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/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.
297 panfrost_batch_fence_is_signaled(struct panfrost_batch_fence
*fence
)
302 /* Batch has not been submitted yet. */
306 int ret
= drmSyncobjWait(pan_screen(fence
->ctx
->base
.screen
)->fd
,
307 &fence
->syncobj
, 1, 0, 0, NULL
);
309 /* Cache whether the fence was signaled */
310 fence
->signaled
= ret
>= 0;
311 return fence
->signaled
;
315 panfrost_bo_access_gc_fences(struct panfrost_context
*ctx
,
316 struct panfrost_bo_access
*access
,
317 const struct panfrost_bo
*bo
)
319 if (access
->writer
&& panfrost_batch_fence_is_signaled(access
->writer
)) {
320 panfrost_batch_fence_unreference(access
->writer
);
321 access
->writer
= NULL
;
324 unsigned nreaders
= 0;
325 util_dynarray_foreach(&access
->readers
, struct panfrost_batch_fence
*,
330 if (panfrost_batch_fence_is_signaled(*reader
)) {
331 panfrost_batch_fence_unreference(*reader
);
339 util_dynarray_clear(&access
->readers
);
342 /* Collect signaled fences to keep the kernel-side syncobj-map small. The
343 * idea is to collect those signaled fences at the end of each flush_all
344 * call. This function is likely to collect only fences from previous
345 * batch flushes not the one that have just have just been submitted and
346 * are probably still in flight when we trigger the garbage collection.
347 * Anyway, we need to do this garbage collection at some point if we don't
348 * want the BO access map to keep invalid entries around and retain
352 panfrost_gc_fences(struct panfrost_context
*ctx
)
354 hash_table_foreach(ctx
->accessed_bos
, entry
) {
355 struct panfrost_bo_access
*access
= entry
->data
;
358 panfrost_bo_access_gc_fences(ctx
, access
, entry
->key
);
359 if (!util_dynarray_num_elements(&access
->readers
,
360 struct panfrost_batch_fence
*) &&
362 _mesa_hash_table_remove(ctx
->accessed_bos
, entry
);
368 panfrost_batch_in_readers(struct panfrost_batch
*batch
,
369 struct panfrost_bo_access
*access
)
371 util_dynarray_foreach(&access
->readers
, struct panfrost_batch_fence
*,
373 if (*reader
&& (*reader
)->batch
== batch
)
382 panfrost_batch_update_bo_access(struct panfrost_batch
*batch
,
383 struct panfrost_bo
*bo
, uint32_t access_type
,
384 bool already_accessed
)
386 struct panfrost_context
*ctx
= batch
->ctx
;
387 struct panfrost_bo_access
*access
;
388 uint32_t old_access_type
;
389 struct hash_entry
*entry
;
391 assert(access_type
== PAN_BO_ACCESS_WRITE
||
392 access_type
== PAN_BO_ACCESS_READ
);
394 entry
= _mesa_hash_table_search(ctx
->accessed_bos
, bo
);
395 access
= entry
? entry
->data
: NULL
;
397 old_access_type
= access
->type
;
399 access
= rzalloc(ctx
, struct panfrost_bo_access
);
400 util_dynarray_init(&access
->readers
, access
);
401 _mesa_hash_table_insert(ctx
->accessed_bos
, bo
, access
);
402 /* We are the first to access this BO, let's initialize
403 * old_access_type to our own access type in that case.
405 old_access_type
= access_type
;
406 access
->type
= access_type
;
411 if (access_type
== PAN_BO_ACCESS_WRITE
&&
412 old_access_type
== PAN_BO_ACCESS_READ
) {
413 /* Previous access was a read and we want to write this BO.
414 * We first need to add explicit deps between our batch and
415 * the previous readers.
417 util_dynarray_foreach(&access
->readers
,
418 struct panfrost_batch_fence
*, reader
) {
419 /* We were already reading the BO, no need to add a dep
420 * on ourself (the acyclic check would complain about
423 if (!(*reader
) || (*reader
)->batch
== batch
)
426 panfrost_batch_add_dep(batch
, *reader
);
428 panfrost_batch_fence_reference(batch
->out_sync
);
430 /* We now are the new writer. */
431 access
->writer
= batch
->out_sync
;
432 access
->type
= access_type
;
434 /* Release the previous readers and reset the readers array. */
435 util_dynarray_foreach(&access
->readers
,
436 struct panfrost_batch_fence
*,
440 panfrost_batch_fence_unreference(*reader
);
443 util_dynarray_clear(&access
->readers
);
444 } else if (access_type
== PAN_BO_ACCESS_WRITE
&&
445 old_access_type
== PAN_BO_ACCESS_WRITE
) {
446 /* Previous access was a write and we want to write this BO.
447 * First check if we were the previous writer, in that case
448 * there's nothing to do. Otherwise we need to add a
449 * dependency between the new writer and the old one.
451 if (access
->writer
!= batch
->out_sync
) {
452 if (access
->writer
) {
453 panfrost_batch_add_dep(batch
, access
->writer
);
454 panfrost_batch_fence_unreference(access
->writer
);
456 panfrost_batch_fence_reference(batch
->out_sync
);
457 access
->writer
= batch
->out_sync
;
459 } else if (access_type
== PAN_BO_ACCESS_READ
&&
460 old_access_type
== PAN_BO_ACCESS_WRITE
) {
461 /* Previous access was a write and we want to read this BO.
462 * First check if we were the previous writer, in that case
463 * we want to keep the access type unchanged, as a write is
464 * more constraining than a read.
466 if (access
->writer
!= batch
->out_sync
) {
467 /* Add a dependency on the previous writer. */
468 panfrost_batch_add_dep(batch
, access
->writer
);
470 /* The previous access was a write, there's no reason
471 * to have entries in the readers array.
473 assert(!util_dynarray_num_elements(&access
->readers
,
474 struct panfrost_batch_fence
*));
476 /* Add ourselves to the readers array. */
477 panfrost_batch_fence_reference(batch
->out_sync
);
478 util_dynarray_append(&access
->readers
,
479 struct panfrost_batch_fence
*,
481 access
->type
= PAN_BO_ACCESS_READ
;
484 /* We already accessed this BO before, so we should already be
485 * in the reader array.
487 if (already_accessed
) {
488 assert(panfrost_batch_in_readers(batch
, access
));
492 /* Previous access was a read and we want to read this BO.
493 * Add ourselves to the readers array and add a dependency on
494 * the previous writer if any.
496 panfrost_batch_fence_reference(batch
->out_sync
);
497 util_dynarray_append(&access
->readers
,
498 struct panfrost_batch_fence
*,
502 panfrost_batch_add_dep(batch
, access
->writer
);
507 panfrost_batch_add_bo(struct panfrost_batch
*batch
, struct panfrost_bo
*bo
,
513 struct hash_entry
*entry
;
514 uint32_t old_flags
= 0;
516 entry
= _mesa_hash_table_search(batch
->bos
, bo
);
518 entry
= _mesa_hash_table_insert(batch
->bos
, bo
,
519 (void *)(uintptr_t)flags
);
520 panfrost_bo_reference(bo
);
522 old_flags
= (uintptr_t)entry
->data
;
524 /* All batches have to agree on the shared flag. */
525 assert((old_flags
& PAN_BO_ACCESS_SHARED
) ==
526 (flags
& PAN_BO_ACCESS_SHARED
));
531 if (old_flags
== flags
)
535 entry
->data
= (void *)(uintptr_t)flags
;
537 /* If this is not a shared BO, we don't really care about dependency
540 if (!(flags
& PAN_BO_ACCESS_SHARED
))
543 /* All dependencies should have been flushed before we execute the
544 * wallpaper draw, so it should be harmless to skip the
545 * update_bo_access() call.
547 if (batch
== batch
->ctx
->wallpaper_batch
)
550 /* Only pass R/W flags to the dep tracking logic. */
551 assert(flags
& PAN_BO_ACCESS_RW
);
552 flags
= (flags
& PAN_BO_ACCESS_WRITE
) ?
553 PAN_BO_ACCESS_WRITE
: PAN_BO_ACCESS_READ
;
554 panfrost_batch_update_bo_access(batch
, bo
, flags
, old_flags
!= 0);
557 void panfrost_batch_add_fbo_bos(struct panfrost_batch
*batch
)
559 uint32_t flags
= PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_WRITE
|
560 PAN_BO_ACCESS_VERTEX_TILER
|
561 PAN_BO_ACCESS_FRAGMENT
;
563 for (unsigned i
= 0; i
< batch
->key
.nr_cbufs
; ++i
) {
564 struct panfrost_resource
*rsrc
= pan_resource(batch
->key
.cbufs
[i
]->texture
);
565 panfrost_batch_add_bo(batch
, rsrc
->bo
, flags
);
568 if (batch
->key
.zsbuf
) {
569 struct panfrost_resource
*rsrc
= pan_resource(batch
->key
.zsbuf
->texture
);
570 panfrost_batch_add_bo(batch
, rsrc
->bo
, flags
);
575 panfrost_batch_create_bo(struct panfrost_batch
*batch
, size_t size
,
576 uint32_t create_flags
, uint32_t access_flags
)
578 struct panfrost_bo
*bo
;
580 bo
= panfrost_bo_create(pan_screen(batch
->ctx
->base
.screen
), size
,
582 panfrost_batch_add_bo(batch
, bo
, access_flags
);
584 /* panfrost_batch_add_bo() has retained a reference and
585 * panfrost_bo_create() initialize the refcnt to 1, so let's
586 * unreference the BO here so it gets released when the batch is
587 * destroyed (unless it's retained by someone else in the meantime).
589 panfrost_bo_unreference(bo
);
593 /* Returns the polygon list's GPU address if available, or otherwise allocates
594 * the polygon list. It's perfectly fast to use allocate/free BO directly,
595 * since we'll hit the BO cache and this is one-per-batch anyway. */
598 panfrost_batch_get_polygon_list(struct panfrost_batch
*batch
, unsigned size
)
600 if (batch
->polygon_list
) {
601 assert(batch
->polygon_list
->size
>= size
);
603 /* Create the BO as invisible, as there's no reason to map */
605 batch
->polygon_list
= panfrost_batch_create_bo(batch
, size
,
607 PAN_BO_ACCESS_PRIVATE
|
609 PAN_BO_ACCESS_VERTEX_TILER
|
610 PAN_BO_ACCESS_FRAGMENT
);
613 return batch
->polygon_list
->gpu
;
617 panfrost_batch_get_scratchpad(struct panfrost_batch
*batch
)
619 if (batch
->scratchpad
)
620 return batch
->scratchpad
;
622 batch
->scratchpad
= panfrost_batch_create_bo(batch
, 64 * 4 * 4096,
624 PAN_BO_ACCESS_PRIVATE
|
626 PAN_BO_ACCESS_VERTEX_TILER
|
627 PAN_BO_ACCESS_FRAGMENT
);
628 assert(batch
->scratchpad
);
629 return batch
->scratchpad
;
633 panfrost_batch_get_tiler_heap(struct panfrost_batch
*batch
)
635 if (batch
->tiler_heap
)
636 return batch
->tiler_heap
;
638 batch
->tiler_heap
= panfrost_batch_create_bo(batch
, 4096 * 4096,
641 PAN_BO_ACCESS_PRIVATE
|
643 PAN_BO_ACCESS_VERTEX_TILER
|
644 PAN_BO_ACCESS_FRAGMENT
);
645 assert(batch
->tiler_heap
);
646 return batch
->tiler_heap
;
650 panfrost_batch_get_tiler_dummy(struct panfrost_batch
*batch
)
652 if (batch
->tiler_dummy
)
653 return batch
->tiler_dummy
;
655 batch
->tiler_dummy
= panfrost_batch_create_bo(batch
, 4096,
657 PAN_BO_ACCESS_PRIVATE
|
659 PAN_BO_ACCESS_VERTEX_TILER
|
660 PAN_BO_ACCESS_FRAGMENT
);
661 assert(batch
->tiler_dummy
);
662 return batch
->tiler_dummy
;
666 panfrost_batch_draw_wallpaper(struct panfrost_batch
*batch
)
668 /* Nothing to reload? TODO: MRT wallpapers */
669 if (batch
->key
.cbufs
[0] == NULL
)
672 /* Check if the buffer has any content on it worth preserving */
674 struct pipe_surface
*surf
= batch
->key
.cbufs
[0];
675 struct panfrost_resource
*rsrc
= pan_resource(surf
->texture
);
676 unsigned level
= surf
->u
.tex
.level
;
678 if (!rsrc
->slices
[level
].initialized
)
681 batch
->ctx
->wallpaper_batch
= batch
;
683 /* Clamp the rendering area to the damage extent. The
684 * KHR_partial_update() spec states that trying to render outside of
685 * the damage region is "undefined behavior", so we should be safe.
687 unsigned damage_width
= (rsrc
->damage
.extent
.maxx
- rsrc
->damage
.extent
.minx
);
688 unsigned damage_height
= (rsrc
->damage
.extent
.maxy
- rsrc
->damage
.extent
.miny
);
690 if (damage_width
&& damage_height
) {
691 panfrost_batch_intersection_scissor(batch
,
692 rsrc
->damage
.extent
.minx
,
693 rsrc
->damage
.extent
.miny
,
694 rsrc
->damage
.extent
.maxx
,
695 rsrc
->damage
.extent
.maxy
);
698 /* FIXME: Looks like aligning on a tile is not enough, but
699 * aligning on twice the tile size seems to works. We don't
700 * know exactly what happens here but this deserves extra
701 * investigation to figure it out.
703 batch
->minx
= batch
->minx
& ~((MALI_TILE_LENGTH
* 2) - 1);
704 batch
->miny
= batch
->miny
& ~((MALI_TILE_LENGTH
* 2) - 1);
705 batch
->maxx
= MIN2(ALIGN_POT(batch
->maxx
, MALI_TILE_LENGTH
* 2),
707 batch
->maxy
= MIN2(ALIGN_POT(batch
->maxy
, MALI_TILE_LENGTH
* 2),
710 struct pipe_scissor_state damage
;
711 struct pipe_box rects
[4];
713 /* Clamp the damage box to the rendering area. */
714 damage
.minx
= MAX2(batch
->minx
, rsrc
->damage
.biggest_rect
.x
);
715 damage
.miny
= MAX2(batch
->miny
, rsrc
->damage
.biggest_rect
.y
);
716 damage
.maxx
= MIN2(batch
->maxx
,
717 rsrc
->damage
.biggest_rect
.x
+
718 rsrc
->damage
.biggest_rect
.width
);
719 damage
.maxy
= MIN2(batch
->maxy
,
720 rsrc
->damage
.biggest_rect
.y
+
721 rsrc
->damage
.biggest_rect
.height
);
723 /* One damage rectangle means we can end up with at most 4 reload
725 * 1: left region, only exists if damage.x > 0
726 * 2: right region, only exists if damage.x + damage.width < fb->width
727 * 3: top region, only exists if damage.y > 0. The intersection with
728 * the left and right regions are dropped
729 * 4: bottom region, only exists if damage.y + damage.height < fb->height.
730 * The intersection with the left and right regions are dropped
732 * ____________________________
739 * |_______|___________|______|
741 u_box_2d(batch
->minx
, batch
->miny
, damage
.minx
- batch
->minx
,
742 batch
->maxy
- batch
->miny
, &rects
[0]);
743 u_box_2d(damage
.maxx
, batch
->miny
, batch
->maxx
- damage
.maxx
,
744 batch
->maxy
- batch
->miny
, &rects
[1]);
745 u_box_2d(damage
.minx
, batch
->miny
, damage
.maxx
- damage
.minx
,
746 damage
.miny
- batch
->miny
, &rects
[2]);
747 u_box_2d(damage
.minx
, damage
.maxy
, damage
.maxx
- damage
.minx
,
748 batch
->maxy
- damage
.maxy
, &rects
[3]);
750 for (unsigned i
= 0; i
< 4; i
++) {
751 /* Width and height are always >= 0 even if width is declared as a
752 * signed integer: u_box_2d() helper takes unsigned args and
753 * panfrost_set_damage_region() is taking care of clamping
756 if (!rects
[i
].width
|| !rects
[i
].height
)
759 /* Blit the wallpaper in */
760 panfrost_blit_wallpaper(batch
->ctx
, &rects
[i
]);
762 batch
->ctx
->wallpaper_batch
= NULL
;
766 panfrost_batch_submit_ioctl(struct panfrost_batch
*batch
,
767 mali_ptr first_job_desc
,
770 struct panfrost_context
*ctx
= batch
->ctx
;
771 struct pipe_context
*gallium
= (struct pipe_context
*) ctx
;
772 struct panfrost_screen
*screen
= pan_screen(gallium
->screen
);
773 struct drm_panfrost_submit submit
= {0,};
774 uint32_t *bo_handles
, *in_syncs
= NULL
;
775 bool is_fragment_shader
;
778 is_fragment_shader
= (reqs
& PANFROST_JD_REQ_FS
) && batch
->first_job
.gpu
;
779 if (is_fragment_shader
)
780 submit
.in_sync_count
= 1;
782 submit
.in_sync_count
= util_dynarray_num_elements(&batch
->dependencies
,
783 struct panfrost_batch_fence
*);
785 if (submit
.in_sync_count
) {
786 in_syncs
= calloc(submit
.in_sync_count
, sizeof(*in_syncs
));
790 /* The fragment job always depends on the vertex/tiler job if there's
793 if (is_fragment_shader
) {
794 in_syncs
[0] = batch
->out_sync
->syncobj
;
798 util_dynarray_foreach(&batch
->dependencies
,
799 struct panfrost_batch_fence
*, dep
)
800 in_syncs
[i
++] = (*dep
)->syncobj
;
803 submit
.in_syncs
= (uintptr_t)in_syncs
;
804 submit
.out_sync
= batch
->out_sync
->syncobj
;
805 submit
.jc
= first_job_desc
;
806 submit
.requirements
= reqs
;
808 bo_handles
= calloc(batch
->bos
->entries
, sizeof(*bo_handles
));
811 hash_table_foreach(batch
->bos
, entry
) {
812 struct panfrost_bo
*bo
= (struct panfrost_bo
*)entry
->key
;
813 assert(bo
->gem_handle
> 0);
814 bo_handles
[submit
.bo_handle_count
++] = bo
->gem_handle
;
817 submit
.bo_handles
= (u64
) (uintptr_t) bo_handles
;
818 ret
= drmIoctl(screen
->fd
, DRM_IOCTL_PANFROST_SUBMIT
, &submit
);
823 fprintf(stderr
, "Error submitting: %m\n");
827 /* Trace the job if we're doing that */
828 if (pan_debug
& PAN_DBG_TRACE
) {
829 /* Wait so we can get errors reported back */
830 drmSyncobjWait(screen
->fd
, &batch
->out_sync
->syncobj
, 1,
832 pandecode_jc(submit
.jc
, FALSE
);
839 panfrost_batch_submit_jobs(struct panfrost_batch
*batch
)
841 bool has_draws
= batch
->first_job
.gpu
;
845 ret
= panfrost_batch_submit_ioctl(batch
, batch
->first_job
.gpu
, 0);
849 if (batch
->first_tiler
.gpu
|| batch
->clear
) {
850 mali_ptr fragjob
= panfrost_fragment_job(batch
, has_draws
);
852 ret
= panfrost_batch_submit_ioctl(batch
, fragjob
, PANFROST_JD_REQ_FS
);
860 panfrost_batch_submit(struct panfrost_batch
*batch
)
864 /* Submit the dependencies first. */
865 util_dynarray_foreach(&batch
->dependencies
,
866 struct panfrost_batch_fence
*, dep
) {
868 panfrost_batch_submit((*dep
)->batch
);
871 struct panfrost_context
*ctx
= batch
->ctx
;
875 if (!batch
->last_job
.gpu
&& !batch
->clear
) {
876 /* Mark the fence as signaled so the fence logic does not try
879 batch
->out_sync
->signaled
= true;
883 if (!batch
->clear
&& batch
->last_tiler
.gpu
)
884 panfrost_batch_draw_wallpaper(batch
);
886 panfrost_scoreboard_link_batch(batch
);
888 ret
= panfrost_batch_submit_jobs(batch
);
891 fprintf(stderr
, "panfrost_batch_submit failed: %d\n", ret
);
894 panfrost_freeze_batch(batch
);
896 /* We always stall the pipeline for correct results since pipelined
897 * rendering is quite broken right now (to be fixed by the panfrost_job
898 * refactor, just take the perf hit for correctness)
900 if (!batch
->out_sync
->signaled
)
901 drmSyncobjWait(pan_screen(ctx
->base
.screen
)->fd
,
902 &batch
->out_sync
->syncobj
, 1, INT64_MAX
, 0,
905 panfrost_free_batch(batch
);
910 panfrost_flush_all_batches(struct panfrost_context
*ctx
, bool wait
)
912 struct util_dynarray fences
, syncobjs
;
915 util_dynarray_init(&fences
, NULL
);
916 util_dynarray_init(&syncobjs
, NULL
);
919 hash_table_foreach(ctx
->batches
, hentry
) {
920 struct panfrost_batch
*batch
= hentry
->data
;
925 panfrost_batch_fence_reference(batch
->out_sync
);
926 util_dynarray_append(&fences
, struct panfrost_batch_fence
*,
928 util_dynarray_append(&syncobjs
, uint32_t,
929 batch
->out_sync
->syncobj
);
932 panfrost_batch_submit(batch
);
935 assert(!ctx
->batches
->entries
);
937 /* Collect batch fences before returning */
938 panfrost_gc_fences(ctx
);
943 drmSyncobjWait(pan_screen(ctx
->base
.screen
)->fd
,
944 util_dynarray_begin(&syncobjs
),
945 util_dynarray_num_elements(&syncobjs
, uint32_t),
946 INT64_MAX
, DRM_SYNCOBJ_WAIT_FLAGS_WAIT_ALL
, NULL
);
948 util_dynarray_foreach(&fences
, struct panfrost_batch_fence
*, fence
)
949 panfrost_batch_fence_unreference(*fence
);
951 util_dynarray_fini(&fences
);
952 util_dynarray_fini(&syncobjs
);
956 panfrost_flush_batches_accessing_bo(struct panfrost_context
*ctx
,
957 struct panfrost_bo
*bo
,
958 uint32_t access_type
)
960 struct panfrost_bo_access
*access
;
961 struct hash_entry
*hentry
;
963 /* It doesn't make any to flush only the readers. */
964 assert(access_type
== PAN_BO_ACCESS_WRITE
||
965 access_type
== PAN_BO_ACCESS_RW
);
967 hentry
= _mesa_hash_table_search(ctx
->accessed_bos
, bo
);
968 access
= hentry
? hentry
->data
: NULL
;
972 if (access_type
& PAN_BO_ACCESS_WRITE
&& access
->writer
&&
973 access
->writer
->batch
)
974 panfrost_batch_submit(access
->writer
->batch
);
976 if (!(access_type
& PAN_BO_ACCESS_READ
))
979 util_dynarray_foreach(&access
->readers
, struct panfrost_batch_fence
*,
981 if (*reader
&& (*reader
)->batch
)
982 panfrost_batch_submit((*reader
)->batch
);
987 panfrost_batch_set_requirements(struct panfrost_batch
*batch
)
989 struct panfrost_context
*ctx
= batch
->ctx
;
991 if (ctx
->rasterizer
&& ctx
->rasterizer
->base
.multisample
)
992 batch
->requirements
|= PAN_REQ_MSAA
;
994 if (ctx
->depth_stencil
&& ctx
->depth_stencil
->depth
.writemask
)
995 batch
->requirements
|= PAN_REQ_DEPTH_WRITE
;
998 /* Helper to smear a 32-bit color across 128-bit components */
1001 pan_pack_color_32(uint32_t *packed
, uint32_t v
)
1003 for (unsigned i
= 0; i
< 4; ++i
)
1008 pan_pack_color_64(uint32_t *packed
, uint32_t lo
, uint32_t hi
)
1010 for (unsigned i
= 0; i
< 4; i
+= 2) {
1017 pan_pack_color(uint32_t *packed
, const union pipe_color_union
*color
, enum pipe_format format
)
1019 /* Alpha magicked to 1.0 if there is no alpha */
1021 bool has_alpha
= util_format_has_alpha(format
);
1022 float clear_alpha
= has_alpha
? color
->f
[3] : 1.0f
;
1024 /* Packed color depends on the framebuffer format */
1026 const struct util_format_description
*desc
=
1027 util_format_description(format
);
1029 if (util_format_is_rgba8_variant(desc
)) {
1030 pan_pack_color_32(packed
,
1031 (float_to_ubyte(clear_alpha
) << 24) |
1032 (float_to_ubyte(color
->f
[2]) << 16) |
1033 (float_to_ubyte(color
->f
[1]) << 8) |
1034 (float_to_ubyte(color
->f
[0]) << 0));
1035 } else if (format
== PIPE_FORMAT_B5G6R5_UNORM
) {
1036 /* First, we convert the components to R5, G6, B5 separately */
1037 unsigned r5
= CLAMP(color
->f
[0], 0.0, 1.0) * 31.0;
1038 unsigned g6
= CLAMP(color
->f
[1], 0.0, 1.0) * 63.0;
1039 unsigned b5
= CLAMP(color
->f
[2], 0.0, 1.0) * 31.0;
1041 /* Then we pack into a sparse u32. TODO: Why these shifts? */
1042 pan_pack_color_32(packed
, (b5
<< 25) | (g6
<< 14) | (r5
<< 5));
1043 } else if (format
== PIPE_FORMAT_B4G4R4A4_UNORM
) {
1044 /* We scale the components against 0xF0 (=240.0), rather than 0xFF */
1045 unsigned r4
= CLAMP(color
->f
[0], 0.0, 1.0) * 240.0;
1046 unsigned g4
= CLAMP(color
->f
[1], 0.0, 1.0) * 240.0;
1047 unsigned b4
= CLAMP(color
->f
[2], 0.0, 1.0) * 240.0;
1048 unsigned a4
= CLAMP(clear_alpha
, 0.0, 1.0) * 240.0;
1050 /* Pack on *byte* intervals */
1051 pan_pack_color_32(packed
, (a4
<< 24) | (b4
<< 16) | (g4
<< 8) | r4
);
1052 } else if (format
== PIPE_FORMAT_B5G5R5A1_UNORM
) {
1053 /* Scale as expected but shift oddly */
1054 unsigned r5
= round(CLAMP(color
->f
[0], 0.0, 1.0)) * 31.0;
1055 unsigned g5
= round(CLAMP(color
->f
[1], 0.0, 1.0)) * 31.0;
1056 unsigned b5
= round(CLAMP(color
->f
[2], 0.0, 1.0)) * 31.0;
1057 unsigned a1
= round(CLAMP(clear_alpha
, 0.0, 1.0)) * 1.0;
1059 pan_pack_color_32(packed
, (a1
<< 31) | (b5
<< 25) | (g5
<< 15) | (r5
<< 5));
1061 /* Try Gallium's generic default path. Doesn't work for all
1062 * formats but it's a good guess. */
1064 union util_color out
;
1066 if (util_format_is_pure_integer(format
)) {
1067 memcpy(out
.ui
, color
->ui
, 16);
1069 util_pack_color(color
->f
, format
, &out
);
1072 unsigned size
= util_format_get_blocksize(format
);
1075 unsigned b
= out
.ui
[0];
1076 unsigned s
= b
| (b
<< 8);
1077 pan_pack_color_32(packed
, s
| (s
<< 16));
1078 } else if (size
== 2)
1079 pan_pack_color_32(packed
, out
.ui
[0] | (out
.ui
[0] << 16));
1081 pan_pack_color_32(packed
, out
.ui
[0]);
1083 pan_pack_color_64(packed
, out
.ui
[0], out
.ui
[1]);
1084 else if (size
== 16)
1085 memcpy(packed
, out
.ui
, 16);
1087 unreachable("Unknown generic format size packing clear colour");
1092 panfrost_batch_clear(struct panfrost_batch
*batch
,
1094 const union pipe_color_union
*color
,
1095 double depth
, unsigned stencil
)
1097 struct panfrost_context
*ctx
= batch
->ctx
;
1099 if (buffers
& PIPE_CLEAR_COLOR
) {
1100 for (unsigned i
= 0; i
< PIPE_MAX_COLOR_BUFS
; ++i
) {
1101 if (!(buffers
& (PIPE_CLEAR_COLOR0
<< i
)))
1104 enum pipe_format format
= ctx
->pipe_framebuffer
.cbufs
[i
]->format
;
1105 pan_pack_color(batch
->clear_color
[i
], color
, format
);
1109 if (buffers
& PIPE_CLEAR_DEPTH
) {
1110 batch
->clear_depth
= depth
;
1113 if (buffers
& PIPE_CLEAR_STENCIL
) {
1114 batch
->clear_stencil
= stencil
;
1117 batch
->clear
|= buffers
;
1119 /* Clearing affects the entire framebuffer (by definition -- this is
1120 * the Gallium clear callback, which clears the whole framebuffer. If
1121 * the scissor test were enabled from the GL side, the state tracker
1122 * would emit a quad instead and we wouldn't go down this code path) */
1124 panfrost_batch_union_scissor(batch
, 0, 0,
1125 ctx
->pipe_framebuffer
.width
,
1126 ctx
->pipe_framebuffer
.height
);
1130 panfrost_batch_compare(const void *a
, const void *b
)
1132 return util_framebuffer_state_equal(a
, b
);
1136 panfrost_batch_hash(const void *key
)
1138 return _mesa_hash_data(key
, sizeof(struct pipe_framebuffer_state
));
1141 /* Given a new bounding rectangle (scissor), let the job cover the union of the
1142 * new and old bounding rectangles */
1145 panfrost_batch_union_scissor(struct panfrost_batch
*batch
,
1146 unsigned minx
, unsigned miny
,
1147 unsigned maxx
, unsigned maxy
)
1149 batch
->minx
= MIN2(batch
->minx
, minx
);
1150 batch
->miny
= MIN2(batch
->miny
, miny
);
1151 batch
->maxx
= MAX2(batch
->maxx
, maxx
);
1152 batch
->maxy
= MAX2(batch
->maxy
, maxy
);
1156 panfrost_batch_intersection_scissor(struct panfrost_batch
*batch
,
1157 unsigned minx
, unsigned miny
,
1158 unsigned maxx
, unsigned maxy
)
1160 batch
->minx
= MAX2(batch
->minx
, minx
);
1161 batch
->miny
= MAX2(batch
->miny
, miny
);
1162 batch
->maxx
= MIN2(batch
->maxx
, maxx
);
1163 batch
->maxy
= MIN2(batch
->maxy
, maxy
);
1166 /* Are we currently rendering to the screen (rather than an FBO)? */
1169 panfrost_batch_is_scanout(struct panfrost_batch
*batch
)
1171 /* If there is no color buffer, it's an FBO */
1172 if (batch
->key
.nr_cbufs
!= 1)
1175 /* If we're too early that no framebuffer was sent, it's scanout */
1176 if (!batch
->key
.cbufs
[0])
1179 return batch
->key
.cbufs
[0]->texture
->bind
& PIPE_BIND_DISPLAY_TARGET
||
1180 batch
->key
.cbufs
[0]->texture
->bind
& PIPE_BIND_SCANOUT
||
1181 batch
->key
.cbufs
[0]->texture
->bind
& PIPE_BIND_SHARED
;
1185 panfrost_batch_init(struct panfrost_context
*ctx
)
1187 ctx
->batches
= _mesa_hash_table_create(ctx
,
1188 panfrost_batch_hash
,
1189 panfrost_batch_compare
);
1190 ctx
->accessed_bos
= _mesa_hash_table_create(ctx
, _mesa_hash_pointer
,
1191 _mesa_key_pointer_equal
);