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 if (other_list
->deps
) {
90 list
->deps
= _mesa_set_clone(other_list
->deps
, NULL
);
92 vk_free(alloc
, list
->relocs
);
93 vk_free(alloc
, list
->reloc_bos
);
94 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
104 anv_reloc_list_finish(struct anv_reloc_list
*list
,
105 const VkAllocationCallbacks
*alloc
)
107 vk_free(alloc
, list
->relocs
);
108 vk_free(alloc
, list
->reloc_bos
);
109 if (list
->deps
!= NULL
)
110 _mesa_set_destroy(list
->deps
, NULL
);
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(256, 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_add(struct anv_reloc_list
*list
,
148 const VkAllocationCallbacks
*alloc
,
149 uint32_t offset
, struct anv_bo
*target_bo
, uint32_t delta
)
151 struct drm_i915_gem_relocation_entry
*entry
;
154 if (target_bo
->flags
& EXEC_OBJECT_PINNED
) {
155 if (list
->deps
== NULL
) {
156 list
->deps
= _mesa_pointer_set_create(NULL
);
157 if (unlikely(list
->deps
== NULL
))
158 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
160 _mesa_set_add(list
->deps
, target_bo
);
164 VkResult result
= anv_reloc_list_grow(list
, alloc
, 1);
165 if (result
!= VK_SUCCESS
)
168 /* XXX: Can we use I915_EXEC_HANDLE_LUT? */
169 index
= list
->num_relocs
++;
170 list
->reloc_bos
[index
] = target_bo
;
171 entry
= &list
->relocs
[index
];
172 entry
->target_handle
= target_bo
->gem_handle
;
173 entry
->delta
= delta
;
174 entry
->offset
= offset
;
175 entry
->presumed_offset
= target_bo
->offset
;
176 entry
->read_domains
= 0;
177 entry
->write_domain
= 0;
178 VG(VALGRIND_CHECK_MEM_IS_DEFINED(entry
, sizeof(*entry
)));
184 anv_reloc_list_append(struct anv_reloc_list
*list
,
185 const VkAllocationCallbacks
*alloc
,
186 struct anv_reloc_list
*other
, uint32_t offset
)
188 VkResult result
= anv_reloc_list_grow(list
, alloc
, other
->num_relocs
);
189 if (result
!= VK_SUCCESS
)
192 if (other
->num_relocs
> 0) {
193 memcpy(&list
->relocs
[list
->num_relocs
], &other
->relocs
[0],
194 other
->num_relocs
* sizeof(other
->relocs
[0]));
195 memcpy(&list
->reloc_bos
[list
->num_relocs
], &other
->reloc_bos
[0],
196 other
->num_relocs
* sizeof(other
->reloc_bos
[0]));
198 for (uint32_t i
= 0; i
< other
->num_relocs
; i
++)
199 list
->relocs
[i
+ list
->num_relocs
].offset
+= offset
;
201 list
->num_relocs
+= other
->num_relocs
;
205 if (list
->deps
== NULL
) {
206 list
->deps
= _mesa_pointer_set_create(NULL
);
207 if (unlikely(list
->deps
== NULL
))
208 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
210 set_foreach(other
->deps
, entry
)
211 _mesa_set_add_pre_hashed(list
->deps
, entry
->hash
, entry
->key
);
217 /*-----------------------------------------------------------------------*
218 * Functions related to anv_batch
219 *-----------------------------------------------------------------------*/
222 anv_batch_emit_dwords(struct anv_batch
*batch
, int num_dwords
)
224 if (batch
->next
+ num_dwords
* 4 > batch
->end
) {
225 VkResult result
= batch
->extend_cb(batch
, batch
->user_data
);
226 if (result
!= VK_SUCCESS
) {
227 anv_batch_set_error(batch
, result
);
232 void *p
= batch
->next
;
234 batch
->next
+= num_dwords
* 4;
235 assert(batch
->next
<= batch
->end
);
241 anv_batch_emit_reloc(struct anv_batch
*batch
,
242 void *location
, struct anv_bo
*bo
, uint32_t delta
)
244 VkResult result
= anv_reloc_list_add(batch
->relocs
, batch
->alloc
,
245 location
- batch
->start
, bo
, delta
);
246 if (result
!= VK_SUCCESS
) {
247 anv_batch_set_error(batch
, result
);
251 return bo
->offset
+ delta
;
255 anv_batch_emit_batch(struct anv_batch
*batch
, struct anv_batch
*other
)
257 uint32_t size
, offset
;
259 size
= other
->next
- other
->start
;
260 assert(size
% 4 == 0);
262 if (batch
->next
+ size
> batch
->end
) {
263 VkResult result
= batch
->extend_cb(batch
, batch
->user_data
);
264 if (result
!= VK_SUCCESS
) {
265 anv_batch_set_error(batch
, result
);
270 assert(batch
->next
+ size
<= batch
->end
);
272 VG(VALGRIND_CHECK_MEM_IS_DEFINED(other
->start
, size
));
273 memcpy(batch
->next
, other
->start
, size
);
275 offset
= batch
->next
- batch
->start
;
276 VkResult result
= anv_reloc_list_append(batch
->relocs
, batch
->alloc
,
277 other
->relocs
, offset
);
278 if (result
!= VK_SUCCESS
) {
279 anv_batch_set_error(batch
, result
);
286 /*-----------------------------------------------------------------------*
287 * Functions related to anv_batch_bo
288 *-----------------------------------------------------------------------*/
291 anv_batch_bo_create(struct anv_cmd_buffer
*cmd_buffer
,
292 struct anv_batch_bo
**bbo_out
)
296 struct anv_batch_bo
*bbo
= vk_alloc(&cmd_buffer
->pool
->alloc
, sizeof(*bbo
),
297 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
299 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
301 result
= anv_bo_pool_alloc(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
,
302 ANV_CMD_BUFFER_BATCH_SIZE
);
303 if (result
!= VK_SUCCESS
)
306 result
= anv_reloc_list_init(&bbo
->relocs
, &cmd_buffer
->pool
->alloc
);
307 if (result
!= VK_SUCCESS
)
315 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
317 vk_free(&cmd_buffer
->pool
->alloc
, bbo
);
323 anv_batch_bo_clone(struct anv_cmd_buffer
*cmd_buffer
,
324 const struct anv_batch_bo
*other_bbo
,
325 struct anv_batch_bo
**bbo_out
)
329 struct anv_batch_bo
*bbo
= vk_alloc(&cmd_buffer
->pool
->alloc
, sizeof(*bbo
),
330 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
332 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
334 result
= anv_bo_pool_alloc(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
,
336 if (result
!= VK_SUCCESS
)
339 result
= anv_reloc_list_init_clone(&bbo
->relocs
, &cmd_buffer
->pool
->alloc
,
341 if (result
!= VK_SUCCESS
)
344 bbo
->length
= other_bbo
->length
;
345 memcpy(bbo
->bo
.map
, other_bbo
->bo
.map
, other_bbo
->length
);
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_start(struct anv_batch_bo
*bbo
, struct anv_batch
*batch
,
361 size_t batch_padding
)
363 batch
->next
= batch
->start
= bbo
->bo
.map
;
364 batch
->end
= bbo
->bo
.map
+ bbo
->bo
.size
- batch_padding
;
365 batch
->relocs
= &bbo
->relocs
;
366 bbo
->relocs
.num_relocs
= 0;
367 _mesa_set_clear(bbo
->relocs
.deps
, NULL
);
371 anv_batch_bo_continue(struct anv_batch_bo
*bbo
, struct anv_batch
*batch
,
372 size_t batch_padding
)
374 batch
->start
= bbo
->bo
.map
;
375 batch
->next
= bbo
->bo
.map
+ bbo
->length
;
376 batch
->end
= bbo
->bo
.map
+ bbo
->bo
.size
- batch_padding
;
377 batch
->relocs
= &bbo
->relocs
;
381 anv_batch_bo_finish(struct anv_batch_bo
*bbo
, struct anv_batch
*batch
)
383 assert(batch
->start
== bbo
->bo
.map
);
384 bbo
->length
= batch
->next
- batch
->start
;
385 VG(VALGRIND_CHECK_MEM_IS_DEFINED(batch
->start
, bbo
->length
));
389 anv_batch_bo_grow(struct anv_cmd_buffer
*cmd_buffer
, struct anv_batch_bo
*bbo
,
390 struct anv_batch
*batch
, size_t aditional
,
391 size_t batch_padding
)
393 assert(batch
->start
== bbo
->bo
.map
);
394 bbo
->length
= batch
->next
- batch
->start
;
396 size_t new_size
= bbo
->bo
.size
;
397 while (new_size
<= bbo
->length
+ aditional
+ batch_padding
)
400 if (new_size
== bbo
->bo
.size
)
403 struct anv_bo new_bo
;
404 VkResult result
= anv_bo_pool_alloc(&cmd_buffer
->device
->batch_bo_pool
,
406 if (result
!= VK_SUCCESS
)
409 memcpy(new_bo
.map
, bbo
->bo
.map
, bbo
->length
);
411 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
414 anv_batch_bo_continue(bbo
, batch
, batch_padding
);
420 anv_batch_bo_link(struct anv_cmd_buffer
*cmd_buffer
,
421 struct anv_batch_bo
*prev_bbo
,
422 struct anv_batch_bo
*next_bbo
,
423 uint32_t next_bbo_offset
)
425 const uint32_t bb_start_offset
=
426 prev_bbo
->length
- GEN8_MI_BATCH_BUFFER_START_length
* 4;
427 ASSERTED
const uint32_t *bb_start
= prev_bbo
->bo
.map
+ bb_start_offset
;
429 /* Make sure we're looking at a MI_BATCH_BUFFER_START */
430 assert(((*bb_start
>> 29) & 0x07) == 0);
431 assert(((*bb_start
>> 23) & 0x3f) == 49);
433 if (cmd_buffer
->device
->instance
->physicalDevice
.use_softpin
) {
434 assert(prev_bbo
->bo
.flags
& EXEC_OBJECT_PINNED
);
435 assert(next_bbo
->bo
.flags
& EXEC_OBJECT_PINNED
);
437 write_reloc(cmd_buffer
->device
,
438 prev_bbo
->bo
.map
+ bb_start_offset
+ 4,
439 next_bbo
->bo
.offset
+ next_bbo_offset
, true);
441 uint32_t reloc_idx
= prev_bbo
->relocs
.num_relocs
- 1;
442 assert(prev_bbo
->relocs
.relocs
[reloc_idx
].offset
== bb_start_offset
+ 4);
444 prev_bbo
->relocs
.reloc_bos
[reloc_idx
] = &next_bbo
->bo
;
445 prev_bbo
->relocs
.relocs
[reloc_idx
].delta
= next_bbo_offset
;
447 /* Use a bogus presumed offset to force a relocation */
448 prev_bbo
->relocs
.relocs
[reloc_idx
].presumed_offset
= -1;
453 anv_batch_bo_destroy(struct anv_batch_bo
*bbo
,
454 struct anv_cmd_buffer
*cmd_buffer
)
456 anv_reloc_list_finish(&bbo
->relocs
, &cmd_buffer
->pool
->alloc
);
457 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
458 vk_free(&cmd_buffer
->pool
->alloc
, bbo
);
462 anv_batch_bo_list_clone(const struct list_head
*list
,
463 struct anv_cmd_buffer
*cmd_buffer
,
464 struct list_head
*new_list
)
466 VkResult result
= VK_SUCCESS
;
468 list_inithead(new_list
);
470 struct anv_batch_bo
*prev_bbo
= NULL
;
471 list_for_each_entry(struct anv_batch_bo
, bbo
, list
, link
) {
472 struct anv_batch_bo
*new_bbo
= NULL
;
473 result
= anv_batch_bo_clone(cmd_buffer
, bbo
, &new_bbo
);
474 if (result
!= VK_SUCCESS
)
476 list_addtail(&new_bbo
->link
, new_list
);
479 anv_batch_bo_link(cmd_buffer
, prev_bbo
, new_bbo
, 0);
484 if (result
!= VK_SUCCESS
) {
485 list_for_each_entry_safe(struct anv_batch_bo
, bbo
, new_list
, link
)
486 anv_batch_bo_destroy(bbo
, cmd_buffer
);
492 /*-----------------------------------------------------------------------*
493 * Functions related to anv_batch_bo
494 *-----------------------------------------------------------------------*/
496 static struct anv_batch_bo
*
497 anv_cmd_buffer_current_batch_bo(struct anv_cmd_buffer
*cmd_buffer
)
499 return LIST_ENTRY(struct anv_batch_bo
, cmd_buffer
->batch_bos
.prev
, link
);
503 anv_cmd_buffer_surface_base_address(struct anv_cmd_buffer
*cmd_buffer
)
505 struct anv_state
*bt_block
= u_vector_head(&cmd_buffer
->bt_block_states
);
506 return (struct anv_address
) {
507 .bo
= anv_binding_table_pool(cmd_buffer
->device
)->block_pool
.bo
,
508 .offset
= bt_block
->offset
,
513 emit_batch_buffer_start(struct anv_cmd_buffer
*cmd_buffer
,
514 struct anv_bo
*bo
, uint32_t offset
)
516 /* In gen8+ the address field grew to two dwords to accomodate 48 bit
517 * offsets. The high 16 bits are in the last dword, so we can use the gen8
518 * version in either case, as long as we set the instruction length in the
519 * header accordingly. This means that we always emit three dwords here
520 * and all the padding and adjustment we do in this file works for all
524 #define GEN7_MI_BATCH_BUFFER_START_length 2
525 #define GEN7_MI_BATCH_BUFFER_START_length_bias 2
527 const uint32_t gen7_length
=
528 GEN7_MI_BATCH_BUFFER_START_length
- GEN7_MI_BATCH_BUFFER_START_length_bias
;
529 const uint32_t gen8_length
=
530 GEN8_MI_BATCH_BUFFER_START_length
- GEN8_MI_BATCH_BUFFER_START_length_bias
;
532 anv_batch_emit(&cmd_buffer
->batch
, GEN8_MI_BATCH_BUFFER_START
, bbs
) {
533 bbs
.DWordLength
= cmd_buffer
->device
->info
.gen
< 8 ?
534 gen7_length
: gen8_length
;
535 bbs
.SecondLevelBatchBuffer
= Firstlevelbatch
;
536 bbs
.AddressSpaceIndicator
= ASI_PPGTT
;
537 bbs
.BatchBufferStartAddress
= (struct anv_address
) { bo
, offset
};
542 cmd_buffer_chain_to_batch_bo(struct anv_cmd_buffer
*cmd_buffer
,
543 struct anv_batch_bo
*bbo
)
545 struct anv_batch
*batch
= &cmd_buffer
->batch
;
546 struct anv_batch_bo
*current_bbo
=
547 anv_cmd_buffer_current_batch_bo(cmd_buffer
);
549 /* We set the end of the batch a little short so we would be sure we
550 * have room for the chaining command. Since we're about to emit the
551 * chaining command, let's set it back where it should go.
553 batch
->end
+= GEN8_MI_BATCH_BUFFER_START_length
* 4;
554 assert(batch
->end
== current_bbo
->bo
.map
+ current_bbo
->bo
.size
);
556 emit_batch_buffer_start(cmd_buffer
, &bbo
->bo
, 0);
558 anv_batch_bo_finish(current_bbo
, batch
);
562 anv_cmd_buffer_chain_batch(struct anv_batch
*batch
, void *_data
)
564 struct anv_cmd_buffer
*cmd_buffer
= _data
;
565 struct anv_batch_bo
*new_bbo
;
567 VkResult result
= anv_batch_bo_create(cmd_buffer
, &new_bbo
);
568 if (result
!= VK_SUCCESS
)
571 struct anv_batch_bo
**seen_bbo
= u_vector_add(&cmd_buffer
->seen_bbos
);
572 if (seen_bbo
== NULL
) {
573 anv_batch_bo_destroy(new_bbo
, cmd_buffer
);
574 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
578 cmd_buffer_chain_to_batch_bo(cmd_buffer
, new_bbo
);
580 list_addtail(&new_bbo
->link
, &cmd_buffer
->batch_bos
);
582 anv_batch_bo_start(new_bbo
, batch
, GEN8_MI_BATCH_BUFFER_START_length
* 4);
588 anv_cmd_buffer_grow_batch(struct anv_batch
*batch
, void *_data
)
590 struct anv_cmd_buffer
*cmd_buffer
= _data
;
591 struct anv_batch_bo
*bbo
= anv_cmd_buffer_current_batch_bo(cmd_buffer
);
593 anv_batch_bo_grow(cmd_buffer
, bbo
, &cmd_buffer
->batch
, 4096,
594 GEN8_MI_BATCH_BUFFER_START_length
* 4);
599 /** Allocate a binding table
601 * This function allocates a binding table. This is a bit more complicated
602 * than one would think due to a combination of Vulkan driver design and some
603 * unfortunate hardware restrictions.
605 * The 3DSTATE_BINDING_TABLE_POINTERS_* packets only have a 16-bit field for
606 * the binding table pointer which means that all binding tables need to live
607 * in the bottom 64k of surface state base address. The way the GL driver has
608 * classically dealt with this restriction is to emit all surface states
609 * on-the-fly into the batch and have a batch buffer smaller than 64k. This
610 * isn't really an option in Vulkan for a couple of reasons:
612 * 1) In Vulkan, we have growing (or chaining) batches so surface states have
613 * to live in their own buffer and we have to be able to re-emit
614 * STATE_BASE_ADDRESS as needed which requires a full pipeline stall. In
615 * order to avoid emitting STATE_BASE_ADDRESS any more often than needed
616 * (it's not that hard to hit 64k of just binding tables), we allocate
617 * surface state objects up-front when VkImageView is created. In order
618 * for this to work, surface state objects need to be allocated from a
621 * 2) We tried to design the surface state system in such a way that it's
622 * already ready for bindless texturing. The way bindless texturing works
623 * on our hardware is that you have a big pool of surface state objects
624 * (with its own state base address) and the bindless handles are simply
625 * offsets into that pool. With the architecture we chose, we already
626 * have that pool and it's exactly the same pool that we use for regular
627 * surface states so we should already be ready for bindless.
629 * 3) For render targets, we need to be able to fill out the surface states
630 * later in vkBeginRenderPass so that we can assign clear colors
631 * correctly. One way to do this would be to just create the surface
632 * state data and then repeatedly copy it into the surface state BO every
633 * time we have to re-emit STATE_BASE_ADDRESS. While this works, it's
634 * rather annoying and just being able to allocate them up-front and
635 * re-use them for the entire render pass.
637 * While none of these are technically blockers for emitting state on the fly
638 * like we do in GL, the ability to have a single surface state pool is
639 * simplifies things greatly. Unfortunately, it comes at a cost...
641 * Because of the 64k limitation of 3DSTATE_BINDING_TABLE_POINTERS_*, we can't
642 * place the binding tables just anywhere in surface state base address.
643 * Because 64k isn't a whole lot of space, we can't simply restrict the
644 * surface state buffer to 64k, we have to be more clever. The solution we've
645 * chosen is to have a block pool with a maximum size of 2G that starts at
646 * zero and grows in both directions. All surface states are allocated from
647 * the top of the pool (positive offsets) and we allocate blocks (< 64k) of
648 * binding tables from the bottom of the pool (negative offsets). Every time
649 * we allocate a new binding table block, we set surface state base address to
650 * point to the bottom of the binding table block. This way all of the
651 * binding tables in the block are in the bottom 64k of surface state base
652 * address. When we fill out the binding table, we add the distance between
653 * the bottom of our binding table block and zero of the block pool to the
654 * surface state offsets so that they are correct relative to out new surface
655 * state base address at the bottom of the binding table block.
657 * \see adjust_relocations_from_block_pool()
658 * \see adjust_relocations_too_block_pool()
660 * \param[in] entries The number of surface state entries the binding
661 * table should be able to hold.
663 * \param[out] state_offset The offset surface surface state base address
664 * where the surface states live. This must be
665 * added to the surface state offset when it is
666 * written into the binding table entry.
668 * \return An anv_state representing the binding table
671 anv_cmd_buffer_alloc_binding_table(struct anv_cmd_buffer
*cmd_buffer
,
672 uint32_t entries
, uint32_t *state_offset
)
674 struct anv_device
*device
= cmd_buffer
->device
;
675 struct anv_state_pool
*state_pool
= &device
->surface_state_pool
;
676 struct anv_state
*bt_block
= u_vector_head(&cmd_buffer
->bt_block_states
);
677 struct anv_state state
;
679 state
.alloc_size
= align_u32(entries
* 4, 32);
681 if (cmd_buffer
->bt_next
+ state
.alloc_size
> state_pool
->block_size
)
682 return (struct anv_state
) { 0 };
684 state
.offset
= cmd_buffer
->bt_next
;
685 state
.map
= anv_block_pool_map(&anv_binding_table_pool(device
)->block_pool
,
686 bt_block
->offset
+ state
.offset
);
688 cmd_buffer
->bt_next
+= state
.alloc_size
;
690 if (device
->instance
->physicalDevice
.use_softpin
) {
691 assert(bt_block
->offset
>= 0);
692 *state_offset
= device
->surface_state_pool
.block_pool
.start_address
-
693 device
->binding_table_pool
.block_pool
.start_address
- bt_block
->offset
;
695 assert(bt_block
->offset
< 0);
696 *state_offset
= -bt_block
->offset
;
703 anv_cmd_buffer_alloc_surface_state(struct anv_cmd_buffer
*cmd_buffer
)
705 struct isl_device
*isl_dev
= &cmd_buffer
->device
->isl_dev
;
706 return anv_state_stream_alloc(&cmd_buffer
->surface_state_stream
,
707 isl_dev
->ss
.size
, isl_dev
->ss
.align
);
711 anv_cmd_buffer_alloc_dynamic_state(struct anv_cmd_buffer
*cmd_buffer
,
712 uint32_t size
, uint32_t alignment
)
714 return anv_state_stream_alloc(&cmd_buffer
->dynamic_state_stream
,
719 anv_cmd_buffer_new_binding_table_block(struct anv_cmd_buffer
*cmd_buffer
)
721 struct anv_state
*bt_block
= u_vector_add(&cmd_buffer
->bt_block_states
);
722 if (bt_block
== NULL
) {
723 anv_batch_set_error(&cmd_buffer
->batch
, VK_ERROR_OUT_OF_HOST_MEMORY
);
724 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
727 *bt_block
= anv_binding_table_pool_alloc(cmd_buffer
->device
);
728 cmd_buffer
->bt_next
= 0;
734 anv_cmd_buffer_init_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
)
736 struct anv_batch_bo
*batch_bo
;
739 list_inithead(&cmd_buffer
->batch_bos
);
741 result
= anv_batch_bo_create(cmd_buffer
, &batch_bo
);
742 if (result
!= VK_SUCCESS
)
745 list_addtail(&batch_bo
->link
, &cmd_buffer
->batch_bos
);
747 cmd_buffer
->batch
.alloc
= &cmd_buffer
->pool
->alloc
;
748 cmd_buffer
->batch
.user_data
= cmd_buffer
;
750 if (cmd_buffer
->device
->can_chain_batches
) {
751 cmd_buffer
->batch
.extend_cb
= anv_cmd_buffer_chain_batch
;
753 cmd_buffer
->batch
.extend_cb
= anv_cmd_buffer_grow_batch
;
756 anv_batch_bo_start(batch_bo
, &cmd_buffer
->batch
,
757 GEN8_MI_BATCH_BUFFER_START_length
* 4);
759 int success
= u_vector_init(&cmd_buffer
->seen_bbos
,
760 sizeof(struct anv_bo
*),
761 8 * sizeof(struct anv_bo
*));
765 *(struct anv_batch_bo
**)u_vector_add(&cmd_buffer
->seen_bbos
) = batch_bo
;
767 /* u_vector requires power-of-two size elements */
768 unsigned pow2_state_size
= util_next_power_of_two(sizeof(struct anv_state
));
769 success
= u_vector_init(&cmd_buffer
->bt_block_states
,
770 pow2_state_size
, 8 * pow2_state_size
);
774 result
= anv_reloc_list_init(&cmd_buffer
->surface_relocs
,
775 &cmd_buffer
->pool
->alloc
);
776 if (result
!= VK_SUCCESS
)
778 cmd_buffer
->last_ss_pool_center
= 0;
780 result
= anv_cmd_buffer_new_binding_table_block(cmd_buffer
);
781 if (result
!= VK_SUCCESS
)
787 u_vector_finish(&cmd_buffer
->bt_block_states
);
789 u_vector_finish(&cmd_buffer
->seen_bbos
);
791 anv_batch_bo_destroy(batch_bo
, cmd_buffer
);
797 anv_cmd_buffer_fini_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
)
799 struct anv_state
*bt_block
;
800 u_vector_foreach(bt_block
, &cmd_buffer
->bt_block_states
)
801 anv_binding_table_pool_free(cmd_buffer
->device
, *bt_block
);
802 u_vector_finish(&cmd_buffer
->bt_block_states
);
804 anv_reloc_list_finish(&cmd_buffer
->surface_relocs
, &cmd_buffer
->pool
->alloc
);
806 u_vector_finish(&cmd_buffer
->seen_bbos
);
808 /* Destroy all of the batch buffers */
809 list_for_each_entry_safe(struct anv_batch_bo
, bbo
,
810 &cmd_buffer
->batch_bos
, link
) {
811 anv_batch_bo_destroy(bbo
, cmd_buffer
);
816 anv_cmd_buffer_reset_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
)
818 /* Delete all but the first batch bo */
819 assert(!list_is_empty(&cmd_buffer
->batch_bos
));
820 while (cmd_buffer
->batch_bos
.next
!= cmd_buffer
->batch_bos
.prev
) {
821 struct anv_batch_bo
*bbo
= anv_cmd_buffer_current_batch_bo(cmd_buffer
);
822 list_del(&bbo
->link
);
823 anv_batch_bo_destroy(bbo
, cmd_buffer
);
825 assert(!list_is_empty(&cmd_buffer
->batch_bos
));
827 anv_batch_bo_start(anv_cmd_buffer_current_batch_bo(cmd_buffer
),
829 GEN8_MI_BATCH_BUFFER_START_length
* 4);
831 while (u_vector_length(&cmd_buffer
->bt_block_states
) > 1) {
832 struct anv_state
*bt_block
= u_vector_remove(&cmd_buffer
->bt_block_states
);
833 anv_binding_table_pool_free(cmd_buffer
->device
, *bt_block
);
835 assert(u_vector_length(&cmd_buffer
->bt_block_states
) == 1);
836 cmd_buffer
->bt_next
= 0;
838 cmd_buffer
->surface_relocs
.num_relocs
= 0;
839 _mesa_set_clear(cmd_buffer
->surface_relocs
.deps
, NULL
);
840 cmd_buffer
->last_ss_pool_center
= 0;
842 /* Reset the list of seen buffers */
843 cmd_buffer
->seen_bbos
.head
= 0;
844 cmd_buffer
->seen_bbos
.tail
= 0;
846 *(struct anv_batch_bo
**)u_vector_add(&cmd_buffer
->seen_bbos
) =
847 anv_cmd_buffer_current_batch_bo(cmd_buffer
);
851 anv_cmd_buffer_end_batch_buffer(struct anv_cmd_buffer
*cmd_buffer
)
853 struct anv_batch_bo
*batch_bo
= anv_cmd_buffer_current_batch_bo(cmd_buffer
);
855 if (cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
) {
856 /* When we start a batch buffer, we subtract a certain amount of
857 * padding from the end to ensure that we always have room to emit a
858 * BATCH_BUFFER_START to chain to the next BO. We need to remove
859 * that padding before we end the batch; otherwise, we may end up
860 * with our BATCH_BUFFER_END in another BO.
862 cmd_buffer
->batch
.end
+= GEN8_MI_BATCH_BUFFER_START_length
* 4;
863 assert(cmd_buffer
->batch
.end
== batch_bo
->bo
.map
+ batch_bo
->bo
.size
);
865 anv_batch_emit(&cmd_buffer
->batch
, GEN8_MI_BATCH_BUFFER_END
, bbe
);
867 /* Round batch up to an even number of dwords. */
868 if ((cmd_buffer
->batch
.next
- cmd_buffer
->batch
.start
) & 4)
869 anv_batch_emit(&cmd_buffer
->batch
, GEN8_MI_NOOP
, noop
);
871 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_PRIMARY
;
873 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_SECONDARY
);
874 /* If this is a secondary command buffer, we need to determine the
875 * mode in which it will be executed with vkExecuteCommands. We
876 * determine this statically here so that this stays in sync with the
877 * actual ExecuteCommands implementation.
879 const uint32_t length
= cmd_buffer
->batch
.next
- cmd_buffer
->batch
.start
;
880 if (!cmd_buffer
->device
->can_chain_batches
) {
881 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_GROW_AND_EMIT
;
882 } else if ((cmd_buffer
->batch_bos
.next
== cmd_buffer
->batch_bos
.prev
) &&
883 (length
< ANV_CMD_BUFFER_BATCH_SIZE
/ 2)) {
884 /* If the secondary has exactly one batch buffer in its list *and*
885 * that batch buffer is less than half of the maximum size, we're
886 * probably better of simply copying it into our batch.
888 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_EMIT
;
889 } else if (!(cmd_buffer
->usage_flags
&
890 VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT
)) {
891 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_CHAIN
;
893 /* In order to chain, we need this command buffer to contain an
894 * MI_BATCH_BUFFER_START which will jump back to the calling batch.
895 * It doesn't matter where it points now so long as has a valid
896 * relocation. We'll adjust it later as part of the chaining
899 * We set the end of the batch a little short so we would be sure we
900 * have room for the chaining command. Since we're about to emit the
901 * chaining command, let's set it back where it should go.
903 cmd_buffer
->batch
.end
+= GEN8_MI_BATCH_BUFFER_START_length
* 4;
904 assert(cmd_buffer
->batch
.start
== batch_bo
->bo
.map
);
905 assert(cmd_buffer
->batch
.end
== batch_bo
->bo
.map
+ batch_bo
->bo
.size
);
907 emit_batch_buffer_start(cmd_buffer
, &batch_bo
->bo
, 0);
908 assert(cmd_buffer
->batch
.start
== batch_bo
->bo
.map
);
910 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN
;
914 anv_batch_bo_finish(batch_bo
, &cmd_buffer
->batch
);
918 anv_cmd_buffer_add_seen_bbos(struct anv_cmd_buffer
*cmd_buffer
,
919 struct list_head
*list
)
921 list_for_each_entry(struct anv_batch_bo
, bbo
, list
, link
) {
922 struct anv_batch_bo
**bbo_ptr
= u_vector_add(&cmd_buffer
->seen_bbos
);
924 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
933 anv_cmd_buffer_add_secondary(struct anv_cmd_buffer
*primary
,
934 struct anv_cmd_buffer
*secondary
)
936 switch (secondary
->exec_mode
) {
937 case ANV_CMD_BUFFER_EXEC_MODE_EMIT
:
938 anv_batch_emit_batch(&primary
->batch
, &secondary
->batch
);
940 case ANV_CMD_BUFFER_EXEC_MODE_GROW_AND_EMIT
: {
941 struct anv_batch_bo
*bbo
= anv_cmd_buffer_current_batch_bo(primary
);
942 unsigned length
= secondary
->batch
.end
- secondary
->batch
.start
;
943 anv_batch_bo_grow(primary
, bbo
, &primary
->batch
, length
,
944 GEN8_MI_BATCH_BUFFER_START_length
* 4);
945 anv_batch_emit_batch(&primary
->batch
, &secondary
->batch
);
948 case ANV_CMD_BUFFER_EXEC_MODE_CHAIN
: {
949 struct anv_batch_bo
*first_bbo
=
950 list_first_entry(&secondary
->batch_bos
, struct anv_batch_bo
, link
);
951 struct anv_batch_bo
*last_bbo
=
952 list_last_entry(&secondary
->batch_bos
, struct anv_batch_bo
, link
);
954 emit_batch_buffer_start(primary
, &first_bbo
->bo
, 0);
956 struct anv_batch_bo
*this_bbo
= anv_cmd_buffer_current_batch_bo(primary
);
957 assert(primary
->batch
.start
== this_bbo
->bo
.map
);
958 uint32_t offset
= primary
->batch
.next
- primary
->batch
.start
;
960 /* Make the tail of the secondary point back to right after the
961 * MI_BATCH_BUFFER_START in the primary batch.
963 anv_batch_bo_link(primary
, last_bbo
, this_bbo
, offset
);
965 anv_cmd_buffer_add_seen_bbos(primary
, &secondary
->batch_bos
);
968 case ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN
: {
969 struct list_head copy_list
;
970 VkResult result
= anv_batch_bo_list_clone(&secondary
->batch_bos
,
973 if (result
!= VK_SUCCESS
)
976 anv_cmd_buffer_add_seen_bbos(primary
, ©_list
);
978 struct anv_batch_bo
*first_bbo
=
979 list_first_entry(©_list
, struct anv_batch_bo
, link
);
980 struct anv_batch_bo
*last_bbo
=
981 list_last_entry(©_list
, struct anv_batch_bo
, link
);
983 cmd_buffer_chain_to_batch_bo(primary
, first_bbo
);
985 list_splicetail(©_list
, &primary
->batch_bos
);
987 anv_batch_bo_continue(last_bbo
, &primary
->batch
,
988 GEN8_MI_BATCH_BUFFER_START_length
* 4);
992 assert(!"Invalid execution mode");
995 anv_reloc_list_append(&primary
->surface_relocs
, &primary
->pool
->alloc
,
996 &secondary
->surface_relocs
, 0);
1000 struct drm_i915_gem_execbuffer2 execbuf
;
1002 struct drm_i915_gem_exec_object2
* objects
;
1004 struct anv_bo
** bos
;
1006 /* Allocated length of the 'objects' and 'bos' arrays */
1007 uint32_t array_length
;
1011 uint32_t fence_count
;
1012 uint32_t fence_array_length
;
1013 struct drm_i915_gem_exec_fence
* fences
;
1014 struct anv_syncobj
** syncobjs
;
1018 anv_execbuf_init(struct anv_execbuf
*exec
)
1020 memset(exec
, 0, sizeof(*exec
));
1024 anv_execbuf_finish(struct anv_execbuf
*exec
,
1025 const VkAllocationCallbacks
*alloc
)
1027 vk_free(alloc
, exec
->objects
);
1028 vk_free(alloc
, exec
->bos
);
1029 vk_free(alloc
, exec
->fences
);
1030 vk_free(alloc
, exec
->syncobjs
);
1034 _compare_bo_handles(const void *_bo1
, const void *_bo2
)
1036 struct anv_bo
* const *bo1
= _bo1
;
1037 struct anv_bo
* const *bo2
= _bo2
;
1039 return (*bo1
)->gem_handle
- (*bo2
)->gem_handle
;
1043 anv_execbuf_add_bo_set(struct anv_execbuf
*exec
,
1045 uint32_t extra_flags
,
1046 const VkAllocationCallbacks
*alloc
);
1049 anv_execbuf_add_bo(struct anv_execbuf
*exec
,
1051 struct anv_reloc_list
*relocs
,
1052 uint32_t extra_flags
,
1053 const VkAllocationCallbacks
*alloc
)
1055 struct drm_i915_gem_exec_object2
*obj
= NULL
;
1057 if (bo
->index
< exec
->bo_count
&& exec
->bos
[bo
->index
] == bo
)
1058 obj
= &exec
->objects
[bo
->index
];
1061 /* We've never seen this one before. Add it to the list and assign
1062 * an id that we can use later.
1064 if (exec
->bo_count
>= exec
->array_length
) {
1065 uint32_t new_len
= exec
->objects
? exec
->array_length
* 2 : 64;
1067 struct drm_i915_gem_exec_object2
*new_objects
=
1068 vk_alloc(alloc
, new_len
* sizeof(*new_objects
),
1069 8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND
);
1070 if (new_objects
== NULL
)
1071 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1073 struct anv_bo
**new_bos
=
1074 vk_alloc(alloc
, new_len
* sizeof(*new_bos
),
1075 8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND
);
1076 if (new_bos
== NULL
) {
1077 vk_free(alloc
, new_objects
);
1078 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1081 if (exec
->objects
) {
1082 memcpy(new_objects
, exec
->objects
,
1083 exec
->bo_count
* sizeof(*new_objects
));
1084 memcpy(new_bos
, exec
->bos
,
1085 exec
->bo_count
* sizeof(*new_bos
));
1088 vk_free(alloc
, exec
->objects
);
1089 vk_free(alloc
, exec
->bos
);
1091 exec
->objects
= new_objects
;
1092 exec
->bos
= new_bos
;
1093 exec
->array_length
= new_len
;
1096 assert(exec
->bo_count
< exec
->array_length
);
1098 bo
->index
= exec
->bo_count
++;
1099 obj
= &exec
->objects
[bo
->index
];
1100 exec
->bos
[bo
->index
] = bo
;
1102 obj
->handle
= bo
->gem_handle
;
1103 obj
->relocation_count
= 0;
1104 obj
->relocs_ptr
= 0;
1106 obj
->offset
= bo
->offset
;
1107 obj
->flags
= (bo
->flags
& ~ANV_BO_FLAG_MASK
) | extra_flags
;
1112 if (relocs
!= NULL
) {
1113 assert(obj
->relocation_count
== 0);
1115 if (relocs
->num_relocs
> 0) {
1116 /* This is the first time we've ever seen a list of relocations for
1117 * this BO. Go ahead and set the relocations and then walk the list
1118 * of relocations and add them all.
1120 exec
->has_relocs
= true;
1121 obj
->relocation_count
= relocs
->num_relocs
;
1122 obj
->relocs_ptr
= (uintptr_t) relocs
->relocs
;
1124 for (size_t i
= 0; i
< relocs
->num_relocs
; i
++) {
1127 /* A quick sanity check on relocations */
1128 assert(relocs
->relocs
[i
].offset
< bo
->size
);
1129 result
= anv_execbuf_add_bo(exec
, relocs
->reloc_bos
[i
], NULL
,
1130 extra_flags
, alloc
);
1132 if (result
!= VK_SUCCESS
)
1137 return anv_execbuf_add_bo_set(exec
, relocs
->deps
, extra_flags
, alloc
);
1143 /* Add BO dependencies to execbuf */
1145 anv_execbuf_add_bo_set(struct anv_execbuf
*exec
,
1147 uint32_t extra_flags
,
1148 const VkAllocationCallbacks
*alloc
)
1150 if (!deps
|| deps
->entries
<= 0)
1153 const uint32_t entries
= deps
->entries
;
1154 struct anv_bo
**bos
=
1155 vk_alloc(alloc
, entries
* sizeof(*bos
),
1156 8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND
);
1158 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1160 struct anv_bo
**bo
= bos
;
1161 set_foreach(deps
, entry
) {
1162 *bo
++ = (void *)entry
->key
;
1165 qsort(bos
, entries
, sizeof(struct anv_bo
*), _compare_bo_handles
);
1167 VkResult result
= VK_SUCCESS
;
1168 for (bo
= bos
; bo
< bos
+ entries
; bo
++) {
1169 result
= anv_execbuf_add_bo(exec
, *bo
, NULL
, extra_flags
, alloc
);
1170 if (result
!= VK_SUCCESS
)
1174 vk_free(alloc
, bos
);
1180 anv_execbuf_add_syncobj(struct anv_execbuf
*exec
,
1181 uint32_t handle
, uint32_t flags
,
1182 const VkAllocationCallbacks
*alloc
)
1186 if (exec
->fence_count
>= exec
->fence_array_length
) {
1187 uint32_t new_len
= MAX2(exec
->fence_array_length
* 2, 64);
1189 exec
->fences
= vk_realloc(alloc
, exec
->fences
,
1190 new_len
* sizeof(*exec
->fences
),
1191 8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND
);
1192 if (exec
->fences
== NULL
)
1193 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1195 exec
->fence_array_length
= new_len
;
1198 exec
->fences
[exec
->fence_count
] = (struct drm_i915_gem_exec_fence
) {
1203 exec
->fence_count
++;
1209 anv_cmd_buffer_process_relocs(struct anv_cmd_buffer
*cmd_buffer
,
1210 struct anv_reloc_list
*list
)
1212 for (size_t i
= 0; i
< list
->num_relocs
; i
++)
1213 list
->relocs
[i
].target_handle
= list
->reloc_bos
[i
]->index
;
1217 adjust_relocations_from_state_pool(struct anv_state_pool
*pool
,
1218 struct anv_reloc_list
*relocs
,
1219 uint32_t last_pool_center_bo_offset
)
1221 assert(last_pool_center_bo_offset
<= pool
->block_pool
.center_bo_offset
);
1222 uint32_t delta
= pool
->block_pool
.center_bo_offset
- last_pool_center_bo_offset
;
1224 for (size_t i
= 0; i
< relocs
->num_relocs
; i
++) {
1225 /* All of the relocations from this block pool to other BO's should
1226 * have been emitted relative to the surface block pool center. We
1227 * need to add the center offset to make them relative to the
1228 * beginning of the actual GEM bo.
1230 relocs
->relocs
[i
].offset
+= delta
;
1235 adjust_relocations_to_state_pool(struct anv_state_pool
*pool
,
1236 struct anv_bo
*from_bo
,
1237 struct anv_reloc_list
*relocs
,
1238 uint32_t last_pool_center_bo_offset
)
1240 assert(last_pool_center_bo_offset
<= pool
->block_pool
.center_bo_offset
);
1241 uint32_t delta
= pool
->block_pool
.center_bo_offset
- last_pool_center_bo_offset
;
1243 /* When we initially emit relocations into a block pool, we don't
1244 * actually know what the final center_bo_offset will be so we just emit
1245 * it as if center_bo_offset == 0. Now that we know what the center
1246 * offset is, we need to walk the list of relocations and adjust any
1247 * relocations that point to the pool bo with the correct offset.
1249 for (size_t i
= 0; i
< relocs
->num_relocs
; i
++) {
1250 if (relocs
->reloc_bos
[i
] == pool
->block_pool
.bo
) {
1251 /* Adjust the delta value in the relocation to correctly
1252 * correspond to the new delta. Initially, this value may have
1253 * been negative (if treated as unsigned), but we trust in
1254 * uint32_t roll-over to fix that for us at this point.
1256 relocs
->relocs
[i
].delta
+= delta
;
1258 /* Since the delta has changed, we need to update the actual
1259 * relocated value with the new presumed value. This function
1260 * should only be called on batch buffers, so we know it isn't in
1261 * use by the GPU at the moment.
1263 assert(relocs
->relocs
[i
].offset
< from_bo
->size
);
1264 write_reloc(pool
->block_pool
.device
,
1265 from_bo
->map
+ relocs
->relocs
[i
].offset
,
1266 relocs
->relocs
[i
].presumed_offset
+
1267 relocs
->relocs
[i
].delta
, false);
1273 anv_reloc_list_apply(struct anv_device
*device
,
1274 struct anv_reloc_list
*list
,
1276 bool always_relocate
)
1278 for (size_t i
= 0; i
< list
->num_relocs
; i
++) {
1279 struct anv_bo
*target_bo
= list
->reloc_bos
[i
];
1280 if (list
->relocs
[i
].presumed_offset
== target_bo
->offset
&&
1284 void *p
= bo
->map
+ list
->relocs
[i
].offset
;
1285 write_reloc(device
, p
, target_bo
->offset
+ list
->relocs
[i
].delta
, true);
1286 list
->relocs
[i
].presumed_offset
= target_bo
->offset
;
1291 * This function applies the relocation for a command buffer and writes the
1292 * actual addresses into the buffers as per what we were told by the kernel on
1293 * the previous execbuf2 call. This should be safe to do because, for each
1294 * relocated address, we have two cases:
1296 * 1) The target BO is inactive (as seen by the kernel). In this case, it is
1297 * not in use by the GPU so updating the address is 100% ok. It won't be
1298 * in-use by the GPU (from our context) again until the next execbuf2
1299 * happens. If the kernel decides to move it in the next execbuf2, it
1300 * will have to do the relocations itself, but that's ok because it should
1301 * have all of the information needed to do so.
1303 * 2) The target BO is active (as seen by the kernel). In this case, it
1304 * hasn't moved since the last execbuffer2 call because GTT shuffling
1305 * *only* happens when the BO is idle. (From our perspective, it only
1306 * happens inside the execbuffer2 ioctl, but the shuffling may be
1307 * triggered by another ioctl, with full-ppgtt this is limited to only
1308 * execbuffer2 ioctls on the same context, or memory pressure.) Since the
1309 * target BO hasn't moved, our anv_bo::offset exactly matches the BO's GTT
1310 * address and the relocated value we are writing into the BO will be the
1311 * same as the value that is already there.
1313 * There is also a possibility that the target BO is active but the exact
1314 * RENDER_SURFACE_STATE object we are writing the relocation into isn't in
1315 * use. In this case, the address currently in the RENDER_SURFACE_STATE
1316 * may be stale but it's still safe to write the relocation because that
1317 * particular RENDER_SURFACE_STATE object isn't in-use by the GPU and
1318 * won't be until the next execbuf2 call.
1320 * By doing relocations on the CPU, we can tell the kernel that it doesn't
1321 * need to bother. We want to do this because the surface state buffer is
1322 * used by every command buffer so, if the kernel does the relocations, it
1323 * will always be busy and the kernel will always stall. This is also
1324 * probably the fastest mechanism for doing relocations since the kernel would
1325 * have to make a full copy of all the relocations lists.
1328 relocate_cmd_buffer(struct anv_cmd_buffer
*cmd_buffer
,
1329 struct anv_execbuf
*exec
)
1331 if (!exec
->has_relocs
)
1334 static int userspace_relocs
= -1;
1335 if (userspace_relocs
< 0)
1336 userspace_relocs
= env_var_as_boolean("ANV_USERSPACE_RELOCS", true);
1337 if (!userspace_relocs
)
1340 /* First, we have to check to see whether or not we can even do the
1341 * relocation. New buffers which have never been submitted to the kernel
1342 * don't have a valid offset so we need to let the kernel do relocations so
1343 * that we can get offsets for them. On future execbuf2 calls, those
1344 * buffers will have offsets and we will be able to skip relocating.
1345 * Invalid offsets are indicated by anv_bo::offset == (uint64_t)-1.
1347 for (uint32_t i
= 0; i
< exec
->bo_count
; i
++) {
1348 if (exec
->bos
[i
]->offset
== (uint64_t)-1)
1352 /* Since surface states are shared between command buffers and we don't
1353 * know what order they will be submitted to the kernel, we don't know
1354 * what address is actually written in the surface state object at any
1355 * given time. The only option is to always relocate them.
1357 anv_reloc_list_apply(cmd_buffer
->device
, &cmd_buffer
->surface_relocs
,
1358 cmd_buffer
->device
->surface_state_pool
.block_pool
.bo
,
1359 true /* always relocate surface states */);
1361 /* Since we own all of the batch buffers, we know what values are stored
1362 * in the relocated addresses and only have to update them if the offsets
1365 struct anv_batch_bo
**bbo
;
1366 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
) {
1367 anv_reloc_list_apply(cmd_buffer
->device
,
1368 &(*bbo
)->relocs
, &(*bbo
)->bo
, false);
1371 for (uint32_t i
= 0; i
< exec
->bo_count
; i
++)
1372 exec
->objects
[i
].offset
= exec
->bos
[i
]->offset
;
1378 setup_execbuf_for_cmd_buffer(struct anv_execbuf
*execbuf
,
1379 struct anv_cmd_buffer
*cmd_buffer
)
1381 struct anv_batch
*batch
= &cmd_buffer
->batch
;
1382 struct anv_state_pool
*ss_pool
=
1383 &cmd_buffer
->device
->surface_state_pool
;
1385 adjust_relocations_from_state_pool(ss_pool
, &cmd_buffer
->surface_relocs
,
1386 cmd_buffer
->last_ss_pool_center
);
1389 if (cmd_buffer
->device
->instance
->physicalDevice
.use_softpin
) {
1390 anv_block_pool_foreach_bo(bo
, &ss_pool
->block_pool
) {
1391 result
= anv_execbuf_add_bo(execbuf
, bo
, NULL
, 0,
1392 &cmd_buffer
->device
->alloc
);
1393 if (result
!= VK_SUCCESS
)
1396 /* Add surface dependencies (BOs) to the execbuf */
1397 anv_execbuf_add_bo_set(execbuf
, cmd_buffer
->surface_relocs
.deps
, 0,
1398 &cmd_buffer
->device
->alloc
);
1400 /* Add the BOs for all memory objects */
1401 list_for_each_entry(struct anv_device_memory
, mem
,
1402 &cmd_buffer
->device
->memory_objects
, link
) {
1403 result
= anv_execbuf_add_bo(execbuf
, mem
->bo
, NULL
, 0,
1404 &cmd_buffer
->device
->alloc
);
1405 if (result
!= VK_SUCCESS
)
1409 struct anv_block_pool
*pool
;
1410 pool
= &cmd_buffer
->device
->dynamic_state_pool
.block_pool
;
1411 anv_block_pool_foreach_bo(bo
, pool
) {
1412 result
= anv_execbuf_add_bo(execbuf
, bo
, NULL
, 0,
1413 &cmd_buffer
->device
->alloc
);
1414 if (result
!= VK_SUCCESS
)
1418 pool
= &cmd_buffer
->device
->instruction_state_pool
.block_pool
;
1419 anv_block_pool_foreach_bo(bo
, pool
) {
1420 result
= anv_execbuf_add_bo(execbuf
, bo
, NULL
, 0,
1421 &cmd_buffer
->device
->alloc
);
1422 if (result
!= VK_SUCCESS
)
1426 pool
= &cmd_buffer
->device
->binding_table_pool
.block_pool
;
1427 anv_block_pool_foreach_bo(bo
, pool
) {
1428 result
= anv_execbuf_add_bo(execbuf
, bo
, NULL
, 0,
1429 &cmd_buffer
->device
->alloc
);
1430 if (result
!= VK_SUCCESS
)
1434 /* Since we aren't in the softpin case, all of our STATE_BASE_ADDRESS BOs
1435 * will get added automatically by processing relocations on the batch
1436 * buffer. We have to add the surface state BO manually because it has
1437 * relocations of its own that we need to be sure are processsed.
1439 result
= anv_execbuf_add_bo(execbuf
, ss_pool
->block_pool
.bo
,
1440 &cmd_buffer
->surface_relocs
, 0,
1441 &cmd_buffer
->device
->alloc
);
1442 if (result
!= VK_SUCCESS
)
1446 /* First, we walk over all of the bos we've seen and add them and their
1447 * relocations to the validate list.
1449 struct anv_batch_bo
**bbo
;
1450 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
) {
1451 adjust_relocations_to_state_pool(ss_pool
, &(*bbo
)->bo
, &(*bbo
)->relocs
,
1452 cmd_buffer
->last_ss_pool_center
);
1454 result
= anv_execbuf_add_bo(execbuf
, &(*bbo
)->bo
, &(*bbo
)->relocs
, 0,
1455 &cmd_buffer
->device
->alloc
);
1456 if (result
!= VK_SUCCESS
)
1460 /* Now that we've adjusted all of the surface state relocations, we need to
1461 * record the surface state pool center so future executions of the command
1462 * buffer can adjust correctly.
1464 cmd_buffer
->last_ss_pool_center
= ss_pool
->block_pool
.center_bo_offset
;
1466 struct anv_batch_bo
*first_batch_bo
=
1467 list_first_entry(&cmd_buffer
->batch_bos
, struct anv_batch_bo
, link
);
1469 /* The kernel requires that the last entry in the validation list be the
1470 * batch buffer to execute. We can simply swap the element
1471 * corresponding to the first batch_bo in the chain with the last
1472 * element in the list.
1474 if (first_batch_bo
->bo
.index
!= execbuf
->bo_count
- 1) {
1475 uint32_t idx
= first_batch_bo
->bo
.index
;
1476 uint32_t last_idx
= execbuf
->bo_count
- 1;
1478 struct drm_i915_gem_exec_object2 tmp_obj
= execbuf
->objects
[idx
];
1479 assert(execbuf
->bos
[idx
] == &first_batch_bo
->bo
);
1481 execbuf
->objects
[idx
] = execbuf
->objects
[last_idx
];
1482 execbuf
->bos
[idx
] = execbuf
->bos
[last_idx
];
1483 execbuf
->bos
[idx
]->index
= idx
;
1485 execbuf
->objects
[last_idx
] = tmp_obj
;
1486 execbuf
->bos
[last_idx
] = &first_batch_bo
->bo
;
1487 first_batch_bo
->bo
.index
= last_idx
;
1490 /* If we are pinning our BOs, we shouldn't have to relocate anything */
1491 if (cmd_buffer
->device
->instance
->physicalDevice
.use_softpin
)
1492 assert(!execbuf
->has_relocs
);
1494 /* Now we go through and fixup all of the relocation lists to point to
1495 * the correct indices in the object array. We have to do this after we
1496 * reorder the list above as some of the indices may have changed.
1498 if (execbuf
->has_relocs
) {
1499 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
)
1500 anv_cmd_buffer_process_relocs(cmd_buffer
, &(*bbo
)->relocs
);
1502 anv_cmd_buffer_process_relocs(cmd_buffer
, &cmd_buffer
->surface_relocs
);
1505 if (!cmd_buffer
->device
->info
.has_llc
) {
1506 __builtin_ia32_mfence();
1507 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
) {
1508 for (uint32_t i
= 0; i
< (*bbo
)->length
; i
+= CACHELINE_SIZE
)
1509 __builtin_ia32_clflush((*bbo
)->bo
.map
+ i
);
1513 execbuf
->execbuf
= (struct drm_i915_gem_execbuffer2
) {
1514 .buffers_ptr
= (uintptr_t) execbuf
->objects
,
1515 .buffer_count
= execbuf
->bo_count
,
1516 .batch_start_offset
= 0,
1517 .batch_len
= batch
->next
- batch
->start
,
1522 .flags
= I915_EXEC_HANDLE_LUT
| I915_EXEC_RENDER
,
1523 .rsvd1
= cmd_buffer
->device
->context_id
,
1527 if (relocate_cmd_buffer(cmd_buffer
, execbuf
)) {
1528 /* If we were able to successfully relocate everything, tell the kernel
1529 * that it can skip doing relocations. The requirement for using
1532 * 1) The addresses written in the objects must match the corresponding
1533 * reloc.presumed_offset which in turn must match the corresponding
1534 * execobject.offset.
1536 * 2) To avoid stalling, execobject.offset should match the current
1537 * address of that object within the active context.
1539 * In order to satisfy all of the invariants that make userspace
1540 * relocations to be safe (see relocate_cmd_buffer()), we need to
1541 * further ensure that the addresses we use match those used by the
1542 * kernel for the most recent execbuf2.
1544 * The kernel may still choose to do relocations anyway if something has
1545 * moved in the GTT. In this case, the relocation list still needs to be
1546 * valid. All relocations on the batch buffers are already valid and
1547 * kept up-to-date. For surface state relocations, by applying the
1548 * relocations in relocate_cmd_buffer, we ensured that the address in
1549 * the RENDER_SURFACE_STATE matches presumed_offset, so it should be
1550 * safe for the kernel to relocate them as needed.
1552 execbuf
->execbuf
.flags
|= I915_EXEC_NO_RELOC
;
1554 /* In the case where we fall back to doing kernel relocations, we need
1555 * to ensure that the relocation list is valid. All relocations on the
1556 * batch buffers are already valid and kept up-to-date. Since surface
1557 * states are shared between command buffers and we don't know what
1558 * order they will be submitted to the kernel, we don't know what
1559 * address is actually written in the surface state object at any given
1560 * time. The only option is to set a bogus presumed offset and let the
1561 * kernel relocate them.
1563 for (size_t i
= 0; i
< cmd_buffer
->surface_relocs
.num_relocs
; i
++)
1564 cmd_buffer
->surface_relocs
.relocs
[i
].presumed_offset
= -1;
1571 setup_empty_execbuf(struct anv_execbuf
*execbuf
, struct anv_device
*device
)
1573 VkResult result
= anv_execbuf_add_bo(execbuf
, &device
->trivial_batch_bo
,
1574 NULL
, 0, &device
->alloc
);
1575 if (result
!= VK_SUCCESS
)
1578 execbuf
->execbuf
= (struct drm_i915_gem_execbuffer2
) {
1579 .buffers_ptr
= (uintptr_t) execbuf
->objects
,
1580 .buffer_count
= execbuf
->bo_count
,
1581 .batch_start_offset
= 0,
1582 .batch_len
= 8, /* GEN7_MI_BATCH_BUFFER_END and NOOP */
1583 .flags
= I915_EXEC_HANDLE_LUT
| I915_EXEC_RENDER
,
1584 .rsvd1
= device
->context_id
,
1592 anv_cmd_buffer_execbuf(struct anv_device
*device
,
1593 struct anv_cmd_buffer
*cmd_buffer
,
1594 const VkSemaphore
*in_semaphores
,
1595 uint32_t num_in_semaphores
,
1596 const VkSemaphore
*out_semaphores
,
1597 uint32_t num_out_semaphores
,
1600 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
1601 UNUSED
struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1603 struct anv_execbuf execbuf
;
1604 anv_execbuf_init(&execbuf
);
1607 VkResult result
= VK_SUCCESS
;
1608 for (uint32_t i
= 0; i
< num_in_semaphores
; i
++) {
1609 ANV_FROM_HANDLE(anv_semaphore
, semaphore
, in_semaphores
[i
]);
1610 struct anv_semaphore_impl
*impl
=
1611 semaphore
->temporary
.type
!= ANV_SEMAPHORE_TYPE_NONE
?
1612 &semaphore
->temporary
: &semaphore
->permanent
;
1614 switch (impl
->type
) {
1615 case ANV_SEMAPHORE_TYPE_BO
:
1616 assert(!pdevice
->has_syncobj
);
1617 result
= anv_execbuf_add_bo(&execbuf
, impl
->bo
, NULL
,
1619 if (result
!= VK_SUCCESS
)
1623 case ANV_SEMAPHORE_TYPE_SYNC_FILE
:
1624 assert(!pdevice
->has_syncobj
);
1625 if (in_fence
== -1) {
1626 in_fence
= impl
->fd
;
1628 int merge
= anv_gem_sync_file_merge(device
, in_fence
, impl
->fd
);
1630 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1640 case ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ
:
1641 result
= anv_execbuf_add_syncobj(&execbuf
, impl
->syncobj
,
1642 I915_EXEC_FENCE_WAIT
,
1644 if (result
!= VK_SUCCESS
)
1653 bool need_out_fence
= false;
1654 for (uint32_t i
= 0; i
< num_out_semaphores
; i
++) {
1655 ANV_FROM_HANDLE(anv_semaphore
, semaphore
, out_semaphores
[i
]);
1657 /* Under most circumstances, out fences won't be temporary. However,
1658 * the spec does allow it for opaque_fd. From the Vulkan 1.0.53 spec:
1660 * "If the import is temporary, the implementation must restore the
1661 * semaphore to its prior permanent state after submitting the next
1662 * semaphore wait operation."
1664 * The spec says nothing whatsoever about signal operations on
1665 * temporarily imported semaphores so it appears they are allowed.
1666 * There are also CTS tests that require this to work.
1668 struct anv_semaphore_impl
*impl
=
1669 semaphore
->temporary
.type
!= ANV_SEMAPHORE_TYPE_NONE
?
1670 &semaphore
->temporary
: &semaphore
->permanent
;
1672 switch (impl
->type
) {
1673 case ANV_SEMAPHORE_TYPE_BO
:
1674 assert(!pdevice
->has_syncobj
);
1675 result
= anv_execbuf_add_bo(&execbuf
, impl
->bo
, NULL
,
1676 EXEC_OBJECT_WRITE
, &device
->alloc
);
1677 if (result
!= VK_SUCCESS
)
1681 case ANV_SEMAPHORE_TYPE_SYNC_FILE
:
1682 assert(!pdevice
->has_syncobj
);
1683 need_out_fence
= true;
1686 case ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ
:
1687 result
= anv_execbuf_add_syncobj(&execbuf
, impl
->syncobj
,
1688 I915_EXEC_FENCE_SIGNAL
,
1690 if (result
!= VK_SUCCESS
)
1700 /* Under most circumstances, out fences won't be temporary. However,
1701 * the spec does allow it for opaque_fd. From the Vulkan 1.0.53 spec:
1703 * "If the import is temporary, the implementation must restore the
1704 * semaphore to its prior permanent state after submitting the next
1705 * semaphore wait operation."
1707 * The spec says nothing whatsoever about signal operations on
1708 * temporarily imported semaphores so it appears they are allowed.
1709 * There are also CTS tests that require this to work.
1711 struct anv_fence_impl
*impl
=
1712 fence
->temporary
.type
!= ANV_FENCE_TYPE_NONE
?
1713 &fence
->temporary
: &fence
->permanent
;
1715 switch (impl
->type
) {
1716 case ANV_FENCE_TYPE_BO
:
1717 assert(!pdevice
->has_syncobj_wait
);
1718 result
= anv_execbuf_add_bo(&execbuf
, &impl
->bo
.bo
, NULL
,
1719 EXEC_OBJECT_WRITE
, &device
->alloc
);
1720 if (result
!= VK_SUCCESS
)
1724 case ANV_FENCE_TYPE_SYNCOBJ
:
1725 result
= anv_execbuf_add_syncobj(&execbuf
, impl
->syncobj
,
1726 I915_EXEC_FENCE_SIGNAL
,
1728 if (result
!= VK_SUCCESS
)
1733 unreachable("Invalid fence type");
1738 if (unlikely(INTEL_DEBUG
& DEBUG_BATCH
)) {
1739 struct anv_batch_bo
**bo
= u_vector_tail(&cmd_buffer
->seen_bbos
);
1741 device
->cmd_buffer_being_decoded
= cmd_buffer
;
1742 gen_print_batch(&device
->decoder_ctx
, (*bo
)->bo
.map
,
1743 (*bo
)->bo
.size
, (*bo
)->bo
.offset
, false);
1744 device
->cmd_buffer_being_decoded
= NULL
;
1747 result
= setup_execbuf_for_cmd_buffer(&execbuf
, cmd_buffer
);
1749 result
= setup_empty_execbuf(&execbuf
, device
);
1752 if (result
!= VK_SUCCESS
)
1755 if (execbuf
.fence_count
> 0) {
1756 assert(device
->instance
->physicalDevice
.has_syncobj
);
1757 execbuf
.execbuf
.flags
|= I915_EXEC_FENCE_ARRAY
;
1758 execbuf
.execbuf
.num_cliprects
= execbuf
.fence_count
;
1759 execbuf
.execbuf
.cliprects_ptr
= (uintptr_t) execbuf
.fences
;
1762 if (in_fence
!= -1) {
1763 execbuf
.execbuf
.flags
|= I915_EXEC_FENCE_IN
;
1764 execbuf
.execbuf
.rsvd2
|= (uint32_t)in_fence
;
1768 execbuf
.execbuf
.flags
|= I915_EXEC_FENCE_OUT
;
1770 result
= anv_device_execbuf(device
, &execbuf
.execbuf
, execbuf
.bos
);
1772 /* Execbuf does not consume the in_fence. It's our job to close it. */
1776 for (uint32_t i
= 0; i
< num_in_semaphores
; i
++) {
1777 ANV_FROM_HANDLE(anv_semaphore
, semaphore
, in_semaphores
[i
]);
1778 /* From the Vulkan 1.0.53 spec:
1780 * "If the import is temporary, the implementation must restore the
1781 * semaphore to its prior permanent state after submitting the next
1782 * semaphore wait operation."
1784 * This has to happen after the execbuf in case we close any syncobjs in
1787 anv_semaphore_reset_temporary(device
, semaphore
);
1790 if (fence
&& fence
->permanent
.type
== ANV_FENCE_TYPE_BO
) {
1791 assert(!pdevice
->has_syncobj_wait
);
1792 /* BO fences can't be shared, so they can't be temporary. */
1793 assert(fence
->temporary
.type
== ANV_FENCE_TYPE_NONE
);
1795 /* Once the execbuf has returned, we need to set the fence state to
1796 * SUBMITTED. We can't do this before calling execbuf because
1797 * anv_GetFenceStatus does take the global device lock before checking
1800 * We set the fence state to SUBMITTED regardless of whether or not the
1801 * execbuf succeeds because we need to ensure that vkWaitForFences() and
1802 * vkGetFenceStatus() return a valid result (VK_ERROR_DEVICE_LOST or
1803 * VK_SUCCESS) in a finite amount of time even if execbuf fails.
1805 fence
->permanent
.bo
.state
= ANV_BO_FENCE_STATE_SUBMITTED
;
1808 if (result
== VK_SUCCESS
&& need_out_fence
) {
1809 assert(!pdevice
->has_syncobj_wait
);
1810 int out_fence
= execbuf
.execbuf
.rsvd2
>> 32;
1811 for (uint32_t i
= 0; i
< num_out_semaphores
; i
++) {
1812 ANV_FROM_HANDLE(anv_semaphore
, semaphore
, out_semaphores
[i
]);
1813 /* Out fences can't have temporary state because that would imply
1814 * that we imported a sync file and are trying to signal it.
1816 assert(semaphore
->temporary
.type
== ANV_SEMAPHORE_TYPE_NONE
);
1817 struct anv_semaphore_impl
*impl
= &semaphore
->permanent
;
1819 if (impl
->type
== ANV_SEMAPHORE_TYPE_SYNC_FILE
) {
1820 assert(impl
->fd
== -1);
1821 impl
->fd
= dup(out_fence
);
1827 anv_execbuf_finish(&execbuf
, &device
->alloc
);