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 "gen7_pack.h"
33 #include "gen8_pack.h"
35 /** \file anv_batch_chain.c
37 * This file contains functions related to anv_cmd_buffer as a data
38 * structure. This involves everything required to create and destroy
39 * the actual batch buffers as well as link them together and handle
40 * relocations and surface state. It specifically does *not* contain any
41 * handling of actual vkCmd calls beyond vkCmdExecuteCommands.
44 /*-----------------------------------------------------------------------*
45 * Functions related to anv_reloc_list
46 *-----------------------------------------------------------------------*/
49 anv_reloc_list_init_clone(struct anv_reloc_list
*list
,
50 const VkAllocationCallbacks
*alloc
,
51 const struct anv_reloc_list
*other_list
)
54 list
->num_relocs
= other_list
->num_relocs
;
55 list
->array_length
= other_list
->array_length
;
58 list
->array_length
= 256;
62 anv_alloc(alloc
, list
->array_length
* sizeof(*list
->relocs
), 8,
63 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
65 if (list
->relocs
== NULL
)
66 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
69 anv_alloc(alloc
, list
->array_length
* sizeof(*list
->reloc_bos
), 8,
70 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
72 if (list
->reloc_bos
== NULL
) {
73 anv_free(alloc
, list
->relocs
);
74 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
78 memcpy(list
->relocs
, other_list
->relocs
,
79 list
->array_length
* sizeof(*list
->relocs
));
80 memcpy(list
->reloc_bos
, other_list
->reloc_bos
,
81 list
->array_length
* sizeof(*list
->reloc_bos
));
88 anv_reloc_list_init(struct anv_reloc_list
*list
,
89 const VkAllocationCallbacks
*alloc
)
91 return anv_reloc_list_init_clone(list
, alloc
, NULL
);
95 anv_reloc_list_finish(struct anv_reloc_list
*list
,
96 const VkAllocationCallbacks
*alloc
)
98 anv_free(alloc
, list
->relocs
);
99 anv_free(alloc
, list
->reloc_bos
);
103 anv_reloc_list_grow(struct anv_reloc_list
*list
,
104 const VkAllocationCallbacks
*alloc
,
105 size_t num_additional_relocs
)
107 if (list
->num_relocs
+ num_additional_relocs
<= list
->array_length
)
110 size_t new_length
= list
->array_length
* 2;
111 while (new_length
< list
->num_relocs
+ num_additional_relocs
)
114 struct drm_i915_gem_relocation_entry
*new_relocs
=
115 anv_alloc(alloc
, new_length
* sizeof(*list
->relocs
), 8,
116 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
117 if (new_relocs
== NULL
)
118 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
120 struct anv_bo
**new_reloc_bos
=
121 anv_alloc(alloc
, new_length
* sizeof(*list
->reloc_bos
), 8,
122 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
123 if (new_relocs
== NULL
) {
124 anv_free(alloc
, new_relocs
);
125 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
128 memcpy(new_relocs
, list
->relocs
, list
->num_relocs
* sizeof(*list
->relocs
));
129 memcpy(new_reloc_bos
, list
->reloc_bos
,
130 list
->num_relocs
* sizeof(*list
->reloc_bos
));
132 anv_free(alloc
, list
->relocs
);
133 anv_free(alloc
, list
->reloc_bos
);
135 list
->array_length
= new_length
;
136 list
->relocs
= new_relocs
;
137 list
->reloc_bos
= new_reloc_bos
;
143 anv_reloc_list_add(struct anv_reloc_list
*list
,
144 const VkAllocationCallbacks
*alloc
,
145 uint32_t offset
, struct anv_bo
*target_bo
, uint32_t delta
)
147 struct drm_i915_gem_relocation_entry
*entry
;
150 const uint32_t domain
=
151 target_bo
->is_winsys_bo
? I915_GEM_DOMAIN_RENDER
: 0;
153 anv_reloc_list_grow(list
, alloc
, 1);
154 /* TODO: Handle failure */
156 /* XXX: Can we use I915_EXEC_HANDLE_LUT? */
157 index
= list
->num_relocs
++;
158 list
->reloc_bos
[index
] = target_bo
;
159 entry
= &list
->relocs
[index
];
160 entry
->target_handle
= target_bo
->gem_handle
;
161 entry
->delta
= delta
;
162 entry
->offset
= offset
;
163 entry
->presumed_offset
= target_bo
->offset
;
164 entry
->read_domains
= domain
;
165 entry
->write_domain
= domain
;
166 VG(VALGRIND_CHECK_MEM_IS_DEFINED(entry
, sizeof(*entry
)));
168 return target_bo
->offset
+ delta
;
172 anv_reloc_list_append(struct anv_reloc_list
*list
,
173 const VkAllocationCallbacks
*alloc
,
174 struct anv_reloc_list
*other
, uint32_t offset
)
176 anv_reloc_list_grow(list
, alloc
, other
->num_relocs
);
177 /* TODO: Handle failure */
179 memcpy(&list
->relocs
[list
->num_relocs
], &other
->relocs
[0],
180 other
->num_relocs
* sizeof(other
->relocs
[0]));
181 memcpy(&list
->reloc_bos
[list
->num_relocs
], &other
->reloc_bos
[0],
182 other
->num_relocs
* sizeof(other
->reloc_bos
[0]));
184 for (uint32_t i
= 0; i
< other
->num_relocs
; i
++)
185 list
->relocs
[i
+ list
->num_relocs
].offset
+= offset
;
187 list
->num_relocs
+= other
->num_relocs
;
190 /*-----------------------------------------------------------------------*
191 * Functions related to anv_batch
192 *-----------------------------------------------------------------------*/
195 anv_batch_emit_dwords(struct anv_batch
*batch
, int num_dwords
)
197 if (batch
->next
+ num_dwords
* 4 > batch
->end
)
198 batch
->extend_cb(batch
, batch
->user_data
);
200 void *p
= batch
->next
;
202 batch
->next
+= num_dwords
* 4;
203 assert(batch
->next
<= batch
->end
);
209 anv_batch_emit_reloc(struct anv_batch
*batch
,
210 void *location
, struct anv_bo
*bo
, uint32_t delta
)
212 return anv_reloc_list_add(batch
->relocs
, batch
->alloc
,
213 location
- batch
->start
, bo
, delta
);
217 anv_batch_emit_batch(struct anv_batch
*batch
, struct anv_batch
*other
)
219 uint32_t size
, offset
;
221 size
= other
->next
- other
->start
;
222 assert(size
% 4 == 0);
224 if (batch
->next
+ size
> batch
->end
)
225 batch
->extend_cb(batch
, batch
->user_data
);
227 assert(batch
->next
+ size
<= batch
->end
);
229 VG(VALGRIND_CHECK_MEM_IS_DEFINED(other
->start
, size
));
230 memcpy(batch
->next
, other
->start
, size
);
232 offset
= batch
->next
- batch
->start
;
233 anv_reloc_list_append(batch
->relocs
, batch
->alloc
,
234 other
->relocs
, offset
);
239 /*-----------------------------------------------------------------------*
240 * Functions related to anv_batch_bo
241 *-----------------------------------------------------------------------*/
244 anv_batch_bo_create(struct anv_cmd_buffer
*cmd_buffer
,
245 struct anv_batch_bo
**bbo_out
)
249 struct anv_batch_bo
*bbo
= anv_alloc(&cmd_buffer
->pool
->alloc
, sizeof(*bbo
),
250 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
252 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
254 result
= anv_bo_pool_alloc(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
255 if (result
!= VK_SUCCESS
)
258 result
= anv_reloc_list_init(&bbo
->relocs
, &cmd_buffer
->pool
->alloc
);
259 if (result
!= VK_SUCCESS
)
267 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
269 anv_free(&cmd_buffer
->pool
->alloc
, bbo
);
275 anv_batch_bo_clone(struct anv_cmd_buffer
*cmd_buffer
,
276 const struct anv_batch_bo
*other_bbo
,
277 struct anv_batch_bo
**bbo_out
)
281 struct anv_batch_bo
*bbo
= anv_alloc(&cmd_buffer
->pool
->alloc
, sizeof(*bbo
),
282 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
284 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
286 result
= anv_bo_pool_alloc(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
287 if (result
!= VK_SUCCESS
)
290 result
= anv_reloc_list_init_clone(&bbo
->relocs
, &cmd_buffer
->pool
->alloc
,
292 if (result
!= VK_SUCCESS
)
295 bbo
->length
= other_bbo
->length
;
296 memcpy(bbo
->bo
.map
, other_bbo
->bo
.map
, other_bbo
->length
);
298 bbo
->last_ss_pool_bo_offset
= other_bbo
->last_ss_pool_bo_offset
;
305 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
307 anv_free(&cmd_buffer
->pool
->alloc
, bbo
);
313 anv_batch_bo_start(struct anv_batch_bo
*bbo
, struct anv_batch
*batch
,
314 size_t batch_padding
)
316 batch
->next
= batch
->start
= bbo
->bo
.map
;
317 batch
->end
= bbo
->bo
.map
+ bbo
->bo
.size
- batch_padding
;
318 batch
->relocs
= &bbo
->relocs
;
319 bbo
->last_ss_pool_bo_offset
= 0;
320 bbo
->relocs
.num_relocs
= 0;
324 anv_batch_bo_continue(struct anv_batch_bo
*bbo
, struct anv_batch
*batch
,
325 size_t batch_padding
)
327 batch
->start
= bbo
->bo
.map
;
328 batch
->next
= bbo
->bo
.map
+ bbo
->length
;
329 batch
->end
= bbo
->bo
.map
+ bbo
->bo
.size
- batch_padding
;
330 batch
->relocs
= &bbo
->relocs
;
334 anv_batch_bo_finish(struct anv_batch_bo
*bbo
, struct anv_batch
*batch
)
336 assert(batch
->start
== bbo
->bo
.map
);
337 bbo
->length
= batch
->next
- batch
->start
;
338 VG(VALGRIND_CHECK_MEM_IS_DEFINED(batch
->start
, bbo
->length
));
342 anv_batch_bo_destroy(struct anv_batch_bo
*bbo
,
343 struct anv_cmd_buffer
*cmd_buffer
)
345 anv_reloc_list_finish(&bbo
->relocs
, &cmd_buffer
->pool
->alloc
);
346 anv_bo_pool_free(&cmd_buffer
->device
->batch_bo_pool
, &bbo
->bo
);
347 anv_free(&cmd_buffer
->pool
->alloc
, bbo
);
351 anv_batch_bo_list_clone(const struct list_head
*list
,
352 struct anv_cmd_buffer
*cmd_buffer
,
353 struct list_head
*new_list
)
355 VkResult result
= VK_SUCCESS
;
357 list_inithead(new_list
);
359 struct anv_batch_bo
*prev_bbo
= NULL
;
360 list_for_each_entry(struct anv_batch_bo
, bbo
, list
, link
) {
361 struct anv_batch_bo
*new_bbo
= NULL
;
362 result
= anv_batch_bo_clone(cmd_buffer
, bbo
, &new_bbo
);
363 if (result
!= VK_SUCCESS
)
365 list_addtail(&new_bbo
->link
, new_list
);
368 /* As we clone this list of batch_bo's, they chain one to the
369 * other using MI_BATCH_BUFFER_START commands. We need to fix up
370 * those relocations as we go. Fortunately, this is pretty easy
371 * as it will always be the last relocation in the list.
373 uint32_t last_idx
= prev_bbo
->relocs
.num_relocs
- 1;
374 assert(prev_bbo
->relocs
.reloc_bos
[last_idx
] == &bbo
->bo
);
375 prev_bbo
->relocs
.reloc_bos
[last_idx
] = &new_bbo
->bo
;
381 if (result
!= VK_SUCCESS
) {
382 list_for_each_entry_safe(struct anv_batch_bo
, bbo
, new_list
, link
)
383 anv_batch_bo_destroy(bbo
, cmd_buffer
);
389 /*-----------------------------------------------------------------------*
390 * Functions related to anv_batch_bo
391 *-----------------------------------------------------------------------*/
393 static inline struct anv_batch_bo
*
394 anv_cmd_buffer_current_batch_bo(struct anv_cmd_buffer
*cmd_buffer
)
396 return LIST_ENTRY(struct anv_batch_bo
, cmd_buffer
->batch_bos
.prev
, link
);
400 anv_cmd_buffer_surface_base_address(struct anv_cmd_buffer
*cmd_buffer
)
402 return (struct anv_address
) {
403 .bo
= &cmd_buffer
->device
->surface_state_block_pool
.bo
,
404 .offset
= *(int32_t *)anv_vector_head(&cmd_buffer
->bt_blocks
),
409 emit_batch_buffer_start(struct anv_cmd_buffer
*cmd_buffer
,
410 struct anv_bo
*bo
, uint32_t offset
)
412 /* In gen8+ the address field grew to two dwords to accomodate 48 bit
413 * offsets. The high 16 bits are in the last dword, so we can use the gen8
414 * version in either case, as long as we set the instruction length in the
415 * header accordingly. This means that we always emit three dwords here
416 * and all the padding and adjustment we do in this file works for all
420 const uint32_t gen7_length
=
421 GEN7_MI_BATCH_BUFFER_START_length
- GEN7_MI_BATCH_BUFFER_START_length_bias
;
422 const uint32_t gen8_length
=
423 GEN8_MI_BATCH_BUFFER_START_length
- GEN8_MI_BATCH_BUFFER_START_length_bias
;
425 anv_batch_emit(&cmd_buffer
->batch
, GEN8_MI_BATCH_BUFFER_START
,
426 .DWordLength
= cmd_buffer
->device
->info
.gen
< 8 ?
427 gen7_length
: gen8_length
,
428 ._2ndLevelBatchBuffer
= _1stlevelbatch
,
429 .AddressSpaceIndicator
= ASI_PPGTT
,
430 .BatchBufferStartAddress
= { bo
, offset
});
434 cmd_buffer_chain_to_batch_bo(struct anv_cmd_buffer
*cmd_buffer
,
435 struct anv_batch_bo
*bbo
)
437 struct anv_batch
*batch
= &cmd_buffer
->batch
;
438 struct anv_batch_bo
*current_bbo
=
439 anv_cmd_buffer_current_batch_bo(cmd_buffer
);
441 /* We set the end of the batch a little short so we would be sure we
442 * have room for the chaining command. Since we're about to emit the
443 * chaining command, let's set it back where it should go.
445 batch
->end
+= GEN8_MI_BATCH_BUFFER_START_length
* 4;
446 assert(batch
->end
== current_bbo
->bo
.map
+ current_bbo
->bo
.size
);
448 emit_batch_buffer_start(cmd_buffer
, &bbo
->bo
, 0);
450 anv_batch_bo_finish(current_bbo
, batch
);
454 anv_cmd_buffer_chain_batch(struct anv_batch
*batch
, void *_data
)
456 struct anv_cmd_buffer
*cmd_buffer
= _data
;
457 struct anv_batch_bo
*new_bbo
;
459 VkResult result
= anv_batch_bo_create(cmd_buffer
, &new_bbo
);
460 if (result
!= VK_SUCCESS
)
463 struct anv_batch_bo
**seen_bbo
= anv_vector_add(&cmd_buffer
->seen_bbos
);
464 if (seen_bbo
== NULL
) {
465 anv_batch_bo_destroy(new_bbo
, cmd_buffer
);
466 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
470 cmd_buffer_chain_to_batch_bo(cmd_buffer
, new_bbo
);
472 list_addtail(&new_bbo
->link
, &cmd_buffer
->batch_bos
);
474 anv_batch_bo_start(new_bbo
, batch
, GEN8_MI_BATCH_BUFFER_START_length
* 4);
480 anv_cmd_buffer_alloc_binding_table(struct anv_cmd_buffer
*cmd_buffer
,
481 uint32_t entries
, uint32_t *state_offset
)
483 struct anv_block_pool
*block_pool
=
484 &cmd_buffer
->device
->surface_state_block_pool
;
485 int32_t *bt_block
= anv_vector_head(&cmd_buffer
->bt_blocks
);
486 struct anv_state state
;
488 state
.alloc_size
= align_u32(entries
* 4, 32);
490 if (cmd_buffer
->bt_next
+ state
.alloc_size
> block_pool
->block_size
)
491 return (struct anv_state
) { 0 };
493 state
.offset
= cmd_buffer
->bt_next
;
494 state
.map
= block_pool
->map
+ *bt_block
+ state
.offset
;
496 cmd_buffer
->bt_next
+= state
.alloc_size
;
498 assert(*bt_block
< 0);
499 *state_offset
= -(*bt_block
);
505 anv_cmd_buffer_alloc_surface_state(struct anv_cmd_buffer
*cmd_buffer
)
507 return anv_state_stream_alloc(&cmd_buffer
->surface_state_stream
, 64, 64);
511 anv_cmd_buffer_alloc_dynamic_state(struct anv_cmd_buffer
*cmd_buffer
,
512 uint32_t size
, uint32_t alignment
)
514 return anv_state_stream_alloc(&cmd_buffer
->dynamic_state_stream
,
519 anv_cmd_buffer_new_binding_table_block(struct anv_cmd_buffer
*cmd_buffer
)
521 struct anv_block_pool
*block_pool
=
522 &cmd_buffer
->device
->surface_state_block_pool
;
524 int32_t *offset
= anv_vector_add(&cmd_buffer
->bt_blocks
);
526 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
528 *offset
= anv_block_pool_alloc_back(block_pool
);
529 cmd_buffer
->bt_next
= 0;
535 anv_cmd_buffer_init_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
)
537 struct anv_batch_bo
*batch_bo
;
540 list_inithead(&cmd_buffer
->batch_bos
);
542 result
= anv_batch_bo_create(cmd_buffer
, &batch_bo
);
543 if (result
!= VK_SUCCESS
)
546 list_addtail(&batch_bo
->link
, &cmd_buffer
->batch_bos
);
548 cmd_buffer
->batch
.alloc
= &cmd_buffer
->pool
->alloc
;
549 cmd_buffer
->batch
.extend_cb
= anv_cmd_buffer_chain_batch
;
550 cmd_buffer
->batch
.user_data
= cmd_buffer
;
552 anv_batch_bo_start(batch_bo
, &cmd_buffer
->batch
,
553 GEN8_MI_BATCH_BUFFER_START_length
* 4);
555 int success
= anv_vector_init(&cmd_buffer
->seen_bbos
,
556 sizeof(struct anv_bo
*),
557 8 * sizeof(struct anv_bo
*));
561 *(struct anv_batch_bo
**)anv_vector_add(&cmd_buffer
->seen_bbos
) = batch_bo
;
563 success
= anv_vector_init(&cmd_buffer
->bt_blocks
, sizeof(int32_t),
564 8 * sizeof(int32_t));
568 result
= anv_reloc_list_init(&cmd_buffer
->surface_relocs
,
569 &cmd_buffer
->pool
->alloc
);
570 if (result
!= VK_SUCCESS
)
573 anv_cmd_buffer_new_binding_table_block(cmd_buffer
);
575 cmd_buffer
->execbuf2
.objects
= NULL
;
576 cmd_buffer
->execbuf2
.bos
= NULL
;
577 cmd_buffer
->execbuf2
.array_length
= 0;
582 anv_vector_finish(&cmd_buffer
->bt_blocks
);
584 anv_vector_finish(&cmd_buffer
->seen_bbos
);
586 anv_batch_bo_destroy(batch_bo
, cmd_buffer
);
592 anv_cmd_buffer_fini_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
)
595 anv_vector_foreach(bt_block
, &cmd_buffer
->bt_blocks
) {
596 anv_block_pool_free(&cmd_buffer
->device
->surface_state_block_pool
,
599 anv_vector_finish(&cmd_buffer
->bt_blocks
);
601 anv_reloc_list_finish(&cmd_buffer
->surface_relocs
, &cmd_buffer
->pool
->alloc
);
603 anv_vector_finish(&cmd_buffer
->seen_bbos
);
605 /* Destroy all of the batch buffers */
606 list_for_each_entry_safe(struct anv_batch_bo
, bbo
,
607 &cmd_buffer
->batch_bos
, link
) {
608 anv_batch_bo_destroy(bbo
, cmd_buffer
);
611 anv_free(&cmd_buffer
->pool
->alloc
, cmd_buffer
->execbuf2
.objects
);
612 anv_free(&cmd_buffer
->pool
->alloc
, cmd_buffer
->execbuf2
.bos
);
616 anv_cmd_buffer_reset_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
)
618 /* Delete all but the first batch bo */
619 assert(!list_empty(&cmd_buffer
->batch_bos
));
620 while (cmd_buffer
->batch_bos
.next
!= cmd_buffer
->batch_bos
.prev
) {
621 struct anv_batch_bo
*bbo
= anv_cmd_buffer_current_batch_bo(cmd_buffer
);
622 list_del(&bbo
->link
);
623 anv_batch_bo_destroy(bbo
, cmd_buffer
);
625 assert(!list_empty(&cmd_buffer
->batch_bos
));
627 anv_batch_bo_start(anv_cmd_buffer_current_batch_bo(cmd_buffer
),
629 GEN8_MI_BATCH_BUFFER_START_length
* 4);
631 while (anv_vector_length(&cmd_buffer
->bt_blocks
) > 1) {
632 int32_t *bt_block
= anv_vector_remove(&cmd_buffer
->bt_blocks
);
633 anv_block_pool_free(&cmd_buffer
->device
->surface_state_block_pool
,
636 assert(anv_vector_length(&cmd_buffer
->bt_blocks
) == 1);
637 cmd_buffer
->bt_next
= 0;
639 cmd_buffer
->surface_relocs
.num_relocs
= 0;
641 /* Reset the list of seen buffers */
642 cmd_buffer
->seen_bbos
.head
= 0;
643 cmd_buffer
->seen_bbos
.tail
= 0;
645 *(struct anv_batch_bo
**)anv_vector_add(&cmd_buffer
->seen_bbos
) =
646 anv_cmd_buffer_current_batch_bo(cmd_buffer
);
650 anv_cmd_buffer_end_batch_buffer(struct anv_cmd_buffer
*cmd_buffer
)
652 struct anv_batch_bo
*batch_bo
= anv_cmd_buffer_current_batch_bo(cmd_buffer
);
654 if (cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
) {
655 /* When we start a batch buffer, we subtract a certain amount of
656 * padding from the end to ensure that we always have room to emit a
657 * BATCH_BUFFER_START to chain to the next BO. We need to remove
658 * that padding before we end the batch; otherwise, we may end up
659 * with our BATCH_BUFFER_END in another BO.
661 cmd_buffer
->batch
.end
+= GEN8_MI_BATCH_BUFFER_START_length
* 4;
662 assert(cmd_buffer
->batch
.end
== batch_bo
->bo
.map
+ batch_bo
->bo
.size
);
664 anv_batch_emit(&cmd_buffer
->batch
, GEN7_MI_BATCH_BUFFER_END
);
666 /* Round batch up to an even number of dwords. */
667 if ((cmd_buffer
->batch
.next
- cmd_buffer
->batch
.start
) & 4)
668 anv_batch_emit(&cmd_buffer
->batch
, GEN7_MI_NOOP
);
670 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_PRIMARY
;
673 anv_batch_bo_finish(batch_bo
, &cmd_buffer
->batch
);
675 if (cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_SECONDARY
) {
676 /* If this is a secondary command buffer, we need to determine the
677 * mode in which it will be executed with vkExecuteCommands. We
678 * determine this statically here so that this stays in sync with the
679 * actual ExecuteCommands implementation.
681 if ((cmd_buffer
->batch_bos
.next
== cmd_buffer
->batch_bos
.prev
) &&
682 (batch_bo
->length
< ANV_CMD_BUFFER_BATCH_SIZE
/ 2)) {
683 /* If the secondary has exactly one batch buffer in its list *and*
684 * that batch buffer is less than half of the maximum size, we're
685 * probably better of simply copying it into our batch.
687 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_EMIT
;
688 } else if (!(cmd_buffer
->usage_flags
&
689 VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT
)) {
690 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_CHAIN
;
692 /* When we chain, we need to add an MI_BATCH_BUFFER_START command
693 * with its relocation. In order to handle this we'll increment here
694 * so we can unconditionally decrement right before adding the
695 * MI_BATCH_BUFFER_START command.
697 batch_bo
->relocs
.num_relocs
++;
698 cmd_buffer
->batch
.next
+= GEN8_MI_BATCH_BUFFER_START_length
* 4;
700 cmd_buffer
->exec_mode
= ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN
;
705 static inline VkResult
706 anv_cmd_buffer_add_seen_bbos(struct anv_cmd_buffer
*cmd_buffer
,
707 struct list_head
*list
)
709 list_for_each_entry(struct anv_batch_bo
, bbo
, list
, link
) {
710 struct anv_batch_bo
**bbo_ptr
= anv_vector_add(&cmd_buffer
->seen_bbos
);
712 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
721 anv_cmd_buffer_add_secondary(struct anv_cmd_buffer
*primary
,
722 struct anv_cmd_buffer
*secondary
)
724 switch (secondary
->exec_mode
) {
725 case ANV_CMD_BUFFER_EXEC_MODE_EMIT
:
726 anv_batch_emit_batch(&primary
->batch
, &secondary
->batch
);
727 anv_cmd_buffer_emit_state_base_address(primary
);
729 case ANV_CMD_BUFFER_EXEC_MODE_CHAIN
: {
730 struct anv_batch_bo
*first_bbo
=
731 list_first_entry(&secondary
->batch_bos
, struct anv_batch_bo
, link
);
732 struct anv_batch_bo
*last_bbo
=
733 list_last_entry(&secondary
->batch_bos
, struct anv_batch_bo
, link
);
735 emit_batch_buffer_start(primary
, &first_bbo
->bo
, 0);
737 struct anv_batch_bo
*this_bbo
= anv_cmd_buffer_current_batch_bo(primary
);
738 assert(primary
->batch
.start
== this_bbo
->bo
.map
);
739 uint32_t offset
= primary
->batch
.next
- primary
->batch
.start
;
740 const uint32_t inst_size
= GEN8_MI_BATCH_BUFFER_START_length
* 4;
742 /* Roll back the previous MI_BATCH_BUFFER_START and its relocation so we
743 * can emit a new command and relocation for the current splice. In
744 * order to handle the initial-use case, we incremented next and
745 * num_relocs in end_batch_buffer() so we can alyways just subtract
748 last_bbo
->relocs
.num_relocs
--;
749 secondary
->batch
.next
-= inst_size
;
750 emit_batch_buffer_start(secondary
, &this_bbo
->bo
, offset
);
751 anv_cmd_buffer_add_seen_bbos(primary
, &secondary
->batch_bos
);
753 /* After patching up the secondary buffer, we need to clflush the
754 * modified instruction in case we're on a !llc platform. We use a
755 * little loop to handle the case where the instruction crosses a cache
758 if (!primary
->device
->info
.has_llc
) {
759 void *inst
= secondary
->batch
.next
- inst_size
;
760 void *p
= (void *) (((uintptr_t) inst
) & ~CACHELINE_MASK
);
761 __builtin_ia32_mfence();
762 while (p
< secondary
->batch
.next
) {
763 __builtin_ia32_clflush(p
);
768 anv_cmd_buffer_emit_state_base_address(primary
);
771 case ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN
: {
772 struct list_head copy_list
;
773 VkResult result
= anv_batch_bo_list_clone(&secondary
->batch_bos
,
776 if (result
!= VK_SUCCESS
)
779 anv_cmd_buffer_add_seen_bbos(primary
, ©_list
);
781 struct anv_batch_bo
*first_bbo
=
782 list_first_entry(©_list
, struct anv_batch_bo
, link
);
783 struct anv_batch_bo
*last_bbo
=
784 list_last_entry(©_list
, struct anv_batch_bo
, link
);
786 cmd_buffer_chain_to_batch_bo(primary
, first_bbo
);
788 list_splicetail(©_list
, &primary
->batch_bos
);
790 anv_batch_bo_continue(last_bbo
, &primary
->batch
,
791 GEN8_MI_BATCH_BUFFER_START_length
* 4);
793 anv_cmd_buffer_emit_state_base_address(primary
);
797 assert(!"Invalid execution mode");
800 anv_reloc_list_append(&primary
->surface_relocs
, &primary
->pool
->alloc
,
801 &secondary
->surface_relocs
, 0);
805 anv_cmd_buffer_add_bo(struct anv_cmd_buffer
*cmd_buffer
,
807 struct anv_reloc_list
*relocs
)
809 struct drm_i915_gem_exec_object2
*obj
= NULL
;
811 if (bo
->index
< cmd_buffer
->execbuf2
.bo_count
&&
812 cmd_buffer
->execbuf2
.bos
[bo
->index
] == bo
)
813 obj
= &cmd_buffer
->execbuf2
.objects
[bo
->index
];
816 /* We've never seen this one before. Add it to the list and assign
817 * an id that we can use later.
819 if (cmd_buffer
->execbuf2
.bo_count
>= cmd_buffer
->execbuf2
.array_length
) {
820 uint32_t new_len
= cmd_buffer
->execbuf2
.objects
?
821 cmd_buffer
->execbuf2
.array_length
* 2 : 64;
823 struct drm_i915_gem_exec_object2
*new_objects
=
824 anv_alloc(&cmd_buffer
->pool
->alloc
, new_len
* sizeof(*new_objects
),
825 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
826 if (new_objects
== NULL
)
827 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
829 struct anv_bo
**new_bos
=
830 anv_alloc(&cmd_buffer
->pool
->alloc
, new_len
* sizeof(*new_bos
),
831 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
832 if (new_objects
== NULL
) {
833 anv_free(&cmd_buffer
->pool
->alloc
, new_objects
);
834 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
837 if (cmd_buffer
->execbuf2
.objects
) {
838 memcpy(new_objects
, cmd_buffer
->execbuf2
.objects
,
839 cmd_buffer
->execbuf2
.bo_count
* sizeof(*new_objects
));
840 memcpy(new_bos
, cmd_buffer
->execbuf2
.bos
,
841 cmd_buffer
->execbuf2
.bo_count
* sizeof(*new_bos
));
844 cmd_buffer
->execbuf2
.objects
= new_objects
;
845 cmd_buffer
->execbuf2
.bos
= new_bos
;
846 cmd_buffer
->execbuf2
.array_length
= new_len
;
849 assert(cmd_buffer
->execbuf2
.bo_count
< cmd_buffer
->execbuf2
.array_length
);
851 bo
->index
= cmd_buffer
->execbuf2
.bo_count
++;
852 obj
= &cmd_buffer
->execbuf2
.objects
[bo
->index
];
853 cmd_buffer
->execbuf2
.bos
[bo
->index
] = bo
;
855 obj
->handle
= bo
->gem_handle
;
856 obj
->relocation_count
= 0;
859 obj
->offset
= bo
->offset
;
860 obj
->flags
= bo
->is_winsys_bo
? EXEC_OBJECT_WRITE
: 0;
865 if (relocs
!= NULL
&& obj
->relocation_count
== 0) {
866 /* This is the first time we've ever seen a list of relocations for
867 * this BO. Go ahead and set the relocations and then walk the list
868 * of relocations and add them all.
870 obj
->relocation_count
= relocs
->num_relocs
;
871 obj
->relocs_ptr
= (uintptr_t) relocs
->relocs
;
873 for (size_t i
= 0; i
< relocs
->num_relocs
; i
++) {
874 /* A quick sanity check on relocations */
875 assert(relocs
->relocs
[i
].offset
< bo
->size
);
876 anv_cmd_buffer_add_bo(cmd_buffer
, relocs
->reloc_bos
[i
], NULL
);
884 anv_cmd_buffer_process_relocs(struct anv_cmd_buffer
*cmd_buffer
,
885 struct anv_reloc_list
*list
)
889 /* If the kernel supports I915_EXEC_NO_RELOC, it will compare offset in
890 * struct drm_i915_gem_exec_object2 against the bos current offset and if
891 * all bos haven't moved it will skip relocation processing alltogether.
892 * If I915_EXEC_NO_RELOC is not supported, the kernel ignores the incoming
893 * value of offset so we can set it either way. For that to work we need
894 * to make sure all relocs use the same presumed offset.
897 for (size_t i
= 0; i
< list
->num_relocs
; i
++) {
898 bo
= list
->reloc_bos
[i
];
899 if (bo
->offset
!= list
->relocs
[i
].presumed_offset
)
900 cmd_buffer
->execbuf2
.need_reloc
= true;
902 list
->relocs
[i
].target_handle
= bo
->index
;
907 read_reloc(const struct anv_device
*device
, const void *p
)
909 if (device
->info
.gen
>= 8)
910 return *(uint64_t *)p
;
912 return *(uint32_t *)p
;
916 write_reloc(const struct anv_device
*device
, void *p
, uint64_t v
)
918 if (device
->info
.gen
>= 8)
925 adjust_relocations_from_block_pool(struct anv_block_pool
*pool
,
926 struct anv_reloc_list
*relocs
)
928 for (size_t i
= 0; i
< relocs
->num_relocs
; i
++) {
929 /* In general, we don't know how stale the relocated value is. It
930 * may have been used last time or it may not. Since we don't want
931 * to stomp it while the GPU may be accessing it, we haven't updated
932 * it anywhere else in the code. Instead, we just set the presumed
933 * offset to what it is now based on the delta and the data in the
934 * block pool. Then the kernel will update it for us if needed.
936 assert(relocs
->relocs
[i
].offset
< pool
->state
.end
);
937 const void *p
= pool
->map
+ relocs
->relocs
[i
].offset
;
939 /* We're reading back the relocated value from potentially incoherent
940 * memory here. However, any change to the value will be from the kernel
941 * writing out relocations, which will keep the CPU cache up to date.
943 relocs
->relocs
[i
].presumed_offset
=
944 read_reloc(pool
->device
, p
) - relocs
->relocs
[i
].delta
;
946 /* All of the relocations from this block pool to other BO's should
947 * have been emitted relative to the surface block pool center. We
948 * need to add the center offset to make them relative to the
949 * beginning of the actual GEM bo.
951 relocs
->relocs
[i
].offset
+= pool
->center_bo_offset
;
956 adjust_relocations_to_block_pool(struct anv_block_pool
*pool
,
957 struct anv_bo
*from_bo
,
958 struct anv_reloc_list
*relocs
,
959 uint32_t *last_pool_center_bo_offset
)
961 assert(*last_pool_center_bo_offset
<= pool
->center_bo_offset
);
962 uint32_t delta
= pool
->center_bo_offset
- *last_pool_center_bo_offset
;
964 /* When we initially emit relocations into a block pool, we don't
965 * actually know what the final center_bo_offset will be so we just emit
966 * it as if center_bo_offset == 0. Now that we know what the center
967 * offset is, we need to walk the list of relocations and adjust any
968 * relocations that point to the pool bo with the correct offset.
970 for (size_t i
= 0; i
< relocs
->num_relocs
; i
++) {
971 if (relocs
->reloc_bos
[i
] == &pool
->bo
) {
972 /* Adjust the delta value in the relocation to correctly
973 * correspond to the new delta. Initially, this value may have
974 * been negative (if treated as unsigned), but we trust in
975 * uint32_t roll-over to fix that for us at this point.
977 relocs
->relocs
[i
].delta
+= delta
;
979 /* Since the delta has changed, we need to update the actual
980 * relocated value with the new presumed value. This function
981 * should only be called on batch buffers, so we know it isn't in
982 * use by the GPU at the moment.
984 assert(relocs
->relocs
[i
].offset
< from_bo
->size
);
985 write_reloc(pool
->device
, from_bo
->map
+ relocs
->relocs
[i
].offset
,
986 relocs
->relocs
[i
].presumed_offset
+
987 relocs
->relocs
[i
].delta
);
991 *last_pool_center_bo_offset
= pool
->center_bo_offset
;
995 anv_cmd_buffer_prepare_execbuf(struct anv_cmd_buffer
*cmd_buffer
)
997 struct anv_batch
*batch
= &cmd_buffer
->batch
;
998 struct anv_block_pool
*ss_pool
=
999 &cmd_buffer
->device
->surface_state_block_pool
;
1001 cmd_buffer
->execbuf2
.bo_count
= 0;
1002 cmd_buffer
->execbuf2
.need_reloc
= false;
1004 adjust_relocations_from_block_pool(ss_pool
, &cmd_buffer
->surface_relocs
);
1005 anv_cmd_buffer_add_bo(cmd_buffer
, &ss_pool
->bo
, &cmd_buffer
->surface_relocs
);
1007 /* First, we walk over all of the bos we've seen and add them and their
1008 * relocations to the validate list.
1010 struct anv_batch_bo
**bbo
;
1011 anv_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
) {
1012 adjust_relocations_to_block_pool(ss_pool
, &(*bbo
)->bo
, &(*bbo
)->relocs
,
1013 &(*bbo
)->last_ss_pool_bo_offset
);
1015 anv_cmd_buffer_add_bo(cmd_buffer
, &(*bbo
)->bo
, &(*bbo
)->relocs
);
1018 struct anv_batch_bo
*first_batch_bo
=
1019 list_first_entry(&cmd_buffer
->batch_bos
, struct anv_batch_bo
, link
);
1021 /* The kernel requires that the last entry in the validation list be the
1022 * batch buffer to execute. We can simply swap the element
1023 * corresponding to the first batch_bo in the chain with the last
1024 * element in the list.
1026 if (first_batch_bo
->bo
.index
!= cmd_buffer
->execbuf2
.bo_count
- 1) {
1027 uint32_t idx
= first_batch_bo
->bo
.index
;
1028 uint32_t last_idx
= cmd_buffer
->execbuf2
.bo_count
- 1;
1030 struct drm_i915_gem_exec_object2 tmp_obj
=
1031 cmd_buffer
->execbuf2
.objects
[idx
];
1032 assert(cmd_buffer
->execbuf2
.bos
[idx
] == &first_batch_bo
->bo
);
1034 cmd_buffer
->execbuf2
.objects
[idx
] = cmd_buffer
->execbuf2
.objects
[last_idx
];
1035 cmd_buffer
->execbuf2
.bos
[idx
] = cmd_buffer
->execbuf2
.bos
[last_idx
];
1036 cmd_buffer
->execbuf2
.bos
[idx
]->index
= idx
;
1038 cmd_buffer
->execbuf2
.objects
[last_idx
] = tmp_obj
;
1039 cmd_buffer
->execbuf2
.bos
[last_idx
] = &first_batch_bo
->bo
;
1040 first_batch_bo
->bo
.index
= last_idx
;
1043 /* Now we go through and fixup all of the relocation lists to point to
1044 * the correct indices in the object array. We have to do this after we
1045 * reorder the list above as some of the indices may have changed.
1047 anv_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
)
1048 anv_cmd_buffer_process_relocs(cmd_buffer
, &(*bbo
)->relocs
);
1050 anv_cmd_buffer_process_relocs(cmd_buffer
, &cmd_buffer
->surface_relocs
);
1052 if (!cmd_buffer
->device
->info
.has_llc
) {
1053 __builtin_ia32_mfence();
1054 anv_vector_foreach(bbo
, &cmd_buffer
->seen_bbos
) {
1055 for (uint32_t i
= 0; i
< (*bbo
)->length
; i
+= CACHELINE_SIZE
)
1056 __builtin_ia32_clflush((*bbo
)->bo
.map
+ i
);
1060 cmd_buffer
->execbuf2
.execbuf
= (struct drm_i915_gem_execbuffer2
) {
1061 .buffers_ptr
= (uintptr_t) cmd_buffer
->execbuf2
.objects
,
1062 .buffer_count
= cmd_buffer
->execbuf2
.bo_count
,
1063 .batch_start_offset
= 0,
1064 .batch_len
= batch
->next
- batch
->start
,
1069 .flags
= I915_EXEC_HANDLE_LUT
| I915_EXEC_RENDER
|
1070 I915_EXEC_CONSTANTS_REL_GENERAL
,
1071 .rsvd1
= cmd_buffer
->device
->context_id
,
1075 if (!cmd_buffer
->execbuf2
.need_reloc
)
1076 cmd_buffer
->execbuf2
.execbuf
.flags
|= I915_EXEC_NO_RELOC
;