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 "spirv_info.h"
31 static struct vtn_access_chain
*
32 vtn_access_chain_extend(struct vtn_builder
*b
, struct vtn_access_chain
*old
,
35 struct vtn_access_chain
*chain
;
37 unsigned new_len
= old
->length
+ new_ids
;
38 /* TODO: don't use rzalloc */
39 chain
= rzalloc_size(b
, sizeof(*chain
) + new_len
* sizeof(chain
->link
[0]));
41 chain
->var
= old
->var
;
42 chain
->length
= new_len
;
44 for (unsigned i
= 0; i
< old
->length
; i
++)
45 chain
->link
[i
] = old
->link
[i
];
51 vtn_access_link_as_ssa(struct vtn_builder
*b
, struct vtn_access_link link
,
55 if (link
.mode
== vtn_access_mode_literal
) {
56 return nir_imm_int(&b
->nb
, link
.id
* stride
);
57 } else if (stride
== 1) {
58 return vtn_ssa_value(b
, link
.id
)->def
;
60 return nir_imul(&b
->nb
, vtn_ssa_value(b
, link
.id
)->def
,
61 nir_imm_int(&b
->nb
, stride
));
65 static struct vtn_type
*
66 vtn_access_chain_tail_type(struct vtn_builder
*b
,
67 struct vtn_access_chain
*chain
)
69 struct vtn_type
*type
= chain
->var
->type
;
70 for (unsigned i
= 0; i
< chain
->length
; i
++) {
71 if (glsl_type_is_struct(type
->type
)) {
72 assert(chain
->link
[i
].mode
== vtn_access_mode_literal
);
73 type
= type
->members
[chain
->link
[i
].id
];
75 type
= type
->array_element
;
81 /* Crawls a chain of array derefs and rewrites the types so that the
82 * lengths stay the same but the terminal type is the one given by
83 * tail_type. This is useful for split structures.
86 rewrite_deref_types(nir_deref
*deref
, const struct glsl_type
*type
)
90 assert(deref
->child
->deref_type
== nir_deref_type_array
);
91 assert(glsl_type_is_array(deref
->type
));
92 rewrite_deref_types(deref
->child
, glsl_get_array_element(type
));
97 vtn_access_chain_to_deref(struct vtn_builder
*b
, struct vtn_access_chain
*chain
)
99 nir_deref_var
*deref_var
;
100 if (chain
->var
->var
) {
101 deref_var
= nir_deref_var_create(b
, chain
->var
->var
);
103 assert(chain
->var
->members
);
104 /* Create the deref_var manually. It will get filled out later. */
105 deref_var
= rzalloc(b
, nir_deref_var
);
106 deref_var
->deref
.deref_type
= nir_deref_type_var
;
109 struct vtn_type
*deref_type
= chain
->var
->type
;
110 nir_deref
*tail
= &deref_var
->deref
;
111 nir_variable
**members
= chain
->var
->members
;
113 for (unsigned i
= 0; i
< chain
->length
; i
++) {
114 enum glsl_base_type base_type
= glsl_get_base_type(deref_type
->type
);
118 case GLSL_TYPE_UINT64
:
119 case GLSL_TYPE_INT64
:
120 case GLSL_TYPE_FLOAT
:
121 case GLSL_TYPE_DOUBLE
:
123 case GLSL_TYPE_ARRAY
: {
124 deref_type
= deref_type
->array_element
;
126 nir_deref_array
*deref_arr
= nir_deref_array_create(b
);
127 deref_arr
->deref
.type
= deref_type
->type
;
129 if (chain
->link
[i
].mode
== vtn_access_mode_literal
) {
130 deref_arr
->deref_array_type
= nir_deref_array_type_direct
;
131 deref_arr
->base_offset
= chain
->link
[i
].id
;
133 assert(chain
->link
[i
].mode
== vtn_access_mode_id
);
134 deref_arr
->deref_array_type
= nir_deref_array_type_indirect
;
135 deref_arr
->base_offset
= 0;
136 deref_arr
->indirect
=
137 nir_src_for_ssa(vtn_ssa_value(b
, chain
->link
[i
].id
)->def
);
139 tail
->child
= &deref_arr
->deref
;
144 case GLSL_TYPE_STRUCT
: {
145 assert(chain
->link
[i
].mode
== vtn_access_mode_literal
);
146 unsigned idx
= chain
->link
[i
].id
;
147 deref_type
= deref_type
->members
[idx
];
149 /* This is a pre-split structure. */
150 deref_var
->var
= members
[idx
];
151 rewrite_deref_types(&deref_var
->deref
, members
[idx
]->type
);
152 assert(tail
->type
== deref_type
->type
);
155 nir_deref_struct
*deref_struct
= nir_deref_struct_create(b
, idx
);
156 deref_struct
->deref
.type
= deref_type
->type
;
157 tail
->child
= &deref_struct
->deref
;
163 unreachable("Invalid type for deref");
167 assert(members
== NULL
);
172 _vtn_local_load_store(struct vtn_builder
*b
, bool load
, nir_deref_var
*deref
,
173 nir_deref
*tail
, struct vtn_ssa_value
*inout
)
175 /* The deref tail may contain a deref to select a component of a vector (in
176 * other words, it might not be an actual tail) so we have to save it away
177 * here since we overwrite it later.
179 nir_deref
*old_child
= tail
->child
;
181 if (glsl_type_is_vector_or_scalar(tail
->type
)) {
182 /* Terminate the deref chain in case there is one more link to pick
183 * off a component of the vector.
187 nir_intrinsic_op op
= load
? nir_intrinsic_load_var
:
188 nir_intrinsic_store_var
;
190 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
191 intrin
->variables
[0] = nir_deref_var_clone(deref
, intrin
);
192 intrin
->num_components
= glsl_get_vector_elements(tail
->type
);
195 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
,
196 intrin
->num_components
,
197 glsl_get_bit_size(tail
->type
),
199 inout
->def
= &intrin
->dest
.ssa
;
201 nir_intrinsic_set_write_mask(intrin
, (1 << intrin
->num_components
) - 1);
202 intrin
->src
[0] = nir_src_for_ssa(inout
->def
);
205 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
206 } else if (glsl_get_base_type(tail
->type
) == GLSL_TYPE_ARRAY
||
207 glsl_type_is_matrix(tail
->type
)) {
208 unsigned elems
= glsl_get_length(tail
->type
);
209 nir_deref_array
*deref_arr
= nir_deref_array_create(b
);
210 deref_arr
->deref_array_type
= nir_deref_array_type_direct
;
211 deref_arr
->deref
.type
= glsl_get_array_element(tail
->type
);
212 tail
->child
= &deref_arr
->deref
;
213 for (unsigned i
= 0; i
< elems
; i
++) {
214 deref_arr
->base_offset
= i
;
215 _vtn_local_load_store(b
, load
, deref
, tail
->child
, inout
->elems
[i
]);
218 assert(glsl_get_base_type(tail
->type
) == GLSL_TYPE_STRUCT
);
219 unsigned elems
= glsl_get_length(tail
->type
);
220 nir_deref_struct
*deref_struct
= nir_deref_struct_create(b
, 0);
221 tail
->child
= &deref_struct
->deref
;
222 for (unsigned i
= 0; i
< elems
; i
++) {
223 deref_struct
->index
= i
;
224 deref_struct
->deref
.type
= glsl_get_struct_field(tail
->type
, i
);
225 _vtn_local_load_store(b
, load
, deref
, tail
->child
, inout
->elems
[i
]);
229 tail
->child
= old_child
;
233 vtn_nir_deref(struct vtn_builder
*b
, uint32_t id
)
235 struct vtn_access_chain
*chain
=
236 vtn_value(b
, id
, vtn_value_type_access_chain
)->access_chain
;
238 return vtn_access_chain_to_deref(b
, chain
);
242 * Gets the NIR-level deref tail, which may have as a child an array deref
243 * selecting which component due to OpAccessChain supporting per-component
244 * indexing in SPIR-V.
247 get_deref_tail(nir_deref_var
*deref
)
249 nir_deref
*cur
= &deref
->deref
;
250 while (!glsl_type_is_vector_or_scalar(cur
->type
) && cur
->child
)
256 struct vtn_ssa_value
*
257 vtn_local_load(struct vtn_builder
*b
, nir_deref_var
*src
)
259 nir_deref
*src_tail
= get_deref_tail(src
);
260 struct vtn_ssa_value
*val
= vtn_create_ssa_value(b
, src_tail
->type
);
261 _vtn_local_load_store(b
, true, src
, src_tail
, val
);
263 if (src_tail
->child
) {
264 nir_deref_array
*vec_deref
= nir_deref_as_array(src_tail
->child
);
265 assert(vec_deref
->deref
.child
== NULL
);
266 val
->type
= vec_deref
->deref
.type
;
267 if (vec_deref
->deref_array_type
== nir_deref_array_type_direct
)
268 val
->def
= vtn_vector_extract(b
, val
->def
, vec_deref
->base_offset
);
270 val
->def
= vtn_vector_extract_dynamic(b
, val
->def
,
271 vec_deref
->indirect
.ssa
);
278 vtn_local_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
281 nir_deref
*dest_tail
= get_deref_tail(dest
);
283 if (dest_tail
->child
) {
284 struct vtn_ssa_value
*val
= vtn_create_ssa_value(b
, dest_tail
->type
);
285 _vtn_local_load_store(b
, true, dest
, dest_tail
, val
);
286 nir_deref_array
*deref
= nir_deref_as_array(dest_tail
->child
);
287 assert(deref
->deref
.child
== NULL
);
288 if (deref
->deref_array_type
== nir_deref_array_type_direct
)
289 val
->def
= vtn_vector_insert(b
, val
->def
, src
->def
,
292 val
->def
= vtn_vector_insert_dynamic(b
, val
->def
, src
->def
,
293 deref
->indirect
.ssa
);
294 _vtn_local_load_store(b
, false, dest
, dest_tail
, val
);
296 _vtn_local_load_store(b
, false, dest
, dest_tail
, src
);
301 get_vulkan_resource_index(struct vtn_builder
*b
, struct vtn_access_chain
*chain
,
302 struct vtn_type
**type
, unsigned *chain_idx
)
304 /* Push constants have no explicit binding */
305 if (chain
->var
->mode
== vtn_variable_mode_push_constant
) {
307 *type
= chain
->var
->type
;
311 nir_ssa_def
*array_index
;
312 if (glsl_type_is_array(chain
->var
->type
->type
)) {
313 assert(chain
->length
> 0);
314 array_index
= vtn_access_link_as_ssa(b
, chain
->link
[0], 1);
316 *type
= chain
->var
->type
->array_element
;
318 array_index
= nir_imm_int(&b
->nb
, 0);
320 *type
= chain
->var
->type
;
323 nir_intrinsic_instr
*instr
=
324 nir_intrinsic_instr_create(b
->nb
.shader
,
325 nir_intrinsic_vulkan_resource_index
);
326 instr
->src
[0] = nir_src_for_ssa(array_index
);
327 nir_intrinsic_set_desc_set(instr
, chain
->var
->descriptor_set
);
328 nir_intrinsic_set_binding(instr
, chain
->var
->binding
);
330 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
331 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
333 return &instr
->dest
.ssa
;
337 vtn_access_chain_to_offset(struct vtn_builder
*b
,
338 struct vtn_access_chain
*chain
,
339 nir_ssa_def
**index_out
, struct vtn_type
**type_out
,
340 unsigned *end_idx_out
, bool stop_at_matrix
)
343 struct vtn_type
*type
;
344 *index_out
= get_vulkan_resource_index(b
, chain
, &type
, &idx
);
346 nir_ssa_def
*offset
= nir_imm_int(&b
->nb
, 0);
347 for (; idx
< chain
->length
; idx
++) {
348 enum glsl_base_type base_type
= glsl_get_base_type(type
->type
);
352 case GLSL_TYPE_UINT64
:
353 case GLSL_TYPE_INT64
:
354 case GLSL_TYPE_FLOAT
:
355 case GLSL_TYPE_DOUBLE
:
357 /* Some users may not want matrix or vector derefs */
362 case GLSL_TYPE_ARRAY
:
363 offset
= nir_iadd(&b
->nb
, offset
,
364 vtn_access_link_as_ssa(b
, chain
->link
[idx
],
367 type
= type
->array_element
;
370 case GLSL_TYPE_STRUCT
: {
371 assert(chain
->link
[idx
].mode
== vtn_access_mode_literal
);
372 unsigned member
= chain
->link
[idx
].id
;
373 offset
= nir_iadd(&b
->nb
, offset
,
374 nir_imm_int(&b
->nb
, type
->offsets
[member
]));
375 type
= type
->members
[member
];
380 unreachable("Invalid type for deref");
392 /* Tries to compute the size of an interface block based on the strides and
393 * offsets that are provided to us in the SPIR-V source.
396 vtn_type_block_size(struct vtn_type
*type
)
398 enum glsl_base_type base_type
= glsl_get_base_type(type
->type
);
402 case GLSL_TYPE_UINT64
:
403 case GLSL_TYPE_INT64
:
404 case GLSL_TYPE_FLOAT
:
406 case GLSL_TYPE_DOUBLE
: {
407 unsigned cols
= type
->row_major
? glsl_get_vector_elements(type
->type
) :
408 glsl_get_matrix_columns(type
->type
);
410 assert(type
->stride
> 0);
411 return type
->stride
* cols
;
412 } else if (base_type
== GLSL_TYPE_DOUBLE
||
413 base_type
== GLSL_TYPE_UINT64
||
414 base_type
== GLSL_TYPE_INT64
) {
415 return glsl_get_vector_elements(type
->type
) * 8;
417 return glsl_get_vector_elements(type
->type
) * 4;
421 case GLSL_TYPE_STRUCT
:
422 case GLSL_TYPE_INTERFACE
: {
424 unsigned num_fields
= glsl_get_length(type
->type
);
425 for (unsigned f
= 0; f
< num_fields
; f
++) {
426 unsigned field_end
= type
->offsets
[f
] +
427 vtn_type_block_size(type
->members
[f
]);
428 size
= MAX2(size
, field_end
);
433 case GLSL_TYPE_ARRAY
:
434 assert(type
->stride
> 0);
435 assert(glsl_get_length(type
->type
) > 0);
436 return type
->stride
* glsl_get_length(type
->type
);
439 assert(!"Invalid block type");
445 vtn_access_chain_get_offset_size(struct vtn_access_chain
*chain
,
446 unsigned *access_offset
,
447 unsigned *access_size
)
449 /* Only valid for push constants accesses now. */
450 assert(chain
->var
->mode
== vtn_variable_mode_push_constant
);
452 struct vtn_type
*type
= chain
->var
->type
;
456 for (unsigned i
= 0; i
< chain
->length
; i
++) {
457 if (chain
->link
[i
].mode
!= vtn_access_mode_literal
)
460 if (glsl_type_is_struct(type
->type
)) {
461 *access_offset
+= type
->offsets
[chain
->link
[i
].id
];
462 type
= type
->members
[chain
->link
[i
].id
];
464 *access_offset
+= type
->stride
* chain
->link
[i
].id
;
465 type
= type
->array_element
;
469 *access_size
= vtn_type_block_size(type
);
473 _vtn_load_store_tail(struct vtn_builder
*b
, nir_intrinsic_op op
, bool load
,
474 nir_ssa_def
*index
, nir_ssa_def
*offset
,
475 unsigned access_offset
, unsigned access_size
,
476 struct vtn_ssa_value
**inout
, const struct glsl_type
*type
)
478 nir_intrinsic_instr
*instr
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
479 instr
->num_components
= glsl_get_vector_elements(type
);
483 nir_intrinsic_set_write_mask(instr
, (1 << instr
->num_components
) - 1);
484 instr
->src
[src
++] = nir_src_for_ssa((*inout
)->def
);
487 if (op
== nir_intrinsic_load_push_constant
) {
488 assert(access_offset
% 4 == 0);
490 nir_intrinsic_set_base(instr
, access_offset
);
491 nir_intrinsic_set_range(instr
, access_size
);
495 instr
->src
[src
++] = nir_src_for_ssa(index
);
497 if (op
== nir_intrinsic_load_push_constant
) {
498 /* We need to subtract the offset from where the intrinsic will load the
501 nir_src_for_ssa(nir_isub(&b
->nb
, offset
,
502 nir_imm_int(&b
->nb
, access_offset
)));
504 instr
->src
[src
++] = nir_src_for_ssa(offset
);
508 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
509 instr
->num_components
,
510 glsl_get_bit_size(type
), NULL
);
511 (*inout
)->def
= &instr
->dest
.ssa
;
514 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
516 if (load
&& glsl_get_base_type(type
) == GLSL_TYPE_BOOL
)
517 (*inout
)->def
= nir_ine(&b
->nb
, (*inout
)->def
, nir_imm_int(&b
->nb
, 0));
521 _vtn_block_load_store(struct vtn_builder
*b
, nir_intrinsic_op op
, bool load
,
522 nir_ssa_def
*index
, nir_ssa_def
*offset
,
523 unsigned access_offset
, unsigned access_size
,
524 struct vtn_access_chain
*chain
, unsigned chain_idx
,
525 struct vtn_type
*type
, struct vtn_ssa_value
**inout
)
527 if (chain
&& chain_idx
>= chain
->length
)
530 if (load
&& chain
== NULL
&& *inout
== NULL
)
531 *inout
= vtn_create_ssa_value(b
, type
->type
);
533 enum glsl_base_type base_type
= glsl_get_base_type(type
->type
);
537 case GLSL_TYPE_UINT64
:
538 case GLSL_TYPE_INT64
:
539 case GLSL_TYPE_FLOAT
:
540 case GLSL_TYPE_DOUBLE
:
542 /* This is where things get interesting. At this point, we've hit
543 * a vector, a scalar, or a matrix.
545 if (glsl_type_is_matrix(type
->type
)) {
547 /* Loading the whole matrix */
548 struct vtn_ssa_value
*transpose
;
549 unsigned num_ops
, vec_width
;
550 if (type
->row_major
) {
551 num_ops
= glsl_get_vector_elements(type
->type
);
552 vec_width
= glsl_get_matrix_columns(type
->type
);
554 const struct glsl_type
*transpose_type
=
555 glsl_matrix_type(base_type
, vec_width
, num_ops
);
556 *inout
= vtn_create_ssa_value(b
, transpose_type
);
558 transpose
= vtn_ssa_transpose(b
, *inout
);
562 num_ops
= glsl_get_matrix_columns(type
->type
);
563 vec_width
= glsl_get_vector_elements(type
->type
);
566 for (unsigned i
= 0; i
< num_ops
; i
++) {
567 nir_ssa_def
*elem_offset
=
568 nir_iadd(&b
->nb
, offset
,
569 nir_imm_int(&b
->nb
, i
* type
->stride
));
570 _vtn_load_store_tail(b
, op
, load
, index
, elem_offset
,
571 access_offset
, access_size
,
573 glsl_vector_type(base_type
, vec_width
));
576 if (load
&& type
->row_major
)
577 *inout
= vtn_ssa_transpose(b
, *inout
);
578 } else if (type
->row_major
) {
579 /* Row-major but with an access chiain. */
580 nir_ssa_def
*col_offset
=
581 vtn_access_link_as_ssa(b
, chain
->link
[chain_idx
],
582 type
->array_element
->stride
);
583 offset
= nir_iadd(&b
->nb
, offset
, col_offset
);
585 if (chain_idx
+ 1 < chain
->length
) {
586 /* Picking off a single element */
587 nir_ssa_def
*row_offset
=
588 vtn_access_link_as_ssa(b
, chain
->link
[chain_idx
+ 1],
590 offset
= nir_iadd(&b
->nb
, offset
, row_offset
);
592 *inout
= vtn_create_ssa_value(b
, glsl_scalar_type(base_type
));
593 _vtn_load_store_tail(b
, op
, load
, index
, offset
,
594 access_offset
, access_size
,
595 inout
, glsl_scalar_type(base_type
));
597 /* Grabbing a column; picking one element off each row */
598 unsigned num_comps
= glsl_get_vector_elements(type
->type
);
599 const struct glsl_type
*column_type
=
600 glsl_get_column_type(type
->type
);
602 nir_ssa_def
*comps
[4];
603 for (unsigned i
= 0; i
< num_comps
; i
++) {
604 nir_ssa_def
*elem_offset
=
605 nir_iadd(&b
->nb
, offset
,
606 nir_imm_int(&b
->nb
, i
* type
->stride
));
608 struct vtn_ssa_value
*comp
, temp_val
;
610 temp_val
.def
= nir_channel(&b
->nb
, (*inout
)->def
, i
);
611 temp_val
.type
= glsl_scalar_type(base_type
);
614 _vtn_load_store_tail(b
, op
, load
, index
, elem_offset
,
615 access_offset
, access_size
,
616 &comp
, glsl_scalar_type(base_type
));
617 comps
[i
] = comp
->def
;
622 *inout
= vtn_create_ssa_value(b
, column_type
);
624 (*inout
)->def
= nir_vec(&b
->nb
, comps
, num_comps
);
628 /* Column-major with a deref. Fall through to array case. */
629 nir_ssa_def
*col_offset
=
630 vtn_access_link_as_ssa(b
, chain
->link
[chain_idx
], type
->stride
);
631 offset
= nir_iadd(&b
->nb
, offset
, col_offset
);
633 _vtn_block_load_store(b
, op
, load
, index
, offset
,
634 access_offset
, access_size
,
635 chain
, chain_idx
+ 1,
636 type
->array_element
, inout
);
638 } else if (chain
== NULL
) {
639 /* Single whole vector */
640 assert(glsl_type_is_vector_or_scalar(type
->type
));
641 _vtn_load_store_tail(b
, op
, load
, index
, offset
,
642 access_offset
, access_size
,
645 /* Single component of a vector. Fall through to array case. */
646 nir_ssa_def
*elem_offset
=
647 vtn_access_link_as_ssa(b
, chain
->link
[chain_idx
], type
->stride
);
648 offset
= nir_iadd(&b
->nb
, offset
, elem_offset
);
650 _vtn_block_load_store(b
, op
, load
, index
, offset
,
651 access_offset
, access_size
,
653 type
->array_element
, inout
);
657 case GLSL_TYPE_ARRAY
: {
658 unsigned elems
= glsl_get_length(type
->type
);
659 for (unsigned i
= 0; i
< elems
; i
++) {
660 nir_ssa_def
*elem_off
=
661 nir_iadd(&b
->nb
, offset
, nir_imm_int(&b
->nb
, i
* type
->stride
));
662 _vtn_block_load_store(b
, op
, load
, index
, elem_off
,
663 access_offset
, access_size
,
665 type
->array_element
, &(*inout
)->elems
[i
]);
670 case GLSL_TYPE_STRUCT
: {
671 unsigned elems
= glsl_get_length(type
->type
);
672 for (unsigned i
= 0; i
< elems
; i
++) {
673 nir_ssa_def
*elem_off
=
674 nir_iadd(&b
->nb
, offset
, nir_imm_int(&b
->nb
, type
->offsets
[i
]));
675 _vtn_block_load_store(b
, op
, load
, index
, elem_off
,
676 access_offset
, access_size
,
678 type
->members
[i
], &(*inout
)->elems
[i
]);
684 unreachable("Invalid block member type");
688 static struct vtn_ssa_value
*
689 vtn_block_load(struct vtn_builder
*b
, struct vtn_access_chain
*src
)
692 unsigned access_offset
= 0, access_size
= 0;
693 switch (src
->var
->mode
) {
694 case vtn_variable_mode_ubo
:
695 op
= nir_intrinsic_load_ubo
;
697 case vtn_variable_mode_ssbo
:
698 op
= nir_intrinsic_load_ssbo
;
700 case vtn_variable_mode_push_constant
:
701 op
= nir_intrinsic_load_push_constant
;
702 vtn_access_chain_get_offset_size(src
, &access_offset
, &access_size
);
705 assert(!"Invalid block variable mode");
708 nir_ssa_def
*offset
, *index
= NULL
;
709 struct vtn_type
*type
;
711 offset
= vtn_access_chain_to_offset(b
, src
, &index
, &type
, &chain_idx
, true);
713 struct vtn_ssa_value
*value
= NULL
;
714 _vtn_block_load_store(b
, op
, true, index
, offset
,
715 access_offset
, access_size
,
716 src
, chain_idx
, type
, &value
);
721 vtn_block_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
722 struct vtn_access_chain
*dst
)
724 nir_ssa_def
*offset
, *index
= NULL
;
725 struct vtn_type
*type
;
727 offset
= vtn_access_chain_to_offset(b
, dst
, &index
, &type
, &chain_idx
, true);
729 _vtn_block_load_store(b
, nir_intrinsic_store_ssbo
, false, index
, offset
,
730 0, 0, dst
, chain_idx
, type
, &src
);
734 vtn_variable_is_external_block(struct vtn_variable
*var
)
736 return var
->mode
== vtn_variable_mode_ssbo
||
737 var
->mode
== vtn_variable_mode_ubo
||
738 var
->mode
== vtn_variable_mode_push_constant
;
742 _vtn_variable_load_store(struct vtn_builder
*b
, bool load
,
743 struct vtn_access_chain
*chain
,
744 struct vtn_type
*tail_type
,
745 struct vtn_ssa_value
**inout
)
747 enum glsl_base_type base_type
= glsl_get_base_type(tail_type
->type
);
751 case GLSL_TYPE_UINT64
:
752 case GLSL_TYPE_INT64
:
753 case GLSL_TYPE_FLOAT
:
755 case GLSL_TYPE_DOUBLE
:
756 /* At this point, we have a scalar, vector, or matrix so we know that
757 * there cannot be any structure splitting still in the way. By
758 * stopping at the matrix level rather than the vector level, we
759 * ensure that matrices get loaded in the optimal way even if they
760 * are storred row-major in a UBO.
763 *inout
= vtn_local_load(b
, vtn_access_chain_to_deref(b
, chain
));
765 vtn_local_store(b
, *inout
, vtn_access_chain_to_deref(b
, chain
));
769 case GLSL_TYPE_ARRAY
:
770 case GLSL_TYPE_STRUCT
: {
771 struct vtn_access_chain
*new_chain
=
772 vtn_access_chain_extend(b
, chain
, 1);
773 new_chain
->link
[chain
->length
].mode
= vtn_access_mode_literal
;
774 unsigned elems
= glsl_get_length(tail_type
->type
);
776 assert(*inout
== NULL
);
777 *inout
= rzalloc(b
, struct vtn_ssa_value
);
778 (*inout
)->type
= tail_type
->type
;
779 (*inout
)->elems
= rzalloc_array(b
, struct vtn_ssa_value
*, elems
);
781 for (unsigned i
= 0; i
< elems
; i
++) {
782 new_chain
->link
[chain
->length
].id
= i
;
783 struct vtn_type
*elem_type
= base_type
== GLSL_TYPE_ARRAY
?
784 tail_type
->array_element
: tail_type
->members
[i
];
785 _vtn_variable_load_store(b
, load
, new_chain
, elem_type
,
786 &(*inout
)->elems
[i
]);
792 unreachable("Invalid access chain type");
796 struct vtn_ssa_value
*
797 vtn_variable_load(struct vtn_builder
*b
, struct vtn_access_chain
*src
)
799 if (vtn_variable_is_external_block(src
->var
)) {
800 return vtn_block_load(b
, src
);
802 struct vtn_type
*tail_type
= vtn_access_chain_tail_type(b
, src
);
803 struct vtn_ssa_value
*val
= NULL
;
804 _vtn_variable_load_store(b
, true, src
, tail_type
, &val
);
810 vtn_variable_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
811 struct vtn_access_chain
*dest
)
813 if (vtn_variable_is_external_block(dest
->var
)) {
814 assert(dest
->var
->mode
== vtn_variable_mode_ssbo
);
815 vtn_block_store(b
, src
, dest
);
817 struct vtn_type
*tail_type
= vtn_access_chain_tail_type(b
, dest
);
818 _vtn_variable_load_store(b
, false, dest
, tail_type
, &src
);
823 _vtn_variable_copy(struct vtn_builder
*b
, struct vtn_access_chain
*dest
,
824 struct vtn_access_chain
*src
, struct vtn_type
*tail_type
)
826 enum glsl_base_type base_type
= glsl_get_base_type(tail_type
->type
);
830 case GLSL_TYPE_UINT64
:
831 case GLSL_TYPE_INT64
:
832 case GLSL_TYPE_FLOAT
:
833 case GLSL_TYPE_DOUBLE
:
835 /* At this point, we have a scalar, vector, or matrix so we know that
836 * there cannot be any structure splitting still in the way. By
837 * stopping at the matrix level rather than the vector level, we
838 * ensure that matrices get loaded in the optimal way even if they
839 * are storred row-major in a UBO.
841 vtn_variable_store(b
, vtn_variable_load(b
, src
), dest
);
844 case GLSL_TYPE_ARRAY
:
845 case GLSL_TYPE_STRUCT
: {
846 struct vtn_access_chain
*new_src
, *new_dest
;
847 new_src
= vtn_access_chain_extend(b
, src
, 1);
848 new_dest
= vtn_access_chain_extend(b
, dest
, 1);
849 new_src
->link
[src
->length
].mode
= vtn_access_mode_literal
;
850 new_dest
->link
[dest
->length
].mode
= vtn_access_mode_literal
;
851 unsigned elems
= glsl_get_length(tail_type
->type
);
852 for (unsigned i
= 0; i
< elems
; i
++) {
853 new_src
->link
[src
->length
].id
= i
;
854 new_dest
->link
[dest
->length
].id
= i
;
855 struct vtn_type
*elem_type
= base_type
== GLSL_TYPE_ARRAY
?
856 tail_type
->array_element
: tail_type
->members
[i
];
857 _vtn_variable_copy(b
, new_dest
, new_src
, elem_type
);
863 unreachable("Invalid access chain type");
868 vtn_variable_copy(struct vtn_builder
*b
, struct vtn_access_chain
*dest
,
869 struct vtn_access_chain
*src
)
871 struct vtn_type
*tail_type
= vtn_access_chain_tail_type(b
, src
);
872 assert(vtn_access_chain_tail_type(b
, dest
)->type
== tail_type
->type
);
874 /* TODO: At some point, we should add a special-case for when we can
875 * just emit a copy_var intrinsic.
877 _vtn_variable_copy(b
, dest
, src
, tail_type
);
881 set_mode_system_value(nir_variable_mode
*mode
)
883 assert(*mode
== nir_var_system_value
|| *mode
== nir_var_shader_in
);
884 *mode
= nir_var_system_value
;
888 vtn_get_builtin_location(struct vtn_builder
*b
,
889 SpvBuiltIn builtin
, int *location
,
890 nir_variable_mode
*mode
)
893 case SpvBuiltInPosition
:
894 *location
= VARYING_SLOT_POS
;
896 case SpvBuiltInPointSize
:
897 *location
= VARYING_SLOT_PSIZ
;
899 case SpvBuiltInClipDistance
:
900 *location
= VARYING_SLOT_CLIP_DIST0
; /* XXX CLIP_DIST1? */
902 case SpvBuiltInCullDistance
:
903 *location
= VARYING_SLOT_CULL_DIST0
;
905 case SpvBuiltInVertexIndex
:
906 *location
= SYSTEM_VALUE_VERTEX_ID
;
907 set_mode_system_value(mode
);
909 case SpvBuiltInVertexId
:
910 /* Vulkan defines VertexID to be zero-based and reserves the new
911 * builtin keyword VertexIndex to indicate the non-zero-based value.
913 *location
= SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
;
914 set_mode_system_value(mode
);
916 case SpvBuiltInInstanceIndex
:
917 *location
= SYSTEM_VALUE_INSTANCE_INDEX
;
918 set_mode_system_value(mode
);
920 case SpvBuiltInInstanceId
:
921 *location
= SYSTEM_VALUE_INSTANCE_ID
;
922 set_mode_system_value(mode
);
924 case SpvBuiltInPrimitiveId
:
925 if (b
->shader
->stage
== MESA_SHADER_FRAGMENT
) {
926 assert(*mode
== nir_var_shader_in
);
927 *location
= VARYING_SLOT_PRIMITIVE_ID
;
928 } else if (*mode
== nir_var_shader_out
) {
929 *location
= VARYING_SLOT_PRIMITIVE_ID
;
931 *location
= SYSTEM_VALUE_PRIMITIVE_ID
;
932 set_mode_system_value(mode
);
935 case SpvBuiltInInvocationId
:
936 *location
= SYSTEM_VALUE_INVOCATION_ID
;
937 set_mode_system_value(mode
);
939 case SpvBuiltInLayer
:
940 *location
= VARYING_SLOT_LAYER
;
941 if (b
->shader
->stage
== MESA_SHADER_FRAGMENT
)
942 *mode
= nir_var_shader_in
;
943 else if (b
->shader
->stage
== MESA_SHADER_GEOMETRY
)
944 *mode
= nir_var_shader_out
;
946 unreachable("invalid stage for SpvBuiltInLayer");
948 case SpvBuiltInViewportIndex
:
949 *location
= VARYING_SLOT_VIEWPORT
;
950 if (b
->shader
->stage
== MESA_SHADER_GEOMETRY
)
951 *mode
= nir_var_shader_out
;
952 else if (b
->shader
->stage
== MESA_SHADER_FRAGMENT
)
953 *mode
= nir_var_shader_in
;
955 unreachable("invalid stage for SpvBuiltInViewportIndex");
957 case SpvBuiltInTessLevelOuter
:
958 *location
= VARYING_SLOT_TESS_LEVEL_OUTER
;
960 case SpvBuiltInTessLevelInner
:
961 *location
= VARYING_SLOT_TESS_LEVEL_INNER
;
963 case SpvBuiltInTessCoord
:
964 *location
= SYSTEM_VALUE_TESS_COORD
;
965 set_mode_system_value(mode
);
967 case SpvBuiltInPatchVertices
:
968 *location
= SYSTEM_VALUE_VERTICES_IN
;
969 set_mode_system_value(mode
);
971 case SpvBuiltInFragCoord
:
972 *location
= VARYING_SLOT_POS
;
973 assert(*mode
== nir_var_shader_in
);
975 case SpvBuiltInPointCoord
:
976 *location
= VARYING_SLOT_PNTC
;
977 assert(*mode
== nir_var_shader_in
);
979 case SpvBuiltInFrontFacing
:
980 *location
= SYSTEM_VALUE_FRONT_FACE
;
981 set_mode_system_value(mode
);
983 case SpvBuiltInSampleId
:
984 *location
= SYSTEM_VALUE_SAMPLE_ID
;
985 set_mode_system_value(mode
);
987 case SpvBuiltInSamplePosition
:
988 *location
= SYSTEM_VALUE_SAMPLE_POS
;
989 set_mode_system_value(mode
);
991 case SpvBuiltInSampleMask
:
992 if (*mode
== nir_var_shader_out
) {
993 *location
= FRAG_RESULT_SAMPLE_MASK
;
995 *location
= SYSTEM_VALUE_SAMPLE_MASK_IN
;
996 set_mode_system_value(mode
);
999 case SpvBuiltInFragDepth
:
1000 *location
= FRAG_RESULT_DEPTH
;
1001 assert(*mode
== nir_var_shader_out
);
1003 case SpvBuiltInNumWorkgroups
:
1004 *location
= SYSTEM_VALUE_NUM_WORK_GROUPS
;
1005 set_mode_system_value(mode
);
1007 case SpvBuiltInWorkgroupSize
:
1008 /* This should already be handled */
1009 unreachable("unsupported builtin");
1011 case SpvBuiltInWorkgroupId
:
1012 *location
= SYSTEM_VALUE_WORK_GROUP_ID
;
1013 set_mode_system_value(mode
);
1015 case SpvBuiltInLocalInvocationId
:
1016 *location
= SYSTEM_VALUE_LOCAL_INVOCATION_ID
;
1017 set_mode_system_value(mode
);
1019 case SpvBuiltInLocalInvocationIndex
:
1020 *location
= SYSTEM_VALUE_LOCAL_INVOCATION_INDEX
;
1021 set_mode_system_value(mode
);
1023 case SpvBuiltInGlobalInvocationId
:
1024 *location
= SYSTEM_VALUE_GLOBAL_INVOCATION_ID
;
1025 set_mode_system_value(mode
);
1027 case SpvBuiltInBaseVertex
:
1028 *location
= SYSTEM_VALUE_BASE_VERTEX
;
1029 set_mode_system_value(mode
);
1031 case SpvBuiltInBaseInstance
:
1032 *location
= SYSTEM_VALUE_BASE_INSTANCE
;
1033 set_mode_system_value(mode
);
1035 case SpvBuiltInDrawIndex
:
1036 *location
= SYSTEM_VALUE_DRAW_ID
;
1037 set_mode_system_value(mode
);
1039 case SpvBuiltInViewIndex
:
1040 *location
= SYSTEM_VALUE_VIEW_INDEX
;
1041 set_mode_system_value(mode
);
1043 case SpvBuiltInHelperInvocation
:
1045 unreachable("unsupported builtin");
1050 apply_var_decoration(struct vtn_builder
*b
, nir_variable
*nir_var
,
1051 const struct vtn_decoration
*dec
)
1053 switch (dec
->decoration
) {
1054 case SpvDecorationRelaxedPrecision
:
1055 break; /* FIXME: Do nothing with this for now. */
1056 case SpvDecorationNoPerspective
:
1057 nir_var
->data
.interpolation
= INTERP_MODE_NOPERSPECTIVE
;
1059 case SpvDecorationFlat
:
1060 nir_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1062 case SpvDecorationCentroid
:
1063 nir_var
->data
.centroid
= true;
1065 case SpvDecorationSample
:
1066 nir_var
->data
.sample
= true;
1068 case SpvDecorationInvariant
:
1069 nir_var
->data
.invariant
= true;
1071 case SpvDecorationConstant
:
1072 assert(nir_var
->constant_initializer
!= NULL
);
1073 nir_var
->data
.read_only
= true;
1075 case SpvDecorationNonReadable
:
1076 nir_var
->data
.image
.write_only
= true;
1078 case SpvDecorationNonWritable
:
1079 nir_var
->data
.read_only
= true;
1080 nir_var
->data
.image
.read_only
= true;
1082 case SpvDecorationComponent
:
1083 nir_var
->data
.location_frac
= dec
->literals
[0];
1085 case SpvDecorationIndex
:
1086 nir_var
->data
.index
= dec
->literals
[0];
1088 case SpvDecorationBuiltIn
: {
1089 SpvBuiltIn builtin
= dec
->literals
[0];
1091 if (builtin
== SpvBuiltInWorkgroupSize
) {
1092 /* This shouldn't be a builtin. It's actually a constant. */
1093 nir_var
->data
.mode
= nir_var_global
;
1094 nir_var
->data
.read_only
= true;
1096 nir_constant
*c
= rzalloc(nir_var
, nir_constant
);
1097 c
->values
[0].u32
[0] = b
->shader
->info
->cs
.local_size
[0];
1098 c
->values
[0].u32
[1] = b
->shader
->info
->cs
.local_size
[1];
1099 c
->values
[0].u32
[2] = b
->shader
->info
->cs
.local_size
[2];
1100 nir_var
->constant_initializer
= c
;
1104 nir_variable_mode mode
= nir_var
->data
.mode
;
1105 vtn_get_builtin_location(b
, builtin
, &nir_var
->data
.location
, &mode
);
1106 nir_var
->data
.mode
= mode
;
1109 case SpvBuiltInTessLevelOuter
:
1110 case SpvBuiltInTessLevelInner
:
1111 nir_var
->data
.compact
= true;
1113 case SpvBuiltInSamplePosition
:
1114 nir_var
->data
.origin_upper_left
= b
->origin_upper_left
;
1116 case SpvBuiltInFragCoord
:
1117 nir_var
->data
.pixel_center_integer
= b
->pixel_center_integer
;
1124 case SpvDecorationSpecId
:
1125 case SpvDecorationRowMajor
:
1126 case SpvDecorationColMajor
:
1127 case SpvDecorationMatrixStride
:
1128 case SpvDecorationRestrict
:
1129 case SpvDecorationAliased
:
1130 case SpvDecorationVolatile
:
1131 case SpvDecorationCoherent
:
1132 case SpvDecorationUniform
:
1133 case SpvDecorationStream
:
1134 case SpvDecorationOffset
:
1135 case SpvDecorationLinkageAttributes
:
1136 break; /* Do nothing with these here */
1138 case SpvDecorationPatch
:
1139 nir_var
->data
.patch
= true;
1142 case SpvDecorationLocation
:
1143 unreachable("Handled above");
1145 case SpvDecorationBlock
:
1146 case SpvDecorationBufferBlock
:
1147 case SpvDecorationArrayStride
:
1148 case SpvDecorationGLSLShared
:
1149 case SpvDecorationGLSLPacked
:
1150 break; /* These can apply to a type but we don't care about them */
1152 case SpvDecorationBinding
:
1153 case SpvDecorationDescriptorSet
:
1154 case SpvDecorationNoContraction
:
1155 case SpvDecorationInputAttachmentIndex
:
1156 vtn_warn("Decoration not allowed for variable or structure member: %s",
1157 spirv_decoration_to_string(dec
->decoration
));
1160 case SpvDecorationXfbBuffer
:
1161 case SpvDecorationXfbStride
:
1162 vtn_warn("Vulkan does not have transform feedback: %s",
1163 spirv_decoration_to_string(dec
->decoration
));
1166 case SpvDecorationCPacked
:
1167 case SpvDecorationSaturatedConversion
:
1168 case SpvDecorationFuncParamAttr
:
1169 case SpvDecorationFPRoundingMode
:
1170 case SpvDecorationFPFastMathMode
:
1171 case SpvDecorationAlignment
:
1172 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1173 spirv_decoration_to_string(dec
->decoration
));
1177 unreachable("Unhandled decoration");
1182 var_is_patch_cb(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
1183 const struct vtn_decoration
*dec
, void *out_is_patch
)
1185 if (dec
->decoration
== SpvDecorationPatch
) {
1186 *((bool *) out_is_patch
) = true;
1191 var_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
1192 const struct vtn_decoration
*dec
, void *void_var
)
1194 struct vtn_variable
*vtn_var
= void_var
;
1196 /* Handle decorations that apply to a vtn_variable as a whole */
1197 switch (dec
->decoration
) {
1198 case SpvDecorationBinding
:
1199 vtn_var
->binding
= dec
->literals
[0];
1201 case SpvDecorationDescriptorSet
:
1202 vtn_var
->descriptor_set
= dec
->literals
[0];
1204 case SpvDecorationInputAttachmentIndex
:
1205 vtn_var
->input_attachment_index
= dec
->literals
[0];
1207 case SpvDecorationPatch
:
1208 vtn_var
->patch
= true;
1214 if (val
->value_type
== vtn_value_type_access_chain
) {
1215 assert(val
->access_chain
->length
== 0);
1216 assert(val
->access_chain
->var
== void_var
);
1217 assert(member
== -1);
1219 assert(val
->value_type
== vtn_value_type_type
);
1222 /* Location is odd. If applied to a split structure, we have to walk the
1223 * whole thing and accumulate the location. It's easier to handle as a
1226 if (dec
->decoration
== SpvDecorationLocation
) {
1227 unsigned location
= dec
->literals
[0];
1228 bool is_vertex_input
;
1229 if (b
->shader
->stage
== MESA_SHADER_FRAGMENT
&&
1230 vtn_var
->mode
== vtn_variable_mode_output
) {
1231 is_vertex_input
= false;
1232 location
+= FRAG_RESULT_DATA0
;
1233 } else if (b
->shader
->stage
== MESA_SHADER_VERTEX
&&
1234 vtn_var
->mode
== vtn_variable_mode_input
) {
1235 is_vertex_input
= true;
1236 location
+= VERT_ATTRIB_GENERIC0
;
1237 } else if (vtn_var
->mode
== vtn_variable_mode_input
||
1238 vtn_var
->mode
== vtn_variable_mode_output
) {
1239 is_vertex_input
= false;
1240 location
+= vtn_var
->patch
? VARYING_SLOT_PATCH0
: VARYING_SLOT_VAR0
;
1242 vtn_warn("Location must be on input or output variable");
1247 /* This handles the member and lone variable cases */
1248 vtn_var
->var
->data
.location
= location
;
1250 /* This handles the structure member case */
1251 assert(vtn_var
->members
);
1253 glsl_get_length(glsl_without_array(vtn_var
->type
->type
));
1254 for (unsigned i
= 0; i
< length
; i
++) {
1255 vtn_var
->members
[i
]->data
.location
= location
;
1257 glsl_count_attribute_slots(vtn_var
->members
[i
]->interface_type
,
1264 assert(member
<= 0);
1265 apply_var_decoration(b
, vtn_var
->var
, dec
);
1266 } else if (vtn_var
->members
) {
1268 assert(vtn_var
->members
);
1269 apply_var_decoration(b
, vtn_var
->members
[member
], dec
);
1272 glsl_get_length(glsl_without_array(vtn_var
->type
->type
));
1273 for (unsigned i
= 0; i
< length
; i
++)
1274 apply_var_decoration(b
, vtn_var
->members
[i
], dec
);
1277 /* A few variables, those with external storage, have no actual
1278 * nir_variables associated with them. Fortunately, all decorations
1279 * we care about for those variables are on the type only.
1281 assert(vtn_var
->mode
== vtn_variable_mode_ubo
||
1282 vtn_var
->mode
== vtn_variable_mode_ssbo
||
1283 vtn_var
->mode
== vtn_variable_mode_push_constant
);
1289 is_per_vertex_inout(const struct vtn_variable
*var
, gl_shader_stage stage
)
1291 if (var
->patch
|| !glsl_type_is_array(var
->type
->type
))
1294 if (var
->mode
== vtn_variable_mode_input
) {
1295 return stage
== MESA_SHADER_TESS_CTRL
||
1296 stage
== MESA_SHADER_TESS_EVAL
||
1297 stage
== MESA_SHADER_GEOMETRY
;
1300 if (var
->mode
== vtn_variable_mode_output
)
1301 return stage
== MESA_SHADER_TESS_CTRL
;
1307 vtn_handle_variables(struct vtn_builder
*b
, SpvOp opcode
,
1308 const uint32_t *w
, unsigned count
)
1312 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
1313 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1317 case SpvOpVariable
: {
1318 struct vtn_variable
*var
= rzalloc(b
, struct vtn_variable
);
1319 var
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1321 var
->chain
.var
= var
;
1322 var
->chain
.length
= 0;
1324 struct vtn_value
*val
=
1325 vtn_push_value(b
, w
[2], vtn_value_type_access_chain
);
1326 val
->access_chain
= &var
->chain
;
1328 struct vtn_type
*without_array
= var
->type
;
1329 while(glsl_type_is_array(without_array
->type
))
1330 without_array
= without_array
->array_element
;
1332 nir_variable_mode nir_mode
;
1333 switch ((SpvStorageClass
)w
[3]) {
1334 case SpvStorageClassUniform
:
1335 case SpvStorageClassUniformConstant
:
1336 if (without_array
->block
) {
1337 var
->mode
= vtn_variable_mode_ubo
;
1338 b
->shader
->info
->num_ubos
++;
1339 } else if (without_array
->buffer_block
) {
1340 var
->mode
= vtn_variable_mode_ssbo
;
1341 b
->shader
->info
->num_ssbos
++;
1342 } else if (glsl_type_is_image(without_array
->type
)) {
1343 var
->mode
= vtn_variable_mode_image
;
1344 nir_mode
= nir_var_uniform
;
1345 b
->shader
->info
->num_images
++;
1346 } else if (glsl_type_is_sampler(without_array
->type
)) {
1347 var
->mode
= vtn_variable_mode_sampler
;
1348 nir_mode
= nir_var_uniform
;
1349 b
->shader
->info
->num_textures
++;
1351 assert(!"Invalid uniform variable type");
1354 case SpvStorageClassPushConstant
:
1355 var
->mode
= vtn_variable_mode_push_constant
;
1356 assert(b
->shader
->num_uniforms
== 0);
1357 b
->shader
->num_uniforms
= vtn_type_block_size(var
->type
);
1359 case SpvStorageClassInput
:
1360 var
->mode
= vtn_variable_mode_input
;
1361 nir_mode
= nir_var_shader_in
;
1363 case SpvStorageClassOutput
:
1364 var
->mode
= vtn_variable_mode_output
;
1365 nir_mode
= nir_var_shader_out
;
1367 case SpvStorageClassPrivate
:
1368 var
->mode
= vtn_variable_mode_global
;
1369 nir_mode
= nir_var_global
;
1371 case SpvStorageClassFunction
:
1372 var
->mode
= vtn_variable_mode_local
;
1373 nir_mode
= nir_var_local
;
1375 case SpvStorageClassWorkgroup
:
1376 var
->mode
= vtn_variable_mode_workgroup
;
1377 nir_mode
= nir_var_shared
;
1379 case SpvStorageClassCrossWorkgroup
:
1380 case SpvStorageClassGeneric
:
1381 case SpvStorageClassAtomicCounter
:
1383 unreachable("Unhandled variable storage class");
1386 switch (var
->mode
) {
1387 case vtn_variable_mode_local
:
1388 case vtn_variable_mode_global
:
1389 case vtn_variable_mode_image
:
1390 case vtn_variable_mode_sampler
:
1391 case vtn_variable_mode_workgroup
:
1392 /* For these, we create the variable normally */
1393 var
->var
= rzalloc(b
->shader
, nir_variable
);
1394 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
1395 var
->var
->type
= var
->type
->type
;
1396 var
->var
->data
.mode
= nir_mode
;
1398 switch (var
->mode
) {
1399 case vtn_variable_mode_image
:
1400 case vtn_variable_mode_sampler
:
1401 var
->var
->interface_type
= without_array
->type
;
1404 var
->var
->interface_type
= NULL
;
1409 case vtn_variable_mode_input
:
1410 case vtn_variable_mode_output
: {
1411 /* In order to know whether or not we're a per-vertex inout, we need
1412 * the patch qualifier. This means walking the variable decorations
1413 * early before we actually create any variables. Not a big deal.
1415 * GLSLang really likes to place decorations in the most interior
1416 * thing it possibly can. In particular, if you have a struct, it
1417 * will place the patch decorations on the struct members. This
1418 * should be handled by the variable splitting below just fine.
1420 * If you have an array-of-struct, things get even more weird as it
1421 * will place the patch decorations on the struct even though it's
1422 * inside an array and some of the members being patch and others not
1423 * makes no sense whatsoever. Since the only sensible thing is for
1424 * it to be all or nothing, we'll call it patch if any of the members
1425 * are declared patch.
1428 vtn_foreach_decoration(b
, val
, var_is_patch_cb
, &var
->patch
);
1429 if (glsl_type_is_array(var
->type
->type
) &&
1430 glsl_type_is_struct(without_array
->type
)) {
1431 vtn_foreach_decoration(b
, without_array
->val
,
1432 var_is_patch_cb
, &var
->patch
);
1435 /* For inputs and outputs, we immediately split structures. This
1436 * is for a couple of reasons. For one, builtins may all come in
1437 * a struct and we really want those split out into separate
1438 * variables. For another, interpolation qualifiers can be
1439 * applied to members of the top-level struct ane we need to be
1440 * able to preserve that information.
1443 int array_length
= -1;
1444 struct vtn_type
*interface_type
= var
->type
;
1445 if (is_per_vertex_inout(var
, b
->shader
->stage
)) {
1446 /* In Geometry shaders (and some tessellation), inputs come
1447 * in per-vertex arrays. However, some builtins come in
1448 * non-per-vertex, hence the need for the is_array check. In
1449 * any case, there are no non-builtin arrays allowed so this
1450 * check should be sufficient.
1452 interface_type
= var
->type
->array_element
;
1453 array_length
= glsl_get_length(var
->type
->type
);
1456 if (glsl_type_is_struct(interface_type
->type
)) {
1457 /* It's a struct. Split it. */
1458 unsigned num_members
= glsl_get_length(interface_type
->type
);
1459 var
->members
= ralloc_array(b
, nir_variable
*, num_members
);
1461 for (unsigned i
= 0; i
< num_members
; i
++) {
1462 const struct glsl_type
*mtype
= interface_type
->members
[i
]->type
;
1463 if (array_length
>= 0)
1464 mtype
= glsl_array_type(mtype
, array_length
);
1466 var
->members
[i
] = rzalloc(b
->shader
, nir_variable
);
1467 var
->members
[i
]->name
=
1468 ralloc_asprintf(var
->members
[i
], "%s.%d", val
->name
, i
);
1469 var
->members
[i
]->type
= mtype
;
1470 var
->members
[i
]->interface_type
=
1471 interface_type
->members
[i
]->type
;
1472 var
->members
[i
]->data
.mode
= nir_mode
;
1473 var
->members
[i
]->data
.patch
= var
->patch
;
1476 var
->var
= rzalloc(b
->shader
, nir_variable
);
1477 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
1478 var
->var
->type
= var
->type
->type
;
1479 var
->var
->interface_type
= interface_type
->type
;
1480 var
->var
->data
.mode
= nir_mode
;
1481 var
->var
->data
.patch
= var
->patch
;
1484 /* For inputs and outputs, we need to grab locations and builtin
1485 * information from the interface type.
1487 vtn_foreach_decoration(b
, interface_type
->val
, var_decoration_cb
, var
);
1490 case vtn_variable_mode_param
:
1491 unreachable("Not created through OpVariable");
1494 case vtn_variable_mode_ubo
:
1495 case vtn_variable_mode_ssbo
:
1496 case vtn_variable_mode_push_constant
:
1497 /* These don't need actual variables. */
1503 nir_constant
*constant
=
1504 vtn_value(b
, w
[4], vtn_value_type_constant
)->constant
;
1505 var
->var
->constant_initializer
=
1506 nir_constant_clone(constant
, var
->var
);
1509 vtn_foreach_decoration(b
, val
, var_decoration_cb
, var
);
1511 if (var
->mode
== vtn_variable_mode_image
||
1512 var
->mode
== vtn_variable_mode_sampler
) {
1513 /* XXX: We still need the binding information in the nir_variable
1514 * for these. We should fix that.
1516 var
->var
->data
.binding
= var
->binding
;
1517 var
->var
->data
.descriptor_set
= var
->descriptor_set
;
1518 var
->var
->data
.index
= var
->input_attachment_index
;
1520 if (var
->mode
== vtn_variable_mode_image
)
1521 var
->var
->data
.image
.format
= without_array
->image_format
;
1524 if (var
->mode
== vtn_variable_mode_local
) {
1525 assert(var
->members
== NULL
&& var
->var
!= NULL
);
1526 nir_function_impl_add_variable(b
->impl
, var
->var
);
1527 } else if (var
->var
) {
1528 nir_shader_add_variable(b
->shader
, var
->var
);
1529 } else if (var
->members
) {
1530 unsigned count
= glsl_get_length(without_array
->type
);
1531 for (unsigned i
= 0; i
< count
; i
++) {
1532 assert(var
->members
[i
]->data
.mode
!= nir_var_local
);
1533 nir_shader_add_variable(b
->shader
, var
->members
[i
]);
1536 assert(var
->mode
== vtn_variable_mode_ubo
||
1537 var
->mode
== vtn_variable_mode_ssbo
||
1538 var
->mode
== vtn_variable_mode_push_constant
);
1543 case SpvOpAccessChain
:
1544 case SpvOpInBoundsAccessChain
: {
1545 struct vtn_access_chain
*base
, *chain
;
1546 struct vtn_value
*base_val
= vtn_untyped_value(b
, w
[3]);
1547 if (base_val
->value_type
== vtn_value_type_sampled_image
) {
1548 /* This is rather insane. SPIR-V allows you to use OpSampledImage
1549 * to combine an array of images with a single sampler to get an
1550 * array of sampled images that all share the same sampler.
1551 * Fortunately, this means that we can more-or-less ignore the
1552 * sampler when crawling the access chain, but it does leave us
1553 * with this rather awkward little special-case.
1555 base
= base_val
->sampled_image
->image
;
1557 assert(base_val
->value_type
== vtn_value_type_access_chain
);
1558 base
= base_val
->access_chain
;
1561 chain
= vtn_access_chain_extend(b
, base
, count
- 4);
1563 unsigned idx
= base
->length
;
1564 for (int i
= 4; i
< count
; i
++) {
1565 struct vtn_value
*link_val
= vtn_untyped_value(b
, w
[i
]);
1566 if (link_val
->value_type
== vtn_value_type_constant
) {
1567 chain
->link
[idx
].mode
= vtn_access_mode_literal
;
1568 chain
->link
[idx
].id
= link_val
->constant
->values
[0].u32
[0];
1570 chain
->link
[idx
].mode
= vtn_access_mode_id
;
1571 chain
->link
[idx
].id
= w
[i
];
1576 if (base_val
->value_type
== vtn_value_type_sampled_image
) {
1577 struct vtn_value
*val
=
1578 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
1579 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
1580 val
->sampled_image
->image
= chain
;
1581 val
->sampled_image
->sampler
= base_val
->sampled_image
->sampler
;
1583 struct vtn_value
*val
=
1584 vtn_push_value(b
, w
[2], vtn_value_type_access_chain
);
1585 val
->access_chain
= chain
;
1590 case SpvOpCopyMemory
: {
1591 struct vtn_value
*dest
= vtn_value(b
, w
[1], vtn_value_type_access_chain
);
1592 struct vtn_value
*src
= vtn_value(b
, w
[2], vtn_value_type_access_chain
);
1594 vtn_variable_copy(b
, dest
->access_chain
, src
->access_chain
);
1599 struct vtn_access_chain
*src
=
1600 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1602 if (src
->var
->mode
== vtn_variable_mode_image
||
1603 src
->var
->mode
== vtn_variable_mode_sampler
) {
1604 vtn_push_value(b
, w
[2], vtn_value_type_access_chain
)->access_chain
= src
;
1608 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1609 val
->ssa
= vtn_variable_load(b
, src
);
1614 struct vtn_access_chain
*dest
=
1615 vtn_value(b
, w
[1], vtn_value_type_access_chain
)->access_chain
;
1616 struct vtn_ssa_value
*src
= vtn_ssa_value(b
, w
[2]);
1617 vtn_variable_store(b
, src
, dest
);
1621 case SpvOpArrayLength
: {
1622 struct vtn_access_chain
*chain
=
1623 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1625 const uint32_t offset
= chain
->var
->type
->offsets
[w
[4]];
1626 const uint32_t stride
= chain
->var
->type
->members
[w
[4]]->stride
;
1629 struct vtn_type
*type
;
1630 nir_ssa_def
*index
=
1631 get_vulkan_resource_index(b
, chain
, &type
, &chain_idx
);
1633 nir_intrinsic_instr
*instr
=
1634 nir_intrinsic_instr_create(b
->nb
.shader
,
1635 nir_intrinsic_get_buffer_size
);
1636 instr
->src
[0] = nir_src_for_ssa(index
);
1637 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
1638 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
1639 nir_ssa_def
*buf_size
= &instr
->dest
.ssa
;
1641 /* array_length = max(buffer_size - offset, 0) / stride */
1642 nir_ssa_def
*array_length
=
1647 nir_imm_int(&b
->nb
, offset
)),
1648 nir_imm_int(&b
->nb
, 0u)),
1649 nir_imm_int(&b
->nb
, stride
));
1651 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1652 val
->ssa
= vtn_create_ssa_value(b
, glsl_uint_type());
1653 val
->ssa
->def
= array_length
;
1657 case SpvOpCopyMemorySized
:
1659 unreachable("Unhandled opcode");