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
30 #include "anv_private.h"
32 #include "genxml/gen8_pack.h"
34 #include "util/debug.h"
36 /** \file anv_batch_chain.c
38 * This file contains functions related to anv_cmd_buffer as a data
39 * structure. This involves everything required to create and destroy
40 * the actual batch buffers as well as link them together and handle
41 * relocations and surface state. It specifically does *not* contain any
42 * handling of actual vkCmd calls beyond vkCmdExecuteCommands.
45 /*-----------------------------------------------------------------------*
46 * Functions related to anv_reloc_list
47 *-----------------------------------------------------------------------*/
50 anv_reloc_list_init(struct anv_reloc_list
*list
,
51 const VkAllocationCallbacks
*alloc
)
53 memset(list
, 0, sizeof(*list
));
58 anv_reloc_list_init_clone(struct anv_reloc_list
*list
,
59 const VkAllocationCallbacks
*alloc
,
60 const struct anv_reloc_list
*other_list
)
62 list
->num_relocs
= other_list
->num_relocs
;
63 list
->array_length
= other_list
->array_length
;
65 if (list
->num_relocs
> 0) {
67 vk_alloc(alloc
, list
->array_length
* sizeof(*list
->relocs
), 8,
68 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
69 if (list
->relocs
== NULL
)
70 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
73 vk_alloc(alloc
, list
->array_length
* sizeof(*list
->reloc_bos
), 8,
74 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
75 if (list
->reloc_bos
== NULL
) {
76 vk_free(alloc
, list
->relocs
);
77 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
80 memcpy(list
->relocs
, other_list
->relocs
,
81 list
->array_length
* sizeof(*list
->relocs
));
82 memcpy(list
->reloc_bos
, other_list
->reloc_bos
,
83 list
->array_length
* sizeof(*list
->reloc_bos
));
86 list
->reloc_bos
= NULL
;
89 list
->dep_words
= other_list
->dep_words
;
91 if (list
->dep_words
> 0) {
93 vk_alloc(alloc
, list
->dep_words
* sizeof(BITSET_WORD
), 8,
94 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
95 memcpy(list
->deps
, other_list
->deps
,
96 list
->dep_words
* sizeof(BITSET_WORD
));
105 anv_reloc_list_finish(struct anv_reloc_list
*list
,
106 const VkAllocationCallbacks
*alloc
)
108 vk_free(alloc
, list
->relocs
);
109 vk_free(alloc
, list
->reloc_bos
);
110 vk_free(alloc
, list
->deps
);
114 anv_reloc_list_grow(struct anv_reloc_list
*list
,
115 const VkAllocationCallbacks
*alloc
,
116 size_t num_additional_relocs
)
118 if (list
->num_relocs
+ num_additional_relocs
<= list
->array_length
)
121 size_t new_length
= MAX2(16, list
->array_length
* 2);
122 while (new_length
< list
->num_relocs
+ num_additional_relocs
)
125 struct drm_i915_gem_relocation_entry
*new_relocs
=
126 vk_realloc(alloc
, list
->relocs
,
127 new_length
* sizeof(*list
->relocs
), 8,
128 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
129 if (new_relocs
== NULL
)
130 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
131 list
->relocs
= new_relocs
;
133 struct anv_bo
**new_reloc_bos
=
134 vk_realloc(alloc
, list
->reloc_bos
,
135 new_length
* sizeof(*list
->reloc_bos
), 8,
136 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
137 if (new_reloc_bos
== NULL
)
138 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
139 list
->reloc_bos
= new_reloc_bos
;
141 list
->array_length
= new_length
;
147 anv_reloc_list_grow_deps(struct anv_reloc_list
*list
,
148 const VkAllocationCallbacks
*alloc
,
149 uint32_t min_num_words
)
151 if (min_num_words
<= list
->dep_words
)
154 uint32_t new_length
= MAX2(32, list
->dep_words
* 2);
155 while (new_length
< min_num_words
)
158 BITSET_WORD
*new_deps
=
159 vk_realloc(alloc
, list
->deps
, new_length
* sizeof(BITSET_WORD
), 8,
160 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
161 if (new_deps
== NULL
)
162 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
163 list
->deps
= new_deps
;
165 /* Zero out the new data */
166 memset(list
->deps
+ list
->dep_words
, 0,
167 (new_length
- list
->dep_words
) * sizeof(BITSET_WORD
));
168 list
->dep_words
= new_length
;
173 #define READ_ONCE(x) (*(volatile __typeof__(x) *)&(x))
176 anv_reloc_list_add(struct anv_reloc_list
*list
,
177 const VkAllocationCallbacks
*alloc
,
178 uint32_t offset
, struct anv_bo
*target_bo
, uint32_t delta
,
179 uint64_t *address_u64_out
)
181 struct drm_i915_gem_relocation_entry
*entry
;
184 struct anv_bo
*unwrapped_target_bo
= anv_bo_unwrap(target_bo
);
185 uint64_t target_bo_offset
= READ_ONCE(unwrapped_target_bo
->offset
);
187 *address_u64_out
= target_bo_offset
+ delta
;
189 if (unwrapped_target_bo
->flags
& EXEC_OBJECT_PINNED
) {
190 assert(!target_bo
->is_wrapper
);
191 uint32_t idx
= unwrapped_target_bo
->gem_handle
;
192 anv_reloc_list_grow_deps(list
, alloc
, (idx
/ BITSET_WORDBITS
) + 1);
193 BITSET_SET(list
->deps
, unwrapped_target_bo
->gem_handle
);
197 VkResult result
= anv_reloc_list_grow(list
, alloc
, 1);
198 if (result
!= VK_SUCCESS
)
201 /* XXX: Can we use I915_EXEC_HANDLE_LUT? */
202 index
= list
->num_relocs
++;
203 list
->reloc_bos
[index
] = target_bo
;
204 entry
= &list
->relocs
[index
];
205 entry
->target_handle
= -1; /* See also anv_cmd_buffer_process_relocs() */
206 entry
->delta
= delta
;
207 entry
->offset
= offset
;
208 entry
->presumed_offset
= target_bo_offset
;
209 entry
->read_domains
= 0;
210 entry
->write_domain
= 0;
211 VG(VALGRIND_CHECK_MEM_IS_DEFINED(entry
, sizeof(*entry
)));
217 anv_reloc_list_clear(struct anv_reloc_list
*list
)
219 list
->num_relocs
= 0;
220 if (list
->dep_words
> 0)
221 memset(list
->deps
, 0, list
->dep_words
* sizeof(BITSET_WORD
));
225 anv_reloc_list_append(struct anv_reloc_list
*list
,
226 const VkAllocationCallbacks
*alloc
,
227 struct anv_reloc_list
*other
, uint32_t offset
)
229 VkResult result
= anv_reloc_list_grow(list
, alloc
, other
->num_relocs
);
230 if (result
!= VK_SUCCESS
)
233 if (other
->num_relocs
> 0) {
234 memcpy(&list
->relocs
[list
->num_relocs
], &other
->relocs
[0],
235 other
->num_relocs
* sizeof(other
->relocs
[0]));
236 memcpy(&list
->reloc_bos
[list
->num_relocs
], &other
->reloc_bos
[0],
237 other
->num_relocs
* sizeof(other
->reloc_bos
[0]));
239 for (uint32_t i
= 0; i
< other
->num_relocs
; i
++)
240 list
->relocs
[i
+ list
->num_relocs
].offset
+= offset
;
242 list
->num_relocs
+= other
->num_relocs
;
245 anv_reloc_list_grow_deps(list
, alloc
, other
->dep_words
);
246 for (uint32_t w
= 0; w
< other
->dep_words
; w
++)
247 list
->deps
[w
] |= other
->deps
[w
];
252 /*-----------------------------------------------------------------------*
253 * Functions related to anv_batch
254 *-----------------------------------------------------------------------*/
257 anv_batch_emit_dwords(struct anv_batch
*batch
, int num_dwords
)
259 if (batch
->next
+ num_dwords
* 4 > batch
->end
) {
260 VkResult result
= batch
->extend_cb(batch
, batch
->user_data
);
261 if (result
!= VK_SUCCESS
) {
262 anv_batch_set_error(batch
, result
);
267 void *p
= batch
->next
;
269 batch
->next
+= num_dwords
* 4;
270 assert(batch
->next
<= batch
->end
);
276 anv_batch_emit_reloc(struct anv_batch
*batch
,
277 void *location
, struct anv_bo
*bo
, uint32_t delta
)
279 uint64_t address_u64
= 0;
280 VkResult result
= anv_reloc_list_add(batch
->relocs
, batch
->alloc
,
281 location
- batch
->start
, bo
, delta
,
283 if (result
!= VK_SUCCESS
) {
284 anv_batch_set_error(batch
, result
);
292 anv_batch_emit_batch(struct anv_batch
*batch
, struct anv_batch
*other
)
294 uint32_t size
, offset
;
296 size
= other
->next
- other
->start
;
297 assert(size
% 4 == 0);
299 if (batch
->next
+ size
> batch
->end
) {
300 VkResult result
= batch
->extend_cb(batch
, batch
->user_data
);
301 if (result
!= VK_SUCCESS
) {
302 anv_batch_set_error(batch
, result
);
307 assert(batch
->next
+ size
<= batch
->end
);
309 VG(VALGRIND_CHECK_MEM_IS_DEFINED(other
->start
, size
));
310 memcpy(batch
->next
, other
->start
, size
);
312 offset
= batch
->next
- batch
->start
;
313 VkResult result
= anv_reloc_list_append(batch
->relocs
, batch
->alloc
,
314 other
->relocs
, offset
);
315 if (result
!= VK_SUCCESS
) {
316 anv_batch_set_error(batch
, result
);
323 /*-----------------------------------------------------------------------*
324 * Functions related to anv_batch_bo
325 *-----------------------------------------------------------------------*/
328 anv_batch_bo_create(struct anv_cmd_buffer
*cmd_buffer
,
329 struct anv_batch_bo
**bbo_out
)
333 struct anv_batch_bo
*bbo
= vk_alloc(&cmd_buffer
->pool
->alloc
, sizeof(*bbo
),
334 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
336 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
338 result
= anv_bo_pool_alloc(&cmd_buffer
->device
->batch_bo_pool
,
339 ANV_CMD_BUFFER_BATCH_SIZE
, &bbo
->bo
);
340 if (result
!= VK_SUCCESS
)
343 result
= anv_reloc_list_init(&bbo
->relocs
, &cmd_buffer
->pool
->alloc
);
344 if (result
!= VK_SUCCESS
)
352 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, bbo
->bo
);
354 vk_free(&cmd_buffer
->pool
->alloc
, bbo
);
360 anv_batch_bo_clone(struct anv_cmd_buffer
*cmd_buffer
,
361 const struct anv_batch_bo
*other_bbo
,
362 struct anv_batch_bo
**bbo_out
)
366 struct anv_batch_bo
*bbo
= vk_alloc(&cmd_buffer
->pool
->alloc
, sizeof(*bbo
),
367 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
369 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
371 result
= anv_bo_pool_alloc(&cmd_buffer
->device
->batch_bo_pool
,
372 other_bbo
->bo
->size
, &bbo
->bo
);
373 if (result
!= VK_SUCCESS
)
376 result
= anv_reloc_list_init_clone(&bbo
->relocs
, &cmd_buffer
->pool
->alloc
,
378 if (result
!= VK_SUCCESS
)
381 bbo
->length
= other_bbo
->length
;
382 memcpy(bbo
->bo
->map
, other_bbo
->bo
->map
, other_bbo
->length
);
388 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, bbo
->bo
);
390 vk_free(&cmd_buffer
->pool
->alloc
, bbo
);
396 anv_batch_bo_start(struct anv_batch_bo
*bbo
, struct anv_batch
*batch
,
397 size_t batch_padding
)
399 batch
->next
= batch
->start
= bbo
->bo
->map
;
400 batch
->end
= bbo
->bo
->map
+ bbo
->bo
->size
- batch_padding
;
401 batch
->relocs
= &bbo
->relocs
;
402 anv_reloc_list_clear(&bbo
->relocs
);
406 anv_batch_bo_continue(struct anv_batch_bo
*bbo
, struct anv_batch
*batch
,
407 size_t batch_padding
)
409 batch
->start
= bbo
->bo
->map
;
410 batch
->next
= bbo
->bo
->map
+ bbo
->length
;
411 batch
->end
= bbo
->bo
->map
+ bbo
->bo
->size
- batch_padding
;
412 batch
->relocs
= &bbo
->relocs
;
416 anv_batch_bo_finish(struct anv_batch_bo
*bbo
, struct anv_batch
*batch
)
418 assert(batch
->start
== bbo
->bo
->map
);
419 bbo
->length
= batch
->next
- batch
->start
;
420 VG(VALGRIND_CHECK_MEM_IS_DEFINED(batch
->start
, bbo
->length
));
424 anv_batch_bo_grow(struct anv_cmd_buffer
*cmd_buffer
, struct anv_batch_bo
*bbo
,
425 struct anv_batch
*batch
, size_t aditional
,
426 size_t batch_padding
)
428 assert(batch
->start
== bbo
->bo
->map
);
429 bbo
->length
= batch
->next
- batch
->start
;
431 size_t new_size
= bbo
->bo
->size
;
432 while (new_size
<= bbo
->length
+ aditional
+ batch_padding
)
435 if (new_size
== bbo
->bo
->size
)
438 struct anv_bo
*new_bo
;
439 VkResult result
= anv_bo_pool_alloc(&cmd_buffer
->device
->batch_bo_pool
,
441 if (result
!= VK_SUCCESS
)
444 memcpy(new_bo
->map
, bbo
->bo
->map
, bbo
->length
);
446 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, bbo
->bo
);
449 anv_batch_bo_continue(bbo
, batch
, batch_padding
);
455 anv_batch_bo_link(struct anv_cmd_buffer
*cmd_buffer
,
456 struct anv_batch_bo
*prev_bbo
,
457 struct anv_batch_bo
*next_bbo
,
458 uint32_t next_bbo_offset
)
460 const uint32_t bb_start_offset
=
461 prev_bbo
->length
- GEN8_MI_BATCH_BUFFER_START_length
* 4;
462 ASSERTED
const uint32_t *bb_start
= prev_bbo
->bo
->map
+ bb_start_offset
;
464 /* Make sure we're looking at a MI_BATCH_BUFFER_START */
465 assert(((*bb_start
>> 29) & 0x07) == 0);
466 assert(((*bb_start
>> 23) & 0x3f) == 49);
468 if (cmd_buffer
->device
->instance
->physicalDevice
.use_softpin
) {
469 assert(prev_bbo
->bo
->flags
& EXEC_OBJECT_PINNED
);
470 assert(next_bbo
->bo
->flags
& EXEC_OBJECT_PINNED
);
472 write_reloc(cmd_buffer
->device
,
473 prev_bbo
->bo
->map
+ bb_start_offset
+ 4,
474 next_bbo
->bo
->offset
+ next_bbo_offset
, true);
476 uint32_t reloc_idx
= prev_bbo
->relocs
.num_relocs
- 1;
477 assert(prev_bbo
->relocs
.relocs
[reloc_idx
].offset
== bb_start_offset
+ 4);
479 prev_bbo
->relocs
.reloc_bos
[reloc_idx
] = next_bbo
->bo
;
480 prev_bbo
->relocs
.relocs
[reloc_idx
].delta
= next_bbo_offset
;
482 /* Use a bogus presumed offset to force a relocation */
483 prev_bbo
->relocs
.relocs
[reloc_idx
].presumed_offset
= -1;
488 anv_batch_bo_destroy(struct anv_batch_bo
*bbo
,
489 struct anv_cmd_buffer
*cmd_buffer
)
491 anv_reloc_list_finish(&bbo
->relocs
, &cmd_buffer
->pool
->alloc
);
492 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, bbo
->bo
);
493 vk_free(&cmd_buffer
->pool
->alloc
, bbo
);
497 anv_batch_bo_list_clone(const struct list_head
*list
,
498 struct anv_cmd_buffer
*cmd_buffer
,
499 struct list_head
*new_list
)
501 VkResult result
= VK_SUCCESS
;
503 list_inithead(new_list
);
505 struct anv_batch_bo
*prev_bbo
= NULL
;
506 list_for_each_entry(struct anv_batch_bo
, bbo
, list
, link
) {
507 struct anv_batch_bo
*new_bbo
= NULL
;
508 result
= anv_batch_bo_clone(cmd_buffer
, bbo
, &new_bbo
);
509 if (result
!= VK_SUCCESS
)
511 list_addtail(&new_bbo
->link
, new_list
);
514 anv_batch_bo_link(cmd_buffer
, prev_bbo
, new_bbo
, 0);
519 if (result
!= VK_SUCCESS
) {
520 list_for_each_entry_safe(struct anv_batch_bo
, bbo
, new_list
, link
) {
521 list_del(&bbo
->link
);
522 anv_batch_bo_destroy(bbo
, cmd_buffer
);
529 /*-----------------------------------------------------------------------*
530 * Functions related to anv_batch_bo
531 *-----------------------------------------------------------------------*/
533 static struct anv_batch_bo
*
534 anv_cmd_buffer_current_batch_bo(struct anv_cmd_buffer
*cmd_buffer
)
536 return LIST_ENTRY(struct anv_batch_bo
, cmd_buffer
->batch_bos
.prev
, link
);
540 anv_cmd_buffer_surface_base_address(struct anv_cmd_buffer
*cmd_buffer
)
542 struct anv_state
*bt_block
= u_vector_head(&cmd_buffer
->bt_block_states
);
543 return (struct anv_address
) {
544 .bo
= anv_binding_table_pool(cmd_buffer
->device
)->block_pool
.bo
,
545 .offset
= bt_block
->offset
,
550 emit_batch_buffer_start(struct anv_cmd_buffer
*cmd_buffer
,
551 struct anv_bo
*bo
, uint32_t offset
)
553 /* In gen8+ the address field grew to two dwords to accomodate 48 bit
554 * offsets. The high 16 bits are in the last dword, so we can use the gen8
555 * version in either case, as long as we set the instruction length in the
556 * header accordingly. This means that we always emit three dwords here
557 * and all the padding and adjustment we do in this file works for all
561 #define GEN7_MI_BATCH_BUFFER_START_length 2
562 #define GEN7_MI_BATCH_BUFFER_START_length_bias 2
564 const uint32_t gen7_length
=
565 GEN7_MI_BATCH_BUFFER_START_length
- GEN7_MI_BATCH_BUFFER_START_length_bias
;
566 const uint32_t gen8_length
=
567 GEN8_MI_BATCH_BUFFER_START_length
- GEN8_MI_BATCH_BUFFER_START_length_bias
;
569 anv_batch_emit(&cmd_buffer
->batch
, GEN8_MI_BATCH_BUFFER_START
, bbs
) {
570 bbs
.DWordLength
= cmd_buffer
->device
->info
.gen
< 8 ?
571 gen7_length
: gen8_length
;
572 bbs
.SecondLevelBatchBuffer
= Firstlevelbatch
;
573 bbs
.AddressSpaceIndicator
= ASI_PPGTT
;
574 bbs
.BatchBufferStartAddress
= (struct anv_address
) { bo
, offset
};
579 cmd_buffer_chain_to_batch_bo(struct anv_cmd_buffer
*cmd_buffer
,
580 struct anv_batch_bo
*bbo
)
582 struct anv_batch
*batch
= &cmd_buffer
->batch
;
583 struct anv_batch_bo
*current_bbo
=
584 anv_cmd_buffer_current_batch_bo(cmd_buffer
);
586 /* We set the end of the batch a little short so we would be sure we
587 * have room for the chaining command. Since we're about to emit the
588 * chaining command, let's set it back where it should go.
590 batch
->end
+= GEN8_MI_BATCH_BUFFER_START_length
* 4;
591 assert(batch
->end
== current_bbo
->bo
->map
+ current_bbo
->bo
->size
);
593 emit_batch_buffer_start(cmd_buffer
, bbo
->bo
, 0);
595 anv_batch_bo_finish(current_bbo
, batch
);
599 anv_cmd_buffer_chain_batch(struct anv_batch
*batch
, void *_data
)
601 struct anv_cmd_buffer
*cmd_buffer
= _data
;
602 struct anv_batch_bo
*new_bbo
;
604 VkResult result
= anv_batch_bo_create(cmd_buffer
, &new_bbo
);
605 if (result
!= VK_SUCCESS
)
608 struct anv_batch_bo
**seen_bbo
= u_vector_add(&cmd_buffer
->seen_bbos
);
609 if (seen_bbo
== NULL
) {
610 anv_batch_bo_destroy(new_bbo
, cmd_buffer
);
611 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
615 cmd_buffer_chain_to_batch_bo(cmd_buffer
, new_bbo
);
617 list_addtail(&new_bbo
->link
, &cmd_buffer
->batch_bos
);
619 anv_batch_bo_start(new_bbo
, batch
, GEN8_MI_BATCH_BUFFER_START_length
* 4);
625 anv_cmd_buffer_grow_batch(struct anv_batch
*batch
, void *_data
)
627 struct anv_cmd_buffer
*cmd_buffer
= _data
;
628 struct anv_batch_bo
*bbo
= anv_cmd_buffer_current_batch_bo(cmd_buffer
);
630 anv_batch_bo_grow(cmd_buffer
, bbo
, &cmd_buffer
->batch
, 4096,
631 GEN8_MI_BATCH_BUFFER_START_length
* 4);
636 /** Allocate a binding table
638 * This function allocates a binding table. This is a bit more complicated
639 * than one would think due to a combination of Vulkan driver design and some
640 * unfortunate hardware restrictions.
642 * The 3DSTATE_BINDING_TABLE_POINTERS_* packets only have a 16-bit field for
643 * the binding table pointer which means that all binding tables need to live
644 * in the bottom 64k of surface state base address. The way the GL driver has
645 * classically dealt with this restriction is to emit all surface states
646 * on-the-fly into the batch and have a batch buffer smaller than 64k. This
647 * isn't really an option in Vulkan for a couple of reasons:
649 * 1) In Vulkan, we have growing (or chaining) batches so surface states have
650 * to live in their own buffer and we have to be able to re-emit
651 * STATE_BASE_ADDRESS as needed which requires a full pipeline stall. In
652 * order to avoid emitting STATE_BASE_ADDRESS any more often than needed
653 * (it's not that hard to hit 64k of just binding tables), we allocate
654 * surface state objects up-front when VkImageView is created. In order
655 * for this to work, surface state objects need to be allocated from a
658 * 2) We tried to design the surface state system in such a way that it's
659 * already ready for bindless texturing. The way bindless texturing works
660 * on our hardware is that you have a big pool of surface state objects
661 * (with its own state base address) and the bindless handles are simply
662 * offsets into that pool. With the architecture we chose, we already
663 * have that pool and it's exactly the same pool that we use for regular
664 * surface states so we should already be ready for bindless.
666 * 3) For render targets, we need to be able to fill out the surface states
667 * later in vkBeginRenderPass so that we can assign clear colors
668 * correctly. One way to do this would be to just create the surface
669 * state data and then repeatedly copy it into the surface state BO every
670 * time we have to re-emit STATE_BASE_ADDRESS. While this works, it's
671 * rather annoying and just being able to allocate them up-front and
672 * re-use them for the entire render pass.
674 * While none of these are technically blockers for emitting state on the fly
675 * like we do in GL, the ability to have a single surface state pool is
676 * simplifies things greatly. Unfortunately, it comes at a cost...
678 * Because of the 64k limitation of 3DSTATE_BINDING_TABLE_POINTERS_*, we can't
679 * place the binding tables just anywhere in surface state base address.
680 * Because 64k isn't a whole lot of space, we can't simply restrict the
681 * surface state buffer to 64k, we have to be more clever. The solution we've
682 * chosen is to have a block pool with a maximum size of 2G that starts at
683 * zero and grows in both directions. All surface states are allocated from
684 * the top of the pool (positive offsets) and we allocate blocks (< 64k) of
685 * binding tables from the bottom of the pool (negative offsets). Every time
686 * we allocate a new binding table block, we set surface state base address to
687 * point to the bottom of the binding table block. This way all of the
688 * binding tables in the block are in the bottom 64k of surface state base
689 * address. When we fill out the binding table, we add the distance between
690 * the bottom of our binding table block and zero of the block pool to the
691 * surface state offsets so that they are correct relative to out new surface
692 * state base address at the bottom of the binding table block.
694 * \see adjust_relocations_from_block_pool()
695 * \see adjust_relocations_too_block_pool()
697 * \param[in] entries The number of surface state entries the binding
698 * table should be able to hold.
700 * \param[out] state_offset The offset surface surface state base address
701 * where the surface states live. This must be
702 * added to the surface state offset when it is
703 * written into the binding table entry.
705 * \return An anv_state representing the binding table
708 anv_cmd_buffer_alloc_binding_table(struct anv_cmd_buffer
*cmd_buffer
,
709 uint32_t entries
, uint32_t *state_offset
)
711 struct anv_device
*device
= cmd_buffer
->device
;
712 struct anv_state_pool
*state_pool
= &device
->surface_state_pool
;
713 struct anv_state
*bt_block
= u_vector_head(&cmd_buffer
->bt_block_states
);
714 struct anv_state state
;
716 state
.alloc_size
= align_u32(entries
* 4, 32);
718 if (cmd_buffer
->bt_next
+ state
.alloc_size
> state_pool
->block_size
)
719 return (struct anv_state
) { 0 };
721 state
.offset
= cmd_buffer
->bt_next
;
722 state
.map
= anv_block_pool_map(&anv_binding_table_pool(device
)->block_pool
,
723 bt_block
->offset
+ state
.offset
);
725 cmd_buffer
->bt_next
+= state
.alloc_size
;
727 if (device
->instance
->physicalDevice
.use_softpin
) {
728 assert(bt_block
->offset
>= 0);
729 *state_offset
= device
->surface_state_pool
.block_pool
.start_address
-
730 device
->binding_table_pool
.block_pool
.start_address
- bt_block
->offset
;
732 assert(bt_block
->offset
< 0);
733 *state_offset
= -bt_block
->offset
;
740 anv_cmd_buffer_alloc_surface_state(struct anv_cmd_buffer
*cmd_buffer
)
742 struct isl_device
*isl_dev
= &cmd_buffer
->device
->isl_dev
;
743 return anv_state_stream_alloc(&cmd_buffer
->surface_state_stream
,
744 isl_dev
->ss
.size
, isl_dev
->ss
.align
);
748 anv_cmd_buffer_alloc_dynamic_state(struct anv_cmd_buffer
*cmd_buffer
,
749 uint32_t size
, uint32_t alignment
)
751 return anv_state_stream_alloc(&cmd_buffer
->dynamic_state_stream
,
756 anv_cmd_buffer_new_binding_table_block(struct anv_cmd_buffer
*cmd_buffer
)
758 struct anv_state
*bt_block
= u_vector_add(&cmd_buffer
->bt_block_states
);
759 if (bt_block
== NULL
) {
760 anv_batch_set_error(&cmd_buffer
->batch
, VK_ERROR_OUT_OF_HOST_MEMORY
);
761 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
764 *bt_block
= anv_binding_table_pool_alloc(cmd_buffer
->device
);
765 cmd_buffer
->bt_next
= 0;
771 anv_cmd_buffer_init_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
)
773 struct anv_batch_bo
*batch_bo
;
776 list_inithead(&cmd_buffer
->batch_bos
);
778 result
= anv_batch_bo_create(cmd_buffer
, &batch_bo
);
779 if (result
!= VK_SUCCESS
)
782 list_addtail(&batch_bo
->link
, &cmd_buffer
->batch_bos
);
784 cmd_buffer
->batch
.alloc
= &cmd_buffer
->pool
->alloc
;
785 cmd_buffer
->batch
.user_data
= cmd_buffer
;
787 if (cmd_buffer
->device
->can_chain_batches
) {
788 cmd_buffer
->batch
.extend_cb
= anv_cmd_buffer_chain_batch
;
790 cmd_buffer
->batch
.extend_cb
= anv_cmd_buffer_grow_batch
;
793 anv_batch_bo_start(batch_bo
, &cmd_buffer
->batch
,
794 GEN8_MI_BATCH_BUFFER_START_length
* 4);
796 int success
= u_vector_init(&cmd_buffer
->seen_bbos
,
797 sizeof(struct anv_bo
*),
798 8 * sizeof(struct anv_bo
*));
802 *(struct anv_batch_bo
**)u_vector_add(&cmd_buffer
->seen_bbos
) = batch_bo
;
804 /* u_vector requires power-of-two size elements */
805 unsigned pow2_state_size
= util_next_power_of_two(sizeof(struct anv_state
));
806 success
= u_vector_init(&cmd_buffer
->bt_block_states
,
807 pow2_state_size
, 8 * pow2_state_size
);
811 result
= anv_reloc_list_init(&cmd_buffer
->surface_relocs
,
812 &cmd_buffer
->pool
->alloc
);
813 if (result
!= VK_SUCCESS
)
815 cmd_buffer
->last_ss_pool_center
= 0;
817 result
= anv_cmd_buffer_new_binding_table_block(cmd_buffer
);
818 if (result
!= VK_SUCCESS
)
824 u_vector_finish(&cmd_buffer
->bt_block_states
);
826 u_vector_finish(&cmd_buffer
->seen_bbos
);
828 anv_batch_bo_destroy(batch_bo
, cmd_buffer
);
834 anv_cmd_buffer_fini_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
)
836 struct anv_state
*bt_block
;
837 u_vector_foreach(bt_block
, &cmd_buffer
->bt_block_states
)
838 anv_binding_table_pool_free(cmd_buffer
->device
, *bt_block
);
839 u_vector_finish(&cmd_buffer
->bt_block_states
);
841 anv_reloc_list_finish(&cmd_buffer
->surface_relocs
, &cmd_buffer
->pool
->alloc
);
843 u_vector_finish(&cmd_buffer
->seen_bbos
);
845 /* Destroy all of the batch buffers */
846 list_for_each_entry_safe(struct anv_batch_bo
, bbo
,
847 &cmd_buffer
->batch_bos
, link
) {
848 list_del(&bbo
->link
);
849 anv_batch_bo_destroy(bbo
, cmd_buffer
);
854 anv_cmd_buffer_reset_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
)
856 /* Delete all but the first batch bo */
857 assert(!list_is_empty(&cmd_buffer
->batch_bos
));
858 while (cmd_buffer
->batch_bos
.next
!= cmd_buffer
->batch_bos
.prev
) {
859 struct anv_batch_bo
*bbo
= anv_cmd_buffer_current_batch_bo(cmd_buffer
);
860 list_del(&bbo
->link
);
861 anv_batch_bo_destroy(bbo
, cmd_buffer
);
863 assert(!list_is_empty(&cmd_buffer
->batch_bos
));
865 anv_batch_bo_start(anv_cmd_buffer_current_batch_bo(cmd_buffer
),
867 GEN8_MI_BATCH_BUFFER_START_length
* 4);
869 while (u_vector_length(&cmd_buffer
->bt_block_states
) > 1) {
870 struct anv_state
*bt_block
= u_vector_remove(&cmd_buffer
->bt_block_states
);
871 anv_binding_table_pool_free(cmd_buffer
->device
, *bt_block
);
873 assert(u_vector_length(&cmd_buffer
->bt_block_states
) == 1);
874 cmd_buffer
->bt_next
= 0;
876 anv_reloc_list_clear(&cmd_buffer
->surface_relocs
);
877 cmd_buffer
->last_ss_pool_center
= 0;
879 /* Reset the list of seen buffers */
880 cmd_buffer
->seen_bbos
.head
= 0;
881 cmd_buffer
->seen_bbos
.tail
= 0;
883 *(struct anv_batch_bo
**)u_vector_add(&cmd_buffer
->seen_bbos
) =
884 anv_cmd_buffer_current_batch_bo(cmd_buffer
);
888 anv_cmd_buffer_end_batch_buffer(struct anv_cmd_buffer
*cmd_buffer
)
890 struct anv_batch_bo
*batch_bo
= anv_cmd_buffer_current_batch_bo(cmd_buffer
);
892 if (cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
) {
893 /* When we start a batch buffer, we subtract a certain amount of
894 * padding from the end to ensure that we always have room to emit a
895 * BATCH_BUFFER_START to chain to the next BO. We need to remove
896 * that padding before we end the batch; otherwise, we may end up
897 * with our BATCH_BUFFER_END in another BO.
899 cmd_buffer
->batch
.end
+= GEN8_MI_BATCH_BUFFER_START_length
* 4;
900 assert(cmd_buffer
->batch
.end
== batch_bo
->bo
->map
+ batch_bo
->bo
->size
);
902 anv_batch_emit(&cmd_buffer
->batch
, GEN8_MI_BATCH_BUFFER_END
, bbe
);
904 /* Round batch up to an even number of dwords. */
905 if ((cmd_buffer
->batch
.next
- cmd_buffer
->batch
.start
) & 4)
906 anv_batch_emit(&cmd_buffer
->batch
, GEN8_MI_NOOP
, noop
);
908 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_PRIMARY
;
910 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_SECONDARY
);
911 /* If this is a secondary command buffer, we need to determine the
912 * mode in which it will be executed with vkExecuteCommands. We
913 * determine this statically here so that this stays in sync with the
914 * actual ExecuteCommands implementation.
916 const uint32_t length
= cmd_buffer
->batch
.next
- cmd_buffer
->batch
.start
;
917 if (!cmd_buffer
->device
->can_chain_batches
) {
918 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_GROW_AND_EMIT
;
919 } else if ((cmd_buffer
->batch_bos
.next
== cmd_buffer
->batch_bos
.prev
) &&
920 (length
< ANV_CMD_BUFFER_BATCH_SIZE
/ 2)) {
921 /* If the secondary has exactly one batch buffer in its list *and*
922 * that batch buffer is less than half of the maximum size, we're
923 * probably better of simply copying it into our batch.
925 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_EMIT
;
926 } else if (!(cmd_buffer
->usage_flags
&
927 VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT
)) {
928 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_CHAIN
;
930 /* In order to chain, we need this command buffer to contain an
931 * MI_BATCH_BUFFER_START which will jump back to the calling batch.
932 * It doesn't matter where it points now so long as has a valid
933 * relocation. We'll adjust it later as part of the chaining
936 * We set the end of the batch a little short so we would be sure we
937 * have room for the chaining command. Since we're about to emit the
938 * chaining command, let's set it back where it should go.
940 cmd_buffer
->batch
.end
+= GEN8_MI_BATCH_BUFFER_START_length
* 4;
941 assert(cmd_buffer
->batch
.start
== batch_bo
->bo
->map
);
942 assert(cmd_buffer
->batch
.end
== batch_bo
->bo
->map
+ batch_bo
->bo
->size
);
944 emit_batch_buffer_start(cmd_buffer
, batch_bo
->bo
, 0);
945 assert(cmd_buffer
->batch
.start
== batch_bo
->bo
->map
);
947 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN
;
951 anv_batch_bo_finish(batch_bo
, &cmd_buffer
->batch
);
955 anv_cmd_buffer_add_seen_bbos(struct anv_cmd_buffer
*cmd_buffer
,
956 struct list_head
*list
)
958 list_for_each_entry(struct anv_batch_bo
, bbo
, list
, link
) {
959 struct anv_batch_bo
**bbo_ptr
= u_vector_add(&cmd_buffer
->seen_bbos
);
961 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
970 anv_cmd_buffer_add_secondary(struct anv_cmd_buffer
*primary
,
971 struct anv_cmd_buffer
*secondary
)
973 switch (secondary
->exec_mode
) {
974 case ANV_CMD_BUFFER_EXEC_MODE_EMIT
:
975 anv_batch_emit_batch(&primary
->batch
, &secondary
->batch
);
977 case ANV_CMD_BUFFER_EXEC_MODE_GROW_AND_EMIT
: {
978 struct anv_batch_bo
*bbo
= anv_cmd_buffer_current_batch_bo(primary
);
979 unsigned length
= secondary
->batch
.end
- secondary
->batch
.start
;
980 anv_batch_bo_grow(primary
, bbo
, &primary
->batch
, length
,
981 GEN8_MI_BATCH_BUFFER_START_length
* 4);
982 anv_batch_emit_batch(&primary
->batch
, &secondary
->batch
);
985 case ANV_CMD_BUFFER_EXEC_MODE_CHAIN
: {
986 struct anv_batch_bo
*first_bbo
=
987 list_first_entry(&secondary
->batch_bos
, struct anv_batch_bo
, link
);
988 struct anv_batch_bo
*last_bbo
=
989 list_last_entry(&secondary
->batch_bos
, struct anv_batch_bo
, link
);
991 emit_batch_buffer_start(primary
, first_bbo
->bo
, 0);
993 struct anv_batch_bo
*this_bbo
= anv_cmd_buffer_current_batch_bo(primary
);
994 assert(primary
->batch
.start
== this_bbo
->bo
->map
);
995 uint32_t offset
= primary
->batch
.next
- primary
->batch
.start
;
997 /* Make the tail of the secondary point back to right after the
998 * MI_BATCH_BUFFER_START in the primary batch.
1000 anv_batch_bo_link(primary
, last_bbo
, this_bbo
, offset
);
1002 anv_cmd_buffer_add_seen_bbos(primary
, &secondary
->batch_bos
);
1005 case ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN
: {
1006 struct list_head copy_list
;
1007 VkResult result
= anv_batch_bo_list_clone(&secondary
->batch_bos
,
1010 if (result
!= VK_SUCCESS
)
1013 anv_cmd_buffer_add_seen_bbos(primary
, ©_list
);
1015 struct anv_batch_bo
*first_bbo
=
1016 list_first_entry(©_list
, struct anv_batch_bo
, link
);
1017 struct anv_batch_bo
*last_bbo
=
1018 list_last_entry(©_list
, struct anv_batch_bo
, link
);
1020 cmd_buffer_chain_to_batch_bo(primary
, first_bbo
);
1022 list_splicetail(©_list
, &primary
->batch_bos
);
1024 anv_batch_bo_continue(last_bbo
, &primary
->batch
,
1025 GEN8_MI_BATCH_BUFFER_START_length
* 4);
1029 assert(!"Invalid execution mode");
1032 anv_reloc_list_append(&primary
->surface_relocs
, &primary
->pool
->alloc
,
1033 &secondary
->surface_relocs
, 0);
1036 struct anv_execbuf
{
1037 struct drm_i915_gem_execbuffer2 execbuf
;
1039 struct drm_i915_gem_exec_object2
* objects
;
1041 struct anv_bo
** bos
;
1043 /* Allocated length of the 'objects' and 'bos' arrays */
1044 uint32_t array_length
;
1048 const VkAllocationCallbacks
* alloc
;
1049 VkSystemAllocationScope alloc_scope
;
1050 uint32_t fence_count
;
1051 uint32_t fence_array_length
;
1052 struct drm_i915_gem_exec_fence
* fences
;
1053 struct anv_syncobj
** syncobjs
;
1057 anv_execbuf_init(struct anv_execbuf
*exec
)
1059 memset(exec
, 0, sizeof(*exec
));
1063 anv_execbuf_finish(struct anv_execbuf
*exec
)
1065 vk_free(exec
->alloc
, exec
->objects
);
1066 vk_free(exec
->alloc
, exec
->bos
);
1067 vk_free(exec
->alloc
, exec
->fences
);
1068 vk_free(exec
->alloc
, exec
->syncobjs
);
1072 anv_execbuf_add_bo_bitset(struct anv_device
*device
,
1073 struct anv_execbuf
*exec
,
1076 uint32_t extra_flags
);
1079 anv_execbuf_add_bo(struct anv_device
*device
,
1080 struct anv_execbuf
*exec
,
1082 struct anv_reloc_list
*relocs
,
1083 uint32_t extra_flags
)
1085 struct drm_i915_gem_exec_object2
*obj
= NULL
;
1087 bo
= anv_bo_unwrap(bo
);
1089 if (bo
->index
< exec
->bo_count
&& exec
->bos
[bo
->index
] == bo
)
1090 obj
= &exec
->objects
[bo
->index
];
1093 /* We've never seen this one before. Add it to the list and assign
1094 * an id that we can use later.
1096 if (exec
->bo_count
>= exec
->array_length
) {
1097 uint32_t new_len
= exec
->objects
? exec
->array_length
* 2 : 64;
1099 struct drm_i915_gem_exec_object2
*new_objects
=
1100 vk_alloc(exec
->alloc
, new_len
* sizeof(*new_objects
), 8, exec
->alloc_scope
);
1101 if (new_objects
== NULL
)
1102 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1104 struct anv_bo
**new_bos
=
1105 vk_alloc(exec
->alloc
, new_len
* sizeof(*new_bos
), 8, exec
->alloc_scope
);
1106 if (new_bos
== NULL
) {
1107 vk_free(exec
->alloc
, new_objects
);
1108 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1111 if (exec
->objects
) {
1112 memcpy(new_objects
, exec
->objects
,
1113 exec
->bo_count
* sizeof(*new_objects
));
1114 memcpy(new_bos
, exec
->bos
,
1115 exec
->bo_count
* sizeof(*new_bos
));
1118 vk_free(exec
->alloc
, exec
->objects
);
1119 vk_free(exec
->alloc
, exec
->bos
);
1121 exec
->objects
= new_objects
;
1122 exec
->bos
= new_bos
;
1123 exec
->array_length
= new_len
;
1126 assert(exec
->bo_count
< exec
->array_length
);
1128 bo
->index
= exec
->bo_count
++;
1129 obj
= &exec
->objects
[bo
->index
];
1130 exec
->bos
[bo
->index
] = bo
;
1132 obj
->handle
= bo
->gem_handle
;
1133 obj
->relocation_count
= 0;
1134 obj
->relocs_ptr
= 0;
1136 obj
->offset
= bo
->offset
;
1137 obj
->flags
= bo
->flags
| extra_flags
;
1142 if (relocs
!= NULL
) {
1143 assert(obj
->relocation_count
== 0);
1145 if (relocs
->num_relocs
> 0) {
1146 /* This is the first time we've ever seen a list of relocations for
1147 * this BO. Go ahead and set the relocations and then walk the list
1148 * of relocations and add them all.
1150 exec
->has_relocs
= true;
1151 obj
->relocation_count
= relocs
->num_relocs
;
1152 obj
->relocs_ptr
= (uintptr_t) relocs
->relocs
;
1154 for (size_t i
= 0; i
< relocs
->num_relocs
; i
++) {
1157 /* A quick sanity check on relocations */
1158 assert(relocs
->relocs
[i
].offset
< bo
->size
);
1159 result
= anv_execbuf_add_bo(device
, exec
, relocs
->reloc_bos
[i
],
1161 if (result
!= VK_SUCCESS
)
1166 return anv_execbuf_add_bo_bitset(device
, exec
, relocs
->dep_words
,
1167 relocs
->deps
, extra_flags
);
1173 /* Add BO dependencies to execbuf */
1175 anv_execbuf_add_bo_bitset(struct anv_device
*device
,
1176 struct anv_execbuf
*exec
,
1179 uint32_t extra_flags
)
1181 for (uint32_t w
= 0; w
< dep_words
; w
++) {
1182 BITSET_WORD mask
= deps
[w
];
1184 int i
= u_bit_scan(&mask
);
1185 uint32_t gem_handle
= w
* BITSET_WORDBITS
+ i
;
1186 struct anv_bo
*bo
= anv_device_lookup_bo(device
, gem_handle
);
1187 assert(bo
->refcount
> 0);
1189 anv_execbuf_add_bo(device
, exec
, bo
, NULL
, extra_flags
);
1190 if (result
!= VK_SUCCESS
)
1199 anv_execbuf_add_syncobj(struct anv_execbuf
*exec
,
1200 uint32_t handle
, uint32_t flags
)
1204 if (exec
->fence_count
>= exec
->fence_array_length
) {
1205 uint32_t new_len
= MAX2(exec
->fence_array_length
* 2, 64);
1207 exec
->fences
= vk_realloc(exec
->alloc
, exec
->fences
,
1208 new_len
* sizeof(*exec
->fences
),
1209 8, exec
->alloc_scope
);
1210 if (exec
->fences
== NULL
)
1211 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1213 exec
->fence_array_length
= new_len
;
1216 exec
->fences
[exec
->fence_count
] = (struct drm_i915_gem_exec_fence
) {
1221 exec
->fence_count
++;
1227 anv_cmd_buffer_process_relocs(struct anv_cmd_buffer
*cmd_buffer
,
1228 struct anv_reloc_list
*list
)
1230 for (size_t i
= 0; i
< list
->num_relocs
; i
++)
1231 list
->relocs
[i
].target_handle
= anv_bo_unwrap(list
->reloc_bos
[i
])->index
;
1235 adjust_relocations_from_state_pool(struct anv_state_pool
*pool
,
1236 struct anv_reloc_list
*relocs
,
1237 uint32_t last_pool_center_bo_offset
)
1239 assert(last_pool_center_bo_offset
<= pool
->block_pool
.center_bo_offset
);
1240 uint32_t delta
= pool
->block_pool
.center_bo_offset
- last_pool_center_bo_offset
;
1242 for (size_t i
= 0; i
< relocs
->num_relocs
; i
++) {
1243 /* All of the relocations from this block pool to other BO's should
1244 * have been emitted relative to the surface block pool center. We
1245 * need to add the center offset to make them relative to the
1246 * beginning of the actual GEM bo.
1248 relocs
->relocs
[i
].offset
+= delta
;
1253 adjust_relocations_to_state_pool(struct anv_state_pool
*pool
,
1254 struct anv_bo
*from_bo
,
1255 struct anv_reloc_list
*relocs
,
1256 uint32_t last_pool_center_bo_offset
)
1258 assert(!from_bo
->is_wrapper
);
1259 assert(last_pool_center_bo_offset
<= pool
->block_pool
.center_bo_offset
);
1260 uint32_t delta
= pool
->block_pool
.center_bo_offset
- last_pool_center_bo_offset
;
1262 /* When we initially emit relocations into a block pool, we don't
1263 * actually know what the final center_bo_offset will be so we just emit
1264 * it as if center_bo_offset == 0. Now that we know what the center
1265 * offset is, we need to walk the list of relocations and adjust any
1266 * relocations that point to the pool bo with the correct offset.
1268 for (size_t i
= 0; i
< relocs
->num_relocs
; i
++) {
1269 if (relocs
->reloc_bos
[i
] == pool
->block_pool
.bo
) {
1270 /* Adjust the delta value in the relocation to correctly
1271 * correspond to the new delta. Initially, this value may have
1272 * been negative (if treated as unsigned), but we trust in
1273 * uint32_t roll-over to fix that for us at this point.
1275 relocs
->relocs
[i
].delta
+= delta
;
1277 /* Since the delta has changed, we need to update the actual
1278 * relocated value with the new presumed value. This function
1279 * should only be called on batch buffers, so we know it isn't in
1280 * use by the GPU at the moment.
1282 assert(relocs
->relocs
[i
].offset
< from_bo
->size
);
1283 write_reloc(pool
->block_pool
.device
,
1284 from_bo
->map
+ relocs
->relocs
[i
].offset
,
1285 relocs
->relocs
[i
].presumed_offset
+
1286 relocs
->relocs
[i
].delta
, false);
1292 anv_reloc_list_apply(struct anv_device
*device
,
1293 struct anv_reloc_list
*list
,
1295 bool always_relocate
)
1297 bo
= anv_bo_unwrap(bo
);
1299 for (size_t i
= 0; i
< list
->num_relocs
; i
++) {
1300 struct anv_bo
*target_bo
= anv_bo_unwrap(list
->reloc_bos
[i
]);
1301 if (list
->relocs
[i
].presumed_offset
== target_bo
->offset
&&
1305 void *p
= bo
->map
+ list
->relocs
[i
].offset
;
1306 write_reloc(device
, p
, target_bo
->offset
+ list
->relocs
[i
].delta
, true);
1307 list
->relocs
[i
].presumed_offset
= target_bo
->offset
;
1312 * This function applies the relocation for a command buffer and writes the
1313 * actual addresses into the buffers as per what we were told by the kernel on
1314 * the previous execbuf2 call. This should be safe to do because, for each
1315 * relocated address, we have two cases:
1317 * 1) The target BO is inactive (as seen by the kernel). In this case, it is
1318 * not in use by the GPU so updating the address is 100% ok. It won't be
1319 * in-use by the GPU (from our context) again until the next execbuf2
1320 * happens. If the kernel decides to move it in the next execbuf2, it
1321 * will have to do the relocations itself, but that's ok because it should
1322 * have all of the information needed to do so.
1324 * 2) The target BO is active (as seen by the kernel). In this case, it
1325 * hasn't moved since the last execbuffer2 call because GTT shuffling
1326 * *only* happens when the BO is idle. (From our perspective, it only
1327 * happens inside the execbuffer2 ioctl, but the shuffling may be
1328 * triggered by another ioctl, with full-ppgtt this is limited to only
1329 * execbuffer2 ioctls on the same context, or memory pressure.) Since the
1330 * target BO hasn't moved, our anv_bo::offset exactly matches the BO's GTT
1331 * address and the relocated value we are writing into the BO will be the
1332 * same as the value that is already there.
1334 * There is also a possibility that the target BO is active but the exact
1335 * RENDER_SURFACE_STATE object we are writing the relocation into isn't in
1336 * use. In this case, the address currently in the RENDER_SURFACE_STATE
1337 * may be stale but it's still safe to write the relocation because that
1338 * particular RENDER_SURFACE_STATE object isn't in-use by the GPU and
1339 * won't be until the next execbuf2 call.
1341 * By doing relocations on the CPU, we can tell the kernel that it doesn't
1342 * need to bother. We want to do this because the surface state buffer is
1343 * used by every command buffer so, if the kernel does the relocations, it
1344 * will always be busy and the kernel will always stall. This is also
1345 * probably the fastest mechanism for doing relocations since the kernel would
1346 * have to make a full copy of all the relocations lists.
1349 relocate_cmd_buffer(struct anv_cmd_buffer
*cmd_buffer
,
1350 struct anv_execbuf
*exec
)
1352 if (!exec
->has_relocs
)
1355 static int userspace_relocs
= -1;
1356 if (userspace_relocs
< 0)
1357 userspace_relocs
= env_var_as_boolean("ANV_USERSPACE_RELOCS", true);
1358 if (!userspace_relocs
)
1361 /* First, we have to check to see whether or not we can even do the
1362 * relocation. New buffers which have never been submitted to the kernel
1363 * don't have a valid offset so we need to let the kernel do relocations so
1364 * that we can get offsets for them. On future execbuf2 calls, those
1365 * buffers will have offsets and we will be able to skip relocating.
1366 * Invalid offsets are indicated by anv_bo::offset == (uint64_t)-1.
1368 for (uint32_t i
= 0; i
< exec
->bo_count
; i
++) {
1369 assert(!exec
->bos
[i
]->is_wrapper
);
1370 if (exec
->bos
[i
]->offset
== (uint64_t)-1)
1374 /* Since surface states are shared between command buffers and we don't
1375 * know what order they will be submitted to the kernel, we don't know
1376 * what address is actually written in the surface state object at any
1377 * given time. The only option is to always relocate them.
1379 struct anv_bo
*surface_state_bo
=
1380 anv_bo_unwrap(cmd_buffer
->device
->surface_state_pool
.block_pool
.bo
);
1381 anv_reloc_list_apply(cmd_buffer
->device
, &cmd_buffer
->surface_relocs
,
1383 true /* always relocate surface states */);
1385 /* Since we own all of the batch buffers, we know what values are stored
1386 * in the relocated addresses and only have to update them if the offsets
1389 struct anv_batch_bo
**bbo
;
1390 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
) {
1391 anv_reloc_list_apply(cmd_buffer
->device
,
1392 &(*bbo
)->relocs
, (*bbo
)->bo
, false);
1395 for (uint32_t i
= 0; i
< exec
->bo_count
; i
++)
1396 exec
->objects
[i
].offset
= exec
->bos
[i
]->offset
;
1402 setup_execbuf_for_cmd_buffer(struct anv_execbuf
*execbuf
,
1403 struct anv_cmd_buffer
*cmd_buffer
)
1405 struct anv_batch
*batch
= &cmd_buffer
->batch
;
1406 struct anv_state_pool
*ss_pool
=
1407 &cmd_buffer
->device
->surface_state_pool
;
1409 adjust_relocations_from_state_pool(ss_pool
, &cmd_buffer
->surface_relocs
,
1410 cmd_buffer
->last_ss_pool_center
);
1412 if (cmd_buffer
->device
->instance
->physicalDevice
.use_softpin
) {
1413 anv_block_pool_foreach_bo(bo
, &ss_pool
->block_pool
) {
1414 result
= anv_execbuf_add_bo(cmd_buffer
->device
, execbuf
,
1416 if (result
!= VK_SUCCESS
)
1419 /* Add surface dependencies (BOs) to the execbuf */
1420 anv_execbuf_add_bo_bitset(cmd_buffer
->device
, execbuf
,
1421 cmd_buffer
->surface_relocs
.dep_words
,
1422 cmd_buffer
->surface_relocs
.deps
, 0);
1424 /* Add the BOs for all memory objects */
1425 list_for_each_entry(struct anv_device_memory
, mem
,
1426 &cmd_buffer
->device
->memory_objects
, link
) {
1427 result
= anv_execbuf_add_bo(cmd_buffer
->device
, execbuf
,
1429 if (result
!= VK_SUCCESS
)
1433 struct anv_block_pool
*pool
;
1434 pool
= &cmd_buffer
->device
->dynamic_state_pool
.block_pool
;
1435 anv_block_pool_foreach_bo(bo
, pool
) {
1436 result
= anv_execbuf_add_bo(cmd_buffer
->device
, execbuf
,
1438 if (result
!= VK_SUCCESS
)
1442 pool
= &cmd_buffer
->device
->instruction_state_pool
.block_pool
;
1443 anv_block_pool_foreach_bo(bo
, pool
) {
1444 result
= anv_execbuf_add_bo(cmd_buffer
->device
, execbuf
,
1446 if (result
!= VK_SUCCESS
)
1450 pool
= &cmd_buffer
->device
->binding_table_pool
.block_pool
;
1451 anv_block_pool_foreach_bo(bo
, pool
) {
1452 result
= anv_execbuf_add_bo(cmd_buffer
->device
, execbuf
,
1454 if (result
!= VK_SUCCESS
)
1458 /* Since we aren't in the softpin case, all of our STATE_BASE_ADDRESS BOs
1459 * will get added automatically by processing relocations on the batch
1460 * buffer. We have to add the surface state BO manually because it has
1461 * relocations of its own that we need to be sure are processsed.
1463 result
= anv_execbuf_add_bo(cmd_buffer
->device
, execbuf
,
1464 ss_pool
->block_pool
.bo
,
1465 &cmd_buffer
->surface_relocs
, 0);
1466 if (result
!= VK_SUCCESS
)
1470 /* First, we walk over all of the bos we've seen and add them and their
1471 * relocations to the validate list.
1473 struct anv_batch_bo
**bbo
;
1474 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
) {
1475 adjust_relocations_to_state_pool(ss_pool
, (*bbo
)->bo
, &(*bbo
)->relocs
,
1476 cmd_buffer
->last_ss_pool_center
);
1478 result
= anv_execbuf_add_bo(cmd_buffer
->device
, execbuf
,
1479 (*bbo
)->bo
, &(*bbo
)->relocs
, 0);
1480 if (result
!= VK_SUCCESS
)
1484 /* Now that we've adjusted all of the surface state relocations, we need to
1485 * record the surface state pool center so future executions of the command
1486 * buffer can adjust correctly.
1488 cmd_buffer
->last_ss_pool_center
= ss_pool
->block_pool
.center_bo_offset
;
1490 struct anv_batch_bo
*first_batch_bo
=
1491 list_first_entry(&cmd_buffer
->batch_bos
, struct anv_batch_bo
, link
);
1493 /* The kernel requires that the last entry in the validation list be the
1494 * batch buffer to execute. We can simply swap the element
1495 * corresponding to the first batch_bo in the chain with the last
1496 * element in the list.
1498 if (first_batch_bo
->bo
->index
!= execbuf
->bo_count
- 1) {
1499 uint32_t idx
= first_batch_bo
->bo
->index
;
1500 uint32_t last_idx
= execbuf
->bo_count
- 1;
1502 struct drm_i915_gem_exec_object2 tmp_obj
= execbuf
->objects
[idx
];
1503 assert(execbuf
->bos
[idx
] == first_batch_bo
->bo
);
1505 execbuf
->objects
[idx
] = execbuf
->objects
[last_idx
];
1506 execbuf
->bos
[idx
] = execbuf
->bos
[last_idx
];
1507 execbuf
->bos
[idx
]->index
= idx
;
1509 execbuf
->objects
[last_idx
] = tmp_obj
;
1510 execbuf
->bos
[last_idx
] = first_batch_bo
->bo
;
1511 first_batch_bo
->bo
->index
= last_idx
;
1514 /* If we are pinning our BOs, we shouldn't have to relocate anything */
1515 if (cmd_buffer
->device
->instance
->physicalDevice
.use_softpin
)
1516 assert(!execbuf
->has_relocs
);
1518 /* Now we go through and fixup all of the relocation lists to point to
1519 * the correct indices in the object array. We have to do this after we
1520 * reorder the list above as some of the indices may have changed.
1522 if (execbuf
->has_relocs
) {
1523 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
)
1524 anv_cmd_buffer_process_relocs(cmd_buffer
, &(*bbo
)->relocs
);
1526 anv_cmd_buffer_process_relocs(cmd_buffer
, &cmd_buffer
->surface_relocs
);
1529 if (!cmd_buffer
->device
->info
.has_llc
) {
1530 __builtin_ia32_mfence();
1531 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
) {
1532 for (uint32_t i
= 0; i
< (*bbo
)->length
; i
+= CACHELINE_SIZE
)
1533 __builtin_ia32_clflush((*bbo
)->bo
->map
+ i
);
1537 execbuf
->execbuf
= (struct drm_i915_gem_execbuffer2
) {
1538 .buffers_ptr
= (uintptr_t) execbuf
->objects
,
1539 .buffer_count
= execbuf
->bo_count
,
1540 .batch_start_offset
= 0,
1541 .batch_len
= batch
->next
- batch
->start
,
1546 .flags
= I915_EXEC_HANDLE_LUT
| I915_EXEC_RENDER
,
1547 .rsvd1
= cmd_buffer
->device
->context_id
,
1551 if (relocate_cmd_buffer(cmd_buffer
, execbuf
)) {
1552 /* If we were able to successfully relocate everything, tell the kernel
1553 * that it can skip doing relocations. The requirement for using
1556 * 1) The addresses written in the objects must match the corresponding
1557 * reloc.presumed_offset which in turn must match the corresponding
1558 * execobject.offset.
1560 * 2) To avoid stalling, execobject.offset should match the current
1561 * address of that object within the active context.
1563 * In order to satisfy all of the invariants that make userspace
1564 * relocations to be safe (see relocate_cmd_buffer()), we need to
1565 * further ensure that the addresses we use match those used by the
1566 * kernel for the most recent execbuf2.
1568 * The kernel may still choose to do relocations anyway if something has
1569 * moved in the GTT. In this case, the relocation list still needs to be
1570 * valid. All relocations on the batch buffers are already valid and
1571 * kept up-to-date. For surface state relocations, by applying the
1572 * relocations in relocate_cmd_buffer, we ensured that the address in
1573 * the RENDER_SURFACE_STATE matches presumed_offset, so it should be
1574 * safe for the kernel to relocate them as needed.
1576 execbuf
->execbuf
.flags
|= I915_EXEC_NO_RELOC
;
1578 /* In the case where we fall back to doing kernel relocations, we need
1579 * to ensure that the relocation list is valid. All relocations on the
1580 * batch buffers are already valid and kept up-to-date. Since surface
1581 * states are shared between command buffers and we don't know what
1582 * order they will be submitted to the kernel, we don't know what
1583 * address is actually written in the surface state object at any given
1584 * time. The only option is to set a bogus presumed offset and let the
1585 * kernel relocate them.
1587 for (size_t i
= 0; i
< cmd_buffer
->surface_relocs
.num_relocs
; i
++)
1588 cmd_buffer
->surface_relocs
.relocs
[i
].presumed_offset
= -1;
1595 setup_empty_execbuf(struct anv_execbuf
*execbuf
, struct anv_device
*device
)
1597 VkResult result
= anv_execbuf_add_bo(device
, execbuf
,
1598 device
->trivial_batch_bo
,
1600 if (result
!= VK_SUCCESS
)
1603 execbuf
->execbuf
= (struct drm_i915_gem_execbuffer2
) {
1604 .buffers_ptr
= (uintptr_t) execbuf
->objects
,
1605 .buffer_count
= execbuf
->bo_count
,
1606 .batch_start_offset
= 0,
1607 .batch_len
= 8, /* GEN7_MI_BATCH_BUFFER_END and NOOP */
1608 .flags
= I915_EXEC_HANDLE_LUT
| I915_EXEC_RENDER
,
1609 .rsvd1
= device
->context_id
,
1617 anv_cmd_buffer_execbuf(struct anv_device
*device
,
1618 struct anv_cmd_buffer
*cmd_buffer
,
1619 const VkSemaphore
*in_semaphores
,
1620 uint32_t num_in_semaphores
,
1621 const VkSemaphore
*out_semaphores
,
1622 uint32_t num_out_semaphores
,
1625 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
1626 UNUSED
struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1628 struct anv_execbuf execbuf
;
1629 anv_execbuf_init(&execbuf
);
1630 execbuf
.alloc
= &device
->alloc
;
1631 execbuf
.alloc_scope
= VK_SYSTEM_ALLOCATION_SCOPE_COMMAND
;
1634 VkResult result
= VK_SUCCESS
;
1635 for (uint32_t i
= 0; i
< num_in_semaphores
; i
++) {
1636 ANV_FROM_HANDLE(anv_semaphore
, semaphore
, in_semaphores
[i
]);
1637 struct anv_semaphore_impl
*impl
=
1638 semaphore
->temporary
.type
!= ANV_SEMAPHORE_TYPE_NONE
?
1639 &semaphore
->temporary
: &semaphore
->permanent
;
1641 switch (impl
->type
) {
1642 case ANV_SEMAPHORE_TYPE_BO
:
1643 assert(!pdevice
->has_syncobj
);
1644 result
= anv_execbuf_add_bo(device
, &execbuf
, impl
->bo
, NULL
, 0);
1645 if (result
!= VK_SUCCESS
)
1649 case ANV_SEMAPHORE_TYPE_SYNC_FILE
:
1650 assert(!pdevice
->has_syncobj
);
1651 if (in_fence
== -1) {
1652 in_fence
= impl
->fd
;
1654 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1657 int merge
= anv_gem_sync_file_merge(device
, in_fence
, impl
->fd
);
1659 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1668 case ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ
:
1669 result
= anv_execbuf_add_syncobj(&execbuf
, impl
->syncobj
,
1670 I915_EXEC_FENCE_WAIT
);
1671 if (result
!= VK_SUCCESS
)
1680 bool need_out_fence
= false;
1681 for (uint32_t i
= 0; i
< num_out_semaphores
; i
++) {
1682 ANV_FROM_HANDLE(anv_semaphore
, semaphore
, out_semaphores
[i
]);
1684 /* Under most circumstances, out fences won't be temporary. However,
1685 * the spec does allow it for opaque_fd. From the Vulkan 1.0.53 spec:
1687 * "If the import is temporary, the implementation must restore the
1688 * semaphore to its prior permanent state after submitting the next
1689 * semaphore wait operation."
1691 * The spec says nothing whatsoever about signal operations on
1692 * temporarily imported semaphores so it appears they are allowed.
1693 * There are also CTS tests that require this to work.
1695 struct anv_semaphore_impl
*impl
=
1696 semaphore
->temporary
.type
!= ANV_SEMAPHORE_TYPE_NONE
?
1697 &semaphore
->temporary
: &semaphore
->permanent
;
1699 switch (impl
->type
) {
1700 case ANV_SEMAPHORE_TYPE_BO
:
1701 assert(!pdevice
->has_syncobj
);
1702 result
= anv_execbuf_add_bo(device
, &execbuf
, impl
->bo
, NULL
,
1704 if (result
!= VK_SUCCESS
)
1708 case ANV_SEMAPHORE_TYPE_SYNC_FILE
:
1709 assert(!pdevice
->has_syncobj
);
1710 need_out_fence
= true;
1713 case ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ
:
1714 result
= anv_execbuf_add_syncobj(&execbuf
, impl
->syncobj
,
1715 I915_EXEC_FENCE_SIGNAL
);
1716 if (result
!= VK_SUCCESS
)
1726 /* Under most circumstances, out fences won't be temporary. However,
1727 * the spec does allow it for opaque_fd. From the Vulkan 1.0.53 spec:
1729 * "If the import is temporary, the implementation must restore the
1730 * semaphore to its prior permanent state after submitting the next
1731 * semaphore wait operation."
1733 * The spec says nothing whatsoever about signal operations on
1734 * temporarily imported semaphores so it appears they are allowed.
1735 * There are also CTS tests that require this to work.
1737 struct anv_fence_impl
*impl
=
1738 fence
->temporary
.type
!= ANV_FENCE_TYPE_NONE
?
1739 &fence
->temporary
: &fence
->permanent
;
1741 switch (impl
->type
) {
1742 case ANV_FENCE_TYPE_BO
:
1743 assert(!pdevice
->has_syncobj_wait
);
1744 result
= anv_execbuf_add_bo(device
, &execbuf
, impl
->bo
.bo
, NULL
,
1746 if (result
!= VK_SUCCESS
)
1750 case ANV_FENCE_TYPE_SYNCOBJ
:
1751 result
= anv_execbuf_add_syncobj(&execbuf
, impl
->syncobj
,
1752 I915_EXEC_FENCE_SIGNAL
);
1753 if (result
!= VK_SUCCESS
)
1758 unreachable("Invalid fence type");
1763 if (unlikely(INTEL_DEBUG
& DEBUG_BATCH
)) {
1764 struct anv_batch_bo
**bo
= u_vector_tail(&cmd_buffer
->seen_bbos
);
1766 device
->cmd_buffer_being_decoded
= cmd_buffer
;
1767 gen_print_batch(&device
->decoder_ctx
, (*bo
)->bo
->map
,
1768 (*bo
)->bo
->size
, (*bo
)->bo
->offset
, false);
1769 device
->cmd_buffer_being_decoded
= NULL
;
1772 result
= setup_execbuf_for_cmd_buffer(&execbuf
, cmd_buffer
);
1774 result
= setup_empty_execbuf(&execbuf
, device
);
1777 if (result
!= VK_SUCCESS
)
1780 if (execbuf
.fence_count
> 0) {
1781 assert(device
->instance
->physicalDevice
.has_syncobj
);
1782 execbuf
.execbuf
.flags
|= I915_EXEC_FENCE_ARRAY
;
1783 execbuf
.execbuf
.num_cliprects
= execbuf
.fence_count
;
1784 execbuf
.execbuf
.cliprects_ptr
= (uintptr_t) execbuf
.fences
;
1787 if (in_fence
!= -1) {
1788 execbuf
.execbuf
.flags
|= I915_EXEC_FENCE_IN
;
1789 execbuf
.execbuf
.rsvd2
|= (uint32_t)in_fence
;
1793 execbuf
.execbuf
.flags
|= I915_EXEC_FENCE_OUT
;
1795 result
= anv_device_execbuf(device
, &execbuf
.execbuf
, execbuf
.bos
);
1797 /* Execbuf does not consume the in_fence. It's our job to close it. */
1801 for (uint32_t i
= 0; i
< num_in_semaphores
; i
++) {
1802 ANV_FROM_HANDLE(anv_semaphore
, semaphore
, in_semaphores
[i
]);
1803 /* From the Vulkan 1.0.53 spec:
1805 * "If the import is temporary, the implementation must restore the
1806 * semaphore to its prior permanent state after submitting the next
1807 * semaphore wait operation."
1809 * This has to happen after the execbuf in case we close any syncobjs in
1812 anv_semaphore_reset_temporary(device
, semaphore
);
1815 if (fence
&& fence
->permanent
.type
== ANV_FENCE_TYPE_BO
) {
1816 assert(!pdevice
->has_syncobj_wait
);
1817 /* BO fences can't be shared, so they can't be temporary. */
1818 assert(fence
->temporary
.type
== ANV_FENCE_TYPE_NONE
);
1820 /* Once the execbuf has returned, we need to set the fence state to
1821 * SUBMITTED. We can't do this before calling execbuf because
1822 * anv_GetFenceStatus does take the global device lock before checking
1825 * We set the fence state to SUBMITTED regardless of whether or not the
1826 * execbuf succeeds because we need to ensure that vkWaitForFences() and
1827 * vkGetFenceStatus() return a valid result (VK_ERROR_DEVICE_LOST or
1828 * VK_SUCCESS) in a finite amount of time even if execbuf fails.
1830 fence
->permanent
.bo
.state
= ANV_BO_FENCE_STATE_SUBMITTED
;
1833 if (result
== VK_SUCCESS
&& need_out_fence
) {
1834 assert(!pdevice
->has_syncobj_wait
);
1835 int out_fence
= execbuf
.execbuf
.rsvd2
>> 32;
1836 for (uint32_t i
= 0; i
< num_out_semaphores
; i
++) {
1837 ANV_FROM_HANDLE(anv_semaphore
, semaphore
, out_semaphores
[i
]);
1838 /* Out fences can't have temporary state because that would imply
1839 * that we imported a sync file and are trying to signal it.
1841 assert(semaphore
->temporary
.type
== ANV_SEMAPHORE_TYPE_NONE
);
1842 struct anv_semaphore_impl
*impl
= &semaphore
->permanent
;
1844 if (impl
->type
== ANV_SEMAPHORE_TYPE_SYNC_FILE
) {
1845 assert(impl
->fd
== -1);
1846 impl
->fd
= dup(out_fence
);
1852 anv_execbuf_finish(&execbuf
);