2 * Copyright (C) 2008 VMware, Inc.
3 * Copyright (C) 2014 Broadcom
4 * Copyright (C) 2018-2019 Alyssa Rosenzweig
5 * Copyright (C) 2019 Collabora, Ltd.
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the "Software"),
9 * to deal in the Software without restriction, including without limitation
10 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
11 * and/or sell copies of the Software, and to permit persons to whom the
12 * Software is furnished to do so, subject to the following conditions:
14 * The above copyright notice and this permission notice (including the next
15 * paragraph) shall be included in all copies or substantial portions of the
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
21 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
22 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
23 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
26 * Authors (Collabora):
27 * Tomeu Vizoso <tomeu.vizoso@collabora.com>
28 * Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
34 #include "drm-uapi/drm_fourcc.h"
36 #include "frontend/winsys_handle.h"
37 #include "util/format/u_format.h"
38 #include "util/u_memory.h"
39 #include "util/u_surface.h"
40 #include "util/u_transfer.h"
41 #include "util/u_transfer_helper.h"
42 #include "util/u_gen_mipmap.h"
45 #include "pan_context.h"
46 #include "pan_screen.h"
47 #include "pan_resource.h"
49 #include "pan_tiling.h"
51 #include "panfrost-quirks.h"
53 static struct pipe_resource
*
54 panfrost_resource_from_handle(struct pipe_screen
*pscreen
,
55 const struct pipe_resource
*templat
,
56 struct winsys_handle
*whandle
,
59 struct panfrost_device
*dev
= pan_device(pscreen
);
60 struct panfrost_resource
*rsc
;
61 struct pipe_resource
*prsc
;
63 assert(whandle
->type
== WINSYS_HANDLE_TYPE_FD
);
65 rsc
= rzalloc(pscreen
, struct panfrost_resource
);
73 pipe_reference_init(&prsc
->reference
, 1);
74 prsc
->screen
= pscreen
;
76 rsc
->bo
= panfrost_bo_import(dev
, whandle
->handle
);
77 rsc
->internal_format
= templat
->format
;
78 rsc
->layout
= MALI_TEXTURE_LINEAR
;
79 rsc
->slices
[0].stride
= whandle
->stride
;
80 rsc
->slices
[0].offset
= whandle
->offset
;
81 rsc
->slices
[0].initialized
= true;
82 panfrost_resource_set_damage_region(NULL
, &rsc
->base
, 0, NULL
);
84 if (dev
->quirks
& IS_BIFROST
&&
85 templat
->bind
& PIPE_BIND_RENDER_TARGET
) {
86 unsigned size
= panfrost_compute_checksum_size(
87 &rsc
->slices
[0], templat
->width0
, templat
->height0
);
88 rsc
->slices
[0].checksum_bo
= panfrost_bo_create(dev
, size
, 0);
89 rsc
->checksummed
= true;
94 renderonly_create_gpu_import_for_resource(prsc
, dev
->ro
, NULL
);
95 /* failure is expected in some cases.. */
102 panfrost_resource_get_handle(struct pipe_screen
*pscreen
,
103 struct pipe_context
*ctx
,
104 struct pipe_resource
*pt
,
105 struct winsys_handle
*handle
,
108 struct panfrost_device
*dev
= pan_device(pscreen
);
109 struct panfrost_resource
*rsrc
= (struct panfrost_resource
*) pt
;
110 struct renderonly_scanout
*scanout
= rsrc
->scanout
;
112 handle
->modifier
= DRM_FORMAT_MOD_INVALID
;
114 if (handle
->type
== WINSYS_HANDLE_TYPE_SHARED
) {
116 } else if (handle
->type
== WINSYS_HANDLE_TYPE_KMS
) {
117 if (renderonly_get_handle(scanout
, handle
))
120 handle
->handle
= rsrc
->bo
->gem_handle
;
121 handle
->stride
= rsrc
->slices
[0].stride
;
122 handle
->offset
= rsrc
->slices
[0].offset
;
124 } else if (handle
->type
== WINSYS_HANDLE_TYPE_FD
) {
126 struct drm_prime_handle args
= {
127 .handle
= scanout
->handle
,
128 .flags
= DRM_CLOEXEC
,
131 int ret
= drmIoctl(dev
->ro
->kms_fd
, DRM_IOCTL_PRIME_HANDLE_TO_FD
, &args
);
135 handle
->stride
= scanout
->stride
;
136 handle
->handle
= args
.fd
;
140 int fd
= panfrost_bo_export(rsrc
->bo
);
146 handle
->stride
= rsrc
->slices
[0].stride
;
147 handle
->offset
= rsrc
->slices
[0].offset
;
156 panfrost_flush_resource(struct pipe_context
*pctx
, struct pipe_resource
*prsc
)
161 static struct pipe_surface
*
162 panfrost_create_surface(struct pipe_context
*pipe
,
163 struct pipe_resource
*pt
,
164 const struct pipe_surface
*surf_tmpl
)
166 struct pipe_surface
*ps
= NULL
;
168 ps
= rzalloc(pipe
, struct pipe_surface
);
171 pipe_reference_init(&ps
->reference
, 1);
172 pipe_resource_reference(&ps
->texture
, pt
);
174 ps
->format
= surf_tmpl
->format
;
176 if (pt
->target
!= PIPE_BUFFER
) {
177 assert(surf_tmpl
->u
.tex
.level
<= pt
->last_level
);
178 ps
->width
= u_minify(pt
->width0
, surf_tmpl
->u
.tex
.level
);
179 ps
->height
= u_minify(pt
->height0
, surf_tmpl
->u
.tex
.level
);
180 ps
->nr_samples
= surf_tmpl
->nr_samples
;
181 ps
->u
.tex
.level
= surf_tmpl
->u
.tex
.level
;
182 ps
->u
.tex
.first_layer
= surf_tmpl
->u
.tex
.first_layer
;
183 ps
->u
.tex
.last_layer
= surf_tmpl
->u
.tex
.last_layer
;
185 /* setting width as number of elements should get us correct renderbuffer width */
186 ps
->width
= surf_tmpl
->u
.buf
.last_element
- surf_tmpl
->u
.buf
.first_element
+ 1;
187 ps
->height
= pt
->height0
;
188 ps
->u
.buf
.first_element
= surf_tmpl
->u
.buf
.first_element
;
189 ps
->u
.buf
.last_element
= surf_tmpl
->u
.buf
.last_element
;
190 assert(ps
->u
.buf
.first_element
<= ps
->u
.buf
.last_element
);
191 assert(ps
->u
.buf
.last_element
< ps
->width
);
199 panfrost_surface_destroy(struct pipe_context
*pipe
,
200 struct pipe_surface
*surf
)
202 assert(surf
->texture
);
203 pipe_resource_reference(&surf
->texture
, NULL
);
207 static struct pipe_resource
*
208 panfrost_create_scanout_res(struct pipe_screen
*screen
,
209 const struct pipe_resource
*template)
211 struct panfrost_device
*dev
= pan_device(screen
);
212 struct pipe_resource scanout_templat
= *template;
213 struct renderonly_scanout
*scanout
;
214 struct winsys_handle handle
;
215 struct pipe_resource
*res
;
217 scanout
= renderonly_scanout_for_resource(&scanout_templat
,
222 assert(handle
.type
== WINSYS_HANDLE_TYPE_FD
);
223 /* TODO: handle modifiers? */
224 res
= screen
->resource_from_handle(screen
, template, &handle
,
225 PIPE_HANDLE_USAGE_FRAMEBUFFER_WRITE
);
226 close(handle
.handle
);
230 struct panfrost_resource
*pres
= pan_resource(res
);
232 pres
->scanout
= scanout
;
237 /* Setup the mip tree given a particular layout, possibly with checksumming */
240 panfrost_setup_slices(struct panfrost_resource
*pres
, size_t *bo_size
)
242 struct pipe_resource
*res
= &pres
->base
;
243 unsigned width
= res
->width0
;
244 unsigned height
= res
->height0
;
245 unsigned depth
= res
->depth0
;
246 unsigned bytes_per_pixel
= util_format_get_blocksize(pres
->internal_format
);
248 /* MSAA is implemented as a 3D texture with z corresponding to the
249 * sample #, horrifyingly enough */
251 bool msaa
= res
->nr_samples
> 1;
255 depth
= res
->nr_samples
;
260 /* Tiled operates blockwise; linear is packed. Also, anything
261 * we render to has to be tile-aligned. Maybe not strictly
262 * necessary, but we're not *that* pressed for memory and it
263 * makes code a lot simpler */
265 bool renderable
= res
->bind
&
266 (PIPE_BIND_RENDER_TARGET
| PIPE_BIND_DEPTH_STENCIL
) &&
267 res
->target
!= PIPE_BUFFER
;
268 bool afbc
= pres
->layout
== MALI_TEXTURE_AFBC
;
269 bool tiled
= pres
->layout
== MALI_TEXTURE_TILED
;
270 bool should_align
= renderable
|| tiled
;
272 /* We don't know how to specify a 2D stride for 3D textures */
274 bool can_align_stride
=
275 res
->target
!= PIPE_TEXTURE_3D
;
277 should_align
&= can_align_stride
;
280 unsigned size_2d
= 0;
282 for (unsigned l
= 0; l
<= res
->last_level
; ++l
) {
283 struct panfrost_slice
*slice
= &pres
->slices
[l
];
285 unsigned effective_width
= width
;
286 unsigned effective_height
= height
;
287 unsigned effective_depth
= depth
;
290 effective_width
= ALIGN_POT(effective_width
, 16);
291 effective_height
= ALIGN_POT(effective_height
, 16);
293 /* We don't need to align depth */
296 /* Align levels to cache-line as a performance improvement for
297 * linear/tiled and as a requirement for AFBC */
299 offset
= ALIGN_POT(offset
, 64);
301 slice
->offset
= offset
;
303 /* Compute the would-be stride */
304 unsigned stride
= bytes_per_pixel
* effective_width
;
306 if (util_format_is_compressed(pres
->internal_format
))
309 /* ..but cache-line align it for performance */
310 if (can_align_stride
&& pres
->layout
== MALI_TEXTURE_LINEAR
)
311 stride
= ALIGN_POT(stride
, 64);
313 slice
->stride
= stride
;
315 unsigned slice_one_size
= slice
->stride
* effective_height
;
316 unsigned slice_full_size
= slice_one_size
* effective_depth
;
318 slice
->size0
= slice_one_size
;
320 /* Report 2D size for 3D texturing */
323 size_2d
= slice_one_size
;
325 /* Compute AFBC sizes if necessary */
328 panfrost_afbc_header_size(width
, height
);
330 offset
+= slice
->header_size
;
333 offset
+= slice_full_size
;
335 /* Add a checksum region if necessary */
336 if (pres
->checksummed
) {
337 slice
->checksum_offset
= offset
;
339 unsigned size
= panfrost_compute_checksum_size(
340 slice
, width
, height
);
345 width
= u_minify(width
, 1);
346 height
= u_minify(height
, 1);
348 /* Don't mipmap the sample count */
350 depth
= u_minify(depth
, 1);
353 assert(res
->array_size
);
355 if (res
->target
!= PIPE_TEXTURE_3D
) {
356 /* Arrays and cubemaps have the entire miptree duplicated */
358 pres
->cubemap_stride
= ALIGN_POT(offset
, 64);
359 *bo_size
= ALIGN_POT(pres
->cubemap_stride
* res
->array_size
, 4096);
361 /* 3D strides across the 2D layers */
362 assert(res
->array_size
== 1);
364 pres
->cubemap_stride
= size_2d
;
365 *bo_size
= ALIGN_POT(offset
, 4096);
370 panfrost_resource_create_bo(struct panfrost_device
*dev
, struct panfrost_resource
*pres
)
372 struct pipe_resource
*res
= &pres
->base
;
374 /* Based on the usage, figure out what storing will be used. There are
377 * Linear: the basic format, bad for memory bandwidth, bad for cache
378 * use. Zero-copy, though. Renderable.
380 * Tiled: Not compressed, but cache-optimized. Expensive to write into
381 * (due to software tiling), but cheap to sample from. Ideal for most
384 * AFBC: Compressed and renderable (so always desirable for non-scanout
385 * rendertargets). Cheap to sample from. The format is black box, so we
386 * can't read/write from software.
388 * Tiling textures is almost always faster, unless we only use it once.
389 * Only a few types of resources can be tiled, ensure the bind is only
390 * (a combination of) one of the following */
392 const unsigned valid_binding
=
393 PIPE_BIND_DEPTH_STENCIL
|
394 PIPE_BIND_RENDER_TARGET
|
395 PIPE_BIND_BLENDABLE
|
396 PIPE_BIND_SAMPLER_VIEW
|
397 PIPE_BIND_DISPLAY_TARGET
;
399 unsigned bpp
= util_format_get_blocksizebits(pres
->internal_format
);
400 bool is_2d
= (res
->target
== PIPE_TEXTURE_2D
) || (res
->target
== PIPE_TEXTURE_RECT
);
401 bool is_sane_bpp
= bpp
== 8 || bpp
== 16 || bpp
== 24 || bpp
== 32 || bpp
== 64 || bpp
== 128;
402 bool should_tile
= (res
->usage
!= PIPE_USAGE_STREAM
);
403 bool must_tile
= (res
->bind
& PIPE_BIND_DEPTH_STENCIL
) &&
404 (dev
->quirks
& (MIDGARD_SFBD
| IS_BIFROST
));
405 bool can_tile
= is_2d
&& is_sane_bpp
&& ((res
->bind
& ~valid_binding
) == 0);
407 /* FBOs we would like to checksum, if at all possible */
408 bool can_checksum
= !(res
->bind
& ~valid_binding
);
409 bool should_checksum
= res
->bind
& PIPE_BIND_RENDER_TARGET
;
411 pres
->checksummed
= can_checksum
&& should_checksum
;
413 /* Set the layout appropriately */
414 assert(!(must_tile
&& !can_tile
)); /* must_tile => can_tile */
415 pres
->layout
= ((can_tile
&& should_tile
) || must_tile
) ? MALI_TEXTURE_TILED
: MALI_TEXTURE_LINEAR
;
416 pres
->layout_constant
= must_tile
|| !can_tile
;
420 panfrost_setup_slices(pres
, &bo_size
);
422 /* We create a BO immediately but don't bother mapping, since we don't
423 * care to map e.g. FBOs which the CPU probably won't touch */
424 pres
->bo
= panfrost_bo_create(dev
, bo_size
, PAN_BO_DELAY_MMAP
);
428 panfrost_resource_set_damage_region(struct pipe_screen
*screen
,
429 struct pipe_resource
*res
,
431 const struct pipe_box
*rects
)
433 struct panfrost_resource
*pres
= pan_resource(res
);
434 struct pipe_scissor_state
*damage_extent
= &pres
->damage
.extent
;
437 if (pres
->damage
.inverted_rects
)
438 ralloc_free(pres
->damage
.inverted_rects
);
440 memset(&pres
->damage
, 0, sizeof(pres
->damage
));
442 pres
->damage
.inverted_rects
=
443 pan_subtract_damage(pres
,
444 res
->width0
, res
->height0
,
445 nrects
, rects
, &pres
->damage
.inverted_len
);
447 /* Track the damage extent: the quad including all damage regions. Will
448 * be used restrict the rendering area */
450 damage_extent
->minx
= 0xffff;
451 damage_extent
->miny
= 0xffff;
453 for (i
= 0; i
< nrects
; i
++) {
454 int x
= rects
[i
].x
, w
= rects
[i
].width
, h
= rects
[i
].height
;
455 int y
= res
->height0
- (rects
[i
].y
+ h
);
457 damage_extent
->minx
= MIN2(damage_extent
->minx
, x
);
458 damage_extent
->miny
= MIN2(damage_extent
->miny
, y
);
459 damage_extent
->maxx
= MAX2(damage_extent
->maxx
,
460 MIN2(x
+ w
, res
->width0
));
461 damage_extent
->maxy
= MAX2(damage_extent
->maxy
,
462 MIN2(y
+ h
, res
->height0
));
466 damage_extent
->minx
= 0;
467 damage_extent
->miny
= 0;
468 damage_extent
->maxx
= res
->width0
;
469 damage_extent
->maxy
= res
->height0
;
474 static struct pipe_resource
*
475 panfrost_resource_create(struct pipe_screen
*screen
,
476 const struct pipe_resource
*template)
478 struct panfrost_device
*dev
= pan_device(screen
);
480 /* Make sure we're familiar */
481 switch (template->target
) {
483 case PIPE_TEXTURE_1D
:
484 case PIPE_TEXTURE_2D
:
485 case PIPE_TEXTURE_3D
:
486 case PIPE_TEXTURE_CUBE
:
487 case PIPE_TEXTURE_RECT
:
488 case PIPE_TEXTURE_1D_ARRAY
:
489 case PIPE_TEXTURE_2D_ARRAY
:
492 unreachable("Unknown texture target\n");
495 if (dev
->ro
&& (template->bind
&
496 (PIPE_BIND_DISPLAY_TARGET
| PIPE_BIND_SCANOUT
| PIPE_BIND_SHARED
)))
497 return panfrost_create_scanout_res(screen
, template);
499 struct panfrost_resource
*so
= rzalloc(screen
, struct panfrost_resource
);
500 so
->base
= *template;
501 so
->base
.screen
= screen
;
502 so
->internal_format
= template->format
;
504 pipe_reference_init(&so
->base
.reference
, 1);
506 util_range_init(&so
->valid_buffer_range
);
508 panfrost_resource_create_bo(dev
, so
);
509 panfrost_resource_set_damage_region(NULL
, &so
->base
, 0, NULL
);
511 if (template->bind
& PIPE_BIND_INDEX_BUFFER
)
512 so
->index_cache
= rzalloc(so
, struct panfrost_minmax_cache
);
514 return (struct pipe_resource
*)so
;
518 panfrost_resource_destroy(struct pipe_screen
*screen
,
519 struct pipe_resource
*pt
)
521 struct panfrost_device
*dev
= pan_device(screen
);
522 struct panfrost_resource
*rsrc
= (struct panfrost_resource
*) pt
;
525 renderonly_scanout_destroy(rsrc
->scanout
, dev
->ro
);
528 panfrost_bo_unreference(rsrc
->bo
);
530 if (rsrc
->slices
[0].checksum_bo
)
531 panfrost_bo_unreference(rsrc
->slices
[0].checksum_bo
);
533 util_range_destroy(&rsrc
->valid_buffer_range
);
539 panfrost_transfer_map(struct pipe_context
*pctx
,
540 struct pipe_resource
*resource
,
542 unsigned usage
, /* a combination of PIPE_TRANSFER_x */
543 const struct pipe_box
*box
,
544 struct pipe_transfer
**out_transfer
)
546 struct panfrost_context
*ctx
= pan_context(pctx
);
547 struct panfrost_device
*dev
= pan_device(pctx
->screen
);
548 struct panfrost_resource
*rsrc
= pan_resource(resource
);
549 int bytes_per_pixel
= util_format_get_blocksize(rsrc
->internal_format
);
550 struct panfrost_bo
*bo
= rsrc
->bo
;
552 struct panfrost_gtransfer
*transfer
= rzalloc(pctx
, struct panfrost_gtransfer
);
553 transfer
->base
.level
= level
;
554 transfer
->base
.usage
= usage
;
555 transfer
->base
.box
= *box
;
557 pipe_resource_reference(&transfer
->base
.resource
, resource
);
559 *out_transfer
= &transfer
->base
;
561 /* If we haven't already mmaped, now's the time */
562 panfrost_bo_mmap(bo
);
564 if (dev
->debug
& (PAN_DBG_TRACE
| PAN_DBG_SYNC
))
565 pandecode_inject_mmap(bo
->gpu
, bo
->cpu
, bo
->size
, NULL
);
567 bool create_new_bo
= usage
& PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE
;
568 bool copy_resource
= false;
570 if (!create_new_bo
&&
571 !(usage
& PIPE_TRANSFER_UNSYNCHRONIZED
) &&
572 (usage
& PIPE_TRANSFER_WRITE
) &&
573 !(resource
->target
== PIPE_BUFFER
574 && !util_ranges_intersect(&rsrc
->valid_buffer_range
, box
->x
, box
->x
+ box
->width
)) &&
575 panfrost_pending_batches_access_bo(ctx
, bo
)) {
577 /* When a resource to be modified is already being used by a
578 * pending batch, it is often faster to copy the whole BO than
579 * to flush and split the frame in two. This also mostly
580 * mitigates broken depth reload.
583 panfrost_flush_batches_accessing_bo(ctx
, bo
, false);
584 panfrost_bo_wait(bo
, INT64_MAX
, false);
586 create_new_bo
= true;
587 copy_resource
= true;
591 /* If the BO is used by one of the pending batches or if it's
592 * not ready yet (still accessed by one of the already flushed
593 * batches), we try to allocate a new one to avoid waiting.
595 if (panfrost_pending_batches_access_bo(ctx
, bo
) ||
596 !panfrost_bo_wait(bo
, 0, true)) {
597 /* We want the BO to be MMAPed. */
598 uint32_t flags
= bo
->flags
& ~PAN_BO_DELAY_MMAP
;
599 struct panfrost_bo
*newbo
= NULL
;
601 /* When the BO has been imported/exported, we can't
602 * replace it by another one, otherwise the
603 * importer/exporter wouldn't see the change we're
606 if (!(bo
->flags
& PAN_BO_SHARED
))
607 newbo
= panfrost_bo_create(dev
, bo
->size
,
612 memcpy(newbo
->cpu
, rsrc
->bo
->cpu
, bo
->size
);
614 panfrost_bo_unreference(bo
);
618 /* Allocation failed or was impossible, let's
619 * fall back on a flush+wait.
621 panfrost_flush_batches_accessing_bo(ctx
, bo
, true);
622 panfrost_bo_wait(bo
, INT64_MAX
, true);
625 } else if ((usage
& PIPE_TRANSFER_WRITE
)
626 && resource
->target
== PIPE_BUFFER
627 && !util_ranges_intersect(&rsrc
->valid_buffer_range
, box
->x
, box
->x
+ box
->width
)) {
628 /* No flush for writes to uninitialized */
629 } else if (!(usage
& PIPE_TRANSFER_UNSYNCHRONIZED
)) {
630 if (usage
& PIPE_TRANSFER_WRITE
) {
631 panfrost_flush_batches_accessing_bo(ctx
, bo
, true);
632 panfrost_bo_wait(bo
, INT64_MAX
, true);
633 } else if (usage
& PIPE_TRANSFER_READ
) {
634 panfrost_flush_batches_accessing_bo(ctx
, bo
, false);
635 panfrost_bo_wait(bo
, INT64_MAX
, false);
639 if (rsrc
->layout
!= MALI_TEXTURE_LINEAR
) {
640 /* Non-linear resources need to be indirectly mapped */
642 if (usage
& PIPE_TRANSFER_MAP_DIRECTLY
)
645 transfer
->base
.stride
= box
->width
* bytes_per_pixel
;
646 transfer
->base
.layer_stride
= transfer
->base
.stride
* box
->height
;
647 transfer
->map
= ralloc_size(transfer
, transfer
->base
.layer_stride
* box
->depth
);
648 assert(box
->depth
== 1);
650 if ((usage
& PIPE_TRANSFER_READ
) && rsrc
->slices
[level
].initialized
) {
651 if (rsrc
->layout
== MALI_TEXTURE_AFBC
) {
652 unreachable("Unimplemented: reads from AFBC");
653 } else if (rsrc
->layout
== MALI_TEXTURE_TILED
) {
654 panfrost_load_tiled_image(
656 bo
->cpu
+ rsrc
->slices
[level
].offset
,
657 box
->x
, box
->y
, box
->width
, box
->height
,
658 transfer
->base
.stride
,
659 rsrc
->slices
[level
].stride
,
660 rsrc
->internal_format
);
664 return transfer
->map
;
666 /* Direct, persistent writes create holes in time for
667 * caching... I don't know if this is actually possible but we
668 * should still get it right */
670 unsigned dpw
= PIPE_TRANSFER_MAP_DIRECTLY
| PIPE_TRANSFER_WRITE
| PIPE_TRANSFER_PERSISTENT
;
672 if ((usage
& dpw
) == dpw
&& rsrc
->index_cache
)
675 transfer
->base
.stride
= rsrc
->slices
[level
].stride
;
676 transfer
->base
.layer_stride
= panfrost_get_layer_stride(
677 rsrc
->slices
, rsrc
->base
.target
== PIPE_TEXTURE_3D
,
678 rsrc
->cubemap_stride
, level
);
680 /* By mapping direct-write, we're implicitly already
681 * initialized (maybe), so be conservative */
683 if ((usage
& PIPE_TRANSFER_WRITE
) && (usage
& PIPE_TRANSFER_MAP_DIRECTLY
)) {
684 rsrc
->slices
[level
].initialized
= true;
685 panfrost_minmax_cache_invalidate(rsrc
->index_cache
, &transfer
->base
);
689 + rsrc
->slices
[level
].offset
690 + transfer
->base
.box
.z
* transfer
->base
.layer_stride
691 + transfer
->base
.box
.y
* rsrc
->slices
[level
].stride
692 + transfer
->base
.box
.x
* bytes_per_pixel
;
697 panfrost_transfer_unmap(struct pipe_context
*pctx
,
698 struct pipe_transfer
*transfer
)
700 /* Gallium expects writeback here, so we tile */
702 struct panfrost_gtransfer
*trans
= pan_transfer(transfer
);
703 struct panfrost_resource
*prsrc
= (struct panfrost_resource
*) transfer
->resource
;
705 /* Mark whatever we wrote as written */
706 if (transfer
->usage
& PIPE_TRANSFER_WRITE
)
707 prsrc
->slices
[transfer
->level
].initialized
= true;
710 struct panfrost_bo
*bo
= prsrc
->bo
;
712 if (transfer
->usage
& PIPE_TRANSFER_WRITE
) {
713 if (prsrc
->layout
== MALI_TEXTURE_AFBC
) {
714 unreachable("Unimplemented: writes to AFBC\n");
715 } else if (prsrc
->layout
== MALI_TEXTURE_TILED
) {
716 assert(transfer
->box
.depth
== 1);
718 /* Do we overwrite the entire resource? If so,
719 * we don't need an intermediate blit so it's a
720 * good time to switch the layout. */
722 bool discards_content
= prsrc
->base
.last_level
== 0
723 && transfer
->box
.width
== prsrc
->base
.width0
724 && transfer
->box
.height
== prsrc
->base
.height0
725 && transfer
->box
.x
== 0
726 && transfer
->box
.y
== 0
727 && !prsrc
->layout_constant
;
729 /* It also serves as a good heuristic for
730 * streaming textures (e.g. in video players),
731 * but we could do better */
733 if (discards_content
)
734 ++prsrc
->layout_updates
;
736 if (prsrc
->layout_updates
>= LAYOUT_CONVERT_THRESHOLD
)
738 prsrc
->layout
= MALI_TEXTURE_LINEAR
;
741 bo
->cpu
+ prsrc
->slices
[0].offset
,
743 prsrc
->slices
[0].stride
,
746 transfer
->box
.height
,
751 panfrost_store_tiled_image(
752 bo
->cpu
+ prsrc
->slices
[transfer
->level
].offset
,
754 transfer
->box
.x
, transfer
->box
.y
,
755 transfer
->box
.width
, transfer
->box
.height
,
756 prsrc
->slices
[transfer
->level
].stride
,
758 prsrc
->internal_format
);
765 util_range_add(&prsrc
->base
, &prsrc
->valid_buffer_range
,
767 transfer
->box
.x
+ transfer
->box
.width
);
769 panfrost_minmax_cache_invalidate(prsrc
->index_cache
, transfer
);
771 /* Derefence the resource */
772 pipe_resource_reference(&transfer
->resource
, NULL
);
774 /* Transfer itself is RALLOCed at the moment */
775 ralloc_free(transfer
);
779 panfrost_transfer_flush_region(struct pipe_context
*pctx
,
780 struct pipe_transfer
*transfer
,
781 const struct pipe_box
*box
)
783 struct panfrost_resource
*rsc
= pan_resource(transfer
->resource
);
785 if (transfer
->resource
->target
== PIPE_BUFFER
) {
786 util_range_add(&rsc
->base
, &rsc
->valid_buffer_range
,
787 transfer
->box
.x
+ box
->x
,
788 transfer
->box
.x
+ box
->x
+ box
->width
);
790 unsigned level
= transfer
->level
;
791 rsc
->slices
[level
].initialized
= true;
796 panfrost_invalidate_resource(struct pipe_context
*pctx
, struct pipe_resource
*prsc
)
801 static enum pipe_format
802 panfrost_resource_get_internal_format(struct pipe_resource
*rsrc
)
804 struct panfrost_resource
*prsrc
= (struct panfrost_resource
*) rsrc
;
805 return prsrc
->internal_format
;
809 panfrost_generate_mipmap(
810 struct pipe_context
*pctx
,
811 struct pipe_resource
*prsrc
,
812 enum pipe_format format
,
815 unsigned first_layer
,
818 struct panfrost_resource
*rsrc
= pan_resource(prsrc
);
820 /* Generating a mipmap invalidates the written levels, so make that
821 * explicit so we don't try to wallpaper them back and end up with
822 * u_blitter recursion */
825 for (unsigned l
= base_level
+ 1; l
<= last_level
; ++l
)
826 rsrc
->slices
[l
].initialized
= false;
828 /* Beyond that, we just delegate the hard stuff. */
830 bool blit_res
= util_gen_mipmap(
832 base_level
, last_level
,
833 first_layer
, last_layer
,
834 PIPE_TEX_FILTER_LINEAR
);
839 /* Computes the address to a texture at a particular slice */
842 panfrost_get_texture_address(
843 struct panfrost_resource
*rsrc
,
844 unsigned level
, unsigned face
, unsigned sample
)
846 bool is_3d
= rsrc
->base
.target
== PIPE_TEXTURE_3D
;
847 return rsrc
->bo
->gpu
+ panfrost_texture_offset(rsrc
->slices
, is_3d
, rsrc
->cubemap_stride
, level
, face
, sample
);
850 /* Given a resource that has already been allocated, hint that it should use a
851 * given layout. These are suggestions, not commands; it is perfectly legal to
852 * stub out this function, but there will be performance implications. */
855 panfrost_resource_hint_layout(
856 struct panfrost_device
*dev
,
857 struct panfrost_resource
*rsrc
,
858 enum mali_texture_layout layout
,
861 /* Nothing to do, although a sophisticated implementation might store
864 if (rsrc
->layout
== layout
)
867 /* We don't use the weight yet, but we should check that it's positive
868 * (semantically meaning that we should choose the given `layout`) */
873 /* Check if the preferred layout is legal for this buffer */
875 if (layout
== MALI_TEXTURE_AFBC
) {
876 bool can_afbc
= panfrost_format_supports_afbc(rsrc
->internal_format
);
877 bool is_scanout
= rsrc
->base
.bind
&
878 (PIPE_BIND_DISPLAY_TARGET
| PIPE_BIND_SCANOUT
| PIPE_BIND_SHARED
);
880 if (!can_afbc
|| is_scanout
)
884 /* Simple heuristic so far: if the resource is uninitialized, switch to
885 * the hinted layout. If it is initialized, keep the original layout.
886 * This misses some cases where it would be beneficial to switch and
889 bool is_initialized
= false;
891 for (unsigned i
= 0; i
< MAX_MIP_LEVELS
; ++i
)
892 is_initialized
|= rsrc
->slices
[i
].initialized
;
897 /* We're uninitialized, so do a layout switch. Reinitialize slices. */
900 rsrc
->layout
= layout
;
901 panfrost_setup_slices(rsrc
, &new_size
);
903 /* If we grew in size, reallocate the BO */
904 if (new_size
> rsrc
->bo
->size
) {
905 panfrost_bo_unreference(rsrc
->bo
);
906 rsrc
->bo
= panfrost_bo_create(dev
, new_size
, PAN_BO_DELAY_MMAP
);
909 /* TODO: If there are textures bound, regenerate their descriptors */
913 panfrost_resource_set_stencil(struct pipe_resource
*prsrc
,
914 struct pipe_resource
*stencil
)
916 pan_resource(prsrc
)->separate_stencil
= pan_resource(stencil
);
919 static struct pipe_resource
*
920 panfrost_resource_get_stencil(struct pipe_resource
*prsrc
)
922 return &pan_resource(prsrc
)->separate_stencil
->base
;
925 static const struct u_transfer_vtbl transfer_vtbl
= {
926 .resource_create
= panfrost_resource_create
,
927 .resource_destroy
= panfrost_resource_destroy
,
928 .transfer_map
= panfrost_transfer_map
,
929 .transfer_unmap
= panfrost_transfer_unmap
,
930 .transfer_flush_region
= panfrost_transfer_flush_region
,
931 .get_internal_format
= panfrost_resource_get_internal_format
,
932 .set_stencil
= panfrost_resource_set_stencil
,
933 .get_stencil
= panfrost_resource_get_stencil
,
937 panfrost_resource_screen_init(struct pipe_screen
*pscreen
)
939 struct panfrost_device
*dev
= pan_device(pscreen
);
941 bool fake_rgtc
= !panfrost_supports_compressed_format(dev
, MALI_BC4_UNORM
);
943 //pscreen->base.resource_create_with_modifiers =
944 // panfrost_resource_create_with_modifiers;
945 pscreen
->resource_create
= u_transfer_helper_resource_create
;
946 pscreen
->resource_destroy
= u_transfer_helper_resource_destroy
;
947 pscreen
->resource_from_handle
= panfrost_resource_from_handle
;
948 pscreen
->resource_get_handle
= panfrost_resource_get_handle
;
949 pscreen
->transfer_helper
= u_transfer_helper_create(&transfer_vtbl
,
955 panfrost_resource_context_init(struct pipe_context
*pctx
)
957 pctx
->transfer_map
= u_transfer_helper_transfer_map
;
958 pctx
->transfer_unmap
= u_transfer_helper_transfer_unmap
;
959 pctx
->create_surface
= panfrost_create_surface
;
960 pctx
->surface_destroy
= panfrost_surface_destroy
;
961 pctx
->resource_copy_region
= util_resource_copy_region
;
962 pctx
->blit
= panfrost_blit
;
963 pctx
->generate_mipmap
= panfrost_generate_mipmap
;
964 pctx
->flush_resource
= panfrost_flush_resource
;
965 pctx
->invalidate_resource
= panfrost_invalidate_resource
;
966 pctx
->transfer_flush_region
= u_transfer_helper_transfer_flush_region
;
967 pctx
->buffer_subdata
= u_default_buffer_subdata
;
968 pctx
->texture_subdata
= u_default_texture_subdata
;