2 * Copyright 2017 Advanced Micro Devices, Inc.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
28 #include "util/u_format.h"
29 #include "util/u_pack_color.h"
30 #include "util/u_surface.h"
33 SI_CLEAR
= SI_SAVE_FRAGMENT_STATE
,
34 SI_CLEAR_SURFACE
= SI_SAVE_FRAMEBUFFER
| SI_SAVE_FRAGMENT_STATE
,
37 enum si_dcc_clear_code
39 DCC_CLEAR_COLOR_0000
= 0x00000000,
40 DCC_CLEAR_COLOR_0001
= 0x40404040,
41 DCC_CLEAR_COLOR_1110
= 0x80808080,
42 DCC_CLEAR_COLOR_1111
= 0xC0C0C0C0,
43 DCC_CLEAR_COLOR_REG
= 0x20202020,
46 static void si_alloc_separate_cmask(struct si_screen
*sscreen
,
47 struct si_texture
*tex
)
49 if (tex
->cmask_buffer
|| !tex
->surface
.cmask_size
)
53 si_aligned_buffer_create(&sscreen
->b
,
54 SI_RESOURCE_FLAG_UNMAPPABLE
,
56 tex
->surface
.cmask_size
,
57 tex
->surface
.cmask_alignment
);
58 if (tex
->cmask_buffer
== NULL
)
61 tex
->cmask_base_address_reg
= tex
->cmask_buffer
->gpu_address
>> 8;
62 tex
->cb_color_info
|= S_028C70_FAST_CLEAR(1);
64 p_atomic_inc(&sscreen
->compressed_colortex_counter
);
67 static bool si_set_clear_color(struct si_texture
*tex
,
68 enum pipe_format surface_format
,
69 const union pipe_color_union
*color
)
73 memset(&uc
, 0, sizeof(uc
));
75 if (tex
->surface
.bpe
== 16) {
76 /* DCC fast clear only:
77 * CLEAR_WORD0 = R = G = B
80 assert(color
->ui
[0] == color
->ui
[1] &&
81 color
->ui
[0] == color
->ui
[2]);
82 uc
.ui
[0] = color
->ui
[0];
83 uc
.ui
[1] = color
->ui
[3];
84 } else if (util_format_is_pure_uint(surface_format
)) {
85 util_format_write_4ui(surface_format
, color
->ui
, 0, &uc
, 0, 0, 0, 1, 1);
86 } else if (util_format_is_pure_sint(surface_format
)) {
87 util_format_write_4i(surface_format
, color
->i
, 0, &uc
, 0, 0, 0, 1, 1);
89 util_pack_color(color
->f
, surface_format
, &uc
);
92 if (memcmp(tex
->color_clear_value
, &uc
, 2 * sizeof(uint32_t)) == 0)
95 memcpy(tex
->color_clear_value
, &uc
, 2 * sizeof(uint32_t));
99 /** Linearize and convert luminace/intensity to red. */
100 enum pipe_format
si_simplify_cb_format(enum pipe_format format
)
102 format
= util_format_linear(format
);
103 format
= util_format_luminance_to_red(format
);
104 return util_format_intensity_to_red(format
);
107 bool vi_alpha_is_on_msb(enum pipe_format format
)
109 format
= si_simplify_cb_format(format
);
111 /* Formats with 3 channels can't have alpha. */
112 if (util_format_description(format
)->nr_channels
== 3)
113 return true; /* same as xxxA; is any value OK here? */
115 return si_translate_colorswap(format
, false) <= 1;
118 static bool vi_get_fast_clear_parameters(enum pipe_format base_format
,
119 enum pipe_format surface_format
,
120 const union pipe_color_union
*color
,
121 uint32_t* clear_value
,
122 bool *eliminate_needed
)
124 /* If we want to clear without needing a fast clear eliminate step, we
125 * can set color and alpha independently to 0 or 1 (or 0/max for integer
128 bool values
[4] = {}; /* whether to clear to 0 or 1 */
129 bool color_value
= false; /* clear color to 0 or 1 */
130 bool alpha_value
= false; /* clear alpha to 0 or 1 */
131 int alpha_channel
; /* index of the alpha component */
132 bool has_color
= false;
133 bool has_alpha
= false;
135 const struct util_format_description
*desc
=
136 util_format_description(si_simplify_cb_format(surface_format
));
138 /* 128-bit fast clear with different R,G,B values is unsupported. */
139 if (desc
->block
.bits
== 128 &&
140 (color
->ui
[0] != color
->ui
[1] ||
141 color
->ui
[0] != color
->ui
[2]))
144 *eliminate_needed
= true;
145 *clear_value
= DCC_CLEAR_COLOR_REG
;
147 if (desc
->layout
!= UTIL_FORMAT_LAYOUT_PLAIN
)
148 return true; /* need ELIMINATE_FAST_CLEAR */
150 bool base_alpha_is_on_msb
= vi_alpha_is_on_msb(base_format
);
151 bool surf_alpha_is_on_msb
= vi_alpha_is_on_msb(surface_format
);
153 /* Formats with 3 channels can't have alpha. */
154 if (desc
->nr_channels
== 3)
156 else if (surf_alpha_is_on_msb
)
157 alpha_channel
= desc
->nr_channels
- 1;
161 for (int i
= 0; i
< 4; ++i
) {
162 if (desc
->swizzle
[i
] >= PIPE_SWIZZLE_0
)
165 if (desc
->channel
[i
].pure_integer
&&
166 desc
->channel
[i
].type
== UTIL_FORMAT_TYPE_SIGNED
) {
167 /* Use the maximum value for clamping the clear color. */
168 int max
= u_bit_consecutive(0, desc
->channel
[i
].size
- 1);
170 values
[i
] = color
->i
[i
] != 0;
171 if (color
->i
[i
] != 0 && MIN2(color
->i
[i
], max
) != max
)
172 return true; /* need ELIMINATE_FAST_CLEAR */
173 } else if (desc
->channel
[i
].pure_integer
&&
174 desc
->channel
[i
].type
== UTIL_FORMAT_TYPE_UNSIGNED
) {
175 /* Use the maximum value for clamping the clear color. */
176 unsigned max
= u_bit_consecutive(0, desc
->channel
[i
].size
);
178 values
[i
] = color
->ui
[i
] != 0U;
179 if (color
->ui
[i
] != 0U && MIN2(color
->ui
[i
], max
) != max
)
180 return true; /* need ELIMINATE_FAST_CLEAR */
182 values
[i
] = color
->f
[i
] != 0.0F
;
183 if (color
->f
[i
] != 0.0F
&& color
->f
[i
] != 1.0F
)
184 return true; /* need ELIMINATE_FAST_CLEAR */
187 if (desc
->swizzle
[i
] == alpha_channel
) {
188 alpha_value
= values
[i
];
191 color_value
= values
[i
];
196 /* If alpha isn't present, make it the same as color, and vice versa. */
198 alpha_value
= color_value
;
200 color_value
= alpha_value
;
202 if (color_value
!= alpha_value
&&
203 base_alpha_is_on_msb
!= surf_alpha_is_on_msb
)
204 return true; /* require ELIMINATE_FAST_CLEAR */
206 /* Check if all color values are equal if they are present. */
207 for (int i
= 0; i
< 4; ++i
) {
208 if (desc
->swizzle
[i
] <= PIPE_SWIZZLE_W
&&
209 desc
->swizzle
[i
] != alpha_channel
&&
210 values
[i
] != color_value
)
211 return true; /* require ELIMINATE_FAST_CLEAR */
214 /* This doesn't need ELIMINATE_FAST_CLEAR.
215 * On chips predating Raven2, the DCC clear codes and the CB clear
216 * color registers must match.
218 *eliminate_needed
= false;
222 *clear_value
= DCC_CLEAR_COLOR_1111
;
224 *clear_value
= DCC_CLEAR_COLOR_1110
;
227 *clear_value
= DCC_CLEAR_COLOR_0001
;
229 *clear_value
= DCC_CLEAR_COLOR_0000
;
234 void vi_dcc_clear_level(struct si_context
*sctx
,
235 struct si_texture
*tex
,
236 unsigned level
, unsigned clear_value
)
238 struct pipe_resource
*dcc_buffer
;
239 uint64_t dcc_offset
, clear_size
;
241 assert(vi_dcc_enabled(tex
, level
));
243 if (tex
->dcc_separate_buffer
) {
244 dcc_buffer
= &tex
->dcc_separate_buffer
->b
.b
;
247 dcc_buffer
= &tex
->buffer
.b
.b
;
248 dcc_offset
= tex
->dcc_offset
;
251 if (sctx
->chip_class
>= GFX9
) {
252 /* Mipmap level clears aren't implemented. */
253 assert(tex
->buffer
.b
.b
.last_level
== 0);
254 /* 4x and 8x MSAA needs a sophisticated compute shader for
255 * the clear. See AMDVLK. */
256 assert(tex
->buffer
.b
.b
.nr_storage_samples
<= 2);
257 clear_size
= tex
->surface
.dcc_size
;
259 unsigned num_layers
= util_num_layers(&tex
->buffer
.b
.b
, level
);
261 /* If this is 0, fast clear isn't possible. (can occur with MSAA) */
262 assert(tex
->surface
.u
.legacy
.level
[level
].dcc_fast_clear_size
);
263 /* Layered 4x and 8x MSAA DCC fast clears need to clear
264 * dcc_fast_clear_size bytes for each layer. A compute shader
265 * would be more efficient than separate per-layer clear operations.
267 assert(tex
->buffer
.b
.b
.nr_storage_samples
<= 2 || num_layers
== 1);
269 dcc_offset
+= tex
->surface
.u
.legacy
.level
[level
].dcc_offset
;
270 clear_size
= tex
->surface
.u
.legacy
.level
[level
].dcc_fast_clear_size
*
274 si_clear_buffer(sctx
, dcc_buffer
, dcc_offset
, clear_size
,
275 &clear_value
, 4, SI_COHERENCY_CB_META
);
278 /* Set the same micro tile mode as the destination of the last MSAA resolve.
279 * This allows hitting the MSAA resolve fast path, which requires that both
280 * src and dst micro tile modes match.
282 static void si_set_optimal_micro_tile_mode(struct si_screen
*sscreen
,
283 struct si_texture
*tex
)
285 if (tex
->buffer
.b
.is_shared
||
286 tex
->buffer
.b
.b
.nr_samples
<= 1 ||
287 tex
->surface
.micro_tile_mode
== tex
->last_msaa_resolve_target_micro_mode
)
290 assert(sscreen
->info
.chip_class
>= GFX9
||
291 tex
->surface
.u
.legacy
.level
[0].mode
== RADEON_SURF_MODE_2D
);
292 assert(tex
->buffer
.b
.b
.last_level
== 0);
294 if (sscreen
->info
.chip_class
>= GFX9
) {
295 /* 4K or larger tiles only. 0 is linear. 1-3 are 256B tiles. */
296 assert(tex
->surface
.u
.gfx9
.surf
.swizzle_mode
>= 4);
298 /* If you do swizzle_mode % 4, you'll get:
304 * Depth-sample order isn't allowed:
306 assert(tex
->surface
.u
.gfx9
.surf
.swizzle_mode
% 4 != 0);
308 switch (tex
->last_msaa_resolve_target_micro_mode
) {
309 case RADEON_MICRO_MODE_DISPLAY
:
310 tex
->surface
.u
.gfx9
.surf
.swizzle_mode
&= ~0x3;
311 tex
->surface
.u
.gfx9
.surf
.swizzle_mode
+= 2; /* D */
313 case RADEON_MICRO_MODE_THIN
:
314 tex
->surface
.u
.gfx9
.surf
.swizzle_mode
&= ~0x3;
315 tex
->surface
.u
.gfx9
.surf
.swizzle_mode
+= 1; /* S */
317 case RADEON_MICRO_MODE_ROTATED
:
318 tex
->surface
.u
.gfx9
.surf
.swizzle_mode
&= ~0x3;
319 tex
->surface
.u
.gfx9
.surf
.swizzle_mode
+= 3; /* R */
322 assert(!"unexpected micro mode");
325 } else if (sscreen
->info
.chip_class
>= CIK
) {
326 /* These magic numbers were copied from addrlib. It doesn't use
327 * any definitions for them either. They are all 2D_TILED_THIN1
328 * modes with different bpp and micro tile mode.
330 switch (tex
->last_msaa_resolve_target_micro_mode
) {
331 case RADEON_MICRO_MODE_DISPLAY
:
332 tex
->surface
.u
.legacy
.tiling_index
[0] = 10;
334 case RADEON_MICRO_MODE_THIN
:
335 tex
->surface
.u
.legacy
.tiling_index
[0] = 14;
337 case RADEON_MICRO_MODE_ROTATED
:
338 tex
->surface
.u
.legacy
.tiling_index
[0] = 28;
340 default: /* depth, thick */
341 assert(!"unexpected micro mode");
345 switch (tex
->last_msaa_resolve_target_micro_mode
) {
346 case RADEON_MICRO_MODE_DISPLAY
:
347 switch (tex
->surface
.bpe
) {
349 tex
->surface
.u
.legacy
.tiling_index
[0] = 10;
352 tex
->surface
.u
.legacy
.tiling_index
[0] = 11;
355 tex
->surface
.u
.legacy
.tiling_index
[0] = 12;
359 case RADEON_MICRO_MODE_THIN
:
360 switch (tex
->surface
.bpe
) {
362 tex
->surface
.u
.legacy
.tiling_index
[0] = 14;
365 tex
->surface
.u
.legacy
.tiling_index
[0] = 15;
368 tex
->surface
.u
.legacy
.tiling_index
[0] = 16;
371 tex
->surface
.u
.legacy
.tiling_index
[0] = 17;
375 default: /* depth, thick */
376 assert(!"unexpected micro mode");
381 tex
->surface
.micro_tile_mode
= tex
->last_msaa_resolve_target_micro_mode
;
383 p_atomic_inc(&sscreen
->dirty_tex_counter
);
386 static void si_do_fast_color_clear(struct si_context
*sctx
,
388 const union pipe_color_union
*color
)
390 struct pipe_framebuffer_state
*fb
= &sctx
->framebuffer
.state
;
393 /* This function is broken in BE, so just disable this path for now */
394 #ifdef PIPE_ARCH_BIG_ENDIAN
398 if (sctx
->render_cond
)
401 for (i
= 0; i
< fb
->nr_cbufs
; i
++) {
402 struct si_texture
*tex
;
403 unsigned clear_bit
= PIPE_CLEAR_COLOR0
<< i
;
408 /* if this colorbuffer is not being cleared */
409 if (!(*buffers
& clear_bit
))
412 unsigned level
= fb
->cbufs
[i
]->u
.tex
.level
;
416 tex
= (struct si_texture
*)fb
->cbufs
[i
]->texture
;
418 /* TODO: GFX9: Implement DCC fast clear for level 0 of
419 * mipmapped textures. Mipmapped DCC has to clear a rectangular
420 * area of DCC for level 0 (because the whole miptree is
421 * organized in a 2D plane).
423 if (sctx
->chip_class
>= GFX9
&&
424 tex
->buffer
.b
.b
.last_level
> 0)
427 /* the clear is allowed if all layers are bound */
428 if (fb
->cbufs
[i
]->u
.tex
.first_layer
!= 0 ||
429 fb
->cbufs
[i
]->u
.tex
.last_layer
!= util_max_layer(&tex
->buffer
.b
.b
, 0)) {
433 /* only supported on tiled surfaces */
434 if (tex
->surface
.is_linear
) {
438 /* shared textures can't use fast clear without an explicit flush,
439 * because there is no way to communicate the clear color among
442 if (tex
->buffer
.b
.is_shared
&&
443 !(tex
->buffer
.external_usage
& PIPE_HANDLE_USAGE_EXPLICIT_FLUSH
))
446 if (sctx
->chip_class
<= VI
&&
447 tex
->surface
.u
.legacy
.level
[0].mode
== RADEON_SURF_MODE_1D
&&
448 !sctx
->screen
->info
.htile_cmask_support_1d_tiling
)
451 /* Use a slow clear for small surfaces where the cost of
452 * the eliminate pass can be higher than the benefit of fast
453 * clear. The closed driver does this, but the numbers may differ.
455 * This helps on both dGPUs and APUs, even small APUs like Mullins.
457 bool too_small
= tex
->buffer
.b
.b
.nr_samples
<= 1 &&
458 tex
->buffer
.b
.b
.width0
*
459 tex
->buffer
.b
.b
.height0
<= 512 * 512;
460 bool eliminate_needed
= false;
461 bool fmask_decompress_needed
= false;
463 /* Fast clear is the most appropriate place to enable DCC for
464 * displayable surfaces.
466 if (sctx
->family
== CHIP_STONEY
&& !too_small
) {
467 vi_separate_dcc_try_enable(sctx
, tex
);
469 /* RB+ isn't supported with a CMASK clear only on Stoney,
470 * so all clears are considered to be hypothetically slow
471 * clears, which is weighed when determining whether to
472 * enable separate DCC.
474 if (tex
->dcc_gather_statistics
) /* only for Stoney */
475 tex
->num_slow_clears
++;
478 /* Try to clear DCC first, otherwise try CMASK. */
479 if (vi_dcc_enabled(tex
, 0)) {
480 uint32_t reset_value
;
482 if (sctx
->screen
->debug_flags
& DBG(NO_DCC_CLEAR
))
485 /* This can happen with mipmapping or MSAA. */
486 if (sctx
->chip_class
== VI
&&
487 !tex
->surface
.u
.legacy
.level
[level
].dcc_fast_clear_size
)
490 if (!vi_get_fast_clear_parameters(tex
->buffer
.b
.b
.format
,
491 fb
->cbufs
[i
]->format
,
496 if (eliminate_needed
&& too_small
)
499 /* DCC fast clear with MSAA should clear CMASK to 0xC. */
500 if (tex
->buffer
.b
.b
.nr_samples
>= 2 && tex
->cmask_buffer
) {
501 /* TODO: This doesn't work with MSAA. */
502 if (eliminate_needed
)
505 uint32_t clear_value
= 0xCCCCCCCC;
506 si_clear_buffer(sctx
, &tex
->cmask_buffer
->b
.b
,
507 tex
->cmask_offset
, tex
->surface
.cmask_size
,
508 &clear_value
, 4, SI_COHERENCY_CB_META
);
509 fmask_decompress_needed
= true;
512 vi_dcc_clear_level(sctx
, tex
, 0, reset_value
);
513 tex
->separate_dcc_dirty
= true;
518 /* 128-bit formats are unusupported */
519 if (tex
->surface
.bpe
> 8) {
523 /* RB+ doesn't work with CMASK fast clear on Stoney. */
524 if (sctx
->family
== CHIP_STONEY
)
527 /* ensure CMASK is enabled */
528 si_alloc_separate_cmask(sctx
->screen
, tex
);
529 if (!tex
->cmask_buffer
)
532 /* Do the fast clear. */
533 uint32_t clear_value
= 0;
534 si_clear_buffer(sctx
, &tex
->cmask_buffer
->b
.b
,
535 tex
->cmask_offset
, tex
->surface
.cmask_size
,
536 &clear_value
, 4, SI_COHERENCY_CB_META
);
537 eliminate_needed
= true;
540 if ((eliminate_needed
|| fmask_decompress_needed
) &&
541 !(tex
->dirty_level_mask
& (1 << level
))) {
542 tex
->dirty_level_mask
|= 1 << level
;
543 p_atomic_inc(&sctx
->screen
->compressed_colortex_counter
);
546 /* We can change the micro tile mode before a full clear. */
547 si_set_optimal_micro_tile_mode(sctx
->screen
, tex
);
549 *buffers
&= ~clear_bit
;
551 /* Chips with DCC constant encoding don't need to set the clear
552 * color registers for DCC clear values 0 and 1.
554 if (sctx
->screen
->has_dcc_constant_encode
&& !eliminate_needed
)
557 if (si_set_clear_color(tex
, fb
->cbufs
[i
]->format
, color
)) {
558 sctx
->framebuffer
.dirty_cbufs
|= 1 << i
;
559 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.framebuffer
);
564 static void si_clear(struct pipe_context
*ctx
, unsigned buffers
,
565 const union pipe_color_union
*color
,
566 double depth
, unsigned stencil
)
568 struct si_context
*sctx
= (struct si_context
*)ctx
;
569 struct pipe_framebuffer_state
*fb
= &sctx
->framebuffer
.state
;
570 struct pipe_surface
*zsbuf
= fb
->zsbuf
;
571 struct si_texture
*zstex
=
572 zsbuf
? (struct si_texture
*)zsbuf
->texture
: NULL
;
574 if (buffers
& PIPE_CLEAR_COLOR
) {
575 si_do_fast_color_clear(sctx
, &buffers
, color
);
577 return; /* all buffers have been fast cleared */
579 /* These buffers cannot use fast clear, make sure to disable expansion. */
580 for (unsigned i
= 0; i
< fb
->nr_cbufs
; i
++) {
581 struct si_texture
*tex
;
583 /* If not clearing this buffer, skip. */
584 if (!(buffers
& (PIPE_CLEAR_COLOR0
<< i
)) || !fb
->cbufs
[i
])
587 tex
= (struct si_texture
*)fb
->cbufs
[i
]->texture
;
588 if (tex
->surface
.fmask_size
== 0)
589 tex
->dirty_level_mask
&= ~(1 << fb
->cbufs
[i
]->u
.tex
.level
);
594 si_htile_enabled(zstex
, zsbuf
->u
.tex
.level
) &&
595 zsbuf
->u
.tex
.first_layer
== 0 &&
596 zsbuf
->u
.tex
.last_layer
== util_max_layer(&zstex
->buffer
.b
.b
, 0)) {
597 /* TC-compatible HTILE only supports depth clears to 0 or 1. */
598 if (buffers
& PIPE_CLEAR_DEPTH
&&
599 (!zstex
->tc_compatible_htile
||
600 depth
== 0 || depth
== 1)) {
601 /* Need to disable EXPCLEAR temporarily if clearing
603 if (!zstex
->depth_cleared
|| zstex
->depth_clear_value
!= depth
) {
604 sctx
->db_depth_disable_expclear
= true;
607 if (zstex
->depth_clear_value
!= (float)depth
) {
608 /* Update DB_DEPTH_CLEAR. */
609 zstex
->depth_clear_value
= depth
;
610 sctx
->framebuffer
.dirty_zsbuf
= true;
611 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.framebuffer
);
613 sctx
->db_depth_clear
= true;
614 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.db_render_state
);
617 /* TC-compatible HTILE only supports stencil clears to 0. */
618 if (buffers
& PIPE_CLEAR_STENCIL
&&
619 (!zstex
->tc_compatible_htile
|| stencil
== 0)) {
622 /* Need to disable EXPCLEAR temporarily if clearing
624 if (!zstex
->stencil_cleared
|| zstex
->stencil_clear_value
!= stencil
) {
625 sctx
->db_stencil_disable_expclear
= true;
628 if (zstex
->stencil_clear_value
!= (uint8_t)stencil
) {
629 /* Update DB_STENCIL_CLEAR. */
630 zstex
->stencil_clear_value
= stencil
;
631 sctx
->framebuffer
.dirty_zsbuf
= true;
632 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.framebuffer
);
634 sctx
->db_stencil_clear
= true;
635 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.db_render_state
);
638 /* TODO: Find out what's wrong here. Fast depth clear leads to
639 * corruption in ARK: Survival Evolved, but that may just be
640 * a coincidence and the root cause is elsewhere.
642 * The corruption can be fixed by putting the DB flush before
643 * or after the depth clear. (surprisingly)
645 * https://bugs.freedesktop.org/show_bug.cgi?id=102955 (apitrace)
647 * This hack decreases back-to-back ClearDepth performance.
649 if ((sctx
->db_depth_clear
|| sctx
->db_stencil_clear
) &&
650 sctx
->screen
->clear_db_cache_before_clear
)
651 sctx
->flags
|= SI_CONTEXT_FLUSH_AND_INV_DB
;
654 si_blitter_begin(sctx
, SI_CLEAR
);
655 util_blitter_clear(sctx
->blitter
, fb
->width
, fb
->height
,
656 util_framebuffer_get_num_layers(fb
),
657 buffers
, color
, depth
, stencil
);
658 si_blitter_end(sctx
);
660 if (sctx
->db_depth_clear
) {
661 sctx
->db_depth_clear
= false;
662 sctx
->db_depth_disable_expclear
= false;
663 zstex
->depth_cleared
= true;
664 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.db_render_state
);
667 if (sctx
->db_stencil_clear
) {
668 sctx
->db_stencil_clear
= false;
669 sctx
->db_stencil_disable_expclear
= false;
670 zstex
->stencil_cleared
= true;
671 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.db_render_state
);
675 static void si_clear_render_target(struct pipe_context
*ctx
,
676 struct pipe_surface
*dst
,
677 const union pipe_color_union
*color
,
678 unsigned dstx
, unsigned dsty
,
679 unsigned width
, unsigned height
,
680 bool render_condition_enabled
)
682 struct si_context
*sctx
= (struct si_context
*)ctx
;
684 si_blitter_begin(sctx
, SI_CLEAR_SURFACE
|
685 (render_condition_enabled
? 0 : SI_DISABLE_RENDER_COND
));
686 util_blitter_clear_render_target(sctx
->blitter
, dst
, color
,
687 dstx
, dsty
, width
, height
);
688 si_blitter_end(sctx
);
691 static void si_clear_depth_stencil(struct pipe_context
*ctx
,
692 struct pipe_surface
*dst
,
693 unsigned clear_flags
,
696 unsigned dstx
, unsigned dsty
,
697 unsigned width
, unsigned height
,
698 bool render_condition_enabled
)
700 struct si_context
*sctx
= (struct si_context
*)ctx
;
702 si_blitter_begin(sctx
, SI_CLEAR_SURFACE
|
703 (render_condition_enabled
? 0 : SI_DISABLE_RENDER_COND
));
704 util_blitter_clear_depth_stencil(sctx
->blitter
, dst
, clear_flags
, depth
, stencil
,
705 dstx
, dsty
, width
, height
);
706 si_blitter_end(sctx
);
709 static void si_clear_texture(struct pipe_context
*pipe
,
710 struct pipe_resource
*tex
,
712 const struct pipe_box
*box
,
715 struct pipe_screen
*screen
= pipe
->screen
;
716 struct si_texture
*stex
= (struct si_texture
*)tex
;
717 struct pipe_surface tmpl
= {{0}};
718 struct pipe_surface
*sf
;
719 const struct util_format_description
*desc
=
720 util_format_description(tex
->format
);
722 tmpl
.format
= tex
->format
;
723 tmpl
.u
.tex
.first_layer
= box
->z
;
724 tmpl
.u
.tex
.last_layer
= box
->z
+ box
->depth
- 1;
725 tmpl
.u
.tex
.level
= level
;
726 sf
= pipe
->create_surface(pipe
, tex
, &tmpl
);
730 if (stex
->is_depth
) {
735 /* Depth is always present. */
736 clear
= PIPE_CLEAR_DEPTH
;
737 desc
->unpack_z_float(&depth
, 0, data
, 0, 1, 1);
739 if (stex
->surface
.has_stencil
) {
740 clear
|= PIPE_CLEAR_STENCIL
;
741 desc
->unpack_s_8uint(&stencil
, 0, data
, 0, 1, 1);
744 si_clear_depth_stencil(pipe
, sf
, clear
, depth
, stencil
,
746 box
->width
, box
->height
, false);
748 union pipe_color_union color
;
750 /* pipe_color_union requires the full vec4 representation. */
751 if (util_format_is_pure_uint(tex
->format
))
752 desc
->unpack_rgba_uint(color
.ui
, 0, data
, 0, 1, 1);
753 else if (util_format_is_pure_sint(tex
->format
))
754 desc
->unpack_rgba_sint(color
.i
, 0, data
, 0, 1, 1);
756 desc
->unpack_rgba_float(color
.f
, 0, data
, 0, 1, 1);
758 if (screen
->is_format_supported(screen
, tex
->format
,
760 PIPE_BIND_RENDER_TARGET
)) {
761 si_clear_render_target(pipe
, sf
, &color
,
763 box
->width
, box
->height
, false);
765 /* Software fallback - just for R9G9B9E5_FLOAT */
766 util_clear_render_target(pipe
, sf
, &color
,
768 box
->width
, box
->height
);
771 pipe_surface_reference(&sf
, NULL
);
774 void si_init_clear_functions(struct si_context
*sctx
)
776 sctx
->b
.clear
= si_clear
;
777 sctx
->b
.clear_render_target
= si_clear_render_target
;
778 sctx
->b
.clear_depth_stencil
= si_clear_depth_stencil
;
779 sctx
->b
.clear_texture
= si_clear_texture
;