1 /*
2 * This file is part of FFmpeg.
3 *
4 * FFmpeg is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU Lesser General Public
6 * License as published by the Free Software Foundation; either
7 * version 2.1 of the License, or (at your option) any later version.
8 *
9 * FFmpeg is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * Lesser General Public License for more details.
13 *
14 * You should have received a copy of the GNU Lesser General Public
15 * License along with FFmpeg; if not, write to the Free Software
16 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
17 */
18
19 #include "config.h"
20 #include "pixdesc.h"
21 #include "avstring.h"
22 #include "imgutils.h"
23 #include "hwcontext.h"
24 #include "hwcontext_internal.h"
25 #include "hwcontext_vulkan.h"
26
27 #if CONFIG_LIBDRM
28 #include <unistd.h>
29 #include <xf86drm.h>
30 #include <drm_fourcc.h>
31 #include "hwcontext_drm.h"
32 #if CONFIG_VAAPI
33 #include <va/va_drmcommon.h>
34 #include "hwcontext_vaapi.h"
35 #endif
36 #endif
37
38 #if CONFIG_CUDA
39 #include "hwcontext_cuda_internal.h"
40 #include "cuda_check.h"
41 #define CHECK_CU(x) FF_CUDA_CHECK_DL(cuda_cu, cu, x)
42 #endif
43
44 typedef struct VulkanQueueCtx {
45 VkFence fence;
46 VkQueue queue;
47 int was_synchronous;
48
49 /* Buffer dependencies */
50 AVBufferRef **buf_deps;
51 int nb_buf_deps;
52 int buf_deps_alloc_size;
53 } VulkanQueueCtx;
54
55 typedef struct VulkanExecCtx {
56 VkCommandPool pool;
57 VkCommandBuffer *bufs;
58 VulkanQueueCtx *queues;
59 int nb_queues;
60 int cur_queue_idx;
61 } VulkanExecCtx;
62
63 typedef struct VulkanDevicePriv {
64 /* Properties */
65 VkPhysicalDeviceProperties2 props;
66 VkPhysicalDeviceMemoryProperties mprops;
67 VkPhysicalDeviceExternalMemoryHostPropertiesEXT hprops;
68
69 /* Queues */
70 uint32_t qfs[3];
71 int num_qfs;
72
73 /* Debug callback */
74 VkDebugUtilsMessengerEXT debug_ctx;
75
76 /* Extensions */
77 uint64_t extensions;
78
79 /* Settings */
80 int use_linear_images;
81
82 /* Nvidia */
83 int dev_is_nvidia;
84 } VulkanDevicePriv;
85
86 typedef struct VulkanFramesPriv {
87 /* Image conversions */
88 VulkanExecCtx conv_ctx;
89
90 /* Image transfers */
91 VulkanExecCtx upload_ctx;
92 VulkanExecCtx download_ctx;
93 } VulkanFramesPriv;
94
95 typedef struct AVVkFrameInternal {
96 #if CONFIG_CUDA
97 /* Importing external memory into cuda is really expensive so we keep the
98 * memory imported all the time */
99 AVBufferRef *cuda_fc_ref; /* Need to keep it around for uninit */
100 CUexternalMemory ext_mem[AV_NUM_DATA_POINTERS];
101 CUmipmappedArray cu_mma[AV_NUM_DATA_POINTERS];
102 CUarray cu_array[AV_NUM_DATA_POINTERS];
103 CUexternalSemaphore cu_sem[AV_NUM_DATA_POINTERS];
104 #endif
105 } AVVkFrameInternal;
106
107 #define GET_QUEUE_COUNT(hwctx, graph, comp, tx) ( \
108 graph ? hwctx->nb_graphics_queues : \
109 comp ? (hwctx->nb_comp_queues ? \
110 hwctx->nb_comp_queues : hwctx->nb_graphics_queues) : \
111 tx ? (hwctx->nb_tx_queues ? hwctx->nb_tx_queues : \
112 (hwctx->nb_comp_queues ? \
113 hwctx->nb_comp_queues : hwctx->nb_graphics_queues)) : \
114 0 \
115 )
116
117 #define VK_LOAD_PFN(inst, name) PFN_##name pfn_##name = (PFN_##name) \
118 vkGetInstanceProcAddr(inst, #name)
119
120 #define DEFAULT_USAGE_FLAGS (VK_IMAGE_USAGE_SAMPLED_BIT | \
121 VK_IMAGE_USAGE_STORAGE_BIT | \
122 VK_IMAGE_USAGE_TRANSFER_SRC_BIT | \
123 VK_IMAGE_USAGE_TRANSFER_DST_BIT)
124
125 #define ADD_VAL_TO_LIST(list, count, val) \
126 do { \
127 list = av_realloc_array(list, sizeof(*list), ++count); \
128 if (!list) { \
129 err = AVERROR(ENOMEM); \
130 goto fail; \
131 } \
132 list[count - 1] = av_strdup(val); \
133 if (!list[count - 1]) { \
134 err = AVERROR(ENOMEM); \
135 goto fail; \
136 } \
137 } while(0)
138
139 static const struct {
140 enum AVPixelFormat pixfmt;
141 const VkFormat vkfmts[4];
142 } vk_pixfmt_map[] = {
143 { AV_PIX_FMT_GRAY8, { VK_FORMAT_R8_UNORM } },
144 { AV_PIX_FMT_GRAY16, { VK_FORMAT_R16_UNORM } },
145 { AV_PIX_FMT_GRAYF32, { VK_FORMAT_R32_SFLOAT } },
146
147 { AV_PIX_FMT_NV12, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8G8_UNORM } },
148 { AV_PIX_FMT_NV21, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8G8_UNORM } },
149 { AV_PIX_FMT_P010, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16G16_UNORM } },
150 { AV_PIX_FMT_P016, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16G16_UNORM } },
151
152 { AV_PIX_FMT_NV16, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8G8_UNORM } },
153
154 { AV_PIX_FMT_NV24, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8G8_UNORM } },
155 { AV_PIX_FMT_NV42, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8G8_UNORM } },
156
157 { AV_PIX_FMT_YUV420P, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } },
158 { AV_PIX_FMT_YUV420P10, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
159 { AV_PIX_FMT_YUV420P12, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
160 { AV_PIX_FMT_YUV420P16, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
161
162 { AV_PIX_FMT_YUV422P, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } },
163 { AV_PIX_FMT_YUV422P10, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
164 { AV_PIX_FMT_YUV422P12, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
165 { AV_PIX_FMT_YUV422P16, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
166
167 { AV_PIX_FMT_YUV444P, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } },
168 { AV_PIX_FMT_YUV444P10, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
169 { AV_PIX_FMT_YUV444P12, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
170 { AV_PIX_FMT_YUV444P16, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
171
172 { AV_PIX_FMT_YUVA420P, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } },
173 { AV_PIX_FMT_YUVA420P10, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
174 /* There is no AV_PIX_FMT_YUVA420P12 */
175 { AV_PIX_FMT_YUVA420P16, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
176
177 { AV_PIX_FMT_YUVA422P, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } },
178 { AV_PIX_FMT_YUVA422P10, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
179 { AV_PIX_FMT_YUVA422P12, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
180 { AV_PIX_FMT_YUVA422P16, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
181
182 { AV_PIX_FMT_YUVA444P, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } },
183 { AV_PIX_FMT_YUVA444P10, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
184 { AV_PIX_FMT_YUVA444P12, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
185 { AV_PIX_FMT_YUVA444P16, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
186
187 { AV_PIX_FMT_BGRA, { VK_FORMAT_B8G8R8A8_UNORM } },
188 { AV_PIX_FMT_RGBA, { VK_FORMAT_R8G8B8A8_UNORM } },
189 { AV_PIX_FMT_RGB24, { VK_FORMAT_R8G8B8_UNORM } },
190 { AV_PIX_FMT_BGR24, { VK_FORMAT_B8G8R8_UNORM } },
191 { AV_PIX_FMT_RGB48, { VK_FORMAT_R16G16B16_UNORM } },
192 { AV_PIX_FMT_RGBA64, { VK_FORMAT_R16G16B16A16_UNORM } },
193 { AV_PIX_FMT_RGBA64, { VK_FORMAT_R16G16B16A16_UNORM } },
194 { AV_PIX_FMT_RGB565, { VK_FORMAT_R5G6B5_UNORM_PACK16 } },
195 { AV_PIX_FMT_BGR565, { VK_FORMAT_B5G6R5_UNORM_PACK16 } },
196 { AV_PIX_FMT_BGR0, { VK_FORMAT_B8G8R8A8_UNORM } },
197 { AV_PIX_FMT_RGB0, { VK_FORMAT_R8G8B8A8_UNORM } },
198
199 /* Lower priority as there's an endianess-dependent overlap between these
200 * and rgba/bgr0, and PACK32 formats are more limited */
201 { AV_PIX_FMT_BGR32, { VK_FORMAT_A8B8G8R8_UNORM_PACK32 } },
202 { AV_PIX_FMT_0BGR32, { VK_FORMAT_A8B8G8R8_UNORM_PACK32 } },
203
204 { AV_PIX_FMT_X2RGB10, { VK_FORMAT_A2R10G10B10_UNORM_PACK32 } },
205
206 { AV_PIX_FMT_GBRAP, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } },
207 { AV_PIX_FMT_GBRAP16, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
208 { AV_PIX_FMT_GBRPF32, { VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_SFLOAT } },
209 { AV_PIX_FMT_GBRAPF32, { VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_SFLOAT } },
210 };
211
av_vkfmt_from_pixfmt(enum AVPixelFormat p)212 const VkFormat *av_vkfmt_from_pixfmt(enum AVPixelFormat p)
213 {
214 for (enum AVPixelFormat i = 0; i < FF_ARRAY_ELEMS(vk_pixfmt_map); i++)
215 if (vk_pixfmt_map[i].pixfmt == p)
216 return vk_pixfmt_map[i].vkfmts;
217 return NULL;
218 }
219
pixfmt_is_supported(AVVulkanDeviceContext * hwctx,enum AVPixelFormat p,int linear)220 static int pixfmt_is_supported(AVVulkanDeviceContext *hwctx, enum AVPixelFormat p,
221 int linear)
222 {
223 const VkFormat *fmt = av_vkfmt_from_pixfmt(p);
224 int planes = av_pix_fmt_count_planes(p);
225
226 if (!fmt)
227 return 0;
228
229 for (int i = 0; i < planes; i++) {
230 VkFormatFeatureFlags flags;
231 VkFormatProperties2 prop = {
232 .sType = VK_STRUCTURE_TYPE_FORMAT_PROPERTIES_2,
233 };
234 vkGetPhysicalDeviceFormatProperties2(hwctx->phys_dev, fmt[i], &prop);
235 flags = linear ? prop.formatProperties.linearTilingFeatures :
236 prop.formatProperties.optimalTilingFeatures;
237 if (!(flags & DEFAULT_USAGE_FLAGS))
238 return 0;
239 }
240
241 return 1;
242 }
243
244 enum VulkanExtensions {
245 EXT_EXTERNAL_DMABUF_MEMORY = 1ULL << 0, /* VK_EXT_external_memory_dma_buf */
246 EXT_DRM_MODIFIER_FLAGS = 1ULL << 1, /* VK_EXT_image_drm_format_modifier */
247 EXT_EXTERNAL_FD_MEMORY = 1ULL << 2, /* VK_KHR_external_memory_fd */
248 EXT_EXTERNAL_FD_SEM = 1ULL << 3, /* VK_KHR_external_semaphore_fd */
249 EXT_EXTERNAL_HOST_MEMORY = 1ULL << 4, /* VK_EXT_external_memory_host */
250 EXT_PUSH_DESCRIPTORS = 1ULL << 5, /* VK_KHR_push_descriptor */
251 EXT_HOST_QUERY_RESET = 1ULL << 6, /* VK_EXT_host_query_reset */
252
253 EXT_NO_FLAG = 1ULL << 63,
254 };
255
256 typedef struct VulkanOptExtension {
257 const char *name;
258 uint64_t flag;
259 } VulkanOptExtension;
260
261 static const VulkanOptExtension optional_instance_exts[] = {
262 /* For future use */
263 };
264
265 static const VulkanOptExtension optional_device_exts[] = {
266 { VK_KHR_EXTERNAL_MEMORY_FD_EXTENSION_NAME, EXT_EXTERNAL_FD_MEMORY, },
267 { VK_EXT_EXTERNAL_MEMORY_DMA_BUF_EXTENSION_NAME, EXT_EXTERNAL_DMABUF_MEMORY, },
268 { VK_EXT_IMAGE_DRM_FORMAT_MODIFIER_EXTENSION_NAME, EXT_DRM_MODIFIER_FLAGS, },
269 { VK_KHR_EXTERNAL_SEMAPHORE_FD_EXTENSION_NAME, EXT_EXTERNAL_FD_SEM, },
270 { VK_EXT_EXTERNAL_MEMORY_HOST_EXTENSION_NAME, EXT_EXTERNAL_HOST_MEMORY, },
271 { VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME, EXT_PUSH_DESCRIPTORS, },
272 { VK_EXT_HOST_QUERY_RESET_EXTENSION_NAME, EXT_HOST_QUERY_RESET, },
273 };
274
275 /* Converts return values to strings */
vk_ret2str(VkResult res)276 static const char *vk_ret2str(VkResult res)
277 {
278 #define CASE(VAL) case VAL: return #VAL
279 switch (res) {
280 CASE(VK_SUCCESS);
281 CASE(VK_NOT_READY);
282 CASE(VK_TIMEOUT);
283 CASE(VK_EVENT_SET);
284 CASE(VK_EVENT_RESET);
285 CASE(VK_INCOMPLETE);
286 CASE(VK_ERROR_OUT_OF_HOST_MEMORY);
287 CASE(VK_ERROR_OUT_OF_DEVICE_MEMORY);
288 CASE(VK_ERROR_INITIALIZATION_FAILED);
289 CASE(VK_ERROR_DEVICE_LOST);
290 CASE(VK_ERROR_MEMORY_MAP_FAILED);
291 CASE(VK_ERROR_LAYER_NOT_PRESENT);
292 CASE(VK_ERROR_EXTENSION_NOT_PRESENT);
293 CASE(VK_ERROR_FEATURE_NOT_PRESENT);
294 CASE(VK_ERROR_INCOMPATIBLE_DRIVER);
295 CASE(VK_ERROR_TOO_MANY_OBJECTS);
296 CASE(VK_ERROR_FORMAT_NOT_SUPPORTED);
297 CASE(VK_ERROR_FRAGMENTED_POOL);
298 CASE(VK_ERROR_SURFACE_LOST_KHR);
299 CASE(VK_ERROR_NATIVE_WINDOW_IN_USE_KHR);
300 CASE(VK_SUBOPTIMAL_KHR);
301 CASE(VK_ERROR_OUT_OF_DATE_KHR);
302 CASE(VK_ERROR_INCOMPATIBLE_DISPLAY_KHR);
303 CASE(VK_ERROR_VALIDATION_FAILED_EXT);
304 CASE(VK_ERROR_INVALID_SHADER_NV);
305 CASE(VK_ERROR_OUT_OF_POOL_MEMORY);
306 CASE(VK_ERROR_INVALID_EXTERNAL_HANDLE);
307 CASE(VK_ERROR_NOT_PERMITTED_EXT);
308 CASE(VK_ERROR_INVALID_DRM_FORMAT_MODIFIER_PLANE_LAYOUT_EXT);
309 CASE(VK_ERROR_INVALID_DEVICE_ADDRESS_EXT);
310 CASE(VK_ERROR_FULL_SCREEN_EXCLUSIVE_MODE_LOST_EXT);
311 default: return "Unknown error";
312 }
313 #undef CASE
314 }
315
vk_dbg_callback(VkDebugUtilsMessageSeverityFlagBitsEXT severity,VkDebugUtilsMessageTypeFlagsEXT messageType,const VkDebugUtilsMessengerCallbackDataEXT * data,void * priv)316 static VkBool32 vk_dbg_callback(VkDebugUtilsMessageSeverityFlagBitsEXT severity,
317 VkDebugUtilsMessageTypeFlagsEXT messageType,
318 const VkDebugUtilsMessengerCallbackDataEXT *data,
319 void *priv)
320 {
321 int l;
322 AVHWDeviceContext *ctx = priv;
323
324 switch (severity) {
325 case VK_DEBUG_UTILS_MESSAGE_SEVERITY_VERBOSE_BIT_EXT: l = AV_LOG_VERBOSE; break;
326 case VK_DEBUG_UTILS_MESSAGE_SEVERITY_INFO_BIT_EXT: l = AV_LOG_INFO; break;
327 case VK_DEBUG_UTILS_MESSAGE_SEVERITY_WARNING_BIT_EXT: l = AV_LOG_WARNING; break;
328 case VK_DEBUG_UTILS_MESSAGE_SEVERITY_ERROR_BIT_EXT: l = AV_LOG_ERROR; break;
329 default: l = AV_LOG_DEBUG; break;
330 }
331
332 av_log(ctx, l, "%s\n", data->pMessage);
333 for (int i = 0; i < data->cmdBufLabelCount; i++)
334 av_log(ctx, l, "\t%i: %s\n", i, data->pCmdBufLabels[i].pLabelName);
335
336 return 0;
337 }
338
check_extensions(AVHWDeviceContext * ctx,int dev,AVDictionary * opts,const char * const ** dst,uint32_t * num,int debug)339 static int check_extensions(AVHWDeviceContext *ctx, int dev, AVDictionary *opts,
340 const char * const **dst, uint32_t *num, int debug)
341 {
342 const char *tstr;
343 const char **extension_names = NULL;
344 VulkanDevicePriv *p = ctx->internal->priv;
345 AVVulkanDeviceContext *hwctx = ctx->hwctx;
346 int err = 0, found, extensions_found = 0;
347
348 const char *mod;
349 int optional_exts_num;
350 uint32_t sup_ext_count;
351 char *user_exts_str = NULL;
352 AVDictionaryEntry *user_exts;
353 VkExtensionProperties *sup_ext;
354 const VulkanOptExtension *optional_exts;
355
356 if (!dev) {
357 mod = "instance";
358 optional_exts = optional_instance_exts;
359 optional_exts_num = FF_ARRAY_ELEMS(optional_instance_exts);
360 user_exts = av_dict_get(opts, "instance_extensions", NULL, 0);
361 if (user_exts) {
362 user_exts_str = av_strdup(user_exts->value);
363 if (!user_exts_str) {
364 err = AVERROR(ENOMEM);
365 goto fail;
366 }
367 }
368 vkEnumerateInstanceExtensionProperties(NULL, &sup_ext_count, NULL);
369 sup_ext = av_malloc_array(sup_ext_count, sizeof(VkExtensionProperties));
370 if (!sup_ext)
371 return AVERROR(ENOMEM);
372 vkEnumerateInstanceExtensionProperties(NULL, &sup_ext_count, sup_ext);
373 } else {
374 mod = "device";
375 optional_exts = optional_device_exts;
376 optional_exts_num = FF_ARRAY_ELEMS(optional_device_exts);
377 user_exts = av_dict_get(opts, "device_extensions", NULL, 0);
378 if (user_exts) {
379 user_exts_str = av_strdup(user_exts->value);
380 if (!user_exts_str) {
381 err = AVERROR(ENOMEM);
382 goto fail;
383 }
384 }
385 vkEnumerateDeviceExtensionProperties(hwctx->phys_dev, NULL,
386 &sup_ext_count, NULL);
387 sup_ext = av_malloc_array(sup_ext_count, sizeof(VkExtensionProperties));
388 if (!sup_ext)
389 return AVERROR(ENOMEM);
390 vkEnumerateDeviceExtensionProperties(hwctx->phys_dev, NULL,
391 &sup_ext_count, sup_ext);
392 }
393
394 for (int i = 0; i < optional_exts_num; i++) {
395 tstr = optional_exts[i].name;
396 found = 0;
397 for (int j = 0; j < sup_ext_count; j++) {
398 if (!strcmp(tstr, sup_ext[j].extensionName)) {
399 found = 1;
400 break;
401 }
402 }
403 if (!found)
404 continue;
405
406 av_log(ctx, AV_LOG_VERBOSE, "Using %s extension \"%s\"\n", mod, tstr);
407 p->extensions |= optional_exts[i].flag;
408 ADD_VAL_TO_LIST(extension_names, extensions_found, tstr);
409 }
410
411 if (debug && !dev) {
412 tstr = VK_EXT_DEBUG_UTILS_EXTENSION_NAME;
413 found = 0;
414 for (int j = 0; j < sup_ext_count; j++) {
415 if (!strcmp(tstr, sup_ext[j].extensionName)) {
416 found = 1;
417 break;
418 }
419 }
420 if (found) {
421 av_log(ctx, AV_LOG_VERBOSE, "Using %s extension \"%s\"\n", mod, tstr);
422 ADD_VAL_TO_LIST(extension_names, extensions_found, tstr);
423 } else {
424 av_log(ctx, AV_LOG_ERROR, "Debug extension \"%s\" not found!\n",
425 tstr);
426 err = AVERROR(EINVAL);
427 goto fail;
428 }
429 }
430
431 if (user_exts_str) {
432 char *save, *token = av_strtok(user_exts_str, "+", &save);
433 while (token) {
434 found = 0;
435 for (int j = 0; j < sup_ext_count; j++) {
436 if (!strcmp(token, sup_ext[j].extensionName)) {
437 found = 1;
438 break;
439 }
440 }
441 if (found) {
442 av_log(ctx, AV_LOG_VERBOSE, "Using %s extension \"%s\"\n", mod, token);
443 ADD_VAL_TO_LIST(extension_names, extensions_found, token);
444 } else {
445 av_log(ctx, AV_LOG_WARNING, "%s extension \"%s\" not found, excluding.\n",
446 mod, token);
447 }
448 token = av_strtok(NULL, "+", &save);
449 }
450 }
451
452 *dst = extension_names;
453 *num = extensions_found;
454
455 av_free(user_exts_str);
456 av_free(sup_ext);
457 return 0;
458
459 fail:
460 if (extension_names)
461 for (int i = 0; i < extensions_found; i++)
462 av_free((void *)extension_names[i]);
463 av_free(extension_names);
464 av_free(user_exts_str);
465 av_free(sup_ext);
466 return err;
467 }
468
469 /* Creates a VkInstance */
create_instance(AVHWDeviceContext * ctx,AVDictionary * opts)470 static int create_instance(AVHWDeviceContext *ctx, AVDictionary *opts)
471 {
472 int err = 0;
473 VkResult ret;
474 VulkanDevicePriv *p = ctx->internal->priv;
475 AVVulkanDeviceContext *hwctx = ctx->hwctx;
476 AVDictionaryEntry *debug_opt = av_dict_get(opts, "debug", NULL, 0);
477 const int debug_mode = debug_opt && strtol(debug_opt->value, NULL, 10);
478 VkApplicationInfo application_info = {
479 .sType = VK_STRUCTURE_TYPE_APPLICATION_INFO,
480 .pEngineName = "libavutil",
481 .apiVersion = VK_API_VERSION_1_1,
482 .engineVersion = VK_MAKE_VERSION(LIBAVUTIL_VERSION_MAJOR,
483 LIBAVUTIL_VERSION_MINOR,
484 LIBAVUTIL_VERSION_MICRO),
485 };
486 VkInstanceCreateInfo inst_props = {
487 .sType = VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO,
488 .pApplicationInfo = &application_info,
489 };
490
491 /* Check for present/missing extensions */
492 err = check_extensions(ctx, 0, opts, &inst_props.ppEnabledExtensionNames,
493 &inst_props.enabledExtensionCount, debug_mode);
494 if (err < 0)
495 return err;
496
497 if (debug_mode) {
498 static const char *layers[] = { "VK_LAYER_KHRONOS_validation" };
499 inst_props.ppEnabledLayerNames = layers;
500 inst_props.enabledLayerCount = FF_ARRAY_ELEMS(layers);
501 }
502
503 /* Try to create the instance */
504 ret = vkCreateInstance(&inst_props, hwctx->alloc, &hwctx->inst);
505
506 /* Check for errors */
507 if (ret != VK_SUCCESS) {
508 av_log(ctx, AV_LOG_ERROR, "Instance creation failure: %s\n",
509 vk_ret2str(ret));
510 for (int i = 0; i < inst_props.enabledExtensionCount; i++)
511 av_free((void *)inst_props.ppEnabledExtensionNames[i]);
512 av_free((void *)inst_props.ppEnabledExtensionNames);
513 return AVERROR_EXTERNAL;
514 }
515
516 if (debug_mode) {
517 VkDebugUtilsMessengerCreateInfoEXT dbg = {
518 .sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_MESSENGER_CREATE_INFO_EXT,
519 .messageSeverity = VK_DEBUG_UTILS_MESSAGE_SEVERITY_VERBOSE_BIT_EXT |
520 VK_DEBUG_UTILS_MESSAGE_SEVERITY_INFO_BIT_EXT |
521 VK_DEBUG_UTILS_MESSAGE_SEVERITY_WARNING_BIT_EXT |
522 VK_DEBUG_UTILS_MESSAGE_SEVERITY_ERROR_BIT_EXT,
523 .messageType = VK_DEBUG_UTILS_MESSAGE_TYPE_GENERAL_BIT_EXT |
524 VK_DEBUG_UTILS_MESSAGE_TYPE_VALIDATION_BIT_EXT |
525 VK_DEBUG_UTILS_MESSAGE_TYPE_PERFORMANCE_BIT_EXT,
526 .pfnUserCallback = vk_dbg_callback,
527 .pUserData = ctx,
528 };
529 VK_LOAD_PFN(hwctx->inst, vkCreateDebugUtilsMessengerEXT);
530
531 pfn_vkCreateDebugUtilsMessengerEXT(hwctx->inst, &dbg,
532 hwctx->alloc, &p->debug_ctx);
533 }
534
535 hwctx->enabled_inst_extensions = inst_props.ppEnabledExtensionNames;
536 hwctx->nb_enabled_inst_extensions = inst_props.enabledExtensionCount;
537
538 return 0;
539 }
540
541 typedef struct VulkanDeviceSelection {
542 uint8_t uuid[VK_UUID_SIZE]; /* Will use this first unless !has_uuid */
543 int has_uuid;
544 const char *name; /* Will use this second unless NULL */
545 uint32_t pci_device; /* Will use this third unless 0x0 */
546 uint32_t vendor_id; /* Last resort to find something deterministic */
547 int index; /* Finally fall back to index */
548 } VulkanDeviceSelection;
549
vk_dev_type(enum VkPhysicalDeviceType type)550 static const char *vk_dev_type(enum VkPhysicalDeviceType type)
551 {
552 switch (type) {
553 case VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU: return "integrated";
554 case VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU: return "discrete";
555 case VK_PHYSICAL_DEVICE_TYPE_VIRTUAL_GPU: return "virtual";
556 case VK_PHYSICAL_DEVICE_TYPE_CPU: return "software";
557 default: return "unknown";
558 }
559 }
560
561 /* Finds a device */
find_device(AVHWDeviceContext * ctx,VulkanDeviceSelection * select)562 static int find_device(AVHWDeviceContext *ctx, VulkanDeviceSelection *select)
563 {
564 int err = 0, choice = -1;
565 uint32_t num;
566 VkResult ret;
567 VkPhysicalDevice *devices = NULL;
568 VkPhysicalDeviceIDProperties *idp = NULL;
569 VkPhysicalDeviceProperties2 *prop = NULL;
570 AVVulkanDeviceContext *hwctx = ctx->hwctx;
571
572 ret = vkEnumeratePhysicalDevices(hwctx->inst, &num, NULL);
573 if (ret != VK_SUCCESS || !num) {
574 av_log(ctx, AV_LOG_ERROR, "No devices found: %s!\n", vk_ret2str(ret));
575 return AVERROR(ENODEV);
576 }
577
578 devices = av_malloc_array(num, sizeof(VkPhysicalDevice));
579 if (!devices)
580 return AVERROR(ENOMEM);
581
582 ret = vkEnumeratePhysicalDevices(hwctx->inst, &num, devices);
583 if (ret != VK_SUCCESS) {
584 av_log(ctx, AV_LOG_ERROR, "Failed enumerating devices: %s\n",
585 vk_ret2str(ret));
586 err = AVERROR(ENODEV);
587 goto end;
588 }
589
590 prop = av_mallocz_array(num, sizeof(*prop));
591 if (!prop) {
592 err = AVERROR(ENOMEM);
593 goto end;
594 }
595
596 idp = av_mallocz_array(num, sizeof(*idp));
597 if (!idp) {
598 err = AVERROR(ENOMEM);
599 goto end;
600 }
601
602 av_log(ctx, AV_LOG_VERBOSE, "GPU listing:\n");
603 for (int i = 0; i < num; i++) {
604 idp[i].sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES;
605 prop[i].sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2;
606 prop[i].pNext = &idp[i];
607
608 vkGetPhysicalDeviceProperties2(devices[i], &prop[i]);
609 av_log(ctx, AV_LOG_VERBOSE, " %d: %s (%s) (0x%x)\n", i,
610 prop[i].properties.deviceName,
611 vk_dev_type(prop[i].properties.deviceType),
612 prop[i].properties.deviceID);
613 }
614
615 if (select->has_uuid) {
616 for (int i = 0; i < num; i++) {
617 if (!strncmp(idp[i].deviceUUID, select->uuid, VK_UUID_SIZE)) {
618 choice = i;
619 goto end;
620 }
621 }
622 av_log(ctx, AV_LOG_ERROR, "Unable to find device by given UUID!\n");
623 err = AVERROR(ENODEV);
624 goto end;
625 } else if (select->name) {
626 av_log(ctx, AV_LOG_VERBOSE, "Requested device: %s\n", select->name);
627 for (int i = 0; i < num; i++) {
628 if (strstr(prop[i].properties.deviceName, select->name)) {
629 choice = i;
630 goto end;
631 }
632 }
633 av_log(ctx, AV_LOG_ERROR, "Unable to find device \"%s\"!\n",
634 select->name);
635 err = AVERROR(ENODEV);
636 goto end;
637 } else if (select->pci_device) {
638 av_log(ctx, AV_LOG_VERBOSE, "Requested device: 0x%x\n", select->pci_device);
639 for (int i = 0; i < num; i++) {
640 if (select->pci_device == prop[i].properties.deviceID) {
641 choice = i;
642 goto end;
643 }
644 }
645 av_log(ctx, AV_LOG_ERROR, "Unable to find device with PCI ID 0x%x!\n",
646 select->pci_device);
647 err = AVERROR(EINVAL);
648 goto end;
649 } else if (select->vendor_id) {
650 av_log(ctx, AV_LOG_VERBOSE, "Requested vendor: 0x%x\n", select->vendor_id);
651 for (int i = 0; i < num; i++) {
652 if (select->vendor_id == prop[i].properties.vendorID) {
653 choice = i;
654 goto end;
655 }
656 }
657 av_log(ctx, AV_LOG_ERROR, "Unable to find device with Vendor ID 0x%x!\n",
658 select->vendor_id);
659 err = AVERROR(ENODEV);
660 goto end;
661 } else {
662 if (select->index < num) {
663 choice = select->index;
664 goto end;
665 }
666 av_log(ctx, AV_LOG_ERROR, "Unable to find device with index %i!\n",
667 select->index);
668 err = AVERROR(ENODEV);
669 goto end;
670 }
671
672 end:
673 if (choice > -1)
674 hwctx->phys_dev = devices[choice];
675
676 av_free(devices);
677 av_free(prop);
678 av_free(idp);
679
680 return err;
681 }
682
search_queue_families(AVHWDeviceContext * ctx,VkDeviceCreateInfo * cd)683 static int search_queue_families(AVHWDeviceContext *ctx, VkDeviceCreateInfo *cd)
684 {
685 uint32_t num;
686 float *weights;
687 VkQueueFamilyProperties *qs = NULL;
688 AVVulkanDeviceContext *hwctx = ctx->hwctx;
689 int graph_index = -1, comp_index = -1, tx_index = -1;
690 VkDeviceQueueCreateInfo *pc = (VkDeviceQueueCreateInfo *)cd->pQueueCreateInfos;
691
692 /* First get the number of queue families */
693 vkGetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &num, NULL);
694 if (!num) {
695 av_log(ctx, AV_LOG_ERROR, "Failed to get queues!\n");
696 return AVERROR_EXTERNAL;
697 }
698
699 /* Then allocate memory */
700 qs = av_malloc_array(num, sizeof(VkQueueFamilyProperties));
701 if (!qs)
702 return AVERROR(ENOMEM);
703
704 /* Finally retrieve the queue families */
705 vkGetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &num, qs);
706
707 #define SEARCH_FLAGS(expr, out) \
708 for (int i = 0; i < num; i++) { \
709 const VkQueueFlagBits flags = qs[i].queueFlags; \
710 if (expr) { \
711 out = i; \
712 break; \
713 } \
714 }
715
716 SEARCH_FLAGS(flags & VK_QUEUE_GRAPHICS_BIT, graph_index)
717
718 SEARCH_FLAGS((flags & VK_QUEUE_COMPUTE_BIT) && (i != graph_index),
719 comp_index)
720
721 SEARCH_FLAGS((flags & VK_QUEUE_TRANSFER_BIT) && (i != graph_index) &&
722 (i != comp_index), tx_index)
723
724 #undef SEARCH_FLAGS
725 #define ADD_QUEUE(fidx, graph, comp, tx) \
726 av_log(ctx, AV_LOG_VERBOSE, "Using queue family %i (total queues: %i) for %s%s%s\n", \
727 fidx, qs[fidx].queueCount, graph ? "graphics " : "", \
728 comp ? "compute " : "", tx ? "transfers " : ""); \
729 av_log(ctx, AV_LOG_VERBOSE, " QF %i flags: %s%s%s%s\n", fidx, \
730 ((qs[fidx].queueFlags) & VK_QUEUE_GRAPHICS_BIT) ? "(graphics) " : "", \
731 ((qs[fidx].queueFlags) & VK_QUEUE_COMPUTE_BIT) ? "(compute) " : "", \
732 ((qs[fidx].queueFlags) & VK_QUEUE_TRANSFER_BIT) ? "(transfers) " : "", \
733 ((qs[fidx].queueFlags) & VK_QUEUE_SPARSE_BINDING_BIT) ? "(sparse) " : ""); \
734 pc[cd->queueCreateInfoCount].queueFamilyIndex = fidx; \
735 pc[cd->queueCreateInfoCount].queueCount = qs[fidx].queueCount; \
736 weights = av_malloc(qs[fidx].queueCount * sizeof(float)); \
737 pc[cd->queueCreateInfoCount].pQueuePriorities = weights; \
738 if (!weights) \
739 goto fail; \
740 for (int i = 0; i < qs[fidx].queueCount; i++) \
741 weights[i] = 1.0f; \
742 cd->queueCreateInfoCount++;
743
744 ADD_QUEUE(graph_index, 1, comp_index < 0, tx_index < 0 && comp_index < 0)
745 hwctx->queue_family_index = graph_index;
746 hwctx->queue_family_comp_index = graph_index;
747 hwctx->queue_family_tx_index = graph_index;
748 hwctx->nb_graphics_queues = qs[graph_index].queueCount;
749
750 if (comp_index != -1) {
751 ADD_QUEUE(comp_index, 0, 1, tx_index < 0)
752 hwctx->queue_family_tx_index = comp_index;
753 hwctx->queue_family_comp_index = comp_index;
754 hwctx->nb_comp_queues = qs[comp_index].queueCount;
755 }
756
757 if (tx_index != -1) {
758 ADD_QUEUE(tx_index, 0, 0, 1)
759 hwctx->queue_family_tx_index = tx_index;
760 hwctx->nb_tx_queues = qs[tx_index].queueCount;
761 }
762
763 #undef ADD_QUEUE
764 av_free(qs);
765
766 return 0;
767
768 fail:
769 av_freep(&pc[0].pQueuePriorities);
770 av_freep(&pc[1].pQueuePriorities);
771 av_freep(&pc[2].pQueuePriorities);
772 av_free(qs);
773
774 return AVERROR(ENOMEM);
775 }
776
create_exec_ctx(AVHWFramesContext * hwfc,VulkanExecCtx * cmd,int queue_family_index,int num_queues)777 static int create_exec_ctx(AVHWFramesContext *hwfc, VulkanExecCtx *cmd,
778 int queue_family_index, int num_queues)
779 {
780 VkResult ret;
781 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
782
783 VkCommandPoolCreateInfo cqueue_create = {
784 .sType = VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO,
785 .flags = VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT,
786 .queueFamilyIndex = queue_family_index,
787 };
788 VkCommandBufferAllocateInfo cbuf_create = {
789 .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO,
790 .level = VK_COMMAND_BUFFER_LEVEL_PRIMARY,
791 .commandBufferCount = num_queues,
792 };
793
794 cmd->nb_queues = num_queues;
795
796 /* Create command pool */
797 ret = vkCreateCommandPool(hwctx->act_dev, &cqueue_create,
798 hwctx->alloc, &cmd->pool);
799 if (ret != VK_SUCCESS) {
800 av_log(hwfc, AV_LOG_ERROR, "Command pool creation failure: %s\n",
801 vk_ret2str(ret));
802 return AVERROR_EXTERNAL;
803 }
804
805 cmd->bufs = av_mallocz(num_queues * sizeof(*cmd->bufs));
806 if (!cmd->bufs)
807 return AVERROR(ENOMEM);
808
809 cbuf_create.commandPool = cmd->pool;
810
811 /* Allocate command buffer */
812 ret = vkAllocateCommandBuffers(hwctx->act_dev, &cbuf_create, cmd->bufs);
813 if (ret != VK_SUCCESS) {
814 av_log(hwfc, AV_LOG_ERROR, "Command buffer alloc failure: %s\n",
815 vk_ret2str(ret));
816 av_freep(&cmd->bufs);
817 return AVERROR_EXTERNAL;
818 }
819
820 cmd->queues = av_mallocz(num_queues * sizeof(*cmd->queues));
821 if (!cmd->queues)
822 return AVERROR(ENOMEM);
823
824 for (int i = 0; i < num_queues; i++) {
825 VulkanQueueCtx *q = &cmd->queues[i];
826 vkGetDeviceQueue(hwctx->act_dev, queue_family_index, i, &q->queue);
827 q->was_synchronous = 1;
828 }
829
830 return 0;
831 }
832
free_exec_ctx(AVHWFramesContext * hwfc,VulkanExecCtx * cmd)833 static void free_exec_ctx(AVHWFramesContext *hwfc, VulkanExecCtx *cmd)
834 {
835 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
836
837 if (cmd->queues) {
838 for (int i = 0; i < cmd->nb_queues; i++) {
839 VulkanQueueCtx *q = &cmd->queues[i];
840
841 /* Make sure all queues have finished executing */
842 if (q->fence && !q->was_synchronous) {
843 vkWaitForFences(hwctx->act_dev, 1, &q->fence, VK_TRUE, UINT64_MAX);
844 vkResetFences(hwctx->act_dev, 1, &q->fence);
845 }
846
847 /* Free the fence */
848 if (q->fence)
849 vkDestroyFence(hwctx->act_dev, q->fence, hwctx->alloc);
850
851 /* Free buffer dependencies */
852 for (int j = 0; j < q->nb_buf_deps; j++)
853 av_buffer_unref(&q->buf_deps[j]);
854 av_free(q->buf_deps);
855 }
856 }
857
858 if (cmd->bufs)
859 vkFreeCommandBuffers(hwctx->act_dev, cmd->pool, cmd->nb_queues, cmd->bufs);
860 if (cmd->pool)
861 vkDestroyCommandPool(hwctx->act_dev, cmd->pool, hwctx->alloc);
862
863 av_freep(&cmd->queues);
864 av_freep(&cmd->bufs);
865 cmd->pool = NULL;
866 }
867
get_buf_exec_ctx(AVHWFramesContext * hwfc,VulkanExecCtx * cmd)868 static VkCommandBuffer get_buf_exec_ctx(AVHWFramesContext *hwfc, VulkanExecCtx *cmd)
869 {
870 return cmd->bufs[cmd->cur_queue_idx];
871 }
872
unref_exec_ctx_deps(AVHWFramesContext * hwfc,VulkanExecCtx * cmd)873 static void unref_exec_ctx_deps(AVHWFramesContext *hwfc, VulkanExecCtx *cmd)
874 {
875 VulkanQueueCtx *q = &cmd->queues[cmd->cur_queue_idx];
876
877 for (int j = 0; j < q->nb_buf_deps; j++)
878 av_buffer_unref(&q->buf_deps[j]);
879 q->nb_buf_deps = 0;
880 }
881
wait_start_exec_ctx(AVHWFramesContext * hwfc,VulkanExecCtx * cmd)882 static int wait_start_exec_ctx(AVHWFramesContext *hwfc, VulkanExecCtx *cmd)
883 {
884 VkResult ret;
885 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
886 VulkanQueueCtx *q = &cmd->queues[cmd->cur_queue_idx];
887
888 VkCommandBufferBeginInfo cmd_start = {
889 .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
890 .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
891 };
892
893 /* Create the fence and don't wait for it initially */
894 if (!q->fence) {
895 VkFenceCreateInfo fence_spawn = {
896 .sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO,
897 };
898 ret = vkCreateFence(hwctx->act_dev, &fence_spawn, hwctx->alloc,
899 &q->fence);
900 if (ret != VK_SUCCESS) {
901 av_log(hwfc, AV_LOG_ERROR, "Failed to queue frame fence: %s\n",
902 vk_ret2str(ret));
903 return AVERROR_EXTERNAL;
904 }
905 } else if (!q->was_synchronous) {
906 vkWaitForFences(hwctx->act_dev, 1, &q->fence, VK_TRUE, UINT64_MAX);
907 vkResetFences(hwctx->act_dev, 1, &q->fence);
908 }
909
910 /* Discard queue dependencies */
911 unref_exec_ctx_deps(hwfc, cmd);
912
913 ret = vkBeginCommandBuffer(cmd->bufs[cmd->cur_queue_idx], &cmd_start);
914 if (ret != VK_SUCCESS) {
915 av_log(hwfc, AV_LOG_ERROR, "Unable to init command buffer: %s\n",
916 vk_ret2str(ret));
917 return AVERROR_EXTERNAL;
918 }
919
920 return 0;
921 }
922
add_buf_dep_exec_ctx(AVHWFramesContext * hwfc,VulkanExecCtx * cmd,AVBufferRef * const * deps,int nb_deps)923 static int add_buf_dep_exec_ctx(AVHWFramesContext *hwfc, VulkanExecCtx *cmd,
924 AVBufferRef * const *deps, int nb_deps)
925 {
926 AVBufferRef **dst;
927 VulkanQueueCtx *q = &cmd->queues[cmd->cur_queue_idx];
928
929 if (!deps || !nb_deps)
930 return 0;
931
932 dst = av_fast_realloc(q->buf_deps, &q->buf_deps_alloc_size,
933 (q->nb_buf_deps + nb_deps) * sizeof(*dst));
934 if (!dst)
935 goto err;
936
937 q->buf_deps = dst;
938
939 for (int i = 0; i < nb_deps; i++) {
940 q->buf_deps[q->nb_buf_deps] = av_buffer_ref(deps[i]);
941 if (!q->buf_deps[q->nb_buf_deps])
942 goto err;
943 q->nb_buf_deps++;
944 }
945
946 return 0;
947
948 err:
949 unref_exec_ctx_deps(hwfc, cmd);
950 return AVERROR(ENOMEM);
951 }
952
submit_exec_ctx(AVHWFramesContext * hwfc,VulkanExecCtx * cmd,VkSubmitInfo * s_info,int synchronous)953 static int submit_exec_ctx(AVHWFramesContext *hwfc, VulkanExecCtx *cmd,
954 VkSubmitInfo *s_info, int synchronous)
955 {
956 VkResult ret;
957 VulkanQueueCtx *q = &cmd->queues[cmd->cur_queue_idx];
958
959 ret = vkEndCommandBuffer(cmd->bufs[cmd->cur_queue_idx]);
960 if (ret != VK_SUCCESS) {
961 av_log(hwfc, AV_LOG_ERROR, "Unable to finish command buffer: %s\n",
962 vk_ret2str(ret));
963 unref_exec_ctx_deps(hwfc, cmd);
964 return AVERROR_EXTERNAL;
965 }
966
967 s_info->pCommandBuffers = &cmd->bufs[cmd->cur_queue_idx];
968 s_info->commandBufferCount = 1;
969
970 ret = vkQueueSubmit(q->queue, 1, s_info, q->fence);
971 if (ret != VK_SUCCESS) {
972 unref_exec_ctx_deps(hwfc, cmd);
973 return AVERROR_EXTERNAL;
974 }
975
976 q->was_synchronous = synchronous;
977
978 if (synchronous) {
979 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
980 vkWaitForFences(hwctx->act_dev, 1, &q->fence, VK_TRUE, UINT64_MAX);
981 vkResetFences(hwctx->act_dev, 1, &q->fence);
982 unref_exec_ctx_deps(hwfc, cmd);
983 } else { /* Rotate queues */
984 cmd->cur_queue_idx = (cmd->cur_queue_idx + 1) % cmd->nb_queues;
985 }
986
987 return 0;
988 }
989
vulkan_device_free(AVHWDeviceContext * ctx)990 static void vulkan_device_free(AVHWDeviceContext *ctx)
991 {
992 VulkanDevicePriv *p = ctx->internal->priv;
993 AVVulkanDeviceContext *hwctx = ctx->hwctx;
994
995 vkDestroyDevice(hwctx->act_dev, hwctx->alloc);
996
997 if (p->debug_ctx) {
998 VK_LOAD_PFN(hwctx->inst, vkDestroyDebugUtilsMessengerEXT);
999 pfn_vkDestroyDebugUtilsMessengerEXT(hwctx->inst, p->debug_ctx,
1000 hwctx->alloc);
1001 }
1002
1003 vkDestroyInstance(hwctx->inst, hwctx->alloc);
1004
1005 for (int i = 0; i < hwctx->nb_enabled_inst_extensions; i++)
1006 av_free((void *)hwctx->enabled_inst_extensions[i]);
1007 av_free((void *)hwctx->enabled_inst_extensions);
1008
1009 for (int i = 0; i < hwctx->nb_enabled_dev_extensions; i++)
1010 av_free((void *)hwctx->enabled_dev_extensions[i]);
1011 av_free((void *)hwctx->enabled_dev_extensions);
1012 }
1013
vulkan_device_create_internal(AVHWDeviceContext * ctx,VulkanDeviceSelection * dev_select,AVDictionary * opts,int flags)1014 static int vulkan_device_create_internal(AVHWDeviceContext *ctx,
1015 VulkanDeviceSelection *dev_select,
1016 AVDictionary *opts, int flags)
1017 {
1018 int err = 0;
1019 VkResult ret;
1020 AVDictionaryEntry *opt_d;
1021 VulkanDevicePriv *p = ctx->internal->priv;
1022 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1023 VkPhysicalDeviceFeatures dev_features = { 0 };
1024 VkDeviceQueueCreateInfo queue_create_info[3] = {
1025 { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO, },
1026 { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO, },
1027 { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO, },
1028 };
1029
1030 VkDeviceCreateInfo dev_info = {
1031 .sType = VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO,
1032 .pNext = &hwctx->device_features,
1033 .pQueueCreateInfos = queue_create_info,
1034 .queueCreateInfoCount = 0,
1035 };
1036
1037 hwctx->device_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2;
1038 ctx->free = vulkan_device_free;
1039
1040 /* Create an instance if not given one */
1041 if ((err = create_instance(ctx, opts)))
1042 goto end;
1043
1044 /* Find a device (if not given one) */
1045 if ((err = find_device(ctx, dev_select)))
1046 goto end;
1047
1048 vkGetPhysicalDeviceFeatures(hwctx->phys_dev, &dev_features);
1049 #define COPY_FEATURE(DST, NAME) (DST).features.NAME = dev_features.NAME;
1050 COPY_FEATURE(hwctx->device_features, shaderImageGatherExtended)
1051 COPY_FEATURE(hwctx->device_features, shaderStorageImageReadWithoutFormat)
1052 COPY_FEATURE(hwctx->device_features, shaderStorageImageWriteWithoutFormat)
1053 COPY_FEATURE(hwctx->device_features, fragmentStoresAndAtomics)
1054 COPY_FEATURE(hwctx->device_features, vertexPipelineStoresAndAtomics)
1055 COPY_FEATURE(hwctx->device_features, shaderInt64)
1056 #undef COPY_FEATURE
1057
1058 /* Search queue family */
1059 if ((err = search_queue_families(ctx, &dev_info)))
1060 goto end;
1061
1062 if ((err = check_extensions(ctx, 1, opts, &dev_info.ppEnabledExtensionNames,
1063 &dev_info.enabledExtensionCount, 0))) {
1064 av_free((void *)queue_create_info[0].pQueuePriorities);
1065 av_free((void *)queue_create_info[1].pQueuePriorities);
1066 av_free((void *)queue_create_info[2].pQueuePriorities);
1067 goto end;
1068 }
1069
1070 ret = vkCreateDevice(hwctx->phys_dev, &dev_info, hwctx->alloc,
1071 &hwctx->act_dev);
1072
1073 av_free((void *)queue_create_info[0].pQueuePriorities);
1074 av_free((void *)queue_create_info[1].pQueuePriorities);
1075 av_free((void *)queue_create_info[2].pQueuePriorities);
1076
1077 if (ret != VK_SUCCESS) {
1078 av_log(ctx, AV_LOG_ERROR, "Device creation failure: %s\n",
1079 vk_ret2str(ret));
1080 for (int i = 0; i < dev_info.enabledExtensionCount; i++)
1081 av_free((void *)dev_info.ppEnabledExtensionNames[i]);
1082 av_free((void *)dev_info.ppEnabledExtensionNames);
1083 err = AVERROR_EXTERNAL;
1084 goto end;
1085 }
1086
1087 /* Tiled images setting, use them by default */
1088 opt_d = av_dict_get(opts, "linear_images", NULL, 0);
1089 if (opt_d)
1090 p->use_linear_images = strtol(opt_d->value, NULL, 10);
1091
1092 hwctx->enabled_dev_extensions = dev_info.ppEnabledExtensionNames;
1093 hwctx->nb_enabled_dev_extensions = dev_info.enabledExtensionCount;
1094
1095 end:
1096 return err;
1097 }
1098
vulkan_device_init(AVHWDeviceContext * ctx)1099 static int vulkan_device_init(AVHWDeviceContext *ctx)
1100 {
1101 uint32_t queue_num;
1102 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1103 VulkanDevicePriv *p = ctx->internal->priv;
1104
1105 /* Set device extension flags */
1106 for (int i = 0; i < hwctx->nb_enabled_dev_extensions; i++) {
1107 for (int j = 0; j < FF_ARRAY_ELEMS(optional_device_exts); j++) {
1108 if (!strcmp(hwctx->enabled_dev_extensions[i],
1109 optional_device_exts[j].name)) {
1110 av_log(ctx, AV_LOG_VERBOSE, "Using device extension %s\n",
1111 hwctx->enabled_dev_extensions[i]);
1112 p->extensions |= optional_device_exts[j].flag;
1113 break;
1114 }
1115 }
1116 }
1117
1118 p->props.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2;
1119 p->props.pNext = &p->hprops;
1120 p->hprops.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT;
1121
1122 vkGetPhysicalDeviceProperties2(hwctx->phys_dev, &p->props);
1123 av_log(ctx, AV_LOG_VERBOSE, "Using device: %s\n",
1124 p->props.properties.deviceName);
1125 av_log(ctx, AV_LOG_VERBOSE, "Alignments:\n");
1126 av_log(ctx, AV_LOG_VERBOSE, " optimalBufferCopyRowPitchAlignment: %li\n",
1127 p->props.properties.limits.optimalBufferCopyRowPitchAlignment);
1128 av_log(ctx, AV_LOG_VERBOSE, " minMemoryMapAlignment: %li\n",
1129 p->props.properties.limits.minMemoryMapAlignment);
1130 if (p->extensions & EXT_EXTERNAL_HOST_MEMORY)
1131 av_log(ctx, AV_LOG_VERBOSE, " minImportedHostPointerAlignment: %li\n",
1132 p->hprops.minImportedHostPointerAlignment);
1133
1134 p->dev_is_nvidia = (p->props.properties.vendorID == 0x10de);
1135
1136 vkGetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &queue_num, NULL);
1137 if (!queue_num) {
1138 av_log(ctx, AV_LOG_ERROR, "Failed to get queues!\n");
1139 return AVERROR_EXTERNAL;
1140 }
1141
1142 #define CHECK_QUEUE(type, n) \
1143 if (n >= queue_num) { \
1144 av_log(ctx, AV_LOG_ERROR, "Invalid %s queue index %i (device has %i queues)!\n", \
1145 type, n, queue_num); \
1146 return AVERROR(EINVAL); \
1147 }
1148
1149 CHECK_QUEUE("graphics", hwctx->queue_family_index)
1150 CHECK_QUEUE("upload", hwctx->queue_family_tx_index)
1151 CHECK_QUEUE("compute", hwctx->queue_family_comp_index)
1152
1153 #undef CHECK_QUEUE
1154
1155 p->qfs[p->num_qfs++] = hwctx->queue_family_index;
1156 if ((hwctx->queue_family_tx_index != hwctx->queue_family_index) &&
1157 (hwctx->queue_family_tx_index != hwctx->queue_family_comp_index))
1158 p->qfs[p->num_qfs++] = hwctx->queue_family_tx_index;
1159 if ((hwctx->queue_family_comp_index != hwctx->queue_family_index) &&
1160 (hwctx->queue_family_comp_index != hwctx->queue_family_tx_index))
1161 p->qfs[p->num_qfs++] = hwctx->queue_family_comp_index;
1162
1163 /* Get device capabilities */
1164 vkGetPhysicalDeviceMemoryProperties(hwctx->phys_dev, &p->mprops);
1165
1166 return 0;
1167 }
1168
vulkan_device_create(AVHWDeviceContext * ctx,const char * device,AVDictionary * opts,int flags)1169 static int vulkan_device_create(AVHWDeviceContext *ctx, const char *device,
1170 AVDictionary *opts, int flags)
1171 {
1172 VulkanDeviceSelection dev_select = { 0 };
1173 if (device && device[0]) {
1174 char *end = NULL;
1175 dev_select.index = strtol(device, &end, 10);
1176 if (end == device) {
1177 dev_select.index = 0;
1178 dev_select.name = device;
1179 }
1180 }
1181
1182 return vulkan_device_create_internal(ctx, &dev_select, opts, flags);
1183 }
1184
vulkan_device_derive(AVHWDeviceContext * ctx,AVHWDeviceContext * src_ctx,AVDictionary * opts,int flags)1185 static int vulkan_device_derive(AVHWDeviceContext *ctx,
1186 AVHWDeviceContext *src_ctx,
1187 AVDictionary *opts, int flags)
1188 {
1189 av_unused VulkanDeviceSelection dev_select = { 0 };
1190
1191 /* If there's only one device on the system, then even if its not covered
1192 * by the following checks (e.g. non-PCIe ARM GPU), having an empty
1193 * dev_select will mean it'll get picked. */
1194 switch(src_ctx->type) {
1195 #if CONFIG_LIBDRM
1196 #if CONFIG_VAAPI
1197 case AV_HWDEVICE_TYPE_VAAPI: {
1198 AVVAAPIDeviceContext *src_hwctx = src_ctx->hwctx;
1199
1200 const char *vendor = vaQueryVendorString(src_hwctx->display);
1201 if (!vendor) {
1202 av_log(ctx, AV_LOG_ERROR, "Unable to get device info from VAAPI!\n");
1203 return AVERROR_EXTERNAL;
1204 }
1205
1206 if (strstr(vendor, "Intel"))
1207 dev_select.vendor_id = 0x8086;
1208 if (strstr(vendor, "AMD"))
1209 dev_select.vendor_id = 0x1002;
1210
1211 return vulkan_device_create_internal(ctx, &dev_select, opts, flags);
1212 }
1213 #endif
1214 case AV_HWDEVICE_TYPE_DRM: {
1215 AVDRMDeviceContext *src_hwctx = src_ctx->hwctx;
1216
1217 drmDevice *drm_dev_info;
1218 int err = drmGetDevice(src_hwctx->fd, &drm_dev_info);
1219 if (err) {
1220 av_log(ctx, AV_LOG_ERROR, "Unable to get device info from DRM fd!\n");
1221 return AVERROR_EXTERNAL;
1222 }
1223
1224 if (drm_dev_info->bustype == DRM_BUS_PCI)
1225 dev_select.pci_device = drm_dev_info->deviceinfo.pci->device_id;
1226
1227 drmFreeDevice(&drm_dev_info);
1228
1229 return vulkan_device_create_internal(ctx, &dev_select, opts, flags);
1230 }
1231 #endif
1232 #if CONFIG_CUDA
1233 case AV_HWDEVICE_TYPE_CUDA: {
1234 AVHWDeviceContext *cuda_cu = src_ctx;
1235 AVCUDADeviceContext *src_hwctx = src_ctx->hwctx;
1236 AVCUDADeviceContextInternal *cu_internal = src_hwctx->internal;
1237 CudaFunctions *cu = cu_internal->cuda_dl;
1238
1239 int ret = CHECK_CU(cu->cuDeviceGetUuid((CUuuid *)&dev_select.uuid,
1240 cu_internal->cuda_device));
1241 if (ret < 0) {
1242 av_log(ctx, AV_LOG_ERROR, "Unable to get UUID from CUDA!\n");
1243 return AVERROR_EXTERNAL;
1244 }
1245
1246 dev_select.has_uuid = 1;
1247
1248 return vulkan_device_create_internal(ctx, &dev_select, opts, flags);
1249 }
1250 #endif
1251 default:
1252 return AVERROR(ENOSYS);
1253 }
1254 }
1255
vulkan_frames_get_constraints(AVHWDeviceContext * ctx,const void * hwconfig,AVHWFramesConstraints * constraints)1256 static int vulkan_frames_get_constraints(AVHWDeviceContext *ctx,
1257 const void *hwconfig,
1258 AVHWFramesConstraints *constraints)
1259 {
1260 int count = 0;
1261 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1262 VulkanDevicePriv *p = ctx->internal->priv;
1263
1264 for (enum AVPixelFormat i = 0; i < AV_PIX_FMT_NB; i++)
1265 count += pixfmt_is_supported(hwctx, i, p->use_linear_images);
1266
1267 #if CONFIG_CUDA
1268 if (p->dev_is_nvidia)
1269 count++;
1270 #endif
1271
1272 constraints->valid_sw_formats = av_malloc_array(count + 1,
1273 sizeof(enum AVPixelFormat));
1274 if (!constraints->valid_sw_formats)
1275 return AVERROR(ENOMEM);
1276
1277 count = 0;
1278 for (enum AVPixelFormat i = 0; i < AV_PIX_FMT_NB; i++)
1279 if (pixfmt_is_supported(hwctx, i, p->use_linear_images))
1280 constraints->valid_sw_formats[count++] = i;
1281
1282 #if CONFIG_CUDA
1283 if (p->dev_is_nvidia)
1284 constraints->valid_sw_formats[count++] = AV_PIX_FMT_CUDA;
1285 #endif
1286 constraints->valid_sw_formats[count++] = AV_PIX_FMT_NONE;
1287
1288 constraints->min_width = 0;
1289 constraints->min_height = 0;
1290 constraints->max_width = p->props.properties.limits.maxImageDimension2D;
1291 constraints->max_height = p->props.properties.limits.maxImageDimension2D;
1292
1293 constraints->valid_hw_formats = av_malloc_array(2, sizeof(enum AVPixelFormat));
1294 if (!constraints->valid_hw_formats)
1295 return AVERROR(ENOMEM);
1296
1297 constraints->valid_hw_formats[0] = AV_PIX_FMT_VULKAN;
1298 constraints->valid_hw_formats[1] = AV_PIX_FMT_NONE;
1299
1300 return 0;
1301 }
1302
alloc_mem(AVHWDeviceContext * ctx,VkMemoryRequirements * req,VkMemoryPropertyFlagBits req_flags,const void * alloc_extension,VkMemoryPropertyFlagBits * mem_flags,VkDeviceMemory * mem)1303 static int alloc_mem(AVHWDeviceContext *ctx, VkMemoryRequirements *req,
1304 VkMemoryPropertyFlagBits req_flags, const void *alloc_extension,
1305 VkMemoryPropertyFlagBits *mem_flags, VkDeviceMemory *mem)
1306 {
1307 VkResult ret;
1308 int index = -1;
1309 VulkanDevicePriv *p = ctx->internal->priv;
1310 AVVulkanDeviceContext *dev_hwctx = ctx->hwctx;
1311 VkMemoryAllocateInfo alloc_info = {
1312 .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
1313 .pNext = alloc_extension,
1314 .allocationSize = req->size,
1315 };
1316
1317 /* The vulkan spec requires memory types to be sorted in the "optimal"
1318 * order, so the first matching type we find will be the best/fastest one */
1319 for (int i = 0; i < p->mprops.memoryTypeCount; i++) {
1320 const VkMemoryType *type = &p->mprops.memoryTypes[i];
1321
1322 /* The memory type must be supported by the requirements (bitfield) */
1323 if (!(req->memoryTypeBits & (1 << i)))
1324 continue;
1325
1326 /* The memory type flags must include our properties */
1327 if ((type->propertyFlags & req_flags) != req_flags)
1328 continue;
1329
1330 /* The memory type must be large enough */
1331 if (req->size > p->mprops.memoryHeaps[type->heapIndex].size)
1332 continue;
1333
1334 /* Found a suitable memory type */
1335 index = i;
1336 break;
1337 }
1338
1339 if (index < 0) {
1340 av_log(ctx, AV_LOG_ERROR, "No memory type found for flags 0x%x\n",
1341 req_flags);
1342 return AVERROR(EINVAL);
1343 }
1344
1345 alloc_info.memoryTypeIndex = index;
1346
1347 ret = vkAllocateMemory(dev_hwctx->act_dev, &alloc_info,
1348 dev_hwctx->alloc, mem);
1349 if (ret != VK_SUCCESS) {
1350 av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory: %s\n",
1351 vk_ret2str(ret));
1352 return AVERROR(ENOMEM);
1353 }
1354
1355 *mem_flags |= p->mprops.memoryTypes[index].propertyFlags;
1356
1357 return 0;
1358 }
1359
vulkan_free_internal(AVVkFrameInternal * internal)1360 static void vulkan_free_internal(AVVkFrameInternal *internal)
1361 {
1362 if (!internal)
1363 return;
1364
1365 #if CONFIG_CUDA
1366 if (internal->cuda_fc_ref) {
1367 AVHWFramesContext *cuda_fc = (AVHWFramesContext *)internal->cuda_fc_ref->data;
1368 int planes = av_pix_fmt_count_planes(cuda_fc->sw_format);
1369 AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
1370 AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
1371 AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
1372 CudaFunctions *cu = cu_internal->cuda_dl;
1373
1374 for (int i = 0; i < planes; i++) {
1375 if (internal->cu_sem[i])
1376 CHECK_CU(cu->cuDestroyExternalSemaphore(internal->cu_sem[i]));
1377 if (internal->cu_mma[i])
1378 CHECK_CU(cu->cuMipmappedArrayDestroy(internal->cu_mma[i]));
1379 if (internal->ext_mem[i])
1380 CHECK_CU(cu->cuDestroyExternalMemory(internal->ext_mem[i]));
1381 }
1382
1383 av_buffer_unref(&internal->cuda_fc_ref);
1384 }
1385 #endif
1386
1387 av_free(internal);
1388 }
1389
vulkan_frame_free(void * opaque,uint8_t * data)1390 static void vulkan_frame_free(void *opaque, uint8_t *data)
1391 {
1392 AVVkFrame *f = (AVVkFrame *)data;
1393 AVHWFramesContext *hwfc = opaque;
1394 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1395 int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1396
1397 vulkan_free_internal(f->internal);
1398
1399 for (int i = 0; i < planes; i++) {
1400 vkDestroyImage(hwctx->act_dev, f->img[i], hwctx->alloc);
1401 vkFreeMemory(hwctx->act_dev, f->mem[i], hwctx->alloc);
1402 vkDestroySemaphore(hwctx->act_dev, f->sem[i], hwctx->alloc);
1403 }
1404
1405 av_free(f);
1406 }
1407
alloc_bind_mem(AVHWFramesContext * hwfc,AVVkFrame * f,void * alloc_pnext,size_t alloc_pnext_stride)1408 static int alloc_bind_mem(AVHWFramesContext *hwfc, AVVkFrame *f,
1409 void *alloc_pnext, size_t alloc_pnext_stride)
1410 {
1411 int err;
1412 VkResult ret;
1413 AVHWDeviceContext *ctx = hwfc->device_ctx;
1414 VulkanDevicePriv *p = ctx->internal->priv;
1415 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1416 VkBindImageMemoryInfo bind_info[AV_NUM_DATA_POINTERS] = { { 0 } };
1417
1418 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1419
1420 for (int i = 0; i < planes; i++) {
1421 int use_ded_mem;
1422 VkImageMemoryRequirementsInfo2 req_desc = {
1423 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_REQUIREMENTS_INFO_2,
1424 .image = f->img[i],
1425 };
1426 VkMemoryDedicatedAllocateInfo ded_alloc = {
1427 .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO,
1428 .pNext = (void *)(((uint8_t *)alloc_pnext) + i*alloc_pnext_stride),
1429 };
1430 VkMemoryDedicatedRequirements ded_req = {
1431 .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS,
1432 };
1433 VkMemoryRequirements2 req = {
1434 .sType = VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2,
1435 .pNext = &ded_req,
1436 };
1437
1438 vkGetImageMemoryRequirements2(hwctx->act_dev, &req_desc, &req);
1439
1440 if (f->tiling == VK_IMAGE_TILING_LINEAR)
1441 req.memoryRequirements.size = FFALIGN(req.memoryRequirements.size,
1442 p->props.properties.limits.minMemoryMapAlignment);
1443
1444 /* In case the implementation prefers/requires dedicated allocation */
1445 use_ded_mem = ded_req.prefersDedicatedAllocation |
1446 ded_req.requiresDedicatedAllocation;
1447 if (use_ded_mem)
1448 ded_alloc.image = f->img[i];
1449
1450 /* Allocate memory */
1451 if ((err = alloc_mem(ctx, &req.memoryRequirements,
1452 f->tiling == VK_IMAGE_TILING_LINEAR ?
1453 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT :
1454 VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
1455 use_ded_mem ? &ded_alloc : (void *)ded_alloc.pNext,
1456 &f->flags, &f->mem[i])))
1457 return err;
1458
1459 f->size[i] = req.memoryRequirements.size;
1460 bind_info[i].sType = VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO;
1461 bind_info[i].image = f->img[i];
1462 bind_info[i].memory = f->mem[i];
1463 }
1464
1465 /* Bind the allocated memory to the images */
1466 ret = vkBindImageMemory2(hwctx->act_dev, planes, bind_info);
1467 if (ret != VK_SUCCESS) {
1468 av_log(ctx, AV_LOG_ERROR, "Failed to bind memory: %s\n",
1469 vk_ret2str(ret));
1470 return AVERROR_EXTERNAL;
1471 }
1472
1473 return 0;
1474 }
1475
1476 enum PrepMode {
1477 PREP_MODE_WRITE,
1478 PREP_MODE_RO_SHADER,
1479 PREP_MODE_EXTERNAL_EXPORT,
1480 };
1481
prepare_frame(AVHWFramesContext * hwfc,VulkanExecCtx * ectx,AVVkFrame * frame,enum PrepMode pmode)1482 static int prepare_frame(AVHWFramesContext *hwfc, VulkanExecCtx *ectx,
1483 AVVkFrame *frame, enum PrepMode pmode)
1484 {
1485 int err;
1486 uint32_t dst_qf;
1487 VkImageLayout new_layout;
1488 VkAccessFlags new_access;
1489 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1490
1491 VkImageMemoryBarrier img_bar[AV_NUM_DATA_POINTERS] = { 0 };
1492
1493 VkSubmitInfo s_info = {
1494 .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
1495 .pSignalSemaphores = frame->sem,
1496 .signalSemaphoreCount = planes,
1497 };
1498
1499 VkPipelineStageFlagBits wait_st[AV_NUM_DATA_POINTERS];
1500 for (int i = 0; i < planes; i++)
1501 wait_st[i] = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
1502
1503 switch (pmode) {
1504 case PREP_MODE_WRITE:
1505 new_layout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
1506 new_access = VK_ACCESS_TRANSFER_WRITE_BIT;
1507 dst_qf = VK_QUEUE_FAMILY_IGNORED;
1508 break;
1509 case PREP_MODE_RO_SHADER:
1510 new_layout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL;
1511 new_access = VK_ACCESS_TRANSFER_READ_BIT;
1512 dst_qf = VK_QUEUE_FAMILY_IGNORED;
1513 break;
1514 case PREP_MODE_EXTERNAL_EXPORT:
1515 new_layout = VK_IMAGE_LAYOUT_GENERAL;
1516 new_access = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT;
1517 dst_qf = VK_QUEUE_FAMILY_EXTERNAL_KHR;
1518 s_info.pWaitSemaphores = frame->sem;
1519 s_info.pWaitDstStageMask = wait_st;
1520 s_info.waitSemaphoreCount = planes;
1521 break;
1522 }
1523
1524 if ((err = wait_start_exec_ctx(hwfc, ectx)))
1525 return err;
1526
1527 /* Change the image layout to something more optimal for writes.
1528 * This also signals the newly created semaphore, making it usable
1529 * for synchronization */
1530 for (int i = 0; i < planes; i++) {
1531 img_bar[i].sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
1532 img_bar[i].srcAccessMask = 0x0;
1533 img_bar[i].dstAccessMask = new_access;
1534 img_bar[i].oldLayout = frame->layout[i];
1535 img_bar[i].newLayout = new_layout;
1536 img_bar[i].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
1537 img_bar[i].dstQueueFamilyIndex = dst_qf;
1538 img_bar[i].image = frame->img[i];
1539 img_bar[i].subresourceRange.levelCount = 1;
1540 img_bar[i].subresourceRange.layerCount = 1;
1541 img_bar[i].subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
1542
1543 frame->layout[i] = img_bar[i].newLayout;
1544 frame->access[i] = img_bar[i].dstAccessMask;
1545 }
1546
1547 vkCmdPipelineBarrier(get_buf_exec_ctx(hwfc, ectx),
1548 VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
1549 VK_PIPELINE_STAGE_TRANSFER_BIT,
1550 0, 0, NULL, 0, NULL, planes, img_bar);
1551
1552 return submit_exec_ctx(hwfc, ectx, &s_info, 0);
1553 }
1554
get_plane_wh(int * w,int * h,enum AVPixelFormat format,int frame_w,int frame_h,int plane)1555 static inline void get_plane_wh(int *w, int *h, enum AVPixelFormat format,
1556 int frame_w, int frame_h, int plane)
1557 {
1558 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(format);
1559
1560 /* Currently always true unless gray + alpha support is added */
1561 if (!plane || (plane == 3) || desc->flags & AV_PIX_FMT_FLAG_RGB ||
1562 !(desc->flags & AV_PIX_FMT_FLAG_PLANAR)) {
1563 *w = frame_w;
1564 *h = frame_h;
1565 return;
1566 }
1567
1568 *w = AV_CEIL_RSHIFT(frame_w, desc->log2_chroma_w);
1569 *h = AV_CEIL_RSHIFT(frame_h, desc->log2_chroma_h);
1570 }
1571
create_frame(AVHWFramesContext * hwfc,AVVkFrame ** frame,VkImageTiling tiling,VkImageUsageFlagBits usage,void * create_pnext)1572 static int create_frame(AVHWFramesContext *hwfc, AVVkFrame **frame,
1573 VkImageTiling tiling, VkImageUsageFlagBits usage,
1574 void *create_pnext)
1575 {
1576 int err;
1577 VkResult ret;
1578 AVHWDeviceContext *ctx = hwfc->device_ctx;
1579 VulkanDevicePriv *p = ctx->internal->priv;
1580 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1581 enum AVPixelFormat format = hwfc->sw_format;
1582 const VkFormat *img_fmts = av_vkfmt_from_pixfmt(format);
1583 const int planes = av_pix_fmt_count_planes(format);
1584
1585 VkExportSemaphoreCreateInfo ext_sem_info = {
1586 .sType = VK_STRUCTURE_TYPE_EXPORT_SEMAPHORE_CREATE_INFO,
1587 .handleTypes = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT,
1588 };
1589
1590 VkSemaphoreCreateInfo sem_spawn = {
1591 .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
1592 .pNext = p->extensions & EXT_EXTERNAL_FD_SEM ? &ext_sem_info : NULL,
1593 };
1594
1595 AVVkFrame *f = av_vk_frame_alloc();
1596 if (!f) {
1597 av_log(ctx, AV_LOG_ERROR, "Unable to allocate memory for AVVkFrame!\n");
1598 return AVERROR(ENOMEM);
1599 }
1600
1601 /* Create the images */
1602 for (int i = 0; i < planes; i++) {
1603 VkImageCreateInfo create_info = {
1604 .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
1605 .pNext = create_pnext,
1606 .imageType = VK_IMAGE_TYPE_2D,
1607 .format = img_fmts[i],
1608 .extent.depth = 1,
1609 .mipLevels = 1,
1610 .arrayLayers = 1,
1611 .flags = VK_IMAGE_CREATE_ALIAS_BIT,
1612 .tiling = tiling,
1613 .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED,
1614 .usage = usage,
1615 .samples = VK_SAMPLE_COUNT_1_BIT,
1616 .pQueueFamilyIndices = p->qfs,
1617 .queueFamilyIndexCount = p->num_qfs,
1618 .sharingMode = p->num_qfs > 1 ? VK_SHARING_MODE_CONCURRENT :
1619 VK_SHARING_MODE_EXCLUSIVE,
1620 };
1621
1622 get_plane_wh(&create_info.extent.width, &create_info.extent.height,
1623 format, hwfc->width, hwfc->height, i);
1624
1625 ret = vkCreateImage(hwctx->act_dev, &create_info,
1626 hwctx->alloc, &f->img[i]);
1627 if (ret != VK_SUCCESS) {
1628 av_log(ctx, AV_LOG_ERROR, "Image creation failure: %s\n",
1629 vk_ret2str(ret));
1630 err = AVERROR(EINVAL);
1631 goto fail;
1632 }
1633
1634 /* Create semaphore */
1635 ret = vkCreateSemaphore(hwctx->act_dev, &sem_spawn,
1636 hwctx->alloc, &f->sem[i]);
1637 if (ret != VK_SUCCESS) {
1638 av_log(hwctx, AV_LOG_ERROR, "Failed to create semaphore: %s\n",
1639 vk_ret2str(ret));
1640 return AVERROR_EXTERNAL;
1641 }
1642
1643 f->layout[i] = create_info.initialLayout;
1644 f->access[i] = 0x0;
1645 }
1646
1647 f->flags = 0x0;
1648 f->tiling = tiling;
1649
1650 *frame = f;
1651 return 0;
1652
1653 fail:
1654 vulkan_frame_free(hwfc, (uint8_t *)f);
1655 return err;
1656 }
1657
1658 /* Checks if an export flag is enabled, and if it is ORs it with *iexp */
try_export_flags(AVHWFramesContext * hwfc,VkExternalMemoryHandleTypeFlags * comp_handle_types,VkExternalMemoryHandleTypeFlagBits * iexp,VkExternalMemoryHandleTypeFlagBits exp)1659 static void try_export_flags(AVHWFramesContext *hwfc,
1660 VkExternalMemoryHandleTypeFlags *comp_handle_types,
1661 VkExternalMemoryHandleTypeFlagBits *iexp,
1662 VkExternalMemoryHandleTypeFlagBits exp)
1663 {
1664 VkResult ret;
1665 AVVulkanFramesContext *hwctx = hwfc->hwctx;
1666 AVVulkanDeviceContext *dev_hwctx = hwfc->device_ctx->hwctx;
1667 VkExternalImageFormatProperties eprops = {
1668 .sType = VK_STRUCTURE_TYPE_EXTERNAL_IMAGE_FORMAT_PROPERTIES_KHR,
1669 };
1670 VkImageFormatProperties2 props = {
1671 .sType = VK_STRUCTURE_TYPE_IMAGE_FORMAT_PROPERTIES_2,
1672 .pNext = &eprops,
1673 };
1674 VkPhysicalDeviceExternalImageFormatInfo enext = {
1675 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_IMAGE_FORMAT_INFO,
1676 .handleType = exp,
1677 };
1678 VkPhysicalDeviceImageFormatInfo2 pinfo = {
1679 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_FORMAT_INFO_2,
1680 .pNext = !exp ? NULL : &enext,
1681 .format = av_vkfmt_from_pixfmt(hwfc->sw_format)[0],
1682 .type = VK_IMAGE_TYPE_2D,
1683 .tiling = hwctx->tiling,
1684 .usage = hwctx->usage,
1685 .flags = VK_IMAGE_CREATE_ALIAS_BIT,
1686 };
1687
1688 ret = vkGetPhysicalDeviceImageFormatProperties2(dev_hwctx->phys_dev,
1689 &pinfo, &props);
1690 if (ret == VK_SUCCESS) {
1691 *iexp |= exp;
1692 *comp_handle_types |= eprops.externalMemoryProperties.compatibleHandleTypes;
1693 }
1694 }
1695
vulkan_pool_alloc(void * opaque,buffer_size_t size)1696 static AVBufferRef *vulkan_pool_alloc(void *opaque, buffer_size_t size)
1697 {
1698 int err;
1699 AVVkFrame *f;
1700 AVBufferRef *avbuf = NULL;
1701 AVHWFramesContext *hwfc = opaque;
1702 AVVulkanFramesContext *hwctx = hwfc->hwctx;
1703 VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
1704 VulkanFramesPriv *fp = hwfc->internal->priv;
1705 VkExportMemoryAllocateInfo eminfo[AV_NUM_DATA_POINTERS];
1706 VkExternalMemoryHandleTypeFlags e = 0x0;
1707
1708 VkExternalMemoryImageCreateInfo eiinfo = {
1709 .sType = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO,
1710 .pNext = hwctx->create_pnext,
1711 };
1712
1713 if (p->extensions & EXT_EXTERNAL_FD_MEMORY)
1714 try_export_flags(hwfc, &eiinfo.handleTypes, &e,
1715 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT);
1716
1717 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
1718 try_export_flags(hwfc, &eiinfo.handleTypes, &e,
1719 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
1720
1721 for (int i = 0; i < av_pix_fmt_count_planes(hwfc->sw_format); i++) {
1722 eminfo[i].sType = VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO;
1723 eminfo[i].pNext = hwctx->alloc_pnext[i];
1724 eminfo[i].handleTypes = e;
1725 }
1726
1727 err = create_frame(hwfc, &f, hwctx->tiling, hwctx->usage,
1728 eiinfo.handleTypes ? &eiinfo : NULL);
1729 if (err)
1730 return NULL;
1731
1732 err = alloc_bind_mem(hwfc, f, eminfo, sizeof(*eminfo));
1733 if (err)
1734 goto fail;
1735
1736 err = prepare_frame(hwfc, &fp->conv_ctx, f, PREP_MODE_WRITE);
1737 if (err)
1738 goto fail;
1739
1740 avbuf = av_buffer_create((uint8_t *)f, sizeof(AVVkFrame),
1741 vulkan_frame_free, hwfc, 0);
1742 if (!avbuf)
1743 goto fail;
1744
1745 return avbuf;
1746
1747 fail:
1748 vulkan_frame_free(hwfc, (uint8_t *)f);
1749 return NULL;
1750 }
1751
vulkan_frames_uninit(AVHWFramesContext * hwfc)1752 static void vulkan_frames_uninit(AVHWFramesContext *hwfc)
1753 {
1754 VulkanFramesPriv *fp = hwfc->internal->priv;
1755
1756 free_exec_ctx(hwfc, &fp->conv_ctx);
1757 free_exec_ctx(hwfc, &fp->upload_ctx);
1758 free_exec_ctx(hwfc, &fp->download_ctx);
1759 }
1760
vulkan_frames_init(AVHWFramesContext * hwfc)1761 static int vulkan_frames_init(AVHWFramesContext *hwfc)
1762 {
1763 int err;
1764 AVVkFrame *f;
1765 AVVulkanFramesContext *hwctx = hwfc->hwctx;
1766 VulkanFramesPriv *fp = hwfc->internal->priv;
1767 AVVulkanDeviceContext *dev_hwctx = hwfc->device_ctx->hwctx;
1768 VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
1769
1770 /* Default pool flags */
1771 hwctx->tiling = hwctx->tiling ? hwctx->tiling : p->use_linear_images ?
1772 VK_IMAGE_TILING_LINEAR : VK_IMAGE_TILING_OPTIMAL;
1773
1774 if (!hwctx->usage)
1775 hwctx->usage = DEFAULT_USAGE_FLAGS;
1776
1777 err = create_exec_ctx(hwfc, &fp->conv_ctx,
1778 dev_hwctx->queue_family_comp_index,
1779 GET_QUEUE_COUNT(dev_hwctx, 0, 1, 0));
1780 if (err)
1781 return err;
1782
1783 err = create_exec_ctx(hwfc, &fp->upload_ctx,
1784 dev_hwctx->queue_family_tx_index,
1785 GET_QUEUE_COUNT(dev_hwctx, 0, 0, 1));
1786 if (err)
1787 return err;
1788
1789 err = create_exec_ctx(hwfc, &fp->download_ctx,
1790 dev_hwctx->queue_family_tx_index, 1);
1791 if (err)
1792 return err;
1793
1794 /* Test to see if allocation will fail */
1795 err = create_frame(hwfc, &f, hwctx->tiling, hwctx->usage,
1796 hwctx->create_pnext);
1797 if (err)
1798 return err;
1799
1800 vulkan_frame_free(hwfc, (uint8_t *)f);
1801
1802 /* If user did not specify a pool, hwfc->pool will be set to the internal one
1803 * in hwcontext.c just after this gets called */
1804 if (!hwfc->pool) {
1805 hwfc->internal->pool_internal = av_buffer_pool_init2(sizeof(AVVkFrame),
1806 hwfc, vulkan_pool_alloc,
1807 NULL);
1808 if (!hwfc->internal->pool_internal)
1809 return AVERROR(ENOMEM);
1810 }
1811
1812 return 0;
1813 }
1814
vulkan_get_buffer(AVHWFramesContext * hwfc,AVFrame * frame)1815 static int vulkan_get_buffer(AVHWFramesContext *hwfc, AVFrame *frame)
1816 {
1817 frame->buf[0] = av_buffer_pool_get(hwfc->pool);
1818 if (!frame->buf[0])
1819 return AVERROR(ENOMEM);
1820
1821 frame->data[0] = frame->buf[0]->data;
1822 frame->format = AV_PIX_FMT_VULKAN;
1823 frame->width = hwfc->width;
1824 frame->height = hwfc->height;
1825
1826 return 0;
1827 }
1828
vulkan_transfer_get_formats(AVHWFramesContext * hwfc,enum AVHWFrameTransferDirection dir,enum AVPixelFormat ** formats)1829 static int vulkan_transfer_get_formats(AVHWFramesContext *hwfc,
1830 enum AVHWFrameTransferDirection dir,
1831 enum AVPixelFormat **formats)
1832 {
1833 enum AVPixelFormat *fmts = av_malloc_array(2, sizeof(*fmts));
1834 if (!fmts)
1835 return AVERROR(ENOMEM);
1836
1837 fmts[0] = hwfc->sw_format;
1838 fmts[1] = AV_PIX_FMT_NONE;
1839
1840 *formats = fmts;
1841 return 0;
1842 }
1843
1844 typedef struct VulkanMapping {
1845 AVVkFrame *frame;
1846 int flags;
1847 } VulkanMapping;
1848
vulkan_unmap_frame(AVHWFramesContext * hwfc,HWMapDescriptor * hwmap)1849 static void vulkan_unmap_frame(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
1850 {
1851 VulkanMapping *map = hwmap->priv;
1852 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1853 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1854
1855 /* Check if buffer needs flushing */
1856 if ((map->flags & AV_HWFRAME_MAP_WRITE) &&
1857 !(map->frame->flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)) {
1858 VkResult ret;
1859 VkMappedMemoryRange flush_ranges[AV_NUM_DATA_POINTERS] = { { 0 } };
1860
1861 for (int i = 0; i < planes; i++) {
1862 flush_ranges[i].sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
1863 flush_ranges[i].memory = map->frame->mem[i];
1864 flush_ranges[i].size = VK_WHOLE_SIZE;
1865 }
1866
1867 ret = vkFlushMappedMemoryRanges(hwctx->act_dev, planes,
1868 flush_ranges);
1869 if (ret != VK_SUCCESS) {
1870 av_log(hwfc, AV_LOG_ERROR, "Failed to flush memory: %s\n",
1871 vk_ret2str(ret));
1872 }
1873 }
1874
1875 for (int i = 0; i < planes; i++)
1876 vkUnmapMemory(hwctx->act_dev, map->frame->mem[i]);
1877
1878 av_free(map);
1879 }
1880
vulkan_map_frame_to_mem(AVHWFramesContext * hwfc,AVFrame * dst,const AVFrame * src,int flags)1881 static int vulkan_map_frame_to_mem(AVHWFramesContext *hwfc, AVFrame *dst,
1882 const AVFrame *src, int flags)
1883 {
1884 VkResult ret;
1885 int err, mapped_mem_count = 0;
1886 AVVkFrame *f = (AVVkFrame *)src->data[0];
1887 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1888 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1889
1890 VulkanMapping *map = av_mallocz(sizeof(VulkanMapping));
1891 if (!map)
1892 return AVERROR(EINVAL);
1893
1894 if (src->format != AV_PIX_FMT_VULKAN) {
1895 av_log(hwfc, AV_LOG_ERROR, "Cannot map from pixel format %s!\n",
1896 av_get_pix_fmt_name(src->format));
1897 err = AVERROR(EINVAL);
1898 goto fail;
1899 }
1900
1901 if (!(f->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) ||
1902 !(f->tiling == VK_IMAGE_TILING_LINEAR)) {
1903 av_log(hwfc, AV_LOG_ERROR, "Unable to map frame, not host visible "
1904 "and linear!\n");
1905 err = AVERROR(EINVAL);
1906 goto fail;
1907 }
1908
1909 dst->width = src->width;
1910 dst->height = src->height;
1911
1912 for (int i = 0; i < planes; i++) {
1913 ret = vkMapMemory(hwctx->act_dev, f->mem[i], 0,
1914 VK_WHOLE_SIZE, 0, (void **)&dst->data[i]);
1915 if (ret != VK_SUCCESS) {
1916 av_log(hwfc, AV_LOG_ERROR, "Failed to map image memory: %s\n",
1917 vk_ret2str(ret));
1918 err = AVERROR_EXTERNAL;
1919 goto fail;
1920 }
1921 mapped_mem_count++;
1922 }
1923
1924 /* Check if the memory contents matter */
1925 if (((flags & AV_HWFRAME_MAP_READ) || !(flags & AV_HWFRAME_MAP_OVERWRITE)) &&
1926 !(f->flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)) {
1927 VkMappedMemoryRange map_mem_ranges[AV_NUM_DATA_POINTERS] = { { 0 } };
1928 for (int i = 0; i < planes; i++) {
1929 map_mem_ranges[i].sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
1930 map_mem_ranges[i].size = VK_WHOLE_SIZE;
1931 map_mem_ranges[i].memory = f->mem[i];
1932 }
1933
1934 ret = vkInvalidateMappedMemoryRanges(hwctx->act_dev, planes,
1935 map_mem_ranges);
1936 if (ret != VK_SUCCESS) {
1937 av_log(hwfc, AV_LOG_ERROR, "Failed to invalidate memory: %s\n",
1938 vk_ret2str(ret));
1939 err = AVERROR_EXTERNAL;
1940 goto fail;
1941 }
1942 }
1943
1944 for (int i = 0; i < planes; i++) {
1945 VkImageSubresource sub = {
1946 .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
1947 };
1948 VkSubresourceLayout layout;
1949 vkGetImageSubresourceLayout(hwctx->act_dev, f->img[i], &sub, &layout);
1950 dst->linesize[i] = layout.rowPitch;
1951 }
1952
1953 map->frame = f;
1954 map->flags = flags;
1955
1956 err = ff_hwframe_map_create(src->hw_frames_ctx, dst, src,
1957 &vulkan_unmap_frame, map);
1958 if (err < 0)
1959 goto fail;
1960
1961 return 0;
1962
1963 fail:
1964 for (int i = 0; i < mapped_mem_count; i++)
1965 vkUnmapMemory(hwctx->act_dev, f->mem[i]);
1966
1967 av_free(map);
1968 return err;
1969 }
1970
1971 #if CONFIG_LIBDRM
vulkan_unmap_from(AVHWFramesContext * hwfc,HWMapDescriptor * hwmap)1972 static void vulkan_unmap_from(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
1973 {
1974 VulkanMapping *map = hwmap->priv;
1975 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1976 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1977
1978 for (int i = 0; i < planes; i++) {
1979 vkDestroyImage(hwctx->act_dev, map->frame->img[i], hwctx->alloc);
1980 vkFreeMemory(hwctx->act_dev, map->frame->mem[i], hwctx->alloc);
1981 vkDestroySemaphore(hwctx->act_dev, map->frame->sem[i], hwctx->alloc);
1982 }
1983
1984 av_freep(&map->frame);
1985 }
1986
1987 static const struct {
1988 uint32_t drm_fourcc;
1989 VkFormat vk_format;
1990 } vulkan_drm_format_map[] = {
1991 { DRM_FORMAT_R8, VK_FORMAT_R8_UNORM },
1992 { DRM_FORMAT_R16, VK_FORMAT_R16_UNORM },
1993 { DRM_FORMAT_GR88, VK_FORMAT_R8G8_UNORM },
1994 { DRM_FORMAT_RG88, VK_FORMAT_R8G8_UNORM },
1995 { DRM_FORMAT_GR1616, VK_FORMAT_R16G16_UNORM },
1996 { DRM_FORMAT_RG1616, VK_FORMAT_R16G16_UNORM },
1997 { DRM_FORMAT_ARGB8888, VK_FORMAT_B8G8R8A8_UNORM },
1998 { DRM_FORMAT_XRGB8888, VK_FORMAT_B8G8R8A8_UNORM },
1999 { DRM_FORMAT_ABGR8888, VK_FORMAT_R8G8B8A8_UNORM },
2000 { DRM_FORMAT_XBGR8888, VK_FORMAT_R8G8B8A8_UNORM },
2001 };
2002
drm_to_vulkan_fmt(uint32_t drm_fourcc)2003 static inline VkFormat drm_to_vulkan_fmt(uint32_t drm_fourcc)
2004 {
2005 for (int i = 0; i < FF_ARRAY_ELEMS(vulkan_drm_format_map); i++)
2006 if (vulkan_drm_format_map[i].drm_fourcc == drm_fourcc)
2007 return vulkan_drm_format_map[i].vk_format;
2008 return VK_FORMAT_UNDEFINED;
2009 }
2010
vulkan_map_from_drm_frame_desc(AVHWFramesContext * hwfc,AVVkFrame ** frame,const AVFrame * src)2011 static int vulkan_map_from_drm_frame_desc(AVHWFramesContext *hwfc, AVVkFrame **frame,
2012 const AVFrame *src)
2013 {
2014 int err = 0;
2015 VkResult ret;
2016 AVVkFrame *f;
2017 int bind_counts = 0;
2018 AVHWDeviceContext *ctx = hwfc->device_ctx;
2019 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2020 VulkanDevicePriv *p = ctx->internal->priv;
2021 VulkanFramesPriv *fp = hwfc->internal->priv;
2022 AVVulkanFramesContext *frames_hwctx = hwfc->hwctx;
2023 const AVDRMFrameDescriptor *desc = (AVDRMFrameDescriptor *)src->data[0];
2024 const int has_modifiers = !!(p->extensions & EXT_DRM_MODIFIER_FLAGS);
2025 VkSubresourceLayout plane_data[AV_NUM_DATA_POINTERS] = { 0 };
2026 VkBindImageMemoryInfo bind_info[AV_NUM_DATA_POINTERS] = { 0 };
2027 VkBindImagePlaneMemoryInfo plane_info[AV_NUM_DATA_POINTERS] = { 0 };
2028 VkExternalMemoryHandleTypeFlagBits htype = VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT;
2029
2030 VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdPropertiesKHR);
2031
2032 for (int i = 0; i < desc->nb_layers; i++) {
2033 if (drm_to_vulkan_fmt(desc->layers[i].format) == VK_FORMAT_UNDEFINED) {
2034 av_log(ctx, AV_LOG_ERROR, "Unsupported DMABUF layer format %#08x!\n",
2035 desc->layers[i].format);
2036 return AVERROR(EINVAL);
2037 }
2038 }
2039
2040 if (!(f = av_vk_frame_alloc())) {
2041 av_log(ctx, AV_LOG_ERROR, "Unable to allocate memory for AVVkFrame!\n");
2042 err = AVERROR(ENOMEM);
2043 goto fail;
2044 }
2045
2046 f->tiling = has_modifiers ? VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT :
2047 desc->objects[0].format_modifier == DRM_FORMAT_MOD_LINEAR ?
2048 VK_IMAGE_TILING_LINEAR : VK_IMAGE_TILING_OPTIMAL;
2049
2050 for (int i = 0; i < desc->nb_layers; i++) {
2051 const int planes = desc->layers[i].nb_planes;
2052 VkImageDrmFormatModifierExplicitCreateInfoEXT drm_info = {
2053 .sType = VK_STRUCTURE_TYPE_IMAGE_DRM_FORMAT_MODIFIER_EXPLICIT_CREATE_INFO_EXT,
2054 .drmFormatModifier = desc->objects[0].format_modifier,
2055 .drmFormatModifierPlaneCount = planes,
2056 .pPlaneLayouts = (const VkSubresourceLayout *)&plane_data,
2057 };
2058
2059 VkExternalMemoryImageCreateInfo einfo = {
2060 .sType = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO,
2061 .pNext = has_modifiers ? &drm_info : NULL,
2062 .handleTypes = htype,
2063 };
2064
2065 VkSemaphoreCreateInfo sem_spawn = {
2066 .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
2067 };
2068
2069 VkImageCreateInfo create_info = {
2070 .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
2071 .pNext = &einfo,
2072 .imageType = VK_IMAGE_TYPE_2D,
2073 .format = drm_to_vulkan_fmt(desc->layers[i].format),
2074 .extent.depth = 1,
2075 .mipLevels = 1,
2076 .arrayLayers = 1,
2077 .flags = VK_IMAGE_CREATE_ALIAS_BIT,
2078 .tiling = f->tiling,
2079 .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED, /* specs say so */
2080 .usage = frames_hwctx->usage,
2081 .samples = VK_SAMPLE_COUNT_1_BIT,
2082 .pQueueFamilyIndices = p->qfs,
2083 .queueFamilyIndexCount = p->num_qfs,
2084 .sharingMode = p->num_qfs > 1 ? VK_SHARING_MODE_CONCURRENT :
2085 VK_SHARING_MODE_EXCLUSIVE,
2086 };
2087
2088 get_plane_wh(&create_info.extent.width, &create_info.extent.height,
2089 hwfc->sw_format, src->width, src->height, i);
2090
2091 for (int j = 0; j < planes; j++) {
2092 plane_data[j].offset = desc->layers[i].planes[j].offset;
2093 plane_data[j].rowPitch = desc->layers[i].planes[j].pitch;
2094 plane_data[j].size = 0; /* The specs say so for all 3 */
2095 plane_data[j].arrayPitch = 0;
2096 plane_data[j].depthPitch = 0;
2097 }
2098
2099 /* Create image */
2100 ret = vkCreateImage(hwctx->act_dev, &create_info,
2101 hwctx->alloc, &f->img[i]);
2102 if (ret != VK_SUCCESS) {
2103 av_log(ctx, AV_LOG_ERROR, "Image creation failure: %s\n",
2104 vk_ret2str(ret));
2105 err = AVERROR(EINVAL);
2106 goto fail;
2107 }
2108
2109 ret = vkCreateSemaphore(hwctx->act_dev, &sem_spawn,
2110 hwctx->alloc, &f->sem[i]);
2111 if (ret != VK_SUCCESS) {
2112 av_log(hwctx, AV_LOG_ERROR, "Failed to create semaphore: %s\n",
2113 vk_ret2str(ret));
2114 return AVERROR_EXTERNAL;
2115 }
2116
2117 /* We'd import a semaphore onto the one we created using
2118 * vkImportSemaphoreFdKHR but unfortunately neither DRM nor VAAPI
2119 * offer us anything we could import and sync with, so instead
2120 * just signal the semaphore we created. */
2121
2122 f->layout[i] = create_info.initialLayout;
2123 f->access[i] = 0x0;
2124 }
2125
2126 for (int i = 0; i < desc->nb_objects; i++) {
2127 int use_ded_mem = 0;
2128 VkMemoryFdPropertiesKHR fdmp = {
2129 .sType = VK_STRUCTURE_TYPE_MEMORY_FD_PROPERTIES_KHR,
2130 };
2131 VkMemoryRequirements req = {
2132 .size = desc->objects[i].size,
2133 };
2134 VkImportMemoryFdInfoKHR idesc = {
2135 .sType = VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR,
2136 .handleType = htype,
2137 .fd = dup(desc->objects[i].fd),
2138 };
2139 VkMemoryDedicatedAllocateInfo ded_alloc = {
2140 .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO,
2141 .pNext = &idesc,
2142 };
2143
2144 ret = pfn_vkGetMemoryFdPropertiesKHR(hwctx->act_dev, htype,
2145 idesc.fd, &fdmp);
2146 if (ret != VK_SUCCESS) {
2147 av_log(hwfc, AV_LOG_ERROR, "Failed to get FD properties: %s\n",
2148 vk_ret2str(ret));
2149 err = AVERROR_EXTERNAL;
2150 close(idesc.fd);
2151 goto fail;
2152 }
2153
2154 req.memoryTypeBits = fdmp.memoryTypeBits;
2155
2156 /* Dedicated allocation only makes sense if there's a one to one mapping
2157 * between images and the memory backing them, so only check in this
2158 * case. */
2159 if (desc->nb_layers == desc->nb_objects) {
2160 VkImageMemoryRequirementsInfo2 req_desc = {
2161 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_REQUIREMENTS_INFO_2,
2162 .image = f->img[i],
2163 };
2164 VkMemoryDedicatedRequirements ded_req = {
2165 .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS,
2166 };
2167 VkMemoryRequirements2 req2 = {
2168 .sType = VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2,
2169 .pNext = &ded_req,
2170 };
2171
2172 vkGetImageMemoryRequirements2(hwctx->act_dev, &req_desc, &req2);
2173
2174 use_ded_mem = ded_req.prefersDedicatedAllocation |
2175 ded_req.requiresDedicatedAllocation;
2176 if (use_ded_mem)
2177 ded_alloc.image = f->img[i];
2178 }
2179
2180 err = alloc_mem(ctx, &req, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
2181 use_ded_mem ? &ded_alloc : ded_alloc.pNext,
2182 &f->flags, &f->mem[i]);
2183 if (err) {
2184 close(idesc.fd);
2185 return err;
2186 }
2187
2188 f->size[i] = desc->objects[i].size;
2189 }
2190
2191 for (int i = 0; i < desc->nb_layers; i++) {
2192 const int planes = desc->layers[i].nb_planes;
2193 const int signal_p = has_modifiers && (planes > 1);
2194 for (int j = 0; j < planes; j++) {
2195 VkImageAspectFlagBits aspect = j == 0 ? VK_IMAGE_ASPECT_MEMORY_PLANE_0_BIT_EXT :
2196 j == 1 ? VK_IMAGE_ASPECT_MEMORY_PLANE_1_BIT_EXT :
2197 VK_IMAGE_ASPECT_MEMORY_PLANE_2_BIT_EXT;
2198
2199 plane_info[bind_counts].sType = VK_STRUCTURE_TYPE_BIND_IMAGE_PLANE_MEMORY_INFO;
2200 plane_info[bind_counts].planeAspect = aspect;
2201
2202 bind_info[bind_counts].sType = VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO;
2203 bind_info[bind_counts].pNext = signal_p ? &plane_info[bind_counts] : NULL;
2204 bind_info[bind_counts].image = f->img[i];
2205 bind_info[bind_counts].memory = f->mem[desc->layers[i].planes[j].object_index];
2206 bind_info[bind_counts].memoryOffset = desc->layers[i].planes[j].offset;
2207 bind_counts++;
2208 }
2209 }
2210
2211 /* Bind the allocated memory to the images */
2212 ret = vkBindImageMemory2(hwctx->act_dev, bind_counts, bind_info);
2213 if (ret != VK_SUCCESS) {
2214 av_log(ctx, AV_LOG_ERROR, "Failed to bind memory: %s\n",
2215 vk_ret2str(ret));
2216 return AVERROR_EXTERNAL;
2217 }
2218
2219 /* NOTE: This is completely uneccesary and unneeded once we can import
2220 * semaphores from DRM. Otherwise we have to activate the semaphores.
2221 * We're reusing the exec context that's also used for uploads/downloads. */
2222 err = prepare_frame(hwfc, &fp->conv_ctx, f, PREP_MODE_RO_SHADER);
2223 if (err)
2224 goto fail;
2225
2226 *frame = f;
2227
2228 return 0;
2229
2230 fail:
2231 for (int i = 0; i < desc->nb_layers; i++) {
2232 vkDestroyImage(hwctx->act_dev, f->img[i], hwctx->alloc);
2233 vkDestroySemaphore(hwctx->act_dev, f->sem[i], hwctx->alloc);
2234 }
2235 for (int i = 0; i < desc->nb_objects; i++)
2236 vkFreeMemory(hwctx->act_dev, f->mem[i], hwctx->alloc);
2237
2238 av_free(f);
2239
2240 return err;
2241 }
2242
vulkan_map_from_drm(AVHWFramesContext * hwfc,AVFrame * dst,const AVFrame * src,int flags)2243 static int vulkan_map_from_drm(AVHWFramesContext *hwfc, AVFrame *dst,
2244 const AVFrame *src, int flags)
2245 {
2246 int err = 0;
2247 AVVkFrame *f;
2248 VulkanMapping *map = NULL;
2249
2250 if ((err = vulkan_map_from_drm_frame_desc(hwfc, &f, src)))
2251 return err;
2252
2253 /* The unmapping function will free this */
2254 dst->data[0] = (uint8_t *)f;
2255 dst->width = src->width;
2256 dst->height = src->height;
2257
2258 map = av_mallocz(sizeof(VulkanMapping));
2259 if (!map)
2260 goto fail;
2261
2262 map->frame = f;
2263 map->flags = flags;
2264
2265 err = ff_hwframe_map_create(dst->hw_frames_ctx, dst, src,
2266 &vulkan_unmap_from, map);
2267 if (err < 0)
2268 goto fail;
2269
2270 av_log(hwfc, AV_LOG_DEBUG, "Mapped DRM object to Vulkan!\n");
2271
2272 return 0;
2273
2274 fail:
2275 vulkan_frame_free(hwfc->device_ctx->hwctx, (uint8_t *)f);
2276 av_free(map);
2277 return err;
2278 }
2279
2280 #if CONFIG_VAAPI
vulkan_map_from_vaapi(AVHWFramesContext * dst_fc,AVFrame * dst,const AVFrame * src,int flags)2281 static int vulkan_map_from_vaapi(AVHWFramesContext *dst_fc,
2282 AVFrame *dst, const AVFrame *src,
2283 int flags)
2284 {
2285 int err;
2286 AVFrame *tmp = av_frame_alloc();
2287 AVHWFramesContext *vaapi_fc = (AVHWFramesContext*)src->hw_frames_ctx->data;
2288 AVVAAPIDeviceContext *vaapi_ctx = vaapi_fc->device_ctx->hwctx;
2289 VASurfaceID surface_id = (VASurfaceID)(uintptr_t)src->data[3];
2290
2291 if (!tmp)
2292 return AVERROR(ENOMEM);
2293
2294 /* We have to sync since like the previous comment said, no semaphores */
2295 vaSyncSurface(vaapi_ctx->display, surface_id);
2296
2297 tmp->format = AV_PIX_FMT_DRM_PRIME;
2298
2299 err = av_hwframe_map(tmp, src, flags);
2300 if (err < 0)
2301 goto fail;
2302
2303 err = vulkan_map_from_drm(dst_fc, dst, tmp, flags);
2304 if (err < 0)
2305 goto fail;
2306
2307 err = ff_hwframe_map_replace(dst, src);
2308
2309 fail:
2310 av_frame_free(&tmp);
2311 return err;
2312 }
2313 #endif
2314 #endif
2315
2316 #if CONFIG_CUDA
vulkan_export_to_cuda(AVHWFramesContext * hwfc,AVBufferRef * cuda_hwfc,const AVFrame * frame)2317 static int vulkan_export_to_cuda(AVHWFramesContext *hwfc,
2318 AVBufferRef *cuda_hwfc,
2319 const AVFrame *frame)
2320 {
2321 int err;
2322 VkResult ret;
2323 AVVkFrame *dst_f;
2324 AVVkFrameInternal *dst_int;
2325 AVHWDeviceContext *ctx = hwfc->device_ctx;
2326 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2327 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
2328 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
2329 VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdKHR);
2330 VK_LOAD_PFN(hwctx->inst, vkGetSemaphoreFdKHR);
2331
2332 AVHWFramesContext *cuda_fc = (AVHWFramesContext*)cuda_hwfc->data;
2333 AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
2334 AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
2335 AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
2336 CudaFunctions *cu = cu_internal->cuda_dl;
2337 CUarray_format cufmt = desc->comp[0].depth > 8 ? CU_AD_FORMAT_UNSIGNED_INT16 :
2338 CU_AD_FORMAT_UNSIGNED_INT8;
2339
2340 dst_f = (AVVkFrame *)frame->data[0];
2341
2342 dst_int = dst_f->internal;
2343 if (!dst_int || !dst_int->cuda_fc_ref) {
2344 if (!dst_f->internal)
2345 dst_f->internal = dst_int = av_mallocz(sizeof(*dst_f->internal));
2346
2347 if (!dst_int) {
2348 err = AVERROR(ENOMEM);
2349 goto fail;
2350 }
2351
2352 dst_int->cuda_fc_ref = av_buffer_ref(cuda_hwfc);
2353 if (!dst_int->cuda_fc_ref) {
2354 err = AVERROR(ENOMEM);
2355 goto fail;
2356 }
2357
2358 for (int i = 0; i < planes; i++) {
2359 CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC tex_desc = {
2360 .offset = 0,
2361 .arrayDesc = {
2362 .Depth = 0,
2363 .Format = cufmt,
2364 .NumChannels = 1 + ((planes == 2) && i),
2365 .Flags = 0,
2366 },
2367 .numLevels = 1,
2368 };
2369 CUDA_EXTERNAL_MEMORY_HANDLE_DESC ext_desc = {
2370 .type = CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD,
2371 .size = dst_f->size[i],
2372 };
2373 VkMemoryGetFdInfoKHR export_info = {
2374 .sType = VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR,
2375 .memory = dst_f->mem[i],
2376 .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR,
2377 };
2378 VkSemaphoreGetFdInfoKHR sem_export = {
2379 .sType = VK_STRUCTURE_TYPE_SEMAPHORE_GET_FD_INFO_KHR,
2380 .semaphore = dst_f->sem[i],
2381 .handleType = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT,
2382 };
2383 CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC ext_sem_desc = {
2384 .type = CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD,
2385 };
2386
2387 int p_w, p_h;
2388 get_plane_wh(&p_w, &p_h, hwfc->sw_format, hwfc->width, hwfc->height, i);
2389
2390 tex_desc.arrayDesc.Width = p_w;
2391 tex_desc.arrayDesc.Height = p_h;
2392
2393 ret = pfn_vkGetMemoryFdKHR(hwctx->act_dev, &export_info,
2394 &ext_desc.handle.fd);
2395 if (ret != VK_SUCCESS) {
2396 av_log(hwfc, AV_LOG_ERROR, "Unable to export the image as a FD!\n");
2397 err = AVERROR_EXTERNAL;
2398 goto fail;
2399 }
2400
2401 ret = CHECK_CU(cu->cuImportExternalMemory(&dst_int->ext_mem[i], &ext_desc));
2402 if (ret < 0) {
2403 err = AVERROR_EXTERNAL;
2404 goto fail;
2405 }
2406
2407 ret = CHECK_CU(cu->cuExternalMemoryGetMappedMipmappedArray(&dst_int->cu_mma[i],
2408 dst_int->ext_mem[i],
2409 &tex_desc));
2410 if (ret < 0) {
2411 err = AVERROR_EXTERNAL;
2412 goto fail;
2413 }
2414
2415 ret = CHECK_CU(cu->cuMipmappedArrayGetLevel(&dst_int->cu_array[i],
2416 dst_int->cu_mma[i], 0));
2417 if (ret < 0) {
2418 err = AVERROR_EXTERNAL;
2419 goto fail;
2420 }
2421
2422 ret = pfn_vkGetSemaphoreFdKHR(hwctx->act_dev, &sem_export,
2423 &ext_sem_desc.handle.fd);
2424 if (ret != VK_SUCCESS) {
2425 av_log(ctx, AV_LOG_ERROR, "Failed to export semaphore: %s\n",
2426 vk_ret2str(ret));
2427 err = AVERROR_EXTERNAL;
2428 goto fail;
2429 }
2430
2431 ret = CHECK_CU(cu->cuImportExternalSemaphore(&dst_int->cu_sem[i],
2432 &ext_sem_desc));
2433 if (ret < 0) {
2434 err = AVERROR_EXTERNAL;
2435 goto fail;
2436 }
2437 }
2438 }
2439
2440 return 0;
2441
2442 fail:
2443 return err;
2444 }
2445
vulkan_transfer_data_from_cuda(AVHWFramesContext * hwfc,AVFrame * dst,const AVFrame * src)2446 static int vulkan_transfer_data_from_cuda(AVHWFramesContext *hwfc,
2447 AVFrame *dst, const AVFrame *src)
2448 {
2449 int err;
2450 VkResult ret;
2451 CUcontext dummy;
2452 AVVkFrame *dst_f;
2453 AVVkFrameInternal *dst_int;
2454 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
2455 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
2456
2457 AVHWFramesContext *cuda_fc = (AVHWFramesContext*)src->hw_frames_ctx->data;
2458 AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
2459 AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
2460 AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
2461 CudaFunctions *cu = cu_internal->cuda_dl;
2462 CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS s_w_par[AV_NUM_DATA_POINTERS] = { 0 };
2463 CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS s_s_par[AV_NUM_DATA_POINTERS] = { 0 };
2464
2465 ret = CHECK_CU(cu->cuCtxPushCurrent(cuda_dev->cuda_ctx));
2466 if (ret < 0)
2467 return AVERROR_EXTERNAL;
2468
2469 dst_f = (AVVkFrame *)dst->data[0];
2470
2471 ret = vulkan_export_to_cuda(hwfc, src->hw_frames_ctx, dst);
2472 if (ret < 0) {
2473 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2474 return ret;
2475 }
2476
2477 dst_int = dst_f->internal;
2478
2479 ret = CHECK_CU(cu->cuWaitExternalSemaphoresAsync(dst_int->cu_sem, s_w_par,
2480 planes, cuda_dev->stream));
2481 if (ret < 0) {
2482 err = AVERROR_EXTERNAL;
2483 goto fail;
2484 }
2485
2486 for (int i = 0; i < planes; i++) {
2487 CUDA_MEMCPY2D cpy = {
2488 .srcMemoryType = CU_MEMORYTYPE_DEVICE,
2489 .srcDevice = (CUdeviceptr)src->data[i],
2490 .srcPitch = src->linesize[i],
2491 .srcY = 0,
2492
2493 .dstMemoryType = CU_MEMORYTYPE_ARRAY,
2494 .dstArray = dst_int->cu_array[i],
2495 };
2496
2497 int p_w, p_h;
2498 get_plane_wh(&p_w, &p_h, hwfc->sw_format, hwfc->width, hwfc->height, i);
2499
2500 cpy.WidthInBytes = p_w * desc->comp[i].step;
2501 cpy.Height = p_h;
2502
2503 ret = CHECK_CU(cu->cuMemcpy2DAsync(&cpy, cuda_dev->stream));
2504 if (ret < 0) {
2505 err = AVERROR_EXTERNAL;
2506 goto fail;
2507 }
2508 }
2509
2510 ret = CHECK_CU(cu->cuSignalExternalSemaphoresAsync(dst_int->cu_sem, s_s_par,
2511 planes, cuda_dev->stream));
2512 if (ret < 0) {
2513 err = AVERROR_EXTERNAL;
2514 goto fail;
2515 }
2516
2517 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2518
2519 av_log(hwfc, AV_LOG_VERBOSE, "Transfered CUDA image to Vulkan!\n");
2520
2521 return 0;
2522
2523 fail:
2524 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2525 vulkan_free_internal(dst_int);
2526 dst_f->internal = NULL;
2527 av_buffer_unref(&dst->buf[0]);
2528 return err;
2529 }
2530 #endif
2531
vulkan_map_to(AVHWFramesContext * hwfc,AVFrame * dst,const AVFrame * src,int flags)2532 static int vulkan_map_to(AVHWFramesContext *hwfc, AVFrame *dst,
2533 const AVFrame *src, int flags)
2534 {
2535 av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2536
2537 switch (src->format) {
2538 #if CONFIG_LIBDRM
2539 #if CONFIG_VAAPI
2540 case AV_PIX_FMT_VAAPI:
2541 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2542 return vulkan_map_from_vaapi(hwfc, dst, src, flags);
2543 #endif
2544 case AV_PIX_FMT_DRM_PRIME:
2545 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2546 return vulkan_map_from_drm(hwfc, dst, src, flags);
2547 #endif
2548 default:
2549 return AVERROR(ENOSYS);
2550 }
2551 }
2552
2553 #if CONFIG_LIBDRM
2554 typedef struct VulkanDRMMapping {
2555 AVDRMFrameDescriptor drm_desc;
2556 AVVkFrame *source;
2557 } VulkanDRMMapping;
2558
vulkan_unmap_to_drm(AVHWFramesContext * hwfc,HWMapDescriptor * hwmap)2559 static void vulkan_unmap_to_drm(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
2560 {
2561 AVDRMFrameDescriptor *drm_desc = hwmap->priv;
2562
2563 for (int i = 0; i < drm_desc->nb_objects; i++)
2564 close(drm_desc->objects[i].fd);
2565
2566 av_free(drm_desc);
2567 }
2568
vulkan_fmt_to_drm(VkFormat vkfmt)2569 static inline uint32_t vulkan_fmt_to_drm(VkFormat vkfmt)
2570 {
2571 for (int i = 0; i < FF_ARRAY_ELEMS(vulkan_drm_format_map); i++)
2572 if (vulkan_drm_format_map[i].vk_format == vkfmt)
2573 return vulkan_drm_format_map[i].drm_fourcc;
2574 return DRM_FORMAT_INVALID;
2575 }
2576
vulkan_map_to_drm(AVHWFramesContext * hwfc,AVFrame * dst,const AVFrame * src,int flags)2577 static int vulkan_map_to_drm(AVHWFramesContext *hwfc, AVFrame *dst,
2578 const AVFrame *src, int flags)
2579 {
2580 int err = 0;
2581 VkResult ret;
2582 AVVkFrame *f = (AVVkFrame *)src->data[0];
2583 VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2584 VulkanFramesPriv *fp = hwfc->internal->priv;
2585 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
2586 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
2587 VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdKHR);
2588 VkImageDrmFormatModifierPropertiesEXT drm_mod = {
2589 .sType = VK_STRUCTURE_TYPE_IMAGE_DRM_FORMAT_MODIFIER_PROPERTIES_EXT,
2590 };
2591
2592 AVDRMFrameDescriptor *drm_desc = av_mallocz(sizeof(*drm_desc));
2593 if (!drm_desc)
2594 return AVERROR(ENOMEM);
2595
2596 err = prepare_frame(hwfc, &fp->conv_ctx, f, PREP_MODE_EXTERNAL_EXPORT);
2597 if (err < 0)
2598 goto end;
2599
2600 err = ff_hwframe_map_create(src->hw_frames_ctx, dst, src, &vulkan_unmap_to_drm, drm_desc);
2601 if (err < 0)
2602 goto end;
2603
2604 if (p->extensions & EXT_DRM_MODIFIER_FLAGS) {
2605 VK_LOAD_PFN(hwctx->inst, vkGetImageDrmFormatModifierPropertiesEXT);
2606 ret = pfn_vkGetImageDrmFormatModifierPropertiesEXT(hwctx->act_dev, f->img[0],
2607 &drm_mod);
2608 if (ret != VK_SUCCESS) {
2609 av_log(hwfc, AV_LOG_ERROR, "Failed to retrieve DRM format modifier!\n");
2610 err = AVERROR_EXTERNAL;
2611 goto end;
2612 }
2613 }
2614
2615 for (int i = 0; (i < planes) && (f->mem[i]); i++) {
2616 VkMemoryGetFdInfoKHR export_info = {
2617 .sType = VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR,
2618 .memory = f->mem[i],
2619 .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT,
2620 };
2621
2622 ret = pfn_vkGetMemoryFdKHR(hwctx->act_dev, &export_info,
2623 &drm_desc->objects[i].fd);
2624 if (ret != VK_SUCCESS) {
2625 av_log(hwfc, AV_LOG_ERROR, "Unable to export the image as a FD!\n");
2626 err = AVERROR_EXTERNAL;
2627 goto end;
2628 }
2629
2630 drm_desc->nb_objects++;
2631 drm_desc->objects[i].size = f->size[i];
2632 drm_desc->objects[i].format_modifier = drm_mod.drmFormatModifier;
2633 }
2634
2635 drm_desc->nb_layers = planes;
2636 for (int i = 0; i < drm_desc->nb_layers; i++) {
2637 VkSubresourceLayout layout;
2638 VkImageSubresource sub = {
2639 .aspectMask = p->extensions & EXT_DRM_MODIFIER_FLAGS ?
2640 VK_IMAGE_ASPECT_MEMORY_PLANE_0_BIT_EXT :
2641 VK_IMAGE_ASPECT_COLOR_BIT,
2642 };
2643 VkFormat plane_vkfmt = av_vkfmt_from_pixfmt(hwfc->sw_format)[i];
2644
2645 drm_desc->layers[i].format = vulkan_fmt_to_drm(plane_vkfmt);
2646 drm_desc->layers[i].nb_planes = 1;
2647
2648 if (drm_desc->layers[i].format == DRM_FORMAT_INVALID) {
2649 av_log(hwfc, AV_LOG_ERROR, "Cannot map to DRM layer, unsupported!\n");
2650 err = AVERROR_PATCHWELCOME;
2651 goto end;
2652 }
2653
2654 drm_desc->layers[i].planes[0].object_index = FFMIN(i, drm_desc->nb_objects - 1);
2655
2656 if (f->tiling == VK_IMAGE_TILING_OPTIMAL)
2657 continue;
2658
2659 vkGetImageSubresourceLayout(hwctx->act_dev, f->img[i], &sub, &layout);
2660 drm_desc->layers[i].planes[0].offset = layout.offset;
2661 drm_desc->layers[i].planes[0].pitch = layout.rowPitch;
2662 }
2663
2664 dst->width = src->width;
2665 dst->height = src->height;
2666 dst->data[0] = (uint8_t *)drm_desc;
2667
2668 av_log(hwfc, AV_LOG_VERBOSE, "Mapped AVVkFrame to a DRM object!\n");
2669
2670 return 0;
2671
2672 end:
2673 av_free(drm_desc);
2674 return err;
2675 }
2676
2677 #if CONFIG_VAAPI
vulkan_map_to_vaapi(AVHWFramesContext * hwfc,AVFrame * dst,const AVFrame * src,int flags)2678 static int vulkan_map_to_vaapi(AVHWFramesContext *hwfc, AVFrame *dst,
2679 const AVFrame *src, int flags)
2680 {
2681 int err;
2682 AVFrame *tmp = av_frame_alloc();
2683 if (!tmp)
2684 return AVERROR(ENOMEM);
2685
2686 tmp->format = AV_PIX_FMT_DRM_PRIME;
2687
2688 err = vulkan_map_to_drm(hwfc, tmp, src, flags);
2689 if (err < 0)
2690 goto fail;
2691
2692 err = av_hwframe_map(dst, tmp, flags);
2693 if (err < 0)
2694 goto fail;
2695
2696 err = ff_hwframe_map_replace(dst, src);
2697
2698 fail:
2699 av_frame_free(&tmp);
2700 return err;
2701 }
2702 #endif
2703 #endif
2704
vulkan_map_from(AVHWFramesContext * hwfc,AVFrame * dst,const AVFrame * src,int flags)2705 static int vulkan_map_from(AVHWFramesContext *hwfc, AVFrame *dst,
2706 const AVFrame *src, int flags)
2707 {
2708 av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2709
2710 switch (dst->format) {
2711 #if CONFIG_LIBDRM
2712 case AV_PIX_FMT_DRM_PRIME:
2713 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2714 return vulkan_map_to_drm(hwfc, dst, src, flags);
2715 #if CONFIG_VAAPI
2716 case AV_PIX_FMT_VAAPI:
2717 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2718 return vulkan_map_to_vaapi(hwfc, dst, src, flags);
2719 #endif
2720 #endif
2721 default:
2722 return vulkan_map_frame_to_mem(hwfc, dst, src, flags);
2723 }
2724 }
2725
2726 typedef struct ImageBuffer {
2727 VkBuffer buf;
2728 VkDeviceMemory mem;
2729 VkMemoryPropertyFlagBits flags;
2730 int mapped_mem;
2731 } ImageBuffer;
2732
free_buf(void * opaque,uint8_t * data)2733 static void free_buf(void *opaque, uint8_t *data)
2734 {
2735 AVHWDeviceContext *ctx = opaque;
2736 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2737 ImageBuffer *vkbuf = (ImageBuffer *)data;
2738
2739 if (vkbuf->buf)
2740 vkDestroyBuffer(hwctx->act_dev, vkbuf->buf, hwctx->alloc);
2741 if (vkbuf->mem)
2742 vkFreeMemory(hwctx->act_dev, vkbuf->mem, hwctx->alloc);
2743
2744 av_free(data);
2745 }
2746
get_req_buffer_size(VulkanDevicePriv * p,int * stride,int height)2747 static size_t get_req_buffer_size(VulkanDevicePriv *p, int *stride, int height)
2748 {
2749 size_t size;
2750 *stride = FFALIGN(*stride, p->props.properties.limits.optimalBufferCopyRowPitchAlignment);
2751 size = height*(*stride);
2752 size = FFALIGN(size, p->props.properties.limits.minMemoryMapAlignment);
2753 return size;
2754 }
2755
create_buf(AVHWDeviceContext * ctx,AVBufferRef ** buf,VkBufferUsageFlags usage,VkMemoryPropertyFlagBits flags,size_t size,uint32_t req_memory_bits,int host_mapped,void * create_pnext,void * alloc_pnext)2756 static int create_buf(AVHWDeviceContext *ctx, AVBufferRef **buf,
2757 VkBufferUsageFlags usage, VkMemoryPropertyFlagBits flags,
2758 size_t size, uint32_t req_memory_bits, int host_mapped,
2759 void *create_pnext, void *alloc_pnext)
2760 {
2761 int err;
2762 VkResult ret;
2763 int use_ded_mem;
2764 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2765
2766 VkBufferCreateInfo buf_spawn = {
2767 .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
2768 .pNext = create_pnext,
2769 .usage = usage,
2770 .size = size,
2771 .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
2772 };
2773
2774 VkBufferMemoryRequirementsInfo2 req_desc = {
2775 .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_REQUIREMENTS_INFO_2,
2776 };
2777 VkMemoryDedicatedAllocateInfo ded_alloc = {
2778 .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO,
2779 .pNext = alloc_pnext,
2780 };
2781 VkMemoryDedicatedRequirements ded_req = {
2782 .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS,
2783 };
2784 VkMemoryRequirements2 req = {
2785 .sType = VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2,
2786 .pNext = &ded_req,
2787 };
2788
2789 ImageBuffer *vkbuf = av_mallocz(sizeof(*vkbuf));
2790 if (!vkbuf)
2791 return AVERROR(ENOMEM);
2792
2793 vkbuf->mapped_mem = host_mapped;
2794
2795 ret = vkCreateBuffer(hwctx->act_dev, &buf_spawn, NULL, &vkbuf->buf);
2796 if (ret != VK_SUCCESS) {
2797 av_log(ctx, AV_LOG_ERROR, "Failed to create buffer: %s\n",
2798 vk_ret2str(ret));
2799 err = AVERROR_EXTERNAL;
2800 goto fail;
2801 }
2802
2803 req_desc.buffer = vkbuf->buf;
2804
2805 vkGetBufferMemoryRequirements2(hwctx->act_dev, &req_desc, &req);
2806
2807 /* In case the implementation prefers/requires dedicated allocation */
2808 use_ded_mem = ded_req.prefersDedicatedAllocation |
2809 ded_req.requiresDedicatedAllocation;
2810 if (use_ded_mem)
2811 ded_alloc.buffer = vkbuf->buf;
2812
2813 /* Additional requirements imposed on us */
2814 if (req_memory_bits)
2815 req.memoryRequirements.memoryTypeBits &= req_memory_bits;
2816
2817 err = alloc_mem(ctx, &req.memoryRequirements, flags,
2818 use_ded_mem ? &ded_alloc : (void *)ded_alloc.pNext,
2819 &vkbuf->flags, &vkbuf->mem);
2820 if (err)
2821 goto fail;
2822
2823 ret = vkBindBufferMemory(hwctx->act_dev, vkbuf->buf, vkbuf->mem, 0);
2824 if (ret != VK_SUCCESS) {
2825 av_log(ctx, AV_LOG_ERROR, "Failed to bind memory to buffer: %s\n",
2826 vk_ret2str(ret));
2827 err = AVERROR_EXTERNAL;
2828 goto fail;
2829 }
2830
2831 *buf = av_buffer_create((uint8_t *)vkbuf, sizeof(*vkbuf), free_buf, ctx, 0);
2832 if (!(*buf)) {
2833 err = AVERROR(ENOMEM);
2834 goto fail;
2835 }
2836
2837 return 0;
2838
2839 fail:
2840 free_buf(ctx, (uint8_t *)vkbuf);
2841 return err;
2842 }
2843
2844 /* Skips mapping of host mapped buffers but still invalidates them */
map_buffers(AVHWDeviceContext * ctx,AVBufferRef ** bufs,uint8_t * mem[],int nb_buffers,int invalidate)2845 static int map_buffers(AVHWDeviceContext *ctx, AVBufferRef **bufs, uint8_t *mem[],
2846 int nb_buffers, int invalidate)
2847 {
2848 VkResult ret;
2849 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2850 VkMappedMemoryRange invalidate_ctx[AV_NUM_DATA_POINTERS];
2851 int invalidate_count = 0;
2852
2853 for (int i = 0; i < nb_buffers; i++) {
2854 ImageBuffer *vkbuf = (ImageBuffer *)bufs[i]->data;
2855 if (vkbuf->mapped_mem)
2856 continue;
2857
2858 ret = vkMapMemory(hwctx->act_dev, vkbuf->mem, 0,
2859 VK_WHOLE_SIZE, 0, (void **)&mem[i]);
2860 if (ret != VK_SUCCESS) {
2861 av_log(ctx, AV_LOG_ERROR, "Failed to map buffer memory: %s\n",
2862 vk_ret2str(ret));
2863 return AVERROR_EXTERNAL;
2864 }
2865 }
2866
2867 if (!invalidate)
2868 return 0;
2869
2870 for (int i = 0; i < nb_buffers; i++) {
2871 ImageBuffer *vkbuf = (ImageBuffer *)bufs[i]->data;
2872 const VkMappedMemoryRange ival_buf = {
2873 .sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE,
2874 .memory = vkbuf->mem,
2875 .size = VK_WHOLE_SIZE,
2876 };
2877
2878 /* For host imported memory Vulkan says to use platform-defined
2879 * sync methods, but doesn't really say not to call flush or invalidate
2880 * on original host pointers. It does explicitly allow to do that on
2881 * host-mapped pointers which are then mapped again using vkMapMemory,
2882 * but known implementations return the original pointers when mapped
2883 * again. */
2884 if (vkbuf->flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)
2885 continue;
2886
2887 invalidate_ctx[invalidate_count++] = ival_buf;
2888 }
2889
2890 if (invalidate_count) {
2891 ret = vkInvalidateMappedMemoryRanges(hwctx->act_dev, invalidate_count,
2892 invalidate_ctx);
2893 if (ret != VK_SUCCESS)
2894 av_log(ctx, AV_LOG_WARNING, "Failed to invalidate memory: %s\n",
2895 vk_ret2str(ret));
2896 }
2897
2898 return 0;
2899 }
2900
unmap_buffers(AVHWDeviceContext * ctx,AVBufferRef ** bufs,int nb_buffers,int flush)2901 static int unmap_buffers(AVHWDeviceContext *ctx, AVBufferRef **bufs,
2902 int nb_buffers, int flush)
2903 {
2904 int err = 0;
2905 VkResult ret;
2906 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2907 VkMappedMemoryRange flush_ctx[AV_NUM_DATA_POINTERS];
2908 int flush_count = 0;
2909
2910 if (flush) {
2911 for (int i = 0; i < nb_buffers; i++) {
2912 ImageBuffer *vkbuf = (ImageBuffer *)bufs[i]->data;
2913 const VkMappedMemoryRange flush_buf = {
2914 .sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE,
2915 .memory = vkbuf->mem,
2916 .size = VK_WHOLE_SIZE,
2917 };
2918
2919 if (vkbuf->flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)
2920 continue;
2921
2922 flush_ctx[flush_count++] = flush_buf;
2923 }
2924 }
2925
2926 if (flush_count) {
2927 ret = vkFlushMappedMemoryRanges(hwctx->act_dev, flush_count, flush_ctx);
2928 if (ret != VK_SUCCESS) {
2929 av_log(ctx, AV_LOG_ERROR, "Failed to flush memory: %s\n",
2930 vk_ret2str(ret));
2931 err = AVERROR_EXTERNAL; /* We still want to try to unmap them */
2932 }
2933 }
2934
2935 for (int i = 0; i < nb_buffers; i++) {
2936 ImageBuffer *vkbuf = (ImageBuffer *)bufs[i]->data;
2937 if (vkbuf->mapped_mem)
2938 continue;
2939
2940 vkUnmapMemory(hwctx->act_dev, vkbuf->mem);
2941 }
2942
2943 return err;
2944 }
2945
transfer_image_buf(AVHWFramesContext * hwfc,const AVFrame * f,AVBufferRef ** bufs,size_t * buf_offsets,const int * buf_stride,int w,int h,enum AVPixelFormat pix_fmt,int to_buf)2946 static int transfer_image_buf(AVHWFramesContext *hwfc, const AVFrame *f,
2947 AVBufferRef **bufs, size_t *buf_offsets,
2948 const int *buf_stride, int w,
2949 int h, enum AVPixelFormat pix_fmt, int to_buf)
2950 {
2951 int err;
2952 AVVkFrame *frame = (AVVkFrame *)f->data[0];
2953 VulkanFramesPriv *fp = hwfc->internal->priv;
2954
2955 int bar_num = 0;
2956 VkPipelineStageFlagBits sem_wait_dst[AV_NUM_DATA_POINTERS];
2957
2958 const int planes = av_pix_fmt_count_planes(pix_fmt);
2959 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(pix_fmt);
2960
2961 VkImageMemoryBarrier img_bar[AV_NUM_DATA_POINTERS] = { 0 };
2962 VulkanExecCtx *ectx = to_buf ? &fp->download_ctx : &fp->upload_ctx;
2963 VkCommandBuffer cmd_buf = get_buf_exec_ctx(hwfc, ectx);
2964
2965 VkSubmitInfo s_info = {
2966 .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
2967 .pSignalSemaphores = frame->sem,
2968 .pWaitSemaphores = frame->sem,
2969 .pWaitDstStageMask = sem_wait_dst,
2970 .signalSemaphoreCount = planes,
2971 .waitSemaphoreCount = planes,
2972 };
2973
2974 if ((err = wait_start_exec_ctx(hwfc, ectx)))
2975 return err;
2976
2977 /* Change the image layout to something more optimal for transfers */
2978 for (int i = 0; i < planes; i++) {
2979 VkImageLayout new_layout = to_buf ? VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL :
2980 VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
2981 VkAccessFlags new_access = to_buf ? VK_ACCESS_TRANSFER_READ_BIT :
2982 VK_ACCESS_TRANSFER_WRITE_BIT;
2983
2984 sem_wait_dst[i] = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
2985
2986 /* If the layout matches and we have read access skip the barrier */
2987 if ((frame->layout[i] == new_layout) && (frame->access[i] & new_access))
2988 continue;
2989
2990 img_bar[bar_num].sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
2991 img_bar[bar_num].srcAccessMask = 0x0;
2992 img_bar[bar_num].dstAccessMask = new_access;
2993 img_bar[bar_num].oldLayout = frame->layout[i];
2994 img_bar[bar_num].newLayout = new_layout;
2995 img_bar[bar_num].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
2996 img_bar[bar_num].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
2997 img_bar[bar_num].image = frame->img[i];
2998 img_bar[bar_num].subresourceRange.levelCount = 1;
2999 img_bar[bar_num].subresourceRange.layerCount = 1;
3000 img_bar[bar_num].subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
3001
3002 frame->layout[i] = img_bar[bar_num].newLayout;
3003 frame->access[i] = img_bar[bar_num].dstAccessMask;
3004
3005 bar_num++;
3006 }
3007
3008 if (bar_num)
3009 vkCmdPipelineBarrier(cmd_buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
3010 VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
3011 0, NULL, 0, NULL, bar_num, img_bar);
3012
3013 /* Schedule a copy for each plane */
3014 for (int i = 0; i < planes; i++) {
3015 ImageBuffer *vkbuf = (ImageBuffer *)bufs[i]->data;
3016 VkBufferImageCopy buf_reg = {
3017 .bufferOffset = buf_offsets[i],
3018 .bufferRowLength = buf_stride[i] / desc->comp[i].step,
3019 .imageSubresource.layerCount = 1,
3020 .imageSubresource.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
3021 .imageOffset = { 0, 0, 0, },
3022 };
3023
3024 int p_w, p_h;
3025 get_plane_wh(&p_w, &p_h, pix_fmt, w, h, i);
3026
3027 buf_reg.bufferImageHeight = p_h;
3028 buf_reg.imageExtent = (VkExtent3D){ p_w, p_h, 1, };
3029
3030 if (to_buf)
3031 vkCmdCopyImageToBuffer(cmd_buf, frame->img[i], frame->layout[i],
3032 vkbuf->buf, 1, &buf_reg);
3033 else
3034 vkCmdCopyBufferToImage(cmd_buf, vkbuf->buf, frame->img[i],
3035 frame->layout[i], 1, &buf_reg);
3036 }
3037
3038 /* When uploading, do this asynchronously if the source is refcounted by
3039 * keeping the buffers as a submission dependency.
3040 * The hwcontext is guaranteed to not be freed until all frames are freed
3041 * in the frames_unint function.
3042 * When downloading to buffer, do this synchronously and wait for the
3043 * queue submission to finish executing */
3044 if (!to_buf) {
3045 int ref;
3046 for (ref = 0; ref < AV_NUM_DATA_POINTERS; ref++) {
3047 if (!f->buf[ref])
3048 break;
3049 if ((err = add_buf_dep_exec_ctx(hwfc, ectx, &f->buf[ref], 1)))
3050 return err;
3051 }
3052 if (ref && (err = add_buf_dep_exec_ctx(hwfc, ectx, bufs, planes)))
3053 return err;
3054 return submit_exec_ctx(hwfc, ectx, &s_info, !ref);
3055 } else {
3056 return submit_exec_ctx(hwfc, ectx, &s_info, 1);
3057 }
3058 }
3059
vulkan_transfer_data(AVHWFramesContext * hwfc,const AVFrame * vkf,const AVFrame * swf,int from)3060 static int vulkan_transfer_data(AVHWFramesContext *hwfc, const AVFrame *vkf,
3061 const AVFrame *swf, int from)
3062 {
3063 int err = 0;
3064 VkResult ret;
3065 AVVkFrame *f = (AVVkFrame *)vkf->data[0];
3066 AVHWDeviceContext *dev_ctx = hwfc->device_ctx;
3067 AVVulkanDeviceContext *hwctx = dev_ctx->hwctx;
3068 VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
3069
3070 AVFrame tmp;
3071 AVBufferRef *bufs[AV_NUM_DATA_POINTERS] = { 0 };
3072 size_t buf_offsets[AV_NUM_DATA_POINTERS] = { 0 };
3073
3074 int p_w, p_h;
3075 const int planes = av_pix_fmt_count_planes(swf->format);
3076
3077 int host_mapped[AV_NUM_DATA_POINTERS] = { 0 };
3078 const int map_host = !!(p->extensions & EXT_EXTERNAL_HOST_MEMORY);
3079
3080 VK_LOAD_PFN(hwctx->inst, vkGetMemoryHostPointerPropertiesEXT);
3081
3082 if ((swf->format != AV_PIX_FMT_NONE && !av_vkfmt_from_pixfmt(swf->format))) {
3083 av_log(hwfc, AV_LOG_ERROR, "Unsupported software frame pixel format!\n");
3084 return AVERROR(EINVAL);
3085 }
3086
3087 if (swf->width > hwfc->width || swf->height > hwfc->height)
3088 return AVERROR(EINVAL);
3089
3090 /* For linear, host visiable images */
3091 if (f->tiling == VK_IMAGE_TILING_LINEAR &&
3092 f->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) {
3093 AVFrame *map = av_frame_alloc();
3094 if (!map)
3095 return AVERROR(ENOMEM);
3096 map->format = swf->format;
3097
3098 err = vulkan_map_frame_to_mem(hwfc, map, vkf, AV_HWFRAME_MAP_WRITE);
3099 if (err)
3100 return err;
3101
3102 err = av_frame_copy((AVFrame *)(from ? swf : map), from ? map : swf);
3103 av_frame_free(&map);
3104 return err;
3105 }
3106
3107 /* Create buffers */
3108 for (int i = 0; i < planes; i++) {
3109 size_t req_size;
3110
3111 VkExternalMemoryBufferCreateInfo create_desc = {
3112 .sType = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_BUFFER_CREATE_INFO,
3113 .handleTypes = VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT,
3114 };
3115
3116 VkImportMemoryHostPointerInfoEXT import_desc = {
3117 .sType = VK_STRUCTURE_TYPE_IMPORT_MEMORY_HOST_POINTER_INFO_EXT,
3118 .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT,
3119 };
3120
3121 VkMemoryHostPointerPropertiesEXT p_props = {
3122 .sType = VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT,
3123 };
3124
3125 get_plane_wh(&p_w, &p_h, swf->format, swf->width, swf->height, i);
3126
3127 tmp.linesize[i] = FFABS(swf->linesize[i]);
3128
3129 /* Do not map images with a negative stride */
3130 if (map_host && swf->linesize[i] > 0) {
3131 size_t offs;
3132 offs = (uintptr_t)swf->data[i] % p->hprops.minImportedHostPointerAlignment;
3133 import_desc.pHostPointer = swf->data[i] - offs;
3134
3135 /* We have to compensate for the few extra bytes of padding we
3136 * completely ignore at the start */
3137 req_size = FFALIGN(offs + tmp.linesize[i] * p_h,
3138 p->hprops.minImportedHostPointerAlignment);
3139
3140 ret = pfn_vkGetMemoryHostPointerPropertiesEXT(hwctx->act_dev,
3141 import_desc.handleType,
3142 import_desc.pHostPointer,
3143 &p_props);
3144
3145 if (ret == VK_SUCCESS) {
3146 host_mapped[i] = 1;
3147 buf_offsets[i] = offs;
3148 }
3149 }
3150
3151 if (!host_mapped[i])
3152 req_size = get_req_buffer_size(p, &tmp.linesize[i], p_h);
3153
3154 err = create_buf(dev_ctx, &bufs[i],
3155 from ? VK_BUFFER_USAGE_TRANSFER_DST_BIT :
3156 VK_BUFFER_USAGE_TRANSFER_SRC_BIT,
3157 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT,
3158 req_size, p_props.memoryTypeBits, host_mapped[i],
3159 host_mapped[i] ? &create_desc : NULL,
3160 host_mapped[i] ? &import_desc : NULL);
3161 if (err)
3162 goto end;
3163 }
3164
3165 if (!from) {
3166 /* Map, copy image to buffer, unmap */
3167 if ((err = map_buffers(dev_ctx, bufs, tmp.data, planes, 0)))
3168 goto end;
3169
3170 for (int i = 0; i < planes; i++) {
3171 if (host_mapped[i])
3172 continue;
3173
3174 get_plane_wh(&p_w, &p_h, swf->format, swf->width, swf->height, i);
3175
3176 av_image_copy_plane(tmp.data[i], tmp.linesize[i],
3177 (const uint8_t *)swf->data[i], swf->linesize[i],
3178 FFMIN(tmp.linesize[i], FFABS(swf->linesize[i])),
3179 p_h);
3180 }
3181
3182 if ((err = unmap_buffers(dev_ctx, bufs, planes, 1)))
3183 goto end;
3184 }
3185
3186 /* Copy buffers into/from image */
3187 err = transfer_image_buf(hwfc, vkf, bufs, buf_offsets, tmp.linesize,
3188 swf->width, swf->height, swf->format, from);
3189
3190 if (from) {
3191 /* Map, copy image to buffer, unmap */
3192 if ((err = map_buffers(dev_ctx, bufs, tmp.data, planes, 0)))
3193 goto end;
3194
3195 for (int i = 0; i < planes; i++) {
3196 if (host_mapped[i])
3197 continue;
3198
3199 get_plane_wh(&p_w, &p_h, swf->format, swf->width, swf->height, i);
3200
3201 av_image_copy_plane(swf->data[i], swf->linesize[i],
3202 (const uint8_t *)tmp.data[i], tmp.linesize[i],
3203 FFMIN(tmp.linesize[i], FFABS(swf->linesize[i])),
3204 p_h);
3205 }
3206
3207 if ((err = unmap_buffers(dev_ctx, bufs, planes, 1)))
3208 goto end;
3209 }
3210
3211 end:
3212 for (int i = 0; i < planes; i++)
3213 av_buffer_unref(&bufs[i]);
3214
3215 return err;
3216 }
3217
vulkan_transfer_data_to(AVHWFramesContext * hwfc,AVFrame * dst,const AVFrame * src)3218 static int vulkan_transfer_data_to(AVHWFramesContext *hwfc, AVFrame *dst,
3219 const AVFrame *src)
3220 {
3221 av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
3222
3223 switch (src->format) {
3224 #if CONFIG_CUDA
3225 case AV_PIX_FMT_CUDA:
3226 if ((p->extensions & EXT_EXTERNAL_FD_MEMORY) &&
3227 (p->extensions & EXT_EXTERNAL_FD_SEM))
3228 return vulkan_transfer_data_from_cuda(hwfc, dst, src);
3229 #endif
3230 default:
3231 if (src->hw_frames_ctx)
3232 return AVERROR(ENOSYS);
3233 else
3234 return vulkan_transfer_data(hwfc, dst, src, 0);
3235 }
3236 }
3237
3238 #if CONFIG_CUDA
vulkan_transfer_data_to_cuda(AVHWFramesContext * hwfc,AVFrame * dst,const AVFrame * src)3239 static int vulkan_transfer_data_to_cuda(AVHWFramesContext *hwfc, AVFrame *dst,
3240 const AVFrame *src)
3241 {
3242 int err;
3243 VkResult ret;
3244 CUcontext dummy;
3245 AVVkFrame *dst_f;
3246 AVVkFrameInternal *dst_int;
3247 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
3248 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
3249
3250 AVHWFramesContext *cuda_fc = (AVHWFramesContext*)dst->hw_frames_ctx->data;
3251 AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
3252 AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
3253 AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
3254 CudaFunctions *cu = cu_internal->cuda_dl;
3255 CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS s_w_par[AV_NUM_DATA_POINTERS] = { 0 };
3256 CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS s_s_par[AV_NUM_DATA_POINTERS] = { 0 };
3257
3258 ret = CHECK_CU(cu->cuCtxPushCurrent(cuda_dev->cuda_ctx));
3259 if (ret < 0)
3260 return AVERROR_EXTERNAL;
3261
3262 dst_f = (AVVkFrame *)src->data[0];
3263
3264 err = vulkan_export_to_cuda(hwfc, dst->hw_frames_ctx, src);
3265 if (err < 0) {
3266 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
3267 return err;
3268 }
3269
3270 dst_int = dst_f->internal;
3271
3272 ret = CHECK_CU(cu->cuWaitExternalSemaphoresAsync(dst_int->cu_sem, s_w_par,
3273 planes, cuda_dev->stream));
3274 if (ret < 0) {
3275 err = AVERROR_EXTERNAL;
3276 goto fail;
3277 }
3278
3279 for (int i = 0; i < planes; i++) {
3280 CUDA_MEMCPY2D cpy = {
3281 .dstMemoryType = CU_MEMORYTYPE_DEVICE,
3282 .dstDevice = (CUdeviceptr)dst->data[i],
3283 .dstPitch = dst->linesize[i],
3284 .dstY = 0,
3285
3286 .srcMemoryType = CU_MEMORYTYPE_ARRAY,
3287 .srcArray = dst_int->cu_array[i],
3288 };
3289
3290 int w, h;
3291 get_plane_wh(&w, &h, hwfc->sw_format, hwfc->width, hwfc->height, i);
3292
3293 cpy.WidthInBytes = w * desc->comp[i].step;
3294 cpy.Height = h;
3295
3296 ret = CHECK_CU(cu->cuMemcpy2DAsync(&cpy, cuda_dev->stream));
3297 if (ret < 0) {
3298 err = AVERROR_EXTERNAL;
3299 goto fail;
3300 }
3301 }
3302
3303 ret = CHECK_CU(cu->cuSignalExternalSemaphoresAsync(dst_int->cu_sem, s_s_par,
3304 planes, cuda_dev->stream));
3305 if (ret < 0) {
3306 err = AVERROR_EXTERNAL;
3307 goto fail;
3308 }
3309
3310 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
3311
3312 av_log(hwfc, AV_LOG_VERBOSE, "Transfered Vulkan image to CUDA!\n");
3313
3314 return 0;
3315
3316 fail:
3317 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
3318 vulkan_free_internal(dst_int);
3319 dst_f->internal = NULL;
3320 av_buffer_unref(&dst->buf[0]);
3321 return err;
3322 }
3323 #endif
3324
vulkan_transfer_data_from(AVHWFramesContext * hwfc,AVFrame * dst,const AVFrame * src)3325 static int vulkan_transfer_data_from(AVHWFramesContext *hwfc, AVFrame *dst,
3326 const AVFrame *src)
3327 {
3328 av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
3329
3330 switch (dst->format) {
3331 #if CONFIG_CUDA
3332 case AV_PIX_FMT_CUDA:
3333 if ((p->extensions & EXT_EXTERNAL_FD_MEMORY) &&
3334 (p->extensions & EXT_EXTERNAL_FD_SEM))
3335 return vulkan_transfer_data_to_cuda(hwfc, dst, src);
3336 #endif
3337 default:
3338 if (dst->hw_frames_ctx)
3339 return AVERROR(ENOSYS);
3340 else
3341 return vulkan_transfer_data(hwfc, src, dst, 1);
3342 }
3343 }
3344
vulkan_frames_derive_to(AVHWFramesContext * dst_fc,AVHWFramesContext * src_fc,int flags)3345 static int vulkan_frames_derive_to(AVHWFramesContext *dst_fc,
3346 AVHWFramesContext *src_fc, int flags)
3347 {
3348 return vulkan_frames_init(dst_fc);
3349 }
3350
av_vk_frame_alloc(void)3351 AVVkFrame *av_vk_frame_alloc(void)
3352 {
3353 return av_mallocz(sizeof(AVVkFrame));
3354 }
3355
3356 const HWContextType ff_hwcontext_type_vulkan = {
3357 .type = AV_HWDEVICE_TYPE_VULKAN,
3358 .name = "Vulkan",
3359
3360 .device_hwctx_size = sizeof(AVVulkanDeviceContext),
3361 .device_priv_size = sizeof(VulkanDevicePriv),
3362 .frames_hwctx_size = sizeof(AVVulkanFramesContext),
3363 .frames_priv_size = sizeof(VulkanFramesPriv),
3364
3365 .device_init = &vulkan_device_init,
3366 .device_create = &vulkan_device_create,
3367 .device_derive = &vulkan_device_derive,
3368
3369 .frames_get_constraints = &vulkan_frames_get_constraints,
3370 .frames_init = vulkan_frames_init,
3371 .frames_get_buffer = vulkan_get_buffer,
3372 .frames_uninit = vulkan_frames_uninit,
3373
3374 .transfer_get_formats = vulkan_transfer_get_formats,
3375 .transfer_data_to = vulkan_transfer_data_to,
3376 .transfer_data_from = vulkan_transfer_data_from,
3377
3378 .map_to = vulkan_map_to,
3379 .map_from = vulkan_map_from,
3380 .frames_derive_to = &vulkan_frames_derive_to,
3381
3382 .pix_fmts = (const enum AVPixelFormat []) {
3383 AV_PIX_FMT_VULKAN,
3384 AV_PIX_FMT_NONE
3385 },
3386 };
3387