2 * Copyright © 2011 Red Hat All Rights Reserved.
3 * Copyright © 2017 Advanced Micro Devices, Inc.
6 * Permission is hereby granted, free of charge, to any person obtaining
7 * a copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
15 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
16 * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
17 * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
18 * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
23 * The above copyright notice and this permission notice (including the
24 * next paragraph) shall be included in all copies or substantial portions
28 #include "ac_surface.h"
29 #include "amd_family.h"
30 #include "addrlib/src/amdgpu_asic_addr.h"
31 #include "ac_gpu_info.h"
32 #include "util/macros.h"
33 #include "util/u_atomic.h"
34 #include "util/u_math.h"
40 #include <amdgpu_drm.h>
42 #include "addrlib/inc/addrinterface.h"
44 #ifndef CIASICIDGFXENGINE_SOUTHERNISLAND
45 #define CIASICIDGFXENGINE_SOUTHERNISLAND 0x0000000A
48 #ifndef CIASICIDGFXENGINE_ARCTICISLAND
49 #define CIASICIDGFXENGINE_ARCTICISLAND 0x0000000D
52 static unsigned get_first(unsigned x
, unsigned y
)
57 static void addrlib_family_rev_id(enum radeon_family family
,
58 unsigned *addrlib_family
,
59 unsigned *addrlib_revid
)
63 *addrlib_family
= FAMILY_SI
;
64 *addrlib_revid
= get_first(AMDGPU_TAHITI_RANGE
);
67 *addrlib_family
= FAMILY_SI
;
68 *addrlib_revid
= get_first(AMDGPU_PITCAIRN_RANGE
);
71 *addrlib_family
= FAMILY_SI
;
72 *addrlib_revid
= get_first(AMDGPU_CAPEVERDE_RANGE
);
75 *addrlib_family
= FAMILY_SI
;
76 *addrlib_revid
= get_first(AMDGPU_OLAND_RANGE
);
79 *addrlib_family
= FAMILY_SI
;
80 *addrlib_revid
= get_first(AMDGPU_HAINAN_RANGE
);
83 *addrlib_family
= FAMILY_CI
;
84 *addrlib_revid
= get_first(AMDGPU_BONAIRE_RANGE
);
87 *addrlib_family
= FAMILY_KV
;
88 *addrlib_revid
= get_first(AMDGPU_SPECTRE_RANGE
);
91 *addrlib_family
= FAMILY_KV
;
92 *addrlib_revid
= get_first(AMDGPU_KALINDI_RANGE
);
95 *addrlib_family
= FAMILY_CI
;
96 *addrlib_revid
= get_first(AMDGPU_HAWAII_RANGE
);
99 *addrlib_family
= FAMILY_KV
;
100 *addrlib_revid
= get_first(AMDGPU_GODAVARI_RANGE
);
103 *addrlib_family
= FAMILY_VI
;
104 *addrlib_revid
= get_first(AMDGPU_TONGA_RANGE
);
107 *addrlib_family
= FAMILY_VI
;
108 *addrlib_revid
= get_first(AMDGPU_ICELAND_RANGE
);
111 *addrlib_family
= FAMILY_CZ
;
112 *addrlib_revid
= get_first(AMDGPU_CARRIZO_RANGE
);
115 *addrlib_family
= FAMILY_CZ
;
116 *addrlib_revid
= get_first(AMDGPU_STONEY_RANGE
);
119 *addrlib_family
= FAMILY_VI
;
120 *addrlib_revid
= get_first(AMDGPU_FIJI_RANGE
);
123 *addrlib_family
= FAMILY_VI
;
124 *addrlib_revid
= get_first(AMDGPU_POLARIS10_RANGE
);
127 *addrlib_family
= FAMILY_VI
;
128 *addrlib_revid
= get_first(AMDGPU_POLARIS11_RANGE
);
131 *addrlib_family
= FAMILY_VI
;
132 *addrlib_revid
= get_first(AMDGPU_POLARIS12_RANGE
);
135 *addrlib_family
= FAMILY_VI
;
136 *addrlib_revid
= get_first(AMDGPU_VEGAM_RANGE
);
139 *addrlib_family
= FAMILY_AI
;
140 *addrlib_revid
= get_first(AMDGPU_VEGA10_RANGE
);
143 *addrlib_family
= FAMILY_AI
;
144 *addrlib_revid
= get_first(AMDGPU_VEGA12_RANGE
);
147 *addrlib_family
= FAMILY_AI
;
148 *addrlib_revid
= get_first(AMDGPU_VEGA20_RANGE
);
151 *addrlib_family
= FAMILY_RV
;
152 *addrlib_revid
= get_first(AMDGPU_RAVEN_RANGE
);
155 *addrlib_family
= FAMILY_RV
;
156 *addrlib_revid
= get_first(AMDGPU_RAVEN2_RANGE
);
159 fprintf(stderr
, "amdgpu: Unknown family.\n");
163 static void *ADDR_API
allocSysMem(const ADDR_ALLOCSYSMEM_INPUT
* pInput
)
165 return malloc(pInput
->sizeInBytes
);
168 static ADDR_E_RETURNCODE ADDR_API
freeSysMem(const ADDR_FREESYSMEM_INPUT
* pInput
)
170 free(pInput
->pVirtAddr
);
174 ADDR_HANDLE
amdgpu_addr_create(const struct radeon_info
*info
,
175 const struct amdgpu_gpu_info
*amdinfo
,
176 uint64_t *max_alignment
)
178 ADDR_CREATE_INPUT addrCreateInput
= {0};
179 ADDR_CREATE_OUTPUT addrCreateOutput
= {0};
180 ADDR_REGISTER_VALUE regValue
= {0};
181 ADDR_CREATE_FLAGS createFlags
= {{0}};
182 ADDR_GET_MAX_ALINGMENTS_OUTPUT addrGetMaxAlignmentsOutput
= {0};
183 ADDR_E_RETURNCODE addrRet
;
185 addrCreateInput
.size
= sizeof(ADDR_CREATE_INPUT
);
186 addrCreateOutput
.size
= sizeof(ADDR_CREATE_OUTPUT
);
188 regValue
.gbAddrConfig
= amdinfo
->gb_addr_cfg
;
189 createFlags
.value
= 0;
191 addrlib_family_rev_id(info
->family
, &addrCreateInput
.chipFamily
, &addrCreateInput
.chipRevision
);
192 if (addrCreateInput
.chipFamily
== FAMILY_UNKNOWN
)
195 if (addrCreateInput
.chipFamily
>= FAMILY_AI
) {
196 addrCreateInput
.chipEngine
= CIASICIDGFXENGINE_ARCTICISLAND
;
197 regValue
.blockVarSizeLog2
= 0;
199 regValue
.noOfBanks
= amdinfo
->mc_arb_ramcfg
& 0x3;
200 regValue
.noOfRanks
= (amdinfo
->mc_arb_ramcfg
& 0x4) >> 2;
202 regValue
.backendDisables
= amdinfo
->enabled_rb_pipes_mask
;
203 regValue
.pTileConfig
= amdinfo
->gb_tile_mode
;
204 regValue
.noOfEntries
= ARRAY_SIZE(amdinfo
->gb_tile_mode
);
205 if (addrCreateInput
.chipFamily
== FAMILY_SI
) {
206 regValue
.pMacroTileConfig
= NULL
;
207 regValue
.noOfMacroEntries
= 0;
209 regValue
.pMacroTileConfig
= amdinfo
->gb_macro_tile_mode
;
210 regValue
.noOfMacroEntries
= ARRAY_SIZE(amdinfo
->gb_macro_tile_mode
);
213 createFlags
.useTileIndex
= 1;
214 createFlags
.useHtileSliceAlign
= 1;
216 addrCreateInput
.chipEngine
= CIASICIDGFXENGINE_SOUTHERNISLAND
;
219 addrCreateInput
.callbacks
.allocSysMem
= allocSysMem
;
220 addrCreateInput
.callbacks
.freeSysMem
= freeSysMem
;
221 addrCreateInput
.callbacks
.debugPrint
= 0;
222 addrCreateInput
.createFlags
= createFlags
;
223 addrCreateInput
.regValue
= regValue
;
225 addrRet
= AddrCreate(&addrCreateInput
, &addrCreateOutput
);
226 if (addrRet
!= ADDR_OK
)
230 addrRet
= AddrGetMaxAlignments(addrCreateOutput
.hLib
, &addrGetMaxAlignmentsOutput
);
231 if (addrRet
== ADDR_OK
){
232 *max_alignment
= addrGetMaxAlignmentsOutput
.baseAlign
;
235 return addrCreateOutput
.hLib
;
238 static int surf_config_sanity(const struct ac_surf_config
*config
,
241 /* FMASK is allocated together with the color surface and can't be
242 * allocated separately.
244 assert(!(flags
& RADEON_SURF_FMASK
));
245 if (flags
& RADEON_SURF_FMASK
)
248 /* all dimension must be at least 1 ! */
249 if (!config
->info
.width
|| !config
->info
.height
|| !config
->info
.depth
||
250 !config
->info
.array_size
|| !config
->info
.levels
)
253 switch (config
->info
.samples
) {
261 if (flags
& RADEON_SURF_Z_OR_SBUFFER
)
268 if (!(flags
& RADEON_SURF_Z_OR_SBUFFER
)) {
269 switch (config
->info
.storage_samples
) {
281 if (config
->is_3d
&& config
->info
.array_size
> 1)
283 if (config
->is_cube
&& config
->info
.depth
> 1)
289 static int gfx6_compute_level(ADDR_HANDLE addrlib
,
290 const struct ac_surf_config
*config
,
291 struct radeon_surf
*surf
, bool is_stencil
,
292 unsigned level
, bool compressed
,
293 ADDR_COMPUTE_SURFACE_INFO_INPUT
*AddrSurfInfoIn
,
294 ADDR_COMPUTE_SURFACE_INFO_OUTPUT
*AddrSurfInfoOut
,
295 ADDR_COMPUTE_DCCINFO_INPUT
*AddrDccIn
,
296 ADDR_COMPUTE_DCCINFO_OUTPUT
*AddrDccOut
,
297 ADDR_COMPUTE_HTILE_INFO_INPUT
*AddrHtileIn
,
298 ADDR_COMPUTE_HTILE_INFO_OUTPUT
*AddrHtileOut
)
300 struct legacy_surf_level
*surf_level
;
301 ADDR_E_RETURNCODE ret
;
303 AddrSurfInfoIn
->mipLevel
= level
;
304 AddrSurfInfoIn
->width
= u_minify(config
->info
.width
, level
);
305 AddrSurfInfoIn
->height
= u_minify(config
->info
.height
, level
);
307 /* Make GFX6 linear surfaces compatible with GFX9 for hybrid graphics,
308 * because GFX9 needs linear alignment of 256 bytes.
310 if (config
->info
.levels
== 1 &&
311 AddrSurfInfoIn
->tileMode
== ADDR_TM_LINEAR_ALIGNED
&&
312 AddrSurfInfoIn
->bpp
&&
313 util_is_power_of_two_or_zero(AddrSurfInfoIn
->bpp
)) {
314 unsigned alignment
= 256 / (AddrSurfInfoIn
->bpp
/ 8);
316 AddrSurfInfoIn
->width
= align(AddrSurfInfoIn
->width
, alignment
);
320 AddrSurfInfoIn
->numSlices
= u_minify(config
->info
.depth
, level
);
321 else if (config
->is_cube
)
322 AddrSurfInfoIn
->numSlices
= 6;
324 AddrSurfInfoIn
->numSlices
= config
->info
.array_size
;
327 /* Set the base level pitch. This is needed for calculation
328 * of non-zero levels. */
330 AddrSurfInfoIn
->basePitch
= surf
->u
.legacy
.stencil_level
[0].nblk_x
;
332 AddrSurfInfoIn
->basePitch
= surf
->u
.legacy
.level
[0].nblk_x
;
334 /* Convert blocks to pixels for compressed formats. */
336 AddrSurfInfoIn
->basePitch
*= surf
->blk_w
;
339 ret
= AddrComputeSurfaceInfo(addrlib
,
342 if (ret
!= ADDR_OK
) {
346 surf_level
= is_stencil
? &surf
->u
.legacy
.stencil_level
[level
] : &surf
->u
.legacy
.level
[level
];
347 surf_level
->offset
= align64(surf
->surf_size
, AddrSurfInfoOut
->baseAlign
);
348 surf_level
->slice_size_dw
= AddrSurfInfoOut
->sliceSize
/ 4;
349 surf_level
->nblk_x
= AddrSurfInfoOut
->pitch
;
350 surf_level
->nblk_y
= AddrSurfInfoOut
->height
;
352 switch (AddrSurfInfoOut
->tileMode
) {
353 case ADDR_TM_LINEAR_ALIGNED
:
354 surf_level
->mode
= RADEON_SURF_MODE_LINEAR_ALIGNED
;
356 case ADDR_TM_1D_TILED_THIN1
:
357 surf_level
->mode
= RADEON_SURF_MODE_1D
;
359 case ADDR_TM_2D_TILED_THIN1
:
360 surf_level
->mode
= RADEON_SURF_MODE_2D
;
367 surf
->u
.legacy
.stencil_tiling_index
[level
] = AddrSurfInfoOut
->tileIndex
;
369 surf
->u
.legacy
.tiling_index
[level
] = AddrSurfInfoOut
->tileIndex
;
371 surf
->surf_size
= surf_level
->offset
+ AddrSurfInfoOut
->surfSize
;
373 /* Clear DCC fields at the beginning. */
374 surf_level
->dcc_offset
= 0;
376 /* The previous level's flag tells us if we can use DCC for this level. */
377 if (AddrSurfInfoIn
->flags
.dccCompatible
&&
378 (level
== 0 || AddrDccOut
->subLvlCompressible
)) {
379 bool prev_level_clearable
= level
== 0 ||
380 AddrDccOut
->dccRamSizeAligned
;
382 AddrDccIn
->colorSurfSize
= AddrSurfInfoOut
->surfSize
;
383 AddrDccIn
->tileMode
= AddrSurfInfoOut
->tileMode
;
384 AddrDccIn
->tileInfo
= *AddrSurfInfoOut
->pTileInfo
;
385 AddrDccIn
->tileIndex
= AddrSurfInfoOut
->tileIndex
;
386 AddrDccIn
->macroModeIndex
= AddrSurfInfoOut
->macroModeIndex
;
388 ret
= AddrComputeDccInfo(addrlib
,
392 if (ret
== ADDR_OK
) {
393 surf_level
->dcc_offset
= surf
->dcc_size
;
394 surf
->num_dcc_levels
= level
+ 1;
395 surf
->dcc_size
= surf_level
->dcc_offset
+ AddrDccOut
->dccRamSize
;
396 surf
->dcc_alignment
= MAX2(surf
->dcc_alignment
, AddrDccOut
->dccRamBaseAlign
);
398 /* If the DCC size of a subresource (1 mip level or 1 slice)
399 * is not aligned, the DCC memory layout is not contiguous for
400 * that subresource, which means we can't use fast clear.
402 * We only do fast clears for whole mipmap levels. If we did
403 * per-slice fast clears, the same restriction would apply.
404 * (i.e. only compute the slice size and see if it's aligned)
406 * The last level can be non-contiguous and still be clearable
407 * if it's interleaved with the next level that doesn't exist.
409 if (AddrDccOut
->dccRamSizeAligned
||
410 (prev_level_clearable
&& level
== config
->info
.levels
- 1))
411 surf_level
->dcc_fast_clear_size
= AddrDccOut
->dccFastClearSize
;
413 surf_level
->dcc_fast_clear_size
= 0;
417 /* TC-compatible HTILE. */
419 AddrSurfInfoIn
->flags
.depth
&&
420 surf_level
->mode
== RADEON_SURF_MODE_2D
&&
422 AddrHtileIn
->flags
.tcCompatible
= AddrSurfInfoIn
->flags
.tcCompatible
;
423 AddrHtileIn
->pitch
= AddrSurfInfoOut
->pitch
;
424 AddrHtileIn
->height
= AddrSurfInfoOut
->height
;
425 AddrHtileIn
->numSlices
= AddrSurfInfoOut
->depth
;
426 AddrHtileIn
->blockWidth
= ADDR_HTILE_BLOCKSIZE_8
;
427 AddrHtileIn
->blockHeight
= ADDR_HTILE_BLOCKSIZE_8
;
428 AddrHtileIn
->pTileInfo
= AddrSurfInfoOut
->pTileInfo
;
429 AddrHtileIn
->tileIndex
= AddrSurfInfoOut
->tileIndex
;
430 AddrHtileIn
->macroModeIndex
= AddrSurfInfoOut
->macroModeIndex
;
432 ret
= AddrComputeHtileInfo(addrlib
,
436 if (ret
== ADDR_OK
) {
437 surf
->htile_size
= AddrHtileOut
->htileBytes
;
438 surf
->htile_slice_size
= AddrHtileOut
->sliceSize
;
439 surf
->htile_alignment
= AddrHtileOut
->baseAlign
;
446 #define G_009910_MICRO_TILE_MODE(x) (((x) >> 0) & 0x03)
447 #define V_009910_ADDR_SURF_THICK_MICRO_TILING 0x03
448 #define G_009910_MICRO_TILE_MODE_NEW(x) (((x) >> 22) & 0x07)
450 static void gfx6_set_micro_tile_mode(struct radeon_surf
*surf
,
451 const struct radeon_info
*info
)
453 uint32_t tile_mode
= info
->si_tile_mode_array
[surf
->u
.legacy
.tiling_index
[0]];
455 if (info
->chip_class
>= CIK
)
456 surf
->micro_tile_mode
= G_009910_MICRO_TILE_MODE_NEW(tile_mode
);
458 surf
->micro_tile_mode
= G_009910_MICRO_TILE_MODE(tile_mode
);
461 static unsigned cik_get_macro_tile_index(struct radeon_surf
*surf
)
463 unsigned index
, tileb
;
465 tileb
= 8 * 8 * surf
->bpe
;
466 tileb
= MIN2(surf
->u
.legacy
.tile_split
, tileb
);
468 for (index
= 0; tileb
> 64; index
++)
475 static bool get_display_flag(const struct ac_surf_config
*config
,
476 const struct radeon_surf
*surf
)
478 unsigned num_channels
= config
->info
.num_channels
;
479 unsigned bpe
= surf
->bpe
;
481 if (surf
->flags
& RADEON_SURF_SCANOUT
&&
482 config
->info
.samples
<= 1 &&
483 surf
->blk_w
<= 2 && surf
->blk_h
== 1) {
485 if (surf
->blk_w
== 2 && surf
->blk_h
== 1)
488 if (/* RGBA8 or RGBA16F */
489 (bpe
>= 4 && bpe
<= 8 && num_channels
== 4) ||
490 /* R5G6B5 or R5G5B5A1 */
491 (bpe
== 2 && num_channels
>= 3) ||
493 (bpe
== 1 && num_channels
== 1))
500 * This must be called after the first level is computed.
502 * Copy surface-global settings like pipe/bank config from level 0 surface
503 * computation, and compute tile swizzle.
505 static int gfx6_surface_settings(ADDR_HANDLE addrlib
,
506 const struct radeon_info
*info
,
507 const struct ac_surf_config
*config
,
508 ADDR_COMPUTE_SURFACE_INFO_OUTPUT
* csio
,
509 struct radeon_surf
*surf
)
511 surf
->surf_alignment
= csio
->baseAlign
;
512 surf
->u
.legacy
.pipe_config
= csio
->pTileInfo
->pipeConfig
- 1;
513 gfx6_set_micro_tile_mode(surf
, info
);
515 /* For 2D modes only. */
516 if (csio
->tileMode
>= ADDR_TM_2D_TILED_THIN1
) {
517 surf
->u
.legacy
.bankw
= csio
->pTileInfo
->bankWidth
;
518 surf
->u
.legacy
.bankh
= csio
->pTileInfo
->bankHeight
;
519 surf
->u
.legacy
.mtilea
= csio
->pTileInfo
->macroAspectRatio
;
520 surf
->u
.legacy
.tile_split
= csio
->pTileInfo
->tileSplitBytes
;
521 surf
->u
.legacy
.num_banks
= csio
->pTileInfo
->banks
;
522 surf
->u
.legacy
.macro_tile_index
= csio
->macroModeIndex
;
524 surf
->u
.legacy
.macro_tile_index
= 0;
527 /* Compute tile swizzle. */
528 /* TODO: fix tile swizzle with mipmapping for SI */
529 if ((info
->chip_class
>= CIK
|| config
->info
.levels
== 1) &&
530 config
->info
.surf_index
&&
531 surf
->u
.legacy
.level
[0].mode
== RADEON_SURF_MODE_2D
&&
532 !(surf
->flags
& (RADEON_SURF_Z_OR_SBUFFER
| RADEON_SURF_SHAREABLE
)) &&
533 !get_display_flag(config
, surf
)) {
534 ADDR_COMPUTE_BASE_SWIZZLE_INPUT AddrBaseSwizzleIn
= {0};
535 ADDR_COMPUTE_BASE_SWIZZLE_OUTPUT AddrBaseSwizzleOut
= {0};
537 AddrBaseSwizzleIn
.size
= sizeof(ADDR_COMPUTE_BASE_SWIZZLE_INPUT
);
538 AddrBaseSwizzleOut
.size
= sizeof(ADDR_COMPUTE_BASE_SWIZZLE_OUTPUT
);
540 AddrBaseSwizzleIn
.surfIndex
= p_atomic_inc_return(config
->info
.surf_index
) - 1;
541 AddrBaseSwizzleIn
.tileIndex
= csio
->tileIndex
;
542 AddrBaseSwizzleIn
.macroModeIndex
= csio
->macroModeIndex
;
543 AddrBaseSwizzleIn
.pTileInfo
= csio
->pTileInfo
;
544 AddrBaseSwizzleIn
.tileMode
= csio
->tileMode
;
546 int r
= AddrComputeBaseSwizzle(addrlib
, &AddrBaseSwizzleIn
,
547 &AddrBaseSwizzleOut
);
551 assert(AddrBaseSwizzleOut
.tileSwizzle
<=
552 u_bit_consecutive(0, sizeof(surf
->tile_swizzle
) * 8));
553 surf
->tile_swizzle
= AddrBaseSwizzleOut
.tileSwizzle
;
558 void ac_compute_cmask(const struct radeon_info
*info
,
559 const struct ac_surf_config
*config
,
560 struct radeon_surf
*surf
)
562 unsigned pipe_interleave_bytes
= info
->pipe_interleave_bytes
;
563 unsigned num_pipes
= info
->num_tile_pipes
;
564 unsigned cl_width
, cl_height
;
566 if (surf
->flags
& RADEON_SURF_Z_OR_SBUFFER
)
569 assert(info
->chip_class
<= VI
);
584 case 16: /* Hawaii */
593 unsigned base_align
= num_pipes
* pipe_interleave_bytes
;
595 unsigned width
= align(surf
->u
.legacy
.level
[0].nblk_x
, cl_width
*8);
596 unsigned height
= align(surf
->u
.legacy
.level
[0].nblk_y
, cl_height
*8);
597 unsigned slice_elements
= (width
* height
) / (8*8);
599 /* Each element of CMASK is a nibble. */
600 unsigned slice_bytes
= slice_elements
/ 2;
602 surf
->u
.legacy
.cmask_slice_tile_max
= (width
* height
) / (128*128);
603 if (surf
->u
.legacy
.cmask_slice_tile_max
)
604 surf
->u
.legacy
.cmask_slice_tile_max
-= 1;
608 num_layers
= config
->info
.depth
;
609 else if (config
->is_cube
)
612 num_layers
= config
->info
.array_size
;
614 surf
->cmask_alignment
= MAX2(256, base_align
);
615 surf
->cmask_size
= align(slice_bytes
, base_align
) * num_layers
;
619 * Fill in the tiling information in \p surf based on the given surface config.
621 * The following fields of \p surf must be initialized by the caller:
622 * blk_w, blk_h, bpe, flags.
624 static int gfx6_compute_surface(ADDR_HANDLE addrlib
,
625 const struct radeon_info
*info
,
626 const struct ac_surf_config
*config
,
627 enum radeon_surf_mode mode
,
628 struct radeon_surf
*surf
)
632 ADDR_COMPUTE_SURFACE_INFO_INPUT AddrSurfInfoIn
= {0};
633 ADDR_COMPUTE_SURFACE_INFO_OUTPUT AddrSurfInfoOut
= {0};
634 ADDR_COMPUTE_DCCINFO_INPUT AddrDccIn
= {0};
635 ADDR_COMPUTE_DCCINFO_OUTPUT AddrDccOut
= {0};
636 ADDR_COMPUTE_HTILE_INFO_INPUT AddrHtileIn
= {0};
637 ADDR_COMPUTE_HTILE_INFO_OUTPUT AddrHtileOut
= {0};
638 ADDR_TILEINFO AddrTileInfoIn
= {0};
639 ADDR_TILEINFO AddrTileInfoOut
= {0};
642 AddrSurfInfoIn
.size
= sizeof(ADDR_COMPUTE_SURFACE_INFO_INPUT
);
643 AddrSurfInfoOut
.size
= sizeof(ADDR_COMPUTE_SURFACE_INFO_OUTPUT
);
644 AddrDccIn
.size
= sizeof(ADDR_COMPUTE_DCCINFO_INPUT
);
645 AddrDccOut
.size
= sizeof(ADDR_COMPUTE_DCCINFO_OUTPUT
);
646 AddrHtileIn
.size
= sizeof(ADDR_COMPUTE_HTILE_INFO_INPUT
);
647 AddrHtileOut
.size
= sizeof(ADDR_COMPUTE_HTILE_INFO_OUTPUT
);
648 AddrSurfInfoOut
.pTileInfo
= &AddrTileInfoOut
;
650 compressed
= surf
->blk_w
== 4 && surf
->blk_h
== 4;
652 /* MSAA requires 2D tiling. */
653 if (config
->info
.samples
> 1)
654 mode
= RADEON_SURF_MODE_2D
;
656 /* DB doesn't support linear layouts. */
657 if (surf
->flags
& (RADEON_SURF_Z_OR_SBUFFER
) &&
658 mode
< RADEON_SURF_MODE_1D
)
659 mode
= RADEON_SURF_MODE_1D
;
661 /* Set the requested tiling mode. */
663 case RADEON_SURF_MODE_LINEAR_ALIGNED
:
664 AddrSurfInfoIn
.tileMode
= ADDR_TM_LINEAR_ALIGNED
;
666 case RADEON_SURF_MODE_1D
:
667 AddrSurfInfoIn
.tileMode
= ADDR_TM_1D_TILED_THIN1
;
669 case RADEON_SURF_MODE_2D
:
670 AddrSurfInfoIn
.tileMode
= ADDR_TM_2D_TILED_THIN1
;
676 /* The format must be set correctly for the allocation of compressed
677 * textures to work. In other cases, setting the bpp is sufficient.
682 AddrSurfInfoIn
.format
= ADDR_FMT_BC1
;
685 AddrSurfInfoIn
.format
= ADDR_FMT_BC3
;
692 AddrDccIn
.bpp
= AddrSurfInfoIn
.bpp
= surf
->bpe
* 8;
695 AddrDccIn
.numSamples
= AddrSurfInfoIn
.numSamples
=
696 MAX2(1, config
->info
.samples
);
697 AddrSurfInfoIn
.tileIndex
= -1;
699 if (!(surf
->flags
& RADEON_SURF_Z_OR_SBUFFER
)) {
700 AddrDccIn
.numSamples
= AddrSurfInfoIn
.numFrags
=
701 MAX2(1, config
->info
.storage_samples
);
704 /* Set the micro tile type. */
705 if (surf
->flags
& RADEON_SURF_SCANOUT
)
706 AddrSurfInfoIn
.tileType
= ADDR_DISPLAYABLE
;
707 else if (surf
->flags
& RADEON_SURF_Z_OR_SBUFFER
)
708 AddrSurfInfoIn
.tileType
= ADDR_DEPTH_SAMPLE_ORDER
;
710 AddrSurfInfoIn
.tileType
= ADDR_NON_DISPLAYABLE
;
712 AddrSurfInfoIn
.flags
.color
= !(surf
->flags
& RADEON_SURF_Z_OR_SBUFFER
);
713 AddrSurfInfoIn
.flags
.depth
= (surf
->flags
& RADEON_SURF_ZBUFFER
) != 0;
714 AddrSurfInfoIn
.flags
.cube
= config
->is_cube
;
715 AddrSurfInfoIn
.flags
.display
= get_display_flag(config
, surf
);
716 AddrSurfInfoIn
.flags
.pow2Pad
= config
->info
.levels
> 1;
717 AddrSurfInfoIn
.flags
.tcCompatible
= (surf
->flags
& RADEON_SURF_TC_COMPATIBLE_HTILE
) != 0;
719 /* Only degrade the tile mode for space if TC-compatible HTILE hasn't been
720 * requested, because TC-compatible HTILE requires 2D tiling.
722 AddrSurfInfoIn
.flags
.opt4Space
= !AddrSurfInfoIn
.flags
.tcCompatible
&&
723 !AddrSurfInfoIn
.flags
.fmask
&&
724 config
->info
.samples
<= 1 &&
725 (surf
->flags
& RADEON_SURF_OPTIMIZE_FOR_SPACE
);
728 * - If we add MSAA support, keep in mind that CB can't decompress 8bpp
730 * - Mipmapped array textures have low performance (discovered by a closed
733 AddrSurfInfoIn
.flags
.dccCompatible
=
734 info
->chip_class
>= VI
&&
735 !(surf
->flags
& RADEON_SURF_Z_OR_SBUFFER
) &&
736 !(surf
->flags
& RADEON_SURF_DISABLE_DCC
) &&
738 ((config
->info
.array_size
== 1 && config
->info
.depth
== 1) ||
739 config
->info
.levels
== 1);
741 AddrSurfInfoIn
.flags
.noStencil
= (surf
->flags
& RADEON_SURF_SBUFFER
) == 0;
742 AddrSurfInfoIn
.flags
.compressZ
= !!(surf
->flags
& RADEON_SURF_Z_OR_SBUFFER
);
744 /* On CI/VI, the DB uses the same pitch and tile mode (except tilesplit)
745 * for Z and stencil. This can cause a number of problems which we work
748 * - a depth part that is incompatible with mipmapped texturing
749 * - at least on Stoney, entirely incompatible Z/S aspects (e.g.
750 * incorrect tiling applied to the stencil part, stencil buffer
751 * memory accesses that go out of bounds) even without mipmapping
753 * Some piglit tests that are prone to different types of related
755 * ./bin/ext_framebuffer_multisample-upsample 2 stencil
756 * ./bin/framebuffer-blit-levels {draw,read} stencil
757 * ./bin/ext_framebuffer_multisample-unaligned-blit N {depth,stencil} {msaa,upsample,downsample}
758 * ./bin/fbo-depth-array fs-writes-{depth,stencil} / {depth,stencil}-{clear,layered-clear,draw}
759 * ./bin/depthstencil-render-miplevels 1024 d=s=z24_s8
761 int stencil_tile_idx
= -1;
763 if (AddrSurfInfoIn
.flags
.depth
&& !AddrSurfInfoIn
.flags
.noStencil
&&
764 (config
->info
.levels
> 1 || info
->family
== CHIP_STONEY
)) {
765 /* Compute stencilTileIdx that is compatible with the (depth)
766 * tileIdx. This degrades the depth surface if necessary to
767 * ensure that a matching stencilTileIdx exists. */
768 AddrSurfInfoIn
.flags
.matchStencilTileCfg
= 1;
770 /* Keep the depth mip-tail compatible with texturing. */
771 AddrSurfInfoIn
.flags
.noStencil
= 1;
774 /* Set preferred macrotile parameters. This is usually required
775 * for shared resources. This is for 2D tiling only. */
776 if (AddrSurfInfoIn
.tileMode
>= ADDR_TM_2D_TILED_THIN1
&&
777 surf
->u
.legacy
.bankw
&& surf
->u
.legacy
.bankh
&&
778 surf
->u
.legacy
.mtilea
&& surf
->u
.legacy
.tile_split
) {
779 /* If any of these parameters are incorrect, the calculation
781 AddrTileInfoIn
.banks
= surf
->u
.legacy
.num_banks
;
782 AddrTileInfoIn
.bankWidth
= surf
->u
.legacy
.bankw
;
783 AddrTileInfoIn
.bankHeight
= surf
->u
.legacy
.bankh
;
784 AddrTileInfoIn
.macroAspectRatio
= surf
->u
.legacy
.mtilea
;
785 AddrTileInfoIn
.tileSplitBytes
= surf
->u
.legacy
.tile_split
;
786 AddrTileInfoIn
.pipeConfig
= surf
->u
.legacy
.pipe_config
+ 1; /* +1 compared to GB_TILE_MODE */
787 AddrSurfInfoIn
.flags
.opt4Space
= 0;
788 AddrSurfInfoIn
.pTileInfo
= &AddrTileInfoIn
;
790 /* If AddrSurfInfoIn.pTileInfo is set, Addrlib doesn't set
791 * the tile index, because we are expected to know it if
792 * we know the other parameters.
794 * This is something that can easily be fixed in Addrlib.
795 * For now, just figure it out here.
796 * Note that only 2D_TILE_THIN1 is handled here.
798 assert(!(surf
->flags
& RADEON_SURF_Z_OR_SBUFFER
));
799 assert(AddrSurfInfoIn
.tileMode
== ADDR_TM_2D_TILED_THIN1
);
801 if (info
->chip_class
== SI
) {
802 if (AddrSurfInfoIn
.tileType
== ADDR_DISPLAYABLE
) {
804 AddrSurfInfoIn
.tileIndex
= 11; /* 16bpp */
806 AddrSurfInfoIn
.tileIndex
= 12; /* 32bpp */
809 AddrSurfInfoIn
.tileIndex
= 14; /* 8bpp */
810 else if (surf
->bpe
== 2)
811 AddrSurfInfoIn
.tileIndex
= 15; /* 16bpp */
812 else if (surf
->bpe
== 4)
813 AddrSurfInfoIn
.tileIndex
= 16; /* 32bpp */
815 AddrSurfInfoIn
.tileIndex
= 17; /* 64bpp (and 128bpp) */
819 if (AddrSurfInfoIn
.tileType
== ADDR_DISPLAYABLE
)
820 AddrSurfInfoIn
.tileIndex
= 10; /* 2D displayable */
822 AddrSurfInfoIn
.tileIndex
= 14; /* 2D non-displayable */
824 /* Addrlib doesn't set this if tileIndex is forced like above. */
825 AddrSurfInfoOut
.macroModeIndex
= cik_get_macro_tile_index(surf
);
829 surf
->has_stencil
= !!(surf
->flags
& RADEON_SURF_SBUFFER
);
830 surf
->num_dcc_levels
= 0;
833 surf
->dcc_alignment
= 1;
834 surf
->htile_size
= 0;
835 surf
->htile_slice_size
= 0;
836 surf
->htile_alignment
= 1;
838 const bool only_stencil
= (surf
->flags
& RADEON_SURF_SBUFFER
) &&
839 !(surf
->flags
& RADEON_SURF_ZBUFFER
);
841 /* Calculate texture layout information. */
843 for (level
= 0; level
< config
->info
.levels
; level
++) {
844 r
= gfx6_compute_level(addrlib
, config
, surf
, false, level
, compressed
,
845 &AddrSurfInfoIn
, &AddrSurfInfoOut
,
846 &AddrDccIn
, &AddrDccOut
, &AddrHtileIn
, &AddrHtileOut
);
853 /* Check that we actually got a TC-compatible HTILE if
854 * we requested it (only for level 0, since we're not
855 * supporting HTILE on higher mip levels anyway). */
856 assert(AddrSurfInfoOut
.tcCompatible
||
857 !AddrSurfInfoIn
.flags
.tcCompatible
||
858 AddrSurfInfoIn
.flags
.matchStencilTileCfg
);
860 if (AddrSurfInfoIn
.flags
.matchStencilTileCfg
) {
861 if (!AddrSurfInfoOut
.tcCompatible
) {
862 AddrSurfInfoIn
.flags
.tcCompatible
= 0;
863 surf
->flags
&= ~RADEON_SURF_TC_COMPATIBLE_HTILE
;
866 AddrSurfInfoIn
.flags
.matchStencilTileCfg
= 0;
867 AddrSurfInfoIn
.tileIndex
= AddrSurfInfoOut
.tileIndex
;
868 stencil_tile_idx
= AddrSurfInfoOut
.stencilTileIdx
;
870 assert(stencil_tile_idx
>= 0);
873 r
= gfx6_surface_settings(addrlib
, info
, config
,
874 &AddrSurfInfoOut
, surf
);
880 /* Calculate texture layout information for stencil. */
881 if (surf
->flags
& RADEON_SURF_SBUFFER
) {
882 AddrSurfInfoIn
.tileIndex
= stencil_tile_idx
;
883 AddrSurfInfoIn
.bpp
= 8;
884 AddrSurfInfoIn
.flags
.depth
= 0;
885 AddrSurfInfoIn
.flags
.stencil
= 1;
886 AddrSurfInfoIn
.flags
.tcCompatible
= 0;
887 /* This will be ignored if AddrSurfInfoIn.pTileInfo is NULL. */
888 AddrTileInfoIn
.tileSplitBytes
= surf
->u
.legacy
.stencil_tile_split
;
890 for (level
= 0; level
< config
->info
.levels
; level
++) {
891 r
= gfx6_compute_level(addrlib
, config
, surf
, true, level
, compressed
,
892 &AddrSurfInfoIn
, &AddrSurfInfoOut
,
893 &AddrDccIn
, &AddrDccOut
,
898 /* DB uses the depth pitch for both stencil and depth. */
900 if (surf
->u
.legacy
.stencil_level
[level
].nblk_x
!=
901 surf
->u
.legacy
.level
[level
].nblk_x
)
902 surf
->u
.legacy
.stencil_adjusted
= true;
904 surf
->u
.legacy
.level
[level
].nblk_x
=
905 surf
->u
.legacy
.stencil_level
[level
].nblk_x
;
910 r
= gfx6_surface_settings(addrlib
, info
, config
,
911 &AddrSurfInfoOut
, surf
);
916 /* For 2D modes only. */
917 if (AddrSurfInfoOut
.tileMode
>= ADDR_TM_2D_TILED_THIN1
) {
918 surf
->u
.legacy
.stencil_tile_split
=
919 AddrSurfInfoOut
.pTileInfo
->tileSplitBytes
;
926 if (config
->info
.samples
>= 2 && AddrSurfInfoIn
.flags
.color
) {
927 ADDR_COMPUTE_FMASK_INFO_INPUT fin
= {0};
928 ADDR_COMPUTE_FMASK_INFO_OUTPUT fout
= {0};
929 ADDR_TILEINFO fmask_tile_info
= {};
931 fin
.size
= sizeof(fin
);
932 fout
.size
= sizeof(fout
);
934 fin
.tileMode
= AddrSurfInfoOut
.tileMode
;
935 fin
.pitch
= AddrSurfInfoOut
.pitch
;
936 fin
.height
= config
->info
.height
;
937 fin
.numSlices
= AddrSurfInfoIn
.numSlices
;
938 fin
.numSamples
= AddrSurfInfoIn
.numSamples
;
939 fin
.numFrags
= AddrSurfInfoIn
.numFrags
;
941 fout
.pTileInfo
= &fmask_tile_info
;
943 r
= AddrComputeFmaskInfo(addrlib
, &fin
, &fout
);
947 surf
->fmask_size
= fout
.fmaskBytes
;
948 surf
->fmask_alignment
= fout
.baseAlign
;
949 surf
->fmask_tile_swizzle
= 0;
951 surf
->u
.legacy
.fmask
.slice_tile_max
=
952 (fout
.pitch
* fout
.height
) / 64;
953 if (surf
->u
.legacy
.fmask
.slice_tile_max
)
954 surf
->u
.legacy
.fmask
.slice_tile_max
-= 1;
956 surf
->u
.legacy
.fmask
.tiling_index
= fout
.tileIndex
;
957 surf
->u
.legacy
.fmask
.bankh
= fout
.pTileInfo
->bankHeight
;
958 surf
->u
.legacy
.fmask
.pitch_in_pixels
= fout
.pitch
;
960 /* Compute tile swizzle for FMASK. */
961 if (config
->info
.fmask_surf_index
&&
962 !(surf
->flags
& RADEON_SURF_SHAREABLE
)) {
963 ADDR_COMPUTE_BASE_SWIZZLE_INPUT xin
= {0};
964 ADDR_COMPUTE_BASE_SWIZZLE_OUTPUT xout
= {0};
966 xin
.size
= sizeof(ADDR_COMPUTE_BASE_SWIZZLE_INPUT
);
967 xout
.size
= sizeof(ADDR_COMPUTE_BASE_SWIZZLE_OUTPUT
);
969 /* This counter starts from 1 instead of 0. */
970 xin
.surfIndex
= p_atomic_inc_return(config
->info
.fmask_surf_index
);
971 xin
.tileIndex
= fout
.tileIndex
;
972 xin
.macroModeIndex
= fout
.macroModeIndex
;
973 xin
.pTileInfo
= fout
.pTileInfo
;
974 xin
.tileMode
= fin
.tileMode
;
976 int r
= AddrComputeBaseSwizzle(addrlib
, &xin
, &xout
);
980 assert(xout
.tileSwizzle
<=
981 u_bit_consecutive(0, sizeof(surf
->tile_swizzle
) * 8));
982 surf
->fmask_tile_swizzle
= xout
.tileSwizzle
;
986 /* Recalculate the whole DCC miptree size including disabled levels.
987 * This is what addrlib does, but calling addrlib would be a lot more
990 if (surf
->dcc_size
&& config
->info
.levels
> 1) {
991 /* The smallest miplevels that are never compressed by DCC
992 * still read the DCC buffer via TC if the base level uses DCC,
993 * and for some reason the DCC buffer needs to be larger if
994 * the miptree uses non-zero tile_swizzle. Otherwise there are
997 * "dcc_alignment * 4" was determined by trial and error.
999 surf
->dcc_size
= align64(surf
->surf_size
>> 8,
1000 surf
->dcc_alignment
* 4);
1003 /* Make sure HTILE covers the whole miptree, because the shader reads
1004 * TC-compatible HTILE even for levels where it's disabled by DB.
1006 if (surf
->htile_size
&& config
->info
.levels
> 1 &&
1007 surf
->flags
& RADEON_SURF_TC_COMPATIBLE_HTILE
) {
1008 /* MSAA can't occur with levels > 1, so ignore the sample count. */
1009 const unsigned total_pixels
= surf
->surf_size
/ surf
->bpe
;
1010 const unsigned htile_block_size
= 8 * 8;
1011 const unsigned htile_element_size
= 4;
1013 surf
->htile_size
= (total_pixels
/ htile_block_size
) *
1015 surf
->htile_size
= align(surf
->htile_size
, surf
->htile_alignment
);
1018 surf
->is_linear
= surf
->u
.legacy
.level
[0].mode
== RADEON_SURF_MODE_LINEAR_ALIGNED
;
1019 surf
->is_displayable
= surf
->is_linear
||
1020 surf
->micro_tile_mode
== RADEON_MICRO_MODE_DISPLAY
||
1021 surf
->micro_tile_mode
== RADEON_MICRO_MODE_ROTATED
;
1023 /* The rotated micro tile mode doesn't work if both CMASK and RB+ are
1024 * used at the same time. This case is not currently expected to occur
1025 * because we don't use rotated. Enforce this restriction on all chips
1026 * to facilitate testing.
1028 if (surf
->micro_tile_mode
== RADEON_MICRO_MODE_ROTATED
) {
1029 assert(!"rotate micro tile mode is unsupported");
1033 ac_compute_cmask(info
, config
, surf
);
1037 /* This is only called when expecting a tiled layout. */
1039 gfx9_get_preferred_swizzle_mode(ADDR_HANDLE addrlib
,
1040 ADDR2_COMPUTE_SURFACE_INFO_INPUT
*in
,
1041 bool is_fmask
, AddrSwizzleMode
*swizzle_mode
)
1043 ADDR_E_RETURNCODE ret
;
1044 ADDR2_GET_PREFERRED_SURF_SETTING_INPUT sin
= {0};
1045 ADDR2_GET_PREFERRED_SURF_SETTING_OUTPUT sout
= {0};
1047 sin
.size
= sizeof(ADDR2_GET_PREFERRED_SURF_SETTING_INPUT
);
1048 sout
.size
= sizeof(ADDR2_GET_PREFERRED_SURF_SETTING_OUTPUT
);
1050 sin
.flags
= in
->flags
;
1051 sin
.resourceType
= in
->resourceType
;
1052 sin
.format
= in
->format
;
1053 sin
.resourceLoction
= ADDR_RSRC_LOC_INVIS
;
1054 /* TODO: We could allow some of these: */
1055 sin
.forbiddenBlock
.micro
= 1; /* don't allow the 256B swizzle modes */
1056 sin
.forbiddenBlock
.var
= 1; /* don't allow the variable-sized swizzle modes */
1057 sin
.forbiddenBlock
.linear
= 1; /* don't allow linear swizzle modes */
1059 sin
.width
= in
->width
;
1060 sin
.height
= in
->height
;
1061 sin
.numSlices
= in
->numSlices
;
1062 sin
.numMipLevels
= in
->numMipLevels
;
1063 sin
.numSamples
= in
->numSamples
;
1064 sin
.numFrags
= in
->numFrags
;
1067 sin
.flags
.display
= 0;
1068 sin
.flags
.color
= 0;
1069 sin
.flags
.fmask
= 1;
1072 ret
= Addr2GetPreferredSurfaceSetting(addrlib
, &sin
, &sout
);
1076 *swizzle_mode
= sout
.swizzleMode
;
1080 static int gfx9_compute_miptree(ADDR_HANDLE addrlib
,
1081 const struct ac_surf_config
*config
,
1082 struct radeon_surf
*surf
, bool compressed
,
1083 ADDR2_COMPUTE_SURFACE_INFO_INPUT
*in
)
1085 ADDR2_MIP_INFO mip_info
[RADEON_SURF_MAX_LEVELS
] = {};
1086 ADDR2_COMPUTE_SURFACE_INFO_OUTPUT out
= {0};
1087 ADDR_E_RETURNCODE ret
;
1089 out
.size
= sizeof(ADDR2_COMPUTE_SURFACE_INFO_OUTPUT
);
1090 out
.pMipInfo
= mip_info
;
1092 ret
= Addr2ComputeSurfaceInfo(addrlib
, in
, &out
);
1096 if (in
->flags
.stencil
) {
1097 surf
->u
.gfx9
.stencil
.swizzle_mode
= in
->swizzleMode
;
1098 surf
->u
.gfx9
.stencil
.epitch
= out
.epitchIsHeight
? out
.mipChainHeight
- 1 :
1099 out
.mipChainPitch
- 1;
1100 surf
->surf_alignment
= MAX2(surf
->surf_alignment
, out
.baseAlign
);
1101 surf
->u
.gfx9
.stencil_offset
= align(surf
->surf_size
, out
.baseAlign
);
1102 surf
->surf_size
= surf
->u
.gfx9
.stencil_offset
+ out
.surfSize
;
1106 surf
->u
.gfx9
.surf
.swizzle_mode
= in
->swizzleMode
;
1107 surf
->u
.gfx9
.surf
.epitch
= out
.epitchIsHeight
? out
.mipChainHeight
- 1 :
1108 out
.mipChainPitch
- 1;
1110 /* CMASK fast clear uses these even if FMASK isn't allocated.
1111 * FMASK only supports the Z swizzle modes, whose numbers are multiples of 4.
1113 surf
->u
.gfx9
.fmask
.swizzle_mode
= surf
->u
.gfx9
.surf
.swizzle_mode
& ~0x3;
1114 surf
->u
.gfx9
.fmask
.epitch
= surf
->u
.gfx9
.surf
.epitch
;
1116 surf
->u
.gfx9
.surf_slice_size
= out
.sliceSize
;
1117 surf
->u
.gfx9
.surf_pitch
= out
.pitch
;
1118 surf
->u
.gfx9
.surf_height
= out
.height
;
1119 surf
->surf_size
= out
.surfSize
;
1120 surf
->surf_alignment
= out
.baseAlign
;
1122 if (in
->swizzleMode
== ADDR_SW_LINEAR
) {
1123 for (unsigned i
= 0; i
< in
->numMipLevels
; i
++)
1124 surf
->u
.gfx9
.offset
[i
] = mip_info
[i
].offset
;
1127 if (in
->flags
.depth
) {
1128 assert(in
->swizzleMode
!= ADDR_SW_LINEAR
);
1131 ADDR2_COMPUTE_HTILE_INFO_INPUT hin
= {0};
1132 ADDR2_COMPUTE_HTILE_INFO_OUTPUT hout
= {0};
1134 hin
.size
= sizeof(ADDR2_COMPUTE_HTILE_INFO_INPUT
);
1135 hout
.size
= sizeof(ADDR2_COMPUTE_HTILE_INFO_OUTPUT
);
1137 hin
.hTileFlags
.pipeAligned
= !in
->flags
.metaPipeUnaligned
;
1138 hin
.hTileFlags
.rbAligned
= !in
->flags
.metaRbUnaligned
;
1139 hin
.depthFlags
= in
->flags
;
1140 hin
.swizzleMode
= in
->swizzleMode
;
1141 hin
.unalignedWidth
= in
->width
;
1142 hin
.unalignedHeight
= in
->height
;
1143 hin
.numSlices
= in
->numSlices
;
1144 hin
.numMipLevels
= in
->numMipLevels
;
1145 hin
.firstMipIdInTail
= out
.firstMipIdInTail
;
1147 ret
= Addr2ComputeHtileInfo(addrlib
, &hin
, &hout
);
1151 surf
->u
.gfx9
.htile
.rb_aligned
= hin
.hTileFlags
.rbAligned
;
1152 surf
->u
.gfx9
.htile
.pipe_aligned
= hin
.hTileFlags
.pipeAligned
;
1153 surf
->htile_size
= hout
.htileBytes
;
1154 surf
->htile_slice_size
= hout
.sliceSize
;
1155 surf
->htile_alignment
= hout
.baseAlign
;
1157 /* Compute tile swizzle for the color surface.
1158 * All *_X and *_T modes can use the swizzle.
1160 if (config
->info
.surf_index
&&
1161 in
->swizzleMode
>= ADDR_SW_64KB_Z_T
&&
1162 !out
.mipChainInTail
&&
1163 !(surf
->flags
& RADEON_SURF_SHAREABLE
) &&
1164 !in
->flags
.display
) {
1165 ADDR2_COMPUTE_PIPEBANKXOR_INPUT xin
= {0};
1166 ADDR2_COMPUTE_PIPEBANKXOR_OUTPUT xout
= {0};
1168 xin
.size
= sizeof(ADDR2_COMPUTE_PIPEBANKXOR_INPUT
);
1169 xout
.size
= sizeof(ADDR2_COMPUTE_PIPEBANKXOR_OUTPUT
);
1171 xin
.surfIndex
= p_atomic_inc_return(config
->info
.surf_index
) - 1;
1172 xin
.flags
= in
->flags
;
1173 xin
.swizzleMode
= in
->swizzleMode
;
1174 xin
.resourceType
= in
->resourceType
;
1175 xin
.format
= in
->format
;
1176 xin
.numSamples
= in
->numSamples
;
1177 xin
.numFrags
= in
->numFrags
;
1179 ret
= Addr2ComputePipeBankXor(addrlib
, &xin
, &xout
);
1183 assert(xout
.pipeBankXor
<=
1184 u_bit_consecutive(0, sizeof(surf
->tile_swizzle
) * 8));
1185 surf
->tile_swizzle
= xout
.pipeBankXor
;
1189 if (!(surf
->flags
& RADEON_SURF_DISABLE_DCC
) &&
1191 in
->swizzleMode
!= ADDR_SW_LINEAR
) {
1192 ADDR2_COMPUTE_DCCINFO_INPUT din
= {0};
1193 ADDR2_COMPUTE_DCCINFO_OUTPUT dout
= {0};
1194 ADDR2_META_MIP_INFO meta_mip_info
[RADEON_SURF_MAX_LEVELS
] = {};
1196 din
.size
= sizeof(ADDR2_COMPUTE_DCCINFO_INPUT
);
1197 dout
.size
= sizeof(ADDR2_COMPUTE_DCCINFO_OUTPUT
);
1198 dout
.pMipInfo
= meta_mip_info
;
1200 din
.dccKeyFlags
.pipeAligned
= !in
->flags
.metaPipeUnaligned
;
1201 din
.dccKeyFlags
.rbAligned
= !in
->flags
.metaRbUnaligned
;
1202 din
.colorFlags
= in
->flags
;
1203 din
.resourceType
= in
->resourceType
;
1204 din
.swizzleMode
= in
->swizzleMode
;
1206 din
.unalignedWidth
= in
->width
;
1207 din
.unalignedHeight
= in
->height
;
1208 din
.numSlices
= in
->numSlices
;
1209 din
.numFrags
= in
->numFrags
;
1210 din
.numMipLevels
= in
->numMipLevels
;
1211 din
.dataSurfaceSize
= out
.surfSize
;
1212 din
.firstMipIdInTail
= out
.firstMipIdInTail
;
1214 ret
= Addr2ComputeDccInfo(addrlib
, &din
, &dout
);
1218 surf
->u
.gfx9
.dcc
.rb_aligned
= din
.dccKeyFlags
.rbAligned
;
1219 surf
->u
.gfx9
.dcc
.pipe_aligned
= din
.dccKeyFlags
.pipeAligned
;
1220 surf
->u
.gfx9
.dcc_pitch_max
= dout
.pitch
- 1;
1221 surf
->dcc_size
= dout
.dccRamSize
;
1222 surf
->dcc_alignment
= dout
.dccRamBaseAlign
;
1223 surf
->num_dcc_levels
= in
->numMipLevels
;
1225 /* Disable DCC for levels that are in the mip tail.
1227 * There are two issues that this is intended to
1230 * 1. Multiple mip levels may share a cache line. This
1231 * can lead to corruption when switching between
1232 * rendering to different mip levels because the
1233 * RBs don't maintain coherency.
1235 * 2. Texturing with metadata after rendering sometimes
1236 * fails with corruption, probably for a similar
1239 * Working around these issues for all levels in the
1240 * mip tail may be overly conservative, but it's what
1243 * Alternative solutions that also work but are worse:
1244 * - Disable DCC entirely.
1245 * - Flush TC L2 after rendering.
1247 for (unsigned i
= 0; i
< in
->numMipLevels
; i
++) {
1248 if (meta_mip_info
[i
].inMiptail
) {
1249 surf
->num_dcc_levels
= i
;
1254 if (!surf
->num_dcc_levels
)
1259 if (in
->numSamples
> 1) {
1260 ADDR2_COMPUTE_FMASK_INFO_INPUT fin
= {0};
1261 ADDR2_COMPUTE_FMASK_INFO_OUTPUT fout
= {0};
1263 fin
.size
= sizeof(ADDR2_COMPUTE_FMASK_INFO_INPUT
);
1264 fout
.size
= sizeof(ADDR2_COMPUTE_FMASK_INFO_OUTPUT
);
1266 ret
= gfx9_get_preferred_swizzle_mode(addrlib
, in
,
1267 true, &fin
.swizzleMode
);
1271 fin
.unalignedWidth
= in
->width
;
1272 fin
.unalignedHeight
= in
->height
;
1273 fin
.numSlices
= in
->numSlices
;
1274 fin
.numSamples
= in
->numSamples
;
1275 fin
.numFrags
= in
->numFrags
;
1277 ret
= Addr2ComputeFmaskInfo(addrlib
, &fin
, &fout
);
1281 surf
->u
.gfx9
.fmask
.swizzle_mode
= fin
.swizzleMode
;
1282 surf
->u
.gfx9
.fmask
.epitch
= fout
.pitch
- 1;
1283 surf
->fmask_size
= fout
.fmaskBytes
;
1284 surf
->fmask_alignment
= fout
.baseAlign
;
1286 /* Compute tile swizzle for the FMASK surface. */
1287 if (config
->info
.fmask_surf_index
&&
1288 fin
.swizzleMode
>= ADDR_SW_64KB_Z_T
&&
1289 !(surf
->flags
& RADEON_SURF_SHAREABLE
)) {
1290 ADDR2_COMPUTE_PIPEBANKXOR_INPUT xin
= {0};
1291 ADDR2_COMPUTE_PIPEBANKXOR_OUTPUT xout
= {0};
1293 xin
.size
= sizeof(ADDR2_COMPUTE_PIPEBANKXOR_INPUT
);
1294 xout
.size
= sizeof(ADDR2_COMPUTE_PIPEBANKXOR_OUTPUT
);
1296 /* This counter starts from 1 instead of 0. */
1297 xin
.surfIndex
= p_atomic_inc_return(config
->info
.fmask_surf_index
);
1298 xin
.flags
= in
->flags
;
1299 xin
.swizzleMode
= fin
.swizzleMode
;
1300 xin
.resourceType
= in
->resourceType
;
1301 xin
.format
= in
->format
;
1302 xin
.numSamples
= in
->numSamples
;
1303 xin
.numFrags
= in
->numFrags
;
1305 ret
= Addr2ComputePipeBankXor(addrlib
, &xin
, &xout
);
1309 assert(xout
.pipeBankXor
<=
1310 u_bit_consecutive(0, sizeof(surf
->fmask_tile_swizzle
) * 8));
1311 surf
->fmask_tile_swizzle
= xout
.pipeBankXor
;
1316 if (in
->swizzleMode
!= ADDR_SW_LINEAR
) {
1317 ADDR2_COMPUTE_CMASK_INFO_INPUT cin
= {0};
1318 ADDR2_COMPUTE_CMASK_INFO_OUTPUT cout
= {0};
1320 cin
.size
= sizeof(ADDR2_COMPUTE_CMASK_INFO_INPUT
);
1321 cout
.size
= sizeof(ADDR2_COMPUTE_CMASK_INFO_OUTPUT
);
1323 if (in
->numSamples
> 1) {
1324 /* FMASK is always aligned. */
1325 cin
.cMaskFlags
.pipeAligned
= 1;
1326 cin
.cMaskFlags
.rbAligned
= 1;
1328 cin
.cMaskFlags
.pipeAligned
= !in
->flags
.metaPipeUnaligned
;
1329 cin
.cMaskFlags
.rbAligned
= !in
->flags
.metaRbUnaligned
;
1331 cin
.colorFlags
= in
->flags
;
1332 cin
.resourceType
= in
->resourceType
;
1333 cin
.unalignedWidth
= in
->width
;
1334 cin
.unalignedHeight
= in
->height
;
1335 cin
.numSlices
= in
->numSlices
;
1337 if (in
->numSamples
> 1)
1338 cin
.swizzleMode
= surf
->u
.gfx9
.fmask
.swizzle_mode
;
1340 cin
.swizzleMode
= in
->swizzleMode
;
1342 ret
= Addr2ComputeCmaskInfo(addrlib
, &cin
, &cout
);
1346 surf
->u
.gfx9
.cmask
.rb_aligned
= cin
.cMaskFlags
.rbAligned
;
1347 surf
->u
.gfx9
.cmask
.pipe_aligned
= cin
.cMaskFlags
.pipeAligned
;
1348 surf
->cmask_size
= cout
.cmaskBytes
;
1349 surf
->cmask_alignment
= cout
.baseAlign
;
1356 static int gfx9_compute_surface(ADDR_HANDLE addrlib
,
1357 const struct radeon_info
*info
,
1358 const struct ac_surf_config
*config
,
1359 enum radeon_surf_mode mode
,
1360 struct radeon_surf
*surf
)
1363 ADDR2_COMPUTE_SURFACE_INFO_INPUT AddrSurfInfoIn
= {0};
1366 AddrSurfInfoIn
.size
= sizeof(ADDR2_COMPUTE_SURFACE_INFO_INPUT
);
1368 compressed
= surf
->blk_w
== 4 && surf
->blk_h
== 4;
1370 /* The format must be set correctly for the allocation of compressed
1371 * textures to work. In other cases, setting the bpp is sufficient. */
1373 switch (surf
->bpe
) {
1375 AddrSurfInfoIn
.format
= ADDR_FMT_BC1
;
1378 AddrSurfInfoIn
.format
= ADDR_FMT_BC3
;
1384 switch (surf
->bpe
) {
1386 assert(!(surf
->flags
& RADEON_SURF_ZBUFFER
));
1387 AddrSurfInfoIn
.format
= ADDR_FMT_8
;
1390 assert(surf
->flags
& RADEON_SURF_ZBUFFER
||
1391 !(surf
->flags
& RADEON_SURF_SBUFFER
));
1392 AddrSurfInfoIn
.format
= ADDR_FMT_16
;
1395 assert(surf
->flags
& RADEON_SURF_ZBUFFER
||
1396 !(surf
->flags
& RADEON_SURF_SBUFFER
));
1397 AddrSurfInfoIn
.format
= ADDR_FMT_32
;
1400 assert(!(surf
->flags
& RADEON_SURF_Z_OR_SBUFFER
));
1401 AddrSurfInfoIn
.format
= ADDR_FMT_32_32
;
1404 assert(!(surf
->flags
& RADEON_SURF_Z_OR_SBUFFER
));
1405 AddrSurfInfoIn
.format
= ADDR_FMT_32_32_32
;
1408 assert(!(surf
->flags
& RADEON_SURF_Z_OR_SBUFFER
));
1409 AddrSurfInfoIn
.format
= ADDR_FMT_32_32_32_32
;
1414 AddrSurfInfoIn
.bpp
= surf
->bpe
* 8;
1417 bool is_color_surface
= !(surf
->flags
& RADEON_SURF_Z_OR_SBUFFER
);
1418 AddrSurfInfoIn
.flags
.color
= is_color_surface
&&
1419 !(surf
->flags
& RADEON_SURF_NO_RENDER_TARGET
);
1420 AddrSurfInfoIn
.flags
.depth
= (surf
->flags
& RADEON_SURF_ZBUFFER
) != 0;
1421 AddrSurfInfoIn
.flags
.display
= get_display_flag(config
, surf
);
1422 /* flags.texture currently refers to TC-compatible HTILE */
1423 AddrSurfInfoIn
.flags
.texture
= is_color_surface
||
1424 surf
->flags
& RADEON_SURF_TC_COMPATIBLE_HTILE
;
1425 AddrSurfInfoIn
.flags
.opt4space
= 1;
1427 AddrSurfInfoIn
.numMipLevels
= config
->info
.levels
;
1428 AddrSurfInfoIn
.numSamples
= MAX2(1, config
->info
.samples
);
1429 AddrSurfInfoIn
.numFrags
= AddrSurfInfoIn
.numSamples
;
1431 if (!(surf
->flags
& RADEON_SURF_Z_OR_SBUFFER
))
1432 AddrSurfInfoIn
.numFrags
= MAX2(1, config
->info
.storage_samples
);
1434 /* GFX9 doesn't support 1D depth textures, so allocate all 1D textures
1435 * as 2D to avoid having shader variants for 1D vs 2D, so all shaders
1436 * must sample 1D textures as 2D. */
1438 AddrSurfInfoIn
.resourceType
= ADDR_RSRC_TEX_3D
;
1440 AddrSurfInfoIn
.resourceType
= ADDR_RSRC_TEX_2D
;
1442 AddrSurfInfoIn
.width
= config
->info
.width
;
1443 AddrSurfInfoIn
.height
= config
->info
.height
;
1446 AddrSurfInfoIn
.numSlices
= config
->info
.depth
;
1447 else if (config
->is_cube
)
1448 AddrSurfInfoIn
.numSlices
= 6;
1450 AddrSurfInfoIn
.numSlices
= config
->info
.array_size
;
1452 /* This is propagated to HTILE/DCC/CMASK. */
1453 AddrSurfInfoIn
.flags
.metaPipeUnaligned
= 0;
1454 AddrSurfInfoIn
.flags
.metaRbUnaligned
= 0;
1457 case RADEON_SURF_MODE_LINEAR_ALIGNED
:
1458 assert(config
->info
.samples
<= 1);
1459 assert(!(surf
->flags
& RADEON_SURF_Z_OR_SBUFFER
));
1460 AddrSurfInfoIn
.swizzleMode
= ADDR_SW_LINEAR
;
1463 case RADEON_SURF_MODE_1D
:
1464 case RADEON_SURF_MODE_2D
:
1465 if (surf
->flags
& RADEON_SURF_IMPORTED
) {
1466 AddrSurfInfoIn
.swizzleMode
= surf
->u
.gfx9
.surf
.swizzle_mode
;
1470 r
= gfx9_get_preferred_swizzle_mode(addrlib
, &AddrSurfInfoIn
,
1471 false, &AddrSurfInfoIn
.swizzleMode
);
1480 surf
->u
.gfx9
.resource_type
= AddrSurfInfoIn
.resourceType
;
1481 surf
->has_stencil
= !!(surf
->flags
& RADEON_SURF_SBUFFER
);
1483 surf
->num_dcc_levels
= 0;
1484 surf
->surf_size
= 0;
1485 surf
->fmask_size
= 0;
1487 surf
->htile_size
= 0;
1488 surf
->htile_slice_size
= 0;
1489 surf
->u
.gfx9
.surf_offset
= 0;
1490 surf
->u
.gfx9
.stencil_offset
= 0;
1491 surf
->cmask_size
= 0;
1493 /* Calculate texture layout information. */
1494 r
= gfx9_compute_miptree(addrlib
, config
, surf
, compressed
,
1499 /* Calculate texture layout information for stencil. */
1500 if (surf
->flags
& RADEON_SURF_SBUFFER
) {
1501 AddrSurfInfoIn
.flags
.stencil
= 1;
1502 AddrSurfInfoIn
.bpp
= 8;
1503 AddrSurfInfoIn
.format
= ADDR_FMT_8
;
1505 if (!AddrSurfInfoIn
.flags
.depth
) {
1506 r
= gfx9_get_preferred_swizzle_mode(addrlib
, &AddrSurfInfoIn
,
1507 false, &AddrSurfInfoIn
.swizzleMode
);
1511 AddrSurfInfoIn
.flags
.depth
= 0;
1513 r
= gfx9_compute_miptree(addrlib
, config
, surf
, compressed
,
1519 surf
->is_linear
= surf
->u
.gfx9
.surf
.swizzle_mode
== ADDR_SW_LINEAR
;
1521 /* Query whether the surface is displayable. */
1522 bool displayable
= false;
1523 if (!config
->is_3d
&& !config
->is_cube
) {
1524 r
= Addr2IsValidDisplaySwizzleMode(addrlib
, surf
->u
.gfx9
.surf
.swizzle_mode
,
1525 surf
->bpe
* 8, &displayable
);
1529 surf
->is_displayable
= displayable
;
1531 switch (surf
->u
.gfx9
.surf
.swizzle_mode
) {
1533 case ADDR_SW_256B_S
:
1535 case ADDR_SW_64KB_S
:
1537 case ADDR_SW_64KB_S_T
:
1538 case ADDR_SW_4KB_S_X
:
1539 case ADDR_SW_64KB_S_X
:
1540 case ADDR_SW_VAR_S_X
:
1541 surf
->micro_tile_mode
= RADEON_MICRO_MODE_THIN
;
1545 case ADDR_SW_LINEAR
:
1546 case ADDR_SW_256B_D
:
1548 case ADDR_SW_64KB_D
:
1550 case ADDR_SW_64KB_D_T
:
1551 case ADDR_SW_4KB_D_X
:
1552 case ADDR_SW_64KB_D_X
:
1553 case ADDR_SW_VAR_D_X
:
1554 surf
->micro_tile_mode
= RADEON_MICRO_MODE_DISPLAY
;
1558 case ADDR_SW_256B_R
:
1560 case ADDR_SW_64KB_R
:
1562 case ADDR_SW_64KB_R_T
:
1563 case ADDR_SW_4KB_R_X
:
1564 case ADDR_SW_64KB_R_X
:
1565 case ADDR_SW_VAR_R_X
:
1566 /* The rotated micro tile mode doesn't work if both CMASK and RB+ are
1567 * used at the same time. This case is not currently expected to occur
1568 * because we don't use rotated. Enforce this restriction on all chips
1569 * to facilitate testing.
1571 assert(!"rotate micro tile mode is unsupported");
1576 case ADDR_SW_64KB_Z
:
1578 case ADDR_SW_64KB_Z_T
:
1579 case ADDR_SW_4KB_Z_X
:
1580 case ADDR_SW_64KB_Z_X
:
1581 case ADDR_SW_VAR_Z_X
:
1582 surf
->micro_tile_mode
= RADEON_MICRO_MODE_DEPTH
;
1592 int ac_compute_surface(ADDR_HANDLE addrlib
, const struct radeon_info
*info
,
1593 const struct ac_surf_config
*config
,
1594 enum radeon_surf_mode mode
,
1595 struct radeon_surf
*surf
)
1599 r
= surf_config_sanity(config
, surf
->flags
);
1603 if (info
->chip_class
>= GFX9
)
1604 return gfx9_compute_surface(addrlib
, info
, config
, mode
, surf
);
1606 return gfx6_compute_surface(addrlib
, info
, config
, mode
, surf
);