2 * Copyright (C) 2012 Rob Clark <robclark@freedesktop.org>
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
24 * Rob Clark <robclark@freedesktop.org>
27 #include "util/u_format.h"
28 #include "util/u_format_rgtc.h"
29 #include "util/u_format_zs.h"
30 #include "util/u_inlines.h"
31 #include "util/u_transfer.h"
32 #include "util/u_string.h"
33 #include "util/u_surface.h"
36 #include "freedreno_resource.h"
37 #include "freedreno_batch_cache.h"
38 #include "freedreno_blitter.h"
39 #include "freedreno_fence.h"
40 #include "freedreno_screen.h"
41 #include "freedreno_surface.h"
42 #include "freedreno_context.h"
43 #include "freedreno_query_hw.h"
44 #include "freedreno_util.h"
46 #include <drm_fourcc.h>
49 /* XXX this should go away, needed for 'struct winsys_handle' */
50 #include "state_tracker/drm_driver.h"
53 * Go through the entire state and see if the resource is bound
54 * anywhere. If it is, mark the relevant state as dirty. This is
55 * called on realloc_bo to ensure the neccessary state is re-
56 * emitted so the GPU looks at the new backing bo.
59 rebind_resource(struct fd_context
*ctx
, struct pipe_resource
*prsc
)
62 for (unsigned i
= 0; i
< ctx
->vtx
.vertexbuf
.count
&& !(ctx
->dirty
& FD_DIRTY_VTXBUF
); i
++) {
63 if (ctx
->vtx
.vertexbuf
.vb
[i
].buffer
.resource
== prsc
)
64 ctx
->dirty
|= FD_DIRTY_VTXBUF
;
67 /* per-shader-stage resources: */
68 for (unsigned stage
= 0; stage
< PIPE_SHADER_TYPES
; stage
++) {
69 /* Constbufs.. note that constbuf[0] is normal uniforms emitted in
70 * cmdstream rather than by pointer..
72 const unsigned num_ubos
= util_last_bit(ctx
->constbuf
[stage
].enabled_mask
);
73 for (unsigned i
= 1; i
< num_ubos
; i
++) {
74 if (ctx
->dirty_shader
[stage
] & FD_DIRTY_SHADER_CONST
)
76 if (ctx
->constbuf
[stage
].cb
[i
].buffer
== prsc
)
77 ctx
->dirty_shader
[stage
] |= FD_DIRTY_SHADER_CONST
;
81 for (unsigned i
= 0; i
< ctx
->tex
[stage
].num_textures
; i
++) {
82 if (ctx
->dirty_shader
[stage
] & FD_DIRTY_SHADER_TEX
)
84 if (ctx
->tex
[stage
].textures
[i
] && (ctx
->tex
[stage
].textures
[i
]->texture
== prsc
))
85 ctx
->dirty_shader
[stage
] |= FD_DIRTY_SHADER_TEX
;
89 const unsigned num_ssbos
= util_last_bit(ctx
->shaderbuf
[stage
].enabled_mask
);
90 for (unsigned i
= 0; i
< num_ssbos
; i
++) {
91 if (ctx
->dirty_shader
[stage
] & FD_DIRTY_SHADER_SSBO
)
93 if (ctx
->shaderbuf
[stage
].sb
[i
].buffer
== prsc
)
94 ctx
->dirty_shader
[stage
] |= FD_DIRTY_SHADER_SSBO
;
100 realloc_bo(struct fd_resource
*rsc
, uint32_t size
)
102 struct pipe_resource
*prsc
= &rsc
->base
;
103 struct fd_screen
*screen
= fd_screen(rsc
->base
.screen
);
104 uint32_t flags
= DRM_FREEDRENO_GEM_CACHE_WCOMBINE
|
105 DRM_FREEDRENO_GEM_TYPE_KMEM
|
106 COND(prsc
->bind
& PIPE_BIND_SCANOUT
, DRM_FREEDRENO_GEM_SCANOUT
);
107 /* TODO other flags? */
109 /* if we start using things other than write-combine,
110 * be sure to check for PIPE_RESOURCE_FLAG_MAP_COHERENT
116 rsc
->bo
= fd_bo_new(screen
->dev
, size
, flags
, "%ux%ux%u@%u:%x",
117 prsc
->width0
, prsc
->height0
, prsc
->depth0
, rsc
->cpp
, prsc
->bind
);
118 rsc
->seqno
= p_atomic_inc_return(&screen
->rsc_seqno
);
119 util_range_set_empty(&rsc
->valid_buffer_range
);
120 fd_bc_invalidate_resource(rsc
, true);
124 do_blit(struct fd_context
*ctx
, const struct pipe_blit_info
*blit
, bool fallback
)
126 struct pipe_context
*pctx
= &ctx
->base
;
128 /* TODO size threshold too?? */
130 /* do blit on gpu: */
131 pctx
->blit(pctx
, blit
);
133 /* do blit on cpu: */
134 util_resource_copy_region(pctx
,
135 blit
->dst
.resource
, blit
->dst
.level
, blit
->dst
.box
.x
,
136 blit
->dst
.box
.y
, blit
->dst
.box
.z
,
137 blit
->src
.resource
, blit
->src
.level
, &blit
->src
.box
);
142 fd_try_shadow_resource(struct fd_context
*ctx
, struct fd_resource
*rsc
,
143 unsigned level
, const struct pipe_box
*box
)
145 struct pipe_context
*pctx
= &ctx
->base
;
146 struct pipe_resource
*prsc
= &rsc
->base
;
147 bool fallback
= false;
152 /* TODO: somehow munge dimensions and format to copy unsupported
153 * render target format to something that is supported?
155 if (!pctx
->screen
->is_format_supported(pctx
->screen
,
156 prsc
->format
, prsc
->target
, prsc
->nr_samples
,
157 prsc
->nr_storage_samples
,
158 PIPE_BIND_RENDER_TARGET
))
161 /* do shadowing back-blits on the cpu for buffers: */
162 if (prsc
->target
== PIPE_BUFFER
)
165 bool whole_level
= util_texrange_covers_whole_level(prsc
, level
,
166 box
->x
, box
->y
, box
->z
, box
->width
, box
->height
, box
->depth
);
168 /* TODO need to be more clever about current level */
169 if ((prsc
->target
>= PIPE_TEXTURE_2D
) && !whole_level
)
172 struct pipe_resource
*pshadow
=
173 pctx
->screen
->resource_create(pctx
->screen
, prsc
);
178 assert(!ctx
->in_shadow
);
179 ctx
->in_shadow
= true;
181 /* get rid of any references that batch-cache might have to us (which
182 * should empty/destroy rsc->batches hashset)
184 fd_bc_invalidate_resource(rsc
, false);
186 mtx_lock(&ctx
->screen
->lock
);
188 /* Swap the backing bo's, so shadow becomes the old buffer,
189 * blit from shadow to new buffer. From here on out, we
192 * Note that we need to do it in this order, otherwise if
193 * we go down cpu blit path, the recursive transfer_map()
194 * sees the wrong status..
196 struct fd_resource
*shadow
= fd_resource(pshadow
);
198 DBG("shadow: %p (%d) -> %p (%d)\n", rsc
, rsc
->base
.reference
.count
,
199 shadow
, shadow
->base
.reference
.count
);
201 /* TODO valid_buffer_range?? */
202 swap(rsc
->bo
, shadow
->bo
);
203 swap(rsc
->write_batch
, shadow
->write_batch
);
204 rsc
->seqno
= p_atomic_inc_return(&ctx
->screen
->rsc_seqno
);
206 /* at this point, the newly created shadow buffer is not referenced
207 * by any batches, but the existing rsc (probably) is. We need to
208 * transfer those references over:
210 debug_assert(shadow
->batch_mask
== 0);
211 struct fd_batch
*batch
;
212 foreach_batch(batch
, &ctx
->screen
->batch_cache
, rsc
->batch_mask
) {
213 struct set_entry
*entry
= _mesa_set_search(batch
->resources
, rsc
);
214 _mesa_set_remove(batch
->resources
, entry
);
215 _mesa_set_add(batch
->resources
, shadow
);
217 swap(rsc
->batch_mask
, shadow
->batch_mask
);
219 mtx_unlock(&ctx
->screen
->lock
);
221 struct pipe_blit_info blit
= {};
222 blit
.dst
.resource
= prsc
;
223 blit
.dst
.format
= prsc
->format
;
224 blit
.src
.resource
= pshadow
;
225 blit
.src
.format
= pshadow
->format
;
226 blit
.mask
= util_format_get_mask(prsc
->format
);
227 blit
.filter
= PIPE_TEX_FILTER_NEAREST
;
229 #define set_box(field, val) do { \
230 blit.dst.field = (val); \
231 blit.src.field = (val); \
234 /* blit the other levels in their entirety: */
235 for (unsigned l
= 0; l
<= prsc
->last_level
; l
++) {
239 /* just blit whole level: */
241 set_box(box
.width
, u_minify(prsc
->width0
, l
));
242 set_box(box
.height
, u_minify(prsc
->height0
, l
));
243 set_box(box
.depth
, u_minify(prsc
->depth0
, l
));
245 do_blit(ctx
, &blit
, fallback
);
248 /* deal w/ current level specially, since we might need to split
249 * it up into a couple blits:
252 set_box(level
, level
);
254 switch (prsc
->target
) {
256 case PIPE_TEXTURE_1D
:
259 set_box(box
.height
, 1);
260 set_box(box
.depth
, 1);
264 set_box(box
.width
, box
->x
);
266 do_blit(ctx
, &blit
, fallback
);
268 if ((box
->x
+ box
->width
) < u_minify(prsc
->width0
, level
)) {
269 set_box(box
.x
, box
->x
+ box
->width
);
270 set_box(box
.width
, u_minify(prsc
->width0
, level
) - (box
->x
+ box
->width
));
272 do_blit(ctx
, &blit
, fallback
);
275 case PIPE_TEXTURE_2D
:
282 ctx
->in_shadow
= false;
284 pipe_resource_reference(&pshadow
, NULL
);
289 static struct fd_resource
*
290 fd_alloc_staging(struct fd_context
*ctx
, struct fd_resource
*rsc
,
291 unsigned level
, const struct pipe_box
*box
)
293 struct pipe_context
*pctx
= &ctx
->base
;
294 struct pipe_resource tmpl
= rsc
->base
;
296 tmpl
.width0
= box
->width
;
297 tmpl
.height0
= box
->height
;
298 /* for array textures, box->depth is the array_size, otherwise
299 * for 3d textures, it is the depth:
301 if (tmpl
.array_size
> 1) {
302 tmpl
.array_size
= box
->depth
;
306 tmpl
.depth0
= box
->depth
;
309 tmpl
.bind
|= PIPE_BIND_LINEAR
;
311 struct pipe_resource
*pstaging
=
312 pctx
->screen
->resource_create(pctx
->screen
, &tmpl
);
316 return fd_resource(pstaging
);
320 fd_blit_from_staging(struct fd_context
*ctx
, struct fd_transfer
*trans
)
322 struct pipe_resource
*dst
= trans
->base
.resource
;
323 struct pipe_blit_info blit
= {};
325 blit
.dst
.resource
= dst
;
326 blit
.dst
.format
= dst
->format
;
327 blit
.dst
.level
= trans
->base
.level
;
328 blit
.dst
.box
= trans
->base
.box
;
329 blit
.src
.resource
= trans
->staging_prsc
;
330 blit
.src
.format
= trans
->staging_prsc
->format
;
332 blit
.src
.box
= trans
->staging_box
;
333 blit
.mask
= util_format_get_mask(trans
->staging_prsc
->format
);
334 blit
.filter
= PIPE_TEX_FILTER_NEAREST
;
336 do_blit(ctx
, &blit
, false);
340 fd_blit_to_staging(struct fd_context
*ctx
, struct fd_transfer
*trans
)
342 struct pipe_resource
*src
= trans
->base
.resource
;
343 struct pipe_blit_info blit
= {};
345 blit
.src
.resource
= src
;
346 blit
.src
.format
= src
->format
;
347 blit
.src
.level
= trans
->base
.level
;
348 blit
.src
.box
= trans
->base
.box
;
349 blit
.dst
.resource
= trans
->staging_prsc
;
350 blit
.dst
.format
= trans
->staging_prsc
->format
;
352 blit
.dst
.box
= trans
->staging_box
;
353 blit
.mask
= util_format_get_mask(trans
->staging_prsc
->format
);
354 blit
.filter
= PIPE_TEX_FILTER_NEAREST
;
356 do_blit(ctx
, &blit
, false);
359 static void fd_resource_transfer_flush_region(struct pipe_context
*pctx
,
360 struct pipe_transfer
*ptrans
,
361 const struct pipe_box
*box
)
363 struct fd_resource
*rsc
= fd_resource(ptrans
->resource
);
365 if (ptrans
->resource
->target
== PIPE_BUFFER
)
366 util_range_add(&rsc
->valid_buffer_range
,
367 ptrans
->box
.x
+ box
->x
,
368 ptrans
->box
.x
+ box
->x
+ box
->width
);
372 flush_resource(struct fd_context
*ctx
, struct fd_resource
*rsc
, unsigned usage
)
374 struct fd_batch
*write_batch
= NULL
;
376 fd_batch_reference(&write_batch
, rsc
->write_batch
);
378 if (usage
& PIPE_TRANSFER_WRITE
) {
379 struct fd_batch
*batch
, *batches
[32] = {};
382 /* This is a bit awkward, probably a fd_batch_flush_locked()
383 * would make things simpler.. but we need to hold the lock
384 * to iterate the batches which reference this resource. So
385 * we must first grab references under a lock, then flush.
387 mtx_lock(&ctx
->screen
->lock
);
388 batch_mask
= rsc
->batch_mask
;
389 foreach_batch(batch
, &ctx
->screen
->batch_cache
, batch_mask
)
390 fd_batch_reference(&batches
[batch
->idx
], batch
);
391 mtx_unlock(&ctx
->screen
->lock
);
393 foreach_batch(batch
, &ctx
->screen
->batch_cache
, batch_mask
)
394 fd_batch_flush(batch
, false, false);
396 foreach_batch(batch
, &ctx
->screen
->batch_cache
, batch_mask
) {
397 fd_batch_sync(batch
);
398 fd_batch_reference(&batches
[batch
->idx
], NULL
);
400 assert(rsc
->batch_mask
== 0);
401 } else if (write_batch
) {
402 fd_batch_flush(write_batch
, true, false);
405 fd_batch_reference(&write_batch
, NULL
);
407 assert(!rsc
->write_batch
);
411 fd_flush_resource(struct pipe_context
*pctx
, struct pipe_resource
*prsc
)
413 flush_resource(fd_context(pctx
), fd_resource(prsc
), PIPE_TRANSFER_READ
);
417 fd_resource_transfer_unmap(struct pipe_context
*pctx
,
418 struct pipe_transfer
*ptrans
)
420 struct fd_context
*ctx
= fd_context(pctx
);
421 struct fd_resource
*rsc
= fd_resource(ptrans
->resource
);
422 struct fd_transfer
*trans
= fd_transfer(ptrans
);
424 if (trans
->staging_prsc
) {
425 if (ptrans
->usage
& PIPE_TRANSFER_WRITE
)
426 fd_blit_from_staging(ctx
, trans
);
427 pipe_resource_reference(&trans
->staging_prsc
, NULL
);
430 if (!(ptrans
->usage
& PIPE_TRANSFER_UNSYNCHRONIZED
)) {
431 fd_bo_cpu_fini(rsc
->bo
);
434 util_range_add(&rsc
->valid_buffer_range
,
436 ptrans
->box
.x
+ ptrans
->box
.width
);
438 pipe_resource_reference(&ptrans
->resource
, NULL
);
439 slab_free(&ctx
->transfer_pool
, ptrans
);
443 fd_resource_transfer_map(struct pipe_context
*pctx
,
444 struct pipe_resource
*prsc
,
445 unsigned level
, unsigned usage
,
446 const struct pipe_box
*box
,
447 struct pipe_transfer
**pptrans
)
449 struct fd_context
*ctx
= fd_context(pctx
);
450 struct fd_resource
*rsc
= fd_resource(prsc
);
451 struct fd_resource_slice
*slice
= fd_resource_slice(rsc
, level
);
452 struct fd_transfer
*trans
;
453 struct pipe_transfer
*ptrans
;
454 enum pipe_format format
= prsc
->format
;
460 DBG("prsc=%p, level=%u, usage=%x, box=%dx%d+%d,%d", prsc
, level
, usage
,
461 box
->width
, box
->height
, box
->x
, box
->y
);
463 ptrans
= slab_alloc(&ctx
->transfer_pool
);
467 /* slab_alloc_st() doesn't zero: */
468 trans
= fd_transfer(ptrans
);
469 memset(trans
, 0, sizeof(*trans
));
471 pipe_resource_reference(&ptrans
->resource
, prsc
);
472 ptrans
->level
= level
;
473 ptrans
->usage
= usage
;
475 ptrans
->stride
= util_format_get_nblocksx(format
, slice
->pitch
) * rsc
->cpp
;
476 ptrans
->layer_stride
= rsc
->layer_first
? rsc
->layer_size
: slice
->size0
;
478 /* we always need a staging texture for tiled buffers:
480 * TODO we might sometimes want to *also* shadow the resource to avoid
481 * splitting a batch.. for ex, mid-frame texture uploads to a tiled
484 if (rsc
->tile_mode
) {
485 struct fd_resource
*staging_rsc
;
487 staging_rsc
= fd_alloc_staging(ctx
, rsc
, level
, box
);
489 // TODO for PIPE_TRANSFER_READ, need to do untiling blit..
490 trans
->staging_prsc
= &staging_rsc
->base
;
491 trans
->base
.stride
= util_format_get_nblocksx(format
,
492 staging_rsc
->slices
[0].pitch
) * staging_rsc
->cpp
;
493 trans
->base
.layer_stride
= staging_rsc
->layer_first
?
494 staging_rsc
->layer_size
: staging_rsc
->slices
[0].size0
;
495 trans
->staging_box
= *box
;
496 trans
->staging_box
.x
= 0;
497 trans
->staging_box
.y
= 0;
498 trans
->staging_box
.z
= 0;
500 if (usage
& PIPE_TRANSFER_READ
) {
501 fd_blit_to_staging(ctx
, trans
);
503 struct fd_batch
*batch
= NULL
;
504 fd_batch_reference(&batch
, staging_rsc
->write_batch
);
506 /* we can't fd_bo_cpu_prep() until the blit to staging
507 * is submitted to kernel.. in that case write_batch
508 * wouldn't be NULL yet:
511 fd_batch_sync(batch
);
512 fd_batch_reference(&batch
, NULL
);
515 fd_bo_cpu_prep(staging_rsc
->bo
, ctx
->pipe
,
516 DRM_FREEDRENO_PREP_READ
);
519 buf
= fd_bo_map(staging_rsc
->bo
);
524 ctx
->stats
.staging_uploads
++;
530 if (ctx
->in_shadow
&& !(usage
& PIPE_TRANSFER_READ
))
531 usage
|= PIPE_TRANSFER_UNSYNCHRONIZED
;
533 if (usage
& PIPE_TRANSFER_READ
)
534 op
|= DRM_FREEDRENO_PREP_READ
;
536 if (usage
& PIPE_TRANSFER_WRITE
)
537 op
|= DRM_FREEDRENO_PREP_WRITE
;
539 if (usage
& PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE
) {
540 realloc_bo(rsc
, fd_bo_size(rsc
->bo
));
541 rebind_resource(ctx
, prsc
);
542 } else if ((usage
& PIPE_TRANSFER_WRITE
) &&
543 prsc
->target
== PIPE_BUFFER
&&
544 !util_ranges_intersect(&rsc
->valid_buffer_range
,
545 box
->x
, box
->x
+ box
->width
)) {
546 /* We are trying to write to a previously uninitialized range. No need
549 } else if (!(usage
& PIPE_TRANSFER_UNSYNCHRONIZED
)) {
550 struct fd_batch
*write_batch
= NULL
;
552 /* hold a reference, so it doesn't disappear under us: */
553 fd_batch_reference(&write_batch
, rsc
->write_batch
);
555 if ((usage
& PIPE_TRANSFER_WRITE
) && write_batch
&&
556 write_batch
->back_blit
) {
557 /* if only thing pending is a back-blit, we can discard it: */
558 fd_batch_reset(write_batch
);
561 /* If the GPU is writing to the resource, or if it is reading from the
562 * resource and we're trying to write to it, flush the renders.
564 bool needs_flush
= pending(rsc
, !!(usage
& PIPE_TRANSFER_WRITE
));
565 bool busy
= needs_flush
|| (0 != fd_bo_cpu_prep(rsc
->bo
,
566 ctx
->pipe
, op
| DRM_FREEDRENO_PREP_NOSYNC
));
568 /* if we need to flush/stall, see if we can make a shadow buffer
571 * TODO we could go down this path !reorder && !busy_for_read
572 * ie. we only *don't* want to go down this path if the blit
573 * will trigger a flush!
575 if (ctx
->screen
->reorder
&& busy
&& !(usage
& PIPE_TRANSFER_READ
) &&
576 (usage
& PIPE_TRANSFER_DISCARD_RANGE
)) {
577 /* try shadowing only if it avoids a flush, otherwise staging would
580 if (needs_flush
&& fd_try_shadow_resource(ctx
, rsc
, level
, box
)) {
581 needs_flush
= busy
= false;
582 rebind_resource(ctx
, prsc
);
583 ctx
->stats
.shadow_uploads
++;
585 struct fd_resource
*staging_rsc
;
588 flush_resource(ctx
, rsc
, usage
);
592 /* in this case, we don't need to shadow the whole resource,
593 * since any draw that references the previous contents has
594 * already had rendering flushed for all tiles. So we can
595 * use a staging buffer to do the upload.
597 staging_rsc
= fd_alloc_staging(ctx
, rsc
, level
, box
);
599 trans
->staging_prsc
= &staging_rsc
->base
;
600 trans
->base
.stride
= util_format_get_nblocksx(format
,
601 staging_rsc
->slices
[0].pitch
) * staging_rsc
->cpp
;
602 trans
->base
.layer_stride
= staging_rsc
->layer_first
?
603 staging_rsc
->layer_size
: staging_rsc
->slices
[0].size0
;
604 trans
->staging_box
= *box
;
605 trans
->staging_box
.x
= 0;
606 trans
->staging_box
.y
= 0;
607 trans
->staging_box
.z
= 0;
608 buf
= fd_bo_map(staging_rsc
->bo
);
613 fd_batch_reference(&write_batch
, NULL
);
615 ctx
->stats
.staging_uploads
++;
623 flush_resource(ctx
, rsc
, usage
);
627 fd_batch_reference(&write_batch
, NULL
);
629 /* The GPU keeps track of how the various bo's are being used, and
630 * will wait if necessary for the proper operation to have
634 ret
= fd_bo_cpu_prep(rsc
->bo
, ctx
->pipe
, op
);
640 buf
= fd_bo_map(rsc
->bo
);
642 box
->y
/ util_format_get_blockheight(format
) * ptrans
->stride
+
643 box
->x
/ util_format_get_blockwidth(format
) * rsc
->cpp
+
644 fd_resource_offset(rsc
, level
, box
->z
);
646 if (usage
& PIPE_TRANSFER_WRITE
)
654 fd_resource_transfer_unmap(pctx
, ptrans
);
659 fd_resource_destroy(struct pipe_screen
*pscreen
,
660 struct pipe_resource
*prsc
)
662 struct fd_resource
*rsc
= fd_resource(prsc
);
663 fd_bc_invalidate_resource(rsc
, true);
666 util_range_destroy(&rsc
->valid_buffer_range
);
671 fd_resource_get_handle(struct pipe_screen
*pscreen
,
672 struct pipe_context
*pctx
,
673 struct pipe_resource
*prsc
,
674 struct winsys_handle
*handle
,
677 struct fd_resource
*rsc
= fd_resource(prsc
);
679 return fd_screen_bo_get_handle(pscreen
, rsc
->bo
,
680 rsc
->slices
[0].pitch
* rsc
->cpp
, handle
);
684 setup_slices(struct fd_resource
*rsc
, uint32_t alignment
, enum pipe_format format
)
686 struct pipe_resource
*prsc
= &rsc
->base
;
687 struct fd_screen
*screen
= fd_screen(prsc
->screen
);
688 enum util_format_layout layout
= util_format_description(format
)->layout
;
689 uint32_t pitchalign
= screen
->gmem_alignw
;
690 uint32_t level
, size
= 0;
691 uint32_t width
= prsc
->width0
;
692 uint32_t height
= prsc
->height0
;
693 uint32_t depth
= prsc
->depth0
;
694 /* in layer_first layout, the level (slice) contains just one
695 * layer (since in fact the layer contains the slices)
697 uint32_t layers_in_level
= rsc
->layer_first
? 1 : prsc
->array_size
;
699 for (level
= 0; level
<= prsc
->last_level
; level
++) {
700 struct fd_resource_slice
*slice
= fd_resource_slice(rsc
, level
);
703 if (layout
== UTIL_FORMAT_LAYOUT_ASTC
)
704 slice
->pitch
= width
=
705 util_align_npot(width
, pitchalign
* util_format_get_blockwidth(format
));
707 slice
->pitch
= width
= align(width
, pitchalign
);
708 slice
->offset
= size
;
709 blocks
= util_format_get_nblocks(format
, width
, height
);
710 /* 1d array and 2d array textures must all have the same layer size
711 * for each miplevel on a3xx. 3d textures can have different layer
712 * sizes for high levels, but the hw auto-sizer is buggy (or at least
713 * different than what this code does), so as soon as the layer size
714 * range gets into range, we stop reducing it.
716 if (prsc
->target
== PIPE_TEXTURE_3D
&& (
718 (level
> 1 && rsc
->slices
[level
- 1].size0
> 0xf000)))
719 slice
->size0
= align(blocks
* rsc
->cpp
, alignment
);
720 else if (level
== 0 || rsc
->layer_first
|| alignment
== 1)
721 slice
->size0
= align(blocks
* rsc
->cpp
, alignment
);
723 slice
->size0
= rsc
->slices
[level
- 1].size0
;
725 size
+= slice
->size0
* depth
* layers_in_level
;
727 width
= u_minify(width
, 1);
728 height
= u_minify(height
, 1);
729 depth
= u_minify(depth
, 1);
736 slice_alignment(enum pipe_texture_target target
)
738 /* on a3xx, 2d array and 3d textures seem to want their
739 * layers aligned to page boundaries:
742 case PIPE_TEXTURE_3D
:
743 case PIPE_TEXTURE_1D_ARRAY
:
744 case PIPE_TEXTURE_2D_ARRAY
:
751 /* cross generation texture layout to plug in to screen->setup_slices()..
752 * replace with generation specific one as-needed.
754 * TODO for a4xx probably can extract out the a4xx specific logic int
755 * a small fd4_setup_slices() wrapper that sets up layer_first, and then
759 fd_setup_slices(struct fd_resource
*rsc
)
763 alignment
= slice_alignment(rsc
->base
.target
);
765 struct fd_screen
*screen
= fd_screen(rsc
->base
.screen
);
766 if (is_a4xx(screen
)) {
767 switch (rsc
->base
.target
) {
768 case PIPE_TEXTURE_3D
:
769 rsc
->layer_first
= false;
772 rsc
->layer_first
= true;
778 return setup_slices(rsc
, alignment
, rsc
->base
.format
);
781 /* special case to resize query buf after allocated.. */
783 fd_resource_resize(struct pipe_resource
*prsc
, uint32_t sz
)
785 struct fd_resource
*rsc
= fd_resource(prsc
);
787 debug_assert(prsc
->width0
== 0);
788 debug_assert(prsc
->target
== PIPE_BUFFER
);
789 debug_assert(prsc
->bind
== PIPE_BIND_QUERY_BUFFER
);
792 realloc_bo(rsc
, fd_screen(prsc
->screen
)->setup_slices(rsc
));
795 // TODO common helper?
797 has_depth(enum pipe_format format
)
800 case PIPE_FORMAT_Z16_UNORM
:
801 case PIPE_FORMAT_Z32_UNORM
:
802 case PIPE_FORMAT_Z32_FLOAT
:
803 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
:
804 case PIPE_FORMAT_Z24_UNORM_S8_UINT
:
805 case PIPE_FORMAT_S8_UINT_Z24_UNORM
:
806 case PIPE_FORMAT_Z24X8_UNORM
:
807 case PIPE_FORMAT_X8Z24_UNORM
:
815 find_modifier(uint64_t needle
, const uint64_t *haystack
, int count
)
819 for (i
= 0; i
< count
; i
++) {
820 if (haystack
[i
] == needle
)
828 * Create a new texture object, using the given template info.
830 static struct pipe_resource
*
831 fd_resource_create_with_modifiers(struct pipe_screen
*pscreen
,
832 const struct pipe_resource
*tmpl
,
833 const uint64_t *modifiers
, int count
)
835 struct fd_screen
*screen
= fd_screen(pscreen
);
836 struct fd_resource
*rsc
= CALLOC_STRUCT(fd_resource
);
837 struct pipe_resource
*prsc
= &rsc
->base
;
838 enum pipe_format format
= tmpl
->format
;
841 DBG("%p: target=%d, format=%s, %ux%ux%u, array_size=%u, last_level=%u, "
842 "nr_samples=%u, usage=%u, bind=%x, flags=%x", prsc
,
843 tmpl
->target
, util_format_name(format
),
844 tmpl
->width0
, tmpl
->height0
, tmpl
->depth0
,
845 tmpl
->array_size
, tmpl
->last_level
, tmpl
->nr_samples
,
846 tmpl
->usage
, tmpl
->bind
, tmpl
->flags
);
854 (PIPE_BIND_SCANOUT | \
856 PIPE_BIND_DISPLAY_TARGET)
858 bool linear
= find_modifier(DRM_FORMAT_MOD_LINEAR
, modifiers
, count
);
859 if (tmpl
->bind
& LINEAR
)
862 /* Normally, for non-shared buffers, allow buffer compression if
863 * not shared, otherwise only allow if QCOM_COMPRESSED modifier
866 * TODO we should probably also limit tiled in a similar way,
867 * except we don't have a format modifier for tiled. (We probably
870 bool allow_ubwc
= find_modifier(DRM_FORMAT_MOD_INVALID
, modifiers
, count
);
871 if (tmpl
->bind
& PIPE_BIND_SHARED
)
872 allow_ubwc
= find_modifier(DRM_FORMAT_MOD_QCOM_COMPRESSED
, modifiers
, count
);
874 if (screen
->tile_mode
&&
875 (tmpl
->target
!= PIPE_BUFFER
) &&
877 rsc
->tile_mode
= screen
->tile_mode(tmpl
);
880 pipe_reference_init(&prsc
->reference
, 1);
882 prsc
->screen
= pscreen
;
884 util_range_init(&rsc
->valid_buffer_range
);
886 rsc
->internal_format
= format
;
887 rsc
->cpp
= util_format_get_blocksize(format
);
888 prsc
->nr_samples
= MAX2(1, prsc
->nr_samples
);
889 rsc
->cpp
*= prsc
->nr_samples
;
893 // XXX probably need some extra work if we hit rsc shadowing path w/ lrz..
894 if ((is_a5xx(screen
) || is_a6xx(screen
)) &&
895 (fd_mesa_debug
& FD_DBG_LRZ
) && has_depth(format
)) {
896 const uint32_t flags
= DRM_FREEDRENO_GEM_CACHE_WCOMBINE
|
897 DRM_FREEDRENO_GEM_TYPE_KMEM
; /* TODO */
898 unsigned lrz_pitch
= align(DIV_ROUND_UP(tmpl
->width0
, 8), 64);
899 unsigned lrz_height
= DIV_ROUND_UP(tmpl
->height0
, 8);
901 /* LRZ buffer is super-sampled: */
902 switch (prsc
->nr_samples
) {
909 unsigned size
= lrz_pitch
* lrz_height
* 2;
911 size
+= 0x1000; /* for GRAS_LRZ_FAST_CLEAR_BUFFER */
913 rsc
->lrz_height
= lrz_height
;
914 rsc
->lrz_width
= lrz_pitch
;
915 rsc
->lrz_pitch
= lrz_pitch
;
916 rsc
->lrz
= fd_bo_new(screen
->dev
, size
, flags
, "lrz");
919 size
= screen
->setup_slices(rsc
);
921 if (allow_ubwc
&& screen
->fill_ubwc_buffer_sizes
&& rsc
->tile_mode
)
922 size
+= screen
->fill_ubwc_buffer_sizes(rsc
);
924 /* special case for hw-query buffer, which we need to allocate before we
928 /* note, semi-intention == instead of & */
929 debug_assert(prsc
->bind
== PIPE_BIND_QUERY_BUFFER
);
933 if (rsc
->layer_first
) {
934 rsc
->layer_size
= align(size
, 4096);
935 size
= rsc
->layer_size
* prsc
->array_size
;
938 realloc_bo(rsc
, size
);
944 fd_resource_destroy(pscreen
, prsc
);
948 static struct pipe_resource
*
949 fd_resource_create(struct pipe_screen
*pscreen
,
950 const struct pipe_resource
*tmpl
)
952 const uint64_t mod
= DRM_FORMAT_MOD_INVALID
;
953 return fd_resource_create_with_modifiers(pscreen
, tmpl
, &mod
, 1);
957 is_supported_modifier(struct pipe_screen
*pscreen
, enum pipe_format pfmt
,
962 /* Get the count of supported modifiers: */
963 pscreen
->query_dmabuf_modifiers(pscreen
, pfmt
, 0, NULL
, NULL
, &count
);
965 /* Get the supported modifiers: */
966 uint64_t modifiers
[count
];
967 pscreen
->query_dmabuf_modifiers(pscreen
, pfmt
, 0, modifiers
, NULL
, &count
);
969 for (int i
= 0; i
< count
; i
++)
970 if (modifiers
[i
] == mod
)
977 * Create a texture from a winsys_handle. The handle is often created in
978 * another process by first creating a pipe texture and then calling
979 * resource_get_handle.
981 static struct pipe_resource
*
982 fd_resource_from_handle(struct pipe_screen
*pscreen
,
983 const struct pipe_resource
*tmpl
,
984 struct winsys_handle
*handle
, unsigned usage
)
986 struct fd_screen
*screen
= fd_screen(pscreen
);
987 struct fd_resource
*rsc
= CALLOC_STRUCT(fd_resource
);
988 struct fd_resource_slice
*slice
= &rsc
->slices
[0];
989 struct pipe_resource
*prsc
= &rsc
->base
;
990 uint32_t pitchalign
= fd_screen(pscreen
)->gmem_alignw
;
992 DBG("target=%d, format=%s, %ux%ux%u, array_size=%u, last_level=%u, "
993 "nr_samples=%u, usage=%u, bind=%x, flags=%x",
994 tmpl
->target
, util_format_name(tmpl
->format
),
995 tmpl
->width0
, tmpl
->height0
, tmpl
->depth0
,
996 tmpl
->array_size
, tmpl
->last_level
, tmpl
->nr_samples
,
997 tmpl
->usage
, tmpl
->bind
, tmpl
->flags
);
1004 pipe_reference_init(&prsc
->reference
, 1);
1006 prsc
->screen
= pscreen
;
1008 util_range_init(&rsc
->valid_buffer_range
);
1010 rsc
->bo
= fd_screen_bo_from_handle(pscreen
, handle
);
1014 prsc
->nr_samples
= MAX2(1, prsc
->nr_samples
);
1015 rsc
->internal_format
= tmpl
->format
;
1016 rsc
->cpp
= prsc
->nr_samples
* util_format_get_blocksize(tmpl
->format
);
1017 slice
->pitch
= handle
->stride
/ rsc
->cpp
;
1018 slice
->offset
= handle
->offset
;
1019 slice
->size0
= handle
->stride
* prsc
->height0
;
1021 if ((slice
->pitch
< align(prsc
->width0
, pitchalign
)) ||
1022 (slice
->pitch
& (pitchalign
- 1)))
1025 if (handle
->modifier
== DRM_FORMAT_MOD_QCOM_COMPRESSED
) {
1026 if (!is_supported_modifier(pscreen
, tmpl
->format
,
1027 DRM_FORMAT_MOD_QCOM_COMPRESSED
)) {
1028 DBG("bad modifier: %lx", handle
->modifier
);
1031 debug_assert(screen
->fill_ubwc_buffer_sizes
);
1032 screen
->fill_ubwc_buffer_sizes(rsc
);
1033 } else if (handle
->modifier
&&
1034 (handle
->modifier
!= DRM_FORMAT_MOD_INVALID
)) {
1043 fd_resource_destroy(pscreen
, prsc
);
1048 fd_render_condition_check(struct pipe_context
*pctx
)
1050 struct fd_context
*ctx
= fd_context(pctx
);
1052 if (!ctx
->cond_query
)
1055 union pipe_query_result res
= { 0 };
1057 ctx
->cond_mode
!= PIPE_RENDER_COND_NO_WAIT
&&
1058 ctx
->cond_mode
!= PIPE_RENDER_COND_BY_REGION_NO_WAIT
;
1060 if (pctx
->get_query_result(pctx
, ctx
->cond_query
, wait
, &res
))
1061 return (bool)res
.u64
!= ctx
->cond_cond
;
1067 * Optimal hardware path for blitting pixels.
1068 * Scaling, format conversion, up- and downsampling (resolve) are allowed.
1071 fd_blit(struct pipe_context
*pctx
, const struct pipe_blit_info
*blit_info
)
1073 struct fd_context
*ctx
= fd_context(pctx
);
1074 struct pipe_blit_info info
= *blit_info
;
1076 if (info
.render_condition_enable
&& !fd_render_condition_check(pctx
))
1079 if (info
.mask
& PIPE_MASK_S
) {
1080 DBG("cannot blit stencil, skipping");
1081 info
.mask
&= ~PIPE_MASK_S
;
1084 if (!util_blitter_is_blit_supported(ctx
->blitter
, &info
)) {
1085 DBG("blit unsupported %s -> %s",
1086 util_format_short_name(info
.src
.resource
->format
),
1087 util_format_short_name(info
.dst
.resource
->format
));
1091 if (!(ctx
->blit
&& ctx
->blit(ctx
, &info
)))
1092 fd_blitter_blit(ctx
, &info
);
1096 fd_blitter_pipe_begin(struct fd_context
*ctx
, bool render_cond
, bool discard
,
1097 enum fd_render_stage stage
)
1099 fd_fence_ref(ctx
->base
.screen
, &ctx
->last_fence
, NULL
);
1101 util_blitter_save_fragment_constant_buffer_slot(ctx
->blitter
,
1102 ctx
->constbuf
[PIPE_SHADER_FRAGMENT
].cb
);
1103 util_blitter_save_vertex_buffer_slot(ctx
->blitter
, ctx
->vtx
.vertexbuf
.vb
);
1104 util_blitter_save_vertex_elements(ctx
->blitter
, ctx
->vtx
.vtx
);
1105 util_blitter_save_vertex_shader(ctx
->blitter
, ctx
->prog
.vp
);
1106 util_blitter_save_so_targets(ctx
->blitter
, ctx
->streamout
.num_targets
,
1107 ctx
->streamout
.targets
);
1108 util_blitter_save_rasterizer(ctx
->blitter
, ctx
->rasterizer
);
1109 util_blitter_save_viewport(ctx
->blitter
, &ctx
->viewport
);
1110 util_blitter_save_scissor(ctx
->blitter
, &ctx
->scissor
);
1111 util_blitter_save_fragment_shader(ctx
->blitter
, ctx
->prog
.fp
);
1112 util_blitter_save_blend(ctx
->blitter
, ctx
->blend
);
1113 util_blitter_save_depth_stencil_alpha(ctx
->blitter
, ctx
->zsa
);
1114 util_blitter_save_stencil_ref(ctx
->blitter
, &ctx
->stencil_ref
);
1115 util_blitter_save_sample_mask(ctx
->blitter
, ctx
->sample_mask
);
1116 util_blitter_save_framebuffer(ctx
->blitter
, &ctx
->framebuffer
);
1117 util_blitter_save_fragment_sampler_states(ctx
->blitter
,
1118 ctx
->tex
[PIPE_SHADER_FRAGMENT
].num_samplers
,
1119 (void **)ctx
->tex
[PIPE_SHADER_FRAGMENT
].samplers
);
1120 util_blitter_save_fragment_sampler_views(ctx
->blitter
,
1121 ctx
->tex
[PIPE_SHADER_FRAGMENT
].num_textures
,
1122 ctx
->tex
[PIPE_SHADER_FRAGMENT
].textures
);
1124 util_blitter_save_render_condition(ctx
->blitter
,
1125 ctx
->cond_query
, ctx
->cond_cond
, ctx
->cond_mode
);
1128 fd_batch_set_stage(ctx
->batch
, stage
);
1130 ctx
->in_blit
= discard
;
1134 fd_blitter_pipe_end(struct fd_context
*ctx
)
1137 fd_batch_set_stage(ctx
->batch
, FD_STAGE_NULL
);
1138 ctx
->in_blit
= false;
1142 fd_invalidate_resource(struct pipe_context
*pctx
, struct pipe_resource
*prsc
)
1144 struct fd_resource
*rsc
= fd_resource(prsc
);
1147 * TODO I guess we could track that the resource is invalidated and
1148 * use that as a hint to realloc rather than stall in _transfer_map(),
1149 * even in the non-DISCARD_WHOLE_RESOURCE case?
1152 if (rsc
->write_batch
) {
1153 struct fd_batch
*batch
= rsc
->write_batch
;
1154 struct pipe_framebuffer_state
*pfb
= &batch
->framebuffer
;
1156 if (pfb
->zsbuf
&& pfb
->zsbuf
->texture
== prsc
)
1157 batch
->resolve
&= ~(FD_BUFFER_DEPTH
| FD_BUFFER_STENCIL
);
1159 for (unsigned i
= 0; i
< pfb
->nr_cbufs
; i
++) {
1160 if (pfb
->cbufs
[i
] && pfb
->cbufs
[i
]->texture
== prsc
) {
1161 batch
->resolve
&= ~(PIPE_CLEAR_COLOR0
<< i
);
1169 static enum pipe_format
1170 fd_resource_get_internal_format(struct pipe_resource
*prsc
)
1172 return fd_resource(prsc
)->internal_format
;
1176 fd_resource_set_stencil(struct pipe_resource
*prsc
,
1177 struct pipe_resource
*stencil
)
1179 fd_resource(prsc
)->stencil
= fd_resource(stencil
);
1182 static struct pipe_resource
*
1183 fd_resource_get_stencil(struct pipe_resource
*prsc
)
1185 struct fd_resource
*rsc
= fd_resource(prsc
);
1187 return &rsc
->stencil
->base
;
1191 static const struct u_transfer_vtbl transfer_vtbl
= {
1192 .resource_create
= fd_resource_create
,
1193 .resource_destroy
= fd_resource_destroy
,
1194 .transfer_map
= fd_resource_transfer_map
,
1195 .transfer_flush_region
= fd_resource_transfer_flush_region
,
1196 .transfer_unmap
= fd_resource_transfer_unmap
,
1197 .get_internal_format
= fd_resource_get_internal_format
,
1198 .set_stencil
= fd_resource_set_stencil
,
1199 .get_stencil
= fd_resource_get_stencil
,
1203 fd_resource_screen_init(struct pipe_screen
*pscreen
)
1205 struct fd_screen
*screen
= fd_screen(pscreen
);
1206 bool fake_rgtc
= screen
->gpu_id
< 400;
1208 pscreen
->resource_create
= u_transfer_helper_resource_create
;
1209 /* NOTE: u_transfer_helper does not yet support the _with_modifiers()
1212 pscreen
->resource_create_with_modifiers
= fd_resource_create_with_modifiers
;
1213 pscreen
->resource_from_handle
= fd_resource_from_handle
;
1214 pscreen
->resource_get_handle
= fd_resource_get_handle
;
1215 pscreen
->resource_destroy
= u_transfer_helper_resource_destroy
;
1217 pscreen
->transfer_helper
= u_transfer_helper_create(&transfer_vtbl
,
1218 true, false, fake_rgtc
, true);
1220 if (!screen
->setup_slices
)
1221 screen
->setup_slices
= fd_setup_slices
;
1225 fd_get_sample_position(struct pipe_context
*context
,
1226 unsigned sample_count
, unsigned sample_index
,
1229 /* The following is copied from nouveau/nv50 except for position
1230 * values, which are taken from blob driver */
1231 static const uint8_t pos1
[1][2] = { { 0x8, 0x8 } };
1232 static const uint8_t pos2
[2][2] = {
1233 { 0xc, 0xc }, { 0x4, 0x4 } };
1234 static const uint8_t pos4
[4][2] = {
1235 { 0x6, 0x2 }, { 0xe, 0x6 },
1236 { 0x2, 0xa }, { 0xa, 0xe } };
1237 /* TODO needs to be verified on supported hw */
1238 static const uint8_t pos8
[8][2] = {
1239 { 0x9, 0x5 }, { 0x7, 0xb },
1240 { 0xd, 0x9 }, { 0x5, 0x3 },
1241 { 0x3, 0xd }, { 0x1, 0x7 },
1242 { 0xb, 0xf }, { 0xf, 0x1 } };
1244 const uint8_t (*ptr
)[2];
1246 switch (sample_count
) {
1264 pos_out
[0] = ptr
[sample_index
][0] / 16.0f
;
1265 pos_out
[1] = ptr
[sample_index
][1] / 16.0f
;
1269 fd_resource_context_init(struct pipe_context
*pctx
)
1271 pctx
->transfer_map
= u_transfer_helper_transfer_map
;
1272 pctx
->transfer_flush_region
= u_transfer_helper_transfer_flush_region
;
1273 pctx
->transfer_unmap
= u_transfer_helper_transfer_unmap
;
1274 pctx
->buffer_subdata
= u_default_buffer_subdata
;
1275 pctx
->texture_subdata
= u_default_texture_subdata
;
1276 pctx
->create_surface
= fd_create_surface
;
1277 pctx
->surface_destroy
= fd_surface_destroy
;
1278 pctx
->resource_copy_region
= fd_resource_copy_region
;
1279 pctx
->blit
= fd_blit
;
1280 pctx
->flush_resource
= fd_flush_resource
;
1281 pctx
->invalidate_resource
= fd_invalidate_resource
;
1282 pctx
->get_sample_position
= fd_get_sample_position
;