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"
39 #if UTIL_ARCH_BIG_ENDIAN
44 vtn_log(struct vtn_builder
*b
, enum nir_spirv_debug_level level
,
45 size_t spirv_offset
, const char *message
)
47 if (b
->options
->debug
.func
) {
48 b
->options
->debug
.func(b
->options
->debug
.private_data
,
49 level
, spirv_offset
, message
);
53 if (level
>= NIR_SPIRV_DEBUG_LEVEL_WARNING
)
54 fprintf(stderr
, "%s\n", message
);
59 vtn_logf(struct vtn_builder
*b
, enum nir_spirv_debug_level level
,
60 size_t spirv_offset
, const char *fmt
, ...)
66 msg
= ralloc_vasprintf(NULL
, fmt
, args
);
69 vtn_log(b
, level
, spirv_offset
, msg
);
75 vtn_log_err(struct vtn_builder
*b
,
76 enum nir_spirv_debug_level level
, const char *prefix
,
77 const char *file
, unsigned line
,
78 const char *fmt
, va_list args
)
82 msg
= ralloc_strdup(NULL
, prefix
);
85 ralloc_asprintf_append(&msg
, " In file %s:%u\n", file
, line
);
88 ralloc_asprintf_append(&msg
, " ");
90 ralloc_vasprintf_append(&msg
, fmt
, args
);
92 ralloc_asprintf_append(&msg
, "\n %zu bytes into the SPIR-V binary",
96 ralloc_asprintf_append(&msg
,
97 "\n in SPIR-V source file %s, line %d, col %d",
98 b
->file
, b
->line
, b
->col
);
101 vtn_log(b
, level
, b
->spirv_offset
, msg
);
107 vtn_dump_shader(struct vtn_builder
*b
, const char *path
, const char *prefix
)
112 int len
= snprintf(filename
, sizeof(filename
), "%s/%s-%d.spirv",
113 path
, prefix
, idx
++);
114 if (len
< 0 || len
>= sizeof(filename
))
117 FILE *f
= fopen(filename
, "w");
121 fwrite(b
->spirv
, sizeof(*b
->spirv
), b
->spirv_word_count
, f
);
124 vtn_info("SPIR-V shader dumped to %s", filename
);
128 _vtn_warn(struct vtn_builder
*b
, const char *file
, unsigned line
,
129 const char *fmt
, ...)
134 vtn_log_err(b
, NIR_SPIRV_DEBUG_LEVEL_WARNING
, "SPIR-V WARNING:\n",
135 file
, line
, fmt
, args
);
140 _vtn_err(struct vtn_builder
*b
, const char *file
, unsigned line
,
141 const char *fmt
, ...)
146 vtn_log_err(b
, NIR_SPIRV_DEBUG_LEVEL_ERROR
, "SPIR-V ERROR:\n",
147 file
, line
, fmt
, args
);
152 _vtn_fail(struct vtn_builder
*b
, const char *file
, unsigned line
,
153 const char *fmt
, ...)
158 vtn_log_err(b
, NIR_SPIRV_DEBUG_LEVEL_ERROR
, "SPIR-V parsing FAILED:\n",
159 file
, line
, fmt
, args
);
162 const char *dump_path
= getenv("MESA_SPIRV_FAIL_DUMP_PATH");
164 vtn_dump_shader(b
, dump_path
, "fail");
166 longjmp(b
->fail_jump
, 1);
169 static struct vtn_ssa_value
*
170 vtn_undef_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
172 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
173 val
->type
= glsl_get_bare_type(type
);
175 if (glsl_type_is_vector_or_scalar(type
)) {
176 unsigned num_components
= glsl_get_vector_elements(val
->type
);
177 unsigned bit_size
= glsl_get_bit_size(val
->type
);
178 val
->def
= nir_ssa_undef(&b
->nb
, num_components
, bit_size
);
180 unsigned elems
= glsl_get_length(val
->type
);
181 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
182 if (glsl_type_is_array_or_matrix(type
)) {
183 const struct glsl_type
*elem_type
= glsl_get_array_element(type
);
184 for (unsigned i
= 0; i
< elems
; i
++)
185 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
187 vtn_assert(glsl_type_is_struct_or_ifc(type
));
188 for (unsigned i
= 0; i
< elems
; i
++) {
189 const struct glsl_type
*elem_type
= glsl_get_struct_field(type
, i
);
190 val
->elems
[i
] = vtn_undef_ssa_value(b
, elem_type
);
198 static struct vtn_ssa_value
*
199 vtn_const_ssa_value(struct vtn_builder
*b
, nir_constant
*constant
,
200 const struct glsl_type
*type
)
202 struct hash_entry
*entry
= _mesa_hash_table_search(b
->const_table
, constant
);
207 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
208 val
->type
= glsl_get_bare_type(type
);
210 if (glsl_type_is_vector_or_scalar(type
)) {
211 unsigned num_components
= glsl_get_vector_elements(val
->type
);
212 unsigned bit_size
= glsl_get_bit_size(type
);
213 nir_load_const_instr
*load
=
214 nir_load_const_instr_create(b
->shader
, num_components
, bit_size
);
216 memcpy(load
->value
, constant
->values
,
217 sizeof(nir_const_value
) * num_components
);
219 nir_instr_insert_before_cf_list(&b
->nb
.impl
->body
, &load
->instr
);
220 val
->def
= &load
->def
;
222 unsigned elems
= glsl_get_length(val
->type
);
223 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
224 if (glsl_type_is_array_or_matrix(type
)) {
225 const struct glsl_type
*elem_type
= glsl_get_array_element(type
);
226 for (unsigned i
= 0; i
< elems
; i
++) {
227 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
231 vtn_assert(glsl_type_is_struct_or_ifc(type
));
232 for (unsigned i
= 0; i
< elems
; i
++) {
233 const struct glsl_type
*elem_type
= glsl_get_struct_field(type
, i
);
234 val
->elems
[i
] = vtn_const_ssa_value(b
, constant
->elements
[i
],
243 struct vtn_ssa_value
*
244 vtn_ssa_value(struct vtn_builder
*b
, uint32_t value_id
)
246 struct vtn_value
*val
= vtn_untyped_value(b
, value_id
);
247 switch (val
->value_type
) {
248 case vtn_value_type_undef
:
249 return vtn_undef_ssa_value(b
, val
->type
->type
);
251 case vtn_value_type_constant
:
252 return vtn_const_ssa_value(b
, val
->constant
, val
->type
->type
);
254 case vtn_value_type_ssa
:
257 case vtn_value_type_pointer
:
258 vtn_assert(val
->pointer
->ptr_type
&& val
->pointer
->ptr_type
->type
);
259 struct vtn_ssa_value
*ssa
=
260 vtn_create_ssa_value(b
, val
->pointer
->ptr_type
->type
);
261 ssa
->def
= vtn_pointer_to_ssa(b
, val
->pointer
);
265 vtn_fail("Invalid type for an SSA value");
270 vtn_push_ssa_value(struct vtn_builder
*b
, uint32_t value_id
,
271 struct vtn_ssa_value
*ssa
)
273 struct vtn_type
*type
= vtn_get_value_type(b
, value_id
);
275 /* See vtn_create_ssa_value */
276 vtn_fail_if(ssa
->type
!= glsl_get_bare_type(type
->type
),
277 "Type mismatch for SPIR-V SSA value");
279 struct vtn_value
*val
;
280 if (type
->base_type
== vtn_base_type_pointer
) {
281 val
= vtn_push_pointer(b
, value_id
, vtn_pointer_from_ssa(b
, ssa
->def
, type
));
283 /* Don't trip the value_type_ssa check in vtn_push_value */
284 val
= vtn_push_value(b
, value_id
, vtn_value_type_invalid
);
285 val
->value_type
= vtn_value_type_ssa
;
293 vtn_get_nir_ssa(struct vtn_builder
*b
, uint32_t value_id
)
295 struct vtn_ssa_value
*ssa
= vtn_ssa_value(b
, value_id
);
296 vtn_fail_if(!glsl_type_is_vector_or_scalar(ssa
->type
),
297 "Expected a vector or scalar type");
302 vtn_push_nir_ssa(struct vtn_builder
*b
, uint32_t value_id
, nir_ssa_def
*def
)
304 /* Types for all SPIR-V SSA values are set as part of a pre-pass so the
305 * type will be valid by the time we get here.
307 struct vtn_type
*type
= vtn_get_value_type(b
, value_id
);
308 vtn_fail_if(def
->num_components
!= glsl_get_vector_elements(type
->type
) ||
309 def
->bit_size
!= glsl_get_bit_size(type
->type
),
310 "Mismatch between NIR and SPIR-V type.");
311 struct vtn_ssa_value
*ssa
= vtn_create_ssa_value(b
, type
->type
);
313 return vtn_push_ssa_value(b
, value_id
, ssa
);
316 static nir_deref_instr
*
317 vtn_get_image(struct vtn_builder
*b
, uint32_t value_id
)
319 struct vtn_type
*type
= vtn_get_value_type(b
, value_id
);
320 vtn_assert(type
->base_type
== vtn_base_type_image
);
321 return nir_build_deref_cast(&b
->nb
, vtn_get_nir_ssa(b
, value_id
),
322 nir_var_uniform
, type
->glsl_image
, 0);
326 vtn_push_image(struct vtn_builder
*b
, uint32_t value_id
,
327 nir_deref_instr
*deref
, bool propagate_non_uniform
)
329 struct vtn_type
*type
= vtn_get_value_type(b
, value_id
);
330 vtn_assert(type
->base_type
== vtn_base_type_image
);
331 struct vtn_value
*value
= vtn_push_nir_ssa(b
, value_id
, &deref
->dest
.ssa
);
332 value
->propagated_non_uniform
= propagate_non_uniform
;
335 static nir_deref_instr
*
336 vtn_get_sampler(struct vtn_builder
*b
, uint32_t value_id
)
338 struct vtn_type
*type
= vtn_get_value_type(b
, value_id
);
339 vtn_assert(type
->base_type
== vtn_base_type_sampler
);
340 return nir_build_deref_cast(&b
->nb
, vtn_get_nir_ssa(b
, value_id
),
341 nir_var_uniform
, glsl_bare_sampler_type(), 0);
345 vtn_sampled_image_to_nir_ssa(struct vtn_builder
*b
,
346 struct vtn_sampled_image si
)
348 return nir_vec2(&b
->nb
, &si
.image
->dest
.ssa
, &si
.sampler
->dest
.ssa
);
352 vtn_push_sampled_image(struct vtn_builder
*b
, uint32_t value_id
,
353 struct vtn_sampled_image si
, bool propagate_non_uniform
)
355 struct vtn_type
*type
= vtn_get_value_type(b
, value_id
);
356 vtn_assert(type
->base_type
== vtn_base_type_sampled_image
);
357 struct vtn_value
*value
= vtn_push_nir_ssa(b
, value_id
,
358 vtn_sampled_image_to_nir_ssa(b
, si
));
359 value
->propagated_non_uniform
= propagate_non_uniform
;
362 static struct vtn_sampled_image
363 vtn_get_sampled_image(struct vtn_builder
*b
, uint32_t value_id
)
365 struct vtn_type
*type
= vtn_get_value_type(b
, value_id
);
366 vtn_assert(type
->base_type
== vtn_base_type_sampled_image
);
367 nir_ssa_def
*si_vec2
= vtn_get_nir_ssa(b
, value_id
);
369 struct vtn_sampled_image si
= { NULL
, };
370 si
.image
= nir_build_deref_cast(&b
->nb
, nir_channel(&b
->nb
, si_vec2
, 0),
372 type
->image
->glsl_image
, 0);
373 si
.sampler
= nir_build_deref_cast(&b
->nb
, nir_channel(&b
->nb
, si_vec2
, 1),
375 glsl_bare_sampler_type(), 0);
380 vtn_string_literal(struct vtn_builder
*b
, const uint32_t *words
,
381 unsigned word_count
, unsigned *words_used
)
383 /* From the SPIR-V spec:
385 * "A string is interpreted as a nul-terminated stream of characters.
386 * The character set is Unicode in the UTF-8 encoding scheme. The UTF-8
387 * octets (8-bit bytes) are packed four per word, following the
388 * little-endian convention (i.e., the first octet is in the
389 * lowest-order 8 bits of the word). The final word contains the
390 * string’s nul-termination character (0), and all contents past the
391 * end of the string in the final word are padded with 0."
393 * On big-endian, we need to byte-swap.
395 #if UTIL_ARCH_BIG_ENDIAN
397 uint32_t *copy
= ralloc_array(b
, uint32_t, word_count
);
398 for (unsigned i
= 0; i
< word_count
; i
++)
399 copy
[i
] = bswap_32(words
[i
]);
404 const char *str
= (char *)words
;
405 const char *end
= memchr(str
, 0, word_count
* 4);
406 vtn_fail_if(end
== NULL
, "String is not null-terminated");
409 *words_used
= DIV_ROUND_UP(end
- str
+ 1, sizeof(*words
));
415 vtn_foreach_instruction(struct vtn_builder
*b
, const uint32_t *start
,
416 const uint32_t *end
, vtn_instruction_handler handler
)
422 const uint32_t *w
= start
;
424 SpvOp opcode
= w
[0] & SpvOpCodeMask
;
425 unsigned count
= w
[0] >> SpvWordCountShift
;
426 vtn_assert(count
>= 1 && w
+ count
<= end
);
428 b
->spirv_offset
= (uint8_t *)w
- (uint8_t *)b
->spirv
;
432 break; /* Do nothing */
435 b
->file
= vtn_value(b
, w
[1], vtn_value_type_string
)->str
;
447 if (!handler(b
, opcode
, w
, count
))
465 vtn_handle_non_semantic_instruction(struct vtn_builder
*b
, SpvOp ext_opcode
,
466 const uint32_t *w
, unsigned count
)
473 vtn_handle_extension(struct vtn_builder
*b
, SpvOp opcode
,
474 const uint32_t *w
, unsigned count
)
477 case SpvOpExtInstImport
: {
478 struct vtn_value
*val
= vtn_push_value(b
, w
[1], vtn_value_type_extension
);
479 const char *ext
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
480 if (strcmp(ext
, "GLSL.std.450") == 0) {
481 val
->ext_handler
= vtn_handle_glsl450_instruction
;
482 } else if ((strcmp(ext
, "SPV_AMD_gcn_shader") == 0)
483 && (b
->options
&& b
->options
->caps
.amd_gcn_shader
)) {
484 val
->ext_handler
= vtn_handle_amd_gcn_shader_instruction
;
485 } else if ((strcmp(ext
, "SPV_AMD_shader_ballot") == 0)
486 && (b
->options
&& b
->options
->caps
.amd_shader_ballot
)) {
487 val
->ext_handler
= vtn_handle_amd_shader_ballot_instruction
;
488 } else if ((strcmp(ext
, "SPV_AMD_shader_trinary_minmax") == 0)
489 && (b
->options
&& b
->options
->caps
.amd_trinary_minmax
)) {
490 val
->ext_handler
= vtn_handle_amd_shader_trinary_minmax_instruction
;
491 } else if ((strcmp(ext
, "SPV_AMD_shader_explicit_vertex_parameter") == 0)
492 && (b
->options
&& b
->options
->caps
.amd_shader_explicit_vertex_parameter
)) {
493 val
->ext_handler
= vtn_handle_amd_shader_explicit_vertex_parameter_instruction
;
494 } else if (strcmp(ext
, "OpenCL.std") == 0) {
495 val
->ext_handler
= vtn_handle_opencl_instruction
;
496 } else if (strstr(ext
, "NonSemantic.") == ext
) {
497 val
->ext_handler
= vtn_handle_non_semantic_instruction
;
499 vtn_fail("Unsupported extension: %s", ext
);
505 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
506 bool handled
= val
->ext_handler(b
, w
[4], w
, count
);
512 vtn_fail_with_opcode("Unhandled opcode", opcode
);
517 _foreach_decoration_helper(struct vtn_builder
*b
,
518 struct vtn_value
*base_value
,
520 struct vtn_value
*value
,
521 vtn_decoration_foreach_cb cb
, void *data
)
523 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
525 if (dec
->scope
== VTN_DEC_DECORATION
) {
526 member
= parent_member
;
527 } else if (dec
->scope
>= VTN_DEC_STRUCT_MEMBER0
) {
528 vtn_fail_if(value
->value_type
!= vtn_value_type_type
||
529 value
->type
->base_type
!= vtn_base_type_struct
,
530 "OpMemberDecorate and OpGroupMemberDecorate are only "
531 "allowed on OpTypeStruct");
532 /* This means we haven't recursed yet */
533 assert(value
== base_value
);
535 member
= dec
->scope
- VTN_DEC_STRUCT_MEMBER0
;
537 vtn_fail_if(member
>= base_value
->type
->length
,
538 "OpMemberDecorate specifies member %d but the "
539 "OpTypeStruct has only %u members",
540 member
, base_value
->type
->length
);
542 /* Not a decoration */
543 assert(dec
->scope
== VTN_DEC_EXECUTION_MODE
);
548 assert(dec
->group
->value_type
== vtn_value_type_decoration_group
);
549 _foreach_decoration_helper(b
, base_value
, member
, dec
->group
,
552 cb(b
, base_value
, member
, dec
, data
);
557 /** Iterates (recursively if needed) over all of the decorations on a value
559 * This function iterates over all of the decorations applied to a given
560 * value. If it encounters a decoration group, it recurses into the group
561 * and iterates over all of those decorations as well.
564 vtn_foreach_decoration(struct vtn_builder
*b
, struct vtn_value
*value
,
565 vtn_decoration_foreach_cb cb
, void *data
)
567 _foreach_decoration_helper(b
, value
, -1, value
, cb
, data
);
571 vtn_foreach_execution_mode(struct vtn_builder
*b
, struct vtn_value
*value
,
572 vtn_execution_mode_foreach_cb cb
, void *data
)
574 for (struct vtn_decoration
*dec
= value
->decoration
; dec
; dec
= dec
->next
) {
575 if (dec
->scope
!= VTN_DEC_EXECUTION_MODE
)
578 assert(dec
->group
== NULL
);
579 cb(b
, value
, dec
, data
);
584 vtn_handle_decoration(struct vtn_builder
*b
, SpvOp opcode
,
585 const uint32_t *w
, unsigned count
)
587 const uint32_t *w_end
= w
+ count
;
588 const uint32_t target
= w
[1];
592 case SpvOpDecorationGroup
:
593 vtn_push_value(b
, target
, vtn_value_type_decoration_group
);
597 case SpvOpDecorateId
:
598 case SpvOpMemberDecorate
:
599 case SpvOpDecorateString
:
600 case SpvOpMemberDecorateString
:
601 case SpvOpExecutionMode
:
602 case SpvOpExecutionModeId
: {
603 struct vtn_value
*val
= vtn_untyped_value(b
, target
);
605 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
608 case SpvOpDecorateId
:
609 case SpvOpDecorateString
:
610 dec
->scope
= VTN_DEC_DECORATION
;
612 case SpvOpMemberDecorate
:
613 case SpvOpMemberDecorateString
:
614 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(w
++);
615 vtn_fail_if(dec
->scope
< VTN_DEC_STRUCT_MEMBER0
, /* overflow */
616 "Member argument of OpMemberDecorate too large");
618 case SpvOpExecutionMode
:
619 case SpvOpExecutionModeId
:
620 dec
->scope
= VTN_DEC_EXECUTION_MODE
;
623 unreachable("Invalid decoration opcode");
625 dec
->decoration
= *(w
++);
628 /* Link into the list */
629 dec
->next
= val
->decoration
;
630 val
->decoration
= dec
;
634 case SpvOpGroupMemberDecorate
:
635 case SpvOpGroupDecorate
: {
636 struct vtn_value
*group
=
637 vtn_value(b
, target
, vtn_value_type_decoration_group
);
639 for (; w
< w_end
; w
++) {
640 struct vtn_value
*val
= vtn_untyped_value(b
, *w
);
641 struct vtn_decoration
*dec
= rzalloc(b
, struct vtn_decoration
);
644 if (opcode
== SpvOpGroupDecorate
) {
645 dec
->scope
= VTN_DEC_DECORATION
;
647 dec
->scope
= VTN_DEC_STRUCT_MEMBER0
+ *(++w
);
648 vtn_fail_if(dec
->scope
< 0, /* Check for overflow */
649 "Member argument of OpGroupMemberDecorate too large");
652 /* Link into the list */
653 dec
->next
= val
->decoration
;
654 val
->decoration
= dec
;
660 unreachable("Unhandled opcode");
664 struct member_decoration_ctx
{
666 struct glsl_struct_field
*fields
;
667 struct vtn_type
*type
;
671 * Returns true if the given type contains a struct decorated Block or
675 vtn_type_contains_block(struct vtn_builder
*b
, struct vtn_type
*type
)
677 switch (type
->base_type
) {
678 case vtn_base_type_array
:
679 return vtn_type_contains_block(b
, type
->array_element
);
680 case vtn_base_type_struct
:
681 if (type
->block
|| type
->buffer_block
)
683 for (unsigned i
= 0; i
< type
->length
; i
++) {
684 if (vtn_type_contains_block(b
, type
->members
[i
]))
693 /** Returns true if two types are "compatible", i.e. you can do an OpLoad,
694 * OpStore, or OpCopyMemory between them without breaking anything.
695 * Technically, the SPIR-V rules require the exact same type ID but this lets
696 * us internally be a bit looser.
699 vtn_types_compatible(struct vtn_builder
*b
,
700 struct vtn_type
*t1
, struct vtn_type
*t2
)
702 if (t1
->id
== t2
->id
)
705 if (t1
->base_type
!= t2
->base_type
)
708 switch (t1
->base_type
) {
709 case vtn_base_type_void
:
710 case vtn_base_type_scalar
:
711 case vtn_base_type_vector
:
712 case vtn_base_type_matrix
:
713 case vtn_base_type_image
:
714 case vtn_base_type_sampler
:
715 case vtn_base_type_sampled_image
:
716 return t1
->type
== t2
->type
;
718 case vtn_base_type_array
:
719 return t1
->length
== t2
->length
&&
720 vtn_types_compatible(b
, t1
->array_element
, t2
->array_element
);
722 case vtn_base_type_pointer
:
723 return vtn_types_compatible(b
, t1
->deref
, t2
->deref
);
725 case vtn_base_type_struct
:
726 if (t1
->length
!= t2
->length
)
729 for (unsigned i
= 0; i
< t1
->length
; i
++) {
730 if (!vtn_types_compatible(b
, t1
->members
[i
], t2
->members
[i
]))
735 case vtn_base_type_function
:
736 /* This case shouldn't get hit since you can't copy around function
737 * types. Just require them to be identical.
742 vtn_fail("Invalid base type");
746 vtn_type_without_array(struct vtn_type
*type
)
748 while (type
->base_type
== vtn_base_type_array
)
749 type
= type
->array_element
;
753 /* does a shallow copy of a vtn_type */
755 static struct vtn_type
*
756 vtn_type_copy(struct vtn_builder
*b
, struct vtn_type
*src
)
758 struct vtn_type
*dest
= ralloc(b
, struct vtn_type
);
761 switch (src
->base_type
) {
762 case vtn_base_type_void
:
763 case vtn_base_type_scalar
:
764 case vtn_base_type_vector
:
765 case vtn_base_type_matrix
:
766 case vtn_base_type_array
:
767 case vtn_base_type_pointer
:
768 case vtn_base_type_image
:
769 case vtn_base_type_sampler
:
770 case vtn_base_type_sampled_image
:
771 /* Nothing more to do */
774 case vtn_base_type_struct
:
775 dest
->members
= ralloc_array(b
, struct vtn_type
*, src
->length
);
776 memcpy(dest
->members
, src
->members
,
777 src
->length
* sizeof(src
->members
[0]));
779 dest
->offsets
= ralloc_array(b
, unsigned, src
->length
);
780 memcpy(dest
->offsets
, src
->offsets
,
781 src
->length
* sizeof(src
->offsets
[0]));
784 case vtn_base_type_function
:
785 dest
->params
= ralloc_array(b
, struct vtn_type
*, src
->length
);
786 memcpy(dest
->params
, src
->params
, src
->length
* sizeof(src
->params
[0]));
793 static const struct glsl_type
*
794 wrap_type_in_array(const struct glsl_type
*type
,
795 const struct glsl_type
*array_type
)
797 if (!glsl_type_is_array(array_type
))
800 const struct glsl_type
*elem_type
=
801 wrap_type_in_array(type
, glsl_get_array_element(array_type
));
802 return glsl_array_type(elem_type
, glsl_get_length(array_type
),
803 glsl_get_explicit_stride(array_type
));
807 vtn_type_needs_explicit_layout(struct vtn_builder
*b
, enum vtn_variable_mode mode
)
809 /* For OpenCL we never want to strip the info from the types, and it makes
810 * type comparisons easier in later stages.
812 if (b
->options
->environment
== NIR_SPIRV_OPENCL
)
816 case vtn_variable_mode_input
:
817 case vtn_variable_mode_output
:
818 /* Layout decorations kept because we need offsets for XFB arrays of
821 return b
->shader
->info
.has_transform_feedback_varyings
;
823 case vtn_variable_mode_ssbo
:
824 case vtn_variable_mode_phys_ssbo
:
825 case vtn_variable_mode_ubo
:
833 const struct glsl_type
*
834 vtn_type_get_nir_type(struct vtn_builder
*b
, struct vtn_type
*type
,
835 enum vtn_variable_mode mode
)
837 if (mode
== vtn_variable_mode_atomic_counter
) {
838 vtn_fail_if(glsl_without_array(type
->type
) != glsl_uint_type(),
839 "Variables in the AtomicCounter storage class should be "
840 "(possibly arrays of arrays of) uint.");
841 return wrap_type_in_array(glsl_atomic_uint_type(), type
->type
);
844 if (mode
== vtn_variable_mode_uniform
) {
845 switch (type
->base_type
) {
846 case vtn_base_type_array
: {
847 const struct glsl_type
*elem_type
=
848 vtn_type_get_nir_type(b
, type
->array_element
, mode
);
850 return glsl_array_type(elem_type
, type
->length
,
851 glsl_get_explicit_stride(type
->type
));
854 case vtn_base_type_struct
: {
855 bool need_new_struct
= false;
856 const uint32_t num_fields
= type
->length
;
857 NIR_VLA(struct glsl_struct_field
, fields
, num_fields
);
858 for (unsigned i
= 0; i
< num_fields
; i
++) {
859 fields
[i
] = *glsl_get_struct_field_data(type
->type
, i
);
860 const struct glsl_type
*field_nir_type
=
861 vtn_type_get_nir_type(b
, type
->members
[i
], mode
);
862 if (fields
[i
].type
!= field_nir_type
) {
863 fields
[i
].type
= field_nir_type
;
864 need_new_struct
= true;
867 if (need_new_struct
) {
868 if (glsl_type_is_interface(type
->type
)) {
869 return glsl_interface_type(fields
, num_fields
,
870 /* packing */ 0, false,
871 glsl_get_type_name(type
->type
));
873 return glsl_struct_type(fields
, num_fields
,
874 glsl_get_type_name(type
->type
),
875 glsl_struct_type_is_packed(type
->type
));
878 /* No changes, just pass it on */
883 case vtn_base_type_image
:
884 return type
->glsl_image
;
886 case vtn_base_type_sampler
:
887 return glsl_bare_sampler_type();
889 case vtn_base_type_sampled_image
:
890 return type
->image
->glsl_image
;
897 /* Layout decorations are allowed but ignored in certain conditions,
898 * to allow SPIR-V generators perform type deduplication. Discard
899 * unnecessary ones when passing to NIR.
901 if (!vtn_type_needs_explicit_layout(b
, mode
))
902 return glsl_get_bare_type(type
->type
);
907 static struct vtn_type
*
908 mutable_matrix_member(struct vtn_builder
*b
, struct vtn_type
*type
, int member
)
910 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
911 type
= type
->members
[member
];
913 /* We may have an array of matrices.... Oh, joy! */
914 while (glsl_type_is_array(type
->type
)) {
915 type
->array_element
= vtn_type_copy(b
, type
->array_element
);
916 type
= type
->array_element
;
919 vtn_assert(glsl_type_is_matrix(type
->type
));
925 vtn_handle_access_qualifier(struct vtn_builder
*b
, struct vtn_type
*type
,
926 int member
, enum gl_access_qualifier access
)
928 type
->members
[member
] = vtn_type_copy(b
, type
->members
[member
]);
929 type
= type
->members
[member
];
931 type
->access
|= access
;
935 array_stride_decoration_cb(struct vtn_builder
*b
,
936 struct vtn_value
*val
, int member
,
937 const struct vtn_decoration
*dec
, void *void_ctx
)
939 struct vtn_type
*type
= val
->type
;
941 if (dec
->decoration
== SpvDecorationArrayStride
) {
942 if (vtn_type_contains_block(b
, type
)) {
943 vtn_warn("The ArrayStride decoration cannot be applied to an array "
944 "type which contains a structure type decorated Block "
946 /* Ignore the decoration */
948 vtn_fail_if(dec
->operands
[0] == 0, "ArrayStride must be non-zero");
949 type
->stride
= dec
->operands
[0];
955 struct_member_decoration_cb(struct vtn_builder
*b
,
956 UNUSED
struct vtn_value
*val
, int member
,
957 const struct vtn_decoration
*dec
, void *void_ctx
)
959 struct member_decoration_ctx
*ctx
= void_ctx
;
964 assert(member
< ctx
->num_fields
);
966 switch (dec
->decoration
) {
967 case SpvDecorationRelaxedPrecision
:
968 case SpvDecorationUniform
:
969 case SpvDecorationUniformId
:
970 break; /* FIXME: Do nothing with this for now. */
971 case SpvDecorationNonWritable
:
972 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_NON_WRITEABLE
);
974 case SpvDecorationNonReadable
:
975 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_NON_READABLE
);
977 case SpvDecorationVolatile
:
978 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_VOLATILE
);
980 case SpvDecorationCoherent
:
981 vtn_handle_access_qualifier(b
, ctx
->type
, member
, ACCESS_COHERENT
);
983 case SpvDecorationNoPerspective
:
984 ctx
->fields
[member
].interpolation
= INTERP_MODE_NOPERSPECTIVE
;
986 case SpvDecorationFlat
:
987 ctx
->fields
[member
].interpolation
= INTERP_MODE_FLAT
;
989 case SpvDecorationExplicitInterpAMD
:
990 ctx
->fields
[member
].interpolation
= INTERP_MODE_EXPLICIT
;
992 case SpvDecorationCentroid
:
993 ctx
->fields
[member
].centroid
= true;
995 case SpvDecorationSample
:
996 ctx
->fields
[member
].sample
= true;
998 case SpvDecorationStream
:
999 /* This is handled later by var_decoration_cb in vtn_variables.c */
1001 case SpvDecorationLocation
:
1002 ctx
->fields
[member
].location
= dec
->operands
[0];
1004 case SpvDecorationComponent
:
1005 break; /* FIXME: What should we do with these? */
1006 case SpvDecorationBuiltIn
:
1007 ctx
->type
->members
[member
] = vtn_type_copy(b
, ctx
->type
->members
[member
]);
1008 ctx
->type
->members
[member
]->is_builtin
= true;
1009 ctx
->type
->members
[member
]->builtin
= dec
->operands
[0];
1010 ctx
->type
->builtin_block
= true;
1012 case SpvDecorationOffset
:
1013 ctx
->type
->offsets
[member
] = dec
->operands
[0];
1014 ctx
->fields
[member
].offset
= dec
->operands
[0];
1016 case SpvDecorationMatrixStride
:
1017 /* Handled as a second pass */
1019 case SpvDecorationColMajor
:
1020 break; /* Nothing to do here. Column-major is the default. */
1021 case SpvDecorationRowMajor
:
1022 mutable_matrix_member(b
, ctx
->type
, member
)->row_major
= true;
1025 case SpvDecorationPatch
:
1028 case SpvDecorationSpecId
:
1029 case SpvDecorationBlock
:
1030 case SpvDecorationBufferBlock
:
1031 case SpvDecorationArrayStride
:
1032 case SpvDecorationGLSLShared
:
1033 case SpvDecorationGLSLPacked
:
1034 case SpvDecorationInvariant
:
1035 case SpvDecorationRestrict
:
1036 case SpvDecorationAliased
:
1037 case SpvDecorationConstant
:
1038 case SpvDecorationIndex
:
1039 case SpvDecorationBinding
:
1040 case SpvDecorationDescriptorSet
:
1041 case SpvDecorationLinkageAttributes
:
1042 case SpvDecorationNoContraction
:
1043 case SpvDecorationInputAttachmentIndex
:
1044 case SpvDecorationCPacked
:
1045 vtn_warn("Decoration not allowed on struct members: %s",
1046 spirv_decoration_to_string(dec
->decoration
));
1049 case SpvDecorationXfbBuffer
:
1050 case SpvDecorationXfbStride
:
1051 /* This is handled later by var_decoration_cb in vtn_variables.c */
1054 case SpvDecorationSaturatedConversion
:
1055 case SpvDecorationFuncParamAttr
:
1056 case SpvDecorationFPRoundingMode
:
1057 case SpvDecorationFPFastMathMode
:
1058 case SpvDecorationAlignment
:
1059 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
) {
1060 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1061 spirv_decoration_to_string(dec
->decoration
));
1065 case SpvDecorationUserSemantic
:
1066 case SpvDecorationUserTypeGOOGLE
:
1067 /* User semantic decorations can safely be ignored by the driver. */
1071 vtn_fail_with_decoration("Unhandled decoration", dec
->decoration
);
1075 /** Chases the array type all the way down to the tail and rewrites the
1076 * glsl_types to be based off the tail's glsl_type.
1079 vtn_array_type_rewrite_glsl_type(struct vtn_type
*type
)
1081 if (type
->base_type
!= vtn_base_type_array
)
1084 vtn_array_type_rewrite_glsl_type(type
->array_element
);
1086 type
->type
= glsl_array_type(type
->array_element
->type
,
1087 type
->length
, type
->stride
);
1090 /* Matrix strides are handled as a separate pass because we need to know
1091 * whether the matrix is row-major or not first.
1094 struct_member_matrix_stride_cb(struct vtn_builder
*b
,
1095 UNUSED
struct vtn_value
*val
, int member
,
1096 const struct vtn_decoration
*dec
,
1099 if (dec
->decoration
!= SpvDecorationMatrixStride
)
1102 vtn_fail_if(member
< 0,
1103 "The MatrixStride decoration is only allowed on members "
1105 vtn_fail_if(dec
->operands
[0] == 0, "MatrixStride must be non-zero");
1107 struct member_decoration_ctx
*ctx
= void_ctx
;
1109 struct vtn_type
*mat_type
= mutable_matrix_member(b
, ctx
->type
, member
);
1110 if (mat_type
->row_major
) {
1111 mat_type
->array_element
= vtn_type_copy(b
, mat_type
->array_element
);
1112 mat_type
->stride
= mat_type
->array_element
->stride
;
1113 mat_type
->array_element
->stride
= dec
->operands
[0];
1115 mat_type
->type
= glsl_explicit_matrix_type(mat_type
->type
,
1116 dec
->operands
[0], true);
1117 mat_type
->array_element
->type
= glsl_get_column_type(mat_type
->type
);
1119 vtn_assert(mat_type
->array_element
->stride
> 0);
1120 mat_type
->stride
= dec
->operands
[0];
1122 mat_type
->type
= glsl_explicit_matrix_type(mat_type
->type
,
1123 dec
->operands
[0], false);
1126 /* Now that we've replaced the glsl_type with a properly strided matrix
1127 * type, rewrite the member type so that it's an array of the proper kind
1130 vtn_array_type_rewrite_glsl_type(ctx
->type
->members
[member
]);
1131 ctx
->fields
[member
].type
= ctx
->type
->members
[member
]->type
;
1135 struct_block_decoration_cb(struct vtn_builder
*b
,
1136 struct vtn_value
*val
, int member
,
1137 const struct vtn_decoration
*dec
, void *ctx
)
1142 struct vtn_type
*type
= val
->type
;
1143 if (dec
->decoration
== SpvDecorationBlock
)
1145 else if (dec
->decoration
== SpvDecorationBufferBlock
)
1146 type
->buffer_block
= true;
1150 type_decoration_cb(struct vtn_builder
*b
,
1151 struct vtn_value
*val
, int member
,
1152 const struct vtn_decoration
*dec
, UNUSED
void *ctx
)
1154 struct vtn_type
*type
= val
->type
;
1157 /* This should have been handled by OpTypeStruct */
1158 assert(val
->type
->base_type
== vtn_base_type_struct
);
1159 assert(member
>= 0 && member
< val
->type
->length
);
1163 switch (dec
->decoration
) {
1164 case SpvDecorationArrayStride
:
1165 vtn_assert(type
->base_type
== vtn_base_type_array
||
1166 type
->base_type
== vtn_base_type_pointer
);
1168 case SpvDecorationBlock
:
1169 vtn_assert(type
->base_type
== vtn_base_type_struct
);
1170 vtn_assert(type
->block
);
1172 case SpvDecorationBufferBlock
:
1173 vtn_assert(type
->base_type
== vtn_base_type_struct
);
1174 vtn_assert(type
->buffer_block
);
1176 case SpvDecorationGLSLShared
:
1177 case SpvDecorationGLSLPacked
:
1178 /* Ignore these, since we get explicit offsets anyways */
1181 case SpvDecorationRowMajor
:
1182 case SpvDecorationColMajor
:
1183 case SpvDecorationMatrixStride
:
1184 case SpvDecorationBuiltIn
:
1185 case SpvDecorationNoPerspective
:
1186 case SpvDecorationFlat
:
1187 case SpvDecorationPatch
:
1188 case SpvDecorationCentroid
:
1189 case SpvDecorationSample
:
1190 case SpvDecorationExplicitInterpAMD
:
1191 case SpvDecorationVolatile
:
1192 case SpvDecorationCoherent
:
1193 case SpvDecorationNonWritable
:
1194 case SpvDecorationNonReadable
:
1195 case SpvDecorationUniform
:
1196 case SpvDecorationUniformId
:
1197 case SpvDecorationLocation
:
1198 case SpvDecorationComponent
:
1199 case SpvDecorationOffset
:
1200 case SpvDecorationXfbBuffer
:
1201 case SpvDecorationXfbStride
:
1202 case SpvDecorationUserSemantic
:
1203 vtn_warn("Decoration only allowed for struct members: %s",
1204 spirv_decoration_to_string(dec
->decoration
));
1207 case SpvDecorationStream
:
1208 /* We don't need to do anything here, as stream is filled up when
1209 * aplying the decoration to a variable, just check that if it is not a
1210 * struct member, it should be a struct.
1212 vtn_assert(type
->base_type
== vtn_base_type_struct
);
1215 case SpvDecorationRelaxedPrecision
:
1216 case SpvDecorationSpecId
:
1217 case SpvDecorationInvariant
:
1218 case SpvDecorationRestrict
:
1219 case SpvDecorationAliased
:
1220 case SpvDecorationConstant
:
1221 case SpvDecorationIndex
:
1222 case SpvDecorationBinding
:
1223 case SpvDecorationDescriptorSet
:
1224 case SpvDecorationLinkageAttributes
:
1225 case SpvDecorationNoContraction
:
1226 case SpvDecorationInputAttachmentIndex
:
1227 vtn_warn("Decoration not allowed on types: %s",
1228 spirv_decoration_to_string(dec
->decoration
));
1231 case SpvDecorationCPacked
:
1232 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
)
1233 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1234 spirv_decoration_to_string(dec
->decoration
));
1236 type
->packed
= true;
1239 case SpvDecorationSaturatedConversion
:
1240 case SpvDecorationFuncParamAttr
:
1241 case SpvDecorationFPRoundingMode
:
1242 case SpvDecorationFPFastMathMode
:
1243 case SpvDecorationAlignment
:
1244 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1245 spirv_decoration_to_string(dec
->decoration
));
1248 case SpvDecorationUserTypeGOOGLE
:
1249 /* User semantic decorations can safely be ignored by the driver. */
1253 vtn_fail_with_decoration("Unhandled decoration", dec
->decoration
);
1258 translate_image_format(struct vtn_builder
*b
, SpvImageFormat format
)
1261 case SpvImageFormatUnknown
: return PIPE_FORMAT_NONE
;
1262 case SpvImageFormatRgba32f
: return PIPE_FORMAT_R32G32B32A32_FLOAT
;
1263 case SpvImageFormatRgba16f
: return PIPE_FORMAT_R16G16B16A16_FLOAT
;
1264 case SpvImageFormatR32f
: return PIPE_FORMAT_R32_FLOAT
;
1265 case SpvImageFormatRgba8
: return PIPE_FORMAT_R8G8B8A8_UNORM
;
1266 case SpvImageFormatRgba8Snorm
: return PIPE_FORMAT_R8G8B8A8_SNORM
;
1267 case SpvImageFormatRg32f
: return PIPE_FORMAT_R32G32_FLOAT
;
1268 case SpvImageFormatRg16f
: return PIPE_FORMAT_R16G16_FLOAT
;
1269 case SpvImageFormatR11fG11fB10f
: return PIPE_FORMAT_R11G11B10_FLOAT
;
1270 case SpvImageFormatR16f
: return PIPE_FORMAT_R16_FLOAT
;
1271 case SpvImageFormatRgba16
: return PIPE_FORMAT_R16G16B16A16_UNORM
;
1272 case SpvImageFormatRgb10A2
: return PIPE_FORMAT_R10G10B10A2_UNORM
;
1273 case SpvImageFormatRg16
: return PIPE_FORMAT_R16G16_UNORM
;
1274 case SpvImageFormatRg8
: return PIPE_FORMAT_R8G8_UNORM
;
1275 case SpvImageFormatR16
: return PIPE_FORMAT_R16_UNORM
;
1276 case SpvImageFormatR8
: return PIPE_FORMAT_R8_UNORM
;
1277 case SpvImageFormatRgba16Snorm
: return PIPE_FORMAT_R16G16B16A16_SNORM
;
1278 case SpvImageFormatRg16Snorm
: return PIPE_FORMAT_R16G16_SNORM
;
1279 case SpvImageFormatRg8Snorm
: return PIPE_FORMAT_R8G8_SNORM
;
1280 case SpvImageFormatR16Snorm
: return PIPE_FORMAT_R16_SNORM
;
1281 case SpvImageFormatR8Snorm
: return PIPE_FORMAT_R8_SNORM
;
1282 case SpvImageFormatRgba32i
: return PIPE_FORMAT_R32G32B32A32_SINT
;
1283 case SpvImageFormatRgba16i
: return PIPE_FORMAT_R16G16B16A16_SINT
;
1284 case SpvImageFormatRgba8i
: return PIPE_FORMAT_R8G8B8A8_SINT
;
1285 case SpvImageFormatR32i
: return PIPE_FORMAT_R32_SINT
;
1286 case SpvImageFormatRg32i
: return PIPE_FORMAT_R32G32_SINT
;
1287 case SpvImageFormatRg16i
: return PIPE_FORMAT_R16G16_SINT
;
1288 case SpvImageFormatRg8i
: return PIPE_FORMAT_R8G8_SINT
;
1289 case SpvImageFormatR16i
: return PIPE_FORMAT_R16_SINT
;
1290 case SpvImageFormatR8i
: return PIPE_FORMAT_R8_SINT
;
1291 case SpvImageFormatRgba32ui
: return PIPE_FORMAT_R32G32B32A32_UINT
;
1292 case SpvImageFormatRgba16ui
: return PIPE_FORMAT_R16G16B16A16_UINT
;
1293 case SpvImageFormatRgba8ui
: return PIPE_FORMAT_R8G8B8A8_UINT
;
1294 case SpvImageFormatR32ui
: return PIPE_FORMAT_R32_UINT
;
1295 case SpvImageFormatRgb10a2ui
: return PIPE_FORMAT_R10G10B10A2_UINT
;
1296 case SpvImageFormatRg32ui
: return PIPE_FORMAT_R32G32_UINT
;
1297 case SpvImageFormatRg16ui
: return PIPE_FORMAT_R16G16_UINT
;
1298 case SpvImageFormatRg8ui
: return PIPE_FORMAT_R8G8_UINT
;
1299 case SpvImageFormatR16ui
: return PIPE_FORMAT_R16_UINT
;
1300 case SpvImageFormatR8ui
: return PIPE_FORMAT_R8_UINT
;
1302 vtn_fail("Invalid image format: %s (%u)",
1303 spirv_imageformat_to_string(format
), format
);
1308 vtn_handle_type(struct vtn_builder
*b
, SpvOp opcode
,
1309 const uint32_t *w
, unsigned count
)
1311 struct vtn_value
*val
= NULL
;
1313 /* In order to properly handle forward declarations, we have to defer
1314 * allocation for pointer types.
1316 if (opcode
!= SpvOpTypePointer
&& opcode
!= SpvOpTypeForwardPointer
) {
1317 val
= vtn_push_value(b
, w
[1], vtn_value_type_type
);
1318 vtn_fail_if(val
->type
!= NULL
,
1319 "Only pointers can have forward declarations");
1320 val
->type
= rzalloc(b
, struct vtn_type
);
1321 val
->type
->id
= w
[1];
1326 val
->type
->base_type
= vtn_base_type_void
;
1327 val
->type
->type
= glsl_void_type();
1330 val
->type
->base_type
= vtn_base_type_scalar
;
1331 val
->type
->type
= glsl_bool_type();
1332 val
->type
->length
= 1;
1334 case SpvOpTypeInt
: {
1335 int bit_size
= w
[2];
1336 const bool signedness
= w
[3];
1337 val
->type
->base_type
= vtn_base_type_scalar
;
1340 val
->type
->type
= (signedness
? glsl_int64_t_type() : glsl_uint64_t_type());
1343 val
->type
->type
= (signedness
? glsl_int_type() : glsl_uint_type());
1346 val
->type
->type
= (signedness
? glsl_int16_t_type() : glsl_uint16_t_type());
1349 val
->type
->type
= (signedness
? glsl_int8_t_type() : glsl_uint8_t_type());
1352 vtn_fail("Invalid int bit size: %u", bit_size
);
1354 val
->type
->length
= 1;
1358 case SpvOpTypeFloat
: {
1359 int bit_size
= w
[2];
1360 val
->type
->base_type
= vtn_base_type_scalar
;
1363 val
->type
->type
= glsl_float16_t_type();
1366 val
->type
->type
= glsl_float_type();
1369 val
->type
->type
= glsl_double_type();
1372 vtn_fail("Invalid float bit size: %u", bit_size
);
1374 val
->type
->length
= 1;
1378 case SpvOpTypeVector
: {
1379 struct vtn_type
*base
= vtn_get_type(b
, w
[2]);
1380 unsigned elems
= w
[3];
1382 vtn_fail_if(base
->base_type
!= vtn_base_type_scalar
,
1383 "Base type for OpTypeVector must be a scalar");
1384 vtn_fail_if((elems
< 2 || elems
> 4) && (elems
!= 8) && (elems
!= 16),
1385 "Invalid component count for OpTypeVector");
1387 val
->type
->base_type
= vtn_base_type_vector
;
1388 val
->type
->type
= glsl_vector_type(glsl_get_base_type(base
->type
), elems
);
1389 val
->type
->length
= elems
;
1390 val
->type
->stride
= glsl_type_is_boolean(val
->type
->type
)
1391 ? 4 : glsl_get_bit_size(base
->type
) / 8;
1392 val
->type
->array_element
= base
;
1396 case SpvOpTypeMatrix
: {
1397 struct vtn_type
*base
= vtn_get_type(b
, w
[2]);
1398 unsigned columns
= w
[3];
1400 vtn_fail_if(base
->base_type
!= vtn_base_type_vector
,
1401 "Base type for OpTypeMatrix must be a vector");
1402 vtn_fail_if(columns
< 2 || columns
> 4,
1403 "Invalid column count for OpTypeMatrix");
1405 val
->type
->base_type
= vtn_base_type_matrix
;
1406 val
->type
->type
= glsl_matrix_type(glsl_get_base_type(base
->type
),
1407 glsl_get_vector_elements(base
->type
),
1409 vtn_fail_if(glsl_type_is_error(val
->type
->type
),
1410 "Unsupported base type for OpTypeMatrix");
1411 assert(!glsl_type_is_error(val
->type
->type
));
1412 val
->type
->length
= columns
;
1413 val
->type
->array_element
= base
;
1414 val
->type
->row_major
= false;
1415 val
->type
->stride
= 0;
1419 case SpvOpTypeRuntimeArray
:
1420 case SpvOpTypeArray
: {
1421 struct vtn_type
*array_element
= vtn_get_type(b
, w
[2]);
1423 if (opcode
== SpvOpTypeRuntimeArray
) {
1424 /* A length of 0 is used to denote unsized arrays */
1425 val
->type
->length
= 0;
1427 val
->type
->length
= vtn_constant_uint(b
, w
[3]);
1430 val
->type
->base_type
= vtn_base_type_array
;
1431 val
->type
->array_element
= array_element
;
1432 if (b
->shader
->info
.stage
== MESA_SHADER_KERNEL
)
1433 val
->type
->stride
= glsl_get_cl_size(array_element
->type
);
1435 vtn_foreach_decoration(b
, val
, array_stride_decoration_cb
, NULL
);
1436 val
->type
->type
= glsl_array_type(array_element
->type
, val
->type
->length
,
1441 case SpvOpTypeStruct
: {
1442 unsigned num_fields
= count
- 2;
1443 val
->type
->base_type
= vtn_base_type_struct
;
1444 val
->type
->length
= num_fields
;
1445 val
->type
->members
= ralloc_array(b
, struct vtn_type
*, num_fields
);
1446 val
->type
->offsets
= ralloc_array(b
, unsigned, num_fields
);
1447 val
->type
->packed
= false;
1449 NIR_VLA(struct glsl_struct_field
, fields
, count
);
1450 for (unsigned i
= 0; i
< num_fields
; i
++) {
1451 val
->type
->members
[i
] = vtn_get_type(b
, w
[i
+ 2]);
1452 fields
[i
] = (struct glsl_struct_field
) {
1453 .type
= val
->type
->members
[i
]->type
,
1454 .name
= ralloc_asprintf(b
, "field%d", i
),
1460 if (b
->shader
->info
.stage
== MESA_SHADER_KERNEL
) {
1461 unsigned offset
= 0;
1462 for (unsigned i
= 0; i
< num_fields
; i
++) {
1463 offset
= align(offset
, glsl_get_cl_alignment(fields
[i
].type
));
1464 fields
[i
].offset
= offset
;
1465 offset
+= glsl_get_cl_size(fields
[i
].type
);
1469 struct member_decoration_ctx ctx
= {
1470 .num_fields
= num_fields
,
1475 vtn_foreach_decoration(b
, val
, struct_member_decoration_cb
, &ctx
);
1476 vtn_foreach_decoration(b
, val
, struct_member_matrix_stride_cb
, &ctx
);
1478 vtn_foreach_decoration(b
, val
, struct_block_decoration_cb
, NULL
);
1480 const char *name
= val
->name
;
1482 if (val
->type
->block
|| val
->type
->buffer_block
) {
1483 /* Packing will be ignored since types coming from SPIR-V are
1484 * explicitly laid out.
1486 val
->type
->type
= glsl_interface_type(fields
, num_fields
,
1487 /* packing */ 0, false,
1488 name
? name
: "block");
1490 val
->type
->type
= glsl_struct_type(fields
, num_fields
,
1491 name
? name
: "struct", false);
1496 case SpvOpTypeFunction
: {
1497 val
->type
->base_type
= vtn_base_type_function
;
1498 val
->type
->type
= NULL
;
1500 val
->type
->return_type
= vtn_get_type(b
, w
[2]);
1502 const unsigned num_params
= count
- 3;
1503 val
->type
->length
= num_params
;
1504 val
->type
->params
= ralloc_array(b
, struct vtn_type
*, num_params
);
1505 for (unsigned i
= 0; i
< count
- 3; i
++) {
1506 val
->type
->params
[i
] = vtn_get_type(b
, w
[i
+ 3]);
1511 case SpvOpTypePointer
:
1512 case SpvOpTypeForwardPointer
: {
1513 /* We can't blindly push the value because it might be a forward
1516 val
= vtn_untyped_value(b
, w
[1]);
1518 SpvStorageClass storage_class
= w
[2];
1520 if (val
->value_type
== vtn_value_type_invalid
) {
1521 val
->value_type
= vtn_value_type_type
;
1522 val
->type
= rzalloc(b
, struct vtn_type
);
1523 val
->type
->id
= w
[1];
1524 val
->type
->base_type
= vtn_base_type_pointer
;
1525 val
->type
->storage_class
= storage_class
;
1527 /* These can actually be stored to nir_variables and used as SSA
1528 * values so they need a real glsl_type.
1530 enum vtn_variable_mode mode
= vtn_storage_class_to_mode(
1531 b
, storage_class
, NULL
, NULL
);
1532 val
->type
->type
= nir_address_format_to_glsl_type(
1533 vtn_mode_to_address_format(b
, mode
));
1535 vtn_fail_if(val
->type
->storage_class
!= storage_class
,
1536 "The storage classes of an OpTypePointer and any "
1537 "OpTypeForwardPointers that provide forward "
1538 "declarations of it must match.");
1541 if (opcode
== SpvOpTypePointer
) {
1542 vtn_fail_if(val
->type
->deref
!= NULL
,
1543 "While OpTypeForwardPointer can be used to provide a "
1544 "forward declaration of a pointer, OpTypePointer can "
1545 "only be used once for a given id.");
1547 val
->type
->deref
= vtn_get_type(b
, w
[3]);
1549 /* Only certain storage classes use ArrayStride. The others (in
1550 * particular Workgroup) are expected to be laid out by the driver.
1552 switch (storage_class
) {
1553 case SpvStorageClassUniform
:
1554 case SpvStorageClassPushConstant
:
1555 case SpvStorageClassStorageBuffer
:
1556 case SpvStorageClassPhysicalStorageBuffer
:
1557 vtn_foreach_decoration(b
, val
, array_stride_decoration_cb
, NULL
);
1560 /* Nothing to do. */
1564 if (b
->physical_ptrs
) {
1565 switch (storage_class
) {
1566 case SpvStorageClassFunction
:
1567 case SpvStorageClassWorkgroup
:
1568 case SpvStorageClassCrossWorkgroup
:
1569 case SpvStorageClassUniformConstant
:
1570 val
->type
->stride
= align(glsl_get_cl_size(val
->type
->deref
->type
),
1571 glsl_get_cl_alignment(val
->type
->deref
->type
));
1581 case SpvOpTypeImage
: {
1582 val
->type
->base_type
= vtn_base_type_image
;
1584 /* Images are represented in NIR as a scalar SSA value that is the
1585 * result of a deref instruction. An OpLoad on an OpTypeImage pointer
1586 * from UniformConstant memory just takes the NIR deref from the pointer
1587 * and turns it into an SSA value.
1589 val
->type
->type
= nir_address_format_to_glsl_type(
1590 vtn_mode_to_address_format(b
, vtn_variable_mode_function
));
1592 const struct vtn_type
*sampled_type
= vtn_get_type(b
, w
[2]);
1593 if (b
->shader
->info
.stage
== MESA_SHADER_KERNEL
) {
1594 vtn_fail_if(sampled_type
->base_type
!= vtn_base_type_void
,
1595 "Sampled type of OpTypeImage must be void for kernels");
1597 vtn_fail_if(sampled_type
->base_type
!= vtn_base_type_scalar
||
1598 glsl_get_bit_size(sampled_type
->type
) != 32,
1599 "Sampled type of OpTypeImage must be a 32-bit scalar");
1602 enum glsl_sampler_dim dim
;
1603 switch ((SpvDim
)w
[3]) {
1604 case SpvDim1D
: dim
= GLSL_SAMPLER_DIM_1D
; break;
1605 case SpvDim2D
: dim
= GLSL_SAMPLER_DIM_2D
; break;
1606 case SpvDim3D
: dim
= GLSL_SAMPLER_DIM_3D
; break;
1607 case SpvDimCube
: dim
= GLSL_SAMPLER_DIM_CUBE
; break;
1608 case SpvDimRect
: dim
= GLSL_SAMPLER_DIM_RECT
; break;
1609 case SpvDimBuffer
: dim
= GLSL_SAMPLER_DIM_BUF
; break;
1610 case SpvDimSubpassData
: dim
= GLSL_SAMPLER_DIM_SUBPASS
; break;
1612 vtn_fail("Invalid SPIR-V image dimensionality: %s (%u)",
1613 spirv_dim_to_string((SpvDim
)w
[3]), w
[3]);
1616 /* w[4]: as per Vulkan spec "Validation Rules within a Module",
1617 * The “Depth” operand of OpTypeImage is ignored.
1619 bool is_array
= w
[5];
1620 bool multisampled
= w
[6];
1621 unsigned sampled
= w
[7];
1622 SpvImageFormat format
= w
[8];
1625 val
->type
->access_qualifier
= w
[9];
1626 else if (b
->shader
->info
.stage
== MESA_SHADER_KERNEL
)
1627 /* Per the CL C spec: If no qualifier is provided, read_only is assumed. */
1628 val
->type
->access_qualifier
= SpvAccessQualifierReadOnly
;
1630 val
->type
->access_qualifier
= SpvAccessQualifierReadWrite
;
1633 if (dim
== GLSL_SAMPLER_DIM_2D
)
1634 dim
= GLSL_SAMPLER_DIM_MS
;
1635 else if (dim
== GLSL_SAMPLER_DIM_SUBPASS
)
1636 dim
= GLSL_SAMPLER_DIM_SUBPASS_MS
;
1638 vtn_fail("Unsupported multisampled image type");
1641 val
->type
->image_format
= translate_image_format(b
, format
);
1643 enum glsl_base_type sampled_base_type
=
1644 glsl_get_base_type(sampled_type
->type
);
1646 val
->type
->glsl_image
= glsl_sampler_type(dim
, false, is_array
,
1648 } else if (sampled
== 2) {
1649 val
->type
->glsl_image
= glsl_image_type(dim
, is_array
,
1651 } else if (b
->shader
->info
.stage
== MESA_SHADER_KERNEL
) {
1652 val
->type
->glsl_image
= glsl_image_type(dim
, is_array
,
1655 vtn_fail("We need to know if the image will be sampled");
1660 case SpvOpTypeSampledImage
: {
1661 val
->type
->base_type
= vtn_base_type_sampled_image
;
1662 val
->type
->image
= vtn_get_type(b
, w
[2]);
1664 /* Sampled images are represented NIR as a vec2 SSA value where each
1665 * component is the result of a deref instruction. The first component
1666 * is the image and the second is the sampler. An OpLoad on an
1667 * OpTypeSampledImage pointer from UniformConstant memory just takes
1668 * the NIR deref from the pointer and duplicates it to both vector
1671 nir_address_format addr_format
=
1672 vtn_mode_to_address_format(b
, vtn_variable_mode_function
);
1673 assert(nir_address_format_num_components(addr_format
) == 1);
1674 unsigned bit_size
= nir_address_format_bit_size(addr_format
);
1675 assert(bit_size
== 32 || bit_size
== 64);
1677 enum glsl_base_type base_type
=
1678 bit_size
== 32 ? GLSL_TYPE_UINT
: GLSL_TYPE_UINT64
;
1679 val
->type
->type
= glsl_vector_type(base_type
, 2);
1683 case SpvOpTypeSampler
:
1684 val
->type
->base_type
= vtn_base_type_sampler
;
1686 /* Samplers are represented in NIR as a scalar SSA value that is the
1687 * result of a deref instruction. An OpLoad on an OpTypeSampler pointer
1688 * from UniformConstant memory just takes the NIR deref from the pointer
1689 * and turns it into an SSA value.
1691 val
->type
->type
= nir_address_format_to_glsl_type(
1692 vtn_mode_to_address_format(b
, vtn_variable_mode_function
));
1695 case SpvOpTypeOpaque
:
1696 case SpvOpTypeEvent
:
1697 case SpvOpTypeDeviceEvent
:
1698 case SpvOpTypeReserveId
:
1699 case SpvOpTypeQueue
:
1702 vtn_fail_with_opcode("Unhandled opcode", opcode
);
1705 vtn_foreach_decoration(b
, val
, type_decoration_cb
, NULL
);
1707 if (val
->type
->base_type
== vtn_base_type_struct
&&
1708 (val
->type
->block
|| val
->type
->buffer_block
)) {
1709 for (unsigned i
= 0; i
< val
->type
->length
; i
++) {
1710 vtn_fail_if(vtn_type_contains_block(b
, val
->type
->members
[i
]),
1711 "Block and BufferBlock decorations cannot decorate a "
1712 "structure type that is nested at any level inside "
1713 "another structure type decorated with Block or "
1719 static nir_constant
*
1720 vtn_null_constant(struct vtn_builder
*b
, struct vtn_type
*type
)
1722 nir_constant
*c
= rzalloc(b
, nir_constant
);
1724 switch (type
->base_type
) {
1725 case vtn_base_type_scalar
:
1726 case vtn_base_type_vector
:
1727 /* Nothing to do here. It's already initialized to zero */
1730 case vtn_base_type_pointer
: {
1731 enum vtn_variable_mode mode
= vtn_storage_class_to_mode(
1732 b
, type
->storage_class
, type
->deref
, NULL
);
1733 nir_address_format addr_format
= vtn_mode_to_address_format(b
, mode
);
1735 const nir_const_value
*null_value
= nir_address_format_null_value(addr_format
);
1736 memcpy(c
->values
, null_value
,
1737 sizeof(nir_const_value
) * nir_address_format_num_components(addr_format
));
1741 case vtn_base_type_void
:
1742 case vtn_base_type_image
:
1743 case vtn_base_type_sampler
:
1744 case vtn_base_type_sampled_image
:
1745 case vtn_base_type_function
:
1746 /* For those we have to return something but it doesn't matter what. */
1749 case vtn_base_type_matrix
:
1750 case vtn_base_type_array
:
1751 vtn_assert(type
->length
> 0);
1752 c
->num_elements
= type
->length
;
1753 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1755 c
->elements
[0] = vtn_null_constant(b
, type
->array_element
);
1756 for (unsigned i
= 1; i
< c
->num_elements
; i
++)
1757 c
->elements
[i
] = c
->elements
[0];
1760 case vtn_base_type_struct
:
1761 c
->num_elements
= type
->length
;
1762 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1763 for (unsigned i
= 0; i
< c
->num_elements
; i
++)
1764 c
->elements
[i
] = vtn_null_constant(b
, type
->members
[i
]);
1768 vtn_fail("Invalid type for null constant");
1775 spec_constant_decoration_cb(struct vtn_builder
*b
, UNUSED
struct vtn_value
*val
,
1776 ASSERTED
int member
,
1777 const struct vtn_decoration
*dec
, void *data
)
1779 vtn_assert(member
== -1);
1780 if (dec
->decoration
!= SpvDecorationSpecId
)
1783 nir_const_value
*value
= data
;
1784 for (unsigned i
= 0; i
< b
->num_specializations
; i
++) {
1785 if (b
->specializations
[i
].id
== dec
->operands
[0]) {
1786 *value
= b
->specializations
[i
].value
;
1793 handle_workgroup_size_decoration_cb(struct vtn_builder
*b
,
1794 struct vtn_value
*val
,
1795 ASSERTED
int member
,
1796 const struct vtn_decoration
*dec
,
1799 vtn_assert(member
== -1);
1800 if (dec
->decoration
!= SpvDecorationBuiltIn
||
1801 dec
->operands
[0] != SpvBuiltInWorkgroupSize
)
1804 vtn_assert(val
->type
->type
== glsl_vector_type(GLSL_TYPE_UINT
, 3));
1805 b
->workgroup_size_builtin
= val
;
1809 vtn_handle_constant(struct vtn_builder
*b
, SpvOp opcode
,
1810 const uint32_t *w
, unsigned count
)
1812 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_constant
);
1813 val
->constant
= rzalloc(b
, nir_constant
);
1815 case SpvOpConstantTrue
:
1816 case SpvOpConstantFalse
:
1817 case SpvOpSpecConstantTrue
:
1818 case SpvOpSpecConstantFalse
: {
1819 vtn_fail_if(val
->type
->type
!= glsl_bool_type(),
1820 "Result type of %s must be OpTypeBool",
1821 spirv_op_to_string(opcode
));
1823 bool bval
= (opcode
== SpvOpConstantTrue
||
1824 opcode
== SpvOpSpecConstantTrue
);
1826 nir_const_value u32val
= nir_const_value_for_uint(bval
, 32);
1828 if (opcode
== SpvOpSpecConstantTrue
||
1829 opcode
== SpvOpSpecConstantFalse
)
1830 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &u32val
);
1832 val
->constant
->values
[0].b
= u32val
.u32
!= 0;
1837 case SpvOpSpecConstant
: {
1838 vtn_fail_if(val
->type
->base_type
!= vtn_base_type_scalar
,
1839 "Result type of %s must be a scalar",
1840 spirv_op_to_string(opcode
));
1841 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1844 val
->constant
->values
[0].u64
= vtn_u64_literal(&w
[3]);
1847 val
->constant
->values
[0].u32
= w
[3];
1850 val
->constant
->values
[0].u16
= w
[3];
1853 val
->constant
->values
[0].u8
= w
[3];
1856 vtn_fail("Unsupported SpvOpConstant bit size: %u", bit_size
);
1859 if (opcode
== SpvOpSpecConstant
)
1860 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
,
1861 &val
->constant
->values
[0]);
1865 case SpvOpSpecConstantComposite
:
1866 case SpvOpConstantComposite
: {
1867 unsigned elem_count
= count
- 3;
1868 vtn_fail_if(elem_count
!= val
->type
->length
,
1869 "%s has %u constituents, expected %u",
1870 spirv_op_to_string(opcode
), elem_count
, val
->type
->length
);
1872 nir_constant
**elems
= ralloc_array(b
, nir_constant
*, elem_count
);
1873 for (unsigned i
= 0; i
< elem_count
; i
++) {
1874 struct vtn_value
*val
= vtn_untyped_value(b
, w
[i
+ 3]);
1876 if (val
->value_type
== vtn_value_type_constant
) {
1877 elems
[i
] = val
->constant
;
1879 vtn_fail_if(val
->value_type
!= vtn_value_type_undef
,
1880 "only constants or undefs allowed for "
1881 "SpvOpConstantComposite");
1882 /* to make it easier, just insert a NULL constant for now */
1883 elems
[i
] = vtn_null_constant(b
, val
->type
);
1887 switch (val
->type
->base_type
) {
1888 case vtn_base_type_vector
: {
1889 assert(glsl_type_is_vector(val
->type
->type
));
1890 for (unsigned i
= 0; i
< elem_count
; i
++)
1891 val
->constant
->values
[i
] = elems
[i
]->values
[0];
1895 case vtn_base_type_matrix
:
1896 case vtn_base_type_struct
:
1897 case vtn_base_type_array
:
1898 ralloc_steal(val
->constant
, elems
);
1899 val
->constant
->num_elements
= elem_count
;
1900 val
->constant
->elements
= elems
;
1904 vtn_fail("Result type of %s must be a composite type",
1905 spirv_op_to_string(opcode
));
1910 case SpvOpSpecConstantOp
: {
1911 nir_const_value u32op
= nir_const_value_for_uint(w
[3], 32);
1912 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &u32op
);
1913 SpvOp opcode
= u32op
.u32
;
1915 case SpvOpVectorShuffle
: {
1916 struct vtn_value
*v0
= &b
->values
[w
[4]];
1917 struct vtn_value
*v1
= &b
->values
[w
[5]];
1919 vtn_assert(v0
->value_type
== vtn_value_type_constant
||
1920 v0
->value_type
== vtn_value_type_undef
);
1921 vtn_assert(v1
->value_type
== vtn_value_type_constant
||
1922 v1
->value_type
== vtn_value_type_undef
);
1924 unsigned len0
= glsl_get_vector_elements(v0
->type
->type
);
1925 unsigned len1
= glsl_get_vector_elements(v1
->type
->type
);
1927 vtn_assert(len0
+ len1
< 16);
1929 unsigned bit_size
= glsl_get_bit_size(val
->type
->type
);
1930 unsigned bit_size0
= glsl_get_bit_size(v0
->type
->type
);
1931 unsigned bit_size1
= glsl_get_bit_size(v1
->type
->type
);
1933 vtn_assert(bit_size
== bit_size0
&& bit_size
== bit_size1
);
1934 (void)bit_size0
; (void)bit_size1
;
1936 nir_const_value undef
= { .u64
= 0xdeadbeefdeadbeef };
1937 nir_const_value combined
[NIR_MAX_VEC_COMPONENTS
* 2];
1939 if (v0
->value_type
== vtn_value_type_constant
) {
1940 for (unsigned i
= 0; i
< len0
; i
++)
1941 combined
[i
] = v0
->constant
->values
[i
];
1943 if (v1
->value_type
== vtn_value_type_constant
) {
1944 for (unsigned i
= 0; i
< len1
; i
++)
1945 combined
[len0
+ i
] = v1
->constant
->values
[i
];
1948 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1949 uint32_t comp
= w
[i
+ 6];
1950 if (comp
== (uint32_t)-1) {
1951 /* If component is not used, set the value to a known constant
1952 * to detect if it is wrongly used.
1954 val
->constant
->values
[j
] = undef
;
1956 vtn_fail_if(comp
>= len0
+ len1
,
1957 "All Component literals must either be FFFFFFFF "
1958 "or in [0, N - 1] (inclusive).");
1959 val
->constant
->values
[j
] = combined
[comp
];
1965 case SpvOpCompositeExtract
:
1966 case SpvOpCompositeInsert
: {
1967 struct vtn_value
*comp
;
1968 unsigned deref_start
;
1969 struct nir_constant
**c
;
1970 if (opcode
== SpvOpCompositeExtract
) {
1971 comp
= vtn_value(b
, w
[4], vtn_value_type_constant
);
1973 c
= &comp
->constant
;
1975 comp
= vtn_value(b
, w
[5], vtn_value_type_constant
);
1977 val
->constant
= nir_constant_clone(comp
->constant
,
1983 const struct vtn_type
*type
= comp
->type
;
1984 for (unsigned i
= deref_start
; i
< count
; i
++) {
1985 vtn_fail_if(w
[i
] > type
->length
,
1986 "%uth index of %s is %u but the type has only "
1987 "%u elements", i
- deref_start
,
1988 spirv_op_to_string(opcode
), w
[i
], type
->length
);
1990 switch (type
->base_type
) {
1991 case vtn_base_type_vector
:
1993 type
= type
->array_element
;
1996 case vtn_base_type_matrix
:
1997 case vtn_base_type_array
:
1998 c
= &(*c
)->elements
[w
[i
]];
1999 type
= type
->array_element
;
2002 case vtn_base_type_struct
:
2003 c
= &(*c
)->elements
[w
[i
]];
2004 type
= type
->members
[w
[i
]];
2008 vtn_fail("%s must only index into composite types",
2009 spirv_op_to_string(opcode
));
2013 if (opcode
== SpvOpCompositeExtract
) {
2017 unsigned num_components
= type
->length
;
2018 for (unsigned i
= 0; i
< num_components
; i
++)
2019 val
->constant
->values
[i
] = (*c
)->values
[elem
+ i
];
2022 struct vtn_value
*insert
=
2023 vtn_value(b
, w
[4], vtn_value_type_constant
);
2024 vtn_assert(insert
->type
== type
);
2026 *c
= insert
->constant
;
2028 unsigned num_components
= type
->length
;
2029 for (unsigned i
= 0; i
< num_components
; i
++)
2030 (*c
)->values
[elem
+ i
] = insert
->constant
->values
[i
];
2038 nir_alu_type dst_alu_type
= nir_get_nir_type_for_glsl_type(val
->type
->type
);
2039 nir_alu_type src_alu_type
= dst_alu_type
;
2040 unsigned num_components
= glsl_get_vector_elements(val
->type
->type
);
2043 vtn_assert(count
<= 7);
2049 /* We have a source in a conversion */
2051 nir_get_nir_type_for_glsl_type(vtn_get_value_type(b
, w
[4])->type
);
2052 /* We use the bitsize of the conversion source to evaluate the opcode later */
2053 bit_size
= glsl_get_bit_size(vtn_get_value_type(b
, w
[4])->type
);
2056 bit_size
= glsl_get_bit_size(val
->type
->type
);
2059 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(b
, opcode
, &swap
,
2060 nir_alu_type_get_type_size(src_alu_type
),
2061 nir_alu_type_get_type_size(dst_alu_type
));
2062 nir_const_value src
[3][NIR_MAX_VEC_COMPONENTS
];
2064 for (unsigned i
= 0; i
< count
- 4; i
++) {
2065 struct vtn_value
*src_val
=
2066 vtn_value(b
, w
[4 + i
], vtn_value_type_constant
);
2068 /* If this is an unsized source, pull the bit size from the
2069 * source; otherwise, we'll use the bit size from the destination.
2071 if (!nir_alu_type_get_type_size(nir_op_infos
[op
].input_types
[i
]))
2072 bit_size
= glsl_get_bit_size(src_val
->type
->type
);
2074 unsigned src_comps
= nir_op_infos
[op
].input_sizes
[i
] ?
2075 nir_op_infos
[op
].input_sizes
[i
] :
2078 unsigned j
= swap
? 1 - i
: i
;
2079 for (unsigned c
= 0; c
< src_comps
; c
++)
2080 src
[j
][c
] = src_val
->constant
->values
[c
];
2083 /* fix up fixed size sources */
2090 for (unsigned i
= 0; i
< num_components
; ++i
) {
2092 case 64: src
[1][i
].u32
= src
[1][i
].u64
; break;
2093 case 16: src
[1][i
].u32
= src
[1][i
].u16
; break;
2094 case 8: src
[1][i
].u32
= src
[1][i
].u8
; break;
2103 nir_const_value
*srcs
[3] = {
2104 src
[0], src
[1], src
[2],
2106 nir_eval_const_opcode(op
, val
->constant
->values
,
2107 num_components
, bit_size
, srcs
,
2108 b
->shader
->info
.float_controls_execution_mode
);
2115 case SpvOpConstantNull
:
2116 val
->constant
= vtn_null_constant(b
, val
->type
);
2120 vtn_fail_with_opcode("Unhandled opcode", opcode
);
2123 /* Now that we have the value, update the workgroup size if needed */
2124 vtn_foreach_decoration(b
, val
, handle_workgroup_size_decoration_cb
, NULL
);
2128 vtn_split_barrier_semantics(struct vtn_builder
*b
,
2129 SpvMemorySemanticsMask semantics
,
2130 SpvMemorySemanticsMask
*before
,
2131 SpvMemorySemanticsMask
*after
)
2133 /* For memory semantics embedded in operations, we split them into up to
2134 * two barriers, to be added before and after the operation. This is less
2135 * strict than if we propagated until the final backend stage, but still
2136 * result in correct execution.
2138 * A further improvement could be pipe this information (and use!) into the
2139 * next compiler layers, at the expense of making the handling of barriers
2143 *before
= SpvMemorySemanticsMaskNone
;
2144 *after
= SpvMemorySemanticsMaskNone
;
2146 SpvMemorySemanticsMask order_semantics
=
2147 semantics
& (SpvMemorySemanticsAcquireMask
|
2148 SpvMemorySemanticsReleaseMask
|
2149 SpvMemorySemanticsAcquireReleaseMask
|
2150 SpvMemorySemanticsSequentiallyConsistentMask
);
2152 if (util_bitcount(order_semantics
) > 1) {
2153 /* Old GLSLang versions incorrectly set all the ordering bits. This was
2154 * fixed in c51287d744fb6e7e9ccc09f6f8451e6c64b1dad6 of glslang repo,
2155 * and it is in GLSLang since revision "SPIRV99.1321" (from Jul-2016).
2157 vtn_warn("Multiple memory ordering semantics specified, "
2158 "assuming AcquireRelease.");
2159 order_semantics
= SpvMemorySemanticsAcquireReleaseMask
;
2162 const SpvMemorySemanticsMask av_vis_semantics
=
2163 semantics
& (SpvMemorySemanticsMakeAvailableMask
|
2164 SpvMemorySemanticsMakeVisibleMask
);
2166 const SpvMemorySemanticsMask storage_semantics
=
2167 semantics
& (SpvMemorySemanticsUniformMemoryMask
|
2168 SpvMemorySemanticsSubgroupMemoryMask
|
2169 SpvMemorySemanticsWorkgroupMemoryMask
|
2170 SpvMemorySemanticsCrossWorkgroupMemoryMask
|
2171 SpvMemorySemanticsAtomicCounterMemoryMask
|
2172 SpvMemorySemanticsImageMemoryMask
|
2173 SpvMemorySemanticsOutputMemoryMask
);
2175 const SpvMemorySemanticsMask other_semantics
=
2176 semantics
& ~(order_semantics
| av_vis_semantics
| storage_semantics
|
2177 SpvMemorySemanticsVolatileMask
);
2179 if (other_semantics
)
2180 vtn_warn("Ignoring unhandled memory semantics: %u\n", other_semantics
);
2182 /* SequentiallyConsistent is treated as AcquireRelease. */
2184 /* The RELEASE barrier happens BEFORE the operation, and it is usually
2185 * associated with a Store. All the write operations with a matching
2186 * semantics will not be reordered after the Store.
2188 if (order_semantics
& (SpvMemorySemanticsReleaseMask
|
2189 SpvMemorySemanticsAcquireReleaseMask
|
2190 SpvMemorySemanticsSequentiallyConsistentMask
)) {
2191 *before
|= SpvMemorySemanticsReleaseMask
| storage_semantics
;
2194 /* The ACQUIRE barrier happens AFTER the operation, and it is usually
2195 * associated with a Load. All the operations with a matching semantics
2196 * will not be reordered before the Load.
2198 if (order_semantics
& (SpvMemorySemanticsAcquireMask
|
2199 SpvMemorySemanticsAcquireReleaseMask
|
2200 SpvMemorySemanticsSequentiallyConsistentMask
)) {
2201 *after
|= SpvMemorySemanticsAcquireMask
| storage_semantics
;
2204 if (av_vis_semantics
& SpvMemorySemanticsMakeVisibleMask
)
2205 *before
|= SpvMemorySemanticsMakeVisibleMask
| storage_semantics
;
2207 if (av_vis_semantics
& SpvMemorySemanticsMakeAvailableMask
)
2208 *after
|= SpvMemorySemanticsMakeAvailableMask
| storage_semantics
;
2211 static nir_memory_semantics
2212 vtn_mem_semantics_to_nir_mem_semantics(struct vtn_builder
*b
,
2213 SpvMemorySemanticsMask semantics
)
2215 nir_memory_semantics nir_semantics
= 0;
2217 SpvMemorySemanticsMask order_semantics
=
2218 semantics
& (SpvMemorySemanticsAcquireMask
|
2219 SpvMemorySemanticsReleaseMask
|
2220 SpvMemorySemanticsAcquireReleaseMask
|
2221 SpvMemorySemanticsSequentiallyConsistentMask
);
2223 if (util_bitcount(order_semantics
) > 1) {
2224 /* Old GLSLang versions incorrectly set all the ordering bits. This was
2225 * fixed in c51287d744fb6e7e9ccc09f6f8451e6c64b1dad6 of glslang repo,
2226 * and it is in GLSLang since revision "SPIRV99.1321" (from Jul-2016).
2228 vtn_warn("Multiple memory ordering semantics bits specified, "
2229 "assuming AcquireRelease.");
2230 order_semantics
= SpvMemorySemanticsAcquireReleaseMask
;
2233 switch (order_semantics
) {
2235 /* Not an ordering barrier. */
2238 case SpvMemorySemanticsAcquireMask
:
2239 nir_semantics
= NIR_MEMORY_ACQUIRE
;
2242 case SpvMemorySemanticsReleaseMask
:
2243 nir_semantics
= NIR_MEMORY_RELEASE
;
2246 case SpvMemorySemanticsSequentiallyConsistentMask
:
2247 /* Fall through. Treated as AcquireRelease in Vulkan. */
2248 case SpvMemorySemanticsAcquireReleaseMask
:
2249 nir_semantics
= NIR_MEMORY_ACQUIRE
| NIR_MEMORY_RELEASE
;
2253 unreachable("Invalid memory order semantics");
2256 if (semantics
& SpvMemorySemanticsMakeAvailableMask
) {
2257 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
2258 "To use MakeAvailable memory semantics the VulkanMemoryModel "
2259 "capability must be declared.");
2260 nir_semantics
|= NIR_MEMORY_MAKE_AVAILABLE
;
2263 if (semantics
& SpvMemorySemanticsMakeVisibleMask
) {
2264 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
2265 "To use MakeVisible memory semantics the VulkanMemoryModel "
2266 "capability must be declared.");
2267 nir_semantics
|= NIR_MEMORY_MAKE_VISIBLE
;
2270 return nir_semantics
;
2273 static nir_variable_mode
2274 vtn_mem_sematics_to_nir_var_modes(struct vtn_builder
*b
,
2275 SpvMemorySemanticsMask semantics
)
2277 /* Vulkan Environment for SPIR-V says "SubgroupMemory, CrossWorkgroupMemory,
2278 * and AtomicCounterMemory are ignored".
2280 semantics
&= ~(SpvMemorySemanticsSubgroupMemoryMask
|
2281 SpvMemorySemanticsCrossWorkgroupMemoryMask
|
2282 SpvMemorySemanticsAtomicCounterMemoryMask
);
2284 /* TODO: Consider adding nir_var_mem_image mode to NIR so it can be used
2285 * for SpvMemorySemanticsImageMemoryMask.
2288 nir_variable_mode modes
= 0;
2289 if (semantics
& (SpvMemorySemanticsUniformMemoryMask
|
2290 SpvMemorySemanticsImageMemoryMask
)) {
2291 modes
|= nir_var_uniform
|
2296 if (semantics
& SpvMemorySemanticsWorkgroupMemoryMask
)
2297 modes
|= nir_var_mem_shared
;
2298 if (semantics
& SpvMemorySemanticsOutputMemoryMask
) {
2299 modes
|= nir_var_shader_out
;
2306 vtn_scope_to_nir_scope(struct vtn_builder
*b
, SpvScope scope
)
2308 nir_scope nir_scope
;
2310 case SpvScopeDevice
:
2311 vtn_fail_if(b
->options
->caps
.vk_memory_model
&&
2312 !b
->options
->caps
.vk_memory_model_device_scope
,
2313 "If the Vulkan memory model is declared and any instruction "
2314 "uses Device scope, the VulkanMemoryModelDeviceScope "
2315 "capability must be declared.");
2316 nir_scope
= NIR_SCOPE_DEVICE
;
2319 case SpvScopeQueueFamily
:
2320 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
2321 "To use Queue Family scope, the VulkanMemoryModel capability "
2322 "must be declared.");
2323 nir_scope
= NIR_SCOPE_QUEUE_FAMILY
;
2326 case SpvScopeWorkgroup
:
2327 nir_scope
= NIR_SCOPE_WORKGROUP
;
2330 case SpvScopeSubgroup
:
2331 nir_scope
= NIR_SCOPE_SUBGROUP
;
2334 case SpvScopeInvocation
:
2335 nir_scope
= NIR_SCOPE_INVOCATION
;
2339 vtn_fail("Invalid memory scope");
2346 vtn_emit_scoped_control_barrier(struct vtn_builder
*b
, SpvScope exec_scope
,
2348 SpvMemorySemanticsMask semantics
)
2350 nir_memory_semantics nir_semantics
=
2351 vtn_mem_semantics_to_nir_mem_semantics(b
, semantics
);
2352 nir_variable_mode modes
= vtn_mem_sematics_to_nir_var_modes(b
, semantics
);
2353 nir_scope nir_exec_scope
= vtn_scope_to_nir_scope(b
, exec_scope
);
2355 /* Memory semantics is optional for OpControlBarrier. */
2356 nir_scope nir_mem_scope
;
2357 if (nir_semantics
== 0 || modes
== 0)
2358 nir_mem_scope
= NIR_SCOPE_NONE
;
2360 nir_mem_scope
= vtn_scope_to_nir_scope(b
, mem_scope
);
2362 nir_scoped_barrier(&b
->nb
, nir_exec_scope
, nir_mem_scope
, nir_semantics
, modes
);
2366 vtn_emit_scoped_memory_barrier(struct vtn_builder
*b
, SpvScope scope
,
2367 SpvMemorySemanticsMask semantics
)
2369 nir_variable_mode modes
= vtn_mem_sematics_to_nir_var_modes(b
, semantics
);
2370 nir_memory_semantics nir_semantics
=
2371 vtn_mem_semantics_to_nir_mem_semantics(b
, semantics
);
2373 /* No barrier to add. */
2374 if (nir_semantics
== 0 || modes
== 0)
2377 nir_scope nir_mem_scope
= vtn_scope_to_nir_scope(b
, scope
);
2378 nir_scoped_barrier(&b
->nb
, NIR_SCOPE_NONE
, nir_mem_scope
, nir_semantics
, modes
);
2381 struct vtn_ssa_value
*
2382 vtn_create_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
2384 /* Always use bare types for SSA values for a couple of reasons:
2386 * 1. Code which emits deref chains should never listen to the explicit
2387 * layout information on the SSA value if any exists. If we've
2388 * accidentally been relying on this, we want to find those bugs.
2390 * 2. We want to be able to quickly check that an SSA value being assigned
2391 * to a SPIR-V value has the right type. Using bare types everywhere
2392 * ensures that we can pointer-compare.
2394 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
2395 val
->type
= glsl_get_bare_type(type
);
2398 if (!glsl_type_is_vector_or_scalar(type
)) {
2399 unsigned elems
= glsl_get_length(val
->type
);
2400 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
2401 if (glsl_type_is_array_or_matrix(type
)) {
2402 const struct glsl_type
*elem_type
= glsl_get_array_element(type
);
2403 for (unsigned i
= 0; i
< elems
; i
++)
2404 val
->elems
[i
] = vtn_create_ssa_value(b
, elem_type
);
2406 vtn_assert(glsl_type_is_struct_or_ifc(type
));
2407 for (unsigned i
= 0; i
< elems
; i
++) {
2408 const struct glsl_type
*elem_type
= glsl_get_struct_field(type
, i
);
2409 val
->elems
[i
] = vtn_create_ssa_value(b
, elem_type
);
2418 vtn_tex_src(struct vtn_builder
*b
, unsigned index
, nir_tex_src_type type
)
2421 src
.src
= nir_src_for_ssa(vtn_get_nir_ssa(b
, index
));
2422 src
.src_type
= type
;
2427 image_operand_arg(struct vtn_builder
*b
, const uint32_t *w
, uint32_t count
,
2428 uint32_t mask_idx
, SpvImageOperandsMask op
)
2430 static const SpvImageOperandsMask ops_with_arg
=
2431 SpvImageOperandsBiasMask
|
2432 SpvImageOperandsLodMask
|
2433 SpvImageOperandsGradMask
|
2434 SpvImageOperandsConstOffsetMask
|
2435 SpvImageOperandsOffsetMask
|
2436 SpvImageOperandsConstOffsetsMask
|
2437 SpvImageOperandsSampleMask
|
2438 SpvImageOperandsMinLodMask
|
2439 SpvImageOperandsMakeTexelAvailableMask
|
2440 SpvImageOperandsMakeTexelVisibleMask
;
2442 assert(util_bitcount(op
) == 1);
2443 assert(w
[mask_idx
] & op
);
2444 assert(op
& ops_with_arg
);
2446 uint32_t idx
= util_bitcount(w
[mask_idx
] & (op
- 1) & ops_with_arg
) + 1;
2448 /* Adjust indices for operands with two arguments. */
2449 static const SpvImageOperandsMask ops_with_two_args
=
2450 SpvImageOperandsGradMask
;
2451 idx
+= util_bitcount(w
[mask_idx
] & (op
- 1) & ops_with_two_args
);
2455 vtn_fail_if(idx
+ (op
& ops_with_two_args
? 1 : 0) >= count
,
2456 "Image op claims to have %s but does not enough "
2457 "following operands", spirv_imageoperands_to_string(op
));
2463 non_uniform_decoration_cb(struct vtn_builder
*b
,
2464 struct vtn_value
*val
, int member
,
2465 const struct vtn_decoration
*dec
, void *void_ctx
)
2467 enum gl_access_qualifier
*access
= void_ctx
;
2468 switch (dec
->decoration
) {
2469 case SpvDecorationNonUniformEXT
:
2470 *access
|= ACCESS_NON_UNIFORM
;
2479 vtn_handle_texture(struct vtn_builder
*b
, SpvOp opcode
,
2480 const uint32_t *w
, unsigned count
)
2482 struct vtn_type
*ret_type
= vtn_get_type(b
, w
[1]);
2484 if (opcode
== SpvOpSampledImage
) {
2485 struct vtn_sampled_image si
= {
2486 .image
= vtn_get_image(b
, w
[3]),
2487 .sampler
= vtn_get_sampler(b
, w
[4]),
2490 enum gl_access_qualifier access
= 0;
2491 vtn_foreach_decoration(b
, vtn_untyped_value(b
, w
[3]),
2492 non_uniform_decoration_cb
, &access
);
2493 vtn_foreach_decoration(b
, vtn_untyped_value(b
, w
[4]),
2494 non_uniform_decoration_cb
, &access
);
2496 vtn_push_sampled_image(b
, w
[2], si
, access
& ACCESS_NON_UNIFORM
);
2498 } else if (opcode
== SpvOpImage
) {
2499 struct vtn_sampled_image si
= vtn_get_sampled_image(b
, w
[3]);
2501 enum gl_access_qualifier access
= 0;
2502 vtn_foreach_decoration(b
, vtn_untyped_value(b
, w
[3]),
2503 non_uniform_decoration_cb
, &access
);
2505 vtn_push_image(b
, w
[2], si
.image
, access
& ACCESS_NON_UNIFORM
);
2509 nir_deref_instr
*image
= NULL
, *sampler
= NULL
;
2510 struct vtn_value
*sampled_val
= vtn_untyped_value(b
, w
[3]);
2511 if (sampled_val
->type
->base_type
== vtn_base_type_sampled_image
) {
2512 struct vtn_sampled_image si
= vtn_get_sampled_image(b
, w
[3]);
2514 sampler
= si
.sampler
;
2516 image
= vtn_get_image(b
, w
[3]);
2519 const enum glsl_sampler_dim sampler_dim
= glsl_get_sampler_dim(image
->type
);
2520 const bool is_array
= glsl_sampler_type_is_array(image
->type
);
2521 nir_alu_type dest_type
= nir_type_invalid
;
2523 /* Figure out the base texture operation */
2526 case SpvOpImageSampleImplicitLod
:
2527 case SpvOpImageSampleDrefImplicitLod
:
2528 case SpvOpImageSampleProjImplicitLod
:
2529 case SpvOpImageSampleProjDrefImplicitLod
:
2530 texop
= nir_texop_tex
;
2533 case SpvOpImageSampleExplicitLod
:
2534 case SpvOpImageSampleDrefExplicitLod
:
2535 case SpvOpImageSampleProjExplicitLod
:
2536 case SpvOpImageSampleProjDrefExplicitLod
:
2537 texop
= nir_texop_txl
;
2540 case SpvOpImageFetch
:
2541 if (sampler_dim
== GLSL_SAMPLER_DIM_MS
) {
2542 texop
= nir_texop_txf_ms
;
2544 texop
= nir_texop_txf
;
2548 case SpvOpImageGather
:
2549 case SpvOpImageDrefGather
:
2550 texop
= nir_texop_tg4
;
2553 case SpvOpImageQuerySizeLod
:
2554 case SpvOpImageQuerySize
:
2555 texop
= nir_texop_txs
;
2556 dest_type
= nir_type_int
;
2559 case SpvOpImageQueryLod
:
2560 texop
= nir_texop_lod
;
2561 dest_type
= nir_type_float
;
2564 case SpvOpImageQueryLevels
:
2565 texop
= nir_texop_query_levels
;
2566 dest_type
= nir_type_int
;
2569 case SpvOpImageQuerySamples
:
2570 texop
= nir_texop_texture_samples
;
2571 dest_type
= nir_type_int
;
2574 case SpvOpFragmentFetchAMD
:
2575 texop
= nir_texop_fragment_fetch
;
2578 case SpvOpFragmentMaskFetchAMD
:
2579 texop
= nir_texop_fragment_mask_fetch
;
2580 dest_type
= nir_type_uint
;
2584 vtn_fail_with_opcode("Unhandled opcode", opcode
);
2587 nir_tex_src srcs
[10]; /* 10 should be enough */
2588 nir_tex_src
*p
= srcs
;
2590 p
->src
= nir_src_for_ssa(&image
->dest
.ssa
);
2591 p
->src_type
= nir_tex_src_texture_deref
;
2601 vtn_fail_if(sampler
== NULL
,
2602 "%s requires an image of type OpTypeSampledImage",
2603 spirv_op_to_string(opcode
));
2604 p
->src
= nir_src_for_ssa(&sampler
->dest
.ssa
);
2605 p
->src_type
= nir_tex_src_sampler_deref
;
2609 case nir_texop_txf_ms
:
2611 case nir_texop_query_levels
:
2612 case nir_texop_texture_samples
:
2613 case nir_texop_samples_identical
:
2614 case nir_texop_fragment_fetch
:
2615 case nir_texop_fragment_mask_fetch
:
2618 case nir_texop_txf_ms_fb
:
2619 vtn_fail("unexpected nir_texop_txf_ms_fb");
2621 case nir_texop_txf_ms_mcs
:
2622 vtn_fail("unexpected nir_texop_txf_ms_mcs");
2623 case nir_texop_tex_prefetch
:
2624 vtn_fail("unexpected nir_texop_tex_prefetch");
2629 struct nir_ssa_def
*coord
;
2630 unsigned coord_components
;
2632 case SpvOpImageSampleImplicitLod
:
2633 case SpvOpImageSampleExplicitLod
:
2634 case SpvOpImageSampleDrefImplicitLod
:
2635 case SpvOpImageSampleDrefExplicitLod
:
2636 case SpvOpImageSampleProjImplicitLod
:
2637 case SpvOpImageSampleProjExplicitLod
:
2638 case SpvOpImageSampleProjDrefImplicitLod
:
2639 case SpvOpImageSampleProjDrefExplicitLod
:
2640 case SpvOpImageFetch
:
2641 case SpvOpImageGather
:
2642 case SpvOpImageDrefGather
:
2643 case SpvOpImageQueryLod
:
2644 case SpvOpFragmentFetchAMD
:
2645 case SpvOpFragmentMaskFetchAMD
: {
2646 /* All these types have the coordinate as their first real argument */
2647 coord_components
= glsl_get_sampler_dim_coordinate_components(sampler_dim
);
2649 if (is_array
&& texop
!= nir_texop_lod
)
2652 struct vtn_ssa_value
*coord_val
= vtn_ssa_value(b
, w
[idx
++]);
2653 coord
= coord_val
->def
;
2654 p
->src
= nir_src_for_ssa(nir_channels(&b
->nb
, coord
,
2655 (1 << coord_components
) - 1));
2657 /* OpenCL allows integer sampling coordinates */
2658 if (glsl_type_is_integer(coord_val
->type
) &&
2659 opcode
== SpvOpImageSampleExplicitLod
) {
2660 vtn_fail_if(b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
,
2661 "Unless the Kernel capability is being used, the coordinate parameter "
2662 "OpImageSampleExplicitLod must be floating point.");
2664 p
->src
= nir_src_for_ssa(nir_i2f32(&b
->nb
, p
->src
.ssa
));
2667 p
->src_type
= nir_tex_src_coord
;
2674 coord_components
= 0;
2679 case SpvOpImageSampleProjImplicitLod
:
2680 case SpvOpImageSampleProjExplicitLod
:
2681 case SpvOpImageSampleProjDrefImplicitLod
:
2682 case SpvOpImageSampleProjDrefExplicitLod
:
2683 /* These have the projector as the last coordinate component */
2684 p
->src
= nir_src_for_ssa(nir_channel(&b
->nb
, coord
, coord_components
));
2685 p
->src_type
= nir_tex_src_projector
;
2693 bool is_shadow
= false;
2694 unsigned gather_component
= 0;
2696 case SpvOpImageSampleDrefImplicitLod
:
2697 case SpvOpImageSampleDrefExplicitLod
:
2698 case SpvOpImageSampleProjDrefImplicitLod
:
2699 case SpvOpImageSampleProjDrefExplicitLod
:
2700 case SpvOpImageDrefGather
:
2701 /* These all have an explicit depth value as their next source */
2703 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_comparator
);
2706 case SpvOpImageGather
:
2707 /* This has a component as its next source */
2708 gather_component
= vtn_constant_uint(b
, w
[idx
++]);
2715 /* For OpImageQuerySizeLod, we always have an LOD */
2716 if (opcode
== SpvOpImageQuerySizeLod
)
2717 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
2719 /* For OpFragmentFetchAMD, we always have a multisample index */
2720 if (opcode
== SpvOpFragmentFetchAMD
)
2721 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ms_index
);
2723 /* Now we need to handle some number of optional arguments */
2724 struct vtn_value
*gather_offsets
= NULL
;
2726 uint32_t operands
= w
[idx
];
2728 if (operands
& SpvImageOperandsBiasMask
) {
2729 vtn_assert(texop
== nir_texop_tex
||
2730 texop
== nir_texop_tg4
);
2731 if (texop
== nir_texop_tex
)
2732 texop
= nir_texop_txb
;
2733 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2734 SpvImageOperandsBiasMask
);
2735 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_bias
);
2738 if (operands
& SpvImageOperandsLodMask
) {
2739 vtn_assert(texop
== nir_texop_txl
|| texop
== nir_texop_txf
||
2740 texop
== nir_texop_txs
|| texop
== nir_texop_tg4
);
2741 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2742 SpvImageOperandsLodMask
);
2743 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_lod
);
2746 if (operands
& SpvImageOperandsGradMask
) {
2747 vtn_assert(texop
== nir_texop_txl
);
2748 texop
= nir_texop_txd
;
2749 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2750 SpvImageOperandsGradMask
);
2751 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_ddx
);
2752 (*p
++) = vtn_tex_src(b
, w
[arg
+ 1], nir_tex_src_ddy
);
2755 vtn_fail_if(util_bitcount(operands
& (SpvImageOperandsConstOffsetsMask
|
2756 SpvImageOperandsOffsetMask
|
2757 SpvImageOperandsConstOffsetMask
)) > 1,
2758 "At most one of the ConstOffset, Offset, and ConstOffsets "
2759 "image operands can be used on a given instruction.");
2761 if (operands
& SpvImageOperandsOffsetMask
) {
2762 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2763 SpvImageOperandsOffsetMask
);
2764 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_offset
);
2767 if (operands
& SpvImageOperandsConstOffsetMask
) {
2768 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2769 SpvImageOperandsConstOffsetMask
);
2770 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_offset
);
2773 if (operands
& SpvImageOperandsConstOffsetsMask
) {
2774 vtn_assert(texop
== nir_texop_tg4
);
2775 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2776 SpvImageOperandsConstOffsetsMask
);
2777 gather_offsets
= vtn_value(b
, w
[arg
], vtn_value_type_constant
);
2780 if (operands
& SpvImageOperandsSampleMask
) {
2781 vtn_assert(texop
== nir_texop_txf_ms
);
2782 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2783 SpvImageOperandsSampleMask
);
2784 texop
= nir_texop_txf_ms
;
2785 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_ms_index
);
2788 if (operands
& SpvImageOperandsMinLodMask
) {
2789 vtn_assert(texop
== nir_texop_tex
||
2790 texop
== nir_texop_txb
||
2791 texop
== nir_texop_txd
);
2792 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2793 SpvImageOperandsMinLodMask
);
2794 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_min_lod
);
2798 nir_tex_instr
*instr
= nir_tex_instr_create(b
->shader
, p
- srcs
);
2801 memcpy(instr
->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
2803 instr
->coord_components
= coord_components
;
2804 instr
->sampler_dim
= sampler_dim
;
2805 instr
->is_array
= is_array
;
2806 instr
->is_shadow
= is_shadow
;
2807 instr
->is_new_style_shadow
=
2808 is_shadow
&& glsl_get_components(ret_type
->type
) == 1;
2809 instr
->component
= gather_component
;
2811 /* The Vulkan spec says:
2813 * "If an instruction loads from or stores to a resource (including
2814 * atomics and image instructions) and the resource descriptor being
2815 * accessed is not dynamically uniform, then the operand corresponding
2816 * to that resource (e.g. the pointer or sampled image operand) must be
2817 * decorated with NonUniform."
2819 * It's very careful to specify that the exact operand must be decorated
2820 * NonUniform. The SPIR-V parser is not expected to chase through long
2821 * chains to find the NonUniform decoration. It's either right there or we
2822 * can assume it doesn't exist.
2824 enum gl_access_qualifier access
= 0;
2825 vtn_foreach_decoration(b
, sampled_val
, non_uniform_decoration_cb
, &access
);
2827 if (sampled_val
->propagated_non_uniform
)
2828 access
|= ACCESS_NON_UNIFORM
;
2830 if (image
&& (access
& ACCESS_NON_UNIFORM
))
2831 instr
->texture_non_uniform
= true;
2833 if (sampler
&& (access
& ACCESS_NON_UNIFORM
))
2834 instr
->sampler_non_uniform
= true;
2836 /* for non-query ops, get dest_type from SPIR-V return type */
2837 if (dest_type
== nir_type_invalid
) {
2838 /* the return type should match the image type, unless the image type is
2839 * VOID (CL image), in which case the return type dictates the sampler
2841 enum glsl_base_type sampler_base
=
2842 glsl_get_sampler_result_type(image
->type
);
2843 enum glsl_base_type ret_base
= glsl_get_base_type(ret_type
->type
);
2844 vtn_fail_if(sampler_base
!= ret_base
&& sampler_base
!= GLSL_TYPE_VOID
,
2845 "SPIR-V return type mismatches image type. This is only valid "
2846 "for untyped images (OpenCL).");
2848 case GLSL_TYPE_FLOAT
: dest_type
= nir_type_float
; break;
2849 case GLSL_TYPE_INT
: dest_type
= nir_type_int
; break;
2850 case GLSL_TYPE_UINT
: dest_type
= nir_type_uint
; break;
2851 case GLSL_TYPE_BOOL
: dest_type
= nir_type_bool
; break;
2853 vtn_fail("Invalid base type for sampler result");
2857 instr
->dest_type
= dest_type
;
2859 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
2860 nir_tex_instr_dest_size(instr
), 32, NULL
);
2862 vtn_assert(glsl_get_vector_elements(ret_type
->type
) ==
2863 nir_tex_instr_dest_size(instr
));
2865 if (gather_offsets
) {
2866 vtn_fail_if(gather_offsets
->type
->base_type
!= vtn_base_type_array
||
2867 gather_offsets
->type
->length
!= 4,
2868 "ConstOffsets must be an array of size four of vectors "
2869 "of two integer components");
2871 struct vtn_type
*vec_type
= gather_offsets
->type
->array_element
;
2872 vtn_fail_if(vec_type
->base_type
!= vtn_base_type_vector
||
2873 vec_type
->length
!= 2 ||
2874 !glsl_type_is_integer(vec_type
->type
),
2875 "ConstOffsets must be an array of size four of vectors "
2876 "of two integer components");
2878 unsigned bit_size
= glsl_get_bit_size(vec_type
->type
);
2879 for (uint32_t i
= 0; i
< 4; i
++) {
2880 const nir_const_value
*cvec
=
2881 gather_offsets
->constant
->elements
[i
]->values
;
2882 for (uint32_t j
= 0; j
< 2; j
++) {
2884 case 8: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i8
; break;
2885 case 16: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i16
; break;
2886 case 32: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i32
; break;
2887 case 64: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i64
; break;
2889 vtn_fail("Unsupported bit size: %u", bit_size
);
2895 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
2897 vtn_push_nir_ssa(b
, w
[2], &instr
->dest
.ssa
);
2901 fill_common_atomic_sources(struct vtn_builder
*b
, SpvOp opcode
,
2902 const uint32_t *w
, nir_src
*src
)
2904 const struct glsl_type
*type
= vtn_get_type(b
, w
[1])->type
;
2905 unsigned bit_size
= glsl_get_bit_size(type
);
2908 case SpvOpAtomicIIncrement
:
2909 src
[0] = nir_src_for_ssa(nir_imm_intN_t(&b
->nb
, 1, bit_size
));
2912 case SpvOpAtomicIDecrement
:
2913 src
[0] = nir_src_for_ssa(nir_imm_intN_t(&b
->nb
, -1, bit_size
));
2916 case SpvOpAtomicISub
:
2918 nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_get_nir_ssa(b
, w
[6])));
2921 case SpvOpAtomicCompareExchange
:
2922 case SpvOpAtomicCompareExchangeWeak
:
2923 src
[0] = nir_src_for_ssa(vtn_get_nir_ssa(b
, w
[8]));
2924 src
[1] = nir_src_for_ssa(vtn_get_nir_ssa(b
, w
[7]));
2927 case SpvOpAtomicExchange
:
2928 case SpvOpAtomicIAdd
:
2929 case SpvOpAtomicSMin
:
2930 case SpvOpAtomicUMin
:
2931 case SpvOpAtomicSMax
:
2932 case SpvOpAtomicUMax
:
2933 case SpvOpAtomicAnd
:
2935 case SpvOpAtomicXor
:
2936 case SpvOpAtomicFAddEXT
:
2937 src
[0] = nir_src_for_ssa(vtn_get_nir_ssa(b
, w
[6]));
2941 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
2945 static nir_ssa_def
*
2946 get_image_coord(struct vtn_builder
*b
, uint32_t value
)
2948 nir_ssa_def
*coord
= vtn_get_nir_ssa(b
, value
);
2950 /* The image_load_store intrinsics assume a 4-dim coordinate */
2951 unsigned swizzle
[4];
2952 for (unsigned i
= 0; i
< 4; i
++)
2953 swizzle
[i
] = MIN2(i
, coord
->num_components
- 1);
2955 return nir_swizzle(&b
->nb
, coord
, swizzle
, 4);
2958 static nir_ssa_def
*
2959 expand_to_vec4(nir_builder
*b
, nir_ssa_def
*value
)
2961 if (value
->num_components
== 4)
2965 for (unsigned i
= 0; i
< 4; i
++)
2966 swiz
[i
] = i
< value
->num_components
? i
: 0;
2967 return nir_swizzle(b
, value
, swiz
, 4);
2971 vtn_handle_image(struct vtn_builder
*b
, SpvOp opcode
,
2972 const uint32_t *w
, unsigned count
)
2974 /* Just get this one out of the way */
2975 if (opcode
== SpvOpImageTexelPointer
) {
2976 struct vtn_value
*val
=
2977 vtn_push_value(b
, w
[2], vtn_value_type_image_pointer
);
2978 val
->image
= ralloc(b
, struct vtn_image_pointer
);
2980 val
->image
->image
= vtn_nir_deref(b
, w
[3]);
2981 val
->image
->coord
= get_image_coord(b
, w
[4]);
2982 val
->image
->sample
= vtn_get_nir_ssa(b
, w
[5]);
2983 val
->image
->lod
= nir_imm_int(&b
->nb
, 0);
2987 struct vtn_image_pointer image
;
2988 SpvScope scope
= SpvScopeInvocation
;
2989 SpvMemorySemanticsMask semantics
= 0;
2991 enum gl_access_qualifier access
= 0;
2993 struct vtn_value
*res_val
;
2995 case SpvOpAtomicExchange
:
2996 case SpvOpAtomicCompareExchange
:
2997 case SpvOpAtomicCompareExchangeWeak
:
2998 case SpvOpAtomicIIncrement
:
2999 case SpvOpAtomicIDecrement
:
3000 case SpvOpAtomicIAdd
:
3001 case SpvOpAtomicISub
:
3002 case SpvOpAtomicLoad
:
3003 case SpvOpAtomicSMin
:
3004 case SpvOpAtomicUMin
:
3005 case SpvOpAtomicSMax
:
3006 case SpvOpAtomicUMax
:
3007 case SpvOpAtomicAnd
:
3009 case SpvOpAtomicXor
:
3010 case SpvOpAtomicFAddEXT
:
3011 res_val
= vtn_value(b
, w
[3], vtn_value_type_image_pointer
);
3012 image
= *res_val
->image
;
3013 scope
= vtn_constant_uint(b
, w
[4]);
3014 semantics
= vtn_constant_uint(b
, w
[5]);
3015 access
|= ACCESS_COHERENT
;
3018 case SpvOpAtomicStore
:
3019 res_val
= vtn_value(b
, w
[1], vtn_value_type_image_pointer
);
3020 image
= *res_val
->image
;
3021 scope
= vtn_constant_uint(b
, w
[2]);
3022 semantics
= vtn_constant_uint(b
, w
[3]);
3023 access
|= ACCESS_COHERENT
;
3026 case SpvOpImageQuerySizeLod
:
3027 res_val
= vtn_untyped_value(b
, w
[3]);
3028 image
.image
= vtn_get_image(b
, w
[3]);
3030 image
.sample
= NULL
;
3031 image
.lod
= vtn_ssa_value(b
, w
[4])->def
;
3034 case SpvOpImageQuerySize
:
3035 res_val
= vtn_untyped_value(b
, w
[3]);
3036 image
.image
= vtn_get_image(b
, w
[3]);
3038 image
.sample
= NULL
;
3042 case SpvOpImageQueryFormat
:
3043 case SpvOpImageQueryOrder
:
3044 res_val
= vtn_untyped_value(b
, w
[3]);
3045 image
.image
= vtn_get_image(b
, w
[3]);
3047 image
.sample
= NULL
;
3051 case SpvOpImageRead
: {
3052 res_val
= vtn_untyped_value(b
, w
[3]);
3053 image
.image
= vtn_get_image(b
, w
[3]);
3054 image
.coord
= get_image_coord(b
, w
[4]);
3056 const SpvImageOperandsMask operands
=
3057 count
> 5 ? w
[5] : SpvImageOperandsMaskNone
;
3059 if (operands
& SpvImageOperandsSampleMask
) {
3060 uint32_t arg
= image_operand_arg(b
, w
, count
, 5,
3061 SpvImageOperandsSampleMask
);
3062 image
.sample
= vtn_get_nir_ssa(b
, w
[arg
]);
3064 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
3067 if (operands
& SpvImageOperandsMakeTexelVisibleMask
) {
3068 vtn_fail_if((operands
& SpvImageOperandsNonPrivateTexelMask
) == 0,
3069 "MakeTexelVisible requires NonPrivateTexel to also be set.");
3070 uint32_t arg
= image_operand_arg(b
, w
, count
, 5,
3071 SpvImageOperandsMakeTexelVisibleMask
);
3072 semantics
= SpvMemorySemanticsMakeVisibleMask
;
3073 scope
= vtn_constant_uint(b
, w
[arg
]);
3076 if (operands
& SpvImageOperandsLodMask
) {
3077 uint32_t arg
= image_operand_arg(b
, w
, count
, 5,
3078 SpvImageOperandsLodMask
);
3079 image
.lod
= vtn_get_nir_ssa(b
, w
[arg
]);
3081 image
.lod
= nir_imm_int(&b
->nb
, 0);
3084 if (operands
& SpvImageOperandsVolatileTexelMask
)
3085 access
|= ACCESS_VOLATILE
;
3090 case SpvOpImageWrite
: {
3091 res_val
= vtn_untyped_value(b
, w
[1]);
3092 image
.image
= vtn_get_image(b
, w
[1]);
3093 image
.coord
= get_image_coord(b
, w
[2]);
3097 const SpvImageOperandsMask operands
=
3098 count
> 4 ? w
[4] : SpvImageOperandsMaskNone
;
3100 if (operands
& SpvImageOperandsSampleMask
) {
3101 uint32_t arg
= image_operand_arg(b
, w
, count
, 4,
3102 SpvImageOperandsSampleMask
);
3103 image
.sample
= vtn_get_nir_ssa(b
, w
[arg
]);
3105 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
3108 if (operands
& SpvImageOperandsMakeTexelAvailableMask
) {
3109 vtn_fail_if((operands
& SpvImageOperandsNonPrivateTexelMask
) == 0,
3110 "MakeTexelAvailable requires NonPrivateTexel to also be set.");
3111 uint32_t arg
= image_operand_arg(b
, w
, count
, 4,
3112 SpvImageOperandsMakeTexelAvailableMask
);
3113 semantics
= SpvMemorySemanticsMakeAvailableMask
;
3114 scope
= vtn_constant_uint(b
, w
[arg
]);
3117 if (operands
& SpvImageOperandsLodMask
) {
3118 uint32_t arg
= image_operand_arg(b
, w
, count
, 4,
3119 SpvImageOperandsLodMask
);
3120 image
.lod
= vtn_get_nir_ssa(b
, w
[arg
]);
3122 image
.lod
= nir_imm_int(&b
->nb
, 0);
3125 if (operands
& SpvImageOperandsVolatileTexelMask
)
3126 access
|= ACCESS_VOLATILE
;
3132 vtn_fail_with_opcode("Invalid image opcode", opcode
);
3135 if (semantics
& SpvMemorySemanticsVolatileMask
)
3136 access
|= ACCESS_VOLATILE
;
3138 nir_intrinsic_op op
;
3140 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_deref_##N; break;
3141 OP(ImageQuerySize
, size
)
3142 OP(ImageQuerySizeLod
, size
)
3144 OP(ImageWrite
, store
)
3145 OP(AtomicLoad
, load
)
3146 OP(AtomicStore
, store
)
3147 OP(AtomicExchange
, atomic_exchange
)
3148 OP(AtomicCompareExchange
, atomic_comp_swap
)
3149 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
3150 OP(AtomicIIncrement
, atomic_add
)
3151 OP(AtomicIDecrement
, atomic_add
)
3152 OP(AtomicIAdd
, atomic_add
)
3153 OP(AtomicISub
, atomic_add
)
3154 OP(AtomicSMin
, atomic_imin
)
3155 OP(AtomicUMin
, atomic_umin
)
3156 OP(AtomicSMax
, atomic_imax
)
3157 OP(AtomicUMax
, atomic_umax
)
3158 OP(AtomicAnd
, atomic_and
)
3159 OP(AtomicOr
, atomic_or
)
3160 OP(AtomicXor
, atomic_xor
)
3161 OP(AtomicFAddEXT
, atomic_fadd
)
3162 OP(ImageQueryFormat
, format
)
3163 OP(ImageQueryOrder
, order
)
3166 vtn_fail_with_opcode("Invalid image opcode", opcode
);
3169 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
3171 intrin
->src
[0] = nir_src_for_ssa(&image
.image
->dest
.ssa
);
3174 case SpvOpImageQuerySize
:
3175 case SpvOpImageQuerySizeLod
:
3176 case SpvOpImageQueryFormat
:
3177 case SpvOpImageQueryOrder
:
3180 /* The image coordinate is always 4 components but we may not have that
3181 * many. Swizzle to compensate.
3183 intrin
->src
[1] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, image
.coord
));
3184 intrin
->src
[2] = nir_src_for_ssa(image
.sample
);
3188 /* The Vulkan spec says:
3190 * "If an instruction loads from or stores to a resource (including
3191 * atomics and image instructions) and the resource descriptor being
3192 * accessed is not dynamically uniform, then the operand corresponding
3193 * to that resource (e.g. the pointer or sampled image operand) must be
3194 * decorated with NonUniform."
3196 * It's very careful to specify that the exact operand must be decorated
3197 * NonUniform. The SPIR-V parser is not expected to chase through long
3198 * chains to find the NonUniform decoration. It's either right there or we
3199 * can assume it doesn't exist.
3201 vtn_foreach_decoration(b
, res_val
, non_uniform_decoration_cb
, &access
);
3202 nir_intrinsic_set_access(intrin
, access
);
3205 case SpvOpImageQueryFormat
:
3206 case SpvOpImageQueryOrder
:
3207 /* No additional sources */
3209 case SpvOpImageQuerySize
:
3210 intrin
->src
[1] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 0));
3212 case SpvOpImageQuerySizeLod
:
3213 intrin
->src
[1] = nir_src_for_ssa(image
.lod
);
3215 case SpvOpAtomicLoad
:
3216 case SpvOpImageRead
:
3217 /* Only OpImageRead can support a lod parameter if
3218 * SPV_AMD_shader_image_load_store_lod is used but the current NIR
3219 * intrinsics definition for atomics requires us to set it for
3222 intrin
->src
[3] = nir_src_for_ssa(image
.lod
);
3224 case SpvOpAtomicStore
:
3225 case SpvOpImageWrite
: {
3226 const uint32_t value_id
= opcode
== SpvOpAtomicStore
? w
[4] : w
[3];
3227 struct vtn_ssa_value
*value
= vtn_ssa_value(b
, value_id
);
3228 /* nir_intrinsic_image_deref_store always takes a vec4 value */
3229 assert(op
== nir_intrinsic_image_deref_store
);
3230 intrin
->num_components
= 4;
3231 intrin
->src
[3] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, value
->def
));
3232 /* Only OpImageWrite can support a lod parameter if
3233 * SPV_AMD_shader_image_load_store_lod is used but the current NIR
3234 * intrinsics definition for atomics requires us to set it for
3237 intrin
->src
[4] = nir_src_for_ssa(image
.lod
);
3239 if (opcode
== SpvOpImageWrite
)
3240 nir_intrinsic_set_type(intrin
, nir_get_nir_type_for_glsl_type(value
->type
));
3244 case SpvOpAtomicCompareExchange
:
3245 case SpvOpAtomicCompareExchangeWeak
:
3246 case SpvOpAtomicIIncrement
:
3247 case SpvOpAtomicIDecrement
:
3248 case SpvOpAtomicExchange
:
3249 case SpvOpAtomicIAdd
:
3250 case SpvOpAtomicISub
:
3251 case SpvOpAtomicSMin
:
3252 case SpvOpAtomicUMin
:
3253 case SpvOpAtomicSMax
:
3254 case SpvOpAtomicUMax
:
3255 case SpvOpAtomicAnd
:
3257 case SpvOpAtomicXor
:
3258 case SpvOpAtomicFAddEXT
:
3259 fill_common_atomic_sources(b
, opcode
, w
, &intrin
->src
[3]);
3263 vtn_fail_with_opcode("Invalid image opcode", opcode
);
3266 /* Image operations implicitly have the Image storage memory semantics. */
3267 semantics
|= SpvMemorySemanticsImageMemoryMask
;
3269 SpvMemorySemanticsMask before_semantics
;
3270 SpvMemorySemanticsMask after_semantics
;
3271 vtn_split_barrier_semantics(b
, semantics
, &before_semantics
, &after_semantics
);
3273 if (before_semantics
)
3274 vtn_emit_memory_barrier(b
, scope
, before_semantics
);
3276 if (opcode
!= SpvOpImageWrite
&& opcode
!= SpvOpAtomicStore
) {
3277 struct vtn_type
*type
= vtn_get_type(b
, w
[1]);
3279 unsigned dest_components
= glsl_get_vector_elements(type
->type
);
3280 if (nir_intrinsic_infos
[op
].dest_components
== 0)
3281 intrin
->num_components
= dest_components
;
3283 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
,
3284 nir_intrinsic_dest_components(intrin
), 32, NULL
);
3286 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3288 nir_ssa_def
*result
= &intrin
->dest
.ssa
;
3289 if (nir_intrinsic_dest_components(intrin
) != dest_components
)
3290 result
= nir_channels(&b
->nb
, result
, (1 << dest_components
) - 1);
3292 vtn_push_nir_ssa(b
, w
[2], result
);
3294 if (opcode
== SpvOpImageRead
)
3295 nir_intrinsic_set_type(intrin
, nir_get_nir_type_for_glsl_type(type
->type
));
3297 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3300 if (after_semantics
)
3301 vtn_emit_memory_barrier(b
, scope
, after_semantics
);
3304 static nir_intrinsic_op
3305 get_ssbo_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
3308 case SpvOpAtomicLoad
: return nir_intrinsic_load_ssbo
;
3309 case SpvOpAtomicStore
: return nir_intrinsic_store_ssbo
;
3310 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
3311 OP(AtomicExchange
, atomic_exchange
)
3312 OP(AtomicCompareExchange
, atomic_comp_swap
)
3313 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
3314 OP(AtomicIIncrement
, atomic_add
)
3315 OP(AtomicIDecrement
, atomic_add
)
3316 OP(AtomicIAdd
, atomic_add
)
3317 OP(AtomicISub
, atomic_add
)
3318 OP(AtomicSMin
, atomic_imin
)
3319 OP(AtomicUMin
, atomic_umin
)
3320 OP(AtomicSMax
, atomic_imax
)
3321 OP(AtomicUMax
, atomic_umax
)
3322 OP(AtomicAnd
, atomic_and
)
3323 OP(AtomicOr
, atomic_or
)
3324 OP(AtomicXor
, atomic_xor
)
3325 OP(AtomicFAddEXT
, atomic_fadd
)
3328 vtn_fail_with_opcode("Invalid SSBO atomic", opcode
);
3332 static nir_intrinsic_op
3333 get_uniform_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
3336 #define OP(S, N) case SpvOp##S: return nir_intrinsic_atomic_counter_ ##N;
3337 OP(AtomicLoad
, read_deref
)
3338 OP(AtomicExchange
, exchange
)
3339 OP(AtomicCompareExchange
, comp_swap
)
3340 OP(AtomicCompareExchangeWeak
, comp_swap
)
3341 OP(AtomicIIncrement
, inc_deref
)
3342 OP(AtomicIDecrement
, post_dec_deref
)
3343 OP(AtomicIAdd
, add_deref
)
3344 OP(AtomicISub
, add_deref
)
3345 OP(AtomicUMin
, min_deref
)
3346 OP(AtomicUMax
, max_deref
)
3347 OP(AtomicAnd
, and_deref
)
3348 OP(AtomicOr
, or_deref
)
3349 OP(AtomicXor
, xor_deref
)
3352 /* We left the following out: AtomicStore, AtomicSMin and
3353 * AtomicSmax. Right now there are not nir intrinsics for them. At this
3354 * moment Atomic Counter support is needed for ARB_spirv support, so is
3355 * only need to support GLSL Atomic Counters that are uints and don't
3356 * allow direct storage.
3358 vtn_fail("Invalid uniform atomic");
3362 static nir_intrinsic_op
3363 get_deref_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
3366 case SpvOpAtomicLoad
: return nir_intrinsic_load_deref
;
3367 case SpvOpAtomicStore
: return nir_intrinsic_store_deref
;
3368 #define OP(S, N) case SpvOp##S: return nir_intrinsic_deref_##N;
3369 OP(AtomicExchange
, atomic_exchange
)
3370 OP(AtomicCompareExchange
, atomic_comp_swap
)
3371 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
3372 OP(AtomicIIncrement
, atomic_add
)
3373 OP(AtomicIDecrement
, atomic_add
)
3374 OP(AtomicIAdd
, atomic_add
)
3375 OP(AtomicISub
, atomic_add
)
3376 OP(AtomicSMin
, atomic_imin
)
3377 OP(AtomicUMin
, atomic_umin
)
3378 OP(AtomicSMax
, atomic_imax
)
3379 OP(AtomicUMax
, atomic_umax
)
3380 OP(AtomicAnd
, atomic_and
)
3381 OP(AtomicOr
, atomic_or
)
3382 OP(AtomicXor
, atomic_xor
)
3383 OP(AtomicFAddEXT
, atomic_fadd
)
3386 vtn_fail_with_opcode("Invalid shared atomic", opcode
);
3391 * Handles shared atomics, ssbo atomics and atomic counters.
3394 vtn_handle_atomics(struct vtn_builder
*b
, SpvOp opcode
,
3395 const uint32_t *w
, UNUSED
unsigned count
)
3397 struct vtn_pointer
*ptr
;
3398 nir_intrinsic_instr
*atomic
;
3400 SpvScope scope
= SpvScopeInvocation
;
3401 SpvMemorySemanticsMask semantics
= 0;
3402 enum gl_access_qualifier access
= 0;
3405 case SpvOpAtomicLoad
:
3406 case SpvOpAtomicExchange
:
3407 case SpvOpAtomicCompareExchange
:
3408 case SpvOpAtomicCompareExchangeWeak
:
3409 case SpvOpAtomicIIncrement
:
3410 case SpvOpAtomicIDecrement
:
3411 case SpvOpAtomicIAdd
:
3412 case SpvOpAtomicISub
:
3413 case SpvOpAtomicSMin
:
3414 case SpvOpAtomicUMin
:
3415 case SpvOpAtomicSMax
:
3416 case SpvOpAtomicUMax
:
3417 case SpvOpAtomicAnd
:
3419 case SpvOpAtomicXor
:
3420 case SpvOpAtomicFAddEXT
:
3421 ptr
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
3422 scope
= vtn_constant_uint(b
, w
[4]);
3423 semantics
= vtn_constant_uint(b
, w
[5]);
3426 case SpvOpAtomicStore
:
3427 ptr
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
3428 scope
= vtn_constant_uint(b
, w
[2]);
3429 semantics
= vtn_constant_uint(b
, w
[3]);
3433 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
3436 if (semantics
& SpvMemorySemanticsVolatileMask
)
3437 access
|= ACCESS_VOLATILE
;
3439 /* uniform as "atomic counter uniform" */
3440 if (ptr
->mode
== vtn_variable_mode_atomic_counter
) {
3441 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
3442 nir_intrinsic_op op
= get_uniform_nir_atomic_op(b
, opcode
);
3443 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
3444 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
3446 /* SSBO needs to initialize index/offset. In this case we don't need to,
3447 * as that info is already stored on the ptr->var->var nir_variable (see
3448 * vtn_create_variable)
3452 case SpvOpAtomicLoad
:
3453 case SpvOpAtomicExchange
:
3454 case SpvOpAtomicCompareExchange
:
3455 case SpvOpAtomicCompareExchangeWeak
:
3456 case SpvOpAtomicIIncrement
:
3457 case SpvOpAtomicIDecrement
:
3458 case SpvOpAtomicIAdd
:
3459 case SpvOpAtomicISub
:
3460 case SpvOpAtomicSMin
:
3461 case SpvOpAtomicUMin
:
3462 case SpvOpAtomicSMax
:
3463 case SpvOpAtomicUMax
:
3464 case SpvOpAtomicAnd
:
3466 case SpvOpAtomicXor
:
3467 /* Nothing: we don't need to call fill_common_atomic_sources here, as
3468 * atomic counter uniforms doesn't have sources
3473 unreachable("Invalid SPIR-V atomic");
3476 } else if (vtn_pointer_uses_ssa_offset(b
, ptr
)) {
3477 nir_ssa_def
*offset
, *index
;
3478 offset
= vtn_pointer_to_offset(b
, ptr
, &index
);
3480 assert(ptr
->mode
== vtn_variable_mode_ssbo
);
3482 nir_intrinsic_op op
= get_ssbo_nir_atomic_op(b
, opcode
);
3483 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
3485 nir_intrinsic_set_access(atomic
, access
| ACCESS_COHERENT
);
3489 case SpvOpAtomicLoad
:
3490 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
3491 nir_intrinsic_set_align(atomic
, 4, 0);
3492 if (ptr
->mode
== vtn_variable_mode_ssbo
)
3493 atomic
->src
[src
++] = nir_src_for_ssa(index
);
3494 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
3497 case SpvOpAtomicStore
:
3498 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
3499 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
3500 nir_intrinsic_set_align(atomic
, 4, 0);
3501 atomic
->src
[src
++] = nir_src_for_ssa(vtn_get_nir_ssa(b
, w
[4]));
3502 if (ptr
->mode
== vtn_variable_mode_ssbo
)
3503 atomic
->src
[src
++] = nir_src_for_ssa(index
);
3504 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
3507 case SpvOpAtomicExchange
:
3508 case SpvOpAtomicCompareExchange
:
3509 case SpvOpAtomicCompareExchangeWeak
:
3510 case SpvOpAtomicIIncrement
:
3511 case SpvOpAtomicIDecrement
:
3512 case SpvOpAtomicIAdd
:
3513 case SpvOpAtomicISub
:
3514 case SpvOpAtomicSMin
:
3515 case SpvOpAtomicUMin
:
3516 case SpvOpAtomicSMax
:
3517 case SpvOpAtomicUMax
:
3518 case SpvOpAtomicAnd
:
3520 case SpvOpAtomicXor
:
3521 case SpvOpAtomicFAddEXT
:
3522 if (ptr
->mode
== vtn_variable_mode_ssbo
)
3523 atomic
->src
[src
++] = nir_src_for_ssa(index
);
3524 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
3525 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[src
]);
3529 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
3532 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
3533 const struct glsl_type
*deref_type
= deref
->type
;
3534 nir_intrinsic_op op
= get_deref_nir_atomic_op(b
, opcode
);
3535 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
3536 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
3538 if (ptr
->mode
!= vtn_variable_mode_workgroup
)
3539 access
|= ACCESS_COHERENT
;
3541 nir_intrinsic_set_access(atomic
, access
);
3544 case SpvOpAtomicLoad
:
3545 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
3548 case SpvOpAtomicStore
:
3549 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
3550 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
3551 atomic
->src
[1] = nir_src_for_ssa(vtn_get_nir_ssa(b
, w
[4]));
3554 case SpvOpAtomicExchange
:
3555 case SpvOpAtomicCompareExchange
:
3556 case SpvOpAtomicCompareExchangeWeak
:
3557 case SpvOpAtomicIIncrement
:
3558 case SpvOpAtomicIDecrement
:
3559 case SpvOpAtomicIAdd
:
3560 case SpvOpAtomicISub
:
3561 case SpvOpAtomicSMin
:
3562 case SpvOpAtomicUMin
:
3563 case SpvOpAtomicSMax
:
3564 case SpvOpAtomicUMax
:
3565 case SpvOpAtomicAnd
:
3567 case SpvOpAtomicXor
:
3568 case SpvOpAtomicFAddEXT
:
3569 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[1]);
3573 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
3577 /* Atomic ordering operations will implicitly apply to the atomic operation
3578 * storage class, so include that too.
3580 semantics
|= vtn_mode_to_memory_semantics(ptr
->mode
);
3582 SpvMemorySemanticsMask before_semantics
;
3583 SpvMemorySemanticsMask after_semantics
;
3584 vtn_split_barrier_semantics(b
, semantics
, &before_semantics
, &after_semantics
);
3586 if (before_semantics
)
3587 vtn_emit_memory_barrier(b
, scope
, before_semantics
);
3589 if (opcode
!= SpvOpAtomicStore
) {
3590 struct vtn_type
*type
= vtn_get_type(b
, w
[1]);
3592 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
,
3593 glsl_get_vector_elements(type
->type
),
3594 glsl_get_bit_size(type
->type
), NULL
);
3596 vtn_push_nir_ssa(b
, w
[2], &atomic
->dest
.ssa
);
3599 nir_builder_instr_insert(&b
->nb
, &atomic
->instr
);
3601 if (after_semantics
)
3602 vtn_emit_memory_barrier(b
, scope
, after_semantics
);
3605 static nir_alu_instr
*
3606 create_vec(struct vtn_builder
*b
, unsigned num_components
, unsigned bit_size
)
3608 nir_op op
= nir_op_vec(num_components
);
3609 nir_alu_instr
*vec
= nir_alu_instr_create(b
->shader
, op
);
3610 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, num_components
,
3612 vec
->dest
.write_mask
= (1 << num_components
) - 1;
3617 struct vtn_ssa_value
*
3618 vtn_ssa_transpose(struct vtn_builder
*b
, struct vtn_ssa_value
*src
)
3620 if (src
->transposed
)
3621 return src
->transposed
;
3623 struct vtn_ssa_value
*dest
=
3624 vtn_create_ssa_value(b
, glsl_transposed_type(src
->type
));
3626 for (unsigned i
= 0; i
< glsl_get_matrix_columns(dest
->type
); i
++) {
3627 nir_alu_instr
*vec
= create_vec(b
, glsl_get_matrix_columns(src
->type
),
3628 glsl_get_bit_size(src
->type
));
3629 if (glsl_type_is_vector_or_scalar(src
->type
)) {
3630 vec
->src
[0].src
= nir_src_for_ssa(src
->def
);
3631 vec
->src
[0].swizzle
[0] = i
;
3633 for (unsigned j
= 0; j
< glsl_get_matrix_columns(src
->type
); j
++) {
3634 vec
->src
[j
].src
= nir_src_for_ssa(src
->elems
[j
]->def
);
3635 vec
->src
[j
].swizzle
[0] = i
;
3638 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3639 dest
->elems
[i
]->def
= &vec
->dest
.dest
.ssa
;
3642 dest
->transposed
= src
;
3647 static nir_ssa_def
*
3648 vtn_vector_shuffle(struct vtn_builder
*b
, unsigned num_components
,
3649 nir_ssa_def
*src0
, nir_ssa_def
*src1
,
3650 const uint32_t *indices
)
3652 nir_alu_instr
*vec
= create_vec(b
, num_components
, src0
->bit_size
);
3654 for (unsigned i
= 0; i
< num_components
; i
++) {
3655 uint32_t index
= indices
[i
];
3656 if (index
== 0xffffffff) {
3658 nir_src_for_ssa(nir_ssa_undef(&b
->nb
, 1, src0
->bit_size
));
3659 } else if (index
< src0
->num_components
) {
3660 vec
->src
[i
].src
= nir_src_for_ssa(src0
);
3661 vec
->src
[i
].swizzle
[0] = index
;
3663 vec
->src
[i
].src
= nir_src_for_ssa(src1
);
3664 vec
->src
[i
].swizzle
[0] = index
- src0
->num_components
;
3668 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3670 return &vec
->dest
.dest
.ssa
;
3674 * Concatentates a number of vectors/scalars together to produce a vector
3676 static nir_ssa_def
*
3677 vtn_vector_construct(struct vtn_builder
*b
, unsigned num_components
,
3678 unsigned num_srcs
, nir_ssa_def
**srcs
)
3680 nir_alu_instr
*vec
= create_vec(b
, num_components
, srcs
[0]->bit_size
);
3682 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
3684 * "When constructing a vector, there must be at least two Constituent
3687 vtn_assert(num_srcs
>= 2);
3689 unsigned dest_idx
= 0;
3690 for (unsigned i
= 0; i
< num_srcs
; i
++) {
3691 nir_ssa_def
*src
= srcs
[i
];
3692 vtn_assert(dest_idx
+ src
->num_components
<= num_components
);
3693 for (unsigned j
= 0; j
< src
->num_components
; j
++) {
3694 vec
->src
[dest_idx
].src
= nir_src_for_ssa(src
);
3695 vec
->src
[dest_idx
].swizzle
[0] = j
;
3700 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
3702 * "When constructing a vector, the total number of components in all
3703 * the operands must equal the number of components in Result Type."
3705 vtn_assert(dest_idx
== num_components
);
3707 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3709 return &vec
->dest
.dest
.ssa
;
3712 static struct vtn_ssa_value
*
3713 vtn_composite_copy(void *mem_ctx
, struct vtn_ssa_value
*src
)
3715 struct vtn_ssa_value
*dest
= rzalloc(mem_ctx
, struct vtn_ssa_value
);
3716 dest
->type
= src
->type
;
3718 if (glsl_type_is_vector_or_scalar(src
->type
)) {
3719 dest
->def
= src
->def
;
3721 unsigned elems
= glsl_get_length(src
->type
);
3723 dest
->elems
= ralloc_array(mem_ctx
, struct vtn_ssa_value
*, elems
);
3724 for (unsigned i
= 0; i
< elems
; i
++)
3725 dest
->elems
[i
] = vtn_composite_copy(mem_ctx
, src
->elems
[i
]);
3731 static struct vtn_ssa_value
*
3732 vtn_composite_insert(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
3733 struct vtn_ssa_value
*insert
, const uint32_t *indices
,
3734 unsigned num_indices
)
3736 struct vtn_ssa_value
*dest
= vtn_composite_copy(b
, src
);
3738 struct vtn_ssa_value
*cur
= dest
;
3740 for (i
= 0; i
< num_indices
- 1; i
++) {
3741 /* If we got a vector here, that means the next index will be trying to
3742 * dereference a scalar.
3744 vtn_fail_if(glsl_type_is_vector_or_scalar(cur
->type
),
3745 "OpCompositeInsert has too many indices.");
3746 vtn_fail_if(indices
[i
] >= glsl_get_length(cur
->type
),
3747 "All indices in an OpCompositeInsert must be in-bounds");
3748 cur
= cur
->elems
[indices
[i
]];
3751 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
3752 vtn_fail_if(indices
[i
] >= glsl_get_vector_elements(cur
->type
),
3753 "All indices in an OpCompositeInsert must be in-bounds");
3755 /* According to the SPIR-V spec, OpCompositeInsert may work down to
3756 * the component granularity. In that case, the last index will be
3757 * the index to insert the scalar into the vector.
3760 cur
->def
= nir_vector_insert_imm(&b
->nb
, cur
->def
, insert
->def
, indices
[i
]);
3762 vtn_fail_if(indices
[i
] >= glsl_get_length(cur
->type
),
3763 "All indices in an OpCompositeInsert must be in-bounds");
3764 cur
->elems
[indices
[i
]] = insert
;
3770 static struct vtn_ssa_value
*
3771 vtn_composite_extract(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
3772 const uint32_t *indices
, unsigned num_indices
)
3774 struct vtn_ssa_value
*cur
= src
;
3775 for (unsigned i
= 0; i
< num_indices
; i
++) {
3776 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
3777 vtn_assert(i
== num_indices
- 1);
3778 vtn_fail_if(indices
[i
] >= glsl_get_vector_elements(cur
->type
),
3779 "All indices in an OpCompositeExtract must be in-bounds");
3781 /* According to the SPIR-V spec, OpCompositeExtract may work down to
3782 * the component granularity. The last index will be the index of the
3783 * vector to extract.
3786 const struct glsl_type
*scalar_type
=
3787 glsl_scalar_type(glsl_get_base_type(cur
->type
));
3788 struct vtn_ssa_value
*ret
= vtn_create_ssa_value(b
, scalar_type
);
3789 ret
->def
= nir_channel(&b
->nb
, cur
->def
, indices
[i
]);
3792 vtn_fail_if(indices
[i
] >= glsl_get_length(cur
->type
),
3793 "All indices in an OpCompositeExtract must be in-bounds");
3794 cur
= cur
->elems
[indices
[i
]];
3802 vtn_handle_composite(struct vtn_builder
*b
, SpvOp opcode
,
3803 const uint32_t *w
, unsigned count
)
3805 struct vtn_type
*type
= vtn_get_type(b
, w
[1]);
3806 struct vtn_ssa_value
*ssa
= vtn_create_ssa_value(b
, type
->type
);
3809 case SpvOpVectorExtractDynamic
:
3810 ssa
->def
= nir_vector_extract(&b
->nb
, vtn_get_nir_ssa(b
, w
[3]),
3811 vtn_get_nir_ssa(b
, w
[4]));
3814 case SpvOpVectorInsertDynamic
:
3815 ssa
->def
= nir_vector_insert(&b
->nb
, vtn_get_nir_ssa(b
, w
[3]),
3816 vtn_get_nir_ssa(b
, w
[4]),
3817 vtn_get_nir_ssa(b
, w
[5]));
3820 case SpvOpVectorShuffle
:
3821 ssa
->def
= vtn_vector_shuffle(b
, glsl_get_vector_elements(type
->type
),
3822 vtn_get_nir_ssa(b
, w
[3]),
3823 vtn_get_nir_ssa(b
, w
[4]),
3827 case SpvOpCompositeConstruct
: {
3828 unsigned elems
= count
- 3;
3830 if (glsl_type_is_vector_or_scalar(type
->type
)) {
3831 nir_ssa_def
*srcs
[NIR_MAX_VEC_COMPONENTS
];
3832 for (unsigned i
= 0; i
< elems
; i
++)
3833 srcs
[i
] = vtn_get_nir_ssa(b
, w
[3 + i
]);
3835 vtn_vector_construct(b
, glsl_get_vector_elements(type
->type
),
3838 ssa
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
3839 for (unsigned i
= 0; i
< elems
; i
++)
3840 ssa
->elems
[i
] = vtn_ssa_value(b
, w
[3 + i
]);
3844 case SpvOpCompositeExtract
:
3845 ssa
= vtn_composite_extract(b
, vtn_ssa_value(b
, w
[3]),
3849 case SpvOpCompositeInsert
:
3850 ssa
= vtn_composite_insert(b
, vtn_ssa_value(b
, w
[4]),
3851 vtn_ssa_value(b
, w
[3]),
3855 case SpvOpCopyLogical
:
3856 ssa
= vtn_composite_copy(b
, vtn_ssa_value(b
, w
[3]));
3858 case SpvOpCopyObject
:
3859 vtn_copy_value(b
, w
[3], w
[2]);
3863 vtn_fail_with_opcode("unknown composite operation", opcode
);
3866 vtn_push_ssa_value(b
, w
[2], ssa
);
3870 vtn_emit_barrier(struct vtn_builder
*b
, nir_intrinsic_op op
)
3872 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
3873 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3877 vtn_emit_memory_barrier(struct vtn_builder
*b
, SpvScope scope
,
3878 SpvMemorySemanticsMask semantics
)
3880 if (b
->shader
->options
->use_scoped_barrier
) {
3881 vtn_emit_scoped_memory_barrier(b
, scope
, semantics
);
3885 static const SpvMemorySemanticsMask all_memory_semantics
=
3886 SpvMemorySemanticsUniformMemoryMask
|
3887 SpvMemorySemanticsWorkgroupMemoryMask
|
3888 SpvMemorySemanticsAtomicCounterMemoryMask
|
3889 SpvMemorySemanticsImageMemoryMask
|
3890 SpvMemorySemanticsOutputMemoryMask
;
3892 /* If we're not actually doing a memory barrier, bail */
3893 if (!(semantics
& all_memory_semantics
))
3896 /* GL and Vulkan don't have these */
3897 vtn_assert(scope
!= SpvScopeCrossDevice
);
3899 if (scope
== SpvScopeSubgroup
)
3900 return; /* Nothing to do here */
3902 if (scope
== SpvScopeWorkgroup
) {
3903 vtn_emit_barrier(b
, nir_intrinsic_group_memory_barrier
);
3907 /* There's only two scopes thing left */
3908 vtn_assert(scope
== SpvScopeInvocation
|| scope
== SpvScopeDevice
);
3910 /* Map the GLSL memoryBarrier() construct and any barriers with more than one
3911 * semantic to the corresponding NIR one.
3913 if (util_bitcount(semantics
& all_memory_semantics
) > 1) {
3914 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier
);
3915 if (semantics
& SpvMemorySemanticsOutputMemoryMask
) {
3916 /* GLSL memoryBarrier() (and the corresponding NIR one) doesn't include
3917 * TCS outputs, so we have to emit it's own intrinsic for that. We
3918 * then need to emit another memory_barrier to prevent moving
3919 * non-output operations to before the tcs_patch barrier.
3921 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_tcs_patch
);
3922 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier
);
3927 /* Issue a more specific barrier */
3928 switch (semantics
& all_memory_semantics
) {
3929 case SpvMemorySemanticsUniformMemoryMask
:
3930 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_buffer
);
3932 case SpvMemorySemanticsWorkgroupMemoryMask
:
3933 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_shared
);
3935 case SpvMemorySemanticsAtomicCounterMemoryMask
:
3936 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_atomic_counter
);
3938 case SpvMemorySemanticsImageMemoryMask
:
3939 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_image
);
3941 case SpvMemorySemanticsOutputMemoryMask
:
3942 if (b
->nb
.shader
->info
.stage
== MESA_SHADER_TESS_CTRL
)
3943 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_tcs_patch
);
3951 vtn_handle_barrier(struct vtn_builder
*b
, SpvOp opcode
,
3952 const uint32_t *w
, UNUSED
unsigned count
)
3955 case SpvOpEmitVertex
:
3956 case SpvOpEmitStreamVertex
:
3957 case SpvOpEndPrimitive
:
3958 case SpvOpEndStreamPrimitive
: {
3959 nir_intrinsic_op intrinsic_op
;
3961 case SpvOpEmitVertex
:
3962 case SpvOpEmitStreamVertex
:
3963 intrinsic_op
= nir_intrinsic_emit_vertex
;
3965 case SpvOpEndPrimitive
:
3966 case SpvOpEndStreamPrimitive
:
3967 intrinsic_op
= nir_intrinsic_end_primitive
;
3970 unreachable("Invalid opcode");
3973 nir_intrinsic_instr
*intrin
=
3974 nir_intrinsic_instr_create(b
->shader
, intrinsic_op
);
3977 case SpvOpEmitStreamVertex
:
3978 case SpvOpEndStreamPrimitive
: {
3979 unsigned stream
= vtn_constant_uint(b
, w
[1]);
3980 nir_intrinsic_set_stream_id(intrin
, stream
);
3988 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3992 case SpvOpMemoryBarrier
: {
3993 SpvScope scope
= vtn_constant_uint(b
, w
[1]);
3994 SpvMemorySemanticsMask semantics
= vtn_constant_uint(b
, w
[2]);
3995 vtn_emit_memory_barrier(b
, scope
, semantics
);
3999 case SpvOpControlBarrier
: {
4000 SpvScope execution_scope
= vtn_constant_uint(b
, w
[1]);
4001 SpvScope memory_scope
= vtn_constant_uint(b
, w
[2]);
4002 SpvMemorySemanticsMask memory_semantics
= vtn_constant_uint(b
, w
[3]);
4004 /* GLSLang, prior to commit 8297936dd6eb3, emitted OpControlBarrier with
4005 * memory semantics of None for GLSL barrier().
4006 * And before that, prior to c3f1cdfa, emitted the OpControlBarrier with
4007 * Device instead of Workgroup for execution scope.
4009 if (b
->wa_glslang_cs_barrier
&&
4010 b
->nb
.shader
->info
.stage
== MESA_SHADER_COMPUTE
&&
4011 (execution_scope
== SpvScopeWorkgroup
||
4012 execution_scope
== SpvScopeDevice
) &&
4013 memory_semantics
== SpvMemorySemanticsMaskNone
) {
4014 execution_scope
= SpvScopeWorkgroup
;
4015 memory_scope
= SpvScopeWorkgroup
;
4016 memory_semantics
= SpvMemorySemanticsAcquireReleaseMask
|
4017 SpvMemorySemanticsWorkgroupMemoryMask
;
4020 /* From the SPIR-V spec:
4022 * "When used with the TessellationControl execution model, it also
4023 * implicitly synchronizes the Output Storage Class: Writes to Output
4024 * variables performed by any invocation executed prior to a
4025 * OpControlBarrier will be visible to any other invocation after
4026 * return from that OpControlBarrier."
4028 if (b
->nb
.shader
->info
.stage
== MESA_SHADER_TESS_CTRL
) {
4029 memory_semantics
&= ~(SpvMemorySemanticsAcquireMask
|
4030 SpvMemorySemanticsReleaseMask
|
4031 SpvMemorySemanticsAcquireReleaseMask
|
4032 SpvMemorySemanticsSequentiallyConsistentMask
);
4033 memory_semantics
|= SpvMemorySemanticsAcquireReleaseMask
|
4034 SpvMemorySemanticsOutputMemoryMask
;
4037 if (b
->shader
->options
->use_scoped_barrier
) {
4038 vtn_emit_scoped_control_barrier(b
, execution_scope
, memory_scope
,
4041 vtn_emit_memory_barrier(b
, memory_scope
, memory_semantics
);
4043 if (execution_scope
== SpvScopeWorkgroup
)
4044 vtn_emit_barrier(b
, nir_intrinsic_control_barrier
);
4050 unreachable("unknown barrier instruction");
4055 gl_primitive_from_spv_execution_mode(struct vtn_builder
*b
,
4056 SpvExecutionMode mode
)
4059 case SpvExecutionModeInputPoints
:
4060 case SpvExecutionModeOutputPoints
:
4061 return 0; /* GL_POINTS */
4062 case SpvExecutionModeInputLines
:
4063 return 1; /* GL_LINES */
4064 case SpvExecutionModeInputLinesAdjacency
:
4065 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
4066 case SpvExecutionModeTriangles
:
4067 return 4; /* GL_TRIANGLES */
4068 case SpvExecutionModeInputTrianglesAdjacency
:
4069 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
4070 case SpvExecutionModeQuads
:
4071 return 7; /* GL_QUADS */
4072 case SpvExecutionModeIsolines
:
4073 return 0x8E7A; /* GL_ISOLINES */
4074 case SpvExecutionModeOutputLineStrip
:
4075 return 3; /* GL_LINE_STRIP */
4076 case SpvExecutionModeOutputTriangleStrip
:
4077 return 5; /* GL_TRIANGLE_STRIP */
4079 vtn_fail("Invalid primitive type: %s (%u)",
4080 spirv_executionmode_to_string(mode
), mode
);
4085 vertices_in_from_spv_execution_mode(struct vtn_builder
*b
,
4086 SpvExecutionMode mode
)
4089 case SpvExecutionModeInputPoints
:
4091 case SpvExecutionModeInputLines
:
4093 case SpvExecutionModeInputLinesAdjacency
:
4095 case SpvExecutionModeTriangles
:
4097 case SpvExecutionModeInputTrianglesAdjacency
:
4100 vtn_fail("Invalid GS input mode: %s (%u)",
4101 spirv_executionmode_to_string(mode
), mode
);
4105 static gl_shader_stage
4106 stage_for_execution_model(struct vtn_builder
*b
, SpvExecutionModel model
)
4109 case SpvExecutionModelVertex
:
4110 return MESA_SHADER_VERTEX
;
4111 case SpvExecutionModelTessellationControl
:
4112 return MESA_SHADER_TESS_CTRL
;
4113 case SpvExecutionModelTessellationEvaluation
:
4114 return MESA_SHADER_TESS_EVAL
;
4115 case SpvExecutionModelGeometry
:
4116 return MESA_SHADER_GEOMETRY
;
4117 case SpvExecutionModelFragment
:
4118 return MESA_SHADER_FRAGMENT
;
4119 case SpvExecutionModelGLCompute
:
4120 return MESA_SHADER_COMPUTE
;
4121 case SpvExecutionModelKernel
:
4122 return MESA_SHADER_KERNEL
;
4124 vtn_fail("Unsupported execution model: %s (%u)",
4125 spirv_executionmodel_to_string(model
), model
);
4129 #define spv_check_supported(name, cap) do { \
4130 if (!(b->options && b->options->caps.name)) \
4131 vtn_warn("Unsupported SPIR-V capability: %s (%u)", \
4132 spirv_capability_to_string(cap), cap); \
4137 vtn_handle_entry_point(struct vtn_builder
*b
, const uint32_t *w
,
4140 struct vtn_value
*entry_point
= &b
->values
[w
[2]];
4141 /* Let this be a name label regardless */
4142 unsigned name_words
;
4143 entry_point
->name
= vtn_string_literal(b
, &w
[3], count
- 3, &name_words
);
4145 if (strcmp(entry_point
->name
, b
->entry_point_name
) != 0 ||
4146 stage_for_execution_model(b
, w
[1]) != b
->entry_point_stage
)
4149 vtn_assert(b
->entry_point
== NULL
);
4150 b
->entry_point
= entry_point
;
4154 vtn_handle_preamble_instruction(struct vtn_builder
*b
, SpvOp opcode
,
4155 const uint32_t *w
, unsigned count
)
4162 case SpvSourceLanguageUnknown
: lang
= "unknown"; break;
4163 case SpvSourceLanguageESSL
: lang
= "ESSL"; break;
4164 case SpvSourceLanguageGLSL
: lang
= "GLSL"; break;
4165 case SpvSourceLanguageOpenCL_C
: lang
= "OpenCL C"; break;
4166 case SpvSourceLanguageOpenCL_CPP
: lang
= "OpenCL C++"; break;
4167 case SpvSourceLanguageHLSL
: lang
= "HLSL"; break;
4170 uint32_t version
= w
[2];
4173 (count
> 3) ? vtn_value(b
, w
[3], vtn_value_type_string
)->str
: "";
4175 vtn_info("Parsing SPIR-V from %s %u source file %s", lang
, version
, file
);
4179 case SpvOpSourceExtension
:
4180 case SpvOpSourceContinued
:
4181 case SpvOpExtension
:
4182 case SpvOpModuleProcessed
:
4183 /* Unhandled, but these are for debug so that's ok. */
4186 case SpvOpCapability
: {
4187 SpvCapability cap
= w
[1];
4189 case SpvCapabilityMatrix
:
4190 case SpvCapabilityShader
:
4191 case SpvCapabilityGeometry
:
4192 case SpvCapabilityGeometryPointSize
:
4193 case SpvCapabilityUniformBufferArrayDynamicIndexing
:
4194 case SpvCapabilitySampledImageArrayDynamicIndexing
:
4195 case SpvCapabilityStorageBufferArrayDynamicIndexing
:
4196 case SpvCapabilityStorageImageArrayDynamicIndexing
:
4197 case SpvCapabilityImageRect
:
4198 case SpvCapabilitySampledRect
:
4199 case SpvCapabilitySampled1D
:
4200 case SpvCapabilityImage1D
:
4201 case SpvCapabilitySampledCubeArray
:
4202 case SpvCapabilityImageCubeArray
:
4203 case SpvCapabilitySampledBuffer
:
4204 case SpvCapabilityImageBuffer
:
4205 case SpvCapabilityImageQuery
:
4206 case SpvCapabilityDerivativeControl
:
4207 case SpvCapabilityInterpolationFunction
:
4208 case SpvCapabilityMultiViewport
:
4209 case SpvCapabilitySampleRateShading
:
4210 case SpvCapabilityClipDistance
:
4211 case SpvCapabilityCullDistance
:
4212 case SpvCapabilityInputAttachment
:
4213 case SpvCapabilityImageGatherExtended
:
4214 case SpvCapabilityStorageImageExtendedFormats
:
4215 case SpvCapabilityVector16
:
4218 case SpvCapabilityLinkage
:
4219 case SpvCapabilityFloat16Buffer
:
4220 case SpvCapabilitySparseResidency
:
4221 vtn_warn("Unsupported SPIR-V capability: %s",
4222 spirv_capability_to_string(cap
));
4225 case SpvCapabilityMinLod
:
4226 spv_check_supported(min_lod
, cap
);
4229 case SpvCapabilityAtomicStorage
:
4230 spv_check_supported(atomic_storage
, cap
);
4233 case SpvCapabilityFloat64
:
4234 spv_check_supported(float64
, cap
);
4236 case SpvCapabilityInt64
:
4237 spv_check_supported(int64
, cap
);
4239 case SpvCapabilityInt16
:
4240 spv_check_supported(int16
, cap
);
4242 case SpvCapabilityInt8
:
4243 spv_check_supported(int8
, cap
);
4246 case SpvCapabilityTransformFeedback
:
4247 spv_check_supported(transform_feedback
, cap
);
4250 case SpvCapabilityGeometryStreams
:
4251 spv_check_supported(geometry_streams
, cap
);
4254 case SpvCapabilityInt64Atomics
:
4255 spv_check_supported(int64_atomics
, cap
);
4258 case SpvCapabilityStorageImageMultisample
:
4259 spv_check_supported(storage_image_ms
, cap
);
4262 case SpvCapabilityAddresses
:
4263 spv_check_supported(address
, cap
);
4266 case SpvCapabilityKernel
:
4267 spv_check_supported(kernel
, cap
);
4270 case SpvCapabilityImageBasic
:
4271 spv_check_supported(kernel_image
, cap
);
4274 case SpvCapabilityLiteralSampler
:
4275 spv_check_supported(literal_sampler
, cap
);
4278 case SpvCapabilityImageReadWrite
:
4279 case SpvCapabilityImageMipmap
:
4280 case SpvCapabilityPipes
:
4281 case SpvCapabilityDeviceEnqueue
:
4282 case SpvCapabilityGenericPointer
:
4283 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
4284 spirv_capability_to_string(cap
));
4287 case SpvCapabilityImageMSArray
:
4288 spv_check_supported(image_ms_array
, cap
);
4291 case SpvCapabilityTessellation
:
4292 case SpvCapabilityTessellationPointSize
:
4293 spv_check_supported(tessellation
, cap
);
4296 case SpvCapabilityDrawParameters
:
4297 spv_check_supported(draw_parameters
, cap
);
4300 case SpvCapabilityStorageImageReadWithoutFormat
:
4301 spv_check_supported(image_read_without_format
, cap
);
4304 case SpvCapabilityStorageImageWriteWithoutFormat
:
4305 spv_check_supported(image_write_without_format
, cap
);
4308 case SpvCapabilityDeviceGroup
:
4309 spv_check_supported(device_group
, cap
);
4312 case SpvCapabilityMultiView
:
4313 spv_check_supported(multiview
, cap
);
4316 case SpvCapabilityGroupNonUniform
:
4317 spv_check_supported(subgroup_basic
, cap
);
4320 case SpvCapabilitySubgroupVoteKHR
:
4321 case SpvCapabilityGroupNonUniformVote
:
4322 spv_check_supported(subgroup_vote
, cap
);
4325 case SpvCapabilitySubgroupBallotKHR
:
4326 case SpvCapabilityGroupNonUniformBallot
:
4327 spv_check_supported(subgroup_ballot
, cap
);
4330 case SpvCapabilityGroupNonUniformShuffle
:
4331 case SpvCapabilityGroupNonUniformShuffleRelative
:
4332 spv_check_supported(subgroup_shuffle
, cap
);
4335 case SpvCapabilityGroupNonUniformQuad
:
4336 spv_check_supported(subgroup_quad
, cap
);
4339 case SpvCapabilityGroupNonUniformArithmetic
:
4340 case SpvCapabilityGroupNonUniformClustered
:
4341 spv_check_supported(subgroup_arithmetic
, cap
);
4344 case SpvCapabilityGroups
:
4345 spv_check_supported(amd_shader_ballot
, cap
);
4348 case SpvCapabilityVariablePointersStorageBuffer
:
4349 case SpvCapabilityVariablePointers
:
4350 spv_check_supported(variable_pointers
, cap
);
4351 b
->variable_pointers
= true;
4354 case SpvCapabilityStorageUniformBufferBlock16
:
4355 case SpvCapabilityStorageUniform16
:
4356 case SpvCapabilityStoragePushConstant16
:
4357 case SpvCapabilityStorageInputOutput16
:
4358 spv_check_supported(storage_16bit
, cap
);
4361 case SpvCapabilityShaderLayer
:
4362 case SpvCapabilityShaderViewportIndex
:
4363 case SpvCapabilityShaderViewportIndexLayerEXT
:
4364 spv_check_supported(shader_viewport_index_layer
, cap
);
4367 case SpvCapabilityStorageBuffer8BitAccess
:
4368 case SpvCapabilityUniformAndStorageBuffer8BitAccess
:
4369 case SpvCapabilityStoragePushConstant8
:
4370 spv_check_supported(storage_8bit
, cap
);
4373 case SpvCapabilityShaderNonUniformEXT
:
4374 spv_check_supported(descriptor_indexing
, cap
);
4377 case SpvCapabilityInputAttachmentArrayDynamicIndexingEXT
:
4378 case SpvCapabilityUniformTexelBufferArrayDynamicIndexingEXT
:
4379 case SpvCapabilityStorageTexelBufferArrayDynamicIndexingEXT
:
4380 spv_check_supported(descriptor_array_dynamic_indexing
, cap
);
4383 case SpvCapabilityUniformBufferArrayNonUniformIndexingEXT
:
4384 case SpvCapabilitySampledImageArrayNonUniformIndexingEXT
:
4385 case SpvCapabilityStorageBufferArrayNonUniformIndexingEXT
:
4386 case SpvCapabilityStorageImageArrayNonUniformIndexingEXT
:
4387 case SpvCapabilityInputAttachmentArrayNonUniformIndexingEXT
:
4388 case SpvCapabilityUniformTexelBufferArrayNonUniformIndexingEXT
:
4389 case SpvCapabilityStorageTexelBufferArrayNonUniformIndexingEXT
:
4390 spv_check_supported(descriptor_array_non_uniform_indexing
, cap
);
4393 case SpvCapabilityRuntimeDescriptorArrayEXT
:
4394 spv_check_supported(runtime_descriptor_array
, cap
);
4397 case SpvCapabilityStencilExportEXT
:
4398 spv_check_supported(stencil_export
, cap
);
4401 case SpvCapabilitySampleMaskPostDepthCoverage
:
4402 spv_check_supported(post_depth_coverage
, cap
);
4405 case SpvCapabilityDenormFlushToZero
:
4406 case SpvCapabilityDenormPreserve
:
4407 case SpvCapabilitySignedZeroInfNanPreserve
:
4408 case SpvCapabilityRoundingModeRTE
:
4409 case SpvCapabilityRoundingModeRTZ
:
4410 spv_check_supported(float_controls
, cap
);
4413 case SpvCapabilityPhysicalStorageBufferAddresses
:
4414 spv_check_supported(physical_storage_buffer_address
, cap
);
4417 case SpvCapabilityComputeDerivativeGroupQuadsNV
:
4418 case SpvCapabilityComputeDerivativeGroupLinearNV
:
4419 spv_check_supported(derivative_group
, cap
);
4422 case SpvCapabilityFloat16
:
4423 spv_check_supported(float16
, cap
);
4426 case SpvCapabilityFragmentShaderSampleInterlockEXT
:
4427 spv_check_supported(fragment_shader_sample_interlock
, cap
);
4430 case SpvCapabilityFragmentShaderPixelInterlockEXT
:
4431 spv_check_supported(fragment_shader_pixel_interlock
, cap
);
4434 case SpvCapabilityDemoteToHelperInvocationEXT
:
4435 spv_check_supported(demote_to_helper_invocation
, cap
);
4438 case SpvCapabilityShaderClockKHR
:
4439 spv_check_supported(shader_clock
, cap
);
4442 case SpvCapabilityVulkanMemoryModel
:
4443 spv_check_supported(vk_memory_model
, cap
);
4446 case SpvCapabilityVulkanMemoryModelDeviceScope
:
4447 spv_check_supported(vk_memory_model_device_scope
, cap
);
4450 case SpvCapabilityImageReadWriteLodAMD
:
4451 spv_check_supported(amd_image_read_write_lod
, cap
);
4454 case SpvCapabilityIntegerFunctions2INTEL
:
4455 spv_check_supported(integer_functions2
, cap
);
4458 case SpvCapabilityFragmentMaskAMD
:
4459 spv_check_supported(amd_fragment_mask
, cap
);
4462 case SpvCapabilityImageGatherBiasLodAMD
:
4463 spv_check_supported(amd_image_gather_bias_lod
, cap
);
4466 case SpvCapabilityAtomicFloat32AddEXT
:
4467 spv_check_supported(float32_atomic_add
, cap
);
4470 case SpvCapabilityAtomicFloat64AddEXT
:
4471 spv_check_supported(float64_atomic_add
, cap
);
4475 vtn_fail("Unhandled capability: %s (%u)",
4476 spirv_capability_to_string(cap
), cap
);
4481 case SpvOpExtInstImport
:
4482 vtn_handle_extension(b
, opcode
, w
, count
);
4485 case SpvOpMemoryModel
:
4487 case SpvAddressingModelPhysical32
:
4488 vtn_fail_if(b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
,
4489 "AddressingModelPhysical32 only supported for kernels");
4490 b
->shader
->info
.cs
.ptr_size
= 32;
4491 b
->physical_ptrs
= true;
4492 assert(nir_address_format_bit_size(b
->options
->global_addr_format
) == 32);
4493 assert(nir_address_format_num_components(b
->options
->global_addr_format
) == 1);
4494 assert(nir_address_format_bit_size(b
->options
->shared_addr_format
) == 32);
4495 assert(nir_address_format_num_components(b
->options
->shared_addr_format
) == 1);
4496 assert(nir_address_format_bit_size(b
->options
->constant_addr_format
) == 32);
4497 assert(nir_address_format_num_components(b
->options
->constant_addr_format
) == 1);
4499 case SpvAddressingModelPhysical64
:
4500 vtn_fail_if(b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
,
4501 "AddressingModelPhysical64 only supported for kernels");
4502 b
->shader
->info
.cs
.ptr_size
= 64;
4503 b
->physical_ptrs
= true;
4504 assert(nir_address_format_bit_size(b
->options
->global_addr_format
) == 64);
4505 assert(nir_address_format_num_components(b
->options
->global_addr_format
) == 1);
4506 assert(nir_address_format_bit_size(b
->options
->shared_addr_format
) == 64);
4507 assert(nir_address_format_num_components(b
->options
->shared_addr_format
) == 1);
4508 assert(nir_address_format_bit_size(b
->options
->constant_addr_format
) == 64);
4509 assert(nir_address_format_num_components(b
->options
->constant_addr_format
) == 1);
4511 case SpvAddressingModelLogical
:
4512 vtn_fail_if(b
->shader
->info
.stage
== MESA_SHADER_KERNEL
,
4513 "AddressingModelLogical only supported for shaders");
4514 b
->physical_ptrs
= false;
4516 case SpvAddressingModelPhysicalStorageBuffer64
:
4517 vtn_fail_if(!b
->options
||
4518 !b
->options
->caps
.physical_storage_buffer_address
,
4519 "AddressingModelPhysicalStorageBuffer64 not supported");
4522 vtn_fail("Unknown addressing model: %s (%u)",
4523 spirv_addressingmodel_to_string(w
[1]), w
[1]);
4527 b
->mem_model
= w
[2];
4529 case SpvMemoryModelSimple
:
4530 case SpvMemoryModelGLSL450
:
4531 case SpvMemoryModelOpenCL
:
4533 case SpvMemoryModelVulkan
:
4534 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
4535 "Vulkan memory model is unsupported by this driver");
4538 vtn_fail("Unsupported memory model: %s",
4539 spirv_memorymodel_to_string(w
[2]));
4544 case SpvOpEntryPoint
:
4545 vtn_handle_entry_point(b
, w
, count
);
4549 vtn_push_value(b
, w
[1], vtn_value_type_string
)->str
=
4550 vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
4554 b
->values
[w
[1]].name
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
4557 case SpvOpMemberName
:
4561 case SpvOpExecutionMode
:
4562 case SpvOpExecutionModeId
:
4563 case SpvOpDecorationGroup
:
4565 case SpvOpDecorateId
:
4566 case SpvOpMemberDecorate
:
4567 case SpvOpGroupDecorate
:
4568 case SpvOpGroupMemberDecorate
:
4569 case SpvOpDecorateString
:
4570 case SpvOpMemberDecorateString
:
4571 vtn_handle_decoration(b
, opcode
, w
, count
);
4574 case SpvOpExtInst
: {
4575 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
4576 if (val
->ext_handler
== vtn_handle_non_semantic_instruction
) {
4577 /* NonSemantic extended instructions are acceptable in preamble. */
4578 vtn_handle_non_semantic_instruction(b
, w
[4], w
, count
);
4581 return false; /* End of preamble. */
4586 return false; /* End of preamble */
4593 vtn_handle_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
4594 const struct vtn_decoration
*mode
, UNUSED
void *data
)
4596 vtn_assert(b
->entry_point
== entry_point
);
4598 switch(mode
->exec_mode
) {
4599 case SpvExecutionModeOriginUpperLeft
:
4600 case SpvExecutionModeOriginLowerLeft
:
4601 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4602 b
->shader
->info
.fs
.origin_upper_left
=
4603 (mode
->exec_mode
== SpvExecutionModeOriginUpperLeft
);
4606 case SpvExecutionModeEarlyFragmentTests
:
4607 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4608 b
->shader
->info
.fs
.early_fragment_tests
= true;
4611 case SpvExecutionModePostDepthCoverage
:
4612 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4613 b
->shader
->info
.fs
.post_depth_coverage
= true;
4616 case SpvExecutionModeInvocations
:
4617 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4618 b
->shader
->info
.gs
.invocations
= MAX2(1, mode
->operands
[0]);
4621 case SpvExecutionModeDepthReplacing
:
4622 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4623 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_ANY
;
4625 case SpvExecutionModeDepthGreater
:
4626 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4627 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_GREATER
;
4629 case SpvExecutionModeDepthLess
:
4630 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4631 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_LESS
;
4633 case SpvExecutionModeDepthUnchanged
:
4634 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4635 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
4638 case SpvExecutionModeLocalSize
:
4639 vtn_assert(gl_shader_stage_is_compute(b
->shader
->info
.stage
));
4640 b
->shader
->info
.cs
.local_size
[0] = mode
->operands
[0];
4641 b
->shader
->info
.cs
.local_size
[1] = mode
->operands
[1];
4642 b
->shader
->info
.cs
.local_size
[2] = mode
->operands
[2];
4645 case SpvExecutionModeLocalSizeHint
:
4646 break; /* Nothing to do with this */
4648 case SpvExecutionModeOutputVertices
:
4649 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4650 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
4651 b
->shader
->info
.tess
.tcs_vertices_out
= mode
->operands
[0];
4653 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4654 b
->shader
->info
.gs
.vertices_out
= mode
->operands
[0];
4658 case SpvExecutionModeInputPoints
:
4659 case SpvExecutionModeInputLines
:
4660 case SpvExecutionModeInputLinesAdjacency
:
4661 case SpvExecutionModeTriangles
:
4662 case SpvExecutionModeInputTrianglesAdjacency
:
4663 case SpvExecutionModeQuads
:
4664 case SpvExecutionModeIsolines
:
4665 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4666 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
4667 b
->shader
->info
.tess
.primitive_mode
=
4668 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
4670 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4671 b
->shader
->info
.gs
.vertices_in
=
4672 vertices_in_from_spv_execution_mode(b
, mode
->exec_mode
);
4673 b
->shader
->info
.gs
.input_primitive
=
4674 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
4678 case SpvExecutionModeOutputPoints
:
4679 case SpvExecutionModeOutputLineStrip
:
4680 case SpvExecutionModeOutputTriangleStrip
:
4681 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4682 b
->shader
->info
.gs
.output_primitive
=
4683 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
4686 case SpvExecutionModeSpacingEqual
:
4687 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4688 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4689 b
->shader
->info
.tess
.spacing
= TESS_SPACING_EQUAL
;
4691 case SpvExecutionModeSpacingFractionalEven
:
4692 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4693 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4694 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_EVEN
;
4696 case SpvExecutionModeSpacingFractionalOdd
:
4697 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4698 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4699 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_ODD
;
4701 case SpvExecutionModeVertexOrderCw
:
4702 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4703 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4704 b
->shader
->info
.tess
.ccw
= false;
4706 case SpvExecutionModeVertexOrderCcw
:
4707 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4708 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4709 b
->shader
->info
.tess
.ccw
= true;
4711 case SpvExecutionModePointMode
:
4712 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4713 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4714 b
->shader
->info
.tess
.point_mode
= true;
4717 case SpvExecutionModePixelCenterInteger
:
4718 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4719 b
->shader
->info
.fs
.pixel_center_integer
= true;
4722 case SpvExecutionModeXfb
:
4723 b
->shader
->info
.has_transform_feedback_varyings
= true;
4726 case SpvExecutionModeVecTypeHint
:
4729 case SpvExecutionModeContractionOff
:
4730 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
)
4731 vtn_warn("ExectionMode only allowed for CL-style kernels: %s",
4732 spirv_executionmode_to_string(mode
->exec_mode
));
4737 case SpvExecutionModeStencilRefReplacingEXT
:
4738 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4741 case SpvExecutionModeDerivativeGroupQuadsNV
:
4742 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
4743 b
->shader
->info
.cs
.derivative_group
= DERIVATIVE_GROUP_QUADS
;
4746 case SpvExecutionModeDerivativeGroupLinearNV
:
4747 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
4748 b
->shader
->info
.cs
.derivative_group
= DERIVATIVE_GROUP_LINEAR
;
4751 case SpvExecutionModePixelInterlockOrderedEXT
:
4752 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4753 b
->shader
->info
.fs
.pixel_interlock_ordered
= true;
4756 case SpvExecutionModePixelInterlockUnorderedEXT
:
4757 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4758 b
->shader
->info
.fs
.pixel_interlock_unordered
= true;
4761 case SpvExecutionModeSampleInterlockOrderedEXT
:
4762 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4763 b
->shader
->info
.fs
.sample_interlock_ordered
= true;
4766 case SpvExecutionModeSampleInterlockUnorderedEXT
:
4767 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4768 b
->shader
->info
.fs
.sample_interlock_unordered
= true;
4771 case SpvExecutionModeDenormPreserve
:
4772 case SpvExecutionModeDenormFlushToZero
:
4773 case SpvExecutionModeSignedZeroInfNanPreserve
:
4774 case SpvExecutionModeRoundingModeRTE
:
4775 case SpvExecutionModeRoundingModeRTZ
: {
4776 unsigned execution_mode
= 0;
4777 switch (mode
->exec_mode
) {
4778 case SpvExecutionModeDenormPreserve
:
4779 switch (mode
->operands
[0]) {
4780 case 16: execution_mode
= FLOAT_CONTROLS_DENORM_PRESERVE_FP16
; break;
4781 case 32: execution_mode
= FLOAT_CONTROLS_DENORM_PRESERVE_FP32
; break;
4782 case 64: execution_mode
= FLOAT_CONTROLS_DENORM_PRESERVE_FP64
; break;
4783 default: vtn_fail("Floating point type not supported");
4786 case SpvExecutionModeDenormFlushToZero
:
4787 switch (mode
->operands
[0]) {
4788 case 16: execution_mode
= FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16
; break;
4789 case 32: execution_mode
= FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32
; break;
4790 case 64: execution_mode
= FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP64
; break;
4791 default: vtn_fail("Floating point type not supported");
4794 case SpvExecutionModeSignedZeroInfNanPreserve
:
4795 switch (mode
->operands
[0]) {
4796 case 16: execution_mode
= FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP16
; break;
4797 case 32: execution_mode
= FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP32
; break;
4798 case 64: execution_mode
= FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP64
; break;
4799 default: vtn_fail("Floating point type not supported");
4802 case SpvExecutionModeRoundingModeRTE
:
4803 switch (mode
->operands
[0]) {
4804 case 16: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
; break;
4805 case 32: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
; break;
4806 case 64: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
; break;
4807 default: vtn_fail("Floating point type not supported");
4810 case SpvExecutionModeRoundingModeRTZ
:
4811 switch (mode
->operands
[0]) {
4812 case 16: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
; break;
4813 case 32: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
; break;
4814 case 64: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
; break;
4815 default: vtn_fail("Floating point type not supported");
4822 b
->shader
->info
.float_controls_execution_mode
|= execution_mode
;
4826 case SpvExecutionModeLocalSizeId
:
4827 case SpvExecutionModeLocalSizeHintId
:
4828 /* Handled later by vtn_handle_execution_mode_id(). */
4832 vtn_fail("Unhandled execution mode: %s (%u)",
4833 spirv_executionmode_to_string(mode
->exec_mode
),
4839 vtn_handle_execution_mode_id(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
4840 const struct vtn_decoration
*mode
, UNUSED
void *data
)
4843 vtn_assert(b
->entry_point
== entry_point
);
4845 switch (mode
->exec_mode
) {
4846 case SpvExecutionModeLocalSizeId
:
4847 b
->shader
->info
.cs
.local_size
[0] = vtn_constant_uint(b
, mode
->operands
[0]);
4848 b
->shader
->info
.cs
.local_size
[1] = vtn_constant_uint(b
, mode
->operands
[1]);
4849 b
->shader
->info
.cs
.local_size
[2] = vtn_constant_uint(b
, mode
->operands
[2]);
4852 case SpvExecutionModeLocalSizeHintId
:
4853 /* Nothing to do with this hint. */
4857 /* Nothing to do. Literal execution modes already handled by
4858 * vtn_handle_execution_mode(). */
4864 vtn_handle_variable_or_type_instruction(struct vtn_builder
*b
, SpvOp opcode
,
4865 const uint32_t *w
, unsigned count
)
4867 vtn_set_instruction_result_type(b
, opcode
, w
, count
);
4871 case SpvOpSourceContinued
:
4872 case SpvOpSourceExtension
:
4873 case SpvOpExtension
:
4874 case SpvOpCapability
:
4875 case SpvOpExtInstImport
:
4876 case SpvOpMemoryModel
:
4877 case SpvOpEntryPoint
:
4878 case SpvOpExecutionMode
:
4881 case SpvOpMemberName
:
4882 case SpvOpDecorationGroup
:
4884 case SpvOpDecorateId
:
4885 case SpvOpMemberDecorate
:
4886 case SpvOpGroupDecorate
:
4887 case SpvOpGroupMemberDecorate
:
4888 case SpvOpDecorateString
:
4889 case SpvOpMemberDecorateString
:
4890 vtn_fail("Invalid opcode types and variables section");
4896 case SpvOpTypeFloat
:
4897 case SpvOpTypeVector
:
4898 case SpvOpTypeMatrix
:
4899 case SpvOpTypeImage
:
4900 case SpvOpTypeSampler
:
4901 case SpvOpTypeSampledImage
:
4902 case SpvOpTypeArray
:
4903 case SpvOpTypeRuntimeArray
:
4904 case SpvOpTypeStruct
:
4905 case SpvOpTypeOpaque
:
4906 case SpvOpTypePointer
:
4907 case SpvOpTypeForwardPointer
:
4908 case SpvOpTypeFunction
:
4909 case SpvOpTypeEvent
:
4910 case SpvOpTypeDeviceEvent
:
4911 case SpvOpTypeReserveId
:
4912 case SpvOpTypeQueue
:
4914 vtn_handle_type(b
, opcode
, w
, count
);
4917 case SpvOpConstantTrue
:
4918 case SpvOpConstantFalse
:
4920 case SpvOpConstantComposite
:
4921 case SpvOpConstantNull
:
4922 case SpvOpSpecConstantTrue
:
4923 case SpvOpSpecConstantFalse
:
4924 case SpvOpSpecConstant
:
4925 case SpvOpSpecConstantComposite
:
4926 case SpvOpSpecConstantOp
:
4927 vtn_handle_constant(b
, opcode
, w
, count
);
4932 case SpvOpConstantSampler
:
4933 vtn_handle_variables(b
, opcode
, w
, count
);
4936 case SpvOpExtInst
: {
4937 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
4938 /* NonSemantic extended instructions are acceptable in preamble, others
4939 * will indicate the end of preamble.
4941 return val
->ext_handler
== vtn_handle_non_semantic_instruction
;
4945 return false; /* End of preamble */
4951 static struct vtn_ssa_value
*
4952 vtn_nir_select(struct vtn_builder
*b
, struct vtn_ssa_value
*src0
,
4953 struct vtn_ssa_value
*src1
, struct vtn_ssa_value
*src2
)
4955 struct vtn_ssa_value
*dest
= rzalloc(b
, struct vtn_ssa_value
);
4956 dest
->type
= src1
->type
;
4958 if (glsl_type_is_vector_or_scalar(src1
->type
)) {
4959 dest
->def
= nir_bcsel(&b
->nb
, src0
->def
, src1
->def
, src2
->def
);
4961 unsigned elems
= glsl_get_length(src1
->type
);
4963 dest
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
4964 for (unsigned i
= 0; i
< elems
; i
++) {
4965 dest
->elems
[i
] = vtn_nir_select(b
, src0
,
4966 src1
->elems
[i
], src2
->elems
[i
]);
4974 vtn_handle_select(struct vtn_builder
*b
, SpvOp opcode
,
4975 const uint32_t *w
, unsigned count
)
4977 /* Handle OpSelect up-front here because it needs to be able to handle
4978 * pointers and not just regular vectors and scalars.
4980 struct vtn_value
*res_val
= vtn_untyped_value(b
, w
[2]);
4981 struct vtn_value
*cond_val
= vtn_untyped_value(b
, w
[3]);
4982 struct vtn_value
*obj1_val
= vtn_untyped_value(b
, w
[4]);
4983 struct vtn_value
*obj2_val
= vtn_untyped_value(b
, w
[5]);
4985 vtn_fail_if(obj1_val
->type
!= res_val
->type
||
4986 obj2_val
->type
!= res_val
->type
,
4987 "Object types must match the result type in OpSelect");
4989 vtn_fail_if((cond_val
->type
->base_type
!= vtn_base_type_scalar
&&
4990 cond_val
->type
->base_type
!= vtn_base_type_vector
) ||
4991 !glsl_type_is_boolean(cond_val
->type
->type
),
4992 "OpSelect must have either a vector of booleans or "
4993 "a boolean as Condition type");
4995 vtn_fail_if(cond_val
->type
->base_type
== vtn_base_type_vector
&&
4996 (res_val
->type
->base_type
!= vtn_base_type_vector
||
4997 res_val
->type
->length
!= cond_val
->type
->length
),
4998 "When Condition type in OpSelect is a vector, the Result "
4999 "type must be a vector of the same length");
5001 switch (res_val
->type
->base_type
) {
5002 case vtn_base_type_scalar
:
5003 case vtn_base_type_vector
:
5004 case vtn_base_type_matrix
:
5005 case vtn_base_type_array
:
5006 case vtn_base_type_struct
:
5009 case vtn_base_type_pointer
:
5010 /* We need to have actual storage for pointer types. */
5011 vtn_fail_if(res_val
->type
->type
== NULL
,
5012 "Invalid pointer result type for OpSelect");
5015 vtn_fail("Result type of OpSelect must be a scalar, composite, or pointer");
5018 vtn_push_ssa_value(b
, w
[2],
5019 vtn_nir_select(b
, vtn_ssa_value(b
, w
[3]),
5020 vtn_ssa_value(b
, w
[4]),
5021 vtn_ssa_value(b
, w
[5])));
5025 vtn_handle_ptr(struct vtn_builder
*b
, SpvOp opcode
,
5026 const uint32_t *w
, unsigned count
)
5028 struct vtn_type
*type1
= vtn_get_value_type(b
, w
[3]);
5029 struct vtn_type
*type2
= vtn_get_value_type(b
, w
[4]);
5030 vtn_fail_if(type1
->base_type
!= vtn_base_type_pointer
||
5031 type2
->base_type
!= vtn_base_type_pointer
,
5032 "%s operands must have pointer types",
5033 spirv_op_to_string(opcode
));
5034 vtn_fail_if(type1
->storage_class
!= type2
->storage_class
,
5035 "%s operands must have the same storage class",
5036 spirv_op_to_string(opcode
));
5038 struct vtn_type
*vtn_type
= vtn_get_type(b
, w
[1]);
5039 const struct glsl_type
*type
= vtn_type
->type
;
5041 nir_address_format addr_format
= vtn_mode_to_address_format(
5042 b
, vtn_storage_class_to_mode(b
, type1
->storage_class
, NULL
, NULL
));
5047 case SpvOpPtrDiff
: {
5048 /* OpPtrDiff returns the difference in number of elements (not byte offset). */
5049 unsigned elem_size
, elem_align
;
5050 glsl_get_natural_size_align_bytes(type1
->deref
->type
,
5051 &elem_size
, &elem_align
);
5053 def
= nir_build_addr_isub(&b
->nb
,
5054 vtn_get_nir_ssa(b
, w
[3]),
5055 vtn_get_nir_ssa(b
, w
[4]),
5057 def
= nir_idiv(&b
->nb
, def
, nir_imm_intN_t(&b
->nb
, elem_size
, def
->bit_size
));
5058 def
= nir_i2i(&b
->nb
, def
, glsl_get_bit_size(type
));
5063 case SpvOpPtrNotEqual
: {
5064 def
= nir_build_addr_ieq(&b
->nb
,
5065 vtn_get_nir_ssa(b
, w
[3]),
5066 vtn_get_nir_ssa(b
, w
[4]),
5068 if (opcode
== SpvOpPtrNotEqual
)
5069 def
= nir_inot(&b
->nb
, def
);
5074 unreachable("Invalid ptr operation");
5077 vtn_push_nir_ssa(b
, w
[2], def
);
5081 vtn_handle_body_instruction(struct vtn_builder
*b
, SpvOp opcode
,
5082 const uint32_t *w
, unsigned count
)
5088 case SpvOpLoopMerge
:
5089 case SpvOpSelectionMerge
:
5090 /* This is handled by cfg pre-pass and walk_blocks */
5094 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
5095 val
->type
= vtn_get_type(b
, w
[1]);
5100 vtn_handle_extension(b
, opcode
, w
, count
);
5106 case SpvOpCopyMemory
:
5107 case SpvOpCopyMemorySized
:
5108 case SpvOpAccessChain
:
5109 case SpvOpPtrAccessChain
:
5110 case SpvOpInBoundsAccessChain
:
5111 case SpvOpInBoundsPtrAccessChain
:
5112 case SpvOpArrayLength
:
5113 case SpvOpConvertPtrToU
:
5114 case SpvOpConvertUToPtr
:
5115 vtn_handle_variables(b
, opcode
, w
, count
);
5118 case SpvOpFunctionCall
:
5119 vtn_handle_function_call(b
, opcode
, w
, count
);
5122 case SpvOpSampledImage
:
5124 case SpvOpImageSampleImplicitLod
:
5125 case SpvOpImageSampleExplicitLod
:
5126 case SpvOpImageSampleDrefImplicitLod
:
5127 case SpvOpImageSampleDrefExplicitLod
:
5128 case SpvOpImageSampleProjImplicitLod
:
5129 case SpvOpImageSampleProjExplicitLod
:
5130 case SpvOpImageSampleProjDrefImplicitLod
:
5131 case SpvOpImageSampleProjDrefExplicitLod
:
5132 case SpvOpImageFetch
:
5133 case SpvOpImageGather
:
5134 case SpvOpImageDrefGather
:
5135 case SpvOpImageQueryLod
:
5136 case SpvOpImageQueryLevels
:
5137 case SpvOpImageQuerySamples
:
5138 vtn_handle_texture(b
, opcode
, w
, count
);
5141 case SpvOpImageRead
:
5142 case SpvOpImageWrite
:
5143 case SpvOpImageTexelPointer
:
5144 case SpvOpImageQueryFormat
:
5145 case SpvOpImageQueryOrder
:
5146 vtn_handle_image(b
, opcode
, w
, count
);
5149 case SpvOpImageQuerySizeLod
:
5150 case SpvOpImageQuerySize
: {
5151 struct vtn_type
*image_type
= vtn_get_value_type(b
, w
[3]);
5152 vtn_assert(image_type
->base_type
== vtn_base_type_image
);
5153 if (glsl_type_is_image(image_type
->glsl_image
)) {
5154 vtn_handle_image(b
, opcode
, w
, count
);
5156 vtn_assert(glsl_type_is_sampler(image_type
->glsl_image
));
5157 vtn_handle_texture(b
, opcode
, w
, count
);
5162 case SpvOpFragmentMaskFetchAMD
:
5163 case SpvOpFragmentFetchAMD
:
5164 vtn_handle_texture(b
, opcode
, w
, count
);
5167 case SpvOpAtomicLoad
:
5168 case SpvOpAtomicExchange
:
5169 case SpvOpAtomicCompareExchange
:
5170 case SpvOpAtomicCompareExchangeWeak
:
5171 case SpvOpAtomicIIncrement
:
5172 case SpvOpAtomicIDecrement
:
5173 case SpvOpAtomicIAdd
:
5174 case SpvOpAtomicISub
:
5175 case SpvOpAtomicSMin
:
5176 case SpvOpAtomicUMin
:
5177 case SpvOpAtomicSMax
:
5178 case SpvOpAtomicUMax
:
5179 case SpvOpAtomicAnd
:
5181 case SpvOpAtomicXor
:
5182 case SpvOpAtomicFAddEXT
: {
5183 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[3]);
5184 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
5185 vtn_handle_image(b
, opcode
, w
, count
);
5187 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
5188 vtn_handle_atomics(b
, opcode
, w
, count
);
5193 case SpvOpAtomicStore
: {
5194 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[1]);
5195 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
5196 vtn_handle_image(b
, opcode
, w
, count
);
5198 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
5199 vtn_handle_atomics(b
, opcode
, w
, count
);
5205 vtn_handle_select(b
, opcode
, w
, count
);
5213 case SpvOpConvertFToU
:
5214 case SpvOpConvertFToS
:
5215 case SpvOpConvertSToF
:
5216 case SpvOpConvertUToF
:
5220 case SpvOpQuantizeToF16
:
5221 case SpvOpPtrCastToGeneric
:
5222 case SpvOpGenericCastToPtr
:
5227 case SpvOpSignBitSet
:
5228 case SpvOpLessOrGreater
:
5230 case SpvOpUnordered
:
5245 case SpvOpVectorTimesScalar
:
5247 case SpvOpIAddCarry
:
5248 case SpvOpISubBorrow
:
5249 case SpvOpUMulExtended
:
5250 case SpvOpSMulExtended
:
5251 case SpvOpShiftRightLogical
:
5252 case SpvOpShiftRightArithmetic
:
5253 case SpvOpShiftLeftLogical
:
5254 case SpvOpLogicalEqual
:
5255 case SpvOpLogicalNotEqual
:
5256 case SpvOpLogicalOr
:
5257 case SpvOpLogicalAnd
:
5258 case SpvOpLogicalNot
:
5259 case SpvOpBitwiseOr
:
5260 case SpvOpBitwiseXor
:
5261 case SpvOpBitwiseAnd
:
5263 case SpvOpFOrdEqual
:
5264 case SpvOpFUnordEqual
:
5265 case SpvOpINotEqual
:
5266 case SpvOpFOrdNotEqual
:
5267 case SpvOpFUnordNotEqual
:
5268 case SpvOpULessThan
:
5269 case SpvOpSLessThan
:
5270 case SpvOpFOrdLessThan
:
5271 case SpvOpFUnordLessThan
:
5272 case SpvOpUGreaterThan
:
5273 case SpvOpSGreaterThan
:
5274 case SpvOpFOrdGreaterThan
:
5275 case SpvOpFUnordGreaterThan
:
5276 case SpvOpULessThanEqual
:
5277 case SpvOpSLessThanEqual
:
5278 case SpvOpFOrdLessThanEqual
:
5279 case SpvOpFUnordLessThanEqual
:
5280 case SpvOpUGreaterThanEqual
:
5281 case SpvOpSGreaterThanEqual
:
5282 case SpvOpFOrdGreaterThanEqual
:
5283 case SpvOpFUnordGreaterThanEqual
:
5289 case SpvOpFwidthFine
:
5290 case SpvOpDPdxCoarse
:
5291 case SpvOpDPdyCoarse
:
5292 case SpvOpFwidthCoarse
:
5293 case SpvOpBitFieldInsert
:
5294 case SpvOpBitFieldSExtract
:
5295 case SpvOpBitFieldUExtract
:
5296 case SpvOpBitReverse
:
5298 case SpvOpTranspose
:
5299 case SpvOpOuterProduct
:
5300 case SpvOpMatrixTimesScalar
:
5301 case SpvOpVectorTimesMatrix
:
5302 case SpvOpMatrixTimesVector
:
5303 case SpvOpMatrixTimesMatrix
:
5304 case SpvOpUCountLeadingZerosINTEL
:
5305 case SpvOpUCountTrailingZerosINTEL
:
5306 case SpvOpAbsISubINTEL
:
5307 case SpvOpAbsUSubINTEL
:
5308 case SpvOpIAddSatINTEL
:
5309 case SpvOpUAddSatINTEL
:
5310 case SpvOpIAverageINTEL
:
5311 case SpvOpUAverageINTEL
:
5312 case SpvOpIAverageRoundedINTEL
:
5313 case SpvOpUAverageRoundedINTEL
:
5314 case SpvOpISubSatINTEL
:
5315 case SpvOpUSubSatINTEL
:
5316 case SpvOpIMul32x16INTEL
:
5317 case SpvOpUMul32x16INTEL
:
5318 vtn_handle_alu(b
, opcode
, w
, count
);
5322 vtn_handle_bitcast(b
, w
, count
);
5325 case SpvOpVectorExtractDynamic
:
5326 case SpvOpVectorInsertDynamic
:
5327 case SpvOpVectorShuffle
:
5328 case SpvOpCompositeConstruct
:
5329 case SpvOpCompositeExtract
:
5330 case SpvOpCompositeInsert
:
5331 case SpvOpCopyLogical
:
5332 case SpvOpCopyObject
:
5333 vtn_handle_composite(b
, opcode
, w
, count
);
5336 case SpvOpEmitVertex
:
5337 case SpvOpEndPrimitive
:
5338 case SpvOpEmitStreamVertex
:
5339 case SpvOpEndStreamPrimitive
:
5340 case SpvOpControlBarrier
:
5341 case SpvOpMemoryBarrier
:
5342 vtn_handle_barrier(b
, opcode
, w
, count
);
5345 case SpvOpGroupNonUniformElect
:
5346 case SpvOpGroupNonUniformAll
:
5347 case SpvOpGroupNonUniformAny
:
5348 case SpvOpGroupNonUniformAllEqual
:
5349 case SpvOpGroupNonUniformBroadcast
:
5350 case SpvOpGroupNonUniformBroadcastFirst
:
5351 case SpvOpGroupNonUniformBallot
:
5352 case SpvOpGroupNonUniformInverseBallot
:
5353 case SpvOpGroupNonUniformBallotBitExtract
:
5354 case SpvOpGroupNonUniformBallotBitCount
:
5355 case SpvOpGroupNonUniformBallotFindLSB
:
5356 case SpvOpGroupNonUniformBallotFindMSB
:
5357 case SpvOpGroupNonUniformShuffle
:
5358 case SpvOpGroupNonUniformShuffleXor
:
5359 case SpvOpGroupNonUniformShuffleUp
:
5360 case SpvOpGroupNonUniformShuffleDown
:
5361 case SpvOpGroupNonUniformIAdd
:
5362 case SpvOpGroupNonUniformFAdd
:
5363 case SpvOpGroupNonUniformIMul
:
5364 case SpvOpGroupNonUniformFMul
:
5365 case SpvOpGroupNonUniformSMin
:
5366 case SpvOpGroupNonUniformUMin
:
5367 case SpvOpGroupNonUniformFMin
:
5368 case SpvOpGroupNonUniformSMax
:
5369 case SpvOpGroupNonUniformUMax
:
5370 case SpvOpGroupNonUniformFMax
:
5371 case SpvOpGroupNonUniformBitwiseAnd
:
5372 case SpvOpGroupNonUniformBitwiseOr
:
5373 case SpvOpGroupNonUniformBitwiseXor
:
5374 case SpvOpGroupNonUniformLogicalAnd
:
5375 case SpvOpGroupNonUniformLogicalOr
:
5376 case SpvOpGroupNonUniformLogicalXor
:
5377 case SpvOpGroupNonUniformQuadBroadcast
:
5378 case SpvOpGroupNonUniformQuadSwap
:
5381 case SpvOpGroupBroadcast
:
5382 case SpvOpGroupIAdd
:
5383 case SpvOpGroupFAdd
:
5384 case SpvOpGroupFMin
:
5385 case SpvOpGroupUMin
:
5386 case SpvOpGroupSMin
:
5387 case SpvOpGroupFMax
:
5388 case SpvOpGroupUMax
:
5389 case SpvOpGroupSMax
:
5390 case SpvOpSubgroupBallotKHR
:
5391 case SpvOpSubgroupFirstInvocationKHR
:
5392 case SpvOpSubgroupReadInvocationKHR
:
5393 case SpvOpSubgroupAllKHR
:
5394 case SpvOpSubgroupAnyKHR
:
5395 case SpvOpSubgroupAllEqualKHR
:
5396 case SpvOpGroupIAddNonUniformAMD
:
5397 case SpvOpGroupFAddNonUniformAMD
:
5398 case SpvOpGroupFMinNonUniformAMD
:
5399 case SpvOpGroupUMinNonUniformAMD
:
5400 case SpvOpGroupSMinNonUniformAMD
:
5401 case SpvOpGroupFMaxNonUniformAMD
:
5402 case SpvOpGroupUMaxNonUniformAMD
:
5403 case SpvOpGroupSMaxNonUniformAMD
:
5404 vtn_handle_subgroup(b
, opcode
, w
, count
);
5409 case SpvOpPtrNotEqual
:
5410 vtn_handle_ptr(b
, opcode
, w
, count
);
5413 case SpvOpBeginInvocationInterlockEXT
:
5414 vtn_emit_barrier(b
, nir_intrinsic_begin_invocation_interlock
);
5417 case SpvOpEndInvocationInterlockEXT
:
5418 vtn_emit_barrier(b
, nir_intrinsic_end_invocation_interlock
);
5421 case SpvOpDemoteToHelperInvocationEXT
: {
5422 nir_intrinsic_instr
*intrin
=
5423 nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_demote
);
5424 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
5428 case SpvOpIsHelperInvocationEXT
: {
5429 nir_intrinsic_instr
*intrin
=
5430 nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_is_helper_invocation
);
5431 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
, 1, 1, NULL
);
5432 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
5434 vtn_push_nir_ssa(b
, w
[2], &intrin
->dest
.ssa
);
5438 case SpvOpReadClockKHR
: {
5439 SpvScope scope
= vtn_constant_uint(b
, w
[3]);
5440 nir_scope nir_scope
;
5443 case SpvScopeDevice
:
5444 nir_scope
= NIR_SCOPE_DEVICE
;
5446 case SpvScopeSubgroup
:
5447 nir_scope
= NIR_SCOPE_SUBGROUP
;
5450 vtn_fail("invalid read clock scope");
5453 /* Operation supports two result types: uvec2 and uint64_t. The NIR
5454 * intrinsic gives uvec2, so pack the result for the other case.
5456 nir_intrinsic_instr
*intrin
=
5457 nir_intrinsic_instr_create(b
->nb
.shader
, nir_intrinsic_shader_clock
);
5458 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
, 2, 32, NULL
);
5459 nir_intrinsic_set_memory_scope(intrin
, nir_scope
);
5460 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
5462 struct vtn_type
*type
= vtn_get_type(b
, w
[1]);
5463 const struct glsl_type
*dest_type
= type
->type
;
5464 nir_ssa_def
*result
;
5466 if (glsl_type_is_vector(dest_type
)) {
5467 assert(dest_type
== glsl_vector_type(GLSL_TYPE_UINT
, 2));
5468 result
= &intrin
->dest
.ssa
;
5470 assert(glsl_type_is_scalar(dest_type
));
5471 assert(glsl_get_base_type(dest_type
) == GLSL_TYPE_UINT64
);
5472 result
= nir_pack_64_2x32(&b
->nb
, &intrin
->dest
.ssa
);
5475 vtn_push_nir_ssa(b
, w
[2], result
);
5479 case SpvOpLifetimeStart
:
5480 case SpvOpLifetimeStop
:
5484 vtn_fail_with_opcode("Unhandled opcode", opcode
);
5491 vtn_create_builder(const uint32_t *words
, size_t word_count
,
5492 gl_shader_stage stage
, const char *entry_point_name
,
5493 const struct spirv_to_nir_options
*options
)
5495 /* Initialize the vtn_builder object */
5496 struct vtn_builder
*b
= rzalloc(NULL
, struct vtn_builder
);
5497 struct spirv_to_nir_options
*dup_options
=
5498 ralloc(b
, struct spirv_to_nir_options
);
5499 *dup_options
= *options
;
5502 b
->spirv_word_count
= word_count
;
5506 list_inithead(&b
->functions
);
5507 b
->entry_point_stage
= stage
;
5508 b
->entry_point_name
= entry_point_name
;
5509 b
->options
= dup_options
;
5512 * Handle the SPIR-V header (first 5 dwords).
5513 * Can't use vtx_assert() as the setjmp(3) target isn't initialized yet.
5515 if (word_count
<= 5)
5518 if (words
[0] != SpvMagicNumber
) {
5519 vtn_err("words[0] was 0x%x, want 0x%x", words
[0], SpvMagicNumber
);
5522 if (words
[1] < 0x10000) {
5523 vtn_err("words[1] was 0x%x, want >= 0x10000", words
[1]);
5527 uint16_t generator_id
= words
[2] >> 16;
5528 uint16_t generator_version
= words
[2];
5530 /* In GLSLang commit 8297936dd6eb3, their handling of barrier() was fixed
5531 * to provide correct memory semantics on compute shader barrier()
5532 * commands. Prior to that, we need to fix them up ourselves. This
5533 * GLSLang fix caused them to bump to generator version 3.
5535 b
->wa_glslang_cs_barrier
= (generator_id
== 8 && generator_version
< 3);
5537 /* words[2] == generator magic */
5538 unsigned value_id_bound
= words
[3];
5539 if (words
[4] != 0) {
5540 vtn_err("words[4] was %u, want 0", words
[4]);
5544 b
->value_id_bound
= value_id_bound
;
5545 b
->values
= rzalloc_array(b
, struct vtn_value
, value_id_bound
);
5553 static nir_function
*
5554 vtn_emit_kernel_entry_point_wrapper(struct vtn_builder
*b
,
5555 nir_function
*entry_point
)
5557 vtn_assert(entry_point
== b
->entry_point
->func
->impl
->function
);
5558 vtn_fail_if(!entry_point
->name
, "entry points are required to have a name");
5559 const char *func_name
=
5560 ralloc_asprintf(b
->shader
, "__wrapped_%s", entry_point
->name
);
5562 /* we shouldn't have any inputs yet */
5563 vtn_assert(!entry_point
->shader
->num_inputs
);
5564 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_KERNEL
);
5566 nir_function
*main_entry_point
= nir_function_create(b
->shader
, func_name
);
5567 main_entry_point
->impl
= nir_function_impl_create(main_entry_point
);
5568 nir_builder_init(&b
->nb
, main_entry_point
->impl
);
5569 b
->nb
.cursor
= nir_after_cf_list(&main_entry_point
->impl
->body
);
5570 b
->func_param_idx
= 0;
5572 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, entry_point
);
5574 for (unsigned i
= 0; i
< entry_point
->num_params
; ++i
) {
5575 struct vtn_type
*param_type
= b
->entry_point
->func
->type
->params
[i
];
5577 /* consider all pointers to function memory to be parameters passed
5580 bool is_by_val
= param_type
->base_type
== vtn_base_type_pointer
&&
5581 param_type
->storage_class
== SpvStorageClassFunction
;
5583 /* input variable */
5584 nir_variable
*in_var
= rzalloc(b
->nb
.shader
, nir_variable
);
5585 in_var
->data
.mode
= nir_var_uniform
;
5586 in_var
->data
.read_only
= true;
5587 in_var
->data
.location
= i
;
5588 if (param_type
->base_type
== vtn_base_type_image
) {
5589 in_var
->data
.access
= 0;
5590 if (param_type
->access_qualifier
& SpvAccessQualifierReadOnly
)
5591 in_var
->data
.access
|= ACCESS_NON_WRITEABLE
;
5592 if (param_type
->access_qualifier
& SpvAccessQualifierWriteOnly
)
5593 in_var
->data
.access
|= ACCESS_NON_READABLE
;
5597 in_var
->type
= param_type
->deref
->type
;
5598 else if (param_type
->base_type
== vtn_base_type_image
)
5599 in_var
->type
= param_type
->glsl_image
;
5600 else if (param_type
->base_type
== vtn_base_type_sampler
)
5601 in_var
->type
= glsl_bare_sampler_type();
5603 in_var
->type
= param_type
->type
;
5605 nir_shader_add_variable(b
->nb
.shader
, in_var
);
5606 b
->nb
.shader
->num_inputs
++;
5608 /* we have to copy the entire variable into function memory */
5610 nir_variable
*copy_var
=
5611 nir_local_variable_create(main_entry_point
->impl
, in_var
->type
,
5613 nir_copy_var(&b
->nb
, copy_var
, in_var
);
5615 nir_src_for_ssa(&nir_build_deref_var(&b
->nb
, copy_var
)->dest
.ssa
);
5616 } else if (param_type
->base_type
== vtn_base_type_image
||
5617 param_type
->base_type
== vtn_base_type_sampler
) {
5618 /* Don't load the var, just pass a deref of it */
5619 call
->params
[i
] = nir_src_for_ssa(&nir_build_deref_var(&b
->nb
, in_var
)->dest
.ssa
);
5621 call
->params
[i
] = nir_src_for_ssa(nir_load_var(&b
->nb
, in_var
));
5625 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
5627 return main_entry_point
;
5631 spirv_to_nir(const uint32_t *words
, size_t word_count
,
5632 struct nir_spirv_specialization
*spec
, unsigned num_spec
,
5633 gl_shader_stage stage
, const char *entry_point_name
,
5634 const struct spirv_to_nir_options
*options
,
5635 const nir_shader_compiler_options
*nir_options
)
5638 const uint32_t *word_end
= words
+ word_count
;
5640 struct vtn_builder
*b
= vtn_create_builder(words
, word_count
,
5641 stage
, entry_point_name
,
5647 /* See also _vtn_fail() */
5648 if (setjmp(b
->fail_jump
)) {
5653 /* Skip the SPIR-V header, handled at vtn_create_builder */
5656 b
->shader
= nir_shader_create(b
, stage
, nir_options
, NULL
);
5658 /* Handle all the preamble instructions */
5659 words
= vtn_foreach_instruction(b
, words
, word_end
,
5660 vtn_handle_preamble_instruction
);
5662 if (b
->entry_point
== NULL
) {
5663 vtn_fail("Entry point not found");
5668 /* Ensure a sane address mode is being used for function temps */
5669 assert(nir_address_format_bit_size(b
->options
->temp_addr_format
) == nir_get_ptr_bitsize(b
->shader
));
5670 assert(nir_address_format_num_components(b
->options
->temp_addr_format
) == 1);
5672 /* Set shader info defaults */
5673 if (stage
== MESA_SHADER_GEOMETRY
)
5674 b
->shader
->info
.gs
.invocations
= 1;
5676 /* Parse execution modes. */
5677 vtn_foreach_execution_mode(b
, b
->entry_point
,
5678 vtn_handle_execution_mode
, NULL
);
5680 b
->specializations
= spec
;
5681 b
->num_specializations
= num_spec
;
5683 /* Handle all variable, type, and constant instructions */
5684 words
= vtn_foreach_instruction(b
, words
, word_end
,
5685 vtn_handle_variable_or_type_instruction
);
5687 /* Parse execution modes that depend on IDs. Must happen after we have
5690 vtn_foreach_execution_mode(b
, b
->entry_point
,
5691 vtn_handle_execution_mode_id
, NULL
);
5693 if (b
->workgroup_size_builtin
) {
5694 vtn_assert(b
->workgroup_size_builtin
->type
->type
==
5695 glsl_vector_type(GLSL_TYPE_UINT
, 3));
5697 nir_const_value
*const_size
=
5698 b
->workgroup_size_builtin
->constant
->values
;
5700 b
->shader
->info
.cs
.local_size
[0] = const_size
[0].u32
;
5701 b
->shader
->info
.cs
.local_size
[1] = const_size
[1].u32
;
5702 b
->shader
->info
.cs
.local_size
[2] = const_size
[2].u32
;
5705 /* Set types on all vtn_values */
5706 vtn_foreach_instruction(b
, words
, word_end
, vtn_set_instruction_result_type
);
5708 vtn_build_cfg(b
, words
, word_end
);
5710 assert(b
->entry_point
->value_type
== vtn_value_type_function
);
5711 b
->entry_point
->func
->referenced
= true;
5716 vtn_foreach_cf_node(node
, &b
->functions
) {
5717 struct vtn_function
*func
= vtn_cf_node_as_function(node
);
5718 if (func
->referenced
&& !func
->emitted
) {
5719 b
->const_table
= _mesa_pointer_hash_table_create(b
);
5721 vtn_function_emit(b
, func
, vtn_handle_body_instruction
);
5727 vtn_assert(b
->entry_point
->value_type
== vtn_value_type_function
);
5728 nir_function
*entry_point
= b
->entry_point
->func
->impl
->function
;
5729 vtn_assert(entry_point
);
5731 /* post process entry_points with input params */
5732 if (entry_point
->num_params
&& b
->shader
->info
.stage
== MESA_SHADER_KERNEL
)
5733 entry_point
= vtn_emit_kernel_entry_point_wrapper(b
, entry_point
);
5735 /* structurize the CFG */
5736 nir_lower_goto_ifs(b
->shader
);
5738 entry_point
->is_entrypoint
= true;
5740 /* When multiple shader stages exist in the same SPIR-V module, we
5741 * generate input and output variables for every stage, in the same
5742 * NIR program. These dead variables can be invalid NIR. For example,
5743 * TCS outputs must be per-vertex arrays (or decorated 'patch'), while
5744 * VS output variables wouldn't be.
5746 * To ensure we have valid NIR, we eliminate any dead inputs and outputs
5747 * right away. In order to do so, we must lower any constant initializers
5748 * on outputs so nir_remove_dead_variables sees that they're written to.
5750 nir_lower_variable_initializers(b
->shader
, nir_var_shader_out
);
5751 nir_remove_dead_variables(b
->shader
,
5752 nir_var_shader_in
| nir_var_shader_out
, NULL
);
5754 /* We sometimes generate bogus derefs that, while never used, give the
5755 * validator a bit of heartburn. Run dead code to get rid of them.
5757 nir_opt_dce(b
->shader
);
5759 /* Unparent the shader from the vtn_builder before we delete the builder */
5760 ralloc_steal(NULL
, b
->shader
);
5762 nir_shader
*shader
= b
->shader
;