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_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 vtn_copy_value(struct vtn_builder
*b
, uint32_t src_value_id
,
81 uint32_t dst_value_id
)
83 struct vtn_value
*src
= vtn_untyped_value(b
, src_value_id
);
84 struct vtn_value
*dst
= vtn_untyped_value(b
, dst_value_id
);
85 struct vtn_value src_copy
= *src
;
87 vtn_fail_if(dst
->value_type
!= vtn_value_type_invalid
,
88 "SPIR-V id %u has already been written by another instruction",
91 vtn_fail_if(dst
->type
->id
!= src
->type
->id
,
92 "Result Type must equal Operand type");
94 src_copy
.name
= dst
->name
;
95 src_copy
.decoration
= dst
->decoration
;
96 src_copy
.type
= dst
->type
;
99 if (dst
->value_type
== vtn_value_type_pointer
)
100 dst
->pointer
= vtn_decorate_pointer(b
, dst
, dst
->pointer
);
103 static struct vtn_access_chain
*
104 vtn_access_chain_create(struct vtn_builder
*b
, unsigned length
)
106 struct vtn_access_chain
*chain
;
108 /* Subtract 1 from the length since there's already one built in */
109 size_t size
= sizeof(*chain
) +
110 (MAX2(length
, 1) - 1) * sizeof(chain
->link
[0]);
111 chain
= rzalloc_size(b
, size
);
112 chain
->length
= length
;
118 vtn_mode_uses_ssa_offset(struct vtn_builder
*b
,
119 enum vtn_variable_mode mode
)
121 return ((mode
== vtn_variable_mode_ubo
||
122 mode
== vtn_variable_mode_ssbo
) &&
123 b
->options
->lower_ubo_ssbo_access_to_offsets
) ||
124 mode
== vtn_variable_mode_push_constant
;
128 vtn_mode_is_cross_invocation(struct vtn_builder
*b
,
129 enum vtn_variable_mode mode
)
131 return mode
== vtn_variable_mode_ssbo
||
132 mode
== vtn_variable_mode_ubo
||
133 mode
== vtn_variable_mode_phys_ssbo
||
134 mode
== vtn_variable_mode_push_constant
||
135 mode
== vtn_variable_mode_workgroup
||
136 mode
== vtn_variable_mode_cross_workgroup
;
140 vtn_pointer_is_external_block(struct vtn_builder
*b
,
141 struct vtn_pointer
*ptr
)
143 return ptr
->mode
== vtn_variable_mode_ssbo
||
144 ptr
->mode
== vtn_variable_mode_ubo
||
145 ptr
->mode
== vtn_variable_mode_phys_ssbo
||
146 ptr
->mode
== vtn_variable_mode_push_constant
;
150 vtn_access_link_as_ssa(struct vtn_builder
*b
, struct vtn_access_link link
,
151 unsigned stride
, unsigned bit_size
)
153 vtn_assert(stride
> 0);
154 if (link
.mode
== vtn_access_mode_literal
) {
155 return nir_imm_intN_t(&b
->nb
, link
.id
* stride
, bit_size
);
157 nir_ssa_def
*ssa
= vtn_ssa_value(b
, link
.id
)->def
;
158 if (ssa
->bit_size
!= bit_size
)
159 ssa
= nir_i2i(&b
->nb
, ssa
, bit_size
);
160 return nir_imul_imm(&b
->nb
, ssa
, stride
);
164 static VkDescriptorType
165 vk_desc_type_for_mode(struct vtn_builder
*b
, enum vtn_variable_mode mode
)
168 case vtn_variable_mode_ubo
:
169 return VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER
;
170 case vtn_variable_mode_ssbo
:
171 return VK_DESCRIPTOR_TYPE_STORAGE_BUFFER
;
173 vtn_fail("Invalid mode for vulkan_resource_index");
178 vtn_variable_resource_index(struct vtn_builder
*b
, struct vtn_variable
*var
,
179 nir_ssa_def
*desc_array_index
)
181 vtn_assert(b
->options
->environment
== NIR_SPIRV_VULKAN
);
183 if (!desc_array_index
) {
184 vtn_assert(glsl_type_is_struct_or_ifc(var
->type
->type
));
185 desc_array_index
= nir_imm_int(&b
->nb
, 0);
188 nir_intrinsic_instr
*instr
=
189 nir_intrinsic_instr_create(b
->nb
.shader
,
190 nir_intrinsic_vulkan_resource_index
);
191 instr
->src
[0] = nir_src_for_ssa(desc_array_index
);
192 nir_intrinsic_set_desc_set(instr
, var
->descriptor_set
);
193 nir_intrinsic_set_binding(instr
, var
->binding
);
194 nir_intrinsic_set_desc_type(instr
, vk_desc_type_for_mode(b
, var
->mode
));
196 vtn_fail_if(var
->mode
!= vtn_variable_mode_ubo
&&
197 var
->mode
!= vtn_variable_mode_ssbo
,
198 "Invalid mode for vulkan_resource_index");
200 nir_address_format addr_format
= vtn_mode_to_address_format(b
, var
->mode
);
201 const struct glsl_type
*index_type
=
202 b
->options
->lower_ubo_ssbo_access_to_offsets
?
203 glsl_uint_type() : nir_address_format_to_glsl_type(addr_format
);
205 instr
->num_components
= glsl_get_vector_elements(index_type
);
206 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, instr
->num_components
,
207 glsl_get_bit_size(index_type
), NULL
);
208 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
210 return &instr
->dest
.ssa
;
214 vtn_resource_reindex(struct vtn_builder
*b
, enum vtn_variable_mode mode
,
215 nir_ssa_def
*base_index
, nir_ssa_def
*offset_index
)
217 vtn_assert(b
->options
->environment
== NIR_SPIRV_VULKAN
);
219 nir_intrinsic_instr
*instr
=
220 nir_intrinsic_instr_create(b
->nb
.shader
,
221 nir_intrinsic_vulkan_resource_reindex
);
222 instr
->src
[0] = nir_src_for_ssa(base_index
);
223 instr
->src
[1] = nir_src_for_ssa(offset_index
);
224 nir_intrinsic_set_desc_type(instr
, vk_desc_type_for_mode(b
, mode
));
226 vtn_fail_if(mode
!= vtn_variable_mode_ubo
&& mode
!= vtn_variable_mode_ssbo
,
227 "Invalid mode for vulkan_resource_reindex");
229 nir_address_format addr_format
= vtn_mode_to_address_format(b
, mode
);
230 const struct glsl_type
*index_type
=
231 b
->options
->lower_ubo_ssbo_access_to_offsets
?
232 glsl_uint_type() : nir_address_format_to_glsl_type(addr_format
);
234 instr
->num_components
= glsl_get_vector_elements(index_type
);
235 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, instr
->num_components
,
236 glsl_get_bit_size(index_type
), NULL
);
237 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
239 return &instr
->dest
.ssa
;
243 vtn_descriptor_load(struct vtn_builder
*b
, enum vtn_variable_mode mode
,
244 nir_ssa_def
*desc_index
)
246 vtn_assert(b
->options
->environment
== NIR_SPIRV_VULKAN
);
248 nir_intrinsic_instr
*desc_load
=
249 nir_intrinsic_instr_create(b
->nb
.shader
,
250 nir_intrinsic_load_vulkan_descriptor
);
251 desc_load
->src
[0] = nir_src_for_ssa(desc_index
);
252 nir_intrinsic_set_desc_type(desc_load
, vk_desc_type_for_mode(b
, mode
));
254 vtn_fail_if(mode
!= vtn_variable_mode_ubo
&& mode
!= vtn_variable_mode_ssbo
,
255 "Invalid mode for load_vulkan_descriptor");
257 nir_address_format addr_format
= vtn_mode_to_address_format(b
, mode
);
258 const struct glsl_type
*ptr_type
=
259 nir_address_format_to_glsl_type(addr_format
);
261 desc_load
->num_components
= glsl_get_vector_elements(ptr_type
);
262 nir_ssa_dest_init(&desc_load
->instr
, &desc_load
->dest
,
263 desc_load
->num_components
,
264 glsl_get_bit_size(ptr_type
), NULL
);
265 nir_builder_instr_insert(&b
->nb
, &desc_load
->instr
);
267 return &desc_load
->dest
.ssa
;
270 /* Dereference the given base pointer by the access chain */
271 static struct vtn_pointer
*
272 vtn_nir_deref_pointer_dereference(struct vtn_builder
*b
,
273 struct vtn_pointer
*base
,
274 struct vtn_access_chain
*deref_chain
)
276 struct vtn_type
*type
= base
->type
;
277 enum gl_access_qualifier access
= base
->access
| deref_chain
->access
;
280 nir_deref_instr
*tail
;
283 } else if (b
->options
->environment
== NIR_SPIRV_VULKAN
&&
284 vtn_pointer_is_external_block(b
, base
)) {
285 nir_ssa_def
*block_index
= base
->block_index
;
287 /* We dereferencing an external block pointer. Correctness of this
288 * operation relies on one particular line in the SPIR-V spec, section
289 * entitled "Validation Rules for Shader Capabilities":
291 * "Block and BufferBlock decorations cannot decorate a structure
292 * type that is nested at any level inside another structure type
293 * decorated with Block or BufferBlock."
295 * This means that we can detect the point where we cross over from
296 * descriptor indexing to buffer indexing by looking for the block
297 * decorated struct type. Anything before the block decorated struct
298 * type is a descriptor indexing operation and anything after the block
299 * decorated struct is a buffer offset operation.
302 /* Figure out the descriptor array index if any
304 * Some of the Vulkan CTS tests with hand-rolled SPIR-V have been known
305 * to forget the Block or BufferBlock decoration from time to time.
306 * It's more robust if we check for both !block_index and for the type
307 * to contain a block. This way there's a decent chance that arrays of
308 * UBOs/SSBOs will work correctly even if variable pointers are
311 nir_ssa_def
*desc_arr_idx
= NULL
;
312 if (!block_index
|| vtn_type_contains_block(b
, type
)) {
313 /* If our type contains a block, then we're still outside the block
314 * and we need to process enough levels of dereferences to get inside
317 if (deref_chain
->ptr_as_array
) {
318 unsigned aoa_size
= glsl_get_aoa_size(type
->type
);
319 desc_arr_idx
= vtn_access_link_as_ssa(b
, deref_chain
->link
[idx
],
320 MAX2(aoa_size
, 1), 32);
324 for (; idx
< deref_chain
->length
; idx
++) {
325 if (type
->base_type
!= vtn_base_type_array
) {
326 vtn_assert(type
->base_type
== vtn_base_type_struct
);
330 unsigned aoa_size
= glsl_get_aoa_size(type
->array_element
->type
);
331 nir_ssa_def
*arr_offset
=
332 vtn_access_link_as_ssa(b
, deref_chain
->link
[idx
],
333 MAX2(aoa_size
, 1), 32);
335 desc_arr_idx
= nir_iadd(&b
->nb
, desc_arr_idx
, arr_offset
);
337 desc_arr_idx
= arr_offset
;
339 type
= type
->array_element
;
340 access
|= type
->access
;
345 vtn_assert(base
->var
&& base
->type
);
346 block_index
= vtn_variable_resource_index(b
, base
->var
, desc_arr_idx
);
347 } else if (desc_arr_idx
) {
348 block_index
= vtn_resource_reindex(b
, base
->mode
,
349 block_index
, desc_arr_idx
);
352 if (idx
== deref_chain
->length
) {
353 /* The entire deref was consumed in finding the block index. Return
354 * a pointer which just has a block index and a later access chain
355 * will dereference deeper.
357 struct vtn_pointer
*ptr
= rzalloc(b
, struct vtn_pointer
);
358 ptr
->mode
= base
->mode
;
360 ptr
->block_index
= block_index
;
361 ptr
->access
= access
;
365 /* If we got here, there's more access chain to handle and we have the
366 * final block index. Insert a descriptor load and cast to a deref to
367 * start the deref chain.
369 nir_ssa_def
*desc
= vtn_descriptor_load(b
, base
->mode
, block_index
);
371 assert(base
->mode
== vtn_variable_mode_ssbo
||
372 base
->mode
== vtn_variable_mode_ubo
);
373 nir_variable_mode nir_mode
=
374 base
->mode
== vtn_variable_mode_ssbo
? nir_var_mem_ssbo
: nir_var_mem_ubo
;
376 tail
= nir_build_deref_cast(&b
->nb
, desc
, nir_mode
,
377 vtn_type_get_nir_type(b
, type
, base
->mode
),
378 base
->ptr_type
->stride
);
380 assert(base
->var
&& base
->var
->var
);
381 tail
= nir_build_deref_var(&b
->nb
, base
->var
->var
);
382 if (base
->ptr_type
&& base
->ptr_type
->type
) {
383 tail
->dest
.ssa
.num_components
=
384 glsl_get_vector_elements(base
->ptr_type
->type
);
385 tail
->dest
.ssa
.bit_size
= glsl_get_bit_size(base
->ptr_type
->type
);
389 if (idx
== 0 && deref_chain
->ptr_as_array
) {
390 /* We start with a deref cast to get the stride. Hopefully, we'll be
391 * able to delete that cast eventually.
393 tail
= nir_build_deref_cast(&b
->nb
, &tail
->dest
.ssa
, tail
->mode
,
394 tail
->type
, base
->ptr_type
->stride
);
396 nir_ssa_def
*index
= vtn_access_link_as_ssa(b
, deref_chain
->link
[0], 1,
397 tail
->dest
.ssa
.bit_size
);
398 tail
= nir_build_deref_ptr_as_array(&b
->nb
, tail
, index
);
402 for (; idx
< deref_chain
->length
; idx
++) {
403 if (glsl_type_is_struct_or_ifc(type
->type
)) {
404 vtn_assert(deref_chain
->link
[idx
].mode
== vtn_access_mode_literal
);
405 unsigned field
= deref_chain
->link
[idx
].id
;
406 tail
= nir_build_deref_struct(&b
->nb
, tail
, field
);
407 type
= type
->members
[field
];
409 nir_ssa_def
*arr_index
=
410 vtn_access_link_as_ssa(b
, deref_chain
->link
[idx
], 1,
411 tail
->dest
.ssa
.bit_size
);
412 tail
= nir_build_deref_array(&b
->nb
, tail
, arr_index
);
413 type
= type
->array_element
;
416 access
|= type
->access
;
419 struct vtn_pointer
*ptr
= rzalloc(b
, struct vtn_pointer
);
420 ptr
->mode
= base
->mode
;
422 ptr
->var
= base
->var
;
424 ptr
->access
= access
;
429 static struct vtn_pointer
*
430 vtn_ssa_offset_pointer_dereference(struct vtn_builder
*b
,
431 struct vtn_pointer
*base
,
432 struct vtn_access_chain
*deref_chain
)
434 nir_ssa_def
*block_index
= base
->block_index
;
435 nir_ssa_def
*offset
= base
->offset
;
436 struct vtn_type
*type
= base
->type
;
437 enum gl_access_qualifier access
= base
->access
;
440 if (base
->mode
== vtn_variable_mode_ubo
||
441 base
->mode
== vtn_variable_mode_ssbo
) {
443 vtn_assert(base
->var
&& base
->type
);
444 nir_ssa_def
*desc_arr_idx
;
445 if (glsl_type_is_array(type
->type
)) {
446 if (deref_chain
->length
>= 1) {
448 vtn_access_link_as_ssa(b
, deref_chain
->link
[0], 1, 32);
450 /* This consumes a level of type */
451 type
= type
->array_element
;
452 access
|= type
->access
;
454 /* This is annoying. We've been asked for a pointer to the
455 * array of UBOs/SSBOs and not a specifc buffer. Return a
456 * pointer with a descriptor index of 0 and we'll have to do
457 * a reindex later to adjust it to the right thing.
459 desc_arr_idx
= nir_imm_int(&b
->nb
, 0);
461 } else if (deref_chain
->ptr_as_array
) {
462 /* You can't have a zero-length OpPtrAccessChain */
463 vtn_assert(deref_chain
->length
>= 1);
464 desc_arr_idx
= vtn_access_link_as_ssa(b
, deref_chain
->link
[0], 1, 32);
466 /* We have a regular non-array SSBO. */
469 block_index
= vtn_variable_resource_index(b
, base
->var
, desc_arr_idx
);
470 } else if (deref_chain
->ptr_as_array
&&
471 type
->base_type
== vtn_base_type_struct
&& type
->block
) {
472 /* We are doing an OpPtrAccessChain on a pointer to a struct that is
473 * decorated block. This is an interesting corner in the SPIR-V
474 * spec. One interpretation would be that they client is clearly
475 * trying to treat that block as if it's an implicit array of blocks
476 * repeated in the buffer. However, the SPIR-V spec for the
477 * OpPtrAccessChain says:
479 * "Base is treated as the address of the first element of an
480 * array, and the Element element’s address is computed to be the
481 * base for the Indexes, as per OpAccessChain."
483 * Taken literally, that would mean that your struct type is supposed
484 * to be treated as an array of such a struct and, since it's
485 * decorated block, that means an array of blocks which corresponds
486 * to an array descriptor. Therefore, we need to do a reindex
487 * operation to add the index from the first link in the access chain
488 * to the index we recieved.
490 * The downside to this interpretation (there always is one) is that
491 * this might be somewhat surprising behavior to apps if they expect
492 * the implicit array behavior described above.
494 vtn_assert(deref_chain
->length
>= 1);
495 nir_ssa_def
*offset_index
=
496 vtn_access_link_as_ssa(b
, deref_chain
->link
[0], 1, 32);
499 block_index
= vtn_resource_reindex(b
, base
->mode
,
500 block_index
, offset_index
);
505 if (base
->mode
== vtn_variable_mode_workgroup
) {
506 /* SLM doesn't need nor have a block index */
507 vtn_assert(!block_index
);
509 /* We need the variable for the base offset */
510 vtn_assert(base
->var
);
512 /* We need ptr_type for size and alignment */
513 vtn_assert(base
->ptr_type
);
515 /* Assign location on first use so that we don't end up bloating SLM
516 * address space for variables which are never statically used.
518 if (base
->var
->shared_location
< 0) {
519 vtn_assert(base
->ptr_type
->length
> 0 && base
->ptr_type
->align
> 0);
520 b
->shader
->num_shared
= vtn_align_u32(b
->shader
->num_shared
,
521 base
->ptr_type
->align
);
522 base
->var
->shared_location
= b
->shader
->num_shared
;
523 b
->shader
->num_shared
+= base
->ptr_type
->length
;
526 offset
= nir_imm_int(&b
->nb
, base
->var
->shared_location
);
527 } else if (base
->mode
== vtn_variable_mode_push_constant
) {
528 /* Push constants neither need nor have a block index */
529 vtn_assert(!block_index
);
531 /* Start off with at the start of the push constant block. */
532 offset
= nir_imm_int(&b
->nb
, 0);
534 /* The code above should have ensured a block_index when needed. */
535 vtn_assert(block_index
);
537 /* Start off with at the start of the buffer. */
538 offset
= nir_imm_int(&b
->nb
, 0);
542 if (deref_chain
->ptr_as_array
&& idx
== 0) {
543 /* We need ptr_type for the stride */
544 vtn_assert(base
->ptr_type
);
546 /* We need at least one element in the chain */
547 vtn_assert(deref_chain
->length
>= 1);
549 nir_ssa_def
*elem_offset
=
550 vtn_access_link_as_ssa(b
, deref_chain
->link
[idx
],
551 base
->ptr_type
->stride
, offset
->bit_size
);
552 offset
= nir_iadd(&b
->nb
, offset
, elem_offset
);
556 for (; idx
< deref_chain
->length
; idx
++) {
557 switch (glsl_get_base_type(type
->type
)) {
560 case GLSL_TYPE_UINT16
:
561 case GLSL_TYPE_INT16
:
562 case GLSL_TYPE_UINT8
:
564 case GLSL_TYPE_UINT64
:
565 case GLSL_TYPE_INT64
:
566 case GLSL_TYPE_FLOAT
:
567 case GLSL_TYPE_FLOAT16
:
568 case GLSL_TYPE_DOUBLE
:
570 case GLSL_TYPE_ARRAY
: {
571 nir_ssa_def
*elem_offset
=
572 vtn_access_link_as_ssa(b
, deref_chain
->link
[idx
],
573 type
->stride
, offset
->bit_size
);
574 offset
= nir_iadd(&b
->nb
, offset
, elem_offset
);
575 type
= type
->array_element
;
576 access
|= type
->access
;
580 case GLSL_TYPE_INTERFACE
:
581 case GLSL_TYPE_STRUCT
: {
582 vtn_assert(deref_chain
->link
[idx
].mode
== vtn_access_mode_literal
);
583 unsigned member
= deref_chain
->link
[idx
].id
;
584 offset
= nir_iadd_imm(&b
->nb
, offset
, type
->offsets
[member
]);
585 type
= type
->members
[member
];
586 access
|= type
->access
;
591 vtn_fail("Invalid type for deref");
595 struct vtn_pointer
*ptr
= rzalloc(b
, struct vtn_pointer
);
596 ptr
->mode
= base
->mode
;
598 ptr
->block_index
= block_index
;
599 ptr
->offset
= offset
;
600 ptr
->access
= access
;
605 /* Dereference the given base pointer by the access chain */
606 static struct vtn_pointer
*
607 vtn_pointer_dereference(struct vtn_builder
*b
,
608 struct vtn_pointer
*base
,
609 struct vtn_access_chain
*deref_chain
)
611 if (vtn_pointer_uses_ssa_offset(b
, base
)) {
612 return vtn_ssa_offset_pointer_dereference(b
, base
, deref_chain
);
614 return vtn_nir_deref_pointer_dereference(b
, base
, deref_chain
);
619 vtn_pointer_to_deref(struct vtn_builder
*b
, struct vtn_pointer
*ptr
)
621 vtn_assert(!vtn_pointer_uses_ssa_offset(b
, ptr
));
623 struct vtn_access_chain chain
= {
626 ptr
= vtn_nir_deref_pointer_dereference(b
, ptr
, &chain
);
633 _vtn_local_load_store(struct vtn_builder
*b
, bool load
, nir_deref_instr
*deref
,
634 struct vtn_ssa_value
*inout
,
635 enum gl_access_qualifier access
)
637 if (glsl_type_is_vector_or_scalar(deref
->type
)) {
639 inout
->def
= nir_load_deref_with_access(&b
->nb
, deref
, access
);
641 nir_store_deref_with_access(&b
->nb
, deref
, inout
->def
, ~0, access
);
643 } else if (glsl_type_is_array(deref
->type
) ||
644 glsl_type_is_matrix(deref
->type
)) {
645 unsigned elems
= glsl_get_length(deref
->type
);
646 for (unsigned i
= 0; i
< elems
; i
++) {
647 nir_deref_instr
*child
=
648 nir_build_deref_array_imm(&b
->nb
, deref
, i
);
649 _vtn_local_load_store(b
, load
, child
, inout
->elems
[i
], access
);
652 vtn_assert(glsl_type_is_struct_or_ifc(deref
->type
));
653 unsigned elems
= glsl_get_length(deref
->type
);
654 for (unsigned i
= 0; i
< elems
; i
++) {
655 nir_deref_instr
*child
= nir_build_deref_struct(&b
->nb
, deref
, i
);
656 _vtn_local_load_store(b
, load
, child
, inout
->elems
[i
], access
);
662 vtn_nir_deref(struct vtn_builder
*b
, uint32_t id
)
664 struct vtn_pointer
*ptr
= vtn_value(b
, id
, vtn_value_type_pointer
)->pointer
;
665 return vtn_pointer_to_deref(b
, ptr
);
669 * Gets the NIR-level deref tail, which may have as a child an array deref
670 * selecting which component due to OpAccessChain supporting per-component
671 * indexing in SPIR-V.
673 static nir_deref_instr
*
674 get_deref_tail(nir_deref_instr
*deref
)
676 if (deref
->deref_type
!= nir_deref_type_array
)
679 nir_deref_instr
*parent
=
680 nir_instr_as_deref(deref
->parent
.ssa
->parent_instr
);
682 if (glsl_type_is_vector(parent
->type
))
688 struct vtn_ssa_value
*
689 vtn_local_load(struct vtn_builder
*b
, nir_deref_instr
*src
,
690 enum gl_access_qualifier access
)
692 nir_deref_instr
*src_tail
= get_deref_tail(src
);
693 struct vtn_ssa_value
*val
= vtn_create_ssa_value(b
, src_tail
->type
);
694 _vtn_local_load_store(b
, true, src_tail
, val
, access
);
696 if (src_tail
!= src
) {
697 val
->type
= src
->type
;
698 val
->def
= nir_vector_extract(&b
->nb
, val
->def
, src
->arr
.index
.ssa
);
705 vtn_local_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
706 nir_deref_instr
*dest
, enum gl_access_qualifier access
)
708 nir_deref_instr
*dest_tail
= get_deref_tail(dest
);
710 if (dest_tail
!= dest
) {
711 struct vtn_ssa_value
*val
= vtn_create_ssa_value(b
, dest_tail
->type
);
712 _vtn_local_load_store(b
, true, dest_tail
, val
, access
);
714 val
->def
= nir_vector_insert(&b
->nb
, val
->def
, src
->def
,
715 dest
->arr
.index
.ssa
);
716 _vtn_local_load_store(b
, false, dest_tail
, val
, access
);
718 _vtn_local_load_store(b
, false, dest_tail
, src
, access
);
723 vtn_pointer_to_offset(struct vtn_builder
*b
, struct vtn_pointer
*ptr
,
724 nir_ssa_def
**index_out
)
726 assert(vtn_pointer_uses_ssa_offset(b
, ptr
));
728 struct vtn_access_chain chain
= {
731 ptr
= vtn_ssa_offset_pointer_dereference(b
, ptr
, &chain
);
733 *index_out
= ptr
->block_index
;
737 /* Tries to compute the size of an interface block based on the strides and
738 * offsets that are provided to us in the SPIR-V source.
741 vtn_type_block_size(struct vtn_builder
*b
, struct vtn_type
*type
)
743 enum glsl_base_type base_type
= glsl_get_base_type(type
->type
);
747 case GLSL_TYPE_UINT16
:
748 case GLSL_TYPE_INT16
:
749 case GLSL_TYPE_UINT8
:
751 case GLSL_TYPE_UINT64
:
752 case GLSL_TYPE_INT64
:
753 case GLSL_TYPE_FLOAT
:
754 case GLSL_TYPE_FLOAT16
:
756 case GLSL_TYPE_DOUBLE
: {
757 unsigned cols
= type
->row_major
? glsl_get_vector_elements(type
->type
) :
758 glsl_get_matrix_columns(type
->type
);
760 vtn_assert(type
->stride
> 0);
761 return type
->stride
* cols
;
763 unsigned type_size
= glsl_get_bit_size(type
->type
) / 8;
764 return glsl_get_vector_elements(type
->type
) * type_size
;
768 case GLSL_TYPE_STRUCT
:
769 case GLSL_TYPE_INTERFACE
: {
771 unsigned num_fields
= glsl_get_length(type
->type
);
772 for (unsigned f
= 0; f
< num_fields
; f
++) {
773 unsigned field_end
= type
->offsets
[f
] +
774 vtn_type_block_size(b
, type
->members
[f
]);
775 size
= MAX2(size
, field_end
);
780 case GLSL_TYPE_ARRAY
:
781 vtn_assert(type
->stride
> 0);
782 vtn_assert(glsl_get_length(type
->type
) > 0);
783 return type
->stride
* glsl_get_length(type
->type
);
786 vtn_fail("Invalid block type");
792 _vtn_load_store_tail(struct vtn_builder
*b
, nir_intrinsic_op op
, bool load
,
793 nir_ssa_def
*index
, nir_ssa_def
*offset
,
794 unsigned access_offset
, unsigned access_size
,
795 struct vtn_ssa_value
**inout
, const struct glsl_type
*type
,
796 enum gl_access_qualifier access
)
798 nir_intrinsic_instr
*instr
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
799 instr
->num_components
= glsl_get_vector_elements(type
);
801 /* Booleans usually shouldn't show up in external memory in SPIR-V.
802 * However, they do for certain older GLSLang versions and can for shared
803 * memory when we lower access chains internally.
805 const unsigned data_bit_size
= glsl_type_is_boolean(type
) ? 32 :
806 glsl_get_bit_size(type
);
810 nir_intrinsic_set_write_mask(instr
, (1 << instr
->num_components
) - 1);
811 instr
->src
[src
++] = nir_src_for_ssa((*inout
)->def
);
814 if (op
== nir_intrinsic_load_push_constant
) {
815 nir_intrinsic_set_base(instr
, access_offset
);
816 nir_intrinsic_set_range(instr
, access_size
);
819 if (op
== nir_intrinsic_load_ubo
||
820 op
== nir_intrinsic_load_ssbo
||
821 op
== nir_intrinsic_store_ssbo
) {
822 nir_intrinsic_set_access(instr
, access
);
825 /* With extensions like relaxed_block_layout, we really can't guarantee
826 * much more than scalar alignment.
828 if (op
!= nir_intrinsic_load_push_constant
)
829 nir_intrinsic_set_align(instr
, data_bit_size
/ 8, 0);
832 instr
->src
[src
++] = nir_src_for_ssa(index
);
834 if (op
== nir_intrinsic_load_push_constant
) {
835 /* We need to subtract the offset from where the intrinsic will load the
838 nir_src_for_ssa(nir_isub(&b
->nb
, offset
,
839 nir_imm_int(&b
->nb
, access_offset
)));
841 instr
->src
[src
++] = nir_src_for_ssa(offset
);
845 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
846 instr
->num_components
, data_bit_size
, NULL
);
847 (*inout
)->def
= &instr
->dest
.ssa
;
850 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
852 if (load
&& glsl_get_base_type(type
) == GLSL_TYPE_BOOL
)
853 (*inout
)->def
= nir_ine(&b
->nb
, (*inout
)->def
, nir_imm_int(&b
->nb
, 0));
857 _vtn_block_load_store(struct vtn_builder
*b
, nir_intrinsic_op op
, bool load
,
858 nir_ssa_def
*index
, nir_ssa_def
*offset
,
859 unsigned access_offset
, unsigned access_size
,
860 struct vtn_type
*type
, enum gl_access_qualifier access
,
861 struct vtn_ssa_value
**inout
)
863 enum glsl_base_type base_type
= glsl_get_base_type(type
->type
);
867 case GLSL_TYPE_UINT16
:
868 case GLSL_TYPE_INT16
:
869 case GLSL_TYPE_UINT8
:
871 case GLSL_TYPE_UINT64
:
872 case GLSL_TYPE_INT64
:
873 case GLSL_TYPE_FLOAT
:
874 case GLSL_TYPE_FLOAT16
:
875 case GLSL_TYPE_DOUBLE
:
877 /* This is where things get interesting. At this point, we've hit
878 * a vector, a scalar, or a matrix.
880 if (glsl_type_is_matrix(type
->type
)) {
881 /* Loading the whole matrix */
882 struct vtn_ssa_value
*transpose
;
883 unsigned num_ops
, vec_width
, col_stride
;
884 if (type
->row_major
) {
885 num_ops
= glsl_get_vector_elements(type
->type
);
886 vec_width
= glsl_get_matrix_columns(type
->type
);
887 col_stride
= type
->array_element
->stride
;
889 const struct glsl_type
*transpose_type
=
890 glsl_matrix_type(base_type
, vec_width
, num_ops
);
891 *inout
= vtn_create_ssa_value(b
, transpose_type
);
893 transpose
= vtn_ssa_transpose(b
, *inout
);
897 num_ops
= glsl_get_matrix_columns(type
->type
);
898 vec_width
= glsl_get_vector_elements(type
->type
);
899 col_stride
= type
->stride
;
902 for (unsigned i
= 0; i
< num_ops
; i
++) {
903 nir_ssa_def
*elem_offset
=
904 nir_iadd_imm(&b
->nb
, offset
, i
* col_stride
);
905 _vtn_load_store_tail(b
, op
, load
, index
, elem_offset
,
906 access_offset
, access_size
,
908 glsl_vector_type(base_type
, vec_width
),
909 type
->access
| access
);
912 if (load
&& type
->row_major
)
913 *inout
= vtn_ssa_transpose(b
, *inout
);
915 unsigned elems
= glsl_get_vector_elements(type
->type
);
916 unsigned type_size
= glsl_get_bit_size(type
->type
) / 8;
917 if (elems
== 1 || type
->stride
== type_size
) {
918 /* This is a tightly-packed normal scalar or vector load */
919 vtn_assert(glsl_type_is_vector_or_scalar(type
->type
));
920 _vtn_load_store_tail(b
, op
, load
, index
, offset
,
921 access_offset
, access_size
,
923 type
->access
| access
);
925 /* This is a strided load. We have to load N things separately.
926 * This is the single column of a row-major matrix case.
928 vtn_assert(type
->stride
> type_size
);
929 vtn_assert(type
->stride
% type_size
== 0);
931 nir_ssa_def
*per_comp
[4];
932 for (unsigned i
= 0; i
< elems
; i
++) {
933 nir_ssa_def
*elem_offset
=
934 nir_iadd_imm(&b
->nb
, offset
, i
* type
->stride
);
935 struct vtn_ssa_value
*comp
, temp_val
;
937 temp_val
.def
= nir_channel(&b
->nb
, (*inout
)->def
, i
);
938 temp_val
.type
= glsl_scalar_type(base_type
);
941 _vtn_load_store_tail(b
, op
, load
, index
, elem_offset
,
942 access_offset
, access_size
,
943 &comp
, glsl_scalar_type(base_type
),
944 type
->access
| access
);
945 per_comp
[i
] = comp
->def
;
950 *inout
= vtn_create_ssa_value(b
, type
->type
);
951 (*inout
)->def
= nir_vec(&b
->nb
, per_comp
, elems
);
957 case GLSL_TYPE_ARRAY
: {
958 unsigned elems
= glsl_get_length(type
->type
);
959 for (unsigned i
= 0; i
< elems
; i
++) {
960 nir_ssa_def
*elem_off
=
961 nir_iadd_imm(&b
->nb
, offset
, i
* type
->stride
);
962 _vtn_block_load_store(b
, op
, load
, index
, elem_off
,
963 access_offset
, access_size
,
965 type
->array_element
->access
| access
,
966 &(*inout
)->elems
[i
]);
971 case GLSL_TYPE_INTERFACE
:
972 case GLSL_TYPE_STRUCT
: {
973 unsigned elems
= glsl_get_length(type
->type
);
974 for (unsigned i
= 0; i
< elems
; i
++) {
975 nir_ssa_def
*elem_off
=
976 nir_iadd_imm(&b
->nb
, offset
, type
->offsets
[i
]);
977 _vtn_block_load_store(b
, op
, load
, index
, elem_off
,
978 access_offset
, access_size
,
980 type
->members
[i
]->access
| access
,
981 &(*inout
)->elems
[i
]);
987 vtn_fail("Invalid block member type");
991 static struct vtn_ssa_value
*
992 vtn_block_load(struct vtn_builder
*b
, struct vtn_pointer
*src
)
995 unsigned access_offset
= 0, access_size
= 0;
997 case vtn_variable_mode_ubo
:
998 op
= nir_intrinsic_load_ubo
;
1000 case vtn_variable_mode_ssbo
:
1001 op
= nir_intrinsic_load_ssbo
;
1003 case vtn_variable_mode_push_constant
:
1004 op
= nir_intrinsic_load_push_constant
;
1005 access_size
= b
->shader
->num_uniforms
;
1007 case vtn_variable_mode_workgroup
:
1008 op
= nir_intrinsic_load_shared
;
1011 vtn_fail("Invalid block variable mode");
1014 nir_ssa_def
*offset
, *index
= NULL
;
1015 offset
= vtn_pointer_to_offset(b
, src
, &index
);
1017 struct vtn_ssa_value
*value
= vtn_create_ssa_value(b
, src
->type
->type
);
1018 _vtn_block_load_store(b
, op
, true, index
, offset
,
1019 access_offset
, access_size
,
1020 src
->type
, src
->access
, &value
);
1025 vtn_block_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
1026 struct vtn_pointer
*dst
)
1028 nir_intrinsic_op op
;
1029 switch (dst
->mode
) {
1030 case vtn_variable_mode_ssbo
:
1031 op
= nir_intrinsic_store_ssbo
;
1033 case vtn_variable_mode_workgroup
:
1034 op
= nir_intrinsic_store_shared
;
1037 vtn_fail("Invalid block variable mode");
1040 nir_ssa_def
*offset
, *index
= NULL
;
1041 offset
= vtn_pointer_to_offset(b
, dst
, &index
);
1043 _vtn_block_load_store(b
, op
, false, index
, offset
,
1044 0, 0, dst
->type
, dst
->access
, &src
);
1048 _vtn_variable_load_store(struct vtn_builder
*b
, bool load
,
1049 struct vtn_pointer
*ptr
,
1050 enum gl_access_qualifier access
,
1051 struct vtn_ssa_value
**inout
)
1053 if (ptr
->mode
== vtn_variable_mode_uniform
) {
1054 if (ptr
->type
->base_type
== vtn_base_type_image
||
1055 ptr
->type
->base_type
== vtn_base_type_sampler
) {
1056 /* See also our handling of OpTypeSampler and OpTypeImage */
1058 (*inout
)->def
= vtn_pointer_to_ssa(b
, ptr
);
1060 } else if (ptr
->type
->base_type
== vtn_base_type_sampled_image
) {
1061 /* See also our handling of OpTypeSampledImage */
1063 struct vtn_sampled_image si
= {
1064 .image
= vtn_pointer_to_deref(b
, ptr
),
1065 .sampler
= vtn_pointer_to_deref(b
, ptr
),
1067 (*inout
)->def
= vtn_sampled_image_to_nir_ssa(b
, si
);
1072 enum glsl_base_type base_type
= glsl_get_base_type(ptr
->type
->type
);
1073 switch (base_type
) {
1074 case GLSL_TYPE_UINT
:
1076 case GLSL_TYPE_UINT16
:
1077 case GLSL_TYPE_INT16
:
1078 case GLSL_TYPE_UINT8
:
1079 case GLSL_TYPE_INT8
:
1080 case GLSL_TYPE_UINT64
:
1081 case GLSL_TYPE_INT64
:
1082 case GLSL_TYPE_FLOAT
:
1083 case GLSL_TYPE_FLOAT16
:
1084 case GLSL_TYPE_BOOL
:
1085 case GLSL_TYPE_DOUBLE
:
1086 if (glsl_type_is_vector_or_scalar(ptr
->type
->type
)) {
1087 /* We hit a vector or scalar; go ahead and emit the load[s] */
1088 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
1089 if (vtn_mode_is_cross_invocation(b
, ptr
->mode
)) {
1090 /* If it's cross-invocation, we call nir_load/store_deref
1091 * directly. The vtn_local_load/store helpers are too clever and
1092 * do magic to avoid array derefs of vectors. That magic is both
1093 * less efficient than the direct load/store and, in the case of
1094 * stores, is broken because it creates a race condition if two
1095 * threads are writing to different components of the same vector
1096 * due to the load+insert+store it uses to emulate the array
1100 (*inout
)->def
= nir_load_deref_with_access(&b
->nb
, deref
,
1101 ptr
->type
->access
| access
);
1103 nir_store_deref_with_access(&b
->nb
, deref
, (*inout
)->def
, ~0,
1104 ptr
->type
->access
| access
);
1108 *inout
= vtn_local_load(b
, deref
, ptr
->type
->access
| access
);
1110 vtn_local_store(b
, *inout
, deref
, ptr
->type
->access
| access
);
1117 case GLSL_TYPE_INTERFACE
:
1118 case GLSL_TYPE_ARRAY
:
1119 case GLSL_TYPE_STRUCT
: {
1120 unsigned elems
= glsl_get_length(ptr
->type
->type
);
1121 struct vtn_access_chain chain
= {
1124 { .mode
= vtn_access_mode_literal
, },
1127 for (unsigned i
= 0; i
< elems
; i
++) {
1128 chain
.link
[0].id
= i
;
1129 struct vtn_pointer
*elem
= vtn_pointer_dereference(b
, ptr
, &chain
);
1130 _vtn_variable_load_store(b
, load
, elem
, ptr
->type
->access
| access
,
1131 &(*inout
)->elems
[i
]);
1137 vtn_fail("Invalid access chain type");
1141 struct vtn_ssa_value
*
1142 vtn_variable_load(struct vtn_builder
*b
, struct vtn_pointer
*src
)
1144 if (vtn_pointer_uses_ssa_offset(b
, src
)) {
1145 return vtn_block_load(b
, src
);
1147 struct vtn_ssa_value
*val
= vtn_create_ssa_value(b
, src
->type
->type
);
1148 _vtn_variable_load_store(b
, true, src
, src
->access
, &val
);
1154 vtn_variable_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
1155 struct vtn_pointer
*dest
)
1157 if (vtn_pointer_uses_ssa_offset(b
, dest
)) {
1158 vtn_assert(dest
->mode
== vtn_variable_mode_ssbo
||
1159 dest
->mode
== vtn_variable_mode_workgroup
);
1160 vtn_block_store(b
, src
, dest
);
1162 _vtn_variable_load_store(b
, false, dest
, dest
->access
, &src
);
1167 _vtn_variable_copy(struct vtn_builder
*b
, struct vtn_pointer
*dest
,
1168 struct vtn_pointer
*src
)
1170 vtn_assert(glsl_get_bare_type(src
->type
->type
) ==
1171 glsl_get_bare_type(dest
->type
->type
));
1172 enum glsl_base_type base_type
= glsl_get_base_type(src
->type
->type
);
1173 switch (base_type
) {
1174 case GLSL_TYPE_UINT
:
1176 case GLSL_TYPE_UINT16
:
1177 case GLSL_TYPE_INT16
:
1178 case GLSL_TYPE_UINT8
:
1179 case GLSL_TYPE_INT8
:
1180 case GLSL_TYPE_UINT64
:
1181 case GLSL_TYPE_INT64
:
1182 case GLSL_TYPE_FLOAT
:
1183 case GLSL_TYPE_FLOAT16
:
1184 case GLSL_TYPE_DOUBLE
:
1185 case GLSL_TYPE_BOOL
:
1186 /* At this point, we have a scalar, vector, or matrix so we know that
1187 * there cannot be any structure splitting still in the way. By
1188 * stopping at the matrix level rather than the vector level, we
1189 * ensure that matrices get loaded in the optimal way even if they
1190 * are storred row-major in a UBO.
1192 vtn_variable_store(b
, vtn_variable_load(b
, src
), dest
);
1195 case GLSL_TYPE_INTERFACE
:
1196 case GLSL_TYPE_ARRAY
:
1197 case GLSL_TYPE_STRUCT
: {
1198 struct vtn_access_chain chain
= {
1201 { .mode
= vtn_access_mode_literal
, },
1204 unsigned elems
= glsl_get_length(src
->type
->type
);
1205 for (unsigned i
= 0; i
< elems
; i
++) {
1206 chain
.link
[0].id
= i
;
1207 struct vtn_pointer
*src_elem
=
1208 vtn_pointer_dereference(b
, src
, &chain
);
1209 struct vtn_pointer
*dest_elem
=
1210 vtn_pointer_dereference(b
, dest
, &chain
);
1212 _vtn_variable_copy(b
, dest_elem
, src_elem
);
1218 vtn_fail("Invalid access chain type");
1223 vtn_variable_copy(struct vtn_builder
*b
, struct vtn_pointer
*dest
,
1224 struct vtn_pointer
*src
)
1226 /* TODO: At some point, we should add a special-case for when we can
1227 * just emit a copy_var intrinsic.
1229 _vtn_variable_copy(b
, dest
, src
);
1233 set_mode_system_value(struct vtn_builder
*b
, nir_variable_mode
*mode
)
1235 vtn_assert(*mode
== nir_var_system_value
|| *mode
== nir_var_shader_in
);
1236 *mode
= nir_var_system_value
;
1240 vtn_get_builtin_location(struct vtn_builder
*b
,
1241 SpvBuiltIn builtin
, int *location
,
1242 nir_variable_mode
*mode
)
1245 case SpvBuiltInPosition
:
1246 *location
= VARYING_SLOT_POS
;
1248 case SpvBuiltInPointSize
:
1249 *location
= VARYING_SLOT_PSIZ
;
1251 case SpvBuiltInClipDistance
:
1252 *location
= VARYING_SLOT_CLIP_DIST0
; /* XXX CLIP_DIST1? */
1254 case SpvBuiltInCullDistance
:
1255 *location
= VARYING_SLOT_CULL_DIST0
;
1257 case SpvBuiltInVertexId
:
1258 case SpvBuiltInVertexIndex
:
1259 /* The Vulkan spec defines VertexIndex to be non-zero-based and doesn't
1260 * allow VertexId. The ARB_gl_spirv spec defines VertexId to be the
1261 * same as gl_VertexID, which is non-zero-based, and removes
1262 * VertexIndex. Since they're both defined to be non-zero-based, we use
1263 * SYSTEM_VALUE_VERTEX_ID for both.
1265 *location
= SYSTEM_VALUE_VERTEX_ID
;
1266 set_mode_system_value(b
, mode
);
1268 case SpvBuiltInInstanceIndex
:
1269 *location
= SYSTEM_VALUE_INSTANCE_INDEX
;
1270 set_mode_system_value(b
, mode
);
1272 case SpvBuiltInInstanceId
:
1273 *location
= SYSTEM_VALUE_INSTANCE_ID
;
1274 set_mode_system_value(b
, mode
);
1276 case SpvBuiltInPrimitiveId
:
1277 if (b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
) {
1278 vtn_assert(*mode
== nir_var_shader_in
);
1279 *location
= VARYING_SLOT_PRIMITIVE_ID
;
1280 } else if (*mode
== nir_var_shader_out
) {
1281 *location
= VARYING_SLOT_PRIMITIVE_ID
;
1283 *location
= SYSTEM_VALUE_PRIMITIVE_ID
;
1284 set_mode_system_value(b
, mode
);
1287 case SpvBuiltInInvocationId
:
1288 *location
= SYSTEM_VALUE_INVOCATION_ID
;
1289 set_mode_system_value(b
, mode
);
1291 case SpvBuiltInLayer
:
1292 *location
= VARYING_SLOT_LAYER
;
1293 if (b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
)
1294 *mode
= nir_var_shader_in
;
1295 else if (b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
)
1296 *mode
= nir_var_shader_out
;
1297 else if (b
->options
&& b
->options
->caps
.shader_viewport_index_layer
&&
1298 (b
->shader
->info
.stage
== MESA_SHADER_VERTEX
||
1299 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
))
1300 *mode
= nir_var_shader_out
;
1302 vtn_fail("invalid stage for SpvBuiltInLayer");
1304 case SpvBuiltInViewportIndex
:
1305 *location
= VARYING_SLOT_VIEWPORT
;
1306 if (b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
)
1307 *mode
= nir_var_shader_out
;
1308 else if (b
->options
&& b
->options
->caps
.shader_viewport_index_layer
&&
1309 (b
->shader
->info
.stage
== MESA_SHADER_VERTEX
||
1310 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
))
1311 *mode
= nir_var_shader_out
;
1312 else if (b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
)
1313 *mode
= nir_var_shader_in
;
1315 vtn_fail("invalid stage for SpvBuiltInViewportIndex");
1317 case SpvBuiltInTessLevelOuter
:
1318 *location
= VARYING_SLOT_TESS_LEVEL_OUTER
;
1320 case SpvBuiltInTessLevelInner
:
1321 *location
= VARYING_SLOT_TESS_LEVEL_INNER
;
1323 case SpvBuiltInTessCoord
:
1324 *location
= SYSTEM_VALUE_TESS_COORD
;
1325 set_mode_system_value(b
, mode
);
1327 case SpvBuiltInPatchVertices
:
1328 *location
= SYSTEM_VALUE_VERTICES_IN
;
1329 set_mode_system_value(b
, mode
);
1331 case SpvBuiltInFragCoord
:
1332 vtn_assert(*mode
== nir_var_shader_in
);
1333 if (b
->options
&& b
->options
->frag_coord_is_sysval
) {
1334 *mode
= nir_var_system_value
;
1335 *location
= SYSTEM_VALUE_FRAG_COORD
;
1337 *location
= VARYING_SLOT_POS
;
1340 case SpvBuiltInPointCoord
:
1341 *location
= VARYING_SLOT_PNTC
;
1342 vtn_assert(*mode
== nir_var_shader_in
);
1344 case SpvBuiltInFrontFacing
:
1345 *location
= SYSTEM_VALUE_FRONT_FACE
;
1346 set_mode_system_value(b
, mode
);
1348 case SpvBuiltInSampleId
:
1349 *location
= SYSTEM_VALUE_SAMPLE_ID
;
1350 set_mode_system_value(b
, mode
);
1352 case SpvBuiltInSamplePosition
:
1353 *location
= SYSTEM_VALUE_SAMPLE_POS
;
1354 set_mode_system_value(b
, mode
);
1356 case SpvBuiltInSampleMask
:
1357 if (*mode
== nir_var_shader_out
) {
1358 *location
= FRAG_RESULT_SAMPLE_MASK
;
1360 *location
= SYSTEM_VALUE_SAMPLE_MASK_IN
;
1361 set_mode_system_value(b
, mode
);
1364 case SpvBuiltInFragDepth
:
1365 *location
= FRAG_RESULT_DEPTH
;
1366 vtn_assert(*mode
== nir_var_shader_out
);
1368 case SpvBuiltInHelperInvocation
:
1369 *location
= SYSTEM_VALUE_HELPER_INVOCATION
;
1370 set_mode_system_value(b
, mode
);
1372 case SpvBuiltInNumWorkgroups
:
1373 *location
= SYSTEM_VALUE_NUM_WORK_GROUPS
;
1374 set_mode_system_value(b
, mode
);
1376 case SpvBuiltInWorkgroupSize
:
1377 *location
= SYSTEM_VALUE_LOCAL_GROUP_SIZE
;
1378 set_mode_system_value(b
, mode
);
1380 case SpvBuiltInWorkgroupId
:
1381 *location
= SYSTEM_VALUE_WORK_GROUP_ID
;
1382 set_mode_system_value(b
, mode
);
1384 case SpvBuiltInLocalInvocationId
:
1385 *location
= SYSTEM_VALUE_LOCAL_INVOCATION_ID
;
1386 set_mode_system_value(b
, mode
);
1388 case SpvBuiltInLocalInvocationIndex
:
1389 *location
= SYSTEM_VALUE_LOCAL_INVOCATION_INDEX
;
1390 set_mode_system_value(b
, mode
);
1392 case SpvBuiltInGlobalInvocationId
:
1393 *location
= SYSTEM_VALUE_GLOBAL_INVOCATION_ID
;
1394 set_mode_system_value(b
, mode
);
1396 case SpvBuiltInGlobalLinearId
:
1397 *location
= SYSTEM_VALUE_GLOBAL_INVOCATION_INDEX
;
1398 set_mode_system_value(b
, mode
);
1400 case SpvBuiltInBaseVertex
:
1401 /* OpenGL gl_BaseVertex (SYSTEM_VALUE_BASE_VERTEX) is not the same
1402 * semantic as Vulkan BaseVertex (SYSTEM_VALUE_FIRST_VERTEX).
1404 if (b
->options
->environment
== NIR_SPIRV_OPENGL
)
1405 *location
= SYSTEM_VALUE_BASE_VERTEX
;
1407 *location
= SYSTEM_VALUE_FIRST_VERTEX
;
1408 set_mode_system_value(b
, mode
);
1410 case SpvBuiltInBaseInstance
:
1411 *location
= SYSTEM_VALUE_BASE_INSTANCE
;
1412 set_mode_system_value(b
, mode
);
1414 case SpvBuiltInDrawIndex
:
1415 *location
= SYSTEM_VALUE_DRAW_ID
;
1416 set_mode_system_value(b
, mode
);
1418 case SpvBuiltInSubgroupSize
:
1419 *location
= SYSTEM_VALUE_SUBGROUP_SIZE
;
1420 set_mode_system_value(b
, mode
);
1422 case SpvBuiltInSubgroupId
:
1423 *location
= SYSTEM_VALUE_SUBGROUP_ID
;
1424 set_mode_system_value(b
, mode
);
1426 case SpvBuiltInSubgroupLocalInvocationId
:
1427 *location
= SYSTEM_VALUE_SUBGROUP_INVOCATION
;
1428 set_mode_system_value(b
, mode
);
1430 case SpvBuiltInNumSubgroups
:
1431 *location
= SYSTEM_VALUE_NUM_SUBGROUPS
;
1432 set_mode_system_value(b
, mode
);
1434 case SpvBuiltInDeviceIndex
:
1435 *location
= SYSTEM_VALUE_DEVICE_INDEX
;
1436 set_mode_system_value(b
, mode
);
1438 case SpvBuiltInViewIndex
:
1439 if (b
->options
&& b
->options
->view_index_is_input
) {
1440 *location
= VARYING_SLOT_VIEW_INDEX
;
1441 vtn_assert(*mode
== nir_var_shader_in
);
1443 *location
= SYSTEM_VALUE_VIEW_INDEX
;
1444 set_mode_system_value(b
, mode
);
1447 case SpvBuiltInSubgroupEqMask
:
1448 *location
= SYSTEM_VALUE_SUBGROUP_EQ_MASK
,
1449 set_mode_system_value(b
, mode
);
1451 case SpvBuiltInSubgroupGeMask
:
1452 *location
= SYSTEM_VALUE_SUBGROUP_GE_MASK
,
1453 set_mode_system_value(b
, mode
);
1455 case SpvBuiltInSubgroupGtMask
:
1456 *location
= SYSTEM_VALUE_SUBGROUP_GT_MASK
,
1457 set_mode_system_value(b
, mode
);
1459 case SpvBuiltInSubgroupLeMask
:
1460 *location
= SYSTEM_VALUE_SUBGROUP_LE_MASK
,
1461 set_mode_system_value(b
, mode
);
1463 case SpvBuiltInSubgroupLtMask
:
1464 *location
= SYSTEM_VALUE_SUBGROUP_LT_MASK
,
1465 set_mode_system_value(b
, mode
);
1467 case SpvBuiltInFragStencilRefEXT
:
1468 *location
= FRAG_RESULT_STENCIL
;
1469 vtn_assert(*mode
== nir_var_shader_out
);
1471 case SpvBuiltInWorkDim
:
1472 *location
= SYSTEM_VALUE_WORK_DIM
;
1473 set_mode_system_value(b
, mode
);
1475 case SpvBuiltInGlobalSize
:
1476 *location
= SYSTEM_VALUE_GLOBAL_GROUP_SIZE
;
1477 set_mode_system_value(b
, mode
);
1479 case SpvBuiltInBaryCoordNoPerspAMD
:
1480 *location
= SYSTEM_VALUE_BARYCENTRIC_LINEAR_PIXEL
;
1481 set_mode_system_value(b
, mode
);
1483 case SpvBuiltInBaryCoordNoPerspCentroidAMD
:
1484 *location
= SYSTEM_VALUE_BARYCENTRIC_LINEAR_CENTROID
;
1485 set_mode_system_value(b
, mode
);
1487 case SpvBuiltInBaryCoordNoPerspSampleAMD
:
1488 *location
= SYSTEM_VALUE_BARYCENTRIC_LINEAR_SAMPLE
;
1489 set_mode_system_value(b
, mode
);
1491 case SpvBuiltInBaryCoordSmoothAMD
:
1492 *location
= SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL
;
1493 set_mode_system_value(b
, mode
);
1495 case SpvBuiltInBaryCoordSmoothCentroidAMD
:
1496 *location
= SYSTEM_VALUE_BARYCENTRIC_PERSP_CENTROID
;
1497 set_mode_system_value(b
, mode
);
1499 case SpvBuiltInBaryCoordSmoothSampleAMD
:
1500 *location
= SYSTEM_VALUE_BARYCENTRIC_PERSP_SAMPLE
;
1501 set_mode_system_value(b
, mode
);
1503 case SpvBuiltInBaryCoordPullModelAMD
:
1504 *location
= SYSTEM_VALUE_BARYCENTRIC_PULL_MODEL
;
1505 set_mode_system_value(b
, mode
);
1508 vtn_fail("Unsupported builtin: %s (%u)",
1509 spirv_builtin_to_string(builtin
), builtin
);
1514 apply_var_decoration(struct vtn_builder
*b
,
1515 struct nir_variable_data
*var_data
,
1516 const struct vtn_decoration
*dec
)
1518 switch (dec
->decoration
) {
1519 case SpvDecorationRelaxedPrecision
:
1520 break; /* FIXME: Do nothing with this for now. */
1521 case SpvDecorationNoPerspective
:
1522 var_data
->interpolation
= INTERP_MODE_NOPERSPECTIVE
;
1524 case SpvDecorationFlat
:
1525 var_data
->interpolation
= INTERP_MODE_FLAT
;
1527 case SpvDecorationExplicitInterpAMD
:
1528 var_data
->interpolation
= INTERP_MODE_EXPLICIT
;
1530 case SpvDecorationCentroid
:
1531 var_data
->centroid
= true;
1533 case SpvDecorationSample
:
1534 var_data
->sample
= true;
1536 case SpvDecorationInvariant
:
1537 var_data
->invariant
= true;
1539 case SpvDecorationConstant
:
1540 var_data
->read_only
= true;
1542 case SpvDecorationNonReadable
:
1543 var_data
->access
|= ACCESS_NON_READABLE
;
1545 case SpvDecorationNonWritable
:
1546 var_data
->read_only
= true;
1547 var_data
->access
|= ACCESS_NON_WRITEABLE
;
1549 case SpvDecorationRestrict
:
1550 var_data
->access
|= ACCESS_RESTRICT
;
1552 case SpvDecorationAliased
:
1553 var_data
->access
&= ~ACCESS_RESTRICT
;
1555 case SpvDecorationVolatile
:
1556 var_data
->access
|= ACCESS_VOLATILE
;
1558 case SpvDecorationCoherent
:
1559 var_data
->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 SpvDecorationUniform
:
1591 case SpvDecorationUniformId
:
1592 case SpvDecorationLinkageAttributes
:
1593 break; /* Do nothing with these here */
1595 case SpvDecorationPatch
:
1596 var_data
->patch
= true;
1599 case SpvDecorationLocation
:
1600 vtn_fail("Handled above");
1602 case SpvDecorationBlock
:
1603 case SpvDecorationBufferBlock
:
1604 case SpvDecorationArrayStride
:
1605 case SpvDecorationGLSLShared
:
1606 case SpvDecorationGLSLPacked
:
1607 break; /* These can apply to a type but we don't care about them */
1609 case SpvDecorationBinding
:
1610 case SpvDecorationDescriptorSet
:
1611 case SpvDecorationNoContraction
:
1612 case SpvDecorationInputAttachmentIndex
:
1613 vtn_warn("Decoration not allowed for variable or structure member: %s",
1614 spirv_decoration_to_string(dec
->decoration
));
1617 case SpvDecorationXfbBuffer
:
1618 var_data
->explicit_xfb_buffer
= true;
1619 var_data
->xfb
.buffer
= dec
->operands
[0];
1620 var_data
->always_active_io
= true;
1622 case SpvDecorationXfbStride
:
1623 var_data
->explicit_xfb_stride
= true;
1624 var_data
->xfb
.stride
= dec
->operands
[0];
1626 case SpvDecorationOffset
:
1627 var_data
->explicit_offset
= true;
1628 var_data
->offset
= dec
->operands
[0];
1631 case SpvDecorationStream
:
1632 var_data
->stream
= dec
->operands
[0];
1635 case SpvDecorationCPacked
:
1636 case SpvDecorationSaturatedConversion
:
1637 case SpvDecorationFuncParamAttr
:
1638 case SpvDecorationFPRoundingMode
:
1639 case SpvDecorationFPFastMathMode
:
1640 case SpvDecorationAlignment
:
1641 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
) {
1642 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1643 spirv_decoration_to_string(dec
->decoration
));
1647 case SpvDecorationUserSemantic
:
1648 case SpvDecorationUserTypeGOOGLE
:
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 SpvStorageClassPhysicalStorageBuffer
:
1815 mode
= vtn_variable_mode_phys_ssbo
;
1816 nir_mode
= nir_var_mem_global
;
1818 case SpvStorageClassUniformConstant
:
1819 if (b
->shader
->info
.stage
== MESA_SHADER_KERNEL
) {
1820 if (b
->options
->constant_as_global
) {
1821 mode
= vtn_variable_mode_cross_workgroup
;
1822 nir_mode
= nir_var_mem_global
;
1824 mode
= vtn_variable_mode_ubo
;
1825 nir_mode
= nir_var_mem_ubo
;
1828 mode
= vtn_variable_mode_uniform
;
1829 nir_mode
= nir_var_uniform
;
1832 case SpvStorageClassPushConstant
:
1833 mode
= vtn_variable_mode_push_constant
;
1834 nir_mode
= nir_var_uniform
;
1836 case SpvStorageClassInput
:
1837 mode
= vtn_variable_mode_input
;
1838 nir_mode
= nir_var_shader_in
;
1840 case SpvStorageClassOutput
:
1841 mode
= vtn_variable_mode_output
;
1842 nir_mode
= nir_var_shader_out
;
1844 case SpvStorageClassPrivate
:
1845 mode
= vtn_variable_mode_private
;
1846 nir_mode
= nir_var_shader_temp
;
1848 case SpvStorageClassFunction
:
1849 mode
= vtn_variable_mode_function
;
1850 nir_mode
= nir_var_function_temp
;
1852 case SpvStorageClassWorkgroup
:
1853 mode
= vtn_variable_mode_workgroup
;
1854 nir_mode
= nir_var_mem_shared
;
1856 case SpvStorageClassAtomicCounter
:
1857 mode
= vtn_variable_mode_atomic_counter
;
1858 nir_mode
= nir_var_uniform
;
1860 case SpvStorageClassCrossWorkgroup
:
1861 mode
= vtn_variable_mode_cross_workgroup
;
1862 nir_mode
= nir_var_mem_global
;
1864 case SpvStorageClassImage
:
1865 mode
= vtn_variable_mode_image
;
1866 nir_mode
= nir_var_mem_ubo
;
1868 case SpvStorageClassGeneric
:
1870 vtn_fail("Unhandled variable storage class: %s (%u)",
1871 spirv_storageclass_to_string(class), class);
1875 *nir_mode_out
= nir_mode
;
1881 vtn_mode_to_address_format(struct vtn_builder
*b
, enum vtn_variable_mode mode
)
1884 case vtn_variable_mode_ubo
:
1885 return b
->options
->ubo_addr_format
;
1887 case vtn_variable_mode_ssbo
:
1888 return b
->options
->ssbo_addr_format
;
1890 case vtn_variable_mode_phys_ssbo
:
1891 return b
->options
->phys_ssbo_addr_format
;
1893 case vtn_variable_mode_push_constant
:
1894 return b
->options
->push_const_addr_format
;
1896 case vtn_variable_mode_workgroup
:
1897 return b
->options
->shared_addr_format
;
1899 case vtn_variable_mode_cross_workgroup
:
1900 return b
->options
->global_addr_format
;
1902 case vtn_variable_mode_function
:
1903 if (b
->physical_ptrs
)
1904 return b
->options
->temp_addr_format
;
1907 case vtn_variable_mode_private
:
1908 case vtn_variable_mode_uniform
:
1909 case vtn_variable_mode_atomic_counter
:
1910 case vtn_variable_mode_input
:
1911 case vtn_variable_mode_output
:
1912 case vtn_variable_mode_image
:
1913 return nir_address_format_logical
;
1916 unreachable("Invalid variable mode");
1920 vtn_pointer_to_ssa(struct vtn_builder
*b
, struct vtn_pointer
*ptr
)
1922 if (vtn_pointer_uses_ssa_offset(b
, ptr
)) {
1923 /* This pointer needs to have a pointer type with actual storage */
1924 vtn_assert(ptr
->ptr_type
);
1925 vtn_assert(ptr
->ptr_type
->type
);
1928 /* If we don't have an offset then we must be a pointer to the variable
1931 vtn_assert(!ptr
->offset
&& !ptr
->block_index
);
1933 struct vtn_access_chain chain
= {
1936 ptr
= vtn_ssa_offset_pointer_dereference(b
, ptr
, &chain
);
1939 vtn_assert(ptr
->offset
);
1940 if (ptr
->block_index
) {
1941 vtn_assert(ptr
->mode
== vtn_variable_mode_ubo
||
1942 ptr
->mode
== vtn_variable_mode_ssbo
);
1943 return nir_vec2(&b
->nb
, ptr
->block_index
, ptr
->offset
);
1945 vtn_assert(ptr
->mode
== vtn_variable_mode_workgroup
);
1949 if (vtn_pointer_is_external_block(b
, ptr
) &&
1950 vtn_type_contains_block(b
, ptr
->type
) &&
1951 ptr
->mode
!= vtn_variable_mode_phys_ssbo
) {
1952 /* In this case, we're looking for a block index and not an actual
1955 * For PhysicalStorageBuffer pointers, we don't have a block index
1956 * at all because we get the pointer directly from the client. This
1957 * assumes that there will never be a SSBO binding variable using the
1958 * PhysicalStorageBuffer storage class. This assumption appears
1959 * to be correct according to the Vulkan spec because the table,
1960 * "Shader Resource and Storage Class Correspondence," the only the
1961 * Uniform storage class with BufferBlock or the StorageBuffer
1962 * storage class with Block can be used.
1964 if (!ptr
->block_index
) {
1965 /* If we don't have a block_index then we must be a pointer to the
1968 vtn_assert(!ptr
->deref
);
1970 struct vtn_access_chain chain
= {
1973 ptr
= vtn_nir_deref_pointer_dereference(b
, ptr
, &chain
);
1976 return ptr
->block_index
;
1978 return &vtn_pointer_to_deref(b
, ptr
)->dest
.ssa
;
1983 struct vtn_pointer
*
1984 vtn_pointer_from_ssa(struct vtn_builder
*b
, nir_ssa_def
*ssa
,
1985 struct vtn_type
*ptr_type
)
1987 vtn_assert(ptr_type
->base_type
== vtn_base_type_pointer
);
1989 struct vtn_pointer
*ptr
= rzalloc(b
, struct vtn_pointer
);
1990 struct vtn_type
*without_array
=
1991 vtn_type_without_array(ptr_type
->deref
);
1993 nir_variable_mode nir_mode
;
1994 ptr
->mode
= vtn_storage_class_to_mode(b
, ptr_type
->storage_class
,
1995 without_array
, &nir_mode
);
1996 ptr
->type
= ptr_type
->deref
;
1997 ptr
->ptr_type
= ptr_type
;
1999 if (vtn_pointer_uses_ssa_offset(b
, ptr
)) {
2000 /* This pointer type needs to have actual storage */
2001 vtn_assert(ptr_type
->type
);
2002 if (ptr
->mode
== vtn_variable_mode_ubo
||
2003 ptr
->mode
== vtn_variable_mode_ssbo
) {
2004 vtn_assert(ssa
->num_components
== 2);
2005 ptr
->block_index
= nir_channel(&b
->nb
, ssa
, 0);
2006 ptr
->offset
= nir_channel(&b
->nb
, ssa
, 1);
2008 vtn_assert(ssa
->num_components
== 1);
2009 ptr
->block_index
= NULL
;
2013 const struct glsl_type
*deref_type
=
2014 vtn_type_get_nir_type(b
, ptr_type
->deref
, ptr
->mode
);
2015 if (!vtn_pointer_is_external_block(b
, ptr
)) {
2016 ptr
->deref
= nir_build_deref_cast(&b
->nb
, ssa
, nir_mode
,
2017 deref_type
, ptr_type
->stride
);
2018 } else if (vtn_type_contains_block(b
, ptr
->type
) &&
2019 ptr
->mode
!= vtn_variable_mode_phys_ssbo
) {
2020 /* This is a pointer to somewhere in an array of blocks, not a
2021 * pointer to somewhere inside the block. Set the block index
2022 * instead of making a cast.
2024 ptr
->block_index
= ssa
;
2026 /* This is a pointer to something internal or a pointer inside a
2027 * block. It's just a regular cast.
2029 * For PhysicalStorageBuffer pointers, we don't have a block index
2030 * at all because we get the pointer directly from the client. This
2031 * assumes that there will never be a SSBO binding variable using the
2032 * PhysicalStorageBuffer storage class. This assumption appears
2033 * to be correct according to the Vulkan spec because the table,
2034 * "Shader Resource and Storage Class Correspondence," the only the
2035 * Uniform storage class with BufferBlock or the StorageBuffer
2036 * storage class with Block can be used.
2038 ptr
->deref
= nir_build_deref_cast(&b
->nb
, ssa
, nir_mode
,
2039 deref_type
, ptr_type
->stride
);
2040 ptr
->deref
->dest
.ssa
.num_components
=
2041 glsl_get_vector_elements(ptr_type
->type
);
2042 ptr
->deref
->dest
.ssa
.bit_size
= glsl_get_bit_size(ptr_type
->type
);
2050 is_per_vertex_inout(const struct vtn_variable
*var
, gl_shader_stage stage
)
2052 if (var
->patch
|| !glsl_type_is_array(var
->type
->type
))
2055 if (var
->mode
== vtn_variable_mode_input
) {
2056 return stage
== MESA_SHADER_TESS_CTRL
||
2057 stage
== MESA_SHADER_TESS_EVAL
||
2058 stage
== MESA_SHADER_GEOMETRY
;
2061 if (var
->mode
== vtn_variable_mode_output
)
2062 return stage
== MESA_SHADER_TESS_CTRL
;
2068 assign_missing_member_locations(struct vtn_variable
*var
)
2071 glsl_get_length(glsl_without_array(var
->type
->type
));
2072 int location
= var
->base_location
;
2074 for (unsigned i
= 0; i
< length
; i
++) {
2075 /* From the Vulkan spec:
2077 * “If the structure type is a Block but without a Location, then each
2078 * of its members must have a Location decoration.”
2081 if (var
->type
->block
) {
2082 assert(var
->base_location
!= -1 ||
2083 var
->var
->members
[i
].location
!= -1);
2086 /* From the Vulkan spec:
2088 * “Any member with its own Location decoration is assigned that
2089 * location. Each remaining member is assigned the location after the
2090 * immediately preceding member in declaration order.”
2092 if (var
->var
->members
[i
].location
!= -1)
2093 location
= var
->var
->members
[i
].location
;
2095 var
->var
->members
[i
].location
= location
;
2097 /* Below we use type instead of interface_type, because interface_type
2098 * is only available when it is a Block. This code also supports
2099 * input/outputs that are just structs
2101 const struct glsl_type
*member_type
=
2102 glsl_get_struct_field(glsl_without_array(var
->type
->type
), i
);
2105 glsl_count_attribute_slots(member_type
,
2106 false /* is_gl_vertex_input */);
2112 vtn_create_variable(struct vtn_builder
*b
, struct vtn_value
*val
,
2113 struct vtn_type
*ptr_type
, SpvStorageClass storage_class
,
2114 nir_constant
*const_initializer
, nir_variable
*var_initializer
)
2116 vtn_assert(ptr_type
->base_type
== vtn_base_type_pointer
);
2117 struct vtn_type
*type
= ptr_type
->deref
;
2119 struct vtn_type
*without_array
= vtn_type_without_array(ptr_type
->deref
);
2121 enum vtn_variable_mode mode
;
2122 nir_variable_mode nir_mode
;
2123 mode
= vtn_storage_class_to_mode(b
, storage_class
, without_array
, &nir_mode
);
2126 case vtn_variable_mode_ubo
:
2127 /* There's no other way to get vtn_variable_mode_ubo */
2128 vtn_assert(without_array
->block
);
2129 b
->shader
->info
.num_ubos
++;
2131 case vtn_variable_mode_ssbo
:
2132 if (storage_class
== SpvStorageClassStorageBuffer
&&
2133 !without_array
->block
) {
2134 if (b
->variable_pointers
) {
2135 vtn_fail("Variables in the StorageBuffer storage class must "
2136 "have a struct type with the Block decoration");
2138 /* If variable pointers are not present, it's still malformed
2139 * SPIR-V but we can parse it and do the right thing anyway.
2140 * Since some of the 8-bit storage tests have bugs in this are,
2141 * just make it a warning for now.
2143 vtn_warn("Variables in the StorageBuffer storage class must "
2144 "have a struct type with the Block decoration");
2147 b
->shader
->info
.num_ssbos
++;
2149 case vtn_variable_mode_uniform
:
2150 if (without_array
->base_type
== vtn_base_type_image
) {
2151 if (glsl_type_is_image(without_array
->glsl_image
))
2152 b
->shader
->info
.num_images
++;
2153 else if (glsl_type_is_sampler(without_array
->glsl_image
))
2154 b
->shader
->info
.num_textures
++;
2157 case vtn_variable_mode_push_constant
:
2158 b
->shader
->num_uniforms
= vtn_type_block_size(b
, type
);
2161 case vtn_variable_mode_image
:
2162 vtn_fail("Cannot create a variable with the Image storage class");
2165 case vtn_variable_mode_phys_ssbo
:
2166 vtn_fail("Cannot create a variable with the "
2167 "PhysicalStorageBuffer storage class");
2171 /* No tallying is needed */
2175 struct vtn_variable
*var
= rzalloc(b
, struct vtn_variable
);
2178 var
->base_location
= -1;
2180 val
->pointer
= rzalloc(b
, struct vtn_pointer
);
2181 val
->pointer
->mode
= var
->mode
;
2182 val
->pointer
->type
= var
->type
;
2183 val
->pointer
->ptr_type
= ptr_type
;
2184 val
->pointer
->var
= var
;
2185 val
->pointer
->access
= var
->type
->access
;
2187 switch (var
->mode
) {
2188 case vtn_variable_mode_function
:
2189 case vtn_variable_mode_private
:
2190 case vtn_variable_mode_uniform
:
2191 case vtn_variable_mode_atomic_counter
:
2192 /* For these, we create the variable normally */
2193 var
->var
= rzalloc(b
->shader
, nir_variable
);
2194 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
2195 var
->var
->type
= vtn_type_get_nir_type(b
, var
->type
, var
->mode
);
2196 var
->var
->data
.mode
= nir_mode
;
2197 var
->var
->data
.location
= -1;
2198 var
->var
->interface_type
= NULL
;
2201 case vtn_variable_mode_ubo
:
2202 case vtn_variable_mode_ssbo
:
2203 var
->var
= rzalloc(b
->shader
, nir_variable
);
2204 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
2206 var
->var
->type
= vtn_type_get_nir_type(b
, var
->type
, var
->mode
);
2207 var
->var
->interface_type
= var
->var
->type
;
2209 var
->var
->data
.mode
= nir_mode
;
2210 var
->var
->data
.location
= -1;
2214 case vtn_variable_mode_workgroup
:
2215 /* Create the variable normally */
2216 var
->var
= rzalloc(b
->shader
, nir_variable
);
2217 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
2218 var
->var
->type
= vtn_type_get_nir_type(b
, var
->type
, var
->mode
);
2219 var
->var
->data
.mode
= nir_var_mem_shared
;
2222 case vtn_variable_mode_input
:
2223 case vtn_variable_mode_output
: {
2224 /* In order to know whether or not we're a per-vertex inout, we need
2225 * the patch qualifier. This means walking the variable decorations
2226 * early before we actually create any variables. Not a big deal.
2228 * GLSLang really likes to place decorations in the most interior
2229 * thing it possibly can. In particular, if you have a struct, it
2230 * will place the patch decorations on the struct members. This
2231 * should be handled by the variable splitting below just fine.
2233 * If you have an array-of-struct, things get even more weird as it
2234 * will place the patch decorations on the struct even though it's
2235 * inside an array and some of the members being patch and others not
2236 * makes no sense whatsoever. Since the only sensible thing is for
2237 * it to be all or nothing, we'll call it patch if any of the members
2238 * are declared patch.
2241 vtn_foreach_decoration(b
, val
, var_is_patch_cb
, &var
->patch
);
2242 if (glsl_type_is_array(var
->type
->type
) &&
2243 glsl_type_is_struct_or_ifc(without_array
->type
)) {
2244 vtn_foreach_decoration(b
, vtn_value(b
, without_array
->id
,
2245 vtn_value_type_type
),
2246 var_is_patch_cb
, &var
->patch
);
2249 /* For inputs and outputs, we immediately split structures. This
2250 * is for a couple of reasons. For one, builtins may all come in
2251 * a struct and we really want those split out into separate
2252 * variables. For another, interpolation qualifiers can be
2253 * applied to members of the top-level struct ane we need to be
2254 * able to preserve that information.
2257 struct vtn_type
*per_vertex_type
= var
->type
;
2258 if (is_per_vertex_inout(var
, b
->shader
->info
.stage
)) {
2259 /* In Geometry shaders (and some tessellation), inputs come
2260 * in per-vertex arrays. However, some builtins come in
2261 * non-per-vertex, hence the need for the is_array check. In
2262 * any case, there are no non-builtin arrays allowed so this
2263 * check should be sufficient.
2265 per_vertex_type
= var
->type
->array_element
;
2268 var
->var
= rzalloc(b
->shader
, nir_variable
);
2269 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
2270 var
->var
->type
= vtn_type_get_nir_type(b
, var
->type
, var
->mode
);
2271 var
->var
->data
.mode
= nir_mode
;
2272 var
->var
->data
.patch
= var
->patch
;
2274 /* Figure out the interface block type. */
2275 struct vtn_type
*iface_type
= per_vertex_type
;
2276 if (var
->mode
== vtn_variable_mode_output
&&
2277 (b
->shader
->info
.stage
== MESA_SHADER_VERTEX
||
2278 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
||
2279 b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
)) {
2280 /* For vertex data outputs, we can end up with arrays of blocks for
2281 * transform feedback where each array element corresponds to a
2282 * different XFB output buffer.
2284 while (iface_type
->base_type
== vtn_base_type_array
)
2285 iface_type
= iface_type
->array_element
;
2287 if (iface_type
->base_type
== vtn_base_type_struct
&& iface_type
->block
)
2288 var
->var
->interface_type
= vtn_type_get_nir_type(b
, iface_type
,
2291 if (per_vertex_type
->base_type
== vtn_base_type_struct
&&
2292 per_vertex_type
->block
) {
2293 /* It's a struct. Set it up as per-member. */
2294 var
->var
->num_members
= glsl_get_length(per_vertex_type
->type
);
2295 var
->var
->members
= rzalloc_array(var
->var
, struct nir_variable_data
,
2296 var
->var
->num_members
);
2298 for (unsigned i
= 0; i
< var
->var
->num_members
; i
++) {
2299 var
->var
->members
[i
].mode
= nir_mode
;
2300 var
->var
->members
[i
].patch
= var
->patch
;
2301 var
->var
->members
[i
].location
= -1;
2305 /* For inputs and outputs, we need to grab locations and builtin
2306 * information from the per-vertex type.
2308 vtn_foreach_decoration(b
, vtn_value(b
, per_vertex_type
->id
,
2309 vtn_value_type_type
),
2310 var_decoration_cb
, var
);
2314 case vtn_variable_mode_push_constant
:
2315 case vtn_variable_mode_cross_workgroup
:
2316 /* These don't need actual variables. */
2319 case vtn_variable_mode_image
:
2320 case vtn_variable_mode_phys_ssbo
:
2321 unreachable("Should have been caught before");
2324 /* We can only have one type of initializer */
2325 assert(!(const_initializer
&& var_initializer
));
2326 if (const_initializer
) {
2327 var
->var
->constant_initializer
=
2328 nir_constant_clone(const_initializer
, var
->var
);
2330 if (var_initializer
)
2331 var
->var
->pointer_initializer
= var_initializer
;
2333 if (var
->mode
== vtn_variable_mode_uniform
||
2334 var
->mode
== vtn_variable_mode_ssbo
) {
2335 /* SSBOs and images are assumed to not alias in the Simple, GLSL and Vulkan memory models */
2336 var
->var
->data
.access
|= b
->mem_model
!= SpvMemoryModelOpenCL
? ACCESS_RESTRICT
: 0;
2339 vtn_foreach_decoration(b
, val
, var_decoration_cb
, var
);
2340 vtn_foreach_decoration(b
, val
, ptr_decoration_cb
, val
->pointer
);
2342 /* Propagate access flags from the OpVariable decorations. */
2343 val
->pointer
->access
|= var
->access
;
2345 if ((var
->mode
== vtn_variable_mode_input
||
2346 var
->mode
== vtn_variable_mode_output
) &&
2347 var
->var
->members
) {
2348 assign_missing_member_locations(var
);
2351 if (var
->mode
== vtn_variable_mode_uniform
||
2352 var
->mode
== vtn_variable_mode_ubo
||
2353 var
->mode
== vtn_variable_mode_ssbo
||
2354 var
->mode
== vtn_variable_mode_atomic_counter
) {
2355 /* XXX: We still need the binding information in the nir_variable
2356 * for these. We should fix that.
2358 var
->var
->data
.binding
= var
->binding
;
2359 var
->var
->data
.explicit_binding
= var
->explicit_binding
;
2360 var
->var
->data
.descriptor_set
= var
->descriptor_set
;
2361 var
->var
->data
.index
= var
->input_attachment_index
;
2362 var
->var
->data
.offset
= var
->offset
;
2364 if (glsl_type_is_image(glsl_without_array(var
->var
->type
)))
2365 var
->var
->data
.image
.format
= without_array
->image_format
;
2368 if (var
->mode
== vtn_variable_mode_function
) {
2369 vtn_assert(var
->var
!= NULL
&& var
->var
->members
== NULL
);
2370 nir_function_impl_add_variable(b
->nb
.impl
, var
->var
);
2371 } else if (var
->var
) {
2372 nir_shader_add_variable(b
->shader
, var
->var
);
2374 vtn_assert(vtn_pointer_is_external_block(b
, val
->pointer
));
2379 vtn_assert_types_equal(struct vtn_builder
*b
, SpvOp opcode
,
2380 struct vtn_type
*dst_type
,
2381 struct vtn_type
*src_type
)
2383 if (dst_type
->id
== src_type
->id
)
2386 if (vtn_types_compatible(b
, dst_type
, src_type
)) {
2387 /* Early versions of GLSLang would re-emit types unnecessarily and you
2388 * would end up with OpLoad, OpStore, or OpCopyMemory opcodes which have
2389 * mismatched source and destination types.
2391 * https://github.com/KhronosGroup/glslang/issues/304
2392 * https://github.com/KhronosGroup/glslang/issues/307
2393 * https://bugs.freedesktop.org/show_bug.cgi?id=104338
2394 * https://bugs.freedesktop.org/show_bug.cgi?id=104424
2396 vtn_warn("Source and destination types of %s do not have the same "
2397 "ID (but are compatible): %u vs %u",
2398 spirv_op_to_string(opcode
), dst_type
->id
, src_type
->id
);
2402 vtn_fail("Source and destination types of %s do not match: %s vs. %s",
2403 spirv_op_to_string(opcode
),
2404 glsl_get_type_name(dst_type
->type
),
2405 glsl_get_type_name(src_type
->type
));
2408 static nir_ssa_def
*
2409 nir_shrink_zero_pad_vec(nir_builder
*b
, nir_ssa_def
*val
,
2410 unsigned num_components
)
2412 if (val
->num_components
== num_components
)
2415 nir_ssa_def
*comps
[NIR_MAX_VEC_COMPONENTS
];
2416 for (unsigned i
= 0; i
< num_components
; i
++) {
2417 if (i
< val
->num_components
)
2418 comps
[i
] = nir_channel(b
, val
, i
);
2420 comps
[i
] = nir_imm_intN_t(b
, 0, val
->bit_size
);
2422 return nir_vec(b
, comps
, num_components
);
2425 static nir_ssa_def
*
2426 nir_sloppy_bitcast(nir_builder
*b
, nir_ssa_def
*val
,
2427 const struct glsl_type
*type
)
2429 const unsigned num_components
= glsl_get_vector_elements(type
);
2430 const unsigned bit_size
= glsl_get_bit_size(type
);
2432 /* First, zero-pad to ensure that the value is big enough that when we
2433 * bit-cast it, we don't loose anything.
2435 if (val
->bit_size
< bit_size
) {
2436 const unsigned src_num_components_needed
=
2437 vtn_align_u32(val
->num_components
, bit_size
/ val
->bit_size
);
2438 val
= nir_shrink_zero_pad_vec(b
, val
, src_num_components_needed
);
2441 val
= nir_bitcast_vector(b
, val
, bit_size
);
2443 return nir_shrink_zero_pad_vec(b
, val
, num_components
);
2447 vtn_handle_variables(struct vtn_builder
*b
, SpvOp opcode
,
2448 const uint32_t *w
, unsigned count
)
2452 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
2453 val
->type
= vtn_get_type(b
, w
[1]);
2457 case SpvOpVariable
: {
2458 struct vtn_type
*ptr_type
= vtn_get_type(b
, w
[1]);
2460 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
2462 SpvStorageClass storage_class
= w
[3];
2463 nir_constant
*const_initializer
= NULL
;
2464 nir_variable
*var_initializer
= NULL
;
2466 struct vtn_value
*init
= vtn_untyped_value(b
, w
[4]);
2467 switch (init
->value_type
) {
2468 case vtn_value_type_constant
:
2469 const_initializer
= init
->constant
;
2471 case vtn_value_type_pointer
:
2472 var_initializer
= init
->pointer
->var
->var
;
2475 vtn_fail("SPIR-V variable initializer %u must be constant or pointer",
2480 vtn_create_variable(b
, val
, ptr_type
, storage_class
, const_initializer
, var_initializer
);
2485 case SpvOpAccessChain
:
2486 case SpvOpPtrAccessChain
:
2487 case SpvOpInBoundsAccessChain
:
2488 case SpvOpInBoundsPtrAccessChain
: {
2489 struct vtn_access_chain
*chain
= vtn_access_chain_create(b
, count
- 4);
2490 enum gl_access_qualifier access
= 0;
2491 chain
->ptr_as_array
= (opcode
== SpvOpPtrAccessChain
|| opcode
== SpvOpInBoundsPtrAccessChain
);
2494 for (int i
= 4; i
< count
; i
++) {
2495 struct vtn_value
*link_val
= vtn_untyped_value(b
, w
[i
]);
2496 if (link_val
->value_type
== vtn_value_type_constant
) {
2497 chain
->link
[idx
].mode
= vtn_access_mode_literal
;
2498 chain
->link
[idx
].id
= vtn_constant_int(b
, w
[i
]);
2500 chain
->link
[idx
].mode
= vtn_access_mode_id
;
2501 chain
->link
[idx
].id
= w
[i
];
2506 struct vtn_type
*ptr_type
= vtn_get_type(b
, w
[1]);
2507 struct vtn_pointer
*base
=
2508 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2509 struct vtn_pointer
*ptr
= vtn_pointer_dereference(b
, base
, chain
);
2510 ptr
->ptr_type
= ptr_type
;
2511 ptr
->access
|= access
;
2512 vtn_push_pointer(b
, w
[2], ptr
);
2516 case SpvOpCopyMemory
: {
2517 struct vtn_value
*dest
= vtn_value(b
, w
[1], vtn_value_type_pointer
);
2518 struct vtn_value
*src
= vtn_value(b
, w
[2], vtn_value_type_pointer
);
2520 vtn_assert_types_equal(b
, opcode
, dest
->type
->deref
, src
->type
->deref
);
2522 vtn_variable_copy(b
, dest
->pointer
, src
->pointer
);
2527 struct vtn_type
*res_type
= vtn_get_type(b
, w
[1]);
2528 struct vtn_value
*src_val
= vtn_value(b
, w
[3], vtn_value_type_pointer
);
2529 struct vtn_pointer
*src
= src_val
->pointer
;
2531 vtn_assert_types_equal(b
, opcode
, res_type
, src_val
->type
->deref
);
2535 SpvMemoryAccessMask access
= w
[4];
2536 if (access
& SpvMemoryAccessAlignedMask
)
2539 if (access
& SpvMemoryAccessMakePointerVisibleMask
) {
2540 SpvMemorySemanticsMask semantics
=
2541 SpvMemorySemanticsMakeVisibleMask
|
2542 vtn_storage_class_to_memory_semantics(src
->ptr_type
->storage_class
);
2544 SpvScope scope
= vtn_constant_uint(b
, w
[idx
]);
2545 vtn_emit_memory_barrier(b
, scope
, semantics
);
2549 vtn_push_ssa_value(b
, w
[2], vtn_variable_load(b
, src
));
2554 struct vtn_value
*dest_val
= vtn_value(b
, w
[1], vtn_value_type_pointer
);
2555 struct vtn_pointer
*dest
= dest_val
->pointer
;
2556 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[2]);
2558 /* OpStore requires us to actually have a storage type */
2559 vtn_fail_if(dest
->type
->type
== NULL
,
2560 "Invalid destination type for OpStore");
2562 if (glsl_get_base_type(dest
->type
->type
) == GLSL_TYPE_BOOL
&&
2563 glsl_get_base_type(src_val
->type
->type
) == GLSL_TYPE_UINT
) {
2564 /* Early versions of GLSLang would use uint types for UBOs/SSBOs but
2565 * would then store them to a local variable as bool. Work around
2566 * the issue by doing an implicit conversion.
2568 * https://github.com/KhronosGroup/glslang/issues/170
2569 * https://bugs.freedesktop.org/show_bug.cgi?id=104424
2571 vtn_warn("OpStore of value of type OpTypeInt to a pointer to type "
2572 "OpTypeBool. Doing an implicit conversion to work around "
2574 struct vtn_ssa_value
*bool_ssa
=
2575 vtn_create_ssa_value(b
, dest
->type
->type
);
2576 bool_ssa
->def
= nir_i2b(&b
->nb
, vtn_ssa_value(b
, w
[2])->def
);
2577 vtn_variable_store(b
, bool_ssa
, dest
);
2581 vtn_assert_types_equal(b
, opcode
, dest_val
->type
->deref
, src_val
->type
);
2583 struct vtn_ssa_value
*src
= vtn_ssa_value(b
, w
[2]);
2584 vtn_variable_store(b
, src
, dest
);
2588 SpvMemoryAccessMask access
= w
[3];
2590 if (access
& SpvMemoryAccessAlignedMask
)
2593 if (access
& SpvMemoryAccessMakePointerAvailableMask
) {
2594 SpvMemorySemanticsMask semantics
=
2595 SpvMemorySemanticsMakeAvailableMask
|
2596 vtn_storage_class_to_memory_semantics(dest
->ptr_type
->storage_class
);
2597 SpvScope scope
= vtn_constant_uint(b
, w
[idx
]);
2598 vtn_emit_memory_barrier(b
, scope
, semantics
);
2604 case SpvOpArrayLength
: {
2605 struct vtn_pointer
*ptr
=
2606 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2607 const uint32_t field
= w
[4];
2609 vtn_fail_if(ptr
->type
->base_type
!= vtn_base_type_struct
,
2610 "OpArrayLength must take a pointer to a structure type");
2611 vtn_fail_if(field
!= ptr
->type
->length
- 1 ||
2612 ptr
->type
->members
[field
]->base_type
!= vtn_base_type_array
,
2613 "OpArrayLength must reference the last memeber of the "
2614 "structure and that must be an array");
2616 const uint32_t offset
= ptr
->type
->offsets
[field
];
2617 const uint32_t stride
= ptr
->type
->members
[field
]->stride
;
2619 if (!ptr
->block_index
) {
2620 struct vtn_access_chain chain
= {
2623 ptr
= vtn_pointer_dereference(b
, ptr
, &chain
);
2624 vtn_assert(ptr
->block_index
);
2627 nir_intrinsic_instr
*instr
=
2628 nir_intrinsic_instr_create(b
->nb
.shader
,
2629 nir_intrinsic_get_buffer_size
);
2630 instr
->src
[0] = nir_src_for_ssa(ptr
->block_index
);
2631 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
2632 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
2633 nir_ssa_def
*buf_size
= &instr
->dest
.ssa
;
2635 /* array_length = max(buffer_size - offset, 0) / stride */
2636 nir_ssa_def
*array_length
=
2641 nir_imm_int(&b
->nb
, offset
)),
2642 nir_imm_int(&b
->nb
, 0u)),
2643 nir_imm_int(&b
->nb
, stride
));
2645 vtn_push_nir_ssa(b
, w
[2], array_length
);
2649 case SpvOpConvertPtrToU
: {
2650 struct vtn_type
*u_type
= vtn_get_type(b
, w
[1]);
2651 struct vtn_type
*ptr_type
= vtn_get_value_type(b
, w
[3]);
2653 vtn_fail_if(ptr_type
->base_type
!= vtn_base_type_pointer
||
2654 ptr_type
->type
== NULL
,
2655 "OpConvertPtrToU can only be used on physical pointers");
2657 vtn_fail_if(u_type
->base_type
!= vtn_base_type_vector
&&
2658 u_type
->base_type
!= vtn_base_type_scalar
,
2659 "OpConvertPtrToU can only be used to cast to a vector or "
2662 /* The pointer will be converted to an SSA value automatically */
2663 nir_ssa_def
*ptr
= vtn_get_nir_ssa(b
, w
[3]);
2664 nir_ssa_def
*u
= nir_sloppy_bitcast(&b
->nb
, ptr
, u_type
->type
);
2665 vtn_push_nir_ssa(b
, w
[2], u
);
2669 case SpvOpConvertUToPtr
: {
2670 struct vtn_type
*ptr_type
= vtn_get_type(b
, w
[1]);
2671 struct vtn_type
*u_type
= vtn_get_value_type(b
, w
[3]);
2673 vtn_fail_if(ptr_type
->base_type
!= vtn_base_type_pointer
||
2674 ptr_type
->type
== NULL
,
2675 "OpConvertUToPtr can only be used on physical pointers");
2677 vtn_fail_if(u_type
->base_type
!= vtn_base_type_vector
&&
2678 u_type
->base_type
!= vtn_base_type_scalar
,
2679 "OpConvertUToPtr can only be used to cast from a vector or "
2682 nir_ssa_def
*u
= vtn_get_nir_ssa(b
, w
[3]);
2683 nir_ssa_def
*ptr
= nir_sloppy_bitcast(&b
->nb
, u
, ptr_type
->type
);
2684 vtn_push_pointer(b
, w
[2], vtn_pointer_from_ssa(b
, ptr
, ptr_type
));
2688 case SpvOpCopyMemorySized
:
2690 vtn_fail_with_opcode("Unhandled opcode", opcode
);