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 old_len
= old
? old
->length
: 0;
38 unsigned new_len
= old_len
+ new_ids
;
39 /* TODO: don't use rzalloc */
40 chain
= rzalloc_size(b
, sizeof(*chain
) + new_len
* sizeof(chain
->link
[0]));
42 chain
->length
= new_len
;
44 for (unsigned i
= 0; i
< old_len
; 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 /* Crawls a chain of array derefs and rewrites the types so that the
66 * lengths stay the same but the terminal type is the one given by
67 * tail_type. This is useful for split structures.
70 rewrite_deref_types(nir_deref
*deref
, const struct glsl_type
*type
)
74 assert(deref
->child
->deref_type
== nir_deref_type_array
);
75 assert(glsl_type_is_array(deref
->type
));
76 rewrite_deref_types(deref
->child
, glsl_get_array_element(type
));
81 vtn_pointer_for_variable(struct vtn_builder
*b
,
82 struct vtn_variable
*var
)
84 struct vtn_pointer
*pointer
= rzalloc(b
, struct vtn_pointer
);
86 pointer
->mode
= var
->mode
;
87 pointer
->type
= var
->type
;
94 vtn_pointer_to_deref(struct vtn_builder
*b
, struct vtn_pointer
*ptr
)
96 /* Do on-the-fly copy propagation for samplers. */
97 if (ptr
->var
->copy_prop_sampler
)
98 return vtn_pointer_to_deref(b
, ptr
->var
->copy_prop_sampler
);
100 nir_deref_var
*deref_var
;
102 deref_var
= nir_deref_var_create(b
, ptr
->var
->var
);
103 /* Raw variable access */
107 assert(ptr
->var
->members
);
108 /* Create the deref_var manually. It will get filled out later. */
109 deref_var
= rzalloc(b
, nir_deref_var
);
110 deref_var
->deref
.deref_type
= nir_deref_type_var
;
113 struct vtn_access_chain
*chain
= ptr
->chain
;
116 struct vtn_type
*deref_type
= ptr
->var
->type
;
117 nir_deref
*tail
= &deref_var
->deref
;
118 nir_variable
**members
= ptr
->var
->members
;
120 for (unsigned i
= 0; i
< chain
->length
; i
++) {
121 enum glsl_base_type base_type
= glsl_get_base_type(deref_type
->type
);
125 case GLSL_TYPE_UINT64
:
126 case GLSL_TYPE_INT64
:
127 case GLSL_TYPE_FLOAT
:
128 case GLSL_TYPE_DOUBLE
:
130 case GLSL_TYPE_ARRAY
: {
131 deref_type
= deref_type
->array_element
;
133 nir_deref_array
*deref_arr
= nir_deref_array_create(b
);
134 deref_arr
->deref
.type
= deref_type
->type
;
136 if (chain
->link
[i
].mode
== vtn_access_mode_literal
) {
137 deref_arr
->deref_array_type
= nir_deref_array_type_direct
;
138 deref_arr
->base_offset
= chain
->link
[i
].id
;
140 assert(chain
->link
[i
].mode
== vtn_access_mode_id
);
141 deref_arr
->deref_array_type
= nir_deref_array_type_indirect
;
142 deref_arr
->base_offset
= 0;
143 deref_arr
->indirect
=
144 nir_src_for_ssa(vtn_ssa_value(b
, chain
->link
[i
].id
)->def
);
146 tail
->child
= &deref_arr
->deref
;
151 case GLSL_TYPE_STRUCT
: {
152 assert(chain
->link
[i
].mode
== vtn_access_mode_literal
);
153 unsigned idx
= chain
->link
[i
].id
;
154 deref_type
= deref_type
->members
[idx
];
156 /* This is a pre-split structure. */
157 deref_var
->var
= members
[idx
];
158 rewrite_deref_types(&deref_var
->deref
, members
[idx
]->type
);
159 assert(tail
->type
== deref_type
->type
);
162 nir_deref_struct
*deref_struct
= nir_deref_struct_create(b
, idx
);
163 deref_struct
->deref
.type
= deref_type
->type
;
164 tail
->child
= &deref_struct
->deref
;
170 unreachable("Invalid type for deref");
174 assert(members
== NULL
);
179 _vtn_local_load_store(struct vtn_builder
*b
, bool load
, nir_deref_var
*deref
,
180 nir_deref
*tail
, struct vtn_ssa_value
*inout
)
182 /* The deref tail may contain a deref to select a component of a vector (in
183 * other words, it might not be an actual tail) so we have to save it away
184 * here since we overwrite it later.
186 nir_deref
*old_child
= tail
->child
;
188 if (glsl_type_is_vector_or_scalar(tail
->type
)) {
189 /* Terminate the deref chain in case there is one more link to pick
190 * off a component of the vector.
194 nir_intrinsic_op op
= load
? nir_intrinsic_load_var
:
195 nir_intrinsic_store_var
;
197 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(b
->shader
, op
);
198 intrin
->variables
[0] = nir_deref_var_clone(deref
, intrin
);
199 intrin
->num_components
= glsl_get_vector_elements(tail
->type
);
202 nir_ssa_dest_init(&intrin
->instr
, &intrin
->dest
,
203 intrin
->num_components
,
204 glsl_get_bit_size(tail
->type
),
206 inout
->def
= &intrin
->dest
.ssa
;
208 nir_intrinsic_set_write_mask(intrin
, (1 << intrin
->num_components
) - 1);
209 intrin
->src
[0] = nir_src_for_ssa(inout
->def
);
212 nir_builder_instr_insert(&b
->nb
, &intrin
->instr
);
213 } else if (glsl_get_base_type(tail
->type
) == GLSL_TYPE_ARRAY
||
214 glsl_type_is_matrix(tail
->type
)) {
215 unsigned elems
= glsl_get_length(tail
->type
);
216 nir_deref_array
*deref_arr
= nir_deref_array_create(b
);
217 deref_arr
->deref_array_type
= nir_deref_array_type_direct
;
218 deref_arr
->deref
.type
= glsl_get_array_element(tail
->type
);
219 tail
->child
= &deref_arr
->deref
;
220 for (unsigned i
= 0; i
< elems
; i
++) {
221 deref_arr
->base_offset
= i
;
222 _vtn_local_load_store(b
, load
, deref
, tail
->child
, inout
->elems
[i
]);
225 assert(glsl_get_base_type(tail
->type
) == GLSL_TYPE_STRUCT
);
226 unsigned elems
= glsl_get_length(tail
->type
);
227 nir_deref_struct
*deref_struct
= nir_deref_struct_create(b
, 0);
228 tail
->child
= &deref_struct
->deref
;
229 for (unsigned i
= 0; i
< elems
; i
++) {
230 deref_struct
->index
= i
;
231 deref_struct
->deref
.type
= glsl_get_struct_field(tail
->type
, i
);
232 _vtn_local_load_store(b
, load
, deref
, tail
->child
, inout
->elems
[i
]);
236 tail
->child
= old_child
;
240 vtn_nir_deref(struct vtn_builder
*b
, uint32_t id
)
242 struct vtn_pointer
*ptr
= vtn_value(b
, id
, vtn_value_type_pointer
)->pointer
;
243 return vtn_pointer_to_deref(b
, ptr
);
247 * Gets the NIR-level deref tail, which may have as a child an array deref
248 * selecting which component due to OpAccessChain supporting per-component
249 * indexing in SPIR-V.
252 get_deref_tail(nir_deref_var
*deref
)
254 nir_deref
*cur
= &deref
->deref
;
255 while (!glsl_type_is_vector_or_scalar(cur
->type
) && cur
->child
)
261 struct vtn_ssa_value
*
262 vtn_local_load(struct vtn_builder
*b
, nir_deref_var
*src
)
264 nir_deref
*src_tail
= get_deref_tail(src
);
265 struct vtn_ssa_value
*val
= vtn_create_ssa_value(b
, src_tail
->type
);
266 _vtn_local_load_store(b
, true, src
, src_tail
, val
);
268 if (src_tail
->child
) {
269 nir_deref_array
*vec_deref
= nir_deref_as_array(src_tail
->child
);
270 assert(vec_deref
->deref
.child
== NULL
);
271 val
->type
= vec_deref
->deref
.type
;
272 if (vec_deref
->deref_array_type
== nir_deref_array_type_direct
)
273 val
->def
= vtn_vector_extract(b
, val
->def
, vec_deref
->base_offset
);
275 val
->def
= vtn_vector_extract_dynamic(b
, val
->def
,
276 vec_deref
->indirect
.ssa
);
283 vtn_local_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
286 nir_deref
*dest_tail
= get_deref_tail(dest
);
288 if (dest_tail
->child
) {
289 struct vtn_ssa_value
*val
= vtn_create_ssa_value(b
, dest_tail
->type
);
290 _vtn_local_load_store(b
, true, dest
, dest_tail
, val
);
291 nir_deref_array
*deref
= nir_deref_as_array(dest_tail
->child
);
292 assert(deref
->deref
.child
== NULL
);
293 if (deref
->deref_array_type
== nir_deref_array_type_direct
)
294 val
->def
= vtn_vector_insert(b
, val
->def
, src
->def
,
297 val
->def
= vtn_vector_insert_dynamic(b
, val
->def
, src
->def
,
298 deref
->indirect
.ssa
);
299 _vtn_local_load_store(b
, false, dest
, dest_tail
, val
);
301 _vtn_local_load_store(b
, false, dest
, dest_tail
, src
);
306 get_vulkan_resource_index(struct vtn_builder
*b
, struct vtn_pointer
*ptr
,
307 struct vtn_type
**type
, unsigned *chain_idx
)
309 /* Push constants have no explicit binding */
310 if (ptr
->mode
== vtn_variable_mode_push_constant
) {
312 *type
= ptr
->var
->type
;
316 nir_ssa_def
*array_index
;
317 if (glsl_type_is_array(ptr
->var
->type
->type
)) {
318 assert(ptr
->chain
->length
> 0);
319 array_index
= vtn_access_link_as_ssa(b
, ptr
->chain
->link
[0], 1);
321 *type
= ptr
->var
->type
->array_element
;
323 array_index
= nir_imm_int(&b
->nb
, 0);
325 *type
= ptr
->var
->type
;
328 nir_intrinsic_instr
*instr
=
329 nir_intrinsic_instr_create(b
->nb
.shader
,
330 nir_intrinsic_vulkan_resource_index
);
331 instr
->src
[0] = nir_src_for_ssa(array_index
);
332 nir_intrinsic_set_desc_set(instr
, ptr
->var
->descriptor_set
);
333 nir_intrinsic_set_binding(instr
, ptr
->var
->binding
);
335 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
336 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
338 return &instr
->dest
.ssa
;
342 vtn_pointer_to_offset(struct vtn_builder
*b
, struct vtn_pointer
*ptr
,
343 nir_ssa_def
**index_out
, struct vtn_type
**type_out
,
344 unsigned *end_idx_out
, bool stop_at_matrix
)
347 struct vtn_type
*type
;
348 *index_out
= get_vulkan_resource_index(b
, ptr
, &type
, &idx
);
350 nir_ssa_def
*offset
= nir_imm_int(&b
->nb
, 0);
351 for (; idx
< ptr
->chain
->length
; idx
++) {
352 enum glsl_base_type base_type
= glsl_get_base_type(type
->type
);
356 case GLSL_TYPE_UINT64
:
357 case GLSL_TYPE_INT64
:
358 case GLSL_TYPE_FLOAT
:
359 case GLSL_TYPE_DOUBLE
:
361 /* Some users may not want matrix or vector derefs */
366 case GLSL_TYPE_ARRAY
:
367 offset
= nir_iadd(&b
->nb
, offset
,
368 vtn_access_link_as_ssa(b
, ptr
->chain
->link
[idx
],
371 type
= type
->array_element
;
374 case GLSL_TYPE_STRUCT
: {
375 assert(ptr
->chain
->link
[idx
].mode
== vtn_access_mode_literal
);
376 unsigned member
= ptr
->chain
->link
[idx
].id
;
377 offset
= nir_iadd(&b
->nb
, offset
,
378 nir_imm_int(&b
->nb
, type
->offsets
[member
]));
379 type
= type
->members
[member
];
384 unreachable("Invalid type for deref");
396 /* Tries to compute the size of an interface block based on the strides and
397 * offsets that are provided to us in the SPIR-V source.
400 vtn_type_block_size(struct vtn_type
*type
)
402 enum glsl_base_type base_type
= glsl_get_base_type(type
->type
);
406 case GLSL_TYPE_UINT64
:
407 case GLSL_TYPE_INT64
:
408 case GLSL_TYPE_FLOAT
:
410 case GLSL_TYPE_DOUBLE
: {
411 unsigned cols
= type
->row_major
? glsl_get_vector_elements(type
->type
) :
412 glsl_get_matrix_columns(type
->type
);
414 assert(type
->stride
> 0);
415 return type
->stride
* cols
;
416 } else if (base_type
== GLSL_TYPE_DOUBLE
||
417 base_type
== GLSL_TYPE_UINT64
||
418 base_type
== GLSL_TYPE_INT64
) {
419 return glsl_get_vector_elements(type
->type
) * 8;
421 return glsl_get_vector_elements(type
->type
) * 4;
425 case GLSL_TYPE_STRUCT
:
426 case GLSL_TYPE_INTERFACE
: {
428 unsigned num_fields
= glsl_get_length(type
->type
);
429 for (unsigned f
= 0; f
< num_fields
; f
++) {
430 unsigned field_end
= type
->offsets
[f
] +
431 vtn_type_block_size(type
->members
[f
]);
432 size
= MAX2(size
, field_end
);
437 case GLSL_TYPE_ARRAY
:
438 assert(type
->stride
> 0);
439 assert(glsl_get_length(type
->type
) > 0);
440 return type
->stride
* glsl_get_length(type
->type
);
443 assert(!"Invalid block type");
449 vtn_access_chain_get_offset_size(struct vtn_access_chain
*chain
,
450 struct vtn_type
*type
,
451 unsigned *access_offset
,
452 unsigned *access_size
)
456 for (unsigned i
= 0; i
< chain
->length
; i
++) {
457 if (chain
->link
[i
].mode
!= vtn_access_mode_literal
)
460 if (glsl_type_is_struct(type
->type
)) {
461 *access_offset
+= type
->offsets
[chain
->link
[i
].id
];
462 type
= type
->members
[chain
->link
[i
].id
];
464 *access_offset
+= type
->stride
* chain
->link
[i
].id
;
465 type
= type
->array_element
;
469 *access_size
= vtn_type_block_size(type
);
473 _vtn_load_store_tail(struct vtn_builder
*b
, nir_intrinsic_op op
, bool load
,
474 nir_ssa_def
*index
, nir_ssa_def
*offset
,
475 unsigned access_offset
, unsigned access_size
,
476 struct vtn_ssa_value
**inout
, const struct glsl_type
*type
)
478 nir_intrinsic_instr
*instr
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
479 instr
->num_components
= glsl_get_vector_elements(type
);
483 nir_intrinsic_set_write_mask(instr
, (1 << instr
->num_components
) - 1);
484 instr
->src
[src
++] = nir_src_for_ssa((*inout
)->def
);
487 if (op
== nir_intrinsic_load_push_constant
) {
488 assert(access_offset
% 4 == 0);
490 nir_intrinsic_set_base(instr
, access_offset
);
491 nir_intrinsic_set_range(instr
, access_size
);
495 instr
->src
[src
++] = nir_src_for_ssa(index
);
497 if (op
== nir_intrinsic_load_push_constant
) {
498 /* We need to subtract the offset from where the intrinsic will load the
501 nir_src_for_ssa(nir_isub(&b
->nb
, offset
,
502 nir_imm_int(&b
->nb
, access_offset
)));
504 instr
->src
[src
++] = nir_src_for_ssa(offset
);
508 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
509 instr
->num_components
,
510 glsl_get_bit_size(type
), NULL
);
511 (*inout
)->def
= &instr
->dest
.ssa
;
514 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
516 if (load
&& glsl_get_base_type(type
) == GLSL_TYPE_BOOL
)
517 (*inout
)->def
= nir_ine(&b
->nb
, (*inout
)->def
, nir_imm_int(&b
->nb
, 0));
521 _vtn_block_load_store(struct vtn_builder
*b
, nir_intrinsic_op op
, bool load
,
522 nir_ssa_def
*index
, nir_ssa_def
*offset
,
523 unsigned access_offset
, unsigned access_size
,
524 struct vtn_access_chain
*chain
, unsigned chain_idx
,
525 struct vtn_type
*type
, struct vtn_ssa_value
**inout
)
527 if (chain
&& chain_idx
>= chain
->length
)
530 if (load
&& chain
== NULL
&& *inout
== NULL
)
531 *inout
= vtn_create_ssa_value(b
, type
->type
);
533 enum glsl_base_type base_type
= glsl_get_base_type(type
->type
);
537 case GLSL_TYPE_UINT64
:
538 case GLSL_TYPE_INT64
:
539 case GLSL_TYPE_FLOAT
:
540 case GLSL_TYPE_DOUBLE
:
542 /* This is where things get interesting. At this point, we've hit
543 * a vector, a scalar, or a matrix.
545 if (glsl_type_is_matrix(type
->type
)) {
546 /* Loading the whole matrix */
547 struct vtn_ssa_value
*transpose
;
548 unsigned num_ops
, vec_width
, col_stride
;
549 if (type
->row_major
) {
550 num_ops
= glsl_get_vector_elements(type
->type
);
551 vec_width
= glsl_get_matrix_columns(type
->type
);
552 col_stride
= type
->array_element
->stride
;
554 const struct glsl_type
*transpose_type
=
555 glsl_matrix_type(base_type
, vec_width
, num_ops
);
556 *inout
= vtn_create_ssa_value(b
, transpose_type
);
558 transpose
= vtn_ssa_transpose(b
, *inout
);
562 num_ops
= glsl_get_matrix_columns(type
->type
);
563 vec_width
= glsl_get_vector_elements(type
->type
);
564 col_stride
= type
->stride
;
567 for (unsigned i
= 0; i
< num_ops
; i
++) {
568 nir_ssa_def
*elem_offset
=
569 nir_iadd(&b
->nb
, offset
, nir_imm_int(&b
->nb
, i
* col_stride
));
570 _vtn_load_store_tail(b
, op
, load
, index
, elem_offset
,
571 access_offset
, access_size
,
573 glsl_vector_type(base_type
, vec_width
));
576 if (load
&& type
->row_major
)
577 *inout
= vtn_ssa_transpose(b
, *inout
);
579 unsigned elems
= glsl_get_vector_elements(type
->type
);
580 unsigned type_size
= glsl_get_bit_size(type
->type
) / 8;
581 if (elems
== 1 || type
->stride
== type_size
) {
582 /* This is a tightly-packed normal scalar or vector load */
583 assert(glsl_type_is_vector_or_scalar(type
->type
));
584 _vtn_load_store_tail(b
, op
, load
, index
, offset
,
585 access_offset
, access_size
,
588 /* This is a strided load. We have to load N things separately.
589 * This is the single column of a row-major matrix case.
591 assert(type
->stride
> type_size
);
592 assert(type
->stride
% type_size
== 0);
594 nir_ssa_def
*per_comp
[4];
595 for (unsigned i
= 0; i
< elems
; i
++) {
596 nir_ssa_def
*elem_offset
=
597 nir_iadd(&b
->nb
, offset
,
598 nir_imm_int(&b
->nb
, i
* type
->stride
));
599 struct vtn_ssa_value
*comp
, temp_val
;
601 temp_val
.def
= nir_channel(&b
->nb
, (*inout
)->def
, i
);
602 temp_val
.type
= glsl_scalar_type(base_type
);
605 _vtn_load_store_tail(b
, op
, load
, index
, elem_offset
,
606 access_offset
, access_size
,
607 &comp
, glsl_scalar_type(base_type
));
608 per_comp
[i
] = comp
->def
;
613 *inout
= vtn_create_ssa_value(b
, type
->type
);
614 (*inout
)->def
= nir_vec(&b
->nb
, per_comp
, elems
);
620 case GLSL_TYPE_ARRAY
: {
621 unsigned elems
= glsl_get_length(type
->type
);
622 for (unsigned i
= 0; i
< elems
; i
++) {
623 nir_ssa_def
*elem_off
=
624 nir_iadd(&b
->nb
, offset
, nir_imm_int(&b
->nb
, i
* type
->stride
));
625 _vtn_block_load_store(b
, op
, load
, index
, elem_off
,
626 access_offset
, access_size
,
628 type
->array_element
, &(*inout
)->elems
[i
]);
633 case GLSL_TYPE_STRUCT
: {
634 unsigned elems
= glsl_get_length(type
->type
);
635 for (unsigned i
= 0; i
< elems
; i
++) {
636 nir_ssa_def
*elem_off
=
637 nir_iadd(&b
->nb
, offset
, nir_imm_int(&b
->nb
, type
->offsets
[i
]));
638 _vtn_block_load_store(b
, op
, load
, index
, elem_off
,
639 access_offset
, access_size
,
641 type
->members
[i
], &(*inout
)->elems
[i
]);
647 unreachable("Invalid block member type");
651 static struct vtn_ssa_value
*
652 vtn_block_load(struct vtn_builder
*b
, struct vtn_pointer
*src
)
655 unsigned access_offset
= 0, access_size
= 0;
657 case vtn_variable_mode_ubo
:
658 op
= nir_intrinsic_load_ubo
;
660 case vtn_variable_mode_ssbo
:
661 op
= nir_intrinsic_load_ssbo
;
663 case vtn_variable_mode_push_constant
:
664 op
= nir_intrinsic_load_push_constant
;
665 vtn_access_chain_get_offset_size(src
->chain
, src
->var
->type
,
666 &access_offset
, &access_size
);
669 assert(!"Invalid block variable mode");
672 nir_ssa_def
*offset
, *index
= NULL
;
674 struct vtn_type
*type
;
675 offset
= vtn_pointer_to_offset(b
, src
, &index
, &type
, &chain_idx
, false);
677 struct vtn_ssa_value
*value
= NULL
;
678 _vtn_block_load_store(b
, op
, true, index
, offset
,
679 access_offset
, access_size
,
680 src
->chain
, chain_idx
, type
, &value
);
685 vtn_block_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
686 struct vtn_pointer
*dst
)
688 nir_ssa_def
*offset
, *index
= NULL
;
690 struct vtn_type
*type
;
691 offset
= vtn_pointer_to_offset(b
, dst
, &index
, &type
, &chain_idx
, false);
693 _vtn_block_load_store(b
, nir_intrinsic_store_ssbo
, false, index
, offset
,
694 0, 0, dst
->chain
, chain_idx
, type
, &src
);
698 vtn_variable_is_external_block(struct vtn_variable
*var
)
700 return var
->mode
== vtn_variable_mode_ssbo
||
701 var
->mode
== vtn_variable_mode_ubo
||
702 var
->mode
== vtn_variable_mode_push_constant
;
706 _vtn_variable_load_store(struct vtn_builder
*b
, bool load
,
707 struct vtn_pointer
*ptr
,
708 struct vtn_ssa_value
**inout
)
710 enum glsl_base_type base_type
= glsl_get_base_type(ptr
->type
->type
);
714 case GLSL_TYPE_UINT64
:
715 case GLSL_TYPE_INT64
:
716 case GLSL_TYPE_FLOAT
:
718 case GLSL_TYPE_DOUBLE
:
719 /* At this point, we have a scalar, vector, or matrix so we know that
720 * there cannot be any structure splitting still in the way. By
721 * stopping at the matrix level rather than the vector level, we
722 * ensure that matrices get loaded in the optimal way even if they
723 * are storred row-major in a UBO.
726 *inout
= vtn_local_load(b
, vtn_pointer_to_deref(b
, ptr
));
728 vtn_local_store(b
, *inout
, vtn_pointer_to_deref(b
, ptr
));
732 case GLSL_TYPE_ARRAY
:
733 case GLSL_TYPE_STRUCT
: {
734 unsigned elems
= glsl_get_length(ptr
->type
->type
);
736 assert(*inout
== NULL
);
737 *inout
= rzalloc(b
, struct vtn_ssa_value
);
738 (*inout
)->type
= ptr
->type
->type
;
739 (*inout
)->elems
= rzalloc_array(b
, struct vtn_ssa_value
*, elems
);
742 struct vtn_pointer elem
= *ptr
;
743 elem
.chain
= vtn_access_chain_extend(b
, ptr
->chain
, 1);
744 unsigned link_idx
= ptr
->chain
? ptr
->chain
->length
: 0;
745 elem
.chain
->link
[link_idx
].mode
= vtn_access_mode_literal
;
746 for (unsigned i
= 0; i
< elems
; i
++) {
747 elem
.chain
->link
[link_idx
].id
= i
;
748 elem
.type
= (base_type
== GLSL_TYPE_ARRAY
) ? ptr
->type
->array_element
:
749 ptr
->type
->members
[i
];
750 _vtn_variable_load_store(b
, load
, &elem
, &(*inout
)->elems
[i
]);
756 unreachable("Invalid access chain type");
760 struct vtn_ssa_value
*
761 vtn_variable_load(struct vtn_builder
*b
, struct vtn_pointer
*src
)
763 if (vtn_variable_is_external_block(src
->var
)) {
764 return vtn_block_load(b
, src
);
766 struct vtn_ssa_value
*val
= NULL
;
767 _vtn_variable_load_store(b
, true, src
, &val
);
773 vtn_variable_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
774 struct vtn_pointer
*dest
)
776 if (vtn_variable_is_external_block(dest
->var
)) {
777 assert(dest
->var
->mode
== vtn_variable_mode_ssbo
);
778 vtn_block_store(b
, src
, dest
);
780 _vtn_variable_load_store(b
, false, dest
, &src
);
785 _vtn_variable_copy(struct vtn_builder
*b
, struct vtn_pointer
*dest
,
786 struct vtn_pointer
*src
)
788 assert(src
->type
->type
== dest
->type
->type
);
789 enum glsl_base_type base_type
= glsl_get_base_type(src
->type
->type
);
793 case GLSL_TYPE_UINT64
:
794 case GLSL_TYPE_INT64
:
795 case GLSL_TYPE_FLOAT
:
796 case GLSL_TYPE_DOUBLE
:
798 /* At this point, we have a scalar, vector, or matrix so we know that
799 * there cannot be any structure splitting still in the way. By
800 * stopping at the matrix level rather than the vector level, we
801 * ensure that matrices get loaded in the optimal way even if they
802 * are storred row-major in a UBO.
804 vtn_variable_store(b
, vtn_variable_load(b
, src
), dest
);
807 case GLSL_TYPE_ARRAY
:
808 case GLSL_TYPE_STRUCT
: {
809 struct vtn_pointer src_elem
= *src
, dest_elem
= *dest
;
810 src_elem
.chain
= vtn_access_chain_extend(b
, src
->chain
, 1);
811 dest_elem
.chain
= vtn_access_chain_extend(b
, dest
->chain
, 1);
812 src_elem
.chain
->link
[src_elem
.chain
->length
- 1].mode
= vtn_access_mode_literal
;
813 dest_elem
.chain
->link
[dest_elem
.chain
->length
- 1].mode
= vtn_access_mode_literal
;
815 unsigned elems
= glsl_get_length(src
->type
->type
);
816 for (unsigned i
= 0; i
< elems
; i
++) {
817 src_elem
.chain
->link
[src_elem
.chain
->length
- 1].id
= i
;
818 dest_elem
.chain
->link
[dest_elem
.chain
->length
- 1].id
= i
;
819 if (base_type
== GLSL_TYPE_STRUCT
) {
820 src_elem
.type
= src
->type
->members
[i
];
821 dest_elem
.type
= dest
->type
->members
[i
];
823 src_elem
.type
= src
->type
->array_element
;
824 dest_elem
.type
= dest
->type
->array_element
;
826 _vtn_variable_copy(b
, &dest_elem
, &src_elem
);
832 unreachable("Invalid access chain type");
837 vtn_variable_copy(struct vtn_builder
*b
, struct vtn_pointer
*dest
,
838 struct vtn_pointer
*src
)
840 /* TODO: At some point, we should add a special-case for when we can
841 * just emit a copy_var intrinsic.
843 _vtn_variable_copy(b
, dest
, src
);
847 set_mode_system_value(nir_variable_mode
*mode
)
849 assert(*mode
== nir_var_system_value
|| *mode
== nir_var_shader_in
);
850 *mode
= nir_var_system_value
;
854 vtn_get_builtin_location(struct vtn_builder
*b
,
855 SpvBuiltIn builtin
, int *location
,
856 nir_variable_mode
*mode
)
859 case SpvBuiltInPosition
:
860 *location
= VARYING_SLOT_POS
;
862 case SpvBuiltInPointSize
:
863 *location
= VARYING_SLOT_PSIZ
;
865 case SpvBuiltInClipDistance
:
866 *location
= VARYING_SLOT_CLIP_DIST0
; /* XXX CLIP_DIST1? */
868 case SpvBuiltInCullDistance
:
869 *location
= VARYING_SLOT_CULL_DIST0
;
871 case SpvBuiltInVertexIndex
:
872 *location
= SYSTEM_VALUE_VERTEX_ID
;
873 set_mode_system_value(mode
);
875 case SpvBuiltInVertexId
:
876 /* Vulkan defines VertexID to be zero-based and reserves the new
877 * builtin keyword VertexIndex to indicate the non-zero-based value.
879 *location
= SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
;
880 set_mode_system_value(mode
);
882 case SpvBuiltInInstanceIndex
:
883 *location
= SYSTEM_VALUE_INSTANCE_INDEX
;
884 set_mode_system_value(mode
);
886 case SpvBuiltInInstanceId
:
887 *location
= SYSTEM_VALUE_INSTANCE_ID
;
888 set_mode_system_value(mode
);
890 case SpvBuiltInPrimitiveId
:
891 if (b
->shader
->stage
== MESA_SHADER_FRAGMENT
) {
892 assert(*mode
== nir_var_shader_in
);
893 *location
= VARYING_SLOT_PRIMITIVE_ID
;
894 } else if (*mode
== nir_var_shader_out
) {
895 *location
= VARYING_SLOT_PRIMITIVE_ID
;
897 *location
= SYSTEM_VALUE_PRIMITIVE_ID
;
898 set_mode_system_value(mode
);
901 case SpvBuiltInInvocationId
:
902 *location
= SYSTEM_VALUE_INVOCATION_ID
;
903 set_mode_system_value(mode
);
905 case SpvBuiltInLayer
:
906 *location
= VARYING_SLOT_LAYER
;
907 if (b
->shader
->stage
== MESA_SHADER_FRAGMENT
)
908 *mode
= nir_var_shader_in
;
909 else if (b
->shader
->stage
== MESA_SHADER_GEOMETRY
)
910 *mode
= nir_var_shader_out
;
912 unreachable("invalid stage for SpvBuiltInLayer");
914 case SpvBuiltInViewportIndex
:
915 *location
= VARYING_SLOT_VIEWPORT
;
916 if (b
->shader
->stage
== MESA_SHADER_GEOMETRY
)
917 *mode
= nir_var_shader_out
;
918 else if (b
->shader
->stage
== MESA_SHADER_FRAGMENT
)
919 *mode
= nir_var_shader_in
;
921 unreachable("invalid stage for SpvBuiltInViewportIndex");
923 case SpvBuiltInTessLevelOuter
:
924 *location
= VARYING_SLOT_TESS_LEVEL_OUTER
;
926 case SpvBuiltInTessLevelInner
:
927 *location
= VARYING_SLOT_TESS_LEVEL_INNER
;
929 case SpvBuiltInTessCoord
:
930 *location
= SYSTEM_VALUE_TESS_COORD
;
931 set_mode_system_value(mode
);
933 case SpvBuiltInPatchVertices
:
934 *location
= SYSTEM_VALUE_VERTICES_IN
;
935 set_mode_system_value(mode
);
937 case SpvBuiltInFragCoord
:
938 *location
= VARYING_SLOT_POS
;
939 assert(*mode
== nir_var_shader_in
);
941 case SpvBuiltInPointCoord
:
942 *location
= VARYING_SLOT_PNTC
;
943 assert(*mode
== nir_var_shader_in
);
945 case SpvBuiltInFrontFacing
:
946 *location
= SYSTEM_VALUE_FRONT_FACE
;
947 set_mode_system_value(mode
);
949 case SpvBuiltInSampleId
:
950 *location
= SYSTEM_VALUE_SAMPLE_ID
;
951 set_mode_system_value(mode
);
953 case SpvBuiltInSamplePosition
:
954 *location
= SYSTEM_VALUE_SAMPLE_POS
;
955 set_mode_system_value(mode
);
957 case SpvBuiltInSampleMask
:
958 if (*mode
== nir_var_shader_out
) {
959 *location
= FRAG_RESULT_SAMPLE_MASK
;
961 *location
= SYSTEM_VALUE_SAMPLE_MASK_IN
;
962 set_mode_system_value(mode
);
965 case SpvBuiltInFragDepth
:
966 *location
= FRAG_RESULT_DEPTH
;
967 assert(*mode
== nir_var_shader_out
);
969 case SpvBuiltInNumWorkgroups
:
970 *location
= SYSTEM_VALUE_NUM_WORK_GROUPS
;
971 set_mode_system_value(mode
);
973 case SpvBuiltInWorkgroupSize
:
974 /* This should already be handled */
975 unreachable("unsupported builtin");
977 case SpvBuiltInWorkgroupId
:
978 *location
= SYSTEM_VALUE_WORK_GROUP_ID
;
979 set_mode_system_value(mode
);
981 case SpvBuiltInLocalInvocationId
:
982 *location
= SYSTEM_VALUE_LOCAL_INVOCATION_ID
;
983 set_mode_system_value(mode
);
985 case SpvBuiltInLocalInvocationIndex
:
986 *location
= SYSTEM_VALUE_LOCAL_INVOCATION_INDEX
;
987 set_mode_system_value(mode
);
989 case SpvBuiltInGlobalInvocationId
:
990 *location
= SYSTEM_VALUE_GLOBAL_INVOCATION_ID
;
991 set_mode_system_value(mode
);
993 case SpvBuiltInBaseVertex
:
994 *location
= SYSTEM_VALUE_BASE_VERTEX
;
995 set_mode_system_value(mode
);
997 case SpvBuiltInBaseInstance
:
998 *location
= SYSTEM_VALUE_BASE_INSTANCE
;
999 set_mode_system_value(mode
);
1001 case SpvBuiltInDrawIndex
:
1002 *location
= SYSTEM_VALUE_DRAW_ID
;
1003 set_mode_system_value(mode
);
1005 case SpvBuiltInViewIndex
:
1006 *location
= SYSTEM_VALUE_VIEW_INDEX
;
1007 set_mode_system_value(mode
);
1009 case SpvBuiltInHelperInvocation
:
1011 unreachable("unsupported builtin");
1016 apply_var_decoration(struct vtn_builder
*b
, nir_variable
*nir_var
,
1017 const struct vtn_decoration
*dec
)
1019 switch (dec
->decoration
) {
1020 case SpvDecorationRelaxedPrecision
:
1021 break; /* FIXME: Do nothing with this for now. */
1022 case SpvDecorationNoPerspective
:
1023 nir_var
->data
.interpolation
= INTERP_MODE_NOPERSPECTIVE
;
1025 case SpvDecorationFlat
:
1026 nir_var
->data
.interpolation
= INTERP_MODE_FLAT
;
1028 case SpvDecorationCentroid
:
1029 nir_var
->data
.centroid
= true;
1031 case SpvDecorationSample
:
1032 nir_var
->data
.sample
= true;
1034 case SpvDecorationInvariant
:
1035 nir_var
->data
.invariant
= true;
1037 case SpvDecorationConstant
:
1038 assert(nir_var
->constant_initializer
!= NULL
);
1039 nir_var
->data
.read_only
= true;
1041 case SpvDecorationNonReadable
:
1042 nir_var
->data
.image
.write_only
= true;
1044 case SpvDecorationNonWritable
:
1045 nir_var
->data
.read_only
= true;
1046 nir_var
->data
.image
.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 SpvDecorationUniform
:
1099 case SpvDecorationStream
:
1100 case SpvDecorationOffset
:
1101 case SpvDecorationLinkageAttributes
:
1102 break; /* Do nothing with these here */
1104 case SpvDecorationPatch
:
1105 nir_var
->data
.patch
= true;
1108 case SpvDecorationLocation
:
1109 unreachable("Handled above");
1111 case SpvDecorationBlock
:
1112 case SpvDecorationBufferBlock
:
1113 case SpvDecorationArrayStride
:
1114 case SpvDecorationGLSLShared
:
1115 case SpvDecorationGLSLPacked
:
1116 break; /* These can apply to a type but we don't care about them */
1118 case SpvDecorationBinding
:
1119 case SpvDecorationDescriptorSet
:
1120 case SpvDecorationNoContraction
:
1121 case SpvDecorationInputAttachmentIndex
:
1122 vtn_warn("Decoration not allowed for variable or structure member: %s",
1123 spirv_decoration_to_string(dec
->decoration
));
1126 case SpvDecorationXfbBuffer
:
1127 case SpvDecorationXfbStride
:
1128 vtn_warn("Vulkan does not have transform feedback: %s",
1129 spirv_decoration_to_string(dec
->decoration
));
1132 case SpvDecorationCPacked
:
1133 case SpvDecorationSaturatedConversion
:
1134 case SpvDecorationFuncParamAttr
:
1135 case SpvDecorationFPRoundingMode
:
1136 case SpvDecorationFPFastMathMode
:
1137 case SpvDecorationAlignment
:
1138 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1139 spirv_decoration_to_string(dec
->decoration
));
1143 unreachable("Unhandled decoration");
1148 var_is_patch_cb(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
1149 const struct vtn_decoration
*dec
, void *out_is_patch
)
1151 if (dec
->decoration
== SpvDecorationPatch
) {
1152 *((bool *) out_is_patch
) = true;
1157 var_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
1158 const struct vtn_decoration
*dec
, void *void_var
)
1160 struct vtn_variable
*vtn_var
= void_var
;
1162 /* Handle decorations that apply to a vtn_variable as a whole */
1163 switch (dec
->decoration
) {
1164 case SpvDecorationBinding
:
1165 vtn_var
->binding
= dec
->literals
[0];
1167 case SpvDecorationDescriptorSet
:
1168 vtn_var
->descriptor_set
= dec
->literals
[0];
1170 case SpvDecorationInputAttachmentIndex
:
1171 vtn_var
->input_attachment_index
= dec
->literals
[0];
1173 case SpvDecorationPatch
:
1174 vtn_var
->patch
= true;
1180 if (val
->value_type
== vtn_value_type_pointer
) {
1181 assert(val
->pointer
->var
== void_var
);
1182 assert(val
->pointer
->chain
== NULL
);
1183 assert(member
== -1);
1185 assert(val
->value_type
== vtn_value_type_type
);
1188 /* Location is odd. If applied to a split structure, we have to walk the
1189 * whole thing and accumulate the location. It's easier to handle as a
1192 if (dec
->decoration
== SpvDecorationLocation
) {
1193 unsigned location
= dec
->literals
[0];
1194 bool is_vertex_input
;
1195 if (b
->shader
->stage
== MESA_SHADER_FRAGMENT
&&
1196 vtn_var
->mode
== vtn_variable_mode_output
) {
1197 is_vertex_input
= false;
1198 location
+= FRAG_RESULT_DATA0
;
1199 } else if (b
->shader
->stage
== MESA_SHADER_VERTEX
&&
1200 vtn_var
->mode
== vtn_variable_mode_input
) {
1201 is_vertex_input
= true;
1202 location
+= VERT_ATTRIB_GENERIC0
;
1203 } else if (vtn_var
->mode
== vtn_variable_mode_input
||
1204 vtn_var
->mode
== vtn_variable_mode_output
) {
1205 is_vertex_input
= false;
1206 location
+= vtn_var
->patch
? VARYING_SLOT_PATCH0
: VARYING_SLOT_VAR0
;
1208 vtn_warn("Location must be on input or output variable");
1213 /* This handles the member and lone variable cases */
1214 vtn_var
->var
->data
.location
= location
;
1216 /* This handles the structure member case */
1217 assert(vtn_var
->members
);
1219 glsl_get_length(glsl_without_array(vtn_var
->type
->type
));
1220 for (unsigned i
= 0; i
< length
; i
++) {
1221 vtn_var
->members
[i
]->data
.location
= location
;
1223 glsl_count_attribute_slots(vtn_var
->members
[i
]->interface_type
,
1230 assert(member
<= 0);
1231 apply_var_decoration(b
, vtn_var
->var
, dec
);
1232 } else if (vtn_var
->members
) {
1234 assert(vtn_var
->members
);
1235 apply_var_decoration(b
, vtn_var
->members
[member
], dec
);
1238 glsl_get_length(glsl_without_array(vtn_var
->type
->type
));
1239 for (unsigned i
= 0; i
< length
; i
++)
1240 apply_var_decoration(b
, vtn_var
->members
[i
], dec
);
1243 /* A few variables, those with external storage, have no actual
1244 * nir_variables associated with them. Fortunately, all decorations
1245 * we care about for those variables are on the type only.
1247 assert(vtn_var
->mode
== vtn_variable_mode_ubo
||
1248 vtn_var
->mode
== vtn_variable_mode_ssbo
||
1249 vtn_var
->mode
== vtn_variable_mode_push_constant
);
1254 static enum vtn_variable_mode
1255 vtn_storage_class_to_mode(SpvStorageClass
class,
1256 struct vtn_type
*interface_type
,
1257 nir_variable_mode
*nir_mode_out
)
1259 enum vtn_variable_mode mode
;
1260 nir_variable_mode nir_mode
;
1262 case SpvStorageClassUniform
:
1263 if (interface_type
->block
) {
1264 mode
= vtn_variable_mode_ubo
;
1266 } else if (interface_type
->buffer_block
) {
1267 mode
= vtn_variable_mode_ssbo
;
1270 assert(!"Invalid uniform variable type");
1273 case SpvStorageClassUniformConstant
:
1274 if (glsl_type_is_image(interface_type
->type
)) {
1275 mode
= vtn_variable_mode_image
;
1276 nir_mode
= nir_var_uniform
;
1277 } else if (glsl_type_is_sampler(interface_type
->type
)) {
1278 mode
= vtn_variable_mode_sampler
;
1279 nir_mode
= nir_var_uniform
;
1281 assert(!"Invalid uniform constant variable type");
1284 case SpvStorageClassPushConstant
:
1285 mode
= vtn_variable_mode_push_constant
;
1286 nir_mode
= nir_var_uniform
;
1288 case SpvStorageClassInput
:
1289 mode
= vtn_variable_mode_input
;
1290 nir_mode
= nir_var_shader_in
;
1292 case SpvStorageClassOutput
:
1293 mode
= vtn_variable_mode_output
;
1294 nir_mode
= nir_var_shader_out
;
1296 case SpvStorageClassPrivate
:
1297 mode
= vtn_variable_mode_global
;
1298 nir_mode
= nir_var_global
;
1300 case SpvStorageClassFunction
:
1301 mode
= vtn_variable_mode_local
;
1302 nir_mode
= nir_var_local
;
1304 case SpvStorageClassWorkgroup
:
1305 mode
= vtn_variable_mode_workgroup
;
1306 nir_mode
= nir_var_shared
;
1308 case SpvStorageClassCrossWorkgroup
:
1309 case SpvStorageClassGeneric
:
1310 case SpvStorageClassAtomicCounter
:
1312 unreachable("Unhandled variable storage class");
1316 *nir_mode_out
= nir_mode
;
1322 is_per_vertex_inout(const struct vtn_variable
*var
, gl_shader_stage stage
)
1324 if (var
->patch
|| !glsl_type_is_array(var
->type
->type
))
1327 if (var
->mode
== vtn_variable_mode_input
) {
1328 return stage
== MESA_SHADER_TESS_CTRL
||
1329 stage
== MESA_SHADER_TESS_EVAL
||
1330 stage
== MESA_SHADER_GEOMETRY
;
1333 if (var
->mode
== vtn_variable_mode_output
)
1334 return stage
== MESA_SHADER_TESS_CTRL
;
1340 vtn_handle_variables(struct vtn_builder
*b
, SpvOp opcode
,
1341 const uint32_t *w
, unsigned count
)
1345 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
1346 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1350 case SpvOpVariable
: {
1351 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1353 struct vtn_type
*without_array
= type
;
1354 while(glsl_type_is_array(without_array
->type
))
1355 without_array
= without_array
->array_element
;
1357 enum vtn_variable_mode mode
;
1358 nir_variable_mode nir_mode
;
1359 mode
= vtn_storage_class_to_mode(w
[3], without_array
, &nir_mode
);
1362 case vtn_variable_mode_ubo
:
1363 b
->shader
->info
.num_ubos
++;
1365 case vtn_variable_mode_ssbo
:
1366 b
->shader
->info
.num_ssbos
++;
1368 case vtn_variable_mode_image
:
1369 b
->shader
->info
.num_images
++;
1371 case vtn_variable_mode_sampler
:
1372 b
->shader
->info
.num_textures
++;
1374 case vtn_variable_mode_push_constant
:
1375 b
->shader
->num_uniforms
= vtn_type_block_size(type
);
1378 /* No tallying is needed */
1382 struct vtn_variable
*var
= rzalloc(b
, struct vtn_variable
);
1386 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
1387 val
->pointer
= vtn_pointer_for_variable(b
, var
);
1389 switch (var
->mode
) {
1390 case vtn_variable_mode_local
:
1391 case vtn_variable_mode_global
:
1392 case vtn_variable_mode_image
:
1393 case vtn_variable_mode_sampler
:
1394 case vtn_variable_mode_workgroup
:
1395 /* For these, we create the variable normally */
1396 var
->var
= rzalloc(b
->shader
, nir_variable
);
1397 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
1398 var
->var
->type
= var
->type
->type
;
1399 var
->var
->data
.mode
= nir_mode
;
1401 switch (var
->mode
) {
1402 case vtn_variable_mode_image
:
1403 case vtn_variable_mode_sampler
:
1404 var
->var
->interface_type
= without_array
->type
;
1407 var
->var
->interface_type
= NULL
;
1412 case vtn_variable_mode_input
:
1413 case vtn_variable_mode_output
: {
1414 /* In order to know whether or not we're a per-vertex inout, we need
1415 * the patch qualifier. This means walking the variable decorations
1416 * early before we actually create any variables. Not a big deal.
1418 * GLSLang really likes to place decorations in the most interior
1419 * thing it possibly can. In particular, if you have a struct, it
1420 * will place the patch decorations on the struct members. This
1421 * should be handled by the variable splitting below just fine.
1423 * If you have an array-of-struct, things get even more weird as it
1424 * will place the patch decorations on the struct even though it's
1425 * inside an array and some of the members being patch and others not
1426 * makes no sense whatsoever. Since the only sensible thing is for
1427 * it to be all or nothing, we'll call it patch if any of the members
1428 * are declared patch.
1431 vtn_foreach_decoration(b
, val
, var_is_patch_cb
, &var
->patch
);
1432 if (glsl_type_is_array(var
->type
->type
) &&
1433 glsl_type_is_struct(without_array
->type
)) {
1434 vtn_foreach_decoration(b
, without_array
->val
,
1435 var_is_patch_cb
, &var
->patch
);
1438 /* For inputs and outputs, we immediately split structures. This
1439 * is for a couple of reasons. For one, builtins may all come in
1440 * a struct and we really want those split out into separate
1441 * variables. For another, interpolation qualifiers can be
1442 * applied to members of the top-level struct ane we need to be
1443 * able to preserve that information.
1446 int array_length
= -1;
1447 struct vtn_type
*interface_type
= var
->type
;
1448 if (is_per_vertex_inout(var
, b
->shader
->stage
)) {
1449 /* In Geometry shaders (and some tessellation), inputs come
1450 * in per-vertex arrays. However, some builtins come in
1451 * non-per-vertex, hence the need for the is_array check. In
1452 * any case, there are no non-builtin arrays allowed so this
1453 * check should be sufficient.
1455 interface_type
= var
->type
->array_element
;
1456 array_length
= glsl_get_length(var
->type
->type
);
1459 if (glsl_type_is_struct(interface_type
->type
)) {
1460 /* It's a struct. Split it. */
1461 unsigned num_members
= glsl_get_length(interface_type
->type
);
1462 var
->members
= ralloc_array(b
, nir_variable
*, num_members
);
1464 for (unsigned i
= 0; i
< num_members
; i
++) {
1465 const struct glsl_type
*mtype
= interface_type
->members
[i
]->type
;
1466 if (array_length
>= 0)
1467 mtype
= glsl_array_type(mtype
, array_length
);
1469 var
->members
[i
] = rzalloc(b
->shader
, nir_variable
);
1470 var
->members
[i
]->name
=
1471 ralloc_asprintf(var
->members
[i
], "%s.%d", val
->name
, i
);
1472 var
->members
[i
]->type
= mtype
;
1473 var
->members
[i
]->interface_type
=
1474 interface_type
->members
[i
]->type
;
1475 var
->members
[i
]->data
.mode
= nir_mode
;
1476 var
->members
[i
]->data
.patch
= var
->patch
;
1479 var
->var
= rzalloc(b
->shader
, nir_variable
);
1480 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
1481 var
->var
->type
= var
->type
->type
;
1482 var
->var
->interface_type
= interface_type
->type
;
1483 var
->var
->data
.mode
= nir_mode
;
1484 var
->var
->data
.patch
= var
->patch
;
1487 /* For inputs and outputs, we need to grab locations and builtin
1488 * information from the interface type.
1490 vtn_foreach_decoration(b
, interface_type
->val
, var_decoration_cb
, var
);
1494 case vtn_variable_mode_param
:
1495 unreachable("Not created through OpVariable");
1497 case vtn_variable_mode_ubo
:
1498 case vtn_variable_mode_ssbo
:
1499 case vtn_variable_mode_push_constant
:
1500 /* These don't need actual variables. */
1506 nir_constant
*constant
=
1507 vtn_value(b
, w
[4], vtn_value_type_constant
)->constant
;
1508 var
->var
->constant_initializer
=
1509 nir_constant_clone(constant
, var
->var
);
1512 vtn_foreach_decoration(b
, val
, var_decoration_cb
, var
);
1514 if (var
->mode
== vtn_variable_mode_image
||
1515 var
->mode
== vtn_variable_mode_sampler
) {
1516 /* XXX: We still need the binding information in the nir_variable
1517 * for these. We should fix that.
1519 var
->var
->data
.binding
= var
->binding
;
1520 var
->var
->data
.descriptor_set
= var
->descriptor_set
;
1521 var
->var
->data
.index
= var
->input_attachment_index
;
1523 if (var
->mode
== vtn_variable_mode_image
)
1524 var
->var
->data
.image
.format
= without_array
->image_format
;
1527 if (var
->mode
== vtn_variable_mode_local
) {
1528 assert(var
->members
== NULL
&& var
->var
!= NULL
);
1529 nir_function_impl_add_variable(b
->impl
, var
->var
);
1530 } else if (var
->var
) {
1531 nir_shader_add_variable(b
->shader
, var
->var
);
1532 } else if (var
->members
) {
1533 unsigned count
= glsl_get_length(without_array
->type
);
1534 for (unsigned i
= 0; i
< count
; i
++) {
1535 assert(var
->members
[i
]->data
.mode
!= nir_var_local
);
1536 nir_shader_add_variable(b
->shader
, var
->members
[i
]);
1539 assert(var
->mode
== vtn_variable_mode_ubo
||
1540 var
->mode
== vtn_variable_mode_ssbo
||
1541 var
->mode
== vtn_variable_mode_push_constant
);
1546 case SpvOpAccessChain
:
1547 case SpvOpInBoundsAccessChain
: {
1548 struct vtn_pointer
*base
, *ptr
;
1549 struct vtn_value
*base_val
= vtn_untyped_value(b
, w
[3]);
1550 if (base_val
->value_type
== vtn_value_type_sampled_image
) {
1551 /* This is rather insane. SPIR-V allows you to use OpSampledImage
1552 * to combine an array of images with a single sampler to get an
1553 * array of sampled images that all share the same sampler.
1554 * Fortunately, this means that we can more-or-less ignore the
1555 * sampler when crawling the access chain, but it does leave us
1556 * with this rather awkward little special-case.
1558 base
= base_val
->sampled_image
->image
;
1560 assert(base_val
->value_type
== vtn_value_type_pointer
);
1561 base
= base_val
->pointer
;
1564 struct vtn_access_chain
*chain
=
1565 vtn_access_chain_extend(b
, base
->chain
, count
- 4);
1566 struct vtn_type
*type
= base
->type
;
1568 unsigned idx
= base
->chain
? base
->chain
->length
: 0;
1569 for (int i
= 4; i
< count
; i
++) {
1570 struct vtn_value
*link_val
= vtn_untyped_value(b
, w
[i
]);
1571 if (link_val
->value_type
== vtn_value_type_constant
) {
1572 chain
->link
[idx
].mode
= vtn_access_mode_literal
;
1573 chain
->link
[idx
].id
= link_val
->constant
->values
[0].u32
[0];
1575 chain
->link
[idx
].mode
= vtn_access_mode_id
;
1576 chain
->link
[idx
].id
= w
[i
];
1579 if (glsl_type_is_struct(type
->type
)) {
1580 assert(chain
->link
[idx
].mode
== vtn_access_mode_literal
);
1581 type
= type
->members
[chain
->link
[idx
].id
];
1583 type
= type
->array_element
;
1588 ptr
= ralloc(b
, struct vtn_pointer
);
1593 if (base_val
->value_type
== vtn_value_type_sampled_image
) {
1594 struct vtn_value
*val
=
1595 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
1596 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
1597 val
->sampled_image
->image
= ptr
;
1598 val
->sampled_image
->sampler
= base_val
->sampled_image
->sampler
;
1600 struct vtn_value
*val
=
1601 vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
1607 case SpvOpCopyMemory
: {
1608 struct vtn_value
*dest
= vtn_value(b
, w
[1], vtn_value_type_pointer
);
1609 struct vtn_value
*src
= vtn_value(b
, w
[2], vtn_value_type_pointer
);
1611 vtn_variable_copy(b
, dest
->pointer
, src
->pointer
);
1616 struct vtn_pointer
*src
=
1617 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
1619 if (src
->mode
== vtn_variable_mode_image
||
1620 src
->mode
== vtn_variable_mode_sampler
) {
1621 vtn_push_value(b
, w
[2], vtn_value_type_pointer
)->pointer
= src
;
1625 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1626 val
->ssa
= vtn_variable_load(b
, src
);
1631 struct vtn_pointer
*dest
=
1632 vtn_value(b
, w
[1], vtn_value_type_pointer
)->pointer
;
1634 if (glsl_type_is_sampler(dest
->type
->type
)) {
1635 vtn_warn("OpStore of a sampler detected. Doing on-the-fly copy "
1636 "propagation to workaround the problem.");
1637 assert(dest
->var
->copy_prop_sampler
== NULL
);
1638 dest
->var
->copy_prop_sampler
=
1639 vtn_value(b
, w
[2], vtn_value_type_pointer
)->pointer
;
1643 struct vtn_ssa_value
*src
= vtn_ssa_value(b
, w
[2]);
1644 vtn_variable_store(b
, src
, dest
);
1648 case SpvOpArrayLength
: {
1649 struct vtn_pointer
*ptr
=
1650 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
1652 const uint32_t offset
= ptr
->var
->type
->offsets
[w
[4]];
1653 const uint32_t stride
= ptr
->var
->type
->members
[w
[4]]->stride
;
1656 struct vtn_type
*type
;
1657 nir_ssa_def
*index
=
1658 get_vulkan_resource_index(b
, ptr
, &type
, &chain_idx
);
1660 nir_intrinsic_instr
*instr
=
1661 nir_intrinsic_instr_create(b
->nb
.shader
,
1662 nir_intrinsic_get_buffer_size
);
1663 instr
->src
[0] = nir_src_for_ssa(index
);
1664 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
1665 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
1666 nir_ssa_def
*buf_size
= &instr
->dest
.ssa
;
1668 /* array_length = max(buffer_size - offset, 0) / stride */
1669 nir_ssa_def
*array_length
=
1674 nir_imm_int(&b
->nb
, offset
)),
1675 nir_imm_int(&b
->nb
, 0u)),
1676 nir_imm_int(&b
->nb
, stride
));
1678 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
1679 val
->ssa
= vtn_create_ssa_value(b
, glsl_uint_type());
1680 val
->ssa
->def
= array_length
;
1684 case SpvOpCopyMemorySized
:
1686 unreachable("Unhandled opcode");