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 chain
= ralloc_size(b
, sizeof(*chain
) + new_len
* sizeof(chain
->link
[0]));
40 chain
->var
= old
->var
;
41 chain
->length
= new_len
;
43 for (unsigned i
= 0; i
< old
->length
; i
++)
44 chain
->link
[i
] = old
->link
[i
];
50 vtn_access_link_as_ssa(struct vtn_builder
*b
, struct vtn_access_link link
,
54 if (link
.mode
== vtn_access_mode_literal
) {
55 return nir_imm_int(&b
->nb
, link
.id
* stride
);
56 } else if (stride
== 1) {
57 return vtn_ssa_value(b
, link
.id
)->def
;
59 return nir_imul(&b
->nb
, vtn_ssa_value(b
, link
.id
)->def
,
60 nir_imm_int(&b
->nb
, stride
));
64 static struct vtn_type
*
65 vtn_access_chain_tail_type(struct vtn_builder
*b
,
66 struct vtn_access_chain
*chain
)
68 struct vtn_type
*type
= chain
->var
->type
;
69 for (unsigned i
= 0; i
< chain
->length
; i
++) {
70 if (glsl_type_is_struct(type
->type
)) {
71 assert(chain
->link
[i
].mode
== vtn_access_mode_literal
);
72 type
= type
->members
[chain
->link
[i
].id
];
74 type
= type
->array_element
;
80 /* Crawls a chain of array derefs and rewrites the types so that the
81 * lengths stay the same but the terminal type is the one given by
82 * tail_type. This is useful for split structures.
85 rewrite_deref_types(nir_deref
*deref
, const struct glsl_type
*type
)
89 assert(deref
->child
->deref_type
== nir_deref_type_array
);
90 assert(glsl_type_is_array(deref
->type
));
91 rewrite_deref_types(deref
->child
, glsl_get_array_element(type
));
96 vtn_access_chain_to_deref(struct vtn_builder
*b
, struct vtn_access_chain
*chain
)
98 nir_deref_var
*deref_var
;
99 if (chain
->var
->var
) {
100 deref_var
= nir_deref_var_create(b
, chain
->var
->var
);
102 assert(chain
->var
->members
);
103 /* Create the deref_var manually. It will get filled out later. */
104 deref_var
= rzalloc(b
, nir_deref_var
);
105 deref_var
->deref
.deref_type
= nir_deref_type_var
;
108 struct vtn_type
*deref_type
= chain
->var
->type
;
109 nir_deref
*tail
= &deref_var
->deref
;
110 nir_variable
**members
= chain
->var
->members
;
112 for (unsigned i
= 0; i
< chain
->length
; i
++) {
113 enum glsl_base_type base_type
= glsl_get_base_type(deref_type
->type
);
117 case GLSL_TYPE_FLOAT
:
118 case GLSL_TYPE_DOUBLE
:
120 case GLSL_TYPE_ARRAY
: {
121 deref_type
= deref_type
->array_element
;
123 nir_deref_array
*deref_arr
= nir_deref_array_create(b
);
124 deref_arr
->deref
.type
= deref_type
->type
;
126 if (chain
->link
[i
].mode
== vtn_access_mode_literal
) {
127 deref_arr
->deref_array_type
= nir_deref_array_type_direct
;
128 deref_arr
->base_offset
= chain
->link
[i
].id
;
130 assert(chain
->link
[i
].mode
== vtn_access_mode_id
);
131 deref_arr
->deref_array_type
= nir_deref_array_type_indirect
;
132 deref_arr
->base_offset
= 0;
133 deref_arr
->indirect
=
134 nir_src_for_ssa(vtn_ssa_value(b
, chain
->link
[i
].id
)->def
);
136 tail
->child
= &deref_arr
->deref
;
141 case GLSL_TYPE_STRUCT
: {
142 assert(chain
->link
[i
].mode
== vtn_access_mode_literal
);
143 unsigned idx
= chain
->link
[i
].id
;
144 deref_type
= deref_type
->members
[idx
];
146 /* This is a pre-split structure. */
147 deref_var
->var
= members
[idx
];
148 rewrite_deref_types(&deref_var
->deref
, members
[idx
]->type
);
149 assert(tail
->type
== deref_type
->type
);
152 nir_deref_struct
*deref_struct
= nir_deref_struct_create(b
, idx
);
153 deref_struct
->deref
.type
= deref_type
->type
;
154 tail
->child
= &deref_struct
->deref
;
160 unreachable("Invalid type for deref");
164 assert(members
== NULL
);
169 _vtn_local_load_store(struct vtn_builder
*b
, bool load
, nir_deref_var
*deref
,
170 nir_deref
*tail
, struct vtn_ssa_value
*inout
)
172 /* The deref tail may contain a deref to select a component of a vector (in
173 * other words, it might not be an actual tail) so we have to save it away
174 * here since we overwrite it later.
176 nir_deref
*old_child
= tail
->child
;
178 if (glsl_type_is_vector_or_scalar(tail
->type
)) {
179 /* Terminate the deref chain in case there is one more link to pick
180 * off a component of the vector.
184 nir_intrinsic_op op
= load
? nir_intrinsic_load_var
:
185 nir_intrinsic_store_var
;
187 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
188 intrin
->variables
[0] =
189 nir_deref_as_var(nir_copy_deref(intrin
, &deref
->deref
));
190 intrin
->num_components
= glsl_get_vector_elements(tail
->type
);
193 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
,
194 intrin
->num_components
,
195 glsl_get_bit_size(tail
->type
),
197 inout
->def
= &intrin
->dest
.ssa
;
199 nir_intrinsic_set_write_mask(intrin
, (1 << intrin
->num_components
) - 1);
200 intrin
->src
[0] = nir_src_for_ssa(inout
->def
);
203 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
204 } else if (glsl_get_base_type(tail
->type
) == GLSL_TYPE_ARRAY
||
205 glsl_type_is_matrix(tail
->type
)) {
206 unsigned elems
= glsl_get_length(tail
->type
);
207 nir_deref_array
*deref_arr
= nir_deref_array_create(b
);
208 deref_arr
->deref_array_type
= nir_deref_array_type_direct
;
209 deref_arr
->deref
.type
= glsl_get_array_element(tail
->type
);
210 tail
->child
= &deref_arr
->deref
;
211 for (unsigned i
= 0; i
< elems
; i
++) {
212 deref_arr
->base_offset
= i
;
213 _vtn_local_load_store(b
, load
, deref
, tail
->child
, inout
->elems
[i
]);
216 assert(glsl_get_base_type(tail
->type
) == GLSL_TYPE_STRUCT
);
217 unsigned elems
= glsl_get_length(tail
->type
);
218 nir_deref_struct
*deref_struct
= nir_deref_struct_create(b
, 0);
219 tail
->child
= &deref_struct
->deref
;
220 for (unsigned i
= 0; i
< elems
; i
++) {
221 deref_struct
->index
= i
;
222 deref_struct
->deref
.type
= glsl_get_struct_field(tail
->type
, i
);
223 _vtn_local_load_store(b
, load
, deref
, tail
->child
, inout
->elems
[i
]);
227 tail
->child
= old_child
;
231 vtn_nir_deref(struct vtn_builder
*b
, uint32_t id
)
233 struct vtn_access_chain
*chain
=
234 vtn_value(b
, id
, vtn_value_type_access_chain
)->access_chain
;
236 return vtn_access_chain_to_deref(b
, chain
);
240 * Gets the NIR-level deref tail, which may have as a child an array deref
241 * selecting which component due to OpAccessChain supporting per-component
242 * indexing in SPIR-V.
245 get_deref_tail(nir_deref_var
*deref
)
247 nir_deref
*cur
= &deref
->deref
;
248 while (!glsl_type_is_vector_or_scalar(cur
->type
) && cur
->child
)
254 struct vtn_ssa_value
*
255 vtn_local_load(struct vtn_builder
*b
, nir_deref_var
*src
)
257 nir_deref
*src_tail
= get_deref_tail(src
);
258 struct vtn_ssa_value
*val
= vtn_create_ssa_value(b
, src_tail
->type
);
259 _vtn_local_load_store(b
, true, src
, src_tail
, val
);
261 if (src_tail
->child
) {
262 nir_deref_array
*vec_deref
= nir_deref_as_array(src_tail
->child
);
263 assert(vec_deref
->deref
.child
== NULL
);
264 val
->type
= vec_deref
->deref
.type
;
265 if (vec_deref
->deref_array_type
== nir_deref_array_type_direct
)
266 val
->def
= vtn_vector_extract(b
, val
->def
, vec_deref
->base_offset
);
268 val
->def
= vtn_vector_extract_dynamic(b
, val
->def
,
269 vec_deref
->indirect
.ssa
);
276 vtn_local_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
279 nir_deref
*dest_tail
= get_deref_tail(dest
);
281 if (dest_tail
->child
) {
282 struct vtn_ssa_value
*val
= vtn_create_ssa_value(b
, dest_tail
->type
);
283 _vtn_local_load_store(b
, true, dest
, dest_tail
, val
);
284 nir_deref_array
*deref
= nir_deref_as_array(dest_tail
->child
);
285 assert(deref
->deref
.child
== NULL
);
286 if (deref
->deref_array_type
== nir_deref_array_type_direct
)
287 val
->def
= vtn_vector_insert(b
, val
->def
, src
->def
,
290 val
->def
= vtn_vector_insert_dynamic(b
, val
->def
, src
->def
,
291 deref
->indirect
.ssa
);
292 _vtn_local_load_store(b
, false, dest
, dest_tail
, val
);
294 _vtn_local_load_store(b
, false, dest
, dest_tail
, src
);
299 get_vulkan_resource_index(struct vtn_builder
*b
, struct vtn_access_chain
*chain
,
300 struct vtn_type
**type
, unsigned *chain_idx
)
302 /* Push constants have no explicit binding */
303 if (chain
->var
->mode
== vtn_variable_mode_push_constant
) {
305 *type
= chain
->var
->type
;
309 nir_ssa_def
*array_index
;
310 if (glsl_type_is_array(chain
->var
->type
->type
)) {
311 assert(chain
->length
> 0);
312 array_index
= vtn_access_link_as_ssa(b
, chain
->link
[0], 1);
314 *type
= chain
->var
->type
->array_element
;
316 array_index
= nir_imm_int(&b
->nb
, 0);
318 *type
= chain
->var
->type
;
321 nir_intrinsic_instr
*instr
=
322 nir_intrinsic_instr_create(b
->nb
.shader
,
323 nir_intrinsic_vulkan_resource_index
);
324 instr
->src
[0] = nir_src_for_ssa(array_index
);
325 nir_intrinsic_set_desc_set(instr
, chain
->var
->descriptor_set
);
326 nir_intrinsic_set_binding(instr
, chain
->var
->binding
);
328 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
329 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
331 return &instr
->dest
.ssa
;
335 vtn_access_chain_to_offset(struct vtn_builder
*b
,
336 struct vtn_access_chain
*chain
,
337 nir_ssa_def
**index_out
, struct vtn_type
**type_out
,
338 unsigned *end_idx_out
, bool stop_at_matrix
)
341 struct vtn_type
*type
;
342 *index_out
= get_vulkan_resource_index(b
, chain
, &type
, &idx
);
344 nir_ssa_def
*offset
= nir_imm_int(&b
->nb
, 0);
345 for (; idx
< chain
->length
; idx
++) {
346 enum glsl_base_type base_type
= glsl_get_base_type(type
->type
);
350 case GLSL_TYPE_FLOAT
:
351 case GLSL_TYPE_DOUBLE
:
353 /* Some users may not want matrix or vector derefs */
358 case GLSL_TYPE_ARRAY
:
359 offset
= nir_iadd(&b
->nb
, offset
,
360 vtn_access_link_as_ssa(b
, chain
->link
[idx
],
363 type
= type
->array_element
;
366 case GLSL_TYPE_STRUCT
: {
367 assert(chain
->link
[idx
].mode
== vtn_access_mode_literal
);
368 unsigned member
= chain
->link
[idx
].id
;
369 offset
= nir_iadd(&b
->nb
, offset
,
370 nir_imm_int(&b
->nb
, type
->offsets
[member
]));
371 type
= type
->members
[member
];
376 unreachable("Invalid type for deref");
389 _vtn_load_store_tail(struct vtn_builder
*b
, nir_intrinsic_op op
, bool load
,
390 nir_ssa_def
*index
, nir_ssa_def
*offset
,
391 struct vtn_ssa_value
**inout
, const struct glsl_type
*type
)
393 nir_intrinsic_instr
*instr
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
394 instr
->num_components
= glsl_get_vector_elements(type
);
398 nir_intrinsic_set_write_mask(instr
, (1 << instr
->num_components
) - 1);
399 instr
->src
[src
++] = nir_src_for_ssa((*inout
)->def
);
402 /* We set the base and size for push constant load to the entire push
403 * constant block for now.
405 if (op
== nir_intrinsic_load_push_constant
) {
406 nir_intrinsic_set_base(instr
, 0);
407 nir_intrinsic_set_range(instr
, 128);
411 instr
->src
[src
++] = nir_src_for_ssa(index
);
413 instr
->src
[src
++] = nir_src_for_ssa(offset
);
416 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
417 instr
->num_components
,
418 glsl_get_bit_size(type
), NULL
);
419 (*inout
)->def
= &instr
->dest
.ssa
;
422 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
424 if (load
&& glsl_get_base_type(type
) == GLSL_TYPE_BOOL
)
425 (*inout
)->def
= nir_ine(&b
->nb
, (*inout
)->def
, nir_imm_int(&b
->nb
, 0));
429 _vtn_block_load_store(struct vtn_builder
*b
, nir_intrinsic_op op
, bool load
,
430 nir_ssa_def
*index
, nir_ssa_def
*offset
,
431 struct vtn_access_chain
*chain
, unsigned chain_idx
,
432 struct vtn_type
*type
, struct vtn_ssa_value
**inout
)
434 if (chain
&& chain_idx
>= chain
->length
)
437 if (load
&& chain
== NULL
&& *inout
== NULL
)
438 *inout
= vtn_create_ssa_value(b
, type
->type
);
440 enum glsl_base_type base_type
= glsl_get_base_type(type
->type
);
444 case GLSL_TYPE_FLOAT
:
446 /* This is where things get interesting. At this point, we've hit
447 * a vector, a scalar, or a matrix.
449 if (glsl_type_is_matrix(type
->type
)) {
451 /* Loading the whole matrix */
452 struct vtn_ssa_value
*transpose
;
453 unsigned num_ops
, vec_width
;
454 if (type
->row_major
) {
455 num_ops
= glsl_get_vector_elements(type
->type
);
456 vec_width
= glsl_get_matrix_columns(type
->type
);
458 const struct glsl_type
*transpose_type
=
459 glsl_matrix_type(base_type
, vec_width
, num_ops
);
460 *inout
= vtn_create_ssa_value(b
, transpose_type
);
462 transpose
= vtn_ssa_transpose(b
, *inout
);
466 num_ops
= glsl_get_matrix_columns(type
->type
);
467 vec_width
= glsl_get_vector_elements(type
->type
);
470 for (unsigned i
= 0; i
< num_ops
; i
++) {
471 nir_ssa_def
*elem_offset
=
472 nir_iadd(&b
->nb
, offset
,
473 nir_imm_int(&b
->nb
, i
* type
->stride
));
474 _vtn_load_store_tail(b
, op
, load
, index
, elem_offset
,
476 glsl_vector_type(base_type
, vec_width
));
479 if (load
&& type
->row_major
)
480 *inout
= vtn_ssa_transpose(b
, *inout
);
481 } else if (type
->row_major
) {
482 /* Row-major but with an access chiain. */
483 nir_ssa_def
*col_offset
=
484 vtn_access_link_as_ssa(b
, chain
->link
[chain_idx
],
485 type
->array_element
->stride
);
486 offset
= nir_iadd(&b
->nb
, offset
, col_offset
);
488 if (chain_idx
+ 1 < chain
->length
) {
489 /* Picking off a single element */
490 nir_ssa_def
*row_offset
=
491 vtn_access_link_as_ssa(b
, chain
->link
[chain_idx
+ 1],
493 offset
= nir_iadd(&b
->nb
, offset
, row_offset
);
495 *inout
= vtn_create_ssa_value(b
, glsl_scalar_type(base_type
));
496 _vtn_load_store_tail(b
, op
, load
, index
, offset
, inout
,
497 glsl_scalar_type(base_type
));
499 /* Grabbing a column; picking one element off each row */
500 unsigned num_comps
= glsl_get_vector_elements(type
->type
);
501 const struct glsl_type
*column_type
=
502 glsl_get_column_type(type
->type
);
504 nir_ssa_def
*comps
[4];
505 for (unsigned i
= 0; i
< num_comps
; i
++) {
506 nir_ssa_def
*elem_offset
=
507 nir_iadd(&b
->nb
, offset
,
508 nir_imm_int(&b
->nb
, i
* type
->stride
));
510 struct vtn_ssa_value
*comp
, temp_val
;
512 temp_val
.def
= nir_channel(&b
->nb
, (*inout
)->def
, i
);
513 temp_val
.type
= glsl_scalar_type(base_type
);
516 _vtn_load_store_tail(b
, op
, load
, index
, elem_offset
,
517 &comp
, glsl_scalar_type(base_type
));
518 comps
[i
] = comp
->def
;
523 *inout
= vtn_create_ssa_value(b
, column_type
);
525 (*inout
)->def
= nir_vec(&b
->nb
, comps
, num_comps
);
529 /* Column-major with a deref. Fall through to array case. */
530 nir_ssa_def
*col_offset
=
531 vtn_access_link_as_ssa(b
, chain
->link
[chain_idx
], type
->stride
);
532 offset
= nir_iadd(&b
->nb
, offset
, col_offset
);
534 _vtn_block_load_store(b
, op
, load
, index
, offset
,
535 chain
, chain_idx
+ 1,
536 type
->array_element
, inout
);
538 } else if (chain
== NULL
) {
539 /* Single whole vector */
540 assert(glsl_type_is_vector_or_scalar(type
->type
));
541 _vtn_load_store_tail(b
, op
, load
, index
, offset
, inout
, type
->type
);
543 /* Single component of a vector. Fall through to array case. */
544 nir_ssa_def
*elem_offset
=
545 vtn_access_link_as_ssa(b
, chain
->link
[chain_idx
], type
->stride
);
546 offset
= nir_iadd(&b
->nb
, offset
, elem_offset
);
548 _vtn_block_load_store(b
, op
, load
, index
, offset
, NULL
, 0,
549 type
->array_element
, inout
);
553 case GLSL_TYPE_ARRAY
: {
554 unsigned elems
= glsl_get_length(type
->type
);
555 for (unsigned i
= 0; i
< elems
; i
++) {
556 nir_ssa_def
*elem_off
=
557 nir_iadd(&b
->nb
, offset
, nir_imm_int(&b
->nb
, i
* type
->stride
));
558 _vtn_block_load_store(b
, op
, load
, index
, elem_off
, NULL
, 0,
559 type
->array_element
, &(*inout
)->elems
[i
]);
564 case GLSL_TYPE_STRUCT
: {
565 unsigned elems
= glsl_get_length(type
->type
);
566 for (unsigned i
= 0; i
< elems
; i
++) {
567 nir_ssa_def
*elem_off
=
568 nir_iadd(&b
->nb
, offset
, nir_imm_int(&b
->nb
, type
->offsets
[i
]));
569 _vtn_block_load_store(b
, op
, load
, index
, elem_off
, NULL
, 0,
570 type
->members
[i
], &(*inout
)->elems
[i
]);
576 unreachable("Invalid block member type");
580 static struct vtn_ssa_value
*
581 vtn_block_load(struct vtn_builder
*b
, struct vtn_access_chain
*src
)
584 switch (src
->var
->mode
) {
585 case vtn_variable_mode_ubo
:
586 op
= nir_intrinsic_load_ubo
;
588 case vtn_variable_mode_ssbo
:
589 op
= nir_intrinsic_load_ssbo
;
591 case vtn_variable_mode_push_constant
:
592 op
= nir_intrinsic_load_push_constant
;
595 assert(!"Invalid block variable mode");
598 nir_ssa_def
*offset
, *index
= NULL
;
599 struct vtn_type
*type
;
601 offset
= vtn_access_chain_to_offset(b
, src
, &index
, &type
, &chain_idx
, true);
603 struct vtn_ssa_value
*value
= NULL
;
604 _vtn_block_load_store(b
, op
, true, index
, offset
,
605 src
, chain_idx
, type
, &value
);
610 vtn_block_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
611 struct vtn_access_chain
*dst
)
613 nir_ssa_def
*offset
, *index
= NULL
;
614 struct vtn_type
*type
;
616 offset
= vtn_access_chain_to_offset(b
, dst
, &index
, &type
, &chain_idx
, true);
618 _vtn_block_load_store(b
, nir_intrinsic_store_ssbo
, false, index
, offset
,
619 dst
, chain_idx
, type
, &src
);
623 vtn_variable_is_external_block(struct vtn_variable
*var
)
625 return var
->mode
== vtn_variable_mode_ssbo
||
626 var
->mode
== vtn_variable_mode_ubo
||
627 var
->mode
== vtn_variable_mode_push_constant
;
631 _vtn_variable_load_store(struct vtn_builder
*b
, bool load
,
632 struct vtn_access_chain
*chain
,
633 struct vtn_type
*tail_type
,
634 struct vtn_ssa_value
**inout
)
636 enum glsl_base_type base_type
= glsl_get_base_type(tail_type
->type
);
640 case GLSL_TYPE_FLOAT
:
642 /* At this point, we have a scalar, vector, or matrix so we know that
643 * there cannot be any structure splitting still in the way. By
644 * stopping at the matrix level rather than the vector level, we
645 * ensure that matrices get loaded in the optimal way even if they
646 * are storred row-major in a UBO.
649 *inout
= vtn_local_load(b
, vtn_access_chain_to_deref(b
, chain
));
651 vtn_local_store(b
, *inout
, vtn_access_chain_to_deref(b
, chain
));
655 case GLSL_TYPE_ARRAY
:
656 case GLSL_TYPE_STRUCT
: {
657 struct vtn_access_chain
*new_chain
=
658 vtn_access_chain_extend(b
, chain
, 1);
659 new_chain
->link
[chain
->length
].mode
= vtn_access_mode_literal
;
660 unsigned elems
= glsl_get_length(tail_type
->type
);
662 assert(*inout
== NULL
);
663 *inout
= rzalloc(b
, struct vtn_ssa_value
);
664 (*inout
)->type
= tail_type
->type
;
665 (*inout
)->elems
= rzalloc_array(b
, struct vtn_ssa_value
*, elems
);
667 for (unsigned i
= 0; i
< elems
; i
++) {
668 new_chain
->link
[chain
->length
].id
= i
;
669 struct vtn_type
*elem_type
= base_type
== GLSL_TYPE_ARRAY
?
670 tail_type
->array_element
: tail_type
->members
[i
];
671 _vtn_variable_load_store(b
, load
, new_chain
, elem_type
,
672 &(*inout
)->elems
[i
]);
678 unreachable("Invalid access chain type");
682 struct vtn_ssa_value
*
683 vtn_variable_load(struct vtn_builder
*b
, struct vtn_access_chain
*src
)
685 if (vtn_variable_is_external_block(src
->var
)) {
686 return vtn_block_load(b
, src
);
688 struct vtn_type
*tail_type
= vtn_access_chain_tail_type(b
, src
);
689 struct vtn_ssa_value
*val
= NULL
;
690 _vtn_variable_load_store(b
, true, src
, tail_type
, &val
);
696 vtn_variable_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
697 struct vtn_access_chain
*dest
)
699 if (vtn_variable_is_external_block(dest
->var
)) {
700 assert(dest
->var
->mode
== vtn_variable_mode_ssbo
);
701 vtn_block_store(b
, src
, dest
);
703 struct vtn_type
*tail_type
= vtn_access_chain_tail_type(b
, dest
);
704 _vtn_variable_load_store(b
, false, dest
, tail_type
, &src
);
709 _vtn_variable_copy(struct vtn_builder
*b
, struct vtn_access_chain
*dest
,
710 struct vtn_access_chain
*src
, struct vtn_type
*tail_type
)
712 enum glsl_base_type base_type
= glsl_get_base_type(tail_type
->type
);
716 case GLSL_TYPE_FLOAT
:
718 /* At this point, we have a scalar, vector, or matrix so we know that
719 * there cannot be any structure splitting still in the way. By
720 * stopping at the matrix level rather than the vector level, we
721 * ensure that matrices get loaded in the optimal way even if they
722 * are storred row-major in a UBO.
724 vtn_variable_store(b
, vtn_variable_load(b
, src
), dest
);
727 case GLSL_TYPE_ARRAY
:
728 case GLSL_TYPE_STRUCT
: {
729 struct vtn_access_chain
*new_src
, *new_dest
;
730 new_src
= vtn_access_chain_extend(b
, src
, 1);
731 new_dest
= vtn_access_chain_extend(b
, dest
, 1);
732 new_src
->link
[src
->length
].mode
= vtn_access_mode_literal
;
733 new_dest
->link
[dest
->length
].mode
= vtn_access_mode_literal
;
734 unsigned elems
= glsl_get_length(tail_type
->type
);
735 for (unsigned i
= 0; i
< elems
; i
++) {
736 new_src
->link
[src
->length
].id
= i
;
737 new_dest
->link
[dest
->length
].id
= i
;
738 struct vtn_type
*elem_type
= base_type
== GLSL_TYPE_ARRAY
?
739 tail_type
->array_element
: tail_type
->members
[i
];
740 _vtn_variable_copy(b
, new_dest
, new_src
, elem_type
);
746 unreachable("Invalid access chain type");
751 vtn_variable_copy(struct vtn_builder
*b
, struct vtn_access_chain
*dest
,
752 struct vtn_access_chain
*src
)
754 struct vtn_type
*tail_type
= vtn_access_chain_tail_type(b
, src
);
755 assert(vtn_access_chain_tail_type(b
, dest
)->type
== tail_type
->type
);
757 /* TODO: At some point, we should add a special-case for when we can
758 * just emit a copy_var intrinsic.
760 _vtn_variable_copy(b
, dest
, src
, tail_type
);
764 set_mode_system_value(nir_variable_mode
*mode
)
766 assert(*mode
== nir_var_system_value
|| *mode
== nir_var_shader_in
);
767 *mode
= nir_var_system_value
;
771 vtn_get_builtin_location(struct vtn_builder
*b
,
772 SpvBuiltIn builtin
, int *location
,
773 nir_variable_mode
*mode
)
776 case SpvBuiltInPosition
:
777 *location
= VARYING_SLOT_POS
;
779 case SpvBuiltInPointSize
:
780 *location
= VARYING_SLOT_PSIZ
;
782 case SpvBuiltInClipDistance
:
783 *location
= VARYING_SLOT_CLIP_DIST0
; /* XXX CLIP_DIST1? */
785 case SpvBuiltInCullDistance
:
786 *location
= VARYING_SLOT_CULL_DIST0
;
788 case SpvBuiltInVertexIndex
:
789 *location
= SYSTEM_VALUE_VERTEX_ID
;
790 set_mode_system_value(mode
);
792 case SpvBuiltInVertexId
:
793 /* Vulkan defines VertexID to be zero-based and reserves the new
794 * builtin keyword VertexIndex to indicate the non-zero-based value.
796 *location
= SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
;
797 set_mode_system_value(mode
);
799 case SpvBuiltInInstanceIndex
:
800 *location
= SYSTEM_VALUE_INSTANCE_INDEX
;
801 set_mode_system_value(mode
);
803 case SpvBuiltInInstanceId
:
804 *location
= SYSTEM_VALUE_INSTANCE_ID
;
805 set_mode_system_value(mode
);
807 case SpvBuiltInPrimitiveId
:
808 *location
= VARYING_SLOT_PRIMITIVE_ID
;
809 *mode
= nir_var_shader_out
;
811 case SpvBuiltInInvocationId
:
812 *location
= SYSTEM_VALUE_INVOCATION_ID
;
813 set_mode_system_value(mode
);
815 case SpvBuiltInLayer
:
816 *location
= VARYING_SLOT_LAYER
;
817 *mode
= nir_var_shader_out
;
819 case SpvBuiltInViewportIndex
:
820 *location
= VARYING_SLOT_VIEWPORT
;
821 if (b
->shader
->stage
== MESA_SHADER_GEOMETRY
)
822 *mode
= nir_var_shader_out
;
823 else if (b
->shader
->stage
== MESA_SHADER_FRAGMENT
)
824 *mode
= nir_var_shader_in
;
826 unreachable("invalid stage for SpvBuiltInViewportIndex");
828 case SpvBuiltInTessLevelOuter
:
829 case SpvBuiltInTessLevelInner
:
830 case SpvBuiltInTessCoord
:
831 case SpvBuiltInPatchVertices
:
832 unreachable("no tessellation support");
833 case SpvBuiltInFragCoord
:
834 *location
= VARYING_SLOT_POS
;
835 assert(*mode
== nir_var_shader_in
);
837 case SpvBuiltInPointCoord
:
838 *location
= VARYING_SLOT_PNTC
;
839 assert(*mode
== nir_var_shader_in
);
841 case SpvBuiltInFrontFacing
:
842 *location
= SYSTEM_VALUE_FRONT_FACE
;
843 set_mode_system_value(mode
);
845 case SpvBuiltInSampleId
:
846 *location
= SYSTEM_VALUE_SAMPLE_ID
;
847 set_mode_system_value(mode
);
849 case SpvBuiltInSamplePosition
:
850 *location
= SYSTEM_VALUE_SAMPLE_POS
;
851 set_mode_system_value(mode
);
853 case SpvBuiltInSampleMask
:
854 *location
= SYSTEM_VALUE_SAMPLE_MASK_IN
; /* XXX out? */
855 set_mode_system_value(mode
);
857 case SpvBuiltInFragDepth
:
858 *location
= FRAG_RESULT_DEPTH
;
859 assert(*mode
== nir_var_shader_out
);
861 case SpvBuiltInNumWorkgroups
:
862 *location
= SYSTEM_VALUE_NUM_WORK_GROUPS
;
863 set_mode_system_value(mode
);
865 case SpvBuiltInWorkgroupSize
:
866 /* This should already be handled */
867 unreachable("unsupported builtin");
869 case SpvBuiltInWorkgroupId
:
870 *location
= SYSTEM_VALUE_WORK_GROUP_ID
;
871 set_mode_system_value(mode
);
873 case SpvBuiltInLocalInvocationId
:
874 *location
= SYSTEM_VALUE_LOCAL_INVOCATION_ID
;
875 set_mode_system_value(mode
);
877 case SpvBuiltInLocalInvocationIndex
:
878 *location
= SYSTEM_VALUE_LOCAL_INVOCATION_INDEX
;
879 set_mode_system_value(mode
);
881 case SpvBuiltInGlobalInvocationId
:
882 *location
= SYSTEM_VALUE_GLOBAL_INVOCATION_ID
;
883 set_mode_system_value(mode
);
885 case SpvBuiltInHelperInvocation
:
887 unreachable("unsupported builtin");
892 apply_var_decoration(struct vtn_builder
*b
, nir_variable
*nir_var
,
893 const struct vtn_decoration
*dec
)
895 switch (dec
->decoration
) {
896 case SpvDecorationRelaxedPrecision
:
897 break; /* FIXME: Do nothing with this for now. */
898 case SpvDecorationNoPerspective
:
899 nir_var
->data
.interpolation
= INTERP_MODE_NOPERSPECTIVE
;
901 case SpvDecorationFlat
:
902 nir_var
->data
.interpolation
= INTERP_MODE_FLAT
;
904 case SpvDecorationCentroid
:
905 nir_var
->data
.centroid
= true;
907 case SpvDecorationSample
:
908 nir_var
->data
.sample
= true;
910 case SpvDecorationInvariant
:
911 nir_var
->data
.invariant
= true;
913 case SpvDecorationConstant
:
914 assert(nir_var
->constant_initializer
!= NULL
);
915 nir_var
->data
.read_only
= true;
917 case SpvDecorationNonWritable
:
918 nir_var
->data
.read_only
= true;
920 case SpvDecorationComponent
:
921 nir_var
->data
.location_frac
= dec
->literals
[0];
923 case SpvDecorationIndex
:
924 nir_var
->data
.explicit_index
= true;
925 nir_var
->data
.index
= dec
->literals
[0];
927 case SpvDecorationBuiltIn
: {
928 SpvBuiltIn builtin
= dec
->literals
[0];
930 if (builtin
== SpvBuiltInWorkgroupSize
) {
931 /* This shouldn't be a builtin. It's actually a constant. */
932 nir_var
->data
.mode
= nir_var_global
;
933 nir_var
->data
.read_only
= true;
935 nir_constant
*c
= rzalloc(nir_var
, nir_constant
);
936 c
->value
.u
[0] = b
->shader
->info
.cs
.local_size
[0];
937 c
->value
.u
[1] = b
->shader
->info
.cs
.local_size
[1];
938 c
->value
.u
[2] = b
->shader
->info
.cs
.local_size
[2];
939 nir_var
->constant_initializer
= c
;
943 nir_variable_mode mode
= nir_var
->data
.mode
;
944 vtn_get_builtin_location(b
, builtin
, &nir_var
->data
.location
, &mode
);
945 nir_var
->data
.explicit_location
= true;
946 nir_var
->data
.mode
= mode
;
948 if (builtin
== SpvBuiltInFragCoord
|| builtin
== SpvBuiltInSamplePosition
)
949 nir_var
->data
.origin_upper_left
= b
->origin_upper_left
;
951 if (builtin
== SpvBuiltInFragCoord
)
952 nir_var
->data
.pixel_center_integer
= b
->pixel_center_integer
;
956 case SpvDecorationSpecId
:
957 case SpvDecorationRowMajor
:
958 case SpvDecorationColMajor
:
959 case SpvDecorationMatrixStride
:
960 case SpvDecorationRestrict
:
961 case SpvDecorationAliased
:
962 case SpvDecorationVolatile
:
963 case SpvDecorationCoherent
:
964 case SpvDecorationNonReadable
:
965 case SpvDecorationUniform
:
966 case SpvDecorationStream
:
967 case SpvDecorationOffset
:
968 case SpvDecorationLinkageAttributes
:
969 break; /* Do nothing with these here */
971 case SpvDecorationPatch
:
972 vtn_warn("Tessellation not yet supported");
975 case SpvDecorationLocation
:
976 unreachable("Handled above");
978 case SpvDecorationBlock
:
979 case SpvDecorationBufferBlock
:
980 case SpvDecorationArrayStride
:
981 case SpvDecorationGLSLShared
:
982 case SpvDecorationGLSLPacked
:
983 break; /* These can apply to a type but we don't care about them */
985 case SpvDecorationBinding
:
986 case SpvDecorationDescriptorSet
:
987 case SpvDecorationNoContraction
:
988 case SpvDecorationInputAttachmentIndex
:
989 vtn_warn("Decoration not allowed for variable or structure member: %s",
990 spirv_decoration_to_string(dec
->decoration
));
993 case SpvDecorationXfbBuffer
:
994 case SpvDecorationXfbStride
:
995 vtn_warn("Vulkan does not have transform feedback: %s",
996 spirv_decoration_to_string(dec
->decoration
));
999 case SpvDecorationCPacked
:
1000 case SpvDecorationSaturatedConversion
:
1001 case SpvDecorationFuncParamAttr
:
1002 case SpvDecorationFPRoundingMode
:
1003 case SpvDecorationFPFastMathMode
:
1004 case SpvDecorationAlignment
:
1005 vtn_warn("Decoraiton only allowed for CL-style kernels: %s",
1006 spirv_decoration_to_string(dec
->decoration
));
1012 var_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
1013 const struct vtn_decoration
*dec
, void *void_var
)
1015 struct vtn_variable
*vtn_var
= void_var
;
1017 /* Handle decorations that apply to a vtn_variable as a whole */
1018 switch (dec
->decoration
) {
1019 case SpvDecorationBinding
:
1020 vtn_var
->binding
= dec
->literals
[0];
1022 case SpvDecorationDescriptorSet
:
1023 vtn_var
->descriptor_set
= dec
->literals
[0];
1029 if (val
->value_type
== vtn_value_type_access_chain
) {
1030 assert(val
->access_chain
->length
== 0);
1031 assert(val
->access_chain
->var
== void_var
);
1032 assert(member
== -1);
1034 assert(val
->value_type
== vtn_value_type_type
);
1037 /* Location is odd. If applied to a split structure, we have to walk the
1038 * whole thing and accumulate the location. It's easier to handle as a
1041 if (dec
->decoration
== SpvDecorationLocation
) {
1042 unsigned location
= dec
->literals
[0];
1043 bool is_vertex_input
;
1044 if (b
->shader
->stage
== MESA_SHADER_FRAGMENT
&&
1045 vtn_var
->mode
== vtn_variable_mode_output
) {
1046 is_vertex_input
= false;
1047 location
+= FRAG_RESULT_DATA0
;
1048 } else if (b
->shader
->stage
== MESA_SHADER_VERTEX
&&
1049 vtn_var
->mode
== vtn_variable_mode_input
) {
1050 is_vertex_input
= true;
1051 location
+= VERT_ATTRIB_GENERIC0
;
1052 } else if (vtn_var
->mode
== vtn_variable_mode_input
||
1053 vtn_var
->mode
== vtn_variable_mode_output
) {
1054 is_vertex_input
= false;
1055 location
+= VARYING_SLOT_VAR0
;
1057 assert(!"Location must be on input or output variable");
1061 /* This handles the member and lone variable cases */
1062 vtn_var
->var
->data
.location
= location
;
1063 vtn_var
->var
->data
.explicit_location
= true;
1065 /* This handles the structure member case */
1066 assert(vtn_var
->members
);
1068 glsl_get_length(glsl_without_array(vtn_var
->type
->type
));
1069 for (unsigned i
= 0; i
< length
; i
++) {
1070 vtn_var
->members
[i
]->data
.location
= location
;
1071 vtn_var
->members
[i
]->data
.explicit_location
= true;
1073 glsl_count_attribute_slots(vtn_var
->members
[i
]->interface_type
,
1080 assert(member
<= 0);
1081 apply_var_decoration(b
, vtn_var
->var
, dec
);
1082 } else if (vtn_var
->members
) {
1084 assert(vtn_var
->members
);
1085 apply_var_decoration(b
, vtn_var
->members
[member
], dec
);
1088 glsl_get_length(glsl_without_array(vtn_var
->type
->type
));
1089 for (unsigned i
= 0; i
< length
; i
++)
1090 apply_var_decoration(b
, vtn_var
->members
[i
], dec
);
1093 /* A few variables, those with external storage, have no actual
1094 * nir_variables associated with them. Fortunately, all decorations
1095 * we care about for those variables are on the type only.
1097 assert(vtn_var
->mode
== vtn_variable_mode_ubo
||
1098 vtn_var
->mode
== vtn_variable_mode_ssbo
||
1099 vtn_var
->mode
== vtn_variable_mode_push_constant
);
1104 /* Tries to compute the size of an interface block based on the strides and
1105 * offsets that are provided to us in the SPIR-V source.
1108 vtn_type_block_size(struct vtn_type
*type
)
1110 enum glsl_base_type base_type
= glsl_get_base_type(type
->type
);
1111 switch (base_type
) {
1112 case GLSL_TYPE_UINT
:
1114 case GLSL_TYPE_FLOAT
:
1115 case GLSL_TYPE_BOOL
:
1116 case GLSL_TYPE_DOUBLE
: {
1117 unsigned cols
= type
->row_major
? glsl_get_vector_elements(type
->type
) :
1118 glsl_get_matrix_columns(type
->type
);
1120 assert(type
->stride
> 0);
1121 return type
->stride
* cols
;
1122 } else if (base_type
== GLSL_TYPE_DOUBLE
) {
1123 return glsl_get_vector_elements(type
->type
) * 8;
1125 return glsl_get_vector_elements(type
->type
) * 4;
1129 case GLSL_TYPE_STRUCT
:
1130 case GLSL_TYPE_INTERFACE
: {
1132 unsigned num_fields
= glsl_get_length(type
->type
);
1133 for (unsigned f
= 0; f
< num_fields
; f
++) {
1134 unsigned field_end
= type
->offsets
[f
] +
1135 vtn_type_block_size(type
->members
[f
]);
1136 size
= MAX2(size
, field_end
);
1141 case GLSL_TYPE_ARRAY
:
1142 assert(type
->stride
> 0);
1143 assert(glsl_get_length(type
->type
) > 0);
1144 return type
->stride
* glsl_get_length(type
->type
);
1147 assert(!"Invalid block type");
1153 vtn_handle_variables(struct vtn_builder
*b
, SpvOp opcode
,
1154 const uint32_t *w
, unsigned count
)
1157 case SpvOpVariable
: {
1158 struct vtn_variable
*var
= rzalloc(b
, struct vtn_variable
);
1159 var
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1161 var
->chain
.var
= var
;
1162 var
->chain
.length
= 0;
1164 struct vtn_value
*val
=
1165 vtn_push_value(b
, w
[2], vtn_value_type_access_chain
);
1166 val
->access_chain
= &var
->chain
;
1168 struct vtn_type
*without_array
= var
->type
;
1169 while(glsl_type_is_array(without_array
->type
))
1170 without_array
= without_array
->array_element
;
1172 nir_variable_mode nir_mode
;
1173 switch ((SpvStorageClass
)w
[3]) {
1174 case SpvStorageClassUniform
:
1175 case SpvStorageClassUniformConstant
:
1176 if (without_array
->block
) {
1177 var
->mode
= vtn_variable_mode_ubo
;
1178 b
->shader
->info
.num_ubos
++;
1179 } else if (without_array
->buffer_block
) {
1180 var
->mode
= vtn_variable_mode_ssbo
;
1181 b
->shader
->info
.num_ssbos
++;
1182 } else if (glsl_type_is_image(without_array
->type
)) {
1183 var
->mode
= vtn_variable_mode_image
;
1184 nir_mode
= nir_var_uniform
;
1185 b
->shader
->info
.num_images
++;
1186 } else if (glsl_type_is_sampler(without_array
->type
)) {
1187 var
->mode
= vtn_variable_mode_sampler
;
1188 nir_mode
= nir_var_uniform
;
1189 b
->shader
->info
.num_textures
++;
1191 assert(!"Invalid uniform variable type");
1194 case SpvStorageClassPushConstant
:
1195 var
->mode
= vtn_variable_mode_push_constant
;
1196 assert(b
->shader
->num_uniforms
== 0);
1197 b
->shader
->num_uniforms
= vtn_type_block_size(var
->type
);
1199 case SpvStorageClassInput
:
1200 var
->mode
= vtn_variable_mode_input
;
1201 nir_mode
= nir_var_shader_in
;
1203 case SpvStorageClassOutput
:
1204 var
->mode
= vtn_variable_mode_output
;
1205 nir_mode
= nir_var_shader_out
;
1207 case SpvStorageClassPrivate
:
1208 var
->mode
= vtn_variable_mode_global
;
1209 nir_mode
= nir_var_global
;
1211 case SpvStorageClassFunction
:
1212 var
->mode
= vtn_variable_mode_local
;
1213 nir_mode
= nir_var_local
;
1215 case SpvStorageClassWorkgroup
:
1216 var
->mode
= vtn_variable_mode_workgroup
;
1217 nir_mode
= nir_var_shared
;
1219 case SpvStorageClassCrossWorkgroup
:
1220 case SpvStorageClassGeneric
:
1221 case SpvStorageClassAtomicCounter
:
1223 unreachable("Unhandled variable storage class");
1226 switch (var
->mode
) {
1227 case vtn_variable_mode_local
:
1228 case vtn_variable_mode_global
:
1229 case vtn_variable_mode_image
:
1230 case vtn_variable_mode_sampler
:
1231 case vtn_variable_mode_workgroup
:
1232 /* For these, we create the variable normally */
1233 var
->var
= rzalloc(b
->shader
, nir_variable
);
1234 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
1235 var
->var
->type
= var
->type
->type
;
1236 var
->var
->data
.mode
= nir_mode
;
1238 switch (var
->mode
) {
1239 case vtn_variable_mode_image
:
1240 case vtn_variable_mode_sampler
:
1241 var
->var
->interface_type
= without_array
->type
;
1244 var
->var
->interface_type
= NULL
;
1249 case vtn_variable_mode_input
:
1250 case vtn_variable_mode_output
: {
1251 /* For inputs and outputs, we immediately split structures. This
1252 * is for a couple of reasons. For one, builtins may all come in
1253 * a struct and we really want those split out into separate
1254 * variables. For another, interpolation qualifiers can be
1255 * applied to members of the top-level struct ane we need to be
1256 * able to preserve that information.
1259 int array_length
= -1;
1260 struct vtn_type
*interface_type
= var
->type
;
1261 if (b
->shader
->stage
== MESA_SHADER_GEOMETRY
&&
1262 glsl_type_is_array(var
->type
->type
)) {
1263 /* In Geometry shaders (and some tessellation), inputs come
1264 * in per-vertex arrays. However, some builtins come in
1265 * non-per-vertex, hence the need for the is_array check. In
1266 * any case, there are no non-builtin arrays allowed so this
1267 * check should be sufficient.
1269 interface_type
= var
->type
->array_element
;
1270 array_length
= glsl_get_length(var
->type
->type
);
1273 if (glsl_type_is_struct(interface_type
->type
)) {
1274 /* It's a struct. Split it. */
1275 unsigned num_members
= glsl_get_length(interface_type
->type
);
1276 var
->members
= ralloc_array(b
, nir_variable
*, num_members
);
1278 for (unsigned i
= 0; i
< num_members
; i
++) {
1279 const struct glsl_type
*mtype
= interface_type
->members
[i
]->type
;
1280 if (array_length
>= 0)
1281 mtype
= glsl_array_type(mtype
, array_length
);
1283 var
->members
[i
] = rzalloc(b
->shader
, nir_variable
);
1284 var
->members
[i
]->name
=
1285 ralloc_asprintf(var
->members
[i
], "%s.%d", val
->name
, i
);
1286 var
->members
[i
]->type
= mtype
;
1287 var
->members
[i
]->interface_type
=
1288 interface_type
->members
[i
]->type
;
1289 var
->members
[i
]->data
.mode
= nir_mode
;
1292 var
->var
= rzalloc(b
->shader
, nir_variable
);
1293 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
1294 var
->var
->type
= var
->type
->type
;
1295 var
->var
->interface_type
= interface_type
->type
;
1296 var
->var
->data
.mode
= nir_mode
;
1299 /* For inputs and outputs, we need to grab locations and builtin
1300 * information from the interface type.
1302 vtn_foreach_decoration(b
, interface_type
->val
, var_decoration_cb
, var
);
1305 case vtn_variable_mode_param
:
1306 unreachable("Not created through OpVariable");
1309 case vtn_variable_mode_ubo
:
1310 case vtn_variable_mode_ssbo
:
1311 case vtn_variable_mode_push_constant
:
1312 /* These don't need actual variables. */
1318 nir_constant
*constant
=
1319 vtn_value(b
, w
[4], vtn_value_type_constant
)->constant
;
1320 var
->var
->constant_initializer
=
1321 nir_constant_clone(constant
, var
->var
);
1324 vtn_foreach_decoration(b
, val
, var_decoration_cb
, var
);
1326 if (var
->mode
== vtn_variable_mode_image
||
1327 var
->mode
== vtn_variable_mode_sampler
) {
1328 /* XXX: We still need the binding information in the nir_variable
1329 * for these. We should fix that.
1331 var
->var
->data
.binding
= var
->binding
;
1332 var
->var
->data
.descriptor_set
= var
->descriptor_set
;
1334 if (var
->mode
== vtn_variable_mode_image
)
1335 var
->var
->data
.image
.format
= without_array
->image_format
;
1338 if (var
->mode
== vtn_variable_mode_local
) {
1339 assert(var
->members
== NULL
&& var
->var
!= NULL
);
1340 nir_function_impl_add_variable(b
->impl
, var
->var
);
1341 } else if (var
->var
) {
1342 nir_shader_add_variable(b
->shader
, var
->var
);
1343 } else if (var
->members
) {
1344 unsigned count
= glsl_get_length(without_array
->type
);
1345 for (unsigned i
= 0; i
< count
; i
++) {
1346 assert(var
->members
[i
]->data
.mode
!= nir_var_local
);
1347 nir_shader_add_variable(b
->shader
, var
->members
[i
]);
1350 assert(var
->mode
== vtn_variable_mode_ubo
||
1351 var
->mode
== vtn_variable_mode_ssbo
||
1352 var
->mode
== vtn_variable_mode_push_constant
);
1357 case SpvOpAccessChain
:
1358 case SpvOpInBoundsAccessChain
: {
1359 struct vtn_access_chain
*base
, *chain
;
1360 struct vtn_value
*base_val
= vtn_untyped_value(b
, w
[3]);
1361 if (base_val
->value_type
== vtn_value_type_sampled_image
) {
1362 /* This is rather insane. SPIR-V allows you to use OpSampledImage
1363 * to combine an array of images with a single sampler to get an
1364 * array of sampled images that all share the same sampler.
1365 * Fortunately, this means that we can more-or-less ignore the
1366 * sampler when crawling the access chain, but it does leave us
1367 * with this rather awkward little special-case.
1369 base
= base_val
->sampled_image
->image
;
1371 assert(base_val
->value_type
== vtn_value_type_access_chain
);
1372 base
= base_val
->access_chain
;
1375 chain
= vtn_access_chain_extend(b
, base
, count
- 4);
1377 unsigned idx
= base
->length
;
1378 for (int i
= 4; i
< count
; i
++) {
1379 struct vtn_value
*link_val
= vtn_untyped_value(b
, w
[i
]);
1380 if (link_val
->value_type
== vtn_value_type_constant
) {
1381 chain
->link
[idx
].mode
= vtn_access_mode_literal
;
1382 chain
->link
[idx
].id
= link_val
->constant
->value
.u
[0];
1384 chain
->link
[idx
].mode
= vtn_access_mode_id
;
1385 chain
->link
[idx
].id
= w
[i
];
1390 if (base_val
->value_type
== vtn_value_type_sampled_image
) {
1391 struct vtn_value
*val
=
1392 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
1393 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
1394 val
->sampled_image
->image
= chain
;
1395 val
->sampled_image
->sampler
= base_val
->sampled_image
->sampler
;
1397 struct vtn_value
*val
=
1398 vtn_push_value(b
, w
[2], vtn_value_type_access_chain
);
1399 val
->access_chain
= chain
;
1404 case SpvOpCopyMemory
: {
1405 struct vtn_value
*dest
= vtn_value(b
, w
[1], vtn_value_type_access_chain
);
1406 struct vtn_value
*src
= vtn_value(b
, w
[2], vtn_value_type_access_chain
);
1408 vtn_variable_copy(b
, dest
->access_chain
, src
->access_chain
);
1413 struct vtn_access_chain
*src
=
1414 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1416 if (src
->var
->mode
== vtn_variable_mode_image
||
1417 src
->var
->mode
== vtn_variable_mode_sampler
) {
1418 vtn_push_value(b
, w
[2], vtn_value_type_access_chain
)->access_chain
= src
;
1422 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1423 val
->ssa
= vtn_variable_load(b
, src
);
1428 struct vtn_access_chain
*dest
=
1429 vtn_value(b
, w
[1], vtn_value_type_access_chain
)->access_chain
;
1430 struct vtn_ssa_value
*src
= vtn_ssa_value(b
, w
[2]);
1431 vtn_variable_store(b
, src
, dest
);
1435 case SpvOpArrayLength
: {
1436 struct vtn_access_chain
*chain
=
1437 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1439 const uint32_t offset
= chain
->var
->type
->offsets
[w
[4]];
1440 const uint32_t stride
= chain
->var
->type
->members
[w
[4]]->stride
;
1443 struct vtn_type
*type
;
1444 nir_ssa_def
*index
=
1445 get_vulkan_resource_index(b
, chain
, &type
, &chain_idx
);
1447 nir_intrinsic_instr
*instr
=
1448 nir_intrinsic_instr_create(b
->nb
.shader
,
1449 nir_intrinsic_get_buffer_size
);
1450 instr
->src
[0] = nir_src_for_ssa(index
);
1451 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
1452 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
1453 nir_ssa_def
*buf_size
= &instr
->dest
.ssa
;
1455 /* array_length = max(buffer_size - offset, 0) / stride */
1456 nir_ssa_def
*array_length
=
1461 nir_imm_int(&b
->nb
, offset
)),
1462 nir_imm_int(&b
->nb
, 0u)),
1463 nir_imm_int(&b
->nb
, stride
));
1465 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1466 val
->ssa
= vtn_create_ssa_value(b
, glsl_uint_type());
1467 val
->ssa
->def
= array_length
;
1471 case SpvOpCopyMemorySized
:
1473 unreachable("Unhandled opcode");