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_FLOAT
:
119 case GLSL_TYPE_DOUBLE
:
121 case GLSL_TYPE_ARRAY
: {
122 deref_type
= deref_type
->array_element
;
124 nir_deref_array
*deref_arr
= nir_deref_array_create(b
);
125 deref_arr
->deref
.type
= deref_type
->type
;
127 if (chain
->link
[i
].mode
== vtn_access_mode_literal
) {
128 deref_arr
->deref_array_type
= nir_deref_array_type_direct
;
129 deref_arr
->base_offset
= chain
->link
[i
].id
;
131 assert(chain
->link
[i
].mode
== vtn_access_mode_id
);
132 deref_arr
->deref_array_type
= nir_deref_array_type_indirect
;
133 deref_arr
->base_offset
= 0;
134 deref_arr
->indirect
=
135 nir_src_for_ssa(vtn_ssa_value(b
, chain
->link
[i
].id
)->def
);
137 tail
->child
= &deref_arr
->deref
;
142 case GLSL_TYPE_STRUCT
: {
143 assert(chain
->link
[i
].mode
== vtn_access_mode_literal
);
144 unsigned idx
= chain
->link
[i
].id
;
145 deref_type
= deref_type
->members
[idx
];
147 /* This is a pre-split structure. */
148 deref_var
->var
= members
[idx
];
149 rewrite_deref_types(&deref_var
->deref
, members
[idx
]->type
);
150 assert(tail
->type
== deref_type
->type
);
153 nir_deref_struct
*deref_struct
= nir_deref_struct_create(b
, idx
);
154 deref_struct
->deref
.type
= deref_type
->type
;
155 tail
->child
= &deref_struct
->deref
;
161 unreachable("Invalid type for deref");
165 assert(members
== NULL
);
170 _vtn_local_load_store(struct vtn_builder
*b
, bool load
, nir_deref_var
*deref
,
171 nir_deref
*tail
, struct vtn_ssa_value
*inout
)
173 /* The deref tail may contain a deref to select a component of a vector (in
174 * other words, it might not be an actual tail) so we have to save it away
175 * here since we overwrite it later.
177 nir_deref
*old_child
= tail
->child
;
179 if (glsl_type_is_vector_or_scalar(tail
->type
)) {
180 /* Terminate the deref chain in case there is one more link to pick
181 * off a component of the vector.
185 nir_intrinsic_op op
= load
? nir_intrinsic_load_var
:
186 nir_intrinsic_store_var
;
188 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
189 intrin
->variables
[0] = nir_deref_var_clone(deref
, intrin
);
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");
388 /* Tries to compute the size of an interface block based on the strides and
389 * offsets that are provided to us in the SPIR-V source.
392 vtn_type_block_size(struct vtn_type
*type
)
394 enum glsl_base_type base_type
= glsl_get_base_type(type
->type
);
398 case GLSL_TYPE_FLOAT
:
400 case GLSL_TYPE_DOUBLE
: {
401 unsigned cols
= type
->row_major
? glsl_get_vector_elements(type
->type
) :
402 glsl_get_matrix_columns(type
->type
);
404 assert(type
->stride
> 0);
405 return type
->stride
* cols
;
406 } else if (base_type
== GLSL_TYPE_DOUBLE
) {
407 return glsl_get_vector_elements(type
->type
) * 8;
409 return glsl_get_vector_elements(type
->type
) * 4;
413 case GLSL_TYPE_STRUCT
:
414 case GLSL_TYPE_INTERFACE
: {
416 unsigned num_fields
= glsl_get_length(type
->type
);
417 for (unsigned f
= 0; f
< num_fields
; f
++) {
418 unsigned field_end
= type
->offsets
[f
] +
419 vtn_type_block_size(type
->members
[f
]);
420 size
= MAX2(size
, field_end
);
425 case GLSL_TYPE_ARRAY
:
426 assert(type
->stride
> 0);
427 assert(glsl_get_length(type
->type
) > 0);
428 return type
->stride
* glsl_get_length(type
->type
);
431 assert(!"Invalid block type");
437 vtn_access_chain_get_offset_size(struct vtn_access_chain
*chain
,
438 unsigned *access_offset
,
439 unsigned *access_size
)
441 /* Only valid for push constants accesses now. */
442 assert(chain
->var
->mode
== vtn_variable_mode_push_constant
);
444 struct vtn_type
*type
= chain
->var
->type
;
448 for (unsigned i
= 0; i
< chain
->length
; i
++) {
449 if (chain
->link
[i
].mode
!= vtn_access_mode_literal
)
452 if (glsl_type_is_struct(type
->type
)) {
453 *access_offset
+= type
->offsets
[chain
->link
[i
].id
];
454 type
= type
->members
[chain
->link
[i
].id
];
456 *access_offset
+= type
->stride
* chain
->link
[i
].id
;
457 type
= type
->array_element
;
461 *access_size
= vtn_type_block_size(type
);
465 _vtn_load_store_tail(struct vtn_builder
*b
, nir_intrinsic_op op
, bool load
,
466 nir_ssa_def
*index
, nir_ssa_def
*offset
,
467 unsigned access_offset
, unsigned access_size
,
468 struct vtn_ssa_value
**inout
, const struct glsl_type
*type
)
470 nir_intrinsic_instr
*instr
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
471 instr
->num_components
= glsl_get_vector_elements(type
);
475 nir_intrinsic_set_write_mask(instr
, (1 << instr
->num_components
) - 1);
476 instr
->src
[src
++] = nir_src_for_ssa((*inout
)->def
);
479 if (op
== nir_intrinsic_load_push_constant
) {
480 assert(access_offset
% 4 == 0);
482 nir_intrinsic_set_base(instr
, access_offset
);
483 nir_intrinsic_set_range(instr
, access_size
);
487 instr
->src
[src
++] = nir_src_for_ssa(index
);
489 if (op
== nir_intrinsic_load_push_constant
) {
490 /* We need to subtract the offset from where the intrinsic will load the
493 nir_src_for_ssa(nir_isub(&b
->nb
, offset
,
494 nir_imm_int(&b
->nb
, access_offset
)));
496 instr
->src
[src
++] = nir_src_for_ssa(offset
);
500 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
501 instr
->num_components
,
502 glsl_get_bit_size(type
), NULL
);
503 (*inout
)->def
= &instr
->dest
.ssa
;
506 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
508 if (load
&& glsl_get_base_type(type
) == GLSL_TYPE_BOOL
)
509 (*inout
)->def
= nir_ine(&b
->nb
, (*inout
)->def
, nir_imm_int(&b
->nb
, 0));
513 _vtn_block_load_store(struct vtn_builder
*b
, nir_intrinsic_op op
, bool load
,
514 nir_ssa_def
*index
, nir_ssa_def
*offset
,
515 unsigned access_offset
, unsigned access_size
,
516 struct vtn_access_chain
*chain
, unsigned chain_idx
,
517 struct vtn_type
*type
, struct vtn_ssa_value
**inout
)
519 if (chain
&& chain_idx
>= chain
->length
)
522 if (load
&& chain
== NULL
&& *inout
== NULL
)
523 *inout
= vtn_create_ssa_value(b
, type
->type
);
525 enum glsl_base_type base_type
= glsl_get_base_type(type
->type
);
529 case GLSL_TYPE_FLOAT
:
530 case GLSL_TYPE_DOUBLE
:
532 /* This is where things get interesting. At this point, we've hit
533 * a vector, a scalar, or a matrix.
535 if (glsl_type_is_matrix(type
->type
)) {
537 /* Loading the whole matrix */
538 struct vtn_ssa_value
*transpose
;
539 unsigned num_ops
, vec_width
;
540 if (type
->row_major
) {
541 num_ops
= glsl_get_vector_elements(type
->type
);
542 vec_width
= glsl_get_matrix_columns(type
->type
);
544 const struct glsl_type
*transpose_type
=
545 glsl_matrix_type(base_type
, vec_width
, num_ops
);
546 *inout
= vtn_create_ssa_value(b
, transpose_type
);
548 transpose
= vtn_ssa_transpose(b
, *inout
);
552 num_ops
= glsl_get_matrix_columns(type
->type
);
553 vec_width
= glsl_get_vector_elements(type
->type
);
556 for (unsigned i
= 0; i
< num_ops
; i
++) {
557 nir_ssa_def
*elem_offset
=
558 nir_iadd(&b
->nb
, offset
,
559 nir_imm_int(&b
->nb
, i
* type
->stride
));
560 _vtn_load_store_tail(b
, op
, load
, index
, elem_offset
,
561 access_offset
, access_size
,
563 glsl_vector_type(base_type
, vec_width
));
566 if (load
&& type
->row_major
)
567 *inout
= vtn_ssa_transpose(b
, *inout
);
568 } else if (type
->row_major
) {
569 /* Row-major but with an access chiain. */
570 nir_ssa_def
*col_offset
=
571 vtn_access_link_as_ssa(b
, chain
->link
[chain_idx
],
572 type
->array_element
->stride
);
573 offset
= nir_iadd(&b
->nb
, offset
, col_offset
);
575 if (chain_idx
+ 1 < chain
->length
) {
576 /* Picking off a single element */
577 nir_ssa_def
*row_offset
=
578 vtn_access_link_as_ssa(b
, chain
->link
[chain_idx
+ 1],
580 offset
= nir_iadd(&b
->nb
, offset
, row_offset
);
582 *inout
= vtn_create_ssa_value(b
, glsl_scalar_type(base_type
));
583 _vtn_load_store_tail(b
, op
, load
, index
, offset
,
584 access_offset
, access_size
,
585 inout
, glsl_scalar_type(base_type
));
587 /* Grabbing a column; picking one element off each row */
588 unsigned num_comps
= glsl_get_vector_elements(type
->type
);
589 const struct glsl_type
*column_type
=
590 glsl_get_column_type(type
->type
);
592 nir_ssa_def
*comps
[4];
593 for (unsigned i
= 0; i
< num_comps
; i
++) {
594 nir_ssa_def
*elem_offset
=
595 nir_iadd(&b
->nb
, offset
,
596 nir_imm_int(&b
->nb
, i
* type
->stride
));
598 struct vtn_ssa_value
*comp
, temp_val
;
600 temp_val
.def
= nir_channel(&b
->nb
, (*inout
)->def
, i
);
601 temp_val
.type
= glsl_scalar_type(base_type
);
604 _vtn_load_store_tail(b
, op
, load
, index
, elem_offset
,
605 access_offset
, access_size
,
606 &comp
, glsl_scalar_type(base_type
));
607 comps
[i
] = comp
->def
;
612 *inout
= vtn_create_ssa_value(b
, column_type
);
614 (*inout
)->def
= nir_vec(&b
->nb
, comps
, num_comps
);
618 /* Column-major with a deref. Fall through to array case. */
619 nir_ssa_def
*col_offset
=
620 vtn_access_link_as_ssa(b
, chain
->link
[chain_idx
], type
->stride
);
621 offset
= nir_iadd(&b
->nb
, offset
, col_offset
);
623 _vtn_block_load_store(b
, op
, load
, index
, offset
,
624 access_offset
, access_size
,
625 chain
, chain_idx
+ 1,
626 type
->array_element
, inout
);
628 } else if (chain
== NULL
) {
629 /* Single whole vector */
630 assert(glsl_type_is_vector_or_scalar(type
->type
));
631 _vtn_load_store_tail(b
, op
, load
, index
, offset
,
632 access_offset
, access_size
,
635 /* Single component of a vector. Fall through to array case. */
636 nir_ssa_def
*elem_offset
=
637 vtn_access_link_as_ssa(b
, chain
->link
[chain_idx
], type
->stride
);
638 offset
= nir_iadd(&b
->nb
, offset
, elem_offset
);
640 _vtn_block_load_store(b
, op
, load
, index
, offset
,
641 access_offset
, access_size
,
643 type
->array_element
, inout
);
647 case GLSL_TYPE_ARRAY
: {
648 unsigned elems
= glsl_get_length(type
->type
);
649 for (unsigned i
= 0; i
< elems
; i
++) {
650 nir_ssa_def
*elem_off
=
651 nir_iadd(&b
->nb
, offset
, nir_imm_int(&b
->nb
, i
* type
->stride
));
652 _vtn_block_load_store(b
, op
, load
, index
, elem_off
,
653 access_offset
, access_size
,
655 type
->array_element
, &(*inout
)->elems
[i
]);
660 case GLSL_TYPE_STRUCT
: {
661 unsigned elems
= glsl_get_length(type
->type
);
662 for (unsigned i
= 0; i
< elems
; i
++) {
663 nir_ssa_def
*elem_off
=
664 nir_iadd(&b
->nb
, offset
, nir_imm_int(&b
->nb
, type
->offsets
[i
]));
665 _vtn_block_load_store(b
, op
, load
, index
, elem_off
,
666 access_offset
, access_size
,
668 type
->members
[i
], &(*inout
)->elems
[i
]);
674 unreachable("Invalid block member type");
678 static struct vtn_ssa_value
*
679 vtn_block_load(struct vtn_builder
*b
, struct vtn_access_chain
*src
)
682 unsigned access_offset
= 0, access_size
= 0;
683 switch (src
->var
->mode
) {
684 case vtn_variable_mode_ubo
:
685 op
= nir_intrinsic_load_ubo
;
687 case vtn_variable_mode_ssbo
:
688 op
= nir_intrinsic_load_ssbo
;
690 case vtn_variable_mode_push_constant
:
691 op
= nir_intrinsic_load_push_constant
;
692 vtn_access_chain_get_offset_size(src
, &access_offset
, &access_size
);
695 assert(!"Invalid block variable mode");
698 nir_ssa_def
*offset
, *index
= NULL
;
699 struct vtn_type
*type
;
701 offset
= vtn_access_chain_to_offset(b
, src
, &index
, &type
, &chain_idx
, true);
703 struct vtn_ssa_value
*value
= NULL
;
704 _vtn_block_load_store(b
, op
, true, index
, offset
,
705 access_offset
, access_size
,
706 src
, chain_idx
, type
, &value
);
711 vtn_block_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
712 struct vtn_access_chain
*dst
)
714 nir_ssa_def
*offset
, *index
= NULL
;
715 struct vtn_type
*type
;
717 offset
= vtn_access_chain_to_offset(b
, dst
, &index
, &type
, &chain_idx
, true);
719 _vtn_block_load_store(b
, nir_intrinsic_store_ssbo
, false, index
, offset
,
720 0, 0, dst
, chain_idx
, type
, &src
);
724 vtn_variable_is_external_block(struct vtn_variable
*var
)
726 return var
->mode
== vtn_variable_mode_ssbo
||
727 var
->mode
== vtn_variable_mode_ubo
||
728 var
->mode
== vtn_variable_mode_push_constant
;
732 _vtn_variable_load_store(struct vtn_builder
*b
, bool load
,
733 struct vtn_access_chain
*chain
,
734 struct vtn_type
*tail_type
,
735 struct vtn_ssa_value
**inout
)
737 enum glsl_base_type base_type
= glsl_get_base_type(tail_type
->type
);
741 case GLSL_TYPE_FLOAT
:
743 case GLSL_TYPE_DOUBLE
:
744 /* At this point, we have a scalar, vector, or matrix so we know that
745 * there cannot be any structure splitting still in the way. By
746 * stopping at the matrix level rather than the vector level, we
747 * ensure that matrices get loaded in the optimal way even if they
748 * are storred row-major in a UBO.
751 *inout
= vtn_local_load(b
, vtn_access_chain_to_deref(b
, chain
));
753 vtn_local_store(b
, *inout
, vtn_access_chain_to_deref(b
, chain
));
757 case GLSL_TYPE_ARRAY
:
758 case GLSL_TYPE_STRUCT
: {
759 struct vtn_access_chain
*new_chain
=
760 vtn_access_chain_extend(b
, chain
, 1);
761 new_chain
->link
[chain
->length
].mode
= vtn_access_mode_literal
;
762 unsigned elems
= glsl_get_length(tail_type
->type
);
764 assert(*inout
== NULL
);
765 *inout
= rzalloc(b
, struct vtn_ssa_value
);
766 (*inout
)->type
= tail_type
->type
;
767 (*inout
)->elems
= rzalloc_array(b
, struct vtn_ssa_value
*, elems
);
769 for (unsigned i
= 0; i
< elems
; i
++) {
770 new_chain
->link
[chain
->length
].id
= i
;
771 struct vtn_type
*elem_type
= base_type
== GLSL_TYPE_ARRAY
?
772 tail_type
->array_element
: tail_type
->members
[i
];
773 _vtn_variable_load_store(b
, load
, new_chain
, elem_type
,
774 &(*inout
)->elems
[i
]);
780 unreachable("Invalid access chain type");
784 struct vtn_ssa_value
*
785 vtn_variable_load(struct vtn_builder
*b
, struct vtn_access_chain
*src
)
787 if (vtn_variable_is_external_block(src
->var
)) {
788 return vtn_block_load(b
, src
);
790 struct vtn_type
*tail_type
= vtn_access_chain_tail_type(b
, src
);
791 struct vtn_ssa_value
*val
= NULL
;
792 _vtn_variable_load_store(b
, true, src
, tail_type
, &val
);
798 vtn_variable_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
799 struct vtn_access_chain
*dest
)
801 if (vtn_variable_is_external_block(dest
->var
)) {
802 assert(dest
->var
->mode
== vtn_variable_mode_ssbo
);
803 vtn_block_store(b
, src
, dest
);
805 struct vtn_type
*tail_type
= vtn_access_chain_tail_type(b
, dest
);
806 _vtn_variable_load_store(b
, false, dest
, tail_type
, &src
);
811 _vtn_variable_copy(struct vtn_builder
*b
, struct vtn_access_chain
*dest
,
812 struct vtn_access_chain
*src
, struct vtn_type
*tail_type
)
814 enum glsl_base_type base_type
= glsl_get_base_type(tail_type
->type
);
818 case GLSL_TYPE_FLOAT
:
819 case GLSL_TYPE_DOUBLE
:
821 /* At this point, we have a scalar, vector, or matrix so we know that
822 * there cannot be any structure splitting still in the way. By
823 * stopping at the matrix level rather than the vector level, we
824 * ensure that matrices get loaded in the optimal way even if they
825 * are storred row-major in a UBO.
827 vtn_variable_store(b
, vtn_variable_load(b
, src
), dest
);
830 case GLSL_TYPE_ARRAY
:
831 case GLSL_TYPE_STRUCT
: {
832 struct vtn_access_chain
*new_src
, *new_dest
;
833 new_src
= vtn_access_chain_extend(b
, src
, 1);
834 new_dest
= vtn_access_chain_extend(b
, dest
, 1);
835 new_src
->link
[src
->length
].mode
= vtn_access_mode_literal
;
836 new_dest
->link
[dest
->length
].mode
= vtn_access_mode_literal
;
837 unsigned elems
= glsl_get_length(tail_type
->type
);
838 for (unsigned i
= 0; i
< elems
; i
++) {
839 new_src
->link
[src
->length
].id
= i
;
840 new_dest
->link
[dest
->length
].id
= i
;
841 struct vtn_type
*elem_type
= base_type
== GLSL_TYPE_ARRAY
?
842 tail_type
->array_element
: tail_type
->members
[i
];
843 _vtn_variable_copy(b
, new_dest
, new_src
, elem_type
);
849 unreachable("Invalid access chain type");
854 vtn_variable_copy(struct vtn_builder
*b
, struct vtn_access_chain
*dest
,
855 struct vtn_access_chain
*src
)
857 struct vtn_type
*tail_type
= vtn_access_chain_tail_type(b
, src
);
858 assert(vtn_access_chain_tail_type(b
, dest
)->type
== tail_type
->type
);
860 /* TODO: At some point, we should add a special-case for when we can
861 * just emit a copy_var intrinsic.
863 _vtn_variable_copy(b
, dest
, src
, tail_type
);
867 set_mode_system_value(nir_variable_mode
*mode
)
869 assert(*mode
== nir_var_system_value
|| *mode
== nir_var_shader_in
);
870 *mode
= nir_var_system_value
;
874 vtn_get_builtin_location(struct vtn_builder
*b
,
875 SpvBuiltIn builtin
, int *location
,
876 nir_variable_mode
*mode
)
879 case SpvBuiltInPosition
:
880 *location
= VARYING_SLOT_POS
;
882 case SpvBuiltInPointSize
:
883 *location
= VARYING_SLOT_PSIZ
;
885 case SpvBuiltInClipDistance
:
886 *location
= VARYING_SLOT_CLIP_DIST0
; /* XXX CLIP_DIST1? */
888 case SpvBuiltInCullDistance
:
889 *location
= VARYING_SLOT_CULL_DIST0
;
891 case SpvBuiltInVertexIndex
:
892 *location
= SYSTEM_VALUE_VERTEX_ID
;
893 set_mode_system_value(mode
);
895 case SpvBuiltInVertexId
:
896 /* Vulkan defines VertexID to be zero-based and reserves the new
897 * builtin keyword VertexIndex to indicate the non-zero-based value.
899 *location
= SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
;
900 set_mode_system_value(mode
);
902 case SpvBuiltInInstanceIndex
:
903 *location
= SYSTEM_VALUE_INSTANCE_INDEX
;
904 set_mode_system_value(mode
);
906 case SpvBuiltInInstanceId
:
907 *location
= SYSTEM_VALUE_INSTANCE_ID
;
908 set_mode_system_value(mode
);
910 case SpvBuiltInPrimitiveId
:
911 if (b
->shader
->stage
== MESA_SHADER_FRAGMENT
) {
912 assert(*mode
== nir_var_shader_in
);
913 *location
= VARYING_SLOT_PRIMITIVE_ID
;
914 } else if (*mode
== nir_var_shader_out
) {
915 *location
= VARYING_SLOT_PRIMITIVE_ID
;
917 *location
= SYSTEM_VALUE_PRIMITIVE_ID
;
918 set_mode_system_value(mode
);
921 case SpvBuiltInInvocationId
:
922 *location
= SYSTEM_VALUE_INVOCATION_ID
;
923 set_mode_system_value(mode
);
925 case SpvBuiltInLayer
:
926 *location
= VARYING_SLOT_LAYER
;
927 if (b
->shader
->stage
== MESA_SHADER_FRAGMENT
)
928 *mode
= nir_var_shader_in
;
929 else if (b
->shader
->stage
== MESA_SHADER_GEOMETRY
)
930 *mode
= nir_var_shader_out
;
932 unreachable("invalid stage for SpvBuiltInLayer");
934 case SpvBuiltInViewportIndex
:
935 *location
= VARYING_SLOT_VIEWPORT
;
936 if (b
->shader
->stage
== MESA_SHADER_GEOMETRY
)
937 *mode
= nir_var_shader_out
;
938 else if (b
->shader
->stage
== MESA_SHADER_FRAGMENT
)
939 *mode
= nir_var_shader_in
;
941 unreachable("invalid stage for SpvBuiltInViewportIndex");
943 case SpvBuiltInTessLevelOuter
:
944 *location
= VARYING_SLOT_TESS_LEVEL_OUTER
;
946 case SpvBuiltInTessLevelInner
:
947 *location
= VARYING_SLOT_TESS_LEVEL_INNER
;
949 case SpvBuiltInTessCoord
:
950 *location
= SYSTEM_VALUE_TESS_COORD
;
951 set_mode_system_value(mode
);
953 case SpvBuiltInPatchVertices
:
954 *location
= SYSTEM_VALUE_VERTICES_IN
;
955 set_mode_system_value(mode
);
957 case SpvBuiltInFragCoord
:
958 *location
= VARYING_SLOT_POS
;
959 assert(*mode
== nir_var_shader_in
);
961 case SpvBuiltInPointCoord
:
962 *location
= VARYING_SLOT_PNTC
;
963 assert(*mode
== nir_var_shader_in
);
965 case SpvBuiltInFrontFacing
:
966 *location
= SYSTEM_VALUE_FRONT_FACE
;
967 set_mode_system_value(mode
);
969 case SpvBuiltInSampleId
:
970 *location
= SYSTEM_VALUE_SAMPLE_ID
;
971 set_mode_system_value(mode
);
973 case SpvBuiltInSamplePosition
:
974 *location
= SYSTEM_VALUE_SAMPLE_POS
;
975 set_mode_system_value(mode
);
977 case SpvBuiltInSampleMask
:
978 if (*mode
== nir_var_shader_out
) {
979 *location
= FRAG_RESULT_SAMPLE_MASK
;
981 *location
= SYSTEM_VALUE_SAMPLE_MASK_IN
;
982 set_mode_system_value(mode
);
985 case SpvBuiltInFragDepth
:
986 *location
= FRAG_RESULT_DEPTH
;
987 assert(*mode
== nir_var_shader_out
);
989 case SpvBuiltInNumWorkgroups
:
990 *location
= SYSTEM_VALUE_NUM_WORK_GROUPS
;
991 set_mode_system_value(mode
);
993 case SpvBuiltInWorkgroupSize
:
994 /* This should already be handled */
995 unreachable("unsupported builtin");
997 case SpvBuiltInWorkgroupId
:
998 *location
= SYSTEM_VALUE_WORK_GROUP_ID
;
999 set_mode_system_value(mode
);
1001 case SpvBuiltInLocalInvocationId
:
1002 *location
= SYSTEM_VALUE_LOCAL_INVOCATION_ID
;
1003 set_mode_system_value(mode
);
1005 case SpvBuiltInLocalInvocationIndex
:
1006 *location
= SYSTEM_VALUE_LOCAL_INVOCATION_INDEX
;
1007 set_mode_system_value(mode
);
1009 case SpvBuiltInGlobalInvocationId
:
1010 *location
= SYSTEM_VALUE_GLOBAL_INVOCATION_ID
;
1011 set_mode_system_value(mode
);
1013 case SpvBuiltInHelperInvocation
:
1015 unreachable("unsupported builtin");
1020 apply_var_decoration(struct vtn_builder
*b
, nir_variable
*nir_var
,
1021 const struct vtn_decoration
*dec
)
1023 switch (dec
->decoration
) {
1024 case SpvDecorationRelaxedPrecision
:
1025 break; /* FIXME: Do nothing with this for now. */
1026 case SpvDecorationNoPerspective
:
1027 nir_var
->data
.interpolation
= INTERP_MODE_NOPERSPECTIVE
;
1029 case SpvDecorationFlat
:
1030 nir_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1032 case SpvDecorationCentroid
:
1033 nir_var
->data
.centroid
= true;
1035 case SpvDecorationSample
:
1036 nir_var
->data
.sample
= true;
1038 case SpvDecorationInvariant
:
1039 nir_var
->data
.invariant
= true;
1041 case SpvDecorationConstant
:
1042 assert(nir_var
->constant_initializer
!= NULL
);
1043 nir_var
->data
.read_only
= true;
1045 case SpvDecorationNonWritable
:
1046 nir_var
->data
.read_only
= true;
1048 case SpvDecorationComponent
:
1049 nir_var
->data
.location_frac
= dec
->literals
[0];
1051 case SpvDecorationIndex
:
1052 nir_var
->data
.index
= dec
->literals
[0];
1054 case SpvDecorationBuiltIn
: {
1055 SpvBuiltIn builtin
= dec
->literals
[0];
1057 if (builtin
== SpvBuiltInWorkgroupSize
) {
1058 /* This shouldn't be a builtin. It's actually a constant. */
1059 nir_var
->data
.mode
= nir_var_global
;
1060 nir_var
->data
.read_only
= true;
1062 nir_constant
*c
= rzalloc(nir_var
, nir_constant
);
1063 c
->values
[0].u32
[0] = b
->shader
->info
->cs
.local_size
[0];
1064 c
->values
[0].u32
[1] = b
->shader
->info
->cs
.local_size
[1];
1065 c
->values
[0].u32
[2] = b
->shader
->info
->cs
.local_size
[2];
1066 nir_var
->constant_initializer
= c
;
1070 nir_variable_mode mode
= nir_var
->data
.mode
;
1071 vtn_get_builtin_location(b
, builtin
, &nir_var
->data
.location
, &mode
);
1072 nir_var
->data
.mode
= mode
;
1075 case SpvBuiltInTessLevelOuter
:
1076 case SpvBuiltInTessLevelInner
:
1077 nir_var
->data
.compact
= true;
1079 case SpvBuiltInSamplePosition
:
1080 nir_var
->data
.origin_upper_left
= b
->origin_upper_left
;
1082 case SpvBuiltInFragCoord
:
1083 nir_var
->data
.pixel_center_integer
= b
->pixel_center_integer
;
1090 case SpvDecorationSpecId
:
1091 case SpvDecorationRowMajor
:
1092 case SpvDecorationColMajor
:
1093 case SpvDecorationMatrixStride
:
1094 case SpvDecorationRestrict
:
1095 case SpvDecorationAliased
:
1096 case SpvDecorationVolatile
:
1097 case SpvDecorationCoherent
:
1098 case SpvDecorationNonReadable
:
1099 case SpvDecorationUniform
:
1100 case SpvDecorationStream
:
1101 case SpvDecorationOffset
:
1102 case SpvDecorationLinkageAttributes
:
1103 break; /* Do nothing with these here */
1105 case SpvDecorationPatch
:
1106 nir_var
->data
.patch
= true;
1109 case SpvDecorationLocation
:
1110 unreachable("Handled above");
1112 case SpvDecorationBlock
:
1113 case SpvDecorationBufferBlock
:
1114 case SpvDecorationArrayStride
:
1115 case SpvDecorationGLSLShared
:
1116 case SpvDecorationGLSLPacked
:
1117 break; /* These can apply to a type but we don't care about them */
1119 case SpvDecorationBinding
:
1120 case SpvDecorationDescriptorSet
:
1121 case SpvDecorationNoContraction
:
1122 case SpvDecorationInputAttachmentIndex
:
1123 vtn_warn("Decoration not allowed for variable or structure member: %s",
1124 spirv_decoration_to_string(dec
->decoration
));
1127 case SpvDecorationXfbBuffer
:
1128 case SpvDecorationXfbStride
:
1129 vtn_warn("Vulkan does not have transform feedback: %s",
1130 spirv_decoration_to_string(dec
->decoration
));
1133 case SpvDecorationCPacked
:
1134 case SpvDecorationSaturatedConversion
:
1135 case SpvDecorationFuncParamAttr
:
1136 case SpvDecorationFPRoundingMode
:
1137 case SpvDecorationFPFastMathMode
:
1138 case SpvDecorationAlignment
:
1139 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1140 spirv_decoration_to_string(dec
->decoration
));
1144 unreachable("Unhandled decoration");
1149 var_is_patch_cb(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
1150 const struct vtn_decoration
*dec
, void *out_is_patch
)
1152 if (dec
->decoration
== SpvDecorationPatch
) {
1153 *((bool *) out_is_patch
) = true;
1158 var_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
1159 const struct vtn_decoration
*dec
, void *void_var
)
1161 struct vtn_variable
*vtn_var
= void_var
;
1163 /* Handle decorations that apply to a vtn_variable as a whole */
1164 switch (dec
->decoration
) {
1165 case SpvDecorationBinding
:
1166 vtn_var
->binding
= dec
->literals
[0];
1168 case SpvDecorationDescriptorSet
:
1169 vtn_var
->descriptor_set
= dec
->literals
[0];
1171 case SpvDecorationInputAttachmentIndex
:
1172 vtn_var
->input_attachment_index
= dec
->literals
[0];
1174 case SpvDecorationPatch
:
1175 vtn_var
->patch
= true;
1181 if (val
->value_type
== vtn_value_type_access_chain
) {
1182 assert(val
->access_chain
->length
== 0);
1183 assert(val
->access_chain
->var
== void_var
);
1184 assert(member
== -1);
1186 assert(val
->value_type
== vtn_value_type_type
);
1189 /* Location is odd. If applied to a split structure, we have to walk the
1190 * whole thing and accumulate the location. It's easier to handle as a
1193 if (dec
->decoration
== SpvDecorationLocation
) {
1194 unsigned location
= dec
->literals
[0];
1195 bool is_vertex_input
;
1196 if (b
->shader
->stage
== MESA_SHADER_FRAGMENT
&&
1197 vtn_var
->mode
== vtn_variable_mode_output
) {
1198 is_vertex_input
= false;
1199 location
+= FRAG_RESULT_DATA0
;
1200 } else if (b
->shader
->stage
== MESA_SHADER_VERTEX
&&
1201 vtn_var
->mode
== vtn_variable_mode_input
) {
1202 is_vertex_input
= true;
1203 location
+= VERT_ATTRIB_GENERIC0
;
1204 } else if (vtn_var
->mode
== vtn_variable_mode_input
||
1205 vtn_var
->mode
== vtn_variable_mode_output
) {
1206 is_vertex_input
= false;
1207 location
+= vtn_var
->patch
? VARYING_SLOT_PATCH0
: VARYING_SLOT_VAR0
;
1209 vtn_warn("Location must be on input or output variable");
1214 /* This handles the member and lone variable cases */
1215 vtn_var
->var
->data
.location
= location
;
1217 /* This handles the structure member case */
1218 assert(vtn_var
->members
);
1220 glsl_get_length(glsl_without_array(vtn_var
->type
->type
));
1221 for (unsigned i
= 0; i
< length
; i
++) {
1222 vtn_var
->members
[i
]->data
.location
= location
;
1224 glsl_count_attribute_slots(vtn_var
->members
[i
]->interface_type
,
1231 assert(member
<= 0);
1232 apply_var_decoration(b
, vtn_var
->var
, dec
);
1233 } else if (vtn_var
->members
) {
1235 assert(vtn_var
->members
);
1236 apply_var_decoration(b
, vtn_var
->members
[member
], dec
);
1239 glsl_get_length(glsl_without_array(vtn_var
->type
->type
));
1240 for (unsigned i
= 0; i
< length
; i
++)
1241 apply_var_decoration(b
, vtn_var
->members
[i
], dec
);
1244 /* A few variables, those with external storage, have no actual
1245 * nir_variables associated with them. Fortunately, all decorations
1246 * we care about for those variables are on the type only.
1248 assert(vtn_var
->mode
== vtn_variable_mode_ubo
||
1249 vtn_var
->mode
== vtn_variable_mode_ssbo
||
1250 vtn_var
->mode
== vtn_variable_mode_push_constant
);
1256 is_per_vertex_inout(const struct vtn_variable
*var
, gl_shader_stage stage
)
1258 if (var
->patch
|| !glsl_type_is_array(var
->type
->type
))
1261 if (var
->mode
== vtn_variable_mode_input
) {
1262 return stage
== MESA_SHADER_TESS_CTRL
||
1263 stage
== MESA_SHADER_TESS_EVAL
||
1264 stage
== MESA_SHADER_GEOMETRY
;
1267 if (var
->mode
== vtn_variable_mode_output
)
1268 return stage
== MESA_SHADER_TESS_CTRL
;
1274 vtn_handle_variables(struct vtn_builder
*b
, SpvOp opcode
,
1275 const uint32_t *w
, unsigned count
)
1278 case SpvOpVariable
: {
1279 struct vtn_variable
*var
= rzalloc(b
, struct vtn_variable
);
1280 var
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1282 var
->chain
.var
= var
;
1283 var
->chain
.length
= 0;
1285 struct vtn_value
*val
=
1286 vtn_push_value(b
, w
[2], vtn_value_type_access_chain
);
1287 val
->access_chain
= &var
->chain
;
1289 struct vtn_type
*without_array
= var
->type
;
1290 while(glsl_type_is_array(without_array
->type
))
1291 without_array
= without_array
->array_element
;
1293 nir_variable_mode nir_mode
;
1294 switch ((SpvStorageClass
)w
[3]) {
1295 case SpvStorageClassUniform
:
1296 case SpvStorageClassUniformConstant
:
1297 if (without_array
->block
) {
1298 var
->mode
= vtn_variable_mode_ubo
;
1299 b
->shader
->info
->num_ubos
++;
1300 } else if (without_array
->buffer_block
) {
1301 var
->mode
= vtn_variable_mode_ssbo
;
1302 b
->shader
->info
->num_ssbos
++;
1303 } else if (glsl_type_is_image(without_array
->type
)) {
1304 var
->mode
= vtn_variable_mode_image
;
1305 nir_mode
= nir_var_uniform
;
1306 b
->shader
->info
->num_images
++;
1307 } else if (glsl_type_is_sampler(without_array
->type
)) {
1308 var
->mode
= vtn_variable_mode_sampler
;
1309 nir_mode
= nir_var_uniform
;
1310 b
->shader
->info
->num_textures
++;
1312 assert(!"Invalid uniform variable type");
1315 case SpvStorageClassPushConstant
:
1316 var
->mode
= vtn_variable_mode_push_constant
;
1317 assert(b
->shader
->num_uniforms
== 0);
1318 b
->shader
->num_uniforms
= vtn_type_block_size(var
->type
);
1320 case SpvStorageClassInput
:
1321 var
->mode
= vtn_variable_mode_input
;
1322 nir_mode
= nir_var_shader_in
;
1324 case SpvStorageClassOutput
:
1325 var
->mode
= vtn_variable_mode_output
;
1326 nir_mode
= nir_var_shader_out
;
1328 case SpvStorageClassPrivate
:
1329 var
->mode
= vtn_variable_mode_global
;
1330 nir_mode
= nir_var_global
;
1332 case SpvStorageClassFunction
:
1333 var
->mode
= vtn_variable_mode_local
;
1334 nir_mode
= nir_var_local
;
1336 case SpvStorageClassWorkgroup
:
1337 var
->mode
= vtn_variable_mode_workgroup
;
1338 nir_mode
= nir_var_shared
;
1340 case SpvStorageClassCrossWorkgroup
:
1341 case SpvStorageClassGeneric
:
1342 case SpvStorageClassAtomicCounter
:
1344 unreachable("Unhandled variable storage class");
1347 switch (var
->mode
) {
1348 case vtn_variable_mode_local
:
1349 case vtn_variable_mode_global
:
1350 case vtn_variable_mode_image
:
1351 case vtn_variable_mode_sampler
:
1352 case vtn_variable_mode_workgroup
:
1353 /* For these, we create the variable normally */
1354 var
->var
= rzalloc(b
->shader
, nir_variable
);
1355 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
1356 var
->var
->type
= var
->type
->type
;
1357 var
->var
->data
.mode
= nir_mode
;
1359 switch (var
->mode
) {
1360 case vtn_variable_mode_image
:
1361 case vtn_variable_mode_sampler
:
1362 var
->var
->interface_type
= without_array
->type
;
1365 var
->var
->interface_type
= NULL
;
1370 case vtn_variable_mode_input
:
1371 case vtn_variable_mode_output
: {
1372 /* In order to know whether or not we're a per-vertex inout, we need
1373 * the patch qualifier. This means walking the variable decorations
1374 * early before we actually create any variables. Not a big deal.
1376 * GLSLang really likes to place decorations in the most interior
1377 * thing it possibly can. In particular, if you have a struct, it
1378 * will place the patch decorations on the struct members. This
1379 * should be handled by the variable splitting below just fine.
1381 * If you have an array-of-struct, things get even more weird as it
1382 * will place the patch decorations on the struct even though it's
1383 * inside an array and some of the members being patch and others not
1384 * makes no sense whatsoever. Since the only sensible thing is for
1385 * it to be all or nothing, we'll call it patch if any of the members
1386 * are declared patch.
1389 vtn_foreach_decoration(b
, val
, var_is_patch_cb
, &var
->patch
);
1390 if (glsl_type_is_array(var
->type
->type
) &&
1391 glsl_type_is_struct(without_array
->type
)) {
1392 vtn_foreach_decoration(b
, without_array
->val
,
1393 var_is_patch_cb
, &var
->patch
);
1396 /* For inputs and outputs, we immediately split structures. This
1397 * is for a couple of reasons. For one, builtins may all come in
1398 * a struct and we really want those split out into separate
1399 * variables. For another, interpolation qualifiers can be
1400 * applied to members of the top-level struct ane we need to be
1401 * able to preserve that information.
1404 int array_length
= -1;
1405 struct vtn_type
*interface_type
= var
->type
;
1406 if (is_per_vertex_inout(var
, b
->shader
->stage
)) {
1407 /* In Geometry shaders (and some tessellation), inputs come
1408 * in per-vertex arrays. However, some builtins come in
1409 * non-per-vertex, hence the need for the is_array check. In
1410 * any case, there are no non-builtin arrays allowed so this
1411 * check should be sufficient.
1413 interface_type
= var
->type
->array_element
;
1414 array_length
= glsl_get_length(var
->type
->type
);
1417 if (glsl_type_is_struct(interface_type
->type
)) {
1418 /* It's a struct. Split it. */
1419 unsigned num_members
= glsl_get_length(interface_type
->type
);
1420 var
->members
= ralloc_array(b
, nir_variable
*, num_members
);
1422 for (unsigned i
= 0; i
< num_members
; i
++) {
1423 const struct glsl_type
*mtype
= interface_type
->members
[i
]->type
;
1424 if (array_length
>= 0)
1425 mtype
= glsl_array_type(mtype
, array_length
);
1427 var
->members
[i
] = rzalloc(b
->shader
, nir_variable
);
1428 var
->members
[i
]->name
=
1429 ralloc_asprintf(var
->members
[i
], "%s.%d", val
->name
, i
);
1430 var
->members
[i
]->type
= mtype
;
1431 var
->members
[i
]->interface_type
=
1432 interface_type
->members
[i
]->type
;
1433 var
->members
[i
]->data
.mode
= nir_mode
;
1434 var
->members
[i
]->data
.patch
= var
->patch
;
1437 var
->var
= rzalloc(b
->shader
, nir_variable
);
1438 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
1439 var
->var
->type
= var
->type
->type
;
1440 var
->var
->interface_type
= interface_type
->type
;
1441 var
->var
->data
.mode
= nir_mode
;
1442 var
->var
->data
.patch
= var
->patch
;
1445 /* For inputs and outputs, we need to grab locations and builtin
1446 * information from the interface type.
1448 vtn_foreach_decoration(b
, interface_type
->val
, var_decoration_cb
, var
);
1451 case vtn_variable_mode_param
:
1452 unreachable("Not created through OpVariable");
1455 case vtn_variable_mode_ubo
:
1456 case vtn_variable_mode_ssbo
:
1457 case vtn_variable_mode_push_constant
:
1458 /* These don't need actual variables. */
1464 nir_constant
*constant
=
1465 vtn_value(b
, w
[4], vtn_value_type_constant
)->constant
;
1466 var
->var
->constant_initializer
=
1467 nir_constant_clone(constant
, var
->var
);
1470 vtn_foreach_decoration(b
, val
, var_decoration_cb
, var
);
1472 if (var
->mode
== vtn_variable_mode_image
||
1473 var
->mode
== vtn_variable_mode_sampler
) {
1474 /* XXX: We still need the binding information in the nir_variable
1475 * for these. We should fix that.
1477 var
->var
->data
.binding
= var
->binding
;
1478 var
->var
->data
.descriptor_set
= var
->descriptor_set
;
1479 var
->var
->data
.index
= var
->input_attachment_index
;
1481 if (var
->mode
== vtn_variable_mode_image
)
1482 var
->var
->data
.image
.format
= without_array
->image_format
;
1485 if (var
->mode
== vtn_variable_mode_local
) {
1486 assert(var
->members
== NULL
&& var
->var
!= NULL
);
1487 nir_function_impl_add_variable(b
->impl
, var
->var
);
1488 } else if (var
->var
) {
1489 nir_shader_add_variable(b
->shader
, var
->var
);
1490 } else if (var
->members
) {
1491 unsigned count
= glsl_get_length(without_array
->type
);
1492 for (unsigned i
= 0; i
< count
; i
++) {
1493 assert(var
->members
[i
]->data
.mode
!= nir_var_local
);
1494 nir_shader_add_variable(b
->shader
, var
->members
[i
]);
1497 assert(var
->mode
== vtn_variable_mode_ubo
||
1498 var
->mode
== vtn_variable_mode_ssbo
||
1499 var
->mode
== vtn_variable_mode_push_constant
);
1504 case SpvOpAccessChain
:
1505 case SpvOpInBoundsAccessChain
: {
1506 struct vtn_access_chain
*base
, *chain
;
1507 struct vtn_value
*base_val
= vtn_untyped_value(b
, w
[3]);
1508 if (base_val
->value_type
== vtn_value_type_sampled_image
) {
1509 /* This is rather insane. SPIR-V allows you to use OpSampledImage
1510 * to combine an array of images with a single sampler to get an
1511 * array of sampled images that all share the same sampler.
1512 * Fortunately, this means that we can more-or-less ignore the
1513 * sampler when crawling the access chain, but it does leave us
1514 * with this rather awkward little special-case.
1516 base
= base_val
->sampled_image
->image
;
1518 assert(base_val
->value_type
== vtn_value_type_access_chain
);
1519 base
= base_val
->access_chain
;
1522 chain
= vtn_access_chain_extend(b
, base
, count
- 4);
1524 unsigned idx
= base
->length
;
1525 for (int i
= 4; i
< count
; i
++) {
1526 struct vtn_value
*link_val
= vtn_untyped_value(b
, w
[i
]);
1527 if (link_val
->value_type
== vtn_value_type_constant
) {
1528 chain
->link
[idx
].mode
= vtn_access_mode_literal
;
1529 chain
->link
[idx
].id
= link_val
->constant
->values
[0].u32
[0];
1531 chain
->link
[idx
].mode
= vtn_access_mode_id
;
1532 chain
->link
[idx
].id
= w
[i
];
1537 if (base_val
->value_type
== vtn_value_type_sampled_image
) {
1538 struct vtn_value
*val
=
1539 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
1540 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
1541 val
->sampled_image
->image
= chain
;
1542 val
->sampled_image
->sampler
= base_val
->sampled_image
->sampler
;
1544 struct vtn_value
*val
=
1545 vtn_push_value(b
, w
[2], vtn_value_type_access_chain
);
1546 val
->access_chain
= chain
;
1551 case SpvOpCopyMemory
: {
1552 struct vtn_value
*dest
= vtn_value(b
, w
[1], vtn_value_type_access_chain
);
1553 struct vtn_value
*src
= vtn_value(b
, w
[2], vtn_value_type_access_chain
);
1555 vtn_variable_copy(b
, dest
->access_chain
, src
->access_chain
);
1560 struct vtn_access_chain
*src
=
1561 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1563 if (src
->var
->mode
== vtn_variable_mode_image
||
1564 src
->var
->mode
== vtn_variable_mode_sampler
) {
1565 vtn_push_value(b
, w
[2], vtn_value_type_access_chain
)->access_chain
= src
;
1569 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1570 val
->ssa
= vtn_variable_load(b
, src
);
1575 struct vtn_access_chain
*dest
=
1576 vtn_value(b
, w
[1], vtn_value_type_access_chain
)->access_chain
;
1577 struct vtn_ssa_value
*src
= vtn_ssa_value(b
, w
[2]);
1578 vtn_variable_store(b
, src
, dest
);
1582 case SpvOpArrayLength
: {
1583 struct vtn_access_chain
*chain
=
1584 vtn_value(b
, w
[3], vtn_value_type_access_chain
)->access_chain
;
1586 const uint32_t offset
= chain
->var
->type
->offsets
[w
[4]];
1587 const uint32_t stride
= chain
->var
->type
->members
[w
[4]]->stride
;
1590 struct vtn_type
*type
;
1591 nir_ssa_def
*index
=
1592 get_vulkan_resource_index(b
, chain
, &type
, &chain_idx
);
1594 nir_intrinsic_instr
*instr
=
1595 nir_intrinsic_instr_create(b
->nb
.shader
,
1596 nir_intrinsic_get_buffer_size
);
1597 instr
->src
[0] = nir_src_for_ssa(index
);
1598 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
1599 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
1600 nir_ssa_def
*buf_size
= &instr
->dest
.ssa
;
1602 /* array_length = max(buffer_size - offset, 0) / stride */
1603 nir_ssa_def
*array_length
=
1608 nir_imm_int(&b
->nb
, offset
)),
1609 nir_imm_int(&b
->nb
, 0u)),
1610 nir_imm_int(&b
->nb
, stride
));
1612 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1613 val
->ssa
= vtn_create_ssa_value(b
, glsl_uint_type());
1614 val
->ssa
->def
= array_length
;
1618 case SpvOpCopyMemorySized
:
1620 unreachable("Unhandled opcode");