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_log.h"
32 #include "util/u_memory.h"
33 #include "util/u_pack_color.h"
34 #include "util/u_surface.h"
35 #include "util/os_time.h"
39 static void r600_texture_discard_cmask(struct r600_common_screen
*rscreen
,
40 struct r600_texture
*rtex
);
41 static enum radeon_surf_mode
42 r600_choose_tiling(struct r600_common_screen
*rscreen
,
43 const struct pipe_resource
*templ
);
46 bool r600_prepare_for_dma_blit(struct r600_common_context
*rctx
,
47 struct r600_texture
*rdst
,
48 unsigned dst_level
, unsigned dstx
,
49 unsigned dsty
, unsigned dstz
,
50 struct r600_texture
*rsrc
,
52 const struct pipe_box
*src_box
)
57 if (rdst
->surface
.bpe
!= rsrc
->surface
.bpe
)
60 /* MSAA: Blits don't exist in the real world. */
61 if (rsrc
->resource
.b
.b
.nr_samples
> 1 ||
62 rdst
->resource
.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 (rsrc
->is_depth
|| rdst
->is_depth
)
73 * src: Both texture and SDMA paths need decompression. Use SDMA.
74 * dst: If overwriting the whole texture, discard CMASK and use
75 * SDMA. Otherwise, use the 3D path.
77 if (rdst
->cmask
.size
&& rdst
->dirty_level_mask
& (1 << dst_level
)) {
78 /* The CMASK clear is only enabled for the first level. */
79 assert(dst_level
== 0);
80 if (!util_texrange_covers_whole_level(&rdst
->resource
.b
.b
, dst_level
,
81 dstx
, dsty
, dstz
, src_box
->width
,
82 src_box
->height
, src_box
->depth
))
85 r600_texture_discard_cmask(rctx
->screen
, rdst
);
88 /* All requirements are met. Prepare textures for SDMA. */
89 if (rsrc
->cmask
.size
&& rsrc
->dirty_level_mask
& (1 << src_level
))
90 rctx
->b
.flush_resource(&rctx
->b
, &rsrc
->resource
.b
.b
);
92 assert(!(rsrc
->dirty_level_mask
& (1 << src_level
)));
93 assert(!(rdst
->dirty_level_mask
& (1 << dst_level
)));
98 /* Same as resource_copy_region, except that both upsampling and downsampling are allowed. */
99 static void r600_copy_region_with_blit(struct pipe_context
*pipe
,
100 struct pipe_resource
*dst
,
102 unsigned dstx
, unsigned dsty
, unsigned dstz
,
103 struct pipe_resource
*src
,
105 const struct pipe_box
*src_box
)
107 struct pipe_blit_info blit
;
109 memset(&blit
, 0, sizeof(blit
));
110 blit
.src
.resource
= src
;
111 blit
.src
.format
= src
->format
;
112 blit
.src
.level
= src_level
;
113 blit
.src
.box
= *src_box
;
114 blit
.dst
.resource
= dst
;
115 blit
.dst
.format
= dst
->format
;
116 blit
.dst
.level
= dst_level
;
117 blit
.dst
.box
.x
= dstx
;
118 blit
.dst
.box
.y
= dsty
;
119 blit
.dst
.box
.z
= dstz
;
120 blit
.dst
.box
.width
= src_box
->width
;
121 blit
.dst
.box
.height
= src_box
->height
;
122 blit
.dst
.box
.depth
= src_box
->depth
;
123 blit
.mask
= util_format_get_mask(src
->format
) &
124 util_format_get_mask(dst
->format
);
125 blit
.filter
= PIPE_TEX_FILTER_NEAREST
;
128 pipe
->blit(pipe
, &blit
);
132 /* Copy from a full GPU texture to a transfer's staging one. */
133 static void r600_copy_to_staging_texture(struct pipe_context
*ctx
, struct r600_transfer
*rtransfer
)
135 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
136 struct pipe_transfer
*transfer
= (struct pipe_transfer
*)rtransfer
;
137 struct pipe_resource
*dst
= &rtransfer
->staging
->b
.b
;
138 struct pipe_resource
*src
= transfer
->resource
;
140 if (src
->nr_samples
> 1) {
141 r600_copy_region_with_blit(ctx
, dst
, 0, 0, 0, 0,
142 src
, transfer
->level
, &transfer
->box
);
146 rctx
->dma_copy(ctx
, dst
, 0, 0, 0, 0, src
, transfer
->level
,
150 /* Copy from a transfer's staging texture to a full GPU one. */
151 static void r600_copy_from_staging_texture(struct pipe_context
*ctx
, struct r600_transfer
*rtransfer
)
153 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
154 struct pipe_transfer
*transfer
= (struct pipe_transfer
*)rtransfer
;
155 struct pipe_resource
*dst
= transfer
->resource
;
156 struct pipe_resource
*src
= &rtransfer
->staging
->b
.b
;
157 struct pipe_box sbox
;
159 u_box_3d(0, 0, 0, transfer
->box
.width
, transfer
->box
.height
, transfer
->box
.depth
, &sbox
);
161 if (dst
->nr_samples
> 1) {
162 r600_copy_region_with_blit(ctx
, dst
, transfer
->level
,
163 transfer
->box
.x
, transfer
->box
.y
, transfer
->box
.z
,
168 rctx
->dma_copy(ctx
, dst
, transfer
->level
,
169 transfer
->box
.x
, transfer
->box
.y
, transfer
->box
.z
,
173 static unsigned r600_texture_get_offset(struct r600_common_screen
*rscreen
,
174 struct r600_texture
*rtex
, unsigned level
,
175 const struct pipe_box
*box
,
177 unsigned *layer_stride
)
179 *stride
= rtex
->surface
.u
.legacy
.level
[level
].nblk_x
*
181 assert((uint64_t)rtex
->surface
.u
.legacy
.level
[level
].slice_size_dw
* 4 <= UINT_MAX
);
182 *layer_stride
= (uint64_t)rtex
->surface
.u
.legacy
.level
[level
].slice_size_dw
* 4;
185 return rtex
->surface
.u
.legacy
.level
[level
].offset
;
187 /* Each texture is an array of mipmap levels. Each level is
188 * an array of slices. */
189 return rtex
->surface
.u
.legacy
.level
[level
].offset
+
190 box
->z
* (uint64_t)rtex
->surface
.u
.legacy
.level
[level
].slice_size_dw
* 4 +
191 (box
->y
/ rtex
->surface
.blk_h
*
192 rtex
->surface
.u
.legacy
.level
[level
].nblk_x
+
193 box
->x
/ rtex
->surface
.blk_w
) * rtex
->surface
.bpe
;
196 static int r600_init_surface(struct r600_common_screen
*rscreen
,
197 struct radeon_surf
*surface
,
198 const struct pipe_resource
*ptex
,
199 enum radeon_surf_mode array_mode
,
200 unsigned pitch_in_bytes_override
,
204 bool is_flushed_depth
)
206 const struct util_format_description
*desc
=
207 util_format_description(ptex
->format
);
208 bool is_depth
, is_stencil
;
210 unsigned i
, bpe
, flags
= 0;
212 is_depth
= util_format_has_depth(desc
);
213 is_stencil
= util_format_has_stencil(desc
);
215 if (rscreen
->chip_class
>= EVERGREEN
&& !is_flushed_depth
&&
216 ptex
->format
== PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
) {
217 bpe
= 4; /* stencil is allocated separately on evergreen */
219 bpe
= util_format_get_blocksize(ptex
->format
);
220 assert(util_is_power_of_two_or_zero(bpe
));
223 if (!is_flushed_depth
&& is_depth
) {
224 flags
|= RADEON_SURF_ZBUFFER
;
227 flags
|= RADEON_SURF_SBUFFER
;
230 if (ptex
->bind
& PIPE_BIND_SCANOUT
|| is_scanout
) {
231 /* This should catch bugs in gallium users setting incorrect flags. */
232 assert(ptex
->nr_samples
<= 1 &&
233 ptex
->array_size
== 1 &&
235 ptex
->last_level
== 0 &&
236 !(flags
& RADEON_SURF_Z_OR_SBUFFER
));
238 flags
|= RADEON_SURF_SCANOUT
;
241 if (ptex
->bind
& PIPE_BIND_SHARED
)
242 flags
|= RADEON_SURF_SHAREABLE
;
244 flags
|= RADEON_SURF_IMPORTED
| RADEON_SURF_SHAREABLE
;
245 if (!(ptex
->flags
& R600_RESOURCE_FLAG_FORCE_TILING
))
246 flags
|= RADEON_SURF_OPTIMIZE_FOR_SPACE
;
248 r
= rscreen
->ws
->surface_init(rscreen
->ws
, ptex
,
249 flags
, bpe
, array_mode
, surface
);
254 if (pitch_in_bytes_override
&&
255 pitch_in_bytes_override
!= surface
->u
.legacy
.level
[0].nblk_x
* bpe
) {
256 /* old ddx on evergreen over estimate alignment for 1d, only 1 level
259 surface
->u
.legacy
.level
[0].nblk_x
= pitch_in_bytes_override
/ bpe
;
260 surface
->u
.legacy
.level
[0].slice_size_dw
=
261 ((uint64_t)pitch_in_bytes_override
* surface
->u
.legacy
.level
[0].nblk_y
) / 4;
265 for (i
= 0; i
< ARRAY_SIZE(surface
->u
.legacy
.level
); ++i
)
266 surface
->u
.legacy
.level
[i
].offset
+= offset
;
272 static void r600_texture_init_metadata(struct r600_common_screen
*rscreen
,
273 struct r600_texture
*rtex
,
274 struct radeon_bo_metadata
*metadata
)
276 struct radeon_surf
*surface
= &rtex
->surface
;
278 memset(metadata
, 0, sizeof(*metadata
));
280 metadata
->u
.legacy
.microtile
= surface
->u
.legacy
.level
[0].mode
>= RADEON_SURF_MODE_1D
?
281 RADEON_LAYOUT_TILED
: RADEON_LAYOUT_LINEAR
;
282 metadata
->u
.legacy
.macrotile
= surface
->u
.legacy
.level
[0].mode
>= RADEON_SURF_MODE_2D
?
283 RADEON_LAYOUT_TILED
: RADEON_LAYOUT_LINEAR
;
284 metadata
->u
.legacy
.pipe_config
= surface
->u
.legacy
.pipe_config
;
285 metadata
->u
.legacy
.bankw
= surface
->u
.legacy
.bankw
;
286 metadata
->u
.legacy
.bankh
= surface
->u
.legacy
.bankh
;
287 metadata
->u
.legacy
.tile_split
= surface
->u
.legacy
.tile_split
;
288 metadata
->u
.legacy
.mtilea
= surface
->u
.legacy
.mtilea
;
289 metadata
->u
.legacy
.num_banks
= surface
->u
.legacy
.num_banks
;
290 metadata
->u
.legacy
.stride
= surface
->u
.legacy
.level
[0].nblk_x
* surface
->bpe
;
291 metadata
->u
.legacy
.scanout
= (surface
->flags
& RADEON_SURF_SCANOUT
) != 0;
294 static void r600_surface_import_metadata(struct r600_common_screen
*rscreen
,
295 struct radeon_surf
*surf
,
296 struct radeon_bo_metadata
*metadata
,
297 enum radeon_surf_mode
*array_mode
,
300 surf
->u
.legacy
.pipe_config
= metadata
->u
.legacy
.pipe_config
;
301 surf
->u
.legacy
.bankw
= metadata
->u
.legacy
.bankw
;
302 surf
->u
.legacy
.bankh
= metadata
->u
.legacy
.bankh
;
303 surf
->u
.legacy
.tile_split
= metadata
->u
.legacy
.tile_split
;
304 surf
->u
.legacy
.mtilea
= metadata
->u
.legacy
.mtilea
;
305 surf
->u
.legacy
.num_banks
= metadata
->u
.legacy
.num_banks
;
307 if (metadata
->u
.legacy
.macrotile
== RADEON_LAYOUT_TILED
)
308 *array_mode
= RADEON_SURF_MODE_2D
;
309 else if (metadata
->u
.legacy
.microtile
== RADEON_LAYOUT_TILED
)
310 *array_mode
= RADEON_SURF_MODE_1D
;
312 *array_mode
= RADEON_SURF_MODE_LINEAR_ALIGNED
;
314 *is_scanout
= metadata
->u
.legacy
.scanout
;
317 static void r600_eliminate_fast_color_clear(struct r600_common_context
*rctx
,
318 struct r600_texture
*rtex
)
320 struct r600_common_screen
*rscreen
= rctx
->screen
;
321 struct pipe_context
*ctx
= &rctx
->b
;
323 if (ctx
== rscreen
->aux_context
)
324 mtx_lock(&rscreen
->aux_context_lock
);
326 ctx
->flush_resource(ctx
, &rtex
->resource
.b
.b
);
327 ctx
->flush(ctx
, NULL
, 0);
329 if (ctx
== rscreen
->aux_context
)
330 mtx_unlock(&rscreen
->aux_context_lock
);
333 static void r600_texture_discard_cmask(struct r600_common_screen
*rscreen
,
334 struct r600_texture
*rtex
)
336 if (!rtex
->cmask
.size
)
339 assert(rtex
->resource
.b
.b
.nr_samples
<= 1);
342 memset(&rtex
->cmask
, 0, sizeof(rtex
->cmask
));
343 rtex
->cmask
.base_address_reg
= rtex
->resource
.gpu_address
>> 8;
344 rtex
->dirty_level_mask
= 0;
346 rtex
->cb_color_info
&= ~EG_S_028C70_FAST_CLEAR(1);
348 if (rtex
->cmask_buffer
!= &rtex
->resource
)
349 r600_resource_reference(&rtex
->cmask_buffer
, NULL
);
351 /* Notify all contexts about the change. */
352 p_atomic_inc(&rscreen
->dirty_tex_counter
);
353 p_atomic_inc(&rscreen
->compressed_colortex_counter
);
356 static void r600_reallocate_texture_inplace(struct r600_common_context
*rctx
,
357 struct r600_texture
*rtex
,
358 unsigned new_bind_flag
,
359 bool invalidate_storage
)
361 struct pipe_screen
*screen
= rctx
->b
.screen
;
362 struct r600_texture
*new_tex
;
363 struct pipe_resource templ
= rtex
->resource
.b
.b
;
366 templ
.bind
|= new_bind_flag
;
368 /* r600g doesn't react to dirty_tex_descriptor_counter */
369 if (rctx
->chip_class
< GFX6
)
372 if (rtex
->resource
.b
.is_shared
)
375 if (new_bind_flag
== PIPE_BIND_LINEAR
) {
376 if (rtex
->surface
.is_linear
)
379 /* This fails with MSAA, depth, and compressed textures. */
380 if (r600_choose_tiling(rctx
->screen
, &templ
) !=
381 RADEON_SURF_MODE_LINEAR_ALIGNED
)
385 new_tex
= (struct r600_texture
*)screen
->resource_create(screen
, &templ
);
389 /* Copy the pixels to the new texture. */
390 if (!invalidate_storage
) {
391 for (i
= 0; i
<= templ
.last_level
; i
++) {
395 u_minify(templ
.width0
, i
), u_minify(templ
.height0
, i
),
396 util_num_layers(&templ
, i
), &box
);
398 rctx
->dma_copy(&rctx
->b
, &new_tex
->resource
.b
.b
, i
, 0, 0, 0,
399 &rtex
->resource
.b
.b
, i
, &box
);
403 if (new_bind_flag
== PIPE_BIND_LINEAR
) {
404 r600_texture_discard_cmask(rctx
->screen
, rtex
);
407 /* Replace the structure fields of rtex. */
408 rtex
->resource
.b
.b
.bind
= templ
.bind
;
409 pb_reference(&rtex
->resource
.buf
, new_tex
->resource
.buf
);
410 rtex
->resource
.gpu_address
= new_tex
->resource
.gpu_address
;
411 rtex
->resource
.vram_usage
= new_tex
->resource
.vram_usage
;
412 rtex
->resource
.gart_usage
= new_tex
->resource
.gart_usage
;
413 rtex
->resource
.bo_size
= new_tex
->resource
.bo_size
;
414 rtex
->resource
.bo_alignment
= new_tex
->resource
.bo_alignment
;
415 rtex
->resource
.domains
= new_tex
->resource
.domains
;
416 rtex
->resource
.flags
= new_tex
->resource
.flags
;
417 rtex
->size
= new_tex
->size
;
418 rtex
->db_render_format
= new_tex
->db_render_format
;
419 rtex
->db_compatible
= new_tex
->db_compatible
;
420 rtex
->can_sample_z
= new_tex
->can_sample_z
;
421 rtex
->can_sample_s
= new_tex
->can_sample_s
;
422 rtex
->surface
= new_tex
->surface
;
423 rtex
->fmask
= new_tex
->fmask
;
424 rtex
->cmask
= new_tex
->cmask
;
425 rtex
->cb_color_info
= new_tex
->cb_color_info
;
426 rtex
->last_msaa_resolve_target_micro_mode
= new_tex
->last_msaa_resolve_target_micro_mode
;
427 rtex
->htile_offset
= new_tex
->htile_offset
;
428 rtex
->depth_cleared
= new_tex
->depth_cleared
;
429 rtex
->stencil_cleared
= new_tex
->stencil_cleared
;
430 rtex
->non_disp_tiling
= new_tex
->non_disp_tiling
;
431 rtex
->framebuffers_bound
= new_tex
->framebuffers_bound
;
433 if (new_bind_flag
== PIPE_BIND_LINEAR
) {
434 assert(!rtex
->htile_offset
);
435 assert(!rtex
->cmask
.size
);
436 assert(!rtex
->fmask
.size
);
437 assert(!rtex
->is_depth
);
440 r600_texture_reference(&new_tex
, NULL
);
442 p_atomic_inc(&rctx
->screen
->dirty_tex_counter
);
445 static void r600_texture_get_info(struct pipe_screen
* screen
,
446 struct pipe_resource
*resource
,
450 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
451 struct r600_texture
*rtex
= (struct r600_texture
*)resource
;
455 if (!rscreen
|| !rtex
)
458 if (resource
->target
!= PIPE_BUFFER
) {
459 offset
= rtex
->surface
.u
.legacy
.level
[0].offset
;
460 stride
= rtex
->surface
.u
.legacy
.level
[0].nblk_x
*
471 static boolean
r600_texture_get_handle(struct pipe_screen
* screen
,
472 struct pipe_context
*ctx
,
473 struct pipe_resource
*resource
,
474 struct winsys_handle
*whandle
,
477 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
478 struct r600_common_context
*rctx
;
479 struct r600_resource
*res
= (struct r600_resource
*)resource
;
480 struct r600_texture
*rtex
= (struct r600_texture
*)resource
;
481 struct radeon_bo_metadata metadata
;
482 bool update_metadata
= false;
483 unsigned stride
, offset
, slice_size
;
485 ctx
= threaded_context_unwrap_sync(ctx
);
486 rctx
= (struct r600_common_context
*)(ctx
? ctx
: rscreen
->aux_context
);
488 if (resource
->target
!= PIPE_BUFFER
) {
489 /* This is not supported now, but it might be required for OpenCL
490 * interop in the future.
492 if (resource
->nr_samples
> 1 || rtex
->is_depth
)
495 /* Move a suballocated texture into a non-suballocated allocation. */
496 if (rscreen
->ws
->buffer_is_suballocated(res
->buf
) ||
497 rtex
->surface
.tile_swizzle
) {
498 assert(!res
->b
.is_shared
);
499 r600_reallocate_texture_inplace(rctx
, rtex
,
500 PIPE_BIND_SHARED
, false);
501 rctx
->b
.flush(&rctx
->b
, NULL
, 0);
502 assert(res
->b
.b
.bind
& PIPE_BIND_SHARED
);
503 assert(res
->flags
& RADEON_FLAG_NO_SUBALLOC
);
504 assert(rtex
->surface
.tile_swizzle
== 0);
507 if (!(usage
& PIPE_HANDLE_USAGE_EXPLICIT_FLUSH
) &&
509 /* Eliminate fast clear (CMASK) */
510 r600_eliminate_fast_color_clear(rctx
, rtex
);
512 /* Disable CMASK if flush_resource isn't going
515 if (rtex
->cmask
.size
)
516 r600_texture_discard_cmask(rscreen
, rtex
);
520 if (!res
->b
.is_shared
|| update_metadata
) {
521 r600_texture_init_metadata(rscreen
, rtex
, &metadata
);
522 if (rscreen
->query_opaque_metadata
)
523 rscreen
->query_opaque_metadata(rscreen
, rtex
,
526 rscreen
->ws
->buffer_set_metadata(res
->buf
, &metadata
);
529 slice_size
= (uint64_t)rtex
->surface
.u
.legacy
.level
[0].slice_size_dw
* 4;
531 /* Move a suballocated buffer into a non-suballocated allocation. */
532 if (rscreen
->ws
->buffer_is_suballocated(res
->buf
)) {
533 assert(!res
->b
.is_shared
);
535 /* Allocate a new buffer with PIPE_BIND_SHARED. */
536 struct pipe_resource templ
= res
->b
.b
;
537 templ
.bind
|= PIPE_BIND_SHARED
;
539 struct pipe_resource
*newb
=
540 screen
->resource_create(screen
, &templ
);
544 /* Copy the old buffer contents to the new one. */
546 u_box_1d(0, newb
->width0
, &box
);
547 rctx
->b
.resource_copy_region(&rctx
->b
, newb
, 0, 0, 0, 0,
549 /* Move the new buffer storage to the old pipe_resource. */
550 r600_replace_buffer_storage(&rctx
->b
, &res
->b
.b
, newb
);
551 pipe_resource_reference(&newb
, NULL
);
553 assert(res
->b
.b
.bind
& PIPE_BIND_SHARED
);
554 assert(res
->flags
& RADEON_FLAG_NO_SUBALLOC
);
561 r600_texture_get_info(screen
, resource
, &stride
, &offset
);
563 if (res
->b
.is_shared
) {
564 /* USAGE_EXPLICIT_FLUSH must be cleared if at least one user
567 res
->external_usage
|= usage
& ~PIPE_HANDLE_USAGE_EXPLICIT_FLUSH
;
568 if (!(usage
& PIPE_HANDLE_USAGE_EXPLICIT_FLUSH
))
569 res
->external_usage
&= ~PIPE_HANDLE_USAGE_EXPLICIT_FLUSH
;
571 res
->b
.is_shared
= true;
572 res
->external_usage
= usage
;
575 return rscreen
->ws
->buffer_get_handle(res
->buf
, stride
, offset
,
576 slice_size
, whandle
);
579 static void r600_texture_destroy(struct pipe_screen
*screen
,
580 struct pipe_resource
*ptex
)
582 struct r600_texture
*rtex
= (struct r600_texture
*)ptex
;
583 struct r600_resource
*resource
= &rtex
->resource
;
585 r600_texture_reference(&rtex
->flushed_depth_texture
, NULL
);
586 pipe_resource_reference((struct pipe_resource
**)&resource
->immed_buffer
, NULL
);
588 if (rtex
->cmask_buffer
!= &rtex
->resource
) {
589 r600_resource_reference(&rtex
->cmask_buffer
, NULL
);
591 pb_reference(&resource
->buf
, NULL
);
595 static const struct u_resource_vtbl r600_texture_vtbl
;
597 /* The number of samples can be specified independently of the texture. */
598 void r600_texture_get_fmask_info(struct r600_common_screen
*rscreen
,
599 struct r600_texture
*rtex
,
601 struct r600_fmask_info
*out
)
603 /* FMASK is allocated like an ordinary texture. */
604 struct pipe_resource templ
= rtex
->resource
.b
.b
;
605 struct radeon_surf fmask
= {};
608 memset(out
, 0, sizeof(*out
));
610 templ
.nr_samples
= 1;
611 flags
= rtex
->surface
.flags
| RADEON_SURF_FMASK
;
613 /* Use the same parameters and tile mode. */
614 fmask
.u
.legacy
.bankw
= rtex
->surface
.u
.legacy
.bankw
;
615 fmask
.u
.legacy
.bankh
= rtex
->surface
.u
.legacy
.bankh
;
616 fmask
.u
.legacy
.mtilea
= rtex
->surface
.u
.legacy
.mtilea
;
617 fmask
.u
.legacy
.tile_split
= rtex
->surface
.u
.legacy
.tile_split
;
620 fmask
.u
.legacy
.bankh
= 4;
622 switch (nr_samples
) {
631 R600_ERR("Invalid sample count for FMASK allocation.\n");
635 /* Overallocate FMASK on R600-R700 to fix colorbuffer corruption.
636 * This can be fixed by writing a separate FMASK allocator specifically
637 * for R600-R700 asics. */
638 if (rscreen
->chip_class
<= R700
) {
642 if (rscreen
->ws
->surface_init(rscreen
->ws
, &templ
,
643 flags
, bpe
, RADEON_SURF_MODE_2D
, &fmask
)) {
644 R600_ERR("Got error in surface_init while allocating FMASK.\n");
648 assert(fmask
.u
.legacy
.level
[0].mode
== RADEON_SURF_MODE_2D
);
650 out
->slice_tile_max
= (fmask
.u
.legacy
.level
[0].nblk_x
* fmask
.u
.legacy
.level
[0].nblk_y
) / 64;
651 if (out
->slice_tile_max
)
652 out
->slice_tile_max
-= 1;
654 out
->tile_mode_index
= fmask
.u
.legacy
.tiling_index
[0];
655 out
->pitch_in_pixels
= fmask
.u
.legacy
.level
[0].nblk_x
;
656 out
->bank_height
= fmask
.u
.legacy
.bankh
;
657 out
->tile_swizzle
= fmask
.tile_swizzle
;
658 out
->alignment
= MAX2(256, fmask
.surf_alignment
);
659 out
->size
= fmask
.surf_size
;
662 static void r600_texture_allocate_fmask(struct r600_common_screen
*rscreen
,
663 struct r600_texture
*rtex
)
665 r600_texture_get_fmask_info(rscreen
, rtex
,
666 rtex
->resource
.b
.b
.nr_samples
, &rtex
->fmask
);
668 rtex
->fmask
.offset
= align64(rtex
->size
, rtex
->fmask
.alignment
);
669 rtex
->size
= rtex
->fmask
.offset
+ rtex
->fmask
.size
;
672 void r600_texture_get_cmask_info(struct r600_common_screen
*rscreen
,
673 struct r600_texture
*rtex
,
674 struct r600_cmask_info
*out
)
676 unsigned cmask_tile_width
= 8;
677 unsigned cmask_tile_height
= 8;
678 unsigned cmask_tile_elements
= cmask_tile_width
* cmask_tile_height
;
679 unsigned element_bits
= 4;
680 unsigned cmask_cache_bits
= 1024;
681 unsigned num_pipes
= rscreen
->info
.num_tile_pipes
;
682 unsigned pipe_interleave_bytes
= rscreen
->info
.pipe_interleave_bytes
;
684 unsigned elements_per_macro_tile
= (cmask_cache_bits
/ element_bits
) * num_pipes
;
685 unsigned pixels_per_macro_tile
= elements_per_macro_tile
* cmask_tile_elements
;
686 unsigned sqrt_pixels_per_macro_tile
= sqrt(pixels_per_macro_tile
);
687 unsigned macro_tile_width
= util_next_power_of_two(sqrt_pixels_per_macro_tile
);
688 unsigned macro_tile_height
= pixels_per_macro_tile
/ macro_tile_width
;
690 unsigned pitch_elements
= align(rtex
->resource
.b
.b
.width0
, macro_tile_width
);
691 unsigned height
= align(rtex
->resource
.b
.b
.height0
, macro_tile_height
);
693 unsigned base_align
= num_pipes
* pipe_interleave_bytes
;
694 unsigned slice_bytes
=
695 ((pitch_elements
* height
* element_bits
+ 7) / 8) / cmask_tile_elements
;
697 assert(macro_tile_width
% 128 == 0);
698 assert(macro_tile_height
% 128 == 0);
700 out
->slice_tile_max
= ((pitch_elements
* height
) / (128*128)) - 1;
701 out
->alignment
= MAX2(256, base_align
);
702 out
->size
= util_num_layers(&rtex
->resource
.b
.b
, 0) *
703 align(slice_bytes
, base_align
);
706 static void r600_texture_allocate_cmask(struct r600_common_screen
*rscreen
,
707 struct r600_texture
*rtex
)
709 r600_texture_get_cmask_info(rscreen
, rtex
, &rtex
->cmask
);
711 rtex
->cmask
.offset
= align64(rtex
->size
, rtex
->cmask
.alignment
);
712 rtex
->size
= rtex
->cmask
.offset
+ rtex
->cmask
.size
;
714 rtex
->cb_color_info
|= EG_S_028C70_FAST_CLEAR(1);
717 static void r600_texture_alloc_cmask_separate(struct r600_common_screen
*rscreen
,
718 struct r600_texture
*rtex
)
720 if (rtex
->cmask_buffer
)
723 assert(rtex
->cmask
.size
== 0);
725 r600_texture_get_cmask_info(rscreen
, rtex
, &rtex
->cmask
);
727 rtex
->cmask_buffer
= (struct r600_resource
*)
728 r600_aligned_buffer_create(&rscreen
->b
,
729 R600_RESOURCE_FLAG_UNMAPPABLE
,
732 rtex
->cmask
.alignment
);
733 if (rtex
->cmask_buffer
== NULL
) {
734 rtex
->cmask
.size
= 0;
738 /* update colorbuffer state bits */
739 rtex
->cmask
.base_address_reg
= rtex
->cmask_buffer
->gpu_address
>> 8;
741 rtex
->cb_color_info
|= EG_S_028C70_FAST_CLEAR(1);
743 p_atomic_inc(&rscreen
->compressed_colortex_counter
);
746 void eg_resource_alloc_immed(struct r600_common_screen
*rscreen
,
747 struct r600_resource
*res
,
750 res
->immed_buffer
= (struct r600_resource
*)
751 pipe_buffer_create(&rscreen
->b
, PIPE_BIND_CUSTOM
,
752 PIPE_USAGE_DEFAULT
, immed_size
);
755 static void r600_texture_get_htile_size(struct r600_common_screen
*rscreen
,
756 struct r600_texture
*rtex
)
758 unsigned cl_width
, cl_height
, width
, height
;
759 unsigned slice_elements
, slice_bytes
, pipe_interleave_bytes
, base_align
;
760 unsigned num_pipes
= rscreen
->info
.num_tile_pipes
;
762 rtex
->surface
.htile_size
= 0;
764 if (rscreen
->chip_class
<= EVERGREEN
&&
765 rscreen
->info
.drm_major
== 2 && rscreen
->info
.drm_minor
< 26)
768 /* HW bug on R6xx. */
769 if (rscreen
->chip_class
== R600
&&
770 (rtex
->resource
.b
.b
.width0
> 7680 ||
771 rtex
->resource
.b
.b
.height0
> 7680))
800 width
= align(rtex
->surface
.u
.legacy
.level
[0].nblk_x
, cl_width
* 8);
801 height
= align(rtex
->surface
.u
.legacy
.level
[0].nblk_y
, cl_height
* 8);
803 slice_elements
= (width
* height
) / (8 * 8);
804 slice_bytes
= slice_elements
* 4;
806 pipe_interleave_bytes
= rscreen
->info
.pipe_interleave_bytes
;
807 base_align
= num_pipes
* pipe_interleave_bytes
;
809 rtex
->surface
.htile_alignment
= base_align
;
810 rtex
->surface
.htile_size
=
811 util_num_layers(&rtex
->resource
.b
.b
, 0) *
812 align(slice_bytes
, base_align
);
815 static void r600_texture_allocate_htile(struct r600_common_screen
*rscreen
,
816 struct r600_texture
*rtex
)
818 r600_texture_get_htile_size(rscreen
, rtex
);
820 if (!rtex
->surface
.htile_size
)
823 rtex
->htile_offset
= align(rtex
->size
, rtex
->surface
.htile_alignment
);
824 rtex
->size
= rtex
->htile_offset
+ rtex
->surface
.htile_size
;
827 void r600_print_texture_info(struct r600_common_screen
*rscreen
,
828 struct r600_texture
*rtex
, struct u_log_context
*log
)
832 /* Common parameters. */
833 u_log_printf(log
, " Info: npix_x=%u, npix_y=%u, npix_z=%u, blk_w=%u, "
834 "blk_h=%u, array_size=%u, last_level=%u, "
835 "bpe=%u, nsamples=%u, flags=0x%x, %s\n",
836 rtex
->resource
.b
.b
.width0
, rtex
->resource
.b
.b
.height0
,
837 rtex
->resource
.b
.b
.depth0
, rtex
->surface
.blk_w
,
839 rtex
->resource
.b
.b
.array_size
, rtex
->resource
.b
.b
.last_level
,
840 rtex
->surface
.bpe
, rtex
->resource
.b
.b
.nr_samples
,
841 rtex
->surface
.flags
, util_format_short_name(rtex
->resource
.b
.b
.format
));
843 u_log_printf(log
, " Layout: size=%"PRIu64
", alignment=%u, bankw=%u, "
844 "bankh=%u, nbanks=%u, mtilea=%u, tilesplit=%u, pipeconfig=%u, scanout=%u\n",
845 rtex
->surface
.surf_size
, rtex
->surface
.surf_alignment
, rtex
->surface
.u
.legacy
.bankw
,
846 rtex
->surface
.u
.legacy
.bankh
, rtex
->surface
.u
.legacy
.num_banks
, rtex
->surface
.u
.legacy
.mtilea
,
847 rtex
->surface
.u
.legacy
.tile_split
, rtex
->surface
.u
.legacy
.pipe_config
,
848 (rtex
->surface
.flags
& RADEON_SURF_SCANOUT
) != 0);
850 if (rtex
->fmask
.size
)
851 u_log_printf(log
, " FMask: offset=%"PRIu64
", size=%"PRIu64
", alignment=%u, pitch_in_pixels=%u, "
852 "bankh=%u, slice_tile_max=%u, tile_mode_index=%u\n",
853 rtex
->fmask
.offset
, rtex
->fmask
.size
, rtex
->fmask
.alignment
,
854 rtex
->fmask
.pitch_in_pixels
, rtex
->fmask
.bank_height
,
855 rtex
->fmask
.slice_tile_max
, rtex
->fmask
.tile_mode_index
);
857 if (rtex
->cmask
.size
)
858 u_log_printf(log
, " CMask: offset=%"PRIu64
", size=%"PRIu64
", alignment=%u, "
859 "slice_tile_max=%u\n",
860 rtex
->cmask
.offset
, rtex
->cmask
.size
, rtex
->cmask
.alignment
,
861 rtex
->cmask
.slice_tile_max
);
863 if (rtex
->htile_offset
)
864 u_log_printf(log
, " HTile: offset=%"PRIu64
", size=%u "
866 rtex
->htile_offset
, rtex
->surface
.htile_size
,
867 rtex
->surface
.htile_alignment
);
869 for (i
= 0; i
<= rtex
->resource
.b
.b
.last_level
; i
++)
870 u_log_printf(log
, " Level[%i]: offset=%"PRIu64
", slice_size=%"PRIu64
", "
871 "npix_x=%u, npix_y=%u, npix_z=%u, nblk_x=%u, nblk_y=%u, "
872 "mode=%u, tiling_index = %u\n",
873 i
, rtex
->surface
.u
.legacy
.level
[i
].offset
,
874 (uint64_t)rtex
->surface
.u
.legacy
.level
[i
].slice_size_dw
* 4,
875 u_minify(rtex
->resource
.b
.b
.width0
, i
),
876 u_minify(rtex
->resource
.b
.b
.height0
, i
),
877 u_minify(rtex
->resource
.b
.b
.depth0
, i
),
878 rtex
->surface
.u
.legacy
.level
[i
].nblk_x
,
879 rtex
->surface
.u
.legacy
.level
[i
].nblk_y
,
880 rtex
->surface
.u
.legacy
.level
[i
].mode
,
881 rtex
->surface
.u
.legacy
.tiling_index
[i
]);
883 if (rtex
->surface
.has_stencil
) {
884 u_log_printf(log
, " StencilLayout: tilesplit=%u\n",
885 rtex
->surface
.u
.legacy
.stencil_tile_split
);
886 for (i
= 0; i
<= rtex
->resource
.b
.b
.last_level
; i
++) {
887 u_log_printf(log
, " StencilLevel[%i]: offset=%"PRIu64
", "
888 "slice_size=%"PRIu64
", npix_x=%u, "
889 "npix_y=%u, npix_z=%u, nblk_x=%u, nblk_y=%u, "
890 "mode=%u, tiling_index = %u\n",
891 i
, rtex
->surface
.u
.legacy
.stencil_level
[i
].offset
,
892 (uint64_t)rtex
->surface
.u
.legacy
.stencil_level
[i
].slice_size_dw
* 4,
893 u_minify(rtex
->resource
.b
.b
.width0
, i
),
894 u_minify(rtex
->resource
.b
.b
.height0
, i
),
895 u_minify(rtex
->resource
.b
.b
.depth0
, i
),
896 rtex
->surface
.u
.legacy
.stencil_level
[i
].nblk_x
,
897 rtex
->surface
.u
.legacy
.stencil_level
[i
].nblk_y
,
898 rtex
->surface
.u
.legacy
.stencil_level
[i
].mode
,
899 rtex
->surface
.u
.legacy
.stencil_tiling_index
[i
]);
904 /* Common processing for r600_texture_create and r600_texture_from_handle */
905 static struct r600_texture
*
906 r600_texture_create_object(struct pipe_screen
*screen
,
907 const struct pipe_resource
*base
,
908 struct pb_buffer
*buf
,
909 struct radeon_surf
*surface
)
911 struct r600_texture
*rtex
;
912 struct r600_resource
*resource
;
913 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
915 rtex
= CALLOC_STRUCT(r600_texture
);
919 resource
= &rtex
->resource
;
920 resource
->b
.b
= *base
;
921 resource
->b
.b
.next
= NULL
;
922 resource
->b
.vtbl
= &r600_texture_vtbl
;
923 pipe_reference_init(&resource
->b
.b
.reference
, 1);
924 resource
->b
.b
.screen
= screen
;
926 /* don't include stencil-only formats which we don't support for rendering */
927 rtex
->is_depth
= util_format_has_depth(util_format_description(rtex
->resource
.b
.b
.format
));
929 rtex
->surface
= *surface
;
930 rtex
->size
= rtex
->surface
.surf_size
;
931 rtex
->db_render_format
= base
->format
;
933 /* Tiled depth textures utilize the non-displayable tile order.
934 * This must be done after r600_setup_surface.
935 * Applies to R600-Cayman. */
936 rtex
->non_disp_tiling
= rtex
->is_depth
&& rtex
->surface
.u
.legacy
.level
[0].mode
>= RADEON_SURF_MODE_1D
;
937 /* Applies to GCN. */
938 rtex
->last_msaa_resolve_target_micro_mode
= rtex
->surface
.micro_tile_mode
;
940 if (rtex
->is_depth
) {
941 if (base
->flags
& (R600_RESOURCE_FLAG_TRANSFER
|
942 R600_RESOURCE_FLAG_FLUSHED_DEPTH
) ||
943 rscreen
->chip_class
>= EVERGREEN
) {
944 rtex
->can_sample_z
= !rtex
->surface
.u
.legacy
.depth_adjusted
;
945 rtex
->can_sample_s
= !rtex
->surface
.u
.legacy
.stencil_adjusted
;
947 if (rtex
->resource
.b
.b
.nr_samples
<= 1 &&
948 (rtex
->resource
.b
.b
.format
== PIPE_FORMAT_Z16_UNORM
||
949 rtex
->resource
.b
.b
.format
== PIPE_FORMAT_Z32_FLOAT
))
950 rtex
->can_sample_z
= true;
953 if (!(base
->flags
& (R600_RESOURCE_FLAG_TRANSFER
|
954 R600_RESOURCE_FLAG_FLUSHED_DEPTH
))) {
955 rtex
->db_compatible
= true;
957 if (!(rscreen
->debug_flags
& DBG_NO_HYPERZ
))
958 r600_texture_allocate_htile(rscreen
, rtex
);
961 if (base
->nr_samples
> 1) {
963 r600_texture_allocate_fmask(rscreen
, rtex
);
964 r600_texture_allocate_cmask(rscreen
, rtex
);
965 rtex
->cmask_buffer
= &rtex
->resource
;
967 if (!rtex
->fmask
.size
|| !rtex
->cmask
.size
) {
974 /* Now create the backing buffer. */
976 r600_init_resource_fields(rscreen
, resource
, rtex
->size
,
977 rtex
->surface
.surf_alignment
);
979 if (!r600_alloc_resource(rscreen
, resource
)) {
985 resource
->gpu_address
= rscreen
->ws
->buffer_get_virtual_address(resource
->buf
);
986 resource
->bo_size
= buf
->size
;
987 resource
->bo_alignment
= buf
->alignment
;
988 resource
->domains
= rscreen
->ws
->buffer_get_initial_domain(resource
->buf
);
989 if (resource
->domains
& RADEON_DOMAIN_VRAM
)
990 resource
->vram_usage
= buf
->size
;
991 else if (resource
->domains
& RADEON_DOMAIN_GTT
)
992 resource
->gart_usage
= buf
->size
;
995 if (rtex
->cmask
.size
) {
996 /* Initialize the cmask to 0xCC (= compressed state). */
997 r600_screen_clear_buffer(rscreen
, &rtex
->cmask_buffer
->b
.b
,
998 rtex
->cmask
.offset
, rtex
->cmask
.size
,
1001 if (rtex
->htile_offset
) {
1002 uint32_t clear_value
= 0;
1004 r600_screen_clear_buffer(rscreen
, &rtex
->resource
.b
.b
,
1006 rtex
->surface
.htile_size
,
1010 /* Initialize the CMASK base register value. */
1011 rtex
->cmask
.base_address_reg
=
1012 (rtex
->resource
.gpu_address
+ rtex
->cmask
.offset
) >> 8;
1014 if (rscreen
->debug_flags
& DBG_VM
) {
1015 fprintf(stderr
, "VM start=0x%"PRIX64
" end=0x%"PRIX64
" | Texture %ix%ix%i, %i levels, %i samples, %s\n",
1016 rtex
->resource
.gpu_address
,
1017 rtex
->resource
.gpu_address
+ rtex
->resource
.buf
->size
,
1018 base
->width0
, base
->height0
, util_num_layers(base
, 0), base
->last_level
+1,
1019 base
->nr_samples
? base
->nr_samples
: 1, util_format_short_name(base
->format
));
1022 if (rscreen
->debug_flags
& DBG_TEX
) {
1024 struct u_log_context log
;
1025 u_log_context_init(&log
);
1026 r600_print_texture_info(rscreen
, rtex
, &log
);
1027 u_log_new_page_print(&log
, stdout
);
1029 u_log_context_destroy(&log
);
1035 static enum radeon_surf_mode
1036 r600_choose_tiling(struct r600_common_screen
*rscreen
,
1037 const struct pipe_resource
*templ
)
1039 const struct util_format_description
*desc
= util_format_description(templ
->format
);
1040 bool force_tiling
= templ
->flags
& R600_RESOURCE_FLAG_FORCE_TILING
;
1041 bool is_depth_stencil
= util_format_is_depth_or_stencil(templ
->format
) &&
1042 !(templ
->flags
& R600_RESOURCE_FLAG_FLUSHED_DEPTH
);
1044 /* MSAA resources must be 2D tiled. */
1045 if (templ
->nr_samples
> 1)
1046 return RADEON_SURF_MODE_2D
;
1048 /* Transfer resources should be linear. */
1049 if (templ
->flags
& R600_RESOURCE_FLAG_TRANSFER
)
1050 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1052 /* r600g: force tiling on TEXTURE_2D and TEXTURE_3D compute resources. */
1053 if (rscreen
->chip_class
>= R600
&& rscreen
->chip_class
<= CAYMAN
&&
1054 (templ
->bind
& PIPE_BIND_COMPUTE_RESOURCE
) &&
1055 (templ
->target
== PIPE_TEXTURE_2D
||
1056 templ
->target
== PIPE_TEXTURE_3D
))
1057 force_tiling
= true;
1059 /* Handle common candidates for the linear mode.
1060 * Compressed textures and DB surfaces must always be tiled.
1062 if (!force_tiling
&&
1063 !is_depth_stencil
&&
1064 !util_format_is_compressed(templ
->format
)) {
1065 if (rscreen
->debug_flags
& DBG_NO_TILING
)
1066 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1068 /* Tiling doesn't work with the 422 (SUBSAMPLED) formats on R600+. */
1069 if (desc
->layout
== UTIL_FORMAT_LAYOUT_SUBSAMPLED
)
1070 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1072 if (templ
->bind
& PIPE_BIND_LINEAR
)
1073 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1075 /* 1D textures should be linear - fixes image operations on 1d */
1076 if (templ
->target
== PIPE_TEXTURE_1D
||
1077 templ
->target
== PIPE_TEXTURE_1D_ARRAY
)
1078 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1080 /* Textures likely to be mapped often. */
1081 if (templ
->usage
== PIPE_USAGE_STAGING
||
1082 templ
->usage
== PIPE_USAGE_STREAM
)
1083 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
1086 /* Make small textures 1D tiled. */
1087 if (templ
->width0
<= 16 || templ
->height0
<= 16 ||
1088 (rscreen
->debug_flags
& DBG_NO_2D_TILING
))
1089 return RADEON_SURF_MODE_1D
;
1091 /* The allocator will switch to 1D if needed. */
1092 return RADEON_SURF_MODE_2D
;
1095 struct pipe_resource
*r600_texture_create(struct pipe_screen
*screen
,
1096 const struct pipe_resource
*templ
)
1098 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
1099 struct radeon_surf surface
= {0};
1100 bool is_flushed_depth
= templ
->flags
& R600_RESOURCE_FLAG_FLUSHED_DEPTH
;
1103 r
= r600_init_surface(rscreen
, &surface
, templ
,
1104 r600_choose_tiling(rscreen
, templ
), 0, 0,
1105 false, false, is_flushed_depth
);
1110 return (struct pipe_resource
*)
1111 r600_texture_create_object(screen
, templ
, NULL
, &surface
);
1114 static struct pipe_resource
*r600_texture_from_handle(struct pipe_screen
*screen
,
1115 const struct pipe_resource
*templ
,
1116 struct winsys_handle
*whandle
,
1119 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
1120 struct pb_buffer
*buf
= NULL
;
1121 unsigned stride
= 0, offset
= 0;
1122 enum radeon_surf_mode array_mode
;
1123 struct radeon_surf surface
= {};
1125 struct radeon_bo_metadata metadata
= {};
1126 struct r600_texture
*rtex
;
1129 /* Support only 2D textures without mipmaps */
1130 if ((templ
->target
!= PIPE_TEXTURE_2D
&& templ
->target
!= PIPE_TEXTURE_RECT
) ||
1131 templ
->depth0
!= 1 || templ
->last_level
!= 0)
1134 buf
= rscreen
->ws
->buffer_from_handle(rscreen
->ws
, whandle
,
1135 rscreen
->info
.max_alignment
,
1140 rscreen
->ws
->buffer_get_metadata(buf
, &metadata
);
1141 r600_surface_import_metadata(rscreen
, &surface
, &metadata
,
1142 &array_mode
, &is_scanout
);
1144 r
= r600_init_surface(rscreen
, &surface
, templ
, array_mode
, stride
,
1145 offset
, true, is_scanout
, false);
1150 rtex
= r600_texture_create_object(screen
, templ
, buf
, &surface
);
1154 rtex
->resource
.b
.is_shared
= true;
1155 rtex
->resource
.external_usage
= usage
;
1157 if (rscreen
->apply_opaque_metadata
)
1158 rscreen
->apply_opaque_metadata(rscreen
, rtex
, &metadata
);
1160 assert(rtex
->surface
.tile_swizzle
== 0);
1161 return &rtex
->resource
.b
.b
;
1164 bool r600_init_flushed_depth_texture(struct pipe_context
*ctx
,
1165 struct pipe_resource
*texture
,
1166 struct r600_texture
**staging
)
1168 struct r600_texture
*rtex
= (struct r600_texture
*)texture
;
1169 struct pipe_resource resource
;
1170 struct r600_texture
**flushed_depth_texture
= staging
?
1171 staging
: &rtex
->flushed_depth_texture
;
1172 enum pipe_format pipe_format
= texture
->format
;
1175 if (rtex
->flushed_depth_texture
)
1176 return true; /* it's ready */
1178 if (!rtex
->can_sample_z
&& rtex
->can_sample_s
) {
1179 switch (pipe_format
) {
1180 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
:
1181 /* Save memory by not allocating the S plane. */
1182 pipe_format
= PIPE_FORMAT_Z32_FLOAT
;
1184 case PIPE_FORMAT_Z24_UNORM_S8_UINT
:
1185 case PIPE_FORMAT_S8_UINT_Z24_UNORM
:
1186 /* Save memory bandwidth by not copying the
1187 * stencil part during flush.
1189 * This potentially increases memory bandwidth
1190 * if an application uses both Z and S texturing
1191 * simultaneously (a flushed Z24S8 texture
1192 * would be stored compactly), but how often
1193 * does that really happen?
1195 pipe_format
= PIPE_FORMAT_Z24X8_UNORM
;
1199 } else if (!rtex
->can_sample_s
&& rtex
->can_sample_z
) {
1200 assert(util_format_has_stencil(util_format_description(pipe_format
)));
1202 /* DB->CB copies to an 8bpp surface don't work. */
1203 pipe_format
= PIPE_FORMAT_X24S8_UINT
;
1207 memset(&resource
, 0, sizeof(resource
));
1208 resource
.target
= texture
->target
;
1209 resource
.format
= pipe_format
;
1210 resource
.width0
= texture
->width0
;
1211 resource
.height0
= texture
->height0
;
1212 resource
.depth0
= texture
->depth0
;
1213 resource
.array_size
= texture
->array_size
;
1214 resource
.last_level
= texture
->last_level
;
1215 resource
.nr_samples
= texture
->nr_samples
;
1216 resource
.usage
= staging
? PIPE_USAGE_STAGING
: PIPE_USAGE_DEFAULT
;
1217 resource
.bind
= texture
->bind
& ~PIPE_BIND_DEPTH_STENCIL
;
1218 resource
.flags
= texture
->flags
| R600_RESOURCE_FLAG_FLUSHED_DEPTH
;
1221 resource
.flags
|= R600_RESOURCE_FLAG_TRANSFER
;
1223 *flushed_depth_texture
= (struct r600_texture
*)ctx
->screen
->resource_create(ctx
->screen
, &resource
);
1224 if (*flushed_depth_texture
== NULL
) {
1225 R600_ERR("failed to create temporary texture to hold flushed depth\n");
1229 (*flushed_depth_texture
)->non_disp_tiling
= false;
1234 * Initialize the pipe_resource descriptor to be of the same size as the box,
1235 * which is supposed to hold a subregion of the texture "orig" at the given
1238 static void r600_init_temp_resource_from_box(struct pipe_resource
*res
,
1239 struct pipe_resource
*orig
,
1240 const struct pipe_box
*box
,
1241 unsigned level
, unsigned flags
)
1243 memset(res
, 0, sizeof(*res
));
1244 res
->format
= orig
->format
;
1245 res
->width0
= box
->width
;
1246 res
->height0
= box
->height
;
1248 res
->array_size
= 1;
1249 res
->usage
= flags
& R600_RESOURCE_FLAG_TRANSFER
? PIPE_USAGE_STAGING
: PIPE_USAGE_DEFAULT
;
1252 /* We must set the correct texture target and dimensions for a 3D box. */
1253 if (box
->depth
> 1 && util_max_layer(orig
, level
) > 0) {
1254 res
->target
= PIPE_TEXTURE_2D_ARRAY
;
1255 res
->array_size
= box
->depth
;
1257 res
->target
= PIPE_TEXTURE_2D
;
1261 static bool r600_can_invalidate_texture(struct r600_common_screen
*rscreen
,
1262 struct r600_texture
*rtex
,
1263 unsigned transfer_usage
,
1264 const struct pipe_box
*box
)
1266 /* r600g doesn't react to dirty_tex_descriptor_counter */
1267 return rscreen
->chip_class
>= GFX6
&&
1268 !rtex
->resource
.b
.is_shared
&&
1269 !(transfer_usage
& PIPE_TRANSFER_READ
) &&
1270 rtex
->resource
.b
.b
.last_level
== 0 &&
1271 util_texrange_covers_whole_level(&rtex
->resource
.b
.b
, 0,
1272 box
->x
, box
->y
, box
->z
,
1273 box
->width
, box
->height
,
1277 static void r600_texture_invalidate_storage(struct r600_common_context
*rctx
,
1278 struct r600_texture
*rtex
)
1280 struct r600_common_screen
*rscreen
= rctx
->screen
;
1282 /* There is no point in discarding depth and tiled buffers. */
1283 assert(!rtex
->is_depth
);
1284 assert(rtex
->surface
.is_linear
);
1286 /* Reallocate the buffer in the same pipe_resource. */
1287 r600_alloc_resource(rscreen
, &rtex
->resource
);
1289 /* Initialize the CMASK base address (needed even without CMASK). */
1290 rtex
->cmask
.base_address_reg
=
1291 (rtex
->resource
.gpu_address
+ rtex
->cmask
.offset
) >> 8;
1293 p_atomic_inc(&rscreen
->dirty_tex_counter
);
1295 rctx
->num_alloc_tex_transfer_bytes
+= rtex
->size
;
1298 static void *r600_texture_transfer_map(struct pipe_context
*ctx
,
1299 struct pipe_resource
*texture
,
1302 const struct pipe_box
*box
,
1303 struct pipe_transfer
**ptransfer
)
1305 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
1306 struct r600_texture
*rtex
= (struct r600_texture
*)texture
;
1307 struct r600_transfer
*trans
;
1308 struct r600_resource
*buf
;
1309 unsigned offset
= 0;
1311 bool use_staging_texture
= false;
1313 assert(!(texture
->flags
& R600_RESOURCE_FLAG_TRANSFER
));
1314 assert(box
->width
&& box
->height
&& box
->depth
);
1316 /* Depth textures use staging unconditionally. */
1317 if (!rtex
->is_depth
) {
1318 /* Degrade the tile mode if we get too many transfers on APUs.
1319 * On dGPUs, the staging texture is always faster.
1320 * Only count uploads that are at least 4x4 pixels large.
1322 if (!rctx
->screen
->info
.has_dedicated_vram
&&
1324 box
->width
>= 4 && box
->height
>= 4 &&
1325 p_atomic_inc_return(&rtex
->num_level0_transfers
) == 10) {
1326 bool can_invalidate
=
1327 r600_can_invalidate_texture(rctx
->screen
, rtex
,
1330 r600_reallocate_texture_inplace(rctx
, rtex
,
1335 /* Tiled textures need to be converted into a linear texture for CPU
1336 * access. The staging texture is always linear and is placed in GART.
1338 * Reading from VRAM or GTT WC is slow, always use the staging
1339 * texture in this case.
1341 * Use the staging texture for uploads if the underlying BO
1344 if (!rtex
->surface
.is_linear
)
1345 use_staging_texture
= true;
1346 else if (usage
& PIPE_TRANSFER_READ
)
1347 use_staging_texture
=
1348 rtex
->resource
.domains
& RADEON_DOMAIN_VRAM
||
1349 rtex
->resource
.flags
& RADEON_FLAG_GTT_WC
;
1350 /* Write & linear only: */
1351 else if (r600_rings_is_buffer_referenced(rctx
, rtex
->resource
.buf
,
1352 RADEON_USAGE_READWRITE
) ||
1353 !rctx
->ws
->buffer_wait(rtex
->resource
.buf
, 0,
1354 RADEON_USAGE_READWRITE
)) {
1356 if (r600_can_invalidate_texture(rctx
->screen
, rtex
,
1358 r600_texture_invalidate_storage(rctx
, rtex
);
1360 use_staging_texture
= true;
1364 trans
= CALLOC_STRUCT(r600_transfer
);
1367 pipe_resource_reference(&trans
->b
.b
.resource
, texture
);
1368 trans
->b
.b
.level
= level
;
1369 trans
->b
.b
.usage
= usage
;
1370 trans
->b
.b
.box
= *box
;
1372 if (rtex
->is_depth
) {
1373 struct r600_texture
*staging_depth
;
1375 if (rtex
->resource
.b
.b
.nr_samples
> 1) {
1376 /* MSAA depth buffers need to be converted to single sample buffers.
1378 * Mapping MSAA depth buffers can occur if ReadPixels is called
1379 * with a multisample GLX visual.
1381 * First downsample the depth buffer to a temporary texture,
1382 * then decompress the temporary one to staging.
1384 * Only the region being mapped is transfered.
1386 struct pipe_resource resource
;
1388 r600_init_temp_resource_from_box(&resource
, texture
, box
, level
, 0);
1390 if (!r600_init_flushed_depth_texture(ctx
, &resource
, &staging_depth
)) {
1391 R600_ERR("failed to create temporary texture to hold untiled copy\n");
1396 if (usage
& PIPE_TRANSFER_READ
) {
1397 struct pipe_resource
*temp
= ctx
->screen
->resource_create(ctx
->screen
, &resource
);
1399 R600_ERR("failed to create a temporary depth texture\n");
1404 r600_copy_region_with_blit(ctx
, temp
, 0, 0, 0, 0, texture
, level
, box
);
1405 rctx
->blit_decompress_depth(ctx
, (struct r600_texture
*)temp
, staging_depth
,
1406 0, 0, 0, box
->depth
, 0, 0);
1407 pipe_resource_reference(&temp
, NULL
);
1410 /* Just get the strides. */
1411 r600_texture_get_offset(rctx
->screen
, staging_depth
, level
, NULL
,
1413 &trans
->b
.b
.layer_stride
);
1415 /* XXX: only readback the rectangle which is being mapped? */
1416 /* XXX: when discard is true, no need to read back from depth texture */
1417 if (!r600_init_flushed_depth_texture(ctx
, texture
, &staging_depth
)) {
1418 R600_ERR("failed to create temporary texture to hold untiled copy\n");
1423 rctx
->blit_decompress_depth(ctx
, rtex
, staging_depth
,
1425 box
->z
, box
->z
+ box
->depth
- 1,
1428 offset
= r600_texture_get_offset(rctx
->screen
, staging_depth
,
1431 &trans
->b
.b
.layer_stride
);
1434 trans
->staging
= (struct r600_resource
*)staging_depth
;
1435 buf
= trans
->staging
;
1436 } else if (use_staging_texture
) {
1437 struct pipe_resource resource
;
1438 struct r600_texture
*staging
;
1440 r600_init_temp_resource_from_box(&resource
, texture
, box
, level
,
1441 R600_RESOURCE_FLAG_TRANSFER
);
1442 resource
.usage
= (usage
& PIPE_TRANSFER_READ
) ?
1443 PIPE_USAGE_STAGING
: PIPE_USAGE_STREAM
;
1445 /* Create the temporary texture. */
1446 staging
= (struct r600_texture
*)ctx
->screen
->resource_create(ctx
->screen
, &resource
);
1448 R600_ERR("failed to create temporary texture to hold untiled copy\n");
1452 trans
->staging
= &staging
->resource
;
1454 /* Just get the strides. */
1455 r600_texture_get_offset(rctx
->screen
, staging
, 0, NULL
,
1457 &trans
->b
.b
.layer_stride
);
1459 if (usage
& PIPE_TRANSFER_READ
)
1460 r600_copy_to_staging_texture(ctx
, trans
);
1462 usage
|= PIPE_TRANSFER_UNSYNCHRONIZED
;
1464 buf
= trans
->staging
;
1466 /* the resource is mapped directly */
1467 offset
= r600_texture_get_offset(rctx
->screen
, rtex
, level
, box
,
1469 &trans
->b
.b
.layer_stride
);
1470 buf
= &rtex
->resource
;
1473 if (!(map
= r600_buffer_map_sync_with_rings(rctx
, buf
, usage
))) {
1474 r600_resource_reference(&trans
->staging
, NULL
);
1479 *ptransfer
= &trans
->b
.b
;
1480 return map
+ offset
;
1483 static void r600_texture_transfer_unmap(struct pipe_context
*ctx
,
1484 struct pipe_transfer
* transfer
)
1486 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
1487 struct r600_transfer
*rtransfer
= (struct r600_transfer
*)transfer
;
1488 struct pipe_resource
*texture
= transfer
->resource
;
1489 struct r600_texture
*rtex
= (struct r600_texture
*)texture
;
1491 if ((transfer
->usage
& PIPE_TRANSFER_WRITE
) && rtransfer
->staging
) {
1492 if (rtex
->is_depth
&& rtex
->resource
.b
.b
.nr_samples
<= 1) {
1493 ctx
->resource_copy_region(ctx
, texture
, transfer
->level
,
1494 transfer
->box
.x
, transfer
->box
.y
, transfer
->box
.z
,
1495 &rtransfer
->staging
->b
.b
, transfer
->level
,
1498 r600_copy_from_staging_texture(ctx
, rtransfer
);
1502 if (rtransfer
->staging
) {
1503 rctx
->num_alloc_tex_transfer_bytes
+= rtransfer
->staging
->buf
->size
;
1504 r600_resource_reference(&rtransfer
->staging
, NULL
);
1507 /* Heuristic for {upload, draw, upload, draw, ..}:
1509 * Flush the gfx IB if we've allocated too much texture storage.
1511 * The idea is that we don't want to build IBs that use too much
1512 * memory and put pressure on the kernel memory manager and we also
1513 * want to make temporary and invalidated buffers go idle ASAP to
1514 * decrease the total memory usage or make them reusable. The memory
1515 * usage will be slightly higher than given here because of the buffer
1516 * cache in the winsys.
1518 * The result is that the kernel memory manager is never a bottleneck.
1520 if (rctx
->num_alloc_tex_transfer_bytes
> rctx
->screen
->info
.gart_size
/ 4) {
1521 rctx
->gfx
.flush(rctx
, PIPE_FLUSH_ASYNC
, NULL
);
1522 rctx
->num_alloc_tex_transfer_bytes
= 0;
1525 pipe_resource_reference(&transfer
->resource
, NULL
);
1529 static const struct u_resource_vtbl r600_texture_vtbl
=
1531 NULL
, /* get_handle */
1532 r600_texture_destroy
, /* resource_destroy */
1533 r600_texture_transfer_map
, /* transfer_map */
1534 u_default_transfer_flush_region
, /* transfer_flush_region */
1535 r600_texture_transfer_unmap
, /* transfer_unmap */
1538 struct pipe_surface
*r600_create_surface_custom(struct pipe_context
*pipe
,
1539 struct pipe_resource
*texture
,
1540 const struct pipe_surface
*templ
,
1541 unsigned width0
, unsigned height0
,
1542 unsigned width
, unsigned height
)
1544 struct r600_surface
*surface
= CALLOC_STRUCT(r600_surface
);
1549 assert(templ
->u
.tex
.first_layer
<= util_max_layer(texture
, templ
->u
.tex
.level
));
1550 assert(templ
->u
.tex
.last_layer
<= util_max_layer(texture
, templ
->u
.tex
.level
));
1552 pipe_reference_init(&surface
->base
.reference
, 1);
1553 pipe_resource_reference(&surface
->base
.texture
, texture
);
1554 surface
->base
.context
= pipe
;
1555 surface
->base
.format
= templ
->format
;
1556 surface
->base
.width
= width
;
1557 surface
->base
.height
= height
;
1558 surface
->base
.u
= templ
->u
;
1560 surface
->width0
= width0
;
1561 surface
->height0
= height0
;
1563 return &surface
->base
;
1566 static struct pipe_surface
*r600_create_surface(struct pipe_context
*pipe
,
1567 struct pipe_resource
*tex
,
1568 const struct pipe_surface
*templ
)
1570 unsigned level
= templ
->u
.tex
.level
;
1571 unsigned width
= u_minify(tex
->width0
, level
);
1572 unsigned height
= u_minify(tex
->height0
, level
);
1573 unsigned width0
= tex
->width0
;
1574 unsigned height0
= tex
->height0
;
1576 if (tex
->target
!= PIPE_BUFFER
&& templ
->format
!= tex
->format
) {
1577 const struct util_format_description
*tex_desc
1578 = util_format_description(tex
->format
);
1579 const struct util_format_description
*templ_desc
1580 = util_format_description(templ
->format
);
1582 assert(tex_desc
->block
.bits
== templ_desc
->block
.bits
);
1584 /* Adjust size of surface if and only if the block width or
1585 * height is changed. */
1586 if (tex_desc
->block
.width
!= templ_desc
->block
.width
||
1587 tex_desc
->block
.height
!= templ_desc
->block
.height
) {
1588 unsigned nblks_x
= util_format_get_nblocksx(tex
->format
, width
);
1589 unsigned nblks_y
= util_format_get_nblocksy(tex
->format
, height
);
1591 width
= nblks_x
* templ_desc
->block
.width
;
1592 height
= nblks_y
* templ_desc
->block
.height
;
1594 width0
= util_format_get_nblocksx(tex
->format
, width0
);
1595 height0
= util_format_get_nblocksy(tex
->format
, height0
);
1599 return r600_create_surface_custom(pipe
, tex
, templ
,
1604 static void r600_surface_destroy(struct pipe_context
*pipe
,
1605 struct pipe_surface
*surface
)
1607 struct r600_surface
*surf
= (struct r600_surface
*)surface
;
1608 r600_resource_reference(&surf
->cb_buffer_fmask
, NULL
);
1609 r600_resource_reference(&surf
->cb_buffer_cmask
, NULL
);
1610 pipe_resource_reference(&surface
->texture
, NULL
);
1614 static void r600_clear_texture(struct pipe_context
*pipe
,
1615 struct pipe_resource
*tex
,
1617 const struct pipe_box
*box
,
1620 struct pipe_screen
*screen
= pipe
->screen
;
1621 struct r600_texture
*rtex
= (struct r600_texture
*)tex
;
1622 struct pipe_surface tmpl
= {{0}};
1623 struct pipe_surface
*sf
;
1624 const struct util_format_description
*desc
=
1625 util_format_description(tex
->format
);
1627 tmpl
.format
= tex
->format
;
1628 tmpl
.u
.tex
.first_layer
= box
->z
;
1629 tmpl
.u
.tex
.last_layer
= box
->z
+ box
->depth
- 1;
1630 tmpl
.u
.tex
.level
= level
;
1631 sf
= pipe
->create_surface(pipe
, tex
, &tmpl
);
1635 if (rtex
->is_depth
) {
1638 uint8_t stencil
= 0;
1640 /* Depth is always present. */
1641 clear
= PIPE_CLEAR_DEPTH
;
1642 desc
->unpack_z_float(&depth
, 0, data
, 0, 1, 1);
1644 if (rtex
->surface
.has_stencil
) {
1645 clear
|= PIPE_CLEAR_STENCIL
;
1646 desc
->unpack_s_8uint(&stencil
, 0, data
, 0, 1, 1);
1649 pipe
->clear_depth_stencil(pipe
, sf
, clear
, depth
, stencil
,
1651 box
->width
, box
->height
, false);
1653 union pipe_color_union color
;
1655 /* pipe_color_union requires the full vec4 representation. */
1656 if (util_format_is_pure_uint(tex
->format
))
1657 desc
->unpack_rgba_uint(color
.ui
, 0, data
, 0, 1, 1);
1658 else if (util_format_is_pure_sint(tex
->format
))
1659 desc
->unpack_rgba_sint(color
.i
, 0, data
, 0, 1, 1);
1661 desc
->unpack_rgba_float(color
.f
, 0, data
, 0, 1, 1);
1663 if (screen
->is_format_supported(screen
, tex
->format
,
1665 PIPE_BIND_RENDER_TARGET
)) {
1666 pipe
->clear_render_target(pipe
, sf
, &color
,
1668 box
->width
, box
->height
, false);
1670 /* Software fallback - just for R9G9B9E5_FLOAT */
1671 util_clear_render_target(pipe
, sf
, &color
,
1673 box
->width
, box
->height
);
1676 pipe_surface_reference(&sf
, NULL
);
1679 unsigned r600_translate_colorswap(enum pipe_format format
, bool do_endian_swap
)
1681 const struct util_format_description
*desc
= util_format_description(format
);
1683 #define HAS_SWIZZLE(chan,swz) (desc->swizzle[chan] == PIPE_SWIZZLE_##swz)
1685 if (format
== PIPE_FORMAT_R11G11B10_FLOAT
) /* isn't plain */
1686 return V_0280A0_SWAP_STD
;
1688 if (desc
->layout
!= UTIL_FORMAT_LAYOUT_PLAIN
)
1691 switch (desc
->nr_channels
) {
1693 if (HAS_SWIZZLE(0,X
))
1694 return V_0280A0_SWAP_STD
; /* X___ */
1695 else if (HAS_SWIZZLE(3,X
))
1696 return V_0280A0_SWAP_ALT_REV
; /* ___X */
1699 if ((HAS_SWIZZLE(0,X
) && HAS_SWIZZLE(1,Y
)) ||
1700 (HAS_SWIZZLE(0,X
) && HAS_SWIZZLE(1,NONE
)) ||
1701 (HAS_SWIZZLE(0,NONE
) && HAS_SWIZZLE(1,Y
)))
1702 return V_0280A0_SWAP_STD
; /* XY__ */
1703 else if ((HAS_SWIZZLE(0,Y
) && HAS_SWIZZLE(1,X
)) ||
1704 (HAS_SWIZZLE(0,Y
) && HAS_SWIZZLE(1,NONE
)) ||
1705 (HAS_SWIZZLE(0,NONE
) && HAS_SWIZZLE(1,X
)))
1707 return (do_endian_swap
? V_0280A0_SWAP_STD
: V_0280A0_SWAP_STD_REV
);
1708 else if (HAS_SWIZZLE(0,X
) && HAS_SWIZZLE(3,Y
))
1709 return V_0280A0_SWAP_ALT
; /* X__Y */
1710 else if (HAS_SWIZZLE(0,Y
) && HAS_SWIZZLE(3,X
))
1711 return V_0280A0_SWAP_ALT_REV
; /* Y__X */
1714 if (HAS_SWIZZLE(0,X
))
1715 return (do_endian_swap
? V_0280A0_SWAP_STD_REV
: V_0280A0_SWAP_STD
);
1716 else if (HAS_SWIZZLE(0,Z
))
1717 return V_0280A0_SWAP_STD_REV
; /* ZYX */
1720 /* check the middle channels, the 1st and 4th channel can be NONE */
1721 if (HAS_SWIZZLE(1,Y
) && HAS_SWIZZLE(2,Z
)) {
1722 return V_0280A0_SWAP_STD
; /* XYZW */
1723 } else if (HAS_SWIZZLE(1,Z
) && HAS_SWIZZLE(2,Y
)) {
1724 return V_0280A0_SWAP_STD_REV
; /* WZYX */
1725 } else if (HAS_SWIZZLE(1,Y
) && HAS_SWIZZLE(2,X
)) {
1726 return V_0280A0_SWAP_ALT
; /* ZYXW */
1727 } else if (HAS_SWIZZLE(1,Z
) && HAS_SWIZZLE(2,W
)) {
1730 return V_0280A0_SWAP_ALT_REV
;
1732 return (do_endian_swap
? V_0280A0_SWAP_ALT
: V_0280A0_SWAP_ALT_REV
);
1739 /* FAST COLOR CLEAR */
1741 static void evergreen_set_clear_color(struct r600_texture
*rtex
,
1742 enum pipe_format surface_format
,
1743 const union pipe_color_union
*color
)
1745 union util_color uc
;
1747 memset(&uc
, 0, sizeof(uc
));
1749 if (rtex
->surface
.bpe
== 16) {
1750 /* DCC fast clear only:
1751 * CLEAR_WORD0 = R = G = B
1754 assert(color
->ui
[0] == color
->ui
[1] &&
1755 color
->ui
[0] == color
->ui
[2]);
1756 uc
.ui
[0] = color
->ui
[0];
1757 uc
.ui
[1] = color
->ui
[3];
1758 } else if (util_format_is_pure_uint(surface_format
)) {
1759 util_format_write_4ui(surface_format
, color
->ui
, 0, &uc
, 0, 0, 0, 1, 1);
1760 } else if (util_format_is_pure_sint(surface_format
)) {
1761 util_format_write_4i(surface_format
, color
->i
, 0, &uc
, 0, 0, 0, 1, 1);
1763 util_pack_color(color
->f
, surface_format
, &uc
);
1766 memcpy(rtex
->color_clear_value
, &uc
, 2 * sizeof(uint32_t));
1769 void evergreen_do_fast_color_clear(struct r600_common_context
*rctx
,
1770 struct pipe_framebuffer_state
*fb
,
1771 struct r600_atom
*fb_state
,
1772 unsigned *buffers
, ubyte
*dirty_cbufs
,
1773 const union pipe_color_union
*color
)
1777 /* This function is broken in BE, so just disable this path for now */
1778 #ifdef PIPE_ARCH_BIG_ENDIAN
1782 if (rctx
->render_cond
)
1785 for (i
= 0; i
< fb
->nr_cbufs
; i
++) {
1786 struct r600_texture
*tex
;
1787 unsigned clear_bit
= PIPE_CLEAR_COLOR0
<< i
;
1792 /* if this colorbuffer is not being cleared */
1793 if (!(*buffers
& clear_bit
))
1796 tex
= (struct r600_texture
*)fb
->cbufs
[i
]->texture
;
1798 /* the clear is allowed if all layers are bound */
1799 if (fb
->cbufs
[i
]->u
.tex
.first_layer
!= 0 ||
1800 fb
->cbufs
[i
]->u
.tex
.last_layer
!= util_max_layer(&tex
->resource
.b
.b
, 0)) {
1804 /* cannot clear mipmapped textures */
1805 if (fb
->cbufs
[i
]->texture
->last_level
!= 0) {
1809 /* only supported on tiled surfaces */
1810 if (tex
->surface
.is_linear
) {
1814 /* shared textures can't use fast clear without an explicit flush,
1815 * because there is no way to communicate the clear color among
1818 if (tex
->resource
.b
.is_shared
&&
1819 !(tex
->resource
.external_usage
& PIPE_HANDLE_USAGE_EXPLICIT_FLUSH
))
1822 /* Use a slow clear for small surfaces where the cost of
1823 * the eliminate pass can be higher than the benefit of fast
1824 * clear. AMDGPU-pro does this, but the numbers may differ.
1826 * This helps on both dGPUs and APUs, even small ones.
1828 if (tex
->resource
.b
.b
.nr_samples
<= 1 &&
1829 tex
->resource
.b
.b
.width0
* tex
->resource
.b
.b
.height0
<= 300 * 300)
1833 /* 128-bit formats are unusupported */
1834 if (tex
->surface
.bpe
> 8) {
1838 /* ensure CMASK is enabled */
1839 r600_texture_alloc_cmask_separate(rctx
->screen
, tex
);
1840 if (tex
->cmask
.size
== 0) {
1844 /* Do the fast clear. */
1845 rctx
->clear_buffer(&rctx
->b
, &tex
->cmask_buffer
->b
.b
,
1846 tex
->cmask
.offset
, tex
->cmask
.size
, 0,
1847 R600_COHERENCY_CB_META
);
1849 bool need_compressed_update
= !tex
->dirty_level_mask
;
1851 tex
->dirty_level_mask
|= 1 << fb
->cbufs
[i
]->u
.tex
.level
;
1853 if (need_compressed_update
)
1854 p_atomic_inc(&rctx
->screen
->compressed_colortex_counter
);
1857 evergreen_set_clear_color(tex
, fb
->cbufs
[i
]->format
, color
);
1860 *dirty_cbufs
|= 1 << i
;
1861 rctx
->set_atom_dirty(rctx
, fb_state
, true);
1862 *buffers
&= ~clear_bit
;
1866 static struct pipe_memory_object
*
1867 r600_memobj_from_handle(struct pipe_screen
*screen
,
1868 struct winsys_handle
*whandle
,
1871 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
1872 struct r600_memory_object
*memobj
= CALLOC_STRUCT(r600_memory_object
);
1873 struct pb_buffer
*buf
= NULL
;
1874 uint32_t stride
, offset
;
1879 buf
= rscreen
->ws
->buffer_from_handle(rscreen
->ws
, whandle
,
1880 rscreen
->info
.max_alignment
,
1887 memobj
->b
.dedicated
= dedicated
;
1889 memobj
->stride
= stride
;
1890 memobj
->offset
= offset
;
1892 return (struct pipe_memory_object
*)memobj
;
1897 r600_memobj_destroy(struct pipe_screen
*screen
,
1898 struct pipe_memory_object
*_memobj
)
1900 struct r600_memory_object
*memobj
= (struct r600_memory_object
*)_memobj
;
1902 pb_reference(&memobj
->buf
, NULL
);
1906 static struct pipe_resource
*
1907 r600_texture_from_memobj(struct pipe_screen
*screen
,
1908 const struct pipe_resource
*templ
,
1909 struct pipe_memory_object
*_memobj
,
1913 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
1914 struct r600_memory_object
*memobj
= (struct r600_memory_object
*)_memobj
;
1915 struct r600_texture
*rtex
;
1916 struct radeon_surf surface
= {};
1917 struct radeon_bo_metadata metadata
= {};
1918 enum radeon_surf_mode array_mode
;
1920 struct pb_buffer
*buf
= NULL
;
1922 if (memobj
->b
.dedicated
) {
1923 rscreen
->ws
->buffer_get_metadata(memobj
->buf
, &metadata
);
1924 r600_surface_import_metadata(rscreen
, &surface
, &metadata
,
1925 &array_mode
, &is_scanout
);
1928 * The bo metadata is unset for un-dedicated images. So we fall
1929 * back to linear. See answer to question 5 of the
1930 * VK_KHX_external_memory spec for some details.
1932 * It is possible that this case isn't going to work if the
1933 * surface pitch isn't correctly aligned by default.
1935 * In order to support it correctly we require multi-image
1936 * metadata to be syncrhonized between radv and radeonsi. The
1937 * semantics of associating multiple image metadata to a memory
1938 * object on the vulkan export side are not concretely defined
1941 * All the use cases we are aware of at the moment for memory
1942 * objects use dedicated allocations. So lets keep the initial
1943 * implementation simple.
1945 * A possible alternative is to attempt to reconstruct the
1946 * tiling information when the TexParameter TEXTURE_TILING_EXT
1949 array_mode
= RADEON_SURF_MODE_LINEAR_ALIGNED
;
1954 r
= r600_init_surface(rscreen
, &surface
, templ
,
1955 array_mode
, memobj
->stride
,
1956 offset
, true, is_scanout
,
1961 rtex
= r600_texture_create_object(screen
, templ
, memobj
->buf
, &surface
);
1965 /* r600_texture_create_object doesn't increment refcount of
1966 * memobj->buf, so increment it here.
1968 pb_reference(&buf
, memobj
->buf
);
1970 rtex
->resource
.b
.is_shared
= true;
1971 rtex
->resource
.external_usage
= PIPE_HANDLE_USAGE_FRAMEBUFFER_WRITE
;
1973 if (rscreen
->apply_opaque_metadata
)
1974 rscreen
->apply_opaque_metadata(rscreen
, rtex
, &metadata
);
1976 return &rtex
->resource
.b
.b
;
1979 void r600_init_screen_texture_functions(struct r600_common_screen
*rscreen
)
1981 rscreen
->b
.resource_from_handle
= r600_texture_from_handle
;
1982 rscreen
->b
.resource_get_handle
= r600_texture_get_handle
;
1983 rscreen
->b
.resource_get_info
= r600_texture_get_info
;
1984 rscreen
->b
.resource_from_memobj
= r600_texture_from_memobj
;
1985 rscreen
->b
.memobj_create_from_handle
= r600_memobj_from_handle
;
1986 rscreen
->b
.memobj_destroy
= r600_memobj_destroy
;
1989 void r600_init_context_texture_functions(struct r600_common_context
*rctx
)
1991 rctx
->b
.create_surface
= r600_create_surface
;
1992 rctx
->b
.surface_destroy
= r600_surface_destroy
;
1993 rctx
->b
.clear_texture
= r600_clear_texture
;