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;
57 VkDebugUtilsMessengerEXT debug_ctx;
66 int use_linear_images;
72 typedef struct VulkanFramesPriv {
76 typedef struct AVVkFrameInternal {
78 /* Importing external memory into cuda is really expensive so we keep the
79 * memory imported all the time */
80 AVBufferRef *cuda_fc_ref; /* Need to keep it around for uninit */
81 CUexternalMemory ext_mem[AV_NUM_DATA_POINTERS];
82 CUmipmappedArray cu_mma[AV_NUM_DATA_POINTERS];
83 CUarray cu_array[AV_NUM_DATA_POINTERS];
84 CUexternalSemaphore cu_sem[AV_NUM_DATA_POINTERS];
88 #define VK_LOAD_PFN(inst, name) PFN_##name pfn_##name = (PFN_##name) \
89 vkGetInstanceProcAddr(inst, #name)
91 #define DEFAULT_USAGE_FLAGS (VK_IMAGE_USAGE_SAMPLED_BIT | \
92 VK_IMAGE_USAGE_STORAGE_BIT | \
93 VK_IMAGE_USAGE_TRANSFER_SRC_BIT | \
94 VK_IMAGE_USAGE_TRANSFER_DST_BIT)
96 #define ADD_VAL_TO_LIST(list, count, val) \
98 list = av_realloc_array(list, sizeof(*list), ++count); \
100 err = AVERROR(ENOMEM); \
103 list[count - 1] = val; \
106 static const struct {
107 enum AVPixelFormat pixfmt;
108 const VkFormat vkfmts[3];
109 } vk_pixfmt_map[] = {
110 { AV_PIX_FMT_GRAY8, { VK_FORMAT_R8_UNORM } },
111 { AV_PIX_FMT_GRAY16, { VK_FORMAT_R16_UNORM } },
112 { AV_PIX_FMT_GRAYF32, { VK_FORMAT_R32_SFLOAT } },
114 { AV_PIX_FMT_NV12, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8G8_UNORM } },
115 { AV_PIX_FMT_P010, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16G16_UNORM } },
116 { AV_PIX_FMT_P016, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16G16_UNORM } },
118 { AV_PIX_FMT_YUV420P, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } },
119 { AV_PIX_FMT_YUV422P, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } },
120 { AV_PIX_FMT_YUV444P, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } },
122 { AV_PIX_FMT_YUV420P16, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
123 { AV_PIX_FMT_YUV422P16, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
124 { AV_PIX_FMT_YUV444P16, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
126 { AV_PIX_FMT_ABGR, { VK_FORMAT_A8B8G8R8_UNORM_PACK32 } },
127 { AV_PIX_FMT_BGRA, { VK_FORMAT_B8G8R8A8_UNORM } },
128 { AV_PIX_FMT_RGBA, { VK_FORMAT_R8G8B8A8_UNORM } },
129 { AV_PIX_FMT_RGB24, { VK_FORMAT_R8G8B8_UNORM } },
130 { AV_PIX_FMT_BGR24, { VK_FORMAT_B8G8R8_UNORM } },
131 { AV_PIX_FMT_RGB48, { VK_FORMAT_R16G16B16_UNORM } },
132 { AV_PIX_FMT_RGBA64, { VK_FORMAT_R16G16B16A16_UNORM } },
133 { AV_PIX_FMT_RGB565, { VK_FORMAT_R5G6B5_UNORM_PACK16 } },
134 { AV_PIX_FMT_BGR565, { VK_FORMAT_B5G6R5_UNORM_PACK16 } },
135 { AV_PIX_FMT_BGR0, { VK_FORMAT_B8G8R8A8_UNORM } },
136 { AV_PIX_FMT_0BGR, { VK_FORMAT_A8B8G8R8_UNORM_PACK32 } },
137 { AV_PIX_FMT_RGB0, { VK_FORMAT_R8G8B8A8_UNORM } },
139 { AV_PIX_FMT_GBRPF32, { VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_SFLOAT } },
142 const VkFormat *av_vkfmt_from_pixfmt(enum AVPixelFormat p)
144 for (enum AVPixelFormat i = 0; i < FF_ARRAY_ELEMS(vk_pixfmt_map); i++)
145 if (vk_pixfmt_map[i].pixfmt == p)
146 return vk_pixfmt_map[i].vkfmts;
150 static int pixfmt_is_supported(AVVulkanDeviceContext *hwctx, enum AVPixelFormat p,
153 const VkFormat *fmt = av_vkfmt_from_pixfmt(p);
154 int planes = av_pix_fmt_count_planes(p);
159 for (int i = 0; i < planes; i++) {
160 VkFormatFeatureFlags flags;
161 VkFormatProperties2 prop = {
162 .sType = VK_STRUCTURE_TYPE_FORMAT_PROPERTIES_2,
164 vkGetPhysicalDeviceFormatProperties2(hwctx->phys_dev, fmt[i], &prop);
165 flags = linear ? prop.formatProperties.linearTilingFeatures :
166 prop.formatProperties.optimalTilingFeatures;
167 if (!(flags & DEFAULT_USAGE_FLAGS))
174 enum VulkanExtensions {
175 EXT_EXTERNAL_DMABUF_MEMORY = 1ULL << 0, /* VK_EXT_external_memory_dma_buf */
176 EXT_DRM_MODIFIER_FLAGS = 1ULL << 1, /* VK_EXT_image_drm_format_modifier */
177 EXT_EXTERNAL_FD_MEMORY = 1ULL << 2, /* VK_KHR_external_memory_fd */
178 EXT_EXTERNAL_FD_SEM = 1ULL << 3, /* VK_KHR_external_semaphore_fd */
180 EXT_OPTIONAL = 1ULL << 62,
181 EXT_REQUIRED = 1ULL << 63,
184 typedef struct VulkanOptExtension {
187 } VulkanOptExtension;
189 static const VulkanOptExtension optional_instance_exts[] = {
193 static const VulkanOptExtension optional_device_exts[] = {
194 { VK_KHR_EXTERNAL_MEMORY_FD_EXTENSION_NAME, EXT_EXTERNAL_FD_MEMORY, },
195 { VK_EXT_EXTERNAL_MEMORY_DMA_BUF_EXTENSION_NAME, EXT_EXTERNAL_DMABUF_MEMORY, },
196 { VK_EXT_IMAGE_DRM_FORMAT_MODIFIER_EXTENSION_NAME, EXT_DRM_MODIFIER_FLAGS, },
197 { VK_KHR_EXTERNAL_SEMAPHORE_FD_EXTENSION_NAME, EXT_EXTERNAL_FD_SEM, },
200 /* Converts return values to strings */
201 static const char *vk_ret2str(VkResult res)
203 #define CASE(VAL) case VAL: return #VAL
209 CASE(VK_EVENT_RESET);
211 CASE(VK_ERROR_OUT_OF_HOST_MEMORY);
212 CASE(VK_ERROR_OUT_OF_DEVICE_MEMORY);
213 CASE(VK_ERROR_INITIALIZATION_FAILED);
214 CASE(VK_ERROR_DEVICE_LOST);
215 CASE(VK_ERROR_MEMORY_MAP_FAILED);
216 CASE(VK_ERROR_LAYER_NOT_PRESENT);
217 CASE(VK_ERROR_EXTENSION_NOT_PRESENT);
218 CASE(VK_ERROR_FEATURE_NOT_PRESENT);
219 CASE(VK_ERROR_INCOMPATIBLE_DRIVER);
220 CASE(VK_ERROR_TOO_MANY_OBJECTS);
221 CASE(VK_ERROR_FORMAT_NOT_SUPPORTED);
222 CASE(VK_ERROR_FRAGMENTED_POOL);
223 CASE(VK_ERROR_SURFACE_LOST_KHR);
224 CASE(VK_ERROR_NATIVE_WINDOW_IN_USE_KHR);
225 CASE(VK_SUBOPTIMAL_KHR);
226 CASE(VK_ERROR_OUT_OF_DATE_KHR);
227 CASE(VK_ERROR_INCOMPATIBLE_DISPLAY_KHR);
228 CASE(VK_ERROR_VALIDATION_FAILED_EXT);
229 CASE(VK_ERROR_INVALID_SHADER_NV);
230 CASE(VK_ERROR_OUT_OF_POOL_MEMORY);
231 CASE(VK_ERROR_INVALID_EXTERNAL_HANDLE);
232 CASE(VK_ERROR_NOT_PERMITTED_EXT);
233 CASE(VK_ERROR_INVALID_DRM_FORMAT_MODIFIER_PLANE_LAYOUT_EXT);
234 CASE(VK_ERROR_INVALID_DEVICE_ADDRESS_EXT);
235 CASE(VK_ERROR_FULL_SCREEN_EXCLUSIVE_MODE_LOST_EXT);
236 default: return "Unknown error";
241 static VkBool32 vk_dbg_callback(VkDebugUtilsMessageSeverityFlagBitsEXT severity,
242 VkDebugUtilsMessageTypeFlagsEXT messageType,
243 const VkDebugUtilsMessengerCallbackDataEXT *data,
247 AVHWDeviceContext *ctx = priv;
250 case VK_DEBUG_UTILS_MESSAGE_SEVERITY_VERBOSE_BIT_EXT: l = AV_LOG_VERBOSE; break;
251 case VK_DEBUG_UTILS_MESSAGE_SEVERITY_INFO_BIT_EXT: l = AV_LOG_INFO; break;
252 case VK_DEBUG_UTILS_MESSAGE_SEVERITY_WARNING_BIT_EXT: l = AV_LOG_WARNING; break;
253 case VK_DEBUG_UTILS_MESSAGE_SEVERITY_ERROR_BIT_EXT: l = AV_LOG_ERROR; break;
254 default: l = AV_LOG_DEBUG; break;
257 av_log(ctx, l, "%s\n", data->pMessage);
258 for (int i = 0; i < data->cmdBufLabelCount; i++)
259 av_log(ctx, l, "\t%i: %s\n", i, data->pCmdBufLabels[i].pLabelName);
264 static int check_extensions(AVHWDeviceContext *ctx, int dev,
265 const char * const **dst, uint32_t *num, int debug)
268 const char **extension_names = NULL;
269 VulkanDevicePriv *p = ctx->internal->priv;
270 AVVulkanDeviceContext *hwctx = ctx->hwctx;
271 int err = 0, found, extensions_found = 0;
274 int optional_exts_num;
275 uint32_t sup_ext_count;
276 VkExtensionProperties *sup_ext;
277 const VulkanOptExtension *optional_exts;
281 optional_exts = optional_instance_exts;
282 optional_exts_num = FF_ARRAY_ELEMS(optional_instance_exts);
283 vkEnumerateInstanceExtensionProperties(NULL, &sup_ext_count, NULL);
284 sup_ext = av_malloc_array(sup_ext_count, sizeof(VkExtensionProperties));
286 return AVERROR(ENOMEM);
287 vkEnumerateInstanceExtensionProperties(NULL, &sup_ext_count, sup_ext);
290 optional_exts = optional_device_exts;
291 optional_exts_num = FF_ARRAY_ELEMS(optional_device_exts);
292 vkEnumerateDeviceExtensionProperties(hwctx->phys_dev, NULL,
293 &sup_ext_count, NULL);
294 sup_ext = av_malloc_array(sup_ext_count, sizeof(VkExtensionProperties));
296 return AVERROR(ENOMEM);
297 vkEnumerateDeviceExtensionProperties(hwctx->phys_dev, NULL,
298 &sup_ext_count, sup_ext);
301 for (int i = 0; i < optional_exts_num; i++) {
302 int req = optional_exts[i].flag & EXT_REQUIRED;
303 tstr = optional_exts[i].name;
306 for (int j = 0; j < sup_ext_count; j++) {
307 if (!strcmp(tstr, sup_ext[j].extensionName)) {
313 int lvl = req ? AV_LOG_ERROR : AV_LOG_VERBOSE;
314 av_log(ctx, lvl, "Extension \"%s\" not found!\n", tstr);
316 err = AVERROR(EINVAL);
322 p->extensions |= optional_exts[i].flag;
324 av_log(ctx, AV_LOG_VERBOSE, "Using %s extension \"%s\"\n", mod, tstr);
326 ADD_VAL_TO_LIST(extension_names, extensions_found, tstr);
330 tstr = VK_EXT_DEBUG_UTILS_EXTENSION_NAME;
332 for (int j = 0; j < sup_ext_count; j++) {
333 if (!strcmp(tstr, sup_ext[j].extensionName)) {
339 ADD_VAL_TO_LIST(extension_names, extensions_found, tstr);
341 av_log(ctx, AV_LOG_ERROR, "Debug extension \"%s\" not found!\n",
343 err = AVERROR(EINVAL);
348 *dst = extension_names;
349 *num = extensions_found;
356 /* Creates a VkInstance */
357 static int create_instance(AVHWDeviceContext *ctx, AVDictionary *opts)
361 VulkanDevicePriv *p = ctx->internal->priv;
362 AVVulkanDeviceContext *hwctx = ctx->hwctx;
363 AVDictionaryEntry *debug_opt = av_dict_get(opts, "debug", NULL, 0);
364 const int debug_mode = debug_opt && strtol(debug_opt->value, NULL, 10);
365 VkApplicationInfo application_info = {
366 .sType = VK_STRUCTURE_TYPE_APPLICATION_INFO,
367 .pEngineName = "libavutil",
368 .apiVersion = VK_API_VERSION_1_1,
369 .engineVersion = VK_MAKE_VERSION(LIBAVUTIL_VERSION_MAJOR,
370 LIBAVUTIL_VERSION_MINOR,
371 LIBAVUTIL_VERSION_MICRO),
373 VkInstanceCreateInfo inst_props = {
374 .sType = VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO,
375 .pApplicationInfo = &application_info,
378 /* Check for present/missing extensions */
379 err = check_extensions(ctx, 0, &inst_props.ppEnabledExtensionNames,
380 &inst_props.enabledExtensionCount, debug_mode);
385 static const char *layers[] = { "VK_LAYER_LUNARG_standard_validation" };
386 inst_props.ppEnabledLayerNames = layers;
387 inst_props.enabledLayerCount = FF_ARRAY_ELEMS(layers);
390 /* Try to create the instance */
391 ret = vkCreateInstance(&inst_props, hwctx->alloc, &hwctx->inst);
393 /* Free used memory */
394 av_free((void *)inst_props.ppEnabledExtensionNames);
396 /* Check for errors */
397 if (ret != VK_SUCCESS) {
398 av_log(ctx, AV_LOG_ERROR, "Instance creation failure: %s\n",
400 return AVERROR_EXTERNAL;
404 VkDebugUtilsMessengerCreateInfoEXT dbg = {
405 .sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_MESSENGER_CREATE_INFO_EXT,
406 .messageSeverity = VK_DEBUG_UTILS_MESSAGE_SEVERITY_VERBOSE_BIT_EXT |
407 VK_DEBUG_UTILS_MESSAGE_SEVERITY_INFO_BIT_EXT |
408 VK_DEBUG_UTILS_MESSAGE_SEVERITY_WARNING_BIT_EXT |
409 VK_DEBUG_UTILS_MESSAGE_SEVERITY_ERROR_BIT_EXT,
410 .messageType = VK_DEBUG_UTILS_MESSAGE_TYPE_GENERAL_BIT_EXT |
411 VK_DEBUG_UTILS_MESSAGE_TYPE_VALIDATION_BIT_EXT |
412 VK_DEBUG_UTILS_MESSAGE_TYPE_PERFORMANCE_BIT_EXT,
413 .pfnUserCallback = vk_dbg_callback,
416 VK_LOAD_PFN(hwctx->inst, vkCreateDebugUtilsMessengerEXT);
418 pfn_vkCreateDebugUtilsMessengerEXT(hwctx->inst, &dbg,
419 hwctx->alloc, &p->debug_ctx);
425 typedef struct VulkanDeviceSelection {
426 uint8_t uuid[VK_UUID_SIZE]; /* Will use this first unless !has_uuid */
428 const char *name; /* Will use this second unless NULL */
429 uint32_t pci_device; /* Will use this third unless 0x0 */
430 uint32_t vendor_id; /* Last resort to find something deterministic */
431 int index; /* Finally fall back to index */
432 } VulkanDeviceSelection;
434 static const char *vk_dev_type(enum VkPhysicalDeviceType type)
437 case VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU: return "integrated";
438 case VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU: return "discrete";
439 case VK_PHYSICAL_DEVICE_TYPE_VIRTUAL_GPU: return "virtual";
440 case VK_PHYSICAL_DEVICE_TYPE_CPU: return "software";
441 default: return "unknown";
446 static int find_device(AVHWDeviceContext *ctx, VulkanDeviceSelection *select)
448 int err = 0, choice = -1;
451 VkPhysicalDevice *devices = NULL;
452 VkPhysicalDeviceIDProperties *idp = NULL;
453 VkPhysicalDeviceProperties2 *prop = NULL;
454 VulkanDevicePriv *p = ctx->internal->priv;
455 AVVulkanDeviceContext *hwctx = ctx->hwctx;
457 ret = vkEnumeratePhysicalDevices(hwctx->inst, &num, NULL);
458 if (ret != VK_SUCCESS || !num) {
459 av_log(ctx, AV_LOG_ERROR, "No devices found: %s!\n", vk_ret2str(ret));
460 return AVERROR(ENODEV);
463 devices = av_malloc_array(num, sizeof(VkPhysicalDevice));
465 return AVERROR(ENOMEM);
467 ret = vkEnumeratePhysicalDevices(hwctx->inst, &num, devices);
468 if (ret != VK_SUCCESS) {
469 av_log(ctx, AV_LOG_ERROR, "Failed enumerating devices: %s\n",
471 err = AVERROR(ENODEV);
475 prop = av_mallocz_array(num, sizeof(*prop));
477 err = AVERROR(ENOMEM);
481 idp = av_mallocz_array(num, sizeof(*idp));
483 err = AVERROR(ENOMEM);
487 av_log(ctx, AV_LOG_VERBOSE, "GPU listing:\n");
488 for (int i = 0; i < num; i++) {
489 idp[i].sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES;
490 prop[i].sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2;
491 prop[i].pNext = &idp[i];
493 vkGetPhysicalDeviceProperties2(devices[i], &prop[i]);
494 av_log(ctx, AV_LOG_VERBOSE, " %d: %s (%s) (0x%x)\n", i,
495 prop[i].properties.deviceName,
496 vk_dev_type(prop[i].properties.deviceType),
497 prop[i].properties.deviceID);
500 if (select->has_uuid) {
501 for (int i = 0; i < num; i++) {
502 if (!strncmp(idp[i].deviceUUID, select->uuid, VK_UUID_SIZE)) {
507 av_log(ctx, AV_LOG_ERROR, "Unable to find device by given UUID!\n");
508 err = AVERROR(ENODEV);
510 } else if (select->name) {
511 av_log(ctx, AV_LOG_VERBOSE, "Requested device: %s\n", select->name);
512 for (int i = 0; i < num; i++) {
513 if (strstr(prop[i].properties.deviceName, select->name)) {
518 av_log(ctx, AV_LOG_ERROR, "Unable to find device \"%s\"!\n",
520 err = AVERROR(ENODEV);
522 } else if (select->pci_device) {
523 av_log(ctx, AV_LOG_VERBOSE, "Requested device: 0x%x\n", select->pci_device);
524 for (int i = 0; i < num; i++) {
525 if (select->pci_device == prop[i].properties.deviceID) {
530 av_log(ctx, AV_LOG_ERROR, "Unable to find device with PCI ID 0x%x!\n",
532 err = AVERROR(EINVAL);
534 } else if (select->vendor_id) {
535 av_log(ctx, AV_LOG_VERBOSE, "Requested vendor: 0x%x\n", select->vendor_id);
536 for (int i = 0; i < num; i++) {
537 if (select->vendor_id == prop[i].properties.vendorID) {
542 av_log(ctx, AV_LOG_ERROR, "Unable to find device with Vendor ID 0x%x!\n",
544 err = AVERROR(ENODEV);
547 if (select->index < num) {
548 choice = select->index;
551 av_log(ctx, AV_LOG_ERROR, "Unable to find device with index %i!\n",
553 err = AVERROR(ENODEV);
559 p->dev_is_nvidia = (prop[choice].properties.vendorID == 0x10de);
560 hwctx->phys_dev = devices[choice];
569 static int search_queue_families(AVHWDeviceContext *ctx, VkDeviceCreateInfo *cd)
572 VkQueueFamilyProperties *qs = NULL;
573 AVVulkanDeviceContext *hwctx = ctx->hwctx;
574 int graph_index = -1, comp_index = -1, tx_index = -1;
575 VkDeviceQueueCreateInfo *pc = (VkDeviceQueueCreateInfo *)cd->pQueueCreateInfos;
577 /* First get the number of queue families */
578 vkGetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &num, NULL);
580 av_log(ctx, AV_LOG_ERROR, "Failed to get queues!\n");
581 return AVERROR_EXTERNAL;
584 /* Then allocate memory */
585 qs = av_malloc_array(num, sizeof(VkQueueFamilyProperties));
587 return AVERROR(ENOMEM);
589 /* Finally retrieve the queue families */
590 vkGetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &num, qs);
592 #define SEARCH_FLAGS(expr, out) \
593 for (int i = 0; i < num; i++) { \
594 const VkQueueFlagBits flags = qs[i].queueFlags; \
601 SEARCH_FLAGS(flags & VK_QUEUE_GRAPHICS_BIT, graph_index)
603 SEARCH_FLAGS((flags & VK_QUEUE_COMPUTE_BIT) && (i != graph_index),
606 SEARCH_FLAGS((flags & VK_QUEUE_TRANSFER_BIT) && (i != graph_index) &&
607 (i != comp_index), tx_index)
610 #define QF_FLAGS(flags) \
611 ((flags) & VK_QUEUE_GRAPHICS_BIT ) ? "(graphics) " : "", \
612 ((flags) & VK_QUEUE_COMPUTE_BIT ) ? "(compute) " : "", \
613 ((flags) & VK_QUEUE_TRANSFER_BIT ) ? "(transfer) " : "", \
614 ((flags) & VK_QUEUE_SPARSE_BINDING_BIT) ? "(sparse) " : ""
616 av_log(ctx, AV_LOG_VERBOSE, "Using queue family %i for graphics, "
617 "flags: %s%s%s%s\n", graph_index, QF_FLAGS(qs[graph_index].queueFlags));
619 hwctx->queue_family_index = graph_index;
620 hwctx->queue_family_tx_index = graph_index;
621 hwctx->queue_family_comp_index = graph_index;
623 pc[cd->queueCreateInfoCount++].queueFamilyIndex = graph_index;
625 if (comp_index != -1) {
626 av_log(ctx, AV_LOG_VERBOSE, "Using queue family %i for compute, "
627 "flags: %s%s%s%s\n", comp_index, QF_FLAGS(qs[comp_index].queueFlags));
628 hwctx->queue_family_tx_index = comp_index;
629 hwctx->queue_family_comp_index = comp_index;
630 pc[cd->queueCreateInfoCount++].queueFamilyIndex = comp_index;
633 if (tx_index != -1) {
634 av_log(ctx, AV_LOG_VERBOSE, "Using queue family %i for transfers, "
635 "flags: %s%s%s%s\n", tx_index, QF_FLAGS(qs[tx_index].queueFlags));
636 hwctx->queue_family_tx_index = tx_index;
637 pc[cd->queueCreateInfoCount++].queueFamilyIndex = tx_index;
647 static int create_exec_ctx(AVHWDeviceContext *ctx, VulkanExecCtx *cmd,
648 int queue_family_index)
651 AVVulkanDeviceContext *hwctx = ctx->hwctx;
653 VkCommandPoolCreateInfo cqueue_create = {
654 .sType = VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO,
655 .flags = VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT,
656 .queueFamilyIndex = queue_family_index,
658 VkCommandBufferAllocateInfo cbuf_create = {
659 .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO,
660 .level = VK_COMMAND_BUFFER_LEVEL_PRIMARY,
661 .commandBufferCount = 1,
664 VkFenceCreateInfo fence_spawn = {
665 .sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO,
668 ret = vkCreateFence(hwctx->act_dev, &fence_spawn,
669 hwctx->alloc, &cmd->fence);
670 if (ret != VK_SUCCESS) {
671 av_log(ctx, AV_LOG_ERROR, "Failed to create frame fence: %s\n",
673 return AVERROR_EXTERNAL;
676 ret = vkCreateCommandPool(hwctx->act_dev, &cqueue_create,
677 hwctx->alloc, &cmd->pool);
678 if (ret != VK_SUCCESS) {
679 av_log(ctx, AV_LOG_ERROR, "Command pool creation failure: %s\n",
681 return AVERROR_EXTERNAL;
684 cbuf_create.commandPool = cmd->pool;
686 ret = vkAllocateCommandBuffers(hwctx->act_dev, &cbuf_create, &cmd->buf);
687 if (ret != VK_SUCCESS) {
688 av_log(ctx, AV_LOG_ERROR, "Command buffer alloc failure: %s\n",
690 return AVERROR_EXTERNAL;
693 vkGetDeviceQueue(hwctx->act_dev, cqueue_create.queueFamilyIndex, 0,
699 static void free_exec_ctx(AVHWDeviceContext *ctx, VulkanExecCtx *cmd)
701 AVVulkanDeviceContext *hwctx = ctx->hwctx;
704 vkDestroyFence(hwctx->act_dev, cmd->fence, hwctx->alloc);
706 vkFreeCommandBuffers(hwctx->act_dev, cmd->pool, 1, &cmd->buf);
708 vkDestroyCommandPool(hwctx->act_dev, cmd->pool, hwctx->alloc);
711 static void vulkan_device_free(AVHWDeviceContext *ctx)
713 VulkanDevicePriv *p = ctx->internal->priv;
714 AVVulkanDeviceContext *hwctx = ctx->hwctx;
716 free_exec_ctx(ctx, &p->cmd);
718 vkDestroyDevice(hwctx->act_dev, hwctx->alloc);
721 VK_LOAD_PFN(hwctx->inst, vkDestroyDebugUtilsMessengerEXT);
722 pfn_vkDestroyDebugUtilsMessengerEXT(hwctx->inst, p->debug_ctx,
726 vkDestroyInstance(hwctx->inst, hwctx->alloc);
729 static int vulkan_device_create_internal(AVHWDeviceContext *ctx,
730 VulkanDeviceSelection *dev_select,
731 AVDictionary *opts, int flags)
735 AVDictionaryEntry *opt_d;
736 VulkanDevicePriv *p = ctx->internal->priv;
737 AVVulkanDeviceContext *hwctx = ctx->hwctx;
738 VkDeviceQueueCreateInfo queue_create_info[3] = {
739 { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
740 .pQueuePriorities = (float []){ 1.0f },
742 { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
743 .pQueuePriorities = (float []){ 1.0f },
745 { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
746 .pQueuePriorities = (float []){ 1.0f },
750 VkDeviceCreateInfo dev_info = {
751 .sType = VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO,
752 .pQueueCreateInfos = queue_create_info,
753 .queueCreateInfoCount = 0,
756 ctx->free = vulkan_device_free;
758 /* Create an instance if not given one */
759 if ((err = create_instance(ctx, opts)))
762 /* Find a device (if not given one) */
763 if ((err = find_device(ctx, dev_select)))
766 vkGetPhysicalDeviceProperties(hwctx->phys_dev, &p->props);
767 av_log(ctx, AV_LOG_VERBOSE, "Using device: %s\n", p->props.deviceName);
768 av_log(ctx, AV_LOG_VERBOSE, "Alignments:\n");
769 av_log(ctx, AV_LOG_VERBOSE, " optimalBufferCopyOffsetAlignment: %li\n",
770 p->props.limits.optimalBufferCopyOffsetAlignment);
771 av_log(ctx, AV_LOG_VERBOSE, " optimalBufferCopyRowPitchAlignment: %li\n",
772 p->props.limits.optimalBufferCopyRowPitchAlignment);
773 av_log(ctx, AV_LOG_VERBOSE, " minMemoryMapAlignment: %li\n",
774 p->props.limits.minMemoryMapAlignment);
776 /* Search queue family */
777 if ((err = search_queue_families(ctx, &dev_info)))
780 if ((err = check_extensions(ctx, 1, &dev_info.ppEnabledExtensionNames,
781 &dev_info.enabledExtensionCount, 0)))
784 ret = vkCreateDevice(hwctx->phys_dev, &dev_info, hwctx->alloc,
787 av_free((void *)dev_info.ppEnabledExtensionNames);
789 if (ret != VK_SUCCESS) {
790 av_log(ctx, AV_LOG_ERROR, "Device creation failure: %s\n",
792 err = AVERROR_EXTERNAL;
796 /* Tiled images setting, use them by default */
797 opt_d = av_dict_get(opts, "linear_images", NULL, 0);
799 p->use_linear_images = strtol(opt_d->value, NULL, 10);
805 static int vulkan_device_init(AVHWDeviceContext *ctx)
809 AVVulkanDeviceContext *hwctx = ctx->hwctx;
810 VulkanDevicePriv *p = ctx->internal->priv;
812 vkGetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &queue_num, NULL);
814 av_log(ctx, AV_LOG_ERROR, "Failed to get queues!\n");
815 return AVERROR_EXTERNAL;
818 #define CHECK_QUEUE(type, n) \
819 if (n >= queue_num) { \
820 av_log(ctx, AV_LOG_ERROR, "Invalid %s queue index %i (device has %i queues)!\n", \
821 type, n, queue_num); \
822 return AVERROR(EINVAL); \
825 CHECK_QUEUE("graphics", hwctx->queue_family_index)
826 CHECK_QUEUE("upload", hwctx->queue_family_tx_index)
827 CHECK_QUEUE("compute", hwctx->queue_family_comp_index)
831 /* Create exec context - if there's something invalid this will error out */
832 err = create_exec_ctx(ctx, &p->cmd, hwctx->queue_family_tx_index);
836 /* Get device capabilities */
837 vkGetPhysicalDeviceMemoryProperties(hwctx->phys_dev, &p->mprops);
842 static int vulkan_device_create(AVHWDeviceContext *ctx, const char *device,
843 AVDictionary *opts, int flags)
845 VulkanDeviceSelection dev_select = { 0 };
846 if (device && device[0]) {
848 dev_select.index = strtol(device, &end, 10);
850 dev_select.index = 0;
851 dev_select.name = device;
855 return vulkan_device_create_internal(ctx, &dev_select, opts, flags);
858 static int vulkan_device_derive(AVHWDeviceContext *ctx,
859 AVHWDeviceContext *src_ctx, int flags)
861 av_unused VulkanDeviceSelection dev_select = { 0 };
863 /* If there's only one device on the system, then even if its not covered
864 * by the following checks (e.g. non-PCIe ARM GPU), having an empty
865 * dev_select will mean it'll get picked. */
866 switch(src_ctx->type) {
869 case AV_HWDEVICE_TYPE_VAAPI: {
870 AVVAAPIDeviceContext *src_hwctx = src_ctx->hwctx;
872 const char *vendor = vaQueryVendorString(src_hwctx->display);
874 av_log(ctx, AV_LOG_ERROR, "Unable to get device info from VAAPI!\n");
875 return AVERROR_EXTERNAL;
878 if (strstr(vendor, "Intel"))
879 dev_select.vendor_id = 0x8086;
880 if (strstr(vendor, "AMD"))
881 dev_select.vendor_id = 0x1002;
883 return vulkan_device_create_internal(ctx, &dev_select, NULL, flags);
886 case AV_HWDEVICE_TYPE_DRM: {
887 AVDRMDeviceContext *src_hwctx = src_ctx->hwctx;
889 drmDevice *drm_dev_info;
890 int err = drmGetDevice(src_hwctx->fd, &drm_dev_info);
892 av_log(ctx, AV_LOG_ERROR, "Unable to get device info from DRM fd!\n");
893 return AVERROR_EXTERNAL;
896 if (drm_dev_info->bustype == DRM_BUS_PCI)
897 dev_select.pci_device = drm_dev_info->deviceinfo.pci->device_id;
899 drmFreeDevice(&drm_dev_info);
901 return vulkan_device_create_internal(ctx, &dev_select, NULL, flags);
905 case AV_HWDEVICE_TYPE_CUDA: {
906 AVHWDeviceContext *cuda_cu = src_ctx;
907 AVCUDADeviceContext *src_hwctx = src_ctx->hwctx;
908 AVCUDADeviceContextInternal *cu_internal = src_hwctx->internal;
909 CudaFunctions *cu = cu_internal->cuda_dl;
911 int ret = CHECK_CU(cu->cuDeviceGetUuid((CUuuid *)&dev_select.uuid,
912 cu_internal->cuda_device));
914 av_log(ctx, AV_LOG_ERROR, "Unable to get UUID from CUDA!\n");
915 return AVERROR_EXTERNAL;
918 dev_select.has_uuid = 1;
920 return vulkan_device_create_internal(ctx, &dev_select, NULL, flags);
924 return AVERROR(ENOSYS);
928 static int vulkan_frames_get_constraints(AVHWDeviceContext *ctx,
929 const void *hwconfig,
930 AVHWFramesConstraints *constraints)
933 AVVulkanDeviceContext *hwctx = ctx->hwctx;
934 VulkanDevicePriv *p = ctx->internal->priv;
936 for (enum AVPixelFormat i = 0; i < AV_PIX_FMT_NB; i++)
937 count += pixfmt_is_supported(hwctx, i, p->use_linear_images);
940 if (p->dev_is_nvidia)
944 constraints->valid_sw_formats = av_malloc_array(count + 1,
945 sizeof(enum AVPixelFormat));
946 if (!constraints->valid_sw_formats)
947 return AVERROR(ENOMEM);
950 for (enum AVPixelFormat i = 0; i < AV_PIX_FMT_NB; i++)
951 if (pixfmt_is_supported(hwctx, i, p->use_linear_images))
952 constraints->valid_sw_formats[count++] = i;
955 if (p->dev_is_nvidia)
956 constraints->valid_sw_formats[count++] = AV_PIX_FMT_CUDA;
958 constraints->valid_sw_formats[count++] = AV_PIX_FMT_NONE;
960 constraints->min_width = 0;
961 constraints->min_height = 0;
962 constraints->max_width = p->props.limits.maxImageDimension2D;
963 constraints->max_height = p->props.limits.maxImageDimension2D;
965 constraints->valid_hw_formats = av_malloc_array(2, sizeof(enum AVPixelFormat));
966 if (!constraints->valid_hw_formats)
967 return AVERROR(ENOMEM);
969 constraints->valid_hw_formats[0] = AV_PIX_FMT_VULKAN;
970 constraints->valid_hw_formats[1] = AV_PIX_FMT_NONE;
975 static int alloc_mem(AVHWDeviceContext *ctx, VkMemoryRequirements *req,
976 VkMemoryPropertyFlagBits req_flags, void *alloc_extension,
977 VkMemoryPropertyFlagBits *mem_flags, VkDeviceMemory *mem)
981 VulkanDevicePriv *p = ctx->internal->priv;
982 AVVulkanDeviceContext *dev_hwctx = ctx->hwctx;
983 VkMemoryAllocateInfo alloc_info = {
984 .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
985 .pNext = alloc_extension,
988 /* Align if we need to */
989 if (req_flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT)
990 req->size = FFALIGN(req->size, p->props.limits.minMemoryMapAlignment);
992 alloc_info.allocationSize = req->size;
994 /* The vulkan spec requires memory types to be sorted in the "optimal"
995 * order, so the first matching type we find will be the best/fastest one */
996 for (int i = 0; i < p->mprops.memoryTypeCount; i++) {
997 /* The memory type must be supported by the requirements (bitfield) */
998 if (!(req->memoryTypeBits & (1 << i)))
1001 /* The memory type flags must include our properties */
1002 if ((p->mprops.memoryTypes[i].propertyFlags & req_flags) != req_flags)
1005 /* Found a suitable memory type */
1011 av_log(ctx, AV_LOG_ERROR, "No memory type found for flags 0x%x\n",
1013 return AVERROR(EINVAL);
1016 alloc_info.memoryTypeIndex = index;
1018 ret = vkAllocateMemory(dev_hwctx->act_dev, &alloc_info,
1019 dev_hwctx->alloc, mem);
1020 if (ret != VK_SUCCESS) {
1021 av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory: %s\n",
1023 return AVERROR(ENOMEM);
1026 *mem_flags |= p->mprops.memoryTypes[index].propertyFlags;
1031 static void vulkan_free_internal(AVVkFrameInternal *internal)
1037 if (internal->cuda_fc_ref) {
1038 AVHWFramesContext *cuda_fc = (AVHWFramesContext *)internal->cuda_fc_ref->data;
1039 int planes = av_pix_fmt_count_planes(cuda_fc->sw_format);
1040 AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
1041 AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
1042 AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
1043 CudaFunctions *cu = cu_internal->cuda_dl;
1045 for (int i = 0; i < planes; i++) {
1046 if (internal->cu_sem[i])
1047 CHECK_CU(cu->cuDestroyExternalSemaphore(internal->cu_sem[i]));
1048 if (internal->cu_mma[i])
1049 CHECK_CU(cu->cuMipmappedArrayDestroy(internal->cu_mma[i]));
1050 if (internal->ext_mem[i])
1051 CHECK_CU(cu->cuDestroyExternalMemory(internal->ext_mem[i]));
1054 av_buffer_unref(&internal->cuda_fc_ref);
1061 static void vulkan_frame_free(void *opaque, uint8_t *data)
1063 AVVkFrame *f = (AVVkFrame *)data;
1064 AVHWFramesContext *hwfc = opaque;
1065 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1066 int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1068 vulkan_free_internal(f->internal);
1070 for (int i = 0; i < planes; i++) {
1071 vkDestroyImage(hwctx->act_dev, f->img[i], hwctx->alloc);
1072 vkFreeMemory(hwctx->act_dev, f->mem[i], hwctx->alloc);
1073 vkDestroySemaphore(hwctx->act_dev, f->sem[i], hwctx->alloc);
1079 static int alloc_bind_mem(AVHWFramesContext *hwfc, AVVkFrame *f,
1080 void *alloc_pnext, size_t alloc_pnext_stride)
1084 AVHWDeviceContext *ctx = hwfc->device_ctx;
1085 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1086 VkBindImageMemoryInfo bind_info[AV_NUM_DATA_POINTERS] = { { 0 } };
1088 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1090 for (int i = 0; i < planes; i++) {
1092 VkImageMemoryRequirementsInfo2 req_desc = {
1093 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_REQUIREMENTS_INFO_2,
1096 VkMemoryDedicatedAllocateInfo ded_alloc = {
1097 .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO,
1098 .pNext = (void *)(((uint8_t *)alloc_pnext) + i*alloc_pnext_stride),
1100 VkMemoryDedicatedRequirements ded_req = {
1101 .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS,
1103 VkMemoryRequirements2 req = {
1104 .sType = VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2,
1108 vkGetImageMemoryRequirements2(hwctx->act_dev, &req_desc, &req);
1110 /* In case the implementation prefers/requires dedicated allocation */
1111 use_ded_mem = ded_req.prefersDedicatedAllocation |
1112 ded_req.requiresDedicatedAllocation;
1114 ded_alloc.image = f->img[i];
1116 /* Allocate memory */
1117 if ((err = alloc_mem(ctx, &req.memoryRequirements,
1118 f->tiling == VK_IMAGE_TILING_LINEAR ?
1119 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT :
1120 VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
1121 use_ded_mem ? &ded_alloc : (void *)ded_alloc.pNext,
1122 &f->flags, &f->mem[i])))
1125 f->size[i] = req.memoryRequirements.size;
1126 bind_info[i].sType = VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO;
1127 bind_info[i].image = f->img[i];
1128 bind_info[i].memory = f->mem[i];
1131 /* Bind the allocated memory to the images */
1132 ret = vkBindImageMemory2(hwctx->act_dev, planes, bind_info);
1133 if (ret != VK_SUCCESS) {
1134 av_log(ctx, AV_LOG_ERROR, "Failed to bind memory: %s\n",
1136 return AVERROR_EXTERNAL;
1144 PREP_MODE_RO_SHADER,
1147 static int prepare_frame(AVHWFramesContext *hwfc, VulkanExecCtx *ectx,
1148 AVVkFrame *frame, enum PrepMode pmode)
1151 VkImageLayout new_layout;
1152 VkAccessFlags new_access;
1153 AVHWDeviceContext *ctx = hwfc->device_ctx;
1154 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1155 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1157 VkImageMemoryBarrier img_bar[AV_NUM_DATA_POINTERS] = { 0 };
1159 VkCommandBufferBeginInfo cmd_start = {
1160 .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
1161 .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
1164 VkSubmitInfo s_info = {
1165 .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
1166 .commandBufferCount = 1,
1167 .pCommandBuffers = &ectx->buf,
1169 .pSignalSemaphores = frame->sem,
1170 .signalSemaphoreCount = planes,
1174 case PREP_MODE_WRITE:
1175 new_layout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
1176 new_access = VK_ACCESS_TRANSFER_WRITE_BIT;
1178 case PREP_MODE_RO_SHADER:
1179 new_layout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL;
1180 new_access = VK_ACCESS_TRANSFER_READ_BIT;
1184 ret = vkBeginCommandBuffer(ectx->buf, &cmd_start);
1185 if (ret != VK_SUCCESS)
1186 return AVERROR_EXTERNAL;
1188 /* Change the image layout to something more optimal for writes.
1189 * This also signals the newly created semaphore, making it usable
1190 * for synchronization */
1191 for (int i = 0; i < planes; i++) {
1192 img_bar[i].sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
1193 img_bar[i].srcAccessMask = 0x0;
1194 img_bar[i].dstAccessMask = new_access;
1195 img_bar[i].oldLayout = frame->layout[i];
1196 img_bar[i].newLayout = new_layout;
1197 img_bar[i].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
1198 img_bar[i].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
1199 img_bar[i].image = frame->img[i];
1200 img_bar[i].subresourceRange.levelCount = 1;
1201 img_bar[i].subresourceRange.layerCount = 1;
1202 img_bar[i].subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
1204 frame->layout[i] = img_bar[i].newLayout;
1205 frame->access[i] = img_bar[i].dstAccessMask;
1208 vkCmdPipelineBarrier(ectx->buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
1209 VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
1210 0, NULL, 0, NULL, planes, img_bar);
1212 ret = vkEndCommandBuffer(ectx->buf);
1213 if (ret != VK_SUCCESS)
1214 return AVERROR_EXTERNAL;
1216 ret = vkQueueSubmit(ectx->queue, 1, &s_info, ectx->fence);
1217 if (ret != VK_SUCCESS) {
1218 return AVERROR_EXTERNAL;
1220 vkWaitForFences(hwctx->act_dev, 1, &ectx->fence, VK_TRUE, UINT64_MAX);
1221 vkResetFences(hwctx->act_dev, 1, &ectx->fence);
1227 static int create_frame(AVHWFramesContext *hwfc, AVVkFrame **frame,
1228 VkImageTiling tiling, VkImageUsageFlagBits usage,
1233 AVHWDeviceContext *ctx = hwfc->device_ctx;
1234 VulkanDevicePriv *p = ctx->internal->priv;
1235 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1236 enum AVPixelFormat format = hwfc->sw_format;
1237 const VkFormat *img_fmts = av_vkfmt_from_pixfmt(format);
1238 const int planes = av_pix_fmt_count_planes(format);
1240 VkExportSemaphoreCreateInfo ext_sem_info = {
1241 .sType = VK_STRUCTURE_TYPE_EXPORT_SEMAPHORE_CREATE_INFO,
1242 .handleTypes = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT,
1245 VkSemaphoreCreateInfo sem_spawn = {
1246 .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
1247 .pNext = p->extensions & EXT_EXTERNAL_FD_SEM ? &ext_sem_info : NULL,
1250 AVVkFrame *f = av_vk_frame_alloc();
1252 av_log(ctx, AV_LOG_ERROR, "Unable to allocate memory for AVVkFrame!\n");
1253 return AVERROR(ENOMEM);
1256 /* Create the images */
1257 for (int i = 0; i < planes; i++) {
1258 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(format);
1259 int w = hwfc->width;
1260 int h = hwfc->height;
1261 const int p_w = i > 0 ? AV_CEIL_RSHIFT(w, desc->log2_chroma_w) : w;
1262 const int p_h = i > 0 ? AV_CEIL_RSHIFT(h, desc->log2_chroma_h) : h;
1264 VkImageCreateInfo image_create_info = {
1265 .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
1266 .pNext = create_pnext,
1267 .imageType = VK_IMAGE_TYPE_2D,
1268 .format = img_fmts[i],
1269 .extent.width = p_w,
1270 .extent.height = p_h,
1274 .flags = VK_IMAGE_CREATE_ALIAS_BIT,
1276 .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED,
1278 .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
1279 .samples = VK_SAMPLE_COUNT_1_BIT,
1282 ret = vkCreateImage(hwctx->act_dev, &image_create_info,
1283 hwctx->alloc, &f->img[i]);
1284 if (ret != VK_SUCCESS) {
1285 av_log(ctx, AV_LOG_ERROR, "Image creation failure: %s\n",
1287 err = AVERROR(EINVAL);
1291 /* Create semaphore */
1292 ret = vkCreateSemaphore(hwctx->act_dev, &sem_spawn,
1293 hwctx->alloc, &f->sem[i]);
1294 if (ret != VK_SUCCESS) {
1295 av_log(hwctx, AV_LOG_ERROR, "Failed to create semaphore: %s\n",
1297 return AVERROR_EXTERNAL;
1300 f->layout[i] = image_create_info.initialLayout;
1311 vulkan_frame_free(hwfc, (uint8_t *)f);
1315 /* Checks if an export flag is enabled, and if it is ORs it with *iexp */
1316 static void try_export_flags(AVHWFramesContext *hwfc,
1317 VkExternalMemoryHandleTypeFlags *comp_handle_types,
1318 VkExternalMemoryHandleTypeFlagBits *iexp,
1319 VkExternalMemoryHandleTypeFlagBits exp)
1322 AVVulkanFramesContext *hwctx = hwfc->hwctx;
1323 AVVulkanDeviceContext *dev_hwctx = hwfc->device_ctx->hwctx;
1324 VkExternalImageFormatProperties eprops = {
1325 .sType = VK_STRUCTURE_TYPE_EXTERNAL_IMAGE_FORMAT_PROPERTIES_KHR,
1327 VkImageFormatProperties2 props = {
1328 .sType = VK_STRUCTURE_TYPE_IMAGE_FORMAT_PROPERTIES_2,
1331 VkPhysicalDeviceExternalImageFormatInfo enext = {
1332 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_IMAGE_FORMAT_INFO,
1335 VkPhysicalDeviceImageFormatInfo2 pinfo = {
1336 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_FORMAT_INFO_2,
1337 .pNext = !exp ? NULL : &enext,
1338 .format = av_vkfmt_from_pixfmt(hwfc->sw_format)[0],
1339 .type = VK_IMAGE_TYPE_2D,
1340 .tiling = hwctx->tiling,
1341 .usage = hwctx->usage,
1342 .flags = VK_IMAGE_CREATE_ALIAS_BIT,
1345 ret = vkGetPhysicalDeviceImageFormatProperties2(dev_hwctx->phys_dev,
1347 if (ret == VK_SUCCESS) {
1349 *comp_handle_types |= eprops.externalMemoryProperties.compatibleHandleTypes;
1353 static AVBufferRef *vulkan_pool_alloc(void *opaque, int size)
1357 AVBufferRef *avbuf = NULL;
1358 AVHWFramesContext *hwfc = opaque;
1359 AVVulkanFramesContext *hwctx = hwfc->hwctx;
1360 VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
1361 VkExportMemoryAllocateInfo eminfo[AV_NUM_DATA_POINTERS];
1362 VkExternalMemoryHandleTypeFlags e = 0x0;
1364 VkExternalMemoryImageCreateInfo eiinfo = {
1365 .sType = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO,
1366 .pNext = hwctx->create_pnext,
1369 if (p->extensions & EXT_EXTERNAL_FD_MEMORY)
1370 try_export_flags(hwfc, &eiinfo.handleTypes, &e,
1371 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT);
1373 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
1374 try_export_flags(hwfc, &eiinfo.handleTypes, &e,
1375 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
1377 for (int i = 0; i < av_pix_fmt_count_planes(hwfc->sw_format); i++) {
1378 eminfo[i].sType = VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO;
1379 eminfo[i].pNext = hwctx->alloc_pnext[i];
1380 eminfo[i].handleTypes = e;
1383 err = create_frame(hwfc, &f, hwctx->tiling, hwctx->usage,
1384 eiinfo.handleTypes ? &eiinfo : NULL);
1388 err = alloc_bind_mem(hwfc, f, eminfo, sizeof(*eminfo));
1392 err = prepare_frame(hwfc, &p->cmd, f, PREP_MODE_WRITE);
1396 avbuf = av_buffer_create((uint8_t *)f, sizeof(AVVkFrame),
1397 vulkan_frame_free, hwfc, 0);
1404 vulkan_frame_free(hwfc, (uint8_t *)f);
1408 static void vulkan_frames_uninit(AVHWFramesContext *hwfc)
1410 VulkanFramesPriv *fp = hwfc->internal->priv;
1412 free_exec_ctx(hwfc->device_ctx, &fp->cmd);
1415 static int vulkan_frames_init(AVHWFramesContext *hwfc)
1419 AVVulkanFramesContext *hwctx = hwfc->hwctx;
1420 VulkanFramesPriv *fp = hwfc->internal->priv;
1421 AVVulkanDeviceContext *dev_hwctx = hwfc->device_ctx->hwctx;
1422 VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
1427 /* Default pool flags */
1428 hwctx->tiling = hwctx->tiling ? hwctx->tiling : p->use_linear_images ?
1429 VK_IMAGE_TILING_LINEAR : VK_IMAGE_TILING_OPTIMAL;
1431 hwctx->usage |= DEFAULT_USAGE_FLAGS;
1433 err = create_exec_ctx(hwfc->device_ctx, &fp->cmd,
1434 dev_hwctx->queue_family_tx_index);
1438 /* Test to see if allocation will fail */
1439 err = create_frame(hwfc, &f, hwctx->tiling, hwctx->usage,
1440 hwctx->create_pnext);
1442 free_exec_ctx(hwfc->device_ctx, &p->cmd);
1446 vulkan_frame_free(hwfc, (uint8_t *)f);
1448 hwfc->internal->pool_internal = av_buffer_pool_init2(sizeof(AVVkFrame),
1449 hwfc, vulkan_pool_alloc,
1451 if (!hwfc->internal->pool_internal) {
1452 free_exec_ctx(hwfc->device_ctx, &p->cmd);
1453 return AVERROR(ENOMEM);
1459 static int vulkan_get_buffer(AVHWFramesContext *hwfc, AVFrame *frame)
1461 frame->buf[0] = av_buffer_pool_get(hwfc->pool);
1463 return AVERROR(ENOMEM);
1465 frame->data[0] = frame->buf[0]->data;
1466 frame->format = AV_PIX_FMT_VULKAN;
1467 frame->width = hwfc->width;
1468 frame->height = hwfc->height;
1473 static int vulkan_transfer_get_formats(AVHWFramesContext *hwfc,
1474 enum AVHWFrameTransferDirection dir,
1475 enum AVPixelFormat **formats)
1477 enum AVPixelFormat *fmts = av_malloc_array(2, sizeof(*fmts));
1479 return AVERROR(ENOMEM);
1481 fmts[0] = hwfc->sw_format;
1482 fmts[1] = AV_PIX_FMT_NONE;
1488 typedef struct VulkanMapping {
1493 static void vulkan_unmap_frame(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
1495 VulkanMapping *map = hwmap->priv;
1496 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1497 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1499 /* Check if buffer needs flushing */
1500 if ((map->flags & AV_HWFRAME_MAP_WRITE) &&
1501 !(map->frame->flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)) {
1503 VkMappedMemoryRange flush_ranges[AV_NUM_DATA_POINTERS] = { { 0 } };
1505 for (int i = 0; i < planes; i++) {
1506 flush_ranges[i].sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
1507 flush_ranges[i].memory = map->frame->mem[i];
1508 flush_ranges[i].size = VK_WHOLE_SIZE;
1511 ret = vkFlushMappedMemoryRanges(hwctx->act_dev, planes,
1513 if (ret != VK_SUCCESS) {
1514 av_log(hwfc, AV_LOG_ERROR, "Failed to flush memory: %s\n",
1519 for (int i = 0; i < planes; i++)
1520 vkUnmapMemory(hwctx->act_dev, map->frame->mem[i]);
1525 static int vulkan_map_frame_to_mem(AVHWFramesContext *hwfc, AVFrame *dst,
1526 const AVFrame *src, int flags)
1529 int err, mapped_mem_count = 0;
1530 AVVkFrame *f = (AVVkFrame *)src->data[0];
1531 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1532 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1534 VulkanMapping *map = av_mallocz(sizeof(VulkanMapping));
1536 return AVERROR(EINVAL);
1538 if (src->format != AV_PIX_FMT_VULKAN) {
1539 av_log(hwfc, AV_LOG_ERROR, "Cannot map from pixel format %s!\n",
1540 av_get_pix_fmt_name(src->format));
1541 err = AVERROR(EINVAL);
1545 if (!(f->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) ||
1546 !(f->tiling == VK_IMAGE_TILING_LINEAR)) {
1547 av_log(hwfc, AV_LOG_ERROR, "Unable to map frame, not host visible "
1549 err = AVERROR(EINVAL);
1553 dst->width = src->width;
1554 dst->height = src->height;
1556 for (int i = 0; i < planes; i++) {
1557 ret = vkMapMemory(hwctx->act_dev, f->mem[i], 0,
1558 VK_WHOLE_SIZE, 0, (void **)&dst->data[i]);
1559 if (ret != VK_SUCCESS) {
1560 av_log(hwfc, AV_LOG_ERROR, "Failed to map image memory: %s\n",
1562 err = AVERROR_EXTERNAL;
1568 /* Check if the memory contents matter */
1569 if (((flags & AV_HWFRAME_MAP_READ) || !(flags & AV_HWFRAME_MAP_OVERWRITE)) &&
1570 !(f->flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)) {
1571 VkMappedMemoryRange map_mem_ranges[AV_NUM_DATA_POINTERS] = { { 0 } };
1572 for (int i = 0; i < planes; i++) {
1573 map_mem_ranges[i].sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
1574 map_mem_ranges[i].size = VK_WHOLE_SIZE;
1575 map_mem_ranges[i].memory = f->mem[i];
1578 ret = vkInvalidateMappedMemoryRanges(hwctx->act_dev, planes,
1580 if (ret != VK_SUCCESS) {
1581 av_log(hwfc, AV_LOG_ERROR, "Failed to invalidate memory: %s\n",
1583 err = AVERROR_EXTERNAL;
1588 for (int i = 0; i < planes; i++) {
1589 VkImageSubresource sub = {
1590 .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
1592 VkSubresourceLayout layout;
1593 vkGetImageSubresourceLayout(hwctx->act_dev, f->img[i], &sub, &layout);
1594 dst->linesize[i] = layout.rowPitch;
1600 err = ff_hwframe_map_create(src->hw_frames_ctx, dst, src,
1601 &vulkan_unmap_frame, map);
1608 for (int i = 0; i < mapped_mem_count; i++)
1609 vkUnmapMemory(hwctx->act_dev, f->mem[i]);
1616 static void vulkan_unmap_from(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
1618 VulkanMapping *map = hwmap->priv;
1619 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1620 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1622 for (int i = 0; i < planes; i++) {
1623 vkDestroyImage(hwctx->act_dev, map->frame->img[i], hwctx->alloc);
1624 vkFreeMemory(hwctx->act_dev, map->frame->mem[i], hwctx->alloc);
1625 vkDestroySemaphore(hwctx->act_dev, map->frame->sem[i], hwctx->alloc);
1628 av_freep(&map->frame);
1631 static const struct {
1632 uint32_t drm_fourcc;
1634 } vulkan_drm_format_map[] = {
1635 { DRM_FORMAT_R8, VK_FORMAT_R8_UNORM },
1636 { DRM_FORMAT_R16, VK_FORMAT_R16_UNORM },
1637 { DRM_FORMAT_GR88, VK_FORMAT_R8G8_UNORM },
1638 { DRM_FORMAT_RG88, VK_FORMAT_R8G8_UNORM },
1639 { DRM_FORMAT_GR1616, VK_FORMAT_R16G16_UNORM },
1640 { DRM_FORMAT_RG1616, VK_FORMAT_R16G16_UNORM },
1641 { DRM_FORMAT_ARGB8888, VK_FORMAT_B8G8R8A8_UNORM },
1642 { DRM_FORMAT_XRGB8888, VK_FORMAT_B8G8R8A8_UNORM },
1643 { DRM_FORMAT_ABGR8888, VK_FORMAT_R8G8B8A8_UNORM },
1644 { DRM_FORMAT_XBGR8888, VK_FORMAT_R8G8B8A8_UNORM },
1647 static inline VkFormat drm_to_vulkan_fmt(uint32_t drm_fourcc)
1649 for (int i = 0; i < FF_ARRAY_ELEMS(vulkan_drm_format_map); i++)
1650 if (vulkan_drm_format_map[i].drm_fourcc == drm_fourcc)
1651 return vulkan_drm_format_map[i].vk_format;
1652 return VK_FORMAT_UNDEFINED;
1655 static int vulkan_map_from_drm_frame_desc(AVHWFramesContext *hwfc, AVVkFrame **frame,
1656 AVDRMFrameDescriptor *desc)
1661 AVHWDeviceContext *ctx = hwfc->device_ctx;
1662 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1663 VulkanDevicePriv *p = ctx->internal->priv;
1664 const AVPixFmtDescriptor *fmt_desc = av_pix_fmt_desc_get(hwfc->sw_format);
1665 const int has_modifiers = p->extensions & EXT_DRM_MODIFIER_FLAGS;
1666 VkSubresourceLayout plane_data[AV_NUM_DATA_POINTERS];
1667 VkBindImageMemoryInfo bind_info[AV_NUM_DATA_POINTERS];
1668 VkExternalMemoryHandleTypeFlagBits htype = VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT;
1670 VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdPropertiesKHR);
1672 for (int i = 0; i < desc->nb_layers; i++) {
1673 if (desc->layers[i].nb_planes > 1) {
1674 av_log(ctx, AV_LOG_ERROR, "Cannot import DMABUFS with more than 1 "
1675 "plane per layer!\n");
1676 return AVERROR(EINVAL);
1679 if (drm_to_vulkan_fmt(desc->layers[i].format) == VK_FORMAT_UNDEFINED) {
1680 av_log(ctx, AV_LOG_ERROR, "Unsupported DMABUF layer format %#08x!\n",
1681 desc->layers[i].format);
1682 return AVERROR(EINVAL);
1686 if (!(f = av_vk_frame_alloc())) {
1687 av_log(ctx, AV_LOG_ERROR, "Unable to allocate memory for AVVkFrame!\n");
1688 err = AVERROR(ENOMEM);
1692 for (int i = 0; i < desc->nb_objects; i++) {
1693 VkMemoryFdPropertiesKHR fdmp = {
1694 .sType = VK_STRUCTURE_TYPE_MEMORY_FD_PROPERTIES_KHR,
1696 VkMemoryRequirements req = {
1697 .size = desc->objects[i].size,
1699 VkImportMemoryFdInfoKHR idesc = {
1700 .sType = VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR,
1701 .handleType = htype,
1702 .fd = dup(desc->objects[i].fd),
1705 ret = pfn_vkGetMemoryFdPropertiesKHR(hwctx->act_dev, htype,
1707 if (ret != VK_SUCCESS) {
1708 av_log(hwfc, AV_LOG_ERROR, "Failed to get FD properties: %s\n",
1710 err = AVERROR_EXTERNAL;
1715 req.memoryTypeBits = fdmp.memoryTypeBits;
1717 err = alloc_mem(ctx, &req, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
1718 &idesc, &f->flags, &f->mem[i]);
1724 f->size[i] = desc->objects[i].size;
1727 f->tiling = has_modifiers ? VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT :
1728 desc->objects[0].format_modifier == DRM_FORMAT_MOD_LINEAR ?
1729 VK_IMAGE_TILING_LINEAR : VK_IMAGE_TILING_OPTIMAL;
1731 for (int i = 0; i < desc->nb_layers; i++) {
1732 VkImageDrmFormatModifierExplicitCreateInfoEXT drm_info = {
1733 .sType = VK_STRUCTURE_TYPE_IMAGE_DRM_FORMAT_MODIFIER_EXPLICIT_CREATE_INFO_EXT,
1734 .drmFormatModifier = desc->objects[0].format_modifier,
1735 .drmFormatModifierPlaneCount = desc->layers[i].nb_planes,
1736 .pPlaneLayouts = (const VkSubresourceLayout *)&plane_data,
1739 VkExternalMemoryImageCreateInfo einfo = {
1740 .sType = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO,
1741 .pNext = has_modifiers ? &drm_info : NULL,
1742 .handleTypes = htype,
1745 VkSemaphoreCreateInfo sem_spawn = {
1746 .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
1749 const int p_w = i > 0 ? AV_CEIL_RSHIFT(hwfc->width, fmt_desc->log2_chroma_w) : hwfc->width;
1750 const int p_h = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, fmt_desc->log2_chroma_h) : hwfc->height;
1752 VkImageCreateInfo image_create_info = {
1753 .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
1755 .imageType = VK_IMAGE_TYPE_2D,
1756 .format = drm_to_vulkan_fmt(desc->layers[i].format),
1757 .extent.width = p_w,
1758 .extent.height = p_h,
1762 .flags = VK_IMAGE_CREATE_ALIAS_BIT,
1763 .tiling = f->tiling,
1764 .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED, /* specs say so */
1765 .usage = DEFAULT_USAGE_FLAGS,
1766 .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
1767 .samples = VK_SAMPLE_COUNT_1_BIT,
1770 for (int j = 0; j < desc->layers[i].nb_planes; j++) {
1771 plane_data[j].offset = desc->layers[i].planes[j].offset;
1772 plane_data[j].rowPitch = desc->layers[i].planes[j].pitch;
1773 plane_data[j].size = 0; /* The specs say so for all 3 */
1774 plane_data[j].arrayPitch = 0;
1775 plane_data[j].depthPitch = 0;
1779 ret = vkCreateImage(hwctx->act_dev, &image_create_info,
1780 hwctx->alloc, &f->img[i]);
1781 if (ret != VK_SUCCESS) {
1782 av_log(ctx, AV_LOG_ERROR, "Image creation failure: %s\n",
1784 err = AVERROR(EINVAL);
1788 ret = vkCreateSemaphore(hwctx->act_dev, &sem_spawn,
1789 hwctx->alloc, &f->sem[i]);
1790 if (ret != VK_SUCCESS) {
1791 av_log(hwctx, AV_LOG_ERROR, "Failed to create semaphore: %s\n",
1793 return AVERROR_EXTERNAL;
1796 /* We'd import a semaphore onto the one we created using
1797 * vkImportSemaphoreFdKHR but unfortunately neither DRM nor VAAPI
1798 * offer us anything we could import and sync with, so instead
1799 * just signal the semaphore we created. */
1801 f->layout[i] = image_create_info.initialLayout;
1804 /* TODO: Fix to support more than 1 plane per layer */
1805 bind_info[i].sType = VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO;
1806 bind_info[i].pNext = NULL;
1807 bind_info[i].image = f->img[i];
1808 bind_info[i].memory = f->mem[desc->layers[i].planes[0].object_index];
1809 bind_info[i].memoryOffset = desc->layers[i].planes[0].offset;
1812 /* Bind the allocated memory to the images */
1813 ret = vkBindImageMemory2(hwctx->act_dev, desc->nb_layers, bind_info);
1814 if (ret != VK_SUCCESS) {
1815 av_log(ctx, AV_LOG_ERROR, "Failed to bind memory: %s\n",
1817 return AVERROR_EXTERNAL;
1820 /* NOTE: This is completely uneccesary and unneeded once we can import
1821 * semaphores from DRM. Otherwise we have to activate the semaphores.
1822 * We're reusing the exec context that's also used for uploads/downloads. */
1823 err = prepare_frame(hwfc, &p->cmd, f, PREP_MODE_RO_SHADER);
1832 for (int i = 0; i < desc->nb_layers; i++) {
1833 vkDestroyImage(hwctx->act_dev, f->img[i], hwctx->alloc);
1834 vkDestroySemaphore(hwctx->act_dev, f->sem[i], hwctx->alloc);
1836 for (int i = 0; i < desc->nb_objects; i++)
1837 vkFreeMemory(hwctx->act_dev, f->mem[i], hwctx->alloc);
1844 static int vulkan_map_from_drm(AVHWFramesContext *hwfc, AVFrame *dst,
1845 const AVFrame *src, int flags)
1849 VulkanMapping *map = NULL;
1851 err = vulkan_map_from_drm_frame_desc(hwfc, &f,
1852 (AVDRMFrameDescriptor *)src->data[0]);
1856 /* The unmapping function will free this */
1857 dst->data[0] = (uint8_t *)f;
1858 dst->width = src->width;
1859 dst->height = src->height;
1861 map = av_mallocz(sizeof(VulkanMapping));
1868 err = ff_hwframe_map_create(dst->hw_frames_ctx, dst, src,
1869 &vulkan_unmap_from, map);
1873 av_log(hwfc, AV_LOG_DEBUG, "Mapped DRM object to Vulkan!\n");
1878 vulkan_frame_free(hwfc->device_ctx->hwctx, (uint8_t *)f);
1884 static int vulkan_map_from_vaapi(AVHWFramesContext *dst_fc,
1885 AVFrame *dst, const AVFrame *src,
1889 AVFrame *tmp = av_frame_alloc();
1890 AVHWFramesContext *vaapi_fc = (AVHWFramesContext*)src->hw_frames_ctx->data;
1891 AVVAAPIDeviceContext *vaapi_ctx = vaapi_fc->device_ctx->hwctx;
1892 VASurfaceID surface_id = (VASurfaceID)(uintptr_t)src->data[3];
1895 return AVERROR(ENOMEM);
1897 /* We have to sync since like the previous comment said, no semaphores */
1898 vaSyncSurface(vaapi_ctx->display, surface_id);
1900 tmp->format = AV_PIX_FMT_DRM_PRIME;
1902 err = av_hwframe_map(tmp, src, flags);
1906 err = vulkan_map_from_drm(dst_fc, dst, tmp, flags);
1910 err = ff_hwframe_map_replace(dst, src);
1913 av_frame_free(&tmp);
1920 static int vulkan_export_to_cuda(AVHWFramesContext *hwfc,
1921 AVBufferRef *cuda_hwfc,
1922 const AVFrame *frame)
1927 AVVkFrameInternal *dst_int;
1928 AVHWDeviceContext *ctx = hwfc->device_ctx;
1929 AVVulkanDeviceContext *hwctx = ctx->hwctx;
1930 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1931 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
1932 VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdKHR);
1933 VK_LOAD_PFN(hwctx->inst, vkGetSemaphoreFdKHR);
1935 AVHWFramesContext *cuda_fc = (AVHWFramesContext*)cuda_hwfc->data;
1936 AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
1937 AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
1938 AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
1939 CudaFunctions *cu = cu_internal->cuda_dl;
1940 CUarray_format cufmt = desc->comp[0].depth > 8 ? CU_AD_FORMAT_UNSIGNED_INT16 :
1941 CU_AD_FORMAT_UNSIGNED_INT8;
1943 dst_f = (AVVkFrame *)frame->data[0];
1945 dst_int = dst_f->internal;
1946 if (!dst_int || !dst_int->cuda_fc_ref) {
1947 if (!dst_f->internal)
1948 dst_f->internal = dst_int = av_mallocz(sizeof(*dst_f->internal));
1951 err = AVERROR(ENOMEM);
1955 dst_int->cuda_fc_ref = av_buffer_ref(cuda_hwfc);
1956 if (!dst_int->cuda_fc_ref) {
1957 err = AVERROR(ENOMEM);
1961 for (int i = 0; i < planes; i++) {
1962 CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC tex_desc = {
1965 .Width = i > 0 ? AV_CEIL_RSHIFT(hwfc->width, desc->log2_chroma_w)
1967 .Height = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, desc->log2_chroma_h)
1971 .NumChannels = 1 + ((planes == 2) && i),
1976 CUDA_EXTERNAL_MEMORY_HANDLE_DESC ext_desc = {
1977 .type = CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD,
1978 .size = dst_f->size[i],
1980 VkMemoryGetFdInfoKHR export_info = {
1981 .sType = VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR,
1982 .memory = dst_f->mem[i],
1983 .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR,
1985 VkSemaphoreGetFdInfoKHR sem_export = {
1986 .sType = VK_STRUCTURE_TYPE_SEMAPHORE_GET_FD_INFO_KHR,
1987 .semaphore = dst_f->sem[i],
1988 .handleType = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT,
1990 CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC ext_sem_desc = {
1991 .type = CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD,
1994 ret = pfn_vkGetMemoryFdKHR(hwctx->act_dev, &export_info,
1995 &ext_desc.handle.fd);
1996 if (ret != VK_SUCCESS) {
1997 av_log(hwfc, AV_LOG_ERROR, "Unable to export the image as a FD!\n");
1998 err = AVERROR_EXTERNAL;
2002 ret = CHECK_CU(cu->cuImportExternalMemory(&dst_int->ext_mem[i], &ext_desc));
2004 err = AVERROR_EXTERNAL;
2008 ret = CHECK_CU(cu->cuExternalMemoryGetMappedMipmappedArray(&dst_int->cu_mma[i],
2009 dst_int->ext_mem[i],
2012 err = AVERROR_EXTERNAL;
2016 ret = CHECK_CU(cu->cuMipmappedArrayGetLevel(&dst_int->cu_array[i],
2017 dst_int->cu_mma[i], 0));
2019 err = AVERROR_EXTERNAL;
2023 ret = pfn_vkGetSemaphoreFdKHR(hwctx->act_dev, &sem_export,
2024 &ext_sem_desc.handle.fd);
2025 if (ret != VK_SUCCESS) {
2026 av_log(ctx, AV_LOG_ERROR, "Failed to export semaphore: %s\n",
2028 err = AVERROR_EXTERNAL;
2032 ret = CHECK_CU(cu->cuImportExternalSemaphore(&dst_int->cu_sem[i],
2035 err = AVERROR_EXTERNAL;
2047 static int vulkan_transfer_data_from_cuda(AVHWFramesContext *hwfc,
2048 AVFrame *dst, const AVFrame *src)
2054 AVVkFrameInternal *dst_int;
2055 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
2056 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
2058 AVHWFramesContext *cuda_fc = (AVHWFramesContext*)src->hw_frames_ctx->data;
2059 AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
2060 AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
2061 AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
2062 CudaFunctions *cu = cu_internal->cuda_dl;
2063 CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS s_w_par[AV_NUM_DATA_POINTERS] = { 0 };
2064 CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS s_s_par[AV_NUM_DATA_POINTERS] = { 0 };
2066 ret = CHECK_CU(cu->cuCtxPushCurrent(cuda_dev->cuda_ctx));
2068 err = AVERROR_EXTERNAL;
2072 dst_f = (AVVkFrame *)dst->data[0];
2074 ret = vulkan_export_to_cuda(hwfc, src->hw_frames_ctx, dst);
2078 dst_int = dst_f->internal;
2080 ret = CHECK_CU(cu->cuWaitExternalSemaphoresAsync(dst_int->cu_sem, s_w_par,
2081 planes, cuda_dev->stream));
2083 err = AVERROR_EXTERNAL;
2087 for (int i = 0; i < planes; i++) {
2088 CUDA_MEMCPY2D cpy = {
2089 .srcMemoryType = CU_MEMORYTYPE_DEVICE,
2090 .srcDevice = (CUdeviceptr)src->data[i],
2091 .srcPitch = src->linesize[i],
2094 .dstMemoryType = CU_MEMORYTYPE_ARRAY,
2095 .dstArray = dst_int->cu_array[i],
2096 .WidthInBytes = (i > 0 ? AV_CEIL_RSHIFT(hwfc->width, desc->log2_chroma_w)
2097 : hwfc->width) * desc->comp[i].step,
2098 .Height = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, desc->log2_chroma_h)
2102 ret = CHECK_CU(cu->cuMemcpy2DAsync(&cpy, cuda_dev->stream));
2104 err = AVERROR_EXTERNAL;
2109 ret = CHECK_CU(cu->cuSignalExternalSemaphoresAsync(dst_int->cu_sem, s_s_par,
2110 planes, cuda_dev->stream));
2112 err = AVERROR_EXTERNAL;
2116 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2118 av_log(hwfc, AV_LOG_VERBOSE, "Transfered CUDA image to Vulkan!\n");
2123 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2124 vulkan_free_internal(dst_int);
2125 dst_f->internal = NULL;
2126 av_buffer_unref(&dst->buf[0]);
2131 static int vulkan_map_to(AVHWFramesContext *hwfc, AVFrame *dst,
2132 const AVFrame *src, int flags)
2134 av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2136 switch (src->format) {
2139 case AV_PIX_FMT_VAAPI:
2140 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2141 return vulkan_map_from_vaapi(hwfc, dst, src, flags);
2143 case AV_PIX_FMT_DRM_PRIME:
2144 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2145 return vulkan_map_from_drm(hwfc, dst, src, flags);
2148 return AVERROR(ENOSYS);
2153 typedef struct VulkanDRMMapping {
2154 AVDRMFrameDescriptor drm_desc;
2158 static void vulkan_unmap_to_drm(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
2160 AVDRMFrameDescriptor *drm_desc = hwmap->priv;
2162 for (int i = 0; i < drm_desc->nb_objects; i++)
2163 close(drm_desc->objects[i].fd);
2168 static inline uint32_t vulkan_fmt_to_drm(VkFormat vkfmt)
2170 for (int i = 0; i < FF_ARRAY_ELEMS(vulkan_drm_format_map); i++)
2171 if (vulkan_drm_format_map[i].vk_format == vkfmt)
2172 return vulkan_drm_format_map[i].drm_fourcc;
2173 return DRM_FORMAT_INVALID;
2176 static int vulkan_map_to_drm(AVHWFramesContext *hwfc, AVFrame *dst,
2177 const AVFrame *src, int flags)
2181 AVVkFrame *f = (AVVkFrame *)src->data[0];
2182 VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2183 AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
2184 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
2185 VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdKHR);
2186 VkImageDrmFormatModifierPropertiesEXT drm_mod = {
2187 .sType = VK_STRUCTURE_TYPE_IMAGE_DRM_FORMAT_MODIFIER_PROPERTIES_EXT,
2190 AVDRMFrameDescriptor *drm_desc = av_mallocz(sizeof(*drm_desc));
2192 return AVERROR(ENOMEM);
2194 err = ff_hwframe_map_create(src->hw_frames_ctx, dst, src, &vulkan_unmap_to_drm, drm_desc);
2198 if (p->extensions & EXT_DRM_MODIFIER_FLAGS) {
2199 VK_LOAD_PFN(hwctx->inst, vkGetImageDrmFormatModifierPropertiesEXT);
2200 ret = pfn_vkGetImageDrmFormatModifierPropertiesEXT(hwctx->act_dev, f->img[0],
2202 if (ret != VK_SUCCESS) {
2203 av_log(hwfc, AV_LOG_ERROR, "Failed to retrieve DRM format modifier!\n");
2204 err = AVERROR_EXTERNAL;
2209 for (int i = 0; (i < planes) && (f->mem[i]); i++) {
2210 VkMemoryGetFdInfoKHR export_info = {
2211 .sType = VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR,
2212 .memory = f->mem[i],
2213 .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT,
2216 ret = pfn_vkGetMemoryFdKHR(hwctx->act_dev, &export_info,
2217 &drm_desc->objects[i].fd);
2218 if (ret != VK_SUCCESS) {
2219 av_log(hwfc, AV_LOG_ERROR, "Unable to export the image as a FD!\n");
2220 err = AVERROR_EXTERNAL;
2224 drm_desc->nb_objects++;
2225 drm_desc->objects[i].size = f->size[i];
2226 drm_desc->objects[i].format_modifier = drm_mod.drmFormatModifier;
2229 drm_desc->nb_layers = planes;
2230 for (int i = 0; i < drm_desc->nb_layers; i++) {
2231 VkSubresourceLayout layout;
2232 VkImageSubresource sub = {
2233 .aspectMask = p->extensions & EXT_DRM_MODIFIER_FLAGS ?
2234 VK_IMAGE_ASPECT_MEMORY_PLANE_0_BIT_EXT :
2235 VK_IMAGE_ASPECT_COLOR_BIT,
2237 VkFormat plane_vkfmt = av_vkfmt_from_pixfmt(hwfc->sw_format)[i];
2239 drm_desc->layers[i].format = vulkan_fmt_to_drm(plane_vkfmt);
2240 drm_desc->layers[i].nb_planes = 1;
2242 if (drm_desc->layers[i].format == DRM_FORMAT_INVALID) {
2243 av_log(hwfc, AV_LOG_ERROR, "Cannot map to DRM layer, unsupported!\n");
2244 err = AVERROR_PATCHWELCOME;
2248 drm_desc->layers[i].planes[0].object_index = FFMIN(i, drm_desc->nb_objects - 1);
2250 if (f->tiling != VK_IMAGE_TILING_OPTIMAL)
2253 vkGetImageSubresourceLayout(hwctx->act_dev, f->img[i], &sub, &layout);
2254 drm_desc->layers[i].planes[0].offset = layout.offset;
2255 drm_desc->layers[i].planes[0].pitch = layout.rowPitch;
2258 dst->width = src->width;
2259 dst->height = src->height;
2260 dst->data[0] = (uint8_t *)drm_desc;
2262 av_log(hwfc, AV_LOG_VERBOSE, "Mapped AVVkFrame to a DRM object!\n");
2272 static int vulkan_map_to_vaapi(AVHWFramesContext *hwfc, AVFrame *dst,
2273 const AVFrame *src, int flags)
2276 AVFrame *tmp = av_frame_alloc();
2278 return AVERROR(ENOMEM);
2280 tmp->format = AV_PIX_FMT_DRM_PRIME;
2282 err = vulkan_map_to_drm(hwfc, tmp, src, flags);
2286 err = av_hwframe_map(dst, tmp, flags);
2290 err = ff_hwframe_map_replace(dst, src);
2293 av_frame_free(&tmp);
2299 static int vulkan_map_from(AVHWFramesContext *hwfc, AVFrame *dst,
2300 const AVFrame *src, int flags)
2302 av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2304 switch (dst->format) {
2306 case AV_PIX_FMT_DRM_PRIME:
2307 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2308 return vulkan_map_to_drm(hwfc, dst, src, flags);
2310 case AV_PIX_FMT_VAAPI:
2311 if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2312 return vulkan_map_to_vaapi(hwfc, dst, src, flags);
2316 return vulkan_map_frame_to_mem(hwfc, dst, src, flags);
2320 typedef struct ImageBuffer {
2323 VkMemoryPropertyFlagBits flags;
2326 static void free_buf(AVHWDeviceContext *ctx, ImageBuffer *buf)
2328 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2332 vkDestroyBuffer(hwctx->act_dev, buf->buf, hwctx->alloc);
2333 vkFreeMemory(hwctx->act_dev, buf->mem, hwctx->alloc);
2336 static int create_buf(AVHWDeviceContext *ctx, ImageBuffer *buf, int height,
2337 int *stride, VkBufferUsageFlags usage,
2338 VkMemoryPropertyFlagBits flags, void *create_pnext,
2343 VkMemoryRequirements req;
2344 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2345 VulkanDevicePriv *p = ctx->internal->priv;
2347 VkBufferCreateInfo buf_spawn = {
2348 .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
2349 .pNext = create_pnext,
2351 .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
2354 *stride = FFALIGN(*stride, p->props.limits.optimalBufferCopyRowPitchAlignment);
2355 buf_spawn.size = height*(*stride);
2357 ret = vkCreateBuffer(hwctx->act_dev, &buf_spawn, NULL, &buf->buf);
2358 if (ret != VK_SUCCESS) {
2359 av_log(ctx, AV_LOG_ERROR, "Failed to create buffer: %s\n",
2361 return AVERROR_EXTERNAL;
2364 vkGetBufferMemoryRequirements(hwctx->act_dev, buf->buf, &req);
2366 err = alloc_mem(ctx, &req, flags, alloc_pnext, &buf->flags, &buf->mem);
2370 ret = vkBindBufferMemory(hwctx->act_dev, buf->buf, buf->mem, 0);
2371 if (ret != VK_SUCCESS) {
2372 av_log(ctx, AV_LOG_ERROR, "Failed to bind memory to buffer: %s\n",
2375 return AVERROR_EXTERNAL;
2381 static int map_buffers(AVHWDeviceContext *ctx, ImageBuffer *buf, uint8_t *mem[],
2382 int nb_buffers, int invalidate)
2385 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2386 VkMappedMemoryRange invalidate_ctx[AV_NUM_DATA_POINTERS];
2387 int invalidate_count = 0;
2389 for (int i = 0; i < nb_buffers; i++) {
2390 ret = vkMapMemory(hwctx->act_dev, buf[i].mem, 0,
2391 VK_WHOLE_SIZE, 0, (void **)&mem[i]);
2392 if (ret != VK_SUCCESS) {
2393 av_log(ctx, AV_LOG_ERROR, "Failed to map buffer memory: %s\n",
2395 return AVERROR_EXTERNAL;
2402 for (int i = 0; i < nb_buffers; i++) {
2403 const VkMappedMemoryRange ival_buf = {
2404 .sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE,
2405 .memory = buf[i].mem,
2406 .size = VK_WHOLE_SIZE,
2408 if (buf[i].flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)
2410 invalidate_ctx[invalidate_count++] = ival_buf;
2413 if (invalidate_count) {
2414 ret = vkInvalidateMappedMemoryRanges(hwctx->act_dev, invalidate_count,
2416 if (ret != VK_SUCCESS)
2417 av_log(ctx, AV_LOG_WARNING, "Failed to invalidate memory: %s\n",
2424 static int unmap_buffers(AVHWDeviceContext *ctx, ImageBuffer *buf,
2425 int nb_buffers, int flush)
2429 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2430 VkMappedMemoryRange flush_ctx[AV_NUM_DATA_POINTERS];
2431 int flush_count = 0;
2434 for (int i = 0; i < nb_buffers; i++) {
2435 const VkMappedMemoryRange flush_buf = {
2436 .sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE,
2437 .memory = buf[i].mem,
2438 .size = VK_WHOLE_SIZE,
2440 if (buf[i].flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)
2442 flush_ctx[flush_count++] = flush_buf;
2447 ret = vkFlushMappedMemoryRanges(hwctx->act_dev, flush_count, flush_ctx);
2448 if (ret != VK_SUCCESS) {
2449 av_log(ctx, AV_LOG_ERROR, "Failed to flush memory: %s\n",
2451 err = AVERROR_EXTERNAL; /* We still want to try to unmap them */
2455 for (int i = 0; i < nb_buffers; i++)
2456 vkUnmapMemory(hwctx->act_dev, buf[i].mem);
2461 static int transfer_image_buf(AVHWDeviceContext *ctx, AVVkFrame *frame,
2462 ImageBuffer *buffer, const int *buf_stride, int w,
2463 int h, enum AVPixelFormat pix_fmt, int to_buf)
2466 AVVulkanDeviceContext *hwctx = ctx->hwctx;
2467 VulkanDevicePriv *s = ctx->internal->priv;
2470 VkPipelineStageFlagBits sem_wait_dst[AV_NUM_DATA_POINTERS];
2472 const int planes = av_pix_fmt_count_planes(pix_fmt);
2473 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(pix_fmt);
2475 VkCommandBufferBeginInfo cmd_start = {
2476 .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
2477 .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
2480 VkImageMemoryBarrier img_bar[AV_NUM_DATA_POINTERS] = { 0 };
2482 VkSubmitInfo s_info = {
2483 .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
2484 .commandBufferCount = 1,
2485 .pCommandBuffers = &s->cmd.buf,
2486 .pSignalSemaphores = frame->sem,
2487 .pWaitSemaphores = frame->sem,
2488 .pWaitDstStageMask = sem_wait_dst,
2489 .signalSemaphoreCount = planes,
2490 .waitSemaphoreCount = planes,
2493 ret = vkBeginCommandBuffer(s->cmd.buf, &cmd_start);
2494 if (ret != VK_SUCCESS) {
2495 av_log(ctx, AV_LOG_ERROR, "Unable to init command buffer: %s\n",
2497 return AVERROR_EXTERNAL;
2500 /* Change the image layout to something more optimal for transfers */
2501 for (int i = 0; i < planes; i++) {
2502 VkImageLayout new_layout = to_buf ? VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL :
2503 VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
2504 VkAccessFlags new_access = to_buf ? VK_ACCESS_TRANSFER_READ_BIT :
2505 VK_ACCESS_TRANSFER_WRITE_BIT;
2507 sem_wait_dst[i] = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
2509 /* If the layout matches and we have read access skip the barrier */
2510 if ((frame->layout[i] == new_layout) && (frame->access[i] & new_access))
2513 img_bar[bar_num].sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
2514 img_bar[bar_num].srcAccessMask = 0x0;
2515 img_bar[bar_num].dstAccessMask = new_access;
2516 img_bar[bar_num].oldLayout = frame->layout[i];
2517 img_bar[bar_num].newLayout = new_layout;
2518 img_bar[bar_num].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
2519 img_bar[bar_num].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
2520 img_bar[bar_num].image = frame->img[i];
2521 img_bar[bar_num].subresourceRange.levelCount = 1;
2522 img_bar[bar_num].subresourceRange.layerCount = 1;
2523 img_bar[bar_num].subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
2525 frame->layout[i] = img_bar[bar_num].newLayout;
2526 frame->access[i] = img_bar[bar_num].dstAccessMask;
2532 vkCmdPipelineBarrier(s->cmd.buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
2533 VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
2534 0, NULL, 0, NULL, bar_num, img_bar);
2536 /* Schedule a copy for each plane */
2537 for (int i = 0; i < planes; i++) {
2538 const int p_w = i > 0 ? AV_CEIL_RSHIFT(w, desc->log2_chroma_w) : w;
2539 const int p_h = i > 0 ? AV_CEIL_RSHIFT(h, desc->log2_chroma_h) : h;
2540 VkBufferImageCopy buf_reg = {
2542 /* Buffer stride isn't in bytes, it's in samples, the implementation
2543 * uses the image's VkFormat to know how many bytes per sample
2544 * the buffer has. So we have to convert by dividing. Stupid.
2545 * Won't work with YUVA or other planar formats with alpha. */
2546 .bufferRowLength = buf_stride[i] / desc->comp[i].step,
2547 .bufferImageHeight = p_h,
2548 .imageSubresource.layerCount = 1,
2549 .imageSubresource.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
2550 .imageOffset = { 0, 0, 0, },
2551 .imageExtent = { p_w, p_h, 1, },
2555 vkCmdCopyImageToBuffer(s->cmd.buf, frame->img[i], frame->layout[i],
2556 buffer[i].buf, 1, &buf_reg);
2558 vkCmdCopyBufferToImage(s->cmd.buf, buffer[i].buf, frame->img[i],
2559 frame->layout[i], 1, &buf_reg);
2562 ret = vkEndCommandBuffer(s->cmd.buf);
2563 if (ret != VK_SUCCESS) {
2564 av_log(ctx, AV_LOG_ERROR, "Unable to finish command buffer: %s\n",
2566 return AVERROR_EXTERNAL;
2569 /* Wait for the download/upload to finish if uploading, otherwise the
2570 * semaphore will take care of synchronization when uploading */
2571 ret = vkQueueSubmit(s->cmd.queue, 1, &s_info, s->cmd.fence);
2572 if (ret != VK_SUCCESS) {
2573 av_log(ctx, AV_LOG_ERROR, "Unable to submit command buffer: %s\n",
2575 return AVERROR_EXTERNAL;
2577 vkWaitForFences(hwctx->act_dev, 1, &s->cmd.fence, VK_TRUE, UINT64_MAX);
2578 vkResetFences(hwctx->act_dev, 1, &s->cmd.fence);
2584 /* Technically we can use VK_EXT_external_memory_host to upload and download,
2585 * however the alignment requirements make this unfeasible as both the pointer
2586 * and the size of each plane need to be aligned to the minimum alignment
2587 * requirement, which on all current implementations (anv, radv) is 4096.
2588 * If the requirement gets relaxed (unlikely) this can easily be implemented. */
2589 static int vulkan_transfer_data_from_mem(AVHWFramesContext *hwfc, AVFrame *dst,
2594 AVVkFrame *f = (AVVkFrame *)dst->data[0];
2595 AVHWDeviceContext *dev_ctx = hwfc->device_ctx;
2596 ImageBuffer buf[AV_NUM_DATA_POINTERS] = { { 0 } };
2597 const int planes = av_pix_fmt_count_planes(src->format);
2598 int log2_chroma = av_pix_fmt_desc_get(src->format)->log2_chroma_h;
2600 if ((src->format != AV_PIX_FMT_NONE && !av_vkfmt_from_pixfmt(src->format))) {
2601 av_log(hwfc, AV_LOG_ERROR, "Unsupported source pixel format!\n");
2602 return AVERROR(EINVAL);
2605 if (src->width > hwfc->width || src->height > hwfc->height)
2606 return AVERROR(EINVAL);
2608 /* For linear, host visiable images */
2609 if (f->tiling == VK_IMAGE_TILING_LINEAR &&
2610 f->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) {
2611 AVFrame *map = av_frame_alloc();
2613 return AVERROR(ENOMEM);
2614 map->format = src->format;
2616 err = vulkan_map_frame_to_mem(hwfc, map, dst, AV_HWFRAME_MAP_WRITE);
2620 err = av_frame_copy(map, src);
2621 av_frame_free(&map);
2625 /* Create buffers */
2626 for (int i = 0; i < planes; i++) {
2627 int h = src->height;
2628 int p_height = i > 0 ? AV_CEIL_RSHIFT(h, log2_chroma) : h;
2630 tmp.linesize[i] = src->linesize[i];
2631 err = create_buf(dev_ctx, &buf[i], p_height,
2632 &tmp.linesize[i], VK_BUFFER_USAGE_TRANSFER_SRC_BIT,
2633 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT, NULL, NULL);
2638 /* Map, copy image to buffer, unmap */
2639 if ((err = map_buffers(dev_ctx, buf, tmp.data, planes, 0)))
2642 av_image_copy(tmp.data, tmp.linesize, (const uint8_t **)src->data,
2643 src->linesize, src->format, src->width, src->height);
2645 if ((err = unmap_buffers(dev_ctx, buf, planes, 1)))
2648 /* Copy buffers to image */
2649 err = transfer_image_buf(dev_ctx, f, buf, tmp.linesize,
2650 src->width, src->height, src->format, 0);
2653 for (int i = 0; i < planes; i++)
2654 free_buf(dev_ctx, &buf[i]);
2659 static int vulkan_transfer_data_to(AVHWFramesContext *hwfc, AVFrame *dst,
2662 av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2664 switch (src->format) {
2666 case AV_PIX_FMT_CUDA:
2667 if ((p->extensions & EXT_EXTERNAL_FD_MEMORY) &&
2668 (p->extensions & EXT_EXTERNAL_FD_SEM))
2669 return vulkan_transfer_data_from_cuda(hwfc, dst, src);
2672 if (src->hw_frames_ctx)
2673 return AVERROR(ENOSYS);
2675 return vulkan_transfer_data_from_mem(hwfc, dst, src);
2680 static int vulkan_transfer_data_to_cuda(AVHWFramesContext *hwfc, AVFrame *dst,
2687 AVVkFrameInternal *dst_int;
2688 const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
2689 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
2691 AVHWFramesContext *cuda_fc = (AVHWFramesContext*)dst->hw_frames_ctx->data;
2692 AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
2693 AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
2694 AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
2695 CudaFunctions *cu = cu_internal->cuda_dl;
2697 ret = CHECK_CU(cu->cuCtxPushCurrent(cuda_dev->cuda_ctx));
2699 err = AVERROR_EXTERNAL;
2703 dst_f = (AVVkFrame *)src->data[0];
2705 err = vulkan_export_to_cuda(hwfc, dst->hw_frames_ctx, src);
2710 dst_int = dst_f->internal;
2712 for (int i = 0; i < planes; i++) {
2713 CUDA_MEMCPY2D cpy = {
2714 .dstMemoryType = CU_MEMORYTYPE_DEVICE,
2715 .dstDevice = (CUdeviceptr)dst->data[i],
2716 .dstPitch = dst->linesize[i],
2719 .srcMemoryType = CU_MEMORYTYPE_ARRAY,
2720 .srcArray = dst_int->cu_array[i],
2721 .WidthInBytes = (i > 0 ? AV_CEIL_RSHIFT(hwfc->width, desc->log2_chroma_w)
2722 : hwfc->width) * desc->comp[i].step,
2723 .Height = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, desc->log2_chroma_h)
2727 ret = CHECK_CU(cu->cuMemcpy2DAsync(&cpy, cuda_dev->stream));
2729 err = AVERROR_EXTERNAL;
2734 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2736 av_log(hwfc, AV_LOG_VERBOSE, "Transfered Vulkan image to CUDA!\n");
2741 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2742 vulkan_free_internal(dst_int);
2743 dst_f->internal = NULL;
2744 av_buffer_unref(&dst->buf[0]);
2749 static int vulkan_transfer_data_to_mem(AVHWFramesContext *hwfc, AVFrame *dst,
2754 AVVkFrame *f = (AVVkFrame *)src->data[0];
2755 AVHWDeviceContext *dev_ctx = hwfc->device_ctx;
2756 ImageBuffer buf[AV_NUM_DATA_POINTERS] = { { 0 } };
2757 const int planes = av_pix_fmt_count_planes(dst->format);
2758 int log2_chroma = av_pix_fmt_desc_get(dst->format)->log2_chroma_h;
2760 if (dst->width > hwfc->width || dst->height > hwfc->height)
2761 return AVERROR(EINVAL);
2763 /* For linear, host visiable images */
2764 if (f->tiling == VK_IMAGE_TILING_LINEAR &&
2765 f->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) {
2766 AVFrame *map = av_frame_alloc();
2768 return AVERROR(ENOMEM);
2769 map->format = dst->format;
2771 err = vulkan_map_frame_to_mem(hwfc, map, src, AV_HWFRAME_MAP_READ);
2775 err = av_frame_copy(dst, map);
2776 av_frame_free(&map);
2780 /* Create buffers */
2781 for (int i = 0; i < planes; i++) {
2782 int h = dst->height;
2783 int p_height = i > 0 ? AV_CEIL_RSHIFT(h, log2_chroma) : h;
2785 tmp.linesize[i] = dst->linesize[i];
2786 err = create_buf(dev_ctx, &buf[i], p_height,
2787 &tmp.linesize[i], VK_BUFFER_USAGE_TRANSFER_DST_BIT,
2788 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT, NULL, NULL);
2791 /* Copy image to buffer */
2792 if ((err = transfer_image_buf(dev_ctx, f, buf, tmp.linesize,
2793 dst->width, dst->height, dst->format, 1)))
2796 /* Map, copy buffer to frame, unmap */
2797 if ((err = map_buffers(dev_ctx, buf, tmp.data, planes, 1)))
2800 av_image_copy(dst->data, dst->linesize, (const uint8_t **)tmp.data,
2801 tmp.linesize, dst->format, dst->width, dst->height);
2803 err = unmap_buffers(dev_ctx, buf, planes, 0);
2806 for (int i = 0; i < planes; i++)
2807 free_buf(dev_ctx, &buf[i]);
2812 static int vulkan_transfer_data_from(AVHWFramesContext *hwfc, AVFrame *dst,
2815 av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2817 switch (dst->format) {
2819 case AV_PIX_FMT_CUDA:
2820 if ((p->extensions & EXT_EXTERNAL_FD_MEMORY) &&
2821 (p->extensions & EXT_EXTERNAL_FD_SEM))
2822 return vulkan_transfer_data_to_cuda(hwfc, dst, src);
2825 if (dst->hw_frames_ctx)
2826 return AVERROR(ENOSYS);
2828 return vulkan_transfer_data_to_mem(hwfc, dst, src);
2832 AVVkFrame *av_vk_frame_alloc(void)
2834 return av_mallocz(sizeof(AVVkFrame));
2837 const HWContextType ff_hwcontext_type_vulkan = {
2838 .type = AV_HWDEVICE_TYPE_VULKAN,
2841 .device_hwctx_size = sizeof(AVVulkanDeviceContext),
2842 .device_priv_size = sizeof(VulkanDevicePriv),
2843 .frames_hwctx_size = sizeof(AVVulkanFramesContext),
2844 .frames_priv_size = sizeof(VulkanFramesPriv),
2846 .device_init = &vulkan_device_init,
2847 .device_create = &vulkan_device_create,
2848 .device_derive = &vulkan_device_derive,
2850 .frames_get_constraints = &vulkan_frames_get_constraints,
2851 .frames_init = vulkan_frames_init,
2852 .frames_get_buffer = vulkan_get_buffer,
2853 .frames_uninit = vulkan_frames_uninit,
2855 .transfer_get_formats = vulkan_transfer_get_formats,
2856 .transfer_data_to = vulkan_transfer_data_to,
2857 .transfer_data_from = vulkan_transfer_data_from,
2859 .map_to = vulkan_map_to,
2860 .map_from = vulkan_map_from,
2862 .pix_fmts = (const enum AVPixelFormat []) {