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 struct radv_image_create_info
*create_info
)
41 const VkImageCreateInfo
*pCreateInfo
= create_info
->vk_info
;
43 if (pCreateInfo
->tiling
== VK_IMAGE_TILING_LINEAR
) {
44 assert(pCreateInfo
->samples
<= 1);
45 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
48 if (!vk_format_is_compressed(pCreateInfo
->format
) &&
49 !vk_format_is_depth_or_stencil(pCreateInfo
->format
)
50 && device
->physical_device
->rad_info
.chip_class
<= GFX8
) {
51 /* this causes hangs in some VK CTS tests on GFX9. */
52 /* Textures with a very small height are recommended to be linear. */
53 if (pCreateInfo
->imageType
== VK_IMAGE_TYPE_1D
||
54 /* Only very thin and long 2D textures should benefit from
56 (pCreateInfo
->extent
.width
> 8 && pCreateInfo
->extent
.height
<= 2))
57 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
60 /* MSAA resources must be 2D tiled. */
61 if (pCreateInfo
->samples
> 1)
62 return RADEON_SURF_MODE_2D
;
64 return RADEON_SURF_MODE_2D
;
68 radv_use_tc_compat_htile_for_image(struct radv_device
*device
,
69 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
) ||
76 (pCreateInfo
->flags
& VK_IMAGE_CREATE_EXTENDED_USAGE_BIT
))
79 if (pCreateInfo
->tiling
== VK_IMAGE_TILING_LINEAR
)
82 if (pCreateInfo
->mipLevels
> 1)
85 /* FIXME: for some reason TC compat with 2/4/8 samples breaks some cts
86 * tests - disable for now. On GFX10 D32_SFLOAT is affected as well.
88 if (pCreateInfo
->samples
>= 2 &&
89 (pCreateInfo
->format
== VK_FORMAT_D32_SFLOAT_S8_UINT
||
90 (pCreateInfo
->format
== VK_FORMAT_D32_SFLOAT
&&
91 device
->physical_device
->rad_info
.chip_class
== GFX10
)))
94 /* GFX9 supports both 32-bit and 16-bit depth surfaces, while GFX8 only
95 * supports 32-bit. Though, it's possible to enable TC-compat for
96 * 16-bit depth surfaces if no Z planes are compressed.
98 if (pCreateInfo
->format
!= VK_FORMAT_D32_SFLOAT_S8_UINT
&&
99 pCreateInfo
->format
!= VK_FORMAT_D32_SFLOAT
&&
100 pCreateInfo
->format
!= VK_FORMAT_D16_UNORM
)
103 if (pCreateInfo
->flags
& VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT
) {
104 const struct VkImageFormatListCreateInfoKHR
*format_list
=
105 (const struct VkImageFormatListCreateInfoKHR
*)
106 vk_find_struct_const(pCreateInfo
->pNext
,
107 IMAGE_FORMAT_LIST_CREATE_INFO_KHR
);
109 /* We have to ignore the existence of the list if viewFormatCount = 0 */
110 if (format_list
&& format_list
->viewFormatCount
) {
111 /* compatibility is transitive, so we only need to check
112 * one format with everything else.
114 for (unsigned i
= 0; i
< format_list
->viewFormatCount
; ++i
) {
115 if (format_list
->pViewFormats
[i
] == VK_FORMAT_UNDEFINED
)
118 if (pCreateInfo
->format
!= format_list
->pViewFormats
[i
])
130 radv_surface_has_scanout(struct radv_device
*device
, const struct radv_image_create_info
*info
)
135 if (!info
->bo_metadata
)
138 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
139 return info
->bo_metadata
->u
.gfx9
.swizzle_mode
== 0 || info
->bo_metadata
->u
.gfx9
.swizzle_mode
% 4 == 2;
141 return info
->bo_metadata
->u
.legacy
.scanout
;
146 radv_support_storage_dcc(const struct radv_physical_device
*pdevice
)
148 return pdevice
->rad_info
.chip_class
>= GFX10
&&
149 (pdevice
->instance
->perftest_flags
& RADV_PERFTEST_STORAGE_DCC
);
153 radv_use_dcc_for_image(struct radv_device
*device
,
154 const struct radv_image
*image
,
155 const struct radv_image_create_info
*create_info
,
156 const VkImageCreateInfo
*pCreateInfo
)
158 bool dcc_compatible_formats
;
161 /* DCC (Delta Color Compression) is only available for GFX8+. */
162 if (device
->physical_device
->rad_info
.chip_class
< GFX8
)
165 if (device
->instance
->debug_flags
& RADV_DEBUG_NO_DCC
)
168 if (image
->shareable
)
171 if (((pCreateInfo
->usage
& VK_IMAGE_USAGE_STORAGE_BIT
) &&
172 !radv_support_storage_dcc(device
->physical_device
))||
173 (pCreateInfo
->flags
& VK_IMAGE_CREATE_EXTENDED_USAGE_BIT
))
176 if (pCreateInfo
->tiling
== VK_IMAGE_TILING_LINEAR
)
179 if (vk_format_is_subsampled(pCreateInfo
->format
) ||
180 vk_format_get_plane_count(pCreateInfo
->format
) > 1)
183 /* TODO: Enable DCC for mipmaps on GFX9+. */
184 if ((pCreateInfo
->arrayLayers
> 1 || pCreateInfo
->mipLevels
> 1) &&
185 device
->physical_device
->rad_info
.chip_class
>= GFX9
)
188 /* Do not enable DCC for mipmapped arrays because performance is worse. */
189 if (pCreateInfo
->arrayLayers
> 1 && pCreateInfo
->mipLevels
> 1)
192 if (radv_surface_has_scanout(device
, create_info
))
195 /* FIXME: DCC for MSAA with 4x and 8x samples doesn't work yet, while
196 * 2x can be enabled with an option.
198 if (pCreateInfo
->samples
> 2 ||
199 (pCreateInfo
->samples
== 2 &&
200 !device
->physical_device
->dcc_msaa_allowed
))
203 /* Determine if the formats are DCC compatible. */
204 dcc_compatible_formats
=
205 radv_is_colorbuffer_format_supported(pCreateInfo
->format
,
208 if (pCreateInfo
->flags
& VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT
) {
209 const struct VkImageFormatListCreateInfoKHR
*format_list
=
210 (const struct VkImageFormatListCreateInfoKHR
*)
211 vk_find_struct_const(pCreateInfo
->pNext
,
212 IMAGE_FORMAT_LIST_CREATE_INFO_KHR
);
214 /* We have to ignore the existence of the list if viewFormatCount = 0 */
215 if (format_list
&& format_list
->viewFormatCount
) {
216 /* compatibility is transitive, so we only need to check
217 * one format with everything else. */
218 for (unsigned i
= 0; i
< format_list
->viewFormatCount
; ++i
) {
219 if (format_list
->pViewFormats
[i
] == VK_FORMAT_UNDEFINED
)
222 if (!radv_dcc_formats_compatible(pCreateInfo
->format
,
223 format_list
->pViewFormats
[i
]))
224 dcc_compatible_formats
= false;
227 dcc_compatible_formats
= false;
231 if (!dcc_compatible_formats
)
238 radv_use_tc_compat_cmask_for_image(struct radv_device
*device
,
239 struct radv_image
*image
)
241 if (!(device
->instance
->perftest_flags
& RADV_PERFTEST_TC_COMPAT_CMASK
))
244 /* TC-compat CMASK is only available for GFX8+. */
245 if (device
->physical_device
->rad_info
.chip_class
< GFX8
)
248 if (image
->usage
& VK_IMAGE_USAGE_STORAGE_BIT
)
251 if (radv_image_has_dcc(image
))
254 if (!radv_image_has_cmask(image
))
261 radv_prefill_surface_from_metadata(struct radv_device
*device
,
262 struct radeon_surf
*surface
,
263 const struct radv_image_create_info
*create_info
)
265 const struct radeon_bo_metadata
*md
= create_info
->bo_metadata
;
266 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
267 if (md
->u
.gfx9
.swizzle_mode
> 0)
268 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_MODE_2D
, MODE
);
270 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_MODE_LINEAR_ALIGNED
, MODE
);
272 surface
->u
.gfx9
.surf
.swizzle_mode
= md
->u
.gfx9
.swizzle_mode
;
274 surface
->u
.legacy
.pipe_config
= md
->u
.legacy
.pipe_config
;
275 surface
->u
.legacy
.bankw
= md
->u
.legacy
.bankw
;
276 surface
->u
.legacy
.bankh
= md
->u
.legacy
.bankh
;
277 surface
->u
.legacy
.tile_split
= md
->u
.legacy
.tile_split
;
278 surface
->u
.legacy
.mtilea
= md
->u
.legacy
.mtilea
;
279 surface
->u
.legacy
.num_banks
= md
->u
.legacy
.num_banks
;
281 if (md
->u
.legacy
.macrotile
== RADEON_LAYOUT_TILED
)
282 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_MODE_2D
, MODE
);
283 else if (md
->u
.legacy
.microtile
== RADEON_LAYOUT_TILED
)
284 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_MODE_1D
, MODE
);
286 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_MODE_LINEAR_ALIGNED
, MODE
);
292 radv_init_surface(struct radv_device
*device
,
293 const struct radv_image
*image
,
294 struct radeon_surf
*surface
,
296 const struct radv_image_create_info
*create_info
)
298 const VkImageCreateInfo
*pCreateInfo
= create_info
->vk_info
;
299 unsigned array_mode
= radv_choose_tiling(device
, create_info
);
300 VkFormat format
= vk_format_get_plane_format(pCreateInfo
->format
, plane_id
);
301 const struct vk_format_description
*desc
= vk_format_description(format
);
302 bool is_depth
, is_stencil
;
304 is_depth
= vk_format_has_depth(desc
);
305 is_stencil
= vk_format_has_stencil(desc
);
307 surface
->blk_w
= vk_format_get_blockwidth(format
);
308 surface
->blk_h
= vk_format_get_blockheight(format
);
310 surface
->bpe
= vk_format_get_blocksize(vk_format_depth_only(format
));
311 /* align byte per element on dword */
312 if (surface
->bpe
== 3) {
315 if (create_info
->bo_metadata
) {
316 radv_prefill_surface_from_metadata(device
, surface
, create_info
);
318 surface
->flags
= RADEON_SURF_SET(array_mode
, MODE
);
321 switch (pCreateInfo
->imageType
){
322 case VK_IMAGE_TYPE_1D
:
323 if (pCreateInfo
->arrayLayers
> 1)
324 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_TYPE_1D_ARRAY
, TYPE
);
326 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_TYPE_1D
, TYPE
);
328 case VK_IMAGE_TYPE_2D
:
329 if (pCreateInfo
->arrayLayers
> 1)
330 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_TYPE_2D_ARRAY
, TYPE
);
332 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_TYPE_2D
, TYPE
);
334 case VK_IMAGE_TYPE_3D
:
335 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_TYPE_3D
, TYPE
);
338 unreachable("unhandled image type");
342 surface
->flags
|= RADEON_SURF_ZBUFFER
;
343 if (radv_use_tc_compat_htile_for_image(device
, pCreateInfo
))
344 surface
->flags
|= RADEON_SURF_TC_COMPATIBLE_HTILE
;
348 surface
->flags
|= RADEON_SURF_SBUFFER
;
350 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
&&
351 pCreateInfo
->imageType
== VK_IMAGE_TYPE_3D
&&
352 vk_format_get_blocksizebits(pCreateInfo
->format
) == 128 &&
353 vk_format_is_compressed(pCreateInfo
->format
))
354 surface
->flags
|= RADEON_SURF_NO_RENDER_TARGET
;
356 surface
->flags
|= RADEON_SURF_OPTIMIZE_FOR_SPACE
;
358 if (!radv_use_dcc_for_image(device
, image
, create_info
, pCreateInfo
))
359 surface
->flags
|= RADEON_SURF_DISABLE_DCC
;
361 if (radv_surface_has_scanout(device
, create_info
))
362 surface
->flags
|= RADEON_SURF_SCANOUT
;
367 static uint32_t si_get_bo_metadata_word1(struct radv_device
*device
)
369 return (ATI_VENDOR_ID
<< 16) | device
->physical_device
->rad_info
.pci_id
;
372 static inline unsigned
373 si_tile_mode_index(const struct radv_image_plane
*plane
, unsigned level
, bool stencil
)
376 return plane
->surface
.u
.legacy
.stencil_tiling_index
[level
];
378 return plane
->surface
.u
.legacy
.tiling_index
[level
];
381 static unsigned radv_map_swizzle(unsigned swizzle
)
385 return V_008F0C_SQ_SEL_Y
;
387 return V_008F0C_SQ_SEL_Z
;
389 return V_008F0C_SQ_SEL_W
;
391 return V_008F0C_SQ_SEL_0
;
393 return V_008F0C_SQ_SEL_1
;
394 default: /* VK_SWIZZLE_X */
395 return V_008F0C_SQ_SEL_X
;
400 radv_make_buffer_descriptor(struct radv_device
*device
,
401 struct radv_buffer
*buffer
,
407 const struct vk_format_description
*desc
;
409 uint64_t gpu_address
= radv_buffer_get_va(buffer
->bo
);
410 uint64_t va
= gpu_address
+ buffer
->offset
;
411 unsigned num_format
, data_format
;
413 desc
= vk_format_description(vk_format
);
414 first_non_void
= vk_format_get_first_non_void_channel(vk_format
);
415 stride
= desc
->block
.bits
/ 8;
419 state
[1] = S_008F04_BASE_ADDRESS_HI(va
>> 32) |
420 S_008F04_STRIDE(stride
);
422 if (device
->physical_device
->rad_info
.chip_class
!= GFX8
&& stride
) {
427 state
[3] = S_008F0C_DST_SEL_X(radv_map_swizzle(desc
->swizzle
[0])) |
428 S_008F0C_DST_SEL_Y(radv_map_swizzle(desc
->swizzle
[1])) |
429 S_008F0C_DST_SEL_Z(radv_map_swizzle(desc
->swizzle
[2])) |
430 S_008F0C_DST_SEL_W(radv_map_swizzle(desc
->swizzle
[3]));
432 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
433 const struct gfx10_format
*fmt
= &gfx10_format_table
[vk_format
];
435 /* OOB_SELECT chooses the out-of-bounds check:
436 * - 0: (index >= NUM_RECORDS) || (offset >= STRIDE)
437 * - 1: index >= NUM_RECORDS
438 * - 2: NUM_RECORDS == 0
439 * - 3: if SWIZZLE_ENABLE == 0: offset >= NUM_RECORDS
440 * else: swizzle_address >= NUM_RECORDS
442 state
[3] |= S_008F0C_FORMAT(fmt
->img_format
) |
443 S_008F0C_OOB_SELECT(0) |
444 S_008F0C_RESOURCE_LEVEL(1);
446 num_format
= radv_translate_buffer_numformat(desc
, first_non_void
);
447 data_format
= radv_translate_buffer_dataformat(desc
, first_non_void
);
449 assert(data_format
!= V_008F0C_BUF_DATA_FORMAT_INVALID
);
450 assert(num_format
!= ~0);
452 state
[3] |= S_008F0C_NUM_FORMAT(num_format
) |
453 S_008F0C_DATA_FORMAT(data_format
);
458 si_set_mutable_tex_desc_fields(struct radv_device
*device
,
459 struct radv_image
*image
,
460 const struct legacy_surf_level
*base_level_info
,
462 unsigned base_level
, unsigned first_level
,
463 unsigned block_width
, bool is_stencil
,
464 bool is_storage_image
, bool disable_compression
,
467 struct radv_image_plane
*plane
= &image
->planes
[plane_id
];
468 uint64_t gpu_address
= image
->bo
? radv_buffer_get_va(image
->bo
) + image
->offset
: 0;
469 uint64_t va
= gpu_address
+ plane
->offset
;
470 enum chip_class chip_class
= device
->physical_device
->rad_info
.chip_class
;
471 uint64_t meta_va
= 0;
472 if (chip_class
>= GFX9
) {
474 va
+= plane
->surface
.u
.gfx9
.stencil_offset
;
476 va
+= plane
->surface
.u
.gfx9
.surf_offset
;
478 va
+= base_level_info
->offset
;
481 if (chip_class
>= GFX9
||
482 base_level_info
->mode
== RADEON_SURF_MODE_2D
)
483 state
[0] |= plane
->surface
.tile_swizzle
;
484 state
[1] &= C_008F14_BASE_ADDRESS_HI
;
485 state
[1] |= S_008F14_BASE_ADDRESS_HI(va
>> 40);
487 if (chip_class
>= GFX8
) {
488 state
[6] &= C_008F28_COMPRESSION_EN
;
490 if (!disable_compression
&& radv_dcc_enabled(image
, first_level
)) {
491 meta_va
= gpu_address
+ image
->dcc_offset
;
492 if (chip_class
<= GFX8
)
493 meta_va
+= base_level_info
->dcc_offset
;
495 unsigned dcc_tile_swizzle
= plane
->surface
.tile_swizzle
<< 8;
496 dcc_tile_swizzle
&= plane
->surface
.dcc_alignment
- 1;
497 meta_va
|= dcc_tile_swizzle
;
498 } else if (!disable_compression
&&
499 radv_image_is_tc_compat_htile(image
)) {
500 meta_va
= gpu_address
+ image
->htile_offset
;
504 state
[6] |= S_008F28_COMPRESSION_EN(1);
505 if (chip_class
<= GFX9
)
506 state
[7] = meta_va
>> 8;
510 if (chip_class
>= GFX10
) {
511 state
[3] &= C_00A00C_SW_MODE
;
514 state
[3] |= S_00A00C_SW_MODE(plane
->surface
.u
.gfx9
.stencil
.swizzle_mode
);
516 state
[3] |= S_00A00C_SW_MODE(plane
->surface
.u
.gfx9
.surf
.swizzle_mode
);
519 state
[6] &= C_00A018_META_DATA_ADDRESS_LO
&
520 C_00A018_META_PIPE_ALIGNED
;
523 struct gfx9_surf_meta_flags meta
;
525 if (image
->dcc_offset
)
526 meta
= plane
->surface
.u
.gfx9
.dcc
;
528 meta
= plane
->surface
.u
.gfx9
.htile
;
530 state
[6] |= S_00A018_META_PIPE_ALIGNED(meta
.pipe_aligned
) |
531 S_00A018_META_DATA_ADDRESS_LO(meta_va
>> 8);
534 state
[7] = meta_va
>> 16;
535 } else if (chip_class
== GFX9
) {
536 state
[3] &= C_008F1C_SW_MODE
;
537 state
[4] &= C_008F20_PITCH
;
540 state
[3] |= S_008F1C_SW_MODE(plane
->surface
.u
.gfx9
.stencil
.swizzle_mode
);
541 state
[4] |= S_008F20_PITCH(plane
->surface
.u
.gfx9
.stencil
.epitch
);
543 state
[3] |= S_008F1C_SW_MODE(plane
->surface
.u
.gfx9
.surf
.swizzle_mode
);
544 state
[4] |= S_008F20_PITCH(plane
->surface
.u
.gfx9
.surf
.epitch
);
547 state
[5] &= C_008F24_META_DATA_ADDRESS
&
548 C_008F24_META_PIPE_ALIGNED
&
549 C_008F24_META_RB_ALIGNED
;
551 struct gfx9_surf_meta_flags meta
;
553 if (image
->dcc_offset
)
554 meta
= plane
->surface
.u
.gfx9
.dcc
;
556 meta
= plane
->surface
.u
.gfx9
.htile
;
558 state
[5] |= S_008F24_META_DATA_ADDRESS(meta_va
>> 40) |
559 S_008F24_META_PIPE_ALIGNED(meta
.pipe_aligned
) |
560 S_008F24_META_RB_ALIGNED(meta
.rb_aligned
);
564 unsigned pitch
= base_level_info
->nblk_x
* block_width
;
565 unsigned index
= si_tile_mode_index(plane
, base_level
, is_stencil
);
567 state
[3] &= C_008F1C_TILING_INDEX
;
568 state
[3] |= S_008F1C_TILING_INDEX(index
);
569 state
[4] &= C_008F20_PITCH
;
570 state
[4] |= S_008F20_PITCH(pitch
- 1);
574 static unsigned radv_tex_dim(VkImageType image_type
, VkImageViewType view_type
,
575 unsigned nr_layers
, unsigned nr_samples
, bool is_storage_image
, bool gfx9
)
577 if (view_type
== VK_IMAGE_VIEW_TYPE_CUBE
|| view_type
== VK_IMAGE_VIEW_TYPE_CUBE_ARRAY
)
578 return is_storage_image
? V_008F1C_SQ_RSRC_IMG_2D_ARRAY
: V_008F1C_SQ_RSRC_IMG_CUBE
;
580 /* GFX9 allocates 1D textures as 2D. */
581 if (gfx9
&& image_type
== VK_IMAGE_TYPE_1D
)
582 image_type
= VK_IMAGE_TYPE_2D
;
583 switch (image_type
) {
584 case VK_IMAGE_TYPE_1D
:
585 return nr_layers
> 1 ? V_008F1C_SQ_RSRC_IMG_1D_ARRAY
: V_008F1C_SQ_RSRC_IMG_1D
;
586 case VK_IMAGE_TYPE_2D
:
588 return nr_layers
> 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY
: V_008F1C_SQ_RSRC_IMG_2D_MSAA
;
590 return nr_layers
> 1 ? V_008F1C_SQ_RSRC_IMG_2D_ARRAY
: V_008F1C_SQ_RSRC_IMG_2D
;
591 case VK_IMAGE_TYPE_3D
:
592 if (view_type
== VK_IMAGE_VIEW_TYPE_3D
)
593 return V_008F1C_SQ_RSRC_IMG_3D
;
595 return V_008F1C_SQ_RSRC_IMG_2D_ARRAY
;
597 unreachable("illegal image type");
601 static unsigned gfx9_border_color_swizzle(const enum vk_swizzle swizzle
[4])
603 unsigned bc_swizzle
= V_008F20_BC_SWIZZLE_XYZW
;
605 if (swizzle
[3] == VK_SWIZZLE_X
) {
606 /* For the pre-defined border color values (white, opaque
607 * black, transparent black), the only thing that matters is
608 * that the alpha channel winds up in the correct place
609 * (because the RGB channels are all the same) so either of
610 * these enumerations will work.
612 if (swizzle
[2] == VK_SWIZZLE_Y
)
613 bc_swizzle
= V_008F20_BC_SWIZZLE_WZYX
;
615 bc_swizzle
= V_008F20_BC_SWIZZLE_WXYZ
;
616 } else if (swizzle
[0] == VK_SWIZZLE_X
) {
617 if (swizzle
[1] == VK_SWIZZLE_Y
)
618 bc_swizzle
= V_008F20_BC_SWIZZLE_XYZW
;
620 bc_swizzle
= V_008F20_BC_SWIZZLE_XWYZ
;
621 } else if (swizzle
[1] == VK_SWIZZLE_X
) {
622 bc_swizzle
= V_008F20_BC_SWIZZLE_YXWZ
;
623 } else if (swizzle
[2] == VK_SWIZZLE_X
) {
624 bc_swizzle
= V_008F20_BC_SWIZZLE_ZYXW
;
630 static bool vi_alpha_is_on_msb(struct radv_device
*device
, VkFormat format
)
632 const struct vk_format_description
*desc
= vk_format_description(format
);
634 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
&& desc
->nr_channels
== 1)
635 return desc
->swizzle
[3] == VK_SWIZZLE_X
;
637 return radv_translate_colorswap(format
, false) <= 1;
640 * Build the sampler view descriptor for a texture (GFX10).
643 gfx10_make_texture_descriptor(struct radv_device
*device
,
644 struct radv_image
*image
,
645 bool is_storage_image
,
646 VkImageViewType view_type
,
648 const VkComponentMapping
*mapping
,
649 unsigned first_level
, unsigned last_level
,
650 unsigned first_layer
, unsigned last_layer
,
651 unsigned width
, unsigned height
, unsigned depth
,
653 uint32_t *fmask_state
)
655 const struct vk_format_description
*desc
;
656 enum vk_swizzle swizzle
[4];
660 desc
= vk_format_description(vk_format
);
661 img_format
= gfx10_format_table
[vk_format
].img_format
;
663 if (desc
->colorspace
== VK_FORMAT_COLORSPACE_ZS
) {
664 const unsigned char swizzle_xxxx
[4] = {0, 0, 0, 0};
665 vk_format_compose_swizzles(mapping
, swizzle_xxxx
, swizzle
);
667 vk_format_compose_swizzles(mapping
, desc
->swizzle
, swizzle
);
670 type
= radv_tex_dim(image
->type
, view_type
, image
->info
.array_size
, image
->info
.samples
,
671 is_storage_image
, device
->physical_device
->rad_info
.chip_class
== GFX9
);
672 if (type
== V_008F1C_SQ_RSRC_IMG_1D_ARRAY
) {
674 depth
= image
->info
.array_size
;
675 } else if (type
== V_008F1C_SQ_RSRC_IMG_2D_ARRAY
||
676 type
== V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY
) {
677 if (view_type
!= VK_IMAGE_VIEW_TYPE_3D
)
678 depth
= image
->info
.array_size
;
679 } else if (type
== V_008F1C_SQ_RSRC_IMG_CUBE
)
680 depth
= image
->info
.array_size
/ 6;
683 state
[1] = S_00A004_FORMAT(img_format
) |
684 S_00A004_WIDTH_LO(width
- 1);
685 state
[2] = S_00A008_WIDTH_HI((width
- 1) >> 2) |
686 S_00A008_HEIGHT(height
- 1) |
687 S_00A008_RESOURCE_LEVEL(1);
688 state
[3] = S_00A00C_DST_SEL_X(radv_map_swizzle(swizzle
[0])) |
689 S_00A00C_DST_SEL_Y(radv_map_swizzle(swizzle
[1])) |
690 S_00A00C_DST_SEL_Z(radv_map_swizzle(swizzle
[2])) |
691 S_00A00C_DST_SEL_W(radv_map_swizzle(swizzle
[3])) |
692 S_00A00C_BASE_LEVEL(image
->info
.samples
> 1 ?
694 S_00A00C_LAST_LEVEL(image
->info
.samples
> 1 ?
695 util_logbase2(image
->info
.samples
) :
697 S_00A00C_BC_SWIZZLE(gfx9_border_color_swizzle(swizzle
)) |
699 /* Depth is the the last accessible layer on gfx9+. The hw doesn't need
700 * to know the total number of layers.
702 state
[4] = S_00A010_DEPTH(type
== V_008F1C_SQ_RSRC_IMG_3D
? depth
- 1 : last_layer
) |
703 S_00A010_BASE_ARRAY(first_layer
);
704 state
[5] = S_00A014_ARRAY_PITCH(!!(type
== V_008F1C_SQ_RSRC_IMG_3D
)) |
705 S_00A014_MAX_MIP(image
->info
.samples
> 1 ?
706 util_logbase2(image
->info
.samples
) :
707 image
->info
.levels
- 1) |
708 S_00A014_PERF_MOD(4);
712 if (radv_dcc_enabled(image
, first_level
)) {
713 state
[6] |= S_00A018_MAX_UNCOMPRESSED_BLOCK_SIZE(V_028C78_MAX_BLOCK_SIZE_256B
) |
714 S_00A018_MAX_COMPRESSED_BLOCK_SIZE(V_028C78_MAX_BLOCK_SIZE_128B
) |
715 S_00A018_ALPHA_IS_ON_MSB(vi_alpha_is_on_msb(device
, vk_format
));
718 /* Initialize the sampler view for FMASK. */
719 if (radv_image_has_fmask(image
)) {
720 uint64_t gpu_address
= radv_buffer_get_va(image
->bo
);
724 assert(image
->plane_count
== 1);
726 va
= gpu_address
+ image
->offset
+ image
->fmask_offset
;
728 switch (image
->info
.samples
) {
730 format
= V_008F0C_IMG_FORMAT_FMASK8_S2_F2
;
733 format
= V_008F0C_IMG_FORMAT_FMASK8_S4_F4
;
736 format
= V_008F0C_IMG_FORMAT_FMASK32_S8_F8
;
739 unreachable("invalid nr_samples");
742 fmask_state
[0] = (va
>> 8) | image
->planes
[0].surface
.fmask_tile_swizzle
;
743 fmask_state
[1] = S_00A004_BASE_ADDRESS_HI(va
>> 40) |
744 S_00A004_FORMAT(format
) |
745 S_00A004_WIDTH_LO(width
- 1);
746 fmask_state
[2] = S_00A008_WIDTH_HI((width
- 1) >> 2) |
747 S_00A008_HEIGHT(height
- 1) |
748 S_00A008_RESOURCE_LEVEL(1);
749 fmask_state
[3] = S_00A00C_DST_SEL_X(V_008F1C_SQ_SEL_X
) |
750 S_00A00C_DST_SEL_Y(V_008F1C_SQ_SEL_X
) |
751 S_00A00C_DST_SEL_Z(V_008F1C_SQ_SEL_X
) |
752 S_00A00C_DST_SEL_W(V_008F1C_SQ_SEL_X
) |
753 S_00A00C_SW_MODE(image
->planes
[0].surface
.u
.gfx9
.fmask
.swizzle_mode
) |
754 S_00A00C_TYPE(radv_tex_dim(image
->type
, view_type
, image
->info
.array_size
, 0, false, false));
755 fmask_state
[4] = S_00A010_DEPTH(last_layer
) |
756 S_00A010_BASE_ARRAY(first_layer
);
758 fmask_state
[6] = S_00A018_META_PIPE_ALIGNED(image
->planes
[0].surface
.u
.gfx9
.cmask
.pipe_aligned
);
760 } else if (fmask_state
)
761 memset(fmask_state
, 0, 8 * 4);
765 * Build the sampler view descriptor for a texture (SI-GFX9)
768 si_make_texture_descriptor(struct radv_device
*device
,
769 struct radv_image
*image
,
770 bool is_storage_image
,
771 VkImageViewType view_type
,
773 const VkComponentMapping
*mapping
,
774 unsigned first_level
, unsigned last_level
,
775 unsigned first_layer
, unsigned last_layer
,
776 unsigned width
, unsigned height
, unsigned depth
,
778 uint32_t *fmask_state
)
780 const struct vk_format_description
*desc
;
781 enum vk_swizzle swizzle
[4];
783 unsigned num_format
, data_format
, type
;
785 desc
= vk_format_description(vk_format
);
787 if (desc
->colorspace
== VK_FORMAT_COLORSPACE_ZS
) {
788 const unsigned char swizzle_xxxx
[4] = {0, 0, 0, 0};
789 vk_format_compose_swizzles(mapping
, swizzle_xxxx
, swizzle
);
791 vk_format_compose_swizzles(mapping
, desc
->swizzle
, swizzle
);
794 first_non_void
= vk_format_get_first_non_void_channel(vk_format
);
796 num_format
= radv_translate_tex_numformat(vk_format
, desc
, first_non_void
);
797 if (num_format
== ~0) {
801 data_format
= radv_translate_tex_dataformat(vk_format
, desc
, first_non_void
);
802 if (data_format
== ~0) {
806 /* S8 with either Z16 or Z32 HTILE need a special format. */
807 if (device
->physical_device
->rad_info
.chip_class
== GFX9
&&
808 vk_format
== VK_FORMAT_S8_UINT
&&
809 radv_image_is_tc_compat_htile(image
)) {
810 if (image
->vk_format
== VK_FORMAT_D32_SFLOAT_S8_UINT
)
811 data_format
= V_008F14_IMG_DATA_FORMAT_S8_32
;
812 else if (image
->vk_format
== VK_FORMAT_D16_UNORM_S8_UINT
)
813 data_format
= V_008F14_IMG_DATA_FORMAT_S8_16
;
815 type
= radv_tex_dim(image
->type
, view_type
, image
->info
.array_size
, image
->info
.samples
,
816 is_storage_image
, device
->physical_device
->rad_info
.chip_class
== GFX9
);
817 if (type
== V_008F1C_SQ_RSRC_IMG_1D_ARRAY
) {
819 depth
= image
->info
.array_size
;
820 } else if (type
== V_008F1C_SQ_RSRC_IMG_2D_ARRAY
||
821 type
== V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY
) {
822 if (view_type
!= VK_IMAGE_VIEW_TYPE_3D
)
823 depth
= image
->info
.array_size
;
824 } else if (type
== V_008F1C_SQ_RSRC_IMG_CUBE
)
825 depth
= image
->info
.array_size
/ 6;
828 state
[1] = (S_008F14_DATA_FORMAT(data_format
) |
829 S_008F14_NUM_FORMAT(num_format
));
830 state
[2] = (S_008F18_WIDTH(width
- 1) |
831 S_008F18_HEIGHT(height
- 1) |
832 S_008F18_PERF_MOD(4));
833 state
[3] = (S_008F1C_DST_SEL_X(radv_map_swizzle(swizzle
[0])) |
834 S_008F1C_DST_SEL_Y(radv_map_swizzle(swizzle
[1])) |
835 S_008F1C_DST_SEL_Z(radv_map_swizzle(swizzle
[2])) |
836 S_008F1C_DST_SEL_W(radv_map_swizzle(swizzle
[3])) |
837 S_008F1C_BASE_LEVEL(image
->info
.samples
> 1 ?
839 S_008F1C_LAST_LEVEL(image
->info
.samples
> 1 ?
840 util_logbase2(image
->info
.samples
) :
842 S_008F1C_TYPE(type
));
844 state
[5] = S_008F24_BASE_ARRAY(first_layer
);
848 if (device
->physical_device
->rad_info
.chip_class
== GFX9
) {
849 unsigned bc_swizzle
= gfx9_border_color_swizzle(swizzle
);
851 /* Depth is the last accessible layer on Gfx9.
852 * The hw doesn't need to know the total number of layers.
854 if (type
== V_008F1C_SQ_RSRC_IMG_3D
)
855 state
[4] |= S_008F20_DEPTH(depth
- 1);
857 state
[4] |= S_008F20_DEPTH(last_layer
);
859 state
[4] |= S_008F20_BC_SWIZZLE(bc_swizzle
);
860 state
[5] |= S_008F24_MAX_MIP(image
->info
.samples
> 1 ?
861 util_logbase2(image
->info
.samples
) :
862 image
->info
.levels
- 1);
864 state
[3] |= S_008F1C_POW2_PAD(image
->info
.levels
> 1);
865 state
[4] |= S_008F20_DEPTH(depth
- 1);
866 state
[5] |= S_008F24_LAST_ARRAY(last_layer
);
868 if (image
->dcc_offset
) {
869 state
[6] = S_008F28_ALPHA_IS_ON_MSB(vi_alpha_is_on_msb(device
, vk_format
));
871 /* The last dword is unused by hw. The shader uses it to clear
872 * bits in the first dword of sampler state.
874 if (device
->physical_device
->rad_info
.chip_class
<= GFX7
&& image
->info
.samples
<= 1) {
875 if (first_level
== last_level
)
876 state
[7] = C_008F30_MAX_ANISO_RATIO
;
878 state
[7] = 0xffffffff;
882 /* Initialize the sampler view for FMASK. */
883 if (radv_image_has_fmask(image
)) {
884 uint32_t fmask_format
, num_format
;
885 uint64_t gpu_address
= radv_buffer_get_va(image
->bo
);
888 assert(image
->plane_count
== 1);
890 va
= gpu_address
+ image
->offset
+ image
->fmask_offset
;
892 if (device
->physical_device
->rad_info
.chip_class
== GFX9
) {
893 fmask_format
= V_008F14_IMG_DATA_FORMAT_FMASK
;
894 switch (image
->info
.samples
) {
896 num_format
= V_008F14_IMG_FMASK_8_2_2
;
899 num_format
= V_008F14_IMG_FMASK_8_4_4
;
902 num_format
= V_008F14_IMG_FMASK_32_8_8
;
905 unreachable("invalid nr_samples");
908 switch (image
->info
.samples
) {
910 fmask_format
= V_008F14_IMG_DATA_FORMAT_FMASK8_S2_F2
;
913 fmask_format
= V_008F14_IMG_DATA_FORMAT_FMASK8_S4_F4
;
916 fmask_format
= V_008F14_IMG_DATA_FORMAT_FMASK32_S8_F8
;
920 fmask_format
= V_008F14_IMG_DATA_FORMAT_INVALID
;
922 num_format
= V_008F14_IMG_NUM_FORMAT_UINT
;
925 fmask_state
[0] = va
>> 8;
926 fmask_state
[0] |= image
->planes
[0].surface
.fmask_tile_swizzle
;
927 fmask_state
[1] = S_008F14_BASE_ADDRESS_HI(va
>> 40) |
928 S_008F14_DATA_FORMAT(fmask_format
) |
929 S_008F14_NUM_FORMAT(num_format
);
930 fmask_state
[2] = S_008F18_WIDTH(width
- 1) |
931 S_008F18_HEIGHT(height
- 1);
932 fmask_state
[3] = S_008F1C_DST_SEL_X(V_008F1C_SQ_SEL_X
) |
933 S_008F1C_DST_SEL_Y(V_008F1C_SQ_SEL_X
) |
934 S_008F1C_DST_SEL_Z(V_008F1C_SQ_SEL_X
) |
935 S_008F1C_DST_SEL_W(V_008F1C_SQ_SEL_X
) |
936 S_008F1C_TYPE(radv_tex_dim(image
->type
, view_type
, image
->info
.array_size
, 0, false, false));
938 fmask_state
[5] = S_008F24_BASE_ARRAY(first_layer
);
942 if (device
->physical_device
->rad_info
.chip_class
== GFX9
) {
943 fmask_state
[3] |= S_008F1C_SW_MODE(image
->planes
[0].surface
.u
.gfx9
.fmask
.swizzle_mode
);
944 fmask_state
[4] |= S_008F20_DEPTH(last_layer
) |
945 S_008F20_PITCH(image
->planes
[0].surface
.u
.gfx9
.fmask
.epitch
);
946 fmask_state
[5] |= S_008F24_META_PIPE_ALIGNED(image
->planes
[0].surface
.u
.gfx9
.cmask
.pipe_aligned
) |
947 S_008F24_META_RB_ALIGNED(image
->planes
[0].surface
.u
.gfx9
.cmask
.rb_aligned
);
949 if (radv_image_is_tc_compat_cmask(image
)) {
950 va
= gpu_address
+ image
->offset
+ image
->cmask_offset
;
952 fmask_state
[5] |= S_008F24_META_DATA_ADDRESS(va
>> 40);
953 fmask_state
[6] |= S_008F28_COMPRESSION_EN(1);
954 fmask_state
[7] |= va
>> 8;
957 fmask_state
[3] |= S_008F1C_TILING_INDEX(image
->planes
[0].surface
.u
.legacy
.fmask
.tiling_index
);
958 fmask_state
[4] |= S_008F20_DEPTH(depth
- 1) |
959 S_008F20_PITCH(image
->planes
[0].surface
.u
.legacy
.fmask
.pitch_in_pixels
- 1);
960 fmask_state
[5] |= S_008F24_LAST_ARRAY(last_layer
);
962 if (radv_image_is_tc_compat_cmask(image
)) {
963 va
= gpu_address
+ image
->offset
+ image
->cmask_offset
;
965 fmask_state
[6] |= S_008F28_COMPRESSION_EN(1);
966 fmask_state
[7] |= va
>> 8;
969 } else if (fmask_state
)
970 memset(fmask_state
, 0, 8 * 4);
974 radv_make_texture_descriptor(struct radv_device
*device
,
975 struct radv_image
*image
,
976 bool is_storage_image
,
977 VkImageViewType view_type
,
979 const VkComponentMapping
*mapping
,
980 unsigned first_level
, unsigned last_level
,
981 unsigned first_layer
, unsigned last_layer
,
982 unsigned width
, unsigned height
, unsigned depth
,
984 uint32_t *fmask_state
)
986 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
987 gfx10_make_texture_descriptor(device
, image
, is_storage_image
,
988 view_type
, vk_format
, mapping
,
989 first_level
, last_level
,
990 first_layer
, last_layer
,
991 width
, height
, depth
,
994 si_make_texture_descriptor(device
, image
, is_storage_image
,
995 view_type
, vk_format
, mapping
,
996 first_level
, last_level
,
997 first_layer
, last_layer
,
998 width
, height
, depth
,
1004 radv_query_opaque_metadata(struct radv_device
*device
,
1005 struct radv_image
*image
,
1006 struct radeon_bo_metadata
*md
)
1008 static const VkComponentMapping fixedmapping
;
1009 uint32_t desc
[8], i
;
1011 assert(image
->plane_count
== 1);
1013 /* Metadata image format format version 1:
1014 * [0] = 1 (metadata format identifier)
1015 * [1] = (VENDOR_ID << 16) | PCI_ID
1016 * [2:9] = image descriptor for the whole resource
1017 * [2] is always 0, because the base address is cleared
1018 * [9] is the DCC offset bits [39:8] from the beginning of
1020 * [10:10+LAST_LEVEL] = mipmap level offset bits [39:8] for each level
1022 md
->metadata
[0] = 1; /* metadata image format version 1 */
1024 /* TILE_MODE_INDEX is ambiguous without a PCI ID. */
1025 md
->metadata
[1] = si_get_bo_metadata_word1(device
);
1028 radv_make_texture_descriptor(device
, image
, false,
1029 (VkImageViewType
)image
->type
, image
->vk_format
,
1030 &fixedmapping
, 0, image
->info
.levels
- 1, 0,
1031 image
->info
.array_size
- 1,
1032 image
->info
.width
, image
->info
.height
,
1036 si_set_mutable_tex_desc_fields(device
, image
, &image
->planes
[0].surface
.u
.legacy
.level
[0], 0, 0, 0,
1037 image
->planes
[0].surface
.blk_w
, false, false, false, desc
);
1039 /* Clear the base address and set the relative DCC offset. */
1041 desc
[1] &= C_008F14_BASE_ADDRESS_HI
;
1042 desc
[7] = image
->dcc_offset
>> 8;
1044 /* Dwords [2:9] contain the image descriptor. */
1045 memcpy(&md
->metadata
[2], desc
, sizeof(desc
));
1047 /* Dwords [10:..] contain the mipmap level offsets. */
1048 if (device
->physical_device
->rad_info
.chip_class
<= GFX8
) {
1049 for (i
= 0; i
<= image
->info
.levels
- 1; i
++)
1050 md
->metadata
[10+i
] = image
->planes
[0].surface
.u
.legacy
.level
[i
].offset
>> 8;
1051 md
->size_metadata
= (11 + image
->info
.levels
- 1) * 4;
1053 md
->size_metadata
= 10 * 4;
1057 radv_init_metadata(struct radv_device
*device
,
1058 struct radv_image
*image
,
1059 struct radeon_bo_metadata
*metadata
)
1061 struct radeon_surf
*surface
= &image
->planes
[0].surface
;
1063 memset(metadata
, 0, sizeof(*metadata
));
1065 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
1066 metadata
->u
.gfx9
.swizzle_mode
= surface
->u
.gfx9
.surf
.swizzle_mode
;
1068 metadata
->u
.legacy
.microtile
= surface
->u
.legacy
.level
[0].mode
>= RADEON_SURF_MODE_1D
?
1069 RADEON_LAYOUT_TILED
: RADEON_LAYOUT_LINEAR
;
1070 metadata
->u
.legacy
.macrotile
= surface
->u
.legacy
.level
[0].mode
>= RADEON_SURF_MODE_2D
?
1071 RADEON_LAYOUT_TILED
: RADEON_LAYOUT_LINEAR
;
1072 metadata
->u
.legacy
.pipe_config
= surface
->u
.legacy
.pipe_config
;
1073 metadata
->u
.legacy
.bankw
= surface
->u
.legacy
.bankw
;
1074 metadata
->u
.legacy
.bankh
= surface
->u
.legacy
.bankh
;
1075 metadata
->u
.legacy
.tile_split
= surface
->u
.legacy
.tile_split
;
1076 metadata
->u
.legacy
.mtilea
= surface
->u
.legacy
.mtilea
;
1077 metadata
->u
.legacy
.num_banks
= surface
->u
.legacy
.num_banks
;
1078 metadata
->u
.legacy
.stride
= surface
->u
.legacy
.level
[0].nblk_x
* surface
->bpe
;
1079 metadata
->u
.legacy
.scanout
= (surface
->flags
& RADEON_SURF_SCANOUT
) != 0;
1081 radv_query_opaque_metadata(device
, image
, metadata
);
1085 radv_image_override_offset_stride(struct radv_device
*device
,
1086 struct radv_image
*image
,
1087 uint64_t offset
, uint32_t stride
)
1089 struct radeon_surf
*surface
= &image
->planes
[0].surface
;
1090 unsigned bpe
= vk_format_get_blocksizebits(image
->vk_format
) / 8;
1092 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
1094 surface
->u
.gfx9
.surf_pitch
= stride
;
1095 surface
->u
.gfx9
.surf_slice_size
=
1096 (uint64_t)stride
* surface
->u
.gfx9
.surf_height
* bpe
;
1098 surface
->u
.gfx9
.surf_offset
= offset
;
1100 surface
->u
.legacy
.level
[0].nblk_x
= stride
;
1101 surface
->u
.legacy
.level
[0].slice_size_dw
=
1102 ((uint64_t)stride
* surface
->u
.legacy
.level
[0].nblk_y
* bpe
) / 4;
1105 for (unsigned i
= 0; i
< ARRAY_SIZE(surface
->u
.legacy
.level
); ++i
)
1106 surface
->u
.legacy
.level
[i
].offset
+= offset
;
1113 radv_image_alloc_fmask(struct radv_device
*device
,
1114 struct radv_image
*image
)
1116 unsigned fmask_alignment
= image
->planes
[0].surface
.fmask_alignment
;
1118 image
->fmask_offset
= align64(image
->size
, fmask_alignment
);
1119 image
->size
= image
->fmask_offset
+ image
->planes
[0].surface
.fmask_size
;
1120 image
->alignment
= MAX2(image
->alignment
, fmask_alignment
);
1124 radv_image_alloc_cmask(struct radv_device
*device
,
1125 struct radv_image
*image
)
1127 unsigned cmask_alignment
= image
->planes
[0].surface
.cmask_alignment
;
1128 unsigned cmask_size
= image
->planes
[0].surface
.cmask_size
;
1129 uint32_t clear_value_size
= 0;
1134 assert(cmask_alignment
);
1136 image
->cmask_offset
= align64(image
->size
, cmask_alignment
);
1137 /* + 8 for storing the clear values */
1138 if (!image
->clear_value_offset
) {
1139 image
->clear_value_offset
= image
->cmask_offset
+ cmask_size
;
1140 clear_value_size
= 8;
1142 image
->size
= image
->cmask_offset
+ cmask_size
+ clear_value_size
;
1143 image
->alignment
= MAX2(image
->alignment
, cmask_alignment
);
1147 radv_image_alloc_dcc(struct radv_image
*image
)
1149 assert(image
->plane_count
== 1);
1151 image
->dcc_offset
= align64(image
->size
, image
->planes
[0].surface
.dcc_alignment
);
1152 /* + 24 for storing the clear values + fce pred + dcc pred for each mip */
1153 image
->clear_value_offset
= image
->dcc_offset
+ image
->planes
[0].surface
.dcc_size
;
1154 image
->fce_pred_offset
= image
->clear_value_offset
+ 8 * image
->info
.levels
;
1155 image
->dcc_pred_offset
= image
->clear_value_offset
+ 16 * image
->info
.levels
;
1156 image
->size
= image
->dcc_offset
+ image
->planes
[0].surface
.dcc_size
+ 24 * image
->info
.levels
;
1157 image
->alignment
= MAX2(image
->alignment
, image
->planes
[0].surface
.dcc_alignment
);
1161 radv_image_alloc_htile(struct radv_device
*device
, struct radv_image
*image
)
1163 image
->htile_offset
= align64(image
->size
, image
->planes
[0].surface
.htile_alignment
);
1165 /* + 8 for storing the clear values */
1166 image
->clear_value_offset
= image
->htile_offset
+ image
->planes
[0].surface
.htile_size
;
1167 image
->size
= image
->clear_value_offset
+ 8;
1168 if (radv_image_is_tc_compat_htile(image
) &&
1169 device
->physical_device
->has_tc_compat_zrange_bug
) {
1170 /* Metadata for the TC-compatible HTILE hardware bug which
1171 * have to be fixed by updating ZRANGE_PRECISION when doing
1172 * fast depth clears to 0.0f.
1174 image
->tc_compat_zrange_offset
= image
->size
;
1175 image
->size
= image
->tc_compat_zrange_offset
+ 4;
1177 image
->alignment
= align64(image
->alignment
, image
->planes
[0].surface
.htile_alignment
);
1181 radv_image_can_enable_dcc_or_cmask(struct radv_image
*image
)
1183 if (image
->info
.samples
<= 1 &&
1184 image
->info
.width
* image
->info
.height
<= 512 * 512) {
1185 /* Do not enable CMASK or DCC for small surfaces where the cost
1186 * of the eliminate pass can be higher than the benefit of fast
1187 * clear. RadeonSI does this, but the image threshold is
1193 return image
->usage
& VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT
&&
1194 (image
->exclusive
|| image
->queue_family_mask
== 1);
1198 radv_image_can_enable_dcc(struct radv_device
*device
, struct radv_image
*image
)
1200 if (!radv_image_can_enable_dcc_or_cmask(image
) ||
1201 !radv_image_has_dcc(image
))
1204 /* On GFX8, DCC layers can be interleaved and it's currently only
1205 * enabled if slice size is equal to the per slice fast clear size
1206 * because the driver assumes that portions of multiple layers are
1207 * contiguous during fast clears.
1209 if (image
->info
.array_size
> 1) {
1210 const struct legacy_surf_level
*surf_level
=
1211 &image
->planes
[0].surface
.u
.legacy
.level
[0];
1213 assert(device
->physical_device
->rad_info
.chip_class
== GFX8
);
1215 if (image
->planes
[0].surface
.dcc_slice_size
!= surf_level
->dcc_fast_clear_size
)
1223 radv_image_can_enable_cmask(struct radv_image
*image
)
1225 if (image
->planes
[0].surface
.bpe
> 8 && image
->info
.samples
== 1) {
1226 /* Do not enable CMASK for non-MSAA images (fast color clear)
1227 * because 128 bit formats are not supported, but FMASK might
1233 return radv_image_can_enable_dcc_or_cmask(image
) &&
1234 image
->info
.levels
== 1 &&
1235 image
->info
.depth
== 1 &&
1236 !image
->planes
[0].surface
.is_linear
;
1240 radv_image_can_enable_fmask(struct radv_image
*image
)
1242 return image
->info
.samples
> 1 && vk_format_is_color(image
->vk_format
);
1246 radv_image_can_enable_htile(struct radv_image
*image
)
1248 return radv_image_has_htile(image
) &&
1249 image
->info
.levels
== 1 &&
1250 image
->info
.width
* image
->info
.height
>= 8 * 8;
1253 static void radv_image_disable_dcc(struct radv_image
*image
)
1255 for (unsigned i
= 0; i
< image
->plane_count
; ++i
)
1256 image
->planes
[i
].surface
.dcc_size
= 0;
1259 static void radv_image_disable_htile(struct radv_image
*image
)
1261 for (unsigned i
= 0; i
< image
->plane_count
; ++i
)
1262 image
->planes
[i
].surface
.htile_size
= 0;
1266 radv_image_create(VkDevice _device
,
1267 const struct radv_image_create_info
*create_info
,
1268 const VkAllocationCallbacks
* alloc
,
1271 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1272 const VkImageCreateInfo
*pCreateInfo
= create_info
->vk_info
;
1273 struct radv_image
*image
= NULL
;
1274 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO
);
1276 const unsigned plane_count
= vk_format_get_plane_count(pCreateInfo
->format
);
1277 const size_t image_struct_size
= sizeof(*image
) + sizeof(struct radv_image_plane
) * plane_count
;
1279 radv_assert(pCreateInfo
->mipLevels
> 0);
1280 radv_assert(pCreateInfo
->arrayLayers
> 0);
1281 radv_assert(pCreateInfo
->samples
> 0);
1282 radv_assert(pCreateInfo
->extent
.width
> 0);
1283 radv_assert(pCreateInfo
->extent
.height
> 0);
1284 radv_assert(pCreateInfo
->extent
.depth
> 0);
1286 image
= vk_zalloc2(&device
->alloc
, alloc
, image_struct_size
, 8,
1287 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1289 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
1291 image
->type
= pCreateInfo
->imageType
;
1292 image
->info
.width
= pCreateInfo
->extent
.width
;
1293 image
->info
.height
= pCreateInfo
->extent
.height
;
1294 image
->info
.depth
= pCreateInfo
->extent
.depth
;
1295 image
->info
.samples
= pCreateInfo
->samples
;
1296 image
->info
.storage_samples
= pCreateInfo
->samples
;
1297 image
->info
.array_size
= pCreateInfo
->arrayLayers
;
1298 image
->info
.levels
= pCreateInfo
->mipLevels
;
1299 image
->info
.num_channels
= vk_format_get_nr_components(pCreateInfo
->format
);
1301 image
->vk_format
= pCreateInfo
->format
;
1302 image
->tiling
= pCreateInfo
->tiling
;
1303 image
->usage
= pCreateInfo
->usage
;
1304 image
->flags
= pCreateInfo
->flags
;
1306 image
->exclusive
= pCreateInfo
->sharingMode
== VK_SHARING_MODE_EXCLUSIVE
;
1307 if (pCreateInfo
->sharingMode
== VK_SHARING_MODE_CONCURRENT
) {
1308 for (uint32_t i
= 0; i
< pCreateInfo
->queueFamilyIndexCount
; ++i
)
1309 if (pCreateInfo
->pQueueFamilyIndices
[i
] == VK_QUEUE_FAMILY_EXTERNAL
||
1310 pCreateInfo
->pQueueFamilyIndices
[i
] == VK_QUEUE_FAMILY_FOREIGN_EXT
)
1311 image
->queue_family_mask
|= (1u << RADV_MAX_QUEUE_FAMILIES
) - 1u;
1313 image
->queue_family_mask
|= 1u << pCreateInfo
->pQueueFamilyIndices
[i
];
1316 image
->shareable
= vk_find_struct_const(pCreateInfo
->pNext
,
1317 EXTERNAL_MEMORY_IMAGE_CREATE_INFO
) != NULL
;
1318 if (!vk_format_is_depth_or_stencil(pCreateInfo
->format
) &&
1319 !radv_surface_has_scanout(device
, create_info
) && !image
->shareable
) {
1320 image
->info
.surf_index
= &device
->image_mrt_offset_counter
;
1323 image
->plane_count
= plane_count
;
1325 image
->alignment
= 1;
1326 for (unsigned plane
= 0; plane
< plane_count
; ++plane
) {
1327 struct ac_surf_info info
= image
->info
;
1328 radv_init_surface(device
, image
, &image
->planes
[plane
].surface
, plane
, create_info
);
1331 const struct vk_format_description
*desc
= vk_format_description(pCreateInfo
->format
);
1332 assert(info
.width
% desc
->width_divisor
== 0);
1333 assert(info
.height
% desc
->height_divisor
== 0);
1335 info
.width
/= desc
->width_divisor
;
1336 info
.height
/= desc
->height_divisor
;
1339 device
->ws
->surface_init(device
->ws
, &info
, &image
->planes
[plane
].surface
);
1341 image
->planes
[plane
].offset
= align(image
->size
, image
->planes
[plane
].surface
.surf_alignment
);
1342 image
->size
= image
->planes
[plane
].offset
+ image
->planes
[plane
].surface
.surf_size
;
1343 image
->alignment
= image
->planes
[plane
].surface
.surf_alignment
;
1345 image
->planes
[plane
].format
= vk_format_get_plane_format(image
->vk_format
, plane
);
1348 if (!create_info
->no_metadata_planes
) {
1349 /* Try to enable DCC first. */
1350 if (radv_image_can_enable_dcc(device
, image
)) {
1351 radv_image_alloc_dcc(image
);
1352 if (image
->info
.samples
> 1) {
1353 /* CMASK should be enabled because DCC fast
1354 * clear with MSAA needs it.
1356 assert(radv_image_can_enable_cmask(image
));
1357 radv_image_alloc_cmask(device
, image
);
1360 /* When DCC cannot be enabled, try CMASK. */
1361 radv_image_disable_dcc(image
);
1362 if (radv_image_can_enable_cmask(image
)) {
1363 radv_image_alloc_cmask(device
, image
);
1367 /* Try to enable FMASK for multisampled images. */
1368 if (radv_image_can_enable_fmask(image
)) {
1369 radv_image_alloc_fmask(device
, image
);
1371 if (radv_use_tc_compat_cmask_for_image(device
, image
))
1372 image
->tc_compatible_cmask
= true;
1374 /* Otherwise, try to enable HTILE for depth surfaces. */
1375 if (radv_image_can_enable_htile(image
) &&
1376 !(device
->instance
->debug_flags
& RADV_DEBUG_NO_HIZ
)) {
1377 image
->tc_compatible_htile
= image
->planes
[0].surface
.flags
& RADEON_SURF_TC_COMPATIBLE_HTILE
;
1378 radv_image_alloc_htile(device
, image
);
1380 radv_image_disable_htile(image
);
1384 radv_image_disable_dcc(image
);
1385 radv_image_disable_htile(image
);
1388 if (pCreateInfo
->flags
& VK_IMAGE_CREATE_SPARSE_BINDING_BIT
) {
1389 image
->alignment
= MAX2(image
->alignment
, 4096);
1390 image
->size
= align64(image
->size
, image
->alignment
);
1393 image
->bo
= device
->ws
->buffer_create(device
->ws
, image
->size
, image
->alignment
,
1394 0, RADEON_FLAG_VIRTUAL
, RADV_BO_PRIORITY_VIRTUAL
);
1396 vk_free2(&device
->alloc
, alloc
, image
);
1397 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1401 *pImage
= radv_image_to_handle(image
);
1407 radv_image_view_make_descriptor(struct radv_image_view
*iview
,
1408 struct radv_device
*device
,
1410 const VkComponentMapping
*components
,
1411 bool is_storage_image
, bool disable_compression
,
1412 unsigned plane_id
, unsigned descriptor_plane_id
)
1414 struct radv_image
*image
= iview
->image
;
1415 struct radv_image_plane
*plane
= &image
->planes
[plane_id
];
1416 const struct vk_format_description
*format_desc
= vk_format_description(image
->vk_format
);
1417 bool is_stencil
= iview
->aspect_mask
== VK_IMAGE_ASPECT_STENCIL_BIT
;
1419 union radv_descriptor
*descriptor
;
1420 uint32_t hw_level
= 0;
1422 if (is_storage_image
) {
1423 descriptor
= &iview
->storage_descriptor
;
1425 descriptor
= &iview
->descriptor
;
1428 assert(vk_format_get_plane_count(vk_format
) == 1);
1429 assert(plane
->surface
.blk_w
% vk_format_get_blockwidth(plane
->format
) == 0);
1430 blk_w
= plane
->surface
.blk_w
/ vk_format_get_blockwidth(plane
->format
) * vk_format_get_blockwidth(vk_format
);
1432 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
)
1433 hw_level
= iview
->base_mip
;
1434 radv_make_texture_descriptor(device
, image
, is_storage_image
,
1438 hw_level
, hw_level
+ iview
->level_count
- 1,
1440 iview
->base_layer
+ iview
->layer_count
- 1,
1441 iview
->extent
.width
/ (plane_id
? format_desc
->width_divisor
: 1),
1442 iview
->extent
.height
/ (plane_id
? format_desc
->height_divisor
: 1),
1443 iview
->extent
.depth
,
1444 descriptor
->plane_descriptors
[descriptor_plane_id
],
1445 descriptor_plane_id
? NULL
: descriptor
->fmask_descriptor
);
1447 const struct legacy_surf_level
*base_level_info
= NULL
;
1448 if (device
->physical_device
->rad_info
.chip_class
<= GFX9
) {
1450 base_level_info
= &plane
->surface
.u
.legacy
.stencil_level
[iview
->base_mip
];
1452 base_level_info
= &plane
->surface
.u
.legacy
.level
[iview
->base_mip
];
1455 if (is_storage_image
&& radv_image_has_dcc(iview
->image
) &&
1456 !radv_support_storage_dcc(device
->physical_device
))
1457 disable_compression
= true;
1458 si_set_mutable_tex_desc_fields(device
, image
,
1463 blk_w
, is_stencil
, is_storage_image
,
1464 disable_compression
,
1465 descriptor
->plane_descriptors
[descriptor_plane_id
]);
1469 radv_plane_from_aspect(VkImageAspectFlags mask
)
1472 case VK_IMAGE_ASPECT_PLANE_1_BIT
:
1474 case VK_IMAGE_ASPECT_PLANE_2_BIT
:
1482 radv_get_aspect_format(struct radv_image
*image
, VkImageAspectFlags mask
)
1485 case VK_IMAGE_ASPECT_PLANE_0_BIT
:
1486 return image
->planes
[0].format
;
1487 case VK_IMAGE_ASPECT_PLANE_1_BIT
:
1488 return image
->planes
[1].format
;
1489 case VK_IMAGE_ASPECT_PLANE_2_BIT
:
1490 return image
->planes
[2].format
;
1491 case VK_IMAGE_ASPECT_STENCIL_BIT
:
1492 return vk_format_stencil_only(image
->vk_format
);
1493 case VK_IMAGE_ASPECT_DEPTH_BIT
:
1494 return vk_format_depth_only(image
->vk_format
);
1495 case VK_IMAGE_ASPECT_DEPTH_BIT
| VK_IMAGE_ASPECT_STENCIL_BIT
:
1496 return vk_format_depth_only(image
->vk_format
);
1498 return image
->vk_format
;
1503 radv_image_view_init(struct radv_image_view
*iview
,
1504 struct radv_device
*device
,
1505 const VkImageViewCreateInfo
* pCreateInfo
,
1506 const struct radv_image_view_extra_create_info
* extra_create_info
)
1508 RADV_FROM_HANDLE(radv_image
, image
, pCreateInfo
->image
);
1509 const VkImageSubresourceRange
*range
= &pCreateInfo
->subresourceRange
;
1511 switch (image
->type
) {
1512 case VK_IMAGE_TYPE_1D
:
1513 case VK_IMAGE_TYPE_2D
:
1514 assert(range
->baseArrayLayer
+ radv_get_layerCount(image
, range
) - 1 <= image
->info
.array_size
);
1516 case VK_IMAGE_TYPE_3D
:
1517 assert(range
->baseArrayLayer
+ radv_get_layerCount(image
, range
) - 1
1518 <= radv_minify(image
->info
.depth
, range
->baseMipLevel
));
1521 unreachable("bad VkImageType");
1523 iview
->image
= image
;
1524 iview
->bo
= image
->bo
;
1525 iview
->type
= pCreateInfo
->viewType
;
1526 iview
->plane_id
= radv_plane_from_aspect(pCreateInfo
->subresourceRange
.aspectMask
);
1527 iview
->aspect_mask
= pCreateInfo
->subresourceRange
.aspectMask
;
1528 iview
->multiple_planes
= vk_format_get_plane_count(image
->vk_format
) > 1 && iview
->aspect_mask
== VK_IMAGE_ASPECT_COLOR_BIT
;
1529 iview
->vk_format
= pCreateInfo
->format
;
1531 if (iview
->aspect_mask
== VK_IMAGE_ASPECT_STENCIL_BIT
) {
1532 iview
->vk_format
= vk_format_stencil_only(iview
->vk_format
);
1533 } else if (iview
->aspect_mask
== VK_IMAGE_ASPECT_DEPTH_BIT
) {
1534 iview
->vk_format
= vk_format_depth_only(iview
->vk_format
);
1537 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
1538 iview
->extent
= (VkExtent3D
) {
1539 .width
= image
->info
.width
,
1540 .height
= image
->info
.height
,
1541 .depth
= image
->info
.depth
,
1544 iview
->extent
= (VkExtent3D
) {
1545 .width
= radv_minify(image
->info
.width
, range
->baseMipLevel
),
1546 .height
= radv_minify(image
->info
.height
, range
->baseMipLevel
),
1547 .depth
= radv_minify(image
->info
.depth
, range
->baseMipLevel
),
1551 if (iview
->vk_format
!= image
->planes
[iview
->plane_id
].format
) {
1552 unsigned view_bw
= vk_format_get_blockwidth(iview
->vk_format
);
1553 unsigned view_bh
= vk_format_get_blockheight(iview
->vk_format
);
1554 unsigned img_bw
= vk_format_get_blockwidth(image
->vk_format
);
1555 unsigned img_bh
= vk_format_get_blockheight(image
->vk_format
);
1557 iview
->extent
.width
= round_up_u32(iview
->extent
.width
* view_bw
, img_bw
);
1558 iview
->extent
.height
= round_up_u32(iview
->extent
.height
* view_bh
, img_bh
);
1560 /* Comment ported from amdvlk -
1561 * If we have the following image:
1562 * Uncompressed pixels Compressed block sizes (4x4)
1563 * mip0: 22 x 22 6 x 6
1564 * mip1: 11 x 11 3 x 3
1569 * On GFX9 the descriptor is always programmed with the WIDTH and HEIGHT of the base level and the HW is
1570 * calculating the degradation of the block sizes down the mip-chain as follows (straight-up
1571 * divide-by-two integer math):
1577 * This means that mip2 will be missing texels.
1579 * Fix this by calculating the base mip's width and height, then convert that, and round it
1580 * back up to get the level 0 size.
1581 * Clamp the converted size between the original values, and next power of two, which
1582 * means we don't oversize the image.
1584 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
&&
1585 vk_format_is_compressed(image
->vk_format
) &&
1586 !vk_format_is_compressed(iview
->vk_format
)) {
1587 unsigned lvl_width
= radv_minify(image
->info
.width
, range
->baseMipLevel
);
1588 unsigned lvl_height
= radv_minify(image
->info
.height
, range
->baseMipLevel
);
1590 lvl_width
= round_up_u32(lvl_width
* view_bw
, img_bw
);
1591 lvl_height
= round_up_u32(lvl_height
* view_bh
, img_bh
);
1593 lvl_width
<<= range
->baseMipLevel
;
1594 lvl_height
<<= range
->baseMipLevel
;
1596 iview
->extent
.width
= CLAMP(lvl_width
, iview
->extent
.width
, iview
->image
->planes
[0].surface
.u
.gfx9
.surf_pitch
);
1597 iview
->extent
.height
= CLAMP(lvl_height
, iview
->extent
.height
, iview
->image
->planes
[0].surface
.u
.gfx9
.surf_height
);
1601 iview
->base_layer
= range
->baseArrayLayer
;
1602 iview
->layer_count
= radv_get_layerCount(image
, range
);
1603 iview
->base_mip
= range
->baseMipLevel
;
1604 iview
->level_count
= radv_get_levelCount(image
, range
);
1606 bool disable_compression
= extra_create_info
? extra_create_info
->disable_compression
: false;
1607 for (unsigned i
= 0; i
< (iview
->multiple_planes
? vk_format_get_plane_count(image
->vk_format
) : 1); ++i
) {
1608 VkFormat format
= vk_format_get_plane_format(iview
->vk_format
, i
);
1609 radv_image_view_make_descriptor(iview
, device
, format
,
1610 &pCreateInfo
->components
,
1611 false, disable_compression
,
1612 iview
->plane_id
+ i
, i
);
1613 radv_image_view_make_descriptor(iview
, device
,
1614 format
, &pCreateInfo
->components
,
1615 true, disable_compression
,
1616 iview
->plane_id
+ i
, i
);
1620 bool radv_layout_has_htile(const struct radv_image
*image
,
1621 VkImageLayout layout
,
1622 bool in_render_loop
,
1623 unsigned queue_mask
)
1625 if (radv_image_is_tc_compat_htile(image
))
1626 return layout
!= VK_IMAGE_LAYOUT_GENERAL
;
1628 return radv_image_has_htile(image
) &&
1629 (layout
== VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL
||
1630 (layout
== VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL
&&
1631 queue_mask
== (1u << RADV_QUEUE_GENERAL
)));
1634 bool radv_layout_is_htile_compressed(const struct radv_image
*image
,
1635 VkImageLayout layout
,
1636 bool in_render_loop
,
1637 unsigned queue_mask
)
1639 if (radv_image_is_tc_compat_htile(image
))
1640 return layout
!= VK_IMAGE_LAYOUT_GENERAL
;
1642 return radv_image_has_htile(image
) &&
1643 (layout
== VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL
||
1644 (layout
== VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL
&&
1645 queue_mask
== (1u << RADV_QUEUE_GENERAL
)));
1648 bool radv_layout_can_fast_clear(const struct radv_image
*image
,
1649 VkImageLayout layout
,
1650 bool in_render_loop
,
1651 unsigned queue_mask
)
1653 return layout
== VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL
;
1656 bool radv_layout_dcc_compressed(const struct radv_device
*device
,
1657 const struct radv_image
*image
,
1658 VkImageLayout layout
,
1659 bool in_render_loop
,
1660 unsigned queue_mask
)
1662 /* Don't compress compute transfer dst, as image stores are not supported. */
1663 if (layout
== VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL
&&
1664 (queue_mask
& (1u << RADV_QUEUE_COMPUTE
)) &&
1665 !radv_support_storage_dcc(device
->physical_device
))
1668 return radv_image_has_dcc(image
) && layout
!= VK_IMAGE_LAYOUT_GENERAL
;
1672 unsigned radv_image_queue_family_mask(const struct radv_image
*image
, uint32_t family
, uint32_t queue_family
)
1674 if (!image
->exclusive
)
1675 return image
->queue_family_mask
;
1676 if (family
== VK_QUEUE_FAMILY_EXTERNAL
||
1677 family
== VK_QUEUE_FAMILY_FOREIGN_EXT
)
1678 return (1u << RADV_MAX_QUEUE_FAMILIES
) - 1u;
1679 if (family
== VK_QUEUE_FAMILY_IGNORED
)
1680 return 1u << queue_family
;
1681 return 1u << family
;
1685 radv_CreateImage(VkDevice device
,
1686 const VkImageCreateInfo
*pCreateInfo
,
1687 const VkAllocationCallbacks
*pAllocator
,
1691 const VkNativeBufferANDROID
*gralloc_info
=
1692 vk_find_struct_const(pCreateInfo
->pNext
, NATIVE_BUFFER_ANDROID
);
1695 return radv_image_from_gralloc(device
, pCreateInfo
, gralloc_info
,
1696 pAllocator
, pImage
);
1699 const struct wsi_image_create_info
*wsi_info
=
1700 vk_find_struct_const(pCreateInfo
->pNext
, WSI_IMAGE_CREATE_INFO_MESA
);
1701 bool scanout
= wsi_info
&& wsi_info
->scanout
;
1703 return radv_image_create(device
,
1704 &(struct radv_image_create_info
) {
1705 .vk_info
= pCreateInfo
,
1713 radv_DestroyImage(VkDevice _device
, VkImage _image
,
1714 const VkAllocationCallbacks
*pAllocator
)
1716 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1717 RADV_FROM_HANDLE(radv_image
, image
, _image
);
1722 if (image
->flags
& VK_IMAGE_CREATE_SPARSE_BINDING_BIT
)
1723 device
->ws
->buffer_destroy(image
->bo
);
1725 if (image
->owned_memory
!= VK_NULL_HANDLE
)
1726 radv_FreeMemory(_device
, image
->owned_memory
, pAllocator
);
1728 vk_free2(&device
->alloc
, pAllocator
, image
);
1731 void radv_GetImageSubresourceLayout(
1734 const VkImageSubresource
* pSubresource
,
1735 VkSubresourceLayout
* pLayout
)
1737 RADV_FROM_HANDLE(radv_image
, image
, _image
);
1738 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1739 int level
= pSubresource
->mipLevel
;
1740 int layer
= pSubresource
->arrayLayer
;
1742 unsigned plane_id
= radv_plane_from_aspect(pSubresource
->aspectMask
);
1744 struct radv_image_plane
*plane
= &image
->planes
[plane_id
];
1745 struct radeon_surf
*surface
= &plane
->surface
;
1747 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
1748 pLayout
->offset
= plane
->offset
+ surface
->u
.gfx9
.offset
[level
] + surface
->u
.gfx9
.surf_slice_size
* layer
;
1749 if (image
->vk_format
== VK_FORMAT_R32G32B32_UINT
||
1750 image
->vk_format
== VK_FORMAT_R32G32B32_SINT
||
1751 image
->vk_format
== VK_FORMAT_R32G32B32_SFLOAT
) {
1752 /* Adjust the number of bytes between each row because
1753 * the pitch is actually the number of components per
1756 pLayout
->rowPitch
= surface
->u
.gfx9
.surf_pitch
* surface
->bpe
/ 3;
1758 assert(util_is_power_of_two_nonzero(surface
->bpe
));
1759 pLayout
->rowPitch
= surface
->u
.gfx9
.surf_pitch
* surface
->bpe
;
1762 pLayout
->arrayPitch
= surface
->u
.gfx9
.surf_slice_size
;
1763 pLayout
->depthPitch
= surface
->u
.gfx9
.surf_slice_size
;
1764 pLayout
->size
= surface
->u
.gfx9
.surf_slice_size
;
1765 if (image
->type
== VK_IMAGE_TYPE_3D
)
1766 pLayout
->size
*= u_minify(image
->info
.depth
, level
);
1768 pLayout
->offset
= plane
->offset
+ surface
->u
.legacy
.level
[level
].offset
+ (uint64_t)surface
->u
.legacy
.level
[level
].slice_size_dw
* 4 * layer
;
1769 pLayout
->rowPitch
= surface
->u
.legacy
.level
[level
].nblk_x
* surface
->bpe
;
1770 pLayout
->arrayPitch
= (uint64_t)surface
->u
.legacy
.level
[level
].slice_size_dw
* 4;
1771 pLayout
->depthPitch
= (uint64_t)surface
->u
.legacy
.level
[level
].slice_size_dw
* 4;
1772 pLayout
->size
= (uint64_t)surface
->u
.legacy
.level
[level
].slice_size_dw
* 4;
1773 if (image
->type
== VK_IMAGE_TYPE_3D
)
1774 pLayout
->size
*= u_minify(image
->info
.depth
, level
);
1780 radv_CreateImageView(VkDevice _device
,
1781 const VkImageViewCreateInfo
*pCreateInfo
,
1782 const VkAllocationCallbacks
*pAllocator
,
1785 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1786 struct radv_image_view
*view
;
1788 view
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*view
), 8,
1789 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1791 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
1793 radv_image_view_init(view
, device
, pCreateInfo
, NULL
);
1795 *pView
= radv_image_view_to_handle(view
);
1801 radv_DestroyImageView(VkDevice _device
, VkImageView _iview
,
1802 const VkAllocationCallbacks
*pAllocator
)
1804 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1805 RADV_FROM_HANDLE(radv_image_view
, iview
, _iview
);
1809 vk_free2(&device
->alloc
, pAllocator
, iview
);
1812 void radv_buffer_view_init(struct radv_buffer_view
*view
,
1813 struct radv_device
*device
,
1814 const VkBufferViewCreateInfo
* pCreateInfo
)
1816 RADV_FROM_HANDLE(radv_buffer
, buffer
, pCreateInfo
->buffer
);
1818 view
->bo
= buffer
->bo
;
1819 view
->range
= pCreateInfo
->range
== VK_WHOLE_SIZE
?
1820 buffer
->size
- pCreateInfo
->offset
: pCreateInfo
->range
;
1821 view
->vk_format
= pCreateInfo
->format
;
1823 radv_make_buffer_descriptor(device
, buffer
, view
->vk_format
,
1824 pCreateInfo
->offset
, view
->range
, view
->state
);
1828 radv_CreateBufferView(VkDevice _device
,
1829 const VkBufferViewCreateInfo
*pCreateInfo
,
1830 const VkAllocationCallbacks
*pAllocator
,
1831 VkBufferView
*pView
)
1833 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1834 struct radv_buffer_view
*view
;
1836 view
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*view
), 8,
1837 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1839 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
1841 radv_buffer_view_init(view
, device
, pCreateInfo
);
1843 *pView
= radv_buffer_view_to_handle(view
);
1849 radv_DestroyBufferView(VkDevice _device
, VkBufferView bufferView
,
1850 const VkAllocationCallbacks
*pAllocator
)
1852 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1853 RADV_FROM_HANDLE(radv_buffer_view
, view
, bufferView
);
1858 vk_free2(&device
->alloc
, pAllocator
, view
);