERIVATIVE_GROUP_LINEAR
;
4634 case SpvExecutionModePixelInterlockOrderedEXT
:
4635 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4636 b
->shader
->info
.fs
.pixel_interlock_ordered
= true;
4639 case SpvExecutionModePixelInterlockUnorderedEXT
:
4640 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4641 b
->shader
->info
.fs
.pixel_interlock_unordered
= true;
4644 case SpvExecutionModeSampleInterlockOrderedEXT
:
4645 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4646 b
->shader
->info
.fs
.sample_interlock_ordered
= true;
4649 case SpvExecutionModeSampleInterlockUnorderedEXT
:
4650 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4651 b
->shader
->info
.fs
.sample_interlock_unordered
= true;
4654 case SpvExecutionModeDenormPreserve
:
4655 case SpvExecutionModeDenormFlushToZero
:
4656 case SpvExecutionModeSignedZeroInfNanPreserve
:
4657 case SpvExecutionModeRoundingModeRTE
:
4658 case SpvExecutionModeRoundingModeRTZ
: {
4659 unsigned execution_mode
= 0;
4660 switch (mode
->exec_mode
) {
4661 case SpvExecutionModeDenormPreserve
:
4662 switch (mode
->operands
[0]) {
4663 case 16: execution_mode
= FLOAT_CONTROLS_DENORM_PRESERVE_FP16
; break;
4664 case 32: execution_mode
= FLOAT_CONTROLS_DENORM_PRESERVE_FP32
; break;
4665 case 64: execution_mode
= FLOAT_CONTROLS_DENORM_PRESERVE_FP64
; break;
4666 default: vtn_fail("Floating point type not supported");
4669 case SpvExecutionModeDenormFlushToZero
:
4670 switch (mode
->operands
[0]) {
4671 case 16: execution_mode
= FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16
; break;
4672 case 32: execution_mode
= FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32
; break;
4673 case 64: execution_mode
= FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP64
; break;
4674 default: vtn_fail("Floating point type not supported");
4677 case SpvExecutionModeSignedZeroInfNanPreserve
:
4678 switch (mode
->operands
[0]) {
4679 case 16: execution_mode
= FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP16
; break;
4680 case 32: execution_mode
= FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP32
; break;
4681 case 64: execution_mode
= FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP64
; break;
4682 default: vtn_fail("Floating point type not supported");
4685 case SpvExecutionModeRoundingModeRTE
:
4686 switch (mode
->operands
[0]) {
4687 case 16: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
; break;
4688 case 32: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
; break;
4689 case 64: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
; break;
4690 default: vtn_fail("Floating point type not supported");
4693 case SpvExecutionModeRoundingModeRTZ
:
4694 switch (mode
->operands
[0]) {
4695 case 16: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
; break;
4696 case 32: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
; break;
4697 case 64: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
; break;
4698 default: vtn_fail("Floating point type not supported");
4705 b
->shader
->info
.float_controls_execution_mode
|= execution_mode
;
4709 case SpvExecutionModeLocalSizeId
:
4710 case SpvExecutionModeLocalSizeHintId
:
4711 /* Handled later by vtn_handle_execution_mode_id(). */
4715 vtn_fail("Unhandled execution mode: %s (%u)",
4716 spirv_executionmode_to_string(mode
->exec_mode
),
4722 vtn_handle_execution_mode_id(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
4723 const struct vtn_decoration
*mode
, UNUSED
void *data
)
4726 vtn_assert(b
->entry_point
== entry_point
);
4728 switch (mode
->exec_mode
) {
4729 case SpvExecutionModeLocalSizeId
:
4730 b
->shader
->info
.cs
.local_size
[0] = vtn_constant_uint(b
, mode
->operands
[0]);
4731 b
->shader
->info
.cs
.local_size
[1] = vtn_constant_uint(b
, mode
->operands
[1]);
4732 b
->shader
->info
.cs
.local_size
[2] = vtn_constant_uint(b
, mode
->operands
[2]);
4735 case SpvExecutionModeLocalSizeHintId
:
4736 /* Nothing to do with this hint. */
4740 /* Nothing to do. Literal execution modes already handled by
4741 * vtn_handle_execution_mode(). */
4747 vtn_handle_variable_or_type_instruction(struct vtn_builder
*b
, SpvOp opcode
,
4748 const uint32_t *w
, unsigned count
)
4750 vtn_set_instruction_result_type(b
, opcode
, w
, count
);
4754 case SpvOpSourceContinued
:
4755 case SpvOpSourceExtension
:
4756 case SpvOpExtension
:
4757 case SpvOpCapability
:
4758 case SpvOpExtInstImport
:
4759 case SpvOpMemoryModel
:
4760 case SpvOpEntryPoint
:
4761 case SpvOpExecutionMode
:
4764 case SpvOpMemberName
:
4765 case SpvOpDecorationGroup
:
4767 case SpvOpDecorateId
:
4768 case SpvOpMemberDecorate
:
4769 case SpvOpGroupDecorate
:
4770 case SpvOpGroupMemberDecorate
:
4771 case SpvOpDecorateString
:
4772 case SpvOpMemberDecorateString
:
4773 vtn_fail("Invalid opcode types and variables section");
4779 case SpvOpTypeFloat
:
4780 case SpvOpTypeVector
:
4781 case SpvOpTypeMatrix
:
4782 case SpvOpTypeImage
:
4783 case SpvOpTypeSampler
:
4784 case SpvOpTypeSampledImage
:
4785 case SpvOpTypeArray
:
4786 case SpvOpTypeRuntimeArray
:
4787 case SpvOpTypeStruct
:
4788 case SpvOpTypeOpaque
:
4789 case SpvOpTypePointer
:
4790 case SpvOpTypeForwardPointer
:
4791 case SpvOpTypeFunction
:
4792 case SpvOpTypeEvent
:
4793 case SpvOpTypeDeviceEvent
:
4794 case SpvOpTypeReserveId
:
4795 case SpvOpTypeQueue
:
4797 vtn_handle_type(b
, opcode
, w
, count
);
4800 case SpvOpConstantTrue
:
4801 case SpvOpConstantFalse
:
4803 case SpvOpConstantComposite
:
4804 case SpvOpConstantSampler
:
4805 case SpvOpConstantNull
:
4806 case SpvOpSpecConstantTrue
:
4807 case SpvOpSpecConstantFalse
:
4808 case SpvOpSpecConstant
:
4809 case SpvOpSpecConstantComposite
:
4810 case SpvOpSpecConstantOp
:
4811 vtn_handle_constant(b
, opcode
, w
, count
);
4816 vtn_handle_variables(b
, opcode
, w
, count
);
4819 case SpvOpExtInst
: {
4820 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
4821 /* NonSemantic extended instructions are acceptable in preamble, others
4822 * will indicate the end of preamble.
4824 return val
->ext_handler
== vtn_handle_non_semantic_instruction
;
4828 return false; /* End of preamble */
4834 static struct vtn_ssa_value
*
4835 vtn_nir_select(struct vtn_builder
*b
, struct vtn_ssa_value
*src0
,
4836 struct vtn_ssa_value
*src1
, struct vtn_ssa_value
*src2
)
4838 struct vtn_ssa_value
*dest
= rzalloc(b
, struct vtn_ssa_value
);
4839 dest
->type
= src1
->type
;
4841 if (glsl_type_is_vector_or_scalar(src1
->type
)) {
4842 dest
->def
= nir_bcsel(&b
->nb
, src0
->def
, src1
->def
, src2
->def
);
4844 unsigned elems
= glsl_get_length(src1
->type
);
4846 dest
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
4847 for (unsigned i
= 0; i
< elems
; i
++) {
4848 dest
->elems
[i
] = vtn_nir_select(b
, src0
,
4849 src1
->elems
[i
], src2
->elems
[i
]);
4857 vtn_handle_select(struct vtn_builder
*b
, SpvOp opcode
,
4858 const uint32_t *w
, unsigned count
)
4860 /* Handle OpSelect up-front here because it needs to be able to handle
4861 * pointers and not just regular vectors and scalars.
4863 struct vtn_value
*res_val
= vtn_untyped_value(b
, w
[2]);
4864 struct vtn_value
*cond_val
= vtn_untyped_value(b
, w
[3]);
4865 struct vtn_value
*obj1_val
= vtn_untyped_value(b
, w
[4]);
4866 struct vtn_value
*obj2_val
= vtn_untyped_value(b
, w
[5]);
4868 vtn_fail_if(obj1_val
->type
!= res_val
->type
||
4869 obj2_val
->type
!= res_val
->type
,
4870 "Object types must match the result type in OpSelect");
4872 vtn_fail_if((cond_val
->type
->base_type
!= vtn_base_type_scalar
&&
4873 cond_val
->type
->base_type
!= vtn_base_type_vector
) ||
4874 !glsl_type_is_boolean(cond_val
->type
->type
),
4875 "OpSelect must have either a vector of booleans or "
4876 "a boolean as Condition type");
4878 vtn_fail_if(cond_val
->type
->base_type
== vtn_base_type_vector
&&
4879 (res_val
->type
->base_type
!= vtn_base_type_vector
||
4880 res_val
->type
->length
!= cond_val
->type
->length
),
4881 "When Condition type in OpSelect is a vector, the Result "
4882 "type must be a vector of the same length");
4884 switch (res_val
->type
->base_type
) {
4885 case vtn_base_type_scalar
:
4886 case vtn_base_type_vector
:
4887 case vtn_base_type_matrix
:
4888 case vtn_base_type_array
:
4889 case vtn_base_type_struct
:
4892 case vtn_base_type_pointer
:
4893 /* We need to have actual storage for pointer types. */
4894 vtn_fail_if(res_val
->type
->type
== NULL
,
4895 "Invalid pointer result type for OpSelect");
4898 vtn_fail("Result type of OpSelect must be a scalar, composite, or pointer");
4901 vtn_push_ssa_value(b
, w
[2],
4902 vtn_nir_select(b
, vtn_ssa_value(b
, w
[3]),
4903 vtn_ssa_value(b
, w
[4]),
4904 vtn_ssa_value(b
, w
[5])));
4908 vtn_handle_ptr(struct vtn_builder
*b
, SpvOp opcode
,
4909 const uint32_t *w
, unsigned count
)
4911 struct vtn_type
*type1
= vtn_get_value_type(b
, w
[3]);
4912 struct vtn_type
*type2
= vtn_get_value_type(b
, w
[4]);
4913 vtn_fail_if(type1
->base_type
!= vtn_base_type_pointer
||
4914 type2
->base_type
!= vtn_base_type_pointer
,
4915 "%s operands must have pointer types",
4916 spirv_op_to_string(opcode
));
4917 vtn_fail_if(type1
->storage_class
!= type2
->storage_class
,
4918 "%s operands must have the same storage class",
4919 spirv_op_to_string(opcode
));
4921 struct vtn_type
*vtn_type
= vtn_get_type(b
, w
[1]);
4922 const struct glsl_type
*type
= vtn_type
->type
;
4924 nir_address_format addr_format
= vtn_mode_to_address_format(
4925 b
, vtn_storage_class_to_mode(b
, type1
->storage_class
, NULL
, NULL
));
4930 case SpvOpPtrDiff
: {
4931 /* OpPtrDiff returns the difference in number of elements (not byte offset). */
4932 unsigned elem_size
, elem_align
;
4933 glsl_get_natural_size_align_bytes(type1
->deref
->type
,
4934 &elem_size
, &elem_align
);
4936 def
= nir_build_addr_isub(&b
->nb
,
4937 vtn_get_nir_ssa(b
, w
[3]),
4938 vtn_get_nir_ssa(b
, w
[4]),
4940 def
= nir_idiv(&b
->nb
, def
, nir_imm_intN_t(&b
->nb
, elem_size
, def
->bit_size
));
4941 def
= nir_i2i(&b
->nb
, def
, glsl_get_bit_size(type
));
4946 case SpvOpPtrNotEqual
: {
4947 def
= nir_build_addr_ieq(&b
->nb
,
4948 vtn_get_nir_ssa(b
, w
[3]),
4949 vtn_get_nir_ssa(b
, w
[4]),
4951 if (opcode
== SpvOpPtrNotEqual
)
4952 def
= nir_inot(&b
->nb
, def
);
4957 unreachable("Invalid ptr operation");
4960 vtn_push_nir_ssa(b
, w
[2], def
);
4964 vtn_handle_body_instruction(struct vtn_builder
*b
, SpvOp opcode
,
4965 const uint32_t *w
, unsigned count
)
4971 case SpvOpLoopMerge
:
4972 case SpvOpSelectionMerge
:
4973 /* This is handled by cfg pre-pass and walk_blocks */
4977 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
4978 val
->type
= vtn_get_type(b
, w
[1]);
4983 vtn_handle_extension(b
, opcode
, w
, count
);
4989 case SpvOpCopyMemory
:
4990 case SpvOpCopyMemorySized
:
4991 case SpvOpAccessChain
:
4992 case SpvOpPtrAccessChain
:
4993 case SpvOpInBoundsAccessChain
:
4994 case SpvOpInBoundsPtrAccessChain
:
4995 case SpvOpArrayLength
:
4996 case SpvOpConvertPtrToU
:
4997 case SpvOpConvertUToPtr
:
4998 vtn_handle_variables(b
, opcode
, w
, count
);
5001 case SpvOpFunctionCall
:
5002 vtn_handle_function_call(b
, opcode
, w
, count
);
5005 case SpvOpSampledImage
:
5007 case SpvOpImageSampleImplicitLod
:
5008 case SpvOpImageSampleExplicitLod
:
5009 case SpvOpImageSampleDrefImplicitLod
:
5010 case SpvOpImageSampleDrefExplicitLod
:
5011 case SpvOpImageSampleProjImplicitLod
:
5012 case SpvOpImageSampleProjExplicitLod
:
5013 case SpvOpImageSampleProjDrefImplicitLod
:
5014 case SpvOpImageSampleProjDrefExplicitLod
:
5015 case SpvOpImageFetch
:
5016 case SpvOpImageGather
:
5017 case SpvOpImageDrefGather
:
5018 case SpvOpImageQuerySizeLod
:
5019 case SpvOpImageQueryLod
:
5020 case SpvOpImageQueryLevels
:
5021 case SpvOpImageQuerySamples
:
5022 vtn_handle_texture(b
, opcode
, w
, count
);
5025 case SpvOpImageRead
:
5026 case SpvOpImageWrite
:
5027 case SpvOpImageTexelPointer
:
5028 vtn_handle_image(b
, opcode
, w
, count
);
5031 case SpvOpImageQuerySize
: {
5032 struct vtn_type
*image_type
= vtn_get_value_type(b
, w
[3]);
5033 vtn_assert(image_type
->base_type
== vtn_base_type_image
);
5034 if (glsl_type_is_image(image_type
->glsl_image
)) {
5035 vtn_handle_image(b
, opcode
, w
, count
);
5037 vtn_assert(glsl_type_is_sampler(image_type
->glsl_image
));
5038 vtn_handle_texture(b
, opcode
, w
, count
);
5043 case SpvOpFragmentMaskFetchAMD
:
5044 case SpvOpFragmentFetchAMD
:
5045 vtn_handle_texture(b
, opcode
, w
, count
);
5048 case SpvOpAtomicLoad
:
5049 case SpvOpAtomicExchange
:
5050 case SpvOpAtomicCompareExchange
:
5051 case SpvOpAtomicCompareExchangeWeak
:
5052 case SpvOpAtomicIIncrement
:
5053 case SpvOpAtomicIDecrement
:
5054 case SpvOpAtomicIAdd
:
5055 case SpvOpAtomicISub
:
5056 case SpvOpAtomicSMin
:
5057 case SpvOpAtomicUMin
:
5058 case SpvOpAtomicSMax
:
5059 case SpvOpAtomicUMax
:
5060 case SpvOpAtomicAnd
:
5062 case SpvOpAtomicXor
:
5063 case SpvOpAtomicFAddEXT
: {
5064 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[3]);
5065 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
5066 vtn_handle_image(b
, opcode
, w
, count
);
5068 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
5069 vtn_handle_atomics(b
, opcode
, w
, count
);
5074 case SpvOpAtomicStore
: {
5075 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[1]);
5076 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
5077 vtn_handle_image(b
, opcode
, w
, count
);
5079 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
5080 vtn_handle_atomics(b
, opcode
, w
, count
);
5086 vtn_handle_select(b
, opcode
, w
, count
);
5094 case SpvOpConvertFToU
:
5095 case SpvOpConvertFToS
:
5096 case SpvOpConvertSToF
:
5097 case SpvOpConvertUToF
:
5101 case SpvOpQuantizeToF16
:
5102 case SpvOpPtrCastToGeneric
:
5103 case SpvOpGenericCastToPtr
:
5108 case SpvOpSignBitSet
:
5109 case SpvOpLessOrGreater
:
5111 case SpvOpUnordered
:
5126 case SpvOpVectorTimesScalar
:
5128 case SpvOpIAddCarry
:
5129 case SpvOpISubBorrow
:
5130 case SpvOpUMulExtended
:
5131 case SpvOpSMulExtended
:
5132 case SpvOpShiftRightLogical
:
5133 case SpvOpShiftRightArithmetic
:
5134 case SpvOpShiftLeftLogical
:
5135 case SpvOpLogicalEqual
:
5136 case SpvOpLogicalNotEqual
:
5137 case SpvOpLogicalOr
:
5138 case SpvOpLogicalAnd
:
5139 case SpvOpLogicalNot
:
5140 case SpvOpBitwiseOr
:
5141 case SpvOpBitwiseXor
:
5142 case SpvOpBitwiseAnd
:
5144 case SpvOpFOrdEqual
:
5145 case SpvOpFUnordEqual
:
5146 case SpvOpINotEqual
:
5147 case SpvOpFOrdNotEqual
:
5148 case SpvOpFUnordNotEqual
:
5149 case SpvOpULessThan
:
5150 case SpvOpSLessThan
:
5151 case SpvOpFOrdLessThan
:
5152 case SpvOpFUnordLessThan
:
5153 case SpvOpUGreaterThan
:
5154 case SpvOpSGreaterThan
:
5155 case SpvOpFOrdGreaterThan
:
5156 case SpvOpFUnordGreaterThan
:
5157 case SpvOpULessThanEqual
:
5158 case SpvOpSLessThanEqual
:
5159 case SpvOpFOrdLessThanEqual
:
5160 case SpvOpFUnordLessThanEqual
:
5161 case SpvOpUGreaterThanEqual
:
5162 case SpvOpSGreaterThanEqual
:
5163 case SpvOpFOrdGreaterThanEqual
:
5164 case SpvOpFUnordGreaterThanEqual
:
5170 case SpvOpFwidthFine
:
5171 case SpvOpDPdxCoarse
:
5172 case SpvOpDPdyCoarse
:
5173 case SpvOpFwidthCoarse
:
5174 case SpvOpBitFieldInsert
:
5175 case SpvOpBitFieldSExtract
:
5176 case SpvOpBitFieldUExtract
:
5177 case SpvOpBitReverse
:
5179 case SpvOpTranspose
:
5180 case SpvOpOuterProduct
:
5181 case SpvOpMatrixTimesScalar
:
5182 case SpvOpVectorTimesMatrix
:
5183 case SpvOpMatrixTimesVector
:
5184 case SpvOpMatrixTimesMatrix
:
5185 case SpvOpUCountLeadingZerosINTEL
:
5186 case SpvOpUCountTrailingZerosINTEL
:
5187 case SpvOpAbsISubINTEL
:
5188 case SpvOpAbsUSubINTEL
:
5189 case SpvOpIAddSatINTEL
:
5190 case SpvOpUAddSatINTEL
:
5191 case SpvOpIAverageINTEL
:
5192 case SpvOpUAverageINTEL
:
5193 case SpvOpIAverageRoundedINTEL
:
5194 case SpvOpUAverageRoundedINTEL
:
5195 case SpvOpISubSatINTEL
:
5196 case SpvOpUSubSatINTEL
:
5197 case SpvOpIMul32x16INTEL
:
5198 case SpvOpUMul32x16INTEL
:
5199 vtn_handle_alu(b
, opcode
, w
, count
);
5203 vtn_handle_bitcast(b
, w
, count
);
5206 case SpvOpVectorExtractDynamic
:
5207 case SpvOpVectorInsertDynamic
:
5208 case SpvOpVectorShuffle
:
5209 case SpvOpCompositeConstruct
:
5210 case SpvOpCompositeExtract
:
5211 case SpvOpCompositeInsert
:
5212 case SpvOpCopyLogical
:
5213 case SpvOpCopyObject
:
5214 vtn_handle_composite(b
, opcode
, w
, count
);
5217 case SpvOpEmitVertex
:
5218 case SpvOpEndPrimitive
:
5219 case SpvOpEmitStreamVertex
:
5220 case SpvOpEndStreamPrimitive
:
5221 case SpvOpControlBarrier
:
5222 case SpvOpMemoryBarrier
:
5223 vtn_handle_barrier(b
, opcode
, w
, count
);
5226 case SpvOpGroupNonUniformElect
:
5227 case SpvOpGroupNonUniformAll
:
5228 case SpvOpGroupNonUniformAny
:
5229 case SpvOpGroupNonUniformAllEqual
:
5230 case SpvOpGroupNonUniformBroadcast
:
5231 case SpvOpGroupNonUniformBroadcastFirst
:
5232 case SpvOpGroupNonUniformBallot
:
5233 case SpvOpGroupNonUniformInverseBallot
:
5234 case SpvOpGroupNonUniformBallotBitExtract
:
5235 case SpvOpGroupNonUniformBallotBitCount
:
5236 case SpvOpGroupNonUniformBallotFindLSB
:
5237 case SpvOpGroupNonUniformBallotFindMSB
:
5238 case SpvOpGroupNonUniformShuffle
:
5239 case SpvOpGroupNonUniformShuffleXor
:
5240 case SpvOpGroupNonUniformShuffleUp
:
5241 case SpvOpGroupNonUniformShuffleDown
:
5242 case SpvOpGroupNonUniformIAdd
:
5243 case SpvOpGroupNonUniformFAdd
:
5244 case SpvOpGroupNonUniformIMul
:
5245 case SpvOpGroupNonUniformFMul
:
5246 case SpvOpGroupNonUniformSMin
:
5247 case SpvOpGroupNonUniformUMin
:
5248 case SpvOpGroupNonUniformFMin
:
5249 case SpvOpGroupNonUniformSMax
:
5250 case SpvOpGroupNonUniformUMax
:
5251 case SpvOpGroupNonUniformFMax
:
5252 case SpvOpGroupNonUniformBitwiseAnd
:
5253 case SpvOpGroupNonUniformBitwiseOr
:
5254 case SpvOpGroupNonUniformBitwiseXor
:
5255 case SpvOpGroupNonUniformLogicalAnd
:
5256 case SpvOpGroupNonUniformLogicalOr
:
5257 case SpvOpGroupNonUniformLogicalXor
:
5258 case SpvOpGroupNonUniformQuadBroadcast
:
5259 case SpvOpGroupNonUniformQuadSwap
:
5262 case SpvOpGroupBroadcast
:
5263 case SpvOpGroupIAdd
:
5264 case SpvOpGroupFAdd
:
5265 case SpvOpGroupFMin
:
5266 case SpvOpGroupUMin
:
5267 case SpvOpGroupSMin
:
5268 case SpvOpGroupFMax
:
5269 case SpvOpGroupUMax
:
5270 case SpvOpGroupSMax
:
5271 case SpvOpSubgroupBallotKHR
:
5272 case SpvOpSubgroupFirstInvocationKHR
:
5273 case SpvOpSubgroupReadInvocationKHR
:
5274 case SpvOpSubgroupAllKHR
:
5275 case SpvOpSubgroupAnyKHR
:
5276 case SpvOpSubgroupAllEqualKHR
:
5277 case SpvOpGroupIAddNonUniformAMD
:
5278 case SpvOpGroupFAddNonUniformAMD
:
5279 case SpvOpGroupFMinNonUniformAMD
:
5280 case SpvOpGroupUMinNonUniformAMD
:
5281 case SpvOpGroupSMinNonUniformAMD
:
5282 case SpvOpGroupFMaxNonUniformAMD
:
5283 case SpvOpGroupUMaxNonUniformAMD
:
5284 case SpvOpGroupSMaxNonUniformAMD
:
5285 vtn_handle_subgroup(b
, opcode
, w
, count
);
5290 case SpvOpPtrNotEqual
:
5291 vtn_handle_ptr(b
, opcode
, w
, count
);
5294 case SpvOpBeginInvocationInterlockEXT
:
5295 vtn_emit_barrier(b
, nir_intrinsic_begin_invocation_interlock
);
5298 case SpvOpEndInvocationInterlockEXT
:
5299 vtn_emit_barrier(b
, nir_intrinsic_end_invocation_interlock
);
5302 case SpvOpDemoteToHelperInvocationEXT
: {
5303 nir_intrinsic_instr
*intrin
=
5304 nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_demote
);
5305 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
5309 case SpvOpIsHelperInvocationEXT
: {
5310 nir_intrinsic_instr
*intrin
=
5311 nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_is_helper_invocation
);
5312 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
, 1, 1, NULL
);
5313 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
5315 vtn_push_nir_ssa(b
, w
[2], &intrin
->dest
.ssa
);
5319 case SpvOpReadClockKHR
: {
5320 SpvScope scope
= vtn_constant_uint(b
, w
[3]);
5321 nir_scope nir_scope
;
5324 case SpvScopeDevice
:
5325 nir_scope
= NIR_SCOPE_DEVICE
;
5327 case SpvScopeSubgroup
:
5328 nir_scope
= NIR_SCOPE_SUBGROUP
;
5331 vtn_fail("invalid read clock scope");
5334 /* Operation supports two result types: uvec2 and uint64_t. The NIR
5335 * intrinsic gives uvec2, so pack the result for the other case.
5337 nir_intrinsic_instr
*intrin
=
5338 nir_intrinsic_instr_create(b
->nb
.shader
, nir_intrinsic_shader_clock
);
5339 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
, 2, 32, NULL
);
5340 nir_intrinsic_set_memory_scope(intrin
, nir_scope
);
5341 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
5343 struct vtn_type
*type
= vtn_get_type(b
, w
[1]);
5344 const struct glsl_type
*dest_type
= type
->type
;
5345 nir_ssa_def
*result
;
5347 if (glsl_type_is_vector(dest_type
)) {
5348 assert(dest_type
== glsl_vector_type(GLSL_TYPE_UINT
, 2));
5349 result
= &intrin
->dest
.ssa
;
5351 assert(glsl_type_is_scalar(dest_type
));
5352 assert(glsl_get_base_type(dest_type
) == GLSL_TYPE_UINT64
);
5353 result
= nir_pack_64_2x32(&b
->nb
, &intrin
->dest
.ssa
);
5356 vtn_push_nir_ssa(b
, w
[2], result
);
5360 case SpvOpLifetimeStart
:
5361 case SpvOpLifetimeStop
:
5365 vtn_fail_with_opcode("Unhandled opcode", opcode
);
5372 vtn_create_builder(const uint32_t *words
, size_t word_count
,
5373 gl_shader_stage stage
, const char *entry_point_name
,
5374 const struct spirv_to_nir_options
*options
)
5376 /* Initialize the vtn_builder object */
5377 struct vtn_builder
*b
= rzalloc(NULL
, struct vtn_builder
);
5378 struct spirv_to_nir_options
*dup_options
=
5379 ralloc(b
, struct spirv_to_nir_options
);
5380 *dup_options
= *options
;
5383 b
->spirv_word_count
= word_count
;
5387 list_inithead(&b
->functions
);
5388 b
->entry_point_stage
= stage
;
5389 b
->entry_point_name
= entry_point_name
;
5390 b
->options
= dup_options
;
5393 * Handle the SPIR-V header (first 5 dwords).
5394 * Can't use vtx_assert() as the setjmp(3) target isn't initialized yet.
5396 if (word_count
<= 5)
5399 if (words
[0] != SpvMagicNumber
) {
5400 vtn_err("words[0] was 0x%x, want 0x%x", words
[0], SpvMagicNumber
);
5403 if (words
[1] < 0x10000) {
5404 vtn_err("words[1] was 0x%x, want >= 0x10000", words
[1]);
5408 uint16_t generator_id
= words
[2] >> 16;
5409 uint16_t generator_version
= words
[2];
5411 /* In GLSLang commit 8297936dd6eb3, their handling of barrier() was fixed
5412 * to provide correct memory semantics on compute shader barrier()
5413 * commands. Prior to that, we need to fix them up ourselves. This
5414 * GLSLang fix caused them to bump to generator version 3.
5416 b
->wa_glslang_cs_barrier
= (generator_id
== 8 && generator_version
< 3);
5418 /* words[2] == generator magic */
5419 unsigned value_id_bound
= words
[3];
5420 if (words
[4] != 0) {
5421 vtn_err("words[4] was %u, want 0", words
[4]);
5425 b
->value_id_bound
= value_id_bound
;
5426 b
->values
= rzalloc_array(b
, struct vtn_value
, value_id_bound
);
5434 static nir_function
*
5435 vtn_emit_kernel_entry_point_wrapper(struct vtn_builder
*b
,
5436 nir_function
*entry_point
)
5438 vtn_assert(entry_point
== b
->entry_point
->func
->impl
->function
);
5439 vtn_fail_if(!entry_point
->name
, "entry points are required to have a name");
5440 const char *func_name
=
5441 ralloc_asprintf(b
->shader
, "__wrapped_%s", entry_point
->name
);
5443 /* we shouldn't have any inputs yet */
5444 vtn_assert(!entry_point
->shader
->num_inputs
);
5445 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_KERNEL
);
5447 nir_function
*main_entry_point
= nir_function_create(b
->shader
, func_name
);
5448 main_entry_point
->impl
= nir_function_impl_create(main_entry_point
);
5449 nir_builder_init(&b
->nb
, main_entry_point
->impl
);
5450 b
->nb
.cursor
= nir_after_cf_list(&main_entry_point
->impl
->body
);
5451 b
->func_param_idx
= 0;
5453 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, entry_point
);
5455 for (unsigned i
= 0; i
< entry_point
->num_params
; ++i
) {
5456 struct vtn_type
*param_type
= b
->entry_point
->func
->type
->params
[i
];
5458 /* consider all pointers to function memory to be parameters passed
5461 bool is_by_val
= param_type
->base_type
== vtn_base_type_pointer
&&
5462 param_type
->storage_class
== SpvStorageClassFunction
;
5464 /* input variable */
5465 nir_variable
*in_var
= rzalloc(b
->nb
.shader
, nir_variable
);
5466 in_var
->data
.mode
= nir_var_shader_in
;
5467 in_var
->data
.read_only
= true;
5468 in_var
->data
.location
= i
;
5471 in_var
->type
= param_type
->deref
->type
;
5473 in_var
->type
= param_type
->type
;
5475 nir_shader_add_variable(b
->nb
.shader
, in_var
);
5476 b
->nb
.shader
->num_inputs
++;
5478 /* we have to copy the entire variable into function memory */
5480 nir_variable
*copy_var
=
5481 nir_local_variable_create(main_entry_point
->impl
, in_var
->type
,
5483 nir_copy_var(&b
->nb
, copy_var
, in_var
);
5485 nir_src_for_ssa(&nir_build_deref_var(&b
->nb
, copy_var
)->dest
.ssa
);
5487 call
->params
[i
] = nir_src_for_ssa(nir_load_var(&b
->nb
, in_var
));
5491 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
5493 return main_entry_point
;
5497 spirv_to_nir(const uint32_t *words
, size_t word_count
,
5498 struct nir_spirv_specialization
*spec
, unsigned num_spec
,
5499 gl_shader_stage stage
, const char *entry_point_name
,
5500 const struct spirv_to_nir_options
*options
,
5501 const nir_shader_compiler_options
*nir_options
)
5504 const uint32_t *word_end
= words
+ word_count
;
5506 struct vtn_builder
*b
= vtn_create_builder(words
, word_count
,
5507 stage
, entry_point_name
,
5513 /* See also _vtn_fail() */
5514 if (setjmp(b
->fail_jump
)) {
5519 /* Skip the SPIR-V header, handled at vtn_create_builder */
5522 b
->shader
= nir_shader_create(b
, stage
, nir_options
, NULL
);
5524 /* Handle all the preamble instructions */
5525 words
= vtn_foreach_instruction(b
, words
, word_end
,
5526 vtn_handle_preamble_instruction
);
5528 if (b
->entry_point
== NULL
) {
5529 vtn_fail("Entry point not found");
5534 /* Set shader info defaults */
5535 if (stage
== MESA_SHADER_GEOMETRY
)
5536 b
->shader
->info
.gs
.invocations
= 1;
5538 /* Parse execution modes. */
5539 vtn_foreach_execution_mode(b
, b
->entry_point
,
5540 vtn_handle_execution_mode
, NULL
);
5542 b
->specializations
= spec
;
5543 b
->num_specializations
= num_spec
;
5545 /* Handle all variable, type, and constant instructions */
5546 words
= vtn_foreach_instruction(b
, words
, word_end
,
5547 vtn_handle_variable_or_type_instruction
);
5549 /* Parse execution modes that depend on IDs. Must happen after we have
5552 vtn_foreach_execution_mode(b
, b
->entry_point
,
5553 vtn_handle_execution_mode_id
, NULL
);
5555 if (b
->workgroup_size_builtin
) {
5556 vtn_assert(b
->workgroup_size_builtin
->type
->type
==
5557 glsl_vector_type(GLSL_TYPE_UINT
, 3));
5559 nir_const_value
*const_size
=
5560 b
->workgroup_size_builtin
->constant
->values
;
5562 b
->shader
->info
.cs
.local_size
[0] = const_size
[0].u32
;
5563 b
->shader
->info
.cs
.local_size
[1] = const_size
[1].u32
;
5564 b
->shader
->info
.cs
.local_size
[2] = const_size
[2].u32
;
5567 /* Set types on all vtn_values */
5568 vtn_foreach_instruction(b
, words
, word_end
, vtn_set_instruction_result_type
);
5570 vtn_build_cfg(b
, words
, word_end
);
5572 assert(b
->entry_point
->value_type
== vtn_value_type_function
);
5573 b
->entry_point
->func
->referenced
= true;
5578 vtn_foreach_cf_node(node
, &b
->functions
) {
5579 struct vtn_function
*func
= vtn_cf_node_as_function(node
);
5580 if (func
->referenced
&& !func
->emitted
) {
5581 b
->const_table
= _mesa_pointer_hash_table_create(b
);
5583 vtn_function_emit(b
, func
, vtn_handle_body_instruction
);
5589 vtn_assert(b
->entry_point
->value_type
== vtn_value_type_function
);
5590 nir_function
*entry_point
= b
->entry_point
->func
->impl
->function
;
5591 vtn_assert(entry_point
);
5593 /* post process entry_points with input params */
5594 if (entry_point
->num_params
&& b
->shader
->info
.stage
== MESA_SHADER_KERNEL
)
5595 entry_point
= vtn_emit_kernel_entry_point_wrapper(b
, entry_point
);
5597 /* structurize the CFG */
5598 nir_lower_goto_ifs(b
->shader
);
5600 entry_point
->is_entrypoint
= true;
5602 /* When multiple shader stages exist in the same SPIR-V module, we
5603 * generate input and output variables for every stage, in the same
5604 * NIR program. These dead variables can be invalid NIR. For example,
5605 * TCS outputs must be per-vertex arrays (or decorated 'patch'), while
5606 * VS output variables wouldn't be.
5608 * To ensure we have valid NIR, we eliminate any dead inputs and outputs
5609 * right away. In order to do so, we must lower any constant initializers
5610 * on outputs so nir_remove_dead_variables sees that they're written to.
5612 nir_lower_variable_initializers(b
->shader
, nir_var_shader_out
);
5613 nir_remove_dead_variables(b
->shader
,
5614 nir_var_shader_in
| nir_var_shader_out
, NULL
);
5616 /* We sometimes generate bogus derefs that, while never used, give the
5617 * validator a bit of heartburn. Run dead code to get rid of them.
5619 nir_opt_dce(b
->shader
);
5621 /* Unparent the shader from the vtn_builder before we delete the builder */
5622 ralloc_steal(NULL
, b
->shader
);
5624 nir_shader
*shader
= b
->shader
;