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_buffer
&& 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_buffer
&& 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 si_transfer
*stransfer
)
143 struct si_context
*sctx
= (struct si_context
*)ctx
;
144 struct pipe_transfer
*transfer
= (struct pipe_transfer
*)stransfer
;
145 struct pipe_resource
*dst
= &stransfer
->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 si_transfer
*stransfer
)
161 struct si_context
*sctx
= (struct si_context
*)ctx
;
162 struct pipe_transfer
*transfer
= (struct pipe_transfer
*)stransfer
;
163 struct pipe_resource
*dst
= transfer
->resource
;
164 struct pipe_resource
*src
= &stransfer
->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 enum radeon_surf_mode array_mode
,
224 unsigned pitch_in_bytes_override
,
228 bool is_flushed_depth
,
229 bool tc_compatible_htile
)
231 const struct util_format_description
*desc
=
232 util_format_description(ptex
->format
);
233 bool is_depth
, is_stencil
;
235 unsigned i
, bpe
, flags
= 0;
237 is_depth
= util_format_has_depth(desc
);
238 is_stencil
= util_format_has_stencil(desc
);
240 if (!is_flushed_depth
&&
241 ptex
->format
== PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
) {
242 bpe
= 4; /* stencil is allocated separately */
244 bpe
= util_format_get_blocksize(ptex
->format
);
245 assert(util_is_power_of_two_or_zero(bpe
));
248 if (!is_flushed_depth
&& is_depth
) {
249 flags
|= RADEON_SURF_ZBUFFER
;
251 if (tc_compatible_htile
&&
252 (sscreen
->info
.chip_class
>= GFX9
||
253 array_mode
== RADEON_SURF_MODE_2D
)) {
254 /* TC-compatible HTILE only supports Z32_FLOAT.
255 * GFX9 also supports Z16_UNORM.
256 * On VI, promote Z16 to Z32. DB->CB copies will convert
257 * the format for transfers.
259 if (sscreen
->info
.chip_class
== VI
)
262 flags
|= RADEON_SURF_TC_COMPATIBLE_HTILE
;
266 flags
|= RADEON_SURF_SBUFFER
;
269 if (sscreen
->info
.chip_class
>= VI
&&
270 (ptex
->flags
& SI_RESOURCE_FLAG_DISABLE_DCC
||
271 ptex
->format
== PIPE_FORMAT_R9G9B9E5_FLOAT
||
272 (ptex
->nr_samples
>= 2 && !sscreen
->dcc_msaa_allowed
)))
273 flags
|= RADEON_SURF_DISABLE_DCC
;
275 /* Stoney: 128bpp MSAA textures randomly fail piglit tests with DCC. */
276 if (sscreen
->info
.family
== CHIP_STONEY
&&
277 bpe
== 16 && ptex
->nr_samples
>= 2)
278 flags
|= RADEON_SURF_DISABLE_DCC
;
280 /* VI: DCC clear for 4x and 8x MSAA array textures unimplemented. */
281 if (sscreen
->info
.chip_class
== VI
&&
282 ptex
->nr_storage_samples
>= 4 &&
283 ptex
->array_size
> 1)
284 flags
|= RADEON_SURF_DISABLE_DCC
;
286 /* GFX9: DCC clear for 4x and 8x MSAA textures unimplemented. */
287 if (sscreen
->info
.chip_class
>= GFX9
&&
288 ptex
->nr_storage_samples
>= 4)
289 flags
|= RADEON_SURF_DISABLE_DCC
;
291 if (ptex
->bind
& PIPE_BIND_SCANOUT
|| is_scanout
) {
292 /* This should catch bugs in gallium users setting incorrect flags. */
293 assert(ptex
->nr_samples
<= 1 &&
294 ptex
->array_size
== 1 &&
296 ptex
->last_level
== 0 &&
297 !(flags
& RADEON_SURF_Z_OR_SBUFFER
));
299 flags
|= RADEON_SURF_SCANOUT
;
302 if (ptex
->bind
& PIPE_BIND_SHARED
)
303 flags
|= RADEON_SURF_SHAREABLE
;
305 flags
|= RADEON_SURF_IMPORTED
| RADEON_SURF_SHAREABLE
;
306 if (!(ptex
->flags
& SI_RESOURCE_FLAG_FORCE_MSAA_TILING
))
307 flags
|= RADEON_SURF_OPTIMIZE_FOR_SPACE
;
309 r
= sscreen
->ws
->surface_init(sscreen
->ws
, ptex
, flags
, bpe
,
310 array_mode
, surface
);
315 unsigned pitch
= pitch_in_bytes_override
/ bpe
;
317 if (sscreen
->info
.chip_class
>= GFX9
) {
319 surface
->u
.gfx9
.surf_pitch
= pitch
;
320 surface
->u
.gfx9
.surf_slice_size
=
321 (uint64_t)pitch
* surface
->u
.gfx9
.surf_height
* bpe
;
323 surface
->u
.gfx9
.surf_offset
= offset
;
326 surface
->u
.legacy
.level
[0].nblk_x
= pitch
;
327 surface
->u
.legacy
.level
[0].slice_size_dw
=
328 ((uint64_t)pitch
* surface
->u
.legacy
.level
[0].nblk_y
* bpe
) / 4;
331 for (i
= 0; i
< ARRAY_SIZE(surface
->u
.legacy
.level
); ++i
)
332 surface
->u
.legacy
.level
[i
].offset
+= offset
;
338 static void si_get_display_metadata(struct si_screen
*sscreen
,
339 struct radeon_surf
*surf
,
340 struct radeon_bo_metadata
*metadata
,
341 enum radeon_surf_mode
*array_mode
,
344 if (sscreen
->info
.chip_class
>= GFX9
) {
345 if (metadata
->u
.gfx9
.swizzle_mode
> 0)
346 *array_mode
= RADEON_SURF_MODE_2D
;
348 *array_mode
= RADEON_SURF_MODE_LINEAR_ALIGNED
;
350 *is_scanout
= metadata
->u
.gfx9
.swizzle_mode
== 0 ||
351 metadata
->u
.gfx9
.swizzle_mode
% 4 == 2;
353 surf
->u
.gfx9
.surf
.swizzle_mode
= metadata
->u
.gfx9
.swizzle_mode
;
355 surf
->u
.legacy
.pipe_config
= metadata
->u
.legacy
.pipe_config
;
356 surf
->u
.legacy
.bankw
= metadata
->u
.legacy
.bankw
;
357 surf
->u
.legacy
.bankh
= metadata
->u
.legacy
.bankh
;
358 surf
->u
.legacy
.tile_split
= metadata
->u
.legacy
.tile_split
;
359 surf
->u
.legacy
.mtilea
= metadata
->u
.legacy
.mtilea
;
360 surf
->u
.legacy
.num_banks
= metadata
->u
.legacy
.num_banks
;
362 if (metadata
->u
.legacy
.macrotile
== RADEON_LAYOUT_TILED
)
363 *array_mode
= RADEON_SURF_MODE_2D
;
364 else if (metadata
->u
.legacy
.microtile
== RADEON_LAYOUT_TILED
)
365 *array_mode
= RADEON_SURF_MODE_1D
;
367 *array_mode
= RADEON_SURF_MODE_LINEAR_ALIGNED
;
369 *is_scanout
= metadata
->u
.legacy
.scanout
;
373 void si_eliminate_fast_color_clear(struct si_context
*sctx
,
374 struct si_texture
*tex
)
376 struct si_screen
*sscreen
= sctx
->screen
;
377 struct pipe_context
*ctx
= &sctx
->b
;
379 if (ctx
== sscreen
->aux_context
)
380 mtx_lock(&sscreen
->aux_context_lock
);
382 unsigned n
= sctx
->num_decompress_calls
;
383 ctx
->flush_resource(ctx
, &tex
->buffer
.b
.b
);
385 /* Flush only if any fast clear elimination took place. */
386 if (n
!= sctx
->num_decompress_calls
)
387 ctx
->flush(ctx
, NULL
, 0);
389 if (ctx
== sscreen
->aux_context
)
390 mtx_unlock(&sscreen
->aux_context_lock
);
393 void si_texture_discard_cmask(struct si_screen
*sscreen
,
394 struct si_texture
*tex
)
396 if (!tex
->cmask_buffer
)
399 assert(tex
->buffer
.b
.b
.nr_samples
<= 1);
402 tex
->cmask_base_address_reg
= tex
->buffer
.gpu_address
>> 8;
403 tex
->dirty_level_mask
= 0;
405 tex
->cb_color_info
&= ~S_028C70_FAST_CLEAR(1);
407 if (tex
->cmask_buffer
!= &tex
->buffer
)
408 si_resource_reference(&tex
->cmask_buffer
, NULL
);
410 tex
->cmask_buffer
= NULL
;
412 /* Notify all contexts about the change. */
413 p_atomic_inc(&sscreen
->dirty_tex_counter
);
414 p_atomic_inc(&sscreen
->compressed_colortex_counter
);
417 static bool si_can_disable_dcc(struct si_texture
*tex
)
419 /* We can't disable DCC if it can be written by another process. */
420 return tex
->dcc_offset
&&
421 (!tex
->buffer
.b
.is_shared
||
422 !(tex
->buffer
.external_usage
& PIPE_HANDLE_USAGE_FRAMEBUFFER_WRITE
));
425 static bool si_texture_discard_dcc(struct si_screen
*sscreen
,
426 struct si_texture
*tex
)
428 if (!si_can_disable_dcc(tex
))
431 assert(tex
->dcc_separate_buffer
== NULL
);
436 /* Notify all contexts about the change. */
437 p_atomic_inc(&sscreen
->dirty_tex_counter
);
442 * Disable DCC for the texture. (first decompress, then discard metadata).
444 * There is unresolved multi-context synchronization issue between
445 * screen::aux_context and the current context. If applications do this with
446 * multiple contexts, it's already undefined behavior for them and we don't
447 * have to worry about that. The scenario is:
449 * If context 1 disables DCC and context 2 has queued commands that write
450 * to the texture via CB with DCC enabled, and the order of operations is
452 * context 2 queues draw calls rendering to the texture, but doesn't flush
453 * context 1 disables DCC and flushes
454 * context 1 & 2 reset descriptors and FB state
455 * context 2 flushes (new compressed tiles written by the draw calls)
456 * context 1 & 2 read garbage, because DCC is disabled, yet there are
459 * \param sctx the current context if you have one, or sscreen->aux_context
462 bool si_texture_disable_dcc(struct si_context
*sctx
,
463 struct si_texture
*tex
)
465 struct si_screen
*sscreen
= sctx
->screen
;
467 if (!sctx
->has_graphics
)
468 return si_texture_discard_dcc(sscreen
, tex
);
470 if (!si_can_disable_dcc(tex
))
473 if (&sctx
->b
== sscreen
->aux_context
)
474 mtx_lock(&sscreen
->aux_context_lock
);
476 /* Decompress DCC. */
477 si_decompress_dcc(sctx
, tex
);
478 sctx
->b
.flush(&sctx
->b
, NULL
, 0);
480 if (&sctx
->b
== sscreen
->aux_context
)
481 mtx_unlock(&sscreen
->aux_context_lock
);
483 return si_texture_discard_dcc(sscreen
, tex
);
486 static void si_reallocate_texture_inplace(struct si_context
*sctx
,
487 struct si_texture
*tex
,
488 unsigned new_bind_flag
,
489 bool invalidate_storage
)
491 struct pipe_screen
*screen
= sctx
->b
.screen
;
492 struct si_texture
*new_tex
;
493 struct pipe_resource templ
= tex
->buffer
.b
.b
;
496 templ
.bind
|= new_bind_flag
;
498 if (tex
->buffer
.b
.is_shared
)
501 if (new_bind_flag
== PIPE_BIND_LINEAR
) {
502 if (tex
->surface
.is_linear
)
505 /* This fails with MSAA, depth, and compressed textures. */
506 if (si_choose_tiling(sctx
->screen
, &templ
, false) !=
507 RADEON_SURF_MODE_LINEAR_ALIGNED
)
511 new_tex
= (struct si_texture
*)screen
->resource_create(screen
, &templ
);
515 /* Copy the pixels to the new texture. */
516 if (!invalidate_storage
) {
517 for (i
= 0; i
<= templ
.last_level
; i
++) {
521 u_minify(templ
.width0
, i
), u_minify(templ
.height0
, i
),
522 util_num_layers(&templ
, i
), &box
);
524 sctx
->dma_copy(&sctx
->b
, &new_tex
->buffer
.b
.b
, i
, 0, 0, 0,
525 &tex
->buffer
.b
.b
, i
, &box
);
529 if (new_bind_flag
== PIPE_BIND_LINEAR
) {
530 si_texture_discard_cmask(sctx
->screen
, tex
);
531 si_texture_discard_dcc(sctx
->screen
, tex
);
534 /* Replace the structure fields of tex. */
535 tex
->buffer
.b
.b
.bind
= templ
.bind
;
536 pb_reference(&tex
->buffer
.buf
, new_tex
->buffer
.buf
);
537 tex
->buffer
.gpu_address
= new_tex
->buffer
.gpu_address
;
538 tex
->buffer
.vram_usage
= new_tex
->buffer
.vram_usage
;
539 tex
->buffer
.gart_usage
= new_tex
->buffer
.gart_usage
;
540 tex
->buffer
.bo_size
= new_tex
->buffer
.bo_size
;
541 tex
->buffer
.bo_alignment
= new_tex
->buffer
.bo_alignment
;
542 tex
->buffer
.domains
= new_tex
->buffer
.domains
;
543 tex
->buffer
.flags
= new_tex
->buffer
.flags
;
545 tex
->surface
= new_tex
->surface
;
546 tex
->size
= new_tex
->size
;
547 si_texture_reference(&tex
->flushed_depth_texture
,
548 new_tex
->flushed_depth_texture
);
550 tex
->fmask_offset
= new_tex
->fmask_offset
;
551 tex
->cmask_offset
= new_tex
->cmask_offset
;
552 tex
->cmask_base_address_reg
= new_tex
->cmask_base_address_reg
;
554 if (tex
->cmask_buffer
== &tex
->buffer
)
555 tex
->cmask_buffer
= NULL
;
557 si_resource_reference(&tex
->cmask_buffer
, NULL
);
559 if (new_tex
->cmask_buffer
== &new_tex
->buffer
)
560 tex
->cmask_buffer
= &tex
->buffer
;
562 si_resource_reference(&tex
->cmask_buffer
, new_tex
->cmask_buffer
);
564 tex
->dcc_offset
= new_tex
->dcc_offset
;
565 tex
->cb_color_info
= new_tex
->cb_color_info
;
566 memcpy(tex
->color_clear_value
, new_tex
->color_clear_value
,
567 sizeof(tex
->color_clear_value
));
568 tex
->last_msaa_resolve_target_micro_mode
= new_tex
->last_msaa_resolve_target_micro_mode
;
570 tex
->htile_offset
= new_tex
->htile_offset
;
571 tex
->depth_clear_value
= new_tex
->depth_clear_value
;
572 tex
->dirty_level_mask
= new_tex
->dirty_level_mask
;
573 tex
->stencil_dirty_level_mask
= new_tex
->stencil_dirty_level_mask
;
574 tex
->db_render_format
= new_tex
->db_render_format
;
575 tex
->stencil_clear_value
= new_tex
->stencil_clear_value
;
576 tex
->tc_compatible_htile
= new_tex
->tc_compatible_htile
;
577 tex
->depth_cleared
= new_tex
->depth_cleared
;
578 tex
->stencil_cleared
= new_tex
->stencil_cleared
;
579 tex
->upgraded_depth
= new_tex
->upgraded_depth
;
580 tex
->db_compatible
= new_tex
->db_compatible
;
581 tex
->can_sample_z
= new_tex
->can_sample_z
;
582 tex
->can_sample_s
= new_tex
->can_sample_s
;
584 tex
->separate_dcc_dirty
= new_tex
->separate_dcc_dirty
;
585 tex
->dcc_gather_statistics
= new_tex
->dcc_gather_statistics
;
586 si_resource_reference(&tex
->dcc_separate_buffer
,
587 new_tex
->dcc_separate_buffer
);
588 si_resource_reference(&tex
->last_dcc_separate_buffer
,
589 new_tex
->last_dcc_separate_buffer
);
591 if (new_bind_flag
== PIPE_BIND_LINEAR
) {
592 assert(!tex
->htile_offset
);
593 assert(!tex
->cmask_buffer
);
594 assert(!tex
->surface
.fmask_size
);
595 assert(!tex
->dcc_offset
);
596 assert(!tex
->is_depth
);
599 si_texture_reference(&new_tex
, NULL
);
601 p_atomic_inc(&sctx
->screen
->dirty_tex_counter
);
604 static uint32_t si_get_bo_metadata_word1(struct si_screen
*sscreen
)
606 return (ATI_VENDOR_ID
<< 16) | sscreen
->info
.pci_id
;
609 static void si_set_tex_bo_metadata(struct si_screen
*sscreen
,
610 struct si_texture
*tex
)
612 struct radeon_surf
*surface
= &tex
->surface
;
613 struct pipe_resource
*res
= &tex
->buffer
.b
.b
;
614 struct radeon_bo_metadata md
;
616 memset(&md
, 0, sizeof(md
));
618 if (sscreen
->info
.chip_class
>= GFX9
) {
619 md
.u
.gfx9
.swizzle_mode
= surface
->u
.gfx9
.surf
.swizzle_mode
;
621 md
.u
.legacy
.microtile
= surface
->u
.legacy
.level
[0].mode
>= RADEON_SURF_MODE_1D
?
622 RADEON_LAYOUT_TILED
: RADEON_LAYOUT_LINEAR
;
623 md
.u
.legacy
.macrotile
= surface
->u
.legacy
.level
[0].mode
>= RADEON_SURF_MODE_2D
?
624 RADEON_LAYOUT_TILED
: RADEON_LAYOUT_LINEAR
;
625 md
.u
.legacy
.pipe_config
= surface
->u
.legacy
.pipe_config
;
626 md
.u
.legacy
.bankw
= surface
->u
.legacy
.bankw
;
627 md
.u
.legacy
.bankh
= surface
->u
.legacy
.bankh
;
628 md
.u
.legacy
.tile_split
= surface
->u
.legacy
.tile_split
;
629 md
.u
.legacy
.mtilea
= surface
->u
.legacy
.mtilea
;
630 md
.u
.legacy
.num_banks
= surface
->u
.legacy
.num_banks
;
631 md
.u
.legacy
.stride
= surface
->u
.legacy
.level
[0].nblk_x
* surface
->bpe
;
632 md
.u
.legacy
.scanout
= (surface
->flags
& RADEON_SURF_SCANOUT
) != 0;
635 assert(tex
->dcc_separate_buffer
== NULL
);
636 assert(tex
->surface
.fmask_size
== 0);
638 /* Metadata image format format version 1:
639 * [0] = 1 (metadata format identifier)
640 * [1] = (VENDOR_ID << 16) | PCI_ID
641 * [2:9] = image descriptor for the whole resource
642 * [2] is always 0, because the base address is cleared
643 * [9] is the DCC offset bits [39:8] from the beginning of
645 * [10:10+LAST_LEVEL] = mipmap level offset bits [39:8] for each level
648 md
.metadata
[0] = 1; /* metadata image format version 1 */
650 /* TILE_MODE_INDEX is ambiguous without a PCI ID. */
651 md
.metadata
[1] = si_get_bo_metadata_word1(sscreen
);
653 static const unsigned char swizzle
[] = {
659 bool is_array
= util_texture_is_array(res
->target
);
662 si_make_texture_descriptor(sscreen
, tex
, true,
663 res
->target
, res
->format
,
664 swizzle
, 0, res
->last_level
, 0,
665 is_array
? res
->array_size
- 1 : 0,
666 res
->width0
, res
->height0
, res
->depth0
,
669 si_set_mutable_tex_desc_fields(sscreen
, tex
, &tex
->surface
.u
.legacy
.level
[0],
670 0, 0, tex
->surface
.blk_w
, false, desc
);
672 /* Clear the base address and set the relative DCC offset. */
674 desc
[1] &= C_008F14_BASE_ADDRESS_HI
;
675 desc
[7] = tex
->dcc_offset
>> 8;
677 /* Dwords [2:9] contain the image descriptor. */
678 memcpy(&md
.metadata
[2], desc
, sizeof(desc
));
679 md
.size_metadata
= 10 * 4;
681 /* Dwords [10:..] contain the mipmap level offsets. */
682 if (sscreen
->info
.chip_class
<= VI
) {
683 for (unsigned i
= 0; i
<= res
->last_level
; i
++)
684 md
.metadata
[10+i
] = tex
->surface
.u
.legacy
.level
[i
].offset
>> 8;
686 md
.size_metadata
+= (1 + res
->last_level
) * 4;
689 sscreen
->ws
->buffer_set_metadata(tex
->buffer
.buf
, &md
);
692 static void si_get_opaque_metadata(struct si_screen
*sscreen
,
693 struct si_texture
*tex
,
694 struct radeon_bo_metadata
*md
)
696 uint32_t *desc
= &md
->metadata
[2];
698 if (sscreen
->info
.chip_class
< VI
)
701 /* Return if DCC is enabled. The texture should be set up with it
704 if (md
->size_metadata
>= 10 * 4 && /* at least 2(header) + 8(desc) dwords */
705 md
->metadata
[0] != 0 &&
706 md
->metadata
[1] == si_get_bo_metadata_word1(sscreen
) &&
707 G_008F28_COMPRESSION_EN(desc
[6])) {
708 tex
->dcc_offset
= (uint64_t)desc
[7] << 8;
712 /* Disable DCC. These are always set by texture_from_handle and must
718 static boolean
si_texture_get_handle(struct pipe_screen
* screen
,
719 struct pipe_context
*ctx
,
720 struct pipe_resource
*resource
,
721 struct winsys_handle
*whandle
,
724 struct si_screen
*sscreen
= (struct si_screen
*)screen
;
725 struct si_context
*sctx
;
726 struct si_resource
*res
= si_resource(resource
);
727 struct si_texture
*tex
= (struct si_texture
*)resource
;
728 bool update_metadata
= false;
729 unsigned stride
, offset
, slice_size
;
732 ctx
= threaded_context_unwrap_sync(ctx
);
733 sctx
= (struct si_context
*)(ctx
? ctx
: sscreen
->aux_context
);
735 if (resource
->target
!= PIPE_BUFFER
) {
736 /* This is not supported now, but it might be required for OpenCL
737 * interop in the future.
739 if (resource
->nr_samples
> 1 || tex
->is_depth
)
742 /* Move a suballocated texture into a non-suballocated allocation. */
743 if (sscreen
->ws
->buffer_is_suballocated(res
->buf
) ||
744 tex
->surface
.tile_swizzle
||
745 (tex
->buffer
.flags
& RADEON_FLAG_NO_INTERPROCESS_SHARING
&&
746 sscreen
->info
.has_local_buffers
&&
747 whandle
->type
!= WINSYS_HANDLE_TYPE_KMS
)) {
748 assert(!res
->b
.is_shared
);
749 si_reallocate_texture_inplace(sctx
, tex
,
750 PIPE_BIND_SHARED
, false);
752 assert(res
->b
.b
.bind
& PIPE_BIND_SHARED
);
753 assert(res
->flags
& RADEON_FLAG_NO_SUBALLOC
);
754 assert(!(res
->flags
& RADEON_FLAG_NO_INTERPROCESS_SHARING
));
755 assert(tex
->surface
.tile_swizzle
== 0);
758 /* Since shader image stores don't support DCC on VI,
759 * disable it for external clients that want write
762 if (usage
& PIPE_HANDLE_USAGE_SHADER_WRITE
&& tex
->dcc_offset
) {
763 if (si_texture_disable_dcc(sctx
, tex
)) {
764 update_metadata
= true;
765 /* si_texture_disable_dcc flushes the context */
770 if (!(usage
& PIPE_HANDLE_USAGE_EXPLICIT_FLUSH
) &&
771 (tex
->cmask_buffer
|| tex
->dcc_offset
)) {
772 /* Eliminate fast clear (both CMASK and DCC) */
773 si_eliminate_fast_color_clear(sctx
, tex
);
774 /* eliminate_fast_color_clear flushes the context */
777 /* Disable CMASK if flush_resource isn't going
780 if (tex
->cmask_buffer
)
781 si_texture_discard_cmask(sscreen
, tex
);
785 if (!res
->b
.is_shared
|| update_metadata
)
786 si_set_tex_bo_metadata(sscreen
, tex
);
788 if (sscreen
->info
.chip_class
>= GFX9
) {
789 offset
= tex
->surface
.u
.gfx9
.surf_offset
;
790 stride
= tex
->surface
.u
.gfx9
.surf_pitch
*
792 slice_size
= tex
->surface
.u
.gfx9
.surf_slice_size
;
794 offset
= tex
->surface
.u
.legacy
.level
[0].offset
;
795 stride
= tex
->surface
.u
.legacy
.level
[0].nblk_x
*
797 slice_size
= (uint64_t)tex
->surface
.u
.legacy
.level
[0].slice_size_dw
* 4;
800 /* Buffer exports are for the OpenCL interop. */
801 /* Move a suballocated buffer into a non-suballocated allocation. */
802 if (sscreen
->ws
->buffer_is_suballocated(res
->buf
) ||
803 /* A DMABUF export always fails if the BO is local. */
804 (tex
->buffer
.flags
& RADEON_FLAG_NO_INTERPROCESS_SHARING
&&
805 sscreen
->info
.has_local_buffers
)) {
806 assert(!res
->b
.is_shared
);
808 /* Allocate a new buffer with PIPE_BIND_SHARED. */
809 struct pipe_resource templ
= res
->b
.b
;
810 templ
.bind
|= PIPE_BIND_SHARED
;
812 struct pipe_resource
*newb
=
813 screen
->resource_create(screen
, &templ
);
817 /* Copy the old buffer contents to the new one. */
819 u_box_1d(0, newb
->width0
, &box
);
820 sctx
->b
.resource_copy_region(&sctx
->b
, newb
, 0, 0, 0, 0,
823 /* Move the new buffer storage to the old pipe_resource. */
824 si_replace_buffer_storage(&sctx
->b
, &res
->b
.b
, newb
);
825 pipe_resource_reference(&newb
, NULL
);
827 assert(res
->b
.b
.bind
& PIPE_BIND_SHARED
);
828 assert(res
->flags
& RADEON_FLAG_NO_SUBALLOC
);
838 sctx
->b
.flush(&sctx
->b
, NULL
, 0);
840 if (res
->b
.is_shared
) {
841 /* USAGE_EXPLICIT_FLUSH must be cleared if at least one user
844 res
->external_usage
|= usage
& ~PIPE_HANDLE_USAGE_EXPLICIT_FLUSH
;
845 if (!(usage
& PIPE_HANDLE_USAGE_EXPLICIT_FLUSH
))
846 res
->external_usage
&= ~PIPE_HANDLE_USAGE_EXPLICIT_FLUSH
;
848 res
->b
.is_shared
= true;
849 res
->external_usage
= usage
;
852 return sscreen
->ws
->buffer_get_handle(res
->buf
, stride
, offset
,
853 slice_size
, whandle
);
856 static void si_texture_destroy(struct pipe_screen
*screen
,
857 struct pipe_resource
*ptex
)
859 struct si_texture
*tex
= (struct si_texture
*)ptex
;
860 struct si_resource
*resource
= &tex
->buffer
;
862 si_texture_reference(&tex
->flushed_depth_texture
, NULL
);
864 if (tex
->cmask_buffer
!= &tex
->buffer
) {
865 si_resource_reference(&tex
->cmask_buffer
, NULL
);
867 pb_reference(&resource
->buf
, NULL
);
868 si_resource_reference(&tex
->dcc_separate_buffer
, NULL
);
869 si_resource_reference(&tex
->last_dcc_separate_buffer
, NULL
);
873 static const struct u_resource_vtbl si_texture_vtbl
;
875 static void si_texture_get_htile_size(struct si_screen
*sscreen
,
876 struct si_texture
*tex
)
878 unsigned cl_width
, cl_height
, width
, height
;
879 unsigned slice_elements
, slice_bytes
, pipe_interleave_bytes
, base_align
;
880 unsigned num_pipes
= sscreen
->info
.num_tile_pipes
;
882 assert(sscreen
->info
.chip_class
<= VI
);
884 tex
->surface
.htile_size
= 0;
886 if (tex
->surface
.u
.legacy
.level
[0].mode
== RADEON_SURF_MODE_1D
&&
887 !sscreen
->info
.htile_cmask_support_1d_tiling
)
890 /* Overalign HTILE on P2 configs to work around GPU hangs in
891 * piglit/depthstencil-render-miplevels 585.
893 * This has been confirmed to help Kabini & Stoney, where the hangs
894 * are always reproducible. I think I have seen the test hang
895 * on Carrizo too, though it was very rare there.
897 if (sscreen
->info
.chip_class
>= CIK
&& num_pipes
< 4)
926 width
= align(tex
->surface
.u
.legacy
.level
[0].nblk_x
, cl_width
* 8);
927 height
= align(tex
->surface
.u
.legacy
.level
[0].nblk_y
, cl_height
* 8);
929 slice_elements
= (width
* height
) / (8 * 8);
930 slice_bytes
= slice_elements
* 4;
932 pipe_interleave_bytes
= sscreen
->info
.pipe_interleave_bytes
;
933 base_align
= num_pipes
* pipe_interleave_bytes
;
935 tex
->surface
.htile_alignment
= base_align
;
936 tex
->surface
.htile_size
=
937 util_num_layers(&tex
->buffer
.b
.b
, 0) *
938 align(slice_bytes
, base_align
);
941 static void si_texture_allocate_htile(struct si_screen
*sscreen
,
942 struct si_texture
*tex
)
944 if (sscreen
->info
.chip_class
<= VI
&& !tex
->tc_compatible_htile
)
945 si_texture_get_htile_size(sscreen
, tex
);
947 if (!tex
->surface
.htile_size
)
950 tex
->htile_offset
= align(tex
->size
, tex
->surface
.htile_alignment
);
951 tex
->size
= tex
->htile_offset
+ tex
->surface
.htile_size
;
954 void si_print_texture_info(struct si_screen
*sscreen
,
955 struct si_texture
*tex
, struct u_log_context
*log
)
959 /* Common parameters. */
960 u_log_printf(log
, " Info: npix_x=%u, npix_y=%u, npix_z=%u, blk_w=%u, "
961 "blk_h=%u, array_size=%u, last_level=%u, "
962 "bpe=%u, nsamples=%u, flags=0x%x, %s\n",
963 tex
->buffer
.b
.b
.width0
, tex
->buffer
.b
.b
.height0
,
964 tex
->buffer
.b
.b
.depth0
, tex
->surface
.blk_w
,
966 tex
->buffer
.b
.b
.array_size
, tex
->buffer
.b
.b
.last_level
,
967 tex
->surface
.bpe
, tex
->buffer
.b
.b
.nr_samples
,
968 tex
->surface
.flags
, util_format_short_name(tex
->buffer
.b
.b
.format
));
970 if (sscreen
->info
.chip_class
>= GFX9
) {
971 u_log_printf(log
, " Surf: size=%"PRIu64
", slice_size=%"PRIu64
", "
972 "alignment=%u, swmode=%u, epitch=%u, pitch=%u\n",
973 tex
->surface
.surf_size
,
974 tex
->surface
.u
.gfx9
.surf_slice_size
,
975 tex
->surface
.surf_alignment
,
976 tex
->surface
.u
.gfx9
.surf
.swizzle_mode
,
977 tex
->surface
.u
.gfx9
.surf
.epitch
,
978 tex
->surface
.u
.gfx9
.surf_pitch
);
980 if (tex
->surface
.fmask_size
) {
981 u_log_printf(log
, " FMASK: offset=%"PRIu64
", size=%"PRIu64
", "
982 "alignment=%u, swmode=%u, epitch=%u\n",
984 tex
->surface
.fmask_size
,
985 tex
->surface
.fmask_alignment
,
986 tex
->surface
.u
.gfx9
.fmask
.swizzle_mode
,
987 tex
->surface
.u
.gfx9
.fmask
.epitch
);
990 if (tex
->cmask_buffer
) {
991 u_log_printf(log
, " CMask: offset=%"PRIu64
", size=%u, "
992 "alignment=%u, rb_aligned=%u, pipe_aligned=%u\n",
994 tex
->surface
.cmask_size
,
995 tex
->surface
.cmask_alignment
,
996 tex
->surface
.u
.gfx9
.cmask
.rb_aligned
,
997 tex
->surface
.u
.gfx9
.cmask
.pipe_aligned
);
1000 if (tex
->htile_offset
) {
1001 u_log_printf(log
, " HTile: offset=%"PRIu64
", size=%u, alignment=%u, "
1002 "rb_aligned=%u, pipe_aligned=%u\n",
1004 tex
->surface
.htile_size
,
1005 tex
->surface
.htile_alignment
,
1006 tex
->surface
.u
.gfx9
.htile
.rb_aligned
,
1007 tex
->surface
.u
.gfx9
.htile
.pipe_aligned
);
1010 if (tex
->dcc_offset
) {
1011 u_log_printf(log
, " DCC: offset=%"PRIu64
", size=%u, "
1012 "alignment=%u, pitch_max=%u, num_dcc_levels=%u\n",
1013 tex
->dcc_offset
, tex
->surface
.dcc_size
,
1014 tex
->surface
.dcc_alignment
,
1015 tex
->surface
.u
.gfx9
.dcc_pitch_max
,
1016 tex
->surface
.num_dcc_levels
);
1019 if (tex
->surface
.u
.gfx9
.stencil_offset
) {
1020 u_log_printf(log
, " Stencil: offset=%"PRIu64
", swmode=%u, epitch=%u\n",
1021 tex
->surface
.u
.gfx9
.stencil_offset
,
1022 tex
->surface
.u
.gfx9
.stencil
.swizzle_mode
,
1023 tex
->surface
.u
.gfx9
.stencil
.epitch
);
1028 u_log_printf(log
, " Layout: size=%"PRIu64
", alignment=%u, bankw=%u, "
1029 "bankh=%u, nbanks=%u, mtilea=%u, tilesplit=%u, pipeconfig=%u, scanout=%u\n",
1030 tex
->surface
.surf_size
, tex
->surface
.surf_alignment
, tex
->surface
.u
.legacy
.bankw
,
1031 tex
->surface
.u
.legacy
.bankh
, tex
->surface
.u
.legacy
.num_banks
, tex
->surface
.u
.legacy
.mtilea
,
1032 tex
->surface
.u
.legacy
.tile_split
, tex
->surface
.u
.legacy
.pipe_config
,
1033 (tex
->surface
.flags
& RADEON_SURF_SCANOUT
) != 0);
1035 if (tex
->surface
.fmask_size
)
1036 u_log_printf(log
, " FMask: offset=%"PRIu64
", size=%"PRIu64
", alignment=%u, pitch_in_pixels=%u, "
1037 "bankh=%u, slice_tile_max=%u, tile_mode_index=%u\n",
1038 tex
->fmask_offset
, tex
->surface
.fmask_size
, tex
->surface
.fmask_alignment
,
1039 tex
->surface
.u
.legacy
.fmask
.pitch_in_pixels
,
1040 tex
->surface
.u
.legacy
.fmask
.bankh
,
1041 tex
->surface
.u
.legacy
.fmask
.slice_tile_max
,
1042 tex
->surface
.u
.legacy
.fmask
.tiling_index
);
1044 if (tex
->cmask_buffer
)
1045 u_log_printf(log
, " CMask: offset=%"PRIu64
", size=%u, alignment=%u, "
1046 "slice_tile_max=%u\n",
1047 tex
->cmask_offset
, tex
->surface
.cmask_size
, tex
->surface
.cmask_alignment
,
1048 tex
->surface
.u
.legacy
.cmask_slice_tile_max
);
1050 if (tex
->htile_offset
)
1051 u_log_printf(log
, " HTile: offset=%"PRIu64
", size=%u, "
1052 "alignment=%u, TC_compatible = %u\n",
1053 tex
->htile_offset
, tex
->surface
.htile_size
,
1054 tex
->surface
.htile_alignment
,
1055 tex
->tc_compatible_htile
);
1057 if (tex
->dcc_offset
) {
1058 u_log_printf(log
, " DCC: offset=%"PRIu64
", size=%u, alignment=%u\n",
1059 tex
->dcc_offset
, tex
->surface
.dcc_size
,
1060 tex
->surface
.dcc_alignment
);
1061 for (i
= 0; i
<= tex
->buffer
.b
.b
.last_level
; i
++)
1062 u_log_printf(log
, " DCCLevel[%i]: enabled=%u, offset=%u, "
1063 "fast_clear_size=%u\n",
1064 i
, i
< tex
->surface
.num_dcc_levels
,
1065 tex
->surface
.u
.legacy
.level
[i
].dcc_offset
,
1066 tex
->surface
.u
.legacy
.level
[i
].dcc_fast_clear_size
);
1069 for (i
= 0; i
<= tex
->buffer
.b
.b
.last_level
; i
++)
1070 u_log_printf(log
, " Level[%i]: offset=%"PRIu64
", slice_size=%"PRIu64
", "
1071 "npix_x=%u, npix_y=%u, npix_z=%u, nblk_x=%u, nblk_y=%u, "
1072 "mode=%u, tiling_index = %u\n",
1073 i
, tex
->surface
.u
.legacy
.level
[i
].offset
,
1074 (uint64_t)tex
->surface
.u
.legacy
.level
[i
].slice_size_dw
* 4,
1075 u_minify(tex
->buffer
.b
.b
.width0
, i
),
1076 u_minify(tex
->buffer
.b
.b
.height0
, i
),
1077 u_minify(tex
->buffer
.b
.b
.depth0
, i
),
1078 tex
->surface
.u
.legacy
.level
[i
].nblk_x
,
1079 tex
->surface
.u
.legacy
.level
[i
].nblk_y
,
1080 tex
->surface
.u
.legacy
.level
[i
].mode
,
1081 tex
->surface
.u
.legacy
.tiling_index
[i
]);
1083 if (tex
->surface
.has_stencil
) {
1084 u_log_printf(log
, " StencilLayout: tilesplit=%u\n",
1085 tex
->surface
.u
.legacy
.stencil_tile_split
);
1086 for (i
= 0; i
<= tex
->buffer
.b
.b
.last_level
; i
++) {
1087 u_log_printf(log
, " StencilLevel[%i]: offset=%"PRIu64
", "
1088 "slice_size=%"PRIu64
", npix_x=%u, "
1089 "npix_y=%u, npix_z=%u, nblk_x=%u, nblk_y=%u, "
1090 "mode=%u, tiling_index = %u\n",
1091 i
, tex
->surface
.u
.legacy
.stencil_level
[i
].offset
,
1092 (uint64_t)tex
->surface
.u
.legacy
.stencil_level
[i
].slice_size_dw
* 4,
1093 u_minify(tex
->buffer
.b
.b
.width0
, i
),
1094 u_minify(tex
->buffer
.b
.b
.height0
, i
),
1095 u_minify(tex
->buffer
.b
.b
.depth0
, i
),
1096 tex
->surface
.u
.legacy
.stencil_level
[i
].nblk_x
,
1097 tex
->surface
.u
.legacy
.stencil_level
[i
].nblk_y
,
1098 tex
->surface
.u
.legacy
.stencil_level
[i
].mode
,
1099 tex
->surface
.u
.legacy
.stencil_tiling_index
[i
]);
1104 /* Common processing for si_texture_create and si_texture_from_handle */
1105 static struct si_texture
*
1106 si_texture_create_object(struct pipe_screen
*screen
,
1107 const struct pipe_resource
*base
,
1108 struct pb_buffer
*buf
,
1109 struct radeon_surf
*surface
)
1111 struct si_texture
*tex
;
1112 struct si_resource
*resource
;
1113 struct si_screen
*sscreen
= (struct si_screen
*)screen
;
1115 tex
= CALLOC_STRUCT(si_texture
);
1119 resource
= &tex
->buffer
;
1120 resource
->b
.b
= *base
;
1121 resource
->b
.b
.next
= NULL
;
1122 resource
->b
.vtbl
= &si_texture_vtbl
;
1123 pipe_reference_init(&resource
->b
.b
.reference
, 1);
1124 resource
->b
.b
.screen
= screen
;
1126 /* don't include stencil-only formats which we don't support for rendering */
1127 tex
->is_depth
= util_format_has_depth(util_format_description(tex
->buffer
.b
.b
.format
));
1129 tex
->surface
= *surface
;
1130 tex
->size
= tex
->surface
.surf_size
;
1132 tex
->tc_compatible_htile
= tex
->surface
.htile_size
!= 0 &&
1133 (tex
->surface
.flags
&
1134 RADEON_SURF_TC_COMPATIBLE_HTILE
);
1136 /* TC-compatible HTILE:
1137 * - VI only supports Z32_FLOAT.
1138 * - GFX9 only supports Z32_FLOAT and Z16_UNORM. */
1139 if (tex
->tc_compatible_htile
) {
1140 if (sscreen
->info
.chip_class
>= GFX9
&&
1141 base
->format
== PIPE_FORMAT_Z16_UNORM
)
1142 tex
->db_render_format
= base
->format
;
1144 tex
->db_render_format
= PIPE_FORMAT_Z32_FLOAT
;
1145 tex
->upgraded_depth
= base
->format
!= PIPE_FORMAT_Z32_FLOAT
&&
1146 base
->format
!= PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
;
1149 tex
->db_render_format
= base
->format
;
1152 /* Applies to GCN. */
1153 tex
->last_msaa_resolve_target_micro_mode
= tex
->surface
.micro_tile_mode
;
1155 /* Disable separate DCC at the beginning. DRI2 doesn't reuse buffers
1156 * between frames, so the only thing that can enable separate DCC
1157 * with DRI2 is multiple slow clears within a frame.
1159 tex
->ps_draw_ratio
= 0;
1161 if (tex
->is_depth
) {
1162 if (sscreen
->info
.chip_class
>= GFX9
) {
1163 tex
->can_sample_z
= true;
1164 tex
->can_sample_s
= true;
1166 tex
->can_sample_z
= !tex
->surface
.u
.legacy
.depth_adjusted
;
1167 tex
->can_sample_s
= !tex
->surface
.u
.legacy
.stencil_adjusted
;
1170 if (!(base
->flags
& (SI_RESOURCE_FLAG_TRANSFER
|
1171 SI_RESOURCE_FLAG_FLUSHED_DEPTH
))) {
1172 tex
->db_compatible
= true;
1174 if (!(sscreen
->debug_flags
& DBG(NO_HYPERZ
)))
1175 si_texture_allocate_htile(sscreen
, tex
);
1178 if (base
->nr_samples
> 1 &&
1180 !(sscreen
->debug_flags
& DBG(NO_FMASK
))) {
1181 /* Allocate FMASK. */
1182 tex
->fmask_offset
= align64(tex
->size
,
1183 tex
->surface
.fmask_alignment
);
1184 tex
->size
= tex
->fmask_offset
+ tex
->surface
.fmask_size
;
1186 /* Allocate CMASK. */
1187 tex
->cmask_offset
= align64(tex
->size
, tex
->surface
.cmask_alignment
);
1188 tex
->size
= tex
->cmask_offset
+ tex
->surface
.cmask_size
;
1189 tex
->cb_color_info
|= S_028C70_FAST_CLEAR(1);
1190 tex
->cmask_buffer
= &tex
->buffer
;
1192 if (!tex
->surface
.fmask_size
|| !tex
->surface
.cmask_size
)
1196 /* Shared textures must always set up DCC here.
1197 * If it's not present, it will be disabled by
1198 * apply_opaque_metadata later.
1200 if (tex
->surface
.dcc_size
&&
1201 (buf
|| !(sscreen
->debug_flags
& DBG(NO_DCC
))) &&
1202 !(tex
->surface
.flags
& RADEON_SURF_SCANOUT
)) {
1203 /* Reserve space for the DCC buffer. */
1204 tex
->dcc_offset
= align64(tex
->size
, tex
->surface
.dcc_alignment
);
1205 tex
->size
= tex
->dcc_offset
+ tex
->surface
.dcc_size
;
1209 /* Now create the backing buffer. */
1211 si_init_resource_fields(sscreen
, resource
, tex
->size
,
1212 tex
->surface
.surf_alignment
);
1214 if (!si_alloc_resource(sscreen
, resource
))
1217 resource
->buf
= buf
;
1218 resource
->gpu_address
= sscreen
->ws
->buffer_get_virtual_address(resource
->buf
);
1219 resource
->bo_size
= buf
->size
;
1220 resource
->bo_alignment
= buf
->alignment
;
1221 resource
->domains
= sscreen
->ws
->buffer_get_initial_domain(resource
->buf
);
1222 if (resource
->domains
& RADEON_DOMAIN_VRAM
)
1223 resource
->vram_usage
= buf
->size
;
1224 else if (resource
->domains
& RADEON_DOMAIN_GTT
)
1225 resource
->gart_usage
= buf
->size
;
1228 if (tex
->cmask_buffer
) {
1229 /* Initialize the cmask to 0xCC (= compressed state). */
1230 si_screen_clear_buffer(sscreen
, &tex
->cmask_buffer
->b
.b
,
1231 tex
->cmask_offset
, tex
->surface
.cmask_size
,
1234 if (tex
->htile_offset
) {
1235 uint32_t clear_value
= 0;
1237 if (sscreen
->info
.chip_class
>= GFX9
|| tex
->tc_compatible_htile
)
1238 clear_value
= 0x0000030F;
1240 si_screen_clear_buffer(sscreen
, &tex
->buffer
.b
.b
,
1242 tex
->surface
.htile_size
,
1246 /* Initialize DCC only if the texture is not being imported. */
1247 if (!buf
&& tex
->dcc_offset
) {
1248 /* Clear DCC to black for all tiles with DCC enabled.
1250 * This fixes corruption in 3DMark Slingshot Extreme, which
1251 * uses uninitialized textures, causing corruption.
1253 if (tex
->surface
.num_dcc_levels
== tex
->buffer
.b
.b
.last_level
+ 1 &&
1254 tex
->buffer
.b
.b
.nr_samples
<= 2) {
1255 /* Simple case - all tiles have DCC enabled. */
1256 si_screen_clear_buffer(sscreen
, &tex
->buffer
.b
.b
,
1258 tex
->surface
.dcc_size
,
1259 DCC_CLEAR_COLOR_0000
);
1260 } else if (sscreen
->info
.chip_class
>= GFX9
) {
1261 /* Clear to uncompressed. Clearing this to black is complicated. */
1262 si_screen_clear_buffer(sscreen
, &tex
->buffer
.b
.b
,
1264 tex
->surface
.dcc_size
,
1267 /* GFX8: Initialize mipmap levels and multisamples separately. */
1268 if (tex
->buffer
.b
.b
.nr_samples
>= 2) {
1269 /* Clearing this to black is complicated. */
1270 si_screen_clear_buffer(sscreen
, &tex
->buffer
.b
.b
,
1272 tex
->surface
.dcc_size
,
1275 /* Clear the enabled mipmap levels to black. */
1278 for (unsigned i
= 0; i
< tex
->surface
.num_dcc_levels
; i
++) {
1279 if (!tex
->surface
.u
.legacy
.level
[i
].dcc_fast_clear_size
)
1282 size
= tex
->surface
.u
.legacy
.level
[i
].dcc_offset
+
1283 tex
->surface
.u
.legacy
.level
[i
].dcc_fast_clear_size
;
1286 /* Mipmap levels with DCC. */
1288 si_screen_clear_buffer(sscreen
, &tex
->buffer
.b
.b
,
1289 tex
->dcc_offset
, size
,
1290 DCC_CLEAR_COLOR_0000
);
1292 /* Mipmap levels without DCC. */
1293 if (size
!= tex
->surface
.dcc_size
) {
1294 si_screen_clear_buffer(sscreen
, &tex
->buffer
.b
.b
,
1295 tex
->dcc_offset
+ size
,
1296 tex
->surface
.dcc_size
- size
,
1303 /* Initialize the CMASK base register value. */
1304 tex
->cmask_base_address_reg
=
1305 (tex
->buffer
.gpu_address
+ tex
->cmask_offset
) >> 8;
1307 if (sscreen
->debug_flags
& DBG(VM
)) {
1308 fprintf(stderr
, "VM start=0x%"PRIX64
" end=0x%"PRIX64
" | Texture %ix%ix%i, %i levels, %i samples, %s\n",
1309 tex
->buffer
.gpu_address
,
1310 tex
->buffer
.gpu_address
+ tex
->buffer
.buf
->size
,
1311 base
->width0
, base
->height0
, util_num_layers(base
, 0), base
->last_level
+1,
1312 base
->nr_samples
? base
->nr_samples
: 1, util_format_short_name(base
->format
));
1315 if (sscreen
->debug_flags
& DBG(TEX
)) {
1317 struct u_log_context log
;
1318 u_log_context_init(&log
);
1319 si_print_texture_info(sscreen
, tex
, &log
);
1320 u_log_new_page_print(&log
, stdout
);
1322 u_log_context_destroy(&log
);
1332 static enum radeon_surf_mode
1333 si_choose_tiling(struct si_screen
*sscreen
,
1334 const struct pipe_resource
*templ
, bool tc_compatible_htile
)
1336 const struct util_format_description
*desc
= util_format_description(templ
->format
);
1337 bool force_tiling
= templ
->flags
& SI_RESOURCE_FLAG_FORCE_MSAA_TILING
;
1338 bool is_depth_stencil
= util_format_is_depth_or_stencil(templ
->format
) &&
1339 !(templ
->flags
& SI_RESOURCE_FLAG_FLUSHED_DEPTH
);
1341 /* MSAA resources must be 2D tiled. */
1342 if (templ
->nr_samples
> 1)
1343 return RADEON_SURF_MODE_2D
;
1345 /* Transfer resources should be linear. */
1346 if (templ
->flags
& SI_RESOURCE_FLAG_TRANSFER
)
1347 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1349 /* Avoid Z/S decompress blits by forcing TC-compatible HTILE on VI,
1350 * which requires 2D tiling.
1352 if (sscreen
->info
.chip_class
== VI
&& tc_compatible_htile
)
1353 return RADEON_SURF_MODE_2D
;
1355 /* Handle common candidates for the linear mode.
1356 * Compressed textures and DB surfaces must always be tiled.
1358 if (!force_tiling
&&
1359 !is_depth_stencil
&&
1360 !util_format_is_compressed(templ
->format
)) {
1361 if (sscreen
->debug_flags
& DBG(NO_TILING
))
1362 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1364 /* Tiling doesn't work with the 422 (SUBSAMPLED) formats. */
1365 if (desc
->layout
== UTIL_FORMAT_LAYOUT_SUBSAMPLED
)
1366 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1368 /* Cursors are linear on SI.
1369 * (XXX double-check, maybe also use RADEON_SURF_SCANOUT) */
1370 if (templ
->bind
& PIPE_BIND_CURSOR
)
1371 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1373 if (templ
->bind
& PIPE_BIND_LINEAR
)
1374 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1376 /* Textures with a very small height are recommended to be linear. */
1377 if (templ
->target
== PIPE_TEXTURE_1D
||
1378 templ
->target
== PIPE_TEXTURE_1D_ARRAY
||
1379 /* Only very thin and long 2D textures should benefit from
1380 * linear_aligned. */
1381 (templ
->width0
> 8 && templ
->height0
<= 2))
1382 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1384 /* Textures likely to be mapped often. */
1385 if (templ
->usage
== PIPE_USAGE_STAGING
||
1386 templ
->usage
== PIPE_USAGE_STREAM
)
1387 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1390 /* Make small textures 1D tiled. */
1391 if (templ
->width0
<= 16 || templ
->height0
<= 16 ||
1392 (sscreen
->debug_flags
& DBG(NO_2D_TILING
)))
1393 return RADEON_SURF_MODE_1D
;
1395 /* The allocator will switch to 1D if needed. */
1396 return RADEON_SURF_MODE_2D
;
1399 struct pipe_resource
*si_texture_create(struct pipe_screen
*screen
,
1400 const struct pipe_resource
*templ
)
1402 struct si_screen
*sscreen
= (struct si_screen
*)screen
;
1403 bool is_zs
= util_format_is_depth_or_stencil(templ
->format
);
1405 if (templ
->nr_samples
>= 2) {
1406 /* This is hackish (overwriting the const pipe_resource template),
1407 * but should be harmless and state trackers can also see
1408 * the overriden number of samples in the created pipe_resource.
1410 if (is_zs
&& sscreen
->eqaa_force_z_samples
) {
1411 ((struct pipe_resource
*)templ
)->nr_samples
=
1412 ((struct pipe_resource
*)templ
)->nr_storage_samples
=
1413 sscreen
->eqaa_force_z_samples
;
1414 } else if (!is_zs
&& sscreen
->eqaa_force_color_samples
) {
1415 ((struct pipe_resource
*)templ
)->nr_samples
=
1416 sscreen
->eqaa_force_coverage_samples
;
1417 ((struct pipe_resource
*)templ
)->nr_storage_samples
=
1418 sscreen
->eqaa_force_color_samples
;
1422 struct radeon_surf surface
= {0};
1423 bool is_flushed_depth
= templ
->flags
& SI_RESOURCE_FLAG_FLUSHED_DEPTH
;
1424 bool tc_compatible_htile
=
1425 sscreen
->info
.chip_class
>= VI
&&
1426 /* There are issues with TC-compatible HTILE on Tonga (and
1427 * Iceland is the same design), and documented bug workarounds
1428 * don't help. For example, this fails:
1429 * piglit/bin/tex-miplevel-selection 'texture()' 2DShadow -auto
1431 sscreen
->info
.family
!= CHIP_TONGA
&&
1432 sscreen
->info
.family
!= CHIP_ICELAND
&&
1433 (templ
->flags
& PIPE_RESOURCE_FLAG_TEXTURING_MORE_LIKELY
) &&
1434 !(sscreen
->debug_flags
& DBG(NO_HYPERZ
)) &&
1435 !is_flushed_depth
&&
1436 templ
->nr_samples
<= 1 && /* TC-compat HTILE is less efficient with MSAA */
1440 r
= si_init_surface(sscreen
, &surface
, templ
,
1441 si_choose_tiling(sscreen
, templ
, tc_compatible_htile
),
1442 0, 0, false, false, is_flushed_depth
,
1443 tc_compatible_htile
);
1448 return (struct pipe_resource
*)
1449 si_texture_create_object(screen
, templ
, NULL
, &surface
);
1452 static struct pipe_resource
*si_texture_from_winsys_buffer(struct si_screen
*sscreen
,
1453 const struct pipe_resource
*templ
,
1454 struct pb_buffer
*buf
,
1460 enum radeon_surf_mode array_mode
;
1461 struct radeon_surf surface
= {};
1462 struct radeon_bo_metadata metadata
= {};
1463 struct si_texture
*tex
;
1468 sscreen
->ws
->buffer_get_metadata(buf
, &metadata
);
1469 si_get_display_metadata(sscreen
, &surface
, &metadata
,
1470 &array_mode
, &is_scanout
);
1473 * The bo metadata is unset for un-dedicated images. So we fall
1474 * back to linear. See answer to question 5 of the
1475 * VK_KHX_external_memory spec for some details.
1477 * It is possible that this case isn't going to work if the
1478 * surface pitch isn't correctly aligned by default.
1480 * In order to support it correctly we require multi-image
1481 * metadata to be syncrhonized between radv and radeonsi. The
1482 * semantics of associating multiple image metadata to a memory
1483 * object on the vulkan export side are not concretely defined
1486 * All the use cases we are aware of at the moment for memory
1487 * objects use dedicated allocations. So lets keep the initial
1488 * implementation simple.
1490 * A possible alternative is to attempt to reconstruct the
1491 * tiling information when the TexParameter TEXTURE_TILING_EXT
1494 array_mode
= RADEON_SURF_MODE_LINEAR_ALIGNED
;
1498 r
= si_init_surface(sscreen
, &surface
, templ
,
1499 array_mode
, stride
, offset
, true, is_scanout
,
1504 tex
= si_texture_create_object(&sscreen
->b
, templ
, buf
, &surface
);
1508 tex
->buffer
.b
.is_shared
= true;
1509 tex
->buffer
.external_usage
= usage
;
1511 si_get_opaque_metadata(sscreen
, tex
, &metadata
);
1513 assert(tex
->surface
.tile_swizzle
== 0);
1514 return &tex
->buffer
.b
.b
;
1517 static struct pipe_resource
*si_texture_from_handle(struct pipe_screen
*screen
,
1518 const struct pipe_resource
*templ
,
1519 struct winsys_handle
*whandle
,
1522 struct si_screen
*sscreen
= (struct si_screen
*)screen
;
1523 struct pb_buffer
*buf
= NULL
;
1524 unsigned stride
= 0, offset
= 0;
1526 /* Support only 2D textures without mipmaps */
1527 if ((templ
->target
!= PIPE_TEXTURE_2D
&& templ
->target
!= PIPE_TEXTURE_RECT
) ||
1528 templ
->depth0
!= 1 || templ
->last_level
!= 0)
1531 buf
= sscreen
->ws
->buffer_from_handle(sscreen
->ws
, whandle
,
1532 sscreen
->info
.max_alignment
,
1537 return si_texture_from_winsys_buffer(sscreen
, templ
, buf
, stride
,
1538 offset
, usage
, true);
1541 bool si_init_flushed_depth_texture(struct pipe_context
*ctx
,
1542 struct pipe_resource
*texture
,
1543 struct si_texture
**staging
)
1545 struct si_texture
*tex
= (struct si_texture
*)texture
;
1546 struct pipe_resource resource
;
1547 struct si_texture
**flushed_depth_texture
= staging
?
1548 staging
: &tex
->flushed_depth_texture
;
1549 enum pipe_format pipe_format
= texture
->format
;
1552 if (tex
->flushed_depth_texture
)
1553 return true; /* it's ready */
1555 if (!tex
->can_sample_z
&& tex
->can_sample_s
) {
1556 switch (pipe_format
) {
1557 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
:
1558 /* Save memory by not allocating the S plane. */
1559 pipe_format
= PIPE_FORMAT_Z32_FLOAT
;
1561 case PIPE_FORMAT_Z24_UNORM_S8_UINT
:
1562 case PIPE_FORMAT_S8_UINT_Z24_UNORM
:
1563 /* Save memory bandwidth by not copying the
1564 * stencil part during flush.
1566 * This potentially increases memory bandwidth
1567 * if an application uses both Z and S texturing
1568 * simultaneously (a flushed Z24S8 texture
1569 * would be stored compactly), but how often
1570 * does that really happen?
1572 pipe_format
= PIPE_FORMAT_Z24X8_UNORM
;
1576 } else if (!tex
->can_sample_s
&& tex
->can_sample_z
) {
1577 assert(util_format_has_stencil(util_format_description(pipe_format
)));
1579 /* DB->CB copies to an 8bpp surface don't work. */
1580 pipe_format
= PIPE_FORMAT_X24S8_UINT
;
1584 memset(&resource
, 0, sizeof(resource
));
1585 resource
.target
= texture
->target
;
1586 resource
.format
= pipe_format
;
1587 resource
.width0
= texture
->width0
;
1588 resource
.height0
= texture
->height0
;
1589 resource
.depth0
= texture
->depth0
;
1590 resource
.array_size
= texture
->array_size
;
1591 resource
.last_level
= texture
->last_level
;
1592 resource
.nr_samples
= texture
->nr_samples
;
1593 resource
.usage
= staging
? PIPE_USAGE_STAGING
: PIPE_USAGE_DEFAULT
;
1594 resource
.bind
= texture
->bind
& ~PIPE_BIND_DEPTH_STENCIL
;
1595 resource
.flags
= texture
->flags
| SI_RESOURCE_FLAG_FLUSHED_DEPTH
;
1598 resource
.flags
|= SI_RESOURCE_FLAG_TRANSFER
;
1600 *flushed_depth_texture
= (struct si_texture
*)ctx
->screen
->resource_create(ctx
->screen
, &resource
);
1601 if (*flushed_depth_texture
== NULL
) {
1602 PRINT_ERR("failed to create temporary texture to hold flushed depth\n");
1609 * Initialize the pipe_resource descriptor to be of the same size as the box,
1610 * which is supposed to hold a subregion of the texture "orig" at the given
1613 static void si_init_temp_resource_from_box(struct pipe_resource
*res
,
1614 struct pipe_resource
*orig
,
1615 const struct pipe_box
*box
,
1616 unsigned level
, unsigned flags
)
1618 memset(res
, 0, sizeof(*res
));
1619 res
->format
= orig
->format
;
1620 res
->width0
= box
->width
;
1621 res
->height0
= box
->height
;
1623 res
->array_size
= 1;
1624 res
->usage
= flags
& SI_RESOURCE_FLAG_TRANSFER
? PIPE_USAGE_STAGING
: PIPE_USAGE_DEFAULT
;
1627 /* We must set the correct texture target and dimensions for a 3D box. */
1628 if (box
->depth
> 1 && util_max_layer(orig
, level
) > 0) {
1629 res
->target
= PIPE_TEXTURE_2D_ARRAY
;
1630 res
->array_size
= box
->depth
;
1632 res
->target
= PIPE_TEXTURE_2D
;
1636 static bool si_can_invalidate_texture(struct si_screen
*sscreen
,
1637 struct si_texture
*tex
,
1638 unsigned transfer_usage
,
1639 const struct pipe_box
*box
)
1641 return !tex
->buffer
.b
.is_shared
&&
1642 !(transfer_usage
& PIPE_TRANSFER_READ
) &&
1643 tex
->buffer
.b
.b
.last_level
== 0 &&
1644 util_texrange_covers_whole_level(&tex
->buffer
.b
.b
, 0,
1645 box
->x
, box
->y
, box
->z
,
1646 box
->width
, box
->height
,
1650 static void si_texture_invalidate_storage(struct si_context
*sctx
,
1651 struct si_texture
*tex
)
1653 struct si_screen
*sscreen
= sctx
->screen
;
1655 /* There is no point in discarding depth and tiled buffers. */
1656 assert(!tex
->is_depth
);
1657 assert(tex
->surface
.is_linear
);
1659 /* Reallocate the buffer in the same pipe_resource. */
1660 si_alloc_resource(sscreen
, &tex
->buffer
);
1662 /* Initialize the CMASK base address (needed even without CMASK). */
1663 tex
->cmask_base_address_reg
=
1664 (tex
->buffer
.gpu_address
+ tex
->cmask_offset
) >> 8;
1666 p_atomic_inc(&sscreen
->dirty_tex_counter
);
1668 sctx
->num_alloc_tex_transfer_bytes
+= tex
->size
;
1671 static void *si_texture_transfer_map(struct pipe_context
*ctx
,
1672 struct pipe_resource
*texture
,
1675 const struct pipe_box
*box
,
1676 struct pipe_transfer
**ptransfer
)
1678 struct si_context
*sctx
= (struct si_context
*)ctx
;
1679 struct si_texture
*tex
= (struct si_texture
*)texture
;
1680 struct si_transfer
*trans
;
1681 struct si_resource
*buf
;
1682 unsigned offset
= 0;
1684 bool use_staging_texture
= false;
1686 assert(!(texture
->flags
& SI_RESOURCE_FLAG_TRANSFER
));
1687 assert(box
->width
&& box
->height
&& box
->depth
);
1689 /* Depth textures use staging unconditionally. */
1690 if (!tex
->is_depth
) {
1691 /* Degrade the tile mode if we get too many transfers on APUs.
1692 * On dGPUs, the staging texture is always faster.
1693 * Only count uploads that are at least 4x4 pixels large.
1695 if (!sctx
->screen
->info
.has_dedicated_vram
&&
1697 box
->width
>= 4 && box
->height
>= 4 &&
1698 p_atomic_inc_return(&tex
->num_level0_transfers
) == 10) {
1699 bool can_invalidate
=
1700 si_can_invalidate_texture(sctx
->screen
, tex
,
1703 si_reallocate_texture_inplace(sctx
, tex
,
1708 /* Tiled textures need to be converted into a linear texture for CPU
1709 * access. The staging texture is always linear and is placed in GART.
1711 * Reading from VRAM or GTT WC is slow, always use the staging
1712 * texture in this case.
1714 * Use the staging texture for uploads if the underlying BO
1717 if (!tex
->surface
.is_linear
)
1718 use_staging_texture
= true;
1719 else if (usage
& PIPE_TRANSFER_READ
)
1720 use_staging_texture
=
1721 tex
->buffer
.domains
& RADEON_DOMAIN_VRAM
||
1722 tex
->buffer
.flags
& RADEON_FLAG_GTT_WC
;
1723 /* Write & linear only: */
1724 else if (si_rings_is_buffer_referenced(sctx
, tex
->buffer
.buf
,
1725 RADEON_USAGE_READWRITE
) ||
1726 !sctx
->ws
->buffer_wait(tex
->buffer
.buf
, 0,
1727 RADEON_USAGE_READWRITE
)) {
1729 if (si_can_invalidate_texture(sctx
->screen
, tex
,
1731 si_texture_invalidate_storage(sctx
, tex
);
1733 use_staging_texture
= true;
1737 trans
= CALLOC_STRUCT(si_transfer
);
1740 pipe_resource_reference(&trans
->b
.b
.resource
, texture
);
1741 trans
->b
.b
.level
= level
;
1742 trans
->b
.b
.usage
= usage
;
1743 trans
->b
.b
.box
= *box
;
1745 if (tex
->is_depth
) {
1746 struct si_texture
*staging_depth
;
1748 if (tex
->buffer
.b
.b
.nr_samples
> 1) {
1749 /* MSAA depth buffers need to be converted to single sample buffers.
1751 * Mapping MSAA depth buffers can occur if ReadPixels is called
1752 * with a multisample GLX visual.
1754 * First downsample the depth buffer to a temporary texture,
1755 * then decompress the temporary one to staging.
1757 * Only the region being mapped is transfered.
1759 struct pipe_resource resource
;
1761 si_init_temp_resource_from_box(&resource
, texture
, box
, level
, 0);
1763 if (!si_init_flushed_depth_texture(ctx
, &resource
, &staging_depth
)) {
1764 PRINT_ERR("failed to create temporary texture to hold untiled copy\n");
1768 if (usage
& PIPE_TRANSFER_READ
) {
1769 struct pipe_resource
*temp
= ctx
->screen
->resource_create(ctx
->screen
, &resource
);
1771 PRINT_ERR("failed to create a temporary depth texture\n");
1775 si_copy_region_with_blit(ctx
, temp
, 0, 0, 0, 0, texture
, level
, box
);
1776 si_blit_decompress_depth(ctx
, (struct si_texture
*)temp
, staging_depth
,
1777 0, 0, 0, box
->depth
, 0, 0);
1778 pipe_resource_reference(&temp
, NULL
);
1781 /* Just get the strides. */
1782 si_texture_get_offset(sctx
->screen
, staging_depth
, level
, NULL
,
1784 &trans
->b
.b
.layer_stride
);
1786 /* XXX: only readback the rectangle which is being mapped? */
1787 /* XXX: when discard is true, no need to read back from depth texture */
1788 if (!si_init_flushed_depth_texture(ctx
, texture
, &staging_depth
)) {
1789 PRINT_ERR("failed to create temporary texture to hold untiled copy\n");
1793 si_blit_decompress_depth(ctx
, tex
, staging_depth
,
1795 box
->z
, box
->z
+ box
->depth
- 1,
1798 offset
= si_texture_get_offset(sctx
->screen
, staging_depth
,
1801 &trans
->b
.b
.layer_stride
);
1804 trans
->staging
= &staging_depth
->buffer
;
1805 buf
= trans
->staging
;
1806 } else if (use_staging_texture
) {
1807 struct pipe_resource resource
;
1808 struct si_texture
*staging
;
1810 si_init_temp_resource_from_box(&resource
, texture
, box
, level
,
1811 SI_RESOURCE_FLAG_TRANSFER
);
1812 resource
.usage
= (usage
& PIPE_TRANSFER_READ
) ?
1813 PIPE_USAGE_STAGING
: PIPE_USAGE_STREAM
;
1815 /* Create the temporary texture. */
1816 staging
= (struct si_texture
*)ctx
->screen
->resource_create(ctx
->screen
, &resource
);
1818 PRINT_ERR("failed to create temporary texture to hold untiled copy\n");
1821 trans
->staging
= &staging
->buffer
;
1823 /* Just get the strides. */
1824 si_texture_get_offset(sctx
->screen
, staging
, 0, NULL
,
1826 &trans
->b
.b
.layer_stride
);
1828 if (usage
& PIPE_TRANSFER_READ
)
1829 si_copy_to_staging_texture(ctx
, trans
);
1831 usage
|= PIPE_TRANSFER_UNSYNCHRONIZED
;
1833 buf
= trans
->staging
;
1835 /* the resource is mapped directly */
1836 offset
= si_texture_get_offset(sctx
->screen
, tex
, level
, box
,
1838 &trans
->b
.b
.layer_stride
);
1842 /* Always unmap texture CPU mappings on 32-bit architectures, so that
1843 * we don't run out of the CPU address space.
1845 if (sizeof(void*) == 4)
1846 usage
|= RADEON_TRANSFER_TEMPORARY
;
1848 if (!(map
= si_buffer_map_sync_with_rings(sctx
, buf
, usage
)))
1851 *ptransfer
= &trans
->b
.b
;
1852 return map
+ offset
;
1855 si_resource_reference(&trans
->staging
, NULL
);
1856 pipe_resource_reference(&trans
->b
.b
.resource
, NULL
);
1861 static void si_texture_transfer_unmap(struct pipe_context
*ctx
,
1862 struct pipe_transfer
* transfer
)
1864 struct si_context
*sctx
= (struct si_context
*)ctx
;
1865 struct si_transfer
*stransfer
= (struct si_transfer
*)transfer
;
1866 struct pipe_resource
*texture
= transfer
->resource
;
1867 struct si_texture
*tex
= (struct si_texture
*)texture
;
1869 /* Always unmap texture CPU mappings on 32-bit architectures, so that
1870 * we don't run out of the CPU address space.
1872 if (sizeof(void*) == 4) {
1873 struct si_resource
*buf
=
1874 stransfer
->staging
? stransfer
->staging
: &tex
->buffer
;
1876 sctx
->ws
->buffer_unmap(buf
->buf
);
1879 if ((transfer
->usage
& PIPE_TRANSFER_WRITE
) && stransfer
->staging
) {
1880 if (tex
->is_depth
&& tex
->buffer
.b
.b
.nr_samples
<= 1) {
1881 ctx
->resource_copy_region(ctx
, texture
, transfer
->level
,
1882 transfer
->box
.x
, transfer
->box
.y
, transfer
->box
.z
,
1883 &stransfer
->staging
->b
.b
, transfer
->level
,
1886 si_copy_from_staging_texture(ctx
, stransfer
);
1890 if (stransfer
->staging
) {
1891 sctx
->num_alloc_tex_transfer_bytes
+= stransfer
->staging
->buf
->size
;
1892 si_resource_reference(&stransfer
->staging
, NULL
);
1895 /* Heuristic for {upload, draw, upload, draw, ..}:
1897 * Flush the gfx IB if we've allocated too much texture storage.
1899 * The idea is that we don't want to build IBs that use too much
1900 * memory and put pressure on the kernel memory manager and we also
1901 * want to make temporary and invalidated buffers go idle ASAP to
1902 * decrease the total memory usage or make them reusable. The memory
1903 * usage will be slightly higher than given here because of the buffer
1904 * cache in the winsys.
1906 * The result is that the kernel memory manager is never a bottleneck.
1908 if (sctx
->num_alloc_tex_transfer_bytes
> sctx
->screen
->info
.gart_size
/ 4) {
1909 si_flush_gfx_cs(sctx
, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW
, NULL
);
1910 sctx
->num_alloc_tex_transfer_bytes
= 0;
1913 pipe_resource_reference(&transfer
->resource
, NULL
);
1917 static const struct u_resource_vtbl si_texture_vtbl
=
1919 NULL
, /* get_handle */
1920 si_texture_destroy
, /* resource_destroy */
1921 si_texture_transfer_map
, /* transfer_map */
1922 u_default_transfer_flush_region
, /* transfer_flush_region */
1923 si_texture_transfer_unmap
, /* transfer_unmap */
1926 /* Return if it's allowed to reinterpret one format as another with DCC enabled.
1928 bool vi_dcc_formats_compatible(enum pipe_format format1
,
1929 enum pipe_format format2
)
1931 const struct util_format_description
*desc1
, *desc2
;
1933 /* No format change - exit early. */
1934 if (format1
== format2
)
1937 format1
= si_simplify_cb_format(format1
);
1938 format2
= si_simplify_cb_format(format2
);
1940 /* Check again after format adjustments. */
1941 if (format1
== format2
)
1944 desc1
= util_format_description(format1
);
1945 desc2
= util_format_description(format2
);
1947 if (desc1
->layout
!= UTIL_FORMAT_LAYOUT_PLAIN
||
1948 desc2
->layout
!= UTIL_FORMAT_LAYOUT_PLAIN
)
1951 /* Float and non-float are totally incompatible. */
1952 if ((desc1
->channel
[0].type
== UTIL_FORMAT_TYPE_FLOAT
) !=
1953 (desc2
->channel
[0].type
== UTIL_FORMAT_TYPE_FLOAT
))
1956 /* Channel sizes must match across DCC formats.
1957 * Comparing just the first 2 channels should be enough.
1959 if (desc1
->channel
[0].size
!= desc2
->channel
[0].size
||
1960 (desc1
->nr_channels
>= 2 &&
1961 desc1
->channel
[1].size
!= desc2
->channel
[1].size
))
1964 /* Everything below is not needed if the driver never uses the DCC
1965 * clear code with the value of 1.
1968 /* If the clear values are all 1 or all 0, this constraint can be
1970 if (vi_alpha_is_on_msb(format1
) != vi_alpha_is_on_msb(format2
))
1973 /* Channel types must match if the clear value of 1 is used.
1974 * The type categories are only float, signed, unsigned.
1975 * NORM and INT are always compatible.
1977 if (desc1
->channel
[0].type
!= desc2
->channel
[0].type
||
1978 (desc1
->nr_channels
>= 2 &&
1979 desc1
->channel
[1].type
!= desc2
->channel
[1].type
))
1985 bool vi_dcc_formats_are_incompatible(struct pipe_resource
*tex
,
1987 enum pipe_format view_format
)
1989 struct si_texture
*stex
= (struct si_texture
*)tex
;
1991 return vi_dcc_enabled(stex
, level
) &&
1992 !vi_dcc_formats_compatible(tex
->format
, view_format
);
1995 /* This can't be merged with the above function, because
1996 * vi_dcc_formats_compatible should be called only when DCC is enabled. */
1997 void vi_disable_dcc_if_incompatible_format(struct si_context
*sctx
,
1998 struct pipe_resource
*tex
,
2000 enum pipe_format view_format
)
2002 struct si_texture
*stex
= (struct si_texture
*)tex
;
2004 if (vi_dcc_formats_are_incompatible(tex
, level
, view_format
))
2005 if (!si_texture_disable_dcc(sctx
, stex
))
2006 si_decompress_dcc(sctx
, stex
);
2009 struct pipe_surface
*si_create_surface_custom(struct pipe_context
*pipe
,
2010 struct pipe_resource
*texture
,
2011 const struct pipe_surface
*templ
,
2012 unsigned width0
, unsigned height0
,
2013 unsigned width
, unsigned height
)
2015 struct si_surface
*surface
= CALLOC_STRUCT(si_surface
);
2020 assert(templ
->u
.tex
.first_layer
<= util_max_layer(texture
, templ
->u
.tex
.level
));
2021 assert(templ
->u
.tex
.last_layer
<= util_max_layer(texture
, templ
->u
.tex
.level
));
2023 pipe_reference_init(&surface
->base
.reference
, 1);
2024 pipe_resource_reference(&surface
->base
.texture
, texture
);
2025 surface
->base
.context
= pipe
;
2026 surface
->base
.format
= templ
->format
;
2027 surface
->base
.width
= width
;
2028 surface
->base
.height
= height
;
2029 surface
->base
.u
= templ
->u
;
2031 surface
->width0
= width0
;
2032 surface
->height0
= height0
;
2034 surface
->dcc_incompatible
=
2035 texture
->target
!= PIPE_BUFFER
&&
2036 vi_dcc_formats_are_incompatible(texture
, templ
->u
.tex
.level
,
2038 return &surface
->base
;
2041 static struct pipe_surface
*si_create_surface(struct pipe_context
*pipe
,
2042 struct pipe_resource
*tex
,
2043 const struct pipe_surface
*templ
)
2045 unsigned level
= templ
->u
.tex
.level
;
2046 unsigned width
= u_minify(tex
->width0
, level
);
2047 unsigned height
= u_minify(tex
->height0
, level
);
2048 unsigned width0
= tex
->width0
;
2049 unsigned height0
= tex
->height0
;
2051 if (tex
->target
!= PIPE_BUFFER
&& templ
->format
!= tex
->format
) {
2052 const struct util_format_description
*tex_desc
2053 = util_format_description(tex
->format
);
2054 const struct util_format_description
*templ_desc
2055 = util_format_description(templ
->format
);
2057 assert(tex_desc
->block
.bits
== templ_desc
->block
.bits
);
2059 /* Adjust size of surface if and only if the block width or
2060 * height is changed. */
2061 if (tex_desc
->block
.width
!= templ_desc
->block
.width
||
2062 tex_desc
->block
.height
!= templ_desc
->block
.height
) {
2063 unsigned nblks_x
= util_format_get_nblocksx(tex
->format
, width
);
2064 unsigned nblks_y
= util_format_get_nblocksy(tex
->format
, height
);
2066 width
= nblks_x
* templ_desc
->block
.width
;
2067 height
= nblks_y
* templ_desc
->block
.height
;
2069 width0
= util_format_get_nblocksx(tex
->format
, width0
);
2070 height0
= util_format_get_nblocksy(tex
->format
, height0
);
2074 return si_create_surface_custom(pipe
, tex
, templ
,
2079 static void si_surface_destroy(struct pipe_context
*pipe
,
2080 struct pipe_surface
*surface
)
2082 pipe_resource_reference(&surface
->texture
, NULL
);
2086 unsigned si_translate_colorswap(enum pipe_format format
, bool do_endian_swap
)
2088 const struct util_format_description
*desc
= util_format_description(format
);
2090 #define HAS_SWIZZLE(chan,swz) (desc->swizzle[chan] == PIPE_SWIZZLE_##swz)
2092 if (format
== PIPE_FORMAT_R11G11B10_FLOAT
) /* isn't plain */
2093 return V_028C70_SWAP_STD
;
2095 if (desc
->layout
!= UTIL_FORMAT_LAYOUT_PLAIN
)
2098 switch (desc
->nr_channels
) {
2100 if (HAS_SWIZZLE(0,X
))
2101 return V_028C70_SWAP_STD
; /* X___ */
2102 else if (HAS_SWIZZLE(3,X
))
2103 return V_028C70_SWAP_ALT_REV
; /* ___X */
2106 if ((HAS_SWIZZLE(0,X
) && HAS_SWIZZLE(1,Y
)) ||
2107 (HAS_SWIZZLE(0,X
) && HAS_SWIZZLE(1,NONE
)) ||
2108 (HAS_SWIZZLE(0,NONE
) && HAS_SWIZZLE(1,Y
)))
2109 return V_028C70_SWAP_STD
; /* XY__ */
2110 else if ((HAS_SWIZZLE(0,Y
) && HAS_SWIZZLE(1,X
)) ||
2111 (HAS_SWIZZLE(0,Y
) && HAS_SWIZZLE(1,NONE
)) ||
2112 (HAS_SWIZZLE(0,NONE
) && HAS_SWIZZLE(1,X
)))
2114 return (do_endian_swap
? V_028C70_SWAP_STD
: V_028C70_SWAP_STD_REV
);
2115 else if (HAS_SWIZZLE(0,X
) && HAS_SWIZZLE(3,Y
))
2116 return V_028C70_SWAP_ALT
; /* X__Y */
2117 else if (HAS_SWIZZLE(0,Y
) && HAS_SWIZZLE(3,X
))
2118 return V_028C70_SWAP_ALT_REV
; /* Y__X */
2121 if (HAS_SWIZZLE(0,X
))
2122 return (do_endian_swap
? V_028C70_SWAP_STD_REV
: V_028C70_SWAP_STD
);
2123 else if (HAS_SWIZZLE(0,Z
))
2124 return V_028C70_SWAP_STD_REV
; /* ZYX */
2127 /* check the middle channels, the 1st and 4th channel can be NONE */
2128 if (HAS_SWIZZLE(1,Y
) && HAS_SWIZZLE(2,Z
)) {
2129 return V_028C70_SWAP_STD
; /* XYZW */
2130 } else if (HAS_SWIZZLE(1,Z
) && HAS_SWIZZLE(2,Y
)) {
2131 return V_028C70_SWAP_STD_REV
; /* WZYX */
2132 } else if (HAS_SWIZZLE(1,Y
) && HAS_SWIZZLE(2,X
)) {
2133 return V_028C70_SWAP_ALT
; /* ZYXW */
2134 } else if (HAS_SWIZZLE(1,Z
) && HAS_SWIZZLE(2,W
)) {
2137 return V_028C70_SWAP_ALT_REV
;
2139 return (do_endian_swap
? V_028C70_SWAP_ALT
: V_028C70_SWAP_ALT_REV
);
2146 /* PIPELINE_STAT-BASED DCC ENABLEMENT FOR DISPLAYABLE SURFACES */
2148 static void vi_dcc_clean_up_context_slot(struct si_context
*sctx
,
2153 if (sctx
->dcc_stats
[slot
].query_active
)
2154 vi_separate_dcc_stop_query(sctx
,
2155 sctx
->dcc_stats
[slot
].tex
);
2157 for (i
= 0; i
< ARRAY_SIZE(sctx
->dcc_stats
[slot
].ps_stats
); i
++)
2158 if (sctx
->dcc_stats
[slot
].ps_stats
[i
]) {
2159 sctx
->b
.destroy_query(&sctx
->b
,
2160 sctx
->dcc_stats
[slot
].ps_stats
[i
]);
2161 sctx
->dcc_stats
[slot
].ps_stats
[i
] = NULL
;
2164 si_texture_reference(&sctx
->dcc_stats
[slot
].tex
, NULL
);
2168 * Return the per-context slot where DCC statistics queries for the texture live.
2170 static unsigned vi_get_context_dcc_stats_index(struct si_context
*sctx
,
2171 struct si_texture
*tex
)
2173 int i
, empty_slot
= -1;
2175 /* Remove zombie textures (textures kept alive by this array only). */
2176 for (i
= 0; i
< ARRAY_SIZE(sctx
->dcc_stats
); i
++)
2177 if (sctx
->dcc_stats
[i
].tex
&&
2178 sctx
->dcc_stats
[i
].tex
->buffer
.b
.b
.reference
.count
== 1)
2179 vi_dcc_clean_up_context_slot(sctx
, i
);
2181 /* Find the texture. */
2182 for (i
= 0; i
< ARRAY_SIZE(sctx
->dcc_stats
); i
++) {
2183 /* Return if found. */
2184 if (sctx
->dcc_stats
[i
].tex
== tex
) {
2185 sctx
->dcc_stats
[i
].last_use_timestamp
= os_time_get();
2189 /* Record the first seen empty slot. */
2190 if (empty_slot
== -1 && !sctx
->dcc_stats
[i
].tex
)
2194 /* Not found. Remove the oldest member to make space in the array. */
2195 if (empty_slot
== -1) {
2196 int oldest_slot
= 0;
2198 /* Find the oldest slot. */
2199 for (i
= 1; i
< ARRAY_SIZE(sctx
->dcc_stats
); i
++)
2200 if (sctx
->dcc_stats
[oldest_slot
].last_use_timestamp
>
2201 sctx
->dcc_stats
[i
].last_use_timestamp
)
2204 /* Clean up the oldest slot. */
2205 vi_dcc_clean_up_context_slot(sctx
, oldest_slot
);
2206 empty_slot
= oldest_slot
;
2209 /* Add the texture to the new slot. */
2210 si_texture_reference(&sctx
->dcc_stats
[empty_slot
].tex
, tex
);
2211 sctx
->dcc_stats
[empty_slot
].last_use_timestamp
= os_time_get();
2215 static struct pipe_query
*
2216 vi_create_resuming_pipestats_query(struct si_context
*sctx
)
2218 struct si_query_hw
*query
= (struct si_query_hw
*)
2219 sctx
->b
.create_query(&sctx
->b
, PIPE_QUERY_PIPELINE_STATISTICS
, 0);
2221 query
->flags
|= SI_QUERY_HW_FLAG_BEGIN_RESUMES
;
2222 return (struct pipe_query
*)query
;
2226 * Called when binding a color buffer.
2228 void vi_separate_dcc_start_query(struct si_context
*sctx
,
2229 struct si_texture
*tex
)
2231 unsigned i
= vi_get_context_dcc_stats_index(sctx
, tex
);
2233 assert(!sctx
->dcc_stats
[i
].query_active
);
2235 if (!sctx
->dcc_stats
[i
].ps_stats
[0])
2236 sctx
->dcc_stats
[i
].ps_stats
[0] = vi_create_resuming_pipestats_query(sctx
);
2238 /* begin or resume the query */
2239 sctx
->b
.begin_query(&sctx
->b
, sctx
->dcc_stats
[i
].ps_stats
[0]);
2240 sctx
->dcc_stats
[i
].query_active
= true;
2244 * Called when unbinding a color buffer.
2246 void vi_separate_dcc_stop_query(struct si_context
*sctx
,
2247 struct si_texture
*tex
)
2249 unsigned i
= vi_get_context_dcc_stats_index(sctx
, tex
);
2251 assert(sctx
->dcc_stats
[i
].query_active
);
2252 assert(sctx
->dcc_stats
[i
].ps_stats
[0]);
2254 /* pause or end the query */
2255 sctx
->b
.end_query(&sctx
->b
, sctx
->dcc_stats
[i
].ps_stats
[0]);
2256 sctx
->dcc_stats
[i
].query_active
= false;
2259 static bool vi_should_enable_separate_dcc(struct si_texture
*tex
)
2261 /* The minimum number of fullscreen draws per frame that is required
2263 return tex
->ps_draw_ratio
+ tex
->num_slow_clears
>= 5;
2266 /* Called by fast clear. */
2267 void vi_separate_dcc_try_enable(struct si_context
*sctx
,
2268 struct si_texture
*tex
)
2270 /* The intent is to use this with shared displayable back buffers,
2271 * but it's not strictly limited only to them.
2273 if (!tex
->buffer
.b
.is_shared
||
2274 !(tex
->buffer
.external_usage
& PIPE_HANDLE_USAGE_EXPLICIT_FLUSH
) ||
2275 tex
->buffer
.b
.b
.target
!= PIPE_TEXTURE_2D
||
2276 tex
->buffer
.b
.b
.last_level
> 0 ||
2277 !tex
->surface
.dcc_size
||
2278 sctx
->screen
->debug_flags
& DBG(NO_DCC
) ||
2279 sctx
->screen
->debug_flags
& DBG(NO_DCC_FB
))
2282 assert(sctx
->chip_class
>= VI
);
2284 if (tex
->dcc_offset
)
2285 return; /* already enabled */
2287 /* Enable the DCC stat gathering. */
2288 if (!tex
->dcc_gather_statistics
) {
2289 tex
->dcc_gather_statistics
= true;
2290 vi_separate_dcc_start_query(sctx
, tex
);
2293 if (!vi_should_enable_separate_dcc(tex
))
2294 return; /* stats show that DCC decompression is too expensive */
2296 assert(tex
->surface
.num_dcc_levels
);
2297 assert(!tex
->dcc_separate_buffer
);
2299 si_texture_discard_cmask(sctx
->screen
, tex
);
2301 /* Get a DCC buffer. */
2302 if (tex
->last_dcc_separate_buffer
) {
2303 assert(tex
->dcc_gather_statistics
);
2304 assert(!tex
->dcc_separate_buffer
);
2305 tex
->dcc_separate_buffer
= tex
->last_dcc_separate_buffer
;
2306 tex
->last_dcc_separate_buffer
= NULL
;
2308 tex
->dcc_separate_buffer
=
2309 si_aligned_buffer_create(sctx
->b
.screen
,
2310 SI_RESOURCE_FLAG_UNMAPPABLE
,
2312 tex
->surface
.dcc_size
,
2313 tex
->surface
.dcc_alignment
);
2314 if (!tex
->dcc_separate_buffer
)
2318 /* dcc_offset is the absolute GPUVM address. */
2319 tex
->dcc_offset
= tex
->dcc_separate_buffer
->gpu_address
;
2321 /* no need to flag anything since this is called by fast clear that
2322 * flags framebuffer state
2327 * Called by pipe_context::flush_resource, the place where DCC decompression
2330 void vi_separate_dcc_process_and_reset_stats(struct pipe_context
*ctx
,
2331 struct si_texture
*tex
)
2333 struct si_context
*sctx
= (struct si_context
*)ctx
;
2334 struct pipe_query
*tmp
;
2335 unsigned i
= vi_get_context_dcc_stats_index(sctx
, tex
);
2336 bool query_active
= sctx
->dcc_stats
[i
].query_active
;
2337 bool disable
= false;
2339 if (sctx
->dcc_stats
[i
].ps_stats
[2]) {
2340 union pipe_query_result result
;
2342 /* Read the results. */
2343 struct pipe_query
*query
= sctx
->dcc_stats
[i
].ps_stats
[2];
2344 ctx
->get_query_result(ctx
, query
,
2346 si_query_buffer_reset(sctx
, &((struct si_query_hw
*)query
)->buffer
);
2348 /* Compute the approximate number of fullscreen draws. */
2349 tex
->ps_draw_ratio
=
2350 result
.pipeline_statistics
.ps_invocations
/
2351 (tex
->buffer
.b
.b
.width0
* tex
->buffer
.b
.b
.height0
);
2352 sctx
->last_tex_ps_draw_ratio
= tex
->ps_draw_ratio
;
2354 disable
= tex
->dcc_separate_buffer
&&
2355 !vi_should_enable_separate_dcc(tex
);
2358 tex
->num_slow_clears
= 0;
2360 /* stop the statistics query for ps_stats[0] */
2362 vi_separate_dcc_stop_query(sctx
, tex
);
2364 /* Move the queries in the queue by one. */
2365 tmp
= sctx
->dcc_stats
[i
].ps_stats
[2];
2366 sctx
->dcc_stats
[i
].ps_stats
[2] = sctx
->dcc_stats
[i
].ps_stats
[1];
2367 sctx
->dcc_stats
[i
].ps_stats
[1] = sctx
->dcc_stats
[i
].ps_stats
[0];
2368 sctx
->dcc_stats
[i
].ps_stats
[0] = tmp
;
2370 /* create and start a new query as ps_stats[0] */
2372 vi_separate_dcc_start_query(sctx
, tex
);
2375 assert(!tex
->last_dcc_separate_buffer
);
2376 tex
->last_dcc_separate_buffer
= tex
->dcc_separate_buffer
;
2377 tex
->dcc_separate_buffer
= NULL
;
2378 tex
->dcc_offset
= 0;
2379 /* no need to flag anything since this is called after
2380 * decompression that re-sets framebuffer state
2385 static struct pipe_memory_object
*
2386 si_memobj_from_handle(struct pipe_screen
*screen
,
2387 struct winsys_handle
*whandle
,
2390 struct si_screen
*sscreen
= (struct si_screen
*)screen
;
2391 struct si_memory_object
*memobj
= CALLOC_STRUCT(si_memory_object
);
2392 struct pb_buffer
*buf
= NULL
;
2393 uint32_t stride
, offset
;
2398 buf
= sscreen
->ws
->buffer_from_handle(sscreen
->ws
, whandle
,
2399 sscreen
->info
.max_alignment
,
2406 memobj
->b
.dedicated
= dedicated
;
2408 memobj
->stride
= stride
;
2410 return (struct pipe_memory_object
*)memobj
;
2415 si_memobj_destroy(struct pipe_screen
*screen
,
2416 struct pipe_memory_object
*_memobj
)
2418 struct si_memory_object
*memobj
= (struct si_memory_object
*)_memobj
;
2420 pb_reference(&memobj
->buf
, NULL
);
2424 static struct pipe_resource
*
2425 si_texture_from_memobj(struct pipe_screen
*screen
,
2426 const struct pipe_resource
*templ
,
2427 struct pipe_memory_object
*_memobj
,
2430 struct si_screen
*sscreen
= (struct si_screen
*)screen
;
2431 struct si_memory_object
*memobj
= (struct si_memory_object
*)_memobj
;
2432 struct pipe_resource
*tex
=
2433 si_texture_from_winsys_buffer(sscreen
, templ
, memobj
->buf
,
2434 memobj
->stride
, offset
,
2435 PIPE_HANDLE_USAGE_FRAMEBUFFER_WRITE
|
2436 PIPE_HANDLE_USAGE_SHADER_WRITE
,
2437 memobj
->b
.dedicated
);
2441 /* si_texture_from_winsys_buffer doesn't increment refcount of
2442 * memobj->buf, so increment it here.
2444 struct pb_buffer
*buf
= NULL
;
2445 pb_reference(&buf
, memobj
->buf
);
2449 static bool si_check_resource_capability(struct pipe_screen
*screen
,
2450 struct pipe_resource
*resource
,
2453 struct si_texture
*tex
= (struct si_texture
*)resource
;
2455 /* Buffers only support the linear flag. */
2456 if (resource
->target
== PIPE_BUFFER
)
2457 return (bind
& ~PIPE_BIND_LINEAR
) == 0;
2459 if (bind
& PIPE_BIND_LINEAR
&& !tex
->surface
.is_linear
)
2462 if (bind
& PIPE_BIND_SCANOUT
&& !tex
->surface
.is_displayable
)
2465 /* TODO: PIPE_BIND_CURSOR - do we care? */
2469 void si_init_screen_texture_functions(struct si_screen
*sscreen
)
2471 sscreen
->b
.resource_from_handle
= si_texture_from_handle
;
2472 sscreen
->b
.resource_get_handle
= si_texture_get_handle
;
2473 sscreen
->b
.resource_from_memobj
= si_texture_from_memobj
;
2474 sscreen
->b
.memobj_create_from_handle
= si_memobj_from_handle
;
2475 sscreen
->b
.memobj_destroy
= si_memobj_destroy
;
2476 sscreen
->b
.check_resource_capability
= si_check_resource_capability
;
2479 void si_init_context_texture_functions(struct si_context
*sctx
)
2481 sctx
->b
.create_surface
= si_create_surface
;
2482 sctx
->b
.surface_destroy
= si_surface_destroy
;