projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radv: pass FMASK alignment to application
[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_private.h"
29 #include "vk_format.h"
30 #include "radv_radeon_winsys.h"
31 #include "sid.h"
32 #include "util/debug.h"
33 static unsigned
34 radv_choose_tiling(struct radv_device *Device,
35 const struct radv_image_create_info *create_info)
36 {
37 const VkImageCreateInfo *pCreateInfo = create_info->vk_info;
38
39 if (pCreateInfo->tiling == VK_IMAGE_TILING_LINEAR) {
40 assert(pCreateInfo->samples <= 1);
41 return RADEON_SURF_MODE_LINEAR_ALIGNED;
42 }
43
44 /* MSAA resources must be 2D tiled. */
45 if (pCreateInfo->samples > 1)
46 return RADEON_SURF_MODE_2D;
47
48 return RADEON_SURF_MODE_2D;
49 }
50 static int
51 radv_init_surface(struct radv_device *device,
52 struct radeon_surf *surface,
53 const struct radv_image_create_info *create_info)
54 {
55 const VkImageCreateInfo *pCreateInfo = create_info->vk_info;
56 unsigned array_mode = radv_choose_tiling(device, create_info);
57 const struct vk_format_description *desc =
58 vk_format_description(pCreateInfo->format);
59 bool is_depth, is_stencil, blendable;
60
61 is_depth = vk_format_has_depth(desc);
62 is_stencil = vk_format_has_stencil(desc);
63 surface->npix_x = pCreateInfo->extent.width;
64 surface->npix_y = pCreateInfo->extent.height;
65 surface->npix_z = pCreateInfo->extent.depth;
66
67 surface->blk_w = vk_format_get_blockwidth(pCreateInfo->format);
68 surface->blk_h = vk_format_get_blockheight(pCreateInfo->format);
69 surface->blk_d = 1;
70 surface->array_size = pCreateInfo->arrayLayers;
71 surface->last_level = pCreateInfo->mipLevels - 1;
72
73 surface->bpe = vk_format_get_blocksize(pCreateInfo->format);
74 /* align byte per element on dword */
75 if (surface->bpe == 3) {
76 surface->bpe = 4;
77 }
78 surface->nsamples = pCreateInfo->samples ? pCreateInfo->samples : 1;
79 surface->flags = RADEON_SURF_SET(array_mode, MODE);
80
81 switch (pCreateInfo->imageType){
82 case VK_IMAGE_TYPE_1D:
83 if (pCreateInfo->arrayLayers > 1)
84 surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_1D_ARRAY, TYPE);
85 else
86 surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_1D, TYPE);
87 break;
88 case VK_IMAGE_TYPE_2D:
89 if (pCreateInfo->arrayLayers > 1)
90 surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_2D_ARRAY, TYPE);
91 else
92 surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_2D, TYPE);
93 break;
94 case VK_IMAGE_TYPE_3D:
95 surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_3D, TYPE);
96 break;
97 default:
98 unreachable("unhandled image type");
99 }
100
101 if (is_depth) {
102 surface->flags |= RADEON_SURF_ZBUFFER;
103 }
104
105 if (is_stencil)
106 surface->flags |= RADEON_SURF_SBUFFER |
107 RADEON_SURF_HAS_SBUFFER_MIPTREE;
108
109 surface->flags |= RADEON_SURF_HAS_TILE_MODE_INDEX;
110
111 if ((pCreateInfo->usage & (VK_IMAGE_USAGE_TRANSFER_SRC_BIT |
112 VK_IMAGE_USAGE_STORAGE_BIT)) ||
113 (pCreateInfo->flags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) ||
114 (pCreateInfo->tiling == VK_IMAGE_TILING_LINEAR) ||
115 device->physical_device->rad_info.chip_class < VI ||
116 create_info->scanout || (device->debug_flags & RADV_DEBUG_NO_DCC) ||
117 !radv_is_colorbuffer_format_supported(pCreateInfo->format, &blendable))
118 surface->flags |= RADEON_SURF_DISABLE_DCC;
119 if (create_info->scanout)
120 surface->flags |= RADEON_SURF_SCANOUT;
121 return 0;
122 }
123 #define ATI_VENDOR_ID 0x1002
124 static uint32_t si_get_bo_metadata_word1(struct radv_device *device)
125 {
126 return (ATI_VENDOR_ID << 16) | device->physical_device->rad_info.pci_id;
127 }
128
129 static inline unsigned
130 si_tile_mode_index(const struct radv_image *image, unsigned level, bool stencil)
131 {
132 if (stencil)
133 return image->surface.stencil_tiling_index[level];
134 else
135 return image->surface.tiling_index[level];
136 }
137
138 static unsigned radv_map_swizzle(unsigned swizzle)
139 {
140 switch (swizzle) {
141 case VK_SWIZZLE_Y:
142 return V_008F0C_SQ_SEL_Y;
143 case VK_SWIZZLE_Z:
144 return V_008F0C_SQ_SEL_Z;
145 case VK_SWIZZLE_W:
146 return V_008F0C_SQ_SEL_W;
147 case VK_SWIZZLE_0:
148 return V_008F0C_SQ_SEL_0;
149 case VK_SWIZZLE_1:
150 return V_008F0C_SQ_SEL_1;
151 default: /* VK_SWIZZLE_X */
152 return V_008F0C_SQ_SEL_X;
153 }
154 }
155
156 static void
157 radv_make_buffer_descriptor(struct radv_device *device,
158 struct radv_buffer *buffer,
159 VkFormat vk_format,
160 unsigned offset,
161 unsigned range,
162 uint32_t *state)
163 {
164 const struct vk_format_description *desc;
165 unsigned stride;
166 uint64_t gpu_address = device->ws->buffer_get_va(buffer->bo);
167 uint64_t va = gpu_address + buffer->offset;
168 unsigned num_format, data_format;
169 int first_non_void;
170 desc = vk_format_description(vk_format);
171 first_non_void = vk_format_get_first_non_void_channel(vk_format);
172 stride = desc->block.bits / 8;
173
174 num_format = radv_translate_buffer_numformat(desc, first_non_void);
175 data_format = radv_translate_buffer_dataformat(desc, first_non_void);
176
177 va += offset;
178 state[0] = va;
179 state[1] = S_008F04_BASE_ADDRESS_HI(va >> 32) |
180 S_008F04_STRIDE(stride);
181 state[2] = range;
182 state[3] = S_008F0C_DST_SEL_X(radv_map_swizzle(desc->swizzle[0])) |
183 S_008F0C_DST_SEL_Y(radv_map_swizzle(desc->swizzle[1])) |
184 S_008F0C_DST_SEL_Z(radv_map_swizzle(desc->swizzle[2])) |
185 S_008F0C_DST_SEL_W(radv_map_swizzle(desc->swizzle[3])) |
186 S_008F0C_NUM_FORMAT(num_format) |
187 S_008F0C_DATA_FORMAT(data_format);
188 }
189
190 static void
191 si_set_mutable_tex_desc_fields(struct radv_device *device,
192 struct radv_image *image,
193 const struct radeon_surf_level *base_level_info,
194 unsigned base_level, unsigned first_level,
195 unsigned block_width, bool is_stencil,
196 uint32_t *state)
197 {
198 uint64_t gpu_address = device->ws->buffer_get_va(image->bo) + image->offset;
199 uint64_t va = gpu_address + base_level_info->offset;
200 unsigned pitch = base_level_info->nblk_x * block_width;
201
202 state[1] &= C_008F14_BASE_ADDRESS_HI;
203 state[3] &= C_008F1C_TILING_INDEX;
204 state[4] &= C_008F20_PITCH;
205 state[6] &= C_008F28_COMPRESSION_EN;
206
207 assert(!(va & 255));
208
209 state[0] = va >> 8;
210 state[1] |= S_008F14_BASE_ADDRESS_HI(va >> 40);
211 state[3] |= S_008F1C_TILING_INDEX(si_tile_mode_index(image, base_level,
212 is_stencil));
213 state[4] |= S_008F20_PITCH(pitch - 1);
214
215 if (image->surface.dcc_size && image->surface.level[first_level].dcc_enabled) {
216 state[6] |= S_008F28_COMPRESSION_EN(1);
217 state[7] = (gpu_address +
218 image->dcc_offset +
219 base_level_info->dcc_offset) >> 8;
220 }
221 }
222
223 static unsigned radv_tex_dim(VkImageType image_type, VkImageViewType view_type,
224 unsigned nr_layers, unsigned nr_samples, bool is_storage_image)
225 {
226 if (view_type == VK_IMAGE_VIEW_TYPE_CUBE || view_type == VK_IMAGE_VIEW_TYPE_CUBE_ARRAY)
227 return is_storage_image ? V_008F1C_SQ_RSRC_IMG_2D_ARRAY : V_008F1C_SQ_RSRC_IMG_CUBE;
228 switch (image_type) {
229 case VK_IMAGE_TYPE_1D:
230 return nr_layers > 1 ? V_008F1C_SQ_RSRC_IMG_1D_ARRAY : V_008F1C_SQ_RSRC_IMG_1D;
231 case VK_IMAGE_TYPE_2D:
232 if (nr_samples > 1)
233 return nr_layers > 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY : V_008F1C_SQ_RSRC_IMG_2D_MSAA;
234 else
235 return nr_layers > 1 ? V_008F1C_SQ_RSRC_IMG_2D_ARRAY : V_008F1C_SQ_RSRC_IMG_2D;
236 case VK_IMAGE_TYPE_3D:
237 if (view_type == VK_IMAGE_VIEW_TYPE_3D)
238 return V_008F1C_SQ_RSRC_IMG_3D;
239 else
240 return V_008F1C_SQ_RSRC_IMG_2D_ARRAY;
241 default:
242 unreachable("illegale image type");
243 }
244 }
245 /**
246 * Build the sampler view descriptor for a texture.
247 */
248 static void
249 si_make_texture_descriptor(struct radv_device *device,
250 struct radv_image *image,
251 bool sampler,
252 VkImageViewType view_type,
253 VkFormat vk_format,
254 const VkComponentMapping *mapping,
255 unsigned first_level, unsigned last_level,
256 unsigned first_layer, unsigned last_layer,
257 unsigned width, unsigned height, unsigned depth,
258 uint32_t *state,
259 uint32_t *fmask_state)
260 {
261 const struct vk_format_description *desc;
262 enum vk_swizzle swizzle[4];
263 int first_non_void;
264 unsigned num_format, data_format, type;
265
266 desc = vk_format_description(vk_format);
267
268 if (desc->colorspace == VK_FORMAT_COLORSPACE_ZS) {
269 const unsigned char swizzle_xxxx[4] = {0, 0, 0, 0};
270 vk_format_compose_swizzles(mapping, swizzle_xxxx, swizzle);
271 } else {
272 vk_format_compose_swizzles(mapping, desc->swizzle, swizzle);
273 }
274
275 first_non_void = vk_format_get_first_non_void_channel(vk_format);
276
277 num_format = radv_translate_tex_numformat(vk_format, desc, first_non_void);
278 if (num_format == ~0) {
279 num_format = 0;
280 }
281
282 data_format = radv_translate_tex_dataformat(vk_format, desc, first_non_void);
283 if (data_format == ~0) {
284 data_format = 0;
285 }
286
287 type = radv_tex_dim(image->type, view_type, image->array_size, image->samples,
288 (image->usage & VK_IMAGE_USAGE_STORAGE_BIT));
289 if (type == V_008F1C_SQ_RSRC_IMG_1D_ARRAY) {
290 height = 1;
291 depth = image->array_size;
292 } else if (type == V_008F1C_SQ_RSRC_IMG_2D_ARRAY ||
293 type == V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY) {
294 if (view_type != VK_IMAGE_VIEW_TYPE_3D)
295 depth = image->array_size;
296 } else if (type == V_008F1C_SQ_RSRC_IMG_CUBE)
297 depth = image->array_size / 6;
298
299 state[0] = 0;
300 state[1] = (S_008F14_DATA_FORMAT(data_format) |
301 S_008F14_NUM_FORMAT(num_format));
302 state[2] = (S_008F18_WIDTH(width - 1) |
303 S_008F18_HEIGHT(height - 1));
304 state[3] = (S_008F1C_DST_SEL_X(radv_map_swizzle(swizzle[0])) |
305 S_008F1C_DST_SEL_Y(radv_map_swizzle(swizzle[1])) |
306 S_008F1C_DST_SEL_Z(radv_map_swizzle(swizzle[2])) |
307 S_008F1C_DST_SEL_W(radv_map_swizzle(swizzle[3])) |
308 S_008F1C_BASE_LEVEL(image->samples > 1 ?
309 0 : first_level) |
310 S_008F1C_LAST_LEVEL(image->samples > 1 ?
311 util_logbase2(image->samples) :
312 last_level) |
313 S_008F1C_POW2_PAD(image->levels > 1) |
314 S_008F1C_TYPE(type));
315 state[4] = S_008F20_DEPTH(depth - 1);
316 state[5] = (S_008F24_BASE_ARRAY(first_layer) |
317 S_008F24_LAST_ARRAY(last_layer));
318 state[6] = 0;
319 state[7] = 0;
320
321 if (image->dcc_offset) {
322 unsigned swap = radv_translate_colorswap(vk_format, FALSE);
323
324 state[6] = S_008F28_ALPHA_IS_ON_MSB(swap <= 1);
325 } else {
326 /* The last dword is unused by hw. The shader uses it to clear
327 * bits in the first dword of sampler state.
328 */
329 if (device->physical_device->rad_info.chip_class <= CIK && image->samples <= 1) {
330 if (first_level == last_level)
331 state[7] = C_008F30_MAX_ANISO_RATIO;
332 else
333 state[7] = 0xffffffff;
334 }
335 }
336
337 /* Initialize the sampler view for FMASK. */
338 if (image->fmask.size) {
339 uint32_t fmask_format;
340 uint64_t gpu_address = device->ws->buffer_get_va(image->bo);
341 uint64_t va;
342
343 va = gpu_address + image->offset + image->fmask.offset;
344
345 switch (image->samples) {
346 case 2:
347 fmask_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S2_F2;
348 break;
349 case 4:
350 fmask_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S4_F4;
351 break;
352 case 8:
353 fmask_format = V_008F14_IMG_DATA_FORMAT_FMASK32_S8_F8;
354 break;
355 default:
356 assert(0);
357 fmask_format = V_008F14_IMG_DATA_FORMAT_INVALID;
358 }
359
360 fmask_state[0] = va >> 8;
361 fmask_state[1] = S_008F14_BASE_ADDRESS_HI(va >> 40) |
362 S_008F14_DATA_FORMAT(fmask_format) |
363 S_008F14_NUM_FORMAT(V_008F14_IMG_NUM_FORMAT_UINT);
364 fmask_state[2] = S_008F18_WIDTH(width - 1) |
365 S_008F18_HEIGHT(height - 1);
366 fmask_state[3] = S_008F1C_DST_SEL_X(V_008F1C_SQ_SEL_X) |
367 S_008F1C_DST_SEL_Y(V_008F1C_SQ_SEL_X) |
368 S_008F1C_DST_SEL_Z(V_008F1C_SQ_SEL_X) |
369 S_008F1C_DST_SEL_W(V_008F1C_SQ_SEL_X) |
370 S_008F1C_TILING_INDEX(image->fmask.tile_mode_index) |
371 S_008F1C_TYPE(radv_tex_dim(image->type, view_type, 1, 0, false));
372 fmask_state[4] = S_008F20_DEPTH(depth - 1) |
373 S_008F20_PITCH(image->fmask.pitch_in_pixels - 1);
374 fmask_state[5] = S_008F24_BASE_ARRAY(first_layer) |
375 S_008F24_LAST_ARRAY(last_layer);
376 fmask_state[6] = 0;
377 fmask_state[7] = 0;
378 }
379 }
380
381 static void
382 radv_query_opaque_metadata(struct radv_device *device,
383 struct radv_image *image,
384 struct radeon_bo_metadata *md)
385 {
386 static const VkComponentMapping fixedmapping;
387 uint32_t desc[8], i;
388
389 /* Metadata image format format version 1:
390 * [0] = 1 (metadata format identifier)
391 * [1] = (VENDOR_ID << 16) | PCI_ID
392 * [2:9] = image descriptor for the whole resource
393 * [2] is always 0, because the base address is cleared
394 * [9] is the DCC offset bits [39:8] from the beginning of
395 * the buffer
396 * [10:10+LAST_LEVEL] = mipmap level offset bits [39:8] for each level
397 */
398 md->metadata[0] = 1; /* metadata image format version 1 */
399
400 /* TILE_MODE_INDEX is ambiguous without a PCI ID. */
401 md->metadata[1] = si_get_bo_metadata_word1(device);
402
403
404 si_make_texture_descriptor(device, image, true,
405 (VkImageViewType)image->type, image->vk_format,
406 &fixedmapping, 0, image->levels - 1, 0,
407 image->array_size,
408 image->extent.width, image->extent.height,
409 image->extent.depth,
410 desc, NULL);
411
412 si_set_mutable_tex_desc_fields(device, image, &image->surface.level[0], 0, 0,
413 image->surface.blk_w, false, desc);
414
415 /* Clear the base address and set the relative DCC offset. */
416 desc[0] = 0;
417 desc[1] &= C_008F14_BASE_ADDRESS_HI;
418 desc[7] = image->dcc_offset >> 8;
419
420 /* Dwords [2:9] contain the image descriptor. */
421 memcpy(&md->metadata[2], desc, sizeof(desc));
422
423 /* Dwords [10:..] contain the mipmap level offsets. */
424 for (i = 0; i <= image->levels - 1; i++)
425 md->metadata[10+i] = image->surface.level[i].offset >> 8;
426
427 md->size_metadata = (11 + image->levels - 1) * 4;
428 }
429
430 void
431 radv_init_metadata(struct radv_device *device,
432 struct radv_image *image,
433 struct radeon_bo_metadata *metadata)
434 {
435 struct radeon_surf *surface = &image->surface;
436
437 memset(metadata, 0, sizeof(*metadata));
438 metadata->microtile = surface->level[0].mode >= RADEON_SURF_MODE_1D ?
439 RADEON_LAYOUT_TILED : RADEON_LAYOUT_LINEAR;
440 metadata->macrotile = surface->level[0].mode >= RADEON_SURF_MODE_2D ?
441 RADEON_LAYOUT_TILED : RADEON_LAYOUT_LINEAR;
442 metadata->pipe_config = surface->pipe_config;
443 metadata->bankw = surface->bankw;
444 metadata->bankh = surface->bankh;
445 metadata->tile_split = surface->tile_split;
446 metadata->mtilea = surface->mtilea;
447 metadata->num_banks = surface->num_banks;
448 metadata->stride = surface->level[0].pitch_bytes;
449 metadata->scanout = (surface->flags & RADEON_SURF_SCANOUT) != 0;
450
451 radv_query_opaque_metadata(device, image, metadata);
452 }
453
454 /* The number of samples can be specified independently of the texture. */
455 static void
456 radv_image_get_fmask_info(struct radv_device *device,
457 struct radv_image *image,
458 unsigned nr_samples,
459 struct radv_fmask_info *out)
460 {
461 /* FMASK is allocated like an ordinary texture. */
462 struct radeon_surf fmask = image->surface;
463
464 memset(out, 0, sizeof(*out));
465
466 fmask.bo_alignment = 0;
467 fmask.bo_size = 0;
468 fmask.nsamples = 1;
469 fmask.flags |= RADEON_SURF_FMASK;
470
471 /* Force 2D tiling if it wasn't set. This may occur when creating
472 * FMASK for MSAA resolve on R6xx. On R6xx, the single-sample
473 * destination buffer must have an FMASK too. */
474 fmask.flags = RADEON_SURF_CLR(fmask.flags, MODE);
475 fmask.flags |= RADEON_SURF_SET(RADEON_SURF_MODE_2D, MODE);
476
477 fmask.flags |= RADEON_SURF_HAS_TILE_MODE_INDEX;
478
479 switch (nr_samples) {
480 case 2:
481 case 4:
482 fmask.bpe = 1;
483 break;
484 case 8:
485 fmask.bpe = 4;
486 break;
487 default:
488 return;
489 }
490
491 device->ws->surface_init(device->ws, &fmask);
492 assert(fmask.level[0].mode == RADEON_SURF_MODE_2D);
493
494 out->slice_tile_max = (fmask.level[0].nblk_x * fmask.level[0].nblk_y) / 64;
495 if (out->slice_tile_max)
496 out->slice_tile_max -= 1;
497
498 out->tile_mode_index = fmask.tiling_index[0];
499 out->pitch_in_pixels = fmask.level[0].nblk_x;
500 out->bank_height = fmask.bankh;
501 out->alignment = MAX2(256, fmask.bo_alignment);
502 out->size = fmask.bo_size;
503 }
504
505 static void
506 radv_image_alloc_fmask(struct radv_device *device,
507 struct radv_image *image)
508 {
509 radv_image_get_fmask_info(device, image, image->samples, &image->fmask);
510
511 image->fmask.offset = align64(image->size, image->fmask.alignment);
512 image->size = image->fmask.offset + image->fmask.size;
513 image->alignment = MAX2(image->alignment, image->fmask.alignment);
514 }
515
516 static void
517 radv_image_get_cmask_info(struct radv_device *device,
518 struct radv_image *image,
519 struct radv_cmask_info *out)
520 {
521 unsigned pipe_interleave_bytes = device->physical_device->rad_info.pipe_interleave_bytes;
522 unsigned num_pipes = device->physical_device->rad_info.num_tile_pipes;
523 unsigned cl_width, cl_height;
524
525 switch (num_pipes) {
526 case 2:
527 cl_width = 32;
528 cl_height = 16;
529 break;
530 case 4:
531 cl_width = 32;
532 cl_height = 32;
533 break;
534 case 8:
535 cl_width = 64;
536 cl_height = 32;
537 break;
538 case 16: /* Hawaii */
539 cl_width = 64;
540 cl_height = 64;
541 break;
542 default:
543 assert(0);
544 return;
545 }
546
547 unsigned base_align = num_pipes * pipe_interleave_bytes;
548
549 unsigned width = align(image->surface.npix_x, cl_width*8);
550 unsigned height = align(image->surface.npix_y, cl_height*8);
551 unsigned slice_elements = (width * height) / (8*8);
552
553 /* Each element of CMASK is a nibble. */
554 unsigned slice_bytes = slice_elements / 2;
555
556 out->slice_tile_max = (width * height) / (128*128);
557 if (out->slice_tile_max)
558 out->slice_tile_max -= 1;
559
560 out->alignment = MAX2(256, base_align);
561 out->size = (image->type == VK_IMAGE_TYPE_3D ? image->extent.depth : image->array_size) *
562 align(slice_bytes, base_align);
563 }
564
565 static void
566 radv_image_alloc_cmask(struct radv_device *device,
567 struct radv_image *image)
568 {
569 radv_image_get_cmask_info(device, image, &image->cmask);
570
571 image->cmask.offset = align64(image->size, image->cmask.alignment);
572 /* + 8 for storing the clear values */
573 image->clear_value_offset = image->cmask.offset + image->cmask.size;
574 image->size = image->cmask.offset + image->cmask.size + 8;
575 image->alignment = MAX2(image->alignment, image->cmask.alignment);
576 }
577
578 static void
579 radv_image_alloc_dcc(struct radv_device *device,
580 struct radv_image *image)
581 {
582 image->dcc_offset = align64(image->size, image->surface.dcc_alignment);
583 /* + 8 for storing the clear values */
584 image->clear_value_offset = image->dcc_offset + image->surface.dcc_size;
585 image->size = image->dcc_offset + image->surface.dcc_size + 8;
586 image->alignment = MAX2(image->alignment, image->surface.dcc_alignment);
587 }
588
589 static unsigned
590 radv_image_get_htile_size(struct radv_device *device,
591 struct radv_image *image)
592 {
593 unsigned cl_width, cl_height, width, height;
594 unsigned slice_elements, slice_bytes, base_align;
595 unsigned num_pipes = device->physical_device->rad_info.num_tile_pipes;
596 unsigned pipe_interleave_bytes = device->physical_device->rad_info.pipe_interleave_bytes;
597
598 /* Overalign HTILE on P2 configs to work around GPU hangs in
599 * piglit/depthstencil-render-miplevels 585.
600 *
601 * This has been confirmed to help Kabini & Stoney, where the hangs
602 * are always reproducible. I think I have seen the test hang
603 * on Carrizo too, though it was very rare there.
604 */
605 if (device->physical_device->rad_info.chip_class >= CIK && num_pipes < 4)
606 num_pipes = 4;
607
608 switch (num_pipes) {
609 case 1:
610 cl_width = 32;
611 cl_height = 16;
612 break;
613 case 2:
614 cl_width = 32;
615 cl_height = 32;
616 break;
617 case 4:
618 cl_width = 64;
619 cl_height = 32;
620 break;
621 case 8:
622 cl_width = 64;
623 cl_height = 64;
624 break;
625 case 16:
626 cl_width = 128;
627 cl_height = 64;
628 break;
629 default:
630 assert(0);
631 return 0;
632 }
633
634 width = align(image->surface.npix_x, cl_width * 8);
635 height = align(image->surface.npix_y, cl_height * 8);
636
637 slice_elements = (width * height) / (8 * 8);
638 slice_bytes = slice_elements * 4;
639
640 base_align = num_pipes * pipe_interleave_bytes;
641
642 image->htile.pitch = width;
643 image->htile.height = height;
644 image->htile.xalign = cl_width * 8;
645 image->htile.yalign = cl_height * 8;
646
647 return image->array_size *
648 align(slice_bytes, base_align);
649 }
650
651 static void
652 radv_image_alloc_htile(struct radv_device *device,
653 struct radv_image *image)
654 {
655 if (device->debug_flags & RADV_DEBUG_NO_HIZ)
656 return;
657
658 image->htile.size = radv_image_get_htile_size(device, image);
659
660 if (!image->htile.size)
661 return;
662
663 image->htile.offset = align64(image->size, 32768);
664
665 /* + 8 for storing the clear values */
666 image->clear_value_offset = image->htile.offset + image->htile.size;
667 image->size = image->htile.offset + image->htile.size + 8;
668 image->alignment = align64(image->alignment, 32768);
669 }
670
671 VkResult
672 radv_image_create(VkDevice _device,
673 const struct radv_image_create_info *create_info,
674 const VkAllocationCallbacks* alloc,
675 VkImage *pImage)
676 {
677 RADV_FROM_HANDLE(radv_device, device, _device);
678 const VkImageCreateInfo *pCreateInfo = create_info->vk_info;
679 struct radv_image *image = NULL;
680 bool can_cmask_dcc = false;
681 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO);
682
683 radv_assert(pCreateInfo->mipLevels > 0);
684 radv_assert(pCreateInfo->arrayLayers > 0);
685 radv_assert(pCreateInfo->samples > 0);
686 radv_assert(pCreateInfo->extent.width > 0);
687 radv_assert(pCreateInfo->extent.height > 0);
688 radv_assert(pCreateInfo->extent.depth > 0);
689
690 image = vk_alloc2(&device->alloc, alloc, sizeof(*image), 8,
691 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
692 if (!image)
693 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
694
695 memset(image, 0, sizeof(*image));
696 image->type = pCreateInfo->imageType;
697 image->extent = pCreateInfo->extent;
698 image->vk_format = pCreateInfo->format;
699 image->levels = pCreateInfo->mipLevels;
700 image->array_size = pCreateInfo->arrayLayers;
701 image->samples = pCreateInfo->samples;
702 image->tiling = pCreateInfo->tiling;
703 image->usage = pCreateInfo->usage;
704
705 image->exclusive = pCreateInfo->sharingMode == VK_SHARING_MODE_EXCLUSIVE;
706 if (pCreateInfo->sharingMode == VK_SHARING_MODE_CONCURRENT) {
707 for (uint32_t i = 0; i < pCreateInfo->queueFamilyIndexCount; ++i)
708 image->queue_family_mask |= 1u << pCreateInfo->pQueueFamilyIndices[i];
709 }
710
711 radv_init_surface(device, &image->surface, create_info);
712
713 device->ws->surface_init(device->ws, &image->surface);
714
715 image->size = image->surface.bo_size;
716 image->alignment = image->surface.bo_alignment;
717
718 if (image->exclusive || image->queue_family_mask == 1)
719 can_cmask_dcc = true;
720
721 if ((pCreateInfo->usage & VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT) &&
722 image->surface.dcc_size && can_cmask_dcc)
723 radv_image_alloc_dcc(device, image);
724 else
725 image->surface.dcc_size = 0;
726
727 if ((pCreateInfo->usage & VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT) &&
728 pCreateInfo->mipLevels == 1 &&
729 !image->surface.dcc_size && image->extent.depth == 1 && can_cmask_dcc)
730 radv_image_alloc_cmask(device, image);
731 if (image->samples > 1 && vk_format_is_color(pCreateInfo->format)) {
732 radv_image_alloc_fmask(device, image);
733 } else if (vk_format_is_depth(pCreateInfo->format)) {
734
735 radv_image_alloc_htile(device, image);
736 }
737
738
739 if (create_info->stride && create_info->stride != image->surface.level[0].pitch_bytes) {
740 image->surface.level[0].nblk_x = create_info->stride / image->surface.bpe;
741 image->surface.level[0].pitch_bytes = create_info->stride;
742 image->surface.level[0].slice_size = create_info->stride * image->surface.level[0].nblk_y;
743 }
744 *pImage = radv_image_to_handle(image);
745
746 return VK_SUCCESS;
747 }
748
749 void
750 radv_image_view_init(struct radv_image_view *iview,
751 struct radv_device *device,
752 const VkImageViewCreateInfo* pCreateInfo,
753 struct radv_cmd_buffer *cmd_buffer,
754 VkImageUsageFlags usage_mask)
755 {
756 RADV_FROM_HANDLE(radv_image, image, pCreateInfo->image);
757 const VkImageSubresourceRange *range = &pCreateInfo->subresourceRange;
758 uint32_t blk_w;
759 bool is_stencil = false;
760 switch (image->type) {
761 case VK_IMAGE_TYPE_1D:
762 case VK_IMAGE_TYPE_2D:
763 assert(range->baseArrayLayer + radv_get_layerCount(image, range) - 1 <= image->array_size);
764 break;
765 case VK_IMAGE_TYPE_3D:
766 assert(range->baseArrayLayer + radv_get_layerCount(image, range) - 1
767 <= radv_minify(image->extent.depth, range->baseMipLevel));
768 break;
769 default:
770 unreachable("bad VkImageType");
771 }
772 iview->image = image;
773 iview->bo = image->bo;
774 iview->type = pCreateInfo->viewType;
775 iview->vk_format = pCreateInfo->format;
776 iview->aspect_mask = pCreateInfo->subresourceRange.aspectMask;
777
778 if (iview->aspect_mask == VK_IMAGE_ASPECT_STENCIL_BIT) {
779 is_stencil = true;
780 iview->vk_format = vk_format_stencil_only(iview->vk_format);
781 } else if (iview->aspect_mask == VK_IMAGE_ASPECT_DEPTH_BIT) {
782 iview->vk_format = vk_format_depth_only(iview->vk_format);
783 }
784
785 iview->extent = (VkExtent3D) {
786 .width = radv_minify(image->extent.width , range->baseMipLevel),
787 .height = radv_minify(image->extent.height, range->baseMipLevel),
788 .depth = radv_minify(image->extent.depth , range->baseMipLevel),
789 };
790
791 iview->extent.width = round_up_u32(iview->extent.width * vk_format_get_blockwidth(iview->vk_format),
792 vk_format_get_blockwidth(image->vk_format));
793 iview->extent.height = round_up_u32(iview->extent.height * vk_format_get_blockheight(iview->vk_format),
794 vk_format_get_blockheight(image->vk_format));
795
796 assert(image->surface.blk_w % vk_format_get_blockwidth(image->vk_format) == 0);
797 blk_w = image->surface.blk_w / vk_format_get_blockwidth(image->vk_format) * vk_format_get_blockwidth(iview->vk_format);
798 iview->base_layer = range->baseArrayLayer;
799 iview->layer_count = radv_get_layerCount(image, range);
800 iview->base_mip = range->baseMipLevel;
801
802 si_make_texture_descriptor(device, image, false,
803 iview->type,
804 iview->vk_format,
805 &pCreateInfo->components,
806 0, radv_get_levelCount(image, range) - 1,
807 range->baseArrayLayer,
808 range->baseArrayLayer + radv_get_layerCount(image, range) - 1,
809 iview->extent.width,
810 iview->extent.height,
811 iview->extent.depth,
812 iview->descriptor,
813 iview->fmask_descriptor);
814 si_set_mutable_tex_desc_fields(device, image,
815 is_stencil ? &image->surface.stencil_level[range->baseMipLevel] : &image->surface.level[range->baseMipLevel], range->baseMipLevel,
816 range->baseMipLevel,
817 blk_w, is_stencil, iview->descriptor);
818 }
819
820 void radv_image_set_optimal_micro_tile_mode(struct radv_device *device,
821 struct radv_image *image, uint32_t micro_tile_mode)
822 {
823 /* These magic numbers were copied from addrlib. It doesn't use any
824 * definitions for them either. They are all 2D_TILED_THIN1 modes with
825 * different bpp and micro tile mode.
826 */
827 if (device->physical_device->rad_info.chip_class >= CIK) {
828 switch (micro_tile_mode) {
829 case 0: /* displayable */
830 image->surface.tiling_index[0] = 10;
831 break;
832 case 1: /* thin */
833 image->surface.tiling_index[0] = 14;
834 break;
835 case 3: /* rotated */
836 image->surface.tiling_index[0] = 28;
837 break;
838 default: /* depth, thick */
839 assert(!"unexpected micro mode");
840 return;
841 }
842 } else { /* SI */
843 switch (micro_tile_mode) {
844 case 0: /* displayable */
845 switch (image->surface.bpe) {
846 case 1:
847 image->surface.tiling_index[0] = 10;
848 break;
849 case 2:
850 image->surface.tiling_index[0] = 11;
851 break;
852 default: /* 4, 8 */
853 image->surface.tiling_index[0] = 12;
854 break;
855 }
856 break;
857 case 1: /* thin */
858 switch (image->surface.bpe) {
859 case 1:
860 image->surface.tiling_index[0] = 14;
861 break;
862 case 2:
863 image->surface.tiling_index[0] = 15;
864 break;
865 case 4:
866 image->surface.tiling_index[0] = 16;
867 break;
868 default: /* 8, 16 */
869 image->surface.tiling_index[0] = 17;
870 break;
871 }
872 break;
873 default: /* depth, thick */
874 assert(!"unexpected micro mode");
875 return;
876 }
877 }
878
879 image->surface.micro_tile_mode = micro_tile_mode;
880 }
881
882 bool radv_layout_has_htile(const struct radv_image *image,
883 VkImageLayout layout)
884 {
885 return (layout == VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL ||
886 layout == VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL);
887 }
888
889 bool radv_layout_is_htile_compressed(const struct radv_image *image,
890 VkImageLayout layout)
891 {
892 return layout == VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL;
893 }
894
895 bool radv_layout_can_expclear(const struct radv_image *image,
896 VkImageLayout layout)
897 {
898 return (layout == VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL ||
899 layout == VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL);
900 }
901
902 bool radv_layout_can_fast_clear(const struct radv_image *image,
903 VkImageLayout layout,
904 unsigned queue_mask)
905 {
906 return layout == VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL &&
907 queue_mask == (1u << RADV_QUEUE_GENERAL);
908 }
909
910
911 unsigned radv_image_queue_family_mask(const struct radv_image *image, uint32_t family, uint32_t queue_family)
912 {
913 if (!image->exclusive)
914 return image->queue_family_mask;
915 if (family == VK_QUEUE_FAMILY_IGNORED)
916 return 1u << queue_family;
917 return 1u << family;
918 }
919
920 VkResult
921 radv_CreateImage(VkDevice device,
922 const VkImageCreateInfo *pCreateInfo,
923 const VkAllocationCallbacks *pAllocator,
924 VkImage *pImage)
925 {
926 return radv_image_create(device,
927 &(struct radv_image_create_info) {
928 .vk_info = pCreateInfo,
929 .scanout = false,
930 },
931 pAllocator,
932 pImage);
933 }
934
935 void
936 radv_DestroyImage(VkDevice _device, VkImage _image,
937 const VkAllocationCallbacks *pAllocator)
938 {
939 RADV_FROM_HANDLE(radv_device, device, _device);
940
941 if (!_image)
942 return;
943
944 vk_free2(&device->alloc, pAllocator, radv_image_from_handle(_image));
945 }
946
947 void radv_GetImageSubresourceLayout(
948 VkDevice device,
949 VkImage _image,
950 const VkImageSubresource* pSubresource,
951 VkSubresourceLayout* pLayout)
952 {
953 RADV_FROM_HANDLE(radv_image, image, _image);
954 int level = pSubresource->mipLevel;
955 int layer = pSubresource->arrayLayer;
956
957 pLayout->offset = image->surface.level[level].offset + image->surface.level[level].slice_size * layer;
958 pLayout->rowPitch = image->surface.level[level].pitch_bytes;
959 pLayout->arrayPitch = image->surface.level[level].slice_size;
960 pLayout->depthPitch = image->surface.level[level].slice_size;
961 pLayout->size = image->surface.level[level].slice_size;
962 if (image->type == VK_IMAGE_TYPE_3D)
963 pLayout->size *= image->surface.level[level].nblk_z;
964 }
965
966
967 VkResult
968 radv_CreateImageView(VkDevice _device,
969 const VkImageViewCreateInfo *pCreateInfo,
970 const VkAllocationCallbacks *pAllocator,
971 VkImageView *pView)
972 {
973 RADV_FROM_HANDLE(radv_device, device, _device);
974 struct radv_image_view *view;
975
976 view = vk_alloc2(&device->alloc, pAllocator, sizeof(*view), 8,
977 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
978 if (view == NULL)
979 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
980
981 radv_image_view_init(view, device, pCreateInfo, NULL, ~0);
982
983 *pView = radv_image_view_to_handle(view);
984
985 return VK_SUCCESS;
986 }
987
988 void
989 radv_DestroyImageView(VkDevice _device, VkImageView _iview,
990 const VkAllocationCallbacks *pAllocator)
991 {
992 RADV_FROM_HANDLE(radv_device, device, _device);
993 RADV_FROM_HANDLE(radv_image_view, iview, _iview);
994
995 if (!iview)
996 return;
997 vk_free2(&device->alloc, pAllocator, iview);
998 }
999
1000 void radv_buffer_view_init(struct radv_buffer_view *view,
1001 struct radv_device *device,
1002 const VkBufferViewCreateInfo* pCreateInfo,
1003 struct radv_cmd_buffer *cmd_buffer)
1004 {
1005 RADV_FROM_HANDLE(radv_buffer, buffer, pCreateInfo->buffer);
1006
1007 view->bo = buffer->bo;
1008 view->range = pCreateInfo->range == VK_WHOLE_SIZE ?
1009 buffer->size - pCreateInfo->offset : pCreateInfo->range;
1010 view->vk_format = pCreateInfo->format;
1011
1012 radv_make_buffer_descriptor(device, buffer, view->vk_format,
1013 pCreateInfo->offset, view->range, view->state);
1014 }
1015
1016 VkResult
1017 radv_CreateBufferView(VkDevice _device,
1018 const VkBufferViewCreateInfo *pCreateInfo,
1019 const VkAllocationCallbacks *pAllocator,
1020 VkBufferView *pView)
1021 {
1022 RADV_FROM_HANDLE(radv_device, device, _device);
1023 struct radv_buffer_view *view;
1024
1025 view = vk_alloc2(&device->alloc, pAllocator, sizeof(*view), 8,
1026 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1027 if (!view)
1028 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1029
1030 radv_buffer_view_init(view, device, pCreateInfo, NULL);
1031
1032 *pView = radv_buffer_view_to_handle(view);
1033
1034 return VK_SUCCESS;
1035 }
1036
1037 void
1038 radv_DestroyBufferView(VkDevice _device, VkBufferView bufferView,
1039 const VkAllocationCallbacks *pAllocator)
1040 {
1041 RADV_FROM_HANDLE(radv_device, device, _device);
1042 RADV_FROM_HANDLE(radv_buffer_view, view, bufferView);
1043
1044 if (!view)
1045 return;
1046
1047 vk_free2(&device->alloc, pAllocator, view);
1048 }