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 util_dynarray_init(&batch
->headers
, batch
);
118 util_dynarray_init(&batch
->gpu_headers
, batch
);
119 util_dynarray_init(&batch
->dependencies
, batch
);
120 batch
->out_sync
= panfrost_create_batch_fence(batch
);
121 util_copy_framebuffer_state(&batch
->key
, key
);
127 panfrost_freeze_batch(struct panfrost_batch
*batch
)
129 struct panfrost_context
*ctx
= batch
->ctx
;
130 struct hash_entry
*entry
;
132 /* Remove the entry in the FBO -> batch hash table if the batch
133 * matches. This way, next draws/clears targeting this FBO will trigger
134 * the creation of a new batch.
136 entry
= _mesa_hash_table_search(ctx
->batches
, &batch
->key
);
137 if (entry
&& entry
->data
== batch
)
138 _mesa_hash_table_remove(ctx
->batches
, entry
);
140 /* If this is the bound batch, the panfrost_context parameters are
141 * relevant so submitting it invalidates those parameters, but if it's
142 * not bound, the context parameters are for some other batch so we
143 * can't invalidate them.
145 if (ctx
->batch
== batch
) {
146 panfrost_invalidate_frame(ctx
);
152 static bool panfrost_batch_is_frozen(struct panfrost_batch
*batch
)
154 struct panfrost_context
*ctx
= batch
->ctx
;
155 struct hash_entry
*entry
;
157 entry
= _mesa_hash_table_search(ctx
->batches
, &batch
->key
);
158 if (entry
&& entry
->data
== batch
)
161 if (ctx
->batch
== batch
)
169 panfrost_free_batch(struct panfrost_batch
*batch
)
174 assert(panfrost_batch_is_frozen(batch
));
176 hash_table_foreach(batch
->bos
, entry
)
177 panfrost_bo_unreference((struct panfrost_bo
*)entry
->key
);
179 util_dynarray_foreach(&batch
->dependencies
,
180 struct panfrost_batch_fence
*, dep
) {
181 panfrost_batch_fence_unreference(*dep
);
184 util_dynarray_fini(&batch
->headers
);
185 util_dynarray_fini(&batch
->gpu_headers
);
187 /* The out_sync fence lifetime is different from the the batch one
188 * since other batches might want to wait on a fence of already
189 * submitted/signaled batch. All we need to do here is make sure the
190 * fence does not point to an invalid batch, which the core will
191 * interpret as 'batch is already submitted'.
193 batch
->out_sync
->batch
= NULL
;
194 panfrost_batch_fence_unreference(batch
->out_sync
);
196 util_unreference_framebuffer_state(&batch
->key
);
202 panfrost_dep_graph_contains_batch(struct panfrost_batch
*root
,
203 struct panfrost_batch
*batch
)
208 util_dynarray_foreach(&root
->dependencies
,
209 struct panfrost_batch_fence
*, dep
) {
210 if ((*dep
)->batch
== batch
||
211 panfrost_dep_graph_contains_batch((*dep
)->batch
, batch
))
220 panfrost_batch_add_dep(struct panfrost_batch
*batch
,
221 struct panfrost_batch_fence
*newdep
)
223 if (batch
== newdep
->batch
)
226 /* We might want to turn ->dependencies into a set if the number of
227 * deps turns out to be big enough to make this 'is dep already there'
228 * search inefficient.
230 util_dynarray_foreach(&batch
->dependencies
,
231 struct panfrost_batch_fence
*, dep
) {
236 /* Make sure the dependency graph is acyclic. */
237 assert(!panfrost_dep_graph_contains_batch(newdep
->batch
, batch
));
239 panfrost_batch_fence_reference(newdep
);
240 util_dynarray_append(&batch
->dependencies
,
241 struct panfrost_batch_fence
*, newdep
);
243 /* We now have a batch depending on us, let's make sure new draw/clear
244 * calls targeting the same FBO use a new batch object.
247 panfrost_freeze_batch(newdep
->batch
);
250 static struct panfrost_batch
*
251 panfrost_get_batch(struct panfrost_context
*ctx
,
252 const struct pipe_framebuffer_state
*key
)
254 /* Lookup the job first */
255 struct hash_entry
*entry
= _mesa_hash_table_search(ctx
->batches
, key
);
260 /* Otherwise, let's create a job */
262 struct panfrost_batch
*batch
= panfrost_create_batch(ctx
, key
);
264 /* Save the created job */
265 _mesa_hash_table_insert(ctx
->batches
, &batch
->key
, batch
);
270 /* Get the job corresponding to the FBO we're currently rendering into */
272 struct panfrost_batch
*
273 panfrost_get_batch_for_fbo(struct panfrost_context
*ctx
)
275 /* If we're wallpapering, we special case to workaround
278 if (ctx
->wallpaper_batch
)
279 return ctx
->wallpaper_batch
;
281 /* If we already began rendering, use that */
284 assert(util_framebuffer_state_equal(&ctx
->batch
->key
,
285 &ctx
->pipe_framebuffer
));
289 /* If not, look up the job */
290 struct panfrost_batch
*batch
= panfrost_get_batch(ctx
,
291 &ctx
->pipe_framebuffer
);
293 /* Set this job as the current FBO job. Will be reset when updating the
294 * FB state and when submitting or releasing a job.
300 struct panfrost_batch
*
301 panfrost_get_fresh_batch_for_fbo(struct panfrost_context
*ctx
)
303 struct panfrost_batch
*batch
;
305 batch
= panfrost_get_batch(ctx
, &ctx
->pipe_framebuffer
);
307 /* The batch has no draw/clear queued, let's return it directly.
308 * Note that it's perfectly fine to re-use a batch with an
309 * existing clear, we'll just update it with the new clear request.
311 if (!batch
->last_job
.gpu
)
314 /* Otherwise, we need to freeze the existing one and instantiate a new
317 panfrost_freeze_batch(batch
);
318 return panfrost_get_batch(ctx
, &ctx
->pipe_framebuffer
);
322 panfrost_batch_fence_is_signaled(struct panfrost_batch_fence
*fence
)
327 /* Batch has not been submitted yet. */
331 int ret
= drmSyncobjWait(pan_screen(fence
->ctx
->base
.screen
)->fd
,
332 &fence
->syncobj
, 1, 0, 0, NULL
);
334 /* Cache whether the fence was signaled */
335 fence
->signaled
= ret
>= 0;
336 return fence
->signaled
;
340 panfrost_bo_access_gc_fences(struct panfrost_context
*ctx
,
341 struct panfrost_bo_access
*access
,
342 const struct panfrost_bo
*bo
)
344 if (access
->writer
&& panfrost_batch_fence_is_signaled(access
->writer
)) {
345 panfrost_batch_fence_unreference(access
->writer
);
346 access
->writer
= NULL
;
349 struct panfrost_batch_fence
**readers_array
= util_dynarray_begin(&access
->readers
);
350 struct panfrost_batch_fence
**new_readers
= readers_array
;
352 util_dynarray_foreach(&access
->readers
, struct panfrost_batch_fence
*,
357 if (panfrost_batch_fence_is_signaled(*reader
)) {
358 panfrost_batch_fence_unreference(*reader
);
361 /* Build a new array of only unsignaled fences in-place */
362 *(new_readers
++) = *reader
;
366 if (!util_dynarray_resize(&access
->readers
, struct panfrost_batch_fence
*,
367 new_readers
- readers_array
) &&
368 new_readers
!= readers_array
)
369 unreachable("Invalid dynarray access->readers");
372 /* Collect signaled fences to keep the kernel-side syncobj-map small. The
373 * idea is to collect those signaled fences at the end of each flush_all
374 * call. This function is likely to collect only fences from previous
375 * batch flushes not the one that have just have just been submitted and
376 * are probably still in flight when we trigger the garbage collection.
377 * Anyway, we need to do this garbage collection at some point if we don't
378 * want the BO access map to keep invalid entries around and retain
382 panfrost_gc_fences(struct panfrost_context
*ctx
)
384 hash_table_foreach(ctx
->accessed_bos
, entry
) {
385 struct panfrost_bo_access
*access
= entry
->data
;
388 panfrost_bo_access_gc_fences(ctx
, access
, entry
->key
);
389 if (!util_dynarray_num_elements(&access
->readers
,
390 struct panfrost_batch_fence
*) &&
393 _mesa_hash_table_remove(ctx
->accessed_bos
, entry
);
400 panfrost_batch_in_readers(struct panfrost_batch
*batch
,
401 struct panfrost_bo_access
*access
)
403 util_dynarray_foreach(&access
->readers
, struct panfrost_batch_fence
*,
405 if (*reader
&& (*reader
)->batch
== batch
)
414 panfrost_batch_update_bo_access(struct panfrost_batch
*batch
,
415 struct panfrost_bo
*bo
, uint32_t access_type
,
416 bool already_accessed
)
418 struct panfrost_context
*ctx
= batch
->ctx
;
419 struct panfrost_bo_access
*access
;
420 uint32_t old_access_type
;
421 struct hash_entry
*entry
;
423 assert(access_type
== PAN_BO_ACCESS_WRITE
||
424 access_type
== PAN_BO_ACCESS_READ
);
426 entry
= _mesa_hash_table_search(ctx
->accessed_bos
, bo
);
427 access
= entry
? entry
->data
: NULL
;
429 old_access_type
= access
->type
;
431 access
= rzalloc(ctx
, struct panfrost_bo_access
);
432 util_dynarray_init(&access
->readers
, access
);
433 _mesa_hash_table_insert(ctx
->accessed_bos
, bo
, access
);
434 /* We are the first to access this BO, let's initialize
435 * old_access_type to our own access type in that case.
437 old_access_type
= access_type
;
438 access
->type
= access_type
;
443 if (access_type
== PAN_BO_ACCESS_WRITE
&&
444 old_access_type
== PAN_BO_ACCESS_READ
) {
445 /* Previous access was a read and we want to write this BO.
446 * We first need to add explicit deps between our batch and
447 * the previous readers.
449 util_dynarray_foreach(&access
->readers
,
450 struct panfrost_batch_fence
*, reader
) {
451 /* We were already reading the BO, no need to add a dep
452 * on ourself (the acyclic check would complain about
455 if (!(*reader
) || (*reader
)->batch
== batch
)
458 panfrost_batch_add_dep(batch
, *reader
);
460 panfrost_batch_fence_reference(batch
->out_sync
);
462 /* We now are the new writer. */
463 access
->writer
= batch
->out_sync
;
464 access
->type
= access_type
;
466 /* Release the previous readers and reset the readers array. */
467 util_dynarray_foreach(&access
->readers
,
468 struct panfrost_batch_fence
*,
472 panfrost_batch_fence_unreference(*reader
);
475 util_dynarray_clear(&access
->readers
);
476 } else if (access_type
== PAN_BO_ACCESS_WRITE
&&
477 old_access_type
== PAN_BO_ACCESS_WRITE
) {
478 /* Previous access was a write and we want to write this BO.
479 * First check if we were the previous writer, in that case
480 * there's nothing to do. Otherwise we need to add a
481 * dependency between the new writer and the old one.
483 if (access
->writer
!= batch
->out_sync
) {
484 if (access
->writer
) {
485 panfrost_batch_add_dep(batch
, access
->writer
);
486 panfrost_batch_fence_unreference(access
->writer
);
488 panfrost_batch_fence_reference(batch
->out_sync
);
489 access
->writer
= batch
->out_sync
;
491 } else if (access_type
== PAN_BO_ACCESS_READ
&&
492 old_access_type
== PAN_BO_ACCESS_WRITE
) {
493 /* Previous access was a write and we want to read this BO.
494 * First check if we were the previous writer, in that case
495 * we want to keep the access type unchanged, as a write is
496 * more constraining than a read.
498 if (access
->writer
!= batch
->out_sync
) {
499 /* Add a dependency on the previous writer. */
500 panfrost_batch_add_dep(batch
, access
->writer
);
502 /* The previous access was a write, there's no reason
503 * to have entries in the readers array.
505 assert(!util_dynarray_num_elements(&access
->readers
,
506 struct panfrost_batch_fence
*));
508 /* Add ourselves to the readers array. */
509 panfrost_batch_fence_reference(batch
->out_sync
);
510 util_dynarray_append(&access
->readers
,
511 struct panfrost_batch_fence
*,
513 access
->type
= PAN_BO_ACCESS_READ
;
516 /* We already accessed this BO before, so we should already be
517 * in the reader array.
519 if (already_accessed
) {
520 assert(panfrost_batch_in_readers(batch
, access
));
524 /* Previous access was a read and we want to read this BO.
525 * Add ourselves to the readers array and add a dependency on
526 * the previous writer if any.
528 panfrost_batch_fence_reference(batch
->out_sync
);
529 util_dynarray_append(&access
->readers
,
530 struct panfrost_batch_fence
*,
534 panfrost_batch_add_dep(batch
, access
->writer
);
539 panfrost_batch_add_bo(struct panfrost_batch
*batch
, struct panfrost_bo
*bo
,
545 struct hash_entry
*entry
;
546 uint32_t old_flags
= 0;
548 entry
= _mesa_hash_table_search(batch
->bos
, bo
);
550 entry
= _mesa_hash_table_insert(batch
->bos
, bo
,
551 (void *)(uintptr_t)flags
);
552 panfrost_bo_reference(bo
);
554 old_flags
= (uintptr_t)entry
->data
;
556 /* All batches have to agree on the shared flag. */
557 assert((old_flags
& PAN_BO_ACCESS_SHARED
) ==
558 (flags
& PAN_BO_ACCESS_SHARED
));
563 if (old_flags
== flags
)
567 entry
->data
= (void *)(uintptr_t)flags
;
569 /* If this is not a shared BO, we don't really care about dependency
572 if (!(flags
& PAN_BO_ACCESS_SHARED
))
575 /* All dependencies should have been flushed before we execute the
576 * wallpaper draw, so it should be harmless to skip the
577 * update_bo_access() call.
579 if (batch
== batch
->ctx
->wallpaper_batch
)
582 /* Only pass R/W flags to the dep tracking logic. */
583 assert(flags
& PAN_BO_ACCESS_RW
);
584 flags
= (flags
& PAN_BO_ACCESS_WRITE
) ?
585 PAN_BO_ACCESS_WRITE
: PAN_BO_ACCESS_READ
;
586 panfrost_batch_update_bo_access(batch
, bo
, flags
, old_flags
!= 0);
589 void panfrost_batch_add_fbo_bos(struct panfrost_batch
*batch
)
591 uint32_t flags
= PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_WRITE
|
592 PAN_BO_ACCESS_VERTEX_TILER
|
593 PAN_BO_ACCESS_FRAGMENT
;
595 for (unsigned i
= 0; i
< batch
->key
.nr_cbufs
; ++i
) {
596 struct panfrost_resource
*rsrc
= pan_resource(batch
->key
.cbufs
[i
]->texture
);
597 panfrost_batch_add_bo(batch
, rsrc
->bo
, flags
);
600 if (batch
->key
.zsbuf
) {
601 struct panfrost_resource
*rsrc
= pan_resource(batch
->key
.zsbuf
->texture
);
602 panfrost_batch_add_bo(batch
, rsrc
->bo
, flags
);
607 panfrost_batch_create_bo(struct panfrost_batch
*batch
, size_t size
,
608 uint32_t create_flags
, uint32_t access_flags
)
610 struct panfrost_bo
*bo
;
612 bo
= panfrost_bo_create(pan_screen(batch
->ctx
->base
.screen
), size
,
614 panfrost_batch_add_bo(batch
, bo
, access_flags
);
616 /* panfrost_batch_add_bo() has retained a reference and
617 * panfrost_bo_create() initialize the refcnt to 1, so let's
618 * unreference the BO here so it gets released when the batch is
619 * destroyed (unless it's retained by someone else in the meantime).
621 panfrost_bo_unreference(bo
);
625 /* Returns the polygon list's GPU address if available, or otherwise allocates
626 * the polygon list. It's perfectly fast to use allocate/free BO directly,
627 * since we'll hit the BO cache and this is one-per-batch anyway. */
630 panfrost_batch_get_polygon_list(struct panfrost_batch
*batch
, unsigned size
)
632 if (batch
->polygon_list
) {
633 assert(batch
->polygon_list
->size
>= size
);
635 /* Create the BO as invisible, as there's no reason to map */
636 size
= util_next_power_of_two(size
);
638 batch
->polygon_list
= panfrost_batch_create_bo(batch
, size
,
640 PAN_BO_ACCESS_PRIVATE
|
642 PAN_BO_ACCESS_VERTEX_TILER
|
643 PAN_BO_ACCESS_FRAGMENT
);
646 return batch
->polygon_list
->gpu
;
650 panfrost_batch_get_scratchpad(struct panfrost_batch
*batch
,
652 unsigned thread_tls_alloc
,
655 unsigned size
= panfrost_get_total_stack_size(shift
,
659 if (batch
->scratchpad
) {
660 assert(batch
->scratchpad
->size
>= size
);
662 batch
->scratchpad
= panfrost_batch_create_bo(batch
, size
,
664 PAN_BO_ACCESS_PRIVATE
|
666 PAN_BO_ACCESS_VERTEX_TILER
|
667 PAN_BO_ACCESS_FRAGMENT
);
670 return batch
->scratchpad
;
674 panfrost_batch_get_shared_memory(struct panfrost_batch
*batch
,
676 unsigned workgroup_count
)
678 if (batch
->shared_memory
) {
679 assert(batch
->shared_memory
->size
>= size
);
681 batch
->shared_memory
= panfrost_batch_create_bo(batch
, size
,
683 PAN_BO_ACCESS_PRIVATE
|
685 PAN_BO_ACCESS_VERTEX_TILER
);
688 return batch
->shared_memory
;
692 panfrost_batch_get_tiler_heap(struct panfrost_batch
*batch
)
694 if (batch
->tiler_heap
)
695 return batch
->tiler_heap
;
697 batch
->tiler_heap
= panfrost_batch_create_bo(batch
, 4096 * 4096,
700 PAN_BO_ACCESS_PRIVATE
|
702 PAN_BO_ACCESS_VERTEX_TILER
|
703 PAN_BO_ACCESS_FRAGMENT
);
704 assert(batch
->tiler_heap
);
705 return batch
->tiler_heap
;
709 panfrost_batch_get_tiler_dummy(struct panfrost_batch
*batch
)
711 struct panfrost_screen
*screen
= pan_screen(batch
->ctx
->base
.screen
);
713 uint32_t create_flags
= 0;
715 if (batch
->tiler_dummy
)
716 return batch
->tiler_dummy
;
718 if (!(screen
->quirks
& MIDGARD_NO_HIER_TILING
))
719 create_flags
= PAN_BO_INVISIBLE
;
721 batch
->tiler_dummy
= panfrost_batch_create_bo(batch
, 4096,
723 PAN_BO_ACCESS_PRIVATE
|
725 PAN_BO_ACCESS_VERTEX_TILER
|
726 PAN_BO_ACCESS_FRAGMENT
);
727 assert(batch
->tiler_dummy
);
728 return batch
->tiler_dummy
;
732 panfrost_batch_draw_wallpaper(struct panfrost_batch
*batch
)
734 /* Color 0 is cleared, no need to draw the wallpaper.
735 * TODO: MRT wallpapers.
737 if (batch
->clear
& PIPE_CLEAR_COLOR0
)
740 /* Nothing to reload? TODO: MRT wallpapers */
741 if (batch
->key
.cbufs
[0] == NULL
)
744 /* No draw calls, and no clear on the depth/stencil bufs.
745 * Drawing the wallpaper would be useless.
747 if (!batch
->last_tiler
.gpu
&&
748 !(batch
->clear
& PIPE_CLEAR_DEPTHSTENCIL
))
751 /* Check if the buffer has any content on it worth preserving */
753 struct pipe_surface
*surf
= batch
->key
.cbufs
[0];
754 struct panfrost_resource
*rsrc
= pan_resource(surf
->texture
);
755 unsigned level
= surf
->u
.tex
.level
;
757 if (!rsrc
->slices
[level
].initialized
)
760 batch
->ctx
->wallpaper_batch
= batch
;
762 /* Clamp the rendering area to the damage extent. The
763 * KHR_partial_update() spec states that trying to render outside of
764 * the damage region is "undefined behavior", so we should be safe.
766 unsigned damage_width
= (rsrc
->damage
.extent
.maxx
- rsrc
->damage
.extent
.minx
);
767 unsigned damage_height
= (rsrc
->damage
.extent
.maxy
- rsrc
->damage
.extent
.miny
);
769 if (damage_width
&& damage_height
) {
770 panfrost_batch_intersection_scissor(batch
,
771 rsrc
->damage
.extent
.minx
,
772 rsrc
->damage
.extent
.miny
,
773 rsrc
->damage
.extent
.maxx
,
774 rsrc
->damage
.extent
.maxy
);
777 /* FIXME: Looks like aligning on a tile is not enough, but
778 * aligning on twice the tile size seems to works. We don't
779 * know exactly what happens here but this deserves extra
780 * investigation to figure it out.
782 batch
->minx
= batch
->minx
& ~((MALI_TILE_LENGTH
* 2) - 1);
783 batch
->miny
= batch
->miny
& ~((MALI_TILE_LENGTH
* 2) - 1);
784 batch
->maxx
= MIN2(ALIGN_POT(batch
->maxx
, MALI_TILE_LENGTH
* 2),
786 batch
->maxy
= MIN2(ALIGN_POT(batch
->maxy
, MALI_TILE_LENGTH
* 2),
789 struct pipe_scissor_state damage
;
790 struct pipe_box rects
[4];
792 /* Clamp the damage box to the rendering area. */
793 damage
.minx
= MAX2(batch
->minx
, rsrc
->damage
.biggest_rect
.x
);
794 damage
.miny
= MAX2(batch
->miny
, rsrc
->damage
.biggest_rect
.y
);
795 damage
.maxx
= MIN2(batch
->maxx
,
796 rsrc
->damage
.biggest_rect
.x
+
797 rsrc
->damage
.biggest_rect
.width
);
798 damage
.maxx
= MAX2(damage
.maxx
, damage
.minx
);
799 damage
.maxy
= MIN2(batch
->maxy
,
800 rsrc
->damage
.biggest_rect
.y
+
801 rsrc
->damage
.biggest_rect
.height
);
802 damage
.maxy
= MAX2(damage
.maxy
, damage
.miny
);
804 /* One damage rectangle means we can end up with at most 4 reload
806 * 1: left region, only exists if damage.x > 0
807 * 2: right region, only exists if damage.x + damage.width < fb->width
808 * 3: top region, only exists if damage.y > 0. The intersection with
809 * the left and right regions are dropped
810 * 4: bottom region, only exists if damage.y + damage.height < fb->height.
811 * The intersection with the left and right regions are dropped
813 * ____________________________
820 * |_______|___________|______|
822 u_box_2d(batch
->minx
, batch
->miny
, damage
.minx
- batch
->minx
,
823 batch
->maxy
- batch
->miny
, &rects
[0]);
824 u_box_2d(damage
.maxx
, batch
->miny
, batch
->maxx
- damage
.maxx
,
825 batch
->maxy
- batch
->miny
, &rects
[1]);
826 u_box_2d(damage
.minx
, batch
->miny
, damage
.maxx
- damage
.minx
,
827 damage
.miny
- batch
->miny
, &rects
[2]);
828 u_box_2d(damage
.minx
, damage
.maxy
, damage
.maxx
- damage
.minx
,
829 batch
->maxy
- damage
.maxy
, &rects
[3]);
831 for (unsigned i
= 0; i
< 4; i
++) {
832 /* Width and height are always >= 0 even if width is declared as a
833 * signed integer: u_box_2d() helper takes unsigned args and
834 * panfrost_set_damage_region() is taking care of clamping
837 if (!rects
[i
].width
|| !rects
[i
].height
)
840 /* Blit the wallpaper in */
841 panfrost_blit_wallpaper(batch
->ctx
, &rects
[i
]);
843 batch
->ctx
->wallpaper_batch
= NULL
;
847 panfrost_batch_submit_ioctl(struct panfrost_batch
*batch
,
848 mali_ptr first_job_desc
,
850 struct mali_job_descriptor_header
*header
)
852 struct panfrost_context
*ctx
= batch
->ctx
;
853 struct pipe_context
*gallium
= (struct pipe_context
*) ctx
;
854 struct panfrost_screen
*screen
= pan_screen(gallium
->screen
);
855 struct drm_panfrost_submit submit
= {0,};
856 uint32_t *bo_handles
, *in_syncs
= NULL
;
857 bool is_fragment_shader
;
860 is_fragment_shader
= (reqs
& PANFROST_JD_REQ_FS
) && batch
->first_job
.gpu
;
861 if (is_fragment_shader
)
862 submit
.in_sync_count
= 1;
864 submit
.in_sync_count
= util_dynarray_num_elements(&batch
->dependencies
,
865 struct panfrost_batch_fence
*);
867 if (submit
.in_sync_count
) {
868 in_syncs
= calloc(submit
.in_sync_count
, sizeof(*in_syncs
));
872 /* The fragment job always depends on the vertex/tiler job if there's
875 if (is_fragment_shader
) {
876 in_syncs
[0] = batch
->out_sync
->syncobj
;
880 util_dynarray_foreach(&batch
->dependencies
,
881 struct panfrost_batch_fence
*, dep
)
882 in_syncs
[i
++] = (*dep
)->syncobj
;
885 submit
.in_syncs
= (uintptr_t)in_syncs
;
886 submit
.out_sync
= batch
->out_sync
->syncobj
;
887 submit
.jc
= first_job_desc
;
888 submit
.requirements
= reqs
;
890 bo_handles
= calloc(batch
->bos
->entries
, sizeof(*bo_handles
));
893 hash_table_foreach(batch
->bos
, entry
) {
894 struct panfrost_bo
*bo
= (struct panfrost_bo
*)entry
->key
;
895 uint32_t flags
= (uintptr_t)entry
->data
;
897 assert(bo
->gem_handle
> 0);
898 bo_handles
[submit
.bo_handle_count
++] = bo
->gem_handle
;
900 /* Update the BO access flags so that panfrost_bo_wait() knows
901 * about all pending accesses.
902 * We only keep the READ/WRITE info since this is all the BO
903 * wait logic cares about.
904 * We also preserve existing flags as this batch might not
905 * be the first one to access the BO.
907 bo
->gpu_access
|= flags
& (PAN_BO_ACCESS_RW
);
910 submit
.bo_handles
= (u64
) (uintptr_t) bo_handles
;
911 ret
= drmIoctl(screen
->fd
, DRM_IOCTL_PANFROST_SUBMIT
, &submit
);
916 DBG("Error submitting: %m\n");
920 if (pan_debug
& PAN_DBG_SYNC
) {
923 /* Wait so we can get errors reported back */
924 drmSyncobjWait(screen
->fd
, &batch
->out_sync
->syncobj
, 1,
927 status
= header
->exception_status
;
929 if (status
&& status
!= 0x1) {
930 DBG("Job %" PRIx64
" failed: source ID: 0x%x access: %s exception: 0x%x (exception_status 0x%x) fault_pointer 0x%" PRIx64
" \n",
932 (status
>> 16) & 0xFFFF,
933 pandecode_exception_access((status
>> 8) & 0x3),
936 header
->fault_pointer
);
940 /* Trace the job if we're doing that */
941 if (pan_debug
& PAN_DBG_TRACE
) {
942 /* Wait so we can get errors reported back */
943 drmSyncobjWait(screen
->fd
, &batch
->out_sync
->syncobj
, 1,
945 pandecode_jc(submit
.jc
, FALSE
, screen
->gpu_id
, false);
952 panfrost_batch_submit_jobs(struct panfrost_batch
*batch
)
954 bool has_draws
= batch
->first_job
.gpu
;
955 struct mali_job_descriptor_header
*header
;
959 header
= (struct mali_job_descriptor_header
*)batch
->first_job
.cpu
;
960 ret
= panfrost_batch_submit_ioctl(batch
, batch
->first_job
.gpu
, 0, header
);
964 if (batch
->first_tiler
.gpu
|| batch
->clear
) {
965 mali_ptr fragjob
= panfrost_fragment_job(batch
, has_draws
, &header
);
967 ret
= panfrost_batch_submit_ioctl(batch
, fragjob
, PANFROST_JD_REQ_FS
, header
);
975 panfrost_batch_submit(struct panfrost_batch
*batch
)
979 /* Submit the dependencies first. */
980 util_dynarray_foreach(&batch
->dependencies
,
981 struct panfrost_batch_fence
*, dep
) {
983 panfrost_batch_submit((*dep
)->batch
);
989 if (!batch
->last_job
.gpu
&& !batch
->clear
) {
990 /* Mark the fence as signaled so the fence logic does not try
993 batch
->out_sync
->signaled
= true;
997 panfrost_batch_draw_wallpaper(batch
);
999 /* Now that all draws are in, we can finally prepare the
1000 * FBD for the batch */
1002 if (batch
->framebuffer
.gpu
&& batch
->first_job
.gpu
) {
1003 struct panfrost_context
*ctx
= batch
->ctx
;
1004 struct pipe_context
*gallium
= (struct pipe_context
*) ctx
;
1005 struct panfrost_screen
*screen
= pan_screen(gallium
->screen
);
1007 if (screen
->quirks
& MIDGARD_SFBD
)
1008 panfrost_attach_sfbd(batch
, ~0);
1010 panfrost_attach_mfbd(batch
, ~0);
1013 panfrost_scoreboard_link_batch(batch
);
1015 ret
= panfrost_batch_submit_jobs(batch
);
1018 DBG("panfrost_batch_submit failed: %d\n", ret
);
1020 /* We must reset the damage info of our render targets here even
1021 * though a damage reset normally happens when the DRI layer swaps
1022 * buffers. That's because there can be implicit flushes the GL
1023 * app is not aware of, and those might impact the damage region: if
1024 * part of the damaged portion is drawn during those implicit flushes,
1025 * you have to reload those areas before next draws are pushed, and
1026 * since the driver can't easily know what's been modified by the draws
1027 * it flushed, the easiest solution is to reload everything.
1029 for (unsigned i
= 0; i
< batch
->key
.nr_cbufs
; i
++) {
1030 struct panfrost_resource
*res
;
1032 if (!batch
->key
.cbufs
[i
])
1035 res
= pan_resource(batch
->key
.cbufs
[i
]->texture
);
1036 panfrost_resource_reset_damage(res
);
1040 panfrost_freeze_batch(batch
);
1041 panfrost_free_batch(batch
);
1045 panfrost_flush_all_batches(struct panfrost_context
*ctx
, bool wait
)
1047 struct util_dynarray fences
, syncobjs
;
1050 util_dynarray_init(&fences
, NULL
);
1051 util_dynarray_init(&syncobjs
, NULL
);
1054 hash_table_foreach(ctx
->batches
, hentry
) {
1055 struct panfrost_batch
*batch
= hentry
->data
;
1060 panfrost_batch_fence_reference(batch
->out_sync
);
1061 util_dynarray_append(&fences
, struct panfrost_batch_fence
*,
1063 util_dynarray_append(&syncobjs
, uint32_t,
1064 batch
->out_sync
->syncobj
);
1067 panfrost_batch_submit(batch
);
1070 assert(!ctx
->batches
->entries
);
1072 /* Collect batch fences before returning */
1073 panfrost_gc_fences(ctx
);
1078 drmSyncobjWait(pan_screen(ctx
->base
.screen
)->fd
,
1079 util_dynarray_begin(&syncobjs
),
1080 util_dynarray_num_elements(&syncobjs
, uint32_t),
1081 INT64_MAX
, DRM_SYNCOBJ_WAIT_FLAGS_WAIT_ALL
, NULL
);
1083 util_dynarray_foreach(&fences
, struct panfrost_batch_fence
*, fence
)
1084 panfrost_batch_fence_unreference(*fence
);
1086 util_dynarray_fini(&fences
);
1087 util_dynarray_fini(&syncobjs
);
1091 panfrost_pending_batches_access_bo(struct panfrost_context
*ctx
,
1092 const struct panfrost_bo
*bo
)
1094 struct panfrost_bo_access
*access
;
1095 struct hash_entry
*hentry
;
1097 hentry
= _mesa_hash_table_search(ctx
->accessed_bos
, bo
);
1098 access
= hentry
? hentry
->data
: NULL
;
1102 if (access
->writer
&& access
->writer
->batch
)
1105 util_dynarray_foreach(&access
->readers
, struct panfrost_batch_fence
*,
1107 if (*reader
&& (*reader
)->batch
)
1115 panfrost_flush_batches_accessing_bo(struct panfrost_context
*ctx
,
1116 struct panfrost_bo
*bo
,
1117 uint32_t access_type
)
1119 struct panfrost_bo_access
*access
;
1120 struct hash_entry
*hentry
;
1122 /* It doesn't make any to flush only the readers. */
1123 assert(access_type
== PAN_BO_ACCESS_WRITE
||
1124 access_type
== PAN_BO_ACCESS_RW
);
1126 hentry
= _mesa_hash_table_search(ctx
->accessed_bos
, bo
);
1127 access
= hentry
? hentry
->data
: NULL
;
1131 if (access_type
& PAN_BO_ACCESS_WRITE
&& access
->writer
&&
1132 access
->writer
->batch
)
1133 panfrost_batch_submit(access
->writer
->batch
);
1135 if (!(access_type
& PAN_BO_ACCESS_READ
))
1138 util_dynarray_foreach(&access
->readers
, struct panfrost_batch_fence
*,
1140 if (*reader
&& (*reader
)->batch
)
1141 panfrost_batch_submit((*reader
)->batch
);
1146 panfrost_batch_set_requirements(struct panfrost_batch
*batch
)
1148 struct panfrost_context
*ctx
= batch
->ctx
;
1150 if (ctx
->rasterizer
&& ctx
->rasterizer
->base
.multisample
)
1151 batch
->requirements
|= PAN_REQ_MSAA
;
1153 if (ctx
->depth_stencil
&& ctx
->depth_stencil
->depth
.writemask
)
1154 batch
->requirements
|= PAN_REQ_DEPTH_WRITE
;
1157 /* Helper to smear a 32-bit color across 128-bit components */
1160 pan_pack_color_32(uint32_t *packed
, uint32_t v
)
1162 for (unsigned i
= 0; i
< 4; ++i
)
1167 pan_pack_color_64(uint32_t *packed
, uint32_t lo
, uint32_t hi
)
1169 for (unsigned i
= 0; i
< 4; i
+= 2) {
1176 pan_pack_color(uint32_t *packed
, const union pipe_color_union
*color
, enum pipe_format format
)
1178 /* Alpha magicked to 1.0 if there is no alpha */
1180 bool has_alpha
= util_format_has_alpha(format
);
1181 float clear_alpha
= has_alpha
? color
->f
[3] : 1.0f
;
1183 /* Packed color depends on the framebuffer format */
1185 const struct util_format_description
*desc
=
1186 util_format_description(format
);
1188 if (util_format_is_rgba8_variant(desc
)) {
1189 pan_pack_color_32(packed
,
1190 ((uint32_t) float_to_ubyte(clear_alpha
) << 24) |
1191 ((uint32_t) float_to_ubyte(color
->f
[2]) << 16) |
1192 ((uint32_t) float_to_ubyte(color
->f
[1]) << 8) |
1193 ((uint32_t) float_to_ubyte(color
->f
[0]) << 0));
1194 } else if (format
== PIPE_FORMAT_B5G6R5_UNORM
) {
1195 /* First, we convert the components to R5, G6, B5 separately */
1196 unsigned r5
= CLAMP(color
->f
[0], 0.0, 1.0) * 31.0;
1197 unsigned g6
= CLAMP(color
->f
[1], 0.0, 1.0) * 63.0;
1198 unsigned b5
= CLAMP(color
->f
[2], 0.0, 1.0) * 31.0;
1200 /* Then we pack into a sparse u32. TODO: Why these shifts? */
1201 pan_pack_color_32(packed
, (b5
<< 25) | (g6
<< 14) | (r5
<< 5));
1202 } else if (format
== PIPE_FORMAT_B4G4R4A4_UNORM
) {
1203 /* We scale the components against 0xF0 (=240.0), rather than 0xFF */
1204 unsigned r4
= CLAMP(color
->f
[0], 0.0, 1.0) * 240.0;
1205 unsigned g4
= CLAMP(color
->f
[1], 0.0, 1.0) * 240.0;
1206 unsigned b4
= CLAMP(color
->f
[2], 0.0, 1.0) * 240.0;
1207 unsigned a4
= CLAMP(clear_alpha
, 0.0, 1.0) * 240.0;
1209 /* Pack on *byte* intervals */
1210 pan_pack_color_32(packed
, (a4
<< 24) | (b4
<< 16) | (g4
<< 8) | r4
);
1211 } else if (format
== PIPE_FORMAT_B5G5R5A1_UNORM
) {
1212 /* Scale as expected but shift oddly */
1213 unsigned r5
= round(CLAMP(color
->f
[0], 0.0, 1.0)) * 31.0;
1214 unsigned g5
= round(CLAMP(color
->f
[1], 0.0, 1.0)) * 31.0;
1215 unsigned b5
= round(CLAMP(color
->f
[2], 0.0, 1.0)) * 31.0;
1216 unsigned a1
= round(CLAMP(clear_alpha
, 0.0, 1.0)) * 1.0;
1218 pan_pack_color_32(packed
, (a1
<< 31) | (b5
<< 25) | (g5
<< 15) | (r5
<< 5));
1220 /* Try Gallium's generic default path. Doesn't work for all
1221 * formats but it's a good guess. */
1223 union util_color out
;
1225 if (util_format_is_pure_integer(format
)) {
1226 memcpy(out
.ui
, color
->ui
, 16);
1228 util_pack_color(color
->f
, format
, &out
);
1231 unsigned size
= util_format_get_blocksize(format
);
1234 unsigned b
= out
.ui
[0];
1235 unsigned s
= b
| (b
<< 8);
1236 pan_pack_color_32(packed
, s
| (s
<< 16));
1237 } else if (size
== 2)
1238 pan_pack_color_32(packed
, out
.ui
[0] | (out
.ui
[0] << 16));
1239 else if (size
== 3 || size
== 4)
1240 pan_pack_color_32(packed
, out
.ui
[0]);
1242 pan_pack_color_64(packed
, out
.ui
[0], out
.ui
[1] | (out
.ui
[1] << 16)); /* RGB16F -- RGBB */
1244 pan_pack_color_64(packed
, out
.ui
[0], out
.ui
[1]);
1245 else if (size
== 16)
1246 memcpy(packed
, out
.ui
, 16);
1248 unreachable("Unknown generic format size packing clear colour");
1253 panfrost_batch_clear(struct panfrost_batch
*batch
,
1255 const union pipe_color_union
*color
,
1256 double depth
, unsigned stencil
)
1258 struct panfrost_context
*ctx
= batch
->ctx
;
1260 if (buffers
& PIPE_CLEAR_COLOR
) {
1261 for (unsigned i
= 0; i
< PIPE_MAX_COLOR_BUFS
; ++i
) {
1262 if (!(buffers
& (PIPE_CLEAR_COLOR0
<< i
)))
1265 enum pipe_format format
= ctx
->pipe_framebuffer
.cbufs
[i
]->format
;
1266 pan_pack_color(batch
->clear_color
[i
], color
, format
);
1270 if (buffers
& PIPE_CLEAR_DEPTH
) {
1271 batch
->clear_depth
= depth
;
1274 if (buffers
& PIPE_CLEAR_STENCIL
) {
1275 batch
->clear_stencil
= stencil
;
1278 batch
->clear
|= buffers
;
1280 /* Clearing affects the entire framebuffer (by definition -- this is
1281 * the Gallium clear callback, which clears the whole framebuffer. If
1282 * the scissor test were enabled from the GL side, the state tracker
1283 * would emit a quad instead and we wouldn't go down this code path) */
1285 panfrost_batch_union_scissor(batch
, 0, 0,
1286 ctx
->pipe_framebuffer
.width
,
1287 ctx
->pipe_framebuffer
.height
);
1291 panfrost_batch_compare(const void *a
, const void *b
)
1293 return util_framebuffer_state_equal(a
, b
);
1297 panfrost_batch_hash(const void *key
)
1299 return _mesa_hash_data(key
, sizeof(struct pipe_framebuffer_state
));
1302 /* Given a new bounding rectangle (scissor), let the job cover the union of the
1303 * new and old bounding rectangles */
1306 panfrost_batch_union_scissor(struct panfrost_batch
*batch
,
1307 unsigned minx
, unsigned miny
,
1308 unsigned maxx
, unsigned maxy
)
1310 batch
->minx
= MIN2(batch
->minx
, minx
);
1311 batch
->miny
= MIN2(batch
->miny
, miny
);
1312 batch
->maxx
= MAX2(batch
->maxx
, maxx
);
1313 batch
->maxy
= MAX2(batch
->maxy
, maxy
);
1317 panfrost_batch_intersection_scissor(struct panfrost_batch
*batch
,
1318 unsigned minx
, unsigned miny
,
1319 unsigned maxx
, unsigned maxy
)
1321 batch
->minx
= MAX2(batch
->minx
, minx
);
1322 batch
->miny
= MAX2(batch
->miny
, miny
);
1323 batch
->maxx
= MIN2(batch
->maxx
, maxx
);
1324 batch
->maxy
= MIN2(batch
->maxy
, maxy
);
1327 /* Are we currently rendering to the screen (rather than an FBO)? */
1330 panfrost_batch_is_scanout(struct panfrost_batch
*batch
)
1332 /* If there is no color buffer, it's an FBO */
1333 if (batch
->key
.nr_cbufs
!= 1)
1336 /* If we're too early that no framebuffer was sent, it's scanout */
1337 if (!batch
->key
.cbufs
[0])
1340 return batch
->key
.cbufs
[0]->texture
->bind
& PIPE_BIND_DISPLAY_TARGET
||
1341 batch
->key
.cbufs
[0]->texture
->bind
& PIPE_BIND_SCANOUT
||
1342 batch
->key
.cbufs
[0]->texture
->bind
& PIPE_BIND_SHARED
;
1346 panfrost_batch_init(struct panfrost_context
*ctx
)
1348 ctx
->batches
= _mesa_hash_table_create(ctx
,
1349 panfrost_batch_hash
,
1350 panfrost_batch_compare
);
1351 ctx
->accessed_bos
= _mesa_hash_table_create(ctx
, _mesa_hash_pointer
,
1352 _mesa_key_pointer_equal
);