2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
30 #include <drm_fourcc.h>
32 #include "anv_private.h"
33 #include "util/debug.h"
35 #include "util/u_math.h"
37 #include "vk_format_info.h"
39 static isl_surf_usage_flags_t
40 choose_isl_surf_usage(VkImageCreateFlags vk_create_flags
,
41 VkImageUsageFlags vk_usage
,
42 isl_surf_usage_flags_t isl_extra_usage
,
43 VkImageAspectFlagBits aspect
)
45 isl_surf_usage_flags_t isl_usage
= isl_extra_usage
;
47 if (vk_usage
& VK_IMAGE_USAGE_SAMPLED_BIT
)
48 isl_usage
|= ISL_SURF_USAGE_TEXTURE_BIT
;
50 if (vk_usage
& VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT
)
51 isl_usage
|= ISL_SURF_USAGE_TEXTURE_BIT
;
53 if (vk_usage
& VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT
)
54 isl_usage
|= ISL_SURF_USAGE_RENDER_TARGET_BIT
;
56 if (vk_create_flags
& VK_IMAGE_CREATE_CUBE_COMPATIBLE_BIT
)
57 isl_usage
|= ISL_SURF_USAGE_CUBE_BIT
;
59 /* Even if we're only using it for transfer operations, clears to depth and
60 * stencil images happen as depth and stencil so they need the right ISL
61 * usage bits or else things will fall apart.
64 case VK_IMAGE_ASPECT_DEPTH_BIT
:
65 isl_usage
|= ISL_SURF_USAGE_DEPTH_BIT
;
67 case VK_IMAGE_ASPECT_STENCIL_BIT
:
68 isl_usage
|= ISL_SURF_USAGE_STENCIL_BIT
;
70 case VK_IMAGE_ASPECT_COLOR_BIT
:
71 case VK_IMAGE_ASPECT_PLANE_0_BIT
:
72 case VK_IMAGE_ASPECT_PLANE_1_BIT
:
73 case VK_IMAGE_ASPECT_PLANE_2_BIT
:
76 unreachable("bad VkImageAspect");
79 if (vk_usage
& VK_IMAGE_USAGE_TRANSFER_SRC_BIT
) {
80 /* blorp implements transfers by sampling from the source image. */
81 isl_usage
|= ISL_SURF_USAGE_TEXTURE_BIT
;
84 if (vk_usage
& VK_IMAGE_USAGE_TRANSFER_DST_BIT
&&
85 aspect
== VK_IMAGE_ASPECT_COLOR_BIT
) {
86 /* blorp implements transfers by rendering into the destination image.
87 * Only request this with color images, as we deal with depth/stencil
88 * formats differently. */
89 isl_usage
|= ISL_SURF_USAGE_RENDER_TARGET_BIT
;
95 static isl_tiling_flags_t
96 choose_isl_tiling_flags(const struct anv_image_create_info
*anv_info
,
97 const struct isl_drm_modifier_info
*isl_mod_info
,
100 const VkImageCreateInfo
*base_info
= anv_info
->vk_info
;
101 isl_tiling_flags_t flags
= 0;
103 switch (base_info
->tiling
) {
105 unreachable("bad VkImageTiling");
106 case VK_IMAGE_TILING_OPTIMAL
:
107 flags
= ISL_TILING_ANY_MASK
;
109 case VK_IMAGE_TILING_LINEAR
:
110 flags
= ISL_TILING_LINEAR_BIT
;
114 if (anv_info
->isl_tiling_flags
)
115 flags
&= anv_info
->isl_tiling_flags
;
118 flags
&= ISL_TILING_LINEAR_BIT
| ISL_TILING_X_BIT
;
121 flags
&= 1 << isl_mod_info
->tiling
;
128 static struct anv_surface
*
129 get_surface(struct anv_image
*image
, VkImageAspectFlagBits aspect
)
131 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
, aspect
);
132 return &image
->planes
[plane
].surface
;
136 add_surface(struct anv_image
*image
, struct anv_surface
*surf
, uint32_t plane
)
138 assert(surf
->isl
.size_B
> 0); /* isl surface must be initialized */
140 if (image
->disjoint
) {
141 surf
->offset
= align_u32(image
->planes
[plane
].size
,
142 surf
->isl
.alignment_B
);
143 /* Plane offset is always 0 when it's disjoint. */
145 surf
->offset
= align_u32(image
->size
, surf
->isl
.alignment_B
);
146 /* Determine plane's offset only once when the first surface is added. */
147 if (image
->planes
[plane
].size
== 0)
148 image
->planes
[plane
].offset
= image
->size
;
151 image
->size
= surf
->offset
+ surf
->isl
.size_B
;
152 image
->planes
[plane
].size
= (surf
->offset
+ surf
->isl
.size_B
) - image
->planes
[plane
].offset
;
154 image
->alignment
= MAX2(image
->alignment
, surf
->isl
.alignment_B
);
155 image
->planes
[plane
].alignment
= MAX2(image
->planes
[plane
].alignment
,
156 surf
->isl
.alignment_B
);
161 all_formats_ccs_e_compatible(const struct gen_device_info
*devinfo
,
162 const struct VkImageCreateInfo
*vk_info
)
164 enum isl_format format
=
165 anv_get_isl_format(devinfo
, vk_info
->format
,
166 VK_IMAGE_ASPECT_COLOR_BIT
, vk_info
->tiling
);
168 if (!isl_format_supports_ccs_e(devinfo
, format
))
171 if (!(vk_info
->flags
& VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT
))
174 const VkImageFormatListCreateInfoKHR
*fmt_list
=
175 vk_find_struct_const(vk_info
->pNext
, IMAGE_FORMAT_LIST_CREATE_INFO_KHR
);
177 if (!fmt_list
|| fmt_list
->viewFormatCount
== 0)
180 for (uint32_t i
= 0; i
< fmt_list
->viewFormatCount
; i
++) {
181 enum isl_format view_format
=
182 anv_get_isl_format(devinfo
, fmt_list
->pViewFormats
[i
],
183 VK_IMAGE_ASPECT_COLOR_BIT
, vk_info
->tiling
);
185 if (!isl_formats_are_ccs_e_compatible(devinfo
, format
, view_format
))
193 * For color images that have an auxiliary surface, request allocation for an
194 * additional buffer that mainly stores fast-clear values. Use of this buffer
195 * allows us to access the image's subresources while being aware of their
196 * fast-clear values in non-trivial cases (e.g., outside of a render pass in
197 * which a fast clear has occurred).
199 * In order to avoid having multiple clear colors for a single plane of an
200 * image (hence a single RENDER_SURFACE_STATE), we only allow fast-clears on
201 * the first slice (level 0, layer 0). At the time of our testing (Jan 17,
202 * 2018), there were no known applications which would benefit from fast-
203 * clearing more than just the first slice.
205 * The fast clear portion of the image is laid out in the following order:
207 * * 1 or 4 dwords (depending on hardware generation) for the clear color
208 * * 1 dword for the anv_fast_clear_type of the clear color
209 * * On gen9+, 1 dword per level and layer of the image (3D levels count
210 * multiple layers) in level-major order for compression state.
212 * For the purpose of discoverability, the algorithm used to manage
213 * compression and fast-clears is described here:
215 * * On a transition from UNDEFINED or PREINITIALIZED to a defined layout,
216 * all of the values in the fast clear portion of the image are initialized
219 * * On fast-clear, the clear value is written into surface state and also
220 * into the buffer and the fast clear type is set appropriately. Both
221 * setting the fast-clear value in the buffer and setting the fast-clear
222 * type happen from the GPU using MI commands.
224 * * Whenever a render or blorp operation is performed with CCS_E, we call
225 * genX(cmd_buffer_mark_image_written) to set the compression state to
226 * true (which is represented by UINT32_MAX).
228 * * On pipeline barrier transitions, the worst-case transition is computed
229 * from the image layouts. The command streamer inspects the fast clear
230 * type and compression state dwords and constructs a predicate. The
231 * worst-case resolve is performed with the given predicate and the fast
232 * clear and compression state is set accordingly.
234 * See anv_layout_to_aux_usage and anv_layout_to_fast_clear_type functions for
235 * details on exactly what is allowed in what layouts.
237 * On gen7-9, we do not have a concept of indirect clear colors in hardware.
238 * In order to deal with this, we have to do some clear color management.
240 * * For LOAD_OP_LOAD at the top of a renderpass, we have to copy the clear
241 * value from the buffer into the surface state with MI commands.
243 * * For any blorp operations, we pass the address to the clear value into
244 * blorp and it knows to copy the clear color.
247 add_aux_state_tracking_buffer(struct anv_image
*image
,
249 const struct anv_device
*device
)
251 assert(image
&& device
);
252 assert(image
->planes
[plane
].aux_surface
.isl
.size_B
> 0 &&
253 image
->aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
);
255 /* Compressed images must be tiled and therefore everything should be 4K
256 * aligned. The CCS has the same alignment requirements. This is good
257 * because we need at least dword-alignment for MI_LOAD/STORE operations.
259 assert(image
->alignment
% 4 == 0);
260 assert((image
->planes
[plane
].offset
+ image
->planes
[plane
].size
) % 4 == 0);
262 /* This buffer should be at the very end of the plane. */
263 if (image
->disjoint
) {
264 assert(image
->planes
[plane
].size
==
265 (image
->planes
[plane
].offset
+ image
->planes
[plane
].size
));
267 assert(image
->size
==
268 (image
->planes
[plane
].offset
+ image
->planes
[plane
].size
));
271 const unsigned clear_color_state_size
= device
->info
.gen
>= 10 ?
272 device
->isl_dev
.ss
.clear_color_state_size
:
273 device
->isl_dev
.ss
.clear_value_size
;
275 /* Clear color and fast clear type */
276 unsigned state_size
= clear_color_state_size
+ 4;
278 /* We only need to track compression on CCS_E surfaces. */
279 if (image
->planes
[plane
].aux_usage
== ISL_AUX_USAGE_CCS_E
) {
280 if (image
->type
== VK_IMAGE_TYPE_3D
) {
281 for (uint32_t l
= 0; l
< image
->levels
; l
++)
282 state_size
+= anv_minify(image
->extent
.depth
, l
) * 4;
284 state_size
+= image
->levels
* image
->array_size
* 4;
288 image
->planes
[plane
].fast_clear_state_offset
=
289 image
->planes
[plane
].offset
+ image
->planes
[plane
].size
;
291 image
->planes
[plane
].size
+= state_size
;
292 image
->size
+= state_size
;
296 * Initialize the anv_image::*_surface selected by \a aspect. Then update the
297 * image's memory requirements (that is, the image's size and alignment).
300 make_surface(const struct anv_device
*dev
,
301 struct anv_image
*image
,
302 const struct anv_image_create_info
*anv_info
,
303 isl_tiling_flags_t tiling_flags
,
304 VkImageAspectFlagBits aspect
)
306 const VkImageCreateInfo
*vk_info
= anv_info
->vk_info
;
309 static const enum isl_surf_dim vk_to_isl_surf_dim
[] = {
310 [VK_IMAGE_TYPE_1D
] = ISL_SURF_DIM_1D
,
311 [VK_IMAGE_TYPE_2D
] = ISL_SURF_DIM_2D
,
312 [VK_IMAGE_TYPE_3D
] = ISL_SURF_DIM_3D
,
315 image
->extent
= anv_sanitize_image_extent(vk_info
->imageType
,
318 const unsigned plane
= anv_image_aspect_to_plane(image
->aspects
, aspect
);
319 const struct anv_format_plane plane_format
=
320 anv_get_format_plane(&dev
->info
, image
->vk_format
, aspect
, image
->tiling
);
321 struct anv_surface
*anv_surf
= &image
->planes
[plane
].surface
;
323 const isl_surf_usage_flags_t usage
=
324 choose_isl_surf_usage(vk_info
->flags
, image
->usage
,
325 anv_info
->isl_extra_usage_flags
, aspect
);
327 /* If an image is created as BLOCK_TEXEL_VIEW_COMPATIBLE, then we need to
328 * fall back to linear on Broadwell and earlier because we aren't
329 * guaranteed that we can handle offsets correctly. On Sky Lake, the
330 * horizontal and vertical alignments are sufficiently high that we can
331 * just use RENDER_SURFACE_STATE::X/Y Offset.
333 bool needs_shadow
= false;
334 if (dev
->info
.gen
<= 8 &&
335 (vk_info
->flags
& VK_IMAGE_CREATE_BLOCK_TEXEL_VIEW_COMPATIBLE_BIT
) &&
336 vk_info
->tiling
== VK_IMAGE_TILING_OPTIMAL
) {
337 assert(isl_format_is_compressed(plane_format
.isl_format
));
338 tiling_flags
= ISL_TILING_LINEAR_BIT
;
342 ok
= isl_surf_init(&dev
->isl_dev
, &anv_surf
->isl
,
343 .dim
= vk_to_isl_surf_dim
[vk_info
->imageType
],
344 .format
= plane_format
.isl_format
,
345 .width
= image
->extent
.width
/ plane_format
.denominator_scales
[0],
346 .height
= image
->extent
.height
/ plane_format
.denominator_scales
[1],
347 .depth
= image
->extent
.depth
,
348 .levels
= vk_info
->mipLevels
,
349 .array_len
= vk_info
->arrayLayers
,
350 .samples
= vk_info
->samples
,
351 .min_alignment_B
= 0,
352 .row_pitch_B
= anv_info
->stride
,
354 .tiling_flags
= tiling_flags
);
357 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
359 image
->planes
[plane
].aux_usage
= ISL_AUX_USAGE_NONE
;
361 add_surface(image
, anv_surf
, plane
);
363 /* If an image is created as BLOCK_TEXEL_VIEW_COMPATIBLE, then we need to
364 * create an identical tiled shadow surface for use while texturing so we
365 * don't get garbage performance.
368 assert(aspect
== VK_IMAGE_ASPECT_COLOR_BIT
);
369 assert(tiling_flags
== ISL_TILING_LINEAR_BIT
);
371 ok
= isl_surf_init(&dev
->isl_dev
, &image
->planes
[plane
].shadow_surface
.isl
,
372 .dim
= vk_to_isl_surf_dim
[vk_info
->imageType
],
373 .format
= plane_format
.isl_format
,
374 .width
= image
->extent
.width
,
375 .height
= image
->extent
.height
,
376 .depth
= image
->extent
.depth
,
377 .levels
= vk_info
->mipLevels
,
378 .array_len
= vk_info
->arrayLayers
,
379 .samples
= vk_info
->samples
,
380 .min_alignment_B
= 0,
381 .row_pitch_B
= anv_info
->stride
,
383 .tiling_flags
= ISL_TILING_ANY_MASK
);
385 /* isl_surf_init() will fail only if provided invalid input. Invalid input
386 * is illegal in Vulkan.
390 add_surface(image
, &image
->planes
[plane
].shadow_surface
, plane
);
393 /* Add a HiZ surface to a depth buffer that will be used for rendering.
395 if (aspect
== VK_IMAGE_ASPECT_DEPTH_BIT
) {
396 /* We don't advertise that depth buffers could be used as storage
399 assert(!(image
->usage
& VK_IMAGE_USAGE_STORAGE_BIT
));
401 /* Allow the user to control HiZ enabling. Disable by default on gen7
402 * because resolves are not currently implemented pre-BDW.
404 if (!(image
->usage
& VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT
)) {
405 /* It will never be used as an attachment, HiZ is pointless. */
406 } else if (dev
->info
.gen
== 7) {
407 anv_perf_warn(dev
->instance
, image
, "Implement gen7 HiZ");
408 } else if (vk_info
->mipLevels
> 1) {
409 anv_perf_warn(dev
->instance
, image
, "Enable multi-LOD HiZ");
410 } else if (vk_info
->arrayLayers
> 1) {
411 anv_perf_warn(dev
->instance
, image
,
412 "Implement multi-arrayLayer HiZ clears and resolves");
413 } else if (dev
->info
.gen
== 8 && vk_info
->samples
> 1) {
414 anv_perf_warn(dev
->instance
, image
, "Enable gen8 multisampled HiZ");
415 } else if (!unlikely(INTEL_DEBUG
& DEBUG_NO_HIZ
)) {
416 assert(image
->planes
[plane
].aux_surface
.isl
.size_B
== 0);
417 ok
= isl_surf_get_hiz_surf(&dev
->isl_dev
,
418 &image
->planes
[plane
].surface
.isl
,
419 &image
->planes
[plane
].aux_surface
.isl
);
421 add_surface(image
, &image
->planes
[plane
].aux_surface
, plane
);
422 image
->planes
[plane
].aux_usage
= ISL_AUX_USAGE_HIZ
;
424 } else if ((aspect
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) && vk_info
->samples
== 1) {
425 /* TODO: Disallow compression with :
427 * 1) non multiplanar images (We appear to hit a sampler bug with
428 * CCS & R16G16 format. Putting the clear state a page/4096bytes
429 * further fixes the issue).
431 * 2) alias images, because they might be aliases of images
434 * 3) compression disabled by debug
436 const bool allow_compression
=
437 image
->n_planes
== 1 &&
438 (vk_info
->flags
& VK_IMAGE_CREATE_ALIAS_BIT
) == 0 &&
439 likely((INTEL_DEBUG
& DEBUG_NO_RBC
) == 0);
441 if (allow_compression
) {
442 assert(image
->planes
[plane
].aux_surface
.isl
.size_B
== 0);
443 ok
= isl_surf_get_ccs_surf(&dev
->isl_dev
,
444 &image
->planes
[plane
].surface
.isl
,
445 &image
->planes
[plane
].aux_surface
.isl
, 0);
448 /* Disable CCS when it is not useful (i.e., when you can't render
449 * to the image with CCS enabled).
451 if (!isl_format_supports_rendering(&dev
->info
,
452 plane_format
.isl_format
)) {
453 /* While it may be technically possible to enable CCS for this
454 * image, we currently don't have things hooked up to get it
457 anv_perf_warn(dev
->instance
, image
,
458 "This image format doesn't support rendering. "
459 "Not allocating an CCS buffer.");
460 image
->planes
[plane
].aux_surface
.isl
.size_B
= 0;
464 add_surface(image
, &image
->planes
[plane
].aux_surface
, plane
);
465 add_aux_state_tracking_buffer(image
, plane
, dev
);
467 /* For images created without MUTABLE_FORMAT_BIT set, we know that
468 * they will always be used with the original format. In
469 * particular, they will always be used with a format that
470 * supports color compression. If it's never used as a storage
471 * image, then it will only be used through the sampler or the as
472 * a render target. This means that it's safe to just leave
473 * compression on at all times for these formats.
475 if (!(vk_info
->usage
& VK_IMAGE_USAGE_STORAGE_BIT
) &&
476 all_formats_ccs_e_compatible(&dev
->info
, vk_info
)) {
477 image
->planes
[plane
].aux_usage
= ISL_AUX_USAGE_CCS_E
;
481 } else if ((aspect
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) && vk_info
->samples
> 1) {
482 assert(!(vk_info
->usage
& VK_IMAGE_USAGE_STORAGE_BIT
));
483 assert(image
->planes
[plane
].aux_surface
.isl
.size_B
== 0);
484 ok
= isl_surf_get_mcs_surf(&dev
->isl_dev
,
485 &image
->planes
[plane
].surface
.isl
,
486 &image
->planes
[plane
].aux_surface
.isl
);
488 add_surface(image
, &image
->planes
[plane
].aux_surface
, plane
);
489 add_aux_state_tracking_buffer(image
, plane
, dev
);
490 image
->planes
[plane
].aux_usage
= ISL_AUX_USAGE_MCS
;
494 assert((image
->planes
[plane
].offset
+ image
->planes
[plane
].size
) == image
->size
);
496 /* Upper bound of the last surface should be smaller than the plane's
499 assert((MAX2(image
->planes
[plane
].surface
.offset
,
500 image
->planes
[plane
].aux_surface
.offset
) +
501 (image
->planes
[plane
].aux_surface
.isl
.size_B
> 0 ?
502 image
->planes
[plane
].aux_surface
.isl
.size_B
:
503 image
->planes
[plane
].surface
.isl
.size_B
)) <=
504 (image
->planes
[plane
].offset
+ image
->planes
[plane
].size
));
506 if (image
->planes
[plane
].aux_surface
.isl
.size_B
) {
507 /* assert(image->planes[plane].fast_clear_state_offset == */
508 /* (image->planes[plane].aux_surface.offset + image->planes[plane].aux_surface.isl.size_B)); */
509 assert(image
->planes
[plane
].fast_clear_state_offset
<
510 (image
->planes
[plane
].offset
+ image
->planes
[plane
].size
));
517 score_drm_format_mod(uint64_t modifier
)
520 case DRM_FORMAT_MOD_LINEAR
: return 1;
521 case I915_FORMAT_MOD_X_TILED
: return 2;
522 case I915_FORMAT_MOD_Y_TILED
: return 3;
523 case I915_FORMAT_MOD_Y_TILED_CCS
: return 4;
524 default: unreachable("bad DRM format modifier");
528 static const struct isl_drm_modifier_info
*
529 choose_drm_format_mod(const struct anv_physical_device
*device
,
530 uint32_t modifier_count
, const uint64_t *modifiers
)
532 uint64_t best_mod
= UINT64_MAX
;
533 uint32_t best_score
= 0;
535 for (uint32_t i
= 0; i
< modifier_count
; ++i
) {
536 uint32_t score
= score_drm_format_mod(modifiers
[i
]);
537 if (score
> best_score
) {
538 best_mod
= modifiers
[i
];
544 return isl_drm_modifier_get_info(best_mod
);
550 anv_image_create(VkDevice _device
,
551 const struct anv_image_create_info
*create_info
,
552 const VkAllocationCallbacks
* alloc
,
555 ANV_FROM_HANDLE(anv_device
, device
, _device
);
556 const VkImageCreateInfo
*pCreateInfo
= create_info
->vk_info
;
557 const struct isl_drm_modifier_info
*isl_mod_info
= NULL
;
558 struct anv_image
*image
= NULL
;
561 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO
);
563 const struct wsi_image_create_info
*wsi_info
=
564 vk_find_struct_const(pCreateInfo
->pNext
, WSI_IMAGE_CREATE_INFO_MESA
);
565 if (wsi_info
&& wsi_info
->modifier_count
> 0) {
566 isl_mod_info
= choose_drm_format_mod(&device
->instance
->physicalDevice
,
567 wsi_info
->modifier_count
,
568 wsi_info
->modifiers
);
569 assert(isl_mod_info
);
572 anv_assert(pCreateInfo
->mipLevels
> 0);
573 anv_assert(pCreateInfo
->arrayLayers
> 0);
574 anv_assert(pCreateInfo
->samples
> 0);
575 anv_assert(pCreateInfo
->extent
.width
> 0);
576 anv_assert(pCreateInfo
->extent
.height
> 0);
577 anv_assert(pCreateInfo
->extent
.depth
> 0);
579 image
= vk_zalloc2(&device
->alloc
, alloc
, sizeof(*image
), 8,
580 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
582 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
584 image
->type
= pCreateInfo
->imageType
;
585 image
->extent
= pCreateInfo
->extent
;
586 image
->vk_format
= pCreateInfo
->format
;
587 image
->format
= anv_get_format(pCreateInfo
->format
);
588 image
->aspects
= vk_format_aspects(image
->vk_format
);
589 image
->levels
= pCreateInfo
->mipLevels
;
590 image
->array_size
= pCreateInfo
->arrayLayers
;
591 image
->samples
= pCreateInfo
->samples
;
592 image
->usage
= pCreateInfo
->usage
;
593 image
->create_flags
= pCreateInfo
->flags
;
594 image
->tiling
= pCreateInfo
->tiling
;
595 image
->disjoint
= pCreateInfo
->flags
& VK_IMAGE_CREATE_DISJOINT_BIT
;
596 image
->needs_set_tiling
= wsi_info
&& wsi_info
->scanout
;
597 image
->drm_format_mod
= isl_mod_info
? isl_mod_info
->modifier
:
598 DRM_FORMAT_MOD_INVALID
;
600 const struct anv_format
*format
= anv_get_format(image
->vk_format
);
601 assert(format
!= NULL
);
603 const isl_tiling_flags_t isl_tiling_flags
=
604 choose_isl_tiling_flags(create_info
, isl_mod_info
,
605 image
->needs_set_tiling
);
607 image
->n_planes
= format
->n_planes
;
610 for_each_bit(b
, image
->aspects
) {
611 r
= make_surface(device
, image
, create_info
, isl_tiling_flags
,
617 *pImage
= anv_image_to_handle(image
);
623 vk_free2(&device
->alloc
, alloc
, image
);
629 anv_CreateImage(VkDevice device
,
630 const VkImageCreateInfo
*pCreateInfo
,
631 const VkAllocationCallbacks
*pAllocator
,
634 const VkNativeBufferANDROID
*gralloc_info
=
635 vk_find_struct_const(pCreateInfo
->pNext
, NATIVE_BUFFER_ANDROID
);
638 return anv_image_from_gralloc(device
, pCreateInfo
, gralloc_info
,
641 return anv_image_create(device
,
642 &(struct anv_image_create_info
) {
643 .vk_info
= pCreateInfo
,
650 anv_DestroyImage(VkDevice _device
, VkImage _image
,
651 const VkAllocationCallbacks
*pAllocator
)
653 ANV_FROM_HANDLE(anv_device
, device
, _device
);
654 ANV_FROM_HANDLE(anv_image
, image
, _image
);
659 for (uint32_t p
= 0; p
< image
->n_planes
; ++p
) {
660 if (image
->planes
[p
].bo_is_owned
) {
661 assert(image
->planes
[p
].address
.bo
!= NULL
);
662 anv_bo_cache_release(device
, &device
->bo_cache
,
663 image
->planes
[p
].address
.bo
);
667 vk_free2(&device
->alloc
, pAllocator
, image
);
670 static void anv_image_bind_memory_plane(struct anv_device
*device
,
671 struct anv_image
*image
,
673 struct anv_device_memory
*memory
,
674 uint32_t memory_offset
)
676 assert(!image
->planes
[plane
].bo_is_owned
);
679 image
->planes
[plane
].address
= ANV_NULL_ADDRESS
;
683 image
->planes
[plane
].address
= (struct anv_address
) {
685 .offset
= memory_offset
,
689 VkResult
anv_BindImageMemory(
692 VkDeviceMemory _memory
,
693 VkDeviceSize memoryOffset
)
695 ANV_FROM_HANDLE(anv_device
, device
, _device
);
696 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
697 ANV_FROM_HANDLE(anv_image
, image
, _image
);
700 anv_foreach_image_aspect_bit(aspect_bit
, image
, image
->aspects
) {
702 anv_image_aspect_to_plane(image
->aspects
, 1UL << aspect_bit
);
703 anv_image_bind_memory_plane(device
, image
, plane
, mem
, memoryOffset
);
709 VkResult
anv_BindImageMemory2(
711 uint32_t bindInfoCount
,
712 const VkBindImageMemoryInfo
* pBindInfos
)
714 ANV_FROM_HANDLE(anv_device
, device
, _device
);
716 for (uint32_t i
= 0; i
< bindInfoCount
; i
++) {
717 const VkBindImageMemoryInfo
*bind_info
= &pBindInfos
[i
];
718 ANV_FROM_HANDLE(anv_device_memory
, mem
, bind_info
->memory
);
719 ANV_FROM_HANDLE(anv_image
, image
, bind_info
->image
);
720 VkImageAspectFlags aspects
= image
->aspects
;
722 vk_foreach_struct_const(s
, bind_info
->pNext
) {
724 case VK_STRUCTURE_TYPE_BIND_IMAGE_PLANE_MEMORY_INFO
: {
725 const VkBindImagePlaneMemoryInfo
*plane_info
=
726 (const VkBindImagePlaneMemoryInfo
*) s
;
728 aspects
= plane_info
->planeAspect
;
732 anv_debug_ignored_stype(s
->sType
);
738 anv_foreach_image_aspect_bit(aspect_bit
, image
, aspects
) {
740 anv_image_aspect_to_plane(image
->aspects
, 1UL << aspect_bit
);
741 anv_image_bind_memory_plane(device
, image
, plane
,
742 mem
, bind_info
->memoryOffset
);
749 void anv_GetImageSubresourceLayout(
752 const VkImageSubresource
* subresource
,
753 VkSubresourceLayout
* layout
)
755 ANV_FROM_HANDLE(anv_image
, image
, _image
);
757 const struct anv_surface
*surface
;
758 if (subresource
->aspectMask
== VK_IMAGE_ASPECT_PLANE_1_BIT_KHR
&&
759 image
->drm_format_mod
!= DRM_FORMAT_MOD_INVALID
&&
760 isl_drm_modifier_has_aux(image
->drm_format_mod
))
761 surface
= &image
->planes
[0].aux_surface
;
763 surface
= get_surface(image
, subresource
->aspectMask
);
765 assert(__builtin_popcount(subresource
->aspectMask
) == 1);
767 layout
->offset
= surface
->offset
;
768 layout
->rowPitch
= surface
->isl
.row_pitch_B
;
769 layout
->depthPitch
= isl_surf_get_array_pitch(&surface
->isl
);
770 layout
->arrayPitch
= isl_surf_get_array_pitch(&surface
->isl
);
772 if (subresource
->mipLevel
> 0 || subresource
->arrayLayer
> 0) {
773 assert(surface
->isl
.tiling
== ISL_TILING_LINEAR
);
776 isl_surf_get_image_offset_B_tile_sa(&surface
->isl
,
777 subresource
->mipLevel
,
778 subresource
->arrayLayer
,
779 0 /* logical_z_offset_px */,
780 &offset_B
, NULL
, NULL
);
781 layout
->offset
+= offset_B
;
782 layout
->size
= layout
->rowPitch
* anv_minify(image
->extent
.height
,
783 subresource
->mipLevel
);
785 layout
->size
= surface
->isl
.size_B
;
790 * This function determines the optimal buffer to use for a given
791 * VkImageLayout and other pieces of information needed to make that
792 * determination. This does not determine the optimal buffer to use
793 * during a resolve operation.
795 * @param devinfo The device information of the Intel GPU.
796 * @param image The image that may contain a collection of buffers.
797 * @param aspect The aspect of the image to be accessed.
798 * @param layout The current layout of the image aspect(s).
800 * @return The primary buffer that should be used for the given layout.
803 anv_layout_to_aux_usage(const struct gen_device_info
* const devinfo
,
804 const struct anv_image
* const image
,
805 const VkImageAspectFlagBits aspect
,
806 const VkImageLayout layout
)
808 /* Validate the inputs. */
810 /* The devinfo is needed as the optimal buffer varies across generations. */
811 assert(devinfo
!= NULL
);
813 /* The layout of a NULL image is not properly defined. */
814 assert(image
!= NULL
);
816 /* The aspect must be exactly one of the image aspects. */
817 assert(util_bitcount(aspect
) == 1 && (aspect
& image
->aspects
));
819 /* Determine the optimal buffer. */
821 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
, aspect
);
823 /* If there is no auxiliary surface allocated, we must use the one and only
826 if (image
->planes
[plane
].aux_surface
.isl
.size_B
== 0)
827 return ISL_AUX_USAGE_NONE
;
829 /* All images that use an auxiliary surface are required to be tiled. */
830 assert(image
->tiling
== VK_IMAGE_TILING_OPTIMAL
);
832 /* Stencil has no aux */
833 assert(aspect
!= VK_IMAGE_ASPECT_STENCIL_BIT
);
837 /* Invalid Layouts */
838 case VK_IMAGE_LAYOUT_RANGE_SIZE
:
839 case VK_IMAGE_LAYOUT_MAX_ENUM
:
840 unreachable("Invalid image layout.");
844 * The pre-initialized layout is equivalent to the undefined layout for
845 * optimally-tiled images. We can only do color compression (CCS or HiZ)
848 case VK_IMAGE_LAYOUT_UNDEFINED
:
849 case VK_IMAGE_LAYOUT_PREINITIALIZED
:
850 return ISL_AUX_USAGE_NONE
;
855 case VK_IMAGE_LAYOUT_GENERAL
:
856 case VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL
:
857 case VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL
:
858 if (aspect
== VK_IMAGE_ASPECT_DEPTH_BIT
) {
859 /* This buffer could be a depth buffer used in a transfer operation.
860 * BLORP currently doesn't use HiZ for transfer operations so we must
861 * use the main buffer for this layout. TODO: Enable HiZ in BLORP.
863 assert(image
->planes
[plane
].aux_usage
== ISL_AUX_USAGE_HIZ
);
864 return ISL_AUX_USAGE_NONE
;
866 assert(image
->aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
);
867 return image
->planes
[plane
].aux_usage
;
871 /* Sampling Layouts */
872 case VK_IMAGE_LAYOUT_DEPTH_STENCIL_READ_ONLY_OPTIMAL
:
873 case VK_IMAGE_LAYOUT_DEPTH_READ_ONLY_STENCIL_ATTACHMENT_OPTIMAL
:
874 assert((image
->aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) == 0);
876 case VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL
:
877 if (aspect
== VK_IMAGE_ASPECT_DEPTH_BIT
) {
878 if (anv_can_sample_with_hiz(devinfo
, image
))
879 return ISL_AUX_USAGE_HIZ
;
881 return ISL_AUX_USAGE_NONE
;
883 return image
->planes
[plane
].aux_usage
;
887 case VK_IMAGE_LAYOUT_PRESENT_SRC_KHR
: {
888 assert(image
->aspects
== VK_IMAGE_ASPECT_COLOR_BIT
);
890 /* When handing the image off to the presentation engine, we need to
891 * ensure that things are properly resolved. For images with no
892 * modifier, we assume that they follow the old rules and always need
893 * a full resolve because the PE doesn't understand any form of
894 * compression. For images with modifiers, we use the aux usage from
897 const struct isl_drm_modifier_info
*mod_info
=
898 isl_drm_modifier_get_info(image
->drm_format_mod
);
899 return mod_info
? mod_info
->aux_usage
: ISL_AUX_USAGE_NONE
;
903 /* Rendering Layouts */
904 case VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL
:
905 assert(aspect
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
);
906 if (image
->planes
[plane
].aux_usage
== ISL_AUX_USAGE_NONE
) {
907 assert(image
->samples
== 1);
908 return ISL_AUX_USAGE_CCS_D
;
910 assert(image
->planes
[plane
].aux_usage
!= ISL_AUX_USAGE_CCS_D
);
911 return image
->planes
[plane
].aux_usage
;
914 case VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL
:
915 case VK_IMAGE_LAYOUT_DEPTH_ATTACHMENT_STENCIL_READ_ONLY_OPTIMAL
:
916 assert(aspect
== VK_IMAGE_ASPECT_DEPTH_BIT
);
917 return ISL_AUX_USAGE_HIZ
;
919 case VK_IMAGE_LAYOUT_SHARED_PRESENT_KHR
:
920 unreachable("VK_KHR_shared_presentable_image is unsupported");
922 case VK_IMAGE_LAYOUT_FRAGMENT_DENSITY_MAP_OPTIMAL_EXT
:
923 unreachable("VK_EXT_fragment_density_map is unsupported");
925 case VK_IMAGE_LAYOUT_SHADING_RATE_OPTIMAL_NV
:
926 unreachable("VK_NV_shading_rate_image is unsupported");
929 /* If the layout isn't recognized in the exhaustive switch above, the
930 * VkImageLayout value is not defined in vulkan.h.
932 unreachable("layout is not a VkImageLayout enumeration member.");
936 * This function returns the level of unresolved fast-clear support of the
937 * given image in the given VkImageLayout.
939 * @param devinfo The device information of the Intel GPU.
940 * @param image The image that may contain a collection of buffers.
941 * @param aspect The aspect of the image to be accessed.
942 * @param layout The current layout of the image aspect(s).
944 enum anv_fast_clear_type
945 anv_layout_to_fast_clear_type(const struct gen_device_info
* const devinfo
,
946 const struct anv_image
* const image
,
947 const VkImageAspectFlagBits aspect
,
948 const VkImageLayout layout
)
950 /* The aspect must be exactly one of the image aspects. */
951 assert(util_bitcount(aspect
) == 1 && (aspect
& image
->aspects
));
953 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
, aspect
);
955 /* If there is no auxiliary surface allocated, there are no fast-clears */
956 if (image
->planes
[plane
].aux_surface
.isl
.size_B
== 0)
957 return ANV_FAST_CLEAR_NONE
;
959 /* All images that use an auxiliary surface are required to be tiled. */
960 assert(image
->tiling
== VK_IMAGE_TILING_OPTIMAL
);
962 /* Stencil has no aux */
963 assert(aspect
!= VK_IMAGE_ASPECT_STENCIL_BIT
);
965 if (aspect
== VK_IMAGE_ASPECT_DEPTH_BIT
) {
966 /* For depth images (with HiZ), the layout supports fast-clears if and
967 * only if it supports HiZ. However, we only support fast-clears to the
968 * default depth value.
970 enum isl_aux_usage aux_usage
=
971 anv_layout_to_aux_usage(devinfo
, image
, aspect
, layout
);
972 return aux_usage
== ISL_AUX_USAGE_HIZ
?
973 ANV_FAST_CLEAR_DEFAULT_VALUE
: ANV_FAST_CLEAR_NONE
;
976 assert(image
->aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
);
978 /* We don't support MSAA fast-clears on Ivybridge or Bay Trail because they
979 * lack the MI ALU which we need to determine the predicates.
981 if (devinfo
->gen
== 7 && !devinfo
->is_haswell
&& image
->samples
> 1)
982 return ANV_FAST_CLEAR_NONE
;
985 case VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL
:
986 return ANV_FAST_CLEAR_ANY
;
988 case VK_IMAGE_LAYOUT_PRESENT_SRC_KHR
: {
989 assert(image
->aspects
== VK_IMAGE_ASPECT_COLOR_BIT
);
991 /* We do not yet support any modifiers which support clear color so we
992 * just always return NONE. One day, this will change.
994 const struct isl_drm_modifier_info
*mod_info
=
995 isl_drm_modifier_get_info(image
->drm_format_mod
);
996 assert(!mod_info
|| !mod_info
->supports_clear_color
);
998 return ANV_FAST_CLEAR_NONE
;
1002 /* If the image has MCS or CCS_E enabled all the time then we can use
1003 * fast-clear as long as the clear color is the default value of zero
1004 * since this is the default value we program into every surface state
1005 * used for texturing.
1007 if (image
->planes
[plane
].aux_usage
== ISL_AUX_USAGE_MCS
||
1008 image
->planes
[plane
].aux_usage
== ISL_AUX_USAGE_CCS_E
)
1009 return ANV_FAST_CLEAR_DEFAULT_VALUE
;
1011 return ANV_FAST_CLEAR_NONE
;
1016 static struct anv_state
1017 alloc_surface_state(struct anv_device
*device
)
1019 return anv_state_pool_alloc(&device
->surface_state_pool
, 64, 64);
1022 static enum isl_channel_select
1023 remap_swizzle(VkComponentSwizzle swizzle
, VkComponentSwizzle component
,
1024 struct isl_swizzle format_swizzle
)
1026 if (swizzle
== VK_COMPONENT_SWIZZLE_IDENTITY
)
1027 swizzle
= component
;
1030 case VK_COMPONENT_SWIZZLE_ZERO
: return ISL_CHANNEL_SELECT_ZERO
;
1031 case VK_COMPONENT_SWIZZLE_ONE
: return ISL_CHANNEL_SELECT_ONE
;
1032 case VK_COMPONENT_SWIZZLE_R
: return format_swizzle
.r
;
1033 case VK_COMPONENT_SWIZZLE_G
: return format_swizzle
.g
;
1034 case VK_COMPONENT_SWIZZLE_B
: return format_swizzle
.b
;
1035 case VK_COMPONENT_SWIZZLE_A
: return format_swizzle
.a
;
1037 unreachable("Invalid swizzle");
1042 anv_image_fill_surface_state(struct anv_device
*device
,
1043 const struct anv_image
*image
,
1044 VkImageAspectFlagBits aspect
,
1045 const struct isl_view
*view_in
,
1046 isl_surf_usage_flags_t view_usage
,
1047 enum isl_aux_usage aux_usage
,
1048 const union isl_color_value
*clear_color
,
1049 enum anv_image_view_state_flags flags
,
1050 struct anv_surface_state
*state_inout
,
1051 struct brw_image_param
*image_param_out
)
1053 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
, aspect
);
1055 const struct anv_surface
*surface
= &image
->planes
[plane
].surface
,
1056 *aux_surface
= &image
->planes
[plane
].aux_surface
;
1058 struct isl_view view
= *view_in
;
1059 view
.usage
|= view_usage
;
1061 /* For texturing with VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL from a
1062 * compressed surface with a shadow surface, we use the shadow instead of
1063 * the primary surface. The shadow surface will be tiled, unlike the main
1064 * surface, so it should get significantly better performance.
1066 if (image
->planes
[plane
].shadow_surface
.isl
.size_B
> 0 &&
1067 isl_format_is_compressed(view
.format
) &&
1068 (flags
& ANV_IMAGE_VIEW_STATE_TEXTURE_OPTIMAL
)) {
1069 assert(isl_format_is_compressed(surface
->isl
.format
));
1070 assert(surface
->isl
.tiling
== ISL_TILING_LINEAR
);
1071 assert(image
->planes
[plane
].shadow_surface
.isl
.tiling
!= ISL_TILING_LINEAR
);
1072 surface
= &image
->planes
[plane
].shadow_surface
;
1075 if (view_usage
== ISL_SURF_USAGE_RENDER_TARGET_BIT
)
1076 view
.swizzle
= anv_swizzle_for_render(view
.swizzle
);
1078 /* If this is a HiZ buffer we can sample from with a programmable clear
1079 * value (SKL+), define the clear value to the optimal constant.
1081 union isl_color_value default_clear_color
= { .u32
= { 0, } };
1082 if (device
->info
.gen
>= 9 && aux_usage
== ISL_AUX_USAGE_HIZ
)
1083 default_clear_color
.f32
[0] = ANV_HZ_FC_VAL
;
1085 clear_color
= &default_clear_color
;
1087 const struct anv_address address
=
1088 anv_address_add(image
->planes
[plane
].address
, surface
->offset
);
1090 if (view_usage
== ISL_SURF_USAGE_STORAGE_BIT
&&
1091 !(flags
& ANV_IMAGE_VIEW_STATE_STORAGE_WRITE_ONLY
) &&
1092 !isl_has_matching_typed_storage_image_format(&device
->info
,
1094 /* In this case, we are a writeable storage buffer which needs to be
1095 * lowered to linear. All tiling and offset calculations will be done in
1098 assert(aux_usage
== ISL_AUX_USAGE_NONE
);
1099 isl_buffer_fill_state(&device
->isl_dev
, state_inout
->state
.map
,
1100 .address
= anv_address_physical(address
),
1101 .size_B
= surface
->isl
.size_B
,
1102 .format
= ISL_FORMAT_RAW
,
1104 .mocs
= anv_mocs_for_bo(device
, address
.bo
));
1105 state_inout
->address
= address
,
1106 state_inout
->aux_address
= ANV_NULL_ADDRESS
;
1107 state_inout
->clear_address
= ANV_NULL_ADDRESS
;
1109 if (view_usage
== ISL_SURF_USAGE_STORAGE_BIT
&&
1110 !(flags
& ANV_IMAGE_VIEW_STATE_STORAGE_WRITE_ONLY
)) {
1111 /* Typed surface reads support a very limited subset of the shader
1112 * image formats. Translate it into the closest format the hardware
1115 assert(aux_usage
== ISL_AUX_USAGE_NONE
);
1116 view
.format
= isl_lower_storage_image_format(&device
->info
,
1120 const struct isl_surf
*isl_surf
= &surface
->isl
;
1122 struct isl_surf tmp_surf
;
1123 uint32_t offset_B
= 0, tile_x_sa
= 0, tile_y_sa
= 0;
1124 if (isl_format_is_compressed(surface
->isl
.format
) &&
1125 !isl_format_is_compressed(view
.format
)) {
1126 /* We're creating an uncompressed view of a compressed surface. This
1127 * is allowed but only for a single level/layer.
1129 assert(surface
->isl
.samples
== 1);
1130 assert(view
.levels
== 1);
1131 assert(view
.array_len
== 1);
1133 isl_surf_get_image_surf(&device
->isl_dev
, isl_surf
,
1135 surface
->isl
.dim
== ISL_SURF_DIM_3D
?
1136 0 : view
.base_array_layer
,
1137 surface
->isl
.dim
== ISL_SURF_DIM_3D
?
1138 view
.base_array_layer
: 0,
1140 &offset_B
, &tile_x_sa
, &tile_y_sa
);
1142 /* The newly created image represents the one subimage we're
1143 * referencing with this view so it only has one array slice and
1146 view
.base_array_layer
= 0;
1147 view
.base_level
= 0;
1149 /* We're making an uncompressed view here. The image dimensions need
1150 * to be scaled down by the block size.
1152 const struct isl_format_layout
*fmtl
=
1153 isl_format_get_layout(surface
->isl
.format
);
1154 tmp_surf
.format
= view
.format
;
1155 tmp_surf
.logical_level0_px
.width
=
1156 DIV_ROUND_UP(tmp_surf
.logical_level0_px
.width
, fmtl
->bw
);
1157 tmp_surf
.logical_level0_px
.height
=
1158 DIV_ROUND_UP(tmp_surf
.logical_level0_px
.height
, fmtl
->bh
);
1159 tmp_surf
.phys_level0_sa
.width
/= fmtl
->bw
;
1160 tmp_surf
.phys_level0_sa
.height
/= fmtl
->bh
;
1161 tile_x_sa
/= fmtl
->bw
;
1162 tile_y_sa
/= fmtl
->bh
;
1164 isl_surf
= &tmp_surf
;
1166 if (device
->info
.gen
<= 8) {
1167 assert(surface
->isl
.tiling
== ISL_TILING_LINEAR
);
1168 assert(tile_x_sa
== 0);
1169 assert(tile_y_sa
== 0);
1173 state_inout
->address
= anv_address_add(address
, offset_B
);
1175 struct anv_address aux_address
= ANV_NULL_ADDRESS
;
1176 if (aux_usage
!= ISL_AUX_USAGE_NONE
) {
1177 aux_address
= anv_address_add(image
->planes
[plane
].address
,
1178 aux_surface
->offset
);
1180 state_inout
->aux_address
= aux_address
;
1182 struct anv_address clear_address
= ANV_NULL_ADDRESS
;
1183 if (device
->info
.gen
>= 10 && aux_usage
!= ISL_AUX_USAGE_NONE
) {
1184 if (aux_usage
== ISL_AUX_USAGE_HIZ
) {
1185 clear_address
= (struct anv_address
) {
1186 .bo
= &device
->hiz_clear_bo
,
1190 clear_address
= anv_image_get_clear_color_addr(device
, image
, aspect
);
1193 state_inout
->clear_address
= clear_address
;
1195 isl_surf_fill_state(&device
->isl_dev
, state_inout
->state
.map
,
1198 .address
= anv_address_physical(state_inout
->address
),
1199 .clear_color
= *clear_color
,
1200 .aux_surf
= &aux_surface
->isl
,
1201 .aux_usage
= aux_usage
,
1202 .aux_address
= anv_address_physical(aux_address
),
1203 .clear_address
= anv_address_physical(clear_address
),
1204 .use_clear_address
= !anv_address_is_null(clear_address
),
1205 .mocs
= anv_mocs_for_bo(device
,
1206 state_inout
->address
.bo
),
1207 .x_offset_sa
= tile_x_sa
,
1208 .y_offset_sa
= tile_y_sa
);
1210 /* With the exception of gen8, the bottom 12 bits of the MCS base address
1211 * are used to store other information. This should be ok, however,
1212 * because the surface buffer addresses are always 4K page aligned.
1214 uint32_t *aux_addr_dw
= state_inout
->state
.map
+
1215 device
->isl_dev
.ss
.aux_addr_offset
;
1216 assert((aux_address
.offset
& 0xfff) == 0);
1217 state_inout
->aux_address
.offset
|= *aux_addr_dw
& 0xfff;
1219 if (device
->info
.gen
>= 10 && clear_address
.bo
) {
1220 uint32_t *clear_addr_dw
= state_inout
->state
.map
+
1221 device
->isl_dev
.ss
.clear_color_state_offset
;
1222 assert((clear_address
.offset
& 0x3f) == 0);
1223 state_inout
->clear_address
.offset
|= *clear_addr_dw
& 0x3f;
1227 anv_state_flush(device
, state_inout
->state
);
1229 if (image_param_out
) {
1230 assert(view_usage
== ISL_SURF_USAGE_STORAGE_BIT
);
1231 isl_surf_fill_image_param(&device
->isl_dev
, image_param_out
,
1232 &surface
->isl
, &view
);
1236 static VkImageAspectFlags
1237 remap_aspect_flags(VkImageAspectFlags view_aspects
)
1239 if (view_aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) {
1240 if (util_bitcount(view_aspects
) == 1)
1241 return VK_IMAGE_ASPECT_COLOR_BIT
;
1243 VkImageAspectFlags color_aspects
= 0;
1244 for (uint32_t i
= 0; i
< util_bitcount(view_aspects
); i
++)
1245 color_aspects
|= VK_IMAGE_ASPECT_PLANE_0_BIT
<< i
;
1246 return color_aspects
;
1248 /* No special remapping needed for depth & stencil aspects. */
1249 return view_aspects
;
1253 anv_image_aspect_get_planes(VkImageAspectFlags aspect_mask
)
1255 uint32_t planes
= 0;
1257 if (aspect_mask
& (VK_IMAGE_ASPECT_COLOR_BIT
|
1258 VK_IMAGE_ASPECT_DEPTH_BIT
|
1259 VK_IMAGE_ASPECT_STENCIL_BIT
|
1260 VK_IMAGE_ASPECT_PLANE_0_BIT
))
1262 if (aspect_mask
& VK_IMAGE_ASPECT_PLANE_1_BIT
)
1264 if (aspect_mask
& VK_IMAGE_ASPECT_PLANE_2_BIT
)
1267 if ((aspect_mask
& VK_IMAGE_ASPECT_DEPTH_BIT
) != 0 &&
1268 (aspect_mask
& VK_IMAGE_ASPECT_STENCIL_BIT
) != 0)
1275 anv_CreateImageView(VkDevice _device
,
1276 const VkImageViewCreateInfo
*pCreateInfo
,
1277 const VkAllocationCallbacks
*pAllocator
,
1280 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1281 ANV_FROM_HANDLE(anv_image
, image
, pCreateInfo
->image
);
1282 struct anv_image_view
*iview
;
1284 iview
= vk_zalloc2(&device
->alloc
, pAllocator
, sizeof(*iview
), 8,
1285 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1287 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1289 const VkImageSubresourceRange
*range
= &pCreateInfo
->subresourceRange
;
1291 assert(range
->layerCount
> 0);
1292 assert(range
->baseMipLevel
< image
->levels
);
1294 const VkImageViewUsageCreateInfo
*usage_info
=
1295 vk_find_struct_const(pCreateInfo
, IMAGE_VIEW_USAGE_CREATE_INFO
);
1296 VkImageUsageFlags view_usage
= usage_info
? usage_info
->usage
: image
->usage
;
1297 /* View usage should be a subset of image usage */
1298 assert((view_usage
& ~image
->usage
) == 0);
1299 assert(view_usage
& (VK_IMAGE_USAGE_SAMPLED_BIT
|
1300 VK_IMAGE_USAGE_STORAGE_BIT
|
1301 VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT
|
1302 VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT
|
1303 VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT
));
1305 switch (image
->type
) {
1307 unreachable("bad VkImageType");
1308 case VK_IMAGE_TYPE_1D
:
1309 case VK_IMAGE_TYPE_2D
:
1310 assert(range
->baseArrayLayer
+ anv_get_layerCount(image
, range
) - 1 <= image
->array_size
);
1312 case VK_IMAGE_TYPE_3D
:
1313 assert(range
->baseArrayLayer
+ anv_get_layerCount(image
, range
) - 1
1314 <= anv_minify(image
->extent
.depth
, range
->baseMipLevel
));
1318 /* First expand aspects to the image's ones (for example
1319 * VK_IMAGE_ASPECT_COLOR_BIT will be converted to
1320 * VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT |
1321 * VK_IMAGE_ASPECT_PLANE_2_BIT for an image of format
1322 * VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM_KHR.
1324 VkImageAspectFlags expanded_aspects
=
1325 anv_image_expand_aspects(image
, range
->aspectMask
);
1327 iview
->image
= image
;
1329 /* Remap the expanded aspects for the image view. For example if only
1330 * VK_IMAGE_ASPECT_PLANE_1_BIT was given in range->aspectMask, we will
1331 * convert it to VK_IMAGE_ASPECT_COLOR_BIT since from the point of view of
1332 * the image view, it only has a single plane.
1334 iview
->aspect_mask
= remap_aspect_flags(expanded_aspects
);
1335 iview
->n_planes
= anv_image_aspect_get_planes(iview
->aspect_mask
);
1336 iview
->vk_format
= pCreateInfo
->format
;
1338 iview
->extent
= (VkExtent3D
) {
1339 .width
= anv_minify(image
->extent
.width
, range
->baseMipLevel
),
1340 .height
= anv_minify(image
->extent
.height
, range
->baseMipLevel
),
1341 .depth
= anv_minify(image
->extent
.depth
, range
->baseMipLevel
),
1344 /* Now go through the underlying image selected planes (computed in
1345 * expanded_aspects) and map them to planes in the image view.
1347 uint32_t iaspect_bit
, vplane
= 0;
1348 anv_foreach_image_aspect_bit(iaspect_bit
, image
, expanded_aspects
) {
1350 anv_image_aspect_to_plane(image
->aspects
, 1UL << iaspect_bit
);
1351 VkImageAspectFlags vplane_aspect
=
1352 anv_plane_to_aspect(iview
->aspect_mask
, vplane
);
1353 struct anv_format_plane format
=
1354 anv_get_format_plane(&device
->info
, pCreateInfo
->format
,
1355 vplane_aspect
, image
->tiling
);
1357 iview
->planes
[vplane
].image_plane
= iplane
;
1359 iview
->planes
[vplane
].isl
= (struct isl_view
) {
1360 .format
= format
.isl_format
,
1361 .base_level
= range
->baseMipLevel
,
1362 .levels
= anv_get_levelCount(image
, range
),
1363 .base_array_layer
= range
->baseArrayLayer
,
1364 .array_len
= anv_get_layerCount(image
, range
),
1366 .r
= remap_swizzle(pCreateInfo
->components
.r
,
1367 VK_COMPONENT_SWIZZLE_R
, format
.swizzle
),
1368 .g
= remap_swizzle(pCreateInfo
->components
.g
,
1369 VK_COMPONENT_SWIZZLE_G
, format
.swizzle
),
1370 .b
= remap_swizzle(pCreateInfo
->components
.b
,
1371 VK_COMPONENT_SWIZZLE_B
, format
.swizzle
),
1372 .a
= remap_swizzle(pCreateInfo
->components
.a
,
1373 VK_COMPONENT_SWIZZLE_A
, format
.swizzle
),
1377 if (pCreateInfo
->viewType
== VK_IMAGE_VIEW_TYPE_3D
) {
1378 iview
->planes
[vplane
].isl
.base_array_layer
= 0;
1379 iview
->planes
[vplane
].isl
.array_len
= iview
->extent
.depth
;
1382 if (pCreateInfo
->viewType
== VK_IMAGE_VIEW_TYPE_CUBE
||
1383 pCreateInfo
->viewType
== VK_IMAGE_VIEW_TYPE_CUBE_ARRAY
) {
1384 iview
->planes
[vplane
].isl
.usage
= ISL_SURF_USAGE_CUBE_BIT
;
1386 iview
->planes
[vplane
].isl
.usage
= 0;
1389 if (view_usage
& VK_IMAGE_USAGE_SAMPLED_BIT
||
1390 (view_usage
& VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT
&&
1391 !(iview
->aspect_mask
& VK_IMAGE_ASPECT_COLOR_BIT
))) {
1392 iview
->planes
[vplane
].optimal_sampler_surface_state
.state
= alloc_surface_state(device
);
1393 iview
->planes
[vplane
].general_sampler_surface_state
.state
= alloc_surface_state(device
);
1395 enum isl_aux_usage general_aux_usage
=
1396 anv_layout_to_aux_usage(&device
->info
, image
, 1UL << iaspect_bit
,
1397 VK_IMAGE_LAYOUT_GENERAL
);
1398 enum isl_aux_usage optimal_aux_usage
=
1399 anv_layout_to_aux_usage(&device
->info
, image
, 1UL << iaspect_bit
,
1400 VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL
);
1402 anv_image_fill_surface_state(device
, image
, 1ULL << iaspect_bit
,
1403 &iview
->planes
[vplane
].isl
,
1404 ISL_SURF_USAGE_TEXTURE_BIT
,
1405 optimal_aux_usage
, NULL
,
1406 ANV_IMAGE_VIEW_STATE_TEXTURE_OPTIMAL
,
1407 &iview
->planes
[vplane
].optimal_sampler_surface_state
,
1410 anv_image_fill_surface_state(device
, image
, 1ULL << iaspect_bit
,
1411 &iview
->planes
[vplane
].isl
,
1412 ISL_SURF_USAGE_TEXTURE_BIT
,
1413 general_aux_usage
, NULL
,
1415 &iview
->planes
[vplane
].general_sampler_surface_state
,
1419 /* NOTE: This one needs to go last since it may stomp isl_view.format */
1420 if (view_usage
& VK_IMAGE_USAGE_STORAGE_BIT
) {
1421 iview
->planes
[vplane
].storage_surface_state
.state
= alloc_surface_state(device
);
1422 iview
->planes
[vplane
].writeonly_storage_surface_state
.state
= alloc_surface_state(device
);
1424 anv_image_fill_surface_state(device
, image
, 1ULL << iaspect_bit
,
1425 &iview
->planes
[vplane
].isl
,
1426 ISL_SURF_USAGE_STORAGE_BIT
,
1427 ISL_AUX_USAGE_NONE
, NULL
,
1429 &iview
->planes
[vplane
].storage_surface_state
,
1430 &iview
->planes
[vplane
].storage_image_param
);
1432 anv_image_fill_surface_state(device
, image
, 1ULL << iaspect_bit
,
1433 &iview
->planes
[vplane
].isl
,
1434 ISL_SURF_USAGE_STORAGE_BIT
,
1435 ISL_AUX_USAGE_NONE
, NULL
,
1436 ANV_IMAGE_VIEW_STATE_STORAGE_WRITE_ONLY
,
1437 &iview
->planes
[vplane
].writeonly_storage_surface_state
,
1444 *pView
= anv_image_view_to_handle(iview
);
1450 anv_DestroyImageView(VkDevice _device
, VkImageView _iview
,
1451 const VkAllocationCallbacks
*pAllocator
)
1453 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1454 ANV_FROM_HANDLE(anv_image_view
, iview
, _iview
);
1459 for (uint32_t plane
= 0; plane
< iview
->n_planes
; plane
++) {
1460 if (iview
->planes
[plane
].optimal_sampler_surface_state
.state
.alloc_size
> 0) {
1461 anv_state_pool_free(&device
->surface_state_pool
,
1462 iview
->planes
[plane
].optimal_sampler_surface_state
.state
);
1465 if (iview
->planes
[plane
].general_sampler_surface_state
.state
.alloc_size
> 0) {
1466 anv_state_pool_free(&device
->surface_state_pool
,
1467 iview
->planes
[plane
].general_sampler_surface_state
.state
);
1470 if (iview
->planes
[plane
].storage_surface_state
.state
.alloc_size
> 0) {
1471 anv_state_pool_free(&device
->surface_state_pool
,
1472 iview
->planes
[plane
].storage_surface_state
.state
);
1475 if (iview
->planes
[plane
].writeonly_storage_surface_state
.state
.alloc_size
> 0) {
1476 anv_state_pool_free(&device
->surface_state_pool
,
1477 iview
->planes
[plane
].writeonly_storage_surface_state
.state
);
1481 vk_free2(&device
->alloc
, pAllocator
, iview
);
1486 anv_CreateBufferView(VkDevice _device
,
1487 const VkBufferViewCreateInfo
*pCreateInfo
,
1488 const VkAllocationCallbacks
*pAllocator
,
1489 VkBufferView
*pView
)
1491 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1492 ANV_FROM_HANDLE(anv_buffer
, buffer
, pCreateInfo
->buffer
);
1493 struct anv_buffer_view
*view
;
1495 view
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*view
), 8,
1496 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1498 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1500 /* TODO: Handle the format swizzle? */
1502 view
->format
= anv_get_isl_format(&device
->info
, pCreateInfo
->format
,
1503 VK_IMAGE_ASPECT_COLOR_BIT
,
1504 VK_IMAGE_TILING_LINEAR
);
1505 const uint32_t format_bs
= isl_format_get_layout(view
->format
)->bpb
/ 8;
1506 view
->range
= anv_buffer_get_range(buffer
, pCreateInfo
->offset
,
1507 pCreateInfo
->range
);
1508 view
->range
= align_down_npot_u32(view
->range
, format_bs
);
1510 view
->address
= anv_address_add(buffer
->address
, pCreateInfo
->offset
);
1512 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT
) {
1513 view
->surface_state
= alloc_surface_state(device
);
1515 anv_fill_buffer_surface_state(device
, view
->surface_state
,
1517 view
->address
, view
->range
, format_bs
);
1519 view
->surface_state
= (struct anv_state
){ 0 };
1522 if (buffer
->usage
& VK_BUFFER_USAGE_STORAGE_TEXEL_BUFFER_BIT
) {
1523 view
->storage_surface_state
= alloc_surface_state(device
);
1524 view
->writeonly_storage_surface_state
= alloc_surface_state(device
);
1526 enum isl_format storage_format
=
1527 isl_has_matching_typed_storage_image_format(&device
->info
,
1529 isl_lower_storage_image_format(&device
->info
, view
->format
) :
1532 anv_fill_buffer_surface_state(device
, view
->storage_surface_state
,
1534 view
->address
, view
->range
,
1535 (storage_format
== ISL_FORMAT_RAW
? 1 :
1536 isl_format_get_layout(storage_format
)->bpb
/ 8));
1538 /* Write-only accesses should use the original format. */
1539 anv_fill_buffer_surface_state(device
, view
->writeonly_storage_surface_state
,
1541 view
->address
, view
->range
,
1542 isl_format_get_layout(view
->format
)->bpb
/ 8);
1544 isl_buffer_fill_image_param(&device
->isl_dev
,
1545 &view
->storage_image_param
,
1546 view
->format
, view
->range
);
1548 view
->storage_surface_state
= (struct anv_state
){ 0 };
1549 view
->writeonly_storage_surface_state
= (struct anv_state
){ 0 };
1552 *pView
= anv_buffer_view_to_handle(view
);
1558 anv_DestroyBufferView(VkDevice _device
, VkBufferView bufferView
,
1559 const VkAllocationCallbacks
*pAllocator
)
1561 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1562 ANV_FROM_HANDLE(anv_buffer_view
, view
, bufferView
);
1567 if (view
->surface_state
.alloc_size
> 0)
1568 anv_state_pool_free(&device
->surface_state_pool
,
1569 view
->surface_state
);
1571 if (view
->storage_surface_state
.alloc_size
> 0)
1572 anv_state_pool_free(&device
->surface_state_pool
,
1573 view
->storage_surface_state
);
1575 if (view
->writeonly_storage_surface_state
.alloc_size
> 0)
1576 anv_state_pool_free(&device
->surface_state_pool
,
1577 view
->writeonly_storage_surface_state
);
1579 vk_free2(&device
->alloc
, pAllocator
, view
);
1582 const struct anv_surface
*
1583 anv_image_get_surface_for_aspect_mask(const struct anv_image
*image
,
1584 VkImageAspectFlags aspect_mask
)
1586 VkImageAspectFlags sanitized_mask
;
1588 switch (aspect_mask
) {
1589 case VK_IMAGE_ASPECT_COLOR_BIT
:
1590 assert(image
->aspects
== VK_IMAGE_ASPECT_COLOR_BIT
);
1591 sanitized_mask
= VK_IMAGE_ASPECT_COLOR_BIT
;
1593 case VK_IMAGE_ASPECT_DEPTH_BIT
:
1594 assert(image
->aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
);
1595 sanitized_mask
= VK_IMAGE_ASPECT_DEPTH_BIT
;
1597 case VK_IMAGE_ASPECT_STENCIL_BIT
:
1598 assert(image
->aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
);
1599 sanitized_mask
= VK_IMAGE_ASPECT_STENCIL_BIT
;
1601 case VK_IMAGE_ASPECT_DEPTH_BIT
| VK_IMAGE_ASPECT_STENCIL_BIT
:
1602 /* FINISHME: The Vulkan spec (git a511ba2) requires support for
1603 * combined depth stencil formats. Specifically, it states:
1605 * At least one of ename:VK_FORMAT_D24_UNORM_S8_UINT or
1606 * ename:VK_FORMAT_D32_SFLOAT_S8_UINT must be supported.
1608 * Image views with both depth and stencil aspects are only valid for
1609 * render target attachments, in which case
1610 * cmd_buffer_emit_depth_stencil() will pick out both the depth and
1611 * stencil surfaces from the underlying surface.
1613 if (image
->aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
) {
1614 sanitized_mask
= VK_IMAGE_ASPECT_DEPTH_BIT
;
1616 assert(image
->aspects
== VK_IMAGE_ASPECT_STENCIL_BIT
);
1617 sanitized_mask
= VK_IMAGE_ASPECT_STENCIL_BIT
;
1620 case VK_IMAGE_ASPECT_PLANE_0_BIT
:
1621 assert((image
->aspects
& ~VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) == 0);
1622 sanitized_mask
= VK_IMAGE_ASPECT_PLANE_0_BIT
;
1624 case VK_IMAGE_ASPECT_PLANE_1_BIT
:
1625 assert((image
->aspects
& ~VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) == 0);
1626 sanitized_mask
= VK_IMAGE_ASPECT_PLANE_1_BIT
;
1628 case VK_IMAGE_ASPECT_PLANE_2_BIT
:
1629 assert((image
->aspects
& ~VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) == 0);
1630 sanitized_mask
= VK_IMAGE_ASPECT_PLANE_2_BIT
;
1633 unreachable("image does not have aspect");
1637 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
, sanitized_mask
);
1638 return &image
->planes
[plane
].surface
;