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"
30 #include "nir_deref.h"
31 #include <vulkan/vulkan_core.h>
34 ptr_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
35 const struct vtn_decoration
*dec
, void *void_ptr
)
37 struct vtn_pointer
*ptr
= void_ptr
;
39 switch (dec
->decoration
) {
40 case SpvDecorationNonUniformEXT
:
41 ptr
->access
|= ACCESS_NON_UNIFORM
;
49 static struct vtn_pointer
*
50 vtn_decorate_pointer(struct vtn_builder
*b
, struct vtn_value
*val
,
51 struct vtn_pointer
*ptr
)
53 struct vtn_pointer dummy
= { .access
= 0 };
54 vtn_foreach_decoration(b
, val
, ptr_decoration_cb
, &dummy
);
56 /* If we're adding access flags, make a copy of the pointer. We could
57 * probably just OR them in without doing so but this prevents us from
58 * leaking them any further than actually specified in the SPIR-V.
60 if (dummy
.access
& ~ptr
->access
) {
61 struct vtn_pointer
*copy
= ralloc(b
, struct vtn_pointer
);
63 copy
->access
|= dummy
.access
;
71 vtn_push_value_pointer(struct vtn_builder
*b
, uint32_t value_id
,
72 struct vtn_pointer
*ptr
)
74 struct vtn_value
*val
= vtn_push_value(b
, value_id
, vtn_value_type_pointer
);
75 val
->pointer
= vtn_decorate_pointer(b
, val
, ptr
);
80 ssa_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
81 const struct vtn_decoration
*dec
, void *void_ssa
)
83 struct vtn_ssa_value
*ssa
= void_ssa
;
85 switch (dec
->decoration
) {
86 case SpvDecorationNonUniformEXT
:
87 ssa
->access
|= ACCESS_NON_UNIFORM
;
96 vtn_push_ssa(struct vtn_builder
*b
, uint32_t value_id
,
97 struct vtn_type
*type
, struct vtn_ssa_value
*ssa
)
99 struct vtn_value
*val
;
100 if (type
->base_type
== vtn_base_type_pointer
) {
101 val
= vtn_push_value_pointer(b
, value_id
, vtn_pointer_from_ssa(b
, ssa
->def
, type
));
103 val
= vtn_push_value(b
, value_id
, vtn_value_type_ssa
);
105 vtn_foreach_decoration(b
, val
, ssa_decoration_cb
, val
->ssa
);
110 static struct vtn_access_chain
*
111 vtn_access_chain_create(struct vtn_builder
*b
, unsigned length
)
113 struct vtn_access_chain
*chain
;
115 /* Subtract 1 from the length since there's already one built in */
116 size_t size
= sizeof(*chain
) +
117 (MAX2(length
, 1) - 1) * sizeof(chain
->link
[0]);
118 chain
= rzalloc_size(b
, size
);
119 chain
->length
= length
;
125 vtn_mode_uses_ssa_offset(struct vtn_builder
*b
,
126 enum vtn_variable_mode mode
)
128 return ((mode
== vtn_variable_mode_ubo
||
129 mode
== vtn_variable_mode_ssbo
) &&
130 b
->options
->lower_ubo_ssbo_access_to_offsets
) ||
131 mode
== vtn_variable_mode_push_constant
;
135 vtn_pointer_is_external_block(struct vtn_builder
*b
,
136 struct vtn_pointer
*ptr
)
138 return ptr
->mode
== vtn_variable_mode_ssbo
||
139 ptr
->mode
== vtn_variable_mode_ubo
||
140 ptr
->mode
== vtn_variable_mode_phys_ssbo
||
141 ptr
->mode
== vtn_variable_mode_push_constant
;
145 vtn_access_link_as_ssa(struct vtn_builder
*b
, struct vtn_access_link link
,
146 unsigned stride
, unsigned bit_size
)
148 vtn_assert(stride
> 0);
149 if (link
.mode
== vtn_access_mode_literal
) {
150 return nir_imm_intN_t(&b
->nb
, link
.id
* stride
, bit_size
);
152 nir_ssa_def
*ssa
= vtn_ssa_value(b
, link
.id
)->def
;
153 if (ssa
->bit_size
!= bit_size
)
154 ssa
= nir_i2i(&b
->nb
, ssa
, bit_size
);
155 return nir_imul_imm(&b
->nb
, ssa
, stride
);
159 static VkDescriptorType
160 vk_desc_type_for_mode(struct vtn_builder
*b
, enum vtn_variable_mode mode
)
163 case vtn_variable_mode_ubo
:
164 return VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER
;
165 case vtn_variable_mode_ssbo
:
166 return VK_DESCRIPTOR_TYPE_STORAGE_BUFFER
;
168 vtn_fail("Invalid mode for vulkan_resource_index");
173 vtn_variable_resource_index(struct vtn_builder
*b
, struct vtn_variable
*var
,
174 nir_ssa_def
*desc_array_index
)
176 vtn_assert(b
->options
->environment
== NIR_SPIRV_VULKAN
);
178 if (!desc_array_index
) {
179 vtn_assert(glsl_type_is_struct_or_ifc(var
->type
->type
));
180 desc_array_index
= nir_imm_int(&b
->nb
, 0);
183 nir_intrinsic_instr
*instr
=
184 nir_intrinsic_instr_create(b
->nb
.shader
,
185 nir_intrinsic_vulkan_resource_index
);
186 instr
->src
[0] = nir_src_for_ssa(desc_array_index
);
187 nir_intrinsic_set_desc_set(instr
, var
->descriptor_set
);
188 nir_intrinsic_set_binding(instr
, var
->binding
);
189 nir_intrinsic_set_desc_type(instr
, vk_desc_type_for_mode(b
, var
->mode
));
191 vtn_fail_if(var
->mode
!= vtn_variable_mode_ubo
&&
192 var
->mode
!= vtn_variable_mode_ssbo
,
193 "Invalid mode for vulkan_resource_index");
195 nir_address_format addr_format
= vtn_mode_to_address_format(b
, var
->mode
);
196 const struct glsl_type
*index_type
=
197 b
->options
->lower_ubo_ssbo_access_to_offsets
?
198 glsl_uint_type() : nir_address_format_to_glsl_type(addr_format
);
200 instr
->num_components
= glsl_get_vector_elements(index_type
);
201 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, instr
->num_components
,
202 glsl_get_bit_size(index_type
), NULL
);
203 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
205 return &instr
->dest
.ssa
;
209 vtn_resource_reindex(struct vtn_builder
*b
, enum vtn_variable_mode mode
,
210 nir_ssa_def
*base_index
, nir_ssa_def
*offset_index
)
212 vtn_assert(b
->options
->environment
== NIR_SPIRV_VULKAN
);
214 nir_intrinsic_instr
*instr
=
215 nir_intrinsic_instr_create(b
->nb
.shader
,
216 nir_intrinsic_vulkan_resource_reindex
);
217 instr
->src
[0] = nir_src_for_ssa(base_index
);
218 instr
->src
[1] = nir_src_for_ssa(offset_index
);
219 nir_intrinsic_set_desc_type(instr
, vk_desc_type_for_mode(b
, mode
));
221 vtn_fail_if(mode
!= vtn_variable_mode_ubo
&& mode
!= vtn_variable_mode_ssbo
,
222 "Invalid mode for vulkan_resource_reindex");
224 nir_address_format addr_format
= vtn_mode_to_address_format(b
, mode
);
225 const struct glsl_type
*index_type
=
226 b
->options
->lower_ubo_ssbo_access_to_offsets
?
227 glsl_uint_type() : nir_address_format_to_glsl_type(addr_format
);
229 instr
->num_components
= glsl_get_vector_elements(index_type
);
230 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, instr
->num_components
,
231 glsl_get_bit_size(index_type
), NULL
);
232 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
234 return &instr
->dest
.ssa
;
238 vtn_descriptor_load(struct vtn_builder
*b
, enum vtn_variable_mode mode
,
239 nir_ssa_def
*desc_index
)
241 vtn_assert(b
->options
->environment
== NIR_SPIRV_VULKAN
);
243 nir_intrinsic_instr
*desc_load
=
244 nir_intrinsic_instr_create(b
->nb
.shader
,
245 nir_intrinsic_load_vulkan_descriptor
);
246 desc_load
->src
[0] = nir_src_for_ssa(desc_index
);
247 nir_intrinsic_set_desc_type(desc_load
, vk_desc_type_for_mode(b
, mode
));
249 vtn_fail_if(mode
!= vtn_variable_mode_ubo
&& mode
!= vtn_variable_mode_ssbo
,
250 "Invalid mode for load_vulkan_descriptor");
252 nir_address_format addr_format
= vtn_mode_to_address_format(b
, mode
);
253 const struct glsl_type
*ptr_type
=
254 nir_address_format_to_glsl_type(addr_format
);
256 desc_load
->num_components
= glsl_get_vector_elements(ptr_type
);
257 nir_ssa_dest_init(&desc_load
->instr
, &desc_load
->dest
,
258 desc_load
->num_components
,
259 glsl_get_bit_size(ptr_type
), NULL
);
260 nir_builder_instr_insert(&b
->nb
, &desc_load
->instr
);
262 return &desc_load
->dest
.ssa
;
265 /* Dereference the given base pointer by the access chain */
266 static struct vtn_pointer
*
267 vtn_nir_deref_pointer_dereference(struct vtn_builder
*b
,
268 struct vtn_pointer
*base
,
269 struct vtn_access_chain
*deref_chain
)
271 struct vtn_type
*type
= base
->type
;
272 enum gl_access_qualifier access
= base
->access
| deref_chain
->access
;
275 nir_deref_instr
*tail
;
278 } else if (b
->options
->environment
== NIR_SPIRV_VULKAN
&&
279 vtn_pointer_is_external_block(b
, base
)) {
280 nir_ssa_def
*block_index
= base
->block_index
;
282 /* We dereferencing an external block pointer. Correctness of this
283 * operation relies on one particular line in the SPIR-V spec, section
284 * entitled "Validation Rules for Shader Capabilities":
286 * "Block and BufferBlock decorations cannot decorate a structure
287 * type that is nested at any level inside another structure type
288 * decorated with Block or BufferBlock."
290 * This means that we can detect the point where we cross over from
291 * descriptor indexing to buffer indexing by looking for the block
292 * decorated struct type. Anything before the block decorated struct
293 * type is a descriptor indexing operation and anything after the block
294 * decorated struct is a buffer offset operation.
297 /* Figure out the descriptor array index if any
299 * Some of the Vulkan CTS tests with hand-rolled SPIR-V have been known
300 * to forget the Block or BufferBlock decoration from time to time.
301 * It's more robust if we check for both !block_index and for the type
302 * to contain a block. This way there's a decent chance that arrays of
303 * UBOs/SSBOs will work correctly even if variable pointers are
306 nir_ssa_def
*desc_arr_idx
= NULL
;
307 if (!block_index
|| vtn_type_contains_block(b
, type
)) {
308 /* If our type contains a block, then we're still outside the block
309 * and we need to process enough levels of dereferences to get inside
312 if (deref_chain
->ptr_as_array
) {
313 unsigned aoa_size
= glsl_get_aoa_size(type
->type
);
314 desc_arr_idx
= vtn_access_link_as_ssa(b
, deref_chain
->link
[idx
],
315 MAX2(aoa_size
, 1), 32);
319 for (; idx
< deref_chain
->length
; idx
++) {
320 if (type
->base_type
!= vtn_base_type_array
) {
321 vtn_assert(type
->base_type
== vtn_base_type_struct
);
325 unsigned aoa_size
= glsl_get_aoa_size(type
->array_element
->type
);
326 nir_ssa_def
*arr_offset
=
327 vtn_access_link_as_ssa(b
, deref_chain
->link
[idx
],
328 MAX2(aoa_size
, 1), 32);
330 desc_arr_idx
= nir_iadd(&b
->nb
, desc_arr_idx
, arr_offset
);
332 desc_arr_idx
= arr_offset
;
334 type
= type
->array_element
;
335 access
|= type
->access
;
340 vtn_assert(base
->var
&& base
->type
);
341 block_index
= vtn_variable_resource_index(b
, base
->var
, desc_arr_idx
);
342 } else if (desc_arr_idx
) {
343 block_index
= vtn_resource_reindex(b
, base
->mode
,
344 block_index
, desc_arr_idx
);
347 if (idx
== deref_chain
->length
) {
348 /* The entire deref was consumed in finding the block index. Return
349 * a pointer which just has a block index and a later access chain
350 * will dereference deeper.
352 struct vtn_pointer
*ptr
= rzalloc(b
, struct vtn_pointer
);
353 ptr
->mode
= base
->mode
;
355 ptr
->block_index
= block_index
;
356 ptr
->access
= access
;
360 /* If we got here, there's more access chain to handle and we have the
361 * final block index. Insert a descriptor load and cast to a deref to
362 * start the deref chain.
364 nir_ssa_def
*desc
= vtn_descriptor_load(b
, base
->mode
, block_index
);
366 assert(base
->mode
== vtn_variable_mode_ssbo
||
367 base
->mode
== vtn_variable_mode_ubo
);
368 nir_variable_mode nir_mode
=
369 base
->mode
== vtn_variable_mode_ssbo
? nir_var_mem_ssbo
: nir_var_mem_ubo
;
371 tail
= nir_build_deref_cast(&b
->nb
, desc
, nir_mode
, type
->type
,
372 base
->ptr_type
->stride
);
374 assert(base
->var
&& base
->var
->var
);
375 tail
= nir_build_deref_var(&b
->nb
, base
->var
->var
);
376 if (base
->ptr_type
&& base
->ptr_type
->type
) {
377 tail
->dest
.ssa
.num_components
=
378 glsl_get_vector_elements(base
->ptr_type
->type
);
379 tail
->dest
.ssa
.bit_size
= glsl_get_bit_size(base
->ptr_type
->type
);
383 if (idx
== 0 && deref_chain
->ptr_as_array
) {
384 /* We start with a deref cast to get the stride. Hopefully, we'll be
385 * able to delete that cast eventually.
387 tail
= nir_build_deref_cast(&b
->nb
, &tail
->dest
.ssa
, tail
->mode
,
388 tail
->type
, base
->ptr_type
->stride
);
390 nir_ssa_def
*index
= vtn_access_link_as_ssa(b
, deref_chain
->link
[0], 1,
391 tail
->dest
.ssa
.bit_size
);
392 tail
= nir_build_deref_ptr_as_array(&b
->nb
, tail
, index
);
396 for (; idx
< deref_chain
->length
; idx
++) {
397 if (glsl_type_is_struct_or_ifc(type
->type
)) {
398 vtn_assert(deref_chain
->link
[idx
].mode
== vtn_access_mode_literal
);
399 unsigned field
= deref_chain
->link
[idx
].id
;
400 tail
= nir_build_deref_struct(&b
->nb
, tail
, field
);
401 type
= type
->members
[field
];
403 nir_ssa_def
*arr_index
=
404 vtn_access_link_as_ssa(b
, deref_chain
->link
[idx
], 1,
405 tail
->dest
.ssa
.bit_size
);
406 tail
= nir_build_deref_array(&b
->nb
, tail
, arr_index
);
407 type
= type
->array_element
;
410 access
|= type
->access
;
413 struct vtn_pointer
*ptr
= rzalloc(b
, struct vtn_pointer
);
414 ptr
->mode
= base
->mode
;
416 ptr
->var
= base
->var
;
418 ptr
->access
= access
;
423 static struct vtn_pointer
*
424 vtn_ssa_offset_pointer_dereference(struct vtn_builder
*b
,
425 struct vtn_pointer
*base
,
426 struct vtn_access_chain
*deref_chain
)
428 nir_ssa_def
*block_index
= base
->block_index
;
429 nir_ssa_def
*offset
= base
->offset
;
430 struct vtn_type
*type
= base
->type
;
431 enum gl_access_qualifier access
= base
->access
;
434 if (base
->mode
== vtn_variable_mode_ubo
||
435 base
->mode
== vtn_variable_mode_ssbo
) {
437 vtn_assert(base
->var
&& base
->type
);
438 nir_ssa_def
*desc_arr_idx
;
439 if (glsl_type_is_array(type
->type
)) {
440 if (deref_chain
->length
>= 1) {
442 vtn_access_link_as_ssa(b
, deref_chain
->link
[0], 1, 32);
444 /* This consumes a level of type */
445 type
= type
->array_element
;
446 access
|= type
->access
;
448 /* This is annoying. We've been asked for a pointer to the
449 * array of UBOs/SSBOs and not a specifc buffer. Return a
450 * pointer with a descriptor index of 0 and we'll have to do
451 * a reindex later to adjust it to the right thing.
453 desc_arr_idx
= nir_imm_int(&b
->nb
, 0);
455 } else if (deref_chain
->ptr_as_array
) {
456 /* You can't have a zero-length OpPtrAccessChain */
457 vtn_assert(deref_chain
->length
>= 1);
458 desc_arr_idx
= vtn_access_link_as_ssa(b
, deref_chain
->link
[0], 1, 32);
460 /* We have a regular non-array SSBO. */
463 block_index
= vtn_variable_resource_index(b
, base
->var
, desc_arr_idx
);
464 } else if (deref_chain
->ptr_as_array
&&
465 type
->base_type
== vtn_base_type_struct
&& type
->block
) {
466 /* We are doing an OpPtrAccessChain on a pointer to a struct that is
467 * decorated block. This is an interesting corner in the SPIR-V
468 * spec. One interpretation would be that they client is clearly
469 * trying to treat that block as if it's an implicit array of blocks
470 * repeated in the buffer. However, the SPIR-V spec for the
471 * OpPtrAccessChain says:
473 * "Base is treated as the address of the first element of an
474 * array, and the Element element’s address is computed to be the
475 * base for the Indexes, as per OpAccessChain."
477 * Taken literally, that would mean that your struct type is supposed
478 * to be treated as an array of such a struct and, since it's
479 * decorated block, that means an array of blocks which corresponds
480 * to an array descriptor. Therefore, we need to do a reindex
481 * operation to add the index from the first link in the access chain
482 * to the index we recieved.
484 * The downside to this interpretation (there always is one) is that
485 * this might be somewhat surprising behavior to apps if they expect
486 * the implicit array behavior described above.
488 vtn_assert(deref_chain
->length
>= 1);
489 nir_ssa_def
*offset_index
=
490 vtn_access_link_as_ssa(b
, deref_chain
->link
[0], 1, 32);
493 block_index
= vtn_resource_reindex(b
, base
->mode
,
494 block_index
, offset_index
);
499 if (base
->mode
== vtn_variable_mode_workgroup
) {
500 /* SLM doesn't need nor have a block index */
501 vtn_assert(!block_index
);
503 /* We need the variable for the base offset */
504 vtn_assert(base
->var
);
506 /* We need ptr_type for size and alignment */
507 vtn_assert(base
->ptr_type
);
509 /* Assign location on first use so that we don't end up bloating SLM
510 * address space for variables which are never statically used.
512 if (base
->var
->shared_location
< 0) {
513 vtn_assert(base
->ptr_type
->length
> 0 && base
->ptr_type
->align
> 0);
514 b
->shader
->num_shared
= vtn_align_u32(b
->shader
->num_shared
,
515 base
->ptr_type
->align
);
516 base
->var
->shared_location
= b
->shader
->num_shared
;
517 b
->shader
->num_shared
+= base
->ptr_type
->length
;
520 offset
= nir_imm_int(&b
->nb
, base
->var
->shared_location
);
521 } else if (base
->mode
== vtn_variable_mode_push_constant
) {
522 /* Push constants neither need nor have a block index */
523 vtn_assert(!block_index
);
525 /* Start off with at the start of the push constant block. */
526 offset
= nir_imm_int(&b
->nb
, 0);
528 /* The code above should have ensured a block_index when needed. */
529 vtn_assert(block_index
);
531 /* Start off with at the start of the buffer. */
532 offset
= nir_imm_int(&b
->nb
, 0);
536 if (deref_chain
->ptr_as_array
&& idx
== 0) {
537 /* We need ptr_type for the stride */
538 vtn_assert(base
->ptr_type
);
540 /* We need at least one element in the chain */
541 vtn_assert(deref_chain
->length
>= 1);
543 nir_ssa_def
*elem_offset
=
544 vtn_access_link_as_ssa(b
, deref_chain
->link
[idx
],
545 base
->ptr_type
->stride
, offset
->bit_size
);
546 offset
= nir_iadd(&b
->nb
, offset
, elem_offset
);
550 for (; idx
< deref_chain
->length
; idx
++) {
551 switch (glsl_get_base_type(type
->type
)) {
554 case GLSL_TYPE_UINT16
:
555 case GLSL_TYPE_INT16
:
556 case GLSL_TYPE_UINT8
:
558 case GLSL_TYPE_UINT64
:
559 case GLSL_TYPE_INT64
:
560 case GLSL_TYPE_FLOAT
:
561 case GLSL_TYPE_FLOAT16
:
562 case GLSL_TYPE_DOUBLE
:
564 case GLSL_TYPE_ARRAY
: {
565 nir_ssa_def
*elem_offset
=
566 vtn_access_link_as_ssa(b
, deref_chain
->link
[idx
],
567 type
->stride
, offset
->bit_size
);
568 offset
= nir_iadd(&b
->nb
, offset
, elem_offset
);
569 type
= type
->array_element
;
570 access
|= type
->access
;
574 case GLSL_TYPE_INTERFACE
:
575 case GLSL_TYPE_STRUCT
: {
576 vtn_assert(deref_chain
->link
[idx
].mode
== vtn_access_mode_literal
);
577 unsigned member
= deref_chain
->link
[idx
].id
;
578 offset
= nir_iadd_imm(&b
->nb
, offset
, type
->offsets
[member
]);
579 type
= type
->members
[member
];
580 access
|= type
->access
;
585 vtn_fail("Invalid type for deref");
589 struct vtn_pointer
*ptr
= rzalloc(b
, struct vtn_pointer
);
590 ptr
->mode
= base
->mode
;
592 ptr
->block_index
= block_index
;
593 ptr
->offset
= offset
;
594 ptr
->access
= access
;
599 /* Dereference the given base pointer by the access chain */
600 static struct vtn_pointer
*
601 vtn_pointer_dereference(struct vtn_builder
*b
,
602 struct vtn_pointer
*base
,
603 struct vtn_access_chain
*deref_chain
)
605 if (vtn_pointer_uses_ssa_offset(b
, base
)) {
606 return vtn_ssa_offset_pointer_dereference(b
, base
, deref_chain
);
608 return vtn_nir_deref_pointer_dereference(b
, base
, deref_chain
);
613 vtn_pointer_for_variable(struct vtn_builder
*b
,
614 struct vtn_variable
*var
, struct vtn_type
*ptr_type
)
616 struct vtn_pointer
*pointer
= rzalloc(b
, struct vtn_pointer
);
618 pointer
->mode
= var
->mode
;
619 pointer
->type
= var
->type
;
620 vtn_assert(ptr_type
->base_type
== vtn_base_type_pointer
);
621 vtn_assert(ptr_type
->deref
->type
== var
->type
->type
);
622 pointer
->ptr_type
= ptr_type
;
624 pointer
->access
= var
->access
| var
->type
->access
;
629 /* Returns an atomic_uint type based on the original uint type. The returned
630 * type will be equivalent to the original one but will have an atomic_uint
631 * type as leaf instead of an uint.
633 * Manages uint scalars, arrays, and arrays of arrays of any nested depth.
635 static const struct glsl_type
*
636 repair_atomic_type(const struct glsl_type
*type
)
638 assert(glsl_get_base_type(glsl_without_array(type
)) == GLSL_TYPE_UINT
);
639 assert(glsl_type_is_scalar(glsl_without_array(type
)));
641 if (glsl_type_is_array(type
)) {
642 const struct glsl_type
*atomic
=
643 repair_atomic_type(glsl_get_array_element(type
));
645 return glsl_array_type(atomic
, glsl_get_length(type
),
646 glsl_get_explicit_stride(type
));
648 return glsl_atomic_uint_type();
653 vtn_pointer_to_deref(struct vtn_builder
*b
, struct vtn_pointer
*ptr
)
655 if (b
->wa_glslang_179
) {
656 /* Do on-the-fly copy propagation for samplers. */
657 if (ptr
->var
&& ptr
->var
->copy_prop_sampler
)
658 return vtn_pointer_to_deref(b
, ptr
->var
->copy_prop_sampler
);
661 vtn_assert(!vtn_pointer_uses_ssa_offset(b
, ptr
));
663 struct vtn_access_chain chain
= {
666 ptr
= vtn_nir_deref_pointer_dereference(b
, ptr
, &chain
);
673 _vtn_local_load_store(struct vtn_builder
*b
, bool load
, nir_deref_instr
*deref
,
674 struct vtn_ssa_value
*inout
,
675 enum gl_access_qualifier access
)
677 if (glsl_type_is_vector_or_scalar(deref
->type
)) {
679 inout
->def
= nir_load_deref_with_access(&b
->nb
, deref
, access
);
681 nir_store_deref_with_access(&b
->nb
, deref
, inout
->def
, ~0, access
);
683 } else if (glsl_type_is_array(deref
->type
) ||
684 glsl_type_is_matrix(deref
->type
)) {
685 unsigned elems
= glsl_get_length(deref
->type
);
686 for (unsigned i
= 0; i
< elems
; i
++) {
687 nir_deref_instr
*child
=
688 nir_build_deref_array_imm(&b
->nb
, deref
, i
);
689 _vtn_local_load_store(b
, load
, child
, inout
->elems
[i
], access
);
692 vtn_assert(glsl_type_is_struct_or_ifc(deref
->type
));
693 unsigned elems
= glsl_get_length(deref
->type
);
694 for (unsigned i
= 0; i
< elems
; i
++) {
695 nir_deref_instr
*child
= nir_build_deref_struct(&b
->nb
, deref
, i
);
696 _vtn_local_load_store(b
, load
, child
, inout
->elems
[i
], access
);
702 vtn_nir_deref(struct vtn_builder
*b
, uint32_t id
)
704 struct vtn_pointer
*ptr
= vtn_value(b
, id
, vtn_value_type_pointer
)->pointer
;
705 return vtn_pointer_to_deref(b
, ptr
);
709 * Gets the NIR-level deref tail, which may have as a child an array deref
710 * selecting which component due to OpAccessChain supporting per-component
711 * indexing in SPIR-V.
713 static nir_deref_instr
*
714 get_deref_tail(nir_deref_instr
*deref
)
716 if (deref
->deref_type
!= nir_deref_type_array
)
719 nir_deref_instr
*parent
=
720 nir_instr_as_deref(deref
->parent
.ssa
->parent_instr
);
722 if (glsl_type_is_vector(parent
->type
))
728 struct vtn_ssa_value
*
729 vtn_local_load(struct vtn_builder
*b
, nir_deref_instr
*src
,
730 enum gl_access_qualifier access
)
732 nir_deref_instr
*src_tail
= get_deref_tail(src
);
733 struct vtn_ssa_value
*val
= vtn_create_ssa_value(b
, src_tail
->type
);
734 _vtn_local_load_store(b
, true, src_tail
, val
, access
);
736 if (src_tail
!= src
) {
737 val
->type
= src
->type
;
738 if (nir_src_is_const(src
->arr
.index
))
739 val
->def
= vtn_vector_extract(b
, val
->def
,
740 nir_src_as_uint(src
->arr
.index
));
742 val
->def
= vtn_vector_extract_dynamic(b
, val
->def
, src
->arr
.index
.ssa
);
749 vtn_local_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
750 nir_deref_instr
*dest
, enum gl_access_qualifier access
)
752 nir_deref_instr
*dest_tail
= get_deref_tail(dest
);
754 if (dest_tail
!= dest
) {
755 struct vtn_ssa_value
*val
= vtn_create_ssa_value(b
, dest_tail
->type
);
756 _vtn_local_load_store(b
, true, dest_tail
, val
, access
);
758 if (nir_src_is_const(dest
->arr
.index
))
759 val
->def
= vtn_vector_insert(b
, val
->def
, src
->def
,
760 nir_src_as_uint(dest
->arr
.index
));
762 val
->def
= vtn_vector_insert_dynamic(b
, val
->def
, src
->def
,
763 dest
->arr
.index
.ssa
);
764 _vtn_local_load_store(b
, false, dest_tail
, val
, access
);
766 _vtn_local_load_store(b
, false, dest_tail
, src
, access
);
771 vtn_pointer_to_offset(struct vtn_builder
*b
, struct vtn_pointer
*ptr
,
772 nir_ssa_def
**index_out
)
774 assert(vtn_pointer_uses_ssa_offset(b
, ptr
));
776 struct vtn_access_chain chain
= {
779 ptr
= vtn_ssa_offset_pointer_dereference(b
, ptr
, &chain
);
781 *index_out
= ptr
->block_index
;
785 /* Tries to compute the size of an interface block based on the strides and
786 * offsets that are provided to us in the SPIR-V source.
789 vtn_type_block_size(struct vtn_builder
*b
, struct vtn_type
*type
)
791 enum glsl_base_type base_type
= glsl_get_base_type(type
->type
);
795 case GLSL_TYPE_UINT16
:
796 case GLSL_TYPE_INT16
:
797 case GLSL_TYPE_UINT8
:
799 case GLSL_TYPE_UINT64
:
800 case GLSL_TYPE_INT64
:
801 case GLSL_TYPE_FLOAT
:
802 case GLSL_TYPE_FLOAT16
:
804 case GLSL_TYPE_DOUBLE
: {
805 unsigned cols
= type
->row_major
? glsl_get_vector_elements(type
->type
) :
806 glsl_get_matrix_columns(type
->type
);
808 vtn_assert(type
->stride
> 0);
809 return type
->stride
* cols
;
811 unsigned type_size
= glsl_get_bit_size(type
->type
) / 8;
812 return glsl_get_vector_elements(type
->type
) * type_size
;
816 case GLSL_TYPE_STRUCT
:
817 case GLSL_TYPE_INTERFACE
: {
819 unsigned num_fields
= glsl_get_length(type
->type
);
820 for (unsigned f
= 0; f
< num_fields
; f
++) {
821 unsigned field_end
= type
->offsets
[f
] +
822 vtn_type_block_size(b
, type
->members
[f
]);
823 size
= MAX2(size
, field_end
);
828 case GLSL_TYPE_ARRAY
:
829 vtn_assert(type
->stride
> 0);
830 vtn_assert(glsl_get_length(type
->type
) > 0);
831 return type
->stride
* glsl_get_length(type
->type
);
834 vtn_fail("Invalid block type");
840 _vtn_load_store_tail(struct vtn_builder
*b
, nir_intrinsic_op op
, bool load
,
841 nir_ssa_def
*index
, nir_ssa_def
*offset
,
842 unsigned access_offset
, unsigned access_size
,
843 struct vtn_ssa_value
**inout
, const struct glsl_type
*type
,
844 enum gl_access_qualifier access
)
846 nir_intrinsic_instr
*instr
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
847 instr
->num_components
= glsl_get_vector_elements(type
);
849 /* Booleans usually shouldn't show up in external memory in SPIR-V.
850 * However, they do for certain older GLSLang versions and can for shared
851 * memory when we lower access chains internally.
853 const unsigned data_bit_size
= glsl_type_is_boolean(type
) ? 32 :
854 glsl_get_bit_size(type
);
858 nir_intrinsic_set_write_mask(instr
, (1 << instr
->num_components
) - 1);
859 instr
->src
[src
++] = nir_src_for_ssa((*inout
)->def
);
862 if (op
== nir_intrinsic_load_push_constant
) {
863 nir_intrinsic_set_base(instr
, access_offset
);
864 nir_intrinsic_set_range(instr
, access_size
);
867 if (op
== nir_intrinsic_load_ubo
||
868 op
== nir_intrinsic_load_ssbo
||
869 op
== nir_intrinsic_store_ssbo
) {
870 nir_intrinsic_set_access(instr
, access
);
873 /* With extensions like relaxed_block_layout, we really can't guarantee
874 * much more than scalar alignment.
876 if (op
!= nir_intrinsic_load_push_constant
)
877 nir_intrinsic_set_align(instr
, data_bit_size
/ 8, 0);
880 instr
->src
[src
++] = nir_src_for_ssa(index
);
882 if (op
== nir_intrinsic_load_push_constant
) {
883 /* We need to subtract the offset from where the intrinsic will load the
886 nir_src_for_ssa(nir_isub(&b
->nb
, offset
,
887 nir_imm_int(&b
->nb
, access_offset
)));
889 instr
->src
[src
++] = nir_src_for_ssa(offset
);
893 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
894 instr
->num_components
, data_bit_size
, NULL
);
895 (*inout
)->def
= &instr
->dest
.ssa
;
898 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
900 if (load
&& glsl_get_base_type(type
) == GLSL_TYPE_BOOL
)
901 (*inout
)->def
= nir_ine(&b
->nb
, (*inout
)->def
, nir_imm_int(&b
->nb
, 0));
905 _vtn_block_load_store(struct vtn_builder
*b
, nir_intrinsic_op op
, bool load
,
906 nir_ssa_def
*index
, nir_ssa_def
*offset
,
907 unsigned access_offset
, unsigned access_size
,
908 struct vtn_type
*type
, enum gl_access_qualifier access
,
909 struct vtn_ssa_value
**inout
)
911 if (load
&& *inout
== NULL
)
912 *inout
= vtn_create_ssa_value(b
, type
->type
);
914 enum glsl_base_type base_type
= glsl_get_base_type(type
->type
);
918 case GLSL_TYPE_UINT16
:
919 case GLSL_TYPE_INT16
:
920 case GLSL_TYPE_UINT8
:
922 case GLSL_TYPE_UINT64
:
923 case GLSL_TYPE_INT64
:
924 case GLSL_TYPE_FLOAT
:
925 case GLSL_TYPE_FLOAT16
:
926 case GLSL_TYPE_DOUBLE
:
928 /* This is where things get interesting. At this point, we've hit
929 * a vector, a scalar, or a matrix.
931 if (glsl_type_is_matrix(type
->type
)) {
932 /* Loading the whole matrix */
933 struct vtn_ssa_value
*transpose
;
934 unsigned num_ops
, vec_width
, col_stride
;
935 if (type
->row_major
) {
936 num_ops
= glsl_get_vector_elements(type
->type
);
937 vec_width
= glsl_get_matrix_columns(type
->type
);
938 col_stride
= type
->array_element
->stride
;
940 const struct glsl_type
*transpose_type
=
941 glsl_matrix_type(base_type
, vec_width
, num_ops
);
942 *inout
= vtn_create_ssa_value(b
, transpose_type
);
944 transpose
= vtn_ssa_transpose(b
, *inout
);
948 num_ops
= glsl_get_matrix_columns(type
->type
);
949 vec_width
= glsl_get_vector_elements(type
->type
);
950 col_stride
= type
->stride
;
953 for (unsigned i
= 0; i
< num_ops
; i
++) {
954 nir_ssa_def
*elem_offset
=
955 nir_iadd_imm(&b
->nb
, offset
, i
* col_stride
);
956 _vtn_load_store_tail(b
, op
, load
, index
, elem_offset
,
957 access_offset
, access_size
,
959 glsl_vector_type(base_type
, vec_width
),
960 type
->access
| access
);
963 if (load
&& type
->row_major
)
964 *inout
= vtn_ssa_transpose(b
, *inout
);
966 unsigned elems
= glsl_get_vector_elements(type
->type
);
967 unsigned type_size
= glsl_get_bit_size(type
->type
) / 8;
968 if (elems
== 1 || type
->stride
== type_size
) {
969 /* This is a tightly-packed normal scalar or vector load */
970 vtn_assert(glsl_type_is_vector_or_scalar(type
->type
));
971 _vtn_load_store_tail(b
, op
, load
, index
, offset
,
972 access_offset
, access_size
,
974 type
->access
| access
);
976 /* This is a strided load. We have to load N things separately.
977 * This is the single column of a row-major matrix case.
979 vtn_assert(type
->stride
> type_size
);
980 vtn_assert(type
->stride
% type_size
== 0);
982 nir_ssa_def
*per_comp
[4];
983 for (unsigned i
= 0; i
< elems
; i
++) {
984 nir_ssa_def
*elem_offset
=
985 nir_iadd_imm(&b
->nb
, offset
, i
* type
->stride
);
986 struct vtn_ssa_value
*comp
, temp_val
;
988 temp_val
.def
= nir_channel(&b
->nb
, (*inout
)->def
, i
);
989 temp_val
.type
= glsl_scalar_type(base_type
);
992 _vtn_load_store_tail(b
, op
, load
, index
, elem_offset
,
993 access_offset
, access_size
,
994 &comp
, glsl_scalar_type(base_type
),
995 type
->access
| access
);
996 per_comp
[i
] = comp
->def
;
1001 *inout
= vtn_create_ssa_value(b
, type
->type
);
1002 (*inout
)->def
= nir_vec(&b
->nb
, per_comp
, elems
);
1008 case GLSL_TYPE_ARRAY
: {
1009 unsigned elems
= glsl_get_length(type
->type
);
1010 for (unsigned i
= 0; i
< elems
; i
++) {
1011 nir_ssa_def
*elem_off
=
1012 nir_iadd_imm(&b
->nb
, offset
, i
* type
->stride
);
1013 _vtn_block_load_store(b
, op
, load
, index
, elem_off
,
1014 access_offset
, access_size
,
1015 type
->array_element
,
1016 type
->array_element
->access
| access
,
1017 &(*inout
)->elems
[i
]);
1022 case GLSL_TYPE_INTERFACE
:
1023 case GLSL_TYPE_STRUCT
: {
1024 unsigned elems
= glsl_get_length(type
->type
);
1025 for (unsigned i
= 0; i
< elems
; i
++) {
1026 nir_ssa_def
*elem_off
=
1027 nir_iadd_imm(&b
->nb
, offset
, type
->offsets
[i
]);
1028 _vtn_block_load_store(b
, op
, load
, index
, elem_off
,
1029 access_offset
, access_size
,
1031 type
->members
[i
]->access
| access
,
1032 &(*inout
)->elems
[i
]);
1038 vtn_fail("Invalid block member type");
1042 static struct vtn_ssa_value
*
1043 vtn_block_load(struct vtn_builder
*b
, struct vtn_pointer
*src
)
1045 nir_intrinsic_op op
;
1046 unsigned access_offset
= 0, access_size
= 0;
1047 switch (src
->mode
) {
1048 case vtn_variable_mode_ubo
:
1049 op
= nir_intrinsic_load_ubo
;
1051 case vtn_variable_mode_ssbo
:
1052 op
= nir_intrinsic_load_ssbo
;
1054 case vtn_variable_mode_push_constant
:
1055 op
= nir_intrinsic_load_push_constant
;
1056 access_size
= b
->shader
->num_uniforms
;
1058 case vtn_variable_mode_workgroup
:
1059 op
= nir_intrinsic_load_shared
;
1062 vtn_fail("Invalid block variable mode");
1065 nir_ssa_def
*offset
, *index
= NULL
;
1066 offset
= vtn_pointer_to_offset(b
, src
, &index
);
1068 struct vtn_ssa_value
*value
= NULL
;
1069 _vtn_block_load_store(b
, op
, true, index
, offset
,
1070 access_offset
, access_size
,
1071 src
->type
, src
->access
, &value
);
1076 vtn_block_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
1077 struct vtn_pointer
*dst
)
1079 nir_intrinsic_op op
;
1080 switch (dst
->mode
) {
1081 case vtn_variable_mode_ssbo
:
1082 op
= nir_intrinsic_store_ssbo
;
1084 case vtn_variable_mode_workgroup
:
1085 op
= nir_intrinsic_store_shared
;
1088 vtn_fail("Invalid block variable mode");
1091 nir_ssa_def
*offset
, *index
= NULL
;
1092 offset
= vtn_pointer_to_offset(b
, dst
, &index
);
1094 _vtn_block_load_store(b
, op
, false, index
, offset
,
1095 0, 0, dst
->type
, dst
->access
, &src
);
1099 _vtn_variable_load_store(struct vtn_builder
*b
, bool load
,
1100 struct vtn_pointer
*ptr
,
1101 enum gl_access_qualifier access
,
1102 struct vtn_ssa_value
**inout
)
1104 enum glsl_base_type base_type
= glsl_get_base_type(ptr
->type
->type
);
1105 switch (base_type
) {
1106 case GLSL_TYPE_UINT
:
1108 case GLSL_TYPE_UINT16
:
1109 case GLSL_TYPE_INT16
:
1110 case GLSL_TYPE_UINT8
:
1111 case GLSL_TYPE_INT8
:
1112 case GLSL_TYPE_UINT64
:
1113 case GLSL_TYPE_INT64
:
1114 case GLSL_TYPE_FLOAT
:
1115 case GLSL_TYPE_FLOAT16
:
1116 case GLSL_TYPE_BOOL
:
1117 case GLSL_TYPE_DOUBLE
:
1118 if (glsl_type_is_vector_or_scalar(ptr
->type
->type
)) {
1119 /* We hit a vector or scalar; go ahead and emit the load[s] */
1120 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
1121 if (vtn_pointer_is_external_block(b
, ptr
)) {
1122 /* If it's external, we call nir_load/store_deref directly. The
1123 * vtn_local_load/store helpers are too clever and do magic to
1124 * avoid array derefs of vectors. That magic is both less
1125 * efficient than the direct load/store and, in the case of
1126 * stores, is broken because it creates a race condition if two
1127 * threads are writing to different components of the same vector
1128 * due to the load+insert+store it uses to emulate the array
1132 *inout
= vtn_create_ssa_value(b
, ptr
->type
->type
);
1133 (*inout
)->def
= nir_load_deref_with_access(&b
->nb
, deref
,
1134 ptr
->type
->access
| access
);
1136 nir_store_deref_with_access(&b
->nb
, deref
, (*inout
)->def
, ~0,
1137 ptr
->type
->access
| access
);
1141 *inout
= vtn_local_load(b
, deref
, ptr
->type
->access
| access
);
1143 vtn_local_store(b
, *inout
, deref
, ptr
->type
->access
| access
);
1150 case GLSL_TYPE_INTERFACE
:
1151 case GLSL_TYPE_ARRAY
:
1152 case GLSL_TYPE_STRUCT
: {
1153 unsigned elems
= glsl_get_length(ptr
->type
->type
);
1155 vtn_assert(*inout
== NULL
);
1156 *inout
= rzalloc(b
, struct vtn_ssa_value
);
1157 (*inout
)->type
= ptr
->type
->type
;
1158 (*inout
)->elems
= rzalloc_array(b
, struct vtn_ssa_value
*, elems
);
1161 struct vtn_access_chain chain
= {
1164 { .mode
= vtn_access_mode_literal
, },
1167 for (unsigned i
= 0; i
< elems
; i
++) {
1168 chain
.link
[0].id
= i
;
1169 struct vtn_pointer
*elem
= vtn_pointer_dereference(b
, ptr
, &chain
);
1170 _vtn_variable_load_store(b
, load
, elem
, ptr
->type
->access
| access
,
1171 &(*inout
)->elems
[i
]);
1177 vtn_fail("Invalid access chain type");
1181 struct vtn_ssa_value
*
1182 vtn_variable_load(struct vtn_builder
*b
, struct vtn_pointer
*src
)
1184 if (vtn_pointer_uses_ssa_offset(b
, src
)) {
1185 return vtn_block_load(b
, src
);
1187 struct vtn_ssa_value
*val
= NULL
;
1188 _vtn_variable_load_store(b
, true, src
, src
->access
, &val
);
1194 vtn_variable_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
1195 struct vtn_pointer
*dest
)
1197 if (vtn_pointer_uses_ssa_offset(b
, dest
)) {
1198 vtn_assert(dest
->mode
== vtn_variable_mode_ssbo
||
1199 dest
->mode
== vtn_variable_mode_workgroup
);
1200 vtn_block_store(b
, src
, dest
);
1202 _vtn_variable_load_store(b
, false, dest
, dest
->access
, &src
);
1207 _vtn_variable_copy(struct vtn_builder
*b
, struct vtn_pointer
*dest
,
1208 struct vtn_pointer
*src
)
1210 vtn_assert(src
->type
->type
== dest
->type
->type
);
1211 enum glsl_base_type base_type
= glsl_get_base_type(src
->type
->type
);
1212 switch (base_type
) {
1213 case GLSL_TYPE_UINT
:
1215 case GLSL_TYPE_UINT16
:
1216 case GLSL_TYPE_INT16
:
1217 case GLSL_TYPE_UINT8
:
1218 case GLSL_TYPE_INT8
:
1219 case GLSL_TYPE_UINT64
:
1220 case GLSL_TYPE_INT64
:
1221 case GLSL_TYPE_FLOAT
:
1222 case GLSL_TYPE_FLOAT16
:
1223 case GLSL_TYPE_DOUBLE
:
1224 case GLSL_TYPE_BOOL
:
1225 /* At this point, we have a scalar, vector, or matrix so we know that
1226 * there cannot be any structure splitting still in the way. By
1227 * stopping at the matrix level rather than the vector level, we
1228 * ensure that matrices get loaded in the optimal way even if they
1229 * are storred row-major in a UBO.
1231 vtn_variable_store(b
, vtn_variable_load(b
, src
), dest
);
1234 case GLSL_TYPE_INTERFACE
:
1235 case GLSL_TYPE_ARRAY
:
1236 case GLSL_TYPE_STRUCT
: {
1237 struct vtn_access_chain chain
= {
1240 { .mode
= vtn_access_mode_literal
, },
1243 unsigned elems
= glsl_get_length(src
->type
->type
);
1244 for (unsigned i
= 0; i
< elems
; i
++) {
1245 chain
.link
[0].id
= i
;
1246 struct vtn_pointer
*src_elem
=
1247 vtn_pointer_dereference(b
, src
, &chain
);
1248 struct vtn_pointer
*dest_elem
=
1249 vtn_pointer_dereference(b
, dest
, &chain
);
1251 _vtn_variable_copy(b
, dest_elem
, src_elem
);
1257 vtn_fail("Invalid access chain type");
1262 vtn_variable_copy(struct vtn_builder
*b
, struct vtn_pointer
*dest
,
1263 struct vtn_pointer
*src
)
1265 /* TODO: At some point, we should add a special-case for when we can
1266 * just emit a copy_var intrinsic.
1268 _vtn_variable_copy(b
, dest
, src
);
1272 set_mode_system_value(struct vtn_builder
*b
, nir_variable_mode
*mode
)
1274 vtn_assert(*mode
== nir_var_system_value
|| *mode
== nir_var_shader_in
);
1275 *mode
= nir_var_system_value
;
1279 vtn_get_builtin_location(struct vtn_builder
*b
,
1280 SpvBuiltIn builtin
, int *location
,
1281 nir_variable_mode
*mode
)
1284 case SpvBuiltInPosition
:
1285 *location
= VARYING_SLOT_POS
;
1287 case SpvBuiltInPointSize
:
1288 *location
= VARYING_SLOT_PSIZ
;
1290 case SpvBuiltInClipDistance
:
1291 *location
= VARYING_SLOT_CLIP_DIST0
; /* XXX CLIP_DIST1? */
1293 case SpvBuiltInCullDistance
:
1294 *location
= VARYING_SLOT_CULL_DIST0
;
1296 case SpvBuiltInVertexId
:
1297 case SpvBuiltInVertexIndex
:
1298 /* The Vulkan spec defines VertexIndex to be non-zero-based and doesn't
1299 * allow VertexId. The ARB_gl_spirv spec defines VertexId to be the
1300 * same as gl_VertexID, which is non-zero-based, and removes
1301 * VertexIndex. Since they're both defined to be non-zero-based, we use
1302 * SYSTEM_VALUE_VERTEX_ID for both.
1304 *location
= SYSTEM_VALUE_VERTEX_ID
;
1305 set_mode_system_value(b
, mode
);
1307 case SpvBuiltInInstanceIndex
:
1308 *location
= SYSTEM_VALUE_INSTANCE_INDEX
;
1309 set_mode_system_value(b
, mode
);
1311 case SpvBuiltInInstanceId
:
1312 *location
= SYSTEM_VALUE_INSTANCE_ID
;
1313 set_mode_system_value(b
, mode
);
1315 case SpvBuiltInPrimitiveId
:
1316 if (b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
) {
1317 vtn_assert(*mode
== nir_var_shader_in
);
1318 *location
= VARYING_SLOT_PRIMITIVE_ID
;
1319 } else if (*mode
== nir_var_shader_out
) {
1320 *location
= VARYING_SLOT_PRIMITIVE_ID
;
1322 *location
= SYSTEM_VALUE_PRIMITIVE_ID
;
1323 set_mode_system_value(b
, mode
);
1326 case SpvBuiltInInvocationId
:
1327 *location
= SYSTEM_VALUE_INVOCATION_ID
;
1328 set_mode_system_value(b
, mode
);
1330 case SpvBuiltInLayer
:
1331 *location
= VARYING_SLOT_LAYER
;
1332 if (b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
)
1333 *mode
= nir_var_shader_in
;
1334 else if (b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
)
1335 *mode
= nir_var_shader_out
;
1336 else if (b
->options
&& b
->options
->caps
.shader_viewport_index_layer
&&
1337 (b
->shader
->info
.stage
== MESA_SHADER_VERTEX
||
1338 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
))
1339 *mode
= nir_var_shader_out
;
1341 vtn_fail("invalid stage for SpvBuiltInLayer");
1343 case SpvBuiltInViewportIndex
:
1344 *location
= VARYING_SLOT_VIEWPORT
;
1345 if (b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
)
1346 *mode
= nir_var_shader_out
;
1347 else if (b
->options
&& b
->options
->caps
.shader_viewport_index_layer
&&
1348 (b
->shader
->info
.stage
== MESA_SHADER_VERTEX
||
1349 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
))
1350 *mode
= nir_var_shader_out
;
1351 else if (b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
)
1352 *mode
= nir_var_shader_in
;
1354 vtn_fail("invalid stage for SpvBuiltInViewportIndex");
1356 case SpvBuiltInTessLevelOuter
:
1357 *location
= VARYING_SLOT_TESS_LEVEL_OUTER
;
1359 case SpvBuiltInTessLevelInner
:
1360 *location
= VARYING_SLOT_TESS_LEVEL_INNER
;
1362 case SpvBuiltInTessCoord
:
1363 *location
= SYSTEM_VALUE_TESS_COORD
;
1364 set_mode_system_value(b
, mode
);
1366 case SpvBuiltInPatchVertices
:
1367 *location
= SYSTEM_VALUE_VERTICES_IN
;
1368 set_mode_system_value(b
, mode
);
1370 case SpvBuiltInFragCoord
:
1371 vtn_assert(*mode
== nir_var_shader_in
);
1372 if (b
->options
&& b
->options
->frag_coord_is_sysval
) {
1373 *mode
= nir_var_system_value
;
1374 *location
= SYSTEM_VALUE_FRAG_COORD
;
1376 *location
= VARYING_SLOT_POS
;
1379 case SpvBuiltInPointCoord
:
1380 *location
= VARYING_SLOT_PNTC
;
1381 vtn_assert(*mode
== nir_var_shader_in
);
1383 case SpvBuiltInFrontFacing
:
1384 *location
= SYSTEM_VALUE_FRONT_FACE
;
1385 set_mode_system_value(b
, mode
);
1387 case SpvBuiltInSampleId
:
1388 *location
= SYSTEM_VALUE_SAMPLE_ID
;
1389 set_mode_system_value(b
, mode
);
1391 case SpvBuiltInSamplePosition
:
1392 *location
= SYSTEM_VALUE_SAMPLE_POS
;
1393 set_mode_system_value(b
, mode
);
1395 case SpvBuiltInSampleMask
:
1396 if (*mode
== nir_var_shader_out
) {
1397 *location
= FRAG_RESULT_SAMPLE_MASK
;
1399 *location
= SYSTEM_VALUE_SAMPLE_MASK_IN
;
1400 set_mode_system_value(b
, mode
);
1403 case SpvBuiltInFragDepth
:
1404 *location
= FRAG_RESULT_DEPTH
;
1405 vtn_assert(*mode
== nir_var_shader_out
);
1407 case SpvBuiltInHelperInvocation
:
1408 *location
= SYSTEM_VALUE_HELPER_INVOCATION
;
1409 set_mode_system_value(b
, mode
);
1411 case SpvBuiltInNumWorkgroups
:
1412 *location
= SYSTEM_VALUE_NUM_WORK_GROUPS
;
1413 set_mode_system_value(b
, mode
);
1415 case SpvBuiltInWorkgroupSize
:
1416 *location
= SYSTEM_VALUE_LOCAL_GROUP_SIZE
;
1417 set_mode_system_value(b
, mode
);
1419 case SpvBuiltInWorkgroupId
:
1420 *location
= SYSTEM_VALUE_WORK_GROUP_ID
;
1421 set_mode_system_value(b
, mode
);
1423 case SpvBuiltInLocalInvocationId
:
1424 *location
= SYSTEM_VALUE_LOCAL_INVOCATION_ID
;
1425 set_mode_system_value(b
, mode
);
1427 case SpvBuiltInLocalInvocationIndex
:
1428 *location
= SYSTEM_VALUE_LOCAL_INVOCATION_INDEX
;
1429 set_mode_system_value(b
, mode
);
1431 case SpvBuiltInGlobalInvocationId
:
1432 *location
= SYSTEM_VALUE_GLOBAL_INVOCATION_ID
;
1433 set_mode_system_value(b
, mode
);
1435 case SpvBuiltInGlobalLinearId
:
1436 *location
= SYSTEM_VALUE_GLOBAL_INVOCATION_INDEX
;
1437 set_mode_system_value(b
, mode
);
1439 case SpvBuiltInBaseVertex
:
1440 /* OpenGL gl_BaseVertex (SYSTEM_VALUE_BASE_VERTEX) is not the same
1441 * semantic as Vulkan BaseVertex (SYSTEM_VALUE_FIRST_VERTEX).
1443 if (b
->options
->environment
== NIR_SPIRV_OPENGL
)
1444 *location
= SYSTEM_VALUE_BASE_VERTEX
;
1446 *location
= SYSTEM_VALUE_FIRST_VERTEX
;
1447 set_mode_system_value(b
, mode
);
1449 case SpvBuiltInBaseInstance
:
1450 *location
= SYSTEM_VALUE_BASE_INSTANCE
;
1451 set_mode_system_value(b
, mode
);
1453 case SpvBuiltInDrawIndex
:
1454 *location
= SYSTEM_VALUE_DRAW_ID
;
1455 set_mode_system_value(b
, mode
);
1457 case SpvBuiltInSubgroupSize
:
1458 *location
= SYSTEM_VALUE_SUBGROUP_SIZE
;
1459 set_mode_system_value(b
, mode
);
1461 case SpvBuiltInSubgroupId
:
1462 *location
= SYSTEM_VALUE_SUBGROUP_ID
;
1463 set_mode_system_value(b
, mode
);
1465 case SpvBuiltInSubgroupLocalInvocationId
:
1466 *location
= SYSTEM_VALUE_SUBGROUP_INVOCATION
;
1467 set_mode_system_value(b
, mode
);
1469 case SpvBuiltInNumSubgroups
:
1470 *location
= SYSTEM_VALUE_NUM_SUBGROUPS
;
1471 set_mode_system_value(b
, mode
);
1473 case SpvBuiltInDeviceIndex
:
1474 *location
= SYSTEM_VALUE_DEVICE_INDEX
;
1475 set_mode_system_value(b
, mode
);
1477 case SpvBuiltInViewIndex
:
1478 *location
= SYSTEM_VALUE_VIEW_INDEX
;
1479 set_mode_system_value(b
, mode
);
1481 case SpvBuiltInSubgroupEqMask
:
1482 *location
= SYSTEM_VALUE_SUBGROUP_EQ_MASK
,
1483 set_mode_system_value(b
, mode
);
1485 case SpvBuiltInSubgroupGeMask
:
1486 *location
= SYSTEM_VALUE_SUBGROUP_GE_MASK
,
1487 set_mode_system_value(b
, mode
);
1489 case SpvBuiltInSubgroupGtMask
:
1490 *location
= SYSTEM_VALUE_SUBGROUP_GT_MASK
,
1491 set_mode_system_value(b
, mode
);
1493 case SpvBuiltInSubgroupLeMask
:
1494 *location
= SYSTEM_VALUE_SUBGROUP_LE_MASK
,
1495 set_mode_system_value(b
, mode
);
1497 case SpvBuiltInSubgroupLtMask
:
1498 *location
= SYSTEM_VALUE_SUBGROUP_LT_MASK
,
1499 set_mode_system_value(b
, mode
);
1501 case SpvBuiltInFragStencilRefEXT
:
1502 *location
= FRAG_RESULT_STENCIL
;
1503 vtn_assert(*mode
== nir_var_shader_out
);
1505 case SpvBuiltInWorkDim
:
1506 *location
= SYSTEM_VALUE_WORK_DIM
;
1507 set_mode_system_value(b
, mode
);
1509 case SpvBuiltInGlobalSize
:
1510 *location
= SYSTEM_VALUE_GLOBAL_GROUP_SIZE
;
1511 set_mode_system_value(b
, mode
);
1514 vtn_fail("Unsupported builtin: %s (%u)",
1515 spirv_builtin_to_string(builtin
), builtin
);
1520 apply_var_decoration(struct vtn_builder
*b
,
1521 struct nir_variable_data
*var_data
,
1522 const struct vtn_decoration
*dec
)
1524 switch (dec
->decoration
) {
1525 case SpvDecorationRelaxedPrecision
:
1526 break; /* FIXME: Do nothing with this for now. */
1527 case SpvDecorationNoPerspective
:
1528 var_data
->interpolation
= INTERP_MODE_NOPERSPECTIVE
;
1530 case SpvDecorationFlat
:
1531 var_data
->interpolation
= INTERP_MODE_FLAT
;
1533 case SpvDecorationCentroid
:
1534 var_data
->centroid
= true;
1536 case SpvDecorationSample
:
1537 var_data
->sample
= true;
1539 case SpvDecorationInvariant
:
1540 var_data
->invariant
= true;
1542 case SpvDecorationConstant
:
1543 var_data
->read_only
= true;
1545 case SpvDecorationNonReadable
:
1546 var_data
->image
.access
|= ACCESS_NON_READABLE
;
1548 case SpvDecorationNonWritable
:
1549 var_data
->read_only
= true;
1550 var_data
->image
.access
|= ACCESS_NON_WRITEABLE
;
1552 case SpvDecorationRestrict
:
1553 var_data
->image
.access
|= ACCESS_RESTRICT
;
1555 case SpvDecorationVolatile
:
1556 var_data
->image
.access
|= ACCESS_VOLATILE
;
1558 case SpvDecorationCoherent
:
1559 var_data
->image
.access
|= ACCESS_COHERENT
;
1561 case SpvDecorationComponent
:
1562 var_data
->location_frac
= dec
->operands
[0];
1564 case SpvDecorationIndex
:
1565 var_data
->index
= dec
->operands
[0];
1567 case SpvDecorationBuiltIn
: {
1568 SpvBuiltIn builtin
= dec
->operands
[0];
1570 nir_variable_mode mode
= var_data
->mode
;
1571 vtn_get_builtin_location(b
, builtin
, &var_data
->location
, &mode
);
1572 var_data
->mode
= mode
;
1575 case SpvBuiltInTessLevelOuter
:
1576 case SpvBuiltInTessLevelInner
:
1577 case SpvBuiltInClipDistance
:
1578 case SpvBuiltInCullDistance
:
1579 var_data
->compact
= true;
1586 case SpvDecorationSpecId
:
1587 case SpvDecorationRowMajor
:
1588 case SpvDecorationColMajor
:
1589 case SpvDecorationMatrixStride
:
1590 case SpvDecorationAliased
:
1591 case SpvDecorationUniform
:
1592 case SpvDecorationUniformId
:
1593 case SpvDecorationLinkageAttributes
:
1594 break; /* Do nothing with these here */
1596 case SpvDecorationPatch
:
1597 var_data
->patch
= true;
1600 case SpvDecorationLocation
:
1601 vtn_fail("Handled above");
1603 case SpvDecorationBlock
:
1604 case SpvDecorationBufferBlock
:
1605 case SpvDecorationArrayStride
:
1606 case SpvDecorationGLSLShared
:
1607 case SpvDecorationGLSLPacked
:
1608 break; /* These can apply to a type but we don't care about them */
1610 case SpvDecorationBinding
:
1611 case SpvDecorationDescriptorSet
:
1612 case SpvDecorationNoContraction
:
1613 case SpvDecorationInputAttachmentIndex
:
1614 vtn_warn("Decoration not allowed for variable or structure member: %s",
1615 spirv_decoration_to_string(dec
->decoration
));
1618 case SpvDecorationXfbBuffer
:
1619 var_data
->explicit_xfb_buffer
= true;
1620 var_data
->xfb
.buffer
= dec
->operands
[0];
1621 var_data
->always_active_io
= true;
1623 case SpvDecorationXfbStride
:
1624 var_data
->explicit_xfb_stride
= true;
1625 var_data
->xfb
.stride
= dec
->operands
[0];
1627 case SpvDecorationOffset
:
1628 var_data
->explicit_offset
= true;
1629 var_data
->offset
= dec
->operands
[0];
1632 case SpvDecorationStream
:
1633 var_data
->stream
= dec
->operands
[0];
1636 case SpvDecorationCPacked
:
1637 case SpvDecorationSaturatedConversion
:
1638 case SpvDecorationFuncParamAttr
:
1639 case SpvDecorationFPRoundingMode
:
1640 case SpvDecorationFPFastMathMode
:
1641 case SpvDecorationAlignment
:
1642 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
) {
1643 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1644 spirv_decoration_to_string(dec
->decoration
));
1648 case SpvDecorationUserSemantic
:
1649 /* User semantic decorations can safely be ignored by the driver. */
1652 case SpvDecorationRestrictPointerEXT
:
1653 case SpvDecorationAliasedPointerEXT
:
1654 /* TODO: We should actually plumb alias information through NIR. */
1658 vtn_fail_with_decoration("Unhandled decoration", dec
->decoration
);
1663 var_is_patch_cb(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
1664 const struct vtn_decoration
*dec
, void *out_is_patch
)
1666 if (dec
->decoration
== SpvDecorationPatch
) {
1667 *((bool *) out_is_patch
) = true;
1672 var_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
1673 const struct vtn_decoration
*dec
, void *void_var
)
1675 struct vtn_variable
*vtn_var
= void_var
;
1677 /* Handle decorations that apply to a vtn_variable as a whole */
1678 switch (dec
->decoration
) {
1679 case SpvDecorationBinding
:
1680 vtn_var
->binding
= dec
->operands
[0];
1681 vtn_var
->explicit_binding
= true;
1683 case SpvDecorationDescriptorSet
:
1684 vtn_var
->descriptor_set
= dec
->operands
[0];
1686 case SpvDecorationInputAttachmentIndex
:
1687 vtn_var
->input_attachment_index
= dec
->operands
[0];
1689 case SpvDecorationPatch
:
1690 vtn_var
->patch
= true;
1692 case SpvDecorationOffset
:
1693 vtn_var
->offset
= dec
->operands
[0];
1695 case SpvDecorationNonWritable
:
1696 vtn_var
->access
|= ACCESS_NON_WRITEABLE
;
1698 case SpvDecorationNonReadable
:
1699 vtn_var
->access
|= ACCESS_NON_READABLE
;
1701 case SpvDecorationVolatile
:
1702 vtn_var
->access
|= ACCESS_VOLATILE
;
1704 case SpvDecorationCoherent
:
1705 vtn_var
->access
|= ACCESS_COHERENT
;
1707 case SpvDecorationCounterBuffer
:
1708 /* Counter buffer decorations can safely be ignored by the driver. */
1714 if (val
->value_type
== vtn_value_type_pointer
) {
1715 assert(val
->pointer
->var
== void_var
);
1716 assert(member
== -1);
1718 assert(val
->value_type
== vtn_value_type_type
);
1721 /* Location is odd. If applied to a split structure, we have to walk the
1722 * whole thing and accumulate the location. It's easier to handle as a
1725 if (dec
->decoration
== SpvDecorationLocation
) {
1726 unsigned location
= dec
->operands
[0];
1727 if (b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
&&
1728 vtn_var
->mode
== vtn_variable_mode_output
) {
1729 location
+= FRAG_RESULT_DATA0
;
1730 } else if (b
->shader
->info
.stage
== MESA_SHADER_VERTEX
&&
1731 vtn_var
->mode
== vtn_variable_mode_input
) {
1732 location
+= VERT_ATTRIB_GENERIC0
;
1733 } else if (vtn_var
->mode
== vtn_variable_mode_input
||
1734 vtn_var
->mode
== vtn_variable_mode_output
) {
1735 location
+= vtn_var
->patch
? VARYING_SLOT_PATCH0
: VARYING_SLOT_VAR0
;
1736 } else if (vtn_var
->mode
!= vtn_variable_mode_uniform
) {
1737 vtn_warn("Location must be on input, output, uniform, sampler or "
1742 if (vtn_var
->var
->num_members
== 0) {
1743 /* This handles the member and lone variable cases */
1744 vtn_var
->var
->data
.location
= location
;
1746 /* This handles the structure member case */
1747 assert(vtn_var
->var
->members
);
1750 vtn_var
->base_location
= location
;
1752 vtn_var
->var
->members
[member
].location
= location
;
1758 if (vtn_var
->var
->num_members
== 0) {
1759 /* We call this function on types as well as variables and not all
1760 * struct types get split so we can end up having stray member
1761 * decorations; just ignore them.
1764 apply_var_decoration(b
, &vtn_var
->var
->data
, dec
);
1765 } else if (member
>= 0) {
1766 /* Member decorations must come from a type */
1767 assert(val
->value_type
== vtn_value_type_type
);
1768 apply_var_decoration(b
, &vtn_var
->var
->members
[member
], dec
);
1771 glsl_get_length(glsl_without_array(vtn_var
->type
->type
));
1772 for (unsigned i
= 0; i
< length
; i
++)
1773 apply_var_decoration(b
, &vtn_var
->var
->members
[i
], dec
);
1776 /* A few variables, those with external storage, have no actual
1777 * nir_variables associated with them. Fortunately, all decorations
1778 * we care about for those variables are on the type only.
1780 vtn_assert(vtn_var
->mode
== vtn_variable_mode_ubo
||
1781 vtn_var
->mode
== vtn_variable_mode_ssbo
||
1782 vtn_var
->mode
== vtn_variable_mode_push_constant
);
1787 enum vtn_variable_mode
1788 vtn_storage_class_to_mode(struct vtn_builder
*b
,
1789 SpvStorageClass
class,
1790 struct vtn_type
*interface_type
,
1791 nir_variable_mode
*nir_mode_out
)
1793 enum vtn_variable_mode mode
;
1794 nir_variable_mode nir_mode
;
1796 case SpvStorageClassUniform
:
1797 /* Assume it's an UBO if we lack the interface_type. */
1798 if (!interface_type
|| interface_type
->block
) {
1799 mode
= vtn_variable_mode_ubo
;
1800 nir_mode
= nir_var_mem_ubo
;
1801 } else if (interface_type
->buffer_block
) {
1802 mode
= vtn_variable_mode_ssbo
;
1803 nir_mode
= nir_var_mem_ssbo
;
1805 /* Default-block uniforms, coming from gl_spirv */
1806 mode
= vtn_variable_mode_uniform
;
1807 nir_mode
= nir_var_uniform
;
1810 case SpvStorageClassStorageBuffer
:
1811 mode
= vtn_variable_mode_ssbo
;
1812 nir_mode
= nir_var_mem_ssbo
;
1814 case SpvStorageClassPhysicalStorageBufferEXT
:
1815 mode
= vtn_variable_mode_phys_ssbo
;
1816 nir_mode
= nir_var_mem_global
;
1818 case SpvStorageClassUniformConstant
:
1819 mode
= vtn_variable_mode_uniform
;
1820 nir_mode
= nir_var_uniform
;
1822 case SpvStorageClassPushConstant
:
1823 mode
= vtn_variable_mode_push_constant
;
1824 nir_mode
= nir_var_uniform
;
1826 case SpvStorageClassInput
:
1827 mode
= vtn_variable_mode_input
;
1828 nir_mode
= nir_var_shader_in
;
1830 case SpvStorageClassOutput
:
1831 mode
= vtn_variable_mode_output
;
1832 nir_mode
= nir_var_shader_out
;
1834 case SpvStorageClassPrivate
:
1835 mode
= vtn_variable_mode_private
;
1836 nir_mode
= nir_var_shader_temp
;
1838 case SpvStorageClassFunction
:
1839 mode
= vtn_variable_mode_function
;
1840 nir_mode
= nir_var_function_temp
;
1842 case SpvStorageClassWorkgroup
:
1843 mode
= vtn_variable_mode_workgroup
;
1844 nir_mode
= nir_var_mem_shared
;
1846 case SpvStorageClassAtomicCounter
:
1847 mode
= vtn_variable_mode_uniform
;
1848 nir_mode
= nir_var_uniform
;
1850 case SpvStorageClassCrossWorkgroup
:
1851 mode
= vtn_variable_mode_cross_workgroup
;
1852 nir_mode
= nir_var_mem_global
;
1854 case SpvStorageClassImage
:
1855 mode
= vtn_variable_mode_image
;
1856 nir_mode
= nir_var_mem_ubo
;
1858 case SpvStorageClassGeneric
:
1860 vtn_fail("Unhandled variable storage class: %s (%u)",
1861 spirv_storageclass_to_string(class), class);
1865 *nir_mode_out
= nir_mode
;
1871 vtn_mode_to_address_format(struct vtn_builder
*b
, enum vtn_variable_mode mode
)
1874 case vtn_variable_mode_ubo
:
1875 return b
->options
->ubo_addr_format
;
1877 case vtn_variable_mode_ssbo
:
1878 return b
->options
->ssbo_addr_format
;
1880 case vtn_variable_mode_phys_ssbo
:
1881 return b
->options
->phys_ssbo_addr_format
;
1883 case vtn_variable_mode_push_constant
:
1884 return b
->options
->push_const_addr_format
;
1886 case vtn_variable_mode_workgroup
:
1887 return b
->options
->shared_addr_format
;
1889 case vtn_variable_mode_cross_workgroup
:
1890 return b
->options
->global_addr_format
;
1892 case vtn_variable_mode_function
:
1893 if (b
->physical_ptrs
)
1894 return b
->options
->temp_addr_format
;
1897 case vtn_variable_mode_private
:
1898 case vtn_variable_mode_uniform
:
1899 case vtn_variable_mode_input
:
1900 case vtn_variable_mode_output
:
1901 case vtn_variable_mode_image
:
1902 return nir_address_format_logical
;
1905 unreachable("Invalid variable mode");
1909 vtn_pointer_to_ssa(struct vtn_builder
*b
, struct vtn_pointer
*ptr
)
1911 if (vtn_pointer_uses_ssa_offset(b
, ptr
)) {
1912 /* This pointer needs to have a pointer type with actual storage */
1913 vtn_assert(ptr
->ptr_type
);
1914 vtn_assert(ptr
->ptr_type
->type
);
1917 /* If we don't have an offset then we must be a pointer to the variable
1920 vtn_assert(!ptr
->offset
&& !ptr
->block_index
);
1922 struct vtn_access_chain chain
= {
1925 ptr
= vtn_ssa_offset_pointer_dereference(b
, ptr
, &chain
);
1928 vtn_assert(ptr
->offset
);
1929 if (ptr
->block_index
) {
1930 vtn_assert(ptr
->mode
== vtn_variable_mode_ubo
||
1931 ptr
->mode
== vtn_variable_mode_ssbo
);
1932 return nir_vec2(&b
->nb
, ptr
->block_index
, ptr
->offset
);
1934 vtn_assert(ptr
->mode
== vtn_variable_mode_workgroup
);
1938 if (vtn_pointer_is_external_block(b
, ptr
) &&
1939 vtn_type_contains_block(b
, ptr
->type
) &&
1940 ptr
->mode
!= vtn_variable_mode_phys_ssbo
) {
1941 /* In this case, we're looking for a block index and not an actual
1944 * For PhysicalStorageBufferEXT pointers, we don't have a block index
1945 * at all because we get the pointer directly from the client. This
1946 * assumes that there will never be a SSBO binding variable using the
1947 * PhysicalStorageBufferEXT storage class. This assumption appears
1948 * to be correct according to the Vulkan spec because the table,
1949 * "Shader Resource and Storage Class Correspondence," the only the
1950 * Uniform storage class with BufferBlock or the StorageBuffer
1951 * storage class with Block can be used.
1953 if (!ptr
->block_index
) {
1954 /* If we don't have a block_index then we must be a pointer to the
1957 vtn_assert(!ptr
->deref
);
1959 struct vtn_access_chain chain
= {
1962 ptr
= vtn_nir_deref_pointer_dereference(b
, ptr
, &chain
);
1965 return ptr
->block_index
;
1967 return &vtn_pointer_to_deref(b
, ptr
)->dest
.ssa
;
1972 struct vtn_pointer
*
1973 vtn_pointer_from_ssa(struct vtn_builder
*b
, nir_ssa_def
*ssa
,
1974 struct vtn_type
*ptr_type
)
1976 vtn_assert(ptr_type
->base_type
== vtn_base_type_pointer
);
1978 struct vtn_pointer
*ptr
= rzalloc(b
, struct vtn_pointer
);
1979 struct vtn_type
*without_array
=
1980 vtn_type_without_array(ptr_type
->deref
);
1982 nir_variable_mode nir_mode
;
1983 ptr
->mode
= vtn_storage_class_to_mode(b
, ptr_type
->storage_class
,
1984 without_array
, &nir_mode
);
1985 ptr
->type
= ptr_type
->deref
;
1986 ptr
->ptr_type
= ptr_type
;
1988 if (b
->wa_glslang_179
) {
1989 /* To work around https://github.com/KhronosGroup/glslang/issues/179 we
1990 * need to whack the mode because it creates a function parameter with
1991 * the Function storage class even though it's a pointer to a sampler.
1992 * If we don't do this, then NIR won't get rid of the deref_cast for us.
1994 if (ptr
->mode
== vtn_variable_mode_function
&&
1995 (ptr
->type
->base_type
== vtn_base_type_sampler
||
1996 ptr
->type
->base_type
== vtn_base_type_sampled_image
)) {
1997 ptr
->mode
= vtn_variable_mode_uniform
;
1998 nir_mode
= nir_var_uniform
;
2002 if (vtn_pointer_uses_ssa_offset(b
, ptr
)) {
2003 /* This pointer type needs to have actual storage */
2004 vtn_assert(ptr_type
->type
);
2005 if (ptr
->mode
== vtn_variable_mode_ubo
||
2006 ptr
->mode
== vtn_variable_mode_ssbo
) {
2007 vtn_assert(ssa
->num_components
== 2);
2008 ptr
->block_index
= nir_channel(&b
->nb
, ssa
, 0);
2009 ptr
->offset
= nir_channel(&b
->nb
, ssa
, 1);
2011 vtn_assert(ssa
->num_components
== 1);
2012 ptr
->block_index
= NULL
;
2016 const struct glsl_type
*deref_type
= ptr_type
->deref
->type
;
2017 if (!vtn_pointer_is_external_block(b
, ptr
)) {
2018 ptr
->deref
= nir_build_deref_cast(&b
->nb
, ssa
, nir_mode
,
2019 deref_type
, ptr_type
->stride
);
2020 } else if (vtn_type_contains_block(b
, ptr
->type
) &&
2021 ptr
->mode
!= vtn_variable_mode_phys_ssbo
) {
2022 /* This is a pointer to somewhere in an array of blocks, not a
2023 * pointer to somewhere inside the block. Set the block index
2024 * instead of making a cast.
2026 ptr
->block_index
= ssa
;
2028 /* This is a pointer to something internal or a pointer inside a
2029 * block. It's just a regular cast.
2031 * For PhysicalStorageBufferEXT pointers, we don't have a block index
2032 * at all because we get the pointer directly from the client. This
2033 * assumes that there will never be a SSBO binding variable using the
2034 * PhysicalStorageBufferEXT storage class. This assumption appears
2035 * to be correct according to the Vulkan spec because the table,
2036 * "Shader Resource and Storage Class Correspondence," the only the
2037 * Uniform storage class with BufferBlock or the StorageBuffer
2038 * storage class with Block can be used.
2040 ptr
->deref
= nir_build_deref_cast(&b
->nb
, ssa
, nir_mode
,
2041 ptr_type
->deref
->type
,
2043 ptr
->deref
->dest
.ssa
.num_components
=
2044 glsl_get_vector_elements(ptr_type
->type
);
2045 ptr
->deref
->dest
.ssa
.bit_size
= glsl_get_bit_size(ptr_type
->type
);
2053 is_per_vertex_inout(const struct vtn_variable
*var
, gl_shader_stage stage
)
2055 if (var
->patch
|| !glsl_type_is_array(var
->type
->type
))
2058 if (var
->mode
== vtn_variable_mode_input
) {
2059 return stage
== MESA_SHADER_TESS_CTRL
||
2060 stage
== MESA_SHADER_TESS_EVAL
||
2061 stage
== MESA_SHADER_GEOMETRY
;
2064 if (var
->mode
== vtn_variable_mode_output
)
2065 return stage
== MESA_SHADER_TESS_CTRL
;
2071 assign_missing_member_locations(struct vtn_variable
*var
)
2074 glsl_get_length(glsl_without_array(var
->type
->type
));
2075 int location
= var
->base_location
;
2077 for (unsigned i
= 0; i
< length
; i
++) {
2078 /* From the Vulkan spec:
2080 * “If the structure type is a Block but without a Location, then each
2081 * of its members must have a Location decoration.”
2084 if (var
->type
->block
) {
2085 assert(var
->base_location
!= -1 ||
2086 var
->var
->members
[i
].location
!= -1);
2089 /* From the Vulkan spec:
2091 * “Any member with its own Location decoration is assigned that
2092 * location. Each remaining member is assigned the location after the
2093 * immediately preceding member in declaration order.”
2095 if (var
->var
->members
[i
].location
!= -1)
2096 location
= var
->var
->members
[i
].location
;
2098 var
->var
->members
[i
].location
= location
;
2100 /* Below we use type instead of interface_type, because interface_type
2101 * is only available when it is a Block. This code also supports
2102 * input/outputs that are just structs
2104 const struct glsl_type
*member_type
=
2105 glsl_get_struct_field(glsl_without_array(var
->type
->type
), i
);
2108 glsl_count_attribute_slots(member_type
,
2109 false /* is_gl_vertex_input */);
2115 vtn_create_variable(struct vtn_builder
*b
, struct vtn_value
*val
,
2116 struct vtn_type
*ptr_type
, SpvStorageClass storage_class
,
2117 nir_constant
*initializer
)
2119 vtn_assert(ptr_type
->base_type
== vtn_base_type_pointer
);
2120 struct vtn_type
*type
= ptr_type
->deref
;
2122 struct vtn_type
*without_array
= vtn_type_without_array(ptr_type
->deref
);
2124 enum vtn_variable_mode mode
;
2125 nir_variable_mode nir_mode
;
2126 mode
= vtn_storage_class_to_mode(b
, storage_class
, without_array
, &nir_mode
);
2129 case vtn_variable_mode_ubo
:
2130 /* There's no other way to get vtn_variable_mode_ubo */
2131 vtn_assert(without_array
->block
);
2132 b
->shader
->info
.num_ubos
++;
2134 case vtn_variable_mode_ssbo
:
2135 if (storage_class
== SpvStorageClassStorageBuffer
&&
2136 !without_array
->block
) {
2137 if (b
->variable_pointers
) {
2138 vtn_fail("Variables in the StorageBuffer storage class must "
2139 "have a struct type with the Block decoration");
2141 /* If variable pointers are not present, it's still malformed
2142 * SPIR-V but we can parse it and do the right thing anyway.
2143 * Since some of the 8-bit storage tests have bugs in this are,
2144 * just make it a warning for now.
2146 vtn_warn("Variables in the StorageBuffer storage class must "
2147 "have a struct type with the Block decoration");
2150 b
->shader
->info
.num_ssbos
++;
2152 case vtn_variable_mode_uniform
:
2153 if (glsl_type_is_image(without_array
->type
))
2154 b
->shader
->info
.num_images
++;
2155 else if (glsl_type_is_sampler(without_array
->type
))
2156 b
->shader
->info
.num_textures
++;
2158 case vtn_variable_mode_push_constant
:
2159 b
->shader
->num_uniforms
= vtn_type_block_size(b
, type
);
2162 case vtn_variable_mode_image
:
2163 vtn_fail("Cannot create a variable with the Image storage class");
2166 case vtn_variable_mode_phys_ssbo
:
2167 vtn_fail("Cannot create a variable with the "
2168 "PhysicalStorageBufferEXT storage class");
2172 /* No tallying is needed */
2176 struct vtn_variable
*var
= rzalloc(b
, struct vtn_variable
);
2179 var
->base_location
= -1;
2181 vtn_assert(val
->value_type
== vtn_value_type_pointer
);
2182 val
->pointer
= vtn_pointer_for_variable(b
, var
, ptr_type
);
2184 switch (var
->mode
) {
2185 case vtn_variable_mode_function
:
2186 case vtn_variable_mode_private
:
2187 case vtn_variable_mode_uniform
:
2188 /* For these, we create the variable normally */
2189 var
->var
= rzalloc(b
->shader
, nir_variable
);
2190 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
2192 if (storage_class
== SpvStorageClassAtomicCounter
) {
2193 /* Need to tweak the nir type here as at vtn_handle_type we don't
2194 * have the access to storage_class, that is the one that points us
2195 * that is an atomic uint.
2197 var
->var
->type
= repair_atomic_type(var
->type
->type
);
2199 /* Private variables don't have any explicit layout but some layouts
2200 * may have leaked through due to type deduplication in the SPIR-V.
2202 var
->var
->type
= var
->type
->type
;
2204 var
->var
->data
.mode
= nir_mode
;
2205 var
->var
->data
.location
= -1;
2206 var
->var
->interface_type
= NULL
;
2209 case vtn_variable_mode_ubo
:
2210 case vtn_variable_mode_ssbo
:
2211 var
->var
= rzalloc(b
->shader
, nir_variable
);
2212 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
2214 var
->var
->type
= var
->type
->type
;
2215 var
->var
->interface_type
= var
->type
->type
;
2217 var
->var
->data
.mode
= nir_mode
;
2218 var
->var
->data
.location
= -1;
2222 case vtn_variable_mode_workgroup
:
2223 /* Create the variable normally */
2224 var
->var
= rzalloc(b
->shader
, nir_variable
);
2225 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
2226 /* Workgroup variables don't have any explicit layout but some
2227 * layouts may have leaked through due to type deduplication in the
2230 var
->var
->type
= var
->type
->type
;
2231 var
->var
->data
.mode
= nir_var_mem_shared
;
2234 case vtn_variable_mode_input
:
2235 case vtn_variable_mode_output
: {
2236 /* In order to know whether or not we're a per-vertex inout, we need
2237 * the patch qualifier. This means walking the variable decorations
2238 * early before we actually create any variables. Not a big deal.
2240 * GLSLang really likes to place decorations in the most interior
2241 * thing it possibly can. In particular, if you have a struct, it
2242 * will place the patch decorations on the struct members. This
2243 * should be handled by the variable splitting below just fine.
2245 * If you have an array-of-struct, things get even more weird as it
2246 * will place the patch decorations on the struct even though it's
2247 * inside an array and some of the members being patch and others not
2248 * makes no sense whatsoever. Since the only sensible thing is for
2249 * it to be all or nothing, we'll call it patch if any of the members
2250 * are declared patch.
2253 vtn_foreach_decoration(b
, val
, var_is_patch_cb
, &var
->patch
);
2254 if (glsl_type_is_array(var
->type
->type
) &&
2255 glsl_type_is_struct_or_ifc(without_array
->type
)) {
2256 vtn_foreach_decoration(b
, vtn_value(b
, without_array
->id
,
2257 vtn_value_type_type
),
2258 var_is_patch_cb
, &var
->patch
);
2261 /* For inputs and outputs, we immediately split structures. This
2262 * is for a couple of reasons. For one, builtins may all come in
2263 * a struct and we really want those split out into separate
2264 * variables. For another, interpolation qualifiers can be
2265 * applied to members of the top-level struct ane we need to be
2266 * able to preserve that information.
2269 struct vtn_type
*per_vertex_type
= var
->type
;
2270 if (is_per_vertex_inout(var
, b
->shader
->info
.stage
)) {
2271 /* In Geometry shaders (and some tessellation), inputs come
2272 * in per-vertex arrays. However, some builtins come in
2273 * non-per-vertex, hence the need for the is_array check. In
2274 * any case, there are no non-builtin arrays allowed so this
2275 * check should be sufficient.
2277 per_vertex_type
= var
->type
->array_element
;
2280 var
->var
= rzalloc(b
->shader
, nir_variable
);
2281 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
2282 /* In Vulkan, shader I/O variables don't have any explicit layout but
2283 * some layouts may have leaked through due to type deduplication in
2284 * the SPIR-V. We do, however, keep the layouts in the variable's
2285 * interface_type because we need offsets for XFB arrays of blocks.
2287 var
->var
->type
= var
->type
->type
;
2288 var
->var
->data
.mode
= nir_mode
;
2289 var
->var
->data
.patch
= var
->patch
;
2291 /* Figure out the interface block type. */
2292 struct vtn_type
*iface_type
= per_vertex_type
;
2293 if (var
->mode
== vtn_variable_mode_output
&&
2294 (b
->shader
->info
.stage
== MESA_SHADER_VERTEX
||
2295 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
||
2296 b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
)) {
2297 /* For vertex data outputs, we can end up with arrays of blocks for
2298 * transform feedback where each array element corresponds to a
2299 * different XFB output buffer.
2301 while (iface_type
->base_type
== vtn_base_type_array
)
2302 iface_type
= iface_type
->array_element
;
2304 if (iface_type
->base_type
== vtn_base_type_struct
&& iface_type
->block
)
2305 var
->var
->interface_type
= iface_type
->type
;
2307 if (per_vertex_type
->base_type
== vtn_base_type_struct
&&
2308 per_vertex_type
->block
) {
2309 /* It's a struct. Set it up as per-member. */
2310 var
->var
->num_members
= glsl_get_length(per_vertex_type
->type
);
2311 var
->var
->members
= rzalloc_array(var
->var
, struct nir_variable_data
,
2312 var
->var
->num_members
);
2314 for (unsigned i
= 0; i
< var
->var
->num_members
; i
++) {
2315 var
->var
->members
[i
].mode
= nir_mode
;
2316 var
->var
->members
[i
].patch
= var
->patch
;
2317 var
->var
->members
[i
].location
= -1;
2321 /* For inputs and outputs, we need to grab locations and builtin
2322 * information from the per-vertex type.
2324 vtn_foreach_decoration(b
, vtn_value(b
, per_vertex_type
->id
,
2325 vtn_value_type_type
),
2326 var_decoration_cb
, var
);
2330 case vtn_variable_mode_push_constant
:
2331 case vtn_variable_mode_cross_workgroup
:
2332 /* These don't need actual variables. */
2335 case vtn_variable_mode_image
:
2336 case vtn_variable_mode_phys_ssbo
:
2337 unreachable("Should have been caught before");
2341 var
->var
->constant_initializer
=
2342 nir_constant_clone(initializer
, var
->var
);
2345 vtn_foreach_decoration(b
, val
, var_decoration_cb
, var
);
2346 vtn_foreach_decoration(b
, val
, ptr_decoration_cb
, val
->pointer
);
2348 if ((var
->mode
== vtn_variable_mode_input
||
2349 var
->mode
== vtn_variable_mode_output
) &&
2350 var
->var
->members
) {
2351 assign_missing_member_locations(var
);
2354 if (var
->mode
== vtn_variable_mode_uniform
||
2355 var
->mode
== vtn_variable_mode_ubo
||
2356 var
->mode
== vtn_variable_mode_ssbo
) {
2357 /* XXX: We still need the binding information in the nir_variable
2358 * for these. We should fix that.
2360 var
->var
->data
.binding
= var
->binding
;
2361 var
->var
->data
.explicit_binding
= var
->explicit_binding
;
2362 var
->var
->data
.descriptor_set
= var
->descriptor_set
;
2363 var
->var
->data
.index
= var
->input_attachment_index
;
2364 var
->var
->data
.offset
= var
->offset
;
2366 if (glsl_type_is_image(without_array
->type
))
2367 var
->var
->data
.image
.format
= without_array
->image_format
;
2370 if (var
->mode
== vtn_variable_mode_function
) {
2371 vtn_assert(var
->var
!= NULL
&& var
->var
->members
== NULL
);
2372 nir_function_impl_add_variable(b
->nb
.impl
, var
->var
);
2373 } else if (var
->var
) {
2374 nir_shader_add_variable(b
->shader
, var
->var
);
2376 vtn_assert(vtn_pointer_is_external_block(b
, val
->pointer
));
2381 vtn_assert_types_equal(struct vtn_builder
*b
, SpvOp opcode
,
2382 struct vtn_type
*dst_type
,
2383 struct vtn_type
*src_type
)
2385 if (dst_type
->id
== src_type
->id
)
2388 if (vtn_types_compatible(b
, dst_type
, src_type
)) {
2389 /* Early versions of GLSLang would re-emit types unnecessarily and you
2390 * would end up with OpLoad, OpStore, or OpCopyMemory opcodes which have
2391 * mismatched source and destination types.
2393 * https://github.com/KhronosGroup/glslang/issues/304
2394 * https://github.com/KhronosGroup/glslang/issues/307
2395 * https://bugs.freedesktop.org/show_bug.cgi?id=104338
2396 * https://bugs.freedesktop.org/show_bug.cgi?id=104424
2398 vtn_warn("Source and destination types of %s do not have the same "
2399 "ID (but are compatible): %u vs %u",
2400 spirv_op_to_string(opcode
), dst_type
->id
, src_type
->id
);
2404 vtn_fail("Source and destination types of %s do not match: %s vs. %s",
2405 spirv_op_to_string(opcode
),
2406 glsl_get_type_name(dst_type
->type
),
2407 glsl_get_type_name(src_type
->type
));
2410 static nir_ssa_def
*
2411 nir_shrink_zero_pad_vec(nir_builder
*b
, nir_ssa_def
*val
,
2412 unsigned num_components
)
2414 if (val
->num_components
== num_components
)
2417 nir_ssa_def
*comps
[NIR_MAX_VEC_COMPONENTS
];
2418 for (unsigned i
= 0; i
< num_components
; i
++) {
2419 if (i
< val
->num_components
)
2420 comps
[i
] = nir_channel(b
, val
, i
);
2422 comps
[i
] = nir_imm_intN_t(b
, 0, val
->bit_size
);
2424 return nir_vec(b
, comps
, num_components
);
2427 static nir_ssa_def
*
2428 nir_sloppy_bitcast(nir_builder
*b
, nir_ssa_def
*val
,
2429 const struct glsl_type
*type
)
2431 const unsigned num_components
= glsl_get_vector_elements(type
);
2432 const unsigned bit_size
= glsl_get_bit_size(type
);
2434 /* First, zero-pad to ensure that the value is big enough that when we
2435 * bit-cast it, we don't loose anything.
2437 if (val
->bit_size
< bit_size
) {
2438 const unsigned src_num_components_needed
=
2439 vtn_align_u32(val
->num_components
, bit_size
/ val
->bit_size
);
2440 val
= nir_shrink_zero_pad_vec(b
, val
, src_num_components_needed
);
2443 val
= nir_bitcast_vector(b
, val
, bit_size
);
2445 return nir_shrink_zero_pad_vec(b
, val
, num_components
);
2449 vtn_handle_variables(struct vtn_builder
*b
, SpvOp opcode
,
2450 const uint32_t *w
, unsigned count
)
2454 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
2455 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2459 case SpvOpVariable
: {
2460 struct vtn_type
*ptr_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2462 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
2464 SpvStorageClass storage_class
= w
[3];
2465 nir_constant
*initializer
= NULL
;
2467 initializer
= vtn_value(b
, w
[4], vtn_value_type_constant
)->constant
;
2469 vtn_create_variable(b
, val
, ptr_type
, storage_class
, initializer
);
2473 case SpvOpAccessChain
:
2474 case SpvOpPtrAccessChain
:
2475 case SpvOpInBoundsAccessChain
:
2476 case SpvOpInBoundsPtrAccessChain
: {
2477 struct vtn_access_chain
*chain
= vtn_access_chain_create(b
, count
- 4);
2478 enum gl_access_qualifier access
= 0;
2479 chain
->ptr_as_array
= (opcode
== SpvOpPtrAccessChain
|| opcode
== SpvOpInBoundsPtrAccessChain
);
2482 for (int i
= 4; i
< count
; i
++) {
2483 struct vtn_value
*link_val
= vtn_untyped_value(b
, w
[i
]);
2484 if (link_val
->value_type
== vtn_value_type_constant
) {
2485 chain
->link
[idx
].mode
= vtn_access_mode_literal
;
2486 chain
->link
[idx
].id
= vtn_constant_int(b
, w
[i
]);
2488 chain
->link
[idx
].mode
= vtn_access_mode_id
;
2489 chain
->link
[idx
].id
= w
[i
];
2491 access
|= vtn_value_access(link_val
);
2495 struct vtn_type
*ptr_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2496 struct vtn_value
*base_val
= vtn_untyped_value(b
, w
[3]);
2497 if (base_val
->value_type
== vtn_value_type_sampled_image
) {
2498 /* This is rather insane. SPIR-V allows you to use OpSampledImage
2499 * to combine an array of images with a single sampler to get an
2500 * array of sampled images that all share the same sampler.
2501 * Fortunately, this means that we can more-or-less ignore the
2502 * sampler when crawling the access chain, but it does leave us
2503 * with this rather awkward little special-case.
2505 struct vtn_value
*val
=
2506 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
2507 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
2508 val
->sampled_image
->image
=
2509 vtn_pointer_dereference(b
, base_val
->sampled_image
->image
, chain
);
2510 val
->sampled_image
->sampler
= base_val
->sampled_image
->sampler
;
2511 val
->sampled_image
->image
=
2512 vtn_decorate_pointer(b
, val
, val
->sampled_image
->image
);
2513 val
->sampled_image
->sampler
=
2514 vtn_decorate_pointer(b
, val
, val
->sampled_image
->sampler
);
2516 vtn_assert(base_val
->value_type
== vtn_value_type_pointer
);
2517 struct vtn_pointer
*ptr
=
2518 vtn_pointer_dereference(b
, base_val
->pointer
, chain
);
2519 ptr
->ptr_type
= ptr_type
;
2520 ptr
->access
|= access
;
2521 vtn_push_value_pointer(b
, w
[2], ptr
);
2526 case SpvOpCopyMemory
: {
2527 struct vtn_value
*dest
= vtn_value(b
, w
[1], vtn_value_type_pointer
);
2528 struct vtn_value
*src
= vtn_value(b
, w
[2], vtn_value_type_pointer
);
2530 vtn_assert_types_equal(b
, opcode
, dest
->type
->deref
, src
->type
->deref
);
2532 vtn_variable_copy(b
, dest
->pointer
, src
->pointer
);
2537 struct vtn_type
*res_type
=
2538 vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2539 struct vtn_value
*src_val
= vtn_value(b
, w
[3], vtn_value_type_pointer
);
2540 struct vtn_pointer
*src
= src_val
->pointer
;
2542 vtn_assert_types_equal(b
, opcode
, res_type
, src_val
->type
->deref
);
2544 if (res_type
->base_type
== vtn_base_type_image
||
2545 res_type
->base_type
== vtn_base_type_sampler
) {
2546 vtn_push_value_pointer(b
, w
[2], src
);
2548 } else if (res_type
->base_type
== vtn_base_type_sampled_image
) {
2549 struct vtn_value
*val
=
2550 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
2551 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
2552 val
->sampled_image
->image
= val
->sampled_image
->sampler
=
2553 vtn_decorate_pointer(b
, val
, src
);
2559 SpvMemoryAccessMask access
= w
[4];
2560 if (access
& SpvMemoryAccessAlignedMask
)
2563 if (access
& SpvMemoryAccessMakePointerVisibleMask
) {
2564 SpvMemorySemanticsMask semantics
=
2565 SpvMemorySemanticsMakeVisibleMask
|
2566 vtn_storage_class_to_memory_semantics(src
->ptr_type
->storage_class
);
2568 SpvScope scope
= vtn_constant_uint(b
, w
[idx
]);
2569 vtn_emit_memory_barrier(b
, scope
, semantics
);
2573 vtn_push_ssa(b
, w
[2], res_type
, vtn_variable_load(b
, src
));
2578 struct vtn_value
*dest_val
= vtn_value(b
, w
[1], vtn_value_type_pointer
);
2579 struct vtn_pointer
*dest
= dest_val
->pointer
;
2580 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[2]);
2582 /* OpStore requires us to actually have a storage type */
2583 vtn_fail_if(dest
->type
->type
== NULL
,
2584 "Invalid destination type for OpStore");
2586 if (glsl_get_base_type(dest
->type
->type
) == GLSL_TYPE_BOOL
&&
2587 glsl_get_base_type(src_val
->type
->type
) == GLSL_TYPE_UINT
) {
2588 /* Early versions of GLSLang would use uint types for UBOs/SSBOs but
2589 * would then store them to a local variable as bool. Work around
2590 * the issue by doing an implicit conversion.
2592 * https://github.com/KhronosGroup/glslang/issues/170
2593 * https://bugs.freedesktop.org/show_bug.cgi?id=104424
2595 vtn_warn("OpStore of value of type OpTypeInt to a pointer to type "
2596 "OpTypeBool. Doing an implicit conversion to work around "
2598 struct vtn_ssa_value
*bool_ssa
=
2599 vtn_create_ssa_value(b
, dest
->type
->type
);
2600 bool_ssa
->def
= nir_i2b(&b
->nb
, vtn_ssa_value(b
, w
[2])->def
);
2601 vtn_variable_store(b
, bool_ssa
, dest
);
2605 vtn_assert_types_equal(b
, opcode
, dest_val
->type
->deref
, src_val
->type
);
2607 if (glsl_type_is_sampler(dest
->type
->type
)) {
2608 if (b
->wa_glslang_179
) {
2609 vtn_warn("OpStore of a sampler detected. Doing on-the-fly copy "
2610 "propagation to workaround the problem.");
2611 vtn_assert(dest
->var
->copy_prop_sampler
== NULL
);
2612 dest
->var
->copy_prop_sampler
=
2613 vtn_value(b
, w
[2], vtn_value_type_pointer
)->pointer
;
2615 vtn_fail("Vulkan does not allow OpStore of a sampler or image.");
2620 struct vtn_ssa_value
*src
= vtn_ssa_value(b
, w
[2]);
2621 vtn_variable_store(b
, src
, dest
);
2625 SpvMemoryAccessMask access
= w
[3];
2627 if (access
& SpvMemoryAccessAlignedMask
)
2630 if (access
& SpvMemoryAccessMakePointerAvailableMask
) {
2631 SpvMemorySemanticsMask semantics
=
2632 SpvMemorySemanticsMakeAvailableMask
|
2633 vtn_storage_class_to_memory_semantics(dest
->ptr_type
->storage_class
);
2634 SpvScope scope
= vtn_constant_uint(b
, w
[idx
]);
2635 vtn_emit_memory_barrier(b
, scope
, semantics
);
2641 case SpvOpArrayLength
: {
2642 struct vtn_pointer
*ptr
=
2643 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2644 const uint32_t field
= w
[4];
2646 vtn_fail_if(ptr
->type
->base_type
!= vtn_base_type_struct
,
2647 "OpArrayLength must take a pointer to a structure type");
2648 vtn_fail_if(field
!= ptr
->type
->length
- 1 ||
2649 ptr
->type
->members
[field
]->base_type
!= vtn_base_type_array
,
2650 "OpArrayLength must reference the last memeber of the "
2651 "structure and that must be an array");
2653 const uint32_t offset
= ptr
->type
->offsets
[field
];
2654 const uint32_t stride
= ptr
->type
->members
[field
]->stride
;
2656 if (!ptr
->block_index
) {
2657 struct vtn_access_chain chain
= {
2660 ptr
= vtn_pointer_dereference(b
, ptr
, &chain
);
2661 vtn_assert(ptr
->block_index
);
2664 nir_intrinsic_instr
*instr
=
2665 nir_intrinsic_instr_create(b
->nb
.shader
,
2666 nir_intrinsic_get_buffer_size
);
2667 instr
->src
[0] = nir_src_for_ssa(ptr
->block_index
);
2668 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
2669 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
2670 nir_ssa_def
*buf_size
= &instr
->dest
.ssa
;
2672 /* array_length = max(buffer_size - offset, 0) / stride */
2673 nir_ssa_def
*array_length
=
2678 nir_imm_int(&b
->nb
, offset
)),
2679 nir_imm_int(&b
->nb
, 0u)),
2680 nir_imm_int(&b
->nb
, stride
));
2682 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2683 val
->ssa
= vtn_create_ssa_value(b
, glsl_uint_type());
2684 val
->ssa
->def
= array_length
;
2688 case SpvOpConvertPtrToU
: {
2689 struct vtn_value
*u_val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2691 vtn_fail_if(u_val
->type
->base_type
!= vtn_base_type_vector
&&
2692 u_val
->type
->base_type
!= vtn_base_type_scalar
,
2693 "OpConvertPtrToU can only be used to cast to a vector or "
2696 /* The pointer will be converted to an SSA value automatically */
2697 struct vtn_ssa_value
*ptr_ssa
= vtn_ssa_value(b
, w
[3]);
2699 u_val
->ssa
= vtn_create_ssa_value(b
, u_val
->type
->type
);
2700 u_val
->ssa
->def
= nir_sloppy_bitcast(&b
->nb
, ptr_ssa
->def
, u_val
->type
->type
);
2701 u_val
->ssa
->access
|= ptr_ssa
->access
;
2705 case SpvOpConvertUToPtr
: {
2706 struct vtn_value
*ptr_val
=
2707 vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
2708 struct vtn_value
*u_val
= vtn_value(b
, w
[3], vtn_value_type_ssa
);
2710 vtn_fail_if(ptr_val
->type
->type
== NULL
,
2711 "OpConvertUToPtr can only be used on physical pointers");
2713 vtn_fail_if(u_val
->type
->base_type
!= vtn_base_type_vector
&&
2714 u_val
->type
->base_type
!= vtn_base_type_scalar
,
2715 "OpConvertUToPtr can only be used to cast from a vector or "
2718 nir_ssa_def
*ptr_ssa
= nir_sloppy_bitcast(&b
->nb
, u_val
->ssa
->def
,
2719 ptr_val
->type
->type
);
2720 ptr_val
->pointer
= vtn_pointer_from_ssa(b
, ptr_ssa
, ptr_val
->type
);
2721 vtn_foreach_decoration(b
, ptr_val
, ptr_decoration_cb
, ptr_val
->pointer
);
2722 ptr_val
->pointer
->access
|= u_val
->ssa
->access
;
2726 case SpvOpCopyMemorySized
:
2728 vtn_fail_with_opcode("Unhandled opcode", opcode
);