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(16, list
->array_length
* 2);
122 while (new_length
< list
->num_relocs
+ num_additional_relocs
)
125 struct drm_i915_gem_relocation_entry
*new_relocs
=
126 vk_realloc(alloc
, list
->relocs
,
127 new_length
* sizeof(*list
->relocs
), 8,
128 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
129 if (new_relocs
== NULL
)
130 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
131 list
->relocs
= new_relocs
;
133 struct anv_bo
**new_reloc_bos
=
134 vk_realloc(alloc
, list
->reloc_bos
,
135 new_length
* sizeof(*list
->reloc_bos
), 8,
136 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
137 if (new_reloc_bos
== NULL
)
138 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
139 list
->reloc_bos
= new_reloc_bos
;
141 list
->array_length
= new_length
;
146 #define READ_ONCE(x) (*(volatile __typeof__(x) *)&(x))
149 anv_reloc_list_add(struct anv_reloc_list
*list
,
150 const VkAllocationCallbacks
*alloc
,
151 uint32_t offset
, struct anv_bo
*target_bo
, uint32_t delta
,
152 uint64_t *address_u64_out
)
154 struct drm_i915_gem_relocation_entry
*entry
;
157 uint64_t target_bo_offset
= READ_ONCE(target_bo
->offset
);
159 *address_u64_out
= target_bo_offset
+ delta
;
161 if (target_bo
->flags
& EXEC_OBJECT_PINNED
) {
162 if (list
->deps
== NULL
) {
163 list
->deps
= _mesa_pointer_set_create(NULL
);
164 if (unlikely(list
->deps
== NULL
))
165 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
167 _mesa_set_add(list
->deps
, target_bo
);
171 VkResult result
= anv_reloc_list_grow(list
, alloc
, 1);
172 if (result
!= VK_SUCCESS
)
175 /* XXX: Can we use I915_EXEC_HANDLE_LUT? */
176 index
= list
->num_relocs
++;
177 list
->reloc_bos
[index
] = target_bo
;
178 entry
= &list
->relocs
[index
];
179 entry
->target_handle
= -1; /* See also anv_cmd_buffer_process_relocs() */
180 entry
->delta
= delta
;
181 entry
->offset
= offset
;
182 entry
->presumed_offset
= target_bo_offset
;
183 entry
->read_domains
= 0;
184 entry
->write_domain
= 0;
185 VG(VALGRIND_CHECK_MEM_IS_DEFINED(entry
, sizeof(*entry
)));
191 anv_reloc_list_append(struct anv_reloc_list
*list
,
192 const VkAllocationCallbacks
*alloc
,
193 struct anv_reloc_list
*other
, uint32_t offset
)
195 VkResult result
= anv_reloc_list_grow(list
, alloc
, other
->num_relocs
);
196 if (result
!= VK_SUCCESS
)
199 if (other
->num_relocs
> 0) {
200 memcpy(&list
->relocs
[list
->num_relocs
], &other
->relocs
[0],
201 other
->num_relocs
* sizeof(other
->relocs
[0]));
202 memcpy(&list
->reloc_bos
[list
->num_relocs
], &other
->reloc_bos
[0],
203 other
->num_relocs
* sizeof(other
->reloc_bos
[0]));
205 for (uint32_t i
= 0; i
< other
->num_relocs
; i
++)
206 list
->relocs
[i
+ list
->num_relocs
].offset
+= offset
;
208 list
->num_relocs
+= other
->num_relocs
;
212 if (list
->deps
== NULL
) {
213 list
->deps
= _mesa_pointer_set_create(NULL
);
214 if (unlikely(list
->deps
== NULL
))
215 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
217 set_foreach(other
->deps
, entry
)
218 _mesa_set_add_pre_hashed(list
->deps
, entry
->hash
, entry
->key
);
224 /*-----------------------------------------------------------------------*
225 * Functions related to anv_batch
226 *-----------------------------------------------------------------------*/
229 anv_batch_emit_dwords(struct anv_batch
*batch
, int num_dwords
)
231 if (batch
->next
+ num_dwords
* 4 > batch
->end
) {
232 VkResult result
= batch
->extend_cb(batch
, batch
->user_data
);
233 if (result
!= VK_SUCCESS
) {
234 anv_batch_set_error(batch
, result
);
239 void *p
= batch
->next
;
241 batch
->next
+= num_dwords
* 4;
242 assert(batch
->next
<= batch
->end
);
248 anv_batch_emit_reloc(struct anv_batch
*batch
,
249 void *location
, struct anv_bo
*bo
, uint32_t delta
)
251 uint64_t address_u64
= 0;
252 VkResult result
= anv_reloc_list_add(batch
->relocs
, batch
->alloc
,
253 location
- batch
->start
, bo
, delta
,
255 if (result
!= VK_SUCCESS
) {
256 anv_batch_set_error(batch
, result
);
264 anv_batch_emit_batch(struct anv_batch
*batch
, struct anv_batch
*other
)
266 uint32_t size
, offset
;
268 size
= other
->next
- other
->start
;
269 assert(size
% 4 == 0);
271 if (batch
->next
+ size
> batch
->end
) {
272 VkResult result
= batch
->extend_cb(batch
, batch
->user_data
);
273 if (result
!= VK_SUCCESS
) {
274 anv_batch_set_error(batch
, result
);
279 assert(batch
->next
+ size
<= batch
->end
);
281 VG(VALGRIND_CHECK_MEM_IS_DEFINED(other
->start
, size
));
282 memcpy(batch
->next
, other
->start
, size
);
284 offset
= batch
->next
- batch
->start
;
285 VkResult result
= anv_reloc_list_append(batch
->relocs
, batch
->alloc
,
286 other
->relocs
, offset
);
287 if (result
!= VK_SUCCESS
) {
288 anv_batch_set_error(batch
, result
);
295 /*-----------------------------------------------------------------------*
296 * Functions related to anv_batch_bo
297 *-----------------------------------------------------------------------*/
300 anv_batch_bo_create(struct anv_cmd_buffer
*cmd_buffer
,
301 struct anv_batch_bo
**bbo_out
)
305 struct anv_batch_bo
*bbo
= vk_alloc(&cmd_buffer
->pool
->alloc
, sizeof(*bbo
),
306 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
308 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
310 result
= anv_bo_pool_alloc(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
,
311 ANV_CMD_BUFFER_BATCH_SIZE
);
312 if (result
!= VK_SUCCESS
)
315 result
= anv_reloc_list_init(&bbo
->relocs
, &cmd_buffer
->pool
->alloc
);
316 if (result
!= VK_SUCCESS
)
324 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
326 vk_free(&cmd_buffer
->pool
->alloc
, bbo
);
332 anv_batch_bo_clone(struct anv_cmd_buffer
*cmd_buffer
,
333 const struct anv_batch_bo
*other_bbo
,
334 struct anv_batch_bo
**bbo_out
)
338 struct anv_batch_bo
*bbo
= vk_alloc(&cmd_buffer
->pool
->alloc
, sizeof(*bbo
),
339 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
341 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
343 result
= anv_bo_pool_alloc(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
,
345 if (result
!= VK_SUCCESS
)
348 result
= anv_reloc_list_init_clone(&bbo
->relocs
, &cmd_buffer
->pool
->alloc
,
350 if (result
!= VK_SUCCESS
)
353 bbo
->length
= other_bbo
->length
;
354 memcpy(bbo
->bo
.map
, other_bbo
->bo
.map
, other_bbo
->length
);
361 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
363 vk_free(&cmd_buffer
->pool
->alloc
, bbo
);
369 anv_batch_bo_start(struct anv_batch_bo
*bbo
, struct anv_batch
*batch
,
370 size_t batch_padding
)
372 batch
->next
= batch
->start
= bbo
->bo
.map
;
373 batch
->end
= bbo
->bo
.map
+ bbo
->bo
.size
- batch_padding
;
374 batch
->relocs
= &bbo
->relocs
;
375 bbo
->relocs
.num_relocs
= 0;
376 _mesa_set_clear(bbo
->relocs
.deps
, NULL
);
380 anv_batch_bo_continue(struct anv_batch_bo
*bbo
, struct anv_batch
*batch
,
381 size_t batch_padding
)
383 batch
->start
= bbo
->bo
.map
;
384 batch
->next
= bbo
->bo
.map
+ bbo
->length
;
385 batch
->end
= bbo
->bo
.map
+ bbo
->bo
.size
- batch_padding
;
386 batch
->relocs
= &bbo
->relocs
;
390 anv_batch_bo_finish(struct anv_batch_bo
*bbo
, struct anv_batch
*batch
)
392 assert(batch
->start
== bbo
->bo
.map
);
393 bbo
->length
= batch
->next
- batch
->start
;
394 VG(VALGRIND_CHECK_MEM_IS_DEFINED(batch
->start
, bbo
->length
));
398 anv_batch_bo_grow(struct anv_cmd_buffer
*cmd_buffer
, struct anv_batch_bo
*bbo
,
399 struct anv_batch
*batch
, size_t aditional
,
400 size_t batch_padding
)
402 assert(batch
->start
== bbo
->bo
.map
);
403 bbo
->length
= batch
->next
- batch
->start
;
405 size_t new_size
= bbo
->bo
.size
;
406 while (new_size
<= bbo
->length
+ aditional
+ batch_padding
)
409 if (new_size
== bbo
->bo
.size
)
412 struct anv_bo new_bo
;
413 VkResult result
= anv_bo_pool_alloc(&cmd_buffer
->device
->batch_bo_pool
,
415 if (result
!= VK_SUCCESS
)
418 memcpy(new_bo
.map
, bbo
->bo
.map
, bbo
->length
);
420 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
423 anv_batch_bo_continue(bbo
, batch
, batch_padding
);
429 anv_batch_bo_link(struct anv_cmd_buffer
*cmd_buffer
,
430 struct anv_batch_bo
*prev_bbo
,
431 struct anv_batch_bo
*next_bbo
,
432 uint32_t next_bbo_offset
)
434 const uint32_t bb_start_offset
=
435 prev_bbo
->length
- GEN8_MI_BATCH_BUFFER_START_length
* 4;
436 ASSERTED
const uint32_t *bb_start
= prev_bbo
->bo
.map
+ bb_start_offset
;
438 /* Make sure we're looking at a MI_BATCH_BUFFER_START */
439 assert(((*bb_start
>> 29) & 0x07) == 0);
440 assert(((*bb_start
>> 23) & 0x3f) == 49);
442 if (cmd_buffer
->device
->instance
->physicalDevice
.use_softpin
) {
443 assert(prev_bbo
->bo
.flags
& EXEC_OBJECT_PINNED
);
444 assert(next_bbo
->bo
.flags
& EXEC_OBJECT_PINNED
);
446 write_reloc(cmd_buffer
->device
,
447 prev_bbo
->bo
.map
+ bb_start_offset
+ 4,
448 next_bbo
->bo
.offset
+ next_bbo_offset
, true);
450 uint32_t reloc_idx
= prev_bbo
->relocs
.num_relocs
- 1;
451 assert(prev_bbo
->relocs
.relocs
[reloc_idx
].offset
== bb_start_offset
+ 4);
453 prev_bbo
->relocs
.reloc_bos
[reloc_idx
] = &next_bbo
->bo
;
454 prev_bbo
->relocs
.relocs
[reloc_idx
].delta
= next_bbo_offset
;
456 /* Use a bogus presumed offset to force a relocation */
457 prev_bbo
->relocs
.relocs
[reloc_idx
].presumed_offset
= -1;
462 anv_batch_bo_destroy(struct anv_batch_bo
*bbo
,
463 struct anv_cmd_buffer
*cmd_buffer
)
465 anv_reloc_list_finish(&bbo
->relocs
, &cmd_buffer
->pool
->alloc
);
466 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
467 vk_free(&cmd_buffer
->pool
->alloc
, bbo
);
471 anv_batch_bo_list_clone(const struct list_head
*list
,
472 struct anv_cmd_buffer
*cmd_buffer
,
473 struct list_head
*new_list
)
475 VkResult result
= VK_SUCCESS
;
477 list_inithead(new_list
);
479 struct anv_batch_bo
*prev_bbo
= NULL
;
480 list_for_each_entry(struct anv_batch_bo
, bbo
, list
, link
) {
481 struct anv_batch_bo
*new_bbo
= NULL
;
482 result
= anv_batch_bo_clone(cmd_buffer
, bbo
, &new_bbo
);
483 if (result
!= VK_SUCCESS
)
485 list_addtail(&new_bbo
->link
, new_list
);
488 anv_batch_bo_link(cmd_buffer
, prev_bbo
, new_bbo
, 0);
493 if (result
!= VK_SUCCESS
) {
494 list_for_each_entry_safe(struct anv_batch_bo
, bbo
, new_list
, link
)
495 anv_batch_bo_destroy(bbo
, cmd_buffer
);
501 /*-----------------------------------------------------------------------*
502 * Functions related to anv_batch_bo
503 *-----------------------------------------------------------------------*/
505 static struct anv_batch_bo
*
506 anv_cmd_buffer_current_batch_bo(struct anv_cmd_buffer
*cmd_buffer
)
508 return LIST_ENTRY(struct anv_batch_bo
, cmd_buffer
->batch_bos
.prev
, link
);
512 anv_cmd_buffer_surface_base_address(struct anv_cmd_buffer
*cmd_buffer
)
514 struct anv_state
*bt_block
= u_vector_head(&cmd_buffer
->bt_block_states
);
515 return (struct anv_address
) {
516 .bo
= anv_binding_table_pool(cmd_buffer
->device
)->block_pool
.bo
,
517 .offset
= bt_block
->offset
,
522 emit_batch_buffer_start(struct anv_cmd_buffer
*cmd_buffer
,
523 struct anv_bo
*bo
, uint32_t offset
)
525 /* In gen8+ the address field grew to two dwords to accomodate 48 bit
526 * offsets. The high 16 bits are in the last dword, so we can use the gen8
527 * version in either case, as long as we set the instruction length in the
528 * header accordingly. This means that we always emit three dwords here
529 * and all the padding and adjustment we do in this file works for all
533 #define GEN7_MI_BATCH_BUFFER_START_length 2
534 #define GEN7_MI_BATCH_BUFFER_START_length_bias 2
536 const uint32_t gen7_length
=
537 GEN7_MI_BATCH_BUFFER_START_length
- GEN7_MI_BATCH_BUFFER_START_length_bias
;
538 const uint32_t gen8_length
=
539 GEN8_MI_BATCH_BUFFER_START_length
- GEN8_MI_BATCH_BUFFER_START_length_bias
;
541 anv_batch_emit(&cmd_buffer
->batch
, GEN8_MI_BATCH_BUFFER_START
, bbs
) {
542 bbs
.DWordLength
= cmd_buffer
->device
->info
.gen
< 8 ?
543 gen7_length
: gen8_length
;
544 bbs
.SecondLevelBatchBuffer
= Firstlevelbatch
;
545 bbs
.AddressSpaceIndicator
= ASI_PPGTT
;
546 bbs
.BatchBufferStartAddress
= (struct anv_address
) { bo
, offset
};
551 cmd_buffer_chain_to_batch_bo(struct anv_cmd_buffer
*cmd_buffer
,
552 struct anv_batch_bo
*bbo
)
554 struct anv_batch
*batch
= &cmd_buffer
->batch
;
555 struct anv_batch_bo
*current_bbo
=
556 anv_cmd_buffer_current_batch_bo(cmd_buffer
);
558 /* We set the end of the batch a little short so we would be sure we
559 * have room for the chaining command. Since we're about to emit the
560 * chaining command, let's set it back where it should go.
562 batch
->end
+= GEN8_MI_BATCH_BUFFER_START_length
* 4;
563 assert(batch
->end
== current_bbo
->bo
.map
+ current_bbo
->bo
.size
);
565 emit_batch_buffer_start(cmd_buffer
, &bbo
->bo
, 0);
567 anv_batch_bo_finish(current_bbo
, batch
);
571 anv_cmd_buffer_chain_batch(struct anv_batch
*batch
, void *_data
)
573 struct anv_cmd_buffer
*cmd_buffer
= _data
;
574 struct anv_batch_bo
*new_bbo
;
576 VkResult result
= anv_batch_bo_create(cmd_buffer
, &new_bbo
);
577 if (result
!= VK_SUCCESS
)
580 struct anv_batch_bo
**seen_bbo
= u_vector_add(&cmd_buffer
->seen_bbos
);
581 if (seen_bbo
== NULL
) {
582 anv_batch_bo_destroy(new_bbo
, cmd_buffer
);
583 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
587 cmd_buffer_chain_to_batch_bo(cmd_buffer
, new_bbo
);
589 list_addtail(&new_bbo
->link
, &cmd_buffer
->batch_bos
);
591 anv_batch_bo_start(new_bbo
, batch
, GEN8_MI_BATCH_BUFFER_START_length
* 4);
597 anv_cmd_buffer_grow_batch(struct anv_batch
*batch
, void *_data
)
599 struct anv_cmd_buffer
*cmd_buffer
= _data
;
600 struct anv_batch_bo
*bbo
= anv_cmd_buffer_current_batch_bo(cmd_buffer
);
602 anv_batch_bo_grow(cmd_buffer
, bbo
, &cmd_buffer
->batch
, 4096,
603 GEN8_MI_BATCH_BUFFER_START_length
* 4);
608 /** Allocate a binding table
610 * This function allocates a binding table. This is a bit more complicated
611 * than one would think due to a combination of Vulkan driver design and some
612 * unfortunate hardware restrictions.
614 * The 3DSTATE_BINDING_TABLE_POINTERS_* packets only have a 16-bit field for
615 * the binding table pointer which means that all binding tables need to live
616 * in the bottom 64k of surface state base address. The way the GL driver has
617 * classically dealt with this restriction is to emit all surface states
618 * on-the-fly into the batch and have a batch buffer smaller than 64k. This
619 * isn't really an option in Vulkan for a couple of reasons:
621 * 1) In Vulkan, we have growing (or chaining) batches so surface states have
622 * to live in their own buffer and we have to be able to re-emit
623 * STATE_BASE_ADDRESS as needed which requires a full pipeline stall. In
624 * order to avoid emitting STATE_BASE_ADDRESS any more often than needed
625 * (it's not that hard to hit 64k of just binding tables), we allocate
626 * surface state objects up-front when VkImageView is created. In order
627 * for this to work, surface state objects need to be allocated from a
630 * 2) We tried to design the surface state system in such a way that it's
631 * already ready for bindless texturing. The way bindless texturing works
632 * on our hardware is that you have a big pool of surface state objects
633 * (with its own state base address) and the bindless handles are simply
634 * offsets into that pool. With the architecture we chose, we already
635 * have that pool and it's exactly the same pool that we use for regular
636 * surface states so we should already be ready for bindless.
638 * 3) For render targets, we need to be able to fill out the surface states
639 * later in vkBeginRenderPass so that we can assign clear colors
640 * correctly. One way to do this would be to just create the surface
641 * state data and then repeatedly copy it into the surface state BO every
642 * time we have to re-emit STATE_BASE_ADDRESS. While this works, it's
643 * rather annoying and just being able to allocate them up-front and
644 * re-use them for the entire render pass.
646 * While none of these are technically blockers for emitting state on the fly
647 * like we do in GL, the ability to have a single surface state pool is
648 * simplifies things greatly. Unfortunately, it comes at a cost...
650 * Because of the 64k limitation of 3DSTATE_BINDING_TABLE_POINTERS_*, we can't
651 * place the binding tables just anywhere in surface state base address.
652 * Because 64k isn't a whole lot of space, we can't simply restrict the
653 * surface state buffer to 64k, we have to be more clever. The solution we've
654 * chosen is to have a block pool with a maximum size of 2G that starts at
655 * zero and grows in both directions. All surface states are allocated from
656 * the top of the pool (positive offsets) and we allocate blocks (< 64k) of
657 * binding tables from the bottom of the pool (negative offsets). Every time
658 * we allocate a new binding table block, we set surface state base address to
659 * point to the bottom of the binding table block. This way all of the
660 * binding tables in the block are in the bottom 64k of surface state base
661 * address. When we fill out the binding table, we add the distance between
662 * the bottom of our binding table block and zero of the block pool to the
663 * surface state offsets so that they are correct relative to out new surface
664 * state base address at the bottom of the binding table block.
666 * \see adjust_relocations_from_block_pool()
667 * \see adjust_relocations_too_block_pool()
669 * \param[in] entries The number of surface state entries the binding
670 * table should be able to hold.
672 * \param[out] state_offset The offset surface surface state base address
673 * where the surface states live. This must be
674 * added to the surface state offset when it is
675 * written into the binding table entry.
677 * \return An anv_state representing the binding table
680 anv_cmd_buffer_alloc_binding_table(struct anv_cmd_buffer
*cmd_buffer
,
681 uint32_t entries
, uint32_t *state_offset
)
683 struct anv_device
*device
= cmd_buffer
->device
;
684 struct anv_state_pool
*state_pool
= &device
->surface_state_pool
;
685 struct anv_state
*bt_block
= u_vector_head(&cmd_buffer
->bt_block_states
);
686 struct anv_state state
;
688 state
.alloc_size
= align_u32(entries
* 4, 32);
690 if (cmd_buffer
->bt_next
+ state
.alloc_size
> state_pool
->block_size
)
691 return (struct anv_state
) { 0 };
693 state
.offset
= cmd_buffer
->bt_next
;
694 state
.map
= anv_block_pool_map(&anv_binding_table_pool(device
)->block_pool
,
695 bt_block
->offset
+ state
.offset
);
697 cmd_buffer
->bt_next
+= state
.alloc_size
;
699 if (device
->instance
->physicalDevice
.use_softpin
) {
700 assert(bt_block
->offset
>= 0);
701 *state_offset
= device
->surface_state_pool
.block_pool
.start_address
-
702 device
->binding_table_pool
.block_pool
.start_address
- bt_block
->offset
;
704 assert(bt_block
->offset
< 0);
705 *state_offset
= -bt_block
->offset
;
712 anv_cmd_buffer_alloc_surface_state(struct anv_cmd_buffer
*cmd_buffer
)
714 struct isl_device
*isl_dev
= &cmd_buffer
->device
->isl_dev
;
715 return anv_state_stream_alloc(&cmd_buffer
->surface_state_stream
,
716 isl_dev
->ss
.size
, isl_dev
->ss
.align
);
720 anv_cmd_buffer_alloc_dynamic_state(struct anv_cmd_buffer
*cmd_buffer
,
721 uint32_t size
, uint32_t alignment
)
723 return anv_state_stream_alloc(&cmd_buffer
->dynamic_state_stream
,
728 anv_cmd_buffer_new_binding_table_block(struct anv_cmd_buffer
*cmd_buffer
)
730 struct anv_state
*bt_block
= u_vector_add(&cmd_buffer
->bt_block_states
);
731 if (bt_block
== NULL
) {
732 anv_batch_set_error(&cmd_buffer
->batch
, VK_ERROR_OUT_OF_HOST_MEMORY
);
733 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
736 *bt_block
= anv_binding_table_pool_alloc(cmd_buffer
->device
);
737 cmd_buffer
->bt_next
= 0;
743 anv_cmd_buffer_init_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
)
745 struct anv_batch_bo
*batch_bo
;
748 list_inithead(&cmd_buffer
->batch_bos
);
750 result
= anv_batch_bo_create(cmd_buffer
, &batch_bo
);
751 if (result
!= VK_SUCCESS
)
754 list_addtail(&batch_bo
->link
, &cmd_buffer
->batch_bos
);
756 cmd_buffer
->batch
.alloc
= &cmd_buffer
->pool
->alloc
;
757 cmd_buffer
->batch
.user_data
= cmd_buffer
;
759 if (cmd_buffer
->device
->can_chain_batches
) {
760 cmd_buffer
->batch
.extend_cb
= anv_cmd_buffer_chain_batch
;
762 cmd_buffer
->batch
.extend_cb
= anv_cmd_buffer_grow_batch
;
765 anv_batch_bo_start(batch_bo
, &cmd_buffer
->batch
,
766 GEN8_MI_BATCH_BUFFER_START_length
* 4);
768 int success
= u_vector_init(&cmd_buffer
->seen_bbos
,
769 sizeof(struct anv_bo
*),
770 8 * sizeof(struct anv_bo
*));
774 *(struct anv_batch_bo
**)u_vector_add(&cmd_buffer
->seen_bbos
) = batch_bo
;
776 /* u_vector requires power-of-two size elements */
777 unsigned pow2_state_size
= util_next_power_of_two(sizeof(struct anv_state
));
778 success
= u_vector_init(&cmd_buffer
->bt_block_states
,
779 pow2_state_size
, 8 * pow2_state_size
);
783 result
= anv_reloc_list_init(&cmd_buffer
->surface_relocs
,
784 &cmd_buffer
->pool
->alloc
);
785 if (result
!= VK_SUCCESS
)
787 cmd_buffer
->last_ss_pool_center
= 0;
789 result
= anv_cmd_buffer_new_binding_table_block(cmd_buffer
);
790 if (result
!= VK_SUCCESS
)
796 u_vector_finish(&cmd_buffer
->bt_block_states
);
798 u_vector_finish(&cmd_buffer
->seen_bbos
);
800 anv_batch_bo_destroy(batch_bo
, cmd_buffer
);
806 anv_cmd_buffer_fini_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
)
808 struct anv_state
*bt_block
;
809 u_vector_foreach(bt_block
, &cmd_buffer
->bt_block_states
)
810 anv_binding_table_pool_free(cmd_buffer
->device
, *bt_block
);
811 u_vector_finish(&cmd_buffer
->bt_block_states
);
813 anv_reloc_list_finish(&cmd_buffer
->surface_relocs
, &cmd_buffer
->pool
->alloc
);
815 u_vector_finish(&cmd_buffer
->seen_bbos
);
817 /* Destroy all of the batch buffers */
818 list_for_each_entry_safe(struct anv_batch_bo
, bbo
,
819 &cmd_buffer
->batch_bos
, link
) {
820 anv_batch_bo_destroy(bbo
, cmd_buffer
);
825 anv_cmd_buffer_reset_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
)
827 /* Delete all but the first batch bo */
828 assert(!list_is_empty(&cmd_buffer
->batch_bos
));
829 while (cmd_buffer
->batch_bos
.next
!= cmd_buffer
->batch_bos
.prev
) {
830 struct anv_batch_bo
*bbo
= anv_cmd_buffer_current_batch_bo(cmd_buffer
);
831 list_del(&bbo
->link
);
832 anv_batch_bo_destroy(bbo
, cmd_buffer
);
834 assert(!list_is_empty(&cmd_buffer
->batch_bos
));
836 anv_batch_bo_start(anv_cmd_buffer_current_batch_bo(cmd_buffer
),
838 GEN8_MI_BATCH_BUFFER_START_length
* 4);
840 while (u_vector_length(&cmd_buffer
->bt_block_states
) > 1) {
841 struct anv_state
*bt_block
= u_vector_remove(&cmd_buffer
->bt_block_states
);
842 anv_binding_table_pool_free(cmd_buffer
->device
, *bt_block
);
844 assert(u_vector_length(&cmd_buffer
->bt_block_states
) == 1);
845 cmd_buffer
->bt_next
= 0;
847 cmd_buffer
->surface_relocs
.num_relocs
= 0;
848 _mesa_set_clear(cmd_buffer
->surface_relocs
.deps
, NULL
);
849 cmd_buffer
->last_ss_pool_center
= 0;
851 /* Reset the list of seen buffers */
852 cmd_buffer
->seen_bbos
.head
= 0;
853 cmd_buffer
->seen_bbos
.tail
= 0;
855 *(struct anv_batch_bo
**)u_vector_add(&cmd_buffer
->seen_bbos
) =
856 anv_cmd_buffer_current_batch_bo(cmd_buffer
);
860 anv_cmd_buffer_end_batch_buffer(struct anv_cmd_buffer
*cmd_buffer
)
862 struct anv_batch_bo
*batch_bo
= anv_cmd_buffer_current_batch_bo(cmd_buffer
);
864 if (cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
) {
865 /* When we start a batch buffer, we subtract a certain amount of
866 * padding from the end to ensure that we always have room to emit a
867 * BATCH_BUFFER_START to chain to the next BO. We need to remove
868 * that padding before we end the batch; otherwise, we may end up
869 * with our BATCH_BUFFER_END in another BO.
871 cmd_buffer
->batch
.end
+= GEN8_MI_BATCH_BUFFER_START_length
* 4;
872 assert(cmd_buffer
->batch
.end
== batch_bo
->bo
.map
+ batch_bo
->bo
.size
);
874 anv_batch_emit(&cmd_buffer
->batch
, GEN8_MI_BATCH_BUFFER_END
, bbe
);
876 /* Round batch up to an even number of dwords. */
877 if ((cmd_buffer
->batch
.next
- cmd_buffer
->batch
.start
) & 4)
878 anv_batch_emit(&cmd_buffer
->batch
, GEN8_MI_NOOP
, noop
);
880 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_PRIMARY
;
882 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_SECONDARY
);
883 /* If this is a secondary command buffer, we need to determine the
884 * mode in which it will be executed with vkExecuteCommands. We
885 * determine this statically here so that this stays in sync with the
886 * actual ExecuteCommands implementation.
888 const uint32_t length
= cmd_buffer
->batch
.next
- cmd_buffer
->batch
.start
;
889 if (!cmd_buffer
->device
->can_chain_batches
) {
890 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_GROW_AND_EMIT
;
891 } else if ((cmd_buffer
->batch_bos
.next
== cmd_buffer
->batch_bos
.prev
) &&
892 (length
< ANV_CMD_BUFFER_BATCH_SIZE
/ 2)) {
893 /* If the secondary has exactly one batch buffer in its list *and*
894 * that batch buffer is less than half of the maximum size, we're
895 * probably better of simply copying it into our batch.
897 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_EMIT
;
898 } else if (!(cmd_buffer
->usage_flags
&
899 VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT
)) {
900 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_CHAIN
;
902 /* In order to chain, we need this command buffer to contain an
903 * MI_BATCH_BUFFER_START which will jump back to the calling batch.
904 * It doesn't matter where it points now so long as has a valid
905 * relocation. We'll adjust it later as part of the chaining
908 * We set the end of the batch a little short so we would be sure we
909 * have room for the chaining command. Since we're about to emit the
910 * chaining command, let's set it back where it should go.
912 cmd_buffer
->batch
.end
+= GEN8_MI_BATCH_BUFFER_START_length
* 4;
913 assert(cmd_buffer
->batch
.start
== batch_bo
->bo
.map
);
914 assert(cmd_buffer
->batch
.end
== batch_bo
->bo
.map
+ batch_bo
->bo
.size
);
916 emit_batch_buffer_start(cmd_buffer
, &batch_bo
->bo
, 0);
917 assert(cmd_buffer
->batch
.start
== batch_bo
->bo
.map
);
919 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN
;
923 anv_batch_bo_finish(batch_bo
, &cmd_buffer
->batch
);
927 anv_cmd_buffer_add_seen_bbos(struct anv_cmd_buffer
*cmd_buffer
,
928 struct list_head
*list
)
930 list_for_each_entry(struct anv_batch_bo
, bbo
, list
, link
) {
931 struct anv_batch_bo
**bbo_ptr
= u_vector_add(&cmd_buffer
->seen_bbos
);
933 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
942 anv_cmd_buffer_add_secondary(struct anv_cmd_buffer
*primary
,
943 struct anv_cmd_buffer
*secondary
)
945 switch (secondary
->exec_mode
) {
946 case ANV_CMD_BUFFER_EXEC_MODE_EMIT
:
947 anv_batch_emit_batch(&primary
->batch
, &secondary
->batch
);
949 case ANV_CMD_BUFFER_EXEC_MODE_GROW_AND_EMIT
: {
950 struct anv_batch_bo
*bbo
= anv_cmd_buffer_current_batch_bo(primary
);
951 unsigned length
= secondary
->batch
.end
- secondary
->batch
.start
;
952 anv_batch_bo_grow(primary
, bbo
, &primary
->batch
, length
,
953 GEN8_MI_BATCH_BUFFER_START_length
* 4);
954 anv_batch_emit_batch(&primary
->batch
, &secondary
->batch
);
957 case ANV_CMD_BUFFER_EXEC_MODE_CHAIN
: {
958 struct anv_batch_bo
*first_bbo
=
959 list_first_entry(&secondary
->batch_bos
, struct anv_batch_bo
, link
);
960 struct anv_batch_bo
*last_bbo
=
961 list_last_entry(&secondary
->batch_bos
, struct anv_batch_bo
, link
);
963 emit_batch_buffer_start(primary
, &first_bbo
->bo
, 0);
965 struct anv_batch_bo
*this_bbo
= anv_cmd_buffer_current_batch_bo(primary
);
966 assert(primary
->batch
.start
== this_bbo
->bo
.map
);
967 uint32_t offset
= primary
->batch
.next
- primary
->batch
.start
;
969 /* Make the tail of the secondary point back to right after the
970 * MI_BATCH_BUFFER_START in the primary batch.
972 anv_batch_bo_link(primary
, last_bbo
, this_bbo
, offset
);
974 anv_cmd_buffer_add_seen_bbos(primary
, &secondary
->batch_bos
);
977 case ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN
: {
978 struct list_head copy_list
;
979 VkResult result
= anv_batch_bo_list_clone(&secondary
->batch_bos
,
982 if (result
!= VK_SUCCESS
)
985 anv_cmd_buffer_add_seen_bbos(primary
, ©_list
);
987 struct anv_batch_bo
*first_bbo
=
988 list_first_entry(©_list
, struct anv_batch_bo
, link
);
989 struct anv_batch_bo
*last_bbo
=
990 list_last_entry(©_list
, struct anv_batch_bo
, link
);
992 cmd_buffer_chain_to_batch_bo(primary
, first_bbo
);
994 list_splicetail(©_list
, &primary
->batch_bos
);
996 anv_batch_bo_continue(last_bbo
, &primary
->batch
,
997 GEN8_MI_BATCH_BUFFER_START_length
* 4);
1001 assert(!"Invalid execution mode");
1004 anv_reloc_list_append(&primary
->surface_relocs
, &primary
->pool
->alloc
,
1005 &secondary
->surface_relocs
, 0);
1008 struct anv_execbuf
{
1009 struct drm_i915_gem_execbuffer2 execbuf
;
1011 struct drm_i915_gem_exec_object2
* objects
;
1013 struct anv_bo
** bos
;
1015 /* Allocated length of the 'objects' and 'bos' arrays */
1016 uint32_t array_length
;
1020 uint32_t fence_count
;
1021 uint32_t fence_array_length
;
1022 struct drm_i915_gem_exec_fence
* fences
;
1023 struct anv_syncobj
** syncobjs
;
1027 anv_execbuf_init(struct anv_execbuf
*exec
)
1029 memset(exec
, 0, sizeof(*exec
));
1033 anv_execbuf_finish(struct anv_execbuf
*exec
,
1034 const VkAllocationCallbacks
*alloc
)
1036 vk_free(alloc
, exec
->objects
);
1037 vk_free(alloc
, exec
->bos
);
1038 vk_free(alloc
, exec
->fences
);
1039 vk_free(alloc
, exec
->syncobjs
);
1043 _compare_bo_handles(const void *_bo1
, const void *_bo2
)
1045 struct anv_bo
* const *bo1
= _bo1
;
1046 struct anv_bo
* const *bo2
= _bo2
;
1048 return (*bo1
)->gem_handle
- (*bo2
)->gem_handle
;
1052 anv_execbuf_add_bo_set(struct anv_execbuf
*exec
,
1054 uint32_t extra_flags
,
1055 const VkAllocationCallbacks
*alloc
);
1058 anv_execbuf_add_bo(struct anv_execbuf
*exec
,
1060 struct anv_reloc_list
*relocs
,
1061 uint32_t extra_flags
,
1062 const VkAllocationCallbacks
*alloc
)
1064 struct drm_i915_gem_exec_object2
*obj
= NULL
;
1066 if (bo
->index
< exec
->bo_count
&& exec
->bos
[bo
->index
] == bo
)
1067 obj
= &exec
->objects
[bo
->index
];
1070 /* We've never seen this one before. Add it to the list and assign
1071 * an id that we can use later.
1073 if (exec
->bo_count
>= exec
->array_length
) {
1074 uint32_t new_len
= exec
->objects
? exec
->array_length
* 2 : 64;
1076 struct drm_i915_gem_exec_object2
*new_objects
=
1077 vk_alloc(alloc
, new_len
* sizeof(*new_objects
),
1078 8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND
);
1079 if (new_objects
== NULL
)
1080 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1082 struct anv_bo
**new_bos
=
1083 vk_alloc(alloc
, new_len
* sizeof(*new_bos
),
1084 8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND
);
1085 if (new_bos
== NULL
) {
1086 vk_free(alloc
, new_objects
);
1087 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1090 if (exec
->objects
) {
1091 memcpy(new_objects
, exec
->objects
,
1092 exec
->bo_count
* sizeof(*new_objects
));
1093 memcpy(new_bos
, exec
->bos
,
1094 exec
->bo_count
* sizeof(*new_bos
));
1097 vk_free(alloc
, exec
->objects
);
1098 vk_free(alloc
, exec
->bos
);
1100 exec
->objects
= new_objects
;
1101 exec
->bos
= new_bos
;
1102 exec
->array_length
= new_len
;
1105 assert(exec
->bo_count
< exec
->array_length
);
1107 bo
->index
= exec
->bo_count
++;
1108 obj
= &exec
->objects
[bo
->index
];
1109 exec
->bos
[bo
->index
] = bo
;
1111 obj
->handle
= bo
->gem_handle
;
1112 obj
->relocation_count
= 0;
1113 obj
->relocs_ptr
= 0;
1115 obj
->offset
= bo
->offset
;
1116 obj
->flags
= bo
->flags
| extra_flags
;
1121 if (relocs
!= NULL
) {
1122 assert(obj
->relocation_count
== 0);
1124 if (relocs
->num_relocs
> 0) {
1125 /* This is the first time we've ever seen a list of relocations for
1126 * this BO. Go ahead and set the relocations and then walk the list
1127 * of relocations and add them all.
1129 exec
->has_relocs
= true;
1130 obj
->relocation_count
= relocs
->num_relocs
;
1131 obj
->relocs_ptr
= (uintptr_t) relocs
->relocs
;
1133 for (size_t i
= 0; i
< relocs
->num_relocs
; i
++) {
1136 /* A quick sanity check on relocations */
1137 assert(relocs
->relocs
[i
].offset
< bo
->size
);
1138 result
= anv_execbuf_add_bo(exec
, relocs
->reloc_bos
[i
], NULL
,
1139 extra_flags
, alloc
);
1141 if (result
!= VK_SUCCESS
)
1146 return anv_execbuf_add_bo_set(exec
, relocs
->deps
, extra_flags
, alloc
);
1152 /* Add BO dependencies to execbuf */
1154 anv_execbuf_add_bo_set(struct anv_execbuf
*exec
,
1156 uint32_t extra_flags
,
1157 const VkAllocationCallbacks
*alloc
)
1159 if (!deps
|| deps
->entries
<= 0)
1162 const uint32_t entries
= deps
->entries
;
1163 struct anv_bo
**bos
=
1164 vk_alloc(alloc
, entries
* sizeof(*bos
),
1165 8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND
);
1167 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1169 struct anv_bo
**bo
= bos
;
1170 set_foreach(deps
, entry
) {
1171 *bo
++ = (void *)entry
->key
;
1174 qsort(bos
, entries
, sizeof(struct anv_bo
*), _compare_bo_handles
);
1176 VkResult result
= VK_SUCCESS
;
1177 for (bo
= bos
; bo
< bos
+ entries
; bo
++) {
1178 result
= anv_execbuf_add_bo(exec
, *bo
, NULL
, extra_flags
, alloc
);
1179 if (result
!= VK_SUCCESS
)
1183 vk_free(alloc
, bos
);
1189 anv_execbuf_add_syncobj(struct anv_execbuf
*exec
,
1190 uint32_t handle
, uint32_t flags
,
1191 const VkAllocationCallbacks
*alloc
)
1195 if (exec
->fence_count
>= exec
->fence_array_length
) {
1196 uint32_t new_len
= MAX2(exec
->fence_array_length
* 2, 64);
1198 exec
->fences
= vk_realloc(alloc
, exec
->fences
,
1199 new_len
* sizeof(*exec
->fences
),
1200 8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND
);
1201 if (exec
->fences
== NULL
)
1202 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1204 exec
->fence_array_length
= new_len
;
1207 exec
->fences
[exec
->fence_count
] = (struct drm_i915_gem_exec_fence
) {
1212 exec
->fence_count
++;
1218 anv_cmd_buffer_process_relocs(struct anv_cmd_buffer
*cmd_buffer
,
1219 struct anv_reloc_list
*list
)
1221 for (size_t i
= 0; i
< list
->num_relocs
; i
++)
1222 list
->relocs
[i
].target_handle
= list
->reloc_bos
[i
]->index
;
1226 adjust_relocations_from_state_pool(struct anv_state_pool
*pool
,
1227 struct anv_reloc_list
*relocs
,
1228 uint32_t last_pool_center_bo_offset
)
1230 assert(last_pool_center_bo_offset
<= pool
->block_pool
.center_bo_offset
);
1231 uint32_t delta
= pool
->block_pool
.center_bo_offset
- last_pool_center_bo_offset
;
1233 for (size_t i
= 0; i
< relocs
->num_relocs
; i
++) {
1234 /* All of the relocations from this block pool to other BO's should
1235 * have been emitted relative to the surface block pool center. We
1236 * need to add the center offset to make them relative to the
1237 * beginning of the actual GEM bo.
1239 relocs
->relocs
[i
].offset
+= delta
;
1244 adjust_relocations_to_state_pool(struct anv_state_pool
*pool
,
1245 struct anv_bo
*from_bo
,
1246 struct anv_reloc_list
*relocs
,
1247 uint32_t last_pool_center_bo_offset
)
1249 assert(last_pool_center_bo_offset
<= pool
->block_pool
.center_bo_offset
);
1250 uint32_t delta
= pool
->block_pool
.center_bo_offset
- last_pool_center_bo_offset
;
1252 /* When we initially emit relocations into a block pool, we don't
1253 * actually know what the final center_bo_offset will be so we just emit
1254 * it as if center_bo_offset == 0. Now that we know what the center
1255 * offset is, we need to walk the list of relocations and adjust any
1256 * relocations that point to the pool bo with the correct offset.
1258 for (size_t i
= 0; i
< relocs
->num_relocs
; i
++) {
1259 if (relocs
->reloc_bos
[i
] == pool
->block_pool
.bo
) {
1260 /* Adjust the delta value in the relocation to correctly
1261 * correspond to the new delta. Initially, this value may have
1262 * been negative (if treated as unsigned), but we trust in
1263 * uint32_t roll-over to fix that for us at this point.
1265 relocs
->relocs
[i
].delta
+= delta
;
1267 /* Since the delta has changed, we need to update the actual
1268 * relocated value with the new presumed value. This function
1269 * should only be called on batch buffers, so we know it isn't in
1270 * use by the GPU at the moment.
1272 assert(relocs
->relocs
[i
].offset
< from_bo
->size
);
1273 write_reloc(pool
->block_pool
.device
,
1274 from_bo
->map
+ relocs
->relocs
[i
].offset
,
1275 relocs
->relocs
[i
].presumed_offset
+
1276 relocs
->relocs
[i
].delta
, false);
1282 anv_reloc_list_apply(struct anv_device
*device
,
1283 struct anv_reloc_list
*list
,
1285 bool always_relocate
)
1287 for (size_t i
= 0; i
< list
->num_relocs
; i
++) {
1288 struct anv_bo
*target_bo
= list
->reloc_bos
[i
];
1289 if (list
->relocs
[i
].presumed_offset
== target_bo
->offset
&&
1293 void *p
= bo
->map
+ list
->relocs
[i
].offset
;
1294 write_reloc(device
, p
, target_bo
->offset
+ list
->relocs
[i
].delta
, true);
1295 list
->relocs
[i
].presumed_offset
= target_bo
->offset
;
1300 * This function applies the relocation for a command buffer and writes the
1301 * actual addresses into the buffers as per what we were told by the kernel on
1302 * the previous execbuf2 call. This should be safe to do because, for each
1303 * relocated address, we have two cases:
1305 * 1) The target BO is inactive (as seen by the kernel). In this case, it is
1306 * not in use by the GPU so updating the address is 100% ok. It won't be
1307 * in-use by the GPU (from our context) again until the next execbuf2
1308 * happens. If the kernel decides to move it in the next execbuf2, it
1309 * will have to do the relocations itself, but that's ok because it should
1310 * have all of the information needed to do so.
1312 * 2) The target BO is active (as seen by the kernel). In this case, it
1313 * hasn't moved since the last execbuffer2 call because GTT shuffling
1314 * *only* happens when the BO is idle. (From our perspective, it only
1315 * happens inside the execbuffer2 ioctl, but the shuffling may be
1316 * triggered by another ioctl, with full-ppgtt this is limited to only
1317 * execbuffer2 ioctls on the same context, or memory pressure.) Since the
1318 * target BO hasn't moved, our anv_bo::offset exactly matches the BO's GTT
1319 * address and the relocated value we are writing into the BO will be the
1320 * same as the value that is already there.
1322 * There is also a possibility that the target BO is active but the exact
1323 * RENDER_SURFACE_STATE object we are writing the relocation into isn't in
1324 * use. In this case, the address currently in the RENDER_SURFACE_STATE
1325 * may be stale but it's still safe to write the relocation because that
1326 * particular RENDER_SURFACE_STATE object isn't in-use by the GPU and
1327 * won't be until the next execbuf2 call.
1329 * By doing relocations on the CPU, we can tell the kernel that it doesn't
1330 * need to bother. We want to do this because the surface state buffer is
1331 * used by every command buffer so, if the kernel does the relocations, it
1332 * will always be busy and the kernel will always stall. This is also
1333 * probably the fastest mechanism for doing relocations since the kernel would
1334 * have to make a full copy of all the relocations lists.
1337 relocate_cmd_buffer(struct anv_cmd_buffer
*cmd_buffer
,
1338 struct anv_execbuf
*exec
)
1340 if (!exec
->has_relocs
)
1343 static int userspace_relocs
= -1;
1344 if (userspace_relocs
< 0)
1345 userspace_relocs
= env_var_as_boolean("ANV_USERSPACE_RELOCS", true);
1346 if (!userspace_relocs
)
1349 /* First, we have to check to see whether or not we can even do the
1350 * relocation. New buffers which have never been submitted to the kernel
1351 * don't have a valid offset so we need to let the kernel do relocations so
1352 * that we can get offsets for them. On future execbuf2 calls, those
1353 * buffers will have offsets and we will be able to skip relocating.
1354 * Invalid offsets are indicated by anv_bo::offset == (uint64_t)-1.
1356 for (uint32_t i
= 0; i
< exec
->bo_count
; i
++) {
1357 if (exec
->bos
[i
]->offset
== (uint64_t)-1)
1361 /* Since surface states are shared between command buffers and we don't
1362 * know what order they will be submitted to the kernel, we don't know
1363 * what address is actually written in the surface state object at any
1364 * given time. The only option is to always relocate them.
1366 anv_reloc_list_apply(cmd_buffer
->device
, &cmd_buffer
->surface_relocs
,
1367 cmd_buffer
->device
->surface_state_pool
.block_pool
.bo
,
1368 true /* always relocate surface states */);
1370 /* Since we own all of the batch buffers, we know what values are stored
1371 * in the relocated addresses and only have to update them if the offsets
1374 struct anv_batch_bo
**bbo
;
1375 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
) {
1376 anv_reloc_list_apply(cmd_buffer
->device
,
1377 &(*bbo
)->relocs
, &(*bbo
)->bo
, false);
1380 for (uint32_t i
= 0; i
< exec
->bo_count
; i
++)
1381 exec
->objects
[i
].offset
= exec
->bos
[i
]->offset
;
1387 setup_execbuf_for_cmd_buffer(struct anv_execbuf
*execbuf
,
1388 struct anv_cmd_buffer
*cmd_buffer
)
1390 struct anv_batch
*batch
= &cmd_buffer
->batch
;
1391 struct anv_state_pool
*ss_pool
=
1392 &cmd_buffer
->device
->surface_state_pool
;
1394 adjust_relocations_from_state_pool(ss_pool
, &cmd_buffer
->surface_relocs
,
1395 cmd_buffer
->last_ss_pool_center
);
1397 if (cmd_buffer
->device
->instance
->physicalDevice
.use_softpin
) {
1398 anv_block_pool_foreach_bo(bo
, &ss_pool
->block_pool
) {
1399 result
= anv_execbuf_add_bo(execbuf
, bo
, NULL
, 0,
1400 &cmd_buffer
->device
->alloc
);
1401 if (result
!= VK_SUCCESS
)
1404 /* Add surface dependencies (BOs) to the execbuf */
1405 anv_execbuf_add_bo_set(execbuf
, cmd_buffer
->surface_relocs
.deps
, 0,
1406 &cmd_buffer
->device
->alloc
);
1408 /* Add the BOs for all memory objects */
1409 list_for_each_entry(struct anv_device_memory
, mem
,
1410 &cmd_buffer
->device
->memory_objects
, link
) {
1411 result
= anv_execbuf_add_bo(execbuf
, mem
->bo
, NULL
, 0,
1412 &cmd_buffer
->device
->alloc
);
1413 if (result
!= VK_SUCCESS
)
1417 struct anv_block_pool
*pool
;
1418 pool
= &cmd_buffer
->device
->dynamic_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
->instruction_state_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 pool
= &cmd_buffer
->device
->binding_table_pool
.block_pool
;
1435 anv_block_pool_foreach_bo(bo
, pool
) {
1436 result
= anv_execbuf_add_bo(execbuf
, bo
, NULL
, 0,
1437 &cmd_buffer
->device
->alloc
);
1438 if (result
!= VK_SUCCESS
)
1442 /* Since we aren't in the softpin case, all of our STATE_BASE_ADDRESS BOs
1443 * will get added automatically by processing relocations on the batch
1444 * buffer. We have to add the surface state BO manually because it has
1445 * relocations of its own that we need to be sure are processsed.
1447 result
= anv_execbuf_add_bo(execbuf
, ss_pool
->block_pool
.bo
,
1448 &cmd_buffer
->surface_relocs
, 0,
1449 &cmd_buffer
->device
->alloc
);
1450 if (result
!= VK_SUCCESS
)
1454 /* First, we walk over all of the bos we've seen and add them and their
1455 * relocations to the validate list.
1457 struct anv_batch_bo
**bbo
;
1458 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
) {
1459 adjust_relocations_to_state_pool(ss_pool
, &(*bbo
)->bo
, &(*bbo
)->relocs
,
1460 cmd_buffer
->last_ss_pool_center
);
1462 result
= anv_execbuf_add_bo(execbuf
, &(*bbo
)->bo
, &(*bbo
)->relocs
, 0,
1463 &cmd_buffer
->device
->alloc
);
1464 if (result
!= VK_SUCCESS
)
1468 /* Now that we've adjusted all of the surface state relocations, we need to
1469 * record the surface state pool center so future executions of the command
1470 * buffer can adjust correctly.
1472 cmd_buffer
->last_ss_pool_center
= ss_pool
->block_pool
.center_bo_offset
;
1474 struct anv_batch_bo
*first_batch_bo
=
1475 list_first_entry(&cmd_buffer
->batch_bos
, struct anv_batch_bo
, link
);
1477 /* The kernel requires that the last entry in the validation list be the
1478 * batch buffer to execute. We can simply swap the element
1479 * corresponding to the first batch_bo in the chain with the last
1480 * element in the list.
1482 if (first_batch_bo
->bo
.index
!= execbuf
->bo_count
- 1) {
1483 uint32_t idx
= first_batch_bo
->bo
.index
;
1484 uint32_t last_idx
= execbuf
->bo_count
- 1;
1486 struct drm_i915_gem_exec_object2 tmp_obj
= execbuf
->objects
[idx
];
1487 assert(execbuf
->bos
[idx
] == &first_batch_bo
->bo
);
1489 execbuf
->objects
[idx
] = execbuf
->objects
[last_idx
];
1490 execbuf
->bos
[idx
] = execbuf
->bos
[last_idx
];
1491 execbuf
->bos
[idx
]->index
= idx
;
1493 execbuf
->objects
[last_idx
] = tmp_obj
;
1494 execbuf
->bos
[last_idx
] = &first_batch_bo
->bo
;
1495 first_batch_bo
->bo
.index
= last_idx
;
1498 /* If we are pinning our BOs, we shouldn't have to relocate anything */
1499 if (cmd_buffer
->device
->instance
->physicalDevice
.use_softpin
)
1500 assert(!execbuf
->has_relocs
);
1502 /* Now we go through and fixup all of the relocation lists to point to
1503 * the correct indices in the object array. We have to do this after we
1504 * reorder the list above as some of the indices may have changed.
1506 if (execbuf
->has_relocs
) {
1507 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
)
1508 anv_cmd_buffer_process_relocs(cmd_buffer
, &(*bbo
)->relocs
);
1510 anv_cmd_buffer_process_relocs(cmd_buffer
, &cmd_buffer
->surface_relocs
);
1513 if (!cmd_buffer
->device
->info
.has_llc
) {
1514 __builtin_ia32_mfence();
1515 u_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
) {
1516 for (uint32_t i
= 0; i
< (*bbo
)->length
; i
+= CACHELINE_SIZE
)
1517 __builtin_ia32_clflush((*bbo
)->bo
.map
+ i
);
1521 execbuf
->execbuf
= (struct drm_i915_gem_execbuffer2
) {
1522 .buffers_ptr
= (uintptr_t) execbuf
->objects
,
1523 .buffer_count
= execbuf
->bo_count
,
1524 .batch_start_offset
= 0,
1525 .batch_len
= batch
->next
- batch
->start
,
1530 .flags
= I915_EXEC_HANDLE_LUT
| I915_EXEC_RENDER
,
1531 .rsvd1
= cmd_buffer
->device
->context_id
,
1535 if (relocate_cmd_buffer(cmd_buffer
, execbuf
)) {
1536 /* If we were able to successfully relocate everything, tell the kernel
1537 * that it can skip doing relocations. The requirement for using
1540 * 1) The addresses written in the objects must match the corresponding
1541 * reloc.presumed_offset which in turn must match the corresponding
1542 * execobject.offset.
1544 * 2) To avoid stalling, execobject.offset should match the current
1545 * address of that object within the active context.
1547 * In order to satisfy all of the invariants that make userspace
1548 * relocations to be safe (see relocate_cmd_buffer()), we need to
1549 * further ensure that the addresses we use match those used by the
1550 * kernel for the most recent execbuf2.
1552 * The kernel may still choose to do relocations anyway if something has
1553 * moved in the GTT. In this case, the relocation list still needs to be
1554 * valid. All relocations on the batch buffers are already valid and
1555 * kept up-to-date. For surface state relocations, by applying the
1556 * relocations in relocate_cmd_buffer, we ensured that the address in
1557 * the RENDER_SURFACE_STATE matches presumed_offset, so it should be
1558 * safe for the kernel to relocate them as needed.
1560 execbuf
->execbuf
.flags
|= I915_EXEC_NO_RELOC
;
1562 /* In the case where we fall back to doing kernel relocations, we need
1563 * to ensure that the relocation list is valid. All relocations on the
1564 * batch buffers are already valid and kept up-to-date. Since surface
1565 * states are shared between command buffers and we don't know what
1566 * order they will be submitted to the kernel, we don't know what
1567 * address is actually written in the surface state object at any given
1568 * time. The only option is to set a bogus presumed offset and let the
1569 * kernel relocate them.
1571 for (size_t i
= 0; i
< cmd_buffer
->surface_relocs
.num_relocs
; i
++)
1572 cmd_buffer
->surface_relocs
.relocs
[i
].presumed_offset
= -1;
1579 setup_empty_execbuf(struct anv_execbuf
*execbuf
, struct anv_device
*device
)
1581 VkResult result
= anv_execbuf_add_bo(execbuf
, &device
->trivial_batch_bo
,
1582 NULL
, 0, &device
->alloc
);
1583 if (result
!= VK_SUCCESS
)
1586 execbuf
->execbuf
= (struct drm_i915_gem_execbuffer2
) {
1587 .buffers_ptr
= (uintptr_t) execbuf
->objects
,
1588 .buffer_count
= execbuf
->bo_count
,
1589 .batch_start_offset
= 0,
1590 .batch_len
= 8, /* GEN7_MI_BATCH_BUFFER_END and NOOP */
1591 .flags
= I915_EXEC_HANDLE_LUT
| I915_EXEC_RENDER
,
1592 .rsvd1
= device
->context_id
,
1600 anv_cmd_buffer_execbuf(struct anv_device
*device
,
1601 struct anv_cmd_buffer
*cmd_buffer
,
1602 const VkSemaphore
*in_semaphores
,
1603 uint32_t num_in_semaphores
,
1604 const VkSemaphore
*out_semaphores
,
1605 uint32_t num_out_semaphores
,
1608 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
1609 UNUSED
struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1611 struct anv_execbuf execbuf
;
1612 anv_execbuf_init(&execbuf
);
1615 VkResult result
= VK_SUCCESS
;
1616 for (uint32_t i
= 0; i
< num_in_semaphores
; i
++) {
1617 ANV_FROM_HANDLE(anv_semaphore
, semaphore
, in_semaphores
[i
]);
1618 struct anv_semaphore_impl
*impl
=
1619 semaphore
->temporary
.type
!= ANV_SEMAPHORE_TYPE_NONE
?
1620 &semaphore
->temporary
: &semaphore
->permanent
;
1622 switch (impl
->type
) {
1623 case ANV_SEMAPHORE_TYPE_BO
:
1624 assert(!pdevice
->has_syncobj
);
1625 result
= anv_execbuf_add_bo(&execbuf
, impl
->bo
, NULL
,
1627 if (result
!= VK_SUCCESS
)
1631 case ANV_SEMAPHORE_TYPE_SYNC_FILE
:
1632 assert(!pdevice
->has_syncobj
);
1633 if (in_fence
== -1) {
1634 in_fence
= impl
->fd
;
1636 int merge
= anv_gem_sync_file_merge(device
, in_fence
, impl
->fd
);
1638 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1648 case ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ
:
1649 result
= anv_execbuf_add_syncobj(&execbuf
, impl
->syncobj
,
1650 I915_EXEC_FENCE_WAIT
,
1652 if (result
!= VK_SUCCESS
)
1661 bool need_out_fence
= false;
1662 for (uint32_t i
= 0; i
< num_out_semaphores
; i
++) {
1663 ANV_FROM_HANDLE(anv_semaphore
, semaphore
, out_semaphores
[i
]);
1665 /* Under most circumstances, out fences won't be temporary. However,
1666 * the spec does allow it for opaque_fd. From the Vulkan 1.0.53 spec:
1668 * "If the import is temporary, the implementation must restore the
1669 * semaphore to its prior permanent state after submitting the next
1670 * semaphore wait operation."
1672 * The spec says nothing whatsoever about signal operations on
1673 * temporarily imported semaphores so it appears they are allowed.
1674 * There are also CTS tests that require this to work.
1676 struct anv_semaphore_impl
*impl
=
1677 semaphore
->temporary
.type
!= ANV_SEMAPHORE_TYPE_NONE
?
1678 &semaphore
->temporary
: &semaphore
->permanent
;
1680 switch (impl
->type
) {
1681 case ANV_SEMAPHORE_TYPE_BO
:
1682 assert(!pdevice
->has_syncobj
);
1683 result
= anv_execbuf_add_bo(&execbuf
, impl
->bo
, NULL
,
1684 EXEC_OBJECT_WRITE
, &device
->alloc
);
1685 if (result
!= VK_SUCCESS
)
1689 case ANV_SEMAPHORE_TYPE_SYNC_FILE
:
1690 assert(!pdevice
->has_syncobj
);
1691 need_out_fence
= true;
1694 case ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ
:
1695 result
= anv_execbuf_add_syncobj(&execbuf
, impl
->syncobj
,
1696 I915_EXEC_FENCE_SIGNAL
,
1698 if (result
!= VK_SUCCESS
)
1708 /* Under most circumstances, out fences won't be temporary. However,
1709 * the spec does allow it for opaque_fd. From the Vulkan 1.0.53 spec:
1711 * "If the import is temporary, the implementation must restore the
1712 * semaphore to its prior permanent state after submitting the next
1713 * semaphore wait operation."
1715 * The spec says nothing whatsoever about signal operations on
1716 * temporarily imported semaphores so it appears they are allowed.
1717 * There are also CTS tests that require this to work.
1719 struct anv_fence_impl
*impl
=
1720 fence
->temporary
.type
!= ANV_FENCE_TYPE_NONE
?
1721 &fence
->temporary
: &fence
->permanent
;
1723 switch (impl
->type
) {
1724 case ANV_FENCE_TYPE_BO
:
1725 assert(!pdevice
->has_syncobj_wait
);
1726 result
= anv_execbuf_add_bo(&execbuf
, &impl
->bo
.bo
, NULL
,
1727 EXEC_OBJECT_WRITE
, &device
->alloc
);
1728 if (result
!= VK_SUCCESS
)
1732 case ANV_FENCE_TYPE_SYNCOBJ
:
1733 result
= anv_execbuf_add_syncobj(&execbuf
, impl
->syncobj
,
1734 I915_EXEC_FENCE_SIGNAL
,
1736 if (result
!= VK_SUCCESS
)
1741 unreachable("Invalid fence type");
1746 if (unlikely(INTEL_DEBUG
& DEBUG_BATCH
)) {
1747 struct anv_batch_bo
**bo
= u_vector_tail(&cmd_buffer
->seen_bbos
);
1749 device
->cmd_buffer_being_decoded
= cmd_buffer
;
1750 gen_print_batch(&device
->decoder_ctx
, (*bo
)->bo
.map
,
1751 (*bo
)->bo
.size
, (*bo
)->bo
.offset
, false);
1752 device
->cmd_buffer_being_decoded
= NULL
;
1755 result
= setup_execbuf_for_cmd_buffer(&execbuf
, cmd_buffer
);
1757 result
= setup_empty_execbuf(&execbuf
, device
);
1760 if (result
!= VK_SUCCESS
)
1763 if (execbuf
.fence_count
> 0) {
1764 assert(device
->instance
->physicalDevice
.has_syncobj
);
1765 execbuf
.execbuf
.flags
|= I915_EXEC_FENCE_ARRAY
;
1766 execbuf
.execbuf
.num_cliprects
= execbuf
.fence_count
;
1767 execbuf
.execbuf
.cliprects_ptr
= (uintptr_t) execbuf
.fences
;
1770 if (in_fence
!= -1) {
1771 execbuf
.execbuf
.flags
|= I915_EXEC_FENCE_IN
;
1772 execbuf
.execbuf
.rsvd2
|= (uint32_t)in_fence
;
1776 execbuf
.execbuf
.flags
|= I915_EXEC_FENCE_OUT
;
1778 result
= anv_device_execbuf(device
, &execbuf
.execbuf
, execbuf
.bos
);
1780 /* Execbuf does not consume the in_fence. It's our job to close it. */
1784 for (uint32_t i
= 0; i
< num_in_semaphores
; i
++) {
1785 ANV_FROM_HANDLE(anv_semaphore
, semaphore
, in_semaphores
[i
]);
1786 /* From the Vulkan 1.0.53 spec:
1788 * "If the import is temporary, the implementation must restore the
1789 * semaphore to its prior permanent state after submitting the next
1790 * semaphore wait operation."
1792 * This has to happen after the execbuf in case we close any syncobjs in
1795 anv_semaphore_reset_temporary(device
, semaphore
);
1798 if (fence
&& fence
->permanent
.type
== ANV_FENCE_TYPE_BO
) {
1799 assert(!pdevice
->has_syncobj_wait
);
1800 /* BO fences can't be shared, so they can't be temporary. */
1801 assert(fence
->temporary
.type
== ANV_FENCE_TYPE_NONE
);
1803 /* Once the execbuf has returned, we need to set the fence state to
1804 * SUBMITTED. We can't do this before calling execbuf because
1805 * anv_GetFenceStatus does take the global device lock before checking
1808 * We set the fence state to SUBMITTED regardless of whether or not the
1809 * execbuf succeeds because we need to ensure that vkWaitForFences() and
1810 * vkGetFenceStatus() return a valid result (VK_ERROR_DEVICE_LOST or
1811 * VK_SUCCESS) in a finite amount of time even if execbuf fails.
1813 fence
->permanent
.bo
.state
= ANV_BO_FENCE_STATE_SUBMITTED
;
1816 if (result
== VK_SUCCESS
&& need_out_fence
) {
1817 assert(!pdevice
->has_syncobj_wait
);
1818 int out_fence
= execbuf
.execbuf
.rsvd2
>> 32;
1819 for (uint32_t i
= 0; i
< num_out_semaphores
; i
++) {
1820 ANV_FROM_HANDLE(anv_semaphore
, semaphore
, out_semaphores
[i
]);
1821 /* Out fences can't have temporary state because that would imply
1822 * that we imported a sync file and are trying to signal it.
1824 assert(semaphore
->temporary
.type
== ANV_SEMAPHORE_TYPE_NONE
);
1825 struct anv_semaphore_impl
*impl
= &semaphore
->permanent
;
1827 if (impl
->type
== ANV_SEMAPHORE_TYPE_SYNC_FILE
) {
1828 assert(impl
->fd
== -1);
1829 impl
->fd
= dup(out_fence
);
1835 anv_execbuf_finish(&execbuf
, &device
->alloc
);