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"
38 #include "panfrost-quirks.h"
40 /* panfrost_bo_access is here to help us keep track of batch accesses to BOs
41 * and build a proper dependency graph such that batches can be pipelined for
42 * better GPU utilization.
44 * Each accessed BO has a corresponding entry in the ->accessed_bos hash table.
45 * A BO is either being written or read at any time, that's what the type field
47 * When the last access is a write, the batch writing the BO might have read
48 * dependencies (readers that have not been executed yet and want to read the
49 * previous BO content), and when the last access is a read, all readers might
50 * depend on another batch to push its results to memory. That's what the
51 * readers/writers keep track off.
52 * There can only be one writer at any given time, if a new batch wants to
53 * write to the same BO, a dependency will be added between the new writer and
54 * the old writer (at the batch level), and panfrost_bo_access->writer will be
55 * updated to point to the new writer.
57 struct panfrost_bo_access
{
59 struct util_dynarray readers
;
60 struct panfrost_batch_fence
*writer
;
63 static struct panfrost_batch_fence
*
64 panfrost_create_batch_fence(struct panfrost_batch
*batch
)
66 struct panfrost_batch_fence
*fence
;
69 fence
= rzalloc(NULL
, struct panfrost_batch_fence
);
71 pipe_reference_init(&fence
->reference
, 1);
72 fence
->ctx
= batch
->ctx
;
74 ret
= drmSyncobjCreate(pan_screen(batch
->ctx
->base
.screen
)->fd
, 0,
82 panfrost_free_batch_fence(struct panfrost_batch_fence
*fence
)
84 drmSyncobjDestroy(pan_screen(fence
->ctx
->base
.screen
)->fd
,
90 panfrost_batch_fence_unreference(struct panfrost_batch_fence
*fence
)
92 if (pipe_reference(&fence
->reference
, NULL
))
93 panfrost_free_batch_fence(fence
);
97 panfrost_batch_fence_reference(struct panfrost_batch_fence
*fence
)
99 pipe_reference(NULL
, &fence
->reference
);
102 static struct panfrost_batch
*
103 panfrost_create_batch(struct panfrost_context
*ctx
,
104 const struct pipe_framebuffer_state
*key
)
106 struct panfrost_batch
*batch
= rzalloc(ctx
, struct panfrost_batch
);
110 batch
->bos
= _mesa_hash_table_create(batch
, _mesa_hash_pointer
,
111 _mesa_key_pointer_equal
);
113 batch
->minx
= batch
->miny
= ~0;
114 batch
->maxx
= batch
->maxy
= 0;
115 batch
->transient_offset
= 0;
117 batch
->out_sync
= panfrost_create_batch_fence(batch
);
118 util_copy_framebuffer_state(&batch
->key
, key
);
124 panfrost_freeze_batch(struct panfrost_batch
*batch
)
126 struct panfrost_context
*ctx
= batch
->ctx
;
127 struct hash_entry
*entry
;
129 /* Remove the entry in the FBO -> batch hash table if the batch
130 * matches. This way, next draws/clears targeting this FBO will trigger
131 * the creation of a new batch.
133 entry
= _mesa_hash_table_search(ctx
->batches
, &batch
->key
);
134 if (entry
&& entry
->data
== batch
)
135 _mesa_hash_table_remove(ctx
->batches
, entry
);
137 /* If this is the bound batch, the panfrost_context parameters are
138 * relevant so submitting it invalidates those parameters, but if it's
139 * not bound, the context parameters are for some other batch so we
140 * can't invalidate them.
142 if (ctx
->batch
== batch
) {
143 panfrost_invalidate_frame(ctx
);
149 static bool panfrost_batch_is_frozen(struct panfrost_batch
*batch
)
151 struct panfrost_context
*ctx
= batch
->ctx
;
152 struct hash_entry
*entry
;
154 entry
= _mesa_hash_table_search(ctx
->batches
, &batch
->key
);
155 if (entry
&& entry
->data
== batch
)
158 if (ctx
->batch
== batch
)
166 panfrost_free_batch(struct panfrost_batch
*batch
)
171 assert(panfrost_batch_is_frozen(batch
));
173 hash_table_foreach(batch
->bos
, entry
)
174 panfrost_bo_unreference((struct panfrost_bo
*)entry
->key
);
176 util_dynarray_foreach(&batch
->dependencies
,
177 struct panfrost_batch_fence
*, dep
) {
178 panfrost_batch_fence_unreference(*dep
);
181 /* The out_sync fence lifetime is different from the the batch one
182 * since other batches might want to wait on a fence of already
183 * submitted/signaled batch. All we need to do here is make sure the
184 * fence does not point to an invalid batch, which the core will
185 * interpret as 'batch is already submitted'.
187 batch
->out_sync
->batch
= NULL
;
188 panfrost_batch_fence_unreference(batch
->out_sync
);
190 util_unreference_framebuffer_state(&batch
->key
);
196 panfrost_dep_graph_contains_batch(struct panfrost_batch
*root
,
197 struct panfrost_batch
*batch
)
202 util_dynarray_foreach(&root
->dependencies
,
203 struct panfrost_batch_fence
*, dep
) {
204 if ((*dep
)->batch
== batch
||
205 panfrost_dep_graph_contains_batch((*dep
)->batch
, batch
))
214 panfrost_batch_add_dep(struct panfrost_batch
*batch
,
215 struct panfrost_batch_fence
*newdep
)
217 if (batch
== newdep
->batch
)
220 /* We might want to turn ->dependencies into a set if the number of
221 * deps turns out to be big enough to make this 'is dep already there'
222 * search inefficient.
224 util_dynarray_foreach(&batch
->dependencies
,
225 struct panfrost_batch_fence
*, dep
) {
230 /* Make sure the dependency graph is acyclic. */
231 assert(!panfrost_dep_graph_contains_batch(newdep
->batch
, batch
));
233 panfrost_batch_fence_reference(newdep
);
234 util_dynarray_append(&batch
->dependencies
,
235 struct panfrost_batch_fence
*, newdep
);
237 /* We now have a batch depending on us, let's make sure new draw/clear
238 * calls targeting the same FBO use a new batch object.
241 panfrost_freeze_batch(newdep
->batch
);
244 static struct panfrost_batch
*
245 panfrost_get_batch(struct panfrost_context
*ctx
,
246 const struct pipe_framebuffer_state
*key
)
248 /* Lookup the job first */
249 struct hash_entry
*entry
= _mesa_hash_table_search(ctx
->batches
, key
);
254 /* Otherwise, let's create a job */
256 struct panfrost_batch
*batch
= panfrost_create_batch(ctx
, key
);
258 /* Save the created job */
259 _mesa_hash_table_insert(ctx
->batches
, &batch
->key
, batch
);
264 /* Get the job corresponding to the FBO we're currently rendering into */
266 struct panfrost_batch
*
267 panfrost_get_batch_for_fbo(struct panfrost_context
*ctx
)
269 /* If we're wallpapering, we special case to workaround
272 if (ctx
->wallpaper_batch
)
273 return ctx
->wallpaper_batch
;
275 /* If we already began rendering, use that */
278 assert(util_framebuffer_state_equal(&ctx
->batch
->key
,
279 &ctx
->pipe_framebuffer
));
283 /* If not, look up the job */
284 struct panfrost_batch
*batch
= panfrost_get_batch(ctx
,
285 &ctx
->pipe_framebuffer
);
287 /* Set this job as the current FBO job. Will be reset when updating the
288 * FB state and when submitting or releasing a job.
294 struct panfrost_batch
*
295 panfrost_get_fresh_batch_for_fbo(struct panfrost_context
*ctx
)
297 struct panfrost_batch
*batch
;
299 batch
= panfrost_get_batch(ctx
, &ctx
->pipe_framebuffer
);
301 /* The batch has no draw/clear queued, let's return it directly.
302 * Note that it's perfectly fine to re-use a batch with an
303 * existing clear, we'll just update it with the new clear request.
305 if (!batch
->first_job
)
308 /* Otherwise, we need to freeze the existing one and instantiate a new
311 panfrost_freeze_batch(batch
);
312 return panfrost_get_batch(ctx
, &ctx
->pipe_framebuffer
);
316 panfrost_batch_fence_is_signaled(struct panfrost_batch_fence
*fence
)
321 /* Batch has not been submitted yet. */
325 int ret
= drmSyncobjWait(pan_screen(fence
->ctx
->base
.screen
)->fd
,
326 &fence
->syncobj
, 1, 0, 0, NULL
);
328 /* Cache whether the fence was signaled */
329 fence
->signaled
= ret
>= 0;
330 return fence
->signaled
;
334 panfrost_bo_access_gc_fences(struct panfrost_context
*ctx
,
335 struct panfrost_bo_access
*access
,
336 const struct panfrost_bo
*bo
)
338 if (access
->writer
&& panfrost_batch_fence_is_signaled(access
->writer
)) {
339 panfrost_batch_fence_unreference(access
->writer
);
340 access
->writer
= NULL
;
343 struct panfrost_batch_fence
**readers_array
= util_dynarray_begin(&access
->readers
);
344 struct panfrost_batch_fence
**new_readers
= readers_array
;
346 util_dynarray_foreach(&access
->readers
, struct panfrost_batch_fence
*,
351 if (panfrost_batch_fence_is_signaled(*reader
)) {
352 panfrost_batch_fence_unreference(*reader
);
355 /* Build a new array of only unsignaled fences in-place */
356 *(new_readers
++) = *reader
;
360 if (!util_dynarray_resize(&access
->readers
, struct panfrost_batch_fence
*,
361 new_readers
- readers_array
) &&
362 new_readers
!= readers_array
)
363 unreachable("Invalid dynarray access->readers");
366 /* Collect signaled fences to keep the kernel-side syncobj-map small. The
367 * idea is to collect those signaled fences at the end of each flush_all
368 * call. This function is likely to collect only fences from previous
369 * batch flushes not the one that have just have just been submitted and
370 * are probably still in flight when we trigger the garbage collection.
371 * Anyway, we need to do this garbage collection at some point if we don't
372 * want the BO access map to keep invalid entries around and retain
376 panfrost_gc_fences(struct panfrost_context
*ctx
)
378 hash_table_foreach(ctx
->accessed_bos
, entry
) {
379 struct panfrost_bo_access
*access
= entry
->data
;
382 panfrost_bo_access_gc_fences(ctx
, access
, entry
->key
);
383 if (!util_dynarray_num_elements(&access
->readers
,
384 struct panfrost_batch_fence
*) &&
387 _mesa_hash_table_remove(ctx
->accessed_bos
, entry
);
394 panfrost_batch_in_readers(struct panfrost_batch
*batch
,
395 struct panfrost_bo_access
*access
)
397 util_dynarray_foreach(&access
->readers
, struct panfrost_batch_fence
*,
399 if (*reader
&& (*reader
)->batch
== batch
)
408 panfrost_batch_update_bo_access(struct panfrost_batch
*batch
,
409 struct panfrost_bo
*bo
, uint32_t access_type
,
410 bool already_accessed
)
412 struct panfrost_context
*ctx
= batch
->ctx
;
413 struct panfrost_bo_access
*access
;
414 uint32_t old_access_type
;
415 struct hash_entry
*entry
;
417 assert(access_type
== PAN_BO_ACCESS_WRITE
||
418 access_type
== PAN_BO_ACCESS_READ
);
420 entry
= _mesa_hash_table_search(ctx
->accessed_bos
, bo
);
421 access
= entry
? entry
->data
: NULL
;
423 old_access_type
= access
->type
;
425 access
= rzalloc(ctx
, struct panfrost_bo_access
);
426 util_dynarray_init(&access
->readers
, access
);
427 _mesa_hash_table_insert(ctx
->accessed_bos
, bo
, access
);
428 /* We are the first to access this BO, let's initialize
429 * old_access_type to our own access type in that case.
431 old_access_type
= access_type
;
432 access
->type
= access_type
;
437 if (access_type
== PAN_BO_ACCESS_WRITE
&&
438 old_access_type
== PAN_BO_ACCESS_READ
) {
439 /* Previous access was a read and we want to write this BO.
440 * We first need to add explicit deps between our batch and
441 * the previous readers.
443 util_dynarray_foreach(&access
->readers
,
444 struct panfrost_batch_fence
*, reader
) {
445 /* We were already reading the BO, no need to add a dep
446 * on ourself (the acyclic check would complain about
449 if (!(*reader
) || (*reader
)->batch
== batch
)
452 panfrost_batch_add_dep(batch
, *reader
);
454 panfrost_batch_fence_reference(batch
->out_sync
);
456 /* We now are the new writer. */
457 access
->writer
= batch
->out_sync
;
458 access
->type
= access_type
;
460 /* Release the previous readers and reset the readers array. */
461 util_dynarray_foreach(&access
->readers
,
462 struct panfrost_batch_fence
*,
466 panfrost_batch_fence_unreference(*reader
);
469 util_dynarray_clear(&access
->readers
);
470 } else if (access_type
== PAN_BO_ACCESS_WRITE
&&
471 old_access_type
== PAN_BO_ACCESS_WRITE
) {
472 /* Previous access was a write and we want to write this BO.
473 * First check if we were the previous writer, in that case
474 * there's nothing to do. Otherwise we need to add a
475 * dependency between the new writer and the old one.
477 if (access
->writer
!= batch
->out_sync
) {
478 if (access
->writer
) {
479 panfrost_batch_add_dep(batch
, access
->writer
);
480 panfrost_batch_fence_unreference(access
->writer
);
482 panfrost_batch_fence_reference(batch
->out_sync
);
483 access
->writer
= batch
->out_sync
;
485 } else if (access_type
== PAN_BO_ACCESS_READ
&&
486 old_access_type
== PAN_BO_ACCESS_WRITE
) {
487 /* Previous access was a write and we want to read this BO.
488 * First check if we were the previous writer, in that case
489 * we want to keep the access type unchanged, as a write is
490 * more constraining than a read.
492 if (access
->writer
!= batch
->out_sync
) {
493 /* Add a dependency on the previous writer. */
494 panfrost_batch_add_dep(batch
, access
->writer
);
496 /* The previous access was a write, there's no reason
497 * to have entries in the readers array.
499 assert(!util_dynarray_num_elements(&access
->readers
,
500 struct panfrost_batch_fence
*));
502 /* Add ourselves to the readers array. */
503 panfrost_batch_fence_reference(batch
->out_sync
);
504 util_dynarray_append(&access
->readers
,
505 struct panfrost_batch_fence
*,
507 access
->type
= PAN_BO_ACCESS_READ
;
510 /* We already accessed this BO before, so we should already be
511 * in the reader array.
513 if (already_accessed
) {
514 assert(panfrost_batch_in_readers(batch
, access
));
518 /* Previous access was a read and we want to read this BO.
519 * Add ourselves to the readers array and add a dependency on
520 * the previous writer if any.
522 panfrost_batch_fence_reference(batch
->out_sync
);
523 util_dynarray_append(&access
->readers
,
524 struct panfrost_batch_fence
*,
528 panfrost_batch_add_dep(batch
, access
->writer
);
533 panfrost_batch_add_bo(struct panfrost_batch
*batch
, struct panfrost_bo
*bo
,
539 struct hash_entry
*entry
;
540 uint32_t old_flags
= 0;
542 entry
= _mesa_hash_table_search(batch
->bos
, bo
);
544 entry
= _mesa_hash_table_insert(batch
->bos
, bo
,
545 (void *)(uintptr_t)flags
);
546 panfrost_bo_reference(bo
);
548 old_flags
= (uintptr_t)entry
->data
;
550 /* All batches have to agree on the shared flag. */
551 assert((old_flags
& PAN_BO_ACCESS_SHARED
) ==
552 (flags
& PAN_BO_ACCESS_SHARED
));
557 if (old_flags
== flags
)
561 entry
->data
= (void *)(uintptr_t)flags
;
563 /* If this is not a shared BO, we don't really care about dependency
566 if (!(flags
& PAN_BO_ACCESS_SHARED
))
569 /* All dependencies should have been flushed before we execute the
570 * wallpaper draw, so it should be harmless to skip the
571 * update_bo_access() call.
573 if (batch
== batch
->ctx
->wallpaper_batch
)
576 /* Only pass R/W flags to the dep tracking logic. */
577 assert(flags
& PAN_BO_ACCESS_RW
);
578 flags
= (flags
& PAN_BO_ACCESS_WRITE
) ?
579 PAN_BO_ACCESS_WRITE
: PAN_BO_ACCESS_READ
;
580 panfrost_batch_update_bo_access(batch
, bo
, flags
, old_flags
!= 0);
583 void panfrost_batch_add_fbo_bos(struct panfrost_batch
*batch
)
585 uint32_t flags
= PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_WRITE
|
586 PAN_BO_ACCESS_VERTEX_TILER
|
587 PAN_BO_ACCESS_FRAGMENT
;
589 for (unsigned i
= 0; i
< batch
->key
.nr_cbufs
; ++i
) {
590 struct panfrost_resource
*rsrc
= pan_resource(batch
->key
.cbufs
[i
]->texture
);
591 panfrost_batch_add_bo(batch
, rsrc
->bo
, flags
);
594 if (batch
->key
.zsbuf
) {
595 struct panfrost_resource
*rsrc
= pan_resource(batch
->key
.zsbuf
->texture
);
596 panfrost_batch_add_bo(batch
, rsrc
->bo
, flags
);
601 panfrost_batch_create_bo(struct panfrost_batch
*batch
, size_t size
,
602 uint32_t create_flags
, uint32_t access_flags
)
604 struct panfrost_bo
*bo
;
606 bo
= panfrost_bo_create(pan_screen(batch
->ctx
->base
.screen
), size
,
608 panfrost_batch_add_bo(batch
, bo
, access_flags
);
610 /* panfrost_batch_add_bo() has retained a reference and
611 * panfrost_bo_create() initialize the refcnt to 1, so let's
612 * unreference the BO here so it gets released when the batch is
613 * destroyed (unless it's retained by someone else in the meantime).
615 panfrost_bo_unreference(bo
);
619 /* Returns the polygon list's GPU address if available, or otherwise allocates
620 * the polygon list. It's perfectly fast to use allocate/free BO directly,
621 * since we'll hit the BO cache and this is one-per-batch anyway. */
624 panfrost_batch_get_polygon_list(struct panfrost_batch
*batch
, unsigned size
)
626 if (batch
->polygon_list
) {
627 assert(batch
->polygon_list
->size
>= size
);
629 /* Create the BO as invisible, as there's no reason to map */
630 size
= util_next_power_of_two(size
);
632 batch
->polygon_list
= panfrost_batch_create_bo(batch
, size
,
634 PAN_BO_ACCESS_PRIVATE
|
636 PAN_BO_ACCESS_VERTEX_TILER
|
637 PAN_BO_ACCESS_FRAGMENT
);
640 return batch
->polygon_list
->gpu
;
644 panfrost_batch_get_scratchpad(struct panfrost_batch
*batch
,
646 unsigned thread_tls_alloc
,
649 unsigned size
= panfrost_get_total_stack_size(shift
,
653 if (batch
->scratchpad
) {
654 assert(batch
->scratchpad
->size
>= size
);
656 batch
->scratchpad
= panfrost_batch_create_bo(batch
, size
,
658 PAN_BO_ACCESS_PRIVATE
|
660 PAN_BO_ACCESS_VERTEX_TILER
|
661 PAN_BO_ACCESS_FRAGMENT
);
664 return batch
->scratchpad
;
668 panfrost_batch_get_shared_memory(struct panfrost_batch
*batch
,
670 unsigned workgroup_count
)
672 if (batch
->shared_memory
) {
673 assert(batch
->shared_memory
->size
>= size
);
675 batch
->shared_memory
= panfrost_batch_create_bo(batch
, size
,
677 PAN_BO_ACCESS_PRIVATE
|
679 PAN_BO_ACCESS_VERTEX_TILER
);
682 return batch
->shared_memory
;
686 panfrost_batch_get_tiler_heap(struct panfrost_batch
*batch
)
688 if (batch
->tiler_heap
)
689 return batch
->tiler_heap
;
691 batch
->tiler_heap
= panfrost_batch_create_bo(batch
, 4096 * 4096,
694 PAN_BO_ACCESS_PRIVATE
|
696 PAN_BO_ACCESS_VERTEX_TILER
|
697 PAN_BO_ACCESS_FRAGMENT
);
698 assert(batch
->tiler_heap
);
699 return batch
->tiler_heap
;
703 panfrost_batch_get_tiler_dummy(struct panfrost_batch
*batch
)
705 struct panfrost_screen
*screen
= pan_screen(batch
->ctx
->base
.screen
);
707 uint32_t create_flags
= 0;
709 if (batch
->tiler_dummy
)
710 return batch
->tiler_dummy
;
712 if (!(screen
->quirks
& MIDGARD_NO_HIER_TILING
))
713 create_flags
= PAN_BO_INVISIBLE
;
715 batch
->tiler_dummy
= panfrost_batch_create_bo(batch
, 4096,
717 PAN_BO_ACCESS_PRIVATE
|
719 PAN_BO_ACCESS_VERTEX_TILER
|
720 PAN_BO_ACCESS_FRAGMENT
);
721 assert(batch
->tiler_dummy
);
722 return batch
->tiler_dummy
;
726 panfrost_batch_draw_wallpaper(struct panfrost_batch
*batch
)
728 /* Color 0 is cleared, no need to draw the wallpaper.
729 * TODO: MRT wallpapers.
731 if (batch
->clear
& PIPE_CLEAR_COLOR0
)
734 /* Nothing to reload? TODO: MRT wallpapers */
735 if (batch
->key
.cbufs
[0] == NULL
)
738 /* No draw calls, and no clear on the depth/stencil bufs.
739 * Drawing the wallpaper would be useless.
741 if (!batch
->tiler_dep
&&
742 !(batch
->clear
& PIPE_CLEAR_DEPTHSTENCIL
))
745 /* Check if the buffer has any content on it worth preserving */
747 struct pipe_surface
*surf
= batch
->key
.cbufs
[0];
748 struct panfrost_resource
*rsrc
= pan_resource(surf
->texture
);
749 unsigned level
= surf
->u
.tex
.level
;
751 if (!rsrc
->slices
[level
].initialized
)
754 batch
->ctx
->wallpaper_batch
= batch
;
756 /* Clamp the rendering area to the damage extent. The
757 * KHR_partial_update() spec states that trying to render outside of
758 * the damage region is "undefined behavior", so we should be safe.
760 unsigned damage_width
= (rsrc
->damage
.extent
.maxx
- rsrc
->damage
.extent
.minx
);
761 unsigned damage_height
= (rsrc
->damage
.extent
.maxy
- rsrc
->damage
.extent
.miny
);
763 if (damage_width
&& damage_height
) {
764 panfrost_batch_intersection_scissor(batch
,
765 rsrc
->damage
.extent
.minx
,
766 rsrc
->damage
.extent
.miny
,
767 rsrc
->damage
.extent
.maxx
,
768 rsrc
->damage
.extent
.maxy
);
771 /* FIXME: Looks like aligning on a tile is not enough, but
772 * aligning on twice the tile size seems to works. We don't
773 * know exactly what happens here but this deserves extra
774 * investigation to figure it out.
776 batch
->minx
= batch
->minx
& ~((MALI_TILE_LENGTH
* 2) - 1);
777 batch
->miny
= batch
->miny
& ~((MALI_TILE_LENGTH
* 2) - 1);
778 batch
->maxx
= MIN2(ALIGN_POT(batch
->maxx
, MALI_TILE_LENGTH
* 2),
780 batch
->maxy
= MIN2(ALIGN_POT(batch
->maxy
, MALI_TILE_LENGTH
* 2),
783 struct pipe_scissor_state damage
;
784 struct pipe_box rects
[4];
786 /* Clamp the damage box to the rendering area. */
787 damage
.minx
= MAX2(batch
->minx
, rsrc
->damage
.biggest_rect
.x
);
788 damage
.miny
= MAX2(batch
->miny
, rsrc
->damage
.biggest_rect
.y
);
789 damage
.maxx
= MIN2(batch
->maxx
,
790 rsrc
->damage
.biggest_rect
.x
+
791 rsrc
->damage
.biggest_rect
.width
);
792 damage
.maxx
= MAX2(damage
.maxx
, damage
.minx
);
793 damage
.maxy
= MIN2(batch
->maxy
,
794 rsrc
->damage
.biggest_rect
.y
+
795 rsrc
->damage
.biggest_rect
.height
);
796 damage
.maxy
= MAX2(damage
.maxy
, damage
.miny
);
798 /* One damage rectangle means we can end up with at most 4 reload
800 * 1: left region, only exists if damage.x > 0
801 * 2: right region, only exists if damage.x + damage.width < fb->width
802 * 3: top region, only exists if damage.y > 0. The intersection with
803 * the left and right regions are dropped
804 * 4: bottom region, only exists if damage.y + damage.height < fb->height.
805 * The intersection with the left and right regions are dropped
807 * ____________________________
814 * |_______|___________|______|
816 u_box_2d(batch
->minx
, batch
->miny
, damage
.minx
- batch
->minx
,
817 batch
->maxy
- batch
->miny
, &rects
[0]);
818 u_box_2d(damage
.maxx
, batch
->miny
, batch
->maxx
- damage
.maxx
,
819 batch
->maxy
- batch
->miny
, &rects
[1]);
820 u_box_2d(damage
.minx
, batch
->miny
, damage
.maxx
- damage
.minx
,
821 damage
.miny
- batch
->miny
, &rects
[2]);
822 u_box_2d(damage
.minx
, damage
.maxy
, damage
.maxx
- damage
.minx
,
823 batch
->maxy
- damage
.maxy
, &rects
[3]);
825 for (unsigned i
= 0; i
< 4; i
++) {
826 /* Width and height are always >= 0 even if width is declared as a
827 * signed integer: u_box_2d() helper takes unsigned args and
828 * panfrost_set_damage_region() is taking care of clamping
831 if (!rects
[i
].width
|| !rects
[i
].height
)
834 /* Blit the wallpaper in */
835 panfrost_blit_wallpaper(batch
->ctx
, &rects
[i
]);
837 batch
->ctx
->wallpaper_batch
= NULL
;
841 panfrost_batch_submit_ioctl(struct panfrost_batch
*batch
,
842 mali_ptr first_job_desc
,
845 struct panfrost_context
*ctx
= batch
->ctx
;
846 struct pipe_context
*gallium
= (struct pipe_context
*) ctx
;
847 struct panfrost_screen
*screen
= pan_screen(gallium
->screen
);
848 struct drm_panfrost_submit submit
= {0,};
849 uint32_t *bo_handles
, *in_syncs
= NULL
;
850 bool is_fragment_shader
;
853 is_fragment_shader
= (reqs
& PANFROST_JD_REQ_FS
) && batch
->first_job
;
854 if (is_fragment_shader
)
855 submit
.in_sync_count
= 1;
857 submit
.in_sync_count
= util_dynarray_num_elements(&batch
->dependencies
,
858 struct panfrost_batch_fence
*);
860 if (submit
.in_sync_count
) {
861 in_syncs
= calloc(submit
.in_sync_count
, sizeof(*in_syncs
));
865 /* The fragment job always depends on the vertex/tiler job if there's
868 if (is_fragment_shader
) {
869 in_syncs
[0] = batch
->out_sync
->syncobj
;
873 util_dynarray_foreach(&batch
->dependencies
,
874 struct panfrost_batch_fence
*, dep
)
875 in_syncs
[i
++] = (*dep
)->syncobj
;
878 submit
.in_syncs
= (uintptr_t)in_syncs
;
879 submit
.out_sync
= batch
->out_sync
->syncobj
;
880 submit
.jc
= first_job_desc
;
881 submit
.requirements
= reqs
;
883 bo_handles
= calloc(batch
->bos
->entries
, sizeof(*bo_handles
));
886 hash_table_foreach(batch
->bos
, entry
) {
887 struct panfrost_bo
*bo
= (struct panfrost_bo
*)entry
->key
;
888 uint32_t flags
= (uintptr_t)entry
->data
;
890 assert(bo
->gem_handle
> 0);
891 bo_handles
[submit
.bo_handle_count
++] = bo
->gem_handle
;
893 /* Update the BO access flags so that panfrost_bo_wait() knows
894 * about all pending accesses.
895 * We only keep the READ/WRITE info since this is all the BO
896 * wait logic cares about.
897 * We also preserve existing flags as this batch might not
898 * be the first one to access the BO.
900 bo
->gpu_access
|= flags
& (PAN_BO_ACCESS_RW
);
903 submit
.bo_handles
= (u64
) (uintptr_t) bo_handles
;
904 ret
= drmIoctl(screen
->fd
, DRM_IOCTL_PANFROST_SUBMIT
, &submit
);
909 DBG("Error submitting: %m\n");
913 /* Trace the job if we're doing that */
914 if (pan_debug
& (PAN_DBG_TRACE
| PAN_DBG_SYNC
)) {
915 /* Wait so we can get errors reported back */
916 drmSyncobjWait(screen
->fd
, &batch
->out_sync
->syncobj
, 1,
919 /* Trace gets priority over sync */
920 bool minimal
= !(pan_debug
& PAN_DBG_TRACE
);
921 pandecode_jc(submit
.jc
, FALSE
, screen
->gpu_id
, minimal
);
928 panfrost_batch_submit_jobs(struct panfrost_batch
*batch
)
930 bool has_draws
= batch
->first_job
;
934 ret
= panfrost_batch_submit_ioctl(batch
, batch
->first_job
, 0);
938 if (batch
->tiler_dep
|| batch
->clear
) {
939 mali_ptr fragjob
= panfrost_fragment_job(batch
, has_draws
);
940 ret
= panfrost_batch_submit_ioctl(batch
, fragjob
, PANFROST_JD_REQ_FS
);
948 panfrost_batch_submit(struct panfrost_batch
*batch
)
952 /* Submit the dependencies first. */
953 util_dynarray_foreach(&batch
->dependencies
,
954 struct panfrost_batch_fence
*, dep
) {
956 panfrost_batch_submit((*dep
)->batch
);
962 if (!batch
->first_job
&& !batch
->clear
) {
963 /* Mark the fence as signaled so the fence logic does not try
966 batch
->out_sync
->signaled
= true;
970 panfrost_batch_draw_wallpaper(batch
);
972 /* Now that all draws are in, we can finally prepare the
973 * FBD for the batch */
975 if (batch
->framebuffer
.gpu
&& batch
->first_job
) {
976 struct panfrost_context
*ctx
= batch
->ctx
;
977 struct pipe_context
*gallium
= (struct pipe_context
*) ctx
;
978 struct panfrost_screen
*screen
= pan_screen(gallium
->screen
);
980 if (screen
->quirks
& MIDGARD_SFBD
)
981 panfrost_attach_sfbd(batch
, ~0);
983 panfrost_attach_mfbd(batch
, ~0);
986 panfrost_scoreboard_initialize_tiler(batch
);
988 ret
= panfrost_batch_submit_jobs(batch
);
991 DBG("panfrost_batch_submit failed: %d\n", ret
);
993 /* We must reset the damage info of our render targets here even
994 * though a damage reset normally happens when the DRI layer swaps
995 * buffers. That's because there can be implicit flushes the GL
996 * app is not aware of, and those might impact the damage region: if
997 * part of the damaged portion is drawn during those implicit flushes,
998 * you have to reload those areas before next draws are pushed, and
999 * since the driver can't easily know what's been modified by the draws
1000 * it flushed, the easiest solution is to reload everything.
1002 for (unsigned i
= 0; i
< batch
->key
.nr_cbufs
; i
++) {
1003 struct panfrost_resource
*res
;
1005 if (!batch
->key
.cbufs
[i
])
1008 res
= pan_resource(batch
->key
.cbufs
[i
]->texture
);
1009 panfrost_resource_reset_damage(res
);
1013 panfrost_freeze_batch(batch
);
1014 panfrost_free_batch(batch
);
1018 panfrost_flush_all_batches(struct panfrost_context
*ctx
, bool wait
)
1020 struct util_dynarray fences
, syncobjs
;
1023 util_dynarray_init(&fences
, NULL
);
1024 util_dynarray_init(&syncobjs
, NULL
);
1027 hash_table_foreach(ctx
->batches
, hentry
) {
1028 struct panfrost_batch
*batch
= hentry
->data
;
1033 panfrost_batch_fence_reference(batch
->out_sync
);
1034 util_dynarray_append(&fences
, struct panfrost_batch_fence
*,
1036 util_dynarray_append(&syncobjs
, uint32_t,
1037 batch
->out_sync
->syncobj
);
1040 panfrost_batch_submit(batch
);
1043 assert(!ctx
->batches
->entries
);
1045 /* Collect batch fences before returning */
1046 panfrost_gc_fences(ctx
);
1051 drmSyncobjWait(pan_screen(ctx
->base
.screen
)->fd
,
1052 util_dynarray_begin(&syncobjs
),
1053 util_dynarray_num_elements(&syncobjs
, uint32_t),
1054 INT64_MAX
, DRM_SYNCOBJ_WAIT_FLAGS_WAIT_ALL
, NULL
);
1056 util_dynarray_foreach(&fences
, struct panfrost_batch_fence
*, fence
)
1057 panfrost_batch_fence_unreference(*fence
);
1059 util_dynarray_fini(&fences
);
1060 util_dynarray_fini(&syncobjs
);
1064 panfrost_pending_batches_access_bo(struct panfrost_context
*ctx
,
1065 const struct panfrost_bo
*bo
)
1067 struct panfrost_bo_access
*access
;
1068 struct hash_entry
*hentry
;
1070 hentry
= _mesa_hash_table_search(ctx
->accessed_bos
, bo
);
1071 access
= hentry
? hentry
->data
: NULL
;
1075 if (access
->writer
&& access
->writer
->batch
)
1078 util_dynarray_foreach(&access
->readers
, struct panfrost_batch_fence
*,
1080 if (*reader
&& (*reader
)->batch
)
1088 panfrost_flush_batches_accessing_bo(struct panfrost_context
*ctx
,
1089 struct panfrost_bo
*bo
,
1090 uint32_t access_type
)
1092 struct panfrost_bo_access
*access
;
1093 struct hash_entry
*hentry
;
1095 /* It doesn't make any to flush only the readers. */
1096 assert(access_type
== PAN_BO_ACCESS_WRITE
||
1097 access_type
== PAN_BO_ACCESS_RW
);
1099 hentry
= _mesa_hash_table_search(ctx
->accessed_bos
, bo
);
1100 access
= hentry
? hentry
->data
: NULL
;
1104 if (access_type
& PAN_BO_ACCESS_WRITE
&& access
->writer
&&
1105 access
->writer
->batch
)
1106 panfrost_batch_submit(access
->writer
->batch
);
1108 if (!(access_type
& PAN_BO_ACCESS_READ
))
1111 util_dynarray_foreach(&access
->readers
, struct panfrost_batch_fence
*,
1113 if (*reader
&& (*reader
)->batch
)
1114 panfrost_batch_submit((*reader
)->batch
);
1119 panfrost_batch_set_requirements(struct panfrost_batch
*batch
)
1121 struct panfrost_context
*ctx
= batch
->ctx
;
1123 if (ctx
->rasterizer
&& ctx
->rasterizer
->base
.multisample
)
1124 batch
->requirements
|= PAN_REQ_MSAA
;
1126 if (ctx
->depth_stencil
&& ctx
->depth_stencil
->depth
.writemask
)
1127 batch
->requirements
|= PAN_REQ_DEPTH_WRITE
;
1130 /* Helper to smear a 32-bit color across 128-bit components */
1133 pan_pack_color_32(uint32_t *packed
, uint32_t v
)
1135 for (unsigned i
= 0; i
< 4; ++i
)
1140 pan_pack_color_64(uint32_t *packed
, uint32_t lo
, uint32_t hi
)
1142 for (unsigned i
= 0; i
< 4; i
+= 2) {
1149 pan_pack_color(uint32_t *packed
, const union pipe_color_union
*color
, enum pipe_format format
)
1151 /* Alpha magicked to 1.0 if there is no alpha */
1153 bool has_alpha
= util_format_has_alpha(format
);
1154 float clear_alpha
= has_alpha
? color
->f
[3] : 1.0f
;
1156 /* Packed color depends on the framebuffer format */
1158 const struct util_format_description
*desc
=
1159 util_format_description(format
);
1161 if (util_format_is_rgba8_variant(desc
)) {
1162 pan_pack_color_32(packed
,
1163 ((uint32_t) float_to_ubyte(clear_alpha
) << 24) |
1164 ((uint32_t) float_to_ubyte(color
->f
[2]) << 16) |
1165 ((uint32_t) float_to_ubyte(color
->f
[1]) << 8) |
1166 ((uint32_t) float_to_ubyte(color
->f
[0]) << 0));
1167 } else if (format
== PIPE_FORMAT_B5G6R5_UNORM
) {
1168 /* First, we convert the components to R5, G6, B5 separately */
1169 unsigned r5
= CLAMP(color
->f
[0], 0.0, 1.0) * 31.0;
1170 unsigned g6
= CLAMP(color
->f
[1], 0.0, 1.0) * 63.0;
1171 unsigned b5
= CLAMP(color
->f
[2], 0.0, 1.0) * 31.0;
1173 /* Then we pack into a sparse u32. TODO: Why these shifts? */
1174 pan_pack_color_32(packed
, (b5
<< 25) | (g6
<< 14) | (r5
<< 5));
1175 } else if (format
== PIPE_FORMAT_B4G4R4A4_UNORM
) {
1176 /* We scale the components against 0xF0 (=240.0), rather than 0xFF */
1177 unsigned r4
= CLAMP(color
->f
[0], 0.0, 1.0) * 240.0;
1178 unsigned g4
= CLAMP(color
->f
[1], 0.0, 1.0) * 240.0;
1179 unsigned b4
= CLAMP(color
->f
[2], 0.0, 1.0) * 240.0;
1180 unsigned a4
= CLAMP(clear_alpha
, 0.0, 1.0) * 240.0;
1182 /* Pack on *byte* intervals */
1183 pan_pack_color_32(packed
, (a4
<< 24) | (b4
<< 16) | (g4
<< 8) | r4
);
1184 } else if (format
== PIPE_FORMAT_B5G5R5A1_UNORM
) {
1185 /* Scale as expected but shift oddly */
1186 unsigned r5
= round(CLAMP(color
->f
[0], 0.0, 1.0)) * 31.0;
1187 unsigned g5
= round(CLAMP(color
->f
[1], 0.0, 1.0)) * 31.0;
1188 unsigned b5
= round(CLAMP(color
->f
[2], 0.0, 1.0)) * 31.0;
1189 unsigned a1
= round(CLAMP(clear_alpha
, 0.0, 1.0)) * 1.0;
1191 pan_pack_color_32(packed
, (a1
<< 31) | (b5
<< 25) | (g5
<< 15) | (r5
<< 5));
1193 /* Try Gallium's generic default path. Doesn't work for all
1194 * formats but it's a good guess. */
1196 union util_color out
;
1198 if (util_format_is_pure_integer(format
)) {
1199 memcpy(out
.ui
, color
->ui
, 16);
1201 util_pack_color(color
->f
, format
, &out
);
1204 unsigned size
= util_format_get_blocksize(format
);
1207 unsigned b
= out
.ui
[0];
1208 unsigned s
= b
| (b
<< 8);
1209 pan_pack_color_32(packed
, s
| (s
<< 16));
1210 } else if (size
== 2)
1211 pan_pack_color_32(packed
, out
.ui
[0] | (out
.ui
[0] << 16));
1212 else if (size
== 3 || size
== 4)
1213 pan_pack_color_32(packed
, out
.ui
[0]);
1215 pan_pack_color_64(packed
, out
.ui
[0], out
.ui
[1] | (out
.ui
[1] << 16)); /* RGB16F -- RGBB */
1217 pan_pack_color_64(packed
, out
.ui
[0], out
.ui
[1]);
1218 else if (size
== 16)
1219 memcpy(packed
, out
.ui
, 16);
1221 unreachable("Unknown generic format size packing clear colour");
1226 panfrost_batch_clear(struct panfrost_batch
*batch
,
1228 const union pipe_color_union
*color
,
1229 double depth
, unsigned stencil
)
1231 struct panfrost_context
*ctx
= batch
->ctx
;
1233 if (buffers
& PIPE_CLEAR_COLOR
) {
1234 for (unsigned i
= 0; i
< PIPE_MAX_COLOR_BUFS
; ++i
) {
1235 if (!(buffers
& (PIPE_CLEAR_COLOR0
<< i
)))
1238 enum pipe_format format
= ctx
->pipe_framebuffer
.cbufs
[i
]->format
;
1239 pan_pack_color(batch
->clear_color
[i
], color
, format
);
1243 if (buffers
& PIPE_CLEAR_DEPTH
) {
1244 batch
->clear_depth
= depth
;
1247 if (buffers
& PIPE_CLEAR_STENCIL
) {
1248 batch
->clear_stencil
= stencil
;
1251 batch
->clear
|= buffers
;
1253 /* Clearing affects the entire framebuffer (by definition -- this is
1254 * the Gallium clear callback, which clears the whole framebuffer. If
1255 * the scissor test were enabled from the GL side, the state tracker
1256 * would emit a quad instead and we wouldn't go down this code path) */
1258 panfrost_batch_union_scissor(batch
, 0, 0,
1259 ctx
->pipe_framebuffer
.width
,
1260 ctx
->pipe_framebuffer
.height
);
1264 panfrost_batch_compare(const void *a
, const void *b
)
1266 return util_framebuffer_state_equal(a
, b
);
1270 panfrost_batch_hash(const void *key
)
1272 return _mesa_hash_data(key
, sizeof(struct pipe_framebuffer_state
));
1275 /* Given a new bounding rectangle (scissor), let the job cover the union of the
1276 * new and old bounding rectangles */
1279 panfrost_batch_union_scissor(struct panfrost_batch
*batch
,
1280 unsigned minx
, unsigned miny
,
1281 unsigned maxx
, unsigned maxy
)
1283 batch
->minx
= MIN2(batch
->minx
, minx
);
1284 batch
->miny
= MIN2(batch
->miny
, miny
);
1285 batch
->maxx
= MAX2(batch
->maxx
, maxx
);
1286 batch
->maxy
= MAX2(batch
->maxy
, maxy
);
1290 panfrost_batch_intersection_scissor(struct panfrost_batch
*batch
,
1291 unsigned minx
, unsigned miny
,
1292 unsigned maxx
, unsigned maxy
)
1294 batch
->minx
= MAX2(batch
->minx
, minx
);
1295 batch
->miny
= MAX2(batch
->miny
, miny
);
1296 batch
->maxx
= MIN2(batch
->maxx
, maxx
);
1297 batch
->maxy
= MIN2(batch
->maxy
, maxy
);
1300 /* Are we currently rendering to the screen (rather than an FBO)? */
1303 panfrost_batch_is_scanout(struct panfrost_batch
*batch
)
1305 /* If there is no color buffer, it's an FBO */
1306 if (batch
->key
.nr_cbufs
!= 1)
1309 /* If we're too early that no framebuffer was sent, it's scanout */
1310 if (!batch
->key
.cbufs
[0])
1313 return batch
->key
.cbufs
[0]->texture
->bind
& PIPE_BIND_DISPLAY_TARGET
||
1314 batch
->key
.cbufs
[0]->texture
->bind
& PIPE_BIND_SCANOUT
||
1315 batch
->key
.cbufs
[0]->texture
->bind
& PIPE_BIND_SHARED
;
1319 panfrost_batch_init(struct panfrost_context
*ctx
)
1321 ctx
->batches
= _mesa_hash_table_create(ctx
,
1322 panfrost_batch_hash
,
1323 panfrost_batch_compare
);
1324 ctx
->accessed_bos
= _mesa_hash_table_create(ctx
, _mesa_hash_pointer
,
1325 _mesa_key_pointer_equal
);