2 * Copyright (C) 2019 Alyssa Rosenzweig
3 * Copyright (C) 2014-2017 Broadcom
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
28 #include "drm-uapi/panfrost_drm.h"
31 #include "pan_context.h"
32 #include "util/hash_table.h"
33 #include "util/ralloc.h"
34 #include "util/format/u_format.h"
35 #include "util/u_pack_color.h"
37 #include "pandecode/decode.h"
39 /* panfrost_bo_access is here to help us keep track of batch accesses to BOs
40 * and build a proper dependency graph such that batches can be pipelined for
41 * better GPU utilization.
43 * Each accessed BO has a corresponding entry in the ->accessed_bos hash table.
44 * A BO is either being written or read at any time, that's what the type field
46 * When the last access is a write, the batch writing the BO might have read
47 * dependencies (readers that have not been executed yet and want to read the
48 * previous BO content), and when the last access is a read, all readers might
49 * depend on another batch to push its results to memory. That's what the
50 * readers/writers keep track off.
51 * There can only be one writer at any given time, if a new batch wants to
52 * write to the same BO, a dependency will be added between the new writer and
53 * the old writer (at the batch level), and panfrost_bo_access->writer will be
54 * updated to point to the new writer.
56 struct panfrost_bo_access
{
58 struct util_dynarray readers
;
59 struct panfrost_batch_fence
*writer
;
62 static struct panfrost_batch_fence
*
63 panfrost_create_batch_fence(struct panfrost_batch
*batch
)
65 struct panfrost_batch_fence
*fence
;
68 fence
= rzalloc(NULL
, struct panfrost_batch_fence
);
70 pipe_reference_init(&fence
->reference
, 1);
71 fence
->ctx
= batch
->ctx
;
73 ret
= drmSyncobjCreate(pan_screen(batch
->ctx
->base
.screen
)->fd
, 0,
81 panfrost_free_batch_fence(struct panfrost_batch_fence
*fence
)
83 drmSyncobjDestroy(pan_screen(fence
->ctx
->base
.screen
)->fd
,
89 panfrost_batch_fence_unreference(struct panfrost_batch_fence
*fence
)
91 if (pipe_reference(&fence
->reference
, NULL
))
92 panfrost_free_batch_fence(fence
);
96 panfrost_batch_fence_reference(struct panfrost_batch_fence
*fence
)
98 pipe_reference(NULL
, &fence
->reference
);
101 static struct panfrost_batch
*
102 panfrost_create_batch(struct panfrost_context
*ctx
,
103 const struct pipe_framebuffer_state
*key
)
105 struct panfrost_batch
*batch
= rzalloc(ctx
, struct panfrost_batch
);
109 batch
->bos
= _mesa_hash_table_create(batch
, _mesa_hash_pointer
,
110 _mesa_key_pointer_equal
);
112 batch
->minx
= batch
->miny
= ~0;
113 batch
->maxx
= batch
->maxy
= 0;
114 batch
->transient_offset
= 0;
116 util_dynarray_init(&batch
->headers
, batch
);
117 util_dynarray_init(&batch
->gpu_headers
, batch
);
118 util_dynarray_init(&batch
->dependencies
, batch
);
119 batch
->out_sync
= panfrost_create_batch_fence(batch
);
120 util_copy_framebuffer_state(&batch
->key
, key
);
126 panfrost_freeze_batch(struct panfrost_batch
*batch
)
128 struct panfrost_context
*ctx
= batch
->ctx
;
129 struct hash_entry
*entry
;
131 /* Remove the entry in the FBO -> batch hash table if the batch
132 * matches. This way, next draws/clears targeting this FBO will trigger
133 * the creation of a new batch.
135 entry
= _mesa_hash_table_search(ctx
->batches
, &batch
->key
);
136 if (entry
&& entry
->data
== batch
)
137 _mesa_hash_table_remove(ctx
->batches
, entry
);
139 /* If this is the bound batch, the panfrost_context parameters are
140 * relevant so submitting it invalidates those parameters, but if it's
141 * not bound, the context parameters are for some other batch so we
142 * can't invalidate them.
144 if (ctx
->batch
== batch
) {
145 panfrost_invalidate_frame(ctx
);
151 static bool panfrost_batch_is_frozen(struct panfrost_batch
*batch
)
153 struct panfrost_context
*ctx
= batch
->ctx
;
154 struct hash_entry
*entry
;
156 entry
= _mesa_hash_table_search(ctx
->batches
, &batch
->key
);
157 if (entry
&& entry
->data
== batch
)
160 if (ctx
->batch
== batch
)
168 panfrost_free_batch(struct panfrost_batch
*batch
)
173 assert(panfrost_batch_is_frozen(batch
));
175 hash_table_foreach(batch
->bos
, entry
)
176 panfrost_bo_unreference((struct panfrost_bo
*)entry
->key
);
178 util_dynarray_foreach(&batch
->dependencies
,
179 struct panfrost_batch_fence
*, dep
) {
180 panfrost_batch_fence_unreference(*dep
);
183 /* The out_sync fence lifetime is different from the the batch one
184 * since other batches might want to wait on a fence of already
185 * submitted/signaled batch. All we need to do here is make sure the
186 * fence does not point to an invalid batch, which the core will
187 * interpret as 'batch is already submitted'.
189 batch
->out_sync
->batch
= NULL
;
190 panfrost_batch_fence_unreference(batch
->out_sync
);
192 util_unreference_framebuffer_state(&batch
->key
);
198 panfrost_dep_graph_contains_batch(struct panfrost_batch
*root
,
199 struct panfrost_batch
*batch
)
204 util_dynarray_foreach(&root
->dependencies
,
205 struct panfrost_batch_fence
*, dep
) {
206 if ((*dep
)->batch
== batch
||
207 panfrost_dep_graph_contains_batch((*dep
)->batch
, batch
))
216 panfrost_batch_add_dep(struct panfrost_batch
*batch
,
217 struct panfrost_batch_fence
*newdep
)
219 if (batch
== newdep
->batch
)
222 /* We might want to turn ->dependencies into a set if the number of
223 * deps turns out to be big enough to make this 'is dep already there'
224 * search inefficient.
226 util_dynarray_foreach(&batch
->dependencies
,
227 struct panfrost_batch_fence
*, dep
) {
232 /* Make sure the dependency graph is acyclic. */
233 assert(!panfrost_dep_graph_contains_batch(newdep
->batch
, batch
));
235 panfrost_batch_fence_reference(newdep
);
236 util_dynarray_append(&batch
->dependencies
,
237 struct panfrost_batch_fence
*, newdep
);
239 /* We now have a batch depending on us, let's make sure new draw/clear
240 * calls targeting the same FBO use a new batch object.
243 panfrost_freeze_batch(newdep
->batch
);
246 static struct panfrost_batch
*
247 panfrost_get_batch(struct panfrost_context
*ctx
,
248 const struct pipe_framebuffer_state
*key
)
250 /* Lookup the job first */
251 struct hash_entry
*entry
= _mesa_hash_table_search(ctx
->batches
, key
);
256 /* Otherwise, let's create a job */
258 struct panfrost_batch
*batch
= panfrost_create_batch(ctx
, key
);
260 /* Save the created job */
261 _mesa_hash_table_insert(ctx
->batches
, &batch
->key
, batch
);
266 /* Get the job corresponding to the FBO we're currently rendering into */
268 struct panfrost_batch
*
269 panfrost_get_batch_for_fbo(struct panfrost_context
*ctx
)
271 /* If we're wallpapering, we special case to workaround
274 if (ctx
->wallpaper_batch
)
275 return ctx
->wallpaper_batch
;
277 /* If we already began rendering, use that */
280 assert(util_framebuffer_state_equal(&ctx
->batch
->key
,
281 &ctx
->pipe_framebuffer
));
285 /* If not, look up the job */
286 struct panfrost_batch
*batch
= panfrost_get_batch(ctx
,
287 &ctx
->pipe_framebuffer
);
289 /* Set this job as the current FBO job. Will be reset when updating the
290 * FB state and when submitting or releasing a job.
296 struct panfrost_batch
*
297 panfrost_get_fresh_batch_for_fbo(struct panfrost_context
*ctx
)
299 struct panfrost_batch
*batch
;
301 batch
= panfrost_get_batch(ctx
, &ctx
->pipe_framebuffer
);
303 /* The batch has no draw/clear queued, let's return it directly.
304 * Note that it's perfectly fine to re-use a batch with an
305 * existing clear, we'll just update it with the new clear request.
307 if (!batch
->last_job
.gpu
)
310 /* Otherwise, we need to freeze the existing one and instantiate a new
313 panfrost_freeze_batch(batch
);
314 return panfrost_get_batch(ctx
, &ctx
->pipe_framebuffer
);
318 panfrost_batch_fence_is_signaled(struct panfrost_batch_fence
*fence
)
323 /* Batch has not been submitted yet. */
327 int ret
= drmSyncobjWait(pan_screen(fence
->ctx
->base
.screen
)->fd
,
328 &fence
->syncobj
, 1, 0, 0, NULL
);
330 /* Cache whether the fence was signaled */
331 fence
->signaled
= ret
>= 0;
332 return fence
->signaled
;
336 panfrost_bo_access_gc_fences(struct panfrost_context
*ctx
,
337 struct panfrost_bo_access
*access
,
338 const struct panfrost_bo
*bo
)
340 if (access
->writer
&& panfrost_batch_fence_is_signaled(access
->writer
)) {
341 panfrost_batch_fence_unreference(access
->writer
);
342 access
->writer
= NULL
;
345 unsigned nreaders
= 0;
346 util_dynarray_foreach(&access
->readers
, struct panfrost_batch_fence
*,
351 if (panfrost_batch_fence_is_signaled(*reader
)) {
352 panfrost_batch_fence_unreference(*reader
);
360 util_dynarray_clear(&access
->readers
);
363 /* Collect signaled fences to keep the kernel-side syncobj-map small. The
364 * idea is to collect those signaled fences at the end of each flush_all
365 * call. This function is likely to collect only fences from previous
366 * batch flushes not the one that have just have just been submitted and
367 * are probably still in flight when we trigger the garbage collection.
368 * Anyway, we need to do this garbage collection at some point if we don't
369 * want the BO access map to keep invalid entries around and retain
373 panfrost_gc_fences(struct panfrost_context
*ctx
)
375 hash_table_foreach(ctx
->accessed_bos
, entry
) {
376 struct panfrost_bo_access
*access
= entry
->data
;
379 panfrost_bo_access_gc_fences(ctx
, access
, entry
->key
);
380 if (!util_dynarray_num_elements(&access
->readers
,
381 struct panfrost_batch_fence
*) &&
383 _mesa_hash_table_remove(ctx
->accessed_bos
, entry
);
389 panfrost_batch_in_readers(struct panfrost_batch
*batch
,
390 struct panfrost_bo_access
*access
)
392 util_dynarray_foreach(&access
->readers
, struct panfrost_batch_fence
*,
394 if (*reader
&& (*reader
)->batch
== batch
)
403 panfrost_batch_update_bo_access(struct panfrost_batch
*batch
,
404 struct panfrost_bo
*bo
, uint32_t access_type
,
405 bool already_accessed
)
407 struct panfrost_context
*ctx
= batch
->ctx
;
408 struct panfrost_bo_access
*access
;
409 uint32_t old_access_type
;
410 struct hash_entry
*entry
;
412 assert(access_type
== PAN_BO_ACCESS_WRITE
||
413 access_type
== PAN_BO_ACCESS_READ
);
415 entry
= _mesa_hash_table_search(ctx
->accessed_bos
, bo
);
416 access
= entry
? entry
->data
: NULL
;
418 old_access_type
= access
->type
;
420 access
= rzalloc(ctx
, struct panfrost_bo_access
);
421 util_dynarray_init(&access
->readers
, access
);
422 _mesa_hash_table_insert(ctx
->accessed_bos
, bo
, access
);
423 /* We are the first to access this BO, let's initialize
424 * old_access_type to our own access type in that case.
426 old_access_type
= access_type
;
427 access
->type
= access_type
;
432 if (access_type
== PAN_BO_ACCESS_WRITE
&&
433 old_access_type
== PAN_BO_ACCESS_READ
) {
434 /* Previous access was a read and we want to write this BO.
435 * We first need to add explicit deps between our batch and
436 * the previous readers.
438 util_dynarray_foreach(&access
->readers
,
439 struct panfrost_batch_fence
*, reader
) {
440 /* We were already reading the BO, no need to add a dep
441 * on ourself (the acyclic check would complain about
444 if (!(*reader
) || (*reader
)->batch
== batch
)
447 panfrost_batch_add_dep(batch
, *reader
);
449 panfrost_batch_fence_reference(batch
->out_sync
);
451 /* We now are the new writer. */
452 access
->writer
= batch
->out_sync
;
453 access
->type
= access_type
;
455 /* Release the previous readers and reset the readers array. */
456 util_dynarray_foreach(&access
->readers
,
457 struct panfrost_batch_fence
*,
461 panfrost_batch_fence_unreference(*reader
);
464 util_dynarray_clear(&access
->readers
);
465 } else if (access_type
== PAN_BO_ACCESS_WRITE
&&
466 old_access_type
== PAN_BO_ACCESS_WRITE
) {
467 /* Previous access was a write and we want to write this BO.
468 * First check if we were the previous writer, in that case
469 * there's nothing to do. Otherwise we need to add a
470 * dependency between the new writer and the old one.
472 if (access
->writer
!= batch
->out_sync
) {
473 if (access
->writer
) {
474 panfrost_batch_add_dep(batch
, access
->writer
);
475 panfrost_batch_fence_unreference(access
->writer
);
477 panfrost_batch_fence_reference(batch
->out_sync
);
478 access
->writer
= batch
->out_sync
;
480 } else if (access_type
== PAN_BO_ACCESS_READ
&&
481 old_access_type
== PAN_BO_ACCESS_WRITE
) {
482 /* Previous access was a write and we want to read this BO.
483 * First check if we were the previous writer, in that case
484 * we want to keep the access type unchanged, as a write is
485 * more constraining than a read.
487 if (access
->writer
!= batch
->out_sync
) {
488 /* Add a dependency on the previous writer. */
489 panfrost_batch_add_dep(batch
, access
->writer
);
491 /* The previous access was a write, there's no reason
492 * to have entries in the readers array.
494 assert(!util_dynarray_num_elements(&access
->readers
,
495 struct panfrost_batch_fence
*));
497 /* Add ourselves to the readers array. */
498 panfrost_batch_fence_reference(batch
->out_sync
);
499 util_dynarray_append(&access
->readers
,
500 struct panfrost_batch_fence
*,
502 access
->type
= PAN_BO_ACCESS_READ
;
505 /* We already accessed this BO before, so we should already be
506 * in the reader array.
508 if (already_accessed
) {
509 assert(panfrost_batch_in_readers(batch
, access
));
513 /* Previous access was a read and we want to read this BO.
514 * Add ourselves to the readers array and add a dependency on
515 * the previous writer if any.
517 panfrost_batch_fence_reference(batch
->out_sync
);
518 util_dynarray_append(&access
->readers
,
519 struct panfrost_batch_fence
*,
523 panfrost_batch_add_dep(batch
, access
->writer
);
528 panfrost_batch_add_bo(struct panfrost_batch
*batch
, struct panfrost_bo
*bo
,
534 struct hash_entry
*entry
;
535 uint32_t old_flags
= 0;
537 entry
= _mesa_hash_table_search(batch
->bos
, bo
);
539 entry
= _mesa_hash_table_insert(batch
->bos
, bo
,
540 (void *)(uintptr_t)flags
);
541 panfrost_bo_reference(bo
);
543 old_flags
= (uintptr_t)entry
->data
;
545 /* All batches have to agree on the shared flag. */
546 assert((old_flags
& PAN_BO_ACCESS_SHARED
) ==
547 (flags
& PAN_BO_ACCESS_SHARED
));
552 if (old_flags
== flags
)
556 entry
->data
= (void *)(uintptr_t)flags
;
558 /* If this is not a shared BO, we don't really care about dependency
561 if (!(flags
& PAN_BO_ACCESS_SHARED
))
564 /* All dependencies should have been flushed before we execute the
565 * wallpaper draw, so it should be harmless to skip the
566 * update_bo_access() call.
568 if (batch
== batch
->ctx
->wallpaper_batch
)
571 /* Only pass R/W flags to the dep tracking logic. */
572 assert(flags
& PAN_BO_ACCESS_RW
);
573 flags
= (flags
& PAN_BO_ACCESS_WRITE
) ?
574 PAN_BO_ACCESS_WRITE
: PAN_BO_ACCESS_READ
;
575 panfrost_batch_update_bo_access(batch
, bo
, flags
, old_flags
!= 0);
578 void panfrost_batch_add_fbo_bos(struct panfrost_batch
*batch
)
580 uint32_t flags
= PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_WRITE
|
581 PAN_BO_ACCESS_VERTEX_TILER
|
582 PAN_BO_ACCESS_FRAGMENT
;
584 for (unsigned i
= 0; i
< batch
->key
.nr_cbufs
; ++i
) {
585 struct panfrost_resource
*rsrc
= pan_resource(batch
->key
.cbufs
[i
]->texture
);
586 panfrost_batch_add_bo(batch
, rsrc
->bo
, flags
);
589 if (batch
->key
.zsbuf
) {
590 struct panfrost_resource
*rsrc
= pan_resource(batch
->key
.zsbuf
->texture
);
591 panfrost_batch_add_bo(batch
, rsrc
->bo
, flags
);
596 panfrost_batch_create_bo(struct panfrost_batch
*batch
, size_t size
,
597 uint32_t create_flags
, uint32_t access_flags
)
599 struct panfrost_bo
*bo
;
601 bo
= panfrost_bo_create(pan_screen(batch
->ctx
->base
.screen
), size
,
603 panfrost_batch_add_bo(batch
, bo
, access_flags
);
605 /* panfrost_batch_add_bo() has retained a reference and
606 * panfrost_bo_create() initialize the refcnt to 1, so let's
607 * unreference the BO here so it gets released when the batch is
608 * destroyed (unless it's retained by someone else in the meantime).
610 panfrost_bo_unreference(bo
);
614 /* Returns the polygon list's GPU address if available, or otherwise allocates
615 * the polygon list. It's perfectly fast to use allocate/free BO directly,
616 * since we'll hit the BO cache and this is one-per-batch anyway. */
619 panfrost_batch_get_polygon_list(struct panfrost_batch
*batch
, unsigned size
)
621 if (batch
->polygon_list
) {
622 assert(batch
->polygon_list
->size
>= size
);
624 /* Create the BO as invisible, as there's no reason to map */
626 batch
->polygon_list
= panfrost_batch_create_bo(batch
, size
,
628 PAN_BO_ACCESS_PRIVATE
|
630 PAN_BO_ACCESS_VERTEX_TILER
|
631 PAN_BO_ACCESS_FRAGMENT
);
634 return batch
->polygon_list
->gpu
;
638 panfrost_batch_get_scratchpad(struct panfrost_batch
*batch
)
640 if (batch
->scratchpad
)
641 return batch
->scratchpad
;
643 batch
->scratchpad
= panfrost_batch_create_bo(batch
, 64 * 4 * 4096,
645 PAN_BO_ACCESS_PRIVATE
|
647 PAN_BO_ACCESS_VERTEX_TILER
|
648 PAN_BO_ACCESS_FRAGMENT
);
649 assert(batch
->scratchpad
);
650 return batch
->scratchpad
;
654 panfrost_batch_get_tiler_heap(struct panfrost_batch
*batch
)
656 if (batch
->tiler_heap
)
657 return batch
->tiler_heap
;
659 batch
->tiler_heap
= panfrost_batch_create_bo(batch
, 4096 * 4096,
662 PAN_BO_ACCESS_PRIVATE
|
664 PAN_BO_ACCESS_VERTEX_TILER
|
665 PAN_BO_ACCESS_FRAGMENT
);
666 assert(batch
->tiler_heap
);
667 return batch
->tiler_heap
;
671 panfrost_batch_get_tiler_dummy(struct panfrost_batch
*batch
)
673 struct panfrost_screen
*screen
= pan_screen(batch
->ctx
->base
.screen
);
675 uint32_t create_flags
= 0;
677 if (batch
->tiler_dummy
)
678 return batch
->tiler_dummy
;
680 if (!screen
->require_sfbd
)
681 create_flags
= PAN_BO_INVISIBLE
;
683 batch
->tiler_dummy
= panfrost_batch_create_bo(batch
, 4096,
685 PAN_BO_ACCESS_PRIVATE
|
687 PAN_BO_ACCESS_VERTEX_TILER
|
688 PAN_BO_ACCESS_FRAGMENT
);
689 assert(batch
->tiler_dummy
);
690 return batch
->tiler_dummy
;
694 panfrost_batch_draw_wallpaper(struct panfrost_batch
*batch
)
696 /* Color 0 is cleared, no need to draw the wallpaper.
697 * TODO: MRT wallpapers.
699 if (batch
->clear
& PIPE_CLEAR_COLOR0
)
702 /* Nothing to reload? TODO: MRT wallpapers */
703 if (batch
->key
.cbufs
[0] == NULL
)
706 /* No draw calls, and no clear on the depth/stencil bufs.
707 * Drawing the wallpaper would be useless.
709 if (!batch
->last_tiler
.gpu
&&
710 !(batch
->clear
& PIPE_CLEAR_DEPTHSTENCIL
))
713 /* Check if the buffer has any content on it worth preserving */
715 struct pipe_surface
*surf
= batch
->key
.cbufs
[0];
716 struct panfrost_resource
*rsrc
= pan_resource(surf
->texture
);
717 unsigned level
= surf
->u
.tex
.level
;
719 if (!rsrc
->slices
[level
].initialized
)
722 batch
->ctx
->wallpaper_batch
= batch
;
724 /* Clamp the rendering area to the damage extent. The
725 * KHR_partial_update() spec states that trying to render outside of
726 * the damage region is "undefined behavior", so we should be safe.
728 unsigned damage_width
= (rsrc
->damage
.extent
.maxx
- rsrc
->damage
.extent
.minx
);
729 unsigned damage_height
= (rsrc
->damage
.extent
.maxy
- rsrc
->damage
.extent
.miny
);
731 if (damage_width
&& damage_height
) {
732 panfrost_batch_intersection_scissor(batch
,
733 rsrc
->damage
.extent
.minx
,
734 rsrc
->damage
.extent
.miny
,
735 rsrc
->damage
.extent
.maxx
,
736 rsrc
->damage
.extent
.maxy
);
739 /* FIXME: Looks like aligning on a tile is not enough, but
740 * aligning on twice the tile size seems to works. We don't
741 * know exactly what happens here but this deserves extra
742 * investigation to figure it out.
744 batch
->minx
= batch
->minx
& ~((MALI_TILE_LENGTH
* 2) - 1);
745 batch
->miny
= batch
->miny
& ~((MALI_TILE_LENGTH
* 2) - 1);
746 batch
->maxx
= MIN2(ALIGN_POT(batch
->maxx
, MALI_TILE_LENGTH
* 2),
748 batch
->maxy
= MIN2(ALIGN_POT(batch
->maxy
, MALI_TILE_LENGTH
* 2),
751 struct pipe_scissor_state damage
;
752 struct pipe_box rects
[4];
754 /* Clamp the damage box to the rendering area. */
755 damage
.minx
= MAX2(batch
->minx
, rsrc
->damage
.biggest_rect
.x
);
756 damage
.miny
= MAX2(batch
->miny
, rsrc
->damage
.biggest_rect
.y
);
757 damage
.maxx
= MIN2(batch
->maxx
,
758 rsrc
->damage
.biggest_rect
.x
+
759 rsrc
->damage
.biggest_rect
.width
);
760 damage
.maxy
= MIN2(batch
->maxy
,
761 rsrc
->damage
.biggest_rect
.y
+
762 rsrc
->damage
.biggest_rect
.height
);
764 /* One damage rectangle means we can end up with at most 4 reload
766 * 1: left region, only exists if damage.x > 0
767 * 2: right region, only exists if damage.x + damage.width < fb->width
768 * 3: top region, only exists if damage.y > 0. The intersection with
769 * the left and right regions are dropped
770 * 4: bottom region, only exists if damage.y + damage.height < fb->height.
771 * The intersection with the left and right regions are dropped
773 * ____________________________
780 * |_______|___________|______|
782 u_box_2d(batch
->minx
, batch
->miny
, damage
.minx
- batch
->minx
,
783 batch
->maxy
- batch
->miny
, &rects
[0]);
784 u_box_2d(damage
.maxx
, batch
->miny
, batch
->maxx
- damage
.maxx
,
785 batch
->maxy
- batch
->miny
, &rects
[1]);
786 u_box_2d(damage
.minx
, batch
->miny
, damage
.maxx
- damage
.minx
,
787 damage
.miny
- batch
->miny
, &rects
[2]);
788 u_box_2d(damage
.minx
, damage
.maxy
, damage
.maxx
- damage
.minx
,
789 batch
->maxy
- damage
.maxy
, &rects
[3]);
791 for (unsigned i
= 0; i
< 4; i
++) {
792 /* Width and height are always >= 0 even if width is declared as a
793 * signed integer: u_box_2d() helper takes unsigned args and
794 * panfrost_set_damage_region() is taking care of clamping
797 if (!rects
[i
].width
|| !rects
[i
].height
)
800 /* Blit the wallpaper in */
801 panfrost_blit_wallpaper(batch
->ctx
, &rects
[i
]);
803 batch
->ctx
->wallpaper_batch
= NULL
;
807 panfrost_batch_submit_ioctl(struct panfrost_batch
*batch
,
808 mali_ptr first_job_desc
,
811 struct panfrost_context
*ctx
= batch
->ctx
;
812 struct pipe_context
*gallium
= (struct pipe_context
*) ctx
;
813 struct panfrost_screen
*screen
= pan_screen(gallium
->screen
);
814 struct drm_panfrost_submit submit
= {0,};
815 uint32_t *bo_handles
, *in_syncs
= NULL
;
816 bool is_fragment_shader
;
819 is_fragment_shader
= (reqs
& PANFROST_JD_REQ_FS
) && batch
->first_job
.gpu
;
820 if (is_fragment_shader
)
821 submit
.in_sync_count
= 1;
823 submit
.in_sync_count
= util_dynarray_num_elements(&batch
->dependencies
,
824 struct panfrost_batch_fence
*);
826 if (submit
.in_sync_count
) {
827 in_syncs
= calloc(submit
.in_sync_count
, sizeof(*in_syncs
));
831 /* The fragment job always depends on the vertex/tiler job if there's
834 if (is_fragment_shader
) {
835 in_syncs
[0] = batch
->out_sync
->syncobj
;
839 util_dynarray_foreach(&batch
->dependencies
,
840 struct panfrost_batch_fence
*, dep
)
841 in_syncs
[i
++] = (*dep
)->syncobj
;
844 submit
.in_syncs
= (uintptr_t)in_syncs
;
845 submit
.out_sync
= batch
->out_sync
->syncobj
;
846 submit
.jc
= first_job_desc
;
847 submit
.requirements
= reqs
;
849 bo_handles
= calloc(batch
->bos
->entries
, sizeof(*bo_handles
));
852 hash_table_foreach(batch
->bos
, entry
) {
853 struct panfrost_bo
*bo
= (struct panfrost_bo
*)entry
->key
;
854 uint32_t flags
= (uintptr_t)entry
->data
;
856 assert(bo
->gem_handle
> 0);
857 bo_handles
[submit
.bo_handle_count
++] = bo
->gem_handle
;
859 /* Update the BO access flags so that panfrost_bo_wait() knows
860 * about all pending accesses.
861 * We only keep the READ/WRITE info since this is all the BO
862 * wait logic cares about.
863 * We also preserve existing flags as this batch might not
864 * be the first one to access the BO.
866 bo
->gpu_access
|= flags
& (PAN_BO_ACCESS_RW
);
869 submit
.bo_handles
= (u64
) (uintptr_t) bo_handles
;
870 ret
= drmIoctl(screen
->fd
, DRM_IOCTL_PANFROST_SUBMIT
, &submit
);
875 fprintf(stderr
, "Error submitting: %m\n");
879 /* Trace the job if we're doing that */
880 if (pan_debug
& PAN_DBG_TRACE
) {
881 /* Wait so we can get errors reported back */
882 drmSyncobjWait(screen
->fd
, &batch
->out_sync
->syncobj
, 1,
884 pandecode_jc(submit
.jc
, FALSE
, screen
->gpu_id
);
891 panfrost_batch_submit_jobs(struct panfrost_batch
*batch
)
893 bool has_draws
= batch
->first_job
.gpu
;
897 ret
= panfrost_batch_submit_ioctl(batch
, batch
->first_job
.gpu
, 0);
901 if (batch
->first_tiler
.gpu
|| batch
->clear
) {
902 mali_ptr fragjob
= panfrost_fragment_job(batch
, has_draws
);
904 ret
= panfrost_batch_submit_ioctl(batch
, fragjob
, PANFROST_JD_REQ_FS
);
912 panfrost_batch_submit(struct panfrost_batch
*batch
)
916 /* Submit the dependencies first. */
917 util_dynarray_foreach(&batch
->dependencies
,
918 struct panfrost_batch_fence
*, dep
) {
920 panfrost_batch_submit((*dep
)->batch
);
926 if (!batch
->last_job
.gpu
&& !batch
->clear
) {
927 /* Mark the fence as signaled so the fence logic does not try
930 batch
->out_sync
->signaled
= true;
934 panfrost_batch_draw_wallpaper(batch
);
936 panfrost_scoreboard_link_batch(batch
);
938 ret
= panfrost_batch_submit_jobs(batch
);
941 fprintf(stderr
, "panfrost_batch_submit failed: %d\n", ret
);
944 panfrost_freeze_batch(batch
);
945 panfrost_free_batch(batch
);
949 panfrost_flush_all_batches(struct panfrost_context
*ctx
, bool wait
)
951 struct util_dynarray fences
, syncobjs
;
954 util_dynarray_init(&fences
, NULL
);
955 util_dynarray_init(&syncobjs
, NULL
);
958 hash_table_foreach(ctx
->batches
, hentry
) {
959 struct panfrost_batch
*batch
= hentry
->data
;
964 panfrost_batch_fence_reference(batch
->out_sync
);
965 util_dynarray_append(&fences
, struct panfrost_batch_fence
*,
967 util_dynarray_append(&syncobjs
, uint32_t,
968 batch
->out_sync
->syncobj
);
971 panfrost_batch_submit(batch
);
974 assert(!ctx
->batches
->entries
);
976 /* Collect batch fences before returning */
977 panfrost_gc_fences(ctx
);
982 drmSyncobjWait(pan_screen(ctx
->base
.screen
)->fd
,
983 util_dynarray_begin(&syncobjs
),
984 util_dynarray_num_elements(&syncobjs
, uint32_t),
985 INT64_MAX
, DRM_SYNCOBJ_WAIT_FLAGS_WAIT_ALL
, NULL
);
987 util_dynarray_foreach(&fences
, struct panfrost_batch_fence
*, fence
)
988 panfrost_batch_fence_unreference(*fence
);
990 util_dynarray_fini(&fences
);
991 util_dynarray_fini(&syncobjs
);
995 panfrost_pending_batches_access_bo(struct panfrost_context
*ctx
,
996 const struct panfrost_bo
*bo
)
998 struct panfrost_bo_access
*access
;
999 struct hash_entry
*hentry
;
1001 hentry
= _mesa_hash_table_search(ctx
->accessed_bos
, bo
);
1002 access
= hentry
? hentry
->data
: NULL
;
1006 if (access
->writer
&& access
->writer
->batch
)
1009 util_dynarray_foreach(&access
->readers
, struct panfrost_batch_fence
*,
1011 if (*reader
&& (*reader
)->batch
)
1019 panfrost_flush_batches_accessing_bo(struct panfrost_context
*ctx
,
1020 struct panfrost_bo
*bo
,
1021 uint32_t access_type
)
1023 struct panfrost_bo_access
*access
;
1024 struct hash_entry
*hentry
;
1026 /* It doesn't make any to flush only the readers. */
1027 assert(access_type
== PAN_BO_ACCESS_WRITE
||
1028 access_type
== PAN_BO_ACCESS_RW
);
1030 hentry
= _mesa_hash_table_search(ctx
->accessed_bos
, bo
);
1031 access
= hentry
? hentry
->data
: NULL
;
1035 if (access_type
& PAN_BO_ACCESS_WRITE
&& access
->writer
&&
1036 access
->writer
->batch
)
1037 panfrost_batch_submit(access
->writer
->batch
);
1039 if (!(access_type
& PAN_BO_ACCESS_READ
))
1042 util_dynarray_foreach(&access
->readers
, struct panfrost_batch_fence
*,
1044 if (*reader
&& (*reader
)->batch
)
1045 panfrost_batch_submit((*reader
)->batch
);
1050 panfrost_batch_set_requirements(struct panfrost_batch
*batch
)
1052 struct panfrost_context
*ctx
= batch
->ctx
;
1054 if (ctx
->rasterizer
&& ctx
->rasterizer
->base
.multisample
)
1055 batch
->requirements
|= PAN_REQ_MSAA
;
1057 if (ctx
->depth_stencil
&& ctx
->depth_stencil
->depth
.writemask
)
1058 batch
->requirements
|= PAN_REQ_DEPTH_WRITE
;
1061 /* Helper to smear a 32-bit color across 128-bit components */
1064 pan_pack_color_32(uint32_t *packed
, uint32_t v
)
1066 for (unsigned i
= 0; i
< 4; ++i
)
1071 pan_pack_color_64(uint32_t *packed
, uint32_t lo
, uint32_t hi
)
1073 for (unsigned i
= 0; i
< 4; i
+= 2) {
1080 pan_pack_color(uint32_t *packed
, const union pipe_color_union
*color
, enum pipe_format format
)
1082 /* Alpha magicked to 1.0 if there is no alpha */
1084 bool has_alpha
= util_format_has_alpha(format
);
1085 float clear_alpha
= has_alpha
? color
->f
[3] : 1.0f
;
1087 /* Packed color depends on the framebuffer format */
1089 const struct util_format_description
*desc
=
1090 util_format_description(format
);
1092 if (util_format_is_rgba8_variant(desc
)) {
1093 pan_pack_color_32(packed
,
1094 (float_to_ubyte(clear_alpha
) << 24) |
1095 (float_to_ubyte(color
->f
[2]) << 16) |
1096 (float_to_ubyte(color
->f
[1]) << 8) |
1097 (float_to_ubyte(color
->f
[0]) << 0));
1098 } else if (format
== PIPE_FORMAT_B5G6R5_UNORM
) {
1099 /* First, we convert the components to R5, G6, B5 separately */
1100 unsigned r5
= CLAMP(color
->f
[0], 0.0, 1.0) * 31.0;
1101 unsigned g6
= CLAMP(color
->f
[1], 0.0, 1.0) * 63.0;
1102 unsigned b5
= CLAMP(color
->f
[2], 0.0, 1.0) * 31.0;
1104 /* Then we pack into a sparse u32. TODO: Why these shifts? */
1105 pan_pack_color_32(packed
, (b5
<< 25) | (g6
<< 14) | (r5
<< 5));
1106 } else if (format
== PIPE_FORMAT_B4G4R4A4_UNORM
) {
1107 /* We scale the components against 0xF0 (=240.0), rather than 0xFF */
1108 unsigned r4
= CLAMP(color
->f
[0], 0.0, 1.0) * 240.0;
1109 unsigned g4
= CLAMP(color
->f
[1], 0.0, 1.0) * 240.0;
1110 unsigned b4
= CLAMP(color
->f
[2], 0.0, 1.0) * 240.0;
1111 unsigned a4
= CLAMP(clear_alpha
, 0.0, 1.0) * 240.0;
1113 /* Pack on *byte* intervals */
1114 pan_pack_color_32(packed
, (a4
<< 24) | (b4
<< 16) | (g4
<< 8) | r4
);
1115 } else if (format
== PIPE_FORMAT_B5G5R5A1_UNORM
) {
1116 /* Scale as expected but shift oddly */
1117 unsigned r5
= round(CLAMP(color
->f
[0], 0.0, 1.0)) * 31.0;
1118 unsigned g5
= round(CLAMP(color
->f
[1], 0.0, 1.0)) * 31.0;
1119 unsigned b5
= round(CLAMP(color
->f
[2], 0.0, 1.0)) * 31.0;
1120 unsigned a1
= round(CLAMP(clear_alpha
, 0.0, 1.0)) * 1.0;
1122 pan_pack_color_32(packed
, (a1
<< 31) | (b5
<< 25) | (g5
<< 15) | (r5
<< 5));
1124 /* Try Gallium's generic default path. Doesn't work for all
1125 * formats but it's a good guess. */
1127 union util_color out
;
1129 if (util_format_is_pure_integer(format
)) {
1130 memcpy(out
.ui
, color
->ui
, 16);
1132 util_pack_color(color
->f
, format
, &out
);
1135 unsigned size
= util_format_get_blocksize(format
);
1138 unsigned b
= out
.ui
[0];
1139 unsigned s
= b
| (b
<< 8);
1140 pan_pack_color_32(packed
, s
| (s
<< 16));
1141 } else if (size
== 2)
1142 pan_pack_color_32(packed
, out
.ui
[0] | (out
.ui
[0] << 16));
1143 else if (size
== 3 || size
== 4)
1144 pan_pack_color_32(packed
, out
.ui
[0]);
1146 pan_pack_color_64(packed
, out
.ui
[0], out
.ui
[1]);
1147 else if (size
== 16)
1148 memcpy(packed
, out
.ui
, 16);
1150 unreachable("Unknown generic format size packing clear colour");
1155 panfrost_batch_clear(struct panfrost_batch
*batch
,
1157 const union pipe_color_union
*color
,
1158 double depth
, unsigned stencil
)
1160 struct panfrost_context
*ctx
= batch
->ctx
;
1162 if (buffers
& PIPE_CLEAR_COLOR
) {
1163 for (unsigned i
= 0; i
< PIPE_MAX_COLOR_BUFS
; ++i
) {
1164 if (!(buffers
& (PIPE_CLEAR_COLOR0
<< i
)))
1167 enum pipe_format format
= ctx
->pipe_framebuffer
.cbufs
[i
]->format
;
1168 pan_pack_color(batch
->clear_color
[i
], color
, format
);
1172 if (buffers
& PIPE_CLEAR_DEPTH
) {
1173 batch
->clear_depth
= depth
;
1176 if (buffers
& PIPE_CLEAR_STENCIL
) {
1177 batch
->clear_stencil
= stencil
;
1180 batch
->clear
|= buffers
;
1182 /* Clearing affects the entire framebuffer (by definition -- this is
1183 * the Gallium clear callback, which clears the whole framebuffer. If
1184 * the scissor test were enabled from the GL side, the state tracker
1185 * would emit a quad instead and we wouldn't go down this code path) */
1187 panfrost_batch_union_scissor(batch
, 0, 0,
1188 ctx
->pipe_framebuffer
.width
,
1189 ctx
->pipe_framebuffer
.height
);
1193 panfrost_batch_compare(const void *a
, const void *b
)
1195 return util_framebuffer_state_equal(a
, b
);
1199 panfrost_batch_hash(const void *key
)
1201 return _mesa_hash_data(key
, sizeof(struct pipe_framebuffer_state
));
1204 /* Given a new bounding rectangle (scissor), let the job cover the union of the
1205 * new and old bounding rectangles */
1208 panfrost_batch_union_scissor(struct panfrost_batch
*batch
,
1209 unsigned minx
, unsigned miny
,
1210 unsigned maxx
, unsigned maxy
)
1212 batch
->minx
= MIN2(batch
->minx
, minx
);
1213 batch
->miny
= MIN2(batch
->miny
, miny
);
1214 batch
->maxx
= MAX2(batch
->maxx
, maxx
);
1215 batch
->maxy
= MAX2(batch
->maxy
, maxy
);
1219 panfrost_batch_intersection_scissor(struct panfrost_batch
*batch
,
1220 unsigned minx
, unsigned miny
,
1221 unsigned maxx
, unsigned maxy
)
1223 batch
->minx
= MAX2(batch
->minx
, minx
);
1224 batch
->miny
= MAX2(batch
->miny
, miny
);
1225 batch
->maxx
= MIN2(batch
->maxx
, maxx
);
1226 batch
->maxy
= MIN2(batch
->maxy
, maxy
);
1229 /* Are we currently rendering to the screen (rather than an FBO)? */
1232 panfrost_batch_is_scanout(struct panfrost_batch
*batch
)
1234 /* If there is no color buffer, it's an FBO */
1235 if (batch
->key
.nr_cbufs
!= 1)
1238 /* If we're too early that no framebuffer was sent, it's scanout */
1239 if (!batch
->key
.cbufs
[0])
1242 return batch
->key
.cbufs
[0]->texture
->bind
& PIPE_BIND_DISPLAY_TARGET
||
1243 batch
->key
.cbufs
[0]->texture
->bind
& PIPE_BIND_SCANOUT
||
1244 batch
->key
.cbufs
[0]->texture
->bind
& PIPE_BIND_SHARED
;
1248 panfrost_batch_init(struct panfrost_context
*ctx
)
1250 ctx
->batches
= _mesa_hash_table_create(ctx
,
1251 panfrost_batch_hash
,
1252 panfrost_batch_compare
);
1253 ctx
->accessed_bos
= _mesa_hash_table_create(ctx
, _mesa_hash_pointer
,
1254 _mesa_key_pointer_equal
);