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_packed_decoration_cb(struct vtn_builder
*b
,
1136 struct vtn_value
*val
, int member
,
1137 const struct vtn_decoration
*dec
, void *void_ctx
)
1139 vtn_assert(val
->type
->base_type
== vtn_base_type_struct
);
1140 if (dec
->decoration
== SpvDecorationCPacked
) {
1141 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
) {
1142 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1143 spirv_decoration_to_string(dec
->decoration
));
1145 val
->type
->packed
= true;
1150 struct_block_decoration_cb(struct vtn_builder
*b
,
1151 struct vtn_value
*val
, int member
,
1152 const struct vtn_decoration
*dec
, void *ctx
)
1157 struct vtn_type
*type
= val
->type
;
1158 if (dec
->decoration
== SpvDecorationBlock
)
1160 else if (dec
->decoration
== SpvDecorationBufferBlock
)
1161 type
->buffer_block
= true;
1165 type_decoration_cb(struct vtn_builder
*b
,
1166 struct vtn_value
*val
, int member
,
1167 const struct vtn_decoration
*dec
, UNUSED
void *ctx
)
1169 struct vtn_type
*type
= val
->type
;
1172 /* This should have been handled by OpTypeStruct */
1173 assert(val
->type
->base_type
== vtn_base_type_struct
);
1174 assert(member
>= 0 && member
< val
->type
->length
);
1178 switch (dec
->decoration
) {
1179 case SpvDecorationArrayStride
:
1180 vtn_assert(type
->base_type
== vtn_base_type_array
||
1181 type
->base_type
== vtn_base_type_pointer
);
1183 case SpvDecorationBlock
:
1184 vtn_assert(type
->base_type
== vtn_base_type_struct
);
1185 vtn_assert(type
->block
);
1187 case SpvDecorationBufferBlock
:
1188 vtn_assert(type
->base_type
== vtn_base_type_struct
);
1189 vtn_assert(type
->buffer_block
);
1191 case SpvDecorationGLSLShared
:
1192 case SpvDecorationGLSLPacked
:
1193 /* Ignore these, since we get explicit offsets anyways */
1196 case SpvDecorationRowMajor
:
1197 case SpvDecorationColMajor
:
1198 case SpvDecorationMatrixStride
:
1199 case SpvDecorationBuiltIn
:
1200 case SpvDecorationNoPerspective
:
1201 case SpvDecorationFlat
:
1202 case SpvDecorationPatch
:
1203 case SpvDecorationCentroid
:
1204 case SpvDecorationSample
:
1205 case SpvDecorationExplicitInterpAMD
:
1206 case SpvDecorationVolatile
:
1207 case SpvDecorationCoherent
:
1208 case SpvDecorationNonWritable
:
1209 case SpvDecorationNonReadable
:
1210 case SpvDecorationUniform
:
1211 case SpvDecorationUniformId
:
1212 case SpvDecorationLocation
:
1213 case SpvDecorationComponent
:
1214 case SpvDecorationOffset
:
1215 case SpvDecorationXfbBuffer
:
1216 case SpvDecorationXfbStride
:
1217 case SpvDecorationUserSemantic
:
1218 vtn_warn("Decoration only allowed for struct members: %s",
1219 spirv_decoration_to_string(dec
->decoration
));
1222 case SpvDecorationStream
:
1223 /* We don't need to do anything here, as stream is filled up when
1224 * aplying the decoration to a variable, just check that if it is not a
1225 * struct member, it should be a struct.
1227 vtn_assert(type
->base_type
== vtn_base_type_struct
);
1230 case SpvDecorationRelaxedPrecision
:
1231 case SpvDecorationSpecId
:
1232 case SpvDecorationInvariant
:
1233 case SpvDecorationRestrict
:
1234 case SpvDecorationAliased
:
1235 case SpvDecorationConstant
:
1236 case SpvDecorationIndex
:
1237 case SpvDecorationBinding
:
1238 case SpvDecorationDescriptorSet
:
1239 case SpvDecorationLinkageAttributes
:
1240 case SpvDecorationNoContraction
:
1241 case SpvDecorationInputAttachmentIndex
:
1242 vtn_warn("Decoration not allowed on types: %s",
1243 spirv_decoration_to_string(dec
->decoration
));
1246 case SpvDecorationCPacked
:
1247 /* Handled when parsing a struct type, nothing to do here. */
1250 case SpvDecorationSaturatedConversion
:
1251 case SpvDecorationFuncParamAttr
:
1252 case SpvDecorationFPRoundingMode
:
1253 case SpvDecorationFPFastMathMode
:
1254 case SpvDecorationAlignment
:
1255 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1256 spirv_decoration_to_string(dec
->decoration
));
1259 case SpvDecorationUserTypeGOOGLE
:
1260 /* User semantic decorations can safely be ignored by the driver. */
1264 vtn_fail_with_decoration("Unhandled decoration", dec
->decoration
);
1269 translate_image_format(struct vtn_builder
*b
, SpvImageFormat format
)
1272 case SpvImageFormatUnknown
: return PIPE_FORMAT_NONE
;
1273 case SpvImageFormatRgba32f
: return PIPE_FORMAT_R32G32B32A32_FLOAT
;
1274 case SpvImageFormatRgba16f
: return PIPE_FORMAT_R16G16B16A16_FLOAT
;
1275 case SpvImageFormatR32f
: return PIPE_FORMAT_R32_FLOAT
;
1276 case SpvImageFormatRgba8
: return PIPE_FORMAT_R8G8B8A8_UNORM
;
1277 case SpvImageFormatRgba8Snorm
: return PIPE_FORMAT_R8G8B8A8_SNORM
;
1278 case SpvImageFormatRg32f
: return PIPE_FORMAT_R32G32_FLOAT
;
1279 case SpvImageFormatRg16f
: return PIPE_FORMAT_R16G16_FLOAT
;
1280 case SpvImageFormatR11fG11fB10f
: return PIPE_FORMAT_R11G11B10_FLOAT
;
1281 case SpvImageFormatR16f
: return PIPE_FORMAT_R16_FLOAT
;
1282 case SpvImageFormatRgba16
: return PIPE_FORMAT_R16G16B16A16_UNORM
;
1283 case SpvImageFormatRgb10A2
: return PIPE_FORMAT_R10G10B10A2_UNORM
;
1284 case SpvImageFormatRg16
: return PIPE_FORMAT_R16G16_UNORM
;
1285 case SpvImageFormatRg8
: return PIPE_FORMAT_R8G8_UNORM
;
1286 case SpvImageFormatR16
: return PIPE_FORMAT_R16_UNORM
;
1287 case SpvImageFormatR8
: return PIPE_FORMAT_R8_UNORM
;
1288 case SpvImageFormatRgba16Snorm
: return PIPE_FORMAT_R16G16B16A16_SNORM
;
1289 case SpvImageFormatRg16Snorm
: return PIPE_FORMAT_R16G16_SNORM
;
1290 case SpvImageFormatRg8Snorm
: return PIPE_FORMAT_R8G8_SNORM
;
1291 case SpvImageFormatR16Snorm
: return PIPE_FORMAT_R16_SNORM
;
1292 case SpvImageFormatR8Snorm
: return PIPE_FORMAT_R8_SNORM
;
1293 case SpvImageFormatRgba32i
: return PIPE_FORMAT_R32G32B32A32_SINT
;
1294 case SpvImageFormatRgba16i
: return PIPE_FORMAT_R16G16B16A16_SINT
;
1295 case SpvImageFormatRgba8i
: return PIPE_FORMAT_R8G8B8A8_SINT
;
1296 case SpvImageFormatR32i
: return PIPE_FORMAT_R32_SINT
;
1297 case SpvImageFormatRg32i
: return PIPE_FORMAT_R32G32_SINT
;
1298 case SpvImageFormatRg16i
: return PIPE_FORMAT_R16G16_SINT
;
1299 case SpvImageFormatRg8i
: return PIPE_FORMAT_R8G8_SINT
;
1300 case SpvImageFormatR16i
: return PIPE_FORMAT_R16_SINT
;
1301 case SpvImageFormatR8i
: return PIPE_FORMAT_R8_SINT
;
1302 case SpvImageFormatRgba32ui
: return PIPE_FORMAT_R32G32B32A32_UINT
;
1303 case SpvImageFormatRgba16ui
: return PIPE_FORMAT_R16G16B16A16_UINT
;
1304 case SpvImageFormatRgba8ui
: return PIPE_FORMAT_R8G8B8A8_UINT
;
1305 case SpvImageFormatR32ui
: return PIPE_FORMAT_R32_UINT
;
1306 case SpvImageFormatRgb10a2ui
: return PIPE_FORMAT_R10G10B10A2_UINT
;
1307 case SpvImageFormatRg32ui
: return PIPE_FORMAT_R32G32_UINT
;
1308 case SpvImageFormatRg16ui
: return PIPE_FORMAT_R16G16_UINT
;
1309 case SpvImageFormatRg8ui
: return PIPE_FORMAT_R8G8_UINT
;
1310 case SpvImageFormatR16ui
: return PIPE_FORMAT_R16_UINT
;
1311 case SpvImageFormatR8ui
: return PIPE_FORMAT_R8_UINT
;
1313 vtn_fail("Invalid image format: %s (%u)",
1314 spirv_imageformat_to_string(format
), format
);
1319 vtn_handle_type(struct vtn_builder
*b
, SpvOp opcode
,
1320 const uint32_t *w
, unsigned count
)
1322 struct vtn_value
*val
= NULL
;
1324 /* In order to properly handle forward declarations, we have to defer
1325 * allocation for pointer types.
1327 if (opcode
!= SpvOpTypePointer
&& opcode
!= SpvOpTypeForwardPointer
) {
1328 val
= vtn_push_value(b
, w
[1], vtn_value_type_type
);
1329 vtn_fail_if(val
->type
!= NULL
,
1330 "Only pointers can have forward declarations");
1331 val
->type
= rzalloc(b
, struct vtn_type
);
1332 val
->type
->id
= w
[1];
1337 val
->type
->base_type
= vtn_base_type_void
;
1338 val
->type
->type
= glsl_void_type();
1341 val
->type
->base_type
= vtn_base_type_scalar
;
1342 val
->type
->type
= glsl_bool_type();
1343 val
->type
->length
= 1;
1345 case SpvOpTypeInt
: {
1346 int bit_size
= w
[2];
1347 const bool signedness
= w
[3];
1348 val
->type
->base_type
= vtn_base_type_scalar
;
1351 val
->type
->type
= (signedness
? glsl_int64_t_type() : glsl_uint64_t_type());
1354 val
->type
->type
= (signedness
? glsl_int_type() : glsl_uint_type());
1357 val
->type
->type
= (signedness
? glsl_int16_t_type() : glsl_uint16_t_type());
1360 val
->type
->type
= (signedness
? glsl_int8_t_type() : glsl_uint8_t_type());
1363 vtn_fail("Invalid int bit size: %u", bit_size
);
1365 val
->type
->length
= 1;
1369 case SpvOpTypeFloat
: {
1370 int bit_size
= w
[2];
1371 val
->type
->base_type
= vtn_base_type_scalar
;
1374 val
->type
->type
= glsl_float16_t_type();
1377 val
->type
->type
= glsl_float_type();
1380 val
->type
->type
= glsl_double_type();
1383 vtn_fail("Invalid float bit size: %u", bit_size
);
1385 val
->type
->length
= 1;
1389 case SpvOpTypeVector
: {
1390 struct vtn_type
*base
= vtn_get_type(b
, w
[2]);
1391 unsigned elems
= w
[3];
1393 vtn_fail_if(base
->base_type
!= vtn_base_type_scalar
,
1394 "Base type for OpTypeVector must be a scalar");
1395 vtn_fail_if((elems
< 2 || elems
> 4) && (elems
!= 8) && (elems
!= 16),
1396 "Invalid component count for OpTypeVector");
1398 val
->type
->base_type
= vtn_base_type_vector
;
1399 val
->type
->type
= glsl_vector_type(glsl_get_base_type(base
->type
), elems
);
1400 val
->type
->length
= elems
;
1401 val
->type
->stride
= glsl_type_is_boolean(val
->type
->type
)
1402 ? 4 : glsl_get_bit_size(base
->type
) / 8;
1403 val
->type
->array_element
= base
;
1407 case SpvOpTypeMatrix
: {
1408 struct vtn_type
*base
= vtn_get_type(b
, w
[2]);
1409 unsigned columns
= w
[3];
1411 vtn_fail_if(base
->base_type
!= vtn_base_type_vector
,
1412 "Base type for OpTypeMatrix must be a vector");
1413 vtn_fail_if(columns
< 2 || columns
> 4,
1414 "Invalid column count for OpTypeMatrix");
1416 val
->type
->base_type
= vtn_base_type_matrix
;
1417 val
->type
->type
= glsl_matrix_type(glsl_get_base_type(base
->type
),
1418 glsl_get_vector_elements(base
->type
),
1420 vtn_fail_if(glsl_type_is_error(val
->type
->type
),
1421 "Unsupported base type for OpTypeMatrix");
1422 assert(!glsl_type_is_error(val
->type
->type
));
1423 val
->type
->length
= columns
;
1424 val
->type
->array_element
= base
;
1425 val
->type
->row_major
= false;
1426 val
->type
->stride
= 0;
1430 case SpvOpTypeRuntimeArray
:
1431 case SpvOpTypeArray
: {
1432 struct vtn_type
*array_element
= vtn_get_type(b
, w
[2]);
1434 if (opcode
== SpvOpTypeRuntimeArray
) {
1435 /* A length of 0 is used to denote unsized arrays */
1436 val
->type
->length
= 0;
1438 val
->type
->length
= vtn_constant_uint(b
, w
[3]);
1441 val
->type
->base_type
= vtn_base_type_array
;
1442 val
->type
->array_element
= array_element
;
1444 vtn_foreach_decoration(b
, val
, array_stride_decoration_cb
, NULL
);
1445 val
->type
->type
= glsl_array_type(array_element
->type
, val
->type
->length
,
1450 case SpvOpTypeStruct
: {
1451 unsigned num_fields
= count
- 2;
1452 val
->type
->base_type
= vtn_base_type_struct
;
1453 val
->type
->length
= num_fields
;
1454 val
->type
->members
= ralloc_array(b
, struct vtn_type
*, num_fields
);
1455 val
->type
->offsets
= ralloc_array(b
, unsigned, num_fields
);
1456 val
->type
->packed
= false;
1458 NIR_VLA(struct glsl_struct_field
, fields
, count
);
1459 for (unsigned i
= 0; i
< num_fields
; i
++) {
1460 val
->type
->members
[i
] = vtn_get_type(b
, w
[i
+ 2]);
1461 fields
[i
] = (struct glsl_struct_field
) {
1462 .type
= val
->type
->members
[i
]->type
,
1463 .name
= ralloc_asprintf(b
, "field%d", i
),
1469 vtn_foreach_decoration(b
, val
, struct_packed_decoration_cb
, NULL
);
1471 struct member_decoration_ctx ctx
= {
1472 .num_fields
= num_fields
,
1477 vtn_foreach_decoration(b
, val
, struct_member_decoration_cb
, &ctx
);
1478 vtn_foreach_decoration(b
, val
, struct_member_matrix_stride_cb
, &ctx
);
1480 vtn_foreach_decoration(b
, val
, struct_block_decoration_cb
, NULL
);
1482 const char *name
= val
->name
;
1484 if (val
->type
->block
|| val
->type
->buffer_block
) {
1485 /* Packing will be ignored since types coming from SPIR-V are
1486 * explicitly laid out.
1488 val
->type
->type
= glsl_interface_type(fields
, num_fields
,
1489 /* packing */ 0, false,
1490 name
? name
: "block");
1492 val
->type
->type
= glsl_struct_type(fields
, num_fields
,
1493 name
? name
: "struct",
1499 case SpvOpTypeFunction
: {
1500 val
->type
->base_type
= vtn_base_type_function
;
1501 val
->type
->type
= NULL
;
1503 val
->type
->return_type
= vtn_get_type(b
, w
[2]);
1505 const unsigned num_params
= count
- 3;
1506 val
->type
->length
= num_params
;
1507 val
->type
->params
= ralloc_array(b
, struct vtn_type
*, num_params
);
1508 for (unsigned i
= 0; i
< count
- 3; i
++) {
1509 val
->type
->params
[i
] = vtn_get_type(b
, w
[i
+ 3]);
1514 case SpvOpTypePointer
:
1515 case SpvOpTypeForwardPointer
: {
1516 /* We can't blindly push the value because it might be a forward
1519 val
= vtn_untyped_value(b
, w
[1]);
1521 SpvStorageClass storage_class
= w
[2];
1523 if (val
->value_type
== vtn_value_type_invalid
) {
1524 val
->value_type
= vtn_value_type_type
;
1525 val
->type
= rzalloc(b
, struct vtn_type
);
1526 val
->type
->id
= w
[1];
1527 val
->type
->base_type
= vtn_base_type_pointer
;
1528 val
->type
->storage_class
= storage_class
;
1530 /* These can actually be stored to nir_variables and used as SSA
1531 * values so they need a real glsl_type.
1533 enum vtn_variable_mode mode
= vtn_storage_class_to_mode(
1534 b
, storage_class
, NULL
, NULL
);
1535 val
->type
->type
= nir_address_format_to_glsl_type(
1536 vtn_mode_to_address_format(b
, mode
));
1538 vtn_fail_if(val
->type
->storage_class
!= storage_class
,
1539 "The storage classes of an OpTypePointer and any "
1540 "OpTypeForwardPointers that provide forward "
1541 "declarations of it must match.");
1544 if (opcode
== SpvOpTypePointer
) {
1545 vtn_fail_if(val
->type
->deref
!= NULL
,
1546 "While OpTypeForwardPointer can be used to provide a "
1547 "forward declaration of a pointer, OpTypePointer can "
1548 "only be used once for a given id.");
1550 val
->type
->deref
= vtn_get_type(b
, w
[3]);
1552 /* Only certain storage classes use ArrayStride. The others (in
1553 * particular Workgroup) are expected to be laid out by the driver.
1555 switch (storage_class
) {
1556 case SpvStorageClassUniform
:
1557 case SpvStorageClassPushConstant
:
1558 case SpvStorageClassStorageBuffer
:
1559 case SpvStorageClassPhysicalStorageBuffer
:
1560 vtn_foreach_decoration(b
, val
, array_stride_decoration_cb
, NULL
);
1563 /* Nothing to do. */
1570 case SpvOpTypeImage
: {
1571 val
->type
->base_type
= vtn_base_type_image
;
1573 /* Images are represented in NIR as a scalar SSA value that is the
1574 * result of a deref instruction. An OpLoad on an OpTypeImage pointer
1575 * from UniformConstant memory just takes the NIR deref from the pointer
1576 * and turns it into an SSA value.
1578 val
->type
->type
= nir_address_format_to_glsl_type(
1579 vtn_mode_to_address_format(b
, vtn_variable_mode_function
));
1581 const struct vtn_type
*sampled_type
= vtn_get_type(b
, w
[2]);
1582 if (b
->shader
->info
.stage
== MESA_SHADER_KERNEL
) {
1583 vtn_fail_if(sampled_type
->base_type
!= vtn_base_type_void
,
1584 "Sampled type of OpTypeImage must be void for kernels");
1586 vtn_fail_if(sampled_type
->base_type
!= vtn_base_type_scalar
||
1587 glsl_get_bit_size(sampled_type
->type
) != 32,
1588 "Sampled type of OpTypeImage must be a 32-bit scalar");
1591 enum glsl_sampler_dim dim
;
1592 switch ((SpvDim
)w
[3]) {
1593 case SpvDim1D
: dim
= GLSL_SAMPLER_DIM_1D
; break;
1594 case SpvDim2D
: dim
= GLSL_SAMPLER_DIM_2D
; break;
1595 case SpvDim3D
: dim
= GLSL_SAMPLER_DIM_3D
; break;
1596 case SpvDimCube
: dim
= GLSL_SAMPLER_DIM_CUBE
; break;
1597 case SpvDimRect
: dim
= GLSL_SAMPLER_DIM_RECT
; break;
1598 case SpvDimBuffer
: dim
= GLSL_SAMPLER_DIM_BUF
; break;
1599 case SpvDimSubpassData
: dim
= GLSL_SAMPLER_DIM_SUBPASS
; break;
1601 vtn_fail("Invalid SPIR-V image dimensionality: %s (%u)",
1602 spirv_dim_to_string((SpvDim
)w
[3]), w
[3]);
1605 /* w[4]: as per Vulkan spec "Validation Rules within a Module",
1606 * The “Depth” operand of OpTypeImage is ignored.
1608 bool is_array
= w
[5];
1609 bool multisampled
= w
[6];
1610 unsigned sampled
= w
[7];
1611 SpvImageFormat format
= w
[8];
1614 val
->type
->access_qualifier
= w
[9];
1615 else if (b
->shader
->info
.stage
== MESA_SHADER_KERNEL
)
1616 /* Per the CL C spec: If no qualifier is provided, read_only is assumed. */
1617 val
->type
->access_qualifier
= SpvAccessQualifierReadOnly
;
1619 val
->type
->access_qualifier
= SpvAccessQualifierReadWrite
;
1622 if (dim
== GLSL_SAMPLER_DIM_2D
)
1623 dim
= GLSL_SAMPLER_DIM_MS
;
1624 else if (dim
== GLSL_SAMPLER_DIM_SUBPASS
)
1625 dim
= GLSL_SAMPLER_DIM_SUBPASS_MS
;
1627 vtn_fail("Unsupported multisampled image type");
1630 val
->type
->image_format
= translate_image_format(b
, format
);
1632 enum glsl_base_type sampled_base_type
=
1633 glsl_get_base_type(sampled_type
->type
);
1635 val
->type
->glsl_image
= glsl_sampler_type(dim
, false, is_array
,
1637 } else if (sampled
== 2) {
1638 val
->type
->glsl_image
= glsl_image_type(dim
, is_array
,
1640 } else if (b
->shader
->info
.stage
== MESA_SHADER_KERNEL
) {
1641 val
->type
->glsl_image
= glsl_image_type(dim
, is_array
,
1644 vtn_fail("We need to know if the image will be sampled");
1649 case SpvOpTypeSampledImage
: {
1650 val
->type
->base_type
= vtn_base_type_sampled_image
;
1651 val
->type
->image
= vtn_get_type(b
, w
[2]);
1653 /* Sampled images are represented NIR as a vec2 SSA value where each
1654 * component is the result of a deref instruction. The first component
1655 * is the image and the second is the sampler. An OpLoad on an
1656 * OpTypeSampledImage pointer from UniformConstant memory just takes
1657 * the NIR deref from the pointer and duplicates it to both vector
1660 nir_address_format addr_format
=
1661 vtn_mode_to_address_format(b
, vtn_variable_mode_function
);
1662 assert(nir_address_format_num_components(addr_format
) == 1);
1663 unsigned bit_size
= nir_address_format_bit_size(addr_format
);
1664 assert(bit_size
== 32 || bit_size
== 64);
1666 enum glsl_base_type base_type
=
1667 bit_size
== 32 ? GLSL_TYPE_UINT
: GLSL_TYPE_UINT64
;
1668 val
->type
->type
= glsl_vector_type(base_type
, 2);
1672 case SpvOpTypeSampler
:
1673 val
->type
->base_type
= vtn_base_type_sampler
;
1675 /* Samplers are represented in NIR as a scalar SSA value that is the
1676 * result of a deref instruction. An OpLoad on an OpTypeSampler pointer
1677 * from UniformConstant memory just takes the NIR deref from the pointer
1678 * and turns it into an SSA value.
1680 val
->type
->type
= nir_address_format_to_glsl_type(
1681 vtn_mode_to_address_format(b
, vtn_variable_mode_function
));
1684 case SpvOpTypeOpaque
:
1685 case SpvOpTypeEvent
:
1686 case SpvOpTypeDeviceEvent
:
1687 case SpvOpTypeReserveId
:
1688 case SpvOpTypeQueue
:
1691 vtn_fail_with_opcode("Unhandled opcode", opcode
);
1694 vtn_foreach_decoration(b
, val
, type_decoration_cb
, NULL
);
1696 if (val
->type
->base_type
== vtn_base_type_struct
&&
1697 (val
->type
->block
|| val
->type
->buffer_block
)) {
1698 for (unsigned i
= 0; i
< val
->type
->length
; i
++) {
1699 vtn_fail_if(vtn_type_contains_block(b
, val
->type
->members
[i
]),
1700 "Block and BufferBlock decorations cannot decorate a "
1701 "structure type that is nested at any level inside "
1702 "another structure type decorated with Block or "
1708 static nir_constant
*
1709 vtn_null_constant(struct vtn_builder
*b
, struct vtn_type
*type
)
1711 nir_constant
*c
= rzalloc(b
, nir_constant
);
1713 switch (type
->base_type
) {
1714 case vtn_base_type_scalar
:
1715 case vtn_base_type_vector
:
1716 /* Nothing to do here. It's already initialized to zero */
1719 case vtn_base_type_pointer
: {
1720 enum vtn_variable_mode mode
= vtn_storage_class_to_mode(
1721 b
, type
->storage_class
, type
->deref
, NULL
);
1722 nir_address_format addr_format
= vtn_mode_to_address_format(b
, mode
);
1724 const nir_const_value
*null_value
= nir_address_format_null_value(addr_format
);
1725 memcpy(c
->values
, null_value
,
1726 sizeof(nir_const_value
) * nir_address_format_num_components(addr_format
));
1730 case vtn_base_type_void
:
1731 case vtn_base_type_image
:
1732 case vtn_base_type_sampler
:
1733 case vtn_base_type_sampled_image
:
1734 case vtn_base_type_function
:
1735 /* For those we have to return something but it doesn't matter what. */
1738 case vtn_base_type_matrix
:
1739 case vtn_base_type_array
:
1740 vtn_assert(type
->length
> 0);
1741 c
->num_elements
= type
->length
;
1742 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1744 c
->elements
[0] = vtn_null_constant(b
, type
->array_element
);
1745 for (unsigned i
= 1; i
< c
->num_elements
; i
++)
1746 c
->elements
[i
] = c
->elements
[0];
1749 case vtn_base_type_struct
:
1750 c
->num_elements
= type
->length
;
1751 c
->elements
= ralloc_array(b
, nir_constant
*, c
->num_elements
);
1752 for (unsigned i
= 0; i
< c
->num_elements
; i
++)
1753 c
->elements
[i
] = vtn_null_constant(b
, type
->members
[i
]);
1757 vtn_fail("Invalid type for null constant");
1764 spec_constant_decoration_cb(struct vtn_builder
*b
, UNUSED
struct vtn_value
*val
,
1765 ASSERTED
int member
,
1766 const struct vtn_decoration
*dec
, void *data
)
1768 vtn_assert(member
== -1);
1769 if (dec
->decoration
!= SpvDecorationSpecId
)
1772 nir_const_value
*value
= data
;
1773 for (unsigned i
= 0; i
< b
->num_specializations
; i
++) {
1774 if (b
->specializations
[i
].id
== dec
->operands
[0]) {
1775 *value
= b
->specializations
[i
].value
;
1782 handle_workgroup_size_decoration_cb(struct vtn_builder
*b
,
1783 struct vtn_value
*val
,
1784 ASSERTED
int member
,
1785 const struct vtn_decoration
*dec
,
1788 vtn_assert(member
== -1);
1789 if (dec
->decoration
!= SpvDecorationBuiltIn
||
1790 dec
->operands
[0] != SpvBuiltInWorkgroupSize
)
1793 vtn_assert(val
->type
->type
== glsl_vector_type(GLSL_TYPE_UINT
, 3));
1794 b
->workgroup_size_builtin
= val
;
1798 vtn_handle_constant(struct vtn_builder
*b
, SpvOp opcode
,
1799 const uint32_t *w
, unsigned count
)
1801 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_constant
);
1802 val
->constant
= rzalloc(b
, nir_constant
);
1804 case SpvOpConstantTrue
:
1805 case SpvOpConstantFalse
:
1806 case SpvOpSpecConstantTrue
:
1807 case SpvOpSpecConstantFalse
: {
1808 vtn_fail_if(val
->type
->type
!= glsl_bool_type(),
1809 "Result type of %s must be OpTypeBool",
1810 spirv_op_to_string(opcode
));
1812 bool bval
= (opcode
== SpvOpConstantTrue
||
1813 opcode
== SpvOpSpecConstantTrue
);
1815 nir_const_value u32val
= nir_const_value_for_uint(bval
, 32);
1817 if (opcode
== SpvOpSpecConstantTrue
||
1818 opcode
== SpvOpSpecConstantFalse
)
1819 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &u32val
);
1821 val
->constant
->values
[0].b
= u32val
.u32
!= 0;
1826 case SpvOpSpecConstant
: {
1827 vtn_fail_if(val
->type
->base_type
!= vtn_base_type_scalar
,
1828 "Result type of %s must be a scalar",
1829 spirv_op_to_string(opcode
));
1830 int bit_size
= glsl_get_bit_size(val
->type
->type
);
1833 val
->constant
->values
[0].u64
= vtn_u64_literal(&w
[3]);
1836 val
->constant
->values
[0].u32
= w
[3];
1839 val
->constant
->values
[0].u16
= w
[3];
1842 val
->constant
->values
[0].u8
= w
[3];
1845 vtn_fail("Unsupported SpvOpConstant bit size: %u", bit_size
);
1848 if (opcode
== SpvOpSpecConstant
)
1849 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
,
1850 &val
->constant
->values
[0]);
1854 case SpvOpSpecConstantComposite
:
1855 case SpvOpConstantComposite
: {
1856 unsigned elem_count
= count
- 3;
1857 vtn_fail_if(elem_count
!= val
->type
->length
,
1858 "%s has %u constituents, expected %u",
1859 spirv_op_to_string(opcode
), elem_count
, val
->type
->length
);
1861 nir_constant
**elems
= ralloc_array(b
, nir_constant
*, elem_count
);
1862 for (unsigned i
= 0; i
< elem_count
; i
++) {
1863 struct vtn_value
*val
= vtn_untyped_value(b
, w
[i
+ 3]);
1865 if (val
->value_type
== vtn_value_type_constant
) {
1866 elems
[i
] = val
->constant
;
1868 vtn_fail_if(val
->value_type
!= vtn_value_type_undef
,
1869 "only constants or undefs allowed for "
1870 "SpvOpConstantComposite");
1871 /* to make it easier, just insert a NULL constant for now */
1872 elems
[i
] = vtn_null_constant(b
, val
->type
);
1876 switch (val
->type
->base_type
) {
1877 case vtn_base_type_vector
: {
1878 assert(glsl_type_is_vector(val
->type
->type
));
1879 for (unsigned i
= 0; i
< elem_count
; i
++)
1880 val
->constant
->values
[i
] = elems
[i
]->values
[0];
1884 case vtn_base_type_matrix
:
1885 case vtn_base_type_struct
:
1886 case vtn_base_type_array
:
1887 ralloc_steal(val
->constant
, elems
);
1888 val
->constant
->num_elements
= elem_count
;
1889 val
->constant
->elements
= elems
;
1893 vtn_fail("Result type of %s must be a composite type",
1894 spirv_op_to_string(opcode
));
1899 case SpvOpSpecConstantOp
: {
1900 nir_const_value u32op
= nir_const_value_for_uint(w
[3], 32);
1901 vtn_foreach_decoration(b
, val
, spec_constant_decoration_cb
, &u32op
);
1902 SpvOp opcode
= u32op
.u32
;
1904 case SpvOpVectorShuffle
: {
1905 struct vtn_value
*v0
= &b
->values
[w
[4]];
1906 struct vtn_value
*v1
= &b
->values
[w
[5]];
1908 vtn_assert(v0
->value_type
== vtn_value_type_constant
||
1909 v0
->value_type
== vtn_value_type_undef
);
1910 vtn_assert(v1
->value_type
== vtn_value_type_constant
||
1911 v1
->value_type
== vtn_value_type_undef
);
1913 unsigned len0
= glsl_get_vector_elements(v0
->type
->type
);
1914 unsigned len1
= glsl_get_vector_elements(v1
->type
->type
);
1916 vtn_assert(len0
+ len1
< 16);
1918 unsigned bit_size
= glsl_get_bit_size(val
->type
->type
);
1919 unsigned bit_size0
= glsl_get_bit_size(v0
->type
->type
);
1920 unsigned bit_size1
= glsl_get_bit_size(v1
->type
->type
);
1922 vtn_assert(bit_size
== bit_size0
&& bit_size
== bit_size1
);
1923 (void)bit_size0
; (void)bit_size1
;
1925 nir_const_value undef
= { .u64
= 0xdeadbeefdeadbeef };
1926 nir_const_value combined
[NIR_MAX_VEC_COMPONENTS
* 2];
1928 if (v0
->value_type
== vtn_value_type_constant
) {
1929 for (unsigned i
= 0; i
< len0
; i
++)
1930 combined
[i
] = v0
->constant
->values
[i
];
1932 if (v1
->value_type
== vtn_value_type_constant
) {
1933 for (unsigned i
= 0; i
< len1
; i
++)
1934 combined
[len0
+ i
] = v1
->constant
->values
[i
];
1937 for (unsigned i
= 0, j
= 0; i
< count
- 6; i
++, j
++) {
1938 uint32_t comp
= w
[i
+ 6];
1939 if (comp
== (uint32_t)-1) {
1940 /* If component is not used, set the value to a known constant
1941 * to detect if it is wrongly used.
1943 val
->constant
->values
[j
] = undef
;
1945 vtn_fail_if(comp
>= len0
+ len1
,
1946 "All Component literals must either be FFFFFFFF "
1947 "or in [0, N - 1] (inclusive).");
1948 val
->constant
->values
[j
] = combined
[comp
];
1954 case SpvOpCompositeExtract
:
1955 case SpvOpCompositeInsert
: {
1956 struct vtn_value
*comp
;
1957 unsigned deref_start
;
1958 struct nir_constant
**c
;
1959 if (opcode
== SpvOpCompositeExtract
) {
1960 comp
= vtn_value(b
, w
[4], vtn_value_type_constant
);
1962 c
= &comp
->constant
;
1964 comp
= vtn_value(b
, w
[5], vtn_value_type_constant
);
1966 val
->constant
= nir_constant_clone(comp
->constant
,
1972 const struct vtn_type
*type
= comp
->type
;
1973 for (unsigned i
= deref_start
; i
< count
; i
++) {
1974 vtn_fail_if(w
[i
] > type
->length
,
1975 "%uth index of %s is %u but the type has only "
1976 "%u elements", i
- deref_start
,
1977 spirv_op_to_string(opcode
), w
[i
], type
->length
);
1979 switch (type
->base_type
) {
1980 case vtn_base_type_vector
:
1982 type
= type
->array_element
;
1985 case vtn_base_type_matrix
:
1986 case vtn_base_type_array
:
1987 c
= &(*c
)->elements
[w
[i
]];
1988 type
= type
->array_element
;
1991 case vtn_base_type_struct
:
1992 c
= &(*c
)->elements
[w
[i
]];
1993 type
= type
->members
[w
[i
]];
1997 vtn_fail("%s must only index into composite types",
1998 spirv_op_to_string(opcode
));
2002 if (opcode
== SpvOpCompositeExtract
) {
2006 unsigned num_components
= type
->length
;
2007 for (unsigned i
= 0; i
< num_components
; i
++)
2008 val
->constant
->values
[i
] = (*c
)->values
[elem
+ i
];
2011 struct vtn_value
*insert
=
2012 vtn_value(b
, w
[4], vtn_value_type_constant
);
2013 vtn_assert(insert
->type
== type
);
2015 *c
= insert
->constant
;
2017 unsigned num_components
= type
->length
;
2018 for (unsigned i
= 0; i
< num_components
; i
++)
2019 (*c
)->values
[elem
+ i
] = insert
->constant
->values
[i
];
2027 nir_alu_type dst_alu_type
= nir_get_nir_type_for_glsl_type(val
->type
->type
);
2028 nir_alu_type src_alu_type
= dst_alu_type
;
2029 unsigned num_components
= glsl_get_vector_elements(val
->type
->type
);
2032 vtn_assert(count
<= 7);
2038 /* We have a source in a conversion */
2040 nir_get_nir_type_for_glsl_type(vtn_get_value_type(b
, w
[4])->type
);
2041 /* We use the bitsize of the conversion source to evaluate the opcode later */
2042 bit_size
= glsl_get_bit_size(vtn_get_value_type(b
, w
[4])->type
);
2045 bit_size
= glsl_get_bit_size(val
->type
->type
);
2048 nir_op op
= vtn_nir_alu_op_for_spirv_opcode(b
, opcode
, &swap
,
2049 nir_alu_type_get_type_size(src_alu_type
),
2050 nir_alu_type_get_type_size(dst_alu_type
));
2051 nir_const_value src
[3][NIR_MAX_VEC_COMPONENTS
];
2053 for (unsigned i
= 0; i
< count
- 4; i
++) {
2054 struct vtn_value
*src_val
=
2055 vtn_value(b
, w
[4 + i
], vtn_value_type_constant
);
2057 /* If this is an unsized source, pull the bit size from the
2058 * source; otherwise, we'll use the bit size from the destination.
2060 if (!nir_alu_type_get_type_size(nir_op_infos
[op
].input_types
[i
]))
2061 bit_size
= glsl_get_bit_size(src_val
->type
->type
);
2063 unsigned src_comps
= nir_op_infos
[op
].input_sizes
[i
] ?
2064 nir_op_infos
[op
].input_sizes
[i
] :
2067 unsigned j
= swap
? 1 - i
: i
;
2068 for (unsigned c
= 0; c
< src_comps
; c
++)
2069 src
[j
][c
] = src_val
->constant
->values
[c
];
2072 /* fix up fixed size sources */
2079 for (unsigned i
= 0; i
< num_components
; ++i
) {
2081 case 64: src
[1][i
].u32
= src
[1][i
].u64
; break;
2082 case 16: src
[1][i
].u32
= src
[1][i
].u16
; break;
2083 case 8: src
[1][i
].u32
= src
[1][i
].u8
; break;
2092 nir_const_value
*srcs
[3] = {
2093 src
[0], src
[1], src
[2],
2095 nir_eval_const_opcode(op
, val
->constant
->values
,
2096 num_components
, bit_size
, srcs
,
2097 b
->shader
->info
.float_controls_execution_mode
);
2104 case SpvOpConstantNull
:
2105 val
->constant
= vtn_null_constant(b
, val
->type
);
2109 vtn_fail_with_opcode("Unhandled opcode", opcode
);
2112 /* Now that we have the value, update the workgroup size if needed */
2113 vtn_foreach_decoration(b
, val
, handle_workgroup_size_decoration_cb
, NULL
);
2117 vtn_split_barrier_semantics(struct vtn_builder
*b
,
2118 SpvMemorySemanticsMask semantics
,
2119 SpvMemorySemanticsMask
*before
,
2120 SpvMemorySemanticsMask
*after
)
2122 /* For memory semantics embedded in operations, we split them into up to
2123 * two barriers, to be added before and after the operation. This is less
2124 * strict than if we propagated until the final backend stage, but still
2125 * result in correct execution.
2127 * A further improvement could be pipe this information (and use!) into the
2128 * next compiler layers, at the expense of making the handling of barriers
2132 *before
= SpvMemorySemanticsMaskNone
;
2133 *after
= SpvMemorySemanticsMaskNone
;
2135 SpvMemorySemanticsMask order_semantics
=
2136 semantics
& (SpvMemorySemanticsAcquireMask
|
2137 SpvMemorySemanticsReleaseMask
|
2138 SpvMemorySemanticsAcquireReleaseMask
|
2139 SpvMemorySemanticsSequentiallyConsistentMask
);
2141 if (util_bitcount(order_semantics
) > 1) {
2142 /* Old GLSLang versions incorrectly set all the ordering bits. This was
2143 * fixed in c51287d744fb6e7e9ccc09f6f8451e6c64b1dad6 of glslang repo,
2144 * and it is in GLSLang since revision "SPIRV99.1321" (from Jul-2016).
2146 vtn_warn("Multiple memory ordering semantics specified, "
2147 "assuming AcquireRelease.");
2148 order_semantics
= SpvMemorySemanticsAcquireReleaseMask
;
2151 const SpvMemorySemanticsMask av_vis_semantics
=
2152 semantics
& (SpvMemorySemanticsMakeAvailableMask
|
2153 SpvMemorySemanticsMakeVisibleMask
);
2155 const SpvMemorySemanticsMask storage_semantics
=
2156 semantics
& (SpvMemorySemanticsUniformMemoryMask
|
2157 SpvMemorySemanticsSubgroupMemoryMask
|
2158 SpvMemorySemanticsWorkgroupMemoryMask
|
2159 SpvMemorySemanticsCrossWorkgroupMemoryMask
|
2160 SpvMemorySemanticsAtomicCounterMemoryMask
|
2161 SpvMemorySemanticsImageMemoryMask
|
2162 SpvMemorySemanticsOutputMemoryMask
);
2164 const SpvMemorySemanticsMask other_semantics
=
2165 semantics
& ~(order_semantics
| av_vis_semantics
| storage_semantics
|
2166 SpvMemorySemanticsVolatileMask
);
2168 if (other_semantics
)
2169 vtn_warn("Ignoring unhandled memory semantics: %u\n", other_semantics
);
2171 /* SequentiallyConsistent is treated as AcquireRelease. */
2173 /* The RELEASE barrier happens BEFORE the operation, and it is usually
2174 * associated with a Store. All the write operations with a matching
2175 * semantics will not be reordered after the Store.
2177 if (order_semantics
& (SpvMemorySemanticsReleaseMask
|
2178 SpvMemorySemanticsAcquireReleaseMask
|
2179 SpvMemorySemanticsSequentiallyConsistentMask
)) {
2180 *before
|= SpvMemorySemanticsReleaseMask
| storage_semantics
;
2183 /* The ACQUIRE barrier happens AFTER the operation, and it is usually
2184 * associated with a Load. All the operations with a matching semantics
2185 * will not be reordered before the Load.
2187 if (order_semantics
& (SpvMemorySemanticsAcquireMask
|
2188 SpvMemorySemanticsAcquireReleaseMask
|
2189 SpvMemorySemanticsSequentiallyConsistentMask
)) {
2190 *after
|= SpvMemorySemanticsAcquireMask
| storage_semantics
;
2193 if (av_vis_semantics
& SpvMemorySemanticsMakeVisibleMask
)
2194 *before
|= SpvMemorySemanticsMakeVisibleMask
| storage_semantics
;
2196 if (av_vis_semantics
& SpvMemorySemanticsMakeAvailableMask
)
2197 *after
|= SpvMemorySemanticsMakeAvailableMask
| storage_semantics
;
2200 static nir_memory_semantics
2201 vtn_mem_semantics_to_nir_mem_semantics(struct vtn_builder
*b
,
2202 SpvMemorySemanticsMask semantics
)
2204 nir_memory_semantics nir_semantics
= 0;
2206 SpvMemorySemanticsMask order_semantics
=
2207 semantics
& (SpvMemorySemanticsAcquireMask
|
2208 SpvMemorySemanticsReleaseMask
|
2209 SpvMemorySemanticsAcquireReleaseMask
|
2210 SpvMemorySemanticsSequentiallyConsistentMask
);
2212 if (util_bitcount(order_semantics
) > 1) {
2213 /* Old GLSLang versions incorrectly set all the ordering bits. This was
2214 * fixed in c51287d744fb6e7e9ccc09f6f8451e6c64b1dad6 of glslang repo,
2215 * and it is in GLSLang since revision "SPIRV99.1321" (from Jul-2016).
2217 vtn_warn("Multiple memory ordering semantics bits specified, "
2218 "assuming AcquireRelease.");
2219 order_semantics
= SpvMemorySemanticsAcquireReleaseMask
;
2222 switch (order_semantics
) {
2224 /* Not an ordering barrier. */
2227 case SpvMemorySemanticsAcquireMask
:
2228 nir_semantics
= NIR_MEMORY_ACQUIRE
;
2231 case SpvMemorySemanticsReleaseMask
:
2232 nir_semantics
= NIR_MEMORY_RELEASE
;
2235 case SpvMemorySemanticsSequentiallyConsistentMask
:
2236 /* Fall through. Treated as AcquireRelease in Vulkan. */
2237 case SpvMemorySemanticsAcquireReleaseMask
:
2238 nir_semantics
= NIR_MEMORY_ACQUIRE
| NIR_MEMORY_RELEASE
;
2242 unreachable("Invalid memory order semantics");
2245 if (semantics
& SpvMemorySemanticsMakeAvailableMask
) {
2246 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
2247 "To use MakeAvailable memory semantics the VulkanMemoryModel "
2248 "capability must be declared.");
2249 nir_semantics
|= NIR_MEMORY_MAKE_AVAILABLE
;
2252 if (semantics
& SpvMemorySemanticsMakeVisibleMask
) {
2253 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
2254 "To use MakeVisible memory semantics the VulkanMemoryModel "
2255 "capability must be declared.");
2256 nir_semantics
|= NIR_MEMORY_MAKE_VISIBLE
;
2259 return nir_semantics
;
2262 static nir_variable_mode
2263 vtn_mem_sematics_to_nir_var_modes(struct vtn_builder
*b
,
2264 SpvMemorySemanticsMask semantics
)
2266 /* Vulkan Environment for SPIR-V says "SubgroupMemory, CrossWorkgroupMemory,
2267 * and AtomicCounterMemory are ignored".
2269 semantics
&= ~(SpvMemorySemanticsSubgroupMemoryMask
|
2270 SpvMemorySemanticsCrossWorkgroupMemoryMask
|
2271 SpvMemorySemanticsAtomicCounterMemoryMask
);
2273 /* TODO: Consider adding nir_var_mem_image mode to NIR so it can be used
2274 * for SpvMemorySemanticsImageMemoryMask.
2277 nir_variable_mode modes
= 0;
2278 if (semantics
& (SpvMemorySemanticsUniformMemoryMask
|
2279 SpvMemorySemanticsImageMemoryMask
)) {
2280 modes
|= nir_var_uniform
|
2285 if (semantics
& SpvMemorySemanticsWorkgroupMemoryMask
)
2286 modes
|= nir_var_mem_shared
;
2287 if (semantics
& SpvMemorySemanticsOutputMemoryMask
) {
2288 modes
|= nir_var_shader_out
;
2295 vtn_scope_to_nir_scope(struct vtn_builder
*b
, SpvScope scope
)
2297 nir_scope nir_scope
;
2299 case SpvScopeDevice
:
2300 vtn_fail_if(b
->options
->caps
.vk_memory_model
&&
2301 !b
->options
->caps
.vk_memory_model_device_scope
,
2302 "If the Vulkan memory model is declared and any instruction "
2303 "uses Device scope, the VulkanMemoryModelDeviceScope "
2304 "capability must be declared.");
2305 nir_scope
= NIR_SCOPE_DEVICE
;
2308 case SpvScopeQueueFamily
:
2309 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
2310 "To use Queue Family scope, the VulkanMemoryModel capability "
2311 "must be declared.");
2312 nir_scope
= NIR_SCOPE_QUEUE_FAMILY
;
2315 case SpvScopeWorkgroup
:
2316 nir_scope
= NIR_SCOPE_WORKGROUP
;
2319 case SpvScopeSubgroup
:
2320 nir_scope
= NIR_SCOPE_SUBGROUP
;
2323 case SpvScopeInvocation
:
2324 nir_scope
= NIR_SCOPE_INVOCATION
;
2328 vtn_fail("Invalid memory scope");
2335 vtn_emit_scoped_control_barrier(struct vtn_builder
*b
, SpvScope exec_scope
,
2337 SpvMemorySemanticsMask semantics
)
2339 nir_memory_semantics nir_semantics
=
2340 vtn_mem_semantics_to_nir_mem_semantics(b
, semantics
);
2341 nir_variable_mode modes
= vtn_mem_sematics_to_nir_var_modes(b
, semantics
);
2342 nir_scope nir_exec_scope
= vtn_scope_to_nir_scope(b
, exec_scope
);
2344 /* Memory semantics is optional for OpControlBarrier. */
2345 nir_scope nir_mem_scope
;
2346 if (nir_semantics
== 0 || modes
== 0)
2347 nir_mem_scope
= NIR_SCOPE_NONE
;
2349 nir_mem_scope
= vtn_scope_to_nir_scope(b
, mem_scope
);
2351 nir_scoped_barrier(&b
->nb
, nir_exec_scope
, nir_mem_scope
, nir_semantics
, modes
);
2355 vtn_emit_scoped_memory_barrier(struct vtn_builder
*b
, SpvScope scope
,
2356 SpvMemorySemanticsMask semantics
)
2358 nir_variable_mode modes
= vtn_mem_sematics_to_nir_var_modes(b
, semantics
);
2359 nir_memory_semantics nir_semantics
=
2360 vtn_mem_semantics_to_nir_mem_semantics(b
, semantics
);
2362 /* No barrier to add. */
2363 if (nir_semantics
== 0 || modes
== 0)
2366 nir_scope nir_mem_scope
= vtn_scope_to_nir_scope(b
, scope
);
2367 nir_scoped_barrier(&b
->nb
, NIR_SCOPE_NONE
, nir_mem_scope
, nir_semantics
, modes
);
2370 struct vtn_ssa_value
*
2371 vtn_create_ssa_value(struct vtn_builder
*b
, const struct glsl_type
*type
)
2373 /* Always use bare types for SSA values for a couple of reasons:
2375 * 1. Code which emits deref chains should never listen to the explicit
2376 * layout information on the SSA value if any exists. If we've
2377 * accidentally been relying on this, we want to find those bugs.
2379 * 2. We want to be able to quickly check that an SSA value being assigned
2380 * to a SPIR-V value has the right type. Using bare types everywhere
2381 * ensures that we can pointer-compare.
2383 struct vtn_ssa_value
*val
= rzalloc(b
, struct vtn_ssa_value
);
2384 val
->type
= glsl_get_bare_type(type
);
2387 if (!glsl_type_is_vector_or_scalar(type
)) {
2388 unsigned elems
= glsl_get_length(val
->type
);
2389 val
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
2390 if (glsl_type_is_array_or_matrix(type
)) {
2391 const struct glsl_type
*elem_type
= glsl_get_array_element(type
);
2392 for (unsigned i
= 0; i
< elems
; i
++)
2393 val
->elems
[i
] = vtn_create_ssa_value(b
, elem_type
);
2395 vtn_assert(glsl_type_is_struct_or_ifc(type
));
2396 for (unsigned i
= 0; i
< elems
; i
++) {
2397 const struct glsl_type
*elem_type
= glsl_get_struct_field(type
, i
);
2398 val
->elems
[i
] = vtn_create_ssa_value(b
, elem_type
);
2407 vtn_tex_src(struct vtn_builder
*b
, unsigned index
, nir_tex_src_type type
)
2410 src
.src
= nir_src_for_ssa(vtn_get_nir_ssa(b
, index
));
2411 src
.src_type
= type
;
2416 image_operand_arg(struct vtn_builder
*b
, const uint32_t *w
, uint32_t count
,
2417 uint32_t mask_idx
, SpvImageOperandsMask op
)
2419 static const SpvImageOperandsMask ops_with_arg
=
2420 SpvImageOperandsBiasMask
|
2421 SpvImageOperandsLodMask
|
2422 SpvImageOperandsGradMask
|
2423 SpvImageOperandsConstOffsetMask
|
2424 SpvImageOperandsOffsetMask
|
2425 SpvImageOperandsConstOffsetsMask
|
2426 SpvImageOperandsSampleMask
|
2427 SpvImageOperandsMinLodMask
|
2428 SpvImageOperandsMakeTexelAvailableMask
|
2429 SpvImageOperandsMakeTexelVisibleMask
;
2431 assert(util_bitcount(op
) == 1);
2432 assert(w
[mask_idx
] & op
);
2433 assert(op
& ops_with_arg
);
2435 uint32_t idx
= util_bitcount(w
[mask_idx
] & (op
- 1) & ops_with_arg
) + 1;
2437 /* Adjust indices for operands with two arguments. */
2438 static const SpvImageOperandsMask ops_with_two_args
=
2439 SpvImageOperandsGradMask
;
2440 idx
+= util_bitcount(w
[mask_idx
] & (op
- 1) & ops_with_two_args
);
2444 vtn_fail_if(idx
+ (op
& ops_with_two_args
? 1 : 0) >= count
,
2445 "Image op claims to have %s but does not enough "
2446 "following operands", spirv_imageoperands_to_string(op
));
2452 non_uniform_decoration_cb(struct vtn_builder
*b
,
2453 struct vtn_value
*val
, int member
,
2454 const struct vtn_decoration
*dec
, void *void_ctx
)
2456 enum gl_access_qualifier
*access
= void_ctx
;
2457 switch (dec
->decoration
) {
2458 case SpvDecorationNonUniformEXT
:
2459 *access
|= ACCESS_NON_UNIFORM
;
2468 vtn_handle_texture(struct vtn_builder
*b
, SpvOp opcode
,
2469 const uint32_t *w
, unsigned count
)
2471 struct vtn_type
*ret_type
= vtn_get_type(b
, w
[1]);
2473 if (opcode
== SpvOpSampledImage
) {
2474 struct vtn_sampled_image si
= {
2475 .image
= vtn_get_image(b
, w
[3]),
2476 .sampler
= vtn_get_sampler(b
, w
[4]),
2479 enum gl_access_qualifier access
= 0;
2480 vtn_foreach_decoration(b
, vtn_untyped_value(b
, w
[3]),
2481 non_uniform_decoration_cb
, &access
);
2482 vtn_foreach_decoration(b
, vtn_untyped_value(b
, w
[4]),
2483 non_uniform_decoration_cb
, &access
);
2485 vtn_push_sampled_image(b
, w
[2], si
, access
& ACCESS_NON_UNIFORM
);
2487 } else if (opcode
== SpvOpImage
) {
2488 struct vtn_sampled_image si
= vtn_get_sampled_image(b
, w
[3]);
2490 enum gl_access_qualifier access
= 0;
2491 vtn_foreach_decoration(b
, vtn_untyped_value(b
, w
[3]),
2492 non_uniform_decoration_cb
, &access
);
2494 vtn_push_image(b
, w
[2], si
.image
, access
& ACCESS_NON_UNIFORM
);
2498 nir_deref_instr
*image
= NULL
, *sampler
= NULL
;
2499 struct vtn_value
*sampled_val
= vtn_untyped_value(b
, w
[3]);
2500 if (sampled_val
->type
->base_type
== vtn_base_type_sampled_image
) {
2501 struct vtn_sampled_image si
= vtn_get_sampled_image(b
, w
[3]);
2503 sampler
= si
.sampler
;
2505 image
= vtn_get_image(b
, w
[3]);
2508 const enum glsl_sampler_dim sampler_dim
= glsl_get_sampler_dim(image
->type
);
2509 const bool is_array
= glsl_sampler_type_is_array(image
->type
);
2510 nir_alu_type dest_type
= nir_type_invalid
;
2512 /* Figure out the base texture operation */
2515 case SpvOpImageSampleImplicitLod
:
2516 case SpvOpImageSampleDrefImplicitLod
:
2517 case SpvOpImageSampleProjImplicitLod
:
2518 case SpvOpImageSampleProjDrefImplicitLod
:
2519 texop
= nir_texop_tex
;
2522 case SpvOpImageSampleExplicitLod
:
2523 case SpvOpImageSampleDrefExplicitLod
:
2524 case SpvOpImageSampleProjExplicitLod
:
2525 case SpvOpImageSampleProjDrefExplicitLod
:
2526 texop
= nir_texop_txl
;
2529 case SpvOpImageFetch
:
2530 if (sampler_dim
== GLSL_SAMPLER_DIM_MS
) {
2531 texop
= nir_texop_txf_ms
;
2533 texop
= nir_texop_txf
;
2537 case SpvOpImageGather
:
2538 case SpvOpImageDrefGather
:
2539 texop
= nir_texop_tg4
;
2542 case SpvOpImageQuerySizeLod
:
2543 case SpvOpImageQuerySize
:
2544 texop
= nir_texop_txs
;
2545 dest_type
= nir_type_int
;
2548 case SpvOpImageQueryLod
:
2549 texop
= nir_texop_lod
;
2550 dest_type
= nir_type_float
;
2553 case SpvOpImageQueryLevels
:
2554 texop
= nir_texop_query_levels
;
2555 dest_type
= nir_type_int
;
2558 case SpvOpImageQuerySamples
:
2559 texop
= nir_texop_texture_samples
;
2560 dest_type
= nir_type_int
;
2563 case SpvOpFragmentFetchAMD
:
2564 texop
= nir_texop_fragment_fetch
;
2567 case SpvOpFragmentMaskFetchAMD
:
2568 texop
= nir_texop_fragment_mask_fetch
;
2569 dest_type
= nir_type_uint
;
2573 vtn_fail_with_opcode("Unhandled opcode", opcode
);
2576 nir_tex_src srcs
[10]; /* 10 should be enough */
2577 nir_tex_src
*p
= srcs
;
2579 p
->src
= nir_src_for_ssa(&image
->dest
.ssa
);
2580 p
->src_type
= nir_tex_src_texture_deref
;
2590 vtn_fail_if(sampler
== NULL
,
2591 "%s requires an image of type OpTypeSampledImage",
2592 spirv_op_to_string(opcode
));
2593 p
->src
= nir_src_for_ssa(&sampler
->dest
.ssa
);
2594 p
->src_type
= nir_tex_src_sampler_deref
;
2598 case nir_texop_txf_ms
:
2600 case nir_texop_query_levels
:
2601 case nir_texop_texture_samples
:
2602 case nir_texop_samples_identical
:
2603 case nir_texop_fragment_fetch
:
2604 case nir_texop_fragment_mask_fetch
:
2607 case nir_texop_txf_ms_fb
:
2608 vtn_fail("unexpected nir_texop_txf_ms_fb");
2610 case nir_texop_txf_ms_mcs
:
2611 vtn_fail("unexpected nir_texop_txf_ms_mcs");
2612 case nir_texop_tex_prefetch
:
2613 vtn_fail("unexpected nir_texop_tex_prefetch");
2618 struct nir_ssa_def
*coord
;
2619 unsigned coord_components
;
2621 case SpvOpImageSampleImplicitLod
:
2622 case SpvOpImageSampleExplicitLod
:
2623 case SpvOpImageSampleDrefImplicitLod
:
2624 case SpvOpImageSampleDrefExplicitLod
:
2625 case SpvOpImageSampleProjImplicitLod
:
2626 case SpvOpImageSampleProjExplicitLod
:
2627 case SpvOpImageSampleProjDrefImplicitLod
:
2628 case SpvOpImageSampleProjDrefExplicitLod
:
2629 case SpvOpImageFetch
:
2630 case SpvOpImageGather
:
2631 case SpvOpImageDrefGather
:
2632 case SpvOpImageQueryLod
:
2633 case SpvOpFragmentFetchAMD
:
2634 case SpvOpFragmentMaskFetchAMD
: {
2635 /* All these types have the coordinate as their first real argument */
2636 coord_components
= glsl_get_sampler_dim_coordinate_components(sampler_dim
);
2638 if (is_array
&& texop
!= nir_texop_lod
)
2641 struct vtn_ssa_value
*coord_val
= vtn_ssa_value(b
, w
[idx
++]);
2642 coord
= coord_val
->def
;
2643 p
->src
= nir_src_for_ssa(nir_channels(&b
->nb
, coord
,
2644 (1 << coord_components
) - 1));
2646 /* OpenCL allows integer sampling coordinates */
2647 if (glsl_type_is_integer(coord_val
->type
) &&
2648 opcode
== SpvOpImageSampleExplicitLod
) {
2649 vtn_fail_if(b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
,
2650 "Unless the Kernel capability is being used, the coordinate parameter "
2651 "OpImageSampleExplicitLod must be floating point.");
2653 p
->src
= nir_src_for_ssa(nir_i2f32(&b
->nb
, p
->src
.ssa
));
2656 p
->src_type
= nir_tex_src_coord
;
2663 coord_components
= 0;
2668 case SpvOpImageSampleProjImplicitLod
:
2669 case SpvOpImageSampleProjExplicitLod
:
2670 case SpvOpImageSampleProjDrefImplicitLod
:
2671 case SpvOpImageSampleProjDrefExplicitLod
:
2672 /* These have the projector as the last coordinate component */
2673 p
->src
= nir_src_for_ssa(nir_channel(&b
->nb
, coord
, coord_components
));
2674 p
->src_type
= nir_tex_src_projector
;
2682 bool is_shadow
= false;
2683 unsigned gather_component
= 0;
2685 case SpvOpImageSampleDrefImplicitLod
:
2686 case SpvOpImageSampleDrefExplicitLod
:
2687 case SpvOpImageSampleProjDrefImplicitLod
:
2688 case SpvOpImageSampleProjDrefExplicitLod
:
2689 case SpvOpImageDrefGather
:
2690 /* These all have an explicit depth value as their next source */
2692 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_comparator
);
2695 case SpvOpImageGather
:
2696 /* This has a component as its next source */
2697 gather_component
= vtn_constant_uint(b
, w
[idx
++]);
2704 /* For OpImageQuerySizeLod, we always have an LOD */
2705 if (opcode
== SpvOpImageQuerySizeLod
)
2706 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_lod
);
2708 /* For OpFragmentFetchAMD, we always have a multisample index */
2709 if (opcode
== SpvOpFragmentFetchAMD
)
2710 (*p
++) = vtn_tex_src(b
, w
[idx
++], nir_tex_src_ms_index
);
2712 /* Now we need to handle some number of optional arguments */
2713 struct vtn_value
*gather_offsets
= NULL
;
2715 uint32_t operands
= w
[idx
];
2717 if (operands
& SpvImageOperandsBiasMask
) {
2718 vtn_assert(texop
== nir_texop_tex
||
2719 texop
== nir_texop_tg4
);
2720 if (texop
== nir_texop_tex
)
2721 texop
= nir_texop_txb
;
2722 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2723 SpvImageOperandsBiasMask
);
2724 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_bias
);
2727 if (operands
& SpvImageOperandsLodMask
) {
2728 vtn_assert(texop
== nir_texop_txl
|| texop
== nir_texop_txf
||
2729 texop
== nir_texop_txs
|| texop
== nir_texop_tg4
);
2730 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2731 SpvImageOperandsLodMask
);
2732 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_lod
);
2735 if (operands
& SpvImageOperandsGradMask
) {
2736 vtn_assert(texop
== nir_texop_txl
);
2737 texop
= nir_texop_txd
;
2738 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2739 SpvImageOperandsGradMask
);
2740 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_ddx
);
2741 (*p
++) = vtn_tex_src(b
, w
[arg
+ 1], nir_tex_src_ddy
);
2744 vtn_fail_if(util_bitcount(operands
& (SpvImageOperandsConstOffsetsMask
|
2745 SpvImageOperandsOffsetMask
|
2746 SpvImageOperandsConstOffsetMask
)) > 1,
2747 "At most one of the ConstOffset, Offset, and ConstOffsets "
2748 "image operands can be used on a given instruction.");
2750 if (operands
& SpvImageOperandsOffsetMask
) {
2751 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2752 SpvImageOperandsOffsetMask
);
2753 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_offset
);
2756 if (operands
& SpvImageOperandsConstOffsetMask
) {
2757 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2758 SpvImageOperandsConstOffsetMask
);
2759 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_offset
);
2762 if (operands
& SpvImageOperandsConstOffsetsMask
) {
2763 vtn_assert(texop
== nir_texop_tg4
);
2764 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2765 SpvImageOperandsConstOffsetsMask
);
2766 gather_offsets
= vtn_value(b
, w
[arg
], vtn_value_type_constant
);
2769 if (operands
& SpvImageOperandsSampleMask
) {
2770 vtn_assert(texop
== nir_texop_txf_ms
);
2771 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2772 SpvImageOperandsSampleMask
);
2773 texop
= nir_texop_txf_ms
;
2774 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_ms_index
);
2777 if (operands
& SpvImageOperandsMinLodMask
) {
2778 vtn_assert(texop
== nir_texop_tex
||
2779 texop
== nir_texop_txb
||
2780 texop
== nir_texop_txd
);
2781 uint32_t arg
= image_operand_arg(b
, w
, count
, idx
,
2782 SpvImageOperandsMinLodMask
);
2783 (*p
++) = vtn_tex_src(b
, w
[arg
], nir_tex_src_min_lod
);
2787 nir_tex_instr
*instr
= nir_tex_instr_create(b
->shader
, p
- srcs
);
2790 memcpy(instr
->src
, srcs
, instr
->num_srcs
* sizeof(*instr
->src
));
2792 instr
->coord_components
= coord_components
;
2793 instr
->sampler_dim
= sampler_dim
;
2794 instr
->is_array
= is_array
;
2795 instr
->is_shadow
= is_shadow
;
2796 instr
->is_new_style_shadow
=
2797 is_shadow
&& glsl_get_components(ret_type
->type
) == 1;
2798 instr
->component
= gather_component
;
2800 /* The Vulkan spec says:
2802 * "If an instruction loads from or stores to a resource (including
2803 * atomics and image instructions) and the resource descriptor being
2804 * accessed is not dynamically uniform, then the operand corresponding
2805 * to that resource (e.g. the pointer or sampled image operand) must be
2806 * decorated with NonUniform."
2808 * It's very careful to specify that the exact operand must be decorated
2809 * NonUniform. The SPIR-V parser is not expected to chase through long
2810 * chains to find the NonUniform decoration. It's either right there or we
2811 * can assume it doesn't exist.
2813 enum gl_access_qualifier access
= 0;
2814 vtn_foreach_decoration(b
, sampled_val
, non_uniform_decoration_cb
, &access
);
2816 if (sampled_val
->propagated_non_uniform
)
2817 access
|= ACCESS_NON_UNIFORM
;
2819 if (image
&& (access
& ACCESS_NON_UNIFORM
))
2820 instr
->texture_non_uniform
= true;
2822 if (sampler
&& (access
& ACCESS_NON_UNIFORM
))
2823 instr
->sampler_non_uniform
= true;
2825 /* for non-query ops, get dest_type from SPIR-V return type */
2826 if (dest_type
== nir_type_invalid
) {
2827 /* the return type should match the image type, unless the image type is
2828 * VOID (CL image), in which case the return type dictates the sampler
2830 enum glsl_base_type sampler_base
=
2831 glsl_get_sampler_result_type(image
->type
);
2832 enum glsl_base_type ret_base
= glsl_get_base_type(ret_type
->type
);
2833 vtn_fail_if(sampler_base
!= ret_base
&& sampler_base
!= GLSL_TYPE_VOID
,
2834 "SPIR-V return type mismatches image type. This is only valid "
2835 "for untyped images (OpenCL).");
2837 case GLSL_TYPE_FLOAT
: dest_type
= nir_type_float
; break;
2838 case GLSL_TYPE_INT
: dest_type
= nir_type_int
; break;
2839 case GLSL_TYPE_UINT
: dest_type
= nir_type_uint
; break;
2840 case GLSL_TYPE_BOOL
: dest_type
= nir_type_bool
; break;
2842 vtn_fail("Invalid base type for sampler result");
2846 instr
->dest_type
= dest_type
;
2848 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
2849 nir_tex_instr_dest_size(instr
), 32, NULL
);
2851 vtn_assert(glsl_get_vector_elements(ret_type
->type
) ==
2852 nir_tex_instr_dest_size(instr
));
2854 if (gather_offsets
) {
2855 vtn_fail_if(gather_offsets
->type
->base_type
!= vtn_base_type_array
||
2856 gather_offsets
->type
->length
!= 4,
2857 "ConstOffsets must be an array of size four of vectors "
2858 "of two integer components");
2860 struct vtn_type
*vec_type
= gather_offsets
->type
->array_element
;
2861 vtn_fail_if(vec_type
->base_type
!= vtn_base_type_vector
||
2862 vec_type
->length
!= 2 ||
2863 !glsl_type_is_integer(vec_type
->type
),
2864 "ConstOffsets must be an array of size four of vectors "
2865 "of two integer components");
2867 unsigned bit_size
= glsl_get_bit_size(vec_type
->type
);
2868 for (uint32_t i
= 0; i
< 4; i
++) {
2869 const nir_const_value
*cvec
=
2870 gather_offsets
->constant
->elements
[i
]->values
;
2871 for (uint32_t j
= 0; j
< 2; j
++) {
2873 case 8: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i8
; break;
2874 case 16: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i16
; break;
2875 case 32: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i32
; break;
2876 case 64: instr
->tg4_offsets
[i
][j
] = cvec
[j
].i64
; break;
2878 vtn_fail("Unsupported bit size: %u", bit_size
);
2884 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
2886 vtn_push_nir_ssa(b
, w
[2], &instr
->dest
.ssa
);
2890 fill_common_atomic_sources(struct vtn_builder
*b
, SpvOp opcode
,
2891 const uint32_t *w
, nir_src
*src
)
2893 const struct glsl_type
*type
= vtn_get_type(b
, w
[1])->type
;
2894 unsigned bit_size
= glsl_get_bit_size(type
);
2897 case SpvOpAtomicIIncrement
:
2898 src
[0] = nir_src_for_ssa(nir_imm_intN_t(&b
->nb
, 1, bit_size
));
2901 case SpvOpAtomicIDecrement
:
2902 src
[0] = nir_src_for_ssa(nir_imm_intN_t(&b
->nb
, -1, bit_size
));
2905 case SpvOpAtomicISub
:
2907 nir_src_for_ssa(nir_ineg(&b
->nb
, vtn_get_nir_ssa(b
, w
[6])));
2910 case SpvOpAtomicCompareExchange
:
2911 case SpvOpAtomicCompareExchangeWeak
:
2912 src
[0] = nir_src_for_ssa(vtn_get_nir_ssa(b
, w
[8]));
2913 src
[1] = nir_src_for_ssa(vtn_get_nir_ssa(b
, w
[7]));
2916 case SpvOpAtomicExchange
:
2917 case SpvOpAtomicIAdd
:
2918 case SpvOpAtomicSMin
:
2919 case SpvOpAtomicUMin
:
2920 case SpvOpAtomicSMax
:
2921 case SpvOpAtomicUMax
:
2922 case SpvOpAtomicAnd
:
2924 case SpvOpAtomicXor
:
2925 case SpvOpAtomicFAddEXT
:
2926 src
[0] = nir_src_for_ssa(vtn_get_nir_ssa(b
, w
[6]));
2930 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
2934 static nir_ssa_def
*
2935 get_image_coord(struct vtn_builder
*b
, uint32_t value
)
2937 nir_ssa_def
*coord
= vtn_get_nir_ssa(b
, value
);
2939 /* The image_load_store intrinsics assume a 4-dim coordinate */
2940 unsigned swizzle
[4];
2941 for (unsigned i
= 0; i
< 4; i
++)
2942 swizzle
[i
] = MIN2(i
, coord
->num_components
- 1);
2944 return nir_swizzle(&b
->nb
, coord
, swizzle
, 4);
2947 static nir_ssa_def
*
2948 expand_to_vec4(nir_builder
*b
, nir_ssa_def
*value
)
2950 if (value
->num_components
== 4)
2954 for (unsigned i
= 0; i
< 4; i
++)
2955 swiz
[i
] = i
< value
->num_components
? i
: 0;
2956 return nir_swizzle(b
, value
, swiz
, 4);
2960 vtn_handle_image(struct vtn_builder
*b
, SpvOp opcode
,
2961 const uint32_t *w
, unsigned count
)
2963 /* Just get this one out of the way */
2964 if (opcode
== SpvOpImageTexelPointer
) {
2965 struct vtn_value
*val
=
2966 vtn_push_value(b
, w
[2], vtn_value_type_image_pointer
);
2967 val
->image
= ralloc(b
, struct vtn_image_pointer
);
2969 val
->image
->image
= vtn_nir_deref(b
, w
[3]);
2970 val
->image
->coord
= get_image_coord(b
, w
[4]);
2971 val
->image
->sample
= vtn_get_nir_ssa(b
, w
[5]);
2972 val
->image
->lod
= nir_imm_int(&b
->nb
, 0);
2976 struct vtn_image_pointer image
;
2977 SpvScope scope
= SpvScopeInvocation
;
2978 SpvMemorySemanticsMask semantics
= 0;
2980 enum gl_access_qualifier access
= 0;
2982 struct vtn_value
*res_val
;
2984 case SpvOpAtomicExchange
:
2985 case SpvOpAtomicCompareExchange
:
2986 case SpvOpAtomicCompareExchangeWeak
:
2987 case SpvOpAtomicIIncrement
:
2988 case SpvOpAtomicIDecrement
:
2989 case SpvOpAtomicIAdd
:
2990 case SpvOpAtomicISub
:
2991 case SpvOpAtomicLoad
:
2992 case SpvOpAtomicSMin
:
2993 case SpvOpAtomicUMin
:
2994 case SpvOpAtomicSMax
:
2995 case SpvOpAtomicUMax
:
2996 case SpvOpAtomicAnd
:
2998 case SpvOpAtomicXor
:
2999 case SpvOpAtomicFAddEXT
:
3000 res_val
= vtn_value(b
, w
[3], vtn_value_type_image_pointer
);
3001 image
= *res_val
->image
;
3002 scope
= vtn_constant_uint(b
, w
[4]);
3003 semantics
= vtn_constant_uint(b
, w
[5]);
3004 access
|= ACCESS_COHERENT
;
3007 case SpvOpAtomicStore
:
3008 res_val
= vtn_value(b
, w
[1], vtn_value_type_image_pointer
);
3009 image
= *res_val
->image
;
3010 scope
= vtn_constant_uint(b
, w
[2]);
3011 semantics
= vtn_constant_uint(b
, w
[3]);
3012 access
|= ACCESS_COHERENT
;
3015 case SpvOpImageQuerySizeLod
:
3016 res_val
= vtn_untyped_value(b
, w
[3]);
3017 image
.image
= vtn_get_image(b
, w
[3]);
3019 image
.sample
= NULL
;
3020 image
.lod
= vtn_ssa_value(b
, w
[4])->def
;
3023 case SpvOpImageQuerySize
:
3024 res_val
= vtn_untyped_value(b
, w
[3]);
3025 image
.image
= vtn_get_image(b
, w
[3]);
3027 image
.sample
= NULL
;
3031 case SpvOpImageQueryFormat
:
3032 case SpvOpImageQueryOrder
:
3033 res_val
= vtn_untyped_value(b
, w
[3]);
3034 image
.image
= vtn_get_image(b
, w
[3]);
3036 image
.sample
= NULL
;
3040 case SpvOpImageRead
: {
3041 res_val
= vtn_untyped_value(b
, w
[3]);
3042 image
.image
= vtn_get_image(b
, w
[3]);
3043 image
.coord
= get_image_coord(b
, w
[4]);
3045 const SpvImageOperandsMask operands
=
3046 count
> 5 ? w
[5] : SpvImageOperandsMaskNone
;
3048 if (operands
& SpvImageOperandsSampleMask
) {
3049 uint32_t arg
= image_operand_arg(b
, w
, count
, 5,
3050 SpvImageOperandsSampleMask
);
3051 image
.sample
= vtn_get_nir_ssa(b
, w
[arg
]);
3053 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
3056 if (operands
& SpvImageOperandsMakeTexelVisibleMask
) {
3057 vtn_fail_if((operands
& SpvImageOperandsNonPrivateTexelMask
) == 0,
3058 "MakeTexelVisible requires NonPrivateTexel to also be set.");
3059 uint32_t arg
= image_operand_arg(b
, w
, count
, 5,
3060 SpvImageOperandsMakeTexelVisibleMask
);
3061 semantics
= SpvMemorySemanticsMakeVisibleMask
;
3062 scope
= vtn_constant_uint(b
, w
[arg
]);
3065 if (operands
& SpvImageOperandsLodMask
) {
3066 uint32_t arg
= image_operand_arg(b
, w
, count
, 5,
3067 SpvImageOperandsLodMask
);
3068 image
.lod
= vtn_get_nir_ssa(b
, w
[arg
]);
3070 image
.lod
= nir_imm_int(&b
->nb
, 0);
3073 if (operands
& SpvImageOperandsVolatileTexelMask
)
3074 access
|= ACCESS_VOLATILE
;
3079 case SpvOpImageWrite
: {
3080 res_val
= vtn_untyped_value(b
, w
[1]);
3081 image
.image
= vtn_get_image(b
, w
[1]);
3082 image
.coord
= get_image_coord(b
, w
[2]);
3086 const SpvImageOperandsMask operands
=
3087 count
> 4 ? w
[4] : SpvImageOperandsMaskNone
;
3089 if (operands
& SpvImageOperandsSampleMask
) {
3090 uint32_t arg
= image_operand_arg(b
, w
, count
, 4,
3091 SpvImageOperandsSampleMask
);
3092 image
.sample
= vtn_get_nir_ssa(b
, w
[arg
]);
3094 image
.sample
= nir_ssa_undef(&b
->nb
, 1, 32);
3097 if (operands
& SpvImageOperandsMakeTexelAvailableMask
) {
3098 vtn_fail_if((operands
& SpvImageOperandsNonPrivateTexelMask
) == 0,
3099 "MakeTexelAvailable requires NonPrivateTexel to also be set.");
3100 uint32_t arg
= image_operand_arg(b
, w
, count
, 4,
3101 SpvImageOperandsMakeTexelAvailableMask
);
3102 semantics
= SpvMemorySemanticsMakeAvailableMask
;
3103 scope
= vtn_constant_uint(b
, w
[arg
]);
3106 if (operands
& SpvImageOperandsLodMask
) {
3107 uint32_t arg
= image_operand_arg(b
, w
, count
, 4,
3108 SpvImageOperandsLodMask
);
3109 image
.lod
= vtn_get_nir_ssa(b
, w
[arg
]);
3111 image
.lod
= nir_imm_int(&b
->nb
, 0);
3114 if (operands
& SpvImageOperandsVolatileTexelMask
)
3115 access
|= ACCESS_VOLATILE
;
3121 vtn_fail_with_opcode("Invalid image opcode", opcode
);
3124 if (semantics
& SpvMemorySemanticsVolatileMask
)
3125 access
|= ACCESS_VOLATILE
;
3127 nir_intrinsic_op op
;
3129 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_deref_##N; break;
3130 OP(ImageQuerySize
, size
)
3131 OP(ImageQuerySizeLod
, size
)
3133 OP(ImageWrite
, store
)
3134 OP(AtomicLoad
, load
)
3135 OP(AtomicStore
, store
)
3136 OP(AtomicExchange
, atomic_exchange
)
3137 OP(AtomicCompareExchange
, atomic_comp_swap
)
3138 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
3139 OP(AtomicIIncrement
, atomic_add
)
3140 OP(AtomicIDecrement
, atomic_add
)
3141 OP(AtomicIAdd
, atomic_add
)
3142 OP(AtomicISub
, atomic_add
)
3143 OP(AtomicSMin
, atomic_imin
)
3144 OP(AtomicUMin
, atomic_umin
)
3145 OP(AtomicSMax
, atomic_imax
)
3146 OP(AtomicUMax
, atomic_umax
)
3147 OP(AtomicAnd
, atomic_and
)
3148 OP(AtomicOr
, atomic_or
)
3149 OP(AtomicXor
, atomic_xor
)
3150 OP(AtomicFAddEXT
, atomic_fadd
)
3151 OP(ImageQueryFormat
, format
)
3152 OP(ImageQueryOrder
, order
)
3155 vtn_fail_with_opcode("Invalid image opcode", opcode
);
3158 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
3160 intrin
->src
[0] = nir_src_for_ssa(&image
.image
->dest
.ssa
);
3163 case SpvOpImageQuerySize
:
3164 case SpvOpImageQuerySizeLod
:
3165 case SpvOpImageQueryFormat
:
3166 case SpvOpImageQueryOrder
:
3169 /* The image coordinate is always 4 components but we may not have that
3170 * many. Swizzle to compensate.
3172 intrin
->src
[1] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, image
.coord
));
3173 intrin
->src
[2] = nir_src_for_ssa(image
.sample
);
3177 /* The Vulkan spec says:
3179 * "If an instruction loads from or stores to a resource (including
3180 * atomics and image instructions) and the resource descriptor being
3181 * accessed is not dynamically uniform, then the operand corresponding
3182 * to that resource (e.g. the pointer or sampled image operand) must be
3183 * decorated with NonUniform."
3185 * It's very careful to specify that the exact operand must be decorated
3186 * NonUniform. The SPIR-V parser is not expected to chase through long
3187 * chains to find the NonUniform decoration. It's either right there or we
3188 * can assume it doesn't exist.
3190 vtn_foreach_decoration(b
, res_val
, non_uniform_decoration_cb
, &access
);
3191 nir_intrinsic_set_access(intrin
, access
);
3194 case SpvOpImageQueryFormat
:
3195 case SpvOpImageQueryOrder
:
3196 /* No additional sources */
3198 case SpvOpImageQuerySize
:
3199 intrin
->src
[1] = nir_src_for_ssa(nir_imm_int(&b
->nb
, 0));
3201 case SpvOpImageQuerySizeLod
:
3202 intrin
->src
[1] = nir_src_for_ssa(image
.lod
);
3204 case SpvOpAtomicLoad
:
3205 case SpvOpImageRead
:
3206 /* Only OpImageRead can support a lod parameter if
3207 * SPV_AMD_shader_image_load_store_lod is used but the current NIR
3208 * intrinsics definition for atomics requires us to set it for
3211 intrin
->src
[3] = nir_src_for_ssa(image
.lod
);
3213 case SpvOpAtomicStore
:
3214 case SpvOpImageWrite
: {
3215 const uint32_t value_id
= opcode
== SpvOpAtomicStore
? w
[4] : w
[3];
3216 struct vtn_ssa_value
*value
= vtn_ssa_value(b
, value_id
);
3217 /* nir_intrinsic_image_deref_store always takes a vec4 value */
3218 assert(op
== nir_intrinsic_image_deref_store
);
3219 intrin
->num_components
= 4;
3220 intrin
->src
[3] = nir_src_for_ssa(expand_to_vec4(&b
->nb
, value
->def
));
3221 /* Only OpImageWrite can support a lod parameter if
3222 * SPV_AMD_shader_image_load_store_lod is used but the current NIR
3223 * intrinsics definition for atomics requires us to set it for
3226 intrin
->src
[4] = nir_src_for_ssa(image
.lod
);
3228 if (opcode
== SpvOpImageWrite
)
3229 nir_intrinsic_set_type(intrin
, nir_get_nir_type_for_glsl_type(value
->type
));
3233 case SpvOpAtomicCompareExchange
:
3234 case SpvOpAtomicCompareExchangeWeak
:
3235 case SpvOpAtomicIIncrement
:
3236 case SpvOpAtomicIDecrement
:
3237 case SpvOpAtomicExchange
:
3238 case SpvOpAtomicIAdd
:
3239 case SpvOpAtomicISub
:
3240 case SpvOpAtomicSMin
:
3241 case SpvOpAtomicUMin
:
3242 case SpvOpAtomicSMax
:
3243 case SpvOpAtomicUMax
:
3244 case SpvOpAtomicAnd
:
3246 case SpvOpAtomicXor
:
3247 case SpvOpAtomicFAddEXT
:
3248 fill_common_atomic_sources(b
, opcode
, w
, &intrin
->src
[3]);
3252 vtn_fail_with_opcode("Invalid image opcode", opcode
);
3255 /* Image operations implicitly have the Image storage memory semantics. */
3256 semantics
|= SpvMemorySemanticsImageMemoryMask
;
3258 SpvMemorySemanticsMask before_semantics
;
3259 SpvMemorySemanticsMask after_semantics
;
3260 vtn_split_barrier_semantics(b
, semantics
, &before_semantics
, &after_semantics
);
3262 if (before_semantics
)
3263 vtn_emit_memory_barrier(b
, scope
, before_semantics
);
3265 if (opcode
!= SpvOpImageWrite
&& opcode
!= SpvOpAtomicStore
) {
3266 struct vtn_type
*type
= vtn_get_type(b
, w
[1]);
3268 unsigned dest_components
= glsl_get_vector_elements(type
->type
);
3269 if (nir_intrinsic_infos
[op
].dest_components
== 0)
3270 intrin
->num_components
= dest_components
;
3272 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
,
3273 nir_intrinsic_dest_components(intrin
), 32, NULL
);
3275 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3277 nir_ssa_def
*result
= &intrin
->dest
.ssa
;
3278 if (nir_intrinsic_dest_components(intrin
) != dest_components
)
3279 result
= nir_channels(&b
->nb
, result
, (1 << dest_components
) - 1);
3281 vtn_push_nir_ssa(b
, w
[2], result
);
3283 if (opcode
== SpvOpImageRead
)
3284 nir_intrinsic_set_type(intrin
, nir_get_nir_type_for_glsl_type(type
->type
));
3286 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3289 if (after_semantics
)
3290 vtn_emit_memory_barrier(b
, scope
, after_semantics
);
3293 static nir_intrinsic_op
3294 get_ssbo_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
3297 case SpvOpAtomicLoad
: return nir_intrinsic_load_ssbo
;
3298 case SpvOpAtomicStore
: return nir_intrinsic_store_ssbo
;
3299 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
3300 OP(AtomicExchange
, atomic_exchange
)
3301 OP(AtomicCompareExchange
, atomic_comp_swap
)
3302 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
3303 OP(AtomicIIncrement
, atomic_add
)
3304 OP(AtomicIDecrement
, atomic_add
)
3305 OP(AtomicIAdd
, atomic_add
)
3306 OP(AtomicISub
, atomic_add
)
3307 OP(AtomicSMin
, atomic_imin
)
3308 OP(AtomicUMin
, atomic_umin
)
3309 OP(AtomicSMax
, atomic_imax
)
3310 OP(AtomicUMax
, atomic_umax
)
3311 OP(AtomicAnd
, atomic_and
)
3312 OP(AtomicOr
, atomic_or
)
3313 OP(AtomicXor
, atomic_xor
)
3314 OP(AtomicFAddEXT
, atomic_fadd
)
3317 vtn_fail_with_opcode("Invalid SSBO atomic", opcode
);
3321 static nir_intrinsic_op
3322 get_uniform_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
3325 #define OP(S, N) case SpvOp##S: return nir_intrinsic_atomic_counter_ ##N;
3326 OP(AtomicLoad
, read_deref
)
3327 OP(AtomicExchange
, exchange
)
3328 OP(AtomicCompareExchange
, comp_swap
)
3329 OP(AtomicCompareExchangeWeak
, comp_swap
)
3330 OP(AtomicIIncrement
, inc_deref
)
3331 OP(AtomicIDecrement
, post_dec_deref
)
3332 OP(AtomicIAdd
, add_deref
)
3333 OP(AtomicISub
, add_deref
)
3334 OP(AtomicUMin
, min_deref
)
3335 OP(AtomicUMax
, max_deref
)
3336 OP(AtomicAnd
, and_deref
)
3337 OP(AtomicOr
, or_deref
)
3338 OP(AtomicXor
, xor_deref
)
3341 /* We left the following out: AtomicStore, AtomicSMin and
3342 * AtomicSmax. Right now there are not nir intrinsics for them. At this
3343 * moment Atomic Counter support is needed for ARB_spirv support, so is
3344 * only need to support GLSL Atomic Counters that are uints and don't
3345 * allow direct storage.
3347 vtn_fail("Invalid uniform atomic");
3351 static nir_intrinsic_op
3352 get_deref_nir_atomic_op(struct vtn_builder
*b
, SpvOp opcode
)
3355 case SpvOpAtomicLoad
: return nir_intrinsic_load_deref
;
3356 case SpvOpAtomicStore
: return nir_intrinsic_store_deref
;
3357 #define OP(S, N) case SpvOp##S: return nir_intrinsic_deref_##N;
3358 OP(AtomicExchange
, atomic_exchange
)
3359 OP(AtomicCompareExchange
, atomic_comp_swap
)
3360 OP(AtomicCompareExchangeWeak
, atomic_comp_swap
)
3361 OP(AtomicIIncrement
, atomic_add
)
3362 OP(AtomicIDecrement
, atomic_add
)
3363 OP(AtomicIAdd
, atomic_add
)
3364 OP(AtomicISub
, atomic_add
)
3365 OP(AtomicSMin
, atomic_imin
)
3366 OP(AtomicUMin
, atomic_umin
)
3367 OP(AtomicSMax
, atomic_imax
)
3368 OP(AtomicUMax
, atomic_umax
)
3369 OP(AtomicAnd
, atomic_and
)
3370 OP(AtomicOr
, atomic_or
)
3371 OP(AtomicXor
, atomic_xor
)
3372 OP(AtomicFAddEXT
, atomic_fadd
)
3375 vtn_fail_with_opcode("Invalid shared atomic", opcode
);
3380 * Handles shared atomics, ssbo atomics and atomic counters.
3383 vtn_handle_atomics(struct vtn_builder
*b
, SpvOp opcode
,
3384 const uint32_t *w
, UNUSED
unsigned count
)
3386 struct vtn_pointer
*ptr
;
3387 nir_intrinsic_instr
*atomic
;
3389 SpvScope scope
= SpvScopeInvocation
;
3390 SpvMemorySemanticsMask semantics
= 0;
3391 enum gl_access_qualifier access
= 0;
3394 case SpvOpAtomicLoad
:
3395 case SpvOpAtomicExchange
:
3396 case SpvOpAtomicCompareExchange
:
3397 case SpvOpAtomicCompareExchangeWeak
:
3398 case SpvOpAtomicIIncrement
:
3399 case SpvOpAtomicIDecrement
:
3400 case SpvOpAtomicIAdd
:
3401 case SpvOpAtomicISub
:
3402 case SpvOpAtomicSMin
:
3403 case SpvOpAtomicUMin
:
3404 case SpvOpAtomicSMax
:
3405 case SpvOpAtomicUMax
:
3406 case SpvOpAtomicAnd
:
3408 case SpvOpAtomicXor
:
3409 case SpvOpAtomicFAddEXT
:
3410 ptr
= vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
3411 scope
= vtn_constant_uint(b
, w
[4]);
3412 semantics
= vtn_constant_uint(b
, w
[5]);
3415 case SpvOpAtomicStore
:
3416 ptr
= vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
3417 scope
= vtn_constant_uint(b
, w
[2]);
3418 semantics
= vtn_constant_uint(b
, w
[3]);
3422 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
3425 if (semantics
& SpvMemorySemanticsVolatileMask
)
3426 access
|= ACCESS_VOLATILE
;
3428 /* uniform as "atomic counter uniform" */
3429 if (ptr
->mode
== vtn_variable_mode_atomic_counter
) {
3430 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
3431 nir_intrinsic_op op
= get_uniform_nir_atomic_op(b
, opcode
);
3432 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
3433 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
3435 /* SSBO needs to initialize index/offset. In this case we don't need to,
3436 * as that info is already stored on the ptr->var->var nir_variable (see
3437 * vtn_create_variable)
3441 case SpvOpAtomicLoad
:
3442 case SpvOpAtomicExchange
:
3443 case SpvOpAtomicCompareExchange
:
3444 case SpvOpAtomicCompareExchangeWeak
:
3445 case SpvOpAtomicIIncrement
:
3446 case SpvOpAtomicIDecrement
:
3447 case SpvOpAtomicIAdd
:
3448 case SpvOpAtomicISub
:
3449 case SpvOpAtomicSMin
:
3450 case SpvOpAtomicUMin
:
3451 case SpvOpAtomicSMax
:
3452 case SpvOpAtomicUMax
:
3453 case SpvOpAtomicAnd
:
3455 case SpvOpAtomicXor
:
3456 /* Nothing: we don't need to call fill_common_atomic_sources here, as
3457 * atomic counter uniforms doesn't have sources
3462 unreachable("Invalid SPIR-V atomic");
3465 } else if (vtn_pointer_uses_ssa_offset(b
, ptr
)) {
3466 nir_ssa_def
*offset
, *index
;
3467 offset
= vtn_pointer_to_offset(b
, ptr
, &index
);
3469 assert(ptr
->mode
== vtn_variable_mode_ssbo
);
3471 nir_intrinsic_op op
= get_ssbo_nir_atomic_op(b
, opcode
);
3472 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
3474 nir_intrinsic_set_access(atomic
, access
| ACCESS_COHERENT
);
3478 case SpvOpAtomicLoad
:
3479 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
3480 nir_intrinsic_set_align(atomic
, 4, 0);
3481 if (ptr
->mode
== vtn_variable_mode_ssbo
)
3482 atomic
->src
[src
++] = nir_src_for_ssa(index
);
3483 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
3486 case SpvOpAtomicStore
:
3487 atomic
->num_components
= glsl_get_vector_elements(ptr
->type
->type
);
3488 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
3489 nir_intrinsic_set_align(atomic
, 4, 0);
3490 atomic
->src
[src
++] = nir_src_for_ssa(vtn_get_nir_ssa(b
, w
[4]));
3491 if (ptr
->mode
== vtn_variable_mode_ssbo
)
3492 atomic
->src
[src
++] = nir_src_for_ssa(index
);
3493 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
3496 case SpvOpAtomicExchange
:
3497 case SpvOpAtomicCompareExchange
:
3498 case SpvOpAtomicCompareExchangeWeak
:
3499 case SpvOpAtomicIIncrement
:
3500 case SpvOpAtomicIDecrement
:
3501 case SpvOpAtomicIAdd
:
3502 case SpvOpAtomicISub
:
3503 case SpvOpAtomicSMin
:
3504 case SpvOpAtomicUMin
:
3505 case SpvOpAtomicSMax
:
3506 case SpvOpAtomicUMax
:
3507 case SpvOpAtomicAnd
:
3509 case SpvOpAtomicXor
:
3510 case SpvOpAtomicFAddEXT
:
3511 if (ptr
->mode
== vtn_variable_mode_ssbo
)
3512 atomic
->src
[src
++] = nir_src_for_ssa(index
);
3513 atomic
->src
[src
++] = nir_src_for_ssa(offset
);
3514 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[src
]);
3518 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
3521 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
3522 const struct glsl_type
*deref_type
= deref
->type
;
3523 nir_intrinsic_op op
= get_deref_nir_atomic_op(b
, opcode
);
3524 atomic
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
3525 atomic
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
3527 if (ptr
->mode
!= vtn_variable_mode_workgroup
)
3528 access
|= ACCESS_COHERENT
;
3530 nir_intrinsic_set_access(atomic
, access
);
3533 case SpvOpAtomicLoad
:
3534 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
3537 case SpvOpAtomicStore
:
3538 atomic
->num_components
= glsl_get_vector_elements(deref_type
);
3539 nir_intrinsic_set_write_mask(atomic
, (1 << atomic
->num_components
) - 1);
3540 atomic
->src
[1] = nir_src_for_ssa(vtn_get_nir_ssa(b
, w
[4]));
3543 case SpvOpAtomicExchange
:
3544 case SpvOpAtomicCompareExchange
:
3545 case SpvOpAtomicCompareExchangeWeak
:
3546 case SpvOpAtomicIIncrement
:
3547 case SpvOpAtomicIDecrement
:
3548 case SpvOpAtomicIAdd
:
3549 case SpvOpAtomicISub
:
3550 case SpvOpAtomicSMin
:
3551 case SpvOpAtomicUMin
:
3552 case SpvOpAtomicSMax
:
3553 case SpvOpAtomicUMax
:
3554 case SpvOpAtomicAnd
:
3556 case SpvOpAtomicXor
:
3557 case SpvOpAtomicFAddEXT
:
3558 fill_common_atomic_sources(b
, opcode
, w
, &atomic
->src
[1]);
3562 vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode
);
3566 /* Atomic ordering operations will implicitly apply to the atomic operation
3567 * storage class, so include that too.
3569 semantics
|= vtn_mode_to_memory_semantics(ptr
->mode
);
3571 SpvMemorySemanticsMask before_semantics
;
3572 SpvMemorySemanticsMask after_semantics
;
3573 vtn_split_barrier_semantics(b
, semantics
, &before_semantics
, &after_semantics
);
3575 if (before_semantics
)
3576 vtn_emit_memory_barrier(b
, scope
, before_semantics
);
3578 if (opcode
!= SpvOpAtomicStore
) {
3579 struct vtn_type
*type
= vtn_get_type(b
, w
[1]);
3581 nir_ssa_dest_init(&atomic
->instr
, &atomic
->dest
,
3582 glsl_get_vector_elements(type
->type
),
3583 glsl_get_bit_size(type
->type
), NULL
);
3585 vtn_push_nir_ssa(b
, w
[2], &atomic
->dest
.ssa
);
3588 nir_builder_instr_insert(&b
->nb
, &atomic
->instr
);
3590 if (after_semantics
)
3591 vtn_emit_memory_barrier(b
, scope
, after_semantics
);
3594 static nir_alu_instr
*
3595 create_vec(struct vtn_builder
*b
, unsigned num_components
, unsigned bit_size
)
3597 nir_op op
= nir_op_vec(num_components
);
3598 nir_alu_instr
*vec
= nir_alu_instr_create(b
->shader
, op
);
3599 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, num_components
,
3601 vec
->dest
.write_mask
= (1 << num_components
) - 1;
3606 struct vtn_ssa_value
*
3607 vtn_ssa_transpose(struct vtn_builder
*b
, struct vtn_ssa_value
*src
)
3609 if (src
->transposed
)
3610 return src
->transposed
;
3612 struct vtn_ssa_value
*dest
=
3613 vtn_create_ssa_value(b
, glsl_transposed_type(src
->type
));
3615 for (unsigned i
= 0; i
< glsl_get_matrix_columns(dest
->type
); i
++) {
3616 nir_alu_instr
*vec
= create_vec(b
, glsl_get_matrix_columns(src
->type
),
3617 glsl_get_bit_size(src
->type
));
3618 if (glsl_type_is_vector_or_scalar(src
->type
)) {
3619 vec
->src
[0].src
= nir_src_for_ssa(src
->def
);
3620 vec
->src
[0].swizzle
[0] = i
;
3622 for (unsigned j
= 0; j
< glsl_get_matrix_columns(src
->type
); j
++) {
3623 vec
->src
[j
].src
= nir_src_for_ssa(src
->elems
[j
]->def
);
3624 vec
->src
[j
].swizzle
[0] = i
;
3627 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3628 dest
->elems
[i
]->def
= &vec
->dest
.dest
.ssa
;
3631 dest
->transposed
= src
;
3636 static nir_ssa_def
*
3637 vtn_vector_shuffle(struct vtn_builder
*b
, unsigned num_components
,
3638 nir_ssa_def
*src0
, nir_ssa_def
*src1
,
3639 const uint32_t *indices
)
3641 nir_alu_instr
*vec
= create_vec(b
, num_components
, src0
->bit_size
);
3643 for (unsigned i
= 0; i
< num_components
; i
++) {
3644 uint32_t index
= indices
[i
];
3645 if (index
== 0xffffffff) {
3647 nir_src_for_ssa(nir_ssa_undef(&b
->nb
, 1, src0
->bit_size
));
3648 } else if (index
< src0
->num_components
) {
3649 vec
->src
[i
].src
= nir_src_for_ssa(src0
);
3650 vec
->src
[i
].swizzle
[0] = index
;
3652 vec
->src
[i
].src
= nir_src_for_ssa(src1
);
3653 vec
->src
[i
].swizzle
[0] = index
- src0
->num_components
;
3657 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3659 return &vec
->dest
.dest
.ssa
;
3663 * Concatentates a number of vectors/scalars together to produce a vector
3665 static nir_ssa_def
*
3666 vtn_vector_construct(struct vtn_builder
*b
, unsigned num_components
,
3667 unsigned num_srcs
, nir_ssa_def
**srcs
)
3669 nir_alu_instr
*vec
= create_vec(b
, num_components
, srcs
[0]->bit_size
);
3671 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
3673 * "When constructing a vector, there must be at least two Constituent
3676 vtn_assert(num_srcs
>= 2);
3678 unsigned dest_idx
= 0;
3679 for (unsigned i
= 0; i
< num_srcs
; i
++) {
3680 nir_ssa_def
*src
= srcs
[i
];
3681 vtn_assert(dest_idx
+ src
->num_components
<= num_components
);
3682 for (unsigned j
= 0; j
< src
->num_components
; j
++) {
3683 vec
->src
[dest_idx
].src
= nir_src_for_ssa(src
);
3684 vec
->src
[dest_idx
].swizzle
[0] = j
;
3689 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
3691 * "When constructing a vector, the total number of components in all
3692 * the operands must equal the number of components in Result Type."
3694 vtn_assert(dest_idx
== num_components
);
3696 nir_builder_instr_insert(&b
->nb
, &vec
->instr
);
3698 return &vec
->dest
.dest
.ssa
;
3701 static struct vtn_ssa_value
*
3702 vtn_composite_copy(void *mem_ctx
, struct vtn_ssa_value
*src
)
3704 struct vtn_ssa_value
*dest
= rzalloc(mem_ctx
, struct vtn_ssa_value
);
3705 dest
->type
= src
->type
;
3707 if (glsl_type_is_vector_or_scalar(src
->type
)) {
3708 dest
->def
= src
->def
;
3710 unsigned elems
= glsl_get_length(src
->type
);
3712 dest
->elems
= ralloc_array(mem_ctx
, struct vtn_ssa_value
*, elems
);
3713 for (unsigned i
= 0; i
< elems
; i
++)
3714 dest
->elems
[i
] = vtn_composite_copy(mem_ctx
, src
->elems
[i
]);
3720 static struct vtn_ssa_value
*
3721 vtn_composite_insert(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
3722 struct vtn_ssa_value
*insert
, const uint32_t *indices
,
3723 unsigned num_indices
)
3725 struct vtn_ssa_value
*dest
= vtn_composite_copy(b
, src
);
3727 struct vtn_ssa_value
*cur
= dest
;
3729 for (i
= 0; i
< num_indices
- 1; i
++) {
3730 /* If we got a vector here, that means the next index will be trying to
3731 * dereference a scalar.
3733 vtn_fail_if(glsl_type_is_vector_or_scalar(cur
->type
),
3734 "OpCompositeInsert has too many indices.");
3735 vtn_fail_if(indices
[i
] >= glsl_get_length(cur
->type
),
3736 "All indices in an OpCompositeInsert must be in-bounds");
3737 cur
= cur
->elems
[indices
[i
]];
3740 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
3741 vtn_fail_if(indices
[i
] >= glsl_get_vector_elements(cur
->type
),
3742 "All indices in an OpCompositeInsert must be in-bounds");
3744 /* According to the SPIR-V spec, OpCompositeInsert may work down to
3745 * the component granularity. In that case, the last index will be
3746 * the index to insert the scalar into the vector.
3749 cur
->def
= nir_vector_insert_imm(&b
->nb
, cur
->def
, insert
->def
, indices
[i
]);
3751 vtn_fail_if(indices
[i
] >= glsl_get_length(cur
->type
),
3752 "All indices in an OpCompositeInsert must be in-bounds");
3753 cur
->elems
[indices
[i
]] = insert
;
3759 static struct vtn_ssa_value
*
3760 vtn_composite_extract(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
3761 const uint32_t *indices
, unsigned num_indices
)
3763 struct vtn_ssa_value
*cur
= src
;
3764 for (unsigned i
= 0; i
< num_indices
; i
++) {
3765 if (glsl_type_is_vector_or_scalar(cur
->type
)) {
3766 vtn_assert(i
== num_indices
- 1);
3767 vtn_fail_if(indices
[i
] >= glsl_get_vector_elements(cur
->type
),
3768 "All indices in an OpCompositeExtract must be in-bounds");
3770 /* According to the SPIR-V spec, OpCompositeExtract may work down to
3771 * the component granularity. The last index will be the index of the
3772 * vector to extract.
3775 const struct glsl_type
*scalar_type
=
3776 glsl_scalar_type(glsl_get_base_type(cur
->type
));
3777 struct vtn_ssa_value
*ret
= vtn_create_ssa_value(b
, scalar_type
);
3778 ret
->def
= nir_channel(&b
->nb
, cur
->def
, indices
[i
]);
3781 vtn_fail_if(indices
[i
] >= glsl_get_length(cur
->type
),
3782 "All indices in an OpCompositeExtract must be in-bounds");
3783 cur
= cur
->elems
[indices
[i
]];
3791 vtn_handle_composite(struct vtn_builder
*b
, SpvOp opcode
,
3792 const uint32_t *w
, unsigned count
)
3794 struct vtn_type
*type
= vtn_get_type(b
, w
[1]);
3795 struct vtn_ssa_value
*ssa
= vtn_create_ssa_value(b
, type
->type
);
3798 case SpvOpVectorExtractDynamic
:
3799 ssa
->def
= nir_vector_extract(&b
->nb
, vtn_get_nir_ssa(b
, w
[3]),
3800 vtn_get_nir_ssa(b
, w
[4]));
3803 case SpvOpVectorInsertDynamic
:
3804 ssa
->def
= nir_vector_insert(&b
->nb
, vtn_get_nir_ssa(b
, w
[3]),
3805 vtn_get_nir_ssa(b
, w
[4]),
3806 vtn_get_nir_ssa(b
, w
[5]));
3809 case SpvOpVectorShuffle
:
3810 ssa
->def
= vtn_vector_shuffle(b
, glsl_get_vector_elements(type
->type
),
3811 vtn_get_nir_ssa(b
, w
[3]),
3812 vtn_get_nir_ssa(b
, w
[4]),
3816 case SpvOpCompositeConstruct
: {
3817 unsigned elems
= count
- 3;
3819 if (glsl_type_is_vector_or_scalar(type
->type
)) {
3820 nir_ssa_def
*srcs
[NIR_MAX_VEC_COMPONENTS
];
3821 for (unsigned i
= 0; i
< elems
; i
++)
3822 srcs
[i
] = vtn_get_nir_ssa(b
, w
[3 + i
]);
3824 vtn_vector_construct(b
, glsl_get_vector_elements(type
->type
),
3827 ssa
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
3828 for (unsigned i
= 0; i
< elems
; i
++)
3829 ssa
->elems
[i
] = vtn_ssa_value(b
, w
[3 + i
]);
3833 case SpvOpCompositeExtract
:
3834 ssa
= vtn_composite_extract(b
, vtn_ssa_value(b
, w
[3]),
3838 case SpvOpCompositeInsert
:
3839 ssa
= vtn_composite_insert(b
, vtn_ssa_value(b
, w
[4]),
3840 vtn_ssa_value(b
, w
[3]),
3844 case SpvOpCopyLogical
:
3845 ssa
= vtn_composite_copy(b
, vtn_ssa_value(b
, w
[3]));
3847 case SpvOpCopyObject
:
3848 vtn_copy_value(b
, w
[3], w
[2]);
3852 vtn_fail_with_opcode("unknown composite operation", opcode
);
3855 vtn_push_ssa_value(b
, w
[2], ssa
);
3859 vtn_emit_barrier(struct vtn_builder
*b
, nir_intrinsic_op op
)
3861 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
3862 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3866 vtn_emit_memory_barrier(struct vtn_builder
*b
, SpvScope scope
,
3867 SpvMemorySemanticsMask semantics
)
3869 if (b
->shader
->options
->use_scoped_barrier
) {
3870 vtn_emit_scoped_memory_barrier(b
, scope
, semantics
);
3874 static const SpvMemorySemanticsMask all_memory_semantics
=
3875 SpvMemorySemanticsUniformMemoryMask
|
3876 SpvMemorySemanticsWorkgroupMemoryMask
|
3877 SpvMemorySemanticsAtomicCounterMemoryMask
|
3878 SpvMemorySemanticsImageMemoryMask
|
3879 SpvMemorySemanticsOutputMemoryMask
;
3881 /* If we're not actually doing a memory barrier, bail */
3882 if (!(semantics
& all_memory_semantics
))
3885 /* GL and Vulkan don't have these */
3886 vtn_assert(scope
!= SpvScopeCrossDevice
);
3888 if (scope
== SpvScopeSubgroup
)
3889 return; /* Nothing to do here */
3891 if (scope
== SpvScopeWorkgroup
) {
3892 vtn_emit_barrier(b
, nir_intrinsic_group_memory_barrier
);
3896 /* There's only two scopes thing left */
3897 vtn_assert(scope
== SpvScopeInvocation
|| scope
== SpvScopeDevice
);
3899 /* Map the GLSL memoryBarrier() construct and any barriers with more than one
3900 * semantic to the corresponding NIR one.
3902 if (util_bitcount(semantics
& all_memory_semantics
) > 1) {
3903 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier
);
3904 if (semantics
& SpvMemorySemanticsOutputMemoryMask
) {
3905 /* GLSL memoryBarrier() (and the corresponding NIR one) doesn't include
3906 * TCS outputs, so we have to emit it's own intrinsic for that. We
3907 * then need to emit another memory_barrier to prevent moving
3908 * non-output operations to before the tcs_patch barrier.
3910 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_tcs_patch
);
3911 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier
);
3916 /* Issue a more specific barrier */
3917 switch (semantics
& all_memory_semantics
) {
3918 case SpvMemorySemanticsUniformMemoryMask
:
3919 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_buffer
);
3921 case SpvMemorySemanticsWorkgroupMemoryMask
:
3922 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_shared
);
3924 case SpvMemorySemanticsAtomicCounterMemoryMask
:
3925 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_atomic_counter
);
3927 case SpvMemorySemanticsImageMemoryMask
:
3928 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_image
);
3930 case SpvMemorySemanticsOutputMemoryMask
:
3931 if (b
->nb
.shader
->info
.stage
== MESA_SHADER_TESS_CTRL
)
3932 vtn_emit_barrier(b
, nir_intrinsic_memory_barrier_tcs_patch
);
3940 vtn_handle_barrier(struct vtn_builder
*b
, SpvOp opcode
,
3941 const uint32_t *w
, UNUSED
unsigned count
)
3944 case SpvOpEmitVertex
:
3945 case SpvOpEmitStreamVertex
:
3946 case SpvOpEndPrimitive
:
3947 case SpvOpEndStreamPrimitive
: {
3948 nir_intrinsic_op intrinsic_op
;
3950 case SpvOpEmitVertex
:
3951 case SpvOpEmitStreamVertex
:
3952 intrinsic_op
= nir_intrinsic_emit_vertex
;
3954 case SpvOpEndPrimitive
:
3955 case SpvOpEndStreamPrimitive
:
3956 intrinsic_op
= nir_intrinsic_end_primitive
;
3959 unreachable("Invalid opcode");
3962 nir_intrinsic_instr
*intrin
=
3963 nir_intrinsic_instr_create(b
->shader
, intrinsic_op
);
3966 case SpvOpEmitStreamVertex
:
3967 case SpvOpEndStreamPrimitive
: {
3968 unsigned stream
= vtn_constant_uint(b
, w
[1]);
3969 nir_intrinsic_set_stream_id(intrin
, stream
);
3977 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
3981 case SpvOpMemoryBarrier
: {
3982 SpvScope scope
= vtn_constant_uint(b
, w
[1]);
3983 SpvMemorySemanticsMask semantics
= vtn_constant_uint(b
, w
[2]);
3984 vtn_emit_memory_barrier(b
, scope
, semantics
);
3988 case SpvOpControlBarrier
: {
3989 SpvScope execution_scope
= vtn_constant_uint(b
, w
[1]);
3990 SpvScope memory_scope
= vtn_constant_uint(b
, w
[2]);
3991 SpvMemorySemanticsMask memory_semantics
= vtn_constant_uint(b
, w
[3]);
3993 /* GLSLang, prior to commit 8297936dd6eb3, emitted OpControlBarrier with
3994 * memory semantics of None for GLSL barrier().
3995 * And before that, prior to c3f1cdfa, emitted the OpControlBarrier with
3996 * Device instead of Workgroup for execution scope.
3998 if (b
->wa_glslang_cs_barrier
&&
3999 b
->nb
.shader
->info
.stage
== MESA_SHADER_COMPUTE
&&
4000 (execution_scope
== SpvScopeWorkgroup
||
4001 execution_scope
== SpvScopeDevice
) &&
4002 memory_semantics
== SpvMemorySemanticsMaskNone
) {
4003 execution_scope
= SpvScopeWorkgroup
;
4004 memory_scope
= SpvScopeWorkgroup
;
4005 memory_semantics
= SpvMemorySemanticsAcquireReleaseMask
|
4006 SpvMemorySemanticsWorkgroupMemoryMask
;
4009 /* From the SPIR-V spec:
4011 * "When used with the TessellationControl execution model, it also
4012 * implicitly synchronizes the Output Storage Class: Writes to Output
4013 * variables performed by any invocation executed prior to a
4014 * OpControlBarrier will be visible to any other invocation after
4015 * return from that OpControlBarrier."
4017 if (b
->nb
.shader
->info
.stage
== MESA_SHADER_TESS_CTRL
) {
4018 memory_semantics
&= ~(SpvMemorySemanticsAcquireMask
|
4019 SpvMemorySemanticsReleaseMask
|
4020 SpvMemorySemanticsAcquireReleaseMask
|
4021 SpvMemorySemanticsSequentiallyConsistentMask
);
4022 memory_semantics
|= SpvMemorySemanticsAcquireReleaseMask
|
4023 SpvMemorySemanticsOutputMemoryMask
;
4026 if (b
->shader
->options
->use_scoped_barrier
) {
4027 vtn_emit_scoped_control_barrier(b
, execution_scope
, memory_scope
,
4030 vtn_emit_memory_barrier(b
, memory_scope
, memory_semantics
);
4032 if (execution_scope
== SpvScopeWorkgroup
)
4033 vtn_emit_barrier(b
, nir_intrinsic_control_barrier
);
4039 unreachable("unknown barrier instruction");
4044 gl_primitive_from_spv_execution_mode(struct vtn_builder
*b
,
4045 SpvExecutionMode mode
)
4048 case SpvExecutionModeInputPoints
:
4049 case SpvExecutionModeOutputPoints
:
4050 return 0; /* GL_POINTS */
4051 case SpvExecutionModeInputLines
:
4052 return 1; /* GL_LINES */
4053 case SpvExecutionModeInputLinesAdjacency
:
4054 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
4055 case SpvExecutionModeTriangles
:
4056 return 4; /* GL_TRIANGLES */
4057 case SpvExecutionModeInputTrianglesAdjacency
:
4058 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
4059 case SpvExecutionModeQuads
:
4060 return 7; /* GL_QUADS */
4061 case SpvExecutionModeIsolines
:
4062 return 0x8E7A; /* GL_ISOLINES */
4063 case SpvExecutionModeOutputLineStrip
:
4064 return 3; /* GL_LINE_STRIP */
4065 case SpvExecutionModeOutputTriangleStrip
:
4066 return 5; /* GL_TRIANGLE_STRIP */
4068 vtn_fail("Invalid primitive type: %s (%u)",
4069 spirv_executionmode_to_string(mode
), mode
);
4074 vertices_in_from_spv_execution_mode(struct vtn_builder
*b
,
4075 SpvExecutionMode mode
)
4078 case SpvExecutionModeInputPoints
:
4080 case SpvExecutionModeInputLines
:
4082 case SpvExecutionModeInputLinesAdjacency
:
4084 case SpvExecutionModeTriangles
:
4086 case SpvExecutionModeInputTrianglesAdjacency
:
4089 vtn_fail("Invalid GS input mode: %s (%u)",
4090 spirv_executionmode_to_string(mode
), mode
);
4094 static gl_shader_stage
4095 stage_for_execution_model(struct vtn_builder
*b
, SpvExecutionModel model
)
4098 case SpvExecutionModelVertex
:
4099 return MESA_SHADER_VERTEX
;
4100 case SpvExecutionModelTessellationControl
:
4101 return MESA_SHADER_TESS_CTRL
;
4102 case SpvExecutionModelTessellationEvaluation
:
4103 return MESA_SHADER_TESS_EVAL
;
4104 case SpvExecutionModelGeometry
:
4105 return MESA_SHADER_GEOMETRY
;
4106 case SpvExecutionModelFragment
:
4107 return MESA_SHADER_FRAGMENT
;
4108 case SpvExecutionModelGLCompute
:
4109 return MESA_SHADER_COMPUTE
;
4110 case SpvExecutionModelKernel
:
4111 return MESA_SHADER_KERNEL
;
4113 vtn_fail("Unsupported execution model: %s (%u)",
4114 spirv_executionmodel_to_string(model
), model
);
4118 #define spv_check_supported(name, cap) do { \
4119 if (!(b->options && b->options->caps.name)) \
4120 vtn_warn("Unsupported SPIR-V capability: %s (%u)", \
4121 spirv_capability_to_string(cap), cap); \
4126 vtn_handle_entry_point(struct vtn_builder
*b
, const uint32_t *w
,
4129 struct vtn_value
*entry_point
= &b
->values
[w
[2]];
4130 /* Let this be a name label regardless */
4131 unsigned name_words
;
4132 entry_point
->name
= vtn_string_literal(b
, &w
[3], count
- 3, &name_words
);
4134 if (strcmp(entry_point
->name
, b
->entry_point_name
) != 0 ||
4135 stage_for_execution_model(b
, w
[1]) != b
->entry_point_stage
)
4138 vtn_assert(b
->entry_point
== NULL
);
4139 b
->entry_point
= entry_point
;
4143 vtn_handle_preamble_instruction(struct vtn_builder
*b
, SpvOp opcode
,
4144 const uint32_t *w
, unsigned count
)
4151 case SpvSourceLanguageUnknown
: lang
= "unknown"; break;
4152 case SpvSourceLanguageESSL
: lang
= "ESSL"; break;
4153 case SpvSourceLanguageGLSL
: lang
= "GLSL"; break;
4154 case SpvSourceLanguageOpenCL_C
: lang
= "OpenCL C"; break;
4155 case SpvSourceLanguageOpenCL_CPP
: lang
= "OpenCL C++"; break;
4156 case SpvSourceLanguageHLSL
: lang
= "HLSL"; break;
4159 uint32_t version
= w
[2];
4162 (count
> 3) ? vtn_value(b
, w
[3], vtn_value_type_string
)->str
: "";
4164 vtn_info("Parsing SPIR-V from %s %u source file %s", lang
, version
, file
);
4168 case SpvOpSourceExtension
:
4169 case SpvOpSourceContinued
:
4170 case SpvOpExtension
:
4171 case SpvOpModuleProcessed
:
4172 /* Unhandled, but these are for debug so that's ok. */
4175 case SpvOpCapability
: {
4176 SpvCapability cap
= w
[1];
4178 case SpvCapabilityMatrix
:
4179 case SpvCapabilityShader
:
4180 case SpvCapabilityGeometry
:
4181 case SpvCapabilityGeometryPointSize
:
4182 case SpvCapabilityUniformBufferArrayDynamicIndexing
:
4183 case SpvCapabilitySampledImageArrayDynamicIndexing
:
4184 case SpvCapabilityStorageBufferArrayDynamicIndexing
:
4185 case SpvCapabilityStorageImageArrayDynamicIndexing
:
4186 case SpvCapabilityImageRect
:
4187 case SpvCapabilitySampledRect
:
4188 case SpvCapabilitySampled1D
:
4189 case SpvCapabilityImage1D
:
4190 case SpvCapabilitySampledCubeArray
:
4191 case SpvCapabilityImageCubeArray
:
4192 case SpvCapabilitySampledBuffer
:
4193 case SpvCapabilityImageBuffer
:
4194 case SpvCapabilityImageQuery
:
4195 case SpvCapabilityDerivativeControl
:
4196 case SpvCapabilityInterpolationFunction
:
4197 case SpvCapabilityMultiViewport
:
4198 case SpvCapabilitySampleRateShading
:
4199 case SpvCapabilityClipDistance
:
4200 case SpvCapabilityCullDistance
:
4201 case SpvCapabilityInputAttachment
:
4202 case SpvCapabilityImageGatherExtended
:
4203 case SpvCapabilityStorageImageExtendedFormats
:
4204 case SpvCapabilityVector16
:
4207 case SpvCapabilityLinkage
:
4208 case SpvCapabilityFloat16Buffer
:
4209 case SpvCapabilitySparseResidency
:
4210 vtn_warn("Unsupported SPIR-V capability: %s",
4211 spirv_capability_to_string(cap
));
4214 case SpvCapabilityMinLod
:
4215 spv_check_supported(min_lod
, cap
);
4218 case SpvCapabilityAtomicStorage
:
4219 spv_check_supported(atomic_storage
, cap
);
4222 case SpvCapabilityFloat64
:
4223 spv_check_supported(float64
, cap
);
4225 case SpvCapabilityInt64
:
4226 spv_check_supported(int64
, cap
);
4228 case SpvCapabilityInt16
:
4229 spv_check_supported(int16
, cap
);
4231 case SpvCapabilityInt8
:
4232 spv_check_supported(int8
, cap
);
4235 case SpvCapabilityTransformFeedback
:
4236 spv_check_supported(transform_feedback
, cap
);
4239 case SpvCapabilityGeometryStreams
:
4240 spv_check_supported(geometry_streams
, cap
);
4243 case SpvCapabilityInt64Atomics
:
4244 spv_check_supported(int64_atomics
, cap
);
4247 case SpvCapabilityStorageImageMultisample
:
4248 spv_check_supported(storage_image_ms
, cap
);
4251 case SpvCapabilityAddresses
:
4252 spv_check_supported(address
, cap
);
4255 case SpvCapabilityKernel
:
4256 spv_check_supported(kernel
, cap
);
4259 case SpvCapabilityImageBasic
:
4260 spv_check_supported(kernel_image
, cap
);
4263 case SpvCapabilityLiteralSampler
:
4264 spv_check_supported(literal_sampler
, cap
);
4267 case SpvCapabilityImageReadWrite
:
4268 case SpvCapabilityImageMipmap
:
4269 case SpvCapabilityPipes
:
4270 case SpvCapabilityDeviceEnqueue
:
4271 case SpvCapabilityGenericPointer
:
4272 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
4273 spirv_capability_to_string(cap
));
4276 case SpvCapabilityImageMSArray
:
4277 spv_check_supported(image_ms_array
, cap
);
4280 case SpvCapabilityTessellation
:
4281 case SpvCapabilityTessellationPointSize
:
4282 spv_check_supported(tessellation
, cap
);
4285 case SpvCapabilityDrawParameters
:
4286 spv_check_supported(draw_parameters
, cap
);
4289 case SpvCapabilityStorageImageReadWithoutFormat
:
4290 spv_check_supported(image_read_without_format
, cap
);
4293 case SpvCapabilityStorageImageWriteWithoutFormat
:
4294 spv_check_supported(image_write_without_format
, cap
);
4297 case SpvCapabilityDeviceGroup
:
4298 spv_check_supported(device_group
, cap
);
4301 case SpvCapabilityMultiView
:
4302 spv_check_supported(multiview
, cap
);
4305 case SpvCapabilityGroupNonUniform
:
4306 spv_check_supported(subgroup_basic
, cap
);
4309 case SpvCapabilitySubgroupVoteKHR
:
4310 case SpvCapabilityGroupNonUniformVote
:
4311 spv_check_supported(subgroup_vote
, cap
);
4314 case SpvCapabilitySubgroupBallotKHR
:
4315 case SpvCapabilityGroupNonUniformBallot
:
4316 spv_check_supported(subgroup_ballot
, cap
);
4319 case SpvCapabilityGroupNonUniformShuffle
:
4320 case SpvCapabilityGroupNonUniformShuffleRelative
:
4321 spv_check_supported(subgroup_shuffle
, cap
);
4324 case SpvCapabilityGroupNonUniformQuad
:
4325 spv_check_supported(subgroup_quad
, cap
);
4328 case SpvCapabilityGroupNonUniformArithmetic
:
4329 case SpvCapabilityGroupNonUniformClustered
:
4330 spv_check_supported(subgroup_arithmetic
, cap
);
4333 case SpvCapabilityGroups
:
4334 spv_check_supported(amd_shader_ballot
, cap
);
4337 case SpvCapabilityVariablePointersStorageBuffer
:
4338 case SpvCapabilityVariablePointers
:
4339 spv_check_supported(variable_pointers
, cap
);
4340 b
->variable_pointers
= true;
4343 case SpvCapabilityStorageUniformBufferBlock16
:
4344 case SpvCapabilityStorageUniform16
:
4345 case SpvCapabilityStoragePushConstant16
:
4346 case SpvCapabilityStorageInputOutput16
:
4347 spv_check_supported(storage_16bit
, cap
);
4350 case SpvCapabilityShaderLayer
:
4351 case SpvCapabilityShaderViewportIndex
:
4352 case SpvCapabilityShaderViewportIndexLayerEXT
:
4353 spv_check_supported(shader_viewport_index_layer
, cap
);
4356 case SpvCapabilityStorageBuffer8BitAccess
:
4357 case SpvCapabilityUniformAndStorageBuffer8BitAccess
:
4358 case SpvCapabilityStoragePushConstant8
:
4359 spv_check_supported(storage_8bit
, cap
);
4362 case SpvCapabilityShaderNonUniformEXT
:
4363 spv_check_supported(descriptor_indexing
, cap
);
4366 case SpvCapabilityInputAttachmentArrayDynamicIndexingEXT
:
4367 case SpvCapabilityUniformTexelBufferArrayDynamicIndexingEXT
:
4368 case SpvCapabilityStorageTexelBufferArrayDynamicIndexingEXT
:
4369 spv_check_supported(descriptor_array_dynamic_indexing
, cap
);
4372 case SpvCapabilityUniformBufferArrayNonUniformIndexingEXT
:
4373 case SpvCapabilitySampledImageArrayNonUniformIndexingEXT
:
4374 case SpvCapabilityStorageBufferArrayNonUniformIndexingEXT
:
4375 case SpvCapabilityStorageImageArrayNonUniformIndexingEXT
:
4376 case SpvCapabilityInputAttachmentArrayNonUniformIndexingEXT
:
4377 case SpvCapabilityUniformTexelBufferArrayNonUniformIndexingEXT
:
4378 case SpvCapabilityStorageTexelBufferArrayNonUniformIndexingEXT
:
4379 spv_check_supported(descriptor_array_non_uniform_indexing
, cap
);
4382 case SpvCapabilityRuntimeDescriptorArrayEXT
:
4383 spv_check_supported(runtime_descriptor_array
, cap
);
4386 case SpvCapabilityStencilExportEXT
:
4387 spv_check_supported(stencil_export
, cap
);
4390 case SpvCapabilitySampleMaskPostDepthCoverage
:
4391 spv_check_supported(post_depth_coverage
, cap
);
4394 case SpvCapabilityDenormFlushToZero
:
4395 case SpvCapabilityDenormPreserve
:
4396 case SpvCapabilitySignedZeroInfNanPreserve
:
4397 case SpvCapabilityRoundingModeRTE
:
4398 case SpvCapabilityRoundingModeRTZ
:
4399 spv_check_supported(float_controls
, cap
);
4402 case SpvCapabilityPhysicalStorageBufferAddresses
:
4403 spv_check_supported(physical_storage_buffer_address
, cap
);
4406 case SpvCapabilityComputeDerivativeGroupQuadsNV
:
4407 case SpvCapabilityComputeDerivativeGroupLinearNV
:
4408 spv_check_supported(derivative_group
, cap
);
4411 case SpvCapabilityFloat16
:
4412 spv_check_supported(float16
, cap
);
4415 case SpvCapabilityFragmentShaderSampleInterlockEXT
:
4416 spv_check_supported(fragment_shader_sample_interlock
, cap
);
4419 case SpvCapabilityFragmentShaderPixelInterlockEXT
:
4420 spv_check_supported(fragment_shader_pixel_interlock
, cap
);
4423 case SpvCapabilityDemoteToHelperInvocationEXT
:
4424 spv_check_supported(demote_to_helper_invocation
, cap
);
4427 case SpvCapabilityShaderClockKHR
:
4428 spv_check_supported(shader_clock
, cap
);
4431 case SpvCapabilityVulkanMemoryModel
:
4432 spv_check_supported(vk_memory_model
, cap
);
4435 case SpvCapabilityVulkanMemoryModelDeviceScope
:
4436 spv_check_supported(vk_memory_model_device_scope
, cap
);
4439 case SpvCapabilityImageReadWriteLodAMD
:
4440 spv_check_supported(amd_image_read_write_lod
, cap
);
4443 case SpvCapabilityIntegerFunctions2INTEL
:
4444 spv_check_supported(integer_functions2
, cap
);
4447 case SpvCapabilityFragmentMaskAMD
:
4448 spv_check_supported(amd_fragment_mask
, cap
);
4451 case SpvCapabilityImageGatherBiasLodAMD
:
4452 spv_check_supported(amd_image_gather_bias_lod
, cap
);
4455 case SpvCapabilityAtomicFloat32AddEXT
:
4456 spv_check_supported(float32_atomic_add
, cap
);
4459 case SpvCapabilityAtomicFloat64AddEXT
:
4460 spv_check_supported(float64_atomic_add
, cap
);
4464 vtn_fail("Unhandled capability: %s (%u)",
4465 spirv_capability_to_string(cap
), cap
);
4470 case SpvOpExtInstImport
:
4471 vtn_handle_extension(b
, opcode
, w
, count
);
4474 case SpvOpMemoryModel
:
4476 case SpvAddressingModelPhysical32
:
4477 vtn_fail_if(b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
,
4478 "AddressingModelPhysical32 only supported for kernels");
4479 b
->shader
->info
.cs
.ptr_size
= 32;
4480 b
->physical_ptrs
= true;
4481 assert(nir_address_format_bit_size(b
->options
->global_addr_format
) == 32);
4482 assert(nir_address_format_num_components(b
->options
->global_addr_format
) == 1);
4483 assert(nir_address_format_bit_size(b
->options
->shared_addr_format
) == 32);
4484 assert(nir_address_format_num_components(b
->options
->shared_addr_format
) == 1);
4485 assert(nir_address_format_bit_size(b
->options
->constant_addr_format
) == 32);
4486 assert(nir_address_format_num_components(b
->options
->constant_addr_format
) == 1);
4488 case SpvAddressingModelPhysical64
:
4489 vtn_fail_if(b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
,
4490 "AddressingModelPhysical64 only supported for kernels");
4491 b
->shader
->info
.cs
.ptr_size
= 64;
4492 b
->physical_ptrs
= true;
4493 assert(nir_address_format_bit_size(b
->options
->global_addr_format
) == 64);
4494 assert(nir_address_format_num_components(b
->options
->global_addr_format
) == 1);
4495 assert(nir_address_format_bit_size(b
->options
->shared_addr_format
) == 64);
4496 assert(nir_address_format_num_components(b
->options
->shared_addr_format
) == 1);
4497 assert(nir_address_format_bit_size(b
->options
->constant_addr_format
) == 64);
4498 assert(nir_address_format_num_components(b
->options
->constant_addr_format
) == 1);
4500 case SpvAddressingModelLogical
:
4501 vtn_fail_if(b
->shader
->info
.stage
== MESA_SHADER_KERNEL
,
4502 "AddressingModelLogical only supported for shaders");
4503 b
->physical_ptrs
= false;
4505 case SpvAddressingModelPhysicalStorageBuffer64
:
4506 vtn_fail_if(!b
->options
||
4507 !b
->options
->caps
.physical_storage_buffer_address
,
4508 "AddressingModelPhysicalStorageBuffer64 not supported");
4511 vtn_fail("Unknown addressing model: %s (%u)",
4512 spirv_addressingmodel_to_string(w
[1]), w
[1]);
4516 b
->mem_model
= w
[2];
4518 case SpvMemoryModelSimple
:
4519 case SpvMemoryModelGLSL450
:
4520 case SpvMemoryModelOpenCL
:
4522 case SpvMemoryModelVulkan
:
4523 vtn_fail_if(!b
->options
->caps
.vk_memory_model
,
4524 "Vulkan memory model is unsupported by this driver");
4527 vtn_fail("Unsupported memory model: %s",
4528 spirv_memorymodel_to_string(w
[2]));
4533 case SpvOpEntryPoint
:
4534 vtn_handle_entry_point(b
, w
, count
);
4538 vtn_push_value(b
, w
[1], vtn_value_type_string
)->str
=
4539 vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
4543 b
->values
[w
[1]].name
= vtn_string_literal(b
, &w
[2], count
- 2, NULL
);
4546 case SpvOpMemberName
:
4550 case SpvOpExecutionMode
:
4551 case SpvOpExecutionModeId
:
4552 case SpvOpDecorationGroup
:
4554 case SpvOpDecorateId
:
4555 case SpvOpMemberDecorate
:
4556 case SpvOpGroupDecorate
:
4557 case SpvOpGroupMemberDecorate
:
4558 case SpvOpDecorateString
:
4559 case SpvOpMemberDecorateString
:
4560 vtn_handle_decoration(b
, opcode
, w
, count
);
4563 case SpvOpExtInst
: {
4564 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
4565 if (val
->ext_handler
== vtn_handle_non_semantic_instruction
) {
4566 /* NonSemantic extended instructions are acceptable in preamble. */
4567 vtn_handle_non_semantic_instruction(b
, w
[4], w
, count
);
4570 return false; /* End of preamble. */
4575 return false; /* End of preamble */
4582 vtn_handle_execution_mode(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
4583 const struct vtn_decoration
*mode
, UNUSED
void *data
)
4585 vtn_assert(b
->entry_point
== entry_point
);
4587 switch(mode
->exec_mode
) {
4588 case SpvExecutionModeOriginUpperLeft
:
4589 case SpvExecutionModeOriginLowerLeft
:
4590 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4591 b
->shader
->info
.fs
.origin_upper_left
=
4592 (mode
->exec_mode
== SpvExecutionModeOriginUpperLeft
);
4595 case SpvExecutionModeEarlyFragmentTests
:
4596 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4597 b
->shader
->info
.fs
.early_fragment_tests
= true;
4600 case SpvExecutionModePostDepthCoverage
:
4601 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4602 b
->shader
->info
.fs
.post_depth_coverage
= true;
4605 case SpvExecutionModeInvocations
:
4606 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4607 b
->shader
->info
.gs
.invocations
= MAX2(1, mode
->operands
[0]);
4610 case SpvExecutionModeDepthReplacing
:
4611 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4612 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_ANY
;
4614 case SpvExecutionModeDepthGreater
:
4615 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4616 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_GREATER
;
4618 case SpvExecutionModeDepthLess
:
4619 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4620 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_LESS
;
4622 case SpvExecutionModeDepthUnchanged
:
4623 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4624 b
->shader
->info
.fs
.depth_layout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
4627 case SpvExecutionModeLocalSize
:
4628 vtn_assert(gl_shader_stage_is_compute(b
->shader
->info
.stage
));
4629 b
->shader
->info
.cs
.local_size
[0] = mode
->operands
[0];
4630 b
->shader
->info
.cs
.local_size
[1] = mode
->operands
[1];
4631 b
->shader
->info
.cs
.local_size
[2] = mode
->operands
[2];
4634 case SpvExecutionModeLocalSizeHint
:
4635 break; /* Nothing to do with this */
4637 case SpvExecutionModeOutputVertices
:
4638 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4639 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
4640 b
->shader
->info
.tess
.tcs_vertices_out
= mode
->operands
[0];
4642 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4643 b
->shader
->info
.gs
.vertices_out
= mode
->operands
[0];
4647 case SpvExecutionModeInputPoints
:
4648 case SpvExecutionModeInputLines
:
4649 case SpvExecutionModeInputLinesAdjacency
:
4650 case SpvExecutionModeTriangles
:
4651 case SpvExecutionModeInputTrianglesAdjacency
:
4652 case SpvExecutionModeQuads
:
4653 case SpvExecutionModeIsolines
:
4654 if (b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4655 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
4656 b
->shader
->info
.tess
.primitive_mode
=
4657 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
4659 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4660 b
->shader
->info
.gs
.vertices_in
=
4661 vertices_in_from_spv_execution_mode(b
, mode
->exec_mode
);
4662 b
->shader
->info
.gs
.input_primitive
=
4663 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
4667 case SpvExecutionModeOutputPoints
:
4668 case SpvExecutionModeOutputLineStrip
:
4669 case SpvExecutionModeOutputTriangleStrip
:
4670 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
);
4671 b
->shader
->info
.gs
.output_primitive
=
4672 gl_primitive_from_spv_execution_mode(b
, mode
->exec_mode
);
4675 case SpvExecutionModeSpacingEqual
:
4676 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4677 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4678 b
->shader
->info
.tess
.spacing
= TESS_SPACING_EQUAL
;
4680 case SpvExecutionModeSpacingFractionalEven
:
4681 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4682 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4683 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_EVEN
;
4685 case SpvExecutionModeSpacingFractionalOdd
:
4686 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4687 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4688 b
->shader
->info
.tess
.spacing
= TESS_SPACING_FRACTIONAL_ODD
;
4690 case SpvExecutionModeVertexOrderCw
:
4691 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4692 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4693 b
->shader
->info
.tess
.ccw
= false;
4695 case SpvExecutionModeVertexOrderCcw
:
4696 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4697 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4698 b
->shader
->info
.tess
.ccw
= true;
4700 case SpvExecutionModePointMode
:
4701 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_TESS_CTRL
||
4702 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
);
4703 b
->shader
->info
.tess
.point_mode
= true;
4706 case SpvExecutionModePixelCenterInteger
:
4707 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4708 b
->shader
->info
.fs
.pixel_center_integer
= true;
4711 case SpvExecutionModeXfb
:
4712 b
->shader
->info
.has_transform_feedback_varyings
= true;
4715 case SpvExecutionModeVecTypeHint
:
4718 case SpvExecutionModeContractionOff
:
4719 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
)
4720 vtn_warn("ExectionMode only allowed for CL-style kernels: %s",
4721 spirv_executionmode_to_string(mode
->exec_mode
));
4726 case SpvExecutionModeStencilRefReplacingEXT
:
4727 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4730 case SpvExecutionModeDerivativeGroupQuadsNV
:
4731 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
4732 b
->shader
->info
.cs
.derivative_group
= DERIVATIVE_GROUP_QUADS
;
4735 case SpvExecutionModeDerivativeGroupLinearNV
:
4736 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_COMPUTE
);
4737 b
->shader
->info
.cs
.derivative_group
= DERIVATIVE_GROUP_LINEAR
;
4740 case SpvExecutionModePixelInterlockOrderedEXT
:
4741 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4742 b
->shader
->info
.fs
.pixel_interlock_ordered
= true;
4745 case SpvExecutionModePixelInterlockUnorderedEXT
:
4746 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4747 b
->shader
->info
.fs
.pixel_interlock_unordered
= true;
4750 case SpvExecutionModeSampleInterlockOrderedEXT
:
4751 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4752 b
->shader
->info
.fs
.sample_interlock_ordered
= true;
4755 case SpvExecutionModeSampleInterlockUnorderedEXT
:
4756 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
4757 b
->shader
->info
.fs
.sample_interlock_unordered
= true;
4760 case SpvExecutionModeDenormPreserve
:
4761 case SpvExecutionModeDenormFlushToZero
:
4762 case SpvExecutionModeSignedZeroInfNanPreserve
:
4763 case SpvExecutionModeRoundingModeRTE
:
4764 case SpvExecutionModeRoundingModeRTZ
: {
4765 unsigned execution_mode
= 0;
4766 switch (mode
->exec_mode
) {
4767 case SpvExecutionModeDenormPreserve
:
4768 switch (mode
->operands
[0]) {
4769 case 16: execution_mode
= FLOAT_CONTROLS_DENORM_PRESERVE_FP16
; break;
4770 case 32: execution_mode
= FLOAT_CONTROLS_DENORM_PRESERVE_FP32
; break;
4771 case 64: execution_mode
= FLOAT_CONTROLS_DENORM_PRESERVE_FP64
; break;
4772 default: vtn_fail("Floating point type not supported");
4775 case SpvExecutionModeDenormFlushToZero
:
4776 switch (mode
->operands
[0]) {
4777 case 16: execution_mode
= FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16
; break;
4778 case 32: execution_mode
= FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32
; break;
4779 case 64: execution_mode
= FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP64
; break;
4780 default: vtn_fail("Floating point type not supported");
4783 case SpvExecutionModeSignedZeroInfNanPreserve
:
4784 switch (mode
->operands
[0]) {
4785 case 16: execution_mode
= FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP16
; break;
4786 case 32: execution_mode
= FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP32
; break;
4787 case 64: execution_mode
= FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP64
; break;
4788 default: vtn_fail("Floating point type not supported");
4791 case SpvExecutionModeRoundingModeRTE
:
4792 switch (mode
->operands
[0]) {
4793 case 16: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
; break;
4794 case 32: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
; break;
4795 case 64: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
; break;
4796 default: vtn_fail("Floating point type not supported");
4799 case SpvExecutionModeRoundingModeRTZ
:
4800 switch (mode
->operands
[0]) {
4801 case 16: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
; break;
4802 case 32: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
; break;
4803 case 64: execution_mode
= FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
; break;
4804 default: vtn_fail("Floating point type not supported");
4811 b
->shader
->info
.float_controls_execution_mode
|= execution_mode
;
4815 case SpvExecutionModeLocalSizeId
:
4816 case SpvExecutionModeLocalSizeHintId
:
4817 /* Handled later by vtn_handle_execution_mode_id(). */
4821 vtn_fail("Unhandled execution mode: %s (%u)",
4822 spirv_executionmode_to_string(mode
->exec_mode
),
4828 vtn_handle_execution_mode_id(struct vtn_builder
*b
, struct vtn_value
*entry_point
,
4829 const struct vtn_decoration
*mode
, UNUSED
void *data
)
4832 vtn_assert(b
->entry_point
== entry_point
);
4834 switch (mode
->exec_mode
) {
4835 case SpvExecutionModeLocalSizeId
:
4836 b
->shader
->info
.cs
.local_size
[0] = vtn_constant_uint(b
, mode
->operands
[0]);
4837 b
->shader
->info
.cs
.local_size
[1] = vtn_constant_uint(b
, mode
->operands
[1]);
4838 b
->shader
->info
.cs
.local_size
[2] = vtn_constant_uint(b
, mode
->operands
[2]);
4841 case SpvExecutionModeLocalSizeHintId
:
4842 /* Nothing to do with this hint. */
4846 /* Nothing to do. Literal execution modes already handled by
4847 * vtn_handle_execution_mode(). */
4853 vtn_handle_variable_or_type_instruction(struct vtn_builder
*b
, SpvOp opcode
,
4854 const uint32_t *w
, unsigned count
)
4856 vtn_set_instruction_result_type(b
, opcode
, w
, count
);
4860 case SpvOpSourceContinued
:
4861 case SpvOpSourceExtension
:
4862 case SpvOpExtension
:
4863 case SpvOpCapability
:
4864 case SpvOpExtInstImport
:
4865 case SpvOpMemoryModel
:
4866 case SpvOpEntryPoint
:
4867 case SpvOpExecutionMode
:
4870 case SpvOpMemberName
:
4871 case SpvOpDecorationGroup
:
4873 case SpvOpDecorateId
:
4874 case SpvOpMemberDecorate
:
4875 case SpvOpGroupDecorate
:
4876 case SpvOpGroupMemberDecorate
:
4877 case SpvOpDecorateString
:
4878 case SpvOpMemberDecorateString
:
4879 vtn_fail("Invalid opcode types and variables section");
4885 case SpvOpTypeFloat
:
4886 case SpvOpTypeVector
:
4887 case SpvOpTypeMatrix
:
4888 case SpvOpTypeImage
:
4889 case SpvOpTypeSampler
:
4890 case SpvOpTypeSampledImage
:
4891 case SpvOpTypeArray
:
4892 case SpvOpTypeRuntimeArray
:
4893 case SpvOpTypeStruct
:
4894 case SpvOpTypeOpaque
:
4895 case SpvOpTypePointer
:
4896 case SpvOpTypeForwardPointer
:
4897 case SpvOpTypeFunction
:
4898 case SpvOpTypeEvent
:
4899 case SpvOpTypeDeviceEvent
:
4900 case SpvOpTypeReserveId
:
4901 case SpvOpTypeQueue
:
4903 vtn_handle_type(b
, opcode
, w
, count
);
4906 case SpvOpConstantTrue
:
4907 case SpvOpConstantFalse
:
4909 case SpvOpConstantComposite
:
4910 case SpvOpConstantNull
:
4911 case SpvOpSpecConstantTrue
:
4912 case SpvOpSpecConstantFalse
:
4913 case SpvOpSpecConstant
:
4914 case SpvOpSpecConstantComposite
:
4915 case SpvOpSpecConstantOp
:
4916 vtn_handle_constant(b
, opcode
, w
, count
);
4921 case SpvOpConstantSampler
:
4922 vtn_handle_variables(b
, opcode
, w
, count
);
4925 case SpvOpExtInst
: {
4926 struct vtn_value
*val
= vtn_value(b
, w
[3], vtn_value_type_extension
);
4927 /* NonSemantic extended instructions are acceptable in preamble, others
4928 * will indicate the end of preamble.
4930 return val
->ext_handler
== vtn_handle_non_semantic_instruction
;
4934 return false; /* End of preamble */
4940 static struct vtn_ssa_value
*
4941 vtn_nir_select(struct vtn_builder
*b
, struct vtn_ssa_value
*src0
,
4942 struct vtn_ssa_value
*src1
, struct vtn_ssa_value
*src2
)
4944 struct vtn_ssa_value
*dest
= rzalloc(b
, struct vtn_ssa_value
);
4945 dest
->type
= src1
->type
;
4947 if (glsl_type_is_vector_or_scalar(src1
->type
)) {
4948 dest
->def
= nir_bcsel(&b
->nb
, src0
->def
, src1
->def
, src2
->def
);
4950 unsigned elems
= glsl_get_length(src1
->type
);
4952 dest
->elems
= ralloc_array(b
, struct vtn_ssa_value
*, elems
);
4953 for (unsigned i
= 0; i
< elems
; i
++) {
4954 dest
->elems
[i
] = vtn_nir_select(b
, src0
,
4955 src1
->elems
[i
], src2
->elems
[i
]);
4963 vtn_handle_select(struct vtn_builder
*b
, SpvOp opcode
,
4964 const uint32_t *w
, unsigned count
)
4966 /* Handle OpSelect up-front here because it needs to be able to handle
4967 * pointers and not just regular vectors and scalars.
4969 struct vtn_value
*res_val
= vtn_untyped_value(b
, w
[2]);
4970 struct vtn_value
*cond_val
= vtn_untyped_value(b
, w
[3]);
4971 struct vtn_value
*obj1_val
= vtn_untyped_value(b
, w
[4]);
4972 struct vtn_value
*obj2_val
= vtn_untyped_value(b
, w
[5]);
4974 vtn_fail_if(obj1_val
->type
!= res_val
->type
||
4975 obj2_val
->type
!= res_val
->type
,
4976 "Object types must match the result type in OpSelect");
4978 vtn_fail_if((cond_val
->type
->base_type
!= vtn_base_type_scalar
&&
4979 cond_val
->type
->base_type
!= vtn_base_type_vector
) ||
4980 !glsl_type_is_boolean(cond_val
->type
->type
),
4981 "OpSelect must have either a vector of booleans or "
4982 "a boolean as Condition type");
4984 vtn_fail_if(cond_val
->type
->base_type
== vtn_base_type_vector
&&
4985 (res_val
->type
->base_type
!= vtn_base_type_vector
||
4986 res_val
->type
->length
!= cond_val
->type
->length
),
4987 "When Condition type in OpSelect is a vector, the Result "
4988 "type must be a vector of the same length");
4990 switch (res_val
->type
->base_type
) {
4991 case vtn_base_type_scalar
:
4992 case vtn_base_type_vector
:
4993 case vtn_base_type_matrix
:
4994 case vtn_base_type_array
:
4995 case vtn_base_type_struct
:
4998 case vtn_base_type_pointer
:
4999 /* We need to have actual storage for pointer types. */
5000 vtn_fail_if(res_val
->type
->type
== NULL
,
5001 "Invalid pointer result type for OpSelect");
5004 vtn_fail("Result type of OpSelect must be a scalar, composite, or pointer");
5007 vtn_push_ssa_value(b
, w
[2],
5008 vtn_nir_select(b
, vtn_ssa_value(b
, w
[3]),
5009 vtn_ssa_value(b
, w
[4]),
5010 vtn_ssa_value(b
, w
[5])));
5014 vtn_handle_ptr(struct vtn_builder
*b
, SpvOp opcode
,
5015 const uint32_t *w
, unsigned count
)
5017 struct vtn_type
*type1
= vtn_get_value_type(b
, w
[3]);
5018 struct vtn_type
*type2
= vtn_get_value_type(b
, w
[4]);
5019 vtn_fail_if(type1
->base_type
!= vtn_base_type_pointer
||
5020 type2
->base_type
!= vtn_base_type_pointer
,
5021 "%s operands must have pointer types",
5022 spirv_op_to_string(opcode
));
5023 vtn_fail_if(type1
->storage_class
!= type2
->storage_class
,
5024 "%s operands must have the same storage class",
5025 spirv_op_to_string(opcode
));
5027 struct vtn_type
*vtn_type
= vtn_get_type(b
, w
[1]);
5028 const struct glsl_type
*type
= vtn_type
->type
;
5030 nir_address_format addr_format
= vtn_mode_to_address_format(
5031 b
, vtn_storage_class_to_mode(b
, type1
->storage_class
, NULL
, NULL
));
5036 case SpvOpPtrDiff
: {
5037 /* OpPtrDiff returns the difference in number of elements (not byte offset). */
5038 unsigned elem_size
, elem_align
;
5039 glsl_get_natural_size_align_bytes(type1
->deref
->type
,
5040 &elem_size
, &elem_align
);
5042 def
= nir_build_addr_isub(&b
->nb
,
5043 vtn_get_nir_ssa(b
, w
[3]),
5044 vtn_get_nir_ssa(b
, w
[4]),
5046 def
= nir_idiv(&b
->nb
, def
, nir_imm_intN_t(&b
->nb
, elem_size
, def
->bit_size
));
5047 def
= nir_i2i(&b
->nb
, def
, glsl_get_bit_size(type
));
5052 case SpvOpPtrNotEqual
: {
5053 def
= nir_build_addr_ieq(&b
->nb
,
5054 vtn_get_nir_ssa(b
, w
[3]),
5055 vtn_get_nir_ssa(b
, w
[4]),
5057 if (opcode
== SpvOpPtrNotEqual
)
5058 def
= nir_inot(&b
->nb
, def
);
5063 unreachable("Invalid ptr operation");
5066 vtn_push_nir_ssa(b
, w
[2], def
);
5070 vtn_handle_body_instruction(struct vtn_builder
*b
, SpvOp opcode
,
5071 const uint32_t *w
, unsigned count
)
5077 case SpvOpLoopMerge
:
5078 case SpvOpSelectionMerge
:
5079 /* This is handled by cfg pre-pass and walk_blocks */
5083 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
5084 val
->type
= vtn_get_type(b
, w
[1]);
5089 vtn_handle_extension(b
, opcode
, w
, count
);
5095 case SpvOpCopyMemory
:
5096 case SpvOpCopyMemorySized
:
5097 case SpvOpAccessChain
:
5098 case SpvOpPtrAccessChain
:
5099 case SpvOpInBoundsAccessChain
:
5100 case SpvOpInBoundsPtrAccessChain
:
5101 case SpvOpArrayLength
:
5102 case SpvOpConvertPtrToU
:
5103 case SpvOpConvertUToPtr
:
5104 vtn_handle_variables(b
, opcode
, w
, count
);
5107 case SpvOpFunctionCall
:
5108 vtn_handle_function_call(b
, opcode
, w
, count
);
5111 case SpvOpSampledImage
:
5113 case SpvOpImageSampleImplicitLod
:
5114 case SpvOpImageSampleExplicitLod
:
5115 case SpvOpImageSampleDrefImplicitLod
:
5116 case SpvOpImageSampleDrefExplicitLod
:
5117 case SpvOpImageSampleProjImplicitLod
:
5118 case SpvOpImageSampleProjExplicitLod
:
5119 case SpvOpImageSampleProjDrefImplicitLod
:
5120 case SpvOpImageSampleProjDrefExplicitLod
:
5121 case SpvOpImageFetch
:
5122 case SpvOpImageGather
:
5123 case SpvOpImageDrefGather
:
5124 case SpvOpImageQueryLod
:
5125 case SpvOpImageQueryLevels
:
5126 case SpvOpImageQuerySamples
:
5127 vtn_handle_texture(b
, opcode
, w
, count
);
5130 case SpvOpImageRead
:
5131 case SpvOpImageWrite
:
5132 case SpvOpImageTexelPointer
:
5133 case SpvOpImageQueryFormat
:
5134 case SpvOpImageQueryOrder
:
5135 vtn_handle_image(b
, opcode
, w
, count
);
5138 case SpvOpImageQuerySizeLod
:
5139 case SpvOpImageQuerySize
: {
5140 struct vtn_type
*image_type
= vtn_get_value_type(b
, w
[3]);
5141 vtn_assert(image_type
->base_type
== vtn_base_type_image
);
5142 if (glsl_type_is_image(image_type
->glsl_image
)) {
5143 vtn_handle_image(b
, opcode
, w
, count
);
5145 vtn_assert(glsl_type_is_sampler(image_type
->glsl_image
));
5146 vtn_handle_texture(b
, opcode
, w
, count
);
5151 case SpvOpFragmentMaskFetchAMD
:
5152 case SpvOpFragmentFetchAMD
:
5153 vtn_handle_texture(b
, opcode
, w
, count
);
5156 case SpvOpAtomicLoad
:
5157 case SpvOpAtomicExchange
:
5158 case SpvOpAtomicCompareExchange
:
5159 case SpvOpAtomicCompareExchangeWeak
:
5160 case SpvOpAtomicIIncrement
:
5161 case SpvOpAtomicIDecrement
:
5162 case SpvOpAtomicIAdd
:
5163 case SpvOpAtomicISub
:
5164 case SpvOpAtomicSMin
:
5165 case SpvOpAtomicUMin
:
5166 case SpvOpAtomicSMax
:
5167 case SpvOpAtomicUMax
:
5168 case SpvOpAtomicAnd
:
5170 case SpvOpAtomicXor
:
5171 case SpvOpAtomicFAddEXT
: {
5172 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[3]);
5173 if (pointer
->value_type
== vtn_value_type_image_pointer
) {
5174 vtn_handle_image(b
, opcode
, w
, count
);
5176 vtn_assert(pointer
->value_type
== vtn_value_type_pointer
);
5177 vtn_handle_atomics(b
, opcode
, w
, count
);
5182 case SpvOpAtomicStore
: {
5183 struct vtn_value
*pointer
= vtn_untyped_value(b
, w
[1]);
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
);
5194 vtn_handle_select(b
, opcode
, w
, count
);
5202 case SpvOpConvertFToU
:
5203 case SpvOpConvertFToS
:
5204 case SpvOpConvertSToF
:
5205 case SpvOpConvertUToF
:
5209 case SpvOpQuantizeToF16
:
5210 case SpvOpPtrCastToGeneric
:
5211 case SpvOpGenericCastToPtr
:
5216 case SpvOpSignBitSet
:
5217 case SpvOpLessOrGreater
:
5219 case SpvOpUnordered
:
5234 case SpvOpVectorTimesScalar
:
5236 case SpvOpIAddCarry
:
5237 case SpvOpISubBorrow
:
5238 case SpvOpUMulExtended
:
5239 case SpvOpSMulExtended
:
5240 case SpvOpShiftRightLogical
:
5241 case SpvOpShiftRightArithmetic
:
5242 case SpvOpShiftLeftLogical
:
5243 case SpvOpLogicalEqual
:
5244 case SpvOpLogicalNotEqual
:
5245 case SpvOpLogicalOr
:
5246 case SpvOpLogicalAnd
:
5247 case SpvOpLogicalNot
:
5248 case SpvOpBitwiseOr
:
5249 case SpvOpBitwiseXor
:
5250 case SpvOpBitwiseAnd
:
5252 case SpvOpFOrdEqual
:
5253 case SpvOpFUnordEqual
:
5254 case SpvOpINotEqual
:
5255 case SpvOpFOrdNotEqual
:
5256 case SpvOpFUnordNotEqual
:
5257 case SpvOpULessThan
:
5258 case SpvOpSLessThan
:
5259 case SpvOpFOrdLessThan
:
5260 case SpvOpFUnordLessThan
:
5261 case SpvOpUGreaterThan
:
5262 case SpvOpSGreaterThan
:
5263 case SpvOpFOrdGreaterThan
:
5264 case SpvOpFUnordGreaterThan
:
5265 case SpvOpULessThanEqual
:
5266 case SpvOpSLessThanEqual
:
5267 case SpvOpFOrdLessThanEqual
:
5268 case SpvOpFUnordLessThanEqual
:
5269 case SpvOpUGreaterThanEqual
:
5270 case SpvOpSGreaterThanEqual
:
5271 case SpvOpFOrdGreaterThanEqual
:
5272 case SpvOpFUnordGreaterThanEqual
:
5278 case SpvOpFwidthFine
:
5279 case SpvOpDPdxCoarse
:
5280 case SpvOpDPdyCoarse
:
5281 case SpvOpFwidthCoarse
:
5282 case SpvOpBitFieldInsert
:
5283 case SpvOpBitFieldSExtract
:
5284 case SpvOpBitFieldUExtract
:
5285 case SpvOpBitReverse
:
5287 case SpvOpTranspose
:
5288 case SpvOpOuterProduct
:
5289 case SpvOpMatrixTimesScalar
:
5290 case SpvOpVectorTimesMatrix
:
5291 case SpvOpMatrixTimesVector
:
5292 case SpvOpMatrixTimesMatrix
:
5293 case SpvOpUCountLeadingZerosINTEL
:
5294 case SpvOpUCountTrailingZerosINTEL
:
5295 case SpvOpAbsISubINTEL
:
5296 case SpvOpAbsUSubINTEL
:
5297 case SpvOpIAddSatINTEL
:
5298 case SpvOpUAddSatINTEL
:
5299 case SpvOpIAverageINTEL
:
5300 case SpvOpUAverageINTEL
:
5301 case SpvOpIAverageRoundedINTEL
:
5302 case SpvOpUAverageRoundedINTEL
:
5303 case SpvOpISubSatINTEL
:
5304 case SpvOpUSubSatINTEL
:
5305 case SpvOpIMul32x16INTEL
:
5306 case SpvOpUMul32x16INTEL
:
5307 vtn_handle_alu(b
, opcode
, w
, count
);
5311 vtn_handle_bitcast(b
, w
, count
);
5314 case SpvOpVectorExtractDynamic
:
5315 case SpvOpVectorInsertDynamic
:
5316 case SpvOpVectorShuffle
:
5317 case SpvOpCompositeConstruct
:
5318 case SpvOpCompositeExtract
:
5319 case SpvOpCompositeInsert
:
5320 case SpvOpCopyLogical
:
5321 case SpvOpCopyObject
:
5322 vtn_handle_composite(b
, opcode
, w
, count
);
5325 case SpvOpEmitVertex
:
5326 case SpvOpEndPrimitive
:
5327 case SpvOpEmitStreamVertex
:
5328 case SpvOpEndStreamPrimitive
:
5329 case SpvOpControlBarrier
:
5330 case SpvOpMemoryBarrier
:
5331 vtn_handle_barrier(b
, opcode
, w
, count
);
5334 case SpvOpGroupNonUniformElect
:
5335 case SpvOpGroupNonUniformAll
:
5336 case SpvOpGroupNonUniformAny
:
5337 case SpvOpGroupNonUniformAllEqual
:
5338 case SpvOpGroupNonUniformBroadcast
:
5339 case SpvOpGroupNonUniformBroadcastFirst
:
5340 case SpvOpGroupNonUniformBallot
:
5341 case SpvOpGroupNonUniformInverseBallot
:
5342 case SpvOpGroupNonUniformBallotBitExtract
:
5343 case SpvOpGroupNonUniformBallotBitCount
:
5344 case SpvOpGroupNonUniformBallotFindLSB
:
5345 case SpvOpGroupNonUniformBallotFindMSB
:
5346 case SpvOpGroupNonUniformShuffle
:
5347 case SpvOpGroupNonUniformShuffleXor
:
5348 case SpvOpGroupNonUniformShuffleUp
:
5349 case SpvOpGroupNonUniformShuffleDown
:
5350 case SpvOpGroupNonUniformIAdd
:
5351 case SpvOpGroupNonUniformFAdd
:
5352 case SpvOpGroupNonUniformIMul
:
5353 case SpvOpGroupNonUniformFMul
:
5354 case SpvOpGroupNonUniformSMin
:
5355 case SpvOpGroupNonUniformUMin
:
5356 case SpvOpGroupNonUniformFMin
:
5357 case SpvOpGroupNonUniformSMax
:
5358 case SpvOpGroupNonUniformUMax
:
5359 case SpvOpGroupNonUniformFMax
:
5360 case SpvOpGroupNonUniformBitwiseAnd
:
5361 case SpvOpGroupNonUniformBitwiseOr
:
5362 case SpvOpGroupNonUniformBitwiseXor
:
5363 case SpvOpGroupNonUniformLogicalAnd
:
5364 case SpvOpGroupNonUniformLogicalOr
:
5365 case SpvOpGroupNonUniformLogicalXor
:
5366 case SpvOpGroupNonUniformQuadBroadcast
:
5367 case SpvOpGroupNonUniformQuadSwap
:
5370 case SpvOpGroupBroadcast
:
5371 case SpvOpGroupIAdd
:
5372 case SpvOpGroupFAdd
:
5373 case SpvOpGroupFMin
:
5374 case SpvOpGroupUMin
:
5375 case SpvOpGroupSMin
:
5376 case SpvOpGroupFMax
:
5377 case SpvOpGroupUMax
:
5378 case SpvOpGroupSMax
:
5379 case SpvOpSubgroupBallotKHR
:
5380 case SpvOpSubgroupFirstInvocationKHR
:
5381 case SpvOpSubgroupReadInvocationKHR
:
5382 case SpvOpSubgroupAllKHR
:
5383 case SpvOpSubgroupAnyKHR
:
5384 case SpvOpSubgroupAllEqualKHR
:
5385 case SpvOpGroupIAddNonUniformAMD
:
5386 case SpvOpGroupFAddNonUniformAMD
:
5387 case SpvOpGroupFMinNonUniformAMD
:
5388 case SpvOpGroupUMinNonUniformAMD
:
5389 case SpvOpGroupSMinNonUniformAMD
:
5390 case SpvOpGroupFMaxNonUniformAMD
:
5391 case SpvOpGroupUMaxNonUniformAMD
:
5392 case SpvOpGroupSMaxNonUniformAMD
:
5393 vtn_handle_subgroup(b
, opcode
, w
, count
);
5398 case SpvOpPtrNotEqual
:
5399 vtn_handle_ptr(b
, opcode
, w
, count
);
5402 case SpvOpBeginInvocationInterlockEXT
:
5403 vtn_emit_barrier(b
, nir_intrinsic_begin_invocation_interlock
);
5406 case SpvOpEndInvocationInterlockEXT
:
5407 vtn_emit_barrier(b
, nir_intrinsic_end_invocation_interlock
);
5410 case SpvOpDemoteToHelperInvocationEXT
: {
5411 nir_intrinsic_instr
*intrin
=
5412 nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_demote
);
5413 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
5417 case SpvOpIsHelperInvocationEXT
: {
5418 nir_intrinsic_instr
*intrin
=
5419 nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_is_helper_invocation
);
5420 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
, 1, 1, NULL
);
5421 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
5423 vtn_push_nir_ssa(b
, w
[2], &intrin
->dest
.ssa
);
5427 case SpvOpReadClockKHR
: {
5428 SpvScope scope
= vtn_constant_uint(b
, w
[3]);
5429 nir_scope nir_scope
;
5432 case SpvScopeDevice
:
5433 nir_scope
= NIR_SCOPE_DEVICE
;
5435 case SpvScopeSubgroup
:
5436 nir_scope
= NIR_SCOPE_SUBGROUP
;
5439 vtn_fail("invalid read clock scope");
5442 /* Operation supports two result types: uvec2 and uint64_t. The NIR
5443 * intrinsic gives uvec2, so pack the result for the other case.
5445 nir_intrinsic_instr
*intrin
=
5446 nir_intrinsic_instr_create(b
->nb
.shader
, nir_intrinsic_shader_clock
);
5447 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
, 2, 32, NULL
);
5448 nir_intrinsic_set_memory_scope(intrin
, nir_scope
);
5449 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
5451 struct vtn_type
*type
= vtn_get_type(b
, w
[1]);
5452 const struct glsl_type
*dest_type
= type
->type
;
5453 nir_ssa_def
*result
;
5455 if (glsl_type_is_vector(dest_type
)) {
5456 assert(dest_type
== glsl_vector_type(GLSL_TYPE_UINT
, 2));
5457 result
= &intrin
->dest
.ssa
;
5459 assert(glsl_type_is_scalar(dest_type
));
5460 assert(glsl_get_base_type(dest_type
) == GLSL_TYPE_UINT64
);
5461 result
= nir_pack_64_2x32(&b
->nb
, &intrin
->dest
.ssa
);
5464 vtn_push_nir_ssa(b
, w
[2], result
);
5468 case SpvOpLifetimeStart
:
5469 case SpvOpLifetimeStop
:
5473 vtn_fail_with_opcode("Unhandled opcode", opcode
);
5480 vtn_create_builder(const uint32_t *words
, size_t word_count
,
5481 gl_shader_stage stage
, const char *entry_point_name
,
5482 const struct spirv_to_nir_options
*options
)
5484 /* Initialize the vtn_builder object */
5485 struct vtn_builder
*b
= rzalloc(NULL
, struct vtn_builder
);
5486 struct spirv_to_nir_options
*dup_options
=
5487 ralloc(b
, struct spirv_to_nir_options
);
5488 *dup_options
= *options
;
5491 b
->spirv_word_count
= word_count
;
5495 list_inithead(&b
->functions
);
5496 b
->entry_point_stage
= stage
;
5497 b
->entry_point_name
= entry_point_name
;
5498 b
->options
= dup_options
;
5501 * Handle the SPIR-V header (first 5 dwords).
5502 * Can't use vtx_assert() as the setjmp(3) target isn't initialized yet.
5504 if (word_count
<= 5)
5507 if (words
[0] != SpvMagicNumber
) {
5508 vtn_err("words[0] was 0x%x, want 0x%x", words
[0], SpvMagicNumber
);
5511 if (words
[1] < 0x10000) {
5512 vtn_err("words[1] was 0x%x, want >= 0x10000", words
[1]);
5516 uint16_t generator_id
= words
[2] >> 16;
5517 uint16_t generator_version
= words
[2];
5519 /* In GLSLang commit 8297936dd6eb3, their handling of barrier() was fixed
5520 * to provide correct memory semantics on compute shader barrier()
5521 * commands. Prior to that, we need to fix them up ourselves. This
5522 * GLSLang fix caused them to bump to generator version 3.
5524 b
->wa_glslang_cs_barrier
= (generator_id
== 8 && generator_version
< 3);
5526 /* words[2] == generator magic */
5527 unsigned value_id_bound
= words
[3];
5528 if (words
[4] != 0) {
5529 vtn_err("words[4] was %u, want 0", words
[4]);
5533 b
->value_id_bound
= value_id_bound
;
5534 b
->values
= rzalloc_array(b
, struct vtn_value
, value_id_bound
);
5542 static nir_function
*
5543 vtn_emit_kernel_entry_point_wrapper(struct vtn_builder
*b
,
5544 nir_function
*entry_point
)
5546 vtn_assert(entry_point
== b
->entry_point
->func
->impl
->function
);
5547 vtn_fail_if(!entry_point
->name
, "entry points are required to have a name");
5548 const char *func_name
=
5549 ralloc_asprintf(b
->shader
, "__wrapped_%s", entry_point
->name
);
5551 vtn_assert(b
->shader
->info
.stage
== MESA_SHADER_KERNEL
);
5553 nir_function
*main_entry_point
= nir_function_create(b
->shader
, func_name
);
5554 main_entry_point
->impl
= nir_function_impl_create(main_entry_point
);
5555 nir_builder_init(&b
->nb
, main_entry_point
->impl
);
5556 b
->nb
.cursor
= nir_after_cf_list(&main_entry_point
->impl
->body
);
5557 b
->func_param_idx
= 0;
5559 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, entry_point
);
5561 for (unsigned i
= 0; i
< entry_point
->num_params
; ++i
) {
5562 struct vtn_type
*param_type
= b
->entry_point
->func
->type
->params
[i
];
5564 /* consider all pointers to function memory to be parameters passed
5567 bool is_by_val
= param_type
->base_type
== vtn_base_type_pointer
&&
5568 param_type
->storage_class
== SpvStorageClassFunction
;
5570 /* input variable */
5571 nir_variable
*in_var
= rzalloc(b
->nb
.shader
, nir_variable
);
5572 in_var
->data
.mode
= nir_var_uniform
;
5573 in_var
->data
.read_only
= true;
5574 in_var
->data
.location
= i
;
5575 if (param_type
->base_type
== vtn_base_type_image
) {
5576 in_var
->data
.access
= 0;
5577 if (param_type
->access_qualifier
& SpvAccessQualifierReadOnly
)
5578 in_var
->data
.access
|= ACCESS_NON_WRITEABLE
;
5579 if (param_type
->access_qualifier
& SpvAccessQualifierWriteOnly
)
5580 in_var
->data
.access
|= ACCESS_NON_READABLE
;
5584 in_var
->type
= param_type
->deref
->type
;
5585 else if (param_type
->base_type
== vtn_base_type_image
)
5586 in_var
->type
= param_type
->glsl_image
;
5587 else if (param_type
->base_type
== vtn_base_type_sampler
)
5588 in_var
->type
= glsl_bare_sampler_type();
5590 in_var
->type
= param_type
->type
;
5592 nir_shader_add_variable(b
->nb
.shader
, in_var
);
5594 /* we have to copy the entire variable into function memory */
5596 nir_variable
*copy_var
=
5597 nir_local_variable_create(main_entry_point
->impl
, in_var
->type
,
5599 nir_copy_var(&b
->nb
, copy_var
, in_var
);
5601 nir_src_for_ssa(&nir_build_deref_var(&b
->nb
, copy_var
)->dest
.ssa
);
5602 } else if (param_type
->base_type
== vtn_base_type_image
||
5603 param_type
->base_type
== vtn_base_type_sampler
) {
5604 /* Don't load the var, just pass a deref of it */
5605 call
->params
[i
] = nir_src_for_ssa(&nir_build_deref_var(&b
->nb
, in_var
)->dest
.ssa
);
5607 call
->params
[i
] = nir_src_for_ssa(nir_load_var(&b
->nb
, in_var
));
5611 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
5613 return main_entry_point
;
5617 spirv_to_nir(const uint32_t *words
, size_t word_count
,
5618 struct nir_spirv_specialization
*spec
, unsigned num_spec
,
5619 gl_shader_stage stage
, const char *entry_point_name
,
5620 const struct spirv_to_nir_options
*options
,
5621 const nir_shader_compiler_options
*nir_options
)
5624 const uint32_t *word_end
= words
+ word_count
;
5626 struct vtn_builder
*b
= vtn_create_builder(words
, word_count
,
5627 stage
, entry_point_name
,
5633 /* See also _vtn_fail() */
5634 if (setjmp(b
->fail_jump
)) {
5639 /* Skip the SPIR-V header, handled at vtn_create_builder */
5642 b
->shader
= nir_shader_create(b
, stage
, nir_options
, NULL
);
5644 /* Handle all the preamble instructions */
5645 words
= vtn_foreach_instruction(b
, words
, word_end
,
5646 vtn_handle_preamble_instruction
);
5648 if (b
->entry_point
== NULL
) {
5649 vtn_fail("Entry point not found for %s shader \"%s\"",
5650 _mesa_shader_stage_to_string(stage
), entry_point_name
);
5655 /* Ensure a sane address mode is being used for function temps */
5656 assert(nir_address_format_bit_size(b
->options
->temp_addr_format
) == nir_get_ptr_bitsize(b
->shader
));
5657 assert(nir_address_format_num_components(b
->options
->temp_addr_format
) == 1);
5659 /* Set shader info defaults */
5660 if (stage
== MESA_SHADER_GEOMETRY
)
5661 b
->shader
->info
.gs
.invocations
= 1;
5663 /* Parse execution modes. */
5664 vtn_foreach_execution_mode(b
, b
->entry_point
,
5665 vtn_handle_execution_mode
, NULL
);
5667 b
->specializations
= spec
;
5668 b
->num_specializations
= num_spec
;
5670 /* Handle all variable, type, and constant instructions */
5671 words
= vtn_foreach_instruction(b
, words
, word_end
,
5672 vtn_handle_variable_or_type_instruction
);
5674 /* Parse execution modes that depend on IDs. Must happen after we have
5677 vtn_foreach_execution_mode(b
, b
->entry_point
,
5678 vtn_handle_execution_mode_id
, NULL
);
5680 if (b
->workgroup_size_builtin
) {
5681 vtn_assert(b
->workgroup_size_builtin
->type
->type
==
5682 glsl_vector_type(GLSL_TYPE_UINT
, 3));
5684 nir_const_value
*const_size
=
5685 b
->workgroup_size_builtin
->constant
->values
;
5687 b
->shader
->info
.cs
.local_size
[0] = const_size
[0].u32
;
5688 b
->shader
->info
.cs
.local_size
[1] = const_size
[1].u32
;
5689 b
->shader
->info
.cs
.local_size
[2] = const_size
[2].u32
;
5692 /* Set types on all vtn_values */
5693 vtn_foreach_instruction(b
, words
, word_end
, vtn_set_instruction_result_type
);
5695 vtn_build_cfg(b
, words
, word_end
);
5697 assert(b
->entry_point
->value_type
== vtn_value_type_function
);
5698 b
->entry_point
->func
->referenced
= true;
5703 vtn_foreach_cf_node(node
, &b
->functions
) {
5704 struct vtn_function
*func
= vtn_cf_node_as_function(node
);
5705 if (func
->referenced
&& !func
->emitted
) {
5706 b
->const_table
= _mesa_pointer_hash_table_create(b
);
5708 vtn_function_emit(b
, func
, vtn_handle_body_instruction
);
5714 vtn_assert(b
->entry_point
->value_type
== vtn_value_type_function
);
5715 nir_function
*entry_point
= b
->entry_point
->func
->impl
->function
;
5716 vtn_assert(entry_point
);
5718 /* post process entry_points with input params */
5719 if (entry_point
->num_params
&& b
->shader
->info
.stage
== MESA_SHADER_KERNEL
)
5720 entry_point
= vtn_emit_kernel_entry_point_wrapper(b
, entry_point
);
5722 /* structurize the CFG */
5723 nir_lower_goto_ifs(b
->shader
);
5725 entry_point
->is_entrypoint
= true;
5727 /* When multiple shader stages exist in the same SPIR-V module, we
5728 * generate input and output variables for every stage, in the same
5729 * NIR program. These dead variables can be invalid NIR. For example,
5730 * TCS outputs must be per-vertex arrays (or decorated 'patch'), while
5731 * VS output variables wouldn't be.
5733 * To ensure we have valid NIR, we eliminate any dead inputs and outputs
5734 * right away. In order to do so, we must lower any constant initializers
5735 * on outputs so nir_remove_dead_variables sees that they're written to.
5737 nir_lower_variable_initializers(b
->shader
, nir_var_shader_out
);
5738 nir_remove_dead_variables(b
->shader
,
5739 nir_var_shader_in
| nir_var_shader_out
, NULL
);
5741 /* We sometimes generate bogus derefs that, while never used, give the
5742 * validator a bit of heartburn. Run dead code to get rid of them.
5744 nir_opt_dce(b
->shader
);
5746 /* Unparent the shader from the vtn_builder before we delete the builder */
5747 ralloc_steal(NULL
, b
->shader
);
5749 nir_shader
*shader
= b
->shader
;