2 * This file is part of FFmpeg.
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.
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.
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
23 #include "hwcontext.h"
24 #include "hwcontext_internal.h"
25 #include "hwcontext_vulkan.h"
30 #include <drm_fourcc.h>
31 #include "hwcontext_drm.h"
33 #include <va/va_drmcommon.h>
34 #include "hwcontext_vaapi.h"
39 #include "hwcontext_cuda_internal.h"
40 #include "cuda_check.h"
41 #define CHECK_CU(x) FF_CUDA_CHECK_DL(cuda_cu, cu, x)
44 typedef struct VulkanExecCtx {
51 typedef struct VulkanDevicePriv {
53 VkPhysicalDeviceProperties props;
54 VkPhysicalDeviceMemoryProperties mprops;
61 VkDebugUtilsMessengerEXT debug_ctx;
70 int use_linear_images;
76 typedef struct VulkanFramesPriv {
80 typedef struct AVVkFrameInternal {
82 /* Importing external memory into cuda is really expensive so we keep the
83 * memory imported all the time */
84 AVBufferRef *cuda_fc_ref; /* Need to keep it around for uninit */
85 CUexternalMemory ext_mem[AV_NUM_DATA_POINTERS];
86 CUmipmappedArray cu_mma[AV_NUM_DATA_POINTERS];
87 CUarray cu_array[AV_NUM_DATA_POINTERS];
88 CUexternalSemaphore cu_sem[AV_NUM_DATA_POINTERS];
92 #define VK_LOAD_PFN(inst, name) PFN_##name pfn_##name = (PFN_##name) \
93 vkGetInstanceProcAddr(inst, #name)
95 #define DEFAULT_USAGE_FLAGS (VK_IMAGE_USAGE_SAMPLED_BIT | \
96 VK_IMAGE_USAGE_STORAGE_BIT | \
97 VK_IMAGE_USAGE_TRANSFER_SRC_BIT | \
98 VK_IMAGE_USAGE_TRANSFER_DST_BIT)
100 #define ADD_VAL_TO_LIST(list, count, val) \
102 list = av_realloc_array(list, sizeof(*list), ++count); \
104 err = AVERROR(ENOMEM); \
107 list[count - 1] = av_strdup(val); \
108 if (!list[count - 1]) { \
109 err = AVERROR(ENOMEM); \
114 static const struct {
115 enum AVPixelFormat pixfmt;
116 const VkFormat vkfmts[3];
117 } vk_pixfmt_map[] = {
118 { AV_PIX_FMT_GRAY8, { VK_FORMAT_R8_UNORM } },
119 { AV_PIX_FMT_GRAY16, { VK_FORMAT_R16_UNORM } },
120 { AV_PIX_FMT_GRAYF32, { VK_FORMAT_R32_SFLOAT } },
122 { AV_PIX_FMT_NV12, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8G8_UNORM } },
123 { AV_PIX_FMT_P010, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16G16_UNORM } },
124 { AV_PIX_FMT_P016, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16G16_UNORM } },
126 { AV_PIX_FMT_YUV420P, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } },
127 { AV_PIX_FMT_YUV422P, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } },
128 { AV_PIX_FMT_YUV444P, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } },
130 { AV_PIX_FMT_YUV420P16, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
131 { AV_PIX_FMT_YUV422P16, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
132 { AV_PIX_FMT_YUV444P16, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
134 { AV_PIX_FMT_ABGR, { VK_FORMAT_A8B8G8R8_UNORM_PACK32 } },
135 { AV_PIX_FMT_BGRA, { VK_FORMAT_B8G8R8A8_UNORM } },
136 { AV_PIX_FMT_RGBA, { VK_FORMAT_R8G8B8A8_UNORM } },
137 { AV_PIX_FMT_RGB24, { VK_FORMAT_R8G8B8_UNORM } },
138 { AV_PIX_FMT_BGR24, { VK_FORMAT_B8G8R8_UNORM } },
139 { AV_PIX_FMT_RGB48, { VK_FORMAT_R16G16B16_UNORM } },
140 { AV_PIX_FMT_RGBA64, { VK_FORMAT_R16G16B16A16_UNORM } },
141 { AV_PIX_FMT_RGB565, { VK_FORMAT_R5G6B5_UNORM_PACK16 } },
142 { AV_PIX_FMT_BGR565, { VK_FORMAT_B5G6R5_UNORM_PACK16 } },
143 { AV_PIX_FMT_BGR0, { VK_FORMAT_B8G8R8A8_UNORM } },
144 { AV_PIX_FMT_0BGR, { VK_FORMAT_A8B8G8R8_UNORM_PACK32 } },
145 { AV_PIX_FMT_RGB0, { VK_FORMAT_R8G8B8A8_UNORM } },
147 { AV_PIX_FMT_GBRPF32, { VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_SFLOAT } },
150 const VkFormat *av_vkfmt_from_pixfmt(enum AVPixelFormat p)
152 for (enum AVPixelFormat i = 0; i < FF_ARRAY_ELEMS(vk_pixfmt_map); i++)
153 if (vk_pixfmt_map[i].pixfmt == p)
154 return vk_pixfmt_map[i].vkfmts;
158 static int pixfmt_is_supported(AVVulkanDeviceContext *hwctx, enum AVPixelFormat p,
161 const VkFormat *fmt = av_vkfmt_from_pixfmt(p);
162 int planes = av_pix_fmt_count_planes(p);
167 for (int i = 0; i < planes; i++) {
168 VkFormatFeatureFlags flags;
169 VkFormatProperties2 prop = {
170 .sType = VK_STRUCTURE_TYPE_FORMAT_PROPERTIES_2,
172 vkGetPhysicalDeviceFormatProperties2(hwctx->phys_dev, fmt[i], &prop);
173 flags = linear ? prop.formatProperties.linearTilingFeatures :
174 prop.formatProperties.optimalTilingFeatures;
175 if (!(flags & DEFAULT_USAGE_FLAGS))
182 enum VulkanExtensions {
183 EXT_EXTERNAL_DMABUF_MEMORY = 1ULL << 0, /* VK_EXT_external_memory_dma_buf */
184 EXT_DRM_MODIFIER_FLAGS = 1ULL << 1, /* VK_EXT_image_drm_format_modifier */
185 EXT_EXTERNAL_FD_MEMORY = 1ULL << 2, /* VK_KHR_external_memory_fd */
186 EXT_EXTERNAL_FD_SEM = 1ULL << 3, /* VK_KHR_external_semaphore_fd */
188 EXT_NO_FLAG = 1ULL << 63,
191 typedef struct VulkanOptExtension {
194 } VulkanOptExtension;
196 static const VulkanOptExtension optional_instance_exts[] = {
197 { VK_KHR_SURFACE_EXTENSION_NAME, EXT_NO_FLAG },
200 static const VulkanOptExtension optional_device_exts[] = {
201 { VK_KHR_EXTERNAL_MEMORY_FD_EXTENSION_NAME, EXT_EXTERNAL_FD_MEMORY, },
202 { VK_EXT_EXTERNAL_MEMORY_DMA_BUF_EXTENSION_NAME, EXT_EXTERNAL_DMABUF_MEMORY, },
203 { VK_EXT_IMAGE_DRM_FORMAT_MODIFIER_EXTENSION_NAME, EXT_DRM_MODIFIER_FLAGS, },
204 { VK_KHR_EXTERNAL_SEMAPHORE_FD_EXTENSION_NAME, EXT_EXTERNAL_FD_SEM, },
207 /* Converts return values to strings */
208 static const char *vk_ret2str(VkResult res)
210 #define CASE(VAL) case VAL: return #VAL
216 CASE(VK_EVENT_RESET);
218 CASE(VK_ERROR_OUT_OF_HOST_MEMORY);
219 CASE(VK_ERROR_OUT_OF_DEVICE_MEMORY);
220 CASE(VK_ERROR_INITIALIZATION_FAILED);
221 CASE(VK_ERROR_DEVICE_LOST);
222 CASE(VK_ERROR_MEMORY_MAP_FAILED);
223 CASE(VK_ERROR_LAYER_NOT_PRESENT);
224 CASE(VK_ERROR_EXTENSION_NOT_PRESENT);
225 CASE(VK_ERROR_FEATURE_NOT_PRESENT);
226 CASE(VK_ERROR_INCOMPATIBLE_DRIVER);
227 CASE(VK_ERROR_TOO_MANY_OBJECTS);
228 CASE(VK_ERROR_FORMAT_NOT_SUPPORTED);
229 CASE(VK_ERROR_FRAGMENTED_POOL);
230 CASE(VK_ERROR_SURFACE_LOST_KHR);
231 CASE(VK_ERROR_NATIVE_WINDOW_IN_USE_KHR);
232 CASE(VK_SUBOPTIMAL_KHR);
233 CASE(VK_ERROR_OUT_OF_DATE_KHR);
234 CASE(VK_ERROR_INCOMPATIBLE_DISPLAY_KHR);
235 CASE(VK_ERROR_VALIDATION_FAILED_EXT);
236 CASE(VK_ERROR_INVALID_SHADER_NV);
237 CASE(VK_ERROR_OUT_OF_POOL_MEMORY);
238 CASE(VK_ERROR_INVALID_EXTERNAL_HANDLE);
239 CASE(VK_ERROR_NOT_PERMITTED_EXT);
240 CASE(VK_ERROR_INVALID_DRM_FORMAT_MODIFIER_PLANE_LAYOUT_EXT);
241 CASE(VK_ERROR_INVALID_DEVICE_ADDRESS_EXT);
242 CASE(VK_ERROR_FULL_SCREEN_EXCLUSIVE_MODE_LOST_EXT);
243 default: return "Unknown error";
248 static VkBool32 vk_dbg_callback(VkDebugUtilsMessageSeverityFlagBitsEXT severity,
249 VkDebugUtilsMessageTypeFlagsEXT messageType,
250 const VkDebugUtilsMessengerCallbackDataEXT *data,
254 AVHWDeviceContext *ctx = priv;
257 case VK_DEBUG_UTILS_MESSAGE_SEVERITY_VERBOSE_BIT_EXT: l = AV_LOG_VERBOSE; break;
258 case VK_DEBUG_UTILS_MESSAGE_SEVERITY_INFO_BIT_EXT: l = AV_LOG_INFO; break;
259 case VK_DEBUG_UTILS_MESSAGE_SEVERITY_WARNING_BIT_EXT: l = AV_LOG_WARNING; break;
260 case VK_DEBUG_UTILS_MESSAGE_SEVERITY_ERROR_BIT_EXT: l = AV_LOG_ERROR; break;
261 default: l = AV_LOG_DEBUG; break;
264 av_log(ctx, l, "%s\n", data->pMessage);
265 for (int i = 0; i < data->cmdBufLabelCount; i++)
266 av_log(ctx, l, "\t%i: %s\n", i, data->pCmdBufLabels[i].pLabelName);
271 static int check_extensions(AVHWDeviceContext *ctx, int dev, AVDictionary *opts,
272 const char * const **dst, uint32_t *num, int debug)
275 const char **extension_names = NULL;
276 VulkanDevicePriv *p = ctx->internal->priv;
277 AVVulkanDeviceContext *hwctx = ctx->hwctx;
278 int err = 0, found, extensions_found = 0;
281 int optional_exts_num;
282 uint32_t sup_ext_count;
283 char *user_exts_str = NULL;
284 AVDictionaryEntry *user_exts;
285 VkExtensionProperties *sup_ext;
286 const VulkanOptExtension *optional_exts;
290 optional_exts = optional_instance_exts;
291 optional_exts_num = FF_ARRAY_ELEMS(optional_instance_exts);
292 user_exts = av_dict_get(opts, "instance_extensions", NULL, 0);
294 user_exts_str = av_strdup(user_exts->value);
295 if (!user_exts_str) {
296 err = AVERROR(ENOMEM);
300 vkEnumerateInstanceExtensionProperties(NULL, &sup_ext_count, NULL);
301 sup_ext = av_malloc_array(sup_ext_count, sizeof(VkExtensionProperties));
303 return AVERROR(ENOMEM);
304 vkEnumerateInstanceExtensionProperties(NULL, &sup_ext_count, sup_ext);
307 optional_exts = optional_device_exts;
308 optional_exts_num = FF_ARRAY_ELEMS(optional_device_exts);
309 user_exts = av_dict_get(opts, "device_extensions", NULL, 0);
311 user_exts_str = av_strdup(user_exts->value);
312 if (!user_exts_str) {
313 err = AVERROR(ENOMEM);
317 vkEnumerateDeviceExtensionProperties(hwctx->phys_dev, NULL,
318 &sup_ext_count, NULL);
319 sup_ext = av_malloc_array(sup_ext_count, sizeof(VkExtensionProperties));
321 return AVERROR(ENOMEM);
322 vkEnumerateDeviceExtensionProperties(hwctx->phys_dev, NULL,
323 &sup_ext_count, sup_ext);
326 for (int i = 0; i < optional_exts_num; i++) {
327 tstr = optional_exts[i].name;
329 for (int j = 0; j < sup_ext_count; j++) {
330 if (!strcmp(tstr, sup_ext[j].extensionName)) {
338 av_log(ctx, AV_LOG_VERBOSE, "Using %s extension \"%s\"\n", mod, tstr);
339 p->extensions |= optional_exts[i].flag;
340 ADD_VAL_TO_LIST(extension_names, extensions_found, tstr);
344 tstr = VK_EXT_DEBUG_UTILS_EXTENSION_NAME;
346 for (int j = 0; j < sup_ext_count; j++) {
347 if (!strcmp(tstr, sup_ext[j].extensionName)) {
353 av_log(ctx, AV_LOG_VERBOSE, "Using %s extension \"%s\"\n", mod, tstr);
354 ADD_VAL_TO_LIST(extension_names, extensions_found, tstr);
356 av_log(ctx, AV_LOG_ERROR, "Debug extension \"%s\" not found!\n",
358 err = AVERROR(EINVAL);
364 char *save, *token = av_strtok(user_exts_str, "+", &save);
367 for (int j = 0; j < sup_ext_count; j++) {
368 if (!strcmp(token, sup_ext[j].extensionName)) {
374 av_log(ctx, AV_LOG_VERBOSE, "Using %s extension \"%s\"\n", mod, tstr);
375 ADD_VAL_TO_LIST(extension_names, extensions_found, token);
377 av_log(ctx, AV_LOG_ERROR, "%s extension \"%s\" not found!\n",
379 err = AVERROR(EINVAL);
382 token = av_strtok(NULL, "+", &save);
386 *dst = extension_names;
387 *num = extensions_found;
389 av_free(user_exts_str);
395 for (int i = 0; i < extensions_found; i++)
396 av_free((void *)extension_names[i]);
397 av_free(extension_names);
398 av_free(user_exts_str);
403 /* Creates a VkInstance */
404 static int create_instance(AVHWDeviceContext *ctx, AVDictionary *opts)
408 VulkanDevicePriv *p = ctx->internal->priv;
409 AVVulkanDeviceContext *hwctx = ctx->hwctx;
410 AVDictionaryEntry *debug_opt = av_dict_get(opts, "debug", NULL, 0);
411 const int debug_mode = debug_opt && strtol(debug_opt->value, NULL, 10);
412 VkApplicationInfo application_info = {
413 .sType = VK_STRUCTURE_TYPE_APPLICATION_INFO,
414 .pEngineName = "libavutil",
415 .apiVersion = VK_API_VERSION_1_1,
416 .engineVersion = VK_MAKE_VERSION(LIBAVUTIL_VERSION_MAJOR,
417 LIBAVUTIL_VERSION_MINOR,
418 LIBAVUTIL_VERSION_MICRO),
420 VkInstanceCreateInfo inst_props = {
421 .sType = VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO,
422 .pApplicationInfo = &application_info,
425 /* Check for present/missing extensions */
426 err = check_extensions(ctx, 0, opts, &inst_props.ppEnabledExtensionNames,
427 &inst_props.enabledExtensionCount, debug_mode);
432 static const char *layers[] = { "VK_LAYER_KHRONOS_validation" };
433 inst_props.ppEnabledLayerNames = layers;
434 inst_props.enabledLayerCount = FF_ARRAY_ELEMS(layers);
437 /* Try to create the instance */
438 ret = vkCreateInstance(&inst_props, hwctx->alloc, &hwctx->inst);
440 /* Check for errors */
441 if (ret != VK_SUCCESS) {
442 av_log(ctx, AV_LOG_ERROR, "Instance creation failure: %s\n",
444 for (int i = 0; i < inst_props.enabledExtensionCount; i++)
445 av_free((void *)inst_props.ppEnabledExtensionNames[i]);
446 av_free((void *)inst_props.ppEnabledExtensionNames);
447 return AVERROR_EXTERNAL;
451 VkDebugUtilsMessengerCreateInfoEXT dbg = {
452 .sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_MESSENGER_CREATE_INFO_EXT,
453 .messageSeverity = VK_DEBUG_UTILS_MESSAGE_SEVERITY_VERBOSE_BIT_EXT |
454 VK_DEBUG_UTILS_MESSAGE_SEVERITY_INFO_BIT_EXT |
455 VK_DEBUG_UTILS_MESSAGE_SEVERITY_WARNING_BIT_EXT |
456 VK_DEBUG_UTILS_MESSAGE_SEVERITY_ERROR_BIT_EXT,
457 .messageType = VK_DEBUG_UTILS_MESSAGE_TYPE_GENERAL_BIT_EXT |
458 VK_DEBUG_UTILS_MESSAGE_TYPE_VALIDATION_BIT_EXT |
459 VK_DEBUG_UTILS_MESSAGE_TYPE_PERFORMANCE_BIT_EXT,
460 .pfnUserCallback = vk_dbg_callback,
463 VK_LOAD_PFN(hwctx->inst, vkCreateDebugUtilsMessengerEXT);
465 pfn_vkCreateDebugUtilsMessengerEXT(hwctx->inst, &dbg,
466 hwctx->alloc, &p->debug_ctx);
469 hwctx->enabled_inst_extensions = inst_props.ppEnabledExtensionNames;
470 hwctx->nb_enabled_inst_extensions = inst_props.enabledExtensionCount;
475 typedef struct VulkanDeviceSelection {
476 uint8_t uuid[VK_UUID_SIZE]; /* Will use this first unless !has_uuid */
478 const char *name; /* Will use this second unless NULL */
479 uint32_t pci_device; /* Will use this third unless 0x0 */
480 uint32_t vendor_id; /* Last resort to find something deterministic */
481 int index; /* Finally fall back to index */
482 } VulkanDeviceSelection;
484 static const char *vk_dev_type(enum VkPhysicalDeviceType type)
487 case VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU: return "integrated";
488 case VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU: return "discrete";
489 case VK_PHYSICAL_DEVICE_TYPE_VIRTUAL_GPU: return "virtual";
490 case VK_PHYSICAL_DEVICE_TYPE_CPU: return "software";
491 default: return "unknown";
496 static int find_device(AVHWDeviceContext *ctx, VulkanDeviceSelection *select)
498 int err = 0, choice = -1;
501 VkPhysicalDevice *devices = NULL;
502 VkPhysicalDeviceIDProperties *idp = NULL;
503 VkPhysicalDeviceProperties2 *prop = NULL;
504 VulkanDevicePriv *p = ctx->internal->priv;
505 AVVulkanDeviceContext *hwctx = ctx->hwctx;
507 ret = vkEnumeratePhysicalDevices(hwctx->inst, &num, NULL);
508 if (ret != VK_SUCCESS || !num) {
509 av_log(ctx, AV_LOG_ERROR, "No devices found: %s!\n", vk_ret2str(ret));
510 return AVERROR(ENODEV);
513 devices = av_malloc_array(num, sizeof(VkPhysicalDevice));
515 return AVERROR(ENOMEM);
517 ret = vkEnumeratePhysicalDevices(hwctx->inst, &num, devices);
518 if (ret != VK_SUCCESS) {
519 av_log(ctx, AV_LOG_ERROR, "Failed enumerating devices: %s\n",
521 err = AVERROR(ENODEV);
525 prop = av_mallocz_array(num, sizeof(*prop));
527 err = AVERROR(ENOMEM);
531 idp = av_mallocz_array(num, sizeof(*idp));
533 err = AVERROR(ENOMEM);
537 av_log(ctx, AV_LOG_VERBOSE, "GPU listing:\n");
538 for (int i = 0; i < num; i++) {
539 idp[i].sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES;
540 prop[i].sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2;
541 prop[i].pNext = &idp[i];
543 vkGetPhysicalDeviceProperties2(devices[i], &prop[i]);
544 av_log(ctx, AV_LOG_VERBOSE, " %d: %s (%s) (0x%x)\n", i,
545 prop[i].properties.deviceName,
546 vk_dev_type(prop[i].properties.deviceType),
547 prop[i].properties.deviceID);
550 if (select->has_uuid) {
551 for (int i = 0; i < num; i++) {
552 if (!strncmp(idp[i].deviceUUID, select->uuid, VK_UUID_SIZE)) {
557 av_log(ctx, AV_LOG_ERROR, "Unable to find device by given UUID!\n");
558 err = AVERROR(ENODEV);
560 } else if (select->name) {
561 av_log(ctx, AV_LOG_VERBOSE, "Requested device: %s\n", select->name);
562 for (int i = 0; i < num; i++) {
563 if (strstr(prop[i].properties.deviceName, select->name)) {
568 av_log(ctx, AV_LOG_ERROR, "Unable to find device \"%s\"!\n",
570 err = AVERROR(ENODEV);
572 } else if (select->pci_device) {
573 av_log(ctx, AV_LOG_VERBOSE, "Requested device: 0x%x\n", select->pci_device);
574 for (int i = 0; i < num; i++) {
575 if (select->pci_device == prop[i].properties.deviceID) {
580 av_log(ctx, AV_LOG_ERROR, "Unable to find device with PCI ID 0x%x!\n",
582 err = AVERROR(EINVAL);
584 } else if (select->vendor_id) {
585 av_log(ctx, AV_LOG_VERBOSE, "Requested vendor: 0x%x\n", select->vendor_id);
586 for (int i = 0; i < num; i++) {
587 if (select->vendor_id == prop[i].properties.vendorID) {
592 av_log(ctx, AV_LOG_ERROR, "Unable to find device with Vendor ID 0x%x!\n",
594 err = AVERROR(ENODEV);
597 if (select->index < num) {
598 choice = select->index;
601 av_log(ctx, AV_LOG_ERROR, "Unable to find device with index %i!\n",
603 err = AVERROR(ENODEV);
609 p->dev_is_nvidia = (prop[choice].properties.vendorID == 0x10de);
610 hwctx->phys_dev = devices[choice];
619 static int search_queue_families(AVHWDeviceContext *ctx, VkDeviceCreateInfo *cd)
623 VkQueueFamilyProperties *qs = NULL;
624 AVVulkanDeviceContext *hwctx = ctx->hwctx;
625 int graph_index = -1, comp_index = -1, tx_index = -1;
626 VkDeviceQueueCreateInfo *pc = (VkDeviceQueueCreateInfo *)cd->pQueueCreateInfos;
628 /* First get the number of queue families */
629 vkGetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &num, NULL);
631 av_log(ctx, AV_LOG_ERROR, "Failed to get queues!\n");
632 return AVERROR_EXTERNAL;
635 /* Then allocate memory */
636 qs = av_malloc_array(num, sizeof(VkQueueFamilyProperties));
638 return AVERROR(ENOMEM);
640 /* Finally retrieve the queue families */
641 vkGetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &num, qs);
643 #define SEARCH_FLAGS(expr, out) \
644 for (int i = 0; i < num; i++) { \
645 const VkQueueFlagBits flags = qs[i].queueFlags; \
652 SEARCH_FLAGS(flags & VK_QUEUE_GRAPHICS_BIT, graph_index)
654 SEARCH_FLAGS((flags & VK_QUEUE_COMPUTE_BIT) && (i != graph_index),
657 SEARCH_FLAGS((flags & VK_QUEUE_TRANSFER_BIT) && (i != graph_index) &&
658 (i != comp_index), tx_index)
661 #define ADD_QUEUE(fidx, graph, comp, tx) \
662 av_log(ctx, AV_LOG_VERBOSE, "Using queue family %i (total queues: %i) for %s%s%s\n", \
663 fidx, qs[fidx].queueCount, graph ? "graphics " : "", \
664 comp ? "compute " : "", tx ? "transfers " : ""); \
665 av_log(ctx, AV_LOG_VERBOSE, " QF %i flags: %s%s%s%s\n", fidx, \
666 ((qs[fidx].queueFlags) & VK_QUEUE_GRAPHICS_BIT) ? "(graphics) " : "", \
667 ((qs[fidx].queueFlags) & VK_QUEUE_COMPUTE_BIT) ? "(compute) " : "", \
668 ((qs[fidx].queueFlags) & VK_QUEUE_TRANSFER_BIT) ? "(transfers) " : "", \
669 ((qs[fidx].queueFlags) & VK_QUEUE_SPARSE_BINDING_BIT) ? "(sparse) " : ""); \
670 pc[cd->queueCreateInfoCount].queueFamilyIndex = fidx; \
671 pc[cd->queueCreateInfoCount].queueCount = qs[fidx].queueCount; \
672 weights = av_malloc(qs[fidx].queueCount * sizeof(float)); \
673 pc[cd->queueCreateInfoCount].pQueuePriorities = weights; \
676 for (int i = 0; i < qs[fidx].queueCount; i++) \
678 cd->queueCreateInfoCount++;
680 ADD_QUEUE(graph_index, 1, comp_index < 0, tx_index < 0 && comp_index < 0)
681 hwctx->queue_family_index = graph_index;
682 hwctx->queue_family_comp_index = graph_index;
683 hwctx->queue_family_tx_index = graph_index;
685 if (comp_index != -1) {
686 ADD_QUEUE(comp_index, 0, 1, tx_index < 0)
687 hwctx->queue_family_tx_index = comp_index;
688 hwctx->queue_family_comp_index = comp_index;
691 if (tx_index != -1) {
692 ADD_QUEUE(tx_index, 0, 0, 1)
693 hwctx->queue_family_tx_index = tx_index;
702 av_freep(&pc[0].pQueuePriorities);
703 av_freep(&pc[1].pQueuePriorities);
704 av_freep(&pc[2].pQueuePriorities);
707 return AVERROR(ENOMEM);
710 static int create_exec_ctx(AVHWDeviceContext *ctx, VulkanExecCtx *cmd,
711 int queue_family_index)
714 AVVulkanDeviceContext *hwctx = ctx->hwctx;
716 VkCommandPoolCreateInfo cqueue_create = {
717 .sType = VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO,
718 .flags = VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT,
719 .queueFamilyIndex = queue_family_index,
721 VkCommandBufferAllocateInfo cbuf_create = {
722 .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO,
723 .level = VK_COMMAND_BUFFER_LEVEL_PRIMARY,
724 .commandBufferCount = 1,
727 VkFenceCreateInfo fence_spawn = {
728 .sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO,
731 ret = vkCreateFence(hwctx->act_dev, &fence_spawn,
732 hwctx->alloc, &cmd->fence);
733 if (ret != VK_SUCCESS) {
734 av_log(ctx, AV_LOG_ERROR, "Failed to create frame fence: %s\n",
736 return AVERROR_EXTERNAL;
739 ret = vkCreateCommandPool(hwctx->act_dev, &cqueue_create,
740 hwctx->alloc, &cmd->pool);
741 if (ret != VK_SUCCESS) {
742 av_log(ctx, AV_LOG_ERROR, "Command pool creation failure: %s\n",
744 return AVERROR_EXTERNAL;
747 cbuf_create.commandPool = cmd->pool;
749 ret = vkAllocateCommandBuffers(hwctx->act_dev, &cbuf_create, &cmd->buf);
750 if (ret != VK_SUCCESS) {
751 av_log(ctx, AV_LOG_ERROR, "Command buffer alloc failure: %s\n",
753 return AVERROR_EXTERNAL;
756 vkGetDeviceQueue(hwctx->act_dev, cqueue_create.queueFamilyIndex, 0,
762 static void free_exec_ctx(AVHWDeviceContext *ctx, VulkanExecCtx *cmd)
764 AVVulkanDeviceContext *hwctx = ctx->hwctx;
767 vkDestroyFence(hwctx->act_dev, cmd->fence, hwctx->alloc);
769 vkFreeCommandBuffers(hwctx->act_dev, cmd->pool, 1, &cmd->buf);
771 vkDestroyCommandPool(hwctx->act_dev, cmd->pool, hwctx->alloc);
774 static void vulkan_device_free(AVHWDeviceContext *ctx)
776 VulkanDevicePriv *p = ctx->internal->priv;
777 AVVulkanDeviceContext *hwctx = ctx->hwctx;
779 free_exec_ctx(ctx, &p->cmd);
781 vkDestroyDevice(hwctx->act_dev, hwctx->alloc);
784 VK_LOAD_PFN(hwctx->inst, vkDestroyDebugUtilsMessengerEXT);
785 pfn_vkDestroyDebugUtilsMessengerEXT(hwctx->inst, p->debug_ctx,
789 vkDestroyInstance(hwctx->inst, hwctx->alloc);
791 for (int i = 0; i < hwctx->nb_enabled_inst_extensions; i++)
792 av_free((void *)hwctx->enabled_inst_extensions[i]);
793 av_free((void *)hwctx->enabled_inst_extensions);
795 for (int i = 0; i < hwctx->nb_enabled_dev_extensions; i++)
796 av_free((void *)hwctx->enabled_dev_extensions[i]);
797 av_free((void *)hwctx->enabled_dev_extensions);
800 static int vulkan_device_create_internal(AVHWDeviceContext *ctx,
801 VulkanDeviceSelection *dev_select,
802 AVDictionary *opts, int flags)
806 AVDictionaryEntry *opt_d;
807 VulkanDevicePriv *p = ctx->internal->priv;
808 AVVulkanDeviceContext *hwctx = ctx->hwctx;
809 VkDeviceQueueCreateInfo queue_create_info[3] = {
810 { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO, },
811 { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO, },
812 { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO, },
815 VkDeviceCreateInfo dev_info = {
816 .sType = VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO,
817 .pQueueCreateInfos = queue_create_info,
818 .queueCreateInfoCount = 0,
821 ctx->free = vulkan_device_free;
823 /* Create an instance if not given one */
824 if ((err = create_instance(ctx, opts)))
827 /* Find a device (if not given one) */
828 if ((err = find_device(ctx, dev_select)))
831 vkGetPhysicalDeviceProperties(hwctx->phys_dev, &p->props);
832 av_log(ctx, AV_LOG_VERBOSE, "Using device: %s\n", p->props.deviceName);
833 av_log(ctx, AV_LOG_VERBOSE, "Alignments:\n");
834 av_log(ctx, AV_LOG_VERBOSE, " optimalBufferCopyOffsetAlignment: %li\n",
835 p->props.limits.optimalBufferCopyOffsetAlignment);
836 av_log(ctx, AV_LOG_VERBOSE, " optimalBufferCopyRowPitchAlignment: %li\n",
837 p->props.limits.optimalBufferCopyRowPitchAlignment);
838 av_log(ctx, AV_LOG_VERBOSE, " minMemoryMapAlignment: %li\n",
839 p->props.limits.minMemoryMapAlignment);
841 /* Search queue family */
842 if ((err = search_queue_families(ctx, &dev_info)))
845 if ((err = check_extensions(ctx, 1, opts, &dev_info.ppEnabledExtensionNames,
846 &dev_info.enabledExtensionCount, 0))) {
847 av_free((void *)queue_create_info[0].pQueuePriorities);
848 av_free((void *)queue_create_info[1].pQueuePriorities);
849 av_free((void *)queue_create_info[2].pQueuePriorities);
853 ret = vkCreateDevice(hwctx->phys_dev, &dev_info, hwctx->alloc,
856 av_free((void *)queue_create_info[0].pQueuePriorities);
857 av_free((void *)queue_create_info[1].pQueuePriorities);
858 av_free((void *)queue_create_info[2].pQueuePriorities);
860 if (ret != VK_SUCCESS) {
861 av_log(ctx, AV_LOG_ERROR, "Device creation failure: %s\n",
863 for (int i = 0; i < dev_info.enabledExtensionCount; i++)
864 av_free((void *)dev_info.ppEnabledExtensionNames[i]);
865 av_free((void *)dev_info.ppEnabledExtensionNames);
866 err = AVERROR_EXTERNAL;
870 /* Tiled images setting, use them by default */
871 opt_d = av_dict_get(opts, "linear_images", NULL, 0);
873 p->use_linear_images = strtol(opt_d->value, NULL, 10);
875 hwctx->enabled_dev_extensions = dev_info.ppEnabledExtensionNames;
876 hwctx->nb_enabled_dev_extensions = dev_info.enabledExtensionCount;
882 static int vulkan_device_init(AVHWDeviceContext *ctx)
886 AVVulkanDeviceContext *hwctx = ctx->hwctx;
887 VulkanDevicePriv *p = ctx->internal->priv;
889 /* Set device extension flags */
890 for (int i = 0; i < hwctx->nb_enabled_dev_extensions; i++) {
891 for (int j = 0; j < FF_ARRAY_ELEMS(optional_device_exts); j++) {
892 if (!strcmp(hwctx->enabled_dev_extensions[i],
893 optional_device_exts[j].name)) {
894 p->extensions |= optional_device_exts[j].flag;
900 vkGetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &queue_num, NULL);
902 av_log(ctx, AV_LOG_ERROR, "Failed to get queues!\n");
903 return AVERROR_EXTERNAL;
906 #define CHECK_QUEUE(type, n) \
907 if (n >= queue_num) { \
908 av_log(ctx, AV_LOG_ERROR, "Invalid %s queue index %i (device has %i queues)!\n", \
909 type, n, queue_num); \
910 return AVERROR(EINVAL); \
913 CHECK_QUEUE("graphics", hwctx->queue_family_index)
914 CHECK_QUEUE("upload", hwctx->queue_family_tx_index)
915 CHECK_QUEUE("compute", hwctx->queue_family_comp_index)
919 p->qfs[p->num_qfs++] = hwctx->queue_family_index;
920 if ((hwctx->queue_family_tx_index != hwctx->queue_family_index) &&
921 (hwctx->queue_family_tx_index != hwctx->queue_family_comp_index))
922 p->qfs[p->num_qfs++] = hwctx->queue_family_tx_index;
923 if ((hwctx->queue_family_comp_index != hwctx->queue_family_index) &&
924 (hwctx->queue_family_comp_index != hwctx->queue_family_tx_index))
925 p->qfs[p->num_qfs++] = hwctx->queue_family_comp_index;
927 /* Create exec context - if there's something invalid this will error out */
928 err = create_exec_ctx(ctx, &p->cmd, hwctx->queue_family_tx_index);
932 /* Get device capabilities */
933 vkGetPhysicalDeviceMemoryProperties(hwctx->phys_dev, &p->mprops);
938 static int vulkan_device_create(AVHWDeviceContext *ctx, const char *device,
939 AVDictionary *opts, int flags)
941 VulkanDeviceSelection dev_select = { 0 };
942 if (device && device[0]) {
944 dev_select.index = strtol(device, &end, 10);
946 dev_select.index = 0;
947 dev_select.name = device;
951 return vulkan_device_create_internal(ctx, &dev_select, opts, flags);
954 static int vulkan_device_derive(AVHWDeviceContext *ctx,
955 AVHWDeviceContext *src_ctx, int flags)
957 av_unused VulkanDeviceSelection dev_select = { 0 };
959 /* If there's only one device on the system, then even if its not covered
960 * by the following checks (e.g. non-PCIe ARM GPU), having an empty
961 * dev_select will mean it'll get picked. */
962 switch(src_ctx->type) {
965 case AV_HWDEVICE_TYPE_VAAPI: {
966 AVVAAPIDeviceContext *src_hwctx = src_ctx->hwctx;
968 const char *vendor = vaQueryVendorString(src_hwctx->display);
970 av_log(ctx, AV_LOG_ERROR, "Unable to get device info from VAAPI!\n");
971 return AVERROR_EXTERNAL;
974 if (strstr(vendor, "Intel"))
975 dev_select.vendor_id = 0x8086;
976 if (strstr(vendor, "AMD"))
977 dev_select.vendor_id = 0x1002;
979 return vulkan_device_create_internal(ctx, &dev_select, NULL, flags);
982 case AV_HWDEVICE_TYPE_DRM: {
983 AVDRMDeviceContext *src_hwctx = src_ctx->hwctx;
985 drmDevice *drm_dev_info;
986 int err = drmGetDevice(src_hwctx->fd, &drm_dev_info);
988 av_log(ctx, AV_LOG_ERROR, "Unable to get device info from DRM fd!\n");
989 return AVERROR_EXTERNAL;
992 if (drm_dev_info->bustype == DRM_BUS_PCI)
993 dev_select.pci_device = drm_dev_info->deviceinfo.pci->device_id;
995 drmFreeDevice(&drm_dev_info);
997 return vulkan_device_create_internal(ctx, &dev_select, NULL, flags);
1001 case AV_HWDEVICE_TYPE_CUDA: {
1002 AVHWDeviceContext *cuda_cu = src_ctx;
1003 AVCUDADeviceContext *src_hwctx = src_ctx->hwctx;
1004 AVCUDADeviceContextInternal *cu_internal = src_hwctx->internal;
1005 CudaFunctions *cu = cu_internal->cuda_dl;
1007 int ret = CHECK_CU(cu->cuDeviceGetUuid((CUuuid *)&dev_select.uuid,
1008 cu_internal->cuda_device));
1010 av_log(ctx, AV_LOG_ERROR, "Unable to get UUID from CUDA!\n");
1011 return AVERROR_EXTERNAL;
1014 dev_select.has_uuid = 1;
1016 return vulkan_device_create_internal(ctx, &dev_select, NULL, flags);
1020 return AVERROR(ENOSYS);
1024 static int vulkan_frames_get_constraints(AVHWDeviceContext *ctx,
1025 const void *hwconfig,
1026 AVHWFramesConstraints *constraints)
1029 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1030 VulkanDevicePriv *p = ctx->internal->priv;
1032 for (enum AVPixelFormat i = 0; i < AV_PIX_FMT_NB; i++)
1033 count += pixfmt_is_supported(hwctx, i, p->use_linear_images);
1036 if (p->dev_is_nvidia)
1040 constraints->valid_sw_formats = av_malloc_array(count + 1,
1041 sizeof(enum AVPixelFormat));
1042 if (!constraints->valid_sw_formats)
1043 return AVERROR(ENOMEM);
1046 for (enum AVPixelFormat i = 0; i < AV_PIX_FMT_NB; i++)
1047 if (pixfmt_is_supported(hwctx, i, p->use_linear_images))
1048 constraints->valid_sw_formats[count++] = i;
1051 if (p->dev_is_nvidia)
1052 constraints->valid_sw_formats[count++] = AV_PIX_FMT_CUDA;
1054 constraints->valid_sw_formats[count++] = AV_PIX_FMT_NONE;
1056 constraints->min_width = 0;
1057 constraints->min_height = 0;
1058 constraints->max_width = p->props.limits.maxImageDimension2D;
1059 constraints->max_height = p->props.limits.maxImageDimension2D;
1061 constraints->valid_hw_formats = av_malloc_array(2, sizeof(enum AVPixelFormat));
1062 if (!constraints->valid_hw_formats)
1063 return AVERROR(ENOMEM);
1065 constraints->valid_hw_formats[0] = AV_PIX_FMT_VULKAN;
1066 constraints->valid_hw_formats[1] = AV_PIX_FMT_NONE;
1071 static int alloc_mem(AVHWDeviceContext *ctx, VkMemoryRequirements *req,
1072 VkMemoryPropertyFlagBits req_flags, void *alloc_extension,
1073 VkMemoryPropertyFlagBits *mem_flags, VkDeviceMemory *mem)
1077 VulkanDevicePriv *p = ctx->internal->priv;
1078 AVVulkanDeviceContext *dev_hwctx = ctx->hwctx;
1079 VkMemoryAllocateInfo alloc_info = {
1080 .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
1081 .pNext = alloc_extension,
1084 /* Align if we need to */
1085 if (req_flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT)
1086 req->size = FFALIGN(req->size, p->props.limits.minMemoryMapAlignment);
1088 alloc_info.allocationSize = req->size;
1090 /* The vulkan spec requires memory types to be sorted in the "optimal"
1091 * order, so the first matching type we find will be the best/fastest one */
1092 for (int i = 0; i < p->mprops.memoryTypeCount; i++) {
1093 /* The memory type must be supported by the requirements (bitfield) */
1094 if (!(req->memoryTypeBits & (1 << i)))
1097 /* The memory type flags must include our properties */
1098 if ((p->mprops.memoryTypes[i].propertyFlags & req_flags) != req_flags)
1101 /* Found a suitable memory type */
1107 av_log(ctx, AV_LOG_ERROR, "No memory type found for flags 0x%x\n",
1109 return AVERROR(EINVAL);
1112 alloc_info.memoryTypeIndex = index;
1114 ret = vkAllocateMemory(dev_hwctx->act_dev, &alloc_info,
1115 dev_hwctx->alloc, mem);
1116 if (ret != VK_SUCCESS) {
1117 av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory: %s\n",
1119 return AVERROR(ENOMEM);
1122 *mem_flags |= p->mprops.memoryTypes[index].propertyFlags;
1127 static void vulkan_free_internal(AVVkFrameInternal *internal)
1133 if (internal->cuda_fc_ref) {
1134 AVHWFramesContext *cuda_fc = (AVHWFramesContext *)internal->cuda_fc_ref->data;
1135 int planes = av_pix_fmt_count_planes(cuda_fc->sw_format);
1136 AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
1137 AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
1138 AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
1139 CudaFunctions *cu = cu_internal->cuda_dl;
1141 for (int i = 0; i < planes; i++) {
1142 if (internal->cu_sem[i])
1143 CHECK_CU(cu->cuDestroyExternalSemaphore(internal->cu_sem[i]));
1144 if (internal->cu_mma[i])
1145 CHECK_CU(cu->cuMipmappedArrayDestroy(internal->cu_mma[i]));
1146 if (internal->ext_mem[i])
1147 CHECK_CU(cu->cuDestroyExternalMemory(internal->ext_mem[i]));
1150 av_buffer_unref(&internal->cuda_fc_ref);
1157 static void vulkan_frame_free(void *opaque, uint8_t *data)
1159 AVVkFrame *f = (AVVkFrame *)data;
1160 AVHWFramesContext *hwfc = opaque;
1161 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1162 int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1164 vulkan_free_internal(f->internal);
1166 for (int i = 0; i < planes; i++) {
1167 vkDestroyImage(hwctx->act_dev, f->img[i], hwctx->alloc);
1168 vkFreeMemory(hwctx->act_dev, f->mem[i], hwctx->alloc);
1169 vkDestroySemaphore(hwctx->act_dev, f->sem[i], hwctx->alloc);
1175 static int alloc_bind_mem(AVHWFramesContext *hwfc, AVVkFrame *f,
1176 void *alloc_pnext, size_t alloc_pnext_stride)
1180 AVHWDeviceContext *ctx = hwfc->device_ctx;
1181 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1182 VkBindImageMemoryInfo bind_info[AV_NUM_DATA_POINTERS] = { { 0 } };
1184 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1186 for (int i = 0; i < planes; i++) {
1188 VkImageMemoryRequirementsInfo2 req_desc = {
1189 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_REQUIREMENTS_INFO_2,
1192 VkMemoryDedicatedAllocateInfo ded_alloc = {
1193 .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO,
1194 .pNext = (void *)(((uint8_t *)alloc_pnext) + i*alloc_pnext_stride),
1196 VkMemoryDedicatedRequirements ded_req = {
1197 .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS,
1199 VkMemoryRequirements2 req = {
1200 .sType = VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2,
1204 vkGetImageMemoryRequirements2(hwctx->act_dev, &req_desc, &req);
1206 /* In case the implementation prefers/requires dedicated allocation */
1207 use_ded_mem = ded_req.prefersDedicatedAllocation |
1208 ded_req.requiresDedicatedAllocation;
1210 ded_alloc.image = f->img[i];
1212 /* Allocate memory */
1213 if ((err = alloc_mem(ctx, &req.memoryRequirements,
1214 f->tiling == VK_IMAGE_TILING_LINEAR ?
1215 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT :
1216 VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
1217 use_ded_mem ? &ded_alloc : (void *)ded_alloc.pNext,
1218 &f->flags, &f->mem[i])))
1221 f->size[i] = req.memoryRequirements.size;
1222 bind_info[i].sType = VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO;
1223 bind_info[i].image = f->img[i];
1224 bind_info[i].memory = f->mem[i];
1227 /* Bind the allocated memory to the images */
1228 ret = vkBindImageMemory2(hwctx->act_dev, planes, bind_info);
1229 if (ret != VK_SUCCESS) {
1230 av_log(ctx, AV_LOG_ERROR, "Failed to bind memory: %s\n",
1232 return AVERROR_EXTERNAL;
1240 PREP_MODE_RO_SHADER,
1241 PREP_MODE_EXTERNAL_EXPORT,
1244 static int prepare_frame(AVHWFramesContext *hwfc, VulkanExecCtx *ectx,
1245 AVVkFrame *frame, enum PrepMode pmode)
1249 VkImageLayout new_layout;
1250 VkAccessFlags new_access;
1251 AVHWDeviceContext *ctx = hwfc->device_ctx;
1252 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1253 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1255 VkImageMemoryBarrier img_bar[AV_NUM_DATA_POINTERS] = { 0 };
1257 VkCommandBufferBeginInfo cmd_start = {
1258 .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
1259 .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
1262 VkSubmitInfo s_info = {
1263 .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
1264 .commandBufferCount = 1,
1265 .pCommandBuffers = &ectx->buf,
1267 .pSignalSemaphores = frame->sem,
1268 .signalSemaphoreCount = planes,
1271 VkPipelineStageFlagBits wait_st[AV_NUM_DATA_POINTERS];
1272 for (int i = 0; i < planes; i++)
1273 wait_st[i] = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
1276 case PREP_MODE_WRITE:
1277 new_layout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
1278 new_access = VK_ACCESS_TRANSFER_WRITE_BIT;
1279 dst_qf = VK_QUEUE_FAMILY_IGNORED;
1281 case PREP_MODE_RO_SHADER:
1282 new_layout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL;
1283 new_access = VK_ACCESS_TRANSFER_READ_BIT;
1284 dst_qf = VK_QUEUE_FAMILY_IGNORED;
1286 case PREP_MODE_EXTERNAL_EXPORT:
1287 new_layout = VK_IMAGE_LAYOUT_GENERAL;
1288 new_access = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT;
1289 dst_qf = VK_QUEUE_FAMILY_EXTERNAL_KHR;
1290 s_info.pWaitSemaphores = frame->sem;
1291 s_info.pWaitDstStageMask = wait_st;
1292 s_info.waitSemaphoreCount = planes;
1296 ret = vkBeginCommandBuffer(ectx->buf, &cmd_start);
1297 if (ret != VK_SUCCESS)
1298 return AVERROR_EXTERNAL;
1300 /* Change the image layout to something more optimal for writes.
1301 * This also signals the newly created semaphore, making it usable
1302 * for synchronization */
1303 for (int i = 0; i < planes; i++) {
1304 img_bar[i].sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
1305 img_bar[i].srcAccessMask = 0x0;
1306 img_bar[i].dstAccessMask = new_access;
1307 img_bar[i].oldLayout = frame->layout[i];
1308 img_bar[i].newLayout = new_layout;
1309 img_bar[i].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
1310 img_bar[i].dstQueueFamilyIndex = dst_qf;
1311 img_bar[i].image = frame->img[i];
1312 img_bar[i].subresourceRange.levelCount = 1;
1313 img_bar[i].subresourceRange.layerCount = 1;
1314 img_bar[i].subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
1316 frame->layout[i] = img_bar[i].newLayout;
1317 frame->access[i] = img_bar[i].dstAccessMask;
1320 vkCmdPipelineBarrier(ectx->buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
1321 VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
1322 0, NULL, 0, NULL, planes, img_bar);
1324 ret = vkEndCommandBuffer(ectx->buf);
1325 if (ret != VK_SUCCESS)
1326 return AVERROR_EXTERNAL;
1328 ret = vkQueueSubmit(ectx->queue, 1, &s_info, ectx->fence);
1329 if (ret != VK_SUCCESS) {
1330 return AVERROR_EXTERNAL;
1332 vkWaitForFences(hwctx->act_dev, 1, &ectx->fence, VK_TRUE, UINT64_MAX);
1333 vkResetFences(hwctx->act_dev, 1, &ectx->fence);
1339 static int create_frame(AVHWFramesContext *hwfc, AVVkFrame **frame,
1340 VkImageTiling tiling, VkImageUsageFlagBits usage,
1345 AVHWDeviceContext *ctx = hwfc->device_ctx;
1346 VulkanDevicePriv *p = ctx->internal->priv;
1347 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1348 enum AVPixelFormat format = hwfc->sw_format;
1349 const VkFormat *img_fmts = av_vkfmt_from_pixfmt(format);
1350 const int planes = av_pix_fmt_count_planes(format);
1352 VkExportSemaphoreCreateInfo ext_sem_info = {
1353 .sType = VK_STRUCTURE_TYPE_EXPORT_SEMAPHORE_CREATE_INFO,
1354 .handleTypes = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT,
1357 VkSemaphoreCreateInfo sem_spawn = {
1358 .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
1359 .pNext = p->extensions & EXT_EXTERNAL_FD_SEM ? &ext_sem_info : NULL,
1362 AVVkFrame *f = av_vk_frame_alloc();
1364 av_log(ctx, AV_LOG_ERROR, "Unable to allocate memory for AVVkFrame!\n");
1365 return AVERROR(ENOMEM);
1368 /* Create the images */
1369 for (int i = 0; i < planes; i++) {
1370 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(format);
1371 int w = hwfc->width;
1372 int h = hwfc->height;
1373 const int p_w = i > 0 ? AV_CEIL_RSHIFT(w, desc->log2_chroma_w) : w;
1374 const int p_h = i > 0 ? AV_CEIL_RSHIFT(h, desc->log2_chroma_h) : h;
1376 VkImageCreateInfo image_create_info = {
1377 .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
1378 .pNext = create_pnext,
1379 .imageType = VK_IMAGE_TYPE_2D,
1380 .format = img_fmts[i],
1381 .extent.width = p_w,
1382 .extent.height = p_h,
1386 .flags = VK_IMAGE_CREATE_ALIAS_BIT,
1388 .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED,
1390 .samples = VK_SAMPLE_COUNT_1_BIT,
1391 .pQueueFamilyIndices = p->qfs,
1392 .queueFamilyIndexCount = p->num_qfs,
1393 .sharingMode = p->num_qfs > 1 ? VK_SHARING_MODE_CONCURRENT :
1394 VK_SHARING_MODE_EXCLUSIVE,
1397 ret = vkCreateImage(hwctx->act_dev, &image_create_info,
1398 hwctx->alloc, &f->img[i]);
1399 if (ret != VK_SUCCESS) {
1400 av_log(ctx, AV_LOG_ERROR, "Image creation failure: %s\n",
1402 err = AVERROR(EINVAL);
1406 /* Create semaphore */
1407 ret = vkCreateSemaphore(hwctx->act_dev, &sem_spawn,
1408 hwctx->alloc, &f->sem[i]);
1409 if (ret != VK_SUCCESS) {
1410 av_log(hwctx, AV_LOG_ERROR, "Failed to create semaphore: %s\n",
1412 return AVERROR_EXTERNAL;
1415 f->layout[i] = image_create_info.initialLayout;
1426 vulkan_frame_free(hwfc, (uint8_t *)f);
1430 /* Checks if an export flag is enabled, and if it is ORs it with *iexp */
1431 static void try_export_flags(AVHWFramesContext *hwfc,
1432 VkExternalMemoryHandleTypeFlags *comp_handle_types,
1433 VkExternalMemoryHandleTypeFlagBits *iexp,
1434 VkExternalMemoryHandleTypeFlagBits exp)
1437 AVVulkanFramesContext *hwctx = hwfc->hwctx;
1438 AVVulkanDeviceContext *dev_hwctx = hwfc->device_ctx->hwctx;
1439 VkExternalImageFormatProperties eprops = {
1440 .sType = VK_STRUCTURE_TYPE_EXTERNAL_IMAGE_FORMAT_PROPERTIES_KHR,
1442 VkImageFormatProperties2 props = {
1443 .sType = VK_STRUCTURE_TYPE_IMAGE_FORMAT_PROPERTIES_2,
1446 VkPhysicalDeviceExternalImageFormatInfo enext = {
1447 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_IMAGE_FORMAT_INFO,
1450 VkPhysicalDeviceImageFormatInfo2 pinfo = {
1451 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_FORMAT_INFO_2,
1452 .pNext = !exp ? NULL : &enext,
1453 .format = av_vkfmt_from_pixfmt(hwfc->sw_format)[0],
1454 .type = VK_IMAGE_TYPE_2D,
1455 .tiling = hwctx->tiling,
1456 .usage = hwctx->usage,
1457 .flags = VK_IMAGE_CREATE_ALIAS_BIT,
1460 ret = vkGetPhysicalDeviceImageFormatProperties2(dev_hwctx->phys_dev,
1462 if (ret == VK_SUCCESS) {
1464 *comp_handle_types |= eprops.externalMemoryProperties.compatibleHandleTypes;
1468 static AVBufferRef *vulkan_pool_alloc(void *opaque, int size)
1472 AVBufferRef *avbuf = NULL;
1473 AVHWFramesContext *hwfc = opaque;
1474 AVVulkanFramesContext *hwctx = hwfc->hwctx;
1475 VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
1476 VkExportMemoryAllocateInfo eminfo[AV_NUM_DATA_POINTERS];
1477 VkExternalMemoryHandleTypeFlags e = 0x0;
1479 VkExternalMemoryImageCreateInfo eiinfo = {
1480 .sType = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO,
1481 .pNext = hwctx->create_pnext,
1484 if (p->extensions & EXT_EXTERNAL_FD_MEMORY)
1485 try_export_flags(hwfc, &eiinfo.handleTypes, &e,
1486 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT);
1488 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
1489 try_export_flags(hwfc, &eiinfo.handleTypes, &e,
1490 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
1492 for (int i = 0; i < av_pix_fmt_count_planes(hwfc->sw_format); i++) {
1493 eminfo[i].sType = VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO;
1494 eminfo[i].pNext = hwctx->alloc_pnext[i];
1495 eminfo[i].handleTypes = e;
1498 err = create_frame(hwfc, &f, hwctx->tiling, hwctx->usage,
1499 eiinfo.handleTypes ? &eiinfo : NULL);
1503 err = alloc_bind_mem(hwfc, f, eminfo, sizeof(*eminfo));
1507 err = prepare_frame(hwfc, &p->cmd, f, PREP_MODE_WRITE);
1511 avbuf = av_buffer_create((uint8_t *)f, sizeof(AVVkFrame),
1512 vulkan_frame_free, hwfc, 0);
1519 vulkan_frame_free(hwfc, (uint8_t *)f);
1523 static void vulkan_frames_uninit(AVHWFramesContext *hwfc)
1525 VulkanFramesPriv *fp = hwfc->internal->priv;
1527 free_exec_ctx(hwfc->device_ctx, &fp->cmd);
1530 static int vulkan_frames_init(AVHWFramesContext *hwfc)
1534 AVVulkanFramesContext *hwctx = hwfc->hwctx;
1535 VulkanFramesPriv *fp = hwfc->internal->priv;
1536 AVVulkanDeviceContext *dev_hwctx = hwfc->device_ctx->hwctx;
1537 VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
1542 /* Default pool flags */
1543 hwctx->tiling = hwctx->tiling ? hwctx->tiling : p->use_linear_images ?
1544 VK_IMAGE_TILING_LINEAR : VK_IMAGE_TILING_OPTIMAL;
1546 hwctx->usage |= DEFAULT_USAGE_FLAGS;
1548 err = create_exec_ctx(hwfc->device_ctx, &fp->cmd,
1549 dev_hwctx->queue_family_tx_index);
1553 /* Test to see if allocation will fail */
1554 err = create_frame(hwfc, &f, hwctx->tiling, hwctx->usage,
1555 hwctx->create_pnext);
1557 free_exec_ctx(hwfc->device_ctx, &p->cmd);
1561 vulkan_frame_free(hwfc, (uint8_t *)f);
1563 hwfc->internal->pool_internal = av_buffer_pool_init2(sizeof(AVVkFrame),
1564 hwfc, vulkan_pool_alloc,
1566 if (!hwfc->internal->pool_internal) {
1567 free_exec_ctx(hwfc->device_ctx, &p->cmd);
1568 return AVERROR(ENOMEM);
1574 static int vulkan_get_buffer(AVHWFramesContext *hwfc, AVFrame *frame)
1576 frame->buf[0] = av_buffer_pool_get(hwfc->pool);
1578 return AVERROR(ENOMEM);
1580 frame->data[0] = frame->buf[0]->data;
1581 frame->format = AV_PIX_FMT_VULKAN;
1582 frame->width = hwfc->width;
1583 frame->height = hwfc->height;
1588 static int vulkan_transfer_get_formats(AVHWFramesContext *hwfc,
1589 enum AVHWFrameTransferDirection dir,
1590 enum AVPixelFormat **formats)
1592 enum AVPixelFormat *fmts = av_malloc_array(2, sizeof(*fmts));
1594 return AVERROR(ENOMEM);
1596 fmts[0] = hwfc->sw_format;
1597 fmts[1] = AV_PIX_FMT_NONE;
1603 typedef struct VulkanMapping {
1608 static void vulkan_unmap_frame(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
1610 VulkanMapping *map = hwmap->priv;
1611 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1612 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1614 /* Check if buffer needs flushing */
1615 if ((map->flags & AV_HWFRAME_MAP_WRITE) &&
1616 !(map->frame->flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)) {
1618 VkMappedMemoryRange flush_ranges[AV_NUM_DATA_POINTERS] = { { 0 } };
1620 for (int i = 0; i < planes; i++) {
1621 flush_ranges[i].sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
1622 flush_ranges[i].memory = map->frame->mem[i];
1623 flush_ranges[i].size = VK_WHOLE_SIZE;
1626 ret = vkFlushMappedMemoryRanges(hwctx->act_dev, planes,
1628 if (ret != VK_SUCCESS) {
1629 av_log(hwfc, AV_LOG_ERROR, "Failed to flush memory: %s\n",
1634 for (int i = 0; i < planes; i++)
1635 vkUnmapMemory(hwctx->act_dev, map->frame->mem[i]);
1640 static int vulkan_map_frame_to_mem(AVHWFramesContext *hwfc, AVFrame *dst,
1641 const AVFrame *src, int flags)
1644 int err, mapped_mem_count = 0;
1645 AVVkFrame *f = (AVVkFrame *)src->data[0];
1646 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1647 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1649 VulkanMapping *map = av_mallocz(sizeof(VulkanMapping));
1651 return AVERROR(EINVAL);
1653 if (src->format != AV_PIX_FMT_VULKAN) {
1654 av_log(hwfc, AV_LOG_ERROR, "Cannot map from pixel format %s!\n",
1655 av_get_pix_fmt_name(src->format));
1656 err = AVERROR(EINVAL);
1660 if (!(f->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) ||
1661 !(f->tiling == VK_IMAGE_TILING_LINEAR)) {
1662 av_log(hwfc, AV_LOG_ERROR, "Unable to map frame, not host visible "
1664 err = AVERROR(EINVAL);
1668 dst->width = src->width;
1669 dst->height = src->height;
1671 for (int i = 0; i < planes; i++) {
1672 ret = vkMapMemory(hwctx->act_dev, f->mem[i], 0,
1673 VK_WHOLE_SIZE, 0, (void **)&dst->data[i]);
1674 if (ret != VK_SUCCESS) {
1675 av_log(hwfc, AV_LOG_ERROR, "Failed to map image memory: %s\n",
1677 err = AVERROR_EXTERNAL;
1683 /* Check if the memory contents matter */
1684 if (((flags & AV_HWFRAME_MAP_READ) || !(flags & AV_HWFRAME_MAP_OVERWRITE)) &&
1685 !(f->flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)) {
1686 VkMappedMemoryRange map_mem_ranges[AV_NUM_DATA_POINTERS] = { { 0 } };
1687 for (int i = 0; i < planes; i++) {
1688 map_mem_ranges[i].sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
1689 map_mem_ranges[i].size = VK_WHOLE_SIZE;
1690 map_mem_ranges[i].memory = f->mem[i];
1693 ret = vkInvalidateMappedMemoryRanges(hwctx->act_dev, planes,
1695 if (ret != VK_SUCCESS) {
1696 av_log(hwfc, AV_LOG_ERROR, "Failed to invalidate memory: %s\n",
1698 err = AVERROR_EXTERNAL;
1703 for (int i = 0; i < planes; i++) {
1704 VkImageSubresource sub = {
1705 .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
1707 VkSubresourceLayout layout;
1708 vkGetImageSubresourceLayout(hwctx->act_dev, f->img[i], &sub, &layout);
1709 dst->linesize[i] = layout.rowPitch;
1715 err = ff_hwframe_map_create(src->hw_frames_ctx, dst, src,
1716 &vulkan_unmap_frame, map);
1723 for (int i = 0; i < mapped_mem_count; i++)
1724 vkUnmapMemory(hwctx->act_dev, f->mem[i]);
1731 static void vulkan_unmap_from(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
1733 VulkanMapping *map = hwmap->priv;
1734 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1735 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1737 for (int i = 0; i < planes; i++) {
1738 vkDestroyImage(hwctx->act_dev, map->frame->img[i], hwctx->alloc);
1739 vkFreeMemory(hwctx->act_dev, map->frame->mem[i], hwctx->alloc);
1740 vkDestroySemaphore(hwctx->act_dev, map->frame->sem[i], hwctx->alloc);
1743 av_freep(&map->frame);
1746 static const struct {
1747 uint32_t drm_fourcc;
1749 } vulkan_drm_format_map[] = {
1750 { DRM_FORMAT_R8, VK_FORMAT_R8_UNORM },
1751 { DRM_FORMAT_R16, VK_FORMAT_R16_UNORM },
1752 { DRM_FORMAT_GR88, VK_FORMAT_R8G8_UNORM },
1753 { DRM_FORMAT_RG88, VK_FORMAT_R8G8_UNORM },
1754 { DRM_FORMAT_GR1616, VK_FORMAT_R16G16_UNORM },
1755 { DRM_FORMAT_RG1616, VK_FORMAT_R16G16_UNORM },
1756 { DRM_FORMAT_ARGB8888, VK_FORMAT_B8G8R8A8_UNORM },
1757 { DRM_FORMAT_XRGB8888, VK_FORMAT_B8G8R8A8_UNORM },
1758 { DRM_FORMAT_ABGR8888, VK_FORMAT_R8G8B8A8_UNORM },
1759 { DRM_FORMAT_XBGR8888, VK_FORMAT_R8G8B8A8_UNORM },
1762 static inline VkFormat drm_to_vulkan_fmt(uint32_t drm_fourcc)
1764 for (int i = 0; i < FF_ARRAY_ELEMS(vulkan_drm_format_map); i++)
1765 if (vulkan_drm_format_map[i].drm_fourcc == drm_fourcc)
1766 return vulkan_drm_format_map[i].vk_format;
1767 return VK_FORMAT_UNDEFINED;
1770 static int vulkan_map_from_drm_frame_desc(AVHWFramesContext *hwfc, AVVkFrame **frame,
1771 AVDRMFrameDescriptor *desc)
1776 int bind_counts = 0;
1777 AVHWDeviceContext *ctx = hwfc->device_ctx;
1778 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1779 VulkanDevicePriv *p = ctx->internal->priv;
1780 const AVPixFmtDescriptor *fmt_desc = av_pix_fmt_desc_get(hwfc->sw_format);
1781 const int has_modifiers = p->extensions & EXT_DRM_MODIFIER_FLAGS;
1782 VkSubresourceLayout plane_data[AV_NUM_DATA_POINTERS] = { 0 };
1783 VkBindImageMemoryInfo bind_info[AV_NUM_DATA_POINTERS] = { 0 };
1784 VkBindImagePlaneMemoryInfo plane_info[AV_NUM_DATA_POINTERS] = { 0 };
1785 VkExternalMemoryHandleTypeFlagBits htype = VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT;
1787 VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdPropertiesKHR);
1789 for (int i = 0; i < desc->nb_layers; i++) {
1790 if (drm_to_vulkan_fmt(desc->layers[i].format) == VK_FORMAT_UNDEFINED) {
1791 av_log(ctx, AV_LOG_ERROR, "Unsupported DMABUF layer format %#08x!\n",
1792 desc->layers[i].format);
1793 return AVERROR(EINVAL);
1797 if (!(f = av_vk_frame_alloc())) {
1798 av_log(ctx, AV_LOG_ERROR, "Unable to allocate memory for AVVkFrame!\n");
1799 err = AVERROR(ENOMEM);
1803 for (int i = 0; i < desc->nb_objects; i++) {
1804 VkMemoryFdPropertiesKHR fdmp = {
1805 .sType = VK_STRUCTURE_TYPE_MEMORY_FD_PROPERTIES_KHR,
1807 VkMemoryRequirements req = {
1808 .size = desc->objects[i].size,
1810 VkImportMemoryFdInfoKHR idesc = {
1811 .sType = VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR,
1812 .handleType = htype,
1813 .fd = dup(desc->objects[i].fd),
1816 ret = pfn_vkGetMemoryFdPropertiesKHR(hwctx->act_dev, htype,
1818 if (ret != VK_SUCCESS) {
1819 av_log(hwfc, AV_LOG_ERROR, "Failed to get FD properties: %s\n",
1821 err = AVERROR_EXTERNAL;
1826 req.memoryTypeBits = fdmp.memoryTypeBits;
1828 err = alloc_mem(ctx, &req, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
1829 &idesc, &f->flags, &f->mem[i]);
1835 f->size[i] = desc->objects[i].size;
1838 f->tiling = has_modifiers ? VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT :
1839 desc->objects[0].format_modifier == DRM_FORMAT_MOD_LINEAR ?
1840 VK_IMAGE_TILING_LINEAR : VK_IMAGE_TILING_OPTIMAL;
1842 for (int i = 0; i < desc->nb_layers; i++) {
1843 const int planes = desc->layers[i].nb_planes;
1844 const int signal_p = has_modifiers && (planes > 1);
1846 VkImageDrmFormatModifierExplicitCreateInfoEXT drm_info = {
1847 .sType = VK_STRUCTURE_TYPE_IMAGE_DRM_FORMAT_MODIFIER_EXPLICIT_CREATE_INFO_EXT,
1848 .drmFormatModifier = desc->objects[0].format_modifier,
1849 .drmFormatModifierPlaneCount = planes,
1850 .pPlaneLayouts = (const VkSubresourceLayout *)&plane_data,
1853 VkExternalMemoryImageCreateInfo einfo = {
1854 .sType = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO,
1855 .pNext = has_modifiers ? &drm_info : NULL,
1856 .handleTypes = htype,
1859 VkSemaphoreCreateInfo sem_spawn = {
1860 .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
1863 const int p_w = i > 0 ? AV_CEIL_RSHIFT(hwfc->width, fmt_desc->log2_chroma_w) : hwfc->width;
1864 const int p_h = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, fmt_desc->log2_chroma_h) : hwfc->height;
1866 VkImageCreateInfo image_create_info = {
1867 .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
1869 .imageType = VK_IMAGE_TYPE_2D,
1870 .format = drm_to_vulkan_fmt(desc->layers[i].format),
1871 .extent.width = p_w,
1872 .extent.height = p_h,
1876 .flags = VK_IMAGE_CREATE_ALIAS_BIT,
1877 .tiling = f->tiling,
1878 .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED, /* specs say so */
1879 .usage = DEFAULT_USAGE_FLAGS,
1880 .samples = VK_SAMPLE_COUNT_1_BIT,
1881 .pQueueFamilyIndices = p->qfs,
1882 .queueFamilyIndexCount = p->num_qfs,
1883 .sharingMode = p->num_qfs > 1 ? VK_SHARING_MODE_CONCURRENT :
1884 VK_SHARING_MODE_EXCLUSIVE,
1887 for (int j = 0; j < planes; j++) {
1888 plane_data[j].offset = desc->layers[i].planes[j].offset;
1889 plane_data[j].rowPitch = desc->layers[i].planes[j].pitch;
1890 plane_data[j].size = 0; /* The specs say so for all 3 */
1891 plane_data[j].arrayPitch = 0;
1892 plane_data[j].depthPitch = 0;
1896 ret = vkCreateImage(hwctx->act_dev, &image_create_info,
1897 hwctx->alloc, &f->img[i]);
1898 if (ret != VK_SUCCESS) {
1899 av_log(ctx, AV_LOG_ERROR, "Image creation failure: %s\n",
1901 err = AVERROR(EINVAL);
1905 ret = vkCreateSemaphore(hwctx->act_dev, &sem_spawn,
1906 hwctx->alloc, &f->sem[i]);
1907 if (ret != VK_SUCCESS) {
1908 av_log(hwctx, AV_LOG_ERROR, "Failed to create semaphore: %s\n",
1910 return AVERROR_EXTERNAL;
1913 /* We'd import a semaphore onto the one we created using
1914 * vkImportSemaphoreFdKHR but unfortunately neither DRM nor VAAPI
1915 * offer us anything we could import and sync with, so instead
1916 * just signal the semaphore we created. */
1918 f->layout[i] = image_create_info.initialLayout;
1921 for (int j = 0; j < planes; j++) {
1922 VkImageAspectFlagBits aspect = j == 0 ? VK_IMAGE_ASPECT_MEMORY_PLANE_0_BIT_EXT :
1923 j == 1 ? VK_IMAGE_ASPECT_MEMORY_PLANE_1_BIT_EXT :
1924 VK_IMAGE_ASPECT_MEMORY_PLANE_2_BIT_EXT;
1926 plane_info[bind_counts].sType = VK_STRUCTURE_TYPE_BIND_IMAGE_PLANE_MEMORY_INFO;
1927 plane_info[bind_counts].planeAspect = aspect;
1929 bind_info[bind_counts].sType = VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO;
1930 bind_info[bind_counts].pNext = signal_p ? &plane_info[bind_counts] : NULL;
1931 bind_info[bind_counts].image = f->img[i];
1932 bind_info[bind_counts].memory = f->mem[desc->layers[i].planes[j].object_index];
1933 bind_info[bind_counts].memoryOffset = desc->layers[i].planes[j].offset;
1938 /* Bind the allocated memory to the images */
1939 ret = vkBindImageMemory2(hwctx->act_dev, bind_counts, bind_info);
1940 if (ret != VK_SUCCESS) {
1941 av_log(ctx, AV_LOG_ERROR, "Failed to bind memory: %s\n",
1943 return AVERROR_EXTERNAL;
1946 /* NOTE: This is completely uneccesary and unneeded once we can import
1947 * semaphores from DRM. Otherwise we have to activate the semaphores.
1948 * We're reusing the exec context that's also used for uploads/downloads. */
1949 err = prepare_frame(hwfc, &p->cmd, f, PREP_MODE_RO_SHADER);
1958 for (int i = 0; i < desc->nb_layers; i++) {
1959 vkDestroyImage(hwctx->act_dev, f->img[i], hwctx->alloc);
1960 vkDestroySemaphore(hwctx->act_dev, f->sem[i], hwctx->alloc);
1962 for (int i = 0; i < desc->nb_objects; i++)
1963 vkFreeMemory(hwctx->act_dev, f->mem[i], hwctx->alloc);
1970 static int vulkan_map_from_drm(AVHWFramesContext *hwfc, AVFrame *dst,
1971 const AVFrame *src, int flags)
1975 VulkanMapping *map = NULL;
1977 err = vulkan_map_from_drm_frame_desc(hwfc, &f,
1978 (AVDRMFrameDescriptor *)src->data[0]);
1982 /* The unmapping function will free this */
1983 dst->data[0] = (uint8_t *)f;
1984 dst->width = src->width;
1985 dst->height = src->height;
1987 map = av_mallocz(sizeof(VulkanMapping));
1994 err = ff_hwframe_map_create(dst->hw_frames_ctx, dst, src,
1995 &vulkan_unmap_from, map);
1999 av_log(hwfc, AV_LOG_DEBUG, "Mapped DRM object to Vulkan!\n");
2004 vulkan_frame_free(hwfc->device_ctx->hwctx, (uint8_t *)f);
2010 static int vulkan_map_from_vaapi(AVHWFramesContext *dst_fc,
2011 AVFrame *dst, const AVFrame *src,
2015 AVFrame *tmp = av_frame_alloc();
2016 AVHWFramesContext *vaapi_fc = (AVHWFramesContext*)src->hw_frames_ctx->data;
2017 AVVAAPIDeviceContext *vaapi_ctx = vaapi_fc->device_ctx->hwctx;
2018 VASurfaceID surface_id = (VASurfaceID)(uintptr_t)src->data[3];
2021 return AVERROR(ENOMEM);
2023 /* We have to sync since like the previous comment said, no semaphores */
2024 vaSyncSurface(vaapi_ctx->display, surface_id);
2026 tmp->format = AV_PIX_FMT_DRM_PRIME;
2028 err = av_hwframe_map(tmp, src, flags);
2032 err = vulkan_map_from_drm(dst_fc, dst, tmp, flags);
2036 err = ff_hwframe_map_replace(dst, src);
2039 av_frame_free(&tmp);
2046 static int vulkan_export_to_cuda(AVHWFramesContext *hwfc,
2047 AVBufferRef *cuda_hwfc,
2048 const AVFrame *frame)
2053 AVVkFrameInternal *dst_int;
2054 AVHWDeviceContext *ctx = hwfc->device_ctx;
2055 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2056 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
2057 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
2058 VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdKHR);
2059 VK_LOAD_PFN(hwctx->inst, vkGetSemaphoreFdKHR);
2061 AVHWFramesContext *cuda_fc = (AVHWFramesContext*)cuda_hwfc->data;
2062 AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
2063 AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
2064 AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
2065 CudaFunctions *cu = cu_internal->cuda_dl;
2066 CUarray_format cufmt = desc->comp[0].depth > 8 ? CU_AD_FORMAT_UNSIGNED_INT16 :
2067 CU_AD_FORMAT_UNSIGNED_INT8;
2069 dst_f = (AVVkFrame *)frame->data[0];
2071 dst_int = dst_f->internal;
2072 if (!dst_int || !dst_int->cuda_fc_ref) {
2073 if (!dst_f->internal)
2074 dst_f->internal = dst_int = av_mallocz(sizeof(*dst_f->internal));
2077 err = AVERROR(ENOMEM);
2081 dst_int->cuda_fc_ref = av_buffer_ref(cuda_hwfc);
2082 if (!dst_int->cuda_fc_ref) {
2083 err = AVERROR(ENOMEM);
2087 for (int i = 0; i < planes; i++) {
2088 CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC tex_desc = {
2091 .Width = i > 0 ? AV_CEIL_RSHIFT(hwfc->width, desc->log2_chroma_w)
2093 .Height = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, desc->log2_chroma_h)
2097 .NumChannels = 1 + ((planes == 2) && i),
2102 CUDA_EXTERNAL_MEMORY_HANDLE_DESC ext_desc = {
2103 .type = CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD,
2104 .size = dst_f->size[i],
2106 VkMemoryGetFdInfoKHR export_info = {
2107 .sType = VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR,
2108 .memory = dst_f->mem[i],
2109 .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR,
2111 VkSemaphoreGetFdInfoKHR sem_export = {
2112 .sType = VK_STRUCTURE_TYPE_SEMAPHORE_GET_FD_INFO_KHR,
2113 .semaphore = dst_f->sem[i],
2114 .handleType = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT,
2116 CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC ext_sem_desc = {
2117 .type = CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD,
2120 ret = pfn_vkGetMemoryFdKHR(hwctx->act_dev, &export_info,
2121 &ext_desc.handle.fd);
2122 if (ret != VK_SUCCESS) {
2123 av_log(hwfc, AV_LOG_ERROR, "Unable to export the image as a FD!\n");
2124 err = AVERROR_EXTERNAL;
2128 ret = CHECK_CU(cu->cuImportExternalMemory(&dst_int->ext_mem[i], &ext_desc));
2130 err = AVERROR_EXTERNAL;
2134 ret = CHECK_CU(cu->cuExternalMemoryGetMappedMipmappedArray(&dst_int->cu_mma[i],
2135 dst_int->ext_mem[i],
2138 err = AVERROR_EXTERNAL;
2142 ret = CHECK_CU(cu->cuMipmappedArrayGetLevel(&dst_int->cu_array[i],
2143 dst_int->cu_mma[i], 0));
2145 err = AVERROR_EXTERNAL;
2149 ret = pfn_vkGetSemaphoreFdKHR(hwctx->act_dev, &sem_export,
2150 &ext_sem_desc.handle.fd);
2151 if (ret != VK_SUCCESS) {
2152 av_log(ctx, AV_LOG_ERROR, "Failed to export semaphore: %s\n",
2154 err = AVERROR_EXTERNAL;
2158 ret = CHECK_CU(cu->cuImportExternalSemaphore(&dst_int->cu_sem[i],
2161 err = AVERROR_EXTERNAL;
2173 static int vulkan_transfer_data_from_cuda(AVHWFramesContext *hwfc,
2174 AVFrame *dst, const AVFrame *src)
2180 AVVkFrameInternal *dst_int;
2181 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
2182 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
2184 AVHWFramesContext *cuda_fc = (AVHWFramesContext*)src->hw_frames_ctx->data;
2185 AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
2186 AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
2187 AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
2188 CudaFunctions *cu = cu_internal->cuda_dl;
2189 CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS s_w_par[AV_NUM_DATA_POINTERS] = { 0 };
2190 CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS s_s_par[AV_NUM_DATA_POINTERS] = { 0 };
2192 ret = CHECK_CU(cu->cuCtxPushCurrent(cuda_dev->cuda_ctx));
2194 err = AVERROR_EXTERNAL;
2198 dst_f = (AVVkFrame *)dst->data[0];
2200 ret = vulkan_export_to_cuda(hwfc, src->hw_frames_ctx, dst);
2204 dst_int = dst_f->internal;
2206 ret = CHECK_CU(cu->cuWaitExternalSemaphoresAsync(dst_int->cu_sem, s_w_par,
2207 planes, cuda_dev->stream));
2209 err = AVERROR_EXTERNAL;
2213 for (int i = 0; i < planes; i++) {
2214 CUDA_MEMCPY2D cpy = {
2215 .srcMemoryType = CU_MEMORYTYPE_DEVICE,
2216 .srcDevice = (CUdeviceptr)src->data[i],
2217 .srcPitch = src->linesize[i],
2220 .dstMemoryType = CU_MEMORYTYPE_ARRAY,
2221 .dstArray = dst_int->cu_array[i],
2222 .WidthInBytes = (i > 0 ? AV_CEIL_RSHIFT(hwfc->width, desc->log2_chroma_w)
2223 : hwfc->width) * desc->comp[i].step,
2224 .Height = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, desc->log2_chroma_h)
2228 ret = CHECK_CU(cu->cuMemcpy2DAsync(&cpy, cuda_dev->stream));
2230 err = AVERROR_EXTERNAL;
2235 ret = CHECK_CU(cu->cuSignalExternalSemaphoresAsync(dst_int->cu_sem, s_s_par,
2236 planes, cuda_dev->stream));
2238 err = AVERROR_EXTERNAL;
2242 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2244 av_log(hwfc, AV_LOG_VERBOSE, "Transfered CUDA image to Vulkan!\n");
2249 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2250 vulkan_free_internal(dst_int);
2251 dst_f->internal = NULL;
2252 av_buffer_unref(&dst->buf[0]);
2257 static int vulkan_map_to(AVHWFramesContext *hwfc, AVFrame *dst,
2258 const AVFrame *src, int flags)
2260 av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2262 switch (src->format) {
2265 case AV_PIX_FMT_VAAPI:
2266 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2267 return vulkan_map_from_vaapi(hwfc, dst, src, flags);
2269 case AV_PIX_FMT_DRM_PRIME:
2270 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2271 return vulkan_map_from_drm(hwfc, dst, src, flags);
2274 return AVERROR(ENOSYS);
2279 typedef struct VulkanDRMMapping {
2280 AVDRMFrameDescriptor drm_desc;
2284 static void vulkan_unmap_to_drm(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
2286 AVDRMFrameDescriptor *drm_desc = hwmap->priv;
2288 for (int i = 0; i < drm_desc->nb_objects; i++)
2289 close(drm_desc->objects[i].fd);
2294 static inline uint32_t vulkan_fmt_to_drm(VkFormat vkfmt)
2296 for (int i = 0; i < FF_ARRAY_ELEMS(vulkan_drm_format_map); i++)
2297 if (vulkan_drm_format_map[i].vk_format == vkfmt)
2298 return vulkan_drm_format_map[i].drm_fourcc;
2299 return DRM_FORMAT_INVALID;
2302 static int vulkan_map_to_drm(AVHWFramesContext *hwfc, AVFrame *dst,
2303 const AVFrame *src, int flags)
2307 AVVkFrame *f = (AVVkFrame *)src->data[0];
2308 VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2309 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
2310 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
2311 VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdKHR);
2312 VkImageDrmFormatModifierPropertiesEXT drm_mod = {
2313 .sType = VK_STRUCTURE_TYPE_IMAGE_DRM_FORMAT_MODIFIER_PROPERTIES_EXT,
2316 AVDRMFrameDescriptor *drm_desc = av_mallocz(sizeof(*drm_desc));
2318 return AVERROR(ENOMEM);
2320 err = prepare_frame(hwfc, &p->cmd, f, PREP_MODE_EXTERNAL_EXPORT);
2324 err = ff_hwframe_map_create(src->hw_frames_ctx, dst, src, &vulkan_unmap_to_drm, drm_desc);
2328 if (p->extensions & EXT_DRM_MODIFIER_FLAGS) {
2329 VK_LOAD_PFN(hwctx->inst, vkGetImageDrmFormatModifierPropertiesEXT);
2330 ret = pfn_vkGetImageDrmFormatModifierPropertiesEXT(hwctx->act_dev, f->img[0],
2332 if (ret != VK_SUCCESS) {
2333 av_log(hwfc, AV_LOG_ERROR, "Failed to retrieve DRM format modifier!\n");
2334 err = AVERROR_EXTERNAL;
2339 for (int i = 0; (i < planes) && (f->mem[i]); i++) {
2340 VkMemoryGetFdInfoKHR export_info = {
2341 .sType = VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR,
2342 .memory = f->mem[i],
2343 .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT,
2346 ret = pfn_vkGetMemoryFdKHR(hwctx->act_dev, &export_info,
2347 &drm_desc->objects[i].fd);
2348 if (ret != VK_SUCCESS) {
2349 av_log(hwfc, AV_LOG_ERROR, "Unable to export the image as a FD!\n");
2350 err = AVERROR_EXTERNAL;
2354 drm_desc->nb_objects++;
2355 drm_desc->objects[i].size = f->size[i];
2356 drm_desc->objects[i].format_modifier = drm_mod.drmFormatModifier;
2359 drm_desc->nb_layers = planes;
2360 for (int i = 0; i < drm_desc->nb_layers; i++) {
2361 VkSubresourceLayout layout;
2362 VkImageSubresource sub = {
2363 .aspectMask = p->extensions & EXT_DRM_MODIFIER_FLAGS ?
2364 VK_IMAGE_ASPECT_MEMORY_PLANE_0_BIT_EXT :
2365 VK_IMAGE_ASPECT_COLOR_BIT,
2367 VkFormat plane_vkfmt = av_vkfmt_from_pixfmt(hwfc->sw_format)[i];
2369 drm_desc->layers[i].format = vulkan_fmt_to_drm(plane_vkfmt);
2370 drm_desc->layers[i].nb_planes = 1;
2372 if (drm_desc->layers[i].format == DRM_FORMAT_INVALID) {
2373 av_log(hwfc, AV_LOG_ERROR, "Cannot map to DRM layer, unsupported!\n");
2374 err = AVERROR_PATCHWELCOME;
2378 drm_desc->layers[i].planes[0].object_index = FFMIN(i, drm_desc->nb_objects - 1);
2380 if (f->tiling == VK_IMAGE_TILING_OPTIMAL)
2383 vkGetImageSubresourceLayout(hwctx->act_dev, f->img[i], &sub, &layout);
2384 drm_desc->layers[i].planes[0].offset = layout.offset;
2385 drm_desc->layers[i].planes[0].pitch = layout.rowPitch;
2388 dst->width = src->width;
2389 dst->height = src->height;
2390 dst->data[0] = (uint8_t *)drm_desc;
2392 av_log(hwfc, AV_LOG_VERBOSE, "Mapped AVVkFrame to a DRM object!\n");
2402 static int vulkan_map_to_vaapi(AVHWFramesContext *hwfc, AVFrame *dst,
2403 const AVFrame *src, int flags)
2406 AVFrame *tmp = av_frame_alloc();
2408 return AVERROR(ENOMEM);
2410 tmp->format = AV_PIX_FMT_DRM_PRIME;
2412 err = vulkan_map_to_drm(hwfc, tmp, src, flags);
2416 err = av_hwframe_map(dst, tmp, flags);
2420 err = ff_hwframe_map_replace(dst, src);
2423 av_frame_free(&tmp);
2429 static int vulkan_map_from(AVHWFramesContext *hwfc, AVFrame *dst,
2430 const AVFrame *src, int flags)
2432 av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2434 switch (dst->format) {
2436 case AV_PIX_FMT_DRM_PRIME:
2437 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2438 return vulkan_map_to_drm(hwfc, dst, src, flags);
2440 case AV_PIX_FMT_VAAPI:
2441 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2442 return vulkan_map_to_vaapi(hwfc, dst, src, flags);
2446 return vulkan_map_frame_to_mem(hwfc, dst, src, flags);
2450 typedef struct ImageBuffer {
2453 VkMemoryPropertyFlagBits flags;
2456 static void free_buf(AVHWDeviceContext *ctx, ImageBuffer *buf)
2458 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2462 vkDestroyBuffer(hwctx->act_dev, buf->buf, hwctx->alloc);
2463 vkFreeMemory(hwctx->act_dev, buf->mem, hwctx->alloc);
2466 static int create_buf(AVHWDeviceContext *ctx, ImageBuffer *buf, int height,
2467 int *stride, VkBufferUsageFlags usage,
2468 VkMemoryPropertyFlagBits flags, void *create_pnext,
2473 VkMemoryRequirements req;
2474 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2475 VulkanDevicePriv *p = ctx->internal->priv;
2477 VkBufferCreateInfo buf_spawn = {
2478 .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
2479 .pNext = create_pnext,
2481 .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
2484 *stride = FFALIGN(*stride, p->props.limits.optimalBufferCopyRowPitchAlignment);
2485 buf_spawn.size = height*(*stride);
2487 ret = vkCreateBuffer(hwctx->act_dev, &buf_spawn, NULL, &buf->buf);
2488 if (ret != VK_SUCCESS) {
2489 av_log(ctx, AV_LOG_ERROR, "Failed to create buffer: %s\n",
2491 return AVERROR_EXTERNAL;
2494 vkGetBufferMemoryRequirements(hwctx->act_dev, buf->buf, &req);
2496 err = alloc_mem(ctx, &req, flags, alloc_pnext, &buf->flags, &buf->mem);
2500 ret = vkBindBufferMemory(hwctx->act_dev, buf->buf, buf->mem, 0);
2501 if (ret != VK_SUCCESS) {
2502 av_log(ctx, AV_LOG_ERROR, "Failed to bind memory to buffer: %s\n",
2505 return AVERROR_EXTERNAL;
2511 static int map_buffers(AVHWDeviceContext *ctx, ImageBuffer *buf, uint8_t *mem[],
2512 int nb_buffers, int invalidate)
2515 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2516 VkMappedMemoryRange invalidate_ctx[AV_NUM_DATA_POINTERS];
2517 int invalidate_count = 0;
2519 for (int i = 0; i < nb_buffers; i++) {
2520 ret = vkMapMemory(hwctx->act_dev, buf[i].mem, 0,
2521 VK_WHOLE_SIZE, 0, (void **)&mem[i]);
2522 if (ret != VK_SUCCESS) {
2523 av_log(ctx, AV_LOG_ERROR, "Failed to map buffer memory: %s\n",
2525 return AVERROR_EXTERNAL;
2532 for (int i = 0; i < nb_buffers; i++) {
2533 const VkMappedMemoryRange ival_buf = {
2534 .sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE,
2535 .memory = buf[i].mem,
2536 .size = VK_WHOLE_SIZE,
2538 if (buf[i].flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)
2540 invalidate_ctx[invalidate_count++] = ival_buf;
2543 if (invalidate_count) {
2544 ret = vkInvalidateMappedMemoryRanges(hwctx->act_dev, invalidate_count,
2546 if (ret != VK_SUCCESS)
2547 av_log(ctx, AV_LOG_WARNING, "Failed to invalidate memory: %s\n",
2554 static int unmap_buffers(AVHWDeviceContext *ctx, ImageBuffer *buf,
2555 int nb_buffers, int flush)
2559 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2560 VkMappedMemoryRange flush_ctx[AV_NUM_DATA_POINTERS];
2561 int flush_count = 0;
2564 for (int i = 0; i < nb_buffers; i++) {
2565 const VkMappedMemoryRange flush_buf = {
2566 .sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE,
2567 .memory = buf[i].mem,
2568 .size = VK_WHOLE_SIZE,
2570 if (buf[i].flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)
2572 flush_ctx[flush_count++] = flush_buf;
2577 ret = vkFlushMappedMemoryRanges(hwctx->act_dev, flush_count, flush_ctx);
2578 if (ret != VK_SUCCESS) {
2579 av_log(ctx, AV_LOG_ERROR, "Failed to flush memory: %s\n",
2581 err = AVERROR_EXTERNAL; /* We still want to try to unmap them */
2585 for (int i = 0; i < nb_buffers; i++)
2586 vkUnmapMemory(hwctx->act_dev, buf[i].mem);
2591 static int transfer_image_buf(AVHWDeviceContext *ctx, AVVkFrame *frame,
2592 ImageBuffer *buffer, const int *buf_stride, int w,
2593 int h, enum AVPixelFormat pix_fmt, int to_buf)
2596 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2597 VulkanDevicePriv *s = ctx->internal->priv;
2600 VkPipelineStageFlagBits sem_wait_dst[AV_NUM_DATA_POINTERS];
2602 const int planes = av_pix_fmt_count_planes(pix_fmt);
2603 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(pix_fmt);
2605 VkCommandBufferBeginInfo cmd_start = {
2606 .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
2607 .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
2610 VkImageMemoryBarrier img_bar[AV_NUM_DATA_POINTERS] = { 0 };
2612 VkSubmitInfo s_info = {
2613 .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
2614 .commandBufferCount = 1,
2615 .pCommandBuffers = &s->cmd.buf,
2616 .pSignalSemaphores = frame->sem,
2617 .pWaitSemaphores = frame->sem,
2618 .pWaitDstStageMask = sem_wait_dst,
2619 .signalSemaphoreCount = planes,
2620 .waitSemaphoreCount = planes,
2623 ret = vkBeginCommandBuffer(s->cmd.buf, &cmd_start);
2624 if (ret != VK_SUCCESS) {
2625 av_log(ctx, AV_LOG_ERROR, "Unable to init command buffer: %s\n",
2627 return AVERROR_EXTERNAL;
2630 /* Change the image layout to something more optimal for transfers */
2631 for (int i = 0; i < planes; i++) {
2632 VkImageLayout new_layout = to_buf ? VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL :
2633 VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
2634 VkAccessFlags new_access = to_buf ? VK_ACCESS_TRANSFER_READ_BIT :
2635 VK_ACCESS_TRANSFER_WRITE_BIT;
2637 sem_wait_dst[i] = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
2639 /* If the layout matches and we have read access skip the barrier */
2640 if ((frame->layout[i] == new_layout) && (frame->access[i] & new_access))
2643 img_bar[bar_num].sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
2644 img_bar[bar_num].srcAccessMask = 0x0;
2645 img_bar[bar_num].dstAccessMask = new_access;
2646 img_bar[bar_num].oldLayout = frame->layout[i];
2647 img_bar[bar_num].newLayout = new_layout;
2648 img_bar[bar_num].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
2649 img_bar[bar_num].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
2650 img_bar[bar_num].image = frame->img[i];
2651 img_bar[bar_num].subresourceRange.levelCount = 1;
2652 img_bar[bar_num].subresourceRange.layerCount = 1;
2653 img_bar[bar_num].subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
2655 frame->layout[i] = img_bar[bar_num].newLayout;
2656 frame->access[i] = img_bar[bar_num].dstAccessMask;
2662 vkCmdPipelineBarrier(s->cmd.buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
2663 VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
2664 0, NULL, 0, NULL, bar_num, img_bar);
2666 /* Schedule a copy for each plane */
2667 for (int i = 0; i < planes; i++) {
2668 const int p_w = i > 0 ? AV_CEIL_RSHIFT(w, desc->log2_chroma_w) : w;
2669 const int p_h = i > 0 ? AV_CEIL_RSHIFT(h, desc->log2_chroma_h) : h;
2670 VkBufferImageCopy buf_reg = {
2672 /* Buffer stride isn't in bytes, it's in samples, the implementation
2673 * uses the image's VkFormat to know how many bytes per sample
2674 * the buffer has. So we have to convert by dividing. Stupid.
2675 * Won't work with YUVA or other planar formats with alpha. */
2676 .bufferRowLength = buf_stride[i] / desc->comp[i].step,
2677 .bufferImageHeight = p_h,
2678 .imageSubresource.layerCount = 1,
2679 .imageSubresource.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
2680 .imageOffset = { 0, 0, 0, },
2681 .imageExtent = { p_w, p_h, 1, },
2685 vkCmdCopyImageToBuffer(s->cmd.buf, frame->img[i], frame->layout[i],
2686 buffer[i].buf, 1, &buf_reg);
2688 vkCmdCopyBufferToImage(s->cmd.buf, buffer[i].buf, frame->img[i],
2689 frame->layout[i], 1, &buf_reg);
2692 ret = vkEndCommandBuffer(s->cmd.buf);
2693 if (ret != VK_SUCCESS) {
2694 av_log(ctx, AV_LOG_ERROR, "Unable to finish command buffer: %s\n",
2696 return AVERROR_EXTERNAL;
2699 /* Wait for the download/upload to finish if uploading, otherwise the
2700 * semaphore will take care of synchronization when uploading */
2701 ret = vkQueueSubmit(s->cmd.queue, 1, &s_info, s->cmd.fence);
2702 if (ret != VK_SUCCESS) {
2703 av_log(ctx, AV_LOG_ERROR, "Unable to submit command buffer: %s\n",
2705 return AVERROR_EXTERNAL;
2707 vkWaitForFences(hwctx->act_dev, 1, &s->cmd.fence, VK_TRUE, UINT64_MAX);
2708 vkResetFences(hwctx->act_dev, 1, &s->cmd.fence);
2714 /* Technically we can use VK_EXT_external_memory_host to upload and download,
2715 * however the alignment requirements make this unfeasible as both the pointer
2716 * and the size of each plane need to be aligned to the minimum alignment
2717 * requirement, which on all current implementations (anv, radv) is 4096.
2718 * If the requirement gets relaxed (unlikely) this can easily be implemented. */
2719 static int vulkan_transfer_data_from_mem(AVHWFramesContext *hwfc, AVFrame *dst,
2724 AVVkFrame *f = (AVVkFrame *)dst->data[0];
2725 AVHWDeviceContext *dev_ctx = hwfc->device_ctx;
2726 ImageBuffer buf[AV_NUM_DATA_POINTERS] = { { 0 } };
2727 const int planes = av_pix_fmt_count_planes(src->format);
2728 int log2_chroma = av_pix_fmt_desc_get(src->format)->log2_chroma_h;
2730 if ((src->format != AV_PIX_FMT_NONE && !av_vkfmt_from_pixfmt(src->format))) {
2731 av_log(hwfc, AV_LOG_ERROR, "Unsupported source pixel format!\n");
2732 return AVERROR(EINVAL);
2735 if (src->width > hwfc->width || src->height > hwfc->height)
2736 return AVERROR(EINVAL);
2738 /* For linear, host visiable images */
2739 if (f->tiling == VK_IMAGE_TILING_LINEAR &&
2740 f->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) {
2741 AVFrame *map = av_frame_alloc();
2743 return AVERROR(ENOMEM);
2744 map->format = src->format;
2746 err = vulkan_map_frame_to_mem(hwfc, map, dst, AV_HWFRAME_MAP_WRITE);
2750 err = av_frame_copy(map, src);
2751 av_frame_free(&map);
2755 /* Create buffers */
2756 for (int i = 0; i < planes; i++) {
2757 int h = src->height;
2758 int p_height = i > 0 ? AV_CEIL_RSHIFT(h, log2_chroma) : h;
2760 tmp.linesize[i] = FFABS(src->linesize[i]);
2761 err = create_buf(dev_ctx, &buf[i], p_height,
2762 &tmp.linesize[i], VK_BUFFER_USAGE_TRANSFER_SRC_BIT,
2763 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT, NULL, NULL);
2768 /* Map, copy image to buffer, unmap */
2769 if ((err = map_buffers(dev_ctx, buf, tmp.data, planes, 0)))
2772 av_image_copy(tmp.data, tmp.linesize, (const uint8_t **)src->data,
2773 src->linesize, src->format, src->width, src->height);
2775 if ((err = unmap_buffers(dev_ctx, buf, planes, 1)))
2778 /* Copy buffers to image */
2779 err = transfer_image_buf(dev_ctx, f, buf, tmp.linesize,
2780 src->width, src->height, src->format, 0);
2783 for (int i = 0; i < planes; i++)
2784 free_buf(dev_ctx, &buf[i]);
2789 static int vulkan_transfer_data_to(AVHWFramesContext *hwfc, AVFrame *dst,
2792 av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2794 switch (src->format) {
2796 case AV_PIX_FMT_CUDA:
2797 if ((p->extensions & EXT_EXTERNAL_FD_MEMORY) &&
2798 (p->extensions & EXT_EXTERNAL_FD_SEM))
2799 return vulkan_transfer_data_from_cuda(hwfc, dst, src);
2802 if (src->hw_frames_ctx)
2803 return AVERROR(ENOSYS);
2805 return vulkan_transfer_data_from_mem(hwfc, dst, src);
2810 static int vulkan_transfer_data_to_cuda(AVHWFramesContext *hwfc, AVFrame *dst,
2817 AVVkFrameInternal *dst_int;
2818 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
2819 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
2821 AVHWFramesContext *cuda_fc = (AVHWFramesContext*)dst->hw_frames_ctx->data;
2822 AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
2823 AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
2824 AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
2825 CudaFunctions *cu = cu_internal->cuda_dl;
2827 ret = CHECK_CU(cu->cuCtxPushCurrent(cuda_dev->cuda_ctx));
2829 err = AVERROR_EXTERNAL;
2833 dst_f = (AVVkFrame *)src->data[0];
2835 err = vulkan_export_to_cuda(hwfc, dst->hw_frames_ctx, src);
2840 dst_int = dst_f->internal;
2842 for (int i = 0; i < planes; i++) {
2843 CUDA_MEMCPY2D cpy = {
2844 .dstMemoryType = CU_MEMORYTYPE_DEVICE,
2845 .dstDevice = (CUdeviceptr)dst->data[i],
2846 .dstPitch = dst->linesize[i],
2849 .srcMemoryType = CU_MEMORYTYPE_ARRAY,
2850 .srcArray = dst_int->cu_array[i],
2851 .WidthInBytes = (i > 0 ? AV_CEIL_RSHIFT(hwfc->width, desc->log2_chroma_w)
2852 : hwfc->width) * desc->comp[i].step,
2853 .Height = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, desc->log2_chroma_h)
2857 ret = CHECK_CU(cu->cuMemcpy2DAsync(&cpy, cuda_dev->stream));
2859 err = AVERROR_EXTERNAL;
2864 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2866 av_log(hwfc, AV_LOG_VERBOSE, "Transfered Vulkan image to CUDA!\n");
2871 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2872 vulkan_free_internal(dst_int);
2873 dst_f->internal = NULL;
2874 av_buffer_unref(&dst->buf[0]);
2879 static int vulkan_transfer_data_to_mem(AVHWFramesContext *hwfc, AVFrame *dst,
2884 AVVkFrame *f = (AVVkFrame *)src->data[0];
2885 AVHWDeviceContext *dev_ctx = hwfc->device_ctx;
2886 ImageBuffer buf[AV_NUM_DATA_POINTERS] = { { 0 } };
2887 const int planes = av_pix_fmt_count_planes(dst->format);
2888 int log2_chroma = av_pix_fmt_desc_get(dst->format)->log2_chroma_h;
2890 if (dst->width > hwfc->width || dst->height > hwfc->height)
2891 return AVERROR(EINVAL);
2893 /* For linear, host visiable images */
2894 if (f->tiling == VK_IMAGE_TILING_LINEAR &&
2895 f->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) {
2896 AVFrame *map = av_frame_alloc();
2898 return AVERROR(ENOMEM);
2899 map->format = dst->format;
2901 err = vulkan_map_frame_to_mem(hwfc, map, src, AV_HWFRAME_MAP_READ);
2905 err = av_frame_copy(dst, map);
2906 av_frame_free(&map);
2910 /* Create buffers */
2911 for (int i = 0; i < planes; i++) {
2912 int h = dst->height;
2913 int p_height = i > 0 ? AV_CEIL_RSHIFT(h, log2_chroma) : h;
2915 tmp.linesize[i] = FFABS(dst->linesize[i]);
2916 err = create_buf(dev_ctx, &buf[i], p_height,
2917 &tmp.linesize[i], VK_BUFFER_USAGE_TRANSFER_DST_BIT,
2918 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT, NULL, NULL);
2921 /* Copy image to buffer */
2922 if ((err = transfer_image_buf(dev_ctx, f, buf, tmp.linesize,
2923 dst->width, dst->height, dst->format, 1)))
2926 /* Map, copy buffer to frame, unmap */
2927 if ((err = map_buffers(dev_ctx, buf, tmp.data, planes, 1)))
2930 av_image_copy(dst->data, dst->linesize, (const uint8_t **)tmp.data,
2931 tmp.linesize, dst->format, dst->width, dst->height);
2933 err = unmap_buffers(dev_ctx, buf, planes, 0);
2936 for (int i = 0; i < planes; i++)
2937 free_buf(dev_ctx, &buf[i]);
2942 static int vulkan_transfer_data_from(AVHWFramesContext *hwfc, AVFrame *dst,
2945 av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2947 switch (dst->format) {
2949 case AV_PIX_FMT_CUDA:
2950 if ((p->extensions & EXT_EXTERNAL_FD_MEMORY) &&
2951 (p->extensions & EXT_EXTERNAL_FD_SEM))
2952 return vulkan_transfer_data_to_cuda(hwfc, dst, src);
2955 if (dst->hw_frames_ctx)
2956 return AVERROR(ENOSYS);
2958 return vulkan_transfer_data_to_mem(hwfc, dst, src);
2962 AVVkFrame *av_vk_frame_alloc(void)
2964 return av_mallocz(sizeof(AVVkFrame));
2967 const HWContextType ff_hwcontext_type_vulkan = {
2968 .type = AV_HWDEVICE_TYPE_VULKAN,
2971 .device_hwctx_size = sizeof(AVVulkanDeviceContext),
2972 .device_priv_size = sizeof(VulkanDevicePriv),
2973 .frames_hwctx_size = sizeof(AVVulkanFramesContext),
2974 .frames_priv_size = sizeof(VulkanFramesPriv),
2976 .device_init = &vulkan_device_init,
2977 .device_create = &vulkan_device_create,
2978 .device_derive = &vulkan_device_derive,
2980 .frames_get_constraints = &vulkan_frames_get_constraints,
2981 .frames_init = vulkan_frames_init,
2982 .frames_get_buffer = vulkan_get_buffer,
2983 .frames_uninit = vulkan_frames_uninit,
2985 .transfer_get_formats = vulkan_transfer_get_formats,
2986 .transfer_data_to = vulkan_transfer_data_to,
2987 .transfer_data_from = vulkan_transfer_data_from,
2989 .map_to = vulkan_map_to,
2990 .map_from = vulkan_map_from,
2992 .pix_fmts = (const enum AVPixelFormat []) {