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_clone(struct anv_reloc_list
*list
,
51 const VkAllocationCallbacks
*alloc
,
52 const struct anv_reloc_list
*other_list
)
55 list
->num_relocs
= other_list
->num_relocs
;
56 list
->array_length
= other_list
->array_length
;
59 list
->array_length
= 256;
63 vk_alloc(alloc
, list
->array_length
* sizeof(*list
->relocs
), 8,
64 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
66 if (list
->relocs
== NULL
)
67 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
70 vk_alloc(alloc
, list
->array_length
* sizeof(*list
->reloc_bos
), 8,
71 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
73 if (list
->reloc_bos
== NULL
) {
74 vk_free(alloc
, list
->relocs
);
75 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
78 list
->deps
= _mesa_set_create(NULL
, _mesa_hash_pointer
,
79 _mesa_key_pointer_equal
);
82 vk_free(alloc
, list
->relocs
);
83 vk_free(alloc
, list
->reloc_bos
);
84 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
88 memcpy(list
->relocs
, other_list
->relocs
,
89 list
->array_length
* sizeof(*list
->relocs
));
90 memcpy(list
->reloc_bos
, other_list
->reloc_bos
,
91 list
->array_length
* sizeof(*list
->reloc_bos
));
92 struct set_entry
*entry
;
93 set_foreach(other_list
->deps
, entry
) {
94 _mesa_set_add_pre_hashed(list
->deps
, entry
->hash
, entry
->key
);
102 anv_reloc_list_init(struct anv_reloc_list
*list
,
103 const VkAllocationCallbacks
*alloc
)
105 return anv_reloc_list_init_clone(list
, alloc
, NULL
);
109 anv_reloc_list_finish(struct anv_reloc_list
*list
,
110 const VkAllocationCallbacks
*alloc
)
112 vk_free(alloc
, list
->relocs
);
113 vk_free(alloc
, list
->reloc_bos
);
114 _mesa_set_destroy(list
->deps
, NULL
);
118 anv_reloc_list_grow(struct anv_reloc_list
*list
,
119 const VkAllocationCallbacks
*alloc
,
120 size_t num_additional_relocs
)
122 if (list
->num_relocs
+ num_additional_relocs
<= list
->array_length
)
125 size_t new_length
= list
->array_length
* 2;
126 while (new_length
< list
->num_relocs
+ num_additional_relocs
)
129 struct drm_i915_gem_relocation_entry
*new_relocs
=
130 vk_alloc(alloc
, new_length
* sizeof(*list
->relocs
), 8,
131 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
132 if (new_relocs
== NULL
)
133 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
135 struct anv_bo
**new_reloc_bos
=
136 vk_alloc(alloc
, new_length
* sizeof(*list
->reloc_bos
), 8,
137 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
138 if (new_reloc_bos
== NULL
) {
139 vk_free(alloc
, new_relocs
);
140 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
143 memcpy(new_relocs
, list
->relocs
, list
->num_relocs
* sizeof(*list
->relocs
));
144 memcpy(new_reloc_bos
, list
->reloc_bos
,
145 list
->num_relocs
* sizeof(*list
->reloc_bos
));
147 vk_free(alloc
, list
->relocs
);
148 vk_free(alloc
, list
->reloc_bos
);
150 list
->array_length
= new_length
;
151 list
->relocs
= new_relocs
;
152 list
->reloc_bos
= new_reloc_bos
;
158 anv_reloc_list_add(struct anv_reloc_list
*list
,
159 const VkAllocationCallbacks
*alloc
,
160 uint32_t offset
, struct anv_bo
*target_bo
, uint32_t delta
)
162 struct drm_i915_gem_relocation_entry
*entry
;
165 if (target_bo
->flags
& EXEC_OBJECT_PINNED
) {
166 _mesa_set_add(list
->deps
, target_bo
);
170 VkResult result
= anv_reloc_list_grow(list
, alloc
, 1);
171 if (result
!= VK_SUCCESS
)
174 /* XXX: Can we use I915_EXEC_HANDLE_LUT? */
175 index
= list
->num_relocs
++;
176 list
->reloc_bos
[index
] = target_bo
;
177 entry
= &list
->relocs
[index
];
178 entry
->target_handle
= target_bo
->gem_handle
;
179 entry
->delta
= delta
;
180 entry
->offset
= offset
;
181 entry
->presumed_offset
= target_bo
->offset
;
182 entry
->read_domains
= 0;
183 entry
->write_domain
= 0;
184 VG(VALGRIND_CHECK_MEM_IS_DEFINED(entry
, sizeof(*entry
)));
190 anv_reloc_list_append(struct anv_reloc_list
*list
,
191 const VkAllocationCallbacks
*alloc
,
192 struct anv_reloc_list
*other
, uint32_t offset
)
194 VkResult result
= anv_reloc_list_grow(list
, alloc
, other
->num_relocs
);
195 if (result
!= VK_SUCCESS
)
198 memcpy(&list
->relocs
[list
->num_relocs
], &other
->relocs
[0],
199 other
->num_relocs
* sizeof(other
->relocs
[0]));
200 memcpy(&list
->reloc_bos
[list
->num_relocs
], &other
->reloc_bos
[0],
201 other
->num_relocs
* sizeof(other
->reloc_bos
[0]));
203 for (uint32_t i
= 0; i
< other
->num_relocs
; i
++)
204 list
->relocs
[i
+ list
->num_relocs
].offset
+= offset
;
206 list
->num_relocs
+= other
->num_relocs
;
208 struct set_entry
*entry
;
209 set_foreach(other
->deps
, entry
) {
210 _mesa_set_add_pre_hashed(list
->deps
, entry
->hash
, entry
->key
);
216 /*-----------------------------------------------------------------------*
217 * Functions related to anv_batch
218 *-----------------------------------------------------------------------*/
221 anv_batch_emit_dwords(struct anv_batch
*batch
, int num_dwords
)
223 if (batch
->next
+ num_dwords
* 4 > batch
->end
) {
224 VkResult result
= batch
->extend_cb(batch
, batch
->user_data
);
225 if (result
!= VK_SUCCESS
) {
226 anv_batch_set_error(batch
, result
);
231 void *p
= batch
->next
;
233 batch
->next
+= num_dwords
* 4;
234 assert(batch
->next
<= batch
->end
);
240 anv_batch_emit_reloc(struct anv_batch
*batch
,
241 void *location
, struct anv_bo
*bo
, uint32_t delta
)
243 VkResult result
= anv_reloc_list_add(batch
->relocs
, batch
->alloc
,
244 location
- batch
->start
, bo
, delta
);
245 if (result
!= VK_SUCCESS
) {
246 anv_batch_set_error(batch
, result
);
250 return bo
->offset
+ delta
;
254 anv_batch_emit_batch(struct anv_batch
*batch
, struct anv_batch
*other
)
256 uint32_t size
, offset
;
258 size
= other
->next
- other
->start
;
259 assert(size
% 4 == 0);
261 if (batch
->next
+ size
> batch
->end
) {
262 VkResult result
= batch
->extend_cb(batch
, batch
->user_data
);
263 if (result
!= VK_SUCCESS
) {
264 anv_batch_set_error(batch
, result
);
269 assert(batch
->next
+ size
<= batch
->end
);
271 VG(VALGRIND_CHECK_MEM_IS_DEFINED(other
->start
, size
));
272 memcpy(batch
->next
, other
->start
, size
);
274 offset
= batch
->next
- batch
->start
;
275 VkResult result
= anv_reloc_list_append(batch
->relocs
, batch
->alloc
,
276 other
->relocs
, offset
);
277 if (result
!= VK_SUCCESS
) {
278 anv_batch_set_error(batch
, result
);
285 /*-----------------------------------------------------------------------*
286 * Functions related to anv_batch_bo
287 *-----------------------------------------------------------------------*/
290 anv_batch_bo_create(struct anv_cmd_buffer
*cmd_buffer
,
291 struct anv_batch_bo
**bbo_out
)
295 struct anv_batch_bo
*bbo
= vk_alloc(&cmd_buffer
->pool
->alloc
, sizeof(*bbo
),
296 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
298 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
300 result
= anv_bo_pool_alloc(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
,
301 ANV_CMD_BUFFER_BATCH_SIZE
);
302 if (result
!= VK_SUCCESS
)
305 result
= anv_reloc_list_init(&bbo
->relocs
, &cmd_buffer
->pool
->alloc
);
306 if (result
!= VK_SUCCESS
)
314 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
316 vk_free(&cmd_buffer
->pool
->alloc
, bbo
);
322 anv_batch_bo_clone(struct anv_cmd_buffer
*cmd_buffer
,
323 const struct anv_batch_bo
*other_bbo
,
324 struct anv_batch_bo
**bbo_out
)
328 struct anv_batch_bo
*bbo
= vk_alloc(&cmd_buffer
->pool
->alloc
, sizeof(*bbo
),
329 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
331 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
333 result
= anv_bo_pool_alloc(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
,
335 if (result
!= VK_SUCCESS
)
338 result
= anv_reloc_list_init_clone(&bbo
->relocs
, &cmd_buffer
->pool
->alloc
,
340 if (result
!= VK_SUCCESS
)
343 bbo
->length
= other_bbo
->length
;
344 memcpy(bbo
->bo
.map
, other_bbo
->bo
.map
, other_bbo
->length
);
351 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
353 vk_free(&cmd_buffer
->pool
->alloc
, bbo
);
359 anv_batch_bo_start(struct anv_batch_bo
*bbo
, struct anv_batch
*batch
,
360 size_t batch_padding
)
362 batch
->next
= batch
->start
= bbo
->bo
.map
;
363 batch
->end
= bbo
->bo
.map
+ bbo
->bo
.size
- batch_padding
;
364 batch
->relocs
= &bbo
->relocs
;
365 bbo
->relocs
.num_relocs
= 0;
366 _mesa_set_clear(bbo
->relocs
.deps
, NULL
);
370 anv_batch_bo_continue(struct anv_batch_bo
*bbo
, struct anv_batch
*batch
,
371 size_t batch_padding
)
373 batch
->start
= bbo
->bo
.map
;
374 batch
->next
= bbo
->bo
.map
+ bbo
->length
;
375 batch
->end
= bbo
->bo
.map
+ bbo
->bo
.size
- batch_padding
;
376 batch
->relocs
= &bbo
->relocs
;
380 anv_batch_bo_finish(struct anv_batch_bo
*bbo
, struct anv_batch
*batch
)
382 assert(batch
->start
== bbo
->bo
.map
);
383 bbo
->length
= batch
->next
- batch
->start
;
384 VG(VALGRIND_CHECK_MEM_IS_DEFINED(batch
->start
, bbo
->length
));
388 anv_batch_bo_grow(struct anv_cmd_buffer
*cmd_buffer
, struct anv_batch_bo
*bbo
,
389 struct anv_batch
*batch
, size_t aditional
,
390 size_t batch_padding
)
392 assert(batch
->start
== bbo
->bo
.map
);
393 bbo
->length
= batch
->next
- batch
->start
;
395 size_t new_size
= bbo
->bo
.size
;
396 while (new_size
<= bbo
->length
+ aditional
+ batch_padding
)
399 if (new_size
== bbo
->bo
.size
)
402 struct anv_bo new_bo
;
403 VkResult result
= anv_bo_pool_alloc(&cmd_buffer
->device
->batch_bo_pool
,
405 if (result
!= VK_SUCCESS
)
408 memcpy(new_bo
.map
, bbo
->bo
.map
, bbo
->length
);
410 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
413 anv_batch_bo_continue(bbo
, batch
, batch_padding
);
419 anv_batch_bo_destroy(struct anv_batch_bo
*bbo
,
420 struct anv_cmd_buffer
*cmd_buffer
)
422 anv_reloc_list_finish(&bbo
->relocs
, &cmd_buffer
->pool
->alloc
);
423 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
424 vk_free(&cmd_buffer
->pool
->alloc
, bbo
);
428 anv_batch_bo_list_clone(const struct list_head
*list
,
429 struct anv_cmd_buffer
*cmd_buffer
,
430 struct list_head
*new_list
)
432 VkResult result
= VK_SUCCESS
;
434 list_inithead(new_list
);
436 struct anv_batch_bo
*prev_bbo
= NULL
;
437 list_for_each_entry(struct anv_batch_bo
, bbo
, list
, link
) {
438 struct anv_batch_bo
*new_bbo
= NULL
;
439 result
= anv_batch_bo_clone(cmd_buffer
, bbo
, &new_bbo
);
440 if (result
!= VK_SUCCESS
)
442 list_addtail(&new_bbo
->link
, new_list
);
445 /* As we clone this list of batch_bo's, they chain one to the
446 * other using MI_BATCH_BUFFER_START commands. We need to fix up
447 * those relocations as we go. Fortunately, this is pretty easy
448 * as it will always be the last relocation in the list.
450 uint32_t last_idx
= prev_bbo
->relocs
.num_relocs
- 1;
451 assert(prev_bbo
->relocs
.reloc_bos
[last_idx
] == &bbo
->bo
);
452 prev_bbo
->relocs
.reloc_bos
[last_idx
] = &new_bbo
->bo
;
458 if (result
!= VK_SUCCESS
) {
459 list_for_each_entry_safe(struct anv_batch_bo
, bbo
, new_list
, link
)
460 anv_batch_bo_destroy(bbo
, cmd_buffer
);
466 /*-----------------------------------------------------------------------*
467 * Functions related to anv_batch_bo
468 *-----------------------------------------------------------------------*/
470 static struct anv_batch_bo
*
471 anv_cmd_buffer_current_batch_bo(struct anv_cmd_buffer
*cmd_buffer
)
473 return LIST_ENTRY(struct anv_batch_bo
, cmd_buffer
->batch_bos
.prev
, link
);
477 anv_cmd_buffer_surface_base_address(struct anv_cmd_buffer
*cmd_buffer
)
479 struct anv_state
*bt_block
= u_vector_head(&cmd_buffer
->bt_block_states
);
480 return (struct anv_address
) {
481 .bo
= &anv_binding_table_pool(cmd_buffer
->device
)->block_pool
.bo
,
482 .offset
= bt_block
->offset
,
487 emit_batch_buffer_start(struct anv_cmd_buffer
*cmd_buffer
,
488 struct anv_bo
*bo
, uint32_t offset
)
490 /* In gen8+ the address field grew to two dwords to accomodate 48 bit
491 * offsets. The high 16 bits are in the last dword, so we can use the gen8
492 * version in either case, as long as we set the instruction length in the
493 * header accordingly. This means that we always emit three dwords here
494 * and all the padding and adjustment we do in this file works for all
498 #define GEN7_MI_BATCH_BUFFER_START_length 2
499 #define GEN7_MI_BATCH_BUFFER_START_length_bias 2
501 const uint32_t gen7_length
=
502 GEN7_MI_BATCH_BUFFER_START_length
- GEN7_MI_BATCH_BUFFER_START_length_bias
;
503 const uint32_t gen8_length
=
504 GEN8_MI_BATCH_BUFFER_START_length
- GEN8_MI_BATCH_BUFFER_START_length_bias
;
506 anv_batch_emit(&cmd_buffer
->batch
, GEN8_MI_BATCH_BUFFER_START
, bbs
) {
507 bbs
.DWordLength
= cmd_buffer
->device
->info
.gen
< 8 ?
508 gen7_length
: gen8_length
;
509 bbs
._2ndLevelBatchBuffer
= _1stlevelbatch
;
510 bbs
.AddressSpaceIndicator
= ASI_PPGTT
;
511 bbs
.BatchBufferStartAddress
= (struct anv_address
) { bo
, offset
};
516 cmd_buffer_chain_to_batch_bo(struct anv_cmd_buffer
*cmd_buffer
,
517 struct anv_batch_bo
*bbo
)
519 struct anv_batch
*batch
= &cmd_buffer
->batch
;
520 struct anv_batch_bo
*current_bbo
=
521 anv_cmd_buffer_current_batch_bo(cmd_buffer
);
523 /* We set the end of the batch a little short so we would be sure we
524 * have room for the chaining command. Since we're about to emit the
525 * chaining command, let's set it back where it should go.
527 batch
->end
+= GEN8_MI_BATCH_BUFFER_START_length
* 4;
528 assert(batch
->end
== current_bbo
->bo
.map
+ current_bbo
->bo
.size
);
530 emit_batch_buffer_start(cmd_buffer
, &bbo
->bo
, 0);
532 anv_batch_bo_finish(current_bbo
, batch
);
536 anv_cmd_buffer_chain_batch(struct anv_batch
*batch
, void *_data
)
538 struct anv_cmd_buffer
*cmd_buffer
= _data
;
539 struct anv_batch_bo
*new_bbo
;
541 VkResult result
= anv_batch_bo_create(cmd_buffer
, &new_bbo
);
542 if (result
!= VK_SUCCESS
)
545 struct anv_batch_bo
**seen_bbo
= u_vector_add(&cmd_buffer
->seen_bbos
);
546 if (seen_bbo
== NULL
) {
547 anv_batch_bo_destroy(new_bbo
, cmd_buffer
);
548 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
552 cmd_buffer_chain_to_batch_bo(cmd_buffer
, new_bbo
);
554 list_addtail(&new_bbo
->link
, &cmd_buffer
->batch_bos
);
556 anv_batch_bo_start(new_bbo
, batch
, GEN8_MI_BATCH_BUFFER_START_length
* 4);
562 anv_cmd_buffer_grow_batch(struct anv_batch
*batch
, void *_data
)
564 struct anv_cmd_buffer
*cmd_buffer
= _data
;
565 struct anv_batch_bo
*bbo
= anv_cmd_buffer_current_batch_bo(cmd_buffer
);
567 anv_batch_bo_grow(cmd_buffer
, bbo
, &cmd_buffer
->batch
, 4096,
568 GEN8_MI_BATCH_BUFFER_START_length
* 4);
573 /** Allocate a binding table
575 * This function allocates a binding table. This is a bit more complicated
576 * than one would think due to a combination of Vulkan driver design and some
577 * unfortunate hardware restrictions.
579 * The 3DSTATE_BINDING_TABLE_POINTERS_* packets only have a 16-bit field for
580 * the binding table pointer which means that all binding tables need to live
581 * in the bottom 64k of surface state base address. The way the GL driver has
582 * classically dealt with this restriction is to emit all surface states
583 * on-the-fly into the batch and have a batch buffer smaller than 64k. This
584 * isn't really an option in Vulkan for a couple of reasons:
586 * 1) In Vulkan, we have growing (or chaining) batches so surface states have
587 * to live in their own buffer and we have to be able to re-emit
588 * STATE_BASE_ADDRESS as needed which requires a full pipeline stall. In
589 * order to avoid emitting STATE_BASE_ADDRESS any more often than needed
590 * (it's not that hard to hit 64k of just binding tables), we allocate
591 * surface state objects up-front when VkImageView is created. In order
592 * for this to work, surface state objects need to be allocated from a
595 * 2) We tried to design the surface state system in such a way that it's
596 * already ready for bindless texturing. The way bindless texturing works
597 * on our hardware is that you have a big pool of surface state objects
598 * (with its own state base address) and the bindless handles are simply
599 * offsets into that pool. With the architecture we chose, we already
600 * have that pool and it's exactly the same pool that we use for regular
601 * surface states so we should already be ready for bindless.
603 * 3) For render targets, we need to be able to fill out the surface states
604 * later in vkBeginRenderPass so that we can assign clear colors
605 * correctly. One way to do this would be to just create the surface
606 * state data and then repeatedly copy it into the surface state BO every
607 * time we have to re-emit STATE_BASE_ADDRESS. While this works, it's
608 * rather annoying and just being able to allocate them up-front and
609 * re-use them for the entire render pass.
611 * While none of these are technically blockers for emitting state on the fly
612 * like we do in GL, the ability to have a single surface state pool is
613 * simplifies things greatly. Unfortunately, it comes at a cost...
615 * Because of the 64k limitation of 3DSTATE_BINDING_TABLE_POINTERS_*, we can't
616 * place the binding tables just anywhere in surface state base address.
617 * Because 64k isn't a whole lot of space, we can't simply restrict the
618 * surface state buffer to 64k, we have to be more clever. The solution we've
619 * chosen is to have a block pool with a maximum size of 2G that starts at
620 * zero and grows in both directions. All surface states are allocated from
621 * the top of the pool (positive offsets) and we allocate blocks (< 64k) of
622 * binding tables from the bottom of the pool (negative offsets). Every time
623 * we allocate a new binding table block, we set surface state base address to
624 * point to the bottom of the binding table block. This way all of the
625 * binding tables in the block are in the bottom 64k of surface state base
626 * address. When we fill out the binding table, we add the distance between
627 * the bottom of our binding table block and zero of the block pool to the
628 * surface state offsets so that they are correct relative to out new surface
629 * state base address at the bottom of the binding table block.
631 * \see adjust_relocations_from_block_pool()
632 * \see adjust_relocations_too_block_pool()
634 * \param[in] entries The number of surface state entries the binding
635 * table should be able to hold.
637 * \param[out] state_offset The offset surface surface state base address
638 * where the surface states live. This must be
639 * added to the surface state offset when it is
640 * written into the binding table entry.
642 * \return An anv_state representing the binding table
645 anv_cmd_buffer_alloc_binding_table(struct anv_cmd_buffer
*cmd_buffer
,
646 uint32_t entries
, uint32_t *state_offset
)
648 struct anv_device
*device
= cmd_buffer
->device
;
649 struct anv_state_pool
*state_pool
= &device
->surface_state_pool
;
650 struct anv_state
*bt_block
= u_vector_head(&cmd_buffer
->bt_block_states
);
651 struct anv_state state
;
653 state
.alloc_size
= align_u32(entries
* 4, 32);
655 if (cmd_buffer
->bt_next
+ state
.alloc_size
> state_pool
->block_size
)
656 return (struct anv_state
) { 0 };
658 state
.offset
= cmd_buffer
->bt_next
;
659 state
.map
= anv_binding_table_pool(device
)->block_pool
.map
+
660 bt_block
->offset
+ state
.offset
;
662 cmd_buffer
->bt_next
+= state
.alloc_size
;
664 if (device
->instance
->physicalDevice
.use_softpin
) {
665 assert(bt_block
->offset
>= 0);
666 *state_offset
= device
->surface_state_pool
.block_pool
.start_address
-
667 device
->binding_table_pool
.block_pool
.start_address
- bt_block
->offset
;
669 assert(bt_block
->offset
< 0);
670 *state_offset
= -bt_block
->offset
;
677 anv_cmd_buffer_alloc_surface_state(struct anv_cmd_buffer
*cmd_buffer
)
679 struct isl_device
*isl_dev
= &cmd_buffer
->device
->isl_dev
;
680 return anv_state_stream_alloc(&cmd_buffer
->surface_state_stream
,
681 isl_dev
->ss
.size
, isl_dev
->ss
.align
);
685 anv_cmd_buffer_alloc_dynamic_state(struct anv_cmd_buffer
*cmd_buffer
,
686 uint32_t size
, uint32_t alignment
)
688 return anv_state_stream_alloc(&cmd_buffer
->dynamic_state_stream
,
693 anv_cmd_buffer_new_binding_table_block(struct anv_cmd_buffer
*cmd_buffer
)
695 struct anv_state
*bt_block
= u_vector_add(&cmd_buffer
->bt_block_states
);
696 if (bt_block
== NULL
) {
697 anv_batch_set_error(&cmd_buffer
->batch
, VK_ERROR_OUT_OF_HOST_MEMORY
);
698 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
701 *bt_block
= anv_binding_table_pool_alloc(cmd_buffer
->device
);
702 cmd_buffer
->bt_next
= 0;
708 anv_cmd_buffer_init_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
)
710 struct anv_batch_bo
*batch_bo
;
713 list_inithead(&cmd_buffer
->batch_bos
);
715 result
= anv_batch_bo_create(cmd_buffer
, &batch_bo
);
716 if (result
!= VK_SUCCESS
)
719 list_addtail(&batch_bo
->link
, &cmd_buffer
->batch_bos
);
721 cmd_buffer
->batch
.alloc
= &cmd_buffer
->pool
->alloc
;
722 cmd_buffer
->batch
.user_data
= cmd_buffer
;
724 if (cmd_buffer
->device
->can_chain_batches
) {
725 cmd_buffer
->batch
.extend_cb
= anv_cmd_buffer_chain_batch
;
727 cmd_buffer
->batch
.extend_cb
= anv_cmd_buffer_grow_batch
;
730 anv_batch_bo_start(batch_bo
, &cmd_buffer
->batch
,
731 GEN8_MI_BATCH_BUFFER_START_length
* 4);
733 int success
= u_vector_init(&cmd_buffer
->seen_bbos
,
734 sizeof(struct anv_bo
*),
735 8 * sizeof(struct anv_bo
*));
739 *(struct anv_batch_bo
**)u_vector_add(&cmd_buffer
->seen_bbos
) = batch_bo
;
741 /* u_vector requires power-of-two size elements */
742 unsigned pow2_state_size
= util_next_power_of_two(sizeof(struct anv_state
));
743 success
= u_vector_init(&cmd_buffer
->bt_block_states
,
744 pow2_state_size
, 8 * pow2_state_size
);
748 result
= anv_reloc_list_init(&cmd_buffer
->surface_relocs
,
749 &cmd_buffer
->pool
->alloc
);
750 if (result
!= VK_SUCCESS
)
752 cmd_buffer
->last_ss_pool_center
= 0;
754 result
= anv_cmd_buffer_new_binding_table_block(cmd_buffer
);
755 if (result
!= VK_SUCCESS
)
761 u_vector_finish(&cmd_buffer
->bt_block_states
);
763 u_vector_finish(&cmd_buffer
->seen_bbos
);
765 anv_batch_bo_destroy(batch_bo
, cmd_buffer
);
771 anv_cmd_buffer_fini_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
)
773 struct anv_state
*bt_block
;
774 u_vector_foreach(bt_block
, &cmd_buffer
->bt_block_states
)
775 anv_binding_table_pool_free(cmd_buffer
->device
, *bt_block
);
776 u_vector_finish(&cmd_buffer
->bt_block_states
);
778 anv_reloc_list_finish(&cmd_buffer
->surface_relocs
, &cmd_buffer
->pool
->alloc
);
780 u_vector_finish(&cmd_buffer
->seen_bbos
);
782 /* Destroy all of the batch buffers */
783 list_for_each_entry_safe(struct anv_batch_bo
, bbo
,
784 &cmd_buffer
->batch_bos
, link
) {
785 anv_batch_bo_destroy(bbo
, cmd_buffer
);
790 anv_cmd_buffer_reset_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
)
792 /* Delete all but the first batch bo */
793 assert(!list_empty(&cmd_buffer
->batch_bos
));
794 while (cmd_buffer
->batch_bos
.next
!= cmd_buffer
->batch_bos
.prev
) {
795 struct anv_batch_bo
*bbo
= anv_cmd_buffer_current_batch_bo(cmd_buffer
);
796 list_del(&bbo
->link
);
797 anv_batch_bo_destroy(bbo
, cmd_buffer
);
799 assert(!list_empty(&cmd_buffer
->batch_bos
));
801 anv_batch_bo_start(anv_cmd_buffer_current_batch_bo(cmd_buffer
),
803 GEN8_MI_BATCH_BUFFER_START_length
* 4);
805 while (u_vector_length(&cmd_buffer
->bt_block_states
) > 1) {
806 struct anv_state
*bt_block
= u_vector_remove(&cmd_buffer
->bt_block_states
);
807 anv_binding_table_pool_free(cmd_buffer
->device
, *bt_block
);
809 assert(u_vector_length(&cmd_buffer
->bt_block_states
) == 1);
810 cmd_buffer
->bt_next
= 0;
812 cmd_buffer
->surface_relocs
.num_relocs
= 0;
813 _mesa_set_clear(cmd_buffer
->surface_relocs
.deps
, NULL
);
814 cmd_buffer
->last_ss_pool_center
= 0;
816 /* Reset the list of seen buffers */
817 cmd_buffer
->seen_bbos
.head
= 0;
818 cmd_buffer
->seen_bbos
.tail
= 0;
820 *(struct anv_batch_bo
**)u_vector_add(&cmd_buffer
->seen_bbos
) =
821 anv_cmd_buffer_current_batch_bo(cmd_buffer
);
825 anv_cmd_buffer_end_batch_buffer(struct anv_cmd_buffer
*cmd_buffer
)
827 struct anv_batch_bo
*batch_bo
= anv_cmd_buffer_current_batch_bo(cmd_buffer
);
829 if (cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
) {
830 /* When we start a batch buffer, we subtract a certain amount of
831 * padding from the end to ensure that we always have room to emit a
832 * BATCH_BUFFER_START to chain to the next BO. We need to remove
833 * that padding before we end the batch; otherwise, we may end up
834 * with our BATCH_BUFFER_END in another BO.
836 cmd_buffer
->batch
.end
+= GEN8_MI_BATCH_BUFFER_START_length
* 4;
837 assert(cmd_buffer
->batch
.end
== batch_bo
->bo
.map
+ batch_bo
->bo
.size
);
839 anv_batch_emit(&cmd_buffer
->batch
, GEN8_MI_BATCH_BUFFER_END
, bbe
);
841 /* Round batch up to an even number of dwords. */
842 if ((cmd_buffer
->batch
.next
- cmd_buffer
->batch
.start
) & 4)
843 anv_batch_emit(&cmd_buffer
->batch
, GEN8_MI_NOOP
, noop
);
845 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_PRIMARY
;
847 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_SECONDARY
);
848 /* If this is a secondary command buffer, we need to determine the
849 * mode in which it will be executed with vkExecuteCommands. We
850 * determine this statically here so that this stays in sync with the
851 * actual ExecuteCommands implementation.
853 const uint32_t length
= cmd_buffer
->batch
.next
- cmd_buffer
->batch
.start
;
854 if (!cmd_buffer
->device
->can_chain_batches
) {
855 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_GROW_AND_EMIT
;
856 } else if ((cmd_buffer
->batch_bos
.next
== cmd_buffer
->batch_bos
.prev
) &&
857 (length
< ANV_CMD_BUFFER_BATCH_SIZE
/ 2)) {
858 /* If the secondary has exactly one batch buffer in its list *and*
859 * that batch buffer is less than half of the maximum size, we're
860 * probably better of simply copying it into our batch.
862 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_EMIT
;
863 } else if (!(cmd_buffer
->usage_flags
&
864 VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT
)) {
865 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_CHAIN
;
867 /* When we chain, we need to add an MI_BATCH_BUFFER_START command
868 * with its relocation. In order to handle this we'll increment here
869 * so we can unconditionally decrement right before adding the
870 * MI_BATCH_BUFFER_START command.
872 batch_bo
->relocs
.num_relocs
++;
873 cmd_buffer
->batch
.next
+= GEN8_MI_BATCH_BUFFER_START_length
* 4;
875 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN
;
879 anv_batch_bo_finish(batch_bo
, &cmd_buffer
->batch
);
883 anv_cmd_buffer_add_seen_bbos(struct anv_cmd_buffer
*cmd_buffer
,
884 struct list_head
*list
)
886 list_for_each_entry(struct anv_batch_bo
, bbo
, list
, link
) {
887 struct anv_batch_bo
**bbo_ptr
= u_vector_add(&cmd_buffer
->seen_bbos
);
889 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
898 anv_cmd_buffer_add_secondary(struct anv_cmd_buffer
*primary
,
899 struct anv_cmd_buffer
*secondary
)
901 switch (secondary
->exec_mode
) {
902 case ANV_CMD_BUFFER_EXEC_MODE_EMIT
:
903 anv_batch_emit_batch(&primary
->batch
, &secondary
->batch
);
905 case ANV_CMD_BUFFER_EXEC_MODE_GROW_AND_EMIT
: {
906 struct anv_batch_bo
*bbo
= anv_cmd_buffer_current_batch_bo(primary
);
907 unsigned length
= secondary
->batch
.end
- secondary
->batch
.start
;
908 anv_batch_bo_grow(primary
, bbo
, &primary
->batch
, length
,
909 GEN8_MI_BATCH_BUFFER_START_length
* 4);
910 anv_batch_emit_batch(&primary
->batch
, &secondary
->batch
);
913 case ANV_CMD_BUFFER_EXEC_MODE_CHAIN
: {
914 struct anv_batch_bo
*first_bbo
=
915 list_first_entry(&secondary
->batch_bos
, struct anv_batch_bo
, link
);
916 struct anv_batch_bo
*last_bbo
=
917 list_last_entry(&secondary
->batch_bos
, struct anv_batch_bo
, link
);
919 emit_batch_buffer_start(primary
, &first_bbo
->bo
, 0);
921 struct anv_batch_bo
*this_bbo
= anv_cmd_buffer_current_batch_bo(primary
);
922 assert(primary
->batch
.start
== this_bbo
->bo
.map
);
923 uint32_t offset
= primary
->batch
.next
- primary
->batch
.start
;
924 const uint32_t inst_size
= GEN8_MI_BATCH_BUFFER_START_length
* 4;
926 /* Roll back the previous MI_BATCH_BUFFER_START and its relocation so we
927 * can emit a new command and relocation for the current splice. In
928 * order to handle the initial-use case, we incremented next and
929 * num_relocs in end_batch_buffer() so we can alyways just subtract
932 last_bbo
->relocs
.num_relocs
--;
933 secondary
->batch
.next
-= inst_size
;
934 emit_batch_buffer_start(secondary
, &this_bbo
->bo
, offset
);
935 anv_cmd_buffer_add_seen_bbos(primary
, &secondary
->batch_bos
);
937 /* After patching up the secondary buffer, we need to clflush the
938 * modified instruction in case we're on a !llc platform. We use a
939 * little loop to handle the case where the instruction crosses a cache
942 if (!primary
->device
->info
.has_llc
) {
943 void *inst
= secondary
->batch
.next
- inst_size
;
944 void *p
= (void *) (((uintptr_t) inst
) & ~CACHELINE_MASK
);
945 __builtin_ia32_mfence();
946 while (p
< secondary
->batch
.next
) {
947 __builtin_ia32_clflush(p
);
953 case ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN
: {
954 struct list_head copy_list
;
955 VkResult result
= anv_batch_bo_list_clone(&secondary
->batch_bos
,
958 if (result
!= VK_SUCCESS
)
961 anv_cmd_buffer_add_seen_bbos(primary
, ©_list
);
963 struct anv_batch_bo
*first_bbo
=
964 list_first_entry(©_list
, struct anv_batch_bo
, link
);
965 struct anv_batch_bo
*last_bbo
=
966 list_last_entry(©_list
, struct anv_batch_bo
, link
);
968 cmd_buffer_chain_to_batch_bo(primary
, first_bbo
);
970 list_splicetail(©_list
, &primary
->batch_bos
);
972 anv_batch_bo_continue(last_bbo
, &primary
->batch
,
973 GEN8_MI_BATCH_BUFFER_START_length
* 4);
977 assert(!"Invalid execution mode");
980 anv_reloc_list_append(&primary
->surface_relocs
, &primary
->pool
->alloc
,
981 &secondary
->surface_relocs
, 0);
985 struct drm_i915_gem_execbuffer2 execbuf
;
987 struct drm_i915_gem_exec_object2
* objects
;
989 struct anv_bo
** bos
;
991 /* Allocated length of the 'objects' and 'bos' arrays */
992 uint32_t array_length
;
994 uint32_t fence_count
;
995 uint32_t fence_array_length
;
996 struct drm_i915_gem_exec_fence
* fences
;
997 struct anv_syncobj
** syncobjs
;
1001 anv_execbuf_init(struct anv_execbuf
*exec
)
1003 memset(exec
, 0, sizeof(*exec
));
1007 anv_execbuf_finish(struct anv_execbuf
*exec
,
1008 const VkAllocationCallbacks
*alloc
)
1010 vk_free(alloc
, exec
->objects
);
1011 vk_free(alloc
, exec
->bos
);
1012 vk_free(alloc
, exec
->fences
);
1013 vk_free(alloc
, exec
->syncobjs
);
1017 _compare_bo_handles(const void *_bo1
, const void *_bo2
)
1019 struct anv_bo
* const *bo1
= _bo1
;
1020 struct anv_bo
* const *bo2
= _bo2
;
1022 return (*bo1
)->gem_handle
- (*bo2
)->gem_handle
;
1026 anv_execbuf_add_bo(struct anv_execbuf
*exec
,
1028 struct anv_reloc_list
*relocs
,
1029 uint32_t extra_flags
,
1030 const VkAllocationCallbacks
*alloc
)
1032 struct drm_i915_gem_exec_object2
*obj
= NULL
;
1034 if (bo
->index
< exec
->bo_count
&& exec
->bos
[bo
->index
] == bo
)
1035 obj
= &exec
->objects
[bo
->index
];
1038 /* We've never seen this one before. Add it to the list and assign
1039 * an id that we can use later.
1041 if (exec
->bo_count
>= exec
->array_length
) {
1042 uint32_t new_len
= exec
->objects
? exec
->array_length
* 2 : 64;
1044 struct drm_i915_gem_exec_object2
*new_objects
=
1045 vk_alloc(alloc
, new_len
* sizeof(*new_objects
),
1046 8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND
);
1047 if (new_objects
== NULL
)
1048 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1050 struct anv_bo
**new_bos
=
1051 vk_alloc(alloc
, new_len
* sizeof(*new_bos
),
1052 8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND
);
1053 if (new_bos
== NULL
) {
1054 vk_free(alloc
, new_objects
);
1055 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1058 if (exec
->objects
) {
1059 memcpy(new_objects
, exec
->objects
,
1060 exec
->bo_count
* sizeof(*new_objects
));
1061 memcpy(new_bos
, exec
->bos
,
1062 exec
->bo_count
* sizeof(*new_bos
));
1065 vk_free(alloc
, exec
->objects
);
1066 vk_free(alloc
, exec
->bos
);
1068 exec
->objects
= new_objects
;
1069 exec
->bos
= new_bos
;
1070 exec
->array_length
= new_len
;
1073 assert(exec
->bo_count
< exec
->array_length
);
1075 bo
->index
= exec
->bo_count
++;
1076 obj
= &exec
->objects
[bo
->index
];
1077 exec
->bos
[bo
->index
] = bo
;
1079 obj
->handle
= bo
->gem_handle
;
1080 obj
->relocation_count
= 0;
1081 obj
->relocs_ptr
= 0;
1083 obj
->offset
= bo
->offset
;
1084 obj
->flags
= bo
->flags
| extra_flags
;
1089 if (relocs
!= NULL
&& obj
->relocation_count
== 0) {
1090 /* This is the first time we've ever seen a list of relocations for
1091 * this BO. Go ahead and set the relocations and then walk the list
1092 * of relocations and add them all.
1094 obj
->relocation_count
= relocs
->num_relocs
;
1095 obj
->relocs_ptr
= (uintptr_t) relocs
->relocs
;
1097 for (size_t i
= 0; i
< relocs
->num_relocs
; i
++) {
1100 /* A quick sanity check on relocations */
1101 assert(relocs
->relocs
[i
].offset
< bo
->size
);
1102 result
= anv_execbuf_add_bo(exec
, relocs
->reloc_bos
[i
], NULL
,
1103 extra_flags
, alloc
);
1105 if (result
!= VK_SUCCESS
)
1109 const uint32_t entries
= relocs
->deps
->entries
;
1110 struct anv_bo
**bos
=
1111 vk_alloc(alloc
, entries
* sizeof(*bos
),
1112 8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND
);
1114 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1116 struct set_entry
*entry
;
1117 struct anv_bo
**bo
= bos
;
1118 set_foreach(relocs
->deps
, entry
) {
1119 *bo
++ = (void *)entry
->key
;
1122 qsort(bos
, entries
, sizeof(struct anv_bo
*), _compare_bo_handles
);
1124 VkResult result
= VK_SUCCESS
;
1125 for (bo
= bos
; bo
< bos
+ entries
; bo
++) {
1126 result
= anv_execbuf_add_bo(exec
, *bo
, NULL
, extra_flags
, alloc
);
1127 if (result
!= VK_SUCCESS
)
1131 vk_free(alloc
, bos
);
1133 if (result
!= VK_SUCCESS
)
1141 anv_execbuf_add_syncobj(struct anv_execbuf
*exec
,
1142 uint32_t handle
, uint32_t flags
,
1143 const VkAllocationCallbacks
*alloc
)
1147 if (exec
->fence_count
>= exec
->fence_array_length
) {
1148 uint32_t new_len
= MAX2(exec
->fence_array_length
* 2, 64);
1150 exec
->fences
= vk_realloc(alloc
, exec
->fences
,
1151 new_len
* sizeof(*exec
->fences
),
1152 8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND
);
1153 if (exec
->fences
== NULL
)
1154 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1156 exec
->fence_array_length
= new_len
;
1159 exec
->fences
[exec
->fence_count
] = (struct drm_i915_gem_exec_fence
) {
1164 exec
->fence_count
++;
1170 anv_cmd_buffer_process_relocs(struct anv_cmd_buffer
*cmd_buffer
,
1171 struct anv_reloc_list
*list
)
1173 for (size_t i
= 0; i
< list
->num_relocs
; i
++)
1174 list
->relocs
[i
].target_handle
= list
->reloc_bos
[i
]->index
;
1178 adjust_relocations_from_state_pool(struct anv_state_pool
*pool
,
1179 struct anv_reloc_list
*relocs
,
1180 uint32_t last_pool_center_bo_offset
)
1182 assert(last_pool_center_bo_offset
<= pool
->block_pool
.center_bo_offset
);
1183 uint32_t delta
= pool
->block_pool
.center_bo_offset
- last_pool_center_bo_offset
;
1185 for (size_t i
= 0; i
< relocs
->num_relocs
; i
++) {
1186 /* All of the relocations from this block pool to other BO's should
1187 * have been emitted relative to the surface block pool center. We
1188 * need to add the center offset to make them relative to the
1189 * beginning of the actual GEM bo.
1191 relocs
->relocs
[i
].offset
+= delta
;
1196 adjust_relocations_to_state_pool(struct anv_state_pool
*pool
,
1197 struct anv_bo
*from_bo
,
1198 struct anv_reloc_list
*relocs
,
1199 uint32_t last_pool_center_bo_offset
)
1201 assert(last_pool_center_bo_offset
<= pool
->block_pool
.center_bo_offset
);
1202 uint32_t delta
= pool
->block_pool
.center_bo_offset
- last_pool_center_bo_offset
;
1204 /* When we initially emit relocations into a block pool, we don't
1205 * actually know what the final center_bo_offset will be so we just emit
1206 * it as if center_bo_offset == 0. Now that we know what the center
1207 * offset is, we need to walk the list of relocations and adjust any
1208 * relocations that point to the pool bo with the correct offset.
1210 for (size_t i
= 0; i
< relocs
->num_relocs
; i
++) {
1211 if (relocs
->reloc_bos
[i
] == &pool
->block_pool
.bo
) {
1212 /* Adjust the delta value in the relocation to correctly
1213 * correspond to the new delta. Initially, this value may have
1214 * been negative (if treated as unsigned), but we trust in
1215 * uint32_t roll-over to fix that for us at this point.
1217 relocs
->relocs
[i
].delta
+= delta
;
1219 /* Since the delta has changed, we need to update the actual
1220 * relocated value with the new presumed value. This function
1221 * should only be called on batch buffers, so we know it isn't in
1222 * use by the GPU at the moment.
1224 assert(relocs
->relocs
[i
].offset
< from_bo
->size
);
1225 write_reloc(pool
->block_pool
.device
,
1226 from_bo
->map
+ relocs
->relocs
[i
].offset
,
1227 relocs
->relocs
[i
].presumed_offset
+
1228 relocs
->relocs
[i
].delta
, false);
1234 anv_reloc_list_apply(struct anv_device
*device
,
1235 struct anv_reloc_list
*list
,
1237 bool always_relocate
)
1239 for (size_t i
= 0; i
< list
->num_relocs
; i
++) {
1240 struct anv_bo
*target_bo
= list
->reloc_bos
[i
];
1241 if (list
->relocs
[i
].presumed_offset
== target_bo
->offset
&&
1245 void *p
= bo
->map
+ list
->relocs
[i
].offset
;
1246 write_reloc(device
, p
, target_bo
->offset
+ list
->relocs
[i
].delta
, true);
1247 list
->relocs
[i
].presumed_offset
= target_bo
->offset
;
1252 * This function applies the relocation for a command buffer and writes the
1253 * actual addresses into the buffers as per what we were told by the kernel on
1254 * the previous execbuf2 call. This should be safe to do because, for each
1255 * relocated address, we have two cases:
1257 * 1) The target BO is inactive (as seen by the kernel). In this case, it is
1258 * not in use by the GPU so updating the address is 100% ok. It won't be
1259 * in-use by the GPU (from our context) again until the next execbuf2
1260 * happens. If the kernel decides to move it in the next execbuf2, it
1261 * will have to do the relocations itself, but that's ok because it should
1262 * have all of the information needed to do so.
1264 * 2) The target BO is active (as seen by the kernel). In this case, it
1265 * hasn't moved since the last execbuffer2 call because GTT shuffling
1266 * *only* happens when the BO is idle. (From our perspective, it only
1267 * happens inside the execbuffer2 ioctl, but the shuffling may be
1268 * triggered by another ioctl, with full-ppgtt this is limited to only
1269 * execbuffer2 ioctls on the same context, or memory pressure.) Since the
1270 * target BO hasn't moved, our anv_bo::offset exactly matches the BO's GTT
1271 * address and the relocated value we are writing into the BO will be the
1272 * same as the value that is already there.
1274 * There is also a possibility that the target BO is active but the exact
1275 * RENDER_SURFACE_STATE object we are writing the relocation into isn't in
1276 * use. In this case, the address currently in the RENDER_SURFACE_STATE
1277 * may be stale but it's still safe to write the relocation because that
1278 * particular RENDER_SURFACE_STATE object isn't in-use by the GPU and
1279 * won't be until the next execbuf2 call.
1281 * By doing relocations on the CPU, we can tell the kernel that it doesn't
1282 * need to bother. We want to do this because the surface state buffer is
1283 * used by every command buffer so, if the kernel does the relocations, it
1284 * will always be busy and the kernel will always stall. This is also
1285 * probably the fastest mechanism for doing relocations since the kernel would
1286 * have to make a full copy of all the relocations lists.
1289 relocate_cmd_buffer(struct anv_cmd_buffer
*cmd_buffer
,
1290 struct anv_execbuf
*exec
)
1292 static int userspace_relocs
= -1;
1293 if (userspace_relocs
< 0)
1294 userspace_relocs
= env_var_as_boolean("ANV_USERSPACE_RELOCS", true);
1295 if (!userspace_relocs
)
1298 /* First, we have to check to see whether or not we can even do the
1299 * relocation. New buffers which have never been submitted to the kernel
1300 * don't have a valid offset so we need to let the kernel do relocations so
1301 * that we can get offsets for them. On future execbuf2 calls, those
1302 * buffers will have offsets and we will be able to skip relocating.
1303 * Invalid offsets are indicated by anv_bo::offset == (uint64_t)-1.
1305 for (uint32_t i
= 0; i
< exec
->bo_count
; i
++) {
1306 if (exec
->bos
[i
]->offset
== (uint64_t)-1)
1310 /* Since surface states are shared between command buffers and we don't
1311 * know what order they will be submitted to the kernel, we don't know
1312 * what address is actually written in the surface state object at any
1313 * given time. The only option is to always relocate them.
1315 anv_reloc_list_apply(cmd_buffer
->device
, &cmd_buffer
->surface_relocs
,
1316 &cmd_buffer
->device
->surface_state_pool
.block_pool
.bo
,
1317 true /* always relocate surface states */);
1319 /* Since we own all of the batch buffers, we know what values are stored
1320 * in the relocated addresses and only have to update them if the offsets
1323 struct anv_batch_bo
**bbo
;
1324 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
) {
1325 anv_reloc_list_apply(cmd_buffer
->device
,
1326 &(*bbo
)->relocs
, &(*bbo
)->bo
, false);
1329 for (uint32_t i
= 0; i
< exec
->bo_count
; i
++)
1330 exec
->objects
[i
].offset
= exec
->bos
[i
]->offset
;
1336 setup_execbuf_for_cmd_buffer(struct anv_execbuf
*execbuf
,
1337 struct anv_cmd_buffer
*cmd_buffer
)
1339 struct anv_batch
*batch
= &cmd_buffer
->batch
;
1340 struct anv_state_pool
*ss_pool
=
1341 &cmd_buffer
->device
->surface_state_pool
;
1343 adjust_relocations_from_state_pool(ss_pool
, &cmd_buffer
->surface_relocs
,
1344 cmd_buffer
->last_ss_pool_center
);
1345 VkResult result
= anv_execbuf_add_bo(execbuf
, &ss_pool
->block_pool
.bo
,
1346 &cmd_buffer
->surface_relocs
, 0,
1347 &cmd_buffer
->device
->alloc
);
1348 if (result
!= VK_SUCCESS
)
1351 /* First, we walk over all of the bos we've seen and add them and their
1352 * relocations to the validate list.
1354 struct anv_batch_bo
**bbo
;
1355 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
) {
1356 adjust_relocations_to_state_pool(ss_pool
, &(*bbo
)->bo
, &(*bbo
)->relocs
,
1357 cmd_buffer
->last_ss_pool_center
);
1359 result
= anv_execbuf_add_bo(execbuf
, &(*bbo
)->bo
, &(*bbo
)->relocs
, 0,
1360 &cmd_buffer
->device
->alloc
);
1361 if (result
!= VK_SUCCESS
)
1365 /* Now that we've adjusted all of the surface state relocations, we need to
1366 * record the surface state pool center so future executions of the command
1367 * buffer can adjust correctly.
1369 cmd_buffer
->last_ss_pool_center
= ss_pool
->block_pool
.center_bo_offset
;
1371 struct anv_batch_bo
*first_batch_bo
=
1372 list_first_entry(&cmd_buffer
->batch_bos
, struct anv_batch_bo
, link
);
1374 /* The kernel requires that the last entry in the validation list be the
1375 * batch buffer to execute. We can simply swap the element
1376 * corresponding to the first batch_bo in the chain with the last
1377 * element in the list.
1379 if (first_batch_bo
->bo
.index
!= execbuf
->bo_count
- 1) {
1380 uint32_t idx
= first_batch_bo
->bo
.index
;
1381 uint32_t last_idx
= execbuf
->bo_count
- 1;
1383 struct drm_i915_gem_exec_object2 tmp_obj
= execbuf
->objects
[idx
];
1384 assert(execbuf
->bos
[idx
] == &first_batch_bo
->bo
);
1386 execbuf
->objects
[idx
] = execbuf
->objects
[last_idx
];
1387 execbuf
->bos
[idx
] = execbuf
->bos
[last_idx
];
1388 execbuf
->bos
[idx
]->index
= idx
;
1390 execbuf
->objects
[last_idx
] = tmp_obj
;
1391 execbuf
->bos
[last_idx
] = &first_batch_bo
->bo
;
1392 first_batch_bo
->bo
.index
= last_idx
;
1395 /* Now we go through and fixup all of the relocation lists to point to
1396 * the correct indices in the object array. We have to do this after we
1397 * reorder the list above as some of the indices may have changed.
1399 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
)
1400 anv_cmd_buffer_process_relocs(cmd_buffer
, &(*bbo
)->relocs
);
1402 anv_cmd_buffer_process_relocs(cmd_buffer
, &cmd_buffer
->surface_relocs
);
1404 if (!cmd_buffer
->device
->info
.has_llc
) {
1405 __builtin_ia32_mfence();
1406 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
) {
1407 for (uint32_t i
= 0; i
< (*bbo
)->length
; i
+= CACHELINE_SIZE
)
1408 __builtin_ia32_clflush((*bbo
)->bo
.map
+ i
);
1412 execbuf
->execbuf
= (struct drm_i915_gem_execbuffer2
) {
1413 .buffers_ptr
= (uintptr_t) execbuf
->objects
,
1414 .buffer_count
= execbuf
->bo_count
,
1415 .batch_start_offset
= 0,
1416 .batch_len
= batch
->next
- batch
->start
,
1421 .flags
= I915_EXEC_HANDLE_LUT
| I915_EXEC_RENDER
,
1422 .rsvd1
= cmd_buffer
->device
->context_id
,
1426 if (relocate_cmd_buffer(cmd_buffer
, execbuf
)) {
1427 /* If we were able to successfully relocate everything, tell the kernel
1428 * that it can skip doing relocations. The requirement for using
1431 * 1) The addresses written in the objects must match the corresponding
1432 * reloc.presumed_offset which in turn must match the corresponding
1433 * execobject.offset.
1435 * 2) To avoid stalling, execobject.offset should match the current
1436 * address of that object within the active context.
1438 * In order to satisfy all of the invariants that make userspace
1439 * relocations to be safe (see relocate_cmd_buffer()), we need to
1440 * further ensure that the addresses we use match those used by the
1441 * kernel for the most recent execbuf2.
1443 * The kernel may still choose to do relocations anyway if something has
1444 * moved in the GTT. In this case, the relocation list still needs to be
1445 * valid. All relocations on the batch buffers are already valid and
1446 * kept up-to-date. For surface state relocations, by applying the
1447 * relocations in relocate_cmd_buffer, we ensured that the address in
1448 * the RENDER_SURFACE_STATE matches presumed_offset, so it should be
1449 * safe for the kernel to relocate them as needed.
1451 execbuf
->execbuf
.flags
|= I915_EXEC_NO_RELOC
;
1453 /* In the case where we fall back to doing kernel relocations, we need
1454 * to ensure that the relocation list is valid. All relocations on the
1455 * batch buffers are already valid and kept up-to-date. Since surface
1456 * states are shared between command buffers and we don't know what
1457 * order they will be submitted to the kernel, we don't know what
1458 * address is actually written in the surface state object at any given
1459 * time. The only option is to set a bogus presumed offset and let the
1460 * kernel relocate them.
1462 for (size_t i
= 0; i
< cmd_buffer
->surface_relocs
.num_relocs
; i
++)
1463 cmd_buffer
->surface_relocs
.relocs
[i
].presumed_offset
= -1;
1470 setup_empty_execbuf(struct anv_execbuf
*execbuf
, struct anv_device
*device
)
1472 VkResult result
= anv_execbuf_add_bo(execbuf
, &device
->trivial_batch_bo
,
1473 NULL
, 0, &device
->alloc
);
1474 if (result
!= VK_SUCCESS
)
1477 execbuf
->execbuf
= (struct drm_i915_gem_execbuffer2
) {
1478 .buffers_ptr
= (uintptr_t) execbuf
->objects
,
1479 .buffer_count
= execbuf
->bo_count
,
1480 .batch_start_offset
= 0,
1481 .batch_len
= 8, /* GEN7_MI_BATCH_BUFFER_END and NOOP */
1482 .flags
= I915_EXEC_HANDLE_LUT
| I915_EXEC_RENDER
,
1483 .rsvd1
= device
->context_id
,
1491 anv_cmd_buffer_execbuf(struct anv_device
*device
,
1492 struct anv_cmd_buffer
*cmd_buffer
,
1493 const VkSemaphore
*in_semaphores
,
1494 uint32_t num_in_semaphores
,
1495 const VkSemaphore
*out_semaphores
,
1496 uint32_t num_out_semaphores
,
1499 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
1501 struct anv_execbuf execbuf
;
1502 anv_execbuf_init(&execbuf
);
1505 VkResult result
= VK_SUCCESS
;
1506 for (uint32_t i
= 0; i
< num_in_semaphores
; i
++) {
1507 ANV_FROM_HANDLE(anv_semaphore
, semaphore
, in_semaphores
[i
]);
1508 struct anv_semaphore_impl
*impl
=
1509 semaphore
->temporary
.type
!= ANV_SEMAPHORE_TYPE_NONE
?
1510 &semaphore
->temporary
: &semaphore
->permanent
;
1512 switch (impl
->type
) {
1513 case ANV_SEMAPHORE_TYPE_BO
:
1514 result
= anv_execbuf_add_bo(&execbuf
, impl
->bo
, NULL
,
1516 if (result
!= VK_SUCCESS
)
1520 case ANV_SEMAPHORE_TYPE_SYNC_FILE
:
1521 if (in_fence
== -1) {
1522 in_fence
= impl
->fd
;
1524 int merge
= anv_gem_sync_file_merge(device
, in_fence
, impl
->fd
);
1526 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1536 case ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ
:
1537 result
= anv_execbuf_add_syncobj(&execbuf
, impl
->syncobj
,
1538 I915_EXEC_FENCE_WAIT
,
1540 if (result
!= VK_SUCCESS
)
1549 bool need_out_fence
= false;
1550 for (uint32_t i
= 0; i
< num_out_semaphores
; i
++) {
1551 ANV_FROM_HANDLE(anv_semaphore
, semaphore
, out_semaphores
[i
]);
1553 /* Under most circumstances, out fences won't be temporary. However,
1554 * the spec does allow it for opaque_fd. From the Vulkan 1.0.53 spec:
1556 * "If the import is temporary, the implementation must restore the
1557 * semaphore to its prior permanent state after submitting the next
1558 * semaphore wait operation."
1560 * The spec says nothing whatsoever about signal operations on
1561 * temporarily imported semaphores so it appears they are allowed.
1562 * There are also CTS tests that require this to work.
1564 struct anv_semaphore_impl
*impl
=
1565 semaphore
->temporary
.type
!= ANV_SEMAPHORE_TYPE_NONE
?
1566 &semaphore
->temporary
: &semaphore
->permanent
;
1568 switch (impl
->type
) {
1569 case ANV_SEMAPHORE_TYPE_BO
:
1570 result
= anv_execbuf_add_bo(&execbuf
, impl
->bo
, NULL
,
1571 EXEC_OBJECT_WRITE
, &device
->alloc
);
1572 if (result
!= VK_SUCCESS
)
1576 case ANV_SEMAPHORE_TYPE_SYNC_FILE
:
1577 need_out_fence
= true;
1580 case ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ
:
1581 result
= anv_execbuf_add_syncobj(&execbuf
, impl
->syncobj
,
1582 I915_EXEC_FENCE_SIGNAL
,
1584 if (result
!= VK_SUCCESS
)
1594 /* Under most circumstances, out fences won't be temporary. However,
1595 * the spec does allow it for opaque_fd. From the Vulkan 1.0.53 spec:
1597 * "If the import is temporary, the implementation must restore the
1598 * semaphore to its prior permanent state after submitting the next
1599 * semaphore wait operation."
1601 * The spec says nothing whatsoever about signal operations on
1602 * temporarily imported semaphores so it appears they are allowed.
1603 * There are also CTS tests that require this to work.
1605 struct anv_fence_impl
*impl
=
1606 fence
->temporary
.type
!= ANV_FENCE_TYPE_NONE
?
1607 &fence
->temporary
: &fence
->permanent
;
1609 switch (impl
->type
) {
1610 case ANV_FENCE_TYPE_BO
:
1611 result
= anv_execbuf_add_bo(&execbuf
, &impl
->bo
.bo
, NULL
,
1612 EXEC_OBJECT_WRITE
, &device
->alloc
);
1613 if (result
!= VK_SUCCESS
)
1617 case ANV_FENCE_TYPE_SYNCOBJ
:
1618 result
= anv_execbuf_add_syncobj(&execbuf
, impl
->syncobj
,
1619 I915_EXEC_FENCE_SIGNAL
,
1621 if (result
!= VK_SUCCESS
)
1626 unreachable("Invalid fence type");
1631 result
= setup_execbuf_for_cmd_buffer(&execbuf
, cmd_buffer
);
1633 result
= setup_empty_execbuf(&execbuf
, device
);
1635 if (result
!= VK_SUCCESS
)
1638 if (execbuf
.fence_count
> 0) {
1639 assert(device
->instance
->physicalDevice
.has_syncobj
);
1640 execbuf
.execbuf
.flags
|= I915_EXEC_FENCE_ARRAY
;
1641 execbuf
.execbuf
.num_cliprects
= execbuf
.fence_count
;
1642 execbuf
.execbuf
.cliprects_ptr
= (uintptr_t) execbuf
.fences
;
1645 if (in_fence
!= -1) {
1646 execbuf
.execbuf
.flags
|= I915_EXEC_FENCE_IN
;
1647 execbuf
.execbuf
.rsvd2
|= (uint32_t)in_fence
;
1651 execbuf
.execbuf
.flags
|= I915_EXEC_FENCE_OUT
;
1653 result
= anv_device_execbuf(device
, &execbuf
.execbuf
, execbuf
.bos
);
1655 /* Execbuf does not consume the in_fence. It's our job to close it. */
1659 for (uint32_t i
= 0; i
< num_in_semaphores
; i
++) {
1660 ANV_FROM_HANDLE(anv_semaphore
, semaphore
, in_semaphores
[i
]);
1661 /* From the Vulkan 1.0.53 spec:
1663 * "If the import is temporary, the implementation must restore the
1664 * semaphore to its prior permanent state after submitting the next
1665 * semaphore wait operation."
1667 * This has to happen after the execbuf in case we close any syncobjs in
1670 anv_semaphore_reset_temporary(device
, semaphore
);
1673 if (fence
&& fence
->permanent
.type
== ANV_FENCE_TYPE_BO
) {
1674 /* BO fences can't be shared, so they can't be temporary. */
1675 assert(fence
->temporary
.type
== ANV_FENCE_TYPE_NONE
);
1677 /* Once the execbuf has returned, we need to set the fence state to
1678 * SUBMITTED. We can't do this before calling execbuf because
1679 * anv_GetFenceStatus does take the global device lock before checking
1682 * We set the fence state to SUBMITTED regardless of whether or not the
1683 * execbuf succeeds because we need to ensure that vkWaitForFences() and
1684 * vkGetFenceStatus() return a valid result (VK_ERROR_DEVICE_LOST or
1685 * VK_SUCCESS) in a finite amount of time even if execbuf fails.
1687 fence
->permanent
.bo
.state
= ANV_BO_FENCE_STATE_SUBMITTED
;
1690 if (result
== VK_SUCCESS
&& need_out_fence
) {
1691 int out_fence
= execbuf
.execbuf
.rsvd2
>> 32;
1692 for (uint32_t i
= 0; i
< num_out_semaphores
; i
++) {
1693 ANV_FROM_HANDLE(anv_semaphore
, semaphore
, out_semaphores
[i
]);
1694 /* Out fences can't have temporary state because that would imply
1695 * that we imported a sync file and are trying to signal it.
1697 assert(semaphore
->temporary
.type
== ANV_SEMAPHORE_TYPE_NONE
);
1698 struct anv_semaphore_impl
*impl
= &semaphore
->permanent
;
1700 if (impl
->type
== ANV_SEMAPHORE_TYPE_SYNC_FILE
) {
1701 assert(impl
->fd
== -1);
1702 impl
->fd
= dup(out_fence
);
1708 anv_execbuf_finish(&execbuf
, &device
->alloc
);