projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radv: set fmask_surf_index on fmask surfaces.
[mesa.git] / src / amd / vulkan / radv_image.c
1 /*
2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
4 *
5 * based in part on anv driver which is:
6 * Copyright © 2015 Intel Corporation
7 *
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:
14 *
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
17 * Software.
18 *
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
25 * IN THE SOFTWARE.
26 */
27
28 #include "radv_debug.h"
29 #include "radv_private.h"
30 #include "vk_format.h"
31 #include "vk_util.h"
32 #include "radv_radeon_winsys.h"
33 #include "sid.h"
34 #include "gfx9d.h"
35 #include "util/debug.h"
36 #include "util/u_atomic.h"
37 static unsigned
38 radv_choose_tiling(struct radv_device *device,
39 const struct radv_image_create_info *create_info)
40 {
41 const VkImageCreateInfo *pCreateInfo = create_info->vk_info;
42
43 if (pCreateInfo->tiling == VK_IMAGE_TILING_LINEAR) {
44 assert(pCreateInfo->samples <= 1);
45 return RADEON_SURF_MODE_LINEAR_ALIGNED;
46 }
47
48 if (!vk_format_is_compressed(pCreateInfo->format) &&
49 !vk_format_is_depth_or_stencil(pCreateInfo->format)
50 && device->physical_device->rad_info.chip_class <= VI) {
51 /* this causes hangs in some VK CTS tests on GFX9. */
52 /* Textures with a very small height are recommended to be linear. */
53 if (pCreateInfo->imageType == VK_IMAGE_TYPE_1D ||
54 /* Only very thin and long 2D textures should benefit from
55 * linear_aligned. */
56 (pCreateInfo->extent.width > 8 && pCreateInfo->extent.height <= 2))
57 return RADEON_SURF_MODE_LINEAR_ALIGNED;
58 }
59
60 /* MSAA resources must be 2D tiled. */
61 if (pCreateInfo->samples > 1)
62 return RADEON_SURF_MODE_2D;
63
64 return RADEON_SURF_MODE_2D;
65 }
66
67 static bool
68 radv_use_tc_compat_htile_for_image(struct radv_device *device,
69 const VkImageCreateInfo *pCreateInfo)
70 {
71 /* TC-compat HTILE is only available for GFX8+. */
72 if (device->physical_device->rad_info.chip_class < VI)
73 return false;
74
75 if (pCreateInfo->usage & VK_IMAGE_USAGE_STORAGE_BIT)
76 return false;
77
78 if (pCreateInfo->flags & (VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT |
79 VK_IMAGE_CREATE_EXTENDED_USAGE_BIT_KHR))
80 return false;
81
82 if (pCreateInfo->tiling == VK_IMAGE_TILING_LINEAR)
83 return false;
84
85 if (pCreateInfo->mipLevels > 1)
86 return false;
87
88 /* FIXME: for some reason TC compat with 2/4/8 samples breaks some cts
89 * tests - disable for now */
90 if (pCreateInfo->samples >= 2 &&
91 pCreateInfo->format == VK_FORMAT_D32_SFLOAT_S8_UINT)
92 return false;
93
94 /* GFX9 supports both 32-bit and 16-bit depth surfaces, while GFX8 only
95 * supports 32-bit. Though, it's possible to enable TC-compat for
96 * 16-bit depth surfaces if no Z planes are compressed.
97 */
98 if (pCreateInfo->format != VK_FORMAT_D32_SFLOAT_S8_UINT &&
99 pCreateInfo->format != VK_FORMAT_D32_SFLOAT &&
100 pCreateInfo->format != VK_FORMAT_D16_UNORM)
101 return false;
102
103 return true;
104 }
105
106 static bool
107 radv_use_dcc_for_image(struct radv_device *device,
108 const struct radv_image_create_info *create_info,
109 const VkImageCreateInfo *pCreateInfo)
110 {
111 bool dcc_compatible_formats;
112 bool blendable;
113
114 /* DCC (Delta Color Compression) is only available for GFX8+. */
115 if (device->physical_device->rad_info.chip_class < VI)
116 return false;
117
118 if (device->instance->debug_flags & RADV_DEBUG_NO_DCC)
119 return false;
120
121 /* TODO: Enable DCC for storage images. */
122 if ((pCreateInfo->usage & VK_IMAGE_USAGE_STORAGE_BIT) ||
123 (pCreateInfo->flags & VK_IMAGE_CREATE_EXTENDED_USAGE_BIT_KHR))
124 return false;
125
126 if (pCreateInfo->tiling == VK_IMAGE_TILING_LINEAR)
127 return false;
128
129 /* TODO: Enable DCC for mipmaps and array layers. */
130 if (pCreateInfo->mipLevels > 1 || pCreateInfo->arrayLayers > 1)
131 return false;
132
133 if (create_info->scanout)
134 return false;
135
136 /* FIXME: DCC for MSAA with 4x and 8x samples doesn't work yet, while
137 * 2x can be enabled with an option.
138 */
139 if (pCreateInfo->samples > 2 ||
140 (pCreateInfo->samples == 2 &&
141 !device->physical_device->dcc_msaa_allowed))
142 return false;
143
144 /* Determine if the formats are DCC compatible. */
145 dcc_compatible_formats =
146 radv_is_colorbuffer_format_supported(pCreateInfo->format,
147 &blendable);
148
149 if (pCreateInfo->flags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) {
150 const struct VkImageFormatListCreateInfoKHR *format_list =
151 (const struct VkImageFormatListCreateInfoKHR *)
152 vk_find_struct_const(pCreateInfo->pNext,
153 IMAGE_FORMAT_LIST_CREATE_INFO_KHR);
154
155 /* We have to ignore the existence of the list if viewFormatCount = 0 */
156 if (format_list && format_list->viewFormatCount) {
157 /* compatibility is transitive, so we only need to check
158 * one format with everything else. */
159 for (unsigned i = 0; i < format_list->viewFormatCount; ++i) {
160 if (!radv_dcc_formats_compatible(pCreateInfo->format,
161 format_list->pViewFormats[i]))
162 dcc_compatible_formats = false;
163 }
164 } else {
165 dcc_compatible_formats = false;
166 }
167 }
168
169 if (!dcc_compatible_formats)
170 return false;
171
172 return true;
173 }
174
175 static int
176 radv_init_surface(struct radv_device *device,
177 struct radeon_surf *surface,
178 const struct radv_image_create_info *create_info)
179 {
180 const VkImageCreateInfo *pCreateInfo = create_info->vk_info;
181 unsigned array_mode = radv_choose_tiling(device, create_info);
182 const struct vk_format_description *desc =
183 vk_format_description(pCreateInfo->format);
184 bool is_depth, is_stencil;
185
186 is_depth = vk_format_has_depth(desc);
187 is_stencil = vk_format_has_stencil(desc);
188
189 surface->blk_w = vk_format_get_blockwidth(pCreateInfo->format);
190 surface->blk_h = vk_format_get_blockheight(pCreateInfo->format);
191
192 surface->bpe = vk_format_get_blocksize(vk_format_depth_only(pCreateInfo->format));
193 /* align byte per element on dword */
194 if (surface->bpe == 3) {
195 surface->bpe = 4;
196 }
197 surface->flags = RADEON_SURF_SET(array_mode, MODE);
198
199 switch (pCreateInfo->imageType){
200 case VK_IMAGE_TYPE_1D:
201 if (pCreateInfo->arrayLayers > 1)
202 surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_1D_ARRAY, TYPE);
203 else
204 surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_1D, TYPE);
205 break;
206 case VK_IMAGE_TYPE_2D:
207 if (pCreateInfo->arrayLayers > 1)
208 surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_2D_ARRAY, TYPE);
209 else
210 surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_2D, TYPE);
211 break;
212 case VK_IMAGE_TYPE_3D:
213 surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_3D, TYPE);
214 break;
215 default:
216 unreachable("unhandled image type");
217 }
218
219 if (is_depth) {
220 surface->flags |= RADEON_SURF_ZBUFFER;
221 if (radv_use_tc_compat_htile_for_image(device, pCreateInfo))
222 surface->flags |= RADEON_SURF_TC_COMPATIBLE_HTILE;
223 }
224
225 if (is_stencil)
226 surface->flags |= RADEON_SURF_SBUFFER;
227
228 surface->flags |= RADEON_SURF_OPTIMIZE_FOR_SPACE;
229
230 if (!radv_use_dcc_for_image(device, create_info, pCreateInfo))
231 surface->flags |= RADEON_SURF_DISABLE_DCC;
232
233 if (create_info->scanout)
234 surface->flags |= RADEON_SURF_SCANOUT;
235 return 0;
236 }
237
238 static uint32_t si_get_bo_metadata_word1(struct radv_device *device)
239 {
240 return (ATI_VENDOR_ID << 16) | device->physical_device->rad_info.pci_id;
241 }
242
243 static inline unsigned
244 si_tile_mode_index(const struct radv_image *image, unsigned level, bool stencil)
245 {
246 if (stencil)
247 return image->surface.u.legacy.stencil_tiling_index[level];
248 else
249 return image->surface.u.legacy.tiling_index[level];
250 }
251
252 static unsigned radv_map_swizzle(unsigned swizzle)
253 {
254 switch (swizzle) {
255 case VK_SWIZZLE_Y:
256 return V_008F0C_SQ_SEL_Y;
257 case VK_SWIZZLE_Z:
258 return V_008F0C_SQ_SEL_Z;
259 case VK_SWIZZLE_W:
260 return V_008F0C_SQ_SEL_W;
261 case VK_SWIZZLE_0:
262 return V_008F0C_SQ_SEL_0;
263 case VK_SWIZZLE_1:
264 return V_008F0C_SQ_SEL_1;
265 default: /* VK_SWIZZLE_X */
266 return V_008F0C_SQ_SEL_X;
267 }
268 }
269
270 static void
271 radv_make_buffer_descriptor(struct radv_device *device,
272 struct radv_buffer *buffer,
273 VkFormat vk_format,
274 unsigned offset,
275 unsigned range,
276 uint32_t *state)
277 {
278 const struct vk_format_description *desc;
279 unsigned stride;
280 uint64_t gpu_address = radv_buffer_get_va(buffer->bo);
281 uint64_t va = gpu_address + buffer->offset;
282 unsigned num_format, data_format;
283 int first_non_void;
284 desc = vk_format_description(vk_format);
285 first_non_void = vk_format_get_first_non_void_channel(vk_format);
286 stride = desc->block.bits / 8;
287
288 num_format = radv_translate_buffer_numformat(desc, first_non_void);
289 data_format = radv_translate_buffer_dataformat(desc, first_non_void);
290
291 va += offset;
292 state[0] = va;
293 state[1] = S_008F04_BASE_ADDRESS_HI(va >> 32) |
294 S_008F04_STRIDE(stride);
295
296 if (device->physical_device->rad_info.chip_class != VI && stride) {
297 range /= stride;
298 }
299
300 state[2] = range;
301 state[3] = S_008F0C_DST_SEL_X(radv_map_swizzle(desc->swizzle[0])) |
302 S_008F0C_DST_SEL_Y(radv_map_swizzle(desc->swizzle[1])) |
303 S_008F0C_DST_SEL_Z(radv_map_swizzle(desc->swizzle[2])) |
304 S_008F0C_DST_SEL_W(radv_map_swizzle(desc->swizzle[3])) |
305 S_008F0C_NUM_FORMAT(num_format) |
306 S_008F0C_DATA_FORMAT(data_format);
307 }
308
309 static void
310 si_set_mutable_tex_desc_fields(struct radv_device *device,
311 struct radv_image *image,
312 const struct legacy_surf_level *base_level_info,
313 unsigned base_level, unsigned first_level,
314 unsigned block_width, bool is_stencil,
315 bool is_storage_image, uint32_t *state)
316 {
317 uint64_t gpu_address = image->bo ? radv_buffer_get_va(image->bo) + image->offset : 0;
318 uint64_t va = gpu_address;
319 enum chip_class chip_class = device->physical_device->rad_info.chip_class;
320 uint64_t meta_va = 0;
321 if (chip_class >= GFX9) {
322 if (is_stencil)
323 va += image->surface.u.gfx9.stencil_offset;
324 else
325 va += image->surface.u.gfx9.surf_offset;
326 } else
327 va += base_level_info->offset;
328
329 state[0] = va >> 8;
330 if (chip_class >= GFX9 ||
331 base_level_info->mode == RADEON_SURF_MODE_2D)
332 state[0] |= image->surface.tile_swizzle;
333 state[1] &= C_008F14_BASE_ADDRESS_HI;
334 state[1] |= S_008F14_BASE_ADDRESS_HI(va >> 40);
335
336 if (chip_class >= VI) {
337 state[6] &= C_008F28_COMPRESSION_EN;
338 state[7] = 0;
339 if (!is_storage_image && radv_dcc_enabled(image, first_level)) {
340 meta_va = gpu_address + image->dcc_offset;
341 if (chip_class <= VI)
342 meta_va += base_level_info->dcc_offset;
343 } else if (!is_storage_image &&
344 radv_image_is_tc_compat_htile(image)) {
345 meta_va = gpu_address + image->htile_offset;
346 }
347
348 if (meta_va) {
349 state[6] |= S_008F28_COMPRESSION_EN(1);
350 state[7] = meta_va >> 8;
351 state[7] |= image->surface.tile_swizzle;
352 }
353 }
354
355 if (chip_class >= GFX9) {
356 state[3] &= C_008F1C_SW_MODE;
357 state[4] &= C_008F20_PITCH_GFX9;
358
359 if (is_stencil) {
360 state[3] |= S_008F1C_SW_MODE(image->surface.u.gfx9.stencil.swizzle_mode);
361 state[4] |= S_008F20_PITCH_GFX9(image->surface.u.gfx9.stencil.epitch);
362 } else {
363 state[3] |= S_008F1C_SW_MODE(image->surface.u.gfx9.surf.swizzle_mode);
364 state[4] |= S_008F20_PITCH_GFX9(image->surface.u.gfx9.surf.epitch);
365 }
366
367 state[5] &= C_008F24_META_DATA_ADDRESS &
368 C_008F24_META_PIPE_ALIGNED &
369 C_008F24_META_RB_ALIGNED;
370 if (meta_va) {
371 struct gfx9_surf_meta_flags meta;
372
373 if (image->dcc_offset)
374 meta = image->surface.u.gfx9.dcc;
375 else
376 meta = image->surface.u.gfx9.htile;
377
378 state[5] |= S_008F24_META_DATA_ADDRESS(meta_va >> 40) |
379 S_008F24_META_PIPE_ALIGNED(meta.pipe_aligned) |
380 S_008F24_META_RB_ALIGNED(meta.rb_aligned);
381 }
382 } else {
383 /* SI-CI-VI */
384 unsigned pitch = base_level_info->nblk_x * block_width;
385 unsigned index = si_tile_mode_index(image, base_level, is_stencil);
386
387 state[3] &= C_008F1C_TILING_INDEX;
388 state[3] |= S_008F1C_TILING_INDEX(index);
389 state[4] &= C_008F20_PITCH_GFX6;
390 state[4] |= S_008F20_PITCH_GFX6(pitch - 1);
391 }
392 }
393
394 static unsigned radv_tex_dim(VkImageType image_type, VkImageViewType view_type,
395 unsigned nr_layers, unsigned nr_samples, bool is_storage_image, bool gfx9)
396 {
397 if (view_type == VK_IMAGE_VIEW_TYPE_CUBE || view_type == VK_IMAGE_VIEW_TYPE_CUBE_ARRAY)
398 return is_storage_image ? V_008F1C_SQ_RSRC_IMG_2D_ARRAY : V_008F1C_SQ_RSRC_IMG_CUBE;
399
400 /* GFX9 allocates 1D textures as 2D. */
401 if (gfx9 && image_type == VK_IMAGE_TYPE_1D)
402 image_type = VK_IMAGE_TYPE_2D;
403 switch (image_type) {
404 case VK_IMAGE_TYPE_1D:
405 return nr_layers > 1 ? V_008F1C_SQ_RSRC_IMG_1D_ARRAY : V_008F1C_SQ_RSRC_IMG_1D;
406 case VK_IMAGE_TYPE_2D:
407 if (nr_samples > 1)
408 return nr_layers > 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY : V_008F1C_SQ_RSRC_IMG_2D_MSAA;
409 else
410 return nr_layers > 1 ? V_008F1C_SQ_RSRC_IMG_2D_ARRAY : V_008F1C_SQ_RSRC_IMG_2D;
411 case VK_IMAGE_TYPE_3D:
412 if (view_type == VK_IMAGE_VIEW_TYPE_3D)
413 return V_008F1C_SQ_RSRC_IMG_3D;
414 else
415 return V_008F1C_SQ_RSRC_IMG_2D_ARRAY;
416 default:
417 unreachable("illegale image type");
418 }
419 }
420
421 static unsigned gfx9_border_color_swizzle(const enum vk_swizzle swizzle[4])
422 {
423 unsigned bc_swizzle = V_008F20_BC_SWIZZLE_XYZW;
424
425 if (swizzle[3] == VK_SWIZZLE_X) {
426 /* For the pre-defined border color values (white, opaque
427 * black, transparent black), the only thing that matters is
428 * that the alpha channel winds up in the correct place
429 * (because the RGB channels are all the same) so either of
430 * these enumerations will work.
431 */
432 if (swizzle[2] == VK_SWIZZLE_Y)
433 bc_swizzle = V_008F20_BC_SWIZZLE_WZYX;
434 else
435 bc_swizzle = V_008F20_BC_SWIZZLE_WXYZ;
436 } else if (swizzle[0] == VK_SWIZZLE_X) {
437 if (swizzle[1] == VK_SWIZZLE_Y)
438 bc_swizzle = V_008F20_BC_SWIZZLE_XYZW;
439 else
440 bc_swizzle = V_008F20_BC_SWIZZLE_XWYZ;
441 } else if (swizzle[1] == VK_SWIZZLE_X) {
442 bc_swizzle = V_008F20_BC_SWIZZLE_YXWZ;
443 } else if (swizzle[2] == VK_SWIZZLE_X) {
444 bc_swizzle = V_008F20_BC_SWIZZLE_ZYXW;
445 }
446
447 return bc_swizzle;
448 }
449
450 /**
451 * Build the sampler view descriptor for a texture.
452 */
453 static void
454 si_make_texture_descriptor(struct radv_device *device,
455 struct radv_image *image,
456 bool is_storage_image,
457 VkImageViewType view_type,
458 VkFormat vk_format,
459 const VkComponentMapping *mapping,
460 unsigned first_level, unsigned last_level,
461 unsigned first_layer, unsigned last_layer,
462 unsigned width, unsigned height, unsigned depth,
463 uint32_t *state,
464 uint32_t *fmask_state)
465 {
466 const struct vk_format_description *desc;
467 enum vk_swizzle swizzle[4];
468 int first_non_void;
469 unsigned num_format, data_format, type;
470
471 desc = vk_format_description(vk_format);
472
473 if (desc->colorspace == VK_FORMAT_COLORSPACE_ZS) {
474 const unsigned char swizzle_xxxx[4] = {0, 0, 0, 0};
475 vk_format_compose_swizzles(mapping, swizzle_xxxx, swizzle);
476 } else {
477 vk_format_compose_swizzles(mapping, desc->swizzle, swizzle);
478 }
479
480 first_non_void = vk_format_get_first_non_void_channel(vk_format);
481
482 num_format = radv_translate_tex_numformat(vk_format, desc, first_non_void);
483 if (num_format == ~0) {
484 num_format = 0;
485 }
486
487 data_format = radv_translate_tex_dataformat(vk_format, desc, first_non_void);
488 if (data_format == ~0) {
489 data_format = 0;
490 }
491
492 /* S8 with either Z16 or Z32 HTILE need a special format. */
493 if (device->physical_device->rad_info.chip_class >= GFX9 &&
494 vk_format == VK_FORMAT_S8_UINT &&
495 radv_image_is_tc_compat_htile(image)) {
496 if (image->vk_format == VK_FORMAT_D32_SFLOAT_S8_UINT)
497 data_format = V_008F14_IMG_DATA_FORMAT_S8_32;
498 else if (image->vk_format == VK_FORMAT_D16_UNORM_S8_UINT)
499 data_format = V_008F14_IMG_DATA_FORMAT_S8_16;
500 }
501 type = radv_tex_dim(image->type, view_type, image->info.array_size, image->info.samples,
502 is_storage_image, device->physical_device->rad_info.chip_class >= GFX9);
503 if (type == V_008F1C_SQ_RSRC_IMG_1D_ARRAY) {
504 height = 1;
505 depth = image->info.array_size;
506 } else if (type == V_008F1C_SQ_RSRC_IMG_2D_ARRAY ||
507 type == V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY) {
508 if (view_type != VK_IMAGE_VIEW_TYPE_3D)
509 depth = image->info.array_size;
510 } else if (type == V_008F1C_SQ_RSRC_IMG_CUBE)
511 depth = image->info.array_size / 6;
512
513 state[0] = 0;
514 state[1] = (S_008F14_DATA_FORMAT_GFX6(data_format) |
515 S_008F14_NUM_FORMAT_GFX6(num_format));
516 state[2] = (S_008F18_WIDTH(width - 1) |
517 S_008F18_HEIGHT(height - 1) |
518 S_008F18_PERF_MOD(4));
519 state[3] = (S_008F1C_DST_SEL_X(radv_map_swizzle(swizzle[0])) |
520 S_008F1C_DST_SEL_Y(radv_map_swizzle(swizzle[1])) |
521 S_008F1C_DST_SEL_Z(radv_map_swizzle(swizzle[2])) |
522 S_008F1C_DST_SEL_W(radv_map_swizzle(swizzle[3])) |
523 S_008F1C_BASE_LEVEL(image->info.samples > 1 ?
524 0 : first_level) |
525 S_008F1C_LAST_LEVEL(image->info.samples > 1 ?
526 util_logbase2(image->info.samples) :
527 last_level) |
528 S_008F1C_TYPE(type));
529 state[4] = 0;
530 state[5] = S_008F24_BASE_ARRAY(first_layer);
531 state[6] = 0;
532 state[7] = 0;
533
534 if (device->physical_device->rad_info.chip_class >= GFX9) {
535 unsigned bc_swizzle = gfx9_border_color_swizzle(swizzle);
536
537 /* Depth is the the last accessible layer on Gfx9.
538 * The hw doesn't need to know the total number of layers.
539 */
540 if (type == V_008F1C_SQ_RSRC_IMG_3D)
541 state[4] |= S_008F20_DEPTH(depth - 1);
542 else
543 state[4] |= S_008F20_DEPTH(last_layer);
544
545 state[4] |= S_008F20_BC_SWIZZLE(bc_swizzle);
546 state[5] |= S_008F24_MAX_MIP(image->info.samples > 1 ?
547 util_logbase2(image->info.samples) :
548 image->info.levels - 1);
549 } else {
550 state[3] |= S_008F1C_POW2_PAD(image->info.levels > 1);
551 state[4] |= S_008F20_DEPTH(depth - 1);
552 state[5] |= S_008F24_LAST_ARRAY(last_layer);
553 }
554 if (image->dcc_offset) {
555 unsigned swap = radv_translate_colorswap(vk_format, FALSE);
556
557 state[6] = S_008F28_ALPHA_IS_ON_MSB(swap <= 1);
558 } else {
559 /* The last dword is unused by hw. The shader uses it to clear
560 * bits in the first dword of sampler state.
561 */
562 if (device->physical_device->rad_info.chip_class <= CIK && image->info.samples <= 1) {
563 if (first_level == last_level)
564 state[7] = C_008F30_MAX_ANISO_RATIO;
565 else
566 state[7] = 0xffffffff;
567 }
568 }
569
570 /* Initialize the sampler view for FMASK. */
571 if (radv_image_has_fmask(image)) {
572 uint32_t fmask_format, num_format;
573 uint64_t gpu_address = radv_buffer_get_va(image->bo);
574 uint64_t va;
575
576 va = gpu_address + image->offset + image->fmask.offset;
577
578 if (device->physical_device->rad_info.chip_class >= GFX9) {
579 fmask_format = V_008F14_IMG_DATA_FORMAT_FMASK;
580 switch (image->info.samples) {
581 case 2:
582 num_format = V_008F14_IMG_FMASK_8_2_2;
583 break;
584 case 4:
585 num_format = V_008F14_IMG_FMASK_8_4_4;
586 break;
587 case 8:
588 num_format = V_008F14_IMG_FMASK_32_8_8;
589 break;
590 default:
591 unreachable("invalid nr_samples");
592 }
593 } else {
594 switch (image->info.samples) {
595 case 2:
596 fmask_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S2_F2;
597 break;
598 case 4:
599 fmask_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S4_F4;
600 break;
601 case 8:
602 fmask_format = V_008F14_IMG_DATA_FORMAT_FMASK32_S8_F8;
603 break;
604 default:
605 assert(0);
606 fmask_format = V_008F14_IMG_DATA_FORMAT_INVALID;
607 }
608 num_format = V_008F14_IMG_NUM_FORMAT_UINT;
609 }
610
611 fmask_state[0] = va >> 8;
612 fmask_state[0] |= image->fmask.tile_swizzle;
613 fmask_state[1] = S_008F14_BASE_ADDRESS_HI(va >> 40) |
614 S_008F14_DATA_FORMAT_GFX6(fmask_format) |
615 S_008F14_NUM_FORMAT_GFX6(num_format);
616 fmask_state[2] = S_008F18_WIDTH(width - 1) |
617 S_008F18_HEIGHT(height - 1);
618 fmask_state[3] = S_008F1C_DST_SEL_X(V_008F1C_SQ_SEL_X) |
619 S_008F1C_DST_SEL_Y(V_008F1C_SQ_SEL_X) |
620 S_008F1C_DST_SEL_Z(V_008F1C_SQ_SEL_X) |
621 S_008F1C_DST_SEL_W(V_008F1C_SQ_SEL_X) |
622 S_008F1C_TYPE(radv_tex_dim(image->type, view_type, 1, 0, false, false));
623 fmask_state[4] = 0;
624 fmask_state[5] = S_008F24_BASE_ARRAY(first_layer);
625 fmask_state[6] = 0;
626 fmask_state[7] = 0;
627
628 if (device->physical_device->rad_info.chip_class >= GFX9) {
629 fmask_state[3] |= S_008F1C_SW_MODE(image->surface.u.gfx9.fmask.swizzle_mode);
630 fmask_state[4] |= S_008F20_DEPTH(last_layer) |
631 S_008F20_PITCH_GFX9(image->surface.u.gfx9.fmask.epitch);
632 fmask_state[5] |= S_008F24_META_PIPE_ALIGNED(image->surface.u.gfx9.cmask.pipe_aligned) |
633 S_008F24_META_RB_ALIGNED(image->surface.u.gfx9.cmask.rb_aligned);
634 } else {
635 fmask_state[3] |= S_008F1C_TILING_INDEX(image->fmask.tile_mode_index);
636 fmask_state[4] |= S_008F20_DEPTH(depth - 1) |
637 S_008F20_PITCH_GFX6(image->fmask.pitch_in_pixels - 1);
638 fmask_state[5] |= S_008F24_LAST_ARRAY(last_layer);
639 }
640 } else if (fmask_state)
641 memset(fmask_state, 0, 8 * 4);
642 }
643
644 static void
645 radv_query_opaque_metadata(struct radv_device *device,
646 struct radv_image *image,
647 struct radeon_bo_metadata *md)
648 {
649 static const VkComponentMapping fixedmapping;
650 uint32_t desc[8], i;
651
652 /* Metadata image format format version 1:
653 * [0] = 1 (metadata format identifier)
654 * [1] = (VENDOR_ID << 16) | PCI_ID
655 * [2:9] = image descriptor for the whole resource
656 * [2] is always 0, because the base address is cleared
657 * [9] is the DCC offset bits [39:8] from the beginning of
658 * the buffer
659 * [10:10+LAST_LEVEL] = mipmap level offset bits [39:8] for each level
660 */
661 md->metadata[0] = 1; /* metadata image format version 1 */
662
663 /* TILE_MODE_INDEX is ambiguous without a PCI ID. */
664 md->metadata[1] = si_get_bo_metadata_word1(device);
665
666
667 si_make_texture_descriptor(device, image, false,
668 (VkImageViewType)image->type, image->vk_format,
669 &fixedmapping, 0, image->info.levels - 1, 0,
670 image->info.array_size,
671 image->info.width, image->info.height,
672 image->info.depth,
673 desc, NULL);
674
675 si_set_mutable_tex_desc_fields(device, image, &image->surface.u.legacy.level[0], 0, 0,
676 image->surface.blk_w, false, false, desc);
677
678 /* Clear the base address and set the relative DCC offset. */
679 desc[0] = 0;
680 desc[1] &= C_008F14_BASE_ADDRESS_HI;
681 desc[7] = image->dcc_offset >> 8;
682
683 /* Dwords [2:9] contain the image descriptor. */
684 memcpy(&md->metadata[2], desc, sizeof(desc));
685
686 /* Dwords [10:..] contain the mipmap level offsets. */
687 if (device->physical_device->rad_info.chip_class <= VI) {
688 for (i = 0; i <= image->info.levels - 1; i++)
689 md->metadata[10+i] = image->surface.u.legacy.level[i].offset >> 8;
690 md->size_metadata = (11 + image->info.levels - 1) * 4;
691 }
692 }
693
694 void
695 radv_init_metadata(struct radv_device *device,
696 struct radv_image *image,
697 struct radeon_bo_metadata *metadata)
698 {
699 struct radeon_surf *surface = &image->surface;
700
701 memset(metadata, 0, sizeof(*metadata));
702
703 if (device->physical_device->rad_info.chip_class >= GFX9) {
704 metadata->u.gfx9.swizzle_mode = surface->u.gfx9.surf.swizzle_mode;
705 } else {
706 metadata->u.legacy.microtile = surface->u.legacy.level[0].mode >= RADEON_SURF_MODE_1D ?
707 RADEON_LAYOUT_TILED : RADEON_LAYOUT_LINEAR;
708 metadata->u.legacy.macrotile = surface->u.legacy.level[0].mode >= RADEON_SURF_MODE_2D ?
709 RADEON_LAYOUT_TILED : RADEON_LAYOUT_LINEAR;
710 metadata->u.legacy.pipe_config = surface->u.legacy.pipe_config;
711 metadata->u.legacy.bankw = surface->u.legacy.bankw;
712 metadata->u.legacy.bankh = surface->u.legacy.bankh;
713 metadata->u.legacy.tile_split = surface->u.legacy.tile_split;
714 metadata->u.legacy.mtilea = surface->u.legacy.mtilea;
715 metadata->u.legacy.num_banks = surface->u.legacy.num_banks;
716 metadata->u.legacy.stride = surface->u.legacy.level[0].nblk_x * surface->bpe;
717 metadata->u.legacy.scanout = (surface->flags & RADEON_SURF_SCANOUT) != 0;
718 }
719 radv_query_opaque_metadata(device, image, metadata);
720 }
721
722 /* The number of samples can be specified independently of the texture. */
723 static void
724 radv_image_get_fmask_info(struct radv_device *device,
725 struct radv_image *image,
726 unsigned nr_samples,
727 struct radv_fmask_info *out)
728 {
729 /* FMASK is allocated like an ordinary texture. */
730 struct radeon_surf fmask = {};
731 struct ac_surf_info info = image->info;
732 memset(out, 0, sizeof(*out));
733
734 if (device->physical_device->rad_info.chip_class >= GFX9) {
735 out->alignment = image->surface.u.gfx9.fmask_alignment;
736 out->size = image->surface.u.gfx9.fmask_size;
737 return;
738 }
739
740 fmask.blk_w = image->surface.blk_w;
741 fmask.blk_h = image->surface.blk_h;
742 info.samples = 1;
743 fmask.flags = image->surface.flags | RADEON_SURF_FMASK;
744
745 if (!image->shareable) {
746 info.fmask_surf_index = &device->fmask_mrt_offset_counter;
747 info.surf_index = &device->fmask_mrt_offset_counter;
748 }
749
750 /* Force 2D tiling if it wasn't set. This may occur when creating
751 * FMASK for MSAA resolve on R6xx. On R6xx, the single-sample
752 * destination buffer must have an FMASK too. */
753 fmask.flags = RADEON_SURF_CLR(fmask.flags, MODE);
754 fmask.flags |= RADEON_SURF_SET(RADEON_SURF_MODE_2D, MODE);
755
756 switch (nr_samples) {
757 case 2:
758 case 4:
759 fmask.bpe = 1;
760 break;
761 case 8:
762 fmask.bpe = 4;
763 break;
764 default:
765 return;
766 }
767
768 device->ws->surface_init(device->ws, &info, &fmask);
769 assert(fmask.u.legacy.level[0].mode == RADEON_SURF_MODE_2D);
770
771 out->slice_tile_max = (fmask.u.legacy.level[0].nblk_x * fmask.u.legacy.level[0].nblk_y) / 64;
772 if (out->slice_tile_max)
773 out->slice_tile_max -= 1;
774
775 out->tile_mode_index = fmask.u.legacy.tiling_index[0];
776 out->pitch_in_pixels = fmask.u.legacy.level[0].nblk_x;
777 out->bank_height = fmask.u.legacy.bankh;
778 out->tile_swizzle = fmask.tile_swizzle;
779 out->alignment = MAX2(256, fmask.surf_alignment);
780 out->size = fmask.surf_size;
781
782 assert(!out->tile_swizzle || !image->shareable);
783 }
784
785 static void
786 radv_image_alloc_fmask(struct radv_device *device,
787 struct radv_image *image)
788 {
789 radv_image_get_fmask_info(device, image, image->info.samples, &image->fmask);
790
791 image->fmask.offset = align64(image->size, image->fmask.alignment);
792 image->size = image->fmask.offset + image->fmask.size;
793 image->alignment = MAX2(image->alignment, image->fmask.alignment);
794 }
795
796 static void
797 radv_image_get_cmask_info(struct radv_device *device,
798 struct radv_image *image,
799 struct radv_cmask_info *out)
800 {
801 unsigned pipe_interleave_bytes = device->physical_device->rad_info.pipe_interleave_bytes;
802 unsigned num_pipes = device->physical_device->rad_info.num_tile_pipes;
803 unsigned cl_width, cl_height;
804
805 if (device->physical_device->rad_info.chip_class >= GFX9) {
806 out->alignment = image->surface.u.gfx9.cmask_alignment;
807 out->size = image->surface.u.gfx9.cmask_size;
808 return;
809 }
810
811 switch (num_pipes) {
812 case 2:
813 cl_width = 32;
814 cl_height = 16;
815 break;
816 case 4:
817 cl_width = 32;
818 cl_height = 32;
819 break;
820 case 8:
821 cl_width = 64;
822 cl_height = 32;
823 break;
824 case 16: /* Hawaii */
825 cl_width = 64;
826 cl_height = 64;
827 break;
828 default:
829 assert(0);
830 return;
831 }
832
833 unsigned base_align = num_pipes * pipe_interleave_bytes;
834
835 unsigned width = align(image->info.width, cl_width*8);
836 unsigned height = align(image->info.height, cl_height*8);
837 unsigned slice_elements = (width * height) / (8*8);
838
839 /* Each element of CMASK is a nibble. */
840 unsigned slice_bytes = slice_elements / 2;
841
842 out->slice_tile_max = (width * height) / (128*128);
843 if (out->slice_tile_max)
844 out->slice_tile_max -= 1;
845
846 out->alignment = MAX2(256, base_align);
847 out->size = (image->type == VK_IMAGE_TYPE_3D ? image->info.depth : image->info.array_size) *
848 align(slice_bytes, base_align);
849 }
850
851 static void
852 radv_image_alloc_cmask(struct radv_device *device,
853 struct radv_image *image)
854 {
855 uint32_t clear_value_size = 0;
856 radv_image_get_cmask_info(device, image, &image->cmask);
857
858 image->cmask.offset = align64(image->size, image->cmask.alignment);
859 /* + 8 for storing the clear values */
860 if (!image->clear_value_offset) {
861 image->clear_value_offset = image->cmask.offset + image->cmask.size;
862 clear_value_size = 8;
863 }
864 image->size = image->cmask.offset + image->cmask.size + clear_value_size;
865 image->alignment = MAX2(image->alignment, image->cmask.alignment);
866 }
867
868 static void
869 radv_image_alloc_dcc(struct radv_image *image)
870 {
871 image->dcc_offset = align64(image->size, image->surface.dcc_alignment);
872 /* + 16 for storing the clear values + dcc pred */
873 image->clear_value_offset = image->dcc_offset + image->surface.dcc_size;
874 image->dcc_pred_offset = image->clear_value_offset + 8;
875 image->size = image->dcc_offset + image->surface.dcc_size + 16;
876 image->alignment = MAX2(image->alignment, image->surface.dcc_alignment);
877 }
878
879 static void
880 radv_image_alloc_htile(struct radv_image *image)
881 {
882 image->htile_offset = align64(image->size, image->surface.htile_alignment);
883
884 /* + 8 for storing the clear values */
885 image->clear_value_offset = image->htile_offset + image->surface.htile_size;
886 image->size = image->clear_value_offset + 8;
887 image->alignment = align64(image->alignment, image->surface.htile_alignment);
888 }
889
890 static inline bool
891 radv_image_can_enable_dcc_or_cmask(struct radv_image *image)
892 {
893 if (image->info.samples <= 1 &&
894 image->info.width * image->info.height <= 512 * 512) {
895 /* Do not enable CMASK or DCC for small surfaces where the cost
896 * of the eliminate pass can be higher than the benefit of fast
897 * clear. RadeonSI does this, but the image threshold is
898 * different.
899 */
900 return false;
901 }
902
903 return image->usage & VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT &&
904 (image->exclusive || image->queue_family_mask == 1);
905 }
906
907 static inline bool
908 radv_image_can_enable_dcc(struct radv_image *image)
909 {
910 return radv_image_can_enable_dcc_or_cmask(image) &&
911 radv_image_has_dcc(image);
912 }
913
914 static inline bool
915 radv_image_can_enable_cmask(struct radv_image *image)
916 {
917 if (image->surface.bpe > 8 && image->info.samples == 1) {
918 /* Do not enable CMASK for non-MSAA images (fast color clear)
919 * because 128 bit formats are not supported, but FMASK might
920 * still be used.
921 */
922 return false;
923 }
924
925 return radv_image_can_enable_dcc_or_cmask(image) &&
926 image->info.levels == 1 &&
927 image->info.depth == 1 &&
928 !image->surface.is_linear;
929 }
930
931 static inline bool
932 radv_image_can_enable_fmask(struct radv_image *image)
933 {
934 return image->info.samples > 1 && vk_format_is_color(image->vk_format);
935 }
936
937 static inline bool
938 radv_image_can_enable_htile(struct radv_image *image)
939 {
940 return image->info.levels == 1 && vk_format_is_depth(image->vk_format);
941 }
942
943 VkResult
944 radv_image_create(VkDevice _device,
945 const struct radv_image_create_info *create_info,
946 const VkAllocationCallbacks* alloc,
947 VkImage *pImage)
948 {
949 RADV_FROM_HANDLE(radv_device, device, _device);
950 const VkImageCreateInfo *pCreateInfo = create_info->vk_info;
951 struct radv_image *image = NULL;
952 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO);
953
954 radv_assert(pCreateInfo->mipLevels > 0);
955 radv_assert(pCreateInfo->arrayLayers > 0);
956 radv_assert(pCreateInfo->samples > 0);
957 radv_assert(pCreateInfo->extent.width > 0);
958 radv_assert(pCreateInfo->extent.height > 0);
959 radv_assert(pCreateInfo->extent.depth > 0);
960
961 image = vk_zalloc2(&device->alloc, alloc, sizeof(*image), 8,
962 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
963 if (!image)
964 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
965
966 image->type = pCreateInfo->imageType;
967 image->info.width = pCreateInfo->extent.width;
968 image->info.height = pCreateInfo->extent.height;
969 image->info.depth = pCreateInfo->extent.depth;
970 image->info.samples = pCreateInfo->samples;
971 image->info.array_size = pCreateInfo->arrayLayers;
972 image->info.levels = pCreateInfo->mipLevels;
973 image->info.num_channels = vk_format_get_nr_components(pCreateInfo->format);
974
975 image->vk_format = pCreateInfo->format;
976 image->tiling = pCreateInfo->tiling;
977 image->usage = pCreateInfo->usage;
978 image->flags = pCreateInfo->flags;
979
980 image->exclusive = pCreateInfo->sharingMode == VK_SHARING_MODE_EXCLUSIVE;
981 if (pCreateInfo->sharingMode == VK_SHARING_MODE_CONCURRENT) {
982 for (uint32_t i = 0; i < pCreateInfo->queueFamilyIndexCount; ++i)
983 if (pCreateInfo->pQueueFamilyIndices[i] == VK_QUEUE_FAMILY_EXTERNAL_KHR)
984 image->queue_family_mask |= (1u << RADV_MAX_QUEUE_FAMILIES) - 1u;
985 else
986 image->queue_family_mask |= 1u << pCreateInfo->pQueueFamilyIndices[i];
987 }
988
989 image->shareable = vk_find_struct_const(pCreateInfo->pNext,
990 EXTERNAL_MEMORY_IMAGE_CREATE_INFO_KHR) != NULL;
991 if (!vk_format_is_depth(pCreateInfo->format) && !create_info->scanout && !image->shareable) {
992 image->info.surf_index = &device->image_mrt_offset_counter;
993 }
994
995 radv_init_surface(device, &image->surface, create_info);
996
997 device->ws->surface_init(device->ws, &image->info, &image->surface);
998
999 image->size = image->surface.surf_size;
1000 image->alignment = image->surface.surf_alignment;
1001
1002 if (!create_info->no_metadata_planes) {
1003 /* Try to enable DCC first. */
1004 if (radv_image_can_enable_dcc(image)) {
1005 radv_image_alloc_dcc(image);
1006 if (image->info.samples > 1) {
1007 /* CMASK should be enabled because DCC fast
1008 * clear with MSAA needs it.
1009 */
1010 assert(radv_image_can_enable_cmask(image));
1011 radv_image_alloc_cmask(device, image);
1012 }
1013 } else {
1014 /* When DCC cannot be enabled, try CMASK. */
1015 image->surface.dcc_size = 0;
1016 if (radv_image_can_enable_cmask(image)) {
1017 radv_image_alloc_cmask(device, image);
1018 }
1019 }
1020
1021 /* Try to enable FMASK for multisampled images. */
1022 if (radv_image_can_enable_fmask(image)) {
1023 radv_image_alloc_fmask(device, image);
1024 } else {
1025 /* Otherwise, try to enable HTILE for depth surfaces. */
1026 if (radv_image_can_enable_htile(image) &&
1027 !(device->instance->debug_flags & RADV_DEBUG_NO_HIZ)) {
1028 radv_image_alloc_htile(image);
1029 image->tc_compatible_htile = image->surface.flags & RADEON_SURF_TC_COMPATIBLE_HTILE;
1030 } else {
1031 image->surface.htile_size = 0;
1032 }
1033 }
1034 } else {
1035 image->surface.dcc_size = 0;
1036 image->surface.htile_size = 0;
1037 }
1038
1039 if (pCreateInfo->flags & VK_IMAGE_CREATE_SPARSE_BINDING_BIT) {
1040 image->alignment = MAX2(image->alignment, 4096);
1041 image->size = align64(image->size, image->alignment);
1042 image->offset = 0;
1043
1044 image->bo = device->ws->buffer_create(device->ws, image->size, image->alignment,
1045 0, RADEON_FLAG_VIRTUAL);
1046 if (!image->bo) {
1047 vk_free2(&device->alloc, alloc, image);
1048 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
1049 }
1050 }
1051
1052 *pImage = radv_image_to_handle(image);
1053
1054 return VK_SUCCESS;
1055 }
1056
1057 static void
1058 radv_image_view_make_descriptor(struct radv_image_view *iview,
1059 struct radv_device *device,
1060 const VkComponentMapping *components,
1061 bool is_storage_image)
1062 {
1063 struct radv_image *image = iview->image;
1064 bool is_stencil = iview->aspect_mask == VK_IMAGE_ASPECT_STENCIL_BIT;
1065 uint32_t blk_w;
1066 uint32_t *descriptor;
1067 uint32_t hw_level = 0;
1068
1069 if (is_storage_image) {
1070 descriptor = iview->storage_descriptor;
1071 } else {
1072 descriptor = iview->descriptor;
1073 }
1074
1075 assert(image->surface.blk_w % vk_format_get_blockwidth(image->vk_format) == 0);
1076 blk_w = image->surface.blk_w / vk_format_get_blockwidth(image->vk_format) * vk_format_get_blockwidth(iview->vk_format);
1077
1078 if (device->physical_device->rad_info.chip_class >= GFX9)
1079 hw_level = iview->base_mip;
1080 si_make_texture_descriptor(device, image, is_storage_image,
1081 iview->type,
1082 iview->vk_format,
1083 components,
1084 hw_level, hw_level + iview->level_count - 1,
1085 iview->base_layer,
1086 iview->base_layer + iview->layer_count - 1,
1087 iview->extent.width,
1088 iview->extent.height,
1089 iview->extent.depth,
1090 descriptor,
1091 descriptor + 8);
1092
1093 const struct legacy_surf_level *base_level_info = NULL;
1094 if (device->physical_device->rad_info.chip_class <= GFX9) {
1095 if (is_stencil)
1096 base_level_info = &image->surface.u.legacy.stencil_level[iview->base_mip];
1097 else
1098 base_level_info = &image->surface.u.legacy.level[iview->base_mip];
1099 }
1100 si_set_mutable_tex_desc_fields(device, image,
1101 base_level_info,
1102 iview->base_mip,
1103 iview->base_mip,
1104 blk_w, is_stencil, is_storage_image, descriptor);
1105 }
1106
1107 void
1108 radv_image_view_init(struct radv_image_view *iview,
1109 struct radv_device *device,
1110 const VkImageViewCreateInfo* pCreateInfo)
1111 {
1112 RADV_FROM_HANDLE(radv_image, image, pCreateInfo->image);
1113 const VkImageSubresourceRange *range = &pCreateInfo->subresourceRange;
1114
1115 switch (image->type) {
1116 case VK_IMAGE_TYPE_1D:
1117 case VK_IMAGE_TYPE_2D:
1118 assert(range->baseArrayLayer + radv_get_layerCount(image, range) - 1 <= image->info.array_size);
1119 break;
1120 case VK_IMAGE_TYPE_3D:
1121 assert(range->baseArrayLayer + radv_get_layerCount(image, range) - 1
1122 <= radv_minify(image->info.depth, range->baseMipLevel));
1123 break;
1124 default:
1125 unreachable("bad VkImageType");
1126 }
1127 iview->image = image;
1128 iview->bo = image->bo;
1129 iview->type = pCreateInfo->viewType;
1130 iview->vk_format = pCreateInfo->format;
1131 iview->aspect_mask = pCreateInfo->subresourceRange.aspectMask;
1132
1133 if (iview->aspect_mask == VK_IMAGE_ASPECT_STENCIL_BIT) {
1134 iview->vk_format = vk_format_stencil_only(iview->vk_format);
1135 } else if (iview->aspect_mask == VK_IMAGE_ASPECT_DEPTH_BIT) {
1136 iview->vk_format = vk_format_depth_only(iview->vk_format);
1137 }
1138
1139 if (device->physical_device->rad_info.chip_class >= GFX9) {
1140 iview->extent = (VkExtent3D) {
1141 .width = image->info.width,
1142 .height = image->info.height,
1143 .depth = image->info.depth,
1144 };
1145 } else {
1146 iview->extent = (VkExtent3D) {
1147 .width = radv_minify(image->info.width , range->baseMipLevel),
1148 .height = radv_minify(image->info.height, range->baseMipLevel),
1149 .depth = radv_minify(image->info.depth , range->baseMipLevel),
1150 };
1151 }
1152
1153 if (iview->vk_format != image->vk_format) {
1154 unsigned view_bw = vk_format_get_blockwidth(iview->vk_format);
1155 unsigned view_bh = vk_format_get_blockheight(iview->vk_format);
1156 unsigned img_bw = vk_format_get_blockwidth(image->vk_format);
1157 unsigned img_bh = vk_format_get_blockheight(image->vk_format);
1158
1159 iview->extent.width = round_up_u32(iview->extent.width * view_bw, img_bw);
1160 iview->extent.height = round_up_u32(iview->extent.height * view_bh, img_bh);
1161
1162 /* Comment ported from amdvlk -
1163 * If we have the following image:
1164 * Uncompressed pixels Compressed block sizes (4x4)
1165 * mip0: 22 x 22 6 x 6
1166 * mip1: 11 x 11 3 x 3
1167 * mip2: 5 x 5 2 x 2
1168 * mip3: 2 x 2 1 x 1
1169 * mip4: 1 x 1 1 x 1
1170 *
1171 * On GFX9 the descriptor is always programmed with the WIDTH and HEIGHT of the base level and the HW is
1172 * calculating the degradation of the block sizes down the mip-chain as follows (straight-up
1173 * divide-by-two integer math):
1174 * mip0: 6x6
1175 * mip1: 3x3
1176 * mip2: 1x1
1177 * mip3: 1x1
1178 *
1179 * This means that mip2 will be missing texels.
1180 *
1181 * Fix this by calculating the base mip's width and height, then convert that, and round it
1182 * back up to get the level 0 size.
1183 * Clamp the converted size between the original values, and next power of two, which
1184 * means we don't oversize the image.
1185 */
1186 if (device->physical_device->rad_info.chip_class >= GFX9 &&
1187 vk_format_is_compressed(image->vk_format) &&
1188 !vk_format_is_compressed(iview->vk_format)) {
1189 unsigned rounded_img_w = util_next_power_of_two(iview->extent.width);
1190 unsigned rounded_img_h = util_next_power_of_two(iview->extent.height);
1191 unsigned lvl_width = radv_minify(image->info.width , range->baseMipLevel);
1192 unsigned lvl_height = radv_minify(image->info.height, range->baseMipLevel);
1193
1194 lvl_width = round_up_u32(lvl_width * view_bw, img_bw);
1195 lvl_height = round_up_u32(lvl_height * view_bh, img_bh);
1196
1197 lvl_width <<= range->baseMipLevel;
1198 lvl_height <<= range->baseMipLevel;
1199
1200 iview->extent.width = CLAMP(lvl_width, iview->extent.width, rounded_img_w);
1201 iview->extent.height = CLAMP(lvl_height, iview->extent.height, rounded_img_h);
1202 }
1203 }
1204
1205 iview->base_layer = range->baseArrayLayer;
1206 iview->layer_count = radv_get_layerCount(image, range);
1207 iview->base_mip = range->baseMipLevel;
1208 iview->level_count = radv_get_levelCount(image, range);
1209
1210 radv_image_view_make_descriptor(iview, device, &pCreateInfo->components, false);
1211 radv_image_view_make_descriptor(iview, device, &pCreateInfo->components, true);
1212 }
1213
1214 bool radv_layout_has_htile(const struct radv_image *image,
1215 VkImageLayout layout,
1216 unsigned queue_mask)
1217 {
1218 if (radv_image_is_tc_compat_htile(image))
1219 return layout != VK_IMAGE_LAYOUT_GENERAL;
1220
1221 return radv_image_has_htile(image) &&
1222 (layout == VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL ||
1223 layout == VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL) &&
1224 queue_mask == (1u << RADV_QUEUE_GENERAL);
1225 }
1226
1227 bool radv_layout_is_htile_compressed(const struct radv_image *image,
1228 VkImageLayout layout,
1229 unsigned queue_mask)
1230 {
1231 if (radv_image_is_tc_compat_htile(image))
1232 return layout != VK_IMAGE_LAYOUT_GENERAL;
1233
1234 return radv_image_has_htile(image) &&
1235 (layout == VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL ||
1236 layout == VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL) &&
1237 queue_mask == (1u << RADV_QUEUE_GENERAL);
1238 }
1239
1240 bool radv_layout_can_fast_clear(const struct radv_image *image,
1241 VkImageLayout layout,
1242 unsigned queue_mask)
1243 {
1244 return layout == VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL &&
1245 queue_mask == (1u << RADV_QUEUE_GENERAL);
1246 }
1247
1248 bool radv_layout_dcc_compressed(const struct radv_image *image,
1249 VkImageLayout layout,
1250 unsigned queue_mask)
1251 {
1252 /* Don't compress compute transfer dst, as image stores are not supported. */
1253 if (layout == VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL &&
1254 (queue_mask & (1u << RADV_QUEUE_COMPUTE)))
1255 return false;
1256
1257 return radv_image_has_dcc(image) && layout != VK_IMAGE_LAYOUT_GENERAL;
1258 }
1259
1260
1261 unsigned radv_image_queue_family_mask(const struct radv_image *image, uint32_t family, uint32_t queue_family)
1262 {
1263 if (!image->exclusive)
1264 return image->queue_family_mask;
1265 if (family == VK_QUEUE_FAMILY_EXTERNAL_KHR)
1266 return (1u << RADV_MAX_QUEUE_FAMILIES) - 1u;
1267 if (family == VK_QUEUE_FAMILY_IGNORED)
1268 return 1u << queue_family;
1269 return 1u << family;
1270 }
1271
1272 VkResult
1273 radv_CreateImage(VkDevice device,
1274 const VkImageCreateInfo *pCreateInfo,
1275 const VkAllocationCallbacks *pAllocator,
1276 VkImage *pImage)
1277 {
1278 #ifdef ANDROID
1279 const VkNativeBufferANDROID *gralloc_info =
1280 vk_find_struct_const(pCreateInfo->pNext, NATIVE_BUFFER_ANDROID);
1281
1282 if (gralloc_info)
1283 return radv_image_from_gralloc(device, pCreateInfo, gralloc_info,
1284 pAllocator, pImage);
1285 #endif
1286
1287 const struct wsi_image_create_info *wsi_info =
1288 vk_find_struct_const(pCreateInfo->pNext, WSI_IMAGE_CREATE_INFO_MESA);
1289 bool scanout = wsi_info && wsi_info->scanout;
1290
1291 return radv_image_create(device,
1292 &(struct radv_image_create_info) {
1293 .vk_info = pCreateInfo,
1294 .scanout = scanout,
1295 },
1296 pAllocator,
1297 pImage);
1298 }
1299
1300 void
1301 radv_DestroyImage(VkDevice _device, VkImage _image,
1302 const VkAllocationCallbacks *pAllocator)
1303 {
1304 RADV_FROM_HANDLE(radv_device, device, _device);
1305 RADV_FROM_HANDLE(radv_image, image, _image);
1306
1307 if (!image)
1308 return;
1309
1310 if (image->flags & VK_IMAGE_CREATE_SPARSE_BINDING_BIT)
1311 device->ws->buffer_destroy(image->bo);
1312
1313 if (image->owned_memory != VK_NULL_HANDLE)
1314 radv_FreeMemory(_device, image->owned_memory, pAllocator);
1315
1316 vk_free2(&device->alloc, pAllocator, image);
1317 }
1318
1319 void radv_GetImageSubresourceLayout(
1320 VkDevice _device,
1321 VkImage _image,
1322 const VkImageSubresource* pSubresource,
1323 VkSubresourceLayout* pLayout)
1324 {
1325 RADV_FROM_HANDLE(radv_image, image, _image);
1326 RADV_FROM_HANDLE(radv_device, device, _device);
1327 int level = pSubresource->mipLevel;
1328 int layer = pSubresource->arrayLayer;
1329 struct radeon_surf *surface = &image->surface;
1330
1331 if (device->physical_device->rad_info.chip_class >= GFX9) {
1332 pLayout->offset = surface->u.gfx9.offset[level] + surface->u.gfx9.surf_slice_size * layer;
1333 pLayout->rowPitch = surface->u.gfx9.surf_pitch * surface->bpe;
1334 pLayout->arrayPitch = surface->u.gfx9.surf_slice_size;
1335 pLayout->depthPitch = surface->u.gfx9.surf_slice_size;
1336 pLayout->size = surface->u.gfx9.surf_slice_size;
1337 if (image->type == VK_IMAGE_TYPE_3D)
1338 pLayout->size *= u_minify(image->info.depth, level);
1339 } else {
1340 pLayout->offset = surface->u.legacy.level[level].offset + (uint64_t)surface->u.legacy.level[level].slice_size_dw * 4 * layer;
1341 pLayout->rowPitch = surface->u.legacy.level[level].nblk_x * surface->bpe;
1342 pLayout->arrayPitch = (uint64_t)surface->u.legacy.level[level].slice_size_dw * 4;
1343 pLayout->depthPitch = (uint64_t)surface->u.legacy.level[level].slice_size_dw * 4;
1344 pLayout->size = (uint64_t)surface->u.legacy.level[level].slice_size_dw * 4;
1345 if (image->type == VK_IMAGE_TYPE_3D)
1346 pLayout->size *= u_minify(image->info.depth, level);
1347 }
1348 }
1349
1350
1351 VkResult
1352 radv_CreateImageView(VkDevice _device,
1353 const VkImageViewCreateInfo *pCreateInfo,
1354 const VkAllocationCallbacks *pAllocator,
1355 VkImageView *pView)
1356 {
1357 RADV_FROM_HANDLE(radv_device, device, _device);
1358 struct radv_image_view *view;
1359
1360 view = vk_alloc2(&device->alloc, pAllocator, sizeof(*view), 8,
1361 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1362 if (view == NULL)
1363 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1364
1365 radv_image_view_init(view, device, pCreateInfo);
1366
1367 *pView = radv_image_view_to_handle(view);
1368
1369 return VK_SUCCESS;
1370 }
1371
1372 void
1373 radv_DestroyImageView(VkDevice _device, VkImageView _iview,
1374 const VkAllocationCallbacks *pAllocator)
1375 {
1376 RADV_FROM_HANDLE(radv_device, device, _device);
1377 RADV_FROM_HANDLE(radv_image_view, iview, _iview);
1378
1379 if (!iview)
1380 return;
1381 vk_free2(&device->alloc, pAllocator, iview);
1382 }
1383
1384 void radv_buffer_view_init(struct radv_buffer_view *view,
1385 struct radv_device *device,
1386 const VkBufferViewCreateInfo* pCreateInfo)
1387 {
1388 RADV_FROM_HANDLE(radv_buffer, buffer, pCreateInfo->buffer);
1389
1390 view->bo = buffer->bo;
1391 view->range = pCreateInfo->range == VK_WHOLE_SIZE ?
1392 buffer->size - pCreateInfo->offset : pCreateInfo->range;
1393 view->vk_format = pCreateInfo->format;
1394
1395 radv_make_buffer_descriptor(device, buffer, view->vk_format,
1396 pCreateInfo->offset, view->range, view->state);
1397 }
1398
1399 VkResult
1400 radv_CreateBufferView(VkDevice _device,
1401 const VkBufferViewCreateInfo *pCreateInfo,
1402 const VkAllocationCallbacks *pAllocator,
1403 VkBufferView *pView)
1404 {
1405 RADV_FROM_HANDLE(radv_device, device, _device);
1406 struct radv_buffer_view *view;
1407
1408 view = vk_alloc2(&device->alloc, pAllocator, sizeof(*view), 8,
1409 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1410 if (!view)
1411 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1412
1413 radv_buffer_view_init(view, device, pCreateInfo);
1414
1415 *pView = radv_buffer_view_to_handle(view);
1416
1417 return VK_SUCCESS;
1418 }
1419
1420 void
1421 radv_DestroyBufferView(VkDevice _device, VkBufferView bufferView,
1422 const VkAllocationCallbacks *pAllocator)
1423 {
1424 RADV_FROM_HANDLE(radv_device, device, _device);
1425 RADV_FROM_HANDLE(radv_buffer_view, view, bufferView);
1426
1427 if (!view)
1428 return;
1429
1430 vk_free2(&device->alloc, pAllocator, view);
1431 }