2 * Copyright 2010 Jerome Glisse <glisse@freedesktop.org>
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
27 #include "r600_pipe_common.h"
29 #include "r600_query.h"
30 #include "util/u_format.h"
31 #include "util/u_memory.h"
32 #include "util/u_pack_color.h"
33 #include "util/u_surface.h"
34 #include "os/os_time.h"
38 static void r600_texture_discard_cmask(struct r600_common_screen
*rscreen
,
39 struct r600_texture
*rtex
);
40 static enum radeon_surf_mode
41 r600_choose_tiling(struct r600_common_screen
*rscreen
,
42 const struct pipe_resource
*templ
);
45 bool r600_prepare_for_dma_blit(struct r600_common_context
*rctx
,
46 struct r600_texture
*rdst
,
47 unsigned dst_level
, unsigned dstx
,
48 unsigned dsty
, unsigned dstz
,
49 struct r600_texture
*rsrc
,
51 const struct pipe_box
*src_box
)
56 if (rdst
->surface
.bpe
!= rsrc
->surface
.bpe
)
59 /* MSAA: Blits don't exist in the real world. */
60 if (rsrc
->resource
.b
.b
.nr_samples
> 1 ||
61 rdst
->resource
.b
.b
.nr_samples
> 1)
64 /* Depth-stencil surfaces:
65 * When dst is linear, the DB->CB copy preserves HTILE.
66 * When dst is tiled, the 3D path must be used to update HTILE.
68 if (rsrc
->is_depth
|| rdst
->is_depth
)
72 * src: Use the 3D path. DCC decompression is expensive.
73 * dst: Use the 3D path to compress the pixels with DCC.
75 if (vi_dcc_enabled(rsrc
, src_level
) ||
76 vi_dcc_enabled(rdst
, dst_level
))
80 * src: Both texture and SDMA paths need decompression. Use SDMA.
81 * dst: If overwriting the whole texture, discard CMASK and use
82 * SDMA. Otherwise, use the 3D path.
84 if (rdst
->cmask
.size
&& rdst
->dirty_level_mask
& (1 << dst_level
)) {
85 /* The CMASK clear is only enabled for the first level. */
86 assert(dst_level
== 0);
87 if (!util_texrange_covers_whole_level(&rdst
->resource
.b
.b
, dst_level
,
88 dstx
, dsty
, dstz
, src_box
->width
,
89 src_box
->height
, src_box
->depth
))
92 r600_texture_discard_cmask(rctx
->screen
, rdst
);
95 /* All requirements are met. Prepare textures for SDMA. */
96 if (rsrc
->cmask
.size
&& rsrc
->dirty_level_mask
& (1 << src_level
))
97 rctx
->b
.flush_resource(&rctx
->b
, &rsrc
->resource
.b
.b
);
99 assert(!(rsrc
->dirty_level_mask
& (1 << src_level
)));
100 assert(!(rdst
->dirty_level_mask
& (1 << dst_level
)));
105 /* Same as resource_copy_region, except that both upsampling and downsampling are allowed. */
106 static void r600_copy_region_with_blit(struct pipe_context
*pipe
,
107 struct pipe_resource
*dst
,
109 unsigned dstx
, unsigned dsty
, unsigned dstz
,
110 struct pipe_resource
*src
,
112 const struct pipe_box
*src_box
)
114 struct pipe_blit_info blit
;
116 memset(&blit
, 0, sizeof(blit
));
117 blit
.src
.resource
= src
;
118 blit
.src
.format
= src
->format
;
119 blit
.src
.level
= src_level
;
120 blit
.src
.box
= *src_box
;
121 blit
.dst
.resource
= dst
;
122 blit
.dst
.format
= dst
->format
;
123 blit
.dst
.level
= dst_level
;
124 blit
.dst
.box
.x
= dstx
;
125 blit
.dst
.box
.y
= dsty
;
126 blit
.dst
.box
.z
= dstz
;
127 blit
.dst
.box
.width
= src_box
->width
;
128 blit
.dst
.box
.height
= src_box
->height
;
129 blit
.dst
.box
.depth
= src_box
->depth
;
130 blit
.mask
= util_format_get_mask(src
->format
) &
131 util_format_get_mask(dst
->format
);
132 blit
.filter
= PIPE_TEX_FILTER_NEAREST
;
135 pipe
->blit(pipe
, &blit
);
139 /* Copy from a full GPU texture to a transfer's staging one. */
140 static void r600_copy_to_staging_texture(struct pipe_context
*ctx
, struct r600_transfer
*rtransfer
)
142 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
143 struct pipe_transfer
*transfer
= (struct pipe_transfer
*)rtransfer
;
144 struct pipe_resource
*dst
= &rtransfer
->staging
->b
.b
;
145 struct pipe_resource
*src
= transfer
->resource
;
147 if (src
->nr_samples
> 1) {
148 r600_copy_region_with_blit(ctx
, dst
, 0, 0, 0, 0,
149 src
, transfer
->level
, &transfer
->box
);
153 rctx
->dma_copy(ctx
, dst
, 0, 0, 0, 0, src
, transfer
->level
,
157 /* Copy from a transfer's staging texture to a full GPU one. */
158 static void r600_copy_from_staging_texture(struct pipe_context
*ctx
, struct r600_transfer
*rtransfer
)
160 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
161 struct pipe_transfer
*transfer
= (struct pipe_transfer
*)rtransfer
;
162 struct pipe_resource
*dst
= transfer
->resource
;
163 struct pipe_resource
*src
= &rtransfer
->staging
->b
.b
;
164 struct pipe_box sbox
;
166 u_box_3d(0, 0, 0, transfer
->box
.width
, transfer
->box
.height
, transfer
->box
.depth
, &sbox
);
168 if (dst
->nr_samples
> 1) {
169 r600_copy_region_with_blit(ctx
, dst
, transfer
->level
,
170 transfer
->box
.x
, transfer
->box
.y
, transfer
->box
.z
,
175 rctx
->dma_copy(ctx
, dst
, transfer
->level
,
176 transfer
->box
.x
, transfer
->box
.y
, transfer
->box
.z
,
180 static unsigned r600_texture_get_offset(struct r600_common_screen
*rscreen
,
181 struct r600_texture
*rtex
, unsigned level
,
182 const struct pipe_box
*box
,
184 unsigned *layer_stride
)
186 if (rscreen
->chip_class
>= GFX9
) {
187 *stride
= rtex
->surface
.u
.gfx9
.surf_pitch
* rtex
->surface
.bpe
;
188 *layer_stride
= rtex
->surface
.u
.gfx9
.surf_slice_size
;
193 /* Each texture is an array of slices. Each slice is an array
194 * of mipmap levels. */
195 return box
->z
* rtex
->surface
.u
.gfx9
.surf_slice_size
+
196 rtex
->surface
.u
.gfx9
.offset
[level
] +
197 (box
->y
/ rtex
->surface
.blk_h
*
198 rtex
->surface
.u
.gfx9
.surf_pitch
+
199 box
->x
/ rtex
->surface
.blk_w
) * rtex
->surface
.bpe
;
201 *stride
= rtex
->surface
.u
.legacy
.level
[level
].nblk_x
*
203 *layer_stride
= rtex
->surface
.u
.legacy
.level
[level
].slice_size
;
206 return rtex
->surface
.u
.legacy
.level
[level
].offset
;
208 /* Each texture is an array of mipmap levels. Each level is
209 * an array of slices. */
210 return rtex
->surface
.u
.legacy
.level
[level
].offset
+
211 box
->z
* rtex
->surface
.u
.legacy
.level
[level
].slice_size
+
212 (box
->y
/ rtex
->surface
.blk_h
*
213 rtex
->surface
.u
.legacy
.level
[level
].nblk_x
+
214 box
->x
/ rtex
->surface
.blk_w
) * rtex
->surface
.bpe
;
218 static int r600_init_surface(struct r600_common_screen
*rscreen
,
219 struct radeon_surf
*surface
,
220 const struct pipe_resource
*ptex
,
221 enum radeon_surf_mode array_mode
,
222 unsigned pitch_in_bytes_override
,
226 bool is_flushed_depth
,
227 bool tc_compatible_htile
)
229 const struct util_format_description
*desc
=
230 util_format_description(ptex
->format
);
231 bool is_depth
, is_stencil
;
233 unsigned i
, bpe
, flags
= 0;
235 is_depth
= util_format_has_depth(desc
);
236 is_stencil
= util_format_has_stencil(desc
);
238 if (rscreen
->chip_class
>= EVERGREEN
&& !is_flushed_depth
&&
239 ptex
->format
== PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
) {
240 bpe
= 4; /* stencil is allocated separately on evergreen */
242 bpe
= util_format_get_blocksize(ptex
->format
);
243 assert(util_is_power_of_two(bpe
));
246 if (!is_flushed_depth
&& is_depth
) {
247 flags
|= RADEON_SURF_ZBUFFER
;
249 if (tc_compatible_htile
&&
250 (rscreen
->chip_class
>= GFX9
||
251 array_mode
== RADEON_SURF_MODE_2D
)) {
252 /* TC-compatible HTILE only supports Z32_FLOAT.
253 * GFX9 also supports Z16_UNORM.
254 * On VI, promote Z16 to Z32. DB->CB copies will convert
255 * the format for transfers.
257 if (rscreen
->chip_class
== VI
)
260 flags
|= RADEON_SURF_TC_COMPATIBLE_HTILE
;
264 flags
|= RADEON_SURF_SBUFFER
;
267 if (rscreen
->chip_class
>= VI
&&
268 (ptex
->flags
& R600_RESOURCE_FLAG_DISABLE_DCC
||
269 ptex
->format
== PIPE_FORMAT_R9G9B9E5_FLOAT
))
270 flags
|= RADEON_SURF_DISABLE_DCC
;
272 if (ptex
->bind
& PIPE_BIND_SCANOUT
|| is_scanout
) {
273 /* This should catch bugs in gallium users setting incorrect flags. */
274 assert(ptex
->nr_samples
<= 1 &&
275 ptex
->array_size
== 1 &&
277 ptex
->last_level
== 0 &&
278 !(flags
& RADEON_SURF_Z_OR_SBUFFER
));
280 flags
|= RADEON_SURF_SCANOUT
;
284 flags
|= RADEON_SURF_IMPORTED
;
285 if (!(ptex
->flags
& R600_RESOURCE_FLAG_FORCE_TILING
))
286 flags
|= RADEON_SURF_OPTIMIZE_FOR_SPACE
;
288 r
= rscreen
->ws
->surface_init(rscreen
->ws
, ptex
, flags
, bpe
,
289 array_mode
, surface
);
294 if (rscreen
->chip_class
>= GFX9
) {
295 assert(!pitch_in_bytes_override
||
296 pitch_in_bytes_override
== surface
->u
.gfx9
.surf_pitch
* bpe
);
297 surface
->u
.gfx9
.surf_offset
= offset
;
299 if (pitch_in_bytes_override
&&
300 pitch_in_bytes_override
!= surface
->u
.legacy
.level
[0].nblk_x
* bpe
) {
301 /* old ddx on evergreen over estimate alignment for 1d, only 1 level
304 surface
->u
.legacy
.level
[0].nblk_x
= pitch_in_bytes_override
/ bpe
;
305 surface
->u
.legacy
.level
[0].slice_size
= pitch_in_bytes_override
*
306 surface
->u
.legacy
.level
[0].nblk_y
;
310 for (i
= 0; i
< ARRAY_SIZE(surface
->u
.legacy
.level
); ++i
)
311 surface
->u
.legacy
.level
[i
].offset
+= offset
;
317 static void r600_texture_init_metadata(struct r600_common_screen
*rscreen
,
318 struct r600_texture
*rtex
,
319 struct radeon_bo_metadata
*metadata
)
321 struct radeon_surf
*surface
= &rtex
->surface
;
323 memset(metadata
, 0, sizeof(*metadata
));
325 if (rscreen
->chip_class
>= GFX9
) {
326 metadata
->u
.gfx9
.swizzle_mode
= surface
->u
.gfx9
.surf
.swizzle_mode
;
328 metadata
->u
.legacy
.microtile
= surface
->u
.legacy
.level
[0].mode
>= RADEON_SURF_MODE_1D
?
329 RADEON_LAYOUT_TILED
: RADEON_LAYOUT_LINEAR
;
330 metadata
->u
.legacy
.macrotile
= surface
->u
.legacy
.level
[0].mode
>= RADEON_SURF_MODE_2D
?
331 RADEON_LAYOUT_TILED
: RADEON_LAYOUT_LINEAR
;
332 metadata
->u
.legacy
.pipe_config
= surface
->u
.legacy
.pipe_config
;
333 metadata
->u
.legacy
.bankw
= surface
->u
.legacy
.bankw
;
334 metadata
->u
.legacy
.bankh
= surface
->u
.legacy
.bankh
;
335 metadata
->u
.legacy
.tile_split
= surface
->u
.legacy
.tile_split
;
336 metadata
->u
.legacy
.mtilea
= surface
->u
.legacy
.mtilea
;
337 metadata
->u
.legacy
.num_banks
= surface
->u
.legacy
.num_banks
;
338 metadata
->u
.legacy
.stride
= surface
->u
.legacy
.level
[0].nblk_x
* surface
->bpe
;
339 metadata
->u
.legacy
.scanout
= (surface
->flags
& RADEON_SURF_SCANOUT
) != 0;
343 static void r600_eliminate_fast_color_clear(struct r600_common_context
*rctx
,
344 struct r600_texture
*rtex
)
346 struct r600_common_screen
*rscreen
= rctx
->screen
;
347 struct pipe_context
*ctx
= &rctx
->b
;
349 if (ctx
== rscreen
->aux_context
)
350 mtx_lock(&rscreen
->aux_context_lock
);
352 ctx
->flush_resource(ctx
, &rtex
->resource
.b
.b
);
353 ctx
->flush(ctx
, NULL
, 0);
355 if (ctx
== rscreen
->aux_context
)
356 mtx_unlock(&rscreen
->aux_context_lock
);
359 static void r600_texture_discard_cmask(struct r600_common_screen
*rscreen
,
360 struct r600_texture
*rtex
)
362 if (!rtex
->cmask
.size
)
365 assert(rtex
->resource
.b
.b
.nr_samples
<= 1);
368 memset(&rtex
->cmask
, 0, sizeof(rtex
->cmask
));
369 rtex
->cmask
.base_address_reg
= rtex
->resource
.gpu_address
>> 8;
370 rtex
->dirty_level_mask
= 0;
372 if (rscreen
->chip_class
>= SI
)
373 rtex
->cb_color_info
&= ~SI_S_028C70_FAST_CLEAR(1);
375 rtex
->cb_color_info
&= ~EG_S_028C70_FAST_CLEAR(1);
377 if (rtex
->cmask_buffer
!= &rtex
->resource
)
378 r600_resource_reference(&rtex
->cmask_buffer
, NULL
);
380 /* Notify all contexts about the change. */
381 p_atomic_inc(&rscreen
->dirty_tex_counter
);
382 p_atomic_inc(&rscreen
->compressed_colortex_counter
);
385 static bool r600_can_disable_dcc(struct r600_texture
*rtex
)
387 /* We can't disable DCC if it can be written by another process. */
388 return rtex
->dcc_offset
&&
389 (!rtex
->resource
.b
.is_shared
||
390 !(rtex
->resource
.external_usage
& PIPE_HANDLE_USAGE_WRITE
));
393 static bool r600_texture_discard_dcc(struct r600_common_screen
*rscreen
,
394 struct r600_texture
*rtex
)
396 if (!r600_can_disable_dcc(rtex
))
399 assert(rtex
->dcc_separate_buffer
== NULL
);
402 rtex
->dcc_offset
= 0;
404 /* Notify all contexts about the change. */
405 p_atomic_inc(&rscreen
->dirty_tex_counter
);
410 * Disable DCC for the texture. (first decompress, then discard metadata).
412 * There is unresolved multi-context synchronization issue between
413 * screen::aux_context and the current context. If applications do this with
414 * multiple contexts, it's already undefined behavior for them and we don't
415 * have to worry about that. The scenario is:
417 * If context 1 disables DCC and context 2 has queued commands that write
418 * to the texture via CB with DCC enabled, and the order of operations is
420 * context 2 queues draw calls rendering to the texture, but doesn't flush
421 * context 1 disables DCC and flushes
422 * context 1 & 2 reset descriptors and FB state
423 * context 2 flushes (new compressed tiles written by the draw calls)
424 * context 1 & 2 read garbage, because DCC is disabled, yet there are
427 * \param rctx the current context if you have one, or rscreen->aux_context
430 bool r600_texture_disable_dcc(struct r600_common_context
*rctx
,
431 struct r600_texture
*rtex
)
433 struct r600_common_screen
*rscreen
= rctx
->screen
;
435 if (!r600_can_disable_dcc(rtex
))
438 if (&rctx
->b
== rscreen
->aux_context
)
439 mtx_lock(&rscreen
->aux_context_lock
);
441 /* Decompress DCC. */
442 rctx
->decompress_dcc(&rctx
->b
, rtex
);
443 rctx
->b
.flush(&rctx
->b
, NULL
, 0);
445 if (&rctx
->b
== rscreen
->aux_context
)
446 mtx_unlock(&rscreen
->aux_context_lock
);
448 return r600_texture_discard_dcc(rscreen
, rtex
);
451 static void r600_degrade_tile_mode_to_linear(struct r600_common_context
*rctx
,
452 struct r600_texture
*rtex
,
453 bool invalidate_storage
)
455 struct pipe_screen
*screen
= rctx
->b
.screen
;
456 struct r600_texture
*new_tex
;
457 struct pipe_resource templ
= rtex
->resource
.b
.b
;
460 templ
.bind
|= PIPE_BIND_LINEAR
;
462 /* r600g doesn't react to dirty_tex_descriptor_counter */
463 if (rctx
->chip_class
< SI
)
466 if (rtex
->resource
.b
.is_shared
||
467 rtex
->surface
.is_linear
)
470 /* This fails with MSAA, depth, and compressed textures. */
471 if (r600_choose_tiling(rctx
->screen
, &templ
) !=
472 RADEON_SURF_MODE_LINEAR_ALIGNED
)
475 new_tex
= (struct r600_texture
*)screen
->resource_create(screen
, &templ
);
479 /* Copy the pixels to the new texture. */
480 if (!invalidate_storage
) {
481 for (i
= 0; i
<= templ
.last_level
; i
++) {
485 u_minify(templ
.width0
, i
), u_minify(templ
.height0
, i
),
486 util_max_layer(&templ
, i
) + 1, &box
);
488 rctx
->dma_copy(&rctx
->b
, &new_tex
->resource
.b
.b
, i
, 0, 0, 0,
489 &rtex
->resource
.b
.b
, i
, &box
);
493 r600_texture_discard_cmask(rctx
->screen
, rtex
);
494 r600_texture_discard_dcc(rctx
->screen
, rtex
);
496 /* Replace the structure fields of rtex. */
497 rtex
->resource
.b
.b
.bind
= templ
.bind
;
498 pb_reference(&rtex
->resource
.buf
, new_tex
->resource
.buf
);
499 rtex
->resource
.gpu_address
= new_tex
->resource
.gpu_address
;
500 rtex
->resource
.vram_usage
= new_tex
->resource
.vram_usage
;
501 rtex
->resource
.gart_usage
= new_tex
->resource
.gart_usage
;
502 rtex
->resource
.bo_size
= new_tex
->resource
.bo_size
;
503 rtex
->resource
.bo_alignment
= new_tex
->resource
.bo_alignment
;
504 rtex
->resource
.domains
= new_tex
->resource
.domains
;
505 rtex
->resource
.flags
= new_tex
->resource
.flags
;
506 rtex
->size
= new_tex
->size
;
507 rtex
->surface
= new_tex
->surface
;
508 rtex
->non_disp_tiling
= new_tex
->non_disp_tiling
;
509 rtex
->cb_color_info
= new_tex
->cb_color_info
;
510 rtex
->cmask
= new_tex
->cmask
; /* needed even without CMASK */
512 assert(!rtex
->htile_offset
);
513 assert(!rtex
->cmask
.size
);
514 assert(!rtex
->fmask
.size
);
515 assert(!rtex
->dcc_offset
);
516 assert(!rtex
->is_depth
);
518 r600_texture_reference(&new_tex
, NULL
);
520 p_atomic_inc(&rctx
->screen
->dirty_tex_counter
);
523 static boolean
r600_texture_get_handle(struct pipe_screen
* screen
,
524 struct pipe_context
*ctx
,
525 struct pipe_resource
*resource
,
526 struct winsys_handle
*whandle
,
529 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
530 struct r600_common_context
*rctx
;
531 struct r600_resource
*res
= (struct r600_resource
*)resource
;
532 struct r600_texture
*rtex
= (struct r600_texture
*)resource
;
533 struct radeon_bo_metadata metadata
;
534 bool update_metadata
= false;
535 unsigned stride
, offset
, slice_size
;
537 ctx
= threaded_context_unwrap_sync(ctx
);
538 rctx
= (struct r600_common_context
*)(ctx
? ctx
: rscreen
->aux_context
);
540 if (resource
->target
!= PIPE_BUFFER
) {
541 /* This is not supported now, but it might be required for OpenCL
542 * interop in the future.
544 if (resource
->nr_samples
> 1 || rtex
->is_depth
)
547 /* Since shader image stores don't support DCC on VI,
548 * disable it for external clients that want write
551 if (usage
& PIPE_HANDLE_USAGE_WRITE
&& rtex
->dcc_offset
) {
552 if (r600_texture_disable_dcc(rctx
, rtex
))
553 update_metadata
= true;
556 if (!(usage
& PIPE_HANDLE_USAGE_EXPLICIT_FLUSH
) &&
557 (rtex
->cmask
.size
|| rtex
->dcc_offset
)) {
558 /* Eliminate fast clear (both CMASK and DCC) */
559 r600_eliminate_fast_color_clear(rctx
, rtex
);
561 /* Disable CMASK if flush_resource isn't going
564 if (rtex
->cmask
.size
)
565 r600_texture_discard_cmask(rscreen
, rtex
);
569 if (!res
->b
.is_shared
|| update_metadata
) {
570 r600_texture_init_metadata(rscreen
, rtex
, &metadata
);
571 if (rscreen
->query_opaque_metadata
)
572 rscreen
->query_opaque_metadata(rscreen
, rtex
,
575 rscreen
->ws
->buffer_set_metadata(res
->buf
, &metadata
);
578 if (rscreen
->chip_class
>= GFX9
) {
579 offset
= rtex
->surface
.u
.gfx9
.surf_offset
;
580 stride
= rtex
->surface
.u
.gfx9
.surf_pitch
*
582 slice_size
= rtex
->surface
.u
.gfx9
.surf_slice_size
;
584 offset
= rtex
->surface
.u
.legacy
.level
[0].offset
;
585 stride
= rtex
->surface
.u
.legacy
.level
[0].nblk_x
*
587 slice_size
= rtex
->surface
.u
.legacy
.level
[0].slice_size
;
596 if (res
->b
.is_shared
) {
597 /* USAGE_EXPLICIT_FLUSH must be cleared if at least one user
600 res
->external_usage
|= usage
& ~PIPE_HANDLE_USAGE_EXPLICIT_FLUSH
;
601 if (!(usage
& PIPE_HANDLE_USAGE_EXPLICIT_FLUSH
))
602 res
->external_usage
&= ~PIPE_HANDLE_USAGE_EXPLICIT_FLUSH
;
604 res
->b
.is_shared
= true;
605 res
->external_usage
= usage
;
608 return rscreen
->ws
->buffer_get_handle(res
->buf
, stride
, offset
,
609 slice_size
, whandle
);
612 static void r600_texture_destroy(struct pipe_screen
*screen
,
613 struct pipe_resource
*ptex
)
615 struct r600_texture
*rtex
= (struct r600_texture
*)ptex
;
616 struct r600_resource
*resource
= &rtex
->resource
;
618 r600_texture_reference(&rtex
->flushed_depth_texture
, NULL
);
620 if (rtex
->cmask_buffer
!= &rtex
->resource
) {
621 r600_resource_reference(&rtex
->cmask_buffer
, NULL
);
623 pb_reference(&resource
->buf
, NULL
);
624 r600_resource_reference(&rtex
->dcc_separate_buffer
, NULL
);
625 r600_resource_reference(&rtex
->last_dcc_separate_buffer
, NULL
);
629 static const struct u_resource_vtbl r600_texture_vtbl
;
631 /* The number of samples can be specified independently of the texture. */
632 void r600_texture_get_fmask_info(struct r600_common_screen
*rscreen
,
633 struct r600_texture
*rtex
,
635 struct r600_fmask_info
*out
)
637 /* FMASK is allocated like an ordinary texture. */
638 struct pipe_resource templ
= rtex
->resource
.b
.b
;
639 struct radeon_surf fmask
= {};
642 memset(out
, 0, sizeof(*out
));
644 if (rscreen
->chip_class
>= GFX9
) {
645 out
->alignment
= rtex
->surface
.u
.gfx9
.fmask_alignment
;
646 out
->size
= rtex
->surface
.u
.gfx9
.fmask_size
;
650 templ
.nr_samples
= 1;
651 flags
= rtex
->surface
.flags
| RADEON_SURF_FMASK
;
653 if (rscreen
->chip_class
<= CAYMAN
) {
654 /* Use the same parameters and tile mode. */
655 fmask
.u
.legacy
.bankw
= rtex
->surface
.u
.legacy
.bankw
;
656 fmask
.u
.legacy
.bankh
= rtex
->surface
.u
.legacy
.bankh
;
657 fmask
.u
.legacy
.mtilea
= rtex
->surface
.u
.legacy
.mtilea
;
658 fmask
.u
.legacy
.tile_split
= rtex
->surface
.u
.legacy
.tile_split
;
661 fmask
.u
.legacy
.bankh
= 4;
664 switch (nr_samples
) {
673 R600_ERR("Invalid sample count for FMASK allocation.\n");
677 /* Overallocate FMASK on R600-R700 to fix colorbuffer corruption.
678 * This can be fixed by writing a separate FMASK allocator specifically
679 * for R600-R700 asics. */
680 if (rscreen
->chip_class
<= R700
) {
684 if (rscreen
->ws
->surface_init(rscreen
->ws
, &templ
, flags
, bpe
,
685 RADEON_SURF_MODE_2D
, &fmask
)) {
686 R600_ERR("Got error in surface_init while allocating FMASK.\n");
690 assert(fmask
.u
.legacy
.level
[0].mode
== RADEON_SURF_MODE_2D
);
692 out
->slice_tile_max
= (fmask
.u
.legacy
.level
[0].nblk_x
* fmask
.u
.legacy
.level
[0].nblk_y
) / 64;
693 if (out
->slice_tile_max
)
694 out
->slice_tile_max
-= 1;
696 out
->tile_mode_index
= fmask
.u
.legacy
.tiling_index
[0];
697 out
->pitch_in_pixels
= fmask
.u
.legacy
.level
[0].nblk_x
;
698 out
->bank_height
= fmask
.u
.legacy
.bankh
;
699 out
->alignment
= MAX2(256, fmask
.surf_alignment
);
700 out
->size
= fmask
.surf_size
;
703 static void r600_texture_allocate_fmask(struct r600_common_screen
*rscreen
,
704 struct r600_texture
*rtex
)
706 r600_texture_get_fmask_info(rscreen
, rtex
,
707 rtex
->resource
.b
.b
.nr_samples
, &rtex
->fmask
);
709 rtex
->fmask
.offset
= align64(rtex
->size
, rtex
->fmask
.alignment
);
710 rtex
->size
= rtex
->fmask
.offset
+ rtex
->fmask
.size
;
713 void r600_texture_get_cmask_info(struct r600_common_screen
*rscreen
,
714 struct r600_texture
*rtex
,
715 struct r600_cmask_info
*out
)
717 unsigned cmask_tile_width
= 8;
718 unsigned cmask_tile_height
= 8;
719 unsigned cmask_tile_elements
= cmask_tile_width
* cmask_tile_height
;
720 unsigned element_bits
= 4;
721 unsigned cmask_cache_bits
= 1024;
722 unsigned num_pipes
= rscreen
->info
.num_tile_pipes
;
723 unsigned pipe_interleave_bytes
= rscreen
->info
.pipe_interleave_bytes
;
725 unsigned elements_per_macro_tile
= (cmask_cache_bits
/ element_bits
) * num_pipes
;
726 unsigned pixels_per_macro_tile
= elements_per_macro_tile
* cmask_tile_elements
;
727 unsigned sqrt_pixels_per_macro_tile
= sqrt(pixels_per_macro_tile
);
728 unsigned macro_tile_width
= util_next_power_of_two(sqrt_pixels_per_macro_tile
);
729 unsigned macro_tile_height
= pixels_per_macro_tile
/ macro_tile_width
;
731 unsigned pitch_elements
= align(rtex
->resource
.b
.b
.width0
, macro_tile_width
);
732 unsigned height
= align(rtex
->resource
.b
.b
.height0
, macro_tile_height
);
734 unsigned base_align
= num_pipes
* pipe_interleave_bytes
;
735 unsigned slice_bytes
=
736 ((pitch_elements
* height
* element_bits
+ 7) / 8) / cmask_tile_elements
;
738 assert(macro_tile_width
% 128 == 0);
739 assert(macro_tile_height
% 128 == 0);
741 out
->slice_tile_max
= ((pitch_elements
* height
) / (128*128)) - 1;
742 out
->alignment
= MAX2(256, base_align
);
743 out
->size
= (util_max_layer(&rtex
->resource
.b
.b
, 0) + 1) *
744 align(slice_bytes
, base_align
);
747 static void si_texture_get_cmask_info(struct r600_common_screen
*rscreen
,
748 struct r600_texture
*rtex
,
749 struct r600_cmask_info
*out
)
751 unsigned pipe_interleave_bytes
= rscreen
->info
.pipe_interleave_bytes
;
752 unsigned num_pipes
= rscreen
->info
.num_tile_pipes
;
753 unsigned cl_width
, cl_height
;
755 if (rscreen
->chip_class
>= GFX9
) {
756 out
->alignment
= rtex
->surface
.u
.gfx9
.cmask_alignment
;
757 out
->size
= rtex
->surface
.u
.gfx9
.cmask_size
;
774 case 16: /* Hawaii */
783 unsigned base_align
= num_pipes
* pipe_interleave_bytes
;
785 unsigned width
= align(rtex
->resource
.b
.b
.width0
, cl_width
*8);
786 unsigned height
= align(rtex
->resource
.b
.b
.height0
, cl_height
*8);
787 unsigned slice_elements
= (width
* height
) / (8*8);
789 /* Each element of CMASK is a nibble. */
790 unsigned slice_bytes
= slice_elements
/ 2;
792 out
->slice_tile_max
= (width
* height
) / (128*128);
793 if (out
->slice_tile_max
)
794 out
->slice_tile_max
-= 1;
796 out
->alignment
= MAX2(256, base_align
);
797 out
->size
= (util_max_layer(&rtex
->resource
.b
.b
, 0) + 1) *
798 align(slice_bytes
, base_align
);
801 static void r600_texture_allocate_cmask(struct r600_common_screen
*rscreen
,
802 struct r600_texture
*rtex
)
804 if (rscreen
->chip_class
>= SI
) {
805 si_texture_get_cmask_info(rscreen
, rtex
, &rtex
->cmask
);
807 r600_texture_get_cmask_info(rscreen
, rtex
, &rtex
->cmask
);
810 rtex
->cmask
.offset
= align64(rtex
->size
, rtex
->cmask
.alignment
);
811 rtex
->size
= rtex
->cmask
.offset
+ rtex
->cmask
.size
;
813 if (rscreen
->chip_class
>= SI
)
814 rtex
->cb_color_info
|= SI_S_028C70_FAST_CLEAR(1);
816 rtex
->cb_color_info
|= EG_S_028C70_FAST_CLEAR(1);
819 static void r600_texture_alloc_cmask_separate(struct r600_common_screen
*rscreen
,
820 struct r600_texture
*rtex
)
822 if (rtex
->cmask_buffer
)
825 assert(rtex
->cmask
.size
== 0);
827 if (rscreen
->chip_class
>= SI
) {
828 si_texture_get_cmask_info(rscreen
, rtex
, &rtex
->cmask
);
830 r600_texture_get_cmask_info(rscreen
, rtex
, &rtex
->cmask
);
833 rtex
->cmask_buffer
= (struct r600_resource
*)
834 r600_aligned_buffer_create(&rscreen
->b
,
835 R600_RESOURCE_FLAG_UNMAPPABLE
,
838 rtex
->cmask
.alignment
);
839 if (rtex
->cmask_buffer
== NULL
) {
840 rtex
->cmask
.size
= 0;
844 /* update colorbuffer state bits */
845 rtex
->cmask
.base_address_reg
= rtex
->cmask_buffer
->gpu_address
>> 8;
847 if (rscreen
->chip_class
>= SI
)
848 rtex
->cb_color_info
|= SI_S_028C70_FAST_CLEAR(1);
850 rtex
->cb_color_info
|= EG_S_028C70_FAST_CLEAR(1);
852 p_atomic_inc(&rscreen
->compressed_colortex_counter
);
855 static void r600_texture_get_htile_size(struct r600_common_screen
*rscreen
,
856 struct r600_texture
*rtex
)
858 unsigned cl_width
, cl_height
, width
, height
;
859 unsigned slice_elements
, slice_bytes
, pipe_interleave_bytes
, base_align
;
860 unsigned num_pipes
= rscreen
->info
.num_tile_pipes
;
862 assert(rscreen
->chip_class
<= VI
);
864 rtex
->surface
.htile_size
= 0;
866 if (rscreen
->chip_class
<= EVERGREEN
&&
867 rscreen
->info
.drm_major
== 2 && rscreen
->info
.drm_minor
< 26)
870 /* HW bug on R6xx. */
871 if (rscreen
->chip_class
== R600
&&
872 (rtex
->resource
.b
.b
.width0
> 7680 ||
873 rtex
->resource
.b
.b
.height0
> 7680))
876 /* HTILE is broken with 1D tiling on old kernels and CIK. */
877 if (rscreen
->chip_class
>= CIK
&&
878 rtex
->surface
.u
.legacy
.level
[0].mode
== RADEON_SURF_MODE_1D
&&
879 rscreen
->info
.drm_major
== 2 && rscreen
->info
.drm_minor
< 38)
882 /* Overalign HTILE on P2 configs to work around GPU hangs in
883 * piglit/depthstencil-render-miplevels 585.
885 * This has been confirmed to help Kabini & Stoney, where the hangs
886 * are always reproducible. I think I have seen the test hang
887 * on Carrizo too, though it was very rare there.
889 if (rscreen
->chip_class
>= CIK
&& num_pipes
< 4)
918 width
= align(rtex
->resource
.b
.b
.width0
, cl_width
* 8);
919 height
= align(rtex
->resource
.b
.b
.height0
, cl_height
* 8);
921 slice_elements
= (width
* height
) / (8 * 8);
922 slice_bytes
= slice_elements
* 4;
924 pipe_interleave_bytes
= rscreen
->info
.pipe_interleave_bytes
;
925 base_align
= num_pipes
* pipe_interleave_bytes
;
927 rtex
->surface
.htile_alignment
= base_align
;
928 rtex
->surface
.htile_size
=
929 (util_max_layer(&rtex
->resource
.b
.b
, 0) + 1) *
930 align(slice_bytes
, base_align
);
933 static void r600_texture_allocate_htile(struct r600_common_screen
*rscreen
,
934 struct r600_texture
*rtex
)
936 if (rscreen
->chip_class
<= VI
&& !rtex
->tc_compatible_htile
)
937 r600_texture_get_htile_size(rscreen
, rtex
);
939 if (!rtex
->surface
.htile_size
)
942 rtex
->htile_offset
= align(rtex
->size
, rtex
->surface
.htile_alignment
);
943 rtex
->size
= rtex
->htile_offset
+ rtex
->surface
.htile_size
;
946 void r600_print_texture_info(struct r600_common_screen
*rscreen
,
947 struct r600_texture
*rtex
, FILE *f
)
951 /* Common parameters. */
952 fprintf(f
, " Info: npix_x=%u, npix_y=%u, npix_z=%u, blk_w=%u, "
953 "blk_h=%u, array_size=%u, last_level=%u, "
954 "bpe=%u, nsamples=%u, flags=0x%x, %s\n",
955 rtex
->resource
.b
.b
.width0
, rtex
->resource
.b
.b
.height0
,
956 rtex
->resource
.b
.b
.depth0
, rtex
->surface
.blk_w
,
958 rtex
->resource
.b
.b
.array_size
, rtex
->resource
.b
.b
.last_level
,
959 rtex
->surface
.bpe
, rtex
->resource
.b
.b
.nr_samples
,
960 rtex
->surface
.flags
, util_format_short_name(rtex
->resource
.b
.b
.format
));
962 if (rscreen
->chip_class
>= GFX9
) {
963 fprintf(f
, " Surf: size=%"PRIu64
", slice_size=%"PRIu64
", "
964 "alignment=%u, swmode=%u, epitch=%u, pitch=%u\n",
965 rtex
->surface
.surf_size
,
966 rtex
->surface
.u
.gfx9
.surf_slice_size
,
967 rtex
->surface
.surf_alignment
,
968 rtex
->surface
.u
.gfx9
.surf
.swizzle_mode
,
969 rtex
->surface
.u
.gfx9
.surf
.epitch
,
970 rtex
->surface
.u
.gfx9
.surf_pitch
);
972 if (rtex
->fmask
.size
) {
973 fprintf(f
, " FMASK: offset=%"PRIu64
", size=%"PRIu64
", "
974 "alignment=%u, swmode=%u, epitch=%u\n",
976 rtex
->surface
.u
.gfx9
.fmask_size
,
977 rtex
->surface
.u
.gfx9
.fmask_alignment
,
978 rtex
->surface
.u
.gfx9
.fmask
.swizzle_mode
,
979 rtex
->surface
.u
.gfx9
.fmask
.epitch
);
982 if (rtex
->cmask
.size
) {
983 fprintf(f
, " CMask: offset=%"PRIu64
", size=%"PRIu64
", "
984 "alignment=%u, rb_aligned=%u, pipe_aligned=%u\n",
986 rtex
->surface
.u
.gfx9
.cmask_size
,
987 rtex
->surface
.u
.gfx9
.cmask_alignment
,
988 rtex
->surface
.u
.gfx9
.cmask
.rb_aligned
,
989 rtex
->surface
.u
.gfx9
.cmask
.pipe_aligned
);
992 if (rtex
->htile_offset
) {
993 fprintf(f
, " HTile: offset=%"PRIu64
", size=%"PRIu64
", alignment=%u, "
994 "rb_aligned=%u, pipe_aligned=%u\n",
996 rtex
->surface
.htile_size
,
997 rtex
->surface
.htile_alignment
,
998 rtex
->surface
.u
.gfx9
.htile
.rb_aligned
,
999 rtex
->surface
.u
.gfx9
.htile
.pipe_aligned
);
1002 if (rtex
->dcc_offset
) {
1003 fprintf(f
, " DCC: offset=%"PRIu64
", size=%"PRIu64
", "
1004 "alignment=%u, pitch_max=%u, num_dcc_levels=%u\n",
1005 rtex
->dcc_offset
, rtex
->surface
.dcc_size
,
1006 rtex
->surface
.dcc_alignment
,
1007 rtex
->surface
.u
.gfx9
.dcc_pitch_max
,
1008 rtex
->surface
.num_dcc_levels
);
1011 if (rtex
->surface
.u
.gfx9
.stencil_offset
) {
1012 fprintf(f
, " Stencil: offset=%"PRIu64
", swmode=%u, epitch=%u\n",
1013 rtex
->surface
.u
.gfx9
.stencil_offset
,
1014 rtex
->surface
.u
.gfx9
.stencil
.swizzle_mode
,
1015 rtex
->surface
.u
.gfx9
.stencil
.epitch
);
1020 fprintf(f
, " Layout: size=%"PRIu64
", alignment=%u, bankw=%u, "
1021 "bankh=%u, nbanks=%u, mtilea=%u, tilesplit=%u, pipeconfig=%u, scanout=%u\n",
1022 rtex
->surface
.surf_size
, rtex
->surface
.surf_alignment
, rtex
->surface
.u
.legacy
.bankw
,
1023 rtex
->surface
.u
.legacy
.bankh
, rtex
->surface
.u
.legacy
.num_banks
, rtex
->surface
.u
.legacy
.mtilea
,
1024 rtex
->surface
.u
.legacy
.tile_split
, rtex
->surface
.u
.legacy
.pipe_config
,
1025 (rtex
->surface
.flags
& RADEON_SURF_SCANOUT
) != 0);
1027 if (rtex
->fmask
.size
)
1028 fprintf(f
, " FMask: offset=%"PRIu64
", size=%"PRIu64
", alignment=%u, pitch_in_pixels=%u, "
1029 "bankh=%u, slice_tile_max=%u, tile_mode_index=%u\n",
1030 rtex
->fmask
.offset
, rtex
->fmask
.size
, rtex
->fmask
.alignment
,
1031 rtex
->fmask
.pitch_in_pixels
, rtex
->fmask
.bank_height
,
1032 rtex
->fmask
.slice_tile_max
, rtex
->fmask
.tile_mode_index
);
1034 if (rtex
->cmask
.size
)
1035 fprintf(f
, " CMask: offset=%"PRIu64
", size=%"PRIu64
", alignment=%u, "
1036 "slice_tile_max=%u\n",
1037 rtex
->cmask
.offset
, rtex
->cmask
.size
, rtex
->cmask
.alignment
,
1038 rtex
->cmask
.slice_tile_max
);
1040 if (rtex
->htile_offset
)
1041 fprintf(f
, " HTile: offset=%"PRIu64
", size=%"PRIu64
", "
1042 "alignment=%u, TC_compatible = %u\n",
1043 rtex
->htile_offset
, rtex
->surface
.htile_size
,
1044 rtex
->surface
.htile_alignment
,
1045 rtex
->tc_compatible_htile
);
1047 if (rtex
->dcc_offset
) {
1048 fprintf(f
, " DCC: offset=%"PRIu64
", size=%"PRIu64
", alignment=%u\n",
1049 rtex
->dcc_offset
, rtex
->surface
.dcc_size
,
1050 rtex
->surface
.dcc_alignment
);
1051 for (i
= 0; i
<= rtex
->resource
.b
.b
.last_level
; i
++)
1052 fprintf(f
, " DCCLevel[%i]: enabled=%u, offset=%"PRIu64
", "
1053 "fast_clear_size=%"PRIu64
"\n",
1054 i
, i
< rtex
->surface
.num_dcc_levels
,
1055 rtex
->surface
.u
.legacy
.level
[i
].dcc_offset
,
1056 rtex
->surface
.u
.legacy
.level
[i
].dcc_fast_clear_size
);
1059 for (i
= 0; i
<= rtex
->resource
.b
.b
.last_level
; i
++)
1060 fprintf(f
, " Level[%i]: offset=%"PRIu64
", slice_size=%"PRIu64
", "
1061 "npix_x=%u, npix_y=%u, npix_z=%u, nblk_x=%u, nblk_y=%u, "
1062 "mode=%u, tiling_index = %u\n",
1063 i
, rtex
->surface
.u
.legacy
.level
[i
].offset
,
1064 rtex
->surface
.u
.legacy
.level
[i
].slice_size
,
1065 u_minify(rtex
->resource
.b
.b
.width0
, i
),
1066 u_minify(rtex
->resource
.b
.b
.height0
, i
),
1067 u_minify(rtex
->resource
.b
.b
.depth0
, i
),
1068 rtex
->surface
.u
.legacy
.level
[i
].nblk_x
,
1069 rtex
->surface
.u
.legacy
.level
[i
].nblk_y
,
1070 rtex
->surface
.u
.legacy
.level
[i
].mode
,
1071 rtex
->surface
.u
.legacy
.tiling_index
[i
]);
1073 if (rtex
->surface
.flags
& RADEON_SURF_SBUFFER
) {
1074 fprintf(f
, " StencilLayout: tilesplit=%u\n",
1075 rtex
->surface
.u
.legacy
.stencil_tile_split
);
1076 for (i
= 0; i
<= rtex
->resource
.b
.b
.last_level
; i
++) {
1077 fprintf(f
, " StencilLevel[%i]: offset=%"PRIu64
", "
1078 "slice_size=%"PRIu64
", npix_x=%u, "
1079 "npix_y=%u, npix_z=%u, nblk_x=%u, nblk_y=%u, "
1080 "mode=%u, tiling_index = %u\n",
1081 i
, rtex
->surface
.u
.legacy
.stencil_level
[i
].offset
,
1082 rtex
->surface
.u
.legacy
.stencil_level
[i
].slice_size
,
1083 u_minify(rtex
->resource
.b
.b
.width0
, i
),
1084 u_minify(rtex
->resource
.b
.b
.height0
, i
),
1085 u_minify(rtex
->resource
.b
.b
.depth0
, i
),
1086 rtex
->surface
.u
.legacy
.stencil_level
[i
].nblk_x
,
1087 rtex
->surface
.u
.legacy
.stencil_level
[i
].nblk_y
,
1088 rtex
->surface
.u
.legacy
.stencil_level
[i
].mode
,
1089 rtex
->surface
.u
.legacy
.stencil_tiling_index
[i
]);
1094 /* Common processing for r600_texture_create and r600_texture_from_handle */
1095 static struct r600_texture
*
1096 r600_texture_create_object(struct pipe_screen
*screen
,
1097 const struct pipe_resource
*base
,
1098 struct pb_buffer
*buf
,
1099 struct radeon_surf
*surface
)
1101 struct r600_texture
*rtex
;
1102 struct r600_resource
*resource
;
1103 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
1105 rtex
= CALLOC_STRUCT(r600_texture
);
1109 resource
= &rtex
->resource
;
1110 resource
->b
.b
= *base
;
1111 resource
->b
.b
.next
= NULL
;
1112 resource
->b
.vtbl
= &r600_texture_vtbl
;
1113 pipe_reference_init(&resource
->b
.b
.reference
, 1);
1114 resource
->b
.b
.screen
= screen
;
1116 /* don't include stencil-only formats which we don't support for rendering */
1117 rtex
->is_depth
= util_format_has_depth(util_format_description(rtex
->resource
.b
.b
.format
));
1119 rtex
->surface
= *surface
;
1120 rtex
->size
= rtex
->surface
.surf_size
;
1122 rtex
->tc_compatible_htile
= rtex
->surface
.htile_size
!= 0 &&
1123 (rtex
->surface
.flags
&
1124 RADEON_SURF_TC_COMPATIBLE_HTILE
);
1126 /* TC-compatible HTILE:
1127 * - VI only supports Z32_FLOAT.
1128 * - GFX9 only supports Z32_FLOAT and Z16_UNORM. */
1129 if (rtex
->tc_compatible_htile
) {
1130 if (rscreen
->chip_class
>= GFX9
&&
1131 base
->format
== PIPE_FORMAT_Z16_UNORM
)
1132 rtex
->db_render_format
= base
->format
;
1134 rtex
->db_render_format
= PIPE_FORMAT_Z32_FLOAT
;
1136 rtex
->db_render_format
= base
->format
;
1139 /* Tiled depth textures utilize the non-displayable tile order.
1140 * This must be done after r600_setup_surface.
1141 * Applies to R600-Cayman. */
1142 rtex
->non_disp_tiling
= rtex
->is_depth
&& rtex
->surface
.u
.legacy
.level
[0].mode
>= RADEON_SURF_MODE_1D
;
1143 /* Applies to GCN. */
1144 rtex
->last_msaa_resolve_target_micro_mode
= rtex
->surface
.micro_tile_mode
;
1146 /* Disable separate DCC at the beginning. DRI2 doesn't reuse buffers
1147 * between frames, so the only thing that can enable separate DCC
1148 * with DRI2 is multiple slow clears within a frame.
1150 rtex
->ps_draw_ratio
= 0;
1152 if (rtex
->is_depth
) {
1153 if (base
->flags
& (R600_RESOURCE_FLAG_TRANSFER
|
1154 R600_RESOURCE_FLAG_FLUSHED_DEPTH
) ||
1155 rscreen
->chip_class
>= EVERGREEN
) {
1156 if (rscreen
->chip_class
>= GFX9
) {
1157 rtex
->can_sample_z
= true;
1158 rtex
->can_sample_s
= true;
1160 rtex
->can_sample_z
= !rtex
->surface
.u
.legacy
.depth_adjusted
;
1161 rtex
->can_sample_s
= !rtex
->surface
.u
.legacy
.stencil_adjusted
;
1164 if (rtex
->resource
.b
.b
.nr_samples
<= 1 &&
1165 (rtex
->resource
.b
.b
.format
== PIPE_FORMAT_Z16_UNORM
||
1166 rtex
->resource
.b
.b
.format
== PIPE_FORMAT_Z32_FLOAT
))
1167 rtex
->can_sample_z
= true;
1170 if (!(base
->flags
& (R600_RESOURCE_FLAG_TRANSFER
|
1171 R600_RESOURCE_FLAG_FLUSHED_DEPTH
))) {
1172 rtex
->db_compatible
= true;
1174 if (!(rscreen
->debug_flags
& DBG_NO_HYPERZ
))
1175 r600_texture_allocate_htile(rscreen
, rtex
);
1178 if (base
->nr_samples
> 1) {
1180 r600_texture_allocate_fmask(rscreen
, rtex
);
1181 r600_texture_allocate_cmask(rscreen
, rtex
);
1182 rtex
->cmask_buffer
= &rtex
->resource
;
1184 if (!rtex
->fmask
.size
|| !rtex
->cmask
.size
) {
1190 /* Shared textures must always set up DCC here.
1191 * If it's not present, it will be disabled by
1192 * apply_opaque_metadata later.
1194 if (rtex
->surface
.dcc_size
&&
1195 (buf
|| !(rscreen
->debug_flags
& DBG_NO_DCC
)) &&
1196 !(rtex
->surface
.flags
& RADEON_SURF_SCANOUT
)) {
1197 /* Reserve space for the DCC buffer. */
1198 rtex
->dcc_offset
= align64(rtex
->size
, rtex
->surface
.dcc_alignment
);
1199 rtex
->size
= rtex
->dcc_offset
+ rtex
->surface
.dcc_size
;
1203 /* Now create the backing buffer. */
1205 r600_init_resource_fields(rscreen
, resource
, rtex
->size
,
1206 rtex
->surface
.surf_alignment
);
1208 resource
->flags
|= RADEON_FLAG_NO_SUBALLOC
;
1210 if (!r600_alloc_resource(rscreen
, resource
)) {
1215 resource
->buf
= buf
;
1216 resource
->gpu_address
= rscreen
->ws
->buffer_get_virtual_address(resource
->buf
);
1217 resource
->bo_size
= buf
->size
;
1218 resource
->bo_alignment
= buf
->alignment
;
1219 resource
->domains
= rscreen
->ws
->buffer_get_initial_domain(resource
->buf
);
1220 if (resource
->domains
& RADEON_DOMAIN_VRAM
)
1221 resource
->vram_usage
= buf
->size
;
1222 else if (resource
->domains
& RADEON_DOMAIN_GTT
)
1223 resource
->gart_usage
= buf
->size
;
1226 if (rtex
->cmask
.size
) {
1227 /* Initialize the cmask to 0xCC (= compressed state). */
1228 r600_screen_clear_buffer(rscreen
, &rtex
->cmask_buffer
->b
.b
,
1229 rtex
->cmask
.offset
, rtex
->cmask
.size
,
1232 if (rtex
->htile_offset
) {
1233 uint32_t clear_value
= 0;
1235 if (rscreen
->chip_class
>= GFX9
|| rtex
->tc_compatible_htile
)
1236 clear_value
= 0x0000030F;
1238 r600_screen_clear_buffer(rscreen
, &rtex
->resource
.b
.b
,
1240 rtex
->surface
.htile_size
,
1244 /* Initialize DCC only if the texture is not being imported. */
1245 if (!buf
&& rtex
->dcc_offset
) {
1246 r600_screen_clear_buffer(rscreen
, &rtex
->resource
.b
.b
,
1248 rtex
->surface
.dcc_size
,
1252 /* Initialize the CMASK base register value. */
1253 rtex
->cmask
.base_address_reg
=
1254 (rtex
->resource
.gpu_address
+ rtex
->cmask
.offset
) >> 8;
1256 if (rscreen
->debug_flags
& DBG_VM
) {
1257 fprintf(stderr
, "VM start=0x%"PRIX64
" end=0x%"PRIX64
" | Texture %ix%ix%i, %i levels, %i samples, %s\n",
1258 rtex
->resource
.gpu_address
,
1259 rtex
->resource
.gpu_address
+ rtex
->resource
.buf
->size
,
1260 base
->width0
, base
->height0
, util_max_layer(base
, 0)+1, base
->last_level
+1,
1261 base
->nr_samples
? base
->nr_samples
: 1, util_format_short_name(base
->format
));
1264 if (rscreen
->debug_flags
& DBG_TEX
) {
1266 r600_print_texture_info(rscreen
, rtex
, stdout
);
1273 static enum radeon_surf_mode
1274 r600_choose_tiling(struct r600_common_screen
*rscreen
,
1275 const struct pipe_resource
*templ
)
1277 const struct util_format_description
*desc
= util_format_description(templ
->format
);
1278 bool force_tiling
= templ
->flags
& R600_RESOURCE_FLAG_FORCE_TILING
;
1279 bool is_depth_stencil
= util_format_is_depth_or_stencil(templ
->format
) &&
1280 !(templ
->flags
& R600_RESOURCE_FLAG_FLUSHED_DEPTH
);
1282 /* MSAA resources must be 2D tiled. */
1283 if (templ
->nr_samples
> 1)
1284 return RADEON_SURF_MODE_2D
;
1286 /* Transfer resources should be linear. */
1287 if (templ
->flags
& R600_RESOURCE_FLAG_TRANSFER
)
1288 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1290 /* Avoid Z/S decompress blits by forcing TC-compatible HTILE on VI,
1291 * which requires 2D tiling.
1293 if (rscreen
->chip_class
== VI
&&
1295 (templ
->flags
& PIPE_RESOURCE_FLAG_TEXTURING_MORE_LIKELY
))
1296 return RADEON_SURF_MODE_2D
;
1298 /* r600g: force tiling on TEXTURE_2D and TEXTURE_3D compute resources. */
1299 if (rscreen
->chip_class
>= R600
&& rscreen
->chip_class
<= CAYMAN
&&
1300 (templ
->bind
& PIPE_BIND_COMPUTE_RESOURCE
) &&
1301 (templ
->target
== PIPE_TEXTURE_2D
||
1302 templ
->target
== PIPE_TEXTURE_3D
))
1303 force_tiling
= true;
1305 /* Handle common candidates for the linear mode.
1306 * Compressed textures and DB surfaces must always be tiled.
1308 if (!force_tiling
&&
1309 !is_depth_stencil
&&
1310 !util_format_is_compressed(templ
->format
)) {
1311 if (rscreen
->debug_flags
& DBG_NO_TILING
)
1312 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1314 /* Tiling doesn't work with the 422 (SUBSAMPLED) formats on R600+. */
1315 if (desc
->layout
== UTIL_FORMAT_LAYOUT_SUBSAMPLED
)
1316 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1318 /* Cursors are linear on SI.
1319 * (XXX double-check, maybe also use RADEON_SURF_SCANOUT) */
1320 if (rscreen
->chip_class
>= SI
&&
1321 (templ
->bind
& PIPE_BIND_CURSOR
))
1322 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1324 if (templ
->bind
& PIPE_BIND_LINEAR
)
1325 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1327 /* Textures with a very small height are recommended to be linear. */
1328 if (templ
->target
== PIPE_TEXTURE_1D
||
1329 templ
->target
== PIPE_TEXTURE_1D_ARRAY
||
1330 /* Only very thin and long 2D textures should benefit from
1331 * linear_aligned. */
1332 (templ
->width0
> 8 && templ
->height0
<= 2))
1333 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1335 /* Textures likely to be mapped often. */
1336 if (templ
->usage
== PIPE_USAGE_STAGING
||
1337 templ
->usage
== PIPE_USAGE_STREAM
)
1338 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1341 /* Make small textures 1D tiled. */
1342 if (templ
->width0
<= 16 || templ
->height0
<= 16 ||
1343 (rscreen
->debug_flags
& DBG_NO_2D_TILING
))
1344 return RADEON_SURF_MODE_1D
;
1346 /* The allocator will switch to 1D if needed. */
1347 return RADEON_SURF_MODE_2D
;
1350 struct pipe_resource
*r600_texture_create(struct pipe_screen
*screen
,
1351 const struct pipe_resource
*templ
)
1353 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
1354 struct radeon_surf surface
= {0};
1355 bool is_flushed_depth
= templ
->flags
& R600_RESOURCE_FLAG_FLUSHED_DEPTH
;
1356 bool tc_compatible_htile
=
1357 rscreen
->chip_class
>= VI
&&
1358 (templ
->flags
& PIPE_RESOURCE_FLAG_TEXTURING_MORE_LIKELY
) &&
1359 !(rscreen
->debug_flags
& DBG_NO_HYPERZ
) &&
1360 !is_flushed_depth
&&
1361 templ
->nr_samples
<= 1 && /* TC-compat HTILE is less efficient with MSAA */
1362 util_format_is_depth_or_stencil(templ
->format
);
1366 r
= r600_init_surface(rscreen
, &surface
, templ
,
1367 r600_choose_tiling(rscreen
, templ
), 0, 0,
1368 false, false, is_flushed_depth
,
1369 tc_compatible_htile
);
1374 return (struct pipe_resource
*)
1375 r600_texture_create_object(screen
, templ
, NULL
, &surface
);
1378 static struct pipe_resource
*r600_texture_from_handle(struct pipe_screen
*screen
,
1379 const struct pipe_resource
*templ
,
1380 struct winsys_handle
*whandle
,
1383 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
1384 struct pb_buffer
*buf
= NULL
;
1385 unsigned stride
= 0, offset
= 0;
1386 unsigned array_mode
;
1387 struct radeon_surf surface
;
1389 struct radeon_bo_metadata metadata
= {};
1390 struct r600_texture
*rtex
;
1393 /* Support only 2D textures without mipmaps */
1394 if ((templ
->target
!= PIPE_TEXTURE_2D
&& templ
->target
!= PIPE_TEXTURE_RECT
) ||
1395 templ
->depth0
!= 1 || templ
->last_level
!= 0)
1398 buf
= rscreen
->ws
->buffer_from_handle(rscreen
->ws
, whandle
, &stride
, &offset
);
1402 rscreen
->ws
->buffer_get_metadata(buf
, &metadata
);
1404 if (rscreen
->chip_class
>= GFX9
) {
1405 if (metadata
.u
.gfx9
.swizzle_mode
> 0)
1406 array_mode
= RADEON_SURF_MODE_2D
;
1408 array_mode
= RADEON_SURF_MODE_LINEAR_ALIGNED
;
1410 is_scanout
= metadata
.u
.gfx9
.swizzle_mode
== 0 ||
1411 metadata
.u
.gfx9
.swizzle_mode
% 4 == 2;
1413 surface
.u
.legacy
.pipe_config
= metadata
.u
.legacy
.pipe_config
;
1414 surface
.u
.legacy
.bankw
= metadata
.u
.legacy
.bankw
;
1415 surface
.u
.legacy
.bankh
= metadata
.u
.legacy
.bankh
;
1416 surface
.u
.legacy
.tile_split
= metadata
.u
.legacy
.tile_split
;
1417 surface
.u
.legacy
.mtilea
= metadata
.u
.legacy
.mtilea
;
1418 surface
.u
.legacy
.num_banks
= metadata
.u
.legacy
.num_banks
;
1420 if (metadata
.u
.legacy
.macrotile
== RADEON_LAYOUT_TILED
)
1421 array_mode
= RADEON_SURF_MODE_2D
;
1422 else if (metadata
.u
.legacy
.microtile
== RADEON_LAYOUT_TILED
)
1423 array_mode
= RADEON_SURF_MODE_1D
;
1425 array_mode
= RADEON_SURF_MODE_LINEAR_ALIGNED
;
1427 is_scanout
= metadata
.u
.legacy
.scanout
;
1430 r
= r600_init_surface(rscreen
, &surface
, templ
, array_mode
, stride
,
1431 offset
, true, is_scanout
, false, false);
1436 rtex
= r600_texture_create_object(screen
, templ
, buf
, &surface
);
1440 rtex
->resource
.b
.is_shared
= true;
1441 rtex
->resource
.external_usage
= usage
;
1443 if (rscreen
->apply_opaque_metadata
)
1444 rscreen
->apply_opaque_metadata(rscreen
, rtex
, &metadata
);
1446 /* Validate that addrlib arrived at the same surface parameters. */
1447 if (rscreen
->chip_class
>= GFX9
) {
1448 assert(metadata
.u
.gfx9
.swizzle_mode
== surface
.u
.gfx9
.surf
.swizzle_mode
);
1451 return &rtex
->resource
.b
.b
;
1454 bool r600_init_flushed_depth_texture(struct pipe_context
*ctx
,
1455 struct pipe_resource
*texture
,
1456 struct r600_texture
**staging
)
1458 struct r600_texture
*rtex
= (struct r600_texture
*)texture
;
1459 struct pipe_resource resource
;
1460 struct r600_texture
**flushed_depth_texture
= staging
?
1461 staging
: &rtex
->flushed_depth_texture
;
1462 enum pipe_format pipe_format
= texture
->format
;
1465 if (rtex
->flushed_depth_texture
)
1466 return true; /* it's ready */
1468 if (!rtex
->can_sample_z
&& rtex
->can_sample_s
) {
1469 switch (pipe_format
) {
1470 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
:
1471 /* Save memory by not allocating the S plane. */
1472 pipe_format
= PIPE_FORMAT_Z32_FLOAT
;
1474 case PIPE_FORMAT_Z24_UNORM_S8_UINT
:
1475 case PIPE_FORMAT_S8_UINT_Z24_UNORM
:
1476 /* Save memory bandwidth by not copying the
1477 * stencil part during flush.
1479 * This potentially increases memory bandwidth
1480 * if an application uses both Z and S texturing
1481 * simultaneously (a flushed Z24S8 texture
1482 * would be stored compactly), but how often
1483 * does that really happen?
1485 pipe_format
= PIPE_FORMAT_Z24X8_UNORM
;
1489 } else if (!rtex
->can_sample_s
&& rtex
->can_sample_z
) {
1490 assert(util_format_has_stencil(util_format_description(pipe_format
)));
1492 /* DB->CB copies to an 8bpp surface don't work. */
1493 pipe_format
= PIPE_FORMAT_X24S8_UINT
;
1497 memset(&resource
, 0, sizeof(resource
));
1498 resource
.target
= texture
->target
;
1499 resource
.format
= pipe_format
;
1500 resource
.width0
= texture
->width0
;
1501 resource
.height0
= texture
->height0
;
1502 resource
.depth0
= texture
->depth0
;
1503 resource
.array_size
= texture
->array_size
;
1504 resource
.last_level
= texture
->last_level
;
1505 resource
.nr_samples
= texture
->nr_samples
;
1506 resource
.usage
= staging
? PIPE_USAGE_STAGING
: PIPE_USAGE_DEFAULT
;
1507 resource
.bind
= texture
->bind
& ~PIPE_BIND_DEPTH_STENCIL
;
1508 resource
.flags
= texture
->flags
| R600_RESOURCE_FLAG_FLUSHED_DEPTH
;
1511 resource
.flags
|= R600_RESOURCE_FLAG_TRANSFER
;
1513 *flushed_depth_texture
= (struct r600_texture
*)ctx
->screen
->resource_create(ctx
->screen
, &resource
);
1514 if (*flushed_depth_texture
== NULL
) {
1515 R600_ERR("failed to create temporary texture to hold flushed depth\n");
1519 (*flushed_depth_texture
)->non_disp_tiling
= false;
1524 * Initialize the pipe_resource descriptor to be of the same size as the box,
1525 * which is supposed to hold a subregion of the texture "orig" at the given
1528 static void r600_init_temp_resource_from_box(struct pipe_resource
*res
,
1529 struct pipe_resource
*orig
,
1530 const struct pipe_box
*box
,
1531 unsigned level
, unsigned flags
)
1533 memset(res
, 0, sizeof(*res
));
1534 res
->format
= orig
->format
;
1535 res
->width0
= box
->width
;
1536 res
->height0
= box
->height
;
1538 res
->array_size
= 1;
1539 res
->usage
= flags
& R600_RESOURCE_FLAG_TRANSFER
? PIPE_USAGE_STAGING
: PIPE_USAGE_DEFAULT
;
1542 /* We must set the correct texture target and dimensions for a 3D box. */
1543 if (box
->depth
> 1 && util_max_layer(orig
, level
) > 0) {
1544 res
->target
= PIPE_TEXTURE_2D_ARRAY
;
1545 res
->array_size
= box
->depth
;
1547 res
->target
= PIPE_TEXTURE_2D
;
1551 static bool r600_can_invalidate_texture(struct r600_common_screen
*rscreen
,
1552 struct r600_texture
*rtex
,
1553 unsigned transfer_usage
,
1554 const struct pipe_box
*box
)
1556 /* r600g doesn't react to dirty_tex_descriptor_counter */
1557 return rscreen
->chip_class
>= SI
&&
1558 !rtex
->resource
.b
.is_shared
&&
1559 !(transfer_usage
& PIPE_TRANSFER_READ
) &&
1560 rtex
->resource
.b
.b
.last_level
== 0 &&
1561 util_texrange_covers_whole_level(&rtex
->resource
.b
.b
, 0,
1562 box
->x
, box
->y
, box
->z
,
1563 box
->width
, box
->height
,
1567 static void r600_texture_invalidate_storage(struct r600_common_context
*rctx
,
1568 struct r600_texture
*rtex
)
1570 struct r600_common_screen
*rscreen
= rctx
->screen
;
1572 /* There is no point in discarding depth and tiled buffers. */
1573 assert(!rtex
->is_depth
);
1574 assert(rtex
->surface
.is_linear
);
1576 /* Reallocate the buffer in the same pipe_resource. */
1577 r600_alloc_resource(rscreen
, &rtex
->resource
);
1579 /* Initialize the CMASK base address (needed even without CMASK). */
1580 rtex
->cmask
.base_address_reg
=
1581 (rtex
->resource
.gpu_address
+ rtex
->cmask
.offset
) >> 8;
1583 p_atomic_inc(&rscreen
->dirty_tex_counter
);
1585 rctx
->num_alloc_tex_transfer_bytes
+= rtex
->size
;
1588 static void *r600_texture_transfer_map(struct pipe_context
*ctx
,
1589 struct pipe_resource
*texture
,
1592 const struct pipe_box
*box
,
1593 struct pipe_transfer
**ptransfer
)
1595 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
1596 struct r600_texture
*rtex
= (struct r600_texture
*)texture
;
1597 struct r600_transfer
*trans
;
1598 struct r600_resource
*buf
;
1599 unsigned offset
= 0;
1601 bool use_staging_texture
= false;
1603 assert(!(texture
->flags
& R600_RESOURCE_FLAG_TRANSFER
));
1604 assert(box
->width
&& box
->height
&& box
->depth
);
1606 /* Depth textures use staging unconditionally. */
1607 if (!rtex
->is_depth
) {
1608 /* Degrade the tile mode if we get too many transfers on APUs.
1609 * On dGPUs, the staging texture is always faster.
1610 * Only count uploads that are at least 4x4 pixels large.
1612 if (!rctx
->screen
->info
.has_dedicated_vram
&&
1614 box
->width
>= 4 && box
->height
>= 4 &&
1615 p_atomic_inc_return(&rtex
->num_level0_transfers
) == 10) {
1616 bool can_invalidate
=
1617 r600_can_invalidate_texture(rctx
->screen
, rtex
,
1620 r600_degrade_tile_mode_to_linear(rctx
, rtex
,
1624 /* Tiled textures need to be converted into a linear texture for CPU
1625 * access. The staging texture is always linear and is placed in GART.
1627 * Reading from VRAM or GTT WC is slow, always use the staging
1628 * texture in this case.
1630 * Use the staging texture for uploads if the underlying BO
1633 if (!rtex
->surface
.is_linear
)
1634 use_staging_texture
= true;
1635 else if (usage
& PIPE_TRANSFER_READ
)
1636 use_staging_texture
=
1637 rtex
->resource
.domains
& RADEON_DOMAIN_VRAM
||
1638 rtex
->resource
.flags
& RADEON_FLAG_GTT_WC
;
1639 /* Write & linear only: */
1640 else if (r600_rings_is_buffer_referenced(rctx
, rtex
->resource
.buf
,
1641 RADEON_USAGE_READWRITE
) ||
1642 !rctx
->ws
->buffer_wait(rtex
->resource
.buf
, 0,
1643 RADEON_USAGE_READWRITE
)) {
1645 if (r600_can_invalidate_texture(rctx
->screen
, rtex
,
1647 r600_texture_invalidate_storage(rctx
, rtex
);
1649 use_staging_texture
= true;
1653 trans
= CALLOC_STRUCT(r600_transfer
);
1656 pipe_resource_reference(&trans
->b
.b
.resource
, texture
);
1657 trans
->b
.b
.level
= level
;
1658 trans
->b
.b
.usage
= usage
;
1659 trans
->b
.b
.box
= *box
;
1661 if (rtex
->is_depth
) {
1662 struct r600_texture
*staging_depth
;
1664 if (rtex
->resource
.b
.b
.nr_samples
> 1) {
1665 /* MSAA depth buffers need to be converted to single sample buffers.
1667 * Mapping MSAA depth buffers can occur if ReadPixels is called
1668 * with a multisample GLX visual.
1670 * First downsample the depth buffer to a temporary texture,
1671 * then decompress the temporary one to staging.
1673 * Only the region being mapped is transfered.
1675 struct pipe_resource resource
;
1677 r600_init_temp_resource_from_box(&resource
, texture
, box
, level
, 0);
1679 if (!r600_init_flushed_depth_texture(ctx
, &resource
, &staging_depth
)) {
1680 R600_ERR("failed to create temporary texture to hold untiled copy\n");
1685 if (usage
& PIPE_TRANSFER_READ
) {
1686 struct pipe_resource
*temp
= ctx
->screen
->resource_create(ctx
->screen
, &resource
);
1688 R600_ERR("failed to create a temporary depth texture\n");
1693 r600_copy_region_with_blit(ctx
, temp
, 0, 0, 0, 0, texture
, level
, box
);
1694 rctx
->blit_decompress_depth(ctx
, (struct r600_texture
*)temp
, staging_depth
,
1695 0, 0, 0, box
->depth
, 0, 0);
1696 pipe_resource_reference(&temp
, NULL
);
1699 /* Just get the strides. */
1700 r600_texture_get_offset(rctx
->screen
, staging_depth
, level
, NULL
,
1702 &trans
->b
.b
.layer_stride
);
1704 /* XXX: only readback the rectangle which is being mapped? */
1705 /* XXX: when discard is true, no need to read back from depth texture */
1706 if (!r600_init_flushed_depth_texture(ctx
, texture
, &staging_depth
)) {
1707 R600_ERR("failed to create temporary texture to hold untiled copy\n");
1712 rctx
->blit_decompress_depth(ctx
, rtex
, staging_depth
,
1714 box
->z
, box
->z
+ box
->depth
- 1,
1717 offset
= r600_texture_get_offset(rctx
->screen
, staging_depth
,
1720 &trans
->b
.b
.layer_stride
);
1723 trans
->staging
= (struct r600_resource
*)staging_depth
;
1724 buf
= trans
->staging
;
1725 } else if (use_staging_texture
) {
1726 struct pipe_resource resource
;
1727 struct r600_texture
*staging
;
1729 r600_init_temp_resource_from_box(&resource
, texture
, box
, level
,
1730 R600_RESOURCE_FLAG_TRANSFER
);
1731 resource
.usage
= (usage
& PIPE_TRANSFER_READ
) ?
1732 PIPE_USAGE_STAGING
: PIPE_USAGE_STREAM
;
1734 /* Create the temporary texture. */
1735 staging
= (struct r600_texture
*)ctx
->screen
->resource_create(ctx
->screen
, &resource
);
1737 R600_ERR("failed to create temporary texture to hold untiled copy\n");
1741 trans
->staging
= &staging
->resource
;
1743 /* Just get the strides. */
1744 r600_texture_get_offset(rctx
->screen
, staging
, 0, NULL
,
1746 &trans
->b
.b
.layer_stride
);
1748 if (usage
& PIPE_TRANSFER_READ
)
1749 r600_copy_to_staging_texture(ctx
, trans
);
1751 usage
|= PIPE_TRANSFER_UNSYNCHRONIZED
;
1753 buf
= trans
->staging
;
1755 /* the resource is mapped directly */
1756 offset
= r600_texture_get_offset(rctx
->screen
, rtex
, level
, box
,
1758 &trans
->b
.b
.layer_stride
);
1759 buf
= &rtex
->resource
;
1762 if (!(map
= r600_buffer_map_sync_with_rings(rctx
, buf
, usage
))) {
1763 r600_resource_reference(&trans
->staging
, NULL
);
1768 *ptransfer
= &trans
->b
.b
;
1769 return map
+ offset
;
1772 static void r600_texture_transfer_unmap(struct pipe_context
*ctx
,
1773 struct pipe_transfer
* transfer
)
1775 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
1776 struct r600_transfer
*rtransfer
= (struct r600_transfer
*)transfer
;
1777 struct pipe_resource
*texture
= transfer
->resource
;
1778 struct r600_texture
*rtex
= (struct r600_texture
*)texture
;
1780 if ((transfer
->usage
& PIPE_TRANSFER_WRITE
) && rtransfer
->staging
) {
1781 if (rtex
->is_depth
&& rtex
->resource
.b
.b
.nr_samples
<= 1) {
1782 ctx
->resource_copy_region(ctx
, texture
, transfer
->level
,
1783 transfer
->box
.x
, transfer
->box
.y
, transfer
->box
.z
,
1784 &rtransfer
->staging
->b
.b
, transfer
->level
,
1787 r600_copy_from_staging_texture(ctx
, rtransfer
);
1791 if (rtransfer
->staging
) {
1792 rctx
->num_alloc_tex_transfer_bytes
+= rtransfer
->staging
->buf
->size
;
1793 r600_resource_reference(&rtransfer
->staging
, NULL
);
1796 /* Heuristic for {upload, draw, upload, draw, ..}:
1798 * Flush the gfx IB if we've allocated too much texture storage.
1800 * The idea is that we don't want to build IBs that use too much
1801 * memory and put pressure on the kernel memory manager and we also
1802 * want to make temporary and invalidated buffers go idle ASAP to
1803 * decrease the total memory usage or make them reusable. The memory
1804 * usage will be slightly higher than given here because of the buffer
1805 * cache in the winsys.
1807 * The result is that the kernel memory manager is never a bottleneck.
1809 if (rctx
->num_alloc_tex_transfer_bytes
> rctx
->screen
->info
.gart_size
/ 4) {
1810 rctx
->gfx
.flush(rctx
, RADEON_FLUSH_ASYNC
, NULL
);
1811 rctx
->num_alloc_tex_transfer_bytes
= 0;
1814 pipe_resource_reference(&transfer
->resource
, NULL
);
1818 static const struct u_resource_vtbl r600_texture_vtbl
=
1820 NULL
, /* get_handle */
1821 r600_texture_destroy
, /* resource_destroy */
1822 r600_texture_transfer_map
, /* transfer_map */
1823 u_default_transfer_flush_region
, /* transfer_flush_region */
1824 r600_texture_transfer_unmap
, /* transfer_unmap */
1827 /* DCC channel type categories within which formats can be reinterpreted
1828 * while keeping the same DCC encoding. The swizzle must also match. */
1829 enum dcc_channel_type
{
1830 dcc_channel_float32
,
1833 dcc_channel_float16
,
1836 dcc_channel_uint_10_10_10_2
,
1839 dcc_channel_incompatible
,
1842 /* Return the type of DCC encoding. */
1843 static enum dcc_channel_type
1844 vi_get_dcc_channel_type(const struct util_format_description
*desc
)
1848 /* Find the first non-void channel. */
1849 for (i
= 0; i
< desc
->nr_channels
; i
++)
1850 if (desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
1852 if (i
== desc
->nr_channels
)
1853 return dcc_channel_incompatible
;
1855 switch (desc
->channel
[i
].size
) {
1857 if (desc
->channel
[i
].type
== UTIL_FORMAT_TYPE_FLOAT
)
1858 return dcc_channel_float32
;
1859 if (desc
->channel
[i
].type
== UTIL_FORMAT_TYPE_UNSIGNED
)
1860 return dcc_channel_uint32
;
1861 return dcc_channel_sint32
;
1863 if (desc
->channel
[i
].type
== UTIL_FORMAT_TYPE_FLOAT
)
1864 return dcc_channel_float16
;
1865 if (desc
->channel
[i
].type
== UTIL_FORMAT_TYPE_UNSIGNED
)
1866 return dcc_channel_uint16
;
1867 return dcc_channel_sint16
;
1869 return dcc_channel_uint_10_10_10_2
;
1871 if (desc
->channel
[i
].type
== UTIL_FORMAT_TYPE_UNSIGNED
)
1872 return dcc_channel_uint8
;
1873 return dcc_channel_sint8
;
1875 return dcc_channel_incompatible
;
1879 /* Return if it's allowed to reinterpret one format as another with DCC enabled. */
1880 bool vi_dcc_formats_compatible(enum pipe_format format1
,
1881 enum pipe_format format2
)
1883 const struct util_format_description
*desc1
, *desc2
;
1884 enum dcc_channel_type type1
, type2
;
1887 if (format1
== format2
)
1890 desc1
= util_format_description(format1
);
1891 desc2
= util_format_description(format2
);
1893 if (desc1
->nr_channels
!= desc2
->nr_channels
)
1896 /* Swizzles must be the same. */
1897 for (i
= 0; i
< desc1
->nr_channels
; i
++)
1898 if (desc1
->swizzle
[i
] <= PIPE_SWIZZLE_W
&&
1899 desc2
->swizzle
[i
] <= PIPE_SWIZZLE_W
&&
1900 desc1
->swizzle
[i
] != desc2
->swizzle
[i
])
1903 type1
= vi_get_dcc_channel_type(desc1
);
1904 type2
= vi_get_dcc_channel_type(desc2
);
1906 return type1
!= dcc_channel_incompatible
&&
1907 type2
!= dcc_channel_incompatible
&&
1911 bool vi_dcc_formats_are_incompatible(struct pipe_resource
*tex
,
1913 enum pipe_format view_format
)
1915 struct r600_texture
*rtex
= (struct r600_texture
*)tex
;
1917 return vi_dcc_enabled(rtex
, level
) &&
1918 !vi_dcc_formats_compatible(tex
->format
, view_format
);
1921 /* This can't be merged with the above function, because
1922 * vi_dcc_formats_compatible should be called only when DCC is enabled. */
1923 void vi_disable_dcc_if_incompatible_format(struct r600_common_context
*rctx
,
1924 struct pipe_resource
*tex
,
1926 enum pipe_format view_format
)
1928 struct r600_texture
*rtex
= (struct r600_texture
*)tex
;
1930 if (vi_dcc_enabled(rtex
, level
) &&
1931 !vi_dcc_formats_compatible(tex
->format
, view_format
))
1932 if (!r600_texture_disable_dcc(rctx
, (struct r600_texture
*)tex
))
1933 rctx
->decompress_dcc(&rctx
->b
, rtex
);
1936 struct pipe_surface
*r600_create_surface_custom(struct pipe_context
*pipe
,
1937 struct pipe_resource
*texture
,
1938 const struct pipe_surface
*templ
,
1939 unsigned width0
, unsigned height0
,
1940 unsigned width
, unsigned height
)
1942 struct r600_surface
*surface
= CALLOC_STRUCT(r600_surface
);
1947 assert(templ
->u
.tex
.first_layer
<= util_max_layer(texture
, templ
->u
.tex
.level
));
1948 assert(templ
->u
.tex
.last_layer
<= util_max_layer(texture
, templ
->u
.tex
.level
));
1950 pipe_reference_init(&surface
->base
.reference
, 1);
1951 pipe_resource_reference(&surface
->base
.texture
, texture
);
1952 surface
->base
.context
= pipe
;
1953 surface
->base
.format
= templ
->format
;
1954 surface
->base
.width
= width
;
1955 surface
->base
.height
= height
;
1956 surface
->base
.u
= templ
->u
;
1958 surface
->width0
= width0
;
1959 surface
->height0
= height0
;
1961 surface
->dcc_incompatible
=
1962 texture
->target
!= PIPE_BUFFER
&&
1963 vi_dcc_formats_are_incompatible(texture
, templ
->u
.tex
.level
,
1965 return &surface
->base
;
1968 static struct pipe_surface
*r600_create_surface(struct pipe_context
*pipe
,
1969 struct pipe_resource
*tex
,
1970 const struct pipe_surface
*templ
)
1972 unsigned level
= templ
->u
.tex
.level
;
1973 unsigned width
= u_minify(tex
->width0
, level
);
1974 unsigned height
= u_minify(tex
->height0
, level
);
1975 unsigned width0
= tex
->width0
;
1976 unsigned height0
= tex
->height0
;
1978 if (tex
->target
!= PIPE_BUFFER
&& templ
->format
!= tex
->format
) {
1979 const struct util_format_description
*tex_desc
1980 = util_format_description(tex
->format
);
1981 const struct util_format_description
*templ_desc
1982 = util_format_description(templ
->format
);
1984 assert(tex_desc
->block
.bits
== templ_desc
->block
.bits
);
1986 /* Adjust size of surface if and only if the block width or
1987 * height is changed. */
1988 if (tex_desc
->block
.width
!= templ_desc
->block
.width
||
1989 tex_desc
->block
.height
!= templ_desc
->block
.height
) {
1990 unsigned nblks_x
= util_format_get_nblocksx(tex
->format
, width
);
1991 unsigned nblks_y
= util_format_get_nblocksy(tex
->format
, height
);
1993 width
= nblks_x
* templ_desc
->block
.width
;
1994 height
= nblks_y
* templ_desc
->block
.height
;
1996 width0
= util_format_get_nblocksx(tex
->format
, width0
);
1997 height0
= util_format_get_nblocksy(tex
->format
, height0
);
2001 return r600_create_surface_custom(pipe
, tex
, templ
,
2006 static void r600_surface_destroy(struct pipe_context
*pipe
,
2007 struct pipe_surface
*surface
)
2009 struct r600_surface
*surf
= (struct r600_surface
*)surface
;
2010 r600_resource_reference(&surf
->cb_buffer_fmask
, NULL
);
2011 r600_resource_reference(&surf
->cb_buffer_cmask
, NULL
);
2012 pipe_resource_reference(&surface
->texture
, NULL
);
2016 static void r600_clear_texture(struct pipe_context
*pipe
,
2017 struct pipe_resource
*tex
,
2019 const struct pipe_box
*box
,
2022 struct pipe_screen
*screen
= pipe
->screen
;
2023 struct r600_texture
*rtex
= (struct r600_texture
*)tex
;
2024 struct pipe_surface tmpl
= {{0}};
2025 struct pipe_surface
*sf
;
2026 const struct util_format_description
*desc
=
2027 util_format_description(tex
->format
);
2029 tmpl
.format
= tex
->format
;
2030 tmpl
.u
.tex
.first_layer
= box
->z
;
2031 tmpl
.u
.tex
.last_layer
= box
->z
+ box
->depth
- 1;
2032 tmpl
.u
.tex
.level
= level
;
2033 sf
= pipe
->create_surface(pipe
, tex
, &tmpl
);
2037 if (rtex
->is_depth
) {
2040 uint8_t stencil
= 0;
2042 /* Depth is always present. */
2043 clear
= PIPE_CLEAR_DEPTH
;
2044 desc
->unpack_z_float(&depth
, 0, data
, 0, 1, 1);
2046 if (rtex
->surface
.flags
& RADEON_SURF_SBUFFER
) {
2047 clear
|= PIPE_CLEAR_STENCIL
;
2048 desc
->unpack_s_8uint(&stencil
, 0, data
, 0, 1, 1);
2051 pipe
->clear_depth_stencil(pipe
, sf
, clear
, depth
, stencil
,
2053 box
->width
, box
->height
, false);
2055 union pipe_color_union color
;
2057 /* pipe_color_union requires the full vec4 representation. */
2058 if (util_format_is_pure_uint(tex
->format
))
2059 desc
->unpack_rgba_uint(color
.ui
, 0, data
, 0, 1, 1);
2060 else if (util_format_is_pure_sint(tex
->format
))
2061 desc
->unpack_rgba_sint(color
.i
, 0, data
, 0, 1, 1);
2063 desc
->unpack_rgba_float(color
.f
, 0, data
, 0, 1, 1);
2065 if (screen
->is_format_supported(screen
, tex
->format
,
2067 PIPE_BIND_RENDER_TARGET
)) {
2068 pipe
->clear_render_target(pipe
, sf
, &color
,
2070 box
->width
, box
->height
, false);
2072 /* Software fallback - just for R9G9B9E5_FLOAT */
2073 util_clear_render_target(pipe
, sf
, &color
,
2075 box
->width
, box
->height
);
2078 pipe_surface_reference(&sf
, NULL
);
2081 unsigned r600_translate_colorswap(enum pipe_format format
, bool do_endian_swap
)
2083 const struct util_format_description
*desc
= util_format_description(format
);
2085 #define HAS_SWIZZLE(chan,swz) (desc->swizzle[chan] == PIPE_SWIZZLE_##swz)
2087 if (format
== PIPE_FORMAT_R11G11B10_FLOAT
) /* isn't plain */
2088 return V_0280A0_SWAP_STD
;
2090 if (desc
->layout
!= UTIL_FORMAT_LAYOUT_PLAIN
)
2093 switch (desc
->nr_channels
) {
2095 if (HAS_SWIZZLE(0,X
))
2096 return V_0280A0_SWAP_STD
; /* X___ */
2097 else if (HAS_SWIZZLE(3,X
))
2098 return V_0280A0_SWAP_ALT_REV
; /* ___X */
2101 if ((HAS_SWIZZLE(0,X
) && HAS_SWIZZLE(1,Y
)) ||
2102 (HAS_SWIZZLE(0,X
) && HAS_SWIZZLE(1,NONE
)) ||
2103 (HAS_SWIZZLE(0,NONE
) && HAS_SWIZZLE(1,Y
)))
2104 return V_0280A0_SWAP_STD
; /* XY__ */
2105 else if ((HAS_SWIZZLE(0,Y
) && HAS_SWIZZLE(1,X
)) ||
2106 (HAS_SWIZZLE(0,Y
) && HAS_SWIZZLE(1,NONE
)) ||
2107 (HAS_SWIZZLE(0,NONE
) && HAS_SWIZZLE(1,X
)))
2109 return (do_endian_swap
? V_0280A0_SWAP_STD
: V_0280A0_SWAP_STD_REV
);
2110 else if (HAS_SWIZZLE(0,X
) && HAS_SWIZZLE(3,Y
))
2111 return V_0280A0_SWAP_ALT
; /* X__Y */
2112 else if (HAS_SWIZZLE(0,Y
) && HAS_SWIZZLE(3,X
))
2113 return V_0280A0_SWAP_ALT_REV
; /* Y__X */
2116 if (HAS_SWIZZLE(0,X
))
2117 return (do_endian_swap
? V_0280A0_SWAP_STD_REV
: V_0280A0_SWAP_STD
);
2118 else if (HAS_SWIZZLE(0,Z
))
2119 return V_0280A0_SWAP_STD_REV
; /* ZYX */
2122 /* check the middle channels, the 1st and 4th channel can be NONE */
2123 if (HAS_SWIZZLE(1,Y
) && HAS_SWIZZLE(2,Z
)) {
2124 return V_0280A0_SWAP_STD
; /* XYZW */
2125 } else if (HAS_SWIZZLE(1,Z
) && HAS_SWIZZLE(2,Y
)) {
2126 return V_0280A0_SWAP_STD_REV
; /* WZYX */
2127 } else if (HAS_SWIZZLE(1,Y
) && HAS_SWIZZLE(2,X
)) {
2128 return V_0280A0_SWAP_ALT
; /* ZYXW */
2129 } else if (HAS_SWIZZLE(1,Z
) && HAS_SWIZZLE(2,W
)) {
2132 return V_0280A0_SWAP_ALT_REV
;
2134 return (do_endian_swap
? V_0280A0_SWAP_ALT
: V_0280A0_SWAP_ALT_REV
);
2141 /* PIPELINE_STAT-BASED DCC ENABLEMENT FOR DISPLAYABLE SURFACES */
2143 static void vi_dcc_clean_up_context_slot(struct r600_common_context
*rctx
,
2148 if (rctx
->dcc_stats
[slot
].query_active
)
2149 vi_separate_dcc_stop_query(&rctx
->b
,
2150 rctx
->dcc_stats
[slot
].tex
);
2152 for (i
= 0; i
< ARRAY_SIZE(rctx
->dcc_stats
[slot
].ps_stats
); i
++)
2153 if (rctx
->dcc_stats
[slot
].ps_stats
[i
]) {
2154 rctx
->b
.destroy_query(&rctx
->b
,
2155 rctx
->dcc_stats
[slot
].ps_stats
[i
]);
2156 rctx
->dcc_stats
[slot
].ps_stats
[i
] = NULL
;
2159 r600_texture_reference(&rctx
->dcc_stats
[slot
].tex
, NULL
);
2163 * Return the per-context slot where DCC statistics queries for the texture live.
2165 static unsigned vi_get_context_dcc_stats_index(struct r600_common_context
*rctx
,
2166 struct r600_texture
*tex
)
2168 int i
, empty_slot
= -1;
2170 /* Remove zombie textures (textures kept alive by this array only). */
2171 for (i
= 0; i
< ARRAY_SIZE(rctx
->dcc_stats
); i
++)
2172 if (rctx
->dcc_stats
[i
].tex
&&
2173 rctx
->dcc_stats
[i
].tex
->resource
.b
.b
.reference
.count
== 1)
2174 vi_dcc_clean_up_context_slot(rctx
, i
);
2176 /* Find the texture. */
2177 for (i
= 0; i
< ARRAY_SIZE(rctx
->dcc_stats
); i
++) {
2178 /* Return if found. */
2179 if (rctx
->dcc_stats
[i
].tex
== tex
) {
2180 rctx
->dcc_stats
[i
].last_use_timestamp
= os_time_get();
2184 /* Record the first seen empty slot. */
2185 if (empty_slot
== -1 && !rctx
->dcc_stats
[i
].tex
)
2189 /* Not found. Remove the oldest member to make space in the array. */
2190 if (empty_slot
== -1) {
2191 int oldest_slot
= 0;
2193 /* Find the oldest slot. */
2194 for (i
= 1; i
< ARRAY_SIZE(rctx
->dcc_stats
); i
++)
2195 if (rctx
->dcc_stats
[oldest_slot
].last_use_timestamp
>
2196 rctx
->dcc_stats
[i
].last_use_timestamp
)
2199 /* Clean up the oldest slot. */
2200 vi_dcc_clean_up_context_slot(rctx
, oldest_slot
);
2201 empty_slot
= oldest_slot
;
2204 /* Add the texture to the new slot. */
2205 r600_texture_reference(&rctx
->dcc_stats
[empty_slot
].tex
, tex
);
2206 rctx
->dcc_stats
[empty_slot
].last_use_timestamp
= os_time_get();
2210 static struct pipe_query
*
2211 vi_create_resuming_pipestats_query(struct pipe_context
*ctx
)
2213 struct r600_query_hw
*query
= (struct r600_query_hw
*)
2214 ctx
->create_query(ctx
, PIPE_QUERY_PIPELINE_STATISTICS
, 0);
2216 query
->flags
|= R600_QUERY_HW_FLAG_BEGIN_RESUMES
;
2217 return (struct pipe_query
*)query
;
2221 * Called when binding a color buffer.
2223 void vi_separate_dcc_start_query(struct pipe_context
*ctx
,
2224 struct r600_texture
*tex
)
2226 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
2227 unsigned i
= vi_get_context_dcc_stats_index(rctx
, tex
);
2229 assert(!rctx
->dcc_stats
[i
].query_active
);
2231 if (!rctx
->dcc_stats
[i
].ps_stats
[0])
2232 rctx
->dcc_stats
[i
].ps_stats
[0] = vi_create_resuming_pipestats_query(ctx
);
2234 /* begin or resume the query */
2235 ctx
->begin_query(ctx
, rctx
->dcc_stats
[i
].ps_stats
[0]);
2236 rctx
->dcc_stats
[i
].query_active
= true;
2240 * Called when unbinding a color buffer.
2242 void vi_separate_dcc_stop_query(struct pipe_context
*ctx
,
2243 struct r600_texture
*tex
)
2245 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
2246 unsigned i
= vi_get_context_dcc_stats_index(rctx
, tex
);
2248 assert(rctx
->dcc_stats
[i
].query_active
);
2249 assert(rctx
->dcc_stats
[i
].ps_stats
[0]);
2251 /* pause or end the query */
2252 ctx
->end_query(ctx
, rctx
->dcc_stats
[i
].ps_stats
[0]);
2253 rctx
->dcc_stats
[i
].query_active
= false;
2256 static bool vi_should_enable_separate_dcc(struct r600_texture
*tex
)
2258 /* The minimum number of fullscreen draws per frame that is required
2260 return tex
->ps_draw_ratio
+ tex
->num_slow_clears
>= 5;
2263 /* Called by fast clear. */
2264 static void vi_separate_dcc_try_enable(struct r600_common_context
*rctx
,
2265 struct r600_texture
*tex
)
2267 /* The intent is to use this with shared displayable back buffers,
2268 * but it's not strictly limited only to them.
2270 if (!tex
->resource
.b
.is_shared
||
2271 !(tex
->resource
.external_usage
& PIPE_HANDLE_USAGE_EXPLICIT_FLUSH
) ||
2272 tex
->resource
.b
.b
.target
!= PIPE_TEXTURE_2D
||
2273 tex
->resource
.b
.b
.last_level
> 0 ||
2274 !tex
->surface
.dcc_size
)
2277 if (tex
->dcc_offset
)
2278 return; /* already enabled */
2280 /* Enable the DCC stat gathering. */
2281 if (!tex
->dcc_gather_statistics
) {
2282 tex
->dcc_gather_statistics
= true;
2283 vi_separate_dcc_start_query(&rctx
->b
, tex
);
2286 if (!vi_should_enable_separate_dcc(tex
))
2287 return; /* stats show that DCC decompression is too expensive */
2289 assert(tex
->surface
.num_dcc_levels
);
2290 assert(!tex
->dcc_separate_buffer
);
2292 r600_texture_discard_cmask(rctx
->screen
, tex
);
2294 /* Get a DCC buffer. */
2295 if (tex
->last_dcc_separate_buffer
) {
2296 assert(tex
->dcc_gather_statistics
);
2297 assert(!tex
->dcc_separate_buffer
);
2298 tex
->dcc_separate_buffer
= tex
->last_dcc_separate_buffer
;
2299 tex
->last_dcc_separate_buffer
= NULL
;
2301 tex
->dcc_separate_buffer
= (struct r600_resource
*)
2302 r600_aligned_buffer_create(rctx
->b
.screen
,
2303 R600_RESOURCE_FLAG_UNMAPPABLE
,
2305 tex
->surface
.dcc_size
,
2306 tex
->surface
.dcc_alignment
);
2307 if (!tex
->dcc_separate_buffer
)
2311 /* dcc_offset is the absolute GPUVM address. */
2312 tex
->dcc_offset
= tex
->dcc_separate_buffer
->gpu_address
;
2314 /* no need to flag anything since this is called by fast clear that
2315 * flags framebuffer state
2320 * Called by pipe_context::flush_resource, the place where DCC decompression
2323 void vi_separate_dcc_process_and_reset_stats(struct pipe_context
*ctx
,
2324 struct r600_texture
*tex
)
2326 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
2327 struct pipe_query
*tmp
;
2328 unsigned i
= vi_get_context_dcc_stats_index(rctx
, tex
);
2329 bool query_active
= rctx
->dcc_stats
[i
].query_active
;
2330 bool disable
= false;
2332 if (rctx
->dcc_stats
[i
].ps_stats
[2]) {
2333 union pipe_query_result result
;
2335 /* Read the results. */
2336 ctx
->get_query_result(ctx
, rctx
->dcc_stats
[i
].ps_stats
[2],
2338 r600_query_hw_reset_buffers(rctx
,
2339 (struct r600_query_hw
*)
2340 rctx
->dcc_stats
[i
].ps_stats
[2]);
2342 /* Compute the approximate number of fullscreen draws. */
2343 tex
->ps_draw_ratio
=
2344 result
.pipeline_statistics
.ps_invocations
/
2345 (tex
->resource
.b
.b
.width0
* tex
->resource
.b
.b
.height0
);
2346 rctx
->last_tex_ps_draw_ratio
= tex
->ps_draw_ratio
;
2348 disable
= tex
->dcc_separate_buffer
&&
2349 !vi_should_enable_separate_dcc(tex
);
2352 tex
->num_slow_clears
= 0;
2354 /* stop the statistics query for ps_stats[0] */
2356 vi_separate_dcc_stop_query(ctx
, tex
);
2358 /* Move the queries in the queue by one. */
2359 tmp
= rctx
->dcc_stats
[i
].ps_stats
[2];
2360 rctx
->dcc_stats
[i
].ps_stats
[2] = rctx
->dcc_stats
[i
].ps_stats
[1];
2361 rctx
->dcc_stats
[i
].ps_stats
[1] = rctx
->dcc_stats
[i
].ps_stats
[0];
2362 rctx
->dcc_stats
[i
].ps_stats
[0] = tmp
;
2364 /* create and start a new query as ps_stats[0] */
2366 vi_separate_dcc_start_query(ctx
, tex
);
2369 assert(!tex
->last_dcc_separate_buffer
);
2370 tex
->last_dcc_separate_buffer
= tex
->dcc_separate_buffer
;
2371 tex
->dcc_separate_buffer
= NULL
;
2372 tex
->dcc_offset
= 0;
2373 /* no need to flag anything since this is called after
2374 * decompression that re-sets framebuffer state
2379 /* FAST COLOR CLEAR */
2381 static void evergreen_set_clear_color(struct r600_texture
*rtex
,
2382 enum pipe_format surface_format
,
2383 const union pipe_color_union
*color
)
2385 union util_color uc
;
2387 memset(&uc
, 0, sizeof(uc
));
2389 if (rtex
->surface
.bpe
== 16) {
2390 /* DCC fast clear only:
2391 * CLEAR_WORD0 = R = G = B
2394 assert(color
->ui
[0] == color
->ui
[1] &&
2395 color
->ui
[0] == color
->ui
[2]);
2396 uc
.ui
[0] = color
->ui
[0];
2397 uc
.ui
[1] = color
->ui
[3];
2398 } else if (util_format_is_pure_uint(surface_format
)) {
2399 util_format_write_4ui(surface_format
, color
->ui
, 0, &uc
, 0, 0, 0, 1, 1);
2400 } else if (util_format_is_pure_sint(surface_format
)) {
2401 util_format_write_4i(surface_format
, color
->i
, 0, &uc
, 0, 0, 0, 1, 1);
2403 util_pack_color(color
->f
, surface_format
, &uc
);
2406 memcpy(rtex
->color_clear_value
, &uc
, 2 * sizeof(uint32_t));
2409 static bool vi_get_fast_clear_parameters(enum pipe_format surface_format
,
2410 const union pipe_color_union
*color
,
2411 uint32_t* reset_value
,
2412 bool* clear_words_needed
)
2414 bool values
[4] = {};
2416 bool main_value
= false;
2417 bool extra_value
= false;
2420 /* This is needed to get the correct DCC clear value for luminance formats.
2421 * 1) Get the linear format (because the next step can't handle L8_SRGB).
2422 * 2) Convert luminance to red. (the real hw format for luminance)
2424 surface_format
= util_format_linear(surface_format
);
2425 surface_format
= util_format_luminance_to_red(surface_format
);
2427 const struct util_format_description
*desc
= util_format_description(surface_format
);
2429 if (desc
->block
.bits
== 128 &&
2430 (color
->ui
[0] != color
->ui
[1] ||
2431 color
->ui
[0] != color
->ui
[2]))
2434 *clear_words_needed
= true;
2435 *reset_value
= 0x20202020U
;
2437 /* If we want to clear without needing a fast clear eliminate step, we
2438 * can set each channel to 0 or 1 (or 0/max for integer formats). We
2439 * have two sets of flags, one for the last or first channel(extra) and
2440 * one for the other channels(main).
2443 if (surface_format
== PIPE_FORMAT_R11G11B10_FLOAT
||
2444 surface_format
== PIPE_FORMAT_B5G6R5_UNORM
||
2445 surface_format
== PIPE_FORMAT_B5G6R5_SRGB
||
2446 util_format_is_alpha(surface_format
)) {
2448 } else if (desc
->layout
== UTIL_FORMAT_LAYOUT_PLAIN
) {
2449 if(r600_translate_colorswap(surface_format
, false) <= 1)
2450 extra_channel
= desc
->nr_channels
- 1;
2456 for (i
= 0; i
< 4; ++i
) {
2457 int index
= desc
->swizzle
[i
] - PIPE_SWIZZLE_X
;
2459 if (desc
->swizzle
[i
] < PIPE_SWIZZLE_X
||
2460 desc
->swizzle
[i
] > PIPE_SWIZZLE_W
)
2463 if (desc
->channel
[i
].pure_integer
&&
2464 desc
->channel
[i
].type
== UTIL_FORMAT_TYPE_SIGNED
) {
2465 /* Use the maximum value for clamping the clear color. */
2466 int max
= u_bit_consecutive(0, desc
->channel
[i
].size
- 1);
2468 values
[i
] = color
->i
[i
] != 0;
2469 if (color
->i
[i
] != 0 && MIN2(color
->i
[i
], max
) != max
)
2471 } else if (desc
->channel
[i
].pure_integer
&&
2472 desc
->channel
[i
].type
== UTIL_FORMAT_TYPE_UNSIGNED
) {
2473 /* Use the maximum value for clamping the clear color. */
2474 unsigned max
= u_bit_consecutive(0, desc
->channel
[i
].size
);
2476 values
[i
] = color
->ui
[i
] != 0U;
2477 if (color
->ui
[i
] != 0U && MIN2(color
->ui
[i
], max
) != max
)
2480 values
[i
] = color
->f
[i
] != 0.0F
;
2481 if (color
->f
[i
] != 0.0F
&& color
->f
[i
] != 1.0F
)
2485 if (index
== extra_channel
)
2486 extra_value
= values
[i
];
2488 main_value
= values
[i
];
2491 for (int i
= 0; i
< 4; ++i
)
2492 if (values
[i
] != main_value
&&
2493 desc
->swizzle
[i
] - PIPE_SWIZZLE_X
!= extra_channel
&&
2494 desc
->swizzle
[i
] >= PIPE_SWIZZLE_X
&&
2495 desc
->swizzle
[i
] <= PIPE_SWIZZLE_W
)
2498 *clear_words_needed
= false;
2500 *reset_value
|= 0x80808080U
;
2503 *reset_value
|= 0x40404040U
;
2507 void vi_dcc_clear_level(struct r600_common_context
*rctx
,
2508 struct r600_texture
*rtex
,
2509 unsigned level
, unsigned clear_value
)
2511 struct pipe_resource
*dcc_buffer
;
2512 uint64_t dcc_offset
, clear_size
;
2514 assert(vi_dcc_enabled(rtex
, level
));
2516 if (rtex
->dcc_separate_buffer
) {
2517 dcc_buffer
= &rtex
->dcc_separate_buffer
->b
.b
;
2520 dcc_buffer
= &rtex
->resource
.b
.b
;
2521 dcc_offset
= rtex
->dcc_offset
;
2524 if (rctx
->chip_class
>= GFX9
) {
2525 /* Mipmap level clears aren't implemented. */
2526 assert(rtex
->resource
.b
.b
.last_level
== 0);
2527 /* MSAA needs a different clear size. */
2528 assert(rtex
->resource
.b
.b
.nr_samples
<= 1);
2529 clear_size
= rtex
->surface
.dcc_size
;
2531 dcc_offset
+= rtex
->surface
.u
.legacy
.level
[level
].dcc_offset
;
2532 clear_size
= rtex
->surface
.u
.legacy
.level
[level
].dcc_fast_clear_size
;
2535 rctx
->clear_buffer(&rctx
->b
, dcc_buffer
, dcc_offset
, clear_size
,
2536 clear_value
, R600_COHERENCY_CB_META
);
2539 /* Set the same micro tile mode as the destination of the last MSAA resolve.
2540 * This allows hitting the MSAA resolve fast path, which requires that both
2541 * src and dst micro tile modes match.
2543 static void si_set_optimal_micro_tile_mode(struct r600_common_screen
*rscreen
,
2544 struct r600_texture
*rtex
)
2546 if (rtex
->resource
.b
.is_shared
||
2547 rtex
->resource
.b
.b
.nr_samples
<= 1 ||
2548 rtex
->surface
.micro_tile_mode
== rtex
->last_msaa_resolve_target_micro_mode
)
2551 assert(rscreen
->chip_class
>= GFX9
||
2552 rtex
->surface
.u
.legacy
.level
[0].mode
== RADEON_SURF_MODE_2D
);
2553 assert(rtex
->resource
.b
.b
.last_level
== 0);
2555 if (rscreen
->chip_class
>= GFX9
) {
2556 /* 4K or larger tiles only. 0 is linear. 1-3 are 256B tiles. */
2557 assert(rtex
->surface
.u
.gfx9
.surf
.swizzle_mode
>= 4);
2559 /* If you do swizzle_mode % 4, you'll get:
2565 * Depth-sample order isn't allowed:
2567 assert(rtex
->surface
.u
.gfx9
.surf
.swizzle_mode
% 4 != 0);
2569 switch (rtex
->last_msaa_resolve_target_micro_mode
) {
2570 case RADEON_MICRO_MODE_DISPLAY
:
2571 rtex
->surface
.u
.gfx9
.surf
.swizzle_mode
&= ~0x3;
2572 rtex
->surface
.u
.gfx9
.surf
.swizzle_mode
+= 2; /* D */
2574 case RADEON_MICRO_MODE_THIN
:
2575 rtex
->surface
.u
.gfx9
.surf
.swizzle_mode
&= ~0x3;
2576 rtex
->surface
.u
.gfx9
.surf
.swizzle_mode
+= 1; /* S */
2578 case RADEON_MICRO_MODE_ROTATED
:
2579 rtex
->surface
.u
.gfx9
.surf
.swizzle_mode
&= ~0x3;
2580 rtex
->surface
.u
.gfx9
.surf
.swizzle_mode
+= 3; /* R */
2582 default: /* depth */
2583 assert(!"unexpected micro mode");
2586 } else if (rscreen
->chip_class
>= CIK
) {
2587 /* These magic numbers were copied from addrlib. It doesn't use
2588 * any definitions for them either. They are all 2D_TILED_THIN1
2589 * modes with different bpp and micro tile mode.
2591 switch (rtex
->last_msaa_resolve_target_micro_mode
) {
2592 case RADEON_MICRO_MODE_DISPLAY
:
2593 rtex
->surface
.u
.legacy
.tiling_index
[0] = 10;
2595 case RADEON_MICRO_MODE_THIN
:
2596 rtex
->surface
.u
.legacy
.tiling_index
[0] = 14;
2598 case RADEON_MICRO_MODE_ROTATED
:
2599 rtex
->surface
.u
.legacy
.tiling_index
[0] = 28;
2601 default: /* depth, thick */
2602 assert(!"unexpected micro mode");
2606 switch (rtex
->last_msaa_resolve_target_micro_mode
) {
2607 case RADEON_MICRO_MODE_DISPLAY
:
2608 switch (rtex
->surface
.bpe
) {
2610 rtex
->surface
.u
.legacy
.tiling_index
[0] = 10;
2613 rtex
->surface
.u
.legacy
.tiling_index
[0] = 11;
2616 rtex
->surface
.u
.legacy
.tiling_index
[0] = 12;
2620 case RADEON_MICRO_MODE_THIN
:
2621 switch (rtex
->surface
.bpe
) {
2623 rtex
->surface
.u
.legacy
.tiling_index
[0] = 14;
2626 rtex
->surface
.u
.legacy
.tiling_index
[0] = 15;
2629 rtex
->surface
.u
.legacy
.tiling_index
[0] = 16;
2631 default: /* 8, 16 */
2632 rtex
->surface
.u
.legacy
.tiling_index
[0] = 17;
2636 default: /* depth, thick */
2637 assert(!"unexpected micro mode");
2642 rtex
->surface
.micro_tile_mode
= rtex
->last_msaa_resolve_target_micro_mode
;
2644 p_atomic_inc(&rscreen
->dirty_tex_counter
);
2647 void evergreen_do_fast_color_clear(struct r600_common_context
*rctx
,
2648 struct pipe_framebuffer_state
*fb
,
2649 struct r600_atom
*fb_state
,
2650 unsigned *buffers
, ubyte
*dirty_cbufs
,
2651 const union pipe_color_union
*color
)
2655 /* This function is broken in BE, so just disable this path for now */
2656 #ifdef PIPE_ARCH_BIG_ENDIAN
2660 if (rctx
->render_cond
)
2663 for (i
= 0; i
< fb
->nr_cbufs
; i
++) {
2664 struct r600_texture
*tex
;
2665 unsigned clear_bit
= PIPE_CLEAR_COLOR0
<< i
;
2670 /* if this colorbuffer is not being cleared */
2671 if (!(*buffers
& clear_bit
))
2674 tex
= (struct r600_texture
*)fb
->cbufs
[i
]->texture
;
2676 /* the clear is allowed if all layers are bound */
2677 if (fb
->cbufs
[i
]->u
.tex
.first_layer
!= 0 ||
2678 fb
->cbufs
[i
]->u
.tex
.last_layer
!= util_max_layer(&tex
->resource
.b
.b
, 0)) {
2682 /* cannot clear mipmapped textures */
2683 if (fb
->cbufs
[i
]->texture
->last_level
!= 0) {
2687 /* only supported on tiled surfaces */
2688 if (tex
->surface
.is_linear
) {
2692 /* shared textures can't use fast clear without an explicit flush,
2693 * because there is no way to communicate the clear color among
2696 if (tex
->resource
.b
.is_shared
&&
2697 !(tex
->resource
.external_usage
& PIPE_HANDLE_USAGE_EXPLICIT_FLUSH
))
2700 /* fast color clear with 1D tiling doesn't work on old kernels and CIK */
2701 if (rctx
->chip_class
== CIK
&&
2702 tex
->surface
.u
.legacy
.level
[0].mode
== RADEON_SURF_MODE_1D
&&
2703 rctx
->screen
->info
.drm_major
== 2 &&
2704 rctx
->screen
->info
.drm_minor
< 38) {
2708 /* Fast clear is the most appropriate place to enable DCC for
2709 * displayable surfaces.
2711 if (rctx
->chip_class
>= VI
&&
2712 !(rctx
->screen
->debug_flags
& DBG_NO_DCC_FB
)) {
2713 vi_separate_dcc_try_enable(rctx
, tex
);
2715 /* RB+ isn't supported with a CMASK clear only on Stoney,
2716 * so all clears are considered to be hypothetically slow
2717 * clears, which is weighed when determining whether to
2718 * enable separate DCC.
2720 if (tex
->dcc_gather_statistics
&&
2721 rctx
->family
== CHIP_STONEY
)
2722 tex
->num_slow_clears
++;
2725 /* Try to clear DCC first, otherwise try CMASK. */
2726 if (vi_dcc_enabled(tex
, 0)) {
2727 uint32_t reset_value
;
2728 bool clear_words_needed
;
2730 if (rctx
->screen
->debug_flags
& DBG_NO_DCC_CLEAR
)
2733 if (!vi_get_fast_clear_parameters(fb
->cbufs
[i
]->format
,
2734 color
, &reset_value
,
2735 &clear_words_needed
))
2738 vi_dcc_clear_level(rctx
, tex
, 0, reset_value
);
2740 unsigned level_bit
= 1 << fb
->cbufs
[i
]->u
.tex
.level
;
2741 if (clear_words_needed
) {
2742 bool need_compressed_update
= !tex
->dirty_level_mask
;
2744 tex
->dirty_level_mask
|= level_bit
;
2746 if (need_compressed_update
)
2747 p_atomic_inc(&rctx
->screen
->compressed_colortex_counter
);
2749 tex
->separate_dcc_dirty
= true;
2751 /* 128-bit formats are unusupported */
2752 if (tex
->surface
.bpe
> 8) {
2756 /* RB+ doesn't work with CMASK fast clear on Stoney. */
2757 if (rctx
->family
== CHIP_STONEY
)
2760 /* ensure CMASK is enabled */
2761 r600_texture_alloc_cmask_separate(rctx
->screen
, tex
);
2762 if (tex
->cmask
.size
== 0) {
2766 /* Do the fast clear. */
2767 rctx
->clear_buffer(&rctx
->b
, &tex
->cmask_buffer
->b
.b
,
2768 tex
->cmask
.offset
, tex
->cmask
.size
, 0,
2769 R600_COHERENCY_CB_META
);
2771 bool need_compressed_update
= !tex
->dirty_level_mask
;
2773 tex
->dirty_level_mask
|= 1 << fb
->cbufs
[i
]->u
.tex
.level
;
2775 if (need_compressed_update
)
2776 p_atomic_inc(&rctx
->screen
->compressed_colortex_counter
);
2779 /* We can change the micro tile mode before a full clear. */
2780 if (rctx
->screen
->chip_class
>= SI
)
2781 si_set_optimal_micro_tile_mode(rctx
->screen
, tex
);
2783 evergreen_set_clear_color(tex
, fb
->cbufs
[i
]->format
, color
);
2786 *dirty_cbufs
|= 1 << i
;
2787 rctx
->set_atom_dirty(rctx
, fb_state
, true);
2788 *buffers
&= ~clear_bit
;
2792 void r600_init_screen_texture_functions(struct r600_common_screen
*rscreen
)
2794 rscreen
->b
.resource_from_handle
= r600_texture_from_handle
;
2795 rscreen
->b
.resource_get_handle
= r600_texture_get_handle
;
2798 void r600_init_context_texture_functions(struct r600_common_context
*rctx
)
2800 rctx
->b
.create_surface
= r600_create_surface
;
2801 rctx
->b
.surface_destroy
= r600_surface_destroy
;
2802 rctx
->b
.clear_texture
= r600_clear_texture
;