1 /* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
4 * Copyright (C) 2012 Rob Clark <robclark@freedesktop.org>
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
26 * Rob Clark <robclark@freedesktop.org>
29 #include "util/u_format.h"
30 #include "util/u_format_rgtc.h"
31 #include "util/u_format_zs.h"
32 #include "util/u_inlines.h"
33 #include "util/u_transfer.h"
34 #include "util/u_string.h"
35 #include "util/u_surface.h"
38 #include "freedreno_resource.h"
39 #include "freedreno_batch_cache.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"
48 /* XXX this should go away, needed for 'struct winsys_handle' */
49 #include "state_tracker/drm_driver.h"
52 * Go through the entire state and see if the resource is bound
53 * anywhere. If it is, mark the relevant state as dirty. This is
54 * called on realloc_bo to ensure the neccessary state is re-
55 * emitted so the GPU looks at the new backing bo.
58 rebind_resource(struct fd_context
*ctx
, struct pipe_resource
*prsc
)
61 for (unsigned i
= 0; i
< ctx
->vtx
.vertexbuf
.count
&& !(ctx
->dirty
& FD_DIRTY_VTXBUF
); i
++) {
62 if (ctx
->vtx
.vertexbuf
.vb
[i
].buffer
.resource
== prsc
)
63 ctx
->dirty
|= FD_DIRTY_VTXBUF
;
66 /* per-shader-stage resources: */
67 for (unsigned stage
= 0; stage
< PIPE_SHADER_TYPES
; stage
++) {
68 /* Constbufs.. note that constbuf[0] is normal uniforms emitted in
69 * cmdstream rather than by pointer..
71 const unsigned num_ubos
= util_last_bit(ctx
->constbuf
[stage
].enabled_mask
);
72 for (unsigned i
= 1; i
< num_ubos
; i
++) {
73 if (ctx
->dirty_shader
[stage
] & FD_DIRTY_SHADER_CONST
)
75 if (ctx
->constbuf
[stage
].cb
[i
].buffer
== prsc
)
76 ctx
->dirty_shader
[stage
] |= FD_DIRTY_SHADER_CONST
;
80 for (unsigned i
= 0; i
< ctx
->tex
[stage
].num_textures
; i
++) {
81 if (ctx
->dirty_shader
[stage
] & FD_DIRTY_SHADER_TEX
)
83 if (ctx
->tex
[stage
].textures
[i
] && (ctx
->tex
[stage
].textures
[i
]->texture
== prsc
))
84 ctx
->dirty_shader
[stage
] |= FD_DIRTY_SHADER_TEX
;
88 const unsigned num_ssbos
= util_last_bit(ctx
->shaderbuf
[stage
].enabled_mask
);
89 for (unsigned i
= 0; i
< num_ssbos
; i
++) {
90 if (ctx
->dirty_shader
[stage
] & FD_DIRTY_SHADER_SSBO
)
92 if (ctx
->shaderbuf
[stage
].sb
[i
].buffer
== prsc
)
93 ctx
->dirty_shader
[stage
] |= FD_DIRTY_SHADER_SSBO
;
99 realloc_bo(struct fd_resource
*rsc
, uint32_t size
)
101 struct fd_screen
*screen
= fd_screen(rsc
->base
.b
.screen
);
102 uint32_t flags
= DRM_FREEDRENO_GEM_CACHE_WCOMBINE
|
103 DRM_FREEDRENO_GEM_TYPE_KMEM
; /* TODO */
105 /* if we start using things other than write-combine,
106 * be sure to check for PIPE_RESOURCE_FLAG_MAP_COHERENT
112 rsc
->bo
= fd_bo_new(screen
->dev
, size
, flags
);
113 util_range_set_empty(&rsc
->valid_buffer_range
);
114 fd_bc_invalidate_resource(rsc
, true);
118 do_blit(struct fd_context
*ctx
, const struct pipe_blit_info
*blit
, bool fallback
)
120 /* TODO size threshold too?? */
121 if ((blit
->src
.resource
->target
!= PIPE_BUFFER
) && !fallback
) {
122 /* do blit on gpu: */
123 fd_blitter_pipe_begin(ctx
, false, true, FD_STAGE_BLIT
);
124 util_blitter_blit(ctx
->blitter
, blit
);
125 fd_blitter_pipe_end(ctx
);
127 /* do blit on cpu: */
128 util_resource_copy_region(&ctx
->base
,
129 blit
->dst
.resource
, blit
->dst
.level
, blit
->dst
.box
.x
,
130 blit
->dst
.box
.y
, blit
->dst
.box
.z
,
131 blit
->src
.resource
, blit
->src
.level
, &blit
->src
.box
);
136 fd_try_shadow_resource(struct fd_context
*ctx
, struct fd_resource
*rsc
,
137 unsigned level
, unsigned usage
, const struct pipe_box
*box
)
139 struct pipe_context
*pctx
= &ctx
->base
;
140 struct pipe_resource
*prsc
= &rsc
->base
.b
;
141 bool fallback
= false;
146 /* TODO: somehow munge dimensions and format to copy unsupported
147 * render target format to something that is supported?
149 if (!pctx
->screen
->is_format_supported(pctx
->screen
,
150 prsc
->format
, prsc
->target
, prsc
->nr_samples
,
151 PIPE_BIND_RENDER_TARGET
))
154 /* these cases should be handled elsewhere.. just for future
155 * reference in case this gets split into a more generic(ish)
158 debug_assert(!(usage
& PIPE_TRANSFER_READ
));
159 debug_assert(!(usage
& PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE
));
161 /* if we do a gpu blit to clone the whole resource, we'll just
162 * end up stalling on that.. so only allow if we can discard
163 * current range (and blit, possibly cpu or gpu, the rest)
165 if (!(usage
& PIPE_TRANSFER_DISCARD_RANGE
))
168 bool whole_level
= util_texrange_covers_whole_level(prsc
, level
,
169 box
->x
, box
->y
, box
->z
, box
->width
, box
->height
, box
->depth
);
171 /* TODO need to be more clever about current level */
172 if ((prsc
->target
>= PIPE_TEXTURE_2D
) && !whole_level
)
175 struct pipe_resource
*pshadow
=
176 pctx
->screen
->resource_create(pctx
->screen
, prsc
);
181 assert(!ctx
->in_shadow
);
182 ctx
->in_shadow
= true;
184 /* get rid of any references that batch-cache might have to us (which
185 * should empty/destroy rsc->batches hashset)
187 fd_bc_invalidate_resource(rsc
, false);
189 mtx_lock(&ctx
->screen
->lock
);
191 /* Swap the backing bo's, so shadow becomes the old buffer,
192 * blit from shadow to new buffer. From here on out, we
195 * Note that we need to do it in this order, otherwise if
196 * we go down cpu blit path, the recursive transfer_map()
197 * sees the wrong status..
199 struct fd_resource
*shadow
= fd_resource(pshadow
);
201 DBG("shadow: %p (%d) -> %p (%d)\n", rsc
, rsc
->base
.b
.reference
.count
,
202 shadow
, shadow
->base
.b
.reference
.count
);
204 /* TODO valid_buffer_range?? */
205 swap(rsc
->bo
, shadow
->bo
);
206 swap(rsc
->write_batch
, shadow
->write_batch
);
208 /* at this point, the newly created shadow buffer is not referenced
209 * by any batches, but the existing rsc (probably) is. We need to
210 * transfer those references over:
212 debug_assert(shadow
->batch_mask
== 0);
213 struct fd_batch
*batch
;
214 foreach_batch(batch
, &ctx
->screen
->batch_cache
, rsc
->batch_mask
) {
215 struct set_entry
*entry
= _mesa_set_search(batch
->resources
, rsc
);
216 _mesa_set_remove(batch
->resources
, entry
);
217 _mesa_set_add(batch
->resources
, shadow
);
219 swap(rsc
->batch_mask
, shadow
->batch_mask
);
221 mtx_unlock(&ctx
->screen
->lock
);
223 struct pipe_blit_info blit
= {0};
224 blit
.dst
.resource
= prsc
;
225 blit
.dst
.format
= prsc
->format
;
226 blit
.src
.resource
= pshadow
;
227 blit
.src
.format
= pshadow
->format
;
228 blit
.mask
= util_format_get_mask(prsc
->format
);
229 blit
.filter
= PIPE_TEX_FILTER_NEAREST
;
231 #define set_box(field, val) do { \
232 blit.dst.field = (val); \
233 blit.src.field = (val); \
236 /* blit the other levels in their entirety: */
237 for (unsigned l
= 0; l
<= prsc
->last_level
; l
++) {
241 /* just blit whole level: */
243 set_box(box
.width
, u_minify(prsc
->width0
, l
));
244 set_box(box
.height
, u_minify(prsc
->height0
, l
));
245 set_box(box
.depth
, u_minify(prsc
->depth0
, l
));
247 do_blit(ctx
, &blit
, fallback
);
250 /* deal w/ current level specially, since we might need to split
251 * it up into a couple blits:
254 set_box(level
, level
);
256 switch (prsc
->target
) {
258 case PIPE_TEXTURE_1D
:
261 set_box(box
.height
, 1);
262 set_box(box
.depth
, 1);
266 set_box(box
.width
, box
->x
);
268 do_blit(ctx
, &blit
, fallback
);
270 if ((box
->x
+ box
->width
) < u_minify(prsc
->width0
, level
)) {
271 set_box(box
.x
, box
->x
+ box
->width
);
272 set_box(box
.width
, u_minify(prsc
->width0
, level
) - (box
->x
+ box
->width
));
274 do_blit(ctx
, &blit
, fallback
);
277 case PIPE_TEXTURE_2D
:
284 ctx
->in_shadow
= false;
286 pipe_resource_reference(&pshadow
, NULL
);
292 fd_resource_layer_offset(struct fd_resource
*rsc
,
293 struct fd_resource_slice
*slice
,
296 if (rsc
->layer_first
)
297 return layer
* rsc
->layer_size
;
299 return layer
* slice
->size0
;
303 fd_resource_flush_z32s8(struct fd_transfer
*trans
, const struct pipe_box
*box
)
305 struct fd_resource
*rsc
= fd_resource(trans
->base
.resource
);
306 struct fd_resource_slice
*slice
= fd_resource_slice(rsc
, trans
->base
.level
);
307 struct fd_resource_slice
*sslice
= fd_resource_slice(rsc
->stencil
, trans
->base
.level
);
308 enum pipe_format format
= trans
->base
.resource
->format
;
310 float *depth
= fd_bo_map(rsc
->bo
) + slice
->offset
+
311 fd_resource_layer_offset(rsc
, slice
, trans
->base
.box
.z
) +
312 (trans
->base
.box
.y
+ box
->y
) * slice
->pitch
* 4 + (trans
->base
.box
.x
+ box
->x
) * 4;
313 uint8_t *stencil
= fd_bo_map(rsc
->stencil
->bo
) + sslice
->offset
+
314 fd_resource_layer_offset(rsc
->stencil
, sslice
, trans
->base
.box
.z
) +
315 (trans
->base
.box
.y
+ box
->y
) * sslice
->pitch
+ trans
->base
.box
.x
+ box
->x
;
317 if (format
!= PIPE_FORMAT_X32_S8X24_UINT
)
318 util_format_z32_float_s8x24_uint_unpack_z_float(
319 depth
, slice
->pitch
* 4,
320 trans
->staging
, trans
->base
.stride
,
321 box
->width
, box
->height
);
323 util_format_z32_float_s8x24_uint_unpack_s_8uint(
324 stencil
, sslice
->pitch
,
325 trans
->staging
, trans
->base
.stride
,
326 box
->width
, box
->height
);
330 fd_resource_flush_rgtc(struct fd_transfer
*trans
, const struct pipe_box
*box
)
332 struct fd_resource
*rsc
= fd_resource(trans
->base
.resource
);
333 struct fd_resource_slice
*slice
= fd_resource_slice(rsc
, trans
->base
.level
);
334 enum pipe_format format
= trans
->base
.resource
->format
;
336 uint8_t *data
= fd_bo_map(rsc
->bo
) + slice
->offset
+
337 fd_resource_layer_offset(rsc
, slice
, trans
->base
.box
.z
) +
338 ((trans
->base
.box
.y
+ box
->y
) * slice
->pitch
+
339 trans
->base
.box
.x
+ box
->x
) * rsc
->cpp
;
341 uint8_t *source
= trans
->staging
+
342 util_format_get_nblocksy(format
, box
->y
) * trans
->base
.stride
+
343 util_format_get_stride(format
, box
->x
);
346 case PIPE_FORMAT_RGTC1_UNORM
:
347 case PIPE_FORMAT_RGTC1_SNORM
:
348 case PIPE_FORMAT_LATC1_UNORM
:
349 case PIPE_FORMAT_LATC1_SNORM
:
350 util_format_rgtc1_unorm_unpack_rgba_8unorm(
351 data
, slice
->pitch
* rsc
->cpp
,
352 source
, trans
->base
.stride
,
353 box
->width
, box
->height
);
355 case PIPE_FORMAT_RGTC2_UNORM
:
356 case PIPE_FORMAT_RGTC2_SNORM
:
357 case PIPE_FORMAT_LATC2_UNORM
:
358 case PIPE_FORMAT_LATC2_SNORM
:
359 util_format_rgtc2_unorm_unpack_rgba_8unorm(
360 data
, slice
->pitch
* rsc
->cpp
,
361 source
, trans
->base
.stride
,
362 box
->width
, box
->height
);
365 assert(!"Unexpected format\n");
371 fd_resource_flush(struct fd_transfer
*trans
, const struct pipe_box
*box
)
373 enum pipe_format format
= trans
->base
.resource
->format
;
376 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
:
377 case PIPE_FORMAT_X32_S8X24_UINT
:
378 fd_resource_flush_z32s8(trans
, box
);
380 case PIPE_FORMAT_RGTC1_UNORM
:
381 case PIPE_FORMAT_RGTC1_SNORM
:
382 case PIPE_FORMAT_RGTC2_UNORM
:
383 case PIPE_FORMAT_RGTC2_SNORM
:
384 case PIPE_FORMAT_LATC1_UNORM
:
385 case PIPE_FORMAT_LATC1_SNORM
:
386 case PIPE_FORMAT_LATC2_UNORM
:
387 case PIPE_FORMAT_LATC2_SNORM
:
388 fd_resource_flush_rgtc(trans
, box
);
391 assert(!"Unexpected staging transfer type");
396 static void fd_resource_transfer_flush_region(struct pipe_context
*pctx
,
397 struct pipe_transfer
*ptrans
,
398 const struct pipe_box
*box
)
400 struct fd_resource
*rsc
= fd_resource(ptrans
->resource
);
401 struct fd_transfer
*trans
= fd_transfer(ptrans
);
403 if (ptrans
->resource
->target
== PIPE_BUFFER
)
404 util_range_add(&rsc
->valid_buffer_range
,
405 ptrans
->box
.x
+ box
->x
,
406 ptrans
->box
.x
+ box
->x
+ box
->width
);
409 fd_resource_flush(trans
, box
);
413 fd_resource_transfer_unmap(struct pipe_context
*pctx
,
414 struct pipe_transfer
*ptrans
)
416 struct fd_context
*ctx
= fd_context(pctx
);
417 struct fd_resource
*rsc
= fd_resource(ptrans
->resource
);
418 struct fd_transfer
*trans
= fd_transfer(ptrans
);
420 if (trans
->staging
&& !(ptrans
->usage
& PIPE_TRANSFER_FLUSH_EXPLICIT
)) {
422 u_box_2d(0, 0, ptrans
->box
.width
, ptrans
->box
.height
, &box
);
423 fd_resource_flush(trans
, &box
);
426 if (!(ptrans
->usage
& PIPE_TRANSFER_UNSYNCHRONIZED
)) {
427 fd_bo_cpu_fini(rsc
->bo
);
429 fd_bo_cpu_fini(rsc
->stencil
->bo
);
432 util_range_add(&rsc
->valid_buffer_range
,
434 ptrans
->box
.x
+ ptrans
->box
.width
);
436 pipe_resource_reference(&ptrans
->resource
, NULL
);
437 slab_free(&ctx
->transfer_pool
, ptrans
);
439 free(trans
->staging
);
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 if (ctx
->in_shadow
&& !(usage
& PIPE_TRANSFER_READ
))
479 usage
|= PIPE_TRANSFER_UNSYNCHRONIZED
;
481 if (usage
& PIPE_TRANSFER_READ
)
482 op
|= DRM_FREEDRENO_PREP_READ
;
484 if (usage
& PIPE_TRANSFER_WRITE
)
485 op
|= DRM_FREEDRENO_PREP_WRITE
;
487 if (usage
& PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE
) {
488 realloc_bo(rsc
, fd_bo_size(rsc
->bo
));
490 realloc_bo(rsc
->stencil
, fd_bo_size(rsc
->stencil
->bo
));
491 rebind_resource(ctx
, prsc
);
492 } else if ((usage
& PIPE_TRANSFER_WRITE
) &&
493 prsc
->target
== PIPE_BUFFER
&&
494 !util_ranges_intersect(&rsc
->valid_buffer_range
,
495 box
->x
, box
->x
+ box
->width
)) {
496 /* We are trying to write to a previously uninitialized range. No need
499 } else if (!(usage
& PIPE_TRANSFER_UNSYNCHRONIZED
)) {
500 struct fd_batch
*write_batch
= NULL
;
502 /* hold a reference, so it doesn't disappear under us: */
503 fd_batch_reference(&write_batch
, rsc
->write_batch
);
505 if ((usage
& PIPE_TRANSFER_WRITE
) && write_batch
&&
506 write_batch
->back_blit
) {
507 /* if only thing pending is a back-blit, we can discard it: */
508 fd_batch_reset(write_batch
);
511 /* If the GPU is writing to the resource, or if it is reading from the
512 * resource and we're trying to write to it, flush the renders.
514 bool needs_flush
= pending(rsc
, !!(usage
& PIPE_TRANSFER_WRITE
));
515 bool busy
= needs_flush
|| (0 != fd_bo_cpu_prep(rsc
->bo
,
516 ctx
->pipe
, op
| DRM_FREEDRENO_PREP_NOSYNC
));
518 /* if we need to flush/stall, see if we can make a shadow buffer
521 * TODO we could go down this path !reorder && !busy_for_read
522 * ie. we only *don't* want to go down this path if the blit
523 * will trigger a flush!
525 if (ctx
->screen
->reorder
&& busy
&& !(usage
& PIPE_TRANSFER_READ
)) {
526 if (fd_try_shadow_resource(ctx
, rsc
, level
, usage
, box
)) {
527 needs_flush
= busy
= false;
528 rebind_resource(ctx
, prsc
);
533 if (usage
& PIPE_TRANSFER_WRITE
) {
534 struct fd_batch
*batch
, *last_batch
= NULL
;
535 foreach_batch(batch
, &ctx
->screen
->batch_cache
, rsc
->batch_mask
) {
536 fd_batch_reference(&last_batch
, batch
);
537 fd_batch_flush(batch
, false);
540 fd_batch_sync(last_batch
);
541 fd_batch_reference(&last_batch
, NULL
);
543 assert(rsc
->batch_mask
== 0);
545 fd_batch_flush(write_batch
, true);
547 assert(!rsc
->write_batch
);
550 fd_batch_reference(&write_batch
, NULL
);
552 /* The GPU keeps track of how the various bo's are being used, and
553 * will wait if necessary for the proper operation to have
557 ret
= fd_bo_cpu_prep(rsc
->bo
, ctx
->pipe
, op
);
563 buf
= fd_bo_map(rsc
->bo
);
567 offset
= slice
->offset
+
568 box
->y
/ util_format_get_blockheight(format
) * ptrans
->stride
+
569 box
->x
/ util_format_get_blockwidth(format
) * rsc
->cpp
+
570 fd_resource_layer_offset(rsc
, slice
, box
->z
);
572 if (prsc
->format
== PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
||
573 prsc
->format
== PIPE_FORMAT_X32_S8X24_UINT
) {
574 assert(trans
->base
.box
.depth
== 1);
576 trans
->base
.stride
= trans
->base
.box
.width
* rsc
->cpp
* 2;
577 trans
->staging
= malloc(trans
->base
.stride
* trans
->base
.box
.height
);
581 /* if we're not discarding the whole range (or resource), we must copy
584 if (!(usage
& (PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE
|
585 PIPE_TRANSFER_DISCARD_RANGE
))) {
586 struct fd_resource_slice
*sslice
=
587 fd_resource_slice(rsc
->stencil
, level
);
588 void *sbuf
= fd_bo_map(rsc
->stencil
->bo
);
592 float *depth
= (float *)(buf
+ slice
->offset
+
593 fd_resource_layer_offset(rsc
, slice
, box
->z
) +
594 box
->y
* slice
->pitch
* 4 + box
->x
* 4);
595 uint8_t *stencil
= sbuf
+ sslice
->offset
+
596 fd_resource_layer_offset(rsc
->stencil
, sslice
, box
->z
) +
597 box
->y
* sslice
->pitch
+ box
->x
;
599 if (format
!= PIPE_FORMAT_X32_S8X24_UINT
)
600 util_format_z32_float_s8x24_uint_pack_z_float(
601 trans
->staging
, trans
->base
.stride
,
602 depth
, slice
->pitch
* 4,
603 box
->width
, box
->height
);
605 util_format_z32_float_s8x24_uint_pack_s_8uint(
606 trans
->staging
, trans
->base
.stride
,
607 stencil
, sslice
->pitch
,
608 box
->width
, box
->height
);
611 buf
= trans
->staging
;
613 } else if (rsc
->internal_format
!= format
&&
614 util_format_description(format
)->layout
== UTIL_FORMAT_LAYOUT_RGTC
) {
615 assert(trans
->base
.box
.depth
== 1);
617 trans
->base
.stride
= util_format_get_stride(
618 format
, trans
->base
.box
.width
);
619 trans
->staging
= malloc(
620 util_format_get_2d_size(format
, trans
->base
.stride
,
621 trans
->base
.box
.height
));
625 /* if we're not discarding the whole range (or resource), we must copy
628 if (!(usage
& (PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE
|
629 PIPE_TRANSFER_DISCARD_RANGE
))) {
630 uint8_t *rgba8
= (uint8_t *)buf
+ slice
->offset
+
631 fd_resource_layer_offset(rsc
, slice
, box
->z
) +
632 box
->y
* slice
->pitch
* rsc
->cpp
+ box
->x
* rsc
->cpp
;
635 case PIPE_FORMAT_RGTC1_UNORM
:
636 case PIPE_FORMAT_RGTC1_SNORM
:
637 case PIPE_FORMAT_LATC1_UNORM
:
638 case PIPE_FORMAT_LATC1_SNORM
:
639 util_format_rgtc1_unorm_pack_rgba_8unorm(
640 trans
->staging
, trans
->base
.stride
,
641 rgba8
, slice
->pitch
* rsc
->cpp
,
642 box
->width
, box
->height
);
644 case PIPE_FORMAT_RGTC2_UNORM
:
645 case PIPE_FORMAT_RGTC2_SNORM
:
646 case PIPE_FORMAT_LATC2_UNORM
:
647 case PIPE_FORMAT_LATC2_SNORM
:
648 util_format_rgtc2_unorm_pack_rgba_8unorm(
649 trans
->staging
, trans
->base
.stride
,
650 rgba8
, slice
->pitch
* rsc
->cpp
,
651 box
->width
, box
->height
);
654 assert(!"Unexpected format");
659 buf
= trans
->staging
;
668 fd_resource_transfer_unmap(pctx
, ptrans
);
673 fd_resource_destroy(struct pipe_screen
*pscreen
,
674 struct pipe_resource
*prsc
)
676 struct fd_resource
*rsc
= fd_resource(prsc
);
677 fd_bc_invalidate_resource(rsc
, true);
680 util_range_destroy(&rsc
->valid_buffer_range
);
685 fd_resource_get_handle(struct pipe_screen
*pscreen
,
686 struct pipe_resource
*prsc
,
687 struct winsys_handle
*handle
)
689 struct fd_resource
*rsc
= fd_resource(prsc
);
691 return fd_screen_bo_get_handle(pscreen
, rsc
->bo
,
692 rsc
->slices
[0].pitch
* rsc
->cpp
, handle
);
696 static const struct u_resource_vtbl fd_resource_vtbl
= {
697 .resource_get_handle
= fd_resource_get_handle
,
698 .resource_destroy
= fd_resource_destroy
,
699 .transfer_map
= fd_resource_transfer_map
,
700 .transfer_flush_region
= fd_resource_transfer_flush_region
,
701 .transfer_unmap
= fd_resource_transfer_unmap
,
705 setup_slices(struct fd_resource
*rsc
, uint32_t alignment
, enum pipe_format format
)
707 struct pipe_resource
*prsc
= &rsc
->base
.b
;
708 struct fd_screen
*screen
= fd_screen(prsc
->screen
);
709 enum util_format_layout layout
= util_format_description(format
)->layout
;
710 uint32_t pitchalign
= screen
->gmem_alignw
;
711 uint32_t level
, size
= 0;
712 uint32_t width
= prsc
->width0
;
713 uint32_t height
= prsc
->height0
;
714 uint32_t depth
= prsc
->depth0
;
715 /* in layer_first layout, the level (slice) contains just one
716 * layer (since in fact the layer contains the slices)
718 uint32_t layers_in_level
= rsc
->layer_first
? 1 : prsc
->array_size
;
720 if (is_a5xx(screen
) && (rsc
->base
.b
.target
>= PIPE_TEXTURE_2D
))
721 height
= align(height
, screen
->gmem_alignh
);
723 for (level
= 0; level
<= prsc
->last_level
; level
++) {
724 struct fd_resource_slice
*slice
= fd_resource_slice(rsc
, level
);
727 if (layout
== UTIL_FORMAT_LAYOUT_ASTC
)
728 slice
->pitch
= width
=
729 util_align_npot(width
, pitchalign
* util_format_get_blockwidth(format
));
731 slice
->pitch
= width
= align(width
, pitchalign
);
732 slice
->offset
= size
;
733 blocks
= util_format_get_nblocks(format
, width
, height
);
734 /* 1d array and 2d array textures must all have the same layer size
735 * for each miplevel on a3xx. 3d textures can have different layer
736 * sizes for high levels, but the hw auto-sizer is buggy (or at least
737 * different than what this code does), so as soon as the layer size
738 * range gets into range, we stop reducing it.
740 if (prsc
->target
== PIPE_TEXTURE_3D
&& (
742 (level
> 1 && rsc
->slices
[level
- 1].size0
> 0xf000)))
743 slice
->size0
= align(blocks
* rsc
->cpp
, alignment
);
744 else if (level
== 0 || rsc
->layer_first
|| alignment
== 1)
745 slice
->size0
= align(blocks
* rsc
->cpp
, alignment
);
747 slice
->size0
= rsc
->slices
[level
- 1].size0
;
749 size
+= slice
->size0
* depth
* layers_in_level
;
751 width
= u_minify(width
, 1);
752 height
= u_minify(height
, 1);
753 depth
= u_minify(depth
, 1);
760 slice_alignment(struct pipe_screen
*pscreen
, const struct pipe_resource
*tmpl
)
762 /* on a3xx, 2d array and 3d textures seem to want their
763 * layers aligned to page boundaries:
765 switch (tmpl
->target
) {
766 case PIPE_TEXTURE_3D
:
767 case PIPE_TEXTURE_1D_ARRAY
:
768 case PIPE_TEXTURE_2D_ARRAY
:
775 /* special case to resize query buf after allocated.. */
777 fd_resource_resize(struct pipe_resource
*prsc
, uint32_t sz
)
779 struct fd_resource
*rsc
= fd_resource(prsc
);
781 debug_assert(prsc
->width0
== 0);
782 debug_assert(prsc
->target
== PIPE_BUFFER
);
783 debug_assert(prsc
->bind
== PIPE_BIND_QUERY_BUFFER
);
786 realloc_bo(rsc
, setup_slices(rsc
, 1, prsc
->format
));
789 // TODO common helper?
791 has_depth(enum pipe_format format
)
794 case PIPE_FORMAT_Z16_UNORM
:
795 case PIPE_FORMAT_Z32_UNORM
:
796 case PIPE_FORMAT_Z32_FLOAT
:
797 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
:
798 case PIPE_FORMAT_Z24_UNORM_S8_UINT
:
799 case PIPE_FORMAT_S8_UINT_Z24_UNORM
:
800 case PIPE_FORMAT_Z24X8_UNORM
:
801 case PIPE_FORMAT_X8Z24_UNORM
:
809 * Create a new texture object, using the given template info.
811 static struct pipe_resource
*
812 fd_resource_create(struct pipe_screen
*pscreen
,
813 const struct pipe_resource
*tmpl
)
815 struct fd_screen
*screen
= fd_screen(pscreen
);
816 struct fd_resource
*rsc
= CALLOC_STRUCT(fd_resource
);
817 struct pipe_resource
*prsc
= &rsc
->base
.b
;
818 enum pipe_format format
= tmpl
->format
;
819 uint32_t size
, alignment
;
821 DBG("%p: target=%d, format=%s, %ux%ux%u, array_size=%u, last_level=%u, "
822 "nr_samples=%u, usage=%u, bind=%x, flags=%x", prsc
,
823 tmpl
->target
, util_format_name(format
),
824 tmpl
->width0
, tmpl
->height0
, tmpl
->depth0
,
825 tmpl
->array_size
, tmpl
->last_level
, tmpl
->nr_samples
,
826 tmpl
->usage
, tmpl
->bind
, tmpl
->flags
);
833 pipe_reference_init(&prsc
->reference
, 1);
835 prsc
->screen
= pscreen
;
837 util_range_init(&rsc
->valid_buffer_range
);
839 rsc
->base
.vtbl
= &fd_resource_vtbl
;
841 if (format
== PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
)
842 format
= PIPE_FORMAT_Z32_FLOAT
;
843 else if (screen
->gpu_id
< 400 &&
844 util_format_description(format
)->layout
== UTIL_FORMAT_LAYOUT_RGTC
)
845 format
= PIPE_FORMAT_R8G8B8A8_UNORM
;
846 rsc
->internal_format
= format
;
847 rsc
->cpp
= util_format_get_blocksize(format
);
851 // XXX probably need some extra work if we hit rsc shadowing path w/ lrz..
852 if (is_a5xx(screen
) && (fd_mesa_debug
& FD_DBG_LRZ
) && has_depth(format
)) {
853 const uint32_t flags
= DRM_FREEDRENO_GEM_CACHE_WCOMBINE
|
854 DRM_FREEDRENO_GEM_TYPE_KMEM
; /* TODO */
855 unsigned lrz_pitch
= align(DIV_ROUND_UP(tmpl
->width0
, 8), 32);
856 unsigned lrz_height
= DIV_ROUND_UP(tmpl
->height0
, 8);
857 unsigned size
= lrz_pitch
* lrz_height
* 2;
859 size
+= 0x1000; /* for GRAS_LRZ_FAST_CLEAR_BUFFER */
861 rsc
->lrz_height
= lrz_height
;
862 rsc
->lrz_width
= lrz_pitch
;
863 rsc
->lrz_pitch
= lrz_pitch
;
864 rsc
->lrz
= fd_bo_new(screen
->dev
, size
, flags
);
867 alignment
= slice_alignment(pscreen
, tmpl
);
868 if (is_a4xx(screen
) || is_a5xx(screen
)) {
869 switch (tmpl
->target
) {
870 case PIPE_TEXTURE_3D
:
871 rsc
->layer_first
= false;
874 rsc
->layer_first
= true;
880 size
= setup_slices(rsc
, alignment
, format
);
882 /* special case for hw-query buffer, which we need to allocate before we
886 /* note, semi-intention == instead of & */
887 debug_assert(prsc
->bind
== PIPE_BIND_QUERY_BUFFER
);
891 if (rsc
->layer_first
) {
892 rsc
->layer_size
= align(size
, 4096);
893 size
= rsc
->layer_size
* prsc
->array_size
;
896 realloc_bo(rsc
, size
);
900 /* There is no native Z32F_S8 sampling or rendering format, so this must
901 * be emulated via two separate textures. The depth texture still keeps
902 * its Z32F_S8 format though, and we also keep a reference to a separate
905 if (tmpl
->format
== PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
) {
906 struct pipe_resource stencil
= *tmpl
;
907 stencil
.format
= PIPE_FORMAT_S8_UINT
;
908 rsc
->stencil
= fd_resource(fd_resource_create(pscreen
, &stencil
));
915 fd_resource_destroy(pscreen
, prsc
);
920 * Create a texture from a winsys_handle. The handle is often created in
921 * another process by first creating a pipe texture and then calling
922 * resource_get_handle.
924 static struct pipe_resource
*
925 fd_resource_from_handle(struct pipe_screen
*pscreen
,
926 const struct pipe_resource
*tmpl
,
927 struct winsys_handle
*handle
, unsigned usage
)
929 struct fd_resource
*rsc
= CALLOC_STRUCT(fd_resource
);
930 struct fd_resource_slice
*slice
= &rsc
->slices
[0];
931 struct pipe_resource
*prsc
= &rsc
->base
.b
;
932 uint32_t pitchalign
= fd_screen(pscreen
)->gmem_alignw
;
934 DBG("target=%d, format=%s, %ux%ux%u, array_size=%u, last_level=%u, "
935 "nr_samples=%u, usage=%u, bind=%x, flags=%x",
936 tmpl
->target
, util_format_name(tmpl
->format
),
937 tmpl
->width0
, tmpl
->height0
, tmpl
->depth0
,
938 tmpl
->array_size
, tmpl
->last_level
, tmpl
->nr_samples
,
939 tmpl
->usage
, tmpl
->bind
, tmpl
->flags
);
946 pipe_reference_init(&prsc
->reference
, 1);
948 prsc
->screen
= pscreen
;
950 util_range_init(&rsc
->valid_buffer_range
);
952 rsc
->bo
= fd_screen_bo_from_handle(pscreen
, handle
);
956 rsc
->base
.vtbl
= &fd_resource_vtbl
;
957 rsc
->cpp
= util_format_get_blocksize(tmpl
->format
);
958 slice
->pitch
= handle
->stride
/ rsc
->cpp
;
959 slice
->offset
= handle
->offset
;
960 slice
->size0
= handle
->stride
* prsc
->height0
;
962 if ((slice
->pitch
< align(prsc
->width0
, pitchalign
)) ||
963 (slice
->pitch
& (pitchalign
- 1)))
971 fd_resource_destroy(pscreen
, prsc
);
976 * _copy_region using pipe (3d engine)
979 fd_blitter_pipe_copy_region(struct fd_context
*ctx
,
980 struct pipe_resource
*dst
,
982 unsigned dstx
, unsigned dsty
, unsigned dstz
,
983 struct pipe_resource
*src
,
985 const struct pipe_box
*src_box
)
987 /* not until we allow rendertargets to be buffers */
988 if (dst
->target
== PIPE_BUFFER
|| src
->target
== PIPE_BUFFER
)
991 if (!util_blitter_is_copy_supported(ctx
->blitter
, dst
, src
))
994 /* TODO we could discard if dst box covers dst level fully.. */
995 fd_blitter_pipe_begin(ctx
, false, false, FD_STAGE_BLIT
);
996 util_blitter_copy_texture(ctx
->blitter
,
997 dst
, dst_level
, dstx
, dsty
, dstz
,
998 src
, src_level
, src_box
);
999 fd_blitter_pipe_end(ctx
);
1005 * Copy a block of pixels from one resource to another.
1006 * The resource must be of the same format.
1007 * Resources with nr_samples > 1 are not allowed.
1010 fd_resource_copy_region(struct pipe_context
*pctx
,
1011 struct pipe_resource
*dst
,
1013 unsigned dstx
, unsigned dsty
, unsigned dstz
,
1014 struct pipe_resource
*src
,
1016 const struct pipe_box
*src_box
)
1018 struct fd_context
*ctx
= fd_context(pctx
);
1020 /* TODO if we have 2d core, or other DMA engine that could be used
1021 * for simple copies and reasonably easily synchronized with the 3d
1022 * core, this is where we'd plug it in..
1025 /* try blit on 3d pipe: */
1026 if (fd_blitter_pipe_copy_region(ctx
,
1027 dst
, dst_level
, dstx
, dsty
, dstz
,
1028 src
, src_level
, src_box
))
1031 /* else fallback to pure sw: */
1032 util_resource_copy_region(pctx
,
1033 dst
, dst_level
, dstx
, dsty
, dstz
,
1034 src
, src_level
, src_box
);
1038 fd_render_condition_check(struct pipe_context
*pctx
)
1040 struct fd_context
*ctx
= fd_context(pctx
);
1042 if (!ctx
->cond_query
)
1045 union pipe_query_result res
= { 0 };
1047 ctx
->cond_mode
!= PIPE_RENDER_COND_NO_WAIT
&&
1048 ctx
->cond_mode
!= PIPE_RENDER_COND_BY_REGION_NO_WAIT
;
1050 if (pctx
->get_query_result(pctx
, ctx
->cond_query
, wait
, &res
))
1051 return (bool)res
.u64
!= ctx
->cond_cond
;
1057 * Optimal hardware path for blitting pixels.
1058 * Scaling, format conversion, up- and downsampling (resolve) are allowed.
1061 fd_blit(struct pipe_context
*pctx
, const struct pipe_blit_info
*blit_info
)
1063 struct fd_context
*ctx
= fd_context(pctx
);
1064 struct pipe_blit_info info
= *blit_info
;
1065 bool discard
= false;
1067 if (info
.src
.resource
->nr_samples
> 1 &&
1068 info
.dst
.resource
->nr_samples
<= 1 &&
1069 !util_format_is_depth_or_stencil(info
.src
.resource
->format
) &&
1070 !util_format_is_pure_integer(info
.src
.resource
->format
)) {
1071 DBG("color resolve unimplemented");
1075 if (info
.render_condition_enable
&& !fd_render_condition_check(pctx
))
1078 if (!info
.scissor_enable
&& !info
.alpha_blend
) {
1079 discard
= util_texrange_covers_whole_level(info
.dst
.resource
,
1080 info
.dst
.level
, info
.dst
.box
.x
, info
.dst
.box
.y
,
1081 info
.dst
.box
.z
, info
.dst
.box
.width
,
1082 info
.dst
.box
.height
, info
.dst
.box
.depth
);
1085 if (util_try_blit_via_copy_region(pctx
, &info
)) {
1089 if (info
.mask
& PIPE_MASK_S
) {
1090 DBG("cannot blit stencil, skipping");
1091 info
.mask
&= ~PIPE_MASK_S
;
1094 if (!util_blitter_is_blit_supported(ctx
->blitter
, &info
)) {
1095 DBG("blit unsupported %s -> %s",
1096 util_format_short_name(info
.src
.resource
->format
),
1097 util_format_short_name(info
.dst
.resource
->format
));
1101 fd_blitter_pipe_begin(ctx
, info
.render_condition_enable
, discard
, FD_STAGE_BLIT
);
1102 util_blitter_blit(ctx
->blitter
, &info
);
1103 fd_blitter_pipe_end(ctx
);
1107 fd_blitter_pipe_begin(struct fd_context
*ctx
, bool render_cond
, bool discard
,
1108 enum fd_render_stage stage
)
1110 util_blitter_save_fragment_constant_buffer_slot(ctx
->blitter
,
1111 ctx
->constbuf
[PIPE_SHADER_FRAGMENT
].cb
);
1112 util_blitter_save_vertex_buffer_slot(ctx
->blitter
, ctx
->vtx
.vertexbuf
.vb
);
1113 util_blitter_save_vertex_elements(ctx
->blitter
, ctx
->vtx
.vtx
);
1114 util_blitter_save_vertex_shader(ctx
->blitter
, ctx
->prog
.vp
);
1115 util_blitter_save_so_targets(ctx
->blitter
, ctx
->streamout
.num_targets
,
1116 ctx
->streamout
.targets
);
1117 util_blitter_save_rasterizer(ctx
->blitter
, ctx
->rasterizer
);
1118 util_blitter_save_viewport(ctx
->blitter
, &ctx
->viewport
);
1119 util_blitter_save_scissor(ctx
->blitter
, &ctx
->scissor
);
1120 util_blitter_save_fragment_shader(ctx
->blitter
, ctx
->prog
.fp
);
1121 util_blitter_save_blend(ctx
->blitter
, ctx
->blend
);
1122 util_blitter_save_depth_stencil_alpha(ctx
->blitter
, ctx
->zsa
);
1123 util_blitter_save_stencil_ref(ctx
->blitter
, &ctx
->stencil_ref
);
1124 util_blitter_save_sample_mask(ctx
->blitter
, ctx
->sample_mask
);
1125 util_blitter_save_framebuffer(ctx
->blitter
,
1126 ctx
->batch
? &ctx
->batch
->framebuffer
: NULL
);
1127 util_blitter_save_fragment_sampler_states(ctx
->blitter
,
1128 ctx
->tex
[PIPE_SHADER_FRAGMENT
].num_samplers
,
1129 (void **)ctx
->tex
[PIPE_SHADER_FRAGMENT
].samplers
);
1130 util_blitter_save_fragment_sampler_views(ctx
->blitter
,
1131 ctx
->tex
[PIPE_SHADER_FRAGMENT
].num_textures
,
1132 ctx
->tex
[PIPE_SHADER_FRAGMENT
].textures
);
1134 util_blitter_save_render_condition(ctx
->blitter
,
1135 ctx
->cond_query
, ctx
->cond_cond
, ctx
->cond_mode
);
1138 fd_batch_set_stage(ctx
->batch
, stage
);
1140 ctx
->in_blit
= discard
;
1144 fd_blitter_pipe_end(struct fd_context
*ctx
)
1147 fd_batch_set_stage(ctx
->batch
, FD_STAGE_NULL
);
1148 ctx
->in_blit
= false;
1152 fd_flush_resource(struct pipe_context
*pctx
, struct pipe_resource
*prsc
)
1154 struct fd_resource
*rsc
= fd_resource(prsc
);
1156 if (rsc
->write_batch
)
1157 fd_batch_flush(rsc
->write_batch
, true);
1159 assert(!rsc
->write_batch
);
1163 fd_invalidate_resource(struct pipe_context
*pctx
, struct pipe_resource
*prsc
)
1165 struct fd_resource
*rsc
= fd_resource(prsc
);
1168 * TODO I guess we could track that the resource is invalidated and
1169 * use that as a hint to realloc rather than stall in _transfer_map(),
1170 * even in the non-DISCARD_WHOLE_RESOURCE case?
1173 if (rsc
->write_batch
) {
1174 struct fd_batch
*batch
= rsc
->write_batch
;
1175 struct pipe_framebuffer_state
*pfb
= &batch
->framebuffer
;
1177 if (pfb
->zsbuf
&& pfb
->zsbuf
->texture
== prsc
)
1178 batch
->resolve
&= ~(FD_BUFFER_DEPTH
| FD_BUFFER_STENCIL
);
1180 for (unsigned i
= 0; i
< pfb
->nr_cbufs
; i
++) {
1181 if (pfb
->cbufs
[i
] && pfb
->cbufs
[i
]->texture
== prsc
) {
1182 batch
->resolve
&= ~(PIPE_CLEAR_COLOR0
<< i
);
1189 fd_resource_screen_init(struct pipe_screen
*pscreen
)
1191 pscreen
->resource_create
= fd_resource_create
;
1192 pscreen
->resource_from_handle
= fd_resource_from_handle
;
1193 pscreen
->resource_get_handle
= u_resource_get_handle_vtbl
;
1194 pscreen
->resource_destroy
= u_resource_destroy_vtbl
;
1198 fd_resource_context_init(struct pipe_context
*pctx
)
1200 pctx
->transfer_map
= u_transfer_map_vtbl
;
1201 pctx
->transfer_flush_region
= u_transfer_flush_region_vtbl
;
1202 pctx
->transfer_unmap
= u_transfer_unmap_vtbl
;
1203 pctx
->buffer_subdata
= u_default_buffer_subdata
;
1204 pctx
->texture_subdata
= u_default_texture_subdata
;
1205 pctx
->create_surface
= fd_create_surface
;
1206 pctx
->surface_destroy
= fd_surface_destroy
;
1207 pctx
->resource_copy_region
= fd_resource_copy_region
;
1208 pctx
->blit
= fd_blit
;
1209 pctx
->flush_resource
= fd_flush_resource
;
1210 pctx
->invalidate_resource
= fd_invalidate_resource
;