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[3];
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_P010, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16G16_UNORM } },
149 { AV_PIX_FMT_P016, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16G16_UNORM } },
150
151 { AV_PIX_FMT_YUV420P, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } },
152 { AV_PIX_FMT_YUV422P, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } },
153 { AV_PIX_FMT_YUV444P, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } },
154
155 { AV_PIX_FMT_YUV420P16, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
156 { AV_PIX_FMT_YUV422P16, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
157 { AV_PIX_FMT_YUV444P16, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
158
159 { AV_PIX_FMT_ABGR, { VK_FORMAT_A8B8G8R8_UNORM_PACK32 } },
160 { AV_PIX_FMT_BGRA, { VK_FORMAT_B8G8R8A8_UNORM } },
161 { AV_PIX_FMT_RGBA, { VK_FORMAT_R8G8B8A8_UNORM } },
162 { AV_PIX_FMT_RGB24, { VK_FORMAT_R8G8B8_UNORM } },
163 { AV_PIX_FMT_BGR24, { VK_FORMAT_B8G8R8_UNORM } },
164 { AV_PIX_FMT_RGB48, { VK_FORMAT_R16G16B16_UNORM } },
165 { AV_PIX_FMT_RGBA64, { VK_FORMAT_R16G16B16A16_UNORM } },
166 { AV_PIX_FMT_RGB565, { VK_FORMAT_R5G6B5_UNORM_PACK16 } },
167 { AV_PIX_FMT_BGR565, { VK_FORMAT_B5G6R5_UNORM_PACK16 } },
168 { AV_PIX_FMT_BGR0, { VK_FORMAT_B8G8R8A8_UNORM } },
169 { AV_PIX_FMT_0BGR, { VK_FORMAT_A8B8G8R8_UNORM_PACK32 } },
170 { AV_PIX_FMT_RGB0, { VK_FORMAT_R8G8B8A8_UNORM } },
171
172 { AV_PIX_FMT_GBRPF32, { VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_SFLOAT } },
173 };
174
av_vkfmt_from_pixfmt(enum AVPixelFormat p)175 const VkFormat *av_vkfmt_from_pixfmt(enum AVPixelFormat p)
176 {
177 for (enum AVPixelFormat i = 0; i < FF_ARRAY_ELEMS(vk_pixfmt_map); i++)
178 if (vk_pixfmt_map[i].pixfmt == p)
179 return vk_pixfmt_map[i].vkfmts;
180 return NULL;
181 }
182
pixfmt_is_supported(AVVulkanDeviceContext * hwctx,enum AVPixelFormat p,int linear)183 static int pixfmt_is_supported(AVVulkanDeviceContext *hwctx, enum AVPixelFormat p,
184 int linear)
185 {
186 const VkFormat *fmt = av_vkfmt_from_pixfmt(p);
187 int planes = av_pix_fmt_count_planes(p);
188
189 if (!fmt)
190 return 0;
191
192 for (int i = 0; i < planes; i++) {
193 VkFormatFeatureFlags flags;
194 VkFormatProperties2 prop = {
195 .sType = VK_STRUCTURE_TYPE_FORMAT_PROPERTIES_2,
196 };
197 vkGetPhysicalDeviceFormatProperties2(hwctx->phys_dev, fmt[i], &prop);
198 flags = linear ? prop.formatProperties.linearTilingFeatures :
199 prop.formatProperties.optimalTilingFeatures;
200 if (!(flags & DEFAULT_USAGE_FLAGS))
201 return 0;
202 }
203
204 return 1;
205 }
206
207 enum VulkanExtensions {
208 EXT_EXTERNAL_DMABUF_MEMORY = 1ULL << 0, /* VK_EXT_external_memory_dma_buf */
209 EXT_DRM_MODIFIER_FLAGS = 1ULL << 1, /* VK_EXT_image_drm_format_modifier */
210 EXT_EXTERNAL_FD_MEMORY = 1ULL << 2, /* VK_KHR_external_memory_fd */
211 EXT_EXTERNAL_FD_SEM = 1ULL << 3, /* VK_KHR_external_semaphore_fd */
212 EXT_EXTERNAL_HOST_MEMORY = 1ULL << 4, /* VK_EXT_external_memory_host */
213
214 EXT_NO_FLAG = 1ULL << 63,
215 };
216
217 typedef struct VulkanOptExtension {
218 const char *name;
219 uint64_t flag;
220 } VulkanOptExtension;
221
222 static const VulkanOptExtension optional_instance_exts[] = {
223 /* For future use */
224 };
225
226 static const VulkanOptExtension optional_device_exts[] = {
227 { VK_KHR_EXTERNAL_MEMORY_FD_EXTENSION_NAME, EXT_EXTERNAL_FD_MEMORY, },
228 { VK_EXT_EXTERNAL_MEMORY_DMA_BUF_EXTENSION_NAME, EXT_EXTERNAL_DMABUF_MEMORY, },
229 { VK_EXT_IMAGE_DRM_FORMAT_MODIFIER_EXTENSION_NAME, EXT_DRM_MODIFIER_FLAGS, },
230 { VK_KHR_EXTERNAL_SEMAPHORE_FD_EXTENSION_NAME, EXT_EXTERNAL_FD_SEM, },
231 { VK_EXT_EXTERNAL_MEMORY_HOST_EXTENSION_NAME, EXT_EXTERNAL_HOST_MEMORY, },
232 };
233
234 /* Converts return values to strings */
vk_ret2str(VkResult res)235 static const char *vk_ret2str(VkResult res)
236 {
237 #define CASE(VAL) case VAL: return #VAL
238 switch (res) {
239 CASE(VK_SUCCESS);
240 CASE(VK_NOT_READY);
241 CASE(VK_TIMEOUT);
242 CASE(VK_EVENT_SET);
243 CASE(VK_EVENT_RESET);
244 CASE(VK_INCOMPLETE);
245 CASE(VK_ERROR_OUT_OF_HOST_MEMORY);
246 CASE(VK_ERROR_OUT_OF_DEVICE_MEMORY);
247 CASE(VK_ERROR_INITIALIZATION_FAILED);
248 CASE(VK_ERROR_DEVICE_LOST);
249 CASE(VK_ERROR_MEMORY_MAP_FAILED);
250 CASE(VK_ERROR_LAYER_NOT_PRESENT);
251 CASE(VK_ERROR_EXTENSION_NOT_PRESENT);
252 CASE(VK_ERROR_FEATURE_NOT_PRESENT);
253 CASE(VK_ERROR_INCOMPATIBLE_DRIVER);
254 CASE(VK_ERROR_TOO_MANY_OBJECTS);
255 CASE(VK_ERROR_FORMAT_NOT_SUPPORTED);
256 CASE(VK_ERROR_FRAGMENTED_POOL);
257 CASE(VK_ERROR_SURFACE_LOST_KHR);
258 CASE(VK_ERROR_NATIVE_WINDOW_IN_USE_KHR);
259 CASE(VK_SUBOPTIMAL_KHR);
260 CASE(VK_ERROR_OUT_OF_DATE_KHR);
261 CASE(VK_ERROR_INCOMPATIBLE_DISPLAY_KHR);
262 CASE(VK_ERROR_VALIDATION_FAILED_EXT);
263 CASE(VK_ERROR_INVALID_SHADER_NV);
264 CASE(VK_ERROR_OUT_OF_POOL_MEMORY);
265 CASE(VK_ERROR_INVALID_EXTERNAL_HANDLE);
266 CASE(VK_ERROR_NOT_PERMITTED_EXT);
267 CASE(VK_ERROR_INVALID_DRM_FORMAT_MODIFIER_PLANE_LAYOUT_EXT);
268 CASE(VK_ERROR_INVALID_DEVICE_ADDRESS_EXT);
269 CASE(VK_ERROR_FULL_SCREEN_EXCLUSIVE_MODE_LOST_EXT);
270 default: return "Unknown error";
271 }
272 #undef CASE
273 }
274
vk_dbg_callback(VkDebugUtilsMessageSeverityFlagBitsEXT severity,VkDebugUtilsMessageTypeFlagsEXT messageType,const VkDebugUtilsMessengerCallbackDataEXT * data,void * priv)275 static VkBool32 vk_dbg_callback(VkDebugUtilsMessageSeverityFlagBitsEXT severity,
276 VkDebugUtilsMessageTypeFlagsEXT messageType,
277 const VkDebugUtilsMessengerCallbackDataEXT *data,
278 void *priv)
279 {
280 int l;
281 AVHWDeviceContext *ctx = priv;
282
283 switch (severity) {
284 case VK_DEBUG_UTILS_MESSAGE_SEVERITY_VERBOSE_BIT_EXT: l = AV_LOG_VERBOSE; break;
285 case VK_DEBUG_UTILS_MESSAGE_SEVERITY_INFO_BIT_EXT: l = AV_LOG_INFO; break;
286 case VK_DEBUG_UTILS_MESSAGE_SEVERITY_WARNING_BIT_EXT: l = AV_LOG_WARNING; break;
287 case VK_DEBUG_UTILS_MESSAGE_SEVERITY_ERROR_BIT_EXT: l = AV_LOG_ERROR; break;
288 default: l = AV_LOG_DEBUG; break;
289 }
290
291 av_log(ctx, l, "%s\n", data->pMessage);
292 for (int i = 0; i < data->cmdBufLabelCount; i++)
293 av_log(ctx, l, "\t%i: %s\n", i, data->pCmdBufLabels[i].pLabelName);
294
295 return 0;
296 }
297
check_extensions(AVHWDeviceContext * ctx,int dev,AVDictionary * opts,const char * const ** dst,uint32_t * num,int debug)298 static int check_extensions(AVHWDeviceContext *ctx, int dev, AVDictionary *opts,
299 const char * const **dst, uint32_t *num, int debug)
300 {
301 const char *tstr;
302 const char **extension_names = NULL;
303 VulkanDevicePriv *p = ctx->internal->priv;
304 AVVulkanDeviceContext *hwctx = ctx->hwctx;
305 int err = 0, found, extensions_found = 0;
306
307 const char *mod;
308 int optional_exts_num;
309 uint32_t sup_ext_count;
310 char *user_exts_str = NULL;
311 AVDictionaryEntry *user_exts;
312 VkExtensionProperties *sup_ext;
313 const VulkanOptExtension *optional_exts;
314
315 if (!dev) {
316 mod = "instance";
317 optional_exts = optional_instance_exts;
318 optional_exts_num = FF_ARRAY_ELEMS(optional_instance_exts);
319 user_exts = av_dict_get(opts, "instance_extensions", NULL, 0);
320 if (user_exts) {
321 user_exts_str = av_strdup(user_exts->value);
322 if (!user_exts_str) {
323 err = AVERROR(ENOMEM);
324 goto fail;
325 }
326 }
327 vkEnumerateInstanceExtensionProperties(NULL, &sup_ext_count, NULL);
328 sup_ext = av_malloc_array(sup_ext_count, sizeof(VkExtensionProperties));
329 if (!sup_ext)
330 return AVERROR(ENOMEM);
331 vkEnumerateInstanceExtensionProperties(NULL, &sup_ext_count, sup_ext);
332 } else {
333 mod = "device";
334 optional_exts = optional_device_exts;
335 optional_exts_num = FF_ARRAY_ELEMS(optional_device_exts);
336 user_exts = av_dict_get(opts, "device_extensions", NULL, 0);
337 if (user_exts) {
338 user_exts_str = av_strdup(user_exts->value);
339 if (!user_exts_str) {
340 err = AVERROR(ENOMEM);
341 goto fail;
342 }
343 }
344 vkEnumerateDeviceExtensionProperties(hwctx->phys_dev, NULL,
345 &sup_ext_count, NULL);
346 sup_ext = av_malloc_array(sup_ext_count, sizeof(VkExtensionProperties));
347 if (!sup_ext)
348 return AVERROR(ENOMEM);
349 vkEnumerateDeviceExtensionProperties(hwctx->phys_dev, NULL,
350 &sup_ext_count, sup_ext);
351 }
352
353 for (int i = 0; i < optional_exts_num; i++) {
354 tstr = optional_exts[i].name;
355 found = 0;
356 for (int j = 0; j < sup_ext_count; j++) {
357 if (!strcmp(tstr, sup_ext[j].extensionName)) {
358 found = 1;
359 break;
360 }
361 }
362 if (!found)
363 continue;
364
365 av_log(ctx, AV_LOG_VERBOSE, "Using %s extension \"%s\"\n", mod, tstr);
366 p->extensions |= optional_exts[i].flag;
367 ADD_VAL_TO_LIST(extension_names, extensions_found, tstr);
368 }
369
370 if (debug && !dev) {
371 tstr = VK_EXT_DEBUG_UTILS_EXTENSION_NAME;
372 found = 0;
373 for (int j = 0; j < sup_ext_count; j++) {
374 if (!strcmp(tstr, sup_ext[j].extensionName)) {
375 found = 1;
376 break;
377 }
378 }
379 if (found) {
380 av_log(ctx, AV_LOG_VERBOSE, "Using %s extension \"%s\"\n", mod, tstr);
381 ADD_VAL_TO_LIST(extension_names, extensions_found, tstr);
382 } else {
383 av_log(ctx, AV_LOG_ERROR, "Debug extension \"%s\" not found!\n",
384 tstr);
385 err = AVERROR(EINVAL);
386 goto fail;
387 }
388 }
389
390 if (user_exts_str) {
391 char *save, *token = av_strtok(user_exts_str, "+", &save);
392 while (token) {
393 found = 0;
394 for (int j = 0; j < sup_ext_count; j++) {
395 if (!strcmp(token, sup_ext[j].extensionName)) {
396 found = 1;
397 break;
398 }
399 }
400 if (found) {
401 av_log(ctx, AV_LOG_VERBOSE, "Using %s extension \"%s\"\n", mod, token);
402 ADD_VAL_TO_LIST(extension_names, extensions_found, token);
403 } else {
404 av_log(ctx, AV_LOG_WARNING, "%s extension \"%s\" not found, excluding.\n",
405 mod, token);
406 }
407 token = av_strtok(NULL, "+", &save);
408 }
409 }
410
411 *dst = extension_names;
412 *num = extensions_found;
413
414 av_free(user_exts_str);
415 av_free(sup_ext);
416 return 0;
417
418 fail:
419 if (extension_names)
420 for (int i = 0; i < extensions_found; i++)
421 av_free((void *)extension_names[i]);
422 av_free(extension_names);
423 av_free(user_exts_str);
424 av_free(sup_ext);
425 return err;
426 }
427
428 /* Creates a VkInstance */
create_instance(AVHWDeviceContext * ctx,AVDictionary * opts)429 static int create_instance(AVHWDeviceContext *ctx, AVDictionary *opts)
430 {
431 int err = 0;
432 VkResult ret;
433 VulkanDevicePriv *p = ctx->internal->priv;
434 AVVulkanDeviceContext *hwctx = ctx->hwctx;
435 AVDictionaryEntry *debug_opt = av_dict_get(opts, "debug", NULL, 0);
436 const int debug_mode = debug_opt && strtol(debug_opt->value, NULL, 10);
437 VkApplicationInfo application_info = {
438 .sType = VK_STRUCTURE_TYPE_APPLICATION_INFO,
439 .pEngineName = "libavutil",
440 .apiVersion = VK_API_VERSION_1_1,
441 .engineVersion = VK_MAKE_VERSION(LIBAVUTIL_VERSION_MAJOR,
442 LIBAVUTIL_VERSION_MINOR,
443 LIBAVUTIL_VERSION_MICRO),
444 };
445 VkInstanceCreateInfo inst_props = {
446 .sType = VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO,
447 .pApplicationInfo = &application_info,
448 };
449
450 /* Check for present/missing extensions */
451 err = check_extensions(ctx, 0, opts, &inst_props.ppEnabledExtensionNames,
452 &inst_props.enabledExtensionCount, debug_mode);
453 if (err < 0)
454 return err;
455
456 if (debug_mode) {
457 static const char *layers[] = { "VK_LAYER_KHRONOS_validation" };
458 inst_props.ppEnabledLayerNames = layers;
459 inst_props.enabledLayerCount = FF_ARRAY_ELEMS(layers);
460 }
461
462 /* Try to create the instance */
463 ret = vkCreateInstance(&inst_props, hwctx->alloc, &hwctx->inst);
464
465 /* Check for errors */
466 if (ret != VK_SUCCESS) {
467 av_log(ctx, AV_LOG_ERROR, "Instance creation failure: %s\n",
468 vk_ret2str(ret));
469 for (int i = 0; i < inst_props.enabledExtensionCount; i++)
470 av_free((void *)inst_props.ppEnabledExtensionNames[i]);
471 av_free((void *)inst_props.ppEnabledExtensionNames);
472 return AVERROR_EXTERNAL;
473 }
474
475 if (debug_mode) {
476 VkDebugUtilsMessengerCreateInfoEXT dbg = {
477 .sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_MESSENGER_CREATE_INFO_EXT,
478 .messageSeverity = VK_DEBUG_UTILS_MESSAGE_SEVERITY_VERBOSE_BIT_EXT |
479 VK_DEBUG_UTILS_MESSAGE_SEVERITY_INFO_BIT_EXT |
480 VK_DEBUG_UTILS_MESSAGE_SEVERITY_WARNING_BIT_EXT |
481 VK_DEBUG_UTILS_MESSAGE_SEVERITY_ERROR_BIT_EXT,
482 .messageType = VK_DEBUG_UTILS_MESSAGE_TYPE_GENERAL_BIT_EXT |
483 VK_DEBUG_UTILS_MESSAGE_TYPE_VALIDATION_BIT_EXT |
484 VK_DEBUG_UTILS_MESSAGE_TYPE_PERFORMANCE_BIT_EXT,
485 .pfnUserCallback = vk_dbg_callback,
486 .pUserData = ctx,
487 };
488 VK_LOAD_PFN(hwctx->inst, vkCreateDebugUtilsMessengerEXT);
489
490 pfn_vkCreateDebugUtilsMessengerEXT(hwctx->inst, &dbg,
491 hwctx->alloc, &p->debug_ctx);
492 }
493
494 hwctx->enabled_inst_extensions = inst_props.ppEnabledExtensionNames;
495 hwctx->nb_enabled_inst_extensions = inst_props.enabledExtensionCount;
496
497 return 0;
498 }
499
500 typedef struct VulkanDeviceSelection {
501 uint8_t uuid[VK_UUID_SIZE]; /* Will use this first unless !has_uuid */
502 int has_uuid;
503 const char *name; /* Will use this second unless NULL */
504 uint32_t pci_device; /* Will use this third unless 0x0 */
505 uint32_t vendor_id; /* Last resort to find something deterministic */
506 int index; /* Finally fall back to index */
507 } VulkanDeviceSelection;
508
vk_dev_type(enum VkPhysicalDeviceType type)509 static const char *vk_dev_type(enum VkPhysicalDeviceType type)
510 {
511 switch (type) {
512 case VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU: return "integrated";
513 case VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU: return "discrete";
514 case VK_PHYSICAL_DEVICE_TYPE_VIRTUAL_GPU: return "virtual";
515 case VK_PHYSICAL_DEVICE_TYPE_CPU: return "software";
516 default: return "unknown";
517 }
518 }
519
520 /* Finds a device */
find_device(AVHWDeviceContext * ctx,VulkanDeviceSelection * select)521 static int find_device(AVHWDeviceContext *ctx, VulkanDeviceSelection *select)
522 {
523 int err = 0, choice = -1;
524 uint32_t num;
525 VkResult ret;
526 VkPhysicalDevice *devices = NULL;
527 VkPhysicalDeviceIDProperties *idp = NULL;
528 VkPhysicalDeviceProperties2 *prop = NULL;
529 AVVulkanDeviceContext *hwctx = ctx->hwctx;
530
531 ret = vkEnumeratePhysicalDevices(hwctx->inst, &num, NULL);
532 if (ret != VK_SUCCESS || !num) {
533 av_log(ctx, AV_LOG_ERROR, "No devices found: %s!\n", vk_ret2str(ret));
534 return AVERROR(ENODEV);
535 }
536
537 devices = av_malloc_array(num, sizeof(VkPhysicalDevice));
538 if (!devices)
539 return AVERROR(ENOMEM);
540
541 ret = vkEnumeratePhysicalDevices(hwctx->inst, &num, devices);
542 if (ret != VK_SUCCESS) {
543 av_log(ctx, AV_LOG_ERROR, "Failed enumerating devices: %s\n",
544 vk_ret2str(ret));
545 err = AVERROR(ENODEV);
546 goto end;
547 }
548
549 prop = av_mallocz_array(num, sizeof(*prop));
550 if (!prop) {
551 err = AVERROR(ENOMEM);
552 goto end;
553 }
554
555 idp = av_mallocz_array(num, sizeof(*idp));
556 if (!idp) {
557 err = AVERROR(ENOMEM);
558 goto end;
559 }
560
561 av_log(ctx, AV_LOG_VERBOSE, "GPU listing:\n");
562 for (int i = 0; i < num; i++) {
563 idp[i].sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES;
564 prop[i].sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2;
565 prop[i].pNext = &idp[i];
566
567 vkGetPhysicalDeviceProperties2(devices[i], &prop[i]);
568 av_log(ctx, AV_LOG_VERBOSE, " %d: %s (%s) (0x%x)\n", i,
569 prop[i].properties.deviceName,
570 vk_dev_type(prop[i].properties.deviceType),
571 prop[i].properties.deviceID);
572 }
573
574 if (select->has_uuid) {
575 for (int i = 0; i < num; i++) {
576 if (!strncmp(idp[i].deviceUUID, select->uuid, VK_UUID_SIZE)) {
577 choice = i;
578 goto end;
579 }
580 }
581 av_log(ctx, AV_LOG_ERROR, "Unable to find device by given UUID!\n");
582 err = AVERROR(ENODEV);
583 goto end;
584 } else if (select->name) {
585 av_log(ctx, AV_LOG_VERBOSE, "Requested device: %s\n", select->name);
586 for (int i = 0; i < num; i++) {
587 if (strstr(prop[i].properties.deviceName, select->name)) {
588 choice = i;
589 goto end;
590 }
591 }
592 av_log(ctx, AV_LOG_ERROR, "Unable to find device \"%s\"!\n",
593 select->name);
594 err = AVERROR(ENODEV);
595 goto end;
596 } else if (select->pci_device) {
597 av_log(ctx, AV_LOG_VERBOSE, "Requested device: 0x%x\n", select->pci_device);
598 for (int i = 0; i < num; i++) {
599 if (select->pci_device == prop[i].properties.deviceID) {
600 choice = i;
601 goto end;
602 }
603 }
604 av_log(ctx, AV_LOG_ERROR, "Unable to find device with PCI ID 0x%x!\n",
605 select->pci_device);
606 err = AVERROR(EINVAL);
607 goto end;
608 } else if (select->vendor_id) {
609 av_log(ctx, AV_LOG_VERBOSE, "Requested vendor: 0x%x\n", select->vendor_id);
610 for (int i = 0; i < num; i++) {
611 if (select->vendor_id == prop[i].properties.vendorID) {
612 choice = i;
613 goto end;
614 }
615 }
616 av_log(ctx, AV_LOG_ERROR, "Unable to find device with Vendor ID 0x%x!\n",
617 select->vendor_id);
618 err = AVERROR(ENODEV);
619 goto end;
620 } else {
621 if (select->index < num) {
622 choice = select->index;
623 goto end;
624 }
625 av_log(ctx, AV_LOG_ERROR, "Unable to find device with index %i!\n",
626 select->index);
627 err = AVERROR(ENODEV);
628 goto end;
629 }
630
631 end:
632 if (choice > -1)
633 hwctx->phys_dev = devices[choice];
634
635 av_free(devices);
636 av_free(prop);
637 av_free(idp);
638
639 return err;
640 }
641
search_queue_families(AVHWDeviceContext * ctx,VkDeviceCreateInfo * cd)642 static int search_queue_families(AVHWDeviceContext *ctx, VkDeviceCreateInfo *cd)
643 {
644 uint32_t num;
645 float *weights;
646 VkQueueFamilyProperties *qs = NULL;
647 AVVulkanDeviceContext *hwctx = ctx->hwctx;
648 int graph_index = -1, comp_index = -1, tx_index = -1;
649 VkDeviceQueueCreateInfo *pc = (VkDeviceQueueCreateInfo *)cd->pQueueCreateInfos;
650
651 /* First get the number of queue families */
652 vkGetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &num, NULL);
653 if (!num) {
654 av_log(ctx, AV_LOG_ERROR, "Failed to get queues!\n");
655 return AVERROR_EXTERNAL;
656 }
657
658 /* Then allocate memory */
659 qs = av_malloc_array(num, sizeof(VkQueueFamilyProperties));
660 if (!qs)
661 return AVERROR(ENOMEM);
662
663 /* Finally retrieve the queue families */
664 vkGetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &num, qs);
665
666 #define SEARCH_FLAGS(expr, out) \
667 for (int i = 0; i < num; i++) { \
668 const VkQueueFlagBits flags = qs[i].queueFlags; \
669 if (expr) { \
670 out = i; \
671 break; \
672 } \
673 }
674
675 SEARCH_FLAGS(flags & VK_QUEUE_GRAPHICS_BIT, graph_index)
676
677 SEARCH_FLAGS((flags & VK_QUEUE_COMPUTE_BIT) && (i != graph_index),
678 comp_index)
679
680 SEARCH_FLAGS((flags & VK_QUEUE_TRANSFER_BIT) && (i != graph_index) &&
681 (i != comp_index), tx_index)
682
683 #undef SEARCH_FLAGS
684 #define ADD_QUEUE(fidx, graph, comp, tx) \
685 av_log(ctx, AV_LOG_VERBOSE, "Using queue family %i (total queues: %i) for %s%s%s\n", \
686 fidx, qs[fidx].queueCount, graph ? "graphics " : "", \
687 comp ? "compute " : "", tx ? "transfers " : ""); \
688 av_log(ctx, AV_LOG_VERBOSE, " QF %i flags: %s%s%s%s\n", fidx, \
689 ((qs[fidx].queueFlags) & VK_QUEUE_GRAPHICS_BIT) ? "(graphics) " : "", \
690 ((qs[fidx].queueFlags) & VK_QUEUE_COMPUTE_BIT) ? "(compute) " : "", \
691 ((qs[fidx].queueFlags) & VK_QUEUE_TRANSFER_BIT) ? "(transfers) " : "", \
692 ((qs[fidx].queueFlags) & VK_QUEUE_SPARSE_BINDING_BIT) ? "(sparse) " : ""); \
693 pc[cd->queueCreateInfoCount].queueFamilyIndex = fidx; \
694 pc[cd->queueCreateInfoCount].queueCount = qs[fidx].queueCount; \
695 weights = av_malloc(qs[fidx].queueCount * sizeof(float)); \
696 pc[cd->queueCreateInfoCount].pQueuePriorities = weights; \
697 if (!weights) \
698 goto fail; \
699 for (int i = 0; i < qs[fidx].queueCount; i++) \
700 weights[i] = 1.0f; \
701 cd->queueCreateInfoCount++;
702
703 ADD_QUEUE(graph_index, 1, comp_index < 0, tx_index < 0 && comp_index < 0)
704 hwctx->queue_family_index = graph_index;
705 hwctx->queue_family_comp_index = graph_index;
706 hwctx->queue_family_tx_index = graph_index;
707 hwctx->nb_graphics_queues = qs[graph_index].queueCount;
708
709 if (comp_index != -1) {
710 ADD_QUEUE(comp_index, 0, 1, tx_index < 0)
711 hwctx->queue_family_tx_index = comp_index;
712 hwctx->queue_family_comp_index = comp_index;
713 hwctx->nb_comp_queues = qs[comp_index].queueCount;
714 }
715
716 if (tx_index != -1) {
717 ADD_QUEUE(tx_index, 0, 0, 1)
718 hwctx->queue_family_tx_index = tx_index;
719 hwctx->nb_tx_queues = qs[tx_index].queueCount;
720 }
721
722 #undef ADD_QUEUE
723 av_free(qs);
724
725 return 0;
726
727 fail:
728 av_freep(&pc[0].pQueuePriorities);
729 av_freep(&pc[1].pQueuePriorities);
730 av_freep(&pc[2].pQueuePriorities);
731 av_free(qs);
732
733 return AVERROR(ENOMEM);
734 }
735
create_exec_ctx(AVHWFramesContext * hwfc,VulkanExecCtx * cmd,int queue_family_index,int num_queues)736 static int create_exec_ctx(AVHWFramesContext *hwfc, VulkanExecCtx *cmd,
737 int queue_family_index, int num_queues)
738 {
739 VkResult ret;
740 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
741
742 VkCommandPoolCreateInfo cqueue_create = {
743 .sType = VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO,
744 .flags = VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT,
745 .queueFamilyIndex = queue_family_index,
746 };
747 VkCommandBufferAllocateInfo cbuf_create = {
748 .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO,
749 .level = VK_COMMAND_BUFFER_LEVEL_PRIMARY,
750 .commandBufferCount = num_queues,
751 };
752
753 cmd->nb_queues = num_queues;
754
755 cmd->queues = av_mallocz(num_queues * sizeof(*cmd->queues));
756 if (!cmd->queues)
757 return AVERROR(ENOMEM);
758
759 cmd->bufs = av_mallocz(num_queues * sizeof(*cmd->bufs));
760 if (!cmd->bufs)
761 return AVERROR(ENOMEM);
762
763 /* Create command pool */
764 ret = vkCreateCommandPool(hwctx->act_dev, &cqueue_create,
765 hwctx->alloc, &cmd->pool);
766 if (ret != VK_SUCCESS) {
767 av_log(hwfc, AV_LOG_ERROR, "Command pool creation failure: %s\n",
768 vk_ret2str(ret));
769 return AVERROR_EXTERNAL;
770 }
771
772 cbuf_create.commandPool = cmd->pool;
773
774 /* Allocate command buffer */
775 ret = vkAllocateCommandBuffers(hwctx->act_dev, &cbuf_create, cmd->bufs);
776 if (ret != VK_SUCCESS) {
777 av_log(hwfc, AV_LOG_ERROR, "Command buffer alloc failure: %s\n",
778 vk_ret2str(ret));
779 return AVERROR_EXTERNAL;
780 }
781
782 for (int i = 0; i < num_queues; i++) {
783 VulkanQueueCtx *q = &cmd->queues[i];
784 vkGetDeviceQueue(hwctx->act_dev, queue_family_index, i, &q->queue);
785 q->was_synchronous = 1;
786 }
787
788 return 0;
789 }
790
free_exec_ctx(AVHWFramesContext * hwfc,VulkanExecCtx * cmd)791 static void free_exec_ctx(AVHWFramesContext *hwfc, VulkanExecCtx *cmd)
792 {
793 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
794
795 /* Make sure all queues have finished executing */
796 for (int i = 0; i < cmd->nb_queues; i++) {
797 VulkanQueueCtx *q = &cmd->queues[i];
798
799 if (q->fence && !q->was_synchronous) {
800 vkWaitForFences(hwctx->act_dev, 1, &q->fence, VK_TRUE, UINT64_MAX);
801 vkResetFences(hwctx->act_dev, 1, &q->fence);
802 }
803
804 /* Free the fence */
805 if (q->fence)
806 vkDestroyFence(hwctx->act_dev, q->fence, hwctx->alloc);
807
808 /* Free buffer dependencies */
809 for (int j = 0; j < q->nb_buf_deps; j++)
810 av_buffer_unref(&q->buf_deps[j]);
811 av_free(q->buf_deps);
812 }
813
814 if (cmd->bufs)
815 vkFreeCommandBuffers(hwctx->act_dev, cmd->pool, cmd->nb_queues, cmd->bufs);
816 if (cmd->pool)
817 vkDestroyCommandPool(hwctx->act_dev, cmd->pool, hwctx->alloc);
818
819 av_freep(&cmd->bufs);
820 av_freep(&cmd->queues);
821 }
822
get_buf_exec_ctx(AVHWFramesContext * hwfc,VulkanExecCtx * cmd)823 static VkCommandBuffer get_buf_exec_ctx(AVHWFramesContext *hwfc, VulkanExecCtx *cmd)
824 {
825 return cmd->bufs[cmd->cur_queue_idx];
826 }
827
unref_exec_ctx_deps(AVHWFramesContext * hwfc,VulkanExecCtx * cmd)828 static void unref_exec_ctx_deps(AVHWFramesContext *hwfc, VulkanExecCtx *cmd)
829 {
830 VulkanQueueCtx *q = &cmd->queues[cmd->cur_queue_idx];
831
832 for (int j = 0; j < q->nb_buf_deps; j++)
833 av_buffer_unref(&q->buf_deps[j]);
834 q->nb_buf_deps = 0;
835 }
836
wait_start_exec_ctx(AVHWFramesContext * hwfc,VulkanExecCtx * cmd)837 static int wait_start_exec_ctx(AVHWFramesContext *hwfc, VulkanExecCtx *cmd)
838 {
839 VkResult ret;
840 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
841 VulkanQueueCtx *q = &cmd->queues[cmd->cur_queue_idx];
842
843 VkCommandBufferBeginInfo cmd_start = {
844 .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
845 .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
846 };
847
848 /* Create the fence and don't wait for it initially */
849 if (!q->fence) {
850 VkFenceCreateInfo fence_spawn = {
851 .sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO,
852 };
853 ret = vkCreateFence(hwctx->act_dev, &fence_spawn, hwctx->alloc,
854 &q->fence);
855 if (ret != VK_SUCCESS) {
856 av_log(hwfc, AV_LOG_ERROR, "Failed to queue frame fence: %s\n",
857 vk_ret2str(ret));
858 return AVERROR_EXTERNAL;
859 }
860 } else if (!q->was_synchronous) {
861 vkWaitForFences(hwctx->act_dev, 1, &q->fence, VK_TRUE, UINT64_MAX);
862 vkResetFences(hwctx->act_dev, 1, &q->fence);
863 }
864
865 /* Discard queue dependencies */
866 unref_exec_ctx_deps(hwfc, cmd);
867
868 ret = vkBeginCommandBuffer(cmd->bufs[cmd->cur_queue_idx], &cmd_start);
869 if (ret != VK_SUCCESS) {
870 av_log(hwfc, AV_LOG_ERROR, "Unable to init command buffer: %s\n",
871 vk_ret2str(ret));
872 return AVERROR_EXTERNAL;
873 }
874
875 return 0;
876 }
877
add_buf_dep_exec_ctx(AVHWFramesContext * hwfc,VulkanExecCtx * cmd,AVBufferRef * const * deps,int nb_deps)878 static int add_buf_dep_exec_ctx(AVHWFramesContext *hwfc, VulkanExecCtx *cmd,
879 AVBufferRef * const *deps, int nb_deps)
880 {
881 AVBufferRef **dst;
882 VulkanQueueCtx *q = &cmd->queues[cmd->cur_queue_idx];
883
884 if (!deps || !nb_deps)
885 return 0;
886
887 dst = av_fast_realloc(q->buf_deps, &q->buf_deps_alloc_size,
888 (q->nb_buf_deps + nb_deps) * sizeof(*dst));
889 if (!dst)
890 goto err;
891
892 q->buf_deps = dst;
893
894 for (int i = 0; i < nb_deps; i++) {
895 q->buf_deps[q->nb_buf_deps] = av_buffer_ref(deps[i]);
896 if (!q->buf_deps[q->nb_buf_deps])
897 goto err;
898 q->nb_buf_deps++;
899 }
900
901 return 0;
902
903 err:
904 unref_exec_ctx_deps(hwfc, cmd);
905 return AVERROR(ENOMEM);
906 }
907
submit_exec_ctx(AVHWFramesContext * hwfc,VulkanExecCtx * cmd,VkSubmitInfo * s_info,int synchronous)908 static int submit_exec_ctx(AVHWFramesContext *hwfc, VulkanExecCtx *cmd,
909 VkSubmitInfo *s_info, int synchronous)
910 {
911 VkResult ret;
912 VulkanQueueCtx *q = &cmd->queues[cmd->cur_queue_idx];
913
914 ret = vkEndCommandBuffer(cmd->bufs[cmd->cur_queue_idx]);
915 if (ret != VK_SUCCESS) {
916 av_log(hwfc, AV_LOG_ERROR, "Unable to finish command buffer: %s\n",
917 vk_ret2str(ret));
918 unref_exec_ctx_deps(hwfc, cmd);
919 return AVERROR_EXTERNAL;
920 }
921
922 s_info->pCommandBuffers = &cmd->bufs[cmd->cur_queue_idx];
923 s_info->commandBufferCount = 1;
924
925 ret = vkQueueSubmit(q->queue, 1, s_info, q->fence);
926 if (ret != VK_SUCCESS) {
927 unref_exec_ctx_deps(hwfc, cmd);
928 return AVERROR_EXTERNAL;
929 }
930
931 q->was_synchronous = synchronous;
932
933 if (synchronous) {
934 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
935 vkWaitForFences(hwctx->act_dev, 1, &q->fence, VK_TRUE, UINT64_MAX);
936 vkResetFences(hwctx->act_dev, 1, &q->fence);
937 unref_exec_ctx_deps(hwfc, cmd);
938 } else { /* Rotate queues */
939 cmd->cur_queue_idx = (cmd->cur_queue_idx + 1) % cmd->nb_queues;
940 }
941
942 return 0;
943 }
944
vulkan_device_free(AVHWDeviceContext * ctx)945 static void vulkan_device_free(AVHWDeviceContext *ctx)
946 {
947 VulkanDevicePriv *p = ctx->internal->priv;
948 AVVulkanDeviceContext *hwctx = ctx->hwctx;
949
950 vkDestroyDevice(hwctx->act_dev, hwctx->alloc);
951
952 if (p->debug_ctx) {
953 VK_LOAD_PFN(hwctx->inst, vkDestroyDebugUtilsMessengerEXT);
954 pfn_vkDestroyDebugUtilsMessengerEXT(hwctx->inst, p->debug_ctx,
955 hwctx->alloc);
956 }
957
958 vkDestroyInstance(hwctx->inst, hwctx->alloc);
959
960 for (int i = 0; i < hwctx->nb_enabled_inst_extensions; i++)
961 av_free((void *)hwctx->enabled_inst_extensions[i]);
962 av_free((void *)hwctx->enabled_inst_extensions);
963
964 for (int i = 0; i < hwctx->nb_enabled_dev_extensions; i++)
965 av_free((void *)hwctx->enabled_dev_extensions[i]);
966 av_free((void *)hwctx->enabled_dev_extensions);
967 }
968
vulkan_device_create_internal(AVHWDeviceContext * ctx,VulkanDeviceSelection * dev_select,AVDictionary * opts,int flags)969 static int vulkan_device_create_internal(AVHWDeviceContext *ctx,
970 VulkanDeviceSelection *dev_select,
971 AVDictionary *opts, int flags)
972 {
973 int err = 0;
974 VkResult ret;
975 AVDictionaryEntry *opt_d;
976 VulkanDevicePriv *p = ctx->internal->priv;
977 AVVulkanDeviceContext *hwctx = ctx->hwctx;
978 VkPhysicalDeviceFeatures dev_features = { 0 };
979 VkDeviceQueueCreateInfo queue_create_info[3] = {
980 { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO, },
981 { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO, },
982 { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO, },
983 };
984
985 VkDeviceCreateInfo dev_info = {
986 .sType = VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO,
987 .pNext = &hwctx->device_features,
988 .pQueueCreateInfos = queue_create_info,
989 .queueCreateInfoCount = 0,
990 };
991
992 hwctx->device_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2;
993 ctx->free = vulkan_device_free;
994
995 /* Create an instance if not given one */
996 if ((err = create_instance(ctx, opts)))
997 goto end;
998
999 /* Find a device (if not given one) */
1000 if ((err = find_device(ctx, dev_select)))
1001 goto end;
1002
1003 vkGetPhysicalDeviceFeatures(hwctx->phys_dev, &dev_features);
1004 #define COPY_FEATURE(DST, NAME) (DST).features.NAME = dev_features.NAME;
1005 COPY_FEATURE(hwctx->device_features, shaderImageGatherExtended)
1006 COPY_FEATURE(hwctx->device_features, fragmentStoresAndAtomics)
1007 COPY_FEATURE(hwctx->device_features, vertexPipelineStoresAndAtomics)
1008 COPY_FEATURE(hwctx->device_features, shaderInt64)
1009 #undef COPY_FEATURE
1010
1011 /* Search queue family */
1012 if ((err = search_queue_families(ctx, &dev_info)))
1013 goto end;
1014
1015 if ((err = check_extensions(ctx, 1, opts, &dev_info.ppEnabledExtensionNames,
1016 &dev_info.enabledExtensionCount, 0))) {
1017 av_free((void *)queue_create_info[0].pQueuePriorities);
1018 av_free((void *)queue_create_info[1].pQueuePriorities);
1019 av_free((void *)queue_create_info[2].pQueuePriorities);
1020 goto end;
1021 }
1022
1023 ret = vkCreateDevice(hwctx->phys_dev, &dev_info, hwctx->alloc,
1024 &hwctx->act_dev);
1025
1026 av_free((void *)queue_create_info[0].pQueuePriorities);
1027 av_free((void *)queue_create_info[1].pQueuePriorities);
1028 av_free((void *)queue_create_info[2].pQueuePriorities);
1029
1030 if (ret != VK_SUCCESS) {
1031 av_log(ctx, AV_LOG_ERROR, "Device creation failure: %s\n",
1032 vk_ret2str(ret));
1033 for (int i = 0; i < dev_info.enabledExtensionCount; i++)
1034 av_free((void *)dev_info.ppEnabledExtensionNames[i]);
1035 av_free((void *)dev_info.ppEnabledExtensionNames);
1036 err = AVERROR_EXTERNAL;
1037 goto end;
1038 }
1039
1040 /* Tiled images setting, use them by default */
1041 opt_d = av_dict_get(opts, "linear_images", NULL, 0);
1042 if (opt_d)
1043 p->use_linear_images = strtol(opt_d->value, NULL, 10);
1044
1045 hwctx->enabled_dev_extensions = dev_info.ppEnabledExtensionNames;
1046 hwctx->nb_enabled_dev_extensions = dev_info.enabledExtensionCount;
1047
1048 end:
1049 return err;
1050 }
1051
vulkan_device_init(AVHWDeviceContext * ctx)1052 static int vulkan_device_init(AVHWDeviceContext *ctx)
1053 {
1054 uint32_t queue_num;
1055 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1056 VulkanDevicePriv *p = ctx->internal->priv;
1057
1058 /* Set device extension flags */
1059 for (int i = 0; i < hwctx->nb_enabled_dev_extensions; i++) {
1060 for (int j = 0; j < FF_ARRAY_ELEMS(optional_device_exts); j++) {
1061 if (!strcmp(hwctx->enabled_dev_extensions[i],
1062 optional_device_exts[j].name)) {
1063 av_log(ctx, AV_LOG_VERBOSE, "Using device extension %s\n",
1064 hwctx->enabled_dev_extensions[i]);
1065 p->extensions |= optional_device_exts[j].flag;
1066 break;
1067 }
1068 }
1069 }
1070
1071 p->props.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2;
1072 p->props.pNext = &p->hprops;
1073 p->hprops.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT;
1074
1075 vkGetPhysicalDeviceProperties2(hwctx->phys_dev, &p->props);
1076 av_log(ctx, AV_LOG_VERBOSE, "Using device: %s\n",
1077 p->props.properties.deviceName);
1078 av_log(ctx, AV_LOG_VERBOSE, "Alignments:\n");
1079 av_log(ctx, AV_LOG_VERBOSE, " optimalBufferCopyRowPitchAlignment: %li\n",
1080 p->props.properties.limits.optimalBufferCopyRowPitchAlignment);
1081 av_log(ctx, AV_LOG_VERBOSE, " minMemoryMapAlignment: %li\n",
1082 p->props.properties.limits.minMemoryMapAlignment);
1083 if (p->extensions & EXT_EXTERNAL_HOST_MEMORY)
1084 av_log(ctx, AV_LOG_VERBOSE, " minImportedHostPointerAlignment: %li\n",
1085 p->hprops.minImportedHostPointerAlignment);
1086
1087 p->dev_is_nvidia = (p->props.properties.vendorID == 0x10de);
1088
1089 vkGetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &queue_num, NULL);
1090 if (!queue_num) {
1091 av_log(ctx, AV_LOG_ERROR, "Failed to get queues!\n");
1092 return AVERROR_EXTERNAL;
1093 }
1094
1095 #define CHECK_QUEUE(type, n) \
1096 if (n >= queue_num) { \
1097 av_log(ctx, AV_LOG_ERROR, "Invalid %s queue index %i (device has %i queues)!\n", \
1098 type, n, queue_num); \
1099 return AVERROR(EINVAL); \
1100 }
1101
1102 CHECK_QUEUE("graphics", hwctx->queue_family_index)
1103 CHECK_QUEUE("upload", hwctx->queue_family_tx_index)
1104 CHECK_QUEUE("compute", hwctx->queue_family_comp_index)
1105
1106 #undef CHECK_QUEUE
1107
1108 p->qfs[p->num_qfs++] = hwctx->queue_family_index;
1109 if ((hwctx->queue_family_tx_index != hwctx->queue_family_index) &&
1110 (hwctx->queue_family_tx_index != hwctx->queue_family_comp_index))
1111 p->qfs[p->num_qfs++] = hwctx->queue_family_tx_index;
1112 if ((hwctx->queue_family_comp_index != hwctx->queue_family_index) &&
1113 (hwctx->queue_family_comp_index != hwctx->queue_family_tx_index))
1114 p->qfs[p->num_qfs++] = hwctx->queue_family_comp_index;
1115
1116 /* Get device capabilities */
1117 vkGetPhysicalDeviceMemoryProperties(hwctx->phys_dev, &p->mprops);
1118
1119 return 0;
1120 }
1121
vulkan_device_create(AVHWDeviceContext * ctx,const char * device,AVDictionary * opts,int flags)1122 static int vulkan_device_create(AVHWDeviceContext *ctx, const char *device,
1123 AVDictionary *opts, int flags)
1124 {
1125 VulkanDeviceSelection dev_select = { 0 };
1126 if (device && device[0]) {
1127 char *end = NULL;
1128 dev_select.index = strtol(device, &end, 10);
1129 if (end == device) {
1130 dev_select.index = 0;
1131 dev_select.name = device;
1132 }
1133 }
1134
1135 return vulkan_device_create_internal(ctx, &dev_select, opts, flags);
1136 }
1137
vulkan_device_derive(AVHWDeviceContext * ctx,AVHWDeviceContext * src_ctx,AVDictionary * opts,int flags)1138 static int vulkan_device_derive(AVHWDeviceContext *ctx,
1139 AVHWDeviceContext *src_ctx,
1140 AVDictionary *opts, int flags)
1141 {
1142 av_unused VulkanDeviceSelection dev_select = { 0 };
1143
1144 /* If there's only one device on the system, then even if its not covered
1145 * by the following checks (e.g. non-PCIe ARM GPU), having an empty
1146 * dev_select will mean it'll get picked. */
1147 switch(src_ctx->type) {
1148 #if CONFIG_LIBDRM
1149 #if CONFIG_VAAPI
1150 case AV_HWDEVICE_TYPE_VAAPI: {
1151 AVVAAPIDeviceContext *src_hwctx = src_ctx->hwctx;
1152
1153 const char *vendor = vaQueryVendorString(src_hwctx->display);
1154 if (!vendor) {
1155 av_log(ctx, AV_LOG_ERROR, "Unable to get device info from VAAPI!\n");
1156 return AVERROR_EXTERNAL;
1157 }
1158
1159 if (strstr(vendor, "Intel"))
1160 dev_select.vendor_id = 0x8086;
1161 if (strstr(vendor, "AMD"))
1162 dev_select.vendor_id = 0x1002;
1163
1164 return vulkan_device_create_internal(ctx, &dev_select, opts, flags);
1165 }
1166 #endif
1167 case AV_HWDEVICE_TYPE_DRM: {
1168 AVDRMDeviceContext *src_hwctx = src_ctx->hwctx;
1169
1170 drmDevice *drm_dev_info;
1171 int err = drmGetDevice(src_hwctx->fd, &drm_dev_info);
1172 if (err) {
1173 av_log(ctx, AV_LOG_ERROR, "Unable to get device info from DRM fd!\n");
1174 return AVERROR_EXTERNAL;
1175 }
1176
1177 if (drm_dev_info->bustype == DRM_BUS_PCI)
1178 dev_select.pci_device = drm_dev_info->deviceinfo.pci->device_id;
1179
1180 drmFreeDevice(&drm_dev_info);
1181
1182 return vulkan_device_create_internal(ctx, &dev_select, opts, flags);
1183 }
1184 #endif
1185 #if CONFIG_CUDA
1186 case AV_HWDEVICE_TYPE_CUDA: {
1187 AVHWDeviceContext *cuda_cu = src_ctx;
1188 AVCUDADeviceContext *src_hwctx = src_ctx->hwctx;
1189 AVCUDADeviceContextInternal *cu_internal = src_hwctx->internal;
1190 CudaFunctions *cu = cu_internal->cuda_dl;
1191
1192 int ret = CHECK_CU(cu->cuDeviceGetUuid((CUuuid *)&dev_select.uuid,
1193 cu_internal->cuda_device));
1194 if (ret < 0) {
1195 av_log(ctx, AV_LOG_ERROR, "Unable to get UUID from CUDA!\n");
1196 return AVERROR_EXTERNAL;
1197 }
1198
1199 dev_select.has_uuid = 1;
1200
1201 return vulkan_device_create_internal(ctx, &dev_select, opts, flags);
1202 }
1203 #endif
1204 default:
1205 return AVERROR(ENOSYS);
1206 }
1207 }
1208
vulkan_frames_get_constraints(AVHWDeviceContext * ctx,const void * hwconfig,AVHWFramesConstraints * constraints)1209 static int vulkan_frames_get_constraints(AVHWDeviceContext *ctx,
1210 const void *hwconfig,
1211 AVHWFramesConstraints *constraints)
1212 {
1213 int count = 0;
1214 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1215 VulkanDevicePriv *p = ctx->internal->priv;
1216
1217 for (enum AVPixelFormat i = 0; i < AV_PIX_FMT_NB; i++)
1218 count += pixfmt_is_supported(hwctx, i, p->use_linear_images);
1219
1220 #if CONFIG_CUDA
1221 if (p->dev_is_nvidia)
1222 count++;
1223 #endif
1224
1225 constraints->valid_sw_formats = av_malloc_array(count + 1,
1226 sizeof(enum AVPixelFormat));
1227 if (!constraints->valid_sw_formats)
1228 return AVERROR(ENOMEM);
1229
1230 count = 0;
1231 for (enum AVPixelFormat i = 0; i < AV_PIX_FMT_NB; i++)
1232 if (pixfmt_is_supported(hwctx, i, p->use_linear_images))
1233 constraints->valid_sw_formats[count++] = i;
1234
1235 #if CONFIG_CUDA
1236 if (p->dev_is_nvidia)
1237 constraints->valid_sw_formats[count++] = AV_PIX_FMT_CUDA;
1238 #endif
1239 constraints->valid_sw_formats[count++] = AV_PIX_FMT_NONE;
1240
1241 constraints->min_width = 0;
1242 constraints->min_height = 0;
1243 constraints->max_width = p->props.properties.limits.maxImageDimension2D;
1244 constraints->max_height = p->props.properties.limits.maxImageDimension2D;
1245
1246 constraints->valid_hw_formats = av_malloc_array(2, sizeof(enum AVPixelFormat));
1247 if (!constraints->valid_hw_formats)
1248 return AVERROR(ENOMEM);
1249
1250 constraints->valid_hw_formats[0] = AV_PIX_FMT_VULKAN;
1251 constraints->valid_hw_formats[1] = AV_PIX_FMT_NONE;
1252
1253 return 0;
1254 }
1255
alloc_mem(AVHWDeviceContext * ctx,VkMemoryRequirements * req,VkMemoryPropertyFlagBits req_flags,const void * alloc_extension,VkMemoryPropertyFlagBits * mem_flags,VkDeviceMemory * mem)1256 static int alloc_mem(AVHWDeviceContext *ctx, VkMemoryRequirements *req,
1257 VkMemoryPropertyFlagBits req_flags, const void *alloc_extension,
1258 VkMemoryPropertyFlagBits *mem_flags, VkDeviceMemory *mem)
1259 {
1260 VkResult ret;
1261 int index = -1;
1262 VulkanDevicePriv *p = ctx->internal->priv;
1263 AVVulkanDeviceContext *dev_hwctx = ctx->hwctx;
1264 VkMemoryAllocateInfo alloc_info = {
1265 .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
1266 .pNext = alloc_extension,
1267 .allocationSize = req->size,
1268 };
1269
1270 /* The vulkan spec requires memory types to be sorted in the "optimal"
1271 * order, so the first matching type we find will be the best/fastest one */
1272 for (int i = 0; i < p->mprops.memoryTypeCount; i++) {
1273 /* The memory type must be supported by the requirements (bitfield) */
1274 if (!(req->memoryTypeBits & (1 << i)))
1275 continue;
1276
1277 /* The memory type flags must include our properties */
1278 if ((p->mprops.memoryTypes[i].propertyFlags & req_flags) != req_flags)
1279 continue;
1280
1281 /* Found a suitable memory type */
1282 index = i;
1283 break;
1284 }
1285
1286 if (index < 0) {
1287 av_log(ctx, AV_LOG_ERROR, "No memory type found for flags 0x%x\n",
1288 req_flags);
1289 return AVERROR(EINVAL);
1290 }
1291
1292 alloc_info.memoryTypeIndex = index;
1293
1294 ret = vkAllocateMemory(dev_hwctx->act_dev, &alloc_info,
1295 dev_hwctx->alloc, mem);
1296 if (ret != VK_SUCCESS) {
1297 av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory: %s\n",
1298 vk_ret2str(ret));
1299 return AVERROR(ENOMEM);
1300 }
1301
1302 *mem_flags |= p->mprops.memoryTypes[index].propertyFlags;
1303
1304 return 0;
1305 }
1306
vulkan_free_internal(AVVkFrameInternal * internal)1307 static void vulkan_free_internal(AVVkFrameInternal *internal)
1308 {
1309 if (!internal)
1310 return;
1311
1312 #if CONFIG_CUDA
1313 if (internal->cuda_fc_ref) {
1314 AVHWFramesContext *cuda_fc = (AVHWFramesContext *)internal->cuda_fc_ref->data;
1315 int planes = av_pix_fmt_count_planes(cuda_fc->sw_format);
1316 AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
1317 AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
1318 AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
1319 CudaFunctions *cu = cu_internal->cuda_dl;
1320
1321 for (int i = 0; i < planes; i++) {
1322 if (internal->cu_sem[i])
1323 CHECK_CU(cu->cuDestroyExternalSemaphore(internal->cu_sem[i]));
1324 if (internal->cu_mma[i])
1325 CHECK_CU(cu->cuMipmappedArrayDestroy(internal->cu_mma[i]));
1326 if (internal->ext_mem[i])
1327 CHECK_CU(cu->cuDestroyExternalMemory(internal->ext_mem[i]));
1328 }
1329
1330 av_buffer_unref(&internal->cuda_fc_ref);
1331 }
1332 #endif
1333
1334 av_free(internal);
1335 }
1336
vulkan_frame_free(void * opaque,uint8_t * data)1337 static void vulkan_frame_free(void *opaque, uint8_t *data)
1338 {
1339 AVVkFrame *f = (AVVkFrame *)data;
1340 AVHWFramesContext *hwfc = opaque;
1341 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1342 int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1343
1344 vulkan_free_internal(f->internal);
1345
1346 for (int i = 0; i < planes; i++) {
1347 vkDestroyImage(hwctx->act_dev, f->img[i], hwctx->alloc);
1348 vkFreeMemory(hwctx->act_dev, f->mem[i], hwctx->alloc);
1349 vkDestroySemaphore(hwctx->act_dev, f->sem[i], hwctx->alloc);
1350 }
1351
1352 av_free(f);
1353 }
1354
alloc_bind_mem(AVHWFramesContext * hwfc,AVVkFrame * f,void * alloc_pnext,size_t alloc_pnext_stride)1355 static int alloc_bind_mem(AVHWFramesContext *hwfc, AVVkFrame *f,
1356 void *alloc_pnext, size_t alloc_pnext_stride)
1357 {
1358 int err;
1359 VkResult ret;
1360 AVHWDeviceContext *ctx = hwfc->device_ctx;
1361 VulkanDevicePriv *p = ctx->internal->priv;
1362 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1363 VkBindImageMemoryInfo bind_info[AV_NUM_DATA_POINTERS] = { { 0 } };
1364
1365 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1366
1367 for (int i = 0; i < planes; i++) {
1368 int use_ded_mem;
1369 VkImageMemoryRequirementsInfo2 req_desc = {
1370 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_REQUIREMENTS_INFO_2,
1371 .image = f->img[i],
1372 };
1373 VkMemoryDedicatedAllocateInfo ded_alloc = {
1374 .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO,
1375 .pNext = (void *)(((uint8_t *)alloc_pnext) + i*alloc_pnext_stride),
1376 };
1377 VkMemoryDedicatedRequirements ded_req = {
1378 .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS,
1379 };
1380 VkMemoryRequirements2 req = {
1381 .sType = VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2,
1382 .pNext = &ded_req,
1383 };
1384
1385 vkGetImageMemoryRequirements2(hwctx->act_dev, &req_desc, &req);
1386
1387 if (f->tiling == VK_IMAGE_TILING_LINEAR)
1388 req.memoryRequirements.size = FFALIGN(req.memoryRequirements.size,
1389 p->props.properties.limits.minMemoryMapAlignment);
1390
1391 /* In case the implementation prefers/requires dedicated allocation */
1392 use_ded_mem = ded_req.prefersDedicatedAllocation |
1393 ded_req.requiresDedicatedAllocation;
1394 if (use_ded_mem)
1395 ded_alloc.image = f->img[i];
1396
1397 /* Allocate memory */
1398 if ((err = alloc_mem(ctx, &req.memoryRequirements,
1399 f->tiling == VK_IMAGE_TILING_LINEAR ?
1400 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT :
1401 VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
1402 use_ded_mem ? &ded_alloc : (void *)ded_alloc.pNext,
1403 &f->flags, &f->mem[i])))
1404 return err;
1405
1406 f->size[i] = req.memoryRequirements.size;
1407 bind_info[i].sType = VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO;
1408 bind_info[i].image = f->img[i];
1409 bind_info[i].memory = f->mem[i];
1410 }
1411
1412 /* Bind the allocated memory to the images */
1413 ret = vkBindImageMemory2(hwctx->act_dev, planes, bind_info);
1414 if (ret != VK_SUCCESS) {
1415 av_log(ctx, AV_LOG_ERROR, "Failed to bind memory: %s\n",
1416 vk_ret2str(ret));
1417 return AVERROR_EXTERNAL;
1418 }
1419
1420 return 0;
1421 }
1422
1423 enum PrepMode {
1424 PREP_MODE_WRITE,
1425 PREP_MODE_RO_SHADER,
1426 PREP_MODE_EXTERNAL_EXPORT,
1427 };
1428
prepare_frame(AVHWFramesContext * hwfc,VulkanExecCtx * ectx,AVVkFrame * frame,enum PrepMode pmode)1429 static int prepare_frame(AVHWFramesContext *hwfc, VulkanExecCtx *ectx,
1430 AVVkFrame *frame, enum PrepMode pmode)
1431 {
1432 int err;
1433 uint32_t dst_qf;
1434 VkImageLayout new_layout;
1435 VkAccessFlags new_access;
1436 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1437
1438 VkImageMemoryBarrier img_bar[AV_NUM_DATA_POINTERS] = { 0 };
1439
1440 VkSubmitInfo s_info = {
1441 .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
1442 .pSignalSemaphores = frame->sem,
1443 .signalSemaphoreCount = planes,
1444 };
1445
1446 VkPipelineStageFlagBits wait_st[AV_NUM_DATA_POINTERS];
1447 for (int i = 0; i < planes; i++)
1448 wait_st[i] = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
1449
1450 switch (pmode) {
1451 case PREP_MODE_WRITE:
1452 new_layout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
1453 new_access = VK_ACCESS_TRANSFER_WRITE_BIT;
1454 dst_qf = VK_QUEUE_FAMILY_IGNORED;
1455 break;
1456 case PREP_MODE_RO_SHADER:
1457 new_layout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL;
1458 new_access = VK_ACCESS_TRANSFER_READ_BIT;
1459 dst_qf = VK_QUEUE_FAMILY_IGNORED;
1460 break;
1461 case PREP_MODE_EXTERNAL_EXPORT:
1462 new_layout = VK_IMAGE_LAYOUT_GENERAL;
1463 new_access = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT;
1464 dst_qf = VK_QUEUE_FAMILY_EXTERNAL_KHR;
1465 s_info.pWaitSemaphores = frame->sem;
1466 s_info.pWaitDstStageMask = wait_st;
1467 s_info.waitSemaphoreCount = planes;
1468 break;
1469 }
1470
1471 if ((err = wait_start_exec_ctx(hwfc, ectx)))
1472 return err;
1473
1474 /* Change the image layout to something more optimal for writes.
1475 * This also signals the newly created semaphore, making it usable
1476 * for synchronization */
1477 for (int i = 0; i < planes; i++) {
1478 img_bar[i].sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
1479 img_bar[i].srcAccessMask = 0x0;
1480 img_bar[i].dstAccessMask = new_access;
1481 img_bar[i].oldLayout = frame->layout[i];
1482 img_bar[i].newLayout = new_layout;
1483 img_bar[i].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
1484 img_bar[i].dstQueueFamilyIndex = dst_qf;
1485 img_bar[i].image = frame->img[i];
1486 img_bar[i].subresourceRange.levelCount = 1;
1487 img_bar[i].subresourceRange.layerCount = 1;
1488 img_bar[i].subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
1489
1490 frame->layout[i] = img_bar[i].newLayout;
1491 frame->access[i] = img_bar[i].dstAccessMask;
1492 }
1493
1494 vkCmdPipelineBarrier(get_buf_exec_ctx(hwfc, ectx),
1495 VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
1496 VK_PIPELINE_STAGE_TRANSFER_BIT,
1497 0, 0, NULL, 0, NULL, planes, img_bar);
1498
1499 return submit_exec_ctx(hwfc, ectx, &s_info, 0);
1500 }
1501
create_frame(AVHWFramesContext * hwfc,AVVkFrame ** frame,VkImageTiling tiling,VkImageUsageFlagBits usage,void * create_pnext)1502 static int create_frame(AVHWFramesContext *hwfc, AVVkFrame **frame,
1503 VkImageTiling tiling, VkImageUsageFlagBits usage,
1504 void *create_pnext)
1505 {
1506 int err;
1507 VkResult ret;
1508 AVHWDeviceContext *ctx = hwfc->device_ctx;
1509 VulkanDevicePriv *p = ctx->internal->priv;
1510 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1511 enum AVPixelFormat format = hwfc->sw_format;
1512 const VkFormat *img_fmts = av_vkfmt_from_pixfmt(format);
1513 const int planes = av_pix_fmt_count_planes(format);
1514
1515 VkExportSemaphoreCreateInfo ext_sem_info = {
1516 .sType = VK_STRUCTURE_TYPE_EXPORT_SEMAPHORE_CREATE_INFO,
1517 .handleTypes = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT,
1518 };
1519
1520 VkSemaphoreCreateInfo sem_spawn = {
1521 .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
1522 .pNext = p->extensions & EXT_EXTERNAL_FD_SEM ? &ext_sem_info : NULL,
1523 };
1524
1525 AVVkFrame *f = av_vk_frame_alloc();
1526 if (!f) {
1527 av_log(ctx, AV_LOG_ERROR, "Unable to allocate memory for AVVkFrame!\n");
1528 return AVERROR(ENOMEM);
1529 }
1530
1531 /* Create the images */
1532 for (int i = 0; i < planes; i++) {
1533 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(format);
1534 int w = hwfc->width;
1535 int h = hwfc->height;
1536 const int p_w = i > 0 ? AV_CEIL_RSHIFT(w, desc->log2_chroma_w) : w;
1537 const int p_h = i > 0 ? AV_CEIL_RSHIFT(h, desc->log2_chroma_h) : h;
1538
1539 VkImageCreateInfo image_create_info = {
1540 .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
1541 .pNext = create_pnext,
1542 .imageType = VK_IMAGE_TYPE_2D,
1543 .format = img_fmts[i],
1544 .extent.width = p_w,
1545 .extent.height = p_h,
1546 .extent.depth = 1,
1547 .mipLevels = 1,
1548 .arrayLayers = 1,
1549 .flags = VK_IMAGE_CREATE_ALIAS_BIT,
1550 .tiling = tiling,
1551 .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED,
1552 .usage = usage,
1553 .samples = VK_SAMPLE_COUNT_1_BIT,
1554 .pQueueFamilyIndices = p->qfs,
1555 .queueFamilyIndexCount = p->num_qfs,
1556 .sharingMode = p->num_qfs > 1 ? VK_SHARING_MODE_CONCURRENT :
1557 VK_SHARING_MODE_EXCLUSIVE,
1558 };
1559
1560 ret = vkCreateImage(hwctx->act_dev, &image_create_info,
1561 hwctx->alloc, &f->img[i]);
1562 if (ret != VK_SUCCESS) {
1563 av_log(ctx, AV_LOG_ERROR, "Image creation failure: %s\n",
1564 vk_ret2str(ret));
1565 err = AVERROR(EINVAL);
1566 goto fail;
1567 }
1568
1569 /* Create semaphore */
1570 ret = vkCreateSemaphore(hwctx->act_dev, &sem_spawn,
1571 hwctx->alloc, &f->sem[i]);
1572 if (ret != VK_SUCCESS) {
1573 av_log(hwctx, AV_LOG_ERROR, "Failed to create semaphore: %s\n",
1574 vk_ret2str(ret));
1575 return AVERROR_EXTERNAL;
1576 }
1577
1578 f->layout[i] = image_create_info.initialLayout;
1579 f->access[i] = 0x0;
1580 }
1581
1582 f->flags = 0x0;
1583 f->tiling = tiling;
1584
1585 *frame = f;
1586 return 0;
1587
1588 fail:
1589 vulkan_frame_free(hwfc, (uint8_t *)f);
1590 return err;
1591 }
1592
1593 /* 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)1594 static void try_export_flags(AVHWFramesContext *hwfc,
1595 VkExternalMemoryHandleTypeFlags *comp_handle_types,
1596 VkExternalMemoryHandleTypeFlagBits *iexp,
1597 VkExternalMemoryHandleTypeFlagBits exp)
1598 {
1599 VkResult ret;
1600 AVVulkanFramesContext *hwctx = hwfc->hwctx;
1601 AVVulkanDeviceContext *dev_hwctx = hwfc->device_ctx->hwctx;
1602 VkExternalImageFormatProperties eprops = {
1603 .sType = VK_STRUCTURE_TYPE_EXTERNAL_IMAGE_FORMAT_PROPERTIES_KHR,
1604 };
1605 VkImageFormatProperties2 props = {
1606 .sType = VK_STRUCTURE_TYPE_IMAGE_FORMAT_PROPERTIES_2,
1607 .pNext = &eprops,
1608 };
1609 VkPhysicalDeviceExternalImageFormatInfo enext = {
1610 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_IMAGE_FORMAT_INFO,
1611 .handleType = exp,
1612 };
1613 VkPhysicalDeviceImageFormatInfo2 pinfo = {
1614 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_FORMAT_INFO_2,
1615 .pNext = !exp ? NULL : &enext,
1616 .format = av_vkfmt_from_pixfmt(hwfc->sw_format)[0],
1617 .type = VK_IMAGE_TYPE_2D,
1618 .tiling = hwctx->tiling,
1619 .usage = hwctx->usage,
1620 .flags = VK_IMAGE_CREATE_ALIAS_BIT,
1621 };
1622
1623 ret = vkGetPhysicalDeviceImageFormatProperties2(dev_hwctx->phys_dev,
1624 &pinfo, &props);
1625 if (ret == VK_SUCCESS) {
1626 *iexp |= exp;
1627 *comp_handle_types |= eprops.externalMemoryProperties.compatibleHandleTypes;
1628 }
1629 }
1630
vulkan_pool_alloc(void * opaque,int size)1631 static AVBufferRef *vulkan_pool_alloc(void *opaque, int size)
1632 {
1633 int err;
1634 AVVkFrame *f;
1635 AVBufferRef *avbuf = NULL;
1636 AVHWFramesContext *hwfc = opaque;
1637 AVVulkanFramesContext *hwctx = hwfc->hwctx;
1638 VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
1639 VulkanFramesPriv *fp = hwfc->internal->priv;
1640 VkExportMemoryAllocateInfo eminfo[AV_NUM_DATA_POINTERS];
1641 VkExternalMemoryHandleTypeFlags e = 0x0;
1642
1643 VkExternalMemoryImageCreateInfo eiinfo = {
1644 .sType = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO,
1645 .pNext = hwctx->create_pnext,
1646 };
1647
1648 if (p->extensions & EXT_EXTERNAL_FD_MEMORY)
1649 try_export_flags(hwfc, &eiinfo.handleTypes, &e,
1650 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT);
1651
1652 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
1653 try_export_flags(hwfc, &eiinfo.handleTypes, &e,
1654 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
1655
1656 for (int i = 0; i < av_pix_fmt_count_planes(hwfc->sw_format); i++) {
1657 eminfo[i].sType = VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO;
1658 eminfo[i].pNext = hwctx->alloc_pnext[i];
1659 eminfo[i].handleTypes = e;
1660 }
1661
1662 err = create_frame(hwfc, &f, hwctx->tiling, hwctx->usage,
1663 eiinfo.handleTypes ? &eiinfo : NULL);
1664 if (err)
1665 return NULL;
1666
1667 err = alloc_bind_mem(hwfc, f, eminfo, sizeof(*eminfo));
1668 if (err)
1669 goto fail;
1670
1671 err = prepare_frame(hwfc, &fp->conv_ctx, f, PREP_MODE_WRITE);
1672 if (err)
1673 goto fail;
1674
1675 avbuf = av_buffer_create((uint8_t *)f, sizeof(AVVkFrame),
1676 vulkan_frame_free, hwfc, 0);
1677 if (!avbuf)
1678 goto fail;
1679
1680 return avbuf;
1681
1682 fail:
1683 vulkan_frame_free(hwfc, (uint8_t *)f);
1684 return NULL;
1685 }
1686
vulkan_frames_uninit(AVHWFramesContext * hwfc)1687 static void vulkan_frames_uninit(AVHWFramesContext *hwfc)
1688 {
1689 VulkanFramesPriv *fp = hwfc->internal->priv;
1690
1691 free_exec_ctx(hwfc, &fp->conv_ctx);
1692 free_exec_ctx(hwfc, &fp->upload_ctx);
1693 free_exec_ctx(hwfc, &fp->download_ctx);
1694 }
1695
vulkan_frames_init(AVHWFramesContext * hwfc)1696 static int vulkan_frames_init(AVHWFramesContext *hwfc)
1697 {
1698 int err;
1699 AVVkFrame *f;
1700 AVVulkanFramesContext *hwctx = hwfc->hwctx;
1701 VulkanFramesPriv *fp = hwfc->internal->priv;
1702 AVVulkanDeviceContext *dev_hwctx = hwfc->device_ctx->hwctx;
1703 VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
1704
1705 /* Default pool flags */
1706 hwctx->tiling = hwctx->tiling ? hwctx->tiling : p->use_linear_images ?
1707 VK_IMAGE_TILING_LINEAR : VK_IMAGE_TILING_OPTIMAL;
1708
1709 if (!hwctx->usage)
1710 hwctx->usage = DEFAULT_USAGE_FLAGS;
1711
1712 err = create_exec_ctx(hwfc, &fp->conv_ctx,
1713 dev_hwctx->queue_family_comp_index,
1714 GET_QUEUE_COUNT(dev_hwctx, 0, 1, 0));
1715 if (err)
1716 goto fail;
1717
1718 err = create_exec_ctx(hwfc, &fp->upload_ctx,
1719 dev_hwctx->queue_family_tx_index,
1720 GET_QUEUE_COUNT(dev_hwctx, 0, 0, 1));
1721 if (err)
1722 goto fail;
1723
1724 err = create_exec_ctx(hwfc, &fp->download_ctx,
1725 dev_hwctx->queue_family_tx_index, 1);
1726 if (err)
1727 goto fail;
1728
1729 /* Test to see if allocation will fail */
1730 err = create_frame(hwfc, &f, hwctx->tiling, hwctx->usage,
1731 hwctx->create_pnext);
1732 if (err)
1733 goto fail;
1734
1735 vulkan_frame_free(hwfc, (uint8_t *)f);
1736
1737 /* If user did not specify a pool, hwfc->pool will be set to the internal one
1738 * in hwcontext.c just after this gets called */
1739 if (!hwfc->pool) {
1740 hwfc->internal->pool_internal = av_buffer_pool_init2(sizeof(AVVkFrame),
1741 hwfc, vulkan_pool_alloc,
1742 NULL);
1743 if (!hwfc->internal->pool_internal) {
1744 err = AVERROR(ENOMEM);
1745 goto fail;
1746 }
1747 }
1748
1749 return 0;
1750
1751 fail:
1752 free_exec_ctx(hwfc, &fp->conv_ctx);
1753 free_exec_ctx(hwfc, &fp->upload_ctx);
1754 free_exec_ctx(hwfc, &fp->download_ctx);
1755
1756 return err;
1757 }
1758
vulkan_get_buffer(AVHWFramesContext * hwfc,AVFrame * frame)1759 static int vulkan_get_buffer(AVHWFramesContext *hwfc, AVFrame *frame)
1760 {
1761 frame->buf[0] = av_buffer_pool_get(hwfc->pool);
1762 if (!frame->buf[0])
1763 return AVERROR(ENOMEM);
1764
1765 frame->data[0] = frame->buf[0]->data;
1766 frame->format = AV_PIX_FMT_VULKAN;
1767 frame->width = hwfc->width;
1768 frame->height = hwfc->height;
1769
1770 return 0;
1771 }
1772
vulkan_transfer_get_formats(AVHWFramesContext * hwfc,enum AVHWFrameTransferDirection dir,enum AVPixelFormat ** formats)1773 static int vulkan_transfer_get_formats(AVHWFramesContext *hwfc,
1774 enum AVHWFrameTransferDirection dir,
1775 enum AVPixelFormat **formats)
1776 {
1777 enum AVPixelFormat *fmts = av_malloc_array(2, sizeof(*fmts));
1778 if (!fmts)
1779 return AVERROR(ENOMEM);
1780
1781 fmts[0] = hwfc->sw_format;
1782 fmts[1] = AV_PIX_FMT_NONE;
1783
1784 *formats = fmts;
1785 return 0;
1786 }
1787
1788 typedef struct VulkanMapping {
1789 AVVkFrame *frame;
1790 int flags;
1791 } VulkanMapping;
1792
vulkan_unmap_frame(AVHWFramesContext * hwfc,HWMapDescriptor * hwmap)1793 static void vulkan_unmap_frame(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
1794 {
1795 VulkanMapping *map = hwmap->priv;
1796 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1797 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1798
1799 /* Check if buffer needs flushing */
1800 if ((map->flags & AV_HWFRAME_MAP_WRITE) &&
1801 !(map->frame->flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)) {
1802 VkResult ret;
1803 VkMappedMemoryRange flush_ranges[AV_NUM_DATA_POINTERS] = { { 0 } };
1804
1805 for (int i = 0; i < planes; i++) {
1806 flush_ranges[i].sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
1807 flush_ranges[i].memory = map->frame->mem[i];
1808 flush_ranges[i].size = VK_WHOLE_SIZE;
1809 }
1810
1811 ret = vkFlushMappedMemoryRanges(hwctx->act_dev, planes,
1812 flush_ranges);
1813 if (ret != VK_SUCCESS) {
1814 av_log(hwfc, AV_LOG_ERROR, "Failed to flush memory: %s\n",
1815 vk_ret2str(ret));
1816 }
1817 }
1818
1819 for (int i = 0; i < planes; i++)
1820 vkUnmapMemory(hwctx->act_dev, map->frame->mem[i]);
1821
1822 av_free(map);
1823 }
1824
vulkan_map_frame_to_mem(AVHWFramesContext * hwfc,AVFrame * dst,const AVFrame * src,int flags)1825 static int vulkan_map_frame_to_mem(AVHWFramesContext *hwfc, AVFrame *dst,
1826 const AVFrame *src, int flags)
1827 {
1828 VkResult ret;
1829 int err, mapped_mem_count = 0;
1830 AVVkFrame *f = (AVVkFrame *)src->data[0];
1831 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1832 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1833
1834 VulkanMapping *map = av_mallocz(sizeof(VulkanMapping));
1835 if (!map)
1836 return AVERROR(EINVAL);
1837
1838 if (src->format != AV_PIX_FMT_VULKAN) {
1839 av_log(hwfc, AV_LOG_ERROR, "Cannot map from pixel format %s!\n",
1840 av_get_pix_fmt_name(src->format));
1841 err = AVERROR(EINVAL);
1842 goto fail;
1843 }
1844
1845 if (!(f->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) ||
1846 !(f->tiling == VK_IMAGE_TILING_LINEAR)) {
1847 av_log(hwfc, AV_LOG_ERROR, "Unable to map frame, not host visible "
1848 "and linear!\n");
1849 err = AVERROR(EINVAL);
1850 goto fail;
1851 }
1852
1853 dst->width = src->width;
1854 dst->height = src->height;
1855
1856 for (int i = 0; i < planes; i++) {
1857 ret = vkMapMemory(hwctx->act_dev, f->mem[i], 0,
1858 VK_WHOLE_SIZE, 0, (void **)&dst->data[i]);
1859 if (ret != VK_SUCCESS) {
1860 av_log(hwfc, AV_LOG_ERROR, "Failed to map image memory: %s\n",
1861 vk_ret2str(ret));
1862 err = AVERROR_EXTERNAL;
1863 goto fail;
1864 }
1865 mapped_mem_count++;
1866 }
1867
1868 /* Check if the memory contents matter */
1869 if (((flags & AV_HWFRAME_MAP_READ) || !(flags & AV_HWFRAME_MAP_OVERWRITE)) &&
1870 !(f->flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)) {
1871 VkMappedMemoryRange map_mem_ranges[AV_NUM_DATA_POINTERS] = { { 0 } };
1872 for (int i = 0; i < planes; i++) {
1873 map_mem_ranges[i].sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
1874 map_mem_ranges[i].size = VK_WHOLE_SIZE;
1875 map_mem_ranges[i].memory = f->mem[i];
1876 }
1877
1878 ret = vkInvalidateMappedMemoryRanges(hwctx->act_dev, planes,
1879 map_mem_ranges);
1880 if (ret != VK_SUCCESS) {
1881 av_log(hwfc, AV_LOG_ERROR, "Failed to invalidate memory: %s\n",
1882 vk_ret2str(ret));
1883 err = AVERROR_EXTERNAL;
1884 goto fail;
1885 }
1886 }
1887
1888 for (int i = 0; i < planes; i++) {
1889 VkImageSubresource sub = {
1890 .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
1891 };
1892 VkSubresourceLayout layout;
1893 vkGetImageSubresourceLayout(hwctx->act_dev, f->img[i], &sub, &layout);
1894 dst->linesize[i] = layout.rowPitch;
1895 }
1896
1897 map->frame = f;
1898 map->flags = flags;
1899
1900 err = ff_hwframe_map_create(src->hw_frames_ctx, dst, src,
1901 &vulkan_unmap_frame, map);
1902 if (err < 0)
1903 goto fail;
1904
1905 return 0;
1906
1907 fail:
1908 for (int i = 0; i < mapped_mem_count; i++)
1909 vkUnmapMemory(hwctx->act_dev, f->mem[i]);
1910
1911 av_free(map);
1912 return err;
1913 }
1914
1915 #if CONFIG_LIBDRM
vulkan_unmap_from(AVHWFramesContext * hwfc,HWMapDescriptor * hwmap)1916 static void vulkan_unmap_from(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
1917 {
1918 VulkanMapping *map = hwmap->priv;
1919 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1920 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1921
1922 for (int i = 0; i < planes; i++) {
1923 vkDestroyImage(hwctx->act_dev, map->frame->img[i], hwctx->alloc);
1924 vkFreeMemory(hwctx->act_dev, map->frame->mem[i], hwctx->alloc);
1925 vkDestroySemaphore(hwctx->act_dev, map->frame->sem[i], hwctx->alloc);
1926 }
1927
1928 av_freep(&map->frame);
1929 }
1930
1931 static const struct {
1932 uint32_t drm_fourcc;
1933 VkFormat vk_format;
1934 } vulkan_drm_format_map[] = {
1935 { DRM_FORMAT_R8, VK_FORMAT_R8_UNORM },
1936 { DRM_FORMAT_R16, VK_FORMAT_R16_UNORM },
1937 { DRM_FORMAT_GR88, VK_FORMAT_R8G8_UNORM },
1938 { DRM_FORMAT_RG88, VK_FORMAT_R8G8_UNORM },
1939 { DRM_FORMAT_GR1616, VK_FORMAT_R16G16_UNORM },
1940 { DRM_FORMAT_RG1616, VK_FORMAT_R16G16_UNORM },
1941 { DRM_FORMAT_ARGB8888, VK_FORMAT_B8G8R8A8_UNORM },
1942 { DRM_FORMAT_XRGB8888, VK_FORMAT_B8G8R8A8_UNORM },
1943 { DRM_FORMAT_ABGR8888, VK_FORMAT_R8G8B8A8_UNORM },
1944 { DRM_FORMAT_XBGR8888, VK_FORMAT_R8G8B8A8_UNORM },
1945 };
1946
drm_to_vulkan_fmt(uint32_t drm_fourcc)1947 static inline VkFormat drm_to_vulkan_fmt(uint32_t drm_fourcc)
1948 {
1949 for (int i = 0; i < FF_ARRAY_ELEMS(vulkan_drm_format_map); i++)
1950 if (vulkan_drm_format_map[i].drm_fourcc == drm_fourcc)
1951 return vulkan_drm_format_map[i].vk_format;
1952 return VK_FORMAT_UNDEFINED;
1953 }
1954
vulkan_map_from_drm_frame_desc(AVHWFramesContext * hwfc,AVVkFrame ** frame,AVDRMFrameDescriptor * desc)1955 static int vulkan_map_from_drm_frame_desc(AVHWFramesContext *hwfc, AVVkFrame **frame,
1956 AVDRMFrameDescriptor *desc)
1957 {
1958 int err = 0;
1959 VkResult ret;
1960 AVVkFrame *f;
1961 int bind_counts = 0;
1962 AVHWDeviceContext *ctx = hwfc->device_ctx;
1963 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1964 VulkanDevicePriv *p = ctx->internal->priv;
1965 VulkanFramesPriv *fp = hwfc->internal->priv;
1966 AVVulkanFramesContext *frames_hwctx = hwfc->hwctx;
1967 const AVPixFmtDescriptor *fmt_desc = av_pix_fmt_desc_get(hwfc->sw_format);
1968 const int has_modifiers = p->extensions & EXT_DRM_MODIFIER_FLAGS;
1969 VkSubresourceLayout plane_data[AV_NUM_DATA_POINTERS] = { 0 };
1970 VkBindImageMemoryInfo bind_info[AV_NUM_DATA_POINTERS] = { 0 };
1971 VkBindImagePlaneMemoryInfo plane_info[AV_NUM_DATA_POINTERS] = { 0 };
1972 VkExternalMemoryHandleTypeFlagBits htype = VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT;
1973
1974 VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdPropertiesKHR);
1975
1976 for (int i = 0; i < desc->nb_layers; i++) {
1977 if (drm_to_vulkan_fmt(desc->layers[i].format) == VK_FORMAT_UNDEFINED) {
1978 av_log(ctx, AV_LOG_ERROR, "Unsupported DMABUF layer format %#08x!\n",
1979 desc->layers[i].format);
1980 return AVERROR(EINVAL);
1981 }
1982 }
1983
1984 if (!(f = av_vk_frame_alloc())) {
1985 av_log(ctx, AV_LOG_ERROR, "Unable to allocate memory for AVVkFrame!\n");
1986 err = AVERROR(ENOMEM);
1987 goto fail;
1988 }
1989
1990 f->tiling = has_modifiers ? VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT :
1991 desc->objects[0].format_modifier == DRM_FORMAT_MOD_LINEAR ?
1992 VK_IMAGE_TILING_LINEAR : VK_IMAGE_TILING_OPTIMAL;
1993
1994 for (int i = 0; i < desc->nb_layers; i++) {
1995 const int planes = desc->layers[i].nb_planes;
1996 VkImageDrmFormatModifierExplicitCreateInfoEXT drm_info = {
1997 .sType = VK_STRUCTURE_TYPE_IMAGE_DRM_FORMAT_MODIFIER_EXPLICIT_CREATE_INFO_EXT,
1998 .drmFormatModifier = desc->objects[0].format_modifier,
1999 .drmFormatModifierPlaneCount = planes,
2000 .pPlaneLayouts = (const VkSubresourceLayout *)&plane_data,
2001 };
2002
2003 VkExternalMemoryImageCreateInfo einfo = {
2004 .sType = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO,
2005 .pNext = has_modifiers ? &drm_info : NULL,
2006 .handleTypes = htype,
2007 };
2008
2009 VkSemaphoreCreateInfo sem_spawn = {
2010 .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
2011 };
2012
2013 const int p_w = i > 0 ? AV_CEIL_RSHIFT(hwfc->width, fmt_desc->log2_chroma_w) : hwfc->width;
2014 const int p_h = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, fmt_desc->log2_chroma_h) : hwfc->height;
2015
2016 VkImageCreateInfo image_create_info = {
2017 .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
2018 .pNext = &einfo,
2019 .imageType = VK_IMAGE_TYPE_2D,
2020 .format = drm_to_vulkan_fmt(desc->layers[i].format),
2021 .extent.width = p_w,
2022 .extent.height = p_h,
2023 .extent.depth = 1,
2024 .mipLevels = 1,
2025 .arrayLayers = 1,
2026 .flags = VK_IMAGE_CREATE_ALIAS_BIT,
2027 .tiling = f->tiling,
2028 .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED, /* specs say so */
2029 .usage = frames_hwctx->usage,
2030 .samples = VK_SAMPLE_COUNT_1_BIT,
2031 .pQueueFamilyIndices = p->qfs,
2032 .queueFamilyIndexCount = p->num_qfs,
2033 .sharingMode = p->num_qfs > 1 ? VK_SHARING_MODE_CONCURRENT :
2034 VK_SHARING_MODE_EXCLUSIVE,
2035 };
2036
2037 for (int j = 0; j < planes; j++) {
2038 plane_data[j].offset = desc->layers[i].planes[j].offset;
2039 plane_data[j].rowPitch = desc->layers[i].planes[j].pitch;
2040 plane_data[j].size = 0; /* The specs say so for all 3 */
2041 plane_data[j].arrayPitch = 0;
2042 plane_data[j].depthPitch = 0;
2043 }
2044
2045 /* Create image */
2046 ret = vkCreateImage(hwctx->act_dev, &image_create_info,
2047 hwctx->alloc, &f->img[i]);
2048 if (ret != VK_SUCCESS) {
2049 av_log(ctx, AV_LOG_ERROR, "Image creation failure: %s\n",
2050 vk_ret2str(ret));
2051 err = AVERROR(EINVAL);
2052 goto fail;
2053 }
2054
2055 ret = vkCreateSemaphore(hwctx->act_dev, &sem_spawn,
2056 hwctx->alloc, &f->sem[i]);
2057 if (ret != VK_SUCCESS) {
2058 av_log(hwctx, AV_LOG_ERROR, "Failed to create semaphore: %s\n",
2059 vk_ret2str(ret));
2060 return AVERROR_EXTERNAL;
2061 }
2062
2063 /* We'd import a semaphore onto the one we created using
2064 * vkImportSemaphoreFdKHR but unfortunately neither DRM nor VAAPI
2065 * offer us anything we could import and sync with, so instead
2066 * just signal the semaphore we created. */
2067
2068 f->layout[i] = image_create_info.initialLayout;
2069 f->access[i] = 0x0;
2070 }
2071
2072 for (int i = 0; i < desc->nb_objects; i++) {
2073 int use_ded_mem = 0;
2074 VkMemoryFdPropertiesKHR fdmp = {
2075 .sType = VK_STRUCTURE_TYPE_MEMORY_FD_PROPERTIES_KHR,
2076 };
2077 VkMemoryRequirements req = {
2078 .size = desc->objects[i].size,
2079 };
2080 VkImportMemoryFdInfoKHR idesc = {
2081 .sType = VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR,
2082 .handleType = htype,
2083 .fd = dup(desc->objects[i].fd),
2084 };
2085 VkMemoryDedicatedAllocateInfo ded_alloc = {
2086 .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO,
2087 .pNext = &idesc,
2088 };
2089
2090 ret = pfn_vkGetMemoryFdPropertiesKHR(hwctx->act_dev, htype,
2091 idesc.fd, &fdmp);
2092 if (ret != VK_SUCCESS) {
2093 av_log(hwfc, AV_LOG_ERROR, "Failed to get FD properties: %s\n",
2094 vk_ret2str(ret));
2095 err = AVERROR_EXTERNAL;
2096 close(idesc.fd);
2097 goto fail;
2098 }
2099
2100 req.memoryTypeBits = fdmp.memoryTypeBits;
2101
2102 /* Dedicated allocation only makes sense if there's a one to one mapping
2103 * between images and the memory backing them, so only check in this
2104 * case. */
2105 if (desc->nb_layers == desc->nb_objects) {
2106 VkImageMemoryRequirementsInfo2 req_desc = {
2107 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_REQUIREMENTS_INFO_2,
2108 .image = f->img[i],
2109 };
2110 VkMemoryDedicatedRequirements ded_req = {
2111 .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS,
2112 };
2113 VkMemoryRequirements2 req2 = {
2114 .sType = VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2,
2115 .pNext = &ded_req,
2116 };
2117
2118 vkGetImageMemoryRequirements2(hwctx->act_dev, &req_desc, &req2);
2119
2120 use_ded_mem = ded_req.prefersDedicatedAllocation |
2121 ded_req.requiresDedicatedAllocation;
2122 if (use_ded_mem)
2123 ded_alloc.image = f->img[i];
2124 }
2125
2126 err = alloc_mem(ctx, &req, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
2127 use_ded_mem ? &ded_alloc : ded_alloc.pNext,
2128 &f->flags, &f->mem[i]);
2129 if (err) {
2130 close(idesc.fd);
2131 return err;
2132 }
2133
2134 f->size[i] = desc->objects[i].size;
2135 }
2136
2137 for (int i = 0; i < desc->nb_layers; i++) {
2138 const int planes = desc->layers[i].nb_planes;
2139 const int signal_p = has_modifiers && (planes > 1);
2140 for (int j = 0; j < planes; j++) {
2141 VkImageAspectFlagBits aspect = j == 0 ? VK_IMAGE_ASPECT_MEMORY_PLANE_0_BIT_EXT :
2142 j == 1 ? VK_IMAGE_ASPECT_MEMORY_PLANE_1_BIT_EXT :
2143 VK_IMAGE_ASPECT_MEMORY_PLANE_2_BIT_EXT;
2144
2145 plane_info[bind_counts].sType = VK_STRUCTURE_TYPE_BIND_IMAGE_PLANE_MEMORY_INFO;
2146 plane_info[bind_counts].planeAspect = aspect;
2147
2148 bind_info[bind_counts].sType = VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO;
2149 bind_info[bind_counts].pNext = signal_p ? &plane_info[bind_counts] : NULL;
2150 bind_info[bind_counts].image = f->img[i];
2151 bind_info[bind_counts].memory = f->mem[desc->layers[i].planes[j].object_index];
2152 bind_info[bind_counts].memoryOffset = desc->layers[i].planes[j].offset;
2153 bind_counts++;
2154 }
2155 }
2156
2157 /* Bind the allocated memory to the images */
2158 ret = vkBindImageMemory2(hwctx->act_dev, bind_counts, bind_info);
2159 if (ret != VK_SUCCESS) {
2160 av_log(ctx, AV_LOG_ERROR, "Failed to bind memory: %s\n",
2161 vk_ret2str(ret));
2162 return AVERROR_EXTERNAL;
2163 }
2164
2165 /* NOTE: This is completely uneccesary and unneeded once we can import
2166 * semaphores from DRM. Otherwise we have to activate the semaphores.
2167 * We're reusing the exec context that's also used for uploads/downloads. */
2168 err = prepare_frame(hwfc, &fp->conv_ctx, f, PREP_MODE_RO_SHADER);
2169 if (err)
2170 goto fail;
2171
2172 *frame = f;
2173
2174 return 0;
2175
2176 fail:
2177 for (int i = 0; i < desc->nb_layers; i++) {
2178 vkDestroyImage(hwctx->act_dev, f->img[i], hwctx->alloc);
2179 vkDestroySemaphore(hwctx->act_dev, f->sem[i], hwctx->alloc);
2180 }
2181 for (int i = 0; i < desc->nb_objects; i++)
2182 vkFreeMemory(hwctx->act_dev, f->mem[i], hwctx->alloc);
2183
2184 av_free(f);
2185
2186 return err;
2187 }
2188
vulkan_map_from_drm(AVHWFramesContext * hwfc,AVFrame * dst,const AVFrame * src,int flags)2189 static int vulkan_map_from_drm(AVHWFramesContext *hwfc, AVFrame *dst,
2190 const AVFrame *src, int flags)
2191 {
2192 int err = 0;
2193 AVVkFrame *f;
2194 VulkanMapping *map = NULL;
2195
2196 err = vulkan_map_from_drm_frame_desc(hwfc, &f,
2197 (AVDRMFrameDescriptor *)src->data[0]);
2198 if (err)
2199 return err;
2200
2201 /* The unmapping function will free this */
2202 dst->data[0] = (uint8_t *)f;
2203 dst->width = src->width;
2204 dst->height = src->height;
2205
2206 map = av_mallocz(sizeof(VulkanMapping));
2207 if (!map)
2208 goto fail;
2209
2210 map->frame = f;
2211 map->flags = flags;
2212
2213 err = ff_hwframe_map_create(dst->hw_frames_ctx, dst, src,
2214 &vulkan_unmap_from, map);
2215 if (err < 0)
2216 goto fail;
2217
2218 av_log(hwfc, AV_LOG_DEBUG, "Mapped DRM object to Vulkan!\n");
2219
2220 return 0;
2221
2222 fail:
2223 vulkan_frame_free(hwfc->device_ctx->hwctx, (uint8_t *)f);
2224 av_free(map);
2225 return err;
2226 }
2227
2228 #if CONFIG_VAAPI
vulkan_map_from_vaapi(AVHWFramesContext * dst_fc,AVFrame * dst,const AVFrame * src,int flags)2229 static int vulkan_map_from_vaapi(AVHWFramesContext *dst_fc,
2230 AVFrame *dst, const AVFrame *src,
2231 int flags)
2232 {
2233 int err;
2234 AVFrame *tmp = av_frame_alloc();
2235 AVHWFramesContext *vaapi_fc = (AVHWFramesContext*)src->hw_frames_ctx->data;
2236 AVVAAPIDeviceContext *vaapi_ctx = vaapi_fc->device_ctx->hwctx;
2237 VASurfaceID surface_id = (VASurfaceID)(uintptr_t)src->data[3];
2238
2239 if (!tmp)
2240 return AVERROR(ENOMEM);
2241
2242 /* We have to sync since like the previous comment said, no semaphores */
2243 vaSyncSurface(vaapi_ctx->display, surface_id);
2244
2245 tmp->format = AV_PIX_FMT_DRM_PRIME;
2246
2247 err = av_hwframe_map(tmp, src, flags);
2248 if (err < 0)
2249 goto fail;
2250
2251 err = vulkan_map_from_drm(dst_fc, dst, tmp, flags);
2252 if (err < 0)
2253 goto fail;
2254
2255 err = ff_hwframe_map_replace(dst, src);
2256
2257 fail:
2258 av_frame_free(&tmp);
2259 return err;
2260 }
2261 #endif
2262 #endif
2263
2264 #if CONFIG_CUDA
vulkan_export_to_cuda(AVHWFramesContext * hwfc,AVBufferRef * cuda_hwfc,const AVFrame * frame)2265 static int vulkan_export_to_cuda(AVHWFramesContext *hwfc,
2266 AVBufferRef *cuda_hwfc,
2267 const AVFrame *frame)
2268 {
2269 int err;
2270 VkResult ret;
2271 AVVkFrame *dst_f;
2272 AVVkFrameInternal *dst_int;
2273 AVHWDeviceContext *ctx = hwfc->device_ctx;
2274 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2275 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
2276 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
2277 VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdKHR);
2278 VK_LOAD_PFN(hwctx->inst, vkGetSemaphoreFdKHR);
2279
2280 AVHWFramesContext *cuda_fc = (AVHWFramesContext*)cuda_hwfc->data;
2281 AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
2282 AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
2283 AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
2284 CudaFunctions *cu = cu_internal->cuda_dl;
2285 CUarray_format cufmt = desc->comp[0].depth > 8 ? CU_AD_FORMAT_UNSIGNED_INT16 :
2286 CU_AD_FORMAT_UNSIGNED_INT8;
2287
2288 dst_f = (AVVkFrame *)frame->data[0];
2289
2290 dst_int = dst_f->internal;
2291 if (!dst_int || !dst_int->cuda_fc_ref) {
2292 if (!dst_f->internal)
2293 dst_f->internal = dst_int = av_mallocz(sizeof(*dst_f->internal));
2294
2295 if (!dst_int) {
2296 err = AVERROR(ENOMEM);
2297 goto fail;
2298 }
2299
2300 dst_int->cuda_fc_ref = av_buffer_ref(cuda_hwfc);
2301 if (!dst_int->cuda_fc_ref) {
2302 err = AVERROR(ENOMEM);
2303 goto fail;
2304 }
2305
2306 for (int i = 0; i < planes; i++) {
2307 CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC tex_desc = {
2308 .offset = 0,
2309 .arrayDesc = {
2310 .Width = i > 0 ? AV_CEIL_RSHIFT(hwfc->width, desc->log2_chroma_w)
2311 : hwfc->width,
2312 .Height = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, desc->log2_chroma_h)
2313 : hwfc->height,
2314 .Depth = 0,
2315 .Format = cufmt,
2316 .NumChannels = 1 + ((planes == 2) && i),
2317 .Flags = 0,
2318 },
2319 .numLevels = 1,
2320 };
2321 CUDA_EXTERNAL_MEMORY_HANDLE_DESC ext_desc = {
2322 .type = CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD,
2323 .size = dst_f->size[i],
2324 };
2325 VkMemoryGetFdInfoKHR export_info = {
2326 .sType = VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR,
2327 .memory = dst_f->mem[i],
2328 .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR,
2329 };
2330 VkSemaphoreGetFdInfoKHR sem_export = {
2331 .sType = VK_STRUCTURE_TYPE_SEMAPHORE_GET_FD_INFO_KHR,
2332 .semaphore = dst_f->sem[i],
2333 .handleType = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT,
2334 };
2335 CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC ext_sem_desc = {
2336 .type = CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD,
2337 };
2338
2339 ret = pfn_vkGetMemoryFdKHR(hwctx->act_dev, &export_info,
2340 &ext_desc.handle.fd);
2341 if (ret != VK_SUCCESS) {
2342 av_log(hwfc, AV_LOG_ERROR, "Unable to export the image as a FD!\n");
2343 err = AVERROR_EXTERNAL;
2344 goto fail;
2345 }
2346
2347 ret = CHECK_CU(cu->cuImportExternalMemory(&dst_int->ext_mem[i], &ext_desc));
2348 if (ret < 0) {
2349 err = AVERROR_EXTERNAL;
2350 goto fail;
2351 }
2352
2353 ret = CHECK_CU(cu->cuExternalMemoryGetMappedMipmappedArray(&dst_int->cu_mma[i],
2354 dst_int->ext_mem[i],
2355 &tex_desc));
2356 if (ret < 0) {
2357 err = AVERROR_EXTERNAL;
2358 goto fail;
2359 }
2360
2361 ret = CHECK_CU(cu->cuMipmappedArrayGetLevel(&dst_int->cu_array[i],
2362 dst_int->cu_mma[i], 0));
2363 if (ret < 0) {
2364 err = AVERROR_EXTERNAL;
2365 goto fail;
2366 }
2367
2368 ret = pfn_vkGetSemaphoreFdKHR(hwctx->act_dev, &sem_export,
2369 &ext_sem_desc.handle.fd);
2370 if (ret != VK_SUCCESS) {
2371 av_log(ctx, AV_LOG_ERROR, "Failed to export semaphore: %s\n",
2372 vk_ret2str(ret));
2373 err = AVERROR_EXTERNAL;
2374 goto fail;
2375 }
2376
2377 ret = CHECK_CU(cu->cuImportExternalSemaphore(&dst_int->cu_sem[i],
2378 &ext_sem_desc));
2379 if (ret < 0) {
2380 err = AVERROR_EXTERNAL;
2381 goto fail;
2382 }
2383 }
2384 }
2385
2386 return 0;
2387
2388 fail:
2389 return err;
2390 }
2391
vulkan_transfer_data_from_cuda(AVHWFramesContext * hwfc,AVFrame * dst,const AVFrame * src)2392 static int vulkan_transfer_data_from_cuda(AVHWFramesContext *hwfc,
2393 AVFrame *dst, const AVFrame *src)
2394 {
2395 int err;
2396 VkResult ret;
2397 CUcontext dummy;
2398 AVVkFrame *dst_f;
2399 AVVkFrameInternal *dst_int;
2400 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
2401 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
2402
2403 AVHWFramesContext *cuda_fc = (AVHWFramesContext*)src->hw_frames_ctx->data;
2404 AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
2405 AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
2406 AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
2407 CudaFunctions *cu = cu_internal->cuda_dl;
2408 CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS s_w_par[AV_NUM_DATA_POINTERS] = { 0 };
2409 CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS s_s_par[AV_NUM_DATA_POINTERS] = { 0 };
2410
2411 ret = CHECK_CU(cu->cuCtxPushCurrent(cuda_dev->cuda_ctx));
2412 if (ret < 0) {
2413 err = AVERROR_EXTERNAL;
2414 goto fail;
2415 }
2416
2417 dst_f = (AVVkFrame *)dst->data[0];
2418
2419 ret = vulkan_export_to_cuda(hwfc, src->hw_frames_ctx, dst);
2420 if (ret < 0) {
2421 goto fail;
2422 }
2423 dst_int = dst_f->internal;
2424
2425 ret = CHECK_CU(cu->cuWaitExternalSemaphoresAsync(dst_int->cu_sem, s_w_par,
2426 planes, cuda_dev->stream));
2427 if (ret < 0) {
2428 err = AVERROR_EXTERNAL;
2429 goto fail;
2430 }
2431
2432 for (int i = 0; i < planes; i++) {
2433 CUDA_MEMCPY2D cpy = {
2434 .srcMemoryType = CU_MEMORYTYPE_DEVICE,
2435 .srcDevice = (CUdeviceptr)src->data[i],
2436 .srcPitch = src->linesize[i],
2437 .srcY = 0,
2438
2439 .dstMemoryType = CU_MEMORYTYPE_ARRAY,
2440 .dstArray = dst_int->cu_array[i],
2441 .WidthInBytes = (i > 0 ? AV_CEIL_RSHIFT(hwfc->width, desc->log2_chroma_w)
2442 : hwfc->width) * desc->comp[i].step,
2443 .Height = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, desc->log2_chroma_h)
2444 : hwfc->height,
2445 };
2446
2447 ret = CHECK_CU(cu->cuMemcpy2DAsync(&cpy, cuda_dev->stream));
2448 if (ret < 0) {
2449 err = AVERROR_EXTERNAL;
2450 goto fail;
2451 }
2452 }
2453
2454 ret = CHECK_CU(cu->cuSignalExternalSemaphoresAsync(dst_int->cu_sem, s_s_par,
2455 planes, cuda_dev->stream));
2456 if (ret < 0) {
2457 err = AVERROR_EXTERNAL;
2458 goto fail;
2459 }
2460
2461 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2462
2463 av_log(hwfc, AV_LOG_VERBOSE, "Transfered CUDA image to Vulkan!\n");
2464
2465 return 0;
2466
2467 fail:
2468 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2469 vulkan_free_internal(dst_int);
2470 dst_f->internal = NULL;
2471 av_buffer_unref(&dst->buf[0]);
2472 return err;
2473 }
2474 #endif
2475
vulkan_map_to(AVHWFramesContext * hwfc,AVFrame * dst,const AVFrame * src,int flags)2476 static int vulkan_map_to(AVHWFramesContext *hwfc, AVFrame *dst,
2477 const AVFrame *src, int flags)
2478 {
2479 av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2480
2481 switch (src->format) {
2482 #if CONFIG_LIBDRM
2483 #if CONFIG_VAAPI
2484 case AV_PIX_FMT_VAAPI:
2485 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2486 return vulkan_map_from_vaapi(hwfc, dst, src, flags);
2487 #endif
2488 case AV_PIX_FMT_DRM_PRIME:
2489 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2490 return vulkan_map_from_drm(hwfc, dst, src, flags);
2491 #endif
2492 default:
2493 return AVERROR(ENOSYS);
2494 }
2495 }
2496
2497 #if CONFIG_LIBDRM
2498 typedef struct VulkanDRMMapping {
2499 AVDRMFrameDescriptor drm_desc;
2500 AVVkFrame *source;
2501 } VulkanDRMMapping;
2502
vulkan_unmap_to_drm(AVHWFramesContext * hwfc,HWMapDescriptor * hwmap)2503 static void vulkan_unmap_to_drm(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
2504 {
2505 AVDRMFrameDescriptor *drm_desc = hwmap->priv;
2506
2507 for (int i = 0; i < drm_desc->nb_objects; i++)
2508 close(drm_desc->objects[i].fd);
2509
2510 av_free(drm_desc);
2511 }
2512
vulkan_fmt_to_drm(VkFormat vkfmt)2513 static inline uint32_t vulkan_fmt_to_drm(VkFormat vkfmt)
2514 {
2515 for (int i = 0; i < FF_ARRAY_ELEMS(vulkan_drm_format_map); i++)
2516 if (vulkan_drm_format_map[i].vk_format == vkfmt)
2517 return vulkan_drm_format_map[i].drm_fourcc;
2518 return DRM_FORMAT_INVALID;
2519 }
2520
vulkan_map_to_drm(AVHWFramesContext * hwfc,AVFrame * dst,const AVFrame * src,int flags)2521 static int vulkan_map_to_drm(AVHWFramesContext *hwfc, AVFrame *dst,
2522 const AVFrame *src, int flags)
2523 {
2524 int err = 0;
2525 VkResult ret;
2526 AVVkFrame *f = (AVVkFrame *)src->data[0];
2527 VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2528 VulkanFramesPriv *fp = hwfc->internal->priv;
2529 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
2530 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
2531 VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdKHR);
2532 VkImageDrmFormatModifierPropertiesEXT drm_mod = {
2533 .sType = VK_STRUCTURE_TYPE_IMAGE_DRM_FORMAT_MODIFIER_PROPERTIES_EXT,
2534 };
2535
2536 AVDRMFrameDescriptor *drm_desc = av_mallocz(sizeof(*drm_desc));
2537 if (!drm_desc)
2538 return AVERROR(ENOMEM);
2539
2540 err = prepare_frame(hwfc, &fp->conv_ctx, f, PREP_MODE_EXTERNAL_EXPORT);
2541 if (err < 0)
2542 goto end;
2543
2544 err = ff_hwframe_map_create(src->hw_frames_ctx, dst, src, &vulkan_unmap_to_drm, drm_desc);
2545 if (err < 0)
2546 goto end;
2547
2548 if (p->extensions & EXT_DRM_MODIFIER_FLAGS) {
2549 VK_LOAD_PFN(hwctx->inst, vkGetImageDrmFormatModifierPropertiesEXT);
2550 ret = pfn_vkGetImageDrmFormatModifierPropertiesEXT(hwctx->act_dev, f->img[0],
2551 &drm_mod);
2552 if (ret != VK_SUCCESS) {
2553 av_log(hwfc, AV_LOG_ERROR, "Failed to retrieve DRM format modifier!\n");
2554 err = AVERROR_EXTERNAL;
2555 goto end;
2556 }
2557 }
2558
2559 for (int i = 0; (i < planes) && (f->mem[i]); i++) {
2560 VkMemoryGetFdInfoKHR export_info = {
2561 .sType = VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR,
2562 .memory = f->mem[i],
2563 .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT,
2564 };
2565
2566 ret = pfn_vkGetMemoryFdKHR(hwctx->act_dev, &export_info,
2567 &drm_desc->objects[i].fd);
2568 if (ret != VK_SUCCESS) {
2569 av_log(hwfc, AV_LOG_ERROR, "Unable to export the image as a FD!\n");
2570 err = AVERROR_EXTERNAL;
2571 goto end;
2572 }
2573
2574 drm_desc->nb_objects++;
2575 drm_desc->objects[i].size = f->size[i];
2576 drm_desc->objects[i].format_modifier = drm_mod.drmFormatModifier;
2577 }
2578
2579 drm_desc->nb_layers = planes;
2580 for (int i = 0; i < drm_desc->nb_layers; i++) {
2581 VkSubresourceLayout layout;
2582 VkImageSubresource sub = {
2583 .aspectMask = p->extensions & EXT_DRM_MODIFIER_FLAGS ?
2584 VK_IMAGE_ASPECT_MEMORY_PLANE_0_BIT_EXT :
2585 VK_IMAGE_ASPECT_COLOR_BIT,
2586 };
2587 VkFormat plane_vkfmt = av_vkfmt_from_pixfmt(hwfc->sw_format)[i];
2588
2589 drm_desc->layers[i].format = vulkan_fmt_to_drm(plane_vkfmt);
2590 drm_desc->layers[i].nb_planes = 1;
2591
2592 if (drm_desc->layers[i].format == DRM_FORMAT_INVALID) {
2593 av_log(hwfc, AV_LOG_ERROR, "Cannot map to DRM layer, unsupported!\n");
2594 err = AVERROR_PATCHWELCOME;
2595 goto end;
2596 }
2597
2598 drm_desc->layers[i].planes[0].object_index = FFMIN(i, drm_desc->nb_objects - 1);
2599
2600 if (f->tiling == VK_IMAGE_TILING_OPTIMAL)
2601 continue;
2602
2603 vkGetImageSubresourceLayout(hwctx->act_dev, f->img[i], &sub, &layout);
2604 drm_desc->layers[i].planes[0].offset = layout.offset;
2605 drm_desc->layers[i].planes[0].pitch = layout.rowPitch;
2606 }
2607
2608 dst->width = src->width;
2609 dst->height = src->height;
2610 dst->data[0] = (uint8_t *)drm_desc;
2611
2612 av_log(hwfc, AV_LOG_VERBOSE, "Mapped AVVkFrame to a DRM object!\n");
2613
2614 return 0;
2615
2616 end:
2617 av_free(drm_desc);
2618 return err;
2619 }
2620
2621 #if CONFIG_VAAPI
vulkan_map_to_vaapi(AVHWFramesContext * hwfc,AVFrame * dst,const AVFrame * src,int flags)2622 static int vulkan_map_to_vaapi(AVHWFramesContext *hwfc, AVFrame *dst,
2623 const AVFrame *src, int flags)
2624 {
2625 int err;
2626 AVFrame *tmp = av_frame_alloc();
2627 if (!tmp)
2628 return AVERROR(ENOMEM);
2629
2630 tmp->format = AV_PIX_FMT_DRM_PRIME;
2631
2632 err = vulkan_map_to_drm(hwfc, tmp, src, flags);
2633 if (err < 0)
2634 goto fail;
2635
2636 err = av_hwframe_map(dst, tmp, flags);
2637 if (err < 0)
2638 goto fail;
2639
2640 err = ff_hwframe_map_replace(dst, src);
2641
2642 fail:
2643 av_frame_free(&tmp);
2644 return err;
2645 }
2646 #endif
2647 #endif
2648
vulkan_map_from(AVHWFramesContext * hwfc,AVFrame * dst,const AVFrame * src,int flags)2649 static int vulkan_map_from(AVHWFramesContext *hwfc, AVFrame *dst,
2650 const AVFrame *src, int flags)
2651 {
2652 av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2653
2654 switch (dst->format) {
2655 #if CONFIG_LIBDRM
2656 case AV_PIX_FMT_DRM_PRIME:
2657 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2658 return vulkan_map_to_drm(hwfc, dst, src, flags);
2659 #if CONFIG_VAAPI
2660 case AV_PIX_FMT_VAAPI:
2661 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2662 return vulkan_map_to_vaapi(hwfc, dst, src, flags);
2663 #endif
2664 #endif
2665 default:
2666 return vulkan_map_frame_to_mem(hwfc, dst, src, flags);
2667 }
2668 }
2669
2670 typedef struct ImageBuffer {
2671 VkBuffer buf;
2672 VkDeviceMemory mem;
2673 VkMemoryPropertyFlagBits flags;
2674 int mapped_mem;
2675 } ImageBuffer;
2676
free_buf(void * opaque,uint8_t * data)2677 static void free_buf(void *opaque, uint8_t *data)
2678 {
2679 AVHWDeviceContext *ctx = opaque;
2680 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2681 ImageBuffer *vkbuf = (ImageBuffer *)data;
2682
2683 if (vkbuf->buf)
2684 vkDestroyBuffer(hwctx->act_dev, vkbuf->buf, hwctx->alloc);
2685 if (vkbuf->mem)
2686 vkFreeMemory(hwctx->act_dev, vkbuf->mem, hwctx->alloc);
2687
2688 av_free(data);
2689 }
2690
create_buf(AVHWDeviceContext * ctx,AVBufferRef ** buf,size_t imp_size,int height,int * stride,VkBufferUsageFlags usage,VkMemoryPropertyFlagBits flags,void * create_pnext,void * alloc_pnext)2691 static int create_buf(AVHWDeviceContext *ctx, AVBufferRef **buf, size_t imp_size,
2692 int height, int *stride, VkBufferUsageFlags usage,
2693 VkMemoryPropertyFlagBits flags, void *create_pnext,
2694 void *alloc_pnext)
2695 {
2696 int err;
2697 VkResult ret;
2698 int use_ded_mem;
2699 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2700 VulkanDevicePriv *p = ctx->internal->priv;
2701
2702 VkBufferCreateInfo buf_spawn = {
2703 .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
2704 .pNext = create_pnext,
2705 .usage = usage,
2706 .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
2707 };
2708
2709 VkBufferMemoryRequirementsInfo2 req_desc = {
2710 .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_REQUIREMENTS_INFO_2,
2711 };
2712 VkMemoryDedicatedAllocateInfo ded_alloc = {
2713 .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO,
2714 .pNext = alloc_pnext,
2715 };
2716 VkMemoryDedicatedRequirements ded_req = {
2717 .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS,
2718 };
2719 VkMemoryRequirements2 req = {
2720 .sType = VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2,
2721 .pNext = &ded_req,
2722 };
2723
2724 ImageBuffer *vkbuf = av_mallocz(sizeof(*vkbuf));
2725 if (!vkbuf)
2726 return AVERROR(ENOMEM);
2727
2728 vkbuf->mapped_mem = !!imp_size;
2729
2730 if (!vkbuf->mapped_mem) {
2731 *stride = FFALIGN(*stride, p->props.properties.limits.optimalBufferCopyRowPitchAlignment);
2732 buf_spawn.size = height*(*stride);
2733 buf_spawn.size = FFALIGN(buf_spawn.size, p->props.properties.limits.minMemoryMapAlignment);
2734 } else {
2735 buf_spawn.size = imp_size;
2736 }
2737
2738 ret = vkCreateBuffer(hwctx->act_dev, &buf_spawn, NULL, &vkbuf->buf);
2739 if (ret != VK_SUCCESS) {
2740 av_log(ctx, AV_LOG_ERROR, "Failed to create buffer: %s\n",
2741 vk_ret2str(ret));
2742 return AVERROR_EXTERNAL;
2743 }
2744
2745 req_desc.buffer = vkbuf->buf;
2746
2747 vkGetBufferMemoryRequirements2(hwctx->act_dev, &req_desc, &req);
2748
2749 /* In case the implementation prefers/requires dedicated allocation */
2750 use_ded_mem = ded_req.prefersDedicatedAllocation |
2751 ded_req.requiresDedicatedAllocation;
2752 if (use_ded_mem)
2753 ded_alloc.buffer = vkbuf->buf;
2754
2755 err = alloc_mem(ctx, &req.memoryRequirements, flags,
2756 use_ded_mem ? &ded_alloc : (void *)ded_alloc.pNext,
2757 &vkbuf->flags, &vkbuf->mem);
2758 if (err)
2759 return err;
2760
2761 ret = vkBindBufferMemory(hwctx->act_dev, vkbuf->buf, vkbuf->mem, 0);
2762 if (ret != VK_SUCCESS) {
2763 av_log(ctx, AV_LOG_ERROR, "Failed to bind memory to buffer: %s\n",
2764 vk_ret2str(ret));
2765 free_buf(ctx, (uint8_t *)vkbuf);
2766 return AVERROR_EXTERNAL;
2767 }
2768
2769 *buf = av_buffer_create((uint8_t *)vkbuf, sizeof(*vkbuf), free_buf, ctx, 0);
2770 if (!(*buf)) {
2771 free_buf(ctx, (uint8_t *)vkbuf);
2772 return AVERROR(ENOMEM);
2773 }
2774
2775 return 0;
2776 }
2777
2778 /* Skips mapping of host mapped buffers but still invalidates them */
map_buffers(AVHWDeviceContext * ctx,AVBufferRef ** bufs,uint8_t * mem[],int nb_buffers,int invalidate)2779 static int map_buffers(AVHWDeviceContext *ctx, AVBufferRef **bufs, uint8_t *mem[],
2780 int nb_buffers, int invalidate)
2781 {
2782 VkResult ret;
2783 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2784 VkMappedMemoryRange invalidate_ctx[AV_NUM_DATA_POINTERS];
2785 int invalidate_count = 0;
2786
2787 for (int i = 0; i < nb_buffers; i++) {
2788 ImageBuffer *vkbuf = (ImageBuffer *)bufs[i]->data;
2789 if (vkbuf->mapped_mem)
2790 continue;
2791
2792 ret = vkMapMemory(hwctx->act_dev, vkbuf->mem, 0,
2793 VK_WHOLE_SIZE, 0, (void **)&mem[i]);
2794 if (ret != VK_SUCCESS) {
2795 av_log(ctx, AV_LOG_ERROR, "Failed to map buffer memory: %s\n",
2796 vk_ret2str(ret));
2797 return AVERROR_EXTERNAL;
2798 }
2799 }
2800
2801 if (!invalidate)
2802 return 0;
2803
2804 for (int i = 0; i < nb_buffers; i++) {
2805 ImageBuffer *vkbuf = (ImageBuffer *)bufs[i]->data;
2806 const VkMappedMemoryRange ival_buf = {
2807 .sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE,
2808 .memory = vkbuf->mem,
2809 .size = VK_WHOLE_SIZE,
2810 };
2811 if (vkbuf->flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)
2812 continue;
2813 invalidate_ctx[invalidate_count++] = ival_buf;
2814 }
2815
2816 if (invalidate_count) {
2817 ret = vkInvalidateMappedMemoryRanges(hwctx->act_dev, invalidate_count,
2818 invalidate_ctx);
2819 if (ret != VK_SUCCESS)
2820 av_log(ctx, AV_LOG_WARNING, "Failed to invalidate memory: %s\n",
2821 vk_ret2str(ret));
2822 }
2823
2824 return 0;
2825 }
2826
unmap_buffers(AVHWDeviceContext * ctx,AVBufferRef ** bufs,int nb_buffers,int flush)2827 static int unmap_buffers(AVHWDeviceContext *ctx, AVBufferRef **bufs,
2828 int nb_buffers, int flush)
2829 {
2830 int err = 0;
2831 VkResult ret;
2832 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2833 VkMappedMemoryRange flush_ctx[AV_NUM_DATA_POINTERS];
2834 int flush_count = 0;
2835
2836 if (flush) {
2837 for (int i = 0; i < nb_buffers; i++) {
2838 ImageBuffer *vkbuf = (ImageBuffer *)bufs[i]->data;
2839 const VkMappedMemoryRange flush_buf = {
2840 .sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE,
2841 .memory = vkbuf->mem,
2842 .size = VK_WHOLE_SIZE,
2843 };
2844 if (vkbuf->flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)
2845 continue;
2846 flush_ctx[flush_count++] = flush_buf;
2847 }
2848 }
2849
2850 if (flush_count) {
2851 ret = vkFlushMappedMemoryRanges(hwctx->act_dev, flush_count, flush_ctx);
2852 if (ret != VK_SUCCESS) {
2853 av_log(ctx, AV_LOG_ERROR, "Failed to flush memory: %s\n",
2854 vk_ret2str(ret));
2855 err = AVERROR_EXTERNAL; /* We still want to try to unmap them */
2856 }
2857 }
2858
2859 for (int i = 0; i < nb_buffers; i++) {
2860 ImageBuffer *vkbuf = (ImageBuffer *)bufs[i]->data;
2861 if (vkbuf->mapped_mem)
2862 continue;
2863
2864 vkUnmapMemory(hwctx->act_dev, vkbuf->mem);
2865 }
2866
2867 return err;
2868 }
2869
transfer_image_buf(AVHWFramesContext * hwfc,const AVFrame * f,AVBufferRef ** bufs,const int * buf_stride,int w,int h,enum AVPixelFormat pix_fmt,int to_buf)2870 static int transfer_image_buf(AVHWFramesContext *hwfc, const AVFrame *f,
2871 AVBufferRef **bufs, const int *buf_stride, int w,
2872 int h, enum AVPixelFormat pix_fmt, int to_buf)
2873 {
2874 int err;
2875 AVVkFrame *frame = (AVVkFrame *)f->data[0];
2876 VulkanFramesPriv *fp = hwfc->internal->priv;
2877
2878 int bar_num = 0;
2879 VkPipelineStageFlagBits sem_wait_dst[AV_NUM_DATA_POINTERS];
2880
2881 const int planes = av_pix_fmt_count_planes(pix_fmt);
2882 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(pix_fmt);
2883
2884 VkImageMemoryBarrier img_bar[AV_NUM_DATA_POINTERS] = { 0 };
2885 VulkanExecCtx *ectx = to_buf ? &fp->download_ctx : &fp->upload_ctx;
2886 VkCommandBuffer cmd_buf = get_buf_exec_ctx(hwfc, ectx);
2887
2888 VkSubmitInfo s_info = {
2889 .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
2890 .pSignalSemaphores = frame->sem,
2891 .pWaitSemaphores = frame->sem,
2892 .pWaitDstStageMask = sem_wait_dst,
2893 .signalSemaphoreCount = planes,
2894 .waitSemaphoreCount = planes,
2895 };
2896
2897 if ((err = wait_start_exec_ctx(hwfc, ectx)))
2898 return err;
2899
2900 /* Change the image layout to something more optimal for transfers */
2901 for (int i = 0; i < planes; i++) {
2902 VkImageLayout new_layout = to_buf ? VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL :
2903 VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
2904 VkAccessFlags new_access = to_buf ? VK_ACCESS_TRANSFER_READ_BIT :
2905 VK_ACCESS_TRANSFER_WRITE_BIT;
2906
2907 sem_wait_dst[i] = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
2908
2909 /* If the layout matches and we have read access skip the barrier */
2910 if ((frame->layout[i] == new_layout) && (frame->access[i] & new_access))
2911 continue;
2912
2913 img_bar[bar_num].sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
2914 img_bar[bar_num].srcAccessMask = 0x0;
2915 img_bar[bar_num].dstAccessMask = new_access;
2916 img_bar[bar_num].oldLayout = frame->layout[i];
2917 img_bar[bar_num].newLayout = new_layout;
2918 img_bar[bar_num].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
2919 img_bar[bar_num].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
2920 img_bar[bar_num].image = frame->img[i];
2921 img_bar[bar_num].subresourceRange.levelCount = 1;
2922 img_bar[bar_num].subresourceRange.layerCount = 1;
2923 img_bar[bar_num].subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
2924
2925 frame->layout[i] = img_bar[bar_num].newLayout;
2926 frame->access[i] = img_bar[bar_num].dstAccessMask;
2927
2928 bar_num++;
2929 }
2930
2931 if (bar_num)
2932 vkCmdPipelineBarrier(cmd_buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
2933 VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
2934 0, NULL, 0, NULL, bar_num, img_bar);
2935
2936 /* Schedule a copy for each plane */
2937 for (int i = 0; i < planes; i++) {
2938 ImageBuffer *vkbuf = (ImageBuffer *)bufs[i]->data;
2939 const int p_w = i > 0 ? AV_CEIL_RSHIFT(w, desc->log2_chroma_w) : w;
2940 const int p_h = i > 0 ? AV_CEIL_RSHIFT(h, desc->log2_chroma_h) : h;
2941 VkBufferImageCopy buf_reg = {
2942 .bufferOffset = 0,
2943 /* Buffer stride isn't in bytes, it's in samples, the implementation
2944 * uses the image's VkFormat to know how many bytes per sample
2945 * the buffer has. So we have to convert by dividing. Stupid.
2946 * Won't work with YUVA or other planar formats with alpha. */
2947 .bufferRowLength = buf_stride[i] / desc->comp[i].step,
2948 .bufferImageHeight = p_h,
2949 .imageSubresource.layerCount = 1,
2950 .imageSubresource.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
2951 .imageOffset = { 0, 0, 0, },
2952 .imageExtent = { p_w, p_h, 1, },
2953 };
2954
2955 if (to_buf)
2956 vkCmdCopyImageToBuffer(cmd_buf, frame->img[i], frame->layout[i],
2957 vkbuf->buf, 1, &buf_reg);
2958 else
2959 vkCmdCopyBufferToImage(cmd_buf, vkbuf->buf, frame->img[i],
2960 frame->layout[i], 1, &buf_reg);
2961 }
2962
2963 /* When uploading, do this asynchronously if the source is refcounted by
2964 * keeping the buffers as a submission dependency.
2965 * The hwcontext is guaranteed to not be freed until all frames are freed
2966 * in the frames_unint function.
2967 * When downloading to buffer, do this synchronously and wait for the
2968 * queue submission to finish executing */
2969 if (!to_buf) {
2970 int ref;
2971 for (ref = 0; ref < AV_NUM_DATA_POINTERS; ref++) {
2972 if (!f->buf[ref])
2973 break;
2974 if ((err = add_buf_dep_exec_ctx(hwfc, ectx, &f->buf[ref], 1)))
2975 return err;
2976 }
2977 if (ref && (err = add_buf_dep_exec_ctx(hwfc, ectx, bufs, planes)))
2978 return err;
2979 return submit_exec_ctx(hwfc, ectx, &s_info, !ref);
2980 } else {
2981 return submit_exec_ctx(hwfc, ectx, &s_info, 1);
2982 }
2983 }
2984
vulkan_transfer_data_from_mem(AVHWFramesContext * hwfc,AVFrame * dst,const AVFrame * src)2985 static int vulkan_transfer_data_from_mem(AVHWFramesContext *hwfc, AVFrame *dst,
2986 const AVFrame *src)
2987 {
2988 int err = 0;
2989 AVFrame tmp;
2990 AVVkFrame *f = (AVVkFrame *)dst->data[0];
2991 AVHWDeviceContext *dev_ctx = hwfc->device_ctx;
2992 AVBufferRef *bufs[AV_NUM_DATA_POINTERS] = { 0 };
2993 const int planes = av_pix_fmt_count_planes(src->format);
2994 int log2_chroma = av_pix_fmt_desc_get(src->format)->log2_chroma_h;
2995 VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2996 int host_mapped[AV_NUM_DATA_POINTERS] = { 0 };
2997 int map_host = p->extensions & EXT_EXTERNAL_HOST_MEMORY;
2998
2999 if ((src->format != AV_PIX_FMT_NONE && !av_vkfmt_from_pixfmt(src->format))) {
3000 av_log(hwfc, AV_LOG_ERROR, "Unsupported source pixel format!\n");
3001 return AVERROR(EINVAL);
3002 }
3003
3004 if (src->width > hwfc->width || src->height > hwfc->height)
3005 return AVERROR(EINVAL);
3006
3007 /* For linear, host visiable images */
3008 if (f->tiling == VK_IMAGE_TILING_LINEAR &&
3009 f->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) {
3010 AVFrame *map = av_frame_alloc();
3011 if (!map)
3012 return AVERROR(ENOMEM);
3013 map->format = src->format;
3014
3015 err = vulkan_map_frame_to_mem(hwfc, map, dst, AV_HWFRAME_MAP_WRITE);
3016 if (err)
3017 return err;
3018
3019 err = av_frame_copy(map, src);
3020 av_frame_free(&map);
3021 return err;
3022 }
3023
3024 /* Create buffers */
3025 for (int i = 0; i < planes; i++) {
3026 int h = src->height;
3027 int p_height = i > 0 ? AV_CEIL_RSHIFT(h, log2_chroma) : h;
3028 size_t p_size = FFABS(src->linesize[i]) * p_height;
3029
3030 VkImportMemoryHostPointerInfoEXT import_desc = {
3031 .sType = VK_STRUCTURE_TYPE_IMPORT_MEMORY_HOST_POINTER_INFO_EXT,
3032 .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT,
3033 .pHostPointer = src->data[i],
3034 };
3035
3036 /* We can only map images with positive stride and alignment appropriate
3037 * for the device. */
3038 host_mapped[i] = map_host && src->linesize[i] > 0 &&
3039 !(p_size % p->hprops.minImportedHostPointerAlignment) &&
3040 !(((uintptr_t)import_desc.pHostPointer) %
3041 p->hprops.minImportedHostPointerAlignment);
3042 p_size = host_mapped[i] ? p_size : 0;
3043
3044 tmp.linesize[i] = FFABS(src->linesize[i]);
3045 err = create_buf(dev_ctx, &bufs[i], p_size, p_height, &tmp.linesize[i],
3046 VK_BUFFER_USAGE_TRANSFER_SRC_BIT,
3047 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT, NULL,
3048 host_mapped[i] ? &import_desc : NULL);
3049 if (err)
3050 goto end;
3051 }
3052
3053 /* Map, copy image to buffer, unmap */
3054 if ((err = map_buffers(dev_ctx, bufs, tmp.data, planes, 0)))
3055 goto end;
3056
3057 for (int i = 0; i < planes; i++) {
3058 int h = src->height;
3059 int p_height = i > 0 ? AV_CEIL_RSHIFT(h, log2_chroma) : h;
3060
3061 if (host_mapped[i])
3062 continue;
3063
3064 av_image_copy_plane(tmp.data[i], tmp.linesize[i],
3065 (const uint8_t *)src->data[i], src->linesize[i],
3066 FFMIN(tmp.linesize[i], FFABS(src->linesize[i])),
3067 p_height);
3068 }
3069
3070 if ((err = unmap_buffers(dev_ctx, bufs, planes, 1)))
3071 goto end;
3072
3073 /* Copy buffers to image */
3074 err = transfer_image_buf(hwfc, dst, bufs, tmp.linesize,
3075 src->width, src->height, src->format, 0);
3076
3077 end:
3078 for (int i = 0; i < planes; i++)
3079 av_buffer_unref(&bufs[i]);
3080
3081 return err;
3082 }
3083
vulkan_transfer_data_to(AVHWFramesContext * hwfc,AVFrame * dst,const AVFrame * src)3084 static int vulkan_transfer_data_to(AVHWFramesContext *hwfc, AVFrame *dst,
3085 const AVFrame *src)
3086 {
3087 av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
3088
3089 switch (src->format) {
3090 #if CONFIG_CUDA
3091 case AV_PIX_FMT_CUDA:
3092 if ((p->extensions & EXT_EXTERNAL_FD_MEMORY) &&
3093 (p->extensions & EXT_EXTERNAL_FD_SEM))
3094 return vulkan_transfer_data_from_cuda(hwfc, dst, src);
3095 #endif
3096 default:
3097 if (src->hw_frames_ctx)
3098 return AVERROR(ENOSYS);
3099 else
3100 return vulkan_transfer_data_from_mem(hwfc, dst, src);
3101 }
3102 }
3103
3104 #if CONFIG_CUDA
vulkan_transfer_data_to_cuda(AVHWFramesContext * hwfc,AVFrame * dst,const AVFrame * src)3105 static int vulkan_transfer_data_to_cuda(AVHWFramesContext *hwfc, AVFrame *dst,
3106 const AVFrame *src)
3107 {
3108 int err;
3109 VkResult ret;
3110 CUcontext dummy;
3111 AVVkFrame *dst_f;
3112 AVVkFrameInternal *dst_int;
3113 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
3114 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
3115
3116 AVHWFramesContext *cuda_fc = (AVHWFramesContext*)dst->hw_frames_ctx->data;
3117 AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
3118 AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
3119 AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
3120 CudaFunctions *cu = cu_internal->cuda_dl;
3121
3122 ret = CHECK_CU(cu->cuCtxPushCurrent(cuda_dev->cuda_ctx));
3123 if (ret < 0) {
3124 err = AVERROR_EXTERNAL;
3125 goto fail;
3126 }
3127
3128 dst_f = (AVVkFrame *)src->data[0];
3129
3130 err = vulkan_export_to_cuda(hwfc, dst->hw_frames_ctx, src);
3131 if (err < 0) {
3132 goto fail;
3133 }
3134
3135 dst_int = dst_f->internal;
3136
3137 for (int i = 0; i < planes; i++) {
3138 CUDA_MEMCPY2D cpy = {
3139 .dstMemoryType = CU_MEMORYTYPE_DEVICE,
3140 .dstDevice = (CUdeviceptr)dst->data[i],
3141 .dstPitch = dst->linesize[i],
3142 .dstY = 0,
3143
3144 .srcMemoryType = CU_MEMORYTYPE_ARRAY,
3145 .srcArray = dst_int->cu_array[i],
3146 .WidthInBytes = (i > 0 ? AV_CEIL_RSHIFT(hwfc->width, desc->log2_chroma_w)
3147 : hwfc->width) * desc->comp[i].step,
3148 .Height = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, desc->log2_chroma_h)
3149 : hwfc->height,
3150 };
3151
3152 ret = CHECK_CU(cu->cuMemcpy2DAsync(&cpy, cuda_dev->stream));
3153 if (ret < 0) {
3154 err = AVERROR_EXTERNAL;
3155 goto fail;
3156 }
3157 }
3158
3159 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
3160
3161 av_log(hwfc, AV_LOG_VERBOSE, "Transfered Vulkan image to CUDA!\n");
3162
3163 return 0;
3164
3165 fail:
3166 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
3167 vulkan_free_internal(dst_int);
3168 dst_f->internal = NULL;
3169 av_buffer_unref(&dst->buf[0]);
3170 return err;
3171 }
3172 #endif
3173
vulkan_transfer_data_to_mem(AVHWFramesContext * hwfc,AVFrame * dst,const AVFrame * src)3174 static int vulkan_transfer_data_to_mem(AVHWFramesContext *hwfc, AVFrame *dst,
3175 const AVFrame *src)
3176 {
3177 int err = 0;
3178 AVFrame tmp;
3179 AVVkFrame *f = (AVVkFrame *)src->data[0];
3180 AVHWDeviceContext *dev_ctx = hwfc->device_ctx;
3181 AVBufferRef *bufs[AV_NUM_DATA_POINTERS] = { 0 };
3182 const int planes = av_pix_fmt_count_planes(dst->format);
3183 int log2_chroma = av_pix_fmt_desc_get(dst->format)->log2_chroma_h;
3184 VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
3185 int host_mapped[AV_NUM_DATA_POINTERS] = { 0 };
3186 int map_host = p->extensions & EXT_EXTERNAL_HOST_MEMORY;
3187
3188 if (dst->width > hwfc->width || dst->height > hwfc->height)
3189 return AVERROR(EINVAL);
3190
3191 /* For linear, host visiable images */
3192 if (f->tiling == VK_IMAGE_TILING_LINEAR &&
3193 f->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) {
3194 AVFrame *map = av_frame_alloc();
3195 if (!map)
3196 return AVERROR(ENOMEM);
3197 map->format = dst->format;
3198
3199 err = vulkan_map_frame_to_mem(hwfc, map, src, AV_HWFRAME_MAP_READ);
3200 if (err)
3201 return err;
3202
3203 err = av_frame_copy(dst, map);
3204 av_frame_free(&map);
3205 return err;
3206 }
3207
3208 /* Create buffers */
3209 for (int i = 0; i < planes; i++) {
3210 int h = dst->height;
3211 int p_height = i > 0 ? AV_CEIL_RSHIFT(h, log2_chroma) : h;
3212 size_t p_size = FFABS(dst->linesize[i]) * p_height;
3213
3214 VkImportMemoryHostPointerInfoEXT import_desc = {
3215 .sType = VK_STRUCTURE_TYPE_IMPORT_MEMORY_HOST_POINTER_INFO_EXT,
3216 .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT,
3217 .pHostPointer = dst->data[i],
3218 };
3219
3220 /* We can only map images with positive stride and alignment appropriate
3221 * for the device. */
3222 host_mapped[i] = map_host && dst->linesize[i] > 0 &&
3223 !(p_size % p->hprops.minImportedHostPointerAlignment) &&
3224 !(((uintptr_t)import_desc.pHostPointer) %
3225 p->hprops.minImportedHostPointerAlignment);
3226 p_size = host_mapped[i] ? p_size : 0;
3227
3228 tmp.linesize[i] = FFABS(dst->linesize[i]);
3229 err = create_buf(dev_ctx, &bufs[i], p_size, p_height,
3230 &tmp.linesize[i], VK_BUFFER_USAGE_TRANSFER_DST_BIT,
3231 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT, NULL,
3232 host_mapped[i] ? &import_desc : NULL);
3233 if (err)
3234 goto end;
3235 }
3236
3237 /* Copy image to buffer */
3238 if ((err = transfer_image_buf(hwfc, src, bufs, tmp.linesize,
3239 dst->width, dst->height, dst->format, 1)))
3240 goto end;
3241
3242 /* Map, copy buffer to frame, unmap */
3243 if ((err = map_buffers(dev_ctx, bufs, tmp.data, planes, 1)))
3244 goto end;
3245
3246 for (int i = 0; i < planes; i++) {
3247 int h = dst->height;
3248 int p_height = i > 0 ? AV_CEIL_RSHIFT(h, log2_chroma) : h;
3249
3250 if (host_mapped[i])
3251 continue;
3252
3253 av_image_copy_plane(dst->data[i], dst->linesize[i],
3254 (const uint8_t *)tmp.data[i], tmp.linesize[i],
3255 FFMIN(tmp.linesize[i], FFABS(dst->linesize[i])),
3256 p_height);
3257 }
3258
3259 err = unmap_buffers(dev_ctx, bufs, planes, 0);
3260
3261 end:
3262 for (int i = 0; i < planes; i++)
3263 av_buffer_unref(&bufs[i]);
3264
3265 return err;
3266 }
3267
vulkan_transfer_data_from(AVHWFramesContext * hwfc,AVFrame * dst,const AVFrame * src)3268 static int vulkan_transfer_data_from(AVHWFramesContext *hwfc, AVFrame *dst,
3269 const AVFrame *src)
3270 {
3271 av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
3272
3273 switch (dst->format) {
3274 #if CONFIG_CUDA
3275 case AV_PIX_FMT_CUDA:
3276 if ((p->extensions & EXT_EXTERNAL_FD_MEMORY) &&
3277 (p->extensions & EXT_EXTERNAL_FD_SEM))
3278 return vulkan_transfer_data_to_cuda(hwfc, dst, src);
3279 #endif
3280 default:
3281 if (dst->hw_frames_ctx)
3282 return AVERROR(ENOSYS);
3283 else
3284 return vulkan_transfer_data_to_mem(hwfc, dst, src);
3285 }
3286 }
3287
vulkan_frames_derive_to(AVHWFramesContext * dst_fc,AVHWFramesContext * src_fc,int flags)3288 static int vulkan_frames_derive_to(AVHWFramesContext *dst_fc,
3289 AVHWFramesContext *src_fc, int flags)
3290 {
3291 return vulkan_frames_init(dst_fc);
3292 }
3293
av_vk_frame_alloc(void)3294 AVVkFrame *av_vk_frame_alloc(void)
3295 {
3296 return av_mallocz(sizeof(AVVkFrame));
3297 }
3298
3299 const HWContextType ff_hwcontext_type_vulkan = {
3300 .type = AV_HWDEVICE_TYPE_VULKAN,
3301 .name = "Vulkan",
3302
3303 .device_hwctx_size = sizeof(AVVulkanDeviceContext),
3304 .device_priv_size = sizeof(VulkanDevicePriv),
3305 .frames_hwctx_size = sizeof(AVVulkanFramesContext),
3306 .frames_priv_size = sizeof(VulkanFramesPriv),
3307
3308 .device_init = &vulkan_device_init,
3309 .device_create = &vulkan_device_create,
3310 .device_derive = &vulkan_device_derive,
3311
3312 .frames_get_constraints = &vulkan_frames_get_constraints,
3313 .frames_init = vulkan_frames_init,
3314 .frames_get_buffer = vulkan_get_buffer,
3315 .frames_uninit = vulkan_frames_uninit,
3316
3317 .transfer_get_formats = vulkan_transfer_get_formats,
3318 .transfer_data_to = vulkan_transfer_data_to,
3319 .transfer_data_from = vulkan_transfer_data_from,
3320
3321 .map_to = vulkan_map_to,
3322 .map_from = vulkan_map_from,
3323 .frames_derive_to = &vulkan_frames_derive_to,
3324
3325 .pix_fmts = (const enum AVPixelFormat []) {
3326 AV_PIX_FMT_VULKAN,
3327 AV_PIX_FMT_NONE
3328 },
3329 };
3330