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)
622 VkQueueFamilyProperties *qs = NULL;
623 AVVulkanDeviceContext *hwctx = ctx->hwctx;
624 int graph_index = -1, comp_index = -1, tx_index = -1;
625 VkDeviceQueueCreateInfo *pc = (VkDeviceQueueCreateInfo *)cd->pQueueCreateInfos;
627 /* First get the number of queue families */
628 vkGetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &num, NULL);
630 av_log(ctx, AV_LOG_ERROR, "Failed to get queues!\n");
631 return AVERROR_EXTERNAL;
634 /* Then allocate memory */
635 qs = av_malloc_array(num, sizeof(VkQueueFamilyProperties));
637 return AVERROR(ENOMEM);
639 /* Finally retrieve the queue families */
640 vkGetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &num, qs);
642 #define SEARCH_FLAGS(expr, out) \
643 for (int i = 0; i < num; i++) { \
644 const VkQueueFlagBits flags = qs[i].queueFlags; \
651 SEARCH_FLAGS(flags & VK_QUEUE_GRAPHICS_BIT, graph_index)
653 SEARCH_FLAGS((flags & VK_QUEUE_COMPUTE_BIT) && (i != graph_index),
656 SEARCH_FLAGS((flags & VK_QUEUE_TRANSFER_BIT) && (i != graph_index) &&
657 (i != comp_index), tx_index)
660 #define QF_FLAGS(flags) \
661 ((flags) & VK_QUEUE_GRAPHICS_BIT ) ? "(graphics) " : "", \
662 ((flags) & VK_QUEUE_COMPUTE_BIT ) ? "(compute) " : "", \
663 ((flags) & VK_QUEUE_TRANSFER_BIT ) ? "(transfer) " : "", \
664 ((flags) & VK_QUEUE_SPARSE_BINDING_BIT) ? "(sparse) " : ""
666 av_log(ctx, AV_LOG_VERBOSE, "Using queue family %i for graphics, "
667 "flags: %s%s%s%s\n", graph_index, QF_FLAGS(qs[graph_index].queueFlags));
669 hwctx->queue_family_index = graph_index;
670 hwctx->queue_family_tx_index = graph_index;
671 hwctx->queue_family_comp_index = graph_index;
673 pc[cd->queueCreateInfoCount++].queueFamilyIndex = graph_index;
675 if (comp_index != -1) {
676 av_log(ctx, AV_LOG_VERBOSE, "Using queue family %i for compute, "
677 "flags: %s%s%s%s\n", comp_index, QF_FLAGS(qs[comp_index].queueFlags));
678 hwctx->queue_family_tx_index = comp_index;
679 hwctx->queue_family_comp_index = comp_index;
680 pc[cd->queueCreateInfoCount++].queueFamilyIndex = comp_index;
683 if (tx_index != -1) {
684 av_log(ctx, AV_LOG_VERBOSE, "Using queue family %i for transfers, "
685 "flags: %s%s%s%s\n", tx_index, QF_FLAGS(qs[tx_index].queueFlags));
686 hwctx->queue_family_tx_index = tx_index;
687 pc[cd->queueCreateInfoCount++].queueFamilyIndex = tx_index;
697 static int create_exec_ctx(AVHWDeviceContext *ctx, VulkanExecCtx *cmd,
698 int queue_family_index)
701 AVVulkanDeviceContext *hwctx = ctx->hwctx;
703 VkCommandPoolCreateInfo cqueue_create = {
704 .sType = VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO,
705 .flags = VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT,
706 .queueFamilyIndex = queue_family_index,
708 VkCommandBufferAllocateInfo cbuf_create = {
709 .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO,
710 .level = VK_COMMAND_BUFFER_LEVEL_PRIMARY,
711 .commandBufferCount = 1,
714 VkFenceCreateInfo fence_spawn = {
715 .sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO,
718 ret = vkCreateFence(hwctx->act_dev, &fence_spawn,
719 hwctx->alloc, &cmd->fence);
720 if (ret != VK_SUCCESS) {
721 av_log(ctx, AV_LOG_ERROR, "Failed to create frame fence: %s\n",
723 return AVERROR_EXTERNAL;
726 ret = vkCreateCommandPool(hwctx->act_dev, &cqueue_create,
727 hwctx->alloc, &cmd->pool);
728 if (ret != VK_SUCCESS) {
729 av_log(ctx, AV_LOG_ERROR, "Command pool creation failure: %s\n",
731 return AVERROR_EXTERNAL;
734 cbuf_create.commandPool = cmd->pool;
736 ret = vkAllocateCommandBuffers(hwctx->act_dev, &cbuf_create, &cmd->buf);
737 if (ret != VK_SUCCESS) {
738 av_log(ctx, AV_LOG_ERROR, "Command buffer alloc failure: %s\n",
740 return AVERROR_EXTERNAL;
743 vkGetDeviceQueue(hwctx->act_dev, cqueue_create.queueFamilyIndex, 0,
749 static void free_exec_ctx(AVHWDeviceContext *ctx, VulkanExecCtx *cmd)
751 AVVulkanDeviceContext *hwctx = ctx->hwctx;
754 vkDestroyFence(hwctx->act_dev, cmd->fence, hwctx->alloc);
756 vkFreeCommandBuffers(hwctx->act_dev, cmd->pool, 1, &cmd->buf);
758 vkDestroyCommandPool(hwctx->act_dev, cmd->pool, hwctx->alloc);
761 static void vulkan_device_free(AVHWDeviceContext *ctx)
763 VulkanDevicePriv *p = ctx->internal->priv;
764 AVVulkanDeviceContext *hwctx = ctx->hwctx;
766 free_exec_ctx(ctx, &p->cmd);
768 vkDestroyDevice(hwctx->act_dev, hwctx->alloc);
771 VK_LOAD_PFN(hwctx->inst, vkDestroyDebugUtilsMessengerEXT);
772 pfn_vkDestroyDebugUtilsMessengerEXT(hwctx->inst, p->debug_ctx,
776 vkDestroyInstance(hwctx->inst, hwctx->alloc);
778 for (int i = 0; i < hwctx->nb_enabled_inst_extensions; i++)
779 av_free((void *)hwctx->enabled_inst_extensions[i]);
780 av_free((void *)hwctx->enabled_inst_extensions);
782 for (int i = 0; i < hwctx->nb_enabled_dev_extensions; i++)
783 av_free((void *)hwctx->enabled_dev_extensions[i]);
784 av_free((void *)hwctx->enabled_dev_extensions);
787 static int vulkan_device_create_internal(AVHWDeviceContext *ctx,
788 VulkanDeviceSelection *dev_select,
789 AVDictionary *opts, int flags)
793 AVDictionaryEntry *opt_d;
794 VulkanDevicePriv *p = ctx->internal->priv;
795 AVVulkanDeviceContext *hwctx = ctx->hwctx;
796 VkDeviceQueueCreateInfo queue_create_info[3] = {
797 { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
798 .pQueuePriorities = (float []){ 1.0f },
800 { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
801 .pQueuePriorities = (float []){ 1.0f },
803 { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
804 .pQueuePriorities = (float []){ 1.0f },
808 VkDeviceCreateInfo dev_info = {
809 .sType = VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO,
810 .pQueueCreateInfos = queue_create_info,
811 .queueCreateInfoCount = 0,
814 ctx->free = vulkan_device_free;
816 /* Create an instance if not given one */
817 if ((err = create_instance(ctx, opts)))
820 /* Find a device (if not given one) */
821 if ((err = find_device(ctx, dev_select)))
824 vkGetPhysicalDeviceProperties(hwctx->phys_dev, &p->props);
825 av_log(ctx, AV_LOG_VERBOSE, "Using device: %s\n", p->props.deviceName);
826 av_log(ctx, AV_LOG_VERBOSE, "Alignments:\n");
827 av_log(ctx, AV_LOG_VERBOSE, " optimalBufferCopyOffsetAlignment: %li\n",
828 p->props.limits.optimalBufferCopyOffsetAlignment);
829 av_log(ctx, AV_LOG_VERBOSE, " optimalBufferCopyRowPitchAlignment: %li\n",
830 p->props.limits.optimalBufferCopyRowPitchAlignment);
831 av_log(ctx, AV_LOG_VERBOSE, " minMemoryMapAlignment: %li\n",
832 p->props.limits.minMemoryMapAlignment);
834 /* Search queue family */
835 if ((err = search_queue_families(ctx, &dev_info)))
838 if ((err = check_extensions(ctx, 1, opts, &dev_info.ppEnabledExtensionNames,
839 &dev_info.enabledExtensionCount, 0)))
842 ret = vkCreateDevice(hwctx->phys_dev, &dev_info, hwctx->alloc,
845 if (ret != VK_SUCCESS) {
846 av_log(ctx, AV_LOG_ERROR, "Device creation failure: %s\n",
848 for (int i = 0; i < dev_info.enabledExtensionCount; i++)
849 av_free((void *)dev_info.ppEnabledExtensionNames[i]);
850 av_free((void *)dev_info.ppEnabledExtensionNames);
851 err = AVERROR_EXTERNAL;
855 /* Tiled images setting, use them by default */
856 opt_d = av_dict_get(opts, "linear_images", NULL, 0);
858 p->use_linear_images = strtol(opt_d->value, NULL, 10);
860 hwctx->enabled_dev_extensions = dev_info.ppEnabledExtensionNames;
861 hwctx->nb_enabled_dev_extensions = dev_info.enabledExtensionCount;
867 static int vulkan_device_init(AVHWDeviceContext *ctx)
871 AVVulkanDeviceContext *hwctx = ctx->hwctx;
872 VulkanDevicePriv *p = ctx->internal->priv;
874 /* Set device extension flags */
875 for (int i = 0; i < hwctx->nb_enabled_dev_extensions; i++) {
876 for (int j = 0; j < FF_ARRAY_ELEMS(optional_device_exts); j++) {
877 if (!strcmp(hwctx->enabled_dev_extensions[i],
878 optional_device_exts[j].name)) {
879 p->extensions |= optional_device_exts[j].flag;
885 vkGetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &queue_num, NULL);
887 av_log(ctx, AV_LOG_ERROR, "Failed to get queues!\n");
888 return AVERROR_EXTERNAL;
891 #define CHECK_QUEUE(type, n) \
892 if (n >= queue_num) { \
893 av_log(ctx, AV_LOG_ERROR, "Invalid %s queue index %i (device has %i queues)!\n", \
894 type, n, queue_num); \
895 return AVERROR(EINVAL); \
898 CHECK_QUEUE("graphics", hwctx->queue_family_index)
899 CHECK_QUEUE("upload", hwctx->queue_family_tx_index)
900 CHECK_QUEUE("compute", hwctx->queue_family_comp_index)
904 p->qfs[p->num_qfs++] = hwctx->queue_family_index;
905 if ((hwctx->queue_family_tx_index != hwctx->queue_family_index) &&
906 (hwctx->queue_family_tx_index != hwctx->queue_family_comp_index))
907 p->qfs[p->num_qfs++] = hwctx->queue_family_tx_index;
908 if ((hwctx->queue_family_comp_index != hwctx->queue_family_index) &&
909 (hwctx->queue_family_comp_index != hwctx->queue_family_tx_index))
910 p->qfs[p->num_qfs++] = hwctx->queue_family_comp_index;
912 /* Create exec context - if there's something invalid this will error out */
913 err = create_exec_ctx(ctx, &p->cmd, hwctx->queue_family_tx_index);
917 /* Get device capabilities */
918 vkGetPhysicalDeviceMemoryProperties(hwctx->phys_dev, &p->mprops);
923 static int vulkan_device_create(AVHWDeviceContext *ctx, const char *device,
924 AVDictionary *opts, int flags)
926 VulkanDeviceSelection dev_select = { 0 };
927 if (device && device[0]) {
929 dev_select.index = strtol(device, &end, 10);
931 dev_select.index = 0;
932 dev_select.name = device;
936 return vulkan_device_create_internal(ctx, &dev_select, opts, flags);
939 static int vulkan_device_derive(AVHWDeviceContext *ctx,
940 AVHWDeviceContext *src_ctx, int flags)
942 av_unused VulkanDeviceSelection dev_select = { 0 };
944 /* If there's only one device on the system, then even if its not covered
945 * by the following checks (e.g. non-PCIe ARM GPU), having an empty
946 * dev_select will mean it'll get picked. */
947 switch(src_ctx->type) {
950 case AV_HWDEVICE_TYPE_VAAPI: {
951 AVVAAPIDeviceContext *src_hwctx = src_ctx->hwctx;
953 const char *vendor = vaQueryVendorString(src_hwctx->display);
955 av_log(ctx, AV_LOG_ERROR, "Unable to get device info from VAAPI!\n");
956 return AVERROR_EXTERNAL;
959 if (strstr(vendor, "Intel"))
960 dev_select.vendor_id = 0x8086;
961 if (strstr(vendor, "AMD"))
962 dev_select.vendor_id = 0x1002;
964 return vulkan_device_create_internal(ctx, &dev_select, NULL, flags);
967 case AV_HWDEVICE_TYPE_DRM: {
968 AVDRMDeviceContext *src_hwctx = src_ctx->hwctx;
970 drmDevice *drm_dev_info;
971 int err = drmGetDevice(src_hwctx->fd, &drm_dev_info);
973 av_log(ctx, AV_LOG_ERROR, "Unable to get device info from DRM fd!\n");
974 return AVERROR_EXTERNAL;
977 if (drm_dev_info->bustype == DRM_BUS_PCI)
978 dev_select.pci_device = drm_dev_info->deviceinfo.pci->device_id;
980 drmFreeDevice(&drm_dev_info);
982 return vulkan_device_create_internal(ctx, &dev_select, NULL, flags);
986 case AV_HWDEVICE_TYPE_CUDA: {
987 AVHWDeviceContext *cuda_cu = src_ctx;
988 AVCUDADeviceContext *src_hwctx = src_ctx->hwctx;
989 AVCUDADeviceContextInternal *cu_internal = src_hwctx->internal;
990 CudaFunctions *cu = cu_internal->cuda_dl;
992 int ret = CHECK_CU(cu->cuDeviceGetUuid((CUuuid *)&dev_select.uuid,
993 cu_internal->cuda_device));
995 av_log(ctx, AV_LOG_ERROR, "Unable to get UUID from CUDA!\n");
996 return AVERROR_EXTERNAL;
999 dev_select.has_uuid = 1;
1001 return vulkan_device_create_internal(ctx, &dev_select, NULL, flags);
1005 return AVERROR(ENOSYS);
1009 static int vulkan_frames_get_constraints(AVHWDeviceContext *ctx,
1010 const void *hwconfig,
1011 AVHWFramesConstraints *constraints)
1014 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1015 VulkanDevicePriv *p = ctx->internal->priv;
1017 for (enum AVPixelFormat i = 0; i < AV_PIX_FMT_NB; i++)
1018 count += pixfmt_is_supported(hwctx, i, p->use_linear_images);
1021 if (p->dev_is_nvidia)
1025 constraints->valid_sw_formats = av_malloc_array(count + 1,
1026 sizeof(enum AVPixelFormat));
1027 if (!constraints->valid_sw_formats)
1028 return AVERROR(ENOMEM);
1031 for (enum AVPixelFormat i = 0; i < AV_PIX_FMT_NB; i++)
1032 if (pixfmt_is_supported(hwctx, i, p->use_linear_images))
1033 constraints->valid_sw_formats[count++] = i;
1036 if (p->dev_is_nvidia)
1037 constraints->valid_sw_formats[count++] = AV_PIX_FMT_CUDA;
1039 constraints->valid_sw_formats[count++] = AV_PIX_FMT_NONE;
1041 constraints->min_width = 0;
1042 constraints->min_height = 0;
1043 constraints->max_width = p->props.limits.maxImageDimension2D;
1044 constraints->max_height = p->props.limits.maxImageDimension2D;
1046 constraints->valid_hw_formats = av_malloc_array(2, sizeof(enum AVPixelFormat));
1047 if (!constraints->valid_hw_formats)
1048 return AVERROR(ENOMEM);
1050 constraints->valid_hw_formats[0] = AV_PIX_FMT_VULKAN;
1051 constraints->valid_hw_formats[1] = AV_PIX_FMT_NONE;
1056 static int alloc_mem(AVHWDeviceContext *ctx, VkMemoryRequirements *req,
1057 VkMemoryPropertyFlagBits req_flags, void *alloc_extension,
1058 VkMemoryPropertyFlagBits *mem_flags, VkDeviceMemory *mem)
1062 VulkanDevicePriv *p = ctx->internal->priv;
1063 AVVulkanDeviceContext *dev_hwctx = ctx->hwctx;
1064 VkMemoryAllocateInfo alloc_info = {
1065 .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
1066 .pNext = alloc_extension,
1069 /* Align if we need to */
1070 if (req_flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT)
1071 req->size = FFALIGN(req->size, p->props.limits.minMemoryMapAlignment);
1073 alloc_info.allocationSize = req->size;
1075 /* The vulkan spec requires memory types to be sorted in the "optimal"
1076 * order, so the first matching type we find will be the best/fastest one */
1077 for (int i = 0; i < p->mprops.memoryTypeCount; i++) {
1078 /* The memory type must be supported by the requirements (bitfield) */
1079 if (!(req->memoryTypeBits & (1 << i)))
1082 /* The memory type flags must include our properties */
1083 if ((p->mprops.memoryTypes[i].propertyFlags & req_flags) != req_flags)
1086 /* Found a suitable memory type */
1092 av_log(ctx, AV_LOG_ERROR, "No memory type found for flags 0x%x\n",
1094 return AVERROR(EINVAL);
1097 alloc_info.memoryTypeIndex = index;
1099 ret = vkAllocateMemory(dev_hwctx->act_dev, &alloc_info,
1100 dev_hwctx->alloc, mem);
1101 if (ret != VK_SUCCESS) {
1102 av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory: %s\n",
1104 return AVERROR(ENOMEM);
1107 *mem_flags |= p->mprops.memoryTypes[index].propertyFlags;
1112 static void vulkan_free_internal(AVVkFrameInternal *internal)
1118 if (internal->cuda_fc_ref) {
1119 AVHWFramesContext *cuda_fc = (AVHWFramesContext *)internal->cuda_fc_ref->data;
1120 int planes = av_pix_fmt_count_planes(cuda_fc->sw_format);
1121 AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
1122 AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
1123 AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
1124 CudaFunctions *cu = cu_internal->cuda_dl;
1126 for (int i = 0; i < planes; i++) {
1127 if (internal->cu_sem[i])
1128 CHECK_CU(cu->cuDestroyExternalSemaphore(internal->cu_sem[i]));
1129 if (internal->cu_mma[i])
1130 CHECK_CU(cu->cuMipmappedArrayDestroy(internal->cu_mma[i]));
1131 if (internal->ext_mem[i])
1132 CHECK_CU(cu->cuDestroyExternalMemory(internal->ext_mem[i]));
1135 av_buffer_unref(&internal->cuda_fc_ref);
1142 static void vulkan_frame_free(void *opaque, uint8_t *data)
1144 AVVkFrame *f = (AVVkFrame *)data;
1145 AVHWFramesContext *hwfc = opaque;
1146 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1147 int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1149 vulkan_free_internal(f->internal);
1151 for (int i = 0; i < planes; i++) {
1152 vkDestroyImage(hwctx->act_dev, f->img[i], hwctx->alloc);
1153 vkFreeMemory(hwctx->act_dev, f->mem[i], hwctx->alloc);
1154 vkDestroySemaphore(hwctx->act_dev, f->sem[i], hwctx->alloc);
1160 static int alloc_bind_mem(AVHWFramesContext *hwfc, AVVkFrame *f,
1161 void *alloc_pnext, size_t alloc_pnext_stride)
1165 AVHWDeviceContext *ctx = hwfc->device_ctx;
1166 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1167 VkBindImageMemoryInfo bind_info[AV_NUM_DATA_POINTERS] = { { 0 } };
1169 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1171 for (int i = 0; i < planes; i++) {
1173 VkImageMemoryRequirementsInfo2 req_desc = {
1174 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_REQUIREMENTS_INFO_2,
1177 VkMemoryDedicatedAllocateInfo ded_alloc = {
1178 .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO,
1179 .pNext = (void *)(((uint8_t *)alloc_pnext) + i*alloc_pnext_stride),
1181 VkMemoryDedicatedRequirements ded_req = {
1182 .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS,
1184 VkMemoryRequirements2 req = {
1185 .sType = VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2,
1189 vkGetImageMemoryRequirements2(hwctx->act_dev, &req_desc, &req);
1191 /* In case the implementation prefers/requires dedicated allocation */
1192 use_ded_mem = ded_req.prefersDedicatedAllocation |
1193 ded_req.requiresDedicatedAllocation;
1195 ded_alloc.image = f->img[i];
1197 /* Allocate memory */
1198 if ((err = alloc_mem(ctx, &req.memoryRequirements,
1199 f->tiling == VK_IMAGE_TILING_LINEAR ?
1200 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT :
1201 VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
1202 use_ded_mem ? &ded_alloc : (void *)ded_alloc.pNext,
1203 &f->flags, &f->mem[i])))
1206 f->size[i] = req.memoryRequirements.size;
1207 bind_info[i].sType = VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO;
1208 bind_info[i].image = f->img[i];
1209 bind_info[i].memory = f->mem[i];
1212 /* Bind the allocated memory to the images */
1213 ret = vkBindImageMemory2(hwctx->act_dev, planes, bind_info);
1214 if (ret != VK_SUCCESS) {
1215 av_log(ctx, AV_LOG_ERROR, "Failed to bind memory: %s\n",
1217 return AVERROR_EXTERNAL;
1225 PREP_MODE_RO_SHADER,
1226 PREP_MODE_EXTERNAL_EXPORT,
1229 static int prepare_frame(AVHWFramesContext *hwfc, VulkanExecCtx *ectx,
1230 AVVkFrame *frame, enum PrepMode pmode)
1234 VkImageLayout new_layout;
1235 VkAccessFlags new_access;
1236 AVHWDeviceContext *ctx = hwfc->device_ctx;
1237 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1238 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1240 VkImageMemoryBarrier img_bar[AV_NUM_DATA_POINTERS] = { 0 };
1242 VkCommandBufferBeginInfo cmd_start = {
1243 .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
1244 .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
1247 VkSubmitInfo s_info = {
1248 .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
1249 .commandBufferCount = 1,
1250 .pCommandBuffers = &ectx->buf,
1252 .pSignalSemaphores = frame->sem,
1253 .signalSemaphoreCount = planes,
1256 VkPipelineStageFlagBits wait_st[AV_NUM_DATA_POINTERS];
1257 for (int i = 0; i < planes; i++)
1258 wait_st[i] = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
1261 case PREP_MODE_WRITE:
1262 new_layout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
1263 new_access = VK_ACCESS_TRANSFER_WRITE_BIT;
1264 dst_qf = VK_QUEUE_FAMILY_IGNORED;
1266 case PREP_MODE_RO_SHADER:
1267 new_layout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL;
1268 new_access = VK_ACCESS_TRANSFER_READ_BIT;
1269 dst_qf = VK_QUEUE_FAMILY_IGNORED;
1271 case PREP_MODE_EXTERNAL_EXPORT:
1272 new_layout = VK_IMAGE_LAYOUT_GENERAL;
1273 new_access = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT;
1274 dst_qf = VK_QUEUE_FAMILY_EXTERNAL_KHR;
1275 s_info.pWaitSemaphores = frame->sem;
1276 s_info.pWaitDstStageMask = wait_st;
1277 s_info.waitSemaphoreCount = planes;
1281 ret = vkBeginCommandBuffer(ectx->buf, &cmd_start);
1282 if (ret != VK_SUCCESS)
1283 return AVERROR_EXTERNAL;
1285 /* Change the image layout to something more optimal for writes.
1286 * This also signals the newly created semaphore, making it usable
1287 * for synchronization */
1288 for (int i = 0; i < planes; i++) {
1289 img_bar[i].sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
1290 img_bar[i].srcAccessMask = 0x0;
1291 img_bar[i].dstAccessMask = new_access;
1292 img_bar[i].oldLayout = frame->layout[i];
1293 img_bar[i].newLayout = new_layout;
1294 img_bar[i].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
1295 img_bar[i].dstQueueFamilyIndex = dst_qf;
1296 img_bar[i].image = frame->img[i];
1297 img_bar[i].subresourceRange.levelCount = 1;
1298 img_bar[i].subresourceRange.layerCount = 1;
1299 img_bar[i].subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
1301 frame->layout[i] = img_bar[i].newLayout;
1302 frame->access[i] = img_bar[i].dstAccessMask;
1305 vkCmdPipelineBarrier(ectx->buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
1306 VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
1307 0, NULL, 0, NULL, planes, img_bar);
1309 ret = vkEndCommandBuffer(ectx->buf);
1310 if (ret != VK_SUCCESS)
1311 return AVERROR_EXTERNAL;
1313 ret = vkQueueSubmit(ectx->queue, 1, &s_info, ectx->fence);
1314 if (ret != VK_SUCCESS) {
1315 return AVERROR_EXTERNAL;
1317 vkWaitForFences(hwctx->act_dev, 1, &ectx->fence, VK_TRUE, UINT64_MAX);
1318 vkResetFences(hwctx->act_dev, 1, &ectx->fence);
1324 static int create_frame(AVHWFramesContext *hwfc, AVVkFrame **frame,
1325 VkImageTiling tiling, VkImageUsageFlagBits usage,
1330 AVHWDeviceContext *ctx = hwfc->device_ctx;
1331 VulkanDevicePriv *p = ctx->internal->priv;
1332 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1333 enum AVPixelFormat format = hwfc->sw_format;
1334 const VkFormat *img_fmts = av_vkfmt_from_pixfmt(format);
1335 const int planes = av_pix_fmt_count_planes(format);
1337 VkExportSemaphoreCreateInfo ext_sem_info = {
1338 .sType = VK_STRUCTURE_TYPE_EXPORT_SEMAPHORE_CREATE_INFO,
1339 .handleTypes = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT,
1342 VkSemaphoreCreateInfo sem_spawn = {
1343 .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
1344 .pNext = p->extensions & EXT_EXTERNAL_FD_SEM ? &ext_sem_info : NULL,
1347 AVVkFrame *f = av_vk_frame_alloc();
1349 av_log(ctx, AV_LOG_ERROR, "Unable to allocate memory for AVVkFrame!\n");
1350 return AVERROR(ENOMEM);
1353 /* Create the images */
1354 for (int i = 0; i < planes; i++) {
1355 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(format);
1356 int w = hwfc->width;
1357 int h = hwfc->height;
1358 const int p_w = i > 0 ? AV_CEIL_RSHIFT(w, desc->log2_chroma_w) : w;
1359 const int p_h = i > 0 ? AV_CEIL_RSHIFT(h, desc->log2_chroma_h) : h;
1361 VkImageCreateInfo image_create_info = {
1362 .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
1363 .pNext = create_pnext,
1364 .imageType = VK_IMAGE_TYPE_2D,
1365 .format = img_fmts[i],
1366 .extent.width = p_w,
1367 .extent.height = p_h,
1371 .flags = VK_IMAGE_CREATE_ALIAS_BIT,
1373 .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED,
1375 .samples = VK_SAMPLE_COUNT_1_BIT,
1376 .pQueueFamilyIndices = p->qfs,
1377 .queueFamilyIndexCount = p->num_qfs,
1378 .sharingMode = p->num_qfs > 1 ? VK_SHARING_MODE_CONCURRENT :
1379 VK_SHARING_MODE_EXCLUSIVE,
1382 ret = vkCreateImage(hwctx->act_dev, &image_create_info,
1383 hwctx->alloc, &f->img[i]);
1384 if (ret != VK_SUCCESS) {
1385 av_log(ctx, AV_LOG_ERROR, "Image creation failure: %s\n",
1387 err = AVERROR(EINVAL);
1391 /* Create semaphore */
1392 ret = vkCreateSemaphore(hwctx->act_dev, &sem_spawn,
1393 hwctx->alloc, &f->sem[i]);
1394 if (ret != VK_SUCCESS) {
1395 av_log(hwctx, AV_LOG_ERROR, "Failed to create semaphore: %s\n",
1397 return AVERROR_EXTERNAL;
1400 f->layout[i] = image_create_info.initialLayout;
1411 vulkan_frame_free(hwfc, (uint8_t *)f);
1415 /* Checks if an export flag is enabled, and if it is ORs it with *iexp */
1416 static void try_export_flags(AVHWFramesContext *hwfc,
1417 VkExternalMemoryHandleTypeFlags *comp_handle_types,
1418 VkExternalMemoryHandleTypeFlagBits *iexp,
1419 VkExternalMemoryHandleTypeFlagBits exp)
1422 AVVulkanFramesContext *hwctx = hwfc->hwctx;
1423 AVVulkanDeviceContext *dev_hwctx = hwfc->device_ctx->hwctx;
1424 VkExternalImageFormatProperties eprops = {
1425 .sType = VK_STRUCTURE_TYPE_EXTERNAL_IMAGE_FORMAT_PROPERTIES_KHR,
1427 VkImageFormatProperties2 props = {
1428 .sType = VK_STRUCTURE_TYPE_IMAGE_FORMAT_PROPERTIES_2,
1431 VkPhysicalDeviceExternalImageFormatInfo enext = {
1432 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_IMAGE_FORMAT_INFO,
1435 VkPhysicalDeviceImageFormatInfo2 pinfo = {
1436 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_FORMAT_INFO_2,
1437 .pNext = !exp ? NULL : &enext,
1438 .format = av_vkfmt_from_pixfmt(hwfc->sw_format)[0],
1439 .type = VK_IMAGE_TYPE_2D,
1440 .tiling = hwctx->tiling,
1441 .usage = hwctx->usage,
1442 .flags = VK_IMAGE_CREATE_ALIAS_BIT,
1445 ret = vkGetPhysicalDeviceImageFormatProperties2(dev_hwctx->phys_dev,
1447 if (ret == VK_SUCCESS) {
1449 *comp_handle_types |= eprops.externalMemoryProperties.compatibleHandleTypes;
1453 static AVBufferRef *vulkan_pool_alloc(void *opaque, int size)
1457 AVBufferRef *avbuf = NULL;
1458 AVHWFramesContext *hwfc = opaque;
1459 AVVulkanFramesContext *hwctx = hwfc->hwctx;
1460 VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
1461 VkExportMemoryAllocateInfo eminfo[AV_NUM_DATA_POINTERS];
1462 VkExternalMemoryHandleTypeFlags e = 0x0;
1464 VkExternalMemoryImageCreateInfo eiinfo = {
1465 .sType = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO,
1466 .pNext = hwctx->create_pnext,
1469 if (p->extensions & EXT_EXTERNAL_FD_MEMORY)
1470 try_export_flags(hwfc, &eiinfo.handleTypes, &e,
1471 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT);
1473 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
1474 try_export_flags(hwfc, &eiinfo.handleTypes, &e,
1475 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
1477 for (int i = 0; i < av_pix_fmt_count_planes(hwfc->sw_format); i++) {
1478 eminfo[i].sType = VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO;
1479 eminfo[i].pNext = hwctx->alloc_pnext[i];
1480 eminfo[i].handleTypes = e;
1483 err = create_frame(hwfc, &f, hwctx->tiling, hwctx->usage,
1484 eiinfo.handleTypes ? &eiinfo : NULL);
1488 err = alloc_bind_mem(hwfc, f, eminfo, sizeof(*eminfo));
1492 err = prepare_frame(hwfc, &p->cmd, f, PREP_MODE_WRITE);
1496 avbuf = av_buffer_create((uint8_t *)f, sizeof(AVVkFrame),
1497 vulkan_frame_free, hwfc, 0);
1504 vulkan_frame_free(hwfc, (uint8_t *)f);
1508 static void vulkan_frames_uninit(AVHWFramesContext *hwfc)
1510 VulkanFramesPriv *fp = hwfc->internal->priv;
1512 free_exec_ctx(hwfc->device_ctx, &fp->cmd);
1515 static int vulkan_frames_init(AVHWFramesContext *hwfc)
1519 AVVulkanFramesContext *hwctx = hwfc->hwctx;
1520 VulkanFramesPriv *fp = hwfc->internal->priv;
1521 AVVulkanDeviceContext *dev_hwctx = hwfc->device_ctx->hwctx;
1522 VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
1527 /* Default pool flags */
1528 hwctx->tiling = hwctx->tiling ? hwctx->tiling : p->use_linear_images ?
1529 VK_IMAGE_TILING_LINEAR : VK_IMAGE_TILING_OPTIMAL;
1531 hwctx->usage |= DEFAULT_USAGE_FLAGS;
1533 err = create_exec_ctx(hwfc->device_ctx, &fp->cmd,
1534 dev_hwctx->queue_family_tx_index);
1538 /* Test to see if allocation will fail */
1539 err = create_frame(hwfc, &f, hwctx->tiling, hwctx->usage,
1540 hwctx->create_pnext);
1542 free_exec_ctx(hwfc->device_ctx, &p->cmd);
1546 vulkan_frame_free(hwfc, (uint8_t *)f);
1548 hwfc->internal->pool_internal = av_buffer_pool_init2(sizeof(AVVkFrame),
1549 hwfc, vulkan_pool_alloc,
1551 if (!hwfc->internal->pool_internal) {
1552 free_exec_ctx(hwfc->device_ctx, &p->cmd);
1553 return AVERROR(ENOMEM);
1559 static int vulkan_get_buffer(AVHWFramesContext *hwfc, AVFrame *frame)
1561 frame->buf[0] = av_buffer_pool_get(hwfc->pool);
1563 return AVERROR(ENOMEM);
1565 frame->data[0] = frame->buf[0]->data;
1566 frame->format = AV_PIX_FMT_VULKAN;
1567 frame->width = hwfc->width;
1568 frame->height = hwfc->height;
1573 static int vulkan_transfer_get_formats(AVHWFramesContext *hwfc,
1574 enum AVHWFrameTransferDirection dir,
1575 enum AVPixelFormat **formats)
1577 enum AVPixelFormat *fmts = av_malloc_array(2, sizeof(*fmts));
1579 return AVERROR(ENOMEM);
1581 fmts[0] = hwfc->sw_format;
1582 fmts[1] = AV_PIX_FMT_NONE;
1588 typedef struct VulkanMapping {
1593 static void vulkan_unmap_frame(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
1595 VulkanMapping *map = hwmap->priv;
1596 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1597 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1599 /* Check if buffer needs flushing */
1600 if ((map->flags & AV_HWFRAME_MAP_WRITE) &&
1601 !(map->frame->flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)) {
1603 VkMappedMemoryRange flush_ranges[AV_NUM_DATA_POINTERS] = { { 0 } };
1605 for (int i = 0; i < planes; i++) {
1606 flush_ranges[i].sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
1607 flush_ranges[i].memory = map->frame->mem[i];
1608 flush_ranges[i].size = VK_WHOLE_SIZE;
1611 ret = vkFlushMappedMemoryRanges(hwctx->act_dev, planes,
1613 if (ret != VK_SUCCESS) {
1614 av_log(hwfc, AV_LOG_ERROR, "Failed to flush memory: %s\n",
1619 for (int i = 0; i < planes; i++)
1620 vkUnmapMemory(hwctx->act_dev, map->frame->mem[i]);
1625 static int vulkan_map_frame_to_mem(AVHWFramesContext *hwfc, AVFrame *dst,
1626 const AVFrame *src, int flags)
1629 int err, mapped_mem_count = 0;
1630 AVVkFrame *f = (AVVkFrame *)src->data[0];
1631 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1632 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1634 VulkanMapping *map = av_mallocz(sizeof(VulkanMapping));
1636 return AVERROR(EINVAL);
1638 if (src->format != AV_PIX_FMT_VULKAN) {
1639 av_log(hwfc, AV_LOG_ERROR, "Cannot map from pixel format %s!\n",
1640 av_get_pix_fmt_name(src->format));
1641 err = AVERROR(EINVAL);
1645 if (!(f->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) ||
1646 !(f->tiling == VK_IMAGE_TILING_LINEAR)) {
1647 av_log(hwfc, AV_LOG_ERROR, "Unable to map frame, not host visible "
1649 err = AVERROR(EINVAL);
1653 dst->width = src->width;
1654 dst->height = src->height;
1656 for (int i = 0; i < planes; i++) {
1657 ret = vkMapMemory(hwctx->act_dev, f->mem[i], 0,
1658 VK_WHOLE_SIZE, 0, (void **)&dst->data[i]);
1659 if (ret != VK_SUCCESS) {
1660 av_log(hwfc, AV_LOG_ERROR, "Failed to map image memory: %s\n",
1662 err = AVERROR_EXTERNAL;
1668 /* Check if the memory contents matter */
1669 if (((flags & AV_HWFRAME_MAP_READ) || !(flags & AV_HWFRAME_MAP_OVERWRITE)) &&
1670 !(f->flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)) {
1671 VkMappedMemoryRange map_mem_ranges[AV_NUM_DATA_POINTERS] = { { 0 } };
1672 for (int i = 0; i < planes; i++) {
1673 map_mem_ranges[i].sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
1674 map_mem_ranges[i].size = VK_WHOLE_SIZE;
1675 map_mem_ranges[i].memory = f->mem[i];
1678 ret = vkInvalidateMappedMemoryRanges(hwctx->act_dev, planes,
1680 if (ret != VK_SUCCESS) {
1681 av_log(hwfc, AV_LOG_ERROR, "Failed to invalidate memory: %s\n",
1683 err = AVERROR_EXTERNAL;
1688 for (int i = 0; i < planes; i++) {
1689 VkImageSubresource sub = {
1690 .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
1692 VkSubresourceLayout layout;
1693 vkGetImageSubresourceLayout(hwctx->act_dev, f->img[i], &sub, &layout);
1694 dst->linesize[i] = layout.rowPitch;
1700 err = ff_hwframe_map_create(src->hw_frames_ctx, dst, src,
1701 &vulkan_unmap_frame, map);
1708 for (int i = 0; i < mapped_mem_count; i++)
1709 vkUnmapMemory(hwctx->act_dev, f->mem[i]);
1716 static void vulkan_unmap_from(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
1718 VulkanMapping *map = hwmap->priv;
1719 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1720 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1722 for (int i = 0; i < planes; i++) {
1723 vkDestroyImage(hwctx->act_dev, map->frame->img[i], hwctx->alloc);
1724 vkFreeMemory(hwctx->act_dev, map->frame->mem[i], hwctx->alloc);
1725 vkDestroySemaphore(hwctx->act_dev, map->frame->sem[i], hwctx->alloc);
1728 av_freep(&map->frame);
1731 static const struct {
1732 uint32_t drm_fourcc;
1734 } vulkan_drm_format_map[] = {
1735 { DRM_FORMAT_R8, VK_FORMAT_R8_UNORM },
1736 { DRM_FORMAT_R16, VK_FORMAT_R16_UNORM },
1737 { DRM_FORMAT_GR88, VK_FORMAT_R8G8_UNORM },
1738 { DRM_FORMAT_RG88, VK_FORMAT_R8G8_UNORM },
1739 { DRM_FORMAT_GR1616, VK_FORMAT_R16G16_UNORM },
1740 { DRM_FORMAT_RG1616, VK_FORMAT_R16G16_UNORM },
1741 { DRM_FORMAT_ARGB8888, VK_FORMAT_B8G8R8A8_UNORM },
1742 { DRM_FORMAT_XRGB8888, VK_FORMAT_B8G8R8A8_UNORM },
1743 { DRM_FORMAT_ABGR8888, VK_FORMAT_R8G8B8A8_UNORM },
1744 { DRM_FORMAT_XBGR8888, VK_FORMAT_R8G8B8A8_UNORM },
1747 static inline VkFormat drm_to_vulkan_fmt(uint32_t drm_fourcc)
1749 for (int i = 0; i < FF_ARRAY_ELEMS(vulkan_drm_format_map); i++)
1750 if (vulkan_drm_format_map[i].drm_fourcc == drm_fourcc)
1751 return vulkan_drm_format_map[i].vk_format;
1752 return VK_FORMAT_UNDEFINED;
1755 static int vulkan_map_from_drm_frame_desc(AVHWFramesContext *hwfc, AVVkFrame **frame,
1756 AVDRMFrameDescriptor *desc)
1761 int bind_counts = 0;
1762 AVHWDeviceContext *ctx = hwfc->device_ctx;
1763 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1764 VulkanDevicePriv *p = ctx->internal->priv;
1765 const AVPixFmtDescriptor *fmt_desc = av_pix_fmt_desc_get(hwfc->sw_format);
1766 const int has_modifiers = p->extensions & EXT_DRM_MODIFIER_FLAGS;
1767 VkSubresourceLayout plane_data[AV_NUM_DATA_POINTERS] = { 0 };
1768 VkBindImageMemoryInfo bind_info[AV_NUM_DATA_POINTERS] = { 0 };
1769 VkBindImagePlaneMemoryInfo plane_info[AV_NUM_DATA_POINTERS] = { 0 };
1770 VkExternalMemoryHandleTypeFlagBits htype = VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT;
1772 VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdPropertiesKHR);
1774 for (int i = 0; i < desc->nb_layers; i++) {
1775 if (drm_to_vulkan_fmt(desc->layers[i].format) == VK_FORMAT_UNDEFINED) {
1776 av_log(ctx, AV_LOG_ERROR, "Unsupported DMABUF layer format %#08x!\n",
1777 desc->layers[i].format);
1778 return AVERROR(EINVAL);
1782 if (!(f = av_vk_frame_alloc())) {
1783 av_log(ctx, AV_LOG_ERROR, "Unable to allocate memory for AVVkFrame!\n");
1784 err = AVERROR(ENOMEM);
1788 for (int i = 0; i < desc->nb_objects; i++) {
1789 VkMemoryFdPropertiesKHR fdmp = {
1790 .sType = VK_STRUCTURE_TYPE_MEMORY_FD_PROPERTIES_KHR,
1792 VkMemoryRequirements req = {
1793 .size = desc->objects[i].size,
1795 VkImportMemoryFdInfoKHR idesc = {
1796 .sType = VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR,
1797 .handleType = htype,
1798 .fd = dup(desc->objects[i].fd),
1801 ret = pfn_vkGetMemoryFdPropertiesKHR(hwctx->act_dev, htype,
1803 if (ret != VK_SUCCESS) {
1804 av_log(hwfc, AV_LOG_ERROR, "Failed to get FD properties: %s\n",
1806 err = AVERROR_EXTERNAL;
1811 req.memoryTypeBits = fdmp.memoryTypeBits;
1813 err = alloc_mem(ctx, &req, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
1814 &idesc, &f->flags, &f->mem[i]);
1820 f->size[i] = desc->objects[i].size;
1823 f->tiling = has_modifiers ? VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT :
1824 desc->objects[0].format_modifier == DRM_FORMAT_MOD_LINEAR ?
1825 VK_IMAGE_TILING_LINEAR : VK_IMAGE_TILING_OPTIMAL;
1827 for (int i = 0; i < desc->nb_layers; i++) {
1828 const int planes = desc->layers[i].nb_planes;
1829 const int signal_p = has_modifiers && (planes > 1);
1831 VkImageDrmFormatModifierExplicitCreateInfoEXT drm_info = {
1832 .sType = VK_STRUCTURE_TYPE_IMAGE_DRM_FORMAT_MODIFIER_EXPLICIT_CREATE_INFO_EXT,
1833 .drmFormatModifier = desc->objects[0].format_modifier,
1834 .drmFormatModifierPlaneCount = planes,
1835 .pPlaneLayouts = (const VkSubresourceLayout *)&plane_data,
1838 VkExternalMemoryImageCreateInfo einfo = {
1839 .sType = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO,
1840 .pNext = has_modifiers ? &drm_info : NULL,
1841 .handleTypes = htype,
1844 VkSemaphoreCreateInfo sem_spawn = {
1845 .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
1848 const int p_w = i > 0 ? AV_CEIL_RSHIFT(hwfc->width, fmt_desc->log2_chroma_w) : hwfc->width;
1849 const int p_h = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, fmt_desc->log2_chroma_h) : hwfc->height;
1851 VkImageCreateInfo image_create_info = {
1852 .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
1854 .imageType = VK_IMAGE_TYPE_2D,
1855 .format = drm_to_vulkan_fmt(desc->layers[i].format),
1856 .extent.width = p_w,
1857 .extent.height = p_h,
1861 .flags = VK_IMAGE_CREATE_ALIAS_BIT,
1862 .tiling = f->tiling,
1863 .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED, /* specs say so */
1864 .usage = DEFAULT_USAGE_FLAGS,
1865 .samples = VK_SAMPLE_COUNT_1_BIT,
1866 .pQueueFamilyIndices = p->qfs,
1867 .queueFamilyIndexCount = p->num_qfs,
1868 .sharingMode = p->num_qfs > 1 ? VK_SHARING_MODE_CONCURRENT :
1869 VK_SHARING_MODE_EXCLUSIVE,
1872 for (int j = 0; j < planes; j++) {
1873 plane_data[j].offset = desc->layers[i].planes[j].offset;
1874 plane_data[j].rowPitch = desc->layers[i].planes[j].pitch;
1875 plane_data[j].size = 0; /* The specs say so for all 3 */
1876 plane_data[j].arrayPitch = 0;
1877 plane_data[j].depthPitch = 0;
1881 ret = vkCreateImage(hwctx->act_dev, &image_create_info,
1882 hwctx->alloc, &f->img[i]);
1883 if (ret != VK_SUCCESS) {
1884 av_log(ctx, AV_LOG_ERROR, "Image creation failure: %s\n",
1886 err = AVERROR(EINVAL);
1890 ret = vkCreateSemaphore(hwctx->act_dev, &sem_spawn,
1891 hwctx->alloc, &f->sem[i]);
1892 if (ret != VK_SUCCESS) {
1893 av_log(hwctx, AV_LOG_ERROR, "Failed to create semaphore: %s\n",
1895 return AVERROR_EXTERNAL;
1898 /* We'd import a semaphore onto the one we created using
1899 * vkImportSemaphoreFdKHR but unfortunately neither DRM nor VAAPI
1900 * offer us anything we could import and sync with, so instead
1901 * just signal the semaphore we created. */
1903 f->layout[i] = image_create_info.initialLayout;
1906 for (int j = 0; j < planes; j++) {
1907 VkImageAspectFlagBits aspect = j == 0 ? VK_IMAGE_ASPECT_MEMORY_PLANE_0_BIT_EXT :
1908 j == 1 ? VK_IMAGE_ASPECT_MEMORY_PLANE_1_BIT_EXT :
1909 VK_IMAGE_ASPECT_MEMORY_PLANE_2_BIT_EXT;
1911 plane_info[bind_counts].sType = VK_STRUCTURE_TYPE_BIND_IMAGE_PLANE_MEMORY_INFO;
1912 plane_info[bind_counts].planeAspect = aspect;
1914 bind_info[bind_counts].sType = VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO;
1915 bind_info[bind_counts].pNext = signal_p ? &plane_info[bind_counts] : NULL;
1916 bind_info[bind_counts].image = f->img[i];
1917 bind_info[bind_counts].memory = f->mem[desc->layers[i].planes[j].object_index];
1918 bind_info[bind_counts].memoryOffset = desc->layers[i].planes[j].offset;
1923 /* Bind the allocated memory to the images */
1924 ret = vkBindImageMemory2(hwctx->act_dev, bind_counts, bind_info);
1925 if (ret != VK_SUCCESS) {
1926 av_log(ctx, AV_LOG_ERROR, "Failed to bind memory: %s\n",
1928 return AVERROR_EXTERNAL;
1931 /* NOTE: This is completely uneccesary and unneeded once we can import
1932 * semaphores from DRM. Otherwise we have to activate the semaphores.
1933 * We're reusing the exec context that's also used for uploads/downloads. */
1934 err = prepare_frame(hwfc, &p->cmd, f, PREP_MODE_RO_SHADER);
1943 for (int i = 0; i < desc->nb_layers; i++) {
1944 vkDestroyImage(hwctx->act_dev, f->img[i], hwctx->alloc);
1945 vkDestroySemaphore(hwctx->act_dev, f->sem[i], hwctx->alloc);
1947 for (int i = 0; i < desc->nb_objects; i++)
1948 vkFreeMemory(hwctx->act_dev, f->mem[i], hwctx->alloc);
1955 static int vulkan_map_from_drm(AVHWFramesContext *hwfc, AVFrame *dst,
1956 const AVFrame *src, int flags)
1960 VulkanMapping *map = NULL;
1962 err = vulkan_map_from_drm_frame_desc(hwfc, &f,
1963 (AVDRMFrameDescriptor *)src->data[0]);
1967 /* The unmapping function will free this */
1968 dst->data[0] = (uint8_t *)f;
1969 dst->width = src->width;
1970 dst->height = src->height;
1972 map = av_mallocz(sizeof(VulkanMapping));
1979 err = ff_hwframe_map_create(dst->hw_frames_ctx, dst, src,
1980 &vulkan_unmap_from, map);
1984 av_log(hwfc, AV_LOG_DEBUG, "Mapped DRM object to Vulkan!\n");
1989 vulkan_frame_free(hwfc->device_ctx->hwctx, (uint8_t *)f);
1995 static int vulkan_map_from_vaapi(AVHWFramesContext *dst_fc,
1996 AVFrame *dst, const AVFrame *src,
2000 AVFrame *tmp = av_frame_alloc();
2001 AVHWFramesContext *vaapi_fc = (AVHWFramesContext*)src->hw_frames_ctx->data;
2002 AVVAAPIDeviceContext *vaapi_ctx = vaapi_fc->device_ctx->hwctx;
2003 VASurfaceID surface_id = (VASurfaceID)(uintptr_t)src->data[3];
2006 return AVERROR(ENOMEM);
2008 /* We have to sync since like the previous comment said, no semaphores */
2009 vaSyncSurface(vaapi_ctx->display, surface_id);
2011 tmp->format = AV_PIX_FMT_DRM_PRIME;
2013 err = av_hwframe_map(tmp, src, flags);
2017 err = vulkan_map_from_drm(dst_fc, dst, tmp, flags);
2021 err = ff_hwframe_map_replace(dst, src);
2024 av_frame_free(&tmp);
2031 static int vulkan_export_to_cuda(AVHWFramesContext *hwfc,
2032 AVBufferRef *cuda_hwfc,
2033 const AVFrame *frame)
2038 AVVkFrameInternal *dst_int;
2039 AVHWDeviceContext *ctx = hwfc->device_ctx;
2040 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2041 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
2042 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
2043 VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdKHR);
2044 VK_LOAD_PFN(hwctx->inst, vkGetSemaphoreFdKHR);
2046 AVHWFramesContext *cuda_fc = (AVHWFramesContext*)cuda_hwfc->data;
2047 AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
2048 AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
2049 AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
2050 CudaFunctions *cu = cu_internal->cuda_dl;
2051 CUarray_format cufmt = desc->comp[0].depth > 8 ? CU_AD_FORMAT_UNSIGNED_INT16 :
2052 CU_AD_FORMAT_UNSIGNED_INT8;
2054 dst_f = (AVVkFrame *)frame->data[0];
2056 dst_int = dst_f->internal;
2057 if (!dst_int || !dst_int->cuda_fc_ref) {
2058 if (!dst_f->internal)
2059 dst_f->internal = dst_int = av_mallocz(sizeof(*dst_f->internal));
2062 err = AVERROR(ENOMEM);
2066 dst_int->cuda_fc_ref = av_buffer_ref(cuda_hwfc);
2067 if (!dst_int->cuda_fc_ref) {
2068 err = AVERROR(ENOMEM);
2072 for (int i = 0; i < planes; i++) {
2073 CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC tex_desc = {
2076 .Width = i > 0 ? AV_CEIL_RSHIFT(hwfc->width, desc->log2_chroma_w)
2078 .Height = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, desc->log2_chroma_h)
2082 .NumChannels = 1 + ((planes == 2) && i),
2087 CUDA_EXTERNAL_MEMORY_HANDLE_DESC ext_desc = {
2088 .type = CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD,
2089 .size = dst_f->size[i],
2091 VkMemoryGetFdInfoKHR export_info = {
2092 .sType = VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR,
2093 .memory = dst_f->mem[i],
2094 .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR,
2096 VkSemaphoreGetFdInfoKHR sem_export = {
2097 .sType = VK_STRUCTURE_TYPE_SEMAPHORE_GET_FD_INFO_KHR,
2098 .semaphore = dst_f->sem[i],
2099 .handleType = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT,
2101 CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC ext_sem_desc = {
2102 .type = CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD,
2105 ret = pfn_vkGetMemoryFdKHR(hwctx->act_dev, &export_info,
2106 &ext_desc.handle.fd);
2107 if (ret != VK_SUCCESS) {
2108 av_log(hwfc, AV_LOG_ERROR, "Unable to export the image as a FD!\n");
2109 err = AVERROR_EXTERNAL;
2113 ret = CHECK_CU(cu->cuImportExternalMemory(&dst_int->ext_mem[i], &ext_desc));
2115 err = AVERROR_EXTERNAL;
2119 ret = CHECK_CU(cu->cuExternalMemoryGetMappedMipmappedArray(&dst_int->cu_mma[i],
2120 dst_int->ext_mem[i],
2123 err = AVERROR_EXTERNAL;
2127 ret = CHECK_CU(cu->cuMipmappedArrayGetLevel(&dst_int->cu_array[i],
2128 dst_int->cu_mma[i], 0));
2130 err = AVERROR_EXTERNAL;
2134 ret = pfn_vkGetSemaphoreFdKHR(hwctx->act_dev, &sem_export,
2135 &ext_sem_desc.handle.fd);
2136 if (ret != VK_SUCCESS) {
2137 av_log(ctx, AV_LOG_ERROR, "Failed to export semaphore: %s\n",
2139 err = AVERROR_EXTERNAL;
2143 ret = CHECK_CU(cu->cuImportExternalSemaphore(&dst_int->cu_sem[i],
2146 err = AVERROR_EXTERNAL;
2158 static int vulkan_transfer_data_from_cuda(AVHWFramesContext *hwfc,
2159 AVFrame *dst, const AVFrame *src)
2165 AVVkFrameInternal *dst_int;
2166 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
2167 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
2169 AVHWFramesContext *cuda_fc = (AVHWFramesContext*)src->hw_frames_ctx->data;
2170 AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
2171 AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
2172 AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
2173 CudaFunctions *cu = cu_internal->cuda_dl;
2174 CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS s_w_par[AV_NUM_DATA_POINTERS] = { 0 };
2175 CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS s_s_par[AV_NUM_DATA_POINTERS] = { 0 };
2177 ret = CHECK_CU(cu->cuCtxPushCurrent(cuda_dev->cuda_ctx));
2179 err = AVERROR_EXTERNAL;
2183 dst_f = (AVVkFrame *)dst->data[0];
2185 ret = vulkan_export_to_cuda(hwfc, src->hw_frames_ctx, dst);
2189 dst_int = dst_f->internal;
2191 ret = CHECK_CU(cu->cuWaitExternalSemaphoresAsync(dst_int->cu_sem, s_w_par,
2192 planes, cuda_dev->stream));
2194 err = AVERROR_EXTERNAL;
2198 for (int i = 0; i < planes; i++) {
2199 CUDA_MEMCPY2D cpy = {
2200 .srcMemoryType = CU_MEMORYTYPE_DEVICE,
2201 .srcDevice = (CUdeviceptr)src->data[i],
2202 .srcPitch = src->linesize[i],
2205 .dstMemoryType = CU_MEMORYTYPE_ARRAY,
2206 .dstArray = dst_int->cu_array[i],
2207 .WidthInBytes = (i > 0 ? AV_CEIL_RSHIFT(hwfc->width, desc->log2_chroma_w)
2208 : hwfc->width) * desc->comp[i].step,
2209 .Height = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, desc->log2_chroma_h)
2213 ret = CHECK_CU(cu->cuMemcpy2DAsync(&cpy, cuda_dev->stream));
2215 err = AVERROR_EXTERNAL;
2220 ret = CHECK_CU(cu->cuSignalExternalSemaphoresAsync(dst_int->cu_sem, s_s_par,
2221 planes, cuda_dev->stream));
2223 err = AVERROR_EXTERNAL;
2227 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2229 av_log(hwfc, AV_LOG_VERBOSE, "Transfered CUDA image to Vulkan!\n");
2234 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2235 vulkan_free_internal(dst_int);
2236 dst_f->internal = NULL;
2237 av_buffer_unref(&dst->buf[0]);
2242 static int vulkan_map_to(AVHWFramesContext *hwfc, AVFrame *dst,
2243 const AVFrame *src, int flags)
2245 av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2247 switch (src->format) {
2250 case AV_PIX_FMT_VAAPI:
2251 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2252 return vulkan_map_from_vaapi(hwfc, dst, src, flags);
2254 case AV_PIX_FMT_DRM_PRIME:
2255 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2256 return vulkan_map_from_drm(hwfc, dst, src, flags);
2259 return AVERROR(ENOSYS);
2264 typedef struct VulkanDRMMapping {
2265 AVDRMFrameDescriptor drm_desc;
2269 static void vulkan_unmap_to_drm(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
2271 AVDRMFrameDescriptor *drm_desc = hwmap->priv;
2273 for (int i = 0; i < drm_desc->nb_objects; i++)
2274 close(drm_desc->objects[i].fd);
2279 static inline uint32_t vulkan_fmt_to_drm(VkFormat vkfmt)
2281 for (int i = 0; i < FF_ARRAY_ELEMS(vulkan_drm_format_map); i++)
2282 if (vulkan_drm_format_map[i].vk_format == vkfmt)
2283 return vulkan_drm_format_map[i].drm_fourcc;
2284 return DRM_FORMAT_INVALID;
2287 static int vulkan_map_to_drm(AVHWFramesContext *hwfc, AVFrame *dst,
2288 const AVFrame *src, int flags)
2292 AVVkFrame *f = (AVVkFrame *)src->data[0];
2293 VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2294 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
2295 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
2296 VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdKHR);
2297 VkImageDrmFormatModifierPropertiesEXT drm_mod = {
2298 .sType = VK_STRUCTURE_TYPE_IMAGE_DRM_FORMAT_MODIFIER_PROPERTIES_EXT,
2301 AVDRMFrameDescriptor *drm_desc = av_mallocz(sizeof(*drm_desc));
2303 return AVERROR(ENOMEM);
2305 err = prepare_frame(hwfc, &p->cmd, f, PREP_MODE_EXTERNAL_EXPORT);
2309 err = ff_hwframe_map_create(src->hw_frames_ctx, dst, src, &vulkan_unmap_to_drm, drm_desc);
2313 if (p->extensions & EXT_DRM_MODIFIER_FLAGS) {
2314 VK_LOAD_PFN(hwctx->inst, vkGetImageDrmFormatModifierPropertiesEXT);
2315 ret = pfn_vkGetImageDrmFormatModifierPropertiesEXT(hwctx->act_dev, f->img[0],
2317 if (ret != VK_SUCCESS) {
2318 av_log(hwfc, AV_LOG_ERROR, "Failed to retrieve DRM format modifier!\n");
2319 err = AVERROR_EXTERNAL;
2324 for (int i = 0; (i < planes) && (f->mem[i]); i++) {
2325 VkMemoryGetFdInfoKHR export_info = {
2326 .sType = VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR,
2327 .memory = f->mem[i],
2328 .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT,
2331 ret = pfn_vkGetMemoryFdKHR(hwctx->act_dev, &export_info,
2332 &drm_desc->objects[i].fd);
2333 if (ret != VK_SUCCESS) {
2334 av_log(hwfc, AV_LOG_ERROR, "Unable to export the image as a FD!\n");
2335 err = AVERROR_EXTERNAL;
2339 drm_desc->nb_objects++;
2340 drm_desc->objects[i].size = f->size[i];
2341 drm_desc->objects[i].format_modifier = drm_mod.drmFormatModifier;
2344 drm_desc->nb_layers = planes;
2345 for (int i = 0; i < drm_desc->nb_layers; i++) {
2346 VkSubresourceLayout layout;
2347 VkImageSubresource sub = {
2348 .aspectMask = p->extensions & EXT_DRM_MODIFIER_FLAGS ?
2349 VK_IMAGE_ASPECT_MEMORY_PLANE_0_BIT_EXT :
2350 VK_IMAGE_ASPECT_COLOR_BIT,
2352 VkFormat plane_vkfmt = av_vkfmt_from_pixfmt(hwfc->sw_format)[i];
2354 drm_desc->layers[i].format = vulkan_fmt_to_drm(plane_vkfmt);
2355 drm_desc->layers[i].nb_planes = 1;
2357 if (drm_desc->layers[i].format == DRM_FORMAT_INVALID) {
2358 av_log(hwfc, AV_LOG_ERROR, "Cannot map to DRM layer, unsupported!\n");
2359 err = AVERROR_PATCHWELCOME;
2363 drm_desc->layers[i].planes[0].object_index = FFMIN(i, drm_desc->nb_objects - 1);
2365 if (f->tiling == VK_IMAGE_TILING_OPTIMAL)
2368 vkGetImageSubresourceLayout(hwctx->act_dev, f->img[i], &sub, &layout);
2369 drm_desc->layers[i].planes[0].offset = layout.offset;
2370 drm_desc->layers[i].planes[0].pitch = layout.rowPitch;
2373 dst->width = src->width;
2374 dst->height = src->height;
2375 dst->data[0] = (uint8_t *)drm_desc;
2377 av_log(hwfc, AV_LOG_VERBOSE, "Mapped AVVkFrame to a DRM object!\n");
2387 static int vulkan_map_to_vaapi(AVHWFramesContext *hwfc, AVFrame *dst,
2388 const AVFrame *src, int flags)
2391 AVFrame *tmp = av_frame_alloc();
2393 return AVERROR(ENOMEM);
2395 tmp->format = AV_PIX_FMT_DRM_PRIME;
2397 err = vulkan_map_to_drm(hwfc, tmp, src, flags);
2401 err = av_hwframe_map(dst, tmp, flags);
2405 err = ff_hwframe_map_replace(dst, src);
2408 av_frame_free(&tmp);
2414 static int vulkan_map_from(AVHWFramesContext *hwfc, AVFrame *dst,
2415 const AVFrame *src, int flags)
2417 av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2419 switch (dst->format) {
2421 case AV_PIX_FMT_DRM_PRIME:
2422 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2423 return vulkan_map_to_drm(hwfc, dst, src, flags);
2425 case AV_PIX_FMT_VAAPI:
2426 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2427 return vulkan_map_to_vaapi(hwfc, dst, src, flags);
2431 return vulkan_map_frame_to_mem(hwfc, dst, src, flags);
2435 typedef struct ImageBuffer {
2438 VkMemoryPropertyFlagBits flags;
2441 static void free_buf(AVHWDeviceContext *ctx, ImageBuffer *buf)
2443 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2447 vkDestroyBuffer(hwctx->act_dev, buf->buf, hwctx->alloc);
2448 vkFreeMemory(hwctx->act_dev, buf->mem, hwctx->alloc);
2451 static int create_buf(AVHWDeviceContext *ctx, ImageBuffer *buf, int height,
2452 int *stride, VkBufferUsageFlags usage,
2453 VkMemoryPropertyFlagBits flags, void *create_pnext,
2458 VkMemoryRequirements req;
2459 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2460 VulkanDevicePriv *p = ctx->internal->priv;
2462 VkBufferCreateInfo buf_spawn = {
2463 .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
2464 .pNext = create_pnext,
2466 .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
2469 *stride = FFALIGN(*stride, p->props.limits.optimalBufferCopyRowPitchAlignment);
2470 buf_spawn.size = height*(*stride);
2472 ret = vkCreateBuffer(hwctx->act_dev, &buf_spawn, NULL, &buf->buf);
2473 if (ret != VK_SUCCESS) {
2474 av_log(ctx, AV_LOG_ERROR, "Failed to create buffer: %s\n",
2476 return AVERROR_EXTERNAL;
2479 vkGetBufferMemoryRequirements(hwctx->act_dev, buf->buf, &req);
2481 err = alloc_mem(ctx, &req, flags, alloc_pnext, &buf->flags, &buf->mem);
2485 ret = vkBindBufferMemory(hwctx->act_dev, buf->buf, buf->mem, 0);
2486 if (ret != VK_SUCCESS) {
2487 av_log(ctx, AV_LOG_ERROR, "Failed to bind memory to buffer: %s\n",
2490 return AVERROR_EXTERNAL;
2496 static int map_buffers(AVHWDeviceContext *ctx, ImageBuffer *buf, uint8_t *mem[],
2497 int nb_buffers, int invalidate)
2500 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2501 VkMappedMemoryRange invalidate_ctx[AV_NUM_DATA_POINTERS];
2502 int invalidate_count = 0;
2504 for (int i = 0; i < nb_buffers; i++) {
2505 ret = vkMapMemory(hwctx->act_dev, buf[i].mem, 0,
2506 VK_WHOLE_SIZE, 0, (void **)&mem[i]);
2507 if (ret != VK_SUCCESS) {
2508 av_log(ctx, AV_LOG_ERROR, "Failed to map buffer memory: %s\n",
2510 return AVERROR_EXTERNAL;
2517 for (int i = 0; i < nb_buffers; i++) {
2518 const VkMappedMemoryRange ival_buf = {
2519 .sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE,
2520 .memory = buf[i].mem,
2521 .size = VK_WHOLE_SIZE,
2523 if (buf[i].flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)
2525 invalidate_ctx[invalidate_count++] = ival_buf;
2528 if (invalidate_count) {
2529 ret = vkInvalidateMappedMemoryRanges(hwctx->act_dev, invalidate_count,
2531 if (ret != VK_SUCCESS)
2532 av_log(ctx, AV_LOG_WARNING, "Failed to invalidate memory: %s\n",
2539 static int unmap_buffers(AVHWDeviceContext *ctx, ImageBuffer *buf,
2540 int nb_buffers, int flush)
2544 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2545 VkMappedMemoryRange flush_ctx[AV_NUM_DATA_POINTERS];
2546 int flush_count = 0;
2549 for (int i = 0; i < nb_buffers; i++) {
2550 const VkMappedMemoryRange flush_buf = {
2551 .sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE,
2552 .memory = buf[i].mem,
2553 .size = VK_WHOLE_SIZE,
2555 if (buf[i].flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)
2557 flush_ctx[flush_count++] = flush_buf;
2562 ret = vkFlushMappedMemoryRanges(hwctx->act_dev, flush_count, flush_ctx);
2563 if (ret != VK_SUCCESS) {
2564 av_log(ctx, AV_LOG_ERROR, "Failed to flush memory: %s\n",
2566 err = AVERROR_EXTERNAL; /* We still want to try to unmap them */
2570 for (int i = 0; i < nb_buffers; i++)
2571 vkUnmapMemory(hwctx->act_dev, buf[i].mem);
2576 static int transfer_image_buf(AVHWDeviceContext *ctx, AVVkFrame *frame,
2577 ImageBuffer *buffer, const int *buf_stride, int w,
2578 int h, enum AVPixelFormat pix_fmt, int to_buf)
2581 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2582 VulkanDevicePriv *s = ctx->internal->priv;
2585 VkPipelineStageFlagBits sem_wait_dst[AV_NUM_DATA_POINTERS];
2587 const int planes = av_pix_fmt_count_planes(pix_fmt);
2588 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(pix_fmt);
2590 VkCommandBufferBeginInfo cmd_start = {
2591 .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
2592 .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
2595 VkImageMemoryBarrier img_bar[AV_NUM_DATA_POINTERS] = { 0 };
2597 VkSubmitInfo s_info = {
2598 .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
2599 .commandBufferCount = 1,
2600 .pCommandBuffers = &s->cmd.buf,
2601 .pSignalSemaphores = frame->sem,
2602 .pWaitSemaphores = frame->sem,
2603 .pWaitDstStageMask = sem_wait_dst,
2604 .signalSemaphoreCount = planes,
2605 .waitSemaphoreCount = planes,
2608 ret = vkBeginCommandBuffer(s->cmd.buf, &cmd_start);
2609 if (ret != VK_SUCCESS) {
2610 av_log(ctx, AV_LOG_ERROR, "Unable to init command buffer: %s\n",
2612 return AVERROR_EXTERNAL;
2615 /* Change the image layout to something more optimal for transfers */
2616 for (int i = 0; i < planes; i++) {
2617 VkImageLayout new_layout = to_buf ? VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL :
2618 VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
2619 VkAccessFlags new_access = to_buf ? VK_ACCESS_TRANSFER_READ_BIT :
2620 VK_ACCESS_TRANSFER_WRITE_BIT;
2622 sem_wait_dst[i] = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
2624 /* If the layout matches and we have read access skip the barrier */
2625 if ((frame->layout[i] == new_layout) && (frame->access[i] & new_access))
2628 img_bar[bar_num].sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
2629 img_bar[bar_num].srcAccessMask = 0x0;
2630 img_bar[bar_num].dstAccessMask = new_access;
2631 img_bar[bar_num].oldLayout = frame->layout[i];
2632 img_bar[bar_num].newLayout = new_layout;
2633 img_bar[bar_num].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
2634 img_bar[bar_num].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
2635 img_bar[bar_num].image = frame->img[i];
2636 img_bar[bar_num].subresourceRange.levelCount = 1;
2637 img_bar[bar_num].subresourceRange.layerCount = 1;
2638 img_bar[bar_num].subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
2640 frame->layout[i] = img_bar[bar_num].newLayout;
2641 frame->access[i] = img_bar[bar_num].dstAccessMask;
2647 vkCmdPipelineBarrier(s->cmd.buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
2648 VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
2649 0, NULL, 0, NULL, bar_num, img_bar);
2651 /* Schedule a copy for each plane */
2652 for (int i = 0; i < planes; i++) {
2653 const int p_w = i > 0 ? AV_CEIL_RSHIFT(w, desc->log2_chroma_w) : w;
2654 const int p_h = i > 0 ? AV_CEIL_RSHIFT(h, desc->log2_chroma_h) : h;
2655 VkBufferImageCopy buf_reg = {
2657 /* Buffer stride isn't in bytes, it's in samples, the implementation
2658 * uses the image's VkFormat to know how many bytes per sample
2659 * the buffer has. So we have to convert by dividing. Stupid.
2660 * Won't work with YUVA or other planar formats with alpha. */
2661 .bufferRowLength = buf_stride[i] / desc->comp[i].step,
2662 .bufferImageHeight = p_h,
2663 .imageSubresource.layerCount = 1,
2664 .imageSubresource.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
2665 .imageOffset = { 0, 0, 0, },
2666 .imageExtent = { p_w, p_h, 1, },
2670 vkCmdCopyImageToBuffer(s->cmd.buf, frame->img[i], frame->layout[i],
2671 buffer[i].buf, 1, &buf_reg);
2673 vkCmdCopyBufferToImage(s->cmd.buf, buffer[i].buf, frame->img[i],
2674 frame->layout[i], 1, &buf_reg);
2677 ret = vkEndCommandBuffer(s->cmd.buf);
2678 if (ret != VK_SUCCESS) {
2679 av_log(ctx, AV_LOG_ERROR, "Unable to finish command buffer: %s\n",
2681 return AVERROR_EXTERNAL;
2684 /* Wait for the download/upload to finish if uploading, otherwise the
2685 * semaphore will take care of synchronization when uploading */
2686 ret = vkQueueSubmit(s->cmd.queue, 1, &s_info, s->cmd.fence);
2687 if (ret != VK_SUCCESS) {
2688 av_log(ctx, AV_LOG_ERROR, "Unable to submit command buffer: %s\n",
2690 return AVERROR_EXTERNAL;
2692 vkWaitForFences(hwctx->act_dev, 1, &s->cmd.fence, VK_TRUE, UINT64_MAX);
2693 vkResetFences(hwctx->act_dev, 1, &s->cmd.fence);
2699 /* Technically we can use VK_EXT_external_memory_host to upload and download,
2700 * however the alignment requirements make this unfeasible as both the pointer
2701 * and the size of each plane need to be aligned to the minimum alignment
2702 * requirement, which on all current implementations (anv, radv) is 4096.
2703 * If the requirement gets relaxed (unlikely) this can easily be implemented. */
2704 static int vulkan_transfer_data_from_mem(AVHWFramesContext *hwfc, AVFrame *dst,
2709 AVVkFrame *f = (AVVkFrame *)dst->data[0];
2710 AVHWDeviceContext *dev_ctx = hwfc->device_ctx;
2711 ImageBuffer buf[AV_NUM_DATA_POINTERS] = { { 0 } };
2712 const int planes = av_pix_fmt_count_planes(src->format);
2713 int log2_chroma = av_pix_fmt_desc_get(src->format)->log2_chroma_h;
2715 if ((src->format != AV_PIX_FMT_NONE && !av_vkfmt_from_pixfmt(src->format))) {
2716 av_log(hwfc, AV_LOG_ERROR, "Unsupported source pixel format!\n");
2717 return AVERROR(EINVAL);
2720 if (src->width > hwfc->width || src->height > hwfc->height)
2721 return AVERROR(EINVAL);
2723 /* For linear, host visiable images */
2724 if (f->tiling == VK_IMAGE_TILING_LINEAR &&
2725 f->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) {
2726 AVFrame *map = av_frame_alloc();
2728 return AVERROR(ENOMEM);
2729 map->format = src->format;
2731 err = vulkan_map_frame_to_mem(hwfc, map, dst, AV_HWFRAME_MAP_WRITE);
2735 err = av_frame_copy(map, src);
2736 av_frame_free(&map);
2740 /* Create buffers */
2741 for (int i = 0; i < planes; i++) {
2742 int h = src->height;
2743 int p_height = i > 0 ? AV_CEIL_RSHIFT(h, log2_chroma) : h;
2745 tmp.linesize[i] = FFABS(src->linesize[i]);
2746 err = create_buf(dev_ctx, &buf[i], p_height,
2747 &tmp.linesize[i], VK_BUFFER_USAGE_TRANSFER_SRC_BIT,
2748 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT, NULL, NULL);
2753 /* Map, copy image to buffer, unmap */
2754 if ((err = map_buffers(dev_ctx, buf, tmp.data, planes, 0)))
2757 av_image_copy(tmp.data, tmp.linesize, (const uint8_t **)src->data,
2758 src->linesize, src->format, src->width, src->height);
2760 if ((err = unmap_buffers(dev_ctx, buf, planes, 1)))
2763 /* Copy buffers to image */
2764 err = transfer_image_buf(dev_ctx, f, buf, tmp.linesize,
2765 src->width, src->height, src->format, 0);
2768 for (int i = 0; i < planes; i++)
2769 free_buf(dev_ctx, &buf[i]);
2774 static int vulkan_transfer_data_to(AVHWFramesContext *hwfc, AVFrame *dst,
2777 av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2779 switch (src->format) {
2781 case AV_PIX_FMT_CUDA:
2782 if ((p->extensions & EXT_EXTERNAL_FD_MEMORY) &&
2783 (p->extensions & EXT_EXTERNAL_FD_SEM))
2784 return vulkan_transfer_data_from_cuda(hwfc, dst, src);
2787 if (src->hw_frames_ctx)
2788 return AVERROR(ENOSYS);
2790 return vulkan_transfer_data_from_mem(hwfc, dst, src);
2795 static int vulkan_transfer_data_to_cuda(AVHWFramesContext *hwfc, AVFrame *dst,
2802 AVVkFrameInternal *dst_int;
2803 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
2804 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
2806 AVHWFramesContext *cuda_fc = (AVHWFramesContext*)dst->hw_frames_ctx->data;
2807 AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
2808 AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
2809 AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
2810 CudaFunctions *cu = cu_internal->cuda_dl;
2812 ret = CHECK_CU(cu->cuCtxPushCurrent(cuda_dev->cuda_ctx));
2814 err = AVERROR_EXTERNAL;
2818 dst_f = (AVVkFrame *)src->data[0];
2820 err = vulkan_export_to_cuda(hwfc, dst->hw_frames_ctx, src);
2825 dst_int = dst_f->internal;
2827 for (int i = 0; i < planes; i++) {
2828 CUDA_MEMCPY2D cpy = {
2829 .dstMemoryType = CU_MEMORYTYPE_DEVICE,
2830 .dstDevice = (CUdeviceptr)dst->data[i],
2831 .dstPitch = dst->linesize[i],
2834 .srcMemoryType = CU_MEMORYTYPE_ARRAY,
2835 .srcArray = dst_int->cu_array[i],
2836 .WidthInBytes = (i > 0 ? AV_CEIL_RSHIFT(hwfc->width, desc->log2_chroma_w)
2837 : hwfc->width) * desc->comp[i].step,
2838 .Height = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, desc->log2_chroma_h)
2842 ret = CHECK_CU(cu->cuMemcpy2DAsync(&cpy, cuda_dev->stream));
2844 err = AVERROR_EXTERNAL;
2849 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2851 av_log(hwfc, AV_LOG_VERBOSE, "Transfered Vulkan image to CUDA!\n");
2856 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2857 vulkan_free_internal(dst_int);
2858 dst_f->internal = NULL;
2859 av_buffer_unref(&dst->buf[0]);
2864 static int vulkan_transfer_data_to_mem(AVHWFramesContext *hwfc, AVFrame *dst,
2869 AVVkFrame *f = (AVVkFrame *)src->data[0];
2870 AVHWDeviceContext *dev_ctx = hwfc->device_ctx;
2871 ImageBuffer buf[AV_NUM_DATA_POINTERS] = { { 0 } };
2872 const int planes = av_pix_fmt_count_planes(dst->format);
2873 int log2_chroma = av_pix_fmt_desc_get(dst->format)->log2_chroma_h;
2875 if (dst->width > hwfc->width || dst->height > hwfc->height)
2876 return AVERROR(EINVAL);
2878 /* For linear, host visiable images */
2879 if (f->tiling == VK_IMAGE_TILING_LINEAR &&
2880 f->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) {
2881 AVFrame *map = av_frame_alloc();
2883 return AVERROR(ENOMEM);
2884 map->format = dst->format;
2886 err = vulkan_map_frame_to_mem(hwfc, map, src, AV_HWFRAME_MAP_READ);
2890 err = av_frame_copy(dst, map);
2891 av_frame_free(&map);
2895 /* Create buffers */
2896 for (int i = 0; i < planes; i++) {
2897 int h = dst->height;
2898 int p_height = i > 0 ? AV_CEIL_RSHIFT(h, log2_chroma) : h;
2900 tmp.linesize[i] = FFABS(dst->linesize[i]);
2901 err = create_buf(dev_ctx, &buf[i], p_height,
2902 &tmp.linesize[i], VK_BUFFER_USAGE_TRANSFER_DST_BIT,
2903 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT, NULL, NULL);
2906 /* Copy image to buffer */
2907 if ((err = transfer_image_buf(dev_ctx, f, buf, tmp.linesize,
2908 dst->width, dst->height, dst->format, 1)))
2911 /* Map, copy buffer to frame, unmap */
2912 if ((err = map_buffers(dev_ctx, buf, tmp.data, planes, 1)))
2915 av_image_copy(dst->data, dst->linesize, (const uint8_t **)tmp.data,
2916 tmp.linesize, dst->format, dst->width, dst->height);
2918 err = unmap_buffers(dev_ctx, buf, planes, 0);
2921 for (int i = 0; i < planes; i++)
2922 free_buf(dev_ctx, &buf[i]);
2927 static int vulkan_transfer_data_from(AVHWFramesContext *hwfc, AVFrame *dst,
2930 av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2932 switch (dst->format) {
2934 case AV_PIX_FMT_CUDA:
2935 if ((p->extensions & EXT_EXTERNAL_FD_MEMORY) &&
2936 (p->extensions & EXT_EXTERNAL_FD_SEM))
2937 return vulkan_transfer_data_to_cuda(hwfc, dst, src);
2940 if (dst->hw_frames_ctx)
2941 return AVERROR(ENOSYS);
2943 return vulkan_transfer_data_to_mem(hwfc, dst, src);
2947 AVVkFrame *av_vk_frame_alloc(void)
2949 return av_mallocz(sizeof(AVVkFrame));
2952 const HWContextType ff_hwcontext_type_vulkan = {
2953 .type = AV_HWDEVICE_TYPE_VULKAN,
2956 .device_hwctx_size = sizeof(AVVulkanDeviceContext),
2957 .device_priv_size = sizeof(VulkanDevicePriv),
2958 .frames_hwctx_size = sizeof(AVVulkanFramesContext),
2959 .frames_priv_size = sizeof(VulkanFramesPriv),
2961 .device_init = &vulkan_device_init,
2962 .device_create = &vulkan_device_create,
2963 .device_derive = &vulkan_device_derive,
2965 .frames_get_constraints = &vulkan_frames_get_constraints,
2966 .frames_init = vulkan_frames_init,
2967 .frames_get_buffer = vulkan_get_buffer,
2968 .frames_uninit = vulkan_frames_uninit,
2970 .transfer_get_formats = vulkan_transfer_get_formats,
2971 .transfer_data_to = vulkan_transfer_data_to,
2972 .transfer_data_from = vulkan_transfer_data_from,
2974 .map_to = vulkan_map_to,
2975 .map_from = vulkan_map_from,
2977 .pix_fmts = (const enum AVPixelFormat []) {