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_private.h"
29 #include "vk_format.h"
31 #include "radv_radeon_winsys.h"
34 #include "util/debug.h"
35 #include "util/u_atomic.h"
37 radv_choose_tiling(struct radv_device
*Device
,
38 const struct radv_image_create_info
*create_info
)
40 const VkImageCreateInfo
*pCreateInfo
= create_info
->vk_info
;
42 if (pCreateInfo
->tiling
== VK_IMAGE_TILING_LINEAR
) {
43 assert(pCreateInfo
->samples
<= 1);
44 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
47 /* Textures with a very small height are recommended to be linear. */
48 if (pCreateInfo
->imageType
== VK_IMAGE_TYPE_1D
||
49 /* Only very thin and long 2D textures should benefit from
51 (pCreateInfo
->extent
.width
> 8 && pCreateInfo
->extent
.height
<= 2))
52 return RADEON_SURF_MODE_LINEAR_ALIGNED
;
54 /* MSAA resources must be 2D tiled. */
55 if (pCreateInfo
->samples
> 1)
56 return RADEON_SURF_MODE_2D
;
58 return RADEON_SURF_MODE_2D
;
61 radv_init_surface(struct radv_device
*device
,
62 struct radeon_surf
*surface
,
63 const struct radv_image_create_info
*create_info
)
65 const VkImageCreateInfo
*pCreateInfo
= create_info
->vk_info
;
66 unsigned array_mode
= radv_choose_tiling(device
, create_info
);
67 const struct vk_format_description
*desc
=
68 vk_format_description(pCreateInfo
->format
);
69 bool is_depth
, is_stencil
, blendable
;
71 is_depth
= vk_format_has_depth(desc
);
72 is_stencil
= vk_format_has_stencil(desc
);
74 surface
->blk_w
= vk_format_get_blockwidth(pCreateInfo
->format
);
75 surface
->blk_h
= vk_format_get_blockheight(pCreateInfo
->format
);
77 surface
->bpe
= vk_format_get_blocksize(vk_format_depth_only(pCreateInfo
->format
));
78 /* align byte per element on dword */
79 if (surface
->bpe
== 3) {
82 surface
->flags
= RADEON_SURF_SET(array_mode
, MODE
);
84 switch (pCreateInfo
->imageType
){
85 case VK_IMAGE_TYPE_1D
:
86 if (pCreateInfo
->arrayLayers
> 1)
87 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_TYPE_1D_ARRAY
, TYPE
);
89 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_TYPE_1D
, TYPE
);
91 case VK_IMAGE_TYPE_2D
:
92 if (pCreateInfo
->arrayLayers
> 1)
93 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_TYPE_2D_ARRAY
, TYPE
);
95 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_TYPE_2D
, TYPE
);
97 case VK_IMAGE_TYPE_3D
:
98 surface
->flags
|= RADEON_SURF_SET(RADEON_SURF_TYPE_3D
, TYPE
);
101 unreachable("unhandled image type");
105 surface
->flags
|= RADEON_SURF_ZBUFFER
;
109 surface
->flags
|= RADEON_SURF_SBUFFER
;
111 surface
->flags
|= RADEON_SURF_OPTIMIZE_FOR_SPACE
;
113 if ((pCreateInfo
->usage
& (VK_IMAGE_USAGE_TRANSFER_SRC_BIT
|
114 VK_IMAGE_USAGE_STORAGE_BIT
)) ||
115 (pCreateInfo
->flags
& VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT
) ||
116 (pCreateInfo
->tiling
== VK_IMAGE_TILING_LINEAR
) ||
117 device
->physical_device
->rad_info
.chip_class
< VI
||
118 create_info
->scanout
|| (device
->debug_flags
& RADV_DEBUG_NO_DCC
) ||
119 !radv_is_colorbuffer_format_supported(pCreateInfo
->format
, &blendable
))
120 surface
->flags
|= RADEON_SURF_DISABLE_DCC
;
121 if (create_info
->scanout
)
122 surface
->flags
|= RADEON_SURF_SCANOUT
;
125 #define ATI_VENDOR_ID 0x1002
126 static uint32_t si_get_bo_metadata_word1(struct radv_device
*device
)
128 return (ATI_VENDOR_ID
<< 16) | device
->physical_device
->rad_info
.pci_id
;
131 static inline unsigned
132 si_tile_mode_index(const struct radv_image
*image
, unsigned level
, bool stencil
)
135 return image
->surface
.u
.legacy
.stencil_tiling_index
[level
];
137 return image
->surface
.u
.legacy
.tiling_index
[level
];
140 static unsigned radv_map_swizzle(unsigned swizzle
)
144 return V_008F0C_SQ_SEL_Y
;
146 return V_008F0C_SQ_SEL_Z
;
148 return V_008F0C_SQ_SEL_W
;
150 return V_008F0C_SQ_SEL_0
;
152 return V_008F0C_SQ_SEL_1
;
153 default: /* VK_SWIZZLE_X */
154 return V_008F0C_SQ_SEL_X
;
159 radv_make_buffer_descriptor(struct radv_device
*device
,
160 struct radv_buffer
*buffer
,
166 const struct vk_format_description
*desc
;
168 uint64_t gpu_address
= device
->ws
->buffer_get_va(buffer
->bo
);
169 uint64_t va
= gpu_address
+ buffer
->offset
;
170 unsigned num_format
, data_format
;
172 desc
= vk_format_description(vk_format
);
173 first_non_void
= vk_format_get_first_non_void_channel(vk_format
);
174 stride
= desc
->block
.bits
/ 8;
176 num_format
= radv_translate_buffer_numformat(desc
, first_non_void
);
177 data_format
= radv_translate_buffer_dataformat(desc
, first_non_void
);
181 state
[1] = S_008F04_BASE_ADDRESS_HI(va
>> 32) |
182 S_008F04_STRIDE(stride
);
184 if (device
->physical_device
->rad_info
.chip_class
< VI
&& stride
) {
189 state
[3] = S_008F0C_DST_SEL_X(radv_map_swizzle(desc
->swizzle
[0])) |
190 S_008F0C_DST_SEL_Y(radv_map_swizzle(desc
->swizzle
[1])) |
191 S_008F0C_DST_SEL_Z(radv_map_swizzle(desc
->swizzle
[2])) |
192 S_008F0C_DST_SEL_W(radv_map_swizzle(desc
->swizzle
[3])) |
193 S_008F0C_NUM_FORMAT(num_format
) |
194 S_008F0C_DATA_FORMAT(data_format
);
198 si_set_mutable_tex_desc_fields(struct radv_device
*device
,
199 struct radv_image
*image
,
200 const struct legacy_surf_level
*base_level_info
,
201 unsigned base_level
, unsigned first_level
,
202 unsigned block_width
, bool is_stencil
,
205 uint64_t gpu_address
= image
->bo
? device
->ws
->buffer_get_va(image
->bo
) + image
->offset
: 0;
206 uint64_t va
= gpu_address
;
207 enum chip_class chip_class
= device
->physical_device
->rad_info
.chip_class
;
208 uint64_t meta_va
= 0;
209 if (chip_class
>= GFX9
) {
211 va
+= image
->surface
.u
.gfx9
.stencil_offset
;
213 va
+= image
->surface
.u
.gfx9
.surf_offset
;
215 va
+= base_level_info
->offset
;
218 if (chip_class
< GFX9
)
219 if (base_level_info
->mode
== RADEON_SURF_MODE_2D
)
220 state
[0] |= image
->surface
.tile_swizzle
;
221 state
[1] &= C_008F14_BASE_ADDRESS_HI
;
222 state
[1] |= S_008F14_BASE_ADDRESS_HI(va
>> 40);
224 if (chip_class
>= VI
) {
225 state
[6] &= C_008F28_COMPRESSION_EN
;
227 if (image
->surface
.dcc_size
&& first_level
< image
->surface
.num_dcc_levels
) {
228 meta_va
= gpu_address
+ image
->dcc_offset
;
229 if (chip_class
<= VI
)
230 meta_va
+= base_level_info
->dcc_offset
;
231 state
[6] |= S_008F28_COMPRESSION_EN(1);
232 state
[7] = meta_va
>> 8;
233 if (chip_class
< GFX9
)
234 state
[7] |= image
->surface
.tile_swizzle
;
238 if (chip_class
>= GFX9
) {
239 state
[3] &= C_008F1C_SW_MODE
;
240 state
[4] &= C_008F20_PITCH_GFX9
;
243 state
[3] |= S_008F1C_SW_MODE(image
->surface
.u
.gfx9
.stencil
.swizzle_mode
);
244 state
[4] |= S_008F20_PITCH_GFX9(image
->surface
.u
.gfx9
.stencil
.epitch
);
246 state
[3] |= S_008F1C_SW_MODE(image
->surface
.u
.gfx9
.surf
.swizzle_mode
);
247 state
[4] |= S_008F20_PITCH_GFX9(image
->surface
.u
.gfx9
.surf
.epitch
);
250 state
[5] &= C_008F24_META_DATA_ADDRESS
&
251 C_008F24_META_PIPE_ALIGNED
&
252 C_008F24_META_RB_ALIGNED
;
254 struct gfx9_surf_meta_flags meta
;
256 if (image
->dcc_offset
)
257 meta
= image
->surface
.u
.gfx9
.dcc
;
259 meta
= image
->surface
.u
.gfx9
.htile
;
261 state
[5] |= S_008F24_META_DATA_ADDRESS(meta_va
>> 40) |
262 S_008F24_META_PIPE_ALIGNED(meta
.pipe_aligned
) |
263 S_008F24_META_RB_ALIGNED(meta
.rb_aligned
);
267 unsigned pitch
= base_level_info
->nblk_x
* block_width
;
268 unsigned index
= si_tile_mode_index(image
, base_level
, is_stencil
);
270 state
[3] &= C_008F1C_TILING_INDEX
;
271 state
[3] |= S_008F1C_TILING_INDEX(index
);
272 state
[4] &= C_008F20_PITCH_GFX6
;
273 state
[4] |= S_008F20_PITCH_GFX6(pitch
- 1);
277 static unsigned radv_tex_dim(VkImageType image_type
, VkImageViewType view_type
,
278 unsigned nr_layers
, unsigned nr_samples
, bool is_storage_image
)
280 if (view_type
== VK_IMAGE_VIEW_TYPE_CUBE
|| view_type
== VK_IMAGE_VIEW_TYPE_CUBE_ARRAY
)
281 return is_storage_image
? V_008F1C_SQ_RSRC_IMG_2D_ARRAY
: V_008F1C_SQ_RSRC_IMG_CUBE
;
282 switch (image_type
) {
283 case VK_IMAGE_TYPE_1D
:
284 return nr_layers
> 1 ? V_008F1C_SQ_RSRC_IMG_1D_ARRAY
: V_008F1C_SQ_RSRC_IMG_1D
;
285 case VK_IMAGE_TYPE_2D
:
287 return nr_layers
> 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY
: V_008F1C_SQ_RSRC_IMG_2D_MSAA
;
289 return nr_layers
> 1 ? V_008F1C_SQ_RSRC_IMG_2D_ARRAY
: V_008F1C_SQ_RSRC_IMG_2D
;
290 case VK_IMAGE_TYPE_3D
:
291 if (view_type
== VK_IMAGE_VIEW_TYPE_3D
)
292 return V_008F1C_SQ_RSRC_IMG_3D
;
294 return V_008F1C_SQ_RSRC_IMG_2D_ARRAY
;
296 unreachable("illegale image type");
300 static unsigned gfx9_border_color_swizzle(const unsigned char swizzle
[4])
302 unsigned bc_swizzle
= V_008F20_BC_SWIZZLE_XYZW
;
304 if (swizzle
[3] == VK_SWIZZLE_X
) {
305 /* For the pre-defined border color values (white, opaque
306 * black, transparent black), the only thing that matters is
307 * that the alpha channel winds up in the correct place
308 * (because the RGB channels are all the same) so either of
309 * these enumerations will work.
311 if (swizzle
[2] == VK_SWIZZLE_Y
)
312 bc_swizzle
= V_008F20_BC_SWIZZLE_WZYX
;
314 bc_swizzle
= V_008F20_BC_SWIZZLE_WXYZ
;
315 } else if (swizzle
[0] == VK_SWIZZLE_X
) {
316 if (swizzle
[1] == VK_SWIZZLE_Y
)
317 bc_swizzle
= V_008F20_BC_SWIZZLE_XYZW
;
319 bc_swizzle
= V_008F20_BC_SWIZZLE_XWYZ
;
320 } else if (swizzle
[1] == VK_SWIZZLE_X
) {
321 bc_swizzle
= V_008F20_BC_SWIZZLE_YXWZ
;
322 } else if (swizzle
[2] == VK_SWIZZLE_X
) {
323 bc_swizzle
= V_008F20_BC_SWIZZLE_ZYXW
;
330 * Build the sampler view descriptor for a texture.
333 si_make_texture_descriptor(struct radv_device
*device
,
334 struct radv_image
*image
,
335 bool is_storage_image
,
336 VkImageViewType view_type
,
338 const VkComponentMapping
*mapping
,
339 unsigned first_level
, unsigned last_level
,
340 unsigned first_layer
, unsigned last_layer
,
341 unsigned width
, unsigned height
, unsigned depth
,
343 uint32_t *fmask_state
)
345 const struct vk_format_description
*desc
;
346 enum vk_swizzle swizzle
[4];
348 unsigned num_format
, data_format
, type
;
350 desc
= vk_format_description(vk_format
);
352 if (desc
->colorspace
== VK_FORMAT_COLORSPACE_ZS
) {
353 const unsigned char swizzle_xxxx
[4] = {0, 0, 0, 0};
354 vk_format_compose_swizzles(mapping
, swizzle_xxxx
, swizzle
);
356 vk_format_compose_swizzles(mapping
, desc
->swizzle
, swizzle
);
359 first_non_void
= vk_format_get_first_non_void_channel(vk_format
);
361 num_format
= radv_translate_tex_numformat(vk_format
, desc
, first_non_void
);
362 if (num_format
== ~0) {
366 data_format
= radv_translate_tex_dataformat(vk_format
, desc
, first_non_void
);
367 if (data_format
== ~0) {
371 type
= radv_tex_dim(image
->type
, view_type
, image
->info
.array_size
, image
->info
.samples
,
373 if (type
== V_008F1C_SQ_RSRC_IMG_1D_ARRAY
) {
375 depth
= image
->info
.array_size
;
376 } else if (type
== V_008F1C_SQ_RSRC_IMG_2D_ARRAY
||
377 type
== V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY
) {
378 if (view_type
!= VK_IMAGE_VIEW_TYPE_3D
)
379 depth
= image
->info
.array_size
;
380 } else if (type
== V_008F1C_SQ_RSRC_IMG_CUBE
)
381 depth
= image
->info
.array_size
/ 6;
384 state
[1] = (S_008F14_DATA_FORMAT_GFX6(data_format
) |
385 S_008F14_NUM_FORMAT_GFX6(num_format
));
386 state
[2] = (S_008F18_WIDTH(width
- 1) |
387 S_008F18_HEIGHT(height
- 1) |
388 S_008F18_PERF_MOD(4));
389 state
[3] = (S_008F1C_DST_SEL_X(radv_map_swizzle(swizzle
[0])) |
390 S_008F1C_DST_SEL_Y(radv_map_swizzle(swizzle
[1])) |
391 S_008F1C_DST_SEL_Z(radv_map_swizzle(swizzle
[2])) |
392 S_008F1C_DST_SEL_W(radv_map_swizzle(swizzle
[3])) |
393 S_008F1C_BASE_LEVEL(image
->info
.samples
> 1 ?
395 S_008F1C_LAST_LEVEL(image
->info
.samples
> 1 ?
396 util_logbase2(image
->info
.samples
) :
398 S_008F1C_TYPE(type
));
400 state
[5] = S_008F24_BASE_ARRAY(first_layer
);
404 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
405 unsigned bc_swizzle
= gfx9_border_color_swizzle(desc
->swizzle
);
407 /* Depth is the the last accessible layer on Gfx9.
408 * The hw doesn't need to know the total number of layers.
410 if (type
== V_008F1C_SQ_RSRC_IMG_3D
)
411 state
[4] |= S_008F20_DEPTH(depth
- 1);
413 state
[4] |= S_008F20_DEPTH(last_layer
);
415 state
[4] |= S_008F20_BC_SWIZZLE(bc_swizzle
);
416 state
[5] |= S_008F24_MAX_MIP(image
->info
.samples
> 1 ?
417 util_logbase2(image
->info
.samples
) :
420 state
[3] |= S_008F1C_POW2_PAD(image
->info
.levels
> 1);
421 state
[4] |= S_008F20_DEPTH(depth
- 1);
422 state
[5] |= S_008F24_LAST_ARRAY(last_layer
);
424 if (image
->dcc_offset
) {
425 unsigned swap
= radv_translate_colorswap(vk_format
, FALSE
);
427 state
[6] = S_008F28_ALPHA_IS_ON_MSB(swap
<= 1);
429 /* The last dword is unused by hw. The shader uses it to clear
430 * bits in the first dword of sampler state.
432 if (device
->physical_device
->rad_info
.chip_class
<= CIK
&& image
->info
.samples
<= 1) {
433 if (first_level
== last_level
)
434 state
[7] = C_008F30_MAX_ANISO_RATIO
;
436 state
[7] = 0xffffffff;
440 /* Initialize the sampler view for FMASK. */
441 if (image
->fmask
.size
) {
442 uint32_t fmask_format
, num_format
;
443 uint64_t gpu_address
= device
->ws
->buffer_get_va(image
->bo
);
446 va
= gpu_address
+ image
->offset
+ image
->fmask
.offset
;
448 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
449 fmask_format
= V_008F14_IMG_DATA_FORMAT_FMASK
;
450 switch (image
->info
.samples
) {
452 num_format
= V_008F14_IMG_FMASK_8_2_2
;
455 num_format
= V_008F14_IMG_FMASK_8_4_4
;
458 num_format
= V_008F14_IMG_FMASK_32_8_8
;
461 unreachable("invalid nr_samples");
464 switch (image
->info
.samples
) {
466 fmask_format
= V_008F14_IMG_DATA_FORMAT_FMASK8_S2_F2
;
469 fmask_format
= V_008F14_IMG_DATA_FORMAT_FMASK8_S4_F4
;
472 fmask_format
= V_008F14_IMG_DATA_FORMAT_FMASK32_S8_F8
;
476 fmask_format
= V_008F14_IMG_DATA_FORMAT_INVALID
;
478 num_format
= V_008F14_IMG_NUM_FORMAT_UINT
;
481 fmask_state
[0] = va
>> 8;
482 if (device
->physical_device
->rad_info
.chip_class
< GFX9
)
483 fmask_state
[0] |= image
->fmask
.tile_swizzle
;
484 fmask_state
[1] = S_008F14_BASE_ADDRESS_HI(va
>> 40) |
485 S_008F14_DATA_FORMAT_GFX6(fmask_format
) |
486 S_008F14_NUM_FORMAT_GFX6(num_format
);
487 fmask_state
[2] = S_008F18_WIDTH(width
- 1) |
488 S_008F18_HEIGHT(height
- 1);
489 fmask_state
[3] = S_008F1C_DST_SEL_X(V_008F1C_SQ_SEL_X
) |
490 S_008F1C_DST_SEL_Y(V_008F1C_SQ_SEL_X
) |
491 S_008F1C_DST_SEL_Z(V_008F1C_SQ_SEL_X
) |
492 S_008F1C_DST_SEL_W(V_008F1C_SQ_SEL_X
) |
493 S_008F1C_TYPE(radv_tex_dim(image
->type
, view_type
, 1, 0, false));
495 fmask_state
[5] = S_008F24_BASE_ARRAY(first_layer
);
499 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
500 fmask_state
[3] |= S_008F1C_SW_MODE(image
->surface
.u
.gfx9
.fmask
.swizzle_mode
);
501 fmask_state
[4] |= S_008F20_DEPTH(last_layer
) |
502 S_008F20_PITCH_GFX9(image
->surface
.u
.gfx9
.fmask
.epitch
);
503 fmask_state
[5] |= S_008F24_META_PIPE_ALIGNED(image
->surface
.u
.gfx9
.cmask
.pipe_aligned
) |
504 S_008F24_META_RB_ALIGNED(image
->surface
.u
.gfx9
.cmask
.rb_aligned
);
506 fmask_state
[3] |= S_008F1C_TILING_INDEX(image
->fmask
.tile_mode_index
);
507 fmask_state
[4] |= S_008F20_DEPTH(depth
- 1) |
508 S_008F20_PITCH_GFX6(image
->fmask
.pitch_in_pixels
- 1);
509 fmask_state
[5] |= S_008F24_LAST_ARRAY(last_layer
);
511 } else if (fmask_state
)
512 memset(fmask_state
, 0, 8 * 4);
516 radv_query_opaque_metadata(struct radv_device
*device
,
517 struct radv_image
*image
,
518 struct radeon_bo_metadata
*md
)
520 static const VkComponentMapping fixedmapping
;
523 /* Metadata image format format version 1:
524 * [0] = 1 (metadata format identifier)
525 * [1] = (VENDOR_ID << 16) | PCI_ID
526 * [2:9] = image descriptor for the whole resource
527 * [2] is always 0, because the base address is cleared
528 * [9] is the DCC offset bits [39:8] from the beginning of
530 * [10:10+LAST_LEVEL] = mipmap level offset bits [39:8] for each level
532 md
->metadata
[0] = 1; /* metadata image format version 1 */
534 /* TILE_MODE_INDEX is ambiguous without a PCI ID. */
535 md
->metadata
[1] = si_get_bo_metadata_word1(device
);
538 si_make_texture_descriptor(device
, image
, false,
539 (VkImageViewType
)image
->type
, image
->vk_format
,
540 &fixedmapping
, 0, image
->info
.levels
- 1, 0,
541 image
->info
.array_size
,
542 image
->info
.width
, image
->info
.height
,
546 si_set_mutable_tex_desc_fields(device
, image
, &image
->surface
.u
.legacy
.level
[0], 0, 0,
547 image
->surface
.blk_w
, false, desc
);
549 /* Clear the base address and set the relative DCC offset. */
551 desc
[1] &= C_008F14_BASE_ADDRESS_HI
;
552 desc
[7] = image
->dcc_offset
>> 8;
554 /* Dwords [2:9] contain the image descriptor. */
555 memcpy(&md
->metadata
[2], desc
, sizeof(desc
));
557 /* Dwords [10:..] contain the mipmap level offsets. */
558 if (device
->physical_device
->rad_info
.chip_class
<= VI
) {
559 for (i
= 0; i
<= image
->info
.levels
- 1; i
++)
560 md
->metadata
[10+i
] = image
->surface
.u
.legacy
.level
[i
].offset
>> 8;
561 md
->size_metadata
= (11 + image
->info
.levels
- 1) * 4;
566 radv_init_metadata(struct radv_device
*device
,
567 struct radv_image
*image
,
568 struct radeon_bo_metadata
*metadata
)
570 struct radeon_surf
*surface
= &image
->surface
;
572 memset(metadata
, 0, sizeof(*metadata
));
574 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
575 metadata
->u
.gfx9
.swizzle_mode
= surface
->u
.gfx9
.surf
.swizzle_mode
;
577 metadata
->u
.legacy
.microtile
= surface
->u
.legacy
.level
[0].mode
>= RADEON_SURF_MODE_1D
?
578 RADEON_LAYOUT_TILED
: RADEON_LAYOUT_LINEAR
;
579 metadata
->u
.legacy
.macrotile
= surface
->u
.legacy
.level
[0].mode
>= RADEON_SURF_MODE_2D
?
580 RADEON_LAYOUT_TILED
: RADEON_LAYOUT_LINEAR
;
581 metadata
->u
.legacy
.pipe_config
= surface
->u
.legacy
.pipe_config
;
582 metadata
->u
.legacy
.bankw
= surface
->u
.legacy
.bankw
;
583 metadata
->u
.legacy
.bankh
= surface
->u
.legacy
.bankh
;
584 metadata
->u
.legacy
.tile_split
= surface
->u
.legacy
.tile_split
;
585 metadata
->u
.legacy
.mtilea
= surface
->u
.legacy
.mtilea
;
586 metadata
->u
.legacy
.num_banks
= surface
->u
.legacy
.num_banks
;
587 metadata
->u
.legacy
.stride
= surface
->u
.legacy
.level
[0].nblk_x
* surface
->bpe
;
588 metadata
->u
.legacy
.scanout
= (surface
->flags
& RADEON_SURF_SCANOUT
) != 0;
590 radv_query_opaque_metadata(device
, image
, metadata
);
593 /* The number of samples can be specified independently of the texture. */
595 radv_image_get_fmask_info(struct radv_device
*device
,
596 struct radv_image
*image
,
598 struct radv_fmask_info
*out
)
600 /* FMASK is allocated like an ordinary texture. */
601 struct radeon_surf fmask
= {};
602 struct ac_surf_info info
= image
->info
;
603 memset(out
, 0, sizeof(*out
));
605 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
606 out
->alignment
= image
->surface
.u
.gfx9
.fmask_alignment
;
607 out
->size
= image
->surface
.u
.gfx9
.fmask_size
;
611 fmask
.blk_w
= image
->surface
.blk_w
;
612 fmask
.blk_h
= image
->surface
.blk_h
;
614 fmask
.flags
= image
->surface
.flags
| RADEON_SURF_FMASK
;
616 if (!image
->shareable
)
617 info
.surf_index
= &device
->fmask_mrt_offset_counter
;
619 /* Force 2D tiling if it wasn't set. This may occur when creating
620 * FMASK for MSAA resolve on R6xx. On R6xx, the single-sample
621 * destination buffer must have an FMASK too. */
622 fmask
.flags
= RADEON_SURF_CLR(fmask
.flags
, MODE
);
623 fmask
.flags
|= RADEON_SURF_SET(RADEON_SURF_MODE_2D
, MODE
);
625 switch (nr_samples
) {
637 device
->ws
->surface_init(device
->ws
, &info
, &fmask
);
638 assert(fmask
.u
.legacy
.level
[0].mode
== RADEON_SURF_MODE_2D
);
640 out
->slice_tile_max
= (fmask
.u
.legacy
.level
[0].nblk_x
* fmask
.u
.legacy
.level
[0].nblk_y
) / 64;
641 if (out
->slice_tile_max
)
642 out
->slice_tile_max
-= 1;
644 out
->tile_mode_index
= fmask
.u
.legacy
.tiling_index
[0];
645 out
->pitch_in_pixels
= fmask
.u
.legacy
.level
[0].nblk_x
;
646 out
->bank_height
= fmask
.u
.legacy
.bankh
;
647 out
->tile_swizzle
= fmask
.tile_swizzle
;
648 out
->alignment
= MAX2(256, fmask
.surf_alignment
);
649 out
->size
= fmask
.surf_size
;
651 assert(!out
->tile_swizzle
|| !image
->shareable
);
655 radv_image_alloc_fmask(struct radv_device
*device
,
656 struct radv_image
*image
)
658 radv_image_get_fmask_info(device
, image
, image
->info
.samples
, &image
->fmask
);
660 image
->fmask
.offset
= align64(image
->size
, image
->fmask
.alignment
);
661 image
->size
= image
->fmask
.offset
+ image
->fmask
.size
;
662 image
->alignment
= MAX2(image
->alignment
, image
->fmask
.alignment
);
666 radv_image_get_cmask_info(struct radv_device
*device
,
667 struct radv_image
*image
,
668 struct radv_cmask_info
*out
)
670 unsigned pipe_interleave_bytes
= device
->physical_device
->rad_info
.pipe_interleave_bytes
;
671 unsigned num_pipes
= device
->physical_device
->rad_info
.num_tile_pipes
;
672 unsigned cl_width
, cl_height
;
674 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
675 out
->alignment
= image
->surface
.u
.gfx9
.cmask_alignment
;
676 out
->size
= image
->surface
.u
.gfx9
.cmask_size
;
693 case 16: /* Hawaii */
702 unsigned base_align
= num_pipes
* pipe_interleave_bytes
;
704 unsigned width
= align(image
->info
.width
, cl_width
*8);
705 unsigned height
= align(image
->info
.height
, cl_height
*8);
706 unsigned slice_elements
= (width
* height
) / (8*8);
708 /* Each element of CMASK is a nibble. */
709 unsigned slice_bytes
= slice_elements
/ 2;
711 out
->slice_tile_max
= (width
* height
) / (128*128);
712 if (out
->slice_tile_max
)
713 out
->slice_tile_max
-= 1;
715 out
->alignment
= MAX2(256, base_align
);
716 out
->size
= (image
->type
== VK_IMAGE_TYPE_3D
? image
->info
.depth
: image
->info
.array_size
) *
717 align(slice_bytes
, base_align
);
721 radv_image_alloc_cmask(struct radv_device
*device
,
722 struct radv_image
*image
)
724 uint32_t clear_value_size
= 0;
725 radv_image_get_cmask_info(device
, image
, &image
->cmask
);
727 image
->cmask
.offset
= align64(image
->size
, image
->cmask
.alignment
);
728 /* + 8 for storing the clear values */
729 if (!image
->clear_value_offset
) {
730 image
->clear_value_offset
= image
->cmask
.offset
+ image
->cmask
.size
;
731 clear_value_size
= 8;
733 image
->size
= image
->cmask
.offset
+ image
->cmask
.size
+ clear_value_size
;
734 image
->alignment
= MAX2(image
->alignment
, image
->cmask
.alignment
);
738 radv_image_alloc_dcc(struct radv_device
*device
,
739 struct radv_image
*image
)
741 image
->dcc_offset
= align64(image
->size
, image
->surface
.dcc_alignment
);
742 /* + 16 for storing the clear values + dcc pred */
743 image
->clear_value_offset
= image
->dcc_offset
+ image
->surface
.dcc_size
;
744 image
->dcc_pred_offset
= image
->clear_value_offset
+ 8;
745 image
->size
= image
->dcc_offset
+ image
->surface
.dcc_size
+ 16;
746 image
->alignment
= MAX2(image
->alignment
, image
->surface
.dcc_alignment
);
750 radv_image_alloc_htile(struct radv_device
*device
,
751 struct radv_image
*image
)
753 if ((device
->debug_flags
& RADV_DEBUG_NO_HIZ
) || image
->info
.levels
> 1) {
754 image
->surface
.htile_size
= 0;
758 image
->htile_offset
= align64(image
->size
, image
->surface
.htile_alignment
);
760 /* + 8 for storing the clear values */
761 image
->clear_value_offset
= image
->htile_offset
+ image
->surface
.htile_size
;
762 image
->size
= image
->clear_value_offset
+ 8;
763 image
->alignment
= align64(image
->alignment
, image
->surface
.htile_alignment
);
767 radv_image_create(VkDevice _device
,
768 const struct radv_image_create_info
*create_info
,
769 const VkAllocationCallbacks
* alloc
,
772 RADV_FROM_HANDLE(radv_device
, device
, _device
);
773 const VkImageCreateInfo
*pCreateInfo
= create_info
->vk_info
;
774 struct radv_image
*image
= NULL
;
775 bool can_cmask_dcc
= false;
776 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO
);
778 radv_assert(pCreateInfo
->mipLevels
> 0);
779 radv_assert(pCreateInfo
->arrayLayers
> 0);
780 radv_assert(pCreateInfo
->samples
> 0);
781 radv_assert(pCreateInfo
->extent
.width
> 0);
782 radv_assert(pCreateInfo
->extent
.height
> 0);
783 radv_assert(pCreateInfo
->extent
.depth
> 0);
785 image
= vk_alloc2(&device
->alloc
, alloc
, sizeof(*image
), 8,
786 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
788 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
790 memset(image
, 0, sizeof(*image
));
791 image
->type
= pCreateInfo
->imageType
;
792 image
->info
.width
= pCreateInfo
->extent
.width
;
793 image
->info
.height
= pCreateInfo
->extent
.height
;
794 image
->info
.depth
= pCreateInfo
->extent
.depth
;
795 image
->info
.samples
= pCreateInfo
->samples
;
796 image
->info
.array_size
= pCreateInfo
->arrayLayers
;
797 image
->info
.levels
= pCreateInfo
->mipLevels
;
799 image
->vk_format
= pCreateInfo
->format
;
800 image
->tiling
= pCreateInfo
->tiling
;
801 image
->usage
= pCreateInfo
->usage
;
802 image
->flags
= pCreateInfo
->flags
;
804 image
->exclusive
= pCreateInfo
->sharingMode
== VK_SHARING_MODE_EXCLUSIVE
;
805 if (pCreateInfo
->sharingMode
== VK_SHARING_MODE_CONCURRENT
) {
806 for (uint32_t i
= 0; i
< pCreateInfo
->queueFamilyIndexCount
; ++i
)
807 if (pCreateInfo
->pQueueFamilyIndices
[i
] == VK_QUEUE_FAMILY_EXTERNAL_KHR
)
808 image
->queue_family_mask
|= (1u << RADV_MAX_QUEUE_FAMILIES
) - 1u;
810 image
->queue_family_mask
|= 1u << pCreateInfo
->pQueueFamilyIndices
[i
];
813 image
->shareable
= vk_find_struct_const(pCreateInfo
->pNext
,
814 EXTERNAL_MEMORY_IMAGE_CREATE_INFO_KHR
) != NULL
;
815 if (!vk_format_is_depth(pCreateInfo
->format
) && !create_info
->scanout
&& !image
->shareable
) {
816 image
->info
.surf_index
= &device
->image_mrt_offset_counter
;
819 radv_init_surface(device
, &image
->surface
, create_info
);
821 device
->ws
->surface_init(device
->ws
, &image
->info
, &image
->surface
);
823 image
->size
= image
->surface
.surf_size
;
824 image
->alignment
= image
->surface
.surf_alignment
;
826 if (image
->exclusive
|| image
->queue_family_mask
== 1)
827 can_cmask_dcc
= true;
829 if ((pCreateInfo
->usage
& VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT
) &&
830 image
->surface
.dcc_size
&& can_cmask_dcc
)
831 radv_image_alloc_dcc(device
, image
);
833 image
->surface
.dcc_size
= 0;
835 if ((pCreateInfo
->usage
& VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT
) &&
836 pCreateInfo
->mipLevels
== 1 &&
837 !image
->surface
.dcc_size
&& image
->info
.depth
== 1 && can_cmask_dcc
)
838 radv_image_alloc_cmask(device
, image
);
839 if (image
->info
.samples
> 1 && vk_format_is_color(pCreateInfo
->format
)) {
840 radv_image_alloc_fmask(device
, image
);
841 } else if (vk_format_is_depth(pCreateInfo
->format
)) {
843 radv_image_alloc_htile(device
, image
);
846 if (pCreateInfo
->flags
& VK_IMAGE_CREATE_SPARSE_BINDING_BIT
) {
847 image
->alignment
= MAX2(image
->alignment
, 4096);
848 image
->size
= align64(image
->size
, image
->alignment
);
851 image
->bo
= device
->ws
->buffer_create(device
->ws
, image
->size
, image
->alignment
,
852 0, RADEON_FLAG_VIRTUAL
);
854 vk_free2(&device
->alloc
, alloc
, image
);
855 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
859 *pImage
= radv_image_to_handle(image
);
865 radv_image_view_make_descriptor(struct radv_image_view
*iview
,
866 struct radv_device
*device
,
867 const VkImageViewCreateInfo
* pCreateInfo
,
868 bool is_storage_image
)
870 RADV_FROM_HANDLE(radv_image
, image
, pCreateInfo
->image
);
871 const VkImageSubresourceRange
*range
= &pCreateInfo
->subresourceRange
;
872 bool is_stencil
= iview
->aspect_mask
== VK_IMAGE_ASPECT_STENCIL_BIT
;
874 uint32_t *descriptor
;
875 uint32_t *fmask_descriptor
;
877 if (is_storage_image
) {
878 descriptor
= iview
->storage_descriptor
;
879 fmask_descriptor
= iview
->storage_fmask_descriptor
;
881 descriptor
= iview
->descriptor
;
882 fmask_descriptor
= iview
->fmask_descriptor
;
885 assert(image
->surface
.blk_w
% vk_format_get_blockwidth(image
->vk_format
) == 0);
886 blk_w
= image
->surface
.blk_w
/ vk_format_get_blockwidth(image
->vk_format
) * vk_format_get_blockwidth(iview
->vk_format
);
888 si_make_texture_descriptor(device
, image
, is_storage_image
,
891 &pCreateInfo
->components
,
892 0, radv_get_levelCount(image
, range
) - 1,
893 range
->baseArrayLayer
,
894 range
->baseArrayLayer
+ radv_get_layerCount(image
, range
) - 1,
896 iview
->extent
.height
,
900 si_set_mutable_tex_desc_fields(device
, image
,
901 is_stencil
? &image
->surface
.u
.legacy
.stencil_level
[range
->baseMipLevel
]
902 : &image
->surface
.u
.legacy
.level
[range
->baseMipLevel
],
905 blk_w
, is_stencil
, descriptor
);
909 radv_image_view_init(struct radv_image_view
*iview
,
910 struct radv_device
*device
,
911 const VkImageViewCreateInfo
* pCreateInfo
)
913 RADV_FROM_HANDLE(radv_image
, image
, pCreateInfo
->image
);
914 const VkImageSubresourceRange
*range
= &pCreateInfo
->subresourceRange
;
916 switch (image
->type
) {
917 case VK_IMAGE_TYPE_1D
:
918 case VK_IMAGE_TYPE_2D
:
919 assert(range
->baseArrayLayer
+ radv_get_layerCount(image
, range
) - 1 <= image
->info
.array_size
);
921 case VK_IMAGE_TYPE_3D
:
922 assert(range
->baseArrayLayer
+ radv_get_layerCount(image
, range
) - 1
923 <= radv_minify(image
->info
.depth
, range
->baseMipLevel
));
926 unreachable("bad VkImageType");
928 iview
->image
= image
;
929 iview
->bo
= image
->bo
;
930 iview
->type
= pCreateInfo
->viewType
;
931 iview
->vk_format
= pCreateInfo
->format
;
932 iview
->aspect_mask
= pCreateInfo
->subresourceRange
.aspectMask
;
934 if (iview
->aspect_mask
== VK_IMAGE_ASPECT_STENCIL_BIT
) {
935 iview
->vk_format
= vk_format_stencil_only(iview
->vk_format
);
936 } else if (iview
->aspect_mask
== VK_IMAGE_ASPECT_DEPTH_BIT
) {
937 iview
->vk_format
= vk_format_depth_only(iview
->vk_format
);
940 iview
->extent
= (VkExtent3D
) {
941 .width
= radv_minify(image
->info
.width
, range
->baseMipLevel
),
942 .height
= radv_minify(image
->info
.height
, range
->baseMipLevel
),
943 .depth
= radv_minify(image
->info
.depth
, range
->baseMipLevel
),
946 iview
->extent
.width
= round_up_u32(iview
->extent
.width
* vk_format_get_blockwidth(iview
->vk_format
),
947 vk_format_get_blockwidth(image
->vk_format
));
948 iview
->extent
.height
= round_up_u32(iview
->extent
.height
* vk_format_get_blockheight(iview
->vk_format
),
949 vk_format_get_blockheight(image
->vk_format
));
951 iview
->base_layer
= range
->baseArrayLayer
;
952 iview
->layer_count
= radv_get_layerCount(image
, range
);
953 iview
->base_mip
= range
->baseMipLevel
;
955 radv_image_view_make_descriptor(iview
, device
, pCreateInfo
, false);
956 radv_image_view_make_descriptor(iview
, device
, pCreateInfo
, true);
959 bool radv_layout_has_htile(const struct radv_image
*image
,
960 VkImageLayout layout
,
963 return image
->surface
.htile_size
&&
964 (layout
== VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL
||
965 layout
== VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL
) &&
966 queue_mask
== (1u << RADV_QUEUE_GENERAL
);
969 bool radv_layout_is_htile_compressed(const struct radv_image
*image
,
970 VkImageLayout layout
,
973 return image
->surface
.htile_size
&&
974 (layout
== VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL
||
975 layout
== VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL
) &&
976 queue_mask
== (1u << RADV_QUEUE_GENERAL
);
979 bool radv_layout_can_fast_clear(const struct radv_image
*image
,
980 VkImageLayout layout
,
983 return layout
== VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL
&&
984 queue_mask
== (1u << RADV_QUEUE_GENERAL
);
988 unsigned radv_image_queue_family_mask(const struct radv_image
*image
, uint32_t family
, uint32_t queue_family
)
990 if (!image
->exclusive
)
991 return image
->queue_family_mask
;
992 if (family
== VK_QUEUE_FAMILY_EXTERNAL_KHR
)
993 return (1u << RADV_MAX_QUEUE_FAMILIES
) - 1u;
994 if (family
== VK_QUEUE_FAMILY_IGNORED
)
995 return 1u << queue_family
;
1000 radv_CreateImage(VkDevice device
,
1001 const VkImageCreateInfo
*pCreateInfo
,
1002 const VkAllocationCallbacks
*pAllocator
,
1005 return radv_image_create(device
,
1006 &(struct radv_image_create_info
) {
1007 .vk_info
= pCreateInfo
,
1015 radv_DestroyImage(VkDevice _device
, VkImage _image
,
1016 const VkAllocationCallbacks
*pAllocator
)
1018 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1019 RADV_FROM_HANDLE(radv_image
, image
, _image
);
1024 if (image
->flags
& VK_IMAGE_CREATE_SPARSE_BINDING_BIT
)
1025 device
->ws
->buffer_destroy(image
->bo
);
1027 vk_free2(&device
->alloc
, pAllocator
, image
);
1030 void radv_GetImageSubresourceLayout(
1033 const VkImageSubresource
* pSubresource
,
1034 VkSubresourceLayout
* pLayout
)
1036 RADV_FROM_HANDLE(radv_image
, image
, _image
);
1037 int level
= pSubresource
->mipLevel
;
1038 int layer
= pSubresource
->arrayLayer
;
1039 struct radeon_surf
*surface
= &image
->surface
;
1041 pLayout
->offset
= surface
->u
.legacy
.level
[level
].offset
+ surface
->u
.legacy
.level
[level
].slice_size
* layer
;
1042 pLayout
->rowPitch
= surface
->u
.legacy
.level
[level
].nblk_x
* surface
->bpe
;
1043 pLayout
->arrayPitch
= surface
->u
.legacy
.level
[level
].slice_size
;
1044 pLayout
->depthPitch
= surface
->u
.legacy
.level
[level
].slice_size
;
1045 pLayout
->size
= surface
->u
.legacy
.level
[level
].slice_size
;
1046 if (image
->type
== VK_IMAGE_TYPE_3D
)
1047 pLayout
->size
*= u_minify(image
->info
.depth
, level
);
1052 radv_CreateImageView(VkDevice _device
,
1053 const VkImageViewCreateInfo
*pCreateInfo
,
1054 const VkAllocationCallbacks
*pAllocator
,
1057 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1058 struct radv_image_view
*view
;
1060 view
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*view
), 8,
1061 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1063 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1065 radv_image_view_init(view
, device
, pCreateInfo
);
1067 *pView
= radv_image_view_to_handle(view
);
1073 radv_DestroyImageView(VkDevice _device
, VkImageView _iview
,
1074 const VkAllocationCallbacks
*pAllocator
)
1076 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1077 RADV_FROM_HANDLE(radv_image_view
, iview
, _iview
);
1081 vk_free2(&device
->alloc
, pAllocator
, iview
);
1084 void radv_buffer_view_init(struct radv_buffer_view
*view
,
1085 struct radv_device
*device
,
1086 const VkBufferViewCreateInfo
* pCreateInfo
,
1087 struct radv_cmd_buffer
*cmd_buffer
)
1089 RADV_FROM_HANDLE(radv_buffer
, buffer
, pCreateInfo
->buffer
);
1091 view
->bo
= buffer
->bo
;
1092 view
->range
= pCreateInfo
->range
== VK_WHOLE_SIZE
?
1093 buffer
->size
- pCreateInfo
->offset
: pCreateInfo
->range
;
1094 view
->vk_format
= pCreateInfo
->format
;
1096 radv_make_buffer_descriptor(device
, buffer
, view
->vk_format
,
1097 pCreateInfo
->offset
, view
->range
, view
->state
);
1101 radv_CreateBufferView(VkDevice _device
,
1102 const VkBufferViewCreateInfo
*pCreateInfo
,
1103 const VkAllocationCallbacks
*pAllocator
,
1104 VkBufferView
*pView
)
1106 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1107 struct radv_buffer_view
*view
;
1109 view
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*view
), 8,
1110 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1112 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1114 radv_buffer_view_init(view
, device
, pCreateInfo
, NULL
);
1116 *pView
= radv_buffer_view_to_handle(view
);
1122 radv_DestroyBufferView(VkDevice _device
, VkBufferView bufferView
,
1123 const VkAllocationCallbacks
*pAllocator
)
1125 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1126 RADV_FROM_HANDLE(radv_buffer_view
, view
, bufferView
);
1131 vk_free2(&device
->alloc
, pAllocator
, view
);