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"
35 #include "util/u_drm.h"
37 #include "freedreno_resource.h"
38 #include "freedreno_batch_cache.h"
39 #include "freedreno_blitter.h"
40 #include "freedreno_fence.h"
41 #include "freedreno_screen.h"
42 #include "freedreno_surface.h"
43 #include "freedreno_context.h"
44 #include "freedreno_query_hw.h"
45 #include "freedreno_util.h"
47 #include "drm-uapi/drm_fourcc.h"
50 /* XXX this should go away, needed for 'struct winsys_handle' */
51 #include "state_tracker/drm_driver.h"
54 * Go through the entire state and see if the resource is bound
55 * anywhere. If it is, mark the relevant state as dirty. This is
56 * called on realloc_bo to ensure the neccessary state is re-
57 * emitted so the GPU looks at the new backing bo.
60 rebind_resource(struct fd_context
*ctx
, struct pipe_resource
*prsc
)
63 for (unsigned i
= 0; i
< ctx
->vtx
.vertexbuf
.count
&& !(ctx
->dirty
& FD_DIRTY_VTXBUF
); i
++) {
64 if (ctx
->vtx
.vertexbuf
.vb
[i
].buffer
.resource
== prsc
)
65 ctx
->dirty
|= FD_DIRTY_VTXBUF
;
68 /* per-shader-stage resources: */
69 for (unsigned stage
= 0; stage
< PIPE_SHADER_TYPES
; stage
++) {
70 /* Constbufs.. note that constbuf[0] is normal uniforms emitted in
71 * cmdstream rather than by pointer..
73 const unsigned num_ubos
= util_last_bit(ctx
->constbuf
[stage
].enabled_mask
);
74 for (unsigned i
= 1; i
< num_ubos
; i
++) {
75 if (ctx
->dirty_shader
[stage
] & FD_DIRTY_SHADER_CONST
)
77 if (ctx
->constbuf
[stage
].cb
[i
].buffer
== prsc
)
78 ctx
->dirty_shader
[stage
] |= FD_DIRTY_SHADER_CONST
;
82 for (unsigned i
= 0; i
< ctx
->tex
[stage
].num_textures
; i
++) {
83 if (ctx
->dirty_shader
[stage
] & FD_DIRTY_SHADER_TEX
)
85 if (ctx
->tex
[stage
].textures
[i
] && (ctx
->tex
[stage
].textures
[i
]->texture
== prsc
))
86 ctx
->dirty_shader
[stage
] |= FD_DIRTY_SHADER_TEX
;
90 const unsigned num_ssbos
= util_last_bit(ctx
->shaderbuf
[stage
].enabled_mask
);
91 for (unsigned i
= 0; i
< num_ssbos
; i
++) {
92 if (ctx
->dirty_shader
[stage
] & FD_DIRTY_SHADER_SSBO
)
94 if (ctx
->shaderbuf
[stage
].sb
[i
].buffer
== prsc
)
95 ctx
->dirty_shader
[stage
] |= FD_DIRTY_SHADER_SSBO
;
101 realloc_bo(struct fd_resource
*rsc
, uint32_t size
)
103 struct pipe_resource
*prsc
= &rsc
->base
;
104 struct fd_screen
*screen
= fd_screen(rsc
->base
.screen
);
105 uint32_t flags
= DRM_FREEDRENO_GEM_CACHE_WCOMBINE
|
106 DRM_FREEDRENO_GEM_TYPE_KMEM
|
107 COND(prsc
->bind
& PIPE_BIND_SCANOUT
, DRM_FREEDRENO_GEM_SCANOUT
);
108 /* TODO other flags? */
110 /* if we start using things other than write-combine,
111 * be sure to check for PIPE_RESOURCE_FLAG_MAP_COHERENT
117 rsc
->bo
= fd_bo_new(screen
->dev
, size
, flags
, "%ux%ux%u@%u:%x",
118 prsc
->width0
, prsc
->height0
, prsc
->depth0
, rsc
->cpp
, prsc
->bind
);
119 rsc
->seqno
= p_atomic_inc_return(&screen
->rsc_seqno
);
120 util_range_set_empty(&rsc
->valid_buffer_range
);
121 fd_bc_invalidate_resource(rsc
, true);
125 do_blit(struct fd_context
*ctx
, const struct pipe_blit_info
*blit
, bool fallback
)
127 struct pipe_context
*pctx
= &ctx
->base
;
129 /* TODO size threshold too?? */
130 if (fallback
|| !fd_blit(pctx
, blit
)) {
131 /* do blit on cpu: */
132 util_resource_copy_region(pctx
,
133 blit
->dst
.resource
, blit
->dst
.level
, blit
->dst
.box
.x
,
134 blit
->dst
.box
.y
, blit
->dst
.box
.z
,
135 blit
->src
.resource
, blit
->src
.level
, &blit
->src
.box
);
140 fd_try_shadow_resource(struct fd_context
*ctx
, struct fd_resource
*rsc
,
141 unsigned level
, const struct pipe_box
*box
, uint64_t modifier
)
143 struct pipe_context
*pctx
= &ctx
->base
;
144 struct pipe_resource
*prsc
= &rsc
->base
;
145 bool fallback
= false;
150 /* TODO: somehow munge dimensions and format to copy unsupported
151 * render target format to something that is supported?
153 if (!pctx
->screen
->is_format_supported(pctx
->screen
,
154 prsc
->format
, prsc
->target
, prsc
->nr_samples
,
155 prsc
->nr_storage_samples
,
156 PIPE_BIND_RENDER_TARGET
))
159 /* do shadowing back-blits on the cpu for buffers: */
160 if (prsc
->target
== PIPE_BUFFER
)
163 bool whole_level
= util_texrange_covers_whole_level(prsc
, level
,
164 box
->x
, box
->y
, box
->z
, box
->width
, box
->height
, box
->depth
);
166 /* TODO need to be more clever about current level */
167 if ((prsc
->target
>= PIPE_TEXTURE_2D
) && !whole_level
)
170 struct pipe_resource
*pshadow
=
171 pctx
->screen
->resource_create_with_modifiers(pctx
->screen
,
177 assert(!ctx
->in_shadow
);
178 ctx
->in_shadow
= true;
180 /* get rid of any references that batch-cache might have to us (which
181 * should empty/destroy rsc->batches hashset)
183 fd_bc_invalidate_resource(rsc
, false);
185 mtx_lock(&ctx
->screen
->lock
);
187 /* Swap the backing bo's, so shadow becomes the old buffer,
188 * blit from shadow to new buffer. From here on out, we
191 * Note that we need to do it in this order, otherwise if
192 * we go down cpu blit path, the recursive transfer_map()
193 * sees the wrong status..
195 struct fd_resource
*shadow
= fd_resource(pshadow
);
197 DBG("shadow: %p (%d) -> %p (%d)\n", rsc
, rsc
->base
.reference
.count
,
198 shadow
, shadow
->base
.reference
.count
);
200 /* TODO valid_buffer_range?? */
201 swap(rsc
->bo
, shadow
->bo
);
202 swap(rsc
->write_batch
, shadow
->write_batch
);
203 rsc
->seqno
= p_atomic_inc_return(&ctx
->screen
->rsc_seqno
);
205 /* at this point, the newly created shadow buffer is not referenced
206 * by any batches, but the existing rsc (probably) is. We need to
207 * transfer those references over:
209 debug_assert(shadow
->batch_mask
== 0);
210 struct fd_batch
*batch
;
211 foreach_batch(batch
, &ctx
->screen
->batch_cache
, rsc
->batch_mask
) {
212 struct set_entry
*entry
= _mesa_set_search(batch
->resources
, rsc
);
213 _mesa_set_remove(batch
->resources
, entry
);
214 _mesa_set_add(batch
->resources
, shadow
);
216 swap(rsc
->batch_mask
, shadow
->batch_mask
);
218 mtx_unlock(&ctx
->screen
->lock
);
220 struct pipe_blit_info blit
= {};
221 blit
.dst
.resource
= prsc
;
222 blit
.dst
.format
= prsc
->format
;
223 blit
.src
.resource
= pshadow
;
224 blit
.src
.format
= pshadow
->format
;
225 blit
.mask
= util_format_get_mask(prsc
->format
);
226 blit
.filter
= PIPE_TEX_FILTER_NEAREST
;
228 #define set_box(field, val) do { \
229 blit.dst.field = (val); \
230 blit.src.field = (val); \
233 /* blit the other levels in their entirety: */
234 for (unsigned l
= 0; l
<= prsc
->last_level
; l
++) {
238 /* just blit whole level: */
240 set_box(box
.width
, u_minify(prsc
->width0
, l
));
241 set_box(box
.height
, u_minify(prsc
->height0
, l
));
242 set_box(box
.depth
, u_minify(prsc
->depth0
, l
));
244 do_blit(ctx
, &blit
, fallback
);
247 /* deal w/ current level specially, since we might need to split
248 * it up into a couple blits:
251 set_box(level
, level
);
253 switch (prsc
->target
) {
255 case PIPE_TEXTURE_1D
:
258 set_box(box
.height
, 1);
259 set_box(box
.depth
, 1);
263 set_box(box
.width
, box
->x
);
265 do_blit(ctx
, &blit
, fallback
);
267 if ((box
->x
+ box
->width
) < u_minify(prsc
->width0
, level
)) {
268 set_box(box
.x
, box
->x
+ box
->width
);
269 set_box(box
.width
, u_minify(prsc
->width0
, level
) - (box
->x
+ box
->width
));
271 do_blit(ctx
, &blit
, fallback
);
274 case PIPE_TEXTURE_2D
:
281 ctx
->in_shadow
= false;
283 pipe_resource_reference(&pshadow
, NULL
);
288 static struct fd_resource
*
289 fd_alloc_staging(struct fd_context
*ctx
, struct fd_resource
*rsc
,
290 unsigned level
, const struct pipe_box
*box
)
292 struct pipe_context
*pctx
= &ctx
->base
;
293 struct pipe_resource tmpl
= rsc
->base
;
295 tmpl
.width0
= box
->width
;
296 tmpl
.height0
= box
->height
;
297 /* for array textures, box->depth is the array_size, otherwise
298 * for 3d textures, it is the depth:
300 if (tmpl
.array_size
> 1) {
301 tmpl
.array_size
= box
->depth
;
305 tmpl
.depth0
= box
->depth
;
308 tmpl
.bind
|= PIPE_BIND_LINEAR
;
310 struct pipe_resource
*pstaging
=
311 pctx
->screen
->resource_create(pctx
->screen
, &tmpl
);
315 return fd_resource(pstaging
);
319 fd_blit_from_staging(struct fd_context
*ctx
, struct fd_transfer
*trans
)
321 struct pipe_resource
*dst
= trans
->base
.resource
;
322 struct pipe_blit_info blit
= {};
324 blit
.dst
.resource
= dst
;
325 blit
.dst
.format
= dst
->format
;
326 blit
.dst
.level
= trans
->base
.level
;
327 blit
.dst
.box
= trans
->base
.box
;
328 blit
.src
.resource
= trans
->staging_prsc
;
329 blit
.src
.format
= trans
->staging_prsc
->format
;
331 blit
.src
.box
= trans
->staging_box
;
332 blit
.mask
= util_format_get_mask(trans
->staging_prsc
->format
);
333 blit
.filter
= PIPE_TEX_FILTER_NEAREST
;
335 do_blit(ctx
, &blit
, false);
339 fd_blit_to_staging(struct fd_context
*ctx
, struct fd_transfer
*trans
)
341 struct pipe_resource
*src
= trans
->base
.resource
;
342 struct pipe_blit_info blit
= {};
344 blit
.src
.resource
= src
;
345 blit
.src
.format
= src
->format
;
346 blit
.src
.level
= trans
->base
.level
;
347 blit
.src
.box
= trans
->base
.box
;
348 blit
.dst
.resource
= trans
->staging_prsc
;
349 blit
.dst
.format
= trans
->staging_prsc
->format
;
351 blit
.dst
.box
= trans
->staging_box
;
352 blit
.mask
= util_format_get_mask(trans
->staging_prsc
->format
);
353 blit
.filter
= PIPE_TEX_FILTER_NEAREST
;
355 do_blit(ctx
, &blit
, false);
358 static void fd_resource_transfer_flush_region(struct pipe_context
*pctx
,
359 struct pipe_transfer
*ptrans
,
360 const struct pipe_box
*box
)
362 struct fd_resource
*rsc
= fd_resource(ptrans
->resource
);
364 if (ptrans
->resource
->target
== PIPE_BUFFER
)
365 util_range_add(&rsc
->valid_buffer_range
,
366 ptrans
->box
.x
+ box
->x
,
367 ptrans
->box
.x
+ box
->x
+ box
->width
);
371 flush_resource(struct fd_context
*ctx
, struct fd_resource
*rsc
, unsigned usage
)
373 struct fd_batch
*write_batch
= NULL
;
375 mtx_lock(&ctx
->screen
->lock
);
376 fd_batch_reference_locked(&write_batch
, rsc
->write_batch
);
377 mtx_unlock(&ctx
->screen
->lock
);
379 if (usage
& PIPE_TRANSFER_WRITE
) {
380 struct fd_batch
*batch
, *batches
[32] = {};
383 /* This is a bit awkward, probably a fd_batch_flush_locked()
384 * would make things simpler.. but we need to hold the lock
385 * to iterate the batches which reference this resource. So
386 * we must first grab references under a lock, then flush.
388 mtx_lock(&ctx
->screen
->lock
);
389 batch_mask
= rsc
->batch_mask
;
390 foreach_batch(batch
, &ctx
->screen
->batch_cache
, batch_mask
)
391 fd_batch_reference_locked(&batches
[batch
->idx
], batch
);
392 mtx_unlock(&ctx
->screen
->lock
);
394 foreach_batch(batch
, &ctx
->screen
->batch_cache
, batch_mask
)
395 fd_batch_flush(batch
, false, false);
397 foreach_batch(batch
, &ctx
->screen
->batch_cache
, batch_mask
) {
398 fd_batch_sync(batch
);
399 fd_batch_reference(&batches
[batch
->idx
], NULL
);
401 assert(rsc
->batch_mask
== 0);
402 } else if (write_batch
) {
403 fd_batch_flush(write_batch
, true, false);
406 fd_batch_reference(&write_batch
, NULL
);
408 assert(!rsc
->write_batch
);
412 fd_flush_resource(struct pipe_context
*pctx
, struct pipe_resource
*prsc
)
414 flush_resource(fd_context(pctx
), fd_resource(prsc
), PIPE_TRANSFER_READ
);
418 fd_resource_transfer_unmap(struct pipe_context
*pctx
,
419 struct pipe_transfer
*ptrans
)
421 struct fd_context
*ctx
= fd_context(pctx
);
422 struct fd_resource
*rsc
= fd_resource(ptrans
->resource
);
423 struct fd_transfer
*trans
= fd_transfer(ptrans
);
425 if (trans
->staging_prsc
) {
426 if (ptrans
->usage
& PIPE_TRANSFER_WRITE
)
427 fd_blit_from_staging(ctx
, trans
);
428 pipe_resource_reference(&trans
->staging_prsc
, NULL
);
431 if (!(ptrans
->usage
& PIPE_TRANSFER_UNSYNCHRONIZED
)) {
432 fd_bo_cpu_fini(rsc
->bo
);
435 util_range_add(&rsc
->valid_buffer_range
,
437 ptrans
->box
.x
+ ptrans
->box
.width
);
439 pipe_resource_reference(&ptrans
->resource
, NULL
);
440 slab_free(&ctx
->transfer_pool
, ptrans
);
444 fd_resource_transfer_map(struct pipe_context
*pctx
,
445 struct pipe_resource
*prsc
,
446 unsigned level
, unsigned usage
,
447 const struct pipe_box
*box
,
448 struct pipe_transfer
**pptrans
)
450 struct fd_context
*ctx
= fd_context(pctx
);
451 struct fd_resource
*rsc
= fd_resource(prsc
);
452 struct fd_resource_slice
*slice
= fd_resource_slice(rsc
, level
);
453 struct fd_transfer
*trans
;
454 struct pipe_transfer
*ptrans
;
455 enum pipe_format format
= prsc
->format
;
461 DBG("prsc=%p, level=%u, usage=%x, box=%dx%d+%d,%d", prsc
, level
, usage
,
462 box
->width
, box
->height
, box
->x
, box
->y
);
464 ptrans
= slab_alloc(&ctx
->transfer_pool
);
468 /* slab_alloc_st() doesn't zero: */
469 trans
= fd_transfer(ptrans
);
470 memset(trans
, 0, sizeof(*trans
));
472 pipe_resource_reference(&ptrans
->resource
, prsc
);
473 ptrans
->level
= level
;
474 ptrans
->usage
= usage
;
476 ptrans
->stride
= util_format_get_nblocksx(format
, slice
->pitch
) * rsc
->cpp
;
477 ptrans
->layer_stride
= rsc
->layer_first
? rsc
->layer_size
: slice
->size0
;
479 /* we always need a staging texture for tiled buffers:
481 * TODO we might sometimes want to *also* shadow the resource to avoid
482 * splitting a batch.. for ex, mid-frame texture uploads to a tiled
485 if (rsc
->tile_mode
) {
486 struct fd_resource
*staging_rsc
;
488 staging_rsc
= fd_alloc_staging(ctx
, rsc
, level
, box
);
490 // TODO for PIPE_TRANSFER_READ, need to do untiling blit..
491 trans
->staging_prsc
= &staging_rsc
->base
;
492 trans
->base
.stride
= util_format_get_nblocksx(format
,
493 staging_rsc
->slices
[0].pitch
) * staging_rsc
->cpp
;
494 trans
->base
.layer_stride
= staging_rsc
->layer_first
?
495 staging_rsc
->layer_size
: staging_rsc
->slices
[0].size0
;
496 trans
->staging_box
= *box
;
497 trans
->staging_box
.x
= 0;
498 trans
->staging_box
.y
= 0;
499 trans
->staging_box
.z
= 0;
501 if (usage
& PIPE_TRANSFER_READ
) {
502 fd_blit_to_staging(ctx
, trans
);
504 struct fd_batch
*batch
= NULL
;
506 fd_context_lock(ctx
);
507 fd_batch_reference_locked(&batch
, staging_rsc
->write_batch
);
508 fd_context_unlock(ctx
);
510 /* we can't fd_bo_cpu_prep() until the blit to staging
511 * is submitted to kernel.. in that case write_batch
512 * wouldn't be NULL yet:
515 fd_batch_sync(batch
);
516 fd_batch_reference(&batch
, NULL
);
519 fd_bo_cpu_prep(staging_rsc
->bo
, ctx
->pipe
,
520 DRM_FREEDRENO_PREP_READ
);
523 buf
= fd_bo_map(staging_rsc
->bo
);
528 ctx
->stats
.staging_uploads
++;
534 if (ctx
->in_shadow
&& !(usage
& PIPE_TRANSFER_READ
))
535 usage
|= PIPE_TRANSFER_UNSYNCHRONIZED
;
537 if (usage
& PIPE_TRANSFER_READ
)
538 op
|= DRM_FREEDRENO_PREP_READ
;
540 if (usage
& PIPE_TRANSFER_WRITE
)
541 op
|= DRM_FREEDRENO_PREP_WRITE
;
543 if (usage
& PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE
) {
544 realloc_bo(rsc
, fd_bo_size(rsc
->bo
));
545 rebind_resource(ctx
, prsc
);
546 } else if ((usage
& PIPE_TRANSFER_WRITE
) &&
547 prsc
->target
== PIPE_BUFFER
&&
548 !util_ranges_intersect(&rsc
->valid_buffer_range
,
549 box
->x
, box
->x
+ box
->width
)) {
550 /* We are trying to write to a previously uninitialized range. No need
553 } else if (!(usage
& PIPE_TRANSFER_UNSYNCHRONIZED
)) {
554 struct fd_batch
*write_batch
= NULL
;
556 /* hold a reference, so it doesn't disappear under us: */
557 fd_context_lock(ctx
);
558 fd_batch_reference_locked(&write_batch
, rsc
->write_batch
);
559 fd_context_unlock(ctx
);
561 if ((usage
& PIPE_TRANSFER_WRITE
) && write_batch
&&
562 write_batch
->back_blit
) {
563 /* if only thing pending is a back-blit, we can discard it: */
564 fd_batch_reset(write_batch
);
567 /* If the GPU is writing to the resource, or if it is reading from the
568 * resource and we're trying to write to it, flush the renders.
570 bool needs_flush
= pending(rsc
, !!(usage
& PIPE_TRANSFER_WRITE
));
571 bool busy
= needs_flush
|| (0 != fd_bo_cpu_prep(rsc
->bo
,
572 ctx
->pipe
, op
| DRM_FREEDRENO_PREP_NOSYNC
));
574 /* if we need to flush/stall, see if we can make a shadow buffer
577 * TODO we could go down this path !reorder && !busy_for_read
578 * ie. we only *don't* want to go down this path if the blit
579 * will trigger a flush!
581 if (ctx
->screen
->reorder
&& busy
&& !(usage
& PIPE_TRANSFER_READ
) &&
582 (usage
& PIPE_TRANSFER_DISCARD_RANGE
)) {
583 /* try shadowing only if it avoids a flush, otherwise staging would
586 if (needs_flush
&& fd_try_shadow_resource(ctx
, rsc
, level
,
587 box
, DRM_FORMAT_MOD_LINEAR
)) {
588 needs_flush
= busy
= false;
589 rebind_resource(ctx
, prsc
);
590 ctx
->stats
.shadow_uploads
++;
592 struct fd_resource
*staging_rsc
;
595 flush_resource(ctx
, rsc
, usage
);
599 /* in this case, we don't need to shadow the whole resource,
600 * since any draw that references the previous contents has
601 * already had rendering flushed for all tiles. So we can
602 * use a staging buffer to do the upload.
604 staging_rsc
= fd_alloc_staging(ctx
, rsc
, level
, box
);
606 trans
->staging_prsc
= &staging_rsc
->base
;
607 trans
->base
.stride
= util_format_get_nblocksx(format
,
608 staging_rsc
->slices
[0].pitch
) * staging_rsc
->cpp
;
609 trans
->base
.layer_stride
= staging_rsc
->layer_first
?
610 staging_rsc
->layer_size
: staging_rsc
->slices
[0].size0
;
611 trans
->staging_box
= *box
;
612 trans
->staging_box
.x
= 0;
613 trans
->staging_box
.y
= 0;
614 trans
->staging_box
.z
= 0;
615 buf
= fd_bo_map(staging_rsc
->bo
);
620 fd_batch_reference(&write_batch
, NULL
);
622 ctx
->stats
.staging_uploads
++;
630 flush_resource(ctx
, rsc
, usage
);
634 fd_batch_reference(&write_batch
, NULL
);
636 /* The GPU keeps track of how the various bo's are being used, and
637 * will wait if necessary for the proper operation to have
641 ret
= fd_bo_cpu_prep(rsc
->bo
, ctx
->pipe
, op
);
647 buf
= fd_bo_map(rsc
->bo
);
649 box
->y
/ util_format_get_blockheight(format
) * ptrans
->stride
+
650 box
->x
/ util_format_get_blockwidth(format
) * rsc
->cpp
+
651 fd_resource_offset(rsc
, level
, box
->z
);
653 if (usage
& PIPE_TRANSFER_WRITE
)
661 fd_resource_transfer_unmap(pctx
, ptrans
);
666 fd_resource_destroy(struct pipe_screen
*pscreen
,
667 struct pipe_resource
*prsc
)
669 struct fd_resource
*rsc
= fd_resource(prsc
);
670 fd_bc_invalidate_resource(rsc
, true);
674 renderonly_scanout_destroy(rsc
->scanout
, fd_screen(pscreen
)->ro
);
676 util_range_destroy(&rsc
->valid_buffer_range
);
681 fd_resource_modifier(struct fd_resource
*rsc
)
684 return DRM_FORMAT_MOD_LINEAR
;
687 return DRM_FORMAT_MOD_QCOM_COMPRESSED
;
689 /* TODO invent a modifier for tiled but not UBWC buffers: */
690 return DRM_FORMAT_MOD_INVALID
;
694 fd_resource_get_handle(struct pipe_screen
*pscreen
,
695 struct pipe_context
*pctx
,
696 struct pipe_resource
*prsc
,
697 struct winsys_handle
*handle
,
700 struct fd_resource
*rsc
= fd_resource(prsc
);
702 handle
->modifier
= fd_resource_modifier(rsc
);
704 return fd_screen_bo_get_handle(pscreen
, rsc
->bo
, rsc
->scanout
,
705 rsc
->slices
[0].pitch
* rsc
->cpp
, handle
);
709 setup_slices(struct fd_resource
*rsc
, uint32_t alignment
, enum pipe_format format
)
711 struct pipe_resource
*prsc
= &rsc
->base
;
712 struct fd_screen
*screen
= fd_screen(prsc
->screen
);
713 enum util_format_layout layout
= util_format_description(format
)->layout
;
714 uint32_t pitchalign
= screen
->gmem_alignw
;
715 uint32_t level
, size
= 0;
716 uint32_t width
= prsc
->width0
;
717 uint32_t height
= prsc
->height0
;
718 uint32_t depth
= prsc
->depth0
;
719 /* in layer_first layout, the level (slice) contains just one
720 * layer (since in fact the layer contains the slices)
722 uint32_t layers_in_level
= rsc
->layer_first
? 1 : prsc
->array_size
;
724 for (level
= 0; level
<= prsc
->last_level
; level
++) {
725 struct fd_resource_slice
*slice
= fd_resource_slice(rsc
, level
);
728 if (layout
== UTIL_FORMAT_LAYOUT_ASTC
)
729 slice
->pitch
= width
=
730 util_align_npot(width
, pitchalign
* util_format_get_blockwidth(format
));
732 slice
->pitch
= width
= align(width
, pitchalign
);
733 slice
->offset
= size
;
734 blocks
= util_format_get_nblocks(format
, width
, height
);
735 /* 1d array and 2d array textures must all have the same layer size
736 * for each miplevel on a3xx. 3d textures can have different layer
737 * sizes for high levels, but the hw auto-sizer is buggy (or at least
738 * different than what this code does), so as soon as the layer size
739 * range gets into range, we stop reducing it.
741 if (prsc
->target
== PIPE_TEXTURE_3D
&& (
743 (level
> 1 && rsc
->slices
[level
- 1].size0
> 0xf000)))
744 slice
->size0
= align(blocks
* rsc
->cpp
, alignment
);
745 else if (level
== 0 || rsc
->layer_first
|| alignment
== 1)
746 slice
->size0
= align(blocks
* rsc
->cpp
, alignment
);
748 slice
->size0
= rsc
->slices
[level
- 1].size0
;
750 size
+= slice
->size0
* depth
* layers_in_level
;
752 width
= u_minify(width
, 1);
753 height
= u_minify(height
, 1);
754 depth
= u_minify(depth
, 1);
761 slice_alignment(enum pipe_texture_target target
)
763 /* on a3xx, 2d array and 3d textures seem to want their
764 * layers aligned to page boundaries:
767 case PIPE_TEXTURE_3D
:
768 case PIPE_TEXTURE_1D_ARRAY
:
769 case PIPE_TEXTURE_2D_ARRAY
:
776 /* cross generation texture layout to plug in to screen->setup_slices()..
777 * replace with generation specific one as-needed.
779 * TODO for a4xx probably can extract out the a4xx specific logic int
780 * a small fd4_setup_slices() wrapper that sets up layer_first, and then
784 fd_setup_slices(struct fd_resource
*rsc
)
788 alignment
= slice_alignment(rsc
->base
.target
);
790 struct fd_screen
*screen
= fd_screen(rsc
->base
.screen
);
791 if (is_a4xx(screen
)) {
792 switch (rsc
->base
.target
) {
793 case PIPE_TEXTURE_3D
:
794 rsc
->layer_first
= false;
797 rsc
->layer_first
= true;
803 return setup_slices(rsc
, alignment
, rsc
->base
.format
);
806 /* special case to resize query buf after allocated.. */
808 fd_resource_resize(struct pipe_resource
*prsc
, uint32_t sz
)
810 struct fd_resource
*rsc
= fd_resource(prsc
);
812 debug_assert(prsc
->width0
== 0);
813 debug_assert(prsc
->target
== PIPE_BUFFER
);
814 debug_assert(prsc
->bind
== PIPE_BIND_QUERY_BUFFER
);
817 realloc_bo(rsc
, fd_screen(prsc
->screen
)->setup_slices(rsc
));
820 // TODO common helper?
822 has_depth(enum pipe_format format
)
825 case PIPE_FORMAT_Z16_UNORM
:
826 case PIPE_FORMAT_Z32_UNORM
:
827 case PIPE_FORMAT_Z32_FLOAT
:
828 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
:
829 case PIPE_FORMAT_Z24_UNORM_S8_UINT
:
830 case PIPE_FORMAT_S8_UINT_Z24_UNORM
:
831 case PIPE_FORMAT_Z24X8_UNORM
:
832 case PIPE_FORMAT_X8Z24_UNORM
:
840 * Create a new texture object, using the given template info.
842 static struct pipe_resource
*
843 fd_resource_create_with_modifiers(struct pipe_screen
*pscreen
,
844 const struct pipe_resource
*tmpl
,
845 const uint64_t *modifiers
, int count
)
847 struct fd_screen
*screen
= fd_screen(pscreen
);
848 struct fd_resource
*rsc
;
849 struct pipe_resource
*prsc
;
850 enum pipe_format format
= tmpl
->format
;
853 /* when using kmsro, scanout buffers are allocated on the display device
854 * create_with_modifiers() doesn't give us usage flags, so we have to
855 * assume that all calls with modifiers are scanout-possible
858 ((tmpl
->bind
& PIPE_BIND_SCANOUT
) ||
859 !(count
== 1 && modifiers
[0] == DRM_FORMAT_MOD_INVALID
))) {
860 struct pipe_resource scanout_templat
= *tmpl
;
861 struct renderonly_scanout
*scanout
;
862 struct winsys_handle handle
;
864 scanout
= renderonly_scanout_for_resource(&scanout_templat
,
865 screen
->ro
, &handle
);
869 renderonly_scanout_destroy(scanout
, screen
->ro
);
871 assert(handle
.type
== WINSYS_HANDLE_TYPE_FD
);
872 rsc
= fd_resource(pscreen
->resource_from_handle(pscreen
, tmpl
,
874 PIPE_HANDLE_USAGE_FRAMEBUFFER_WRITE
));
875 close(handle
.handle
);
882 rsc
= CALLOC_STRUCT(fd_resource
);
885 DBG("%p: target=%d, format=%s, %ux%ux%u, array_size=%u, last_level=%u, "
886 "nr_samples=%u, usage=%u, bind=%x, flags=%x", prsc
,
887 tmpl
->target
, util_format_name(format
),
888 tmpl
->width0
, tmpl
->height0
, tmpl
->depth0
,
889 tmpl
->array_size
, tmpl
->last_level
, tmpl
->nr_samples
,
890 tmpl
->usage
, tmpl
->bind
, tmpl
->flags
);
898 (PIPE_BIND_SCANOUT | \
900 PIPE_BIND_DISPLAY_TARGET)
902 bool linear
= drm_find_modifier(DRM_FORMAT_MOD_LINEAR
, modifiers
, count
);
903 if (tmpl
->bind
& LINEAR
)
906 /* Normally, for non-shared buffers, allow buffer compression if
907 * not shared, otherwise only allow if QCOM_COMPRESSED modifier
910 * TODO we should probably also limit tiled in a similar way,
911 * except we don't have a format modifier for tiled. (We probably
914 bool allow_ubwc
= drm_find_modifier(DRM_FORMAT_MOD_INVALID
, modifiers
, count
);
915 if (tmpl
->bind
& PIPE_BIND_SHARED
)
916 allow_ubwc
= drm_find_modifier(DRM_FORMAT_MOD_QCOM_COMPRESSED
, modifiers
, count
);
918 /* TODO turn on UBWC for all internal buffers
920 * There are still some regressions in deqp with UBWC enabled. I
921 * think it is mostly related to sampler/image views using a format
922 * that doesn't support compression with a resource created with
923 * a format that does. We need to track the compression state of
924 * a buffer and do an (in-place, hopefully?) resolve if it is re-
925 * interpreted with a format that does not support compression.
927 * It is possible (likely?) that we can't do atomic ops on a
928 * compressed buffer as well, so this would also require transition
929 * to a compressed state.
931 allow_ubwc
&= !!(fd_mesa_debug
& FD_DBG_UBWC
);
933 if (screen
->tile_mode
&&
934 (tmpl
->target
!= PIPE_BUFFER
) &&
936 rsc
->tile_mode
= screen
->tile_mode(tmpl
);
939 pipe_reference_init(&prsc
->reference
, 1);
941 prsc
->screen
= pscreen
;
943 util_range_init(&rsc
->valid_buffer_range
);
945 rsc
->internal_format
= format
;
946 rsc
->cpp
= util_format_get_blocksize(format
);
947 rsc
->cpp
*= fd_resource_nr_samples(prsc
);
951 // XXX probably need some extra work if we hit rsc shadowing path w/ lrz..
952 if ((is_a5xx(screen
) || is_a6xx(screen
)) &&
953 (fd_mesa_debug
& FD_DBG_LRZ
) && has_depth(format
)) {
954 const uint32_t flags
= DRM_FREEDRENO_GEM_CACHE_WCOMBINE
|
955 DRM_FREEDRENO_GEM_TYPE_KMEM
; /* TODO */
956 unsigned lrz_pitch
= align(DIV_ROUND_UP(tmpl
->width0
, 8), 64);
957 unsigned lrz_height
= DIV_ROUND_UP(tmpl
->height0
, 8);
959 /* LRZ buffer is super-sampled: */
960 switch (prsc
->nr_samples
) {
967 unsigned size
= lrz_pitch
* lrz_height
* 2;
969 size
+= 0x1000; /* for GRAS_LRZ_FAST_CLEAR_BUFFER */
971 rsc
->lrz_height
= lrz_height
;
972 rsc
->lrz_width
= lrz_pitch
;
973 rsc
->lrz_pitch
= lrz_pitch
;
974 rsc
->lrz
= fd_bo_new(screen
->dev
, size
, flags
, "lrz");
977 size
= screen
->setup_slices(rsc
);
979 if (allow_ubwc
&& screen
->fill_ubwc_buffer_sizes
&& rsc
->tile_mode
)
980 size
+= screen
->fill_ubwc_buffer_sizes(rsc
);
982 /* special case for hw-query buffer, which we need to allocate before we
986 /* note, semi-intention == instead of & */
987 debug_assert(prsc
->bind
== PIPE_BIND_QUERY_BUFFER
);
991 if (rsc
->layer_first
) {
992 rsc
->layer_size
= align(size
, 4096);
993 size
= rsc
->layer_size
* prsc
->array_size
;
996 realloc_bo(rsc
, size
);
1002 fd_resource_destroy(pscreen
, prsc
);
1006 static struct pipe_resource
*
1007 fd_resource_create(struct pipe_screen
*pscreen
,
1008 const struct pipe_resource
*tmpl
)
1010 const uint64_t mod
= DRM_FORMAT_MOD_INVALID
;
1011 return fd_resource_create_with_modifiers(pscreen
, tmpl
, &mod
, 1);
1015 is_supported_modifier(struct pipe_screen
*pscreen
, enum pipe_format pfmt
,
1020 /* Get the count of supported modifiers: */
1021 pscreen
->query_dmabuf_modifiers(pscreen
, pfmt
, 0, NULL
, NULL
, &count
);
1023 /* Get the supported modifiers: */
1024 uint64_t modifiers
[count
];
1025 pscreen
->query_dmabuf_modifiers(pscreen
, pfmt
, count
, modifiers
, NULL
, &count
);
1027 for (int i
= 0; i
< count
; i
++)
1028 if (modifiers
[i
] == mod
)
1035 * Create a texture from a winsys_handle. The handle is often created in
1036 * another process by first creating a pipe texture and then calling
1037 * resource_get_handle.
1039 static struct pipe_resource
*
1040 fd_resource_from_handle(struct pipe_screen
*pscreen
,
1041 const struct pipe_resource
*tmpl
,
1042 struct winsys_handle
*handle
, unsigned usage
)
1044 struct fd_screen
*screen
= fd_screen(pscreen
);
1045 struct fd_resource
*rsc
= CALLOC_STRUCT(fd_resource
);
1046 struct fd_resource_slice
*slice
= &rsc
->slices
[0];
1047 struct pipe_resource
*prsc
= &rsc
->base
;
1048 uint32_t pitchalign
= fd_screen(pscreen
)->gmem_alignw
;
1050 DBG("target=%d, format=%s, %ux%ux%u, array_size=%u, last_level=%u, "
1051 "nr_samples=%u, usage=%u, bind=%x, flags=%x",
1052 tmpl
->target
, util_format_name(tmpl
->format
),
1053 tmpl
->width0
, tmpl
->height0
, tmpl
->depth0
,
1054 tmpl
->array_size
, tmpl
->last_level
, tmpl
->nr_samples
,
1055 tmpl
->usage
, tmpl
->bind
, tmpl
->flags
);
1062 pipe_reference_init(&prsc
->reference
, 1);
1064 prsc
->screen
= pscreen
;
1066 util_range_init(&rsc
->valid_buffer_range
);
1068 rsc
->bo
= fd_screen_bo_from_handle(pscreen
, handle
);
1072 rsc
->internal_format
= tmpl
->format
;
1073 rsc
->cpp
= util_format_get_blocksize(tmpl
->format
);
1074 rsc
->cpp
*= fd_resource_nr_samples(prsc
);
1075 slice
->pitch
= handle
->stride
/ rsc
->cpp
;
1076 slice
->offset
= handle
->offset
;
1077 slice
->size0
= handle
->stride
* prsc
->height0
;
1079 if ((slice
->pitch
< align(prsc
->width0
, pitchalign
)) ||
1080 (slice
->pitch
& (pitchalign
- 1)))
1083 if (handle
->modifier
== DRM_FORMAT_MOD_QCOM_COMPRESSED
) {
1084 if (!is_supported_modifier(pscreen
, tmpl
->format
,
1085 DRM_FORMAT_MOD_QCOM_COMPRESSED
)) {
1086 DBG("bad modifier: %"PRIx64
, handle
->modifier
);
1089 debug_assert(screen
->fill_ubwc_buffer_sizes
);
1090 screen
->fill_ubwc_buffer_sizes(rsc
);
1091 } else if (handle
->modifier
&&
1092 (handle
->modifier
!= DRM_FORMAT_MOD_INVALID
)) {
1100 renderonly_create_gpu_import_for_resource(prsc
, screen
->ro
, NULL
);
1101 /* failure is expected in some cases.. */
1109 fd_resource_destroy(pscreen
, prsc
);
1114 fd_render_condition_check(struct pipe_context
*pctx
)
1116 struct fd_context
*ctx
= fd_context(pctx
);
1118 if (!ctx
->cond_query
)
1121 union pipe_query_result res
= { 0 };
1123 ctx
->cond_mode
!= PIPE_RENDER_COND_NO_WAIT
&&
1124 ctx
->cond_mode
!= PIPE_RENDER_COND_BY_REGION_NO_WAIT
;
1126 if (pctx
->get_query_result(pctx
, ctx
->cond_query
, wait
, &res
))
1127 return (bool)res
.u64
!= ctx
->cond_cond
;
1133 fd_invalidate_resource(struct pipe_context
*pctx
, struct pipe_resource
*prsc
)
1135 struct fd_context
*ctx
= fd_context(pctx
);
1136 struct fd_resource
*rsc
= fd_resource(prsc
);
1139 * TODO I guess we could track that the resource is invalidated and
1140 * use that as a hint to realloc rather than stall in _transfer_map(),
1141 * even in the non-DISCARD_WHOLE_RESOURCE case?
1143 * Note: we set dirty bits to trigger invalidate logic fd_draw_vbo
1146 if (rsc
->write_batch
) {
1147 struct fd_batch
*batch
= rsc
->write_batch
;
1148 struct pipe_framebuffer_state
*pfb
= &batch
->framebuffer
;
1150 if (pfb
->zsbuf
&& pfb
->zsbuf
->texture
== prsc
) {
1151 batch
->resolve
&= ~(FD_BUFFER_DEPTH
| FD_BUFFER_STENCIL
);
1152 ctx
->dirty
|= FD_DIRTY_ZSA
;
1155 for (unsigned i
= 0; i
< pfb
->nr_cbufs
; i
++) {
1156 if (pfb
->cbufs
[i
] && pfb
->cbufs
[i
]->texture
== prsc
) {
1157 batch
->resolve
&= ~(PIPE_CLEAR_COLOR0
<< i
);
1158 ctx
->dirty
|= FD_DIRTY_FRAMEBUFFER
;
1166 static enum pipe_format
1167 fd_resource_get_internal_format(struct pipe_resource
*prsc
)
1169 return fd_resource(prsc
)->internal_format
;
1173 fd_resource_set_stencil(struct pipe_resource
*prsc
,
1174 struct pipe_resource
*stencil
)
1176 fd_resource(prsc
)->stencil
= fd_resource(stencil
);
1179 static struct pipe_resource
*
1180 fd_resource_get_stencil(struct pipe_resource
*prsc
)
1182 struct fd_resource
*rsc
= fd_resource(prsc
);
1184 return &rsc
->stencil
->base
;
1188 static const struct u_transfer_vtbl transfer_vtbl
= {
1189 .resource_create
= fd_resource_create
,
1190 .resource_destroy
= fd_resource_destroy
,
1191 .transfer_map
= fd_resource_transfer_map
,
1192 .transfer_flush_region
= fd_resource_transfer_flush_region
,
1193 .transfer_unmap
= fd_resource_transfer_unmap
,
1194 .get_internal_format
= fd_resource_get_internal_format
,
1195 .set_stencil
= fd_resource_set_stencil
,
1196 .get_stencil
= fd_resource_get_stencil
,
1200 fd_resource_screen_init(struct pipe_screen
*pscreen
)
1202 struct fd_screen
*screen
= fd_screen(pscreen
);
1203 bool fake_rgtc
= screen
->gpu_id
< 400;
1205 pscreen
->resource_create
= u_transfer_helper_resource_create
;
1206 /* NOTE: u_transfer_helper does not yet support the _with_modifiers()
1209 pscreen
->resource_create_with_modifiers
= fd_resource_create_with_modifiers
;
1210 pscreen
->resource_from_handle
= fd_resource_from_handle
;
1211 pscreen
->resource_get_handle
= fd_resource_get_handle
;
1212 pscreen
->resource_destroy
= u_transfer_helper_resource_destroy
;
1214 pscreen
->transfer_helper
= u_transfer_helper_create(&transfer_vtbl
,
1215 true, false, fake_rgtc
, true);
1217 if (!screen
->setup_slices
)
1218 screen
->setup_slices
= fd_setup_slices
;
1222 fd_get_sample_position(struct pipe_context
*context
,
1223 unsigned sample_count
, unsigned sample_index
,
1226 /* The following is copied from nouveau/nv50 except for position
1227 * values, which are taken from blob driver */
1228 static const uint8_t pos1
[1][2] = { { 0x8, 0x8 } };
1229 static const uint8_t pos2
[2][2] = {
1230 { 0xc, 0xc }, { 0x4, 0x4 } };
1231 static const uint8_t pos4
[4][2] = {
1232 { 0x6, 0x2 }, { 0xe, 0x6 },
1233 { 0x2, 0xa }, { 0xa, 0xe } };
1234 /* TODO needs to be verified on supported hw */
1235 static const uint8_t pos8
[8][2] = {
1236 { 0x9, 0x5 }, { 0x7, 0xb },
1237 { 0xd, 0x9 }, { 0x5, 0x3 },
1238 { 0x3, 0xd }, { 0x1, 0x7 },
1239 { 0xb, 0xf }, { 0xf, 0x1 } };
1241 const uint8_t (*ptr
)[2];
1243 switch (sample_count
) {
1261 pos_out
[0] = ptr
[sample_index
][0] / 16.0f
;
1262 pos_out
[1] = ptr
[sample_index
][1] / 16.0f
;
1266 fd_blit_pipe(struct pipe_context
*pctx
, const struct pipe_blit_info
*blit_info
)
1268 /* wrap fd_blit to return void */
1269 fd_blit(pctx
, blit_info
);
1273 fd_resource_context_init(struct pipe_context
*pctx
)
1275 pctx
->transfer_map
= u_transfer_helper_transfer_map
;
1276 pctx
->transfer_flush_region
= u_transfer_helper_transfer_flush_region
;
1277 pctx
->transfer_unmap
= u_transfer_helper_transfer_unmap
;
1278 pctx
->buffer_subdata
= u_default_buffer_subdata
;
1279 pctx
->texture_subdata
= u_default_texture_subdata
;
1280 pctx
->create_surface
= fd_create_surface
;
1281 pctx
->surface_destroy
= fd_surface_destroy
;
1282 pctx
->resource_copy_region
= fd_resource_copy_region
;
1283 pctx
->blit
= fd_blit_pipe
;
1284 pctx
->flush_resource
= fd_flush_resource
;
1285 pctx
->invalidate_resource
= fd_invalidate_resource
;
1286 pctx
->get_sample_position
= fd_get_sample_position
;