2 * Copyright 2010 Jerome Glisse <glisse@freedesktop.org>
3 * Copyright 2018 Advanced Micro Devices, Inc.
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 * on the rights to use, copy, modify, merge, publish, distribute, sub
10 * license, and/or sell copies of the Software, and to permit persons to whom
11 * the 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 NON-INFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
21 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
22 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
23 * USE OR OTHER DEALINGS IN THE SOFTWARE.
26 #include "radeonsi/si_pipe.h"
27 #include "radeonsi/si_query.h"
28 #include "util/u_format.h"
29 #include "util/u_log.h"
30 #include "util/u_memory.h"
31 #include "util/u_pack_color.h"
32 #include "util/u_resource.h"
33 #include "util/u_surface.h"
34 #include "util/u_transfer.h"
35 #include "util/os_time.h"
38 #include "state_tracker/drm_driver.h"
39 #include "amd/common/sid.h"
41 static enum radeon_surf_mode
42 si_choose_tiling(struct si_screen
*sscreen
,
43 const struct pipe_resource
*templ
, bool tc_compatible_htile
);
46 bool si_prepare_for_dma_blit(struct si_context
*sctx
,
47 struct si_texture
*dst
,
48 unsigned dst_level
, unsigned dstx
,
49 unsigned dsty
, unsigned dstz
,
50 struct si_texture
*src
,
52 const struct pipe_box
*src_box
)
57 if (dst
->surface
.bpe
!= src
->surface
.bpe
)
60 /* MSAA: Blits don't exist in the real world. */
61 if (src
->buffer
.b
.b
.nr_samples
> 1 ||
62 dst
->buffer
.b
.b
.nr_samples
> 1)
65 /* Depth-stencil surfaces:
66 * When dst is linear, the DB->CB copy preserves HTILE.
67 * When dst is tiled, the 3D path must be used to update HTILE.
69 if (src
->is_depth
|| dst
->is_depth
)
73 * src: Use the 3D path. DCC decompression is expensive.
74 * dst: Use the 3D path to compress the pixels with DCC.
76 if (vi_dcc_enabled(src
, src_level
) ||
77 vi_dcc_enabled(dst
, dst_level
))
81 * src: Both texture and SDMA paths need decompression. Use SDMA.
82 * dst: If overwriting the whole texture, discard CMASK and use
83 * SDMA. Otherwise, use the 3D path.
85 if (dst
->cmask
.size
&& dst
->dirty_level_mask
& (1 << dst_level
)) {
86 /* The CMASK clear is only enabled for the first level. */
87 assert(dst_level
== 0);
88 if (!util_texrange_covers_whole_level(&dst
->buffer
.b
.b
, dst_level
,
89 dstx
, dsty
, dstz
, src_box
->width
,
90 src_box
->height
, src_box
->depth
))
93 si_texture_discard_cmask(sctx
->screen
, dst
);
96 /* All requirements are met. Prepare textures for SDMA. */
97 if (src
->cmask
.size
&& src
->dirty_level_mask
& (1 << src_level
))
98 sctx
->b
.flush_resource(&sctx
->b
, &src
->buffer
.b
.b
);
100 assert(!(src
->dirty_level_mask
& (1 << src_level
)));
101 assert(!(dst
->dirty_level_mask
& (1 << dst_level
)));
106 /* Same as resource_copy_region, except that both upsampling and downsampling are allowed. */
107 static void si_copy_region_with_blit(struct pipe_context
*pipe
,
108 struct pipe_resource
*dst
,
110 unsigned dstx
, unsigned dsty
, unsigned dstz
,
111 struct pipe_resource
*src
,
113 const struct pipe_box
*src_box
)
115 struct pipe_blit_info blit
;
117 memset(&blit
, 0, sizeof(blit
));
118 blit
.src
.resource
= src
;
119 blit
.src
.format
= src
->format
;
120 blit
.src
.level
= src_level
;
121 blit
.src
.box
= *src_box
;
122 blit
.dst
.resource
= dst
;
123 blit
.dst
.format
= dst
->format
;
124 blit
.dst
.level
= dst_level
;
125 blit
.dst
.box
.x
= dstx
;
126 blit
.dst
.box
.y
= dsty
;
127 blit
.dst
.box
.z
= dstz
;
128 blit
.dst
.box
.width
= src_box
->width
;
129 blit
.dst
.box
.height
= src_box
->height
;
130 blit
.dst
.box
.depth
= src_box
->depth
;
131 blit
.mask
= util_format_get_mask(src
->format
) &
132 util_format_get_mask(dst
->format
);
133 blit
.filter
= PIPE_TEX_FILTER_NEAREST
;
136 pipe
->blit(pipe
, &blit
);
140 /* Copy from a full GPU texture to a transfer's staging one. */
141 static void si_copy_to_staging_texture(struct pipe_context
*ctx
, struct r600_transfer
*rtransfer
)
143 struct si_context
*sctx
= (struct si_context
*)ctx
;
144 struct pipe_transfer
*transfer
= (struct pipe_transfer
*)rtransfer
;
145 struct pipe_resource
*dst
= &rtransfer
->staging
->b
.b
;
146 struct pipe_resource
*src
= transfer
->resource
;
148 if (src
->nr_samples
> 1) {
149 si_copy_region_with_blit(ctx
, dst
, 0, 0, 0, 0,
150 src
, transfer
->level
, &transfer
->box
);
154 sctx
->dma_copy(ctx
, dst
, 0, 0, 0, 0, src
, transfer
->level
,
158 /* Copy from a transfer's staging texture to a full GPU one. */
159 static void si_copy_from_staging_texture(struct pipe_context
*ctx
, struct r600_transfer
*rtransfer
)
161 struct si_context
*sctx
= (struct si_context
*)ctx
;
162 struct pipe_transfer
*transfer
= (struct pipe_transfer
*)rtransfer
;
163 struct pipe_resource
*dst
= transfer
->resource
;
164 struct pipe_resource
*src
= &rtransfer
->staging
->b
.b
;
165 struct pipe_box sbox
;
167 u_box_3d(0, 0, 0, transfer
->box
.width
, transfer
->box
.height
, transfer
->box
.depth
, &sbox
);
169 if (dst
->nr_samples
> 1) {
170 si_copy_region_with_blit(ctx
, dst
, transfer
->level
,
171 transfer
->box
.x
, transfer
->box
.y
, transfer
->box
.z
,
176 sctx
->dma_copy(ctx
, dst
, transfer
->level
,
177 transfer
->box
.x
, transfer
->box
.y
, transfer
->box
.z
,
181 static unsigned si_texture_get_offset(struct si_screen
*sscreen
,
182 struct si_texture
*tex
, unsigned level
,
183 const struct pipe_box
*box
,
185 unsigned *layer_stride
)
187 if (sscreen
->info
.chip_class
>= GFX9
) {
188 *stride
= tex
->surface
.u
.gfx9
.surf_pitch
* tex
->surface
.bpe
;
189 *layer_stride
= tex
->surface
.u
.gfx9
.surf_slice_size
;
194 /* Each texture is an array of slices. Each slice is an array
195 * of mipmap levels. */
196 return box
->z
* tex
->surface
.u
.gfx9
.surf_slice_size
+
197 tex
->surface
.u
.gfx9
.offset
[level
] +
198 (box
->y
/ tex
->surface
.blk_h
*
199 tex
->surface
.u
.gfx9
.surf_pitch
+
200 box
->x
/ tex
->surface
.blk_w
) * tex
->surface
.bpe
;
202 *stride
= tex
->surface
.u
.legacy
.level
[level
].nblk_x
*
204 assert((uint64_t)tex
->surface
.u
.legacy
.level
[level
].slice_size_dw
* 4 <= UINT_MAX
);
205 *layer_stride
= (uint64_t)tex
->surface
.u
.legacy
.level
[level
].slice_size_dw
* 4;
208 return tex
->surface
.u
.legacy
.level
[level
].offset
;
210 /* Each texture is an array of mipmap levels. Each level is
211 * an array of slices. */
212 return tex
->surface
.u
.legacy
.level
[level
].offset
+
213 box
->z
* (uint64_t)tex
->surface
.u
.legacy
.level
[level
].slice_size_dw
* 4 +
214 (box
->y
/ tex
->surface
.blk_h
*
215 tex
->surface
.u
.legacy
.level
[level
].nblk_x
+
216 box
->x
/ tex
->surface
.blk_w
) * tex
->surface
.bpe
;
220 static int si_init_surface(struct si_screen
*sscreen
,
221 struct radeon_surf
*surface
,
222 const struct pipe_resource
*ptex
,
223 unsigned num_color_samples
,
224 enum radeon_surf_mode array_mode
,
225 unsigned pitch_in_bytes_override
,
229 bool is_flushed_depth
,
230 bool tc_compatible_htile
)
232 const struct util_format_description
*desc
=
233 util_format_description(ptex
->format
);
234 bool is_depth
, is_stencil
;
236 unsigned i
, bpe
, flags
= 0;
238 is_depth
= util_format_has_depth(desc
);
239 is_stencil
= util_format_has_stencil(desc
);
241 if (!is_flushed_depth
&&
242 ptex
->format
== PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
) {
243 bpe
= 4; /* stencil is allocated separately on evergreen */
245 bpe
= util_format_get_blocksize(ptex
->format
);
246 assert(util_is_power_of_two_or_zero(bpe
));
249 if (!is_flushed_depth
&& is_depth
) {
250 flags
|= RADEON_SURF_ZBUFFER
;
252 if (tc_compatible_htile
&&
253 (sscreen
->info
.chip_class
>= GFX9
||
254 array_mode
== RADEON_SURF_MODE_2D
)) {
255 /* TC-compatible HTILE only supports Z32_FLOAT.
256 * GFX9 also supports Z16_UNORM.
257 * On VI, promote Z16 to Z32. DB->CB copies will convert
258 * the format for transfers.
260 if (sscreen
->info
.chip_class
== VI
)
263 flags
|= RADEON_SURF_TC_COMPATIBLE_HTILE
;
267 flags
|= RADEON_SURF_SBUFFER
;
270 if (sscreen
->info
.chip_class
>= VI
&&
271 (ptex
->flags
& SI_RESOURCE_FLAG_DISABLE_DCC
||
272 ptex
->format
== PIPE_FORMAT_R9G9B9E5_FLOAT
||
273 (ptex
->nr_samples
>= 2 && !sscreen
->dcc_msaa_allowed
)))
274 flags
|= RADEON_SURF_DISABLE_DCC
;
276 /* Stoney: 128bpp MSAA textures randomly fail piglit tests with DCC. */
277 if (sscreen
->info
.family
== CHIP_STONEY
&&
278 bpe
== 16 && ptex
->nr_samples
>= 2)
279 flags
|= RADEON_SURF_DISABLE_DCC
;
281 /* VI: DCC clear for 4x and 8x MSAA array textures unimplemented. */
282 if (sscreen
->info
.chip_class
== VI
&&
283 num_color_samples
>= 4 &&
284 ptex
->array_size
> 1)
285 flags
|= RADEON_SURF_DISABLE_DCC
;
287 /* GFX9: DCC clear for 4x and 8x MSAA textures unimplemented. */
288 if (sscreen
->info
.chip_class
>= GFX9
&&
289 num_color_samples
>= 4)
290 flags
|= RADEON_SURF_DISABLE_DCC
;
292 if (ptex
->bind
& PIPE_BIND_SCANOUT
|| is_scanout
) {
293 /* This should catch bugs in gallium users setting incorrect flags. */
294 assert(ptex
->nr_samples
<= 1 &&
295 ptex
->array_size
== 1 &&
297 ptex
->last_level
== 0 &&
298 !(flags
& RADEON_SURF_Z_OR_SBUFFER
));
300 flags
|= RADEON_SURF_SCANOUT
;
303 if (ptex
->bind
& PIPE_BIND_SHARED
)
304 flags
|= RADEON_SURF_SHAREABLE
;
306 flags
|= RADEON_SURF_IMPORTED
| RADEON_SURF_SHAREABLE
;
307 if (!(ptex
->flags
& SI_RESOURCE_FLAG_FORCE_TILING
))
308 flags
|= RADEON_SURF_OPTIMIZE_FOR_SPACE
;
310 r
= sscreen
->ws
->surface_init(sscreen
->ws
, ptex
, num_color_samples
,
311 flags
, bpe
, array_mode
, surface
);
316 unsigned pitch
= pitch_in_bytes_override
/ bpe
;
318 if (sscreen
->info
.chip_class
>= GFX9
) {
320 surface
->u
.gfx9
.surf_pitch
= pitch
;
321 surface
->u
.gfx9
.surf_slice_size
=
322 (uint64_t)pitch
* surface
->u
.gfx9
.surf_height
* bpe
;
324 surface
->u
.gfx9
.surf_offset
= offset
;
327 surface
->u
.legacy
.level
[0].nblk_x
= pitch
;
328 surface
->u
.legacy
.level
[0].slice_size_dw
=
329 ((uint64_t)pitch
* surface
->u
.legacy
.level
[0].nblk_y
* bpe
) / 4;
332 for (i
= 0; i
< ARRAY_SIZE(surface
->u
.legacy
.level
); ++i
)
333 surface
->u
.legacy
.level
[i
].offset
+= offset
;
339 static void si_texture_init_metadata(struct si_screen
*sscreen
,
340 struct si_texture
*tex
,
341 struct radeon_bo_metadata
*metadata
)
343 struct radeon_surf
*surface
= &tex
->surface
;
345 memset(metadata
, 0, sizeof(*metadata
));
347 if (sscreen
->info
.chip_class
>= GFX9
) {
348 metadata
->u
.gfx9
.swizzle_mode
= surface
->u
.gfx9
.surf
.swizzle_mode
;
350 metadata
->u
.legacy
.microtile
= surface
->u
.legacy
.level
[0].mode
>= RADEON_SURF_MODE_1D
?
351 RADEON_LAYOUT_TILED
: RADEON_LAYOUT_LINEAR
;
352 metadata
->u
.legacy
.macrotile
= surface
->u
.legacy
.level
[0].mode
>= RADEON_SURF_MODE_2D
?
353 RADEON_LAYOUT_TILED
: RADEON_LAYOUT_LINEAR
;
354 metadata
->u
.legacy
.pipe_config
= surface
->u
.legacy
.pipe_config
;
355 metadata
->u
.legacy
.bankw
= surface
->u
.legacy
.bankw
;
356 metadata
->u
.legacy
.bankh
= surface
->u
.legacy
.bankh
;
357 metadata
->u
.legacy
.tile_split
= surface
->u
.legacy
.tile_split
;
358 metadata
->u
.legacy
.mtilea
= surface
->u
.legacy
.mtilea
;
359 metadata
->u
.legacy
.num_banks
= surface
->u
.legacy
.num_banks
;
360 metadata
->u
.legacy
.stride
= surface
->u
.legacy
.level
[0].nblk_x
* surface
->bpe
;
361 metadata
->u
.legacy
.scanout
= (surface
->flags
& RADEON_SURF_SCANOUT
) != 0;
365 static void si_surface_import_metadata(struct si_screen
*sscreen
,
366 struct radeon_surf
*surf
,
367 struct radeon_bo_metadata
*metadata
,
368 enum radeon_surf_mode
*array_mode
,
371 if (sscreen
->info
.chip_class
>= GFX9
) {
372 if (metadata
->u
.gfx9
.swizzle_mode
> 0)
373 *array_mode
= RADEON_SURF_MODE_2D
;
375 *array_mode
= RADEON_SURF_MODE_LINEAR_ALIGNED
;
377 *is_scanout
= metadata
->u
.gfx9
.swizzle_mode
== 0 ||
378 metadata
->u
.gfx9
.swizzle_mode
% 4 == 2;
380 surf
->u
.gfx9
.surf
.swizzle_mode
= metadata
->u
.gfx9
.swizzle_mode
;
382 surf
->u
.legacy
.pipe_config
= metadata
->u
.legacy
.pipe_config
;
383 surf
->u
.legacy
.bankw
= metadata
->u
.legacy
.bankw
;
384 surf
->u
.legacy
.bankh
= metadata
->u
.legacy
.bankh
;
385 surf
->u
.legacy
.tile_split
= metadata
->u
.legacy
.tile_split
;
386 surf
->u
.legacy
.mtilea
= metadata
->u
.legacy
.mtilea
;
387 surf
->u
.legacy
.num_banks
= metadata
->u
.legacy
.num_banks
;
389 if (metadata
->u
.legacy
.macrotile
== RADEON_LAYOUT_TILED
)
390 *array_mode
= RADEON_SURF_MODE_2D
;
391 else if (metadata
->u
.legacy
.microtile
== RADEON_LAYOUT_TILED
)
392 *array_mode
= RADEON_SURF_MODE_1D
;
394 *array_mode
= RADEON_SURF_MODE_LINEAR_ALIGNED
;
396 *is_scanout
= metadata
->u
.legacy
.scanout
;
400 void si_eliminate_fast_color_clear(struct si_context
*sctx
,
401 struct si_texture
*tex
)
403 struct si_screen
*sscreen
= sctx
->screen
;
404 struct pipe_context
*ctx
= &sctx
->b
;
406 if (ctx
== sscreen
->aux_context
)
407 mtx_lock(&sscreen
->aux_context_lock
);
409 unsigned n
= sctx
->num_decompress_calls
;
410 ctx
->flush_resource(ctx
, &tex
->buffer
.b
.b
);
412 /* Flush only if any fast clear elimination took place. */
413 if (n
!= sctx
->num_decompress_calls
)
414 ctx
->flush(ctx
, NULL
, 0);
416 if (ctx
== sscreen
->aux_context
)
417 mtx_unlock(&sscreen
->aux_context_lock
);
420 void si_texture_discard_cmask(struct si_screen
*sscreen
,
421 struct si_texture
*tex
)
423 if (!tex
->cmask
.size
)
426 assert(tex
->buffer
.b
.b
.nr_samples
<= 1);
429 memset(&tex
->cmask
, 0, sizeof(tex
->cmask
));
430 tex
->cmask
.base_address_reg
= tex
->buffer
.gpu_address
>> 8;
431 tex
->dirty_level_mask
= 0;
433 tex
->cb_color_info
&= ~S_028C70_FAST_CLEAR(1);
435 if (tex
->cmask_buffer
!= &tex
->buffer
)
436 r600_resource_reference(&tex
->cmask_buffer
, NULL
);
438 /* Notify all contexts about the change. */
439 p_atomic_inc(&sscreen
->dirty_tex_counter
);
440 p_atomic_inc(&sscreen
->compressed_colortex_counter
);
443 static bool si_can_disable_dcc(struct si_texture
*tex
)
445 /* We can't disable DCC if it can be written by another process. */
446 return tex
->dcc_offset
&&
447 (!tex
->buffer
.b
.is_shared
||
448 !(tex
->buffer
.external_usage
& PIPE_HANDLE_USAGE_WRITE
));
451 static bool si_texture_discard_dcc(struct si_screen
*sscreen
,
452 struct si_texture
*tex
)
454 if (!si_can_disable_dcc(tex
))
457 assert(tex
->dcc_separate_buffer
== NULL
);
462 /* Notify all contexts about the change. */
463 p_atomic_inc(&sscreen
->dirty_tex_counter
);
468 * Disable DCC for the texture. (first decompress, then discard metadata).
470 * There is unresolved multi-context synchronization issue between
471 * screen::aux_context and the current context. If applications do this with
472 * multiple contexts, it's already undefined behavior for them and we don't
473 * have to worry about that. The scenario is:
475 * If context 1 disables DCC and context 2 has queued commands that write
476 * to the texture via CB with DCC enabled, and the order of operations is
478 * context 2 queues draw calls rendering to the texture, but doesn't flush
479 * context 1 disables DCC and flushes
480 * context 1 & 2 reset descriptors and FB state
481 * context 2 flushes (new compressed tiles written by the draw calls)
482 * context 1 & 2 read garbage, because DCC is disabled, yet there are
485 * \param sctx the current context if you have one, or rscreen->aux_context
488 bool si_texture_disable_dcc(struct si_context
*sctx
,
489 struct si_texture
*tex
)
491 struct si_screen
*sscreen
= sctx
->screen
;
493 if (!si_can_disable_dcc(tex
))
496 if (&sctx
->b
== sscreen
->aux_context
)
497 mtx_lock(&sscreen
->aux_context_lock
);
499 /* Decompress DCC. */
500 si_decompress_dcc(sctx
, tex
);
501 sctx
->b
.flush(&sctx
->b
, NULL
, 0);
503 if (&sctx
->b
== sscreen
->aux_context
)
504 mtx_unlock(&sscreen
->aux_context_lock
);
506 return si_texture_discard_dcc(sscreen
, tex
);
509 static void si_reallocate_texture_inplace(struct si_context
*sctx
,
510 struct si_texture
*tex
,
511 unsigned new_bind_flag
,
512 bool invalidate_storage
)
514 struct pipe_screen
*screen
= sctx
->b
.screen
;
515 struct si_texture
*new_tex
;
516 struct pipe_resource templ
= tex
->buffer
.b
.b
;
519 templ
.bind
|= new_bind_flag
;
521 if (tex
->buffer
.b
.is_shared
)
524 if (new_bind_flag
== PIPE_BIND_LINEAR
) {
525 if (tex
->surface
.is_linear
)
528 /* This fails with MSAA, depth, and compressed textures. */
529 if (si_choose_tiling(sctx
->screen
, &templ
, false) !=
530 RADEON_SURF_MODE_LINEAR_ALIGNED
)
534 new_tex
= (struct si_texture
*)screen
->resource_create(screen
, &templ
);
538 /* Copy the pixels to the new texture. */
539 if (!invalidate_storage
) {
540 for (i
= 0; i
<= templ
.last_level
; i
++) {
544 u_minify(templ
.width0
, i
), u_minify(templ
.height0
, i
),
545 util_num_layers(&templ
, i
), &box
);
547 sctx
->dma_copy(&sctx
->b
, &new_tex
->buffer
.b
.b
, i
, 0, 0, 0,
548 &tex
->buffer
.b
.b
, i
, &box
);
552 if (new_bind_flag
== PIPE_BIND_LINEAR
) {
553 si_texture_discard_cmask(sctx
->screen
, tex
);
554 si_texture_discard_dcc(sctx
->screen
, tex
);
557 /* Replace the structure fields of tex. */
558 tex
->buffer
.b
.b
.bind
= templ
.bind
;
559 pb_reference(&tex
->buffer
.buf
, new_tex
->buffer
.buf
);
560 tex
->buffer
.gpu_address
= new_tex
->buffer
.gpu_address
;
561 tex
->buffer
.vram_usage
= new_tex
->buffer
.vram_usage
;
562 tex
->buffer
.gart_usage
= new_tex
->buffer
.gart_usage
;
563 tex
->buffer
.bo_size
= new_tex
->buffer
.bo_size
;
564 tex
->buffer
.bo_alignment
= new_tex
->buffer
.bo_alignment
;
565 tex
->buffer
.domains
= new_tex
->buffer
.domains
;
566 tex
->buffer
.flags
= new_tex
->buffer
.flags
;
567 tex
->size
= new_tex
->size
;
568 tex
->db_render_format
= new_tex
->db_render_format
;
569 tex
->db_compatible
= new_tex
->db_compatible
;
570 tex
->can_sample_z
= new_tex
->can_sample_z
;
571 tex
->can_sample_s
= new_tex
->can_sample_s
;
572 tex
->surface
= new_tex
->surface
;
573 tex
->fmask_offset
= new_tex
->fmask_offset
;
574 tex
->cmask
= new_tex
->cmask
;
575 tex
->cb_color_info
= new_tex
->cb_color_info
;
576 tex
->last_msaa_resolve_target_micro_mode
= new_tex
->last_msaa_resolve_target_micro_mode
;
577 tex
->htile_offset
= new_tex
->htile_offset
;
578 tex
->tc_compatible_htile
= new_tex
->tc_compatible_htile
;
579 tex
->depth_cleared
= new_tex
->depth_cleared
;
580 tex
->stencil_cleared
= new_tex
->stencil_cleared
;
581 tex
->dcc_gather_statistics
= new_tex
->dcc_gather_statistics
;
582 tex
->framebuffers_bound
= new_tex
->framebuffers_bound
;
584 if (new_bind_flag
== PIPE_BIND_LINEAR
) {
585 assert(!tex
->htile_offset
);
586 assert(!tex
->cmask
.size
);
587 assert(!tex
->surface
.fmask_size
);
588 assert(!tex
->dcc_offset
);
589 assert(!tex
->is_depth
);
592 si_texture_reference(&new_tex
, NULL
);
594 p_atomic_inc(&sctx
->screen
->dirty_tex_counter
);
597 static uint32_t si_get_bo_metadata_word1(struct si_screen
*sscreen
)
599 return (ATI_VENDOR_ID
<< 16) | sscreen
->info
.pci_id
;
602 static void si_query_opaque_metadata(struct si_screen
*sscreen
,
603 struct si_texture
*tex
,
604 struct radeon_bo_metadata
*md
)
606 struct pipe_resource
*res
= &tex
->buffer
.b
.b
;
607 static const unsigned char swizzle
[] = {
614 bool is_array
= util_texture_is_array(res
->target
);
616 if (!sscreen
->info
.has_bo_metadata
)
619 assert(tex
->dcc_separate_buffer
== NULL
);
620 assert(tex
->surface
.fmask_size
== 0);
622 /* Metadata image format format version 1:
623 * [0] = 1 (metadata format identifier)
624 * [1] = (VENDOR_ID << 16) | PCI_ID
625 * [2:9] = image descriptor for the whole resource
626 * [2] is always 0, because the base address is cleared
627 * [9] is the DCC offset bits [39:8] from the beginning of
629 * [10:10+LAST_LEVEL] = mipmap level offset bits [39:8] for each level
632 md
->metadata
[0] = 1; /* metadata image format version 1 */
634 /* TILE_MODE_INDEX is ambiguous without a PCI ID. */
635 md
->metadata
[1] = si_get_bo_metadata_word1(sscreen
);
637 si_make_texture_descriptor(sscreen
, tex
, true,
638 res
->target
, res
->format
,
639 swizzle
, 0, res
->last_level
, 0,
640 is_array
? res
->array_size
- 1 : 0,
641 res
->width0
, res
->height0
, res
->depth0
,
644 si_set_mutable_tex_desc_fields(sscreen
, tex
, &tex
->surface
.u
.legacy
.level
[0],
645 0, 0, tex
->surface
.blk_w
, false, desc
);
647 /* Clear the base address and set the relative DCC offset. */
649 desc
[1] &= C_008F14_BASE_ADDRESS_HI
;
650 desc
[7] = tex
->dcc_offset
>> 8;
652 /* Dwords [2:9] contain the image descriptor. */
653 memcpy(&md
->metadata
[2], desc
, sizeof(desc
));
654 md
->size_metadata
= 10 * 4;
656 /* Dwords [10:..] contain the mipmap level offsets. */
657 if (sscreen
->info
.chip_class
<= VI
) {
658 for (i
= 0; i
<= res
->last_level
; i
++)
659 md
->metadata
[10+i
] = tex
->surface
.u
.legacy
.level
[i
].offset
>> 8;
661 md
->size_metadata
+= (1 + res
->last_level
) * 4;
665 static void si_apply_opaque_metadata(struct si_screen
*sscreen
,
666 struct si_texture
*tex
,
667 struct radeon_bo_metadata
*md
)
669 uint32_t *desc
= &md
->metadata
[2];
671 if (sscreen
->info
.chip_class
< VI
)
674 /* Return if DCC is enabled. The texture should be set up with it
677 if (md
->size_metadata
>= 10 * 4 && /* at least 2(header) + 8(desc) dwords */
678 md
->metadata
[0] != 0 &&
679 md
->metadata
[1] == si_get_bo_metadata_word1(sscreen
) &&
680 G_008F28_COMPRESSION_EN(desc
[6])) {
681 tex
->dcc_offset
= (uint64_t)desc
[7] << 8;
685 /* Disable DCC. These are always set by texture_from_handle and must
691 static boolean
si_texture_get_handle(struct pipe_screen
* screen
,
692 struct pipe_context
*ctx
,
693 struct pipe_resource
*resource
,
694 struct winsys_handle
*whandle
,
697 struct si_screen
*sscreen
= (struct si_screen
*)screen
;
698 struct si_context
*sctx
;
699 struct r600_resource
*res
= r600_resource(resource
);
700 struct si_texture
*tex
= (struct si_texture
*)resource
;
701 struct radeon_bo_metadata metadata
;
702 bool update_metadata
= false;
703 unsigned stride
, offset
, slice_size
;
706 ctx
= threaded_context_unwrap_sync(ctx
);
707 sctx
= (struct si_context
*)(ctx
? ctx
: sscreen
->aux_context
);
709 if (resource
->target
!= PIPE_BUFFER
) {
710 /* This is not supported now, but it might be required for OpenCL
711 * interop in the future.
713 if (resource
->nr_samples
> 1 || tex
->is_depth
)
716 /* Move a suballocated texture into a non-suballocated allocation. */
717 if (sscreen
->ws
->buffer_is_suballocated(res
->buf
) ||
718 tex
->surface
.tile_swizzle
||
719 (tex
->buffer
.flags
& RADEON_FLAG_NO_INTERPROCESS_SHARING
&&
720 sscreen
->info
.has_local_buffers
&&
721 whandle
->type
!= WINSYS_HANDLE_TYPE_KMS
)) {
722 assert(!res
->b
.is_shared
);
723 si_reallocate_texture_inplace(sctx
, tex
,
724 PIPE_BIND_SHARED
, false);
726 assert(res
->b
.b
.bind
& PIPE_BIND_SHARED
);
727 assert(res
->flags
& RADEON_FLAG_NO_SUBALLOC
);
728 assert(!(res
->flags
& RADEON_FLAG_NO_INTERPROCESS_SHARING
));
729 assert(tex
->surface
.tile_swizzle
== 0);
732 /* Since shader image stores don't support DCC on VI,
733 * disable it for external clients that want write
736 if (usage
& PIPE_HANDLE_USAGE_WRITE
&& tex
->dcc_offset
) {
737 if (si_texture_disable_dcc(sctx
, tex
)) {
738 update_metadata
= true;
739 /* si_texture_disable_dcc flushes the context */
744 if (!(usage
& PIPE_HANDLE_USAGE_EXPLICIT_FLUSH
) &&
745 (tex
->cmask
.size
|| tex
->dcc_offset
)) {
746 /* Eliminate fast clear (both CMASK and DCC) */
747 si_eliminate_fast_color_clear(sctx
, tex
);
748 /* eliminate_fast_color_clear flushes the context */
751 /* Disable CMASK if flush_resource isn't going
755 si_texture_discard_cmask(sscreen
, tex
);
759 if (!res
->b
.is_shared
|| update_metadata
) {
760 si_texture_init_metadata(sscreen
, tex
, &metadata
);
761 si_query_opaque_metadata(sscreen
, tex
, &metadata
);
763 sscreen
->ws
->buffer_set_metadata(res
->buf
, &metadata
);
766 if (sscreen
->info
.chip_class
>= GFX9
) {
767 offset
= tex
->surface
.u
.gfx9
.surf_offset
;
768 stride
= tex
->surface
.u
.gfx9
.surf_pitch
*
770 slice_size
= tex
->surface
.u
.gfx9
.surf_slice_size
;
772 offset
= tex
->surface
.u
.legacy
.level
[0].offset
;
773 stride
= tex
->surface
.u
.legacy
.level
[0].nblk_x
*
775 slice_size
= (uint64_t)tex
->surface
.u
.legacy
.level
[0].slice_size_dw
* 4;
778 /* Buffer exports are for the OpenCL interop. */
779 /* Move a suballocated buffer into a non-suballocated allocation. */
780 if (sscreen
->ws
->buffer_is_suballocated(res
->buf
) ||
781 /* A DMABUF export always fails if the BO is local. */
782 (tex
->buffer
.flags
& RADEON_FLAG_NO_INTERPROCESS_SHARING
&&
783 sscreen
->info
.has_local_buffers
)) {
784 assert(!res
->b
.is_shared
);
786 /* Allocate a new buffer with PIPE_BIND_SHARED. */
787 struct pipe_resource templ
= res
->b
.b
;
788 templ
.bind
|= PIPE_BIND_SHARED
;
790 struct pipe_resource
*newb
=
791 screen
->resource_create(screen
, &templ
);
795 /* Copy the old buffer contents to the new one. */
797 u_box_1d(0, newb
->width0
, &box
);
798 sctx
->b
.resource_copy_region(&sctx
->b
, newb
, 0, 0, 0, 0,
801 /* Move the new buffer storage to the old pipe_resource. */
802 si_replace_buffer_storage(&sctx
->b
, &res
->b
.b
, newb
);
803 pipe_resource_reference(&newb
, NULL
);
805 assert(res
->b
.b
.bind
& PIPE_BIND_SHARED
);
806 assert(res
->flags
& RADEON_FLAG_NO_SUBALLOC
);
816 sctx
->b
.flush(&sctx
->b
, NULL
, 0);
818 if (res
->b
.is_shared
) {
819 /* USAGE_EXPLICIT_FLUSH must be cleared if at least one user
822 res
->external_usage
|= usage
& ~PIPE_HANDLE_USAGE_EXPLICIT_FLUSH
;
823 if (!(usage
& PIPE_HANDLE_USAGE_EXPLICIT_FLUSH
))
824 res
->external_usage
&= ~PIPE_HANDLE_USAGE_EXPLICIT_FLUSH
;
826 res
->b
.is_shared
= true;
827 res
->external_usage
= usage
;
830 return sscreen
->ws
->buffer_get_handle(res
->buf
, stride
, offset
,
831 slice_size
, whandle
);
834 static void si_texture_destroy(struct pipe_screen
*screen
,
835 struct pipe_resource
*ptex
)
837 struct si_texture
*tex
= (struct si_texture
*)ptex
;
838 struct r600_resource
*resource
= &tex
->buffer
;
840 si_texture_reference(&tex
->flushed_depth_texture
, NULL
);
842 if (tex
->cmask_buffer
!= &tex
->buffer
) {
843 r600_resource_reference(&tex
->cmask_buffer
, NULL
);
845 pb_reference(&resource
->buf
, NULL
);
846 r600_resource_reference(&tex
->dcc_separate_buffer
, NULL
);
847 r600_resource_reference(&tex
->last_dcc_separate_buffer
, NULL
);
851 static const struct u_resource_vtbl si_texture_vtbl
;
853 void si_texture_get_cmask_info(struct si_screen
*sscreen
,
854 struct si_texture
*tex
,
855 struct r600_cmask_info
*out
)
857 unsigned pipe_interleave_bytes
= sscreen
->info
.pipe_interleave_bytes
;
858 unsigned num_pipes
= sscreen
->info
.num_tile_pipes
;
859 unsigned cl_width
, cl_height
;
861 if (sscreen
->info
.chip_class
>= GFX9
) {
862 out
->alignment
= tex
->surface
.u
.gfx9
.cmask_alignment
;
863 out
->size
= tex
->surface
.u
.gfx9
.cmask_size
;
880 case 16: /* Hawaii */
889 unsigned base_align
= num_pipes
* pipe_interleave_bytes
;
891 unsigned width
= align(tex
->buffer
.b
.b
.width0
, cl_width
*8);
892 unsigned height
= align(tex
->buffer
.b
.b
.height0
, cl_height
*8);
893 unsigned slice_elements
= (width
* height
) / (8*8);
895 /* Each element of CMASK is a nibble. */
896 unsigned slice_bytes
= slice_elements
/ 2;
898 out
->slice_tile_max
= (width
* height
) / (128*128);
899 if (out
->slice_tile_max
)
900 out
->slice_tile_max
-= 1;
902 out
->alignment
= MAX2(256, base_align
);
903 out
->size
= util_num_layers(&tex
->buffer
.b
.b
, 0) *
904 align(slice_bytes
, base_align
);
907 static void si_texture_allocate_cmask(struct si_screen
*sscreen
,
908 struct si_texture
*tex
)
910 si_texture_get_cmask_info(sscreen
, tex
, &tex
->cmask
);
912 tex
->cmask
.offset
= align64(tex
->size
, tex
->cmask
.alignment
);
913 tex
->size
= tex
->cmask
.offset
+ tex
->cmask
.size
;
915 tex
->cb_color_info
|= S_028C70_FAST_CLEAR(1);
918 static void si_texture_get_htile_size(struct si_screen
*sscreen
,
919 struct si_texture
*tex
)
921 unsigned cl_width
, cl_height
, width
, height
;
922 unsigned slice_elements
, slice_bytes
, pipe_interleave_bytes
, base_align
;
923 unsigned num_pipes
= sscreen
->info
.num_tile_pipes
;
925 assert(sscreen
->info
.chip_class
<= VI
);
927 tex
->surface
.htile_size
= 0;
929 if (tex
->surface
.u
.legacy
.level
[0].mode
== RADEON_SURF_MODE_1D
&&
930 !sscreen
->info
.htile_cmask_support_1d_tiling
)
933 /* Overalign HTILE on P2 configs to work around GPU hangs in
934 * piglit/depthstencil-render-miplevels 585.
936 * This has been confirmed to help Kabini & Stoney, where the hangs
937 * are always reproducible. I think I have seen the test hang
938 * on Carrizo too, though it was very rare there.
940 if (sscreen
->info
.chip_class
>= CIK
&& num_pipes
< 4)
969 width
= align(tex
->buffer
.b
.b
.width0
, cl_width
* 8);
970 height
= align(tex
->buffer
.b
.b
.height0
, cl_height
* 8);
972 slice_elements
= (width
* height
) / (8 * 8);
973 slice_bytes
= slice_elements
* 4;
975 pipe_interleave_bytes
= sscreen
->info
.pipe_interleave_bytes
;
976 base_align
= num_pipes
* pipe_interleave_bytes
;
978 tex
->surface
.htile_alignment
= base_align
;
979 tex
->surface
.htile_size
=
980 util_num_layers(&tex
->buffer
.b
.b
, 0) *
981 align(slice_bytes
, base_align
);
984 static void si_texture_allocate_htile(struct si_screen
*sscreen
,
985 struct si_texture
*tex
)
987 if (sscreen
->info
.chip_class
<= VI
&& !tex
->tc_compatible_htile
)
988 si_texture_get_htile_size(sscreen
, tex
);
990 if (!tex
->surface
.htile_size
)
993 tex
->htile_offset
= align(tex
->size
, tex
->surface
.htile_alignment
);
994 tex
->size
= tex
->htile_offset
+ tex
->surface
.htile_size
;
997 void si_print_texture_info(struct si_screen
*sscreen
,
998 struct si_texture
*tex
, struct u_log_context
*log
)
1002 /* Common parameters. */
1003 u_log_printf(log
, " Info: npix_x=%u, npix_y=%u, npix_z=%u, blk_w=%u, "
1004 "blk_h=%u, array_size=%u, last_level=%u, "
1005 "bpe=%u, nsamples=%u, flags=0x%x, %s\n",
1006 tex
->buffer
.b
.b
.width0
, tex
->buffer
.b
.b
.height0
,
1007 tex
->buffer
.b
.b
.depth0
, tex
->surface
.blk_w
,
1009 tex
->buffer
.b
.b
.array_size
, tex
->buffer
.b
.b
.last_level
,
1010 tex
->surface
.bpe
, tex
->buffer
.b
.b
.nr_samples
,
1011 tex
->surface
.flags
, util_format_short_name(tex
->buffer
.b
.b
.format
));
1013 if (sscreen
->info
.chip_class
>= GFX9
) {
1014 u_log_printf(log
, " Surf: size=%"PRIu64
", slice_size=%"PRIu64
", "
1015 "alignment=%u, swmode=%u, epitch=%u, pitch=%u\n",
1016 tex
->surface
.surf_size
,
1017 tex
->surface
.u
.gfx9
.surf_slice_size
,
1018 tex
->surface
.surf_alignment
,
1019 tex
->surface
.u
.gfx9
.surf
.swizzle_mode
,
1020 tex
->surface
.u
.gfx9
.surf
.epitch
,
1021 tex
->surface
.u
.gfx9
.surf_pitch
);
1023 if (tex
->surface
.fmask_size
) {
1024 u_log_printf(log
, " FMASK: offset=%"PRIu64
", size=%"PRIu64
", "
1025 "alignment=%u, swmode=%u, epitch=%u\n",
1027 tex
->surface
.fmask_size
,
1028 tex
->surface
.fmask_alignment
,
1029 tex
->surface
.u
.gfx9
.fmask
.swizzle_mode
,
1030 tex
->surface
.u
.gfx9
.fmask
.epitch
);
1033 if (tex
->cmask
.size
) {
1034 u_log_printf(log
, " CMask: offset=%"PRIu64
", size=%"PRIu64
", "
1035 "alignment=%u, rb_aligned=%u, pipe_aligned=%u\n",
1037 tex
->surface
.u
.gfx9
.cmask_size
,
1038 tex
->surface
.u
.gfx9
.cmask_alignment
,
1039 tex
->surface
.u
.gfx9
.cmask
.rb_aligned
,
1040 tex
->surface
.u
.gfx9
.cmask
.pipe_aligned
);
1043 if (tex
->htile_offset
) {
1044 u_log_printf(log
, " HTile: offset=%"PRIu64
", size=%u, alignment=%u, "
1045 "rb_aligned=%u, pipe_aligned=%u\n",
1047 tex
->surface
.htile_size
,
1048 tex
->surface
.htile_alignment
,
1049 tex
->surface
.u
.gfx9
.htile
.rb_aligned
,
1050 tex
->surface
.u
.gfx9
.htile
.pipe_aligned
);
1053 if (tex
->dcc_offset
) {
1054 u_log_printf(log
, " DCC: offset=%"PRIu64
", size=%u, "
1055 "alignment=%u, pitch_max=%u, num_dcc_levels=%u\n",
1056 tex
->dcc_offset
, tex
->surface
.dcc_size
,
1057 tex
->surface
.dcc_alignment
,
1058 tex
->surface
.u
.gfx9
.dcc_pitch_max
,
1059 tex
->surface
.num_dcc_levels
);
1062 if (tex
->surface
.u
.gfx9
.stencil_offset
) {
1063 u_log_printf(log
, " Stencil: offset=%"PRIu64
", swmode=%u, epitch=%u\n",
1064 tex
->surface
.u
.gfx9
.stencil_offset
,
1065 tex
->surface
.u
.gfx9
.stencil
.swizzle_mode
,
1066 tex
->surface
.u
.gfx9
.stencil
.epitch
);
1071 u_log_printf(log
, " Layout: size=%"PRIu64
", alignment=%u, bankw=%u, "
1072 "bankh=%u, nbanks=%u, mtilea=%u, tilesplit=%u, pipeconfig=%u, scanout=%u\n",
1073 tex
->surface
.surf_size
, tex
->surface
.surf_alignment
, tex
->surface
.u
.legacy
.bankw
,
1074 tex
->surface
.u
.legacy
.bankh
, tex
->surface
.u
.legacy
.num_banks
, tex
->surface
.u
.legacy
.mtilea
,
1075 tex
->surface
.u
.legacy
.tile_split
, tex
->surface
.u
.legacy
.pipe_config
,
1076 (tex
->surface
.flags
& RADEON_SURF_SCANOUT
) != 0);
1078 if (tex
->surface
.fmask_size
)
1079 u_log_printf(log
, " FMask: offset=%"PRIu64
", size=%"PRIu64
", alignment=%u, pitch_in_pixels=%u, "
1080 "bankh=%u, slice_tile_max=%u, tile_mode_index=%u\n",
1081 tex
->fmask_offset
, tex
->surface
.fmask_size
, tex
->surface
.fmask_alignment
,
1082 tex
->surface
.u
.legacy
.fmask
.pitch_in_pixels
,
1083 tex
->surface
.u
.legacy
.fmask
.bankh
,
1084 tex
->surface
.u
.legacy
.fmask
.slice_tile_max
,
1085 tex
->surface
.u
.legacy
.fmask
.tiling_index
);
1087 if (tex
->cmask
.size
)
1088 u_log_printf(log
, " CMask: offset=%"PRIu64
", size=%"PRIu64
", alignment=%u, "
1089 "slice_tile_max=%u\n",
1090 tex
->cmask
.offset
, tex
->cmask
.size
, tex
->cmask
.alignment
,
1091 tex
->cmask
.slice_tile_max
);
1093 if (tex
->htile_offset
)
1094 u_log_printf(log
, " HTile: offset=%"PRIu64
", size=%u, "
1095 "alignment=%u, TC_compatible = %u\n",
1096 tex
->htile_offset
, tex
->surface
.htile_size
,
1097 tex
->surface
.htile_alignment
,
1098 tex
->tc_compatible_htile
);
1100 if (tex
->dcc_offset
) {
1101 u_log_printf(log
, " DCC: offset=%"PRIu64
", size=%u, alignment=%u\n",
1102 tex
->dcc_offset
, tex
->surface
.dcc_size
,
1103 tex
->surface
.dcc_alignment
);
1104 for (i
= 0; i
<= tex
->buffer
.b
.b
.last_level
; i
++)
1105 u_log_printf(log
, " DCCLevel[%i]: enabled=%u, offset=%u, "
1106 "fast_clear_size=%u\n",
1107 i
, i
< tex
->surface
.num_dcc_levels
,
1108 tex
->surface
.u
.legacy
.level
[i
].dcc_offset
,
1109 tex
->surface
.u
.legacy
.level
[i
].dcc_fast_clear_size
);
1112 for (i
= 0; i
<= tex
->buffer
.b
.b
.last_level
; i
++)
1113 u_log_printf(log
, " Level[%i]: offset=%"PRIu64
", slice_size=%"PRIu64
", "
1114 "npix_x=%u, npix_y=%u, npix_z=%u, nblk_x=%u, nblk_y=%u, "
1115 "mode=%u, tiling_index = %u\n",
1116 i
, tex
->surface
.u
.legacy
.level
[i
].offset
,
1117 (uint64_t)tex
->surface
.u
.legacy
.level
[i
].slice_size_dw
* 4,
1118 u_minify(tex
->buffer
.b
.b
.width0
, i
),
1119 u_minify(tex
->buffer
.b
.b
.height0
, i
),
1120 u_minify(tex
->buffer
.b
.b
.depth0
, i
),
1121 tex
->surface
.u
.legacy
.level
[i
].nblk_x
,
1122 tex
->surface
.u
.legacy
.level
[i
].nblk_y
,
1123 tex
->surface
.u
.legacy
.level
[i
].mode
,
1124 tex
->surface
.u
.legacy
.tiling_index
[i
]);
1126 if (tex
->surface
.has_stencil
) {
1127 u_log_printf(log
, " StencilLayout: tilesplit=%u\n",
1128 tex
->surface
.u
.legacy
.stencil_tile_split
);
1129 for (i
= 0; i
<= tex
->buffer
.b
.b
.last_level
; i
++) {
1130 u_log_printf(log
, " StencilLevel[%i]: offset=%"PRIu64
", "
1131 "slice_size=%"PRIu64
", npix_x=%u, "
1132 "npix_y=%u, npix_z=%u, nblk_x=%u, nblk_y=%u, "
1133 "mode=%u, tiling_index = %u\n",
1134 i
, tex
->surface
.u
.legacy
.stencil_level
[i
].offset
,
1135 (uint64_t)tex
->surface
.u
.legacy
.stencil_level
[i
].slice_size_dw
* 4,
1136 u_minify(tex
->buffer
.b
.b
.width0
, i
),
1137 u_minify(tex
->buffer
.b
.b
.height0
, i
),
1138 u_minify(tex
->buffer
.b
.b
.depth0
, i
),
1139 tex
->surface
.u
.legacy
.stencil_level
[i
].nblk_x
,
1140 tex
->surface
.u
.legacy
.stencil_level
[i
].nblk_y
,
1141 tex
->surface
.u
.legacy
.stencil_level
[i
].mode
,
1142 tex
->surface
.u
.legacy
.stencil_tiling_index
[i
]);
1147 /* Common processing for si_texture_create and si_texture_from_handle */
1148 static struct si_texture
*
1149 si_texture_create_object(struct pipe_screen
*screen
,
1150 const struct pipe_resource
*base
,
1151 unsigned num_color_samples
,
1152 struct pb_buffer
*buf
,
1153 struct radeon_surf
*surface
)
1155 struct si_texture
*tex
;
1156 struct r600_resource
*resource
;
1157 struct si_screen
*sscreen
= (struct si_screen
*)screen
;
1159 tex
= CALLOC_STRUCT(si_texture
);
1163 resource
= &tex
->buffer
;
1164 resource
->b
.b
= *base
;
1165 resource
->b
.b
.next
= NULL
;
1166 resource
->b
.vtbl
= &si_texture_vtbl
;
1167 pipe_reference_init(&resource
->b
.b
.reference
, 1);
1168 resource
->b
.b
.screen
= screen
;
1170 /* don't include stencil-only formats which we don't support for rendering */
1171 tex
->is_depth
= util_format_has_depth(util_format_description(tex
->buffer
.b
.b
.format
));
1173 tex
->surface
= *surface
;
1174 tex
->size
= tex
->surface
.surf_size
;
1175 tex
->num_color_samples
= num_color_samples
;
1177 tex
->tc_compatible_htile
= tex
->surface
.htile_size
!= 0 &&
1178 (tex
->surface
.flags
&
1179 RADEON_SURF_TC_COMPATIBLE_HTILE
);
1181 /* TC-compatible HTILE:
1182 * - VI only supports Z32_FLOAT.
1183 * - GFX9 only supports Z32_FLOAT and Z16_UNORM. */
1184 if (tex
->tc_compatible_htile
) {
1185 if (sscreen
->info
.chip_class
>= GFX9
&&
1186 base
->format
== PIPE_FORMAT_Z16_UNORM
)
1187 tex
->db_render_format
= base
->format
;
1189 tex
->db_render_format
= PIPE_FORMAT_Z32_FLOAT
;
1190 tex
->upgraded_depth
= base
->format
!= PIPE_FORMAT_Z32_FLOAT
&&
1191 base
->format
!= PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
;
1194 tex
->db_render_format
= base
->format
;
1197 /* Applies to GCN. */
1198 tex
->last_msaa_resolve_target_micro_mode
= tex
->surface
.micro_tile_mode
;
1200 /* Disable separate DCC at the beginning. DRI2 doesn't reuse buffers
1201 * between frames, so the only thing that can enable separate DCC
1202 * with DRI2 is multiple slow clears within a frame.
1204 tex
->ps_draw_ratio
= 0;
1206 if (tex
->is_depth
) {
1207 if (sscreen
->info
.chip_class
>= GFX9
) {
1208 tex
->can_sample_z
= true;
1209 tex
->can_sample_s
= true;
1211 tex
->can_sample_z
= !tex
->surface
.u
.legacy
.depth_adjusted
;
1212 tex
->can_sample_s
= !tex
->surface
.u
.legacy
.stencil_adjusted
;
1215 if (!(base
->flags
& (SI_RESOURCE_FLAG_TRANSFER
|
1216 SI_RESOURCE_FLAG_FLUSHED_DEPTH
))) {
1217 tex
->db_compatible
= true;
1219 if (!(sscreen
->debug_flags
& DBG(NO_HYPERZ
)))
1220 si_texture_allocate_htile(sscreen
, tex
);
1223 if (base
->nr_samples
> 1 &&
1225 !(sscreen
->debug_flags
& DBG(NO_FMASK
))) {
1226 /* Allocate FMASK. */
1227 tex
->fmask_offset
= align64(tex
->size
,
1228 tex
->surface
.fmask_alignment
);
1229 tex
->size
= tex
->fmask_offset
+ tex
->surface
.fmask_size
;
1231 si_texture_allocate_cmask(sscreen
, tex
);
1232 tex
->cmask_buffer
= &tex
->buffer
;
1234 if (!tex
->surface
.fmask_size
|| !tex
->cmask
.size
) {
1240 /* Shared textures must always set up DCC here.
1241 * If it's not present, it will be disabled by
1242 * apply_opaque_metadata later.
1244 if (tex
->surface
.dcc_size
&&
1245 (buf
|| !(sscreen
->debug_flags
& DBG(NO_DCC
))) &&
1246 !(tex
->surface
.flags
& RADEON_SURF_SCANOUT
)) {
1247 /* Reserve space for the DCC buffer. */
1248 tex
->dcc_offset
= align64(tex
->size
, tex
->surface
.dcc_alignment
);
1249 tex
->size
= tex
->dcc_offset
+ tex
->surface
.dcc_size
;
1253 /* Now create the backing buffer. */
1255 si_init_resource_fields(sscreen
, resource
, tex
->size
,
1256 tex
->surface
.surf_alignment
);
1258 if (!si_alloc_resource(sscreen
, resource
)) {
1263 resource
->buf
= buf
;
1264 resource
->gpu_address
= sscreen
->ws
->buffer_get_virtual_address(resource
->buf
);
1265 resource
->bo_size
= buf
->size
;
1266 resource
->bo_alignment
= buf
->alignment
;
1267 resource
->domains
= sscreen
->ws
->buffer_get_initial_domain(resource
->buf
);
1268 if (resource
->domains
& RADEON_DOMAIN_VRAM
)
1269 resource
->vram_usage
= buf
->size
;
1270 else if (resource
->domains
& RADEON_DOMAIN_GTT
)
1271 resource
->gart_usage
= buf
->size
;
1274 if (tex
->cmask
.size
) {
1275 /* Initialize the cmask to 0xCC (= compressed state). */
1276 si_screen_clear_buffer(sscreen
, &tex
->cmask_buffer
->b
.b
,
1277 tex
->cmask
.offset
, tex
->cmask
.size
,
1280 if (tex
->htile_offset
) {
1281 uint32_t clear_value
= 0;
1283 if (sscreen
->info
.chip_class
>= GFX9
|| tex
->tc_compatible_htile
)
1284 clear_value
= 0x0000030F;
1286 si_screen_clear_buffer(sscreen
, &tex
->buffer
.b
.b
,
1288 tex
->surface
.htile_size
,
1292 /* Initialize DCC only if the texture is not being imported. */
1293 if (!buf
&& tex
->dcc_offset
) {
1294 si_screen_clear_buffer(sscreen
, &tex
->buffer
.b
.b
,
1296 tex
->surface
.dcc_size
,
1300 /* Initialize the CMASK base register value. */
1301 tex
->cmask
.base_address_reg
=
1302 (tex
->buffer
.gpu_address
+ tex
->cmask
.offset
) >> 8;
1304 if (sscreen
->debug_flags
& DBG(VM
)) {
1305 fprintf(stderr
, "VM start=0x%"PRIX64
" end=0x%"PRIX64
" | Texture %ix%ix%i, %i levels, %i samples, %s\n",
1306 tex
->buffer
.gpu_address
,
1307 tex
->buffer
.gpu_address
+ tex
->buffer
.buf
->size
,
1308 base
->width0
, base
->height0
, util_num_layers(base
, 0), base
->last_level
+1,
1309 base
->nr_samples
? base
->nr_samples
: 1, util_format_short_name(base
->format
));
1312 if (sscreen
->debug_flags
& DBG(TEX
)) {
1314 struct u_log_context log
;
1315 u_log_context_init(&log
);
1316 si_print_texture_info(sscreen
, tex
, &log
);
1317 u_log_new_page_print(&log
, stdout
);
1319 u_log_context_destroy(&log
);
1325 static enum radeon_surf_mode
1326 si_choose_tiling(struct si_screen
*sscreen
,
1327 const struct pipe_resource
*templ
, bool tc_compatible_htile
)
1329 const struct util_format_description
*desc
= util_format_description(templ
->format
);
1330 bool force_tiling
= templ
->flags
& SI_RESOURCE_FLAG_FORCE_TILING
;
1331 bool is_depth_stencil
= util_format_is_depth_or_stencil(templ
->format
) &&
1332 !(templ
->flags
& SI_RESOURCE_FLAG_FLUSHED_DEPTH
);
1334 /* MSAA resources must be 2D tiled. */
1335 if (templ
->nr_samples
> 1)
1336 return RADEON_SURF_MODE_2D
;
1338 /* Transfer resources should be linear. */
1339 if (templ
->flags
& SI_RESOURCE_FLAG_TRANSFER
)
1340 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1342 /* Avoid Z/S decompress blits by forcing TC-compatible HTILE on VI,
1343 * which requires 2D tiling.
1345 if (sscreen
->info
.chip_class
== VI
&& tc_compatible_htile
)
1346 return RADEON_SURF_MODE_2D
;
1348 /* Handle common candidates for the linear mode.
1349 * Compressed textures and DB surfaces must always be tiled.
1351 if (!force_tiling
&&
1352 !is_depth_stencil
&&
1353 !util_format_is_compressed(templ
->format
)) {
1354 if (sscreen
->debug_flags
& DBG(NO_TILING
))
1355 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1357 /* Tiling doesn't work with the 422 (SUBSAMPLED) formats. */
1358 if (desc
->layout
== UTIL_FORMAT_LAYOUT_SUBSAMPLED
)
1359 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1361 /* Cursors are linear on SI.
1362 * (XXX double-check, maybe also use RADEON_SURF_SCANOUT) */
1363 if (templ
->bind
& PIPE_BIND_CURSOR
)
1364 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1366 if (templ
->bind
& PIPE_BIND_LINEAR
)
1367 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1369 /* Textures with a very small height are recommended to be linear. */
1370 if (templ
->target
== PIPE_TEXTURE_1D
||
1371 templ
->target
== PIPE_TEXTURE_1D_ARRAY
||
1372 /* Only very thin and long 2D textures should benefit from
1373 * linear_aligned. */
1374 (templ
->width0
> 8 && templ
->height0
<= 2))
1375 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1377 /* Textures likely to be mapped often. */
1378 if (templ
->usage
== PIPE_USAGE_STAGING
||
1379 templ
->usage
== PIPE_USAGE_STREAM
)
1380 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1383 /* Make small textures 1D tiled. */
1384 if (templ
->width0
<= 16 || templ
->height0
<= 16 ||
1385 (sscreen
->debug_flags
& DBG(NO_2D_TILING
)))
1386 return RADEON_SURF_MODE_1D
;
1388 /* The allocator will switch to 1D if needed. */
1389 return RADEON_SURF_MODE_2D
;
1392 static unsigned si_get_num_color_samples(struct si_screen
*sscreen
,
1393 const struct pipe_resource
*templ
,
1396 if (!imported
&& templ
->nr_samples
>= 2 &&
1397 sscreen
->eqaa_force_color_samples
)
1398 return sscreen
->eqaa_force_color_samples
;
1400 return CLAMP(templ
->nr_samples
, 1, 8);
1403 struct pipe_resource
*si_texture_create(struct pipe_screen
*screen
,
1404 const struct pipe_resource
*templ
)
1406 struct si_screen
*sscreen
= (struct si_screen
*)screen
;
1407 bool is_zs
= util_format_is_depth_or_stencil(templ
->format
);
1409 if (templ
->nr_samples
>= 2) {
1410 /* This is hackish (overwriting the const pipe_resource template),
1411 * but should be harmless and state trackers can also see
1412 * the overriden number of samples in the created pipe_resource.
1414 if (is_zs
&& sscreen
->eqaa_force_z_samples
) {
1415 ((struct pipe_resource
*)templ
)->nr_samples
=
1416 sscreen
->eqaa_force_z_samples
;
1417 } else if (!is_zs
&& sscreen
->eqaa_force_color_samples
) {
1418 ((struct pipe_resource
*)templ
)->nr_samples
=
1419 sscreen
->eqaa_force_coverage_samples
;
1423 struct radeon_surf surface
= {0};
1424 bool is_flushed_depth
= templ
->flags
& SI_RESOURCE_FLAG_FLUSHED_DEPTH
;
1425 bool tc_compatible_htile
=
1426 sscreen
->info
.chip_class
>= VI
&&
1427 /* There are issues with TC-compatible HTILE on Tonga (and
1428 * Iceland is the same design), and documented bug workarounds
1429 * don't help. For example, this fails:
1430 * piglit/bin/tex-miplevel-selection 'texture()' 2DShadow -auto
1432 sscreen
->info
.family
!= CHIP_TONGA
&&
1433 sscreen
->info
.family
!= CHIP_ICELAND
&&
1434 (templ
->flags
& PIPE_RESOURCE_FLAG_TEXTURING_MORE_LIKELY
) &&
1435 !(sscreen
->debug_flags
& DBG(NO_HYPERZ
)) &&
1436 !is_flushed_depth
&&
1437 templ
->nr_samples
<= 1 && /* TC-compat HTILE is less efficient with MSAA */
1439 unsigned num_color_samples
= si_get_num_color_samples(sscreen
, templ
, false);
1442 r
= si_init_surface(sscreen
, &surface
, templ
, num_color_samples
,
1443 si_choose_tiling(sscreen
, templ
, tc_compatible_htile
),
1444 0, 0, false, false, is_flushed_depth
,
1445 tc_compatible_htile
);
1450 return (struct pipe_resource
*)
1451 si_texture_create_object(screen
, templ
, num_color_samples
,
1455 static struct pipe_resource
*si_texture_from_handle(struct pipe_screen
*screen
,
1456 const struct pipe_resource
*templ
,
1457 struct winsys_handle
*whandle
,
1460 struct si_screen
*sscreen
= (struct si_screen
*)screen
;
1461 struct pb_buffer
*buf
= NULL
;
1462 unsigned stride
= 0, offset
= 0;
1463 enum radeon_surf_mode array_mode
;
1464 struct radeon_surf surface
= {};
1466 struct radeon_bo_metadata metadata
= {};
1467 struct si_texture
*tex
;
1470 /* Support only 2D textures without mipmaps */
1471 if ((templ
->target
!= PIPE_TEXTURE_2D
&& templ
->target
!= PIPE_TEXTURE_RECT
) ||
1472 templ
->depth0
!= 1 || templ
->last_level
!= 0)
1475 buf
= sscreen
->ws
->buffer_from_handle(sscreen
->ws
, whandle
, &stride
, &offset
);
1479 sscreen
->ws
->buffer_get_metadata(buf
, &metadata
);
1480 si_surface_import_metadata(sscreen
, &surface
, &metadata
,
1481 &array_mode
, &is_scanout
);
1483 unsigned num_color_samples
= si_get_num_color_samples(sscreen
, templ
, true);
1485 r
= si_init_surface(sscreen
, &surface
, templ
, num_color_samples
,
1486 array_mode
, stride
, offset
, true, is_scanout
,
1492 tex
= si_texture_create_object(screen
, templ
, num_color_samples
,
1497 tex
->buffer
.b
.is_shared
= true;
1498 tex
->buffer
.external_usage
= usage
;
1500 si_apply_opaque_metadata(sscreen
, tex
, &metadata
);
1502 assert(tex
->surface
.tile_swizzle
== 0);
1503 return &tex
->buffer
.b
.b
;
1506 bool si_init_flushed_depth_texture(struct pipe_context
*ctx
,
1507 struct pipe_resource
*texture
,
1508 struct si_texture
**staging
)
1510 struct si_texture
*tex
= (struct si_texture
*)texture
;
1511 struct pipe_resource resource
;
1512 struct si_texture
**flushed_depth_texture
= staging
?
1513 staging
: &tex
->flushed_depth_texture
;
1514 enum pipe_format pipe_format
= texture
->format
;
1517 if (tex
->flushed_depth_texture
)
1518 return true; /* it's ready */
1520 if (!tex
->can_sample_z
&& tex
->can_sample_s
) {
1521 switch (pipe_format
) {
1522 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
:
1523 /* Save memory by not allocating the S plane. */
1524 pipe_format
= PIPE_FORMAT_Z32_FLOAT
;
1526 case PIPE_FORMAT_Z24_UNORM_S8_UINT
:
1527 case PIPE_FORMAT_S8_UINT_Z24_UNORM
:
1528 /* Save memory bandwidth by not copying the
1529 * stencil part during flush.
1531 * This potentially increases memory bandwidth
1532 * if an application uses both Z and S texturing
1533 * simultaneously (a flushed Z24S8 texture
1534 * would be stored compactly), but how often
1535 * does that really happen?
1537 pipe_format
= PIPE_FORMAT_Z24X8_UNORM
;
1541 } else if (!tex
->can_sample_s
&& tex
->can_sample_z
) {
1542 assert(util_format_has_stencil(util_format_description(pipe_format
)));
1544 /* DB->CB copies to an 8bpp surface don't work. */
1545 pipe_format
= PIPE_FORMAT_X24S8_UINT
;
1549 memset(&resource
, 0, sizeof(resource
));
1550 resource
.target
= texture
->target
;
1551 resource
.format
= pipe_format
;
1552 resource
.width0
= texture
->width0
;
1553 resource
.height0
= texture
->height0
;
1554 resource
.depth0
= texture
->depth0
;
1555 resource
.array_size
= texture
->array_size
;
1556 resource
.last_level
= texture
->last_level
;
1557 resource
.nr_samples
= texture
->nr_samples
;
1558 resource
.usage
= staging
? PIPE_USAGE_STAGING
: PIPE_USAGE_DEFAULT
;
1559 resource
.bind
= texture
->bind
& ~PIPE_BIND_DEPTH_STENCIL
;
1560 resource
.flags
= texture
->flags
| SI_RESOURCE_FLAG_FLUSHED_DEPTH
;
1563 resource
.flags
|= SI_RESOURCE_FLAG_TRANSFER
;
1565 *flushed_depth_texture
= (struct si_texture
*)ctx
->screen
->resource_create(ctx
->screen
, &resource
);
1566 if (*flushed_depth_texture
== NULL
) {
1567 PRINT_ERR("failed to create temporary texture to hold flushed depth\n");
1574 * Initialize the pipe_resource descriptor to be of the same size as the box,
1575 * which is supposed to hold a subregion of the texture "orig" at the given
1578 static void si_init_temp_resource_from_box(struct pipe_resource
*res
,
1579 struct pipe_resource
*orig
,
1580 const struct pipe_box
*box
,
1581 unsigned level
, unsigned flags
)
1583 memset(res
, 0, sizeof(*res
));
1584 res
->format
= orig
->format
;
1585 res
->width0
= box
->width
;
1586 res
->height0
= box
->height
;
1588 res
->array_size
= 1;
1589 res
->usage
= flags
& SI_RESOURCE_FLAG_TRANSFER
? PIPE_USAGE_STAGING
: PIPE_USAGE_DEFAULT
;
1592 /* We must set the correct texture target and dimensions for a 3D box. */
1593 if (box
->depth
> 1 && util_max_layer(orig
, level
) > 0) {
1594 res
->target
= PIPE_TEXTURE_2D_ARRAY
;
1595 res
->array_size
= box
->depth
;
1597 res
->target
= PIPE_TEXTURE_2D
;
1601 static bool si_can_invalidate_texture(struct si_screen
*sscreen
,
1602 struct si_texture
*tex
,
1603 unsigned transfer_usage
,
1604 const struct pipe_box
*box
)
1606 return !tex
->buffer
.b
.is_shared
&&
1607 !(transfer_usage
& PIPE_TRANSFER_READ
) &&
1608 tex
->buffer
.b
.b
.last_level
== 0 &&
1609 util_texrange_covers_whole_level(&tex
->buffer
.b
.b
, 0,
1610 box
->x
, box
->y
, box
->z
,
1611 box
->width
, box
->height
,
1615 static void si_texture_invalidate_storage(struct si_context
*sctx
,
1616 struct si_texture
*tex
)
1618 struct si_screen
*sscreen
= sctx
->screen
;
1620 /* There is no point in discarding depth and tiled buffers. */
1621 assert(!tex
->is_depth
);
1622 assert(tex
->surface
.is_linear
);
1624 /* Reallocate the buffer in the same pipe_resource. */
1625 si_alloc_resource(sscreen
, &tex
->buffer
);
1627 /* Initialize the CMASK base address (needed even without CMASK). */
1628 tex
->cmask
.base_address_reg
=
1629 (tex
->buffer
.gpu_address
+ tex
->cmask
.offset
) >> 8;
1631 p_atomic_inc(&sscreen
->dirty_tex_counter
);
1633 sctx
->num_alloc_tex_transfer_bytes
+= tex
->size
;
1636 static void *si_texture_transfer_map(struct pipe_context
*ctx
,
1637 struct pipe_resource
*texture
,
1640 const struct pipe_box
*box
,
1641 struct pipe_transfer
**ptransfer
)
1643 struct si_context
*sctx
= (struct si_context
*)ctx
;
1644 struct si_texture
*tex
= (struct si_texture
*)texture
;
1645 struct r600_transfer
*trans
;
1646 struct r600_resource
*buf
;
1647 unsigned offset
= 0;
1649 bool use_staging_texture
= false;
1651 assert(!(texture
->flags
& SI_RESOURCE_FLAG_TRANSFER
));
1652 assert(box
->width
&& box
->height
&& box
->depth
);
1654 /* Depth textures use staging unconditionally. */
1655 if (!tex
->is_depth
) {
1656 /* Degrade the tile mode if we get too many transfers on APUs.
1657 * On dGPUs, the staging texture is always faster.
1658 * Only count uploads that are at least 4x4 pixels large.
1660 if (!sctx
->screen
->info
.has_dedicated_vram
&&
1662 box
->width
>= 4 && box
->height
>= 4 &&
1663 p_atomic_inc_return(&tex
->num_level0_transfers
) == 10) {
1664 bool can_invalidate
=
1665 si_can_invalidate_texture(sctx
->screen
, tex
,
1668 si_reallocate_texture_inplace(sctx
, tex
,
1673 /* Tiled textures need to be converted into a linear texture for CPU
1674 * access. The staging texture is always linear and is placed in GART.
1676 * Reading from VRAM or GTT WC is slow, always use the staging
1677 * texture in this case.
1679 * Use the staging texture for uploads if the underlying BO
1682 if (!tex
->surface
.is_linear
)
1683 use_staging_texture
= true;
1684 else if (usage
& PIPE_TRANSFER_READ
)
1685 use_staging_texture
=
1686 tex
->buffer
.domains
& RADEON_DOMAIN_VRAM
||
1687 tex
->buffer
.flags
& RADEON_FLAG_GTT_WC
;
1688 /* Write & linear only: */
1689 else if (si_rings_is_buffer_referenced(sctx
, tex
->buffer
.buf
,
1690 RADEON_USAGE_READWRITE
) ||
1691 !sctx
->ws
->buffer_wait(tex
->buffer
.buf
, 0,
1692 RADEON_USAGE_READWRITE
)) {
1694 if (si_can_invalidate_texture(sctx
->screen
, tex
,
1696 si_texture_invalidate_storage(sctx
, tex
);
1698 use_staging_texture
= true;
1702 trans
= CALLOC_STRUCT(r600_transfer
);
1705 pipe_resource_reference(&trans
->b
.b
.resource
, texture
);
1706 trans
->b
.b
.level
= level
;
1707 trans
->b
.b
.usage
= usage
;
1708 trans
->b
.b
.box
= *box
;
1710 if (tex
->is_depth
) {
1711 struct si_texture
*staging_depth
;
1713 if (tex
->buffer
.b
.b
.nr_samples
> 1) {
1714 /* MSAA depth buffers need to be converted to single sample buffers.
1716 * Mapping MSAA depth buffers can occur if ReadPixels is called
1717 * with a multisample GLX visual.
1719 * First downsample the depth buffer to a temporary texture,
1720 * then decompress the temporary one to staging.
1722 * Only the region being mapped is transfered.
1724 struct pipe_resource resource
;
1726 si_init_temp_resource_from_box(&resource
, texture
, box
, level
, 0);
1728 if (!si_init_flushed_depth_texture(ctx
, &resource
, &staging_depth
)) {
1729 PRINT_ERR("failed to create temporary texture to hold untiled copy\n");
1733 if (usage
& PIPE_TRANSFER_READ
) {
1734 struct pipe_resource
*temp
= ctx
->screen
->resource_create(ctx
->screen
, &resource
);
1736 PRINT_ERR("failed to create a temporary depth texture\n");
1740 si_copy_region_with_blit(ctx
, temp
, 0, 0, 0, 0, texture
, level
, box
);
1741 si_blit_decompress_depth(ctx
, (struct si_texture
*)temp
, staging_depth
,
1742 0, 0, 0, box
->depth
, 0, 0);
1743 pipe_resource_reference(&temp
, NULL
);
1746 /* Just get the strides. */
1747 si_texture_get_offset(sctx
->screen
, staging_depth
, level
, NULL
,
1749 &trans
->b
.b
.layer_stride
);
1751 /* XXX: only readback the rectangle which is being mapped? */
1752 /* XXX: when discard is true, no need to read back from depth texture */
1753 if (!si_init_flushed_depth_texture(ctx
, texture
, &staging_depth
)) {
1754 PRINT_ERR("failed to create temporary texture to hold untiled copy\n");
1758 si_blit_decompress_depth(ctx
, tex
, staging_depth
,
1760 box
->z
, box
->z
+ box
->depth
- 1,
1763 offset
= si_texture_get_offset(sctx
->screen
, staging_depth
,
1766 &trans
->b
.b
.layer_stride
);
1769 trans
->staging
= &staging_depth
->buffer
;
1770 buf
= trans
->staging
;
1771 } else if (use_staging_texture
) {
1772 struct pipe_resource resource
;
1773 struct si_texture
*staging
;
1775 si_init_temp_resource_from_box(&resource
, texture
, box
, level
,
1776 SI_RESOURCE_FLAG_TRANSFER
);
1777 resource
.usage
= (usage
& PIPE_TRANSFER_READ
) ?
1778 PIPE_USAGE_STAGING
: PIPE_USAGE_STREAM
;
1780 /* Create the temporary texture. */
1781 staging
= (struct si_texture
*)ctx
->screen
->resource_create(ctx
->screen
, &resource
);
1783 PRINT_ERR("failed to create temporary texture to hold untiled copy\n");
1786 trans
->staging
= &staging
->buffer
;
1788 /* Just get the strides. */
1789 si_texture_get_offset(sctx
->screen
, staging
, 0, NULL
,
1791 &trans
->b
.b
.layer_stride
);
1793 if (usage
& PIPE_TRANSFER_READ
)
1794 si_copy_to_staging_texture(ctx
, trans
);
1796 usage
|= PIPE_TRANSFER_UNSYNCHRONIZED
;
1798 buf
= trans
->staging
;
1800 /* the resource is mapped directly */
1801 offset
= si_texture_get_offset(sctx
->screen
, tex
, level
, box
,
1803 &trans
->b
.b
.layer_stride
);
1807 if (!(map
= si_buffer_map_sync_with_rings(sctx
, buf
, usage
)))
1810 *ptransfer
= &trans
->b
.b
;
1811 return map
+ offset
;
1814 r600_resource_reference(&trans
->staging
, NULL
);
1815 pipe_resource_reference(&trans
->b
.b
.resource
, NULL
);
1820 static void si_texture_transfer_unmap(struct pipe_context
*ctx
,
1821 struct pipe_transfer
* transfer
)
1823 struct si_context
*sctx
= (struct si_context
*)ctx
;
1824 struct r600_transfer
*rtransfer
= (struct r600_transfer
*)transfer
;
1825 struct pipe_resource
*texture
= transfer
->resource
;
1826 struct si_texture
*tex
= (struct si_texture
*)texture
;
1828 if ((transfer
->usage
& PIPE_TRANSFER_WRITE
) && rtransfer
->staging
) {
1829 if (tex
->is_depth
&& tex
->buffer
.b
.b
.nr_samples
<= 1) {
1830 ctx
->resource_copy_region(ctx
, texture
, transfer
->level
,
1831 transfer
->box
.x
, transfer
->box
.y
, transfer
->box
.z
,
1832 &rtransfer
->staging
->b
.b
, transfer
->level
,
1835 si_copy_from_staging_texture(ctx
, rtransfer
);
1839 if (rtransfer
->staging
) {
1840 sctx
->num_alloc_tex_transfer_bytes
+= rtransfer
->staging
->buf
->size
;
1841 r600_resource_reference(&rtransfer
->staging
, NULL
);
1844 /* Heuristic for {upload, draw, upload, draw, ..}:
1846 * Flush the gfx IB if we've allocated too much texture storage.
1848 * The idea is that we don't want to build IBs that use too much
1849 * memory and put pressure on the kernel memory manager and we also
1850 * want to make temporary and invalidated buffers go idle ASAP to
1851 * decrease the total memory usage or make them reusable. The memory
1852 * usage will be slightly higher than given here because of the buffer
1853 * cache in the winsys.
1855 * The result is that the kernel memory manager is never a bottleneck.
1857 if (sctx
->num_alloc_tex_transfer_bytes
> sctx
->screen
->info
.gart_size
/ 4) {
1858 si_flush_gfx_cs(sctx
, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW
, NULL
);
1859 sctx
->num_alloc_tex_transfer_bytes
= 0;
1862 pipe_resource_reference(&transfer
->resource
, NULL
);
1866 static const struct u_resource_vtbl si_texture_vtbl
=
1868 NULL
, /* get_handle */
1869 si_texture_destroy
, /* resource_destroy */
1870 si_texture_transfer_map
, /* transfer_map */
1871 u_default_transfer_flush_region
, /* transfer_flush_region */
1872 si_texture_transfer_unmap
, /* transfer_unmap */
1875 /* Return if it's allowed to reinterpret one format as another with DCC enabled.
1877 bool vi_dcc_formats_compatible(enum pipe_format format1
,
1878 enum pipe_format format2
)
1880 const struct util_format_description
*desc1
, *desc2
;
1882 /* No format change - exit early. */
1883 if (format1
== format2
)
1886 format1
= si_simplify_cb_format(format1
);
1887 format2
= si_simplify_cb_format(format2
);
1889 /* Check again after format adjustments. */
1890 if (format1
== format2
)
1893 desc1
= util_format_description(format1
);
1894 desc2
= util_format_description(format2
);
1896 if (desc1
->layout
!= UTIL_FORMAT_LAYOUT_PLAIN
||
1897 desc2
->layout
!= UTIL_FORMAT_LAYOUT_PLAIN
)
1900 /* Float and non-float are totally incompatible. */
1901 if ((desc1
->channel
[0].type
== UTIL_FORMAT_TYPE_FLOAT
) !=
1902 (desc2
->channel
[0].type
== UTIL_FORMAT_TYPE_FLOAT
))
1905 /* Channel sizes must match across DCC formats.
1906 * Comparing just the first 2 channels should be enough.
1908 if (desc1
->channel
[0].size
!= desc2
->channel
[0].size
||
1909 (desc1
->nr_channels
>= 2 &&
1910 desc1
->channel
[1].size
!= desc2
->channel
[1].size
))
1913 /* Everything below is not needed if the driver never uses the DCC
1914 * clear code with the value of 1.
1917 /* If the clear values are all 1 or all 0, this constraint can be
1919 if (vi_alpha_is_on_msb(format1
) != vi_alpha_is_on_msb(format2
))
1922 /* Channel types must match if the clear value of 1 is used.
1923 * The type categories are only float, signed, unsigned.
1924 * NORM and INT are always compatible.
1926 if (desc1
->channel
[0].type
!= desc2
->channel
[0].type
||
1927 (desc1
->nr_channels
>= 2 &&
1928 desc1
->channel
[1].type
!= desc2
->channel
[1].type
))
1934 bool vi_dcc_formats_are_incompatible(struct pipe_resource
*tex
,
1936 enum pipe_format view_format
)
1938 struct si_texture
*stex
= (struct si_texture
*)tex
;
1940 return vi_dcc_enabled(stex
, level
) &&
1941 !vi_dcc_formats_compatible(tex
->format
, view_format
);
1944 /* This can't be merged with the above function, because
1945 * vi_dcc_formats_compatible should be called only when DCC is enabled. */
1946 void vi_disable_dcc_if_incompatible_format(struct si_context
*sctx
,
1947 struct pipe_resource
*tex
,
1949 enum pipe_format view_format
)
1951 struct si_texture
*stex
= (struct si_texture
*)tex
;
1953 if (vi_dcc_formats_are_incompatible(tex
, level
, view_format
))
1954 if (!si_texture_disable_dcc(sctx
, stex
))
1955 si_decompress_dcc(sctx
, stex
);
1958 struct pipe_surface
*si_create_surface_custom(struct pipe_context
*pipe
,
1959 struct pipe_resource
*texture
,
1960 const struct pipe_surface
*templ
,
1961 unsigned width0
, unsigned height0
,
1962 unsigned width
, unsigned height
)
1964 struct r600_surface
*surface
= CALLOC_STRUCT(r600_surface
);
1969 assert(templ
->u
.tex
.first_layer
<= util_max_layer(texture
, templ
->u
.tex
.level
));
1970 assert(templ
->u
.tex
.last_layer
<= util_max_layer(texture
, templ
->u
.tex
.level
));
1972 pipe_reference_init(&surface
->base
.reference
, 1);
1973 pipe_resource_reference(&surface
->base
.texture
, texture
);
1974 surface
->base
.context
= pipe
;
1975 surface
->base
.format
= templ
->format
;
1976 surface
->base
.width
= width
;
1977 surface
->base
.height
= height
;
1978 surface
->base
.u
= templ
->u
;
1980 surface
->width0
= width0
;
1981 surface
->height0
= height0
;
1983 surface
->dcc_incompatible
=
1984 texture
->target
!= PIPE_BUFFER
&&
1985 vi_dcc_formats_are_incompatible(texture
, templ
->u
.tex
.level
,
1987 return &surface
->base
;
1990 static struct pipe_surface
*si_create_surface(struct pipe_context
*pipe
,
1991 struct pipe_resource
*tex
,
1992 const struct pipe_surface
*templ
)
1994 unsigned level
= templ
->u
.tex
.level
;
1995 unsigned width
= u_minify(tex
->width0
, level
);
1996 unsigned height
= u_minify(tex
->height0
, level
);
1997 unsigned width0
= tex
->width0
;
1998 unsigned height0
= tex
->height0
;
2000 if (tex
->target
!= PIPE_BUFFER
&& templ
->format
!= tex
->format
) {
2001 const struct util_format_description
*tex_desc
2002 = util_format_description(tex
->format
);
2003 const struct util_format_description
*templ_desc
2004 = util_format_description(templ
->format
);
2006 assert(tex_desc
->block
.bits
== templ_desc
->block
.bits
);
2008 /* Adjust size of surface if and only if the block width or
2009 * height is changed. */
2010 if (tex_desc
->block
.width
!= templ_desc
->block
.width
||
2011 tex_desc
->block
.height
!= templ_desc
->block
.height
) {
2012 unsigned nblks_x
= util_format_get_nblocksx(tex
->format
, width
);
2013 unsigned nblks_y
= util_format_get_nblocksy(tex
->format
, height
);
2015 width
= nblks_x
* templ_desc
->block
.width
;
2016 height
= nblks_y
* templ_desc
->block
.height
;
2018 width0
= util_format_get_nblocksx(tex
->format
, width0
);
2019 height0
= util_format_get_nblocksy(tex
->format
, height0
);
2023 return si_create_surface_custom(pipe
, tex
, templ
,
2028 static void si_surface_destroy(struct pipe_context
*pipe
,
2029 struct pipe_surface
*surface
)
2031 pipe_resource_reference(&surface
->texture
, NULL
);
2035 unsigned si_translate_colorswap(enum pipe_format format
, bool do_endian_swap
)
2037 const struct util_format_description
*desc
= util_format_description(format
);
2039 #define HAS_SWIZZLE(chan,swz) (desc->swizzle[chan] == PIPE_SWIZZLE_##swz)
2041 if (format
== PIPE_FORMAT_R11G11B10_FLOAT
) /* isn't plain */
2042 return V_028C70_SWAP_STD
;
2044 if (desc
->layout
!= UTIL_FORMAT_LAYOUT_PLAIN
)
2047 switch (desc
->nr_channels
) {
2049 if (HAS_SWIZZLE(0,X
))
2050 return V_028C70_SWAP_STD
; /* X___ */
2051 else if (HAS_SWIZZLE(3,X
))
2052 return V_028C70_SWAP_ALT_REV
; /* ___X */
2055 if ((HAS_SWIZZLE(0,X
) && HAS_SWIZZLE(1,Y
)) ||
2056 (HAS_SWIZZLE(0,X
) && HAS_SWIZZLE(1,NONE
)) ||
2057 (HAS_SWIZZLE(0,NONE
) && HAS_SWIZZLE(1,Y
)))
2058 return V_028C70_SWAP_STD
; /* XY__ */
2059 else if ((HAS_SWIZZLE(0,Y
) && HAS_SWIZZLE(1,X
)) ||
2060 (HAS_SWIZZLE(0,Y
) && HAS_SWIZZLE(1,NONE
)) ||
2061 (HAS_SWIZZLE(0,NONE
) && HAS_SWIZZLE(1,X
)))
2063 return (do_endian_swap
? V_028C70_SWAP_STD
: V_028C70_SWAP_STD_REV
);
2064 else if (HAS_SWIZZLE(0,X
) && HAS_SWIZZLE(3,Y
))
2065 return V_028C70_SWAP_ALT
; /* X__Y */
2066 else if (HAS_SWIZZLE(0,Y
) && HAS_SWIZZLE(3,X
))
2067 return V_028C70_SWAP_ALT_REV
; /* Y__X */
2070 if (HAS_SWIZZLE(0,X
))
2071 return (do_endian_swap
? V_028C70_SWAP_STD_REV
: V_028C70_SWAP_STD
);
2072 else if (HAS_SWIZZLE(0,Z
))
2073 return V_028C70_SWAP_STD_REV
; /* ZYX */
2076 /* check the middle channels, the 1st and 4th channel can be NONE */
2077 if (HAS_SWIZZLE(1,Y
) && HAS_SWIZZLE(2,Z
)) {
2078 return V_028C70_SWAP_STD
; /* XYZW */
2079 } else if (HAS_SWIZZLE(1,Z
) && HAS_SWIZZLE(2,Y
)) {
2080 return V_028C70_SWAP_STD_REV
; /* WZYX */
2081 } else if (HAS_SWIZZLE(1,Y
) && HAS_SWIZZLE(2,X
)) {
2082 return V_028C70_SWAP_ALT
; /* ZYXW */
2083 } else if (HAS_SWIZZLE(1,Z
) && HAS_SWIZZLE(2,W
)) {
2086 return V_028C70_SWAP_ALT_REV
;
2088 return (do_endian_swap
? V_028C70_SWAP_ALT
: V_028C70_SWAP_ALT_REV
);
2095 /* PIPELINE_STAT-BASED DCC ENABLEMENT FOR DISPLAYABLE SURFACES */
2097 static void vi_dcc_clean_up_context_slot(struct si_context
*sctx
,
2102 if (sctx
->dcc_stats
[slot
].query_active
)
2103 vi_separate_dcc_stop_query(sctx
,
2104 sctx
->dcc_stats
[slot
].tex
);
2106 for (i
= 0; i
< ARRAY_SIZE(sctx
->dcc_stats
[slot
].ps_stats
); i
++)
2107 if (sctx
->dcc_stats
[slot
].ps_stats
[i
]) {
2108 sctx
->b
.destroy_query(&sctx
->b
,
2109 sctx
->dcc_stats
[slot
].ps_stats
[i
]);
2110 sctx
->dcc_stats
[slot
].ps_stats
[i
] = NULL
;
2113 si_texture_reference(&sctx
->dcc_stats
[slot
].tex
, NULL
);
2117 * Return the per-context slot where DCC statistics queries for the texture live.
2119 static unsigned vi_get_context_dcc_stats_index(struct si_context
*sctx
,
2120 struct si_texture
*tex
)
2122 int i
, empty_slot
= -1;
2124 /* Remove zombie textures (textures kept alive by this array only). */
2125 for (i
= 0; i
< ARRAY_SIZE(sctx
->dcc_stats
); i
++)
2126 if (sctx
->dcc_stats
[i
].tex
&&
2127 sctx
->dcc_stats
[i
].tex
->buffer
.b
.b
.reference
.count
== 1)
2128 vi_dcc_clean_up_context_slot(sctx
, i
);
2130 /* Find the texture. */
2131 for (i
= 0; i
< ARRAY_SIZE(sctx
->dcc_stats
); i
++) {
2132 /* Return if found. */
2133 if (sctx
->dcc_stats
[i
].tex
== tex
) {
2134 sctx
->dcc_stats
[i
].last_use_timestamp
= os_time_get();
2138 /* Record the first seen empty slot. */
2139 if (empty_slot
== -1 && !sctx
->dcc_stats
[i
].tex
)
2143 /* Not found. Remove the oldest member to make space in the array. */
2144 if (empty_slot
== -1) {
2145 int oldest_slot
= 0;
2147 /* Find the oldest slot. */
2148 for (i
= 1; i
< ARRAY_SIZE(sctx
->dcc_stats
); i
++)
2149 if (sctx
->dcc_stats
[oldest_slot
].last_use_timestamp
>
2150 sctx
->dcc_stats
[i
].last_use_timestamp
)
2153 /* Clean up the oldest slot. */
2154 vi_dcc_clean_up_context_slot(sctx
, oldest_slot
);
2155 empty_slot
= oldest_slot
;
2158 /* Add the texture to the new slot. */
2159 si_texture_reference(&sctx
->dcc_stats
[empty_slot
].tex
, tex
);
2160 sctx
->dcc_stats
[empty_slot
].last_use_timestamp
= os_time_get();
2164 static struct pipe_query
*
2165 vi_create_resuming_pipestats_query(struct si_context
*sctx
)
2167 struct si_query_hw
*query
= (struct si_query_hw
*)
2168 sctx
->b
.create_query(&sctx
->b
, PIPE_QUERY_PIPELINE_STATISTICS
, 0);
2170 query
->flags
|= SI_QUERY_HW_FLAG_BEGIN_RESUMES
;
2171 return (struct pipe_query
*)query
;
2175 * Called when binding a color buffer.
2177 void vi_separate_dcc_start_query(struct si_context
*sctx
,
2178 struct si_texture
*tex
)
2180 unsigned i
= vi_get_context_dcc_stats_index(sctx
, tex
);
2182 assert(!sctx
->dcc_stats
[i
].query_active
);
2184 if (!sctx
->dcc_stats
[i
].ps_stats
[0])
2185 sctx
->dcc_stats
[i
].ps_stats
[0] = vi_create_resuming_pipestats_query(sctx
);
2187 /* begin or resume the query */
2188 sctx
->b
.begin_query(&sctx
->b
, sctx
->dcc_stats
[i
].ps_stats
[0]);
2189 sctx
->dcc_stats
[i
].query_active
= true;
2193 * Called when unbinding a color buffer.
2195 void vi_separate_dcc_stop_query(struct si_context
*sctx
,
2196 struct si_texture
*tex
)
2198 unsigned i
= vi_get_context_dcc_stats_index(sctx
, tex
);
2200 assert(sctx
->dcc_stats
[i
].query_active
);
2201 assert(sctx
->dcc_stats
[i
].ps_stats
[0]);
2203 /* pause or end the query */
2204 sctx
->b
.end_query(&sctx
->b
, sctx
->dcc_stats
[i
].ps_stats
[0]);
2205 sctx
->dcc_stats
[i
].query_active
= false;
2208 static bool vi_should_enable_separate_dcc(struct si_texture
*tex
)
2210 /* The minimum number of fullscreen draws per frame that is required
2212 return tex
->ps_draw_ratio
+ tex
->num_slow_clears
>= 5;
2215 /* Called by fast clear. */
2216 void vi_separate_dcc_try_enable(struct si_context
*sctx
,
2217 struct si_texture
*tex
)
2219 /* The intent is to use this with shared displayable back buffers,
2220 * but it's not strictly limited only to them.
2222 if (!tex
->buffer
.b
.is_shared
||
2223 !(tex
->buffer
.external_usage
& PIPE_HANDLE_USAGE_EXPLICIT_FLUSH
) ||
2224 tex
->buffer
.b
.b
.target
!= PIPE_TEXTURE_2D
||
2225 tex
->buffer
.b
.b
.last_level
> 0 ||
2226 !tex
->surface
.dcc_size
)
2229 if (tex
->dcc_offset
)
2230 return; /* already enabled */
2232 /* Enable the DCC stat gathering. */
2233 if (!tex
->dcc_gather_statistics
) {
2234 tex
->dcc_gather_statistics
= true;
2235 vi_separate_dcc_start_query(sctx
, tex
);
2238 if (!vi_should_enable_separate_dcc(tex
))
2239 return; /* stats show that DCC decompression is too expensive */
2241 assert(tex
->surface
.num_dcc_levels
);
2242 assert(!tex
->dcc_separate_buffer
);
2244 si_texture_discard_cmask(sctx
->screen
, tex
);
2246 /* Get a DCC buffer. */
2247 if (tex
->last_dcc_separate_buffer
) {
2248 assert(tex
->dcc_gather_statistics
);
2249 assert(!tex
->dcc_separate_buffer
);
2250 tex
->dcc_separate_buffer
= tex
->last_dcc_separate_buffer
;
2251 tex
->last_dcc_separate_buffer
= NULL
;
2253 tex
->dcc_separate_buffer
=
2254 si_aligned_buffer_create(sctx
->b
.screen
,
2255 SI_RESOURCE_FLAG_UNMAPPABLE
,
2257 tex
->surface
.dcc_size
,
2258 tex
->surface
.dcc_alignment
);
2259 if (!tex
->dcc_separate_buffer
)
2263 /* dcc_offset is the absolute GPUVM address. */
2264 tex
->dcc_offset
= tex
->dcc_separate_buffer
->gpu_address
;
2266 /* no need to flag anything since this is called by fast clear that
2267 * flags framebuffer state
2272 * Called by pipe_context::flush_resource, the place where DCC decompression
2275 void vi_separate_dcc_process_and_reset_stats(struct pipe_context
*ctx
,
2276 struct si_texture
*tex
)
2278 struct si_context
*sctx
= (struct si_context
*)ctx
;
2279 struct pipe_query
*tmp
;
2280 unsigned i
= vi_get_context_dcc_stats_index(sctx
, tex
);
2281 bool query_active
= sctx
->dcc_stats
[i
].query_active
;
2282 bool disable
= false;
2284 if (sctx
->dcc_stats
[i
].ps_stats
[2]) {
2285 union pipe_query_result result
;
2287 /* Read the results. */
2288 ctx
->get_query_result(ctx
, sctx
->dcc_stats
[i
].ps_stats
[2],
2290 si_query_hw_reset_buffers(sctx
,
2291 (struct si_query_hw
*)
2292 sctx
->dcc_stats
[i
].ps_stats
[2]);
2294 /* Compute the approximate number of fullscreen draws. */
2295 tex
->ps_draw_ratio
=
2296 result
.pipeline_statistics
.ps_invocations
/
2297 (tex
->buffer
.b
.b
.width0
* tex
->buffer
.b
.b
.height0
);
2298 sctx
->last_tex_ps_draw_ratio
= tex
->ps_draw_ratio
;
2300 disable
= tex
->dcc_separate_buffer
&&
2301 !vi_should_enable_separate_dcc(tex
);
2304 tex
->num_slow_clears
= 0;
2306 /* stop the statistics query for ps_stats[0] */
2308 vi_separate_dcc_stop_query(sctx
, tex
);
2310 /* Move the queries in the queue by one. */
2311 tmp
= sctx
->dcc_stats
[i
].ps_stats
[2];
2312 sctx
->dcc_stats
[i
].ps_stats
[2] = sctx
->dcc_stats
[i
].ps_stats
[1];
2313 sctx
->dcc_stats
[i
].ps_stats
[1] = sctx
->dcc_stats
[i
].ps_stats
[0];
2314 sctx
->dcc_stats
[i
].ps_stats
[0] = tmp
;
2316 /* create and start a new query as ps_stats[0] */
2318 vi_separate_dcc_start_query(sctx
, tex
);
2321 assert(!tex
->last_dcc_separate_buffer
);
2322 tex
->last_dcc_separate_buffer
= tex
->dcc_separate_buffer
;
2323 tex
->dcc_separate_buffer
= NULL
;
2324 tex
->dcc_offset
= 0;
2325 /* no need to flag anything since this is called after
2326 * decompression that re-sets framebuffer state
2331 static struct pipe_memory_object
*
2332 si_memobj_from_handle(struct pipe_screen
*screen
,
2333 struct winsys_handle
*whandle
,
2336 struct si_screen
*sscreen
= (struct si_screen
*)screen
;
2337 struct r600_memory_object
*memobj
= CALLOC_STRUCT(r600_memory_object
);
2338 struct pb_buffer
*buf
= NULL
;
2339 uint32_t stride
, offset
;
2344 buf
= sscreen
->ws
->buffer_from_handle(sscreen
->ws
, whandle
,
2351 memobj
->b
.dedicated
= dedicated
;
2353 memobj
->stride
= stride
;
2354 memobj
->offset
= offset
;
2356 return (struct pipe_memory_object
*)memobj
;
2361 si_memobj_destroy(struct pipe_screen
*screen
,
2362 struct pipe_memory_object
*_memobj
)
2364 struct r600_memory_object
*memobj
= (struct r600_memory_object
*)_memobj
;
2366 pb_reference(&memobj
->buf
, NULL
);
2370 static struct pipe_resource
*
2371 si_texture_from_memobj(struct pipe_screen
*screen
,
2372 const struct pipe_resource
*templ
,
2373 struct pipe_memory_object
*_memobj
,
2377 struct si_screen
*sscreen
= (struct si_screen
*)screen
;
2378 struct r600_memory_object
*memobj
= (struct r600_memory_object
*)_memobj
;
2379 struct si_texture
*tex
;
2380 struct radeon_surf surface
= {};
2381 struct radeon_bo_metadata metadata
= {};
2382 enum radeon_surf_mode array_mode
;
2384 struct pb_buffer
*buf
= NULL
;
2386 if (memobj
->b
.dedicated
) {
2387 sscreen
->ws
->buffer_get_metadata(memobj
->buf
, &metadata
);
2388 si_surface_import_metadata(sscreen
, &surface
, &metadata
,
2389 &array_mode
, &is_scanout
);
2392 * The bo metadata is unset for un-dedicated images. So we fall
2393 * back to linear. See answer to question 5 of the
2394 * VK_KHX_external_memory spec for some details.
2396 * It is possible that this case isn't going to work if the
2397 * surface pitch isn't correctly aligned by default.
2399 * In order to support it correctly we require multi-image
2400 * metadata to be syncrhonized between radv and radeonsi. The
2401 * semantics of associating multiple image metadata to a memory
2402 * object on the vulkan export side are not concretely defined
2405 * All the use cases we are aware of at the moment for memory
2406 * objects use dedicated allocations. So lets keep the initial
2407 * implementation simple.
2409 * A possible alternative is to attempt to reconstruct the
2410 * tiling information when the TexParameter TEXTURE_TILING_EXT
2413 array_mode
= RADEON_SURF_MODE_LINEAR_ALIGNED
;
2417 unsigned num_color_samples
= si_get_num_color_samples(sscreen
, templ
, true);
2419 r
= si_init_surface(sscreen
, &surface
, templ
, num_color_samples
,
2420 array_mode
, memobj
->stride
, offset
, true,
2421 is_scanout
, false, false);
2425 tex
= si_texture_create_object(screen
, templ
, num_color_samples
,
2426 memobj
->buf
, &surface
);
2430 /* si_texture_create_object doesn't increment refcount of
2431 * memobj->buf, so increment it here.
2433 pb_reference(&buf
, memobj
->buf
);
2435 tex
->buffer
.b
.is_shared
= true;
2436 tex
->buffer
.external_usage
= PIPE_HANDLE_USAGE_READ_WRITE
;
2438 si_apply_opaque_metadata(sscreen
, tex
, &metadata
);
2440 return &tex
->buffer
.b
.b
;
2443 static bool si_check_resource_capability(struct pipe_screen
*screen
,
2444 struct pipe_resource
*resource
,
2447 struct si_texture
*tex
= (struct si_texture
*)resource
;
2449 /* Buffers only support the linear flag. */
2450 if (resource
->target
== PIPE_BUFFER
)
2451 return (bind
& ~PIPE_BIND_LINEAR
) == 0;
2453 if (bind
& PIPE_BIND_LINEAR
&& !tex
->surface
.is_linear
)
2456 if (bind
& PIPE_BIND_SCANOUT
&& !tex
->surface
.is_displayable
)
2459 /* TODO: PIPE_BIND_CURSOR - do we care? */
2463 void si_init_screen_texture_functions(struct si_screen
*sscreen
)
2465 sscreen
->b
.resource_from_handle
= si_texture_from_handle
;
2466 sscreen
->b
.resource_get_handle
= si_texture_get_handle
;
2467 sscreen
->b
.resource_from_memobj
= si_texture_from_memobj
;
2468 sscreen
->b
.memobj_create_from_handle
= si_memobj_from_handle
;
2469 sscreen
->b
.memobj_destroy
= si_memobj_destroy
;
2470 sscreen
->b
.check_resource_capability
= si_check_resource_capability
;
2473 void si_init_context_texture_functions(struct si_context
*sctx
)
2475 sctx
->b
.create_surface
= si_create_surface
;
2476 sctx
->b
.surface_destroy
= si_surface_destroy
;