2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
5 * based in part on anv driver which is:
6 * Copyright © 2015 Intel Corporation
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 #include "radv_debug.h"
29 #include "radv_private.h"
30 #include "vk_format.h"
32 #include "radv_radeon_winsys.h"
34 #include "util/debug.h"
35 #include "util/u_atomic.h"
38 radv_choose_tiling(struct radv_device
*device
,
39 const VkImageCreateInfo
*pCreateInfo
,
42 if (pCreateInfo
->tiling
== VK_IMAGE_TILING_LINEAR
) {
43 assert(pCreateInfo
->samples
<= 1);
44 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
47 if (!vk_format_is_compressed(format
) &&
48 !vk_format_is_depth_or_stencil(format
)
49 && device
->physical_device
->rad_info
.chip_class
<= GFX8
) {
50 /* this causes hangs in some VK CTS tests on GFX9. */
51 /* Textures with a very small height are recommended to be linear. */
52 if (pCreateInfo
->imageType
== VK_IMAGE_TYPE_1D
||
53 /* Only very thin and long 2D textures should benefit from
55 (pCreateInfo
->extent
.width
> 8 && pCreateInfo
->extent
.height
<= 2))
56 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
59 /* MSAA resources must be 2D tiled. */
60 if (pCreateInfo
->samples
> 1)
61 return RADEON_SURF_MODE_2D
;
63 return RADEON_SURF_MODE_2D
;
67 radv_use_tc_compat_htile_for_image(struct radv_device
*device
,
68 const VkImageCreateInfo
*pCreateInfo
,
71 /* TC-compat HTILE is only available for GFX8+. */
72 if (device
->physical_device
->rad_info
.chip_class
< GFX8
)
75 if ((pCreateInfo
->usage
& VK_IMAGE_USAGE_STORAGE_BIT
))
78 if (pCreateInfo
->tiling
== VK_IMAGE_TILING_LINEAR
)
81 if (pCreateInfo
->mipLevels
> 1)
84 /* Do not enable TC-compatible HTILE if the image isn't readable by a
85 * shader because no texture fetches will happen.
87 if (!(pCreateInfo
->usage
& (VK_IMAGE_USAGE_SAMPLED_BIT
|
88 VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT
|
89 VK_IMAGE_USAGE_TRANSFER_SRC_BIT
)))
92 /* FIXME: for some reason TC compat with 2/4/8 samples breaks some cts
93 * tests - disable for now. On GFX10 D32_SFLOAT is affected as well.
95 if (pCreateInfo
->samples
>= 2 &&
96 (format
== VK_FORMAT_D32_SFLOAT_S8_UINT
||
97 (format
== VK_FORMAT_D32_SFLOAT
&&
98 device
->physical_device
->rad_info
.chip_class
== GFX10
)))
101 /* GFX9 supports both 32-bit and 16-bit depth surfaces, while GFX8 only
102 * supports 32-bit. Though, it's possible to enable TC-compat for
103 * 16-bit depth surfaces if no Z planes are compressed.
105 if (format
!= VK_FORMAT_D32_SFLOAT_S8_UINT
&&
106 format
!= VK_FORMAT_D32_SFLOAT
&&
107 format
!= VK_FORMAT_D16_UNORM
)
110 if (pCreateInfo
->flags
& VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT
) {
111 const struct VkImageFormatListCreateInfo
*format_list
=
112 (const struct VkImageFormatListCreateInfo
*)
113 vk_find_struct_const(pCreateInfo
->pNext
,
114 IMAGE_FORMAT_LIST_CREATE_INFO
);
116 /* We have to ignore the existence of the list if viewFormatCount = 0 */
117 if (format_list
&& format_list
->viewFormatCount
) {
118 /* compatibility is transitive, so we only need to check
119 * one format with everything else.
121 for (unsigned i
= 0; i
< format_list
->viewFormatCount
; ++i
) {
122 if (format_list
->pViewFormats
[i
] == VK_FORMAT_UNDEFINED
)
125 if (format
!= format_list
->pViewFormats
[i
])
137 radv_surface_has_scanout(struct radv_device
*device
, const struct radv_image_create_info
*info
)
139 if (info
->bo_metadata
) {
140 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
)
141 return info
->bo_metadata
->u
.gfx9
.scanout
;
143 return info
->bo_metadata
->u
.legacy
.scanout
;
146 return info
->scanout
;
150 radv_use_dcc_for_image(struct radv_device
*device
,
151 const struct radv_image
*image
,
152 const VkImageCreateInfo
*pCreateInfo
,
155 bool dcc_compatible_formats
;
158 /* DCC (Delta Color Compression) is only available for GFX8+. */
159 if (device
->physical_device
->rad_info
.chip_class
< GFX8
)
162 if (device
->instance
->debug_flags
& RADV_DEBUG_NO_DCC
)
165 if (image
->shareable
)
168 /* TODO: Enable DCC for storage images. */
169 if ((pCreateInfo
->usage
& VK_IMAGE_USAGE_STORAGE_BIT
))
172 if (pCreateInfo
->tiling
== VK_IMAGE_TILING_LINEAR
)
175 if (vk_format_is_subsampled(format
) ||
176 vk_format_get_plane_count(format
) > 1)
179 /* TODO: Enable DCC for mipmaps on GFX9+. */
180 if ((pCreateInfo
->arrayLayers
> 1 || pCreateInfo
->mipLevels
> 1) &&
181 device
->physical_device
->rad_info
.chip_class
>= GFX9
)
184 /* Do not enable DCC for mipmapped arrays because performance is worse. */
185 if (pCreateInfo
->arrayLayers
> 1 && pCreateInfo
->mipLevels
> 1)
188 /* FIXME: DCC for MSAA with 4x and 8x samples doesn't work yet, while
189 * 2x can be enabled with an option.
191 if (pCreateInfo
->samples
> 2 ||
192 (pCreateInfo
->samples
== 2 &&
193 !device
->physical_device
->dcc_msaa_allowed
))
196 /* Determine if the formats are DCC compatible. */
197 dcc_compatible_formats
=
198 radv_is_colorbuffer_format_supported(format
,
201 if (pCreateInfo
->flags
& VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT
) {
202 const struct VkImageFormatListCreateInfo
*format_list
=
203 (const struct VkImageFormatListCreateInfo
*)
204 vk_find_struct_const(pCreateInfo
->pNext
,
205 IMAGE_FORMAT_LIST_CREATE_INFO
);
207 /* We have to ignore the existence of the list if viewFormatCount = 0 */
208 if (format_list
&& format_list
->viewFormatCount
) {
209 /* compatibility is transitive, so we only need to check
210 * one format with everything else. */
211 for (unsigned i
= 0; i
< format_list
->viewFormatCount
; ++i
) {
212 if (format_list
->pViewFormats
[i
] == VK_FORMAT_UNDEFINED
)
215 if (!radv_dcc_formats_compatible(format
,
216 format_list
->pViewFormats
[i
]))
217 dcc_compatible_formats
= false;
220 dcc_compatible_formats
= false;
224 if (!dcc_compatible_formats
)
231 radv_use_tc_compat_cmask_for_image(struct radv_device
*device
,
232 struct radv_image
*image
)
234 if (!(device
->instance
->perftest_flags
& RADV_PERFTEST_TC_COMPAT_CMASK
))
237 /* TC-compat CMASK is only available for GFX8+. */
238 if (device
->physical_device
->rad_info
.chip_class
< GFX8
)
241 if (image
->usage
& VK_IMAGE_USAGE_STORAGE_BIT
)
244 if (radv_image_has_dcc(image
))
247 if (!radv_image_has_cmask(image
))
253 static uint32_t si_get_bo_metadata_word1(const struct radv_device
*device
)
255 return (ATI_VENDOR_ID
<< 16) | device
->physical_device
->rad_info
.pci_id
;
259 radv_is_valid_opaque_metadata(const struct radv_device
*device
,
260 const struct radeon_bo_metadata
*md
)
262 if (md
->metadata
[0] != 1 ||
263 md
->metadata
[1] != si_get_bo_metadata_word1(device
))
266 if (md
->size_metadata
< 40)
273 radv_patch_surface_from_metadata(struct radv_device
*device
,
274 struct radeon_surf
*surface
,
275 const struct radeon_bo_metadata
*md
)
277 surface
->flags
= RADEON_SURF_CLR(surface
->flags
, MODE
);
279 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
280 if (md
->u
.gfx9
.swizzle_mode
> 0)
281 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_MODE_2D
, MODE
);
283 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_MODE_LINEAR_ALIGNED
, MODE
);
285 surface
->u
.gfx9
.surf
.swizzle_mode
= md
->u
.gfx9
.swizzle_mode
;
287 surface
->u
.legacy
.pipe_config
= md
->u
.legacy
.pipe_config
;
288 surface
->u
.legacy
.bankw
= md
->u
.legacy
.bankw
;
289 surface
->u
.legacy
.bankh
= md
->u
.legacy
.bankh
;
290 surface
->u
.legacy
.tile_split
= md
->u
.legacy
.tile_split
;
291 surface
->u
.legacy
.mtilea
= md
->u
.legacy
.mtilea
;
292 surface
->u
.legacy
.num_banks
= md
->u
.legacy
.num_banks
;
294 if (md
->u
.legacy
.macrotile
== RADEON_LAYOUT_TILED
)
295 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_MODE_2D
, MODE
);
296 else if (md
->u
.legacy
.microtile
== RADEON_LAYOUT_TILED
)
297 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_MODE_1D
, MODE
);
299 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_MODE_LINEAR_ALIGNED
, MODE
);
305 radv_patch_image_dimensions(struct radv_device
*device
,
306 struct radv_image
*image
,
307 const struct radv_image_create_info
*create_info
,
308 struct ac_surf_info
*image_info
)
310 unsigned width
= image
->info
.width
;
311 unsigned height
= image
->info
.height
;
314 * minigbm sometimes allocates bigger images which is going to result in
315 * weird strides and other properties. Lets be lenient where possible and
316 * fail it on GFX10 (as we cannot cope there).
318 * Example hack: https://chromium-review.googlesource.com/c/chromiumos/platform/minigbm/+/1457777/
320 if (create_info
->bo_metadata
&&
321 radv_is_valid_opaque_metadata(device
, create_info
->bo_metadata
)) {
322 const struct radeon_bo_metadata
*md
= create_info
->bo_metadata
;
324 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
325 width
= G_00A004_WIDTH_LO(md
->metadata
[3]) +
326 (G_00A008_WIDTH_HI(md
->metadata
[4]) << 2) + 1;
327 height
= S_00A008_HEIGHT(md
->metadata
[4]) + 1;
329 width
= G_008F18_WIDTH(md
->metadata
[4]) + 1;
330 height
= G_008F18_HEIGHT(md
->metadata
[4]) + 1;
334 if (image
->info
.width
== width
&& image
->info
.height
== height
)
337 if (width
< image
->info
.width
|| height
< image
->info
.height
) {
339 "The imported image has smaller dimensions than the internal\n"
340 "dimensions. Using it is going to fail badly, so we reject\n"
342 "(internal dimensions: %d x %d, external dimensions: %d x %d)\n",
343 image
->info
.width
, image
->info
.height
, width
, height
);
344 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
345 } else if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
347 "Tried to import an image with inconsistent width on GFX10.\n"
348 "As GFX10 has no separate stride fields we cannot cope with\n"
349 "an inconsistency in width and will fail this import.\n"
350 "(internal dimensions: %d x %d, external dimensions: %d x %d)\n",
351 image
->info
.width
, image
->info
.height
, width
, height
);
352 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
355 "Tried to import an image with inconsistent width on pre-GFX10.\n"
356 "As GFX10 has no separate stride fields we cannot cope with\n"
357 "an inconsistency and would fail on GFX10.\n"
358 "(internal dimensions: %d x %d, external dimensions: %d x %d)\n",
359 image
->info
.width
, image
->info
.height
, width
, height
);
361 image_info
->width
= width
;
362 image_info
->height
= height
;
368 radv_patch_image_from_extra_info(struct radv_device
*device
,
369 struct radv_image
*image
,
370 const struct radv_image_create_info
*create_info
,
371 struct ac_surf_info
*image_info
)
373 VkResult result
= radv_patch_image_dimensions(device
, image
, create_info
, image_info
);
374 if (result
!= VK_SUCCESS
)
377 for (unsigned plane
= 0; plane
< image
->plane_count
; ++plane
) {
378 if (create_info
->bo_metadata
) {
379 radv_patch_surface_from_metadata(device
, &image
->planes
[plane
].surface
,
380 create_info
->bo_metadata
);
383 if (radv_surface_has_scanout(device
, create_info
)) {
384 image
->planes
[plane
].surface
.flags
|= RADEON_SURF_SCANOUT
;
385 image
->planes
[plane
].surface
.flags
|= RADEON_SURF_DISABLE_DCC
;
387 image
->info
.surf_index
= NULL
;
394 radv_init_surface(struct radv_device
*device
,
395 const struct radv_image
*image
,
396 struct radeon_surf
*surface
,
398 const VkImageCreateInfo
*pCreateInfo
,
399 VkFormat image_format
)
401 unsigned array_mode
= radv_choose_tiling(device
, pCreateInfo
, image_format
);
402 VkFormat format
= vk_format_get_plane_format(image_format
, plane_id
);
403 const struct vk_format_description
*desc
= vk_format_description(format
);
404 bool is_depth
, is_stencil
;
406 is_depth
= vk_format_has_depth(desc
);
407 is_stencil
= vk_format_has_stencil(desc
);
409 surface
->blk_w
= vk_format_get_blockwidth(format
);
410 surface
->blk_h
= vk_format_get_blockheight(format
);
412 surface
->bpe
= vk_format_get_blocksize(vk_format_depth_only(format
));
413 /* align byte per element on dword */
414 if (surface
->bpe
== 3) {
418 surface
->flags
= RADEON_SURF_SET(array_mode
, MODE
);
420 switch (pCreateInfo
->imageType
){
421 case VK_IMAGE_TYPE_1D
:
422 if (pCreateInfo
->arrayLayers
> 1)
423 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_TYPE_1D_ARRAY
, TYPE
);
425 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_TYPE_1D
, TYPE
);
427 case VK_IMAGE_TYPE_2D
:
428 if (pCreateInfo
->arrayLayers
> 1)
429 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_TYPE_2D_ARRAY
, TYPE
);
431 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_TYPE_2D
, TYPE
);
433 case VK_IMAGE_TYPE_3D
:
434 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_TYPE_3D
, TYPE
);
437 unreachable("unhandled image type");
441 surface
->flags
|= RADEON_SURF_ZBUFFER
;
442 if (radv_use_tc_compat_htile_for_image(device
, pCreateInfo
, image_format
))
443 surface
->flags
|= RADEON_SURF_TC_COMPATIBLE_HTILE
;
447 surface
->flags
|= RADEON_SURF_SBUFFER
;
449 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
&&
450 pCreateInfo
->imageType
== VK_IMAGE_TYPE_3D
&&
451 vk_format_get_blocksizebits(image_format
) == 128 &&
452 vk_format_is_compressed(image_format
))
453 surface
->flags
|= RADEON_SURF_NO_RENDER_TARGET
;
455 surface
->flags
|= RADEON_SURF_OPTIMIZE_FOR_SPACE
;
457 if (!radv_use_dcc_for_image(device
, image
, pCreateInfo
, image_format
))
458 surface
->flags
|= RADEON_SURF_DISABLE_DCC
;
463 static inline unsigned
464 si_tile_mode_index(const struct radv_image_plane
*plane
, unsigned level
, bool stencil
)
467 return plane
->surface
.u
.legacy
.stencil_tiling_index
[level
];
469 return plane
->surface
.u
.legacy
.tiling_index
[level
];
472 static unsigned radv_map_swizzle(unsigned swizzle
)
476 return V_008F0C_SQ_SEL_Y
;
478 return V_008F0C_SQ_SEL_Z
;
480 return V_008F0C_SQ_SEL_W
;
482 return V_008F0C_SQ_SEL_0
;
484 return V_008F0C_SQ_SEL_1
;
485 default: /* VK_SWIZZLE_X */
486 return V_008F0C_SQ_SEL_X
;
491 radv_make_buffer_descriptor(struct radv_device
*device
,
492 struct radv_buffer
*buffer
,
498 const struct vk_format_description
*desc
;
500 uint64_t gpu_address
= radv_buffer_get_va(buffer
->bo
);
501 uint64_t va
= gpu_address
+ buffer
->offset
;
502 unsigned num_format
, data_format
;
504 desc
= vk_format_description(vk_format
);
505 first_non_void
= vk_format_get_first_non_void_channel(vk_format
);
506 stride
= desc
->block
.bits
/ 8;
510 state
[1] = S_008F04_BASE_ADDRESS_HI(va
>> 32) |
511 S_008F04_STRIDE(stride
);
513 if (device
->physical_device
->rad_info
.chip_class
!= GFX8
&& stride
) {
518 state
[3] = S_008F0C_DST_SEL_X(radv_map_swizzle(desc
->swizzle
[0])) |
519 S_008F0C_DST_SEL_Y(radv_map_swizzle(desc
->swizzle
[1])) |
520 S_008F0C_DST_SEL_Z(radv_map_swizzle(desc
->swizzle
[2])) |
521 S_008F0C_DST_SEL_W(radv_map_swizzle(desc
->swizzle
[3]));
523 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
524 const struct gfx10_format
*fmt
= &gfx10_format_table
[vk_format
];
526 /* OOB_SELECT chooses the out-of-bounds check:
527 * - 0: (index >= NUM_RECORDS) || (offset >= STRIDE)
528 * - 1: index >= NUM_RECORDS
529 * - 2: NUM_RECORDS == 0
530 * - 3: if SWIZZLE_ENABLE == 0: offset >= NUM_RECORDS
531 * else: swizzle_address >= NUM_RECORDS
533 state
[3] |= S_008F0C_FORMAT(fmt
->img_format
) |
534 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_STRUCTURED_WITH_OFFSET
) |
535 S_008F0C_RESOURCE_LEVEL(1);
537 num_format
= radv_translate_buffer_numformat(desc
, first_non_void
);
538 data_format
= radv_translate_buffer_dataformat(desc
, first_non_void
);
540 assert(data_format
!= V_008F0C_BUF_DATA_FORMAT_INVALID
);
541 assert(num_format
!= ~0);
543 state
[3] |= S_008F0C_NUM_FORMAT(num_format
) |
544 S_008F0C_DATA_FORMAT(data_format
);
549 si_set_mutable_tex_desc_fields(struct radv_device
*device
,
550 struct radv_image
*image
,
551 const struct legacy_surf_level
*base_level_info
,
553 unsigned base_level
, unsigned first_level
,
554 unsigned block_width
, bool is_stencil
,
555 bool is_storage_image
, bool disable_compression
,
558 struct radv_image_plane
*plane
= &image
->planes
[plane_id
];
559 uint64_t gpu_address
= image
->bo
? radv_buffer_get_va(image
->bo
) + image
->offset
: 0;
560 uint64_t va
= gpu_address
+ plane
->offset
;
561 enum chip_class chip_class
= device
->physical_device
->rad_info
.chip_class
;
562 uint64_t meta_va
= 0;
563 if (chip_class
>= GFX9
) {
565 va
+= plane
->surface
.u
.gfx9
.stencil_offset
;
567 va
+= plane
->surface
.u
.gfx9
.surf_offset
;
569 va
+= base_level_info
->offset
;
572 if (chip_class
>= GFX9
||
573 base_level_info
->mode
== RADEON_SURF_MODE_2D
)
574 state
[0] |= plane
->surface
.tile_swizzle
;
575 state
[1] &= C_008F14_BASE_ADDRESS_HI
;
576 state
[1] |= S_008F14_BASE_ADDRESS_HI(va
>> 40);
578 if (chip_class
>= GFX8
) {
579 state
[6] &= C_008F28_COMPRESSION_EN
;
581 if (!disable_compression
&& radv_dcc_enabled(image
, first_level
)) {
582 meta_va
= gpu_address
+ image
->dcc_offset
;
583 if (chip_class
<= GFX8
)
584 meta_va
+= base_level_info
->dcc_offset
;
586 unsigned dcc_tile_swizzle
= plane
->surface
.tile_swizzle
<< 8;
587 dcc_tile_swizzle
&= plane
->surface
.dcc_alignment
- 1;
588 meta_va
|= dcc_tile_swizzle
;
589 } else if (!disable_compression
&&
590 radv_image_is_tc_compat_htile(image
)) {
591 meta_va
= gpu_address
+ image
->htile_offset
;
595 state
[6] |= S_008F28_COMPRESSION_EN(1);
596 if (chip_class
<= GFX9
)
597 state
[7] = meta_va
>> 8;
601 if (chip_class
>= GFX10
) {
602 state
[3] &= C_00A00C_SW_MODE
;
605 state
[3] |= S_00A00C_SW_MODE(plane
->surface
.u
.gfx9
.stencil
.swizzle_mode
);
607 state
[3] |= S_00A00C_SW_MODE(plane
->surface
.u
.gfx9
.surf
.swizzle_mode
);
610 state
[6] &= C_00A018_META_DATA_ADDRESS_LO
&
611 C_00A018_META_PIPE_ALIGNED
;
614 struct gfx9_surf_meta_flags meta
;
616 if (image
->dcc_offset
)
617 meta
= plane
->surface
.u
.gfx9
.dcc
;
619 meta
= plane
->surface
.u
.gfx9
.htile
;
621 state
[6] |= S_00A018_META_PIPE_ALIGNED(meta
.pipe_aligned
) |
622 S_00A018_META_DATA_ADDRESS_LO(meta_va
>> 8);
625 state
[7] = meta_va
>> 16;
626 } else if (chip_class
== GFX9
) {
627 state
[3] &= C_008F1C_SW_MODE
;
628 state
[4] &= C_008F20_PITCH
;
631 state
[3] |= S_008F1C_SW_MODE(plane
->surface
.u
.gfx9
.stencil
.swizzle_mode
);
632 state
[4] |= S_008F20_PITCH(plane
->surface
.u
.gfx9
.stencil
.epitch
);
634 state
[3] |= S_008F1C_SW_MODE(plane
->surface
.u
.gfx9
.surf
.swizzle_mode
);
635 state
[4] |= S_008F20_PITCH(plane
->surface
.u
.gfx9
.surf
.epitch
);
638 state
[5] &= C_008F24_META_DATA_ADDRESS
&
639 C_008F24_META_PIPE_ALIGNED
&
640 C_008F24_META_RB_ALIGNED
;
642 struct gfx9_surf_meta_flags meta
;
644 if (image
->dcc_offset
)
645 meta
= plane
->surface
.u
.gfx9
.dcc
;
647 meta
= plane
->surface
.u
.gfx9
.htile
;
649 state
[5] |= S_008F24_META_DATA_ADDRESS(meta_va
>> 40) |
650 S_008F24_META_PIPE_ALIGNED(meta
.pipe_aligned
) |
651 S_008F24_META_RB_ALIGNED(meta
.rb_aligned
);
655 unsigned pitch
= base_level_info
->nblk_x
* block_width
;
656 unsigned index
= si_tile_mode_index(plane
, base_level
, is_stencil
);
658 state
[3] &= C_008F1C_TILING_INDEX
;
659 state
[3] |= S_008F1C_TILING_INDEX(index
);
660 state
[4] &= C_008F20_PITCH
;
661 state
[4] |= S_008F20_PITCH(pitch
- 1);
665 static unsigned radv_tex_dim(VkImageType image_type
, VkImageViewType view_type
,
666 unsigned nr_layers
, unsigned nr_samples
, bool is_storage_image
, bool gfx9
)
668 if (view_type
== VK_IMAGE_VIEW_TYPE_CUBE
|| view_type
== VK_IMAGE_VIEW_TYPE_CUBE_ARRAY
)
669 return is_storage_image
? V_008F1C_SQ_RSRC_IMG_2D_ARRAY
: V_008F1C_SQ_RSRC_IMG_CUBE
;
671 /* GFX9 allocates 1D textures as 2D. */
672 if (gfx9
&& image_type
== VK_IMAGE_TYPE_1D
)
673 image_type
= VK_IMAGE_TYPE_2D
;
674 switch (image_type
) {
675 case VK_IMAGE_TYPE_1D
:
676 return nr_layers
> 1 ? V_008F1C_SQ_RSRC_IMG_1D_ARRAY
: V_008F1C_SQ_RSRC_IMG_1D
;
677 case VK_IMAGE_TYPE_2D
:
679 return nr_layers
> 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY
: V_008F1C_SQ_RSRC_IMG_2D_MSAA
;
681 return nr_layers
> 1 ? V_008F1C_SQ_RSRC_IMG_2D_ARRAY
: V_008F1C_SQ_RSRC_IMG_2D
;
682 case VK_IMAGE_TYPE_3D
:
683 if (view_type
== VK_IMAGE_VIEW_TYPE_3D
)
684 return V_008F1C_SQ_RSRC_IMG_3D
;
686 return V_008F1C_SQ_RSRC_IMG_2D_ARRAY
;
688 unreachable("illegal image type");
692 static unsigned gfx9_border_color_swizzle(const enum vk_swizzle swizzle
[4])
694 unsigned bc_swizzle
= V_008F20_BC_SWIZZLE_XYZW
;
696 if (swizzle
[3] == VK_SWIZZLE_X
) {
697 /* For the pre-defined border color values (white, opaque
698 * black, transparent black), the only thing that matters is
699 * that the alpha channel winds up in the correct place
700 * (because the RGB channels are all the same) so either of
701 * these enumerations will work.
703 if (swizzle
[2] == VK_SWIZZLE_Y
)
704 bc_swizzle
= V_008F20_BC_SWIZZLE_WZYX
;
706 bc_swizzle
= V_008F20_BC_SWIZZLE_WXYZ
;
707 } else if (swizzle
[0] == VK_SWIZZLE_X
) {
708 if (swizzle
[1] == VK_SWIZZLE_Y
)
709 bc_swizzle
= V_008F20_BC_SWIZZLE_XYZW
;
711 bc_swizzle
= V_008F20_BC_SWIZZLE_XWYZ
;
712 } else if (swizzle
[1] == VK_SWIZZLE_X
) {
713 bc_swizzle
= V_008F20_BC_SWIZZLE_YXWZ
;
714 } else if (swizzle
[2] == VK_SWIZZLE_X
) {
715 bc_swizzle
= V_008F20_BC_SWIZZLE_ZYXW
;
721 bool vi_alpha_is_on_msb(struct radv_device
*device
, VkFormat format
)
723 const struct vk_format_description
*desc
= vk_format_description(format
);
725 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
&& desc
->nr_channels
== 1)
726 return desc
->swizzle
[3] == VK_SWIZZLE_X
;
728 return radv_translate_colorswap(format
, false) <= 1;
731 * Build the sampler view descriptor for a texture (GFX10).
734 gfx10_make_texture_descriptor(struct radv_device
*device
,
735 struct radv_image
*image
,
736 bool is_storage_image
,
737 VkImageViewType view_type
,
739 const VkComponentMapping
*mapping
,
740 unsigned first_level
, unsigned last_level
,
741 unsigned first_layer
, unsigned last_layer
,
742 unsigned width
, unsigned height
, unsigned depth
,
744 uint32_t *fmask_state
)
746 const struct vk_format_description
*desc
;
747 enum vk_swizzle swizzle
[4];
751 desc
= vk_format_description(vk_format
);
752 img_format
= gfx10_format_table
[vk_format
].img_format
;
754 if (desc
->colorspace
== VK_FORMAT_COLORSPACE_ZS
) {
755 const unsigned char swizzle_xxxx
[4] = {0, 0, 0, 0};
756 vk_format_compose_swizzles(mapping
, swizzle_xxxx
, swizzle
);
758 vk_format_compose_swizzles(mapping
, desc
->swizzle
, swizzle
);
761 type
= radv_tex_dim(image
->type
, view_type
, image
->info
.array_size
, image
->info
.samples
,
762 is_storage_image
, device
->physical_device
->rad_info
.chip_class
== GFX9
);
763 if (type
== V_008F1C_SQ_RSRC_IMG_1D_ARRAY
) {
765 depth
= image
->info
.array_size
;
766 } else if (type
== V_008F1C_SQ_RSRC_IMG_2D_ARRAY
||
767 type
== V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY
) {
768 if (view_type
!= VK_IMAGE_VIEW_TYPE_3D
)
769 depth
= image
->info
.array_size
;
770 } else if (type
== V_008F1C_SQ_RSRC_IMG_CUBE
)
771 depth
= image
->info
.array_size
/ 6;
774 state
[1] = S_00A004_FORMAT(img_format
) |
775 S_00A004_WIDTH_LO(width
- 1);
776 state
[2] = S_00A008_WIDTH_HI((width
- 1) >> 2) |
777 S_00A008_HEIGHT(height
- 1) |
778 S_00A008_RESOURCE_LEVEL(1);
779 state
[3] = S_00A00C_DST_SEL_X(radv_map_swizzle(swizzle
[0])) |
780 S_00A00C_DST_SEL_Y(radv_map_swizzle(swizzle
[1])) |
781 S_00A00C_DST_SEL_Z(radv_map_swizzle(swizzle
[2])) |
782 S_00A00C_DST_SEL_W(radv_map_swizzle(swizzle
[3])) |
783 S_00A00C_BASE_LEVEL(image
->info
.samples
> 1 ?
785 S_00A00C_LAST_LEVEL(image
->info
.samples
> 1 ?
786 util_logbase2(image
->info
.samples
) :
788 S_00A00C_BC_SWIZZLE(gfx9_border_color_swizzle(swizzle
)) |
790 /* Depth is the the last accessible layer on gfx9+. The hw doesn't need
791 * to know the total number of layers.
793 state
[4] = S_00A010_DEPTH(type
== V_008F1C_SQ_RSRC_IMG_3D
? depth
- 1 : last_layer
) |
794 S_00A010_BASE_ARRAY(first_layer
);
795 state
[5] = S_00A014_ARRAY_PITCH(0) |
796 S_00A014_MAX_MIP(image
->info
.samples
> 1 ?
797 util_logbase2(image
->info
.samples
) :
798 image
->info
.levels
- 1) |
799 S_00A014_PERF_MOD(4);
803 if (radv_dcc_enabled(image
, first_level
)) {
804 state
[6] |= S_00A018_MAX_UNCOMPRESSED_BLOCK_SIZE(V_028C78_MAX_BLOCK_SIZE_256B
) |
805 S_00A018_MAX_COMPRESSED_BLOCK_SIZE(V_028C78_MAX_BLOCK_SIZE_128B
) |
806 S_00A018_ALPHA_IS_ON_MSB(vi_alpha_is_on_msb(device
, vk_format
));
809 /* Initialize the sampler view for FMASK. */
810 if (radv_image_has_fmask(image
)) {
811 uint64_t gpu_address
= radv_buffer_get_va(image
->bo
);
815 assert(image
->plane_count
== 1);
817 va
= gpu_address
+ image
->offset
+ image
->fmask_offset
;
819 switch (image
->info
.samples
) {
821 format
= V_008F0C_IMG_FORMAT_FMASK8_S2_F2
;
824 format
= V_008F0C_IMG_FORMAT_FMASK8_S4_F4
;
827 format
= V_008F0C_IMG_FORMAT_FMASK32_S8_F8
;
830 unreachable("invalid nr_samples");
833 fmask_state
[0] = (va
>> 8) | image
->planes
[0].surface
.fmask_tile_swizzle
;
834 fmask_state
[1] = S_00A004_BASE_ADDRESS_HI(va
>> 40) |
835 S_00A004_FORMAT(format
) |
836 S_00A004_WIDTH_LO(width
- 1);
837 fmask_state
[2] = S_00A008_WIDTH_HI((width
- 1) >> 2) |
838 S_00A008_HEIGHT(height
- 1) |
839 S_00A008_RESOURCE_LEVEL(1);
840 fmask_state
[3] = S_00A00C_DST_SEL_X(V_008F1C_SQ_SEL_X
) |
841 S_00A00C_DST_SEL_Y(V_008F1C_SQ_SEL_X
) |
842 S_00A00C_DST_SEL_Z(V_008F1C_SQ_SEL_X
) |
843 S_00A00C_DST_SEL_W(V_008F1C_SQ_SEL_X
) |
844 S_00A00C_SW_MODE(image
->planes
[0].surface
.u
.gfx9
.fmask
.swizzle_mode
) |
845 S_00A00C_TYPE(radv_tex_dim(image
->type
, view_type
, image
->info
.array_size
, 0, false, false));
846 fmask_state
[4] = S_00A010_DEPTH(last_layer
) |
847 S_00A010_BASE_ARRAY(first_layer
);
849 fmask_state
[6] = S_00A018_META_PIPE_ALIGNED(image
->planes
[0].surface
.u
.gfx9
.cmask
.pipe_aligned
);
851 } else if (fmask_state
)
852 memset(fmask_state
, 0, 8 * 4);
856 * Build the sampler view descriptor for a texture (SI-GFX9)
859 si_make_texture_descriptor(struct radv_device
*device
,
860 struct radv_image
*image
,
861 bool is_storage_image
,
862 VkImageViewType view_type
,
864 const VkComponentMapping
*mapping
,
865 unsigned first_level
, unsigned last_level
,
866 unsigned first_layer
, unsigned last_layer
,
867 unsigned width
, unsigned height
, unsigned depth
,
869 uint32_t *fmask_state
)
871 const struct vk_format_description
*desc
;
872 enum vk_swizzle swizzle
[4];
874 unsigned num_format
, data_format
, type
;
876 desc
= vk_format_description(vk_format
);
878 if (desc
->colorspace
== VK_FORMAT_COLORSPACE_ZS
) {
879 const unsigned char swizzle_xxxx
[4] = {0, 0, 0, 0};
880 vk_format_compose_swizzles(mapping
, swizzle_xxxx
, swizzle
);
882 vk_format_compose_swizzles(mapping
, desc
->swizzle
, swizzle
);
885 first_non_void
= vk_format_get_first_non_void_channel(vk_format
);
887 num_format
= radv_translate_tex_numformat(vk_format
, desc
, first_non_void
);
888 if (num_format
== ~0) {
892 data_format
= radv_translate_tex_dataformat(vk_format
, desc
, first_non_void
);
893 if (data_format
== ~0) {
897 /* S8 with either Z16 or Z32 HTILE need a special format. */
898 if (device
->physical_device
->rad_info
.chip_class
== GFX9
&&
899 vk_format
== VK_FORMAT_S8_UINT
&&
900 radv_image_is_tc_compat_htile(image
)) {
901 if (image
->vk_format
== VK_FORMAT_D32_SFLOAT_S8_UINT
)
902 data_format
= V_008F14_IMG_DATA_FORMAT_S8_32
;
903 else if (image
->vk_format
== VK_FORMAT_D16_UNORM_S8_UINT
)
904 data_format
= V_008F14_IMG_DATA_FORMAT_S8_16
;
906 type
= radv_tex_dim(image
->type
, view_type
, image
->info
.array_size
, image
->info
.samples
,
907 is_storage_image
, device
->physical_device
->rad_info
.chip_class
== GFX9
);
908 if (type
== V_008F1C_SQ_RSRC_IMG_1D_ARRAY
) {
910 depth
= image
->info
.array_size
;
911 } else if (type
== V_008F1C_SQ_RSRC_IMG_2D_ARRAY
||
912 type
== V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY
) {
913 if (view_type
!= VK_IMAGE_VIEW_TYPE_3D
)
914 depth
= image
->info
.array_size
;
915 } else if (type
== V_008F1C_SQ_RSRC_IMG_CUBE
)
916 depth
= image
->info
.array_size
/ 6;
919 state
[1] = (S_008F14_DATA_FORMAT(data_format
) |
920 S_008F14_NUM_FORMAT(num_format
));
921 state
[2] = (S_008F18_WIDTH(width
- 1) |
922 S_008F18_HEIGHT(height
- 1) |
923 S_008F18_PERF_MOD(4));
924 state
[3] = (S_008F1C_DST_SEL_X(radv_map_swizzle(swizzle
[0])) |
925 S_008F1C_DST_SEL_Y(radv_map_swizzle(swizzle
[1])) |
926 S_008F1C_DST_SEL_Z(radv_map_swizzle(swizzle
[2])) |
927 S_008F1C_DST_SEL_W(radv_map_swizzle(swizzle
[3])) |
928 S_008F1C_BASE_LEVEL(image
->info
.samples
> 1 ?
930 S_008F1C_LAST_LEVEL(image
->info
.samples
> 1 ?
931 util_logbase2(image
->info
.samples
) :
933 S_008F1C_TYPE(type
));
935 state
[5] = S_008F24_BASE_ARRAY(first_layer
);
939 if (device
->physical_device
->rad_info
.chip_class
== GFX9
) {
940 unsigned bc_swizzle
= gfx9_border_color_swizzle(swizzle
);
942 /* Depth is the last accessible layer on Gfx9.
943 * The hw doesn't need to know the total number of layers.
945 if (type
== V_008F1C_SQ_RSRC_IMG_3D
)
946 state
[4] |= S_008F20_DEPTH(depth
- 1);
948 state
[4] |= S_008F20_DEPTH(last_layer
);
950 state
[4] |= S_008F20_BC_SWIZZLE(bc_swizzle
);
951 state
[5] |= S_008F24_MAX_MIP(image
->info
.samples
> 1 ?
952 util_logbase2(image
->info
.samples
) :
953 image
->info
.levels
- 1);
955 state
[3] |= S_008F1C_POW2_PAD(image
->info
.levels
> 1);
956 state
[4] |= S_008F20_DEPTH(depth
- 1);
957 state
[5] |= S_008F24_LAST_ARRAY(last_layer
);
959 if (image
->dcc_offset
) {
960 state
[6] = S_008F28_ALPHA_IS_ON_MSB(vi_alpha_is_on_msb(device
, vk_format
));
962 /* The last dword is unused by hw. The shader uses it to clear
963 * bits in the first dword of sampler state.
965 if (device
->physical_device
->rad_info
.chip_class
<= GFX7
&& image
->info
.samples
<= 1) {
966 if (first_level
== last_level
)
967 state
[7] = C_008F30_MAX_ANISO_RATIO
;
969 state
[7] = 0xffffffff;
973 /* Initialize the sampler view for FMASK. */
974 if (radv_image_has_fmask(image
)) {
975 uint32_t fmask_format
, num_format
;
976 uint64_t gpu_address
= radv_buffer_get_va(image
->bo
);
979 assert(image
->plane_count
== 1);
981 va
= gpu_address
+ image
->offset
+ image
->fmask_offset
;
983 if (device
->physical_device
->rad_info
.chip_class
== GFX9
) {
984 fmask_format
= V_008F14_IMG_DATA_FORMAT_FMASK
;
985 switch (image
->info
.samples
) {
987 num_format
= V_008F14_IMG_FMASK_8_2_2
;
990 num_format
= V_008F14_IMG_FMASK_8_4_4
;
993 num_format
= V_008F14_IMG_FMASK_32_8_8
;
996 unreachable("invalid nr_samples");
999 switch (image
->info
.samples
) {
1001 fmask_format
= V_008F14_IMG_DATA_FORMAT_FMASK8_S2_F2
;
1004 fmask_format
= V_008F14_IMG_DATA_FORMAT_FMASK8_S4_F4
;
1007 fmask_format
= V_008F14_IMG_DATA_FORMAT_FMASK32_S8_F8
;
1011 fmask_format
= V_008F14_IMG_DATA_FORMAT_INVALID
;
1013 num_format
= V_008F14_IMG_NUM_FORMAT_UINT
;
1016 fmask_state
[0] = va
>> 8;
1017 fmask_state
[0] |= image
->planes
[0].surface
.fmask_tile_swizzle
;
1018 fmask_state
[1] = S_008F14_BASE_ADDRESS_HI(va
>> 40) |
1019 S_008F14_DATA_FORMAT(fmask_format
) |
1020 S_008F14_NUM_FORMAT(num_format
);
1021 fmask_state
[2] = S_008F18_WIDTH(width
- 1) |
1022 S_008F18_HEIGHT(height
- 1);
1023 fmask_state
[3] = S_008F1C_DST_SEL_X(V_008F1C_SQ_SEL_X
) |
1024 S_008F1C_DST_SEL_Y(V_008F1C_SQ_SEL_X
) |
1025 S_008F1C_DST_SEL_Z(V_008F1C_SQ_SEL_X
) |
1026 S_008F1C_DST_SEL_W(V_008F1C_SQ_SEL_X
) |
1027 S_008F1C_TYPE(radv_tex_dim(image
->type
, view_type
, image
->info
.array_size
, 0, false, false));
1029 fmask_state
[5] = S_008F24_BASE_ARRAY(first_layer
);
1033 if (device
->physical_device
->rad_info
.chip_class
== GFX9
) {
1034 fmask_state
[3] |= S_008F1C_SW_MODE(image
->planes
[0].surface
.u
.gfx9
.fmask
.swizzle_mode
);
1035 fmask_state
[4] |= S_008F20_DEPTH(last_layer
) |
1036 S_008F20_PITCH(image
->planes
[0].surface
.u
.gfx9
.fmask
.epitch
);
1037 fmask_state
[5] |= S_008F24_META_PIPE_ALIGNED(image
->planes
[0].surface
.u
.gfx9
.cmask
.pipe_aligned
) |
1038 S_008F24_META_RB_ALIGNED(image
->planes
[0].surface
.u
.gfx9
.cmask
.rb_aligned
);
1040 if (radv_image_is_tc_compat_cmask(image
)) {
1041 va
= gpu_address
+ image
->offset
+ image
->cmask_offset
;
1043 fmask_state
[5] |= S_008F24_META_DATA_ADDRESS(va
>> 40);
1044 fmask_state
[6] |= S_008F28_COMPRESSION_EN(1);
1045 fmask_state
[7] |= va
>> 8;
1048 fmask_state
[3] |= S_008F1C_TILING_INDEX(image
->planes
[0].surface
.u
.legacy
.fmask
.tiling_index
);
1049 fmask_state
[4] |= S_008F20_DEPTH(depth
- 1) |
1050 S_008F20_PITCH(image
->planes
[0].surface
.u
.legacy
.fmask
.pitch_in_pixels
- 1);
1051 fmask_state
[5] |= S_008F24_LAST_ARRAY(last_layer
);
1053 if (radv_image_is_tc_compat_cmask(image
)) {
1054 va
= gpu_address
+ image
->offset
+ image
->cmask_offset
;
1056 fmask_state
[6] |= S_008F28_COMPRESSION_EN(1);
1057 fmask_state
[7] |= va
>> 8;
1060 } else if (fmask_state
)
1061 memset(fmask_state
, 0, 8 * 4);
1065 radv_make_texture_descriptor(struct radv_device
*device
,
1066 struct radv_image
*image
,
1067 bool is_storage_image
,
1068 VkImageViewType view_type
,
1070 const VkComponentMapping
*mapping
,
1071 unsigned first_level
, unsigned last_level
,
1072 unsigned first_layer
, unsigned last_layer
,
1073 unsigned width
, unsigned height
, unsigned depth
,
1075 uint32_t *fmask_state
)
1077 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
1078 gfx10_make_texture_descriptor(device
, image
, is_storage_image
,
1079 view_type
, vk_format
, mapping
,
1080 first_level
, last_level
,
1081 first_layer
, last_layer
,
1082 width
, height
, depth
,
1083 state
, fmask_state
);
1085 si_make_texture_descriptor(device
, image
, is_storage_image
,
1086 view_type
, vk_format
, mapping
,
1087 first_level
, last_level
,
1088 first_layer
, last_layer
,
1089 width
, height
, depth
,
1090 state
, fmask_state
);
1095 radv_query_opaque_metadata(struct radv_device
*device
,
1096 struct radv_image
*image
,
1097 struct radeon_bo_metadata
*md
)
1099 static const VkComponentMapping fixedmapping
;
1100 uint32_t desc
[8], i
;
1102 assert(image
->plane_count
== 1);
1104 /* Metadata image format format version 1:
1105 * [0] = 1 (metadata format identifier)
1106 * [1] = (VENDOR_ID << 16) | PCI_ID
1107 * [2:9] = image descriptor for the whole resource
1108 * [2] is always 0, because the base address is cleared
1109 * [9] is the DCC offset bits [39:8] from the beginning of
1111 * [10:10+LAST_LEVEL] = mipmap level offset bits [39:8] for each level
1113 md
->metadata
[0] = 1; /* metadata image format version 1 */
1115 /* TILE_MODE_INDEX is ambiguous without a PCI ID. */
1116 md
->metadata
[1] = si_get_bo_metadata_word1(device
);
1119 radv_make_texture_descriptor(device
, image
, false,
1120 (VkImageViewType
)image
->type
, image
->vk_format
,
1121 &fixedmapping
, 0, image
->info
.levels
- 1, 0,
1122 image
->info
.array_size
- 1,
1123 image
->info
.width
, image
->info
.height
,
1127 si_set_mutable_tex_desc_fields(device
, image
, &image
->planes
[0].surface
.u
.legacy
.level
[0], 0, 0, 0,
1128 image
->planes
[0].surface
.blk_w
, false, false, false, desc
);
1130 /* Clear the base address and set the relative DCC offset. */
1132 desc
[1] &= C_008F14_BASE_ADDRESS_HI
;
1133 desc
[7] = image
->dcc_offset
>> 8;
1135 /* Dwords [2:9] contain the image descriptor. */
1136 memcpy(&md
->metadata
[2], desc
, sizeof(desc
));
1138 /* Dwords [10:..] contain the mipmap level offsets. */
1139 if (device
->physical_device
->rad_info
.chip_class
<= GFX8
) {
1140 for (i
= 0; i
<= image
->info
.levels
- 1; i
++)
1141 md
->metadata
[10+i
] = image
->planes
[0].surface
.u
.legacy
.level
[i
].offset
>> 8;
1142 md
->size_metadata
= (11 + image
->info
.levels
- 1) * 4;
1144 md
->size_metadata
= 10 * 4;
1148 radv_init_metadata(struct radv_device
*device
,
1149 struct radv_image
*image
,
1150 struct radeon_bo_metadata
*metadata
)
1152 struct radeon_surf
*surface
= &image
->planes
[0].surface
;
1154 memset(metadata
, 0, sizeof(*metadata
));
1156 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
1157 metadata
->u
.gfx9
.swizzle_mode
= surface
->u
.gfx9
.surf
.swizzle_mode
;
1158 metadata
->u
.gfx9
.scanout
= (surface
->flags
& RADEON_SURF_SCANOUT
) != 0;
1160 metadata
->u
.legacy
.microtile
= surface
->u
.legacy
.level
[0].mode
>= RADEON_SURF_MODE_1D
?
1161 RADEON_LAYOUT_TILED
: RADEON_LAYOUT_LINEAR
;
1162 metadata
->u
.legacy
.macrotile
= surface
->u
.legacy
.level
[0].mode
>= RADEON_SURF_MODE_2D
?
1163 RADEON_LAYOUT_TILED
: RADEON_LAYOUT_LINEAR
;
1164 metadata
->u
.legacy
.pipe_config
= surface
->u
.legacy
.pipe_config
;
1165 metadata
->u
.legacy
.bankw
= surface
->u
.legacy
.bankw
;
1166 metadata
->u
.legacy
.bankh
= surface
->u
.legacy
.bankh
;
1167 metadata
->u
.legacy
.tile_split
= surface
->u
.legacy
.tile_split
;
1168 metadata
->u
.legacy
.mtilea
= surface
->u
.legacy
.mtilea
;
1169 metadata
->u
.legacy
.num_banks
= surface
->u
.legacy
.num_banks
;
1170 metadata
->u
.legacy
.stride
= surface
->u
.legacy
.level
[0].nblk_x
* surface
->bpe
;
1171 metadata
->u
.legacy
.scanout
= (surface
->flags
& RADEON_SURF_SCANOUT
) != 0;
1173 radv_query_opaque_metadata(device
, image
, metadata
);
1177 radv_image_override_offset_stride(struct radv_device
*device
,
1178 struct radv_image
*image
,
1179 uint64_t offset
, uint32_t stride
)
1181 struct radeon_surf
*surface
= &image
->planes
[0].surface
;
1182 unsigned bpe
= vk_format_get_blocksizebits(image
->vk_format
) / 8;
1184 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
1186 surface
->u
.gfx9
.surf_pitch
= stride
;
1187 surface
->u
.gfx9
.surf_slice_size
=
1188 (uint64_t)stride
* surface
->u
.gfx9
.surf_height
* bpe
;
1190 surface
->u
.gfx9
.surf_offset
= offset
;
1192 surface
->u
.legacy
.level
[0].nblk_x
= stride
;
1193 surface
->u
.legacy
.level
[0].slice_size_dw
=
1194 ((uint64_t)stride
* surface
->u
.legacy
.level
[0].nblk_y
* bpe
) / 4;
1197 for (unsigned i
= 0; i
< ARRAY_SIZE(surface
->u
.legacy
.level
); ++i
)
1198 surface
->u
.legacy
.level
[i
].offset
+= offset
;
1205 radv_image_alloc_fmask(struct radv_device
*device
,
1206 struct radv_image
*image
)
1208 unsigned fmask_alignment
= image
->planes
[0].surface
.fmask_alignment
;
1210 image
->fmask_offset
= align64(image
->size
, fmask_alignment
);
1211 image
->size
= image
->fmask_offset
+ image
->planes
[0].surface
.fmask_size
;
1212 image
->alignment
= MAX2(image
->alignment
, fmask_alignment
);
1216 radv_image_alloc_cmask(struct radv_device
*device
,
1217 struct radv_image
*image
)
1219 unsigned cmask_alignment
= image
->planes
[0].surface
.cmask_alignment
;
1220 unsigned cmask_size
= image
->planes
[0].surface
.cmask_size
;
1221 uint32_t clear_value_size
= 0;
1226 assert(cmask_alignment
);
1228 image
->cmask_offset
= align64(image
->size
, cmask_alignment
);
1229 /* + 8 for storing the clear values */
1230 if (!image
->clear_value_offset
) {
1231 image
->clear_value_offset
= image
->cmask_offset
+ cmask_size
;
1232 clear_value_size
= 8;
1234 image
->size
= image
->cmask_offset
+ cmask_size
+ clear_value_size
;
1235 image
->alignment
= MAX2(image
->alignment
, cmask_alignment
);
1239 radv_image_alloc_dcc(struct radv_image
*image
)
1241 assert(image
->plane_count
== 1);
1243 image
->dcc_offset
= align64(image
->size
, image
->planes
[0].surface
.dcc_alignment
);
1244 /* + 24 for storing the clear values + fce pred + dcc pred for each mip */
1245 image
->clear_value_offset
= image
->dcc_offset
+ image
->planes
[0].surface
.dcc_size
;
1246 image
->fce_pred_offset
= image
->clear_value_offset
+ 8 * image
->info
.levels
;
1247 image
->dcc_pred_offset
= image
->clear_value_offset
+ 16 * image
->info
.levels
;
1248 image
->size
= image
->dcc_offset
+ image
->planes
[0].surface
.dcc_size
+ 24 * image
->info
.levels
;
1249 image
->alignment
= MAX2(image
->alignment
, image
->planes
[0].surface
.dcc_alignment
);
1253 radv_image_alloc_htile(struct radv_device
*device
, struct radv_image
*image
)
1255 image
->htile_offset
= align64(image
->size
, image
->planes
[0].surface
.htile_alignment
);
1257 /* + 8 for storing the clear values */
1258 image
->clear_value_offset
= image
->htile_offset
+ image
->planes
[0].surface
.htile_size
;
1259 image
->size
= image
->clear_value_offset
+ image
->info
.levels
* 8;
1260 if (radv_image_is_tc_compat_htile(image
) &&
1261 device
->physical_device
->rad_info
.has_tc_compat_zrange_bug
) {
1262 /* Metadata for the TC-compatible HTILE hardware bug which
1263 * have to be fixed by updating ZRANGE_PRECISION when doing
1264 * fast depth clears to 0.0f.
1266 image
->tc_compat_zrange_offset
= image
->size
;
1267 image
->size
= image
->tc_compat_zrange_offset
+ image
->info
.levels
* 4;
1269 image
->alignment
= align64(image
->alignment
, image
->planes
[0].surface
.htile_alignment
);
1273 radv_image_can_enable_dcc_or_cmask(struct radv_image
*image
)
1275 if (image
->info
.samples
<= 1 &&
1276 image
->info
.width
* image
->info
.height
<= 512 * 512) {
1277 /* Do not enable CMASK or DCC for small surfaces where the cost
1278 * of the eliminate pass can be higher than the benefit of fast
1279 * clear. RadeonSI does this, but the image threshold is
1285 return image
->usage
& VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT
&&
1286 (image
->exclusive
|| image
->queue_family_mask
== 1);
1290 radv_image_can_enable_dcc(struct radv_device
*device
, struct radv_image
*image
)
1292 if (!radv_image_can_enable_dcc_or_cmask(image
) ||
1293 !radv_image_has_dcc(image
))
1296 /* On GFX8, DCC layers can be interleaved and it's currently only
1297 * enabled if slice size is equal to the per slice fast clear size
1298 * because the driver assumes that portions of multiple layers are
1299 * contiguous during fast clears.
1301 if (image
->info
.array_size
> 1) {
1302 const struct legacy_surf_level
*surf_level
=
1303 &image
->planes
[0].surface
.u
.legacy
.level
[0];
1305 assert(device
->physical_device
->rad_info
.chip_class
== GFX8
);
1307 if (image
->planes
[0].surface
.dcc_slice_size
!= surf_level
->dcc_fast_clear_size
)
1315 radv_image_can_enable_cmask(struct radv_image
*image
)
1317 if (image
->planes
[0].surface
.bpe
> 8 && image
->info
.samples
== 1) {
1318 /* Do not enable CMASK for non-MSAA images (fast color clear)
1319 * because 128 bit formats are not supported, but FMASK might
1325 return radv_image_can_enable_dcc_or_cmask(image
) &&
1326 image
->info
.levels
== 1 &&
1327 image
->info
.depth
== 1 &&
1328 !image
->planes
[0].surface
.is_linear
;
1332 radv_image_can_enable_fmask(struct radv_image
*image
)
1334 return image
->info
.samples
> 1 && vk_format_is_color(image
->vk_format
);
1338 radv_image_can_enable_htile(struct radv_image
*image
)
1340 return radv_image_has_htile(image
) &&
1341 image
->info
.levels
== 1 &&
1342 image
->info
.width
* image
->info
.height
>= 8 * 8;
1345 static void radv_image_disable_dcc(struct radv_image
*image
)
1347 for (unsigned i
= 0; i
< image
->plane_count
; ++i
)
1348 image
->planes
[i
].surface
.dcc_size
= 0;
1351 static void radv_image_disable_htile(struct radv_image
*image
)
1353 for (unsigned i
= 0; i
< image
->plane_count
; ++i
)
1354 image
->planes
[i
].surface
.htile_size
= 0;
1358 radv_image_create_layout(struct radv_device
*device
,
1359 struct radv_image_create_info create_info
,
1360 struct radv_image
*image
)
1362 /* Check that we did not initialize things earlier */
1363 assert(!image
->planes
[0].surface
.surf_size
);
1365 /* Clear the pCreateInfo pointer so we catch issues in the delayed case when we test in the
1366 * common internal case. */
1367 create_info
.vk_info
= NULL
;
1369 struct ac_surf_info image_info
= image
->info
;
1370 VkResult result
= radv_patch_image_from_extra_info(device
, image
, &create_info
, &image_info
);
1371 if (result
!= VK_SUCCESS
)
1375 image
->alignment
= 1;
1376 for (unsigned plane
= 0; plane
< image
->plane_count
; ++plane
) {
1377 struct ac_surf_info info
= image_info
;
1380 const struct vk_format_description
*desc
= vk_format_description(image
->vk_format
);
1381 assert(info
.width
% desc
->width_divisor
== 0);
1382 assert(info
.height
% desc
->height_divisor
== 0);
1384 info
.width
/= desc
->width_divisor
;
1385 info
.height
/= desc
->height_divisor
;
1388 device
->ws
->surface_init(device
->ws
, &info
, &image
->planes
[plane
].surface
);
1390 image
->planes
[plane
].offset
= align(image
->size
, image
->planes
[plane
].surface
.surf_alignment
);
1391 image
->size
= image
->planes
[plane
].offset
+ image
->planes
[plane
].surface
.surf_size
;
1392 image
->alignment
= image
->planes
[plane
].surface
.surf_alignment
;
1394 image
->planes
[plane
].format
= vk_format_get_plane_format(image
->vk_format
, plane
);
1397 if (!create_info
.no_metadata_planes
) {
1398 /* Try to enable DCC first. */
1399 if (radv_image_can_enable_dcc(device
, image
)) {
1400 radv_image_alloc_dcc(image
);
1401 if (image
->info
.samples
> 1) {
1402 /* CMASK should be enabled because DCC fast
1403 * clear with MSAA needs it.
1405 assert(radv_image_can_enable_cmask(image
));
1406 radv_image_alloc_cmask(device
, image
);
1409 /* When DCC cannot be enabled, try CMASK. */
1410 radv_image_disable_dcc(image
);
1411 if (radv_image_can_enable_cmask(image
)) {
1412 radv_image_alloc_cmask(device
, image
);
1416 /* Try to enable FMASK for multisampled images. */
1417 if (radv_image_can_enable_fmask(image
)) {
1418 radv_image_alloc_fmask(device
, image
);
1420 if (radv_use_tc_compat_cmask_for_image(device
, image
))
1421 image
->tc_compatible_cmask
= true;
1423 /* Otherwise, try to enable HTILE for depth surfaces. */
1424 if (radv_image_can_enable_htile(image
) &&
1425 !(device
->instance
->debug_flags
& RADV_DEBUG_NO_HIZ
)) {
1426 image
->tc_compatible_htile
= image
->planes
[0].surface
.flags
& RADEON_SURF_TC_COMPATIBLE_HTILE
;
1427 radv_image_alloc_htile(device
, image
);
1429 radv_image_disable_htile(image
);
1433 radv_image_disable_dcc(image
);
1434 radv_image_disable_htile(image
);
1437 assert(image
->planes
[0].surface
.surf_size
);
1442 radv_image_create(VkDevice _device
,
1443 const struct radv_image_create_info
*create_info
,
1444 const VkAllocationCallbacks
* alloc
,
1447 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1448 const VkImageCreateInfo
*pCreateInfo
= create_info
->vk_info
;
1449 struct radv_image
*image
= NULL
;
1450 VkFormat format
= radv_select_android_external_format(pCreateInfo
->pNext
,
1451 pCreateInfo
->format
);
1452 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO
);
1454 const unsigned plane_count
= vk_format_get_plane_count(format
);
1455 const size_t image_struct_size
= sizeof(*image
) + sizeof(struct radv_image_plane
) * plane_count
;
1457 radv_assert(pCreateInfo
->mipLevels
> 0);
1458 radv_assert(pCreateInfo
->arrayLayers
> 0);
1459 radv_assert(pCreateInfo
->samples
> 0);
1460 radv_assert(pCreateInfo
->extent
.width
> 0);
1461 radv_assert(pCreateInfo
->extent
.height
> 0);
1462 radv_assert(pCreateInfo
->extent
.depth
> 0);
1464 image
= vk_zalloc2(&device
->alloc
, alloc
, image_struct_size
, 8,
1465 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1467 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
1469 image
->type
= pCreateInfo
->imageType
;
1470 image
->info
.width
= pCreateInfo
->extent
.width
;
1471 image
->info
.height
= pCreateInfo
->extent
.height
;
1472 image
->info
.depth
= pCreateInfo
->extent
.depth
;
1473 image
->info
.samples
= pCreateInfo
->samples
;
1474 image
->info
.storage_samples
= pCreateInfo
->samples
;
1475 image
->info
.array_size
= pCreateInfo
->arrayLayers
;
1476 image
->info
.levels
= pCreateInfo
->mipLevels
;
1477 image
->info
.num_channels
= vk_format_get_nr_components(format
);
1479 image
->vk_format
= format
;
1480 image
->tiling
= pCreateInfo
->tiling
;
1481 image
->usage
= pCreateInfo
->usage
;
1482 image
->flags
= pCreateInfo
->flags
;
1483 image
->plane_count
= plane_count
;
1485 image
->exclusive
= pCreateInfo
->sharingMode
== VK_SHARING_MODE_EXCLUSIVE
;
1486 if (pCreateInfo
->sharingMode
== VK_SHARING_MODE_CONCURRENT
) {
1487 for (uint32_t i
= 0; i
< pCreateInfo
->queueFamilyIndexCount
; ++i
)
1488 if (pCreateInfo
->pQueueFamilyIndices
[i
] == VK_QUEUE_FAMILY_EXTERNAL
||
1489 pCreateInfo
->pQueueFamilyIndices
[i
] == VK_QUEUE_FAMILY_FOREIGN_EXT
)
1490 image
->queue_family_mask
|= (1u << RADV_MAX_QUEUE_FAMILIES
) - 1u;
1492 image
->queue_family_mask
|= 1u << pCreateInfo
->pQueueFamilyIndices
[i
];
1495 const VkExternalMemoryImageCreateInfo
*external_info
=
1496 vk_find_struct_const(pCreateInfo
->pNext
,
1497 EXTERNAL_MEMORY_IMAGE_CREATE_INFO
) ;
1499 image
->shareable
= external_info
;
1500 if (!vk_format_is_depth_or_stencil(format
) && !image
->shareable
) {
1501 image
->info
.surf_index
= &device
->image_mrt_offset_counter
;
1504 for (unsigned plane
= 0; plane
< image
->plane_count
; ++plane
) {
1505 radv_init_surface(device
, image
, &image
->planes
[plane
].surface
, plane
, pCreateInfo
, format
);
1508 bool delay_layout
= external_info
&&
1509 (external_info
->handleTypes
& VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
);
1512 *pImage
= radv_image_to_handle(image
);
1513 assert (!(image
->flags
& VK_IMAGE_CREATE_SPARSE_BINDING_BIT
));
1517 ASSERTED VkResult result
= radv_image_create_layout(device
, *create_info
, image
);
1518 assert(result
== VK_SUCCESS
);
1520 if (image
->flags
& VK_IMAGE_CREATE_SPARSE_BINDING_BIT
) {
1521 image
->alignment
= MAX2(image
->alignment
, 4096);
1522 image
->size
= align64(image
->size
, image
->alignment
);
1525 image
->bo
= device
->ws
->buffer_create(device
->ws
, image
->size
, image
->alignment
,
1526 0, RADEON_FLAG_VIRTUAL
, RADV_BO_PRIORITY_VIRTUAL
);
1528 vk_free2(&device
->alloc
, alloc
, image
);
1529 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1533 *pImage
= radv_image_to_handle(image
);
1539 radv_image_view_make_descriptor(struct radv_image_view
*iview
,
1540 struct radv_device
*device
,
1542 const VkComponentMapping
*components
,
1543 bool is_storage_image
, bool disable_compression
,
1544 unsigned plane_id
, unsigned descriptor_plane_id
)
1546 struct radv_image
*image
= iview
->image
;
1547 struct radv_image_plane
*plane
= &image
->planes
[plane_id
];
1548 const struct vk_format_description
*format_desc
= vk_format_description(image
->vk_format
);
1549 bool is_stencil
= iview
->aspect_mask
== VK_IMAGE_ASPECT_STENCIL_BIT
;
1551 union radv_descriptor
*descriptor
;
1552 uint32_t hw_level
= 0;
1554 if (is_storage_image
) {
1555 descriptor
= &iview
->storage_descriptor
;
1557 descriptor
= &iview
->descriptor
;
1560 assert(vk_format_get_plane_count(vk_format
) == 1);
1561 assert(plane
->surface
.blk_w
% vk_format_get_blockwidth(plane
->format
) == 0);
1562 blk_w
= plane
->surface
.blk_w
/ vk_format_get_blockwidth(plane
->format
) * vk_format_get_blockwidth(vk_format
);
1564 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
)
1565 hw_level
= iview
->base_mip
;
1566 radv_make_texture_descriptor(device
, image
, is_storage_image
,
1570 hw_level
, hw_level
+ iview
->level_count
- 1,
1572 iview
->base_layer
+ iview
->layer_count
- 1,
1573 iview
->extent
.width
/ (plane_id
? format_desc
->width_divisor
: 1),
1574 iview
->extent
.height
/ (plane_id
? format_desc
->height_divisor
: 1),
1575 iview
->extent
.depth
,
1576 descriptor
->plane_descriptors
[descriptor_plane_id
],
1577 descriptor_plane_id
? NULL
: descriptor
->fmask_descriptor
);
1579 const struct legacy_surf_level
*base_level_info
= NULL
;
1580 if (device
->physical_device
->rad_info
.chip_class
<= GFX9
) {
1582 base_level_info
= &plane
->surface
.u
.legacy
.stencil_level
[iview
->base_mip
];
1584 base_level_info
= &plane
->surface
.u
.legacy
.level
[iview
->base_mip
];
1586 si_set_mutable_tex_desc_fields(device
, image
,
1591 blk_w
, is_stencil
, is_storage_image
,
1592 is_storage_image
|| disable_compression
,
1593 descriptor
->plane_descriptors
[descriptor_plane_id
]);
1597 radv_plane_from_aspect(VkImageAspectFlags mask
)
1600 case VK_IMAGE_ASPECT_PLANE_1_BIT
:
1602 case VK_IMAGE_ASPECT_PLANE_2_BIT
:
1610 radv_get_aspect_format(struct radv_image
*image
, VkImageAspectFlags mask
)
1613 case VK_IMAGE_ASPECT_PLANE_0_BIT
:
1614 return image
->planes
[0].format
;
1615 case VK_IMAGE_ASPECT_PLANE_1_BIT
:
1616 return image
->planes
[1].format
;
1617 case VK_IMAGE_ASPECT_PLANE_2_BIT
:
1618 return image
->planes
[2].format
;
1619 case VK_IMAGE_ASPECT_STENCIL_BIT
:
1620 return vk_format_stencil_only(image
->vk_format
);
1621 case VK_IMAGE_ASPECT_DEPTH_BIT
:
1622 return vk_format_depth_only(image
->vk_format
);
1623 case VK_IMAGE_ASPECT_DEPTH_BIT
| VK_IMAGE_ASPECT_STENCIL_BIT
:
1624 return vk_format_depth_only(image
->vk_format
);
1626 return image
->vk_format
;
1631 radv_image_view_init(struct radv_image_view
*iview
,
1632 struct radv_device
*device
,
1633 const VkImageViewCreateInfo
* pCreateInfo
,
1634 const struct radv_image_view_extra_create_info
* extra_create_info
)
1636 RADV_FROM_HANDLE(radv_image
, image
, pCreateInfo
->image
);
1637 const VkImageSubresourceRange
*range
= &pCreateInfo
->subresourceRange
;
1639 switch (image
->type
) {
1640 case VK_IMAGE_TYPE_1D
:
1641 case VK_IMAGE_TYPE_2D
:
1642 assert(range
->baseArrayLayer
+ radv_get_layerCount(image
, range
) - 1 <= image
->info
.array_size
);
1644 case VK_IMAGE_TYPE_3D
:
1645 assert(range
->baseArrayLayer
+ radv_get_layerCount(image
, range
) - 1
1646 <= radv_minify(image
->info
.depth
, range
->baseMipLevel
));
1649 unreachable("bad VkImageType");
1651 iview
->image
= image
;
1652 iview
->bo
= image
->bo
;
1653 iview
->type
= pCreateInfo
->viewType
;
1654 iview
->plane_id
= radv_plane_from_aspect(pCreateInfo
->subresourceRange
.aspectMask
);
1655 iview
->aspect_mask
= pCreateInfo
->subresourceRange
.aspectMask
;
1656 iview
->multiple_planes
= vk_format_get_plane_count(image
->vk_format
) > 1 && iview
->aspect_mask
== VK_IMAGE_ASPECT_COLOR_BIT
;
1658 iview
->vk_format
= pCreateInfo
->format
;
1660 /* If the image has an Android external format, pCreateInfo->format will be
1661 * VK_FORMAT_UNDEFINED. */
1662 if (iview
->vk_format
== VK_FORMAT_UNDEFINED
)
1663 iview
->vk_format
= image
->vk_format
;
1665 if (iview
->aspect_mask
== VK_IMAGE_ASPECT_STENCIL_BIT
) {
1666 iview
->vk_format
= vk_format_stencil_only(iview
->vk_format
);
1667 } else if (iview
->aspect_mask
== VK_IMAGE_ASPECT_DEPTH_BIT
) {
1668 iview
->vk_format
= vk_format_depth_only(iview
->vk_format
);
1671 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
1672 iview
->extent
= (VkExtent3D
) {
1673 .width
= image
->info
.width
,
1674 .height
= image
->info
.height
,
1675 .depth
= image
->info
.depth
,
1678 iview
->extent
= (VkExtent3D
) {
1679 .width
= radv_minify(image
->info
.width
, range
->baseMipLevel
),
1680 .height
= radv_minify(image
->info
.height
, range
->baseMipLevel
),
1681 .depth
= radv_minify(image
->info
.depth
, range
->baseMipLevel
),
1685 if (iview
->vk_format
!= image
->planes
[iview
->plane_id
].format
) {
1686 unsigned view_bw
= vk_format_get_blockwidth(iview
->vk_format
);
1687 unsigned view_bh
= vk_format_get_blockheight(iview
->vk_format
);
1688 unsigned img_bw
= vk_format_get_blockwidth(image
->vk_format
);
1689 unsigned img_bh
= vk_format_get_blockheight(image
->vk_format
);
1691 iview
->extent
.width
= round_up_u32(iview
->extent
.width
* view_bw
, img_bw
);
1692 iview
->extent
.height
= round_up_u32(iview
->extent
.height
* view_bh
, img_bh
);
1694 /* Comment ported from amdvlk -
1695 * If we have the following image:
1696 * Uncompressed pixels Compressed block sizes (4x4)
1697 * mip0: 22 x 22 6 x 6
1698 * mip1: 11 x 11 3 x 3
1703 * On GFX9 the descriptor is always programmed with the WIDTH and HEIGHT of the base level and the HW is
1704 * calculating the degradation of the block sizes down the mip-chain as follows (straight-up
1705 * divide-by-two integer math):
1711 * This means that mip2 will be missing texels.
1713 * Fix this by calculating the base mip's width and height, then convert that, and round it
1714 * back up to get the level 0 size.
1715 * Clamp the converted size between the original values, and next power of two, which
1716 * means we don't oversize the image.
1718 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
&&
1719 vk_format_is_compressed(image
->vk_format
) &&
1720 !vk_format_is_compressed(iview
->vk_format
)) {
1721 unsigned lvl_width
= radv_minify(image
->info
.width
, range
->baseMipLevel
);
1722 unsigned lvl_height
= radv_minify(image
->info
.height
, range
->baseMipLevel
);
1724 lvl_width
= round_up_u32(lvl_width
* view_bw
, img_bw
);
1725 lvl_height
= round_up_u32(lvl_height
* view_bh
, img_bh
);
1727 lvl_width
<<= range
->baseMipLevel
;
1728 lvl_height
<<= range
->baseMipLevel
;
1730 iview
->extent
.width
= CLAMP(lvl_width
, iview
->extent
.width
, iview
->image
->planes
[0].surface
.u
.gfx9
.surf_pitch
);
1731 iview
->extent
.height
= CLAMP(lvl_height
, iview
->extent
.height
, iview
->image
->planes
[0].surface
.u
.gfx9
.surf_height
);
1735 iview
->base_layer
= range
->baseArrayLayer
;
1736 iview
->layer_count
= radv_get_layerCount(image
, range
);
1737 iview
->base_mip
= range
->baseMipLevel
;
1738 iview
->level_count
= radv_get_levelCount(image
, range
);
1740 bool disable_compression
= extra_create_info
? extra_create_info
->disable_compression
: false;
1741 for (unsigned i
= 0; i
< (iview
->multiple_planes
? vk_format_get_plane_count(image
->vk_format
) : 1); ++i
) {
1742 VkFormat format
= vk_format_get_plane_format(iview
->vk_format
, i
);
1743 radv_image_view_make_descriptor(iview
, device
, format
,
1744 &pCreateInfo
->components
,
1745 false, disable_compression
,
1746 iview
->plane_id
+ i
, i
);
1747 radv_image_view_make_descriptor(iview
, device
,
1748 format
, &pCreateInfo
->components
,
1749 true, disable_compression
,
1750 iview
->plane_id
+ i
, i
);
1754 bool radv_layout_is_htile_compressed(const struct radv_image
*image
,
1755 VkImageLayout layout
,
1756 bool in_render_loop
,
1757 unsigned queue_mask
)
1759 if (radv_image_is_tc_compat_htile(image
))
1760 return layout
!= VK_IMAGE_LAYOUT_GENERAL
;
1762 return radv_image_has_htile(image
) &&
1763 (layout
== VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL
||
1764 layout
== VK_IMAGE_LAYOUT_DEPTH_ATTACHMENT_OPTIMAL_KHR
||
1765 layout
== VK_IMAGE_LAYOUT_STENCIL_ATTACHMENT_OPTIMAL_KHR
||
1766 (layout
== VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL
&&
1767 queue_mask
== (1u << RADV_QUEUE_GENERAL
)));
1770 bool radv_layout_can_fast_clear(const struct radv_image
*image
,
1771 VkImageLayout layout
,
1772 bool in_render_loop
,
1773 unsigned queue_mask
)
1775 return layout
== VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL
;
1778 bool radv_layout_dcc_compressed(const struct radv_device
*device
,
1779 const struct radv_image
*image
,
1780 VkImageLayout layout
,
1781 bool in_render_loop
,
1782 unsigned queue_mask
)
1784 /* Don't compress compute transfer dst, as image stores are not supported. */
1785 if (layout
== VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL
&&
1786 (queue_mask
& (1u << RADV_QUEUE_COMPUTE
)))
1789 return radv_image_has_dcc(image
) && layout
!= VK_IMAGE_LAYOUT_GENERAL
;
1793 unsigned radv_image_queue_family_mask(const struct radv_image
*image
, uint32_t family
, uint32_t queue_family
)
1795 if (!image
->exclusive
)
1796 return image
->queue_family_mask
;
1797 if (family
== VK_QUEUE_FAMILY_EXTERNAL
||
1798 family
== VK_QUEUE_FAMILY_FOREIGN_EXT
)
1799 return (1u << RADV_MAX_QUEUE_FAMILIES
) - 1u;
1800 if (family
== VK_QUEUE_FAMILY_IGNORED
)
1801 return 1u << queue_family
;
1802 return 1u << family
;
1806 radv_CreateImage(VkDevice device
,
1807 const VkImageCreateInfo
*pCreateInfo
,
1808 const VkAllocationCallbacks
*pAllocator
,
1812 const VkNativeBufferANDROID
*gralloc_info
=
1813 vk_find_struct_const(pCreateInfo
->pNext
, NATIVE_BUFFER_ANDROID
);
1816 return radv_image_from_gralloc(device
, pCreateInfo
, gralloc_info
,
1817 pAllocator
, pImage
);
1820 const struct wsi_image_create_info
*wsi_info
=
1821 vk_find_struct_const(pCreateInfo
->pNext
, WSI_IMAGE_CREATE_INFO_MESA
);
1822 bool scanout
= wsi_info
&& wsi_info
->scanout
;
1824 return radv_image_create(device
,
1825 &(struct radv_image_create_info
) {
1826 .vk_info
= pCreateInfo
,
1834 radv_DestroyImage(VkDevice _device
, VkImage _image
,
1835 const VkAllocationCallbacks
*pAllocator
)
1837 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1838 RADV_FROM_HANDLE(radv_image
, image
, _image
);
1843 if (image
->flags
& VK_IMAGE_CREATE_SPARSE_BINDING_BIT
)
1844 device
->ws
->buffer_destroy(image
->bo
);
1846 if (image
->owned_memory
!= VK_NULL_HANDLE
)
1847 radv_FreeMemory(_device
, image
->owned_memory
, pAllocator
);
1849 vk_free2(&device
->alloc
, pAllocator
, image
);
1852 void radv_GetImageSubresourceLayout(
1855 const VkImageSubresource
* pSubresource
,
1856 VkSubresourceLayout
* pLayout
)
1858 RADV_FROM_HANDLE(radv_image
, image
, _image
);
1859 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1860 int level
= pSubresource
->mipLevel
;
1861 int layer
= pSubresource
->arrayLayer
;
1863 unsigned plane_id
= radv_plane_from_aspect(pSubresource
->aspectMask
);
1865 struct radv_image_plane
*plane
= &image
->planes
[plane_id
];
1866 struct radeon_surf
*surface
= &plane
->surface
;
1868 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
1869 uint64_t level_offset
= surface
->is_linear
? surface
->u
.gfx9
.offset
[level
] : 0;
1871 pLayout
->offset
= plane
->offset
+ level_offset
+ surface
->u
.gfx9
.surf_slice_size
* layer
;
1872 if (image
->vk_format
== VK_FORMAT_R32G32B32_UINT
||
1873 image
->vk_format
== VK_FORMAT_R32G32B32_SINT
||
1874 image
->vk_format
== VK_FORMAT_R32G32B32_SFLOAT
) {
1875 /* Adjust the number of bytes between each row because
1876 * the pitch is actually the number of components per
1879 pLayout
->rowPitch
= surface
->u
.gfx9
.surf_pitch
* surface
->bpe
/ 3;
1881 uint32_t pitch
= surface
->is_linear
? surface
->u
.gfx9
.pitch
[level
] : surface
->u
.gfx9
.surf_pitch
;
1883 assert(util_is_power_of_two_nonzero(surface
->bpe
));
1884 pLayout
->rowPitch
= pitch
* surface
->bpe
;
1887 pLayout
->arrayPitch
= surface
->u
.gfx9
.surf_slice_size
;
1888 pLayout
->depthPitch
= surface
->u
.gfx9
.surf_slice_size
;
1889 pLayout
->size
= surface
->u
.gfx9
.surf_slice_size
;
1890 if (image
->type
== VK_IMAGE_TYPE_3D
)
1891 pLayout
->size
*= u_minify(image
->info
.depth
, level
);
1893 pLayout
->offset
= plane
->offset
+ surface
->u
.legacy
.level
[level
].offset
+ (uint64_t)surface
->u
.legacy
.level
[level
].slice_size_dw
* 4 * layer
;
1894 pLayout
->rowPitch
= surface
->u
.legacy
.level
[level
].nblk_x
* surface
->bpe
;
1895 pLayout
->arrayPitch
= (uint64_t)surface
->u
.legacy
.level
[level
].slice_size_dw
* 4;
1896 pLayout
->depthPitch
= (uint64_t)surface
->u
.legacy
.level
[level
].slice_size_dw
* 4;
1897 pLayout
->size
= (uint64_t)surface
->u
.legacy
.level
[level
].slice_size_dw
* 4;
1898 if (image
->type
== VK_IMAGE_TYPE_3D
)
1899 pLayout
->size
*= u_minify(image
->info
.depth
, level
);
1905 radv_CreateImageView(VkDevice _device
,
1906 const VkImageViewCreateInfo
*pCreateInfo
,
1907 const VkAllocationCallbacks
*pAllocator
,
1910 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1911 struct radv_image_view
*view
;
1913 view
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*view
), 8,
1914 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1916 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
1918 radv_image_view_init(view
, device
, pCreateInfo
, NULL
);
1920 *pView
= radv_image_view_to_handle(view
);
1926 radv_DestroyImageView(VkDevice _device
, VkImageView _iview
,
1927 const VkAllocationCallbacks
*pAllocator
)
1929 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1930 RADV_FROM_HANDLE(radv_image_view
, iview
, _iview
);
1934 vk_free2(&device
->alloc
, pAllocator
, iview
);
1937 void radv_buffer_view_init(struct radv_buffer_view
*view
,
1938 struct radv_device
*device
,
1939 const VkBufferViewCreateInfo
* pCreateInfo
)
1941 RADV_FROM_HANDLE(radv_buffer
, buffer
, pCreateInfo
->buffer
);
1943 view
->bo
= buffer
->bo
;
1944 view
->range
= pCreateInfo
->range
== VK_WHOLE_SIZE
?
1945 buffer
->size
- pCreateInfo
->offset
: pCreateInfo
->range
;
1946 view
->vk_format
= pCreateInfo
->format
;
1948 radv_make_buffer_descriptor(device
, buffer
, view
->vk_format
,
1949 pCreateInfo
->offset
, view
->range
, view
->state
);
1953 radv_CreateBufferView(VkDevice _device
,
1954 const VkBufferViewCreateInfo
*pCreateInfo
,
1955 const VkAllocationCallbacks
*pAllocator
,
1956 VkBufferView
*pView
)
1958 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1959 struct radv_buffer_view
*view
;
1961 view
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*view
), 8,
1962 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1964 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
1966 radv_buffer_view_init(view
, device
, pCreateInfo
);
1968 *pView
= radv_buffer_view_to_handle(view
);
1974 radv_DestroyBufferView(VkDevice _device
, VkBufferView bufferView
,
1975 const VkAllocationCallbacks
*pAllocator
)
1977 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1978 RADV_FROM_HANDLE(radv_buffer_view
, view
, bufferView
);
1983 vk_free2(&device
->alloc
, pAllocator
, view
);